diff --git a/.gitattributes b/.gitattributes index 118c229dee4e4533c53327de9ee3848cb3699f61..a140832fa522eb71f99d570aec3e58ca9b43f36e 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5212,3 +5212,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Video[[:space:]]Recognition[[:space:]]in[[:space:]]Portrait[[:space:]]Mode/1430817f-2fb2-4e8e-aaf4-dc2cf17e83e3_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Video[[:space:]]Super-Resolution[[:space:]]Transformer[[:space:]]with[[:space:]]Masked[[:space:]]Inter&Intra-Frame[[:space:]]Attention/1e69759c-d8e8-49da-aff4-8c9d96153686_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Video-Based[[:space:]]Human[[:space:]]Pose[[:space:]]Regression[[:space:]]via[[:space:]]Decoupled[[:space:]]Space-Time[[:space:]]Aggregation/f3738d37-ef8c-4352-b1e1-4943413c7dec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video-P2P_[[:space:]]Video[[:space:]]Editing[[:space:]]with[[:space:]]Cross-attention[[:space:]]Control/81ad8533-a31f-4ffc-839e-f07d4869e500_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video2Game_[[:space:]]Real-time[[:space:]]Interactive[[:space:]]Realistic[[:space:]]and[[:space:]]Browser-Compatible[[:space:]]Environment[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Video/5bc42c40-19c8-4162-b372-460cfcad74f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoBooth_[[:space:]]Diffusion-based[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Image[[:space:]]Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoCon_[[:space:]]Robust[[:space:]]Video-Language[[:space:]]Alignment[[:space:]]via[[:space:]]Contrast[[:space:]]Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoCrafter2_[[:space:]]Overcoming[[:space:]]Data[[:space:]]Limitations[[:space:]]for[[:space:]]High-Quality[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models/ab56d7df-62cd-4c60-8578-721928a2487b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoCutLER_[[:space:]]Surprisingly[[:space:]]Simple[[:space:]]Unsupervised[[:space:]]Video[[:space:]]Instance[[:space:]]Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoGrounding-DINO_[[:space:]]Towards[[:space:]]Open-Vocabulary[[:space:]]Spatio-Temporal[[:space:]]Video[[:space:]]Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoLLM-online_[[:space:]]Online[[:space:]]Video[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Streaming[[:space:]]Video/c25a98ca-87a4-45af-9e10-acd055b937ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoMAC_[[:space:]]Video[[:space:]]Masked[[:space:]]Autoencoders[[:space:]]Meet[[:space:]]ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoRF_[[:space:]]Rendering[[:space:]]Dynamic[[:space:]]Radiance[[:space:]]Fields[[:space:]]as[[:space:]]2D[[:space:]]Feature[[:space:]]Video[[:space:]]Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VideoSwap_[[:space:]]Customized[[:space:]]Video[[:space:]]Subject[[:space:]]Swapping[[:space:]]with[[:space:]]Interactive[[:space:]]Semantic[[:space:]]Point[[:space:]]Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/View[[:space:]]From[[:space:]]Above_[[:space:]]Orthogonal-View[[:space:]]aware[[:space:]]Cross-view[[:space:]]Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/View-Category[[:space:]]Interactive[[:space:]]Sharing[[:space:]]Transformer[[:space:]]for[[:space:]]Incomplete[[:space:]]Multi-View[[:space:]]Multi-Label[[:space:]]Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/View-decoupled[[:space:]]Transformer[[:space:]]for[[:space:]]Person[[:space:]]Re-identification[[:space:]]under[[:space:]]Aerial-ground[[:space:]]Camera[[:space:]]Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViewDiff_[[:space:]]3D-Consistent[[:space:]]Image[[:space:]]Generation[[:space:]]with[[:space:]]Text-to-Image[[:space:]]Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViewFusion_[[:space:]]Towards[[:space:]]Multi-View[[:space:]]Consistency[[:space:]]via[[:space:]]Interpolated[[:space:]]Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Viewpoint-Aware[[:space:]]Visual[[:space:]]Grounding[[:space:]]in[[:space:]]3D[[:space:]]Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Virtual[[:space:]]Immunohistochemistry[[:space:]]Staining[[:space:]]for[[:space:]]Histological[[:space:]]Images[[:space:]]Assisted[[:space:]]by[[:space:]]Weakly-supervised[[:space:]]Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Vision-and-Language[[:space:]]Navigation[[:space:]]via[[:space:]]Causal[[:space:]]Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Anagrams_[[:space:]]Generating[[:space:]]Multi-View[[:space:]]Optical[[:space:]]Illusions[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Concept[[:space:]]Connectome[[:space:]](VCC)_[[:space:]]Open[[:space:]]World[[:space:]]Concept[[:space:]]Discovery[[:space:]]and[[:space:]]their[[:space:]]Interlayer[[:space:]]Connections[[:space:]]in[[:space:]]Deep[[:space:]]Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Delta[[:space:]]Generator[[:space:]]with[[:space:]]Large[[:space:]]Multi-modal[[:space:]]Models[[:space:]]for[[:space:]]Semi-supervised[[:space:]]Composed[[:space:]]Image[[:space:]]Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Fact[[:space:]]Checker_[[:space:]]Enabling[[:space:]]High-Fidelity[[:space:]]Detailed[[:space:]]Caption[[:space:]]Generation/41baf37b-e81b-4a82-a385-a78a616c7868_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]In-Context[[:space:]]Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Layout[[:space:]]Composer_[[:space:]]Image-Vector[[:space:]]Dual[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Design[[:space:]]Layout[[:space:]]Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Objectification[[:space:]]in[[:space:]]Films_[[:space:]]Towards[[:space:]]a[[:space:]]New[[:space:]]AI[[:space:]]Task[[:space:]]for[[:space:]]Video[[:space:]]Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Point[[:space:]]Cloud[[:space:]]Forecasting[[:space:]]enables[[:space:]]Scalable[[:space:]]Autonomous[[:space:]]Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Program[[:space:]]Distillation_[[:space:]]Distilling[[:space:]]Tools[[:space:]]and[[:space:]]Programmatic[[:space:]]Reasoning[[:space:]]into[[:space:]]Vision-Language[[:space:]]Models/d3d264df-7a0f-49c2-be47-325ff84855b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Programming[[:space:]]for[[:space:]]Zero-shot[[:space:]]Open-Vocabulary[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual[[:space:]]Prompting[[:space:]]for[[:space:]]Generalized[[:space:]]Few-shot[[:space:]]Segmentation_[[:space:]]A[[:space:]]Multi-scale[[:space:]]Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Visual-Augmented[[:space:]]Dynamic[[:space:]]Semantic[[:space:]]Prototype[[:space:]]for[[:space:]]Generative[[:space:]]Zero-Shot[[:space:]]Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VkD_[[:space:]]Improving[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]using[[:space:]]Orthogonal[[:space:]]Projections/260ea0d3-c32f-4988-98e9-ada076a48353_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Vlogger_[[:space:]]Make[[:space:]]Your[[:space:]]Dream[[:space:]]A[[:space:]]Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VoCo_[[:space:]]A[[:space:]]Simple-yet-Effective[[:space:]]Volume[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]3D[[:space:]]Medical[[:space:]]Image[[:space:]]Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Volumetric[[:space:]]Environment[[:space:]]Representation[[:space:]]for[[:space:]]Vision-Language[[:space:]]Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WALT3D_[[:space:]]Generating[[:space:]]Realistic[[:space:]]Training[[:space:]]Data[[:space:]]from[[:space:]]Time-Lapse[[:space:]]Imagery[[:space:]]for[[:space:]]Reconstructing[[:space:]]Dynamic[[:space:]]Objects[[:space:]]Under[[:space:]]Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WANDR_[[:space:]]Intention-guided[[:space:]]Human[[:space:]]Motion[[:space:]]Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WHAM_[[:space:]]Reconstructing[[:space:]]World-grounded[[:space:]]Humans[[:space:]]with[[:space:]]Accurate[[:space:]]3D[[:space:]]Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WOUAF_[[:space:]]Weight[[:space:]]Modulation[[:space:]]for[[:space:]]User[[:space:]]Attribution[[:space:]]and[[:space:]]Fingerprinting[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/92d9125f-3522-4960-9c33-6d71a67f7675_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WWW_[[:space:]]A[[:space:]]Unified[[:space:]]Framework[[:space:]]for[[:space:]]Explaining[[:space:]]What[[:space:]]Where[[:space:]]and[[:space:]]Why[[:space:]]of[[:space:]]Neural[[:space:]]Networks[[:space:]]by[[:space:]]Interpretation[[:space:]]of[[:space:]]Neuron[[:space:]]Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WateRF_[[:space:]]Robust[[:space:]]Watermarks[[:space:]]in[[:space:]]Radiance[[:space:]]Fields[[:space:]]for[[:space:]]Protection[[:space:]]of[[:space:]]Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Watermark-embedded[[:space:]]Adversarial[[:space:]]Examples[[:space:]]for[[:space:]]Copyright[[:space:]]Protection[[:space:]]against[[:space:]]Diffusion[[:space:]]Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WaveFace_[[:space:]]Authentic[[:space:]]Face[[:space:]]Restoration[[:space:]]with[[:space:]]Efficient[[:space:]]Frequency[[:space:]]Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WaveMo_[[:space:]]Learning[[:space:]]Wavefront[[:space:]]Modulations[[:space:]]to[[:space:]]See[[:space:]]Through[[:space:]]Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Wavelet-based[[:space:]]Fourier[[:space:]]Information[[:space:]]Interaction[[:space:]]with[[:space:]]Frequency[[:space:]]Diffusion[[:space:]]Adjustment[[:space:]]for[[:space:]]Underwater[[:space:]]Image[[:space:]]Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weak-to-Strong[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]X-Ray[[:space:]]Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weakly[[:space:]]Misalignment-free[[:space:]]Adaptive[[:space:]]Feature[[:space:]]Alignment[[:space:]]for[[:space:]]UAVs-based[[:space:]]Multimodal[[:space:]]Object[[:space:]]Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weakly[[:space:]]Supervised[[:space:]]Monocular[[:space:]]3D[[:space:]]Detection[[:space:]]with[[:space:]]a[[:space:]]Single-View[[:space:]]Image/e8def396-7aa9-4446-a151-70f8d14a292d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weakly[[:space:]]Supervised[[:space:]]Point[[:space:]]Cloud[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]via[[:space:]]Artificial[[:space:]]Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weakly[[:space:]]Supervised[[:space:]]Video[[:space:]]Individual[[:space:]]Counting/357e37d5-1951-48de-9fbc-83246528162f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weakly-Supervised[[:space:]]Audio-Visual[[:space:]]Video[[:space:]]Parsing[[:space:]]with[[:space:]]Prototype-based[[:space:]]Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Weakly-Supervised[[:space:]]Emotion[[:space:]]Transition[[:space:]]Learning[[:space:]]for[[:space:]]Diverse[[:space:]]3D[[:space:]]Co-speech[[:space:]]Gesture[[:space:]]Generation/82233940-816a-43f0-81c1-3331488ee413_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/What[[:space:]]Do[[:space:]]You[[:space:]]See[[:space:]]in[[:space:]]Vehicle_[[:space:]]Comprehensive[[:space:]]Vision[[:space:]]Solution[[:space:]]for[[:space:]]In-Vehicle[[:space:]]Gaze[[:space:]]Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/What[[:space:]]How[[:space:]]and[[:space:]]When[[:space:]]Should[[:space:]]Object[[:space:]]Detectors[[:space:]]Update[[:space:]]in[[:space:]]Continually[[:space:]]Changing[[:space:]]Test[[:space:]]Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/What[[:space:]]If[[:space:]]the[[:space:]]TV[[:space:]]Was[[:space:]]Off_[[:space:]]Examining[[:space:]]Counterfactual[[:space:]]Reasoning[[:space:]]Abilities[[:space:]]of[[:space:]]Multi-modal[[:space:]]Language[[:space:]]Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/What[[:space:]]Sketch[[:space:]]Explainability[[:space:]]Really[[:space:]]Means[[:space:]]for[[:space:]]Downstream[[:space:]]Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/What[[:space:]]When[[:space:]]and[[:space:]]Where_[[:space:]]Self-Supervised[[:space:]]Spatio-Temporal[[:space:]]Grounding[[:space:]]in[[:space:]]Untrimmed[[:space:]]Multi-Action[[:space:]]Videos[[:space:]]from[[:space:]]Narrated[[:space:]]Instructions/59de70a6-beec-4552-9d70-f207c0df32db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/What[[:space:]]You[[:space:]]See[[:space:]]is[[:space:]]What[[:space:]]You[[:space:]]GAN_[[:space:]]Rendering[[:space:]]Every[[:space:]]Pixel[[:space:]]for[[:space:]]High-Fidelity[[:space:]]Geometry[[:space:]]in[[:space:]]3D[[:space:]]GANs/2c134100-ed06-4b65-8341-cd48f2b74244_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/When[[:space:]]StyleGAN[[:space:]]Meets[[:space:]]Stable[[:space:]]Diffusion_[[:space:]]a[[:space:]]W+[[:space:]]Adapter[[:space:]]for[[:space:]]Personalized[[:space:]]Image[[:space:]]Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/When[[:space:]]Visual[[:space:]]Grounding[[:space:]]Meets[[:space:]]Gigapixel-level[[:space:]]Large-scale[[:space:]]Scenes_[[:space:]]Benchmark[[:space:]]and[[:space:]]Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Why[[:space:]]Not[[:space:]]Use[[:space:]]Your[[:space:]]Textbook_[[:space:]]Knowledge-Enhanced[[:space:]]Procedure[[:space:]]Planning[[:space:]]of[[:space:]]Instructional[[:space:]]Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WildlifeMapper_[[:space:]]Aerial[[:space:]]Image[[:space:]]Analysis[[:space:]]for[[:space:]]Multi-Species[[:space:]]Detection[[:space:]]and[[:space:]]Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/WinSyn_[[:space:]]_[[:space:]]A[[:space:]]High[[:space:]]Resolution[[:space:]]Testbed[[:space:]]for[[:space:]]Synthetic[[:space:]]Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Wired[[:space:]]Perspectives_[[:space:]]Multi-View[[:space:]]Wire[[:space:]]Art[[:space:]]Embraces[[:space:]]Generative[[:space:]]AI/2a286243-d96a-406c-bb14-a19955ff2a60_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_content_list.json b/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7902759e2e156c5b1efcdc99618a5a4010f6a8c9 --- /dev/null +++ b/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_content_list.json @@ -0,0 +1,2998 @@ +[ + { + "type": "text", + "text": "Video-P2P: Video Editing with Cross-attention Control", + "text_level": 1, + "bbox": [ + 205, + 130, + 763, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaoteng Liu", + "bbox": [ + 148, + 179, + 263, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuechen Zhang", + "bbox": [ + 307, + 180, + 439, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenbo Li", + "bbox": [ + 483, + 180, + 570, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhe Lin3", + "bbox": [ + 614, + 180, + 687, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaya Jia1,2", + "bbox": [ + 730, + 180, + 818, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong 2SmartMore 3 Adobe", + "bbox": [ + 228, + 198, + 738, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83784f0f7af0642761ccb26b1d60a832c3369678821e7de39fa7f11de19df8c9.jpg", + "image_caption": [ + "[Input Video] A child is riding a bike on the road.", + "Figure 1. Video-P2P generates new characters while optimally maintaining the pose and environment in videos." + ], + "image_footnote": [], + "bbox": [ + 84, + 247, + 885, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 652, + 313, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-P2P is the first framework for real-world video editing with cross-attention control. While attention control has proven effective for image editing with pre-trained image generation models, there are currently no large-scale video generation models publicly available. Video-P2P addresses this limitation by adapting an image generation diffusion model to complete various video editing tasks. Specifically, we propose to first tune a Text-to-Set (T2S) model to complete an approximate inversion and then optimize a shared unconditional embedding to achieve accurate video inversion with a small memory cost. We further prove that it is crucial for consistent video editing. For attention control, we introduce a novel decoupled-guidance strategy, which uses different guidance strategies for the source and target prompts. The optimized unconditional embedding for", + "bbox": [ + 75, + 675, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the source prompt improves reconstruction ability, while an initialized unconditional embedding for the target prompt enhances editability. Incorporating the attention maps of these two branches enables detailed editing. These technical designs enable various text-driven editing applications, including word swap, prompt refinement, and attention reweighting. Video-P2P works well on real-world videos for generating new characters while optimally preserving their original poses and scenes. It significantly outperforms previous approaches.", + "bbox": [ + 496, + 654, + 893, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 816, + 632, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video creation and editing are key tasks [16, 17, 25, 37, 42]. Text-driven editing becomes one promising pipeline. Several methods have demonstrated the ability to edit generated or real-world images with text prompts [13, 19, 24].", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8599", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[Input Video] A penguin is running on the ice.", + "bbox": [ + 140, + 89, + 392, + 101 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b37b6980737bf96d85a327052f5d39fb3d04acb9d28bde6473d41972e8e0de2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 104, + 171, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b93c05dfa17a99af84a4f64c2ac9ef4918987ae77712d476dd8eb1ec1564ec3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 104, + 269, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f6a423f5441f3998f6074973130d25cfc81007b59aee8133d02a582ec56348e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 104, + 364, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/429ccc9f43f35c525c434644027b2825fdc78dcba46dead1a9035a6a949d9bc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 104, + 460, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[Image-P2P] A Robotic penguin is running on the ice.", + "bbox": [ + 133, + 175, + 421, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c0d6d3ff201311d99c4d451a6b02eacc96865fb29825406bc21f7ab2e688f91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 189, + 171, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3ab97766f73b0a33fd19a59604f56644a2fd68d53c8a6fd7bf124eddb4e070b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 188, + 269, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cad2ca82be901f2723d6cd79c2c6c53b1abc9b713a2f14c00a60540881c58182.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 188, + 364, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c906a9290bd92858a094f4bcbf2169d410f7aa2a750feee31c843c11180a00dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 188, + 460, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[Video-P2P] A Robotic penguin is running on the ice.", + "bbox": [ + 133, + 258, + 416, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39c6eb391474c84ec39901ff43d5440df94bf674ae61234100001570524dbe6b.jpg", + "image_caption": [ + "Figure 2. Video-P2P vs Image-P2P. Editing a video frame-by-frame (Image-P2P) cannot guarantee semantic consistency across frames. Video-P2P enables changing the penguin into the same robotic type in every frame." + ], + "image_footnote": [], + "bbox": [ + 84, + 272, + 171, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/77922d7b571d70ac66c3223449fbb30850500f481e26cdb1d888959b775461b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 272, + 269, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1ef304f30049add608b912260b996b12b6a3e4a15b76d22462b3c0e197afacf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 272, + 364, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c067812dd4a8a7b41a7396ef114ef469c93d82907696721247e312b3e1d0d998.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 272, + 460, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Till now, it is still challenging to edit only local objects in a video, such as changing a running \"dog\" into a \"cat\" without influencing the environment. This paper proposes a pipeline that can edit a video both locally and globally, as shown in Figs. 1 and 5.", + "bbox": [ + 75, + 426, + 468, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-driven image editing requires a model capable of generating target content, such as changing the category or property of an object. Diffusion models have demonstrated outstanding generation capabilities in this area [3, 13, 19, 40]. Among these methods, attention control emerges as the most effective pipeline for detailed image editing [13, 24]. In order to edit a real image, this pipeline includes two necessary steps: (1) inverting the image into latent features with a pre-trained diffusion model, and (2) controlling attention maps in the denoising process to edit the corresponding parts of the image. For example, by swapping their attention maps, we can replace a \"child\" with a \"panda\".", + "bbox": [ + 75, + 503, + 470, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to build an attention control-based pipeline for video editing. Since no large-scale pre-trained video generation models are publicly available, we propose a novel framework to show that a pre-trained image diffusion model can be adapted for detailed video editing.", + "bbox": [ + 75, + 686, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While a pre-trained image diffusion model can be utilized for video editing by processing frames individually (Image-P2P), it lacks semantic consistency across frames (the 2nd row of Fig. 2). To maintain semantic consistency, we propose using a structure on inversion and attention control for all frames, by transforming the Text-to-image diffusion model (T2I) into a Text-to-set model (T2S). This approach is effective, as illustrated in the 3rd row, where the robotic penguin maintains its consistency across frames.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We adopt the method proposed in [47] to convert a T2I model into a T2S model by altering the convolution kernels and replacing the self-attention with frame-attentions. This conversion yields a model that generates a set of semantically consistent images. The generation quality will be degraded with the inflation step but it can be recovered after tuning on the original video. Although the tuned T2S model is not an ideal video generation model, it suffices to create an approximate inversion for a video as shown in Fig. 3 (c). It is just an approximation because errors are accumulated in the denoising pass, consistent with conclusions in [24, 43].", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the inversion quality, we propose to optimize a shared unconditional embedding for all frames to align the denoising latent features with the diffusion latent features. Our experiments show that shared embedding is efficient for video inversion and crucial for video editing. Comparisons are shown in Fig. 3 and Fig. 9.", + "bbox": [ + 496, + 289, + 890, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As discussed in [13], successful attention control requires a model to have both reconstruction ability and editability. While image inversion has been argued to possess both abilities in [24], we find that video editing presents different challenges. The T2S model, as an inflation model not trained on any videos, is not robust to the perturbations caused by various unconditional embeddings. Although our optimized embedding can achieve reconstruction, changing prompts can destabilize the model and result in a low-quality generation. On the other hand, we find that the approximate inversion with an initialized unconditional embedding (the text embedding of an empty prompt) is editable but cannot reconstruct well. To address this issue, we propose a decoupled-guidance strategy in attention control, utilizing different guidance strategies for the source and target prompts. Specifically, we use the optimized unconditional embedding for the source prompt and the initialized unconditional embedding for the target prompt. We incorporate the attention maps from these two branches to generate the target video. These two simple designs prove effective and successfully complete video editing. Our contributions can be summarized as:", + "bbox": [ + 496, + 396, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first framework for video editing with attention control. A decoupled-guidance strategy is designed to further improve performance.", + "- We introduce an effective and efficient video inversion method with shared unconditional embedding optimization to improve video editing substantially.", + "- We conduct extensive ablation studies and comparisons to show the effectiveness of our video editing framework." + ], + "bbox": [ + 500, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8600", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb07cff4b777789dd136a51f13c5a28d4eaae1f761c91acbf382fc162bdcc82d.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 84, + 103, + 277, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e7fdadcd8c74bc72fa00a8af6bf73f17a2403c9e52a9c5102c95a41fdb7028e8.jpg", + "image_caption": [ + "(b) DDIM", + "Figure 3. Inversion Comparison. (b) The inflated model cannot generate high-quality results. (c) Tuning the model can create an approximate video inversion. (d) Optimizing a shared unconditional embedding can accurately reconstruct the input video." + ], + "image_footnote": [], + "bbox": [ + 287, + 103, + 480, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9df2f657e8072787d9f9ef0ba23cffd0e0cb61f8178ed334849a358bad59e968.jpg", + "image_caption": [ + "(c) Tune+DDIM" + ], + "image_footnote": [], + "bbox": [ + 488, + 103, + 681, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/437e49b274129baf479c0f39ed63117bdfdf3c0b5b8650f976234841d9b1869d.jpg", + "image_caption": [ + "(d) Ours (Tune+Shared Null)" + ], + "image_footnote": [], + "bbox": [ + 692, + 103, + 883, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 234, + 218, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Text Driven Generation", + "text_level": 1, + "bbox": [ + 76, + 260, + 294, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DALL-E [32] first considers the text-to-image (T2I) generation task as a sequence-to-sequence translation problem, with subsequent research improving generation quality [6, 10, 50]. Denoising Diffusion Probabilistic Models (DDPMs)[15] have recently gained popularity for T2I. GLIDE[26] utilizes classifier-free guidance to improve text conditioning. DALLE-2 [33] leverages CLIP [31] for better text-image alignment. Latent Diffusion Models (LDMs) [34] propose processing in the latent space to enhance training efficiency. In our work, we employ a pretrained image diffusion model based on LDMs.", + "bbox": [ + 75, + 285, + 468, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-video (T2V) generation is a nascent research area. GODIVA [46] first introduces VQ-VAE [41] to T2V. CogVideo [18] combines T2V with CogView-2 [6], utilizing pre-trained text-to-image models. Video Diffusion Models (VDM) [17] propose a space-time U-Net for performing diffusion on pixels. Imagen Video [16] successfully generates high-quality videos with cascaded diffusion models and v-prediction parameterization. Phenaki [42] generates videos with time-variable prompts. Make-A-Video [37] combines the appearance generation of T2I models with movement information from video data. While these approaches generate reasonable short videos, they still contain artifacts and do not support real-world video editing. Additionally, most of these approaches are not publicly available at this time. Tune-A-Video [47] inflates an image diffusion model into a video model and tunes it to reconstruct the input video. It allows for changes in semantic content but with limited temporal consistency. We find that using DDIM inversion results can improve its temporal consistency. However, it cannot avoid altering unrelated regions. We adapt some designs of TAV to do our model initialization.", + "bbox": [ + 75, + 453, + 470, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text Driven Editing", + "text_level": 1, + "bbox": [ + 76, + 800, + 264, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative models have demonstrated impressive performance in image editing, with approaches ranging from GANs [12, 27, 28, 44] to diffusion models [1, 19]. SDEdit [22] adds noise to an input image and uses the diffusion process to recover an edited version. Prompt-to", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt [13] uses attention control to minimize changes to unrelated parts, while Null-Text Inversion [24] improves real image editing. Plug-and-Play [40] uses feature and self-attention injection, which are unable to avoid editing the whole image when changing a local part. InstructPix2Pix [3] enables flexible text-driven editing with user-provided instructions. Textual Inversion [11], DreamBooth [35], and Custom-Diffusion [20] learn special tokens for personalized concepts and generate related images. Some improved inversion methods [7, 23] are proposed after our work.", + "bbox": [ + 496, + 234, + 890, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video editing with generative models has seen several advances recently. Text2Live [2] employs CLIP to edit textures in videos but struggles with significant semantic changes. Dreamix [25] uses a pre-trainedImagen Video [16] backbone to perform image-to-video and video-to-video editing, with the ability to change motion as well. Gen-1 [8] trains models jointly on images and videos for tasks such as stylization and customization. While these methods enable modifying video content, they operate like guided generation and tend to modify all regions together when editing an object. Our proposed method allows for local editing with a diffusion model pre-trained on images.", + "bbox": [ + 496, + 404, + 892, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides, there have been some concurrent works and followers after our work, such as [5, 9, 21, 30, 45, 49, 52]. Video-P2P focuses more on introducing the image diffusion models to detailed video editing, instead of conditional video generation.", + "bbox": [ + 496, + 589, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 686, + 589, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{V}$ be a real video containing $n$ frames. We adopt the Prompt-to-Prompt setting by introducing a source prompt $\\mathcal{P}$ and an edited prompt $\\mathcal{P}^*$ which together generate an edited video $\\mathcal{V}^*$ containing $n$ frames. The prompts are provided by the user. Similar to TAV [47], we assume that the object of interest is present in the first frame.", + "bbox": [ + 496, + 715, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve cross-attention control in video editing, we propose Video-P2P, a framework with two key technical designs: (1) optimizing a shared unconditional embedding for video inversion, and (2) using different guidance for the source and edited prompts, and incorporating their attention maps. The framework is illustrated in Fig. 4.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 90, + 207, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs). LDMs generate an image latent $z_{0}$ using a random noise vector $z_{t}$ and a textual condition $P$ as inputs. As variants of DDPMs, these models aim to predict artificial noise by minimizing the following objective:", + "bbox": [ + 75, + 117, + 468, + 194 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} E _ {z _ {0}, \\varepsilon \\sim N (0, I), t \\sim \\text {U n i f o r m} (1, T)} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 202, + 468, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C} = \\psi(\\mathcal{P})$ is the embedding of the text prompt, and noise $\\varepsilon$ is added to $z_0$ according to step $t$ to obtain $z_t$ . During inference, the model predicts noise $\\varepsilon_\\theta(\\cdot)$ for $T$ steps to generate an image from $z_T$ .", + "bbox": [ + 75, + 234, + 468, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDIM sampling and inversion. Deterministic DDIM sampling can be used to generate an image from latent features in a small number of denoising steps:", + "bbox": [ + 76, + 301, + 468, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\varepsilon_ {\\theta} (z _ {t}, t, \\mathcal {C}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 354, + 468, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use an encoder to encode the real image before the diffusion process and a decoder to decode after the denoising process. DDIM sampling can be reversed in a few steps through the equation:", + "bbox": [ + 76, + 402, + 468, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\varepsilon_ {\\theta} (z _ {t}, t, \\mathcal {C}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 472, + 468, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "known as DDIM inversion [38]. This can be used to obtain the corresponding latent features of a real image.", + "bbox": [ + 75, + 520, + 468, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Null-text inversion. To mitigate the amplification effect of text conditioning during image generation, classifier-free guidance is proposed, which performs unconditional prediction [14]:", + "bbox": [ + 75, + 554, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = w \\cdot \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) + (1 - w) \\cdot \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 626, + 467, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varnothing = \\psi(\"\")$ is the embedding of a null text and $w$ is the guidance weight. However, the classifier-free guidance increases errors accumulated in the denoising process, leading to imperfect image reconstruction using the DDIM inversion. [24] proposes to align the diffusion latent trajectory $z_{T}^{*},\\ldots ,z_{0}^{*}$ with the denoising latent trajectory $z_{T},\\ldots ,z_{0}$ by optimizing a step-wise unconditional embedding $\\varnothing_{t}$ :", + "bbox": [ + 75, + 656, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\varnothing_ {t}} \\left\\| z _ {t - 1} ^ {*} - z _ {t - 1} \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 772, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video Inversion", + "text_level": 1, + "bbox": [ + 76, + 816, + 233, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by constructing a T2S model that is capable of performing an approximate inversion. Following the VDM baselines [16, 17] and TAV [47], we employ $1 \\times 3 \\times 3$ pattern convolution kernels and temporal attention. Moreover, we", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9b9c12e560e4ca378f1536dfc72f7ecbdc928da9faf8afdce1b59ee5adc6b31.jpg", + "image_caption": [ + "Figure 4. Framework. We optimize one shared unconditional embedding for the reconstruct branch (orange). The initialized unconditional embedding is utilized for the editable branch (green). Their attention maps are incorporated to create the target video." + ], + "image_footnote": [], + "bbox": [ + 504, + 92, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "replace the self-attention with frame-attentions, which take the first frames $v_{0}$ and the current frame $v_{i}$ as inputs and update features for the frame $v_{i}$ . The formulation of the frame-attention is as follows:", + "bbox": [ + 496, + 353, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = W ^ {Q} \\mathbf {v} _ {i}, K = W ^ {K} \\mathbf {v} _ {0}, V = W ^ {V} \\mathbf {v} _ {0}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 420, + 890, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W$ are the projection matrices in attention. Then, attention maps are calculated as:", + "bbox": [ + 496, + 446, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 482, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ is the latent projection dimension. The model processes a video pair-by-pair and computes $n$ times to obtain the prediction for every frame. We find that the simple design suffices for video inversion since the reversed latent features can capture temporal information. Besides, frame-attention conserves memory and speeds up the process.", + "bbox": [ + 496, + 523, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While model inflation can aid in preserving semantic consistency across frames, it adversely impacts the generation quality of the T2I model. This is because the self-attention parameters are utilized to compute frame correlations, which have not been pre-trained. Consequently, the T2S model, generated through inflation, is insufficient for the approximate inversion, as demonstrated in Fig. 2. To address this, we fine-tune the query projection matrices $W^{Q}$ of the frame- and cross-attentions, as well as additional temporal attention, to perform noise prediction based on the input video following [47]. After this initialization, the T2S model is capable of generating semantically consistent image sets while maintaining the quality of each frame, resulting in successful approximate inversion.", + "bbox": [ + 496, + 613, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the fine-tuned T2S model, we perform video inversion by optimizing a shared unconditional embedding. During inversion, each latent feature $z_{t}$ contains a channel for the frames with dimension $n$ , where $z_{t,i}$ denotes the latent feature for the $i$ -th frame. We employ DDIM inversion", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8602", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to generate latent features $z_0^*, \\ldots, z_T^*$ . The unconditional embedding is defined as follows:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\varnothing_ {t}} \\sum_ {i = 1} ^ {n} \\left| \\left| z _ {t - 1, i} ^ {*} - z _ {t - 1, i} \\left(\\bar {z} _ {t, i}, \\bar {z} _ {t, 0}, \\varnothing_ {t}, \\mathcal {C}\\right) \\right| \\right| _ {2} ^ {2}, \\text {w h e r e} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 128, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {t - 1, i} = z _ {t - 1, i} \\left(\\bar {z} _ {t, i}, \\bar {z} _ {t, 0}, \\emptyset_ {t}, \\mathcal {C}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 179, + 468, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is updated at each step. The T2S model's frame-attentions use two latent features to calculate the corresponding feature for the next step. Notice $\\varnothing_{t}$ is shared by all frames $(i = 1,\\dots ,n)$ which minimizes the memory usage. Besides, using the same unconditional embedding for all frames avoids destabilizing the semantic consistency in attention control.", + "bbox": [ + 76, + 199, + 468, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Decoupled-guidance Attention Control", + "text_level": 1, + "bbox": [ + 76, + 297, + 410, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To perform attention control on real images, existing works [13, 24] require an inference pipeline with both reconstruction ability and editability. However, achieving such a pipeline for a T2S model is challenging. Video inversion allows us to establish an inference pipeline to reconstruct the original video well. However, the T2S model is not as robust as T2I models due to a lack of pre-training with videos. As a result, its editability is compromised with the optimized unconditional embedding, leading to degraded generation quality when changing prompts. In contrast, we find that using an initialized unconditional embedding $\\varnothing$ makes the model more editable while it cannot reconstruct perfectly. This inspires us to combine the abilities of two inference pipelines. For the source prompt, we use the optimized unconditional embedding in the classifier-free guidance. For the target prompt, we choose the initialized unconditional embedding. We then incorporate attention maps from these two branches to obtain the edited video, where the unchanged parts are influenced by the source branch and the edited parts are influenced by the target branch.", + "bbox": [ + 75, + 320, + 468, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pseudo algorithm is shown in Alg. 1. We adopt the attention control methods from Image-P2P to Video-P2P. For example, to perform word swap, the Edit function can be represented as:", + "bbox": [ + 75, + 622, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE d i t \\left(M _ {t}, M _ {t} ^ {*}, t\\right) := \\left\\{ \\begin{array}{l l} M _ {t} ^ {*} & \\text {i f} t < \\tau \\\\ M _ {t} & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 689, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$M_{t}$ and $M_{t}^{*}$ are the cross-attention maps for every frame at every step, and $DM$ is the tuned T2S model. Changing the frame-attention maps has a small influence on the final results. Attention maps are swapped only for the first $\\tau$ steps because attentions are formed in the early period. $\\overline{M}_{t,w}$ is the average attention map of the word $w$ calculated at step $t$ . It is averaged over steps $T,\\ldots ,t$ independently for every frame. For the $j$ -th frame, we calculate:", + "bbox": [ + 76, + 734, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {M} _ {t, w, j} = \\frac {1}{T - t} \\sum_ {i = t} ^ {T} \\bar {M} _ {i, w, j} \\quad j = 1, \\dots , n. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 863, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Prompt-to-Prompt video editing" + ], + "code_body": "1 Input: A source prompt $\\mathcal{P}$ a target prompt $\\mathcal{P}^*$ source video $V_{src}$ \n2 Output: Edited video $V_{dst}$ \n3 Latent features from DDIM inversion: $z_{T}$ \n4 $z_{T}^{*}\\gets z_{T}$ \n5 Initialized unconditional embedding $\\varnothing^{*}$ and optimized unconditional embedding $\\varnothing$ . \n6 for $t = T,T - 1,\\ldots ,1$ do \n7 $z_{t - 1},M_t\\gets DM(z_t,\\mathcal{P},t,\\emptyset)$ \n8 $M_t^* \\leftarrow DM(z_t^*,\\mathcal{P}^*,t,\\emptyset^*)$ \n9 $\\widehat{M}_t\\gets \\operatorname {Edit}(M_t,M_t^*,t)$ \n10 $z_{t - 1}^{*}\\gets DM(z_{t}^{*},\\mathcal{P}^{*},t,\\emptyset^{*})\\{M_{t}^{*}\\gets \\widehat{M}_{t}\\}$ \n11 $\\alpha \\gets B(\\overline{M}_{t,w})\\cup B(\\overline{M}_{t,w}^{*})$ \n12 $z_{t - 1}^{*}\\gets (1 - \\alpha)\\odot z_{t - 1} + \\alpha \\odot z_{t - 1}^{*}$ \n13 end \n14 Return $(z_0,z_0^*)$", + "bbox": [ + 506, + 113, + 834, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$B\\left(\\overline{M}_{t,w}\\right)$ represents the binary mask obtained from the attention map. A value is set to 1 when larger than a threshold.", + "bbox": [ + 500, + 383, + 890, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 428, + 630, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 453, + 715, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We develop our method based on CompVis Stable Diffusion (v1-5). Similar to TAV [47], we fix the image autoencoder and sample 8 or 24 frames at the resolution of $512 \\times 512$ from a video. To initialize the model, we fine-tune the T2S model for 500 steps to reconstruct the original video. During attention control, we set the cross-attention replacing ratio to 0.4 and the attention threshold to 0.3. For prompt refinement, we set the refinement ratio to 0.4. Parameters can be adjusted to control the editing fidelity for different examples. All 8-frame experiments are conducted on a single V100 GPU, with 5 minutes for initialization (tuning), 6 minutes for inversion, and 1 minute for inference. It is difficult to avoid the inversion cost when preserving the background faithfully according to [24]. To speed up to less than one minute, one can replace the null-text inversion with recent inversion methods [7, 23]. In the appendix, we show that a shared unconditional embedding works on 24 frames. Our inversion process extends to over 100 frames. However, editing longer videos (1min) remains challenging. VideoP2P and most existing works concentrate on shorter videos.", + "bbox": [ + 496, + 474, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Applications", + "text_level": 1, + "bbox": [ + 500, + 786, + 633, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video-P2P can be utilized for a range of editing applications, including prompt refinement, attention re-weighting, and word swapping. Video-P2P is able to maintain semantic consistency across different frames and preserve the temporal coherence of the original video during the editing process. More examples can be found in the appendix.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8603", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[Input Video] A car is driving on the road.", + "bbox": [ + 84, + 89, + 339, + 102 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d041f0dc691e815d07d43bb27ac3d4388932856f29f2c8767775e4a1a61036ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 104, + 181, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/424b100a8c1eff47e2d61b38e5422e7e94eb92ec8699ebc48c79275008374eab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 104, + 279, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/466a4d96970a114f7fa2ab8fdcf444b9ef667a5e84615b0cf4cc9e707b658f6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 104, + 380, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89c0496cf3125144f4efa6c4d4846fba160be89950a59e061de466be9d803890.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 104, + 482, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A car is driving on the flooded road.", + "bbox": [ + 84, + 181, + 299, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/526e11e454db70c56885a81ec7c68bd2e02d47ea3dd52e4ff8eef900d2291696.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 196, + 179, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/542b1a64aa5dc9e6892e5717be1c39e6209a98c2fabff79f76f5e01b75e3267d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 196, + 279, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ce6484cc75d9b254c1ebb01583e4b2478efbfc8fea02e5d1687c8a09adec475.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 196, + 380, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee2db35a88659841986f543fe31609d400fce813c17a67ae86c92d92a9fbc720.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 196, + 482, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Input Video] A man drives a motorbike in the forest.", + "bbox": [ + 84, + 277, + 406, + 290 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/450e6c5c0d5b59204572be19b7838591492fcbd5298c024b550dc21fb7eb40b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 292, + 179, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d018e862a6d53557a848da1367e00f2c867156159411515e8d943f844b1a978.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 292, + 279, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4eda895ca1663d2ff13acff297a478ce69f2f2d691c5baf2bd8a1e499ea15211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 292, + 380, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee7f7b5fde13fd494fc2c3cd331b5d599d0154482a719072766efda644df1399.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 292, + 482, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A man Bat Man drives a motorbike in the forest.", + "bbox": [ + 84, + 369, + 372, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8a9af957f6359e04aba3601a2dae483d4e3984da2d769ca7047b71b1499347e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 382, + 179, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d61b009e0be12b74b2071726c3a0f03a78cce723ebd3207c91337c0785804faa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 382, + 279, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62f1de7a0f07a89ea7eeecce0010a7de316a06080b4529be255e158d2a451506.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 382, + 380, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ed69f4db463c83b5ede3a0484259d05aee2273625fd7df42751a24a5829d6c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 382, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Input Video] A dog is running on the grass.", + "bbox": [ + 84, + 467, + 352, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31e2ed7092a4f51d00dc41353bef6c3e5cdec06692bded3ca6189c354b8bd335.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 481, + 179, + 554 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d23fa563541f6e0f41956f1bd86883b0c472adb351e26d930dc1da575e93abe1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 481, + 279, + 554 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c241ba8214ee7849a4de823ba1d386cddc0ec81b77199cbe40faa65b2430e320.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 481, + 380, + 554 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9aeaaa0dcde9fbfc4555d2a07e0c47c76d799fd563b8b3abcf77750d30c87777.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 481, + 482, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A robotic dog is running on the grass.", + "bbox": [ + 84, + 558, + 313, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8df7d95a573c3577bb5c625f4f4cb28c7729c4a406a727daff8d92bf0480c982.jpg", + "image_caption": [ + "Figure 5. Videos edited by Video-P2P with text prompts. Video-P2P can do both word swaps and prompt refinement." + ], + "image_footnote": [], + "bbox": [ + 84, + 573, + 179, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bd63a0690d6e3a9a1367b06b81b2d653a4d5c6ca82d90f3e2e52e34968042f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 574, + 279, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0938b734dd91c8c58dc07c4a80d575f39745ff41699115c11f014d12fa5be87e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 574, + 380, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbd8d637c0642b7f918bef4749d72518611f4c2d10f4c7a8b7c826532908aefd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 574, + 482, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A car is driving on the road railway.", + "bbox": [ + 488, + 90, + 702, + 102 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c0561a773a7ad13eba901e9981878279c92894b7144e3bbe1499c6138e844c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 104, + 584, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29c081042cc8388248e695d5517cde1129ecf5d5ad4bf9a7b9426596d2c71ae8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 104, + 684, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70f87d2b08a5002336c7dd58f76466b8a4d8ca8982a0e3e21c1ff9270355c147.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 104, + 785, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/885f9c8e9ae5858269988a7610bf7c7cbf452af985c929fa0174d95b96b9d9e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 104, + 885, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A car is driving on the road at sunset.", + "bbox": [ + 491, + 181, + 715, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc667d6683e2643ad402feac4a1604edacd2a8f4d4a38492ce997d17e653266e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 196, + 584, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38779350364a4198756abaee8242f7ba825a07e65f7834aee885ffabc7a50925.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 196, + 684, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4b28a78f4b1e69b5d9200496e7d91cf1211c71b00e52539ebc538089dfef04d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 196, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f6144bfdd5f849dde10a2db9439c60aa083276237339ab558a28c25c13913c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 196, + 885, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A man drives a Lego motorbike in the forest.", + "bbox": [ + 488, + 276, + 754, + 289 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21e2aa76875d25aed32e4f31ba398c6ff8e183c8755d19ead1741589407bf69d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 292, + 584, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17ef09c21f078b5e003af01a14afb06e3ffbd54997180c011cbfadf3658ea0fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 292, + 684, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4137e19e1ae31754272c2de9a8010a19f70a53df2fab79a007fffa52ec8763ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 292, + 785, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40e9cb9081a2ba7698eced0a87dff02dc9d0595c59e02e408a621ba68188bffb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 292, + 885, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A man Spider Man drives a motorbike in the forest.", + "bbox": [ + 488, + 369, + 792, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c41fd342c23d913120c85d6070e697852bfe03aa6e027d6c34dd3a9c954ec39f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 382, + 584, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6533b27b862229ad1f1f31aacc7cca91fb53e02e352a96fd27393b4b0a92c172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 382, + 684, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb8e620954ea7c5c7dd18584ef72b8b264f93c55b5b536990f67335f7b5ac6a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 382, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20a50f60ecb91869ef762bcdba895c738cfb57265161360d0aa0325b80b8af99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 382, + 885, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A dog cat is running on the grass.", + "bbox": [ + 488, + 465, + 694, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b233b6f264d011190749631b76dd0bf1102b113211cb9e0c06bf06d9def44609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 481, + 584, + 555 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd46e9e85d1e6a625a30ac03e9200a5a17276eb2a22f02945dd2b94cea2e6c4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 481, + 684, + 555 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9335c60824c9d469c767134cfab8445bc5a53d69061b3eab1a466937de72e4bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 481, + 785, + 555 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3fc67b288ac36e0c0fd21a3239eb2fd7f583636be608f0db9d081b5c5331cf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 481, + 885, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A fluffy $\\uparrow$ dog is running on the grass.", + "bbox": [ + 488, + 558, + 714, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55e82f37575debf34f5527138a68a5c576ad5b5a62d585e1b9c903f57c99cf42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 574, + 584, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50b820aba89d2cb15d3c9ab424c57af90e732f765732413cbb91dfdc4c711d86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 574, + 684, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2bfc7b2dc05fdec65b5915f4ee8adfa5096383e284de366eb8c01b9d8f5537f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 574, + 785, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a87f466a72fc87f915cf31f5784f6ff037e21b097b963c0b35eb5ab57676f52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 574, + 885, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Word swap. Video-P2P enables the replacement of entities based on word swapping while maintaining the coherence of unrelated regions. As illustrated in Fig. 5, Video-P2P seamlessly replaces the man on the motorbike with Spider-Man while minimizing the changes to the motorbike's appearance (the 4th row). The generated Spider-Man exhibits a consistent appearance across frames, and the background remains unchanged. Furthermore, we can replace a dog with a cat while preserving its gesture and the surrounding grass (the 5th row).", + "bbox": [ + 75, + 686, + 470, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt refinement. Video-P2P is able to do prompt refinement, such as modifying object properties. For example, we can transform the running dog into a robotic one (the", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6th row in Fig. 5), and convert a motorbike into a Lego toy with the same motion (the 3rd row). Notice the grass and sky are almost not influenced. Additionally, Video-P2P can perform global editing like changing the weather to sunset or flooding the road with water (2nd row). Style transfer can also be accomplished by Video-P2P, as exemplified by transforming the video into a watercolor painting.", + "bbox": [ + 496, + 686, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attention re-weighting. Similar to Image-P2P, Video-P2P also enables attention re-weighting. By adjusting the cross-attention of specific words, we can manipulate the extent of the corresponding generation. For instance, we can regulate how fluffy a dog is in the video (the 6th row of Fig. 5).", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8604", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73b34461955180fb977c1aa7f225be47abc7959ac0e8eb7c578e2cd5d4903ed4.jpg", + "image_caption": [ + "[TAV+DDIM] King-Kong is roaring." + ], + "image_footnote": [], + "bbox": [ + 83, + 102, + 173, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b19744c9cd86d509b8804e9ed804d8adf8ea2648fb369ee6163072d40b28b36b.jpg", + "image_caption": [ + "[Input Video] A lion is roaring." + ], + "image_footnote": [], + "bbox": [ + 183, + 102, + 272, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24f0601bbf8180941b1e4ad2ae01433a4ceb308127f0bbc99b57688c4a158df1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 102, + 372, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9592dfa5fb777d4bd5f446ae6bcddd175b4d89596b561a56b5c05c00a275d4c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 102, + 473, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17b3ef8917c72755dc3d0066b0e76618c27df64c6cdc2bd3e9cc3ff5981725ab.jpg", + "image_caption": [ + "[Video-P2P] King-Kong is roaring." + ], + "image_footnote": [], + "bbox": [ + 83, + 186, + 173, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc64099e94ce3a956a5ceffab8cf4250ac4d4aeea2578648a0b5f2efb5981508.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 186, + 274, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9449b45a0a9567af94c930c382ad8bc1e45ed5b80523ce6a66bd9d50a5afb29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 186, + 372, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e842132d49e2d25a900fadeba48bb13caee16835dfbdd8ed0e72d8872a095325.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 186, + 473, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f64dc1cc4265792ca7f709725c7ad3ec6b5e01264dfa19dcc6ecc15674996d4e.jpg", + "image_caption": [ + "Figure 6. Video-P2P v.s. Tune-A-Video (TAV). Video-P2P offers the ability to edit content locally, while TAV+DDIM cannot avoid influencing unrelated regions." + ], + "image_footnote": [], + "bbox": [ + 83, + 272, + 173, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f8337cc8ab4d0c97df70aa754f7de5e165fd7951bec03dc0ce3afc286ff01f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 272, + 272, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79760badc0cf109ed6153b54a52c7a38690752895dfededcfb9fbced70308b61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 272, + 372, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/382ba663ac22251733534e7307ef727e9862e79811d40e33e42d802451bbb2a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 272, + 473, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4522730b2ed31cc1cca85a7bcd0d6af825fd6b5a7136c7469050d4b93cb205f.jpg", + "image_caption": [ + "[TAV+DDIM] A Lego tiger is walking in the forest." + ], + "image_footnote": [], + "bbox": [ + 496, + 102, + 586, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d008b7946f8122401c3fe16308379bc92517a56241128c7c87e14fdf427cda99.jpg", + "image_caption": [ + "[Input Video] A tiger is walking in the forest." + ], + "image_footnote": [], + "bbox": [ + 594, + 103, + 687, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50a6e8885e4ca64df6bfdf3c7e60ee0a7f68cb9d28a4922b46581a10844cd08a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 103, + 785, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b439aa2bc86b7ce4ad3f42daefa8b7916780eb85392e8e8904fe4840cd052c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 103, + 885, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7782192d9c05e64cda6725ce3d84c515aefc8eb806e14805cfdf274e0fec6943.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 186, + 586, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/190248bd0954fb26d448252f9922ad667205186b116c46487bcbbadf82f69eec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 186, + 687, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/076c1115f5f76469ef84b5a6a83aa6514462306d8f8115ed7bf073e2de036b0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 186, + 785, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63ba69ad9298ab3044dbf95a65c11833f9b0dc8ca8bb8dba05487e7d65f87d9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 186, + 885, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51975775d897c93dd1b7e7c5812018a75922dc77ee0b5c7f2c4e1cc3592f9f7e.jpg", + "image_caption": [ + "[Video-P2P] A Lego tiger is walking in the forest." + ], + "image_footnote": [], + "bbox": [ + 496, + 272, + 586, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcc7a614e32dc413920c3ea1bc131a2ce1087be8ad7a8bcd4124822786c4d7d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 272, + 687, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82495e4a28e42d7b325a974dbc6033d8b6b18e5b11952aa99898b2d3cdf013a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 272, + 785, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c028d4891a5a3bffbd29516bc107010da27052ce5ad63c5fa3703e88f6e8e0e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 272, + 885, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Input Video] a brown dog and a white dog on the kitchen floor.", + "bbox": [ + 88, + 391, + 464, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/213e52013709e7a91620e907ec629fc1de890b81d3f2450b163e981e977d7018.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 407, + 176, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97c6802bd66efb6f675affed8023f0c8abea1a33c075ae1269a7d3756edf57b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 407, + 276, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c6700b3759eb09a9b0eab10d35c19b3558526d4054c79f0058559c7262bfb0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 407, + 375, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98339615b3b21c121f796ce67b78058e5e22772c89e3fa23d0ea9e9ca6aeac02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 407, + 462, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c4b3a02753a59567a9e5497b22d90aa46b44aff47bafbf2013245a8555f6990.jpg", + "image_caption": [ + "[Dreamix] a brown cat and a white cat on the kitchen floor." + ], + "image_footnote": [], + "bbox": [ + 81, + 472, + 176, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a98e31f799ca3ce95990d8a6940b37f52628f827d62715eef13a368cf5feba25.jpg", + "image_caption": [ + "[Video-P2P] a brown cat and a white cat on the kitchen floor." + ], + "image_footnote": [], + "bbox": [ + 181, + 472, + 276, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2dcbfc1b2fb50ba8069a0b4262ed1ee097ffd8e94607f5dc1df383827dc77897.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 472, + 375, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7931fb2f1c4e456ffac76f676cc98d4711df3fa0425a83b9cb07992e0a4de6c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 472, + 462, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8037d67c72d3f815b6e0c47c39f81da7a9f740cfb779c84dd6fe840d990fc5ff.jpg", + "image_caption": [ + "Figure 7. Video-P2P v.s. Dreamix. Both methods can change the dogs to cats. Video-P2P can preserve background's details." + ], + "image_footnote": [], + "bbox": [ + 81, + 537, + 176, + 584 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4330f52597e182425ee890153dcf8d0ff448678efe91c3fa302919007b1206a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 537, + 274, + 584 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9d15f53ba02e1af9d095394ce0ee360f706b54a8eb2716da5d8715728727356.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 537, + 375, + 584 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1bbb4b0354897c5b69534c203038cfdbcfb06fd2db7f60c2846cd22a32ede99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 537, + 462, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison", + "text_level": 1, + "bbox": [ + 76, + 625, + 209, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Tune-A-Video. Both TAV+DDIM [47] and our Video-P2P allow for video editing with text prompts. However, TAV+DDIM cannot avoid altering the entire video content when editing specific objects, while Video-P2P can edit a local area and minimize the influence on other regions. Fig. 6 (Left) demonstrates that Video-P2P preserves the complex shape of the cloud when replacing a lion with King Kong, whereas TAV+DDIM can only maintain the color tone of the sky in this case.", + "bbox": [ + 75, + 655, + 468, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although our model initialization is similar to TAV, Video-P2P can still generate temporal-consistent results where TAV+DDIM fails. As demonstrated in Fig. 6 (Right), TAV struggles to generate a temporally consistent sequence in the second row, even when the inputs are features from DDIM inversion. In contrast, our method can produce better structure-preserved results, as shown in the third row.", + "bbox": [ + 75, + 794, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Dreamix. In contrast to Dreamix [25], which uses a pre-trained video diffusion model that is not publicly available, our method yields superior results for subject replacement. Although our method cannot perform video motion editing due to the lack of temporal priors, we outperform Dreamix in preserving details and motion consistency. As it is not open-sourced, we conducted our evaluation on its released demo. As demonstrated in Fig. 7, both methods can transform two dogs into two cats, but our method preserves the details of the drawer in the background (the 3rd row). Furthermore, Dreamix may affect the time sequence to some extent, as the generated cat moves more slowly than the original dog in the video. Our method completely preserves the motion of the original video.", + "bbox": [ + 496, + 393, + 890, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. We create a test set of 45 samples based on the DAVIS [29] dataset and report five metrics for quantitative analysis. The CLIP Score measures the textual similarity between the text prompt and video, while Masked PSNR and LPIPS [51] evaluate the quality of structure preservation. We also proposed a novel metric, Object Semantic Variance (OSV), to measure semantic consistency across frames. In addition, we report an average CLIP similarity (Temp) between consecutive frames (as in Gen-1 [8] and FateZero [30]) to evaluate the temporal coherence. For detailed explanations of these metrics, please refer to the appendix. Our results, as shown in Table 1, demonstrate that Video-P2P performs well on all metrics. Compared to TAV+DDIM, Video-P2P achieves higher Masked PSNR and lower LPIPS, indicating better preservation of unchanged regions. Compared to the other two methods, Video-P2P has much lower OSV and Temp, indicating its superior ability to maintain semantic and temporal consistency across frames. When comparing to Image-P2P, we", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63f13cb8eb4b0de5cfb2f6ab703cd4770241a7f5f65ece74054a0e6cd63a6550.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 101, + 176, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20ae2dd82cce5c0158053af32ffd6702a0b78bc6c79217896315e5c560a6f114.jpg", + "image_caption": [ + "A dog-cat is running on the grass." + ], + "image_footnote": [], + "bbox": [ + 178, + 101, + 272, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81e707be5ed6e1e121f94b9035ca81f74932db19e29fa59caeec1f258f7897aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 101, + 369, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ef134070c02eab4e058de8e5c502a80675853cdd2ada111a0733108898f91b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 101, + 465, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb35601c5af7e22c08ba850afdfc7c1f1fc59e585782830772771c739f8853cc.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 81, + 159, + 176, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f3658fa5fe44419984fafeb680a0cd1ba9307eb888761eb85d2d3a91962b663.jpg", + "image_caption": [ + "Image-P2P" + ], + "image_footnote": [], + "bbox": [ + 178, + 159, + 272, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9a5d8d30685c62a97f0cba66c3529862cf207800c4236aaf61faad0ad46c5d6.jpg", + "image_caption": [ + "w/o model init" + ], + "image_footnote": [], + "bbox": [ + 274, + 159, + 369, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41d19275b1d888db267b3c33fe9650b2e453093a5946327f7759b8d6b5658b26.jpg", + "image_caption": [ + "w/ model init" + ], + "image_footnote": [], + "bbox": [ + 372, + 159, + 465, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0b298690c88cbbb66acaf1b791dd345bd7a4dedf09a306fc38ce495e12a6e76.jpg", + "image_caption": [ + "Figure 8. Model initialization improves the generation quality.", + "Figure 9. Ablation on shared unconditional embeddings." + ], + "image_footnote": [], + "bbox": [ + 78, + 255, + 467, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e45c31d65878decc4ae696d1408b4d427a2874b76f1d56fb22a641a0131d344e.jpg", + "image_caption": [ + "Figure 10. Ablation on decoupled-guidance attention control." + ], + "image_footnote": [], + "bbox": [ + 81, + 328, + 467, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consider it as a reference point rather than a direct competitor because it does not account for the consistency between different frames. Image-P2P represents the upper bound of PSNR and LPIPS, which serve as metrics to evaluate per-frame quality. Among all video editing methods, Video-P2P is the best one. Moreover, in Tab. 3, we report the user study results, where Video-P2P ranks first on average and has a high preference rate compared to other methods.", + "bbox": [ + 75, + 407, + 468, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 541, + 228, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model initialization. While the inflated image diffusion model can generate semantically consistent images, the T2S model's generation ability is compromised during inflation. As seen in Fig.8 (the 3rd column), directly using the inflated T2S model produces unrealistic results with an inaccurate background. To mitigate this, we initialize the T2S model by fine-tuning the given video. This is evident in Fig.8 (4th column), where the cat's appearance improves, and the grass reconstruction becomes more accurate.", + "bbox": [ + 75, + 561, + 468, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shared unconditional embedding. The use of shared unconditional embeddings is crucial to ensure stable and consistent generations. On our test set, we found that $10\\%$ of the videos exhibit instability when multiple unconditional embeddings are employed. A visual example is included in Fig. 9 (Right). As shown in Tab. 1 (row 5), semantic and temporal consistency are diminished without shared unconditional embedding, reminiscent of the content code in [39]. We observe that optimizing a shared unconditional embedding can significantly improve the PSNR compared to TAV+DDIM in Tab. 2. However, using multiple unconditional embeddings for each frame only increases the PSNR by 0.2 but results in higher parameter usage ( $n$ times)", + "bbox": [ + 75, + 704, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/898483488d5eae74cf60aa996f6b4f96a1069b9ede66ba057d98b1440192b093.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Temp ↑
TAV+DDIM0.334317.370.465155.670.9697
Image-P2P0.327222.920.308276.920.9021
Ours (w/o MI)0.321320.620.325550.270.9460
Ours (w/o DG)0.322618.970.386668.750.9451
Ours (w/o SU)0.335120.640.322148.780.9665
Ours0.336720.650.321347.380.9725
", + "bbox": [ + 508, + 88, + 883, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/82479551bb7bfc56151fb901b621786b40680cf2ccfee854f74f79ea1e526ec2.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation. We evaluate textual similarity (CLIP), region preservation (Masked PSNR, LPIPS), semantic consistency (OSV), and temporal consistency (Temp). DG and SU mean Decoupled-Guidance and shared unconditional embeddings." + ], + "table_footnote": [], + "table_body": "
VQVAETAV +DDIMMulti- uncondShared- uncond
PSNR(dB) ↑24.7315.4322.9722.75
#Param. ↓/0.13M22.68M2.94M
", + "bbox": [ + 501, + 234, + 893, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c42e757a53e7ee892c2b241e0490cf47f1c21c5065a2c57848d9e19e8d72ecce.jpg", + "table_caption": [ + "Table 2. Video reconstruction quality. A shared unconditional embedding can reconstruct a high-quality video with a small size." + ], + "table_footnote": [], + "table_body": "
Image-P2PTAVTAV+DDIMVideo-P2P
Structure2.673.332.611.39
Preserving13.59%6.52%10.87%69.02%
Text3.402.782.281.54
Alignment3.80%14.13%19.57%62.50%
Realism & Quality3.382.982.211.43
4.35%7.61%19.02%69.02%
", + "bbox": [ + 501, + 324, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. User study of average ranking $\\downarrow$ and preference rate $\\uparrow$ . and leads to a lower Masked PSNR after attention control. Thus, we conclude that shared unconditional embedding is the most effective and efficient method for video inversion.", + "bbox": [ + 500, + 426, + 890, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoupled-guidance attention control. To obtain the latent features of the input video, we optimize an unconditional embedding using the source prompt. It is important to note that this embedding is only suitable for the source prompt during the prompt-to-prompt process. Using the optimized embedding for the target prompt may negatively impact the quality of the generated results, as shown in Fig. 10 (Left). Instead, we utilize the initialized unconditional embedding for the target prompt and incorporate attention maps from two branches. The decoupled-guidance attention control approach significantly improves the editing quality, as shown in Fig.10 (Right). Quantitative ablations can be found in Tab. 1 (the 4th row and 6th row).", + "bbox": [ + 496, + 494, + 890, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 704, + 617, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video-P2P provides the first framework for video editing with cross-attention control. We optimize a shared unconditional embedding based on a well-initialized T2S model for video inversion. We also propose the decoupled-guidance strategy for attention control. These techniques enable Video-P2P to perform various applications, such as word swap, prompt refinement, and attention re-weighting.", + "bbox": [ + 496, + 729, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement Supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8606", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3", + "[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723, 2022. 3, 2", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 3", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1", + "[5] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3", + "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3", + "[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. arXiv preprint arXiv:2305.04441, 2023. 3, 5", + "[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3, 7", + "[9] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3", + "[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106, 2022. 3", + "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3", + "[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), pages 1-13, 2022. 3", + "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 5", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3", + "[16] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3, 4", + "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1, 3, 4", + "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3", + "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 3", + "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3", + "[21] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3", + "[22] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3", + "[23] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3, 5", + "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1, 2, 3, 4, 5", + "[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 3, 7", + "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3", + "[28] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF Inter" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Conference on Computer Vision, pages 2085-2094, 2021.3", + "[29] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7", + "[30] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3, 7", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 3", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831, 2021. 3", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3", + "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4", + "[39] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 8", + "[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 3", + "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017. 3", + "[42] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3", + "[43] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2", + "[44] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 3", + "[45] Wen Wang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3", + "[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3", + "[47] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 5, 7", + "[48] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In ECCV, 2022. 1", + "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3", + "[50] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 3", + "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7, 1", + "[52] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 712 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8608", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_model.json b/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22e444936eb85e2eaf1553c7d464d3843101df5f --- /dev/null +++ b/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_model.json @@ -0,0 +1,3597 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.764, + 0.152 + ], + "angle": 0, + "content": "Video-P2P: Video Editing with Cross-attention Control" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.18, + 0.264, + 0.198 + ], + "angle": 0, + "content": "Shaoteng Liu" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.181, + 0.441, + 0.199 + ], + "angle": 0, + "content": "Yuechen Zhang" + }, + { + "type": "text", + "bbox": [ + 0.485, + 0.181, + 0.571, + 0.198 + ], + "angle": 0, + "content": "Wenbo Li" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.181, + 0.689, + 0.198 + ], + "angle": 0, + "content": "Zhe Lin3" + }, + { + "type": "text", + "bbox": [ + 0.732, + 0.181, + 0.82, + 0.198 + ], + "angle": 0, + "content": "Jiaya Jia1,2" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.199, + 0.74, + 0.216 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong 2SmartMore 3 Adobe" + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.235, + 0.615, + 0.248 + ], + "angle": 0, + "content": "[Input Video] A child is riding a bike on the road." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.248, + 0.887, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.619, + 0.816, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Video-P2P generates new characters while optimally maintaining the pose and environment in videos." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.654, + 0.314, + 0.669 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.676, + 0.474, + 0.904 + ], + "angle": 0, + "content": "Video-P2P is the first framework for real-world video editing with cross-attention control. While attention control has proven effective for image editing with pre-trained image generation models, there are currently no large-scale video generation models publicly available. Video-P2P addresses this limitation by adapting an image generation diffusion model to complete various video editing tasks. Specifically, we propose to first tune a Text-to-Set (T2S) model to complete an approximate inversion and then optimize a shared unconditional embedding to achieve accurate video inversion with a small memory cost. We further prove that it is crucial for consistent video editing. For attention control, we introduce a novel decoupled-guidance strategy, which uses different guidance strategies for the source and target prompts. The optimized unconditional embedding for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.807 + ], + "angle": 0, + "content": "the source prompt improves reconstruction ability, while an initialized unconditional embedding for the target prompt enhances editability. Incorporating the attention maps of these two branches enables detailed editing. These technical designs enable various text-driven editing applications, including word swap, prompt refinement, and attention reweighting. Video-P2P works well on real-world videos for generating new characters while optimally preserving their original poses and scenes. It significantly outperforms previous approaches." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.817, + 0.633, + 0.832 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Video creation and editing are key tasks [16, 17, 25, 37, 42]. Text-driven editing becomes one promising pipeline. Several methods have demonstrated the ability to edit generated or real-world images with text prompts [13, 19, 24]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.142, + 0.09, + 0.393, + 0.102 + ], + "angle": 0, + "content": "[Input Video] A penguin is running on the ice." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.105, + 0.173, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.105, + 0.27, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.105, + 0.366, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.105, + 0.462, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.176, + 0.422, + 0.188 + ], + "angle": 0, + "content": "[Image-P2P] A Robotic penguin is running on the ice." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.19, + 0.173, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.189, + 0.27, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.189, + 0.365, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.189, + 0.462, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.26, + 0.418, + 0.272 + ], + "angle": 0, + "content": "[Video-P2P] A Robotic penguin is running on the ice." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.273, + 0.173, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.273, + 0.27, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.273, + 0.365, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.273, + 0.462, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Video-P2P vs Image-P2P. Editing a video frame-by-frame (Image-P2P) cannot guarantee semantic consistency across frames. Video-P2P enables changing the penguin into the same robotic type in every frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Till now, it is still challenging to edit only local objects in a video, such as changing a running \"dog\" into a \"cat\" without influencing the environment. This paper proposes a pipeline that can edit a video both locally and globally, as shown in Figs. 1 and 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.687 + ], + "angle": 0, + "content": "Text-driven image editing requires a model capable of generating target content, such as changing the category or property of an object. Diffusion models have demonstrated outstanding generation capabilities in this area [3, 13, 19, 40]. Among these methods, attention control emerges as the most effective pipeline for detailed image editing [13, 24]. In order to edit a real image, this pipeline includes two necessary steps: (1) inverting the image into latent features with a pre-trained diffusion model, and (2) controlling attention maps in the denoising process to edit the corresponding parts of the image. For example, by swapping their attention maps, we can replace a \"child\" with a \"panda\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.764 + ], + "angle": 0, + "content": "In this paper, we aim to build an attention control-based pipeline for video editing. Since no large-scale pre-trained video generation models are publicly available, we propose a novel framework to show that a pre-trained image diffusion model can be adapted for detailed video editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "While a pre-trained image diffusion model can be utilized for video editing by processing frames individually (Image-P2P), it lacks semantic consistency across frames (the 2nd row of Fig. 2). To maintain semantic consistency, we propose using a structure on inversion and attention control for all frames, by transforming the Text-to-image diffusion model (T2I) into a Text-to-set model (T2S). This approach is effective, as illustrated in the 3rd row, where the robotic penguin maintains its consistency across frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "We adopt the method proposed in [47] to convert a T2I model into a T2S model by altering the convolution kernels and replacing the self-attention with frame-attentions. This conversion yields a model that generates a set of semantically consistent images. The generation quality will be degraded with the inflation step but it can be recovered after tuning on the original video. Although the tuned T2S model is not an ideal video generation model, it suffices to create an approximate inversion for a video as shown in Fig. 3 (c). It is just an approximation because errors are accumulated in the denoising pass, consistent with conclusions in [24, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.381 + ], + "angle": 0, + "content": "To improve the inversion quality, we propose to optimize a shared unconditional embedding for all frames to align the denoising latent features with the diffusion latent features. Our experiments show that shared embedding is efficient for video inversion and crucial for video editing. Comparisons are shown in Fig. 3 and Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.73 + ], + "angle": 0, + "content": "As discussed in [13], successful attention control requires a model to have both reconstruction ability and editability. While image inversion has been argued to possess both abilities in [24], we find that video editing presents different challenges. The T2S model, as an inflation model not trained on any videos, is not robust to the perturbations caused by various unconditional embeddings. Although our optimized embedding can achieve reconstruction, changing prompts can destabilize the model and result in a low-quality generation. On the other hand, we find that the approximate inversion with an initialized unconditional embedding (the text embedding of an empty prompt) is editable but cannot reconstruct well. To address this issue, we propose a decoupled-guidance strategy in attention control, utilizing different guidance strategies for the source and target prompts. Specifically, we use the optimized unconditional embedding for the source prompt and the initialized unconditional embedding for the target prompt. We incorporate the attention maps from these two branches to generate the target video. These two simple designs prove effective and successfully complete video editing. Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.811 + ], + "angle": 0, + "content": "- We propose the first framework for video editing with attention control. A decoupled-guidance strategy is designed to further improve performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.818, + 0.892, + 0.864 + ], + "angle": 0, + "content": "- We introduce an effective and efficient video inversion method with shared unconditional embedding optimization to improve video editing substantially." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We conduct extensive ablation studies and comparisons to show the effectiveness of our video editing framework." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8600" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.09, + 0.203, + 0.102 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.104, + 0.278, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.09, + 0.411, + 0.102 + ], + "angle": 0, + "content": "(b) DDIM" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.104, + 0.481, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.09, + 0.627, + 0.102 + ], + "angle": 0, + "content": "(c) Tune+DDIM" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.104, + 0.683, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.09, + 0.866, + 0.102 + ], + "angle": 0, + "content": "(d) Ours (Tune+Shared Null)" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.104, + 0.885, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.194, + 0.89, + 0.222 + ], + "angle": 0, + "content": "Figure 3. Inversion Comparison. (b) The inflated model cannot generate high-quality results. (c) Tuning the model can create an approximate video inversion. (d) Optimizing a shared unconditional embedding can accurately reconstruct the input video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.235, + 0.219, + 0.25 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.261, + 0.295, + 0.275 + ], + "angle": 0, + "content": "2.1. Text Driven Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.47, + 0.451 + ], + "angle": 0, + "content": "DALL-E [32] first considers the text-to-image (T2I) generation task as a sequence-to-sequence translation problem, with subsequent research improving generation quality [6, 10, 50]. Denoising Diffusion Probabilistic Models (DDPMs)[15] have recently gained popularity for T2I. GLIDE[26] utilizes classifier-free guidance to improve text conditioning. DALLE-2 [33] leverages CLIP [31] for better text-image alignment. Latent Diffusion Models (LDMs) [34] propose processing in the latent space to enhance training efficiency. In our work, we employ a pretrained image diffusion model based on LDMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.471, + 0.786 + ], + "angle": 0, + "content": "Text-to-video (T2V) generation is a nascent research area. GODIVA [46] first introduces VQ-VAE [41] to T2V. CogVideo [18] combines T2V with CogView-2 [6], utilizing pre-trained text-to-image models. Video Diffusion Models (VDM) [17] propose a space-time U-Net for performing diffusion on pixels. Imagen Video [16] successfully generates high-quality videos with cascaded diffusion models and v-prediction parameterization. Phenaki [42] generates videos with time-variable prompts. Make-A-Video [37] combines the appearance generation of T2I models with movement information from video data. While these approaches generate reasonable short videos, they still contain artifacts and do not support real-world video editing. Additionally, most of these approaches are not publicly available at this time. Tune-A-Video [47] inflates an image diffusion model into a video model and tunes it to reconstruct the input video. It allows for changes in semantic content but with limited temporal consistency. We find that using DDIM inversion results can improve its temporal consistency. However, it cannot avoid altering unrelated regions. We adapt some designs of TAV to do our model initialization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.266, + 0.816 + ], + "angle": 0, + "content": "2.2. Text Driven Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Generative models have demonstrated impressive performance in image editing, with approaches ranging from GANs [12, 27, 28, 44] to diffusion models [1, 19]. SDEdit [22] adds noise to an input image and uses the diffusion process to recover an edited version. Prompt-to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Prompt [13] uses attention control to minimize changes to unrelated parts, while Null-Text Inversion [24] improves real image editing. Plug-and-Play [40] uses feature and self-attention injection, which are unable to avoid editing the whole image when changing a local part. InstructPix2Pix [3] enables flexible text-driven editing with user-provided instructions. Textual Inversion [11], DreamBooth [35], and Custom-Diffusion [20] learn special tokens for personalized concepts and generate related images. Some improved inversion methods [7, 23] are proposed after our work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Video editing with generative models has seen several advances recently. Text2Live [2] employs CLIP to edit textures in videos but struggles with significant semantic changes. Dreamix [25] uses a pre-trainedImagen Video [16] backbone to perform image-to-video and video-to-video editing, with the ability to change motion as well. Gen-1 [8] trains models jointly on images and videos for tasks such as stylization and customization. While these methods enable modifying video content, they operate like guided generation and tend to modify all regions together when editing an object. Our proposed method allows for local editing with a diffusion model pre-trained on images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Besides, there have been some concurrent works and followers after our work, such as [5, 9, 21, 30, 45, 49, 52]. Video-P2P focuses more on introducing the image diffusion models to detailed video editing, instead of conditional video generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.59, + 0.703 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Let \\(\\mathcal{V}\\) be a real video containing \\(n\\) frames. We adopt the Prompt-to-Prompt setting by introducing a source prompt \\(\\mathcal{P}\\) and an edited prompt \\(\\mathcal{P}^*\\) which together generate an edited video \\(\\mathcal{V}^*\\) containing \\(n\\) frames. The prompts are provided by the user. Similar to TAV [47], we assume that the object of interest is present in the first frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To achieve cross-attention control in video editing, we propose Video-P2P, a framework with two key technical designs: (1) optimizing a shared unconditional embedding for video inversion, and (2) using different guidance for the source and edited prompts, and incorporating their attention maps. The framework is illustrated in Fig. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8601" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.208, + 0.108 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.118, + 0.47, + 0.195 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs). LDMs generate an image latent \\( z_{0} \\) using a random noise vector \\( z_{t} \\) and a textual condition \\( P \\) as inputs. As variants of DDPMs, these models aim to predict artificial noise by minimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.203, + 0.47, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} E _ {z _ {0}, \\varepsilon \\sim N (0, I), t \\sim \\text {U n i f o r m} (1, T)} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.469, + 0.299 + ], + "angle": 0, + "content": "where \\(\\mathcal{C} = \\psi(\\mathcal{P})\\) is the embedding of the text prompt, and noise \\(\\varepsilon\\) is added to \\(z_0\\) according to step \\(t\\) to obtain \\(z_t\\). During inference, the model predicts noise \\(\\varepsilon_\\theta(\\cdot)\\) for \\(T\\) steps to generate an image from \\(z_T\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.469, + 0.348 + ], + "angle": 0, + "content": "DDIM sampling and inversion. Deterministic DDIM sampling can be used to generate an image from latent features in a small number of denoising steps:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.355, + 0.469, + 0.402 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\varepsilon_ {\\theta} (z _ {t}, t, \\mathcal {C}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.404, + 0.469, + 0.465 + ], + "angle": 0, + "content": "We use an encoder to encode the real image before the diffusion process and a decoder to decode after the denoising process. DDIM sampling can be reversed in a few steps through the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.473, + 0.469, + 0.519 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\varepsilon_ {\\theta} (z _ {t}, t, \\mathcal {C}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.469, + 0.551 + ], + "angle": 0, + "content": "known as DDIM inversion [38]. This can be used to obtain the corresponding latent features of a real image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.469, + 0.616 + ], + "angle": 0, + "content": "Null-text inversion. To mitigate the amplification effect of text conditioning during image generation, classifier-free guidance is proposed, which performs unconditional prediction [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.627, + 0.468, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = w \\cdot \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) + (1 - w) \\cdot \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.469, + 0.763 + ], + "angle": 0, + "content": "where \\(\\varnothing = \\psi(\"\")\\) is the embedding of a null text and \\(w\\) is the guidance weight. However, the classifier-free guidance increases errors accumulated in the denoising process, leading to imperfect image reconstruction using the DDIM inversion. [24] proposes to align the diffusion latent trajectory \\(z_{T}^{*},\\ldots ,z_{0}^{*}\\) with the denoising latent trajectory \\(z_{T},\\ldots ,z_{0}\\) by optimizing a step-wise unconditional embedding \\(\\varnothing_{t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.773, + 0.469, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\varnothing_ {t}} \\left\\| z _ {t - 1} ^ {*} - z _ {t - 1} \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.235, + 0.832 + ], + "angle": 0, + "content": "3.2. Video Inversion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We begin by constructing a T2S model that is capable of performing an approximate inversion. Following the VDM baselines [16, 17] and TAV [47], we employ \\(1 \\times 3 \\times 3\\) pattern convolution kernels and temporal attention. Moreover, we" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.093, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.344 + ], + "angle": 0, + "content": "Figure 4. Framework. We optimize one shared unconditional embedding for the reconstruct branch (orange). The initialized unconditional embedding is utilized for the editable branch (green). Their attention maps are incorporated to create the target video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.413 + ], + "angle": 0, + "content": "replace the self-attention with frame-attentions, which take the first frames \\( v_{0} \\) and the current frame \\( v_{i} \\) as inputs and update features for the frame \\( v_{i} \\). The formulation of the frame-attention is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.421, + 0.891, + 0.439 + ], + "angle": 0, + "content": "\\[\nQ = W ^ {Q} \\mathbf {v} _ {i}, K = W ^ {K} \\mathbf {v} _ {0}, V = W ^ {V} \\mathbf {v} _ {0}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.891, + 0.476 + ], + "angle": 0, + "content": "where \\(W\\) are the projection matrices in attention. Then, attention maps are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.483, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\[\nM = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.614 + ], + "angle": 0, + "content": "where \\( d \\) is the latent projection dimension. The model processes a video pair-by-pair and computes \\( n \\) times to obtain the prediction for every frame. We find that the simple design suffices for video inversion since the reversed latent features can capture temporal information. Besides, frame-attention conserves memory and speeds up the process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.825 + ], + "angle": 0, + "content": "While model inflation can aid in preserving semantic consistency across frames, it adversely impacts the generation quality of the T2I model. This is because the self-attention parameters are utilized to compute frame correlations, which have not been pre-trained. Consequently, the T2S model, generated through inflation, is insufficient for the approximate inversion, as demonstrated in Fig. 2. To address this, we fine-tune the query projection matrices \\( W^{Q} \\) of the frame- and cross-attentions, as well as additional temporal attention, to perform noise prediction based on the input video following [47]. After this initialization, the T2S model is capable of generating semantically consistent image sets while maintaining the quality of each frame, resulting in successful approximate inversion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Using the fine-tuned T2S model, we perform video inversion by optimizing a shared unconditional embedding. During inversion, each latent feature \\( z_{t} \\) contains a channel for the frames with dimension \\( n \\), where \\( z_{t,i} \\) denotes the latent feature for the \\( i \\)-th frame. We employ DDIM inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "to generate latent features \\(z_0^*, \\ldots, z_T^*\\). The unconditional embedding is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.129, + 0.47, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\varnothing_ {t}} \\sum_ {i = 1} ^ {n} \\left| \\left| z _ {t - 1, i} ^ {*} - z _ {t - 1, i} \\left(\\bar {z} _ {t, i}, \\bar {z} _ {t, 0}, \\varnothing_ {t}, \\mathcal {C}\\right) \\right| \\right| _ {2} ^ {2}, \\text {w h e r e} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.18, + 0.469, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {t - 1, i} = z _ {t - 1, i} \\left(\\bar {z} _ {t, i}, \\bar {z} _ {t, 0}, \\emptyset_ {t}, \\mathcal {C}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.2, + 0.469, + 0.29 + ], + "angle": 0, + "content": "is updated at each step. The T2S model's frame-attentions use two latent features to calculate the corresponding feature for the next step. Notice \\(\\varnothing_{t}\\) is shared by all frames \\((i = 1,\\dots ,n)\\) which minimizes the memory usage. Besides, using the same unconditional embedding for all frames avoids destabilizing the semantic consistency in attention control." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.298, + 0.411, + 0.314 + ], + "angle": 0, + "content": "3.3. Decoupled-guidance Attention Control" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.622 + ], + "angle": 0, + "content": "To perform attention control on real images, existing works [13, 24] require an inference pipeline with both reconstruction ability and editability. However, achieving such a pipeline for a T2S model is challenging. Video inversion allows us to establish an inference pipeline to reconstruct the original video well. However, the T2S model is not as robust as T2I models due to a lack of pre-training with videos. As a result, its editability is compromised with the optimized unconditional embedding, leading to degraded generation quality when changing prompts. In contrast, we find that using an initialized unconditional embedding \\(\\varnothing\\) makes the model more editable while it cannot reconstruct perfectly. This inspires us to combine the abilities of two inference pipelines. For the source prompt, we use the optimized unconditional embedding in the classifier-free guidance. For the target prompt, we choose the initialized unconditional embedding. We then incorporate attention maps from these two branches to obtain the edited video, where the unchanged parts are influenced by the source branch and the edited parts are influenced by the target branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.469, + 0.683 + ], + "angle": 0, + "content": "The pseudo algorithm is shown in Alg. 1. We adopt the attention control methods from Image-P2P to Video-P2P. For example, to perform word swap, the Edit function can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.69, + 0.469, + 0.731 + ], + "angle": 0, + "content": "\\[\nE d i t \\left(M _ {t}, M _ {t} ^ {*}, t\\right) := \\left\\{ \\begin{array}{l l} M _ {t} ^ {*} & \\text {i f} t < \\tau \\\\ M _ {t} & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.469, + 0.856 + ], + "angle": 0, + "content": "\\(M_{t}\\) and \\(M_{t}^{*}\\) are the cross-attention maps for every frame at every step, and \\(DM\\) is the tuned T2S model. Changing the frame-attention maps has a small influence on the final results. Attention maps are swapped only for the first \\(\\tau\\) steps because attentions are formed in the early period. \\(\\overline{M}_{t,w}\\) is the average attention map of the word \\(w\\) calculated at step \\(t\\). It is averaged over steps \\(T,\\ldots ,t\\) independently for every frame. For the \\(j\\)-th frame, we calculate:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.864, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\bar {M} _ {t, w, j} = \\frac {1}{T - t} \\sum_ {i = t} ^ {T} \\bar {M} _ {i, w, j} \\quad j = 1, \\dots , n. \\tag {11}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.819, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Prompt-to-Prompt video editing" + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.114, + 0.836, + 0.37 + ], + "angle": 0, + "content": "1 Input: A source prompt \\(\\mathcal{P}\\) a target prompt \\(\\mathcal{P}^*\\) source video \\(V_{src}\\) \n2 Output: Edited video \\(V_{dst}\\) \n3 Latent features from DDIM inversion: \\(z_{T}\\) \n4 \\(z_{T}^{*}\\gets z_{T}\\) \n5 Initialized unconditional embedding \\(\\varnothing^{*}\\) and optimized unconditional embedding \\(\\varnothing\\) . \n6 for \\(t = T,T - 1,\\ldots ,1\\) do \n7 \\(z_{t - 1},M_t\\gets DM(z_t,\\mathcal{P},t,\\emptyset)\\) \n8 \\(M_t^* \\leftarrow DM(z_t^*,\\mathcal{P}^*,t,\\emptyset^*)\\) \n9 \\(\\widehat{M}_t\\gets \\operatorname {Edit}(M_t,M_t^*,t)\\) \n10 \\(z_{t - 1}^{*}\\gets DM(z_{t}^{*},\\mathcal{P}^{*},t,\\emptyset^{*})\\{M_{t}^{*}\\gets \\widehat{M}_{t}\\}\\) \n11 \\(\\alpha \\gets B(\\overline{M}_{t,w})\\cup B(\\overline{M}_{t,w}^{*})\\) \n12 \\(z_{t - 1}^{*}\\gets (1 - \\alpha)\\odot z_{t - 1} + \\alpha \\odot z_{t - 1}^{*}\\) \n13 end \n14 Return \\((z_0,z_0^*)\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.385, + 0.891, + 0.417 + ], + "angle": 0, + "content": "\\(B\\left(\\overline{M}_{t,w}\\right)\\) represents the binary mask obtained from the attention map. A value is set to 1 when larger than a threshold." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.632, + 0.446 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.454, + 0.717, + 0.469 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We develop our method based on CompVis Stable Diffusion (v1-5). Similar to TAV [47], we fix the image autoencoder and sample 8 or 24 frames at the resolution of \\(512 \\times 512\\) from a video. To initialize the model, we fine-tune the T2S model for 500 steps to reconstruct the original video. During attention control, we set the cross-attention replacing ratio to 0.4 and the attention threshold to 0.3. For prompt refinement, we set the refinement ratio to 0.4. Parameters can be adjusted to control the editing fidelity for different examples. All 8-frame experiments are conducted on a single V100 GPU, with 5 minutes for initialization (tuning), 6 minutes for inversion, and 1 minute for inference. It is difficult to avoid the inversion cost when preserving the background faithfully according to [24]. To speed up to less than one minute, one can replace the null-text inversion with recent inversion methods [7, 23]. In the appendix, we show that a shared unconditional embedding works on 24 frames. Our inversion process extends to over 100 frames. However, editing longer videos (1min) remains challenging. VideoP2P and most existing works concentrate on shorter videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.634, + 0.803 + ], + "angle": 0, + "content": "4.2. Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video-P2P can be utilized for a range of editing applications, including prompt refinement, attention re-weighting, and word swapping. Video-P2P is able to maintain semantic consistency across different frames and preserve the temporal coherence of the original video during the editing process. More examples can be found in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.085, + 0.09, + 0.341, + 0.103 + ], + "angle": 0, + "content": "[Input Video] A car is driving on the road." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.106, + 0.182, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.106, + 0.281, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.106, + 0.381, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.106, + 0.483, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.182, + 0.3, + 0.195 + ], + "angle": 0, + "content": "A car is driving on the flooded road." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.197, + 0.181, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.197, + 0.281, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.197, + 0.381, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.197, + 0.483, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.279, + 0.407, + 0.291 + ], + "angle": 0, + "content": "[Input Video] A man drives a motorbike in the forest." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.293, + 0.181, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.293, + 0.281, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.293, + 0.381, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.293, + 0.483, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.37, + 0.373, + 0.382 + ], + "angle": 0, + "content": "A man Bat Man drives a motorbike in the forest." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.383, + 0.181, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.383, + 0.281, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.383, + 0.381, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.383, + 0.483, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.468, + 0.354, + 0.481 + ], + "angle": 0, + "content": "[Input Video] A dog is running on the grass." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.482, + 0.181, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.482, + 0.281, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.482, + 0.381, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.482, + 0.483, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.559, + 0.314, + 0.573 + ], + "angle": 0, + "content": "A robotic dog is running on the grass." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.574, + 0.181, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.575, + 0.281, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.575, + 0.381, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.575, + 0.483, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.091, + 0.703, + 0.103 + ], + "angle": 0, + "content": "A car is driving on the road railway." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.106, + 0.586, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.106, + 0.686, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.106, + 0.787, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.106, + 0.887, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.183, + 0.716, + 0.195 + ], + "angle": 0, + "content": "A car is driving on the road at sunset." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.197, + 0.586, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.197, + 0.686, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.197, + 0.787, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.197, + 0.887, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.277, + 0.756, + 0.29 + ], + "angle": 0, + "content": "A man drives a Lego motorbike in the forest." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.293, + 0.586, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.293, + 0.686, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.293, + 0.787, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.293, + 0.887, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.37, + 0.793, + 0.382 + ], + "angle": 0, + "content": "A man Spider Man drives a motorbike in the forest." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.383, + 0.586, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.383, + 0.686, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.383, + 0.787, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.383, + 0.887, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.467, + 0.695, + 0.48 + ], + "angle": 0, + "content": "A dog cat is running on the grass." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.482, + 0.586, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.482, + 0.686, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.482, + 0.787, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.482, + 0.887, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.559, + 0.715, + 0.573 + ], + "angle": 0, + "content": "A fluffy \\(\\uparrow\\) dog is running on the grass." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.575, + 0.586, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.575, + 0.686, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.575, + 0.787, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.575, + 0.887, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.659, + 0.831, + 0.674 + ], + "angle": 0, + "content": "Figure 5. Videos edited by Video-P2P with text prompts. Video-P2P can do both word swaps and prompt refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Word swap. Video-P2P enables the replacement of entities based on word swapping while maintaining the coherence of unrelated regions. As illustrated in Fig. 5, Video-P2P seamlessly replaces the man on the motorbike with Spider-Man while minimizing the changes to the motorbike's appearance (the 4th row). The generated Spider-Man exhibits a consistent appearance across frames, and the background remains unchanged. Furthermore, we can replace a dog with a cat while preserving its gesture and the surrounding grass (the 5th row)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Prompt refinement. Video-P2P is able to do prompt refinement, such as modifying object properties. For example, we can transform the running dog into a robotic one (the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.793 + ], + "angle": 0, + "content": "6th row in Fig. 5), and convert a motorbike into a Lego toy with the same motion (the 3rd row). Notice the grass and sky are almost not influenced. Additionally, Video-P2P can perform global editing like changing the weather to sunset or flooding the road with water (2nd row). Style transfer can also be accomplished by Video-P2P, as exemplified by transforming the video into a watercolor painting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Attention re-weighting. Similar to Image-P2P, Video-P2P also enables attention re-weighting. By adjusting the cross-attention of specific words, we can manipulate the extent of the corresponding generation. For instance, we can regulate how fluffy a dog is in the video (the 6th row of Fig. 5)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8604" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.09, + 0.352, + 0.102 + ], + "angle": 0, + "content": "[Input Video] A lion is roaring." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.103, + 0.174, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.103, + 0.273, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.103, + 0.374, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.103, + 0.474, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.174, + 0.365, + 0.186 + ], + "angle": 0, + "content": "[TAV+DDIM] King-Kong is roaring." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.187, + 0.174, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.187, + 0.275, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.187, + 0.374, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.187, + 0.474, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.259, + 0.36, + 0.271 + ], + "angle": 0, + "content": "[Video-P2P] King-Kong is roaring." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.273, + 0.174, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.273, + 0.274, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.273, + 0.374, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.273, + 0.474, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.09, + 0.808, + 0.102 + ], + "angle": 0, + "content": "[Input Video] A tiger is walking in the forest." + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.103, + 0.588, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.104, + 0.688, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.104, + 0.786, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.104, + 0.887, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.174, + 0.823, + 0.186 + ], + "angle": 0, + "content": "[TAV+DDIM] A Lego tiger is walking in the forest." + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.187, + 0.588, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.187, + 0.688, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.187, + 0.787, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.187, + 0.887, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.259, + 0.818, + 0.271 + ], + "angle": 0, + "content": "[Video-P2P] A Lego tiger is walking in the forest." + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.273, + 0.588, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.273, + 0.688, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.273, + 0.787, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.273, + 0.887, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Figure 6. Video-P2P v.s. Tune-A-Video (TAV). Video-P2P offers the ability to edit content locally, while TAV+DDIM cannot avoid influencing unrelated regions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.392, + 0.465, + 0.406 + ], + "angle": 0, + "content": "[Input Video] a brown dog and a white dog on the kitchen floor." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.408, + 0.177, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.183, + 0.409, + 0.277, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.409, + 0.377, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.409, + 0.463, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.46, + 0.456, + 0.472 + ], + "angle": 0, + "content": "[Dreamix] a brown cat and a white cat on the kitchen floor." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.473, + 0.177, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.473, + 0.277, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.473, + 0.377, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.473, + 0.463, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.524, + 0.466, + 0.536 + ], + "angle": 0, + "content": "[Video-P2P] a brown cat and a white cat on the kitchen floor." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.538, + 0.177, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.538, + 0.275, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.538, + 0.377, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.538, + 0.463, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.594, + 0.469, + 0.623 + ], + "angle": 0, + "content": "Figure 7. Video-P2P v.s. Dreamix. Both methods can change the dogs to cats. Video-P2P can preserve background's details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.626, + 0.21, + 0.641 + ], + "angle": 0, + "content": "4.3. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.469, + 0.791 + ], + "angle": 0, + "content": "Comparison with Tune-A-Video. Both TAV+DDIM [47] and our Video-P2P allow for video editing with text prompts. However, TAV+DDIM cannot avoid altering the entire video content when editing specific objects, while Video-P2P can edit a local area and minimize the influence on other regions. Fig. 6 (Left) demonstrates that Video-P2P preserves the complex shape of the cloud when replacing a lion with King Kong, whereas TAV+DDIM can only maintain the color tone of the sky in this case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Although our model initialization is similar to TAV, Video-P2P can still generate temporal-consistent results where TAV+DDIM fails. As demonstrated in Fig. 6 (Right), TAV struggles to generate a temporally consistent sequence in the second row, even when the inputs are features from DDIM inversion. In contrast, our method can produce better structure-preserved results, as shown in the third row." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Comparison with Dreamix. In contrast to Dreamix [25], which uses a pre-trained video diffusion model that is not publicly available, our method yields superior results for subject replacement. Although our method cannot perform video motion editing due to the lack of temporal priors, we outperform Dreamix in preserving details and motion consistency. As it is not open-sourced, we conducted our evaluation on its released demo. As demonstrated in Fig. 7, both methods can transform two dogs into two cats, but our method preserves the details of the drawer in the background (the 3rd row). Furthermore, Dreamix may affect the time sequence to some extent, as the generated cat moves more slowly than the original dog in the video. Our method completely preserves the motion of the original video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. We create a test set of 45 samples based on the DAVIS [29] dataset and report five metrics for quantitative analysis. The CLIP Score measures the textual similarity between the text prompt and video, while Masked PSNR and LPIPS [51] evaluate the quality of structure preservation. We also proposed a novel metric, Object Semantic Variance (OSV), to measure semantic consistency across frames. In addition, we report an average CLIP similarity (Temp) between consecutive frames (as in Gen-1 [8] and FateZero [30]) to evaluate the temporal coherence. For detailed explanations of these metrics, please refer to the appendix. Our results, as shown in Table 1, demonstrate that Video-P2P performs well on all metrics. Compared to TAV+DDIM, Video-P2P achieves higher Masked PSNR and lower LPIPS, indicating better preservation of unchanged regions. Compared to the other two methods, Video-P2P has much lower OSV and Temp, indicating its superior ability to maintain semantic and temporal consistency across frames. When comparing to Image-P2P, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8605" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.089, + 0.366, + 0.101 + ], + "angle": 0, + "content": "A dog-cat is running on the grass." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.102, + 0.177, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.102, + 0.273, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.102, + 0.37, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.102, + 0.467, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.16, + 0.177, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.217, + 0.144, + 0.226 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.16, + 0.273, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.217, + 0.252, + 0.226 + ], + "angle": 0, + "content": "Image-P2P" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.16, + 0.37, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.217, + 0.355, + 0.225 + ], + "angle": 0, + "content": "w/o model init" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.16, + 0.467, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.217, + 0.452, + 0.225 + ], + "angle": 0, + "content": "w/ model init" + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.234, + 0.458, + 0.248 + ], + "angle": 0, + "content": "Figure 8. Model initialization improves the generation quality." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.256, + 0.468, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.306, + 0.441, + 0.32 + ], + "angle": 0, + "content": "Figure 9. Ablation on shared unconditional embeddings." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.329, + 0.468, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.455, + 0.404 + ], + "angle": 0, + "content": "Figure 10. Ablation on decoupled-guidance attention control." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.47, + 0.528 + ], + "angle": 0, + "content": "consider it as a reference point rather than a direct competitor because it does not account for the consistency between different frames. Image-P2P represents the upper bound of PSNR and LPIPS, which serve as metrics to evaluate per-frame quality. Among all video editing methods, Video-P2P is the best one. Moreover, in Tab. 3, we report the user study results, where Video-P2P ranks first on average and has a high preference rate compared to other methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.23, + 0.557 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.699 + ], + "angle": 0, + "content": "Model initialization. While the inflated image diffusion model can generate semantically consistent images, the T2S model's generation ability is compromised during inflation. As seen in Fig.8 (the 3rd column), directly using the inflated T2S model produces unrealistic results with an inaccurate background. To mitigate this, we initialize the T2S model by fine-tuning the given video. This is evident in Fig.8 (4th column), where the cat's appearance improves, and the grass reconstruction becomes more accurate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Shared unconditional embedding. The use of shared unconditional embeddings is crucial to ensure stable and consistent generations. On our test set, we found that \\(10\\%\\) of the videos exhibit instability when multiple unconditional embeddings are employed. A visual example is included in Fig. 9 (Right). As shown in Tab. 1 (row 5), semantic and temporal consistency are diminished without shared unconditional embedding, reminiscent of the content code in [39]. We observe that optimizing a shared unconditional embedding can significantly improve the PSNR compared to TAV+DDIM in Tab. 2. However, using multiple unconditional embeddings for each frame only increases the PSNR by 0.2 but results in higher parameter usage (\\(n\\) times)" + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.089, + 0.885, + 0.171 + ], + "angle": 0, + "content": "
CLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Temp ↑
TAV+DDIM0.334317.370.465155.670.9697
Image-P2P0.327222.920.308276.920.9021
Ours (w/o MI)0.321320.620.325550.270.9460
Ours (w/o DG)0.322618.970.386668.750.9451
Ours (w/o SU)0.335120.640.322148.780.9665
Ours0.336720.650.321347.380.9725
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation. We evaluate textual similarity (CLIP), region preservation (Masked PSNR, LPIPS), semantic consistency (OSV), and temporal consistency (Temp). DG and SU mean Decoupled-Guidance and shared unconditional embeddings." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.235, + 0.895, + 0.29 + ], + "angle": 0, + "content": "
VQVAETAV +DDIMMulti- uncondShared- uncond
PSNR(dB) ↑24.7315.4322.9722.75
#Param. ↓/0.13M22.68M2.94M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.296, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Table 2. Video reconstruction quality. A shared unconditional embedding can reconstruct a high-quality video with a small size." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.325, + 0.891, + 0.422 + ], + "angle": 0, + "content": "
Image-P2PTAVTAV+DDIMVideo-P2P
Structure2.673.332.611.39
Preserving13.59%6.52%10.87%69.02%
Text3.402.782.281.54
Alignment3.80%14.13%19.57%62.50%
Realism & Quality3.382.982.211.43
4.35%7.61%19.02%69.02%
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.428, + 0.892, + 0.491 + ], + "angle": 0, + "content": "Table 3. User study of average ranking \\(\\downarrow\\) and preference rate \\(\\uparrow\\). and leads to a lower Masked PSNR after attention control. Thus, we conclude that shared unconditional embedding is the most effective and efficient method for video inversion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Decoupled-guidance attention control. To obtain the latent features of the input video, we optimize an unconditional embedding using the source prompt. It is important to note that this embedding is only suitable for the source prompt during the prompt-to-prompt process. Using the optimized embedding for the target prompt may negatively impact the quality of the generated results, as shown in Fig. 10 (Left). Instead, we utilize the initialized unconditional embedding for the target prompt and incorporate attention maps from two branches. The decoupled-guidance attention control approach significantly improves the editing quality, as shown in Fig.10 (Right). Quantitative ablations can be found in Tab. 1 (the 4th row and 6th row)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.618, + 0.72 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Video-P2P provides the first framework for video editing with cross-attention control. We optimize a shared unconditional embedding based on a well-initialized T2S model for video inversion. We also propose the decoupled-guidance strategy for attention control. These techniques enable Video-P2P to perform various applications, such as word swap, prompt refinement, and attention re-weighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement Supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8606" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.212 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723, 2022. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.325 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[5] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.491 + ], + "angle": 0, + "content": "[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. arXiv preprint arXiv:2305.04441, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[9] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), pages 1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[21] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[22] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[23] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF Inter" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8607" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "national Conference on Computer Vision, pages 2085-2094, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[29] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[30] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[39] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[44] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[45] Wen Wang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[47] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[48] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[50] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[52] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8608" + } + ] +] \ No newline at end of file diff --git a/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_origin.pdf b/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a2487d583bf44d75f2262f44947b9e6e107abee --- /dev/null +++ b/2024/Video-P2P_ Video Editing with Cross-attention Control/81ad8533-a31f-4ffc-839e-f07d4869e500_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7f1138f4d0e6362d6074553cd4917135cdb8c59443ac04be1157cce2b37536e +size 9623230 diff --git a/2024/Video-P2P_ Video Editing with Cross-attention Control/full.md b/2024/Video-P2P_ Video Editing with Cross-attention Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c30d39c5461eb81d2f9c7468681a2748dea29bf --- /dev/null +++ b/2024/Video-P2P_ Video Editing with Cross-attention Control/full.md @@ -0,0 +1,569 @@ +# Video-P2P: Video Editing with Cross-attention Control + +Shaoteng Liu + +Yuechen Zhang + +Wenbo Li + +Zhe Lin3 + +Jiaya Jia1,2 + +1The Chinese University of Hong Kong 2SmartMore 3 Adobe + +![](images/83784f0f7af0642761ccb26b1d60a832c3369678821e7de39fa7f11de19df8c9.jpg) +[Input Video] A child is riding a bike on the road. +Figure 1. Video-P2P generates new characters while optimally maintaining the pose and environment in videos. + +# Abstract + +Video-P2P is the first framework for real-world video editing with cross-attention control. While attention control has proven effective for image editing with pre-trained image generation models, there are currently no large-scale video generation models publicly available. Video-P2P addresses this limitation by adapting an image generation diffusion model to complete various video editing tasks. Specifically, we propose to first tune a Text-to-Set (T2S) model to complete an approximate inversion and then optimize a shared unconditional embedding to achieve accurate video inversion with a small memory cost. We further prove that it is crucial for consistent video editing. For attention control, we introduce a novel decoupled-guidance strategy, which uses different guidance strategies for the source and target prompts. The optimized unconditional embedding for + +the source prompt improves reconstruction ability, while an initialized unconditional embedding for the target prompt enhances editability. Incorporating the attention maps of these two branches enables detailed editing. These technical designs enable various text-driven editing applications, including word swap, prompt refinement, and attention reweighting. Video-P2P works well on real-world videos for generating new characters while optimally preserving their original poses and scenes. It significantly outperforms previous approaches. + +# 1. Introduction + +Video creation and editing are key tasks [16, 17, 25, 37, 42]. Text-driven editing becomes one promising pipeline. Several methods have demonstrated the ability to edit generated or real-world images with text prompts [13, 19, 24]. + +[Input Video] A penguin is running on the ice. + +![](images/b37b6980737bf96d85a327052f5d39fb3d04acb9d28bde6473d41972e8e0de2f.jpg) + +![](images/b93c05dfa17a99af84a4f64c2ac9ef4918987ae77712d476dd8eb1ec1564ec3c.jpg) + +![](images/3f6a423f5441f3998f6074973130d25cfc81007b59aee8133d02a582ec56348e.jpg) + +![](images/429ccc9f43f35c525c434644027b2825fdc78dcba46dead1a9035a6a949d9bc2.jpg) + +[Image-P2P] A Robotic penguin is running on the ice. + +![](images/4c0d6d3ff201311d99c4d451a6b02eacc96865fb29825406bc21f7ab2e688f91.jpg) + +![](images/e3ab97766f73b0a33fd19a59604f56644a2fd68d53c8a6fd7bf124eddb4e070b.jpg) + +![](images/cad2ca82be901f2723d6cd79c2c6c53b1abc9b713a2f14c00a60540881c58182.jpg) + +![](images/c906a9290bd92858a094f4bcbf2169d410f7aa2a750feee31c843c11180a00dd.jpg) + +[Video-P2P] A Robotic penguin is running on the ice. + +![](images/39c6eb391474c84ec39901ff43d5440df94bf674ae61234100001570524dbe6b.jpg) +Figure 2. Video-P2P vs Image-P2P. Editing a video frame-by-frame (Image-P2P) cannot guarantee semantic consistency across frames. Video-P2P enables changing the penguin into the same robotic type in every frame. + +![](images/77922d7b571d70ac66c3223449fbb30850500f481e26cdb1d888959b775461b3.jpg) + +![](images/1ef304f30049add608b912260b996b12b6a3e4a15b76d22462b3c0e197afacf8.jpg) + +![](images/c067812dd4a8a7b41a7396ef114ef469c93d82907696721247e312b3e1d0d998.jpg) + +Till now, it is still challenging to edit only local objects in a video, such as changing a running "dog" into a "cat" without influencing the environment. This paper proposes a pipeline that can edit a video both locally and globally, as shown in Figs. 1 and 5. + +Text-driven image editing requires a model capable of generating target content, such as changing the category or property of an object. Diffusion models have demonstrated outstanding generation capabilities in this area [3, 13, 19, 40]. Among these methods, attention control emerges as the most effective pipeline for detailed image editing [13, 24]. In order to edit a real image, this pipeline includes two necessary steps: (1) inverting the image into latent features with a pre-trained diffusion model, and (2) controlling attention maps in the denoising process to edit the corresponding parts of the image. For example, by swapping their attention maps, we can replace a "child" with a "panda". + +In this paper, we aim to build an attention control-based pipeline for video editing. Since no large-scale pre-trained video generation models are publicly available, we propose a novel framework to show that a pre-trained image diffusion model can be adapted for detailed video editing. + +While a pre-trained image diffusion model can be utilized for video editing by processing frames individually (Image-P2P), it lacks semantic consistency across frames (the 2nd row of Fig. 2). To maintain semantic consistency, we propose using a structure on inversion and attention control for all frames, by transforming the Text-to-image diffusion model (T2I) into a Text-to-set model (T2S). This approach is effective, as illustrated in the 3rd row, where the robotic penguin maintains its consistency across frames. + +We adopt the method proposed in [47] to convert a T2I model into a T2S model by altering the convolution kernels and replacing the self-attention with frame-attentions. This conversion yields a model that generates a set of semantically consistent images. The generation quality will be degraded with the inflation step but it can be recovered after tuning on the original video. Although the tuned T2S model is not an ideal video generation model, it suffices to create an approximate inversion for a video as shown in Fig. 3 (c). It is just an approximation because errors are accumulated in the denoising pass, consistent with conclusions in [24, 43]. + +To improve the inversion quality, we propose to optimize a shared unconditional embedding for all frames to align the denoising latent features with the diffusion latent features. Our experiments show that shared embedding is efficient for video inversion and crucial for video editing. Comparisons are shown in Fig. 3 and Fig. 9. + +As discussed in [13], successful attention control requires a model to have both reconstruction ability and editability. While image inversion has been argued to possess both abilities in [24], we find that video editing presents different challenges. The T2S model, as an inflation model not trained on any videos, is not robust to the perturbations caused by various unconditional embeddings. Although our optimized embedding can achieve reconstruction, changing prompts can destabilize the model and result in a low-quality generation. On the other hand, we find that the approximate inversion with an initialized unconditional embedding (the text embedding of an empty prompt) is editable but cannot reconstruct well. To address this issue, we propose a decoupled-guidance strategy in attention control, utilizing different guidance strategies for the source and target prompts. Specifically, we use the optimized unconditional embedding for the source prompt and the initialized unconditional embedding for the target prompt. We incorporate the attention maps from these two branches to generate the target video. These two simple designs prove effective and successfully complete video editing. Our contributions can be summarized as: + +- We propose the first framework for video editing with attention control. A decoupled-guidance strategy is designed to further improve performance. +- We introduce an effective and efficient video inversion method with shared unconditional embedding optimization to improve video editing substantially. +- We conduct extensive ablation studies and comparisons to show the effectiveness of our video editing framework. + +![](images/fb07cff4b777789dd136a51f13c5a28d4eaae1f761c91acbf382fc162bdcc82d.jpg) +(a) Input + +![](images/e7fdadcd8c74bc72fa00a8af6bf73f17a2403c9e52a9c5102c95a41fdb7028e8.jpg) +(b) DDIM +Figure 3. Inversion Comparison. (b) The inflated model cannot generate high-quality results. (c) Tuning the model can create an approximate video inversion. (d) Optimizing a shared unconditional embedding can accurately reconstruct the input video. + +![](images/9df2f657e8072787d9f9ef0ba23cffd0e0cb61f8178ed334849a358bad59e968.jpg) +(c) Tune+DDIM + +![](images/437e49b274129baf479c0f39ed63117bdfdf3c0b5b8650f976234841d9b1869d.jpg) +(d) Ours (Tune+Shared Null) + +# 2. Related Work + +# 2.1. Text Driven Generation + +DALL-E [32] first considers the text-to-image (T2I) generation task as a sequence-to-sequence translation problem, with subsequent research improving generation quality [6, 10, 50]. Denoising Diffusion Probabilistic Models (DDPMs)[15] have recently gained popularity for T2I. GLIDE[26] utilizes classifier-free guidance to improve text conditioning. DALLE-2 [33] leverages CLIP [31] for better text-image alignment. Latent Diffusion Models (LDMs) [34] propose processing in the latent space to enhance training efficiency. In our work, we employ a pretrained image diffusion model based on LDMs. + +Text-to-video (T2V) generation is a nascent research area. GODIVA [46] first introduces VQ-VAE [41] to T2V. CogVideo [18] combines T2V with CogView-2 [6], utilizing pre-trained text-to-image models. Video Diffusion Models (VDM) [17] propose a space-time U-Net for performing diffusion on pixels. Imagen Video [16] successfully generates high-quality videos with cascaded diffusion models and v-prediction parameterization. Phenaki [42] generates videos with time-variable prompts. Make-A-Video [37] combines the appearance generation of T2I models with movement information from video data. While these approaches generate reasonable short videos, they still contain artifacts and do not support real-world video editing. Additionally, most of these approaches are not publicly available at this time. Tune-A-Video [47] inflates an image diffusion model into a video model and tunes it to reconstruct the input video. It allows for changes in semantic content but with limited temporal consistency. We find that using DDIM inversion results can improve its temporal consistency. However, it cannot avoid altering unrelated regions. We adapt some designs of TAV to do our model initialization. + +# 2.2. Text Driven Editing + +Generative models have demonstrated impressive performance in image editing, with approaches ranging from GANs [12, 27, 28, 44] to diffusion models [1, 19]. SDEdit [22] adds noise to an input image and uses the diffusion process to recover an edited version. Prompt-to + +Prompt [13] uses attention control to minimize changes to unrelated parts, while Null-Text Inversion [24] improves real image editing. Plug-and-Play [40] uses feature and self-attention injection, which are unable to avoid editing the whole image when changing a local part. InstructPix2Pix [3] enables flexible text-driven editing with user-provided instructions. Textual Inversion [11], DreamBooth [35], and Custom-Diffusion [20] learn special tokens for personalized concepts and generate related images. Some improved inversion methods [7, 23] are proposed after our work. + +Video editing with generative models has seen several advances recently. Text2Live [2] employs CLIP to edit textures in videos but struggles with significant semantic changes. Dreamix [25] uses a pre-trainedImagen Video [16] backbone to perform image-to-video and video-to-video editing, with the ability to change motion as well. Gen-1 [8] trains models jointly on images and videos for tasks such as stylization and customization. While these methods enable modifying video content, they operate like guided generation and tend to modify all regions together when editing an object. Our proposed method allows for local editing with a diffusion model pre-trained on images. + +Besides, there have been some concurrent works and followers after our work, such as [5, 9, 21, 30, 45, 49, 52]. Video-P2P focuses more on introducing the image diffusion models to detailed video editing, instead of conditional video generation. + +# 3. Method + +Let $\mathcal{V}$ be a real video containing $n$ frames. We adopt the Prompt-to-Prompt setting by introducing a source prompt $\mathcal{P}$ and an edited prompt $\mathcal{P}^*$ which together generate an edited video $\mathcal{V}^*$ containing $n$ frames. The prompts are provided by the user. Similar to TAV [47], we assume that the object of interest is present in the first frame. + +To achieve cross-attention control in video editing, we propose Video-P2P, a framework with two key technical designs: (1) optimizing a shared unconditional embedding for video inversion, and (2) using different guidance for the source and edited prompts, and incorporating their attention maps. The framework is illustrated in Fig. 4. + +# 3.1. Preliminary + +Latent Diffusion Models (LDMs). LDMs generate an image latent $z_{0}$ using a random noise vector $z_{t}$ and a textual condition $P$ as inputs. As variants of DDPMs, these models aim to predict artificial noise by minimizing the following objective: + +$$ +\min _ {\theta} E _ {z _ {0}, \varepsilon \sim N (0, I), t \sim \text {U n i f o r m} (1, T)} \| \varepsilon - \varepsilon_ {\theta} \left(z _ {t}, t, \mathcal {C}\right) \| _ {2} ^ {2}, \tag {1} +$$ + +where $\mathcal{C} = \psi(\mathcal{P})$ is the embedding of the text prompt, and noise $\varepsilon$ is added to $z_0$ according to step $t$ to obtain $z_t$ . During inference, the model predicts noise $\varepsilon_\theta(\cdot)$ for $T$ steps to generate an image from $z_T$ . + +DDIM sampling and inversion. Deterministic DDIM sampling can be used to generate an image from latent features in a small number of denoising steps: + +$$ +z _ {t - 1} = \sqrt {\frac {\alpha_ {t - 1}}{\alpha_ {t}}} z _ {t} + \left(\sqrt {\frac {1}{\alpha_ {t - 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \cdot \varepsilon_ {\theta} (z _ {t}, t, \mathcal {C}). \tag {2} +$$ + +We use an encoder to encode the real image before the diffusion process and a decoder to decode after the denoising process. DDIM sampling can be reversed in a few steps through the equation: + +$$ +z _ {t + 1} = \sqrt {\frac {\alpha_ {t + 1}}{\alpha_ {t}}} z _ {t} + \left(\sqrt {\frac {1}{\alpha_ {t + 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \cdot \varepsilon_ {\theta} (z _ {t}, t, \mathcal {C}), \tag {3} +$$ + +known as DDIM inversion [38]. This can be used to obtain the corresponding latent features of a real image. + +Null-text inversion. To mitigate the amplification effect of text conditioning during image generation, classifier-free guidance is proposed, which performs unconditional prediction [14]: + +$$ +\tilde {\varepsilon} _ {\theta} \left(z _ {t}, t, \mathcal {C}, \varnothing\right) = w \cdot \varepsilon_ {\theta} \left(z _ {t}, t, \mathcal {C}\right) + (1 - w) \cdot \varepsilon_ {\theta} \left(z _ {t}, t, \varnothing\right), \tag {4} +$$ + +where $\varnothing = \psi("")$ is the embedding of a null text and $w$ is the guidance weight. However, the classifier-free guidance increases errors accumulated in the denoising process, leading to imperfect image reconstruction using the DDIM inversion. [24] proposes to align the diffusion latent trajectory $z_{T}^{*},\ldots ,z_{0}^{*}$ with the denoising latent trajectory $z_{T},\ldots ,z_{0}$ by optimizing a step-wise unconditional embedding $\varnothing_{t}$ : + +$$ +\min _ {\varnothing_ {t}} \left\| z _ {t - 1} ^ {*} - z _ {t - 1} \right\| _ {2} ^ {2}. \tag {5} +$$ + +# 3.2. Video Inversion + +We begin by constructing a T2S model that is capable of performing an approximate inversion. Following the VDM baselines [16, 17] and TAV [47], we employ $1 \times 3 \times 3$ pattern convolution kernels and temporal attention. Moreover, we + +![](images/e9b9c12e560e4ca378f1536dfc72f7ecbdc928da9faf8afdce1b59ee5adc6b31.jpg) +Figure 4. Framework. We optimize one shared unconditional embedding for the reconstruct branch (orange). The initialized unconditional embedding is utilized for the editable branch (green). Their attention maps are incorporated to create the target video. + +replace the self-attention with frame-attentions, which take the first frames $v_{0}$ and the current frame $v_{i}$ as inputs and update features for the frame $v_{i}$ . The formulation of the frame-attention is as follows: + +$$ +Q = W ^ {Q} \mathbf {v} _ {i}, K = W ^ {K} \mathbf {v} _ {0}, V = W ^ {V} \mathbf {v} _ {0}, \tag {6} +$$ + +where $W$ are the projection matrices in attention. Then, attention maps are calculated as: + +$$ +M = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right), \tag {7} +$$ + +where $d$ is the latent projection dimension. The model processes a video pair-by-pair and computes $n$ times to obtain the prediction for every frame. We find that the simple design suffices for video inversion since the reversed latent features can capture temporal information. Besides, frame-attention conserves memory and speeds up the process. + +While model inflation can aid in preserving semantic consistency across frames, it adversely impacts the generation quality of the T2I model. This is because the self-attention parameters are utilized to compute frame correlations, which have not been pre-trained. Consequently, the T2S model, generated through inflation, is insufficient for the approximate inversion, as demonstrated in Fig. 2. To address this, we fine-tune the query projection matrices $W^{Q}$ of the frame- and cross-attentions, as well as additional temporal attention, to perform noise prediction based on the input video following [47]. After this initialization, the T2S model is capable of generating semantically consistent image sets while maintaining the quality of each frame, resulting in successful approximate inversion. + +Using the fine-tuned T2S model, we perform video inversion by optimizing a shared unconditional embedding. During inversion, each latent feature $z_{t}$ contains a channel for the frames with dimension $n$ , where $z_{t,i}$ denotes the latent feature for the $i$ -th frame. We employ DDIM inversion + +to generate latent features $z_0^*, \ldots, z_T^*$ . The unconditional embedding is defined as follows: + +$$ +\min _ {\varnothing_ {t}} \sum_ {i = 1} ^ {n} \left| \left| z _ {t - 1, i} ^ {*} - z _ {t - 1, i} \left(\bar {z} _ {t, i}, \bar {z} _ {t, 0}, \varnothing_ {t}, \mathcal {C}\right) \right| \right| _ {2} ^ {2}, \text {w h e r e} \tag {8} +$$ + +$$ +\bar {z} _ {t - 1, i} = z _ {t - 1, i} \left(\bar {z} _ {t, i}, \bar {z} _ {t, 0}, \emptyset_ {t}, \mathcal {C}\right) \tag {9} +$$ + +is updated at each step. The T2S model's frame-attentions use two latent features to calculate the corresponding feature for the next step. Notice $\varnothing_{t}$ is shared by all frames $(i = 1,\dots ,n)$ which minimizes the memory usage. Besides, using the same unconditional embedding for all frames avoids destabilizing the semantic consistency in attention control. + +# 3.3. Decoupled-guidance Attention Control + +To perform attention control on real images, existing works [13, 24] require an inference pipeline with both reconstruction ability and editability. However, achieving such a pipeline for a T2S model is challenging. Video inversion allows us to establish an inference pipeline to reconstruct the original video well. However, the T2S model is not as robust as T2I models due to a lack of pre-training with videos. As a result, its editability is compromised with the optimized unconditional embedding, leading to degraded generation quality when changing prompts. In contrast, we find that using an initialized unconditional embedding $\varnothing$ makes the model more editable while it cannot reconstruct perfectly. This inspires us to combine the abilities of two inference pipelines. For the source prompt, we use the optimized unconditional embedding in the classifier-free guidance. For the target prompt, we choose the initialized unconditional embedding. We then incorporate attention maps from these two branches to obtain the edited video, where the unchanged parts are influenced by the source branch and the edited parts are influenced by the target branch. + +The pseudo algorithm is shown in Alg. 1. We adopt the attention control methods from Image-P2P to Video-P2P. For example, to perform word swap, the Edit function can be represented as: + +$$ +E d i t \left(M _ {t}, M _ {t} ^ {*}, t\right) := \left\{ \begin{array}{l l} M _ {t} ^ {*} & \text {i f} t < \tau \\ M _ {t} & \text {o t h e r w i s e} \end{array} , \right. \tag {10} +$$ + +$M_{t}$ and $M_{t}^{*}$ are the cross-attention maps for every frame at every step, and $DM$ is the tuned T2S model. Changing the frame-attention maps has a small influence on the final results. Attention maps are swapped only for the first $\tau$ steps because attentions are formed in the early period. $\overline{M}_{t,w}$ is the average attention map of the word $w$ calculated at step $t$ . It is averaged over steps $T,\ldots ,t$ independently for every frame. For the $j$ -th frame, we calculate: + +$$ +\bar {M} _ {t, w, j} = \frac {1}{T - t} \sum_ {i = t} ^ {T} \bar {M} _ {i, w, j} \quad j = 1, \dots , n. \tag {11} +$$ + +Algorithm 1: Prompt-to-Prompt video editing +1 Input: A source prompt $\mathcal{P}$ a target prompt $\mathcal{P}^*$ source video $V_{src}$ +2 Output: Edited video $V_{dst}$ +3 Latent features from DDIM inversion: $z_{T}$ +4 $z_{T}^{*}\gets z_{T}$ +5 Initialized unconditional embedding $\varnothing^{*}$ and optimized unconditional embedding $\varnothing$ . +6 for $t = T,T - 1,\ldots ,1$ do +7 $z_{t - 1},M_t\gets DM(z_t,\mathcal{P},t,\emptyset)$ +8 $M_t^* \leftarrow DM(z_t^*,\mathcal{P}^*,t,\emptyset^*)$ +9 $\widehat{M}_t\gets \operatorname {Edit}(M_t,M_t^*,t)$ +10 $z_{t - 1}^{*}\gets DM(z_{t}^{*},\mathcal{P}^{*},t,\emptyset^{*})\{M_{t}^{*}\gets \widehat{M}_{t}\}$ +11 $\alpha \gets B(\overline{M}_{t,w})\cup B(\overline{M}_{t,w}^{*})$ +12 $z_{t - 1}^{*}\gets (1 - \alpha)\odot z_{t - 1} + \alpha \odot z_{t - 1}^{*}$ +13 end +14 Return $(z_0,z_0^*)$ + +$B\left(\overline{M}_{t,w}\right)$ represents the binary mask obtained from the attention map. A value is set to 1 when larger than a threshold. + +# 4. Experiments + +# 4.1. Implementation Details + +We develop our method based on CompVis Stable Diffusion (v1-5). Similar to TAV [47], we fix the image autoencoder and sample 8 or 24 frames at the resolution of $512 \times 512$ from a video. To initialize the model, we fine-tune the T2S model for 500 steps to reconstruct the original video. During attention control, we set the cross-attention replacing ratio to 0.4 and the attention threshold to 0.3. For prompt refinement, we set the refinement ratio to 0.4. Parameters can be adjusted to control the editing fidelity for different examples. All 8-frame experiments are conducted on a single V100 GPU, with 5 minutes for initialization (tuning), 6 minutes for inversion, and 1 minute for inference. It is difficult to avoid the inversion cost when preserving the background faithfully according to [24]. To speed up to less than one minute, one can replace the null-text inversion with recent inversion methods [7, 23]. In the appendix, we show that a shared unconditional embedding works on 24 frames. Our inversion process extends to over 100 frames. However, editing longer videos (1min) remains challenging. VideoP2P and most existing works concentrate on shorter videos. + +# 4.2. Applications + +Video-P2P can be utilized for a range of editing applications, including prompt refinement, attention re-weighting, and word swapping. Video-P2P is able to maintain semantic consistency across different frames and preserve the temporal coherence of the original video during the editing process. More examples can be found in the appendix. + +[Input Video] A car is driving on the road. + +![](images/d041f0dc691e815d07d43bb27ac3d4388932856f29f2c8767775e4a1a61036ea.jpg) + +![](images/424b100a8c1eff47e2d61b38e5422e7e94eb92ec8699ebc48c79275008374eab.jpg) + +![](images/466a4d96970a114f7fa2ab8fdcf444b9ef667a5e84615b0cf4cc9e707b658f6c.jpg) + +![](images/89c0496cf3125144f4efa6c4d4846fba160be89950a59e061de466be9d803890.jpg) + +A car is driving on the flooded road. + +![](images/526e11e454db70c56885a81ec7c68bd2e02d47ea3dd52e4ff8eef900d2291696.jpg) + +![](images/542b1a64aa5dc9e6892e5717be1c39e6209a98c2fabff79f76f5e01b75e3267d.jpg) + +![](images/4ce6484cc75d9b254c1ebb01583e4b2478efbfc8fea02e5d1687c8a09adec475.jpg) + +![](images/ee2db35a88659841986f543fe31609d400fce813c17a67ae86c92d92a9fbc720.jpg) + +[Input Video] A man drives a motorbike in the forest. + +![](images/450e6c5c0d5b59204572be19b7838591492fcbd5298c024b550dc21fb7eb40b2.jpg) + +![](images/1d018e862a6d53557a848da1367e00f2c867156159411515e8d943f844b1a978.jpg) + +![](images/4eda895ca1663d2ff13acff297a478ce69f2f2d691c5baf2bd8a1e499ea15211.jpg) + +![](images/ee7f7b5fde13fd494fc2c3cd331b5d599d0154482a719072766efda644df1399.jpg) + +A man Bat Man drives a motorbike in the forest. + +![](images/c8a9af957f6359e04aba3601a2dae483d4e3984da2d769ca7047b71b1499347e.jpg) + +![](images/d61b009e0be12b74b2071726c3a0f03a78cce723ebd3207c91337c0785804faa.jpg) + +![](images/62f1de7a0f07a89ea7eeecce0010a7de316a06080b4529be255e158d2a451506.jpg) + +![](images/4ed69f4db463c83b5ede3a0484259d05aee2273625fd7df42751a24a5829d6c2.jpg) + +[Input Video] A dog is running on the grass. + +![](images/31e2ed7092a4f51d00dc41353bef6c3e5cdec06692bded3ca6189c354b8bd335.jpg) + +![](images/d23fa563541f6e0f41956f1bd86883b0c472adb351e26d930dc1da575e93abe1.jpg) + +![](images/c241ba8214ee7849a4de823ba1d386cddc0ec81b77199cbe40faa65b2430e320.jpg) + +![](images/9aeaaa0dcde9fbfc4555d2a07e0c47c76d799fd563b8b3abcf77750d30c87777.jpg) + +A robotic dog is running on the grass. + +![](images/8df7d95a573c3577bb5c625f4f4cb28c7729c4a406a727daff8d92bf0480c982.jpg) +Figure 5. Videos edited by Video-P2P with text prompts. Video-P2P can do both word swaps and prompt refinement. + +![](images/1bd63a0690d6e3a9a1367b06b81b2d653a4d5c6ca82d90f3e2e52e34968042f4.jpg) + +![](images/0938b734dd91c8c58dc07c4a80d575f39745ff41699115c11f014d12fa5be87e.jpg) + +![](images/cbd8d637c0642b7f918bef4749d72518611f4c2d10f4c7a8b7c826532908aefd.jpg) + +A car is driving on the road railway. + +![](images/5c0561a773a7ad13eba901e9981878279c92894b7144e3bbe1499c6138e844c5.jpg) + +![](images/29c081042cc8388248e695d5517cde1129ecf5d5ad4bf9a7b9426596d2c71ae8.jpg) + +![](images/70f87d2b08a5002336c7dd58f76466b8a4d8ca8982a0e3e21c1ff9270355c147.jpg) + +![](images/885f9c8e9ae5858269988a7610bf7c7cbf452af985c929fa0174d95b96b9d9e5.jpg) + +A car is driving on the road at sunset. + +![](images/cc667d6683e2643ad402feac4a1604edacd2a8f4d4a38492ce997d17e653266e.jpg) + +![](images/38779350364a4198756abaee8242f7ba825a07e65f7834aee885ffabc7a50925.jpg) + +![](images/f4b28a78f4b1e69b5d9200496e7d91cf1211c71b00e52539ebc538089dfef04d.jpg) + +![](images/0f6144bfdd5f849dde10a2db9439c60aa083276237339ab558a28c25c13913c6.jpg) + +A man drives a Lego motorbike in the forest. + +![](images/21e2aa76875d25aed32e4f31ba398c6ff8e183c8755d19ead1741589407bf69d.jpg) + +![](images/17ef09c21f078b5e003af01a14afb06e3ffbd54997180c011cbfadf3658ea0fc.jpg) + +![](images/4137e19e1ae31754272c2de9a8010a19f70a53df2fab79a007fffa52ec8763ac.jpg) + +![](images/40e9cb9081a2ba7698eced0a87dff02dc9d0595c59e02e408a621ba68188bffb.jpg) + +A man Spider Man drives a motorbike in the forest. + +![](images/c41fd342c23d913120c85d6070e697852bfe03aa6e027d6c34dd3a9c954ec39f.jpg) + +![](images/6533b27b862229ad1f1f31aacc7cca91fb53e02e352a96fd27393b4b0a92c172.jpg) + +![](images/bb8e620954ea7c5c7dd18584ef72b8b264f93c55b5b536990f67335f7b5ac6a3.jpg) + +![](images/20a50f60ecb91869ef762bcdba895c738cfb57265161360d0aa0325b80b8af99.jpg) + +A dog cat is running on the grass. + +![](images/b233b6f264d011190749631b76dd0bf1102b113211cb9e0c06bf06d9def44609.jpg) + +![](images/cd46e9e85d1e6a625a30ac03e9200a5a17276eb2a22f02945dd2b94cea2e6c4d.jpg) + +![](images/9335c60824c9d469c767134cfab8445bc5a53d69061b3eab1a466937de72e4bf.jpg) + +![](images/a3fc67b288ac36e0c0fd21a3239eb2fd7f583636be608f0db9d081b5c5331cf5.jpg) + +A fluffy $\uparrow$ dog is running on the grass. + +![](images/55e82f37575debf34f5527138a68a5c576ad5b5a62d585e1b9c903f57c99cf42.jpg) + +![](images/50b820aba89d2cb15d3c9ab424c57af90e732f765732413cbb91dfdc4c711d86.jpg) + +![](images/2bfc7b2dc05fdec65b5915f4ee8adfa5096383e284de366eb8c01b9d8f5537f9.jpg) + +![](images/1a87f466a72fc87f915cf31f5784f6ff037e21b097b963c0b35eb5ab57676f52.jpg) + +Word swap. Video-P2P enables the replacement of entities based on word swapping while maintaining the coherence of unrelated regions. As illustrated in Fig. 5, Video-P2P seamlessly replaces the man on the motorbike with Spider-Man while minimizing the changes to the motorbike's appearance (the 4th row). The generated Spider-Man exhibits a consistent appearance across frames, and the background remains unchanged. Furthermore, we can replace a dog with a cat while preserving its gesture and the surrounding grass (the 5th row). + +Prompt refinement. Video-P2P is able to do prompt refinement, such as modifying object properties. For example, we can transform the running dog into a robotic one (the + +6th row in Fig. 5), and convert a motorbike into a Lego toy with the same motion (the 3rd row). Notice the grass and sky are almost not influenced. Additionally, Video-P2P can perform global editing like changing the weather to sunset or flooding the road with water (2nd row). Style transfer can also be accomplished by Video-P2P, as exemplified by transforming the video into a watercolor painting. + +Attention re-weighting. Similar to Image-P2P, Video-P2P also enables attention re-weighting. By adjusting the cross-attention of specific words, we can manipulate the extent of the corresponding generation. For instance, we can regulate how fluffy a dog is in the video (the 6th row of Fig. 5). + +![](images/73b34461955180fb977c1aa7f225be47abc7959ac0e8eb7c578e2cd5d4903ed4.jpg) +[TAV+DDIM] King-Kong is roaring. + +![](images/b19744c9cd86d509b8804e9ed804d8adf8ea2648fb369ee6163072d40b28b36b.jpg) +[Input Video] A lion is roaring. + +![](images/24f0601bbf8180941b1e4ad2ae01433a4ceb308127f0bbc99b57688c4a158df1.jpg) + +![](images/9592dfa5fb777d4bd5f446ae6bcddd175b4d89596b561a56b5c05c00a275d4c6.jpg) + +![](images/17b3ef8917c72755dc3d0066b0e76618c27df64c6cdc2bd3e9cc3ff5981725ab.jpg) +[Video-P2P] King-Kong is roaring. + +![](images/bc64099e94ce3a956a5ceffab8cf4250ac4d4aeea2578648a0b5f2efb5981508.jpg) + +![](images/b9449b45a0a9567af94c930c382ad8bc1e45ed5b80523ce6a66bd9d50a5afb29.jpg) + +![](images/e842132d49e2d25a900fadeba48bb13caee16835dfbdd8ed0e72d8872a095325.jpg) + +![](images/f64dc1cc4265792ca7f709725c7ad3ec6b5e01264dfa19dcc6ecc15674996d4e.jpg) +Figure 6. Video-P2P v.s. Tune-A-Video (TAV). Video-P2P offers the ability to edit content locally, while TAV+DDIM cannot avoid influencing unrelated regions. + +![](images/2f8337cc8ab4d0c97df70aa754f7de5e165fd7951bec03dc0ce3afc286ff01f0.jpg) + +![](images/79760badc0cf109ed6153b54a52c7a38690752895dfededcfb9fbced70308b61.jpg) + +![](images/382ba663ac22251733534e7307ef727e9862e79811d40e33e42d802451bbb2a5.jpg) + +![](images/f4522730b2ed31cc1cca85a7bcd0d6af825fd6b5a7136c7469050d4b93cb205f.jpg) +[TAV+DDIM] A Lego tiger is walking in the forest. + +![](images/d008b7946f8122401c3fe16308379bc92517a56241128c7c87e14fdf427cda99.jpg) +[Input Video] A tiger is walking in the forest. + +![](images/50a6e8885e4ca64df6bfdf3c7e60ee0a7f68cb9d28a4922b46581a10844cd08a.jpg) + +![](images/1b439aa2bc86b7ce4ad3f42daefa8b7916780eb85392e8e8904fe4840cd052c6.jpg) + +![](images/7782192d9c05e64cda6725ce3d84c515aefc8eb806e14805cfdf274e0fec6943.jpg) + +![](images/190248bd0954fb26d448252f9922ad667205186b116c46487bcbbadf82f69eec.jpg) + +![](images/076c1115f5f76469ef84b5a6a83aa6514462306d8f8115ed7bf073e2de036b0b.jpg) + +![](images/63ba69ad9298ab3044dbf95a65c11833f9b0dc8ca8bb8dba05487e7d65f87d9b.jpg) + +![](images/51975775d897c93dd1b7e7c5812018a75922dc77ee0b5c7f2c4e1cc3592f9f7e.jpg) +[Video-P2P] A Lego tiger is walking in the forest. + +![](images/fcc7a614e32dc413920c3ea1bc131a2ce1087be8ad7a8bcd4124822786c4d7d9.jpg) + +![](images/82495e4a28e42d7b325a974dbc6033d8b6b18e5b11952aa99898b2d3cdf013a6.jpg) + +![](images/c028d4891a5a3bffbd29516bc107010da27052ce5ad63c5fa3703e88f6e8e0e4.jpg) + +[Input Video] a brown dog and a white dog on the kitchen floor. + +![](images/213e52013709e7a91620e907ec629fc1de890b81d3f2450b163e981e977d7018.jpg) + +![](images/97c6802bd66efb6f675affed8023f0c8abea1a33c075ae1269a7d3756edf57b6.jpg) + +![](images/3c6700b3759eb09a9b0eab10d35c19b3558526d4054c79f0058559c7262bfb0b.jpg) + +![](images/98339615b3b21c121f796ce67b78058e5e22772c89e3fa23d0ea9e9ca6aeac02.jpg) + +![](images/9c4b3a02753a59567a9e5497b22d90aa46b44aff47bafbf2013245a8555f6990.jpg) +[Dreamix] a brown cat and a white cat on the kitchen floor. + +![](images/a98e31f799ca3ce95990d8a6940b37f52628f827d62715eef13a368cf5feba25.jpg) +[Video-P2P] a brown cat and a white cat on the kitchen floor. + +![](images/2dcbfc1b2fb50ba8069a0b4262ed1ee097ffd8e94607f5dc1df383827dc77897.jpg) + +![](images/7931fb2f1c4e456ffac76f676cc98d4711df3fa0425a83b9cb07992e0a4de6c0.jpg) + +![](images/8037d67c72d3f815b6e0c47c39f81da7a9f740cfb779c84dd6fe840d990fc5ff.jpg) +Figure 7. Video-P2P v.s. Dreamix. Both methods can change the dogs to cats. Video-P2P can preserve background's details. + +![](images/4330f52597e182425ee890153dcf8d0ff448678efe91c3fa302919007b1206a5.jpg) + +![](images/b9d15f53ba02e1af9d095394ce0ee360f706b54a8eb2716da5d8715728727356.jpg) + +![](images/d1bbb4b0354897c5b69534c203038cfdbcfb06fd2db7f60c2846cd22a32ede99.jpg) + +# 4.3. Comparison + +Comparison with Tune-A-Video. Both TAV+DDIM [47] and our Video-P2P allow for video editing with text prompts. However, TAV+DDIM cannot avoid altering the entire video content when editing specific objects, while Video-P2P can edit a local area and minimize the influence on other regions. Fig. 6 (Left) demonstrates that Video-P2P preserves the complex shape of the cloud when replacing a lion with King Kong, whereas TAV+DDIM can only maintain the color tone of the sky in this case. + +Although our model initialization is similar to TAV, Video-P2P can still generate temporal-consistent results where TAV+DDIM fails. As demonstrated in Fig. 6 (Right), TAV struggles to generate a temporally consistent sequence in the second row, even when the inputs are features from DDIM inversion. In contrast, our method can produce better structure-preserved results, as shown in the third row. + +Comparison with Dreamix. In contrast to Dreamix [25], which uses a pre-trained video diffusion model that is not publicly available, our method yields superior results for subject replacement. Although our method cannot perform video motion editing due to the lack of temporal priors, we outperform Dreamix in preserving details and motion consistency. As it is not open-sourced, we conducted our evaluation on its released demo. As demonstrated in Fig. 7, both methods can transform two dogs into two cats, but our method preserves the details of the drawer in the background (the 3rd row). Furthermore, Dreamix may affect the time sequence to some extent, as the generated cat moves more slowly than the original dog in the video. Our method completely preserves the motion of the original video. + +Quantitative results. We create a test set of 45 samples based on the DAVIS [29] dataset and report five metrics for quantitative analysis. The CLIP Score measures the textual similarity between the text prompt and video, while Masked PSNR and LPIPS [51] evaluate the quality of structure preservation. We also proposed a novel metric, Object Semantic Variance (OSV), to measure semantic consistency across frames. In addition, we report an average CLIP similarity (Temp) between consecutive frames (as in Gen-1 [8] and FateZero [30]) to evaluate the temporal coherence. For detailed explanations of these metrics, please refer to the appendix. Our results, as shown in Table 1, demonstrate that Video-P2P performs well on all metrics. Compared to TAV+DDIM, Video-P2P achieves higher Masked PSNR and lower LPIPS, indicating better preservation of unchanged regions. Compared to the other two methods, Video-P2P has much lower OSV and Temp, indicating its superior ability to maintain semantic and temporal consistency across frames. When comparing to Image-P2P, we + +![](images/63f13cb8eb4b0de5cfb2f6ab703cd4770241a7f5f65ece74054a0e6cd63a6550.jpg) + +![](images/20ae2dd82cce5c0158053af32ffd6702a0b78bc6c79217896315e5c560a6f114.jpg) +A dog-cat is running on the grass. + +![](images/81e707be5ed6e1e121f94b9035ca81f74932db19e29fa59caeec1f258f7897aa.jpg) + +![](images/5ef134070c02eab4e058de8e5c502a80675853cdd2ada111a0733108898f91b2.jpg) + +![](images/eb35601c5af7e22c08ba850afdfc7c1f1fc59e585782830772771c739f8853cc.jpg) +Input + +![](images/6f3658fa5fe44419984fafeb680a0cd1ba9307eb888761eb85d2d3a91962b663.jpg) +Image-P2P + +![](images/f9a5d8d30685c62a97f0cba66c3529862cf207800c4236aaf61faad0ad46c5d6.jpg) +w/o model init + +![](images/41d19275b1d888db267b3c33fe9650b2e453093a5946327f7759b8d6b5658b26.jpg) +w/ model init + +![](images/e0b298690c88cbbb66acaf1b791dd345bd7a4dedf09a306fc38ce495e12a6e76.jpg) +Figure 8. Model initialization improves the generation quality. +Figure 9. Ablation on shared unconditional embeddings. + +![](images/e45c31d65878decc4ae696d1408b4d427a2874b76f1d56fb22a641a0131d344e.jpg) +Figure 10. Ablation on decoupled-guidance attention control. + +consider it as a reference point rather than a direct competitor because it does not account for the consistency between different frames. Image-P2P represents the upper bound of PSNR and LPIPS, which serve as metrics to evaluate per-frame quality. Among all video editing methods, Video-P2P is the best one. Moreover, in Tab. 3, we report the user study results, where Video-P2P ranks first on average and has a high preference rate compared to other methods. + +# 4.4. Ablation Study + +Model initialization. While the inflated image diffusion model can generate semantically consistent images, the T2S model's generation ability is compromised during inflation. As seen in Fig.8 (the 3rd column), directly using the inflated T2S model produces unrealistic results with an inaccurate background. To mitigate this, we initialize the T2S model by fine-tuning the given video. This is evident in Fig.8 (4th column), where the cat's appearance improves, and the grass reconstruction becomes more accurate. + +Shared unconditional embedding. The use of shared unconditional embeddings is crucial to ensure stable and consistent generations. On our test set, we found that $10\%$ of the videos exhibit instability when multiple unconditional embeddings are employed. A visual example is included in Fig. 9 (Right). As shown in Tab. 1 (row 5), semantic and temporal consistency are diminished without shared unconditional embedding, reminiscent of the content code in [39]. We observe that optimizing a shared unconditional embedding can significantly improve the PSNR compared to TAV+DDIM in Tab. 2. However, using multiple unconditional embeddings for each frame only increases the PSNR by 0.2 but results in higher parameter usage ( $n$ times) + +
CLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Temp ↑
TAV+DDIM0.334317.370.465155.670.9697
Image-P2P0.327222.920.308276.920.9021
Ours (w/o MI)0.321320.620.325550.270.9460
Ours (w/o DG)0.322618.970.386668.750.9451
Ours (w/o SU)0.335120.640.322148.780.9665
Ours0.336720.650.321347.380.9725
+ +Table 1. Quantitative evaluation. We evaluate textual similarity (CLIP), region preservation (Masked PSNR, LPIPS), semantic consistency (OSV), and temporal consistency (Temp). DG and SU mean Decoupled-Guidance and shared unconditional embeddings. + +
VQVAETAV +DDIMMulti- uncondShared- uncond
PSNR(dB) ↑24.7315.4322.9722.75
#Param. ↓/0.13M22.68M2.94M
+ +Table 2. Video reconstruction quality. A shared unconditional embedding can reconstruct a high-quality video with a small size. + +
Image-P2PTAVTAV+DDIMVideo-P2P
Structure2.673.332.611.39
Preserving13.59%6.52%10.87%69.02%
Text3.402.782.281.54
Alignment3.80%14.13%19.57%62.50%
Realism & Quality3.382.982.211.43
4.35%7.61%19.02%69.02%
+ +Table 3. User study of average ranking $\downarrow$ and preference rate $\uparrow$ . and leads to a lower Masked PSNR after attention control. Thus, we conclude that shared unconditional embedding is the most effective and efficient method for video inversion. + +Decoupled-guidance attention control. To obtain the latent features of the input video, we optimize an unconditional embedding using the source prompt. It is important to note that this embedding is only suitable for the source prompt during the prompt-to-prompt process. Using the optimized embedding for the target prompt may negatively impact the quality of the generated results, as shown in Fig. 10 (Left). Instead, we utilize the initialized unconditional embedding for the target prompt and incorporate attention maps from two branches. The decoupled-guidance attention control approach significantly improves the editing quality, as shown in Fig.10 (Right). Quantitative ablations can be found in Tab. 1 (the 4th row and 6th row). + +# 5. Conclusion + +Video-P2P provides the first framework for video editing with cross-attention control. We optimize a shared unconditional embedding based on a well-initialized T2S model for video inversion. We also propose the decoupled-guidance strategy for attention control. These techniques enable Video-P2P to perform various applications, such as word swap, prompt refinement, and attention re-weighting. + +Acknowledgement Supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095. + +# References + +[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3 +[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723, 2022. 3, 2 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 3 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1 +[5] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3 +[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3 +[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. arXiv preprint arXiv:2305.04441, 2023. 3, 5 +[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3, 7 +[9] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3 +[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106, 2022. 3 +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3 +[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), pages 1-13, 2022. 3 +[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 5 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3 +[16] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3, 4 +[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1, 3, 4 +[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 3 +[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3 +[21] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3 +[22] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3 +[23] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3, 5 +[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1, 2, 3, 4, 5 +[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 3, 7 +[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3 +[28] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF Inter + +national Conference on Computer Vision, pages 2085-2094, 2021.3 +[29] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7 +[30] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3, 7 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 3 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831, 2021. 3 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3 +[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4 +[39] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 8 +[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 3 +[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017. 3 +[42] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi + +Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3 +[43] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2 +[44] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 3 +[45] Wen Wang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3 +[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3 +[47] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 5, 7 +[48] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In ECCV, 2022. 1 +[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3 +[50] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 3 +[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7, 1 +[52] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3 \ No newline at end of file diff --git a/2024/Video-P2P_ Video Editing with Cross-attention Control/images.zip b/2024/Video-P2P_ Video Editing with Cross-attention Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6441393402551f8a8cbe97b16d8281f40ec61005 --- /dev/null +++ b/2024/Video-P2P_ Video Editing with Cross-attention Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2040f1b608d6896c60014d1e7503ce4840ee037ee8a890e27ffc45b23c192b38 +size 926148 diff --git a/2024/Video-P2P_ Video Editing with Cross-attention Control/layout.json b/2024/Video-P2P_ Video Editing with Cross-attention Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..836c980fdd23d46e74766a983b6d75f8a4f4efd7 --- /dev/null +++ b/2024/Video-P2P_ Video Editing with Cross-attention Control/layout.json @@ -0,0 +1,13523 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 467, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 467, + 120 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 467, + 120 + ], + "type": "text", + "content": "Video-P2P: Video Editing with Cross-attention Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 161, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 161, + 156 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 161, + 156 + ], + "type": "text", + "content": "Shaoteng Liu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 143, + 269, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 143, + 269, + 157 + ], + "spans": [ + { + "bbox": [ + 188, + 143, + 269, + 157 + ], + "type": "text", + "content": "Yuechen Zhang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 143, + 349, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 143, + 349, + 156 + ], + "spans": [ + { + "bbox": [ + 296, + 143, + 349, + 156 + ], + "type": "text", + "content": "Wenbo Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 376, + 143, + 421, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 143, + 421, + 156 + ], + "spans": [ + { + "bbox": [ + 376, + 143, + 421, + 156 + ], + "type": "text", + "content": "Zhe Lin3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 447, + 143, + 501, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 447, + 143, + 501, + 156 + ], + "spans": [ + { + "bbox": [ + 447, + 143, + 501, + 156 + ], + "type": "text", + "content": "Jiaya Jia1,2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 157, + 452, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 157, + 452, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 452, + 171 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong 2SmartMore 3 Adobe" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 52, + 196, + 542, + 482 + ], + "blocks": [ + { + "bbox": [ + 208, + 186, + 376, + 196 + ], + "lines": [ + { + "bbox": [ + 208, + 186, + 376, + 196 + ], + "spans": [ + { + "bbox": [ + 208, + 186, + 376, + 196 + ], + "type": "text", + "content": "[Input Video] A child is riding a bike on the road." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 196, + 542, + 482 + ], + "lines": [ + { + "bbox": [ + 52, + 196, + 542, + 482 + ], + "spans": [ + { + "bbox": [ + 52, + 196, + 542, + 482 + ], + "type": "image", + "image_path": "83784f0f7af0642761ccb26b1d60a832c3369678821e7de39fa7f11de19df8c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 490, + 499, + 502 + ], + "lines": [ + { + "bbox": [ + 92, + 490, + 499, + 502 + ], + "spans": [ + { + "bbox": [ + 92, + 490, + 499, + 502 + ], + "type": "text", + "content": "Figure 1. Video-P2P generates new characters while optimally maintaining the pose and environment in videos." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 517, + 192, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 517, + 192, + 529 + ], + "spans": [ + { + "bbox": [ + 143, + 517, + 192, + 529 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 535, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 290, + 715 + ], + "type": "text", + "content": "Video-P2P is the first framework for real-world video editing with cross-attention control. While attention control has proven effective for image editing with pre-trained image generation models, there are currently no large-scale video generation models publicly available. Video-P2P addresses this limitation by adapting an image generation diffusion model to complete various video editing tasks. Specifically, we propose to first tune a Text-to-Set (T2S) model to complete an approximate inversion and then optimize a shared unconditional embedding to achieve accurate video inversion with a small memory cost. We further prove that it is crucial for consistent video editing. For attention control, we introduce a novel decoupled-guidance strategy, which uses different guidance strategies for the source and target prompts. The optimized unconditional embedding for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 518, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 639 + ], + "type": "text", + "content": "the source prompt improves reconstruction ability, while an initialized unconditional embedding for the target prompt enhances editability. Incorporating the attention maps of these two branches enables detailed editing. These technical designs enable various text-driven editing applications, including word swap, prompt refinement, and attention reweighting. Video-P2P works well on real-world videos for generating new characters while optimally preserving their original poses and scenes. It significantly outperforms previous approaches." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 387, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 387, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 387, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Video creation and editing are key tasks [16, 17, 25, 37, 42]. Text-driven editing becomes one promising pipeline. Several methods have demonstrated the ability to edit generated or real-world images with text prompts [13, 19, 24]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8599" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 86, + 71, + 240, + 80 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 71, + 240, + 80 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 240, + 80 + ], + "type": "text", + "content": "[Input Video] A penguin is running on the ice." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 83, + 105, + 137 + ], + "blocks": [ + { + "bbox": [ + 52, + 83, + 105, + 137 + ], + "lines": [ + { + "bbox": [ + 52, + 83, + 105, + 137 + ], + "spans": [ + { + "bbox": [ + 52, + 83, + 105, + 137 + ], + "type": "image", + "image_path": "b37b6980737bf96d85a327052f5d39fb3d04acb9d28bde6473d41972e8e0de2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 111, + 83, + 165, + 137 + ], + "blocks": [ + { + "bbox": [ + 111, + 83, + 165, + 137 + ], + "lines": [ + { + "bbox": [ + 111, + 83, + 165, + 137 + ], + "spans": [ + { + "bbox": [ + 111, + 83, + 165, + 137 + ], + "type": "image", + "image_path": "b93c05dfa17a99af84a4f64c2ac9ef4918987ae77712d476dd8eb1ec1564ec3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 83, + 223, + 137 + ], + "blocks": [ + { + "bbox": [ + 170, + 83, + 223, + 137 + ], + "lines": [ + { + "bbox": [ + 170, + 83, + 223, + 137 + ], + "spans": [ + { + "bbox": [ + 170, + 83, + 223, + 137 + ], + "type": "image", + "image_path": "3f6a423f5441f3998f6074973130d25cfc81007b59aee8133d02a582ec56348e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 229, + 83, + 282, + 137 + ], + "blocks": [ + { + "bbox": [ + 229, + 83, + 282, + 137 + ], + "lines": [ + { + "bbox": [ + 229, + 83, + 282, + 137 + ], + "spans": [ + { + "bbox": [ + 229, + 83, + 282, + 137 + ], + "type": "image", + "image_path": "429ccc9f43f35c525c434644027b2825fdc78dcba46dead1a9035a6a949d9bc2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 139, + 258, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 139, + 258, + 148 + ], + "spans": [ + { + "bbox": [ + 82, + 139, + 258, + 148 + ], + "type": "text", + "content": "[Image-P2P] A Robotic penguin is running on the ice." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 150, + 105, + 203 + ], + "blocks": [ + { + "bbox": [ + 52, + 150, + 105, + 203 + ], + "lines": [ + { + "bbox": [ + 52, + 150, + 105, + 203 + ], + "spans": [ + { + "bbox": [ + 52, + 150, + 105, + 203 + ], + "type": "image", + "image_path": "4c0d6d3ff201311d99c4d451a6b02eacc96865fb29825406bc21f7ab2e688f91.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 111, + 149, + 165, + 203 + ], + "blocks": [ + { + "bbox": [ + 111, + 149, + 165, + 203 + ], + "lines": [ + { + "bbox": [ + 111, + 149, + 165, + 203 + ], + "spans": [ + { + "bbox": [ + 111, + 149, + 165, + 203 + ], + "type": "image", + "image_path": "e3ab97766f73b0a33fd19a59604f56644a2fd68d53c8a6fd7bf124eddb4e070b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 170, + 149, + 223, + 203 + ], + "blocks": [ + { + "bbox": [ + 170, + 149, + 223, + 203 + ], + "lines": [ + { + "bbox": [ + 170, + 149, + 223, + 203 + ], + "spans": [ + { + "bbox": [ + 170, + 149, + 223, + 203 + ], + "type": "image", + "image_path": "cad2ca82be901f2723d6cd79c2c6c53b1abc9b713a2f14c00a60540881c58182.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 229, + 149, + 282, + 203 + ], + "blocks": [ + { + "bbox": [ + 229, + 149, + 282, + 203 + ], + "lines": [ + { + "bbox": [ + 229, + 149, + 282, + 203 + ], + "spans": [ + { + "bbox": [ + 229, + 149, + 282, + 203 + ], + "type": "image", + "image_path": "c906a9290bd92858a094f4bcbf2169d410f7aa2a750feee31c843c11180a00dd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 205, + 255, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 205, + 255, + 215 + ], + "spans": [ + { + "bbox": [ + 82, + 205, + 255, + 215 + ], + "type": "text", + "content": "[Video-P2P] A Robotic penguin is running on the ice." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 52, + 216, + 105, + 270 + ], + "blocks": [ + { + "bbox": [ + 52, + 216, + 105, + 270 + ], + "lines": [ + { + "bbox": [ + 52, + 216, + 105, + 270 + ], + "spans": [ + { + "bbox": [ + 52, + 216, + 105, + 270 + ], + "type": "image", + "image_path": "39c6eb391474c84ec39901ff43d5440df94bf674ae61234100001570524dbe6b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 326 + ], + "type": "text", + "content": "Figure 2. Video-P2P vs Image-P2P. Editing a video frame-by-frame (Image-P2P) cannot guarantee semantic consistency across frames. Video-P2P enables changing the penguin into the same robotic type in every frame." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 111, + 216, + 165, + 270 + ], + "blocks": [ + { + "bbox": [ + 111, + 216, + 165, + 270 + ], + "lines": [ + { + "bbox": [ + 111, + 216, + 165, + 270 + ], + "spans": [ + { + "bbox": [ + 111, + 216, + 165, + 270 + ], + "type": "image", + "image_path": "77922d7b571d70ac66c3223449fbb30850500f481e26cdb1d888959b775461b3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 170, + 216, + 223, + 270 + ], + "blocks": [ + { + "bbox": [ + 170, + 216, + 223, + 270 + ], + "lines": [ + { + "bbox": [ + 170, + 216, + 223, + 270 + ], + "spans": [ + { + "bbox": [ + 170, + 216, + 223, + 270 + ], + "type": "image", + "image_path": "1ef304f30049add608b912260b996b12b6a3e4a15b76d22462b3c0e197afacf8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 229, + 216, + 282, + 270 + ], + "blocks": [ + { + "bbox": [ + 229, + 216, + 282, + 270 + ], + "lines": [ + { + "bbox": [ + 229, + 216, + 282, + 270 + ], + "spans": [ + { + "bbox": [ + 229, + 216, + 282, + 270 + ], + "type": "image", + "image_path": "c067812dd4a8a7b41a7396ef114ef469c93d82907696721247e312b3e1d0d998.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": "Till now, it is still challenging to edit only local objects in a video, such as changing a running \"dog\" into a \"cat\" without influencing the environment. This paper proposes a pipeline that can edit a video both locally and globally, as shown in Figs. 1 and 5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 399, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 544 + ], + "type": "text", + "content": "Text-driven image editing requires a model capable of generating target content, such as changing the category or property of an object. Diffusion models have demonstrated outstanding generation capabilities in this area [3, 13, 19, 40]. Among these methods, attention control emerges as the most effective pipeline for detailed image editing [13, 24]. In order to edit a real image, this pipeline includes two necessary steps: (1) inverting the image into latent features with a pre-trained diffusion model, and (2) controlling attention maps in the denoising process to edit the corresponding parts of the image. For example, by swapping their attention maps, we can replace a \"child\" with a \"panda\"." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "content": "In this paper, we aim to build an attention control-based pipeline for video editing. Since no large-scale pre-trained video generation models are publicly available, we propose a novel framework to show that a pre-trained image diffusion model can be adapted for detailed video editing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "While a pre-trained image diffusion model can be utilized for video editing by processing frames individually (Image-P2P), it lacks semantic consistency across frames (the 2nd row of Fig. 2). To maintain semantic consistency, we propose using a structure on inversion and attention control for all frames, by transforming the Text-to-image diffusion model (T2I) into a Text-to-set model (T2S). This approach is effective, as illustrated in the 3rd row, where the robotic penguin maintains its consistency across frames." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "We adopt the method proposed in [47] to convert a T2I model into a T2S model by altering the convolution kernels and replacing the self-attention with frame-attentions. This conversion yields a model that generates a set of semantically consistent images. The generation quality will be degraded with the inflation step but it can be recovered after tuning on the original video. Although the tuned T2S model is not an ideal video generation model, it suffices to create an approximate inversion for a video as shown in Fig. 3 (c). It is just an approximation because errors are accumulated in the denoising pass, consistent with conclusions in [24, 43]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 229, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 301 + ], + "type": "text", + "content": "To improve the inversion quality, we propose to optimize a shared unconditional embedding for all frames to align the denoising latent features with the diffusion latent features. Our experiments show that shared embedding is efficient for video inversion and crucial for video editing. Comparisons are shown in Fig. 3 and Fig. 9." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 314, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 578 + ], + "type": "text", + "content": "As discussed in [13], successful attention control requires a model to have both reconstruction ability and editability. While image inversion has been argued to possess both abilities in [24], we find that video editing presents different challenges. The T2S model, as an inflation model not trained on any videos, is not robust to the perturbations caused by various unconditional embeddings. Although our optimized embedding can achieve reconstruction, changing prompts can destabilize the model and result in a low-quality generation. On the other hand, we find that the approximate inversion with an initialized unconditional embedding (the text embedding of an empty prompt) is editable but cannot reconstruct well. To address this issue, we propose a decoupled-guidance strategy in attention control, utilizing different guidance strategies for the source and target prompts. Specifically, we use the optimized unconditional embedding for the source prompt and the initialized unconditional embedding for the target prompt. We incorporate the attention maps from these two branches to generate the target video. These two simple designs prove effective and successfully complete video editing. Our contributions can be summarized as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 605, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 306, + 605, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 545, + 642 + ], + "type": "text", + "content": "- We propose the first framework for video editing with attention control. A decoupled-guidance strategy is designed to further improve performance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 647, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 545, + 684 + ], + "type": "text", + "content": "- We introduce an effective and efficient video inversion method with shared unconditional embedding optimization to improve video editing substantially." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "- We conduct extensive ablation studies and comparisons to show the effectiveness of our video editing framework." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8600" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 82, + 170, + 141 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 124, + 80 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 124, + 80 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 124, + 80 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 82, + 170, + 141 + ], + "lines": [ + { + "bbox": [ + 52, + 82, + 170, + 141 + ], + "spans": [ + { + "bbox": [ + 52, + 82, + 170, + 141 + ], + "type": "image", + "image_path": "fb07cff4b777789dd136a51f13c5a28d4eaae1f761c91acbf382fc162bdcc82d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 176, + 82, + 294, + 141 + ], + "blocks": [ + { + "bbox": [ + 217, + 71, + 251, + 80 + ], + "lines": [ + { + "bbox": [ + 217, + 71, + 251, + 80 + ], + "spans": [ + { + "bbox": [ + 217, + 71, + 251, + 80 + ], + "type": "text", + "content": "(b) DDIM" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 82, + 294, + 141 + ], + "lines": [ + { + "bbox": [ + 176, + 82, + 294, + 141 + ], + "spans": [ + { + "bbox": [ + 176, + 82, + 294, + 141 + ], + "type": "image", + "image_path": "e7fdadcd8c74bc72fa00a8af6bf73f17a2403c9e52a9c5102c95a41fdb7028e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 153, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 544, + 175 + ], + "type": "text", + "content": "Figure 3. Inversion Comparison. (b) The inflated model cannot generate high-quality results. (c) Tuning the model can create an approximate video inversion. (d) Optimizing a shared unconditional embedding can accurately reconstruct the input video." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 82, + 417, + 141 + ], + "blocks": [ + { + "bbox": [ + 327, + 71, + 383, + 80 + ], + "lines": [ + { + "bbox": [ + 327, + 71, + 383, + 80 + ], + "spans": [ + { + "bbox": [ + 327, + 71, + 383, + 80 + ], + "type": "text", + "content": "(c) Tune+DDIM" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 82, + 417, + 141 + ], + "lines": [ + { + "bbox": [ + 299, + 82, + 417, + 141 + ], + "spans": [ + { + "bbox": [ + 299, + 82, + 417, + 141 + ], + "type": "image", + "image_path": "9df2f657e8072787d9f9ef0ba23cffd0e0cb61f8178ed334849a358bad59e968.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 424, + 82, + 541, + 141 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 529, + 80 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 529, + 80 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 529, + 80 + ], + "type": "text", + "content": "(d) Ours (Tune+Shared Null)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 82, + 541, + 141 + ], + "lines": [ + { + "bbox": [ + 424, + 82, + 541, + 141 + ], + "spans": [ + { + "bbox": [ + 424, + 82, + 541, + 141 + ], + "type": "image", + "image_path": "437e49b274129baf479c0f39ed63117bdfdf3c0b5b8650f976234841d9b1869d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 186, + 134, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 134, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 134, + 198 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 206, + 180, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 180, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 180, + 217 + ], + "type": "text", + "content": "2.1. Text Driven Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 226, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 287, + 357 + ], + "type": "text", + "content": "DALL-E [32] first considers the text-to-image (T2I) generation task as a sequence-to-sequence translation problem, with subsequent research improving generation quality [6, 10, 50]. Denoising Diffusion Probabilistic Models (DDPMs)[15] have recently gained popularity for T2I. GLIDE[26] utilizes classifier-free guidance to improve text conditioning. DALLE-2 [33] leverages CLIP [31] for better text-image alignment. Latent Diffusion Models (LDMs) [34] propose processing in the latent space to enhance training efficiency. In our work, we employ a pretrained image diffusion model based on LDMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 359, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 288, + 622 + ], + "type": "text", + "content": "Text-to-video (T2V) generation is a nascent research area. GODIVA [46] first introduces VQ-VAE [41] to T2V. CogVideo [18] combines T2V with CogView-2 [6], utilizing pre-trained text-to-image models. Video Diffusion Models (VDM) [17] propose a space-time U-Net for performing diffusion on pixels. Imagen Video [16] successfully generates high-quality videos with cascaded diffusion models and v-prediction parameterization. Phenaki [42] generates videos with time-variable prompts. Make-A-Video [37] combines the appearance generation of T2I models with movement information from video data. While these approaches generate reasonable short videos, they still contain artifacts and do not support real-world video editing. Additionally, most of these approaches are not publicly available at this time. Tune-A-Video [47] inflates an image diffusion model into a video model and tunes it to reconstruct the input video. It allows for changes in semantic content but with limited temporal consistency. We find that using DDIM inversion results can improve its temporal consistency. However, it cannot avoid altering unrelated regions. We adapt some designs of TAV to do our model initialization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 634, + 162, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 162, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 162, + 646 + ], + "type": "text", + "content": "2.2. Text Driven Editing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Generative models have demonstrated impressive performance in image editing, with approaches ranging from GANs [12, 27, 28, 44] to diffusion models [1, 19]. SDEdit [22] adds noise to an input image and uses the diffusion process to recover an edited version. Prompt-to" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 186, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 317 + ], + "type": "text", + "content": "Prompt [13] uses attention control to minimize changes to unrelated parts, while Null-Text Inversion [24] improves real image editing. Plug-and-Play [40] uses feature and self-attention injection, which are unable to avoid editing the whole image when changing a local part. InstructPix2Pix [3] enables flexible text-driven editing with user-provided instructions. Textual Inversion [11], DreamBooth [35], and Custom-Diffusion [20] learn special tokens for personalized concepts and generate related images. Some improved inversion methods [7, 23] are proposed after our work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 320, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 546, + 464 + ], + "type": "text", + "content": "Video editing with generative models has seen several advances recently. Text2Live [2] employs CLIP to edit textures in videos but struggles with significant semantic changes. Dreamix [25] uses a pre-trainedImagen Video [16] backbone to perform image-to-video and video-to-video editing, with the ability to change motion as well. Gen-1 [8] trains models jointly on images and videos for tasks such as stylization and customization. While these methods enable modifying video content, they operate like guided generation and tend to modify all regions together when editing an object. Our proposed method allows for local editing with a diffusion model pre-trained on images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 467, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 525 + ], + "type": "text", + "content": "Besides, there have been some concurrent works and followers after our work, such as [5, 9, 21, 30, 45, 49, 52]. Video-P2P focuses more on introducing the image diffusion models to detailed video editing, instead of conditional video generation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 544, + 361, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 361, + 556 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 361, + 556 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " be a real video containing " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " frames. We adopt the Prompt-to-Prompt setting by introducing a source prompt " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " and an edited prompt " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^*" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " which together generate an edited video " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^*" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " frames. The prompts are provided by the user. Similar to TAV [47], we assume that the object of interest is present in the first frame." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "To achieve cross-attention control in video editing, we propose Video-P2P, a framework with two key technical designs: (1) optimizing a shared unconditional embedding for video inversion, and (2) using different guidance for the source and edited prompts, and incorporating their attention maps. The framework is illustrated in Fig. 4." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8601" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 127, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 127, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 127, + 85 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "spans": [ + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs). LDMs generate an image latent " + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "text", + "content": " using a random noise vector " + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "text", + "content": " and a textual condition " + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 93, + 287, + 154 + ], + "type": "text", + "content": " as inputs. As variants of DDPMs, these models aim to predict artificial noise by minimizing the following objective:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 160, + 287, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 160, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 54, + 160, + 287, + 179 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} E _ {z _ {0}, \\varepsilon \\sim N (0, I), t \\sim \\text {U n i f o r m} (1, T)} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "633345d99c7c07c082b0f242675ec17ddcbad3afc6bc7598b82423ffb1a0bd0d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\psi(\\mathcal{P})" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " is the embedding of the text prompt, and noise " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " according to step " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": ". During inference, the model predicts noise " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\varepsilon_\\theta(\\cdot)" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " steps to generate an image from " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "text", + "content": "DDIM sampling and inversion. Deterministic DDIM sampling can be used to generate an image from latent features in a small number of denoising steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\varepsilon_ {\\theta} (z _ {t}, t, \\mathcal {C}). \\tag {2}", + "image_path": "1038664f51ebe70cbb9209c906ea7eab9142d9881f4d84f4c4f33def2e59e8f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 319, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 368 + ], + "type": "text", + "content": "We use an encoder to encode the real image before the diffusion process and a decoder to decode after the denoising process. DDIM sampling can be reversed in a few steps through the equation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 374, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 287, + 411 + ], + "type": "interline_equation", + "content": "z _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\varepsilon_ {\\theta} (z _ {t}, t, \\mathcal {C}), \\tag {3}", + "image_path": "ba256079254c8a73b1e23c150bcd7a0b023821a12b4db1976bfa6e7d974767bc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 412, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 436 + ], + "type": "text", + "content": "known as DDIM inversion [38]. This can be used to obtain the corresponding latent features of a real image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": "Null-text inversion. To mitigate the amplification effect of text conditioning during image generation, classifier-free guidance is proposed, which performs unconditional prediction [14]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 496, + 286, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 496, + 286, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 496, + 286, + 519 + ], + "type": "interline_equation", + "content": "\\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = w \\cdot \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) + (1 - w) \\cdot \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right), \\tag {4}", + "image_path": "293a30a7a788bd28364e41461c8af32c22a134f907640a06f1f1e81b64410d30.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\varnothing = \\psi(\"\")" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": " is the embedding of a null text and " + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": " is the guidance weight. However, the classifier-free guidance increases errors accumulated in the denoising process, leading to imperfect image reconstruction using the DDIM inversion. [24] proposes to align the diffusion latent trajectory " + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "inline_equation", + "content": "z_{T}^{*},\\ldots ,z_{0}^{*}" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": " with the denoising latent trajectory " + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "inline_equation", + "content": "z_{T},\\ldots ,z_{0}" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": " by optimizing a step-wise unconditional embedding " + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\varnothing_{t}" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 121, + 612, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 612, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 121, + 612, + 287, + 632 + ], + "type": "interline_equation", + "content": "\\min _ {\\varnothing_ {t}} \\left\\| z _ {t - 1} ^ {*} - z _ {t - 1} \\right\\| _ {2} ^ {2}. \\tag {5}", + "image_path": "493f30ae16bd391cd78c4f8be69e83d72dcd7b90961de18ee4a351a7aa5564d0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 647, + 143, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 143, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 143, + 658 + ], + "type": "text", + "content": "3.2. Video Inversion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "We begin by constructing a T2S model that is capable of performing an approximate inversion. Following the VDM baselines [16, 17] and TAV [47], we employ " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 3 \\times 3" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " pattern convolution kernels and temporal attention. Moreover, we" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 545, + 216 + ], + "type": "image", + "image_path": "e9b9c12e560e4ca378f1536dfc72f7ecbdc928da9faf8afdce1b59ee5adc6b31.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 227, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 272 + ], + "type": "text", + "content": "Figure 4. Framework. We optimize one shared unconditional embedding for the reconstruct branch (orange). The initialized unconditional embedding is utilized for the editable branch (green). Their attention maps are incorporated to create the target video." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "text", + "content": "replace the self-attention with frame-attentions, which take the first frames " + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "text", + "content": " and the current frame " + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "text", + "content": " as inputs and update features for the frame " + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 280, + 545, + 327 + ], + "type": "text", + "content": ". The formulation of the frame-attention is as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 333, + 545, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 333, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 343, + 333, + 545, + 347 + ], + "type": "interline_equation", + "content": "Q = W ^ {Q} \\mathbf {v} _ {i}, K = W ^ {K} \\mathbf {v} _ {0}, V = W ^ {V} \\mathbf {v} _ {0}, \\tag {6}", + "image_path": "a223ff7ced34e053d1e8c09998781d7fd2e13d0b7aa7023692bfb932060bc028.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 354, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 354, + 545, + 376 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 354, + 545, + 376 + ], + "type": "text", + "content": " are the projection matrices in attention. Then, attention maps are calculated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 370, + 382, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 382, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 370, + 382, + 545, + 410 + ], + "type": "interline_equation", + "content": "M = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right), \\tag {7}", + "image_path": "fa201d70ba01a14052e3fa2f9bcf2fccb26fa89e6f702147b443e70951b43a69.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": " is the latent projection dimension. The model processes a video pair-by-pair and computes " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": " times to obtain the prediction for every frame. We find that the simple design suffices for video inversion since the reversed latent features can capture temporal information. Besides, frame-attention conserves memory and speeds up the process." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": "While model inflation can aid in preserving semantic consistency across frames, it adversely impacts the generation quality of the T2I model. This is because the self-attention parameters are utilized to compute frame correlations, which have not been pre-trained. Consequently, the T2S model, generated through inflation, is insufficient for the approximate inversion, as demonstrated in Fig. 2. To address this, we fine-tune the query projection matrices " + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "inline_equation", + "content": "W^{Q}" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": " of the frame- and cross-attentions, as well as additional temporal attention, to perform noise prediction based on the input video following [47]. After this initialization, the T2S model is capable of generating semantically consistent image sets while maintaining the quality of each frame, resulting in successful approximate inversion." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Using the fine-tuned T2S model, we perform video inversion by optimizing a shared unconditional embedding. During inversion, each latent feature " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " contains a channel for the frames with dimension " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "z_{t,i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " denotes the latent feature for the " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "-th frame. We employ DDIM inversion" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8602" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "to generate latent features " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "z_0^*, \\ldots, z_T^*" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": ". The unconditional embedding is defined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 102, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\min _ {\\varnothing_ {t}} \\sum_ {i = 1} ^ {n} \\left| \\left| z _ {t - 1, i} ^ {*} - z _ {t - 1, i} \\left(\\bar {z} _ {t, i}, \\bar {z} _ {t, 0}, \\varnothing_ {t}, \\mathcal {C}\\right) \\right| \\right| _ {2} ^ {2}, \\text {w h e r e} \\tag {8}", + "image_path": "fb2b22603c44fb191f1aec8f4d2e3a3d53ec2bd25678b7f3ae2ede4d3e9b4e97.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 142, + 287, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 287, + 155 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {t - 1, i} = z _ {t - 1, i} \\left(\\bar {z} _ {t, i}, \\bar {z} _ {t, 0}, \\emptyset_ {t}, \\mathcal {C}\\right) \\tag {9}", + "image_path": "cba5c933b0292352c620c9d951bfe7b7869b7c13f0af1bcf56dd3c07c982f334.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "type": "text", + "content": "is updated at each step. The T2S model's frame-attentions use two latent features to calculate the corresponding feature for the next step. Notice " + }, + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\varnothing_{t}" + }, + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "type": "text", + "content": " is shared by all frames " + }, + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "type": "inline_equation", + "content": "(i = 1,\\dots ,n)" + }, + { + "bbox": [ + 47, + 158, + 287, + 229 + ], + "type": "text", + "content": " which minimizes the memory usage. Besides, using the same unconditional embedding for all frames avoids destabilizing the semantic consistency in attention control." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 236, + 251, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 251, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 251, + 248 + ], + "type": "text", + "content": "3.3. Decoupled-guidance Attention Control" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 254, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 492 + ], + "type": "text", + "content": "To perform attention control on real images, existing works [13, 24] require an inference pipeline with both reconstruction ability and editability. However, achieving such a pipeline for a T2S model is challenging. Video inversion allows us to establish an inference pipeline to reconstruct the original video well. However, the T2S model is not as robust as T2I models due to a lack of pre-training with videos. As a result, its editability is compromised with the optimized unconditional embedding, leading to degraded generation quality when changing prompts. In contrast, we find that using an initialized unconditional embedding " + }, + { + "bbox": [ + 46, + 254, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 46, + 254, + 287, + 492 + ], + "type": "text", + "content": " makes the model more editable while it cannot reconstruct perfectly. This inspires us to combine the abilities of two inference pipelines. For the source prompt, we use the optimized unconditional embedding in the classifier-free guidance. For the target prompt, we choose the initialized unconditional embedding. We then incorporate attention maps from these two branches to obtain the edited video, where the unchanged parts are influenced by the source branch and the edited parts are influenced by the target branch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 493, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 540 + ], + "type": "text", + "content": "The pseudo algorithm is shown in Alg. 1. We adopt the attention control methods from Image-P2P to Video-P2P. For example, to perform word swap, the Edit function can be represented as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 546, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 546, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 80, + 546, + 287, + 578 + ], + "type": "interline_equation", + "content": "E d i t \\left(M _ {t}, M _ {t} ^ {*}, t\\right) := \\left\\{ \\begin{array}{l l} M _ {t} ^ {*} & \\text {i f} t < \\tau \\\\ M _ {t} & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {10}", + "image_path": "b2771703cfa57536e99b68b14b4b9e28fb29e01dcb2734fe8938d0d4596a1b0f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "M_{t}^{*}" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " are the cross-attention maps for every frame at every step, and " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "DM" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " is the tuned T2S model. Changing the frame-attention maps has a small influence on the final results. Attention maps are swapped only for the first " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " steps because attentions are formed in the early period. " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\overline{M}_{t,w}" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " is the average attention map of the word " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " calculated at step " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": ". It is averaged over steps " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "T,\\ldots ,t" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": " independently for every frame. For the " + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": "-th frame, we calculate:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 684, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 684, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 67, + 684, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\bar {M} _ {t, w, j} = \\frac {1}{T - t} \\sum_ {i = t} ^ {T} \\bar {M} _ {i, w, j} \\quad j = 1, \\dots , n. \\tag {11}", + "image_path": "ac0fa3e3b09f2d4fdff2950035fbfcd8e6054f5b1ae1f9dd76e2812b12dbd43d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 310, + 90, + 511, + 293 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 501, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 501, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 501, + 87 + ], + "type": "text", + "content": "Algorithm 1: Prompt-to-Prompt video editing" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "lines": [ + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "spans": [ + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": "1 Input: A source prompt " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " a target prompt " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^*" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " source video " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "V_{src}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n2 Output: Edited video " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "V_{dst}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n3 Latent features from DDIM inversion: " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "z_{T}^{*}\\gets z_{T}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n5 Initialized unconditional embedding " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "\\varnothing^{*}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " and optimized unconditional embedding " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " . \n6 for " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "t = T,T - 1,\\ldots ,1" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " do \n7 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "z_{t - 1},M_t\\gets DM(z_t,\\mathcal{P},t,\\emptyset)" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "M_t^* \\leftarrow DM(z_t^*,\\mathcal{P}^*,t,\\emptyset^*)" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n9 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "\\widehat{M}_t\\gets \\operatorname {Edit}(M_t,M_t^*,t)" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n10 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "z_{t - 1}^{*}\\gets DM(z_{t}^{*},\\mathcal{P}^{*},t,\\emptyset^{*})\\{M_{t}^{*}\\gets \\widehat{M}_{t}\\}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n11 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "\\alpha \\gets B(\\overline{M}_{t,w})\\cup B(\\overline{M}_{t,w}^{*})" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n12 " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "z_{t - 1}^{*}\\gets (1 - \\alpha)\\odot z_{t - 1} + \\alpha \\odot z_{t - 1}^{*}" + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "text", + "content": " \n13 end \n14 Return " + }, + { + "bbox": [ + 310, + 90, + 511, + 293 + ], + "type": "inline_equation", + "content": "(z_0,z_0^*)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 304, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 545, + 330 + ], + "type": "inline_equation", + "content": "B\\left(\\overline{M}_{t,w}\\right)" + }, + { + "bbox": [ + 306, + 304, + 545, + 330 + ], + "type": "text", + "content": " represents the binary mask obtained from the attention map. A value is set to 1 when larger than a threshold." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 339, + 386, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 386, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 386, + 353 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 359, + 438, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 359, + 438, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 359, + 438, + 371 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 376, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 617 + ], + "type": "text", + "content": "We develop our method based on CompVis Stable Diffusion (v1-5). Similar to TAV [47], we fix the image autoencoder and sample 8 or 24 frames at the resolution of " + }, + { + "bbox": [ + 304, + 376, + 545, + 617 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 376, + 545, + 617 + ], + "type": "text", + "content": " from a video. To initialize the model, we fine-tune the T2S model for 500 steps to reconstruct the original video. During attention control, we set the cross-attention replacing ratio to 0.4 and the attention threshold to 0.3. For prompt refinement, we set the refinement ratio to 0.4. Parameters can be adjusted to control the editing fidelity for different examples. All 8-frame experiments are conducted on a single V100 GPU, with 5 minutes for initialization (tuning), 6 minutes for inversion, and 1 minute for inference. It is difficult to avoid the inversion cost when preserving the background faithfully according to [24]. To speed up to less than one minute, one can replace the null-text inversion with recent inversion methods [7, 23]. In the appendix, we show that a shared unconditional embedding works on 24 frames. Our inversion process extends to over 100 frames. However, editing longer videos (1min) remains challenging. VideoP2P and most existing works concentrate on shorter videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 623, + 388, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 388, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 388, + 635 + ], + "type": "text", + "content": "4.2. Applications" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Video-P2P can be utilized for a range of editing applications, including prompt refinement, attention re-weighting, and word swapping. Video-P2P is able to maintain semantic consistency across different frames and preserve the temporal coherence of the original video during the editing process. More examples can be found in the appendix." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8603" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 71, + 208, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 71, + 208, + 81 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 208, + 81 + ], + "type": "text", + "content": "[Input Video] A car is driving on the road." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 83, + 111, + 142 + ], + "blocks": [ + { + "bbox": [ + 52, + 83, + 111, + 142 + ], + "lines": [ + { + "bbox": [ + 52, + 83, + 111, + 142 + ], + "spans": [ + { + "bbox": [ + 52, + 83, + 111, + 142 + ], + "type": "image", + "image_path": "d041f0dc691e815d07d43bb27ac3d4388932856f29f2c8767775e4a1a61036ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 112, + 83, + 171, + 142 + ], + "blocks": [ + { + "bbox": [ + 112, + 83, + 171, + 142 + ], + "lines": [ + { + "bbox": [ + 112, + 83, + 171, + 142 + ], + "spans": [ + { + "bbox": [ + 112, + 83, + 171, + 142 + ], + "type": "image", + "image_path": "424b100a8c1eff47e2d61b38e5422e7e94eb92ec8699ebc48c79275008374eab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 83, + 233, + 142 + ], + "blocks": [ + { + "bbox": [ + 173, + 83, + 233, + 142 + ], + "lines": [ + { + "bbox": [ + 173, + 83, + 233, + 142 + ], + "spans": [ + { + "bbox": [ + 173, + 83, + 233, + 142 + ], + "type": "image", + "image_path": "466a4d96970a114f7fa2ab8fdcf444b9ef667a5e84615b0cf4cc9e707b658f6c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 236, + 83, + 295, + 142 + ], + "blocks": [ + { + "bbox": [ + 236, + 83, + 295, + 142 + ], + "lines": [ + { + "bbox": [ + 236, + 83, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 236, + 83, + 295, + 142 + ], + "type": "image", + "image_path": "89c0496cf3125144f4efa6c4d4846fba160be89950a59e061de466be9d803890.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 144, + 183, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 144, + 183, + 154 + ], + "spans": [ + { + "bbox": [ + 52, + 144, + 183, + 154 + ], + "type": "text", + "content": "A car is driving on the flooded road." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 156, + 110, + 213 + ], + "blocks": [ + { + "bbox": [ + 52, + 156, + 110, + 213 + ], + "lines": [ + { + "bbox": [ + 52, + 156, + 110, + 213 + ], + "spans": [ + { + "bbox": [ + 52, + 156, + 110, + 213 + ], + "type": "image", + "image_path": "526e11e454db70c56885a81ec7c68bd2e02d47ea3dd52e4ff8eef900d2291696.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 112, + 156, + 171, + 213 + ], + "blocks": [ + { + "bbox": [ + 112, + 156, + 171, + 213 + ], + "lines": [ + { + "bbox": [ + 112, + 156, + 171, + 213 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 171, + 213 + ], + "type": "image", + "image_path": "542b1a64aa5dc9e6892e5717be1c39e6209a98c2fabff79f76f5e01b75e3267d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 174, + 156, + 233, + 213 + ], + "blocks": [ + { + "bbox": [ + 174, + 156, + 233, + 213 + ], + "lines": [ + { + "bbox": [ + 174, + 156, + 233, + 213 + ], + "spans": [ + { + "bbox": [ + 174, + 156, + 233, + 213 + ], + "type": "image", + "image_path": "4ce6484cc75d9b254c1ebb01583e4b2478efbfc8fea02e5d1687c8a09adec475.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 156, + 295, + 213 + ], + "blocks": [ + { + "bbox": [ + 236, + 156, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 236, + 156, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 236, + 156, + 295, + 213 + ], + "type": "image", + "image_path": "ee2db35a88659841986f543fe31609d400fce813c17a67ae86c92d92a9fbc720.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 220, + 249, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 220, + 249, + 230 + ], + "spans": [ + { + "bbox": [ + 52, + 220, + 249, + 230 + ], + "type": "text", + "content": "[Input Video] A man drives a motorbike in the forest." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 52, + 232, + 110, + 289 + ], + "blocks": [ + { + "bbox": [ + 52, + 232, + 110, + 289 + ], + "lines": [ + { + "bbox": [ + 52, + 232, + 110, + 289 + ], + "spans": [ + { + "bbox": [ + 52, + 232, + 110, + 289 + ], + "type": "image", + "image_path": "450e6c5c0d5b59204572be19b7838591492fcbd5298c024b550dc21fb7eb40b2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 112, + 232, + 171, + 289 + ], + "blocks": [ + { + "bbox": [ + 112, + 232, + 171, + 289 + ], + "lines": [ + { + "bbox": [ + 112, + 232, + 171, + 289 + ], + "spans": [ + { + "bbox": [ + 112, + 232, + 171, + 289 + ], + "type": "image", + "image_path": "1d018e862a6d53557a848da1367e00f2c867156159411515e8d943f844b1a978.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 174, + 232, + 233, + 289 + ], + "blocks": [ + { + "bbox": [ + 174, + 232, + 233, + 289 + ], + "lines": [ + { + "bbox": [ + 174, + 232, + 233, + 289 + ], + "spans": [ + { + "bbox": [ + 174, + 232, + 233, + 289 + ], + "type": "image", + "image_path": "4eda895ca1663d2ff13acff297a478ce69f2f2d691c5baf2bd8a1e499ea15211.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 236, + 232, + 295, + 289 + ], + "blocks": [ + { + "bbox": [ + 236, + 232, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 236, + 232, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 236, + 232, + 295, + 289 + ], + "type": "image", + "image_path": "ee7f7b5fde13fd494fc2c3cd331b5d599d0154482a719072766efda644df1399.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 52, + 293, + 228, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 293, + 228, + 302 + ], + "spans": [ + { + "bbox": [ + 52, + 293, + 228, + 302 + ], + "type": "text", + "content": "A man Bat Man drives a motorbike in the forest." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 52, + 303, + 110, + 361 + ], + "blocks": [ + { + "bbox": [ + 52, + 303, + 110, + 361 + ], + "lines": [ + { + "bbox": [ + 52, + 303, + 110, + 361 + ], + "spans": [ + { + "bbox": [ + 52, + 303, + 110, + 361 + ], + "type": "image", + "image_path": "c8a9af957f6359e04aba3601a2dae483d4e3984da2d769ca7047b71b1499347e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 112, + 303, + 171, + 361 + ], + "blocks": [ + { + "bbox": [ + 112, + 303, + 171, + 361 + ], + "lines": [ + { + "bbox": [ + 112, + 303, + 171, + 361 + ], + "spans": [ + { + "bbox": [ + 112, + 303, + 171, + 361 + ], + "type": "image", + "image_path": "d61b009e0be12b74b2071726c3a0f03a78cce723ebd3207c91337c0785804faa.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 174, + 303, + 233, + 361 + ], + "blocks": [ + { + "bbox": [ + 174, + 303, + 233, + 361 + ], + "lines": [ + { + "bbox": [ + 174, + 303, + 233, + 361 + ], + "spans": [ + { + "bbox": [ + 174, + 303, + 233, + 361 + ], + "type": "image", + "image_path": "62f1de7a0f07a89ea7eeecce0010a7de316a06080b4529be255e158d2a451506.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 236, + 303, + 295, + 361 + ], + "blocks": [ + { + "bbox": [ + 236, + 303, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 236, + 303, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 236, + 303, + 295, + 361 + ], + "type": "image", + "image_path": "4ed69f4db463c83b5ede3a0484259d05aee2273625fd7df42751a24a5829d6c2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 52, + 370, + 216, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 370, + 216, + 380 + ], + "spans": [ + { + "bbox": [ + 52, + 370, + 216, + 380 + ], + "type": "text", + "content": "[Input Video] A dog is running on the grass." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 52, + 381, + 110, + 439 + ], + "blocks": [ + { + "bbox": [ + 52, + 381, + 110, + 439 + ], + "lines": [ + { + "bbox": [ + 52, + 381, + 110, + 439 + ], + "spans": [ + { + "bbox": [ + 52, + 381, + 110, + 439 + ], + "type": "image", + "image_path": "31e2ed7092a4f51d00dc41353bef6c3e5cdec06692bded3ca6189c354b8bd335.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 112, + 381, + 171, + 439 + ], + "blocks": [ + { + "bbox": [ + 112, + 381, + 171, + 439 + ], + "lines": [ + { + "bbox": [ + 112, + 381, + 171, + 439 + ], + "spans": [ + { + "bbox": [ + 112, + 381, + 171, + 439 + ], + "type": "image", + "image_path": "d23fa563541f6e0f41956f1bd86883b0c472adb351e26d930dc1da575e93abe1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 174, + 381, + 233, + 439 + ], + "blocks": [ + { + "bbox": [ + 174, + 381, + 233, + 439 + ], + "lines": [ + { + "bbox": [ + 174, + 381, + 233, + 439 + ], + "spans": [ + { + "bbox": [ + 174, + 381, + 233, + 439 + ], + "type": "image", + "image_path": "c241ba8214ee7849a4de823ba1d386cddc0ec81b77199cbe40faa65b2430e320.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 236, + 381, + 295, + 439 + ], + "blocks": [ + { + "bbox": [ + 236, + 381, + 295, + 439 + ], + "lines": [ + { + "bbox": [ + 236, + 381, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 236, + 381, + 295, + 439 + ], + "type": "image", + "image_path": "9aeaaa0dcde9fbfc4555d2a07e0c47c76d799fd563b8b3abcf77750d30c87777.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 52, + 442, + 192, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 442, + 192, + 453 + ], + "spans": [ + { + "bbox": [ + 52, + 442, + 192, + 453 + ], + "type": "text", + "content": "A robotic dog is running on the grass." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 52, + 454, + 110, + 513 + ], + "blocks": [ + { + "bbox": [ + 52, + 454, + 110, + 513 + ], + "lines": [ + { + "bbox": [ + 52, + 454, + 110, + 513 + ], + "spans": [ + { + "bbox": [ + 52, + 454, + 110, + 513 + ], + "type": "image", + "image_path": "8df7d95a573c3577bb5c625f4f4cb28c7729c4a406a727daff8d92bf0480c982.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 521, + 508, + 533 + ], + "lines": [ + { + "bbox": [ + 83, + 521, + 508, + 533 + ], + "spans": [ + { + "bbox": [ + 83, + 521, + 508, + 533 + ], + "type": "text", + "content": "Figure 5. Videos edited by Video-P2P with text prompts. Video-P2P can do both word swaps and prompt refinement." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 112, + 455, + 171, + 513 + ], + "blocks": [ + { + "bbox": [ + 112, + 455, + 171, + 513 + ], + "lines": [ + { + "bbox": [ + 112, + 455, + 171, + 513 + ], + "spans": [ + { + "bbox": [ + 112, + 455, + 171, + 513 + ], + "type": "image", + "image_path": "1bd63a0690d6e3a9a1367b06b81b2d653a4d5c6ca82d90f3e2e52e34968042f4.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 174, + 455, + 233, + 513 + ], + "blocks": [ + { + "bbox": [ + 174, + 455, + 233, + 513 + ], + "lines": [ + { + "bbox": [ + 174, + 455, + 233, + 513 + ], + "spans": [ + { + "bbox": [ + 174, + 455, + 233, + 513 + ], + "type": "image", + "image_path": "0938b734dd91c8c58dc07c4a80d575f39745ff41699115c11f014d12fa5be87e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 236, + 455, + 295, + 513 + ], + "blocks": [ + { + "bbox": [ + 236, + 455, + 295, + 513 + ], + "lines": [ + { + "bbox": [ + 236, + 455, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 236, + 455, + 295, + 513 + ], + "type": "image", + "image_path": "cbd8d637c0642b7f918bef4749d72518611f4c2d10f4c7a8b7c826532908aefd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 72, + 430, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 72, + 430, + 81 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 430, + 81 + ], + "type": "text", + "content": "A car is driving on the road railway." + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 299, + 83, + 358, + 142 + ], + "blocks": [ + { + "bbox": [ + 299, + 83, + 358, + 142 + ], + "lines": [ + { + "bbox": [ + 299, + 83, + 358, + 142 + ], + "spans": [ + { + "bbox": [ + 299, + 83, + 358, + 142 + ], + "type": "image", + "image_path": "5c0561a773a7ad13eba901e9981878279c92894b7144e3bbe1499c6138e844c5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 361, + 83, + 419, + 142 + ], + "blocks": [ + { + "bbox": [ + 361, + 83, + 419, + 142 + ], + "lines": [ + { + "bbox": [ + 361, + 83, + 419, + 142 + ], + "spans": [ + { + "bbox": [ + 361, + 83, + 419, + 142 + ], + "type": "image", + "image_path": "29c081042cc8388248e695d5517cde1129ecf5d5ad4bf9a7b9426596d2c71ae8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 422, + 83, + 481, + 142 + ], + "blocks": [ + { + "bbox": [ + 422, + 83, + 481, + 142 + ], + "lines": [ + { + "bbox": [ + 422, + 83, + 481, + 142 + ], + "spans": [ + { + "bbox": [ + 422, + 83, + 481, + 142 + ], + "type": "image", + "image_path": "70f87d2b08a5002336c7dd58f76466b8a4d8ca8982a0e3e21c1ff9270355c147.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 484, + 83, + 542, + 142 + ], + "blocks": [ + { + "bbox": [ + 484, + 83, + 542, + 142 + ], + "lines": [ + { + "bbox": [ + 484, + 83, + 542, + 142 + ], + "spans": [ + { + "bbox": [ + 484, + 83, + 542, + 142 + ], + "type": "image", + "image_path": "885f9c8e9ae5858269988a7610bf7c7cbf452af985c929fa0174d95b96b9d9e5.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 301, + 144, + 438, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 144, + 438, + 154 + ], + "spans": [ + { + "bbox": [ + 301, + 144, + 438, + 154 + ], + "type": "text", + "content": "A car is driving on the road at sunset." + } + ] + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 299, + 156, + 358, + 213 + ], + "blocks": [ + { + "bbox": [ + 299, + 156, + 358, + 213 + ], + "lines": [ + { + "bbox": [ + 299, + 156, + 358, + 213 + ], + "spans": [ + { + "bbox": [ + 299, + 156, + 358, + 213 + ], + "type": "image", + "image_path": "cc667d6683e2643ad402feac4a1604edacd2a8f4d4a38492ce997d17e653266e.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 361, + 156, + 419, + 213 + ], + "blocks": [ + { + "bbox": [ + 361, + 156, + 419, + 213 + ], + "lines": [ + { + "bbox": [ + 361, + 156, + 419, + 213 + ], + "spans": [ + { + "bbox": [ + 361, + 156, + 419, + 213 + ], + "type": "image", + "image_path": "38779350364a4198756abaee8242f7ba825a07e65f7834aee885ffabc7a50925.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 422, + 156, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 422, + 156, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 422, + 156, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 422, + 156, + 481, + 213 + ], + "type": "image", + "image_path": "f4b28a78f4b1e69b5d9200496e7d91cf1211c71b00e52539ebc538089dfef04d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 484, + 156, + 542, + 213 + ], + "blocks": [ + { + "bbox": [ + 484, + 156, + 542, + 213 + ], + "lines": [ + { + "bbox": [ + 484, + 156, + 542, + 213 + ], + "spans": [ + { + "bbox": [ + 484, + 156, + 542, + 213 + ], + "type": "image", + "image_path": "0f6144bfdd5f849dde10a2db9439c60aa083276237339ab558a28c25c13913c6.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 299, + 219, + 462, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 219, + 462, + 229 + ], + "spans": [ + { + "bbox": [ + 299, + 219, + 462, + 229 + ], + "type": "text", + "content": "A man drives a Lego motorbike in the forest." + } + ] + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 299, + 232, + 358, + 289 + ], + "blocks": [ + { + "bbox": [ + 299, + 232, + 358, + 289 + ], + "lines": [ + { + "bbox": [ + 299, + 232, + 358, + 289 + ], + "spans": [ + { + "bbox": [ + 299, + 232, + 358, + 289 + ], + "type": "image", + "image_path": "21e2aa76875d25aed32e4f31ba398c6ff8e183c8755d19ead1741589407bf69d.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 361, + 232, + 419, + 289 + ], + "blocks": [ + { + "bbox": [ + 361, + 232, + 419, + 289 + ], + "lines": [ + { + "bbox": [ + 361, + 232, + 419, + 289 + ], + "spans": [ + { + "bbox": [ + 361, + 232, + 419, + 289 + ], + "type": "image", + "image_path": "17ef09c21f078b5e003af01a14afb06e3ffbd54997180c011cbfadf3658ea0fc.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 422, + 232, + 481, + 289 + ], + "blocks": [ + { + "bbox": [ + 422, + 232, + 481, + 289 + ], + "lines": [ + { + "bbox": [ + 422, + 232, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 422, + 232, + 481, + 289 + ], + "type": "image", + "image_path": "4137e19e1ae31754272c2de9a8010a19f70a53df2fab79a007fffa52ec8763ac.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 484, + 232, + 542, + 289 + ], + "blocks": [ + { + "bbox": [ + 484, + 232, + 542, + 289 + ], + "lines": [ + { + "bbox": [ + 484, + 232, + 542, + 289 + ], + "spans": [ + { + "bbox": [ + 484, + 232, + 542, + 289 + ], + "type": "image", + "image_path": "40e9cb9081a2ba7698eced0a87dff02dc9d0595c59e02e408a621ba68188bffb.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 299, + 293, + 485, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 293, + 485, + 302 + ], + "spans": [ + { + "bbox": [ + 299, + 293, + 485, + 302 + ], + "type": "text", + "content": "A man Spider Man drives a motorbike in the forest." + } + ] + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 299, + 303, + 358, + 361 + ], + "blocks": [ + { + "bbox": [ + 299, + 303, + 358, + 361 + ], + "lines": [ + { + "bbox": [ + 299, + 303, + 358, + 361 + ], + "spans": [ + { + "bbox": [ + 299, + 303, + 358, + 361 + ], + "type": "image", + "image_path": "c41fd342c23d913120c85d6070e697852bfe03aa6e027d6c34dd3a9c954ec39f.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 361, + 303, + 419, + 361 + ], + "blocks": [ + { + "bbox": [ + 361, + 303, + 419, + 361 + ], + "lines": [ + { + "bbox": [ + 361, + 303, + 419, + 361 + ], + "spans": [ + { + "bbox": [ + 361, + 303, + 419, + 361 + ], + "type": "image", + "image_path": "6533b27b862229ad1f1f31aacc7cca91fb53e02e352a96fd27393b4b0a92c172.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 422, + 303, + 481, + 361 + ], + "blocks": [ + { + "bbox": [ + 422, + 303, + 481, + 361 + ], + "lines": [ + { + "bbox": [ + 422, + 303, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 422, + 303, + 481, + 361 + ], + "type": "image", + "image_path": "bb8e620954ea7c5c7dd18584ef72b8b264f93c55b5b536990f67335f7b5ac6a3.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 484, + 303, + 542, + 361 + ], + "blocks": [ + { + "bbox": [ + 484, + 303, + 542, + 361 + ], + "lines": [ + { + "bbox": [ + 484, + 303, + 542, + 361 + ], + "spans": [ + { + "bbox": [ + 484, + 303, + 542, + 361 + ], + "type": "image", + "image_path": "20a50f60ecb91869ef762bcdba895c738cfb57265161360d0aa0325b80b8af99.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 299, + 369, + 425, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 369, + 425, + 380 + ], + "spans": [ + { + "bbox": [ + 299, + 369, + 425, + 380 + ], + "type": "text", + "content": "A dog cat is running on the grass." + } + ] + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 299, + 381, + 358, + 440 + ], + "blocks": [ + { + "bbox": [ + 299, + 381, + 358, + 440 + ], + "lines": [ + { + "bbox": [ + 299, + 381, + 358, + 440 + ], + "spans": [ + { + "bbox": [ + 299, + 381, + 358, + 440 + ], + "type": "image", + "image_path": "b233b6f264d011190749631b76dd0bf1102b113211cb9e0c06bf06d9def44609.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 361, + 381, + 419, + 440 + ], + "blocks": [ + { + "bbox": [ + 361, + 381, + 419, + 440 + ], + "lines": [ + { + "bbox": [ + 361, + 381, + 419, + 440 + ], + "spans": [ + { + "bbox": [ + 361, + 381, + 419, + 440 + ], + "type": "image", + "image_path": "cd46e9e85d1e6a625a30ac03e9200a5a17276eb2a22f02945dd2b94cea2e6c4d.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 422, + 381, + 481, + 440 + ], + "blocks": [ + { + "bbox": [ + 422, + 381, + 481, + 440 + ], + "lines": [ + { + "bbox": [ + 422, + 381, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 422, + 381, + 481, + 440 + ], + "type": "image", + "image_path": "9335c60824c9d469c767134cfab8445bc5a53d69061b3eab1a466937de72e4bf.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 484, + 381, + 542, + 440 + ], + "blocks": [ + { + "bbox": [ + 484, + 381, + 542, + 440 + ], + "lines": [ + { + "bbox": [ + 484, + 381, + 542, + 440 + ], + "spans": [ + { + "bbox": [ + 484, + 381, + 542, + 440 + ], + "type": "image", + "image_path": "a3fc67b288ac36e0c0fd21a3239eb2fd7f583636be608f0db9d081b5c5331cf5.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "bbox": [ + 299, + 442, + 437, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 442, + 437, + 453 + ], + "spans": [ + { + "bbox": [ + 299, + 442, + 437, + 453 + ], + "type": "text", + "content": "A fluffy " + }, + { + "bbox": [ + 299, + 442, + 437, + 453 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 299, + 442, + 437, + 453 + ], + "type": "text", + "content": " dog is running on the grass." + } + ] + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 299, + 455, + 358, + 513 + ], + "blocks": [ + { + "bbox": [ + 299, + 455, + 358, + 513 + ], + "lines": [ + { + "bbox": [ + 299, + 455, + 358, + 513 + ], + "spans": [ + { + "bbox": [ + 299, + 455, + 358, + 513 + ], + "type": "image", + "image_path": "55e82f37575debf34f5527138a68a5c576ad5b5a62d585e1b9c903f57c99cf42.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 361, + 455, + 419, + 513 + ], + "blocks": [ + { + "bbox": [ + 361, + 455, + 419, + 513 + ], + "lines": [ + { + "bbox": [ + 361, + 455, + 419, + 513 + ], + "spans": [ + { + "bbox": [ + 361, + 455, + 419, + 513 + ], + "type": "image", + "image_path": "50b820aba89d2cb15d3c9ab424c57af90e732f765732413cbb91dfdc4c711d86.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 422, + 455, + 481, + 513 + ], + "blocks": [ + { + "bbox": [ + 422, + 455, + 481, + 513 + ], + "lines": [ + { + "bbox": [ + 422, + 455, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 422, + 455, + 481, + 513 + ], + "type": "image", + "image_path": "2bfc7b2dc05fdec65b5915f4ee8adfa5096383e284de366eb8c01b9d8f5537f9.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 484, + 455, + 542, + 513 + ], + "blocks": [ + { + "bbox": [ + 484, + 455, + 542, + 513 + ], + "lines": [ + { + "bbox": [ + 484, + 455, + 542, + 513 + ], + "spans": [ + { + "bbox": [ + 484, + 455, + 542, + 513 + ], + "type": "image", + "image_path": "1a87f466a72fc87f915cf31f5784f6ff037e21b097b963c0b35eb5ab57676f52.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "bbox": [ + 46, + 544, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 662 + ], + "type": "text", + "content": "Word swap. Video-P2P enables the replacement of entities based on word swapping while maintaining the coherence of unrelated regions. As illustrated in Fig. 5, Video-P2P seamlessly replaces the man on the motorbike with Spider-Man while minimizing the changes to the motorbike's appearance (the 4th row). The generated Spider-Man exhibits a consistent appearance across frames, and the background remains unchanged. Furthermore, we can replace a dog with a cat while preserving its gesture and the surrounding grass (the 5th row)." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Prompt refinement. Video-P2P is able to do prompt refinement, such as modifying object properties. For example, we can transform the running dog into a robotic one (the" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": "6th row in Fig. 5), and convert a motorbike into a Lego toy with the same motion (the 3rd row). Notice the grass and sky are almost not influenced. Additionally, Video-P2P can perform global editing like changing the weather to sunset or flooding the road with water (2nd row). Style transfer can also be accomplished by Video-P2P, as exemplified by transforming the video into a watercolor painting." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Attention re-weighting. Similar to Image-P2P, Video-P2P also enables attention re-weighting. By adjusting the cross-attention of specific words, we can manipulate the extent of the corresponding generation. For instance, we can regulate how fluffy a dog is in the video (the 6th row of Fig. 5)." + } + ] + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8604" + } + ] + } + ], + "index": 65 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 81, + 106, + 135 + ], + "blocks": [ + { + "bbox": [ + 51, + 81, + 106, + 135 + ], + "lines": [ + { + "bbox": [ + 51, + 81, + 106, + 135 + ], + "spans": [ + { + "bbox": [ + 51, + 81, + 106, + 135 + ], + "type": "image", + "image_path": "73b34461955180fb977c1aa7f225be47abc7959ac0e8eb7c578e2cd5d4903ed4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 137, + 223, + 147 + ], + "lines": [ + { + "bbox": [ + 107, + 137, + 223, + 147 + ], + "spans": [ + { + "bbox": [ + 107, + 137, + 223, + 147 + ], + "type": "text", + "content": "[TAV+DDIM] King-Kong is roaring." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 112, + 81, + 167, + 135 + ], + "blocks": [ + { + "bbox": [ + 111, + 71, + 215, + 80 + ], + "lines": [ + { + "bbox": [ + 111, + 71, + 215, + 80 + ], + "spans": [ + { + "bbox": [ + 111, + 71, + 215, + 80 + ], + "type": "text", + "content": "[Input Video] A lion is roaring." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 112, + 81, + 167, + 135 + ], + "lines": [ + { + "bbox": [ + 112, + 81, + 167, + 135 + ], + "spans": [ + { + "bbox": [ + 112, + 81, + 167, + 135 + ], + "type": "image", + "image_path": "b19744c9cd86d509b8804e9ed804d8adf8ea2648fb369ee6163072d40b28b36b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 81, + 228, + 135 + ], + "blocks": [ + { + "bbox": [ + 173, + 81, + 228, + 135 + ], + "lines": [ + { + "bbox": [ + 173, + 81, + 228, + 135 + ], + "spans": [ + { + "bbox": [ + 173, + 81, + 228, + 135 + ], + "type": "image", + "image_path": "24f0601bbf8180941b1e4ad2ae01433a4ceb308127f0bbc99b57688c4a158df1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 81, + 290, + 135 + ], + "blocks": [ + { + "bbox": [ + 235, + 81, + 290, + 135 + ], + "lines": [ + { + "bbox": [ + 235, + 81, + 290, + 135 + ], + "spans": [ + { + "bbox": [ + 235, + 81, + 290, + 135 + ], + "type": "image", + "image_path": "9592dfa5fb777d4bd5f446ae6bcddd175b4d89596b561a56b5c05c00a275d4c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 148, + 106, + 202 + ], + "blocks": [ + { + "bbox": [ + 51, + 148, + 106, + 202 + ], + "lines": [ + { + "bbox": [ + 51, + 148, + 106, + 202 + ], + "spans": [ + { + "bbox": [ + 51, + 148, + 106, + 202 + ], + "type": "image", + "image_path": "17b3ef8917c72755dc3d0066b0e76618c27df64c6cdc2bd3e9cc3ff5981725ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 205, + 220, + 214 + ], + "lines": [ + { + "bbox": [ + 107, + 205, + 220, + 214 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 220, + 214 + ], + "type": "text", + "content": "[Video-P2P] King-Kong is roaring." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 112, + 148, + 168, + 202 + ], + "blocks": [ + { + "bbox": [ + 112, + 148, + 168, + 202 + ], + "lines": [ + { + "bbox": [ + 112, + 148, + 168, + 202 + ], + "spans": [ + { + "bbox": [ + 112, + 148, + 168, + 202 + ], + "type": "image", + "image_path": "bc64099e94ce3a956a5ceffab8cf4250ac4d4aeea2578648a0b5f2efb5981508.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 173, + 148, + 228, + 202 + ], + "blocks": [ + { + "bbox": [ + 173, + 148, + 228, + 202 + ], + "lines": [ + { + "bbox": [ + 173, + 148, + 228, + 202 + ], + "spans": [ + { + "bbox": [ + 173, + 148, + 228, + 202 + ], + "type": "image", + "image_path": "b9449b45a0a9567af94c930c382ad8bc1e45ed5b80523ce6a66bd9d50a5afb29.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 148, + 290, + 202 + ], + "blocks": [ + { + "bbox": [ + 235, + 148, + 290, + 202 + ], + "lines": [ + { + "bbox": [ + 235, + 148, + 290, + 202 + ], + "spans": [ + { + "bbox": [ + 235, + 148, + 290, + 202 + ], + "type": "image", + "image_path": "e842132d49e2d25a900fadeba48bb13caee16835dfbdd8ed0e72d8872a095325.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 216, + 106, + 271 + ], + "blocks": [ + { + "bbox": [ + 51, + 216, + 106, + 271 + ], + "lines": [ + { + "bbox": [ + 51, + 216, + 106, + 271 + ], + "spans": [ + { + "bbox": [ + 51, + 216, + 106, + 271 + ], + "type": "image", + "image_path": "f64dc1cc4265792ca7f709725c7ad3ec6b5e01264dfa19dcc6ecc15674996d4e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 280, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 545, + 302 + ], + "type": "text", + "content": "Figure 6. Video-P2P v.s. Tune-A-Video (TAV). Video-P2P offers the ability to edit content locally, while TAV+DDIM cannot avoid influencing unrelated regions." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 112, + 216, + 167, + 271 + ], + "blocks": [ + { + "bbox": [ + 112, + 216, + 167, + 271 + ], + "lines": [ + { + "bbox": [ + 112, + 216, + 167, + 271 + ], + "spans": [ + { + "bbox": [ + 112, + 216, + 167, + 271 + ], + "type": "image", + "image_path": "2f8337cc8ab4d0c97df70aa754f7de5e165fd7951bec03dc0ce3afc286ff01f0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 173, + 216, + 228, + 271 + ], + "blocks": [ + { + "bbox": [ + 173, + 216, + 228, + 271 + ], + "lines": [ + { + "bbox": [ + 173, + 216, + 228, + 271 + ], + "spans": [ + { + "bbox": [ + 173, + 216, + 228, + 271 + ], + "type": "image", + "image_path": "79760badc0cf109ed6153b54a52c7a38690752895dfededcfb9fbced70308b61.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 235, + 216, + 290, + 271 + ], + "blocks": [ + { + "bbox": [ + 235, + 216, + 290, + 271 + ], + "lines": [ + { + "bbox": [ + 235, + 216, + 290, + 271 + ], + "spans": [ + { + "bbox": [ + 235, + 216, + 290, + 271 + ], + "type": "image", + "image_path": "382ba663ac22251733534e7307ef727e9862e79811d40e33e42d802451bbb2a5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 304, + 81, + 359, + 137 + ], + "blocks": [ + { + "bbox": [ + 304, + 81, + 359, + 137 + ], + "lines": [ + { + "bbox": [ + 304, + 81, + 359, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 81, + 359, + 137 + ], + "type": "image", + "image_path": "f4522730b2ed31cc1cca85a7bcd0d6af825fd6b5a7136c7469050d4b93cb205f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 137, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 337, + 137, + 503, + 147 + ], + "spans": [ + { + "bbox": [ + 337, + 137, + 503, + 147 + ], + "type": "text", + "content": "[TAV+DDIM] A Lego tiger is walking in the forest." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 364, + 82, + 421, + 137 + ], + "blocks": [ + { + "bbox": [ + 342, + 71, + 494, + 80 + ], + "lines": [ + { + "bbox": [ + 342, + 71, + 494, + 80 + ], + "spans": [ + { + "bbox": [ + 342, + 71, + 494, + 80 + ], + "type": "text", + "content": "[Input Video] A tiger is walking in the forest." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 82, + 421, + 137 + ], + "lines": [ + { + "bbox": [ + 364, + 82, + 421, + 137 + ], + "spans": [ + { + "bbox": [ + 364, + 82, + 421, + 137 + ], + "type": "image", + "image_path": "d008b7946f8122401c3fe16308379bc92517a56241128c7c87e14fdf427cda99.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 426, + 82, + 481, + 137 + ], + "blocks": [ + { + "bbox": [ + 426, + 82, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 426, + 82, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 426, + 82, + 481, + 137 + ], + "type": "image", + "image_path": "50a6e8885e4ca64df6bfdf3c7e60ee0a7f68cb9d28a4922b46581a10844cd08a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 487, + 82, + 542, + 137 + ], + "blocks": [ + { + "bbox": [ + 487, + 82, + 542, + 137 + ], + "lines": [ + { + "bbox": [ + 487, + 82, + 542, + 137 + ], + "spans": [ + { + "bbox": [ + 487, + 82, + 542, + 137 + ], + "type": "image", + "image_path": "1b439aa2bc86b7ce4ad3f42daefa8b7916780eb85392e8e8904fe4840cd052c6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 304, + 148, + 359, + 203 + ], + "blocks": [ + { + "bbox": [ + 304, + 148, + 359, + 203 + ], + "lines": [ + { + "bbox": [ + 304, + 148, + 359, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 359, + 203 + ], + "type": "image", + "image_path": "7782192d9c05e64cda6725ce3d84c515aefc8eb806e14805cfdf274e0fec6943.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 364, + 148, + 421, + 203 + ], + "blocks": [ + { + "bbox": [ + 364, + 148, + 421, + 203 + ], + "lines": [ + { + "bbox": [ + 364, + 148, + 421, + 203 + ], + "spans": [ + { + "bbox": [ + 364, + 148, + 421, + 203 + ], + "type": "image", + "image_path": "190248bd0954fb26d448252f9922ad667205186b116c46487bcbbadf82f69eec.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 426, + 148, + 481, + 203 + ], + "blocks": [ + { + "bbox": [ + 426, + 148, + 481, + 203 + ], + "lines": [ + { + "bbox": [ + 426, + 148, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 426, + 148, + 481, + 203 + ], + "type": "image", + "image_path": "076c1115f5f76469ef84b5a6a83aa6514462306d8f8115ed7bf073e2de036b0b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 487, + 148, + 542, + 203 + ], + "blocks": [ + { + "bbox": [ + 487, + 148, + 542, + 203 + ], + "lines": [ + { + "bbox": [ + 487, + 148, + 542, + 203 + ], + "spans": [ + { + "bbox": [ + 487, + 148, + 542, + 203 + ], + "type": "image", + "image_path": "63ba69ad9298ab3044dbf95a65c11833f9b0dc8ca8bb8dba05487e7d65f87d9b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 304, + 216, + 359, + 271 + ], + "blocks": [ + { + "bbox": [ + 337, + 205, + 500, + 214 + ], + "lines": [ + { + "bbox": [ + 337, + 205, + 500, + 214 + ], + "spans": [ + { + "bbox": [ + 337, + 205, + 500, + 214 + ], + "type": "text", + "content": "[Video-P2P] A Lego tiger is walking in the forest." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 216, + 359, + 271 + ], + "lines": [ + { + "bbox": [ + 304, + 216, + 359, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 359, + 271 + ], + "type": "image", + "image_path": "51975775d897c93dd1b7e7c5812018a75922dc77ee0b5c7f2c4e1cc3592f9f7e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 365, + 216, + 421, + 271 + ], + "blocks": [ + { + "bbox": [ + 365, + 216, + 421, + 271 + ], + "lines": [ + { + "bbox": [ + 365, + 216, + 421, + 271 + ], + "spans": [ + { + "bbox": [ + 365, + 216, + 421, + 271 + ], + "type": "image", + "image_path": "fcc7a614e32dc413920c3ea1bc131a2ce1087be8ad7a8bcd4124822786c4d7d9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 426, + 216, + 481, + 271 + ], + "blocks": [ + { + "bbox": [ + 426, + 216, + 481, + 271 + ], + "lines": [ + { + "bbox": [ + 426, + 216, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 426, + 216, + 481, + 271 + ], + "type": "image", + "image_path": "82495e4a28e42d7b325a974dbc6033d8b6b18e5b11952aa99898b2d3cdf013a6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 487, + 216, + 542, + 271 + ], + "blocks": [ + { + "bbox": [ + 487, + 216, + 542, + 271 + ], + "lines": [ + { + "bbox": [ + 487, + 216, + 542, + 271 + ], + "spans": [ + { + "bbox": [ + 487, + 216, + 542, + 271 + ], + "type": "image", + "image_path": "c028d4891a5a3bffbd29516bc107010da27052ce5ad63c5fa3703e88f6e8e0e4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 54, + 310, + 284, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 310, + 284, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 284, + 321 + ], + "type": "text", + "content": "[Input Video] a brown dog and a white dog on the kitchen floor." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 50, + 323, + 108, + 361 + ], + "blocks": [ + { + "bbox": [ + 50, + 323, + 108, + 361 + ], + "lines": [ + { + "bbox": [ + 50, + 323, + 108, + 361 + ], + "spans": [ + { + "bbox": [ + 50, + 323, + 108, + 361 + ], + "type": "image", + "image_path": "213e52013709e7a91620e907ec629fc1de890b81d3f2450b163e981e977d7018.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 111, + 323, + 169, + 361 + ], + "blocks": [ + { + "bbox": [ + 111, + 323, + 169, + 361 + ], + "lines": [ + { + "bbox": [ + 111, + 323, + 169, + 361 + ], + "spans": [ + { + "bbox": [ + 111, + 323, + 169, + 361 + ], + "type": "image", + "image_path": "97c6802bd66efb6f675affed8023f0c8abea1a33c075ae1269a7d3756edf57b6.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 173, + 323, + 230, + 361 + ], + "blocks": [ + { + "bbox": [ + 173, + 323, + 230, + 361 + ], + "lines": [ + { + "bbox": [ + 173, + 323, + 230, + 361 + ], + "spans": [ + { + "bbox": [ + 173, + 323, + 230, + 361 + ], + "type": "image", + "image_path": "3c6700b3759eb09a9b0eab10d35c19b3558526d4054c79f0058559c7262bfb0b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 232, + 323, + 283, + 361 + ], + "blocks": [ + { + "bbox": [ + 232, + 323, + 283, + 361 + ], + "lines": [ + { + "bbox": [ + 232, + 323, + 283, + 361 + ], + "spans": [ + { + "bbox": [ + 232, + 323, + 283, + 361 + ], + "type": "image", + "image_path": "98339615b3b21c121f796ce67b78058e5e22772c89e3fa23d0ea9e9ca6aeac02.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 50, + 374, + 108, + 412 + ], + "blocks": [ + { + "bbox": [ + 63, + 364, + 279, + 373 + ], + "lines": [ + { + "bbox": [ + 63, + 364, + 279, + 373 + ], + "spans": [ + { + "bbox": [ + 63, + 364, + 279, + 373 + ], + "type": "text", + "content": "[Dreamix] a brown cat and a white cat on the kitchen floor." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 374, + 108, + 412 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 108, + 412 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 108, + 412 + ], + "type": "image", + "image_path": "9c4b3a02753a59567a9e5497b22d90aa46b44aff47bafbf2013245a8555f6990.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 111, + 374, + 169, + 412 + ], + "blocks": [ + { + "bbox": [ + 111, + 374, + 169, + 412 + ], + "lines": [ + { + "bbox": [ + 111, + 374, + 169, + 412 + ], + "spans": [ + { + "bbox": [ + 111, + 374, + 169, + 412 + ], + "type": "image", + "image_path": "a98e31f799ca3ce95990d8a6940b37f52628f827d62715eef13a368cf5feba25.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 415, + 285, + 424 + ], + "lines": [ + { + "bbox": [ + 61, + 415, + 285, + 424 + ], + "spans": [ + { + "bbox": [ + 61, + 415, + 285, + 424 + ], + "type": "text", + "content": "[Video-P2P] a brown cat and a white cat on the kitchen floor." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 173, + 374, + 230, + 412 + ], + "blocks": [ + { + "bbox": [ + 173, + 374, + 230, + 412 + ], + "lines": [ + { + "bbox": [ + 173, + 374, + 230, + 412 + ], + "spans": [ + { + "bbox": [ + 173, + 374, + 230, + 412 + ], + "type": "image", + "image_path": "2dcbfc1b2fb50ba8069a0b4262ed1ee097ffd8e94607f5dc1df383827dc77897.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 232, + 374, + 283, + 412 + ], + "blocks": [ + { + "bbox": [ + 232, + 374, + 283, + 412 + ], + "lines": [ + { + "bbox": [ + 232, + 374, + 283, + 412 + ], + "spans": [ + { + "bbox": [ + 232, + 374, + 283, + 412 + ], + "type": "image", + "image_path": "7931fb2f1c4e456ffac76f676cc98d4711df3fa0425a83b9cb07992e0a4de6c0.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 50, + 426, + 108, + 463 + ], + "blocks": [ + { + "bbox": [ + 50, + 426, + 108, + 463 + ], + "lines": [ + { + "bbox": [ + 50, + 426, + 108, + 463 + ], + "spans": [ + { + "bbox": [ + 50, + 426, + 108, + 463 + ], + "type": "image", + "image_path": "8037d67c72d3f815b6e0c47c39f81da7a9f740cfb779c84dd6fe840d990fc5ff.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 470, + 287, + 493 + ], + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 493 + ], + "type": "text", + "content": "Figure 7. Video-P2P v.s. Dreamix. Both methods can change the dogs to cats. Video-P2P can preserve background's details." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 111, + 426, + 168, + 463 + ], + "blocks": [ + { + "bbox": [ + 111, + 426, + 168, + 463 + ], + "lines": [ + { + "bbox": [ + 111, + 426, + 168, + 463 + ], + "spans": [ + { + "bbox": [ + 111, + 426, + 168, + 463 + ], + "type": "image", + "image_path": "4330f52597e182425ee890153dcf8d0ff448678efe91c3fa302919007b1206a5.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 173, + 426, + 230, + 463 + ], + "blocks": [ + { + "bbox": [ + 173, + 426, + 230, + 463 + ], + "lines": [ + { + "bbox": [ + 173, + 426, + 230, + 463 + ], + "spans": [ + { + "bbox": [ + 173, + 426, + 230, + 463 + ], + "type": "image", + "image_path": "b9d15f53ba02e1af9d095394ce0ee360f706b54a8eb2716da5d8715728727356.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 232, + 426, + 283, + 463 + ], + "blocks": [ + { + "bbox": [ + 232, + 426, + 283, + 463 + ], + "lines": [ + { + "bbox": [ + 232, + 426, + 283, + 463 + ], + "spans": [ + { + "bbox": [ + 232, + 426, + 283, + 463 + ], + "type": "image", + "image_path": "d1bbb4b0354897c5b69534c203038cfdbcfb06fd2db7f60c2846cd22a32ede99.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 47, + 495, + 128, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 128, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 128, + 507 + ], + "type": "text", + "content": "4.3. Comparison" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 519, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 626 + ], + "type": "text", + "content": "Comparison with Tune-A-Video. Both TAV+DDIM [47] and our Video-P2P allow for video editing with text prompts. However, TAV+DDIM cannot avoid altering the entire video content when editing specific objects, while Video-P2P can edit a local area and minimize the influence on other regions. Fig. 6 (Left) demonstrates that Video-P2P preserves the complex shape of the cloud when replacing a lion with King Kong, whereas TAV+DDIM can only maintain the color tone of the sky in this case." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": "Although our model initialization is similar to TAV, Video-P2P can still generate temporal-consistent results where TAV+DDIM fails. As demonstrated in Fig. 6 (Right), TAV struggles to generate a temporally consistent sequence in the second row, even when the inputs are features from DDIM inversion. In contrast, our method can produce better structure-preserved results, as shown in the third row." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 312, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 479 + ], + "type": "text", + "content": "Comparison with Dreamix. In contrast to Dreamix [25], which uses a pre-trained video diffusion model that is not publicly available, our method yields superior results for subject replacement. Although our method cannot perform video motion editing due to the lack of temporal priors, we outperform Dreamix in preserving details and motion consistency. As it is not open-sourced, we conducted our evaluation on its released demo. As demonstrated in Fig. 7, both methods can transform two dogs into two cats, but our method preserves the details of the drawer in the background (the 3rd row). Furthermore, Dreamix may affect the time sequence to some extent, as the generated cat moves more slowly than the original dog in the video. Our method completely preserves the motion of the original video." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Quantitative results. We create a test set of 45 samples based on the DAVIS [29] dataset and report five metrics for quantitative analysis. The CLIP Score measures the textual similarity between the text prompt and video, while Masked PSNR and LPIPS [51] evaluate the quality of structure preservation. We also proposed a novel metric, Object Semantic Variance (OSV), to measure semantic consistency across frames. In addition, we report an average CLIP similarity (Temp) between consecutive frames (as in Gen-1 [8] and FateZero [30]) to evaluate the temporal coherence. For detailed explanations of these metrics, please refer to the appendix. Our results, as shown in Table 1, demonstrate that Video-P2P performs well on all metrics. Compared to TAV+DDIM, Video-P2P achieves higher Masked PSNR and lower LPIPS, indicating better preservation of unchanged regions. Compared to the other two methods, Video-P2P has much lower OSV and Temp, indicating its superior ability to maintain semantic and temporal consistency across frames. When comparing to Image-P2P, we" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8605" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 80, + 108, + 125 + ], + "blocks": [ + { + "bbox": [ + 50, + 80, + 108, + 125 + ], + "lines": [ + { + "bbox": [ + 50, + 80, + 108, + 125 + ], + "spans": [ + { + "bbox": [ + 50, + 80, + 108, + 125 + ], + "type": "image", + "image_path": "63f13cb8eb4b0de5cfb2f6ab703cd4770241a7f5f65ece74054a0e6cd63a6550.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 109, + 80, + 167, + 125 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 223, + 79 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 223, + 79 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 223, + 79 + ], + "type": "text", + "content": "A dog-cat is running on the grass." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 80, + 167, + 125 + ], + "lines": [ + { + "bbox": [ + 109, + 80, + 167, + 125 + ], + "spans": [ + { + "bbox": [ + 109, + 80, + 167, + 125 + ], + "type": "image", + "image_path": "20ae2dd82cce5c0158053af32ffd6702a0b78bc6c79217896315e5c560a6f114.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 80, + 226, + 125 + ], + "blocks": [ + { + "bbox": [ + 168, + 80, + 226, + 125 + ], + "lines": [ + { + "bbox": [ + 168, + 80, + 226, + 125 + ], + "spans": [ + { + "bbox": [ + 168, + 80, + 226, + 125 + ], + "type": "image", + "image_path": "81e707be5ed6e1e121f94b9035ca81f74932db19e29fa59caeec1f258f7897aa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 228, + 80, + 285, + 125 + ], + "blocks": [ + { + "bbox": [ + 228, + 80, + 285, + 125 + ], + "lines": [ + { + "bbox": [ + 228, + 80, + 285, + 125 + ], + "spans": [ + { + "bbox": [ + 228, + 80, + 285, + 125 + ], + "type": "image", + "image_path": "5ef134070c02eab4e058de8e5c502a80675853cdd2ada111a0733108898f91b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 126, + 108, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 126, + 108, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 126, + 108, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 126, + 108, + 171 + ], + "type": "image", + "image_path": "eb35601c5af7e22c08ba850afdfc7c1f1fc59e585782830772771c739f8853cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 171, + 88, + 178 + ], + "lines": [ + { + "bbox": [ + 71, + 171, + 88, + 178 + ], + "spans": [ + { + "bbox": [ + 71, + 171, + 88, + 178 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 109, + 126, + 167, + 171 + ], + "blocks": [ + { + "bbox": [ + 109, + 126, + 167, + 171 + ], + "lines": [ + { + "bbox": [ + 109, + 126, + 167, + 171 + ], + "spans": [ + { + "bbox": [ + 109, + 126, + 167, + 171 + ], + "type": "image", + "image_path": "6f3658fa5fe44419984fafeb680a0cd1ba9307eb888761eb85d2d3a91962b663.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 171, + 154, + 178 + ], + "lines": [ + { + "bbox": [ + 123, + 171, + 154, + 178 + ], + "spans": [ + { + "bbox": [ + 123, + 171, + 154, + 178 + ], + "type": "text", + "content": "Image-P2P" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 126, + 226, + 171 + ], + "blocks": [ + { + "bbox": [ + 168, + 126, + 226, + 171 + ], + "lines": [ + { + "bbox": [ + 168, + 126, + 226, + 171 + ], + "spans": [ + { + "bbox": [ + 168, + 126, + 226, + 171 + ], + "type": "image", + "image_path": "f9a5d8d30685c62a97f0cba66c3529862cf207800c4236aaf61faad0ad46c5d6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 171, + 217, + 178 + ], + "lines": [ + { + "bbox": [ + 176, + 171, + 217, + 178 + ], + "spans": [ + { + "bbox": [ + 176, + 171, + 217, + 178 + ], + "type": "text", + "content": "w/o model init" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 126, + 285, + 171 + ], + "blocks": [ + { + "bbox": [ + 228, + 126, + 285, + 171 + ], + "lines": [ + { + "bbox": [ + 228, + 126, + 285, + 171 + ], + "spans": [ + { + "bbox": [ + 228, + 126, + 285, + 171 + ], + "type": "image", + "image_path": "41d19275b1d888db267b3c33fe9650b2e453093a5946327f7759b8d6b5658b26.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 171, + 276, + 178 + ], + "lines": [ + { + "bbox": [ + 238, + 171, + 276, + 178 + ], + "spans": [ + { + "bbox": [ + 238, + 171, + 276, + 178 + ], + "type": "text", + "content": "w/ model init" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 48, + 202, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 53, + 185, + 280, + 196 + ], + "lines": [ + { + "bbox": [ + 53, + 185, + 280, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 280, + 196 + ], + "type": "text", + "content": "Figure 8. Model initialization improves the generation quality." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 202, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 48, + 202, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 202, + 286, + 236 + ], + "type": "image", + "image_path": "e0b298690c88cbbb66acaf1b791dd345bd7a4dedf09a306fc38ce495e12a6e76.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 242, + 269, + 253 + ], + "lines": [ + { + "bbox": [ + 64, + 242, + 269, + 253 + ], + "spans": [ + { + "bbox": [ + 64, + 242, + 269, + 253 + ], + "type": "text", + "content": "Figure 9. Ablation on shared unconditional embeddings." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 260, + 286, + 301 + ], + "blocks": [ + { + "bbox": [ + 50, + 260, + 286, + 301 + ], + "lines": [ + { + "bbox": [ + 50, + 260, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 50, + 260, + 286, + 301 + ], + "type": "image", + "image_path": "e45c31d65878decc4ae696d1408b4d427a2874b76f1d56fb22a641a0131d344e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 278, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 278, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 278, + 319 + ], + "type": "text", + "content": "Figure 10. Ablation on decoupled-guidance attention control." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 323, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 418 + ], + "type": "text", + "content": "consider it as a reference point rather than a direct competitor because it does not account for the consistency between different frames. Image-P2P represents the upper bound of PSNR and LPIPS, which serve as metrics to evaluate per-frame quality. Among all video editing methods, Video-P2P is the best one. Moreover, in Tab. 3, we report the user study results, where Video-P2P ranks first on average and has a high preference rate compared to other methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 429, + 140, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 140, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 140, + 441 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 445, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 553 + ], + "type": "text", + "content": "Model initialization. While the inflated image diffusion model can generate semantically consistent images, the T2S model's generation ability is compromised during inflation. As seen in Fig.8 (the 3rd column), directly using the inflated T2S model produces unrealistic results with an inaccurate background. To mitigate this, we initialize the T2S model by fine-tuning the given video. This is evident in Fig.8 (4th column), where the cat's appearance improves, and the grass reconstruction becomes more accurate." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "content": "Shared unconditional embedding. The use of shared unconditional embeddings is crucial to ensure stable and consistent generations. On our test set, we found that " + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "content": " of the videos exhibit instability when multiple unconditional embeddings are employed. A visual example is included in Fig. 9 (Right). As shown in Tab. 1 (row 5), semantic and temporal consistency are diminished without shared unconditional embedding, reminiscent of the content code in [39]. We observe that optimizing a shared unconditional embedding can significantly improve the PSNR compared to TAV+DDIM in Tab. 2. However, using multiple unconditional embeddings for each frame only increases the PSNR by 0.2 but results in higher parameter usage (" + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "content": " times)" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 311, + 70, + 541, + 135 + ], + "blocks": [ + { + "bbox": [ + 311, + 70, + 541, + 135 + ], + "lines": [ + { + "bbox": [ + 311, + 70, + 541, + 135 + ], + "spans": [ + { + "bbox": [ + 311, + 70, + 541, + 135 + ], + "type": "table", + "html": "
CLIP ↑M.PSNR ↑LPIPS ↓OSV ↓Temp ↑
TAV+DDIM0.334317.370.465155.670.9697
Image-P2P0.327222.920.308276.920.9021
Ours (w/o MI)0.321320.620.325550.270.9460
Ours (w/o DG)0.322618.970.386668.750.9451
Ours (w/o SU)0.335120.640.322148.780.9665
Ours0.336720.650.321347.380.9725
", + "image_path": "898483488d5eae74cf60aa996f6b4f96a1069b9ede66ba057d98b1440192b093.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 307, + 186, + 547, + 229 + ], + "blocks": [ + { + "bbox": [ + 305, + 140, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 183 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation. We evaluate textual similarity (CLIP), region preservation (Masked PSNR, LPIPS), semantic consistency (OSV), and temporal consistency (Temp). DG and SU mean Decoupled-Guidance and shared unconditional embeddings." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "type": "table", + "html": "
VQVAETAV +DDIMMulti- uncondShared- uncond
PSNR(dB) ↑24.7315.4322.9722.75
#Param. ↓/0.13M22.68M2.94M
", + "image_path": "82479551bb7bfc56151fb901b621786b40680cf2ccfee854f74f79ea1e526ec2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 307, + 257, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 306, + 234, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 234, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 545, + 255 + ], + "type": "text", + "content": "Table 2. Video reconstruction quality. A shared unconditional embedding can reconstruct a high-quality video with a small size." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 257, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 334 + ], + "type": "table", + "html": "
Image-P2PTAVTAV+DDIMVideo-P2P
Structure2.673.332.611.39
Preserving13.59%6.52%10.87%69.02%
Text3.402.782.281.54
Alignment3.80%14.13%19.57%62.50%
Realism & Quality3.382.982.211.43
4.35%7.61%19.02%69.02%
", + "image_path": "c42e757a53e7ee892c2b241e0490cf47f1c21c5065a2c57848d9e19e8d72ecce.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "type": "text", + "content": "Table 3. User study of average ranking " + }, + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "type": "text", + "content": " and preference rate " + }, + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 338, + 545, + 388 + ], + "type": "text", + "content": ". and leads to a lower Masked PSNR after attention control. Thus, we conclude that shared unconditional embedding is the most effective and efficient method for video inversion." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 392, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 547 + ], + "type": "text", + "content": "Decoupled-guidance attention control. To obtain the latent features of the input video, we optimize an unconditional embedding using the source prompt. It is important to note that this embedding is only suitable for the source prompt during the prompt-to-prompt process. Using the optimized embedding for the target prompt may negatively impact the quality of the generated results, as shown in Fig. 10 (Left). Instead, we utilize the initialized unconditional embedding for the target prompt and incorporate attention maps from two branches. The decoupled-guidance attention control approach significantly improves the editing quality, as shown in Fig.10 (Right). Quantitative ablations can be found in Tab. 1 (the 4th row and 6th row)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 558, + 378, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 378, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 378, + 570 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "type": "text", + "content": "Video-P2P provides the first framework for video editing with cross-attention control. We optimize a shared unconditional embedding based on a well-initialized T2S model for video inversion. We also propose the decoupled-guidance strategy for attention control. These techniques enable Video-P2P to perform various applications, such as word swap, prompt refinement, and attention re-weighting." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement Supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8606" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 167 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723, 2022. 3, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 286, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 257 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 286, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 302 + ], + "type": "text", + "content": "[5] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 286, + 345 + ], + "type": "text", + "content": "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 388 + ], + "type": "text", + "content": "[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. arXiv preprint arXiv:2305.04441, 2023. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "text", + "content": "[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "text", + "content": "[9] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 286, + 545 + ], + "type": "text", + "content": "[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 601 + ], + "type": "text", + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 646 + ], + "type": "text", + "content": "[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), pages 1-13, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[16] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "type": "text", + "content": "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "text", + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "text", + "content": "[21] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[22] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 475 + ], + "type": "text", + "content": "[23] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "text", + "content": "[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "type": "text", + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[28] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF Inter" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8607" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "national Conference on Computer Vision, pages 2085-2094, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[29] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[30] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 248 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 336 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "text", + "content": "[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[39] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[42] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 564 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[43] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[44] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[45] Wen Wang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "text", + "content": "[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "type": "text", + "content": "[47] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[48] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In ECCV, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 397, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 545, + 430 + ], + "type": "text", + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 432, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 486 + ], + "type": "text", + "content": "[50] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 487, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 520 + ], + "type": "text", + "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 521, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 564 + ], + "type": "text", + "content": "[52] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8608" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_content_list.json b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a423b9cb4447ef135aa4e11f8e82ac53bcf80062 --- /dev/null +++ b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "Video2Game: Real-time, Interactive, Realistic and Browser-Compatible Environment from a Single Video", + "text_level": 1, + "bbox": [ + 124, + 128, + 846, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongchi Xia $^{1,2}$ Zhi-Hao Lin $^{1}$ Wei-Chiu Ma $^{3}$ Shenlong Wang $^{1}$", + "bbox": [ + 209, + 202, + 756, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Illinois Urbana-Champaign 2Shanghai Jiao Tong University 3Cornell University https://video2game.github.io/", + "bbox": [ + 94, + 220, + 875, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/019048486c7a00546ae2e401585e3400601abaa9c6a9f98151db88abca89aa68.jpg", + "image_caption": [ + "Figure 1. Video2Game takes an input video of an arbitrary scene and automatically transforms it into a real-time, interactive, realistic and browser-compatible environment. The users can freely explore the environment and interact with the objects in the scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 270, + 300, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e707acea833d3e22c0f7c79611dab4950f098c1a44aef591b207eb8f9fb8fd92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 271, + 517, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/feef627377f78d931bd4a81eac2678d5756bebec9b1db5e28b7fe1f3c0f881cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 271, + 705, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48873a54d0a1a72ec773b77032124140039daa7f62a7ca19aa2f816d9ed228f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 271, + 890, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 539, + 313, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating high-quality and interactive virtual environments, such as games and simulators, often involves complex and costly manual modeling processes. In this paper, we present Video2Game, a novel approach that automatically converts videos of real-world scenes into realistic and interactive game environments. At the heart of our system are three core components: (i) a neural radiance fields (NeRF) module that effectively captures the geometry and visual appearance of the scene; (ii) a mesh module that distills the knowledge from NeRF for faster rendering; and (iii) a physics module that models the interactions and physical dynamics among the objects. By following the carefully designed pipeline, one can construct an interactable and actionable digital replica of the real world. We benchmark our system on both indoor and large-scale outdoor scenes. We show that we can not only produce highly-realistic renderings in real-time, but also build interactive games on top.", + "bbox": [ + 75, + 570, + 472, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 845, + 207, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crafting a visually compelling and interactive environment is crucial for immersive experiences in various domains,", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as video games, virtual reality applications, and self-driving simulators. This process, however, is complex and expensive. It demands the skills of experts in the field and the use of professional software development tools [21, 24]. For instance, Grand Theft Auto V [23], known for its intricately detailed environment, was one of the most expensive video games ever developed, with a budget over $265 million primarily for asset creation. Similarly, the development of the CARLA autonomous driving simulator [19] involves a multidisciplinary team of 3D artists, programmers, and engineers to meticulously craft and texture the virtual cityscapes, creating its lifelike environments.", + "bbox": [ + 496, + 540, + 890, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An appealing alternative to extensive manual modelling is creating environments directly from the real world. For instance, photogrammetry, a technique for constructing digital replicas of objects or scenes from overlapping real-world photographs, has been utilized for environment creation [52, 53]. Success stories also span various games and simulators. However, most use cases are limited to creating object assets and necessitate significant post-processing, such as material creation, texturing, and geometry fixes [66]. People thus turn to neural radiance fields (NeRFs) [46], as it offers a more promising approach to", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4578", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dace2d598dece63e3219702cf837ef9d2f80e770d781e077a3512b9365bfb430.jpg", + "image_caption": [ + "Figure 2. Overview of Video2Game: Given multiple posed images from a single video as input, we first construct a large-scale NeRF model that is realistic and possesses high-quality surface geometry. We then transform this NeRF model into a mesh representation with corresponding rigid-body dynamics to enable interactions. We utilize UV-mapped neural texture, which is both expressive and compatible with game engines. Finally, we obtain an interactive virtual environment that virtual actors can interact with, can respond to user control, and deliver high-resolution rendering from novel camera perspectives – all in real-time." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modeling large scenes. With careful design [13, 22, 39, 48, 61], NeRF is able to render free-viewpoint, photo-realistic images efficiently. However, crafting an interactive environment entails more than just creating a visually high-fidelity digital twin; it also involves building a physically plausible, immersive, real-time and importantly, interactive world tailored to user experiences. Furthermore, we expect such a virtual world to be compatible with real-time interaction interfaces such as common game engines. Despite its promise, the use of NeRF to create interactive environments from real-world videos remains largely unexplored.", + "bbox": [ + 75, + 376, + 472, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce Video2Game, a novel approach to automatically converting a video of a scene into a realistic and interactive virtual environment. Given a video as input, we first construct a NeRF that can effectively capture the geometric and visual information of a large-scale, unbounded) scene. Then we distill the NeRF into a game engine-compatible, neural textured mesh. This significantly improves the rendering efficiency while maintains the overall quality. To model the interactions among the objects, we further decompose the scene into individual actionable entities and equip them with respective physics model. Finally, we import our automatically generated assets into a WebGL-based game engine and create a playable game. The resulting virtual environment is photo-realistic, interactive, and runs in real-time. See Fig. 1 for demonstration. In summary, our key contributions are:", + "bbox": [ + 75, + 549, + 468, + 790 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel 3D modeling algorithm for real-time, free-viewpoint rendering and physical simulation, surpassing state-of-the-art NeRF baking methods with added rigid-body physics for enhanced simulation.", + "- An automated game-creation framework to transform a scene video into an interactive, realistic environment, compatible with current game engines." + ], + "bbox": [ + 76, + 792, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 375, + 650, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a single video, we aim to create a real-time, interactive game where the agents (e.g., the character, the car) can navigate and explore the reconstructed digital world, interact with objects in the scene (e.g., collision and manipulate objects), and achieve their respective tasks (e.g., collecting coins, shooting targets). We draw inspirations from multiple areas and combine the best of all. In this section, we will briefly review those closely related areas which forms the foundation of our work.", + "bbox": [ + 496, + 400, + 890, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS): Our work builds upon the success of novel view synthesis [14, 25, 35, 62], which is crucial for our game since it enables the agents to move freely and view the reconstructed world seamlessly from various perspectives. Among all these approaches [26, 60, 68, 85, 86], we exploit neural radiance field (NeRF) [46] as our underlying representation. NeRF has emerged as one of the most promising tools in NVS since its introduction [49-51], and has great performance across a wide range of scenarios [36, 56, 75, 81]. For instance, it can be easily extended to handle various challenging real-world scenarios such as learning from noisy camera poses [38, 70], reflectance modeling for photo-realistic relighting [69, 83], and real-time rendering [16, 39, 55, 65, 76]. In this work, we combine recent advances in NeRF with physics modeling to build an immersive digital world from one single video, moving from passive NVS to our complete solution for embodied world modeling where agents can actively explore and interact with the scene.", + "bbox": [ + 496, + 537, + 892, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable video generation: Using different control signals to manipulate the output of a visual model has garnered great interest in the community. This has had a profound impact on content creation [57, 58], digital editing [11, 34], and simulation [30, 31, 40]. One could also lever", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4579", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age large foundation models to control video content using text [57, 58]. However, they lack fine-grained and real-time control over the generated content. Alternatively, training (conditional) generative models for each scene enables better disentanglement of dynamics (e.g., foreground vs. background) and supports better control. For instance, one can represent a self-driving scene [31] or a Pacman game [30] as latent codes and generate video frames based on control inputs with a neural network. One can also learn to control the players within tennis games [44, 45, 79, 80]. Our work falls under the second line of research, where the model takes user control signals (e.g., a keystroke from the keyboard) as input and responds by rendering a new scene. However, instead of focusing on a specific scene (e.g., tennis games), we have developed a pipeline that allows the creation of a playable environment from a single video of a generic scene. Additionally, we model everything in 3D, which enables us to effectively capture not only view-dependent appearance but also physical interactions among rigid-body equipped objects. Importantly, we adopt a neural representation compatible with graphics engines, enabling users to play the entire game in their browser at an interactive rate.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-driven simulation: Building a realistic simulation environment has been a longstanding challenge. [19, 29, 67, 71]. While it's promising, we come close to mirror the real world only in recent years [10, 15, 42, 43, 59, 74, 75]. The key insight of these work is to build models by leveraging real-world data. Our work closely aligns with this line of research on building high-fidelity simulators from real-world data, with a few key differences. First, existing works mainly focus on offline training and evaluation [10, 43, 74, 75], whereas our system runs at an interactive rate and allows for online, real-time control. Second, some existing works [41, 43, 72, 87] need additional data modality like LiDAR point clouds for geometry reconstruction, but RGB video is all we need. Third, most photo-realistic simulators don't model physical interactions. However, we supports various physics modeling and allows agents to interact with the environment. Last, existing simulators are typically resource-intensive, while our system is lightweight and can be easily accessible in common engines.", + "bbox": [ + 76, + 422, + 472, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Video2Game", + "text_level": 1, + "bbox": [ + 76, + 723, + 210, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of images or a video of a scene, our goal is to construct an interactable and actionable digital twin, upon which we can build real-time, interactive games or realistic (sensor) simulators. Based on the observations that prevalent approaches to constructing digital replica mainly focus on visual appearance and ignore the underlying physical interactions, we carefully design our system such that it can not only produce high-quality rendering across viewpoints, but also support the modeling of physical actions (e.g., navigation, collision, manipulation, etc). At the heart", + "bbox": [ + 76, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8c2907449af9e07687461d6937d7ce4dfbf40441df49b16200fa95be7f69e0ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 696, + 167 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c116fa750e6d1bde87203919a95252062dd40cec5b5620b283a6dd2278bfaeac.jpg", + "image_caption": [ + "Figure 3. Visualization of automatically computed collision geometry: Sphere collider (green), box collider (yellow), convex polygon collider (purple) and trimesh collider (red)." + ], + "image_footnote": [], + "bbox": [ + 506, + 169, + 696, + 247 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3bb2df923f3aba42d4f67fcad36f98457137003d41783f6e7a092d100a42cba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 885, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of our systems is a compositional implicit-explicit 3D representation that is effective and efficient for both sensor and physics simulation. By decomposing the world into individual entities, we can better model and manipulate their physical properties (e.g., specularity, mass, friction), and simulate the outcomes of interactions more effectively.", + "bbox": [ + 496, + 335, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by introducing a NeRF model that can effectively capture the geometric and visual information of a large-scale, unbounded scene (Sec. 3.1). Next, we present an approach to convert the NeRF into a game-engine compatible mesh with neural texture maps, significantly improving the rendering efficiency while maintaining the quality (Sec. 3.2). To enable physical interactions, we further decompose the scene into individual actionable entities and equip them with respective physics models (Sec. 3.3). Finally, we describe how we integrate our interactive environment into a WebGL-based game engine, allowing users to play and interact with the virtual world in real time on their personal browser. Fig. 2 provides an overview of our proposed framework.", + "bbox": [ + 496, + 426, + 892, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Large-scale NeRF", + "text_level": 1, + "bbox": [ + 500, + 648, + 673, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries: Instant-NGP [48] is a notable variant of NeRF, which represents the radiance field with a combination of spatial hash-based voxels and neural networks: $\\mathbf{c},\\sigma = F_{\\theta}(\\mathbf{x},\\mathbf{d};\\Phi) = \\mathrm{MLP}_{\\theta}(\\mathsf{It}(\\mathbf{x},\\Phi),\\mathbf{d})$ . Given a 3D point $\\mathbf{x}\\in \\mathbb{R}^3$ and a camera direction $\\mathbf{d}\\in \\mathbb{R}^2$ as input, Instant-NGP first interpolate the point feature $\\mathsf{It}(\\mathbf{x},\\Phi)$ from the adjacent voxel features $\\Phi$ . Then the point feature and the camera direction are fed into a light-weight multilayer perception (MLP) to predict the color $\\mathbf{c}\\in \\mathbb{R}^3$ and density $\\sigma \\in \\mathbb{R}^{+}$ . To render the scene appearance, we first cast a ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ from the camera center o through the pixel center in direction d, and sample a set of 3D points $\\{\\mathbf{x}_i\\}$ along the ray. We then query their respective color $\\{\\mathbf{c}_i\\}$ and density $\\{\\sigma_i\\}$ and obtain the color of the pixel through alpha-composition: $\\mathbf{C}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i}w_{i}\\mathbf{c}_{i}$ . Similarly, the", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4580", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c620637932444c9baccae73c88aa25a7245e2059f01a47cdad11e61f6f030b3.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons among NeRF models. The rendering quality of our base NeRF is superior to baselines, and with leveraging monocular cues, we substantially improve rendered geometry compared to other baselines. This significantly facilitates NeRF baking in subsequent stages. Here we consider depths measured by LiDAR point cloud in KITTI-360 and compute normals based on it." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 885, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "expected depth can be computed by: $\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i}w_{i}t_{i}$ Here, $w_{i}$ indicates the blending weight that is derived from the densities $\\{\\sigma_i\\}$ . We refer the readers to [46] for more details. To learn the voxel features $\\Phi$ and the MLP weights $\\theta$ , we compute the difference between the ground truth color and the rendered color: $\\mathcal{L}_{\\mathrm{rgb}} = \\sum_{\\mathbf{r}}\\| \\mathbf{C}_{\\mathrm{GT}}(\\mathbf{r}) - \\mathbf{C}_{\\mathrm{NeRF}}(\\mathbf{r})\\| _2^2$ . Large-scale NeRF: While Instant-NGP [48] has shown promising results on densely observed and bounded scenes, its performance starts to degrade when extending to sparsely-captured, large-scale, unbounded environments. To mitigate these issues, we propose several enhancements:", + "bbox": [ + 75, + 318, + 468, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c}, \\sigma , s, \\mathbf {n} = F _ {\\theta} (\\mathbf {x}, \\mathbf {d}; \\Phi) = \\operatorname {M L P} _ {\\theta} (\\operatorname {I t} (\\operatorname {C t} (\\mathbf {x}), \\Phi), \\mathbf {d}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 501, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First of all, we exploit the contraction function $\\mathrm{Ct}(\\mathbf{x})$ [12] to map the unbounded coordinates into a bounded region. In addition to radiance and density, we predict the semantics $s$ and the surface normal $\\mathbf{n}$ of the 3D points, guided with 2D priors to better regularize the scene geometry. Furthermore, we divide large-scale scenes into several blocks [63] to capture the fine-grained details. We now describe these enhancements in more details.", + "bbox": [ + 75, + 529, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth: High-quality geometry is critical for modeling physical interactions. Inspired by MonoSDF [78], we leverage off-the-shelf monocular depth estimators [20, 27] to guide and improve the underlying geometry. We first predict the depth of the scene from rendered RGB images. Then we minimize the discrepancy between the rendered depth and the predicted depth via $\\mathcal{L}_{\\mathrm{depth}} = \\sum_{\\mathbf{r}}\\|(a\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}) + b) - \\mathbf{D}_{\\mathrm{mono}}(\\mathbf{r})\\|_2^2$ , where $a$ and $b$ are the scale and shift that aligns the two distribution [54].", + "bbox": [ + 75, + 650, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface normals: Similar to depth, we encourage the normal estimated from NeRF to be consistent with the normal predicted by the off-the-shelf estimator [20, 27]. The normal of a 3D point $\\mathbf{x}_i$ can be either analytically derived from the estimated density $\\mathbf{n}_i = (1 - \\frac{\\nabla_{\\mathbf{x}}\\sigma_i}{\\|\\nabla\\sigma_i\\|})$ , or predicted by the MLP header as in Eq. 1. We could aggregate them via volume render: $\\mathbf{N}(\\mathbf{r}) = \\sum_{i}\\mathbf{w}_{i}\\mathbf{n}_{i}$ . Em", + "bbox": [ + 75, + 792, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pirically we find that adopting both normals and promoting their mutual consistency works the best, since the MLP header offers more flexibility. We thus employ $\\mathcal{L}_{\\mathrm{normal}} = \\|\\mathbf{N}_{\\mathrm{mlp}}(\\mathbf{r}) - \\mathbf{N}_{\\mathrm{mono}}(\\mathbf{r})\\|_2^2 + \\|\\mathbf{N}_{\\mathrm{mlp}}(\\mathbf{r}) - \\mathbf{N}_{\\mathrm{density}}(\\mathbf{r})\\|_2^2$ .", + "bbox": [ + 498, + 319, + 890, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics: We also predict semantic logits for each sampled 3D points with our MLP. This helps us capture the correlation between semantics and geometry [36, 84]. We render the semantic map with volume rendering $\\mathbf{S}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i} \\mathbf{w}_{i} \\mathbf{s}_{i}$ and compute the cross-entropy with that of a 2D segmentation model $\\mathcal{L}_{\\text{semantics}} = \\mathrm{CE}(\\mathbf{S}_{\\text{mono}}, \\mathbf{S}_{\\text{NeRF}})$ .", + "bbox": [ + 498, + 385, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularization: We additionally adopt two regularization terms. To reduce floaters in the scene, for each randomly sampled 3D point $\\mathbf{x}$ , we penalize its density by $\\mathcal{L}_{\\mathrm{sp}} = \\sum 1 - \\exp (-\\alpha \\sigma (\\mathbf{x}))$ , where $\\alpha >0$ [77]. For each sky pixel (which we derived from the semantic MLP), we encourage its depth $\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}^{\\mathrm{sky}})$ to be as far as possible. The loss is defined as: $\\mathcal{L}_{\\mathrm{sky}} = \\sum_{\\mathbf{r}^{\\mathrm{sky}}} \\exp (-\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}^{\\mathrm{sky}}))$ .", + "bbox": [ + 498, + 482, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Blocking: Capitalizing on a single Instant-NGP to cover an extraordinarily large scene such as KITTI-360 [37] would often lead to inferior results. We thus adopt a strategy akin to BlockNeRF [63] where we divided the whole scene into numerous blocks and model each region with a separate Instant-NGP. Adjacent regions maintain substantial overlaps to ensure smooth transition.", + "bbox": [ + 498, + 594, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning: We jointly optimize the voxel feature $\\Phi$ and the MLP weights $\\theta$ by minimizing the following loss:", + "bbox": [ + 498, + 705, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} ^ {\\mathrm {N e R F}} = \\mathcal {L} _ {\\mathrm {r g b}} + \\mathcal {L} _ {\\text {n o r m a l}} + \\mathcal {L} _ {\\text {s e m a n t i c s}} + \\mathcal {L} _ {\\text {d e p t h}} + \\mathcal {L} _ {\\text {s k y}} + \\mathcal {L} _ {\\text {s p}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 746, + 890, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. NeRF Baking", + "text_level": 1, + "bbox": [ + 500, + 786, + 640, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to create a digital replica that users (or agents) can freely explore and act upon in real time. Although our large-scale NeRF effectively renders high-quality images and geometry, its efficiency is limited by the computational costs associated with sampling 3D points. The underlying volume density representation further complicates", + "bbox": [ + 498, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4581", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bf3689a37085cba93564561b423ce014855cfad58c85d174192779621e6628a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRepresentationKITTI-360GardenvaseInteractive Compatibility
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Real timeRigid-body physicsScene decomposition
InstantNGP [48]27.460.8530.16525.900.7570.191XXX
Nerfacto [64]Volume23.200.7630.23822.160.5170.283XXX
Video2Game27.620.8710.13126.570.8150.143XXX
Gauss. Spl. [28]Points17.850.6150.42827.500.8580.099✓XX
MobileNeRF [16]19.670.6270.45222.800.5050.365✓XX
BakedSDF* [76]Mesh22.370.7570.30222.680.5140.369✓✓X
Video2Game23.350.7650.24622.810.5080.363✓✓✓
", + "bbox": [ + 81, + 88, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on novel view synthesis and interactive compatibility analysis. Video2Game produces better or comparable results across scenes, suggesting the effectiveness of our NeRF and mesh model. The performance improves the most when tackling unbounded, large-scale scenes in KITTI-360. We note that existing NeRFs cannot reach the interactive rate required for real-time games. While point-based rendering significantly improves the speed, it does not support rigid body physics. BakedSDF [76] represents the whole scene with one single mesh, thus does not support object-level interactions.", + "bbox": [ + 75, + 224, + 892, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the problem. For instance, it's unclear how to define physical interaction with such a representation (e.g., defining collision). The representation is also not compatible with common graphics engines. While recent software, such as the NeRFStudio Blender plugin and LumaAI Unreal add-on, has made some strides, their interaction capabilities and scene geometry quality are still not optimal for real-time user engagement, especially when the scene is large and the observations are relatively sparse. To overcome these challenges, we draw inspiration from recent NeRF meshing advancements and present a novel NeRF baking framework that efficiently transforms our NeRF representation into a game-engine compatible mesh. As we will show in Sec. 4, this conversion greatly enhances rendering efficiency while preserving quality and facilitates physical interactions.", + "bbox": [ + 75, + 321, + 472, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh representation: Our mesh $\\mathcal{M} = (\\mathbf{V},\\mathbf{F},\\mathbf{T})$ comprises vertices $\\mathbf{V}\\in \\mathbb{R}^{|V|\\times 3}$ , faces $\\mathbf{F}\\in \\mathbb{N}^{|F|\\times 3}$ and a UV neural texture map $\\mathbf{T}\\in \\mathbb{R}^{H\\times W\\times 6}$ . Following [65], we store the base color in the first three dimension of $\\mathbf{T}$ , and encode the specular feature in the rest. The initial mesh topology are obtained by marching cubes in the NeRF density field. We further prune the invisible faces. conduct mesh decimation and edge length regularization. The UV coordinate of each vertex is calculated via xatlas [7].", + "bbox": [ + 75, + 555, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Rendering:** We leverage differentiable renderers [33] to render our mesh into RGB images $\\mathbf{C}_{\\mathbb{R}}$ and depth maps $\\mathbf{D}_{\\mathbb{R}}$ . Specifically, we first rasterize the mesh into screen space and obtain the UV coordinate for each pixel $i$ . Then we sample the corresponding texture feature $\\mathbf{T}_i = [\\mathbf{B}_i; \\mathbf{S}_i]$ and feed it into our customized shader. Finally, the shader computes the sum of the view-independent base color $\\mathbf{B}_i$ and the view-dependent MLP $\\mathrm{MLP}_{\\theta}^{\\mathrm{shader}}(\\mathbf{S}_i, \\mathbf{d}_i)$ :", + "bbox": [ + 75, + 698, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathrm {R}} = \\mathbf {B} _ {i} + \\operatorname {M L P} _ {\\theta} ^ {\\text {s h a d e r}} \\left(\\mathbf {S} _ {i}, \\mathbf {d} _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 825, + 468, + 842 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MLP is lightweight and can be baked in GLSL.", + "bbox": [ + 76, + 847, + 419, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning: We train the shader MLP $\\mathbf{MLP}_{\\theta}^{\\mathrm{shader}}$ and the neural texture map $\\mathbf{T}$ by minimizing the color difference", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between the mesh and the ground truth, and the geometry difference between the mesh and the NeRF model:", + "bbox": [ + 498, + 321, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {T}, \\theta} ^ {\\text {m e s h}} = \\sum_ {\\mathbf {r}} \\| \\mathbf {C} _ {\\mathrm {R}} (\\mathbf {r}) - \\mathbf {C} _ {\\mathrm {G T}} (\\mathbf {r}) \\| + \\| \\mathbf {D} _ {\\mathrm {R}} (\\mathbf {r}) - \\mathbf {D} _ {\\mathrm {N e R F}} (\\mathbf {r}) \\|. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 359, + 892, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anti-aliasing: Common differentiable rasterizers only take the center of each pixel into account. This may lead to aliasing in the learned texture map. To resolve this issue, we randomly perturb the optical center of the camera by 0.5 pixels along each axis at every training step. This ensure all the regions within a pixel get rasterized.", + "bbox": [ + 498, + 402, + 890, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relationship to existing work: Our approach is closely related to recent work on NeRF meshing [16, 55, 65, 76], but there exist key differences. While MobileNeRF [16] also adopts an explicit mesh with neural textures, they mainly capitalize on planar primitives. The quality of the reconstructed mesh is thus inferior. BakedSDF [76] offers excellent runtime and rendering quality, but their vertex coloring approach has limited resolution for large scenes. NeRF2Mesh [65] lacks depth distillation and doesn't adopt contraction space for unbounded scenes. They also have a sophisticated multi-stage training and multi-resolution refinement process. Finally, MeRF [55], though efficient, still relies on volume-rendering.", + "bbox": [ + 496, + 503, + 892, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Representation for Physical Interaction", + "text_level": 1, + "bbox": [ + 500, + 710, + 836, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our mesh model facilitates efficient novel-view rendering in real time and allows for basic rigid-body physical interactions. For example, the explicit mesh structure permits an agent to \"stand\" on the ground. Nevertheless, beyond navigation, an agent should be capable of performing various actions including collision and manipulation. Furthermore, a scene comprises not only the background but also interactable foreground objects, each possessing unique physical properties. For instance, a street-bound car is much heavier than a flower vase. When struck by another object, a car may barely move but the vase may fall and shatter.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4582", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance physical interaction realism, we decompose the scene into discrete, actionable entities, each endowed with specific physical characteristics (e.g., mass, friction). This approach, in conjunction with rigid-body physics, allows for the effective simulation that adheres to physical laws.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene decomposition: Directly editing and decomposing a mesh is extremely difficult due to topology change. Fortunately, neural fields are inherently compositional in 3D. By identifying the objects each spatial region belongs to, we can use neural fields to guide the decomposition of the mesh. Specifically, we sample a 3D point $\\mathbf{x}_i$ within each voxel $i$ and determine its semantic category either through the predicted semantic logits $s_i$ or by verifying whether the point is within a specified bounding box. The process is repeated for all voxels to segment the entire scene. Then, for each object, we perform NeRF meshing individually, setting the density of the remaining areas to zero. The intersections between objects are automatically resolved by marching cube. Finally, we initialize the neural texture of these new, individual meshes from the original mesh model. For newly created faces, we employ nearest neighbor inpainting on the neural texture map, which empirically yields satisfactory results. Fig. 1 shows an example where a vase is separated from a table. The middle of the table is original occluded yet we are able to maintain high-quality rendering.", + "bbox": [ + 76, + 179, + 472, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Physical parameters reasoning: The next step is to equip decomposed individual meshes with various physics-related attributes so that we can effectively model and simulate their physical dynamics. In this work, we focus on rigid body physics, where each entity $i$ is represented by a collision geometry $\\text{col}_i$ , mass $m_i$ , and friction parameters $f_i$ . We support four types of collision geometry with different levels of complexity and efficiency: box, sphere, convex polygon, and triangle mesh. Depending on the object and the task of interest, one can select the most suitable collision check for them. For other physical parameters (e.g. mass, friction), one can either set them manually or query large language models (LLMs) for an estimation.", + "bbox": [ + 76, + 494, + 472, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Physical interactions: Rigid body dynamics, while simple, can support a variety of interactions. With the collision check, an user/agent can easily navigate through the environment while respecting the geometry of the scene. The agents will no longer be stuck in a road or cut through a wall. It also allows the agent to interact with the objects in the scene. For instance, one can push the objects towards the location of interest. The object movement will be determined by its mass and other physical properties such as the friction. We can also manipulate the objects by adopting a magnet grasper, following AI2-Thor [32]. This opens the avenue towards automatic creation of realistic, interactive virtual environment for robot learning.", + "bbox": [ + 76, + 704, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Interactive Environment", + "text_level": 1, + "bbox": [ + 500, + 90, + 725, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We deploy our interactive environment within a real-time, browser-based game engine. We manage the underlying logic and assets using Sketchbook [3], a Game Engine based on Three.js that leverages WebGL [4] for rendering. This combination ensures high efficiency while offering the flexibility and sophistication required for intricate rendering tasks. It also allows us to easily integrate content from different scenes together. We have further extended Sketchbook's capabilities by implementing a GLSL-based Shader [2]. This enables real-time computation of our MLP-based specular Shader during deployment. For physics simulation, we use Cannon.js [1], which assures realism and efficiency in the motion within our interactive environment. It supports not only rigid body dynamics but also more sophisticated modeling techniques. For example, we can precompute the fracturing effect for dynamic objects. Upon experiencing a significant force, these objects are realistically simulated by the real-time physics engine, which handles the interactions between the fractured pieces and the rest of the scene, such as their falling and settling on the ground. Besides browser-based engine, the virtual environments from Video2Game pipeline could be also integrated into both Blender [17] and Unreal engines [21] (see Fig. 6).", + "bbox": [ + 496, + 113, + 893, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 473, + 633, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by presenting our experimental setup, followed by a comparison of our model with state-of-the-art approaches. Next, we conduct an extensive analysis of our model's distinctive features and design choices. Then we demonstrate how we constructed a web browser-compatible game capable of delivering a smooth interactive experience exceeding 100 frames per second (FPS), all derived from a single video source. Finally, we showcase the capabilities of our model in robot simulation through two demonstrations.", + "bbox": [ + 496, + 498, + 893, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 636, + 581, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: We evaluate the effectiveness of Video2Game across three distinct scenes in various scenarios, including \"Gardenvase\" [12], an outdoor object-centric scene; the KITTI-360 dataset [37], a large-scale self-driving scene with a sequence that forms a closed loop, suitable for carracing and Temple Run-like games; and finally, an indoor scene from the VR-NeRF [73] dataset, showcasing the potential for robot simulations.", + "bbox": [ + 496, + 657, + 893, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics: To evaluate the quality of the rendered images, we adopt PSNR, SSIM, and LPIPS [82]. For geometry reconstruction, we evaluate with LiDAR point cloud in KITTI-360 dataset. Root mean square deviation (RMSE), mean absolute error (MAE), and outlier rate are applied to measure the disparity existing in estimated geometry.", + "bbox": [ + 496, + 779, + 893, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model: For NeRF, we adopt hashgrid encoding [47] and two-layer MLP for each header. For textured mesh,", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4583", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42d4570516ce8748ec52282b39a31d448c8cd3999f71542fbfebec826cfddc3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 88, + 303, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b79dda2042d703d677b8c2f5ef460727d0c09e83e3170a1b30cfcdd705029ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 88, + 483, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84d04c47b272545d9f18f790d653a2787eb04894274b498b0136ca6fd899ece9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 666, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a3b9e70cfb3dca7ee2df1412cd3eeb0317628929fb2910cf9f516c269603663.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 88, + 846, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7f5c4580edb50b073c204ff1a59d742fdac793b472b94442d7f37573e2478c3.jpg", + "image_caption": [ + "G.T." + ], + "image_footnote": [], + "bbox": [ + 122, + 181, + 303, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84f432847ef02c1ec883e74bfb4e3b1d4623389efdac16630774e2939aa0be56.jpg", + "image_caption": [ + "MobileNeRF [16]" + ], + "image_footnote": [], + "bbox": [ + 303, + 181, + 483, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ad283fe30eb91e3a3a7fb127bf31015209c73279f4fb3974eef7d222aa23a64.jpg", + "image_caption": [ + "BakedSDF [76]" + ], + "image_footnote": [], + "bbox": [ + 483, + 181, + 665, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/433677b2ba6d51393bdfa26b198e208d4b2e54ee7005136c27b532f034c4a512.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 666, + 181, + 846, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51a9e47e3da5629b7ba65f1a62cd8e8c49826fd0986a08c3f874e5c7a375e627.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons among mesh models. We compare our mesh rendering method with others in Garden scene [12].", + "Blender", + "Figure 6. Video2Game in Blender and Unreal Engine." + ], + "image_footnote": [], + "bbox": [ + 83, + 340, + 289, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bda57ece2b41809c4487435009ec56dcbd74e4263cd10d08271ba9c2bafe7821.jpg", + "image_caption": [ + "Unreal" + ], + "image_footnote": [], + "bbox": [ + 290, + 340, + 480, + 433 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/85980f0e309d0c6d6e637f658b6ac4959b51e89cc7a515dbcd1241c5ab05dd45.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOutlier-%↓RMSE↓MAE↓
Instant-NGP [48]22.894.3001.577
Nerfacto [64]50.958.0072.863
Gauss. Spl. [28]91.0811.7688.797
BakedSDF* (offline) [76]43.785.9362.509
Video2Game (Our NeRF)13.233.0281.041
", + "bbox": [ + 120, + 487, + 429, + 566 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5874869482428b7f87ca4d68e32a3a9b6373f061afdece4f0ceb457041bc1ebe.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation on NeRF geometry. Our NeRF renders significantly more accurate depth compared with the baselines. The unit is meter and the outlier threshold is 1.5 meters." + ], + "table_footnote": [], + "table_body": "
MethodVolume RenderingMesh Rastization
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Vanilla NGP27.460.8530.16522.540.7160.350
+ Regularization terms27.520.8610.15722.970.7320.303
+ Monocular cues27.620.8710.13123.350.7650.246
", + "bbox": [ + 81, + 630, + 468, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on KITTI-360.", + "bbox": [ + 148, + 702, + 397, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we conduct marching cubes on the NeRF and post-process it to a fixed precision. We set the texture image size to 4096x4096. For GLSL Shader, we design a light-weight two-layer MLP, which enables efficient real-time rendering. For KITTI-360 (see Sec. 3.1), we divide the whole scene into 16 blocks and create a skydome mesh for the sky.", + "bbox": [ + 75, + 747, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines: To evaluate the visual and geometry quality of our model, we compare against SOTA approaches in neural rendering and neural reconstruction. Instant-NGP [48] is a NeRF-based method that exploits multi-resolution", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hashing encoding. Nerfacto [64] extends the classic NeRF with learnable volumetric sampling and appearance embedding. 3D Gaussian Splatting [28] leverages 3D Gaussians and achieves fast training and rendering. MobileNeRF [16] adopts a hybrid NeRF-mesh representation. It can be baked into a texture map and enable real-time rendering. BakedSDF [76] adopts a volume-surface scene representation. It models view-dependent appearance efficiently by baking spherical Gaussians into the mesh.", + "bbox": [ + 496, + 343, + 890, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental results", + "text_level": 1, + "bbox": [ + 500, + 494, + 694, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel view synthesis: Tab. 1 shows the rendering performance and interactive compatibility of our model against the baselines on KITTI-360 [37] and Gardenvase [12]. Our NeRF achieves superior performance when compared to state-of-the-art neural volume render approaches across different scenes. Though [28] performs best in Gardenvase, it fails to handle the sparse camera settings in KITTI-360, where it learns bad 3D orientations of Gaussians. Our baked mesh outperforms other mesh rendering baselines significantly in KITTI-360 and performs similarly in Gardenvase as shown in Fig. 5. Additionally, our pipeline has the highest interactive compatibility among all baselines.", + "bbox": [ + 496, + 518, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Geometry reconstruction:** Our model performs significantly better than the baseline regarding geometry accuracy (see Tab. 2). We provide some qualitative results in Fig. 4, demonstrating that our model can generate high-quality depth maps and surface normals, whereas those produced by the baselines contain more noise.", + "bbox": [ + 496, + 702, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study: To understand the contribution of each component in our model, we begin with the basic Instant-NGP [48] and sequentially reintroduce other components. The results in Tab. 3 show that our regularization and monocular cues improve the quality of both volume rendering in NeRF and mesh rasterization. Additionally, we do observe a decline in rendering performance when convert", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4584", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing NeRF into game engine-compatible meshes.", + "bbox": [ + 76, + 90, + 395, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Video2Game", + "text_level": 1, + "bbox": [ + 76, + 118, + 212, + 132 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have shown our approach's effectiveness in rendering quality and reconstruction accuracy across various setups. Next, we demonstrate the construction of a web browser-compatible game enabling player control and interaction with the environment at over 100 FPS.", + "bbox": [ + 75, + 142, + 467, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data preparation: We build our environments based on videos in Gardenvase [12], KITTI-360 [37] and VR-NeRF [73] mentioned in Sec. 4.1, using our proposed approach. The outcomes include executable environments with mesh geometry, materials, and rigid-body physics, all encoded in GLB and texture files.", + "bbox": [ + 75, + 218, + 467, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Game engine: We build our game based on several key components in our game engine mentioned in Sec. 3.4. By leveraging them, our game generates a highly realistic visual rendering as well as physical interactions (see Fig. 1) and runs smoothly at an interactive rate across various platforms, browsers, and hardware setups (see Tab. 4). As for other game engines (see Fig. 6), in Blender [17] we showcase the compatibility of our exported assets with other game engines. For Unreal [21], we further demonstrate a real-time game demo where a humanoid robot can freely interact within the Gardenvase scene, such as standing on the table and kicking off the central vase. These prove the compatibility of our proposed pipeline.", + "bbox": [ + 75, + 310, + 467, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interactive game features: Movement in games: Agents can navigate the area freely within the virtual environment where their actions follow real-world physics and are constrained by collision models. Shooting game: For realistic shooting physics, we calculated the rigid-body collision dynamics for both the central vase and the surrounding scene (see Fig. 3), separated using mesh semantic filtering. We used a box collider for the vase and convex polygon colliders for the background. The player shoots footballs with a sphere collider at the vase on the table, causing it to fly off and fall to the ground (see Fig. 1). Temple-Run like game: The agent collects coins while running in the KITTI Loop composed of four streets in KITTI-360. Obstructive chairs on the road can be smashed thanks to pre-computed fracture animations. The agent can also drive and push roadside vehicles existing in the scene forward by crashing into them. This interactivity is achieved through rigid-body dynamics simulation and collision modeling.", + "bbox": [ + 75, + 507, + 467, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robot simulation: We demonstrate the potential of leveraging Video2Game for robot simulation using the VRNeRF dataset. We reconstruct the scene and segment simulatable rigid-body objects (e.g., the fruit bowl on the table). We show two demos in Fig. 7: a Stretch Robot pushing the bowl off the table and a Fetch Robot performing pick-and-place actions. We employ PyBullet [18] to simulate the underlying physics with the help of corresponding collision", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8564a5cc8c8dd4435bc50abc3bf6ae97e1bec9b809cbe1dab034917184f53ddf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PlatformFPS (hz)CPU-Usage (%)GPU-Usage (%)
Mac M1 ProMac OS, Chrome1023470
Intel Core i9 + NV 4060Windows, Edge240674
AMD 5950 + NV 3090Linux, Chrome1442040
", + "bbox": [ + 504, + 80, + 890, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Runtime Analysis. Our interactive environment can run in real-time across various hardware setup and various platforms. User actions may slightly vary, which could lead to minor variations in runtime.", + "bbox": [ + 498, + 138, + 890, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3111f1e2dfca3c5576de0b283bf6c3f82b67ffba0c3e2ad118f672b97ddc492.jpg", + "image_caption": [ + "Figure 7. Robot simulation in VRNeRF [73] dataset. We demonstrate the possibility of conducting robot learning in our virtual environments using Stretch Robot [6] and Fetch Robot [5]." + ], + "image_footnote": [], + "bbox": [ + 516, + 210, + 890, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models. Since real-time grasping simulation is challenging, following existing robot simulation frameworks [8, 9, 32], objects near the Fetch gripper are automatically picked up. This demonstrates our model's ability to convert a real-time video stream into a virtual environment, allowing robots to rehearse before acting in the real environment.", + "bbox": [ + 498, + 492, + 890, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 597, + 617, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel approach to converting real-world video footage into playable, real-time, and interactive game environments. Specifically, we combine the potential of NeRF modeling with physics modeling and integrate them into modern game engines. Our approach enables any individual to transform their surroundings into an interactive digital environment, unlocking exciting possibilities for 3D content creation, with promising implications for future advancements in digital game design and robot simulation.", + "bbox": [ + 498, + 621, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This project is supported by the Intel AI SRS gift, the IBM IIDAI Grant, the InsperIllinois Innovation Grant, the NCSA Faculty Fellowship, and NSF Awards #2331878, #2340254, and #2312102. We greatly appreciate the NCSA for providing computing resources. We thank Derek Hoiem and Albert Zhai for helpful discussions. We thank Jingkang Wang, Joao Marques and Hanxiao Jiang for proofreading.", + "bbox": [ + 498, + 773, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4585", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cannon.js. https://scheppe.github.io/cannon.js/.6", + "[2] GLSL. https://www.khronos.org/opengl/wiki/OpenGL_Shading_Language.6", + "[3] Sketchbook. https://github.com/swift502/Sketchbook.6", + "[4] WebGL. https://www.khronos.org/webgl/.6", + "[5] Fetch Mobile Manipulator. https://fetchrobotics.borealtech.com/robotics-platforms/fetch-mobile-manipulator/?lang=en.8", + "[6] Stretch® Research Edition. https://hello-robot.com/product.8", + "[7] Xatlas. https://github.com/mworchel/xatlaspython.5", + "[8] RoboTHOR: An Open Simulation-to-Real Embodied AI Platform. CVPR, 2020. 8", + "[9] ManipulaTHOR: A Framework for Visual Object Manipulation. In CVPR, 2021. 8", + "[10] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In ICRA, 2022. 3", + "[11] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2", + "[12] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 4, 6, 7, 8", + "[13] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. ECCV, 2022. 2", + "[14] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, 1993. 2", + "[15] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3", + "[16] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. CVPR, 2023. 2, 5, 7", + "[17] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6, 8", + "[18] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 8", + "[19] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 1, 3", + "[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4", + "[21] Epic Games. Unreal engine. 1, 6, 8", + "[22] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[23] Rockstar Games. Grand theft auto v. 2014. 1", + "[24] John K Haas. A history of the unity game engine. 2014. 1", + "[25] Benno Heigl, Reinhard Koch, Marc Pollefeys, Joachim Denzler, and Luc Van Gool. Plenoptic modeling and rendering from image sequences taken by a hand-held camera. In DAGM-Symposium, 1999. 2", + "[26] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In ICCV, 2021. 2", + "[27] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4", + "[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 5, 7", + "[29] Pradeep K Khosla and Takeo Kanade. Parameter identification of robot dynamics. In IEEE conference on decision and control, 1985. 3", + "[30] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2, 3", + "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 3", + "[32] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai, 2022. 6, 8", + "[33] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5", + "[34] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv, 2023. 2", + "[35] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2", + "[36] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv, 2022. 2, 4", + "[37] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 4, 6, 7, 8", + "[38] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4586", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 2", + "[40] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattad, and Shenlong Wang. Urbanir: Large-scale urban scene inverse rendering from a single video, 2023. 2", + "[41] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3", + "[42] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives, 2023. 3", + "[43] Sivabalan Manivasagam, Shenlong Wang, Kelvin Wong, Wenyuan Zeng, Mikita Sazanovich, Shuhan Tan, Bin Yang, Wei-Chiu Ma, and Raquel Urtasun. Lidarsim: Realistic lidar simulation by leveraging the real world. In CVPR, 2020. 3", + "[44] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021. 3", + "[45] Willi Menapace, Stéphane Lathuiliere, Aliaksandr Siarohin, Christian Theobalt, Sergey Tulyakov, Vladislav Golyanik, and Elisa Ricci. Playable environments: Video manipulation in space and time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3584-3593, 2022. 3", + "[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4", + "[47] Thomas Muller. tiny-cuda-nn, 2021. 6", + "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 3, 4, 5, 7", + "[49] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2", + "[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021.", + "[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv, 2021. 2", + "[52] Cristina Portales, José Luis Lerma, and Santiago Navarro. Augmented reality and photogrammetry: A synergy to visualize physical and virtual city environments. ISPRS Journal of Photogrammetry and Remote Sensing, 65(1):134-142, 2010. 1", + "[53] Charalambos Poullis, Andrew Gardner, and Paul Debevec. Photogrammetric modeling and image-based rendering for" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rapid virtual environment creation. In Proceedings of ASC, 2004. 1", + "[54] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 4", + "[55] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv preprint arXiv:2302.12249, 2023. 2, 5", + "[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2", + "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv, 2022. 2, 3", + "[58] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv, 2023. 2, 3", + "[59] Sanghyun Son, Yi-Ling Qiao, Jason Sewall, and Ming C Lin. Differentiable hybrid traffic simulation. TOG, 2022. 3", + "[60] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2", + "[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[62] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision, 1998. 2", + "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 4", + "[64] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv, 2023. 5, 7", + "[65] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2, 5", + "[66] ADRIAN SAVARI Thomas, MOHD FAHRUL Hassan, MUSTAFFA Ibrahim, M Nasrull, A Rahman, SZ Sapuan, and F Ahmad. A study on close-range photogrammetry in image based modelling and rendering (imbr) approaches and post-processing analysis. Journal of Engineering Science and Technology, 14(4):1912-1923, 2019. 1", + "[67] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IROS, 2012. 3" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4587", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2", + "[69] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2", + "[70] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv, 2021. 2", + "[71] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 2000. 3", + "[72] Yuwen Xiong, Jingkang Ma, Wei-Chiu Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. CVPR, 2023. 3", + "[73] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 6, 8", + "[74] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In CVPR, 2020. 3", + "[75] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. CVPR, 2023. 2, 3", + "[76] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. Siggraph, 2023. 2, 5, 7", + "[77] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields, 2021. 4", + "[78] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. in arXiv, 2022. 4", + "[79] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph. 3", + "[80] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Transactions on Graphics (TOG), 40(3):1-16, 2021. 3", + "[81] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2020. 2", + "[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2", + "[84] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation, 2021. 4", + "[85] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv, 2018. 2", + "[86] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. in arXiv, 2022. 2", + "[87] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 290 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4588", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_model.json b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0274b63385c3c1631a72f139a905c25fe609f83c --- /dev/null +++ b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.848, + 0.175 + ], + "angle": 0, + "content": "Video2Game: Real-time, Interactive, Realistic and Browser-Compatible Environment from a Single Video" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.203, + 0.758, + 0.221 + ], + "angle": 0, + "content": "Hongchi Xia\\(^{1,2}\\) Zhi-Hao Lin\\(^{1}\\) Wei-Chiu Ma\\(^{3}\\) Shenlong Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.221, + 0.877, + 0.257 + ], + "angle": 0, + "content": "1University of Illinois Urbana-Champaign 2Shanghai Jiao Tong University 3Cornell University https://video2game.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.271, + 0.301, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.272, + 0.518, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.272, + 0.707, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.272, + 0.891, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.893, + 0.521 + ], + "angle": 0, + "content": "Figure 1. Video2Game takes an input video of an arbitrary scene and automatically transforms it into a real-time, interactive, realistic and browser-compatible environment. The users can freely explore the environment and interact with the objects in the scene." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.54, + 0.314, + 0.555 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.473, + 0.829 + ], + "angle": 0, + "content": "Creating high-quality and interactive virtual environments, such as games and simulators, often involves complex and costly manual modeling processes. In this paper, we present Video2Game, a novel approach that automatically converts videos of real-world scenes into realistic and interactive game environments. At the heart of our system are three core components: (i) a neural radiance fields (NeRF) module that effectively captures the geometry and visual appearance of the scene; (ii) a mesh module that distills the knowledge from NeRF for faster rendering; and (iii) a physics module that models the interactions and physical dynamics among the objects. By following the carefully designed pipeline, one can construct an interactable and actionable digital replica of the real world. We benchmark our system on both indoor and large-scale outdoor scenes. We show that we can not only produce highly-realistic renderings in real-time, but also build interactive games on top." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.208, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Crafting a visually compelling and interactive environment is crucial for immersive experiences in various domains," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.722 + ], + "angle": 0, + "content": "such as video games, virtual reality applications, and self-driving simulators. This process, however, is complex and expensive. It demands the skills of experts in the field and the use of professional software development tools [21, 24]. For instance, Grand Theft Auto V [23], known for its intricately detailed environment, was one of the most expensive video games ever developed, with a budget over $265 million primarily for asset creation. Similarly, the development of the CARLA autonomous driving simulator [19] involves a multidisciplinary team of 3D artists, programmers, and engineers to meticulously craft and texture the virtual cityscapes, creating its lifelike environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "An appealing alternative to extensive manual modelling is creating environments directly from the real world. For instance, photogrammetry, a technique for constructing digital replicas of objects or scenes from overlapping real-world photographs, has been utilized for environment creation [52, 53]. Success stories also span various games and simulators. However, most use cases are limited to creating object assets and necessitate significant post-processing, such as material creation, texturing, and geometry fixes [66]. People thus turn to neural radiance fields (NeRFs) [46], as it offers a more promising approach to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.897, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Overview of Video2Game: Given multiple posed images from a single video as input, we first construct a large-scale NeRF model that is realistic and possesses high-quality surface geometry. We then transform this NeRF model into a mesh representation with corresponding rigid-body dynamics to enable interactions. We utilize UV-mapped neural texture, which is both expressive and compatible with game engines. Finally, we obtain an interactive virtual environment that virtual actors can interact with, can respond to user control, and deliver high-resolution rendering from novel camera perspectives – all in real-time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.473, + 0.543 + ], + "angle": 0, + "content": "modeling large scenes. With careful design [13, 22, 39, 48, 61], NeRF is able to render free-viewpoint, photo-realistic images efficiently. However, crafting an interactive environment entails more than just creating a visually high-fidelity digital twin; it also involves building a physically plausible, immersive, real-time and importantly, interactive world tailored to user experiences. Furthermore, we expect such a virtual world to be compatible with real-time interaction interfaces such as common game engines. Despite its promise, the use of NeRF to create interactive environments from real-world videos remains largely unexplored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.791 + ], + "angle": 0, + "content": "In this paper, we introduce Video2Game, a novel approach to automatically converting a video of a scene into a realistic and interactive virtual environment. Given a video as input, we first construct a NeRF that can effectively capture the geometric and visual information of a large-scale, unbounded) scene. Then we distill the NeRF into a game engine-compatible, neural textured mesh. This significantly improves the rendering efficiency while maintains the overall quality. To model the interactions among the objects, we further decompose the scene into individual actionable entities and equip them with respective physics model. Finally, we import our automatically generated assets into a WebGL-based game engine and create a playable game. The resulting virtual environment is photo-realistic, interactive, and runs in real-time. See Fig. 1 for demonstration. In summary, our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.793, + 0.47, + 0.852 + ], + "angle": 0, + "content": "- A novel 3D modeling algorithm for real-time, free-viewpoint rendering and physical simulation, surpassing state-of-the-art NeRF baking methods with added rigid-body physics for enhanced simulation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.853, + 0.47, + 0.899 + ], + "angle": 0, + "content": "- An automated game-creation framework to transform a scene video into an interactive, realistic environment, compatible with current game engines." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.793, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.651, + 0.391 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Given a single video, we aim to create a real-time, interactive game where the agents (e.g., the character, the car) can navigate and explore the reconstructed digital world, interact with objects in the scene (e.g., collision and manipulate objects), and achieve their respective tasks (e.g., collecting coins, shooting targets). We draw inspirations from multiple areas and combine the best of all. In this section, we will briefly review those closely related areas which forms the foundation of our work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Novel view synthesis (NVS): Our work builds upon the success of novel view synthesis [14, 25, 35, 62], which is crucial for our game since it enables the agents to move freely and view the reconstructed world seamlessly from various perspectives. Among all these approaches [26, 60, 68, 85, 86], we exploit neural radiance field (NeRF) [46] as our underlying representation. NeRF has emerged as one of the most promising tools in NVS since its introduction [49-51], and has great performance across a wide range of scenarios [36, 56, 75, 81]. For instance, it can be easily extended to handle various challenging real-world scenarios such as learning from noisy camera poses [38, 70], reflectance modeling for photo-realistic relighting [69, 83], and real-time rendering [16, 39, 55, 65, 76]. In this work, we combine recent advances in NeRF with physics modeling to build an immersive digital world from one single video, moving from passive NVS to our complete solution for embodied world modeling where agents can actively explore and interact with the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Controllable video generation: Using different control signals to manipulate the output of a visual model has garnered great interest in the community. This has had a profound impact on content creation [57, 58], digital editing [11, 34], and simulation [30, 31, 40]. One could also lever" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "age large foundation models to control video content using text [57, 58]. However, they lack fine-grained and real-time control over the generated content. Alternatively, training (conditional) generative models for each scene enables better disentanglement of dynamics (e.g., foreground vs. background) and supports better control. For instance, one can represent a self-driving scene [31] or a Pacman game [30] as latent codes and generate video frames based on control inputs with a neural network. One can also learn to control the players within tennis games [44, 45, 79, 80]. Our work falls under the second line of research, where the model takes user control signals (e.g., a keystroke from the keyboard) as input and responds by rendering a new scene. However, instead of focusing on a specific scene (e.g., tennis games), we have developed a pipeline that allows the creation of a playable environment from a single video of a generic scene. Additionally, we model everything in 3D, which enables us to effectively capture not only view-dependent appearance but also physical interactions among rigid-body equipped objects. Importantly, we adopt a neural representation compatible with graphics engines, enabling users to play the entire game in their browser at an interactive rate." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.473, + 0.711 + ], + "angle": 0, + "content": "Data-driven simulation: Building a realistic simulation environment has been a longstanding challenge. [19, 29, 67, 71]. While it's promising, we come close to mirror the real world only in recent years [10, 15, 42, 43, 59, 74, 75]. The key insight of these work is to build models by leveraging real-world data. Our work closely aligns with this line of research on building high-fidelity simulators from real-world data, with a few key differences. First, existing works mainly focus on offline training and evaluation [10, 43, 74, 75], whereas our system runs at an interactive rate and allows for online, real-time control. Second, some existing works [41, 43, 72, 87] need additional data modality like LiDAR point clouds for geometry reconstruction, but RGB video is all we need. Third, most photo-realistic simulators don't model physical interactions. However, we supports various physics modeling and allows agents to interact with the environment. Last, existing simulators are typically resource-intensive, while our system is lightweight and can be easily accessible in common engines." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.212, + 0.74 + ], + "angle": 0, + "content": "3. Video2Game" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given a sequence of images or a video of a scene, our goal is to construct an interactable and actionable digital twin, upon which we can build real-time, interactive games or realistic (sensor) simulators. Based on the observations that prevalent approaches to constructing digital replica mainly focus on visual appearance and ignore the underlying physical interactions, we carefully design our system such that it can not only produce high-quality rendering across viewpoints, but also support the modeling of physical actions (e.g., navigation, collision, manipulation, etc). At the heart" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.697, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.17, + 0.697, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.092, + 0.887, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.265, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Visualization of automatically computed collision geometry: Sphere collider (green), box collider (yellow), convex polygon collider (purple) and trimesh collider (red)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.427 + ], + "angle": 0, + "content": "of our systems is a compositional implicit-explicit 3D representation that is effective and efficient for both sensor and physics simulation. By decomposing the world into individual entities, we can better model and manipulate their physical properties (e.g., specularity, mass, friction), and simulate the outcomes of interactions more effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.893, + 0.639 + ], + "angle": 0, + "content": "We start by introducing a NeRF model that can effectively capture the geometric and visual information of a large-scale, unbounded scene (Sec. 3.1). Next, we present an approach to convert the NeRF into a game-engine compatible mesh with neural texture maps, significantly improving the rendering efficiency while maintaining the quality (Sec. 3.2). To enable physical interactions, we further decompose the scene into individual actionable entities and equip them with respective physics models (Sec. 3.3). Finally, we describe how we integrate our interactive environment into a WebGL-based game engine, allowing users to play and interact with the virtual world in real time on their personal browser. Fig. 2 provides an overview of our proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.674, + 0.667 + ], + "angle": 0, + "content": "3.1. Large-scale NeRF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Preliminaries: Instant-NGP [48] is a notable variant of NeRF, which represents the radiance field with a combination of spatial hash-based voxels and neural networks: \\(\\mathbf{c},\\sigma = F_{\\theta}(\\mathbf{x},\\mathbf{d};\\Phi) = \\mathrm{MLP}_{\\theta}(\\mathsf{It}(\\mathbf{x},\\Phi),\\mathbf{d})\\) . Given a 3D point \\(\\mathbf{x}\\in \\mathbb{R}^3\\) and a camera direction \\(\\mathbf{d}\\in \\mathbb{R}^2\\) as input, Instant-NGP first interpolate the point feature \\(\\mathsf{It}(\\mathbf{x},\\Phi)\\) from the adjacent voxel features \\(\\Phi\\) . Then the point feature and the camera direction are fed into a light-weight multilayer perception (MLP) to predict the color \\(\\mathbf{c}\\in \\mathbb{R}^3\\) and density \\(\\sigma \\in \\mathbb{R}^{+}\\) . To render the scene appearance, we first cast a ray \\(\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\) from the camera center o through the pixel center in direction d, and sample a set of 3D points \\(\\{\\mathbf{x}_i\\}\\) along the ray. We then query their respective color \\(\\{\\mathbf{c}_i\\}\\) and density \\(\\{\\sigma_i\\}\\) and obtain the color of the pixel through alpha-composition: \\(\\mathbf{C}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i}w_{i}\\mathbf{c}_{i}\\) . Similarly, the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4580" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.887, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons among NeRF models. The rendering quality of our base NeRF is superior to baselines, and with leveraging monocular cues, we substantially improve rendered geometry compared to other baselines. This significantly facilitates NeRF baking in subsequent stages. Here we consider depths measured by LiDAR point cloud in KITTI-360 and compute normals based on it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.487 + ], + "angle": 0, + "content": "expected depth can be computed by: \\(\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i}w_{i}t_{i}\\) Here, \\(w_{i}\\) indicates the blending weight that is derived from the densities \\(\\{\\sigma_i\\}\\). We refer the readers to [46] for more details. To learn the voxel features \\(\\Phi\\) and the MLP weights \\(\\theta\\), we compute the difference between the ground truth color and the rendered color: \\(\\mathcal{L}_{\\mathrm{rgb}} = \\sum_{\\mathbf{r}}\\| \\mathbf{C}_{\\mathrm{GT}}(\\mathbf{r}) - \\mathbf{C}_{\\mathrm{NeRF}}(\\mathbf{r})\\| _2^2\\). Large-scale NeRF: While Instant-NGP [48] has shown promising results on densely observed and bounded scenes, its performance starts to degrade when extending to sparsely-captured, large-scale, unbounded environments. To mitigate these issues, we propose several enhancements:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.502, + 0.469, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c}, \\sigma , s, \\mathbf {n} = F _ {\\theta} (\\mathbf {x}, \\mathbf {d}; \\Phi) = \\operatorname {M L P} _ {\\theta} (\\operatorname {I t} (\\operatorname {C t} (\\mathbf {x}), \\Phi), \\mathbf {d}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.65 + ], + "angle": 0, + "content": "First of all, we exploit the contraction function \\(\\mathrm{Ct}(\\mathbf{x})\\) [12] to map the unbounded coordinates into a bounded region. In addition to radiance and density, we predict the semantics \\(s\\) and the surface normal \\(\\mathbf{n}\\) of the 3D points, guided with 2D priors to better regularize the scene geometry. Furthermore, we divide large-scale scenes into several blocks [63] to capture the fine-grained details. We now describe these enhancements in more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Depth: High-quality geometry is critical for modeling physical interactions. Inspired by MonoSDF [78], we leverage off-the-shelf monocular depth estimators [20, 27] to guide and improve the underlying geometry. We first predict the depth of the scene from rendered RGB images. Then we minimize the discrepancy between the rendered depth and the predicted depth via \\(\\mathcal{L}_{\\mathrm{depth}} = \\sum_{\\mathbf{r}}\\|(a\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}) + b) - \\mathbf{D}_{\\mathrm{mono}}(\\mathbf{r})\\|_2^2\\), where \\(a\\) and \\(b\\) are the scale and shift that aligns the two distribution [54]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Surface normals: Similar to depth, we encourage the normal estimated from NeRF to be consistent with the normal predicted by the off-the-shelf estimator [20, 27]. The normal of a 3D point \\(\\mathbf{x}_i\\) can be either analytically derived from the estimated density \\(\\mathbf{n}_i = (1 - \\frac{\\nabla_{\\mathbf{x}}\\sigma_i}{\\|\\nabla\\sigma_i\\|})\\), or predicted by the MLP header as in Eq. 1. We could aggregate them via volume render: \\(\\mathbf{N}(\\mathbf{r}) = \\sum_{i}\\mathbf{w}_{i}\\mathbf{n}_{i}\\). Em" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.32, + 0.892, + 0.382 + ], + "angle": 0, + "content": "pirically we find that adopting both normals and promoting their mutual consistency works the best, since the MLP header offers more flexibility. We thus employ \\(\\mathcal{L}_{\\mathrm{normal}} = \\|\\mathbf{N}_{\\mathrm{mlp}}(\\mathbf{r}) - \\mathbf{N}_{\\mathrm{mono}}(\\mathbf{r})\\|_2^2 + \\|\\mathbf{N}_{\\mathrm{mlp}}(\\mathbf{r}) - \\mathbf{N}_{\\mathrm{density}}(\\mathbf{r})\\|_2^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.386, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Semantics: We also predict semantic logits for each sampled 3D points with our MLP. This helps us capture the correlation between semantics and geometry [36, 84]. We render the semantic map with volume rendering \\(\\mathbf{S}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i} \\mathbf{w}_{i} \\mathbf{s}_{i}\\) and compute the cross-entropy with that of a 2D segmentation model \\(\\mathcal{L}_{\\text{semantics}} = \\mathrm{CE}(\\mathbf{S}_{\\text{mono}}, \\mathbf{S}_{\\text{NeRF}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Regularization: We additionally adopt two regularization terms. To reduce floaters in the scene, for each randomly sampled 3D point \\(\\mathbf{x}\\), we penalize its density by \\(\\mathcal{L}_{\\mathrm{sp}} = \\sum 1 - \\exp (-\\alpha \\sigma (\\mathbf{x}))\\), where \\(\\alpha >0\\) [77]. For each sky pixel (which we derived from the semantic MLP), we encourage its depth \\(\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}^{\\mathrm{sky}})\\) to be as far as possible. The loss is defined as: \\(\\mathcal{L}_{\\mathrm{sky}} = \\sum_{\\mathbf{r}^{\\mathrm{sky}}} \\exp (-\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}^{\\mathrm{sky}}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Blocking: Capitalizing on a single Instant-NGP to cover an extraordinarily large scene such as KITTI-360 [37] would often lead to inferior results. We thus adopt a strategy akin to BlockNeRF [63] where we divided the whole scene into numerous blocks and model each region with a separate Instant-NGP. Adjacent regions maintain substantial overlaps to ensure smooth transition." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.707, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Learning: We jointly optimize the voxel feature \\(\\Phi\\) and the MLP weights \\(\\theta\\) by minimizing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.747, + 0.892, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} ^ {\\mathrm {N e R F}} = \\mathcal {L} _ {\\mathrm {r g b}} + \\mathcal {L} _ {\\text {n o r m a l}} + \\mathcal {L} _ {\\text {s e m a n t i c s}} + \\mathcal {L} _ {\\text {d e p t h}} + \\mathcal {L} _ {\\text {s k y}} + \\mathcal {L} _ {\\text {s p}} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.642, + 0.803 + ], + "angle": 0, + "content": "3.2. NeRF Baking" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We aim to create a digital replica that users (or agents) can freely explore and act upon in real time. Although our large-scale NeRF effectively renders high-quality images and geometry, its efficiency is limited by the computational costs associated with sampling 3D points. The underlying volume density representation further complicates" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4581" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.216 + ], + "angle": 0, + "content": "
MethodRepresentationKITTI-360GardenvaseInteractive Compatibility
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Real timeRigid-body physicsScene decomposition
InstantNGP [48]27.460.8530.16525.900.7570.191XXX
Nerfacto [64]Volume23.200.7630.23822.160.5170.283XXX
Video2Game27.620.8710.13126.570.8150.143XXX
Gauss. Spl. [28]Points17.850.6150.42827.500.8580.099✓XX
MobileNeRF [16]19.670.6270.45222.800.5050.365✓XX
BakedSDF* [76]Mesh22.370.7570.30222.680.5140.369✓✓X
Video2Game23.350.7650.24622.810.5080.363✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.226, + 0.893, + 0.298 + ], + "angle": 0, + "content": "Table 1. Quantitative results on novel view synthesis and interactive compatibility analysis. Video2Game produces better or comparable results across scenes, suggesting the effectiveness of our NeRF and mesh model. The performance improves the most when tackling unbounded, large-scale scenes in KITTI-360. We note that existing NeRFs cannot reach the interactive rate required for real-time games. While point-based rendering significantly improves the speed, it does not support rigid body physics. BakedSDF [76] represents the whole scene with one single mesh, thus does not support object-level interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.549 + ], + "angle": 0, + "content": "the problem. For instance, it's unclear how to define physical interaction with such a representation (e.g., defining collision). The representation is also not compatible with common graphics engines. While recent software, such as the NeRFStudio Blender plugin and LumaAI Unreal add-on, has made some strides, their interaction capabilities and scene geometry quality are still not optimal for real-time user engagement, especially when the scene is large and the observations are relatively sparse. To overcome these challenges, we draw inspiration from recent NeRF meshing advancements and present a novel NeRF baking framework that efficiently transforms our NeRF representation into a game-engine compatible mesh. As we will show in Sec. 4, this conversion greatly enhances rendering efficiency while preserving quality and facilitates physical interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Mesh representation: Our mesh \\(\\mathcal{M} = (\\mathbf{V},\\mathbf{F},\\mathbf{T})\\) comprises vertices \\(\\mathbf{V}\\in \\mathbb{R}^{|V|\\times 3}\\), faces \\(\\mathbf{F}\\in \\mathbb{N}^{|F|\\times 3}\\) and a UV neural texture map \\(\\mathbf{T}\\in \\mathbb{R}^{H\\times W\\times 6}\\). Following [65], we store the base color in the first three dimension of \\(\\mathbf{T}\\), and encode the specular feature in the rest. The initial mesh topology are obtained by marching cubes in the NeRF density field. We further prune the invisible faces. conduct mesh decimation and edge length regularization. The UV coordinate of each vertex is calculated via xatlas [7]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.821 + ], + "angle": 0, + "content": "**Rendering:** We leverage differentiable renderers [33] to render our mesh into RGB images \\(\\mathbf{C}_{\\mathbb{R}}\\) and depth maps \\(\\mathbf{D}_{\\mathbb{R}}\\). Specifically, we first rasterize the mesh into screen space and obtain the UV coordinate for each pixel \\(i\\). Then we sample the corresponding texture feature \\(\\mathbf{T}_i = [\\mathbf{B}_i; \\mathbf{S}_i]\\) and feed it into our customized shader. Finally, the shader computes the sum of the view-independent base color \\(\\mathbf{B}_i\\) and the view-dependent MLP \\(\\mathrm{MLP}_{\\theta}^{\\mathrm{shader}}(\\mathbf{S}_i, \\mathbf{d}_i)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.826, + 0.469, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathrm {R}} = \\mathbf {B} _ {i} + \\operatorname {M L P} _ {\\theta} ^ {\\text {s h a d e r}} \\left(\\mathbf {S} _ {i}, \\mathbf {d} _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.848, + 0.421, + 0.864 + ], + "angle": 0, + "content": "The MLP is lightweight and can be baked in GLSL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Learning: We train the shader MLP \\(\\mathbf{MLP}_{\\theta}^{\\mathrm{shader}}\\) and the neural texture map \\(\\mathbf{T}\\) by minimizing the color difference" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.323, + 0.892, + 0.352 + ], + "angle": 0, + "content": "between the mesh and the ground truth, and the geometry difference between the mesh and the NeRF model:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.361, + 0.893, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {T}, \\theta} ^ {\\text {m e s h}} = \\sum_ {\\mathbf {r}} \\| \\mathbf {C} _ {\\mathrm {R}} (\\mathbf {r}) - \\mathbf {C} _ {\\mathrm {G T}} (\\mathbf {r}) \\| + \\| \\mathbf {D} _ {\\mathrm {R}} (\\mathbf {r}) - \\mathbf {D} _ {\\mathrm {N e R F}} (\\mathbf {r}) \\|. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Anti-aliasing: Common differentiable rasterizers only take the center of each pixel into account. This may lead to aliasing in the learned texture map. To resolve this issue, we randomly perturb the optical center of the camera by 0.5 pixels along each axis at every training step. This ensure all the regions within a pixel get rasterized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.702 + ], + "angle": 0, + "content": "Relationship to existing work: Our approach is closely related to recent work on NeRF meshing [16, 55, 65, 76], but there exist key differences. While MobileNeRF [16] also adopts an explicit mesh with neural textures, they mainly capitalize on planar primitives. The quality of the reconstructed mesh is thus inferior. BakedSDF [76] offers excellent runtime and rendering quality, but their vertex coloring approach has limited resolution for large scenes. NeRF2Mesh [65] lacks depth distillation and doesn't adopt contraction space for unbounded scenes. They also have a sophisticated multi-stage training and multi-resolution refinement process. Finally, MeRF [55], though efficient, still relies on volume-rendering." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.838, + 0.727 + ], + "angle": 0, + "content": "3.3. Representation for Physical Interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our mesh model facilitates efficient novel-view rendering in real time and allows for basic rigid-body physical interactions. For example, the explicit mesh structure permits an agent to \"stand\" on the ground. Nevertheless, beyond navigation, an agent should be capable of performing various actions including collision and manipulation. Furthermore, a scene comprises not only the background but also interactable foreground objects, each possessing unique physical properties. For instance, a street-bound car is much heavier than a flower vase. When struck by another object, a car may barely move but the vase may fall and shatter." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "To enhance physical interaction realism, we decompose the scene into discrete, actionable entities, each endowed with specific physical characteristics (e.g., mass, friction). This approach, in conjunction with rigid-body physics, allows for the effective simulation that adheres to physical laws." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.18, + 0.473, + 0.484 + ], + "angle": 0, + "content": "Scene decomposition: Directly editing and decomposing a mesh is extremely difficult due to topology change. Fortunately, neural fields are inherently compositional in 3D. By identifying the objects each spatial region belongs to, we can use neural fields to guide the decomposition of the mesh. Specifically, we sample a 3D point \\(\\mathbf{x}_i\\) within each voxel \\(i\\) and determine its semantic category either through the predicted semantic logits \\(s_i\\) or by verifying whether the point is within a specified bounding box. The process is repeated for all voxels to segment the entire scene. Then, for each object, we perform NeRF meshing individually, setting the density of the remaining areas to zero. The intersections between objects are automatically resolved by marching cube. Finally, we initialize the neural texture of these new, individual meshes from the original mesh model. For newly created faces, we employ nearest neighbor inpainting on the neural texture map, which empirically yields satisfactory results. Fig. 1 shows an example where a vase is separated from a table. The middle of the table is original occluded yet we are able to maintain high-quality rendering." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.496, + 0.473, + 0.693 + ], + "angle": 0, + "content": "Physical parameters reasoning: The next step is to equip decomposed individual meshes with various physics-related attributes so that we can effectively model and simulate their physical dynamics. In this work, we focus on rigid body physics, where each entity \\(i\\) is represented by a collision geometry \\(\\text{col}_i\\), mass \\(m_i\\), and friction parameters \\(f_i\\). We support four types of collision geometry with different levels of complexity and efficiency: box, sphere, convex polygon, and triangle mesh. Depending on the object and the task of interest, one can select the most suitable collision check for them. For other physical parameters (e.g. mass, friction), one can either set them manually or query large language models (LLMs) for an estimation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Physical interactions: Rigid body dynamics, while simple, can support a variety of interactions. With the collision check, an user/agent can easily navigate through the environment while respecting the geometry of the scene. The agents will no longer be stuck in a road or cut through a wall. It also allows the agent to interact with the objects in the scene. For instance, one can push the objects towards the location of interest. The object movement will be determined by its mass and other physical properties such as the friction. We can also manipulate the objects by adopting a magnet grasper, following AI2-Thor [32]. This opens the avenue towards automatic creation of realistic, interactive virtual environment for robot learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.727, + 0.106 + ], + "angle": 0, + "content": "3.4. Interactive Environment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.465 + ], + "angle": 0, + "content": "We deploy our interactive environment within a real-time, browser-based game engine. We manage the underlying logic and assets using Sketchbook [3], a Game Engine based on Three.js that leverages WebGL [4] for rendering. This combination ensures high efficiency while offering the flexibility and sophistication required for intricate rendering tasks. It also allows us to easily integrate content from different scenes together. We have further extended Sketchbook's capabilities by implementing a GLSL-based Shader [2]. This enables real-time computation of our MLP-based specular Shader during deployment. For physics simulation, we use Cannon.js [1], which assures realism and efficiency in the motion within our interactive environment. It supports not only rigid body dynamics but also more sophisticated modeling techniques. For example, we can precompute the fracturing effect for dynamic objects. Upon experiencing a significant force, these objects are realistically simulated by the real-time physics engine, which handles the interactions between the fractured pieces and the rest of the scene, such as their falling and settling on the ground. Besides browser-based engine, the virtual environments from Video2Game pipeline could be also integrated into both Blender [17] and Unreal engines [21] (see Fig. 6)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.634, + 0.491 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.895, + 0.636 + ], + "angle": 0, + "content": "We begin by presenting our experimental setup, followed by a comparison of our model with state-of-the-art approaches. Next, we conduct an extensive analysis of our model's distinctive features and design choices. Then we demonstrate how we constructed a web browser-compatible game capable of delivering a smooth interactive experience exceeding 100 frames per second (FPS), all derived from a single video source. Finally, we showcase the capabilities of our model in robot simulation through two demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.583, + 0.653 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.894, + 0.78 + ], + "angle": 0, + "content": "Dataset: We evaluate the effectiveness of Video2Game across three distinct scenes in various scenarios, including \"Gardenvase\" [12], an outdoor object-centric scene; the KITTI-360 dataset [37], a large-scale self-driving scene with a sequence that forms a closed loop, suitable for carracing and Temple Run-like games; and finally, an indoor scene from the VR-NeRF [73] dataset, showcasing the potential for robot simulations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.871 + ], + "angle": 0, + "content": "Metrics: To evaluate the quality of the rendered images, we adopt PSNR, SSIM, and LPIPS [82]. For geometry reconstruction, we evaluate with LiDAR point cloud in KITTI-360 dataset. Root mean square deviation (RMSE), mean absolute error (MAE), and outlier rate are applied to measure the disparity existing in estimated geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Our model: For NeRF, we adopt hashgrid encoding [47] and two-layer MLP for each header. For textured mesh," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "4583" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.305, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.089, + 0.484, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.667, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.089, + 0.847, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.182, + 0.305, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.277, + 0.23, + 0.289 + ], + "angle": 0, + "content": "G.T." + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.182, + 0.485, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.277, + 0.451, + 0.291 + ], + "angle": 0, + "content": "MobileNeRF [16]" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.182, + 0.666, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.277, + 0.624, + 0.291 + ], + "angle": 0, + "content": "BakedSDF [76]" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.182, + 0.847, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.278, + 0.772, + 0.289 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.303, + 0.88, + 0.318 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons among mesh models. We compare our mesh rendering method with others in Garden scene [12]." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.342, + 0.29, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.437, + 0.213, + 0.449 + ], + "angle": 0, + "content": "Blender" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.342, + 0.482, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.437, + 0.408, + 0.449 + ], + "angle": 0, + "content": "Unreal" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.457, + 0.442, + 0.471 + ], + "angle": 0, + "content": "Figure 6. Video2Game in Blender and Unreal Engine." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.488, + 0.43, + 0.568 + ], + "angle": 0, + "content": "
MethodOutlier-%↓RMSE↓MAE↓
Instant-NGP [48]22.894.3001.577
Nerfacto [64]50.958.0072.863
Gauss. Spl. [28]91.0811.7688.797
BakedSDF* (offline) [76]43.785.9362.509
Video2Game (Our NeRF)13.233.0281.041
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation on NeRF geometry. Our NeRF renders significantly more accurate depth compared with the baselines. The unit is meter and the outlier threshold is 1.5 meters." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.631, + 0.469, + 0.693 + ], + "angle": 0, + "content": "
MethodVolume RenderingMesh Rastization
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Vanilla NGP27.460.8530.16522.540.7160.350
+ Regularization terms27.520.8610.15722.970.7320.303
+ Monocular cues27.620.8710.13123.350.7650.246
" + }, + { + "type": "table_caption", + "bbox": [ + 0.149, + 0.703, + 0.398, + 0.716 + ], + "angle": 0, + "content": "Table 3. Ablation studies on KITTI-360." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.839 + ], + "angle": 0, + "content": "we conduct marching cubes on the NeRF and post-process it to a fixed precision. We set the texture image size to 4096x4096. For GLSL Shader, we design a light-weight two-layer MLP, which enables efficient real-time rendering. For KITTI-360 (see Sec. 3.1), we divide the whole scene into 16 blocks and create a skydome mesh for the sky." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Baselines: To evaluate the visual and geometry quality of our model, we compare against SOTA approaches in neural rendering and neural reconstruction. Instant-NGP [48] is a NeRF-based method that exploits multi-resolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.481 + ], + "angle": 0, + "content": "hashing encoding. Nerfacto [64] extends the classic NeRF with learnable volumetric sampling and appearance embedding. 3D Gaussian Splatting [28] leverages 3D Gaussians and achieves fast training and rendering. MobileNeRF [16] adopts a hybrid NeRF-mesh representation. It can be baked into a texture map and enable real-time rendering. BakedSDF [76] adopts a volume-surface scene representation. It models view-dependent appearance efficiently by baking spherical Gaussians into the mesh." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.696, + 0.511 + ], + "angle": 0, + "content": "4.2. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Novel view synthesis: Tab. 1 shows the rendering performance and interactive compatibility of our model against the baselines on KITTI-360 [37] and Gardenvase [12]. Our NeRF achieves superior performance when compared to state-of-the-art neural volume render approaches across different scenes. Though [28] performs best in Gardenvase, it fails to handle the sparse camera settings in KITTI-360, where it learns bad 3D orientations of Gaussians. Our baked mesh outperforms other mesh rendering baselines significantly in KITTI-360 and performs similarly in Gardenvase as shown in Fig. 5. Additionally, our pipeline has the highest interactive compatibility among all baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.793 + ], + "angle": 0, + "content": "**Geometry reconstruction:** Our model performs significantly better than the baseline regarding geometry accuracy (see Tab. 2). We provide some qualitative results in Fig. 4, demonstrating that our model can generate high-quality depth maps and surface normals, whereas those produced by the baselines contain more noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation study: To understand the contribution of each component in our model, we begin with the basic Instant-NGP [48] and sequentially reintroduce other components. The results in Tab. 3 show that our regularization and monocular cues improve the quality of both volume rendering in NeRF and mesh rasterization. Additionally, we do observe a decline in rendering performance when convert" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4584" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.396, + 0.107 + ], + "angle": 0, + "content": "ing NeRF into game engine-compatible meshes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.119, + 0.214, + 0.133 + ], + "angle": 0, + "content": "4.3. Video2Game" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.143, + 0.468, + 0.218 + ], + "angle": 0, + "content": "We have shown our approach's effectiveness in rendering quality and reconstruction accuracy across various setups. Next, we demonstrate the construction of a web browser-compatible game enabling player control and interaction with the environment at over 100 FPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.468, + 0.308 + ], + "angle": 0, + "content": "Data preparation: We build our environments based on videos in Gardenvase [12], KITTI-360 [37] and VR-NeRF [73] mentioned in Sec. 4.1, using our proposed approach. The outcomes include executable environments with mesh geometry, materials, and rigid-body physics, all encoded in GLB and texture files." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.468, + 0.506 + ], + "angle": 0, + "content": "Game engine: We build our game based on several key components in our game engine mentioned in Sec. 3.4. By leveraging them, our game generates a highly realistic visual rendering as well as physical interactions (see Fig. 1) and runs smoothly at an interactive rate across various platforms, browsers, and hardware setups (see Tab. 4). As for other game engines (see Fig. 6), in Blender [17] we showcase the compatibility of our exported assets with other game engines. For Unreal [21], we further demonstrate a real-time game demo where a humanoid robot can freely interact within the Gardenvase scene, such as standing on the table and kicking off the central vase. These prove the compatibility of our proposed pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.468, + 0.779 + ], + "angle": 0, + "content": "Interactive game features: Movement in games: Agents can navigate the area freely within the virtual environment where their actions follow real-world physics and are constrained by collision models. Shooting game: For realistic shooting physics, we calculated the rigid-body collision dynamics for both the central vase and the surrounding scene (see Fig. 3), separated using mesh semantic filtering. We used a box collider for the vase and convex polygon colliders for the background. The player shoots footballs with a sphere collider at the vase on the table, causing it to fly off and fall to the ground (see Fig. 1). Temple-Run like game: The agent collects coins while running in the KITTI Loop composed of four streets in KITTI-360. Obstructive chairs on the road can be smashed thanks to pre-computed fracture animations. The agent can also drive and push roadside vehicles existing in the scene forward by crashing into them. This interactivity is achieved through rigid-body dynamics simulation and collision modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Robot simulation: We demonstrate the potential of leveraging Video2Game for robot simulation using the VRNeRF dataset. We reconstruct the scene and segment simulatable rigid-body objects (e.g., the fruit bowl on the table). We show two demos in Fig. 7: a Stretch Robot pushing the bowl off the table and a Fetch Robot performing pick-and-place actions. We employ PyBullet [18] to simulate the underlying physics with the help of corresponding collision" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.082, + 0.892, + 0.129 + ], + "angle": 0, + "content": "
PlatformFPS (hz)CPU-Usage (%)GPU-Usage (%)
Mac M1 ProMac OS, Chrome1023470
Intel Core i9 + NV 4060Windows, Edge240674
AMD 5950 + NV 3090Linux, Chrome1442040
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.14, + 0.892, + 0.194 + ], + "angle": 0, + "content": "Table 4. Runtime Analysis. Our interactive environment can run in real-time across various hardware setup and various platforms. User actions may slightly vary, which could lead to minor variations in runtime." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.211, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.467 + ], + "angle": 0, + "content": "Figure 7. Robot simulation in VRNeRF [73] dataset. We demonstrate the possibility of conducting robot learning in our virtual environments using Stretch Robot [6] and Fetch Robot [5]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.584 + ], + "angle": 0, + "content": "models. Since real-time grasping simulation is challenging, following existing robot simulation frameworks [8, 9, 32], objects near the Fetch gripper are automatically picked up. This demonstrates our model's ability to convert a real-time video stream into a virtual environment, allowing robots to rehearse before acting in the real environment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.598, + 0.618, + 0.613 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.622, + 0.892, + 0.759 + ], + "angle": 0, + "content": "We present a novel approach to converting real-world video footage into playable, real-time, and interactive game environments. Specifically, we combine the potential of NeRF modeling with physics modeling and integrate them into modern game engines. Our approach enables any individual to transform their surroundings into an interactive digital environment, unlocking exciting possibilities for 3D content creation, with promising implications for future advancements in digital game design and robot simulation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.775, + 0.892, + 0.888 + ], + "angle": 0, + "content": "Acknowledgements This project is supported by the Intel AI SRS gift, the IBM IIDAI Grant, the InsperIllinois Innovation Grant, the NCSA Faculty Fellowship, and NSF Awards #2331878, #2340254, and #2312102. We greatly appreciate the NCSA for providing computing resources. We thank Derek Hoiem and Albert Zhai for helpful discussions. We thank Jingkang Wang, Joao Marques and Hanxiao Jiang for proofreading." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4585" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Cannon.js. https://scheppe.github.io/cannon.js/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] GLSL. https://www.khronos.org/opengl/wiki/OpenGL_Shading_Language.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[3] Sketchbook. https://github.com/swift502/Sketchbook.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.201, + 0.451, + 0.214 + ], + "angle": 0, + "content": "[4] WebGL. https://www.khronos.org/webgl/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.216, + 0.468, + 0.256 + ], + "angle": 0, + "content": "[5] Fetch Mobile Manipulator. https://fetchrobotics.borealtech.com/robotics-platforms/fetch-mobile-manipulator/?lang=en.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.258, + 0.466, + 0.283 + ], + "angle": 0, + "content": "[6] Stretch® Research Edition. https://hello-robot.com/product.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.286, + 0.468, + 0.312 + ], + "angle": 0, + "content": "[7] Xatlas. https://github.com/mworchel/xatlaspython.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.314, + 0.468, + 0.34 + ], + "angle": 0, + "content": "[8] RoboTHOR: An Open Simulation-to-Real Embodied AI Platform. CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.342, + 0.468, + 0.368 + ], + "angle": 0, + "content": "[9] ManipulaTHOR: A Framework for Visual Object Manipulation. In CVPR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.37, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[10] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In ICRA, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[11] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[12] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[13] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[14] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[15] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[16] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. CVPR, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[17] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[18] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.891, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.746, + 0.134 + ], + "angle": 0, + "content": "[21] Epic Games. Unreal engine. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.891, + 0.175 + ], + "angle": 0, + "content": "[22] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.803, + 0.189 + ], + "angle": 0, + "content": "[23] Rockstar Games. Grand theft auto v. 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.88, + 0.205 + ], + "angle": 0, + "content": "[24] John K Haas. A history of the unity game engine. 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.891, + 0.26 + ], + "angle": 0, + "content": "[25] Benno Heigl, Reinhard Koch, Marc Pollefeys, Joachim Denzler, and Luc Van Gool. Plenoptic modeling and rendering from image sequences taken by a hand-held camera. In DAGM-Symposium, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.891, + 0.302 + ], + "angle": 0, + "content": "[26] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.891, + 0.357 + ], + "angle": 0, + "content": "[27] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.891, + 0.413 + ], + "angle": 0, + "content": "[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.891, + 0.454 + ], + "angle": 0, + "content": "[29] Pradeep K Khosla and Takeo Kanade. Parameter identification of robot dynamics. In IEEE conference on decision and control, 1985. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.891, + 0.524 + ], + "angle": 0, + "content": "[30] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.891, + 0.566 + ], + "angle": 0, + "content": "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.891, + 0.635 + ], + "angle": 0, + "content": "[32] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.891, + 0.691 + ], + "angle": 0, + "content": "[33] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.891, + 0.733 + ], + "angle": 0, + "content": "[34] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.891, + 0.774 + ], + "angle": 0, + "content": "[35] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.891, + 0.816 + ], + "angle": 0, + "content": "[36] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.891, + 0.857 + ], + "angle": 0, + "content": "[37] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.891, + 0.899 + ], + "angle": 0, + "content": "[38] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.891, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4586" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[39] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[40] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattad, and Shenlong Wang. Urbanir: Large-scale urban scene inverse rendering from a single video, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[41] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[42] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[43] Sivabalan Manivasagam, Shenlong Wang, Kelvin Wong, Wenyuan Zeng, Mikita Sazanovich, Shuhan Tan, Bin Yang, Wei-Chiu Ma, and Raquel Urtasun. Lidarsim: Realistic lidar simulation by leveraging the real world. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[44] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[45] Willi Menapace, Stéphane Lathuiliere, Aliaksandr Siarohin, Christian Theobalt, Sergey Tulyakov, Vladislav Golyanik, and Elisa Ricci. Playable environments: Video manipulation in space and time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3584-3593, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.342, + 0.588 + ], + "angle": 0, + "content": "[47] Thomas Muller. tiny-cuda-nn, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[49] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[52] Cristina Portales, José Luis Lerma, and Santiago Navarro. Augmented reality and photogrammetry: A synergy to visualize physical and virtual city environments. ISPRS Journal of Photogrammetry and Remote Sensing, 65(1):134-142, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[53] Charalambos Poullis, Andrew Gardner, and Paul Debevec. Photogrammetric modeling and image-based rendering for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "rapid virtual environment creation. In Proceedings of ASC, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[54] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[55] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv preprint arXiv:2302.12249, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[58] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[59] Sanghyun Son, Yi-Ling Qiao, Jason Sewall, and Ming C Lin. Differentiable hybrid traffic simulation. TOG, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[60] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[62] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[64] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[65] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[66] ADRIAN SAVARI Thomas, MOHD FAHRUL Hassan, MUSTAFFA Ibrahim, M Nasrull, A Rahman, SZ Sapuan, and F Ahmad. A study on close-range photogrammetry in image based modelling and rendering (imbr) approaches and post-processing analysis. Journal of Engineering Science and Technology, 14(4):1912-1923, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[67] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IROS, 2012. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[68] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[69] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[70] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[71] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.296, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[72] Yuwen Xiong, Jingkang Ma, Wei-Chiu Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.34, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[73] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.425, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[74] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[75] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[76] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. Siggraph, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.599, + 0.469, + 0.639 + ], + "angle": 0, + "content": "[77] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.643, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[78] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. in arXiv, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.7, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[79] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[80] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Transactions on Graphics (TOG), 40(3):1-16, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[81] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[83] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[84] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[85] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[86] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. in arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[87] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4588" + } + ] +] \ No newline at end of file diff --git a/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_origin.pdf b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7171dd099f8dd9a0e4823960b897e077e22911fe --- /dev/null +++ b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/5bc42c40-19c8-4162-b372-460cfcad74f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a0247077b792ac747cf5000c37b4b4c90475ad4a98aa5dd082d8070889feba5 +size 2907610 diff --git a/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/full.md b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..80496809830deca10afcc98e3a7647238961491c --- /dev/null +++ b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/full.md @@ -0,0 +1,332 @@ +# Video2Game: Real-time, Interactive, Realistic and Browser-Compatible Environment from a Single Video + +Hongchi Xia $^{1,2}$ Zhi-Hao Lin $^{1}$ Wei-Chiu Ma $^{3}$ Shenlong Wang $^{1}$ + +1University of Illinois Urbana-Champaign 2Shanghai Jiao Tong University 3Cornell University https://video2game.github.io/ + +![](images/019048486c7a00546ae2e401585e3400601abaa9c6a9f98151db88abca89aa68.jpg) +Figure 1. Video2Game takes an input video of an arbitrary scene and automatically transforms it into a real-time, interactive, realistic and browser-compatible environment. The users can freely explore the environment and interact with the objects in the scene. + +![](images/e707acea833d3e22c0f7c79611dab4950f098c1a44aef591b207eb8f9fb8fd92.jpg) + +![](images/feef627377f78d931bd4a81eac2678d5756bebec9b1db5e28b7fe1f3c0f881cd.jpg) + +![](images/48873a54d0a1a72ec773b77032124140039daa7f62a7ca19aa2f816d9ed228f2.jpg) + +# Abstract + +Creating high-quality and interactive virtual environments, such as games and simulators, often involves complex and costly manual modeling processes. In this paper, we present Video2Game, a novel approach that automatically converts videos of real-world scenes into realistic and interactive game environments. At the heart of our system are three core components: (i) a neural radiance fields (NeRF) module that effectively captures the geometry and visual appearance of the scene; (ii) a mesh module that distills the knowledge from NeRF for faster rendering; and (iii) a physics module that models the interactions and physical dynamics among the objects. By following the carefully designed pipeline, one can construct an interactable and actionable digital replica of the real world. We benchmark our system on both indoor and large-scale outdoor scenes. We show that we can not only produce highly-realistic renderings in real-time, but also build interactive games on top. + +# 1. Introduction + +Crafting a visually compelling and interactive environment is crucial for immersive experiences in various domains, + +such as video games, virtual reality applications, and self-driving simulators. This process, however, is complex and expensive. It demands the skills of experts in the field and the use of professional software development tools [21, 24]. For instance, Grand Theft Auto V [23], known for its intricately detailed environment, was one of the most expensive video games ever developed, with a budget over $265 million primarily for asset creation. Similarly, the development of the CARLA autonomous driving simulator [19] involves a multidisciplinary team of 3D artists, programmers, and engineers to meticulously craft and texture the virtual cityscapes, creating its lifelike environments. + +An appealing alternative to extensive manual modelling is creating environments directly from the real world. For instance, photogrammetry, a technique for constructing digital replicas of objects or scenes from overlapping real-world photographs, has been utilized for environment creation [52, 53]. Success stories also span various games and simulators. However, most use cases are limited to creating object assets and necessitate significant post-processing, such as material creation, texturing, and geometry fixes [66]. People thus turn to neural radiance fields (NeRFs) [46], as it offers a more promising approach to + +![](images/dace2d598dece63e3219702cf837ef9d2f80e770d781e077a3512b9365bfb430.jpg) +Figure 2. Overview of Video2Game: Given multiple posed images from a single video as input, we first construct a large-scale NeRF model that is realistic and possesses high-quality surface geometry. We then transform this NeRF model into a mesh representation with corresponding rigid-body dynamics to enable interactions. We utilize UV-mapped neural texture, which is both expressive and compatible with game engines. Finally, we obtain an interactive virtual environment that virtual actors can interact with, can respond to user control, and deliver high-resolution rendering from novel camera perspectives – all in real-time. + +modeling large scenes. With careful design [13, 22, 39, 48, 61], NeRF is able to render free-viewpoint, photo-realistic images efficiently. However, crafting an interactive environment entails more than just creating a visually high-fidelity digital twin; it also involves building a physically plausible, immersive, real-time and importantly, interactive world tailored to user experiences. Furthermore, we expect such a virtual world to be compatible with real-time interaction interfaces such as common game engines. Despite its promise, the use of NeRF to create interactive environments from real-world videos remains largely unexplored. + +In this paper, we introduce Video2Game, a novel approach to automatically converting a video of a scene into a realistic and interactive virtual environment. Given a video as input, we first construct a NeRF that can effectively capture the geometric and visual information of a large-scale, unbounded) scene. Then we distill the NeRF into a game engine-compatible, neural textured mesh. This significantly improves the rendering efficiency while maintains the overall quality. To model the interactions among the objects, we further decompose the scene into individual actionable entities and equip them with respective physics model. Finally, we import our automatically generated assets into a WebGL-based game engine and create a playable game. The resulting virtual environment is photo-realistic, interactive, and runs in real-time. See Fig. 1 for demonstration. In summary, our key contributions are: + +- A novel 3D modeling algorithm for real-time, free-viewpoint rendering and physical simulation, surpassing state-of-the-art NeRF baking methods with added rigid-body physics for enhanced simulation. +- An automated game-creation framework to transform a scene video into an interactive, realistic environment, compatible with current game engines. + +# 2. Related Works + +Given a single video, we aim to create a real-time, interactive game where the agents (e.g., the character, the car) can navigate and explore the reconstructed digital world, interact with objects in the scene (e.g., collision and manipulate objects), and achieve their respective tasks (e.g., collecting coins, shooting targets). We draw inspirations from multiple areas and combine the best of all. In this section, we will briefly review those closely related areas which forms the foundation of our work. + +Novel view synthesis (NVS): Our work builds upon the success of novel view synthesis [14, 25, 35, 62], which is crucial for our game since it enables the agents to move freely and view the reconstructed world seamlessly from various perspectives. Among all these approaches [26, 60, 68, 85, 86], we exploit neural radiance field (NeRF) [46] as our underlying representation. NeRF has emerged as one of the most promising tools in NVS since its introduction [49-51], and has great performance across a wide range of scenarios [36, 56, 75, 81]. For instance, it can be easily extended to handle various challenging real-world scenarios such as learning from noisy camera poses [38, 70], reflectance modeling for photo-realistic relighting [69, 83], and real-time rendering [16, 39, 55, 65, 76]. In this work, we combine recent advances in NeRF with physics modeling to build an immersive digital world from one single video, moving from passive NVS to our complete solution for embodied world modeling where agents can actively explore and interact with the scene. + +Controllable video generation: Using different control signals to manipulate the output of a visual model has garnered great interest in the community. This has had a profound impact on content creation [57, 58], digital editing [11, 34], and simulation [30, 31, 40]. One could also lever + +age large foundation models to control video content using text [57, 58]. However, they lack fine-grained and real-time control over the generated content. Alternatively, training (conditional) generative models for each scene enables better disentanglement of dynamics (e.g., foreground vs. background) and supports better control. For instance, one can represent a self-driving scene [31] or a Pacman game [30] as latent codes and generate video frames based on control inputs with a neural network. One can also learn to control the players within tennis games [44, 45, 79, 80]. Our work falls under the second line of research, where the model takes user control signals (e.g., a keystroke from the keyboard) as input and responds by rendering a new scene. However, instead of focusing on a specific scene (e.g., tennis games), we have developed a pipeline that allows the creation of a playable environment from a single video of a generic scene. Additionally, we model everything in 3D, which enables us to effectively capture not only view-dependent appearance but also physical interactions among rigid-body equipped objects. Importantly, we adopt a neural representation compatible with graphics engines, enabling users to play the entire game in their browser at an interactive rate. + +Data-driven simulation: Building a realistic simulation environment has been a longstanding challenge. [19, 29, 67, 71]. While it's promising, we come close to mirror the real world only in recent years [10, 15, 42, 43, 59, 74, 75]. The key insight of these work is to build models by leveraging real-world data. Our work closely aligns with this line of research on building high-fidelity simulators from real-world data, with a few key differences. First, existing works mainly focus on offline training and evaluation [10, 43, 74, 75], whereas our system runs at an interactive rate and allows for online, real-time control. Second, some existing works [41, 43, 72, 87] need additional data modality like LiDAR point clouds for geometry reconstruction, but RGB video is all we need. Third, most photo-realistic simulators don't model physical interactions. However, we supports various physics modeling and allows agents to interact with the environment. Last, existing simulators are typically resource-intensive, while our system is lightweight and can be easily accessible in common engines. + +# 3. Video2Game + +Given a sequence of images or a video of a scene, our goal is to construct an interactable and actionable digital twin, upon which we can build real-time, interactive games or realistic (sensor) simulators. Based on the observations that prevalent approaches to constructing digital replica mainly focus on visual appearance and ignore the underlying physical interactions, we carefully design our system such that it can not only produce high-quality rendering across viewpoints, but also support the modeling of physical actions (e.g., navigation, collision, manipulation, etc). At the heart + +![](images/8c2907449af9e07687461d6937d7ce4dfbf40441df49b16200fa95be7f69e0ac.jpg) + +![](images/c116fa750e6d1bde87203919a95252062dd40cec5b5620b283a6dd2278bfaeac.jpg) +Figure 3. Visualization of automatically computed collision geometry: Sphere collider (green), box collider (yellow), convex polygon collider (purple) and trimesh collider (red). + +![](images/d3bb2df923f3aba42d4f67fcad36f98457137003d41783f6e7a092d100a42cba.jpg) + +of our systems is a compositional implicit-explicit 3D representation that is effective and efficient for both sensor and physics simulation. By decomposing the world into individual entities, we can better model and manipulate their physical properties (e.g., specularity, mass, friction), and simulate the outcomes of interactions more effectively. + +We start by introducing a NeRF model that can effectively capture the geometric and visual information of a large-scale, unbounded scene (Sec. 3.1). Next, we present an approach to convert the NeRF into a game-engine compatible mesh with neural texture maps, significantly improving the rendering efficiency while maintaining the quality (Sec. 3.2). To enable physical interactions, we further decompose the scene into individual actionable entities and equip them with respective physics models (Sec. 3.3). Finally, we describe how we integrate our interactive environment into a WebGL-based game engine, allowing users to play and interact with the virtual world in real time on their personal browser. Fig. 2 provides an overview of our proposed framework. + +# 3.1. Large-scale NeRF + +Preliminaries: Instant-NGP [48] is a notable variant of NeRF, which represents the radiance field with a combination of spatial hash-based voxels and neural networks: $\mathbf{c},\sigma = F_{\theta}(\mathbf{x},\mathbf{d};\Phi) = \mathrm{MLP}_{\theta}(\mathsf{It}(\mathbf{x},\Phi),\mathbf{d})$ . Given a 3D point $\mathbf{x}\in \mathbb{R}^3$ and a camera direction $\mathbf{d}\in \mathbb{R}^2$ as input, Instant-NGP first interpolate the point feature $\mathsf{It}(\mathbf{x},\Phi)$ from the adjacent voxel features $\Phi$ . Then the point feature and the camera direction are fed into a light-weight multilayer perception (MLP) to predict the color $\mathbf{c}\in \mathbb{R}^3$ and density $\sigma \in \mathbb{R}^{+}$ . To render the scene appearance, we first cast a ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ from the camera center o through the pixel center in direction d, and sample a set of 3D points $\{\mathbf{x}_i\}$ along the ray. We then query their respective color $\{\mathbf{c}_i\}$ and density $\{\sigma_i\}$ and obtain the color of the pixel through alpha-composition: $\mathbf{C}_{\mathrm{NeRF}}(\mathbf{r}) = \sum_{i}w_{i}\mathbf{c}_{i}$ . Similarly, the + +![](images/7c620637932444c9baccae73c88aa25a7245e2059f01a47cdad11e61f6f030b3.jpg) +Figure 4. Qualitative comparisons among NeRF models. The rendering quality of our base NeRF is superior to baselines, and with leveraging monocular cues, we substantially improve rendered geometry compared to other baselines. This significantly facilitates NeRF baking in subsequent stages. Here we consider depths measured by LiDAR point cloud in KITTI-360 and compute normals based on it. + +expected depth can be computed by: $\mathbf{D}_{\mathrm{NeRF}}(\mathbf{r}) = \sum_{i}w_{i}t_{i}$ Here, $w_{i}$ indicates the blending weight that is derived from the densities $\{\sigma_i\}$ . We refer the readers to [46] for more details. To learn the voxel features $\Phi$ and the MLP weights $\theta$ , we compute the difference between the ground truth color and the rendered color: $\mathcal{L}_{\mathrm{rgb}} = \sum_{\mathbf{r}}\| \mathbf{C}_{\mathrm{GT}}(\mathbf{r}) - \mathbf{C}_{\mathrm{NeRF}}(\mathbf{r})\| _2^2$ . Large-scale NeRF: While Instant-NGP [48] has shown promising results on densely observed and bounded scenes, its performance starts to degrade when extending to sparsely-captured, large-scale, unbounded environments. To mitigate these issues, we propose several enhancements: + +$$ +\mathbf {c}, \sigma , s, \mathbf {n} = F _ {\theta} (\mathbf {x}, \mathbf {d}; \Phi) = \operatorname {M L P} _ {\theta} (\operatorname {I t} (\operatorname {C t} (\mathbf {x}), \Phi), \mathbf {d}). \tag {1} +$$ + +First of all, we exploit the contraction function $\mathrm{Ct}(\mathbf{x})$ [12] to map the unbounded coordinates into a bounded region. In addition to radiance and density, we predict the semantics $s$ and the surface normal $\mathbf{n}$ of the 3D points, guided with 2D priors to better regularize the scene geometry. Furthermore, we divide large-scale scenes into several blocks [63] to capture the fine-grained details. We now describe these enhancements in more details. + +Depth: High-quality geometry is critical for modeling physical interactions. Inspired by MonoSDF [78], we leverage off-the-shelf monocular depth estimators [20, 27] to guide and improve the underlying geometry. We first predict the depth of the scene from rendered RGB images. Then we minimize the discrepancy between the rendered depth and the predicted depth via $\mathcal{L}_{\mathrm{depth}} = \sum_{\mathbf{r}}\|(a\mathbf{D}_{\mathrm{NeRF}}(\mathbf{r}) + b) - \mathbf{D}_{\mathrm{mono}}(\mathbf{r})\|_2^2$ , where $a$ and $b$ are the scale and shift that aligns the two distribution [54]. + +Surface normals: Similar to depth, we encourage the normal estimated from NeRF to be consistent with the normal predicted by the off-the-shelf estimator [20, 27]. The normal of a 3D point $\mathbf{x}_i$ can be either analytically derived from the estimated density $\mathbf{n}_i = (1 - \frac{\nabla_{\mathbf{x}}\sigma_i}{\|\nabla\sigma_i\|})$ , or predicted by the MLP header as in Eq. 1. We could aggregate them via volume render: $\mathbf{N}(\mathbf{r}) = \sum_{i}\mathbf{w}_{i}\mathbf{n}_{i}$ . Em + +pirically we find that adopting both normals and promoting their mutual consistency works the best, since the MLP header offers more flexibility. We thus employ $\mathcal{L}_{\mathrm{normal}} = \|\mathbf{N}_{\mathrm{mlp}}(\mathbf{r}) - \mathbf{N}_{\mathrm{mono}}(\mathbf{r})\|_2^2 + \|\mathbf{N}_{\mathrm{mlp}}(\mathbf{r}) - \mathbf{N}_{\mathrm{density}}(\mathbf{r})\|_2^2$ . + +Semantics: We also predict semantic logits for each sampled 3D points with our MLP. This helps us capture the correlation between semantics and geometry [36, 84]. We render the semantic map with volume rendering $\mathbf{S}_{\mathrm{NeRF}}(\mathbf{r}) = \sum_{i} \mathbf{w}_{i} \mathbf{s}_{i}$ and compute the cross-entropy with that of a 2D segmentation model $\mathcal{L}_{\text{semantics}} = \mathrm{CE}(\mathbf{S}_{\text{mono}}, \mathbf{S}_{\text{NeRF}})$ . + +Regularization: We additionally adopt two regularization terms. To reduce floaters in the scene, for each randomly sampled 3D point $\mathbf{x}$ , we penalize its density by $\mathcal{L}_{\mathrm{sp}} = \sum 1 - \exp (-\alpha \sigma (\mathbf{x}))$ , where $\alpha >0$ [77]. For each sky pixel (which we derived from the semantic MLP), we encourage its depth $\mathbf{D}_{\mathrm{NeRF}}(\mathbf{r}^{\mathrm{sky}})$ to be as far as possible. The loss is defined as: $\mathcal{L}_{\mathrm{sky}} = \sum_{\mathbf{r}^{\mathrm{sky}}} \exp (-\mathbf{D}_{\mathrm{NeRF}}(\mathbf{r}^{\mathrm{sky}}))$ . + +Blocking: Capitalizing on a single Instant-NGP to cover an extraordinarily large scene such as KITTI-360 [37] would often lead to inferior results. We thus adopt a strategy akin to BlockNeRF [63] where we divided the whole scene into numerous blocks and model each region with a separate Instant-NGP. Adjacent regions maintain substantial overlaps to ensure smooth transition. + +Learning: We jointly optimize the voxel feature $\Phi$ and the MLP weights $\theta$ by minimizing the following loss: + +$$ +\mathcal {L} _ {\text {t o t a l}} ^ {\mathrm {N e R F}} = \mathcal {L} _ {\mathrm {r g b}} + \mathcal {L} _ {\text {n o r m a l}} + \mathcal {L} _ {\text {s e m a n t i c s}} + \mathcal {L} _ {\text {d e p t h}} + \mathcal {L} _ {\text {s k y}} + \mathcal {L} _ {\text {s p}} \tag {2} +$$ + +# 3.2. NeRF Baking + +We aim to create a digital replica that users (or agents) can freely explore and act upon in real time. Although our large-scale NeRF effectively renders high-quality images and geometry, its efficiency is limited by the computational costs associated with sampling 3D points. The underlying volume density representation further complicates + +
MethodRepresentationKITTI-360GardenvaseInteractive Compatibility
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Real timeRigid-body physicsScene decomposition
InstantNGP [48]27.460.8530.16525.900.7570.191XXX
Nerfacto [64]Volume23.200.7630.23822.160.5170.283XXX
Video2Game27.620.8710.13126.570.8150.143XXX
Gauss. Spl. [28]Points17.850.6150.42827.500.8580.099✓XX
MobileNeRF [16]19.670.6270.45222.800.5050.365✓XX
BakedSDF* [76]Mesh22.370.7570.30222.680.5140.369✓✓X
Video2Game23.350.7650.24622.810.5080.363✓✓✓
+ +Table 1. Quantitative results on novel view synthesis and interactive compatibility analysis. Video2Game produces better or comparable results across scenes, suggesting the effectiveness of our NeRF and mesh model. The performance improves the most when tackling unbounded, large-scale scenes in KITTI-360. We note that existing NeRFs cannot reach the interactive rate required for real-time games. While point-based rendering significantly improves the speed, it does not support rigid body physics. BakedSDF [76] represents the whole scene with one single mesh, thus does not support object-level interactions. + +the problem. For instance, it's unclear how to define physical interaction with such a representation (e.g., defining collision). The representation is also not compatible with common graphics engines. While recent software, such as the NeRFStudio Blender plugin and LumaAI Unreal add-on, has made some strides, their interaction capabilities and scene geometry quality are still not optimal for real-time user engagement, especially when the scene is large and the observations are relatively sparse. To overcome these challenges, we draw inspiration from recent NeRF meshing advancements and present a novel NeRF baking framework that efficiently transforms our NeRF representation into a game-engine compatible mesh. As we will show in Sec. 4, this conversion greatly enhances rendering efficiency while preserving quality and facilitates physical interactions. + +Mesh representation: Our mesh $\mathcal{M} = (\mathbf{V},\mathbf{F},\mathbf{T})$ comprises vertices $\mathbf{V}\in \mathbb{R}^{|V|\times 3}$ , faces $\mathbf{F}\in \mathbb{N}^{|F|\times 3}$ and a UV neural texture map $\mathbf{T}\in \mathbb{R}^{H\times W\times 6}$ . Following [65], we store the base color in the first three dimension of $\mathbf{T}$ , and encode the specular feature in the rest. The initial mesh topology are obtained by marching cubes in the NeRF density field. We further prune the invisible faces. conduct mesh decimation and edge length regularization. The UV coordinate of each vertex is calculated via xatlas [7]. + +**Rendering:** We leverage differentiable renderers [33] to render our mesh into RGB images $\mathbf{C}_{\mathbb{R}}$ and depth maps $\mathbf{D}_{\mathbb{R}}$ . Specifically, we first rasterize the mesh into screen space and obtain the UV coordinate for each pixel $i$ . Then we sample the corresponding texture feature $\mathbf{T}_i = [\mathbf{B}_i; \mathbf{S}_i]$ and feed it into our customized shader. Finally, the shader computes the sum of the view-independent base color $\mathbf{B}_i$ and the view-dependent MLP $\mathrm{MLP}_{\theta}^{\mathrm{shader}}(\mathbf{S}_i, \mathbf{d}_i)$ : + +$$ +\mathbf {C} _ {\mathrm {R}} = \mathbf {B} _ {i} + \operatorname {M L P} _ {\theta} ^ {\text {s h a d e r}} \left(\mathbf {S} _ {i}, \mathbf {d} _ {i}\right). \tag {3} +$$ + +The MLP is lightweight and can be baked in GLSL. + +Learning: We train the shader MLP $\mathbf{MLP}_{\theta}^{\mathrm{shader}}$ and the neural texture map $\mathbf{T}$ by minimizing the color difference + +between the mesh and the ground truth, and the geometry difference between the mesh and the NeRF model: + +$$ +\mathcal {L} _ {\mathbf {T}, \theta} ^ {\text {m e s h}} = \sum_ {\mathbf {r}} \| \mathbf {C} _ {\mathrm {R}} (\mathbf {r}) - \mathbf {C} _ {\mathrm {G T}} (\mathbf {r}) \| + \| \mathbf {D} _ {\mathrm {R}} (\mathbf {r}) - \mathbf {D} _ {\mathrm {N e R F}} (\mathbf {r}) \|. \tag {4} +$$ + +Anti-aliasing: Common differentiable rasterizers only take the center of each pixel into account. This may lead to aliasing in the learned texture map. To resolve this issue, we randomly perturb the optical center of the camera by 0.5 pixels along each axis at every training step. This ensure all the regions within a pixel get rasterized. + +Relationship to existing work: Our approach is closely related to recent work on NeRF meshing [16, 55, 65, 76], but there exist key differences. While MobileNeRF [16] also adopts an explicit mesh with neural textures, they mainly capitalize on planar primitives. The quality of the reconstructed mesh is thus inferior. BakedSDF [76] offers excellent runtime and rendering quality, but their vertex coloring approach has limited resolution for large scenes. NeRF2Mesh [65] lacks depth distillation and doesn't adopt contraction space for unbounded scenes. They also have a sophisticated multi-stage training and multi-resolution refinement process. Finally, MeRF [55], though efficient, still relies on volume-rendering. + +# 3.3. Representation for Physical Interaction + +Our mesh model facilitates efficient novel-view rendering in real time and allows for basic rigid-body physical interactions. For example, the explicit mesh structure permits an agent to "stand" on the ground. Nevertheless, beyond navigation, an agent should be capable of performing various actions including collision and manipulation. Furthermore, a scene comprises not only the background but also interactable foreground objects, each possessing unique physical properties. For instance, a street-bound car is much heavier than a flower vase. When struck by another object, a car may barely move but the vase may fall and shatter. + +To enhance physical interaction realism, we decompose the scene into discrete, actionable entities, each endowed with specific physical characteristics (e.g., mass, friction). This approach, in conjunction with rigid-body physics, allows for the effective simulation that adheres to physical laws. + +Scene decomposition: Directly editing and decomposing a mesh is extremely difficult due to topology change. Fortunately, neural fields are inherently compositional in 3D. By identifying the objects each spatial region belongs to, we can use neural fields to guide the decomposition of the mesh. Specifically, we sample a 3D point $\mathbf{x}_i$ within each voxel $i$ and determine its semantic category either through the predicted semantic logits $s_i$ or by verifying whether the point is within a specified bounding box. The process is repeated for all voxels to segment the entire scene. Then, for each object, we perform NeRF meshing individually, setting the density of the remaining areas to zero. The intersections between objects are automatically resolved by marching cube. Finally, we initialize the neural texture of these new, individual meshes from the original mesh model. For newly created faces, we employ nearest neighbor inpainting on the neural texture map, which empirically yields satisfactory results. Fig. 1 shows an example where a vase is separated from a table. The middle of the table is original occluded yet we are able to maintain high-quality rendering. + +Physical parameters reasoning: The next step is to equip decomposed individual meshes with various physics-related attributes so that we can effectively model and simulate their physical dynamics. In this work, we focus on rigid body physics, where each entity $i$ is represented by a collision geometry $\text{col}_i$ , mass $m_i$ , and friction parameters $f_i$ . We support four types of collision geometry with different levels of complexity and efficiency: box, sphere, convex polygon, and triangle mesh. Depending on the object and the task of interest, one can select the most suitable collision check for them. For other physical parameters (e.g. mass, friction), one can either set them manually or query large language models (LLMs) for an estimation. + +Physical interactions: Rigid body dynamics, while simple, can support a variety of interactions. With the collision check, an user/agent can easily navigate through the environment while respecting the geometry of the scene. The agents will no longer be stuck in a road or cut through a wall. It also allows the agent to interact with the objects in the scene. For instance, one can push the objects towards the location of interest. The object movement will be determined by its mass and other physical properties such as the friction. We can also manipulate the objects by adopting a magnet grasper, following AI2-Thor [32]. This opens the avenue towards automatic creation of realistic, interactive virtual environment for robot learning. + +# 3.4. Interactive Environment + +We deploy our interactive environment within a real-time, browser-based game engine. We manage the underlying logic and assets using Sketchbook [3], a Game Engine based on Three.js that leverages WebGL [4] for rendering. This combination ensures high efficiency while offering the flexibility and sophistication required for intricate rendering tasks. It also allows us to easily integrate content from different scenes together. We have further extended Sketchbook's capabilities by implementing a GLSL-based Shader [2]. This enables real-time computation of our MLP-based specular Shader during deployment. For physics simulation, we use Cannon.js [1], which assures realism and efficiency in the motion within our interactive environment. It supports not only rigid body dynamics but also more sophisticated modeling techniques. For example, we can precompute the fracturing effect for dynamic objects. Upon experiencing a significant force, these objects are realistically simulated by the real-time physics engine, which handles the interactions between the fractured pieces and the rest of the scene, such as their falling and settling on the ground. Besides browser-based engine, the virtual environments from Video2Game pipeline could be also integrated into both Blender [17] and Unreal engines [21] (see Fig. 6). + +# 4. Experiments + +We begin by presenting our experimental setup, followed by a comparison of our model with state-of-the-art approaches. Next, we conduct an extensive analysis of our model's distinctive features and design choices. Then we demonstrate how we constructed a web browser-compatible game capable of delivering a smooth interactive experience exceeding 100 frames per second (FPS), all derived from a single video source. Finally, we showcase the capabilities of our model in robot simulation through two demonstrations. + +# 4.1. Setup + +Dataset: We evaluate the effectiveness of Video2Game across three distinct scenes in various scenarios, including "Gardenvase" [12], an outdoor object-centric scene; the KITTI-360 dataset [37], a large-scale self-driving scene with a sequence that forms a closed loop, suitable for carracing and Temple Run-like games; and finally, an indoor scene from the VR-NeRF [73] dataset, showcasing the potential for robot simulations. + +Metrics: To evaluate the quality of the rendered images, we adopt PSNR, SSIM, and LPIPS [82]. For geometry reconstruction, we evaluate with LiDAR point cloud in KITTI-360 dataset. Root mean square deviation (RMSE), mean absolute error (MAE), and outlier rate are applied to measure the disparity existing in estimated geometry. + +Our model: For NeRF, we adopt hashgrid encoding [47] and two-layer MLP for each header. For textured mesh, + +![](images/42d4570516ce8748ec52282b39a31d448c8cd3999f71542fbfebec826cfddc3b.jpg) + +![](images/3b79dda2042d703d677b8c2f5ef460727d0c09e83e3170a1b30cfcdd705029ee.jpg) + +![](images/84d04c47b272545d9f18f790d653a2787eb04894274b498b0136ca6fd899ece9.jpg) + +![](images/2a3b9e70cfb3dca7ee2df1412cd3eeb0317628929fb2910cf9f516c269603663.jpg) + +![](images/d7f5c4580edb50b073c204ff1a59d742fdac793b472b94442d7f37573e2478c3.jpg) +G.T. + +![](images/84f432847ef02c1ec883e74bfb4e3b1d4623389efdac16630774e2939aa0be56.jpg) +MobileNeRF [16] + +![](images/2ad283fe30eb91e3a3a7fb127bf31015209c73279f4fb3974eef7d222aa23a64.jpg) +BakedSDF [76] + +![](images/433677b2ba6d51393bdfa26b198e208d4b2e54ee7005136c27b532f034c4a512.jpg) +Ours + +![](images/51a9e47e3da5629b7ba65f1a62cd8e8c49826fd0986a08c3f874e5c7a375e627.jpg) +Figure 5. Qualitative comparisons among mesh models. We compare our mesh rendering method with others in Garden scene [12]. +Blender +Figure 6. Video2Game in Blender and Unreal Engine. + +![](images/bda57ece2b41809c4487435009ec56dcbd74e4263cd10d08271ba9c2bafe7821.jpg) +Unreal + +
MethodOutlier-%↓RMSE↓MAE↓
Instant-NGP [48]22.894.3001.577
Nerfacto [64]50.958.0072.863
Gauss. Spl. [28]91.0811.7688.797
BakedSDF* (offline) [76]43.785.9362.509
Video2Game (Our NeRF)13.233.0281.041
+ +Table 2. Quantitative evaluation on NeRF geometry. Our NeRF renders significantly more accurate depth compared with the baselines. The unit is meter and the outlier threshold is 1.5 meters. + +
MethodVolume RenderingMesh Rastization
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Vanilla NGP27.460.8530.16522.540.7160.350
+ Regularization terms27.520.8610.15722.970.7320.303
+ Monocular cues27.620.8710.13123.350.7650.246
+ +Table 3. Ablation studies on KITTI-360. + +we conduct marching cubes on the NeRF and post-process it to a fixed precision. We set the texture image size to 4096x4096. For GLSL Shader, we design a light-weight two-layer MLP, which enables efficient real-time rendering. For KITTI-360 (see Sec. 3.1), we divide the whole scene into 16 blocks and create a skydome mesh for the sky. + +Baselines: To evaluate the visual and geometry quality of our model, we compare against SOTA approaches in neural rendering and neural reconstruction. Instant-NGP [48] is a NeRF-based method that exploits multi-resolution + +hashing encoding. Nerfacto [64] extends the classic NeRF with learnable volumetric sampling and appearance embedding. 3D Gaussian Splatting [28] leverages 3D Gaussians and achieves fast training and rendering. MobileNeRF [16] adopts a hybrid NeRF-mesh representation. It can be baked into a texture map and enable real-time rendering. BakedSDF [76] adopts a volume-surface scene representation. It models view-dependent appearance efficiently by baking spherical Gaussians into the mesh. + +# 4.2. Experimental results + +Novel view synthesis: Tab. 1 shows the rendering performance and interactive compatibility of our model against the baselines on KITTI-360 [37] and Gardenvase [12]. Our NeRF achieves superior performance when compared to state-of-the-art neural volume render approaches across different scenes. Though [28] performs best in Gardenvase, it fails to handle the sparse camera settings in KITTI-360, where it learns bad 3D orientations of Gaussians. Our baked mesh outperforms other mesh rendering baselines significantly in KITTI-360 and performs similarly in Gardenvase as shown in Fig. 5. Additionally, our pipeline has the highest interactive compatibility among all baselines. + +**Geometry reconstruction:** Our model performs significantly better than the baseline regarding geometry accuracy (see Tab. 2). We provide some qualitative results in Fig. 4, demonstrating that our model can generate high-quality depth maps and surface normals, whereas those produced by the baselines contain more noise. + +Ablation study: To understand the contribution of each component in our model, we begin with the basic Instant-NGP [48] and sequentially reintroduce other components. The results in Tab. 3 show that our regularization and monocular cues improve the quality of both volume rendering in NeRF and mesh rasterization. Additionally, we do observe a decline in rendering performance when convert + +ing NeRF into game engine-compatible meshes. + +# 4.3. Video2Game + +We have shown our approach's effectiveness in rendering quality and reconstruction accuracy across various setups. Next, we demonstrate the construction of a web browser-compatible game enabling player control and interaction with the environment at over 100 FPS. + +Data preparation: We build our environments based on videos in Gardenvase [12], KITTI-360 [37] and VR-NeRF [73] mentioned in Sec. 4.1, using our proposed approach. The outcomes include executable environments with mesh geometry, materials, and rigid-body physics, all encoded in GLB and texture files. + +Game engine: We build our game based on several key components in our game engine mentioned in Sec. 3.4. By leveraging them, our game generates a highly realistic visual rendering as well as physical interactions (see Fig. 1) and runs smoothly at an interactive rate across various platforms, browsers, and hardware setups (see Tab. 4). As for other game engines (see Fig. 6), in Blender [17] we showcase the compatibility of our exported assets with other game engines. For Unreal [21], we further demonstrate a real-time game demo where a humanoid robot can freely interact within the Gardenvase scene, such as standing on the table and kicking off the central vase. These prove the compatibility of our proposed pipeline. + +Interactive game features: Movement in games: Agents can navigate the area freely within the virtual environment where their actions follow real-world physics and are constrained by collision models. Shooting game: For realistic shooting physics, we calculated the rigid-body collision dynamics for both the central vase and the surrounding scene (see Fig. 3), separated using mesh semantic filtering. We used a box collider for the vase and convex polygon colliders for the background. The player shoots footballs with a sphere collider at the vase on the table, causing it to fly off and fall to the ground (see Fig. 1). Temple-Run like game: The agent collects coins while running in the KITTI Loop composed of four streets in KITTI-360. Obstructive chairs on the road can be smashed thanks to pre-computed fracture animations. The agent can also drive and push roadside vehicles existing in the scene forward by crashing into them. This interactivity is achieved through rigid-body dynamics simulation and collision modeling. + +Robot simulation: We demonstrate the potential of leveraging Video2Game for robot simulation using the VRNeRF dataset. We reconstruct the scene and segment simulatable rigid-body objects (e.g., the fruit bowl on the table). We show two demos in Fig. 7: a Stretch Robot pushing the bowl off the table and a Fetch Robot performing pick-and-place actions. We employ PyBullet [18] to simulate the underlying physics with the help of corresponding collision + +
PlatformFPS (hz)CPU-Usage (%)GPU-Usage (%)
Mac M1 ProMac OS, Chrome1023470
Intel Core i9 + NV 4060Windows, Edge240674
AMD 5950 + NV 3090Linux, Chrome1442040
+ +Table 4. Runtime Analysis. Our interactive environment can run in real-time across various hardware setup and various platforms. User actions may slightly vary, which could lead to minor variations in runtime. + +![](images/a3111f1e2dfca3c5576de0b283bf6c3f82b67ffba0c3e2ad118f672b97ddc492.jpg) +Figure 7. Robot simulation in VRNeRF [73] dataset. We demonstrate the possibility of conducting robot learning in our virtual environments using Stretch Robot [6] and Fetch Robot [5]. + +models. Since real-time grasping simulation is challenging, following existing robot simulation frameworks [8, 9, 32], objects near the Fetch gripper are automatically picked up. This demonstrates our model's ability to convert a real-time video stream into a virtual environment, allowing robots to rehearse before acting in the real environment. + +# 5. Conclusion + +We present a novel approach to converting real-world video footage into playable, real-time, and interactive game environments. Specifically, we combine the potential of NeRF modeling with physics modeling and integrate them into modern game engines. Our approach enables any individual to transform their surroundings into an interactive digital environment, unlocking exciting possibilities for 3D content creation, with promising implications for future advancements in digital game design and robot simulation. + +Acknowledgements This project is supported by the Intel AI SRS gift, the IBM IIDAI Grant, the InsperIllinois Innovation Grant, the NCSA Faculty Fellowship, and NSF Awards #2331878, #2340254, and #2312102. We greatly appreciate the NCSA for providing computing resources. We thank Derek Hoiem and Albert Zhai for helpful discussions. We thank Jingkang Wang, Joao Marques and Hanxiao Jiang for proofreading. + +# References + +[1] Cannon.js. https://scheppe.github.io/cannon.js/.6 +[2] GLSL. https://www.khronos.org/opengl/wiki/OpenGL_Shading_Language.6 +[3] Sketchbook. https://github.com/swift502/Sketchbook.6 +[4] WebGL. https://www.khronos.org/webgl/.6 +[5] Fetch Mobile Manipulator. https://fetchrobotics.borealtech.com/robotics-platforms/fetch-mobile-manipulator/?lang=en.8 +[6] Stretch® Research Edition. https://hello-robot.com/product.8 +[7] Xatlas. https://github.com/mworchel/xatlaspython.5 +[8] RoboTHOR: An Open Simulation-to-Real Embodied AI Platform. CVPR, 2020. 8 +[9] ManipulaTHOR: A Framework for Visual Object Manipulation. In CVPR, 2021. 8 +[10] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In ICRA, 2022. 3 +[11] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2 +[12] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 4, 6, 7, 8 +[13] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. ECCV, 2022. 2 +[14] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, 1993. 2 +[15] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3 +[16] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. CVPR, 2023. 2, 5, 7 +[17] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6, 8 +[18] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 8 +[19] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 1, 3 +[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings + +of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4 +[21] Epic Games. Unreal engine. 1, 6, 8 +[22] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[23] Rockstar Games. Grand theft auto v. 2014. 1 +[24] John K Haas. A history of the unity game engine. 2014. 1 +[25] Benno Heigl, Reinhard Koch, Marc Pollefeys, Joachim Denzler, and Luc Van Gool. Plenoptic modeling and rendering from image sequences taken by a hand-held camera. In DAGM-Symposium, 1999. 2 +[26] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In ICCV, 2021. 2 +[27] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4 +[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 5, 7 +[29] Pradeep K Khosla and Takeo Kanade. Parameter identification of robot dynamics. In IEEE conference on decision and control, 1985. 3 +[30] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2, 3 +[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 3 +[32] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai, 2022. 6, 8 +[33] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5 +[34] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv, 2023. 2 +[35] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2 +[36] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv, 2022. 2, 4 +[37] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 4, 6, 7, 8 +[38] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2 + +[39] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 2 +[40] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattad, and Shenlong Wang. Urbanir: Large-scale urban scene inverse rendering from a single video, 2023. 2 +[41] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3 +[42] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives, 2023. 3 +[43] Sivabalan Manivasagam, Shenlong Wang, Kelvin Wong, Wenyuan Zeng, Mikita Sazanovich, Shuhan Tan, Bin Yang, Wei-Chiu Ma, and Raquel Urtasun. Lidarsim: Realistic lidar simulation by leveraging the real world. In CVPR, 2020. 3 +[44] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021. 3 +[45] Willi Menapace, Stéphane Lathuiliere, Aliaksandr Siarohin, Christian Theobalt, Sergey Tulyakov, Vladislav Golyanik, and Elisa Ricci. Playable environments: Video manipulation in space and time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3584-3593, 2022. 3 +[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4 +[47] Thomas Muller. tiny-cuda-nn, 2021. 6 +[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 3, 4, 5, 7 +[49] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2 +[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. +[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv, 2021. 2 +[52] Cristina Portales, José Luis Lerma, and Santiago Navarro. Augmented reality and photogrammetry: A synergy to visualize physical and virtual city environments. ISPRS Journal of Photogrammetry and Remote Sensing, 65(1):134-142, 2010. 1 +[53] Charalambos Poullis, Andrew Gardner, and Paul Debevec. Photogrammetric modeling and image-based rendering for + +rapid virtual environment creation. In Proceedings of ASC, 2004. 1 +[54] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 4 +[55] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv preprint arXiv:2302.12249, 2023. 2, 5 +[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2 +[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv, 2022. 2, 3 +[58] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv, 2023. 2, 3 +[59] Sanghyun Son, Yi-Ling Qiao, Jason Sewall, and Ming C Lin. Differentiable hybrid traffic simulation. TOG, 2022. 3 +[60] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2 +[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[62] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision, 1998. 2 +[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 4 +[64] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv, 2023. 5, 7 +[65] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2, 5 +[66] ADRIAN SAVARI Thomas, MOHD FAHRUL Hassan, MUSTAFFA Ibrahim, M Nasrull, A Rahman, SZ Sapuan, and F Ahmad. A study on close-range photogrammetry in image based modelling and rendering (imbr) approaches and post-processing analysis. Journal of Engineering Science and Technology, 14(4):1912-1923, 2019. 1 +[67] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IROS, 2012. 3 + +[68] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2 +[69] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2 +[70] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv, 2021. 2 +[71] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 2000. 3 +[72] Yuwen Xiong, Jingkang Ma, Wei-Chiu Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. CVPR, 2023. 3 +[73] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 6, 8 +[74] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In CVPR, 2020. 3 +[75] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. CVPR, 2023. 2, 3 +[76] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. Siggraph, 2023. 2, 5, 7 +[77] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields, 2021. 4 +[78] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. in arXiv, 2022. 4 +[79] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph. 3 +[80] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Transactions on Graphics (TOG), 40(3):1-16, 2021. 3 +[81] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2020. 2 +[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 + +[83] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2 +[84] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation, 2021. 4 +[85] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv, 2018. 2 +[86] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. in arXiv, 2022. 2 +[87] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 3 \ No newline at end of file diff --git a/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/images.zip b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6acd5a6ac8669753dd71c7e723e61e018ef3cd4b --- /dev/null +++ b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d836e91a57b4f3db75c591b68e5a9706eb2bee8f9cea6d7ebedcf9d69c3da891 +size 628878 diff --git a/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/layout.json b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e747bd549a095c19d16604c3a29e80b632ceee0 --- /dev/null +++ b/2024/Video2Game_ Real-time Interactive Realistic and Browser-Compatible Environment from a Single Video/layout.json @@ -0,0 +1,9123 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 518, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 518, + 138 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 518, + 138 + ], + "type": "text", + "content": "Video2Game: Real-time, Interactive, Realistic and Browser-Compatible Environment from a Single Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "text", + "content": "Hongchi Xia" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "text", + "content": " Zhi-Hao Lin" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "text", + "content": " Wei-Chiu Ma" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "text", + "content": " Shenlong Wang" + }, + { + "bbox": [ + 128, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 175, + 536, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 175, + 536, + 203 + ], + "spans": [ + { + "bbox": [ + 58, + 175, + 536, + 203 + ], + "type": "text", + "content": "1University of Illinois Urbana-Champaign 2Shanghai Jiao Tong University 3Cornell University https://video2game.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 214, + 184, + 384 + ], + "blocks": [ + { + "bbox": [ + 50, + 214, + 184, + 384 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 184, + 384 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 184, + 384 + ], + "type": "image", + "image_path": "019048486c7a00546ae2e401585e3400601abaa9c6a9f98151db88abca89aa68.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 546, + 412 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 546, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 546, + 412 + ], + "type": "text", + "content": "Figure 1. Video2Game takes an input video of an arbitrary scene and automatically transforms it into a real-time, interactive, realistic and browser-compatible environment. The users can freely explore the environment and interact with the objects in the scene." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 184, + 215, + 317, + 384 + ], + "blocks": [ + { + "bbox": [ + 184, + 215, + 317, + 384 + ], + "lines": [ + { + "bbox": [ + 184, + 215, + 317, + 384 + ], + "spans": [ + { + "bbox": [ + 184, + 215, + 317, + 384 + ], + "type": "image", + "image_path": "e707acea833d3e22c0f7c79611dab4950f098c1a44aef591b207eb8f9fb8fd92.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 215, + 432, + 384 + ], + "blocks": [ + { + "bbox": [ + 320, + 215, + 432, + 384 + ], + "lines": [ + { + "bbox": [ + 320, + 215, + 432, + 384 + ], + "spans": [ + { + "bbox": [ + 320, + 215, + 432, + 384 + ], + "type": "image", + "image_path": "feef627377f78d931bd4a81eac2678d5756bebec9b1db5e28b7fe1f3c0f881cd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 215, + 545, + 384 + ], + "blocks": [ + { + "bbox": [ + 434, + 215, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 434, + 215, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 434, + 215, + 545, + 384 + ], + "type": "image", + "image_path": "48873a54d0a1a72ec773b77032124140039daa7f62a7ca19aa2f816d9ed228f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 427, + 192, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 427, + 192, + 439 + ], + "spans": [ + { + "bbox": [ + 143, + 427, + 192, + 439 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 452, + 289, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 289, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 289, + 656 + ], + "type": "text", + "content": "Creating high-quality and interactive virtual environments, such as games and simulators, often involves complex and costly manual modeling processes. In this paper, we present Video2Game, a novel approach that automatically converts videos of real-world scenes into realistic and interactive game environments. At the heart of our system are three core components: (i) a neural radiance fields (NeRF) module that effectively captures the geometry and visual appearance of the scene; (ii) a mesh module that distills the knowledge from NeRF for faster rendering; and (iii) a physics module that models the interactions and physical dynamics among the objects. By following the carefully designed pipeline, one can construct an interactable and actionable digital replica of the real world. We benchmark our system on both indoor and large-scale outdoor scenes. We show that we can not only produce highly-realistic renderings in real-time, but also build interactive games on top." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Crafting a visually compelling and interactive environment is crucial for immersive experiences in various domains," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 428, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 571 + ], + "type": "text", + "content": "such as video games, virtual reality applications, and self-driving simulators. This process, however, is complex and expensive. It demands the skills of experts in the field and the use of professional software development tools [21, 24]. For instance, Grand Theft Auto V [23], known for its intricately detailed environment, was one of the most expensive video games ever developed, with a budget over $265 million primarily for asset creation. Similarly, the development of the CARLA autonomous driving simulator [19] involves a multidisciplinary team of 3D artists, programmers, and engineers to meticulously craft and texture the virtual cityscapes, creating its lifelike environments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "An appealing alternative to extensive manual modelling is creating environments directly from the real world. For instance, photogrammetry, a technique for constructing digital replicas of objects or scenes from overlapping real-world photographs, has been utilized for environment creation [52, 53]. Success stories also span various games and simulators. However, most use cases are limited to creating object assets and necessitate significant post-processing, such as material creation, texturing, and geometry fixes [66]. People thus turn to neural radiance fields (NeRFs) [46], as it offers a more promising approach to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "4578" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 545, + 216 + ], + "type": "image", + "image_path": "dace2d598dece63e3219702cf837ef9d2f80e770d781e077a3512b9365bfb430.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "text", + "content": "Figure 2. Overview of Video2Game: Given multiple posed images from a single video as input, we first construct a large-scale NeRF model that is realistic and possesses high-quality surface geometry. We then transform this NeRF model into a mesh representation with corresponding rigid-body dynamics to enable interactions. We utilize UV-mapped neural texture, which is both expressive and compatible with game engines. Finally, we obtain an interactive virtual environment that virtual actors can interact with, can respond to user control, and deliver high-resolution rendering from novel camera perspectives – all in real-time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 289, + 430 + ], + "type": "text", + "content": "modeling large scenes. With careful design [13, 22, 39, 48, 61], NeRF is able to render free-viewpoint, photo-realistic images efficiently. However, crafting an interactive environment entails more than just creating a visually high-fidelity digital twin; it also involves building a physically plausible, immersive, real-time and importantly, interactive world tailored to user experiences. Furthermore, we expect such a virtual world to be compatible with real-time interaction interfaces such as common game engines. Despite its promise, the use of NeRF to create interactive environments from real-world videos remains largely unexplored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 626 + ], + "type": "text", + "content": "In this paper, we introduce Video2Game, a novel approach to automatically converting a video of a scene into a realistic and interactive virtual environment. Given a video as input, we first construct a NeRF that can effectively capture the geometric and visual information of a large-scale, unbounded) scene. Then we distill the NeRF into a game engine-compatible, neural textured mesh. This significantly improves the rendering efficiency while maintains the overall quality. To model the interactions among the objects, we further decompose the scene into individual actionable entities and equip them with respective physics model. Finally, we import our automatically generated assets into a WebGL-based game engine and create a playable game. The resulting virtual environment is photo-realistic, interactive, and runs in real-time. See Fig. 1 for demonstration. In summary, our key contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 628, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 628, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 674 + ], + "type": "text", + "content": "- A novel 3D modeling algorithm for real-time, free-viewpoint rendering and physical simulation, surpassing state-of-the-art NeRF baking methods with added rigid-body physics for enhanced simulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 675, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 287, + 712 + ], + "type": "text", + "content": "- An automated game-creation framework to transform a scene video into an interactive, realistic environment, compatible with current game engines." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 297, + 398, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 398, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 398, + 309 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 317, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 424 + ], + "type": "text", + "content": "Given a single video, we aim to create a real-time, interactive game where the agents (e.g., the character, the car) can navigate and explore the reconstructed digital world, interact with objects in the scene (e.g., collision and manipulate objects), and achieve their respective tasks (e.g., collecting coins, shooting targets). We draw inspirations from multiple areas and combine the best of all. In this section, we will briefly review those closely related areas which forms the foundation of our work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 652 + ], + "type": "text", + "content": "Novel view synthesis (NVS): Our work builds upon the success of novel view synthesis [14, 25, 35, 62], which is crucial for our game since it enables the agents to move freely and view the reconstructed world seamlessly from various perspectives. Among all these approaches [26, 60, 68, 85, 86], we exploit neural radiance field (NeRF) [46] as our underlying representation. NeRF has emerged as one of the most promising tools in NVS since its introduction [49-51], and has great performance across a wide range of scenarios [36, 56, 75, 81]. For instance, it can be easily extended to handle various challenging real-world scenarios such as learning from noisy camera poses [38, 70], reflectance modeling for photo-realistic relighting [69, 83], and real-time rendering [16, 39, 55, 65, 76]. In this work, we combine recent advances in NeRF with physics modeling to build an immersive digital world from one single video, moving from passive NVS to our complete solution for embodied world modeling where agents can actively explore and interact with the scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Controllable video generation: Using different control signals to manipulate the output of a visual model has garnered great interest in the community. This has had a profound impact on content creation [57, 58], digital editing [11, 34], and simulation [30, 31, 40]. One could also lever" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4579" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "age large foundation models to control video content using text [57, 58]. However, they lack fine-grained and real-time control over the generated content. Alternatively, training (conditional) generative models for each scene enables better disentanglement of dynamics (e.g., foreground vs. background) and supports better control. For instance, one can represent a self-driving scene [31] or a Pacman game [30] as latent codes and generate video frames based on control inputs with a neural network. One can also learn to control the players within tennis games [44, 45, 79, 80]. Our work falls under the second line of research, where the model takes user control signals (e.g., a keystroke from the keyboard) as input and responds by rendering a new scene. However, instead of focusing on a specific scene (e.g., tennis games), we have developed a pipeline that allows the creation of a playable environment from a single video of a generic scene. Additionally, we model everything in 3D, which enables us to effectively capture not only view-dependent appearance but also physical interactions among rigid-body equipped objects. Importantly, we adopt a neural representation compatible with graphics engines, enabling users to play the entire game in their browser at an interactive rate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 335, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 289, + 563 + ], + "type": "text", + "content": "Data-driven simulation: Building a realistic simulation environment has been a longstanding challenge. [19, 29, 67, 71]. While it's promising, we come close to mirror the real world only in recent years [10, 15, 42, 43, 59, 74, 75]. The key insight of these work is to build models by leveraging real-world data. Our work closely aligns with this line of research on building high-fidelity simulators from real-world data, with a few key differences. First, existing works mainly focus on offline training and evaluation [10, 43, 74, 75], whereas our system runs at an interactive rate and allows for online, real-time control. Second, some existing works [41, 43, 72, 87] need additional data modality like LiDAR point clouds for geometry reconstruction, but RGB video is all we need. Third, most photo-realistic simulators don't model physical interactions. However, we supports various physics modeling and allows agents to interact with the environment. Last, existing simulators are typically resource-intensive, while our system is lightweight and can be easily accessible in common engines." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 573, + 129, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 129, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 129, + 586 + ], + "type": "text", + "content": "3. Video2Game" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 289, + 715 + ], + "type": "text", + "content": "Given a sequence of images or a video of a scene, our goal is to construct an interactable and actionable digital twin, upon which we can build real-time, interactive games or realistic (sensor) simulators. Based on the observations that prevalent approaches to constructing digital replica mainly focus on visual appearance and ignore the underlying physical interactions, we carefully design our system such that it can not only produce high-quality rendering across viewpoints, but also support the modeling of physical actions (e.g., navigation, collision, manipulation, etc). At the heart" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 426, + 133 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 426, + 133 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 426, + 133 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 426, + 133 + ], + "type": "image", + "image_path": "8c2907449af9e07687461d6937d7ce4dfbf40441df49b16200fa95be7f69e0ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 134, + 426, + 196 + ], + "blocks": [ + { + "bbox": [ + 310, + 134, + 426, + 196 + ], + "lines": [ + { + "bbox": [ + 310, + 134, + 426, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 134, + 426, + 196 + ], + "type": "image", + "image_path": "c116fa750e6d1bde87203919a95252062dd40cec5b5620b283a6dd2278bfaeac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 209, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 243 + ], + "type": "text", + "content": "Figure 3. Visualization of automatically computed collision geometry: Sphere collider (green), box collider (yellow), convex polygon collider (purple) and trimesh collider (red)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 542, + 196 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 542, + 196 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 542, + 196 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 542, + 196 + ], + "type": "image", + "image_path": "d3bb2df923f3aba42d4f67fcad36f98457137003d41783f6e7a092d100a42cba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 266, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 338 + ], + "type": "text", + "content": "of our systems is a compositional implicit-explicit 3D representation that is effective and efficient for both sensor and physics simulation. By decomposing the world into individual entities, we can better model and manipulate their physical properties (e.g., specularity, mass, friction), and simulate the outcomes of interactions more effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "type": "text", + "content": "We start by introducing a NeRF model that can effectively capture the geometric and visual information of a large-scale, unbounded scene (Sec. 3.1). Next, we present an approach to convert the NeRF into a game-engine compatible mesh with neural texture maps, significantly improving the rendering efficiency while maintaining the quality (Sec. 3.2). To enable physical interactions, we further decompose the scene into individual actionable entities and equip them with respective physics models (Sec. 3.3). Finally, we describe how we integrate our interactive environment into a WebGL-based game engine, allowing users to play and interact with the virtual world in real time on their personal browser. Fig. 2 provides an overview of our proposed framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 514, + 412, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 412, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 412, + 528 + ], + "type": "text", + "content": "3.1. Large-scale NeRF" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Preliminaries: Instant-NGP [48] is a notable variant of NeRF, which represents the radiance field with a combination of spatial hash-based voxels and neural networks: " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c},\\sigma = F_{\\theta}(\\mathbf{x},\\mathbf{d};\\Phi) = \\mathrm{MLP}_{\\theta}(\\mathsf{It}(\\mathbf{x},\\Phi),\\mathbf{d})" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " . Given a 3D point " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and a camera direction " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{d}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " as input, Instant-NGP first interpolate the point feature " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{It}(\\mathbf{x},\\Phi)" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " from the adjacent voxel features " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " . Then the point feature and the camera direction are fed into a light-weight multilayer perception (MLP) to predict the color " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " . To render the scene appearance, we first cast a ray " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " from the camera center o through the pixel center in direction d, and sample a set of 3D points " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " along the ray. We then query their respective color " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i\\}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i\\}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and obtain the color of the pixel through alpha-composition: " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i}w_{i}\\mathbf{c}_{i}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " . Similarly, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4580" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 542, + 190 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 190 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 190 + ], + "type": "image", + "image_path": "7c620637932444c9baccae73c88aa25a7245e2059f01a47cdad11e61f6f030b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons among NeRF models. The rendering quality of our base NeRF is superior to baselines, and with leveraging monocular cues, we substantially improve rendered geometry compared to other baselines. This significantly facilitates NeRF baking in subsequent stages. Here we consider depths measured by LiDAR point cloud in KITTI-360 and compute normals based on it." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": "expected depth can be computed by: " + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i}w_{i}t_{i}" + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": " Here, " + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": " indicates the blending weight that is derived from the densities " + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i\\}" + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": ". We refer the readers to [46] for more details. To learn the voxel features " + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": " and the MLP weights " + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": ", we compute the difference between the ground truth color and the rendered color: " + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}} = \\sum_{\\mathbf{r}}\\| \\mathbf{C}_{\\mathrm{GT}}(\\mathbf{r}) - \\mathbf{C}_{\\mathrm{NeRF}}(\\mathbf{r})\\| _2^2" + }, + { + "bbox": [ + 46, + 252, + 287, + 385 + ], + "type": "text", + "content": ". Large-scale NeRF: While Instant-NGP [48] has shown promising results on densely observed and bounded scenes, its performance starts to degrade when extending to sparsely-captured, large-scale, unbounded environments. To mitigate these issues, we propose several enhancements:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 397, + 287, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 397, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 397, + 287, + 410 + ], + "type": "interline_equation", + "content": "\\mathbf {c}, \\sigma , s, \\mathbf {n} = F _ {\\theta} (\\mathbf {x}, \\mathbf {d}; \\Phi) = \\operatorname {M L P} _ {\\theta} (\\operatorname {I t} (\\operatorname {C t} (\\mathbf {x}), \\Phi), \\mathbf {d}). \\tag {1}", + "image_path": "509700c5cb048b12757b2e938e73f1fe46062cefab124c2e9dc7ff4b34974b36.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "content": "First of all, we exploit the contraction function " + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\mathrm{Ct}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "content": " [12] to map the unbounded coordinates into a bounded region. In addition to radiance and density, we predict the semantics " + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "content": " and the surface normal " + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "content": " of the 3D points, guided with 2D priors to better regularize the scene geometry. Furthermore, we divide large-scale scenes into several blocks [63] to capture the fine-grained details. We now describe these enhancements in more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "content": "Depth: High-quality geometry is critical for modeling physical interactions. Inspired by MonoSDF [78], we leverage off-the-shelf monocular depth estimators [20, 27] to guide and improve the underlying geometry. We first predict the depth of the scene from rendered RGB images. Then we minimize the discrepancy between the rendered depth and the predicted depth via " + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}} = \\sum_{\\mathbf{r}}\\|(a\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}) + b) - \\mathbf{D}_{\\mathrm{mono}}(\\mathbf{r})\\|_2^2" + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "content": " are the scale and shift that aligns the two distribution [54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": "Surface normals: Similar to depth, we encourage the normal estimated from NeRF to be consistent with the normal predicted by the off-the-shelf estimator [20, 27]. The normal of a 3D point " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " can be either analytically derived from the estimated density " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i = (1 - \\frac{\\nabla_{\\mathbf{x}}\\sigma_i}{\\|\\nabla\\sigma_i\\|})" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ", or predicted by the MLP header as in Eq. 1. We could aggregate them via volume render: " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{N}(\\mathbf{r}) = \\sum_{i}\\mathbf{w}_{i}\\mathbf{n}_{i}" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ". Em" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 253, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 253, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 545, + 302 + ], + "type": "text", + "content": "pirically we find that adopting both normals and promoting their mutual consistency works the best, since the MLP header offers more flexibility. We thus employ " + }, + { + "bbox": [ + 305, + 253, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}} = \\|\\mathbf{N}_{\\mathrm{mlp}}(\\mathbf{r}) - \\mathbf{N}_{\\mathrm{mono}}(\\mathbf{r})\\|_2^2 + \\|\\mathbf{N}_{\\mathrm{mlp}}(\\mathbf{r}) - \\mathbf{N}_{\\mathrm{density}}(\\mathbf{r})\\|_2^2" + }, + { + "bbox": [ + 305, + 253, + 545, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "type": "text", + "content": "Semantics: We also predict semantic logits for each sampled 3D points with our MLP. This helps us capture the correlation between semantics and geometry [36, 84]. We render the semantic map with volume rendering " + }, + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{NeRF}}(\\mathbf{r}) = \\sum_{i} \\mathbf{w}_{i} \\mathbf{s}_{i}" + }, + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "type": "text", + "content": " and compute the cross-entropy with that of a 2D segmentation model " + }, + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{semantics}} = \\mathrm{CE}(\\mathbf{S}_{\\text{mono}}, \\mathbf{S}_{\\text{NeRF}})" + }, + { + "bbox": [ + 305, + 305, + 545, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "content": "Regularization: We additionally adopt two regularization terms. To reduce floaters in the scene, for each randomly sampled 3D point " + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "content": ", we penalize its density by " + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sp}} = \\sum 1 - \\exp (-\\alpha \\sigma (\\mathbf{x}))" + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\alpha >0" + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "content": " [77]. For each sky pixel (which we derived from the semantic MLP), we encourage its depth " + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}^{\\mathrm{sky}})" + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "content": " to be as far as possible. The loss is defined as: " + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sky}} = \\sum_{\\mathbf{r}^{\\mathrm{sky}}} \\exp (-\\mathbf{D}_{\\mathrm{NeRF}}(\\mathbf{r}^{\\mathrm{sky}}))" + }, + { + "bbox": [ + 305, + 382, + 545, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": "Blocking: Capitalizing on a single Instant-NGP to cover an extraordinarily large scene such as KITTI-360 [37] would often lead to inferior results. We thus adopt a strategy akin to BlockNeRF [63] where we divided the whole scene into numerous blocks and model each region with a separate Instant-NGP. Adjacent regions maintain substantial overlaps to ensure smooth transition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "type": "text", + "content": "Learning: We jointly optimize the voxel feature " + }, + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "type": "text", + "content": " and the MLP weights " + }, + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 559, + 545, + 583 + ], + "type": "text", + "content": " by minimizing the following loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 591, + 545, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 591, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 310, + 591, + 545, + 607 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} ^ {\\mathrm {N e R F}} = \\mathcal {L} _ {\\mathrm {r g b}} + \\mathcal {L} _ {\\text {n o r m a l}} + \\mathcal {L} _ {\\text {s e m a n t i c s}} + \\mathcal {L} _ {\\text {d e p t h}} + \\mathcal {L} _ {\\text {s k y}} + \\mathcal {L} _ {\\text {s p}} \\tag {2}", + "image_path": "676fd00bd1a69fa2698eea4a7d3485aa924a3d6ff4073e63b39d286fbfcaf9ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 623, + 392, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 392, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 392, + 635 + ], + "type": "text", + "content": "3.2. NeRF Baking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "content": "We aim to create a digital replica that users (or agents) can freely explore and act upon in real time. Although our large-scale NeRF effectively renders high-quality images and geometry, its efficiency is limited by the computational costs associated with sampling 3D points. The underlying volume density representation further complicates" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4581" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 171 + ], + "type": "table", + "html": "
MethodRepresentationKITTI-360GardenvaseInteractive Compatibility
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Real timeRigid-body physicsScene decomposition
InstantNGP [48]27.460.8530.16525.900.7570.191XXX
Nerfacto [64]Volume23.200.7630.23822.160.5170.283XXX
Video2Game27.620.8710.13126.570.8150.143XXX
Gauss. Spl. [28]Points17.850.6150.42827.500.8580.099✓XX
MobileNeRF [16]19.670.6270.45222.800.5050.365✓XX
BakedSDF* [76]Mesh22.370.7570.30222.680.5140.369✓✓X
Video2Game23.350.7650.24622.810.5080.363✓✓✓
", + "image_path": "bf3689a37085cba93564561b423ce014855cfad58c85d174192779621e6628a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 546, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 546, + 236 + ], + "type": "text", + "content": "Table 1. Quantitative results on novel view synthesis and interactive compatibility analysis. Video2Game produces better or comparable results across scenes, suggesting the effectiveness of our NeRF and mesh model. The performance improves the most when tackling unbounded, large-scale scenes in KITTI-360. We note that existing NeRFs cannot reach the interactive rate required for real-time games. While point-based rendering significantly improves the speed, it does not support rigid body physics. BakedSDF [76] represents the whole scene with one single mesh, thus does not support object-level interactions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 255, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 434 + ], + "type": "text", + "content": "the problem. For instance, it's unclear how to define physical interaction with such a representation (e.g., defining collision). The representation is also not compatible with common graphics engines. While recent software, such as the NeRFStudio Blender plugin and LumaAI Unreal add-on, has made some strides, their interaction capabilities and scene geometry quality are still not optimal for real-time user engagement, especially when the scene is large and the observations are relatively sparse. To overcome these challenges, we draw inspiration from recent NeRF meshing advancements and present a novel NeRF baking framework that efficiently transforms our NeRF representation into a game-engine compatible mesh. As we will show in Sec. 4, this conversion greatly enhances rendering efficiency while preserving quality and facilitates physical interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": "Mesh representation: Our mesh " + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathbf{V},\\mathbf{F},\\mathbf{T})" + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": " comprises vertices " + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\in \\mathbb{R}^{|V|\\times 3}" + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": ", faces " + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{F}\\in \\mathbb{N}^{|F|\\times 3}" + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": " and a UV neural texture map " + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{T}\\in \\mathbb{R}^{H\\times W\\times 6}" + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": ". Following [65], we store the base color in the first three dimension of " + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": ", and encode the specular feature in the rest. The initial mesh topology are obtained by marching cubes in the NeRF density field. We further prune the invisible faces. conduct mesh decimation and edge length regularization. The UV coordinate of each vertex is calculated via xatlas [7]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": "**Rendering:** We leverage differentiable renderers [33] to render our mesh into RGB images " + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbb{R}}" + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": " and depth maps " + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbb{R}}" + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": ". Specifically, we first rasterize the mesh into screen space and obtain the UV coordinate for each pixel " + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": ". Then we sample the corresponding texture feature " + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i = [\\mathbf{B}_i; \\mathbf{S}_i]" + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": " and feed it into our customized shader. Finally, the shader computes the sum of the view-independent base color " + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": " and the view-dependent MLP " + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\theta}^{\\mathrm{shader}}(\\mathbf{S}_i, \\mathbf{d}_i)" + }, + { + "bbox": [ + 46, + 553, + 287, + 650 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 654, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 654, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 287, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathrm {R}} = \\mathbf {B} _ {i} + \\operatorname {M L P} _ {\\theta} ^ {\\text {s h a d e r}} \\left(\\mathbf {S} _ {i}, \\mathbf {d} _ {i}\\right). \\tag {3}", + "image_path": "6408e803537ec01d85e76bca999726363c7c2d796c47f027b4fc1473b906c4f4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 257, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 257, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 257, + 684 + ], + "type": "text", + "content": "The MLP is lightweight and can be baked in GLSL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Learning: We train the shader MLP " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{MLP}_{\\theta}^{\\mathrm{shader}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " and the neural texture map " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " by minimizing the color difference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 255, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 278 + ], + "type": "text", + "content": "between the mesh and the ground truth, and the geometry difference between the mesh and the NeRF model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 285, + 546, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 285, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 285, + 546, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {T}, \\theta} ^ {\\text {m e s h}} = \\sum_ {\\mathbf {r}} \\| \\mathbf {C} _ {\\mathrm {R}} (\\mathbf {r}) - \\mathbf {C} _ {\\mathrm {G T}} (\\mathbf {r}) \\| + \\| \\mathbf {D} _ {\\mathrm {R}} (\\mathbf {r}) - \\mathbf {D} _ {\\mathrm {N e R F}} (\\mathbf {r}) \\|. \\tag {4}", + "image_path": "9220a1b62b1c1e396a410a406878a3d1eda29d1f14322a3b05bc380250864f43.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 319, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 392 + ], + "type": "text", + "content": "Anti-aliasing: Common differentiable rasterizers only take the center of each pixel into account. This may lead to aliasing in the learned texture map. To resolve this issue, we randomly perturb the optical center of the camera by 0.5 pixels along each axis at every training step. This ensure all the regions within a pixel get rasterized." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": "Relationship to existing work: Our approach is closely related to recent work on NeRF meshing [16, 55, 65, 76], but there exist key differences. While MobileNeRF [16] also adopts an explicit mesh with neural textures, they mainly capitalize on planar primitives. The quality of the reconstructed mesh is thus inferior. BakedSDF [76] offers excellent runtime and rendering quality, but their vertex coloring approach has limited resolution for large scenes. NeRF2Mesh [65] lacks depth distillation and doesn't adopt contraction space for unbounded scenes. They also have a sophisticated multi-stage training and multi-resolution refinement process. Finally, MeRF [55], though efficient, still relies on volume-rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 563, + 512, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 512, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 512, + 575 + ], + "type": "text", + "content": "3.3. Representation for Physical Interaction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Our mesh model facilitates efficient novel-view rendering in real time and allows for basic rigid-body physical interactions. For example, the explicit mesh structure permits an agent to \"stand\" on the ground. Nevertheless, beyond navigation, an agent should be capable of performing various actions including collision and manipulation. Furthermore, a scene comprises not only the background but also interactable foreground objects, each possessing unique physical properties. For instance, a street-bound car is much heavier than a flower vase. When struck by another object, a car may barely move but the vase may fall and shatter." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4582" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "To enhance physical interaction realism, we decompose the scene into discrete, actionable entities, each endowed with specific physical characteristics (e.g., mass, friction). This approach, in conjunction with rigid-body physics, allows for the effective simulation that adheres to physical laws." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "text", + "content": "Scene decomposition: Directly editing and decomposing a mesh is extremely difficult due to topology change. Fortunately, neural fields are inherently compositional in 3D. By identifying the objects each spatial region belongs to, we can use neural fields to guide the decomposition of the mesh. Specifically, we sample a 3D point " + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "text", + "content": " within each voxel " + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "text", + "content": " and determine its semantic category either through the predicted semantic logits " + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 142, + 289, + 383 + ], + "type": "text", + "content": " or by verifying whether the point is within a specified bounding box. The process is repeated for all voxels to segment the entire scene. Then, for each object, we perform NeRF meshing individually, setting the density of the remaining areas to zero. The intersections between objects are automatically resolved by marching cube. Finally, we initialize the neural texture of these new, individual meshes from the original mesh model. For newly created faces, we employ nearest neighbor inpainting on the neural texture map, which empirically yields satisfactory results. Fig. 1 shows an example where a vase is separated from a table. The middle of the table is original occluded yet we are able to maintain high-quality rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "text", + "content": "Physical parameters reasoning: The next step is to equip decomposed individual meshes with various physics-related attributes so that we can effectively model and simulate their physical dynamics. In this work, we focus on rigid body physics, where each entity " + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "text", + "content": " is represented by a collision geometry " + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "\\text{col}_i" + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "text", + "content": ", mass " + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "text", + "content": ", and friction parameters " + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 47, + 392, + 289, + 548 + ], + "type": "text", + "content": ". We support four types of collision geometry with different levels of complexity and efficiency: box, sphere, convex polygon, and triangle mesh. Depending on the object and the task of interest, one can select the most suitable collision check for them. For other physical parameters (e.g. mass, friction), one can either set them manually or query large language models (LLMs) for an estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "content": "Physical interactions: Rigid body dynamics, while simple, can support a variety of interactions. With the collision check, an user/agent can easily navigate through the environment while respecting the geometry of the scene. The agents will no longer be stuck in a road or cut through a wall. It also allows the agent to interact with the objects in the scene. For instance, one can push the objects towards the location of interest. The object movement will be determined by its mass and other physical properties such as the friction. We can also manipulate the objects by adopting a magnet grasper, following AI2-Thor [32]. This opens the avenue towards automatic creation of realistic, interactive virtual environment for robot learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 444, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 444, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 444, + 83 + ], + "type": "text", + "content": "3.4. Interactive Environment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 90, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 368 + ], + "type": "text", + "content": "We deploy our interactive environment within a real-time, browser-based game engine. We manage the underlying logic and assets using Sketchbook [3], a Game Engine based on Three.js that leverages WebGL [4] for rendering. This combination ensures high efficiency while offering the flexibility and sophistication required for intricate rendering tasks. It also allows us to easily integrate content from different scenes together. We have further extended Sketchbook's capabilities by implementing a GLSL-based Shader [2]. This enables real-time computation of our MLP-based specular Shader during deployment. For physics simulation, we use Cannon.js [1], which assures realism and efficiency in the motion within our interactive environment. It supports not only rigid body dynamics but also more sophisticated modeling techniques. For example, we can precompute the fracturing effect for dynamic objects. Upon experiencing a significant force, these objects are realistically simulated by the real-time physics engine, which handles the interactions between the fractured pieces and the rest of the scene, such as their falling and settling on the ground. Besides browser-based engine, the virtual environments from Video2Game pipeline could be also integrated into both Blender [17] and Unreal engines [21] (see Fig. 6)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 375, + 388, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 388, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 388, + 388 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 395, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 547, + 503 + ], + "type": "text", + "content": "We begin by presenting our experimental setup, followed by a comparison of our model with state-of-the-art approaches. Next, we conduct an extensive analysis of our model's distinctive features and design choices. Then we demonstrate how we constructed a web browser-compatible game capable of delivering a smooth interactive experience exceeding 100 frames per second (FPS), all derived from a single video source. Finally, we showcase the capabilities of our model in robot simulation through two demonstrations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 504, + 356, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 356, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 356, + 517 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 617 + ], + "type": "text", + "content": "Dataset: We evaluate the effectiveness of Video2Game across three distinct scenes in various scenarios, including \"Gardenvase\" [12], an outdoor object-centric scene; the KITTI-360 dataset [37], a large-scale self-driving scene with a sequence that forms a closed loop, suitable for carracing and Temple Run-like games; and finally, an indoor scene from the VR-NeRF [73] dataset, showcasing the potential for robot simulations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 689 + ], + "type": "text", + "content": "Metrics: To evaluate the quality of the rendered images, we adopt PSNR, SSIM, and LPIPS [82]. For geometry reconstruction, we evaluate with LiDAR point cloud in KITTI-360 dataset. Root mean square deviation (RMSE), mean absolute error (MAE), and outlier rate are applied to measure the disparity existing in estimated geometry." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "Our model: For NeRF, we adopt hashgrid encoding [47] and two-layer MLP for each header. For textured mesh," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "4583" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 186, + 143 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 186, + 143 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 186, + 143 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 186, + 143 + ], + "type": "image", + "image_path": "42d4570516ce8748ec52282b39a31d448c8cd3999f71542fbfebec826cfddc3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 296, + 143 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 296, + 143 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 296, + 143 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 296, + 143 + ], + "type": "image", + "image_path": "3b79dda2042d703d677b8c2f5ef460727d0c09e83e3170a1b30cfcdd705029ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 408, + 143 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 408, + 143 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 408, + 143 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 408, + 143 + ], + "type": "image", + "image_path": "84d04c47b272545d9f18f790d653a2787eb04894274b498b0136ca6fd899ece9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 408, + 70, + 518, + 143 + ], + "blocks": [ + { + "bbox": [ + 408, + 70, + 518, + 143 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 518, + 143 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 518, + 143 + ], + "type": "image", + "image_path": "2a3b9e70cfb3dca7ee2df1412cd3eeb0317628929fb2910cf9f516c269603663.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 75, + 144, + 186, + 217 + ], + "blocks": [ + { + "bbox": [ + 75, + 144, + 186, + 217 + ], + "lines": [ + { + "bbox": [ + 75, + 144, + 186, + 217 + ], + "spans": [ + { + "bbox": [ + 75, + 144, + 186, + 217 + ], + "type": "image", + "image_path": "d7f5c4580edb50b073c204ff1a59d742fdac793b472b94442d7f37573e2478c3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 219, + 140, + 228 + ], + "lines": [ + { + "bbox": [ + 123, + 219, + 140, + 228 + ], + "spans": [ + { + "bbox": [ + 123, + 219, + 140, + 228 + ], + "type": "text", + "content": "G.T." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 144, + 296, + 217 + ], + "blocks": [ + { + "bbox": [ + 186, + 144, + 296, + 217 + ], + "lines": [ + { + "bbox": [ + 186, + 144, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 186, + 144, + 296, + 217 + ], + "type": "image", + "image_path": "84f432847ef02c1ec883e74bfb4e3b1d4623389efdac16630774e2939aa0be56.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 219, + 276, + 230 + ], + "lines": [ + { + "bbox": [ + 208, + 219, + 276, + 230 + ], + "spans": [ + { + "bbox": [ + 208, + 219, + 276, + 230 + ], + "type": "text", + "content": "MobileNeRF [16]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 144, + 407, + 217 + ], + "blocks": [ + { + "bbox": [ + 296, + 144, + 407, + 217 + ], + "lines": [ + { + "bbox": [ + 296, + 144, + 407, + 217 + ], + "spans": [ + { + "bbox": [ + 296, + 144, + 407, + 217 + ], + "type": "image", + "image_path": "2ad283fe30eb91e3a3a7fb127bf31015209c73279f4fb3974eef7d222aa23a64.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 219, + 381, + 230 + ], + "lines": [ + { + "bbox": [ + 322, + 219, + 381, + 230 + ], + "spans": [ + { + "bbox": [ + 322, + 219, + 381, + 230 + ], + "type": "text", + "content": "BakedSDF [76]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 408, + 144, + 518, + 217 + ], + "blocks": [ + { + "bbox": [ + 408, + 144, + 518, + 217 + ], + "lines": [ + { + "bbox": [ + 408, + 144, + 518, + 217 + ], + "spans": [ + { + "bbox": [ + 408, + 144, + 518, + 217 + ], + "type": "image", + "image_path": "433677b2ba6d51393bdfa26b198e208d4b2e54ee7005136c27b532f034c4a512.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 220, + 472, + 228 + ], + "lines": [ + { + "bbox": [ + 452, + 220, + 472, + 228 + ], + "spans": [ + { + "bbox": [ + 452, + 220, + 472, + 228 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 51, + 270, + 177, + 343 + ], + "blocks": [ + { + "bbox": [ + 53, + 239, + 538, + 251 + ], + "lines": [ + { + "bbox": [ + 53, + 239, + 538, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 538, + 251 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons among mesh models. We compare our mesh rendering method with others in Garden scene [12]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 270, + 177, + 343 + ], + "lines": [ + { + "bbox": [ + 51, + 270, + 177, + 343 + ], + "spans": [ + { + "bbox": [ + 51, + 270, + 177, + 343 + ], + "type": "image", + "image_path": "51a9e47e3da5629b7ba65f1a62cd8e8c49826fd0986a08c3f874e5c7a375e627.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 346, + 130, + 355 + ], + "lines": [ + { + "bbox": [ + 99, + 346, + 130, + 355 + ], + "spans": [ + { + "bbox": [ + 99, + 346, + 130, + 355 + ], + "type": "text", + "content": "Blender" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 361, + 270, + 373 + ], + "lines": [ + { + "bbox": [ + 63, + 361, + 270, + 373 + ], + "spans": [ + { + "bbox": [ + 63, + 361, + 270, + 373 + ], + "type": "text", + "content": "Figure 6. Video2Game in Blender and Unreal Engine." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 178, + 270, + 294, + 343 + ], + "blocks": [ + { + "bbox": [ + 178, + 270, + 294, + 343 + ], + "lines": [ + { + "bbox": [ + 178, + 270, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 178, + 270, + 294, + 343 + ], + "type": "image", + "image_path": "bda57ece2b41809c4487435009ec56dcbd74e4263cd10d08271ba9c2bafe7821.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 346, + 249, + 355 + ], + "lines": [ + { + "bbox": [ + 222, + 346, + 249, + 355 + ], + "spans": [ + { + "bbox": [ + 222, + 346, + 249, + 355 + ], + "type": "text", + "content": "Unreal" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 74, + 386, + 263, + 449 + ], + "blocks": [ + { + "bbox": [ + 74, + 386, + 263, + 449 + ], + "lines": [ + { + "bbox": [ + 74, + 386, + 263, + 449 + ], + "spans": [ + { + "bbox": [ + 74, + 386, + 263, + 449 + ], + "type": "table", + "html": "
MethodOutlier-%↓RMSE↓MAE↓
Instant-NGP [48]22.894.3001.577
Nerfacto [64]50.958.0072.863
Gauss. Spl. [28]91.0811.7688.797
BakedSDF* (offline) [76]43.785.9362.509
Video2Game (Our NeRF)13.233.0281.041
", + "image_path": "85980f0e309d0c6d6e637f658b6ac4959b51e89cc7a515dbcd1241c5ab05dd45.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 50, + 499, + 287, + 548 + ], + "blocks": [ + { + "bbox": [ + 46, + 458, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 491 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation on NeRF geometry. Our NeRF renders significantly more accurate depth compared with the baselines. The unit is meter and the outlier threshold is 1.5 meters." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 499, + 287, + 548 + ], + "lines": [ + { + "bbox": [ + 50, + 499, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 50, + 499, + 287, + 548 + ], + "type": "table", + "html": "
MethodVolume RenderingMesh Rastization
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Vanilla NGP27.460.8530.16522.540.7160.350
+ Regularization terms27.520.8610.15722.970.7320.303
+ Monocular cues27.620.8710.13123.350.7650.246
", + "image_path": "5874869482428b7f87ca4d68e32a3a9b6373f061afdece4f0ceb457041bc1ebe.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 91, + 556, + 243, + 567 + ], + "lines": [ + { + "bbox": [ + 91, + 556, + 243, + 567 + ], + "spans": [ + { + "bbox": [ + 91, + 556, + 243, + 567 + ], + "type": "text", + "content": "Table 3. Ablation studies on KITTI-360." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 592, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 664 + ], + "type": "text", + "content": "we conduct marching cubes on the NeRF and post-process it to a fixed precision. We set the texture image size to 4096x4096. For GLSL Shader, we design a light-weight two-layer MLP, which enables efficient real-time rendering. For KITTI-360 (see Sec. 3.1), we divide the whole scene into 16 blocks and create a skydome mesh for the sky." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Baselines: To evaluate the visual and geometry quality of our model, we compare against SOTA approaches in neural rendering and neural reconstruction. Instant-NGP [48] is a NeRF-based method that exploits multi-resolution" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "type": "text", + "content": "hashing encoding. Nerfacto [64] extends the classic NeRF with learnable volumetric sampling and appearance embedding. 3D Gaussian Splatting [28] leverages 3D Gaussians and achieves fast training and rendering. MobileNeRF [16] adopts a hybrid NeRF-mesh representation. It can be baked into a texture map and enable real-time rendering. BakedSDF [76] adopts a volume-surface scene representation. It models view-dependent appearance efficiently by baking spherical Gaussians into the mesh." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 392, + 425, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 425, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 425, + 404 + ], + "type": "text", + "content": "4.2. Experimental results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 411, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 555 + ], + "type": "text", + "content": "Novel view synthesis: Tab. 1 shows the rendering performance and interactive compatibility of our model against the baselines on KITTI-360 [37] and Gardenvase [12]. Our NeRF achieves superior performance when compared to state-of-the-art neural volume render approaches across different scenes. Though [28] performs best in Gardenvase, it fails to handle the sparse camera settings in KITTI-360, where it learns bad 3D orientations of Gaussians. Our baked mesh outperforms other mesh rendering baselines significantly in KITTI-360 and performs similarly in Gardenvase as shown in Fig. 5. Additionally, our pipeline has the highest interactive compatibility among all baselines." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "content": "**Geometry reconstruction:** Our model performs significantly better than the baseline regarding geometry accuracy (see Tab. 2). We provide some qualitative results in Fig. 4, demonstrating that our model can generate high-quality depth maps and surface normals, whereas those produced by the baselines contain more noise." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Ablation study: To understand the contribution of each component in our model, we begin with the basic Instant-NGP [48] and sequentially reintroduce other components. The results in Tab. 3 show that our regularization and monocular cues improve the quality of both volume rendering in NeRF and mesh rasterization. Additionally, we do observe a decline in rendering performance when convert" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4584" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 242, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 242, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 242, + 84 + ], + "type": "text", + "content": "ing NeRF into game engine-compatible meshes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 94, + 130, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 94, + 130, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 94, + 130, + 105 + ], + "type": "text", + "content": "4.3. Video2Game" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 113, + 286, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 113, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 113, + 286, + 172 + ], + "type": "text", + "content": "We have shown our approach's effectiveness in rendering quality and reconstruction accuracy across various setups. Next, we demonstrate the construction of a web browser-compatible game enabling player control and interaction with the environment at over 100 FPS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 173, + 286, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 286, + 243 + ], + "type": "text", + "content": "Data preparation: We build our environments based on videos in Gardenvase [12], KITTI-360 [37] and VR-NeRF [73] mentioned in Sec. 4.1, using our proposed approach. The outcomes include executable environments with mesh geometry, materials, and rigid-body physics, all encoded in GLB and texture files." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 246, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 286, + 400 + ], + "type": "text", + "content": "Game engine: We build our game based on several key components in our game engine mentioned in Sec. 3.4. By leveraging them, our game generates a highly realistic visual rendering as well as physical interactions (see Fig. 1) and runs smoothly at an interactive rate across various platforms, browsers, and hardware setups (see Tab. 4). As for other game engines (see Fig. 6), in Blender [17] we showcase the compatibility of our exported assets with other game engines. For Unreal [21], we further demonstrate a real-time game demo where a humanoid robot can freely interact within the Gardenvase scene, such as standing on the table and kicking off the central vase. These prove the compatibility of our proposed pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 286, + 616 + ], + "type": "text", + "content": "Interactive game features: Movement in games: Agents can navigate the area freely within the virtual environment where their actions follow real-world physics and are constrained by collision models. Shooting game: For realistic shooting physics, we calculated the rigid-body collision dynamics for both the central vase and the surrounding scene (see Fig. 3), separated using mesh semantic filtering. We used a box collider for the vase and convex polygon colliders for the background. The player shoots footballs with a sphere collider at the vase on the table, causing it to fly off and fall to the ground (see Fig. 1). Temple-Run like game: The agent collects coins while running in the KITTI Loop composed of four streets in KITTI-360. Obstructive chairs on the road can be smashed thanks to pre-computed fracture animations. The agent can also drive and push roadside vehicles existing in the scene forward by crashing into them. This interactivity is achieved through rigid-body dynamics simulation and collision modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Robot simulation: We demonstrate the potential of leveraging Video2Game for robot simulation using the VRNeRF dataset. We reconstruct the scene and segment simulatable rigid-body objects (e.g., the fruit bowl on the table). We show two demos in Fig. 7: a Stretch Robot pushing the bowl off the table and a Fetch Robot performing pick-and-place actions. We employ PyBullet [18] to simulate the underlying physics with the help of corresponding collision" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 64, + 545, + 102 + ], + "blocks": [ + { + "bbox": [ + 309, + 64, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 309, + 64, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 309, + 64, + 545, + 102 + ], + "type": "table", + "html": "
PlatformFPS (hz)CPU-Usage (%)GPU-Usage (%)
Mac M1 ProMac OS, Chrome1023470
Intel Core i9 + NV 4060Windows, Edge240674
AMD 5950 + NV 3090Linux, Chrome1442040
", + "image_path": "8564a5cc8c8dd4435bc50abc3bf6ae97e1bec9b809cbe1dab034917184f53ddf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 110, + 545, + 153 + ], + "lines": [ + { + "bbox": [ + 305, + 110, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 110, + 545, + 153 + ], + "type": "text", + "content": "Table 4. Runtime Analysis. Our interactive environment can run in real-time across various hardware setup and various platforms. User actions may slightly vary, which could lead to minor variations in runtime." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 167, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 316, + 167, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 316, + 167, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 167, + 545, + 329 + ], + "type": "image", + "image_path": "a3111f1e2dfca3c5576de0b283bf6c3f82b67ffba0c3e2ad118f672b97ddc492.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 336, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 369 + ], + "type": "text", + "content": "Figure 7. Robot simulation in VRNeRF [73] dataset. We demonstrate the possibility of conducting robot learning in our virtual environments using Stretch Robot [6] and Fetch Robot [5]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 390, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 462 + ], + "type": "text", + "content": "models. Since real-time grasping simulation is challenging, following existing robot simulation frameworks [8, 9, 32], objects near the Fetch gripper are automatically picked up. This demonstrates our model's ability to convert a real-time video stream into a virtual environment, allowing robots to rehearse before acting in the real environment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 473, + 378, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 378, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 378, + 485 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 492, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 492, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 492, + 545, + 601 + ], + "type": "text", + "content": "We present a novel approach to converting real-world video footage into playable, real-time, and interactive game environments. Specifically, we combine the potential of NeRF modeling with physics modeling and integrate them into modern game engines. Our approach enables any individual to transform their surroundings into an interactive digital environment, unlocking exciting possibilities for 3D content creation, with promising implications for future advancements in digital game design and robot simulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 613, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 613, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 305, + 613, + 545, + 703 + ], + "type": "text", + "content": "Acknowledgements This project is supported by the Intel AI SRS gift, the IBM IIDAI Grant, the InsperIllinois Innovation Grant, the NCSA Faculty Fellowship, and NSF Awards #2331878, #2340254, and #2312102. We greatly appreciate the NCSA for providing computing resources. We thank Derek Hoiem and Albert Zhai for helpful discussions. We thank Jingkang Wang, Joao Marques and Hanxiao Jiang for proofreading." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4585" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Cannon.js. https://scheppe.github.io/cannon.js/.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] GLSL. https://www.khronos.org/opengl/wiki/OpenGL_Shading_Language.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 157 + ], + "type": "text", + "content": "[3] Sketchbook. https://github.com/swift502/Sketchbook.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 159, + 276, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 276, + 169 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 276, + 169 + ], + "type": "text", + "content": "[4] WebGL. https://www.khronos.org/webgl/.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 171, + 286, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 202 + ], + "type": "text", + "content": "[5] Fetch Mobile Manipulator. https://fetchrobotics.borealtech.com/robotics-platforms/fetch-mobile-manipulator/?lang=en.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 204, + 285, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 285, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 285, + 224 + ], + "type": "text", + "content": "[6] Stretch® Research Edition. https://hello-robot.com/product.8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 226, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 286, + 247 + ], + "type": "text", + "content": "[7] Xatlas. https://github.com/mworchel/xatlaspython.5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 248, + 286, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 286, + 269 + ], + "type": "text", + "content": "[8] RoboTHOR: An Open Simulation-to-Real Embodied AI Platform. CVPR, 2020. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 270, + 286, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 291 + ], + "type": "text", + "content": "[9] ManipulaTHOR: A Framework for Visual Object Manipulation. In CVPR, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 293, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 346 + ], + "type": "text", + "content": "[10] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In ICRA, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "type": "text", + "content": "[11] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 381, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 287, + 413 + ], + "type": "text", + "content": "[12] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 415, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 435 + ], + "type": "text", + "content": "[13] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. ECCV, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 437, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 479 + ], + "type": "text", + "content": "[14] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, 1993. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 481, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 534 + ], + "type": "text", + "content": "[15] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[16] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. CVPR, 2023. 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 581, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 612 + ], + "type": "text", + "content": "[17] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[18] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 41, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 96, + 456, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 456, + 106 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 456, + 106 + ], + "type": "text", + "content": "[21] Epic Games. Unreal engine. 1, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 138 + ], + "type": "text", + "content": "[22] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 140, + 491, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 491, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 491, + 149 + ], + "type": "text", + "content": "[23] Rockstar Games. Grand theft auto v. 2014. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 152, + 538, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 538, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 538, + 162 + ], + "type": "text", + "content": "[24] John K Haas. A history of the unity game engine. 2014. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 205 + ], + "type": "text", + "content": "[25] Benno Heigl, Reinhard Koch, Marc Pollefeys, Joachim Denzler, and Luc Van Gool. Plenoptic modeling and rendering from image sequences taken by a hand-held camera. In DAGM-Symposium, 1999. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 207, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 239 + ], + "type": "text", + "content": "[26] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In ICCV, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "type": "text", + "content": "[27] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 328, + 545, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 359 + ], + "type": "text", + "content": "[29] Pradeep K Khosla and Takeo Kanade. Parameter identification of robot dynamics. In IEEE conference on decision and control, 1985. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[30] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "text", + "content": "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[32] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai, 2022. 6, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 505, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 547 + ], + "type": "text", + "content": "[33] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "text", + "content": "[34] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv, 2023. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "text", + "content": "[35] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "text", + "content": "[36] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv, 2022. 2, 4" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "text", + "content": "[37] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 4, 6, 7, 8" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[38] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 40 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4586" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[39] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[40] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattad, and Shenlong Wang. Urbanir: Large-scale urban scene inverse rendering from a single video, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[41] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[42] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[43] Sivabalan Manivasagam, Shenlong Wang, Kelvin Wong, Wenyuan Zeng, Mikita Sazanovich, Shuhan Tan, Bin Yang, Wei-Chiu Ma, and Raquel Urtasun. Lidarsim: Realistic lidar simulation by leveraging the real world. In CVPR, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "text", + "content": "[44] Willi Menapace, Stephane Lathuiliere, Sergey Tulyakov, Aliaksandr Siarohin, and Elisa Ricci. Playable video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10070, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "type": "text", + "content": "[45] Willi Menapace, Stéphane Lathuiliere, Aliaksandr Siarohin, Christian Theobalt, Sergey Tulyakov, Vladislav Golyanik, and Elisa Ricci. Playable environments: Video manipulation in space and time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3584-3593, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[46] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 209, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 209, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 209, + 465 + ], + "type": "text", + "content": "[47] Thomas Muller. tiny-cuda-nn, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "text", + "content": "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[49] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[52] Cristina Portales, José Luis Lerma, and Santiago Navarro. Augmented reality and photogrammetry: A synergy to visualize physical and virtual city environments. ISPRS Journal of Photogrammetry and Remote Sensing, 65(1):134-142, 2010. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[53] Charalambos Poullis, Andrew Gardner, and Paul Debevec. Photogrammetric modeling and image-based rendering for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "rapid virtual environment creation. In Proceedings of ASC, 2004. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[54] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[55] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv preprint arXiv:2302.12249, 2023. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[58] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "type": "text", + "content": "[59] Sanghyun Son, Yi-Ling Qiao, Jason Sewall, and Ming C Lin. Differentiable hybrid traffic simulation. TOG, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "text", + "content": "[60] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "text", + "content": "[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "text", + "content": "[62] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision, 1998. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "text", + "content": "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 513, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 578 + ], + "type": "text", + "content": "[64] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv, 2023. 5, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[65] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[66] ADRIAN SAVARI Thomas, MOHD FAHRUL Hassan, MUSTAFFA Ibrahim, M Nasrull, A Rahman, SZ Sapuan, and F Ahmad. A study on close-range photogrammetry in image based modelling and rendering (imbr) approaches and post-processing analysis. Journal of Engineering Science and Technology, 14(4):1912-1923, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[67] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IROS, 2012. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4587" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[68] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "type": "text", + "content": "[69] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "type": "text", + "content": "[70] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 287, + 232 + ], + "type": "text", + "content": "[71] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 2000. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 234, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 234, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 49, + 234, + 287, + 267 + ], + "type": "text", + "content": "[72] Yuwen Xiong, Jingkang Ma, Wei-Chiu Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. CVPR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 269, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 269, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 49, + 269, + 287, + 334 + ], + "type": "text", + "content": "[73] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 336, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 336, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 336, + 287, + 380 + ], + "type": "text", + "content": "[74] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In CVPR, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 382, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 287, + 425 + ], + "type": "text", + "content": "[75] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 428, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 287, + 472 + ], + "type": "text", + "content": "[76] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. Siggraph, 2023. 2, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 474, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 474, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 474, + 287, + 506 + ], + "type": "text", + "content": "[77] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 509, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 509, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 509, + 287, + 552 + ], + "type": "text", + "content": "[78] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. in arXiv, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 554, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 554, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 554, + 287, + 597 + ], + "type": "text", + "content": "[79] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 287, + 643 + ], + "type": "text", + "content": "[80] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Transactions on Graphics (TOG), 40(3):1-16, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 287, + 677 + ], + "type": "text", + "content": "[81] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "text", + "content": "[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 230 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[83] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[84] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[85] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[86] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. in arXiv, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 209, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 230 + ], + "type": "text", + "content": "[87] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4588" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_content_list.json b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..feb68688d042e6eeea925854c49cfee5e4e36a5f --- /dev/null +++ b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_content_list.json @@ -0,0 +1,2011 @@ +[ + { + "type": "text", + "text": "VideoBooth: Diffusion-based Video Generation with Image Prompts", + "text_level": 1, + "bbox": [ + 140, + 92, + 831, + 114 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuming Jiang", + "bbox": [ + 225, + 142, + 346, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianxing Wu", + "bbox": [ + 367, + 142, + 478, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuai Yang $^{3*}$", + "bbox": [ + 501, + 142, + 609, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenyang Si", + "bbox": [ + 629, + 142, + 740, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dahua Lin", + "bbox": [ + 241, + 161, + 336, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Qiao2", + "bbox": [ + 357, + 161, + 433, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Change Loy1", + "bbox": [ + 452, + 161, + 606, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziwei Liu", + "bbox": [ + 627, + 161, + 728, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai AI Laboratory", + "bbox": [ + 202, + 178, + 764, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Wangxuan Institute of Computer Technology, Peking University", + "bbox": [ + 227, + 196, + 741, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://vchitec.github.io/VideoBooth-project/", + "bbox": [ + 251, + 215, + 707, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Prompt", + "bbox": [ + 96, + 237, + 200, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A horse eating grass.", + "bbox": [ + 205, + 238, + 359, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ee099b10ef93093afb05ebce906facfcd700ee993c6e0fc47d1555a48328cc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 252, + 194, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ed8eed6772cc25c9c26de9c36835f5dc0c5925cfdbb8c848306503bbdc2841f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 253, + 532, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Horse grazes in snowy meadow", + "bbox": [ + 542, + 237, + 772, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6cde38fbdb97095e5490a53d35d42928e2c79e795dc4506fda21f9229a3b6d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 252, + 864, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Prompt", + "bbox": [ + 96, + 323, + 200, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Portrait of a dog, looks out the car window.", + "bbox": [ + 210, + 323, + 527, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dog walking in the green farm 4k", + "bbox": [ + 542, + 323, + 785, + 337 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f23977d3089665deeae489cc80e070f51668bb3a40054bead28e0362b8599e7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 338, + 176, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3cd9e2f7fbfeaf24cc2a8cae94e3cedb3683d0e01c4d402fbf68e3eab9e3e0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 339, + 532, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d199bbdf497d60bdd0ee3de9e36de32e016d9272b55ab9010f8b18711019365.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 339, + 864, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Prompt", + "bbox": [ + 96, + 411, + 200, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cat is looking at a laptop.", + "bbox": [ + 210, + 411, + 397, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Close up of cat on top of a vintage chair", + "bbox": [ + 542, + 411, + 834, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59453a6f2e07afe2a46e8d9d6cfb2e184a0f56b4bc90702ed376c36b71448562.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 426, + 192, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/196f7f5ab890e3f4cb7f8268c1b58a366095ce4d2c8e9c7406cd41ba3f71f85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 426, + 532, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a897094ff13e8930d3047e0bb97aea526a38579d275e9a79b5345b5c5f50a83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 426, + 864, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Prompt", + "bbox": [ + 96, + 503, + 200, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elephant drinking water in masai mara reserve, kenya", + "bbox": [ + 210, + 494, + 529, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elephant walk in the yellow grass of savannah", + "bbox": [ + 542, + 494, + 807, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7cbbfe9b71d6ab0c358d6b23d921e69b0eb55d377366f6ca858990ea19a0dd2.jpg", + "image_caption": [ + "Figure 1. Videos synthesized by image prompts. Our VideoBooth generates videos with the subjects specified in the image prompts." + ], + "image_footnote": [], + "bbox": [ + 107, + 521, + 192, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdcaec6ff80b79900beb17dd6efaaa37ef9d3a6e3a72517779d25deecf9f63a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 520, + 532, + 583 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80078ebce751182a79e080233234349871584f74028c9dd32f3747b42b4ac0f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 520, + 864, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 606, + 313, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-driven video generation witnesses rapid progress. However, merely using text prompts is not enough to depict the desired subject appearance that accurately aligns with users' intents, especially for customized content creation. In this paper, we study the task of video generation with image prompts, which provide more accurate and direct content control beyond the text prompts. Specifically, we propose a feed-forward framework VideoBooth, with two dedicated designs: 1) We propose to embed image prompts in a coarse-to-fine manner. Coarse visual embeddings from image encoder provide high-level encodings of image prompts, while fine visual embeddings from the proposed attention injection module provide multi-scale and detailed encoding of image prompts. These two complementary embeddings can faithfully capture the desired appearance. 2) In the attention injection module at fine level, multi-scale", + "bbox": [ + 75, + 638, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image prompts are fed into different cross-frame attention layers as additional keys and values. This extra spatial information refines the details in the first frame and then it is propagated to the remaining frames, which maintains temporal consistency. Extensive experiments demonstrate that VideoBooth achieves state-of-the-art performance in generating customized high-quality videos with subjects specified in image prompts. Notably, VideoBooth is a generalizable framework where a single model works for a wide range of image prompts with only feed-forward passes.", + "bbox": [ + 496, + 608, + 895, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 796, + 632, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image models [16, 17, 20, 26, 31, 37, 39, 43-45, 62, 67-69, 73, 75] have attracted substantial attention. With Stable Diffusion [69], we can now easily generate images using texts. Recently, the focus has been shifted to text-to-video models [6, 18, 30, 38, 46, 48, 60, 76, 80, 84,", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Shuai Yang was in S-Lab, NTU.", + "bbox": [ + 99, + 886, + 421, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "85, 97] to generate videos by taking text descriptions as inputs. However, in some user cases, texts alone are not expressive enough to define the specific appearance of subjects [21, 71]. For example, as shown in Fig. 2, if we want to generate a video clip containing the dog as the third row, we need to use several attributive adjuncts to define the appearance of the dog in text prompts. Even with these extensive attributive adjuncts, models still cannot generate the desired appearance. Defining the appearance of the desired object with texts alone has the following flaws: 1) It is hard to enumerate all the desired attributes, and 2) The model cannot capture all attributes accurately with a long text. Compared to using texts, a more straightforward way to define the appearance is to provide reference images, termed image prompts. The image prompts are complementary to the text prompts and enrich the details that are hard to be depicted by text prompts.", + "bbox": [ + 75, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several attempts to introduce image prompts into text-to-image models, which can be roughly divided into two groups. One is to fine-tune parts of parameters using few-shot reference images [15, 21, 28, 51, 71], which contain the same objects captured under different circumstances. However, the requirement for the number of reference images is demanding as sometimes it is not practical to obtain multiple images of the same object. The other category [11, 42, 52, 83, 87, 88], aiming to address this limitation, proposes to embed image prompts into text-to-image models and the inference is tuning-free. Both of these two types of attempts achieve plausible results in generating images containing objects specified in image prompts.", + "bbox": [ + 75, + 351, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we study a more challenging task, i.e., text-to-video generation with image prompts. The task has two main challenges: 1) Similar to text-to-image generation, the attributes of image prompts should be accurately captured and then reflected in the generated videos; 2) Different from text-to-image generation, we aim for the dynamic movement of the object rather than a static one. Directly adapting these methods to video domain results in mismatched appearance or unnatural degraded movements. To address these challenges, we proposed VideoBooth with elaborately designed coarse-to-fine visual embedding components: 1) Coarse visual embeddings via image encoder: An image encoder is trained to inject the image prompts into text embeddings; 2) Fine visual embeddings via attention injection: The image prompts are mapped to multiscale latent representations to control the generation process through cross-frame attentions of text-to-video models.", + "bbox": [ + 75, + 550, + 472, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, inspired by early attempts [42, 83] in text-to-image models with image prompts, we extract the CLIP image features of the provided image prompts using the pretrained CLIP model [66]. Then the extracted features are mapped into the text embedding space, which are inserted to replace parts of the original text embeddings. The well-", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": " eating snack inside big iron cage at home.", + "bbox": [ + 537, + 90, + 875, + 104 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/65f9d4710b0b0a16c5b504f260541f524833c1526dba63867e813ca73eb9594a.jpg", + "image_caption": [ + "Figure 2. The Use of Image Prompts. We generate three video clips using different types of prompts: simple text prompt, long text prompt, and image prompt. We use the LLaVa model [55] to generate a text prompt describing the appearance of the image prompt. Using text prompts alone cannot fully capture the visual characteristics of the image prompt." + ], + "image_footnote": [], + "bbox": [ + 514, + 106, + 879, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained encoder embeds the coarse appearance information of the given image prompts. However, coarse visual embedding is a universal embedding: 1) It only contains high-level semantic information, and 2) It is shared across all blocks with the same scale. As a result, some visual details are missing in the coarse visual embeddings.", + "bbox": [ + 496, + 402, + 893, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further refine the generated details as well as maintain temporal consistency, different from the highly compacted coarse visual embeddings, multi-scale image prompts are injected into cross-frame attention modules in different layers. The image prompts provide spatial information as well as details with different granularities. On the one hand, keeping spatial information of the image prompts can retain more details. On the other hand, different cross-frame attention modules need detailed information at different scales. Specifically, the latent representations of image prompts are appended as additional keys and values to refine the details in the first generated frame. To propagate the refined first frame to the following frames to maintain temporal consistency, we then use the updated values of the first frame as values for the remaining frames.", + "bbox": [ + 496, + 492, + 893, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We set up a dedicated VideoBooth dataset to support the study of the new task. With each video, we provide an image prompt and a text prompt. Extensive experiments demonstrate the effectiveness of our proposed VideoBooth to generate videos with subjects specified in image prompts. As shown in Fig. 1, videos generated by VideoBooth better keep the visual attributes of image prompts. Besides, our proposed VideoBooth is tuning-free at inference time and videos can be generated with feed-forward passes only. The contributions are summarised as follows:", + "bbox": [ + 496, + 719, + 895, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To our knowledge, we are the first to explore the task of video generation using image prompts without finetuning", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6690", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "at inference time. We propose a dedicated dataset to support the task. Our proposed VideoBooth framework can generate consistent videos containing the subjects specified in image prompts.", + "bbox": [ + 89, + 90, + 467, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new coarse-to-fine visual embedding strategy by image encoder and attention injection, which better captures the characteristics of the image prompts.", + "- We propose a novel attention injection method, using the multi-scale image prompts with spatial information to refine the generated details." + ], + "bbox": [ + 76, + 152, + 467, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 256, + 217, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Video Models take the text descriptions as inputs and generate clips of videos. Early explorations [34, 80] on text-to-video models are based on the idea of VQVAE. Make-A-Video [76] proposes to add temporal attention to the architecture of DALLE2 model [68]. Recently, the emergence of diffusion models [32, 70] boosts research on text-to-video models [1, 23, 30, 33, 60, 82, 97]. Video LDM [6] proposes to train the text-to-video models on Stable Diffusion with temporal attention and 3D convolution introduced to handle the temporal generation. Gen-1 [18] introduces depth maps to handle the temporal consistency of text-to-video models. Some methods [27, 95] resort to training separate modules for synthesizing motions. All of the methods initialize their models with pre-trained text-to-image models. Another paradigm of using text-to-image models is to directly apply Stable Diffusion [69] to few-shot or zero-shot settings. Tune-A-Video [84] adapts the self-attention into cross-frame attention and then finetunes the stable diffusion model on a video clip. Models trained in this way have the capability to transfer motions from original videos. Text2Video-Zero [48] proposes to generate videos by using correlated noise maps to improve consistency. Apart from video generation, diffusion models have been applied to video-to-video generations [8, 9, 19, 24, 36, 41, 54, 58, 64, 65, 74, 81, 89, 90, 94].", + "bbox": [ + 76, + 281, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Customized Content Creation aims at generating images and videos using reference images [15]. For customized text-to-image generation, optimization-based methods [28, 35, 51, 53] are proposed to optimize the weights of the diffusion model. For example, Textual Inversion [21] optimizes the word embeddings, while DreamBooth [71] proposes to finetune the weights of Stable Diffusion as well. Optimization-based methods require several reference images with the same subject to avoid the overfitting of the model, which is demanding in real-world applications. The cost of finetuning hampers the practical usage of these methods. To address these limitations, encoder-based methods [61, 88, 91, 98] are proposed to learn a mapping network to embed the reference images. ELITE [83] proposes to learn a global mapping network and local mapping network to encode the images into word embeddings. Jia", + "bbox": [ + 76, + 659, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "et al. [42] propose to use an additional cross-attention to embed the image features. With the trained encoder, the personalized generation can be achieved in a feed-forward pass. Some recent works [2, 11, 22, 25] combine the encoder-based model and finetuning-based model to improve the performance. BLIP-Diffusion [52] proposes to pretrain a multimodal encoder in a large-scale dataset and then finetune the model on the specific subject for inference. Customized image generation is also applied to place the objects into the user-specified scenes [4, 13, 50, 77, 93]. Also, some efforts [12, 14, 40, 72, 79, 86, 87, 92] have been made to personalized face generation. He et al. [29] propose to improve the performance from the data perspective. Some works [3, 47, 59] focus on composing multiple subjects in one image. Apart from works on image generation, there are some early attempts at personalized video manipulation. Make-A-Protagonist [96] edits an existing video in a personalized way using Stable Diffusion 2.1 to embed the image prompts. The motion of the original video is learned from Tune-A-Video [84]. VideoDreamer [10] proposes to generate personalized videos by generating the first frames using a finetuning-based method and then generating the video clip using the Text2Video-Zero [48]. Different from existing works, our proposed VideoBooth does not need to finetune any weights at the inference time.", + "bbox": [ + 496, + 90, + 890, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VideoBooth", + "text_level": 1, + "bbox": [ + 500, + 481, + 624, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed VideoBooth aims at generating videos from an image prompt $I$ and a text prompt $T$ . The image prompt specifies the appearance of the subject. An overview of our proposed VideoBooth is illustrated in Fig. 3. The image prompt is fed into VideoBooth in two levels. At the coarse level, it is fed into a pretrained CLIP Image encoder to extract visual features. An encoder, composed of several MLP layers, is trained to map visual features into the space of text embeddings. The obtained embedding $f_{I}$ will be inserted into text embedding, which is extracted by feeding text prompt $T$ into CLIP text encoder. To further refine the synthesized details, we propose to inject image prompt $I$ into the cross-frame attention module in the pretrained video diffusion model. Specifically, we append latent representation $x_{I}^{t}$ of image prompt $I$ into the cross-frame attention. In this way, multi-scale visual details with spatial information are involved in the calculation of attention maps so that visual characteristics can be better preserved. Two ways of feeding image prompt corporate with each other in a coarse-to-fine manner. The encoder provides coarse visual embeddings of the image prompt, while the attention injection provides fine visual embeddings.", + "bbox": [ + 496, + 506, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Pretrained Text-to-Video Model", + "text_level": 1, + "bbox": [ + 500, + 847, + 885, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed VideoBooth is developed based on the pretrained text-to-video model [30, 82, 84]. In this section, we", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6691", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0f190a5e7a7e5ed2822bbbc8db28c83d41b178c151a9139dbf0e4f5d2492a40.jpg", + "image_caption": [ + "Figure 3. Overview of VideoBooth. VideoBooth generates videos by taking image prompts $I$ and text prompts $T$ as inputs. The image prompt is fed into the CLIP image encoder, followed by MLP layers. The obtained coarse visual embedding $f_{I}$ is then inserted into the text embeddings. The composed embeddings serve as the input for cross attention. The embedding extracted by the encoder provides a coarse encoding of the visual appearance of the image prompt. To further refine the details in the generated videos, at the fine level, we append the latent representation of the image prompt to the cross-frame attention as additional keys and values. Different cross-frame attention layers receive latent representations with different scales. The multi-scale features with spatial details refine the synthesized details." + ], + "image_footnote": [], + "bbox": [ + 119, + 99, + 851, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "will briefly introduce the framework of text-to-video model. Inflated 2D Conv. To handle video data and capture the temporal correlation, 2D conv in the Stable Diffusion model is inflated to 3D conv. In this way, the U-Net can encode 3D features containing the temporal dimension.", + "bbox": [ + 75, + 446, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Frame Attention Module. Stable Diffusion has a self-attention module, where the features are enhanced by attending to themselves. To improve temporal consistency, the self-attention is modified into cross-frame attention. Specifically, the feature of each frame is enhanced by attending and referencing to the first frame and the previous frame. The cross-frame attention operates on both the spatial domain and temporal domain, thus the temporal consistency of the synthesized frames is improved.", + "bbox": [ + 75, + 522, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Attention Module. Apart from cross-frame attention, a temporal attention module is introduced to further improve temporal consistency. Temporal attention operates on temporal domain and attends to all frames.", + "bbox": [ + 75, + 657, + 468, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Coarse Visual Embeddings via Image Encoder", + "text_level": 1, + "bbox": [ + 76, + 726, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an image prompt $I$ and text prompt $T$ , the generated video is supposed to be consistent with visual elements and textual elements. Inspired by previous attempts at image-based customization methods [42, 83], we propose to encode visual information of image prompts by an image encoder. The image prompt and text prompt complement each other. The image prompt provides visual characteristics of the desired subject in the video, and the text prompt provides other orthogonal information. The extracted visual embeddings are combined with text embeddings as the final", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "embeddings for the cross-attention module. Specifically, the CLIP image encoder is employed to extract the visual features $f_{V}$ of image prompt $I$ . Since the discrepancy exists between the CLIP image and text embeddings, $f_{V}$ is then fed into MLP layers $F(\\cdot)$ to map $f_{V}$ to the spaces of text embeddings. The final embedding $f_{I}$ for the image prompt is obtained as follows:", + "bbox": [ + 496, + 446, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {V} = \\operatorname {C L I P} _ {I} (I), f _ {I} = F \\left(f _ {V}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 564, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for the text prompt $T$ , we feed it into the CLIP text encoder to extract the text embedding $f_{T}$ :", + "bbox": [ + 498, + 593, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {T} = \\left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \\dots , f _ {T} ^ {k}, \\dots \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 633, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{T}^{k}$ is the $k$ -th word embedding in the text prompt.", + "bbox": [ + 498, + 662, + 874, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make the diffusion model generate videos conditioning on both text prompts and image prompts, we need to integrate these two embeddings, i.e., $f_{I}$ and $f_{T}$ . The idea is to replace the word embedding of the target subject with $f_{I}$ . Mathematically, $f_{I}$ and $f_{t}$ are fused to obtain the final text condition $c_{t}$ as follows:", + "bbox": [ + 496, + 679, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {T} = \\left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \\dots , f _ {T} ^ {k - 1}, f _ {I}, f _ {T} ^ {k + n}, \\dots \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 779, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k$ is the token index of the target subject in the text embedding, and $n$ is the length of the text tokens for the target subject. For example, an image of a papillon dog is provided as an image prompt. To fuse the information from the image prompt and the text prompt \"Papillon dog celebrates birthday with gifts\", the word embeddings of the", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d2d02e2df2dcfdd3ed7d86c32c188766cc9aa2d352c4ea8e5f1297754d070d3e.jpg", + "image_caption": [ + "Figure 4. Fine Visual Embedding Refinement. We propose to inject the latent representation of image prompt (here we use the image for illustration purpose) directly into the cross-frame attention module. We use the keys and values from the image prompt to update the values of the first frame firstly. Then, the updated values of the first frame are used to update the remaining frames. Injecting the image prompt in the cross-frame attention helps to transfer the detailed visual characteristics of the image prompts to the synthesized frames. We perform the refinement in different cross-attention layers with different scales." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 460, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"papillon dog\" will be replaced with $f_{I}$ before they are fed into the cross-attention module of the diffusion models.", + "bbox": [ + 76, + 412, + 468, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training of the coarse stage, we fix the parameters of the CLIP image encoder, and train the MLP layers. To make the diffusion model accommodate with the composed text embeddings $c_{T}$ , we also finetune K and V projections (linear layers to map the input feature to the corresponding keys and values) in the cross-attention module.", + "bbox": [ + 75, + 444, + 468, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fine Visual Embeddings via Attention Injection", + "text_level": 1, + "bbox": [ + 76, + 544, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The well-trained image encoder embeds the coarse visual embeddings for image prompts and thus the synthesized videos contain the subjects specified in image prompts. However, the image encoder projects the image prompt into a flattened high-level representation, resulting in the loss of its detailed visual cues. Thus, some detailed visual characteristics in the image prompts may not be well preserved. To address this problem, a more effective way to preserve these details is to provide the model with the image prompts with spatial resolutions.", + "bbox": [ + 75, + 568, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further refine the synthesized details, we propose to inject image prompts into the cross-frame attention of the text-to-video models. By injecting image prompts into the cross-frame attention, the image prompts are involved in the updates of the synthesized frames so that the model can directly borrow some visual cues from image prompts.", + "bbox": [ + 75, + 719, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since text-to-video diffusion models operate in the latent space, we first feed the image prompt into the VAE of Stable Diffusion and get its latent representation $x_{I}$ . Moreover, since the sampling of the videos starts from the noise map, the latent in the intermediate timesteps contains the noises. If we append the clean latent $x_{I}$ of the image prompt", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to the cross-frame attention, the domain discrepancy exists. Therefore, we follow the diffusion forward process to add corresponding noises to $x_{I}$ :", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} ^ {I} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} ^ {I} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 141, + 890, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{\\alpha}$ is a hyperparameter determined by the denoising schedule and $\\epsilon \\sim N(0, I)$ .", + "bbox": [ + 496, + 166, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cross-frame attention is used to improve the temporal consistency of the generated frames. For each frame, the key and value are the concatenation of the features of the first frame and the previous frame. Here, we introduce the image prompts as the additional keys and values for the frames. As shown in Fig. 4, we propose to update the values of the first frame firstly using the keys and values of the image prompts and the frame itself. Mathematically, the operation can be expressed as follows:", + "bbox": [ + 496, + 196, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {0} ^ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {0} K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 337, + 890, + 372 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = \\left[ K _ {I}, K _ {0} \\right], V = \\left[ V _ {I}, V _ {0} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 375, + 803, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K_{I}$ and $V_{I}$ are the keys and values obtained from the image prompts. The query, key, and value of the first frame are denoted $Q_{0}, K_{0}$ , and $V_{0}$ , respectively. It should be noted that we use a separately trained $\\mathrm{K}$ and $\\mathrm{V}$ projection for latent representations $x_{t}^{I}$ of image prompts because the image prompts have clean backgrounds, which are different from other frames. The parameters of the newly added $\\mathrm{K}$ and $\\mathrm{V}$ projections are initialized by original $\\mathrm{K}$ and $\\mathrm{V}$ projections.", + "bbox": [ + 496, + 397, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the updated first frame is used to refine the remaining frames. When updating the remaining frames, the keys used for calculating the attention maps are the original keys, while the values are the updated ones. The update is expressed as follows:", + "bbox": [ + 496, + 518, + 890, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} ^ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 598, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = [ K _ {0}, K _ {i - 1} ], V = [ V _ {0} ^ {\\text {n e w}}, V _ {i - 1} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 635, + 820, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To sum up, in the attention injection, we update the values of the first frame using the image prompts first, and then use the updated first frame to update the other frames. In this way, the visual cues from the image prompts can be consistently propagated to all the frames.", + "bbox": [ + 496, + 657, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It should be noted that the diffusion model has multiple cross-frame attention layers with different scales. To inject multi-scale visual cues for better detail refinement, in different cross-frame attention layers, we feed latent representations of the image prompts with corresponding resolutions, which are obtained from different stages of the U-Net.", + "bbox": [ + 496, + 734, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Coarse-to-Fine Training Strategy", + "text_level": 1, + "bbox": [ + 500, + 832, + 790, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The visual details of the image prompts are embedded into the final synthesized results in two stages: coarse visual embeddings using an image encoder and fine visual embedding", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by attention injection. We propose to train these two modules in a coarse-to-fine manner. In other words, we train the coarse image encoder and tune the parameters in the cross-attention first. After the model has the capability of generating videos containing the subjects specified in image prompts, we then train the attention injection module to embed image prompts into cross-frame attention layers. As we will show in the ablation study (Sec. 5.5), if these two modules are trained together, the fine attention injection module leaks the strong visual cues and the coarse encoder learns meaningless representations. As a result, in sampling phase, the image encoder for the coarse visual embedding cannot provide the coarse information and then the fine attention module cannot refine the details. Therefore, it is necessary to train VideoBooth in a coarse-to-fine manner.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. VideoBooth Dataset", + "text_level": 1, + "bbox": [ + 76, + 349, + 269, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We establish the VideoBooth dataset to support the task of video generation using image prompts. We start from the WebVid dataset [5], a well-known open-source dataset for text-to-video generation. In the WebVid dataset, there is a text prompt with each video. In this paper, we study the task of generating a video clip from one text prompt and one image prompt. Hence, in addition to the original text prompt, we need to provide an image prompt for each video. We propose to segment the subjects from the first frame of the video using the Grounded-SAM (Grounded Segment Anything) [49, 57], and the segmented subjects are image prompts. The Grounded-SAM receives word prompts as inputs and generates segmentation masks for the target subjects specified in word prompts. To obtain the word prompt for the input to Grounded-SAM, we use the spaCy library to parse the noun chunks from the original text prompts, which are used as the word prompts. After the segmentation, we perform data filtering to ensure the data quality. We filter out small objects and large objects (those are almost the same size as the original video) according to the ratio of the object to the whole video. Also, since we focus on generating video clips containing moving objects, we further filter the videos containing moving objects. The keywords we used for filtering are dog, cat, bear, car, panda, tiger, horse, elephant, and lion. In the current version, we have processed 2.5M subset of the WebVid dataset. After data filtering, we have 48,724 video data pairs for training. We will process the full set of the WebVid dataset and include the filtered data in our VideoBooth dataset.", + "bbox": [ + 75, + 380, + 472, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance, we also set up a test benchmark. The test benchmark consists of 650 test pairs. For each pair, an image prompt and a text prompt are provided. The test pairs are selected from the rest of the WebVid-10M dataset, which does not overlap with the training set.", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e3ee0f8053c03bbf45e0f2f14e1a73f201cf623630eee432779676cab10b3425.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons. VideoBooth achieves the best image alignment and comparable text alignment performance." + ], + "table_footnote": [], + "table_body": "
MethodCLIP-Text ↑CLIP-Image ↑DINO ↑
Textual Inversion [21]29.974969.799545.3143
DreamBooth [71]30.687771.207852.9661
ELITE [83]30.088173.751858.9522
VideoBooth (Ours)30.096774.797165.0979
", + "bbox": [ + 501, + 122, + 895, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 213, + 632, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison Methods", + "text_level": 1, + "bbox": [ + 500, + 238, + 702, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Textual Inversion [21] is a method for customized text-to-image generation. The appearance of the target subjects is embedded into the text embeddings. Concretely, a text token $S^{*}$ is optimized to represent the subject. We adapt it to the task of text-to-video generation by replacing the image model with the video model.", + "bbox": [ + 496, + 261, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DreamBooth [71] is also proposed for customized text-to-image generation. It injects the target subject into the text tokens as well as model weights. During the training, both the model weights and the special token $S*$ are optimized.", + "bbox": [ + 496, + 352, + 890, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ELITE [83] is an encoder-based method for personalized generation. An encoder is trained to embed the images into the text embeddings. Local mapping and global mapping are employed to transform the CLIP embedding of image prompts into the features, which are injected into the cross-attention module. We adapt and retrain the method using the same pretrained video model we use.", + "bbox": [ + 496, + 412, + 890, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 527, + 683, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use three metrics to evaluate the performance [83]. To measure the alignment of the generated videos and given text prompts, we use the CLIP-Text metric. The metric is calculated using the cosine similarity of the CLIP text embeddings of text prompts and CLIP image embeddings of the generated frames. For each video, the value is obtained by averaging values of the all frames. As for the evaluation of the similarity between the given image prompts and the generated videos, we adopt two metrics: CLIP-Image and DINO [7, 63]. The CLIP-Image metric is calculated by the cosine similarity between the CLIP image embedding of image prompts and generated frames. Since the CLIP model is trained to align image embeddings and text embeddings, we follow the practice in previous methods [13, 71] and use the DINO similarity as another indicator. DINO is trained to differentiate the differences between objects of the same classes. We use the ViT-S/16 model to extract the features of the image prompts and generated frames. The final score is obtained by averaging over all frames.", + "bbox": [ + 496, + 550, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 500, + 847, + 740, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report quantitative results in Table 1. As shown in Table 1, our proposed VideoBooth achieves state-of-the-art", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6694", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/101aaef392f59fa34a50a1f1252c07b23ed9633b77955d6aaffd503e198d7499.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison. VideoBooth effectively preserves the fidelity of image prompts and achieves better visual quality." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 352, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfb987e8911e53d525e68e5645bb2d6a1ecd4a2ee49a6a70241168f4e8008560.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 95, + 612, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47f0a6bf7d23dd377306176abf7f243d2c6d4c70f3446d83ef095402c70cb6bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 95, + 870, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "image alignment performance compared to baseline methods. As for the alignment with the text prompts, our proposed VideoBooth has comparable performance with the baseline models. It should be noted that the CLIP-Text score of DreamBooth is significantly higher than the other methods. The reason lies in that the optimized token $S*$ in DreamBooth is inserted into the text embeddings, rather than replacing the original word embeddings like other methods. This would result in the generated videos of DreamBooth being highly related to the text prompts but having no correlation to the image prompts in some cases. We also conducted a user study, in which 25 users participated. Each user is presented with twelve groups of videos, and each group contains four videos generated by four methods. For each group, users are asked to make three choices: 1) which one has the best image alignment? 2) which one has the best text alignment? 3) which one has the best overall quality. Figure 6 summarizes the results. Our results are preferred by most users in all three dimensions.", + "bbox": [ + 75, + 420, + 473, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 723, + 305, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show three visual comparisons on the generated video frames of our proposed VideoBooth and baseline methods in Fig. 5. In the first example, the model is supposed to generate videos containing the dog specified in the image laying on the ground. Textual Inversion cannot correctly embed the appearance of the dog and results in generating another totally different dog. DreamBooth and ELITE can embed the coarse appearance of the dog but the generated details vary from the image prompt. Our proposed VideoBooth successfully embeds the details of the image prompts into", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0ce31bd0b3b4d54c441ad43f4c9d246a8585a9faa92bc7f203f1d5f5c47c58d.jpg", + "image_caption": [ + "Figure 6. User Study. Our proposed VideoBooth achieves the highest user preference ratios on all three dimensions." + ], + "image_footnote": [], + "bbox": [ + 513, + 420, + 877, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the synthesized videos. In the second example, the condition is to generate a cat playing with the yellow ball. All the methods can generate a yellow ball and a cat but only our proposed VideoBooth can accurately generate the cat having the appearance from the image prompt. As for the last example, Textual Inversion fails to generate a car. DreamBooth generates a distorted car in the bush. ELITE model can generate a car in the bush, but the color differs from the image prompt. By contrast, our model can generate videos having the same car in the bush.", + "bbox": [ + 496, + 580, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 739, + 653, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the proposed components, we perform three ablation studies. Due to the computational resources, we train these models on the subset of our training set. The quantitative metrics are shown in Table 2.", + "bbox": [ + 496, + 763, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Only Coarse Embeddings. This ablation model injects the image prompts with only coarse embeddings via Image Encoder. In the example shown in Fig. 7(a), the ablation model only encodes the coarse appearance of the image prompts. The pattern in the legs of the synthesized dog is different", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6695", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6943cb25f233daa454e7ce53cb62d883dd68912839fffc7f26cee3381f5d9595.jpg", + "image_caption": [ + "Dog as a chef at the table" + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 153, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81c88a84ce53ecd6a5b863733caea3d8c80eccd1d46b9a148babd67f516327a4.jpg", + "image_caption": [ + "(a) Only Coarse Embeddings with Image Encoder" + ], + "image_footnote": [], + "bbox": [ + 83, + 152, + 464, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff789fb811d4423174fefae79ee89ca5b0148ed07b40b0cc25eba116cd8721dd.jpg", + "image_caption": [ + "(b) Only Fine Embeddings with Attention Injection" + ], + "image_footnote": [], + "bbox": [ + 83, + 226, + 464, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d0fcc35c6120ec2947a6b9e368dd970ea376de83ca1ccd765270919345403a0.jpg", + "image_caption": [ + "(c) Unified Training for Image Encoder and Attention Injection" + ], + "image_footnote": [], + "bbox": [ + 83, + 301, + 464, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce2655202b7f6a62180e6d2e91d944faeceeba5bceb2219dcc7842e57c3dd5be.jpg", + "image_caption": [ + "(d) Full Model", + "Figure 7. Ablation Study. (a) With only coarse embeddings from image encoder, generated patterns in the body of dog are different from image prompt. (b) With only fine embeddings from attention injection, there lacks coarse encodings of the dog for attention injection module to refine and thus the generated dog is distorted at later synthesized frames. (c) The unified training degrades the capability of image encoder and thus the dog is distorted. (d) The full model better keeps the all visual details of image prompt." + ], + "image_footnote": [], + "bbox": [ + 83, + 375, + 464, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from that in the image prompt. By contrast, the results of our full model show that our proposed model can transfer all the details in the image prompts to the synthesized videos.", + "bbox": [ + 75, + 575, + 468, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Only Fine Embeddings. In this ablation model, we only have fine embeddings of the image prompt in cross-frame attention layers. The main purpose of using fine embedding is to refine the coarse encoding of image prompts from the Image Encoder. Without coarse embeddings, the model with fine embeddings only cannot refine the details. As shown in Fig. 7(b), the first frame contains the exact appearance as the image prompt, but the temporal consistency cannot be guaranteed. The generated dog is distorted in the following frames. The reason is that the model trained in this way overfits the image prompt. In the first frame, the model can copy the information from image prompts. In the following frames, without the coarse embeddings, the generation of the appearance only relies on the propagation of the appearance from the first frame.", + "bbox": [ + 75, + 625, + 468, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Necessity of Coarse-to-Fine Training. In Video-Booth, we propose the coarse-to-fine training strategy, i.e., train the coarse embeddings first and then train the attention", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/86c1da2bb76c39a390117e05416d9b1b083ba8c039750b6f920610dcf415029f.jpg", + "table_caption": [ + "Table 2. Ablation Study. The full model has the best scores." + ], + "table_footnote": [], + "table_body": "
VariantsCLIP-Image ↑DINO ↑
(a) Coarse Embeddings only75.436664.9568
(b) Fine Embeddings only75.555366.0378
(c) Unified Training75.825467.4201
(d) Full Model76.163169.7374
", + "bbox": [ + 522, + 109, + 859, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "injection module. In this ablation model, we train these two modules within one stage. The unified training makes the model rely heavily on the strong guidance provided in attention injection. In this way, the trained image encoder has limited capability. Thus, the model trained in this way has a similar behavior as the model with only fine embeddings. The model also overfits the image prompts from the attention injection. As shown in the first example in Fig. 7(c), the appearance in the first frame is correct, but the generated dog in the following frames is distorted. Due to the image encoder having limited capability to provide the correct coarse encoding of image prompts, the attention injection cannot refine the details. By contrast, the full model can generate consistent frames with all details well preserved.", + "bbox": [ + 496, + 181, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 407, + 612, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework VideoBooth to generate videos using image prompts and text prompts. The image prompts specify the appearance of the subjects. We inject the image prompts into the model in two modules: Coarse Embeddings via Image Encoder and Fine Embeddings via Attention Injection. The Image Encoder provides the coarse embeddings of image prompts for the refinement of the Attention Injection module. These two modules cooperate with each other and they are trained in a coarse-to-fine manner. Our proposed VideoBooth generates consistent videos containing the desired subjects.", + "bbox": [ + 496, + 431, + 890, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. The model is trained with the image prompts that do not have very divergent views from target videos. Thus, if an image prompt with a back view is provided, we cannot generate videos from the front view. In future work, we can augment image prompts using image-to-3D models [56, 78] so that image prompts used for training will have different views from target videos.", + "bbox": [ + 496, + 598, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Negative Societal Impacts. The model may be applied to generate fake videos, which can be potentially avoided by using deepfake detection methods.", + "bbox": [ + 496, + 704, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and the Google PhD Fellowship. This study is partially supported by the National Key R&D Program of China (2022ZD0160201), and Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6696", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 3", + "[2] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06925, 2023. 3", + "[3] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023. 3", + "[4] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 3", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6", + "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 1, 3", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6", + "[8] Duygu Ceylan, Chun-Hao Huang, and Niloy J. MiTRA. Pix2video: Video editing using image diffusion. arXiv:2303.12688, 2023. 3", + "[9] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3", + "[10] Hong Chen, Xin Wang, Guanning Zeng, Yipeng Zhang, Yuwei Zhou, Feilin Han, and Wenwu Zhu. Videodreamer: Customized multi-subject text-to-video generation with disen-mix finetuning. arXiv preprint arXiv:2311.00990, 2023. 3", + "[11] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Disentangled parameter-efficient tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2, 3", + "[12] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photoverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 3", + "[13] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 3, 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3", + "[15] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3", + "[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1", + "[17] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.1", + "[18] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 1, 3", + "[19] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Ccredit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3", + "[20] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1", + "[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2, 3, 6", + "[22] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3", + "[23] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3", + "[24] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3", + "[25] Yuan Gong, Youxin Pang, Xiaodong Cun, Menghan Xia, Haoxin Chen, Longyue Wang, Yong Zhang, Xintao Wang, Ying Shan, and Yujiu Yang. Talecrafter: Interactive story visualization with multiple characters. arXiv preprint arXiv:2305.18247, 2023. 3", + "[26] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 1", + "[27] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023. 2, 3", + "[29] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 3", + "[30] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 3", + "[31] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1", + "[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3", + "[33] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3", + "[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3", + "[35] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3", + "[36] Zhihao Hu and Dong Xu. Videocontrolnet: A motion-guided video-to-video translation framework by using diffusion model with controlnet. arXiv preprint arXiv:2307.14073, 2023. 3", + "[37] Ziqi Huang, Kelvin CK Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6080-6090, 2023. 1", + "[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. arXiv preprint arXiv:2311.17982, 2023. 1", + "[39] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 1", + "[40] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3", + "[41] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zeroshot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3", + "[42] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2, 3, 4", + "[43] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13799-13808, 2021. 1", + "[44] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022.", + "[45] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[46] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1", + "[47] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Learning object level concepts using multi-concept prompt learning. arXiv preprint arXiv:2310.12274, 2023. 3", + "[48] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 1, 3", + "[49] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6", + "[50] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17089-17099, 2023. 3", + "[51] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. 2023. 2, 3", + "[52] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 2, 3", + "[53] Yuheng Li, Haotian Liu, Yangming Wen, and Yong Jae Lee. Generate anything anywhere in any scene. arXiv preprint arXiv:2306.17154, 2023. 3", + "[54] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. In arXiv, 2023. 3", + "[55] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2", + "[56] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3:" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6698", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 8", + "[57] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6", + "[58] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3", + "[59] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. arXiv preprint arXiv:2305.19327, 2023. 3", + "[60] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 1, 3", + "[61] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 3", + "[62] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1", + "[63] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[64] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3", + "[65] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3", + "[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[67] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1", + "[68] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 3", + "[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3", + "[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3, 6", + "[72] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3", + "[73] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1", + "[74] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. arXiv preprint arXiv:2303.07945, 2023. 3", + "[75] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 1", + "[76] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3", + "[77] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 3", + "[78] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 8", + "[79] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3", + "[80] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3", + "[81] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3" + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6699", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 3", + "[83] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 2, 3, 4, 6", + "[84] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 3", + "[85] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. arXiv preprint arXiv:2312.07537, 2023. 2", + "[86] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. Singleinsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 3", + "[87] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3", + "[88] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking \"text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 2, 3", + "[89] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video editing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 3", + "[90] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3", + "[91] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3", + "[92] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 3", + "[93] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3", + "[94] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3", + "[95] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[96] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3", + "[97] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3", + "[98] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 260 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6700", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_model.json b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d0e96fa363c0944b307abbfb3ecf28ea3a05ccad --- /dev/null +++ b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_model.json @@ -0,0 +1,3084 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.093, + 0.833, + 0.116 + ], + "angle": 0, + "content": "VideoBooth: Diffusion-based Video Generation with Image Prompts" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.143, + 0.348, + 0.162 + ], + "angle": 0, + "content": "Yuming Jiang" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.143, + 0.48, + 0.162 + ], + "angle": 0, + "content": "Tianxing Wu" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.143, + 0.61, + 0.162 + ], + "angle": 0, + "content": "Shuai Yang\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.63, + 0.143, + 0.741, + 0.162 + ], + "angle": 0, + "content": "Chenyang Si" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.162, + 0.338, + 0.178 + ], + "angle": 0, + "content": "Dahua Lin" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.162, + 0.434, + 0.179 + ], + "angle": 0, + "content": "Yu Qiao2" + }, + { + "type": "text", + "bbox": [ + 0.454, + 0.162, + 0.607, + 0.179 + ], + "angle": 0, + "content": "Chen Change Loy1" + }, + { + "type": "text", + "bbox": [ + 0.628, + 0.162, + 0.729, + 0.179 + ], + "angle": 0, + "content": "Ziwei Liu" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.179, + 0.765, + 0.197 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University \\(^{2}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.197, + 0.743, + 0.215 + ], + "angle": 0, + "content": "\\(^{3}\\)Wangxuan Institute of Computer Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.216, + 0.708, + 0.232 + ], + "angle": 0, + "content": "https://vchitec.github.io/VideoBooth-project/" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.238, + 0.202, + 0.251 + ], + "angle": 0, + "content": "Image Prompt" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.239, + 0.361, + 0.252 + ], + "angle": 0, + "content": "A horse eating grass." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.253, + 0.195, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.254, + 0.533, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.238, + 0.773, + 0.252 + ], + "angle": 0, + "content": "Horse grazes in snowy meadow" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.253, + 0.865, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.324, + 0.201, + 0.338 + ], + "angle": 0, + "content": "Image Prompt" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.324, + 0.528, + 0.337 + ], + "angle": 0, + "content": "Portrait of a dog, looks out the car window." + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.324, + 0.786, + 0.338 + ], + "angle": 0, + "content": "Dog walking in the green farm 4k" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.339, + 0.177, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.34, + 0.533, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.34, + 0.865, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.412, + 0.202, + 0.425 + ], + "angle": 0, + "content": "Image Prompt" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.412, + 0.398, + 0.425 + ], + "angle": 0, + "content": "Cat is looking at a laptop." + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.412, + 0.836, + 0.425 + ], + "angle": 0, + "content": "Close up of cat on top of a vintage chair" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.427, + 0.194, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.427, + 0.533, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.427, + 0.865, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.504, + 0.202, + 0.517 + ], + "angle": 0, + "content": "Image Prompt" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.495, + 0.53, + 0.521 + ], + "angle": 0, + "content": "Elephant drinking water in masai mara reserve, kenya" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.496, + 0.808, + 0.52 + ], + "angle": 0, + "content": "Elephant walk in the yellow grass of savannah" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.522, + 0.194, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.521, + 0.533, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.521, + 0.865, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.586, + 0.882, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Videos synthesized by image prompts. Our VideoBooth generates videos with the subjects specified in the image prompts." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.607, + 0.314, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.473, + 0.88 + ], + "angle": 0, + "content": "Text-driven video generation witnesses rapid progress. However, merely using text prompts is not enough to depict the desired subject appearance that accurately aligns with users' intents, especially for customized content creation. In this paper, we study the task of video generation with image prompts, which provide more accurate and direct content control beyond the text prompts. Specifically, we propose a feed-forward framework VideoBooth, with two dedicated designs: 1) We propose to embed image prompts in a coarse-to-fine manner. Coarse visual embeddings from image encoder provide high-level encodings of image prompts, while fine visual embeddings from the proposed attention injection module provide multi-scale and detailed encoding of image prompts. These two complementary embeddings can faithfully capture the desired appearance. 2) In the attention injection module at fine level, multi-scale" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.896, + 0.76 + ], + "angle": 0, + "content": "image prompts are fed into different cross-frame attention layers as additional keys and values. This extra spatial information refines the details in the first frame and then it is propagated to the remaining frames, which maintains temporal consistency. Extensive experiments demonstrate that VideoBooth achieves state-of-the-art performance in generating customized high-quality videos with subjects specified in image prompts. Notably, VideoBooth is a generalizable framework where a single model works for a wide range of image prompts with only feed-forward passes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.797, + 0.633, + 0.813 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Text-to-image models [16, 17, 20, 26, 31, 37, 39, 43-45, 62, 67-69, 73, 75] have attracted substantial attention. With Stable Diffusion [69], we can now easily generate images using texts. Recently, the focus has been shifted to text-to-video models [6, 18, 30, 38, 46, 48, 60, 76, 80, 84," + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.422, + 0.9 + ], + "angle": 0, + "content": "*This work was done when Shuai Yang was in S-Lab, NTU." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6689" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "85, 97] to generate videos by taking text descriptions as inputs. However, in some user cases, texts alone are not expressive enough to define the specific appearance of subjects [21, 71]. For example, as shown in Fig. 2, if we want to generate a video clip containing the dog as the third row, we need to use several attributive adjuncts to define the appearance of the dog in text prompts. Even with these extensive attributive adjuncts, models still cannot generate the desired appearance. Defining the appearance of the desired object with texts alone has the following flaws: 1) It is hard to enumerate all the desired attributes, and 2) The model cannot capture all attributes accurately with a long text. Compared to using texts, a more straightforward way to define the appearance is to provide reference images, termed image prompts. The image prompts are complementary to the text prompts and enrich the details that are hard to be depicted by text prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.548 + ], + "angle": 0, + "content": "There are several attempts to introduce image prompts into text-to-image models, which can be roughly divided into two groups. One is to fine-tune parts of parameters using few-shot reference images [15, 21, 28, 51, 71], which contain the same objects captured under different circumstances. However, the requirement for the number of reference images is demanding as sometimes it is not practical to obtain multiple images of the same object. The other category [11, 42, 52, 83, 87, 88], aiming to address this limitation, proposes to embed image prompts into text-to-image models and the inference is tuning-free. Both of these two types of attempts achieve plausible results in generating images containing objects specified in image prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.807 + ], + "angle": 0, + "content": "In this paper, we study a more challenging task, i.e., text-to-video generation with image prompts. The task has two main challenges: 1) Similar to text-to-image generation, the attributes of image prompts should be accurately captured and then reflected in the generated videos; 2) Different from text-to-image generation, we aim for the dynamic movement of the object rather than a static one. Directly adapting these methods to video domain results in mismatched appearance or unnatural degraded movements. To address these challenges, we proposed VideoBooth with elaborately designed coarse-to-fine visual embedding components: 1) Coarse visual embeddings via image encoder: An image encoder is trained to inject the image prompts into text embeddings; 2) Fine visual embeddings via attention injection: The image prompts are mapped to multiscale latent representations to control the generation process through cross-frame attentions of text-to-video models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Specifically, inspired by early attempts [42, 83] in text-to-image models with image prompts, we extract the CLIP image features of the provided image prompts using the pretrained CLIP model [66]. Then the extracted features are mapped into the text embedding space, which are inserted to replace parts of the original text embeddings. The well-" + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.092, + 0.877, + 0.105 + ], + "angle": 0, + "content": " eating snack inside big iron cage at home." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.107, + 0.88, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 2. The Use of Image Prompts. We generate three video clips using different types of prompts: simple text prompt, long text prompt, and image prompt. We use the LLaVa model [55] to generate a text prompt describing the appearance of the image prompt. Using text prompts alone cannot fully capture the visual characteristics of the image prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.493 + ], + "angle": 0, + "content": "trained encoder embeds the coarse appearance information of the given image prompts. However, coarse visual embedding is a universal embedding: 1) It only contains high-level semantic information, and 2) It is shared across all blocks with the same scale. As a result, some visual details are missing in the coarse visual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.719 + ], + "angle": 0, + "content": "To further refine the generated details as well as maintain temporal consistency, different from the highly compacted coarse visual embeddings, multi-scale image prompts are injected into cross-frame attention modules in different layers. The image prompts provide spatial information as well as details with different granularities. On the one hand, keeping spatial information of the image prompts can retain more details. On the other hand, different cross-frame attention modules need detailed information at different scales. Specifically, the latent representations of image prompts are appended as additional keys and values to refine the details in the first generated frame. To propagate the refined first frame to the following frames to maintain temporal consistency, we then use the updated values of the first frame as values for the remaining frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.87 + ], + "angle": 0, + "content": "We set up a dedicated VideoBooth dataset to support the study of the new task. With each video, we provide an image prompt and a text prompt. Extensive experiments demonstrate the effectiveness of our proposed VideoBooth to generate videos with subjects specified in image prompts. As shown in Fig. 1, videos generated by VideoBooth better keep the visual attributes of image prompts. Besides, our proposed VideoBooth is tuning-free at inference time and videos can be generated with feed-forward passes only. The contributions are summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "- To our knowledge, we are the first to explore the task of video generation using image prompts without finetuning" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "at inference time. We propose a dedicated dataset to support the task. Our proposed VideoBooth framework can generate consistent videos containing the subjects specified in image prompts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.153, + 0.468, + 0.198 + ], + "angle": 0, + "content": "- We introduce a new coarse-to-fine visual embedding strategy by image encoder and attention injection, which better captures the characteristics of the image prompts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.199, + 0.468, + 0.243 + ], + "angle": 0, + "content": "- We propose a novel attention injection method, using the multi-scale image prompts with spatial information to refine the generated details." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.153, + 0.468, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.257, + 0.218, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Text-to-Video Models take the text descriptions as inputs and generate clips of videos. Early explorations [34, 80] on text-to-video models are based on the idea of VQVAE. Make-A-Video [76] proposes to add temporal attention to the architecture of DALLE2 model [68]. Recently, the emergence of diffusion models [32, 70] boosts research on text-to-video models [1, 23, 30, 33, 60, 82, 97]. Video LDM [6] proposes to train the text-to-video models on Stable Diffusion with temporal attention and 3D convolution introduced to handle the temporal generation. Gen-1 [18] introduces depth maps to handle the temporal consistency of text-to-video models. Some methods [27, 95] resort to training separate modules for synthesizing motions. All of the methods initialize their models with pre-trained text-to-image models. Another paradigm of using text-to-image models is to directly apply Stable Diffusion [69] to few-shot or zero-shot settings. Tune-A-Video [84] adapts the self-attention into cross-frame attention and then finetunes the stable diffusion model on a video clip. Models trained in this way have the capability to transfer motions from original videos. Text2Video-Zero [48] proposes to generate videos by using correlated noise maps to improve consistency. Apart from video generation, diffusion models have been applied to video-to-video generations [8, 9, 19, 24, 36, 41, 54, 58, 64, 65, 74, 81, 89, 90, 94]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Customized Content Creation aims at generating images and videos using reference images [15]. For customized text-to-image generation, optimization-based methods [28, 35, 51, 53] are proposed to optimize the weights of the diffusion model. For example, Textual Inversion [21] optimizes the word embeddings, while DreamBooth [71] proposes to finetune the weights of Stable Diffusion as well. Optimization-based methods require several reference images with the same subject to avoid the overfitting of the model, which is demanding in real-world applications. The cost of finetuning hampers the practical usage of these methods. To address these limitations, encoder-based methods [61, 88, 91, 98] are proposed to learn a mapping network to embed the reference images. ELITE [83] proposes to learn a global mapping network and local mapping network to encode the images into word embeddings. Jia" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.469 + ], + "angle": 0, + "content": "et al. [42] propose to use an additional cross-attention to embed the image features. With the trained encoder, the personalized generation can be achieved in a feed-forward pass. Some recent works [2, 11, 22, 25] combine the encoder-based model and finetuning-based model to improve the performance. BLIP-Diffusion [52] proposes to pretrain a multimodal encoder in a large-scale dataset and then finetune the model on the specific subject for inference. Customized image generation is also applied to place the objects into the user-specified scenes [4, 13, 50, 77, 93]. Also, some efforts [12, 14, 40, 72, 79, 86, 87, 92] have been made to personalized face generation. He et al. [29] propose to improve the performance from the data perspective. Some works [3, 47, 59] focus on composing multiple subjects in one image. Apart from works on image generation, there are some early attempts at personalized video manipulation. Make-A-Protagonist [96] edits an existing video in a personalized way using Stable Diffusion 2.1 to embed the image prompts. The motion of the original video is learned from Tune-A-Video [84]. VideoDreamer [10] proposes to generate personalized videos by generating the first frames using a finetuning-based method and then generating the video clip using the Text2Video-Zero [48]. Different from existing works, our proposed VideoBooth does not need to finetune any weights at the inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.625, + 0.497 + ], + "angle": 0, + "content": "3. VideoBooth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Our proposed VideoBooth aims at generating videos from an image prompt \\( I \\) and a text prompt \\( T \\). The image prompt specifies the appearance of the subject. An overview of our proposed VideoBooth is illustrated in Fig. 3. The image prompt is fed into VideoBooth in two levels. At the coarse level, it is fed into a pretrained CLIP Image encoder to extract visual features. An encoder, composed of several MLP layers, is trained to map visual features into the space of text embeddings. The obtained embedding \\( f_{I} \\) will be inserted into text embedding, which is extracted by feeding text prompt \\( T \\) into CLIP text encoder. To further refine the synthesized details, we propose to inject image prompt \\( I \\) into the cross-frame attention module in the pretrained video diffusion model. Specifically, we append latent representation \\( x_{I}^{t} \\) of image prompt \\( I \\) into the cross-frame attention. In this way, multi-scale visual details with spatial information are involved in the calculation of attention maps so that visual characteristics can be better preserved. Two ways of feeding image prompt corporate with each other in a coarse-to-fine manner. The encoder provides coarse visual embeddings of the image prompt, while the attention injection provides fine visual embeddings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.886, + 0.863 + ], + "angle": 0, + "content": "3.1. Preliminary: Pretrained Text-to-Video Model" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our proposed VideoBooth is developed based on the pretrained text-to-video model [30, 82, 84]. In this section, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.1, + 0.852, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.35, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Overview of VideoBooth. VideoBooth generates videos by taking image prompts \\(I\\) and text prompts \\(T\\) as inputs. The image prompt is fed into the CLIP image encoder, followed by MLP layers. The obtained coarse visual embedding \\(f_{I}\\) is then inserted into the text embeddings. The composed embeddings serve as the input for cross attention. The embedding extracted by the encoder provides a coarse encoding of the visual appearance of the image prompt. To further refine the details in the generated videos, at the fine level, we append the latent representation of the image prompt to the cross-frame attention as additional keys and values. Different cross-frame attention layers receive latent representations with different scales. The multi-scale features with spatial details refine the synthesized details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.522 + ], + "angle": 0, + "content": "will briefly introduce the framework of text-to-video model. Inflated 2D Conv. To handle video data and capture the temporal correlation, 2D conv in the Stable Diffusion model is inflated to 3D conv. In this way, the U-Net can encode 3D features containing the temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Cross-Frame Attention Module. Stable Diffusion has a self-attention module, where the features are enhanced by attending to themselves. To improve temporal consistency, the self-attention is modified into cross-frame attention. Specifically, the feature of each frame is enhanced by attending and referencing to the first frame and the previous frame. The cross-frame attention operates on both the spatial domain and temporal domain, thus the temporal consistency of the synthesized frames is improved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Temporal Attention Module. Apart from cross-frame attention, a temporal attention module is introduced to further improve temporal consistency. Temporal attention operates on temporal domain and attends to all frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.469, + 0.743 + ], + "angle": 0, + "content": "3.2. Coarse Visual Embeddings via Image Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Given an image prompt \\( I \\) and text prompt \\( T \\), the generated video is supposed to be consistent with visual elements and textual elements. Inspired by previous attempts at image-based customization methods [42, 83], we propose to encode visual information of image prompts by an image encoder. The image prompt and text prompt complement each other. The image prompt provides visual characteristics of the desired subject in the video, and the text prompt provides other orthogonal information. The extracted visual embeddings are combined with text embeddings as the final" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.552 + ], + "angle": 0, + "content": "embeddings for the cross-attention module. Specifically, the CLIP image encoder is employed to extract the visual features \\( f_{V} \\) of image prompt \\( I \\). Since the discrepancy exists between the CLIP image and text embeddings, \\( f_{V} \\) is then fed into MLP layers \\( F(\\cdot) \\) to map \\( f_{V} \\) to the spaces of text embeddings. The final embedding \\( f_{I} \\) for the image prompt is obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.565, + 0.891, + 0.581 + ], + "angle": 0, + "content": "\\[\nf _ {V} = \\operatorname {C L I P} _ {I} (I), f _ {I} = F \\left(f _ {V}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.594, + 0.891, + 0.624 + ], + "angle": 0, + "content": "As for the text prompt \\( T \\), we feed it into the CLIP text encoder to extract the text embedding \\( f_{T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.635, + 0.891, + 0.653 + ], + "angle": 0, + "content": "\\[\nf _ {T} = \\left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \\dots , f _ {T} ^ {k}, \\dots \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.664, + 0.875, + 0.68 + ], + "angle": 0, + "content": "where \\( f_{T}^{k} \\) is the \\( k \\)-th word embedding in the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.77 + ], + "angle": 0, + "content": "To make the diffusion model generate videos conditioning on both text prompts and image prompts, we need to integrate these two embeddings, i.e., \\( f_{I} \\) and \\( f_{T} \\). The idea is to replace the word embedding of the target subject with \\( f_{I} \\). Mathematically, \\( f_{I} \\) and \\( f_{t} \\) are fused to obtain the final text condition \\( c_{t} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.78, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\nc _ {T} = \\left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \\dots , f _ {T} ^ {k - 1}, f _ {I}, f _ {T} ^ {k + n}, \\dots \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( k \\) is the token index of the target subject in the text embedding, and \\( n \\) is the length of the text tokens for the target subject. For example, an image of a papillon dog is provided as an image prompt. To fuse the information from the image prompt and the text prompt \"Papillon dog celebrates birthday with gifts\", the word embeddings of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6692" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.462, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.473, + 0.394 + ], + "angle": 0, + "content": "Figure 4. Fine Visual Embedding Refinement. We propose to inject the latent representation of image prompt (here we use the image for illustration purpose) directly into the cross-frame attention module. We use the keys and values from the image prompt to update the values of the first frame firstly. Then, the updated values of the first frame are used to update the remaining frames. Injecting the image prompt in the cross-frame attention helps to transfer the detailed visual characteristics of the image prompts to the synthesized frames. We perform the refinement in different cross-attention layers with different scales." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.414, + 0.47, + 0.444 + ], + "angle": 0, + "content": "\"papillon dog\" will be replaced with \\( f_{I} \\) before they are fed into the cross-attention module of the diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.536 + ], + "angle": 0, + "content": "During the training of the coarse stage, we fix the parameters of the CLIP image encoder, and train the MLP layers. To make the diffusion model accommodate with the composed text embeddings \\( c_{T} \\), we also finetune K and V projections (linear layers to map the input feature to the corresponding keys and values) in the cross-attention module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.47, + 0.563 + ], + "angle": 0, + "content": "3.3. Fine Visual Embeddings via Attention Injection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.719 + ], + "angle": 0, + "content": "The well-trained image encoder embeds the coarse visual embeddings for image prompts and thus the synthesized videos contain the subjects specified in image prompts. However, the image encoder projects the image prompt into a flattened high-level representation, resulting in the loss of its detailed visual cues. Thus, some detailed visual characteristics in the image prompts may not be well preserved. To address this problem, a more effective way to preserve these details is to provide the model with the image prompts with spatial resolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.81 + ], + "angle": 0, + "content": "To further refine the synthesized details, we propose to inject image prompts into the cross-frame attention of the text-to-video models. By injecting image prompts into the cross-frame attention, the image prompts are involved in the updates of the synthesized frames so that the model can directly borrow some visual cues from image prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Since text-to-video diffusion models operate in the latent space, we first feed the image prompt into the VAE of Stable Diffusion and get its latent representation \\( x_{I} \\). Moreover, since the sampling of the videos starts from the noise map, the latent in the intermediate timesteps contains the noises. If we append the clean latent \\( x_{I} \\) of the image prompt" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "to the cross-frame attention, the domain discrepancy exists. Therefore, we follow the diffusion forward process to add corresponding noises to \\( x_{I} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.142, + 0.892, + 0.161 + ], + "angle": 0, + "content": "\\[\nx _ {t} ^ {I} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} ^ {I} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.892, + 0.197 + ], + "angle": 0, + "content": "where \\(\\overline{\\alpha}\\) is a hyperparameter determined by the denoising schedule and \\(\\epsilon \\sim N(0, I)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.333 + ], + "angle": 0, + "content": "The cross-frame attention is used to improve the temporal consistency of the generated frames. For each frame, the key and value are the concatenation of the features of the first frame and the previous frame. Here, we introduce the image prompts as the additional keys and values for the frames. As shown in Fig. 4, we propose to update the values of the first frame firstly using the keys and values of the image prompts and the frame itself. Mathematically, the operation can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.338, + 0.892, + 0.373 + ], + "angle": 0, + "content": "\\[\nV _ {0} ^ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {0} K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.375, + 0.805, + 0.393 + ], + "angle": 0, + "content": "\\[\nK = \\left[ K _ {I}, K _ {0} \\right], V = \\left[ V _ {I}, V _ {0} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.519 + ], + "angle": 0, + "content": "where \\( K_{I} \\) and \\( V_{I} \\) are the keys and values obtained from the image prompts. The query, key, and value of the first frame are denoted \\( Q_{0}, K_{0} \\), and \\( V_{0} \\), respectively. It should be noted that we use a separately trained \\( \\mathrm{K} \\) and \\( \\mathrm{V} \\) projection for latent representations \\( x_{t}^{I} \\) of image prompts because the image prompts have clean backgrounds, which are different from other frames. The parameters of the newly added \\( \\mathrm{K} \\) and \\( \\mathrm{V} \\) projections are initialized by original \\( \\mathrm{K} \\) and \\( \\mathrm{V} \\) projections." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Then the updated first frame is used to refine the remaining frames. When updating the remaining frames, the keys used for calculating the attention maps are the original keys, while the values are the updated ones. The update is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.599, + 0.892, + 0.635 + ], + "angle": 0, + "content": "\\[\nV _ {i} ^ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.636, + 0.821, + 0.654 + ], + "angle": 0, + "content": "\\[\nK = [ K _ {0}, K _ {i - 1} ], V = [ V _ {0} ^ {\\text {n e w}}, V _ {i - 1} ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.735 + ], + "angle": 0, + "content": "To sum up, in the attention injection, we update the values of the first frame using the image prompts first, and then use the updated first frame to update the other frames. In this way, the visual cues from the image prompts can be consistently propagated to all the frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.826 + ], + "angle": 0, + "content": "It should be noted that the diffusion model has multiple cross-frame attention layers with different scales. To inject multi-scale visual cues for better detail refinement, in different cross-frame attention layers, we feed latent representations of the image prompts with corresponding resolutions, which are obtained from different stages of the U-Net." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.792, + 0.849 + ], + "angle": 0, + "content": "3.4. Coarse-to-Fine Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The visual details of the image prompts are embedded into the final synthesized results in two stages: coarse visual embeddings using an image encoder and fine visual embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6693" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "by attention injection. We propose to train these two modules in a coarse-to-fine manner. In other words, we train the coarse image encoder and tune the parameters in the cross-attention first. After the model has the capability of generating videos containing the subjects specified in image prompts, we then train the attention injection module to embed image prompts into cross-frame attention layers. As we will show in the ablation study (Sec. 5.5), if these two modules are trained together, the fine attention injection module leaks the strong visual cues and the coarse encoder learns meaningless representations. As a result, in sampling phase, the image encoder for the coarse visual embedding cannot provide the coarse information and then the fine attention module cannot refine the details. Therefore, it is necessary to train VideoBooth in a coarse-to-fine manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.35, + 0.27, + 0.366 + ], + "angle": 0, + "content": "4. VideoBooth Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.82 + ], + "angle": 0, + "content": "We establish the VideoBooth dataset to support the task of video generation using image prompts. We start from the WebVid dataset [5], a well-known open-source dataset for text-to-video generation. In the WebVid dataset, there is a text prompt with each video. In this paper, we study the task of generating a video clip from one text prompt and one image prompt. Hence, in addition to the original text prompt, we need to provide an image prompt for each video. We propose to segment the subjects from the first frame of the video using the Grounded-SAM (Grounded Segment Anything) [49, 57], and the segmented subjects are image prompts. The Grounded-SAM receives word prompts as inputs and generates segmentation masks for the target subjects specified in word prompts. To obtain the word prompt for the input to Grounded-SAM, we use the spaCy library to parse the noun chunks from the original text prompts, which are used as the word prompts. After the segmentation, we perform data filtering to ensure the data quality. We filter out small objects and large objects (those are almost the same size as the original video) according to the ratio of the object to the whole video. Also, since we focus on generating video clips containing moving objects, we further filter the videos containing moving objects. The keywords we used for filtering are dog, cat, bear, car, panda, tiger, horse, elephant, and lion. In the current version, we have processed 2.5M subset of the WebVid dataset. After data filtering, we have 48,724 video data pairs for training. We will process the full set of the WebVid dataset and include the filtered data in our VideoBooth dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To evaluate the performance, we also set up a test benchmark. The test benchmark consists of 650 test pairs. For each pair, an image prompt and a text prompt are provided. The test pairs are selected from the rest of the WebVid-10M dataset, which does not overlap with the training set." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons. VideoBooth achieves the best image alignment and comparable text alignment performance." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.123, + 0.897, + 0.189 + ], + "angle": 0, + "content": "
MethodCLIP-Text ↑CLIP-Image ↑DINO ↑
Textual Inversion [21]29.974969.799545.3143
DreamBooth [71]30.687771.207852.9661
ELITE [83]30.088173.751858.9522
VideoBooth (Ours)30.096774.797165.0979
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.214, + 0.633, + 0.231 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.239, + 0.704, + 0.256 + ], + "angle": 0, + "content": "5.1. Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Textual Inversion [21] is a method for customized text-to-image generation. The appearance of the target subjects is embedded into the text embeddings. Concretely, a text token \\( S^{*} \\) is optimized to represent the subject. We adapt it to the task of text-to-video generation by replacing the image model with the video model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.412 + ], + "angle": 0, + "content": "DreamBooth [71] is also proposed for customized text-to-image generation. It injects the target subject into the text tokens as well as model weights. During the training, both the model weights and the special token \\( S* \\) are optimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.518 + ], + "angle": 0, + "content": "ELITE [83] is an encoder-based method for personalized generation. An encoder is trained to embed the images into the text embeddings. Local mapping and global mapping are employed to transform the CLIP embedding of image prompts into the features, which are injected into the cross-attention module. We adapt and retrain the method using the same pretrained video model we use." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.684, + 0.543 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We use three metrics to evaluate the performance [83]. To measure the alignment of the generated videos and given text prompts, we use the CLIP-Text metric. The metric is calculated using the cosine similarity of the CLIP text embeddings of text prompts and CLIP image embeddings of the generated frames. For each video, the value is obtained by averaging values of the all frames. As for the evaluation of the similarity between the given image prompts and the generated videos, we adopt two metrics: CLIP-Image and DINO [7, 63]. The CLIP-Image metric is calculated by the cosine similarity between the CLIP image embedding of image prompts and generated frames. Since the CLIP model is trained to align image embeddings and text embeddings, we follow the practice in previous methods [13, 71] and use the DINO similarity as another indicator. DINO is trained to differentiate the differences between objects of the same classes. We use the ViT-S/16 model to extract the features of the image prompts and generated frames. The final score is obtained by averaging over all frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.741, + 0.865 + ], + "angle": 0, + "content": "5.3. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We report quantitative results in Table 1. As shown in Table 1, our proposed VideoBooth achieves state-of-the-art" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.353, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.096, + 0.613, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.096, + 0.871, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.395, + 0.88, + 0.411 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison. VideoBooth effectively preserves the fidelity of image prompts and achieves better visual quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.474, + 0.709 + ], + "angle": 0, + "content": "image alignment performance compared to baseline methods. As for the alignment with the text prompts, our proposed VideoBooth has comparable performance with the baseline models. It should be noted that the CLIP-Text score of DreamBooth is significantly higher than the other methods. The reason lies in that the optimized token \\( S* \\) in DreamBooth is inserted into the text embeddings, rather than replacing the original word embeddings like other methods. This would result in the generated videos of DreamBooth being highly related to the text prompts but having no correlation to the image prompts in some cases. We also conducted a user study, in which 25 users participated. Each user is presented with twelve groups of videos, and each group contains four videos generated by four methods. For each group, users are asked to make three choices: 1) which one has the best image alignment? 2) which one has the best text alignment? 3) which one has the best overall quality. Figure 6 summarizes the results. Our results are preferred by most users in all three dimensions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.307, + 0.74 + ], + "angle": 0, + "content": "5.4. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We show three visual comparisons on the generated video frames of our proposed VideoBooth and baseline methods in Fig. 5. In the first example, the model is supposed to generate videos containing the dog specified in the image laying on the ground. Textual Inversion cannot correctly embed the appearance of the dog and results in generating another totally different dog. DreamBooth and ELITE can embed the coarse appearance of the dog but the generated details vary from the image prompt. Our proposed VideoBooth successfully embeds the details of the image prompts into" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.421, + 0.878, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 6. User Study. Our proposed VideoBooth achieves the highest user preference ratios on all three dimensions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.732 + ], + "angle": 0, + "content": "the synthesized videos. In the second example, the condition is to generate a cat playing with the yellow ball. All the methods can generate a yellow ball and a cat but only our proposed VideoBooth can accurately generate the cat having the appearance from the image prompt. As for the last example, Textual Inversion fails to generate a car. DreamBooth generates a distorted car in the bush. ELITE model can generate a car in the bush, but the color differs from the image prompt. By contrast, our model can generate videos having the same car in the bush." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.655, + 0.758 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.825 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the proposed components, we perform three ablation studies. Due to the computational resources, we train these models on the subset of our training set. The quantitative metrics are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Only Coarse Embeddings. This ablation model injects the image prompts with only coarse embeddings via Image Encoder. In the example shown in Fig. 7(a), the ablation model only encodes the coarse appearance of the image prompts. The pattern in the legs of the synthesized dog is different" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.094, + 0.154, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Dog as a chef at the table" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.153, + 0.465, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.214, + 0.409, + 0.226 + ], + "angle": 0, + "content": "(a) Only Coarse Embeddings with Image Encoder" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.227, + 0.465, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.29, + 0.413, + 0.301 + ], + "angle": 0, + "content": "(b) Only Fine Embeddings with Attention Injection" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.302, + 0.465, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.363, + 0.444, + 0.374 + ], + "angle": 0, + "content": "(c) Unified Training for Image Encoder and Attention Injection" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.376, + 0.465, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.439, + 0.315, + 0.45 + ], + "angle": 0, + "content": "(d) Full Model" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.451, + 0.468, + 0.562 + ], + "angle": 0, + "content": "Figure 7. Ablation Study. (a) With only coarse embeddings from image encoder, generated patterns in the body of dog are different from image prompt. (b) With only fine embeddings from attention injection, there lacks coarse encodings of the dog for attention injection module to refine and thus the generated dog is distorted at later synthesized frames. (c) The unified training degrades the capability of image encoder and thus the dog is distorted. (d) The full model better keeps the all visual details of image prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.469, + 0.623 + ], + "angle": 0, + "content": "from that in the image prompt. By contrast, the results of our full model show that our proposed model can transfer all the details in the image prompts to the synthesized videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.469, + 0.852 + ], + "angle": 0, + "content": "Only Fine Embeddings. In this ablation model, we only have fine embeddings of the image prompt in cross-frame attention layers. The main purpose of using fine embedding is to refine the coarse encoding of image prompts from the Image Encoder. Without coarse embeddings, the model with fine embeddings only cannot refine the details. As shown in Fig. 7(b), the first frame contains the exact appearance as the image prompt, but the temporal consistency cannot be guaranteed. The generated dog is distorted in the following frames. The reason is that the model trained in this way overfits the image prompt. In the first frame, the model can copy the information from image prompts. In the following frames, without the coarse embeddings, the generation of the appearance only relies on the propagation of the appearance from the first frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "The Necessity of Coarse-to-Fine Training. In Video-Booth, we propose the coarse-to-fine training strategy, i.e., train the coarse embeddings first and then train the attention" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.09, + 0.875, + 0.103 + ], + "angle": 0, + "content": "Table 2. Ablation Study. The full model has the best scores." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.11, + 0.86, + 0.175 + ], + "angle": 0, + "content": "
VariantsCLIP-Image ↑DINO ↑
(a) Coarse Embeddings only75.436664.9568
(b) Fine Embeddings only75.555366.0378
(c) Unified Training75.825467.4201
(d) Full Model76.163169.7374
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.394 + ], + "angle": 0, + "content": "injection module. In this ablation model, we train these two modules within one stage. The unified training makes the model rely heavily on the strong guidance provided in attention injection. In this way, the trained image encoder has limited capability. Thus, the model trained in this way has a similar behavior as the model with only fine embeddings. The model also overfits the image prompts from the attention injection. As shown in the first example in Fig. 7(c), the appearance in the first frame is correct, but the generated dog in the following frames is distorted. Due to the image encoder having limited capability to provide the correct coarse encoding of image prompts, the attention injection cannot refine the details. By contrast, the full model can generate consistent frames with all details well preserved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.613, + 0.423 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.599 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework VideoBooth to generate videos using image prompts and text prompts. The image prompts specify the appearance of the subjects. We inject the image prompts into the model in two modules: Coarse Embeddings via Image Encoder and Fine Embeddings via Attention Injection. The Image Encoder provides the coarse embeddings of image prompts for the refinement of the Attention Injection module. These two modules cooperate with each other and they are trained in a coarse-to-fine manner. Our proposed VideoBooth generates consistent videos containing the desired subjects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Limitations and Future Work. The model is trained with the image prompts that do not have very divergent views from target videos. Thus, if an image prompt with a back view is provided, we cannot generate videos from the front view. In future work, we can augment image prompts using image-to-3D models [56, 78] so that image prompts used for training will have different views from target videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Potential Negative Societal Impacts. The model may be applied to generate fake videos, which can be potentially avoided by using deepfake detection methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and the Google PhD Fellowship. This study is partially supported by the National Key R&D Program of China (2022ZD0160201), and Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6696" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06925, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[4] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[8] Duygu Ceylan, Chun-Hao Huang, and Niloy J. MiTRA. Pix2video: Video editing using image diffusion. arXiv:2303.12688, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.471, + 0.644 + ], + "angle": 0, + "content": "[9] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[10] Hong Chen, Xin Wang, Guanning Zeng, Yipeng Zhang, Yuwei Zhou, Feilin Han, and Wenwu Zhu. Videodreamer: Customized multi-subject text-to-video generation with disen-mix finetuning. arXiv preprint arXiv:2311.00990, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[11] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Disentangled parameter-efficient tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photoverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[13] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[17] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[19] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Ccredit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[23] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[24] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[25] Yuan Gong, Youxin Pang, Xiaodong Cun, Menghan Xia, Haoxin Chen, Longyue Wang, Yong Zhang, Xintao Wang, Ying Shan, and Yujiu Yang. Talecrafter: Interactive story visualization with multiple characters. arXiv preprint arXiv:2305.18247, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6697" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[29] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[30] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[31] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[33] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[35] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[36] Zhihao Hu and Dong Xu. Videocontrolnet: A motion-guided video-to-video translation framework by using diffusion model with controlnet. arXiv preprint arXiv:2307.14073, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.677 + ], + "angle": 0, + "content": "[37] Ziqi Huang, Kelvin CK Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6080-6090, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. arXiv preprint arXiv:2311.17982, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[39] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[40] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[41] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zeroshot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[42] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[43] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13799-13808, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[44] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[45] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[46] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[47] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Learning object level concepts using multi-concept prompt learning. arXiv preprint arXiv:2310.12274, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[48] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[49] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[50] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17089-17099, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[51] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[52] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[53] Yuheng Li, Haotian Liu, Yangming Wen, and Yong Jae Lee. Generate anything anywhere in any scene. arXiv preprint arXiv:2306.17154, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[54] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. In arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[55] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[56] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6698" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[57] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[58] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[59] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. arXiv preprint arXiv:2305.19327, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[60] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[61] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[62] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[63] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[64] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[65] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[67] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[68] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[72] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[73] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[74] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. arXiv preprint arXiv:2303.07945, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[75] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[76] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[77] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[78] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[79] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[80] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[81] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6699" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[82] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[83] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[84] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[85] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. arXiv preprint arXiv:2312.07537, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[86] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. Singleinsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[87] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[88] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking \"text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[89] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video editing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[90] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[91] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.72 + ], + "angle": 0, + "content": "[92] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.471, + 0.776 + ], + "angle": 0, + "content": "[93] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[94] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[95] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[96] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[97] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[98] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6700" + } + ] +] \ No newline at end of file diff --git a/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_origin.pdf b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f247ef29083a5d318a045e65624123cf39c0df79 --- /dev/null +++ b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/f4a1b767-87c7-478e-8bb4-526b79e30b63_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68ab946648869e22c87dfb00ba3447fb0dcab43bfbf7ab362aaf9e5b67b955cd +size 3077350 diff --git a/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/full.md b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8165f5e72d577456ba076cbc5c5893985d2de05 --- /dev/null +++ b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/full.md @@ -0,0 +1,414 @@ +# VideoBooth: Diffusion-based Video Generation with Image Prompts + +Yuming Jiang + +Tianxing Wu + +Shuai Yang $^{3*}$ + +Chenyang Si + +Dahua Lin + +Yu Qiao2 + +Chen Change Loy1 + +Ziwei Liu + +$^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai AI Laboratory + +$^{3}$ Wangxuan Institute of Computer Technology, Peking University + +https://vchitec.github.io/VideoBooth-project/ + +Image Prompt + +A horse eating grass. + +![](images/0ee099b10ef93093afb05ebce906facfcd700ee993c6e0fc47d1555a48328cc5.jpg) + +![](images/9ed8eed6772cc25c9c26de9c36835f5dc0c5925cfdbb8c848306503bbdc2841f.jpg) + +Horse grazes in snowy meadow + +![](images/c6cde38fbdb97095e5490a53d35d42928e2c79e795dc4506fda21f9229a3b6d4.jpg) + +Image Prompt + +Portrait of a dog, looks out the car window. + +Dog walking in the green farm 4k + +![](images/f23977d3089665deeae489cc80e070f51668bb3a40054bead28e0362b8599e7c.jpg) + +![](images/f3cd9e2f7fbfeaf24cc2a8cae94e3cedb3683d0e01c4d402fbf68e3eab9e3e0b.jpg) + +![](images/7d199bbdf497d60bdd0ee3de9e36de32e016d9272b55ab9010f8b18711019365.jpg) + +Image Prompt + +Cat is looking at a laptop. + +Close up of cat on top of a vintage chair + +![](images/59453a6f2e07afe2a46e8d9d6cfb2e184a0f56b4bc90702ed376c36b71448562.jpg) + +![](images/196f7f5ab890e3f4cb7f8268c1b58a366095ce4d2c8e9c7406cd41ba3f71f85d.jpg) + +![](images/5a897094ff13e8930d3047e0bb97aea526a38579d275e9a79b5345b5c5f50a83.jpg) + +Image Prompt + +Elephant drinking water in masai mara reserve, kenya + +Elephant walk in the yellow grass of savannah + +![](images/b7cbbfe9b71d6ab0c358d6b23d921e69b0eb55d377366f6ca858990ea19a0dd2.jpg) +Figure 1. Videos synthesized by image prompts. Our VideoBooth generates videos with the subjects specified in the image prompts. + +![](images/cdcaec6ff80b79900beb17dd6efaaa37ef9d3a6e3a72517779d25deecf9f63a6.jpg) + +![](images/80078ebce751182a79e080233234349871584f74028c9dd32f3747b42b4ac0f1.jpg) + +# Abstract + +Text-driven video generation witnesses rapid progress. However, merely using text prompts is not enough to depict the desired subject appearance that accurately aligns with users' intents, especially for customized content creation. In this paper, we study the task of video generation with image prompts, which provide more accurate and direct content control beyond the text prompts. Specifically, we propose a feed-forward framework VideoBooth, with two dedicated designs: 1) We propose to embed image prompts in a coarse-to-fine manner. Coarse visual embeddings from image encoder provide high-level encodings of image prompts, while fine visual embeddings from the proposed attention injection module provide multi-scale and detailed encoding of image prompts. These two complementary embeddings can faithfully capture the desired appearance. 2) In the attention injection module at fine level, multi-scale + +image prompts are fed into different cross-frame attention layers as additional keys and values. This extra spatial information refines the details in the first frame and then it is propagated to the remaining frames, which maintains temporal consistency. Extensive experiments demonstrate that VideoBooth achieves state-of-the-art performance in generating customized high-quality videos with subjects specified in image prompts. Notably, VideoBooth is a generalizable framework where a single model works for a wide range of image prompts with only feed-forward passes. + +# 1. Introduction + +Text-to-image models [16, 17, 20, 26, 31, 37, 39, 43-45, 62, 67-69, 73, 75] have attracted substantial attention. With Stable Diffusion [69], we can now easily generate images using texts. Recently, the focus has been shifted to text-to-video models [6, 18, 30, 38, 46, 48, 60, 76, 80, 84, + +85, 97] to generate videos by taking text descriptions as inputs. However, in some user cases, texts alone are not expressive enough to define the specific appearance of subjects [21, 71]. For example, as shown in Fig. 2, if we want to generate a video clip containing the dog as the third row, we need to use several attributive adjuncts to define the appearance of the dog in text prompts. Even with these extensive attributive adjuncts, models still cannot generate the desired appearance. Defining the appearance of the desired object with texts alone has the following flaws: 1) It is hard to enumerate all the desired attributes, and 2) The model cannot capture all attributes accurately with a long text. Compared to using texts, a more straightforward way to define the appearance is to provide reference images, termed image prompts. The image prompts are complementary to the text prompts and enrich the details that are hard to be depicted by text prompts. + +There are several attempts to introduce image prompts into text-to-image models, which can be roughly divided into two groups. One is to fine-tune parts of parameters using few-shot reference images [15, 21, 28, 51, 71], which contain the same objects captured under different circumstances. However, the requirement for the number of reference images is demanding as sometimes it is not practical to obtain multiple images of the same object. The other category [11, 42, 52, 83, 87, 88], aiming to address this limitation, proposes to embed image prompts into text-to-image models and the inference is tuning-free. Both of these two types of attempts achieve plausible results in generating images containing objects specified in image prompts. + +In this paper, we study a more challenging task, i.e., text-to-video generation with image prompts. The task has two main challenges: 1) Similar to text-to-image generation, the attributes of image prompts should be accurately captured and then reflected in the generated videos; 2) Different from text-to-image generation, we aim for the dynamic movement of the object rather than a static one. Directly adapting these methods to video domain results in mismatched appearance or unnatural degraded movements. To address these challenges, we proposed VideoBooth with elaborately designed coarse-to-fine visual embedding components: 1) Coarse visual embeddings via image encoder: An image encoder is trained to inject the image prompts into text embeddings; 2) Fine visual embeddings via attention injection: The image prompts are mapped to multiscale latent representations to control the generation process through cross-frame attentions of text-to-video models. + +Specifically, inspired by early attempts [42, 83] in text-to-image models with image prompts, we extract the CLIP image features of the provided image prompts using the pretrained CLIP model [66]. Then the extracted features are mapped into the text embedding space, which are inserted to replace parts of the original text embeddings. The well- + + eating snack inside big iron cage at home. + +![](images/65f9d4710b0b0a16c5b504f260541f524833c1526dba63867e813ca73eb9594a.jpg) +Figure 2. The Use of Image Prompts. We generate three video clips using different types of prompts: simple text prompt, long text prompt, and image prompt. We use the LLaVa model [55] to generate a text prompt describing the appearance of the image prompt. Using text prompts alone cannot fully capture the visual characteristics of the image prompt. + +trained encoder embeds the coarse appearance information of the given image prompts. However, coarse visual embedding is a universal embedding: 1) It only contains high-level semantic information, and 2) It is shared across all blocks with the same scale. As a result, some visual details are missing in the coarse visual embeddings. + +To further refine the generated details as well as maintain temporal consistency, different from the highly compacted coarse visual embeddings, multi-scale image prompts are injected into cross-frame attention modules in different layers. The image prompts provide spatial information as well as details with different granularities. On the one hand, keeping spatial information of the image prompts can retain more details. On the other hand, different cross-frame attention modules need detailed information at different scales. Specifically, the latent representations of image prompts are appended as additional keys and values to refine the details in the first generated frame. To propagate the refined first frame to the following frames to maintain temporal consistency, we then use the updated values of the first frame as values for the remaining frames. + +We set up a dedicated VideoBooth dataset to support the study of the new task. With each video, we provide an image prompt and a text prompt. Extensive experiments demonstrate the effectiveness of our proposed VideoBooth to generate videos with subjects specified in image prompts. As shown in Fig. 1, videos generated by VideoBooth better keep the visual attributes of image prompts. Besides, our proposed VideoBooth is tuning-free at inference time and videos can be generated with feed-forward passes only. The contributions are summarised as follows: + +- To our knowledge, we are the first to explore the task of video generation using image prompts without finetuning + +at inference time. We propose a dedicated dataset to support the task. Our proposed VideoBooth framework can generate consistent videos containing the subjects specified in image prompts. + +- We introduce a new coarse-to-fine visual embedding strategy by image encoder and attention injection, which better captures the characteristics of the image prompts. +- We propose a novel attention injection method, using the multi-scale image prompts with spatial information to refine the generated details. + +# 2. Related Work + +Text-to-Video Models take the text descriptions as inputs and generate clips of videos. Early explorations [34, 80] on text-to-video models are based on the idea of VQVAE. Make-A-Video [76] proposes to add temporal attention to the architecture of DALLE2 model [68]. Recently, the emergence of diffusion models [32, 70] boosts research on text-to-video models [1, 23, 30, 33, 60, 82, 97]. Video LDM [6] proposes to train the text-to-video models on Stable Diffusion with temporal attention and 3D convolution introduced to handle the temporal generation. Gen-1 [18] introduces depth maps to handle the temporal consistency of text-to-video models. Some methods [27, 95] resort to training separate modules for synthesizing motions. All of the methods initialize their models with pre-trained text-to-image models. Another paradigm of using text-to-image models is to directly apply Stable Diffusion [69] to few-shot or zero-shot settings. Tune-A-Video [84] adapts the self-attention into cross-frame attention and then finetunes the stable diffusion model on a video clip. Models trained in this way have the capability to transfer motions from original videos. Text2Video-Zero [48] proposes to generate videos by using correlated noise maps to improve consistency. Apart from video generation, diffusion models have been applied to video-to-video generations [8, 9, 19, 24, 36, 41, 54, 58, 64, 65, 74, 81, 89, 90, 94]. + +Customized Content Creation aims at generating images and videos using reference images [15]. For customized text-to-image generation, optimization-based methods [28, 35, 51, 53] are proposed to optimize the weights of the diffusion model. For example, Textual Inversion [21] optimizes the word embeddings, while DreamBooth [71] proposes to finetune the weights of Stable Diffusion as well. Optimization-based methods require several reference images with the same subject to avoid the overfitting of the model, which is demanding in real-world applications. The cost of finetuning hampers the practical usage of these methods. To address these limitations, encoder-based methods [61, 88, 91, 98] are proposed to learn a mapping network to embed the reference images. ELITE [83] proposes to learn a global mapping network and local mapping network to encode the images into word embeddings. Jia + +et al. [42] propose to use an additional cross-attention to embed the image features. With the trained encoder, the personalized generation can be achieved in a feed-forward pass. Some recent works [2, 11, 22, 25] combine the encoder-based model and finetuning-based model to improve the performance. BLIP-Diffusion [52] proposes to pretrain a multimodal encoder in a large-scale dataset and then finetune the model on the specific subject for inference. Customized image generation is also applied to place the objects into the user-specified scenes [4, 13, 50, 77, 93]. Also, some efforts [12, 14, 40, 72, 79, 86, 87, 92] have been made to personalized face generation. He et al. [29] propose to improve the performance from the data perspective. Some works [3, 47, 59] focus on composing multiple subjects in one image. Apart from works on image generation, there are some early attempts at personalized video manipulation. Make-A-Protagonist [96] edits an existing video in a personalized way using Stable Diffusion 2.1 to embed the image prompts. The motion of the original video is learned from Tune-A-Video [84]. VideoDreamer [10] proposes to generate personalized videos by generating the first frames using a finetuning-based method and then generating the video clip using the Text2Video-Zero [48]. Different from existing works, our proposed VideoBooth does not need to finetune any weights at the inference time. + +# 3. VideoBooth + +Our proposed VideoBooth aims at generating videos from an image prompt $I$ and a text prompt $T$ . The image prompt specifies the appearance of the subject. An overview of our proposed VideoBooth is illustrated in Fig. 3. The image prompt is fed into VideoBooth in two levels. At the coarse level, it is fed into a pretrained CLIP Image encoder to extract visual features. An encoder, composed of several MLP layers, is trained to map visual features into the space of text embeddings. The obtained embedding $f_{I}$ will be inserted into text embedding, which is extracted by feeding text prompt $T$ into CLIP text encoder. To further refine the synthesized details, we propose to inject image prompt $I$ into the cross-frame attention module in the pretrained video diffusion model. Specifically, we append latent representation $x_{I}^{t}$ of image prompt $I$ into the cross-frame attention. In this way, multi-scale visual details with spatial information are involved in the calculation of attention maps so that visual characteristics can be better preserved. Two ways of feeding image prompt corporate with each other in a coarse-to-fine manner. The encoder provides coarse visual embeddings of the image prompt, while the attention injection provides fine visual embeddings. + +# 3.1. Preliminary: Pretrained Text-to-Video Model + +Our proposed VideoBooth is developed based on the pretrained text-to-video model [30, 82, 84]. In this section, we + +![](images/a0f190a5e7a7e5ed2822bbbc8db28c83d41b178c151a9139dbf0e4f5d2492a40.jpg) +Figure 3. Overview of VideoBooth. VideoBooth generates videos by taking image prompts $I$ and text prompts $T$ as inputs. The image prompt is fed into the CLIP image encoder, followed by MLP layers. The obtained coarse visual embedding $f_{I}$ is then inserted into the text embeddings. The composed embeddings serve as the input for cross attention. The embedding extracted by the encoder provides a coarse encoding of the visual appearance of the image prompt. To further refine the details in the generated videos, at the fine level, we append the latent representation of the image prompt to the cross-frame attention as additional keys and values. Different cross-frame attention layers receive latent representations with different scales. The multi-scale features with spatial details refine the synthesized details. + +will briefly introduce the framework of text-to-video model. Inflated 2D Conv. To handle video data and capture the temporal correlation, 2D conv in the Stable Diffusion model is inflated to 3D conv. In this way, the U-Net can encode 3D features containing the temporal dimension. + +Cross-Frame Attention Module. Stable Diffusion has a self-attention module, where the features are enhanced by attending to themselves. To improve temporal consistency, the self-attention is modified into cross-frame attention. Specifically, the feature of each frame is enhanced by attending and referencing to the first frame and the previous frame. The cross-frame attention operates on both the spatial domain and temporal domain, thus the temporal consistency of the synthesized frames is improved. + +Temporal Attention Module. Apart from cross-frame attention, a temporal attention module is introduced to further improve temporal consistency. Temporal attention operates on temporal domain and attends to all frames. + +# 3.2. Coarse Visual Embeddings via Image Encoder + +Given an image prompt $I$ and text prompt $T$ , the generated video is supposed to be consistent with visual elements and textual elements. Inspired by previous attempts at image-based customization methods [42, 83], we propose to encode visual information of image prompts by an image encoder. The image prompt and text prompt complement each other. The image prompt provides visual characteristics of the desired subject in the video, and the text prompt provides other orthogonal information. The extracted visual embeddings are combined with text embeddings as the final + +embeddings for the cross-attention module. Specifically, the CLIP image encoder is employed to extract the visual features $f_{V}$ of image prompt $I$ . Since the discrepancy exists between the CLIP image and text embeddings, $f_{V}$ is then fed into MLP layers $F(\cdot)$ to map $f_{V}$ to the spaces of text embeddings. The final embedding $f_{I}$ for the image prompt is obtained as follows: + +$$ +f _ {V} = \operatorname {C L I P} _ {I} (I), f _ {I} = F \left(f _ {V}\right). \tag {1} +$$ + +As for the text prompt $T$ , we feed it into the CLIP text encoder to extract the text embedding $f_{T}$ : + +$$ +f _ {T} = \left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \dots , f _ {T} ^ {k}, \dots \right], \tag {2} +$$ + +where $f_{T}^{k}$ is the $k$ -th word embedding in the text prompt. + +To make the diffusion model generate videos conditioning on both text prompts and image prompts, we need to integrate these two embeddings, i.e., $f_{I}$ and $f_{T}$ . The idea is to replace the word embedding of the target subject with $f_{I}$ . Mathematically, $f_{I}$ and $f_{t}$ are fused to obtain the final text condition $c_{t}$ as follows: + +$$ +c _ {T} = \left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \dots , f _ {T} ^ {k - 1}, f _ {I}, f _ {T} ^ {k + n}, \dots \right], \tag {3} +$$ + +where $k$ is the token index of the target subject in the text embedding, and $n$ is the length of the text tokens for the target subject. For example, an image of a papillon dog is provided as an image prompt. To fuse the information from the image prompt and the text prompt "Papillon dog celebrates birthday with gifts", the word embeddings of the + +![](images/d2d02e2df2dcfdd3ed7d86c32c188766cc9aa2d352c4ea8e5f1297754d070d3e.jpg) +Figure 4. Fine Visual Embedding Refinement. We propose to inject the latent representation of image prompt (here we use the image for illustration purpose) directly into the cross-frame attention module. We use the keys and values from the image prompt to update the values of the first frame firstly. Then, the updated values of the first frame are used to update the remaining frames. Injecting the image prompt in the cross-frame attention helps to transfer the detailed visual characteristics of the image prompts to the synthesized frames. We perform the refinement in different cross-attention layers with different scales. + +"papillon dog" will be replaced with $f_{I}$ before they are fed into the cross-attention module of the diffusion models. + +During the training of the coarse stage, we fix the parameters of the CLIP image encoder, and train the MLP layers. To make the diffusion model accommodate with the composed text embeddings $c_{T}$ , we also finetune K and V projections (linear layers to map the input feature to the corresponding keys and values) in the cross-attention module. + +# 3.3. Fine Visual Embeddings via Attention Injection + +The well-trained image encoder embeds the coarse visual embeddings for image prompts and thus the synthesized videos contain the subjects specified in image prompts. However, the image encoder projects the image prompt into a flattened high-level representation, resulting in the loss of its detailed visual cues. Thus, some detailed visual characteristics in the image prompts may not be well preserved. To address this problem, a more effective way to preserve these details is to provide the model with the image prompts with spatial resolutions. + +To further refine the synthesized details, we propose to inject image prompts into the cross-frame attention of the text-to-video models. By injecting image prompts into the cross-frame attention, the image prompts are involved in the updates of the synthesized frames so that the model can directly borrow some visual cues from image prompts. + +Since text-to-video diffusion models operate in the latent space, we first feed the image prompt into the VAE of Stable Diffusion and get its latent representation $x_{I}$ . Moreover, since the sampling of the videos starts from the noise map, the latent in the intermediate timesteps contains the noises. If we append the clean latent $x_{I}$ of the image prompt + +to the cross-frame attention, the domain discrepancy exists. Therefore, we follow the diffusion forward process to add corresponding noises to $x_{I}$ : + +$$ +x _ {t} ^ {I} = \sqrt {\bar {\alpha} _ {t}} x _ {0} ^ {I} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {4} +$$ + +where $\overline{\alpha}$ is a hyperparameter determined by the denoising schedule and $\epsilon \sim N(0, I)$ . + +The cross-frame attention is used to improve the temporal consistency of the generated frames. For each frame, the key and value are the concatenation of the features of the first frame and the previous frame. Here, we introduce the image prompts as the additional keys and values for the frames. As shown in Fig. 4, we propose to update the values of the first frame firstly using the keys and values of the image prompts and the frame itself. Mathematically, the operation can be expressed as follows: + +$$ +V _ {0} ^ {\text {n e w}} = \operatorname {s o f t m a x} \left(\frac {Q _ {0} K ^ {T}}{\sqrt {d}}\right) \cdot V, \tag {5} +$$ + +$$ +K = \left[ K _ {I}, K _ {0} \right], V = \left[ V _ {I}, V _ {0} \right], +$$ + +where $K_{I}$ and $V_{I}$ are the keys and values obtained from the image prompts. The query, key, and value of the first frame are denoted $Q_{0}, K_{0}$ , and $V_{0}$ , respectively. It should be noted that we use a separately trained $\mathrm{K}$ and $\mathrm{V}$ projection for latent representations $x_{t}^{I}$ of image prompts because the image prompts have clean backgrounds, which are different from other frames. The parameters of the newly added $\mathrm{K}$ and $\mathrm{V}$ projections are initialized by original $\mathrm{K}$ and $\mathrm{V}$ projections. + +Then the updated first frame is used to refine the remaining frames. When updating the remaining frames, the keys used for calculating the attention maps are the original keys, while the values are the updated ones. The update is expressed as follows: + +$$ +V _ {i} ^ {\text {n e w}} = \operatorname {s o f t m a x} \left(\frac {Q _ {i} K ^ {T}}{\sqrt {d}}\right) \cdot V, \tag {6} +$$ + +$$ +K = [ K _ {0}, K _ {i - 1} ], V = [ V _ {0} ^ {\text {n e w}}, V _ {i - 1} ]. +$$ + +To sum up, in the attention injection, we update the values of the first frame using the image prompts first, and then use the updated first frame to update the other frames. In this way, the visual cues from the image prompts can be consistently propagated to all the frames. + +It should be noted that the diffusion model has multiple cross-frame attention layers with different scales. To inject multi-scale visual cues for better detail refinement, in different cross-frame attention layers, we feed latent representations of the image prompts with corresponding resolutions, which are obtained from different stages of the U-Net. + +# 3.4. Coarse-to-Fine Training Strategy + +The visual details of the image prompts are embedded into the final synthesized results in two stages: coarse visual embeddings using an image encoder and fine visual embedding + +by attention injection. We propose to train these two modules in a coarse-to-fine manner. In other words, we train the coarse image encoder and tune the parameters in the cross-attention first. After the model has the capability of generating videos containing the subjects specified in image prompts, we then train the attention injection module to embed image prompts into cross-frame attention layers. As we will show in the ablation study (Sec. 5.5), if these two modules are trained together, the fine attention injection module leaks the strong visual cues and the coarse encoder learns meaningless representations. As a result, in sampling phase, the image encoder for the coarse visual embedding cannot provide the coarse information and then the fine attention module cannot refine the details. Therefore, it is necessary to train VideoBooth in a coarse-to-fine manner. + +# 4. VideoBooth Dataset + +We establish the VideoBooth dataset to support the task of video generation using image prompts. We start from the WebVid dataset [5], a well-known open-source dataset for text-to-video generation. In the WebVid dataset, there is a text prompt with each video. In this paper, we study the task of generating a video clip from one text prompt and one image prompt. Hence, in addition to the original text prompt, we need to provide an image prompt for each video. We propose to segment the subjects from the first frame of the video using the Grounded-SAM (Grounded Segment Anything) [49, 57], and the segmented subjects are image prompts. The Grounded-SAM receives word prompts as inputs and generates segmentation masks for the target subjects specified in word prompts. To obtain the word prompt for the input to Grounded-SAM, we use the spaCy library to parse the noun chunks from the original text prompts, which are used as the word prompts. After the segmentation, we perform data filtering to ensure the data quality. We filter out small objects and large objects (those are almost the same size as the original video) according to the ratio of the object to the whole video. Also, since we focus on generating video clips containing moving objects, we further filter the videos containing moving objects. The keywords we used for filtering are dog, cat, bear, car, panda, tiger, horse, elephant, and lion. In the current version, we have processed 2.5M subset of the WebVid dataset. After data filtering, we have 48,724 video data pairs for training. We will process the full set of the WebVid dataset and include the filtered data in our VideoBooth dataset. + +To evaluate the performance, we also set up a test benchmark. The test benchmark consists of 650 test pairs. For each pair, an image prompt and a text prompt are provided. The test pairs are selected from the rest of the WebVid-10M dataset, which does not overlap with the training set. + +Table 1. Quantitative Comparisons. VideoBooth achieves the best image alignment and comparable text alignment performance. + +
MethodCLIP-Text ↑CLIP-Image ↑DINO ↑
Textual Inversion [21]29.974969.799545.3143
DreamBooth [71]30.687771.207852.9661
ELITE [83]30.088173.751858.9522
VideoBooth (Ours)30.096774.797165.0979
+ +# 5. Experiments + +# 5.1. Comparison Methods + +Textual Inversion [21] is a method for customized text-to-image generation. The appearance of the target subjects is embedded into the text embeddings. Concretely, a text token $S^{*}$ is optimized to represent the subject. We adapt it to the task of text-to-video generation by replacing the image model with the video model. + +DreamBooth [71] is also proposed for customized text-to-image generation. It injects the target subject into the text tokens as well as model weights. During the training, both the model weights and the special token $S*$ are optimized. + +ELITE [83] is an encoder-based method for personalized generation. An encoder is trained to embed the images into the text embeddings. Local mapping and global mapping are employed to transform the CLIP embedding of image prompts into the features, which are injected into the cross-attention module. We adapt and retrain the method using the same pretrained video model we use. + +# 5.2. Evaluation Metrics + +We use three metrics to evaluate the performance [83]. To measure the alignment of the generated videos and given text prompts, we use the CLIP-Text metric. The metric is calculated using the cosine similarity of the CLIP text embeddings of text prompts and CLIP image embeddings of the generated frames. For each video, the value is obtained by averaging values of the all frames. As for the evaluation of the similarity between the given image prompts and the generated videos, we adopt two metrics: CLIP-Image and DINO [7, 63]. The CLIP-Image metric is calculated by the cosine similarity between the CLIP image embedding of image prompts and generated frames. Since the CLIP model is trained to align image embeddings and text embeddings, we follow the practice in previous methods [13, 71] and use the DINO similarity as another indicator. DINO is trained to differentiate the differences between objects of the same classes. We use the ViT-S/16 model to extract the features of the image prompts and generated frames. The final score is obtained by averaging over all frames. + +# 5.3. Quantitative Comparisons + +We report quantitative results in Table 1. As shown in Table 1, our proposed VideoBooth achieves state-of-the-art + +![](images/101aaef392f59fa34a50a1f1252c07b23ed9633b77955d6aaffd503e198d7499.jpg) +Figure 5. Qualitative Comparison. VideoBooth effectively preserves the fidelity of image prompts and achieves better visual quality. + +![](images/bfb987e8911e53d525e68e5645bb2d6a1ecd4a2ee49a6a70241168f4e8008560.jpg) + +![](images/47f0a6bf7d23dd377306176abf7f243d2c6d4c70f3446d83ef095402c70cb6bf.jpg) + +image alignment performance compared to baseline methods. As for the alignment with the text prompts, our proposed VideoBooth has comparable performance with the baseline models. It should be noted that the CLIP-Text score of DreamBooth is significantly higher than the other methods. The reason lies in that the optimized token $S*$ in DreamBooth is inserted into the text embeddings, rather than replacing the original word embeddings like other methods. This would result in the generated videos of DreamBooth being highly related to the text prompts but having no correlation to the image prompts in some cases. We also conducted a user study, in which 25 users participated. Each user is presented with twelve groups of videos, and each group contains four videos generated by four methods. For each group, users are asked to make three choices: 1) which one has the best image alignment? 2) which one has the best text alignment? 3) which one has the best overall quality. Figure 6 summarizes the results. Our results are preferred by most users in all three dimensions. + +# 5.4. Qualitative Comparisons + +We show three visual comparisons on the generated video frames of our proposed VideoBooth and baseline methods in Fig. 5. In the first example, the model is supposed to generate videos containing the dog specified in the image laying on the ground. Textual Inversion cannot correctly embed the appearance of the dog and results in generating another totally different dog. DreamBooth and ELITE can embed the coarse appearance of the dog but the generated details vary from the image prompt. Our proposed VideoBooth successfully embeds the details of the image prompts into + +![](images/f0ce31bd0b3b4d54c441ad43f4c9d246a8585a9faa92bc7f203f1d5f5c47c58d.jpg) +Figure 6. User Study. Our proposed VideoBooth achieves the highest user preference ratios on all three dimensions. + +the synthesized videos. In the second example, the condition is to generate a cat playing with the yellow ball. All the methods can generate a yellow ball and a cat but only our proposed VideoBooth can accurately generate the cat having the appearance from the image prompt. As for the last example, Textual Inversion fails to generate a car. DreamBooth generates a distorted car in the bush. ELITE model can generate a car in the bush, but the color differs from the image prompt. By contrast, our model can generate videos having the same car in the bush. + +# 5.5. Ablation Study + +To evaluate the effectiveness of the proposed components, we perform three ablation studies. Due to the computational resources, we train these models on the subset of our training set. The quantitative metrics are shown in Table 2. + +Only Coarse Embeddings. This ablation model injects the image prompts with only coarse embeddings via Image Encoder. In the example shown in Fig. 7(a), the ablation model only encodes the coarse appearance of the image prompts. The pattern in the legs of the synthesized dog is different + +![](images/6943cb25f233daa454e7ce53cb62d883dd68912839fffc7f26cee3381f5d9595.jpg) +Dog as a chef at the table + +![](images/81c88a84ce53ecd6a5b863733caea3d8c80eccd1d46b9a148babd67f516327a4.jpg) +(a) Only Coarse Embeddings with Image Encoder + +![](images/ff789fb811d4423174fefae79ee89ca5b0148ed07b40b0cc25eba116cd8721dd.jpg) +(b) Only Fine Embeddings with Attention Injection + +![](images/0d0fcc35c6120ec2947a6b9e368dd970ea376de83ca1ccd765270919345403a0.jpg) +(c) Unified Training for Image Encoder and Attention Injection + +![](images/ce2655202b7f6a62180e6d2e91d944faeceeba5bceb2219dcc7842e57c3dd5be.jpg) +(d) Full Model +Figure 7. Ablation Study. (a) With only coarse embeddings from image encoder, generated patterns in the body of dog are different from image prompt. (b) With only fine embeddings from attention injection, there lacks coarse encodings of the dog for attention injection module to refine and thus the generated dog is distorted at later synthesized frames. (c) The unified training degrades the capability of image encoder and thus the dog is distorted. (d) The full model better keeps the all visual details of image prompt. + +from that in the image prompt. By contrast, the results of our full model show that our proposed model can transfer all the details in the image prompts to the synthesized videos. + +Only Fine Embeddings. In this ablation model, we only have fine embeddings of the image prompt in cross-frame attention layers. The main purpose of using fine embedding is to refine the coarse encoding of image prompts from the Image Encoder. Without coarse embeddings, the model with fine embeddings only cannot refine the details. As shown in Fig. 7(b), the first frame contains the exact appearance as the image prompt, but the temporal consistency cannot be guaranteed. The generated dog is distorted in the following frames. The reason is that the model trained in this way overfits the image prompt. In the first frame, the model can copy the information from image prompts. In the following frames, without the coarse embeddings, the generation of the appearance only relies on the propagation of the appearance from the first frame. + +The Necessity of Coarse-to-Fine Training. In Video-Booth, we propose the coarse-to-fine training strategy, i.e., train the coarse embeddings first and then train the attention + +Table 2. Ablation Study. The full model has the best scores. + +
VariantsCLIP-Image ↑DINO ↑
(a) Coarse Embeddings only75.436664.9568
(b) Fine Embeddings only75.555366.0378
(c) Unified Training75.825467.4201
(d) Full Model76.163169.7374
+ +injection module. In this ablation model, we train these two modules within one stage. The unified training makes the model rely heavily on the strong guidance provided in attention injection. In this way, the trained image encoder has limited capability. Thus, the model trained in this way has a similar behavior as the model with only fine embeddings. The model also overfits the image prompts from the attention injection. As shown in the first example in Fig. 7(c), the appearance in the first frame is correct, but the generated dog in the following frames is distorted. Due to the image encoder having limited capability to provide the correct coarse encoding of image prompts, the attention injection cannot refine the details. By contrast, the full model can generate consistent frames with all details well preserved. + +# 6. Discussion + +In this paper, we propose a novel framework VideoBooth to generate videos using image prompts and text prompts. The image prompts specify the appearance of the subjects. We inject the image prompts into the model in two modules: Coarse Embeddings via Image Encoder and Fine Embeddings via Attention Injection. The Image Encoder provides the coarse embeddings of image prompts for the refinement of the Attention Injection module. These two modules cooperate with each other and they are trained in a coarse-to-fine manner. Our proposed VideoBooth generates consistent videos containing the desired subjects. + +Limitations and Future Work. The model is trained with the image prompts that do not have very divergent views from target videos. Thus, if an image prompt with a back view is provided, we cannot generate videos from the front view. In future work, we can augment image prompts using image-to-3D models [56, 78] so that image prompts used for training will have different views from target videos. + +Potential Negative Societal Impacts. The model may be applied to generate fake videos, which can be potentially avoided by using deepfake detection methods. + +Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and the Google PhD Fellowship. This study is partially supported by the National Key R&D Program of China (2022ZD0160201), and Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 3 +[2] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06925, 2023. 3 +[3] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023. 3 +[4] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 3 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6 +[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 1, 3 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6 +[8] Duygu Ceylan, Chun-Hao Huang, and Niloy J. MiTRA. Pix2video: Video editing using image diffusion. arXiv:2303.12688, 2023. 3 +[9] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3 +[10] Hong Chen, Xin Wang, Guanning Zeng, Yipeng Zhang, Yuwei Zhou, Feilin Han, and Wenwu Zhu. Videodreamer: Customized multi-subject text-to-video generation with disen-mix finetuning. arXiv preprint arXiv:2311.00990, 2023. 3 +[11] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Disentangled parameter-efficient tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2, 3 +[12] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photoverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 3 +[13] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 3, 6 + +[14] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3 +[15] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3 +[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1 +[17] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.1 +[18] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 1, 3 +[19] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Ccredit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3 +[20] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1 +[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2, 3, 6 +[22] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3 +[23] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3 +[24] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3 +[25] Yuan Gong, Youxin Pang, Xiaodong Cun, Menghan Xia, Haoxin Chen, Longyue Wang, Yong Zhang, Xintao Wang, Ying Shan, and Yujiu Yang. Talecrafter: Interactive story visualization with multiple characters. arXiv preprint arXiv:2305.18247, 2023. 3 +[26] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 1 +[27] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 + +[28] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023. 2, 3 +[29] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 3 +[30] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 3 +[31] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1 +[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3 +[33] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3 +[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[35] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3 +[36] Zhihao Hu and Dong Xu. Videocontrolnet: A motion-guided video-to-video translation framework by using diffusion model with controlnet. arXiv preprint arXiv:2307.14073, 2023. 3 +[37] Ziqi Huang, Kelvin CK Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6080-6090, 2023. 1 +[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. arXiv preprint arXiv:2311.17982, 2023. 1 +[39] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 1 +[40] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3 +[41] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zeroshot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3 +[42] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and + +Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2, 3, 4 +[43] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13799-13808, 2021. 1 +[44] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. +[45] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[46] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1 +[47] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Learning object level concepts using multi-concept prompt learning. arXiv preprint arXiv:2310.12274, 2023. 3 +[48] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 1, 3 +[49] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6 +[50] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17089-17099, 2023. 3 +[51] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. 2023. 2, 3 +[52] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 2, 3 +[53] Yuheng Li, Haotian Liu, Yangming Wen, and Yong Jae Lee. Generate anything anywhere in any scene. arXiv preprint arXiv:2306.17154, 2023. 3 +[54] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. In arXiv, 2023. 3 +[55] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2 +[56] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: + +Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 8 +[57] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6 +[58] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3 +[59] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. arXiv preprint arXiv:2305.19327, 2023. 3 +[60] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 1, 3 +[61] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 3 +[62] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1 +[63] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[64] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3 +[65] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3 +[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[67] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 +[68] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3 + +[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 3 +[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3 +[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3, 6 +[72] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3 +[73] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1 +[74] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. arXiv preprint arXiv:2303.07945, 2023. 3 +[75] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 1 +[76] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3 +[77] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 3 +[78] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 8 +[79] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3 +[80] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3 +[81] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3 + +[82] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 3 +[83] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 2, 3, 4, 6 +[84] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 3 +[85] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. arXiv preprint arXiv:2312.07537, 2023. 2 +[86] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. Singleinsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 3 +[87] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3 +[88] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking "text" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 2, 3 +[89] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video editing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 3 +[90] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3 +[91] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[92] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 3 +[93] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3 +[94] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3 +[95] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 3 + +[96] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3 +[97] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3 +[98] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 3 \ No newline at end of file diff --git a/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/images.zip b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd0325a4fa0721d3a7c768291ddebdc54f5f19cc --- /dev/null +++ b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adf958355395f5554342bcd79e73ca411ef9a59e9f0d8f4677054dc440999522 +size 636819 diff --git a/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/layout.json b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..630d973e00c5bdfba05d348df9f2b216f6e7e16d --- /dev/null +++ b/2024/VideoBooth_ Diffusion-based Video Generation with Image Prompts/layout.json @@ -0,0 +1,10533 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 73, + 509, + 91 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 73, + 509, + 91 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 509, + 91 + ], + "type": "text", + "content": "VideoBooth: Diffusion-based Video Generation with Image Prompts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 113, + 212, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 113, + 212, + 128 + ], + "spans": [ + { + "bbox": [ + 138, + 113, + 212, + 128 + ], + "type": "text", + "content": "Yuming Jiang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 113, + 293, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 113, + 293, + 128 + ], + "spans": [ + { + "bbox": [ + 225, + 113, + 293, + 128 + ], + "type": "text", + "content": "Tianxing Wu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 113, + 373, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 113, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 373, + 128 + ], + "type": "text", + "content": "Shuai Yang" + }, + { + "bbox": [ + 307, + 113, + 373, + 128 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 385, + 113, + 453, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 113, + 453, + 128 + ], + "spans": [ + { + "bbox": [ + 385, + 113, + 453, + 128 + ], + "type": "text", + "content": "Chenyang Si" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 148, + 128, + 206, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 128, + 206, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 128, + 206, + 140 + ], + "type": "text", + "content": "Dahua Lin" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 128, + 265, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 128, + 265, + 141 + ], + "spans": [ + { + "bbox": [ + 219, + 128, + 265, + 141 + ], + "type": "text", + "content": "Yu Qiao2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 277, + 128, + 371, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 128, + 371, + 141 + ], + "spans": [ + { + "bbox": [ + 277, + 128, + 371, + 141 + ], + "type": "text", + "content": "Chen Change Loy1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 384, + 128, + 446, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 128, + 446, + 141 + ], + "spans": [ + { + "bbox": [ + 384, + 128, + 446, + 141 + ], + "type": "text", + "content": "Ziwei Liu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 124, + 141, + 468, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 141, + 468, + 156 + ], + "spans": [ + { + "bbox": [ + 124, + 141, + 468, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 141, + 468, + 156 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 124, + 141, + 468, + 156 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 141, + 468, + 156 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 156, + 454, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 156, + 454, + 170 + ], + "spans": [ + { + "bbox": [ + 139, + 156, + 454, + 170 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 156, + 454, + 170 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 154, + 171, + 433, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 171, + 433, + 183 + ], + "spans": [ + { + "bbox": [ + 154, + 171, + 433, + 183 + ], + "type": "text", + "content": "https://vchitec.github.io/VideoBooth-project/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 188, + 123, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 188, + 123, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 188, + 123, + 198 + ], + "type": "text", + "content": "Image Prompt" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 189, + 220, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 189, + 220, + 199 + ], + "spans": [ + { + "bbox": [ + 126, + 189, + 220, + 199 + ], + "type": "text", + "content": "A horse eating grass." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 65, + 200, + 119, + 253 + ], + "blocks": [ + { + "bbox": [ + 65, + 200, + 119, + 253 + ], + "lines": [ + { + "bbox": [ + 65, + 200, + 119, + 253 + ], + "spans": [ + { + "bbox": [ + 65, + 200, + 119, + 253 + ], + "type": "image", + "image_path": "0ee099b10ef93093afb05ebce906facfcd700ee993c6e0fc47d1555a48328cc5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 124, + 201, + 326, + 253 + ], + "blocks": [ + { + "bbox": [ + 124, + 201, + 326, + 253 + ], + "lines": [ + { + "bbox": [ + 124, + 201, + 326, + 253 + ], + "spans": [ + { + "bbox": [ + 124, + 201, + 326, + 253 + ], + "type": "image", + "image_path": "9ed8eed6772cc25c9c26de9c36835f5dc0c5925cfdbb8c848306503bbdc2841f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 332, + 188, + 473, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 188, + 473, + 199 + ], + "spans": [ + { + "bbox": [ + 332, + 188, + 473, + 199 + ], + "type": "text", + "content": "Horse grazes in snowy meadow" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 329, + 200, + 529, + 253 + ], + "blocks": [ + { + "bbox": [ + 329, + 200, + 529, + 253 + ], + "lines": [ + { + "bbox": [ + 329, + 200, + 529, + 253 + ], + "spans": [ + { + "bbox": [ + 329, + 200, + 529, + 253 + ], + "type": "image", + "image_path": "c6cde38fbdb97095e5490a53d35d42928e2c79e795dc4506fda21f9229a3b6d4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 256, + 123, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 256, + 123, + 267 + ], + "spans": [ + { + "bbox": [ + 59, + 256, + 123, + 267 + ], + "type": "text", + "content": "Image Prompt" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 129, + 256, + 323, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 256, + 323, + 266 + ], + "spans": [ + { + "bbox": [ + 129, + 256, + 323, + 266 + ], + "type": "text", + "content": "Portrait of a dog, looks out the car window." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 332, + 256, + 481, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 256, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 332, + 256, + 481, + 267 + ], + "type": "text", + "content": "Dog walking in the green farm 4k" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 76, + 268, + 108, + 321 + ], + "blocks": [ + { + "bbox": [ + 76, + 268, + 108, + 321 + ], + "lines": [ + { + "bbox": [ + 76, + 268, + 108, + 321 + ], + "spans": [ + { + "bbox": [ + 76, + 268, + 108, + 321 + ], + "type": "image", + "image_path": "f23977d3089665deeae489cc80e070f51668bb3a40054bead28e0362b8599e7c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 124, + 269, + 326, + 323 + ], + "blocks": [ + { + "bbox": [ + 124, + 269, + 326, + 323 + ], + "lines": [ + { + "bbox": [ + 124, + 269, + 326, + 323 + ], + "spans": [ + { + "bbox": [ + 124, + 269, + 326, + 323 + ], + "type": "image", + "image_path": "f3cd9e2f7fbfeaf24cc2a8cae94e3cedb3683d0e01c4d402fbf68e3eab9e3e0b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 329, + 269, + 529, + 321 + ], + "blocks": [ + { + "bbox": [ + 329, + 269, + 529, + 321 + ], + "lines": [ + { + "bbox": [ + 329, + 269, + 529, + 321 + ], + "spans": [ + { + "bbox": [ + 329, + 269, + 529, + 321 + ], + "type": "image", + "image_path": "7d199bbdf497d60bdd0ee3de9e36de32e016d9272b55ab9010f8b18711019365.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 59, + 326, + 123, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 326, + 123, + 336 + ], + "spans": [ + { + "bbox": [ + 59, + 326, + 123, + 336 + ], + "type": "text", + "content": "Image Prompt" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 129, + 326, + 243, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 326, + 243, + 336 + ], + "spans": [ + { + "bbox": [ + 129, + 326, + 243, + 336 + ], + "type": "text", + "content": "Cat is looking at a laptop." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 332, + 326, + 511, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 326, + 511, + 336 + ], + "spans": [ + { + "bbox": [ + 332, + 326, + 511, + 336 + ], + "type": "text", + "content": "Close up of cat on top of a vintage chair" + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 65, + 338, + 118, + 389 + ], + "blocks": [ + { + "bbox": [ + 65, + 338, + 118, + 389 + ], + "lines": [ + { + "bbox": [ + 65, + 338, + 118, + 389 + ], + "spans": [ + { + "bbox": [ + 65, + 338, + 118, + 389 + ], + "type": "image", + "image_path": "59453a6f2e07afe2a46e8d9d6cfb2e184a0f56b4bc90702ed376c36b71448562.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 124, + 338, + 326, + 388 + ], + "blocks": [ + { + "bbox": [ + 124, + 338, + 326, + 388 + ], + "lines": [ + { + "bbox": [ + 124, + 338, + 326, + 388 + ], + "spans": [ + { + "bbox": [ + 124, + 338, + 326, + 388 + ], + "type": "image", + "image_path": "196f7f5ab890e3f4cb7f8268c1b58a366095ce4d2c8e9c7406cd41ba3f71f85d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 329, + 338, + 529, + 389 + ], + "blocks": [ + { + "bbox": [ + 329, + 338, + 529, + 389 + ], + "lines": [ + { + "bbox": [ + 329, + 338, + 529, + 389 + ], + "spans": [ + { + "bbox": [ + 329, + 338, + 529, + 389 + ], + "type": "image", + "image_path": "5a897094ff13e8930d3047e0bb97aea526a38579d275e9a79b5345b5c5f50a83.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 59, + 399, + 123, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 399, + 123, + 409 + ], + "spans": [ + { + "bbox": [ + 59, + 399, + 123, + 409 + ], + "type": "text", + "content": "Image Prompt" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 129, + 392, + 324, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 392, + 324, + 412 + ], + "spans": [ + { + "bbox": [ + 129, + 392, + 324, + 412 + ], + "type": "text", + "content": "Elephant drinking water in masai mara reserve, kenya" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 332, + 392, + 494, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 392, + 494, + 411 + ], + "spans": [ + { + "bbox": [ + 332, + 392, + 494, + 411 + ], + "type": "text", + "content": "Elephant walk in the yellow grass of savannah" + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 66, + 413, + 118, + 459 + ], + "blocks": [ + { + "bbox": [ + 66, + 413, + 118, + 459 + ], + "lines": [ + { + "bbox": [ + 66, + 413, + 118, + 459 + ], + "spans": [ + { + "bbox": [ + 66, + 413, + 118, + 459 + ], + "type": "image", + "image_path": "b7cbbfe9b71d6ab0c358d6b23d921e69b0eb55d377366f6ca858990ea19a0dd2.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 464, + 539, + 475 + ], + "lines": [ + { + "bbox": [ + 53, + 464, + 539, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 539, + 475 + ], + "type": "text", + "content": "Figure 1. Videos synthesized by image prompts. Our VideoBooth generates videos with the subjects specified in the image prompts." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 124, + 412, + 326, + 462 + ], + "blocks": [ + { + "bbox": [ + 124, + 412, + 326, + 462 + ], + "lines": [ + { + "bbox": [ + 124, + 412, + 326, + 462 + ], + "spans": [ + { + "bbox": [ + 124, + 412, + 326, + 462 + ], + "type": "image", + "image_path": "cdcaec6ff80b79900beb17dd6efaaa37ef9d3a6e3a72517779d25deecf9f63a6.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 329, + 412, + 529, + 462 + ], + "blocks": [ + { + "bbox": [ + 329, + 412, + 529, + 462 + ], + "lines": [ + { + "bbox": [ + 329, + 412, + 529, + 462 + ], + "spans": [ + { + "bbox": [ + 329, + 412, + 529, + 462 + ], + "type": "image", + "image_path": "80078ebce751182a79e080233234349871584f74028c9dd32f3747b42b4ac0f1.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 143, + 480, + 192, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 480, + 192, + 493 + ], + "spans": [ + { + "bbox": [ + 143, + 480, + 192, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 506, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 696 + ], + "type": "text", + "content": "Text-driven video generation witnesses rapid progress. However, merely using text prompts is not enough to depict the desired subject appearance that accurately aligns with users' intents, especially for customized content creation. In this paper, we study the task of video generation with image prompts, which provide more accurate and direct content control beyond the text prompts. Specifically, we propose a feed-forward framework VideoBooth, with two dedicated designs: 1) We propose to embed image prompts in a coarse-to-fine manner. Coarse visual embeddings from image encoder provide high-level encodings of image prompts, while fine visual embeddings from the proposed attention injection module provide multi-scale and detailed encoding of image prompts. These two complementary embeddings can faithfully capture the desired appearance. 2) In the attention injection module at fine level, multi-scale" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 482, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 548, + 601 + ], + "type": "text", + "content": "image prompts are fed into different cross-frame attention layers as additional keys and values. This extra spatial information refines the details in the first frame and then it is propagated to the remaining frames, which maintains temporal consistency. Extensive experiments demonstrate that VideoBooth achieves state-of-the-art performance in generating customized high-quality videos with subjects specified in image prompts. Notably, VideoBooth is a generalizable framework where a single model works for a wide range of image prompts with only feed-forward passes." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 631, + 387, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 631, + 387, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 631, + 387, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Text-to-image models [16, 17, 20, 26, 31, 37, 39, 43-45, 62, 67-69, 73, 75] have attracted substantial attention. With Stable Diffusion [69], we can now easily generate images using texts. Recently, the focus has been shifted to text-to-video models [6, 18, 30, 38, 46, 48, 60, 76, 80, 84," + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 702, + 258, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 258, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 258, + 712 + ], + "type": "text", + "content": "*This work was done when Shuai Yang was in S-Lab, NTU." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6689" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "content": "85, 97] to generate videos by taking text descriptions as inputs. However, in some user cases, texts alone are not expressive enough to define the specific appearance of subjects [21, 71]. For example, as shown in Fig. 2, if we want to generate a video clip containing the dog as the third row, we need to use several attributive adjuncts to define the appearance of the dog in text prompts. Even with these extensive attributive adjuncts, models still cannot generate the desired appearance. Defining the appearance of the desired object with texts alone has the following flaws: 1) It is hard to enumerate all the desired attributes, and 2) The model cannot capture all attributes accurately with a long text. Compared to using texts, a more straightforward way to define the appearance is to provide reference images, termed image prompts. The image prompts are complementary to the text prompts and enrich the details that are hard to be depicted by text prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 278, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 434 + ], + "type": "text", + "content": "There are several attempts to introduce image prompts into text-to-image models, which can be roughly divided into two groups. One is to fine-tune parts of parameters using few-shot reference images [15, 21, 28, 51, 71], which contain the same objects captured under different circumstances. However, the requirement for the number of reference images is demanding as sometimes it is not practical to obtain multiple images of the same object. The other category [11, 42, 52, 83, 87, 88], aiming to address this limitation, proposes to embed image prompts into text-to-image models and the inference is tuning-free. Both of these two types of attempts achieve plausible results in generating images containing objects specified in image prompts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 436, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 639 + ], + "type": "text", + "content": "In this paper, we study a more challenging task, i.e., text-to-video generation with image prompts. The task has two main challenges: 1) Similar to text-to-image generation, the attributes of image prompts should be accurately captured and then reflected in the generated videos; 2) Different from text-to-image generation, we aim for the dynamic movement of the object rather than a static one. Directly adapting these methods to video domain results in mismatched appearance or unnatural degraded movements. To address these challenges, we proposed VideoBooth with elaborately designed coarse-to-fine visual embedding components: 1) Coarse visual embeddings via image encoder: An image encoder is trained to inject the image prompts into text embeddings; 2) Fine visual embeddings via attention injection: The image prompts are mapped to multiscale latent representations to control the generation process through cross-frame attentions of text-to-video models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "Specifically, inspired by early attempts [42, 83] in text-to-image models with image prompts, we extract the CLIP image features of the provided image prompts using the pretrained CLIP model [66]. Then the extracted features are mapped into the text embedding space, which are inserted to replace parts of the original text embeddings. The well-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 329, + 72, + 536, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 72, + 536, + 83 + ], + "spans": [ + { + "bbox": [ + 329, + 72, + 536, + 83 + ], + "type": "text", + "content": " eating snack inside big iron cage at home." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 84, + 538, + 244 + ], + "blocks": [ + { + "bbox": [ + 315, + 84, + 538, + 244 + ], + "lines": [ + { + "bbox": [ + 315, + 84, + 538, + 244 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 538, + 244 + ], + "type": "image", + "image_path": "65f9d4710b0b0a16c5b504f260541f524833c1526dba63867e813ca73eb9594a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 246, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 312 + ], + "type": "text", + "content": "Figure 2. The Use of Image Prompts. We generate three video clips using different types of prompts: simple text prompt, long text prompt, and image prompt. We use the LLaVa model [55] to generate a text prompt describing the appearance of the image prompt. Using text prompts alone cannot fully capture the visual characteristics of the image prompt." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "type": "text", + "content": "trained encoder embeds the coarse appearance information of the given image prompts. However, coarse visual embedding is a universal embedding: 1) It only contains high-level semantic information, and 2) It is shared across all blocks with the same scale. As a result, some visual details are missing in the coarse visual embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 390, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 569 + ], + "type": "text", + "content": "To further refine the generated details as well as maintain temporal consistency, different from the highly compacted coarse visual embeddings, multi-scale image prompts are injected into cross-frame attention modules in different layers. The image prompts provide spatial information as well as details with different granularities. On the one hand, keeping spatial information of the image prompts can retain more details. On the other hand, different cross-frame attention modules need detailed information at different scales. Specifically, the latent representations of image prompts are appended as additional keys and values to refine the details in the first generated frame. To propagate the refined first frame to the following frames to maintain temporal consistency, we then use the updated values of the first frame as values for the remaining frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 689 + ], + "type": "text", + "content": "We set up a dedicated VideoBooth dataset to support the study of the new task. With each video, we provide an image prompt and a text prompt. Extensive experiments demonstrate the effectiveness of our proposed VideoBooth to generate videos with subjects specified in image prompts. As shown in Fig. 1, videos generated by VideoBooth better keep the visual attributes of image prompts. Besides, our proposed VideoBooth is tuning-free at inference time and videos can be generated with feed-forward passes only. The contributions are summarised as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "- To our knowledge, we are the first to explore the task of video generation using image prompts without finetuning" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6690" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 120 + ], + "type": "text", + "content": "at inference time. We propose a dedicated dataset to support the task. Our proposed VideoBooth framework can generate consistent videos containing the subjects specified in image prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 286, + 192 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 121, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 286, + 156 + ], + "type": "text", + "content": "- We introduce a new coarse-to-fine visual embedding strategy by image encoder and attention injection, which better captures the characteristics of the image prompts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 157, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 286, + 192 + ], + "type": "text", + "content": "- We propose a novel attention injection method, using the multi-scale image prompts with spatial information to refine the generated details." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 203, + 133, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 133, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 133, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 223, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 521 + ], + "type": "text", + "content": "Text-to-Video Models take the text descriptions as inputs and generate clips of videos. Early explorations [34, 80] on text-to-video models are based on the idea of VQVAE. Make-A-Video [76] proposes to add temporal attention to the architecture of DALLE2 model [68]. Recently, the emergence of diffusion models [32, 70] boosts research on text-to-video models [1, 23, 30, 33, 60, 82, 97]. Video LDM [6] proposes to train the text-to-video models on Stable Diffusion with temporal attention and 3D convolution introduced to handle the temporal generation. Gen-1 [18] introduces depth maps to handle the temporal consistency of text-to-video models. Some methods [27, 95] resort to training separate modules for synthesizing motions. All of the methods initialize their models with pre-trained text-to-image models. Another paradigm of using text-to-image models is to directly apply Stable Diffusion [69] to few-shot or zero-shot settings. Tune-A-Video [84] adapts the self-attention into cross-frame attention and then finetunes the stable diffusion model on a video clip. Models trained in this way have the capability to transfer motions from original videos. Text2Video-Zero [48] proposes to generate videos by using correlated noise maps to improve consistency. Apart from video generation, diffusion models have been applied to video-to-video generations [8, 9, 19, 24, 36, 41, 54, 58, 64, 65, 74, 81, 89, 90, 94]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 287, + 713 + ], + "type": "text", + "content": "Customized Content Creation aims at generating images and videos using reference images [15]. For customized text-to-image generation, optimization-based methods [28, 35, 51, 53] are proposed to optimize the weights of the diffusion model. For example, Textual Inversion [21] optimizes the word embeddings, while DreamBooth [71] proposes to finetune the weights of Stable Diffusion as well. Optimization-based methods require several reference images with the same subject to avoid the overfitting of the model, which is demanding in real-world applications. The cost of finetuning hampers the practical usage of these methods. To address these limitations, encoder-based methods [61, 88, 91, 98] are proposed to learn a mapping network to embed the reference images. ELITE [83] proposes to learn a global mapping network and local mapping network to encode the images into word embeddings. Jia" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": "et al. [42] propose to use an additional cross-attention to embed the image features. With the trained encoder, the personalized generation can be achieved in a feed-forward pass. Some recent works [2, 11, 22, 25] combine the encoder-based model and finetuning-based model to improve the performance. BLIP-Diffusion [52] proposes to pretrain a multimodal encoder in a large-scale dataset and then finetune the model on the specific subject for inference. Customized image generation is also applied to place the objects into the user-specified scenes [4, 13, 50, 77, 93]. Also, some efforts [12, 14, 40, 72, 79, 86, 87, 92] have been made to personalized face generation. He et al. [29] propose to improve the performance from the data perspective. Some works [3, 47, 59] focus on composing multiple subjects in one image. Apart from works on image generation, there are some early attempts at personalized video manipulation. Make-A-Protagonist [96] edits an existing video in a personalized way using Stable Diffusion 2.1 to embed the image prompts. The motion of the original video is learned from Tune-A-Video [84]. VideoDreamer [10] proposes to generate personalized videos by generating the first frames using a finetuning-based method and then generating the video clip using the Text2Video-Zero [48]. Different from existing works, our proposed VideoBooth does not need to finetune any weights at the inference time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 381, + 382, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 382, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 382, + 393 + ], + "type": "text", + "content": "3. VideoBooth" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": "Our proposed VideoBooth aims at generating videos from an image prompt " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": ". The image prompt specifies the appearance of the subject. An overview of our proposed VideoBooth is illustrated in Fig. 3. The image prompt is fed into VideoBooth in two levels. At the coarse level, it is fed into a pretrained CLIP Image encoder to extract visual features. An encoder, composed of several MLP layers, is trained to map visual features into the space of text embeddings. The obtained embedding " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": " will be inserted into text embedding, which is extracted by feeding text prompt " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": " into CLIP text encoder. To further refine the synthesized details, we propose to inject image prompt " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": " into the cross-frame attention module in the pretrained video diffusion model. Specifically, we append latent representation " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "x_{I}^{t}" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": " of image prompt " + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": " into the cross-frame attention. In this way, multi-scale visual details with spatial information are involved in the calculation of attention maps so that visual characteristics can be better preserved. Two ways of feeding image prompt corporate with each other in a coarse-to-fine manner. The encoder provides coarse visual embeddings of the image prompt, while the attention injection provides fine visual embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 671, + 542, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 542, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 542, + 683 + ], + "type": "text", + "content": "3.1. Preliminary: Pretrained Text-to-Video Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our proposed VideoBooth is developed based on the pretrained text-to-video model [30, 82, 84]. In this section, we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6691" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 79, + 521, + 277 + ], + "blocks": [ + { + "bbox": [ + 73, + 79, + 521, + 277 + ], + "lines": [ + { + "bbox": [ + 73, + 79, + 521, + 277 + ], + "spans": [ + { + "bbox": [ + 73, + 79, + 521, + 277 + ], + "type": "image", + "image_path": "a0f190a5e7a7e5ed2822bbbc8db28c83d41b178c151a9139dbf0e4f5d2492a40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "text", + "content": "Figure 3. Overview of VideoBooth. VideoBooth generates videos by taking image prompts " + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "text", + "content": " and text prompts " + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "text", + "content": " as inputs. The image prompt is fed into the CLIP image encoder, followed by MLP layers. The obtained coarse visual embedding " + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 45, + 277, + 545, + 344 + ], + "type": "text", + "content": " is then inserted into the text embeddings. The composed embeddings serve as the input for cross attention. The embedding extracted by the encoder provides a coarse encoding of the visual appearance of the image prompt. To further refine the details in the generated videos, at the fine level, we append the latent representation of the image prompt to the cross-frame attention as additional keys and values. Different cross-frame attention layers receive latent representations with different scales. The multi-scale features with spatial details refine the synthesized details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "text", + "content": "will briefly introduce the framework of text-to-video model. Inflated 2D Conv. To handle video data and capture the temporal correlation, 2D conv in the Stable Diffusion model is inflated to 3D conv. In this way, the U-Net can encode 3D features containing the temporal dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": "Cross-Frame Attention Module. Stable Diffusion has a self-attention module, where the features are enhanced by attending to themselves. To improve temporal consistency, the self-attention is modified into cross-frame attention. Specifically, the feature of each frame is enhanced by attending and referencing to the first frame and the previous frame. The cross-frame attention operates on both the spatial domain and temporal domain, thus the temporal consistency of the synthesized frames is improved." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 569 + ], + "type": "text", + "content": "Temporal Attention Module. Apart from cross-frame attention, a temporal attention module is introduced to further improve temporal consistency. Temporal attention operates on temporal domain and attends to all frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 287, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 588 + ], + "type": "text", + "content": "3.2. Coarse Visual Embeddings via Image Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Given an image prompt " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and text prompt " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", the generated video is supposed to be consistent with visual elements and textual elements. Inspired by previous attempts at image-based customization methods [42, 83], we propose to encode visual information of image prompts by an image encoder. The image prompt and text prompt complement each other. The image prompt provides visual characteristics of the desired subject in the video, and the text prompt provides other orthogonal information. The extracted visual embeddings are combined with text embeddings as the final" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": "embeddings for the cross-attention module. Specifically, the CLIP image encoder is employed to extract the visual features " + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "inline_equation", + "content": "f_{V}" + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": " of image prompt " + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": ". Since the discrepancy exists between the CLIP image and text embeddings, " + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "inline_equation", + "content": "f_{V}" + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": " is then fed into MLP layers " + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": " to map " + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "inline_equation", + "content": "f_{V}" + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": " to the spaces of text embeddings. The final embedding " + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": " for the image prompt is obtained as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 363, + 447, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 447, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 363, + 447, + 545, + 460 + ], + "type": "interline_equation", + "content": "f _ {V} = \\operatorname {C L I P} _ {I} (I), f _ {I} = F \\left(f _ {V}\\right). \\tag {1}", + "image_path": "51b1168999fdfcea7a0325e6e0e01997dd1da1a9d8fd34d1a9c1316e4602ccc8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "type": "text", + "content": "As for the text prompt " + }, + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "type": "text", + "content": ", we feed it into the CLIP text encoder to extract the text embedding " + }, + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 305, + 470, + 545, + 494 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 502, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 502, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 373, + 502, + 545, + 517 + ], + "type": "interline_equation", + "content": "f _ {T} = \\left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \\dots , f _ {T} ^ {k}, \\dots \\right], \\tag {2}", + "image_path": "6d8d8263dd80438da68b76569064a9a67f74c1497b5601a1d5d85ee0a69fbe6e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "type": "inline_equation", + "content": "f_{T}^{k}" + }, + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 525, + 535, + 538 + ], + "type": "text", + "content": "-th word embedding in the text prompt." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": "To make the diffusion model generate videos conditioning on both text prompts and image prompts, we need to integrate these two embeddings, i.e., " + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": ". The idea is to replace the word embedding of the target subject with " + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": ". Mathematically, " + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": " are fused to obtain the final text condition " + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 304, + 538, + 545, + 609 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 617, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 617, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 348, + 617, + 545, + 633 + ], + "type": "interline_equation", + "content": "c _ {T} = \\left[ f _ {T} ^ {0}, f _ {T} ^ {1}, \\dots , f _ {T} ^ {k - 1}, f _ {I}, f _ {T} ^ {k + n}, \\dots \\right], \\tag {3}", + "image_path": "50c36723546146dc23c7c0b9d0e91f5e0222d3eb3a3b4bb75cb8dac5f6fde614.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " is the token index of the target subject in the text embedding, and " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " is the length of the text tokens for the target subject. For example, an image of a papillon dog is provided as an image prompt. To fuse the information from the image prompt and the text prompt \"Papillon dog celebrates birthday with gifts\", the word embeddings of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6692" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 282, + 194 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 282, + 194 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 282, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 282, + 194 + ], + "type": "image", + "image_path": "d2d02e2df2dcfdd3ed7d86c32c188766cc9aa2d352c4ea8e5f1297754d070d3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 289, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 289, + 312 + ], + "type": "text", + "content": "Figure 4. Fine Visual Embedding Refinement. We propose to inject the latent representation of image prompt (here we use the image for illustration purpose) directly into the cross-frame attention module. We use the keys and values from the image prompt to update the values of the first frame firstly. Then, the updated values of the first frame are used to update the remaining frames. Injecting the image prompt in the cross-frame attention helps to transfer the detailed visual characteristics of the image prompts to the synthesized frames. We perform the refinement in different cross-attention layers with different scales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 327, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 351 + ], + "type": "text", + "content": "\"papillon dog\" will be replaced with " + }, + { + "bbox": [ + 47, + 327, + 287, + 351 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 47, + 327, + 287, + 351 + ], + "type": "text", + "content": " before they are fed into the cross-attention module of the diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "type": "text", + "content": "During the training of the coarse stage, we fix the parameters of the CLIP image encoder, and train the MLP layers. To make the diffusion model accommodate with the composed text embeddings " + }, + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "type": "text", + "content": ", we also finetune K and V projections (linear layers to map the input feature to the corresponding keys and values) in the cross-attention module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 287, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 287, + 445 + ], + "type": "text", + "content": "3.3. Fine Visual Embeddings via Attention Injection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 569 + ], + "type": "text", + "content": "The well-trained image encoder embeds the coarse visual embeddings for image prompts and thus the synthesized videos contain the subjects specified in image prompts. However, the image encoder projects the image prompt into a flattened high-level representation, resulting in the loss of its detailed visual cues. Thus, some detailed visual characteristics in the image prompts may not be well preserved. To address this problem, a more effective way to preserve these details is to provide the model with the image prompts with spatial resolutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": "To further refine the synthesized details, we propose to inject image prompts into the cross-frame attention of the text-to-video models. By injecting image prompts into the cross-frame attention, the image prompts are involved in the updates of the synthesized frames so that the model can directly borrow some visual cues from image prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Since text-to-video diffusion models operate in the latent space, we first feed the image prompt into the VAE of Stable Diffusion and get its latent representation " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_{I}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ". Moreover, since the sampling of the videos starts from the noise map, the latent in the intermediate timesteps contains the noises. If we append the clean latent " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_{I}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " of the image prompt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "to the cross-frame attention, the domain discrepancy exists. Therefore, we follow the diffusion forward process to add corresponding noises to " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "x_{I}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 112, + 545, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 112, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 370, + 112, + 545, + 127 + ], + "type": "interline_equation", + "content": "x _ {t} ^ {I} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} ^ {I} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {4}", + "image_path": "8e27baf2603d3740f3be64c3801ba31c768ee0e0cc1747b340e1bc77c9b2bedd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}" + }, + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "type": "text", + "content": " is a hyperparameter determined by the denoising schedule and " + }, + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(0, I)" + }, + { + "bbox": [ + 304, + 132, + 545, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 156, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 263 + ], + "type": "text", + "content": "The cross-frame attention is used to improve the temporal consistency of the generated frames. For each frame, the key and value are the concatenation of the features of the first frame and the previous frame. Here, we introduce the image prompts as the additional keys and values for the frames. As shown in Fig. 4, we propose to update the values of the first frame firstly using the keys and values of the image prompts and the frame itself. Mathematically, the operation can be expressed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 267, + 545, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 267, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 359, + 267, + 545, + 295 + ], + "type": "interline_equation", + "content": "V _ {0} ^ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {0} K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {5}", + "image_path": "885f0640f8425ca0d1c3292885112f4f01ac07fd6f4d4aca1f24792782af8858.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 297, + 492, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 297, + 492, + 311 + ], + "spans": [ + { + "bbox": [ + 372, + 297, + 492, + 311 + ], + "type": "interline_equation", + "content": "K = \\left[ K _ {I}, K _ {0} \\right], V = \\left[ V _ {I}, V _ {0} \\right],", + "image_path": "f648b7b9e98ed2c7e8db16f2fae458898c5361062868fac2495e4063b41282dd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "K_{I}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "V_{I}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " are the keys and values obtained from the image prompts. The query, key, and value of the first frame are denoted " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "Q_{0}, K_{0}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": ", respectively. It should be noted that we use a separately trained " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{V}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " projection for latent representations " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "x_{t}^{I}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " of image prompts because the image prompts have clean backgrounds, which are different from other frames. The parameters of the newly added " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{V}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " projections are initialized by original " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{V}" + }, + { + "bbox": [ + 304, + 315, + 545, + 411 + ], + "type": "text", + "content": " projections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 411, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 470 + ], + "type": "text", + "content": "Then the updated first frame is used to refine the remaining frames. When updating the remaining frames, the keys used for calculating the attention maps are the original keys, while the values are the updated ones. The update is expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 369, + 474, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 474, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 369, + 474, + 545, + 502 + ], + "type": "interline_equation", + "content": "V _ {i} ^ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {6}", + "image_path": "0af91848fe349d7d349fe0a51282ad88f87efb70e89daa1879ebdac929717f98.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 503, + 502, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 503, + 502, + 517 + ], + "spans": [ + { + "bbox": [ + 350, + 503, + 502, + 517 + ], + "type": "interline_equation", + "content": "K = [ K _ {0}, K _ {i - 1} ], V = [ V _ {0} ^ {\\text {n e w}}, V _ {i - 1} ].", + "image_path": "fbc7c785542e0acfae3d7896855dc96a21b709d078144e0aa4870a9699eea1d1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "content": "To sum up, in the attention injection, we update the values of the first frame using the image prompts first, and then use the updated first frame to update the other frames. In this way, the visual cues from the image prompts can be consistently propagated to all the frames." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": "It should be noted that the diffusion model has multiple cross-frame attention layers with different scales. To inject multi-scale visual cues for better detail refinement, in different cross-frame attention layers, we feed latent representations of the image prompts with corresponding resolutions, which are obtained from different stages of the U-Net." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 659, + 484, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 484, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 484, + 672 + ], + "type": "text", + "content": "3.4. Coarse-to-Fine Training Strategy" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "The visual details of the image prompts are embedded into the final synthesized results in two stages: coarse visual embeddings using an image encoder and fine visual embedding" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6693" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "by attention injection. We propose to train these two modules in a coarse-to-fine manner. In other words, we train the coarse image encoder and tune the parameters in the cross-attention first. After the model has the capability of generating videos containing the subjects specified in image prompts, we then train the attention injection module to embed image prompts into cross-frame attention layers. As we will show in the ablation study (Sec. 5.5), if these two modules are trained together, the fine attention injection module leaks the strong visual cues and the coarse encoder learns meaningless representations. As a result, in sampling phase, the image encoder for the coarse visual embedding cannot provide the coarse information and then the fine attention module cannot refine the details. Therefore, it is necessary to train VideoBooth in a coarse-to-fine manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 165, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 165, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 165, + 289 + ], + "type": "text", + "content": "4. VideoBooth Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 301, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 649 + ], + "type": "text", + "content": "We establish the VideoBooth dataset to support the task of video generation using image prompts. We start from the WebVid dataset [5], a well-known open-source dataset for text-to-video generation. In the WebVid dataset, there is a text prompt with each video. In this paper, we study the task of generating a video clip from one text prompt and one image prompt. Hence, in addition to the original text prompt, we need to provide an image prompt for each video. We propose to segment the subjects from the first frame of the video using the Grounded-SAM (Grounded Segment Anything) [49, 57], and the segmented subjects are image prompts. The Grounded-SAM receives word prompts as inputs and generates segmentation masks for the target subjects specified in word prompts. To obtain the word prompt for the input to Grounded-SAM, we use the spaCy library to parse the noun chunks from the original text prompts, which are used as the word prompts. After the segmentation, we perform data filtering to ensure the data quality. We filter out small objects and large objects (those are almost the same size as the original video) according to the ratio of the object to the whole video. Also, since we focus on generating video clips containing moving objects, we further filter the videos containing moving objects. The keywords we used for filtering are dog, cat, bear, car, panda, tiger, horse, elephant, and lion. In the current version, we have processed 2.5M subset of the WebVid dataset. After data filtering, we have 48,724 video data pairs for training. We will process the full set of the WebVid dataset and include the filtered data in our VideoBooth dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "To evaluate the performance, we also set up a test benchmark. The test benchmark consists of 650 test pairs. For each pair, an image prompt and a text prompt are provided. The test pairs are selected from the rest of the WebVid-10M dataset, which does not overlap with the training set." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 97, + 548, + 149 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons. VideoBooth achieves the best image alignment and comparable text alignment performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 97, + 548, + 149 + ], + "lines": [ + { + "bbox": [ + 307, + 97, + 548, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 548, + 149 + ], + "type": "table", + "html": "
MethodCLIP-Text ↑CLIP-Image ↑DINO ↑
Textual Inversion [21]29.974969.799545.3143
DreamBooth [71]30.687771.207852.9661
ELITE [83]30.088173.751858.9522
VideoBooth (Ours)30.096774.797165.0979
", + "image_path": "e3ee0f8053c03bbf45e0f2f14e1a73f201cf623630eee432779676cab10b3425.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 169, + 387, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 169, + 387, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 387, + 182 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 189, + 430, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 430, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 430, + 202 + ], + "type": "text", + "content": "5.1. Comparison Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 207, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 277 + ], + "type": "text", + "content": "Textual Inversion [21] is a method for customized text-to-image generation. The appearance of the target subjects is embedded into the text embeddings. Concretely, a text token " + }, + { + "bbox": [ + 304, + 207, + 545, + 277 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 304, + 207, + 545, + 277 + ], + "type": "text", + "content": " is optimized to represent the subject. We adapt it to the task of text-to-video generation by replacing the image model with the video model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 279, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 326 + ], + "type": "text", + "content": "DreamBooth [71] is also proposed for customized text-to-image generation. It injects the target subject into the text tokens as well as model weights. During the training, both the model weights and the special token " + }, + { + "bbox": [ + 304, + 279, + 545, + 326 + ], + "type": "inline_equation", + "content": "S*" + }, + { + "bbox": [ + 304, + 279, + 545, + 326 + ], + "type": "text", + "content": " are optimized." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 410 + ], + "type": "text", + "content": "ELITE [83] is an encoder-based method for personalized generation. An encoder is trained to embed the images into the text embeddings. Local mapping and global mapping are employed to transform the CLIP embedding of image prompts into the features, which are injected into the cross-attention module. We adapt and retrain the method using the same pretrained video model we use." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 418, + 418, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 418, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 418, + 430 + ], + "type": "text", + "content": "5.2. Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": "We use three metrics to evaluate the performance [83]. To measure the alignment of the generated videos and given text prompts, we use the CLIP-Text metric. The metric is calculated using the cosine similarity of the CLIP text embeddings of text prompts and CLIP image embeddings of the generated frames. For each video, the value is obtained by averaging values of the all frames. As for the evaluation of the similarity between the given image prompts and the generated videos, we adopt two metrics: CLIP-Image and DINO [7, 63]. The CLIP-Image metric is calculated by the cosine similarity between the CLIP image embedding of image prompts and generated frames. Since the CLIP model is trained to align image embeddings and text embeddings, we follow the practice in previous methods [13, 71] and use the DINO similarity as another indicator. DINO is trained to differentiate the differences between objects of the same classes. We use the ViT-S/16 model to extract the features of the image prompts and generated frames. The final score is obtained by averaging over all frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 671, + 453, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 453, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 453, + 685 + ], + "type": "text", + "content": "5.3. Quantitative Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "We report quantitative results in Table 1. As shown in Table 1, our proposed VideoBooth achieves state-of-the-art" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6694" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 216, + 310 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 216, + 310 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 216, + 310 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 216, + 310 + ], + "type": "image", + "image_path": "101aaef392f59fa34a50a1f1252c07b23ed9633b77955d6aaffd503e198d7499.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 312, + 538, + 325 + ], + "lines": [ + { + "bbox": [ + 53, + 312, + 538, + 325 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 538, + 325 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison. VideoBooth effectively preserves the fidelity of image prompts and achieves better visual quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 76, + 375, + 310 + ], + "blocks": [ + { + "bbox": [ + 221, + 76, + 375, + 310 + ], + "lines": [ + { + "bbox": [ + 221, + 76, + 375, + 310 + ], + "spans": [ + { + "bbox": [ + 221, + 76, + 375, + 310 + ], + "type": "image", + "image_path": "bfb987e8911e53d525e68e5645bb2d6a1ecd4a2ee49a6a70241168f4e8008560.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 76, + 533, + 310 + ], + "blocks": [ + { + "bbox": [ + 376, + 76, + 533, + 310 + ], + "lines": [ + { + "bbox": [ + 376, + 76, + 533, + 310 + ], + "spans": [ + { + "bbox": [ + 376, + 76, + 533, + 310 + ], + "type": "image", + "image_path": "47f0a6bf7d23dd377306176abf7f243d2c6d4c70f3446d83ef095402c70cb6bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 333, + 290, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 290, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 290, + 561 + ], + "type": "text", + "content": "image alignment performance compared to baseline methods. As for the alignment with the text prompts, our proposed VideoBooth has comparable performance with the baseline models. It should be noted that the CLIP-Text score of DreamBooth is significantly higher than the other methods. The reason lies in that the optimized token " + }, + { + "bbox": [ + 46, + 333, + 290, + 561 + ], + "type": "inline_equation", + "content": "S*" + }, + { + "bbox": [ + 46, + 333, + 290, + 561 + ], + "type": "text", + "content": " in DreamBooth is inserted into the text embeddings, rather than replacing the original word embeddings like other methods. This would result in the generated videos of DreamBooth being highly related to the text prompts but having no correlation to the image prompts in some cases. We also conducted a user study, in which 25 users participated. Each user is presented with twelve groups of videos, and each group contains four videos generated by four methods. For each group, users are asked to make three choices: 1) which one has the best image alignment? 2) which one has the best text alignment? 3) which one has the best overall quality. Figure 6 summarizes the results. Our results are preferred by most users in all three dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 573, + 187, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 187, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 187, + 586 + ], + "type": "text", + "content": "5.4. Qualitative Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "We show three visual comparisons on the generated video frames of our proposed VideoBooth and baseline methods in Fig. 5. In the first example, the model is supposed to generate videos containing the dog specified in the image laying on the ground. Textual Inversion cannot correctly embed the appearance of the dog and results in generating another totally different dog. DreamBooth and ELITE can embed the coarse appearance of the dog but the generated details vary from the image prompt. Our proposed VideoBooth successfully embeds the details of the image prompts into" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 314, + 333, + 537, + 427 + ], + "blocks": [ + { + "bbox": [ + 314, + 333, + 537, + 427 + ], + "lines": [ + { + "bbox": [ + 314, + 333, + 537, + 427 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 537, + 427 + ], + "type": "image", + "image_path": "f0ce31bd0b3b4d54c441ad43f4c9d246a8585a9faa92bc7f203f1d5f5c47c58d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 429, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 451 + ], + "type": "text", + "content": "Figure 6. User Study. Our proposed VideoBooth achieves the highest user preference ratios on all three dimensions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "content": "the synthesized videos. In the second example, the condition is to generate a cat playing with the yellow ball. All the methods can generate a yellow ball and a cat but only our proposed VideoBooth can accurately generate the cat having the appearance from the image prompt. As for the last example, Textual Inversion fails to generate a car. DreamBooth generates a distorted car in the bush. ELITE model can generate a car in the bush, but the color differs from the image prompt. By contrast, our model can generate videos having the same car in the bush." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 586, + 400, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 400, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 400, + 600 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": "To evaluate the effectiveness of the proposed components, we perform three ablation studies. Due to the computational resources, we train these models on the subset of our training set. The quantitative metrics are shown in Table 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Only Coarse Embeddings. This ablation model injects the image prompts with only coarse embeddings via Image Encoder. In the example shown in Fig. 7(a), the ablation model only encodes the coarse appearance of the image prompts. The pattern in the legs of the synthesized dog is different" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6695" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 94, + 121 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 94, + 121 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 94, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 94, + 121 + ], + "type": "image", + "image_path": "6943cb25f233daa454e7ce53cb62d883dd68912839fffc7f26cee3381f5d9595.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 91, + 219, + 101 + ], + "lines": [ + { + "bbox": [ + 99, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 99, + 91, + 219, + 101 + ], + "type": "text", + "content": "Dog as a chef at the table" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 121, + 284, + 167 + ], + "blocks": [ + { + "bbox": [ + 51, + 121, + 284, + 167 + ], + "lines": [ + { + "bbox": [ + 51, + 121, + 284, + 167 + ], + "spans": [ + { + "bbox": [ + 51, + 121, + 284, + 167 + ], + "type": "image", + "image_path": "81c88a84ce53ecd6a5b863733caea3d8c80eccd1d46b9a148babd67f516327a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 169, + 250, + 178 + ], + "lines": [ + { + "bbox": [ + 85, + 169, + 250, + 178 + ], + "spans": [ + { + "bbox": [ + 85, + 169, + 250, + 178 + ], + "type": "text", + "content": "(a) Only Coarse Embeddings with Image Encoder" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 179, + 284, + 228 + ], + "blocks": [ + { + "bbox": [ + 51, + 179, + 284, + 228 + ], + "lines": [ + { + "bbox": [ + 51, + 179, + 284, + 228 + ], + "spans": [ + { + "bbox": [ + 51, + 179, + 284, + 228 + ], + "type": "image", + "image_path": "ff789fb811d4423174fefae79ee89ca5b0148ed07b40b0cc25eba116cd8721dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 229, + 252, + 238 + ], + "lines": [ + { + "bbox": [ + 83, + 229, + 252, + 238 + ], + "spans": [ + { + "bbox": [ + 83, + 229, + 252, + 238 + ], + "type": "text", + "content": "(b) Only Fine Embeddings with Attention Injection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 239, + 284, + 285 + ], + "blocks": [ + { + "bbox": [ + 51, + 239, + 284, + 285 + ], + "lines": [ + { + "bbox": [ + 51, + 239, + 284, + 285 + ], + "spans": [ + { + "bbox": [ + 51, + 239, + 284, + 285 + ], + "type": "image", + "image_path": "0d0fcc35c6120ec2947a6b9e368dd970ea376de83ca1ccd765270919345403a0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 287, + 271, + 296 + ], + "lines": [ + { + "bbox": [ + 63, + 287, + 271, + 296 + ], + "spans": [ + { + "bbox": [ + 63, + 287, + 271, + 296 + ], + "type": "text", + "content": "(c) Unified Training for Image Encoder and Attention Injection" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 51, + 297, + 284, + 345 + ], + "blocks": [ + { + "bbox": [ + 51, + 297, + 284, + 345 + ], + "lines": [ + { + "bbox": [ + 51, + 297, + 284, + 345 + ], + "spans": [ + { + "bbox": [ + 51, + 297, + 284, + 345 + ], + "type": "image", + "image_path": "ce2655202b7f6a62180e6d2e91d944faeceeba5bceb2219dcc7842e57c3dd5be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 347, + 192, + 356 + ], + "lines": [ + { + "bbox": [ + 143, + 347, + 192, + 356 + ], + "spans": [ + { + "bbox": [ + 143, + 347, + 192, + 356 + ], + "type": "text", + "content": "(d) Full Model" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 357, + 286, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 286, + 445 + ], + "type": "text", + "content": "Figure 7. Ablation Study. (a) With only coarse embeddings from image encoder, generated patterns in the body of dog are different from image prompt. (b) With only fine embeddings from attention injection, there lacks coarse encodings of the dog for attention injection module to refine and thus the generated dog is distorted at later synthesized frames. (c) The unified training degrades the capability of image encoder and thus the dog is distorted. (d) The full model better keeps the all visual details of image prompt." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 456, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 493 + ], + "type": "text", + "content": "from that in the image prompt. By contrast, the results of our full model show that our proposed model can transfer all the details in the image prompts to the synthesized videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 495, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 674 + ], + "type": "text", + "content": "Only Fine Embeddings. In this ablation model, we only have fine embeddings of the image prompt in cross-frame attention layers. The main purpose of using fine embedding is to refine the coarse encoding of image prompts from the Image Encoder. Without coarse embeddings, the model with fine embeddings only cannot refine the details. As shown in Fig. 7(b), the first frame contains the exact appearance as the image prompt, but the temporal consistency cannot be guaranteed. The generated dog is distorted in the following frames. The reason is that the model trained in this way overfits the image prompt. In the first frame, the model can copy the information from image prompts. In the following frames, without the coarse embeddings, the generation of the appearance only relies on the propagation of the appearance from the first frame." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "The Necessity of Coarse-to-Fine Training. In Video-Booth, we propose the coarse-to-fine training strategy, i.e., train the coarse embeddings first and then train the attention" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 320, + 87, + 526, + 138 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 535, + 81 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 535, + 81 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 535, + 81 + ], + "type": "text", + "content": "Table 2. Ablation Study. The full model has the best scores." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 87, + 526, + 138 + ], + "lines": [ + { + "bbox": [ + 320, + 87, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 320, + 87, + 526, + 138 + ], + "type": "table", + "html": "
VariantsCLIP-Image ↑DINO ↑
(a) Coarse Embeddings only75.436664.9568
(b) Fine Embeddings only75.555366.0378
(c) Unified Training75.825467.4201
(d) Full Model76.163169.7374
", + "image_path": "86c1da2bb76c39a390117e05416d9b1b083ba8c039750b6f920610dcf415029f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": "injection module. In this ablation model, we train these two modules within one stage. The unified training makes the model rely heavily on the strong guidance provided in attention injection. In this way, the trained image encoder has limited capability. Thus, the model trained in this way has a similar behavior as the model with only fine embeddings. The model also overfits the image prompts from the attention injection. As shown in the first example in Fig. 7(c), the appearance in the first frame is correct, but the generated dog in the following frames is distorted. Due to the image encoder having limited capability to provide the correct coarse encoding of image prompts, the attention injection cannot refine the details. By contrast, the full model can generate consistent frames with all details well preserved." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 323, + 375, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 375, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 375, + 335 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 342, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 474 + ], + "type": "text", + "content": "In this paper, we propose a novel framework VideoBooth to generate videos using image prompts and text prompts. The image prompts specify the appearance of the subjects. We inject the image prompts into the model in two modules: Coarse Embeddings via Image Encoder and Fine Embeddings via Attention Injection. The Image Encoder provides the coarse embeddings of image prompts for the refinement of the Attention Injection module. These two modules cooperate with each other and they are trained in a coarse-to-fine manner. Our proposed VideoBooth generates consistent videos containing the desired subjects." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": "Limitations and Future Work. The model is trained with the image prompts that do not have very divergent views from target videos. Thus, if an image prompt with a back view is provided, we cannot generate videos from the front view. In future work, we can augment image prompts using image-to-3D models [56, 78] so that image prompts used for training will have different views from target videos." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "content": "Potential Negative Societal Impacts. The model may be applied to generate fake videos, which can be potentially avoided by using deepfake detection methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and the Google PhD Fellowship. This study is partially supported by the National Key R&D Program of China (2022ZD0160201), and Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6696" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06925, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "text", + "content": "[3] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "text", + "content": "[4] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 384 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 385, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 288, + 441 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 441, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 288, + 474 + ], + "type": "text", + "content": "[8] Duygu Ceylan, Chun-Hao Huang, and Niloy J. MiTRA. Pix2video: Video editing using image diffusion. arXiv:2303.12688, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 510 + ], + "type": "text", + "content": "[9] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[10] Hong Chen, Xin Wang, Guanning Zeng, Yipeng Zhang, Yuwei Zhou, Feilin Han, and Wenwu Zhu. Videodreamer: Customized multi-subject text-to-video generation with disen-mix finetuning. arXiv preprint arXiv:2311.00990, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[11] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Disentangled parameter-efficient tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[12] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photoverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[13] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[15] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "type": "text", + "content": "[16] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "text", + "content": "[17] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 260, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 304 + ], + "type": "text", + "content": "[18] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "text", + "content": "[19] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Ccredit: Creative and controllable video editing via diffusion models. arXiv preprint arXiv:2309.16496, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "text", + "content": "[20] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "text", + "content": "[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[22] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 481, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 525 + ], + "type": "text", + "content": "[23] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 559 + ], + "type": "text", + "content": "[24] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[25] Yuan Gong, Youxin Pang, Xiaodong Cun, Menghan Xia, Haoxin Chen, Longyue Wang, Yong Zhang, Xintao Wang, Ying Shan, and Yujiu Yang. Talecrafter: Interactive story visualization with multiple characters. arXiv preprint arXiv:2305.18247, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 669 + ], + "type": "text", + "content": "[26] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[27] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6697" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[29] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[30] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "type": "text", + "content": "[31] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "text", + "content": "[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "type": "text", + "content": "[33] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 435 + ], + "type": "text", + "content": "[35] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "text", + "content": "[36] Zhihao Hu and Dong Xu. Videocontrolnet: A motion-guided video-to-video translation framework by using diffusion model with controlnet. arXiv preprint arXiv:2307.14073, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "type": "text", + "content": "[37] Ziqi Huang, Kelvin CK Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6080-6090, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "type": "text", + "content": "[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. arXiv preprint arXiv:2311.17982, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": "[39] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 658 + ], + "type": "text", + "content": "[40] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "text", + "content": "[41] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zeroshot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[42] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "text", + "content": "[43] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13799-13808, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[44] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[45] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 242, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 296 + ], + "type": "text", + "content": "[46] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "text", + "content": "[47] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Learning object level concepts using multi-concept prompt learning. arXiv preprint arXiv:2310.12274, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 342, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 398 + ], + "type": "text", + "content": "[48] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[49] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 454, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 520 + ], + "type": "text", + "content": "[50] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17089-17099, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 521, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 555 + ], + "type": "text", + "content": "[51] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "text", + "content": "[52] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 600, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 633 + ], + "type": "text", + "content": "[53] Yuheng Li, Haotian Liu, Yangming Wen, and Yong Jae Lee. Generate anything anywhere in any scene. arXiv preprint arXiv:2306.17154, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "text", + "content": "[54] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. In arXiv, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 669, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 690 + ], + "type": "text", + "content": "[55] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[56] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3:" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6698" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "type": "text", + "content": "[57] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "text", + "content": "[58] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[59] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. arXiv preprint arXiv:2305.19327, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 304 + ], + "type": "text", + "content": "[60] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "text", + "content": "[61] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[62] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "text", + "content": "[63] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[64] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "text", + "content": "[65] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "text", + "content": "[67] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[68] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 161 + ], + "type": "text", + "content": "[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "text", + "content": "[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "text", + "content": "[72] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 327 + ], + "type": "text", + "content": "[73] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "text", + "content": "[74] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. arXiv preprint arXiv:2303.07945, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "text", + "content": "[75] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 458 + ], + "type": "text", + "content": "[76] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "text", + "content": "[77] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 557 + ], + "type": "text", + "content": "[78] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[79] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 603, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 667 + ], + "type": "text", + "content": "[80] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[81] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6699" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[82] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[83] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[84] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 271 + ], + "type": "text", + "content": "[85] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. arXiv preprint arXiv:2312.07537, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "text", + "content": "[86] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. Singleinsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[87] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "text", + "content": "[88] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking \"text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[89] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video editing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "text", + "content": "[90] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "type": "text", + "content": "[91] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 288, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 570 + ], + "type": "text", + "content": "[92] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 288, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 288, + 614 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 288, + 614 + ], + "type": "text", + "content": "[93] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 615, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 288, + 658 + ], + "type": "text", + "content": "[94] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[95] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 206 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[96] Yuyang Zhao, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Make-a-protagonist: Generic video editing with an ensemble of experts. arXiv preprint arXiv:2305.08850, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[97] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 206 + ], + "type": "text", + "content": "[98] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6700" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_content_list.json b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2554538e3fd80e5186c931fa452a2acc5510a180 --- /dev/null +++ b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_content_list.json @@ -0,0 +1,1632 @@ +[ + { + "type": "text", + "text": "VideoCon: Robust Video-Language Alignment via Contrast Captions", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hritik Bansal1 Yonatan Bitton2 Idan Szpektor2* Kai-Wei Chang1* Aditya Grover1* \n1UCLA 2Google Research \n{hbansal, kwchang, adityag}@cs.ucla.edu \n{yonatanbitton, szpektor}@google.com", + "bbox": [ + 130, + 179, + 844, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite being (pre)trained on a massive amount of data, state-of-the-art video-language alignment models are not robust to semantically-plausible contrastive changes in the video captions. Our work addresses this by identifying a broad spectrum of contrast misalignments, such as replacing entities, actions, and flipping event order, which alignment models should be robust against. To this end, we introduce the VideoCon, a video-language alignment dataset constructed by a large language model that generates plausible contrast video captions and explanations for differences between original and contrast video captions. Then, a generative video-language model is finetuned with VideoCon to assess video-language entailment and generate explanations. Our VideoCon-based alignment model significantly outperforms current models. It exhibits a 12-point increase in AUC for the video-language alignment task on human-generated contrast captions. Finally, our model sets new state of the art zero-shot performance in temporally-extensive video-language tasks such as text-to-video retrieval (SSv2-Temporal) and video question answering (ATP-Hard). Moreover, our model shows superior performance on novel videos and human-crafted captions and explanations.", + "bbox": [ + 75, + 318, + 472, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantically aligning data points from diverse modalities is a long-standing goal of AI. We focus on video-language alignment, which is challenging due to the complexities involved in understanding of entities, their relationships, and temporal order of the depicted events [17]. Recent models such as VideoCLIP [55], ImageBind [14] learn a shared embedding space. Similarly, generative models such as Flamingo [1], mPLUG-Owl-Video [61] can provide a classification label (e.g., yes/no) when queried about video-language alignment.", + "bbox": [ + 75, + 722, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite large-scale pretraining, prior work [5, 37, 38, 51] highlights that video-language alignment models are not robust to semantically plausible manipulations to an original aligned caption in the form of contrast captions, such as from 'dog runs away before it eats food' to 'dog runs away after it eats food'. Such pitfalls in robustness questions the trustworthiness of alignment models for large-scale deployment. To mitigate these shortcomings, one possible solution is to scale video-language pairs more for increased diversity during pretraining. However, this is challenging due to the difficulties in sourcing new, high-quality and permissible content, as well as the requirement for substantial storage capacity. Several works [11, 13, 16] have shown that naively training models on web-scale data has diminishing returns on downstream tasks, and emphasize the importance of data quality. Furthermore, the recent studies [28, 62] demonstrate that applying a contrastive objective to the pretraining datasets does not encourage the model to grasp the fine-grained details within image/region-caption data.", + "bbox": [ + 496, + 287, + 893, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we take a scalable, active strategy to gather high-quality data that is deliberately enriched with the attributes that we want to instill in alignment models. We create a novel dataset, VideoCon, to improve the robustness of models. Specifically, the dataset consists of a variety of semantically plausible video-language misalignments in contrast captions. These misalignments include altering objects (entities), actions, attributes, relations, counts, event orders, and introducing hallucinations (Figure 2). To construct VideoCon, a large language model (PaLM-2 API) takes video-caption pairs as input and generates high-quality contrast captions for a given misalignment type. To make our dataset temporally-challenging, we skipped \"easy\" video-caption pairs whose alignment could be inferred based on a single frame (image) understanding [9, 26] ( $\\S 3.1$ ). In addition, the LLM generates natural language explanations (NLEs) [42] to the differences between original and altered captions, which are used for further robust training. We performed human verification on a sample of VideoCon and found that it is of high-quality. Finally, to evaluate the model's generalization capabilities, we col", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Advising.", + "bbox": [ + 94, + 886, + 191, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13927", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e8fee37352d8ab1f77e746d84d70b9cfa47f99ed1985abda0ba57454e9c5947.jpg", + "image_caption": [ + "Figure 1. Overview of our VideoCon approach. First, aligned video-language pairs are filtered to retain temporally-challenging instances. Then contrast captions and natural language explanations (NLE) are generated by an LLM to create the VideoCon dataset. Second, a video-language alignment model is finetuned with VideoCon on the alignment and NLE tasks. Finally, the finetuned model is evaluated against the baseline model. Our results show that it outperforms the baseline, achieving state-of-the-art results on downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 883, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lect human-generated contrast captions and NLEs for the videos sourced from external datasets that did not contribute to VideoCon's development.", + "bbox": [ + 75, + 308, + 468, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We finetuned a generative video-language model (mPLUG-Owl-Video) on the VideoCon dataset. The trained model surpasses existing video-language alignment models by a large margin on the LLM-generated test set for both video-language alignment and NLE generation tasks. Interestingly, we observed that our finetuned model generalizes to unseen videos and human-generated contrast captions and NLEs, and outperforms the baseline models. For instance, our model's ROC-AUC exceeds the baseline model by 12 points on the human-generated contrast captions. This indicates that our model has developed a better understanding of the entities, their interactions, action understanding, as well as the temporal order of the events for robust video-language alignment.", + "bbox": [ + 75, + 354, + 468, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further assessed the effectiveness of robust training via contrast captions on zero-shot downstream videolanguage tasks such text-to-video retrieval and video question answering on the temporally-challenging and action-intensive SSv2-Temporal [45] and SSv2-Events [5]. Our model achieves state-of-the-art (SOTA) performance, improving on SSv2-Temporal by $4.3\\mathrm{mAP}$ , SSv2-Events by $3.6\\mathrm{mAP}$ points. In addition, our model also achieves SOTA on temporal and causal video question answering in the ATP-Hard dataset, increasing $4\\%$ accuracy. This suggests that equipping a model with the knowledge of contrast captions is highly data-efficient and effective in improving its robustness in comparison to scaling the pretraining data. The complete pipeline is illustrated in Figure 1. The dataset and the model will be released upon acceptance.", + "bbox": [ + 75, + 566, + 470, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Video Language Alignment", + "text_level": 1, + "bbox": [ + 76, + 806, + 331, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are interested in assessing the semantic alignment between the video1 and text data since it powers many prac", + "bbox": [ + 76, + 833, + 468, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tical applications such as video-text retrieval [57], video generation [7, 47] and video captioning [59]. To this end, [14, 39, 49, 55] designed (image)video-text alignment models that are utilized for evaluating the semantic similarity between the two modalities. However, previous works [5, 37, 38, 51] have questioned their robustness to semantically plausible changes to the video descriptions, termed here contrast captions. Our aim is to improve the robustness of video-text alignment models by training on contrast captions with a wide range of misalignments.", + "bbox": [ + 496, + 308, + 890, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider a dataset $\\mathcal{D} = \\{(V_i, T_i, C_i, E_i)\\}$ where $V_i$ is a video, $T_i$ is an aligned caption, $C_i$ is a contrast caption which is a perturbation of $T_i$ but misaligns with $V_i$ , and $E_i$ is a natural language explanation for the misalignment between $V_i$ and $C_i$ . We consider two video-language alignment tasks: (a) video-language entailment, (b) natural language explanation.", + "bbox": [ + 496, + 458, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Language Entailment (VLE) casts video-text alignment as a Visual Entailment (VE) task. VE was originally defined for images as premises and texts as hypothesis [53, 54]. We extend VE definition also for videos as premises, under which a classification model $A_{vle}(V,T)$ predicts whether a video $V$ entails a text $T$ .", + "bbox": [ + 496, + 584, + 890, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural Language Explanation (NLE) requires a model, $A_{nle}(V,C)$ , to generate an open-ended explanation for the discrepancy between a video $V$ and a non-entailing caption $C$ .", + "bbox": [ + 496, + 691, + 890, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address both VLE and NLE tasks under a multitask setting in which a single video-language generative model generates the binary label for entailment and the open-ended explanation.", + "bbox": [ + 496, + 752, + 890, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. VideoCon: Contrast Captions Generation for Robust Video-Language Alignment", + "text_level": 1, + "bbox": [ + 498, + 825, + 890, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our research goal is to measure the impact of a comprehensive dataset on increasing the robustness of video-text align-", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Like prior works [33, 55], we use only the video frames (the visual channel) without the soundtrack (the audio channel).", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13928", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f58ee2103ef495f101d966eab6030107743d2821b15cb27558c2319b1be4ad19.jpg", + "image_caption": [ + "Figure 2. Overview of the VideoCon data generation process from top to bottom. Specifically, we prompt a large language model (PaLM-2) with the original caption that is grounded in the video, and the intended type of misalignment within the contrast caption. We consider seven kinds of misalignments including object, action, attribute, counting, spatial relation, hallucination, and event order flip. We provide a generated contrast caption and the corresponding natural language explanation for each misalignment type." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 797, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment models. To this end, we first collect video-caption pairs where the caption cannot be derived from a single frame of video. We then categorize a wide range of semantically plausible manipulations of video captions. Using an LLM for large-scale computation, contrast captions and related explanations are generated for the defined categories, constructing the VideoCon dataset. Finally, we extend VideoCon to include human-created contrast captions as held-out evaluation on unseen videos. We detail the dataset construction steps below.", + "bbox": [ + 75, + 412, + 472, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporally-Challenging Instance Selection", + "text_level": 1, + "bbox": [ + 76, + 582, + 442, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To construct VideoCon, we start with existing datasets that include natural (real) videos and associated high-quality human-written captions: MSR-VTT [57], VaTeX [48], and TEMPO [17]. MSR-VTT and VaTeX consist of 20 captions and 10 captions per video, respectively, while TEMPO consists of a single caption per video. More dataset details are in Appendix §B.", + "bbox": [ + 75, + 607, + 468, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TEMPO is designed to create temporally-challenging instances, while MSR-VTT and VaTeX contain more general video-caption pairs. For MSR-VTT and VaTeX, we filter out instances, where the caption is highly associated with a single frame in the video based on an image-text alignment model. In such cases, a video-text alignment can leverage shortcuts and align the video to its caption without understanding the temporal or causal relations depicted in the video. We want to filter such instances.", + "bbox": [ + 75, + 715, + 470, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we employ the End-to-End VNLI model [60] to calculate an alignment score $A_{vle}(V,T)$ between a video $V = \\{I_1,I_2,\\dots ,I_N\\}$ and a text $T$ where $I_{i}$ is a frame from", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the video sampled at a rate of 1 frame per second. Formally,", + "bbox": [ + 498, + 412, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {v l e} (V, T) = \\max _ {i} \\left(V N L I \\left(I _ {i}, T\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 441, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $VNLI(I_i, T)$ is the task of visual natural language inference that assesses the extent to which the text $T$ entails the image $I_i$ . There are 20 and 10 captions per video in the MSR-VTT and VaTeX datasets, respectively. We retain 5 captions per video from these datasets with the lowest $A_{vle}(V, T)$ , and the remaining captions are filtered out. Post-filtering, the percentage of temporally-challenging instances increased from $36.5\\%$ to $81.5\\%$ in MSR-VTT, and from $42.6\\%$ to $71\\%$ in VaTeX.", + "bbox": [ + 496, + 469, + 892, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Categories of Contrast Captions", + "text_level": 1, + "bbox": [ + 498, + 616, + 784, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim for VideoCon to include a wide range of misalignments in its contrast captions. Overall, VideoCon covers seven misalignment types, exemplified in Figure 2. We include replacement of objects (entities) and actions following the analysis in [37, 38], and replacement of attributes, counts, relations, as well as adding unrelated but plausible information to captions as hallucinations following [29, 32, 35]'s study of image/text alignment model brittleness. Since most video-text models rely on pretrained image backbones, they are likely to suffer from similar problems. Finally, following [5]'s analysis that video-text models do not understand temporal order of the events, we include event order flipping as misalignment type.", + "bbox": [ + 496, + 640, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Data Generation using an LLM", + "text_level": 1, + "bbox": [ + 500, + 845, + 779, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate contrast captions and corresponding NLE we first assign one of the seven misalignment types (§3.2) to", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13929", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/645e16ec5010339726cb400434e3a471e9dc3a9adb1256106315c49df13933d4.jpg", + "image_caption": [ + "Figure 3. Distribution of the types of misalignments within the contrast captions of the VideoCon dataset. We observe that the dataset has good representation for all the kinds of misalignments ranging from $8.8\\%$ to $24.2\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 148, + 103, + 397, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each caption in the input video-text datasets (§3.1) (details in Appendix §C). Then, given a video $V$ and a misalignment type $m$ , we prompt PaLM-2 API² [2] to generate a contrast caption and accompanied explanation (our type-specific prompts are detailed in Appendix §D).", + "bbox": [ + 75, + 337, + 467, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analyzing the LLM generations, we found that sometimes the output caption $C$ do not contradict the original caption $T$ . For example, a generated contrast caption \"a person riding a car\" does not contradict the original caption \"a person riding a mustang\". To filter such cases, we employ a Natural Language Inference (NLI) model [19] and remove cases in which the contrast caption is assessed as entailed by the original caption $NLI(T,C) > 0.5$ . Postfiltering, each tuple $(V,T,C,m)$ is converted to the two instances of video/language entailment task: $A_{en}(V,T) = 1$ and $A_{en}(V,C) = 0$ . We present the dataset statistics for the entailment task in Table 1, including train/eval/test splits. In addition, Fig. 3 shows the distribution of misalignment types in the dataset. We observe that VideoCon maintains a high density across the 7 misalignments ranging from $8.8\\%$ to $24.2\\%$ .", + "bbox": [ + 75, + 412, + 467, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also found that some generated explanations do not describe the differences between $T$ and $C$ well. For example, the explanation \"two friends are not traveling together\" does not fully describe the discrepancy between \"three friends traveling together\" and \"two friends are traveling together\". To filter these out, generated examples are removed if $NLI(F(T,C),E) < 0.6$ where $F(T,C)$ is the premise comprising the original and contrast captions. Specifically, premise will be 'Expected Caption: $T$ Actual Caption: $E'$ and hypothesis will be 'Difference between Expected and Actual Caption: $E'$ . This filter indicates that the information in the explanation is not entailed by the difference between the two captions. The dataset statistics for the NLE task is presented in Table 1. We refer to the final", + "bbox": [ + 75, + 655, + 467, + 864 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/16ac4b946daaafb1768824f2b10643f39c6b1f7b30ed130c5ba5b4841b6dc2cb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Video-Language EntailmentNatural Language Explanation
SourceTrainValTestTrainValTest
MSR-VTT3836647816538158882066788
VaTeX664807368110301803453636
TEMPO1071270982708416527391073
Total11555883122735650233329011497
", + "bbox": [ + 500, + 88, + 890, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Statistics for the VLE and NLE tasks in VideoCon.", + "bbox": [ + 517, + 167, + 870, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM-generated dataset as VideoCon (LLM).", + "bbox": [ + 500, + 210, + 799, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess the quality of VideoCon (LLM), we perform human evaluation on 500 contrast captions and NLEs (details in Appendix E). The human evaluator found $91\\%$ of the contrast captions and $89\\%$ of the NLEs to be valid, indicating the high-quality of VideoCon (LLM).", + "bbox": [ + 496, + 228, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Data Generation using Humans", + "text_level": 1, + "bbox": [ + 500, + 316, + 777, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To study whether a model trained on VideoCon (LLM) generalizes to out-of-distribution videos and its performance on human-generated contrast captions, we randomly selected a set of videos from the validation set of ActivityNet [10]. This dataset consists of captions matched with segments in the video, e.g., \"a little boy is climbing on an outside gym\" matched to the first 10 seconds of its related video. We extracted video segments with an associated caption. Human workers3 on Amazon MTurk were then shown the video segments and associated captions and were asked to create a semantically plausible contrast caption and a corresponding NLE (more details in Appendix §F). We did not communicate any type of target misalignments to encourage natural diversity of human created contrast captions.", + "bbox": [ + 496, + 340, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, we collected 570 tuples $(V, T, C_{human}, E_{human})$ where $V$ is the video, $T$ is the original caption, $C_{human}$ is the human-written contrast caption, and $E_{human}$ is the human-written explanations. We denote this dataset by VideoCon (Human). We sample 100 instances from this dataset, and found $93\\%$ to be clean. In addition, we observe that many of the human-generated contrast captions perturbing one or more objects ( $35\\%$ ) and actions ( $35\\%$ ) depicted in the caption. While $8\\% - 10\\%$ of the contrast captions flip the order of the events and attribute of the objects. As this dataset is largely unfiltered, it contains a mix of temporally-easy and challenging instances. We also constructed a more temporally-challenging subset of 290 instances, denoted VideoCon (Human-Hard), by filtering out tuples in which $A_{vle}(V, T) < 0.5$ (Eq. (1)), as in §3.1.", + "bbox": [ + 496, + 555, + 890, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 814, + 691, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We next describe our evaluation setting for measuring the impact of VideoCon on video-text alignment modeling.", + "bbox": [ + 500, + 840, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https : / / developers . generativeai . google / products/palm", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3A shortlist that passed our qualification test.", + "bbox": [ + 514, + 886, + 756, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13930", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Finetuning with VideoCon", + "text_level": 1, + "bbox": [ + 76, + 90, + 316, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal in constructing VideoCon (LLM) is to improve robustness of video-text alignment models by fine-tuning on this dataset. To this end, we start with the mPLUG-OwlVideo model [61], denoted Owl-Base. Its building blocks are CLIP [39] as visual encoder and LLaMA-7B [46] as text encoder/decoder and it was pretrained on VideoChat [27].", + "bbox": [ + 75, + 114, + 468, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Entailment Task:", + "text_level": 1, + "bbox": [ + 107, + 222, + 205, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given: V (Video), T (Caption), C (Contrast Caption)", + "bbox": [ + 109, + 244, + 385, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction (I): [V] Does this video entail the description $[T]$ ? Response (R): Yes", + "bbox": [ + 109, + 268, + 437, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction (I): [V] Does this video entail the description [C]? Response (R): No", + "bbox": [ + 109, + 305, + 437, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4. Entailment task prompt for finetuning.", + "bbox": [ + 127, + 343, + 415, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Natural Language Explanation Generation Task:", + "text_level": 1, + "bbox": [ + 102, + 392, + 401, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given: V (Video), C (Contrast Caption), E (NLE)", + "bbox": [ + 104, + 417, + 392, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction (I): [V] What is the misalignment between this video and the description $[C]$ ? Response (R): [E]", + "bbox": [ + 104, + 444, + 441, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5. NLE generation task prompt for finetuning.", + "bbox": [ + 112, + 500, + 429, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fine-tune $Owl$ -Base on VideoCon (LLM), its $\\{V, T, C, E\\}^4$ tuples were converted into two types of multimodal instruction-response pairs, one for the VLE task $(I_{vle}, R)$ (Fig. 4) and one for the NLE task $(I_{nle}, R)$ (Fig. 5). We then train $Owl$ -Base on all instruction pairs from both the tasks with maximum likelihood loss, resulting in a single model $Owl$ -Con.", + "bbox": [ + 76, + 531, + 468, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. VideoCon Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 648, + 339, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the performance of the $Owl-Con$ on video-text alignment we generate $Owl-Con$ response to prompt $I_{vle}$ for video $V$ and text $Y \\in \\{T, C\\}$ . We then calculate the probability of generating responses $s_y = Owl-Con$ ( $^\\prime \\mathrm{Yes}^\\prime | I_{vle}(V, Y)$ ) and $s_n = Owl-Con$ ( $^\\prime \\mathrm{No}^\\prime | I_{vle}(V, Y)$ ), and based on these scores the probability for class 'Yes': $P_{yes}(V, Y) = \\frac{s_y}{s_y + s_n}$ . Finally, we compute the ROC-AUC score for $P_{yes}(V, Y)$ over the VideoCon (LLM) eval set, with $\\{V, T\\}$ as label 1 and $\\{V, C\\}$ as label 0.", + "bbox": [ + 76, + 672, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate $Owl$ -Con on the NLE task, we prompt it with instruction $I_{nle}$ instantiated on $\\{V,C\\}$ pairs from the VideoCon (LLM) eval set. We compare the generated explanation $\\hat{E}$ to the ground truth $E$ by measuring entailment", + "bbox": [ + 76, + 813, + 468, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "probability $NLI(E, \\hat{E})$ . In our experiments, we experiment with two $NLI$ automatic metrics: (a) $Q^2$ score [19], and (b) PaLM-2 API. We performed human evaluation to measure the agreement between the automatic metrics and the human-rating. We found that both metrics achieve high agreement with human assessment (Appendix §H).", + "bbox": [ + 500, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Video-Text Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 189, + 769, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We complement the VideoCon intrinsic evaluation over the testset with an extrinsic evaluation over two temporal and action difficult downstream tasks.", + "bbox": [ + 498, + 212, + 890, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate alignment model performance for text2video retrieval over SSv2-Temporal [45] and SSv2-Events [5] datasets. We consider the SSv2-Template captions instead of the label captions since they remove the object-centric bias in model evaluation [26]. We compute input-text/candidate-video alignment score, rank videos and report mean Average Precision (mAP). We evaluate alignment model performance for video question answering over the ATP-Hard [9] dataset. We cast each question/candidate-answer pair as an imperative statement using PaLM-2 API, measure alignment to the input video and report Accuracy. More details on the downstream datasets and the evaluation setup are in Appendix §I.", + "bbox": [ + 498, + 258, + 892, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Baselines", + "text_level": 1, + "bbox": [ + 500, + 462, + 607, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the video-text alignment text, we compare Owl-Con with the following baselines: (a) End-to-End VNLI as zero-shot atemporal model since it does not have access to the temporal order of the video frames, (b) VideoCLIP [55], (c) ImageBind [14], (d) Owl-Base, and (e) Owl-Rand: Owl-Base fine-tuned on VideoCon tuples $\\{V,T,\\hat{C},E\\}$ where $\\hat{C}$ is randomly selected from other captions in the dataset. Owl-Rand would indicate if there is merit in the contrast, hard-negative captions in VideoCon. We include additional baselines TACT [5] and VFC [37] for evaluating on the downstream tasks (\\$5.3).", + "bbox": [ + 498, + 484, + 890, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 662, + 632, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present our intrinsic (VideoCon eval set) and extrinsic (downstream tasks) evaluation results, showing the benefits of VideoCon for robust video-language alignment.", + "bbox": [ + 498, + 688, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Performance on VideoCon Entailment Task", + "text_level": 1, + "bbox": [ + 500, + 741, + 869, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the ROC-AUC scores of the tested models in Table 2. From the table we see that the baseline models find the VideoCon testset difficult, as reflected by low AUC scores (e.g. Owl-Base-57.2), close to random. Even training on VideoCon train instances, but with \"easy\" negatives (Owl-Rand-59.7), hardly improves the base models. A significant improvement is achieved with the VNLI-specific model (67), showing that the entailment task is not inherently represented in generic video-language aligned training", + "bbox": [ + 498, + 763, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{4}\\mathrm{V}$ : video, T: original caption, C: contrast caption, E: explanation.", + "bbox": [ + 94, + 886, + 442, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13931", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a64ff6bbdc2bc8e69de5f40836651737118ddce24c4cf8141bf33fb5d11c879d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsVideoCon (LLM) TestVideoCon (Human)VideoCon (Human-Hard)
Random50.050.050.0
VideoCLIP [55]53.247.347.5
ImageBind (Video-Text) [14]57.165.263.0
Owl-Base [61]57.266.864.1
Owl-Rand59.768.965.5
End-to-End VNLI [60]67.072.465.0
Owl-Con (Ours)84.678.374.4
", + "bbox": [ + 112, + 87, + 857, + 214 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b970c390320b1cf036e5184546f902cfbc8291d7f71a907ea2c5b7c8e8748e71.jpg", + "table_caption": [ + "Table 2. ROC-AUC scores of the tested models for the entailment task on VideoCon test sets." + ], + "table_footnote": [], + "table_body": "
VideoCon (LLM)VideoCon (Human)
ModelsQ2entailmentPaLM-2 entailment acc. (%)Q2entailmentPaLM-2 entailment acc.(%)
Owl-Base0.1936.80.2339.6
Owl-Con (Ours)0.5065.40.3247.1
", + "bbox": [ + 78, + 251, + 893, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sets and requires specific training. Yet, the best performance is achieved by training on VideoCon, which addresses the diversity in plausible misalignments and includes \"difficult\" training examples, reaching 84.6 AUC. This demonstrates the merit of VideoCon for improving video-language alignment robustness. We show qualitative examples for the model predictions in §6.2.", + "bbox": [ + 75, + 369, + 468, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When evaluating on out-of-domain (OOD) data around video types and misalignment distribution, we again see that training with VideoCon offers significant improvement to alignment detection, outperforming all baselines, albeit with smaller relative gains: $17\\%$ and $16\\%$ improvement compared to Owl-Base on (Human) and (Human-Hard) respectively compared to $48\\%$ on (LLM) test. In future work, we plan to further diversify the misalignments VideoCon covers to further improve its benefits on OOD cases.", + "bbox": [ + 75, + 477, + 468, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We notice that the performance of the VNLI atemporal model is better than existing video-language alignment models. It might be attributed to its training with contrast captions in [60]. It further highlights that the existing videolanguage models are not robust in comparison to a temporal probe on video-language alignment evaluation, corroborating the findings from [9, 26].", + "bbox": [ + 75, + 612, + 468, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance on NLE Generation Task", + "text_level": 1, + "bbox": [ + 76, + 726, + 410, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents the performance of the tested models against the ground-truth on the NLE task, depicting average $Q^2$ score and PaLM-2 entailment accuracy. The results show that on in-domain VideoCon, Owl-Con outperforms Owl-Base by an impressive $263\\%$ and $178\\%$ relative increase on $Q^2$ score and PaLM-2 accuracy respectively. This indicates the finetuned model can accurately generate NLE that match well with the ground-truth NLE. This indicates that our model can generate accurate NLE for a wide range of misalignments in the video captions, which makes it use", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ful for dense video-language alignment evaluation.", + "bbox": [ + 498, + 369, + 834, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On out-of-domain VideoCon, the improvement is more moderate but still high: $40\\%$ and $20\\%$ relative increase on $Q^2$ and PaLM-2 respectively. This is probably due to the more diverse ways humans express explanations compared to LLM prompting. In future work we plan to further address linguistic diversity in explanations for more robust generation and evaluation.", + "bbox": [ + 496, + 386, + 890, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Performance on Video-Text Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 503, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4f6617e9e52241b4b493574a3e2d2bfc1c4befbadc737ed63d3587ec131c3856.jpg", + "table_caption": [ + "Table 3. Performance of the tested models on the NLE generation task, measured via entailment metrics." + ], + "table_footnote": [], + "table_body": "
ModelsSSv2-Temporal\nmAPSSv2-Events\nmAP
Random7.33.3
VideoCLIP9.86.4
ImageBind (video-language)10.55.5
Owl-Base10.96.8
TACT [5]-7.8
Owl-Rand12.19.9
End-to-End VNLI [60]14.610.4
Owl-Con (Ours)15.211.4
", + "bbox": [ + 501, + 537, + 890, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Mean Average Precision (mAP) scores for the tested models in the zero-shot text-to-video retrieval tasks.", + "bbox": [ + 498, + 685, + 890, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We next present our results on the two downstream tasks, Text2Video Retrieval and Video Question Answering. Starting with the retrieval task, we report mean Average Precision (mAP) of the tested models on the SSv2-Temporal and SSv2-Events datasets in Table 4. The benefits of training with additional examples tailored for temporal video-language alignment is already evident in the performance of Owl-Rand, which improves over the previous SSv2-Events SOTA - TACT with a relative increase of $27\\%$ .", + "bbox": [ + 496, + 733, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, when training on harder negative contrastive instances, $Owl-Con$ achieves a significant improvement,", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13932", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/458d52e5b4438b2d8b32a26a88cf7b87c83d520e6458b892e38f14e26a2b8b4b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SuccessFailure
VideoCon LLMCaption: The adult gets the child's attention after tapping his shoulderECaption: Two people move a large object down a narrow stairwayE
Contrast Caption: The adult gets the child's attention before tappingCContrast Caption: Two people move a large object up a narrow stairwayE
GT NLE: First, an adult hand is seen tapping the child's shoulder, then the adult gets the child's attention, not the other way aroundE GTGT NLE: Men move a large piece of furniture down a narrow staircase, not up a narrow staircaseC GT
Predicted NLE: First, the adult taps the child's shoulder, then the adult gets the child's attention, not the other way aroundPredicted NLE: Three men are moving a large piece of furniture, not two
VideoCon HumanCaption: We see the group making cookiesECaption: A lady puts her hair in a bun on the sideE
Contrast Caption: We see the group eating cookiesCContrast Caption: A lady puts her hair in a braid on the sideE
GT NLE: The group is making cookies, not eating themE GTGT NLE: The woman puts her hair in a bun, not a braidC GT
Predicted NLE: We see the group cooking cookies, not eating themPredicted NLE: The lady puts her hair in a braid on top of her head, not on the side
", + "bbox": [ + 83, + 92, + 911, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "outperforming all baselines, with a relative increase over the best baseline End-to-End VNLI model by $7.5\\%$ on SSv2-Temporal and $9.6\\%$ on SSv2-Events (46% over TACT), setting new SOTA results. This points at the benefits of exposing the model to temporal examples, such as action and event-order.", + "bbox": [ + 75, + 364, + 468, + 454 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6cda5e936b47fd1ad8499354047b43aed06d552804ed4debad0495eca75d6f19.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsAccuracy (%)
CLIP23.8
VideoCLIP23.4
ImageBind (video-language)25.4
TACT [5]27.6
VFC [37]31.4
Owl-Base37.1
Owl-Rand37.2
End-to-End VNLI [60]39.0
Owl-Con (Ours)41.1
", + "bbox": [ + 109, + 462, + 436, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Accuracy scores for the tested models on the zero-shot video question-answering task on ATP-Hard dataset.", + "bbox": [ + 76, + 628, + 468, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the Video Question Answering task, we compare the performance of the various models in Table 5. Here too Owl-Con achieves SOTA results and outperforms the strongest baseline End-to-End VNLI model with a relative increase of $5.1\\%$ . This corroborates the observations in our other experiments, which demonstrate the advantage of the VideoCon datasets, covering various misalignments, especially those pertaining to temporal and causal reasoning over dynamic events. The results also confirm the need for carefully chosen contrastive negative examples, showing that picking negatives at random may mask out the potential benefit of an alignment training set. Finally, the competitive performance of atemporal End-to-End VNLI model on the downstream tasks is surprising and underscores the need for stronger video-language datasets for robust benchmarking.", + "bbox": [ + 75, + 674, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad723f65f8a9060a0ae314f8fac1fbbcd78fa7b8eb947d4dba5d3979b7235912.jpg", + "image_caption": [ + "Figure 6. Qualitative examples for the success (green) and failure (red) modes of our model. In every example, we present a few video frames in an temporal order from top to bottom, its associated caption, contrast caption, ground-truth NLE from the datasets. Additionally, we present the predicted NLE from our model. The small boxes at the end of caption cells indicate whether our model considers that caption to be grounded in the video. E and C indicates that the model predicts the caption to entail and contradict to the video, respectively. E-GT and C-GT indicates the predicted NLE entails and contradicts the ground-truth (GT) NLE, respectively.", + "Figure 7. ROC-AUC of End-to-End VNLI, Owl-Base, and OwlCon across all types of misalignment in VideoCon (LLM) test set." + ], + "image_footnote": [], + "bbox": [ + 540, + 363, + 851, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Analysis", + "text_level": 1, + "bbox": [ + 500, + 616, + 596, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze Owl-Con's performance improvements across the kinds of misalignments in VideoCon. Additionally, we present a few qualitative examples to highlight the success and failure modes of our model.", + "bbox": [ + 498, + 641, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Per-misalignment Entailment Results", + "text_level": 1, + "bbox": [ + 500, + 710, + 823, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compared the ROC-AUC scores of the atemporal End-to-End VNLI, Owl-Base, and Owl-Con on specific misalignments in the contrast captions from VideoCon (LLM) testset in Figure 7. We observed that Owl-Con outperforms the baseline models across all misalignment types. This suggests that our model can reason well about the entities, their relations, and the temporal order of events in the video.", + "bbox": [ + 496, + 734, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The largest improvement of $\\text{Owl-Con}$ compared to the two baselines is on event order flip, indicating that the baselines lack temporal understanding and the VideoCon is efficient in adding this capability to an alignment model. In", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13933", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "addition, on hallucination both Owl-Con and End-to-End VNLI significantly outperform Owl-Base, since both models were explicitly exposed to entailment/non-entailment training data. It is surprising to see that while End-to-End VNLI was trained on significantly more entailment data, much of it human-curated, Owl-Con outperforms it with only automatically generated data. This could be due to the better encoding of video in Owl-Con compared to the atemporal nature of End-to-End VNLI. Finally, the analysis shows other types of atemporal misalignments that are difficult for End-to-End VNLI to sort out, e.g. counting' and relation, where the training data in VideoCon is useful to improve these capabilities as well. This shows that our approach of detailed analysis of misalignment types of generation of examples for them is effective.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Qualitative Examples", + "text_level": 1, + "bbox": [ + 76, + 324, + 279, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We highlight a few classification examples of Owl-Con in Figure 6. The rows refer to the test source of the instances and the columns refer to the success and failure modes, respectively. In Row1/Column1, we observe that our model provides correct predictions for the entailment between the video and original caption while predicting contradiction for the contrast caption that flips the order of the events i.e., grabbing attention and tapping shoulders. Interestingly, our model can also provide the accurate NLE when prompted with the video and the contrast caption. This suggests that our model is useful for providing fine-grained details about the video-language alignment. In Row2/Column2, the model confuses 'buns' with 'braids' in hair and gives a wrong NLE that contradicts the ground-truth. This error, due to its inability to distinguish between objects, might be improved with diverse videos and captions.", + "bbox": [ + 75, + 348, + 472, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Related Work", + "text_level": 1, + "bbox": [ + 76, + 601, + 218, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Foundation Models for Video-Language Understanding. Foundation models have emerged for video-language understanding [1, 4, 49, 55, 56] by pre-training on large amount of video-text pairs scraped from the web [6, 36, 58]. Additionally, prior works have either leveraged the pretrained CLIP model for video-language tasks [12, 33, 34] or adopted a socratic approach [50, 63] to employ LLMs (GPT-3) in reasoning over video captions. We highlight that despite the large-scale training of the video-language foundation models [14, 55, 56], they lack robustness to semantic changes to the captions which severely limits their real-world use for alignment applications. We provide a fix by training models on a novel video-centric VideoCon dataset.", + "bbox": [ + 75, + 626, + 472, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improving Video-Language Robustness. Prior work [37, 38, 51] highlights that the video-text models cannot comprehend the semantics of the text with focus on manipulating the verb, actions, and entities grounded in the video", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "description. To improve the temporal understanding, [5] finetunes a pretrained model with temporal order loss. Despite this, their models do not achieve good zero-shot performance on downstream tasks consistently. In our work, we categorize a wide range of plausible misalignments in the contrast captions, and create a temporally-challenging VideoCon dataset.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video-Language Alignment Evaluation. Many applications such as text-to-video retrieval [15, 48, 57] and text-to-video generation [7, 47] require evaluation of the semantic alignment between the natural language text and raw video. In this work, we indicate that the existing video-text models such as VideoCLIP and ImageBind are not robust to semantic changes in the video captions, which becomes critical for faithful video-text alignment evaluation. In our work, we propose VideoCon and finetune a video-language generative model to perform robust entailment task and provide fine-grained NLE for the observed misalignments between the video and text. In the future, our model can be utilized to enhance alignment through sparse (entailment scores) and dense (fine-grained NLE) feedback [43].", + "bbox": [ + 496, + 215, + 893, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 441, + 619, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a comprehensive dataset, VideoCon, designed for robust video-text alignment. It features various semantic misalignments and explanations for text-video discrepancies. Through finetuning video-language models on this dataset, we enhanced their performance on complex tasks like text-to-video retrieval and video question answering, achieving state-of-the-art results.", + "bbox": [ + 496, + 467, + 890, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One current limitation and an important future direction is to increase the complexity of the generated contrast captions. Specifically, the model may encounter several misalignments within a single contrast caption. Addressing this issue, the model should be equipped to accurately assign low entailment scores to these contrast captions and consequently generate precise NLEs. An important future direction is to scale VideoCon to larger datasets. Here, we create contrast captions for high-quality captions written by humans for every video, however, the web-scale datasets have low-quality captions that are not well grounded in the video. In this regard, using synthetic data followed by VideoCon-like contrast caption generation can be a plausible approach.", + "bbox": [ + 496, + 574, + 893, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Acknowledgement", + "text_level": 1, + "bbox": [ + 498, + 784, + 679, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This material is based on research supported by the ECOLE program under Cooperative Agreement HR00112390060 with the US Defense Advanced Research Projects Agency (DARPA). In addition, the research is partly funded by ONR grant N00014-23-1-2780. Hritik Bansal is supported in part by AFOSR MURI grant FA9550-22-1-0380.", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13934", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 8", + "[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 4", + "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2", + "[4] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 8, 1", + "[5] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 1, 2, 3, 5, 6, 7, 8", + "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 8, 1", + "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 8, 1", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[9] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the \"video\" in video-language understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2917-2927, 2022. 1, 5, 6", + "[10] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 3", + "[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 1", + "[12] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021.8, 1", + "[13] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datcomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1", + "[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1, 2, 5, 6, 8", + "[15] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 8, 1, 5", + "[16] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Sarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 1", + "[17] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 1, 3, 2", + "[18] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 2", + "[19] Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. Q2: Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. arXiv preprint arXiv:2104.08202, 2021. 4, 5", + "[20] https://commoncrawl.org/. 1", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[22] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897, 2023. 1", + "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1", + "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13935", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3", + "[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 5, 6", + "[27] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5", + "[28] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. arXiv preprint arXiv:2306.14060, 2023. 1", + "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3", + "[30] Weixin Liang, James Zou, and Zhou Yu. Alice: Active learning with contrastive natural language explanations. arXiv preprint arXiv:2009.10259, 2020. 1", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6", + "[32] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3", + "[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 8, 1", + "[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 8, 1", + "[35] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 3", + "[36] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 8, 1", + "[37] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 1, 2, 3, 5, 7, 8", + "[38] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, 2022. 1, 2, 3, 8", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5", + "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[42] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8322-8332, 2022. 1", + "[43] Jérémy Scheurer, Jon Ander Campos, Tomasz Korbak, Jun Shern Chan, Angelica Chen, Kyunghyun Cho, and Ethan Perez. Training language models with language feedback at scale. arXiv preprint arXiv:2303.16755, 2023. 8, 1", + "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1", + "[45] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 535-544, 2021. 2, 5", + "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5", + "[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 8, 1", + "[48] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 8, 1", + "[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 8, 1" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13936", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems, 35: 8483-8497, 2022. 8, 1", + "[51] Zhenhailong Wang, Ansel Blume, Sha Li, Genglin Liu, Jaemin Cho, Zineng Tang, Mohit Bansal, and Heng Ji. Paxion: Patching action knowledge in video-language foundation models. arXiv preprint arXiv:2305.10683, 2023. 1, 2, 8", + "[52] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5", + "[53] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment task for visually-grounded language learning. arXiv preprint arXiv:1811.10582, 2018. 2", + "[54] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding. arXiv preprint arXiv:1901.06706, 2019. 2", + "[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 5, 6, 8", + "[56] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023.8.1", + "[57] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 3, 8, 1", + "[58] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8, 1", + "[59] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2", + "[60] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400, 2023. 3, 6, 7, 1, 2", + "[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 5, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1", + "[63] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 8, 1", + "[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI Conference on Artificial Intelligence, pages 7590-7598, 2018. 7" + ], + "bbox": [ + 501, + 92, + 890, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13937", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_model.json b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bfb5c273d636a83962835363d6119f8cb56cf1e4 --- /dev/null +++ b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "VideoCon: Robust Video-Language Alignment via Contrast Captions" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.18, + 0.846, + 0.253 + ], + "angle": 0, + "content": "Hritik Bansal1 Yonatan Bitton2 Idan Szpektor2* Kai-Wei Chang1* Aditya Grover1* \n1UCLA 2Google Research \n{hbansal, kwchang, adityag}@cs.ucla.edu \n{yonatanbitton, szpektor}@google.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Despite being (pre)trained on a massive amount of data, state-of-the-art video-language alignment models are not robust to semantically-plausible contrastive changes in the video captions. Our work addresses this by identifying a broad spectrum of contrast misalignments, such as replacing entities, actions, and flipping event order, which alignment models should be robust against. To this end, we introduce the VideoCon, a video-language alignment dataset constructed by a large language model that generates plausible contrast video captions and explanations for differences between original and contrast video captions. Then, a generative video-language model is finetuned with VideoCon to assess video-language entailment and generate explanations. Our VideoCon-based alignment model significantly outperforms current models. It exhibits a 12-point increase in AUC for the video-language alignment task on human-generated contrast captions. Finally, our model sets new state of the art zero-shot performance in temporally-extensive video-language tasks such as text-to-video retrieval (SSv2-Temporal) and video question answering (ATP-Hard). Moreover, our model shows superior performance on novel videos and human-crafted captions and explanations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Semantically aligning data points from diverse modalities is a long-standing goal of AI. We focus on video-language alignment, which is challenging due to the complexities involved in understanding of entities, their relationships, and temporal order of the depicted events [17]. Recent models such as VideoCLIP [55], ImageBind [14] learn a shared embedding space. Similarly, generative models such as Flamingo [1], mPLUG-Owl-Video [61] can provide a classification label (e.g., yes/no) when queried about video-language alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Despite large-scale pretraining, prior work [5, 37, 38, 51] highlights that video-language alignment models are not robust to semantically plausible manipulations to an original aligned caption in the form of contrast captions, such as from 'dog runs away before it eats food' to 'dog runs away after it eats food'. Such pitfalls in robustness questions the trustworthiness of alignment models for large-scale deployment. To mitigate these shortcomings, one possible solution is to scale video-language pairs more for increased diversity during pretraining. However, this is challenging due to the difficulties in sourcing new, high-quality and permissible content, as well as the requirement for substantial storage capacity. Several works [11, 13, 16] have shown that naively training models on web-scale data has diminishing returns on downstream tasks, and emphasize the importance of data quality. Furthermore, the recent studies [28, 62] demonstrate that applying a contrastive objective to the pretraining datasets does not encourage the model to grasp the fine-grained details within image/region-caption data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To this end, we take a scalable, active strategy to gather high-quality data that is deliberately enriched with the attributes that we want to instill in alignment models. We create a novel dataset, VideoCon, to improve the robustness of models. Specifically, the dataset consists of a variety of semantically plausible video-language misalignments in contrast captions. These misalignments include altering objects (entities), actions, attributes, relations, counts, event orders, and introducing hallucinations (Figure 2). To construct VideoCon, a large language model (PaLM-2 API) takes video-caption pairs as input and generates high-quality contrast captions for a given misalignment type. To make our dataset temporally-challenging, we skipped \"easy\" video-caption pairs whose alignment could be inferred based on a single frame (image) understanding [9, 26] (\\(\\S 3.1\\)). In addition, the LLM generates natural language explanations (NLEs) [42] to the differences between original and altered captions, which are used for further robust training. We performed human verification on a sample of VideoCon and found that it is of high-quality. Finally, to evaluate the model's generalization capabilities, we col" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.192, + 0.901 + ], + "angle": 0, + "content": "*Equal Advising." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.885, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.894, + 0.281 + ], + "angle": 0, + "content": "Figure 1. Overview of our VideoCon approach. First, aligned video-language pairs are filtered to retain temporally-challenging instances. Then contrast captions and natural language explanations (NLE) are generated by an LLM to create the VideoCon dataset. Second, a video-language alignment model is finetuned with VideoCon on the alignment and NLE tasks. Finally, the finetuned model is evaluated against the baseline model. Our results show that it outperforms the baseline, achieving state-of-the-art results on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.47, + 0.354 + ], + "angle": 0, + "content": "lect human-generated contrast captions and NLEs for the videos sourced from external datasets that did not contribute to VideoCon's development." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.566 + ], + "angle": 0, + "content": "We finetuned a generative video-language model (mPLUG-Owl-Video) on the VideoCon dataset. The trained model surpasses existing video-language alignment models by a large margin on the LLM-generated test set for both video-language alignment and NLE generation tasks. Interestingly, we observed that our finetuned model generalizes to unseen videos and human-generated contrast captions and NLEs, and outperforms the baseline models. For instance, our model's ROC-AUC exceeds the baseline model by 12 points on the human-generated contrast captions. This indicates that our model has developed a better understanding of the entities, their interactions, action understanding, as well as the temporal order of the events for robust video-language alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.793 + ], + "angle": 0, + "content": "We further assessed the effectiveness of robust training via contrast captions on zero-shot downstream videolanguage tasks such text-to-video retrieval and video question answering on the temporally-challenging and action-intensive SSv2-Temporal [45] and SSv2-Events [5]. Our model achieves state-of-the-art (SOTA) performance, improving on SSv2-Temporal by \\(4.3\\mathrm{mAP}\\), SSv2-Events by \\(3.6\\mathrm{mAP}\\) points. In addition, our model also achieves SOTA on temporal and causal video question answering in the ATP-Hard dataset, increasing \\(4\\%\\) accuracy. This suggests that equipping a model with the knowledge of contrast captions is highly data-efficient and effective in improving its robustness in comparison to scaling the pretraining data. The complete pipeline is illustrated in Figure 1. The dataset and the model will be released upon acceptance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.332, + 0.826 + ], + "angle": 0, + "content": "2. Video Language Alignment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.834, + 0.469, + 0.865 + ], + "angle": 0, + "content": "We are interested in assessing the semantic alignment between the video1 and text data since it powers many prac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "1Like prior works [33, 55], we use only the video frames (the visual channel) without the soundtrack (the audio channel)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.459 + ], + "angle": 0, + "content": "tical applications such as video-text retrieval [57], video generation [7, 47] and video captioning [59]. To this end, [14, 39, 49, 55] designed (image)video-text alignment models that are utilized for evaluating the semantic similarity between the two modalities. However, previous works [5, 37, 38, 51] have questioned their robustness to semantically plausible changes to the video descriptions, termed here contrast captions. Our aim is to improve the robustness of video-text alignment models by training on contrast captions with a wide range of misalignments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Consider a dataset \\(\\mathcal{D} = \\{(V_i, T_i, C_i, E_i)\\}\\) where \\(V_i\\) is a video, \\(T_i\\) is an aligned caption, \\(C_i\\) is a contrast caption which is a perturbation of \\(T_i\\) but misaligns with \\(V_i\\), and \\(E_i\\) is a natural language explanation for the misalignment between \\(V_i\\) and \\(C_i\\). We consider two video-language alignment tasks: (a) video-language entailment, (b) natural language explanation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Video-Language Entailment (VLE) casts video-text alignment as a Visual Entailment (VE) task. VE was originally defined for images as premises and texts as hypothesis [53, 54]. We extend VE definition also for videos as premises, under which a classification model \\( A_{vle}(V,T) \\) predicts whether a video \\( V \\) entails a text \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Natural Language Explanation (NLE) requires a model, \\( A_{nle}(V,C) \\), to generate an open-ended explanation for the discrepancy between a video \\( V \\) and a non-entailing caption \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.892, + 0.814 + ], + "angle": 0, + "content": "In this paper, we address both VLE and NLE tasks under a multitask setting in which a single video-language generative model generates the binary label for entailment and the open-ended explanation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.827, + 0.892, + 0.863 + ], + "angle": 0, + "content": "3. VideoCon: Contrast Captions Generation for Robust Video-Language Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our research goal is to measure the impact of a comprehensive dataset on increasing the robustness of video-text align-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.089, + 0.798, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.39 + ], + "angle": 0, + "content": "Figure 2. Overview of the VideoCon data generation process from top to bottom. Specifically, we prompt a large language model (PaLM-2) with the original caption that is grounded in the video, and the intended type of misalignment within the contrast caption. We consider seven kinds of misalignments including object, action, attribute, counting, spatial relation, hallucination, and event order flip. We provide a generated contrast caption and the corresponding natural language explanation for each misalignment type." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.473, + 0.566 + ], + "angle": 0, + "content": "ment models. To this end, we first collect video-caption pairs where the caption cannot be derived from a single frame of video. We then categorize a wide range of semantically plausible manipulations of video captions. Using an LLM for large-scale computation, contrast captions and related explanations are generated for the defined categories, constructing the VideoCon dataset. Finally, we extend VideoCon to include human-created contrast captions as held-out evaluation on unseen videos. We detail the dataset construction steps below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.583, + 0.444, + 0.6 + ], + "angle": 0, + "content": "3.1. Temporally-Challenging Instance Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.715 + ], + "angle": 0, + "content": "To construct VideoCon, we start with existing datasets that include natural (real) videos and associated high-quality human-written captions: MSR-VTT [57], VaTeX [48], and TEMPO [17]. MSR-VTT and VaTeX consist of 20 captions and 10 captions per video, respectively, while TEMPO consists of a single caption per video. More dataset details are in Appendix §B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.471, + 0.852 + ], + "angle": 0, + "content": "TEMPO is designed to create temporally-challenging instances, while MSR-VTT and VaTeX contain more general video-caption pairs. For MSR-VTT and VaTeX, we filter out instances, where the caption is highly associated with a single frame in the video based on an image-text alignment model. In such cases, a video-text alignment can leverage shortcuts and align the video to its caption without understanding the temporal or causal relations depicted in the video. We want to filter such instances." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To this end, we employ the End-to-End VNLI model [60] to calculate an alignment score \\( A_{vle}(V,T) \\) between a video \\( V = \\{I_1,I_2,\\dots ,I_N\\} \\) and a text \\( T \\) where \\( I_{i} \\) is a frame from" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.43 + ], + "angle": 0, + "content": "the video sampled at a rate of 1 frame per second. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.442, + 0.892, + 0.46 + ], + "angle": 0, + "content": "\\[\nA _ {v l e} (V, T) = \\max _ {i} \\left(V N L I \\left(I _ {i}, T\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.893, + 0.607 + ], + "angle": 0, + "content": "where \\( VNLI(I_i, T) \\) is the task of visual natural language inference that assesses the extent to which the text \\( T \\) entails the image \\( I_i \\). There are 20 and 10 captions per video in the MSR-VTT and VaTeX datasets, respectively. We retain 5 captions per video from these datasets with the lowest \\( A_{vle}(V, T) \\), and the remaining captions are filtered out. Post-filtering, the percentage of temporally-challenging instances increased from \\( 36.5\\% \\) to \\( 81.5\\% \\) in MSR-VTT, and from \\( 42.6\\% \\) to \\( 71\\% \\) in VaTeX." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.617, + 0.785, + 0.634 + ], + "angle": 0, + "content": "3.2. Categories of Contrast Captions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.838 + ], + "angle": 0, + "content": "We aim for VideoCon to include a wide range of misalignments in its contrast captions. Overall, VideoCon covers seven misalignment types, exemplified in Figure 2. We include replacement of objects (entities) and actions following the analysis in [37, 38], and replacement of attributes, counts, relations, as well as adding unrelated but plausible information to captions as hallucinations following [29, 32, 35]'s study of image/text alignment model brittleness. Since most video-text models rely on pretrained image backbones, they are likely to suffer from similar problems. Finally, following [5]'s analysis that video-text models do not understand temporal order of the events, we include event order flipping as misalignment type." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.78, + 0.864 + ], + "angle": 0, + "content": "3.3. Data Generation using an LLM" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To generate contrast captions and corresponding NLE we first assign one of the seven misalignment types (§3.2) to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.15, + 0.104, + 0.398, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.471, + 0.312 + ], + "angle": 0, + "content": "Figure 3. Distribution of the types of misalignments within the contrast captions of the VideoCon dataset. We observe that the dataset has good representation for all the kinds of misalignments ranging from \\(8.8\\%\\) to \\(24.2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.468, + 0.414 + ], + "angle": 0, + "content": "each caption in the input video-text datasets (§3.1) (details in Appendix §C). Then, given a video \\( V \\) and a misalignment type \\( m \\), we prompt PaLM-2 API² [2] to generate a contrast caption and accompanied explanation (our type-specific prompts are detailed in Appendix §D)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.468, + 0.653 + ], + "angle": 0, + "content": "Analyzing the LLM generations, we found that sometimes the output caption \\( C \\) do not contradict the original caption \\( T \\). For example, a generated contrast caption \"a person riding a car\" does not contradict the original caption \"a person riding a mustang\". To filter such cases, we employ a Natural Language Inference (NLI) model [19] and remove cases in which the contrast caption is assessed as entailed by the original caption \\( NLI(T,C) > 0.5 \\). Postfiltering, each tuple \\( (V,T,C,m) \\) is converted to the two instances of video/language entailment task: \\( A_{en}(V,T) = 1 \\) and \\( A_{en}(V,C) = 0 \\). We present the dataset statistics for the entailment task in Table 1, including train/eval/test splits. In addition, Fig. 3 shows the distribution of misalignment types in the dataset. We observe that VideoCon maintains a high density across the 7 misalignments ranging from \\( 8.8\\% \\) to \\( 24.2\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.468, + 0.866 + ], + "angle": 0, + "content": "We also found that some generated explanations do not describe the differences between \\( T \\) and \\( C \\) well. For example, the explanation \"two friends are not traveling together\" does not fully describe the discrepancy between \"three friends traveling together\" and \"two friends are traveling together\". To filter these out, generated examples are removed if \\( NLI(F(T,C),E) < 0.6 \\) where \\( F(T,C) \\) is the premise comprising the original and contrast captions. Specifically, premise will be 'Expected Caption: \\( T \\) Actual Caption: \\( E' \\) and hypothesis will be 'Difference between Expected and Actual Caption: \\( E' \\). This filter indicates that the information in the explanation is not entailed by the difference between the two captions. The dataset statistics for the NLE task is presented in Table 1. We refer to the final" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.158 + ], + "angle": 0, + "content": "
Video-Language EntailmentNatural Language Explanation
SourceTrainValTestTrainValTest
MSR-VTT3836647816538158882066788
VaTeX664807368110301803453636
TEMPO1071270982708416527391073
Total11555883122735650233329011497
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.168, + 0.872, + 0.181 + ], + "angle": 0, + "content": "Table 1. Statistics for the VLE and NLE tasks in VideoCon." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.212, + 0.8, + 0.227 + ], + "angle": 0, + "content": "LLM-generated dataset as VideoCon (LLM)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.304 + ], + "angle": 0, + "content": "To assess the quality of VideoCon (LLM), we perform human evaluation on 500 contrast captions and NLEs (details in Appendix E). The human evaluator found \\(91\\%\\) of the contrast captions and \\(89\\%\\) of the NLEs to be valid, indicating the high-quality of VideoCon (LLM)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.779, + 0.334 + ], + "angle": 0, + "content": "3.4. Data Generation using Humans" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.553 + ], + "angle": 0, + "content": "To study whether a model trained on VideoCon (LLM) generalizes to out-of-distribution videos and its performance on human-generated contrast captions, we randomly selected a set of videos from the validation set of ActivityNet [10]. This dataset consists of captions matched with segments in the video, e.g., \"a little boy is climbing on an outside gym\" matched to the first 10 seconds of its related video. We extracted video segments with an associated caption. Human workers3 on Amazon MTurk were then shown the video segments and associated captions and were asked to create a semantically plausible contrast caption and a corresponding NLE (more details in Appendix §F). We did not communicate any type of target misalignments to encourage natural diversity of human created contrast captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.797 + ], + "angle": 0, + "content": "Overall, we collected 570 tuples \\((V, T, C_{human}, E_{human})\\) where \\(V\\) is the video, \\(T\\) is the original caption, \\(C_{human}\\) is the human-written contrast caption, and \\(E_{human}\\) is the human-written explanations. We denote this dataset by VideoCon (Human). We sample 100 instances from this dataset, and found \\(93\\%\\) to be clean. In addition, we observe that many of the human-generated contrast captions perturbing one or more objects (\\(35\\%\\)) and actions (\\(35\\%\\)) depicted in the caption. While \\(8\\% - 10\\%\\) of the contrast captions flip the order of the events and attribute of the objects. As this dataset is largely unfiltered, it contains a mix of temporally-easy and challenging instances. We also constructed a more temporally-challenging subset of 290 instances, denoted VideoCon (Human-Hard), by filtering out tuples in which \\(A_{vle}(V, T) < 0.5\\) (Eq. (1)), as in §3.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.692, + 0.833 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.842, + 0.891, + 0.872 + ], + "angle": 0, + "content": "We next describe our evaluation setting for measuring the impact of VideoCon on video-text alignment modeling." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https : / / developers . generativeai . google / products/palm" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.757, + 0.9 + ], + "angle": 0, + "content": "3A shortlist that passed our qualification test." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.317, + 0.108 + ], + "angle": 0, + "content": "4.1. Finetuning with VideoCon" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.206 + ], + "angle": 0, + "content": "Our goal in constructing VideoCon (LLM) is to improve robustness of video-text alignment models by fine-tuning on this dataset. To this end, we start with the mPLUG-OwlVideo model [61], denoted Owl-Base. Its building blocks are CLIP [39] as visual encoder and LLaMA-7B [46] as text encoder/decoder and it was pretrained on VideoChat [27]." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.223, + 0.207, + 0.235 + ], + "angle": 0, + "content": "Entailment Task:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.246, + 0.386, + 0.26 + ], + "angle": 0, + "content": "Given: V (Video), T (Caption), C (Contrast Caption)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.269, + 0.438, + 0.295 + ], + "angle": 0, + "content": "Instruction (I): [V] Does this video entail the description \\([T]\\)? Response (R): Yes" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.306, + 0.438, + 0.331 + ], + "angle": 0, + "content": "Instruction (I): [V] Does this video entail the description [C]? Response (R): No" + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.344, + 0.416, + 0.358 + ], + "angle": 0, + "content": "Figure 4. Entailment task prompt for finetuning." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.393, + 0.402, + 0.407 + ], + "angle": 0, + "content": "Natural Language Explanation Generation Task:" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.419, + 0.393, + 0.433 + ], + "angle": 0, + "content": "Given: V (Video), C (Contrast Caption), E (NLE)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.445, + 0.442, + 0.485 + ], + "angle": 0, + "content": "Instruction (I): [V] What is the misalignment between this video and the description \\([C]\\)? Response (R): [E]" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.5, + 0.431, + 0.514 + ], + "angle": 0, + "content": "Figure 5. NLE generation task prompt for finetuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.532, + 0.47, + 0.638 + ], + "angle": 0, + "content": "To fine-tune \\(Owl\\)-Base on VideoCon (LLM), its \\(\\{V, T, C, E\\}^4\\) tuples were converted into two types of multimodal instruction-response pairs, one for the VLE task \\((I_{vle}, R)\\) (Fig. 4) and one for the NLE task \\((I_{nle}, R)\\) (Fig. 5). We then train \\(Owl\\)-Base on all instruction pairs from both the tasks with maximum likelihood loss, resulting in a single model \\(Owl\\)-Con." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.341, + 0.665 + ], + "angle": 0, + "content": "4.2. VideoCon Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.469, + 0.813 + ], + "angle": 0, + "content": "To evaluate the performance of the \\(Owl-Con\\) on video-text alignment we generate \\(Owl-Con\\) response to prompt \\(I_{vle}\\) for video \\(V\\) and text \\(Y \\in \\{T, C\\}\\). We then calculate the probability of generating responses \\(s_y = Owl-Con\\) (\\(^\\prime \\mathrm{Yes}^\\prime | I_{vle}(V, Y)\\)) and \\(s_n = Owl-Con\\) (\\(^\\prime \\mathrm{No}^\\prime | I_{vle}(V, Y)\\)), and based on these scores the probability for class 'Yes': \\(P_{yes}(V, Y) = \\frac{s_y}{s_y + s_n}\\). Finally, we compute the ROC-AUC score for \\(P_{yes}(V, Y)\\) over the VideoCon (LLM) eval set, with \\(\\{V, T\\}\\) as label 1 and \\(\\{V, C\\}\\) as label 0." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.814, + 0.469, + 0.875 + ], + "angle": 0, + "content": "To evaluate \\(Owl\\)-Con on the NLE task, we prompt it with instruction \\(I_{nle}\\) instantiated on \\(\\{V,C\\}\\) pairs from the VideoCon (LLM) eval set. We compare the generated explanation \\(\\hat{E}\\) to the ground truth \\(E\\) by measuring entailment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.183 + ], + "angle": 0, + "content": "probability \\( NLI(E, \\hat{E}) \\). In our experiments, we experiment with two \\( NLI \\) automatic metrics: (a) \\( Q^2 \\) score [19], and (b) PaLM-2 API. We performed human evaluation to measure the agreement between the automatic metrics and the human-rating. We found that both metrics achieve high agreement with human assessment (Appendix §H)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.19, + 0.77, + 0.205 + ], + "angle": 0, + "content": "4.3. Video-Text Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.257 + ], + "angle": 0, + "content": "We complement the VideoCon intrinsic evaluation over the testset with an extrinsic evaluation over two temporal and action difficult downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.893, + 0.456 + ], + "angle": 0, + "content": "We evaluate alignment model performance for text2video retrieval over SSv2-Temporal [45] and SSv2-Events [5] datasets. We consider the SSv2-Template captions instead of the label captions since they remove the object-centric bias in model evaluation [26]. We compute input-text/candidate-video alignment score, rank videos and report mean Average Precision (mAP). We evaluate alignment model performance for video question answering over the ATP-Hard [9] dataset. We cast each question/candidate-answer pair as an imperative statement using PaLM-2 API, measure alignment to the input video and report Accuracy. More details on the downstream datasets and the evaluation setup are in Appendix §I." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.608, + 0.477 + ], + "angle": 0, + "content": "4.4. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.892, + 0.652 + ], + "angle": 0, + "content": "For the video-text alignment text, we compare Owl-Con with the following baselines: (a) End-to-End VNLI as zero-shot atemporal model since it does not have access to the temporal order of the video frames, (b) VideoCLIP [55], (c) ImageBind [14], (d) Owl-Base, and (e) Owl-Rand: Owl-Base fine-tuned on VideoCon tuples \\(\\{V,T,\\hat{C},E\\}\\) where \\(\\hat{C}\\) is randomly selected from other captions in the dataset. Owl-Rand would indicate if there is merit in the contrast, hard-negative captions in VideoCon. We include additional baselines TACT [5] and VFC [37] for evaluating on the downstream tasks (\\$5.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.633, + 0.681 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.891, + 0.735 + ], + "angle": 0, + "content": "We present our intrinsic (VideoCon eval set) and extrinsic (downstream tasks) evaluation results, showing the benefits of VideoCon for robust video-language alignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.87, + 0.757 + ], + "angle": 0, + "content": "5.1. Performance on VideoCon Entailment Task" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We present the ROC-AUC scores of the tested models in Table 2. From the table we see that the baseline models find the VideoCon testset difficult, as reflected by low AUC scores (e.g. Owl-Base-57.2), close to random. Even training on VideoCon train instances, but with \"easy\" negatives (Owl-Rand-59.7), hardly improves the base models. A significant improvement is achieved with the VNLI-specific model (67), showing that the entailment task is not inherently represented in generic video-language aligned training" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.443, + 0.901 + ], + "angle": 0, + "content": "\\(^{4}\\mathrm{V}\\) : video, T: original caption, C: contrast caption, E: explanation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13931" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.088, + 0.858, + 0.215 + ], + "angle": 0, + "content": "
ModelsVideoCon (LLM) TestVideoCon (Human)VideoCon (Human-Hard)
Random50.050.050.0
VideoCLIP [55]53.247.347.5
ImageBind (Video-Text) [14]57.165.263.0
Owl-Base [61]57.266.864.1
Owl-Rand59.768.965.5
End-to-End VNLI [60]67.072.465.0
Owl-Con (Ours)84.678.374.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.208, + 0.225, + 0.761, + 0.239 + ], + "angle": 0, + "content": "Table 2. ROC-AUC scores of the tested models for the entailment task on VideoCon test sets." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.252, + 0.894, + 0.322 + ], + "angle": 0, + "content": "
VideoCon (LLM)VideoCon (Human)
ModelsQ2entailmentPaLM-2 entailment acc. (%)Q2entailmentPaLM-2 entailment acc.(%)
Owl-Base0.1936.80.2339.6
Owl-Con (Ours)0.5065.40.3247.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.175, + 0.331, + 0.794, + 0.344 + ], + "angle": 0, + "content": "Table 3. Performance of the tested models on the NLE generation task, measured via entailment metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.476 + ], + "angle": 0, + "content": "sets and requires specific training. Yet, the best performance is achieved by training on VideoCon, which addresses the diversity in plausible misalignments and includes \"difficult\" training examples, reaching 84.6 AUC. This demonstrates the merit of VideoCon for improving video-language alignment robustness. We show qualitative examples for the model predictions in §6.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.612 + ], + "angle": 0, + "content": "When evaluating on out-of-domain (OOD) data around video types and misalignment distribution, we again see that training with VideoCon offers significant improvement to alignment detection, outperforming all baselines, albeit with smaller relative gains: \\(17\\%\\) and \\(16\\%\\) improvement compared to Owl-Base on (Human) and (Human-Hard) respectively compared to \\(48\\%\\) on (LLM) test. In future work, we plan to further diversify the misalignments VideoCon covers to further improve its benefits on OOD cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.719 + ], + "angle": 0, + "content": "We notice that the performance of the VNLI atemporal model is better than existing video-language alignment models. It might be attributed to its training with contrast captions in [60]. It further highlights that the existing videolanguage models are not robust in comparison to a temporal probe on video-language alignment evaluation, corroborating the findings from [9, 26]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.411, + 0.743 + ], + "angle": 0, + "content": "5.2. Performance on NLE Generation Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Table 3 presents the performance of the tested models against the ground-truth on the NLE task, depicting average \\( Q^2 \\) score and PaLM-2 entailment accuracy. The results show that on in-domain VideoCon, Owl-Con outperforms Owl-Base by an impressive \\( 263\\% \\) and \\( 178\\% \\) relative increase on \\( Q^2 \\) score and PaLM-2 accuracy respectively. This indicates the finetuned model can accurately generate NLE that match well with the ground-truth NLE. This indicates that our model can generate accurate NLE for a wide range of misalignments in the video captions, which makes it use" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.836, + 0.386 + ], + "angle": 0, + "content": "ful for dense video-language alignment evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.493 + ], + "angle": 0, + "content": "On out-of-domain VideoCon, the improvement is more moderate but still high: \\(40\\%\\) and \\(20\\%\\) relative increase on \\(Q^2\\) and PaLM-2 respectively. This is probably due to the more diverse ways humans express explanations compared to LLM prompting. In future work we plan to further address linguistic diversity in explanations for more robust generation and evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.892, + 0.52 + ], + "angle": 0, + "content": "5.3. Performance on Video-Text Downstream Tasks" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.538, + 0.892, + 0.676 + ], + "angle": 0, + "content": "
ModelsSSv2-Temporal\nmAPSSv2-Events\nmAP
Random7.33.3
VideoCLIP9.86.4
ImageBind (video-language)10.55.5
Owl-Base10.96.8
TACT [5]-7.8
Owl-Rand12.19.9
End-to-End VNLI [60]14.610.4
Owl-Con (Ours)15.211.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.686, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Table 4. Mean Average Precision (mAP) scores for the tested models in the zero-shot text-to-video retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We next present our results on the two downstream tasks, Text2Video Retrieval and Video Question Answering. Starting with the retrieval task, we report mean Average Precision (mAP) of the tested models on the SSv2-Temporal and SSv2-Events datasets in Table 4. The benefits of training with additional examples tailored for temporal video-language alignment is already evident in the performance of Owl-Rand, which improves over the previous SSv2-Events SOTA - TACT with a relative increase of \\(27\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, when training on harder negative contrastive instances, \\(Owl-Con\\) achieves a significant improvement," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13932" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.093, + 0.913, + 0.253 + ], + "angle": 0, + "content": "
SuccessFailure
VideoCon LLMCaption: The adult gets the child's attention after tapping his shoulderECaption: Two people move a large object down a narrow stairwayE
Contrast Caption: The adult gets the child's attention before tappingCContrast Caption: Two people move a large object up a narrow stairwayE
GT NLE: First, an adult hand is seen tapping the child's shoulder, then the adult gets the child's attention, not the other way aroundE GTGT NLE: Men move a large piece of furniture down a narrow staircase, not up a narrow staircaseC GT
Predicted NLE: First, the adult taps the child's shoulder, then the adult gets the child's attention, not the other way aroundPredicted NLE: Three men are moving a large piece of furniture, not two
VideoCon HumanCaption: We see the group making cookiesECaption: A lady puts her hair in a bun on the sideE
Contrast Caption: We see the group eating cookiesCContrast Caption: A lady puts her hair in a braid on the sideE
GT NLE: The group is making cookies, not eating themE GTGT NLE: The woman puts her hair in a bun, not a braidC GT
Predicted NLE: We see the group cooking cookies, not eating themPredicted NLE: The lady puts her hair in a braid on top of her head, not on the side
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Figure 6. Qualitative examples for the success (green) and failure (red) modes of our model. In every example, we present a few video frames in an temporal order from top to bottom, its associated caption, contrast caption, ground-truth NLE from the datasets. Additionally, we present the predicted NLE from our model. The small boxes at the end of caption cells indicate whether our model considers that caption to be grounded in the video. E and C indicates that the model predicts the caption to entail and contradict to the video, respectively. E-GT and C-GT indicates the predicted NLE entails and contradicts the ground-truth (GT) NLE, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.455 + ], + "angle": 0, + "content": "outperforming all baselines, with a relative increase over the best baseline End-to-End VNLI model by \\(7.5\\%\\) on SSv2-Temporal and \\(9.6\\%\\) on SSv2-Events (46% over TACT), setting new SOTA results. This points at the benefits of exposing the model to temporal examples, such as action and event-order." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.463, + 0.437, + 0.619 + ], + "angle": 0, + "content": "
ModelsAccuracy (%)
CLIP23.8
VideoCLIP23.4
ImageBind (video-language)25.4
TACT [5]27.6
VFC [37]31.4
Owl-Base37.1
Owl-Rand37.2
End-to-End VNLI [60]39.0
Owl-Con (Ours)41.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.63, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Table 5. Accuracy scores for the tested models on the zero-shot video question-answering task on ATP-Hard dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.901 + ], + "angle": 0, + "content": "For the Video Question Answering task, we compare the performance of the various models in Table 5. Here too Owl-Con achieves SOTA results and outperforms the strongest baseline End-to-End VNLI model with a relative increase of \\(5.1\\%\\). This corroborates the observations in our other experiments, which demonstrate the advantage of the VideoCon datasets, covering various misalignments, especially those pertaining to temporal and causal reasoning over dynamic events. The results also confirm the need for carefully chosen contrastive negative examples, showing that picking negatives at random may mask out the potential benefit of an alignment training set. Finally, the competitive performance of atemporal End-to-End VNLI model on the downstream tasks is surprising and underscores the need for stronger video-language datasets for robust benchmarking." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.364, + 0.852, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.565, + 0.891, + 0.593 + ], + "angle": 0, + "content": "Figure 7. ROC-AUC of End-to-End VNLI, Owl-Base, and OwlCon across all types of misalignment in VideoCon (LLM) test set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.597, + 0.634 + ], + "angle": 0, + "content": "6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.703 + ], + "angle": 0, + "content": "We analyze Owl-Con's performance improvements across the kinds of misalignments in VideoCon. Additionally, we present a few qualitative examples to highlight the success and failure modes of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.824, + 0.728 + ], + "angle": 0, + "content": "6.1. Per-misalignment Entailment Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.84 + ], + "angle": 0, + "content": "We compared the ROC-AUC scores of the atemporal End-to-End VNLI, Owl-Base, and Owl-Con on specific misalignments in the contrast captions from VideoCon (LLM) testset in Figure 7. We observed that Owl-Con outperforms the baseline models across all misalignment types. This suggests that our model can reason well about the entities, their relations, and the temporal order of events in the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The largest improvement of \\( \\text{Owl-Con} \\) compared to the two baselines is on event order flip, indicating that the baselines lack temporal understanding and the VideoCon is efficient in adding this capability to an alignment model. In" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13933" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "addition, on hallucination both Owl-Con and End-to-End VNLI significantly outperform Owl-Base, since both models were explicitly exposed to entailment/non-entailment training data. It is surprising to see that while End-to-End VNLI was trained on significantly more entailment data, much of it human-curated, Owl-Con outperforms it with only automatically generated data. This could be due to the better encoding of video in Owl-Con compared to the atemporal nature of End-to-End VNLI. Finally, the analysis shows other types of atemporal misalignments that are difficult for End-to-End VNLI to sort out, e.g. counting' and relation, where the training data in VideoCon is useful to improve these capabilities as well. This shows that our approach of detailed analysis of misalignment types of generation of examples for them is effective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.281, + 0.342 + ], + "angle": 0, + "content": "6.2. Qualitative Examples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.473, + 0.591 + ], + "angle": 0, + "content": "We highlight a few classification examples of Owl-Con in Figure 6. The rows refer to the test source of the instances and the columns refer to the success and failure modes, respectively. In Row1/Column1, we observe that our model provides correct predictions for the entailment between the video and original caption while predicting contradiction for the contrast caption that flips the order of the events i.e., grabbing attention and tapping shoulders. Interestingly, our model can also provide the accurate NLE when prompted with the video and the contrast caption. This suggests that our model is useful for providing fine-grained details about the video-language alignment. In Row2/Column2, the model confuses 'buns' with 'braids' in hair and gives a wrong NLE that contradicts the ground-truth. This error, due to its inability to distinguish between objects, might be improved with diverse videos and captions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.22, + 0.618 + ], + "angle": 0, + "content": "7. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.473, + 0.825 + ], + "angle": 0, + "content": "Foundation Models for Video-Language Understanding. Foundation models have emerged for video-language understanding [1, 4, 49, 55, 56] by pre-training on large amount of video-text pairs scraped from the web [6, 36, 58]. Additionally, prior works have either leveraged the pretrained CLIP model for video-language tasks [12, 33, 34] or adopted a socratic approach [50, 63] to employ LLMs (GPT-3) in reasoning over video captions. We highlight that despite the large-scale training of the video-language foundation models [14, 55, 56], they lack robustness to semantic changes to the captions which severely limits their real-world use for alignment applications. We provide a fix by training models on a novel video-centric VideoCon dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Improving Video-Language Robustness. Prior work [37, 38, 51] highlights that the video-text models cannot comprehend the semantics of the text with focus on manipulating the verb, actions, and entities grounded in the video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "description. To improve the temporal understanding, [5] finetunes a pretrained model with temporal order loss. Despite this, their models do not achieve good zero-shot performance on downstream tasks consistently. In our work, we categorize a wide range of plausible misalignments in the contrast captions, and create a temporally-challenging VideoCon dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Video-Language Alignment Evaluation. Many applications such as text-to-video retrieval [15, 48, 57] and text-to-video generation [7, 47] require evaluation of the semantic alignment between the natural language text and raw video. In this work, we indicate that the existing video-text models such as VideoCLIP and ImageBind are not robust to semantic changes in the video captions, which becomes critical for faithful video-text alignment evaluation. In our work, we propose VideoCon and finetune a video-language generative model to perform robust entailment task and provide fine-grained NLE for the observed misalignments between the video and text. In the future, our model can be utilized to enhance alignment through sparse (entailment scores) and dense (fine-grained NLE) feedback [43]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.443, + 0.62, + 0.459 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.574 + ], + "angle": 0, + "content": "We introduced a comprehensive dataset, VideoCon, designed for robust video-text alignment. It features various semantic misalignments and explanations for text-video discrepancies. Through finetuning video-language models on this dataset, we enhanced their performance on complex tasks like text-to-video retrieval and video question answering, achieving state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.895, + 0.773 + ], + "angle": 0, + "content": "One current limitation and an important future direction is to increase the complexity of the generated contrast captions. Specifically, the model may encounter several misalignments within a single contrast caption. Addressing this issue, the model should be equipped to accurately assign low entailment scores to these contrast captions and consequently generate precise NLEs. An important future direction is to scale VideoCon to larger datasets. Here, we create contrast captions for high-quality captions written by humans for every video, however, the web-scale datasets have low-quality captions that are not well grounded in the video. In this regard, using synthetic data followed by VideoCon-like contrast caption generation can be a plausible approach." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.785, + 0.681, + 0.803 + ], + "angle": 0, + "content": "9. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This material is based on research supported by the ECOLE program under Cooperative Agreement HR00112390060 with the US Defense Advanced Research Projects Agency (DARPA). In addition, the research is partly funded by ONR grant N00014-23-1-2780. Hritik Bansal is supported in part by AFOSR MURI grant FA9550-22-1-0380." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13934" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.472, + 0.399 + ], + "angle": 0, + "content": "[4] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[5] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.472, + 0.541 + ], + "angle": 0, + "content": "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.543, + 0.472, + 0.627 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.629, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.701, + 0.472, + 0.77 + ], + "angle": 0, + "content": "[9] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the \"video\" in video-language understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2917-2927, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.772, + 0.472, + 0.842 + ], + "angle": 0, + "content": "[10] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.844, + 0.472, + 0.885 + ], + "angle": 0, + "content": "[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021.8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[13] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datcomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[15] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[16] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Sarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[17] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 1, 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[18] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[19] Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. Q2: Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. arXiv preprint arXiv:2104.08202, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.705, + 0.623 + ], + "angle": 0, + "content": "[20] https://commoncrawl.org/. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.624, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[22] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13935" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[27] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[28] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. arXiv preprint arXiv:2306.14060, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.469, + 0.391 + ], + "angle": 0, + "content": "[30] Weixin Liang, James Zou, and Zhou Yu. Alice: Active learning with contrastive natural language explanations. arXiv preprint arXiv:2009.10259, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.393, + 0.469, + 0.434 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.436, + 0.469, + 0.491 + ], + "angle": 0, + "content": "[32] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.493, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.549, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.621, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[35] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.691, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[36] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[37] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, 2022. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[42] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8322-8332, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[43] Jérémy Scheurer, Jon Ander Campos, Tomasz Korbak, Jun Shern Chan, Angelica Chen, Kyunghyun Cho, and Ethan Perez. Training language models with language feedback at scale. arXiv preprint arXiv:2303.16755, 2023. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[45] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 535-544, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[48] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 8, 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13936" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[50] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems, 35: 8483-8497, 2022. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.176, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[51] Zhenhailong Wang, Ansel Blume, Sha Li, Genglin Liu, Jaemin Cho, Zineng Tang, Mohit Bansal, and Heng Ji. Paxion: Patching action knowledge in video-language foundation models. arXiv preprint arXiv:2305.10683, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[52] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[53] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment task for visually-grounded language learning. arXiv preprint arXiv:1811.10582, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[54] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding. arXiv preprint arXiv:1901.06706, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[56] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023.8.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.47, + 0.595 + ], + "angle": 0, + "content": "[57] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 3, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[58] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[59] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[60] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400, 2023. 3, 6, 7, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[62] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[63] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI Conference on Artificial Intelligence, pages 7590-7598, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13937" + } + ] +] \ No newline at end of file diff --git a/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_origin.pdf b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d8d07c974bacf2a14e8bd3d46d359bb5c98f744 --- /dev/null +++ b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/7aaa3f92-fb68-4a9b-809d-b79e89d5cc93_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ac09f9277d1a713ca590883038fa63ff46b19f10e6215b8f76ff5c28ee435e0 +size 10580059 diff --git a/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/full.md b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..455d2bc8abf75dabfd314c081886403952414ee4 --- /dev/null +++ b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/full.md @@ -0,0 +1,309 @@ +# VideoCon: Robust Video-Language Alignment via Contrast Captions + +Hritik Bansal1 Yonatan Bitton2 Idan Szpektor2* Kai-Wei Chang1* Aditya Grover1* +1UCLA 2Google Research +{hbansal, kwchang, adityag}@cs.ucla.edu +{yonatanbitton, szpektor}@google.com + +# Abstract + +Despite being (pre)trained on a massive amount of data, state-of-the-art video-language alignment models are not robust to semantically-plausible contrastive changes in the video captions. Our work addresses this by identifying a broad spectrum of contrast misalignments, such as replacing entities, actions, and flipping event order, which alignment models should be robust against. To this end, we introduce the VideoCon, a video-language alignment dataset constructed by a large language model that generates plausible contrast video captions and explanations for differences between original and contrast video captions. Then, a generative video-language model is finetuned with VideoCon to assess video-language entailment and generate explanations. Our VideoCon-based alignment model significantly outperforms current models. It exhibits a 12-point increase in AUC for the video-language alignment task on human-generated contrast captions. Finally, our model sets new state of the art zero-shot performance in temporally-extensive video-language tasks such as text-to-video retrieval (SSv2-Temporal) and video question answering (ATP-Hard). Moreover, our model shows superior performance on novel videos and human-crafted captions and explanations. + +# 1. Introduction + +Semantically aligning data points from diverse modalities is a long-standing goal of AI. We focus on video-language alignment, which is challenging due to the complexities involved in understanding of entities, their relationships, and temporal order of the depicted events [17]. Recent models such as VideoCLIP [55], ImageBind [14] learn a shared embedding space. Similarly, generative models such as Flamingo [1], mPLUG-Owl-Video [61] can provide a classification label (e.g., yes/no) when queried about video-language alignment. + +Despite large-scale pretraining, prior work [5, 37, 38, 51] highlights that video-language alignment models are not robust to semantically plausible manipulations to an original aligned caption in the form of contrast captions, such as from 'dog runs away before it eats food' to 'dog runs away after it eats food'. Such pitfalls in robustness questions the trustworthiness of alignment models for large-scale deployment. To mitigate these shortcomings, one possible solution is to scale video-language pairs more for increased diversity during pretraining. However, this is challenging due to the difficulties in sourcing new, high-quality and permissible content, as well as the requirement for substantial storage capacity. Several works [11, 13, 16] have shown that naively training models on web-scale data has diminishing returns on downstream tasks, and emphasize the importance of data quality. Furthermore, the recent studies [28, 62] demonstrate that applying a contrastive objective to the pretraining datasets does not encourage the model to grasp the fine-grained details within image/region-caption data. + +To this end, we take a scalable, active strategy to gather high-quality data that is deliberately enriched with the attributes that we want to instill in alignment models. We create a novel dataset, VideoCon, to improve the robustness of models. Specifically, the dataset consists of a variety of semantically plausible video-language misalignments in contrast captions. These misalignments include altering objects (entities), actions, attributes, relations, counts, event orders, and introducing hallucinations (Figure 2). To construct VideoCon, a large language model (PaLM-2 API) takes video-caption pairs as input and generates high-quality contrast captions for a given misalignment type. To make our dataset temporally-challenging, we skipped "easy" video-caption pairs whose alignment could be inferred based on a single frame (image) understanding [9, 26] ( $\S 3.1$ ). In addition, the LLM generates natural language explanations (NLEs) [42] to the differences between original and altered captions, which are used for further robust training. We performed human verification on a sample of VideoCon and found that it is of high-quality. Finally, to evaluate the model's generalization capabilities, we col + +![](images/5e8fee37352d8ab1f77e746d84d70b9cfa47f99ed1985abda0ba57454e9c5947.jpg) +Figure 1. Overview of our VideoCon approach. First, aligned video-language pairs are filtered to retain temporally-challenging instances. Then contrast captions and natural language explanations (NLE) are generated by an LLM to create the VideoCon dataset. Second, a video-language alignment model is finetuned with VideoCon on the alignment and NLE tasks. Finally, the finetuned model is evaluated against the baseline model. Our results show that it outperforms the baseline, achieving state-of-the-art results on downstream tasks. + +lect human-generated contrast captions and NLEs for the videos sourced from external datasets that did not contribute to VideoCon's development. + +We finetuned a generative video-language model (mPLUG-Owl-Video) on the VideoCon dataset. The trained model surpasses existing video-language alignment models by a large margin on the LLM-generated test set for both video-language alignment and NLE generation tasks. Interestingly, we observed that our finetuned model generalizes to unseen videos and human-generated contrast captions and NLEs, and outperforms the baseline models. For instance, our model's ROC-AUC exceeds the baseline model by 12 points on the human-generated contrast captions. This indicates that our model has developed a better understanding of the entities, their interactions, action understanding, as well as the temporal order of the events for robust video-language alignment. + +We further assessed the effectiveness of robust training via contrast captions on zero-shot downstream videolanguage tasks such text-to-video retrieval and video question answering on the temporally-challenging and action-intensive SSv2-Temporal [45] and SSv2-Events [5]. Our model achieves state-of-the-art (SOTA) performance, improving on SSv2-Temporal by $4.3\mathrm{mAP}$ , SSv2-Events by $3.6\mathrm{mAP}$ points. In addition, our model also achieves SOTA on temporal and causal video question answering in the ATP-Hard dataset, increasing $4\%$ accuracy. This suggests that equipping a model with the knowledge of contrast captions is highly data-efficient and effective in improving its robustness in comparison to scaling the pretraining data. The complete pipeline is illustrated in Figure 1. The dataset and the model will be released upon acceptance. + +# 2. Video Language Alignment + +We are interested in assessing the semantic alignment between the video1 and text data since it powers many prac + +tical applications such as video-text retrieval [57], video generation [7, 47] and video captioning [59]. To this end, [14, 39, 49, 55] designed (image)video-text alignment models that are utilized for evaluating the semantic similarity between the two modalities. However, previous works [5, 37, 38, 51] have questioned their robustness to semantically plausible changes to the video descriptions, termed here contrast captions. Our aim is to improve the robustness of video-text alignment models by training on contrast captions with a wide range of misalignments. + +Consider a dataset $\mathcal{D} = \{(V_i, T_i, C_i, E_i)\}$ where $V_i$ is a video, $T_i$ is an aligned caption, $C_i$ is a contrast caption which is a perturbation of $T_i$ but misaligns with $V_i$ , and $E_i$ is a natural language explanation for the misalignment between $V_i$ and $C_i$ . We consider two video-language alignment tasks: (a) video-language entailment, (b) natural language explanation. + +Video-Language Entailment (VLE) casts video-text alignment as a Visual Entailment (VE) task. VE was originally defined for images as premises and texts as hypothesis [53, 54]. We extend VE definition also for videos as premises, under which a classification model $A_{vle}(V,T)$ predicts whether a video $V$ entails a text $T$ . + +Natural Language Explanation (NLE) requires a model, $A_{nle}(V,C)$ , to generate an open-ended explanation for the discrepancy between a video $V$ and a non-entailing caption $C$ . + +In this paper, we address both VLE and NLE tasks under a multitask setting in which a single video-language generative model generates the binary label for entailment and the open-ended explanation. + +# 3. VideoCon: Contrast Captions Generation for Robust Video-Language Alignment + +Our research goal is to measure the impact of a comprehensive dataset on increasing the robustness of video-text align- + +![](images/f58ee2103ef495f101d966eab6030107743d2821b15cb27558c2319b1be4ad19.jpg) +Figure 2. Overview of the VideoCon data generation process from top to bottom. Specifically, we prompt a large language model (PaLM-2) with the original caption that is grounded in the video, and the intended type of misalignment within the contrast caption. We consider seven kinds of misalignments including object, action, attribute, counting, spatial relation, hallucination, and event order flip. We provide a generated contrast caption and the corresponding natural language explanation for each misalignment type. + +ment models. To this end, we first collect video-caption pairs where the caption cannot be derived from a single frame of video. We then categorize a wide range of semantically plausible manipulations of video captions. Using an LLM for large-scale computation, contrast captions and related explanations are generated for the defined categories, constructing the VideoCon dataset. Finally, we extend VideoCon to include human-created contrast captions as held-out evaluation on unseen videos. We detail the dataset construction steps below. + +# 3.1. Temporally-Challenging Instance Selection + +To construct VideoCon, we start with existing datasets that include natural (real) videos and associated high-quality human-written captions: MSR-VTT [57], VaTeX [48], and TEMPO [17]. MSR-VTT and VaTeX consist of 20 captions and 10 captions per video, respectively, while TEMPO consists of a single caption per video. More dataset details are in Appendix §B. + +TEMPO is designed to create temporally-challenging instances, while MSR-VTT and VaTeX contain more general video-caption pairs. For MSR-VTT and VaTeX, we filter out instances, where the caption is highly associated with a single frame in the video based on an image-text alignment model. In such cases, a video-text alignment can leverage shortcuts and align the video to its caption without understanding the temporal or causal relations depicted in the video. We want to filter such instances. + +To this end, we employ the End-to-End VNLI model [60] to calculate an alignment score $A_{vle}(V,T)$ between a video $V = \{I_1,I_2,\dots ,I_N\}$ and a text $T$ where $I_{i}$ is a frame from + +the video sampled at a rate of 1 frame per second. Formally, + +$$ +A _ {v l e} (V, T) = \max _ {i} \left(V N L I \left(I _ {i}, T\right)\right) \tag {1} +$$ + +where $VNLI(I_i, T)$ is the task of visual natural language inference that assesses the extent to which the text $T$ entails the image $I_i$ . There are 20 and 10 captions per video in the MSR-VTT and VaTeX datasets, respectively. We retain 5 captions per video from these datasets with the lowest $A_{vle}(V, T)$ , and the remaining captions are filtered out. Post-filtering, the percentage of temporally-challenging instances increased from $36.5\%$ to $81.5\%$ in MSR-VTT, and from $42.6\%$ to $71\%$ in VaTeX. + +# 3.2. Categories of Contrast Captions + +We aim for VideoCon to include a wide range of misalignments in its contrast captions. Overall, VideoCon covers seven misalignment types, exemplified in Figure 2. We include replacement of objects (entities) and actions following the analysis in [37, 38], and replacement of attributes, counts, relations, as well as adding unrelated but plausible information to captions as hallucinations following [29, 32, 35]'s study of image/text alignment model brittleness. Since most video-text models rely on pretrained image backbones, they are likely to suffer from similar problems. Finally, following [5]'s analysis that video-text models do not understand temporal order of the events, we include event order flipping as misalignment type. + +# 3.3. Data Generation using an LLM + +To generate contrast captions and corresponding NLE we first assign one of the seven misalignment types (§3.2) to + +![](images/645e16ec5010339726cb400434e3a471e9dc3a9adb1256106315c49df13933d4.jpg) +Figure 3. Distribution of the types of misalignments within the contrast captions of the VideoCon dataset. We observe that the dataset has good representation for all the kinds of misalignments ranging from $8.8\%$ to $24.2\%$ . + +each caption in the input video-text datasets (§3.1) (details in Appendix §C). Then, given a video $V$ and a misalignment type $m$ , we prompt PaLM-2 API² [2] to generate a contrast caption and accompanied explanation (our type-specific prompts are detailed in Appendix §D). + +Analyzing the LLM generations, we found that sometimes the output caption $C$ do not contradict the original caption $T$ . For example, a generated contrast caption "a person riding a car" does not contradict the original caption "a person riding a mustang". To filter such cases, we employ a Natural Language Inference (NLI) model [19] and remove cases in which the contrast caption is assessed as entailed by the original caption $NLI(T,C) > 0.5$ . Postfiltering, each tuple $(V,T,C,m)$ is converted to the two instances of video/language entailment task: $A_{en}(V,T) = 1$ and $A_{en}(V,C) = 0$ . We present the dataset statistics for the entailment task in Table 1, including train/eval/test splits. In addition, Fig. 3 shows the distribution of misalignment types in the dataset. We observe that VideoCon maintains a high density across the 7 misalignments ranging from $8.8\%$ to $24.2\%$ . + +We also found that some generated explanations do not describe the differences between $T$ and $C$ well. For example, the explanation "two friends are not traveling together" does not fully describe the discrepancy between "three friends traveling together" and "two friends are traveling together". To filter these out, generated examples are removed if $NLI(F(T,C),E) < 0.6$ where $F(T,C)$ is the premise comprising the original and contrast captions. Specifically, premise will be 'Expected Caption: $T$ Actual Caption: $E'$ and hypothesis will be 'Difference between Expected and Actual Caption: $E'$ . This filter indicates that the information in the explanation is not entailed by the difference between the two captions. The dataset statistics for the NLE task is presented in Table 1. We refer to the final + +
Video-Language EntailmentNatural Language Explanation
SourceTrainValTestTrainValTest
MSR-VTT3836647816538158882066788
VaTeX664807368110301803453636
TEMPO1071270982708416527391073
Total11555883122735650233329011497
+ +Table 1. Statistics for the VLE and NLE tasks in VideoCon. + +LLM-generated dataset as VideoCon (LLM). + +To assess the quality of VideoCon (LLM), we perform human evaluation on 500 contrast captions and NLEs (details in Appendix E). The human evaluator found $91\%$ of the contrast captions and $89\%$ of the NLEs to be valid, indicating the high-quality of VideoCon (LLM). + +# 3.4. Data Generation using Humans + +To study whether a model trained on VideoCon (LLM) generalizes to out-of-distribution videos and its performance on human-generated contrast captions, we randomly selected a set of videos from the validation set of ActivityNet [10]. This dataset consists of captions matched with segments in the video, e.g., "a little boy is climbing on an outside gym" matched to the first 10 seconds of its related video. We extracted video segments with an associated caption. Human workers3 on Amazon MTurk were then shown the video segments and associated captions and were asked to create a semantically plausible contrast caption and a corresponding NLE (more details in Appendix §F). We did not communicate any type of target misalignments to encourage natural diversity of human created contrast captions. + +Overall, we collected 570 tuples $(V, T, C_{human}, E_{human})$ where $V$ is the video, $T$ is the original caption, $C_{human}$ is the human-written contrast caption, and $E_{human}$ is the human-written explanations. We denote this dataset by VideoCon (Human). We sample 100 instances from this dataset, and found $93\%$ to be clean. In addition, we observe that many of the human-generated contrast captions perturbing one or more objects ( $35\%$ ) and actions ( $35\%$ ) depicted in the caption. While $8\% - 10\%$ of the contrast captions flip the order of the events and attribute of the objects. As this dataset is largely unfiltered, it contains a mix of temporally-easy and challenging instances. We also constructed a more temporally-challenging subset of 290 instances, denoted VideoCon (Human-Hard), by filtering out tuples in which $A_{vle}(V, T) < 0.5$ (Eq. (1)), as in §3.1. + +# 4. Experimental Setup + +We next describe our evaluation setting for measuring the impact of VideoCon on video-text alignment modeling. + +# 4.1. Finetuning with VideoCon + +Our goal in constructing VideoCon (LLM) is to improve robustness of video-text alignment models by fine-tuning on this dataset. To this end, we start with the mPLUG-OwlVideo model [61], denoted Owl-Base. Its building blocks are CLIP [39] as visual encoder and LLaMA-7B [46] as text encoder/decoder and it was pretrained on VideoChat [27]. + +# Entailment Task: + +Given: V (Video), T (Caption), C (Contrast Caption) + +Instruction (I): [V] Does this video entail the description $[T]$ ? Response (R): Yes + +Instruction (I): [V] Does this video entail the description [C]? Response (R): No + +Figure 4. Entailment task prompt for finetuning. + +# Natural Language Explanation Generation Task: + +Given: V (Video), C (Contrast Caption), E (NLE) + +Instruction (I): [V] What is the misalignment between this video and the description $[C]$ ? Response (R): [E] + +Figure 5. NLE generation task prompt for finetuning. + +To fine-tune $Owl$ -Base on VideoCon (LLM), its $\{V, T, C, E\}^4$ tuples were converted into two types of multimodal instruction-response pairs, one for the VLE task $(I_{vle}, R)$ (Fig. 4) and one for the NLE task $(I_{nle}, R)$ (Fig. 5). We then train $Owl$ -Base on all instruction pairs from both the tasks with maximum likelihood loss, resulting in a single model $Owl$ -Con. + +# 4.2. VideoCon Evaluation Metrics + +To evaluate the performance of the $Owl-Con$ on video-text alignment we generate $Owl-Con$ response to prompt $I_{vle}$ for video $V$ and text $Y \in \{T, C\}$ . We then calculate the probability of generating responses $s_y = Owl-Con$ ( $^\prime \mathrm{Yes}^\prime | I_{vle}(V, Y)$ ) and $s_n = Owl-Con$ ( $^\prime \mathrm{No}^\prime | I_{vle}(V, Y)$ ), and based on these scores the probability for class 'Yes': $P_{yes}(V, Y) = \frac{s_y}{s_y + s_n}$ . Finally, we compute the ROC-AUC score for $P_{yes}(V, Y)$ over the VideoCon (LLM) eval set, with $\{V, T\}$ as label 1 and $\{V, C\}$ as label 0. + +To evaluate $Owl$ -Con on the NLE task, we prompt it with instruction $I_{nle}$ instantiated on $\{V,C\}$ pairs from the VideoCon (LLM) eval set. We compare the generated explanation $\hat{E}$ to the ground truth $E$ by measuring entailment + +probability $NLI(E, \hat{E})$ . In our experiments, we experiment with two $NLI$ automatic metrics: (a) $Q^2$ score [19], and (b) PaLM-2 API. We performed human evaluation to measure the agreement between the automatic metrics and the human-rating. We found that both metrics achieve high agreement with human assessment (Appendix §H). + +# 4.3. Video-Text Downstream Tasks + +We complement the VideoCon intrinsic evaluation over the testset with an extrinsic evaluation over two temporal and action difficult downstream tasks. + +We evaluate alignment model performance for text2video retrieval over SSv2-Temporal [45] and SSv2-Events [5] datasets. We consider the SSv2-Template captions instead of the label captions since they remove the object-centric bias in model evaluation [26]. We compute input-text/candidate-video alignment score, rank videos and report mean Average Precision (mAP). We evaluate alignment model performance for video question answering over the ATP-Hard [9] dataset. We cast each question/candidate-answer pair as an imperative statement using PaLM-2 API, measure alignment to the input video and report Accuracy. More details on the downstream datasets and the evaluation setup are in Appendix §I. + +# 4.4. Baselines + +For the video-text alignment text, we compare Owl-Con with the following baselines: (a) End-to-End VNLI as zero-shot atemporal model since it does not have access to the temporal order of the video frames, (b) VideoCLIP [55], (c) ImageBind [14], (d) Owl-Base, and (e) Owl-Rand: Owl-Base fine-tuned on VideoCon tuples $\{V,T,\hat{C},E\}$ where $\hat{C}$ is randomly selected from other captions in the dataset. Owl-Rand would indicate if there is merit in the contrast, hard-negative captions in VideoCon. We include additional baselines TACT [5] and VFC [37] for evaluating on the downstream tasks (\$5.3). + +# 5. Experiments + +We present our intrinsic (VideoCon eval set) and extrinsic (downstream tasks) evaluation results, showing the benefits of VideoCon for robust video-language alignment. + +# 5.1. Performance on VideoCon Entailment Task + +We present the ROC-AUC scores of the tested models in Table 2. From the table we see that the baseline models find the VideoCon testset difficult, as reflected by low AUC scores (e.g. Owl-Base-57.2), close to random. Even training on VideoCon train instances, but with "easy" negatives (Owl-Rand-59.7), hardly improves the base models. A significant improvement is achieved with the VNLI-specific model (67), showing that the entailment task is not inherently represented in generic video-language aligned training + +
ModelsVideoCon (LLM) TestVideoCon (Human)VideoCon (Human-Hard)
Random50.050.050.0
VideoCLIP [55]53.247.347.5
ImageBind (Video-Text) [14]57.165.263.0
Owl-Base [61]57.266.864.1
Owl-Rand59.768.965.5
End-to-End VNLI [60]67.072.465.0
Owl-Con (Ours)84.678.374.4
+ +Table 2. ROC-AUC scores of the tested models for the entailment task on VideoCon test sets. + +
VideoCon (LLM)VideoCon (Human)
ModelsQ2entailmentPaLM-2 entailment acc. (%)Q2entailmentPaLM-2 entailment acc.(%)
Owl-Base0.1936.80.2339.6
Owl-Con (Ours)0.5065.40.3247.1
+ +sets and requires specific training. Yet, the best performance is achieved by training on VideoCon, which addresses the diversity in plausible misalignments and includes "difficult" training examples, reaching 84.6 AUC. This demonstrates the merit of VideoCon for improving video-language alignment robustness. We show qualitative examples for the model predictions in §6.2. + +When evaluating on out-of-domain (OOD) data around video types and misalignment distribution, we again see that training with VideoCon offers significant improvement to alignment detection, outperforming all baselines, albeit with smaller relative gains: $17\%$ and $16\%$ improvement compared to Owl-Base on (Human) and (Human-Hard) respectively compared to $48\%$ on (LLM) test. In future work, we plan to further diversify the misalignments VideoCon covers to further improve its benefits on OOD cases. + +We notice that the performance of the VNLI atemporal model is better than existing video-language alignment models. It might be attributed to its training with contrast captions in [60]. It further highlights that the existing videolanguage models are not robust in comparison to a temporal probe on video-language alignment evaluation, corroborating the findings from [9, 26]. + +# 5.2. Performance on NLE Generation Task + +Table 3 presents the performance of the tested models against the ground-truth on the NLE task, depicting average $Q^2$ score and PaLM-2 entailment accuracy. The results show that on in-domain VideoCon, Owl-Con outperforms Owl-Base by an impressive $263\%$ and $178\%$ relative increase on $Q^2$ score and PaLM-2 accuracy respectively. This indicates the finetuned model can accurately generate NLE that match well with the ground-truth NLE. This indicates that our model can generate accurate NLE for a wide range of misalignments in the video captions, which makes it use + +ful for dense video-language alignment evaluation. + +On out-of-domain VideoCon, the improvement is more moderate but still high: $40\%$ and $20\%$ relative increase on $Q^2$ and PaLM-2 respectively. This is probably due to the more diverse ways humans express explanations compared to LLM prompting. In future work we plan to further address linguistic diversity in explanations for more robust generation and evaluation. + +# 5.3. Performance on Video-Text Downstream Tasks + +Table 3. Performance of the tested models on the NLE generation task, measured via entailment metrics. + +
ModelsSSv2-Temporal +mAPSSv2-Events +mAP
Random7.33.3
VideoCLIP9.86.4
ImageBind (video-language)10.55.5
Owl-Base10.96.8
TACT [5]-7.8
Owl-Rand12.19.9
End-to-End VNLI [60]14.610.4
Owl-Con (Ours)15.211.4
+ +Table 4. Mean Average Precision (mAP) scores for the tested models in the zero-shot text-to-video retrieval tasks. + +We next present our results on the two downstream tasks, Text2Video Retrieval and Video Question Answering. Starting with the retrieval task, we report mean Average Precision (mAP) of the tested models on the SSv2-Temporal and SSv2-Events datasets in Table 4. The benefits of training with additional examples tailored for temporal video-language alignment is already evident in the performance of Owl-Rand, which improves over the previous SSv2-Events SOTA - TACT with a relative increase of $27\%$ . + +However, when training on harder negative contrastive instances, $Owl-Con$ achieves a significant improvement, + +
SuccessFailure
VideoCon LLMCaption: The adult gets the child's attention after tapping his shoulderECaption: Two people move a large object down a narrow stairwayE
Contrast Caption: The adult gets the child's attention before tappingCContrast Caption: Two people move a large object up a narrow stairwayE
GT NLE: First, an adult hand is seen tapping the child's shoulder, then the adult gets the child's attention, not the other way aroundE GTGT NLE: Men move a large piece of furniture down a narrow staircase, not up a narrow staircaseC GT
Predicted NLE: First, the adult taps the child's shoulder, then the adult gets the child's attention, not the other way aroundPredicted NLE: Three men are moving a large piece of furniture, not two
VideoCon HumanCaption: We see the group making cookiesECaption: A lady puts her hair in a bun on the sideE
Contrast Caption: We see the group eating cookiesCContrast Caption: A lady puts her hair in a braid on the sideE
GT NLE: The group is making cookies, not eating themE GTGT NLE: The woman puts her hair in a bun, not a braidC GT
Predicted NLE: We see the group cooking cookies, not eating themPredicted NLE: The lady puts her hair in a braid on top of her head, not on the side
+ +outperforming all baselines, with a relative increase over the best baseline End-to-End VNLI model by $7.5\%$ on SSv2-Temporal and $9.6\%$ on SSv2-Events (46% over TACT), setting new SOTA results. This points at the benefits of exposing the model to temporal examples, such as action and event-order. + +
ModelsAccuracy (%)
CLIP23.8
VideoCLIP23.4
ImageBind (video-language)25.4
TACT [5]27.6
VFC [37]31.4
Owl-Base37.1
Owl-Rand37.2
End-to-End VNLI [60]39.0
Owl-Con (Ours)41.1
+ +Table 5. Accuracy scores for the tested models on the zero-shot video question-answering task on ATP-Hard dataset. + +For the Video Question Answering task, we compare the performance of the various models in Table 5. Here too Owl-Con achieves SOTA results and outperforms the strongest baseline End-to-End VNLI model with a relative increase of $5.1\%$ . This corroborates the observations in our other experiments, which demonstrate the advantage of the VideoCon datasets, covering various misalignments, especially those pertaining to temporal and causal reasoning over dynamic events. The results also confirm the need for carefully chosen contrastive negative examples, showing that picking negatives at random may mask out the potential benefit of an alignment training set. Finally, the competitive performance of atemporal End-to-End VNLI model on the downstream tasks is surprising and underscores the need for stronger video-language datasets for robust benchmarking. + +![](images/ad723f65f8a9060a0ae314f8fac1fbbcd78fa7b8eb947d4dba5d3979b7235912.jpg) +Figure 6. Qualitative examples for the success (green) and failure (red) modes of our model. In every example, we present a few video frames in an temporal order from top to bottom, its associated caption, contrast caption, ground-truth NLE from the datasets. Additionally, we present the predicted NLE from our model. The small boxes at the end of caption cells indicate whether our model considers that caption to be grounded in the video. E and C indicates that the model predicts the caption to entail and contradict to the video, respectively. E-GT and C-GT indicates the predicted NLE entails and contradicts the ground-truth (GT) NLE, respectively. +Figure 7. ROC-AUC of End-to-End VNLI, Owl-Base, and OwlCon across all types of misalignment in VideoCon (LLM) test set. + +# 6. Analysis + +We analyze Owl-Con's performance improvements across the kinds of misalignments in VideoCon. Additionally, we present a few qualitative examples to highlight the success and failure modes of our model. + +# 6.1. Per-misalignment Entailment Results + +We compared the ROC-AUC scores of the atemporal End-to-End VNLI, Owl-Base, and Owl-Con on specific misalignments in the contrast captions from VideoCon (LLM) testset in Figure 7. We observed that Owl-Con outperforms the baseline models across all misalignment types. This suggests that our model can reason well about the entities, their relations, and the temporal order of events in the video. + +The largest improvement of $\text{Owl-Con}$ compared to the two baselines is on event order flip, indicating that the baselines lack temporal understanding and the VideoCon is efficient in adding this capability to an alignment model. In + +addition, on hallucination both Owl-Con and End-to-End VNLI significantly outperform Owl-Base, since both models were explicitly exposed to entailment/non-entailment training data. It is surprising to see that while End-to-End VNLI was trained on significantly more entailment data, much of it human-curated, Owl-Con outperforms it with only automatically generated data. This could be due to the better encoding of video in Owl-Con compared to the atemporal nature of End-to-End VNLI. Finally, the analysis shows other types of atemporal misalignments that are difficult for End-to-End VNLI to sort out, e.g. counting' and relation, where the training data in VideoCon is useful to improve these capabilities as well. This shows that our approach of detailed analysis of misalignment types of generation of examples for them is effective. + +# 6.2. Qualitative Examples + +We highlight a few classification examples of Owl-Con in Figure 6. The rows refer to the test source of the instances and the columns refer to the success and failure modes, respectively. In Row1/Column1, we observe that our model provides correct predictions for the entailment between the video and original caption while predicting contradiction for the contrast caption that flips the order of the events i.e., grabbing attention and tapping shoulders. Interestingly, our model can also provide the accurate NLE when prompted with the video and the contrast caption. This suggests that our model is useful for providing fine-grained details about the video-language alignment. In Row2/Column2, the model confuses 'buns' with 'braids' in hair and gives a wrong NLE that contradicts the ground-truth. This error, due to its inability to distinguish between objects, might be improved with diverse videos and captions. + +# 7. Related Work + +Foundation Models for Video-Language Understanding. Foundation models have emerged for video-language understanding [1, 4, 49, 55, 56] by pre-training on large amount of video-text pairs scraped from the web [6, 36, 58]. Additionally, prior works have either leveraged the pretrained CLIP model for video-language tasks [12, 33, 34] or adopted a socratic approach [50, 63] to employ LLMs (GPT-3) in reasoning over video captions. We highlight that despite the large-scale training of the video-language foundation models [14, 55, 56], they lack robustness to semantic changes to the captions which severely limits their real-world use for alignment applications. We provide a fix by training models on a novel video-centric VideoCon dataset. + +Improving Video-Language Robustness. Prior work [37, 38, 51] highlights that the video-text models cannot comprehend the semantics of the text with focus on manipulating the verb, actions, and entities grounded in the video + +description. To improve the temporal understanding, [5] finetunes a pretrained model with temporal order loss. Despite this, their models do not achieve good zero-shot performance on downstream tasks consistently. In our work, we categorize a wide range of plausible misalignments in the contrast captions, and create a temporally-challenging VideoCon dataset. + +Video-Language Alignment Evaluation. Many applications such as text-to-video retrieval [15, 48, 57] and text-to-video generation [7, 47] require evaluation of the semantic alignment between the natural language text and raw video. In this work, we indicate that the existing video-text models such as VideoCLIP and ImageBind are not robust to semantic changes in the video captions, which becomes critical for faithful video-text alignment evaluation. In our work, we propose VideoCon and finetune a video-language generative model to perform robust entailment task and provide fine-grained NLE for the observed misalignments between the video and text. In the future, our model can be utilized to enhance alignment through sparse (entailment scores) and dense (fine-grained NLE) feedback [43]. + +# 8. Conclusion + +We introduced a comprehensive dataset, VideoCon, designed for robust video-text alignment. It features various semantic misalignments and explanations for text-video discrepancies. Through finetuning video-language models on this dataset, we enhanced their performance on complex tasks like text-to-video retrieval and video question answering, achieving state-of-the-art results. + +One current limitation and an important future direction is to increase the complexity of the generated contrast captions. Specifically, the model may encounter several misalignments within a single contrast caption. Addressing this issue, the model should be equipped to accurately assign low entailment scores to these contrast captions and consequently generate precise NLEs. An important future direction is to scale VideoCon to larger datasets. Here, we create contrast captions for high-quality captions written by humans for every video, however, the web-scale datasets have low-quality captions that are not well grounded in the video. In this regard, using synthetic data followed by VideoCon-like contrast caption generation can be a plausible approach. + +# 9. Acknowledgement + +This material is based on research supported by the ECOLE program under Cooperative Agreement HR00112390060 with the US Defense Advanced Research Projects Agency (DARPA). In addition, the research is partly funded by ONR grant N00014-23-1-2780. Hritik Bansal is supported in part by AFOSR MURI grant FA9550-22-1-0380. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 8 +[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 4 +[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2 +[4] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 8, 1 +[5] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 1, 2, 3, 5, 6, 7, 8 +[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 8, 1 +[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 8, 1 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[9] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the "video" in video-language understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2917-2927, 2022. 1, 5, 6 +[10] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 3 +[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 1 +[12] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. + +Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021.8, 1 +[13] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datcomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1 +[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1, 2, 5, 6, 8 +[15] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 8, 1, 5 +[16] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Sarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 1 +[17] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 1, 3, 2 +[18] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 2 +[19] Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. Q2: Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. arXiv preprint arXiv:2104.08202, 2021. 4, 5 +[20] https://commoncrawl.org/. 1 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[22] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897, 2023. 1 +[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1 +[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1 + +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3 +[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 5, 6 +[27] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5 +[28] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. arXiv preprint arXiv:2306.14060, 2023. 1 +[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3 +[30] Weixin Liang, James Zou, and Zhou Yu. Alice: Active learning with contrastive natural language explanations. arXiv preprint arXiv:2009.10259, 2020. 1 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6 +[32] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3 +[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 8, 1 +[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 8, 1 +[35] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 3 +[36] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 8, 1 +[37] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 1, 2, 3, 5, 7, 8 +[38] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of + +the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, 2022. 1, 2, 3, 8 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5 +[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[42] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8322-8332, 2022. 1 +[43] Jérémy Scheurer, Jon Ander Campos, Tomasz Korbak, Jun Shern Chan, Angelica Chen, Kyunghyun Cho, and Ethan Perez. Training language models with language feedback at scale. arXiv preprint arXiv:2303.16755, 2023. 8, 1 +[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1 +[45] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 535-544, 2021. 2, 5 +[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5 +[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 8, 1 +[48] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 8, 1 +[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 8, 1 + +[50] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems, 35: 8483-8497, 2022. 8, 1 +[51] Zhenhailong Wang, Ansel Blume, Sha Li, Genglin Liu, Jaemin Cho, Zineng Tang, Mohit Bansal, and Heng Ji. Paxion: Patching action knowledge in video-language foundation models. arXiv preprint arXiv:2305.10683, 2023. 1, 2, 8 +[52] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5 +[53] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment task for visually-grounded language learning. arXiv preprint arXiv:1811.10582, 2018. 2 +[54] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding. arXiv preprint arXiv:1901.06706, 2019. 2 +[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 5, 6, 8 +[56] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023.8.1 +[57] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 3, 8, 1 +[58] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8, 1 +[59] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2 +[60] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400, 2023. 3, 6, 7, 1, 2 +[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 5, 6 + +[62] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1 +[63] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 8, 1 +[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI Conference on Artificial Intelligence, pages 7590-7598, 2018. 7 \ No newline at end of file diff --git a/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/images.zip b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45fca7176712d07a72e55033f6b913a92bd0a602 --- /dev/null +++ b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24fb831248fa1ae29b49926ec4d4e4d4ef46a7fabdbb80dfa01d8a60a37a0be1 +size 434789 diff --git a/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/layout.json b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e258f4ada26271cbe28dc50fe0e9ff3be7a48c5e --- /dev/null +++ b/2024/VideoCon_ Robust Video-Language Alignment via Contrast Captions/layout.json @@ -0,0 +1,9204 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "VideoCon: Robust Video-Language Alignment via Contrast Captions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 517, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 517, + 200 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 517, + 200 + ], + "type": "text", + "content": "Hritik Bansal1 Yonatan Bitton2 Idan Szpektor2* Kai-Wei Chang1* Aditya Grover1* \n1UCLA 2Google Research \n{hbansal, kwchang, adityag}@cs.ucla.edu \n{yonatanbitton, szpektor}@google.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "type": "text", + "content": "Despite being (pre)trained on a massive amount of data, state-of-the-art video-language alignment models are not robust to semantically-plausible contrastive changes in the video captions. Our work addresses this by identifying a broad spectrum of contrast misalignments, such as replacing entities, actions, and flipping event order, which alignment models should be robust against. To this end, we introduce the VideoCon, a video-language alignment dataset constructed by a large language model that generates plausible contrast video captions and explanations for differences between original and contrast video captions. Then, a generative video-language model is finetuned with VideoCon to assess video-language entailment and generate explanations. Our VideoCon-based alignment model significantly outperforms current models. It exhibits a 12-point increase in AUC for the video-language alignment task on human-generated contrast captions. Finally, our model sets new state of the art zero-shot performance in temporally-extensive video-language tasks such as text-to-video retrieval (SSv2-Temporal) and video question answering (ATP-Hard). Moreover, our model shows superior performance on novel videos and human-crafted captions and explanations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "content": "Semantically aligning data points from diverse modalities is a long-standing goal of AI. We focus on video-language alignment, which is challenging due to the complexities involved in understanding of entities, their relationships, and temporal order of the depicted events [17]. Recent models such as VideoCLIP [55], ImageBind [14] learn a shared embedding space. Similarly, generative models such as Flamingo [1], mPLUG-Owl-Video [61] can provide a classification label (e.g., yes/no) when queried about video-language alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 456 + ], + "type": "text", + "content": "Despite large-scale pretraining, prior work [5, 37, 38, 51] highlights that video-language alignment models are not robust to semantically plausible manipulations to an original aligned caption in the form of contrast captions, such as from 'dog runs away before it eats food' to 'dog runs away after it eats food'. Such pitfalls in robustness questions the trustworthiness of alignment models for large-scale deployment. To mitigate these shortcomings, one possible solution is to scale video-language pairs more for increased diversity during pretraining. However, this is challenging due to the difficulties in sourcing new, high-quality and permissible content, as well as the requirement for substantial storage capacity. Several works [11, 13, 16] have shown that naively training models on web-scale data has diminishing returns on downstream tasks, and emphasize the importance of data quality. Furthermore, the recent studies [28, 62] demonstrate that applying a contrastive objective to the pretraining datasets does not encourage the model to grasp the fine-grained details within image/region-caption data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "To this end, we take a scalable, active strategy to gather high-quality data that is deliberately enriched with the attributes that we want to instill in alignment models. We create a novel dataset, VideoCon, to improve the robustness of models. Specifically, the dataset consists of a variety of semantically plausible video-language misalignments in contrast captions. These misalignments include altering objects (entities), actions, attributes, relations, counts, event orders, and introducing hallucinations (Figure 2). To construct VideoCon, a large language model (PaLM-2 API) takes video-caption pairs as input and generates high-quality contrast captions for a given misalignment type. To make our dataset temporally-challenging, we skipped \"easy\" video-caption pairs whose alignment could be inferred based on a single frame (image) understanding [9, 26] (" + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\S 3.1" + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "). In addition, the LLM generates natural language explanations (NLEs) [42] to the differences between original and altered captions, which are used for further robust training. We performed human verification on a sample of VideoCon and found that it is of high-quality. Finally, to evaluate the model's generalization capabilities, we col" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 117, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 117, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 117, + 713 + ], + "type": "text", + "content": "*Equal Advising." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13927" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 541, + 159 + ], + "type": "image", + "image_path": "5e8fee37352d8ab1f77e746d84d70b9cfa47f99ed1985abda0ba57454e9c5947.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 222 + ], + "type": "text", + "content": "Figure 1. Overview of our VideoCon approach. First, aligned video-language pairs are filtered to retain temporally-challenging instances. Then contrast captions and natural language explanations (NLE) are generated by an LLM to create the VideoCon dataset. Second, a video-language alignment model is finetuned with VideoCon on the alignment and NLE tasks. Finally, the finetuned model is evaluated against the baseline model. Our results show that it outperforms the baseline, achieving state-of-the-art results on downstream tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 244, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 287, + 280 + ], + "type": "text", + "content": "lect human-generated contrast captions and NLEs for the videos sourced from external datasets that did not contribute to VideoCon's development." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 448 + ], + "type": "text", + "content": "We finetuned a generative video-language model (mPLUG-Owl-Video) on the VideoCon dataset. The trained model surpasses existing video-language alignment models by a large margin on the LLM-generated test set for both video-language alignment and NLE generation tasks. Interestingly, we observed that our finetuned model generalizes to unseen videos and human-generated contrast captions and NLEs, and outperforms the baseline models. For instance, our model's ROC-AUC exceeds the baseline model by 12 points on the human-generated contrast captions. This indicates that our model has developed a better understanding of the entities, their interactions, action understanding, as well as the temporal order of the events for robust video-language alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": "We further assessed the effectiveness of robust training via contrast captions on zero-shot downstream videolanguage tasks such text-to-video retrieval and video question answering on the temporally-challenging and action-intensive SSv2-Temporal [45] and SSv2-Events [5]. Our model achieves state-of-the-art (SOTA) performance, improving on SSv2-Temporal by " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "4.3\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": ", SSv2-Events by " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "3.6\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " points. In addition, our model also achieves SOTA on temporal and causal video question answering in the ATP-Hard dataset, increasing " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " accuracy. This suggests that equipping a model with the knowledge of contrast captions is highly data-efficient and effective in improving its robustness in comparison to scaling the pretraining data. The complete pipeline is illustrated in Figure 1. The dataset and the model will be released upon acceptance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 639, + 203, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 203, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 203, + 654 + ], + "type": "text", + "content": "2. Video Language Alignment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "text", + "content": "We are interested in assessing the semantic alignment between the video1 and text data since it powers many prac" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 244, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 363 + ], + "type": "text", + "content": "tical applications such as video-text retrieval [57], video generation [7, 47] and video captioning [59]. To this end, [14, 39, 49, 55] designed (image)video-text alignment models that are utilized for evaluating the semantic similarity between the two modalities. However, previous works [5, 37, 38, 51] have questioned their robustness to semantically plausible changes to the video descriptions, termed here contrast captions. Our aim is to improve the robustness of video-text alignment models by training on contrast captions with a wide range of misalignments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": "Consider a dataset " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(V_i, T_i, C_i, E_i)\\}" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " is a video, " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " is an aligned caption, " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " is a contrast caption which is a perturbation of " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " but misaligns with " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " is a natural language explanation for the misalignment between " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": ". We consider two video-language alignment tasks: (a) video-language entailment, (b) natural language explanation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": "Video-Language Entailment (VLE) casts video-text alignment as a Visual Entailment (VE) task. VE was originally defined for images as premises and texts as hypothesis [53, 54]. We extend VE definition also for videos as premises, under which a classification model " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "A_{vle}(V,T)" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": " predicts whether a video " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": " entails a text " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "content": "Natural Language Explanation (NLE) requires a model, " + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "inline_equation", + "content": "A_{nle}(V,C)" + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "content": ", to generate an open-ended explanation for the discrepancy between a video " + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "content": " and a non-entailing caption " + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 596, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 545, + 644 + ], + "type": "text", + "content": "In this paper, we address both VLE and NLE tasks under a multitask setting in which a single video-language generative model generates the binary label for entailment and the open-ended explanation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 654, + 545, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 683 + ], + "type": "text", + "content": "3. VideoCon: Contrast Captions Generation for Robust Video-Language Alignment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our research goal is to measure the impact of a comprehensive dataset on increasing the robustness of video-text align-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "1Like prior works [33, 55], we use only the video frames (the visual channel) without the soundtrack (the audio channel)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13928" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 488, + 255 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 488, + 255 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 488, + 255 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 488, + 255 + ], + "type": "image", + "image_path": "f58ee2103ef495f101d966eab6030107743d2821b15cb27558c2319b1be4ad19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "type": "text", + "content": "Figure 2. Overview of the VideoCon data generation process from top to bottom. Specifically, we prompt a large language model (PaLM-2) with the original caption that is grounded in the video, and the intended type of misalignment within the contrast caption. We consider seven kinds of misalignments including object, action, attribute, counting, spatial relation, hallucination, and event order flip. We provide a generated contrast caption and the corresponding natural language explanation for each misalignment type." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 448 + ], + "type": "text", + "content": "ment models. To this end, we first collect video-caption pairs where the caption cannot be derived from a single frame of video. We then categorize a wide range of semantically plausible manipulations of video captions. Using an LLM for large-scale computation, contrast captions and related explanations are generated for the defined categories, constructing the VideoCon dataset. Finally, we extend VideoCon to include human-created contrast captions as held-out evaluation on unseen videos. We detail the dataset construction steps below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 461, + 271, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 271, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 271, + 475 + ], + "type": "text", + "content": "3.1. Temporally-Challenging Instance Selection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "type": "text", + "content": "To construct VideoCon, we start with existing datasets that include natural (real) videos and associated high-quality human-written captions: MSR-VTT [57], VaTeX [48], and TEMPO [17]. MSR-VTT and VaTeX consist of 20 captions and 10 captions per video, respectively, while TEMPO consists of a single caption per video. More dataset details are in Appendix §B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 674 + ], + "type": "text", + "content": "TEMPO is designed to create temporally-challenging instances, while MSR-VTT and VaTeX contain more general video-caption pairs. For MSR-VTT and VaTeX, we filter out instances, where the caption is highly associated with a single frame in the video based on an image-text alignment model. In such cases, a video-text alignment can leverage shortcuts and align the video to its caption without understanding the temporal or causal relations depicted in the video. We want to filter such instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "To this end, we employ the End-to-End VNLI model [60] to calculate an alignment score " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "A_{vle}(V,T)" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " between a video " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "V = \\{I_1,I_2,\\dots ,I_N\\}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " and a text " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " is a frame from" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 327, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 340 + ], + "type": "text", + "content": "the video sampled at a rate of 1 frame per second. Formally," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 353, + 350, + 545, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 350, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 353, + 350, + 545, + 364 + ], + "type": "interline_equation", + "content": "A _ {v l e} (V, T) = \\max _ {i} \\left(V N L I \\left(I _ {i}, T\\right)\\right) \\tag {1}", + "image_path": "e6f833ad75d089be665bd71a80d596334c151aa10298487b07828fe63d2552b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "VNLI(I_i, T)" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": " is the task of visual natural language inference that assesses the extent to which the text " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": " entails the image " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": ". There are 20 and 10 captions per video in the MSR-VTT and VaTeX datasets, respectively. We retain 5 captions per video from these datasets with the lowest " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "A_{vle}(V, T)" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": ", and the remaining captions are filtered out. Post-filtering, the percentage of temporally-challenging instances increased from " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "36.5\\%" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": " in MSR-VTT, and from " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "42.6\\%" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 304, + 372, + 546, + 480 + ], + "type": "text", + "content": " in VaTeX." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 488, + 480, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 480, + 502 + ], + "type": "text", + "content": "3.2. Categories of Contrast Captions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 507, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 663 + ], + "type": "text", + "content": "We aim for VideoCon to include a wide range of misalignments in its contrast captions. Overall, VideoCon covers seven misalignment types, exemplified in Figure 2. We include replacement of objects (entities) and actions following the analysis in [37, 38], and replacement of attributes, counts, relations, as well as adding unrelated but plausible information to captions as hallucinations following [29, 32, 35]'s study of image/text alignment model brittleness. Since most video-text models rely on pretrained image backbones, they are likely to suffer from similar problems. Finally, following [5]'s analysis that video-text models do not understand temporal order of the events, we include event order flipping as misalignment type." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 477, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 477, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 477, + 684 + ], + "type": "text", + "content": "3.3. Data Generation using an LLM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "To generate contrast captions and corresponding NLE we first assign one of the seven misalignment types (§3.2) to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13929" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 82, + 243, + 187 + ], + "blocks": [ + { + "bbox": [ + 91, + 82, + 243, + 187 + ], + "lines": [ + { + "bbox": [ + 91, + 82, + 243, + 187 + ], + "spans": [ + { + "bbox": [ + 91, + 82, + 243, + 187 + ], + "type": "image", + "image_path": "645e16ec5010339726cb400434e3a471e9dc3a9adb1256106315c49df13933d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "type": "text", + "content": "Figure 3. Distribution of the types of misalignments within the contrast captions of the VideoCon dataset. We observe that the dataset has good representation for all the kinds of misalignments ranging from " + }, + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "type": "inline_equation", + "content": "24.2\\%" + }, + { + "bbox": [ + 46, + 202, + 288, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": "each caption in the input video-text datasets (§3.1) (details in Appendix §C). Then, given a video " + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": " and a misalignment type " + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": ", we prompt PaLM-2 API² [2] to generate a contrast caption and accompanied explanation (our type-specific prompts are detailed in Appendix §D)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": "Analyzing the LLM generations, we found that sometimes the output caption " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": " do not contradict the original caption " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": ". For example, a generated contrast caption \"a person riding a car\" does not contradict the original caption \"a person riding a mustang\". To filter such cases, we employ a Natural Language Inference (NLI) model [19] and remove cases in which the contrast caption is assessed as entailed by the original caption " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "NLI(T,C) > 0.5" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": ". Postfiltering, each tuple " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "(V,T,C,m)" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": " is converted to the two instances of video/language entailment task: " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "A_{en}(V,T) = 1" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "A_{en}(V,C) = 0" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": ". We present the dataset statistics for the entailment task in Table 1, including train/eval/test splits. In addition, Fig. 3 shows the distribution of misalignment types in the dataset. We observe that VideoCon maintains a high density across the 7 misalignments ranging from " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "inline_equation", + "content": "24.2\\%" + }, + { + "bbox": [ + 46, + 327, + 286, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": "We also found that some generated explanations do not describe the differences between " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": " well. For example, the explanation \"two friends are not traveling together\" does not fully describe the discrepancy between \"three friends traveling together\" and \"two friends are traveling together\". To filter these out, generated examples are removed if " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "NLI(F(T,C),E) < 0.6" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "F(T,C)" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": " is the premise comprising the original and contrast captions. Specifically, premise will be 'Expected Caption: " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": " Actual Caption: " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": " and hypothesis will be 'Difference between Expected and Actual Caption: " + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 46, + 519, + 286, + 685 + ], + "type": "text", + "content": ". This filter indicates that the information in the explanation is not entailed by the difference between the two captions. The dataset statistics for the NLE task is presented in Table 1. We refer to the final" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 125 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 125 + ], + "type": "table", + "html": "
Video-Language EntailmentNatural Language Explanation
SourceTrainValTestTrainValTest
MSR-VTT3836647816538158882066788
VaTeX664807368110301803453636
TEMPO1071270982708416527391073
Total11555883122735650233329011497
", + "image_path": "16ac4b946daaafb1768824f2b10643f39c6b1f7b30ed130c5ba5b4841b6dc2cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 133, + 533, + 143 + ], + "lines": [ + { + "bbox": [ + 317, + 133, + 533, + 143 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 533, + 143 + ], + "type": "text", + "content": "Table 1. Statistics for the VLE and NLE tasks in VideoCon." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 167, + 489, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 489, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 489, + 179 + ], + "type": "text", + "content": "LLM-generated dataset as VideoCon (LLM)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "text", + "content": "To assess the quality of VideoCon (LLM), we perform human evaluation on 500 contrast captions and NLEs (details in Appendix E). The human evaluator found " + }, + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "text", + "content": " of the contrast captions and " + }, + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "text", + "content": " of the NLEs to be valid, indicating the high-quality of VideoCon (LLM)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 251, + 476, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 476, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 476, + 264 + ], + "type": "text", + "content": "3.4. Data Generation using Humans" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 270, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 437 + ], + "type": "text", + "content": "To study whether a model trained on VideoCon (LLM) generalizes to out-of-distribution videos and its performance on human-generated contrast captions, we randomly selected a set of videos from the validation set of ActivityNet [10]. This dataset consists of captions matched with segments in the video, e.g., \"a little boy is climbing on an outside gym\" matched to the first 10 seconds of its related video. We extracted video segments with an associated caption. Human workers3 on Amazon MTurk were then shown the video segments and associated captions and were asked to create a semantically plausible contrast caption and a corresponding NLE (more details in Appendix §F). We did not communicate any type of target misalignments to encourage natural diversity of human created contrast captions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": "Overall, we collected 570 tuples " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "(V, T, C_{human}, E_{human})" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " is the video, " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " is the original caption, " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "C_{human}" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " is the human-written contrast caption, and " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "E_{human}" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " is the human-written explanations. We denote this dataset by VideoCon (Human). We sample 100 instances from this dataset, and found " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " to be clean. In addition, we observe that many of the human-generated contrast captions perturbing one or more objects (" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": ") and actions (" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": ") depicted in the caption. While " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "8\\% - 10\\%" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " of the contrast captions flip the order of the events and attribute of the objects. As this dataset is largely unfiltered, it contains a mix of temporally-easy and challenging instances. We also constructed a more temporally-challenging subset of 290 instances, denoted VideoCon (Human-Hard), by filtering out tuples in which " + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "inline_equation", + "content": "A_{vle}(V, T) < 0.5" + }, + { + "bbox": [ + 304, + 440, + 545, + 631 + ], + "type": "text", + "content": " (Eq. (1)), as in §3.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 645, + 423, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 423, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 423, + 659 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 666, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 690 + ], + "type": "text", + "content": "We next describe our evaluation setting for measuring the impact of VideoCon on video-text alignment modeling." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "https : / / developers . generativeai . google / products/palm" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 702, + 463, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 463, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 463, + 712 + ], + "type": "text", + "content": "3A shortlist that passed our qualification test." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13930" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 194, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 194, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 194, + 85 + ], + "type": "text", + "content": "4.1. Finetuning with VideoCon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "type": "text", + "content": "Our goal in constructing VideoCon (LLM) is to improve robustness of video-text alignment models by fine-tuning on this dataset. To this end, we start with the mPLUG-OwlVideo model [61], denoted Owl-Base. Its building blocks are CLIP [39] as visual encoder and LLaMA-7B [46] as text encoder/decoder and it was pretrained on VideoChat [27]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 176, + 126, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 176, + 126, + 186 + ], + "spans": [ + { + "bbox": [ + 66, + 176, + 126, + 186 + ], + "type": "text", + "content": "Entailment Task:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 194, + 236, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 194, + 236, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 194, + 236, + 205 + ], + "type": "text", + "content": "Given: V (Video), T (Caption), C (Contrast Caption)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 213, + 268, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 213, + 268, + 233 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 268, + 233 + ], + "type": "text", + "content": "Instruction (I): [V] Does this video entail the description " + }, + { + "bbox": [ + 67, + 213, + 268, + 233 + ], + "type": "inline_equation", + "content": "[T]" + }, + { + "bbox": [ + 67, + 213, + 268, + 233 + ], + "type": "text", + "content": "? Response (R): Yes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 242, + 268, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 242, + 268, + 262 + ], + "spans": [ + { + "bbox": [ + 67, + 242, + 268, + 262 + ], + "type": "text", + "content": "Instruction (I): [V] Does this video entail the description [C]? Response (R): No" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 272, + 254, + 283 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 272, + 254, + 283 + ], + "spans": [ + { + "bbox": [ + 78, + 272, + 254, + 283 + ], + "type": "text", + "content": "Figure 4. Entailment task prompt for finetuning." + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "bbox": [ + 63, + 311, + 246, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 311, + 246, + 322 + ], + "spans": [ + { + "bbox": [ + 63, + 311, + 246, + 322 + ], + "type": "text", + "content": "Natural Language Explanation Generation Task:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 331, + 240, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 331, + 240, + 342 + ], + "spans": [ + { + "bbox": [ + 64, + 331, + 240, + 342 + ], + "type": "text", + "content": "Given: V (Video), C (Contrast Caption), E (NLE)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 352, + 270, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 352, + 270, + 384 + ], + "spans": [ + { + "bbox": [ + 64, + 352, + 270, + 384 + ], + "type": "text", + "content": "Instruction (I): [V] What is the misalignment between this video and the description " + }, + { + "bbox": [ + 64, + 352, + 270, + 384 + ], + "type": "inline_equation", + "content": "[C]" + }, + { + "bbox": [ + 64, + 352, + 270, + 384 + ], + "type": "text", + "content": "? Response (R): [E]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 396, + 263, + 407 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 396, + 263, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 396, + 263, + 407 + ], + "type": "text", + "content": "Figure 5. NLE generation task prompt for finetuning." + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": "To fine-tune " + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "inline_equation", + "content": "Owl" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": "-Base on VideoCon (LLM), its " + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\{V, T, C, E\\}^4" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": " tuples were converted into two types of multimodal instruction-response pairs, one for the VLE task " + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "inline_equation", + "content": "(I_{vle}, R)" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": " (Fig. 4) and one for the NLE task " + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "inline_equation", + "content": "(I_{nle}, R)" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": " (Fig. 5). We then train " + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "inline_equation", + "content": "Owl" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": "-Base on all instruction pairs from both the tasks with maximum likelihood loss, resulting in a single model " + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "inline_equation", + "content": "Owl" + }, + { + "bbox": [ + 47, + 421, + 287, + 505 + ], + "type": "text", + "content": "-Con." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 514, + 208, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 208, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 208, + 526 + ], + "type": "text", + "content": "4.2. VideoCon Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": "To evaluate the performance of the " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "Owl-Con" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " on video-text alignment we generate " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "Owl-Con" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " response to prompt " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "I_{vle}" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " for video " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "Y \\in \\{T, C\\}" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": ". We then calculate the probability of generating responses " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "s_y = Owl-Con" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "^\\prime \\mathrm{Yes}^\\prime | I_{vle}(V, Y)" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "s_n = Owl-Con" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "^\\prime \\mathrm{No}^\\prime | I_{vle}(V, Y)" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": "), and based on these scores the probability for class 'Yes': " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "P_{yes}(V, Y) = \\frac{s_y}{s_y + s_n}" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": ". Finally, we compute the ROC-AUC score for " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "P_{yes}(V, Y)" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " over the VideoCon (LLM) eval set, with " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\{V, T\\}" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " as label 1 and " + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\{V, C\\}" + }, + { + "bbox": [ + 47, + 533, + 287, + 643 + ], + "type": "text", + "content": " as label 0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "content": "To evaluate " + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "inline_equation", + "content": "Owl" + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "content": "-Con on the NLE task, we prompt it with instruction " + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "inline_equation", + "content": "I_{nle}" + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "content": " instantiated on " + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\{V,C\\}" + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "content": " pairs from the VideoCon (LLM) eval set. We compare the generated explanation " + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "content": " to the ground truth " + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 644, + 287, + 693 + ], + "type": "text", + "content": " by measuring entailment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": "probability " + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "NLI(E, \\hat{E})" + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": ". In our experiments, we experiment with two " + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "NLI" + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": " automatic metrics: (a) " + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "Q^2" + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": " score [19], and (b) PaLM-2 API. We performed human evaluation to measure the agreement between the automatic metrics and the human-rating. We found that both metrics achieve high agreement with human assessment (Appendix §H)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 150, + 471, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 471, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 471, + 162 + ], + "type": "text", + "content": "4.3. Video-Text Downstream Tasks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 168, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 203 + ], + "type": "text", + "content": "We complement the VideoCon intrinsic evaluation over the testset with an extrinsic evaluation over two temporal and action difficult downstream tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 205, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 546, + 361 + ], + "type": "text", + "content": "We evaluate alignment model performance for text2video retrieval over SSv2-Temporal [45] and SSv2-Events [5] datasets. We consider the SSv2-Template captions instead of the label captions since they remove the object-centric bias in model evaluation [26]. We compute input-text/candidate-video alignment score, rank videos and report mean Average Precision (mAP). We evaluate alignment model performance for video question answering over the ATP-Hard [9] dataset. We cast each question/candidate-answer pair as an imperative statement using PaLM-2 API, measure alignment to the input video and report Accuracy. More details on the downstream datasets and the evaluation setup are in Appendix §I." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 366, + 372, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 372, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 372, + 377 + ], + "type": "text", + "content": "4.4. Baselines" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "type": "text", + "content": "For the video-text alignment text, we compare Owl-Con with the following baselines: (a) End-to-End VNLI as zero-shot atemporal model since it does not have access to the temporal order of the video frames, (b) VideoCLIP [55], (c) ImageBind [14], (d) Owl-Base, and (e) Owl-Rand: Owl-Base fine-tuned on VideoCon tuples " + }, + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "type": "inline_equation", + "content": "\\{V,T,\\hat{C},E\\}" + }, + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 305, + 384, + 545, + 516 + ], + "type": "text", + "content": " is randomly selected from other captions in the dataset. Owl-Rand would indicate if there is merit in the contrast, hard-negative captions in VideoCon. We include additional baselines TACT [5] and VFC [37] for evaluating on the downstream tasks (\\$5.3)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 525, + 387, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 387, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 387, + 539 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 545, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 582 + ], + "type": "text", + "content": "We present our intrinsic (VideoCon eval set) and extrinsic (downstream tasks) evaluation results, showing the benefits of VideoCon for robust video-language alignment." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 587, + 532, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 532, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 532, + 599 + ], + "type": "text", + "content": "5.1. Performance on VideoCon Entailment Task" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": "We present the ROC-AUC scores of the tested models in Table 2. From the table we see that the baseline models find the VideoCon testset difficult, as reflected by low AUC scores (e.g. Owl-Base-57.2), close to random. Even training on VideoCon train instances, but with \"easy\" negatives (Owl-Rand-59.7), hardly improves the base models. A significant improvement is achieved with the VNLI-specific model (67), showing that the entailment task is not inherently represented in generic video-language aligned training" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "inline_equation", + "content": "^{4}\\mathrm{V}" + }, + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "text", + "content": " : video, T: original caption, C: contrast caption, E: explanation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13931" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 69, + 525, + 170 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 525, + 170 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 525, + 170 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 525, + 170 + ], + "type": "table", + "html": "
ModelsVideoCon (LLM) TestVideoCon (Human)VideoCon (Human-Hard)
Random50.050.050.0
VideoCLIP [55]53.247.347.5
ImageBind (Video-Text) [14]57.165.263.0
Owl-Base [61]57.266.864.1
Owl-Rand59.768.965.5
End-to-End VNLI [60]67.072.465.0
Owl-Con (Ours)84.678.374.4
", + "image_path": "a64ff6bbdc2bc8e69de5f40836651737118ddce24c4cf8141bf33fb5d11c879d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 199, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 127, + 178, + 465, + 189 + ], + "lines": [ + { + "bbox": [ + 127, + 178, + 465, + 189 + ], + "spans": [ + { + "bbox": [ + 127, + 178, + 465, + 189 + ], + "type": "text", + "content": "Table 2. ROC-AUC scores of the tested models for the entailment task on VideoCon test sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 199, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 48, + 199, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 547, + 255 + ], + "type": "table", + "html": "
VideoCon (LLM)VideoCon (Human)
ModelsQ2entailmentPaLM-2 entailment acc. (%)Q2entailmentPaLM-2 entailment acc.(%)
Owl-Base0.1936.80.2339.6
Owl-Con (Ours)0.5065.40.3247.1
", + "image_path": "b970c390320b1cf036e5184546f902cfbc8291d7f71a907ea2c5b7c8e8748e71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "content": "sets and requires specific training. Yet, the best performance is achieved by training on VideoCon, which addresses the diversity in plausible misalignments and includes \"difficult\" training examples, reaching 84.6 AUC. This demonstrates the merit of VideoCon for improving video-language alignment robustness. We show qualitative examples for the model predictions in §6.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": "When evaluating on out-of-domain (OOD) data around video types and misalignment distribution, we again see that training with VideoCon offers significant improvement to alignment detection, outperforming all baselines, albeit with smaller relative gains: " + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": " improvement compared to Owl-Base on (Human) and (Human-Hard) respectively compared to " + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": " on (LLM) test. In future work, we plan to further diversify the misalignments VideoCon covers to further improve its benefits on OOD cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 485, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 569 + ], + "type": "text", + "content": "We notice that the performance of the VNLI atemporal model is better than existing video-language alignment models. It might be attributed to its training with contrast captions in [60]. It further highlights that the existing videolanguage models are not robust in comparison to a temporal probe on video-language alignment evaluation, corroborating the findings from [9, 26]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 251, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 251, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 251, + 588 + ], + "type": "text", + "content": "5.2. Performance on NLE Generation Task" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Table 3 presents the performance of the tested models against the ground-truth on the NLE task, depicting average " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "Q^2" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " score and PaLM-2 entailment accuracy. The results show that on in-domain VideoCon, Owl-Con outperforms Owl-Base by an impressive " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "263\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "178\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " relative increase on " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "Q^2" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " score and PaLM-2 accuracy respectively. This indicates the finetuned model can accurately generate NLE that match well with the ground-truth NLE. This indicates that our model can generate accurate NLE for a wide range of misalignments in the video captions, which makes it use" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 293, + 511, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 511, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 511, + 305 + ], + "type": "text", + "content": "ful for dense video-language alignment evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": "On out-of-domain VideoCon, the improvement is more moderate but still high: " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " relative increase on " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "Q^2" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " and PaLM-2 respectively. This is probably due to the more diverse ways humans express explanations compared to LLM prompting. In future work we plan to further address linguistic diversity in explanations for more robust generation and evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 399, + 545, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 411 + ], + "type": "text", + "content": "5.3. Performance on Video-Text Downstream Tasks" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 426, + 545, + 535 + ], + "blocks": [ + { + "bbox": [ + 107, + 262, + 485, + 272 + ], + "lines": [ + { + "bbox": [ + 107, + 262, + 485, + 272 + ], + "spans": [ + { + "bbox": [ + 107, + 262, + 485, + 272 + ], + "type": "text", + "content": "Table 3. Performance of the tested models on the NLE generation task, measured via entailment metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 426, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 535 + ], + "type": "table", + "html": "
ModelsSSv2-Temporal\nmAPSSv2-Events\nmAP
Random7.33.3
VideoCLIP9.86.4
ImageBind (video-language)10.55.5
Owl-Base10.96.8
TACT [5]-7.8
Owl-Rand12.19.9
End-to-End VNLI [60]14.610.4
Owl-Con (Ours)15.211.4
", + "image_path": "4f6617e9e52241b4b493574a3e2d2bfc1c4befbadc737ed63d3587ec131c3856.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 543, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 305, + 543, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 545, + 564 + ], + "type": "text", + "content": "Table 4. Mean Average Precision (mAP) scores for the tested models in the zero-shot text-to-video retrieval tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": "We next present our results on the two downstream tasks, Text2Video Retrieval and Video Question Answering. Starting with the retrieval task, we report mean Average Precision (mAP) of the tested models on the SSv2-Temporal and SSv2-Events datasets in Table 4. The benefits of training with additional examples tailored for temporal video-language alignment is already evident in the performance of Owl-Rand, which improves over the previous SSv2-Events SOTA - TACT with a relative increase of " + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "However, when training on harder negative contrastive instances, " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "Owl-Con" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " achieves a significant improvement," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13932" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 73, + 558, + 200 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 558, + 200 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 558, + 200 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 558, + 200 + ], + "type": "table", + "html": "
SuccessFailure
VideoCon LLMCaption: The adult gets the child's attention after tapping his shoulderECaption: Two people move a large object down a narrow stairwayE
Contrast Caption: The adult gets the child's attention before tappingCContrast Caption: Two people move a large object up a narrow stairwayE
GT NLE: First, an adult hand is seen tapping the child's shoulder, then the adult gets the child's attention, not the other way aroundE GTGT NLE: Men move a large piece of furniture down a narrow staircase, not up a narrow staircaseC GT
Predicted NLE: First, the adult taps the child's shoulder, then the adult gets the child's attention, not the other way aroundPredicted NLE: Three men are moving a large piece of furniture, not two
VideoCon HumanCaption: We see the group making cookiesECaption: A lady puts her hair in a bun on the sideE
Contrast Caption: We see the group eating cookiesCContrast Caption: A lady puts her hair in a braid on the sideE
GT NLE: The group is making cookies, not eating themE GTGT NLE: The woman puts her hair in a bun, not a braidC GT
Predicted NLE: We see the group cooking cookies, not eating themPredicted NLE: The lady puts her hair in a braid on top of her head, not on the side
", + "image_path": "458d52e5b4438b2d8b32a26a88cf7b87c83d520e6458b892e38f14e26a2b8b4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "type": "text", + "content": "outperforming all baselines, with a relative increase over the best baseline End-to-End VNLI model by " + }, + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "type": "text", + "content": " on SSv2-Temporal and " + }, + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 289, + 287, + 360 + ], + "type": "text", + "content": " on SSv2-Events (46% over TACT), setting new SOTA results. This points at the benefits of exposing the model to temporal examples, such as action and event-order." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 366, + 267, + 490 + ], + "blocks": [ + { + "bbox": [ + 67, + 366, + 267, + 490 + ], + "lines": [ + { + "bbox": [ + 67, + 366, + 267, + 490 + ], + "spans": [ + { + "bbox": [ + 67, + 366, + 267, + 490 + ], + "type": "table", + "html": "
ModelsAccuracy (%)
CLIP23.8
VideoCLIP23.4
ImageBind (video-language)25.4
TACT [5]27.6
VFC [37]31.4
Owl-Base37.1
Owl-Rand37.2
End-to-End VNLI [60]39.0
Owl-Con (Ours)41.1
", + "image_path": "6cda5e936b47fd1ad8499354047b43aed06d552804ed4debad0495eca75d6f19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 521 + ], + "type": "text", + "content": "Table 5. Accuracy scores for the tested models on the zero-shot video question-answering task on ATP-Hard dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": "For the Video Question Answering task, we compare the performance of the various models in Table 5. Here too Owl-Con achieves SOTA results and outperforms the strongest baseline End-to-End VNLI model with a relative increase of " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": ". This corroborates the observations in our other experiments, which demonstrate the advantage of the VideoCon datasets, covering various misalignments, especially those pertaining to temporal and causal reasoning over dynamic events. The results also confirm the need for carefully chosen contrastive negative examples, showing that picking negatives at random may mask out the potential benefit of an alignment training set. Finally, the competitive performance of atemporal End-to-End VNLI model on the downstream tasks is surprising and underscores the need for stronger video-language datasets for robust benchmarking." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 288, + 521, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 213, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 545, + 268 + ], + "type": "text", + "content": "Figure 6. Qualitative examples for the success (green) and failure (red) modes of our model. In every example, we present a few video frames in an temporal order from top to bottom, its associated caption, contrast caption, ground-truth NLE from the datasets. Additionally, we present the predicted NLE from our model. The small boxes at the end of caption cells indicate whether our model considers that caption to be grounded in the video. E and C indicates that the model predicts the caption to entail and contradict to the video, respectively. E-GT and C-GT indicates the predicted NLE entails and contradicts the ground-truth (GT) NLE, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 288, + 521, + 437 + ], + "lines": [ + { + "bbox": [ + 331, + 288, + 521, + 437 + ], + "spans": [ + { + "bbox": [ + 331, + 288, + 521, + 437 + ], + "type": "image", + "image_path": "ad723f65f8a9060a0ae314f8fac1fbbcd78fa7b8eb947d4dba5d3979b7235912.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "text", + "content": "Figure 7. ROC-AUC of End-to-End VNLI, Owl-Base, and OwlCon across all types of misalignment in VideoCon (LLM) test set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 488, + 365, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 365, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 365, + 502 + ], + "type": "text", + "content": "6. Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 508, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 556 + ], + "type": "text", + "content": "We analyze Owl-Con's performance improvements across the kinds of misalignments in VideoCon. Additionally, we present a few qualitative examples to highlight the success and failure modes of our model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 563, + 504, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 504, + 576 + ], + "type": "text", + "content": "6.1. Per-misalignment Entailment Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": "We compared the ROC-AUC scores of the atemporal End-to-End VNLI, Owl-Base, and Owl-Con on specific misalignments in the contrast captions from VideoCon (LLM) testset in Figure 7. We observed that Owl-Con outperforms the baseline models across all misalignment types. This suggests that our model can reason well about the entities, their relations, and the temporal order of events in the video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "The largest improvement of " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\text{Owl-Con}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " compared to the two baselines is on event order flip, indicating that the baselines lack temporal understanding and the VideoCon is efficient in adding this capability to an alignment model. In" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13933" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "addition, on hallucination both Owl-Con and End-to-End VNLI significantly outperform Owl-Base, since both models were explicitly exposed to entailment/non-entailment training data. It is surprising to see that while End-to-End VNLI was trained on significantly more entailment data, much of it human-curated, Owl-Con outperforms it with only automatically generated data. This could be due to the better encoding of video in Owl-Con compared to the atemporal nature of End-to-End VNLI. Finally, the analysis shows other types of atemporal misalignments that are difficult for End-to-End VNLI to sort out, e.g. counting' and relation, where the training data in VideoCon is useful to improve these capabilities as well. This shows that our approach of detailed analysis of misalignment types of generation of examples for them is effective." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 257, + 171, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 171, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 171, + 270 + ], + "type": "text", + "content": "6.2. Qualitative Examples" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 276, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 289, + 468 + ], + "type": "text", + "content": "We highlight a few classification examples of Owl-Con in Figure 6. The rows refer to the test source of the instances and the columns refer to the success and failure modes, respectively. In Row1/Column1, we observe that our model provides correct predictions for the entailment between the video and original caption while predicting contradiction for the contrast caption that flips the order of the events i.e., grabbing attention and tapping shoulders. Interestingly, our model can also provide the accurate NLE when prompted with the video and the contrast caption. This suggests that our model is useful for providing fine-grained details about the video-language alignment. In Row2/Column2, the model confuses 'buns' with 'braids' in hair and gives a wrong NLE that contradicts the ground-truth. This error, due to its inability to distinguish between objects, might be improved with diverse videos and captions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 476, + 134, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 134, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 134, + 489 + ], + "type": "text", + "content": "7. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 496, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 289, + 653 + ], + "type": "text", + "content": "Foundation Models for Video-Language Understanding. Foundation models have emerged for video-language understanding [1, 4, 49, 55, 56] by pre-training on large amount of video-text pairs scraped from the web [6, 36, 58]. Additionally, prior works have either leveraged the pretrained CLIP model for video-language tasks [12, 33, 34] or adopted a socratic approach [50, 63] to employ LLMs (GPT-3) in reasoning over video captions. We highlight that despite the large-scale training of the video-language foundation models [14, 55, 56], they lack robustness to semantic changes to the captions which severely limits their real-world use for alignment applications. We provide a fix by training models on a novel video-centric VideoCon dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Improving Video-Language Robustness. Prior work [37, 38, 51] highlights that the video-text models cannot comprehend the semantics of the text with focus on manipulating the verb, actions, and entities grounded in the video" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "description. To improve the temporal understanding, [5] finetunes a pretrained model with temporal order loss. Despite this, their models do not achieve good zero-shot performance on downstream tasks consistently. In our work, we categorize a wide range of plausible misalignments in the contrast captions, and create a temporally-challenging VideoCon dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 171, + 547, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 340 + ], + "type": "text", + "content": "Video-Language Alignment Evaluation. Many applications such as text-to-video retrieval [15, 48, 57] and text-to-video generation [7, 47] require evaluation of the semantic alignment between the natural language text and raw video. In this work, we indicate that the existing video-text models such as VideoCLIP and ImageBind are not robust to semantic changes in the video captions, which becomes critical for faithful video-text alignment evaluation. In our work, we propose VideoCon and finetune a video-language generative model to perform robust entailment task and provide fine-grained NLE for the observed misalignments between the video and text. In the future, our model can be utilized to enhance alignment through sparse (entailment scores) and dense (fine-grained NLE) feedback [43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 350, + 379, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 379, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 379, + 363 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": "We introduced a comprehensive dataset, VideoCon, designed for robust video-text alignment. It features various semantic misalignments and explanations for text-video discrepancies. Through finetuning video-language models on this dataset, we enhanced their performance on complex tasks like text-to-video retrieval and video question answering, achieving state-of-the-art results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 455, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 612 + ], + "type": "text", + "content": "One current limitation and an important future direction is to increase the complexity of the generated contrast captions. Specifically, the model may encounter several misalignments within a single contrast caption. Addressing this issue, the model should be equipped to accurately assign low entailment scores to these contrast captions and consequently generate precise NLEs. An important future direction is to scale VideoCon to larger datasets. Here, we create contrast captions for high-quality captions written by humans for every video, however, the web-scale datasets have low-quality captions that are not well grounded in the video. In this regard, using synthetic data followed by VideoCon-like contrast caption generation can be a plausible approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 621, + 416, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 621, + 416, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 621, + 416, + 635 + ], + "type": "text", + "content": "9. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "This material is based on research supported by the ECOLE program under Cooperative Agreement HR00112390060 with the US Defense Advanced Research Projects Agency (DARPA). In addition, the research is partly funded by ONR grant N00014-23-1-2780. Hritik Bansal is supported in part by AFOSR MURI grant FA9550-22-1-0380." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13934" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "text", + "content": "[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "text", + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 316 + ], + "type": "text", + "content": "[4] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 8, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "type": "text", + "content": "[5] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 288, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 428 + ], + "type": "text", + "content": "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 288, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 288, + 496 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 609 + ], + "type": "text", + "content": "[9] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the \"video\" in video-language understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2917-2927, 2022. 1, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 611, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 666 + ], + "type": "text", + "content": "[10] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 668, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 288, + 700 + ], + "type": "text", + "content": "[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "text", + "content": "[12] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021.8, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "text", + "content": "[13] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datcomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "type": "text", + "content": "[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 293 + ], + "type": "text", + "content": "[15] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 8, 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "text", + "content": "[16] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Sarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[17] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 1, 3, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "text", + "content": "[18] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "text", + "content": "[19] Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. Q2: Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. arXiv preprint arXiv:2104.08202, 2021. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 482, + 431, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 431, + 493 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 431, + 493 + ], + "type": "text", + "content": "[20] https://commoncrawl.org/. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "type": "text", + "content": "[22] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 592, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 657 + ], + "type": "text", + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13935" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "text", + "content": "[27] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "text", + "content": "[28] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. Desco: Learning object recognition with rich language descriptions. arXiv preprint arXiv:2306.14060, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 274 + ], + "type": "text", + "content": "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 277, + 287, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 287, + 309 + ], + "type": "text", + "content": "[30] Weixin Liang, James Zou, and Zhou Yu. Alice: Active learning with contrastive natural language explanations. arXiv preprint arXiv:2009.10259, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 311, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 311, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 49, + 311, + 287, + 343 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 345, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 345, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 345, + 287, + 388 + ], + "type": "text", + "content": "[32] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 390, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 390, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 49, + 390, + 287, + 433 + ], + "type": "text", + "content": "[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 8, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 434, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 489 + ], + "type": "text", + "content": "[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 8, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 49, + 491, + 287, + 545 + ], + "type": "text", + "content": "[35] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10910-10921, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 547, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 547, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 547, + 287, + 612 + ], + "type": "text", + "content": "[36] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 8, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 668 + ], + "type": "text", + "content": "[37] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[38] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, 2022. 1, 2, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 162 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "text", + "content": "[42] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8322-8332, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "type": "text", + "content": "[43] Jérémy Scheurer, Jon Ander Campos, Tomasz Korbak, Jun Shern Chan, Angelica Chen, Kyunghyun Cho, and Ethan Perez. Training language models with language feedback at scale. arXiv preprint arXiv:2303.16755, 2023. 8, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 378, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 443 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "text", + "content": "[45] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 535-544, 2021. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "text", + "content": "[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 8, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "type": "text", + "content": "[48] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 8, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 8, 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13936" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[50] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems, 35: 8483-8497, 2022. 8, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 139, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 139, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 288, + 194 + ], + "type": "text", + "content": "[51] Zhenhailong Wang, Ansel Blume, Sha Li, Genglin Liu, Jaemin Cho, Zineng Tang, Mohit Bansal, and Heng Ji. Paxion: Patching action knowledge in video-language foundation models. arXiv preprint arXiv:2305.10683, 2023. 1, 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "text", + "content": "[52] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "text", + "content": "[53] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment task for visually-grounded language learning. arXiv preprint arXiv:1811.10582, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "text", + "content": "[54] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding. arXiv preprint arXiv:1901.06706, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "type": "text", + "content": "[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "type": "text", + "content": "[56] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023.8.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 427, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 471 + ], + "type": "text", + "content": "[57] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 3, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 472, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 536 + ], + "type": "text", + "content": "[58] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "type": "text", + "content": "[59] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "type": "text", + "content": "[60] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400, 2023. 3, 6, 7, 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 228 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[62] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[63] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 8, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "text", + "content": "[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI Conference on Artificial Intelligence, pages 7590-7598, 2018. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13937" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_content_list.json b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..96152fea57e307fbfa77a163c910ef8c3526cd20 --- /dev/null +++ b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_content_list.json @@ -0,0 +1,1653 @@ +[ + { + "type": "text", + "text": "VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models", + "text_level": 1, + "bbox": [ + 168, + 128, + 803, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoxin Chen", + "bbox": [ + 205, + 203, + 313, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong Zhang*", + "bbox": [ + 354, + 204, + 459, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaodong Cun", + "bbox": [ + 493, + 204, + 611, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Menghan Xia", + "bbox": [ + 651, + 204, + 763, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xintao Wang", + "bbox": [ + 300, + 222, + 406, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao Weng", + "bbox": [ + 447, + 222, + 542, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ying Shan", + "bbox": [ + 584, + 223, + 666, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tencent AI Lab", + "bbox": [ + 416, + 256, + 545, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In cyberpunk, neonpunk style,", + "bbox": [ + 94, + 529, + 274, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kung Fu Panda, jump and kick.", + "bbox": [ + 99, + 544, + 287, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cinematic photo melting pistachio ice", + "bbox": [ + 346, + 515, + 575, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cream dripping down the cone.", + "bbox": [ + 351, + 530, + 535, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "35mm photograph, film, bokeh.", + "bbox": [ + 352, + 542, + 537, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large motion, surrounded by butterflies,", + "bbox": [ + 625, + 529, + 864, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a girl walks through a lush garden.", + "bbox": [ + 627, + 544, + 838, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Give a text prompt, our method can generate a video with high visual quality and accurate text-video alignment. Note that it is trained with only low-quality videos and high-quality images. No high-quality videos are required. Best viewed with Acrobat Reader. Click the images to play the video clips.", + "bbox": [ + 75, + 566, + 893, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 622, + 313, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video generation aims to produce a video based on a given prompt. Recently, several commercial video models have been able to generate plausible videos with minimal noise, excellent details, and high aesthetic scores. However, these models rely on large-scale, well-filtered, high-quality videos that are not accessible to the community. Many existing research works, which train models using the low-quality WebVid-10M dataset, struggle to generate high-quality videos because the models are optimized to fit WebVid-10M. In this work, we explore the training scheme of video models extended from Stable Diffusion and investigate the feasibility of leveraging low-quality videos and synthesized high-quality images to obtain a high-quality video model. We first analyze the connection between the spatial", + "bbox": [ + 75, + 656, + 472, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and temporal modules of video models and the distribution shift to low-quality videos. We observe that full training of all modules results in a stronger coupling between spatial and temporal modules than only training temporal modules. Based on this stronger coupling, we shift the distribution to higher quality without motion degradation by finetuning spatial modules with high-quality images, resulting in a generic high-quality video model. Evaluations are conducted to demonstrate the superiority of the proposed method, particularly in picture quality, motion, and concept composition.", + "bbox": [ + 500, + 625, + 895, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 811, + 632, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from the development of diffusion models [25, 43], video generation has achieved breakthroughs, particularly in basic text-to-video (T2V) generation models. Most existing methods [14, 23, 26, 30, 48, 63] follow a logic to", + "bbox": [ + 498, + 839, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 76, + 886, + 207, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7310", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "obtain video models, i.e., extending a text-to-image (T2I) backbone to a video model by adding temporal modules and then training it with videos. Several models train video models from scratch, while most start from a pre-trained T2I model, typically Stable Diffusion (SD) [38]. Models can also be categorized into two groups based on the space modeled by diffusion models, i.e., pixel-space models [26, 30, 60] and latent-space models [14, 23, 48, 63]. The latter is the dominant approach. Picture quality, motion consistency, and concept composition are essential dimensions for evaluating a video model.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a few commercial startups have released their T2V models [5, 6, 8, 9] that can produce plausible videos with minimal noise, excellent details, and high aesthetic scores. However, they are trained on a large-scale and well-filtered high-quality video dataset, which is not accessible to the community and academia. Collecting millions of high-quality videos is challenging due to copyright restrictions and post-filtering processing. Though there are a few open-source video datasets collected from the Internet for video understanding, such as HowTo100M [33], HD-VILA100M [56], and InterVid [51], there exist many issues for video generation, e.g., poor picture quality and caption, multiple clips in one video, and static frames or slides. WebVid10M [12] is the most widely used dataset to train video generation models in academia. The clips are well-segmented, and the diversity is good. However, the picture quality is unsatisfactory, and most videos have a resolution of about $320\\mathrm{p}$ . The lack of high-quality datasets poses a significant obstacle to training high-quality video models in academia.", + "bbox": [ + 75, + 261, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we target a quite challenging problem, i.e., training high-quality video models without using high-quality videos. We dive into the training process of SD-based video models to analyze the connection between spatial and temporal modules under different training strategies and investigate the distribution shift to low-quality videos. We make a meaningful observation that the full training of all modules results in a stronger coupling between appearance and motion than just training temporal modules. The full training can achieve more natural motion and tolerate more subsequent modification of spatial modules, which is the key to improving the quality of generated videos. Based on the observation of the connection, we propose a method to overcome the data limitation by disentangling motion from appearance at the data level. Specifically, instead of high-quality videos, we exploit low-quality videos to guarantee motion consistency and use high-quality images to ensure picture quality and concept composition ability. Benefiting from the successful T2I models such as SDXL and Midjourney, it is convenient to obtain a large set of images with high-resolution and complex concept composition. As shown in Fig. 2, we design a pipeline to train a video diffusion model following the guidelines from the analysis. Fig. 1", + "bbox": [ + 75, + 553, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shows visual examples generated by our method.", + "bbox": [ + 500, + 90, + 823, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 517, + 106, + 857, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a method to overcome the data for training high-quality video models by disentangling motion from appearance at the data level.", + "- We investigate the connection between spatial and temporal modules, and the distribution shift. We identify the keys to obtain a high-quality video model.", + "- We design an effective pipeline based on the observations, i.e., obtaining a fully trained video model first and tuning the spatial modules with synthesized high-quality images." + ], + "bbox": [ + 500, + 121, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 271, + 640, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The evolution of video generation techniques goes along with the development of generative models. Generative adversarial networks [17] and variational auto-encoders [18] are the commonly used backbones in early research of video generation, e.g., VGAN [47], TGAN [40], MoCoGAN [44], GODIA [52], StyleGAN-V [41], and MCVD [46]. Then, since transformers have been successfully applied in various fields, they are also introduced for video synthesis, e.g.,CogVideo [28], VideoGPT [57], NUVA-infinity [53], TATS [19], MAGVIT [58], Phenaki [45].", + "bbox": [ + 496, + 297, + 893, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models (DMs) [25, 42, 43] have been a famous star in generative models, especially in text-to-image (T2I) generation [13, 21, 24, 34, 36-39, 62]. For video generation, Video Diffusion Models (VDMs) are proposed to model the distribution of videos. VDM [27] is the first to utilize a space-time factorized U-Net to model videos in pixel space for unconditional video generation. It uses an image-video joint training strategy to avoid concept forgetting.Imagen Video [26] and Make-a-Video [30] are two cascade models that target text-to-video generation in pixel space. Show-1 [60] is another cascade model that uses IF [1] as the base model and LDM extended video model for super-resolution. Subsequently, LVDM [15, 23] and MagicVideo [63] propose to extend LDM [38] to model videos in the latent space of an auto-encoder. Many other methods use the same paradigm, including ModelScope [48], Align Your Latent [14], Hotshot-XL [7], LAVIE [50], PY-OCO [20], VideoFactory [49], VPDM [59], VIDM [32], and Latent-Shift [11]. Besides text-to-video generation, a few methods, such as [16, 55, 61], attempt to generate videos from a given image and a prompt as condition.", + "bbox": [ + 496, + 448, + 895, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several startups release their text-to-video generation services, e.g., Gen-2 [5], Pika Labs [9], Moonvalley [8], and Genmo [6]. Their models can generate plausible results with minimal noise, excellent details, and high aesthetic scores. However, those methods are trained with a large-scale well-filtered high-quality video dataset that is not accessible to researchers. The video models are also not available, leading to the slow development of downstream tasks to a certain extent. The most widely used video dataset is WebVid-10M,", + "bbox": [ + 496, + 765, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7311", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/74ee483ed404f3eed19ed4671a4f5cbd90ee1bcbdff88e7ed86c6cae3756a30a.jpg", + "image_caption": [ + "Figure 2. The efficient training pipeline. In the first stage, we use a large amount of low-quality videos to fully train the VDM to learn motion. In the second stage, we only train the spatial parameters with the high-quality images generated by the T2I model to improve the concept composition and picture quality." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a large-scale dataset of short videos with textual descriptions sourced from stock footage sites. The videos are diverse and rich in their content, and each video is well-segmented, however, the picture quality is unsatisfactory and most videos are $320\\mathrm{p}$ . Training a high-quality video model under the data limitation is quite challenging.", + "bbox": [ + 75, + 320, + 470, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AnimateDiff [22] finds that combining temporal modules from a video model trained on WebVid-10M and a LORA SD model can improve the picture quality of the generated videos. However, this is not a generic model and does not always work. There are a few severe issues. First, the temporal modules can only be combined with a few selected LORA models, which makes it not a generic model. Second, the motion quality degenerates when the modules do not match the LORA model well. UnlikeAnimateDiff, we analyze the connection between spatial and temporal modules instead of direct combination, and design a pipeline to train a generic high-quality video model without high-quality video by disentangling appearance and motion at the data level.", + "bbox": [ + 75, + 412, + 473, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 625, + 168, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an effective method to overcome the data limitation for training high-quality video diffusion models. We first analyze the connection between spatial and temporal modules of SD-based video models under different training strategies. Based on the observations, we then develop a pipeline to train high-quality video models with just low-quality videos and high-quality images, i.e., disentangling appearance from motion at the data level.", + "bbox": [ + 75, + 651, + 470, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spatial-temporal Connection Analyses", + "text_level": 1, + "bbox": [ + 76, + 785, + 408, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Base T2V model. To leverage the prior in SD trained on a large-scale image dataset, most text-to-video diffusion models inflate the SD model to a video model by adding temporal modules, including Align Your Latent [14],AnimateDiff [22], LVDM [23], Magic Video [63], ModelScopeT2V [48], and LAVIE [50]. They follow VDM [27]", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to use a particular type of 3D-UNet that is factorized over space and time.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These models can be categorized into two groups according to their training strategies. One is to use videos to train temporal modules with spatial ones fixed, called partial training. The other is to learn both the spatial and temporal modules with the SD weights as initialization, called full training.", + "bbox": [ + 496, + 121, + 893, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Though these SD-based T2V models have similar architectures, they are trained under different training settings. We use one typical model to investigate the connection between spatial and temporal modules under the two training strategies. We follow the architecture of the open-source VideoCrafter1 [15] with FPS (frames per second) condition. We also incorporate the temporal convolution in ModelScopeT2V [48] to improve temporal consistency.", + "bbox": [ + 496, + 210, + 893, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter Perturbation for Full and Partial Training. We apply the two training strategies to the same architecture using the same data. The model is initialized from pretrained SD weights. WebVid-10M [12] is exploited as the training data. To avoid concept forgetting, LAION-COCO 600M [3] is also used for video and image joint training. The resolution is $512 \\times 320$ . For simplicity, the fully trained video model is denoted as $M_F(\\theta_T, \\theta_S)$ , while the partially trained one is denoted as $M_P(\\theta_T, \\theta_S^0)$ , where $\\theta_T$ and $\\theta_S$ are the learned parameters of the temporal and spatial modules, respectively. $\\theta_S^0$ are the original spatial parameters of SD.", + "bbox": [ + 496, + 349, + 893, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the connection strength between spatial and temporal modules, we perturb the parameters of the specified modules by using another high-quality image dataset $\\mathcal{D}_I$ for finetuning. The image data is JDB [35] that consists of synthesized images from Midjourney [4]. As the JDB has 4 million images, we use LORA [29] for finetuning.", + "bbox": [ + 496, + 517, + 890, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Perturbation. We first perturb the spatial parameters of the two video models using the image dataset. The temporal parameters are frozen. The perturbation process of the fully trained base model $M_F$ can be denoted as:", + "bbox": [ + 496, + 607, + 893, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {F} ^ {\\prime} (\\theta_ {T}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}) \\leftarrow \\mathrm {P E R T B} _ {\\theta_ {S}} ^ {\\mathrm {L O R A}} (M _ {F} (\\theta_ {T}, \\theta_ {S}), \\mathcal {D} _ {I}),\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 678, + 879, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{PERTB}_{\\theta_S}^{\\mathrm{LORA}}$ denote finetuning $M_F$ with respect to $\\theta_{S}$ on the image dataset $\\mathcal{D}_I$ using LORA. $\\Delta_{\\theta_S}$ represents the parameters of the LORA branch. Similarly, we can obtain the perturbed model of the partially trained video model:", + "bbox": [ + 496, + 708, + 890, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {P} ^ {\\prime} \\left(\\theta_ {T}, \\theta_ {S} ^ {0} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {S}} ^ {\\text {L O R A}} \\left(M _ {P} \\left(\\theta_ {T}, \\theta_ {S} ^ {0}\\right), \\mathcal {D} _ {I}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 780, + 879, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For easy understanding, we also use the name 'F-SpALORA' to denote model $M_F^{\\prime}$ and 'P-Spa-LORA' for $M_P^\\prime$ . 'F' denotes the fully trained base model while $\\mathbf{P}^{\\prime}$ stands for the partially trained model. 'Spa' and Temp' mean finetuning spatial and temporal modules, respectively. 'LORA' represents using LORA for finetuning, while DIR' means direct", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7312", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A professional dancer gracefully performs ballet on stage. \nFigure 3. Perturbing spatial modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips.", + "bbox": [ + 76, + 386, + 470, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "finetuning without LORA. For example, 'F-Spatial-LORA' represents perturbing spatial modules of the fully trained T2V model using LORA.", + "bbox": [ + 76, + 446, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparing the synthesized videos of the two resulting models, we have the following observations. First, the motion quality of F-Spa-LORA is more stable than that of P-Spa-LORA (see user study in Table 4). The motion of P-Spa-LORA becomes worse quickly during the finetuning process. The more finetuning steps, the video tends to be more still with local flicker (see Fig. 3). While the motion of F-Spa-LORA slightly degenerates compared to the fully trained base model. Second, P-Spa-LORA achieves much better visual quality than F-Spa-LORA (see Fig. 3). The picture quality and aesthetic score of F-Spa-LORA are greatly improved compared to the partially trained base model (see Table 3). Surprisingly, the watermark is also removed. While F-Spa-LORA obtains a slight improvement in picture quality and aesthetic score, the generated videos are still noisy.", + "bbox": [ + 75, + 492, + 470, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the two observations, we can conclude that the coupling strength between spatial and temporal modules of the fully trained model is stronger than that of the partially trained model. Because the spatial-temporal coupling of the partially trained model can be easily broken, leading to quick motion degeneration and picture quality shift. A stronger connection can tolerate parameter perturbation more than a weak one. Our observation can be used to explain the quality improvement and motion degeneration ofAnimateDiff.AnimateDiff is not a generic model and only works for selected personalized SD models. The reason is that its motion modules are obtained with the partially training strategy, and", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robot dancing in times square.", + "text_level": 1, + "bbox": [ + 599, + 246, + 787, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4. Perturbing temporal modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips.", + "bbox": [ + 500, + 261, + 893, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "they cannot tolerate large parameter perturbations. When the personalized model does not match the temporal modules, both picture and motion quality will degenerate.", + "bbox": [ + 498, + 305, + 892, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Perturbation. The partially trained model has only the temporal modules updated, but the picture quality is shifted to the quality of WebVid-10M. Hence, the temporal modules take responsibility for not only the motion but also the picture quality. We perturb the temporal modules while fixing the spatial modules with the image dataset. The perturbation processes can be denoted as:", + "bbox": [ + 498, + 349, + 893, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {F} ^ {\\prime \\prime} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}} ^ {\\text {L O R A}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 465, + 879, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {P} ^ {\\prime \\prime} (\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} ^ {0}) \\gets \\mathrm {P E R T B} _ {\\theta_ {T}} ^ {\\mathrm {L O R A}} (M _ {P} (\\theta_ {T}, \\theta_ {S} ^ {0}), \\mathcal {D} _ {I}).\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 488, + 879, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the picture quality of P-Temp-LORA $(M_P^{\\prime \\prime})$ is better than F-Temp-LORA $(M_F^{\\prime \\prime})$ . However, the foreground and background of the videos are more shaky, i.e., the temporal consistency becomes worse (see Fig. 4). The picture of F-Temp-LORA is improved, but the watermark is still there. Its motion is close to the base model and much better than P-Temp-LORA (see Table 4). Those observations also support the conclusion obtained from spatial perturbation.", + "bbox": [ + 498, + 518, + 893, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data-level Disentanglement of Appearance and Motion", + "text_level": 1, + "bbox": [ + 500, + 650, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since obtaining a large-scale, high-quality video dataset with high diversity is challenging due to copyright issues, we explore the possibility of training a high-quality video model without using high-quality videos. We have access to low-quality videos such as WebVid-10M and high-quality images such as JDB. We propose to disentangle motion from appearance at the data level, i.e., learning motion from low-quality videos while learning picture quality and aesthetics from high-quality images. We can first train a video model with videos and then fine-tune the video model with images. The keys lie in how to train a video model and how to fine-tune it with images.", + "bbox": [ + 496, + 688, + 893, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the study of the connection between spatial and temporal modules, a fully trained model is more suitable", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "iter $= 1K$", + "bbox": [ + 158, + 89, + 228, + 102 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "iter $= 10K$", + "bbox": [ + 331, + 89, + 408, + 102 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P-Temp-LORA", + "bbox": [ + 560, + 90, + 651, + 104 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F-Temp-LORA", + "bbox": [ + 736, + 90, + 826, + 104 + ], + "page_idx": 3 + }, + { + "type": "aside_text", + "text": "P-Spa-LORA", + "bbox": [ + 84, + 142, + 102, + 205 + ], + "page_idx": 3 + }, + { + "type": "aside_text", + "text": "F-Spa-LORA", + "bbox": [ + 84, + 282, + 102, + 347 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7313", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An astronaut is waving his hands on the moon.", + "text_level": 1, + "bbox": [ + 326, + 258, + 633, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5. Module selection based on the fully trained T2V model. Best viewed with Acrobat Reader. Click the images to play the video clips.", + "bbox": [ + 76, + 279, + 888, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anime illustration of a blue pig, riding a scooter near a lake, with the sun in the sky", + "bbox": [ + 81, + 311, + 411, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "F-Spa-DIR-LAION", + "text_level": 1, + "bbox": [ + 117, + 489, + 238, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "F-Spa-DIR", + "text_level": 1, + "bbox": [ + 330, + 489, + 398, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 6. Influence of image data on concept composition. 'F-Spa-DIR-LAION' uses the LAION aesthetics V2 as the image data while 'F-Spa-DIR' uses JDB. Best viewed with Acrobat Reader. Click the images to play the video clips.", + "bbox": [ + 76, + 506, + 470, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for the subsequent finetuning with high-quality images. This is because the strong spatial-temporal coupling can tolerate the parameter perturbation for both spatial and temporal modules without obvious motion degeneration.", + "bbox": [ + 76, + 575, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we need to investigate how to fine-tune the base model with images. In both spatial and temporal perturbation (Sec. 3.1), the picture quality can be improved but not very significantly. To obtain a greater quality improvement, we evaluate two strategies. One is to involve more parameters, i.e., finetuning both spatial and temporal modules with images. The other is to change the finetuning method, i.e., using direct finetuning without LORA. We can evaluate the following four cases:", + "bbox": [ + 76, + 636, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M _ {F} ^ {A} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}, \\theta_ {S}} ^ {\\text {L O R A}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right), \\\\ M _ {F} ^ {B} \\left(\\theta_ {T}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {S}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right), \\\\ M _ {F} ^ {C} (\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S}) \\leftarrow \\mathrm {P E R T B} _ {\\theta_ {T}} (M _ {F} (\\theta_ {T}, \\theta_ {S}), \\mathcal {D} _ {I}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 780, + 498, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {F} ^ {D} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}, \\theta_ {S}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 842, + 498, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $M_F^A$ (F-Spa&Temp-LORA) is obtained by following the first strategy, $M_F^B$ , $M_F^C$ , and $M_F^D$ are obtained using", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the second strategy. $M_F^B$ (F-Spa-DIR) and $M_F^C$ (F-TempDIR) represent directly finetuning the spatial and temporal modules, respectively. $M_F^D$ (F-Spa&Temp-DIR) represents directly finetuning all modules.", + "bbox": [ + 500, + 311, + 893, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing the generated videos of the four models, we have the following observations. First, F-Spa&Temp-LORA can further improve the picture quality of F-Spa-LORA, but the quality is still unsatisfying. The watermark exists in most generated videos, and the noise is obvious. Second, F-Temp-DIR achieves better picture quality than F-Temp-LORA. It is also better than F-Spa&Temp-LORA. The watermark is removed or lightened in half of the videos. Third, F-Spa-DIR and F-Spa&Temp-DIR achieve the best picture quality among the fine-tuned models. However, the motion of F-Spa-DIR is better (see Fig. 5 and Table 4). The foreground and background of F-Spa&Temp-DIR are flashing in videos generated by $M_F^D$ , especially local textures.", + "bbox": [ + 498, + 375, + 892, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By exploring the finetuning strategies and different modules, we identify that directly finetuning spatial modules with high-quality images is the best way to improve the picture quality without marginal loss of motion quality. As shown in Fig. 2, our data-level disentanglement pipeline can be summarized as follows: fully training a video model with low-quality videos first and then directly finetuning the spatial modules only with high-quality images.", + "bbox": [ + 498, + 573, + 890, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Promotion of Concept Composition", + "text_level": 1, + "bbox": [ + 500, + 709, + 805, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve the concept composition ability of video models, we propose to use synthesized images with complex concepts instead of using real images at the partial finetuning stage. The success of T2I models such as SDXL and Midjourney is built upon large-scale high-quality images. They have the ability to composite concepts that do not appear in the real world. Rather than using their training images, we propose transferring their concept composition ability to video models by synthesizing a set of images with complex concepts. In this way, we can alleviate the burden of capturing both concept and motion well at the same time.", + "bbox": [ + 498, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "F-Spa&Temp-LORA", + "bbox": [ + 125, + 89, + 254, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "F-Temp-DIR", + "bbox": [ + 344, + 89, + 424, + 103 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "F-Spa-DIR", + "bbox": [ + 545, + 89, + 614, + 103 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "F-Spa&Temp-DIR", + "bbox": [ + 718, + 89, + 831, + 104 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7314", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the effectiveness of synthesized images, we use JDB and LAION-aesthetics V2 [2] as image data for the second finetuning stage. LAION-aesthetics V2 consists of web-collected images while JDB contains images synthesized by Midjourney. We observe that the model trained with JDB has much better concept composition ability (see Fig. 6 and Table 3). More results are in the supplementary material.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 210, + 209, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 234, + 174, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data. To overcome data limitations, we utilized WebVid-10M [12] as the source of low-quality video data and JDB [35] for high-quality image data. WebVid-10M is a large-scale, diverse video dataset comprising approximately 10 million text-video pairs. The resolution of most videos is $336 \\times 596$ , and each video consists of a single shot. During training, we sample from videos at varying frame rates. JDB is a large-scale image dataset featuring around 4 million high-resolution images from Midjourney, each annotated with a corresponding text prompt. To prevent concept forgetting during the training of the base T2V model, we also employ LAION-COCO[3], a dataset comprising 600 million generated high-quality captions for publicly available web images, for both image and video training.", + "bbox": [ + 75, + 258, + 470, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We exploit EvalCrafter [31] for quantitative evaluation. EvalCrafter is a benchmark to evaluate text-to-video generation models, which contains around 18 objective metrics for visual quality, content quality, motion quality, and text-caption alignment. It provides about 512 prompts. The objective metrics are aligned to user opinions from five subjective studies, i.e., motion quality, text-video alignment, temporal consistency, visual quality, and user favor. The motion quality considers three metrics: action recognition, average flow, amplitude classification score, while temporal consistency considers warping error, semantic consistency, face consistency. The technical and aesthetic scores in EvalCrafter are adapted from DOVER [54]. Besides, we conduct user studies of human preference since there still lacks a comprehensive objective metric to measure motion quality.", + "bbox": [ + 75, + 488, + 470, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. In Sec 3.1, the two based models share the same architecture, adapted from the open-source VideoCrafter1 [15], and incorporate temporal convolution from ModelScopeT2V [48]. The spatial modules are initialized with weights from SD 2.1, and the outputs of the temporal modules are initialized to zeros. The training resolution is set at $512 \\times 320$ . For joint image and video training, we utilize the low-quality WebVid-10M and LAION-COCO datasets. The models are trained on 32 NVIDIA A100 GPUs for $270K$ iterations with a batch size of 128. The learning rate is set at $5 \\times 10^{-5}$ for all training tasks. When employing", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4d3f9dfadb7e42d749ef75ea59dbdd0c03e7badf7ce6446457a85dd6bc46508f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Visual QualityText-Video AlignmentMotion QualityTemporal Consistency
PikaLab*63.5254.1157.7469.35
Gen2*67.3552.3062.5369.71
VideoCrafter161.6466.7656.0660.36
Show-152.1962.0753.7460.83
AnimateDiff58.8974.7951.3856.61
Ours63.2864.6753.9562.02
", + "bbox": [ + 501, + 89, + 893, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison on the EvalCrafter benchmark. Higher score indicates better performance. * commercial models.", + "bbox": [ + 500, + 202, + 890, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LORA for the perturbation of temporal or spatial modules, we exclusively use JDB for tuning. The finetuning is conducted on 8 A100 GPUs for $30K$ iterations with a batch size of 256. Given that the images from JDB have a square resolution, we adjust the finetuning resolution to $512 \\times 512$ .", + "bbox": [ + 496, + 248, + 893, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art T2V Models", + "text_level": 1, + "bbox": [ + 498, + 335, + 890, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach with several state-of-the-art T2V models, including popular commercial models such as Gen-2 [5] and Pika Labs [9], as well as open-source models like Show-1 [60], VideoCrafter1 [15], andAnimateDiff [22]. Gen-2, Pika Labs, and VideoCrafter1 all utilize high-quality videos for training their T2V models. It is noteworthy thatAnimateDiff and our models use only the videos from WebVid-10M. Show-1 employs additional high-quality videos for finetuning to eliminate the watermark in WebVid-10M.AnimateDiff is not a generic T2V model; it works only when the LORA SD model is compatible with its temporal modules. For our comparison, we use its temporal modules (second version) based on SD v1.5 and employ Realistic Vision V2.0 [10] as its corresponding LORA model.", + "bbox": [ + 496, + 359, + 893, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation. The quantitative results obtained using EvalCrafter are presented in Table 1. Our method achieves visual quality comparable to that of VideoCrafter1 and Pika Labs, which use high-quality videos for training. This underscores the effectiveness of employing high-quality images to enhance picture quality and aesthetic scores. Furthermore, our text-video alignment performance is ranked second. In terms of motion quality, our performance surpasses that of Show-1 but falls short of models that utilize a larger volume of videos to learn motion. This indicates that our method can enhance visual quality without significant motion degradation.", + "bbox": [ + 496, + 592, + 893, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Evaluation. The visual comparison is depicted in Fig. 7. Additional results are provided in the supplementary material. The visual quality of our results is on par with that of commercial models such as Gen-2 and Pika Labs. Since we employ JDB as the image dataset, the picture quality of our synthesized videos shifts from WebVid-10M to JDB. Regarding motion, our motion quality is superior to", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df1cde7bcaf23e87514f277b58a92271c01105c69a01a49cbbe5408142ef5875.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
A bear rummages through a\ndumpster, searching for food\nscapsA Time 1980 painting of a boy\ngoing to schoolA man cruises through the city\non a motorcycle, feeling the\nadrenaline rushA monkey eating a pizza in\ncentral park, GoPro film style
Gen-2
Show-1VideoCrafter1Pika Labs
Ours
", + "bbox": [ + 84, + 87, + 901, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7. Comparison of different text-to-video generation models. Best viewed with Acrobat Reader. Click the images to play the video clips.", + "bbox": [ + 76, + 813, + 890, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that ofAnimateDiff and comparable to Show-1. Although the integration of temporal modules with a LORA SD model can enhance visual quality,AnimateDiff experiences motion", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "degradation in generic scenes.", + "bbox": [ + 500, + 854, + 700, + 869 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7316", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4ced48ca6d5a96b7cb4bf8846060d9da8a30482b0dee4eb2b6b16cba12246d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsText-Video AlignmentMotion QualityVisual Quality
Ours vs Gen256%46%34%
Ours vsAnimateDiff55%64%69%
Ours vs Show-159%59%82%
Ours vs VideoCrafter161%63%61%
", + "bbox": [ + 78, + 88, + 467, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. For further evaluation, we conduct a user study to compare our method with other video models. We select 50 prompts from EvalCrafter, covering diverse scenes, styles, and objects. When comparing a model pair, three video production experts are asked to select their preferred video from three options: method 1, method 2, and comparable results, according to the given subject, i.e., visual quality, motion quality, and text-video alignment. The results are shown in Table 4. Our method has better visual quality thanAnimateDiff and Show-1 and is comparable to VideoCrafter1. Our method is more preferred than Show-1 andAnimateDiff in motion quality.", + "bbox": [ + 75, + 239, + 470, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Strategy Evaluation", + "text_level": 1, + "bbox": [ + 76, + 431, + 264, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial-temporal Connection. In Sec. 3.1, we show a visual comparison of perturbing the spatial and temporal parameters of the fully and partially trained models in Fig. 3 and Fig. 4. Here we provide the quantitative comparisons about the visual quality in Table 3, including aesthetic and technical scores from DOVER [54]. We observe that fin-tuning the partially trained model can always achieve better visual quality than the fully trained model. It means that the distribution of the partially trained model can be shifted more easily. Besides, we conduct a user study asking participants to choose a favorable model that has better performance in motion, in terms of foreground/background flash and motion flicker. The results are shown in Table 4. It can be observed that the motion quality of perturbed fully trained models is better. The fully trained model can tolerate larger parameter perturbations than the partially trained model. These observations show that the fully trained model has stronger spatial-temporal coupling.", + "bbox": [ + 75, + 455, + 470, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Module Selection. After selecting the fully trained model as the base, we use two strategies to identify the most effective module to fine-tune, resulting in four models in Sec. 3.2. The visual quality evaluation of these models is shown in the bottom part of Table 3. The visual quality of F-Spa-DIR and F-Spa&Temp-DIR is much better than the other two models. It reveals that directly finetuning spatial modules is the key to improving picture quality.", + "bbox": [ + 75, + 748, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since F-Spa-DIR and F-Spa&Temp-DIR achieve close visual quality, we conduct a user study on motion quality to", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/11ce327d197de53627bf25823592d484f98dda11123ab8d8d760a110f932650a.jpg", + "table_caption": [ + "Table 2. Human preference. The numbers represent the probability of users choosing our method." + ], + "table_footnote": [], + "table_body": "
MethodAesthetic Score (↑)Technical Score (↑)
P-base34.3242.69
F-base46.5551.76
P-Spa-LORA78.2572.74
F-Spa-LORA77.9759.60
P-Temp-LORA77.4054.85
F-Temp-LORA66.2650.32
F-Spa-DIR82.5770.35
F-Temp-DIR82.7765.34
F-Spa&Temp-DIR83.5967.75
F-Spa&Temp-LORA80.4463.61
F-Spa-DIR-LAION67.8354.26
", + "bbox": [ + 506, + 88, + 883, + 255 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d61f5599b06ca963e4e7db9be2dc7f8e9e956b67facaa26d04cc6f07154d43a7.jpg", + "table_caption": [ + "Table 3. Visual quality evaluation of the perturbed T2V models." + ], + "table_footnote": [], + "table_body": "
MethodsMotion Quality
F-Spa-LORA vs P-Spa-LORA87%
F-Temp-LORA vs P-Temp-LORA73%
F-Spa-DIR vs F-Spa&Temp-DIR67%
", + "bbox": [ + 506, + 289, + 883, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. User study on the motion of the perturbed T2V models.", + "bbox": [ + 504, + 364, + 883, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "determine the final model. The results are shown in the last row of Table 4. Directly finetuning the spatial modules only performs better in motion. As shown in Fig. 5, F-Spa-DIR is more stable and has better temporal consistency than F-Spa&Temp-DIR. The latter has obvious flashes in both the foreground and background.", + "bbox": [ + 496, + 402, + 893, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Image Data. To verify the effectiveness of synthesized images, we use the LAION Aesthetics V2 dataset and JDB to directly fine-tune the spatial modules in the second stage, respectively. The visual examples are shown in Fig. 6. It shows that the model trained with JDB composite concepts better than the model trained with LAION Aesthetics V2. The quantitative evaluation of visual quality is shown in Table 3. F-Spa-DIR is much better than F-Spa-DIR-LAION in both aesthetic and technical scores.", + "bbox": [ + 496, + 512, + 892, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To overcome data limitations, we propose a method for training high-quality video diffusion models without using high-quality videos. We delve into the training schemes of SD-based video models and investigate the coupling strength between spatial and temporal dimensions. We observe that fully trained T2V models exhibit stronger spatial-temporal coupling than partially trained models. Based on this observation, we propose disentangling appearance from motion at the data level, i.e., by exploiting low-quality videos for motion learning and high-quality images for appearance learning. Additionally, we suggest using synthetic images with complex concepts for finetuning, rather than real images. Quantitative and qualitative evaluations are conducted to demonstrate the effectiveness of the proposed method.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] If. Accessed October 22, 2023 [Online] https://github.com/deep-floyd/IF.2", + "[2] Laion-aesthetics. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-aesthetics/, .6", + "[3] Laion-coco. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-coco/, .3,6", + "[4] Midjourney. Accessed October 22, 2023 [Online] https://www.midjourney.com/home.3", + "[5] Gen-2. Accessed October 22, 2023 [Online] https://research.runwayml.com/gen2,.2,6", + "[6] Genmo. Accessed October 22, 2023 [Online] https://www.genmo.ai/, .2", + "[7] Hotshot-x1. Accessed October 22, 2023 [Online] https://github.com/hotshotco/Hotshot-XL.2", + "[8] Moonvalley. Accessed October 22, 2023 [Online] https://moonvalley.ai/.2", + "[9] Pika labs. Accessed October 22, 2023 [Online] https://www.pika.art/.2,6", + "[10] Realistic vision v2.0. Accessed October 22, 2023 [Online] https://huggingface.co/ckpt/realistic-vision-v20.6", + "[11] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2", + "[12] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3, 6", + "[13] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[14] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 3", + "[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 6", + "[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2", + "[17] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, 35(1):53-65, 2018. 2", + "[18] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, pages 102–118. Springer, 2022. 2", + "[20] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2", + "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2", + "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3, 6", + "[23] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 2, 3", + "[24] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. arXiv preprint arXiv:2310.07702, 2023. 2", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2", + "[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 2", + "[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 3", + "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[30] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 1, 2", + "[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models, 2023. 6", + "[32] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In AAAI, pages 9117-9125, 2023. 2", + "[33] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7318", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2", + "[35] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 3, 6", + "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2", + "[40] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, pages 2830-2839, 2017. 2", + "[41] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 2", + "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2", + "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2", + "[44] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2", + "[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2", + "[46] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 2", + "[47] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 29, 2016. 2", + "[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2", + "[50] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3", + "[51] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2", + "[52] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2", + "[53] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. arXiv preprint arXiv:2207.09814, 2022. 2", + "[54] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In CVPR, pages 20144-20154, 2023. 6, 8", + "[55] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2", + "[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 2", + "[57] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2", + "[58] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2", + "[59] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In CVPR, pages 18456-18466, 2023. 2", + "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation, 2023. 2, 6", + "[61] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 2", + "[62] Yuechen Zhang, Jinbo Xing, Eric Lo, and Jiaya Jia. Realworld image variation by aligning diffusion inversion chain. arXiv preprint arXiv:2305.18729, 2023. 2" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7319", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[63] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1, 2, 3", + "bbox": [ + 78, + 90, + 470, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7320", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_model.json b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..345ff251f037a6b544ff885a43f340336ba17ab9 --- /dev/null +++ b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.13, + 0.804, + 0.175 + ], + "angle": 0, + "content": "VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.204, + 0.314, + 0.221 + ], + "angle": 0, + "content": "Haoxin Chen" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.205, + 0.46, + 0.222 + ], + "angle": 0, + "content": "Yong Zhang*" + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.205, + 0.612, + 0.222 + ], + "angle": 0, + "content": "Xiaodong Cun" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.205, + 0.764, + 0.222 + ], + "angle": 0, + "content": "Menghan Xia" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.223, + 0.408, + 0.24 + ], + "angle": 0, + "content": "Xintao Wang" + }, + { + "type": "text", + "bbox": [ + 0.449, + 0.223, + 0.543, + 0.24 + ], + "angle": 0, + "content": "Chao Weng" + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.224, + 0.668, + 0.24 + ], + "angle": 0, + "content": "Ying Shan" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.257, + 0.547, + 0.273 + ], + "angle": 0, + "content": "Tencent AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.53, + 0.275, + 0.544 + ], + "angle": 0, + "content": "In cyberpunk, neonpunk style," + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.545, + 0.288, + 0.557 + ], + "angle": 0, + "content": "Kung Fu Panda, jump and kick." + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.516, + 0.576, + 0.53 + ], + "angle": 0, + "content": "Cinematic photo melting pistachio ice" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.531, + 0.537, + 0.543 + ], + "angle": 0, + "content": "cream dripping down the cone." + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.544, + 0.538, + 0.557 + ], + "angle": 0, + "content": "35mm photograph, film, bokeh." + }, + { + "type": "text", + "bbox": [ + 0.626, + 0.53, + 0.866, + 0.544 + ], + "angle": 0, + "content": "Large motion, surrounded by butterflies," + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.545, + 0.839, + 0.557 + ], + "angle": 0, + "content": "a girl walks through a lush garden." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.568, + 0.895, + 0.611 + ], + "angle": 0, + "content": "Figure 1. Give a text prompt, our method can generate a video with high visual quality and accurate text-video alignment. Note that it is trained with only low-quality videos and high-quality images. No high-quality videos are required. Best viewed with Acrobat Reader. Click the images to play the video clips." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.623, + 0.314, + 0.639 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Text-to-video generation aims to produce a video based on a given prompt. Recently, several commercial video models have been able to generate plausible videos with minimal noise, excellent details, and high aesthetic scores. However, these models rely on large-scale, well-filtered, high-quality videos that are not accessible to the community. Many existing research works, which train models using the low-quality WebVid-10M dataset, struggle to generate high-quality videos because the models are optimized to fit WebVid-10M. In this work, we explore the training scheme of video models extended from Stable Diffusion and investigate the feasibility of leveraging low-quality videos and synthesized high-quality images to obtain a high-quality video model. We first analyze the connection between the spatial" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.625, + 0.896, + 0.777 + ], + "angle": 0, + "content": "and temporal modules of video models and the distribution shift to low-quality videos. We observe that full training of all modules results in a stronger coupling between spatial and temporal modules than only training temporal modules. Based on this stronger coupling, we shift the distribution to higher quality without motion degradation by finetuning spatial modules with high-quality images, resulting in a generic high-quality video model. Evaluations are conducted to demonstrate the superiority of the proposed method, particularly in picture quality, motion, and concept composition." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.813, + 0.633, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Benefiting from the development of diffusion models [25, 43], video generation has achieved breakthroughs, particularly in basic text-to-video (T2V) generation models. Most existing methods [14, 23, 26, 30, 48, 63] follow a logic to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.209, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7310" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "obtain video models, i.e., extending a text-to-image (T2I) backbone to a video model by adding temporal modules and then training it with videos. Several models train video models from scratch, while most start from a pre-trained T2I model, typically Stable Diffusion (SD) [38]. Models can also be categorized into two groups based on the space modeled by diffusion models, i.e., pixel-space models [26, 30, 60] and latent-space models [14, 23, 48, 63]. The latter is the dominant approach. Picture quality, motion consistency, and concept composition are essential dimensions for evaluating a video model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.549 + ], + "angle": 0, + "content": "Recently, a few commercial startups have released their T2V models [5, 6, 8, 9] that can produce plausible videos with minimal noise, excellent details, and high aesthetic scores. However, they are trained on a large-scale and well-filtered high-quality video dataset, which is not accessible to the community and academia. Collecting millions of high-quality videos is challenging due to copyright restrictions and post-filtering processing. Though there are a few open-source video datasets collected from the Internet for video understanding, such as HowTo100M [33], HD-VILA100M [56], and InterVid [51], there exist many issues for video generation, e.g., poor picture quality and caption, multiple clips in one video, and static frames or slides. WebVid10M [12] is the most widely used dataset to train video generation models in academia. The clips are well-segmented, and the diversity is good. However, the picture quality is unsatisfactory, and most videos have a resolution of about \\(320\\mathrm{p}\\). The lack of high-quality datasets poses a significant obstacle to training high-quality video models in academia." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this work, we target a quite challenging problem, i.e., training high-quality video models without using high-quality videos. We dive into the training process of SD-based video models to analyze the connection between spatial and temporal modules under different training strategies and investigate the distribution shift to low-quality videos. We make a meaningful observation that the full training of all modules results in a stronger coupling between appearance and motion than just training temporal modules. The full training can achieve more natural motion and tolerate more subsequent modification of spatial modules, which is the key to improving the quality of generated videos. Based on the observation of the connection, we propose a method to overcome the data limitation by disentangling motion from appearance at the data level. Specifically, instead of high-quality videos, we exploit low-quality videos to guarantee motion consistency and use high-quality images to ensure picture quality and concept composition ability. Benefiting from the successful T2I models such as SDXL and Midjourney, it is convenient to obtain a large set of images with high-resolution and complex concept composition. As shown in Fig. 2, we design a pipeline to train a video diffusion model following the guidelines from the analysis. Fig. 1" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.824, + 0.106 + ], + "angle": 0, + "content": "shows visual examples generated by our method." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.107, + 0.858, + 0.121 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.167 + ], + "angle": 0, + "content": "- We propose a method to overcome the data for training high-quality video models by disentangling motion from appearance at the data level." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.894, + 0.213 + ], + "angle": 0, + "content": "- We investigate the connection between spatial and temporal modules, and the distribution shift. We identify the keys to obtain a high-quality video model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.213, + 0.894, + 0.258 + ], + "angle": 0, + "content": "- We design an effective pipeline based on the observations, i.e., obtaining a fully trained video model first and tuning the spatial modules with synthesized high-quality images." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.272, + 0.642, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.447 + ], + "angle": 0, + "content": "The evolution of video generation techniques goes along with the development of generative models. Generative adversarial networks [17] and variational auto-encoders [18] are the commonly used backbones in early research of video generation, e.g., VGAN [47], TGAN [40], MoCoGAN [44], GODIA [52], StyleGAN-V [41], and MCVD [46]. Then, since transformers have been successfully applied in various fields, they are also introduced for video synthesis, e.g.,CogVideo [28], VideoGPT [57], NUVA-infinity [53], TATS [19], MAGVIT [58], Phenaki [45]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.897, + 0.765 + ], + "angle": 0, + "content": "Recently, diffusion models (DMs) [25, 42, 43] have been a famous star in generative models, especially in text-to-image (T2I) generation [13, 21, 24, 34, 36-39, 62]. For video generation, Video Diffusion Models (VDMs) are proposed to model the distribution of videos. VDM [27] is the first to utilize a space-time factorized U-Net to model videos in pixel space for unconditional video generation. It uses an image-video joint training strategy to avoid concept forgetting.Imagen Video [26] and Make-a-Video [30] are two cascade models that target text-to-video generation in pixel space. Show-1 [60] is another cascade model that uses IF [1] as the base model and LDM extended video model for super-resolution. Subsequently, LVDM [15, 23] and MagicVideo [63] propose to extend LDM [38] to model videos in the latent space of an auto-encoder. Many other methods use the same paradigm, including ModelScope [48], Align Your Latent [14], Hotshot-XL [7], LAVIE [50], PY-OCO [20], VideoFactory [49], VPDM [59], VIDM [32], and Latent-Shift [11]. Besides text-to-video generation, a few methods, such as [16, 55, 61], attempt to generate videos from a given image and a prompt as condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Several startups release their text-to-video generation services, e.g., Gen-2 [5], Pika Labs [9], Moonvalley [8], and Genmo [6]. Their models can generate plausible results with minimal noise, excellent details, and high aesthetic scores. However, those methods are trained with a large-scale well-filtered high-quality video dataset that is not accessible to researchers. The video models are also not available, leading to the slow development of downstream tasks to a certain extent. The most widely used video dataset is WebVid-10M," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7311" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 2. The efficient training pipeline. In the first stage, we use a large amount of low-quality videos to fully train the VDM to learn motion. In the second stage, we only train the spatial parameters with the high-quality images generated by the T2I model to improve the concept composition and picture quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.471, + 0.412 + ], + "angle": 0, + "content": "a large-scale dataset of short videos with textual descriptions sourced from stock footage sites. The videos are diverse and rich in their content, and each video is well-segmented, however, the picture quality is unsatisfactory and most videos are \\(320\\mathrm{p}\\). Training a high-quality video model under the data limitation is quite challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.474, + 0.611 + ], + "angle": 0, + "content": "AnimateDiff [22] finds that combining temporal modules from a video model trained on WebVid-10M and a LORA SD model can improve the picture quality of the generated videos. However, this is not a generic model and does not always work. There are a few severe issues. First, the temporal modules can only be combined with a few selected LORA models, which makes it not a generic model. Second, the motion quality degenerates when the modules do not match the LORA model well. UnlikeAnimateDiff, we analyze the connection between spatial and temporal modules instead of direct combination, and design a pipeline to train a generic high-quality video model without high-quality video by disentangling appearance and motion at the data level." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.626, + 0.169, + 0.642 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.472, + 0.775 + ], + "angle": 0, + "content": "We propose an effective method to overcome the data limitation for training high-quality video diffusion models. We first analyze the connection between spatial and temporal modules of SD-based video models under different training strategies. Based on the observations, we then develop a pipeline to train high-quality video models with just low-quality videos and high-quality images, i.e., disentangling appearance from motion at the data level." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.409, + 0.803 + ], + "angle": 0, + "content": "3.1. Spatial-temporal Connection Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Base T2V model. To leverage the prior in SD trained on a large-scale image dataset, most text-to-video diffusion models inflate the SD model to a video model by adding temporal modules, including Align Your Latent [14],AnimateDiff [22], LVDM [23], Magic Video [63], ModelScopeT2V [48], and LAVIE [50]. They follow VDM [27]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "to use a particular type of 3D-UNet that is factorized over space and time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.212 + ], + "angle": 0, + "content": "These models can be categorized into two groups according to their training strategies. One is to use videos to train temporal modules with spatial ones fixed, called partial training. The other is to learn both the spatial and temporal modules with the SD weights as initialization, called full training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Though these SD-based T2V models have similar architectures, they are trained under different training settings. We use one typical model to investigate the connection between spatial and temporal modules under the two training strategies. We follow the architecture of the open-source VideoCrafter1 [15] with FPS (frames per second) condition. We also incorporate the temporal convolution in ModelScopeT2V [48] to improve temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.518 + ], + "angle": 0, + "content": "Parameter Perturbation for Full and Partial Training. We apply the two training strategies to the same architecture using the same data. The model is initialized from pretrained SD weights. WebVid-10M [12] is exploited as the training data. To avoid concept forgetting, LAION-COCO 600M [3] is also used for video and image joint training. The resolution is \\(512 \\times 320\\). For simplicity, the fully trained video model is denoted as \\(M_F(\\theta_T, \\theta_S)\\), while the partially trained one is denoted as \\(M_P(\\theta_T, \\theta_S^0)\\), where \\(\\theta_T\\) and \\(\\theta_S\\) are the learned parameters of the temporal and spatial modules, respectively. \\(\\theta_S^0\\) are the original spatial parameters of SD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.608 + ], + "angle": 0, + "content": "To evaluate the connection strength between spatial and temporal modules, we perturb the parameters of the specified modules by using another high-quality image dataset \\(\\mathcal{D}_I\\) for finetuning. The image data is JDB [35] that consists of synthesized images from Midjourney [4]. As the JDB has 4 million images, we use LORA [29] for finetuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.894, + 0.669 + ], + "angle": 0, + "content": "Spatial Perturbation. We first perturb the spatial parameters of the two video models using the image dataset. The temporal parameters are frozen. The perturbation process of the fully trained base model \\( M_F \\) can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.679, + 0.88, + 0.699 + ], + "angle": 0, + "content": "\\[\nM _ {F} ^ {\\prime} (\\theta_ {T}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}) \\leftarrow \\mathrm {P E R T B} _ {\\theta_ {S}} ^ {\\mathrm {L O R A}} (M _ {F} (\\theta_ {T}, \\theta_ {S}), \\mathcal {D} _ {I}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.772 + ], + "angle": 0, + "content": "where \\(\\mathrm{PERTB}_{\\theta_S}^{\\mathrm{LORA}}\\) denote finetuning \\(M_F\\) with respect to \\(\\theta_{S}\\) on the image dataset \\(\\mathcal{D}_I\\) using LORA. \\(\\Delta_{\\theta_S}\\) represents the parameters of the LORA branch. Similarly, we can obtain the perturbed model of the partially trained video model:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.781, + 0.88, + 0.801 + ], + "angle": 0, + "content": "\\[\nM _ {P} ^ {\\prime} \\left(\\theta_ {T}, \\theta_ {S} ^ {0} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {S}} ^ {\\text {L O R A}} \\left(M _ {P} \\left(\\theta_ {T}, \\theta_ {S} ^ {0}\\right), \\mathcal {D} _ {I}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "For easy understanding, we also use the name 'F-SpALORA' to denote model \\(M_F^{\\prime}\\) and 'P-Spa-LORA' for \\(M_P^\\prime\\) . 'F' denotes the fully trained base model while \\(\\mathbf{P}^{\\prime}\\) stands for the partially trained model. 'Spa' and Temp' mean finetuning spatial and temporal modules, respectively. 'LORA' represents using LORA for finetuning, while DIR' means direct" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7312" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.159, + 0.09, + 0.23, + 0.103 + ], + "angle": 0, + "content": "iter \\(= 1K\\)" + }, + { + "type": "header", + "bbox": [ + 0.332, + 0.09, + 0.41, + 0.103 + ], + "angle": 0, + "content": "iter \\(= 10K\\)" + }, + { + "type": "header", + "bbox": [ + 0.561, + 0.091, + 0.652, + 0.105 + ], + "angle": 0, + "content": "P-Temp-LORA" + }, + { + "type": "header", + "bbox": [ + 0.738, + 0.091, + 0.828, + 0.105 + ], + "angle": 0, + "content": "F-Temp-LORA" + }, + { + "type": "aside_text", + "bbox": [ + 0.086, + 0.143, + 0.104, + 0.207 + ], + "angle": 270, + "content": "P-Spa-LORA" + }, + { + "type": "aside_text", + "bbox": [ + 0.086, + 0.283, + 0.104, + 0.348 + ], + "angle": 270, + "content": "F-Spa-LORA" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.387, + 0.472, + 0.433 + ], + "angle": 0, + "content": "A professional dancer gracefully performs ballet on stage. \nFigure 3. Perturbing spatial modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.448, + 0.47, + 0.492 + ], + "angle": 0, + "content": "finetuning without LORA. For example, 'F-Spatial-LORA' represents perturbing spatial modules of the fully trained T2V model using LORA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.471, + 0.719 + ], + "angle": 0, + "content": "Comparing the synthesized videos of the two resulting models, we have the following observations. First, the motion quality of F-Spa-LORA is more stable than that of P-Spa-LORA (see user study in Table 4). The motion of P-Spa-LORA becomes worse quickly during the finetuning process. The more finetuning steps, the video tends to be more still with local flicker (see Fig. 3). While the motion of F-Spa-LORA slightly degenerates compared to the fully trained base model. Second, P-Spa-LORA achieves much better visual quality than F-Spa-LORA (see Fig. 3). The picture quality and aesthetic score of F-Spa-LORA are greatly improved compared to the partially trained base model (see Table 3). Surprisingly, the watermark is also removed. While F-Spa-LORA obtains a slight improvement in picture quality and aesthetic score, the generated videos are still noisy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.472, + 0.902 + ], + "angle": 0, + "content": "From the two observations, we can conclude that the coupling strength between spatial and temporal modules of the fully trained model is stronger than that of the partially trained model. Because the spatial-temporal coupling of the partially trained model can be easily broken, leading to quick motion degeneration and picture quality shift. A stronger connection can tolerate parameter perturbation more than a weak one. Our observation can be used to explain the quality improvement and motion degeneration ofAnimateDiff.AnimateDiff is not a generic model and only works for selected personalized SD models. The reason is that its motion modules are obtained with the partially training strategy, and" + }, + { + "type": "title", + "bbox": [ + 0.601, + 0.247, + 0.788, + 0.26 + ], + "angle": 0, + "content": "Robot dancing in times square." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.262, + 0.894, + 0.291 + ], + "angle": 0, + "content": "Figure 4. Perturbing temporal modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.893, + 0.351 + ], + "angle": 0, + "content": "they cannot tolerate large parameter perturbations. When the personalized model does not match the temporal modules, both picture and motion quality will degenerate." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.894, + 0.457 + ], + "angle": 0, + "content": "Temporal Perturbation. The partially trained model has only the temporal modules updated, but the picture quality is shifted to the quality of WebVid-10M. Hence, the temporal modules take responsibility for not only the motion but also the picture quality. We perturb the temporal modules while fixing the spatial modules with the image dataset. The perturbation processes can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.466, + 0.88, + 0.487 + ], + "angle": 0, + "content": "\\[\nM _ {F} ^ {\\prime \\prime} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}} ^ {\\text {L O R A}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.489, + 0.88, + 0.509 + ], + "angle": 0, + "content": "\\[\nM _ {P} ^ {\\prime \\prime} (\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} ^ {0}) \\gets \\mathrm {P E R T B} _ {\\theta_ {T}} ^ {\\mathrm {L O R A}} (M _ {P} (\\theta_ {T}, \\theta_ {S} ^ {0}), \\mathcal {D} _ {I}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.519, + 0.895, + 0.644 + ], + "angle": 0, + "content": "We observe that the picture quality of P-Temp-LORA \\((M_P^{\\prime \\prime})\\) is better than F-Temp-LORA \\((M_F^{\\prime \\prime})\\). However, the foreground and background of the videos are more shaky, i.e., the temporal consistency becomes worse (see Fig. 4). The picture of F-Temp-LORA is improved, but the watermark is still there. Its motion is close to the base model and much better than P-Temp-LORA (see Table 4). Those observations also support the conclusion obtained from spatial perturbation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.892, + 0.681 + ], + "angle": 0, + "content": "3.2. Data-level Disentanglement of Appearance and Motion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Since obtaining a large-scale, high-quality video dataset with high diversity is challenging due to copyright issues, we explore the possibility of training a high-quality video model without using high-quality videos. We have access to low-quality videos such as WebVid-10M and high-quality images such as JDB. We propose to disentangle motion from appearance at the data level, i.e., learning motion from low-quality videos while learning picture quality and aesthetics from high-quality images. We can first train a video model with videos and then fine-tune the video model with images. The keys lie in how to train a video model and how to fine-tune it with images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "According to the study of the connection between spatial and temporal modules, a fully trained model is more suitable" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7313" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.127, + 0.09, + 0.256, + 0.105 + ], + "angle": 0, + "content": "F-Spa&Temp-LORA" + }, + { + "type": "header", + "bbox": [ + 0.346, + 0.09, + 0.426, + 0.104 + ], + "angle": 0, + "content": "F-Temp-DIR" + }, + { + "type": "header", + "bbox": [ + 0.546, + 0.09, + 0.615, + 0.104 + ], + "angle": 0, + "content": "F-Spa-DIR" + }, + { + "type": "header", + "bbox": [ + 0.719, + 0.09, + 0.832, + 0.105 + ], + "angle": 0, + "content": "F-Spa&Temp-DIR" + }, + { + "type": "title", + "bbox": [ + 0.328, + 0.26, + 0.635, + 0.274 + ], + "angle": 0, + "content": "An astronaut is waving his hands on the moon." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.28, + 0.89, + 0.294 + ], + "angle": 0, + "content": "Figure 5. Module selection based on the fully trained T2V model. Best viewed with Acrobat Reader. Click the images to play the video clips." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.313, + 0.412, + 0.342 + ], + "angle": 0, + "content": "Anime illustration of a blue pig, riding a scooter near a lake, with the sun in the sky" + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.491, + 0.239, + 0.504 + ], + "angle": 0, + "content": "F-Spa-DIR-LAION" + }, + { + "type": "title", + "bbox": [ + 0.331, + 0.491, + 0.399, + 0.504 + ], + "angle": 0, + "content": "F-Spa-DIR" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.471, + 0.563 + ], + "angle": 0, + "content": "Figure 6. Influence of image data on concept composition. 'F-Spa-DIR-LAION' uses the LAION aesthetics V2 as the image data while 'F-Spa-DIR' uses JDB. Best viewed with Acrobat Reader. Click the images to play the video clips." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.469, + 0.636 + ], + "angle": 0, + "content": "for the subsequent finetuning with high-quality images. This is because the strong spatial-temporal coupling can tolerate the parameter perturbation for both spatial and temporal modules without obvious motion degeneration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.637, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Next, we need to investigate how to fine-tune the base model with images. In both spatial and temporal perturbation (Sec. 3.1), the picture quality can be improved but not very significantly. To obtain a greater quality improvement, we evaluate two strategies. One is to involve more parameters, i.e., finetuning both spatial and temporal modules with images. The other is to change the finetuning method, i.e., using direct finetuning without LORA. We can evaluate the following four cases:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.781, + 0.499, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M _ {F} ^ {A} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}, \\theta_ {S}} ^ {\\text {L O R A}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right), \\\\ M _ {F} ^ {B} \\left(\\theta_ {T}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {S}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right), \\\\ M _ {F} ^ {C} (\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S}) \\leftarrow \\mathrm {P E R T B} _ {\\theta_ {T}} (M _ {F} (\\theta_ {T}, \\theta_ {S}), \\mathcal {D} _ {I}), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.843, + 0.499, + 0.862 + ], + "angle": 0, + "content": "\\[\nM _ {F} ^ {D} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}, \\theta_ {S}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Where \\(M_F^A\\) (F-Spa&Temp-LORA) is obtained by following the first strategy, \\(M_F^B\\), \\(M_F^C\\), and \\(M_F^D\\) are obtained using" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.313, + 0.894, + 0.374 + ], + "angle": 0, + "content": "the second strategy. \\( M_F^B \\) (F-Spa-DIR) and \\( M_F^C \\) (F-TempDIR) represent directly finetuning the spatial and temporal modules, respectively. \\( M_F^D \\) (F-Spa&Temp-DIR) represents directly finetuning all modules." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.376, + 0.893, + 0.573 + ], + "angle": 0, + "content": "Comparing the generated videos of the four models, we have the following observations. First, F-Spa&Temp-LORA can further improve the picture quality of F-Spa-LORA, but the quality is still unsatisfying. The watermark exists in most generated videos, and the noise is obvious. Second, F-Temp-DIR achieves better picture quality than F-Temp-LORA. It is also better than F-Spa&Temp-LORA. The watermark is removed or lightened in half of the videos. Third, F-Spa-DIR and F-Spa&Temp-DIR achieve the best picture quality among the fine-tuned models. However, the motion of F-Spa-DIR is better (see Fig. 5 and Table 4). The foreground and background of F-Spa&Temp-DIR are flashing in videos generated by \\( M_F^D \\), especially local textures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.574, + 0.892, + 0.696 + ], + "angle": 0, + "content": "By exploring the finetuning strategies and different modules, we identify that directly finetuning spatial modules with high-quality images is the best way to improve the picture quality without marginal loss of motion quality. As shown in Fig. 2, our data-level disentanglement pipeline can be summarized as follows: fully training a video model with low-quality videos first and then directly finetuning the spatial modules only with high-quality images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.807, + 0.726 + ], + "angle": 0, + "content": "3.3. Promotion of Concept Composition" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To improve the concept composition ability of video models, we propose to use synthesized images with complex concepts instead of using real images at the partial finetuning stage. The success of T2I models such as SDXL and Midjourney is built upon large-scale high-quality images. They have the ability to composite concepts that do not appear in the real world. Rather than using their training images, we propose transferring their concept composition ability to video models by synthesizing a set of images with complex concepts. In this way, we can alleviate the burden of capturing both concept and motion well at the same time." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.472, + 0.198 + ], + "angle": 0, + "content": "To validate the effectiveness of synthesized images, we use JDB and LAION-aesthetics V2 [2] as image data for the second finetuning stage. LAION-aesthetics V2 consists of web-collected images while JDB contains images synthesized by Midjourney. We observe that the model trained with JDB has much better concept composition ability (see Fig. 6 and Table 3). More results are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.211, + 0.21, + 0.228 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.236, + 0.176, + 0.252 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.472, + 0.471 + ], + "angle": 0, + "content": "Data. To overcome data limitations, we utilized WebVid-10M [12] as the source of low-quality video data and JDB [35] for high-quality image data. WebVid-10M is a large-scale, diverse video dataset comprising approximately 10 million text-video pairs. The resolution of most videos is \\(336 \\times 596\\), and each video consists of a single shot. During training, we sample from videos at varying frame rates. JDB is a large-scale image dataset featuring around 4 million high-resolution images from Midjourney, each annotated with a corresponding text prompt. To prevent concept forgetting during the training of the base T2V model, we also employ LAION-COCO[3], a dataset comprising 600 million generated high-quality captions for publicly available web images, for both image and video training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.472, + 0.718 + ], + "angle": 0, + "content": "Metrics. We exploit EvalCrafter [31] for quantitative evaluation. EvalCrafter is a benchmark to evaluate text-to-video generation models, which contains around 18 objective metrics for visual quality, content quality, motion quality, and text-caption alignment. It provides about 512 prompts. The objective metrics are aligned to user opinions from five subjective studies, i.e., motion quality, text-video alignment, temporal consistency, visual quality, and user favor. The motion quality considers three metrics: action recognition, average flow, amplitude classification score, while temporal consistency considers warping error, semantic consistency, face consistency. The technical and aesthetic scores in EvalCrafter are adapted from DOVER [54]. Besides, we conduct user studies of human preference since there still lacks a comprehensive objective metric to measure motion quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Training Details. In Sec 3.1, the two based models share the same architecture, adapted from the open-source VideoCrafter1 [15], and incorporate temporal convolution from ModelScopeT2V [48]. The spatial modules are initialized with weights from SD 2.1, and the outputs of the temporal modules are initialized to zeros. The training resolution is set at \\(512 \\times 320\\). For joint image and video training, we utilize the low-quality WebVid-10M and LAION-COCO datasets. The models are trained on 32 NVIDIA A100 GPUs for \\(270K\\) iterations with a batch size of 128. The learning rate is set at \\(5 \\times 10^{-5}\\) for all training tasks. When employing" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.895, + 0.201 + ], + "angle": 0, + "content": "
Visual QualityText-Video AlignmentMotion QualityTemporal Consistency
PikaLab*63.5254.1157.7469.35
Gen2*67.3552.3062.5369.71
VideoCrafter161.6466.7656.0660.36
Show-152.1962.0753.7460.83
AnimateDiff58.8974.7951.3856.61
Ours63.2864.6753.9562.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.203, + 0.892, + 0.231 + ], + "angle": 0, + "content": "Table 1. Comparison on the EvalCrafter benchmark. Higher score indicates better performance. * commercial models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.895, + 0.326 + ], + "angle": 0, + "content": "LORA for the perturbation of temporal or spatial modules, we exclusively use JDB for tuning. The finetuning is conducted on 8 A100 GPUs for \\(30K\\) iterations with a batch size of 256. Given that the images from JDB have a square resolution, we adjust the finetuning resolution to \\(512 \\times 512\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.352 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art T2V Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.571 + ], + "angle": 0, + "content": "We compare our approach with several state-of-the-art T2V models, including popular commercial models such as Gen-2 [5] and Pika Labs [9], as well as open-source models like Show-1 [60], VideoCrafter1 [15], andAnimateDiff [22]. Gen-2, Pika Labs, and VideoCrafter1 all utilize high-quality videos for training their T2V models. It is noteworthy thatAnimateDiff and our models use only the videos from WebVid-10M. Show-1 employs additional high-quality videos for finetuning to eliminate the watermark in WebVid-10M.AnimateDiff is not a generic T2V model; it works only when the LORA SD model is compatible with its temporal modules. For our comparison, we use its temporal modules (second version) based on SD v1.5 and employ Realistic Vision V2.0 [10] as its corresponding LORA model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.895, + 0.774 + ], + "angle": 0, + "content": "Quantitative Evaluation. The quantitative results obtained using EvalCrafter are presented in Table 1. Our method achieves visual quality comparable to that of VideoCrafter1 and Pika Labs, which use high-quality videos for training. This underscores the effectiveness of employing high-quality images to enhance picture quality and aesthetic scores. Furthermore, our text-video alignment performance is ranked second. In terms of motion quality, our performance surpasses that of Show-1 but falls short of models that utilize a larger volume of videos to learn motion. This indicates that our method can enhance visual quality without significant motion degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Qualitative Evaluation. The visual comparison is depicted in Fig. 7. Additional results are provided in the supplementary material. The visual quality of our results is on par with that of commercial models such as Gen-2 and Pika Labs. Since we employ JDB as the image dataset, the picture quality of our synthesized videos shifts from WebVid-10M to JDB. Regarding motion, our motion quality is superior to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7315" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.088, + 0.902, + 0.74 + ], + "angle": 0, + "content": "
A bear rummages through a\ndumpster, searching for food\nscapsA Time 1980 painting of a boy\ngoing to schoolA man cruises through the city\non a motorcycle, feeling the\nadrenaline rushA monkey eating a pizza in\ncentral park, GoPro film style
Gen-2
Show-1VideoCrafter1Pika Labs
Ours
" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.814, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Figure 7. Comparison of different text-to-video generation models. Best viewed with Acrobat Reader. Click the images to play the video clips." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.471, + 0.901 + ], + "angle": 0, + "content": "that ofAnimateDiff and comparable to Show-1. Although the integration of temporal modules with a LORA SD model can enhance visual quality,AnimateDiff experiences motion" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.855, + 0.702, + 0.871 + ], + "angle": 0, + "content": "degradation in generic scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7316" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.177 + ], + "angle": 0, + "content": "
MethodsText-Video AlignmentMotion QualityVisual Quality
Ours vs Gen256%46%34%
Ours vsAnimateDiff55%64%69%
Ours vs Show-159%59%82%
Ours vs VideoCrafter161%63%61%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.189, + 0.47, + 0.218 + ], + "angle": 0, + "content": "Table 2. Human preference. The numbers represent the probability of users choosing our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.471, + 0.422 + ], + "angle": 0, + "content": "User Study. For further evaluation, we conduct a user study to compare our method with other video models. We select 50 prompts from EvalCrafter, covering diverse scenes, styles, and objects. When comparing a model pair, three video production experts are asked to select their preferred video from three options: method 1, method 2, and comparable results, according to the given subject, i.e., visual quality, motion quality, and text-video alignment. The results are shown in Table 4. Our method has better visual quality thanAnimateDiff and Show-1 and is comparable to VideoCrafter1. Our method is more preferred than Show-1 andAnimateDiff in motion quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.266, + 0.449 + ], + "angle": 0, + "content": "4.3. Strategy Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.471, + 0.729 + ], + "angle": 0, + "content": "Spatial-temporal Connection. In Sec. 3.1, we show a visual comparison of perturbing the spatial and temporal parameters of the fully and partially trained models in Fig. 3 and Fig. 4. Here we provide the quantitative comparisons about the visual quality in Table 3, including aesthetic and technical scores from DOVER [54]. We observe that fin-tuning the partially trained model can always achieve better visual quality than the fully trained model. It means that the distribution of the partially trained model can be shifted more easily. Besides, we conduct a user study asking participants to choose a favorable model that has better performance in motion, in terms of foreground/background flash and motion flicker. The results are shown in Table 4. It can be observed that the motion quality of perturbed fully trained models is better. The fully trained model can tolerate larger parameter perturbations than the partially trained model. These observations show that the fully trained model has stronger spatial-temporal coupling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Module Selection. After selecting the fully trained model as the base, we use two strategies to identify the most effective module to fine-tune, resulting in four models in Sec. 3.2. The visual quality evaluation of these models is shown in the bottom part of Table 3. The visual quality of F-Spa-DIR and F-Spa&Temp-DIR is much better than the other two models. It reveals that directly finetuning spatial modules is the key to improving picture quality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Since F-Spa-DIR and F-Spa&Temp-DIR achieve close visual quality, we conduct a user study on motion quality to" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.885, + 0.256 + ], + "angle": 0, + "content": "
MethodAesthetic Score (↑)Technical Score (↑)
P-base34.3242.69
F-base46.5551.76
P-Spa-LORA78.2572.74
F-Spa-LORA77.9759.60
P-Temp-LORA77.4054.85
F-Temp-LORA66.2650.32
F-Spa-DIR82.5770.35
F-Temp-DIR82.7765.34
F-Spa&Temp-DIR83.5967.75
F-Spa&Temp-LORA80.4463.61
F-Spa-DIR-LAION67.8354.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.266, + 0.884, + 0.28 + ], + "angle": 0, + "content": "Table 3. Visual quality evaluation of the perturbed T2V models." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.29, + 0.884, + 0.355 + ], + "angle": 0, + "content": "
MethodsMotion Quality
F-Spa-LORA vs P-Spa-LORA87%
F-Temp-LORA vs P-Temp-LORA73%
F-Spa-DIR vs F-Spa&Temp-DIR67%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.365, + 0.885, + 0.38 + ], + "angle": 0, + "content": "Table 4. User study on the motion of the perturbed T2V models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.894, + 0.493 + ], + "angle": 0, + "content": "determine the final model. The results are shown in the last row of Table 4. Directly finetuning the spatial modules only performs better in motion. As shown in Fig. 5, F-Spa-DIR is more stable and has better temporal consistency than F-Spa&Temp-DIR. The latter has obvious flashes in both the foreground and background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.893, + 0.65 + ], + "angle": 0, + "content": "Influence of Image Data. To verify the effectiveness of synthesized images, we use the LAION Aesthetics V2 dataset and JDB to directly fine-tune the spatial modules in the second stage, respectively. The visual examples are shown in Fig. 6. It shows that the model trained with JDB composite concepts better than the model trained with LAION Aesthetics V2. The quantitative evaluation of visual quality is shown in Table 3. F-Spa-DIR is much better than F-Spa-DIR-LAION in both aesthetic and technical scores." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.619, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To overcome data limitations, we propose a method for training high-quality video diffusion models without using high-quality videos. We delve into the training schemes of SD-based video models and investigate the coupling strength between spatial and temporal dimensions. We observe that fully trained T2V models exhibit stronger spatial-temporal coupling than partially trained models. Based on this observation, we propose disentangling appearance from motion at the data level, i.e., by exploiting low-quality videos for motion learning and high-quality images for appearance learning. Additionally, we suggest using synthetic images with complex concepts for finetuning, rather than real images. Quantitative and qualitative evaluations are conducted to demonstrate the effectiveness of the proposed method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7317" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.143 + ], + "angle": 0, + "content": "[1] If. Accessed October 22, 2023 [Online] https://github.com/deep-floyd/IF.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.472, + 0.186 + ], + "angle": 0, + "content": "[2] Laion-aesthetics. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-aesthetics/, .6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[3] Laion-coco. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-coco/, .3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[4] Midjourney. Accessed October 22, 2023 [Online] https://www.midjourney.com/home.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[5] Gen-2. Accessed October 22, 2023 [Online] https://research.runwayml.com/gen2,.2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.279, + 0.472, + 0.307 + ], + "angle": 0, + "content": "[6] Genmo. Accessed October 22, 2023 [Online] https://www.genmo.ai/, .2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.309, + 0.472, + 0.337 + ], + "angle": 0, + "content": "[7] Hotshot-x1. Accessed October 22, 2023 [Online] https://github.com/hotshotco/Hotshot-XL.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.472, + 0.367 + ], + "angle": 0, + "content": "[8] Moonvalley. Accessed October 22, 2023 [Online] https://moonvalley.ai/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.472, + 0.397 + ], + "angle": 0, + "content": "[9] Pika labs. Accessed October 22, 2023 [Online] https://www.pika.art/.2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.472, + 0.439 + ], + "angle": 0, + "content": "[10] Realistic vision v2.0. Accessed October 22, 2023 [Online] https://huggingface.co/ckpt/realistic-vision-v20.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.498 + ], + "angle": 0, + "content": "[11] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.5, + 0.472, + 0.542 + ], + "angle": 0, + "content": "[12] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.544, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[13] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[14] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[17] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, 35(1):53-65, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[18] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[19] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, pages 102–118. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[20] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.893, + 0.374 + ], + "angle": 0, + "content": "[23] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[24] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. arXiv preprint arXiv:2310.07702, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.894, + 0.474 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.894, + 0.643 + ], + "angle": 0, + "content": "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[30] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[32] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In AAAI, pages 9117-9125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[33] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7318" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[34] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[35] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.472, + 0.333 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.472, + 0.379 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.472, + 0.494 + ], + "angle": 0, + "content": "[40] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, pages 2830-2839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.472, + 0.551 + ], + "angle": 0, + "content": "[41] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.472, + 0.596 + ], + "angle": 0, + "content": "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.472, + 0.653 + ], + "angle": 0, + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.472, + 0.698 + ], + "angle": 0, + "content": "[44] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.472, + 0.769 + ], + "angle": 0, + "content": "[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[46] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[47] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[49] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[50] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[51] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[52] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[53] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. arXiv preprint arXiv:2207.09814, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[54] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In CVPR, pages 20144-20154, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[55] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.894, + 0.644 + ], + "angle": 0, + "content": "[57] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[58] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[59] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In CVPR, pages 18456-18466, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[61] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[62] Yuechen Zhang, Jinbo Xing, Eric Lo, and Jiaya Jia. Realworld image variation by aligning diffusion inversion chain. arXiv preprint arXiv:2305.18729, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7319" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.148 + ], + "angle": 0, + "content": "[63] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1, 2, 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7320" + } + ] +] \ No newline at end of file diff --git a/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_origin.pdf b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8da4f87ef291dac4fb68a803904a19e6f496b1ed --- /dev/null +++ b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/ab56d7df-62cd-4c60-8578-721928a2487b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e38b01951a54b6c97719ba455089e42554ca26ae616ee7051dc755a963bb9b9f +size 9915019 diff --git a/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/full.md b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3afc82f10b75553af7a2da8bc06f67ae47ace864 --- /dev/null +++ b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/full.md @@ -0,0 +1,313 @@ +# VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models + +Haoxin Chen + +Yong Zhang* + +Xiaodong Cun + +Menghan Xia + +Xintao Wang + +Chao Weng + +Ying Shan + +Tencent AI Lab + +In cyberpunk, neonpunk style, + +Kung Fu Panda, jump and kick. + +Cinematic photo melting pistachio ice + +cream dripping down the cone. + +35mm photograph, film, bokeh. + +Large motion, surrounded by butterflies, + +a girl walks through a lush garden. + +Figure 1. Give a text prompt, our method can generate a video with high visual quality and accurate text-video alignment. Note that it is trained with only low-quality videos and high-quality images. No high-quality videos are required. Best viewed with Acrobat Reader. Click the images to play the video clips. + +# Abstract + +Text-to-video generation aims to produce a video based on a given prompt. Recently, several commercial video models have been able to generate plausible videos with minimal noise, excellent details, and high aesthetic scores. However, these models rely on large-scale, well-filtered, high-quality videos that are not accessible to the community. Many existing research works, which train models using the low-quality WebVid-10M dataset, struggle to generate high-quality videos because the models are optimized to fit WebVid-10M. In this work, we explore the training scheme of video models extended from Stable Diffusion and investigate the feasibility of leveraging low-quality videos and synthesized high-quality images to obtain a high-quality video model. We first analyze the connection between the spatial + +and temporal modules of video models and the distribution shift to low-quality videos. We observe that full training of all modules results in a stronger coupling between spatial and temporal modules than only training temporal modules. Based on this stronger coupling, we shift the distribution to higher quality without motion degradation by finetuning spatial modules with high-quality images, resulting in a generic high-quality video model. Evaluations are conducted to demonstrate the superiority of the proposed method, particularly in picture quality, motion, and concept composition. + +# 1. Introduction + +Benefiting from the development of diffusion models [25, 43], video generation has achieved breakthroughs, particularly in basic text-to-video (T2V) generation models. Most existing methods [14, 23, 26, 30, 48, 63] follow a logic to + +obtain video models, i.e., extending a text-to-image (T2I) backbone to a video model by adding temporal modules and then training it with videos. Several models train video models from scratch, while most start from a pre-trained T2I model, typically Stable Diffusion (SD) [38]. Models can also be categorized into two groups based on the space modeled by diffusion models, i.e., pixel-space models [26, 30, 60] and latent-space models [14, 23, 48, 63]. The latter is the dominant approach. Picture quality, motion consistency, and concept composition are essential dimensions for evaluating a video model. + +Recently, a few commercial startups have released their T2V models [5, 6, 8, 9] that can produce plausible videos with minimal noise, excellent details, and high aesthetic scores. However, they are trained on a large-scale and well-filtered high-quality video dataset, which is not accessible to the community and academia. Collecting millions of high-quality videos is challenging due to copyright restrictions and post-filtering processing. Though there are a few open-source video datasets collected from the Internet for video understanding, such as HowTo100M [33], HD-VILA100M [56], and InterVid [51], there exist many issues for video generation, e.g., poor picture quality and caption, multiple clips in one video, and static frames or slides. WebVid10M [12] is the most widely used dataset to train video generation models in academia. The clips are well-segmented, and the diversity is good. However, the picture quality is unsatisfactory, and most videos have a resolution of about $320\mathrm{p}$ . The lack of high-quality datasets poses a significant obstacle to training high-quality video models in academia. + +In this work, we target a quite challenging problem, i.e., training high-quality video models without using high-quality videos. We dive into the training process of SD-based video models to analyze the connection between spatial and temporal modules under different training strategies and investigate the distribution shift to low-quality videos. We make a meaningful observation that the full training of all modules results in a stronger coupling between appearance and motion than just training temporal modules. The full training can achieve more natural motion and tolerate more subsequent modification of spatial modules, which is the key to improving the quality of generated videos. Based on the observation of the connection, we propose a method to overcome the data limitation by disentangling motion from appearance at the data level. Specifically, instead of high-quality videos, we exploit low-quality videos to guarantee motion consistency and use high-quality images to ensure picture quality and concept composition ability. Benefiting from the successful T2I models such as SDXL and Midjourney, it is convenient to obtain a large set of images with high-resolution and complex concept composition. As shown in Fig. 2, we design a pipeline to train a video diffusion model following the guidelines from the analysis. Fig. 1 + +shows visual examples generated by our method. + +Our main contributions are summarized as follows: + +- We propose a method to overcome the data for training high-quality video models by disentangling motion from appearance at the data level. +- We investigate the connection between spatial and temporal modules, and the distribution shift. We identify the keys to obtain a high-quality video model. +- We design an effective pipeline based on the observations, i.e., obtaining a fully trained video model first and tuning the spatial modules with synthesized high-quality images. + +# 2. Related Work + +The evolution of video generation techniques goes along with the development of generative models. Generative adversarial networks [17] and variational auto-encoders [18] are the commonly used backbones in early research of video generation, e.g., VGAN [47], TGAN [40], MoCoGAN [44], GODIA [52], StyleGAN-V [41], and MCVD [46]. Then, since transformers have been successfully applied in various fields, they are also introduced for video synthesis, e.g.,CogVideo [28], VideoGPT [57], NUVA-infinity [53], TATS [19], MAGVIT [58], Phenaki [45]. + +Recently, diffusion models (DMs) [25, 42, 43] have been a famous star in generative models, especially in text-to-image (T2I) generation [13, 21, 24, 34, 36-39, 62]. For video generation, Video Diffusion Models (VDMs) are proposed to model the distribution of videos. VDM [27] is the first to utilize a space-time factorized U-Net to model videos in pixel space for unconditional video generation. It uses an image-video joint training strategy to avoid concept forgetting.Imagen Video [26] and Make-a-Video [30] are two cascade models that target text-to-video generation in pixel space. Show-1 [60] is another cascade model that uses IF [1] as the base model and LDM extended video model for super-resolution. Subsequently, LVDM [15, 23] and MagicVideo [63] propose to extend LDM [38] to model videos in the latent space of an auto-encoder. Many other methods use the same paradigm, including ModelScope [48], Align Your Latent [14], Hotshot-XL [7], LAVIE [50], PY-OCO [20], VideoFactory [49], VPDM [59], VIDM [32], and Latent-Shift [11]. Besides text-to-video generation, a few methods, such as [16, 55, 61], attempt to generate videos from a given image and a prompt as condition. + +Several startups release their text-to-video generation services, e.g., Gen-2 [5], Pika Labs [9], Moonvalley [8], and Genmo [6]. Their models can generate plausible results with minimal noise, excellent details, and high aesthetic scores. However, those methods are trained with a large-scale well-filtered high-quality video dataset that is not accessible to researchers. The video models are also not available, leading to the slow development of downstream tasks to a certain extent. The most widely used video dataset is WebVid-10M, + +![](images/74ee483ed404f3eed19ed4671a4f5cbd90ee1bcbdff88e7ed86c6cae3756a30a.jpg) +Figure 2. The efficient training pipeline. In the first stage, we use a large amount of low-quality videos to fully train the VDM to learn motion. In the second stage, we only train the spatial parameters with the high-quality images generated by the T2I model to improve the concept composition and picture quality. + +a large-scale dataset of short videos with textual descriptions sourced from stock footage sites. The videos are diverse and rich in their content, and each video is well-segmented, however, the picture quality is unsatisfactory and most videos are $320\mathrm{p}$ . Training a high-quality video model under the data limitation is quite challenging. + +AnimateDiff [22] finds that combining temporal modules from a video model trained on WebVid-10M and a LORA SD model can improve the picture quality of the generated videos. However, this is not a generic model and does not always work. There are a few severe issues. First, the temporal modules can only be combined with a few selected LORA models, which makes it not a generic model. Second, the motion quality degenerates when the modules do not match the LORA model well. UnlikeAnimateDiff, we analyze the connection between spatial and temporal modules instead of direct combination, and design a pipeline to train a generic high-quality video model without high-quality video by disentangling appearance and motion at the data level. + +# 3. Method + +We propose an effective method to overcome the data limitation for training high-quality video diffusion models. We first analyze the connection between spatial and temporal modules of SD-based video models under different training strategies. Based on the observations, we then develop a pipeline to train high-quality video models with just low-quality videos and high-quality images, i.e., disentangling appearance from motion at the data level. + +# 3.1. Spatial-temporal Connection Analyses + +Base T2V model. To leverage the prior in SD trained on a large-scale image dataset, most text-to-video diffusion models inflate the SD model to a video model by adding temporal modules, including Align Your Latent [14],AnimateDiff [22], LVDM [23], Magic Video [63], ModelScopeT2V [48], and LAVIE [50]. They follow VDM [27] + +to use a particular type of 3D-UNet that is factorized over space and time. + +These models can be categorized into two groups according to their training strategies. One is to use videos to train temporal modules with spatial ones fixed, called partial training. The other is to learn both the spatial and temporal modules with the SD weights as initialization, called full training. + +Though these SD-based T2V models have similar architectures, they are trained under different training settings. We use one typical model to investigate the connection between spatial and temporal modules under the two training strategies. We follow the architecture of the open-source VideoCrafter1 [15] with FPS (frames per second) condition. We also incorporate the temporal convolution in ModelScopeT2V [48] to improve temporal consistency. + +Parameter Perturbation for Full and Partial Training. We apply the two training strategies to the same architecture using the same data. The model is initialized from pretrained SD weights. WebVid-10M [12] is exploited as the training data. To avoid concept forgetting, LAION-COCO 600M [3] is also used for video and image joint training. The resolution is $512 \times 320$ . For simplicity, the fully trained video model is denoted as $M_F(\theta_T, \theta_S)$ , while the partially trained one is denoted as $M_P(\theta_T, \theta_S^0)$ , where $\theta_T$ and $\theta_S$ are the learned parameters of the temporal and spatial modules, respectively. $\theta_S^0$ are the original spatial parameters of SD. + +To evaluate the connection strength between spatial and temporal modules, we perturb the parameters of the specified modules by using another high-quality image dataset $\mathcal{D}_I$ for finetuning. The image data is JDB [35] that consists of synthesized images from Midjourney [4]. As the JDB has 4 million images, we use LORA [29] for finetuning. + +Spatial Perturbation. We first perturb the spatial parameters of the two video models using the image dataset. The temporal parameters are frozen. The perturbation process of the fully trained base model $M_F$ can be denoted as: + +$$ +M _ {F} ^ {\prime} (\theta_ {T}, \theta_ {S} + \Delta_ {\theta_ {S}}) \leftarrow \mathrm {P E R T B} _ {\theta_ {S}} ^ {\mathrm {L O R A}} (M _ {F} (\theta_ {T}, \theta_ {S}), \mathcal {D} _ {I}), +$$ + +where $\mathrm{PERTB}_{\theta_S}^{\mathrm{LORA}}$ denote finetuning $M_F$ with respect to $\theta_{S}$ on the image dataset $\mathcal{D}_I$ using LORA. $\Delta_{\theta_S}$ represents the parameters of the LORA branch. Similarly, we can obtain the perturbed model of the partially trained video model: + +$$ +M _ {P} ^ {\prime} \left(\theta_ {T}, \theta_ {S} ^ {0} + \Delta_ {\theta_ {S}}\right) \leftarrow \operatorname {P E R T B} _ {\theta_ {S}} ^ {\text {L O R A}} \left(M _ {P} \left(\theta_ {T}, \theta_ {S} ^ {0}\right), \mathcal {D} _ {I}\right). +$$ + +For easy understanding, we also use the name 'F-SpALORA' to denote model $M_F^{\prime}$ and 'P-Spa-LORA' for $M_P^\prime$ . 'F' denotes the fully trained base model while $\mathbf{P}^{\prime}$ stands for the partially trained model. 'Spa' and Temp' mean finetuning spatial and temporal modules, respectively. 'LORA' represents using LORA for finetuning, while DIR' means direct + +A professional dancer gracefully performs ballet on stage. +Figure 3. Perturbing spatial modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips. + +finetuning without LORA. For example, 'F-Spatial-LORA' represents perturbing spatial modules of the fully trained T2V model using LORA. + +Comparing the synthesized videos of the two resulting models, we have the following observations. First, the motion quality of F-Spa-LORA is more stable than that of P-Spa-LORA (see user study in Table 4). The motion of P-Spa-LORA becomes worse quickly during the finetuning process. The more finetuning steps, the video tends to be more still with local flicker (see Fig. 3). While the motion of F-Spa-LORA slightly degenerates compared to the fully trained base model. Second, P-Spa-LORA achieves much better visual quality than F-Spa-LORA (see Fig. 3). The picture quality and aesthetic score of F-Spa-LORA are greatly improved compared to the partially trained base model (see Table 3). Surprisingly, the watermark is also removed. While F-Spa-LORA obtains a slight improvement in picture quality and aesthetic score, the generated videos are still noisy. + +From the two observations, we can conclude that the coupling strength between spatial and temporal modules of the fully trained model is stronger than that of the partially trained model. Because the spatial-temporal coupling of the partially trained model can be easily broken, leading to quick motion degeneration and picture quality shift. A stronger connection can tolerate parameter perturbation more than a weak one. Our observation can be used to explain the quality improvement and motion degeneration ofAnimateDiff.AnimateDiff is not a generic model and only works for selected personalized SD models. The reason is that its motion modules are obtained with the partially training strategy, and + +# Robot dancing in times square. + +Figure 4. Perturbing temporal modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips. + +they cannot tolerate large parameter perturbations. When the personalized model does not match the temporal modules, both picture and motion quality will degenerate. + +Temporal Perturbation. The partially trained model has only the temporal modules updated, but the picture quality is shifted to the quality of WebVid-10M. Hence, the temporal modules take responsibility for not only the motion but also the picture quality. We perturb the temporal modules while fixing the spatial modules with the image dataset. The perturbation processes can be denoted as: + +$$ +M _ {F} ^ {\prime \prime} \left(\theta_ {T} + \Delta_ {\theta_ {T}}, \theta_ {S}\right) \leftarrow \operatorname {P E R T B} _ {\theta_ {T}} ^ {\text {L O R A}} \left(M _ {F} \left(\theta_ {T}, \theta_ {S}\right), \mathcal {D} _ {I}\right), +$$ + +$$ +M _ {P} ^ {\prime \prime} (\theta_ {T} + \Delta_ {\theta_ {T}}, \theta_ {S} ^ {0}) \gets \mathrm {P E R T B} _ {\theta_ {T}} ^ {\mathrm {L O R A}} (M _ {P} (\theta_ {T}, \theta_ {S} ^ {0}), \mathcal {D} _ {I}). +$$ + +We observe that the picture quality of P-Temp-LORA $(M_P^{\prime \prime})$ is better than F-Temp-LORA $(M_F^{\prime \prime})$ . However, the foreground and background of the videos are more shaky, i.e., the temporal consistency becomes worse (see Fig. 4). The picture of F-Temp-LORA is improved, but the watermark is still there. Its motion is close to the base model and much better than P-Temp-LORA (see Table 4). Those observations also support the conclusion obtained from spatial perturbation. + +# 3.2. Data-level Disentanglement of Appearance and Motion + +Since obtaining a large-scale, high-quality video dataset with high diversity is challenging due to copyright issues, we explore the possibility of training a high-quality video model without using high-quality videos. We have access to low-quality videos such as WebVid-10M and high-quality images such as JDB. We propose to disentangle motion from appearance at the data level, i.e., learning motion from low-quality videos while learning picture quality and aesthetics from high-quality images. We can first train a video model with videos and then fine-tune the video model with images. The keys lie in how to train a video model and how to fine-tune it with images. + +According to the study of the connection between spatial and temporal modules, a fully trained model is more suitable + +# An astronaut is waving his hands on the moon. + +Figure 5. Module selection based on the fully trained T2V model. Best viewed with Acrobat Reader. Click the images to play the video clips. + +Anime illustration of a blue pig, riding a scooter near a lake, with the sun in the sky + +# F-Spa-DIR-LAION + +# F-Spa-DIR + +Figure 6. Influence of image data on concept composition. 'F-Spa-DIR-LAION' uses the LAION aesthetics V2 as the image data while 'F-Spa-DIR' uses JDB. Best viewed with Acrobat Reader. Click the images to play the video clips. + +for the subsequent finetuning with high-quality images. This is because the strong spatial-temporal coupling can tolerate the parameter perturbation for both spatial and temporal modules without obvious motion degeneration. + +Next, we need to investigate how to fine-tune the base model with images. In both spatial and temporal perturbation (Sec. 3.1), the picture quality can be improved but not very significantly. To obtain a greater quality improvement, we evaluate two strategies. One is to involve more parameters, i.e., finetuning both spatial and temporal modules with images. The other is to change the finetuning method, i.e., using direct finetuning without LORA. We can evaluate the following four cases: + +$$ +\begin{array}{l} M _ {F} ^ {A} \left(\theta_ {T} + \Delta_ {\theta_ {T}}, \theta_ {S} + \Delta_ {\theta_ {S}}\right) \leftarrow \operatorname {P E R T B} _ {\theta_ {T}, \theta_ {S}} ^ {\text {L O R A}} \left(M _ {F} \left(\theta_ {T}, \theta_ {S}\right), \mathcal {D} _ {I}\right), \\ M _ {F} ^ {B} \left(\theta_ {T}, \theta_ {S} + \Delta_ {\theta_ {S}}\right) \leftarrow \operatorname {P E R T B} _ {\theta_ {S}} \left(M _ {F} \left(\theta_ {T}, \theta_ {S}\right), \mathcal {D} _ {I}\right), \\ M _ {F} ^ {C} (\theta_ {T} + \Delta_ {\theta_ {T}}, \theta_ {S}) \leftarrow \mathrm {P E R T B} _ {\theta_ {T}} (M _ {F} (\theta_ {T}, \theta_ {S}), \mathcal {D} _ {I}), \\ \end{array} +$$ + +$$ +M _ {F} ^ {D} \left(\theta_ {T} + \Delta_ {\theta_ {T}}, \theta_ {S} + \Delta_ {\theta_ {S}}\right) \leftarrow \operatorname {P E R T B} _ {\theta_ {T}, \theta_ {S}} \left(M _ {F} \left(\theta_ {T}, \theta_ {S}\right), \mathcal {D} _ {I}\right), +$$ + +Where $M_F^A$ (F-Spa&Temp-LORA) is obtained by following the first strategy, $M_F^B$ , $M_F^C$ , and $M_F^D$ are obtained using + +the second strategy. $M_F^B$ (F-Spa-DIR) and $M_F^C$ (F-TempDIR) represent directly finetuning the spatial and temporal modules, respectively. $M_F^D$ (F-Spa&Temp-DIR) represents directly finetuning all modules. + +Comparing the generated videos of the four models, we have the following observations. First, F-Spa&Temp-LORA can further improve the picture quality of F-Spa-LORA, but the quality is still unsatisfying. The watermark exists in most generated videos, and the noise is obvious. Second, F-Temp-DIR achieves better picture quality than F-Temp-LORA. It is also better than F-Spa&Temp-LORA. The watermark is removed or lightened in half of the videos. Third, F-Spa-DIR and F-Spa&Temp-DIR achieve the best picture quality among the fine-tuned models. However, the motion of F-Spa-DIR is better (see Fig. 5 and Table 4). The foreground and background of F-Spa&Temp-DIR are flashing in videos generated by $M_F^D$ , especially local textures. + +By exploring the finetuning strategies and different modules, we identify that directly finetuning spatial modules with high-quality images is the best way to improve the picture quality without marginal loss of motion quality. As shown in Fig. 2, our data-level disentanglement pipeline can be summarized as follows: fully training a video model with low-quality videos first and then directly finetuning the spatial modules only with high-quality images. + +# 3.3. Promotion of Concept Composition + +To improve the concept composition ability of video models, we propose to use synthesized images with complex concepts instead of using real images at the partial finetuning stage. The success of T2I models such as SDXL and Midjourney is built upon large-scale high-quality images. They have the ability to composite concepts that do not appear in the real world. Rather than using their training images, we propose transferring their concept composition ability to video models by synthesizing a set of images with complex concepts. In this way, we can alleviate the burden of capturing both concept and motion well at the same time. + +To validate the effectiveness of synthesized images, we use JDB and LAION-aesthetics V2 [2] as image data for the second finetuning stage. LAION-aesthetics V2 consists of web-collected images while JDB contains images synthesized by Midjourney. We observe that the model trained with JDB has much better concept composition ability (see Fig. 6 and Table 3). More results are in the supplementary material. + +# 4. Experiments + +# 4.1. Settings + +Data. To overcome data limitations, we utilized WebVid-10M [12] as the source of low-quality video data and JDB [35] for high-quality image data. WebVid-10M is a large-scale, diverse video dataset comprising approximately 10 million text-video pairs. The resolution of most videos is $336 \times 596$ , and each video consists of a single shot. During training, we sample from videos at varying frame rates. JDB is a large-scale image dataset featuring around 4 million high-resolution images from Midjourney, each annotated with a corresponding text prompt. To prevent concept forgetting during the training of the base T2V model, we also employ LAION-COCO[3], a dataset comprising 600 million generated high-quality captions for publicly available web images, for both image and video training. + +Metrics. We exploit EvalCrafter [31] for quantitative evaluation. EvalCrafter is a benchmark to evaluate text-to-video generation models, which contains around 18 objective metrics for visual quality, content quality, motion quality, and text-caption alignment. It provides about 512 prompts. The objective metrics are aligned to user opinions from five subjective studies, i.e., motion quality, text-video alignment, temporal consistency, visual quality, and user favor. The motion quality considers three metrics: action recognition, average flow, amplitude classification score, while temporal consistency considers warping error, semantic consistency, face consistency. The technical and aesthetic scores in EvalCrafter are adapted from DOVER [54]. Besides, we conduct user studies of human preference since there still lacks a comprehensive objective metric to measure motion quality. + +Training Details. In Sec 3.1, the two based models share the same architecture, adapted from the open-source VideoCrafter1 [15], and incorporate temporal convolution from ModelScopeT2V [48]. The spatial modules are initialized with weights from SD 2.1, and the outputs of the temporal modules are initialized to zeros. The training resolution is set at $512 \times 320$ . For joint image and video training, we utilize the low-quality WebVid-10M and LAION-COCO datasets. The models are trained on 32 NVIDIA A100 GPUs for $270K$ iterations with a batch size of 128. The learning rate is set at $5 \times 10^{-5}$ for all training tasks. When employing + +
Visual QualityText-Video AlignmentMotion QualityTemporal Consistency
PikaLab*63.5254.1157.7469.35
Gen2*67.3552.3062.5369.71
VideoCrafter161.6466.7656.0660.36
Show-152.1962.0753.7460.83
AnimateDiff58.8974.7951.3856.61
Ours63.2864.6753.9562.02
+ +Table 1. Comparison on the EvalCrafter benchmark. Higher score indicates better performance. * commercial models. + +LORA for the perturbation of temporal or spatial modules, we exclusively use JDB for tuning. The finetuning is conducted on 8 A100 GPUs for $30K$ iterations with a batch size of 256. Given that the images from JDB have a square resolution, we adjust the finetuning resolution to $512 \times 512$ . + +# 4.2. Comparison with State-of-the-Art T2V Models + +We compare our approach with several state-of-the-art T2V models, including popular commercial models such as Gen-2 [5] and Pika Labs [9], as well as open-source models like Show-1 [60], VideoCrafter1 [15], andAnimateDiff [22]. Gen-2, Pika Labs, and VideoCrafter1 all utilize high-quality videos for training their T2V models. It is noteworthy thatAnimateDiff and our models use only the videos from WebVid-10M. Show-1 employs additional high-quality videos for finetuning to eliminate the watermark in WebVid-10M.AnimateDiff is not a generic T2V model; it works only when the LORA SD model is compatible with its temporal modules. For our comparison, we use its temporal modules (second version) based on SD v1.5 and employ Realistic Vision V2.0 [10] as its corresponding LORA model. + +Quantitative Evaluation. The quantitative results obtained using EvalCrafter are presented in Table 1. Our method achieves visual quality comparable to that of VideoCrafter1 and Pika Labs, which use high-quality videos for training. This underscores the effectiveness of employing high-quality images to enhance picture quality and aesthetic scores. Furthermore, our text-video alignment performance is ranked second. In terms of motion quality, our performance surpasses that of Show-1 but falls short of models that utilize a larger volume of videos to learn motion. This indicates that our method can enhance visual quality without significant motion degradation. + +Qualitative Evaluation. The visual comparison is depicted in Fig. 7. Additional results are provided in the supplementary material. The visual quality of our results is on par with that of commercial models such as Gen-2 and Pika Labs. Since we employ JDB as the image dataset, the picture quality of our synthesized videos shifts from WebVid-10M to JDB. Regarding motion, our motion quality is superior to + +
A bear rummages through a +dumpster, searching for food +scapsA Time 1980 painting of a boy +going to schoolA man cruises through the city +on a motorcycle, feeling the +adrenaline rushA monkey eating a pizza in +central park, GoPro film style
Gen-2
Show-1VideoCrafter1Pika Labs
Ours
+ +Figure 7. Comparison of different text-to-video generation models. Best viewed with Acrobat Reader. Click the images to play the video clips. + +that ofAnimateDiff and comparable to Show-1. Although the integration of temporal modules with a LORA SD model can enhance visual quality,AnimateDiff experiences motion + +degradation in generic scenes. + +
MethodsText-Video AlignmentMotion QualityVisual Quality
Ours vs Gen256%46%34%
Ours vsAnimateDiff55%64%69%
Ours vs Show-159%59%82%
Ours vs VideoCrafter161%63%61%
+ +User Study. For further evaluation, we conduct a user study to compare our method with other video models. We select 50 prompts from EvalCrafter, covering diverse scenes, styles, and objects. When comparing a model pair, three video production experts are asked to select their preferred video from three options: method 1, method 2, and comparable results, according to the given subject, i.e., visual quality, motion quality, and text-video alignment. The results are shown in Table 4. Our method has better visual quality thanAnimateDiff and Show-1 and is comparable to VideoCrafter1. Our method is more preferred than Show-1 andAnimateDiff in motion quality. + +# 4.3. Strategy Evaluation + +Spatial-temporal Connection. In Sec. 3.1, we show a visual comparison of perturbing the spatial and temporal parameters of the fully and partially trained models in Fig. 3 and Fig. 4. Here we provide the quantitative comparisons about the visual quality in Table 3, including aesthetic and technical scores from DOVER [54]. We observe that fin-tuning the partially trained model can always achieve better visual quality than the fully trained model. It means that the distribution of the partially trained model can be shifted more easily. Besides, we conduct a user study asking participants to choose a favorable model that has better performance in motion, in terms of foreground/background flash and motion flicker. The results are shown in Table 4. It can be observed that the motion quality of perturbed fully trained models is better. The fully trained model can tolerate larger parameter perturbations than the partially trained model. These observations show that the fully trained model has stronger spatial-temporal coupling. + +Module Selection. After selecting the fully trained model as the base, we use two strategies to identify the most effective module to fine-tune, resulting in four models in Sec. 3.2. The visual quality evaluation of these models is shown in the bottom part of Table 3. The visual quality of F-Spa-DIR and F-Spa&Temp-DIR is much better than the other two models. It reveals that directly finetuning spatial modules is the key to improving picture quality. + +Since F-Spa-DIR and F-Spa&Temp-DIR achieve close visual quality, we conduct a user study on motion quality to + +Table 2. Human preference. The numbers represent the probability of users choosing our method. + +
MethodAesthetic Score (↑)Technical Score (↑)
P-base34.3242.69
F-base46.5551.76
P-Spa-LORA78.2572.74
F-Spa-LORA77.9759.60
P-Temp-LORA77.4054.85
F-Temp-LORA66.2650.32
F-Spa-DIR82.5770.35
F-Temp-DIR82.7765.34
F-Spa&Temp-DIR83.5967.75
F-Spa&Temp-LORA80.4463.61
F-Spa-DIR-LAION67.8354.26
+ +Table 3. Visual quality evaluation of the perturbed T2V models. + +
MethodsMotion Quality
F-Spa-LORA vs P-Spa-LORA87%
F-Temp-LORA vs P-Temp-LORA73%
F-Spa-DIR vs F-Spa&Temp-DIR67%
+ +Table 4. User study on the motion of the perturbed T2V models. + +determine the final model. The results are shown in the last row of Table 4. Directly finetuning the spatial modules only performs better in motion. As shown in Fig. 5, F-Spa-DIR is more stable and has better temporal consistency than F-Spa&Temp-DIR. The latter has obvious flashes in both the foreground and background. + +Influence of Image Data. To verify the effectiveness of synthesized images, we use the LAION Aesthetics V2 dataset and JDB to directly fine-tune the spatial modules in the second stage, respectively. The visual examples are shown in Fig. 6. It shows that the model trained with JDB composite concepts better than the model trained with LAION Aesthetics V2. The quantitative evaluation of visual quality is shown in Table 3. F-Spa-DIR is much better than F-Spa-DIR-LAION in both aesthetic and technical scores. + +# 5. Conclusion + +To overcome data limitations, we propose a method for training high-quality video diffusion models without using high-quality videos. We delve into the training schemes of SD-based video models and investigate the coupling strength between spatial and temporal dimensions. We observe that fully trained T2V models exhibit stronger spatial-temporal coupling than partially trained models. Based on this observation, we propose disentangling appearance from motion at the data level, i.e., by exploiting low-quality videos for motion learning and high-quality images for appearance learning. Additionally, we suggest using synthetic images with complex concepts for finetuning, rather than real images. Quantitative and qualitative evaluations are conducted to demonstrate the effectiveness of the proposed method. + +# References + +[1] If. Accessed October 22, 2023 [Online] https://github.com/deep-floyd/IF.2 +[2] Laion-aesthetics. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-aesthetics/, .6 +[3] Laion-coco. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-coco/, .3,6 +[4] Midjourney. Accessed October 22, 2023 [Online] https://www.midjourney.com/home.3 +[5] Gen-2. Accessed October 22, 2023 [Online] https://research.runwayml.com/gen2,.2,6 +[6] Genmo. Accessed October 22, 2023 [Online] https://www.genmo.ai/, .2 +[7] Hotshot-x1. Accessed October 22, 2023 [Online] https://github.com/hotshotco/Hotshot-XL.2 +[8] Moonvalley. Accessed October 22, 2023 [Online] https://moonvalley.ai/.2 +[9] Pika labs. Accessed October 22, 2023 [Online] https://www.pika.art/.2,6 +[10] Realistic vision v2.0. Accessed October 22, 2023 [Online] https://huggingface.co/ckpt/realistic-vision-v20.6 +[11] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2 +[12] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3, 6 +[13] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[14] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 3 +[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 6 +[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2 +[17] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, 35(1):53-65, 2018. 2 +[18] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2 + +[19] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, pages 102–118. Springer, 2022. 2 +[20] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2 +[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2 +[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3, 6 +[23] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 2, 3 +[24] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. arXiv preprint arXiv:2310.07702, 2023. 2 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2 +[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 2 +[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 3 +[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[30] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 1, 2 +[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models, 2023. 6 +[32] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In AAAI, pages 9117-9125, 2023. 2 +[33] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2 + +[34] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2 +[35] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 3, 6 +[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2 +[40] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, pages 2830-2839, 2017. 2 +[41] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 2 +[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2 +[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2 +[44] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2 +[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2 +[46] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 2 +[47] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 29, 2016. 2 +[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 6 + +[49] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2 +[50] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3 +[51] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2 +[52] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2 +[53] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. arXiv preprint arXiv:2207.09814, 2022. 2 +[54] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In CVPR, pages 20144-20154, 2023. 6, 8 +[55] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2 +[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 2 +[57] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2 +[58] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2 +[59] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In CVPR, pages 18456-18466, 2023. 2 +[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation, 2023. 2, 6 +[61] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 2 +[62] Yuechen Zhang, Jinbo Xing, Eric Lo, and Jiaya Jia. Realworld image variation by aligning diffusion inversion chain. arXiv preprint arXiv:2305.18729, 2023. 2 + +[63] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1, 2, 3 \ No newline at end of file diff --git a/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/images.zip b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0917f78cef2a2fabfc9a3f39374dda5c7b60d67b --- /dev/null +++ b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84ccb327988db393b7dc79952ad7b4de2d922a39e2454cfbb1981b475a6434a5 +size 205015 diff --git a/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/layout.json b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8398dee67990d51cfe0a1c631cbdde538538acf1 --- /dev/null +++ b/2024/VideoCrafter2_ Overcoming Data Limitations for High-Quality Video Diffusion Models/layout.json @@ -0,0 +1,7646 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 102, + 492, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 102, + 492, + 138 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 492, + 138 + ], + "type": "text", + "content": "VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 161, + 192, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 161, + 192, + 175 + ], + "spans": [ + { + "bbox": [ + 126, + 161, + 192, + 175 + ], + "type": "text", + "content": "Haoxin Chen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 162, + 281, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 162, + 281, + 175 + ], + "spans": [ + { + "bbox": [ + 217, + 162, + 281, + 175 + ], + "type": "text", + "content": "Yong Zhang*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 162, + 374, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 162, + 374, + 175 + ], + "spans": [ + { + "bbox": [ + 302, + 162, + 374, + 175 + ], + "type": "text", + "content": "Xiaodong Cun" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 399, + 162, + 467, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 162, + 467, + 175 + ], + "spans": [ + { + "bbox": [ + 399, + 162, + 467, + 175 + ], + "type": "text", + "content": "Menghan Xia" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 176, + 249, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 176, + 249, + 190 + ], + "spans": [ + { + "bbox": [ + 184, + 176, + 249, + 190 + ], + "type": "text", + "content": "Xintao Wang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 274, + 176, + 332, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 176, + 332, + 190 + ], + "spans": [ + { + "bbox": [ + 274, + 176, + 332, + 190 + ], + "type": "text", + "content": "Chao Weng" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 358, + 177, + 408, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 177, + 408, + 190 + ], + "spans": [ + { + "bbox": [ + 358, + 177, + 408, + 190 + ], + "type": "text", + "content": "Ying Shan" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 203, + 334, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 203, + 334, + 216 + ], + "spans": [ + { + "bbox": [ + 255, + 203, + 334, + 216 + ], + "type": "text", + "content": "Tencent AI Lab" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 419, + 168, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 419, + 168, + 430 + ], + "spans": [ + { + "bbox": [ + 58, + 419, + 168, + 430 + ], + "type": "text", + "content": "In cyberpunk, neonpunk style," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 431, + 176, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 431, + 176, + 441 + ], + "spans": [ + { + "bbox": [ + 61, + 431, + 176, + 441 + ], + "type": "text", + "content": "Kung Fu Panda, jump and kick." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 212, + 408, + 352, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 408, + 352, + 419 + ], + "spans": [ + { + "bbox": [ + 212, + 408, + 352, + 419 + ], + "type": "text", + "content": "Cinematic photo melting pistachio ice" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 215, + 420, + 328, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 420, + 328, + 430 + ], + "spans": [ + { + "bbox": [ + 215, + 420, + 328, + 430 + ], + "type": "text", + "content": "cream dripping down the cone." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 216, + 430, + 329, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 430, + 329, + 441 + ], + "spans": [ + { + "bbox": [ + 216, + 430, + 329, + 441 + ], + "type": "text", + "content": "35mm photograph, film, bokeh." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 419, + 529, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 419, + 529, + 430 + ], + "spans": [ + { + "bbox": [ + 383, + 419, + 529, + 430 + ], + "type": "text", + "content": "Large motion, surrounded by butterflies," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 384, + 431, + 513, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 431, + 513, + 441 + ], + "spans": [ + { + "bbox": [ + 384, + 431, + 513, + 441 + ], + "type": "text", + "content": "a girl walks through a lush garden." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 449, + 547, + 483 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 547, + 483 + ], + "type": "text", + "content": "Figure 1. Give a text prompt, our method can generate a video with high visual quality and accurate text-video alignment. Note that it is trained with only low-quality videos and high-quality images. No high-quality videos are required. Best viewed with Acrobat Reader. Click the images to play the video clips." + } + ] + } + ], + "index": 18, + "type": "text" + }, + { + "bbox": [ + 143, + 493, + 192, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 493, + 192, + 506 + ], + "spans": [ + { + "bbox": [ + 143, + 493, + 192, + 506 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 520, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 289, + 688 + ], + "type": "text", + "content": "Text-to-video generation aims to produce a video based on a given prompt. Recently, several commercial video models have been able to generate plausible videos with minimal noise, excellent details, and high aesthetic scores. However, these models rely on large-scale, well-filtered, high-quality videos that are not accessible to the community. Many existing research works, which train models using the low-quality WebVid-10M dataset, struggle to generate high-quality videos because the models are optimized to fit WebVid-10M. In this work, we explore the training scheme of video models extended from Stable Diffusion and investigate the feasibility of leveraging low-quality videos and synthesized high-quality images to obtain a high-quality video model. We first analyze the connection between the spatial" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 495, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 548, + 615 + ], + "type": "text", + "content": "and temporal modules of video models and the distribution shift to low-quality videos. We observe that full training of all modules results in a stronger coupling between spatial and temporal modules than only training temporal modules. Based on this stronger coupling, we shift the distribution to higher quality without motion degradation by finetuning spatial modules with high-quality images, resulting in a generic high-quality video model. Evaluations are conducted to demonstrate the superiority of the proposed method, particularly in picture quality, motion, and concept composition." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 643, + 387, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 387, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 387, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 665, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 548, + 715 + ], + "type": "text", + "content": "Benefiting from the development of diffusion models [25, 43], video generation has achieved breakthroughs, particularly in basic text-to-video (T2V) generation models. Most existing methods [14, 23, 26, 30, 48, 63] follow a logic to" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 127, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 127, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 127, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7310" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "obtain video models, i.e., extending a text-to-image (T2I) backbone to a video model by adding temporal modules and then training it with videos. Several models train video models from scratch, while most start from a pre-trained T2I model, typically Stable Diffusion (SD) [38]. Models can also be categorized into two groups based on the space modeled by diffusion models, i.e., pixel-space models [26, 30, 60] and latent-space models [14, 23, 48, 63]. The latter is the dominant approach. Picture quality, motion consistency, and concept composition are essential dimensions for evaluating a video model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 434 + ], + "type": "text", + "content": "Recently, a few commercial startups have released their T2V models [5, 6, 8, 9] that can produce plausible videos with minimal noise, excellent details, and high aesthetic scores. However, they are trained on a large-scale and well-filtered high-quality video dataset, which is not accessible to the community and academia. Collecting millions of high-quality videos is challenging due to copyright restrictions and post-filtering processing. Though there are a few open-source video datasets collected from the Internet for video understanding, such as HowTo100M [33], HD-VILA100M [56], and InterVid [51], there exist many issues for video generation, e.g., poor picture quality and caption, multiple clips in one video, and static frames or slides. WebVid10M [12] is the most widely used dataset to train video generation models in academia. The clips are well-segmented, and the diversity is good. However, the picture quality is unsatisfactory, and most videos have a resolution of about " + }, + { + "bbox": [ + 46, + 207, + 289, + 434 + ], + "type": "inline_equation", + "content": "320\\mathrm{p}" + }, + { + "bbox": [ + 46, + 207, + 289, + 434 + ], + "type": "text", + "content": ". The lack of high-quality datasets poses a significant obstacle to training high-quality video models in academia." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 438, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 714 + ], + "type": "text", + "content": "In this work, we target a quite challenging problem, i.e., training high-quality video models without using high-quality videos. We dive into the training process of SD-based video models to analyze the connection between spatial and temporal modules under different training strategies and investigate the distribution shift to low-quality videos. We make a meaningful observation that the full training of all modules results in a stronger coupling between appearance and motion than just training temporal modules. The full training can achieve more natural motion and tolerate more subsequent modification of spatial modules, which is the key to improving the quality of generated videos. Based on the observation of the connection, we propose a method to overcome the data limitation by disentangling motion from appearance at the data level. Specifically, instead of high-quality videos, we exploit low-quality videos to guarantee motion consistency and use high-quality images to ensure picture quality and concept composition ability. Benefiting from the successful T2I models such as SDXL and Midjourney, it is convenient to obtain a large set of images with high-resolution and complex concept composition. As shown in Fig. 2, we design a pipeline to train a video diffusion model following the guidelines from the analysis. Fig. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 504, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 504, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 504, + 83 + ], + "type": "text", + "content": "shows visual examples generated by our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 84, + 525, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 84, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 84, + 525, + 95 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 96, + 547, + 204 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 96, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 545, + 132 + ], + "type": "text", + "content": "- We propose a method to overcome the data for training high-quality video models by disentangling motion from appearance at the data level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 133, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 547, + 168 + ], + "type": "text", + "content": "- We investigate the connection between spatial and temporal modules, and the distribution shift. We identify the keys to obtain a high-quality video model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 168, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 547, + 204 + ], + "type": "text", + "content": "- We design an effective pipeline based on the observations, i.e., obtaining a fully trained video model first and tuning the spatial modules with synthesized high-quality images." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 236, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 354 + ], + "type": "text", + "content": "The evolution of video generation techniques goes along with the development of generative models. Generative adversarial networks [17] and variational auto-encoders [18] are the commonly used backbones in early research of video generation, e.g., VGAN [47], TGAN [40], MoCoGAN [44], GODIA [52], StyleGAN-V [41], and MCVD [46]. Then, since transformers have been successfully applied in various fields, they are also introduced for video synthesis, e.g.,CogVideo [28], VideoGPT [57], NUVA-infinity [53], TATS [19], MAGVIT [58], Phenaki [45]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 355, + 548, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 548, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 548, + 605 + ], + "type": "text", + "content": "Recently, diffusion models (DMs) [25, 42, 43] have been a famous star in generative models, especially in text-to-image (T2I) generation [13, 21, 24, 34, 36-39, 62]. For video generation, Video Diffusion Models (VDMs) are proposed to model the distribution of videos. VDM [27] is the first to utilize a space-time factorized U-Net to model videos in pixel space for unconditional video generation. It uses an image-video joint training strategy to avoid concept forgetting.Imagen Video [26] and Make-a-Video [30] are two cascade models that target text-to-video generation in pixel space. Show-1 [60] is another cascade model that uses IF [1] as the base model and LDM extended video model for super-resolution. Subsequently, LVDM [15, 23] and MagicVideo [63] propose to extend LDM [38] to model videos in the latent space of an auto-encoder. Many other methods use the same paradigm, including ModelScope [48], Align Your Latent [14], Hotshot-XL [7], LAVIE [50], PY-OCO [20], VideoFactory [49], VPDM [59], VIDM [32], and Latent-Shift [11]. Besides text-to-video generation, a few methods, such as [16, 55, 61], attempt to generate videos from a given image and a prompt as condition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "content": "Several startups release their text-to-video generation services, e.g., Gen-2 [5], Pika Labs [9], Moonvalley [8], and Genmo [6]. Their models can generate plausible results with minimal noise, excellent details, and high aesthetic scores. However, those methods are trained with a large-scale well-filtered high-quality video dataset that is not accessible to researchers. The video models are also not available, leading to the slow development of downstream tasks to a certain extent. The most widely used video dataset is WebVid-10M," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7311" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 175 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 175 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 175 + ], + "type": "image", + "image_path": "74ee483ed404f3eed19ed4671a4f5cbd90ee1bcbdff88e7ed86c6cae3756a30a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 181, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 237 + ], + "type": "text", + "content": "Figure 2. The efficient training pipeline. In the first stage, we use a large amount of low-quality videos to fully train the VDM to learn motion. In the second stage, we only train the spatial parameters with the high-quality images generated by the T2I model to improve the concept composition and picture quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 254, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 288, + 326 + ], + "type": "text", + "content": "a large-scale dataset of short videos with textual descriptions sourced from stock footage sites. The videos are diverse and rich in their content, and each video is well-segmented, however, the picture quality is unsatisfactory and most videos are " + }, + { + "bbox": [ + 46, + 254, + 288, + 326 + ], + "type": "inline_equation", + "content": "320\\mathrm{p}" + }, + { + "bbox": [ + 46, + 254, + 288, + 326 + ], + "type": "text", + "content": ". Training a high-quality video model under the data limitation is quite challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 290, + 483 + ], + "type": "text", + "content": "AnimateDiff [22] finds that combining temporal modules from a video model trained on WebVid-10M and a LORA SD model can improve the picture quality of the generated videos. However, this is not a generic model and does not always work. There are a few severe issues. First, the temporal modules can only be combined with a few selected LORA models, which makes it not a generic model. Second, the motion quality degenerates when the modules do not match the LORA model well. UnlikeAnimateDiff, we analyze the connection between spatial and temporal modules instead of direct combination, and design a pipeline to train a generic high-quality video model without high-quality video by disentangling appearance and motion at the data level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 495, + 103, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 103, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 103, + 508 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 516, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 288, + 613 + ], + "type": "text", + "content": "We propose an effective method to overcome the data limitation for training high-quality video diffusion models. We first analyze the connection between spatial and temporal modules of SD-based video models under different training strategies. Based on the observations, we then develop a pipeline to train high-quality video models with just low-quality videos and high-quality images, i.e., disentangling appearance from motion at the data level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 622, + 250, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 250, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 250, + 635 + ], + "type": "text", + "content": "3.1. Spatial-temporal Connection Analyses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Base T2V model. To leverage the prior in SD trained on a large-scale image dataset, most text-to-video diffusion models inflate the SD model to a video model by adding temporal modules, including Align Your Latent [14],AnimateDiff [22], LVDM [23], Magic Video [63], ModelScopeT2V [48], and LAVIE [50]. They follow VDM [27]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "to use a particular type of 3D-UNet that is factorized over space and time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 167 + ], + "type": "text", + "content": "These models can be categorized into two groups according to their training strategies. One is to use videos to train temporal modules with spatial ones fixed, called partial training. The other is to learn both the spatial and temporal modules with the SD weights as initialization, called full training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 167, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 264 + ], + "type": "text", + "content": "Though these SD-based T2V models have similar architectures, they are trained under different training settings. We use one typical model to investigate the connection between spatial and temporal modules under the two training strategies. We follow the architecture of the open-source VideoCrafter1 [15] with FPS (frames per second) condition. We also incorporate the temporal convolution in ModelScopeT2V [48] to improve temporal consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": "Parameter Perturbation for Full and Partial Training. We apply the two training strategies to the same architecture using the same data. The model is initialized from pretrained SD weights. WebVid-10M [12] is exploited as the training data. To avoid concept forgetting, LAION-COCO 600M [3] is also used for video and image joint training. The resolution is " + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "inline_equation", + "content": "512 \\times 320" + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": ". For simplicity, the fully trained video model is denoted as " + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "inline_equation", + "content": "M_F(\\theta_T, \\theta_S)" + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": ", while the partially trained one is denoted as " + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "inline_equation", + "content": "M_P(\\theta_T, \\theta_S^0)" + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\theta_S" + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": " are the learned parameters of the temporal and spatial modules, respectively. " + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\theta_S^0" + }, + { + "bbox": [ + 304, + 277, + 547, + 410 + ], + "type": "text", + "content": " are the original spatial parameters of SD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 410, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 481 + ], + "type": "text", + "content": "To evaluate the connection strength between spatial and temporal modules, we perturb the parameters of the specified modules by using another high-quality image dataset " + }, + { + "bbox": [ + 304, + 410, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_I" + }, + { + "bbox": [ + 304, + 410, + 545, + 481 + ], + "type": "text", + "content": " for finetuning. The image data is JDB [35] that consists of synthesized images from Midjourney [4]. As the JDB has 4 million images, we use LORA [29] for finetuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 481, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 547, + 529 + ], + "type": "text", + "content": "Spatial Perturbation. We first perturb the spatial parameters of the two video models using the image dataset. The temporal parameters are frozen. The perturbation process of the fully trained base model " + }, + { + "bbox": [ + 304, + 481, + 547, + 529 + ], + "type": "inline_equation", + "content": "M_F" + }, + { + "bbox": [ + 304, + 481, + 547, + 529 + ], + "type": "text", + "content": " can be denoted as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 537, + 538, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 538, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 538, + 553 + ], + "type": "interline_equation", + "content": "M _ {F} ^ {\\prime} (\\theta_ {T}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}) \\leftarrow \\mathrm {P E R T B} _ {\\theta_ {S}} ^ {\\mathrm {L O R A}} (M _ {F} (\\theta_ {T}, \\theta_ {S}), \\mathcal {D} _ {I}),", + "image_path": "421610c9dc370878b083b1242146d730b0707c6ff8760e53ffef1b4ca87ee9f0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{PERTB}_{\\theta_S}^{\\mathrm{LORA}}" + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "content": " denote finetuning " + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "inline_equation", + "content": "M_F" + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\theta_{S}" + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "content": " on the image dataset " + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_I" + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "content": " using LORA. " + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\Delta_{\\theta_S}" + }, + { + "bbox": [ + 304, + 561, + 545, + 611 + ], + "type": "text", + "content": " represents the parameters of the LORA branch. Similarly, we can obtain the perturbed model of the partially trained video model:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 538, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 538, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 538, + 634 + ], + "type": "interline_equation", + "content": "M _ {P} ^ {\\prime} \\left(\\theta_ {T}, \\theta_ {S} ^ {0} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {S}} ^ {\\text {L O R A}} \\left(M _ {P} \\left(\\theta_ {T}, \\theta_ {S} ^ {0}\\right), \\mathcal {D} _ {I}\\right).", + "image_path": "d322b78d49e894d3486f5d20d21f939e9439703ab8bf83b9651fa081e08e2cef.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "For easy understanding, we also use the name 'F-SpALORA' to denote model " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "M_F^{\\prime}" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " and 'P-Spa-LORA' for " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "M_P^\\prime" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " . 'F' denotes the fully trained base model while " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\prime}" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " stands for the partially trained model. 'Spa' and Temp' mean finetuning spatial and temporal modules, respectively. 'LORA' represents using LORA for finetuning, while DIR' means direct" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7312" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 306, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 288, + 342 + ], + "type": "text", + "content": "A professional dancer gracefully performs ballet on stage. \nFigure 3. Perturbing spatial modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "text", + "content": "finetuning without LORA. For example, 'F-Spatial-LORA' represents perturbing spatial modules of the fully trained T2V model using LORA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 288, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 569 + ], + "type": "text", + "content": "Comparing the synthesized videos of the two resulting models, we have the following observations. First, the motion quality of F-Spa-LORA is more stable than that of P-Spa-LORA (see user study in Table 4). The motion of P-Spa-LORA becomes worse quickly during the finetuning process. The more finetuning steps, the video tends to be more still with local flicker (see Fig. 3). While the motion of F-Spa-LORA slightly degenerates compared to the fully trained base model. Second, P-Spa-LORA achieves much better visual quality than F-Spa-LORA (see Fig. 3). The picture quality and aesthetic score of F-Spa-LORA are greatly improved compared to the partially trained base model (see Table 3). Surprisingly, the watermark is also removed. While F-Spa-LORA obtains a slight improvement in picture quality and aesthetic score, the generated videos are still noisy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "From the two observations, we can conclude that the coupling strength between spatial and temporal modules of the fully trained model is stronger than that of the partially trained model. Because the spatial-temporal coupling of the partially trained model can be easily broken, leading to quick motion degeneration and picture quality shift. A stronger connection can tolerate parameter perturbation more than a weak one. Our observation can be used to explain the quality improvement and motion degeneration ofAnimateDiff.AnimateDiff is not a generic model and only works for selected personalized SD models. The reason is that its motion modules are obtained with the partially training strategy, and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 195, + 482, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 195, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 367, + 195, + 482, + 205 + ], + "type": "text", + "content": "Robot dancing in times square." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 207, + 547, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 547, + 230 + ], + "type": "text", + "content": "Figure 4. Perturbing temporal modules using LORA. Best viewed with Acrobat Reader. Click the images to play the video clips." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 242, + 546, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 546, + 277 + ], + "type": "text", + "content": "they cannot tolerate large parameter perturbations. When the personalized model does not match the temporal modules, both picture and motion quality will degenerate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 277, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 547, + 361 + ], + "type": "text", + "content": "Temporal Perturbation. The partially trained model has only the temporal modules updated, but the picture quality is shifted to the quality of WebVid-10M. Hence, the temporal modules take responsibility for not only the motion but also the picture quality. We perturb the temporal modules while fixing the spatial modules with the image dataset. The perturbation processes can be denoted as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 369, + 538, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 369, + 538, + 385 + ], + "spans": [ + { + "bbox": [ + 312, + 369, + 538, + 385 + ], + "type": "interline_equation", + "content": "M _ {F} ^ {\\prime \\prime} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}} ^ {\\text {L O R A}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right),", + "image_path": "88f91e4731ba2039ca18b4a1863ac74bb9a54d98088e7d0ee064013365857151.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 387, + 538, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 538, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 538, + 403 + ], + "type": "interline_equation", + "content": "M _ {P} ^ {\\prime \\prime} (\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} ^ {0}) \\gets \\mathrm {P E R T B} _ {\\theta_ {T}} ^ {\\mathrm {L O R A}} (M _ {P} (\\theta_ {T}, \\theta_ {S} ^ {0}), \\mathcal {D} _ {I}).", + "image_path": "2cb079ddd519e80d2e4d45feab560b06d53b366b3523f843aa7534cf753c2ae1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "type": "text", + "content": "We observe that the picture quality of P-Temp-LORA " + }, + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "type": "inline_equation", + "content": "(M_P^{\\prime \\prime})" + }, + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "type": "text", + "content": " is better than F-Temp-LORA " + }, + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "type": "inline_equation", + "content": "(M_F^{\\prime \\prime})" + }, + { + "bbox": [ + 305, + 411, + 547, + 510 + ], + "type": "text", + "content": ". However, the foreground and background of the videos are more shaky, i.e., the temporal consistency becomes worse (see Fig. 4). The picture of F-Temp-LORA is improved, but the watermark is still there. Its motion is close to the base model and much better than P-Temp-LORA (see Table 4). Those observations also support the conclusion obtained from spatial perturbation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 515, + 545, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 545, + 539 + ], + "type": "text", + "content": "3.2. Data-level Disentanglement of Appearance and Motion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 689 + ], + "type": "text", + "content": "Since obtaining a large-scale, high-quality video dataset with high diversity is challenging due to copyright issues, we explore the possibility of training a high-quality video model without using high-quality videos. We have access to low-quality videos such as WebVid-10M and high-quality images such as JDB. We propose to disentangle motion from appearance at the data level, i.e., learning motion from low-quality videos while learning picture quality and aesthetics from high-quality images. We can first train a video model with videos and then fine-tune the video model with images. The keys lie in how to train a video model and how to fine-tune it with images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "According to the study of the connection between spatial and temporal modules, a fully trained model is more suitable" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 97, + 71, + 140, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 71, + 140, + 81 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 140, + 81 + ], + "type": "text", + "content": "iter " + }, + { + "bbox": [ + 97, + 71, + 140, + 81 + ], + "type": "inline_equation", + "content": "= 1K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 203, + 71, + 250, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 71, + 250, + 81 + ], + "spans": [ + { + "bbox": [ + 203, + 71, + 250, + 81 + ], + "type": "text", + "content": "iter " + }, + { + "bbox": [ + 203, + 71, + 250, + 81 + ], + "type": "inline_equation", + "content": "= 10K" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 343, + 72, + 399, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 72, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 343, + 72, + 399, + 83 + ], + "type": "text", + "content": "P-Temp-LORA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 451, + 72, + 506, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 72, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 451, + 72, + 506, + 83 + ], + "type": "text", + "content": "F-Temp-LORA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 113, + 63, + 163 + ], + "type": "aside_text", + "angle": 270, + "lines": [ + { + "bbox": [ + 52, + 113, + 63, + 163 + ], + "spans": [ + { + "bbox": [ + 52, + 113, + 63, + 163 + ], + "type": "text", + "content": "P-Spa-LORA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 224, + 63, + 275 + ], + "type": "aside_text", + "angle": 270, + "lines": [ + { + "bbox": [ + 52, + 224, + 63, + 275 + ], + "spans": [ + { + "bbox": [ + 52, + 224, + 63, + 275 + ], + "type": "text", + "content": "F-Spa-LORA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7313" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 200, + 205, + 388, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 205, + 388, + 217 + ], + "spans": [ + { + "bbox": [ + 200, + 205, + 388, + 217 + ], + "type": "text", + "content": "An astronaut is waving his hands on the moon." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 221, + 544, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 544, + 232 + ], + "type": "text", + "content": "Figure 5. Module selection based on the fully trained T2V model. Best viewed with Acrobat Reader. Click the images to play the video clips." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 247, + 252, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 247, + 252, + 270 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 252, + 270 + ], + "type": "text", + "content": "Anime illustration of a blue pig, riding a scooter near a lake, with the sun in the sky" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 388, + 146, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 388, + 146, + 399 + ], + "spans": [ + { + "bbox": [ + 72, + 388, + 146, + 399 + ], + "type": "text", + "content": "F-Spa-DIR-LAION" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 202, + 388, + 244, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 388, + 244, + 399 + ], + "spans": [ + { + "bbox": [ + 202, + 388, + 244, + 399 + ], + "type": "text", + "content": "F-Spa-DIR" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 401, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 288, + 445 + ], + "type": "text", + "content": "Figure 6. Influence of image data on concept composition. 'F-Spa-DIR-LAION' uses the LAION aesthetics V2 as the image data while 'F-Spa-DIR' uses JDB. Best viewed with Acrobat Reader. Click the images to play the video clips." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 456, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 503 + ], + "type": "text", + "content": "for the subsequent finetuning with high-quality images. This is because the strong spatial-temporal coupling can tolerate the parameter perturbation for both spatial and temporal modules without obvious motion degeneration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 504, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 287, + 612 + ], + "type": "text", + "content": "Next, we need to investigate how to fine-tune the base model with images. In both spatial and temporal perturbation (Sec. 3.1), the picture quality can be improved but not very significantly. To obtain a greater quality improvement, we evaluate two strategies. One is to involve more parameters, i.e., finetuning both spatial and temporal modules with images. The other is to change the finetuning method, i.e., using direct finetuning without LORA. We can evaluate the following four cases:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 618, + 305, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 618, + 305, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 618, + 305, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M _ {F} ^ {A} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}, \\theta_ {S}} ^ {\\text {L O R A}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right), \\\\ M _ {F} ^ {B} \\left(\\theta_ {T}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {S}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right), \\\\ M _ {F} ^ {C} (\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S}) \\leftarrow \\mathrm {P E R T B} _ {\\theta_ {T}} (M _ {F} (\\theta_ {T}, \\theta_ {S}), \\mathcal {D} _ {I}), \\\\ \\end{array}", + "image_path": "8dd52858550441ed6ed3f8bcc57ed64a75dd1e3d39ec0664f626223118a7a02e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 667, + 305, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 667, + 305, + 682 + ], + "spans": [ + { + "bbox": [ + 49, + 667, + 305, + 682 + ], + "type": "interline_equation", + "content": "M _ {F} ^ {D} \\left(\\theta_ {T} + \\Delta_ {\\theta_ {T}}, \\theta_ {S} + \\Delta_ {\\theta_ {S}}\\right) \\leftarrow \\operatorname {P E R T B} _ {\\theta_ {T}, \\theta_ {S}} \\left(M _ {F} \\left(\\theta_ {T}, \\theta_ {S}\\right), \\mathcal {D} _ {I}\\right),", + "image_path": "c9d0b959ec590c49324d6302708042462c3743c40e29b2457a6358f0137b808f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_F^A" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " (F-Spa&Temp-LORA) is obtained by following the first strategy, " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_F^B" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_F^C" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_F^D" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " are obtained using" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "text", + "content": "the second strategy. " + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "inline_equation", + "content": "M_F^B" + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "text", + "content": " (F-Spa-DIR) and " + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "inline_equation", + "content": "M_F^C" + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "text", + "content": " (F-TempDIR) represent directly finetuning the spatial and temporal modules, respectively. " + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "inline_equation", + "content": "M_F^D" + }, + { + "bbox": [ + 306, + 247, + 547, + 296 + ], + "type": "text", + "content": " (F-Spa&Temp-DIR) represents directly finetuning all modules." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": "Comparing the generated videos of the four models, we have the following observations. First, F-Spa&Temp-LORA can further improve the picture quality of F-Spa-LORA, but the quality is still unsatisfying. The watermark exists in most generated videos, and the noise is obvious. Second, F-Temp-DIR achieves better picture quality than F-Temp-LORA. It is also better than F-Spa&Temp-LORA. The watermark is removed or lightened in half of the videos. Third, F-Spa-DIR and F-Spa&Temp-DIR achieve the best picture quality among the fine-tuned models. However, the motion of F-Spa-DIR is better (see Fig. 5 and Table 4). The foreground and background of F-Spa&Temp-DIR are flashing in videos generated by " + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "inline_equation", + "content": "M_F^D" + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": ", especially local textures." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 454, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 454, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 545, + 551 + ], + "type": "text", + "content": "By exploring the finetuning strategies and different modules, we identify that directly finetuning spatial modules with high-quality images is the best way to improve the picture quality without marginal loss of motion quality. As shown in Fig. 2, our data-level disentanglement pipeline can be summarized as follows: fully training a video model with low-quality videos first and then directly finetuning the spatial modules only with high-quality images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 562, + 493, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 493, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 493, + 574 + ], + "type": "text", + "content": "3.3. Promotion of Concept Composition" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 547, + 713 + ], + "type": "text", + "content": "To improve the concept composition ability of video models, we propose to use synthesized images with complex concepts instead of using real images at the partial finetuning stage. The success of T2I models such as SDXL and Midjourney is built upon large-scale high-quality images. They have the ability to composite concepts that do not appear in the real world. Rather than using their training images, we propose transferring their concept composition ability to video models by synthesizing a set of images with complex concepts. In this way, we can alleviate the burden of capturing both concept and motion well at the same time." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 71, + 156, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 71, + 156, + 83 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 156, + 83 + ], + "type": "text", + "content": "F-Spa&Temp-LORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 211, + 71, + 260, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 71, + 260, + 82 + ], + "spans": [ + { + "bbox": [ + 211, + 71, + 260, + 82 + ], + "type": "text", + "content": "F-Temp-DIR" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 334, + 71, + 376, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 71, + 376, + 82 + ], + "spans": [ + { + "bbox": [ + 334, + 71, + 376, + 82 + ], + "type": "text", + "content": "F-Spa-DIR" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 440, + 71, + 509, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 71, + 509, + 83 + ], + "spans": [ + { + "bbox": [ + 440, + 71, + 509, + 83 + ], + "type": "text", + "content": "F-Spa&Temp-DIR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7314" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "To validate the effectiveness of synthesized images, we use JDB and LAION-aesthetics V2 [2] as image data for the second finetuning stage. LAION-aesthetics V2 consists of web-collected images while JDB contains images synthesized by Midjourney. We observe that the model trained with JDB has much better concept composition ability (see Fig. 6 and Table 3). More results are in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 167, + 128, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 128, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 128, + 180 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 107, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 107, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 107, + 199 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": "Data. To overcome data limitations, we utilized WebVid-10M [12] as the source of low-quality video data and JDB [35] for high-quality image data. WebVid-10M is a large-scale, diverse video dataset comprising approximately 10 million text-video pairs. The resolution of most videos is " + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "inline_equation", + "content": "336 \\times 596" + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": ", and each video consists of a single shot. During training, we sample from videos at varying frame rates. JDB is a large-scale image dataset featuring around 4 million high-resolution images from Midjourney, each annotated with a corresponding text prompt. To prevent concept forgetting during the training of the base T2V model, we also employ LAION-COCO[3], a dataset comprising 600 million generated high-quality captions for publicly available web images, for both image and video training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 387, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 288, + 568 + ], + "type": "text", + "content": "Metrics. We exploit EvalCrafter [31] for quantitative evaluation. EvalCrafter is a benchmark to evaluate text-to-video generation models, which contains around 18 objective metrics for visual quality, content quality, motion quality, and text-caption alignment. It provides about 512 prompts. The objective metrics are aligned to user opinions from five subjective studies, i.e., motion quality, text-video alignment, temporal consistency, visual quality, and user favor. The motion quality considers three metrics: action recognition, average flow, amplitude classification score, while temporal consistency considers warping error, semantic consistency, face consistency. The technical and aesthetic scores in EvalCrafter are adapted from DOVER [54]. Besides, we conduct user studies of human preference since there still lacks a comprehensive objective metric to measure motion quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Training Details. In Sec 3.1, the two based models share the same architecture, adapted from the open-source VideoCrafter1 [15], and incorporate temporal convolution from ModelScopeT2V [48]. The spatial modules are initialized with weights from SD 2.1, and the outputs of the temporal modules are initialized to zeros. The training resolution is set at " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 320" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". For joint image and video training, we utilize the low-quality WebVid-10M and LAION-COCO datasets. The models are trained on 32 NVIDIA A100 GPUs for " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "270K" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " iterations with a batch size of 128. The learning rate is set at " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " for all training tasks. When employing" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 159 + ], + "type": "table", + "html": "
Visual QualityText-Video AlignmentMotion QualityTemporal Consistency
PikaLab*63.5254.1157.7469.35
Gen2*67.3552.3062.5369.71
VideoCrafter161.6466.7656.0660.36
Show-152.1962.0753.7460.83
AnimateDiff58.8974.7951.3856.61
Ours63.2864.6753.9562.02
", + "image_path": "4d3f9dfadb7e42d749ef75ea59dbdd0c03e7badf7ce6446457a85dd6bc46508f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 160, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 160, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 160, + 545, + 182 + ], + "type": "text", + "content": "Table 1. Comparison on the EvalCrafter benchmark. Higher score indicates better performance. * commercial models." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "type": "text", + "content": "LORA for the perturbation of temporal or spatial modules, we exclusively use JDB for tuning. The finetuning is conducted on 8 A100 GPUs for " + }, + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "type": "inline_equation", + "content": "30K" + }, + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "type": "text", + "content": " iterations with a batch size of 256. Given that the images from JDB have a square resolution, we adjust the finetuning resolution to " + }, + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 197, + 547, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 266, + 545, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 278 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art T2V Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 285, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 452 + ], + "type": "text", + "content": "We compare our approach with several state-of-the-art T2V models, including popular commercial models such as Gen-2 [5] and Pika Labs [9], as well as open-source models like Show-1 [60], VideoCrafter1 [15], andAnimateDiff [22]. Gen-2, Pika Labs, and VideoCrafter1 all utilize high-quality videos for training their T2V models. It is noteworthy thatAnimateDiff and our models use only the videos from WebVid-10M. Show-1 employs additional high-quality videos for finetuning to eliminate the watermark in WebVid-10M.AnimateDiff is not a generic T2V model; it works only when the LORA SD model is compatible with its temporal modules. For our comparison, we use its temporal modules (second version) based on SD v1.5 and employ Realistic Vision V2.0 [10] as its corresponding LORA model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 613 + ], + "type": "text", + "content": "Quantitative Evaluation. The quantitative results obtained using EvalCrafter are presented in Table 1. Our method achieves visual quality comparable to that of VideoCrafter1 and Pika Labs, which use high-quality videos for training. This underscores the effectiveness of employing high-quality images to enhance picture quality and aesthetic scores. Furthermore, our text-video alignment performance is ranked second. In terms of motion quality, our performance surpasses that of Show-1 but falls short of models that utilize a larger volume of videos to learn motion. This indicates that our method can enhance visual quality without significant motion degradation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Qualitative Evaluation. The visual comparison is depicted in Fig. 7. Additional results are provided in the supplementary material. The visual quality of our results is on par with that of commercial models such as Gen-2 and Pika Labs. Since we employ JDB as the image dataset, the picture quality of our synthesized videos shifts from WebVid-10M to JDB. Regarding motion, our motion quality is superior to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7315" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 69, + 552, + 586 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 552, + 586 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 552, + 586 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 552, + 586 + ], + "type": "table", + "html": "
A bear rummages through a\ndumpster, searching for food\nscapsA Time 1980 painting of a boy\ngoing to schoolA man cruises through the city\non a motorcycle, feeling the\nadrenaline rushA monkey eating a pizza in\ncentral park, GoPro film style
Gen-2
Show-1VideoCrafter1Pika Labs
Ours
", + "image_path": "df1cde7bcaf23e87514f277b58a92271c01105c69a01a49cbbe5408142ef5875.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 644, + 545, + 657 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 545, + 657 + ], + "type": "text", + "content": "Figure 7. Comparison of different text-to-video generation models. Best viewed with Acrobat Reader. Click the images to play the video clips." + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "that ofAnimateDiff and comparable to Show-1. Although the integration of temporal modules with a LORA SD model can enhance visual quality,AnimateDiff experiences motion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 677, + 429, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 429, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 429, + 689 + ], + "type": "text", + "content": "degradation in generic scenes." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7316" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 140 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 140 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 140 + ], + "type": "table", + "html": "
MethodsText-Video AlignmentMotion QualityVisual Quality
Ours vs Gen256%46%34%
Ours vsAnimateDiff55%64%69%
Ours vs Show-159%59%82%
Ours vs VideoCrafter161%63%61%
", + "image_path": "e4ced48ca6d5a96b7cb4bf8846060d9da8a30482b0dee4eb2b6b16cba12246d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 288, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 288, + 334 + ], + "type": "text", + "content": "User Study. For further evaluation, we conduct a user study to compare our method with other video models. We select 50 prompts from EvalCrafter, covering diverse scenes, styles, and objects. When comparing a model pair, three video production experts are asked to select their preferred video from three options: method 1, method 2, and comparable results, according to the given subject, i.e., visual quality, motion quality, and text-video alignment. The results are shown in Table 4. Our method has better visual quality thanAnimateDiff and Show-1 and is comparable to VideoCrafter1. Our method is more preferred than Show-1 andAnimateDiff in motion quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 162, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 162, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 162, + 355 + ], + "type": "text", + "content": "4.3. Strategy Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 288, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 577 + ], + "type": "text", + "content": "Spatial-temporal Connection. In Sec. 3.1, we show a visual comparison of perturbing the spatial and temporal parameters of the fully and partially trained models in Fig. 3 and Fig. 4. Here we provide the quantitative comparisons about the visual quality in Table 3, including aesthetic and technical scores from DOVER [54]. We observe that fin-tuning the partially trained model can always achieve better visual quality than the fully trained model. It means that the distribution of the partially trained model can be shifted more easily. Besides, we conduct a user study asking participants to choose a favorable model that has better performance in motion, in terms of foreground/background flash and motion flicker. The results are shown in Table 4. It can be observed that the motion quality of perturbed fully trained models is better. The fully trained model can tolerate larger parameter perturbations than the partially trained model. These observations show that the fully trained model has stronger spatial-temporal coupling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": "Module Selection. After selecting the fully trained model as the base, we use two strategies to identify the most effective module to fine-tune, resulting in four models in Sec. 3.2. The visual quality evaluation of these models is shown in the bottom part of Table 3. The visual quality of F-Spa-DIR and F-Spa&Temp-DIR is much better than the other two models. It reveals that directly finetuning spatial modules is the key to improving picture quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Since F-Spa-DIR and F-Spa&Temp-DIR achieve close visual quality, we conduct a user study on motion quality to" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 541, + 202 + ], + "blocks": [ + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "lines": [ + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "type": "text", + "content": "Table 2. Human preference. The numbers represent the probability of users choosing our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 541, + 202 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 541, + 202 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 541, + 202 + ], + "type": "table", + "html": "
MethodAesthetic Score (↑)Technical Score (↑)
P-base34.3242.69
F-base46.5551.76
P-Spa-LORA78.2572.74
F-Spa-LORA77.9759.60
P-Temp-LORA77.4054.85
F-Temp-LORA66.2650.32
F-Spa-DIR82.5770.35
F-Temp-DIR82.7765.34
F-Spa&Temp-DIR83.5967.75
F-Spa&Temp-LORA80.4463.61
F-Spa-DIR-LAION67.8354.26
", + "image_path": "11ce327d197de53627bf25823592d484f98dda11123ab8d8d760a110f932650a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 229, + 541, + 281 + ], + "blocks": [ + { + "bbox": [ + 309, + 210, + 541, + 221 + ], + "lines": [ + { + "bbox": [ + 309, + 210, + 541, + 221 + ], + "spans": [ + { + "bbox": [ + 309, + 210, + 541, + 221 + ], + "type": "text", + "content": "Table 3. Visual quality evaluation of the perturbed T2V models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 229, + 541, + 281 + ], + "lines": [ + { + "bbox": [ + 310, + 229, + 541, + 281 + ], + "spans": [ + { + "bbox": [ + 310, + 229, + 541, + 281 + ], + "type": "table", + "html": "
MethodsMotion Quality
F-Spa-LORA vs P-Spa-LORA87%
F-Temp-LORA vs P-Temp-LORA73%
F-Spa-DIR vs F-Spa&Temp-DIR67%
", + "image_path": "d61f5599b06ca963e4e7db9be2dc7f8e9e956b67facaa26d04cc6f07154d43a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 289, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 309, + 289, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 309, + 289, + 541, + 300 + ], + "type": "text", + "content": "Table 4. User study on the motion of the perturbed T2V models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "type": "text", + "content": "determine the final model. The results are shown in the last row of Table 4. Directly finetuning the spatial modules only performs better in motion. As shown in Fig. 5, F-Spa-DIR is more stable and has better temporal consistency than F-Spa&Temp-DIR. The latter has obvious flashes in both the foreground and background." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 406, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 546, + 514 + ], + "type": "text", + "content": "Influence of Image Data. To verify the effectiveness of synthesized images, we use the LAION Aesthetics V2 dataset and JDB to directly fine-tune the spatial modules in the second stage, respectively. The visual examples are shown in Fig. 6. It shows that the model trained with JDB composite concepts better than the model trained with LAION Aesthetics V2. The quantitative evaluation of visual quality is shown in Table 3. F-Spa-DIR is much better than F-Spa-DIR-LAION in both aesthetic and technical scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "To overcome data limitations, we propose a method for training high-quality video diffusion models without using high-quality videos. We delve into the training schemes of SD-based video models and investigate the coupling strength between spatial and temporal dimensions. We observe that fully trained T2V models exhibit stronger spatial-temporal coupling than partially trained models. Based on this observation, we propose disentangling appearance from motion at the data level, i.e., by exploiting low-quality videos for motion learning and high-quality images for appearance learning. Additionally, we suggest using synthetic images with complex concepts for finetuning, rather than real images. Quantitative and qualitative evaluations are conducted to demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7317" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 113 + ], + "type": "text", + "content": "[1] If. Accessed October 22, 2023 [Online] https://github.com/deep-floyd/IF.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 288, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 288, + 147 + ], + "type": "text", + "content": "[2] Laion-aesthetics. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-aesthetics/, .6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 149, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 288, + 171 + ], + "type": "text", + "content": "[3] Laion-coco. Accessed October 22, 2023 [Online] https://laion.ai/blog/laion-coco/, .3,6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 173, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 288, + 195 + ], + "type": "text", + "content": "[4] Midjourney. Accessed October 22, 2023 [Online] https://www.midjourney.com/home.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 197, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 219 + ], + "type": "text", + "content": "[5] Gen-2. Accessed October 22, 2023 [Online] https://research.runwayml.com/gen2,.2,6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 220, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 243 + ], + "type": "text", + "content": "[6] Genmo. Accessed October 22, 2023 [Online] https://www.genmo.ai/, .2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 244, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 244, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 244, + 288, + 266 + ], + "type": "text", + "content": "[7] Hotshot-x1. Accessed October 22, 2023 [Online] https://github.com/hotshotco/Hotshot-XL.2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 268, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 288, + 290 + ], + "type": "text", + "content": "[8] Moonvalley. Accessed October 22, 2023 [Online] https://moonvalley.ai/.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 292, + 288, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 314 + ], + "type": "text", + "content": "[9] Pika labs. Accessed October 22, 2023 [Online] https://www.pika.art/.2,6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 316, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 347 + ], + "type": "text", + "content": "[10] Realistic vision v2.0. Accessed October 22, 2023 [Online] https://huggingface.co/ckpt/realistic-vision-v20.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 350, + 288, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 394 + ], + "type": "text", + "content": "[11] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 396, + 288, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 288, + 429 + ], + "type": "text", + "content": "[12] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 430, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 288, + 485 + ], + "type": "text", + "content": "[13] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 487, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 531 + ], + "type": "text", + "content": "[14] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "type": "text", + "content": "[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 589, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 643 + ], + "type": "text", + "content": "[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[17] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, 35(1):53-65, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[18] Carl Doersch. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908, 2016. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[19] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, pages 102–118. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[20] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "text", + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "type": "text", + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 253, + 546, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 546, + 296 + ], + "type": "text", + "content": "[23] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "type": "text", + "content": "[24] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. arXiv preprint arXiv:2310.07702, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 354, + 547, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 375 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 431 + ], + "type": "text", + "content": "[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 433, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 464 + ], + "type": "text", + "content": "[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 466, + 547, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 547, + 509 + ], + "type": "text", + "content": "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 511, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 547, + 555 + ], + "type": "text", + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 556, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 547, + 589 + ], + "type": "text", + "content": "[30] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 590, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 635 + ], + "type": "text", + "content": "[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models, 2023. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 657 + ], + "type": "text", + "content": "[32] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In AAAI, pages 9117-9125, 2023. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[33] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7318" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[34] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "text", + "content": "[35] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 219 + ], + "type": "text", + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 288, + 263 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 266, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 288, + 300 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 288, + 356 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 358, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 288, + 391 + ], + "type": "text", + "content": "[40] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, pages 2830-2839, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 436 + ], + "type": "text", + "content": "[41] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 438, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 288, + 472 + ], + "type": "text", + "content": "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 474, + 288, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 288, + 517 + ], + "type": "text", + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 519, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 288, + 552 + ], + "type": "text", + "content": "[44] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 554, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 288, + 609 + ], + "type": "text", + "content": "[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 610, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 288, + 643 + ], + "type": "text", + "content": "[46] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 667 + ], + "type": "text", + "content": "[47] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 29, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "text", + "content": "[49] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[50] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "text", + "content": "[51] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "type": "text", + "content": "[52] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 329 + ], + "type": "text", + "content": "[53] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. arXiv preprint arXiv:2207.09814, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 331, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 386 + ], + "type": "text", + "content": "[54] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In CVPR, pages 20144-20154, 2023. 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "text", + "content": "[55] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 433, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 475 + ], + "type": "text", + "content": "[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 477, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 547, + 510 + ], + "type": "text", + "content": "[57] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "type": "text", + "content": "[58] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 556, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 547, + 589 + ], + "type": "text", + "content": "[59] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In CVPR, pages 18456-18466, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 634 + ], + "type": "text", + "content": "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation, 2023. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "text", + "content": "[61] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "type": "text", + "content": "[62] Yuechen Zhang, Jinbo Xing, Eric Lo, and Jiaya Jia. Realworld image variation by aligning diffusion inversion chain. arXiv preprint arXiv:2305.18729, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7319" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "text", + "content": "[63] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1, 2, 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7320" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_content_list.json b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd29f3eb1963a1002fc347b29ff61b1a006b7e48 --- /dev/null +++ b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_content_list.json @@ -0,0 +1,1382 @@ +[ + { + "type": "text", + "text": "VideoCutLER: Surprisingly Simple Unsupervised Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xudong Wang", + "bbox": [ + 133, + 180, + 251, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ishan Misra \nUC I", + "bbox": [ + 292, + 180, + 388, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyun Zeng \neley FAI", + "bbox": [ + 428, + 181, + 524, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Girdhar \nMeta AI", + "bbox": [ + 565, + 181, + 679, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trevor Darrell", + "bbox": [ + 717, + 181, + 834, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Code: https://github.com/facebookresearch/CutLER", + "bbox": [ + 328, + 219, + 637, + 232 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f68c0f203457877cd08e5f42d4db5c8f6ae6cc3a1767868743dc8aa6bf0ff24.jpg", + "image_caption": [ + "Learning Unsupervised Video Instance Segmentation", + "Figure 1. VideoCutLER is a simple unsupervised video instance segmentation method (UnVIS). We show the first competitive unsupervised results on the challenging YouTube VIS benchmark. Moreover, unlike most prior approaches, we demonstrate that UnVIS models can be learned without relying on natural videos and optical flow estimates. Row 1: We propose VideoCutLER, a simple cut-synthesis-and-learn pipeline that involves three main steps. Firstly, we generate pseudo-masks for multiple objects in an image using MaskCut [35]. Then, we convert a random pair of images in the minibatch into a video with corresponding pseudo mask trajectories using ImageCut2Video. Finally, we train an unsupervised video instance segmentation model using these mask trajectories. Row 2: Despite being trained only on unlabeled images, at inference time VideoCutLER can be directly applied to unseen videos and can segment and track multiple instances across time (Fig. 1a), even for small objects (Fig. 1b), objects that are absent in specific frames (Fig. 1c), and instances with high overlap (Fig. 1d). Column 2: Our method surpasses the previous SOTA method OCLR [37] by a factor of 10 in terms of class-agnostic $\\mathrm{AP}_{50}^{\\mathrm{video}}$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 262, + 883, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 578, + 313, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving $50.7\\%$ AP $_{50}$ ^ $\\text{video}$ , surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by $15.9\\%$ on YouTubeVIS-2019 in terms of AP $_{\\text{video}}$ .", + "bbox": [ + 75, + 626, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 578, + 630, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video instance segmentation is vital for various computer vision applications, e.g. video surveillance, autonomous driving, and video editing, yet labeled videos are costly to obtain. Hence, there is a pressing need to devise an unsupervised video instance segmentation approach that can comprehend video content comprehensively and operate in general domains without labels.", + "bbox": [ + 496, + 608, + 890, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior work in this area typically relies on an optical flow network as an off-the-shelf motion estimator [30, 37, 38]. Although optical flow can be informative in detecting pixel motion between frames, it is not always a reliable technique, particularly in the presence of occlusions, motion blur, complex motion patterns, changes in illuminations, etc. As a result, models that heavily rely on optical flow estimations may fail in several common scenarios. For example, stationary or slowly moving objects may have flow estimates similar to the background, causing them to be omitted in the segmentation process (e.g., the parrot with negligible motion is missed in Fig. 2a). Similarly, non", + "bbox": [ + 496, + 719, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22755", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4147f3b71f819a0116a06564cc88ba474d58ae395e6f08ae4d0409e8c15e59f.jpg", + "image_caption": [ + "Figure 2. Challenges encountered by the previous state-of-the-art OCLR: Within the framework of OCLR [37], a method that heavily relies on optical flows as model inputs, several distinct failure cases emerge. These include situations where the method struggles to accurately segment both moving and static objects (as demonstrated in Fig. 2a), struggles to effectively track non-rigid objects as a coherent unit (Fig. 2b), encounters difficulties in distinguishing overlapping instances (Fig. 2c), and fails to maintain consistent predictions under varying illumination conditions (Fig. 2d). Nonetheless, many of these challenges can be effectively addressed through the application of our proposed approach, VideoCutLER, without being reliant on the optical estimations used by various prior works [37, 38]. We present qualitative comparisons using the YouTubeVIS dataset [39]." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 468, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rigid objects with non-consistent motions for several parts have varying optical flows, leading to a failure in segmenting all parts cohesively as a unit if object motion is presumed constant (Fig. 2b). Also, objects with similar motion patterns and high overlap are complex for optic flow methods to accurately distinguish between them, especially in boundary regions (Fig. 2c). Finally, objects with illumination changes across frames can cause optical-flow based models to produce non-consistent and blurred segmentation masks (Fig. 2d). Given the limitations above, we advocate for unsupervised video segmentation models which do not", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "depend on optical flow estimates. We propose a method to train a video segmentation model by generating simple synthetic videos from individual images, without relying on explicit motion estimates or requiring labeled natural videos.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method, VideoCutLER, is an unsupervised Video instance segmentation model that employs a Cut-synthesis-and-LEaRn pipeline (Fig. 1). First, given unlabeled images, we extract pseudo-masks for multiple objects in an image using MaskCut [35], leveraging a self-supervised DINO [4] and a spectral clustering method Normalized Cuts [28] (details in Sec. 3.1). Second, given unlabeled images and their pseudo-masks in a minibatch, we propose ImageCut2Video, a surprisingly simple video synthesis scheme that generates a video from those with corresponding pseudo mask trajectories (details in Sec. 3.2). Finally, those mask trajectories are used to train a video instance segmentation model, aiming to perform object segmentation with temporal consistency across video frames (details in Sec. 3.3). Our model learns to segment and track object instances based on their appearance (feature) similarities across video frames.", + "bbox": [ + 496, + 152, + 892, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite being learned from only unlabeled images (and the temporally simple synthetic video sequences we construct from them), VideoCutLER succeeds at multi-instance video segmentation, achieving a new state-of-the-art (SOTA) performance of $50.7\\%$ AP $^{\\text{video}}$ on YouTubeVIS2019. This result surpasses the previous SOTA [37] by substantial margins of $45.9\\%$ ( $50.7\\%$ vs. $4.8\\%$ ). This result also considerably narrows the performance gap between supervised and unsupervised learning, reducing it from $29.1\\%$ to $11.0\\%$ in terms of the AP $^{\\text{video}}$", + "bbox": [ + 496, + 397, + 892, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, most prior works on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers being randomly initialized. In contrast, our pretraining strategy takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, our method outperforms DINO's [4] AP $^{\\text{video}}$ on YoutubeVIS-2019 by $15.9\\%$ .", + "bbox": [ + 496, + 551, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Insights: We found that a simple video synthesis method yield surprisingly effective results for training unsupervised multi-instance video segmentation models. This efficacy is achieved without the necessity of explicit motion estimates or the utilization of natural videos (relying solely on unlabeled ImageNet data suffices), a novel aspect that has not been previously demonstrated in the field. Methods: We propose a simple yet effective cut-synthesize-and-learn pipeline VideoCutLER for learning video instance segmentation models, given unlabeled images. Results: Our method shows the first successfully results on challenging unsupervised multi-instance video segmentation benchmark YouTubeVIS, outperforming the previous SOTA model's $\\mathrm{AP}_{50}^{\\mathrm{video}}$ by a large margin.", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22756", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/610b53dddac6bc544d2a14bf6d54bc86567bb77fcbdd0865f36313bc03f98686.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. We compare previous methods on unsupervised video instance segmentation, including CRW [17], DINO [4], and OCLR [37], with our VideoCutLER in term of key properties. Our VideoCutLER is the only approach that fulfills all these desired properties. $\\dagger$ : The optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data and human-annotated data like KITTI-2015 [18] and HD1K [19]." + ], + "table_body": "
CRWDINOOCLROurs
Segment multiple objects✓✗✓✓
Track objects across frames✓✗✓✓
No need for optical flow✓✓✗✓
No 1st-frame ground-truth✗✗✓✓
No human labels at any stage✗✗✓†✓
Pretrained model for sup. learning✗✓✗✓
", + "bbox": [ + 78, + 88, + 470, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 316, + 218, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised video instance segmentation (VIS) requires not only separating and tracking the main moving foreground objects from the background, but also differentiating between different instances, without any human annotations [32]. Previous works [16, 21, 36, 38, 40] on unsupervised video segmentation has primarily centered on unsupervised video object segmentation (VOS), aiming to detect all moving objects as the foreground and to generate a pixel-level binary segmentation mask, regardless of whether the scene contains a single instance or multiple instances. Despite some works exploring unsupervised video instance segmentation (VIS), many of these approaches have resorted to either utilizing first frame annotations [4, 17, 22] to propagate label information throughout the video frames or leveraging supervised learning using large amounts of external labeled data [25, 31, 41, 42]. Furthermore, prior studies typically utilized optical flow networks that were pretrained with human supervision using either synthetic data or labeled natural videos [31, 37, 38, 40].", + "bbox": [ + 75, + 340, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The properties deemed necessary for an unsupervised learning method to excel in video instance segmentation tasks are presented and discussed in Tab. 1. Our proposed method, VideoCutLER, is the only approach that satisfies all these properties, making it an effective and promising solution for unsupervised video instance segmentation.", + "bbox": [ + 75, + 628, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised object discovery aims to automatically discover and segment objects in an image in an unsupervised manner [16, 34-36]. LOST [29] and TokenCut [36] focus on salient object detection and segmentation via leveraging the patch features from a pretrained DINO [4] model. For multi-object discovery, FreeSOLO [34] first generates object pseudo-masks for unlabeled images, then learns an unsupervised instance segmentation model using these pseudo-masks. CutLER [35] presents a straightforward cut-and-learn pipeline for unsupervised detection and segmentation of multiple instances. It has demonstrated promising results on more than eleven different benchmarks, covering", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a wide range of domains.", + "bbox": [ + 500, + 92, + 668, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to previous approaches, our unsupervised learning method focuses on simultaneously tracking objects in a video sequence while identifying correspondences between instances across multiple frames.", + "bbox": [ + 496, + 106, + 890, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised representation learning generates its own supervision signal by exploiting the implicit patterns or structures present in the input data [3, 4, 14, 15]. Unlike most previous self-supervised learning models, which still require fine-tuning on labeled data to be operative on complex computer vision tasks, such as detection and segmentation, VideoCutLER can tackle these complex, challenging tasks with purely unsupervised learning methods.", + "bbox": [ + 496, + 167, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VideoCutLER", + "text_level": 1, + "bbox": [ + 500, + 299, + 645, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present VideoCutLER, a simple cut-synthesis-and-learn pipeline consisting of three main steps. First, we generate pseudo-masks for multiple objects in an image using MaskCut (Sec. 3.1). Next, we convert a random pair of images in the minibatch into a synthetic video with corresponding pseudo mask trajectories using ImageCut2Video (Sec. 3.2). Finally, we train an unsupervised video instance segmentation model using these mask trajectories. As the model inputs do not contain explicit motion estimates, it learns to track objects based on their appearance similarity (Sec. 3.3).", + "bbox": [ + 496, + 325, + 890, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Single-image unsupervised segmentation", + "text_level": 1, + "bbox": [ + 498, + 484, + 846, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ the MaskCut method, introduced in the CutLER [35] method. MaskCut is an efficient spectral clustering approach for unsupervised image instance segmentation and object detection and can discover multiple object masks in a single image without human supervision. MaskCut builds upon a self-supervised DINO model [2] with a backbone of ViT [10] and a cut-based clustering method Normalized Cuts (NCut) [28]. MaskCut first generates a patchwise affinity matrix $W_{ij} = \\frac{K_i K_j}{\\|K_i\\|_2\\|\\bar{K}_j\\|_2}$ using the 'key' features $K_{i}$ for patch $i$ from DINO's last attention layer. Subsequently, the NCut algorithm [28] is employed on the affinity matrix by solving a generalized eigenvalue problem", + "bbox": [ + 496, + 507, + 890, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(D - W) x = \\lambda D x \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 700, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D$ is a diagonal matrix with $d(i) = \\sum_{j} W_{ij}$ and $x$ is the eigenvector that corresponds to the second smallest eigenvalue $\\lambda$ . Then, the foreground masks $M^s$ can be extracted via bi-partitioning $x$ , which segments a single object within the image. To segment more than one object, Mask-Cut uses an iterative process that masks out the values in the affinity matrix using the extracted foreground mask:", + "bbox": [ + 496, + 726, + 890, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i j} ^ {t} = \\frac {\\left(K _ {i} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right) \\left(K _ {j} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right)}{\\| K _ {i} \\| _ {2} \\| K _ {j} \\| _ {2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 840, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and repeats the NCut algorithm. We set $t = 3$ by default.", + "bbox": [ + 500, + 885, + 870, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22757", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although MaskCut can effectively locate and segment multiple objects in an image, it operates only on a single image, lacking temporal consistency in the instance segmentation masks produced across video frames.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. ImageCut2Video Synthesis for Training", + "text_level": 1, + "bbox": [ + 76, + 162, + 418, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a learning-based approach to ensuring temporal consistency in video segmentation masks, based on generating synthetic videos from pairs of individual images and MaskCut masks. Surprisingly, we found that an extremely simple synthetic video generation method yields sufficient training data to learn a powerful video segmentation model that can operate on videos with much greater complexity of motion than is present in the training data.", + "bbox": [ + 75, + 186, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given unlabeled images in the minibatch and their pseudo-masks, our ImageCut2Video method synthesizes corresponding videos and pseudo-mask trajectories, thereby allowing us to train the model in an unsupervised manner while offering the necessary supervision for simultaneous detection, segmentation, and tracking of objects in videos.", + "bbox": [ + 75, + 308, + 468, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, given an image and its corresponding pseudomasks in the mini-batch, we duplicate the image $t$ times and connect its MaskCut pseudo-masks to form the initial trajectories. This synthetic video, however, only contains static foreground objects. To generate additional trajectories with mobile objects, a second image is randomly selected from the mini-batch, and its objects are cropped using its MaskCut pseudo-masks. These objects are then randomly resized, repositioned, and augmented before being pasted onto the first image. The resulting masks are connected along the temporal dimension to generate additional trajectories with mobile objects.", + "bbox": [ + 75, + 398, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given a target image $I_{1}$ , a random source image $I_{2}$ in the mini-batch and its corresponding set of binary pseudo-masks $\\{M_2^1,\\dots,M_2^s\\}$ , we first apply a transformation function $\\mathcal{T}$ to resize and shift these pseudomasks randomly. This gives us a new set of pseudo-masks $\\{\\hat{M}_2^1,\\dots,\\hat{M}_2^s\\}$ , where $\\hat{M}_2^s = \\mathcal{T}(M_2^s)$ . Next, we synthesize a video with $t$ frames by duplicating image $I_{1}$ for $t$ times and pasting the augmented masks onto $I_{1}$ using:", + "bbox": [ + 76, + 580, + 468, + 702 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {1} ^ {t} = I _ {1} \\times \\Pi_ {i = 1} ^ {s} (1 - \\hat {M} _ {2} ^ {i}) + I _ {2} \\times \\left(1 - \\Pi_ {i = 1} ^ {s} (1 - \\hat {M} _ {2} ^ {i})\\right) \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 713, + 468, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\times$ refers to element-wise multiplication.", + "bbox": [ + 76, + 744, + 387, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Video Segmentation Model", + "text_level": 1, + "bbox": [ + 76, + 770, + 320, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, the synthetic videos produced by Image-Cut2Video, comprising both mobile and stationary objects, are used as the inputs to train a video instance segmentation model. The segmentation mask trajectories corresponding to each object in the video serve as 'ground-truth' labels.", + "bbox": [ + 75, + 794, + 468, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize VideoMask2Former [6, 7] with a backbone of ResNet50 [12] as our video instance segmentation (VIS)", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model. It operates by attending to the 3D spatiotemporal features of our synthetic videos and generating 3D volume predictions of pseudo-mask trajectories using shared queries across frames. The shared queries across frames enable the model to segment and track object instances based on their appearance (feature) similarities, making it a powerful framework for analyzing video sequences.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 210, + 720, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VideoCutLER. We first employ the MaskCut approach on images preprocessed to a resolution of $480 \\times 480$ pixels. We then compute a patch-wise cosine similarity matrix using the pretrained ViT-Base/8 DINO [4] model, which serves as input to the MaskCut algorithm for initial segmentation mask generation. We set $t = 3$ , which is the maximum number of masks per image. To refine the segmentation masks, we employ a post-processing step using Conditional Random Fields (CRFs) [20], which enforces smoothness constraints and preserves object boundaries, resulting in improved segmentation masks.", + "bbox": [ + 496, + 234, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we use ImageCut2Video to synthetic videos given images and their pseudo-masks in a mini-batch. We found that synthetic videos with two frames are sufficient to train a video instance segmentation model; therefore, we use $s = 2$ by default. We randomly change the brightness, contrast, and rotation of the masks to create new variations of pseudo-masks. Additionally, we randomly resize the pseudo-masks (scale∈[0.8,1.0]), and shift their positions.", + "bbox": [ + 496, + 402, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training and test data. Our model is trained solely on the unlabeled images from ImageNet [8], which comprises approximately 1.3 million images. Without further fine-tuning on any video datasets, we test our model's zero-shot unsupervised video instance segmentation performance on four multi-instance video segmentation benchmarks, including YouTubeVIS-2019 [39], YouTubeVIS-2021 [39], DAVIS2017 [26], and DAVIS2017-Motion [26, 27].", + "bbox": [ + 496, + 523, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "YoutubeVIS-2019 and YouTube-VIS2021 contain 2,883 high-resolution YouTube videos and 3,859 high-resolution YouTube videos, respectively. We evaluate the zero-shot unsupervised learning performance on their training splits in a class-agnostic manner. For DAVIS-2017, we evaluate our model's performance on the 30 videos from its val set.", + "bbox": [ + 496, + 643, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training settings. 1) Unsupervised Image Model Pretraining: We first pretrain a Mask2Former [7] model with a backbone of ResNet50 [12] on ImageNet using MaskCut's pseudo-masks. The model is optimized for 160k iterations, with a batch size of 16 and a learning rate of 0.00002. The learning rate is decayed by a factor of 20 at iteration 80,000. To prevent overfitting, a dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders. 2) Unsupervised Video Model Learning: We initialize the VideoMask2Former model [6] with model weights from the previous stage, and then fine-tune", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22758", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dbfbca98d37a6e97de77d2e233a886e79fe0b97c1267501f50070fd789a90be2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTraining settingsYouTubeVIS-2019YouTubeVIS-2021
flowvideossup.\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_{S}\\)\\(AP_{M}\\)\\(AP_{L}\\)\\(AR_{10}\\)\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_{S}\\)\\(AP_{M}\\)\\(AP_{L}\\)\\(AR_{10}\\)
MotionGroup* [38]✓✓✘1.30.10.30.20.30.51.71.10.10.20.10.20.51.5
OCLR* [37]✓✓\\(X^†\\)4.80.41.30.01.25.511.04.40.31.20.11.67.19.6
CutLER‡✘✘✘37.514.617.13.313.927.630.429.210.412.83.112.827.822.6
VideoCutLER✘✓*✘50.724.226.05.620.937.942.438.919.017.15.318.337.531.3
vs. prev. SOTA+12.8+9.6+8.9+2.3+7.0+10.3+12.0+9.7+8.6+4.3+2.2+5.5+9.7+8.7
", + "bbox": [ + 80, + 88, + 890, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Zero-shot unsupervised multi-instance video segmentation on YouTubeVIS-2019 and YouTubeVIS-2021. We report the instance segmentation metrics (AP and AR) and training settings. *: reproduced MotionGroup [38] and OCLR [37] results with the official code and checkpoints. †: the optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data [1, 9] and human-annotated data, such as KITTI-2015 [18] and HD1K [19]. ‡: We train a CutLER [35] model with Mask2Former as a detector on ImageNet-1K, following CutLER's official training recipe, and use it as a strong baseline. *: VideoCutLER is trained on synthetic videos generated using ImageNet. Sup and flow denote human supervision and optical flow information, respectively. We evaluate results on YouTubeVIS's train splits in a class-agnostic manner (note: we never train on YouTubeVIS).", + "bbox": [ + 75, + 194, + 892, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "it on the synthetic videos we construct from ImageNet. We train VideoCutLER on 8 A100 GPUs for 80k iterations, using the AdamW optimizer [24]. We set the initial learning rate to 0.000005 and apply a learning rate multiplier of 0.1 to the backbone. A dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders.", + "bbox": [ + 75, + 319, + 467, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metric $\\mathbf{AP}^{\\mathrm{video}}$ and $\\mathbf{AR}^{\\mathrm{video}}$ : The evaluation metrics used in YouTubeVIS are Averaged Precision (AP) and Averaged Recall (AR), which are similar to those used in COCO [23]. The evaluation is specifically conducted at 10 intersection-over-union (IoU) thresholds ranging from $50\\%$ to $95\\%$ with a step of $5\\%$ [39]. However, unlike in image instance segmentation, each instance in a video comprises a sequence of masks, so the IoU computation is performed not only in the spatial domain, but also in the temporal domain by summing the intersections at every single frame over the unions at every single frame.", + "bbox": [ + 75, + 409, + 467, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metric $\\mathcal{J}$ and $\\mathcal{F}$ : For DAVIS [27], we report results using their official evaluation metrics $\\mathcal{J} \\& \\mathcal{F}$ , $\\mathcal{J}$ and $\\mathcal{F}$ . The region measure $(\\mathcal{J})$ [27] is the intersection-over-union (IoU) score between the algorithm's mask and the ground-truth mask. The boundary measure $(\\mathcal{F})$ [27] is the average precision of the boundary of the algorithm's mask. The evaluation metrics are computed separately for each instance, and then the results are averaged over all instances to get the final score. $\\mathcal{J} \\& \\mathcal{F}$ is the mean of $\\mathcal{J}$ and $\\mathcal{F}$ .", + "bbox": [ + 75, + 575, + 467, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of VideoCutLER on several video instance segmentation benchmarks. In Sec. 5.1, we demonstrate that our approach can effectively perform segmentation and tracking of multiple objects in videos, even when trained on unlabeled ImageNet images without any form of supervision. Our experimental results reveal that our method can drastically reduce the performance gap between unsupervised and supervised learning methods for video instance discovery and tracking. Furthermore, Sec. 5.2 demonstrates that fine-tuning VideoCutLER leads", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to further performance gains in video instance segmentation, surpassing previous works such as DINO in both fully supervised learning and semi-supervised learning tasks. In Sec. 5.3, we conduct an ablation study to examine the impact of key components and their hyperparameters.", + "bbox": [ + 496, + 319, + 890, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Unsupervised Zero-shot Evaluations", + "text_level": 1, + "bbox": [ + 498, + 406, + 815, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of our method against previous state-of-the-art approaches on various video instance segmentation benchmarks.", + "bbox": [ + 496, + 430, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluating unsupervised video instance segmentation poses two main challenges. Firstly, as unsupervised learning methods train the model without semantic classes, the class-aware video segmentation setup cannot be used directly for an evaluation. As a result, following previous works, we evaluate video instance segmentation results in a class-agnostic manner. Secondly, video instance segmentation datasets often annotate only a subset of the objects in the video, which makes Average Recall (AR) a valuable metric that does not penalize models for detecting novel objects not labeled in the dataset [35]. Therefore, we report both AR and AP for YouTubeVIS. Regarding DAVIS, we use the official unsupervised learning metrics $\\mathcal{I},\\mathcal{F}$ , and $\\mathcal{J}\\&\\mathcal{F}$ . All these metrics assess the performance of unsupervised video instance segmentation in a class-agnostic manner. Sec. 4 lists more details on evaluation metrics.", + "bbox": [ + 496, + 476, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detailed comparisons on YouTubeVIS. Tab. 2 presents a summary of the results for unsupervised zero-shot video instance segmentation on the YouTubeVIS-2019 and YouTubeVIS-2021 datasets. We compare our method's results with the previous state-of-the-art methods OCLR [37] and motion grouping [38]. We reproduce their results using their official code and checkpoints to ensure fairness.", + "bbox": [ + 496, + 718, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although OCLR [37] is also trained on synthetic videos, it relies on the off-the-shelf optical flow estimator RAFT [30] to compute optical flows for RGB sequences. It is worth noting that RAFT is pretrained on a combination of synthetic videos [1, 9] and human-annotated videos", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22759", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/09b3a8deab7f9414af636a6b73e588373d1c330d4312fa610303ba8598cf108e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Zero-shot unsupervised single/few-instance segmentation. VideoCutLER also outperforms the previous state-of-the-arts on DAVIS2017 and DAVIS2017-Motion. Note: 12 out of 30 videos from DAVIS2017 and 26 out of 30 videos from DAVIS2017-Motion contain only 1 moving instance. Additionally, DAVIS datasets focus solely on the performance of moving prominent objects, even in videos where multiple objects are present. This disadvantages our model since it can segment both static and moving objects and has not been exposed to any downstream videos during training. *: utilize optical flow predictions from RAFT [30], which is pretrained on external videos. All methods are evaluated in a zero-shot manner, i.e. no fine-tuning on target videos." + ], + "table_body": "
MethodsTraining settingsDAVIS2017DAVIS2017-Motion
flowvideossup.training dataJ&FJ(Mean)F(Mean)J&FJ(Mean)F(Mean)
MotionGroup (sup.) [38]✓✓XIN-1K+synthetic---39.544.934.2
Mask R-CNN (w/ flow)* [13, 37]✓✓XIN-1K+synthetic---50.350.450.2
OCLR (w/ flow)* [37]✓✓XIN-1K+synthetic39.638.241.155.154.555.7
VideoCutLERXXXIN-1K43.641.745.557.357.457.2
vs. prev. SOTA+4.0+3.5+4.4+2.2+2.9+1.5
", + "bbox": [ + 78, + 89, + 897, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "such as KITTI-2015 [18] and HD1K [19]. Our approach, VideoCutLER, despite not using any optical flow estimations like many previous works on unsupervised video segmentation, achieves over $10 \\times$ higher $\\mathrm{AP}_{50}$ and $18 \\times$ higher AP than OCLR [37] on YouTubeVIS-2019. Additionally, we achieve over $30\\%$ higher recall. Furthermore, unlike the previous state-of-the-art method OCLR [37], which exhibits poor performance in segmenting small objects (with $0.0\\%$ $\\mathrm{AP}_S$ ), our approach significantly outperforms it. Similar performance gains can be observed on YouTubeVIS-2021. Finally, the performance gains to CutLER [35] demonstrates the effectiveness of VideoCutLER in training unsupervised multi-instance video segmentation models, surpassing CutLER by over $12.8\\%$ on YouTubeVIS-2019.", + "bbox": [ + 75, + 301, + 472, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 3, we present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset.", + "bbox": [ + 75, + 516, + 470, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detailed comparisons on DAVIS. To provide a comprehensive evaluation and comparison with existing unsupervised video instance segmentation approaches, we also assess the performance of our model on the validation sets of DAVIS-2017 and DAVIS2017-Motion [27, 37]. Note that both DAVIS2017 and DAVIS2017-Motion datasets focus only on the performance of instance segmentation on prominent moving objects, even in videos with multiple objects. As a result, only a single or a few objects of interest per video are annotated, which may not reflect the challenges that arise when multiple objects are present.", + "bbox": [ + 75, + 565, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although the evaluation of DAVIS is an unfair assessment for us since VideoCutLER is supposed to segment both static and moving objects, whereas DAVIS focuses on moving prominent objects, with only a single or a few moving objects of interest per video annotated. However, Tab. 3 shows that VideoCutLER yields approximately $4\\%$ higher $\\mathcal{I}, \\mathcal{F}$ , and $\\mathcal{J} \\& \\mathcal{F}$ . The additional results on DAVIS demonstrate that VideoCutLER achieves superior performance not only on static or minimally moving objects but also on dynamic objects, where prior methods relying on optical flow estimates can benefit from additional cues.", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b41978343a5dfba7fb8df916cab7667635ebfd12e22f7b1e42678cbd948a70a8.jpg", + "image_caption": [ + "Figure 3. We present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset. It's noteworthy that VideoCutLER is solely pretrained on image dataset ImageNet-1K, and its evaluation is conducted directly on the video dataset YouTubeVIS (no further fine-tuning required). The visual results provided effectively highlight that VideoCutLER is capable of segmenting and tracking multiple instances, delivering consistent tracking results across video frames, and successfully distinguishing between various instances, even when significant overlapping occurs. We show more demo results in appendix." + ], + "image_footnote": [], + "bbox": [ + 521, + 299, + 870, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of supervised and unsupervised learning in object discovery and tracking abilities is presented in Tab. 4. We train a supervised MaskTrack R-CNN [39] model on the human-annotated training set of YouTubeVIS-2019 dataset, and evaluate it in a class-agnostic manner on the videos that are not shared between YouTubeVIS-2019 and YouTubeVIS-2021 datasets [39]. Tab. 4 shows that our VideoCutLER model significantly narrows the gap between supervised learning and unsupervised learning methods in terms of the averaged precision $\\mathrm{AP}_{50}$ (gaps: $29.1\\% \\rightarrow 11.0\\%$ ) and the averaged recall $\\mathrm{AR}_{100}$ (gaps: $14.9\\% \\rightarrow 3.2\\%$ ), particularly for the $\\mathrm{AR}_{100}$ .", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22760", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fdc73412d3f1da91b35c982f92d49781c16920b70b0762a6462f007c1fed3318.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTraining settingsYouTubeVIS-2021\\YouTubeVIS-2019
flowvideossup.training data\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)\\(AR_{100}\\)
Mask2Former [6]✓✓✘IN-1K+YT201948.922.224.9----
MaskTrack R-CNN* [39]✓✓✓IN-1K+YT201932.413.015.08.424.939.020.3
MaskTrack R-CNN* [39]✓✓✓IN-1K+COCO+YT201935.818.718.710.531.346.824.5
\\(OCLR^*\\)[37]✓✓✘IN-1K+synthetic3.30.21.00.32.77.55.4
VideoCutLER✘✘✘IN-1K21.47.19.04.913.329.617.1
vs. prev. SOTA+18.1+6.9+8.0+4.6+10.6+22.1+11.7
", + "bbox": [ + 80, + 88, + 890, + 199 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fef4583e4be72f779d10df3886701aaeb127110f1cb3ecbcbd15860afa468a9f.jpg", + "table_caption": [ + "Table 4. VideoCutLER greatly narrows the gap between fully-supervised learning and unsupervised learning for multi-instance video segmentation. Results are evaluated in a class-agnostic manner on the relative complement of the set of videos from YouTubeVIS-2021 and the set of videos from YouTubeVIS-2019. VideoCutLER and Mask2Former use a backbone of ResNet50. *: reproduced results with the official code and checkpoints. IN-1K refers to ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
MethodsArchitectureYouTubeVIS-2019YouTubeVIS-2021
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)
DINO [4]Mask2Former [6]23.039.023.76.028.034.224.641.425.98.734.039.9
VideoCutLERMask2Former [6]38.956.743.322.143.151.833.453.836.315.740.954.8
vs. prev. SOTA+15.9+17.7+19.6+16.1+15.1+17.6+8.8+12.4+10.4+7.0+6.9+14.9
", + "bbox": [ + 81, + 266, + 888, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. VideoCutLER can serve as a strong pretrained model for the supervised video instance segmentation task. The video segmentation model, Mask2Former, is initialized with various pretrained models, i.e., DINO or VideoCutLER, and fine-tuned on the training set with human annotations. We report the instance segmentation metrics and evaluate the model performance on the val splits.", + "bbox": [ + 75, + 345, + 892, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6efe8b2c7632c6b706d69504d99e206a5a445cebbdb508260abaa639a101be63.jpg", + "image_caption": [ + "Figure 4. We fine-tune VideoCutLER for semi-supervised video instance segmentation on the YouTubeVIS-2019 dataset, using different percentages of labeled training data. We evaluate the performance of our method by reporting the average precision and recall on the validation set of YouTubeVIS-2019. To establish a strong baseline, we use the self-supervised DINO [4] model and initialize the weights of VideoMask2Former with DINO. To ensure a fair comparison, both baselines and VideoCutLER are trained using the same schedule and recipe." + ], + "image_footnote": [], + "bbox": [ + 86, + 412, + 272, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c637624ef2de954a2e9705afd24045a701cabe73cd7241073af1215c72d1c979.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 412, + 459, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Label-Efficient and Fully-Supervised Learning", + "text_level": 1, + "bbox": [ + 76, + 724, + 470, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate VideoCutLER as a pretraining approach for supervised video instance segmentation models, and evaluate its effectiveness in label-efficient and fully-supervised learning scenarios.", + "bbox": [ + 75, + 750, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup. We use VideoMask2Former with a backbone of ResNet50 for all experiments in this section unless otherwise noted. For our experiments on semi-supervised learning, we randomly sample a subset of videos from the training split with different proportions of labeled videos. After pretraining our VideoCutLER model on ImageNet, we", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fine-tune the model on the YouTubeVIS-2019 [39] dataset with its human annotations. For our experiments on the fully-supervised learning task, we fine-tune the VideoCUTLER model on all available labeled data from the training sets of YouTubeVIS. For baselines, we initialize a VideoMask2Former model with a DINO [4] model pre-trained on ImageNet and fine-tuned on labeled videos. Since DINO has shown strong performance in detection and segmentation tasks, it serves as a strong baseline for our experiments.", + "bbox": [ + 496, + 414, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For semi-supervised learning, both the baselines and our models are trained for $2 \\times$ schedule, with a learning rate of 0.0001 for all model weights, except for the final classification layers, which use a learning rate of 0.0016. We train the models using a batch size of 16 and 8 GPUs. For fully-supervised learning, we use the $1 \\times$ schedule and a learning rate of 0.0002 for the final classification layers. We evaluate their performance on the val split of the YouTubeVIS-2019, and report results from its official evaluation server.", + "bbox": [ + 496, + 551, + 892, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data for fully-/semi-supervised VIS. We fine-tune the pretrained VideoCutLER model on all or a subset of the training split of YouTubeVIS-2019. We then evaluate the resulting models on the validation set. To ensure a fair comparison, we use the same amount of human annotations to train our model and baselines. Specifically, we initialize the baselines with the DINO-pretrained model and fine-tune them on the training set of the respective dataset. We evaluate the model performance on their validation sets and report results from its official evaluation server.", + "bbox": [ + 496, + 688, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Most prior approaches on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers, such as Mask2Former's decoders,", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22761", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d03c2d4a538f9fe12a35560cb7eb60fc1dd56634fdccec330bb129b769550f59.jpg", + "image_caption": [ + "Figure 5. We present qualitative results on videos covering a range of out-of-domain sources, e.g., sketches, 3D computer-generated imagery (CGI) and hybrid (CGI + realistic). VideoCutLER can produce high-quality segmentation and tracking results for small objects that are often difficult to distinguish from the background, as well as for object sketches that lack textual information." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "being randomly initialized. In contrast, VideoCutLER takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, as shown in Fig. 4 and Tab. 5, our method outperforms these prior works significantly, offering a strong pretrained model for fully-/semi-supervised learning tasks.", + "bbox": [ + 75, + 465, + 468, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 shows that VideoCutLER consistently outperforms the strong baseline method DINO [4] across all label-efficient learning settings with varying proportions of labeled YouTubeVIS-2019 videos. The most significant performance gains are observed when $20\\%$ labeled data is provided, where VideoCutLER exceeds DINO by over $12\\% \\mathrm{AP}_{50}$ and $13.2\\%$ AR. As demonstrated in Tab. 5, training the model with all available labeled videos from YouTubeVIS yields considerable performance gains, surpassing DINO by more than $15.9\\%$ AP on YouTubeVIS-2019 and $8.8\\%$ on YouTubeVIS-2021, respectively.", + "bbox": [ + 75, + 574, + 468, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 753, + 230, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameters and design choices. We present an ablation study on several key hyper-parameters and design choices of VideoCutLER in Tab. 6. First, we analyze the impact of varying the size of video frames used for training VideoCutLER. From Tab. 6a, we observe that the shortest edge length of 240 pixels yields the best performance. Using a larger resolution does not always lead to better results. Next, Tab. 6b shows the effect of the number of frames used", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/098114b25bbff4ca222615eb64b5b789735358ea064a0e30499875d2cefd7000.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Size → 180360480#frames → CutLER† [35]234
APvideo5049.950.7APvideo5037.549.850.750.4
(a) Frame size.(b) # frames. z
Augmentations → none+bright+rotation+contrast+cropall
APvideo5047.848.148.948.348.750.7
", + "bbox": [ + 509, + 87, + 888, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Data augmentations for ImageCut2Video.", + "bbox": [ + 549, + 175, + 836, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablations for VideoCutLER. We report video instance segmentation result $\\mathrm{AP}_{50}^{\\mathrm{video}}$ on YoutubeVIS-2019. (a) The impact of varying video frame sizes on training VideoCutLER. (b) The effect of the number of frames used for model training. (c) The impact of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. Default settings are highlighted in gray.", + "bbox": [ + 498, + 194, + 890, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for training video instance segmentation models. We found that synthetic videos with three frames are optimal for learning an unsupervised video instance segmentation model. Increasing the number of frames does not result in a further improved performance, aligning with the findings reported in [6]. Furthermore, Tab. 6c investigates the contribution of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. We found that compared to ImageCut2Video without any data augmentations, adding these augmentations can bring about $3\\%$ performance gains. Generalizability. The results presented in Fig. 5 demonstrate that VideoCutLER can effectively perform video instance segmentation on out-of-domain data sources, e.g. sketches, 3D computer-generated imagery, and hybrid videos that combine CGI. These results show that our model can be applied to a broad range of videos beyond the domains it was initially trained on, i.e., ImageNet.", + "bbox": [ + 496, + 316, + 892, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Summary", + "text_level": 1, + "bbox": [ + 500, + 604, + 607, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a simple unsupervised approach to segment multiple instances in a video. Our approach, VideoCutLER, does not require labels, and does not rely on motion-based learning signals like optical flow. In fact, VideoCutLER does not need real videos for training as we synthesize videos using natural images from the ImageNet-1K. Despite being simpler, VideoCutLER outperforms models that use additional learning signals or video data, achieving $10 \\times$ their performance on benchmarks like YouTubeVIS. Moreover, VideoCutLER is a strong pretrained model for supervised learning. We hope that our approach enables both a wide range of applications in video recognition, as well as its simplicity enables easy future research.", + "bbox": [ + 496, + 628, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: while VideoCutLER demonstrates its capability to achieve the state-of-the-art performance without relying on optical flow estimations, potential further improvements may be obtained by leveraging natural videos and integrating joint training with optical flow estimations.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22762", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 5", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020. 3", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 7, 8", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 7", + "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 4, 7, 8", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 4", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4", + "[9] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 5", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3", + "[11] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 7", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6", + "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 3, 7", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3", + "[16] Yuan-Ting Hu, Jia-Bin Huang, and Alexander G Schwing. Unsupervised video object segmentation using motion saliency-guided spatio-temporal propagation. In Proceedings of the European conference on computer vision (ECCV), pages 786-802, 2018. 3", + "[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 3", + "[18] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In Proceedings of the IEEE international conference on computer vision, pages 3271-3279, 2015. 3, 5, 6", + "[19] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The hci benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 19-28, 2016. 3, 5, 6", + "[20] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 2011. 4", + "[21] Minhyeok Lee, Suhwan Cho, Seunghoon Lee, Chaewon Park, and Sangyoun Lee. Unsupervised video object segmentation via prototype memory network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5924-5934, 2023. 3", + "[22] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[25] Jonathon Luiten, Idil Esen Zulfikar, and Bastian Leibe. Unovost: Unsupervised offline video object segmentation and tracking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2000-2009, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22763", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4", + "[27] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4, 5, 6", + "[28] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2, 3", + "[29] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3", + "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 1, 3, 5, 6", + "[31] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 3", + "[32] Wenguan Wang, Tianfei Zhou, Fatih Porikli, David Crandall, and Luc Van Gool. A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153, 2021. 3", + "[33] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 2, 7", + "[34] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 3", + "[35] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 3, 5, 6, 8", + "[36] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022. 3", + "[37] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 1, 2, 3, 5, 6", + "[39] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2, 4, 5, 6, 7", + "[40] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019. 3", + "[41] Zhao Yang, Qiang Wang, Song Bai, Weiming Hu, and Philip HS Torr. Video segmentation by detection for the 2019 unsupervised davis challenge. 2019. 3", + "[42] Tianfei Zhou, Jianwu Li, Shunzhou Wang, Ran Tao, and Jianbing Shen. Matnet: Motion-attentive transition network for zero-shot video object segmentation. IEEE Transactions on Image Processing, 29:8326-8338, 2020. 3" + ], + "bbox": [ + 501, + 92, + 893, + 388 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22764", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_model.json b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..08db7309ac0b38ec24f914402ef4756001be2b24 --- /dev/null +++ b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.885, + 0.155 + ], + "angle": 0, + "content": "VideoCutLER: Surprisingly Simple Unsupervised Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.181, + 0.253, + 0.2 + ], + "angle": 0, + "content": "Xudong Wang" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.181, + 0.39, + 0.218 + ], + "angle": 0, + "content": "Ishan Misra \nUC I" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.182, + 0.526, + 0.218 + ], + "angle": 0, + "content": "Ziyun Zeng \neley FAI" + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.182, + 0.68, + 0.218 + ], + "angle": 0, + "content": "Rohit Girdhar \nMeta AI" + }, + { + "type": "text", + "bbox": [ + 0.718, + 0.182, + 0.835, + 0.198 + ], + "angle": 0, + "content": "Trevor Darrell" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.22, + 0.638, + 0.233 + ], + "angle": 0, + "content": "Code: https://github.com/facebookresearch/CutLER" + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.246, + 0.388, + 0.261 + ], + "angle": 0, + "content": "Learning Unsupervised Video Instance Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.263, + 0.885, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.438, + 0.893, + 0.565 + ], + "angle": 0, + "content": "Figure 1. VideoCutLER is a simple unsupervised video instance segmentation method (UnVIS). We show the first competitive unsupervised results on the challenging YouTube VIS benchmark. Moreover, unlike most prior approaches, we demonstrate that UnVIS models can be learned without relying on natural videos and optical flow estimates. Row 1: We propose VideoCutLER, a simple cut-synthesis-and-learn pipeline that involves three main steps. Firstly, we generate pseudo-masks for multiple objects in an image using MaskCut [35]. Then, we convert a random pair of images in the minibatch into a video with corresponding pseudo mask trajectories using ImageCut2Video. Finally, we train an unsupervised video instance segmentation model using these mask trajectories. Row 2: Despite being trained only on unlabeled images, at inference time VideoCutLER can be directly applied to unseen videos and can segment and track multiple instances across time (Fig. 1a), even for small objects (Fig. 1b), objects that are absent in specific frames (Fig. 1c), and instances with high overlap (Fig. 1d). Column 2: Our method surpasses the previous SOTA method OCLR [37] by a factor of 10 in terms of class-agnostic \\(\\mathrm{AP}_{50}^{\\mathrm{video}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.579, + 0.314, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving \\(50.7\\%\\) AP\\(_{50}\\)^\\(\\text{video}\\), surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by \\(15.9\\%\\) on YouTubeVIS-2019 in terms of AP\\(_{\\text{video}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.579, + 0.631, + 0.594 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Video instance segmentation is vital for various computer vision applications, e.g. video surveillance, autonomous driving, and video editing, yet labeled videos are costly to obtain. Hence, there is a pressing need to devise an unsupervised video instance segmentation approach that can comprehend video content comprehensively and operate in general domains without labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Prior work in this area typically relies on an optical flow network as an off-the-shelf motion estimator [30, 37, 38]. Although optical flow can be informative in detecting pixel motion between frames, it is not always a reliable technique, particularly in the presence of occlusions, motion blur, complex motion patterns, changes in illuminations, etc. As a result, models that heavily rely on optical flow estimations may fail in several common scenarios. For example, stationary or slowly moving objects may have flow estimates similar to the background, causing them to be omitted in the segmentation process (e.g., the parrot with negligible motion is missed in Fig. 2a). Similarly, non" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22755" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.47, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Figure 2. Challenges encountered by the previous state-of-the-art OCLR: Within the framework of OCLR [37], a method that heavily relies on optical flows as model inputs, several distinct failure cases emerge. These include situations where the method struggles to accurately segment both moving and static objects (as demonstrated in Fig. 2a), struggles to effectively track non-rigid objects as a coherent unit (Fig. 2b), encounters difficulties in distinguishing overlapping instances (Fig. 2c), and fails to maintain consistent predictions under varying illumination conditions (Fig. 2d). Nonetheless, many of these challenges can be effectively addressed through the application of our proposed approach, VideoCutLER, without being reliant on the optical estimations used by various prior works [37, 38]. We present qualitative comparisons using the YouTubeVIS dataset [39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "rigid objects with non-consistent motions for several parts have varying optical flows, leading to a failure in segmenting all parts cohesively as a unit if object motion is presumed constant (Fig. 2b). Also, objects with similar motion patterns and high overlap are complex for optic flow methods to accurately distinguish between them, especially in boundary regions (Fig. 2c). Finally, objects with illumination changes across frames can cause optical-flow based models to produce non-consistent and blurred segmentation masks (Fig. 2d). Given the limitations above, we advocate for unsupervised video segmentation models which do not" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "depend on optical flow estimates. We propose a method to train a video segmentation model by generating simple synthetic videos from individual images, without relying on explicit motion estimates or requiring labeled natural videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Our method, VideoCutLER, is an unsupervised Video instance segmentation model that employs a Cut-synthesis-and-LEaRn pipeline (Fig. 1). First, given unlabeled images, we extract pseudo-masks for multiple objects in an image using MaskCut [35], leveraging a self-supervised DINO [4] and a spectral clustering method Normalized Cuts [28] (details in Sec. 3.1). Second, given unlabeled images and their pseudo-masks in a minibatch, we propose ImageCut2Video, a surprisingly simple video synthesis scheme that generates a video from those with corresponding pseudo mask trajectories (details in Sec. 3.2). Finally, those mask trajectories are used to train a video instance segmentation model, aiming to perform object segmentation with temporal consistency across video frames (details in Sec. 3.3). Our model learns to segment and track object instances based on their appearance (feature) similarities across video frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.893, + 0.55 + ], + "angle": 0, + "content": "Despite being learned from only unlabeled images (and the temporally simple synthetic video sequences we construct from them), VideoCutLER succeeds at multi-instance video segmentation, achieving a new state-of-the-art (SOTA) performance of \\(50.7\\%\\) AP\\(^{\\text{video}}\\) on YouTubeVIS2019. This result surpasses the previous SOTA [37] by substantial margins of \\(45.9\\%\\) (\\(50.7\\%\\) vs. \\(4.8\\%\\)). This result also considerably narrows the performance gap between supervised and unsupervised learning, reducing it from \\(29.1\\%\\) to \\(11.0\\%\\) in terms of the AP\\(^{\\text{video}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Moreover, most prior works on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers being randomly initialized. In contrast, our pretraining strategy takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, our method outperforms DINO's [4] AP\\(^{\\text{video}}\\) on YoutubeVIS-2019 by \\(15.9\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Contributions. Insights: We found that a simple video synthesis method yield surprisingly effective results for training unsupervised multi-instance video segmentation models. This efficacy is achieved without the necessity of explicit motion estimates or the utilization of natural videos (relying solely on unlabeled ImageNet data suffices), a novel aspect that has not been previously demonstrated in the field. Methods: We propose a simple yet effective cut-synthesize-and-learn pipeline VideoCutLER for learning video instance segmentation models, given unlabeled images. Results: Our method shows the first successfully results on challenging unsupervised multi-instance video segmentation benchmark YouTubeVIS, outperforming the previous SOTA model's \\(\\mathrm{AP}_{50}^{\\mathrm{video}}\\) by a large margin." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22756" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.191 + ], + "angle": 0, + "content": "
CRWDINOOCLROurs
Segment multiple objects✓✗✓✓
Track objects across frames✓✗✓✓
No need for optical flow✓✓✗✓
No 1st-frame ground-truth✗✗✓✓
No human labels at any stage✗✗✓†✓
Pretrained model for sup. learning✗✓✗✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.195, + 0.47, + 0.293 + ], + "angle": 0, + "content": "Table 1. We compare previous methods on unsupervised video instance segmentation, including CRW [17], DINO [4], and OCLR [37], with our VideoCutLER in term of key properties. Our VideoCutLER is the only approach that fulfills all these desired properties. \\(\\dagger\\): The optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data and human-annotated data like KITTI-2015 [18] and HD1K [19]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.317, + 0.22, + 0.332 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Unsupervised video instance segmentation (VIS) requires not only separating and tracking the main moving foreground objects from the background, but also differentiating between different instances, without any human annotations [32]. Previous works [16, 21, 36, 38, 40] on unsupervised video segmentation has primarily centered on unsupervised video object segmentation (VOS), aiming to detect all moving objects as the foreground and to generate a pixel-level binary segmentation mask, regardless of whether the scene contains a single instance or multiple instances. Despite some works exploring unsupervised video instance segmentation (VIS), many of these approaches have resorted to either utilizing first frame annotations [4, 17, 22] to propagate label information throughout the video frames or leveraging supervised learning using large amounts of external labeled data [25, 31, 41, 42]. Furthermore, prior studies typically utilized optical flow networks that were pretrained with human supervision using either synthetic data or labeled natural videos [31, 37, 38, 40]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.719 + ], + "angle": 0, + "content": "The properties deemed necessary for an unsupervised learning method to excel in video instance segmentation tasks are presented and discussed in Tab. 1. Our proposed method, VideoCutLER, is the only approach that satisfies all these properties, making it an effective and promising solution for unsupervised video instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Unsupervised object discovery aims to automatically discover and segment objects in an image in an unsupervised manner [16, 34-36]. LOST [29] and TokenCut [36] focus on salient object detection and segmentation via leveraging the patch features from a pretrained DINO [4] model. For multi-object discovery, FreeSOLO [34] first generates object pseudo-masks for unlabeled images, then learns an unsupervised instance segmentation model using these pseudo-masks. CutLER [35] presents a straightforward cut-and-learn pipeline for unsupervised detection and segmentation of multiple instances. It has demonstrated promising results on more than eleven different benchmarks, covering" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.669, + 0.106 + ], + "angle": 0, + "content": "a wide range of domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.167 + ], + "angle": 0, + "content": "In contrast to previous approaches, our unsupervised learning method focuses on simultaneously tracking objects in a video sequence while identifying correspondences between instances across multiple frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Self-supervised representation learning generates its own supervision signal by exploiting the implicit patterns or structures present in the input data [3, 4, 14, 15]. Unlike most previous self-supervised learning models, which still require fine-tuning on labeled data to be operative on complex computer vision tasks, such as detection and segmentation, VideoCutLER can tackle these complex, challenging tasks with purely unsupervised learning methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.3, + 0.646, + 0.315 + ], + "angle": 0, + "content": "3. VideoCutLER" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.478 + ], + "angle": 0, + "content": "We present VideoCutLER, a simple cut-synthesis-and-learn pipeline consisting of three main steps. First, we generate pseudo-masks for multiple objects in an image using MaskCut (Sec. 3.1). Next, we convert a random pair of images in the minibatch into a synthetic video with corresponding pseudo mask trajectories using ImageCut2Video (Sec. 3.2). Finally, we train an unsupervised video instance segmentation model using these mask trajectories. As the model inputs do not contain explicit motion estimates, it learns to track objects based on their appearance similarity (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.485, + 0.848, + 0.502 + ], + "angle": 0, + "content": "3.1. Single-image unsupervised segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.693 + ], + "angle": 0, + "content": "We employ the MaskCut method, introduced in the CutLER [35] method. MaskCut is an efficient spectral clustering approach for unsupervised image instance segmentation and object detection and can discover multiple object masks in a single image without human supervision. MaskCut builds upon a self-supervised DINO model [2] with a backbone of ViT [10] and a cut-based clustering method Normalized Cuts (NCut) [28]. MaskCut first generates a patchwise affinity matrix \\( W_{ij} = \\frac{K_i K_j}{\\|K_i\\|_2\\|\\bar{K}_j\\|_2} \\) using the 'key' features \\( K_{i} \\) for patch \\( i \\) from DINO's last attention layer. Subsequently, the NCut algorithm [28] is employed on the affinity matrix by solving a generalized eigenvalue problem" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.702, + 0.891, + 0.719 + ], + "angle": 0, + "content": "\\[\n(D - W) x = \\lambda D x \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.834 + ], + "angle": 0, + "content": "where \\(D\\) is a diagonal matrix with \\(d(i) = \\sum_{j} W_{ij}\\) and \\(x\\) is the eigenvector that corresponds to the second smallest eigenvalue \\(\\lambda\\). Then, the foreground masks \\(M^s\\) can be extracted via bi-partitioning \\(x\\), which segments a single object within the image. To segment more than one object, Mask-Cut uses an iterative process that masks out the values in the affinity matrix using the extracted foreground mask:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.842, + 0.891, + 0.879 + ], + "angle": 0, + "content": "\\[\nW _ {i j} ^ {t} = \\frac {\\left(K _ {i} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right) \\left(K _ {j} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right)}{\\| K _ {i} \\| _ {2} \\| K _ {j} \\| _ {2}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.871, + 0.902 + ], + "angle": 0, + "content": "and repeats the NCut algorithm. We set \\( t = 3 \\) by default." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22757" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "Although MaskCut can effectively locate and segment multiple objects in an image, it operates only on a single image, lacking temporal consistency in the instance segmentation masks produced across video frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.163, + 0.419, + 0.181 + ], + "angle": 0, + "content": "3.2. ImageCut2Video Synthesis for Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.469, + 0.307 + ], + "angle": 0, + "content": "We propose a learning-based approach to ensuring temporal consistency in video segmentation masks, based on generating synthetic videos from pairs of individual images and MaskCut masks. Surprisingly, we found that an extremely simple synthetic video generation method yields sufficient training data to learn a powerful video segmentation model that can operate on videos with much greater complexity of motion than is present in the training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.469, + 0.399 + ], + "angle": 0, + "content": "Given unlabeled images in the minibatch and their pseudo-masks, our ImageCut2Video method synthesizes corresponding videos and pseudo-mask trajectories, thereby allowing us to train the model in an unsupervised manner while offering the necessary supervision for simultaneous detection, segmentation, and tracking of objects in videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.469, + 0.58 + ], + "angle": 0, + "content": "First, given an image and its corresponding pseudomasks in the mini-batch, we duplicate the image \\( t \\) times and connect its MaskCut pseudo-masks to form the initial trajectories. This synthetic video, however, only contains static foreground objects. To generate additional trajectories with mobile objects, a second image is randomly selected from the mini-batch, and its objects are cropped using its MaskCut pseudo-masks. These objects are then randomly resized, repositioned, and augmented before being pasted onto the first image. The resulting masks are connected along the temporal dimension to generate additional trajectories with mobile objects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.469, + 0.703 + ], + "angle": 0, + "content": "Specifically, given a target image \\(I_{1}\\), a random source image \\(I_{2}\\) in the mini-batch and its corresponding set of binary pseudo-masks \\(\\{M_2^1,\\dots,M_2^s\\}\\), we first apply a transformation function \\(\\mathcal{T}\\) to resize and shift these pseudomasks randomly. This gives us a new set of pseudo-masks \\(\\{\\hat{M}_2^1,\\dots,\\hat{M}_2^s\\}\\), where \\(\\hat{M}_2^s = \\mathcal{T}(M_2^s)\\). Next, we synthesize a video with \\(t\\) frames by duplicating image \\(I_{1}\\) for \\(t\\) times and pasting the augmented masks onto \\(I_{1}\\) using:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.714, + 0.469, + 0.733 + ], + "angle": 0, + "content": "\\[\nI _ {1} ^ {t} = I _ {1} \\times \\Pi_ {i = 1} ^ {s} (1 - \\hat {M} _ {2} ^ {i}) + I _ {2} \\times \\left(1 - \\Pi_ {i = 1} ^ {s} (1 - \\hat {M} _ {2} ^ {i})\\right) \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.745, + 0.388, + 0.76 + ], + "angle": 0, + "content": "where \\(\\times\\) refers to element-wise multiplication." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.321, + 0.787 + ], + "angle": 0, + "content": "3.3. Video Segmentation Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.869 + ], + "angle": 0, + "content": "During training, the synthetic videos produced by Image-Cut2Video, comprising both mobile and stationary objects, are used as the inputs to train a video instance segmentation model. The segmentation mask trajectories corresponding to each object in the video serve as 'ground-truth' labels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We utilize VideoMask2Former [6, 7] with a backbone of ResNet50 [12] as our video instance segmentation (VIS)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "model. It operates by attending to the 3D spatiotemporal features of our synthetic videos and generating 3D volume predictions of pseudo-mask trajectories using shared queries across frames. The shared queries across frames enable the model to segment and track object instances based on their appearance (feature) similarities, making it a powerful framework for analyzing video sequences." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.211, + 0.722, + 0.228 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.403 + ], + "angle": 0, + "content": "VideoCutLER. We first employ the MaskCut approach on images preprocessed to a resolution of \\(480 \\times 480\\) pixels. We then compute a patch-wise cosine similarity matrix using the pretrained ViT-Base/8 DINO [4] model, which serves as input to the MaskCut algorithm for initial segmentation mask generation. We set \\(t = 3\\), which is the maximum number of masks per image. To refine the segmentation masks, we employ a post-processing step using Conditional Random Fields (CRFs) [20], which enforces smoothness constraints and preserves object boundaries, resulting in improved segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.524 + ], + "angle": 0, + "content": "Next, we use ImageCut2Video to synthetic videos given images and their pseudo-masks in a mini-batch. We found that synthetic videos with two frames are sufficient to train a video instance segmentation model; therefore, we use \\( s = 2 \\) by default. We randomly change the brightness, contrast, and rotation of the masks to create new variations of pseudo-masks. Additionally, we randomly resize the pseudo-masks (scale∈[0.8,1.0]), and shift their positions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Training and test data. Our model is trained solely on the unlabeled images from ImageNet [8], which comprises approximately 1.3 million images. Without further fine-tuning on any video datasets, we test our model's zero-shot unsupervised video instance segmentation performance on four multi-instance video segmentation benchmarks, including YouTubeVIS-2019 [39], YouTubeVIS-2021 [39], DAVIS2017 [26], and DAVIS2017-Motion [26, 27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.735 + ], + "angle": 0, + "content": "YoutubeVIS-2019 and YouTube-VIS2021 contain 2,883 high-resolution YouTube videos and 3,859 high-resolution YouTube videos, respectively. We evaluate the zero-shot unsupervised learning performance on their training splits in a class-agnostic manner. For DAVIS-2017, we evaluate our model's performance on the 30 videos from its val set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training settings. 1) Unsupervised Image Model Pretraining: We first pretrain a Mask2Former [7] model with a backbone of ResNet50 [12] on ImageNet using MaskCut's pseudo-masks. The model is optimized for 160k iterations, with a batch size of 16 and a learning rate of 0.00002. The learning rate is decayed by a factor of 20 at iteration 80,000. To prevent overfitting, a dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders. 2) Unsupervised Video Model Learning: We initialize the VideoMask2Former model [6] with model weights from the previous stage, and then fine-tune" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22758" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.193 + ], + "angle": 0, + "content": "
MethodsTraining settingsYouTubeVIS-2019YouTubeVIS-2021
flowvideossup.\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_{S}\\)\\(AP_{M}\\)\\(AP_{L}\\)\\(AR_{10}\\)\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_{S}\\)\\(AP_{M}\\)\\(AP_{L}\\)\\(AR_{10}\\)
MotionGroup* [38]✓✓✘1.30.10.30.20.30.51.71.10.10.20.10.20.51.5
OCLR* [37]✓✓\\(X^†\\)4.80.41.30.01.25.511.04.40.31.20.11.67.19.6
CutLER‡✘✘✘37.514.617.13.313.927.630.429.210.412.83.112.827.822.6
VideoCutLER✘✓*✘50.724.226.05.620.937.942.438.919.017.15.318.337.531.3
vs. prev. SOTA+12.8+9.6+8.9+2.3+7.0+10.3+12.0+9.7+8.6+4.3+2.2+5.5+9.7+8.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.893, + 0.295 + ], + "angle": 0, + "content": "Table 2. Zero-shot unsupervised multi-instance video segmentation on YouTubeVIS-2019 and YouTubeVIS-2021. We report the instance segmentation metrics (AP and AR) and training settings. *: reproduced MotionGroup [38] and OCLR [37] results with the official code and checkpoints. †: the optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data [1, 9] and human-annotated data, such as KITTI-2015 [18] and HD1K [19]. ‡: We train a CutLER [35] model with Mask2Former as a detector on ImageNet-1K, following CutLER's official training recipe, and use it as a strong baseline. *: VideoCutLER is trained on synthetic videos generated using ImageNet. Sup and flow denote human supervision and optical flow information, respectively. We evaluate results on YouTubeVIS's train splits in a class-agnostic manner (note: we never train on YouTubeVIS)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.41 + ], + "angle": 0, + "content": "it on the synthetic videos we construct from ImageNet. We train VideoCutLER on 8 A100 GPUs for 80k iterations, using the AdamW optimizer [24]. We set the initial learning rate to 0.000005 and apply a learning rate multiplier of 0.1 to the backbone. A dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.468, + 0.576 + ], + "angle": 0, + "content": "Evaluation metric \\(\\mathbf{AP}^{\\mathrm{video}}\\) and \\(\\mathbf{AR}^{\\mathrm{video}}\\): The evaluation metrics used in YouTubeVIS are Averaged Precision (AP) and Averaged Recall (AR), which are similar to those used in COCO [23]. The evaluation is specifically conducted at 10 intersection-over-union (IoU) thresholds ranging from \\(50\\%\\) to \\(95\\%\\) with a step of \\(5\\%\\) [39]. However, unlike in image instance segmentation, each instance in a video comprises a sequence of masks, so the IoU computation is performed not only in the spatial domain, but also in the temporal domain by summing the intersections at every single frame over the unions at every single frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.468, + 0.712 + ], + "angle": 0, + "content": "Evaluation metric \\(\\mathcal{J}\\) and \\(\\mathcal{F}\\): For DAVIS [27], we report results using their official evaluation metrics \\(\\mathcal{J} \\& \\mathcal{F}\\), \\(\\mathcal{J}\\) and \\(\\mathcal{F}\\). The region measure \\((\\mathcal{J})\\) [27] is the intersection-over-union (IoU) score between the algorithm's mask and the ground-truth mask. The boundary measure \\((\\mathcal{F})\\) [27] is the average precision of the boundary of the algorithm's mask. The evaluation metrics are computed separately for each instance, and then the results are averaged over all instances to get the final score. \\(\\mathcal{J} \\& \\mathcal{F}\\) is the mean of \\(\\mathcal{J}\\) and \\(\\mathcal{F}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.21, + 0.743 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We evaluate the performance of VideoCutLER on several video instance segmentation benchmarks. In Sec. 5.1, we demonstrate that our approach can effectively perform segmentation and tracking of multiple objects in videos, even when trained on unlabeled ImageNet images without any form of supervision. Our experimental results reveal that our method can drastically reduce the performance gap between unsupervised and supervised learning methods for video instance discovery and tracking. Furthermore, Sec. 5.2 demonstrates that fine-tuning VideoCutLER leads" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.397 + ], + "angle": 0, + "content": "to further performance gains in video instance segmentation, surpassing previous works such as DINO in both fully supervised learning and semi-supervised learning tasks. In Sec. 5.3, we conduct an ablation study to examine the impact of key components and their hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.407, + 0.816, + 0.423 + ], + "angle": 0, + "content": "5.1. Unsupervised Zero-shot Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.476 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of our method against previous state-of-the-art approaches on various video instance segmentation benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Evaluating unsupervised video instance segmentation poses two main challenges. Firstly, as unsupervised learning methods train the model without semantic classes, the class-aware video segmentation setup cannot be used directly for an evaluation. As a result, following previous works, we evaluate video instance segmentation results in a class-agnostic manner. Secondly, video instance segmentation datasets often annotate only a subset of the objects in the video, which makes Average Recall (AR) a valuable metric that does not penalize models for detecting novel objects not labeled in the dataset [35]. Therefore, we report both AR and AP for YouTubeVIS. Regarding DAVIS, we use the official unsupervised learning metrics \\(\\mathcal{I},\\mathcal{F}\\), and \\(\\mathcal{J}\\&\\mathcal{F}\\). All these metrics assess the performance of unsupervised video instance segmentation in a class-agnostic manner. Sec. 4 lists more details on evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Detailed comparisons on YouTubeVIS. Tab. 2 presents a summary of the results for unsupervised zero-shot video instance segmentation on the YouTubeVIS-2019 and YouTubeVIS-2021 datasets. We compare our method's results with the previous state-of-the-art methods OCLR [37] and motion grouping [38]. We reproduce their results using their official code and checkpoints to ensure fairness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although OCLR [37] is also trained on synthetic videos, it relies on the off-the-shelf optical flow estimator RAFT [30] to compute optical flows for RGB sequences. It is worth noting that RAFT is pretrained on a combination of synthetic videos [1, 9] and human-annotated videos" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22759" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.898, + 0.188 + ], + "angle": 0, + "content": "
MethodsTraining settingsDAVIS2017DAVIS2017-Motion
flowvideossup.training dataJ&FJ(Mean)F(Mean)J&FJ(Mean)F(Mean)
MotionGroup (sup.) [38]✓✓XIN-1K+synthetic---39.544.934.2
Mask R-CNN (w/ flow)* [13, 37]✓✓XIN-1K+synthetic---50.350.450.2
OCLR (w/ flow)* [37]✓✓XIN-1K+synthetic39.638.241.155.154.555.7
VideoCutLERXXXIN-1K43.641.745.557.357.457.2
vs. prev. SOTA+4.0+3.5+4.4+2.2+2.9+1.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.192, + 0.894, + 0.276 + ], + "angle": 0, + "content": "Table 3. Zero-shot unsupervised single/few-instance segmentation. VideoCutLER also outperforms the previous state-of-the-arts on DAVIS2017 and DAVIS2017-Motion. Note: 12 out of 30 videos from DAVIS2017 and 26 out of 30 videos from DAVIS2017-Motion contain only 1 moving instance. Additionally, DAVIS datasets focus solely on the performance of moving prominent objects, even in videos where multiple objects are present. This disadvantages our model since it can segment both static and moving objects and has not been exposed to any downstream videos during training. *: utilize optical flow predictions from RAFT [30], which is pretrained on external videos. All methods are evaluated in a zero-shot manner, i.e. no fine-tuning on target videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.514 + ], + "angle": 0, + "content": "such as KITTI-2015 [18] and HD1K [19]. Our approach, VideoCutLER, despite not using any optical flow estimations like many previous works on unsupervised video segmentation, achieves over \\(10 \\times\\) higher \\(\\mathrm{AP}_{50}\\) and \\(18 \\times\\) higher AP than OCLR [37] on YouTubeVIS-2019. Additionally, we achieve over \\(30\\%\\) higher recall. Furthermore, unlike the previous state-of-the-art method OCLR [37], which exhibits poor performance in segmenting small objects (with \\(0.0\\%\\) \\(\\mathrm{AP}_S\\)), our approach significantly outperforms it. Similar performance gains can be observed on YouTubeVIS-2021. Finally, the performance gains to CutLER [35] demonstrates the effectiveness of VideoCutLER in training unsupervised multi-instance video segmentation models, surpassing CutLER by over \\(12.8\\%\\) on YouTubeVIS-2019." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.471, + 0.562 + ], + "angle": 0, + "content": "In Fig. 3, we present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Detailed comparisons on DAVIS. To provide a comprehensive evaluation and comparison with existing unsupervised video instance segmentation approaches, we also assess the performance of our model on the validation sets of DAVIS-2017 and DAVIS2017-Motion [27, 37]. Note that both DAVIS2017 and DAVIS2017-Motion datasets focus only on the performance of instance segmentation on prominent moving objects, even in videos with multiple objects. As a result, only a single or a few objects of interest per video are annotated, which may not reflect the challenges that arise when multiple objects are present." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Although the evaluation of DAVIS is an unfair assessment for us since VideoCutLER is supposed to segment both static and moving objects, whereas DAVIS focuses on moving prominent objects, with only a single or a few moving objects of interest per video annotated. However, Tab. 3 shows that VideoCutLER yields approximately \\(4\\%\\) higher \\(\\mathcal{I}, \\mathcal{F}\\), and \\(\\mathcal{J} \\& \\mathcal{F}\\). The additional results on DAVIS demonstrate that VideoCutLER achieves superior performance not only on static or minimally moving objects but also on dynamic objects, where prior methods relying on optical flow estimates can benefit from additional cues." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.3, + 0.872, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.537, + 0.895, + 0.69 + ], + "angle": 0, + "content": "Figure 3. We present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset. It's noteworthy that VideoCutLER is solely pretrained on image dataset ImageNet-1K, and its evaluation is conducted directly on the video dataset YouTubeVIS (no further fine-tuning required). The visual results provided effectively highlight that VideoCutLER is capable of segmenting and tracking multiple instances, delivering consistent tracking results across video frames, and successfully distinguishing between various instances, even when significant overlapping occurs. We show more demo results in appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Comparison of supervised and unsupervised learning in object discovery and tracking abilities is presented in Tab. 4. We train a supervised MaskTrack R-CNN [39] model on the human-annotated training set of YouTubeVIS-2019 dataset, and evaluate it in a class-agnostic manner on the videos that are not shared between YouTubeVIS-2019 and YouTubeVIS-2021 datasets [39]. Tab. 4 shows that our VideoCutLER model significantly narrows the gap between supervised learning and unsupervised learning methods in terms of the averaged precision \\(\\mathrm{AP}_{50}\\) (gaps: \\(29.1\\% \\rightarrow 11.0\\%\\)) and the averaged recall \\(\\mathrm{AR}_{100}\\) (gaps: \\(14.9\\% \\rightarrow 3.2\\%\\)), particularly for the \\(\\mathrm{AR}_{100}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22760" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.2 + ], + "angle": 0, + "content": "
MethodsTraining settingsYouTubeVIS-2021\\YouTubeVIS-2019
flowvideossup.training data\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)\\(AR_{100}\\)
Mask2Former [6]✓✓✘IN-1K+YT201948.922.224.9----
MaskTrack R-CNN* [39]✓✓✓IN-1K+YT201932.413.015.08.424.939.020.3
MaskTrack R-CNN* [39]✓✓✓IN-1K+COCO+YT201935.818.718.710.531.346.824.5
\\(OCLR^*\\)[37]✓✓✘IN-1K+synthetic3.30.21.00.32.77.55.4
VideoCutLER✘✘✘IN-1K21.47.19.04.913.329.617.1
vs. prev. SOTA+18.1+6.9+8.0+4.6+10.6+22.1+11.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.203, + 0.893, + 0.261 + ], + "angle": 0, + "content": "Table 4. VideoCutLER greatly narrows the gap between fully-supervised learning and unsupervised learning for multi-instance video segmentation. Results are evaluated in a class-agnostic manner on the relative complement of the set of videos from YouTubeVIS-2021 and the set of videos from YouTubeVIS-2019. VideoCutLER and Mask2Former use a backbone of ResNet50. *: reproduced results with the official code and checkpoints. IN-1K refers to ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.267, + 0.889, + 0.344 + ], + "angle": 0, + "content": "
MethodsArchitectureYouTubeVIS-2019YouTubeVIS-2021
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)
DINO [4]Mask2Former [6]23.039.023.76.028.034.224.641.425.98.734.039.9
VideoCutLERMask2Former [6]38.956.743.322.143.151.833.453.836.315.740.954.8
vs. prev. SOTA+15.9+17.7+19.6+16.1+15.1+17.6+8.8+12.4+10.4+7.0+6.9+14.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.346, + 0.893, + 0.391 + ], + "angle": 0, + "content": "Table 5. VideoCutLER can serve as a strong pretrained model for the supervised video instance segmentation task. The video segmentation model, Mask2Former, is initialized with various pretrained models, i.e., DINO or VideoCutLER, and fine-tuned on the training set with human annotations. We report the instance segmentation metrics and evaluate the model performance on the val splits." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.414, + 0.273, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.414, + 0.46, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Figure 4. We fine-tune VideoCutLER for semi-supervised video instance segmentation on the YouTubeVIS-2019 dataset, using different percentages of labeled training data. We evaluate the performance of our method by reporting the average precision and recall on the validation set of YouTubeVIS-2019. To establish a strong baseline, we use the self-supervised DINO [4] model and initialize the weights of VideoMask2Former with DINO. To ensure a fair comparison, both baselines and VideoCutLER are trained using the same schedule and recipe." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.471, + 0.743 + ], + "angle": 0, + "content": "5.2. Label-Efficient and Fully-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.81 + ], + "angle": 0, + "content": "In this section, we investigate VideoCutLER as a pretraining approach for supervised video instance segmentation models, and evaluate its effectiveness in label-efficient and fully-supervised learning scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Setup. We use VideoMask2Former with a backbone of ResNet50 for all experiments in this section unless otherwise noted. For our experiments on semi-supervised learning, we randomly sample a subset of videos from the training split with different proportions of labeled videos. After pretraining our VideoCutLER model on ImageNet, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.552 + ], + "angle": 0, + "content": "fine-tune the model on the YouTubeVIS-2019 [39] dataset with its human annotations. For our experiments on the fully-supervised learning task, we fine-tune the VideoCUTLER model on all available labeled data from the training sets of YouTubeVIS. For baselines, we initialize a VideoMask2Former model with a DINO [4] model pre-trained on ImageNet and fine-tuned on labeled videos. Since DINO has shown strong performance in detection and segmentation tasks, it serves as a strong baseline for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.688 + ], + "angle": 0, + "content": "For semi-supervised learning, both the baselines and our models are trained for \\(2 \\times\\) schedule, with a learning rate of 0.0001 for all model weights, except for the final classification layers, which use a learning rate of 0.0016. We train the models using a batch size of 16 and 8 GPUs. For fully-supervised learning, we use the \\(1 \\times\\) schedule and a learning rate of 0.0002 for the final classification layers. We evaluate their performance on the val split of the YouTubeVIS-2019, and report results from its official evaluation server." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Data for fully-/semi-supervised VIS. We fine-tune the pretrained VideoCutLER model on all or a subset of the training split of YouTubeVIS-2019. We then evaluate the resulting models on the validation set. To ensure a fair comparison, we use the same amount of human annotations to train our model and baselines. Specifically, we initialize the baselines with the DINO-pretrained model and fine-tune them on the training set of the respective dataset. We evaluate the model performance on their validation sets and report results from its official evaluation server." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Results. Most prior approaches on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers, such as Mask2Former's decoders," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22761" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.437 + ], + "angle": 0, + "content": "Figure 5. We present qualitative results on videos covering a range of out-of-domain sources, e.g., sketches, 3D computer-generated imagery (CGI) and hybrid (CGI + realistic). VideoCutLER can produce high-quality segmentation and tracking results for small objects that are often difficult to distinguish from the background, as well as for object sketches that lack textual information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.573 + ], + "angle": 0, + "content": "being randomly initialized. In contrast, VideoCutLER takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, as shown in Fig. 4 and Tab. 5, our method outperforms these prior works significantly, offering a strong pretrained model for fully-/semi-supervised learning tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Fig. 4 shows that VideoCutLER consistently outperforms the strong baseline method DINO [4] across all label-efficient learning settings with varying proportions of labeled YouTubeVIS-2019 videos. The most significant performance gains are observed when \\(20\\%\\) labeled data is provided, where VideoCutLER exceeds DINO by over \\(12\\% \\mathrm{AP}_{50}\\) and \\(13.2\\%\\) AR. As demonstrated in Tab. 5, training the model with all available labeled videos from YouTubeVIS yields considerable performance gains, surpassing DINO by more than \\(15.9\\%\\) AP on YouTubeVIS-2019 and \\(8.8\\%\\) on YouTubeVIS-2021, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.231, + 0.772 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Hyper-parameters and design choices. We present an ablation study on several key hyper-parameters and design choices of VideoCutLER in Tab. 6. First, we analyze the impact of varying the size of video frames used for training VideoCutLER. From Tab. 6a, we observe that the shortest edge length of 240 pixels yields the best performance. Using a larger resolution does not always lead to better results. Next, Tab. 6b shows the effect of the number of frames used" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.088, + 0.89, + 0.174 + ], + "angle": 0, + "content": "
Size → 180360480#frames → CutLER† [35]234
APvideo5049.950.7APvideo5037.549.850.750.4
(a) Frame size.(b) # frames. z
Augmentations → none+bright+rotation+contrast+cropall
APvideo5047.848.148.948.348.750.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.176, + 0.837, + 0.19 + ], + "angle": 0, + "content": "(c) Data augmentations for ImageCut2Video." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.195, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Table 6. Ablations for VideoCutLER. We report video instance segmentation result \\(\\mathrm{AP}_{50}^{\\mathrm{video}}\\) on YoutubeVIS-2019. (a) The impact of varying video frame sizes on training VideoCutLER. (b) The effect of the number of frames used for model training. (c) The impact of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. Default settings are highlighted in gray." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.893, + 0.591 + ], + "angle": 0, + "content": "for training video instance segmentation models. We found that synthetic videos with three frames are optimal for learning an unsupervised video instance segmentation model. Increasing the number of frames does not result in a further improved performance, aligning with the findings reported in [6]. Furthermore, Tab. 6c investigates the contribution of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. We found that compared to ImageCut2Video without any data augmentations, adding these augmentations can bring about \\(3\\%\\) performance gains. Generalizability. The results presented in Fig. 5 demonstrate that VideoCutLER can effectively perform video instance segmentation on out-of-domain data sources, e.g. sketches, 3D computer-generated imagery, and hybrid videos that combine CGI. These results show that our model can be applied to a broad range of videos beyond the domains it was initially trained on, i.e., ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.608, + 0.621 + ], + "angle": 0, + "content": "6. Summary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We presented a simple unsupervised approach to segment multiple instances in a video. Our approach, VideoCutLER, does not require labels, and does not rely on motion-based learning signals like optical flow. In fact, VideoCutLER does not need real videos for training as we synthesize videos using natural images from the ImageNet-1K. Despite being simpler, VideoCutLER outperforms models that use additional learning signals or video data, achieving \\(10 \\times\\) their performance on benchmarks like YouTubeVIS. Moreover, VideoCutLER is a strong pretrained model for supervised learning. We hope that our approach enables both a wide range of applications in video recognition, as well as its simplicity enables easy future research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations: while VideoCutLER demonstrates its capability to achieve the state-of-the-art performance without relying on optical flow estimations, potential further improvements may be obtained by leveraging natural videos and integrating joint training with optical flow estimations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22762" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.47, + 0.537 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.594, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Yuan-Ting Hu, Jia-Bin Huang, and Alexander G Schwing. Unsupervised video object segmentation using motion saliency-guided spatio-temporal propagation. In Proceedings of the European conference on computer vision (ECCV), pages 786-802, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In Proceedings of the IEEE international conference on computer vision, pages 3271-3279, 2015. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[19] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The hci benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 19-28, 2016. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[21] Minhyeok Lee, Suhwan Cho, Seunghoon Lee, Chaewon Park, and Sangyoun Lee. Unsupervised video object segmentation via prototype memory network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5924-5934, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[25] Jonathon Luiten, Idil Esen Zulfikar, and Bastian Leibe. Unovost: Unsupervised offline video object segmentation and tracking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2000-2009, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22763" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[26] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[27] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.472, + 0.276 + ], + "angle": 0, + "content": "[28] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.348 + ], + "angle": 0, + "content": "[29] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[31] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[32] Wenguan Wang, Tianfei Zhou, Fatih Porikli, David Crandall, and Luc Van Gool. A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[33] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[34] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[35] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[36] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[37] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[38] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[39] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[40] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.893, + 0.331 + ], + "angle": 0, + "content": "[41] Zhao Yang, Qiang Wang, Song Bai, Weiming Hu, and Philip HS Torr. Video segmentation by detection for the 2019 unsupervised davis challenge. 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[42] Tianfei Zhou, Jianwu Li, Shunzhou Wang, Ran Tao, and Jianbing Shen. Matnet: Motion-attentive transition network for zero-shot video object segmentation. IEEE Transactions on Image Processing, 29:8326-8338, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22764" + } + ] +] \ No newline at end of file diff --git a/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_origin.pdf b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e8457416ea4a0ef0c007bdd41f491a5a0383354 --- /dev/null +++ b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/d5335ce2-de3e-4c31-a9e0-97b9c027eebd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a62a860a99e9108700fc8d1605a7abbb424e02c7a70b27255407a66d1b663544 +size 2480872 diff --git a/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/full.md b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7d3170f18ed1285a8a3e584f69664b4bde111af --- /dev/null +++ b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/full.md @@ -0,0 +1,264 @@ +# VideoCutLER: Surprisingly Simple Unsupervised Video Instance Segmentation + +Xudong Wang + +Ishan Misra +UC I + +Ziyun Zeng +eley FAI + +Rohit Girdhar +Meta AI + +Trevor Darrell + +Code: https://github.com/facebookresearch/CutLER + +![](images/0f68c0f203457877cd08e5f42d4db5c8f6ae6cc3a1767868743dc8aa6bf0ff24.jpg) +Learning Unsupervised Video Instance Segmentation +Figure 1. VideoCutLER is a simple unsupervised video instance segmentation method (UnVIS). We show the first competitive unsupervised results on the challenging YouTube VIS benchmark. Moreover, unlike most prior approaches, we demonstrate that UnVIS models can be learned without relying on natural videos and optical flow estimates. Row 1: We propose VideoCutLER, a simple cut-synthesis-and-learn pipeline that involves three main steps. Firstly, we generate pseudo-masks for multiple objects in an image using MaskCut [35]. Then, we convert a random pair of images in the minibatch into a video with corresponding pseudo mask trajectories using ImageCut2Video. Finally, we train an unsupervised video instance segmentation model using these mask trajectories. Row 2: Despite being trained only on unlabeled images, at inference time VideoCutLER can be directly applied to unseen videos and can segment and track multiple instances across time (Fig. 1a), even for small objects (Fig. 1b), objects that are absent in specific frames (Fig. 1c), and instances with high overlap (Fig. 1d). Column 2: Our method surpasses the previous SOTA method OCLR [37] by a factor of 10 in terms of class-agnostic $\mathrm{AP}_{50}^{\mathrm{video}}$ . + +# Abstract + +Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving $50.7\%$ AP $_{50}$ ^ $\text{video}$ , surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by $15.9\%$ on YouTubeVIS-2019 in terms of AP $_{\text{video}}$ . + +# 1. Introduction + +Video instance segmentation is vital for various computer vision applications, e.g. video surveillance, autonomous driving, and video editing, yet labeled videos are costly to obtain. Hence, there is a pressing need to devise an unsupervised video instance segmentation approach that can comprehend video content comprehensively and operate in general domains without labels. + +Prior work in this area typically relies on an optical flow network as an off-the-shelf motion estimator [30, 37, 38]. Although optical flow can be informative in detecting pixel motion between frames, it is not always a reliable technique, particularly in the presence of occlusions, motion blur, complex motion patterns, changes in illuminations, etc. As a result, models that heavily rely on optical flow estimations may fail in several common scenarios. For example, stationary or slowly moving objects may have flow estimates similar to the background, causing them to be omitted in the segmentation process (e.g., the parrot with negligible motion is missed in Fig. 2a). Similarly, non + +![](images/e4147f3b71f819a0116a06564cc88ba474d58ae395e6f08ae4d0409e8c15e59f.jpg) +Figure 2. Challenges encountered by the previous state-of-the-art OCLR: Within the framework of OCLR [37], a method that heavily relies on optical flows as model inputs, several distinct failure cases emerge. These include situations where the method struggles to accurately segment both moving and static objects (as demonstrated in Fig. 2a), struggles to effectively track non-rigid objects as a coherent unit (Fig. 2b), encounters difficulties in distinguishing overlapping instances (Fig. 2c), and fails to maintain consistent predictions under varying illumination conditions (Fig. 2d). Nonetheless, many of these challenges can be effectively addressed through the application of our proposed approach, VideoCutLER, without being reliant on the optical estimations used by various prior works [37, 38]. We present qualitative comparisons using the YouTubeVIS dataset [39]. + +rigid objects with non-consistent motions for several parts have varying optical flows, leading to a failure in segmenting all parts cohesively as a unit if object motion is presumed constant (Fig. 2b). Also, objects with similar motion patterns and high overlap are complex for optic flow methods to accurately distinguish between them, especially in boundary regions (Fig. 2c). Finally, objects with illumination changes across frames can cause optical-flow based models to produce non-consistent and blurred segmentation masks (Fig. 2d). Given the limitations above, we advocate for unsupervised video segmentation models which do not + +depend on optical flow estimates. We propose a method to train a video segmentation model by generating simple synthetic videos from individual images, without relying on explicit motion estimates or requiring labeled natural videos. + +Our method, VideoCutLER, is an unsupervised Video instance segmentation model that employs a Cut-synthesis-and-LEaRn pipeline (Fig. 1). First, given unlabeled images, we extract pseudo-masks for multiple objects in an image using MaskCut [35], leveraging a self-supervised DINO [4] and a spectral clustering method Normalized Cuts [28] (details in Sec. 3.1). Second, given unlabeled images and their pseudo-masks in a minibatch, we propose ImageCut2Video, a surprisingly simple video synthesis scheme that generates a video from those with corresponding pseudo mask trajectories (details in Sec. 3.2). Finally, those mask trajectories are used to train a video instance segmentation model, aiming to perform object segmentation with temporal consistency across video frames (details in Sec. 3.3). Our model learns to segment and track object instances based on their appearance (feature) similarities across video frames. + +Despite being learned from only unlabeled images (and the temporally simple synthetic video sequences we construct from them), VideoCutLER succeeds at multi-instance video segmentation, achieving a new state-of-the-art (SOTA) performance of $50.7\%$ AP $^{\text{video}}$ on YouTubeVIS2019. This result surpasses the previous SOTA [37] by substantial margins of $45.9\%$ ( $50.7\%$ vs. $4.8\%$ ). This result also considerably narrows the performance gap between supervised and unsupervised learning, reducing it from $29.1\%$ to $11.0\%$ in terms of the AP $^{\text{video}}$ + +Moreover, most prior works on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers being randomly initialized. In contrast, our pretraining strategy takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, our method outperforms DINO's [4] AP $^{\text{video}}$ on YoutubeVIS-2019 by $15.9\%$ . + +Contributions. Insights: We found that a simple video synthesis method yield surprisingly effective results for training unsupervised multi-instance video segmentation models. This efficacy is achieved without the necessity of explicit motion estimates or the utilization of natural videos (relying solely on unlabeled ImageNet data suffices), a novel aspect that has not been previously demonstrated in the field. Methods: We propose a simple yet effective cut-synthesize-and-learn pipeline VideoCutLER for learning video instance segmentation models, given unlabeled images. Results: Our method shows the first successfully results on challenging unsupervised multi-instance video segmentation benchmark YouTubeVIS, outperforming the previous SOTA model's $\mathrm{AP}_{50}^{\mathrm{video}}$ by a large margin. + +
CRWDINOOCLROurs
Segment multiple objects✓✗✓✓
Track objects across frames✓✗✓✓
No need for optical flow✓✓✗✓
No 1st-frame ground-truth✗✗✓✓
No human labels at any stage✗✗✓†✓
Pretrained model for sup. learning✗✓✗✓
+ +Table 1. We compare previous methods on unsupervised video instance segmentation, including CRW [17], DINO [4], and OCLR [37], with our VideoCutLER in term of key properties. Our VideoCutLER is the only approach that fulfills all these desired properties. $\dagger$ : The optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data and human-annotated data like KITTI-2015 [18] and HD1K [19]. + +# 2. Related Work + +Unsupervised video instance segmentation (VIS) requires not only separating and tracking the main moving foreground objects from the background, but also differentiating between different instances, without any human annotations [32]. Previous works [16, 21, 36, 38, 40] on unsupervised video segmentation has primarily centered on unsupervised video object segmentation (VOS), aiming to detect all moving objects as the foreground and to generate a pixel-level binary segmentation mask, regardless of whether the scene contains a single instance or multiple instances. Despite some works exploring unsupervised video instance segmentation (VIS), many of these approaches have resorted to either utilizing first frame annotations [4, 17, 22] to propagate label information throughout the video frames or leveraging supervised learning using large amounts of external labeled data [25, 31, 41, 42]. Furthermore, prior studies typically utilized optical flow networks that were pretrained with human supervision using either synthetic data or labeled natural videos [31, 37, 38, 40]. + +The properties deemed necessary for an unsupervised learning method to excel in video instance segmentation tasks are presented and discussed in Tab. 1. Our proposed method, VideoCutLER, is the only approach that satisfies all these properties, making it an effective and promising solution for unsupervised video instance segmentation. + +Unsupervised object discovery aims to automatically discover and segment objects in an image in an unsupervised manner [16, 34-36]. LOST [29] and TokenCut [36] focus on salient object detection and segmentation via leveraging the patch features from a pretrained DINO [4] model. For multi-object discovery, FreeSOLO [34] first generates object pseudo-masks for unlabeled images, then learns an unsupervised instance segmentation model using these pseudo-masks. CutLER [35] presents a straightforward cut-and-learn pipeline for unsupervised detection and segmentation of multiple instances. It has demonstrated promising results on more than eleven different benchmarks, covering + +a wide range of domains. + +In contrast to previous approaches, our unsupervised learning method focuses on simultaneously tracking objects in a video sequence while identifying correspondences between instances across multiple frames. + +Self-supervised representation learning generates its own supervision signal by exploiting the implicit patterns or structures present in the input data [3, 4, 14, 15]. Unlike most previous self-supervised learning models, which still require fine-tuning on labeled data to be operative on complex computer vision tasks, such as detection and segmentation, VideoCutLER can tackle these complex, challenging tasks with purely unsupervised learning methods. + +# 3. VideoCutLER + +We present VideoCutLER, a simple cut-synthesis-and-learn pipeline consisting of three main steps. First, we generate pseudo-masks for multiple objects in an image using MaskCut (Sec. 3.1). Next, we convert a random pair of images in the minibatch into a synthetic video with corresponding pseudo mask trajectories using ImageCut2Video (Sec. 3.2). Finally, we train an unsupervised video instance segmentation model using these mask trajectories. As the model inputs do not contain explicit motion estimates, it learns to track objects based on their appearance similarity (Sec. 3.3). + +# 3.1. Single-image unsupervised segmentation + +We employ the MaskCut method, introduced in the CutLER [35] method. MaskCut is an efficient spectral clustering approach for unsupervised image instance segmentation and object detection and can discover multiple object masks in a single image without human supervision. MaskCut builds upon a self-supervised DINO model [2] with a backbone of ViT [10] and a cut-based clustering method Normalized Cuts (NCut) [28]. MaskCut first generates a patchwise affinity matrix $W_{ij} = \frac{K_i K_j}{\|K_i\|_2\|\bar{K}_j\|_2}$ using the 'key' features $K_{i}$ for patch $i$ from DINO's last attention layer. Subsequently, the NCut algorithm [28] is employed on the affinity matrix by solving a generalized eigenvalue problem + +$$ +(D - W) x = \lambda D x \tag {1} +$$ + +where $D$ is a diagonal matrix with $d(i) = \sum_{j} W_{ij}$ and $x$ is the eigenvector that corresponds to the second smallest eigenvalue $\lambda$ . Then, the foreground masks $M^s$ can be extracted via bi-partitioning $x$ , which segments a single object within the image. To segment more than one object, Mask-Cut uses an iterative process that masks out the values in the affinity matrix using the extracted foreground mask: + +$$ +W _ {i j} ^ {t} = \frac {\left(K _ {i} \prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\right) \left(K _ {j} \prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\right)}{\| K _ {i} \| _ {2} \| K _ {j} \| _ {2}} \tag {2} +$$ + +and repeats the NCut algorithm. We set $t = 3$ by default. + +Although MaskCut can effectively locate and segment multiple objects in an image, it operates only on a single image, lacking temporal consistency in the instance segmentation masks produced across video frames. + +# 3.2. ImageCut2Video Synthesis for Training + +We propose a learning-based approach to ensuring temporal consistency in video segmentation masks, based on generating synthetic videos from pairs of individual images and MaskCut masks. Surprisingly, we found that an extremely simple synthetic video generation method yields sufficient training data to learn a powerful video segmentation model that can operate on videos with much greater complexity of motion than is present in the training data. + +Given unlabeled images in the minibatch and their pseudo-masks, our ImageCut2Video method synthesizes corresponding videos and pseudo-mask trajectories, thereby allowing us to train the model in an unsupervised manner while offering the necessary supervision for simultaneous detection, segmentation, and tracking of objects in videos. + +First, given an image and its corresponding pseudomasks in the mini-batch, we duplicate the image $t$ times and connect its MaskCut pseudo-masks to form the initial trajectories. This synthetic video, however, only contains static foreground objects. To generate additional trajectories with mobile objects, a second image is randomly selected from the mini-batch, and its objects are cropped using its MaskCut pseudo-masks. These objects are then randomly resized, repositioned, and augmented before being pasted onto the first image. The resulting masks are connected along the temporal dimension to generate additional trajectories with mobile objects. + +Specifically, given a target image $I_{1}$ , a random source image $I_{2}$ in the mini-batch and its corresponding set of binary pseudo-masks $\{M_2^1,\dots,M_2^s\}$ , we first apply a transformation function $\mathcal{T}$ to resize and shift these pseudomasks randomly. This gives us a new set of pseudo-masks $\{\hat{M}_2^1,\dots,\hat{M}_2^s\}$ , where $\hat{M}_2^s = \mathcal{T}(M_2^s)$ . Next, we synthesize a video with $t$ frames by duplicating image $I_{1}$ for $t$ times and pasting the augmented masks onto $I_{1}$ using: + +$$ +I _ {1} ^ {t} = I _ {1} \times \Pi_ {i = 1} ^ {s} (1 - \hat {M} _ {2} ^ {i}) + I _ {2} \times \left(1 - \Pi_ {i = 1} ^ {s} (1 - \hat {M} _ {2} ^ {i})\right) \quad (3) +$$ + +where $\times$ refers to element-wise multiplication. + +# 3.3. Video Segmentation Model + +During training, the synthetic videos produced by Image-Cut2Video, comprising both mobile and stationary objects, are used as the inputs to train a video instance segmentation model. The segmentation mask trajectories corresponding to each object in the video serve as 'ground-truth' labels. + +We utilize VideoMask2Former [6, 7] with a backbone of ResNet50 [12] as our video instance segmentation (VIS) + +model. It operates by attending to the 3D spatiotemporal features of our synthetic videos and generating 3D volume predictions of pseudo-mask trajectories using shared queries across frames. The shared queries across frames enable the model to segment and track object instances based on their appearance (feature) similarities, making it a powerful framework for analyzing video sequences. + +# 4. Implementation Details + +VideoCutLER. We first employ the MaskCut approach on images preprocessed to a resolution of $480 \times 480$ pixels. We then compute a patch-wise cosine similarity matrix using the pretrained ViT-Base/8 DINO [4] model, which serves as input to the MaskCut algorithm for initial segmentation mask generation. We set $t = 3$ , which is the maximum number of masks per image. To refine the segmentation masks, we employ a post-processing step using Conditional Random Fields (CRFs) [20], which enforces smoothness constraints and preserves object boundaries, resulting in improved segmentation masks. + +Next, we use ImageCut2Video to synthetic videos given images and their pseudo-masks in a mini-batch. We found that synthetic videos with two frames are sufficient to train a video instance segmentation model; therefore, we use $s = 2$ by default. We randomly change the brightness, contrast, and rotation of the masks to create new variations of pseudo-masks. Additionally, we randomly resize the pseudo-masks (scale∈[0.8,1.0]), and shift their positions. + +Training and test data. Our model is trained solely on the unlabeled images from ImageNet [8], which comprises approximately 1.3 million images. Without further fine-tuning on any video datasets, we test our model's zero-shot unsupervised video instance segmentation performance on four multi-instance video segmentation benchmarks, including YouTubeVIS-2019 [39], YouTubeVIS-2021 [39], DAVIS2017 [26], and DAVIS2017-Motion [26, 27]. + +YoutubeVIS-2019 and YouTube-VIS2021 contain 2,883 high-resolution YouTube videos and 3,859 high-resolution YouTube videos, respectively. We evaluate the zero-shot unsupervised learning performance on their training splits in a class-agnostic manner. For DAVIS-2017, we evaluate our model's performance on the 30 videos from its val set. + +Training settings. 1) Unsupervised Image Model Pretraining: We first pretrain a Mask2Former [7] model with a backbone of ResNet50 [12] on ImageNet using MaskCut's pseudo-masks. The model is optimized for 160k iterations, with a batch size of 16 and a learning rate of 0.00002. The learning rate is decayed by a factor of 20 at iteration 80,000. To prevent overfitting, a dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders. 2) Unsupervised Video Model Learning: We initialize the VideoMask2Former model [6] with model weights from the previous stage, and then fine-tune + +
MethodsTraining settingsYouTubeVIS-2019YouTubeVIS-2021
flowvideossup.\(AP_{50}\)\(AP_{75}\)AP\(AP_{S}\)\(AP_{M}\)\(AP_{L}\)\(AR_{10}\)\(AP_{50}\)\(AP_{75}\)AP\(AP_{S}\)\(AP_{M}\)\(AP_{L}\)\(AR_{10}\)
MotionGroup* [38]✓✓✘1.30.10.30.20.30.51.71.10.10.20.10.20.51.5
OCLR* [37]✓✓\(X^†\)4.80.41.30.01.25.511.04.40.31.20.11.67.19.6
CutLER‡✘✘✘37.514.617.13.313.927.630.429.210.412.83.112.827.822.6
VideoCutLER✘✓*✘50.724.226.05.620.937.942.438.919.017.15.318.337.531.3
vs. prev. SOTA+12.8+9.6+8.9+2.3+7.0+10.3+12.0+9.7+8.6+4.3+2.2+5.5+9.7+8.7
+ +Table 2. Zero-shot unsupervised multi-instance video segmentation on YouTubeVIS-2019 and YouTubeVIS-2021. We report the instance segmentation metrics (AP and AR) and training settings. *: reproduced MotionGroup [38] and OCLR [37] results with the official code and checkpoints. †: the optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data [1, 9] and human-annotated data, such as KITTI-2015 [18] and HD1K [19]. ‡: We train a CutLER [35] model with Mask2Former as a detector on ImageNet-1K, following CutLER's official training recipe, and use it as a strong baseline. *: VideoCutLER is trained on synthetic videos generated using ImageNet. Sup and flow denote human supervision and optical flow information, respectively. We evaluate results on YouTubeVIS's train splits in a class-agnostic manner (note: we never train on YouTubeVIS). + +it on the synthetic videos we construct from ImageNet. We train VideoCutLER on 8 A100 GPUs for 80k iterations, using the AdamW optimizer [24]. We set the initial learning rate to 0.000005 and apply a learning rate multiplier of 0.1 to the backbone. A dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders. + +Evaluation metric $\mathbf{AP}^{\mathrm{video}}$ and $\mathbf{AR}^{\mathrm{video}}$ : The evaluation metrics used in YouTubeVIS are Averaged Precision (AP) and Averaged Recall (AR), which are similar to those used in COCO [23]. The evaluation is specifically conducted at 10 intersection-over-union (IoU) thresholds ranging from $50\%$ to $95\%$ with a step of $5\%$ [39]. However, unlike in image instance segmentation, each instance in a video comprises a sequence of masks, so the IoU computation is performed not only in the spatial domain, but also in the temporal domain by summing the intersections at every single frame over the unions at every single frame. + +Evaluation metric $\mathcal{J}$ and $\mathcal{F}$ : For DAVIS [27], we report results using their official evaluation metrics $\mathcal{J} \& \mathcal{F}$ , $\mathcal{J}$ and $\mathcal{F}$ . The region measure $(\mathcal{J})$ [27] is the intersection-over-union (IoU) score between the algorithm's mask and the ground-truth mask. The boundary measure $(\mathcal{F})$ [27] is the average precision of the boundary of the algorithm's mask. The evaluation metrics are computed separately for each instance, and then the results are averaged over all instances to get the final score. $\mathcal{J} \& \mathcal{F}$ is the mean of $\mathcal{J}$ and $\mathcal{F}$ . + +# 5. Experiments + +We evaluate the performance of VideoCutLER on several video instance segmentation benchmarks. In Sec. 5.1, we demonstrate that our approach can effectively perform segmentation and tracking of multiple objects in videos, even when trained on unlabeled ImageNet images without any form of supervision. Our experimental results reveal that our method can drastically reduce the performance gap between unsupervised and supervised learning methods for video instance discovery and tracking. Furthermore, Sec. 5.2 demonstrates that fine-tuning VideoCutLER leads + +to further performance gains in video instance segmentation, surpassing previous works such as DINO in both fully supervised learning and semi-supervised learning tasks. In Sec. 5.3, we conduct an ablation study to examine the impact of key components and their hyperparameters. + +# 5.1. Unsupervised Zero-shot Evaluations + +In this section, we evaluate the performance of our method against previous state-of-the-art approaches on various video instance segmentation benchmarks. + +Evaluating unsupervised video instance segmentation poses two main challenges. Firstly, as unsupervised learning methods train the model without semantic classes, the class-aware video segmentation setup cannot be used directly for an evaluation. As a result, following previous works, we evaluate video instance segmentation results in a class-agnostic manner. Secondly, video instance segmentation datasets often annotate only a subset of the objects in the video, which makes Average Recall (AR) a valuable metric that does not penalize models for detecting novel objects not labeled in the dataset [35]. Therefore, we report both AR and AP for YouTubeVIS. Regarding DAVIS, we use the official unsupervised learning metrics $\mathcal{I},\mathcal{F}$ , and $\mathcal{J}\&\mathcal{F}$ . All these metrics assess the performance of unsupervised video instance segmentation in a class-agnostic manner. Sec. 4 lists more details on evaluation metrics. + +Detailed comparisons on YouTubeVIS. Tab. 2 presents a summary of the results for unsupervised zero-shot video instance segmentation on the YouTubeVIS-2019 and YouTubeVIS-2021 datasets. We compare our method's results with the previous state-of-the-art methods OCLR [37] and motion grouping [38]. We reproduce their results using their official code and checkpoints to ensure fairness. + +Although OCLR [37] is also trained on synthetic videos, it relies on the off-the-shelf optical flow estimator RAFT [30] to compute optical flows for RGB sequences. It is worth noting that RAFT is pretrained on a combination of synthetic videos [1, 9] and human-annotated videos + +
MethodsTraining settingsDAVIS2017DAVIS2017-Motion
flowvideossup.training dataJ&FJ(Mean)F(Mean)J&FJ(Mean)F(Mean)
MotionGroup (sup.) [38]✓✓XIN-1K+synthetic---39.544.934.2
Mask R-CNN (w/ flow)* [13, 37]✓✓XIN-1K+synthetic---50.350.450.2
OCLR (w/ flow)* [37]✓✓XIN-1K+synthetic39.638.241.155.154.555.7
VideoCutLERXXXIN-1K43.641.745.557.357.457.2
vs. prev. SOTA+4.0+3.5+4.4+2.2+2.9+1.5
+ +Table 3. Zero-shot unsupervised single/few-instance segmentation. VideoCutLER also outperforms the previous state-of-the-arts on DAVIS2017 and DAVIS2017-Motion. Note: 12 out of 30 videos from DAVIS2017 and 26 out of 30 videos from DAVIS2017-Motion contain only 1 moving instance. Additionally, DAVIS datasets focus solely on the performance of moving prominent objects, even in videos where multiple objects are present. This disadvantages our model since it can segment both static and moving objects and has not been exposed to any downstream videos during training. *: utilize optical flow predictions from RAFT [30], which is pretrained on external videos. All methods are evaluated in a zero-shot manner, i.e. no fine-tuning on target videos. + +such as KITTI-2015 [18] and HD1K [19]. Our approach, VideoCutLER, despite not using any optical flow estimations like many previous works on unsupervised video segmentation, achieves over $10 \times$ higher $\mathrm{AP}_{50}$ and $18 \times$ higher AP than OCLR [37] on YouTubeVIS-2019. Additionally, we achieve over $30\%$ higher recall. Furthermore, unlike the previous state-of-the-art method OCLR [37], which exhibits poor performance in segmenting small objects (with $0.0\%$ $\mathrm{AP}_S$ ), our approach significantly outperforms it. Similar performance gains can be observed on YouTubeVIS-2021. Finally, the performance gains to CutLER [35] demonstrates the effectiveness of VideoCutLER in training unsupervised multi-instance video segmentation models, surpassing CutLER by over $12.8\%$ on YouTubeVIS-2019. + +In Fig. 3, we present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset. + +Detailed comparisons on DAVIS. To provide a comprehensive evaluation and comparison with existing unsupervised video instance segmentation approaches, we also assess the performance of our model on the validation sets of DAVIS-2017 and DAVIS2017-Motion [27, 37]. Note that both DAVIS2017 and DAVIS2017-Motion datasets focus only on the performance of instance segmentation on prominent moving objects, even in videos with multiple objects. As a result, only a single or a few objects of interest per video are annotated, which may not reflect the challenges that arise when multiple objects are present. + +Although the evaluation of DAVIS is an unfair assessment for us since VideoCutLER is supposed to segment both static and moving objects, whereas DAVIS focuses on moving prominent objects, with only a single or a few moving objects of interest per video annotated. However, Tab. 3 shows that VideoCutLER yields approximately $4\%$ higher $\mathcal{I}, \mathcal{F}$ , and $\mathcal{J} \& \mathcal{F}$ . The additional results on DAVIS demonstrate that VideoCutLER achieves superior performance not only on static or minimally moving objects but also on dynamic objects, where prior methods relying on optical flow estimates can benefit from additional cues. + +![](images/b41978343a5dfba7fb8df916cab7667635ebfd12e22f7b1e42678cbd948a70a8.jpg) +Figure 3. We present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset. It's noteworthy that VideoCutLER is solely pretrained on image dataset ImageNet-1K, and its evaluation is conducted directly on the video dataset YouTubeVIS (no further fine-tuning required). The visual results provided effectively highlight that VideoCutLER is capable of segmenting and tracking multiple instances, delivering consistent tracking results across video frames, and successfully distinguishing between various instances, even when significant overlapping occurs. We show more demo results in appendix. + +Comparison of supervised and unsupervised learning in object discovery and tracking abilities is presented in Tab. 4. We train a supervised MaskTrack R-CNN [39] model on the human-annotated training set of YouTubeVIS-2019 dataset, and evaluate it in a class-agnostic manner on the videos that are not shared between YouTubeVIS-2019 and YouTubeVIS-2021 datasets [39]. Tab. 4 shows that our VideoCutLER model significantly narrows the gap between supervised learning and unsupervised learning methods in terms of the averaged precision $\mathrm{AP}_{50}$ (gaps: $29.1\% \rightarrow 11.0\%$ ) and the averaged recall $\mathrm{AR}_{100}$ (gaps: $14.9\% \rightarrow 3.2\%$ ), particularly for the $\mathrm{AR}_{100}$ . + +
MethodsTraining settingsYouTubeVIS-2021\YouTubeVIS-2019
flowvideossup.training data\(AP_{50}\)\(AP_{75}\)AP\(AP_S\)\(AP_M\)\(AP_L\)\(AR_{100}\)
Mask2Former [6]✓✓✘IN-1K+YT201948.922.224.9----
MaskTrack R-CNN* [39]✓✓✓IN-1K+YT201932.413.015.08.424.939.020.3
MaskTrack R-CNN* [39]✓✓✓IN-1K+COCO+YT201935.818.718.710.531.346.824.5
\(OCLR^*\)[37]✓✓✘IN-1K+synthetic3.30.21.00.32.77.55.4
VideoCutLER✘✘✘IN-1K21.47.19.04.913.329.617.1
vs. prev. SOTA+18.1+6.9+8.0+4.6+10.6+22.1+11.7
+ +Table 4. VideoCutLER greatly narrows the gap between fully-supervised learning and unsupervised learning for multi-instance video segmentation. Results are evaluated in a class-agnostic manner on the relative complement of the set of videos from YouTubeVIS-2021 and the set of videos from YouTubeVIS-2019. VideoCutLER and Mask2Former use a backbone of ResNet50. *: reproduced results with the official code and checkpoints. IN-1K refers to ImageNet-1K. + +
MethodsArchitectureYouTubeVIS-2019YouTubeVIS-2021
AP\(AP_{50}\)\(AP_{75}\)\(AP_S\)\(AP_M\)\(AP_L\)AP\(AP_{50}\)\(AP_{75}\)\(AP_S\)\(AP_M\)\(AP_L\)
DINO [4]Mask2Former [6]23.039.023.76.028.034.224.641.425.98.734.039.9
VideoCutLERMask2Former [6]38.956.743.322.143.151.833.453.836.315.740.954.8
vs. prev. SOTA+15.9+17.7+19.6+16.1+15.1+17.6+8.8+12.4+10.4+7.0+6.9+14.9
+ +Table 5. VideoCutLER can serve as a strong pretrained model for the supervised video instance segmentation task. The video segmentation model, Mask2Former, is initialized with various pretrained models, i.e., DINO or VideoCutLER, and fine-tuned on the training set with human annotations. We report the instance segmentation metrics and evaluate the model performance on the val splits. + +![](images/6efe8b2c7632c6b706d69504d99e206a5a445cebbdb508260abaa639a101be63.jpg) +Figure 4. We fine-tune VideoCutLER for semi-supervised video instance segmentation on the YouTubeVIS-2019 dataset, using different percentages of labeled training data. We evaluate the performance of our method by reporting the average precision and recall on the validation set of YouTubeVIS-2019. To establish a strong baseline, we use the self-supervised DINO [4] model and initialize the weights of VideoMask2Former with DINO. To ensure a fair comparison, both baselines and VideoCutLER are trained using the same schedule and recipe. + +![](images/c637624ef2de954a2e9705afd24045a701cabe73cd7241073af1215c72d1c979.jpg) + +# 5.2. Label-Efficient and Fully-Supervised Learning + +In this section, we investigate VideoCutLER as a pretraining approach for supervised video instance segmentation models, and evaluate its effectiveness in label-efficient and fully-supervised learning scenarios. + +Setup. We use VideoMask2Former with a backbone of ResNet50 for all experiments in this section unless otherwise noted. For our experiments on semi-supervised learning, we randomly sample a subset of videos from the training split with different proportions of labeled videos. After pretraining our VideoCutLER model on ImageNet, we + +fine-tune the model on the YouTubeVIS-2019 [39] dataset with its human annotations. For our experiments on the fully-supervised learning task, we fine-tune the VideoCUTLER model on all available labeled data from the training sets of YouTubeVIS. For baselines, we initialize a VideoMask2Former model with a DINO [4] model pre-trained on ImageNet and fine-tuned on labeled videos. Since DINO has shown strong performance in detection and segmentation tasks, it serves as a strong baseline for our experiments. + +For semi-supervised learning, both the baselines and our models are trained for $2 \times$ schedule, with a learning rate of 0.0001 for all model weights, except for the final classification layers, which use a learning rate of 0.0016. We train the models using a batch size of 16 and 8 GPUs. For fully-supervised learning, we use the $1 \times$ schedule and a learning rate of 0.0002 for the final classification layers. We evaluate their performance on the val split of the YouTubeVIS-2019, and report results from its official evaluation server. + +Data for fully-/semi-supervised VIS. We fine-tune the pretrained VideoCutLER model on all or a subset of the training split of YouTubeVIS-2019. We then evaluate the resulting models on the validation set. To ensure a fair comparison, we use the same amount of human annotations to train our model and baselines. Specifically, we initialize the baselines with the DINO-pretrained model and fine-tune them on the training set of the respective dataset. We evaluate the model performance on their validation sets and report results from its official evaluation server. + +Results. Most prior approaches on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers, such as Mask2Former's decoders, + +![](images/d03c2d4a538f9fe12a35560cb7eb60fc1dd56634fdccec330bb129b769550f59.jpg) +Figure 5. We present qualitative results on videos covering a range of out-of-domain sources, e.g., sketches, 3D computer-generated imagery (CGI) and hybrid (CGI + realistic). VideoCutLER can produce high-quality segmentation and tracking results for small objects that are often difficult to distinguish from the background, as well as for object sketches that lack textual information. + +being randomly initialized. In contrast, VideoCutLER takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, as shown in Fig. 4 and Tab. 5, our method outperforms these prior works significantly, offering a strong pretrained model for fully-/semi-supervised learning tasks. + +Fig. 4 shows that VideoCutLER consistently outperforms the strong baseline method DINO [4] across all label-efficient learning settings with varying proportions of labeled YouTubeVIS-2019 videos. The most significant performance gains are observed when $20\%$ labeled data is provided, where VideoCutLER exceeds DINO by over $12\% \mathrm{AP}_{50}$ and $13.2\%$ AR. As demonstrated in Tab. 5, training the model with all available labeled videos from YouTubeVIS yields considerable performance gains, surpassing DINO by more than $15.9\%$ AP on YouTubeVIS-2019 and $8.8\%$ on YouTubeVIS-2021, respectively. + +# 5.3. Ablation Study + +Hyper-parameters and design choices. We present an ablation study on several key hyper-parameters and design choices of VideoCutLER in Tab. 6. First, we analyze the impact of varying the size of video frames used for training VideoCutLER. From Tab. 6a, we observe that the shortest edge length of 240 pixels yields the best performance. Using a larger resolution does not always lead to better results. Next, Tab. 6b shows the effect of the number of frames used + +
Size → 180360480#frames → CutLER† [35]234
APvideo5049.950.7APvideo5037.549.850.750.4
(a) Frame size.(b) # frames. z
Augmentations → none+bright+rotation+contrast+cropall
APvideo5047.848.148.948.348.750.7
+ +(c) Data augmentations for ImageCut2Video. + +Table 6. Ablations for VideoCutLER. We report video instance segmentation result $\mathrm{AP}_{50}^{\mathrm{video}}$ on YoutubeVIS-2019. (a) The impact of varying video frame sizes on training VideoCutLER. (b) The effect of the number of frames used for model training. (c) The impact of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. Default settings are highlighted in gray. + +for training video instance segmentation models. We found that synthetic videos with three frames are optimal for learning an unsupervised video instance segmentation model. Increasing the number of frames does not result in a further improved performance, aligning with the findings reported in [6]. Furthermore, Tab. 6c investigates the contribution of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. We found that compared to ImageCut2Video without any data augmentations, adding these augmentations can bring about $3\%$ performance gains. Generalizability. The results presented in Fig. 5 demonstrate that VideoCutLER can effectively perform video instance segmentation on out-of-domain data sources, e.g. sketches, 3D computer-generated imagery, and hybrid videos that combine CGI. These results show that our model can be applied to a broad range of videos beyond the domains it was initially trained on, i.e., ImageNet. + +# 6. Summary + +We presented a simple unsupervised approach to segment multiple instances in a video. Our approach, VideoCutLER, does not require labels, and does not rely on motion-based learning signals like optical flow. In fact, VideoCutLER does not need real videos for training as we synthesize videos using natural images from the ImageNet-1K. Despite being simpler, VideoCutLER outperforms models that use additional learning signals or video data, achieving $10 \times$ their performance on benchmarks like YouTubeVIS. Moreover, VideoCutLER is a strong pretrained model for supervised learning. We hope that our approach enables both a wide range of applications in video recognition, as well as its simplicity enables easy future research. + +Limitations: while VideoCutLER demonstrates its capability to achieve the state-of-the-art performance without relying on optical flow estimations, potential further improvements may be obtained by leveraging natural videos and integrating joint training with optical flow estimations. + +# References + +[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 5 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020. 3 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 7, 8 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 7 +[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 4, 7, 8 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 4 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 +[9] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 5 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3 +[11] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 7 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 + +[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6 +[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 3, 7 +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3 +[16] Yuan-Ting Hu, Jia-Bin Huang, and Alexander G Schwing. Unsupervised video object segmentation using motion saliency-guided spatio-temporal propagation. In Proceedings of the European conference on computer vision (ECCV), pages 786-802, 2018. 3 +[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 3 +[18] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In Proceedings of the IEEE international conference on computer vision, pages 3271-3279, 2015. 3, 5, 6 +[19] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The hci benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 19-28, 2016. 3, 5, 6 +[20] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 2011. 4 +[21] Minhyeok Lee, Suhwan Cho, Seunghoon Lee, Chaewon Park, and Sangyoun Lee. Unsupervised video object segmentation via prototype memory network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5924-5934, 2023. 3 +[22] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[25] Jonathon Luiten, Idil Esen Zulfikar, and Bastian Leibe. Unovost: Unsupervised offline video object segmentation and tracking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2000-2009, 2020. 3 + +[26] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4 +[27] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4, 5, 6 +[28] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2, 3 +[29] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3 +[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 1, 3, 5, 6 +[31] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 3 +[32] Wenguan Wang, Tianfei Zhou, Fatih Porikli, David Crandall, and Luc Van Gool. A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153, 2021. 3 +[33] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 2, 7 +[34] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 3 +[35] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 3, 5, 6, 8 +[36] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022. 3 +[37] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6, 7 + +[38] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 1, 2, 3, 5, 6 +[39] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2, 4, 5, 6, 7 +[40] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019. 3 +[41] Zhao Yang, Qiang Wang, Song Bai, Weiming Hu, and Philip HS Torr. Video segmentation by detection for the 2019 unsupervised davis challenge. 2019. 3 +[42] Tianfei Zhou, Jianwu Li, Shunzhou Wang, Ran Tao, and Jianbing Shen. Matnet: Motion-attentive transition network for zero-shot video object segmentation. IEEE Transactions on Image Processing, 29:8326-8338, 2020. 3 \ No newline at end of file diff --git a/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/images.zip b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bad76b2b790a7827016fbba50de9e4a3d018b2a9 --- /dev/null +++ b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80c9d2be376d5d76e843aafad791026f43125a6f3b3b2f9ea60ee37b6737edac +size 621263 diff --git a/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/layout.json b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c9d46481ea2fe9d2b1d800ee185f73705f7a29ec --- /dev/null +++ b/2024/VideoCutLER_ Surprisingly Simple Unsupervised Video Instance Segmentation/layout.json @@ -0,0 +1,7351 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 122 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 122 + ], + "type": "text", + "content": "VideoCutLER: Surprisingly Simple Unsupervised Video Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 143, + 154, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 143, + 154, + 158 + ], + "spans": [ + { + "bbox": [ + 82, + 143, + 154, + 158 + ], + "type": "text", + "content": "Xudong Wang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 143, + 238, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 143, + 238, + 172 + ], + "spans": [ + { + "bbox": [ + 179, + 143, + 238, + 172 + ], + "type": "text", + "content": "Ishan Misra \nUC I" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 262, + 144, + 321, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 144, + 321, + 172 + ], + "spans": [ + { + "bbox": [ + 262, + 144, + 321, + 172 + ], + "type": "text", + "content": "Ziyun Zeng \neley FAI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 346, + 144, + 416, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 144, + 416, + 172 + ], + "spans": [ + { + "bbox": [ + 346, + 144, + 416, + 172 + ], + "type": "text", + "content": "Rohit Girdhar \nMeta AI" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 439, + 144, + 511, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 439, + 144, + 511, + 156 + ], + "spans": [ + { + "bbox": [ + 439, + 144, + 511, + 156 + ], + "type": "text", + "content": "Trevor Darrell" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 201, + 174, + 390, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 174, + 390, + 184 + ], + "spans": [ + { + "bbox": [ + 201, + 174, + 390, + 184 + ], + "type": "text", + "content": "Code: https://github.com/facebookresearch/CutLER" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 208, + 541, + 345 + ], + "blocks": [ + { + "bbox": [ + 53, + 194, + 237, + 206 + ], + "lines": [ + { + "bbox": [ + 53, + 194, + 237, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 237, + 206 + ], + "type": "text", + "content": "Learning Unsupervised Video Instance Segmentation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 208, + 541, + 345 + ], + "lines": [ + { + "bbox": [ + 53, + 208, + 541, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 541, + 345 + ], + "type": "image", + "image_path": "0f68c0f203457877cd08e5f42d4db5c8f6ae6cc3a1767868743dc8aa6bf0ff24.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 546, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 546, + 447 + ], + "type": "text", + "content": "Figure 1. VideoCutLER is a simple unsupervised video instance segmentation method (UnVIS). We show the first competitive unsupervised results on the challenging YouTube VIS benchmark. Moreover, unlike most prior approaches, we demonstrate that UnVIS models can be learned without relying on natural videos and optical flow estimates. Row 1: We propose VideoCutLER, a simple cut-synthesis-and-learn pipeline that involves three main steps. Firstly, we generate pseudo-masks for multiple objects in an image using MaskCut [35]. Then, we convert a random pair of images in the minibatch into a video with corresponding pseudo mask trajectories using ImageCut2Video. Finally, we train an unsupervised video instance segmentation model using these mask trajectories. Row 2: Despite being trained only on unlabeled images, at inference time VideoCutLER can be directly applied to unseen videos and can segment and track multiple instances across time (Fig. 1a), even for small objects (Fig. 1b), objects that are absent in specific frames (Fig. 1c), and instances with high overlap (Fig. 1d). Column 2: Our method surpasses the previous SOTA method OCLR [37] by a factor of 10 in terms of class-agnostic " + }, + { + "bbox": [ + 46, + 346, + 546, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{\\mathrm{video}}" + }, + { + "bbox": [ + 46, + 346, + 546, + 447 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 458, + 192, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 458, + 192, + 470 + ], + "spans": [ + { + "bbox": [ + 143, + 458, + 192, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "content": "Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving " + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "content": " AP" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "inline_equation", + "content": "_{50}" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "content": "^" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "inline_equation", + "content": "\\text{video}" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "content": ", surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by " + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "content": " on YouTubeVIS-2019 in terms of AP" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "inline_equation", + "content": "_{\\text{video}}" + }, + { + "bbox": [ + 46, + 496, + 289, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 458, + 386, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 458, + 386, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 458, + 386, + 470 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 482, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 566 + ], + "type": "text", + "content": "Video instance segmentation is vital for various computer vision applications, e.g. video surveillance, autonomous driving, and video editing, yet labeled videos are costly to obtain. Hence, there is a pressing need to devise an unsupervised video instance segmentation approach that can comprehend video content comprehensively and operate in general domains without labels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": "Prior work in this area typically relies on an optical flow network as an off-the-shelf motion estimator [30, 37, 38]. Although optical flow can be informative in detecting pixel motion between frames, it is not always a reliable technique, particularly in the presence of occlusions, motion blur, complex motion patterns, changes in illuminations, etc. As a result, models that heavily rely on optical flow estimations may fail in several common scenarios. For example, stationary or slowly moving objects may have flow estimates similar to the background, causing them to be omitted in the segmentation process (e.g., the parrot with negligible motion is missed in Fig. 2a). Similarly, non" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22755" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 287, + 395 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 287, + 395 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 287, + 395 + ], + "type": "image", + "image_path": "e4147f3b71f819a0116a06564cc88ba474d58ae395e6f08ae4d0409e8c15e59f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 398, + 287, + 552 + ], + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 552 + ], + "type": "text", + "content": "Figure 2. Challenges encountered by the previous state-of-the-art OCLR: Within the framework of OCLR [37], a method that heavily relies on optical flows as model inputs, several distinct failure cases emerge. These include situations where the method struggles to accurately segment both moving and static objects (as demonstrated in Fig. 2a), struggles to effectively track non-rigid objects as a coherent unit (Fig. 2b), encounters difficulties in distinguishing overlapping instances (Fig. 2c), and fails to maintain consistent predictions under varying illumination conditions (Fig. 2d). Nonetheless, many of these challenges can be effectively addressed through the application of our proposed approach, VideoCutLER, without being reliant on the optical estimations used by various prior works [37, 38]. We present qualitative comparisons using the YouTubeVIS dataset [39]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "rigid objects with non-consistent motions for several parts have varying optical flows, leading to a failure in segmenting all parts cohesively as a unit if object motion is presumed constant (Fig. 2b). Also, objects with similar motion patterns and high overlap are complex for optic flow methods to accurately distinguish between them, especially in boundary regions (Fig. 2c). Finally, objects with illumination changes across frames can cause optical-flow based models to produce non-consistent and blurred segmentation masks (Fig. 2d). Given the limitations above, we advocate for unsupervised video segmentation models which do not" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "depend on optical flow estimates. We propose a method to train a video segmentation model by generating simple synthetic videos from individual images, without relying on explicit motion estimates or requiring labeled natural videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "content": "Our method, VideoCutLER, is an unsupervised Video instance segmentation model that employs a Cut-synthesis-and-LEaRn pipeline (Fig. 1). First, given unlabeled images, we extract pseudo-masks for multiple objects in an image using MaskCut [35], leveraging a self-supervised DINO [4] and a spectral clustering method Normalized Cuts [28] (details in Sec. 3.1). Second, given unlabeled images and their pseudo-masks in a minibatch, we propose ImageCut2Video, a surprisingly simple video synthesis scheme that generates a video from those with corresponding pseudo mask trajectories (details in Sec. 3.2). Finally, those mask trajectories are used to train a video instance segmentation model, aiming to perform object segmentation with temporal consistency across video frames (details in Sec. 3.3). Our model learns to segment and track object instances based on their appearance (feature) similarities across video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": "Despite being learned from only unlabeled images (and the temporally simple synthetic video sequences we construct from them), VideoCutLER succeeds at multi-instance video segmentation, achieving a new state-of-the-art (SOTA) performance of " + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": " AP" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "^{\\text{video}}" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": " on YouTubeVIS2019. This result surpasses the previous SOTA [37] by substantial margins of " + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "45.9\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": "). This result also considerably narrows the performance gap between supervised and unsupervised learning, reducing it from " + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "29.1\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "11.0\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "text", + "content": " in terms of the AP" + }, + { + "bbox": [ + 304, + 315, + 546, + 435 + ], + "type": "inline_equation", + "content": "^{\\text{video}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "text", + "content": "Moreover, most prior works on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers being randomly initialized. In contrast, our pretraining strategy takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, our method outperforms DINO's [4] AP" + }, + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "inline_equation", + "content": "^{\\text{video}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "text", + "content": " on YoutubeVIS-2019 by " + }, + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Contributions. Insights: We found that a simple video synthesis method yield surprisingly effective results for training unsupervised multi-instance video segmentation models. This efficacy is achieved without the necessity of explicit motion estimates or the utilization of natural videos (relying solely on unlabeled ImageNet data suffices), a novel aspect that has not been previously demonstrated in the field. Methods: We propose a simple yet effective cut-synthesize-and-learn pipeline VideoCutLER for learning video instance segmentation models, given unlabeled images. Results: Our method shows the first successfully results on challenging unsupervised multi-instance video segmentation benchmark YouTubeVIS, outperforming the previous SOTA model's " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{\\mathrm{video}}" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " by a large margin." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22756" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 151 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 151 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 151 + ], + "type": "table", + "html": "
CRWDINOOCLROurs
Segment multiple objects✓✗✓✓
Track objects across frames✓✗✓✓
No need for optical flow✓✓✗✓
No 1st-frame ground-truth✗✗✓✓
No human labels at any stage✗✗✓†✓
Pretrained model for sup. learning✗✓✗✓
", + "image_path": "610b53dddac6bc544d2a14bf6d54bc86567bb77fcbdd0865f36313bc03f98686.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 154, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 287, + 232 + ], + "type": "text", + "content": "Table 1. We compare previous methods on unsupervised video instance segmentation, including CRW [17], DINO [4], and OCLR [37], with our VideoCutLER in term of key properties. Our VideoCutLER is the only approach that fulfills all these desired properties. " + }, + { + "bbox": [ + 46, + 154, + 287, + 232 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 154, + 287, + 232 + ], + "type": "text", + "content": ": The optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data and human-annotated data like KITTI-2015 [18] and HD1K [19]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 251, + 134, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 134, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 134, + 262 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 498 + ], + "type": "text", + "content": "Unsupervised video instance segmentation (VIS) requires not only separating and tracking the main moving foreground objects from the background, but also differentiating between different instances, without any human annotations [32]. Previous works [16, 21, 36, 38, 40] on unsupervised video segmentation has primarily centered on unsupervised video object segmentation (VOS), aiming to detect all moving objects as the foreground and to generate a pixel-level binary segmentation mask, regardless of whether the scene contains a single instance or multiple instances. Despite some works exploring unsupervised video instance segmentation (VIS), many of these approaches have resorted to either utilizing first frame annotations [4, 17, 22] to propagate label information throughout the video frames or leveraging supervised learning using large amounts of external labeled data [25, 31, 41, 42]. Furthermore, prior studies typically utilized optical flow networks that were pretrained with human supervision using either synthetic data or labeled natural videos [31, 37, 38, 40]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": "The properties deemed necessary for an unsupervised learning method to excel in video instance segmentation tasks are presented and discussed in Tab. 1. Our proposed method, VideoCutLER, is the only approach that satisfies all these properties, making it an effective and promising solution for unsupervised video instance segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Unsupervised object discovery aims to automatically discover and segment objects in an image in an unsupervised manner [16, 34-36]. LOST [29] and TokenCut [36] focus on salient object detection and segmentation via leveraging the patch features from a pretrained DINO [4] model. For multi-object discovery, FreeSOLO [34] first generates object pseudo-masks for unlabeled images, then learns an unsupervised instance segmentation model using these pseudo-masks. CutLER [35] presents a straightforward cut-and-learn pipeline for unsupervised detection and segmentation of multiple instances. It has demonstrated promising results on more than eleven different benchmarks, covering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 73, + 409, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 409, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 409, + 83 + ], + "type": "text", + "content": "a wide range of domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 84, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 132 + ], + "type": "text", + "content": "In contrast to previous approaches, our unsupervised learning method focuses on simultaneously tracking objects in a video sequence while identifying correspondences between instances across multiple frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": "Self-supervised representation learning generates its own supervision signal by exploiting the implicit patterns or structures present in the input data [3, 4, 14, 15]. Unlike most previous self-supervised learning models, which still require fine-tuning on labeled data to be operative on complex computer vision tasks, such as detection and segmentation, VideoCutLER can tackle these complex, challenging tasks with purely unsupervised learning methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 237, + 395, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 237, + 395, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 395, + 249 + ], + "type": "text", + "content": "3. VideoCutLER" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": "We present VideoCutLER, a simple cut-synthesis-and-learn pipeline consisting of three main steps. First, we generate pseudo-masks for multiple objects in an image using MaskCut (Sec. 3.1). Next, we convert a random pair of images in the minibatch into a synthetic video with corresponding pseudo mask trajectories using ImageCut2Video (Sec. 3.2). Finally, we train an unsupervised video instance segmentation model using these mask trajectories. As the model inputs do not contain explicit motion estimates, it learns to track objects based on their appearance similarity (Sec. 3.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 384, + 518, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 518, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 518, + 397 + ], + "type": "text", + "content": "3.1. Single-image unsupervised segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "text", + "content": "We employ the MaskCut method, introduced in the CutLER [35] method. MaskCut is an efficient spectral clustering approach for unsupervised image instance segmentation and object detection and can discover multiple object masks in a single image without human supervision. MaskCut builds upon a self-supervised DINO model [2] with a backbone of ViT [10] and a cut-based clustering method Normalized Cuts (NCut) [28]. MaskCut first generates a patchwise affinity matrix " + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "inline_equation", + "content": "W_{ij} = \\frac{K_i K_j}{\\|K_i\\|_2\\|\\bar{K}_j\\|_2}" + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "text", + "content": " using the 'key' features " + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "text", + "content": " for patch " + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 402, + 545, + 548 + ], + "type": "text", + "content": " from DINO's last attention layer. Subsequently, the NCut algorithm [28] is employed on the affinity matrix by solving a generalized eigenvalue problem" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 555, + 545, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 555, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 384, + 555, + 545, + 569 + ], + "type": "interline_equation", + "content": "(D - W) x = \\lambda D x \\tag {1}", + "image_path": "f8691b87c659fd25877567fca3cb3f09dc077b40a14eb7c9f334ca521130f13d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": " is a diagonal matrix with " + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "inline_equation", + "content": "d(i) = \\sum_{j} W_{ij}" + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": " is the eigenvector that corresponds to the second smallest eigenvalue " + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": ". Then, the foreground masks " + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "inline_equation", + "content": "M^s" + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": " can be extracted via bi-partitioning " + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 575, + 545, + 660 + ], + "type": "text", + "content": ", which segments a single object within the image. To segment more than one object, Mask-Cut uses an iterative process that masks out the values in the affinity matrix using the extracted foreground mask:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 666, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 666, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 345, + 666, + 545, + 696 + ], + "type": "interline_equation", + "content": "W _ {i j} ^ {t} = \\frac {\\left(K _ {i} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right) \\left(K _ {j} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right)}{\\| K _ {i} \\| _ {2} \\| K _ {j} \\| _ {2}} \\tag {2}", + "image_path": "0f3d0f310b4abd80db59104e361926da528cec7cbd5ba2e2cb1008bf54ea9e39.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 701, + 533, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 533, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 533, + 714 + ], + "type": "text", + "content": "and repeats the NCut algorithm. We set " + }, + { + "bbox": [ + 306, + 701, + 533, + 714 + ], + "type": "inline_equation", + "content": "t = 3" + }, + { + "bbox": [ + 306, + 701, + 533, + 714 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22757" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "Although MaskCut can effectively locate and segment multiple objects in an image, it operates only on a single image, lacking temporal consistency in the instance segmentation masks produced across video frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 256, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 256, + 143 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 256, + 143 + ], + "type": "text", + "content": "3.2. ImageCut2Video Synthesis for Training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 243 + ], + "type": "text", + "content": "We propose a learning-based approach to ensuring temporal consistency in video segmentation masks, based on generating synthetic videos from pairs of individual images and MaskCut masks. Surprisingly, we found that an extremely simple synthetic video generation method yields sufficient training data to learn a powerful video segmentation model that can operate on videos with much greater complexity of motion than is present in the training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 244, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 287, + 316 + ], + "type": "text", + "content": "Given unlabeled images in the minibatch and their pseudo-masks, our ImageCut2Video method synthesizes corresponding videos and pseudo-mask trajectories, thereby allowing us to train the model in an unsupervised manner while offering the necessary supervision for simultaneous detection, segmentation, and tracking of objects in videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "text", + "content": "First, given an image and its corresponding pseudomasks in the mini-batch, we duplicate the image " + }, + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "text", + "content": " times and connect its MaskCut pseudo-masks to form the initial trajectories. This synthetic video, however, only contains static foreground objects. To generate additional trajectories with mobile objects, a second image is randomly selected from the mini-batch, and its objects are cropped using its MaskCut pseudo-masks. These objects are then randomly resized, repositioned, and augmented before being pasted onto the first image. The resulting masks are connected along the temporal dimension to generate additional trajectories with mobile objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": "Specifically, given a target image " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": ", a random source image " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": " in the mini-batch and its corresponding set of binary pseudo-masks " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\{M_2^1,\\dots,M_2^s\\}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": ", we first apply a transformation function " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": " to resize and shift these pseudomasks randomly. This gives us a new set of pseudo-masks " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\{\\hat{M}_2^1,\\dots,\\hat{M}_2^s\\}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{M}_2^s = \\mathcal{T}(M_2^s)" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": ". Next, we synthesize a video with " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": " frames by duplicating image " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": " times and pasting the augmented masks onto " + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 47, + 460, + 287, + 556 + ], + "type": "text", + "content": " using:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 565, + 287, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 287, + 580 + ], + "type": "interline_equation", + "content": "I _ {1} ^ {t} = I _ {1} \\times \\Pi_ {i = 1} ^ {s} (1 - \\hat {M} _ {2} ^ {i}) + I _ {2} \\times \\left(1 - \\Pi_ {i = 1} ^ {s} (1 - \\hat {M} _ {2} ^ {i})\\right) \\quad (3)", + "image_path": "7b3b64fa62757ffcab65de88f27840179e2a3a4a0a278d316b1e85b84e0cef18.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 590, + 237, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 237, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 237, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 590, + 237, + 601 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 47, + 590, + 237, + 601 + ], + "type": "text", + "content": " refers to element-wise multiplication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 196, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 196, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 196, + 623 + ], + "type": "text", + "content": "3.3. Video Segmentation Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 688 + ], + "type": "text", + "content": "During training, the synthetic videos produced by Image-Cut2Video, comprising both mobile and stationary objects, are used as the inputs to train a video instance segmentation model. The segmentation mask trajectories corresponding to each object in the video serve as 'ground-truth' labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We utilize VideoMask2Former [6, 7] with a backbone of ResNet50 [12] as our video instance segmentation (VIS)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "model. It operates by attending to the 3D spatiotemporal features of our synthetic videos and generating 3D volume predictions of pseudo-mask trajectories using shared queries across frames. The shared queries across frames enable the model to segment and track object instances based on their appearance (feature) similarities, making it a powerful framework for analyzing video sequences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 167, + 441, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 441, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 441, + 180 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "type": "text", + "content": "VideoCutLER. We first employ the MaskCut approach on images preprocessed to a resolution of " + }, + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "type": "text", + "content": " pixels. We then compute a patch-wise cosine similarity matrix using the pretrained ViT-Base/8 DINO [4] model, which serves as input to the MaskCut algorithm for initial segmentation mask generation. We set " + }, + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "type": "inline_equation", + "content": "t = 3" + }, + { + "bbox": [ + 304, + 186, + 545, + 319 + ], + "type": "text", + "content": ", which is the maximum number of masks per image. To refine the segmentation masks, we employ a post-processing step using Conditional Random Fields (CRFs) [20], which enforces smoothness constraints and preserves object boundaries, resulting in improved segmentation masks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 319, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 415 + ], + "type": "text", + "content": "Next, we use ImageCut2Video to synthetic videos given images and their pseudo-masks in a mini-batch. We found that synthetic videos with two frames are sufficient to train a video instance segmentation model; therefore, we use " + }, + { + "bbox": [ + 304, + 319, + 545, + 415 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 304, + 319, + 545, + 415 + ], + "type": "text", + "content": " by default. We randomly change the brightness, contrast, and rotation of the masks to create new variations of pseudo-masks. Additionally, we randomly resize the pseudo-masks (scale∈[0.8,1.0]), and shift their positions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "content": "Training and test data. Our model is trained solely on the unlabeled images from ImageNet [8], which comprises approximately 1.3 million images. Without further fine-tuning on any video datasets, we test our model's zero-shot unsupervised video instance segmentation performance on four multi-instance video segmentation benchmarks, including YouTubeVIS-2019 [39], YouTubeVIS-2021 [39], DAVIS2017 [26], and DAVIS2017-Motion [26, 27]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": "YoutubeVIS-2019 and YouTube-VIS2021 contain 2,883 high-resolution YouTube videos and 3,859 high-resolution YouTube videos, respectively. We evaluate the zero-shot unsupervised learning performance on their training splits in a class-agnostic manner. For DAVIS-2017, we evaluate our model's performance on the 30 videos from its val set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Training settings. 1) Unsupervised Image Model Pretraining: We first pretrain a Mask2Former [7] model with a backbone of ResNet50 [12] on ImageNet using MaskCut's pseudo-masks. The model is optimized for 160k iterations, with a batch size of 16 and a learning rate of 0.00002. The learning rate is decayed by a factor of 20 at iteration 80,000. To prevent overfitting, a dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders. 2) Unsupervised Video Model Learning: We initialize the VideoMask2Former model [6] with model weights from the previous stage, and then fine-tune" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22758" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 545, + 152 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 152 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 152 + ], + "type": "table", + "html": "
MethodsTraining settingsYouTubeVIS-2019YouTubeVIS-2021
flowvideossup.\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_{S}\\)\\(AP_{M}\\)\\(AP_{L}\\)\\(AR_{10}\\)\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_{S}\\)\\(AP_{M}\\)\\(AP_{L}\\)\\(AR_{10}\\)
MotionGroup* [38]✓✓✘1.30.10.30.20.30.51.71.10.10.20.10.20.51.5
OCLR* [37]✓✓\\(X^†\\)4.80.41.30.01.25.511.04.40.31.20.11.67.19.6
CutLER‡✘✘✘37.514.617.13.313.927.630.429.210.412.83.112.827.822.6
VideoCutLER✘✓*✘50.724.226.05.620.937.942.438.919.017.15.318.337.531.3
vs. prev. SOTA+12.8+9.6+8.9+2.3+7.0+10.3+12.0+9.7+8.6+4.3+2.2+5.5+9.7+8.7
", + "image_path": "dbfbca98d37a6e97de77d2e233a886e79fe0b97c1267501f50070fd789a90be2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 154, + 546, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 546, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 546, + 233 + ], + "type": "text", + "content": "Table 2. Zero-shot unsupervised multi-instance video segmentation on YouTubeVIS-2019 and YouTubeVIS-2021. We report the instance segmentation metrics (AP and AR) and training settings. *: reproduced MotionGroup [38] and OCLR [37] results with the official code and checkpoints. †: the optical flow estimator OCLR employs (RAFT [30]) is pretrained on both synthetic data [1, 9] and human-annotated data, such as KITTI-2015 [18] and HD1K [19]. ‡: We train a CutLER [35] model with Mask2Former as a detector on ImageNet-1K, following CutLER's official training recipe, and use it as a strong baseline. *: VideoCutLER is trained on synthetic videos generated using ImageNet. Sup and flow denote human supervision and optical flow information, respectively. We evaluate results on YouTubeVIS's train splits in a class-agnostic manner (note: we never train on YouTubeVIS)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "text", + "content": "it on the synthetic videos we construct from ImageNet. We train VideoCutLER on 8 A100 GPUs for 80k iterations, using the AdamW optimizer [24]. We set the initial learning rate to 0.000005 and apply a learning rate multiplier of 0.1 to the backbone. A dropout layer with a rate of 0.3 is added after the self-attention layers of transformer decoders." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "content": "Evaluation metric " + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{AP}^{\\mathrm{video}}" + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{AR}^{\\mathrm{video}}" + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "content": ": The evaluation metrics used in YouTubeVIS are Averaged Precision (AP) and Averaged Recall (AR), which are similar to those used in COCO [23]. The evaluation is specifically conducted at 10 intersection-over-union (IoU) thresholds ranging from " + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "content": " with a step of " + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 324, + 286, + 456 + ], + "type": "text", + "content": " [39]. However, unlike in image instance segmentation, each instance in a video comprises a sequence of masks, so the IoU computation is performed not only in the spatial domain, but also in the temporal domain by summing the intersections at every single frame over the unions at every single frame." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": "Evaluation metric " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": ": For DAVIS [27], we report results using their official evaluation metrics " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": ". The region measure " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "(\\mathcal{J})" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": " [27] is the intersection-over-union (IoU) score between the algorithm's mask and the ground-truth mask. The boundary measure " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": " [27] is the average precision of the boundary of the algorithm's mask. The evaluation metrics are computed separately for each instance, and then the results are averaged over all instances to get the final score. " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": " is the mean of " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 588 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "We evaluate the performance of VideoCutLER on several video instance segmentation benchmarks. In Sec. 5.1, we demonstrate that our approach can effectively perform segmentation and tracking of multiple objects in videos, even when trained on unlabeled ImageNet images without any form of supervision. Our experimental results reveal that our method can drastically reduce the performance gap between unsupervised and supervised learning methods for video instance discovery and tracking. Furthermore, Sec. 5.2 demonstrates that fine-tuning VideoCutLER leads" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 253, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 314 + ], + "type": "text", + "content": "to further performance gains in video instance segmentation, surpassing previous works such as DINO in both fully supervised learning and semi-supervised learning tasks. In Sec. 5.3, we conduct an ablation study to examine the impact of key components and their hyperparameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 322, + 499, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 499, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 499, + 335 + ], + "type": "text", + "content": "5.1. Unsupervised Zero-shot Evaluations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 341, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 376 + ], + "type": "text", + "content": "In this section, we evaluate the performance of our method against previous state-of-the-art approaches on various video instance segmentation benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "type": "text", + "content": "Evaluating unsupervised video instance segmentation poses two main challenges. Firstly, as unsupervised learning methods train the model without semantic classes, the class-aware video segmentation setup cannot be used directly for an evaluation. As a result, following previous works, we evaluate video instance segmentation results in a class-agnostic manner. Secondly, video instance segmentation datasets often annotate only a subset of the objects in the video, which makes Average Recall (AR) a valuable metric that does not penalize models for detecting novel objects not labeled in the dataset [35]. Therefore, we report both AR and AP for YouTubeVIS. Regarding DAVIS, we use the official unsupervised learning metrics " + }, + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{I},\\mathcal{F}" + }, + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 377, + 545, + 567 + ], + "type": "text", + "content": ". All these metrics assess the performance of unsupervised video instance segmentation in a class-agnostic manner. Sec. 4 lists more details on evaluation metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "content": "Detailed comparisons on YouTubeVIS. Tab. 2 presents a summary of the results for unsupervised zero-shot video instance segmentation on the YouTubeVIS-2019 and YouTubeVIS-2021 datasets. We compare our method's results with the previous state-of-the-art methods OCLR [37] and motion grouping [38]. We reproduce their results using their official code and checkpoints to ensure fairness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Although OCLR [37] is also trained on synthetic videos, it relies on the off-the-shelf optical flow estimator RAFT [30] to compute optical flows for RGB sequences. It is worth noting that RAFT is pretrained on a combination of synthetic videos [1, 9] and human-annotated videos" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22759" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 549, + 148 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 549, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 549, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 549, + 148 + ], + "type": "table", + "html": "
MethodsTraining settingsDAVIS2017DAVIS2017-Motion
flowvideossup.training dataJ&FJ(Mean)F(Mean)J&FJ(Mean)F(Mean)
MotionGroup (sup.) [38]✓✓XIN-1K+synthetic---39.544.934.2
Mask R-CNN (w/ flow)* [13, 37]✓✓XIN-1K+synthetic---50.350.450.2
OCLR (w/ flow)* [37]✓✓XIN-1K+synthetic39.638.241.155.154.555.7
VideoCutLERXXXIN-1K43.641.745.557.357.457.2
vs. prev. SOTA+4.0+3.5+4.4+2.2+2.9+1.5
", + "image_path": "09b3a8deab7f9414af636a6b73e588373d1c330d4312fa610303ba8598cf108e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 152, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 218 + ], + "type": "text", + "content": "Table 3. Zero-shot unsupervised single/few-instance segmentation. VideoCutLER also outperforms the previous state-of-the-arts on DAVIS2017 and DAVIS2017-Motion. Note: 12 out of 30 videos from DAVIS2017 and 26 out of 30 videos from DAVIS2017-Motion contain only 1 moving instance. Additionally, DAVIS datasets focus solely on the performance of moving prominent objects, even in videos where multiple objects are present. This disadvantages our model since it can segment both static and moving objects and has not been exposed to any downstream videos during training. *: utilize optical flow predictions from RAFT [30], which is pretrained on external videos. All methods are evaluated in a zero-shot manner, i.e. no fine-tuning on target videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": "such as KITTI-2015 [18] and HD1K [19]. Our approach, VideoCutLER, despite not using any optical flow estimations like many previous works on unsupervised video segmentation, achieves over " + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": " higher " + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "18 \\times" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": " higher AP than OCLR [37] on YouTubeVIS-2019. Additionally, we achieve over " + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": " higher recall. Furthermore, unlike the previous state-of-the-art method OCLR [37], which exhibits poor performance in segmenting small objects (with " + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "0.0\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_S" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": "), our approach significantly outperforms it. Similar performance gains can be observed on YouTubeVIS-2021. Finally, the performance gains to CutLER [35] demonstrates the effectiveness of VideoCutLER in training unsupervised multi-instance video segmentation models, surpassing CutLER by over " + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 407 + ], + "type": "text", + "content": " on YouTubeVIS-2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 409, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 288, + 445 + ], + "type": "text", + "content": "In Fig. 3, we present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 579 + ], + "type": "text", + "content": "Detailed comparisons on DAVIS. To provide a comprehensive evaluation and comparison with existing unsupervised video instance segmentation approaches, we also assess the performance of our model on the validation sets of DAVIS-2017 and DAVIS2017-Motion [27, 37]. Note that both DAVIS2017 and DAVIS2017-Motion datasets focus only on the performance of instance segmentation on prominent moving objects, even in videos with multiple objects. As a result, only a single or a few objects of interest per video are annotated, which may not reflect the challenges that arise when multiple objects are present." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Although the evaluation of DAVIS is an unfair assessment for us since VideoCutLER is supposed to segment both static and moving objects, whereas DAVIS focuses on moving prominent objects, with only a single or a few moving objects of interest per video annotated. However, Tab. 3 shows that VideoCutLER yields approximately " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " higher " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}, \\mathcal{F}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". The additional results on DAVIS demonstrate that VideoCutLER achieves superior performance not only on static or minimally moving objects but also on dynamic objects, where prior methods relying on optical flow estimates can benefit from additional cues." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 237, + 533, + 422 + ], + "blocks": [ + { + "bbox": [ + 319, + 237, + 533, + 422 + ], + "lines": [ + { + "bbox": [ + 319, + 237, + 533, + 422 + ], + "spans": [ + { + "bbox": [ + 319, + 237, + 533, + 422 + ], + "type": "image", + "image_path": "b41978343a5dfba7fb8df916cab7667635ebfd12e22f7b1e42678cbd948a70a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 425, + 547, + 546 + ], + "lines": [ + { + "bbox": [ + 305, + 425, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 547, + 546 + ], + "type": "text", + "content": "Figure 3. We present qualitative visualizations illustrating the zero-shot unsupervised video instance segmentation outcomes of VideoCutLER on YouTubeVIS dataset. It's noteworthy that VideoCutLER is solely pretrained on image dataset ImageNet-1K, and its evaluation is conducted directly on the video dataset YouTubeVIS (no further fine-tuning required). The visual results provided effectively highlight that VideoCutLER is capable of segmenting and tracking multiple instances, delivering consistent tracking results across video frames, and successfully distinguishing between various instances, even when significant overlapping occurs. We show more demo results in appendix." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Comparison of supervised and unsupervised learning in object discovery and tracking abilities is presented in Tab. 4. We train a supervised MaskTrack R-CNN [39] model on the human-annotated training set of YouTubeVIS-2019 dataset, and evaluate it in a class-agnostic manner on the videos that are not shared between YouTubeVIS-2019 and YouTubeVIS-2021 datasets [39]. Tab. 4 shows that our VideoCutLER model significantly narrows the gap between supervised learning and unsupervised learning methods in terms of the averaged precision " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " (gaps: " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "29.1\\% \\rightarrow 11.0\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ") and the averaged recall " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{100}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " (gaps: " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "14.9\\% \\rightarrow 3.2\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "), particularly for the " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{100}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22760" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 158 + ], + "type": "table", + "html": "
MethodsTraining settingsYouTubeVIS-2021\\YouTubeVIS-2019
flowvideossup.training data\\(AP_{50}\\)\\(AP_{75}\\)AP\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)\\(AR_{100}\\)
Mask2Former [6]✓✓✘IN-1K+YT201948.922.224.9----
MaskTrack R-CNN* [39]✓✓✓IN-1K+YT201932.413.015.08.424.939.020.3
MaskTrack R-CNN* [39]✓✓✓IN-1K+COCO+YT201935.818.718.710.531.346.824.5
\\(OCLR^*\\)[37]✓✓✘IN-1K+synthetic3.30.21.00.32.77.55.4
VideoCutLER✘✘✘IN-1K21.47.19.04.913.329.617.1
vs. prev. SOTA+18.1+6.9+8.0+4.6+10.6+22.1+11.7
", + "image_path": "fdc73412d3f1da91b35c982f92d49781c16920b70b0762a6462f007c1fed3318.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 211, + 544, + 272 + ], + "blocks": [ + { + "bbox": [ + 46, + 160, + 546, + 206 + ], + "lines": [ + { + "bbox": [ + 46, + 160, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 546, + 206 + ], + "type": "text", + "content": "Table 4. VideoCutLER greatly narrows the gap between fully-supervised learning and unsupervised learning for multi-instance video segmentation. Results are evaluated in a class-agnostic manner on the relative complement of the set of videos from YouTubeVIS-2021 and the set of videos from YouTubeVIS-2019. VideoCutLER and Mask2Former use a backbone of ResNet50. *: reproduced results with the official code and checkpoints. IN-1K refers to ImageNet-1K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 211, + 544, + 272 + ], + "lines": [ + { + "bbox": [ + 50, + 211, + 544, + 272 + ], + "spans": [ + { + "bbox": [ + 50, + 211, + 544, + 272 + ], + "type": "table", + "html": "
MethodsArchitectureYouTubeVIS-2019YouTubeVIS-2021
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP_S\\)\\(AP_M\\)\\(AP_L\\)
DINO [4]Mask2Former [6]23.039.023.76.028.034.224.641.425.98.734.039.9
VideoCutLERMask2Former [6]38.956.743.322.143.151.833.453.836.315.740.954.8
vs. prev. SOTA+15.9+17.7+19.6+16.1+15.1+17.6+8.8+12.4+10.4+7.0+6.9+14.9
", + "image_path": "fef4583e4be72f779d10df3886701aaeb127110f1cb3ecbcbd15860afa468a9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 546, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 546, + 309 + ], + "type": "text", + "content": "Table 5. VideoCutLER can serve as a strong pretrained model for the supervised video instance segmentation task. The video segmentation model, Mask2Former, is initialized with various pretrained models, i.e., DINO or VideoCutLER, and fine-tuned on the training set with human annotations. We report the instance segmentation metrics and evaluate the model performance on the val splits." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 53, + 327, + 167, + 453 + ], + "blocks": [ + { + "bbox": [ + 53, + 327, + 167, + 453 + ], + "lines": [ + { + "bbox": [ + 53, + 327, + 167, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 167, + 453 + ], + "type": "image", + "image_path": "6efe8b2c7632c6b706d69504d99e206a5a445cebbdb508260abaa639a101be63.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 455, + 287, + 555 + ], + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 555 + ], + "type": "text", + "content": "Figure 4. We fine-tune VideoCutLER for semi-supervised video instance segmentation on the YouTubeVIS-2019 dataset, using different percentages of labeled training data. We evaluate the performance of our method by reporting the average precision and recall on the validation set of YouTubeVIS-2019. To establish a strong baseline, we use the self-supervised DINO [4] model and initialize the weights of VideoMask2Former with DINO. To ensure a fair comparison, both baselines and VideoCutLER are trained using the same schedule and recipe." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 327, + 281, + 453 + ], + "blocks": [ + { + "bbox": [ + 168, + 327, + 281, + 453 + ], + "lines": [ + { + "bbox": [ + 168, + 327, + 281, + 453 + ], + "spans": [ + { + "bbox": [ + 168, + 327, + 281, + 453 + ], + "type": "image", + "image_path": "c637624ef2de954a2e9705afd24045a701cabe73cd7241073af1215c72d1c979.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 288, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 288, + 588 + ], + "type": "text", + "content": "5.2. Label-Efficient and Fully-Supervised Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": "In this section, we investigate VideoCutLER as a pretraining approach for supervised video instance segmentation models, and evaluate its effectiveness in label-efficient and fully-supervised learning scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Setup. We use VideoMask2Former with a backbone of ResNet50 for all experiments in this section unless otherwise noted. For our experiments on semi-supervised learning, we randomly sample a subset of videos from the training split with different proportions of labeled videos. After pretraining our VideoCutLER model on ImageNet, we" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 328, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 437 + ], + "type": "text", + "content": "fine-tune the model on the YouTubeVIS-2019 [39] dataset with its human annotations. For our experiments on the fully-supervised learning task, we fine-tune the VideoCUTLER model on all available labeled data from the training sets of YouTubeVIS. For baselines, we initialize a VideoMask2Former model with a DINO [4] model pre-trained on ImageNet and fine-tuned on labeled videos. Since DINO has shown strong performance in detection and segmentation tasks, it serves as a strong baseline for our experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "text", + "content": "For semi-supervised learning, both the baselines and our models are trained for " + }, + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "text", + "content": " schedule, with a learning rate of 0.0001 for all model weights, except for the final classification layers, which use a learning rate of 0.0016. We train the models using a batch size of 16 and 8 GPUs. For fully-supervised learning, we use the " + }, + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "text", + "content": " schedule and a learning rate of 0.0002 for the final classification layers. We evaluate their performance on the val split of the YouTubeVIS-2019, and report results from its official evaluation server." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 664 + ], + "type": "text", + "content": "Data for fully-/semi-supervised VIS. We fine-tune the pretrained VideoCutLER model on all or a subset of the training split of YouTubeVIS-2019. We then evaluate the resulting models on the validation set. To ensure a fair comparison, we use the same amount of human annotations to train our model and baselines. Specifically, we initialize the baselines with the DINO-pretrained model and fine-tune them on the training set of the respective dataset. We evaluate the model performance on their validation sets and report results from its official evaluation server." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Results. Most prior approaches on self-supervised representation learning [4, 5, 11, 14, 33] are limited to providing initializations only for the model backbones, with the remaining layers, such as Mask2Former's decoders," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22761" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 277 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 277 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 277 + ], + "type": "image", + "image_path": "d03c2d4a538f9fe12a35560cb7eb60fc1dd56634fdccec330bb129b769550f59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 280, + 287, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 346 + ], + "type": "text", + "content": "Figure 5. We present qualitative results on videos covering a range of out-of-domain sources, e.g., sketches, 3D computer-generated imagery (CGI) and hybrid (CGI + realistic). VideoCutLER can produce high-quality segmentation and tracking results for small objects that are often difficult to distinguish from the background, as well as for object sketches that lack textual information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": "being randomly initialized. In contrast, VideoCutLER takes a more comprehensive approach that allows all model weights to be pretrained, resulting in a stronger pretrained model better suited for supervised learning. As a result, as shown in Fig. 4 and Tab. 5, our method outperforms these prior works significantly, offering a strong pretrained model for fully-/semi-supervised learning tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": "Fig. 4 shows that VideoCutLER consistently outperforms the strong baseline method DINO [4] across all label-efficient learning settings with varying proportions of labeled YouTubeVIS-2019 videos. The most significant performance gains are observed when " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": " labeled data is provided, where VideoCutLER exceeds DINO by over " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "12\\% \\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": " AR. As demonstrated in Tab. 5, training the model with all available labeled videos from YouTubeVIS yields considerable performance gains, surpassing DINO by more than " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": " AP on YouTubeVIS-2019 and " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": " on YouTubeVIS-2021, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 141, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 141, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 141, + 611 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Hyper-parameters and design choices. We present an ablation study on several key hyper-parameters and design choices of VideoCutLER in Tab. 6. First, we analyze the impact of varying the size of video frames used for training VideoCutLER. From Tab. 6a, we observe that the shortest edge length of 240 pixels yields the best performance. Using a larger resolution does not always lead to better results. Next, Tab. 6b shows the effect of the number of frames used" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 312, + 69, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 312, + 69, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 312, + 69, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 69, + 544, + 137 + ], + "type": "table", + "html": "
Size → 180360480#frames → CutLER† [35]234
APvideo5049.950.7APvideo5037.549.850.750.4
(a) Frame size.(b) # frames. z
Augmentations → none+bright+rotation+contrast+cropall
APvideo5047.848.148.948.348.750.7
", + "image_path": "098114b25bbff4ca222615eb64b5b789735358ea064a0e30499875d2cefd7000.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 336, + 139, + 512, + 150 + ], + "lines": [ + { + "bbox": [ + 336, + 139, + 512, + 150 + ], + "spans": [ + { + "bbox": [ + 336, + 139, + 512, + 150 + ], + "type": "text", + "content": "(c) Data augmentations for ImageCut2Video." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 154, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 154, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 154, + 545, + 232 + ], + "type": "text", + "content": "Table 6. Ablations for VideoCutLER. We report video instance segmentation result " + }, + { + "bbox": [ + 305, + 154, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{\\mathrm{video}}" + }, + { + "bbox": [ + 305, + 154, + 545, + 232 + ], + "type": "text", + "content": " on YoutubeVIS-2019. (a) The impact of varying video frame sizes on training VideoCutLER. (b) The effect of the number of frames used for model training. (c) The impact of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. Default settings are highlighted in gray." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 251, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 546, + 468 + ], + "type": "text", + "content": "for training video instance segmentation models. We found that synthetic videos with three frames are optimal for learning an unsupervised video instance segmentation model. Increasing the number of frames does not result in a further improved performance, aligning with the findings reported in [6]. Furthermore, Tab. 6c investigates the contribution of several augmentation methods, including brightness, rotation, contrast, and random cropping, which are used as default during model training. We found that compared to ImageCut2Video without any data augmentations, adding these augmentations can bring about " + }, + { + "bbox": [ + 304, + 251, + 546, + 468 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 251, + 546, + 468 + ], + "type": "text", + "content": " performance gains. Generalizability. The results presented in Fig. 5 demonstrate that VideoCutLER can effectively perform video instance segmentation on out-of-domain data sources, e.g. sketches, 3D computer-generated imagery, and hybrid videos that combine CGI. These results show that our model can be applied to a broad range of videos beyond the domains it was initially trained on, i.e., ImageNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 479, + 372, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 372, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 372, + 491 + ], + "type": "text", + "content": "6. Summary" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": "We presented a simple unsupervised approach to segment multiple instances in a video. Our approach, VideoCutLER, does not require labels, and does not rely on motion-based learning signals like optical flow. In fact, VideoCutLER does not need real videos for training as we synthesize videos using natural images from the ImageNet-1K. Despite being simpler, VideoCutLER outperforms models that use additional learning signals or video data, achieving " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": " their performance on benchmarks like YouTubeVIS. Moreover, VideoCutLER is a strong pretrained model for supervised learning. We hope that our approach enables both a wide range of applications in video recognition, as well as its simplicity enables easy future research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Limitations: while VideoCutLER demonstrates its capability to achieve the state-of-the-art performance without relying on optical flow estimations, potential further improvements may be obtained by leveraging natural videos and integrating joint training with optical flow estimations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22762" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "text", + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 324 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "text", + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 4, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 370, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 287, + 425 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 469 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 470, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 287, + 536 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 601 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "text", + "content": "[11] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "text", + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[16] Yuan-Ting Hu, Jia-Bin Huang, and Alexander G Schwing. Unsupervised video object segmentation using motion saliency-guided spatio-temporal propagation. In Proceedings of the European conference on computer vision (ECCV), pages 786-802, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "text", + "content": "[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "text", + "content": "[18] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In Proceedings of the IEEE international conference on computer vision, pages 3271-3279, 2015. 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 361, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 447 + ], + "type": "text", + "content": "[19] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The hci benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 19-28, 2016. 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[20] Philipp Krahenbihl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. NeurIPS, 2011. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[21] Minhyeok Lee, Suhwan Cho, Seunghoon Lee, Chaewon Park, and Sangyoun Lee. Unsupervised video object segmentation via prototype memory network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5924-5934, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[22] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 658 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Jonathon Luiten, Idil Esen Zulfikar, and Bastian Leibe. Unovost: Unsupervised offline video object segmentation and tracking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2000-2009, 2020. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22763" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[26] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "type": "text", + "content": "[27] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 218 + ], + "type": "text", + "content": "[28] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "type": "text", + "content": "[29] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "text", + "content": "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 388 + ], + "type": "text", + "content": "[31] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "text", + "content": "[32] Wenguan Wang, Tianfei Zhou, Fatih Porikli, David Crandall, and Luc Van Gool. A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "text", + "content": "[33] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 554 + ], + "type": "text", + "content": "[34] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[35] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "text", + "content": "[36] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[37] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 308 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[38] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[39] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 547, + 228 + ], + "type": "text", + "content": "[40] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 230, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 546, + 262 + ], + "type": "text", + "content": "[41] Zhao Yang, Qiang Wang, Song Bai, Weiming Hu, and Philip HS Torr. Video segmentation by detection for the 2019 unsupervised davis challenge. 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 263, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 547, + 308 + ], + "type": "text", + "content": "[42] Tianfei Zhou, Jianwu Li, Shunzhou Wang, Ran Tao, and Jianbing Shen. Matnet: Motion-attentive transition network for zero-shot video object segmentation. IEEE Transactions on Image Processing, 29:8326-8338, 2020. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22764" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_content_list.json b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f91c57dbc55b6893d23e821d1866b14de5924c14 --- /dev/null +++ b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_content_list.json @@ -0,0 +1,1660 @@ +[ + { + "type": "text", + "text": "VideoGrounding-DINO: Towards Open-Vocabulary Spatio-Temporal Video Grounding", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Syed Talal Wasim1 Muzammal Naseer1 \nSalman Khan1,2 Ming-Hsuan Yang3,4 Fahad Shahbaz Khan1,5", + "bbox": [ + 218, + 202, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Mohamed bin Zayed University of AI $^{2}$ Australian National University $^{3}$ University of California, Merced $^{4}$ Google Research $^{5}$ Linköping University", + "bbox": [ + 153, + 244, + 790, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video grounding aims to localize a spatio-temporal section in a video corresponding to an input text query. This paper addresses a critical limitation in current video grounding methodologies by introducing an Open-Vocabulary Spatio-Temporal Video Grounding task. Unlike prevalent closed-set approaches that struggle with open-vocabulary scenarios due to limited training data and predefined vocabularies, our model leverages pre-trained representations from foundational spatial grounding models. This empowers it to effectively bridge the semantic gap between natural language and diverse visual content, achieving strong performance in closed-set and open-vocabulary settings. Our contributions include a novel spatio-temporal video grounding model, surpassing state-of-the-art results in closed-set evaluations on multiple datasets and demonstrating superior performance in open-vocabulary scenarios. Notably, the proposed model outperforms state-of-the-art methods in closed-set settings on VidSTG (Declarative and Interrogative) and HC-STVG (V1 and V2) datasets. Furthermore, in open-vocabulary evaluations on HC-STVG V1 and YouCook-Interactions, our model surpasses the recent best-performing models by $4.88\\mathrm{m\\_vIoU}$ and $1.83\\%$ accuracy, demonstrating its efficacy in handling diverse linguistic and visual concepts for improved video understanding. Our codes will be publicly released.", + "bbox": [ + 76, + 348, + 473, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatio-temporal video grounding is pivotal in linking visual content with natural language descriptions, thus facilitating semantics interpretation within visual data. Prevailing approaches in video grounding such as TubeDETR [28], STCAT [9], and STVGFormer [11] focus mainly on supervised closed-set settings, where models are trained on specific datasets [24, 32] with predefined vocabulary and meticulously annotated data. While these current state-of", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83b2e303b48c452304cd569f4a882c31b066146be3d773ffec2a4b4823ab5291.jpg", + "image_caption": [ + "Figure 1. Performance comparison on conventional closed-set and open-vocabulary settings for the video grounding task. We compare our approach with TubeDETR [28] and STCAT [9] in supervised setting for VidSTG [32] declarative/interrogative and HC-STVG V1 [24], along with open-vocabulary evaluation on HC-STVG V1 and YouCook-Interactions [23] datasets." + ], + "image_footnote": [], + "bbox": [ + 506, + 319, + 883, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the-art methods excel in closed-set settings on datasets like VidSTG [32] and HC-STVG [24], their limited generalization beyond training dataset distributions poses a significant challenge. The relatively small scale and restricted sample variety in existing video datasets hinder models from adapting to unseen scenarios effectively.", + "bbox": [ + 496, + 670, + 892, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the inherent limitation of supervised closed-set approaches in terms of their restricted vocabulary, this paper investigates Open-Vocabulary Spatio-Temporal Video Grounding. Unlike conventional methodologies, this paradigm addresses challenges posed by the unrestricted diversity of language and visual concepts in the wild. The goal is to train on a set of base categories and generalize to unseen objects/actions based on the open-vocabulary nature of backbone models. This paper ex", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18909", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plores the challenges and opportunities inherent in open-vocabulary video grounding, laying the groundwork for more robust and versatile video understanding.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, training an effective open-vocabulary video grounding model would require a large enough dataset with a rich set of natural language expressions and corresponding spatio-temporal localizations. Such an extensive dataset can allow the model to learn generalized visual and textual representations and handle out-of-distribution samples. However, video grounding datasets [24, 32] are quite limited in scale, e.g., VidSTG has only $5.4\\mathrm{k}$ training videos with $80.6\\mathrm{k}$ distinct sentences. In contrast, an image grounding model GLIP [10] is trained with $\\sim 26.5\\mathrm{M}$ image-text pairs. Therefore, our research explores the following fundamental question: How spatio-temporal video grounding models can achieve robust performance in both Closed-Set and Open-Vocabulary scenarios without requiring large-scale video annotations, ensuring effective generalization beyond training datasets?", + "bbox": [ + 75, + 141, + 470, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addressing this question, we find inspiration in the accomplishments of foundational models specializing in spatial grounding [4, 8, 12, 14, 29]. These models undergo training on an extensive corpus of image-text data, enabling effective generalization to samples from a given target distribution. We aim to harness this pretrained representation to enhance our video grounding model. Our proposed solution is a spatio-temporal video grounding model adopting a DETR-like architecture enhanced by temporal aggregation modules. The spatial modules are initialized using the pretrained representation of a foundational image model [12]. Meanwhile, the image and text feature extractors remain frozen while the video-specific spatio-temporal adaptations are modeled via learnable adapter blocks. This approach is designed to preserve the nuanced representation of the foundational model, enhancing our model's ability to generalize effectively to novel samples.", + "bbox": [ + 75, + 388, + 470, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A summary of our closed-set and open-vocabulary results is shown in Fig. 1, where the proposed approach excels in both settings by a clear margin. Our major contributions are summarized as follows:", + "bbox": [ + 76, + 651, + 468, + 712 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time, we evaluate spatio-temporal video grounding models in an open-vocabulary setting on HC-STVG V1 [24] and YouCook-Interactions [23] benchmarks in a zero-shot manner. We outperform state-of-the-art methods TubeDETR [28] and STCAT [9] by 4.26 m_vIoU and $1.83\\%$ accuracy, respectively.", + "- By combining the strengths of spatial grounding models with complementary video-specific adapters, our approach consistently outperforms the previous state-of-the-art in closed-set setting on four benchmarks, i.e., VidSTG (Declarative) [32], VidSTG (Interrogative) [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + ], + "bbox": [ + 76, + 717, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spatial Grounding Foundation Models: Recent literature introduces notable spatial grounding models. GLIP [10] unifies object detection and phrase grounding through a language-image pre-training model, leveraging extensive image-text pairs for semantic-rich representations. Grounding DINO [12] integrates language with a transformer-based detector to achieve an open-set detector, excelling in benchmarks like COCO and ODinW. Kosmos-1 [8] and Kosmos-2 [14] contribute Multimodal Large Language Models (MLLMs) with capabilities such as zero-shot and few-shot learning, language understanding, and multimodal tasks. Kosmos-2 [14] specifically integrates grounding into downstream applications, introducing GrIT, a large-scale dataset of Grounded Image-Text pairs. Shikra [4] addresses referential ability in MLLMs by handling spatial coordinates in inputs and outputs, showcasing promising performance in various vision-language tasks. Ferret [29] unifies referring and grounding in the LLM paradigm, achieving superior performance in classical referring and grounding tasks, excelling in region-based multimodal chatting and image description. Recently, GLaMM [16] allows pixel-level grounded conversations with an LLM, showcasing generalizability to several captioning and referring segmentation tasks. However, spatial methods cannot work for grounding objects in videos, a gap addressed by this work.", + "bbox": [ + 496, + 114, + 890, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spatio-Temporal Video Grounding: Several methods tackle the challenge of localizing objects in untrimmed videos based on query sentences. STVGBert [21] presents a one-stage visual-linguistic transformer for simultaneous spatial and temporal localization. TubeDETR [28] introduces a transformer-based architecture to model temporal, spatial, and multi-modal interactions efficiently. Augmented 2D-TAN [22] adopts a two-stage approach, enhancing the 2D-TAN with a temporal context-aware Bi-LSTM Aggregation Module. OMRN [33] addresses the challenge of unaligned data and multi-form sentences in spatiotemporal video grounding, proposing an object-aware multi-branch relation network for effective relation discovery. MMN [26] introduces a Mutual Matching Network as a metric-learning framework for temporal grounding, achieving competitive performance. STCAT [9] is an end-to-end one-stage framework addressing feature alignment and prediction inconsistency. Finally, STVGFformer [11] proposes an effective framework with static and dynamic branches for cross-modal understanding. While the above methods advance video grounding, their generalization to out-of-distribution and open-vocabulary samples is limited due to constrained video datasets [24, 32]. To address this issue, we utilize the generalized representation of spatial grounding foundation models [4, 8, 10, 12, 14, 29] trained on a large corpus of image-text data and can perform well on both closed-set and open-vocabulary evaluations.", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 89, + 210, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed above, the current state-of-the-art spatiotemporal video grounding methods [9, 11, 21, 22, 24, 26, 28, 30, 32, 33] primarily evaluate in a supervised setting on the VidSTG [32] and HC-STVG [24] datasets. However, these methods lack the multimodal spatio-temporal understanding required to perform well on out-of-distribution samples [2]. Therefore, this work aims to achieve improved open-vocabulary performance while maintaining strong closed-set video-grounding performance.", + "bbox": [ + 75, + 114, + 467, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We take inspiration from recent foundation models for spatial grounding [4, 8, 10, 12, 14, 29]. These models are trained on a large corpus of visual-textual data and hence, generalize well to unseen samples. We aim to leverage the strong generalization capabilities of such foundation models to achieve strong open-set spatio-temporal video grounding performance. Our proposed spatio-temporal video grounding method uses DETR-like [1] design, with temporal aggregation and adaptation modules for learning video-specific representations.", + "bbox": [ + 75, + 251, + 467, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Below, we explain our proposed methodology. We formally define the spatio-temporal video grounding problem in Sec. 3.1. We then explain our architecture details in Sec. 3.2. We finally explain the loss formulation used to train the model and model initialization in Sec. 3.3.", + "bbox": [ + 75, + 401, + 467, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 484, + 259, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The spatio-temporal video grounding task involves localizing and recognizing objects and actions in a video sequence by integrating spatial and temporal information. In contrast to spatial grounding, which focuses on recognizing and localizing objects or actions within individual frames, spatiotemporal grounding extends this concept to include the temporal dimension. This means understanding where objects or actions are in each frame and how they evolve and move over time.", + "bbox": [ + 75, + 508, + 467, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a video $V \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ with $T$ frames, $H \\times W$ spatial resolution, and $C$ channels, respectively, along with a text prompt $P$ . The spatial grounding problem can be defined as the localization of one or more objects associated with the prompt $P$ in a frame $V_{t}, t \\in \\{1, \\dots, T\\}$ using a bounding box $B_{i}^{t} = (x_{i}^{t}, y_{i}^{t}, w_{i}^{t}, h_{i}^{t})$ , where $x_{i}^{t}$ and $y_{i}^{t}$ are the coordinates of the top-left corner, $w_{i}^{t}$ and $h_{i}^{t}$ are the width and height of the bounding box, $i \\in \\{1, \\dots, N\\}$ is the object number for the frame $t$ . The temporal grounding problem, on the other hand, involves understanding how objects or actions evolve over time. It aims at localizing the temporal interval $(t_{s}, t_{e})$ where the specific action/interaction happens in the entire temporal duration, where interval $(t_{s}, t_{e})$ indicates the start and end frame of the object occurrence within the total frames $T$ ( $1 \\leq t_{s} < t_{e} \\leq T$ ). Hence, the spatio-temporal grounding problem for object $i$ associated with prompt $P$ can be summarized as a", + "bbox": [ + 75, + 642, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set of spatio-temporal coordinates associated with the subset of frames where the object exists: $(x_i^t,y_i^t,w_i^t,h_i^t,t)$ and $t\\in \\{t_s,\\dots,t_e\\}$ . The interval $(t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}$ and is a subset of the total frames $T$", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spatio-Temporal Video Grounding", + "text_level": 1, + "bbox": [ + 498, + 164, + 803, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we explain our video grounding model in Fig. 2. As discussed earlier, we aim to design a spatio-temporal video grounding model that can perform well in closed-set and open-vocabulary settings. Strong open-vocabulary performance requires learning a rich visual/textual representation, which in turn requires a large amount of training data. Unfortunately, spatio-temporal video grounding datasets are quite limited in scale [24, 32], resulting in current video grounding methods failing in generalizing well to out-of-distribution samples because they lack the requisite strong visual/textual representation.", + "bbox": [ + 496, + 188, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To solve this problem, our approach takes inspiration from recent spatial grounding methods [4, 8, 12, 14, 29], which have strong open-vocabulary performance thanks to the large image-text corpus they are trained on. We can utilize the generalized representations of these models to enrich the weaker representation of video-grounding approaches obtained from the limited number of training samples. Our approach aims to leverage the strong pretrained representations of spatial grounding methods to achieve strong closed-set supervised and open-vocabulary video grounding performance. Our spatio-temporal video grounding approach is based on the state-of-the-art DETR-based [1] object detection framework DINO [31] and also borrows concepts of image-text alignment and grounding from Grounded Language-Image Pre-training (GLIP) [10] and Grounding DINO [12]. We extract initial features from backbone vision and text encoders $\\theta_v$ and $\\theta_p$ . Following that, we model inter-frame and intra-frame features and learn cross-modal visual/textual relations in the Cross-Modality Spatio-Temporal Encoder (Sec. 3.2.1). The result enriched cross-modal features are used to initialize queries for each frame (Sec. 3.2.2). These queries are then decoded to predict the bounding boxes per frame and the temporal grounding start/end frame by aggregating information across spatial/temporal dimensions and injecting information through cross-attention from enriched visual/textual context (Sec. 3.2.3).", + "bbox": [ + 496, + 356, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the video $V$ and text prompt $P$ as described above, we obtain per-frame features $F_{v}^{0}$ and text features $F_{p}^{0}$ from vision and text encoders, $\\theta_v$ and $\\theta_p$ , respectively. The vision encoder is based on a Swin Transformer [13], and the text-encoder is defined as a BERT [6] model. Like other DETR-based detectors [31, 35], image features are extracted from different vision-encoder blocks at multiple scales. The features are then passed to the Cross-Modality Spatio-Temporal Encoder.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aee7204402aeb18acb3ae3d1e5eb4d4b0cbaff7cdc169590e9d1cd00e392b2f5.jpg", + "image_caption": [ + "Figure 2. Overall architecture: We present our video grounding architecture. It consists of vision and text encoders that produce visual and textual features. A cross-modality spatio-temporal encoder which fuses information across spatial/temporal dimensions and visual/textual modalities. A language guided query selection module to initialize cross-modal queries. A cross-modality spatio-temporal decoder to decoder queries while fusing information from visual/textual features. And finally two prediction heads to predict the bounding boxes per frame and the temporal tube. Modules with $(\\bullet)$ are trainable and those with $(\\text{串})$ are frozen." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 893, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Cross-Modality Spatio-Temporal Encoder", + "text_level": 1, + "bbox": [ + 76, + 449, + 423, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The initial vision and text features $F_{v}^{0}$ and $F_{p}^{0}$ neither contain any cross-modal information nor model the temporal relationship across frames. Therefore, we further encode the initial features through the Cross-Modality Spatio-Temporal Encoder to model the temporal information across frames and learn cross-modal features.", + "bbox": [ + 75, + 472, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each layer of the $M$ layer encoder first applies a Multi-Head Self-Attention (MHSA) [25] to the visual features $F_{v}$ along the temporal dimension, followed by a Deformable Attention (DA) [35] along the spatial dimension. This is done to model relations within frames and temporally across frames. Similarly, we apply a MHSA on the text features $F_{p}$ . This is illustrated in Eq. 1.", + "bbox": [ + 75, + 564, + 470, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {v} ^ {m \\prime} = \\mathrm {D A} _ {s p a t i a l} ^ {m} (\\mathrm {M H S A} _ {t e m p o r a l} ^ {m} (F _ {v} ^ {m - 1})), \\\\ F _ {p} ^ {m \\prime} = \\mathrm {M H S A} _ {p} ^ {m} (F _ {p} ^ {m - 1}), \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 681, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{v}^{m - 1}$ and $F_{p}^{m - 1}$ are visual and textual input features to layer $m$ , $F_{v}^{m'}$ and $F_{p}^{m'}$ are the intermediate visual and textual feature representation, and $\\mathrm{DA}_{\\text{spatial}}^m$ , $\\mathrm{MHSA}_{\\text{temporal}}^m$ and $\\mathrm{MHSA}_p^m$ are the spatial-deformable, temporal and textual attentions at layer $m \\in \\{1,\\dots,M\\}$ , respectively. Following initial spatial, temporal, and textual attentions, we fuse features across the visual and textual modalities, as done in GLIP [10].", + "bbox": [ + 75, + 733, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, we calculate the joint visual-textual attention, $\\mathrm{Attn}_{\\mathrm{joint}}^{\\mathrm{m}}$ , using projected intermediate features $F_{v}^{m'}$ and $F_{p}^{m'}$ . This attention is then used alongside the", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "intermediate features $F_{v}^{m'}$ and $F_{p}^{m'}$ , to calculate the image-to-text and text-to-image cross-attentions as shown in Eq. 2 and Eq. 3.", + "bbox": [ + 496, + 448, + 892, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A t t n} _ {\\mathrm {j o i n t}} ^ {\\mathrm {m}} = \\left(\\frac {\\operatorname {p r o j} _ {q , v} ^ {m} \\left(F _ {v} ^ {m \\prime}\\right) \\operatorname {p r o j} _ {q , p} ^ {m} \\left(F _ {p} ^ {m \\prime}\\right) ^ {T}}{\\sqrt {d ^ {k}}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 505, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $proj_{q,v}^{m}$ and $proj_{q,p}^{m}$ are the query projections for the visual and textual features, respectively, at layer $m$ .", + "bbox": [ + 496, + 555, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {v} ^ {m} = \\operatorname {F F N} _ {v} ^ {m} \\left(\\operatorname {s o f t m a x} \\left(\\operatorname {A t t n} _ {\\text {j o i n t}} ^ {\\mathrm {m}}\\right) p r o j _ {p} ^ {m} \\left(F _ {p} ^ {m \\prime}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 595, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {p} ^ {m} = \\mathrm {F F N} _ {p} ^ {m} (\\mathrm {s o f t m a x} (\\mathrm {A t t n} _ {\\mathrm {j o i n t}} ^ {\\mathrm {m} _ {T}}) p r o j _ {v} ^ {m} (F _ {v} ^ {m /}))\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 619, + 859, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{v}^{m}$ and $F_{p}^{m}$ are the final output features at layer $m$ of the encoder, $\\mathrm{FFN}_v^m$ and $\\mathrm{FFN}_v^m$ are the visual and textual Feed Forward Networks (FFN) and $proj_0^m$ and $proj_0^m$ are the linear layers to project the visual and textual features. The final encoded features at layer $m = M$ are then utilized to initialize cross-modal queries.", + "bbox": [ + 496, + 646, + 893, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Language-Guided Query Selection", + "text_level": 1, + "bbox": [ + 500, + 756, + 795, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This module is designed to select features more relevant to the input text as decoder queries for effective language-vision fusion. We combine a DETR/DINO [1, 31] style queries with a sinusoidal temporal positional encoding to the positional part of the queries. The sinusoidal positional encoding added to the positional part of the queries adds important contextual information regarding the sequence of frames, allowing for improved temporal correlation and", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18912", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "grounding [28]. The query selection module takes the encoder's visual and textual features as input and outputs num_query indices that correspond to the most relevant features for object detection per frame, $\\{Q_t^0\\}_{t=1}^T$ , where $Q_t^0$ are the initial queries for the frame $t$ . The module initializes the decoder queries using a combination of the selected indices and dynamic anchor boxes. The content part of the queries is set to be learnable during training, while the positional part is computed using the dynamic anchor boxes initialized using the encoder outputs. We also add a sinusoidal temporal positional encoding to the positional part of the queries.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Cross-Modality Spatio-Temporal Decoder", + "text_level": 1, + "bbox": [ + 76, + 292, + 421, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To decode the above queries into bounding box locations and temporal start/end tubes, we need to transform them into an output embedding, which can then be fed into prediction heads. The decoder allows for the queries to interact globally with others within a frame and across frames while utilizing the entire visual and textual features as context. Formally, the queries produced earlier are fed into a $N$ layer decoder. Each layer starts with a temporal self-attention, a spatial self-attention followed by a visual cross-attention and textual cross-attention, and finally an FFN. This is represented in Eq. 4.", + "bbox": [ + 75, + 316, + 470, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t} ^ {n ^ {\\prime}} = \\mathrm {M H S A} _ {\\text {s p a t i a l}} ^ {n} (\\mathrm {M H S A} _ {\\text {t e m p o r a l}} ^ {n} \\left(Q _ {t} ^ {n - 1}\\right)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 493, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t} ^ {n} = \\mathrm {F F N} ^ {n} (\\mathrm {C A} _ {p} ^ {n} (\\mathrm {C A} _ {v} ^ {n} (Q _ {t} ^ {n \\prime}, F _ {v} ^ {M}), F _ {p} ^ {M})),\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 513, + 413, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q_{t}^{n - 1}$ are the input queries at layer $n\\in \\{1,\\dots,N\\}$ , $Q_{t}^{n^{\\prime}}$ are the intermediate queries after spatial and temporal attention at layer $n$ , $Q_{t}^{n^{\\prime}}$ are the output queries at layer $n$ , and $\\mathrm{CA}_v^n$ and $\\mathrm{CA}_p^n$ are the visual and textual cross-attention at layer $n$ . The cross-attentions are further elaborated in Eq. 5.", + "bbox": [ + 76, + 545, + 468, + 637 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C A} _ {v} ^ {n} \\left(Q _ {t} ^ {n \\prime}, F _ {v} ^ {M}\\right) = \\left(\\frac {\\operatorname {p r o j} _ {q , v} ^ {n} \\left(Q _ {t} ^ {n \\prime}\\right) \\operatorname {p r o j} _ {k , v} ^ {n} \\left(F _ {v} ^ {M}\\right) ^ {T}}{\\sqrt {d ^ {k}}} \\operatorname {p r o j} _ {v} ^ {n} \\left(F _ {v} ^ {M}\\right) ^ {T}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 646, + 496, + 684 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C A} _ {p} ^ {n} \\left(\\mathrm {C A} _ {v} ^ {n}, F _ {p} ^ {M}\\right) = \\left(\\frac {\\operatorname {p r o j} _ {q , p} ^ {n} \\left(\\mathrm {C A} _ {v} ^ {n}\\right) \\operatorname {p r o j} _ {k , p} ^ {n} \\left(F _ {p} ^ {M}\\right) ^ {T}}{\\sqrt {d ^ {k}}} \\operatorname {p r o j} _ {p} ^ {n} \\left(F _ {p} ^ {M}\\right) ^ {T}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 686, + 496, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $proj_{q,v}^n, proj_{k,v}^n$ and $proj_v^n$ are the visual query, key and values projection for layer $n$ and $proj_{q,p}^n, proj_{k,p}^n$ and $proj_p^n$ are the textual query, key and value projections. The final queries from the decoder at layer $N$ , $\\{Q_t^N\\}_{t=1}^T$ , are then used for prediction.", + "bbox": [ + 76, + 734, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.4 Prediction Heads", + "text_level": 1, + "bbox": [ + 76, + 830, + 250, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decoder outputs refined queries per frame $\\{Q_t^N\\}_{t = 1}^T$ We follow the standard DETR-like bounding box regression head implemented as a Multi-Layer Perceptron (MLP),", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which predicts bounding boxes $B_{i}^{t} = (x_{i}^{t},y_{i}^{t},w_{i}^{t},h_{i}^{t})$ , per frame. To predict the temporal interval $(t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}$ , we add a temporal grounding head, implemented as an MLP, alongside the bounding box regression head, similar to existing works like [9, 28]. The new head predicts the probabilities of the start $\\tau_{s}\\in [0,1]^{T}$ and ends $\\tau_{e}\\in [0,1]^{T}$ of the interval. During inference, the start and end interval $(t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}$ is computed by taking the maximum of the joint distribution of $(\\tau_s,\\tau_e)$ . Any invalid combinations with $t_e\\leq t_s$ are masked out.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 500, + 253, + 643, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To leverage the generalized pre-trained representation from spatial-grounding foundation models, we initialize all spatial modules and cross attentions from the Grounding DINO [12] spatial grounding model. To preserve this generalized representation while ensuring effective modeling of the downstream task, we freeze the Vision and Text Encoders $\\theta_v$ and $\\theta_p$ and fine-tune the remaining components.", + "bbox": [ + 496, + 277, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the model receives a batch of videos $V$ with text prompt $P$ . The ground-truth annotation contains the bounding box sequence $\\{B_i^t\\}_{i = t_s}^{t_e}$ , and the corresponding start and end timestamps $(t_{s},t_{e})$ . For spatial grounding, we follow the standard loss formulation used in DETR-like [1, 31, 35], namely the $L_{1}$ loss, $\\mathcal{L}_{L_1}$ , and the Generalized Intersection over Union (GIoU) [17] loss, $\\mathcal{L}_{GIoU}$ . Formally, the spatial grounding loss, $\\mathcal{L}_{spatial}$ is defined in Eq. 6.", + "bbox": [ + 496, + 383, + 890, + 520 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\lambda_ {L _ {1}} \\mathcal {L} _ {L _ {1}} (\\hat {B}, B) + \\lambda_ {G I o U} \\mathcal {L} _ {G I o U} (\\hat {B}, B). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 532, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For temporal grounding, we follow [18, 21, 28] and generate two 1-dimensional gaussian heatmaps $\\pi_s,\\pi_e\\in \\mathcal{R}^T$ for the starting and ending positions. The temporal grounding loss is therefore defined in Eq. 7 as,", + "bbox": [ + 496, + 564, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p o r a l}} = \\mathcal {L} _ {K L} ^ {s} \\left(\\hat {\\pi} _ {s}, \\pi_ {s}\\right) + \\mathcal {L} _ {K L} ^ {e} \\left(\\hat {\\pi} _ {e}, \\pi_ {e}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 638, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{KL}^s$ and $\\mathcal{L}_{KL}^{e}$ are the KL divergence losses for the start and end distributions, respectively.", + "bbox": [ + 496, + 667, + 888, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the model outputs the bounding boxes and starting/ending distributions during inference. We determine the temporal grounding segment, $(t_s,t_e)$ , by taking the segment with the maximal joint start and end probability. Then, we consider the bounding boxes only within that tube for spatial grounding.", + "bbox": [ + 496, + 699, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 806, + 586, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup and Protocols", + "text_level": 1, + "bbox": [ + 500, + 830, + 797, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Below, we first briefly explain the implementation details (Sec. 4.1.1), followed by evaluation settings (Sec. 4.1.2), and datasets (Sec. 4.1.3) used in our work.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dbf4084e7df5bdaf3f1fd5e8e496643b73381c5c76fc149bbd2131079fe88883.jpg", + "table_caption": [ + "Table 1. Performance comparisons of the state-of-the-art on HC-STVG V1 [24] and YouCook-Interactions [23] in open-vocabulary setting." + ], + "table_footnote": [], + "table_body": "
MethodPre-trainingHC-STVG V1YouCook-Interactions
m_vIoUvIoU@0.3vIoU@0.5Accuracy
TubeDETR (CVPR'22) [28]VidSTG16.8422.329.2251.63
STCAT (NeurIPS'22) [9]VidSTG22.5832.1420.8355.90
VideoGrounding-DINOVidSTG27.4640.1329.9257.73
", + "bbox": [ + 122, + 114, + 839, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 227, + 294, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed in the methodology (Sec. 3), we initialize the spatial modules in our model from the Grounding DINO [12] spatial grounding model and keep the vision and text encoders frozen. Our prediction heads for both spatial and temporal predictions are set to be 3-layer Multi-Layer Perceptrons (MLPs). We sample 128 frames during training and inference, resized to a resolution of 448 on the shorter side. We set both $M$ and $N$ to 6, and train the model with a batch size of 8 and learning rate of $1e^{-4}$ , and weight decay if $10^{-4}$ . The number of epochs for VidSTG is set to 10, and for HC-STVG V1/V2 is set to 90.", + "bbox": [ + 75, + 251, + 470, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Evaluation Settings", + "text_level": 1, + "bbox": [ + 76, + 438, + 266, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our video grounding model in two settings, Open-Vocabulary and Closed-Set Supervised.", + "bbox": [ + 76, + 462, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open-Vocabulary Evaluation: In the open-vocabulary setting we train our model on the VidSTG [32] dataset and then evaluate on two different datasets, HC-STVG V1 [24] and YouCook-Interactions [23] to understand how well the model generalizes to new distributions. The reason for choosing these two datasets is that the former provides a relatively minor distribution shift given the similar perspective/objects in the videos compared to the training dataset VidSTG. In contrast, the latter provides a major distribution shift with changes in perspective and annotated objects/interactions.", + "bbox": [ + 75, + 493, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Closed-Set Supervised Evaluation: In the supervised evaluation setting, we train on the training set and evaluate each dataset's respective validation/testing set. This evaluation is conducted for three majorly used datasets in spatio-temporal video grounding, namely VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24].", + "bbox": [ + 75, + 659, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3 Datasets", + "text_level": 1, + "bbox": [ + 76, + 770, + 191, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our approach and compare against the state-of-the-art in two settings: Open-Vocabulary and Closed-Set Supervised, across a total of four grounding datasets, namely: VidSTG [32], HCSTVG V1 [24], HCSTVG V2 [24], and YouCook-Interactions [23].", + "bbox": [ + 75, + 794, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VidSTG: The VidSTG [32] dataset is derived from the VidOR [20] dataset, incorporating object relation annotations.", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It includes 99,943 video-text pairs, encompassing 44,808 declarative sentence queries and 55,135 interrogative sentence queries. The training, validation, and test sets consist of 80,684, 8,956, and 10,303 sentences and 5,436, 602, and 732 videos, respectively. VidSTG's text queries are confined to describing pre-defined object/relation categories in VidOR [20].", + "bbox": [ + 496, + 227, + 890, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HC-STVG V1/V2: The HC-STVG datasets are sourced from movie scenes, each video clip spanning approximately 20 seconds. These datasets pose challenges in spatiotemporal grounding due to video clips featuring multiple individuals engaged in similar actions. HC-STVG V1 comprises 4,500 training and 1,160 testing video-text pairs. HC-STVG V2 expands HC-STVG V1, enhancing annotation quality with 10,131, 2,000, and 4,413 samples for training, validation, and testing, respectively. As HC-STVG V2's test set annotations are unavailable publicly, results are reported on the validation set.", + "bbox": [ + 496, + 333, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YouCook-Interactions: The YouCook-Interactions [23] dataset serves as an expansion of the YouCook2 [34] dataset focused on cooking instructions. This extension includes bounding boxes for 6,000 carefully chosen frames, typically encompassing the hand and the tool specified in the corresponding sentence-level annotations. Our assessment revolves around examining models' spatial grounding capabilities using this dataset.", + "bbox": [ + 496, + 500, + 890, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results and Analysis", + "text_level": 1, + "bbox": [ + 500, + 632, + 803, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present our results across the evaluation mentioned above settings (Sec. 4.1.2) and datasets (Sec. 4.1.3). We start with the closed-set evaluation in Sec. 4.2.2, followed by the open-vocabulary evaluation in Sec. 4.2.1.", + "bbox": [ + 496, + 657, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Open-Vocabulary Evaluation", + "text_level": 1, + "bbox": [ + 500, + 753, + 756, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For open-vocabulary evaluation, we train on VidSTG [32] and present results HC-STVG V1 [24] and YouCook-Interactions [23]. The results are reported jointly in Tab. 1. Results on HC-STVG V1: We report open-vocabulary evaluation on m_vIoU, vIoU@0.3, and vIoU@0.5. We achieve state-of-the-art performance over both TubeDETR [28] and STCAT [9]. We attribute this strong performance to our design, which leverages the strong pre", + "bbox": [ + 496, + 777, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18914", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d4fcfa28fb75831d518a55b06c2bb94d20ee1f9ba269e9f844bd18c8295dd4d.jpg", + "table_caption": [ + "Table 2. Performance comparisons of the state-of-the-art on the VidSTG [32] test set in closed-set supervised setting." + ], + "table_footnote": [], + "table_body": "
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3
Factorized:
GroundRe (ECCV'16) [19]+TALL (ICCV'17) [7]9.7811.044.099.3211.39
STPR (ICCV'17) [27]+TALL (ICCV'17) [7]34.6310.4012.384.2733.739.9811.74
WSSTG (arXiv'19) [5]+TALL (ICCV'17) [7]11.3614.635.9110.6513.90
GroundRe (ECCV'16) [19]+L-Net (AAAI'19) [3]11.8915.325.4511.0514.28
STPR (ICCV'17) [27]+L-Net (AAAI'19) [3]40.8612.9316.275.6839.7911.9414.73
WSSTG (arXiv'19) [5]+L-Net (AAAI'19) [3]14.4518.007.8913.3617.39
Two-Stage:
STGRN (CVPR'20) [32]48.4719.7525.7714.6046.9818.3221.10
STGVT (TCSVT'21) [24]-21.6229.8018.94---
OMRN (IJCAI'21) [33]50.7323.1132.6116.4249.1920.6328.35
One-Stage:
STVGBert (ICCV'21) [21]-23.9730.9118.39-22.5125.97
TubeDETR (CVPR'22) [28]48.1030.4042.5028.2046.9025.7035.70
STCAT (NeurIPS'22) [9]50.8233.1446.2032.5849.6728.2239.24
STVGFormer (CVPR'23) [11]-33.7047.2032.80-28.5039.90
VideoGrounding-DINO51.9734.6748.1133.9650.8329.8941.03
", + "bbox": [ + 80, + 114, + 893, + 398 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e3235c6b2633c1a8d84e537e19b91a7659d5c66cc8e3114fb5ce31132069fcd6.jpg", + "table_caption": [ + "Table 3. Performance comparisons of the state-of-the-art on the HC-STVG V1 [24] test set in closed-set supervised setting." + ], + "table_footnote": [], + "table_body": "
Methodsm_vIoUvIoU@0.3vIoU@0.5
STGVT (TCSVT'21) [24]18.1526.819.48
STVGBert (ICCV'21) [21]20.4229.3711.31
TubeDETR (CVPR'22) [28]32.4049.8023.50
STCAT (NeurIPS'22) [9]35.0957.6730.09
STVGFormer (CVPR'23) [11]36.9062.2034.80
VideoGrounding-DINO38.2562.4736.14
", + "bbox": [ + 81, + 450, + 468, + 558 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c4578b1a417febd51dd28f370e492e29e5eab770f470bb2d8647979b8ce95c1.jpg", + "table_caption": [ + "Table 4. Performance comparisons of the state-of-the-art on the HC-STVG V2 [24] val set in closed-set supervised setting." + ], + "table_footnote": [], + "table_body": "
Methodsm_vIoUvIoU@0.3vIoU@0.5
Yu et al (arXiv'21) [30]30.00--
Aug. 2D-TAN (arXiv'21) [22]30.4050.4018.80
TubeDETR (CVPR'22) [28]36.4058.8030.60
STVGFormer (CVPR'23) [11]38.7065.5033.80
VideoGrounding-DINO39.8867.1334.49
", + "bbox": [ + 81, + 627, + 468, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained generalized features of a spatial grounding foundation model.", + "bbox": [ + 76, + 753, + 468, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on YouCook-Interactions: We evaluate our method further on the YouCook-Interactions [23] dataset, reporting pointing game accuracy for spatial grounding. Our approach gains nearly $2\\%$ in accuracy over STCAT [9] and more than $6\\%$ compared to TubeDETR [28]. This further shows our strong generalization capabilities in the open-vocabulary setting.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Closed-Set Supervised Evaluation", + "text_level": 1, + "bbox": [ + 500, + 414, + 790, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present closed-set evaluations across three datasets, VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24].", + "bbox": [ + 498, + 459, + 890, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on VidSTG: We present results on the VidSTG test set in closed-set setting in Tab. 2, reporting m_tIoU, m_vIoU, vIoU@0.3 and vIoU@0.5. The results show that our method achieves state-of-the-art performance in comparison to both Two-Stage and One-Stage methods. In particular, we achieve more than $1t\\_ IoU$ gain in temporal grounding over the previous best methods OMRN [33] (One-Stage) and STVGFormer [11] (Two-Stage), both for Declarative and Interrogative sentences. Similarly, for m_vIoU, vIoU@0.3 and vIoU@0.5, we achieve a more than 1 unit gain the state-of-the-art methods STVGFormer [11] and STCAT [9]. Note that our method uses a frozen visual and textual encoder. In contrast, those mentioned above previous state-of-the-art methods all train the entire encoder.", + "bbox": [ + 496, + 500, + 892, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on HC-STVG V1: We present results on the HC-STVG V1 dataset in Tab. 3, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. We achieve a nearly 1.5 unit gain in m_vIoU and vIoU@0.5 and a 1 unit gain in vIoU@0.3 over the previous best method STVGFormer [11]. This shows the consistent performance of our method on this dataset.", + "bbox": [ + 496, + 722, + 893, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on HC-STVG V2: We present results on the HC-STVG V2 dataset in Tab. 4, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. Our performance gain on HC-STVG V1 is reflected here as well, with a consistent performance gain on HC-STVG V2, in comparison to the SoTA [11, 28].", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18915", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88e9076221c924f667439eca1135ff62dc4cb7507eea0a4cdc543855e9630a6c.jpg", + "table_caption": [ + "Table 5. Ablation on various design choices for our approach on the VidSTG [32] test set in closed-set supervised setting." + ], + "table_footnote": [], + "table_body": "
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3vIoU@0.5
Naive Solution (Frozen Grounding DINO [12])39.7818.0722.3113.7539.799.6610.423.84
+ Decoder Temporal Aggregation42.8120.7426.5315.4143.8112.3816.718.62
+ Encoder Temporal Aggregation46.2923.1932.3818.9547.1716.1923.2813.04
+ Finetuned Spatial Modules in Decoder48.0628.9741.6026.0649.5824.2732.8520.11
+ Finetuned Spatial Modules in Encoder51.9734.6748.1133.9650.8329.8941.0327.58
", + "bbox": [ + 80, + 114, + 893, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca88162c02f563393391dc4d069b87b269c2104263551c82160189ba87dca90a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 239, + 228, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea612f6d83381b82ff1dfdfe080133886293bc62b3eeb0b2eec3c5adaadddfe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 316, + 227, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98d29cc55f91c5c6a4967598f9e91a788603fe8d0f54a9fd47cb6021d88b7e95.jpg", + "image_caption": [ + "Figure 3. Sample visualization for video grounding result on HC-STVG V1 [24] for TubeDETR [28], STCAT [9] and ours with the prompt The man behind the shirtless man turns and squats. We show bounding boxes for Ground-truth, Closed-Set Supervised, and OpenVocabulary results. Note how both TubeDETR and STCAT are close to the ground truth in the supervised setting (STCAT more so than TubeDETR), they cannot correctly ground the text properly in the open-vocabulary setting." + ], + "image_footnote": [], + "bbox": [ + 81, + 395, + 227, + 472 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ac908981847e7812c09c5e3877a061c747288912af8a2947b4a0df40ee61faf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 239, + 359, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73f969c386dda56ab005bd106858ad39533f7999b610795328bd3e11a132ff52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 316, + 359, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f860a59b7490cbddaf0e2d4894768c07c0ee2fd17c2e1481ce2dab2202d0662.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 395, + 359, + 472 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d83187aa9cd505b9f960cf5bcd00bc52ac638a865df585edc410be35cbd59a3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 239, + 493, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fe8f6d61883e0cc4d194384b5982d276019e540308fc337d149fb16f482d04b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 316, + 491, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9860f1dbe704e5b1ba0cdc49a4717a0d09331b17134a7b8c28264896aa2e25fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 395, + 491, + 472 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fb5ca49031dbe6f986f0952e7aada430234de34986f931ee813c92172a74a8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 239, + 622, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f4885f3d88815f1ab7d25e9196f26cf1ade29cd810562e02f1e98e698b30f2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 316, + 622, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6fee9a4934a0c0182e4bdada01195b6599d659e8ee0bd36fa0fc7d75559e84d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 395, + 622, + 472 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1bbb07d32c6746de6dfebc43ac9cfca10403baf304bee4bf1606fc98ad12710.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 239, + 754, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44e36a493c450f89d73962bdc7d207717185f1adbbf3be658576e04e621fa3f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 316, + 754, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c709ad6d1ce8374ee9b7c39816f002ab03819b31d13c3435199b2924bb0fdb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 395, + 754, + 472 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/317b3d8135774a989b8f0df6d5ba5d4480a274ae7a9d162bef6aeea9cc9d03e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 239, + 887, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3bfc08974eaf1c5604943d9037789109a1d471a36d3f0a266b3ced507a0aab2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 316, + 887, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25ef794d542e9e6bbc238641c65b5c2ef42c79dc304df9b0471659d58cf4617e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 395, + 887, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablative Analysis", + "text_level": 1, + "bbox": [ + 76, + 550, + 246, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform an ablative analysis of the various design choices for our model. In particular, we first evaluate a naive baseline where no additional temporal aggregators are added, and all pre-trained spatial modules are frozen in the encoder and decoder. This baseline is relatively weak in both temporal and spatial grounding. We next add temporal modules in first the decoder, and subsequently in the encoder. We find that it gives significant improvements in the temporal grounding, and additionally also improves the spatial grounding. Finally, we finetune the pre-trained spatial modules in both the decoder and the encoder, which provides a strong improvement in spatial grounding, alongside consistent improvement in temporal grounding.", + "bbox": [ + 75, + 574, + 468, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitation", + "text_level": 1, + "bbox": [ + 76, + 785, + 194, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our video grounding model excels in closed-set and open-vocabulary scenarios, it leverages image-text pretrained models like Grounding DINO [12]. To enhance understanding in open-vocabulary settings, an extension to video-language pre-training on a larger and more diverse dataset, akin to CLIP [15], can help further boost perfor", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance. Building a video-language pre-training dataset with diverse natural language expressions and spatio-temporal localization is imperative, given the constraints of datasets like VidSTG [32] and HC-STVG [24].", + "bbox": [ + 496, + 550, + 890, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 642, + 617, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces an Open-Vocabulary Spatio-Temporal Video Grounding task, enhancing current closed-set methodologies by using pre-trained representations from spatial grounding models. The proposed model performs well in closed-set and open-vocabulary scenarios, surpassing state-of-the-art results in supervised setting on VidSTG and HC-STVG datasets, and outperforming recent models in open-vocabulary on HC-STVG V1 and YouCook-Interactions. Its architecture includes learnable adapter blocks for video-specific adaptation, bridging the semantic gap between natural language queries and visual content. This research addresses open-vocabulary challenges and explores achieving robust performance without extensive video annotations, paving the way for open-vocabulary video grounding.", + "bbox": [ + 496, + 674, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18916", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3, 4, 5", + "[2] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where? - self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 3", + "[3] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, , and Jiebo Luo. Localizing natural language in videos. In AAAI, 2019. 7", + "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3", + "[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549, 2019. 7", + "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2019. 3", + "[7] Jiyang Gao, Chen Sun, Zhenheng Yang, , and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 7", + "[8] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 2, 3", + "[9] Yang Jin, Yongzhi Li, Zehuan Yuan, and Yadong Mu. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In NeurIPS, 2022. 1, 2, 3, 5, 6, 7, 8", + "[10] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2, 3, 4", + "[11] Zihang Lin, Chaolei Tan, Jian-Fang Hu, Zhi Jin, Tiancai Ye, and Wei-Shi Zheng. Collaborative static and dynamic vision-language streams for spatio-temporal video grounding. In CVPR, 2023. 1, 2, 3, 7", + "[12] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 5, 6, 8", + "[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3", + "[14] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3", + "[15] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 8", + "[16] Hanoona Rasheeda, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2", + "[17] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union. In CVPR, 2019. 5", + "[18] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, 2020. 5", + "[19] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, , and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In ECCV, 2016. 7", + "[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In ICMR, 2019. 6", + "[21] Rui Su, Qian Yu., and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In ICCV, 2021. 2, 3, 5, 7", + "[22] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 2, 3, 7", + "[23] Reuben Tan, Bryan A. Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos, 2021. 1, 2, 6, 7", + "[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 2, 3, 6, 7, 8", + "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4", + "[26] Zhenzhi Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2, 3", + "[27] Masataka Yamaguchi, Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Spatio-temporal person retrieval via natural language queries. In ICCV, 2017. 7", + "[28] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18917", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 3", + "[30] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 3, 7", + "[31] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5", + "[32] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In CVPR, 2020. 1, 2, 3, 6, 7, 8", + "[33] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In *IJCAI*, 2021. 2, 3, 7", + "[34] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 6", + "[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 4, 5" + ], + "bbox": [ + 78, + 90, + 470, + 446 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18918", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_model.json b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3493120079c81e67372a59841efeb7db02e88184 --- /dev/null +++ b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "VideoGrounding-DINO: Towards Open-Vocabulary Spatio-Temporal Video Grounding" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.203, + 0.744, + 0.241 + ], + "angle": 0, + "content": "Syed Talal Wasim1 Muzammal Naseer1 \nSalman Khan1,2 Ming-Hsuan Yang3,4 Fahad Shahbaz Khan1,5" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.245, + 0.792, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\)Mohamed bin Zayed University of AI \\(^{2}\\)Australian National University \\(^{3}\\)University of California, Merced \\(^{4}\\)Google Research \\(^{5}\\)Linköping University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.317, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.349, + 0.474, + 0.728 + ], + "angle": 0, + "content": "Video grounding aims to localize a spatio-temporal section in a video corresponding to an input text query. This paper addresses a critical limitation in current video grounding methodologies by introducing an Open-Vocabulary Spatio-Temporal Video Grounding task. Unlike prevalent closed-set approaches that struggle with open-vocabulary scenarios due to limited training data and predefined vocabularies, our model leverages pre-trained representations from foundational spatial grounding models. This empowers it to effectively bridge the semantic gap between natural language and diverse visual content, achieving strong performance in closed-set and open-vocabulary settings. Our contributions include a novel spatio-temporal video grounding model, surpassing state-of-the-art results in closed-set evaluations on multiple datasets and demonstrating superior performance in open-vocabulary scenarios. Notably, the proposed model outperforms state-of-the-art methods in closed-set settings on VidSTG (Declarative and Interrogative) and HC-STVG (V1 and V2) datasets. Furthermore, in open-vocabulary evaluations on HC-STVG V1 and YouCook-Interactions, our model surpasses the recent best-performing models by \\(4.88\\mathrm{m\\_vIoU}\\) and \\(1.83\\%\\) accuracy, demonstrating its efficacy in handling diverse linguistic and visual concepts for improved video understanding. Our codes will be publicly released." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Spatio-temporal video grounding is pivotal in linking visual content with natural language descriptions, thus facilitating semantics interpretation within visual data. Prevailing approaches in video grounding such as TubeDETR [28], STCAT [9], and STVGFormer [11] focus mainly on supervised closed-set settings, where models are trained on specific datasets [24, 32] with predefined vocabulary and meticulously annotated data. While these current state-of" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.32, + 0.885, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.554, + 0.894, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Performance comparison on conventional closed-set and open-vocabulary settings for the video grounding task. We compare our approach with TubeDETR [28] and STCAT [9] in supervised setting for VidSTG [32] declarative/interrogative and HC-STVG V1 [24], along with open-vocabulary evaluation on HC-STVG V1 and YouCook-Interactions [23] datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.893, + 0.763 + ], + "angle": 0, + "content": "the-art methods excel in closed-set settings on datasets like VidSTG [32] and HC-STVG [24], their limited generalization beyond training dataset distributions poses a significant challenge. The relatively small scale and restricted sample variety in existing video datasets hinder models from adapting to unseen scenarios effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Motivated by the inherent limitation of supervised closed-set approaches in terms of their restricted vocabulary, this paper investigates Open-Vocabulary Spatio-Temporal Video Grounding. Unlike conventional methodologies, this paradigm addresses challenges posed by the unrestricted diversity of language and visual concepts in the wild. The goal is to train on a set of base categories and generalize to unseen objects/actions based on the open-vocabulary nature of backbone models. This paper ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "plores the challenges and opportunities inherent in open-vocabulary video grounding, laying the groundwork for more robust and versatile video understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.142, + 0.471, + 0.385 + ], + "angle": 0, + "content": "However, training an effective open-vocabulary video grounding model would require a large enough dataset with a rich set of natural language expressions and corresponding spatio-temporal localizations. Such an extensive dataset can allow the model to learn generalized visual and textual representations and handle out-of-distribution samples. However, video grounding datasets [24, 32] are quite limited in scale, e.g., VidSTG has only \\(5.4\\mathrm{k}\\) training videos with \\(80.6\\mathrm{k}\\) distinct sentences. In contrast, an image grounding model GLIP [10] is trained with \\(\\sim 26.5\\mathrm{M}\\) image-text pairs. Therefore, our research explores the following fundamental question: How spatio-temporal video grounding models can achieve robust performance in both Closed-Set and Open-Vocabulary scenarios without requiring large-scale video annotations, ensuring effective generalization beyond training datasets?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.471, + 0.648 + ], + "angle": 0, + "content": "In addressing this question, we find inspiration in the accomplishments of foundational models specializing in spatial grounding [4, 8, 12, 14, 29]. These models undergo training on an extensive corpus of image-text data, enabling effective generalization to samples from a given target distribution. We aim to harness this pretrained representation to enhance our video grounding model. Our proposed solution is a spatio-temporal video grounding model adopting a DETR-like architecture enhanced by temporal aggregation modules. The spatial modules are initialized using the pretrained representation of a foundational image model [12]. Meanwhile, the image and text feature extractors remain frozen while the video-specific spatio-temporal adaptations are modeled via learnable adapter blocks. This approach is designed to preserve the nuanced representation of the foundational model, enhancing our model's ability to generalize effectively to novel samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.652, + 0.47, + 0.713 + ], + "angle": 0, + "content": "A summary of our closed-set and open-vocabulary results is shown in Fig. 1, where the proposed approach excels in both settings by a clear margin. Our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.718, + 0.47, + 0.809 + ], + "angle": 0, + "content": "- For the first time, we evaluate spatio-temporal video grounding models in an open-vocabulary setting on HC-STVG V1 [24] and YouCook-Interactions [23] benchmarks in a zero-shot manner. We outperform state-of-the-art methods TubeDETR [28] and STCAT [9] by 4.26 m_vIoU and \\(1.83\\%\\) accuracy, respectively." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.809, + 0.471, + 0.9 + ], + "angle": 0, + "content": "- By combining the strengths of spatial grounding models with complementary video-specific adapters, our approach consistently outperforms the previous state-of-the-art in closed-set setting on four benchmarks, i.e., VidSTG (Declarative) [32], VidSTG (Interrogative) [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.718, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Spatial Grounding Foundation Models: Recent literature introduces notable spatial grounding models. GLIP [10] unifies object detection and phrase grounding through a language-image pre-training model, leveraging extensive image-text pairs for semantic-rich representations. Grounding DINO [12] integrates language with a transformer-based detector to achieve an open-set detector, excelling in benchmarks like COCO and ODinW. Kosmos-1 [8] and Kosmos-2 [14] contribute Multimodal Large Language Models (MLLMs) with capabilities such as zero-shot and few-shot learning, language understanding, and multimodal tasks. Kosmos-2 [14] specifically integrates grounding into downstream applications, introducing GrIT, a large-scale dataset of Grounded Image-Text pairs. Shikra [4] addresses referential ability in MLLMs by handling spatial coordinates in inputs and outputs, showcasing promising performance in various vision-language tasks. Ferret [29] unifies referring and grounding in the LLM paradigm, achieving superior performance in classical referring and grounding tasks, excelling in region-based multimodal chatting and image description. Recently, GLaMM [16] allows pixel-level grounded conversations with an LLM, showcasing generalizability to several captioning and referring segmentation tasks. However, spatial methods cannot work for grounding objects in videos, a gap addressed by this work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Spatio-Temporal Video Grounding: Several methods tackle the challenge of localizing objects in untrimmed videos based on query sentences. STVGBert [21] presents a one-stage visual-linguistic transformer for simultaneous spatial and temporal localization. TubeDETR [28] introduces a transformer-based architecture to model temporal, spatial, and multi-modal interactions efficiently. Augmented 2D-TAN [22] adopts a two-stage approach, enhancing the 2D-TAN with a temporal context-aware Bi-LSTM Aggregation Module. OMRN [33] addresses the challenge of unaligned data and multi-form sentences in spatiotemporal video grounding, proposing an object-aware multi-branch relation network for effective relation discovery. MMN [26] introduces a Mutual Matching Network as a metric-learning framework for temporal grounding, achieving competitive performance. STCAT [9] is an end-to-end one-stage framework addressing feature alignment and prediction inconsistency. Finally, STVGFformer [11] proposes an effective framework with static and dynamic branches for cross-modal understanding. While the above methods advance video grounding, their generalization to out-of-distribution and open-vocabulary samples is limited due to constrained video datasets [24, 32]. To address this issue, we utilize the generalized representation of spatial grounding foundation models [4, 8, 10, 12, 14, 29] trained on a large corpus of image-text data and can perform well on both closed-set and open-vocabulary evaluations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18910" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.212, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.468, + 0.252 + ], + "angle": 0, + "content": "As discussed above, the current state-of-the-art spatiotemporal video grounding methods [9, 11, 21, 22, 24, 26, 28, 30, 32, 33] primarily evaluate in a supervised setting on the VidSTG [32] and HC-STVG [24] datasets. However, these methods lack the multimodal spatio-temporal understanding required to perform well on out-of-distribution samples [2]. Therefore, this work aims to achieve improved open-vocabulary performance while maintaining strong closed-set video-grounding performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.468, + 0.401 + ], + "angle": 0, + "content": "We take inspiration from recent foundation models for spatial grounding [4, 8, 10, 12, 14, 29]. These models are trained on a large corpus of visual-textual data and hence, generalize well to unseen samples. We aim to leverage the strong generalization capabilities of such foundation models to achieve strong open-set spatio-temporal video grounding performance. Our proposed spatio-temporal video grounding method uses DETR-like [1] design, with temporal aggregation and adaptation modules for learning video-specific representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.468, + 0.477 + ], + "angle": 0, + "content": "Below, we explain our proposed methodology. We formally define the spatio-temporal video grounding problem in Sec. 3.1. We then explain our architecture details in Sec. 3.2. We finally explain the loss formulation used to train the model and model initialization in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.26, + 0.5 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.468, + 0.643 + ], + "angle": 0, + "content": "The spatio-temporal video grounding task involves localizing and recognizing objects and actions in a video sequence by integrating spatial and temporal information. In contrast to spatial grounding, which focuses on recognizing and localizing objects or actions within individual frames, spatiotemporal grounding extends this concept to include the temporal dimension. This means understanding where objects or actions are in each frame and how they evolve and move over time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Consider a video \\(V \\in \\mathbb{R}^{T \\times H \\times W \\times C}\\) with \\(T\\) frames, \\(H \\times W\\) spatial resolution, and \\(C\\) channels, respectively, along with a text prompt \\(P\\). The spatial grounding problem can be defined as the localization of one or more objects associated with the prompt \\(P\\) in a frame \\(V_{t}, t \\in \\{1, \\dots, T\\}\\) using a bounding box \\(B_{i}^{t} = (x_{i}^{t}, y_{i}^{t}, w_{i}^{t}, h_{i}^{t})\\), where \\(x_{i}^{t}\\) and \\(y_{i}^{t}\\) are the coordinates of the top-left corner, \\(w_{i}^{t}\\) and \\(h_{i}^{t}\\) are the width and height of the bounding box, \\(i \\in \\{1, \\dots, N\\}\\) is the object number for the frame \\(t\\). The temporal grounding problem, on the other hand, involves understanding how objects or actions evolve over time. It aims at localizing the temporal interval \\((t_{s}, t_{e})\\) where the specific action/interaction happens in the entire temporal duration, where interval \\((t_{s}, t_{e})\\) indicates the start and end frame of the object occurrence within the total frames \\(T\\) (\\(1 \\leq t_{s} < t_{e} \\leq T\\)). Hence, the spatio-temporal grounding problem for object \\(i\\) associated with prompt \\(P\\) can be summarized as a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "set of spatio-temporal coordinates associated with the subset of frames where the object exists: \\((x_i^t,y_i^t,w_i^t,h_i^t,t)\\) and \\(t\\in \\{t_s,\\dots,t_e\\}\\). The interval \\((t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}\\) and is a subset of the total frames \\(T\\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.165, + 0.804, + 0.181 + ], + "angle": 0, + "content": "3.2. Spatio-Temporal Video Grounding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Here, we explain our video grounding model in Fig. 2. As discussed earlier, we aim to design a spatio-temporal video grounding model that can perform well in closed-set and open-vocabulary settings. Strong open-vocabulary performance requires learning a rich visual/textual representation, which in turn requires a large amount of training data. Unfortunately, spatio-temporal video grounding datasets are quite limited in scale [24, 32], resulting in current video grounding methods failing in generalizing well to out-of-distribution samples because they lack the requisite strong visual/textual representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.763 + ], + "angle": 0, + "content": "To solve this problem, our approach takes inspiration from recent spatial grounding methods [4, 8, 12, 14, 29], which have strong open-vocabulary performance thanks to the large image-text corpus they are trained on. We can utilize the generalized representations of these models to enrich the weaker representation of video-grounding approaches obtained from the limited number of training samples. Our approach aims to leverage the strong pretrained representations of spatial grounding methods to achieve strong closed-set supervised and open-vocabulary video grounding performance. Our spatio-temporal video grounding approach is based on the state-of-the-art DETR-based [1] object detection framework DINO [31] and also borrows concepts of image-text alignment and grounding from Grounded Language-Image Pre-training (GLIP) [10] and Grounding DINO [12]. We extract initial features from backbone vision and text encoders \\(\\theta_v\\) and \\(\\theta_p\\). Following that, we model inter-frame and intra-frame features and learn cross-modal visual/textual relations in the Cross-Modality Spatio-Temporal Encoder (Sec. 3.2.1). The result enriched cross-modal features are used to initialize queries for each frame (Sec. 3.2.2). These queries are then decoded to predict the bounding boxes per frame and the temporal grounding start/end frame by aggregating information across spatial/temporal dimensions and injecting information through cross-attention from enriched visual/textual context (Sec. 3.2.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given the video \\(V\\) and text prompt \\(P\\) as described above, we obtain per-frame features \\(F_{v}^{0}\\) and text features \\(F_{p}^{0}\\) from vision and text encoders, \\(\\theta_v\\) and \\(\\theta_p\\), respectively. The vision encoder is based on a Swin Transformer [13], and the text-encoder is defined as a BERT [6] model. Like other DETR-based detectors [31, 35], image features are extracted from different vision-encoder blocks at multiple scales. The features are then passed to the Cross-Modality Spatio-Temporal Encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18911" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.895, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Overall architecture: We present our video grounding architecture. It consists of vision and text encoders that produce visual and textual features. A cross-modality spatio-temporal encoder which fuses information across spatial/temporal dimensions and visual/textual modalities. A language guided query selection module to initialize cross-modal queries. A cross-modality spatio-temporal decoder to decoder queries while fusing information from visual/textual features. And finally two prediction heads to predict the bounding boxes per frame and the temporal tube. Modules with \\((\\bullet)\\) are trainable and those with \\((\\text{串})\\) are frozen." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.424, + 0.466 + ], + "angle": 0, + "content": "3.2.1 Cross-Modality Spatio-Temporal Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.563 + ], + "angle": 0, + "content": "The initial vision and text features \\( F_{v}^{0} \\) and \\( F_{p}^{0} \\) neither contain any cross-modal information nor model the temporal relationship across frames. Therefore, we further encode the initial features through the Cross-Modality Spatio-Temporal Encoder to model the temporal information across frames and learn cross-modal features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.673 + ], + "angle": 0, + "content": "Each layer of the \\(M\\) layer encoder first applies a Multi-Head Self-Attention (MHSA) [25] to the visual features \\(F_{v}\\) along the temporal dimension, followed by a Deformable Attention (DA) [35] along the spatial dimension. This is done to model relations within frames and temporally across frames. Similarly, we apply a MHSA on the text features \\(F_{p}\\). This is illustrated in Eq. 1." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.683, + 0.47, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {v} ^ {m \\prime} = \\mathrm {D A} _ {s p a t i a l} ^ {m} (\\mathrm {M H S A} _ {t e m p o r a l} ^ {m} (F _ {v} ^ {m - 1})), \\\\ F _ {p} ^ {m \\prime} = \\mathrm {M H S A} _ {p} ^ {m} (F _ {p} ^ {m - 1}), \\end{array} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.856 + ], + "angle": 0, + "content": "where \\(F_{v}^{m - 1}\\) and \\(F_{p}^{m - 1}\\) are visual and textual input features to layer \\(m\\), \\(F_{v}^{m'}\\) and \\(F_{p}^{m'}\\) are the intermediate visual and textual feature representation, and \\(\\mathrm{DA}_{\\text{spatial}}^m\\), \\(\\mathrm{MHSA}_{\\text{temporal}}^m\\) and \\(\\mathrm{MHSA}_p^m\\) are the spatial-deformable, temporal and textual attentions at layer \\(m \\in \\{1,\\dots,M\\}\\), respectively. Following initial spatial, temporal, and textual attentions, we fuse features across the visual and textual modalities, as done in GLIP [10]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.904 + ], + "angle": 0, + "content": "More specifically, we calculate the joint visual-textual attention, \\(\\mathrm{Attn}_{\\mathrm{joint}}^{\\mathrm{m}}\\), using projected intermediate features \\(F_{v}^{m'}\\) and \\(F_{p}^{m'}\\). This attention is then used alongside the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.893, + 0.495 + ], + "angle": 0, + "content": "intermediate features \\( F_{v}^{m'} \\) and \\( F_{p}^{m'} \\), to calculate the image-to-text and text-to-image cross-attentions as shown in Eq. 2 and Eq. 3." + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.506, + 0.892, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A t t n} _ {\\mathrm {j o i n t}} ^ {\\mathrm {m}} = \\left(\\frac {\\operatorname {p r o j} _ {q , v} ^ {m} \\left(F _ {v} ^ {m \\prime}\\right) \\operatorname {p r o j} _ {q , p} ^ {m} \\left(F _ {p} ^ {m \\prime}\\right) ^ {T}}{\\sqrt {d ^ {k}}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.587 + ], + "angle": 0, + "content": "where \\( proj_{q,v}^{m} \\) and \\( proj_{q,p}^{m} \\) are the query projections for the visual and textual features, respectively, at layer \\( m \\)." + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.597, + 0.892, + 0.624 + ], + "angle": 0, + "content": "\\[\nF _ {v} ^ {m} = \\operatorname {F F N} _ {v} ^ {m} \\left(\\operatorname {s o f t m a x} \\left(\\operatorname {A t t n} _ {\\text {j o i n t}} ^ {\\mathrm {m}}\\right) p r o j _ {p} ^ {m} \\left(F _ {p} ^ {m \\prime}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.621, + 0.86, + 0.639 + ], + "angle": 0, + "content": "\\[\nF _ {p} ^ {m} = \\mathrm {F F N} _ {p} ^ {m} (\\mathrm {s o f t m a x} (\\mathrm {A t t n} _ {\\mathrm {j o i n t}} ^ {\\mathrm {m} _ {T}}) p r o j _ {v} ^ {m} (F _ {v} ^ {m /}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.894, + 0.74 + ], + "angle": 0, + "content": "where \\( F_{v}^{m} \\) and \\( F_{p}^{m} \\) are the final output features at layer \\( m \\) of the encoder, \\( \\mathrm{FFN}_v^m \\) and \\( \\mathrm{FFN}_v^m \\) are the visual and textual Feed Forward Networks (FFN) and \\( proj_0^m \\) and \\( proj_0^m \\) are the linear layers to project the visual and textual features. The final encoded features at layer \\( m = M \\) are then utilized to initialize cross-modal queries." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.796, + 0.773 + ], + "angle": 0, + "content": "3.2.2 Language-Guided Query Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This module is designed to select features more relevant to the input text as decoder queries for effective language-vision fusion. We combine a DETR/DINO [1, 31] style queries with a sinusoidal temporal positional encoding to the positional part of the queries. The sinusoidal positional encoding added to the positional part of the queries adds important contextual information regarding the sequence of frames, allowing for improved temporal correlation and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18912" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "grounding [28]. The query selection module takes the encoder's visual and textual features as input and outputs num_query indices that correspond to the most relevant features for object detection per frame, \\(\\{Q_t^0\\}_{t=1}^T\\), where \\(Q_t^0\\) are the initial queries for the frame \\(t\\). The module initializes the decoder queries using a combination of the selected indices and dynamic anchor boxes. The content part of the queries is set to be learnable during training, while the positional part is computed using the dynamic anchor boxes initialized using the encoder outputs. We also add a sinusoidal temporal positional encoding to the positional part of the queries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.293, + 0.423, + 0.308 + ], + "angle": 0, + "content": "3.2.3 Cross-Modality Spatio-Temporal Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.483 + ], + "angle": 0, + "content": "To decode the above queries into bounding box locations and temporal start/end tubes, we need to transform them into an output embedding, which can then be fed into prediction heads. The decoder allows for the queries to interact globally with others within a frame and across frames while utilizing the entire visual and textual features as context. Formally, the queries produced earlier are fed into a \\(N\\) layer decoder. Each layer starts with a temporal self-attention, a spatial self-attention followed by a visual cross-attention and textual cross-attention, and finally an FFN. This is represented in Eq. 4." + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.494, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\nQ _ {t} ^ {n ^ {\\prime}} = \\mathrm {M H S A} _ {\\text {s p a t i a l}} ^ {n} (\\mathrm {M H S A} _ {\\text {t e m p o r a l}} ^ {n} \\left(Q _ {t} ^ {n - 1}\\right)), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.515, + 0.414, + 0.535 + ], + "angle": 0, + "content": "\\[\nQ _ {t} ^ {n} = \\mathrm {F F N} ^ {n} (\\mathrm {C A} _ {p} ^ {n} (\\mathrm {C A} _ {v} ^ {n} (Q _ {t} ^ {n \\prime}, F _ {v} ^ {M}), F _ {p} ^ {M})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.47, + 0.638 + ], + "angle": 0, + "content": "where \\( Q_{t}^{n - 1} \\) are the input queries at layer \\( n\\in \\{1,\\dots,N\\} \\), \\( Q_{t}^{n^{\\prime}} \\) are the intermediate queries after spatial and temporal attention at layer \\( n \\), \\( Q_{t}^{n^{\\prime}} \\) are the output queries at layer \\( n \\), and \\( \\mathrm{CA}_v^n \\) and \\( \\mathrm{CA}_p^n \\) are the visual and textual cross-attention at layer \\( n \\). The cross-attentions are further elaborated in Eq. 5." + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.647, + 0.498, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C A} _ {v} ^ {n} \\left(Q _ {t} ^ {n \\prime}, F _ {v} ^ {M}\\right) = \\left(\\frac {\\operatorname {p r o j} _ {q , v} ^ {n} \\left(Q _ {t} ^ {n \\prime}\\right) \\operatorname {p r o j} _ {k , v} ^ {n} \\left(F _ {v} ^ {M}\\right) ^ {T}}{\\sqrt {d ^ {k}}} \\operatorname {p r o j} _ {v} ^ {n} \\left(F _ {v} ^ {M}\\right) ^ {T}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.687, + 0.498, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C A} _ {p} ^ {n} \\left(\\mathrm {C A} _ {v} ^ {n}, F _ {p} ^ {M}\\right) = \\left(\\frac {\\operatorname {p r o j} _ {q , p} ^ {n} \\left(\\mathrm {C A} _ {v} ^ {n}\\right) \\operatorname {p r o j} _ {k , p} ^ {n} \\left(F _ {p} ^ {M}\\right) ^ {T}}{\\sqrt {d ^ {k}}} \\operatorname {p r o j} _ {p} ^ {n} \\left(F _ {p} ^ {M}\\right) ^ {T}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.812 + ], + "angle": 0, + "content": "where \\(proj_{q,v}^n, proj_{k,v}^n\\) and \\(proj_v^n\\) are the visual query, key and values projection for layer \\(n\\) and \\(proj_{q,p}^n, proj_{k,p}^n\\) and \\(proj_p^n\\) are the textual query, key and value projections. The final queries from the decoder at layer \\(N\\), \\(\\{Q_t^N\\}_{t=1}^T\\), are then used for prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.251, + 0.846 + ], + "angle": 0, + "content": "3.2.4 Prediction Heads" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The decoder outputs refined queries per frame \\(\\{Q_t^N\\}_{t = 1}^T\\) We follow the standard DETR-like bounding box regression head implemented as a Multi-Layer Perceptron (MLP)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.243 + ], + "angle": 0, + "content": "which predicts bounding boxes \\( B_{i}^{t} = (x_{i}^{t},y_{i}^{t},w_{i}^{t},h_{i}^{t}) \\), per frame. To predict the temporal interval \\( (t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\} \\), we add a temporal grounding head, implemented as an MLP, alongside the bounding box regression head, similar to existing works like [9, 28]. The new head predicts the probabilities of the start \\( \\tau_{s}\\in [0,1]^{T} \\) and ends \\( \\tau_{e}\\in [0,1]^{T} \\) of the interval. During inference, the start and end interval \\( (t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\} \\) is computed by taking the maximum of the joint distribution of \\( (\\tau_s,\\tau_e) \\). Any invalid combinations with \\( t_e\\leq t_s \\) are masked out." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.255, + 0.644, + 0.269 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.385 + ], + "angle": 0, + "content": "To leverage the generalized pre-trained representation from spatial-grounding foundation models, we initialize all spatial modules and cross attentions from the Grounding DINO [12] spatial grounding model. To preserve this generalized representation while ensuring effective modeling of the downstream task, we freeze the Vision and Text Encoders \\(\\theta_v\\) and \\(\\theta_p\\) and fine-tune the remaining components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.521 + ], + "angle": 0, + "content": "During training, the model receives a batch of videos \\(V\\) with text prompt \\(P\\). The ground-truth annotation contains the bounding box sequence \\(\\{B_i^t\\}_{i = t_s}^{t_e}\\), and the corresponding start and end timestamps \\((t_{s},t_{e})\\). For spatial grounding, we follow the standard loss formulation used in DETR-like [1, 31, 35], namely the \\(L_{1}\\) loss, \\(\\mathcal{L}_{L_1}\\), and the Generalized Intersection over Union (GIoU) [17] loss, \\(\\mathcal{L}_{GIoU}\\). Formally, the spatial grounding loss, \\(\\mathcal{L}_{spatial}\\) is defined in Eq. 6." + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.533, + 0.892, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\lambda_ {L _ {1}} \\mathcal {L} _ {L _ {1}} (\\hat {B}, B) + \\lambda_ {G I o U} \\mathcal {L} _ {G I o U} (\\hat {B}, B). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.626 + ], + "angle": 0, + "content": "For temporal grounding, we follow [18, 21, 28] and generate two 1-dimensional gaussian heatmaps \\(\\pi_s,\\pi_e\\in \\mathcal{R}^T\\) for the starting and ending positions. The temporal grounding loss is therefore defined in Eq. 7 as," + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.639, + 0.892, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p o r a l}} = \\mathcal {L} _ {K L} ^ {s} \\left(\\hat {\\pi} _ {s}, \\pi_ {s}\\right) + \\mathcal {L} _ {K L} ^ {e} \\left(\\hat {\\pi} _ {e}, \\pi_ {e}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.89, + 0.699 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{KL}^s\\) and \\(\\mathcal{L}_{KL}^{e}\\) are the KL divergence losses for the start and end distributions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Note that the model outputs the bounding boxes and starting/ending distributions during inference. We determine the temporal grounding segment, \\((t_s,t_e)\\), by taking the segment with the maximal joint start and end probability. Then, we consider the bounding boxes only within that tube for spatial grounding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.587, + 0.822 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.799, + 0.848 + ], + "angle": 0, + "content": "4.1. Experimental Setup and Protocols" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Below, we first briefly explain the implementation details (Sec. 4.1.1), followed by evaluation settings (Sec. 4.1.2), and datasets (Sec. 4.1.3) used in our work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18913" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.891, + 0.104 + ], + "angle": 0, + "content": "Table 1. Performance comparisons of the state-of-the-art on HC-STVG V1 [24] and YouCook-Interactions [23] in open-vocabulary setting." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.115, + 0.84, + 0.213 + ], + "angle": 0, + "content": "
MethodPre-trainingHC-STVG V1YouCook-Interactions
m_vIoUvIoU@0.3vIoU@0.5Accuracy
TubeDETR (CVPR'22) [28]VidSTG16.8422.329.2251.63
STCAT (NeurIPS'22) [9]VidSTG22.5832.1420.8355.90
VideoGrounding-DINOVidSTG27.4640.1329.9257.73
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.228, + 0.295, + 0.243 + ], + "angle": 0, + "content": "4.1.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.471, + 0.418 + ], + "angle": 0, + "content": "As discussed in the methodology (Sec. 3), we initialize the spatial modules in our model from the Grounding DINO [12] spatial grounding model and keep the vision and text encoders frozen. Our prediction heads for both spatial and temporal predictions are set to be 3-layer Multi-Layer Perceptrons (MLPs). We sample 128 frames during training and inference, resized to a resolution of 448 on the shorter side. We set both \\( M \\) and \\( N \\) to 6, and train the model with a batch size of 8 and learning rate of \\( 1e^{-4} \\), and weight decay if \\( 10^{-4} \\). The number of epochs for VidSTG is set to 10, and for HC-STVG V1/V2 is set to 90." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.267, + 0.455 + ], + "angle": 0, + "content": "4.1.2 Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.463, + 0.469, + 0.493 + ], + "angle": 0, + "content": "We evaluate our video grounding model in two settings, Open-Vocabulary and Closed-Set Supervised." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Open-Vocabulary Evaluation: In the open-vocabulary setting we train our model on the VidSTG [32] dataset and then evaluate on two different datasets, HC-STVG V1 [24] and YouCook-Interactions [23] to understand how well the model generalizes to new distributions. The reason for choosing these two datasets is that the former provides a relatively minor distribution shift given the similar perspective/objects in the videos compared to the training dataset VidSTG. In contrast, the latter provides a major distribution shift with changes in perspective and annotated objects/interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.751 + ], + "angle": 0, + "content": "Closed-Set Supervised Evaluation: In the supervised evaluation setting, we train on the training set and evaluate each dataset's respective validation/testing set. This evaluation is conducted for three majorly used datasets in spatio-temporal video grounding, namely VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.192, + 0.785 + ], + "angle": 0, + "content": "4.1.3 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We evaluate our approach and compare against the state-of-the-art in two settings: Open-Vocabulary and Closed-Set Supervised, across a total of four grounding datasets, namely: VidSTG [32], HCSTVG V1 [24], HCSTVG V2 [24], and YouCook-Interactions [23]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "VidSTG: The VidSTG [32] dataset is derived from the VidOR [20] dataset, incorporating object relation annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.333 + ], + "angle": 0, + "content": "It includes 99,943 video-text pairs, encompassing 44,808 declarative sentence queries and 55,135 interrogative sentence queries. The training, validation, and test sets consist of 80,684, 8,956, and 10,303 sentences and 5,436, 602, and 732 videos, respectively. VidSTG's text queries are confined to describing pre-defined object/relation categories in VidOR [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.501 + ], + "angle": 0, + "content": "HC-STVG V1/V2: The HC-STVG datasets are sourced from movie scenes, each video clip spanning approximately 20 seconds. These datasets pose challenges in spatiotemporal grounding due to video clips featuring multiple individuals engaged in similar actions. HC-STVG V1 comprises 4,500 training and 1,160 testing video-text pairs. HC-STVG V2 expands HC-STVG V1, enhancing annotation quality with 10,131, 2,000, and 4,413 samples for training, validation, and testing, respectively. As HC-STVG V2's test set annotations are unavailable publicly, results are reported on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.622 + ], + "angle": 0, + "content": "YouCook-Interactions: The YouCook-Interactions [23] dataset serves as an expansion of the YouCook2 [34] dataset focused on cooking instructions. This extension includes bounding boxes for 6,000 carefully chosen frames, typically encompassing the hand and the tool specified in the corresponding sentence-level annotations. Our assessment revolves around examining models' spatial grounding capabilities using this dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.805, + 0.65 + ], + "angle": 0, + "content": "4.2. Experimental Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.732 + ], + "angle": 0, + "content": "In this section, we present our results across the evaluation mentioned above settings (Sec. 4.1.2) and datasets (Sec. 4.1.3). We start with the closed-set evaluation in Sec. 4.2.2, followed by the open-vocabulary evaluation in Sec. 4.2.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.757, + 0.77 + ], + "angle": 0, + "content": "4.2.1 Open-Vocabulary Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For open-vocabulary evaluation, we train on VidSTG [32] and present results HC-STVG V1 [24] and YouCook-Interactions [23]. The results are reported jointly in Tab. 1. Results on HC-STVG V1: We report open-vocabulary evaluation on m_vIoU, vIoU@0.3, and vIoU@0.5. We achieve state-of-the-art performance over both TubeDETR [28] and STCAT [9]. We attribute this strong performance to our design, which leverages the strong pre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18914" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.09, + 0.832, + 0.105 + ], + "angle": 0, + "content": "Table 2. Performance comparisons of the state-of-the-art on the VidSTG [32] test set in closed-set supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.115, + 0.895, + 0.4 + ], + "angle": 0, + "content": "
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3
Factorized:
GroundRe (ECCV'16) [19]+TALL (ICCV'17) [7]9.7811.044.099.3211.39
STPR (ICCV'17) [27]+TALL (ICCV'17) [7]34.6310.4012.384.2733.739.9811.74
WSSTG (arXiv'19) [5]+TALL (ICCV'17) [7]11.3614.635.9110.6513.90
GroundRe (ECCV'16) [19]+L-Net (AAAI'19) [3]11.8915.325.4511.0514.28
STPR (ICCV'17) [27]+L-Net (AAAI'19) [3]40.8612.9316.275.6839.7911.9414.73
WSSTG (arXiv'19) [5]+L-Net (AAAI'19) [3]14.4518.007.8913.3617.39
Two-Stage:
STGRN (CVPR'20) [32]48.4719.7525.7714.6046.9818.3221.10
STGVT (TCSVT'21) [24]-21.6229.8018.94---
OMRN (IJCAI'21) [33]50.7323.1132.6116.4249.1920.6328.35
One-Stage:
STVGBert (ICCV'21) [21]-23.9730.9118.39-22.5125.97
TubeDETR (CVPR'22) [28]48.1030.4042.5028.2046.9025.7035.70
STCAT (NeurIPS'22) [9]50.8233.1446.2032.5849.6728.2239.24
STVGFormer (CVPR'23) [11]-33.7047.2032.80-28.5039.90
VideoGrounding-DINO51.9734.6748.1133.9650.8329.8941.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.413, + 0.471, + 0.441 + ], + "angle": 0, + "content": "Table 3. Performance comparisons of the state-of-the-art on the HC-STVG V1 [24] test set in closed-set supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.452, + 0.47, + 0.559 + ], + "angle": 0, + "content": "
Methodsm_vIoUvIoU@0.3vIoU@0.5
STGVT (TCSVT'21) [24]18.1526.819.48
STVGBert (ICCV'21) [21]20.4229.3711.31
TubeDETR (CVPR'22) [28]32.4049.8023.50
STCAT (NeurIPS'22) [9]35.0957.6730.09
STVGFormer (CVPR'23) [11]36.9062.2034.80
VideoGrounding-DINO38.2562.4736.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.589, + 0.471, + 0.618 + ], + "angle": 0, + "content": "Table 4. Performance comparisons of the state-of-the-art on the HC-STVG V2 [24] val set in closed-set supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.628, + 0.47, + 0.723 + ], + "angle": 0, + "content": "
Methodsm_vIoUvIoU@0.3vIoU@0.5
Yu et al (arXiv'21) [30]30.00--
Aug. 2D-TAN (arXiv'21) [22]30.4050.4018.80
TubeDETR (CVPR'22) [28]36.4058.8030.60
STVGFormer (CVPR'23) [11]38.7065.5033.80
VideoGrounding-DINO39.8867.1334.49
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.754, + 0.47, + 0.784 + ], + "angle": 0, + "content": "trained generalized features of a spatial grounding foundation model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Results on YouCook-Interactions: We evaluate our method further on the YouCook-Interactions [23] dataset, reporting pointing game accuracy for spatial grounding. Our approach gains nearly \\(2\\%\\) in accuracy over STCAT [9] and more than \\(6\\%\\) compared to TubeDETR [28]. This further shows our strong generalization capabilities in the open-vocabulary setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.415, + 0.792, + 0.43 + ], + "angle": 0, + "content": "4.2.2 Closed-Set Supervised Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.892, + 0.49 + ], + "angle": 0, + "content": "We present closed-set evaluations across three datasets, VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Results on VidSTG: We present results on the VidSTG test set in closed-set setting in Tab. 2, reporting m_tIoU, m_vIoU, vIoU@0.3 and vIoU@0.5. The results show that our method achieves state-of-the-art performance in comparison to both Two-Stage and One-Stage methods. In particular, we achieve more than \\(1t\\_ IoU\\) gain in temporal grounding over the previous best methods OMRN [33] (One-Stage) and STVGFormer [11] (Two-Stage), both for Declarative and Interrogative sentences. Similarly, for m_vIoU, vIoU@0.3 and vIoU@0.5, we achieve a more than 1 unit gain the state-of-the-art methods STVGFormer [11] and STCAT [9]. Note that our method uses a frozen visual and textual encoder. In contrast, those mentioned above previous state-of-the-art methods all train the entire encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.894, + 0.814 + ], + "angle": 0, + "content": "Results on HC-STVG V1: We present results on the HC-STVG V1 dataset in Tab. 3, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. We achieve a nearly 1.5 unit gain in m_vIoU and vIoU@0.5 and a 1 unit gain in vIoU@0.3 over the previous best method STVGFormer [11]. This shows the consistent performance of our method on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Results on HC-STVG V2: We present results on the HC-STVG V2 dataset in Tab. 4, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. Our performance gain on HC-STVG V1 is reflected here as well, with a consistent performance gain on HC-STVG V2, in comparison to the SoTA [11, 28]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18915" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.09, + 0.843, + 0.104 + ], + "angle": 0, + "content": "Table 5. Ablation on various design choices for our approach on the VidSTG [32] test set in closed-set supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.115, + 0.894, + 0.226 + ], + "angle": 0, + "content": "
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3vIoU@0.5
Naive Solution (Frozen Grounding DINO [12])39.7818.0722.3113.7539.799.6610.423.84
+ Decoder Temporal Aggregation42.8120.7426.5315.4143.8112.3816.718.62
+ Encoder Temporal Aggregation46.2923.1932.3818.9547.1716.1923.2813.04
+ Finetuned Spatial Modules in Decoder48.0628.9741.6026.0649.5824.2732.8520.11
+ Finetuned Spatial Modules in Encoder51.9734.6748.1133.9650.8329.8941.0327.58
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.241, + 0.229, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.318, + 0.228, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.396, + 0.228, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.241, + 0.361, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.318, + 0.361, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.396, + 0.361, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.241, + 0.494, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.318, + 0.493, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.396, + 0.493, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.241, + 0.624, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.318, + 0.624, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.396, + 0.624, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.241, + 0.756, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.318, + 0.756, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.396, + 0.756, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.241, + 0.888, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.318, + 0.888, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.396, + 0.888, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.479, + 0.893, + 0.536 + ], + "angle": 0, + "content": "Figure 3. Sample visualization for video grounding result on HC-STVG V1 [24] for TubeDETR [28], STCAT [9] and ours with the prompt The man behind the shirtless man turns and squats. We show bounding boxes for Ground-truth, Closed-Set Supervised, and OpenVocabulary results. Note how both TubeDETR and STCAT are close to the ground truth in the supervised setting (STCAT more so than TubeDETR), they cannot correctly ground the text properly in the open-vocabulary setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.248, + 0.567 + ], + "angle": 0, + "content": "4.3. Ablative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.773 + ], + "angle": 0, + "content": "We perform an ablative analysis of the various design choices for our model. In particular, we first evaluate a naive baseline where no additional temporal aggregators are added, and all pre-trained spatial modules are frozen in the encoder and decoder. This baseline is relatively weak in both temporal and spatial grounding. We next add temporal modules in first the decoder, and subsequently in the encoder. We find that it gives significant improvements in the temporal grounding, and additionally also improves the spatial grounding. Finally, we finetune the pre-trained spatial modules in both the decoder and the encoder, which provides a strong improvement in spatial grounding, alongside consistent improvement in temporal grounding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.196, + 0.8 + ], + "angle": 0, + "content": "4.4. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "While our video grounding model excels in closed-set and open-vocabulary scenarios, it leverages image-text pretrained models like Grounding DINO [12]. To enhance understanding in open-vocabulary settings, an extension to video-language pre-training on a larger and more diverse dataset, akin to CLIP [15], can help further boost perfor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.613 + ], + "angle": 0, + "content": "mance. Building a video-language pre-training dataset with diverse natural language expressions and spatio-temporal localization is imperative, given the constraints of datasets like VidSTG [32] and HC-STVG [24]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.619, + 0.658 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.902 + ], + "angle": 0, + "content": "This paper introduces an Open-Vocabulary Spatio-Temporal Video Grounding task, enhancing current closed-set methodologies by using pre-trained representations from spatial grounding models. The proposed model performs well in closed-set and open-vocabulary scenarios, surpassing state-of-the-art results in supervised setting on VidSTG and HC-STVG datasets, and outperforming recent models in open-vocabulary on HC-STVG V1 and YouCook-Interactions. Its architecture includes learnable adapter blocks for video-specific adaptation, bridging the semantic gap between natural language queries and visual content. This research addresses open-vocabulary challenges and explores achieving robust performance without extensive video annotations, paving the way for open-vocabulary video grounding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18916" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.256 + ], + "angle": 0, + "content": "[2] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where? - self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[3] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, , and Jiebo Luo. Localizing natural language in videos. In AAAI, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Jiyang Gao, Chen Sun, Zhenheng Yang, , and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[8] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.471, + 0.67 + ], + "angle": 0, + "content": "[9] Yang Jin, Yongzhi Li, Zehuan Yuan, and Yadong Mu. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In NeurIPS, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[10] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Zihang Lin, Chaolei Tan, Jian-Fang Hu, Zhi Jin, Tiancai Ye, and Wei-Shi Zheng. Collaborative static and dynamic vision-language streams for spatio-temporal video grounding. In CVPR, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[12] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[15] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Hanoona Rasheeda, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[17] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[18] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[19] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, , and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In ECCV, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In ICMR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Rui Su, Qian Yu., and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In ICCV, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[22] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[23] Reuben Tan, Bryan A. Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[26] Zhenzhi Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[27] Masataka Yamaguchi, Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Spatio-temporal person retrieval via natural language queries. In ICCV, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18917" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[30] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[31] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[32] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In CVPR, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[33] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In *IJCAI*, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[34] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18918" + } + ] +] \ No newline at end of file diff --git a/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_origin.pdf b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1c39ea9b119833ad1950880d5ce2e3225dc60c2 --- /dev/null +++ b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/8669738c-7c6c-4e4e-b34d-5818146c2193_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:950ba461549809aa06bbc59821ca5249c8f697f72bc684d6930c964a10a1af45 +size 4051121 diff --git a/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/full.md b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59141977dcad54a0d885a06528a608dabc22ee2e --- /dev/null +++ b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/full.md @@ -0,0 +1,317 @@ +# VideoGrounding-DINO: Towards Open-Vocabulary Spatio-Temporal Video Grounding + +Syed Talal Wasim1 Muzammal Naseer1 +Salman Khan1,2 Ming-Hsuan Yang3,4 Fahad Shahbaz Khan1,5 + +$^{1}$ Mohamed bin Zayed University of AI $^{2}$ Australian National University $^{3}$ University of California, Merced $^{4}$ Google Research $^{5}$ Linköping University + +# Abstract + +Video grounding aims to localize a spatio-temporal section in a video corresponding to an input text query. This paper addresses a critical limitation in current video grounding methodologies by introducing an Open-Vocabulary Spatio-Temporal Video Grounding task. Unlike prevalent closed-set approaches that struggle with open-vocabulary scenarios due to limited training data and predefined vocabularies, our model leverages pre-trained representations from foundational spatial grounding models. This empowers it to effectively bridge the semantic gap between natural language and diverse visual content, achieving strong performance in closed-set and open-vocabulary settings. Our contributions include a novel spatio-temporal video grounding model, surpassing state-of-the-art results in closed-set evaluations on multiple datasets and demonstrating superior performance in open-vocabulary scenarios. Notably, the proposed model outperforms state-of-the-art methods in closed-set settings on VidSTG (Declarative and Interrogative) and HC-STVG (V1 and V2) datasets. Furthermore, in open-vocabulary evaluations on HC-STVG V1 and YouCook-Interactions, our model surpasses the recent best-performing models by $4.88\mathrm{m\_vIoU}$ and $1.83\%$ accuracy, demonstrating its efficacy in handling diverse linguistic and visual concepts for improved video understanding. Our codes will be publicly released. + +# 1. Introduction + +Spatio-temporal video grounding is pivotal in linking visual content with natural language descriptions, thus facilitating semantics interpretation within visual data. Prevailing approaches in video grounding such as TubeDETR [28], STCAT [9], and STVGFormer [11] focus mainly on supervised closed-set settings, where models are trained on specific datasets [24, 32] with predefined vocabulary and meticulously annotated data. While these current state-of + +![](images/83b2e303b48c452304cd569f4a882c31b066146be3d773ffec2a4b4823ab5291.jpg) +Figure 1. Performance comparison on conventional closed-set and open-vocabulary settings for the video grounding task. We compare our approach with TubeDETR [28] and STCAT [9] in supervised setting for VidSTG [32] declarative/interrogative and HC-STVG V1 [24], along with open-vocabulary evaluation on HC-STVG V1 and YouCook-Interactions [23] datasets. + +the-art methods excel in closed-set settings on datasets like VidSTG [32] and HC-STVG [24], their limited generalization beyond training dataset distributions poses a significant challenge. The relatively small scale and restricted sample variety in existing video datasets hinder models from adapting to unseen scenarios effectively. + +Motivated by the inherent limitation of supervised closed-set approaches in terms of their restricted vocabulary, this paper investigates Open-Vocabulary Spatio-Temporal Video Grounding. Unlike conventional methodologies, this paradigm addresses challenges posed by the unrestricted diversity of language and visual concepts in the wild. The goal is to train on a set of base categories and generalize to unseen objects/actions based on the open-vocabulary nature of backbone models. This paper ex + +plores the challenges and opportunities inherent in open-vocabulary video grounding, laying the groundwork for more robust and versatile video understanding. + +However, training an effective open-vocabulary video grounding model would require a large enough dataset with a rich set of natural language expressions and corresponding spatio-temporal localizations. Such an extensive dataset can allow the model to learn generalized visual and textual representations and handle out-of-distribution samples. However, video grounding datasets [24, 32] are quite limited in scale, e.g., VidSTG has only $5.4\mathrm{k}$ training videos with $80.6\mathrm{k}$ distinct sentences. In contrast, an image grounding model GLIP [10] is trained with $\sim 26.5\mathrm{M}$ image-text pairs. Therefore, our research explores the following fundamental question: How spatio-temporal video grounding models can achieve robust performance in both Closed-Set and Open-Vocabulary scenarios without requiring large-scale video annotations, ensuring effective generalization beyond training datasets? + +In addressing this question, we find inspiration in the accomplishments of foundational models specializing in spatial grounding [4, 8, 12, 14, 29]. These models undergo training on an extensive corpus of image-text data, enabling effective generalization to samples from a given target distribution. We aim to harness this pretrained representation to enhance our video grounding model. Our proposed solution is a spatio-temporal video grounding model adopting a DETR-like architecture enhanced by temporal aggregation modules. The spatial modules are initialized using the pretrained representation of a foundational image model [12]. Meanwhile, the image and text feature extractors remain frozen while the video-specific spatio-temporal adaptations are modeled via learnable adapter blocks. This approach is designed to preserve the nuanced representation of the foundational model, enhancing our model's ability to generalize effectively to novel samples. + +A summary of our closed-set and open-vocabulary results is shown in Fig. 1, where the proposed approach excels in both settings by a clear margin. Our major contributions are summarized as follows: + +- For the first time, we evaluate spatio-temporal video grounding models in an open-vocabulary setting on HC-STVG V1 [24] and YouCook-Interactions [23] benchmarks in a zero-shot manner. We outperform state-of-the-art methods TubeDETR [28] and STCAT [9] by 4.26 m_vIoU and $1.83\%$ accuracy, respectively. +- By combining the strengths of spatial grounding models with complementary video-specific adapters, our approach consistently outperforms the previous state-of-the-art in closed-set setting on four benchmarks, i.e., VidSTG (Declarative) [32], VidSTG (Interrogative) [32], HC-STVG V1 [24] and HC-STVG V2 [24]. + +# 2. Related Work + +Spatial Grounding Foundation Models: Recent literature introduces notable spatial grounding models. GLIP [10] unifies object detection and phrase grounding through a language-image pre-training model, leveraging extensive image-text pairs for semantic-rich representations. Grounding DINO [12] integrates language with a transformer-based detector to achieve an open-set detector, excelling in benchmarks like COCO and ODinW. Kosmos-1 [8] and Kosmos-2 [14] contribute Multimodal Large Language Models (MLLMs) with capabilities such as zero-shot and few-shot learning, language understanding, and multimodal tasks. Kosmos-2 [14] specifically integrates grounding into downstream applications, introducing GrIT, a large-scale dataset of Grounded Image-Text pairs. Shikra [4] addresses referential ability in MLLMs by handling spatial coordinates in inputs and outputs, showcasing promising performance in various vision-language tasks. Ferret [29] unifies referring and grounding in the LLM paradigm, achieving superior performance in classical referring and grounding tasks, excelling in region-based multimodal chatting and image description. Recently, GLaMM [16] allows pixel-level grounded conversations with an LLM, showcasing generalizability to several captioning and referring segmentation tasks. However, spatial methods cannot work for grounding objects in videos, a gap addressed by this work. + +Spatio-Temporal Video Grounding: Several methods tackle the challenge of localizing objects in untrimmed videos based on query sentences. STVGBert [21] presents a one-stage visual-linguistic transformer for simultaneous spatial and temporal localization. TubeDETR [28] introduces a transformer-based architecture to model temporal, spatial, and multi-modal interactions efficiently. Augmented 2D-TAN [22] adopts a two-stage approach, enhancing the 2D-TAN with a temporal context-aware Bi-LSTM Aggregation Module. OMRN [33] addresses the challenge of unaligned data and multi-form sentences in spatiotemporal video grounding, proposing an object-aware multi-branch relation network for effective relation discovery. MMN [26] introduces a Mutual Matching Network as a metric-learning framework for temporal grounding, achieving competitive performance. STCAT [9] is an end-to-end one-stage framework addressing feature alignment and prediction inconsistency. Finally, STVGFformer [11] proposes an effective framework with static and dynamic branches for cross-modal understanding. While the above methods advance video grounding, their generalization to out-of-distribution and open-vocabulary samples is limited due to constrained video datasets [24, 32]. To address this issue, we utilize the generalized representation of spatial grounding foundation models [4, 8, 10, 12, 14, 29] trained on a large corpus of image-text data and can perform well on both closed-set and open-vocabulary evaluations. + +# 3. Methodology + +As discussed above, the current state-of-the-art spatiotemporal video grounding methods [9, 11, 21, 22, 24, 26, 28, 30, 32, 33] primarily evaluate in a supervised setting on the VidSTG [32] and HC-STVG [24] datasets. However, these methods lack the multimodal spatio-temporal understanding required to perform well on out-of-distribution samples [2]. Therefore, this work aims to achieve improved open-vocabulary performance while maintaining strong closed-set video-grounding performance. + +We take inspiration from recent foundation models for spatial grounding [4, 8, 10, 12, 14, 29]. These models are trained on a large corpus of visual-textual data and hence, generalize well to unseen samples. We aim to leverage the strong generalization capabilities of such foundation models to achieve strong open-set spatio-temporal video grounding performance. Our proposed spatio-temporal video grounding method uses DETR-like [1] design, with temporal aggregation and adaptation modules for learning video-specific representations. + +Below, we explain our proposed methodology. We formally define the spatio-temporal video grounding problem in Sec. 3.1. We then explain our architecture details in Sec. 3.2. We finally explain the loss formulation used to train the model and model initialization in Sec. 3.3. + +# 3.1. Problem Definition + +The spatio-temporal video grounding task involves localizing and recognizing objects and actions in a video sequence by integrating spatial and temporal information. In contrast to spatial grounding, which focuses on recognizing and localizing objects or actions within individual frames, spatiotemporal grounding extends this concept to include the temporal dimension. This means understanding where objects or actions are in each frame and how they evolve and move over time. + +Consider a video $V \in \mathbb{R}^{T \times H \times W \times C}$ with $T$ frames, $H \times W$ spatial resolution, and $C$ channels, respectively, along with a text prompt $P$ . The spatial grounding problem can be defined as the localization of one or more objects associated with the prompt $P$ in a frame $V_{t}, t \in \{1, \dots, T\}$ using a bounding box $B_{i}^{t} = (x_{i}^{t}, y_{i}^{t}, w_{i}^{t}, h_{i}^{t})$ , where $x_{i}^{t}$ and $y_{i}^{t}$ are the coordinates of the top-left corner, $w_{i}^{t}$ and $h_{i}^{t}$ are the width and height of the bounding box, $i \in \{1, \dots, N\}$ is the object number for the frame $t$ . The temporal grounding problem, on the other hand, involves understanding how objects or actions evolve over time. It aims at localizing the temporal interval $(t_{s}, t_{e})$ where the specific action/interaction happens in the entire temporal duration, where interval $(t_{s}, t_{e})$ indicates the start and end frame of the object occurrence within the total frames $T$ ( $1 \leq t_{s} < t_{e} \leq T$ ). Hence, the spatio-temporal grounding problem for object $i$ associated with prompt $P$ can be summarized as a + +set of spatio-temporal coordinates associated with the subset of frames where the object exists: $(x_i^t,y_i^t,w_i^t,h_i^t,t)$ and $t\in \{t_s,\dots,t_e\}$ . The interval $(t_s,t_e)|\{1\leq t_s < t_e\leq T\}$ and is a subset of the total frames $T$ + +# 3.2. Spatio-Temporal Video Grounding + +Here, we explain our video grounding model in Fig. 2. As discussed earlier, we aim to design a spatio-temporal video grounding model that can perform well in closed-set and open-vocabulary settings. Strong open-vocabulary performance requires learning a rich visual/textual representation, which in turn requires a large amount of training data. Unfortunately, spatio-temporal video grounding datasets are quite limited in scale [24, 32], resulting in current video grounding methods failing in generalizing well to out-of-distribution samples because they lack the requisite strong visual/textual representation. + +To solve this problem, our approach takes inspiration from recent spatial grounding methods [4, 8, 12, 14, 29], which have strong open-vocabulary performance thanks to the large image-text corpus they are trained on. We can utilize the generalized representations of these models to enrich the weaker representation of video-grounding approaches obtained from the limited number of training samples. Our approach aims to leverage the strong pretrained representations of spatial grounding methods to achieve strong closed-set supervised and open-vocabulary video grounding performance. Our spatio-temporal video grounding approach is based on the state-of-the-art DETR-based [1] object detection framework DINO [31] and also borrows concepts of image-text alignment and grounding from Grounded Language-Image Pre-training (GLIP) [10] and Grounding DINO [12]. We extract initial features from backbone vision and text encoders $\theta_v$ and $\theta_p$ . Following that, we model inter-frame and intra-frame features and learn cross-modal visual/textual relations in the Cross-Modality Spatio-Temporal Encoder (Sec. 3.2.1). The result enriched cross-modal features are used to initialize queries for each frame (Sec. 3.2.2). These queries are then decoded to predict the bounding boxes per frame and the temporal grounding start/end frame by aggregating information across spatial/temporal dimensions and injecting information through cross-attention from enriched visual/textual context (Sec. 3.2.3). + +Given the video $V$ and text prompt $P$ as described above, we obtain per-frame features $F_{v}^{0}$ and text features $F_{p}^{0}$ from vision and text encoders, $\theta_v$ and $\theta_p$ , respectively. The vision encoder is based on a Swin Transformer [13], and the text-encoder is defined as a BERT [6] model. Like other DETR-based detectors [31, 35], image features are extracted from different vision-encoder blocks at multiple scales. The features are then passed to the Cross-Modality Spatio-Temporal Encoder. + +![](images/aee7204402aeb18acb3ae3d1e5eb4d4b0cbaff7cdc169590e9d1cd00e392b2f5.jpg) +Figure 2. Overall architecture: We present our video grounding architecture. It consists of vision and text encoders that produce visual and textual features. A cross-modality spatio-temporal encoder which fuses information across spatial/temporal dimensions and visual/textual modalities. A language guided query selection module to initialize cross-modal queries. A cross-modality spatio-temporal decoder to decoder queries while fusing information from visual/textual features. And finally two prediction heads to predict the bounding boxes per frame and the temporal tube. Modules with $(\bullet)$ are trainable and those with $(\text{串})$ are frozen. + +# 3.2.1 Cross-Modality Spatio-Temporal Encoder + +The initial vision and text features $F_{v}^{0}$ and $F_{p}^{0}$ neither contain any cross-modal information nor model the temporal relationship across frames. Therefore, we further encode the initial features through the Cross-Modality Spatio-Temporal Encoder to model the temporal information across frames and learn cross-modal features. + +Each layer of the $M$ layer encoder first applies a Multi-Head Self-Attention (MHSA) [25] to the visual features $F_{v}$ along the temporal dimension, followed by a Deformable Attention (DA) [35] along the spatial dimension. This is done to model relations within frames and temporally across frames. Similarly, we apply a MHSA on the text features $F_{p}$ . This is illustrated in Eq. 1. + +$$ +\begin{array}{l} F _ {v} ^ {m \prime} = \mathrm {D A} _ {s p a t i a l} ^ {m} (\mathrm {M H S A} _ {t e m p o r a l} ^ {m} (F _ {v} ^ {m - 1})), \\ F _ {p} ^ {m \prime} = \mathrm {M H S A} _ {p} ^ {m} (F _ {p} ^ {m - 1}), \end{array} \tag {1} +$$ + +where $F_{v}^{m - 1}$ and $F_{p}^{m - 1}$ are visual and textual input features to layer $m$ , $F_{v}^{m'}$ and $F_{p}^{m'}$ are the intermediate visual and textual feature representation, and $\mathrm{DA}_{\text{spatial}}^m$ , $\mathrm{MHSA}_{\text{temporal}}^m$ and $\mathrm{MHSA}_p^m$ are the spatial-deformable, temporal and textual attentions at layer $m \in \{1,\dots,M\}$ , respectively. Following initial spatial, temporal, and textual attentions, we fuse features across the visual and textual modalities, as done in GLIP [10]. + +More specifically, we calculate the joint visual-textual attention, $\mathrm{Attn}_{\mathrm{joint}}^{\mathrm{m}}$ , using projected intermediate features $F_{v}^{m'}$ and $F_{p}^{m'}$ . This attention is then used alongside the + +intermediate features $F_{v}^{m'}$ and $F_{p}^{m'}$ , to calculate the image-to-text and text-to-image cross-attentions as shown in Eq. 2 and Eq. 3. + +$$ +\mathrm {A t t n} _ {\mathrm {j o i n t}} ^ {\mathrm {m}} = \left(\frac {\operatorname {p r o j} _ {q , v} ^ {m} \left(F _ {v} ^ {m \prime}\right) \operatorname {p r o j} _ {q , p} ^ {m} \left(F _ {p} ^ {m \prime}\right) ^ {T}}{\sqrt {d ^ {k}}}\right), \tag {2} +$$ + +where $proj_{q,v}^{m}$ and $proj_{q,p}^{m}$ are the query projections for the visual and textual features, respectively, at layer $m$ . + +$$ +F _ {v} ^ {m} = \operatorname {F F N} _ {v} ^ {m} \left(\operatorname {s o f t m a x} \left(\operatorname {A t t n} _ {\text {j o i n t}} ^ {\mathrm {m}}\right) p r o j _ {p} ^ {m} \left(F _ {p} ^ {m \prime}\right)\right), \tag {3} +$$ + +$$ +F _ {p} ^ {m} = \mathrm {F F N} _ {p} ^ {m} (\mathrm {s o f t m a x} (\mathrm {A t t n} _ {\mathrm {j o i n t}} ^ {\mathrm {m} _ {T}}) p r o j _ {v} ^ {m} (F _ {v} ^ {m /})) +$$ + +where $F_{v}^{m}$ and $F_{p}^{m}$ are the final output features at layer $m$ of the encoder, $\mathrm{FFN}_v^m$ and $\mathrm{FFN}_v^m$ are the visual and textual Feed Forward Networks (FFN) and $proj_0^m$ and $proj_0^m$ are the linear layers to project the visual and textual features. The final encoded features at layer $m = M$ are then utilized to initialize cross-modal queries. + +# 3.2.2 Language-Guided Query Selection + +This module is designed to select features more relevant to the input text as decoder queries for effective language-vision fusion. We combine a DETR/DINO [1, 31] style queries with a sinusoidal temporal positional encoding to the positional part of the queries. The sinusoidal positional encoding added to the positional part of the queries adds important contextual information regarding the sequence of frames, allowing for improved temporal correlation and + +grounding [28]. The query selection module takes the encoder's visual and textual features as input and outputs num_query indices that correspond to the most relevant features for object detection per frame, $\{Q_t^0\}_{t=1}^T$ , where $Q_t^0$ are the initial queries for the frame $t$ . The module initializes the decoder queries using a combination of the selected indices and dynamic anchor boxes. The content part of the queries is set to be learnable during training, while the positional part is computed using the dynamic anchor boxes initialized using the encoder outputs. We also add a sinusoidal temporal positional encoding to the positional part of the queries. + +# 3.2.3 Cross-Modality Spatio-Temporal Decoder + +To decode the above queries into bounding box locations and temporal start/end tubes, we need to transform them into an output embedding, which can then be fed into prediction heads. The decoder allows for the queries to interact globally with others within a frame and across frames while utilizing the entire visual and textual features as context. Formally, the queries produced earlier are fed into a $N$ layer decoder. Each layer starts with a temporal self-attention, a spatial self-attention followed by a visual cross-attention and textual cross-attention, and finally an FFN. This is represented in Eq. 4. + +$$ +Q _ {t} ^ {n ^ {\prime}} = \mathrm {M H S A} _ {\text {s p a t i a l}} ^ {n} (\mathrm {M H S A} _ {\text {t e m p o r a l}} ^ {n} \left(Q _ {t} ^ {n - 1}\right)), \tag {4} +$$ + +$$ +Q _ {t} ^ {n} = \mathrm {F F N} ^ {n} (\mathrm {C A} _ {p} ^ {n} (\mathrm {C A} _ {v} ^ {n} (Q _ {t} ^ {n \prime}, F _ {v} ^ {M}), F _ {p} ^ {M})), +$$ + +where $Q_{t}^{n - 1}$ are the input queries at layer $n\in \{1,\dots,N\}$ , $Q_{t}^{n^{\prime}}$ are the intermediate queries after spatial and temporal attention at layer $n$ , $Q_{t}^{n^{\prime}}$ are the output queries at layer $n$ , and $\mathrm{CA}_v^n$ and $\mathrm{CA}_p^n$ are the visual and textual cross-attention at layer $n$ . The cross-attentions are further elaborated in Eq. 5. + +$$ +\mathrm {C A} _ {v} ^ {n} \left(Q _ {t} ^ {n \prime}, F _ {v} ^ {M}\right) = \left(\frac {\operatorname {p r o j} _ {q , v} ^ {n} \left(Q _ {t} ^ {n \prime}\right) \operatorname {p r o j} _ {k , v} ^ {n} \left(F _ {v} ^ {M}\right) ^ {T}}{\sqrt {d ^ {k}}} \operatorname {p r o j} _ {v} ^ {n} \left(F _ {v} ^ {M}\right) ^ {T}\right), +$$ + +$$ +\mathrm {C A} _ {p} ^ {n} \left(\mathrm {C A} _ {v} ^ {n}, F _ {p} ^ {M}\right) = \left(\frac {\operatorname {p r o j} _ {q , p} ^ {n} \left(\mathrm {C A} _ {v} ^ {n}\right) \operatorname {p r o j} _ {k , p} ^ {n} \left(F _ {p} ^ {M}\right) ^ {T}}{\sqrt {d ^ {k}}} \operatorname {p r o j} _ {p} ^ {n} \left(F _ {p} ^ {M}\right) ^ {T}\right) \tag {5} +$$ + +where $proj_{q,v}^n, proj_{k,v}^n$ and $proj_v^n$ are the visual query, key and values projection for layer $n$ and $proj_{q,p}^n, proj_{k,p}^n$ and $proj_p^n$ are the textual query, key and value projections. The final queries from the decoder at layer $N$ , $\{Q_t^N\}_{t=1}^T$ , are then used for prediction. + +# 3.2.4 Prediction Heads + +The decoder outputs refined queries per frame $\{Q_t^N\}_{t = 1}^T$ We follow the standard DETR-like bounding box regression head implemented as a Multi-Layer Perceptron (MLP), + +which predicts bounding boxes $B_{i}^{t} = (x_{i}^{t},y_{i}^{t},w_{i}^{t},h_{i}^{t})$ , per frame. To predict the temporal interval $(t_s,t_e)|\{1\leq t_s < t_e\leq T\}$ , we add a temporal grounding head, implemented as an MLP, alongside the bounding box regression head, similar to existing works like [9, 28]. The new head predicts the probabilities of the start $\tau_{s}\in [0,1]^{T}$ and ends $\tau_{e}\in [0,1]^{T}$ of the interval. During inference, the start and end interval $(t_s,t_e)|\{1\leq t_s < t_e\leq T\}$ is computed by taking the maximum of the joint distribution of $(\tau_s,\tau_e)$ . Any invalid combinations with $t_e\leq t_s$ are masked out. + +# 3.3. Loss Function + +To leverage the generalized pre-trained representation from spatial-grounding foundation models, we initialize all spatial modules and cross attentions from the Grounding DINO [12] spatial grounding model. To preserve this generalized representation while ensuring effective modeling of the downstream task, we freeze the Vision and Text Encoders $\theta_v$ and $\theta_p$ and fine-tune the remaining components. + +During training, the model receives a batch of videos $V$ with text prompt $P$ . The ground-truth annotation contains the bounding box sequence $\{B_i^t\}_{i = t_s}^{t_e}$ , and the corresponding start and end timestamps $(t_{s},t_{e})$ . For spatial grounding, we follow the standard loss formulation used in DETR-like [1, 31, 35], namely the $L_{1}$ loss, $\mathcal{L}_{L_1}$ , and the Generalized Intersection over Union (GIoU) [17] loss, $\mathcal{L}_{GIoU}$ . Formally, the spatial grounding loss, $\mathcal{L}_{spatial}$ is defined in Eq. 6. + +$$ +\mathcal {L} _ {\text {s p a t i a l}} = \lambda_ {L _ {1}} \mathcal {L} _ {L _ {1}} (\hat {B}, B) + \lambda_ {G I o U} \mathcal {L} _ {G I o U} (\hat {B}, B). \tag {6} +$$ + +For temporal grounding, we follow [18, 21, 28] and generate two 1-dimensional gaussian heatmaps $\pi_s,\pi_e\in \mathcal{R}^T$ for the starting and ending positions. The temporal grounding loss is therefore defined in Eq. 7 as, + +$$ +\mathcal {L} _ {\text {t e m p o r a l}} = \mathcal {L} _ {K L} ^ {s} \left(\hat {\pi} _ {s}, \pi_ {s}\right) + \mathcal {L} _ {K L} ^ {e} \left(\hat {\pi} _ {e}, \pi_ {e}\right), \tag {7} +$$ + +where $\mathcal{L}_{KL}^s$ and $\mathcal{L}_{KL}^{e}$ are the KL divergence losses for the start and end distributions, respectively. + +Note that the model outputs the bounding boxes and starting/ending distributions during inference. We determine the temporal grounding segment, $(t_s,t_e)$ , by taking the segment with the maximal joint start and end probability. Then, we consider the bounding boxes only within that tube for spatial grounding. + +# 4. Results + +# 4.1. Experimental Setup and Protocols + +Below, we first briefly explain the implementation details (Sec. 4.1.1), followed by evaluation settings (Sec. 4.1.2), and datasets (Sec. 4.1.3) used in our work. + +Table 1. Performance comparisons of the state-of-the-art on HC-STVG V1 [24] and YouCook-Interactions [23] in open-vocabulary setting. + +
MethodPre-trainingHC-STVG V1YouCook-Interactions
m_vIoUvIoU@0.3vIoU@0.5Accuracy
TubeDETR (CVPR'22) [28]VidSTG16.8422.329.2251.63
STCAT (NeurIPS'22) [9]VidSTG22.5832.1420.8355.90
VideoGrounding-DINOVidSTG27.4640.1329.9257.73
+ +# 4.1.1 Implementation Details + +As discussed in the methodology (Sec. 3), we initialize the spatial modules in our model from the Grounding DINO [12] spatial grounding model and keep the vision and text encoders frozen. Our prediction heads for both spatial and temporal predictions are set to be 3-layer Multi-Layer Perceptrons (MLPs). We sample 128 frames during training and inference, resized to a resolution of 448 on the shorter side. We set both $M$ and $N$ to 6, and train the model with a batch size of 8 and learning rate of $1e^{-4}$ , and weight decay if $10^{-4}$ . The number of epochs for VidSTG is set to 10, and for HC-STVG V1/V2 is set to 90. + +# 4.1.2 Evaluation Settings + +We evaluate our video grounding model in two settings, Open-Vocabulary and Closed-Set Supervised. + +Open-Vocabulary Evaluation: In the open-vocabulary setting we train our model on the VidSTG [32] dataset and then evaluate on two different datasets, HC-STVG V1 [24] and YouCook-Interactions [23] to understand how well the model generalizes to new distributions. The reason for choosing these two datasets is that the former provides a relatively minor distribution shift given the similar perspective/objects in the videos compared to the training dataset VidSTG. In contrast, the latter provides a major distribution shift with changes in perspective and annotated objects/interactions. + +Closed-Set Supervised Evaluation: In the supervised evaluation setting, we train on the training set and evaluate each dataset's respective validation/testing set. This evaluation is conducted for three majorly used datasets in spatio-temporal video grounding, namely VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24]. + +# 4.1.3 Datasets + +We evaluate our approach and compare against the state-of-the-art in two settings: Open-Vocabulary and Closed-Set Supervised, across a total of four grounding datasets, namely: VidSTG [32], HCSTVG V1 [24], HCSTVG V2 [24], and YouCook-Interactions [23]. + +VidSTG: The VidSTG [32] dataset is derived from the VidOR [20] dataset, incorporating object relation annotations. + +It includes 99,943 video-text pairs, encompassing 44,808 declarative sentence queries and 55,135 interrogative sentence queries. The training, validation, and test sets consist of 80,684, 8,956, and 10,303 sentences and 5,436, 602, and 732 videos, respectively. VidSTG's text queries are confined to describing pre-defined object/relation categories in VidOR [20]. + +HC-STVG V1/V2: The HC-STVG datasets are sourced from movie scenes, each video clip spanning approximately 20 seconds. These datasets pose challenges in spatiotemporal grounding due to video clips featuring multiple individuals engaged in similar actions. HC-STVG V1 comprises 4,500 training and 1,160 testing video-text pairs. HC-STVG V2 expands HC-STVG V1, enhancing annotation quality with 10,131, 2,000, and 4,413 samples for training, validation, and testing, respectively. As HC-STVG V2's test set annotations are unavailable publicly, results are reported on the validation set. + +YouCook-Interactions: The YouCook-Interactions [23] dataset serves as an expansion of the YouCook2 [34] dataset focused on cooking instructions. This extension includes bounding boxes for 6,000 carefully chosen frames, typically encompassing the hand and the tool specified in the corresponding sentence-level annotations. Our assessment revolves around examining models' spatial grounding capabilities using this dataset. + +# 4.2. Experimental Results and Analysis + +In this section, we present our results across the evaluation mentioned above settings (Sec. 4.1.2) and datasets (Sec. 4.1.3). We start with the closed-set evaluation in Sec. 4.2.2, followed by the open-vocabulary evaluation in Sec. 4.2.1. + +# 4.2.1 Open-Vocabulary Evaluation + +For open-vocabulary evaluation, we train on VidSTG [32] and present results HC-STVG V1 [24] and YouCook-Interactions [23]. The results are reported jointly in Tab. 1. Results on HC-STVG V1: We report open-vocabulary evaluation on m_vIoU, vIoU@0.3, and vIoU@0.5. We achieve state-of-the-art performance over both TubeDETR [28] and STCAT [9]. We attribute this strong performance to our design, which leverages the strong pre + +Table 2. Performance comparisons of the state-of-the-art on the VidSTG [32] test set in closed-set supervised setting. + +
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3
Factorized:
GroundRe (ECCV'16) [19]+TALL (ICCV'17) [7]9.7811.044.099.3211.39
STPR (ICCV'17) [27]+TALL (ICCV'17) [7]34.6310.4012.384.2733.739.9811.74
WSSTG (arXiv'19) [5]+TALL (ICCV'17) [7]11.3614.635.9110.6513.90
GroundRe (ECCV'16) [19]+L-Net (AAAI'19) [3]11.8915.325.4511.0514.28
STPR (ICCV'17) [27]+L-Net (AAAI'19) [3]40.8612.9316.275.6839.7911.9414.73
WSSTG (arXiv'19) [5]+L-Net (AAAI'19) [3]14.4518.007.8913.3617.39
Two-Stage:
STGRN (CVPR'20) [32]48.4719.7525.7714.6046.9818.3221.10
STGVT (TCSVT'21) [24]-21.6229.8018.94---
OMRN (IJCAI'21) [33]50.7323.1132.6116.4249.1920.6328.35
One-Stage:
STVGBert (ICCV'21) [21]-23.9730.9118.39-22.5125.97
TubeDETR (CVPR'22) [28]48.1030.4042.5028.2046.9025.7035.70
STCAT (NeurIPS'22) [9]50.8233.1446.2032.5849.6728.2239.24
STVGFormer (CVPR'23) [11]-33.7047.2032.80-28.5039.90
VideoGrounding-DINO51.9734.6748.1133.9650.8329.8941.03
+ +Table 3. Performance comparisons of the state-of-the-art on the HC-STVG V1 [24] test set in closed-set supervised setting. + +
Methodsm_vIoUvIoU@0.3vIoU@0.5
STGVT (TCSVT'21) [24]18.1526.819.48
STVGBert (ICCV'21) [21]20.4229.3711.31
TubeDETR (CVPR'22) [28]32.4049.8023.50
STCAT (NeurIPS'22) [9]35.0957.6730.09
STVGFormer (CVPR'23) [11]36.9062.2034.80
VideoGrounding-DINO38.2562.4736.14
+ +Table 4. Performance comparisons of the state-of-the-art on the HC-STVG V2 [24] val set in closed-set supervised setting. + +
Methodsm_vIoUvIoU@0.3vIoU@0.5
Yu et al (arXiv'21) [30]30.00--
Aug. 2D-TAN (arXiv'21) [22]30.4050.4018.80
TubeDETR (CVPR'22) [28]36.4058.8030.60
STVGFormer (CVPR'23) [11]38.7065.5033.80
VideoGrounding-DINO39.8867.1334.49
+ +trained generalized features of a spatial grounding foundation model. + +Results on YouCook-Interactions: We evaluate our method further on the YouCook-Interactions [23] dataset, reporting pointing game accuracy for spatial grounding. Our approach gains nearly $2\%$ in accuracy over STCAT [9] and more than $6\%$ compared to TubeDETR [28]. This further shows our strong generalization capabilities in the open-vocabulary setting. + +# 4.2.2 Closed-Set Supervised Evaluation + +We present closed-set evaluations across three datasets, VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24]. + +Results on VidSTG: We present results on the VidSTG test set in closed-set setting in Tab. 2, reporting m_tIoU, m_vIoU, vIoU@0.3 and vIoU@0.5. The results show that our method achieves state-of-the-art performance in comparison to both Two-Stage and One-Stage methods. In particular, we achieve more than $1t\_ IoU$ gain in temporal grounding over the previous best methods OMRN [33] (One-Stage) and STVGFormer [11] (Two-Stage), both for Declarative and Interrogative sentences. Similarly, for m_vIoU, vIoU@0.3 and vIoU@0.5, we achieve a more than 1 unit gain the state-of-the-art methods STVGFormer [11] and STCAT [9]. Note that our method uses a frozen visual and textual encoder. In contrast, those mentioned above previous state-of-the-art methods all train the entire encoder. + +Results on HC-STVG V1: We present results on the HC-STVG V1 dataset in Tab. 3, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. We achieve a nearly 1.5 unit gain in m_vIoU and vIoU@0.5 and a 1 unit gain in vIoU@0.3 over the previous best method STVGFormer [11]. This shows the consistent performance of our method on this dataset. + +Results on HC-STVG V2: We present results on the HC-STVG V2 dataset in Tab. 4, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. Our performance gain on HC-STVG V1 is reflected here as well, with a consistent performance gain on HC-STVG V2, in comparison to the SoTA [11, 28]. + +Table 5. Ablation on various design choices for our approach on the VidSTG [32] test set in closed-set supervised setting. + +
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3vIoU@0.5
Naive Solution (Frozen Grounding DINO [12])39.7818.0722.3113.7539.799.6610.423.84
+ Decoder Temporal Aggregation42.8120.7426.5315.4143.8112.3816.718.62
+ Encoder Temporal Aggregation46.2923.1932.3818.9547.1716.1923.2813.04
+ Finetuned Spatial Modules in Decoder48.0628.9741.6026.0649.5824.2732.8520.11
+ Finetuned Spatial Modules in Encoder51.9734.6748.1133.9650.8329.8941.0327.58
+ +![](images/ca88162c02f563393391dc4d069b87b269c2104263551c82160189ba87dca90a.jpg) + +![](images/ea612f6d83381b82ff1dfdfe080133886293bc62b3eeb0b2eec3c5adaadddfe8.jpg) + +![](images/98d29cc55f91c5c6a4967598f9e91a788603fe8d0f54a9fd47cb6021d88b7e95.jpg) +Figure 3. Sample visualization for video grounding result on HC-STVG V1 [24] for TubeDETR [28], STCAT [9] and ours with the prompt The man behind the shirtless man turns and squats. We show bounding boxes for Ground-truth, Closed-Set Supervised, and OpenVocabulary results. Note how both TubeDETR and STCAT are close to the ground truth in the supervised setting (STCAT more so than TubeDETR), they cannot correctly ground the text properly in the open-vocabulary setting. + +![](images/2ac908981847e7812c09c5e3877a061c747288912af8a2947b4a0df40ee61faf.jpg) + +![](images/73f969c386dda56ab005bd106858ad39533f7999b610795328bd3e11a132ff52.jpg) + +![](images/8f860a59b7490cbddaf0e2d4894768c07c0ee2fd17c2e1481ce2dab2202d0662.jpg) + +![](images/d83187aa9cd505b9f960cf5bcd00bc52ac638a865df585edc410be35cbd59a3c.jpg) + +![](images/8fe8f6d61883e0cc4d194384b5982d276019e540308fc337d149fb16f482d04b.jpg) + +![](images/9860f1dbe704e5b1ba0cdc49a4717a0d09331b17134a7b8c28264896aa2e25fc.jpg) + +![](images/1fb5ca49031dbe6f986f0952e7aada430234de34986f931ee813c92172a74a8e.jpg) + +![](images/0f4885f3d88815f1ab7d25e9196f26cf1ade29cd810562e02f1e98e698b30f2b.jpg) + +![](images/a6fee9a4934a0c0182e4bdada01195b6599d659e8ee0bd36fa0fc7d75559e84d.jpg) + +![](images/a1bbb07d32c6746de6dfebc43ac9cfca10403baf304bee4bf1606fc98ad12710.jpg) + +![](images/44e36a493c450f89d73962bdc7d207717185f1adbbf3be658576e04e621fa3f3.jpg) + +![](images/2c709ad6d1ce8374ee9b7c39816f002ab03819b31d13c3435199b2924bb0fdb0.jpg) + +![](images/317b3d8135774a989b8f0df6d5ba5d4480a274ae7a9d162bef6aeea9cc9d03e2.jpg) + +![](images/3bfc08974eaf1c5604943d9037789109a1d471a36d3f0a266b3ced507a0aab2f.jpg) + +![](images/25ef794d542e9e6bbc238641c65b5c2ef42c79dc304df9b0471659d58cf4617e.jpg) + +# 4.3. Ablative Analysis + +We perform an ablative analysis of the various design choices for our model. In particular, we first evaluate a naive baseline where no additional temporal aggregators are added, and all pre-trained spatial modules are frozen in the encoder and decoder. This baseline is relatively weak in both temporal and spatial grounding. We next add temporal modules in first the decoder, and subsequently in the encoder. We find that it gives significant improvements in the temporal grounding, and additionally also improves the spatial grounding. Finally, we finetune the pre-trained spatial modules in both the decoder and the encoder, which provides a strong improvement in spatial grounding, alongside consistent improvement in temporal grounding. + +# 4.4. Limitation + +While our video grounding model excels in closed-set and open-vocabulary scenarios, it leverages image-text pretrained models like Grounding DINO [12]. To enhance understanding in open-vocabulary settings, an extension to video-language pre-training on a larger and more diverse dataset, akin to CLIP [15], can help further boost perfor + +mance. Building a video-language pre-training dataset with diverse natural language expressions and spatio-temporal localization is imperative, given the constraints of datasets like VidSTG [32] and HC-STVG [24]. + +# 5. Conclusion + +This paper introduces an Open-Vocabulary Spatio-Temporal Video Grounding task, enhancing current closed-set methodologies by using pre-trained representations from spatial grounding models. The proposed model performs well in closed-set and open-vocabulary scenarios, surpassing state-of-the-art results in supervised setting on VidSTG and HC-STVG datasets, and outperforming recent models in open-vocabulary on HC-STVG V1 and YouCook-Interactions. Its architecture includes learnable adapter blocks for video-specific adaptation, bridging the semantic gap between natural language queries and visual content. This research addresses open-vocabulary challenges and explores achieving robust performance without extensive video annotations, paving the way for open-vocabulary video grounding. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3, 4, 5 +[2] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where? - self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 3 +[3] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, , and Jiebo Luo. Localizing natural language in videos. In AAAI, 2019. 7 +[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3 +[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549, 2019. 7 +[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2019. 3 +[7] Jiyang Gao, Chen Sun, Zhenheng Yang, , and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 7 +[8] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 2, 3 +[9] Yang Jin, Yongzhi Li, Zehuan Yuan, and Yadong Mu. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In NeurIPS, 2022. 1, 2, 3, 5, 6, 7, 8 +[10] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2, 3, 4 +[11] Zihang Lin, Chaolei Tan, Jian-Fang Hu, Zhi Jin, Tiancai Ye, and Wei-Shi Zheng. Collaborative static and dynamic vision-language streams for spatio-temporal video grounding. In CVPR, 2023. 1, 2, 3, 7 +[12] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 5, 6, 8 +[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3 +[14] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3 +[15] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 8 +[16] Hanoona Rasheeda, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2 +[17] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union. In CVPR, 2019. 5 +[18] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, 2020. 5 +[19] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, , and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In ECCV, 2016. 7 +[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In ICMR, 2019. 6 +[21] Rui Su, Qian Yu., and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In ICCV, 2021. 2, 3, 5, 7 +[22] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 2, 3, 7 +[23] Reuben Tan, Bryan A. Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos, 2021. 1, 2, 6, 7 +[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 2, 3, 6, 7, 8 +[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4 +[26] Zhenzhi Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2, 3 +[27] Masataka Yamaguchi, Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Spatio-temporal person retrieval via natural language queries. In ICCV, 2017. 7 +[28] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8 + +[29] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 3 +[30] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 3, 7 +[31] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5 +[32] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In CVPR, 2020. 1, 2, 3, 6, 7, 8 +[33] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In *IJCAI*, 2021. 2, 3, 7 +[34] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 6 +[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 4, 5 \ No newline at end of file diff --git a/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/images.zip b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..85ad934d7457aadfecb2fce902514ceb4c2306a4 --- /dev/null +++ b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af95db23fb675a7bb979ffe192ceebcfd608d77bf42eb9c86b5d22d033099cb4 +size 600794 diff --git a/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/layout.json b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fec5c5fad46831545506567111a2155002c75763 --- /dev/null +++ b/2024/VideoGrounding-DINO_ Towards Open-Vocabulary Spatio-Temporal Video Grounding/layout.json @@ -0,0 +1,8854 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "text", + "content": "VideoGrounding-DINO: Towards Open-Vocabulary Spatio-Temporal Video Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 160, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 160, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 160, + 455, + 190 + ], + "type": "text", + "content": "Syed Talal Wasim1 Muzammal Naseer1 \nSalman Khan1,2 Ming-Hsuan Yang3,4 Fahad Shahbaz Khan1,5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "spans": [ + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "text", + "content": "Mohamed bin Zayed University of AI " + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "text", + "content": "Australian National University " + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "text", + "content": "University of California, Merced " + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "text", + "content": "Google Research " + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 94, + 194, + 484, + 224 + ], + "type": "text", + "content": "Linköping University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "type": "text", + "content": "Video grounding aims to localize a spatio-temporal section in a video corresponding to an input text query. This paper addresses a critical limitation in current video grounding methodologies by introducing an Open-Vocabulary Spatio-Temporal Video Grounding task. Unlike prevalent closed-set approaches that struggle with open-vocabulary scenarios due to limited training data and predefined vocabularies, our model leverages pre-trained representations from foundational spatial grounding models. This empowers it to effectively bridge the semantic gap between natural language and diverse visual content, achieving strong performance in closed-set and open-vocabulary settings. Our contributions include a novel spatio-temporal video grounding model, surpassing state-of-the-art results in closed-set evaluations on multiple datasets and demonstrating superior performance in open-vocabulary scenarios. Notably, the proposed model outperforms state-of-the-art methods in closed-set settings on VidSTG (Declarative and Interrogative) and HC-STVG (V1 and V2) datasets. Furthermore, in open-vocabulary evaluations on HC-STVG V1 and YouCook-Interactions, our model surpasses the recent best-performing models by " + }, + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "type": "inline_equation", + "content": "4.88\\mathrm{m\\_vIoU}" + }, + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 47, + 276, + 290, + 576 + ], + "type": "text", + "content": " accuracy, demonstrating its efficacy in handling diverse linguistic and visual concepts for improved video understanding. Our codes will be publicly released." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "Spatio-temporal video grounding is pivotal in linking visual content with natural language descriptions, thus facilitating semantics interpretation within visual data. Prevailing approaches in video grounding such as TubeDETR [28], STCAT [9], and STVGFormer [11] focus mainly on supervised closed-set settings, where models are trained on specific datasets [24, 32] with predefined vocabulary and meticulously annotated data. While these current state-of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 253, + 541, + 427 + ], + "blocks": [ + { + "bbox": [ + 310, + 253, + 541, + 427 + ], + "lines": [ + { + "bbox": [ + 310, + 253, + 541, + 427 + ], + "spans": [ + { + "bbox": [ + 310, + 253, + 541, + 427 + ], + "type": "image", + "image_path": "83b2e303b48c452304cd569f4a882c31b066146be3d773ffec2a4b4823ab5291.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 438, + 547, + 504 + ], + "lines": [ + { + "bbox": [ + 305, + 438, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 547, + 504 + ], + "type": "text", + "content": "Figure 1. Performance comparison on conventional closed-set and open-vocabulary settings for the video grounding task. We compare our approach with TubeDETR [28] and STCAT [9] in supervised setting for VidSTG [32] declarative/interrogative and HC-STVG V1 [24], along with open-vocabulary evaluation on HC-STVG V1 and YouCook-Interactions [23] datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 531, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 546, + 604 + ], + "type": "text", + "content": "the-art methods excel in closed-set settings on datasets like VidSTG [32] and HC-STVG [24], their limited generalization beyond training dataset distributions poses a significant challenge. The relatively small scale and restricted sample variety in existing video datasets hinder models from adapting to unseen scenarios effectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Motivated by the inherent limitation of supervised closed-set approaches in terms of their restricted vocabulary, this paper investigates Open-Vocabulary Spatio-Temporal Video Grounding. Unlike conventional methodologies, this paradigm addresses challenges posed by the unrestricted diversity of language and visual concepts in the wild. The goal is to train on a set of base categories and generalize to unseen objects/actions based on the open-vocabulary nature of backbone models. This paper ex" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18909" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "plores the challenges and opportunities inherent in open-vocabulary video grounding, laying the groundwork for more robust and versatile video understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "text", + "content": "However, training an effective open-vocabulary video grounding model would require a large enough dataset with a rich set of natural language expressions and corresponding spatio-temporal localizations. Such an extensive dataset can allow the model to learn generalized visual and textual representations and handle out-of-distribution samples. However, video grounding datasets [24, 32] are quite limited in scale, e.g., VidSTG has only " + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "inline_equation", + "content": "5.4\\mathrm{k}" + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "text", + "content": " training videos with " + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "inline_equation", + "content": "80.6\\mathrm{k}" + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "text", + "content": " distinct sentences. In contrast, an image grounding model GLIP [10] is trained with " + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "inline_equation", + "content": "\\sim 26.5\\mathrm{M}" + }, + { + "bbox": [ + 46, + 112, + 288, + 304 + ], + "type": "text", + "content": " image-text pairs. Therefore, our research explores the following fundamental question: How spatio-temporal video grounding models can achieve robust performance in both Closed-Set and Open-Vocabulary scenarios without requiring large-scale video annotations, ensuring effective generalization beyond training datasets?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 308, + 288, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 513 + ], + "type": "text", + "content": "In addressing this question, we find inspiration in the accomplishments of foundational models specializing in spatial grounding [4, 8, 12, 14, 29]. These models undergo training on an extensive corpus of image-text data, enabling effective generalization to samples from a given target distribution. We aim to harness this pretrained representation to enhance our video grounding model. Our proposed solution is a spatio-temporal video grounding model adopting a DETR-like architecture enhanced by temporal aggregation modules. The spatial modules are initialized using the pretrained representation of a foundational image model [12]. Meanwhile, the image and text feature extractors remain frozen while the video-specific spatio-temporal adaptations are modeled via learnable adapter blocks. This approach is designed to preserve the nuanced representation of the foundational model, enhancing our model's ability to generalize effectively to novel samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 516, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 287, + 564 + ], + "type": "text", + "content": "A summary of our closed-set and open-vocabulary results is shown in Fig. 1, where the proposed approach excels in both settings by a clear margin. Our major contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 568, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 568, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 640 + ], + "type": "text", + "content": "- For the first time, we evaluate spatio-temporal video grounding models in an open-vocabulary setting on HC-STVG V1 [24] and YouCook-Interactions [23] benchmarks in a zero-shot manner. We outperform state-of-the-art methods TubeDETR [28] and STCAT [9] by 4.26 m_vIoU and " + }, + { + "bbox": [ + 47, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 47, + 568, + 287, + 640 + ], + "type": "text", + "content": " accuracy, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 640, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 288, + 712 + ], + "type": "text", + "content": "- By combining the strengths of spatial grounding models with complementary video-specific adapters, our approach consistently outperforms the previous state-of-the-art in closed-set setting on four benchmarks, i.e., VidSTG (Declarative) [32], VidSTG (Interrogative) [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 390 + ], + "type": "text", + "content": "Spatial Grounding Foundation Models: Recent literature introduces notable spatial grounding models. GLIP [10] unifies object detection and phrase grounding through a language-image pre-training model, leveraging extensive image-text pairs for semantic-rich representations. Grounding DINO [12] integrates language with a transformer-based detector to achieve an open-set detector, excelling in benchmarks like COCO and ODinW. Kosmos-1 [8] and Kosmos-2 [14] contribute Multimodal Large Language Models (MLLMs) with capabilities such as zero-shot and few-shot learning, language understanding, and multimodal tasks. Kosmos-2 [14] specifically integrates grounding into downstream applications, introducing GrIT, a large-scale dataset of Grounded Image-Text pairs. Shikra [4] addresses referential ability in MLLMs by handling spatial coordinates in inputs and outputs, showcasing promising performance in various vision-language tasks. Ferret [29] unifies referring and grounding in the LLM paradigm, achieving superior performance in classical referring and grounding tasks, excelling in region-based multimodal chatting and image description. Recently, GLaMM [16] allows pixel-level grounded conversations with an LLM, showcasing generalizability to several captioning and referring segmentation tasks. However, spatial methods cannot work for grounding objects in videos, a gap addressed by this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "Spatio-Temporal Video Grounding: Several methods tackle the challenge of localizing objects in untrimmed videos based on query sentences. STVGBert [21] presents a one-stage visual-linguistic transformer for simultaneous spatial and temporal localization. TubeDETR [28] introduces a transformer-based architecture to model temporal, spatial, and multi-modal interactions efficiently. Augmented 2D-TAN [22] adopts a two-stage approach, enhancing the 2D-TAN with a temporal context-aware Bi-LSTM Aggregation Module. OMRN [33] addresses the challenge of unaligned data and multi-form sentences in spatiotemporal video grounding, proposing an object-aware multi-branch relation network for effective relation discovery. MMN [26] introduces a Mutual Matching Network as a metric-learning framework for temporal grounding, achieving competitive performance. STCAT [9] is an end-to-end one-stage framework addressing feature alignment and prediction inconsistency. Finally, STVGFformer [11] proposes an effective framework with static and dynamic branches for cross-modal understanding. While the above methods advance video grounding, their generalization to out-of-distribution and open-vocabulary samples is limited due to constrained video datasets [24, 32]. To address this issue, we utilize the generalized representation of spatial grounding foundation models [4, 8, 10, 12, 14, 29] trained on a large corpus of image-text data and can perform well on both closed-set and open-vocabulary evaluations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18910" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 286, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 286, + 199 + ], + "type": "text", + "content": "As discussed above, the current state-of-the-art spatiotemporal video grounding methods [9, 11, 21, 22, 24, 26, 28, 30, 32, 33] primarily evaluate in a supervised setting on the VidSTG [32] and HC-STVG [24] datasets. However, these methods lack the multimodal spatio-temporal understanding required to perform well on out-of-distribution samples [2]. Therefore, this work aims to achieve improved open-vocabulary performance while maintaining strong closed-set video-grounding performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 199, + 286, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 286, + 317 + ], + "type": "text", + "content": "We take inspiration from recent foundation models for spatial grounding [4, 8, 10, 12, 14, 29]. These models are trained on a large corpus of visual-textual data and hence, generalize well to unseen samples. We aim to leverage the strong generalization capabilities of such foundation models to achieve strong open-set spatio-temporal video grounding performance. Our proposed spatio-temporal video grounding method uses DETR-like [1] design, with temporal aggregation and adaptation modules for learning video-specific representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 286, + 377 + ], + "type": "text", + "content": "Below, we explain our proposed methodology. We formally define the spatio-temporal video grounding problem in Sec. 3.1. We then explain our architecture details in Sec. 3.2. We finally explain the loss formulation used to train the model and model initialization in Sec. 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 384, + 159, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 159, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 159, + 396 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 286, + 509 + ], + "type": "text", + "content": "The spatio-temporal video grounding task involves localizing and recognizing objects and actions in a video sequence by integrating spatial and temporal information. In contrast to spatial grounding, which focuses on recognizing and localizing objects or actions within individual frames, spatiotemporal grounding extends this concept to include the temporal dimension. This means understanding where objects or actions are in each frame and how they evolve and move over time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": "Consider a video " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " frames, " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " spatial resolution, and " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " channels, respectively, along with a text prompt " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": ". The spatial grounding problem can be defined as the localization of one or more objects associated with the prompt " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " in a frame " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "V_{t}, t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " using a bounding box " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "B_{i}^{t} = (x_{i}^{t}, y_{i}^{t}, w_{i}^{t}, h_{i}^{t})" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "y_{i}^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " are the coordinates of the top-left corner, " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "w_{i}^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "h_{i}^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " are the width and height of the bounding box, " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " is the object number for the frame " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": ". The temporal grounding problem, on the other hand, involves understanding how objects or actions evolve over time. It aims at localizing the temporal interval " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "(t_{s}, t_{e})" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " where the specific action/interaction happens in the entire temporal duration, where interval " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "(t_{s}, t_{e})" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " indicates the start and end frame of the object occurrence within the total frames " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "1 \\leq t_{s} < t_{e} \\leq T" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": "). Hence, the spatio-temporal grounding problem for object " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " associated with prompt " + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 509, + 286, + 713 + ], + "type": "text", + "content": " can be summarized as a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "set of spatio-temporal coordinates associated with the subset of frames where the object exists: " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "(x_i^t,y_i^t,w_i^t,h_i^t,t)" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "t\\in \\{t_s,\\dots,t_e\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": ". The interval " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "(t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " and is a subset of the total frames " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "T" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 130, + 492, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 492, + 143 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 492, + 143 + ], + "type": "text", + "content": "3.2. Spatio-Temporal Video Grounding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 149, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 281 + ], + "type": "text", + "content": "Here, we explain our video grounding model in Fig. 2. As discussed earlier, we aim to design a spatio-temporal video grounding model that can perform well in closed-set and open-vocabulary settings. Strong open-vocabulary performance requires learning a rich visual/textual representation, which in turn requires a large amount of training data. Unfortunately, spatio-temporal video grounding datasets are quite limited in scale [24, 32], resulting in current video grounding methods failing in generalizing well to out-of-distribution samples because they lack the requisite strong visual/textual representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "type": "text", + "content": "To solve this problem, our approach takes inspiration from recent spatial grounding methods [4, 8, 12, 14, 29], which have strong open-vocabulary performance thanks to the large image-text corpus they are trained on. We can utilize the generalized representations of these models to enrich the weaker representation of video-grounding approaches obtained from the limited number of training samples. Our approach aims to leverage the strong pretrained representations of spatial grounding methods to achieve strong closed-set supervised and open-vocabulary video grounding performance. Our spatio-temporal video grounding approach is based on the state-of-the-art DETR-based [1] object detection framework DINO [31] and also borrows concepts of image-text alignment and grounding from Grounded Language-Image Pre-training (GLIP) [10] and Grounding DINO [12]. We extract initial features from backbone vision and text encoders " + }, + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 304, + 282, + 545, + 604 + ], + "type": "text", + "content": ". Following that, we model inter-frame and intra-frame features and learn cross-modal visual/textual relations in the Cross-Modality Spatio-Temporal Encoder (Sec. 3.2.1). The result enriched cross-modal features are used to initialize queries for each frame (Sec. 3.2.2). These queries are then decoded to predict the bounding boxes per frame and the temporal grounding start/end frame by aggregating information across spatial/temporal dimensions and injecting information through cross-attention from enriched visual/textual context (Sec. 3.2.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Given the video " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and text prompt " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " as described above, we obtain per-frame features " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{v}^{0}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and text features " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{p}^{0}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " from vision and text encoders, " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", respectively. The vision encoder is based on a Swin Transformer [13], and the text-encoder is defined as a BERT [6] model. Like other DETR-based detectors [31, 35], image features are extracted from different vision-encoder blocks at multiple scales. The features are then passed to the Cross-Modality Spatio-Temporal Encoder." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18911" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 547, + 271 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 547, + 271 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 547, + 271 + ], + "type": "image", + "image_path": "aee7204402aeb18acb3ae3d1e5eb4d4b0cbaff7cdc169590e9d1cd00e392b2f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "type": "text", + "content": "Figure 2. Overall architecture: We present our video grounding architecture. It consists of vision and text encoders that produce visual and textual features. A cross-modality spatio-temporal encoder which fuses information across spatial/temporal dimensions and visual/textual modalities. A language guided query selection module to initialize cross-modal queries. A cross-modality spatio-temporal decoder to decoder queries while fusing information from visual/textual features. And finally two prediction heads to predict the bounding boxes per frame and the temporal tube. Modules with " + }, + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "type": "inline_equation", + "content": "(\\bullet)" + }, + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "type": "text", + "content": " are trainable and those with " + }, + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "type": "inline_equation", + "content": "(\\text{串})" + }, + { + "bbox": [ + 46, + 279, + 547, + 335 + ], + "type": "text", + "content": " are frozen." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 356, + 259, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 259, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 259, + 369 + ], + "type": "text", + "content": "3.2.1 Cross-Modality Spatio-Temporal Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "type": "text", + "content": "The initial vision and text features " + }, + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "type": "inline_equation", + "content": "F_{v}^{0}" + }, + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "type": "inline_equation", + "content": "F_{p}^{0}" + }, + { + "bbox": [ + 46, + 374, + 287, + 445 + ], + "type": "text", + "content": " neither contain any cross-modal information nor model the temporal relationship across frames. Therefore, we further encode the initial features through the Cross-Modality Spatio-Temporal Encoder to model the temporal information across frames and learn cross-modal features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "text", + "content": "Each layer of the " + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "text", + "content": " layer encoder first applies a Multi-Head Self-Attention (MHSA) [25] to the visual features " + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "text", + "content": " along the temporal dimension, followed by a Deformable Attention (DA) [35] along the spatial dimension. This is done to model relations within frames and temporally across frames. Similarly, we apply a MHSA on the text features " + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 46, + 447, + 288, + 533 + ], + "type": "text", + "content": ". This is illustrated in Eq. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 540, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 540, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 75, + 540, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {v} ^ {m \\prime} = \\mathrm {D A} _ {s p a t i a l} ^ {m} (\\mathrm {M H S A} _ {t e m p o r a l} ^ {m} (F _ {v} ^ {m - 1})), \\\\ F _ {p} ^ {m \\prime} = \\mathrm {M H S A} _ {p} ^ {m} (F _ {p} ^ {m - 1}), \\end{array} \\tag {1}", + "image_path": "d82d9b899a299bc4d5eecf6609e67ee89b268d278c1cc35d1c32ca09e352a89a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{v}^{m - 1}" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{p}^{m - 1}" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " are visual and textual input features to layer " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{v}^{m'}" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{p}^{m'}" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " are the intermediate visual and textual feature representation, and " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{DA}_{\\text{spatial}}^m" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{MHSA}_{\\text{temporal}}^m" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{MHSA}_p^m" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " are the spatial-deformable, temporal and textual attentions at layer " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "m \\in \\{1,\\dots,M\\}" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": ", respectively. Following initial spatial, temporal, and textual attentions, we fuse features across the visual and textual modalities, as done in GLIP [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "More specifically, we calculate the joint visual-textual attention, " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Attn}_{\\mathrm{joint}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": ", using projected intermediate features " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{v}^{m'}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{p}^{m'}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": ". This attention is then used alongside the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "type": "text", + "content": "intermediate features " + }, + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "type": "inline_equation", + "content": "F_{v}^{m'}" + }, + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "type": "inline_equation", + "content": "F_{p}^{m'}" + }, + { + "bbox": [ + 304, + 355, + 546, + 392 + ], + "type": "text", + "content": ", to calculate the image-to-text and text-to-image cross-attentions as shown in Eq. 2 and Eq. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 400, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 400, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 329, + 400, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\mathrm {A t t n} _ {\\mathrm {j o i n t}} ^ {\\mathrm {m}} = \\left(\\frac {\\operatorname {p r o j} _ {q , v} ^ {m} \\left(F _ {v} ^ {m \\prime}\\right) \\operatorname {p r o j} _ {q , p} ^ {m} \\left(F _ {p} ^ {m \\prime}\\right) ^ {T}}{\\sqrt {d ^ {k}}}\\right), \\tag {2}", + "image_path": "88266726d2ba7db7843f77bf15d02cf8ef57857dc46d5bafe73a964e3f4a7844.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "inline_equation", + "content": "proj_{q,v}^{m}" + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "inline_equation", + "content": "proj_{q,p}^{m}" + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "text", + "content": " are the query projections for the visual and textual features, respectively, at layer " + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 440, + 545, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 545, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 545, + 494 + ], + "type": "interline_equation", + "content": "F _ {v} ^ {m} = \\operatorname {F F N} _ {v} ^ {m} \\left(\\operatorname {s o f t m a x} \\left(\\operatorname {A t t n} _ {\\text {j o i n t}} ^ {\\mathrm {m}}\\right) p r o j _ {p} ^ {m} \\left(F _ {p} ^ {m \\prime}\\right)\\right), \\tag {3}", + "image_path": "cd0e01f46b9a690d9f2838265745de7f0f96e16671408a5ae86262aac3df74b7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 491, + 526, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 526, + 506 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 526, + 506 + ], + "type": "interline_equation", + "content": "F _ {p} ^ {m} = \\mathrm {F F N} _ {p} ^ {m} (\\mathrm {s o f t m a x} (\\mathrm {A t t n} _ {\\mathrm {j o i n t}} ^ {\\mathrm {m} _ {T}}) p r o j _ {v} ^ {m} (F _ {v} ^ {m /}))", + "image_path": "08fa655fcae5099b68f3667df0b447cfdc76acc917e4595655ee4b9e87f349dd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "F_{v}^{m}" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "F_{p}^{m}" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " are the final output features at layer " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " of the encoder, " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_v^m" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_v^m" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " are the visual and textual Feed Forward Networks (FFN) and " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "proj_0^m" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "proj_0^m" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " are the linear layers to project the visual and textual features. The final encoded features at layer " + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "inline_equation", + "content": "m = M" + }, + { + "bbox": [ + 304, + 512, + 547, + 586 + ], + "type": "text", + "content": " are then utilized to initialize cross-modal queries." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 599, + 487, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 487, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 487, + 612 + ], + "type": "text", + "content": "3.2.2 Language-Guided Query Selection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "This module is designed to select features more relevant to the input text as decoder queries for effective language-vision fusion. We combine a DETR/DINO [1, 31] style queries with a sinusoidal temporal positional encoding to the positional part of the queries. The sinusoidal positional encoding added to the positional part of the queries adds important contextual information regarding the sequence of frames, allowing for improved temporal correlation and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18912" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "grounding [28]. The query selection module takes the encoder's visual and textual features as input and outputs num_query indices that correspond to the most relevant features for object detection per frame, " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\{Q_t^0\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "Q_t^0" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " are the initial queries for the frame " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". The module initializes the decoder queries using a combination of the selected indices and dynamic anchor boxes. The content part of the queries is set to be learnable during training, while the positional part is computed using the dynamic anchor boxes initialized using the encoder outputs. We also add a sinusoidal temporal positional encoding to the positional part of the queries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 232, + 258, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 258, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 258, + 243 + ], + "type": "text", + "content": "3.2.3 Cross-Modality Spatio-Temporal Decoder" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 382 + ], + "type": "text", + "content": "To decode the above queries into bounding box locations and temporal start/end tubes, we need to transform them into an output embedding, which can then be fed into prediction heads. The decoder allows for the queries to interact globally with others within a frame and across frames while utilizing the entire visual and textual features as context. Formally, the queries produced earlier are fed into a " + }, + { + "bbox": [ + 46, + 251, + 288, + 382 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 251, + 288, + 382 + ], + "type": "text", + "content": " layer decoder. Each layer starts with a temporal self-attention, a spatial self-attention followed by a visual cross-attention and textual cross-attention, and finally an FFN. This is represented in Eq. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 391, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 391, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 70, + 391, + 287, + 411 + ], + "type": "interline_equation", + "content": "Q _ {t} ^ {n ^ {\\prime}} = \\mathrm {M H S A} _ {\\text {s p a t i a l}} ^ {n} (\\mathrm {M H S A} _ {\\text {t e m p o r a l}} ^ {n} \\left(Q _ {t} ^ {n - 1}\\right)), \\tag {4}", + "image_path": "84d0bb714cad892df7289efff8b48002589703a61dac2c65464209edc9eed76c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 407, + 253, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 407, + 253, + 423 + ], + "spans": [ + { + "bbox": [ + 75, + 407, + 253, + 423 + ], + "type": "interline_equation", + "content": "Q _ {t} ^ {n} = \\mathrm {F F N} ^ {n} (\\mathrm {C A} _ {p} ^ {n} (\\mathrm {C A} _ {v} ^ {n} (Q _ {t} ^ {n \\prime}, F _ {v} ^ {M}), F _ {p} ^ {M})),", + "image_path": "d603cef229fadef913998011d1602bba4cd9cb31e316edc580e4921dbff03449.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "Q_{t}^{n - 1}" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": " are the input queries at layer " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "n\\in \\{1,\\dots,N\\}" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "Q_{t}^{n^{\\prime}}" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": " are the intermediate queries after spatial and temporal attention at layer " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "Q_{t}^{n^{\\prime}}" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": " are the output queries at layer " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{CA}_v^n" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{CA}_p^n" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": " are the visual and textual cross-attention at layer " + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 432, + 287, + 505 + ], + "type": "text", + "content": ". The cross-attentions are further elaborated in Eq. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 512, + 304, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 512, + 304, + 542 + ], + "spans": [ + { + "bbox": [ + 50, + 512, + 304, + 542 + ], + "type": "interline_equation", + "content": "\\mathrm {C A} _ {v} ^ {n} \\left(Q _ {t} ^ {n \\prime}, F _ {v} ^ {M}\\right) = \\left(\\frac {\\operatorname {p r o j} _ {q , v} ^ {n} \\left(Q _ {t} ^ {n \\prime}\\right) \\operatorname {p r o j} _ {k , v} ^ {n} \\left(F _ {v} ^ {M}\\right) ^ {T}}{\\sqrt {d ^ {k}}} \\operatorname {p r o j} _ {v} ^ {n} \\left(F _ {v} ^ {M}\\right) ^ {T}\\right),", + "image_path": "1288f9c0a815b15f50a5fa0cbb5e2042681de0ef5011a41bbf19d4c96433ae52.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 544, + 304, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 304, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 304, + 581 + ], + "type": "interline_equation", + "content": "\\mathrm {C A} _ {p} ^ {n} \\left(\\mathrm {C A} _ {v} ^ {n}, F _ {p} ^ {M}\\right) = \\left(\\frac {\\operatorname {p r o j} _ {q , p} ^ {n} \\left(\\mathrm {C A} _ {v} ^ {n}\\right) \\operatorname {p r o j} _ {k , p} ^ {n} \\left(F _ {p} ^ {M}\\right) ^ {T}}{\\sqrt {d ^ {k}}} \\operatorname {p r o j} _ {p} ^ {n} \\left(F _ {p} ^ {M}\\right) ^ {T}\\right) \\tag {5}", + "image_path": "07d4ba2e072794657014334d16bd1a544bc3716f4de44134df064ae13501679c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "proj_{q,v}^n, proj_{k,v}^n" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "proj_v^n" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": " are the visual query, key and values projection for layer " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "proj_{q,p}^n, proj_{k,p}^n" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "proj_p^n" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": " are the textual query, key and value projections. The final queries from the decoder at layer " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\{Q_t^N\\}_{t=1}^T" + }, + { + "bbox": [ + 47, + 582, + 287, + 643 + ], + "type": "text", + "content": ", are then used for prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 153, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 153, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 153, + 670 + ], + "type": "text", + "content": "3.2.4 Prediction Heads" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "The decoder outputs refined queries per frame " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{Q_t^N\\}_{t = 1}^T" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " We follow the standard DETR-like bounding box regression head implemented as a Multi-Layer Perceptron (MLP)," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "which predicts bounding boxes " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "B_{i}^{t} = (x_{i}^{t},y_{i}^{t},w_{i}^{t},h_{i}^{t})" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ", per frame. To predict the temporal interval " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "(t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ", we add a temporal grounding head, implemented as an MLP, alongside the bounding box regression head, similar to existing works like [9, 28]. The new head predicts the probabilities of the start " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\tau_{s}\\in [0,1]^{T}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " and ends " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\tau_{e}\\in [0,1]^{T}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " of the interval. During inference, the start and end interval " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "(t_s,t_e)|\\{1\\leq t_s < t_e\\leq T\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " is computed by taking the maximum of the joint distribution of " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "(\\tau_s,\\tau_e)" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ". Any invalid combinations with " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "t_e\\leq t_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " are masked out." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 201, + 394, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 394, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 394, + 213 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": "To leverage the generalized pre-trained representation from spatial-grounding foundation models, we initialize all spatial modules and cross attentions from the Grounding DINO [12] spatial grounding model. To preserve this generalized representation while ensuring effective modeling of the downstream task, we freeze the Vision and Text Encoders " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " and fine-tune the remaining components." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": "During training, the model receives a batch of videos " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": " with text prompt " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": ". The ground-truth annotation contains the bounding box sequence " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\{B_i^t\\}_{i = t_s}^{t_e}" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": ", and the corresponding start and end timestamps " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "(t_{s},t_{e})" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": ". For spatial grounding, we follow the standard loss formulation used in DETR-like [1, 31, 35], namely the " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": " loss, " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{L_1}" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": ", and the Generalized Intersection over Union (GIoU) [17] loss, " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GIoU}" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": ". Formally, the spatial grounding loss, " + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{spatial}" + }, + { + "bbox": [ + 304, + 304, + 545, + 412 + ], + "type": "text", + "content": " is defined in Eq. 6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 422, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 545, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a t i a l}} = \\lambda_ {L _ {1}} \\mathcal {L} _ {L _ {1}} (\\hat {B}, B) + \\lambda_ {G I o U} \\mathcal {L} _ {G I o U} (\\hat {B}, B). \\tag {6}", + "image_path": "b74a718625d1b8c7325a97479736b2bb36adb7c301446e5c1308a924297b143f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 447, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 495 + ], + "type": "text", + "content": "For temporal grounding, we follow [18, 21, 28] and generate two 1-dimensional gaussian heatmaps " + }, + { + "bbox": [ + 304, + 447, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\pi_s,\\pi_e\\in \\mathcal{R}^T" + }, + { + "bbox": [ + 304, + 447, + 545, + 495 + ], + "type": "text", + "content": " for the starting and ending positions. The temporal grounding loss is therefore defined in Eq. 7 as," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 506, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 506, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 337, + 506, + 545, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p o r a l}} = \\mathcal {L} _ {K L} ^ {s} \\left(\\hat {\\pi} _ {s}, \\pi_ {s}\\right) + \\mathcal {L} _ {K L} ^ {e} \\left(\\hat {\\pi} _ {e}, \\pi_ {e}\\right), \\tag {7}", + "image_path": "f490fba380022bda587cceb9dab1239e2ed16186f7808ba18c062a709a39e52d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}^s" + }, + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}^{e}" + }, + { + "bbox": [ + 304, + 529, + 544, + 553 + ], + "type": "text", + "content": " are the KL divergence losses for the start and end distributions, respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 554, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 625 + ], + "type": "text", + "content": "Note that the model outputs the bounding boxes and starting/ending distributions during inference. We determine the temporal grounding segment, " + }, + { + "bbox": [ + 304, + 554, + 545, + 625 + ], + "type": "inline_equation", + "content": "(t_s,t_e)" + }, + { + "bbox": [ + 304, + 554, + 545, + 625 + ], + "type": "text", + "content": ", by taking the segment with the maximal joint start and end probability. Then, we consider the bounding boxes only within that tube for spatial grounding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 639, + 359, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 359, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 359, + 651 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 658, + 488, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 488, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 488, + 671 + ], + "type": "text", + "content": "4.1. Experimental Setup and Protocols" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Below, we first briefly explain the implementation details (Sec. 4.1.1), followed by evaluation settings (Sec. 4.1.2), and datasets (Sec. 4.1.3) used in our work." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18913" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 91, + 514, + 168 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "text", + "content": "Table 1. Performance comparisons of the state-of-the-art on HC-STVG V1 [24] and YouCook-Interactions [23] in open-vocabulary setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 91, + 514, + 168 + ], + "lines": [ + { + "bbox": [ + 75, + 91, + 514, + 168 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 514, + 168 + ], + "type": "table", + "html": "
MethodPre-trainingHC-STVG V1YouCook-Interactions
m_vIoUvIoU@0.3vIoU@0.5Accuracy
TubeDETR (CVPR'22) [28]VidSTG16.8422.329.2251.63
STCAT (NeurIPS'22) [9]VidSTG22.5832.1420.8355.90
VideoGrounding-DINOVidSTG27.4640.1329.9257.73
", + "image_path": "dbf4084e7df5bdaf3f1fd5e8e496643b73381c5c76fc149bbd2131079fe88883.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 180, + 180, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 180, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 180, + 192 + ], + "type": "text", + "content": "4.1.1 Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "text", + "content": "As discussed in the methodology (Sec. 3), we initialize the spatial modules in our model from the Grounding DINO [12] spatial grounding model and keep the vision and text encoders frozen. Our prediction heads for both spatial and temporal predictions are set to be 3-layer Multi-Layer Perceptrons (MLPs). We sample 128 frames during training and inference, resized to a resolution of 448 on the shorter side. We set both " + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "text", + "content": " to 6, and train the model with a batch size of 8 and learning rate of " + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "text", + "content": ", and weight decay if " + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 199, + 288, + 331 + ], + "type": "text", + "content": ". The number of epochs for VidSTG is set to 10, and for HC-STVG V1/V2 is set to 90." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 347, + 163, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 163, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 163, + 360 + ], + "type": "text", + "content": "4.1.2 Evaluation Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 366, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 287, + 390 + ], + "type": "text", + "content": "We evaluate our video grounding model in two settings, Open-Vocabulary and Closed-Set Supervised." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 391, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 521 + ], + "type": "text", + "content": "Open-Vocabulary Evaluation: In the open-vocabulary setting we train our model on the VidSTG [32] dataset and then evaluate on two different datasets, HC-STVG V1 [24] and YouCook-Interactions [23] to understand how well the model generalizes to new distributions. The reason for choosing these two datasets is that the former provides a relatively minor distribution shift given the similar perspective/objects in the videos compared to the training dataset VidSTG. In contrast, the latter provides a major distribution shift with changes in perspective and annotated objects/interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": "Closed-Set Supervised Evaluation: In the supervised evaluation setting, we train on the training set and evaluate each dataset's respective validation/testing set. This evaluation is conducted for three majorly used datasets in spatio-temporal video grounding, namely VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 117, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 117, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 117, + 621 + ], + "type": "text", + "content": "4.1.3 Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": "We evaluate our approach and compare against the state-of-the-art in two settings: Open-Vocabulary and Closed-Set Supervised, across a total of four grounding datasets, namely: VidSTG [32], HCSTVG V1 [24], HCSTVG V2 [24], and YouCook-Interactions [23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "VidSTG: The VidSTG [32] dataset is derived from the VidOR [20] dataset, incorporating object relation annotations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "content": "It includes 99,943 video-text pairs, encompassing 44,808 declarative sentence queries and 55,135 interrogative sentence queries. The training, validation, and test sets consist of 80,684, 8,956, and 10,303 sentences and 5,436, 602, and 732 videos, respectively. VidSTG's text queries are confined to describing pre-defined object/relation categories in VidOR [20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "content": "HC-STVG V1/V2: The HC-STVG datasets are sourced from movie scenes, each video clip spanning approximately 20 seconds. These datasets pose challenges in spatiotemporal grounding due to video clips featuring multiple individuals engaged in similar actions. HC-STVG V1 comprises 4,500 training and 1,160 testing video-text pairs. HC-STVG V2 expands HC-STVG V1, enhancing annotation quality with 10,131, 2,000, and 4,413 samples for training, validation, and testing, respectively. As HC-STVG V2's test set annotations are unavailable publicly, results are reported on the validation set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "type": "text", + "content": "YouCook-Interactions: The YouCook-Interactions [23] dataset serves as an expansion of the YouCook2 [34] dataset focused on cooking instructions. This extension includes bounding boxes for 6,000 carefully chosen frames, typically encompassing the hand and the tool specified in the corresponding sentence-level annotations. Our assessment revolves around examining models' spatial grounding capabilities using this dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 501, + 492, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 492, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 492, + 514 + ], + "type": "text", + "content": "4.2. Experimental Results and Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 579 + ], + "type": "text", + "content": "In this section, we present our results across the evaluation mentioned above settings (Sec. 4.1.2) and datasets (Sec. 4.1.3). We start with the closed-set evaluation in Sec. 4.2.2, followed by the open-vocabulary evaluation in Sec. 4.2.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 597, + 463, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 463, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 463, + 609 + ], + "type": "text", + "content": "4.2.1 Open-Vocabulary Evaluation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "content": "For open-vocabulary evaluation, we train on VidSTG [32] and present results HC-STVG V1 [24] and YouCook-Interactions [23]. The results are reported jointly in Tab. 1. Results on HC-STVG V1: We report open-vocabulary evaluation on m_vIoU, vIoU@0.3, and vIoU@0.5. We achieve state-of-the-art performance over both TubeDETR [28] and STCAT [9]. We attribute this strong performance to our design, which leverages the strong pre" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18914" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 91, + 547, + 316 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 509, + 83 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 509, + 83 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 509, + 83 + ], + "type": "text", + "content": "Table 2. Performance comparisons of the state-of-the-art on the VidSTG [32] test set in closed-set supervised setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 91, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 49, + 91, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 547, + 316 + ], + "type": "table", + "html": "
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3
Factorized:
GroundRe (ECCV'16) [19]+TALL (ICCV'17) [7]9.7811.044.099.3211.39
STPR (ICCV'17) [27]+TALL (ICCV'17) [7]34.6310.4012.384.2733.739.9811.74
WSSTG (arXiv'19) [5]+TALL (ICCV'17) [7]11.3614.635.9110.6513.90
GroundRe (ECCV'16) [19]+L-Net (AAAI'19) [3]11.8915.325.4511.0514.28
STPR (ICCV'17) [27]+L-Net (AAAI'19) [3]40.8612.9316.275.6839.7911.9414.73
WSSTG (arXiv'19) [5]+L-Net (AAAI'19) [3]14.4518.007.8913.3617.39
Two-Stage:
STGRN (CVPR'20) [32]48.4719.7525.7714.6046.9818.3221.10
STGVT (TCSVT'21) [24]-21.6229.8018.94---
OMRN (IJCAI'21) [33]50.7323.1132.6116.4249.1920.6328.35
One-Stage:
STVGBert (ICCV'21) [21]-23.9730.9118.39-22.5125.97
TubeDETR (CVPR'22) [28]48.1030.4042.5028.2046.9025.7035.70
STCAT (NeurIPS'22) [9]50.8233.1446.2032.5849.6728.2239.24
STVGFormer (CVPR'23) [11]-33.7047.2032.80-28.5039.90
VideoGrounding-DINO51.9734.6748.1133.9650.8329.8941.03
", + "image_path": "9d4fcfa28fb75831d518a55b06c2bb94d20ee1f9ba269e9f844bd18c8295dd4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 357, + 287, + 442 + ], + "blocks": [ + { + "bbox": [ + 47, + 327, + 288, + 349 + ], + "lines": [ + { + "bbox": [ + 47, + 327, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 288, + 349 + ], + "type": "text", + "content": "Table 3. Performance comparisons of the state-of-the-art on the HC-STVG V1 [24] test set in closed-set supervised setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 357, + 287, + 442 + ], + "lines": [ + { + "bbox": [ + 50, + 357, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 50, + 357, + 287, + 442 + ], + "type": "table", + "html": "
Methodsm_vIoUvIoU@0.3vIoU@0.5
STGVT (TCSVT'21) [24]18.1526.819.48
STVGBert (ICCV'21) [21]20.4229.3711.31
TubeDETR (CVPR'22) [28]32.4049.8023.50
STCAT (NeurIPS'22) [9]35.0957.6730.09
STVGFormer (CVPR'23) [11]36.9062.2034.80
VideoGrounding-DINO38.2562.4736.14
", + "image_path": "e3235c6b2633c1a8d84e537e19b91a7659d5c66cc8e3114fb5ce31132069fcd6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 497, + 287, + 572 + ], + "blocks": [ + { + "bbox": [ + 47, + 466, + 288, + 489 + ], + "lines": [ + { + "bbox": [ + 47, + 466, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 288, + 489 + ], + "type": "text", + "content": "Table 4. Performance comparisons of the state-of-the-art on the HC-STVG V2 [24] val set in closed-set supervised setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 497, + 287, + 572 + ], + "lines": [ + { + "bbox": [ + 50, + 497, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 50, + 497, + 287, + 572 + ], + "type": "table", + "html": "
Methodsm_vIoUvIoU@0.3vIoU@0.5
Yu et al (arXiv'21) [30]30.00--
Aug. 2D-TAN (arXiv'21) [22]30.4050.4018.80
TubeDETR (CVPR'22) [28]36.4058.8030.60
STVGFormer (CVPR'23) [11]38.7065.5033.80
VideoGrounding-DINO39.8867.1334.49
", + "image_path": "4c4578b1a417febd51dd28f370e492e29e5eab770f470bb2d8647979b8ce95c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 597, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 620 + ], + "type": "text", + "content": "trained generalized features of a spatial grounding foundation model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Results on YouCook-Interactions: We evaluate our method further on the YouCook-Interactions [23] dataset, reporting pointing game accuracy for spatial grounding. Our approach gains nearly " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " in accuracy over STCAT [9] and more than " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " compared to TubeDETR [28]. This further shows our strong generalization capabilities in the open-vocabulary setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 328, + 484, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 328, + 484, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 484, + 340 + ], + "type": "text", + "content": "4.2.2 Closed-Set Supervised Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 364, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 545, + 388 + ], + "type": "text", + "content": "We present closed-set evaluations across three datasets, VidSTG [32], HC-STVG V1 [24] and HC-STVG V2 [24]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 564 + ], + "type": "text", + "content": "Results on VidSTG: We present results on the VidSTG test set in closed-set setting in Tab. 2, reporting m_tIoU, m_vIoU, vIoU@0.3 and vIoU@0.5. The results show that our method achieves state-of-the-art performance in comparison to both Two-Stage and One-Stage methods. In particular, we achieve more than " + }, + { + "bbox": [ + 304, + 396, + 546, + 564 + ], + "type": "inline_equation", + "content": "1t\\_ IoU" + }, + { + "bbox": [ + 304, + 396, + 546, + 564 + ], + "type": "text", + "content": " gain in temporal grounding over the previous best methods OMRN [33] (One-Stage) and STVGFormer [11] (Two-Stage), both for Declarative and Interrogative sentences. Similarly, for m_vIoU, vIoU@0.3 and vIoU@0.5, we achieve a more than 1 unit gain the state-of-the-art methods STVGFormer [11] and STCAT [9]. Note that our method uses a frozen visual and textual encoder. In contrast, those mentioned above previous state-of-the-art methods all train the entire encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 572, + 547, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 547, + 644 + ], + "type": "text", + "content": "Results on HC-STVG V1: We present results on the HC-STVG V1 dataset in Tab. 3, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. We achieve a nearly 1.5 unit gain in m_vIoU and vIoU@0.5 and a 1 unit gain in vIoU@0.3 over the previous best method STVGFormer [11]. This shows the consistent performance of our method on this dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Results on HC-STVG V2: We present results on the HC-STVG V2 dataset in Tab. 4, reporting m_vIoU, vIoU@0.3 and vIoU@0.5. Our performance gain on HC-STVG V1 is reflected here as well, with a consistent performance gain on HC-STVG V2, in comparison to the SoTA [11, 28]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18915" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 91, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 515, + 82 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 515, + 82 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 515, + 82 + ], + "type": "text", + "content": "Table 5. Ablation on various design choices for our approach on the VidSTG [32] test set in closed-set supervised setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 91, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 49, + 91, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 547, + 178 + ], + "type": "table", + "html": "
MethodDeclarative SentencesInterrogative Sentences
m_tIoUm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_vIoUvIoU@0.3vIoU@0.5
Naive Solution (Frozen Grounding DINO [12])39.7818.0722.3113.7539.799.6610.423.84
+ Decoder Temporal Aggregation42.8120.7426.5315.4143.8112.3816.718.62
+ Encoder Temporal Aggregation46.2923.1932.3818.9547.1716.1923.2813.04
+ Finetuned Spatial Modules in Decoder48.0628.9741.6026.0649.5824.2732.8520.11
+ Finetuned Spatial Modules in Encoder51.9734.6748.1133.9650.8329.8941.0327.58
", + "image_path": "88e9076221c924f667439eca1135ff62dc4cb7507eea0a4cdc543855e9630a6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 190, + 140, + 251 + ], + "blocks": [ + { + "bbox": [ + 50, + 190, + 140, + 251 + ], + "lines": [ + { + "bbox": [ + 50, + 190, + 140, + 251 + ], + "spans": [ + { + "bbox": [ + 50, + 190, + 140, + 251 + ], + "type": "image", + "image_path": "ca88162c02f563393391dc4d069b87b269c2104263551c82160189ba87dca90a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 251, + 139, + 312 + ], + "blocks": [ + { + "bbox": [ + 50, + 251, + 139, + 312 + ], + "lines": [ + { + "bbox": [ + 50, + 251, + 139, + 312 + ], + "spans": [ + { + "bbox": [ + 50, + 251, + 139, + 312 + ], + "type": "image", + "image_path": "ea612f6d83381b82ff1dfdfe080133886293bc62b3eeb0b2eec3c5adaadddfe8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 313, + 139, + 374 + ], + "blocks": [ + { + "bbox": [ + 50, + 313, + 139, + 374 + ], + "lines": [ + { + "bbox": [ + 50, + 313, + 139, + 374 + ], + "spans": [ + { + "bbox": [ + 50, + 313, + 139, + 374 + ], + "type": "image", + "image_path": "98d29cc55f91c5c6a4967598f9e91a788603fe8d0f54a9fd47cb6021d88b7e95.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 379, + 546, + 424 + ], + "lines": [ + { + "bbox": [ + 47, + 379, + 546, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 546, + 424 + ], + "type": "text", + "content": "Figure 3. Sample visualization for video grounding result on HC-STVG V1 [24] for TubeDETR [28], STCAT [9] and ours with the prompt The man behind the shirtless man turns and squats. We show bounding boxes for Ground-truth, Closed-Set Supervised, and OpenVocabulary results. Note how both TubeDETR and STCAT are close to the ground truth in the supervised setting (STCAT more so than TubeDETR), they cannot correctly ground the text properly in the open-vocabulary setting." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 141, + 190, + 220, + 251 + ], + "blocks": [ + { + "bbox": [ + 141, + 190, + 220, + 251 + ], + "lines": [ + { + "bbox": [ + 141, + 190, + 220, + 251 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 220, + 251 + ], + "type": "image", + "image_path": "2ac908981847e7812c09c5e3877a061c747288912af8a2947b4a0df40ee61faf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 141, + 251, + 220, + 312 + ], + "blocks": [ + { + "bbox": [ + 141, + 251, + 220, + 312 + ], + "lines": [ + { + "bbox": [ + 141, + 251, + 220, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 251, + 220, + 312 + ], + "type": "image", + "image_path": "73f969c386dda56ab005bd106858ad39533f7999b610795328bd3e11a132ff52.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 141, + 313, + 220, + 374 + ], + "blocks": [ + { + "bbox": [ + 141, + 313, + 220, + 374 + ], + "lines": [ + { + "bbox": [ + 141, + 313, + 220, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 220, + 374 + ], + "type": "image", + "image_path": "8f860a59b7490cbddaf0e2d4894768c07c0ee2fd17c2e1481ce2dab2202d0662.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 222, + 190, + 302, + 251 + ], + "blocks": [ + { + "bbox": [ + 222, + 190, + 302, + 251 + ], + "lines": [ + { + "bbox": [ + 222, + 190, + 302, + 251 + ], + "spans": [ + { + "bbox": [ + 222, + 190, + 302, + 251 + ], + "type": "image", + "image_path": "d83187aa9cd505b9f960cf5bcd00bc52ac638a865df585edc410be35cbd59a3c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 222, + 251, + 301, + 312 + ], + "blocks": [ + { + "bbox": [ + 222, + 251, + 301, + 312 + ], + "lines": [ + { + "bbox": [ + 222, + 251, + 301, + 312 + ], + "spans": [ + { + "bbox": [ + 222, + 251, + 301, + 312 + ], + "type": "image", + "image_path": "8fe8f6d61883e0cc4d194384b5982d276019e540308fc337d149fb16f482d04b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 222, + 313, + 301, + 374 + ], + "blocks": [ + { + "bbox": [ + 222, + 313, + 301, + 374 + ], + "lines": [ + { + "bbox": [ + 222, + 313, + 301, + 374 + ], + "spans": [ + { + "bbox": [ + 222, + 313, + 301, + 374 + ], + "type": "image", + "image_path": "9860f1dbe704e5b1ba0cdc49a4717a0d09331b17134a7b8c28264896aa2e25fc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 302, + 190, + 381, + 251 + ], + "blocks": [ + { + "bbox": [ + 302, + 190, + 381, + 251 + ], + "lines": [ + { + "bbox": [ + 302, + 190, + 381, + 251 + ], + "spans": [ + { + "bbox": [ + 302, + 190, + 381, + 251 + ], + "type": "image", + "image_path": "1fb5ca49031dbe6f986f0952e7aada430234de34986f931ee813c92172a74a8e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 302, + 251, + 381, + 312 + ], + "blocks": [ + { + "bbox": [ + 302, + 251, + 381, + 312 + ], + "lines": [ + { + "bbox": [ + 302, + 251, + 381, + 312 + ], + "spans": [ + { + "bbox": [ + 302, + 251, + 381, + 312 + ], + "type": "image", + "image_path": "0f4885f3d88815f1ab7d25e9196f26cf1ade29cd810562e02f1e98e698b30f2b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 302, + 313, + 381, + 374 + ], + "blocks": [ + { + "bbox": [ + 302, + 313, + 381, + 374 + ], + "lines": [ + { + "bbox": [ + 302, + 313, + 381, + 374 + ], + "spans": [ + { + "bbox": [ + 302, + 313, + 381, + 374 + ], + "type": "image", + "image_path": "a6fee9a4934a0c0182e4bdada01195b6599d659e8ee0bd36fa0fc7d75559e84d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 383, + 190, + 462, + 251 + ], + "blocks": [ + { + "bbox": [ + 383, + 190, + 462, + 251 + ], + "lines": [ + { + "bbox": [ + 383, + 190, + 462, + 251 + ], + "spans": [ + { + "bbox": [ + 383, + 190, + 462, + 251 + ], + "type": "image", + "image_path": "a1bbb07d32c6746de6dfebc43ac9cfca10403baf304bee4bf1606fc98ad12710.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 251, + 462, + 312 + ], + "blocks": [ + { + "bbox": [ + 383, + 251, + 462, + 312 + ], + "lines": [ + { + "bbox": [ + 383, + 251, + 462, + 312 + ], + "spans": [ + { + "bbox": [ + 383, + 251, + 462, + 312 + ], + "type": "image", + "image_path": "44e36a493c450f89d73962bdc7d207717185f1adbbf3be658576e04e621fa3f3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 383, + 313, + 462, + 374 + ], + "blocks": [ + { + "bbox": [ + 383, + 313, + 462, + 374 + ], + "lines": [ + { + "bbox": [ + 383, + 313, + 462, + 374 + ], + "spans": [ + { + "bbox": [ + 383, + 313, + 462, + 374 + ], + "type": "image", + "image_path": "2c709ad6d1ce8374ee9b7c39816f002ab03819b31d13c3435199b2924bb0fdb0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 463, + 190, + 543, + 251 + ], + "blocks": [ + { + "bbox": [ + 463, + 190, + 543, + 251 + ], + "lines": [ + { + "bbox": [ + 463, + 190, + 543, + 251 + ], + "spans": [ + { + "bbox": [ + 463, + 190, + 543, + 251 + ], + "type": "image", + "image_path": "317b3d8135774a989b8f0df6d5ba5d4480a274ae7a9d162bef6aeea9cc9d03e2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 463, + 251, + 543, + 312 + ], + "blocks": [ + { + "bbox": [ + 463, + 251, + 543, + 312 + ], + "lines": [ + { + "bbox": [ + 463, + 251, + 543, + 312 + ], + "spans": [ + { + "bbox": [ + 463, + 251, + 543, + 312 + ], + "type": "image", + "image_path": "3bfc08974eaf1c5604943d9037789109a1d471a36d3f0a266b3ced507a0aab2f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 463, + 313, + 543, + 374 + ], + "blocks": [ + { + "bbox": [ + 463, + 313, + 543, + 374 + ], + "lines": [ + { + "bbox": [ + 463, + 313, + 543, + 374 + ], + "spans": [ + { + "bbox": [ + 463, + 313, + 543, + 374 + ], + "type": "image", + "image_path": "25ef794d542e9e6bbc238641c65b5c2ef42c79dc304df9b0471659d58cf4617e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 436, + 151, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 151, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 151, + 449 + ], + "type": "text", + "content": "4.3. Ablative Analysis" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 455, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 612 + ], + "type": "text", + "content": "We perform an ablative analysis of the various design choices for our model. In particular, we first evaluate a naive baseline where no additional temporal aggregators are added, and all pre-trained spatial modules are frozen in the encoder and decoder. This baseline is relatively weak in both temporal and spatial grounding. We next add temporal modules in first the decoder, and subsequently in the encoder. We find that it gives significant improvements in the temporal grounding, and additionally also improves the spatial grounding. Finally, we finetune the pre-trained spatial modules in both the decoder and the encoder, which provides a strong improvement in spatial grounding, alongside consistent improvement in temporal grounding." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 622, + 119, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 119, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 119, + 633 + ], + "type": "text", + "content": "4.4. Limitation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "While our video grounding model excels in closed-set and open-vocabulary scenarios, it leverages image-text pretrained models like Grounding DINO [12]. To enhance understanding in open-vocabulary settings, an extension to video-language pre-training on a larger and more diverse dataset, akin to CLIP [15], can help further boost perfor" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 436, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 485 + ], + "type": "text", + "content": "mance. Building a video-language pre-training dataset with diverse natural language expressions and spatio-temporal localization is imperative, given the constraints of datasets like VidSTG [32] and HC-STVG [24]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": "This paper introduces an Open-Vocabulary Spatio-Temporal Video Grounding task, enhancing current closed-set methodologies by using pre-trained representations from spatial grounding models. The proposed model performs well in closed-set and open-vocabulary scenarios, surpassing state-of-the-art results in supervised setting on VidSTG and HC-STVG datasets, and outperforming recent models in open-vocabulary on HC-STVG V1 and YouCook-Interactions. Its architecture includes learnable adapter blocks for video-specific adaptation, bridging the semantic gap between natural language queries and visual content. This research addresses open-vocabulary challenges and explores achieving robust performance without extensive video annotations, paving the way for open-vocabulary video grounding." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18916" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "text", + "content": "[2] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where? - self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "type": "text", + "content": "[3] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, , and Jiebo Luo. Localizing natural language in videos. In AAAI, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "type": "text", + "content": "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 372 + ], + "type": "text", + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "type": "text", + "content": "[7] Jiyang Gao, Chen Sun, Zhenheng Yang, , and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 485 + ], + "type": "text", + "content": "[8] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 530 + ], + "type": "text", + "content": "[9] Yang Jin, Yongzhi Li, Zehuan Yuan, and Yadong Mu. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In NeurIPS, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "text", + "content": "[10] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[11] Zihang Lin, Chaolei Tan, Jian-Fang Hu, Zhi Jin, Tiancai Ye, and Wei-Shi Zheng. Collaborative static and dynamic vision-language streams for spatio-temporal video grounding. In CVPR, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[14] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 204 + ], + "type": "text", + "content": "[15] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "text", + "content": "[16] Hanoona Rasheeda, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[17] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union. In CVPR, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "text", + "content": "[18] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[19] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, , and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In ECCV, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "text", + "content": "[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In ICMR, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[21] Rui Su, Qian Yu., and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In ICCV, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 438, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 481 + ], + "type": "text", + "content": "[22] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[23] Reuben Tan, Bryan A. Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 581 + ], + "type": "text", + "content": "[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 647 + ], + "type": "text", + "content": "[26] Zhenzhi Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "text", + "content": "[27] Masataka Yamaguchi, Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Spatio-temporal person retrieval via natural language queries. In ICCV, 2017. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18917" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 354 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[30] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[31] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "type": "text", + "content": "[32] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In CVPR, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 252, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 288, + 285 + ], + "type": "text", + "content": "[33] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In *IJCAI*, 2021. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 319 + ], + "type": "text", + "content": "[34] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 354 + ], + "type": "text", + "content": "[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 4, 5" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18918" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_content_list.json b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..798f66f689cae47232bcf998c15100a6bdf5735f --- /dev/null +++ b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_content_list.json @@ -0,0 +1,1642 @@ +[ + { + "type": "text", + "text": "VideoLLM-online: Online Video Large Language Model for Streaming Video", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joya Chen1", + "bbox": [ + 135, + 179, + 228, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Difei Gao", + "bbox": [ + 153, + 198, + 241, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoyang Lv²", + "bbox": [ + 254, + 180, + 370, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia-Wei Liu1", + "bbox": [ + 267, + 198, + 367, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiwei Wu", + "bbox": [ + 393, + 180, + 491, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziteng Gao", + "bbox": [ + 393, + 198, + 493, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kevin Qinghong Lin1", + "bbox": [ + 517, + 180, + 687, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongxing Mao", + "bbox": [ + 519, + 198, + 648, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "n1 Chenan Song", + "bbox": [ + 673, + 180, + 831, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mike Zheng Shou", + "bbox": [ + 674, + 198, + 836, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Show Lab, National University of Singapore 2Reality Labs Research, Meta", + "bbox": [ + 173, + 219, + 792, + 237 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55bcd345351946a74795ffc9d05af932569b03fe66e2cd880067414b9227d9de.jpg", + "image_caption": [ + "Streaming Dialogue", + "Figure 1. Zero-shot examples of our VideoLLM-online applied to an egocentric video stream from Ego-Exo4D dataset [27]. Our model is designed for temporally aligned, long-context, real-time dialogue in continuous video streams, shedding light on the future always-on, contextual AI assistants (e.g., smart AR glasses). Model responses are appropriately simplified for better visualization." + ], + "image_footnote": [], + "bbox": [ + 78, + 250, + 890, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 584, + 313, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent Large Language Models (LLMs) have been enhanced with vision capabilities, enabling them to comprehend images, videos, and interleaved vision-language content. However, the learning methods of these large multimodal models (LMMs) typically treat videos as predetermined clips, rendering them less effective and efficient at handling streaming video inputs. In this paper, we propose a novel Learning-In-Video-Stream (LIVE) framework, which enables temporally aligned, long-context, and real-time dialogue within a continuous video stream. Our LIVE framework comprises comprehensive approaches to achieve video streaming dialogue, encompassing: (1) a training objective designed to perform language modeling for continuous streaming inputs, (2) a data generation scheme that converts offline temporal annotations into a streaming dialogue format, and (3) an optimized inference pipeline to speed up interactive chat in real-world video streams. With", + "bbox": [ + 73, + 617, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "our LIVE framework, we develop a simplified model called VideoLLM-online and demonstrate its significant advantages in processing streaming videos. For instance, our VideoLLM-online-7B model can operate at over 10 FPS on an A100 GPU for a 5-minute video clip from Ego4D narration. Moreover, VideoLLM-online also showcases state-of-the-art performance on public offline video benchmarks, such as recognition, captioning, and forecasting. The code, model, data, and demo have been made available at showlab.github.io/videoollm-online.", + "bbox": [ + 500, + 582, + 892, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 767, + 630, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building the future of always-on contextual AI agent that can promptly answer any human question, digitizing inputs as episodic memories and forecasting future plans given any query in online continuous setting represents a \"holy grail\" mission in AI research. Powered by advancements in large language models (LLMs) [8, 34, 60, 61, 63, 72], recent large multimodal models (LMMs) have unveiled impressive ca", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author.", + "bbox": [ + 76, + 886, + 210, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18407", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pabilities such as vision-language dialogue [6, 13, 40, 41, 48, 49, 93], spatial understanding [6, 37, 51, 64, 83, 87], processing diverse modalities [19, 28, 55, 77, 85]. Seminal exemplars, like OpenAI's GPT-4V [62], are progressively evolving into highly versatile human AI assistants.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is time to envision an always-on,contextual, J.A.R.V.I.S-like multimodal assistant that supports freeform user-assistant dialogue within the video stream, which we term \"video streaming dialogue\". Unlike existing LMMs for video understanding (i.e., VideoLLMs) [42, 43, 52, 56, 69, 81, 86] that work offline with manually selected short-video clips, an online assistant should continuously receive video frames with visual content that is constantly refreshed. This paradigm shift presents new challenges. First, the user query may come with temporally aligned requirements (e.g., \"alert me when it's time to flip the steak\"), thus the VideoLLM should scan every incoming frame to avoid event missing, instead of only yielding video-level responses. Second, to answer questions regarding summarization and planning, the VideoLLM must retain the longcontext historical vision and language, which poses huge risk of exceeding maximum context window of LLMs, as well as introduces considerable burden to causal decoding speed and GPU memory. Third, the VideoLLM should generate the answer in real-time, keeping pace with the video stream for always-on scenario. These abilities, however, are even partially overlooked by the most advanced AI assistants [3, 62].", + "bbox": [ + 75, + 170, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One possible path towards such an online VideoLLM, inspired by current interleaved vision-language models [3, 6, 40, 62, 82], is to employ a multi-turn dialogue format to achieve per-frame chatting within a video stream. This can be accomplished by facilitating very frequent user interactions, utilizing the visual frame as query at each timestamp to obtain the answer. We follow this to perform prompt engineering for GPT-4V [44, 62], but the results are disappointing: GPT-4V tends to output lengthy content at every frame, leading to significant delays, making it impractical for real-time streaming video. We also explore training baseline models for per-frame chatting. Unfortunately, this approach evidently diminishes the language modeling capability, likely due to harmful language modeling on an excessive number of redundant frames.", + "bbox": [ + 75, + 518, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose Learning-In-Video-strEam (LIVE), a comprehensive framework that encompasses learning, data, and inference methods to develop an online video assistant. Unlike per-frame dialogue approach, LIVE introduces a novel training objective termed Streaming EOS (End-Of-Sequence) prediction that enables the model to learn when to response or remain silent in a video stream. This objective differs from next-token prediction since EOS tokens here will not appear in the input/output sequence. However, it can work well with the autoregressive loss to train", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an online VideoLLM. This design reduces unnecessary context, helping the model to manage much longer streaming videos. Nevertheless, the training still requires data from user queries and assistant responses within video streams, which is scarce in popular video datasets usually used for training offline video models. To address this issue, LIVE presents a streaming dialogue generation scheme that converts offline annotations into online dialogues to support free-form chatting. To enhance inference efficiency, LIVE leverages continuous key-value caching for streaming assistance, and parallelizes the fast visual encoding and slow language decoding to prevent bottlenecks, thus moving towards real-time application.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With LIVE framework, we build a simple VideoLLM-online model upon CLIP [65] vision encoder and Llama2 [73] language model. To evaluate the performance of video streaming dialogue, we utilize the language perplexity metric and design two new metrics to comprehensively assess the model's capabilities in language modeling, temporal responsiveness, and overall fluency. Experiments with real-time Ego4D narration [26] demonstrate that our method shows advantages in all that metrics, with higher speed and lower memory cost. Furthermore, our model achieves state-of-the-art results on numerous offline benchmarks, such as short- and long-term activity recognition and forecasting on the COIN and Ego4D LTA benchmarks [26, 71]. In addition, our model has good speed/memory efficiency, which allows continuous video narration on Ego4D video clips with memory cost less than 20 GB and average speed higher than 10 FPS on a single A100 GPU, paving the way for future real-world usage.", + "bbox": [ + 496, + 287, + 892, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 573, + 640, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Dialogue. Before transformers [74] become mainstream in vision, visual dialogue methods [1, 16] tend to employ a visually enhanced encoder, with a discriminative head to select candidate answers or a recurrent architecture to generate multi-turn responses. For the encoder, a variety of attention mechanism-based approaches [22, 59, 66] have been proposed to learn the interactions between the image, the answers, and the dialogue history. There have also been explorations of encoder-only BERT [18] models for visual dialogue [57, 76]. However, most of them rely solely on a single image/video at the beginning of the conversation, followed by multi-turn pure language dialogue, which makes them less flexible than the current interleaved vision-language dialogue systems.", + "bbox": [ + 496, + 598, + 890, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Multimodal Models. The advent of large language models (LLMs) [8, 60, 63] has inspired a series of LLM multimodal variants, i.e. large multimodal models (LMMs). Early LMMs [2, 13, 41, 49, 93] achieve image dialogue by projecting the image encoding (e.g., from CLIP [65]) to align with LLM embedding space. Then, lots of ef", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "forts [3, 6, 40, 62, 82] explore more free-form interleaved vision-text chatting, spatial understanding [6, 11, 12, 37, 51, 83, 87], video comprehension [42, 43, 52, 56, 69, 81, 86], etc. However, when it comes to online scenario, there is less exploration on how LLMs can fulfill the temporal alignment, long-context, and real-time requirements for streaming video inputs. Our research bridges this gap, offering comprehensive solutions across model training, data, and inference to study the problem.", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online Video Understanding. Typical video understanding benchmarks, such as action recognition [9], temporal action localization [29], video question answering [38], and video dialogue [1], typically allow models to access entire video frames to make predictions, a setting referred to as \"offline\". However, such setting does not align well with many real-time demands (e.g., autonomous driving, AR glasses). Instead, there is a growing focus on \"online\" video understanding problems like online action detection [24] and anticipation [35], which aim to identify the current/future action at each timestamp without seeing the future. Our study pioneers LMMs for online video understanding. Unlike previous online action detection [75, 89] or anticipation models [25, 88], which primarily address one task with a highly customized model, we aim to propose a general solution to achieve free-form dialogue along the online video stream, enabling a model to flexibly handle diverse tasks.", + "bbox": [ + 75, + 227, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Token Decoding. Efficient token decoding for LLMs and LMMs is essential for applying them to real-time online services. To accelerate that, a diversity of strategies has been proposed, such as parallelism on batch dimension [67] or cached key-value sequence [15], computation/memory management optimization [23, 30, 36, 67, 84], and even some lossy approaches [50, 78]. Our focused streaming video scenario has less concerns in large batch processing, but expects faster decoding to avoid excessive frame skipping. We have considered this in training objectives, and further propose some inference schemes to accelerate the decoding efficiency.", + "bbox": [ + 75, + 484, + 472, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 676, + 169, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present Learning-In-Video-strEam (LIVE) framework that enables LMMs to provide temporally aligned response, handle long-context streaming video, and run efficiently towards real-time usage. We will start from the problem definition of \"video streaming dialogue\", analyze its challenges, and introduce our approach to solve the problem.", + "bbox": [ + 75, + 702, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video Streaming Dialogue", + "text_level": 1, + "bbox": [ + 76, + 816, + 313, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Though huge successes have been witnessed in large multimodal models (LMMs), the assistance scenario like a smart AR glass helping the user with cooking, is still far from the capabilities of current LMMs,", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "even for the most advanced version, GPT-4V [62]. For example, despite carefully prompting GPT-4V similarly to MMVid [44]—to perform per-frame dialogue for handling streaming video inputs—the redundancy in responses between frames, a limited context window of $10\\sim 50$ frames, and slow speed, collectively render the current GPT-4V unsuitable for online video understanding, as our prompting analysis demonstrates (see supplementary material).", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To bridge the gap, we define a problem termed \"video streaming dialogue\". Given the context sequence before time $t = t_1$ , denoted as $[\\mathrm{Ctx}^{t < t_1}]$ , which may encompass previous vision-language content (e.g., historical user queries, video frames, assistant responses), and an ongoing continuous video stream from $t_1$ to $t_2$ , denoted as $[\\mathrm{Frame}^{t_1 \\leq t \\leq t_2}]$ , our goal is (1) to determine whether the current time $t_2$ is suitable for language modeling; (2) to carry out language modeling", + "bbox": [ + 496, + 212, + 892, + 348 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t _ {2}} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {1}} \\right], \\left[ \\mathrm {F} ^ {t _ {1} \\leq t \\leq t _ {2}} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t _ {2}} \\right]\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 373, + 890, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "if $t_2$ is determined, where $[\\mathrm{Txt}_i^t]$ denotes the ideal language token in the $i$ -th position at timestamp $t$ . [F] is the abbreviation of [Frame].", + "bbox": [ + 496, + 405, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we analyze if existing techniques have been enough to solve the problem.", + "bbox": [ + 496, + 450, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interleaved/Per-frame Dialogue are Suboptimal. First, we investigate whether the popular approach of interleaved vision-language chatting can address this problem. Related to our formulation above, such a method learns language modeling after given frames between timestamps $t_1$ and $t_2$ . However, if this approach is adopted during inference, it necessitates the manual selection of timestamps $t_1$ and $t_2$ , which does not align with the concept of video streaming dialogue. Current VideoLLMs [42, 43, 52, 56, 69, 81, 86] represent a simplified version of this approach, engaging in single- or multi-turn pure language dialogue following video clip inputs.", + "bbox": [ + 496, + 481, + 892, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If we consider a more free-form format, multi-turn interleaved vision-language chatting [3, 6, 40, 62, 82], we can get a per-frame chatting solution that might be hopeful to solve video streaming dialogue, which performs per-frame language modeling", + "bbox": [ + 496, + 662, + 890, + 737 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {1}} \\right], \\left[ \\mathrm {F r a m e} ^ {t} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t} \\right]\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 762, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for every frame from timestamp $t_1 \\leq t \\leq t_2$ . Since $t_2$ is necessary to output answer, so we can simply learn short text (e.g., \"this is not the time to answer\") between $t_1 \\leq t < t_2$ . However, this approach imposes a significant burden on processing speed. For every frame, performing the slow, recurrent, and lengthy next-token prediction with a billion-scale language model makes it extremely hard to", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18409", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "achieve real-time video streaming dialogue. Then, this will lead to unavoidable frame skipping, which is unexpected and detrimental for temporal alignment. Furthermore, it taxes the limited context window of the language model, presenting challenges for modeling long contexts and managing GPU memory efficiently.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Streaming EOS Prediction. To solve the problem mentioned above, we first consider a more efficient per-frame chatting method: simply assigning the End-of-Sequence (EOS) token as the content for chatting between $t_1 \\leq t < t_2$ . However, this approach remains suboptimal. The dialogue prompt template (e.g., [INST], [/INST], space tokens in Llama [72, 73]) still consumes a considerable number of tokens per frame, which is less favorable due to the numerous frames in streaming video. Furthermore, the excessive number of EOS tokens in the sequence can significantly increase the language model's perplexity, as we have observed in our experiments.", + "bbox": [ + 75, + 181, + 470, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead, we propose a novel training objective named \"streaming EOS prediction\" to address this issue. We still assume $t_2$ is essential for decoding language; thus, we normally learn language modeling here:", + "bbox": [ + 75, + 363, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t _ {2}} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {2}} \\right], \\left[ \\mathrm {F r a m e} ^ {t _ {2}} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t _ {2}} \\right]\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 446, + 468, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, for timestamps $t_1 \\leq t < t_2$ , which are redundant for producing answers, we directly learn the model to predict EOS token on the frame tokens, i.e.", + "bbox": [ + 75, + 479, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max P (\\text {E O S} \\left| [ \\text {C t x} ^ {< t} ], [ \\text {F r a m e} ^ {t} ]\\right), \\text {w h e r e} t _ {1} \\leq t < t _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 550, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we \"skip\" a dialogue turn and learn to determine when it is appropriate to decode language for streaming inputs. During inference, if EOS is predicted on a frame, then we can directly ask the next frame to input. Meanwhile, the EOS token is not appended to the context to prevent it from affecting the language modeling. Therefore, this task is not about next-token prediction; however, it can work with autoregressive loss to train a video streaming dialogue model. In the following, we first introduce how can we get the data of streaming video and timestamped language annotations, then present the details about the model and the training procedure.", + "bbox": [ + 75, + 580, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also note that the EOS token mentioned here is not limited to the real EOS token used in language models (e.g., $\\langle / \\mathbb{S} \\rangle$ in Llama). It is permissible to use any token or to introduce a new token, provided that it is specified in the system prompt. We use this term solely for simplicity.", + "bbox": [ + 75, + 762, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.Data", + "text_level": 1, + "bbox": [ + 76, + 847, + 150, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online Annotations to Video Streaming Dialogue. Some video datasets, such as Ego4D narrations [26], are inher", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10c68d0ebedd7216acfc5738c9b8c84dc1920e547942e2b8504460528aa67ca1.jpg", + "image_caption": [ + "Figure 2. The streaming dialogue data generation method in our LIVE framework. We randomly insert templated questions into the video timeline and \"expose\" the ground-truth video annotations (along with their timestamps) to LLMs, prompting them to answer the queries within a period of time." + ], + "image_footnote": [], + "bbox": [ + 498, + 88, + 890, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ently collected in a streaming manner, with annotators providing real-time narrations as they watch a 5-minute long video clip. However, prior research [45, 90] has predominantly focused on learning from short, discrete clips (e.g. only 32 frames), rather than in a continuous streaming context. For this dataset, we followed the same instructions provided to human annotators [26] as our model training prompt. This prompt instructs the model to simulate human annotators to streamingly generate narrations in a 5-minute video (about 600 frames in 2 FPS). For demo purposes (not for experiments), we also utilize Llama-2-13B-Chat [73] to rephrase the narration text, correcting grammatical errors and typos, and converting it into a more understandable version (e.g., changing \"C does...\" to \"You do...\").", + "bbox": [ + 496, + 551, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Offline Annotations to Video Streaming Dialogue. Despite the Ego4D narration data being collected in a streaming manner, most prevalent video datasets [14, 26, 54, 71] are used to train offline models and only feature temporal segment annotations paired with basic language descriptions (e.g., activities, narrations). To bridge this gap, we propose a method for synthesizing dialogue data from these sources. As shown in Figure 2, our key idea is use LLM to generate user-assistant dialogues based on video annota", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18410", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions, involving the following steps:", + "bbox": [ + 76, + 90, + 316, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First, we prepare a question template library containing various queries about the past, present, and future tenses of the video, totaling $N$ queries. We randomly sample one question from the library, denoted as $Q_{i}$ .", + "- Then, we obtain the video annotation timeline from the offline dataset. This usually includes timestamped language descriptions, which we organize into a language prompt, e.g., \"time $t_a \\sim t_b$ : boiling the water; time $t_c \\sim t_d$ : cutting the vegetables.\", denoted as $A$ . We consider all the state change critical timestamps as the ideal response times. For this example, $t_a, t_b, t_c$ , and $t_d$ are all considered response times.", + "- Third, we prompt the large language model to generate responses at every critical timestamp, e.g., $t_a, t_b, t_c, t_d$ , according to $Q_i$ and $A$ . We can repeat this procedure for each $Q_i$ until all queries have been processed. The responses are saved for loading during training.", + "- Finally, during training, we (1) randomly sample a query and load its responses at critical timestamps, (2) randomly insert a query into a video timestamp $t_r$ , (3) discard the responses that occur before $t_r$ , and add a response at $t_r$ . Here different queries can be inserted into one video, which only requires discarding the responses of the previous query after the new query insertion timestamp." + ], + "bbox": [ + 76, + 116, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we can generate temporally varied and freeform dialogue data within a video stream. We have prepared 50 questions each for past, current, and future events, totaling $N = 150$ queries. We use Llama-2-13B-Chat to generate the responses and insert a maximum of 3 queries per training sample. The offline datasets we used are COIN [71] and Ego4D GoalStep [70] (for demo usage), which belong to the categories of egocentric and instructional video datasets, aligning with our aim to develop online video assistants. Here, we do not consider online action detection benchmarks (e.g., THUMOS14 [33], TVSeries [24]) because they are closed-set online classification benchmarks, and their labels are too brief, which may cause language models to generate hallucinatory responses. Please refer to the supplementary material for the generated dialogue example.", + "bbox": [ + 75, + 489, + 467, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Model Training", + "text_level": 1, + "bbox": [ + 76, + 752, + 233, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Architecture. We illustrate the model architecture in Figure 3. Similar to LLaVA [48, 49], it comprises three key components: an image encoder, an MLP projector, and a language model. For the image encoder, we utilize the CLIP ViT-L [20, 65] encoder (pretrained on DataComp1B [21]) to extract video frame embeddings at 2 FPS. Each video frame embedding has a shape of $(1 + h_p\\times w_p)\\times c$ where $(1 + h_{p}\\times w_{p})$ denotes the CLS token and average", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pooled spatial tokens. The extracted frame embeddings from the image encoder are then fed into MLP projector to frame tokens, as in LLaVA-1.5 [48]. Then frame tokens are interleaved with language tokens as input to an LLM, Llama-2-7B-Chat [73]. Finally, we incorporate LoRA [31] in every linear layer of the LLM for efficient tuning.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Loss. As described in Section 3.1, our learning objective is twofold. The first part focuses on autoregressive language modeling, aiming to maximize the joint probability of input text sequences. The second training objective involves streaming EOS prediction, which requires the model to remain silent when it is unnecessary to output responses. With these two training objectives, we have language modeling (LM) loss and streaming loss terms to minimize, both employing cross-entropy loss:", + "bbox": [ + 496, + 181, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\left(\\underbrace {- \\log l _ {j + 1} P _ {j} ^ {[ T x t _ {j + 1} ]}} _ {L M L o s s} - \\underbrace {w \\log f _ {j} P _ {j} ^ {[ E O S ]}} _ {\\text {S t r e a m i n g L o s s}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 339, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l_{j}$ and $f_{j}$ denote condition indicators. $l_{j}$ is 1 if the $j$ -th token is a language response token, and 0 otherwise. $f_{j}$ is 1 if (1) the $j$ -th token is the last token of a frame $^3$ , and (2) $l_{j+1} = 0$ . In essence, the streaming EOS loss is applied to frames before responding. $P_{j}^{[\\mathrm{Txt}_{j+1}]}$ denotes the probability on $j + 1$ -th text token, output from the language model head of the $j$ -th token, and $P_{j}^{[\\mathrm{EOS}]}$ represents that probability for the EOS token. $w$ is a balance term, set to $w = 1$ by default. As shown in Figuer 3, we visualize the ranges of language loss and streaming loss in an input sequence when we only use 1 token for each frame.", + "bbox": [ + 496, + 398, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference", + "text_level": 1, + "bbox": [ + 500, + 575, + 609, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Probability Correction. The prevalence of EOS token will bias the model towards EOS token prediction. To address this, we introduce a threshold $\\theta$ to correct the output probability on frame tokens: EOS will not be considered as the next token if $P_{j}^{[\\mathrm{EOS}]} < \\theta$ . In practical usage, we find that setting $\\theta$ to 0.5 yields much better results than no threshold here.", + "bbox": [ + 496, + 598, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Continuous Key-Value Cache. As shown in Figure 4, During inference, the video is input as a frame-by-frame stream, with a default FPS 2. Our model takes the current frame as input and generates tokens on-the-fly. In the whole process, we use the key-value cache trick to accelerate token decoding, thus we do not need to manually append the generated", + "bbox": [ + 496, + 704, + 890, + 795 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2The experiments described in this paper are conducted without extra spatial tokens (i.e., $h_p = w_p = 0$ ), which is the most efficient setup and can handle half-hour videos within a 4096 context window. Our released models for demo usage include $1 + h_p \\times w_p = 1 + 3 \\times 3 = 10$ tokens, offering better detail in dialogue but supporting shorter maximum video lengths. Despite more tokens used per frame, all models can still run at over 10 FPS for 5-minute Ego4D narration streams.", + "bbox": [ + 500, + 801, + 890, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3When a frame has multiple patch tokens, loss is only on the last one.", + "bbox": [ + 500, + 887, + 866, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f11632ceebe9151bacb8a426e61456bbb5f741b41c38654aad3375bcb54ee584.jpg", + "image_caption": [ + "Figure 3. The training method in our LIVE framework. We organize the user-assistant dialogue data and video frames in temporal order as the input sequence. To learn the model when to answer or keep silent in a video stream, we employ not only the standard language modeling (LM) loss but also introduce a streaming EOS prediction loss. This additional loss supervises the model when it is necessary to generate language, enabling it to produce temporally aligned responses and reduces the redundant dialogue history." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 869, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a5624900b49eba8690c05eaa1796c72854767a44d7e5fb14362a770626154e3.jpg", + "image_caption": [ + "Figure 4. Inference pipeline in our LIVE framework. During inference, video frames serve as streaming inputs. Our model maintains a continuous key-value cache as the input progresses to speed up the inference. Furthermore, we parallelize the fast video frame encoder and the slower language model to avoid the bottleneck in the latter. Video frame tokens can be always encoded and buffered, no need to wait the language decoding." + ], + "image_footnote": [], + "bbox": [ + 78, + 523, + 465, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tokens for the next frame. As our training encourages the model to keep silence, this continuous inference would be efficient, providing the possibility to pace with the video stream speed.", + "bbox": [ + 75, + 776, + 468, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parallelization of encoding and decoding. Our video frame encoder utilizes CLIP ViT-L (307M), which is significantly smaller than the 7B LLM. This size discrepancy leads to a speed mismatch, potentially resulting in frame", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "skipping when the LLM decodes long sentences. To mitigate this issue, we parallelize the processes and establish a buffer for video frame tokens. Once the language model completes its decoding, it fetches all frame tokens from the buffer and inputs them in batch.", + "bbox": [ + 498, + 526, + 890, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 648, + 632, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 684, + 717, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement VideoLLM-online-7B with our LIVE framework. It uses OpenCLIP-ViT-L [21, 65] as the video frame encoder, 2-layer MLP as multimodal connector, and Llama2-7B-Chat [73] as the language model. We only use 1 token for each video frame, sampled by 2 FPS in the following experiments. We add LoRA [31] to all linear layers of Llama-2, with a rank of 128, scaling factor of 256. For video streaming dialogue experiments, we train 2 epochs for the model. For the downstreaming offline experiments, we train 5 epochs without pre-training for fair comparison with previous methods. By default, we set streaming loss weight $w = 1.0$ during training.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18412", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/18893148faf0f495155dedd2dff82155ce06aaf19c9177c3e9f4b869b78d82e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTraining ObjectiveEgo4D Narration Stream on Validation Set#Training Token↓Training Cost
LM-PPL↓TimeDiff ↓Fluency↑
No Training498.56.500.1%n/an/a
Interleaved DialogueLanguage Modeling2.456.4711.1%169412h
Per-frame Dialogue for StreamingLanguage Modeling (w/ EOS turns)3.342.5237.7 %673722h
Streaming Dialogue (Ours)Language Modeling + Streaming EOS2.432.3242.6%169412h
", + "bbox": [ + 96, + 89, + 872, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d768ec8109d97f49e8fc8621e254d838c30989c993b75fdac038855d68b661f.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Streaming loss function. Standard CE (cross-entropy) is enough for training streaming dialogue; there is no need to specifically to address the class imbalance on EOS token." + ], + "table_body": "
Streaming LossEgo4DNarrationStreamValidation
LM-PPL↓TimeDiff↓Fluency↑
Standard CE2.432.3242.6%
OHEM [68]2.532.3941.0%
Focal Loss [46]2.592.4439.4%
", + "bbox": [ + 96, + 227, + 390, + 295 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cce9a4a7284fb4d4f2b5e9b97f7cab496c6bd81397bacfc8981db723f3770cc1.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Streaming loss weight. Using slightly higher stream- (d) Generation memory/speed. ing loss weight $(\\tau = 2.0)$ achieves the best trade-off Streaming dialogue method has among various metrics. much better efficiency." + ], + "table_body": "
Weight τEgo4D Narration Stream Validation
LM-PPL↓TimeDiff↓Fluency↑
τ = 0.52.442.3242.4%
τ = 1.02.432.3242.6%
τ = 2.02.462.3142.5%
τ = 3.02.472.3242.5%
", + "bbox": [ + 426, + 224, + 689, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/26307c48fd021a9f627ced162ecd8e886ce006f494abd570be27c946c9d07f1a.jpg", + "table_caption": [ + "(a) Learning method for streaming dialogue. Training with streaming dialogue method can achieve much better TimeDiff and Fluency, as well as maintain the language modeling quality. Meanwhile, the streaming dialogue can enjoy much more efficient training than per-frame dialogue for video streaming dialogue." + ], + "table_footnote": [], + "table_body": "
MethodMem↓FPS↑
Interleaved34.4G1.5
Per-frame Streaming24.9G7.5
Streaming18.2G13.5
", + "bbox": [ + 712, + 222, + 885, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e615b43d66f17278869ffc2d2a9ab43f73520801ce7e9bb0c6dda8bd98aa7b4b.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Results on COIN benchmarks (left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal." + ], + "table_body": "
MethodNot use HowTo100MCOIN Benchmark Top-1 Accuracy↑
StepTaskNextProc.Proc.+
ClipBERT [39]✓30.865.4---
TimeSformer [7]✗46.585.334.017.040.1
Paprika [92]✗51.085.843.2--
DistantSup [47]✗54.190.039.4-41.3
VideoTF [58]✗56.591.042.440.246.4
ProcedureVRL [91]✗56.990.846.8--
VideoTaskGraph [5]✗57.290.540.2--
VideoLLM-online✓59.892.148.147.952.9
", + "bbox": [ + 89, + 426, + 504, + 545 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e0406155103279d5189f34a8a5893320d12949d013ecb443970d2ee35892869.jpg", + "table_caption": [ + "Table 1. Ablation experiments on Ego4D Narration Stream. We train VideoLLM-online on Ego4D [26] narration stream training set and evaluate on its validation set. The comparison is based on our designed metrics: the ratio of strictly correct prediction tokens (Fluency), language modeling perplexity (LM-PPL) and time difference (TimeDiff) metrics. #Training Token denotes the average token length during training. TimeDiff refers to difference in second. Default settings are highlighted." + ], + "table_footnote": [ + "(b) Results on Ego4D LTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions." + ], + "table_body": "
MethodNot use EgoVLPEnd-to-end?Ego4D LTA ED@Z=20↓
VerbNounAction
CLIP [17]✓✓0.7390.7690.941
EgoT2 [79]✓✓0.7220.7640.935
I-CVAE [53]✓✓0.7530.7490.931
HierVL [4]✓✓0.7240.7350.928
VideoLLM [10]✗✓0.7210.7250.921
VideoLLM-online✓✓0.6970.6980.897
Palm [32]✗✗0.6960.6510.886
AntGPT [88]✗✗0.6500.6500.877
", + "bbox": [ + 519, + 425, + 880, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Experiments on COIN [71] and Ego4D [26] benchmarks. VideoLLM-online is finetuned on their training set, and strictly evaluated on the test set by generated string comparison with the ground-truth text. It achieves best results among end-to-end models.", + "bbox": [ + 75, + 584, + 892, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation Setting", + "text_level": 1, + "bbox": [ + 76, + 627, + 256, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We use (1) assistance-related, instructional video dataset COIN [71] and (2) continuous, egocentric video dataset Ego4D [26] in various settings:", + "bbox": [ + 75, + 654, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Ego4D Narration Stream: We also leverage the dense Ego4D timestamp-narration to create a streaming set. The goal is to generate narrations timely like Ego4D human annotators [26]. We follow the division of the training, validation, and test set in EgoVLP [45].", + "- COIN+Ego4D Stream: To further evaluate the potential of model's performance to free-form dialogue, we construct a simple COIN+Ego4D Stream set, constructed from COIN annotations using our data generation methods, and the above Ego4D Narration Stream. The model should remind the user when an action starts, summarizes the action when it ends, as well as forecasts the next action. We use the same training/testing splits as COIN" + ], + "bbox": [ + 76, + 704, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "benchmarks. See supplementary material for the results.", + "bbox": [ + 511, + 628, + 883, + 643 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- COIN Benchmarks: Following on previous studies [47, 58, 80, 91], we evaluate our model on six common benchmarks of the COIN dataset: step recognition, step forecasting, task summarization, procedure forecasting, procedure forecasting with a goal.", + "- Ego4D long-term action anticipation (LTA) benchmark: This benchmark requires to predict next $Z = 20$ actions (verbs and nouns) for the given video of previous 8 steps. We use the standard Ego4d v2 splits as in previous studies [32, 88]." + ], + "bbox": [ + 500, + 643, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We use the following metrics to evaluate the model as an online video assistant:", + "bbox": [ + 500, + 799, + 890, + 828 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Language modeling perplexity (LM-PPL). This indicates the language modeling capability at a given times-tamp. A lower LM-PPL is better for correct answering.", + "- Time Difference (TimeDiff). To evaluate the temporal" + ], + "bbox": [ + 500, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "alignment capability of an online assistant, we calculate the discrepancy between the timestamp of its response and the expected timestamp for each response. We average TimeDiff each turn as the metric.", + "bbox": [ + 89, + 90, + 467, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Fluency. Individual LM-PPL or TimeDiff do not entirely evaluate both language and temporal effectiveness in a streaming dialogue. We introduce the Fluency metric, which evaluates the proportion of consecutive successful token prediction within a dialogue turn. As the token also including language tokens, Fluency can comprehensively reflect the aspects of LM-PPL and TimeDiff.", + "bbox": [ + 76, + 152, + 467, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. To our best knowledge, we are the first one to tackle producing temporal aligned, free-form language answer with streaming video settings. To better understand the challenges, we build baseline models for video-text interleaved dialogue, per-frame dialogue, as we described in Section 3.1, with the same model architecture and training details to VideoLLM-online, differing only in their training objective and multi-turn formulation.", + "bbox": [ + 76, + 261, + 467, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 392, + 228, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning Method. Table 1a shows the ablation studies on learning methods in a streaming setting. Both vision-language interleaved and streaming methods exhibit low perplexity loss, indicating that our proposed objective does not hurt language modeling capability. However, learning with per-frame for streaming will produce significant higher LM-PPL than others, which might be attributed to the too more single EOS token in answering that affects the original language modeling.", + "bbox": [ + 76, + 415, + 467, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When we turn to online metrics of TimeDiff and Fluency, streaming dialogue method yields much better results than others. In our observation, the first interleaved dialogue method always outputs language after every frame, and the second multi-turn for streaming dialogue approach tends to answer EOS token after every frame, which decreases their performance for streaming video inputs. Furthermore, per-frame for streaming dialogue method will significantly slow down the training speed due to its lengthy prompts, while our method has no negative impact on the efficiency.", + "bbox": [ + 76, + 551, + 467, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Streaming Loss. We continue to investigate the most suitable strategy to learn the streaming objective. As shown in Table 1b and Table 1c, we find a default setting works surprisingly well (CE loss, $\\tau = 1.0$ ), which demonstrates there is no need to apply more advanced loss (e.g. Focal Loss [46]) to address the imbalance on EOS token.", + "bbox": [ + 76, + 704, + 467, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference Efficiency. In Table 1d, we test the inference efficiency on Ego4D narration stream validation set (5 minute), and report the memory cost and average FPS on a single A100 GPU. The first interleaved dialogue method, which will output language after every video frame, has huge memory cost and slow generation speed. The second one, per-frame dialogue for streaming that formulates", + "bbox": [ + 76, + 795, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all in a multi-turn dialogue, show better efficiency than the first one since it can cost less tokens in redundant frames. However, this approach still lags significantly behind our streaming dialogue approach, which does not cost extra tokens in redundant frames thus maintain smaller key-value cache. We observe for most Ego4D videos, our model can run larger than 10 FPS, providing possibility for AI assistants working in real-time video stream.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Results", + "text_level": 1, + "bbox": [ + 500, + 220, + 591, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Offline language modeling. We show our model can perform well on traditional temporal summarization and forecasting problems. As shown in Table 2(a), our model achieves state-of-the-art performances in step/task summarization and next step/procedure forecasting benchmarks of COIN dataset [71]. Furthermore, we also obtain the best performance among end-to-end models evaluated on Ego4D LTA. We observe the results of Palm [32] and AntGPT [88] are better than us, but they used egocentric pre-trained visual feature [45], and integrates lots of complex cascading methods to improve the forecasting results. Our VideoLLM-online, however, directly outputs language as the results, which performs better than the similar end-to-end VideoLLM [10].", + "bbox": [ + 496, + 243, + 890, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. See Figure 1, we visualize two representative examples, real-time narration and online dialogue. The most distinctive characteristics of our approach are: (1) the dialogue process goes along with the streaming video input, rather than chatting based on the full video. (2) The response will be \"muted\" when it is unnecessary, significantly improving the overall speed of video streaming dialogue.", + "bbox": [ + 496, + 455, + 890, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 573, + 617, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Learning-In-Video-strEam (LIVE), a novel framework empowering LLMs to handle streaming video, to produce temporal aligned answers, hold long-context video duration, and have high inference efficiency. We use LIVE to train a simple VideoLLM-online model, which not only achieves superior capability in online/offline vision-language tasks, but also enable fast inference for an online video streaming setting. We believe enabling such abilities will be an important step to move towards always-on online assistant. In future work, to make our VideoLLM-online be more general and improve its spatial capability in zero-shot prediction for downstream applications, we will explore suitable pre-training data source, and develop models that can employ more spatial tokens but without many trade-off on speed and memory cost.", + "bbox": [ + 496, + 598, + 890, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This work is sponsored by Project Aria Team, Meta. The datasets and processing were acquired and all models were trained at the National University of Singapore (NUS) by NUS authors.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18414", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Huda AlAmri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K. Marks, Chiori Hori, Peter Anderson, Stefan Lee, and Devi Parikh. Audio visual scene-aware dialog. In CVPR, pages 7558-7567, 2019. 2, 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2", + "[3] Rohan Anil, Sebastian Borgeaud, Yonghui Wu, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Slav Petrov, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy P. Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul Ronald Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, and et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2023. 2, 3", + "[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In CVPR, pages 23066-23078, 2023. 7", + "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 7", + "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966, 2023. 2, 3", + "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 7", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In *NeurIPS*, pages 1877–1901, 2020. 1, 2" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 3", + "[10] Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292, 2023. 7, 8", + "[11] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 3", + "[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3", + "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C.H.Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500, 2023. 2", + "[14] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. Int. J. Comput. Vis., 130(1):33-55, 2022. 4", + "[15] Tri Dao, Daniel Haziza, Francisco Massa, and Grigory Sizov. Flash-decoding for long-context inference. https://pytorch.org/blog/flash-decoding/, 2023.3", + "[16] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In CVPR, pages 1080-1089, 2017. 2", + "[17] Srijan Das and Michael S. Ryoo. Video + clip baseline for ego4d long-term action anticipation. arXiv:2207.00579, 2022. 7", + "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2", + "[19] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv:2309.11499, 2023. 2", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5", + "[21] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M. Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. arXiv:2304.14108, 2023. 5, 6", + "[22] Zhe Gan, Yu Cheng, Ahmed El Kholy, Linjie Li, Jingjing Liu, and Jianfeng Gao. Multi-step reasoning via recurrent dual attention for visual dialog. In ACL, pages 6463-6474, 2019. 2", + "[23] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms. arXiv:2310.01801, 2023. 3", + "[24] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In ECCV, pages 269-284, 2016. 3, 5", + "[25] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. Arxiv, 2106.02036, 2021. 3", + "[26] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18973-18990, 2022. 2, 4, 7", + "[27] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, FuJen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abraham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, and et al. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. arXiv:2311.18259, 2023. 1", + "[28] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv:2309.03905, 2023. 2", + "[29] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 3", + "[30] Ke Hong, Guohao Dai, Jiaming Xu, Qiuli Mao, Xiuhong Li, Jun Liu, Kangdi Chen, Hanyu Dong, and Yu Wang. Flashdecoding++: Faster large language model inference on gpus. arXiv:2311.01282, 2023. 3", + "[31] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5, 6", + "[32] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models @ ego4d long-term action anticipation challenge 2023. arXiv:2306.16545, 2023. 7, 8", + "[33] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorbani, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos \"in the wild\". Comput. Vis. Image Underst., 155:1-23, 2017. 5", + "[34] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv:2001.08361, 2020. 1", + "[35] Kris M. Kitani, Brian D. Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In ECCV, pages 201-214, 2012. 3", + "[36] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. arXiv:2309.06180, 2023. 3", + "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv:2308.00692, 2023. 2, 3", + "[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379, 2018. 3", + "[39] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 7", + "[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18416", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[42] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv:2305.06355, 2023. 2, 3", + "[43] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 2, 3", + "[44] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, Ce Liu, and Lijuan Wang. MM-VID: advancing video understanding with gpt-4v(ision). arXiv:2310.19773, 2023. 2, 3", + "[45] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Egocentric video-language pretraining. arXiv:2206.01670, 2022. 4, 7, 8", + "[46] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 7, 8", + "[47] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853, 2022. 7", + "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 5", + "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 2, 5", + "[50] Xiaoxuan Liu, Lanxiang Hu, Peter Bailis, Ion Stoica, Zhijie Deng, Alvin Cheung, and Hao Zhang. Online speculative decoding. arXiv:2310.07177, 2023. 3", + "[51] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv:2306.09093, 2023. 2, 3", + "[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424, 2023. 2, 3", + "[53] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action anticipation. In WACV, pages 6037-6046, 2023. 7", + "[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 4", + "[55] Seungwhan Moon, Andrea Madotto, Zhaojiang Lin, Tushar Nagarajan, Matt Smith, Shashank Jain, Chun-Fu Yeh," + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Prakash Murugesan, Peyman Heidari, Yue Liu, et al. Anymal: An efficient and scalable any-modality augmented language model. arXiv:2309.16058, 2023. 2", + "[56] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv:2305.15021, 2023. 2, 3", + "[57] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for visual dialog: A simple state-of-the-art baseline. In ECCV, pages 336-352, 2020. 2", + "[58] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv:2303.13519, 2023. 7", + "[59] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Efficient attention mechanism for visual dialog that can handle all the interactions between multiple inputs. In ECCV, pages 223–240, 2020. 2", + "[60] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt/, 2023.1, 2", + "[61] OpenAI. GPT-4 technical report. arXiv:2303.08774, 2023. 1", + "[62] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.2.3", + "[63] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 2", + "[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 2", + "[65] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 5, 6", + "[66] Idan Schwartz, Seunghak Yu, Tamir Hazan, and Alexander G. Schwing. Factor graph attention. In CVPR, pages 2039-2048, 2019. 2", + "[67] Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Re, Ion Stoica, and Ce Zhang. Flexgen: High-throughput generative inference of large language models with a singlegpu. arXiv:2303.06865, 2023. 3", + "[68] Abhinav Shrivastava, Abhinav Gupta, and Ross B. Girshick. Training region-based object detectors with online hard example mining. In CVPR, pages 761-769, 2016. 7", + "[69] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv:2307.16449, 2023. 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 5", + "[71] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 2, 4, 5, 7, 8", + "[72] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023. 1, 4", + "[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv:2307.09288, 2023. 2,456", + "[74] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 6000-6010, 2017. 2", + "[75] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Yuanjie Shao, Zhengrong Zuo, Changxin Gao, and Nong Sang. Oadtr: Online action detection with transformers. In ICCV, pages 7545-7555, 2021. 3", + "[76] Yue Wang, Shafiq R. Joty, Michael R. Lyu, Irwin King, Caiming Xiong, and Steven C. H. Hoi. VD-BERT: A unified vision and dialog transformer with BERT. In EMNLP, pages 3325-3338, 2020. 2", + "[77] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv:2309.05519, 2023. 2", + "[78] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv:2309.17453, 2023. 3", + "[79] Zihui Xue, Yale Song, Kristen Grauman, and Lorenzo Torresani. Egocentric video task translation. In CVPR, pages 2310-2320, 2023. 7", + "[80] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Schmid. Unloc: A unified framework for video localization tasks. In ICCV, pages 13623-13633, 2023. 7", + "[81] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2, 3", + "[82] Zhewei Yao, Xiaoxia Wu, Conglong Li, Minjia Zhang, Heyang Qi, Olatunj Ruwase, Ammar Ahmad Awan, Samyam Rajbhandari, and Yuxiong He. Deepspeed-visualchat: Multi-round multi-image interleave chat via multi-modal causal attention. arXiv:2309.14327, 2023. 2, 3", + "[83] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv:2310.07704, 2023. 2, 3", + "[84] Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. Orca: A distributed serving system for {Transformer-Based} generative models. In OSDI, pages 521-538, 2022. 3", + "[85] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv:2309.02591, 2023. 2", + "[86] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858, 2023. 2, 3", + "[87] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 2, 3", + "[88] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? In ICLR, 2024. 3, 7, 8", + "[89] Yue Zhao and Philipp Kähenbuhl. Real-time online video detection with temporal smoothing transformers. In ECCV, pages 485-502, 2022. 3", + "[90] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 4", + "[91] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In CVPR, pages 14825-14835, 2023. 7", + "[92] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 7", + "[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 503, + 92, + 890, + 869 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18418", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_model.json b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b32b814cf27462492faeb677a5dd567ae294c6b5 --- /dev/null +++ b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_model.json @@ -0,0 +1,2776 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "VideoLLM-online: Online Video Large Language Model for Streaming Video" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.18, + 0.23, + 0.199 + ], + "angle": 0, + "content": "Joya Chen1" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.199, + 0.242, + 0.215 + ], + "angle": 0, + "content": "Difei Gao" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.181, + 0.371, + 0.199 + ], + "angle": 0, + "content": "Zhaoyang Lv²" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.199, + 0.369, + 0.216 + ], + "angle": 0, + "content": "Jia-Wei Liu1" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.181, + 0.493, + 0.198 + ], + "angle": 0, + "content": "Shiwei Wu" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.199, + 0.494, + 0.217 + ], + "angle": 0, + "content": "Ziteng Gao" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.181, + 0.689, + 0.199 + ], + "angle": 0, + "content": "Kevin Qinghong Lin1" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.199, + 0.649, + 0.217 + ], + "angle": 0, + "content": "Dongxing Mao" + }, + { + "type": "text", + "bbox": [ + 0.674, + 0.181, + 0.832, + 0.199 + ], + "angle": 0, + "content": "n1 Chenan Song" + }, + { + "type": "text", + "bbox": [ + 0.675, + 0.199, + 0.838, + 0.217 + ], + "angle": 0, + "content": "Mike Zheng Shou" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.22, + 0.793, + 0.238 + ], + "angle": 0, + "content": "1Show Lab, National University of Singapore 2Reality Labs Research, Meta" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.25, + 0.891, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.503, + 0.889, + 0.515 + ], + "angle": 0, + "content": "Streaming Dialogue" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.526, + 0.893, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Zero-shot examples of our VideoLLM-online applied to an egocentric video stream from Ego-Exo4D dataset [27]. Our model is designed for temporally aligned, long-context, real-time dialogue in continuous video streams, shedding light on the future always-on, contextual AI assistants (e.g., smart AR glasses). Model responses are appropriately simplified for better visualization." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.585, + 0.314, + 0.601 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.618, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Recent Large Language Models (LLMs) have been enhanced with vision capabilities, enabling them to comprehend images, videos, and interleaved vision-language content. However, the learning methods of these large multimodal models (LMMs) typically treat videos as predetermined clips, rendering them less effective and efficient at handling streaming video inputs. In this paper, we propose a novel Learning-In-Video-Stream (LIVE) framework, which enables temporally aligned, long-context, and real-time dialogue within a continuous video stream. Our LIVE framework comprises comprehensive approaches to achieve video streaming dialogue, encompassing: (1) a training objective designed to perform language modeling for continuous streaming inputs, (2) a data generation scheme that converts offline temporal annotations into a streaming dialogue format, and (3) an optimized inference pipeline to speed up interactive chat in real-world video streams. With" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.583, + 0.893, + 0.734 + ], + "angle": 0, + "content": "our LIVE framework, we develop a simplified model called VideoLLM-online and demonstrate its significant advantages in processing streaming videos. For instance, our VideoLLM-online-7B model can operate at over 10 FPS on an A100 GPU for a 5-minute video clip from Ego4D narration. Moreover, VideoLLM-online also showcases state-of-the-art performance on public offline video benchmarks, such as recognition, captioning, and forecasting. The code, model, data, and demo have been made available at showlab.github.io/videoollm-online." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.768, + 0.631, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Building the future of always-on contextual AI agent that can promptly answer any human question, digitizing inputs as episodic memories and forecasting future plans given any query in online continuous setting represents a \"holy grail\" mission in AI research. Powered by advancements in large language models (LLMs) [8, 34, 60, 61, 63, 72], recent large multimodal models (LMMs) have unveiled impressive ca" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.212, + 0.9 + ], + "angle": 0, + "content": "Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18407" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "pabilities such as vision-language dialogue [6, 13, 40, 41, 48, 49, 93], spatial understanding [6, 37, 51, 64, 83, 87], processing diverse modalities [19, 28, 55, 77, 85]. Seminal exemplars, like OpenAI's GPT-4V [62], are progressively evolving into highly versatile human AI assistants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.473, + 0.518 + ], + "angle": 0, + "content": "It is time to envision an always-on,contextual, J.A.R.V.I.S-like multimodal assistant that supports freeform user-assistant dialogue within the video stream, which we term \"video streaming dialogue\". Unlike existing LMMs for video understanding (i.e., VideoLLMs) [42, 43, 52, 56, 69, 81, 86] that work offline with manually selected short-video clips, an online assistant should continuously receive video frames with visual content that is constantly refreshed. This paradigm shift presents new challenges. First, the user query may come with temporally aligned requirements (e.g., \"alert me when it's time to flip the steak\"), thus the VideoLLM should scan every incoming frame to avoid event missing, instead of only yielding video-level responses. Second, to answer questions regarding summarization and planning, the VideoLLM must retain the longcontext historical vision and language, which poses huge risk of exceeding maximum context window of LLMs, as well as introduces considerable burden to causal decoding speed and GPU memory. Third, the VideoLLM should generate the answer in real-time, keeping pace with the video stream for always-on scenario. These abilities, however, are even partially overlooked by the most advanced AI assistants [3, 62]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.746 + ], + "angle": 0, + "content": "One possible path towards such an online VideoLLM, inspired by current interleaved vision-language models [3, 6, 40, 62, 82], is to employ a multi-turn dialogue format to achieve per-frame chatting within a video stream. This can be accomplished by facilitating very frequent user interactions, utilizing the visual frame as query at each timestamp to obtain the answer. We follow this to perform prompt engineering for GPT-4V [44, 62], but the results are disappointing: GPT-4V tends to output lengthy content at every frame, leading to significant delays, making it impractical for real-time streaming video. We also explore training baseline models for per-frame chatting. Unfortunately, this approach evidently diminishes the language modeling capability, likely due to harmful language modeling on an excessive number of redundant frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We propose Learning-In-Video-strEam (LIVE), a comprehensive framework that encompasses learning, data, and inference methods to develop an online video assistant. Unlike per-frame dialogue approach, LIVE introduces a novel training objective termed Streaming EOS (End-Of-Sequence) prediction that enables the model to learn when to response or remain silent in a video stream. This objective differs from next-token prediction since EOS tokens here will not appear in the input/output sequence. However, it can work well with the autoregressive loss to train" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "an online VideoLLM. This design reduces unnecessary context, helping the model to manage much longer streaming videos. Nevertheless, the training still requires data from user queries and assistant responses within video streams, which is scarce in popular video datasets usually used for training offline video models. To address this issue, LIVE presents a streaming dialogue generation scheme that converts offline annotations into online dialogues to support free-form chatting. To enhance inference efficiency, LIVE leverages continuous key-value caching for streaming assistance, and parallelizes the fast visual encoding and slow language decoding to prevent bottlenecks, thus moving towards real-time application." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.893, + 0.56 + ], + "angle": 0, + "content": "With LIVE framework, we build a simple VideoLLM-online model upon CLIP [65] vision encoder and Llama2 [73] language model. To evaluate the performance of video streaming dialogue, we utilize the language perplexity metric and design two new metrics to comprehensively assess the model's capabilities in language modeling, temporal responsiveness, and overall fluency. Experiments with real-time Ego4D narration [26] demonstrate that our method shows advantages in all that metrics, with higher speed and lower memory cost. Furthermore, our model achieves state-of-the-art results on numerous offline benchmarks, such as short- and long-term activity recognition and forecasting on the COIN and Ego4D LTA benchmarks [26, 71]. In addition, our model has good speed/memory efficiency, which allows continuous video narration on Ego4D video clips with memory cost less than 20 GB and average speed higher than 10 FPS on a single A100 GPU, paving the way for future real-world usage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.642, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Visual Dialogue. Before transformers [74] become mainstream in vision, visual dialogue methods [1, 16] tend to employ a visually enhanced encoder, with a discriminative head to select candidate answers or a recurrent architecture to generate multi-turn responses. For the encoder, a variety of attention mechanism-based approaches [22, 59, 66] have been proposed to learn the interactions between the image, the answers, and the dialogue history. There have also been explorations of encoder-only BERT [18] models for visual dialogue [57, 76]. However, most of them rely solely on a single image/video at the beginning of the conversation, followed by multi-turn pure language dialogue, which makes them less flexible than the current interleaved vision-language dialogue systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Large Multimodal Models. The advent of large language models (LLMs) [8, 60, 63] has inspired a series of LLM multimodal variants, i.e. large multimodal models (LMMs). Early LMMs [2, 13, 41, 49, 93] achieve image dialogue by projecting the image encoding (e.g., from CLIP [65]) to align with LLM embedding space. Then, lots of ef" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18408" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.227 + ], + "angle": 0, + "content": "forts [3, 6, 40, 62, 82] explore more free-form interleaved vision-text chatting, spatial understanding [6, 11, 12, 37, 51, 83, 87], video comprehension [42, 43, 52, 56, 69, 81, 86], etc. However, when it comes to online scenario, there is less exploration on how LLMs can fulfill the temporal alignment, long-context, and real-time requirements for streaming video inputs. Our research bridges this gap, offering comprehensive solutions across model training, data, and inference to study the problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Online Video Understanding. Typical video understanding benchmarks, such as action recognition [9], temporal action localization [29], video question answering [38], and video dialogue [1], typically allow models to access entire video frames to make predictions, a setting referred to as \"offline\". However, such setting does not align well with many real-time demands (e.g., autonomous driving, AR glasses). Instead, there is a growing focus on \"online\" video understanding problems like online action detection [24] and anticipation [35], which aim to identify the current/future action at each timestamp without seeing the future. Our study pioneers LMMs for online video understanding. Unlike previous online action detection [75, 89] or anticipation models [25, 88], which primarily address one task with a highly customized model, we aim to propose a general solution to achieve free-form dialogue along the online video stream, enabling a model to flexibly handle diverse tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.665 + ], + "angle": 0, + "content": "Efficient Token Decoding. Efficient token decoding for LLMs and LMMs is essential for applying them to real-time online services. To accelerate that, a diversity of strategies has been proposed, such as parallelism on batch dimension [67] or cached key-value sequence [15], computation/memory management optimization [23, 30, 36, 67, 84], and even some lossy approaches [50, 78]. Our focused streaming video scenario has less concerns in large batch processing, but expects faster decoding to avoid excessive frame skipping. We have considered this in training objectives, and further propose some inference schemes to accelerate the decoding efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.17, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.809 + ], + "angle": 0, + "content": "In this section, we present Learning-In-Video-strEam (LIVE) framework that enables LMMs to provide temporally aligned response, handle long-context streaming video, and run efficiently towards real-time usage. We will start from the problem definition of \"video streaming dialogue\", analyze its challenges, and introduce our approach to solve the problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.315, + 0.833 + ], + "angle": 0, + "content": "3.1. Video Streaming Dialogue" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Problem Formulation. Though huge successes have been witnessed in large multimodal models (LMMs), the assistance scenario like a smart AR glass helping the user with cooking, is still far from the capabilities of current LMMs," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "even for the most advanced version, GPT-4V [62]. For example, despite carefully prompting GPT-4V similarly to MMVid [44]—to perform per-frame dialogue for handling streaming video inputs—the redundancy in responses between frames, a limited context window of \\(10\\sim 50\\) frames, and slow speed, collectively render the current GPT-4V unsuitable for online video understanding, as our prompting analysis demonstrates (see supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.893, + 0.349 + ], + "angle": 0, + "content": "To bridge the gap, we define a problem termed \"video streaming dialogue\". Given the context sequence before time \\( t = t_1 \\), denoted as \\( [\\mathrm{Ctx}^{t < t_1}] \\), which may encompass previous vision-language content (e.g., historical user queries, video frames, assistant responses), and an ongoing continuous video stream from \\( t_1 \\) to \\( t_2 \\), denoted as \\( [\\mathrm{Frame}^{t_1 \\leq t \\leq t_2}] \\), our goal is (1) to determine whether the current time \\( t_2 \\) is suitable for language modeling; (2) to carry out language modeling" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.374, + 0.892, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t _ {2}} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {1}} \\right], \\left[ \\mathrm {F} ^ {t _ {1} \\leq t \\leq t _ {2}} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t _ {2}} \\right]\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.451 + ], + "angle": 0, + "content": "if \\( t_2 \\) is determined, where \\( [\\mathrm{Txt}_i^t] \\) denotes the ideal language token in the \\( i \\)-th position at timestamp \\( t \\). [F] is the abbreviation of [Frame]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.481 + ], + "angle": 0, + "content": "In the following, we analyze if existing techniques have been enough to solve the problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.893, + 0.662 + ], + "angle": 0, + "content": "Interleaved/Per-frame Dialogue are Suboptimal. First, we investigate whether the popular approach of interleaved vision-language chatting can address this problem. Related to our formulation above, such a method learns language modeling after given frames between timestamps \\( t_1 \\) and \\( t_2 \\). However, if this approach is adopted during inference, it necessitates the manual selection of timestamps \\( t_1 \\) and \\( t_2 \\), which does not align with the concept of video streaming dialogue. Current VideoLLMs [42, 43, 52, 56, 69, 81, 86] represent a simplified version of this approach, engaging in single- or multi-turn pure language dialogue following video clip inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.738 + ], + "angle": 0, + "content": "If we consider a more free-form format, multi-turn interleaved vision-language chatting [3, 6, 40, 62, 82], we can get a per-frame chatting solution that might be hopeful to solve video streaming dialogue, which performs per-frame language modeling" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.763, + 0.892, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {1}} \\right], \\left[ \\mathrm {F r a m e} ^ {t} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t} \\right]\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "for every frame from timestamp \\( t_1 \\leq t \\leq t_2 \\). Since \\( t_2 \\) is necessary to output answer, so we can simply learn short text (e.g., \"this is not the time to answer\") between \\( t_1 \\leq t < t_2 \\). However, this approach imposes a significant burden on processing speed. For every frame, performing the slow, recurrent, and lengthy next-token prediction with a billion-scale language model makes it extremely hard to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18409" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "achieve real-time video streaming dialogue. Then, this will lead to unavoidable frame skipping, which is unexpected and detrimental for temporal alignment. Furthermore, it taxes the limited context window of the language model, presenting challenges for modeling long contexts and managing GPU memory efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.363 + ], + "angle": 0, + "content": "Streaming EOS Prediction. To solve the problem mentioned above, we first consider a more efficient per-frame chatting method: simply assigning the End-of-Sequence (EOS) token as the content for chatting between \\( t_1 \\leq t < t_2 \\). However, this approach remains suboptimal. The dialogue prompt template (e.g., [INST], [/INST], space tokens in Llama [72, 73]) still consumes a considerable number of tokens per frame, which is less favorable due to the numerous frames in streaming video. Furthermore, the excessive number of EOS tokens in the sequence can significantly increase the language model's perplexity, as we have observed in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Instead, we propose a novel training objective named \"streaming EOS prediction\" to address this issue. We still assume \\( t_2 \\) is essential for decoding language; thus, we normally learn language modeling here:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.448, + 0.469, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t _ {2}} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {2}} \\right], \\left[ \\mathrm {F r a m e} ^ {t _ {2}} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t _ {2}} \\right]\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.526 + ], + "angle": 0, + "content": "However, for timestamps \\( t_1 \\leq t < t_2 \\), which are redundant for producing answers, we directly learn the model to predict EOS token on the frame tokens, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.551, + 0.469, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\max P (\\text {E O S} \\left| [ \\text {C t x} ^ {< t} ], [ \\text {F r a m e} ^ {t} ]\\right), \\text {w h e r e} t _ {1} \\leq t < t _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In this way, we \"skip\" a dialogue turn and learn to determine when it is appropriate to decode language for streaming inputs. During inference, if EOS is predicted on a frame, then we can directly ask the next frame to input. Meanwhile, the EOS token is not appended to the context to prevent it from affecting the language modeling. Therefore, this task is not about next-token prediction; however, it can work with autoregressive loss to train a video streaming dialogue model. In the following, we first introduce how can we get the data of streaming video and timestamped language annotations, then present the details about the model and the training procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We also note that the EOS token mentioned here is not limited to the real EOS token used in language models (e.g., \\( \\langle / \\mathbb{S} \\rangle \\) in Llama). It is permissible to use any token or to introduce a new token, provided that it is specified in the system prompt. We use this term solely for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.151, + 0.861 + ], + "angle": 0, + "content": "3.2.Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Online Annotations to Video Streaming Dialogue. Some video datasets, such as Ego4D narrations [26], are inher" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.891, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.452, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Figure 2. The streaming dialogue data generation method in our LIVE framework. We randomly insert templated questions into the video timeline and \"expose\" the ground-truth video annotations (along with their timestamps) to LLMs, prompting them to answer the queries within a period of time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.763 + ], + "angle": 0, + "content": "ently collected in a streaming manner, with annotators providing real-time narrations as they watch a 5-minute long video clip. However, prior research [45, 90] has predominantly focused on learning from short, discrete clips (e.g. only 32 frames), rather than in a continuous streaming context. For this dataset, we followed the same instructions provided to human annotators [26] as our model training prompt. This prompt instructs the model to simulate human annotators to streamingly generate narrations in a 5-minute video (about 600 frames in 2 FPS). For demo purposes (not for experiments), we also utilize Llama-2-13B-Chat [73] to rephrase the narration text, correcting grammatical errors and typos, and converting it into a more understandable version (e.g., changing \"C does...\" to \"You do...\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Offline Annotations to Video Streaming Dialogue. Despite the Ego4D narration data being collected in a streaming manner, most prevalent video datasets [14, 26, 54, 71] are used to train offline models and only feature temporal segment annotations paired with basic language descriptions (e.g., activities, narrations). To bridge this gap, we propose a method for synthesizing dialogue data from these sources. As shown in Figure 2, our key idea is use LLM to generate user-assistant dialogues based on video annota" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18410" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.317, + 0.107 + ], + "angle": 0, + "content": "tions, involving the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.117, + 0.468, + 0.178 + ], + "angle": 0, + "content": "- First, we prepare a question template library containing various queries about the past, present, and future tenses of the video, totaling \\( N \\) queries. We randomly sample one question from the library, denoted as \\( Q_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.178, + 0.468, + 0.298 + ], + "angle": 0, + "content": "- Then, we obtain the video annotation timeline from the offline dataset. This usually includes timestamped language descriptions, which we organize into a language prompt, e.g., \"time \\( t_a \\sim t_b \\): boiling the water; time \\( t_c \\sim t_d \\): cutting the vegetables.\", denoted as \\( A \\). We consider all the state change critical timestamps as the ideal response times. For this example, \\( t_a, t_b, t_c \\), and \\( t_d \\) are all considered response times." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.299, + 0.468, + 0.374 + ], + "angle": 0, + "content": "- Third, we prompt the large language model to generate responses at every critical timestamp, e.g., \\( t_a, t_b, t_c, t_d \\), according to \\( Q_i \\) and \\( A \\). We can repeat this procedure for each \\( Q_i \\) until all queries have been processed. The responses are saved for loading during training." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.375, + 0.468, + 0.48 + ], + "angle": 0, + "content": "- Finally, during training, we (1) randomly sample a query and load its responses at critical timestamps, (2) randomly insert a query into a video timestamp \\( t_r \\), (3) discard the responses that occur before \\( t_r \\), and add a response at \\( t_r \\). Here different queries can be inserted into one video, which only requires discarding the responses of the previous query after the new query insertion timestamp." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.117, + 0.468, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.468, + 0.732 + ], + "angle": 0, + "content": "In this way, we can generate temporally varied and freeform dialogue data within a video stream. We have prepared 50 questions each for past, current, and future events, totaling \\(N = 150\\) queries. We use Llama-2-13B-Chat to generate the responses and insert a maximum of 3 queries per training sample. The offline datasets we used are COIN [71] and Ego4D GoalStep [70] (for demo usage), which belong to the categories of egocentric and instructional video datasets, aligning with our aim to develop online video assistants. Here, we do not consider online action detection benchmarks (e.g., THUMOS14 [33], TVSeries [24]) because they are closed-set online classification benchmarks, and their labels are too brief, which may cause language models to generate hallucinatory responses. Please refer to the supplementary material for the generated dialogue example." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.235, + 0.769 + ], + "angle": 0, + "content": "3.3. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Model Architecture. We illustrate the model architecture in Figure 3. Similar to LLaVA [48, 49], it comprises three key components: an image encoder, an MLP projector, and a language model. For the image encoder, we utilize the CLIP ViT-L [20, 65] encoder (pretrained on DataComp1B [21]) to extract video frame embeddings at 2 FPS. Each video frame embedding has a shape of \\((1 + h_p\\times w_p)\\times c\\) where \\((1 + h_{p}\\times w_{p})\\) denotes the CLS token and average" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "pooled spatial tokens. The extracted frame embeddings from the image encoder are then fed into MLP projector to frame tokens, as in LLaVA-1.5 [48]. Then frame tokens are interleaved with language tokens as input to an LLM, Llama-2-7B-Chat [73]. Finally, we incorporate LoRA [31] in every linear layer of the LLM for efficient tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Training Loss. As described in Section 3.1, our learning objective is twofold. The first part focuses on autoregressive language modeling, aiming to maximize the joint probability of input text sequences. The second training objective involves streaming EOS prediction, which requires the model to remain silent when it is unnecessary to output responses. With these two training objectives, we have language modeling (LM) loss and streaming loss terms to minimize, both employing cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.34, + 0.892, + 0.39 + ], + "angle": 0, + "content": "\\[\nL = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\left(\\underbrace {- \\log l _ {j + 1} P _ {j} ^ {[ T x t _ {j + 1} ]}} _ {L M L o s s} - \\underbrace {w \\log f _ {j} P _ {j} ^ {[ E O S ]}} _ {\\text {S t r e a m i n g L o s s}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.567 + ], + "angle": 0, + "content": "where \\( l_{j} \\) and \\( f_{j} \\) denote condition indicators. \\( l_{j} \\) is 1 if the \\( j \\)-th token is a language response token, and 0 otherwise. \\( f_{j} \\) is 1 if (1) the \\( j \\)-th token is the last token of a frame\\(^3\\), and (2) \\( l_{j+1} = 0 \\). In essence, the streaming EOS loss is applied to frames before responding. \\( P_{j}^{[\\mathrm{Txt}_{j+1}]} \\) denotes the probability on \\( j + 1 \\)-th text token, output from the language model head of the \\( j \\)-th token, and \\( P_{j}^{[\\mathrm{EOS}]} \\) represents that probability for the EOS token. \\( w \\) is a balance term, set to \\( w = 1 \\) by default. As shown in Figuer 3, we visualize the ranges of language loss and streaming loss in an input sequence when we only use 1 token for each frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.61, + 0.59 + ], + "angle": 0, + "content": "3.4. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Probability Correction. The prevalence of EOS token will bias the model towards EOS token prediction. To address this, we introduce a threshold \\(\\theta\\) to correct the output probability on frame tokens: EOS will not be considered as the next token if \\(P_{j}^{[\\mathrm{EOS}]} < \\theta\\). In practical usage, we find that setting \\(\\theta\\) to 0.5 yields much better results than no threshold here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Continuous Key-Value Cache. As shown in Figure 4, During inference, the video is input as a frame-by-frame stream, with a default FPS 2. Our model takes the current frame as input and generates tokens on-the-fly. In the whole process, we use the key-value cache trick to accelerate token decoding, thus we do not need to manually append the generated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.803, + 0.892, + 0.887 + ], + "angle": 0, + "content": "2The experiments described in this paper are conducted without extra spatial tokens (i.e., \\( h_p = w_p = 0 \\)), which is the most efficient setup and can handle half-hour videos within a 4096 context window. Our released models for demo usage include \\( 1 + h_p \\times w_p = 1 + 3 \\times 3 = 10 \\) tokens, offering better detail in dialogue but supporting shorter maximum video lengths. Despite more tokens used per frame, all models can still run at over 10 FPS for 5-minute Ego4D narration streams." + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.888, + 0.867, + 0.9 + ], + "angle": 0, + "content": "3When a frame has multiple patch tokens, loss is only on the last one." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.803, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.87, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 3. The training method in our LIVE framework. We organize the user-assistant dialogue data and video frames in temporal order as the input sequence. To learn the model when to answer or keep silent in a video stream, we employ not only the standard language modeling (LM) loss but also introduce a streaming EOS prediction loss. This additional loss supervises the model when it is necessary to generate language, enabling it to produce temporally aligned responses and reduces the redundant dialogue history." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.524, + 0.466, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.745 + ], + "angle": 0, + "content": "Figure 4. Inference pipeline in our LIVE framework. During inference, video frames serve as streaming inputs. Our model maintains a continuous key-value cache as the input progresses to speed up the inference. Furthermore, we parallelize the fast video frame encoder and the slower language model to avoid the bottleneck in the latter. Video frame tokens can be always encoded and buffered, no need to wait the language decoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.469, + 0.838 + ], + "angle": 0, + "content": "tokens for the next frame. As our training encourages the model to keep silence, this continuous inference would be efficient, providing the possibility to pace with the video stream speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Parallelization of encoding and decoding. Our video frame encoder utilizes CLIP ViT-L (307M), which is significantly smaller than the 7B LLM. This size discrepancy leads to a speed mismatch, potentially resulting in frame" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.892, + 0.604 + ], + "angle": 0, + "content": "skipping when the LLM decodes long sentences. To mitigate this issue, we parallelize the processes and establish a buffer for video frame tokens. Once the language model completes its decoding, it fetches all frame tokens from the buffer and inputs them in batch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.633, + 0.667 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.685, + 0.718, + 0.701 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We implement VideoLLM-online-7B with our LIVE framework. It uses OpenCLIP-ViT-L [21, 65] as the video frame encoder, 2-layer MLP as multimodal connector, and Llama2-7B-Chat [73] as the language model. We only use 1 token for each video frame, sampled by 2 FPS in the following experiments. We add LoRA [31] to all linear layers of Llama-2, with a rank of 128, scaling factor of 256. For video streaming dialogue experiments, we train 2 epochs for the model. For the downstreaming offline experiments, we train 5 epochs without pre-training for fair comparison with previous methods. By default, we set streaming loss weight \\( w = 1.0 \\) during training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18412" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.09, + 0.874, + 0.184 + ], + "angle": 0, + "content": "
MethodTraining ObjectiveEgo4D Narration Stream on Validation Set#Training Token↓Training Cost
LM-PPL↓TimeDiff ↓Fluency↑
No Training498.56.500.1%n/an/a
Interleaved DialogueLanguage Modeling2.456.4711.1%169412h
Per-frame Dialogue for StreamingLanguage Modeling (w/ EOS turns)3.342.5237.7 %673722h
Streaming Dialogue (Ours)Language Modeling + Streaming EOS2.432.3242.6%169412h
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.186, + 0.893, + 0.224 + ], + "angle": 0, + "content": "(a) Learning method for streaming dialogue. Training with streaming dialogue method can achieve much better TimeDiff and Fluency, as well as maintain the language modeling quality. Meanwhile, the streaming dialogue can enjoy much more efficient training than per-frame dialogue for video streaming dialogue." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.228, + 0.391, + 0.296 + ], + "angle": 0, + "content": "
Streaming LossEgo4DNarrationStreamValidation
LM-PPL↓TimeDiff↓Fluency↑
Standard CE2.432.3242.6%
OHEM [68]2.532.3941.0%
Focal Loss [46]2.592.4439.4%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.298, + 0.405, + 0.335 + ], + "angle": 0, + "content": "(b) Streaming loss function. Standard CE (cross-entropy) is enough for training streaming dialogue; there is no need to specifically to address the class imbalance on EOS token." + }, + { + "type": "table", + "bbox": [ + 0.427, + 0.225, + 0.691, + 0.3 + ], + "angle": 0, + "content": "
Weight τEgo4D Narration Stream Validation
LM-PPL↓TimeDiff↓Fluency↑
τ = 0.52.442.3242.4%
τ = 1.02.432.3242.6%
τ = 2.02.462.3142.5%
τ = 3.02.472.3242.5%
" + }, + { + "type": "table", + "bbox": [ + 0.714, + 0.223, + 0.886, + 0.3 + ], + "angle": 0, + "content": "
MethodMem↓FPS↑
Interleaved34.4G1.5
Per-frame Streaming24.9G7.5
Streaming18.2G13.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.404, + 0.303, + 0.89, + 0.34 + ], + "angle": 0, + "content": "(c) Streaming loss weight. Using slightly higher stream- (d) Generation memory/speed. ing loss weight \\((\\tau = 2.0)\\) achieves the best trade-off Streaming dialogue method has among various metrics. much better efficiency." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.353, + 0.893, + 0.412 + ], + "angle": 0, + "content": "Table 1. Ablation experiments on Ego4D Narration Stream. We train VideoLLM-online on Ego4D [26] narration stream training set and evaluate on its validation set. The comparison is based on our designed metrics: the ratio of strictly correct prediction tokens (Fluency), language modeling perplexity (LM-PPL) and time difference (TimeDiff) metrics. #Training Token denotes the average token length during training. TimeDiff refers to difference in second. Default settings are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.427, + 0.506, + 0.546 + ], + "angle": 0, + "content": "
MethodNot use HowTo100MCOIN Benchmark Top-1 Accuracy↑
StepTaskNextProc.Proc.+
ClipBERT [39]✓30.865.4---
TimeSformer [7]✗46.585.334.017.040.1
Paprika [92]✗51.085.843.2--
DistantSup [47]✗54.190.039.4-41.3
VideoTF [58]✗56.591.042.440.246.4
ProcedureVRL [91]✗56.990.846.8--
VideoTaskGraph [5]✗57.290.540.2--
VideoLLM-online✓59.892.148.147.952.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.079, + 0.547, + 0.512, + 0.573 + ], + "angle": 0, + "content": "(a) Results on COIN benchmarks (left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.426, + 0.882, + 0.547 + ], + "angle": 0, + "content": "
MethodNot use EgoVLPEnd-to-end?Ego4D LTA ED@Z=20↓
VerbNounAction
CLIP [17]✓✓0.7390.7690.941
EgoT2 [79]✓✓0.7220.7640.935
I-CVAE [53]✓✓0.7530.7490.931
HierVL [4]✓✓0.7240.7350.928
VideoLLM [10]✗✓0.7210.7250.921
VideoLLM-online✓✓0.6970.6980.897
Palm [32]✗✗0.6960.6510.886
AntGPT [88]✗✗0.6500.6500.877
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.548, + 0.888, + 0.573 + ], + "angle": 0, + "content": "(b) Results on Ego4D LTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.585, + 0.893, + 0.613 + ], + "angle": 0, + "content": "Table 2. Experiments on COIN [71] and Ego4D [26] benchmarks. VideoLLM-online is finetuned on their training set, and strictly evaluated on the test set by generated string comparison with the ground-truth text. It achieves best results among end-to-end models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.628, + 0.257, + 0.645 + ], + "angle": 0, + "content": "4.2. Evaluation Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Datasets. We use (1) assistance-related, instructional video dataset COIN [71] and (2) continuous, egocentric video dataset Ego4D [26] in various settings:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.705, + 0.469, + 0.78 + ], + "angle": 0, + "content": "- Ego4D Narration Stream: We also leverage the dense Ego4D timestamp-narration to create a streaming set. The goal is to generate narrations timely like Ego4D human annotators [26]. We follow the division of the training, validation, and test set in EgoVLP [45]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- COIN+Ego4D Stream: To further evaluate the potential of model's performance to free-form dialogue, we construct a simple COIN+Ego4D Stream set, constructed from COIN annotations using our data generation methods, and the above Ego4D Narration Stream. The model should remind the user when an action starts, summarizes the action when it ends, as well as forecasts the next action. We use the same training/testing splits as COIN" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.885, + 0.644 + ], + "angle": 0, + "content": "benchmarks. See supplementary material for the results." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.645, + 0.892, + 0.719 + ], + "angle": 0, + "content": "- COIN Benchmarks: Following on previous studies [47, 58, 80, 91], we evaluate our model on six common benchmarks of the COIN dataset: step recognition, step forecasting, task summarization, procedure forecasting, procedure forecasting with a goal." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.72, + 0.892, + 0.795 + ], + "angle": 0, + "content": "- Ego4D long-term action anticipation (LTA) benchmark: This benchmark requires to predict next \\( Z = 20 \\) actions (verbs and nouns) for the given video of previous 8 steps. We use the standard Ego4d v2 splits as in previous studies [32, 88]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.645, + 0.892, + 0.795 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.8, + 0.891, + 0.829 + ], + "angle": 0, + "content": "Evaluation metrics. We use the following metrics to evaluate the model as an online video assistant:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.891, + 0.886 + ], + "angle": 0, + "content": "- Language modeling perplexity (LM-PPL). This indicates the language modeling capability at a given times-tamp. A lower LM-PPL is better for correct answering." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Time Difference (TimeDiff). To evaluate the temporal" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18413" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "alignment capability of an online assistant, we calculate the discrepancy between the timestamp of its response and the expected timestamp for each response. We average TimeDiff each turn as the metric." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.153, + 0.468, + 0.256 + ], + "angle": 0, + "content": "- Fluency. Individual LM-PPL or TimeDiff do not entirely evaluate both language and temporal effectiveness in a streaming dialogue. We introduce the Fluency metric, which evaluates the proportion of consecutive successful token prediction within a dialogue turn. As the token also including language tokens, Fluency can comprehensively reflect the aspects of LM-PPL and TimeDiff." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.262, + 0.468, + 0.381 + ], + "angle": 0, + "content": "Baselines. To our best knowledge, we are the first one to tackle producing temporal aligned, free-form language answer with streaming video settings. To better understand the challenges, we build baseline models for video-text interleaved dialogue, per-frame dialogue, as we described in Section 3.1, with the same model architecture and training details to VideoLLM-online, differing only in their training objective and multi-turn formulation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.393, + 0.229, + 0.409 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.468, + 0.551 + ], + "angle": 0, + "content": "Learning Method. Table 1a shows the ablation studies on learning methods in a streaming setting. Both vision-language interleaved and streaming methods exhibit low perplexity loss, indicating that our proposed objective does not hurt language modeling capability. However, learning with per-frame for streaming will produce significant higher LM-PPL than others, which might be attributed to the too more single EOS token in answering that affects the original language modeling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.468, + 0.704 + ], + "angle": 0, + "content": "When we turn to online metrics of TimeDiff and Fluency, streaming dialogue method yields much better results than others. In our observation, the first interleaved dialogue method always outputs language after every frame, and the second multi-turn for streaming dialogue approach tends to answer EOS token after every frame, which decreases their performance for streaming video inputs. Furthermore, per-frame for streaming dialogue method will significantly slow down the training speed due to its lengthy prompts, while our method has no negative impact on the efficiency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Streaming Loss. We continue to investigate the most suitable strategy to learn the streaming objective. As shown in Table 1b and Table 1c, we find a default setting works surprisingly well (CE loss, \\(\\tau = 1.0\\)), which demonstrates there is no need to apply more advanced loss (e.g. Focal Loss [46]) to address the imbalance on EOS token." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Inference Efficiency. In Table 1d, we test the inference efficiency on Ego4D narration stream validation set (5 minute), and report the memory cost and average FPS on a single A100 GPU. The first interleaved dialogue method, which will output language after every video frame, has huge memory cost and slow generation speed. The second one, per-frame dialogue for streaming that formulates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "all in a multi-turn dialogue, show better efficiency than the first one since it can cost less tokens in redundant frames. However, this approach still lags significantly behind our streaming dialogue approach, which does not cost extra tokens in redundant frames thus maintain smaller key-value cache. We observe for most Ego4D videos, our model can run larger than 10 FPS, providing possibility for AI assistants working in real-time video stream." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.221, + 0.593, + 0.235 + ], + "angle": 0, + "content": "4.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.454 + ], + "angle": 0, + "content": "Offline language modeling. We show our model can perform well on traditional temporal summarization and forecasting problems. As shown in Table 2(a), our model achieves state-of-the-art performances in step/task summarization and next step/procedure forecasting benchmarks of COIN dataset [71]. Furthermore, we also obtain the best performance among end-to-end models evaluated on Ego4D LTA. We observe the results of Palm [32] and AntGPT [88] are better than us, but they used egocentric pre-trained visual feature [45], and integrates lots of complex cascading methods to improve the forecasting results. Our VideoLLM-online, however, directly outputs language as the results, which performs better than the similar end-to-end VideoLLM [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Visualization. See Figure 1, we visualize two representative examples, real-time narration and online dialogue. The most distinctive characteristics of our approach are: (1) the dialogue process goes along with the streaming video input, rather than chatting based on the full video. (2) The response will be \"muted\" when it is unnecessary, significantly improving the overall speed of video streaming dialogue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.619, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.826 + ], + "angle": 0, + "content": "We propose Learning-In-Video-strEam (LIVE), a novel framework empowering LLMs to handle streaming video, to produce temporal aligned answers, hold long-context video duration, and have high inference efficiency. We use LIVE to train a simple VideoLLM-online model, which not only achieves superior capability in online/offline vision-language tasks, but also enable fast inference for an online video streaming setting. We believe enabling such abilities will be an important step to move towards always-on online assistant. In future work, to make our VideoLLM-online be more general and improve its spatial capability in zero-shot prediction for downstream applications, we will explore suitable pre-training data source, and develop models that can employ more spatial tokens but without many trade-off on speed and memory cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment This work is sponsored by Project Aria Team, Meta. The datasets and processing were acquired and all models were trained at the National University of Singapore (NUS) by NUS authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18414" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Huda AlAmri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K. Marks, Chiori Hori, Peter Anderson, Stefan Lee, and Devi Parikh. Audio visual scene-aware dialog. In CVPR, pages 7558-7567, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.547 + ], + "angle": 0, + "content": "[3] Rohan Anil, Sebastian Borgeaud, Yonghui Wu, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Slav Petrov, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy P. Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul Ronald Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, and et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.549, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In CVPR, pages 23066-23078, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.763, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In *NeurIPS*, pages 1877–1901, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[9] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[10] Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[11] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C.H.Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[14] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. Int. J. Comput. Vis., 130(1):33-55, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[15] Tri Dao, Daniel Haziza, Francisco Massa, and Grigory Sizov. Flash-decoding for long-context inference. https://pytorch.org/blog/flash-decoding/, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[16] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In CVPR, pages 1080-1089, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[17] Srijan Das and Michael S. Ryoo. Video + clip baseline for ego4d long-term action anticipation. arXiv:2207.00579, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[19] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv:2309.11499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M. Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18415" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.188 + ], + "angle": 0, + "content": "Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. arXiv:2304.14108, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[22] Zhe Gan, Yu Cheng, Ahmed El Kholy, Linjie Li, Jingjing Liu, and Jianfeng Gao. Multi-step reasoning via recurrent dual attention for visual dialog. In ACL, pages 6463-6474, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[23] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms. arXiv:2310.01801, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[24] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In ECCV, pages 269-284, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[25] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. Arxiv, 2106.02036, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[26] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18973-18990, 2022. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[27] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, FuJen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abraham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, and et al. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. arXiv:2311.18259, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[28] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv:2309.03905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[29] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[30] Ke Hong, Guohao Dai, Jiaming Xu, Qiuli Mao, Xiuhong Li, Jun Liu, Kangdi Chen, Hanyu Dong, and Yu Wang. Flashdecoding++: Faster large language model inference on gpus. arXiv:2311.01282, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[31] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[32] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models @ ego4d long-term action anticipation challenge 2023. arXiv:2306.16545, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[33] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorbani, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos \"in the wild\". Comput. Vis. Image Underst., 155:1-23, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[34] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv:2001.08361, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[35] Kris M. Kitani, Brian D. Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In ECCV, pages 201-214, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[36] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. arXiv:2309.06180, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv:2308.00692, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[39] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18416" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[42] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv:2305.06355, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.469, + 0.251 + ], + "angle": 0, + "content": "[43] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[44] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, Ce Liu, and Lijuan Wang. MM-VID: advancing video understanding with gpt-4v(ision). arXiv:2310.19773, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[45] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Egocentric video-language pretraining. arXiv:2206.01670, 2022. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.409, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[46] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[47] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[50] Xiaoxuan Liu, Lanxiang Hu, Peter Bailis, Ion Stoica, Zhijie Deng, Alvin Cheung, and Hao Zhang. Online speculative decoding. arXiv:2310.07177, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[51] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv:2306.09093, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[53] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action anticipation. In WACV, pages 6037-6046, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[55] Seungwhan Moon, Andrea Madotto, Zhaojiang Lin, Tushar Nagarajan, Matt Smith, Shashank Jain, Chun-Fu Yeh," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Prakash Murugesan, Peyman Heidari, Yue Liu, et al. Anymal: An efficient and scalable any-modality augmented language model. arXiv:2309.16058, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[56] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv:2305.15021, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[57] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for visual dialog: A simple state-of-the-art baseline. In ECCV, pages 336-352, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[58] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv:2303.13519, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[59] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Efficient attention mechanism for visual dialog that can handle all the interactions between multiple inputs. In ECCV, pages 223–240, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[60] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt/, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[61] OpenAI. GPT-4 technical report. arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[62] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.2.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[63] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[65] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[66] Idan Schwartz, Seunghak Yu, Tamir Hazan, and Alexander G. Schwing. Factor graph attention. In CVPR, pages 2039-2048, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[67] Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Re, Ion Stoica, and Ce Zhang. Flexgen: High-throughput generative inference of large language models with a singlegpu. arXiv:2303.06865, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[68] Abhinav Shrivastava, Abhinav Gupta, and Ross B. Girshick. Training region-based object detectors with online hard example mining. In CVPR, pages 761-769, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[69] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv:2307.16449, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18417" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[70] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[71] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[72] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.569 + ], + "angle": 0, + "content": "[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv:2307.09288, 2023. 2,456" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.47, + 0.626 + ], + "angle": 0, + "content": "[74] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 6000-6010, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.47, + 0.683 + ], + "angle": 0, + "content": "[75] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Yuanjie Shao, Zhengrong Zuo, Changxin Gao, and Nong Sang. Oadtr: Online action detection with transformers. In ICCV, pages 7545-7555, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[76] Yue Wang, Shafiq R. Joty, Michael R. Lyu, Irwin King, Caiming Xiong, and Steven C. H. Hoi. VD-BERT: A unified vision and dialog transformer with BERT. In EMNLP, pages 3325-3338, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[77] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv:2309.05519, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.47, + 0.827 + ], + "angle": 0, + "content": "[78] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv:2309.17453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[79] Zihui Xue, Yale Song, Kristen Grauman, and Lorenzo Torresani. Egocentric video task translation. In CVPR, pages 2310-2320, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[80] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Schmid. Unloc: A unified framework for video localization tasks. In ICCV, pages 13623-13633, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[81] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[82] Zhewei Yao, Xiaoxia Wu, Conglong Li, Minjia Zhang, Heyang Qi, Olatunj Ruwase, Ammar Ahmad Awan, Samyam Rajbhandari, and Yuxiong He. Deepspeed-visualchat: Multi-round multi-image interleave chat via multi-modal causal attention. arXiv:2309.14327, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[83] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv:2310.07704, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[84] Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. Orca: A distributed serving system for {Transformer-Based} generative models. In OSDI, pages 521-538, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[85] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv:2309.02591, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[86] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.503, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[87] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[88] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? In ICLR, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[89] Yue Zhao and Philipp Kähenbuhl. Real-time online video detection with temporal smoothing transformers. In ECCV, pages 485-502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[90] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[91] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In CVPR, pages 14825-14835, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[92] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.814, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18418" + } + ] +] \ No newline at end of file diff --git a/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_origin.pdf b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..064e4b4731d7243aa84eb54baedfa8080778a454 --- /dev/null +++ b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/c25a98ca-87a4-45af-9e10-acd055b937ac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c9d82e0ea41a70fb9a11353f282f1657cc2b9271ab236b70f007f9379e452d7 +size 2207504 diff --git a/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/full.md b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..650a1253bd6ff6597b63a40ba730f3f7e4139087 --- /dev/null +++ b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/full.md @@ -0,0 +1,349 @@ +# VideoLLM-online: Online Video Large Language Model for Streaming Video + +Joya Chen1 + +Difei Gao + +Zhaoyang Lv² + +Jia-Wei Liu1 + +Shiwei Wu + +Ziteng Gao + +Kevin Qinghong Lin1 + +Dongxing Mao + +n1 Chenan Song + +Mike Zheng Shou + +1Show Lab, National University of Singapore 2Reality Labs Research, Meta + +![](images/55bcd345351946a74795ffc9d05af932569b03fe66e2cd880067414b9227d9de.jpg) +Streaming Dialogue +Figure 1. Zero-shot examples of our VideoLLM-online applied to an egocentric video stream from Ego-Exo4D dataset [27]. Our model is designed for temporally aligned, long-context, real-time dialogue in continuous video streams, shedding light on the future always-on, contextual AI assistants (e.g., smart AR glasses). Model responses are appropriately simplified for better visualization. + +# Abstract + +Recent Large Language Models (LLMs) have been enhanced with vision capabilities, enabling them to comprehend images, videos, and interleaved vision-language content. However, the learning methods of these large multimodal models (LMMs) typically treat videos as predetermined clips, rendering them less effective and efficient at handling streaming video inputs. In this paper, we propose a novel Learning-In-Video-Stream (LIVE) framework, which enables temporally aligned, long-context, and real-time dialogue within a continuous video stream. Our LIVE framework comprises comprehensive approaches to achieve video streaming dialogue, encompassing: (1) a training objective designed to perform language modeling for continuous streaming inputs, (2) a data generation scheme that converts offline temporal annotations into a streaming dialogue format, and (3) an optimized inference pipeline to speed up interactive chat in real-world video streams. With + +our LIVE framework, we develop a simplified model called VideoLLM-online and demonstrate its significant advantages in processing streaming videos. For instance, our VideoLLM-online-7B model can operate at over 10 FPS on an A100 GPU for a 5-minute video clip from Ego4D narration. Moreover, VideoLLM-online also showcases state-of-the-art performance on public offline video benchmarks, such as recognition, captioning, and forecasting. The code, model, data, and demo have been made available at showlab.github.io/videoollm-online. + +# 1. Introduction + +Building the future of always-on contextual AI agent that can promptly answer any human question, digitizing inputs as episodic memories and forecasting future plans given any query in online continuous setting represents a "holy grail" mission in AI research. Powered by advancements in large language models (LLMs) [8, 34, 60, 61, 63, 72], recent large multimodal models (LMMs) have unveiled impressive ca + +pabilities such as vision-language dialogue [6, 13, 40, 41, 48, 49, 93], spatial understanding [6, 37, 51, 64, 83, 87], processing diverse modalities [19, 28, 55, 77, 85]. Seminal exemplars, like OpenAI's GPT-4V [62], are progressively evolving into highly versatile human AI assistants. + +It is time to envision an always-on,contextual, J.A.R.V.I.S-like multimodal assistant that supports freeform user-assistant dialogue within the video stream, which we term "video streaming dialogue". Unlike existing LMMs for video understanding (i.e., VideoLLMs) [42, 43, 52, 56, 69, 81, 86] that work offline with manually selected short-video clips, an online assistant should continuously receive video frames with visual content that is constantly refreshed. This paradigm shift presents new challenges. First, the user query may come with temporally aligned requirements (e.g., "alert me when it's time to flip the steak"), thus the VideoLLM should scan every incoming frame to avoid event missing, instead of only yielding video-level responses. Second, to answer questions regarding summarization and planning, the VideoLLM must retain the longcontext historical vision and language, which poses huge risk of exceeding maximum context window of LLMs, as well as introduces considerable burden to causal decoding speed and GPU memory. Third, the VideoLLM should generate the answer in real-time, keeping pace with the video stream for always-on scenario. These abilities, however, are even partially overlooked by the most advanced AI assistants [3, 62]. + +One possible path towards such an online VideoLLM, inspired by current interleaved vision-language models [3, 6, 40, 62, 82], is to employ a multi-turn dialogue format to achieve per-frame chatting within a video stream. This can be accomplished by facilitating very frequent user interactions, utilizing the visual frame as query at each timestamp to obtain the answer. We follow this to perform prompt engineering for GPT-4V [44, 62], but the results are disappointing: GPT-4V tends to output lengthy content at every frame, leading to significant delays, making it impractical for real-time streaming video. We also explore training baseline models for per-frame chatting. Unfortunately, this approach evidently diminishes the language modeling capability, likely due to harmful language modeling on an excessive number of redundant frames. + +We propose Learning-In-Video-strEam (LIVE), a comprehensive framework that encompasses learning, data, and inference methods to develop an online video assistant. Unlike per-frame dialogue approach, LIVE introduces a novel training objective termed Streaming EOS (End-Of-Sequence) prediction that enables the model to learn when to response or remain silent in a video stream. This objective differs from next-token prediction since EOS tokens here will not appear in the input/output sequence. However, it can work well with the autoregressive loss to train + +an online VideoLLM. This design reduces unnecessary context, helping the model to manage much longer streaming videos. Nevertheless, the training still requires data from user queries and assistant responses within video streams, which is scarce in popular video datasets usually used for training offline video models. To address this issue, LIVE presents a streaming dialogue generation scheme that converts offline annotations into online dialogues to support free-form chatting. To enhance inference efficiency, LIVE leverages continuous key-value caching for streaming assistance, and parallelizes the fast visual encoding and slow language decoding to prevent bottlenecks, thus moving towards real-time application. + +With LIVE framework, we build a simple VideoLLM-online model upon CLIP [65] vision encoder and Llama2 [73] language model. To evaluate the performance of video streaming dialogue, we utilize the language perplexity metric and design two new metrics to comprehensively assess the model's capabilities in language modeling, temporal responsiveness, and overall fluency. Experiments with real-time Ego4D narration [26] demonstrate that our method shows advantages in all that metrics, with higher speed and lower memory cost. Furthermore, our model achieves state-of-the-art results on numerous offline benchmarks, such as short- and long-term activity recognition and forecasting on the COIN and Ego4D LTA benchmarks [26, 71]. In addition, our model has good speed/memory efficiency, which allows continuous video narration on Ego4D video clips with memory cost less than 20 GB and average speed higher than 10 FPS on a single A100 GPU, paving the way for future real-world usage. + +# 2. Related Work + +Visual Dialogue. Before transformers [74] become mainstream in vision, visual dialogue methods [1, 16] tend to employ a visually enhanced encoder, with a discriminative head to select candidate answers or a recurrent architecture to generate multi-turn responses. For the encoder, a variety of attention mechanism-based approaches [22, 59, 66] have been proposed to learn the interactions between the image, the answers, and the dialogue history. There have also been explorations of encoder-only BERT [18] models for visual dialogue [57, 76]. However, most of them rely solely on a single image/video at the beginning of the conversation, followed by multi-turn pure language dialogue, which makes them less flexible than the current interleaved vision-language dialogue systems. + +Large Multimodal Models. The advent of large language models (LLMs) [8, 60, 63] has inspired a series of LLM multimodal variants, i.e. large multimodal models (LMMs). Early LMMs [2, 13, 41, 49, 93] achieve image dialogue by projecting the image encoding (e.g., from CLIP [65]) to align with LLM embedding space. Then, lots of ef + +forts [3, 6, 40, 62, 82] explore more free-form interleaved vision-text chatting, spatial understanding [6, 11, 12, 37, 51, 83, 87], video comprehension [42, 43, 52, 56, 69, 81, 86], etc. However, when it comes to online scenario, there is less exploration on how LLMs can fulfill the temporal alignment, long-context, and real-time requirements for streaming video inputs. Our research bridges this gap, offering comprehensive solutions across model training, data, and inference to study the problem. + +Online Video Understanding. Typical video understanding benchmarks, such as action recognition [9], temporal action localization [29], video question answering [38], and video dialogue [1], typically allow models to access entire video frames to make predictions, a setting referred to as "offline". However, such setting does not align well with many real-time demands (e.g., autonomous driving, AR glasses). Instead, there is a growing focus on "online" video understanding problems like online action detection [24] and anticipation [35], which aim to identify the current/future action at each timestamp without seeing the future. Our study pioneers LMMs for online video understanding. Unlike previous online action detection [75, 89] or anticipation models [25, 88], which primarily address one task with a highly customized model, we aim to propose a general solution to achieve free-form dialogue along the online video stream, enabling a model to flexibly handle diverse tasks. + +Efficient Token Decoding. Efficient token decoding for LLMs and LMMs is essential for applying them to real-time online services. To accelerate that, a diversity of strategies has been proposed, such as parallelism on batch dimension [67] or cached key-value sequence [15], computation/memory management optimization [23, 30, 36, 67, 84], and even some lossy approaches [50, 78]. Our focused streaming video scenario has less concerns in large batch processing, but expects faster decoding to avoid excessive frame skipping. We have considered this in training objectives, and further propose some inference schemes to accelerate the decoding efficiency. + +# 3. Method + +In this section, we present Learning-In-Video-strEam (LIVE) framework that enables LMMs to provide temporally aligned response, handle long-context streaming video, and run efficiently towards real-time usage. We will start from the problem definition of "video streaming dialogue", analyze its challenges, and introduce our approach to solve the problem. + +# 3.1. Video Streaming Dialogue + +Problem Formulation. Though huge successes have been witnessed in large multimodal models (LMMs), the assistance scenario like a smart AR glass helping the user with cooking, is still far from the capabilities of current LMMs, + +even for the most advanced version, GPT-4V [62]. For example, despite carefully prompting GPT-4V similarly to MMVid [44]—to perform per-frame dialogue for handling streaming video inputs—the redundancy in responses between frames, a limited context window of $10\sim 50$ frames, and slow speed, collectively render the current GPT-4V unsuitable for online video understanding, as our prompting analysis demonstrates (see supplementary material). + +To bridge the gap, we define a problem termed "video streaming dialogue". Given the context sequence before time $t = t_1$ , denoted as $[\mathrm{Ctx}^{t < t_1}]$ , which may encompass previous vision-language content (e.g., historical user queries, video frames, assistant responses), and an ongoing continuous video stream from $t_1$ to $t_2$ , denoted as $[\mathrm{Frame}^{t_1 \leq t \leq t_2}]$ , our goal is (1) to determine whether the current time $t_2$ is suitable for language modeling; (2) to carry out language modeling + +$$ +\max P \left(\left[ \mathrm {T x t} _ {i + 1} ^ {t _ {2}} \right] \mid \left[ \mathrm {C t x} ^ {< t _ {1}} \right], \left[ \mathrm {F} ^ {t _ {1} \leq t \leq t _ {2}} \right], \left[ \mathrm {T x t} _ {\leq i} ^ {t _ {2}} \right]\right) \tag {1} +$$ + +if $t_2$ is determined, where $[\mathrm{Txt}_i^t]$ denotes the ideal language token in the $i$ -th position at timestamp $t$ . [F] is the abbreviation of [Frame]. + +In the following, we analyze if existing techniques have been enough to solve the problem. + +Interleaved/Per-frame Dialogue are Suboptimal. First, we investigate whether the popular approach of interleaved vision-language chatting can address this problem. Related to our formulation above, such a method learns language modeling after given frames between timestamps $t_1$ and $t_2$ . However, if this approach is adopted during inference, it necessitates the manual selection of timestamps $t_1$ and $t_2$ , which does not align with the concept of video streaming dialogue. Current VideoLLMs [42, 43, 52, 56, 69, 81, 86] represent a simplified version of this approach, engaging in single- or multi-turn pure language dialogue following video clip inputs. + +If we consider a more free-form format, multi-turn interleaved vision-language chatting [3, 6, 40, 62, 82], we can get a per-frame chatting solution that might be hopeful to solve video streaming dialogue, which performs per-frame language modeling + +$$ +\max P \left(\left[ \mathrm {T x t} _ {i + 1} ^ {t} \right] \mid \left[ \mathrm {C t x} ^ {< t _ {1}} \right], \left[ \mathrm {F r a m e} ^ {t} \right], \left[ \mathrm {T x t} _ {\leq i} ^ {t} \right]\right) \tag {2} +$$ + +for every frame from timestamp $t_1 \leq t \leq t_2$ . Since $t_2$ is necessary to output answer, so we can simply learn short text (e.g., "this is not the time to answer") between $t_1 \leq t < t_2$ . However, this approach imposes a significant burden on processing speed. For every frame, performing the slow, recurrent, and lengthy next-token prediction with a billion-scale language model makes it extremely hard to + +achieve real-time video streaming dialogue. Then, this will lead to unavoidable frame skipping, which is unexpected and detrimental for temporal alignment. Furthermore, it taxes the limited context window of the language model, presenting challenges for modeling long contexts and managing GPU memory efficiently. + +Streaming EOS Prediction. To solve the problem mentioned above, we first consider a more efficient per-frame chatting method: simply assigning the End-of-Sequence (EOS) token as the content for chatting between $t_1 \leq t < t_2$ . However, this approach remains suboptimal. The dialogue prompt template (e.g., [INST], [/INST], space tokens in Llama [72, 73]) still consumes a considerable number of tokens per frame, which is less favorable due to the numerous frames in streaming video. Furthermore, the excessive number of EOS tokens in the sequence can significantly increase the language model's perplexity, as we have observed in our experiments. + +Instead, we propose a novel training objective named "streaming EOS prediction" to address this issue. We still assume $t_2$ is essential for decoding language; thus, we normally learn language modeling here: + +$$ +\max P \left(\left[ \mathrm {T x t} _ {i + 1} ^ {t _ {2}} \right] \mid \left[ \mathrm {C t x} ^ {< t _ {2}} \right], \left[ \mathrm {F r a m e} ^ {t _ {2}} \right], \left[ \mathrm {T x t} _ {\leq i} ^ {t _ {2}} \right]\right). \tag {3} +$$ + +However, for timestamps $t_1 \leq t < t_2$ , which are redundant for producing answers, we directly learn the model to predict EOS token on the frame tokens, i.e. + +$$ +\max P (\text {E O S} \left| [ \text {C t x} ^ {< t} ], [ \text {F r a m e} ^ {t} ]\right), \text {w h e r e} t _ {1} \leq t < t _ {2}. \tag {4} +$$ + +In this way, we "skip" a dialogue turn and learn to determine when it is appropriate to decode language for streaming inputs. During inference, if EOS is predicted on a frame, then we can directly ask the next frame to input. Meanwhile, the EOS token is not appended to the context to prevent it from affecting the language modeling. Therefore, this task is not about next-token prediction; however, it can work with autoregressive loss to train a video streaming dialogue model. In the following, we first introduce how can we get the data of streaming video and timestamped language annotations, then present the details about the model and the training procedure. + +We also note that the EOS token mentioned here is not limited to the real EOS token used in language models (e.g., $\langle / \mathbb{S} \rangle$ in Llama). It is permissible to use any token or to introduce a new token, provided that it is specified in the system prompt. We use this term solely for simplicity. + +# 3.2.Data + +Online Annotations to Video Streaming Dialogue. Some video datasets, such as Ego4D narrations [26], are inher + +![](images/10c68d0ebedd7216acfc5738c9b8c84dc1920e547942e2b8504460528aa67ca1.jpg) +Figure 2. The streaming dialogue data generation method in our LIVE framework. We randomly insert templated questions into the video timeline and "expose" the ground-truth video annotations (along with their timestamps) to LLMs, prompting them to answer the queries within a period of time. + +ently collected in a streaming manner, with annotators providing real-time narrations as they watch a 5-minute long video clip. However, prior research [45, 90] has predominantly focused on learning from short, discrete clips (e.g. only 32 frames), rather than in a continuous streaming context. For this dataset, we followed the same instructions provided to human annotators [26] as our model training prompt. This prompt instructs the model to simulate human annotators to streamingly generate narrations in a 5-minute video (about 600 frames in 2 FPS). For demo purposes (not for experiments), we also utilize Llama-2-13B-Chat [73] to rephrase the narration text, correcting grammatical errors and typos, and converting it into a more understandable version (e.g., changing "C does..." to "You do..."). + +Offline Annotations to Video Streaming Dialogue. Despite the Ego4D narration data being collected in a streaming manner, most prevalent video datasets [14, 26, 54, 71] are used to train offline models and only feature temporal segment annotations paired with basic language descriptions (e.g., activities, narrations). To bridge this gap, we propose a method for synthesizing dialogue data from these sources. As shown in Figure 2, our key idea is use LLM to generate user-assistant dialogues based on video annota + +tions, involving the following steps: + +- First, we prepare a question template library containing various queries about the past, present, and future tenses of the video, totaling $N$ queries. We randomly sample one question from the library, denoted as $Q_{i}$ . +- Then, we obtain the video annotation timeline from the offline dataset. This usually includes timestamped language descriptions, which we organize into a language prompt, e.g., "time $t_a \sim t_b$ : boiling the water; time $t_c \sim t_d$ : cutting the vegetables.", denoted as $A$ . We consider all the state change critical timestamps as the ideal response times. For this example, $t_a, t_b, t_c$ , and $t_d$ are all considered response times. +- Third, we prompt the large language model to generate responses at every critical timestamp, e.g., $t_a, t_b, t_c, t_d$ , according to $Q_i$ and $A$ . We can repeat this procedure for each $Q_i$ until all queries have been processed. The responses are saved for loading during training. +- Finally, during training, we (1) randomly sample a query and load its responses at critical timestamps, (2) randomly insert a query into a video timestamp $t_r$ , (3) discard the responses that occur before $t_r$ , and add a response at $t_r$ . Here different queries can be inserted into one video, which only requires discarding the responses of the previous query after the new query insertion timestamp. + +In this way, we can generate temporally varied and freeform dialogue data within a video stream. We have prepared 50 questions each for past, current, and future events, totaling $N = 150$ queries. We use Llama-2-13B-Chat to generate the responses and insert a maximum of 3 queries per training sample. The offline datasets we used are COIN [71] and Ego4D GoalStep [70] (for demo usage), which belong to the categories of egocentric and instructional video datasets, aligning with our aim to develop online video assistants. Here, we do not consider online action detection benchmarks (e.g., THUMOS14 [33], TVSeries [24]) because they are closed-set online classification benchmarks, and their labels are too brief, which may cause language models to generate hallucinatory responses. Please refer to the supplementary material for the generated dialogue example. + +# 3.3. Model Training + +Model Architecture. We illustrate the model architecture in Figure 3. Similar to LLaVA [48, 49], it comprises three key components: an image encoder, an MLP projector, and a language model. For the image encoder, we utilize the CLIP ViT-L [20, 65] encoder (pretrained on DataComp1B [21]) to extract video frame embeddings at 2 FPS. Each video frame embedding has a shape of $(1 + h_p\times w_p)\times c$ where $(1 + h_{p}\times w_{p})$ denotes the CLS token and average + +pooled spatial tokens. The extracted frame embeddings from the image encoder are then fed into MLP projector to frame tokens, as in LLaVA-1.5 [48]. Then frame tokens are interleaved with language tokens as input to an LLM, Llama-2-7B-Chat [73]. Finally, we incorporate LoRA [31] in every linear layer of the LLM for efficient tuning. + +Training Loss. As described in Section 3.1, our learning objective is twofold. The first part focuses on autoregressive language modeling, aiming to maximize the joint probability of input text sequences. The second training objective involves streaming EOS prediction, which requires the model to remain silent when it is unnecessary to output responses. With these two training objectives, we have language modeling (LM) loss and streaming loss terms to minimize, both employing cross-entropy loss: + +$$ +L = \frac {1}{N} \sum_ {j = 1} ^ {N} \left(\underbrace {- \log l _ {j + 1} P _ {j} ^ {[ T x t _ {j + 1} ]}} _ {L M L o s s} - \underbrace {w \log f _ {j} P _ {j} ^ {[ E O S ]}} _ {\text {S t r e a m i n g L o s s}}\right), \tag {5} +$$ + +where $l_{j}$ and $f_{j}$ denote condition indicators. $l_{j}$ is 1 if the $j$ -th token is a language response token, and 0 otherwise. $f_{j}$ is 1 if (1) the $j$ -th token is the last token of a frame $^3$ , and (2) $l_{j+1} = 0$ . In essence, the streaming EOS loss is applied to frames before responding. $P_{j}^{[\mathrm{Txt}_{j+1}]}$ denotes the probability on $j + 1$ -th text token, output from the language model head of the $j$ -th token, and $P_{j}^{[\mathrm{EOS}]}$ represents that probability for the EOS token. $w$ is a balance term, set to $w = 1$ by default. As shown in Figuer 3, we visualize the ranges of language loss and streaming loss in an input sequence when we only use 1 token for each frame. + +# 3.4. Inference + +Probability Correction. The prevalence of EOS token will bias the model towards EOS token prediction. To address this, we introduce a threshold $\theta$ to correct the output probability on frame tokens: EOS will not be considered as the next token if $P_{j}^{[\mathrm{EOS}]} < \theta$ . In practical usage, we find that setting $\theta$ to 0.5 yields much better results than no threshold here. + +Continuous Key-Value Cache. As shown in Figure 4, During inference, the video is input as a frame-by-frame stream, with a default FPS 2. Our model takes the current frame as input and generates tokens on-the-fly. In the whole process, we use the key-value cache trick to accelerate token decoding, thus we do not need to manually append the generated + +![](images/f11632ceebe9151bacb8a426e61456bbb5f741b41c38654aad3375bcb54ee584.jpg) +Figure 3. The training method in our LIVE framework. We organize the user-assistant dialogue data and video frames in temporal order as the input sequence. To learn the model when to answer or keep silent in a video stream, we employ not only the standard language modeling (LM) loss but also introduce a streaming EOS prediction loss. This additional loss supervises the model when it is necessary to generate language, enabling it to produce temporally aligned responses and reduces the redundant dialogue history. + +![](images/5a5624900b49eba8690c05eaa1796c72854767a44d7e5fb14362a770626154e3.jpg) +Figure 4. Inference pipeline in our LIVE framework. During inference, video frames serve as streaming inputs. Our model maintains a continuous key-value cache as the input progresses to speed up the inference. Furthermore, we parallelize the fast video frame encoder and the slower language model to avoid the bottleneck in the latter. Video frame tokens can be always encoded and buffered, no need to wait the language decoding. + +tokens for the next frame. As our training encourages the model to keep silence, this continuous inference would be efficient, providing the possibility to pace with the video stream speed. + +Parallelization of encoding and decoding. Our video frame encoder utilizes CLIP ViT-L (307M), which is significantly smaller than the 7B LLM. This size discrepancy leads to a speed mismatch, potentially resulting in frame + +skipping when the LLM decodes long sentences. To mitigate this issue, we parallelize the processes and establish a buffer for video frame tokens. Once the language model completes its decoding, it fetches all frame tokens from the buffer and inputs them in batch. + +# 4. Experiments + +# 4.1. Implementation Details + +We implement VideoLLM-online-7B with our LIVE framework. It uses OpenCLIP-ViT-L [21, 65] as the video frame encoder, 2-layer MLP as multimodal connector, and Llama2-7B-Chat [73] as the language model. We only use 1 token for each video frame, sampled by 2 FPS in the following experiments. We add LoRA [31] to all linear layers of Llama-2, with a rank of 128, scaling factor of 256. For video streaming dialogue experiments, we train 2 epochs for the model. For the downstreaming offline experiments, we train 5 epochs without pre-training for fair comparison with previous methods. By default, we set streaming loss weight $w = 1.0$ during training. + +
MethodTraining ObjectiveEgo4D Narration Stream on Validation Set#Training Token↓Training Cost
LM-PPL↓TimeDiff ↓Fluency↑
No Training498.56.500.1%n/an/a
Interleaved DialogueLanguage Modeling2.456.4711.1%169412h
Per-frame Dialogue for StreamingLanguage Modeling (w/ EOS turns)3.342.5237.7 %673722h
Streaming Dialogue (Ours)Language Modeling + Streaming EOS2.432.3242.6%169412h
+ +
Streaming LossEgo4DNarrationStreamValidation
LM-PPL↓TimeDiff↓Fluency↑
Standard CE2.432.3242.6%
OHEM [68]2.532.3941.0%
Focal Loss [46]2.592.4439.4%
+ +(b) Streaming loss function. Standard CE (cross-entropy) is enough for training streaming dialogue; there is no need to specifically to address the class imbalance on EOS token. + +
Weight τEgo4D Narration Stream Validation
LM-PPL↓TimeDiff↓Fluency↑
τ = 0.52.442.3242.4%
τ = 1.02.432.3242.6%
τ = 2.02.462.3142.5%
τ = 3.02.472.3242.5%
+ +(c) Streaming loss weight. Using slightly higher stream- (d) Generation memory/speed. ing loss weight $(\tau = 2.0)$ achieves the best trade-off Streaming dialogue method has among various metrics. much better efficiency. + +(a) Learning method for streaming dialogue. Training with streaming dialogue method can achieve much better TimeDiff and Fluency, as well as maintain the language modeling quality. Meanwhile, the streaming dialogue can enjoy much more efficient training than per-frame dialogue for video streaming dialogue. + +
MethodMem↓FPS↑
Interleaved34.4G1.5
Per-frame Streaming24.9G7.5
Streaming18.2G13.5
+ +
MethodNot use HowTo100MCOIN Benchmark Top-1 Accuracy↑
StepTaskNextProc.Proc.+
ClipBERT [39]✓30.865.4---
TimeSformer [7]✗46.585.334.017.040.1
Paprika [92]✗51.085.843.2--
DistantSup [47]✗54.190.039.4-41.3
VideoTF [58]✗56.591.042.440.246.4
ProcedureVRL [91]✗56.990.846.8--
VideoTaskGraph [5]✗57.290.540.2--
VideoLLM-online✓59.892.148.147.952.9
+ +(a) Results on COIN benchmarks (left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal. + +Table 1. Ablation experiments on Ego4D Narration Stream. We train VideoLLM-online on Ego4D [26] narration stream training set and evaluate on its validation set. The comparison is based on our designed metrics: the ratio of strictly correct prediction tokens (Fluency), language modeling perplexity (LM-PPL) and time difference (TimeDiff) metrics. #Training Token denotes the average token length during training. TimeDiff refers to difference in second. Default settings are highlighted. + +
MethodNot use EgoVLPEnd-to-end?Ego4D LTA ED@Z=20↓
VerbNounAction
CLIP [17]✓✓0.7390.7690.941
EgoT2 [79]✓✓0.7220.7640.935
I-CVAE [53]✓✓0.7530.7490.931
HierVL [4]✓✓0.7240.7350.928
VideoLLM [10]✗✓0.7210.7250.921
VideoLLM-online✓✓0.6970.6980.897
Palm [32]✗✗0.6960.6510.886
AntGPT [88]✗✗0.6500.6500.877
+ +(b) Results on Ego4D LTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions. + +Table 2. Experiments on COIN [71] and Ego4D [26] benchmarks. VideoLLM-online is finetuned on their training set, and strictly evaluated on the test set by generated string comparison with the ground-truth text. It achieves best results among end-to-end models. + +# 4.2. Evaluation Setting + +Datasets. We use (1) assistance-related, instructional video dataset COIN [71] and (2) continuous, egocentric video dataset Ego4D [26] in various settings: + +- Ego4D Narration Stream: We also leverage the dense Ego4D timestamp-narration to create a streaming set. The goal is to generate narrations timely like Ego4D human annotators [26]. We follow the division of the training, validation, and test set in EgoVLP [45]. +- COIN+Ego4D Stream: To further evaluate the potential of model's performance to free-form dialogue, we construct a simple COIN+Ego4D Stream set, constructed from COIN annotations using our data generation methods, and the above Ego4D Narration Stream. The model should remind the user when an action starts, summarizes the action when it ends, as well as forecasts the next action. We use the same training/testing splits as COIN + +benchmarks. See supplementary material for the results. + +- COIN Benchmarks: Following on previous studies [47, 58, 80, 91], we evaluate our model on six common benchmarks of the COIN dataset: step recognition, step forecasting, task summarization, procedure forecasting, procedure forecasting with a goal. +- Ego4D long-term action anticipation (LTA) benchmark: This benchmark requires to predict next $Z = 20$ actions (verbs and nouns) for the given video of previous 8 steps. We use the standard Ego4d v2 splits as in previous studies [32, 88]. + +Evaluation metrics. We use the following metrics to evaluate the model as an online video assistant: + +- Language modeling perplexity (LM-PPL). This indicates the language modeling capability at a given times-tamp. A lower LM-PPL is better for correct answering. +- Time Difference (TimeDiff). To evaluate the temporal + +alignment capability of an online assistant, we calculate the discrepancy between the timestamp of its response and the expected timestamp for each response. We average TimeDiff each turn as the metric. + +- Fluency. Individual LM-PPL or TimeDiff do not entirely evaluate both language and temporal effectiveness in a streaming dialogue. We introduce the Fluency metric, which evaluates the proportion of consecutive successful token prediction within a dialogue turn. As the token also including language tokens, Fluency can comprehensively reflect the aspects of LM-PPL and TimeDiff. + +Baselines. To our best knowledge, we are the first one to tackle producing temporal aligned, free-form language answer with streaming video settings. To better understand the challenges, we build baseline models for video-text interleaved dialogue, per-frame dialogue, as we described in Section 3.1, with the same model architecture and training details to VideoLLM-online, differing only in their training objective and multi-turn formulation. + +# 4.3. Ablation Study + +Learning Method. Table 1a shows the ablation studies on learning methods in a streaming setting. Both vision-language interleaved and streaming methods exhibit low perplexity loss, indicating that our proposed objective does not hurt language modeling capability. However, learning with per-frame for streaming will produce significant higher LM-PPL than others, which might be attributed to the too more single EOS token in answering that affects the original language modeling. + +When we turn to online metrics of TimeDiff and Fluency, streaming dialogue method yields much better results than others. In our observation, the first interleaved dialogue method always outputs language after every frame, and the second multi-turn for streaming dialogue approach tends to answer EOS token after every frame, which decreases their performance for streaming video inputs. Furthermore, per-frame for streaming dialogue method will significantly slow down the training speed due to its lengthy prompts, while our method has no negative impact on the efficiency. + +Streaming Loss. We continue to investigate the most suitable strategy to learn the streaming objective. As shown in Table 1b and Table 1c, we find a default setting works surprisingly well (CE loss, $\tau = 1.0$ ), which demonstrates there is no need to apply more advanced loss (e.g. Focal Loss [46]) to address the imbalance on EOS token. + +Inference Efficiency. In Table 1d, we test the inference efficiency on Ego4D narration stream validation set (5 minute), and report the memory cost and average FPS on a single A100 GPU. The first interleaved dialogue method, which will output language after every video frame, has huge memory cost and slow generation speed. The second one, per-frame dialogue for streaming that formulates + +all in a multi-turn dialogue, show better efficiency than the first one since it can cost less tokens in redundant frames. However, this approach still lags significantly behind our streaming dialogue approach, which does not cost extra tokens in redundant frames thus maintain smaller key-value cache. We observe for most Ego4D videos, our model can run larger than 10 FPS, providing possibility for AI assistants working in real-time video stream. + +# 4.4. Results + +Offline language modeling. We show our model can perform well on traditional temporal summarization and forecasting problems. As shown in Table 2(a), our model achieves state-of-the-art performances in step/task summarization and next step/procedure forecasting benchmarks of COIN dataset [71]. Furthermore, we also obtain the best performance among end-to-end models evaluated on Ego4D LTA. We observe the results of Palm [32] and AntGPT [88] are better than us, but they used egocentric pre-trained visual feature [45], and integrates lots of complex cascading methods to improve the forecasting results. Our VideoLLM-online, however, directly outputs language as the results, which performs better than the similar end-to-end VideoLLM [10]. + +Visualization. See Figure 1, we visualize two representative examples, real-time narration and online dialogue. The most distinctive characteristics of our approach are: (1) the dialogue process goes along with the streaming video input, rather than chatting based on the full video. (2) The response will be "muted" when it is unnecessary, significantly improving the overall speed of video streaming dialogue. + +# 5. Conclusion + +We propose Learning-In-Video-strEam (LIVE), a novel framework empowering LLMs to handle streaming video, to produce temporal aligned answers, hold long-context video duration, and have high inference efficiency. We use LIVE to train a simple VideoLLM-online model, which not only achieves superior capability in online/offline vision-language tasks, but also enable fast inference for an online video streaming setting. We believe enabling such abilities will be an important step to move towards always-on online assistant. In future work, to make our VideoLLM-online be more general and improve its spatial capability in zero-shot prediction for downstream applications, we will explore suitable pre-training data source, and develop models that can employ more spatial tokens but without many trade-off on speed and memory cost. + +Acknowledgment This work is sponsored by Project Aria Team, Meta. The datasets and processing were acquired and all models were trained at the National University of Singapore (NUS) by NUS authors. + +# References + +[1] Huda AlAmri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K. Marks, Chiori Hori, Peter Anderson, Stefan Lee, and Devi Parikh. Audio visual scene-aware dialog. In CVPR, pages 7558-7567, 2019. 2, 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2 +[3] Rohan Anil, Sebastian Borgeaud, Yonghui Wu, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Slav Petrov, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy P. Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul Ronald Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, and et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2023. 2, 3 +[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In CVPR, pages 23066-23078, 2023. 7 +[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 7 +[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966, 2023. 2, 3 +[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 7 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In *NeurIPS*, pages 1877–1901, 2020. 1, 2 + +[9] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 3 +[10] Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292, 2023. 7, 8 +[11] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 3 +[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3 +[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C.H.Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500, 2023. 2 +[14] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. Int. J. Comput. Vis., 130(1):33-55, 2022. 4 +[15] Tri Dao, Daniel Haziza, Francisco Massa, and Grigory Sizov. Flash-decoding for long-context inference. https://pytorch.org/blog/flash-decoding/, 2023.3 +[16] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In CVPR, pages 1080-1089, 2017. 2 +[17] Srijan Das and Michael S. Ryoo. Video + clip baseline for ego4d long-term action anticipation. arXiv:2207.00579, 2022. 7 +[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2 +[19] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv:2309.11499, 2023. 2 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5 +[21] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M. Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen + +Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. arXiv:2304.14108, 2023. 5, 6 +[22] Zhe Gan, Yu Cheng, Ahmed El Kholy, Linjie Li, Jingjing Liu, and Jianfeng Gao. Multi-step reasoning via recurrent dual attention for visual dialog. In ACL, pages 6463-6474, 2019. 2 +[23] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms. arXiv:2310.01801, 2023. 3 +[24] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In ECCV, pages 269-284, 2016. 3, 5 +[25] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. Arxiv, 2106.02036, 2021. 3 +[26] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18973-18990, 2022. 2, 4, 7 +[27] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, FuJen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abraham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed + +ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, and et al. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. arXiv:2311.18259, 2023. 1 +[28] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv:2309.03905, 2023. 2 +[29] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 3 +[30] Ke Hong, Guohao Dai, Jiaming Xu, Qiuli Mao, Xiuhong Li, Jun Liu, Kangdi Chen, Hanyu Dong, and Yu Wang. Flashdecoding++: Faster large language model inference on gpus. arXiv:2311.01282, 2023. 3 +[31] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5, 6 +[32] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models @ ego4d long-term action anticipation challenge 2023. arXiv:2306.16545, 2023. 7, 8 +[33] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorbani, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos "in the wild". Comput. Vis. Image Underst., 155:1-23, 2017. 5 +[34] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv:2001.08361, 2020. 1 +[35] Kris M. Kitani, Brian D. Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In ECCV, pages 201-214, 2012. 3 +[36] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. arXiv:2309.06180, 2023. 3 +[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv:2308.00692, 2023. 2, 3 +[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379, 2018. 3 +[39] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 7 +[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 2, 3 + +[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[42] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv:2305.06355, 2023. 2, 3 +[43] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 2, 3 +[44] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, Ce Liu, and Lijuan Wang. MM-VID: advancing video understanding with gpt-4v(ision). arXiv:2310.19773, 2023. 2, 3 +[45] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Egocentric video-language pretraining. arXiv:2206.01670, 2022. 4, 7, 8 +[46] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 7, 8 +[47] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853, 2022. 7 +[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 5 +[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 2, 5 +[50] Xiaoxuan Liu, Lanxiang Hu, Peter Bailis, Ion Stoica, Zhijie Deng, Alvin Cheung, and Hao Zhang. Online speculative decoding. arXiv:2310.07177, 2023. 3 +[51] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv:2306.09093, 2023. 2, 3 +[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424, 2023. 2, 3 +[53] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action anticipation. In WACV, pages 6037-6046, 2023. 7 +[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 4 +[55] Seungwhan Moon, Andrea Madotto, Zhaojiang Lin, Tushar Nagarajan, Matt Smith, Shashank Jain, Chun-Fu Yeh, + +Prakash Murugesan, Peyman Heidari, Yue Liu, et al. Anymal: An efficient and scalable any-modality augmented language model. arXiv:2309.16058, 2023. 2 +[56] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv:2305.15021, 2023. 2, 3 +[57] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for visual dialog: A simple state-of-the-art baseline. In ECCV, pages 336-352, 2020. 2 +[58] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv:2303.13519, 2023. 7 +[59] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Efficient attention mechanism for visual dialog that can handle all the interactions between multiple inputs. In ECCV, pages 223–240, 2020. 2 +[60] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt/, 2023.1, 2 +[61] OpenAI. GPT-4 technical report. arXiv:2303.08774, 2023. 1 +[62] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.2.3 +[63] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 2 +[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 2 +[65] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 5, 6 +[66] Idan Schwartz, Seunghak Yu, Tamir Hazan, and Alexander G. Schwing. Factor graph attention. In CVPR, pages 2039-2048, 2019. 2 +[67] Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Re, Ion Stoica, and Ce Zhang. Flexgen: High-throughput generative inference of large language models with a singlegpu. arXiv:2303.06865, 2023. 3 +[68] Abhinav Shrivastava, Abhinav Gupta, and Ross B. Girshick. Training region-based object detectors with online hard example mining. In CVPR, pages 761-769, 2016. 7 +[69] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv:2307.16449, 2023. 2, 3 + +[70] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 5 +[71] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 2, 4, 5, 7, 8 +[72] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023. 1, 4 +[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv:2307.09288, 2023. 2,456 +[74] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 6000-6010, 2017. 2 +[75] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Yuanjie Shao, Zhengrong Zuo, Changxin Gao, and Nong Sang. Oadtr: Online action detection with transformers. In ICCV, pages 7545-7555, 2021. 3 +[76] Yue Wang, Shafiq R. Joty, Michael R. Lyu, Irwin King, Caiming Xiong, and Steven C. H. Hoi. VD-BERT: A unified vision and dialog transformer with BERT. In EMNLP, pages 3325-3338, 2020. 2 +[77] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv:2309.05519, 2023. 2 +[78] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv:2309.17453, 2023. 3 +[79] Zihui Xue, Yale Song, Kristen Grauman, and Lorenzo Torresani. Egocentric video task translation. In CVPR, pages 2310-2320, 2023. 7 +[80] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia + +Schmid. Unloc: A unified framework for video localization tasks. In ICCV, pages 13623-13633, 2023. 7 +[81] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2, 3 +[82] Zhewei Yao, Xiaoxia Wu, Conglong Li, Minjia Zhang, Heyang Qi, Olatunj Ruwase, Ammar Ahmad Awan, Samyam Rajbhandari, and Yuxiong He. Deepspeed-visualchat: Multi-round multi-image interleave chat via multi-modal causal attention. arXiv:2309.14327, 2023. 2, 3 +[83] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv:2310.07704, 2023. 2, 3 +[84] Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. Orca: A distributed serving system for {Transformer-Based} generative models. In OSDI, pages 521-538, 2022. 3 +[85] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv:2309.02591, 2023. 2 +[86] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858, 2023. 2, 3 +[87] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 2, 3 +[88] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? In ICLR, 2024. 3, 7, 8 +[89] Yue Zhao and Philipp Kähenbuhl. Real-time online video detection with temporal smoothing transformers. In ECCV, pages 485-502, 2022. 3 +[90] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 4 +[91] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In CVPR, pages 14825-14835, 2023. 7 +[92] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 7 +[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/images.zip b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..83cff82ae0bed5382078463d5b4739730f2f1801 --- /dev/null +++ b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41232682f638e8066c7620b0495980ea418a79074e50e136233bb3ef53b9f34f +size 448294 diff --git a/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/layout.json b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c38bc22c6789493a11b3fc8f6cffe68ea87d3a07 --- /dev/null +++ b/2024/VideoLLM-online_ Online Video Large Language Model for Streaming Video/layout.json @@ -0,0 +1,9498 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "VideoLLM-online: Online Video Large Language Model for Streaming Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 142, + 140, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 142, + 140, + 157 + ], + "spans": [ + { + "bbox": [ + 83, + 142, + 140, + 157 + ], + "type": "text", + "content": "Joya Chen1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 157, + 148, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 157, + 148, + 170 + ], + "spans": [ + { + "bbox": [ + 94, + 157, + 148, + 170 + ], + "type": "text", + "content": "Difei Gao" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 143, + 227, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 143, + 227, + 157 + ], + "spans": [ + { + "bbox": [ + 156, + 143, + 227, + 157 + ], + "type": "text", + "content": "Zhaoyang Lv²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 164, + 157, + 225, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 157, + 225, + 171 + ], + "spans": [ + { + "bbox": [ + 164, + 157, + 225, + 171 + ], + "type": "text", + "content": "Jia-Wei Liu1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 143, + 301, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 143, + 301, + 156 + ], + "spans": [ + { + "bbox": [ + 241, + 143, + 301, + 156 + ], + "type": "text", + "content": "Shiwei Wu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 157, + 302, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 157, + 302, + 171 + ], + "spans": [ + { + "bbox": [ + 241, + 157, + 302, + 171 + ], + "type": "text", + "content": "Ziteng Gao" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 143, + 421, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 421, + 157 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 421, + 157 + ], + "type": "text", + "content": "Kevin Qinghong Lin1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 157, + 397, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 157, + 397, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 157, + 397, + 171 + ], + "type": "text", + "content": "Dongxing Mao" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 412, + 143, + 509, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 143, + 509, + 157 + ], + "spans": [ + { + "bbox": [ + 412, + 143, + 509, + 157 + ], + "type": "text", + "content": "n1 Chenan Song" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 413, + 157, + 512, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 157, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 413, + 157, + 512, + 171 + ], + "type": "text", + "content": "Mike Zheng Shou" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 174, + 485, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 174, + 485, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 485, + 188 + ], + "type": "text", + "content": "1Show Lab, National University of Singapore 2Reality Labs Research, Meta" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 48, + 198, + 545, + 398 + ], + "blocks": [ + { + "bbox": [ + 48, + 198, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 545, + 398 + ], + "type": "image", + "image_path": "55bcd345351946a74795ffc9d05af932569b03fe66e2cd880067414b9227d9de.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 398, + 544, + 407 + ], + "lines": [ + { + "bbox": [ + 482, + 398, + 544, + 407 + ], + "spans": [ + { + "bbox": [ + 482, + 398, + 544, + 407 + ], + "type": "text", + "content": "Streaming Dialogue" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 416, + 546, + 450 + ], + "lines": [ + { + "bbox": [ + 45, + 416, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 45, + 416, + 546, + 450 + ], + "type": "text", + "content": "Figure 1. Zero-shot examples of our VideoLLM-online applied to an egocentric video stream from Ego-Exo4D dataset [27]. Our model is designed for temporally aligned, long-context, real-time dialogue in continuous video streams, shedding light on the future always-on, contextual AI assistants (e.g., smart AR glasses). Model responses are appropriately simplified for better visualization." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 463, + 192, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 463, + 192, + 475 + ], + "spans": [ + { + "bbox": [ + 143, + 463, + 192, + 475 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 45, + 489, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 489, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 489, + 288, + 693 + ], + "type": "text", + "content": "Recent Large Language Models (LLMs) have been enhanced with vision capabilities, enabling them to comprehend images, videos, and interleaved vision-language content. However, the learning methods of these large multimodal models (LMMs) typically treat videos as predetermined clips, rendering them less effective and efficient at handling streaming video inputs. In this paper, we propose a novel Learning-In-Video-Stream (LIVE) framework, which enables temporally aligned, long-context, and real-time dialogue within a continuous video stream. Our LIVE framework comprises comprehensive approaches to achieve video streaming dialogue, encompassing: (1) a training objective designed to perform language modeling for continuous streaming inputs, (2) a data generation scheme that converts offline temporal annotations into a streaming dialogue format, and (3) an optimized inference pipeline to speed up interactive chat in real-world video streams. With" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 461, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 461, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 546, + 581 + ], + "type": "text", + "content": "our LIVE framework, we develop a simplified model called VideoLLM-online and demonstrate its significant advantages in processing streaming videos. For instance, our VideoLLM-online-7B model can operate at over 10 FPS on an A100 GPU for a 5-minute video clip from Ego4D narration. Moreover, VideoLLM-online also showcases state-of-the-art performance on public offline video benchmarks, such as recognition, captioning, and forecasting. The code, model, data, and demo have been made available at showlab.github.io/videoollm-online." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 608, + 386, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 386, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 386, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Building the future of always-on contextual AI agent that can promptly answer any human question, digitizing inputs as episodic memories and forecasting future plans given any query in online continuous setting represents a \"holy grail\" mission in AI research. Powered by advancements in large language models (LLMs) [8, 34, 60, 61, 63, 72], recent large multimodal models (LMMs) have unveiled impressive ca" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 129, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 129, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 129, + 712 + ], + "type": "text", + "content": "Corresponding Author." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18407" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "pabilities such as vision-language dialogue [6, 13, 40, 41, 48, 49, 93], spatial understanding [6, 37, 51, 64, 83, 87], processing diverse modalities [19, 28, 55, 77, 85]. Seminal exemplars, like OpenAI's GPT-4V [62], are progressively evolving into highly versatile human AI assistants." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 135, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 289, + 410 + ], + "type": "text", + "content": "It is time to envision an always-on,contextual, J.A.R.V.I.S-like multimodal assistant that supports freeform user-assistant dialogue within the video stream, which we term \"video streaming dialogue\". Unlike existing LMMs for video understanding (i.e., VideoLLMs) [42, 43, 52, 56, 69, 81, 86] that work offline with manually selected short-video clips, an online assistant should continuously receive video frames with visual content that is constantly refreshed. This paradigm shift presents new challenges. First, the user query may come with temporally aligned requirements (e.g., \"alert me when it's time to flip the steak\"), thus the VideoLLM should scan every incoming frame to avoid event missing, instead of only yielding video-level responses. Second, to answer questions regarding summarization and planning, the VideoLLM must retain the longcontext historical vision and language, which poses huge risk of exceeding maximum context window of LLMs, as well as introduces considerable burden to causal decoding speed and GPU memory. Third, the VideoLLM should generate the answer in real-time, keeping pace with the video stream for always-on scenario. These abilities, however, are even partially overlooked by the most advanced AI assistants [3, 62]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 590 + ], + "type": "text", + "content": "One possible path towards such an online VideoLLM, inspired by current interleaved vision-language models [3, 6, 40, 62, 82], is to employ a multi-turn dialogue format to achieve per-frame chatting within a video stream. This can be accomplished by facilitating very frequent user interactions, utilizing the visual frame as query at each timestamp to obtain the answer. We follow this to perform prompt engineering for GPT-4V [44, 62], but the results are disappointing: GPT-4V tends to output lengthy content at every frame, leading to significant delays, making it impractical for real-time streaming video. We also explore training baseline models for per-frame chatting. Unfortunately, this approach evidently diminishes the language modeling capability, likely due to harmful language modeling on an excessive number of redundant frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "We propose Learning-In-Video-strEam (LIVE), a comprehensive framework that encompasses learning, data, and inference methods to develop an online video assistant. Unlike per-frame dialogue approach, LIVE introduces a novel training objective termed Streaming EOS (End-Of-Sequence) prediction that enables the model to learn when to response or remain silent in a video stream. This objective differs from next-token prediction since EOS tokens here will not appear in the input/output sequence. However, it can work well with the autoregressive loss to train" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "an online VideoLLM. This design reduces unnecessary context, helping the model to manage much longer streaming videos. Nevertheless, the training still requires data from user queries and assistant responses within video streams, which is scarce in popular video datasets usually used for training offline video models. To address this issue, LIVE presents a streaming dialogue generation scheme that converts offline annotations into online dialogues to support free-form chatting. To enhance inference efficiency, LIVE leverages continuous key-value caching for streaming assistance, and parallelizes the fast visual encoding and slow language decoding to prevent bottlenecks, thus moving towards real-time application." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 228, + 546, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 443 + ], + "type": "text", + "content": "With LIVE framework, we build a simple VideoLLM-online model upon CLIP [65] vision encoder and Llama2 [73] language model. To evaluate the performance of video streaming dialogue, we utilize the language perplexity metric and design two new metrics to comprehensively assess the model's capabilities in language modeling, temporal responsiveness, and overall fluency. Experiments with real-time Ego4D narration [26] demonstrate that our method shows advantages in all that metrics, with higher speed and lower memory cost. Furthermore, our model achieves state-of-the-art results on numerous offline benchmarks, such as short- and long-term activity recognition and forecasting on the COIN and Ego4D LTA benchmarks [26, 71]. In addition, our model has good speed/memory efficiency, which allows continuous video narration on Ego4D video clips with memory cost less than 20 GB and average speed higher than 10 FPS on a single A100 GPU, paving the way for future real-world usage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 454, + 392, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 392, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 392, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 474, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 642 + ], + "type": "text", + "content": "Visual Dialogue. Before transformers [74] become mainstream in vision, visual dialogue methods [1, 16] tend to employ a visually enhanced encoder, with a discriminative head to select candidate answers or a recurrent architecture to generate multi-turn responses. For the encoder, a variety of attention mechanism-based approaches [22, 59, 66] have been proposed to learn the interactions between the image, the answers, and the dialogue history. There have also been explorations of encoder-only BERT [18] models for visual dialogue [57, 76]. However, most of them rely solely on a single image/video at the beginning of the conversation, followed by multi-turn pure language dialogue, which makes them less flexible than the current interleaved vision-language dialogue systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Large Multimodal Models. The advent of large language models (LLMs) [8, 60, 63] has inspired a series of LLM multimodal variants, i.e. large multimodal models (LMMs). Early LMMs [2, 13, 41, 49, 93] achieve image dialogue by projecting the image encoding (e.g., from CLIP [65]) to align with LLM embedding space. Then, lots of ef" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18408" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "forts [3, 6, 40, 62, 82] explore more free-form interleaved vision-text chatting, spatial understanding [6, 11, 12, 37, 51, 83, 87], video comprehension [42, 43, 52, 56, 69, 81, 86], etc. However, when it comes to online scenario, there is less exploration on how LLMs can fulfill the temporal alignment, long-context, and real-time requirements for streaming video inputs. Our research bridges this gap, offering comprehensive solutions across model training, data, and inference to study the problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "type": "text", + "content": "Online Video Understanding. Typical video understanding benchmarks, such as action recognition [9], temporal action localization [29], video question answering [38], and video dialogue [1], typically allow models to access entire video frames to make predictions, a setting referred to as \"offline\". However, such setting does not align well with many real-time demands (e.g., autonomous driving, AR glasses). Instead, there is a growing focus on \"online\" video understanding problems like online action detection [24] and anticipation [35], which aim to identify the current/future action at each timestamp without seeing the future. Our study pioneers LMMs for online video understanding. Unlike previous online action detection [75, 89] or anticipation models [25, 88], which primarily address one task with a highly customized model, we aim to propose a general solution to achieve free-form dialogue along the online video stream, enabling a model to flexibly handle diverse tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 526 + ], + "type": "text", + "content": "Efficient Token Decoding. Efficient token decoding for LLMs and LMMs is essential for applying them to real-time online services. To accelerate that, a diversity of strategies has been proposed, such as parallelism on batch dimension [67] or cached key-value sequence [15], computation/memory management optimization [23, 30, 36, 67, 84], and even some lossy approaches [50, 78]. Our focused streaming video scenario has less concerns in large batch processing, but expects faster decoding to avoid excessive frame skipping. We have considered this in training objectives, and further propose some inference schemes to accelerate the decoding efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 104, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 104, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 104, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": "In this section, we present Learning-In-Video-strEam (LIVE) framework that enables LMMs to provide temporally aligned response, handle long-context streaming video, and run efficiently towards real-time usage. We will start from the problem definition of \"video streaming dialogue\", analyze its challenges, and introduce our approach to solve the problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 192, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 192, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 192, + 659 + ], + "type": "text", + "content": "3.1. Video Streaming Dialogue" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Problem Formulation. Though huge successes have been witnessed in large multimodal models (LMMs), the assistance scenario like a smart AR glass helping the user with cooking, is still far from the capabilities of current LMMs," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "even for the most advanced version, GPT-4V [62]. For example, despite carefully prompting GPT-4V similarly to MMVid [44]—to perform per-frame dialogue for handling streaming video inputs—the redundancy in responses between frames, a limited context window of " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "10\\sim 50" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " frames, and slow speed, collectively render the current GPT-4V unsuitable for online video understanding, as our prompting analysis demonstrates (see supplementary material)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": "To bridge the gap, we define a problem termed \"video streaming dialogue\". Given the context sequence before time " + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "inline_equation", + "content": "t = t_1" + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "inline_equation", + "content": "[\\mathrm{Ctx}^{t < t_1}]" + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": ", which may encompass previous vision-language content (e.g., historical user queries, video frames, assistant responses), and an ongoing continuous video stream from " + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "inline_equation", + "content": "[\\mathrm{Frame}^{t_1 \\leq t \\leq t_2}]" + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": ", our goal is (1) to determine whether the current time " + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 168, + 546, + 276 + ], + "type": "text", + "content": " is suitable for language modeling; (2) to carry out language modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 296, + 545, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 296, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 312, + 296, + 545, + 320 + ], + "type": "interline_equation", + "content": "\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t _ {2}} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {1}} \\right], \\left[ \\mathrm {F} ^ {t _ {1} \\leq t \\leq t _ {2}} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t _ {2}} \\right]\\right) \\tag {1}", + "image_path": "91ca191fb4dae5d03b0fddc3486f45f575ac21876890092d0b14e0bc2ddf54a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "text", + "content": "if " + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "text", + "content": " is determined, where " + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "inline_equation", + "content": "[\\mathrm{Txt}_i^t]" + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "text", + "content": " denotes the ideal language token in the " + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "text", + "content": "-th position at timestamp " + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 321, + 545, + 357 + ], + "type": "text", + "content": ". [F] is the abbreviation of [Frame]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 357, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 380 + ], + "type": "text", + "content": "In the following, we analyze if existing techniques have been enough to solve the problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "text", + "content": "Interleaved/Per-frame Dialogue are Suboptimal. First, we investigate whether the popular approach of interleaved vision-language chatting can address this problem. Related to our formulation above, such a method learns language modeling after given frames between timestamps " + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "text", + "content": ". However, if this approach is adopted during inference, it necessitates the manual selection of timestamps " + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 381, + 546, + 524 + ], + "type": "text", + "content": ", which does not align with the concept of video streaming dialogue. Current VideoLLMs [42, 43, 52, 56, 69, 81, 86] represent a simplified version of this approach, engaging in single- or multi-turn pure language dialogue following video clip inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 525, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 584 + ], + "type": "text", + "content": "If we consider a more free-form format, multi-turn interleaved vision-language chatting [3, 6, 40, 62, 82], we can get a per-frame chatting solution that might be hopeful to solve video streaming dialogue, which performs per-frame language modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 604, + 545, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 604, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 314, + 604, + 545, + 629 + ], + "type": "interline_equation", + "content": "\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {1}} \\right], \\left[ \\mathrm {F r a m e} ^ {t} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t} \\right]\\right) \\tag {2}", + "image_path": "5362225bc228920afa1d8fbe76b48faa839ab1098e0a4a535679b368e2507f34.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "for every frame from timestamp " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "t_1 \\leq t \\leq t_2" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " is necessary to output answer, so we can simply learn short text (e.g., \"this is not the time to answer\") between " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "t_1 \\leq t < t_2" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": ". However, this approach imposes a significant burden on processing speed. For every frame, performing the slow, recurrent, and lengthy next-token prediction with a billion-scale language model makes it extremely hard to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18409" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "achieve real-time video streaming dialogue. Then, this will lead to unavoidable frame skipping, which is unexpected and detrimental for temporal alignment. Furthermore, it taxes the limited context window of the language model, presenting challenges for modeling long contexts and managing GPU memory efficiently." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "type": "text", + "content": "Streaming EOS Prediction. To solve the problem mentioned above, we first consider a more efficient per-frame chatting method: simply assigning the End-of-Sequence (EOS) token as the content for chatting between " + }, + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "type": "inline_equation", + "content": "t_1 \\leq t < t_2" + }, + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "type": "text", + "content": ". However, this approach remains suboptimal. The dialogue prompt template (e.g., [INST], [/INST], space tokens in Llama [72, 73]) still consumes a considerable number of tokens per frame, which is less favorable due to the numerous frames in streaming video. Furthermore, the excessive number of EOS tokens in the sequence can significantly increase the language model's perplexity, as we have observed in our experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 335 + ], + "type": "text", + "content": "Instead, we propose a novel training objective named \"streaming EOS prediction\" to address this issue. We still assume " + }, + { + "bbox": [ + 46, + 288, + 287, + 335 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 288, + 287, + 335 + ], + "type": "text", + "content": " is essential for decoding language; thus, we normally learn language modeling here:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 354, + 287, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 354, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 52, + 354, + 287, + 379 + ], + "type": "interline_equation", + "content": "\\max P \\left(\\left[ \\mathrm {T x t} _ {i + 1} ^ {t _ {2}} \\right] \\mid \\left[ \\mathrm {C t x} ^ {< t _ {2}} \\right], \\left[ \\mathrm {F r a m e} ^ {t _ {2}} \\right], \\left[ \\mathrm {T x t} _ {\\leq i} ^ {t _ {2}} \\right]\\right). \\tag {3}", + "image_path": "bd79f0aba0c5a0b2d6cde713afe949ac96bbb140655d7ef481079426b2121b47.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "content": "However, for timestamps " + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "inline_equation", + "content": "t_1 \\leq t < t_2" + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "content": ", which are redundant for producing answers, we directly learn the model to predict EOS token on the frame tokens, i.e." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 436, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 436, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 51, + 436, + 287, + 460 + ], + "type": "interline_equation", + "content": "\\max P (\\text {E O S} \\left| [ \\text {C t x} ^ {< t} ], [ \\text {F r a m e} ^ {t} ]\\right), \\text {w h e r e} t _ {1} \\leq t < t _ {2}. \\tag {4}", + "image_path": "656d4b0b6bcefbc9942f59d8590fd78b6c225d62778cf6853adf73f35549c8af.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 604 + ], + "type": "text", + "content": "In this way, we \"skip\" a dialogue turn and learn to determine when it is appropriate to decode language for streaming inputs. During inference, if EOS is predicted on a frame, then we can directly ask the next frame to input. Meanwhile, the EOS token is not appended to the context to prevent it from affecting the language modeling. Therefore, this task is not about next-token prediction; however, it can work with autoregressive loss to train a video streaming dialogue model. In the following, we first introduce how can we get the data of streaming video and timestamped language annotations, then present the details about the model and the training procedure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 604, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 665 + ], + "type": "text", + "content": "We also note that the EOS token mentioned here is not limited to the real EOS token used in language models (e.g., " + }, + { + "bbox": [ + 46, + 604, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\langle / \\mathbb{S} \\rangle" + }, + { + "bbox": [ + 46, + 604, + 287, + 665 + ], + "type": "text", + "content": " in Llama). It is permissible to use any token or to introduce a new token, provided that it is specified in the system prompt. We use this term solely for simplicity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 92, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 92, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 92, + 681 + ], + "type": "text", + "content": "3.2.Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Online Annotations to Video Streaming Dialogue. Some video datasets, such as Ego4D narrations [26], are inher" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 545, + 351 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 351 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 351 + ], + "type": "image", + "image_path": "10c68d0ebedd7216acfc5738c9b8c84dc1920e547942e2b8504460528aa67ca1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 357, + 546, + 413 + ], + "lines": [ + { + "bbox": [ + 305, + 357, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 546, + 413 + ], + "type": "text", + "content": "Figure 2. The streaming dialogue data generation method in our LIVE framework. We randomly insert templated questions into the video timeline and \"expose\" the ground-truth video annotations (along with their timestamps) to LLMs, prompting them to answer the queries within a period of time." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 604 + ], + "type": "text", + "content": "ently collected in a streaming manner, with annotators providing real-time narrations as they watch a 5-minute long video clip. However, prior research [45, 90] has predominantly focused on learning from short, discrete clips (e.g. only 32 frames), rather than in a continuous streaming context. For this dataset, we followed the same instructions provided to human annotators [26] as our model training prompt. This prompt instructs the model to simulate human annotators to streamingly generate narrations in a 5-minute video (about 600 frames in 2 FPS). For demo purposes (not for experiments), we also utilize Llama-2-13B-Chat [73] to rephrase the narration text, correcting grammatical errors and typos, and converting it into a more understandable version (e.g., changing \"C does...\" to \"You do...\")." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Offline Annotations to Video Streaming Dialogue. Despite the Ego4D narration data being collected in a streaming manner, most prevalent video datasets [14, 26, 54, 71] are used to train offline models and only feature temporal segment annotations paired with basic language descriptions (e.g., activities, narrations). To bridge this gap, we propose a method for synthesizing dialogue data from these sources. As shown in Figure 2, our key idea is use LLM to generate user-assistant dialogues based on video annota" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18410" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 194, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 194, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 194, + 84 + ], + "type": "text", + "content": "tions, involving the following steps:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 286, + 380 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "type": "text", + "content": "- First, we prepare a question template library containing various queries about the past, present, and future tenses of the video, totaling " + }, + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "type": "text", + "content": " queries. We randomly sample one question from the library, denoted as " + }, + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 47, + 92, + 286, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "content": "- Then, we obtain the video annotation timeline from the offline dataset. This usually includes timestamped language descriptions, which we organize into a language prompt, e.g., \"time " + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "inline_equation", + "content": "t_a \\sim t_b" + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "content": ": boiling the water; time " + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "inline_equation", + "content": "t_c \\sim t_d" + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "content": ": cutting the vegetables.\", denoted as " + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "content": ". We consider all the state change critical timestamps as the ideal response times. For this example, " + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "inline_equation", + "content": "t_a, t_b, t_c" + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "inline_equation", + "content": "t_d" + }, + { + "bbox": [ + 47, + 140, + 286, + 236 + ], + "type": "text", + "content": " are all considered response times." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "text", + "content": "- Third, we prompt the large language model to generate responses at every critical timestamp, e.g., " + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "inline_equation", + "content": "t_a, t_b, t_c, t_d" + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "text", + "content": ", according to " + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "text", + "content": ". We can repeat this procedure for each " + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 47, + 236, + 286, + 296 + ], + "type": "text", + "content": " until all queries have been processed. The responses are saved for loading during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "text", + "content": "- Finally, during training, we (1) randomly sample a query and load its responses at critical timestamps, (2) randomly insert a query into a video timestamp " + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "t_r" + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "text", + "content": ", (3) discard the responses that occur before " + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "t_r" + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "text", + "content": ", and add a response at " + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "t_r" + }, + { + "bbox": [ + 47, + 297, + 286, + 380 + ], + "type": "text", + "content": ". Here different queries can be inserted into one video, which only requires discarding the responses of the previous query after the new query insertion timestamp." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 388, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 579 + ], + "type": "text", + "content": "In this way, we can generate temporally varied and freeform dialogue data within a video stream. We have prepared 50 questions each for past, current, and future events, totaling " + }, + { + "bbox": [ + 46, + 388, + 286, + 579 + ], + "type": "inline_equation", + "content": "N = 150" + }, + { + "bbox": [ + 46, + 388, + 286, + 579 + ], + "type": "text", + "content": " queries. We use Llama-2-13B-Chat to generate the responses and insert a maximum of 3 queries per training sample. The offline datasets we used are COIN [71] and Ego4D GoalStep [70] (for demo usage), which belong to the categories of egocentric and instructional video datasets, aligning with our aim to develop online video assistants. Here, we do not consider online action detection benchmarks (e.g., THUMOS14 [33], TVSeries [24]) because they are closed-set online classification benchmarks, and their labels are too brief, which may cause language models to generate hallucinatory responses. Please refer to the supplementary material for the generated dialogue example." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 143, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 143, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 143, + 609 + ], + "type": "text", + "content": "3.3. Model Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Model Architecture. We illustrate the model architecture in Figure 3. Similar to LLaVA [48, 49], it comprises three key components: an image encoder, an MLP projector, and a language model. For the image encoder, we utilize the CLIP ViT-L [20, 65] encoder (pretrained on DataComp1B [21]) to extract video frame embeddings at 2 FPS. Each video frame embedding has a shape of " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "(1 + h_p\\times w_p)\\times c" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "(1 + h_{p}\\times w_{p})" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " denotes the CLS token and average" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "pooled spatial tokens. The extracted frame embeddings from the image encoder are then fed into MLP projector to frame tokens, as in LLaVA-1.5 [48]. Then frame tokens are interleaved with language tokens as input to an LLM, Llama-2-7B-Chat [73]. Finally, we incorporate LoRA [31] in every linear layer of the LLM for efficient tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "type": "text", + "content": "Training Loss. As described in Section 3.1, our learning objective is twofold. The first part focuses on autoregressive language modeling, aiming to maximize the joint probability of input text sequences. The second training objective involves streaming EOS prediction, which requires the model to remain silent when it is unnecessary to output responses. With these two training objectives, we have language modeling (LM) loss and streaming loss terms to minimize, both employing cross-entropy loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 269, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 269, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 311, + 269, + 545, + 308 + ], + "type": "interline_equation", + "content": "L = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\left(\\underbrace {- \\log l _ {j + 1} P _ {j} ^ {[ T x t _ {j + 1} ]}} _ {L M L o s s} - \\underbrace {w \\log f _ {j} P _ {j} ^ {[ E O S ]}} _ {\\text {S t r e a m i n g L o s s}}\\right), \\tag {5}", + "image_path": "f8f4bf57a2918fe313c6bf4a513733519c870da375d8c7c82e7aabc35ec7874c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "l_{j}" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " denote condition indicators. " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "l_{j}" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " is 1 if the " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": "-th token is a language response token, and 0 otherwise. " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " is 1 if (1) the " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": "-th token is the last token of a frame" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": ", and (2) " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "l_{j+1} = 0" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": ". In essence, the streaming EOS loss is applied to frames before responding. " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "P_{j}^{[\\mathrm{Txt}_{j+1}]}" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " denotes the probability on " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "j + 1" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": "-th text token, output from the language model head of the " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": "-th token, and " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "P_{j}^{[\\mathrm{EOS}]}" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " represents that probability for the EOS token. " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " is a balance term, set to " + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 304, + 316, + 545, + 449 + ], + "type": "text", + "content": " by default. As shown in Figuer 3, we visualize the ranges of language loss and streaming loss in an input sequence when we only use 1 token for each frame." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 456, + 373, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 373, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 373, + 467 + ], + "type": "text", + "content": "3.4. Inference" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": "Probability Correction. The prevalence of EOS token will bias the model towards EOS token prediction. To address this, we introduce a threshold " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": " to correct the output probability on frame tokens: EOS will not be considered as the next token if " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "P_{j}^{[\\mathrm{EOS}]} < \\theta" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": ". In practical usage, we find that setting " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": " to 0.5 yields much better results than no threshold here." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": "Continuous Key-Value Cache. As shown in Figure 4, During inference, the video is input as a frame-by-frame stream, with a default FPS 2. Our model takes the current frame as input and generates tokens on-the-fly. In the whole process, we use the key-value cache trick to accelerate token decoding, thus we do not need to manually append the generated" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "type": "text", + "content": "2The experiments described in this paper are conducted without extra spatial tokens (i.e., " + }, + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "type": "inline_equation", + "content": "h_p = w_p = 0" + }, + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "type": "text", + "content": "), which is the most efficient setup and can handle half-hour videos within a 4096 context window. Our released models for demo usage include " + }, + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "type": "inline_equation", + "content": "1 + h_p \\times w_p = 1 + 3 \\times 3 = 10" + }, + { + "bbox": [ + 306, + 635, + 545, + 702 + ], + "type": "text", + "content": " tokens, offering better detail in dialogue but supporting shorter maximum video lengths. Despite more tokens used per frame, all models can still run at over 10 FPS for 5-minute Ego4D narration streams." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 703, + 530, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 703, + 530, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 703, + 530, + 712 + ], + "type": "text", + "content": "3When a frame has multiple patch tokens, loss is only on the last one." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18411" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 532, + 342 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 532, + 342 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 532, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 532, + 342 + ], + "type": "image", + "image_path": "f11632ceebe9151bacb8a426e61456bbb5f741b41c38654aad3375bcb54ee584.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 352, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 546, + 398 + ], + "type": "text", + "content": "Figure 3. The training method in our LIVE framework. We organize the user-assistant dialogue data and video frames in temporal order as the input sequence. To learn the model when to answer or keep silent in a video stream, we employ not only the standard language modeling (LM) loss but also introduce a streaming EOS prediction loss. This additional loss supervises the model when it is necessary to generate language, enabling it to produce temporally aligned responses and reduces the redundant dialogue history." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 415, + 285, + 504 + ], + "blocks": [ + { + "bbox": [ + 48, + 415, + 285, + 504 + ], + "lines": [ + { + "bbox": [ + 48, + 415, + 285, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 285, + 504 + ], + "type": "image", + "image_path": "5a5624900b49eba8690c05eaa1796c72854767a44d7e5fb14362a770626154e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 511, + 287, + 590 + ], + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 590 + ], + "type": "text", + "content": "Figure 4. Inference pipeline in our LIVE framework. During inference, video frames serve as streaming inputs. Our model maintains a continuous key-value cache as the input progresses to speed up the inference. Furthermore, we parallelize the fast video frame encoder and the slower language model to avoid the bottleneck in the latter. Video frame tokens can be always encoded and buffered, no need to wait the language decoding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 615, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 663 + ], + "type": "text", + "content": "tokens for the next frame. As our training encourages the model to keep silence, this continuous inference would be efficient, providing the possibility to pace with the video stream speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Parallelization of encoding and decoding. Our video frame encoder utilizes CLIP ViT-L (307M), which is significantly smaller than the 7B LLM. This size discrepancy leads to a speed mismatch, potentially resulting in frame" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 417, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 545, + 478 + ], + "type": "text", + "content": "skipping when the LLM decodes long sentences. To mitigate this issue, we parallelize the processes and establish a buffer for video frame tokens. Once the language model completes its decoding, it fetches all frame tokens from the buffer and inputs them in batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 514, + 387, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 387, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 387, + 528 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 542, + 439, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 439, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 439, + 555 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We implement VideoLLM-online-7B with our LIVE framework. It uses OpenCLIP-ViT-L [21, 65] as the video frame encoder, 2-layer MLP as multimodal connector, and Llama2-7B-Chat [73] as the language model. We only use 1 token for each video frame, sampled by 2 FPS in the following experiments. We add LoRA [31] to all linear layers of Llama-2, with a rank of 128, scaling factor of 256. For video streaming dialogue experiments, we train 2 epochs for the model. For the downstreaming offline experiments, we train 5 epochs without pre-training for fair comparison with previous methods. By default, we set streaming loss weight " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "w = 1.0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18412" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 534, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 534, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 534, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 534, + 145 + ], + "type": "table", + "html": "
MethodTraining ObjectiveEgo4D Narration Stream on Validation Set#Training Token↓Training Cost
LM-PPL↓TimeDiff ↓Fluency↑
No Training498.56.500.1%n/an/a
Interleaved DialogueLanguage Modeling2.456.4711.1%169412h
Per-frame Dialogue for StreamingLanguage Modeling (w/ EOS turns)3.342.5237.7 %673722h
Streaming Dialogue (Ours)Language Modeling + Streaming EOS2.432.3242.6%169412h
", + "image_path": "18893148faf0f495155dedd2dff82155ce06aaf19c9177c3e9f4b869b78d82e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 180, + 239, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 180, + 239, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 180, + 239, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 180, + 239, + 234 + ], + "type": "table", + "html": "
Streaming LossEgo4DNarrationStreamValidation
LM-PPL↓TimeDiff↓Fluency↑
Standard CE2.432.3242.6%
OHEM [68]2.532.3941.0%
Focal Loss [46]2.592.4439.4%
", + "image_path": "4d768ec8109d97f49e8fc8621e254d838c30989c993b75fdac038855d68b661f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 236, + 247, + 265 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 247, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 247, + 265 + ], + "type": "text", + "content": "(b) Streaming loss function. Standard CE (cross-entropy) is enough for training streaming dialogue; there is no need to specifically to address the class imbalance on EOS token." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 261, + 178, + 422, + 237 + ], + "blocks": [ + { + "bbox": [ + 261, + 178, + 422, + 237 + ], + "lines": [ + { + "bbox": [ + 261, + 178, + 422, + 237 + ], + "spans": [ + { + "bbox": [ + 261, + 178, + 422, + 237 + ], + "type": "table", + "html": "
Weight τEgo4D Narration Stream Validation
LM-PPL↓TimeDiff↓Fluency↑
τ = 0.52.442.3242.4%
τ = 1.02.432.3242.6%
τ = 2.02.462.3142.5%
τ = 3.02.472.3242.5%
", + "image_path": "cce9a4a7284fb4d4f2b5e9b97f7cab496c6bd81397bacfc8981db723f3770cc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 247, + 239, + 544, + 269 + ], + "lines": [ + { + "bbox": [ + 247, + 239, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 247, + 239, + 544, + 269 + ], + "type": "text", + "content": "(c) Streaming loss weight. Using slightly higher stream- (d) Generation memory/speed. ing loss weight " + }, + { + "bbox": [ + 247, + 239, + 544, + 269 + ], + "type": "inline_equation", + "content": "(\\tau = 2.0)" + }, + { + "bbox": [ + 247, + 239, + 544, + 269 + ], + "type": "text", + "content": " achieves the best trade-off Streaming dialogue method has among various metrics. much better efficiency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 436, + 176, + 542, + 237 + ], + "blocks": [ + { + "bbox": [ + 46, + 147, + 546, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 546, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 546, + 177 + ], + "type": "text", + "content": "(a) Learning method for streaming dialogue. Training with streaming dialogue method can achieve much better TimeDiff and Fluency, as well as maintain the language modeling quality. Meanwhile, the streaming dialogue can enjoy much more efficient training than per-frame dialogue for video streaming dialogue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 436, + 176, + 542, + 237 + ], + "lines": [ + { + "bbox": [ + 436, + 176, + 542, + 237 + ], + "spans": [ + { + "bbox": [ + 436, + 176, + 542, + 237 + ], + "type": "table", + "html": "
MethodMem↓FPS↑
Interleaved34.4G1.5
Per-frame Streaming24.9G7.5
Streaming18.2G13.5
", + "image_path": "26307c48fd021a9f627ced162ecd8e886ce006f494abd570be27c946c9d07f1a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 55, + 338, + 309, + 432 + ], + "blocks": [ + { + "bbox": [ + 55, + 338, + 309, + 432 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 309, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 309, + 432 + ], + "type": "table", + "html": "
MethodNot use HowTo100MCOIN Benchmark Top-1 Accuracy↑
StepTaskNextProc.Proc.+
ClipBERT [39]✓30.865.4---
TimeSformer [7]✗46.585.334.017.040.1
Paprika [92]✗51.085.843.2--
DistantSup [47]✗54.190.039.4-41.3
VideoTF [58]✗56.591.042.440.246.4
ProcedureVRL [91]✗56.990.846.8--
VideoTaskGraph [5]✗57.290.540.2--
VideoLLM-online✓59.892.148.147.952.9
", + "image_path": "e615b43d66f17278869ffc2d2a9ab43f73520801ce7e9bb0c6dda8bd98aa7b4b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 433, + 313, + 453 + ], + "lines": [ + { + "bbox": [ + 48, + 433, + 313, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 313, + 453 + ], + "type": "text", + "content": "(a) Results on COIN benchmarks (left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 337, + 539, + 433 + ], + "blocks": [ + { + "bbox": [ + 46, + 279, + 546, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 546, + 326 + ], + "type": "text", + "content": "Table 1. Ablation experiments on Ego4D Narration Stream. We train VideoLLM-online on Ego4D [26] narration stream training set and evaluate on its validation set. The comparison is based on our designed metrics: the ratio of strictly correct prediction tokens (Fluency), language modeling perplexity (LM-PPL) and time difference (TimeDiff) metrics. #Training Token denotes the average token length during training. TimeDiff refers to difference in second. Default settings are highlighted." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 337, + 539, + 433 + ], + "lines": [ + { + "bbox": [ + 318, + 337, + 539, + 433 + ], + "spans": [ + { + "bbox": [ + 318, + 337, + 539, + 433 + ], + "type": "table", + "html": "
MethodNot use EgoVLPEnd-to-end?Ego4D LTA ED@Z=20↓
VerbNounAction
CLIP [17]✓✓0.7390.7690.941
EgoT2 [79]✓✓0.7220.7640.935
I-CVAE [53]✓✓0.7530.7490.931
HierVL [4]✓✓0.7240.7350.928
VideoLLM [10]✗✓0.7210.7250.921
VideoLLM-online✓✓0.6970.6980.897
Palm [32]✗✗0.6960.6510.886
AntGPT [88]✗✗0.6500.6500.877
", + "image_path": "4e0406155103279d5189f34a8a5893320d12949d013ecb443970d2ee35892869.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 434, + 543, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 543, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 543, + 453 + ], + "type": "text", + "content": "(b) Results on Ego4D LTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 463, + 546, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 463, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 546, + 485 + ], + "type": "text", + "content": "Table 2. Experiments on COIN [71] and Ego4D [26] benchmarks. VideoLLM-online is finetuned on their training set, and strictly evaluated on the test set by generated string comparison with the ground-truth text. It achieves best results among end-to-end models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 497, + 157, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 157, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 157, + 510 + ], + "type": "text", + "content": "4.2. Evaluation Setting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 518, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 555 + ], + "type": "text", + "content": "Datasets. We use (1) assistance-related, instructional video dataset COIN [71] and (2) continuous, egocentric video dataset Ego4D [26] in various settings:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": "- Ego4D Narration Stream: We also leverage the dense Ego4D timestamp-narration to create a streaming set. The goal is to generate narrations timely like Ego4D human annotators [26]. We follow the division of the training, validation, and test set in EgoVLP [45]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "content": "- COIN+Ego4D Stream: To further evaluate the potential of model's performance to free-form dialogue, we construct a simple COIN+Ego4D Stream set, constructed from COIN annotations using our data generation methods, and the above Ego4D Narration Stream. The model should remind the user when an action starts, summarizes the action when it ends, as well as forecasts the next action. We use the same training/testing splits as COIN" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 498, + 541, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 541, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 541, + 510 + ], + "type": "text", + "content": "benchmarks. See supplementary material for the results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 510, + 545, + 629 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 306, + 510, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 545, + 569 + ], + "type": "text", + "content": "- COIN Benchmarks: Following on previous studies [47, 58, 80, 91], we evaluate our model on six common benchmarks of the COIN dataset: step recognition, step forecasting, task summarization, procedure forecasting, procedure forecasting with a goal." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 570, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 545, + 629 + ], + "type": "text", + "content": "- Ego4D long-term action anticipation (LTA) benchmark: This benchmark requires to predict next " + }, + { + "bbox": [ + 306, + 570, + 545, + 629 + ], + "type": "inline_equation", + "content": "Z = 20" + }, + { + "bbox": [ + 306, + 570, + 545, + 629 + ], + "type": "text", + "content": " actions (verbs and nouns) for the given video of previous 8 steps. We use the standard Ego4d v2 splits as in previous studies [32, 88]." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 633, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 545, + 656 + ], + "type": "text", + "content": "Evaluation metrics. We use the following metrics to evaluate the model as an online video assistant:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 306, + 665, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 545, + 701 + ], + "type": "text", + "content": "- Language modeling perplexity (LM-PPL). This indicates the language modeling capability at a given times-tamp. A lower LM-PPL is better for correct answering." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "- Time Difference (TimeDiff). To evaluate the temporal" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18413" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 119 + ], + "type": "text", + "content": "alignment capability of an online assistant, we calculate the discrepancy between the timestamp of its response and the expected timestamp for each response. We average TimeDiff each turn as the metric." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 286, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 286, + 202 + ], + "type": "text", + "content": "- Fluency. Individual LM-PPL or TimeDiff do not entirely evaluate both language and temporal effectiveness in a streaming dialogue. We introduce the Fluency metric, which evaluates the proportion of consecutive successful token prediction within a dialogue turn. As the token also including language tokens, Fluency can comprehensively reflect the aspects of LM-PPL and TimeDiff." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 207, + 286, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 286, + 301 + ], + "type": "text", + "content": "Baselines. To our best knowledge, we are the first one to tackle producing temporal aligned, free-form language answer with streaming video settings. To better understand the challenges, we build baseline models for video-text interleaved dialogue, per-frame dialogue, as we described in Section 3.1, with the same model architecture and training details to VideoLLM-online, differing only in their training objective and multi-turn formulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 311, + 140, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 140, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 140, + 323 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 329, + 286, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 286, + 436 + ], + "type": "text", + "content": "Learning Method. Table 1a shows the ablation studies on learning methods in a streaming setting. Both vision-language interleaved and streaming methods exhibit low perplexity loss, indicating that our proposed objective does not hurt language modeling capability. However, learning with per-frame for streaming will produce significant higher LM-PPL than others, which might be attributed to the too more single EOS token in answering that affects the original language modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 437, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 286, + 557 + ], + "type": "text", + "content": "When we turn to online metrics of TimeDiff and Fluency, streaming dialogue method yields much better results than others. In our observation, the first interleaved dialogue method always outputs language after every frame, and the second multi-turn for streaming dialogue approach tends to answer EOS token after every frame, which decreases their performance for streaming video inputs. Furthermore, per-frame for streaming dialogue method will significantly slow down the training speed due to its lengthy prompts, while our method has no negative impact on the efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 558, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 286, + 628 + ], + "type": "text", + "content": "Streaming Loss. We continue to investigate the most suitable strategy to learn the streaming objective. As shown in Table 1b and Table 1c, we find a default setting works surprisingly well (CE loss, " + }, + { + "bbox": [ + 47, + 558, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\tau = 1.0" + }, + { + "bbox": [ + 47, + 558, + 286, + 628 + ], + "type": "text", + "content": "), which demonstrates there is no need to apply more advanced loss (e.g. Focal Loss [46]) to address the imbalance on EOS token." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 630, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 286, + 712 + ], + "type": "text", + "content": "Inference Efficiency. In Table 1d, we test the inference efficiency on Ego4D narration stream validation set (5 minute), and report the memory cost and average FPS on a single A100 GPU. The first interleaved dialogue method, which will output language after every video frame, has huge memory cost and slow generation speed. The second one, per-frame dialogue for streaming that formulates" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "all in a multi-turn dialogue, show better efficiency than the first one since it can cost less tokens in redundant frames. However, this approach still lags significantly behind our streaming dialogue approach, which does not cost extra tokens in redundant frames thus maintain smaller key-value cache. We observe for most Ego4D videos, our model can run larger than 10 FPS, providing possibility for AI assistants working in real-time video stream." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 175, + 362, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 362, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 362, + 186 + ], + "type": "text", + "content": "4.4. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 193, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 359 + ], + "type": "text", + "content": "Offline language modeling. We show our model can perform well on traditional temporal summarization and forecasting problems. As shown in Table 2(a), our model achieves state-of-the-art performances in step/task summarization and next step/procedure forecasting benchmarks of COIN dataset [71]. Furthermore, we also obtain the best performance among end-to-end models evaluated on Ego4D LTA. We observe the results of Palm [32] and AntGPT [88] are better than us, but they used egocentric pre-trained visual feature [45], and integrates lots of complex cascading methods to improve the forecasting results. Our VideoLLM-online, however, directly outputs language as the results, which performs better than the similar end-to-end VideoLLM [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "text", + "content": "Visualization. See Figure 1, we visualize two representative examples, real-time narration and online dialogue. The most distinctive characteristics of our approach are: (1) the dialogue process goes along with the streaming video input, rather than chatting based on the full video. (2) The response will be \"muted\" when it is unnecessary, significantly improving the overall speed of video streaming dialogue." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 654 + ], + "type": "text", + "content": "We propose Learning-In-Video-strEam (LIVE), a novel framework empowering LLMs to handle streaming video, to produce temporal aligned answers, hold long-context video duration, and have high inference efficiency. We use LIVE to train a simple VideoLLM-online model, which not only achieves superior capability in online/offline vision-language tasks, but also enable fast inference for an online video streaming setting. We believe enabling such abilities will be an important step to move towards always-on online assistant. In future work, to make our VideoLLM-online be more general and improve its spatial capability in zero-shot prediction for downstream applications, we will explore suitable pre-training data source, and develop models that can employ more spatial tokens but without many trade-off on speed and memory cost." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment This work is sponsored by Project Aria Team, Meta. The datasets and processing were acquired and all models were trained at the National University of Singapore (NUS) by NUS authors." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18414" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Huda AlAmri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K. Marks, Chiori Hori, Peter Anderson, Stefan Lee, and Devi Parikh. Audio visual scene-aware dialog. In CVPR, pages 7558-7567, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 256 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 258, + 288, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 433 + ], + "type": "text", + "content": "[3] Rohan Anil, Sebastian Borgeaud, Yonghui Wu, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Slav Petrov, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy P. Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul Ronald Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, and et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 287, + 476 + ], + "type": "text", + "content": "[4] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In CVPR, pages 23066-23078, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 479, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 287, + 522 + ], + "type": "text", + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "type": "text", + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 570, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 287, + 601 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 604, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 604, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 604, + 288, + 713 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In *NeurIPS*, pages 1877–1901, 2020. 1, 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[9] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[10] Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292, 2023. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "text", + "content": "[11] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C.H.Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 332, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 397 + ], + "type": "text", + "content": "[14] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. Int. J. Comput. Vis., 130(1):33-55, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "text", + "content": "[15] Tri Dao, Daniel Haziza, Francisco Massa, and Grigory Sizov. Flash-decoding for long-context inference. https://pytorch.org/blog/flash-decoding/, 2023.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "text", + "content": "[16] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In CVPR, pages 1080-1089, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 468, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 498 + ], + "type": "text", + "content": "[17] Srijan Das and Michael S. Ryoo. Video + clip baseline for ego4d long-term action anticipation. arXiv:2207.00579, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "text", + "content": "[19] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv:2309.11499, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 592, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 656 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[21] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M. Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18415" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 148 + ], + "type": "text", + "content": "Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. arXiv:2304.14108, 2023. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[22] Zhe Gan, Yu Cheng, Ahmed El Kholy, Linjie Li, Jingjing Liu, and Jianfeng Gao. Multi-step reasoning via recurrent dual attention for visual dialog. In ACL, pages 6463-6474, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[23] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms. arXiv:2310.01801, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "text", + "content": "[24] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In ECCV, pages 269-284, 2016. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 294 + ], + "type": "text", + "content": "[25] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. Arxiv, 2106.02036, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 296, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 296, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 296, + 287, + 581 + ], + "type": "text", + "content": "[26] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18973-18990, 2022. 2, 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 582, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 713 + ], + "type": "text", + "content": "[27] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, FuJen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abraham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 138 + ], + "type": "text", + "content": "ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, and et al. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. arXiv:2311.18259, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "text", + "content": "[28] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv:2309.03905, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[29] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[30] Ke Hong, Guohao Dai, Jiaming Xu, Qiuli Mao, Xiuhong Li, Jun Liu, Kangdi Chen, Hanyu Dong, and Yu Wang. Flashdecoding++: Faster large language model inference on gpus. arXiv:2311.01282, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "text", + "content": "[31] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "text", + "content": "[32] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models @ ego4d long-term action anticipation challenge 2023. arXiv:2306.16545, 2023. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[33] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorbani, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos \"in the wild\". Comput. Vis. Image Underst., 155:1-23, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "text", + "content": "[34] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv:2001.08361, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "text", + "content": "[35] Kris M. Kitani, Brian D. Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In ECCV, pages 201-214, 2012. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "text", + "content": "[36] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. arXiv:2309.06180, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "text", + "content": "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv:2308.00692, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[39] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 2, 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18416" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "text", + "content": "[42] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv:2305.06355, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 198 + ], + "type": "text", + "content": "[43] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 200, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 287, + 255 + ], + "type": "text", + "content": "[44] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, Ce Liu, and Lijuan Wang. MM-VID: advancing video understanding with gpt-4v(ision). arXiv:2310.19773, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 322 + ], + "type": "text", + "content": "[45] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Egocentric video-language pretraining. arXiv:2206.01670, 2022. 4, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 323, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 287, + 357 + ], + "type": "text", + "content": "[46] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 358, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 403 + ], + "type": "text", + "content": "[47] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 439, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 287, + 461 + ], + "type": "text", + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "type": "text", + "content": "[50] Xiaoxuan Liu, Lanxiang Hu, Peter Bailis, Ion Stoica, Zhijie Deng, Alvin Cheung, and Hao Zhang. Online speculative decoding. arXiv:2310.07177, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "type": "text", + "content": "[51] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv:2306.09093, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "type": "text", + "content": "[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "text", + "content": "[53] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action anticipation. In WACV, pages 6037-6046, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[55] Seungwhan Moon, Andrea Madotto, Zhaojiang Lin, Tushar Nagarajan, Matt Smith, Shashank Jain, Chun-Fu Yeh," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Prakash Murugesan, Peyman Heidari, Yue Liu, et al. Anymal: An efficient and scalable any-modality augmented language model. arXiv:2309.16058, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[56] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv:2305.15021, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "text", + "content": "[57] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for visual dialog: A simple state-of-the-art baseline. In ECCV, pages 336-352, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "text", + "content": "[58] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv:2303.13519, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[59] Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Efficient attention mechanism for visual dialog that can handle all the interactions between multiple inputs. In ECCV, pages 223–240, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 273, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 294 + ], + "type": "text", + "content": "[60] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt/, 2023.1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 295, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 315 + ], + "type": "text", + "content": "[61] OpenAI. GPT-4 technical report. arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[62] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.2.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 426 + ], + "type": "text", + "content": "[63] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 472, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 536 + ], + "type": "text", + "content": "[65] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "type": "text", + "content": "[66] Idan Schwartz, Seunghak Yu, Tamir Hazan, and Alexander G. Schwing. Factor graph attention. In CVPR, pages 2039-2048, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[67] Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Re, Ion Stoica, and Ce Zhang. Flexgen: High-throughput generative inference of large language models with a singlegpu. arXiv:2303.06865, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "type": "text", + "content": "[68] Abhinav Shrivastava, Abhinav Gupta, and Ross B. Girshick. Training region-based object detectors with online hard example mining. In CVPR, pages 761-769, 2016. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[69] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv:2307.16449, 2023. 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18417" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[70] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 163 + ], + "type": "text", + "content": "[71] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 218 + ], + "type": "text", + "content": "[72] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023. 1, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 450 + ], + "type": "text", + "content": "[73] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv:2307.09288, 2023. 2,456" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "text", + "content": "[74] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 6000-6010, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "type": "text", + "content": "[75] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Yuanjie Shao, Zhengrong Zuo, Changxin Gao, and Nong Sang. Oadtr: Online action detection with transformers. In ICCV, pages 7545-7555, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[76] Yue Wang, Shafiq R. Joty, Michael R. Lyu, Irwin King, Caiming Xiong, and Steven C. H. Hoi. VD-BERT: A unified vision and dialog transformer with BERT. In EMNLP, pages 3325-3338, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "text", + "content": "[77] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv:2309.05519, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "type": "text", + "content": "[78] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv:2309.17453, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "text", + "content": "[79] Zihui Xue, Yale Song, Kristen Grauman, and Lorenzo Torresani. Egocentric video task translation. In CVPR, pages 2310-2320, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[80] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 689 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Schmid. Unloc: A unified framework for video localization tasks. In ICCV, pages 13623-13633, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[81] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 217 + ], + "type": "text", + "content": "[82] Zhewei Yao, Xiaoxia Wu, Conglong Li, Minjia Zhang, Heyang Qi, Olatunj Ruwase, Ammar Ahmad Awan, Samyam Rajbhandari, and Yuxiong He. Deepspeed-visualchat: Multi-round multi-image interleave chat via multi-modal causal attention. arXiv:2309.14327, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[83] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv:2310.07704, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[84] Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. Orca: A distributed serving system for {Transformer-Based} generative models. In OSDI, pages 521-538, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 308, + 545, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 362 + ], + "type": "text", + "content": "[85] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv:2309.02591, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "type": "text", + "content": "[86] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 398, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 398, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 545, + 441 + ], + "type": "text", + "content": "[87] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 442, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 486 + ], + "type": "text", + "content": "[88] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? In ICLR, 2024. 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 487, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 520 + ], + "type": "text", + "content": "[89] Yue Zhao and Philipp Kähenbuhl. Real-time online video detection with temporal smoothing transformers. In ECCV, pages 485-502, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 521, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 554 + ], + "type": "text", + "content": "[90] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "text", + "content": "[91] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In CVPR, pages 14825-14835, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "text", + "content": "[92] Honglu Zhou, Roberto Martín-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 644, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 644, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 644, + 545, + 689 + ], + "type": "text", + "content": "[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18418" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_content_list.json b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0d23c1c67f411ea5d315461e2aece6999b01dbe6 --- /dev/null +++ b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_content_list.json @@ -0,0 +1,2110 @@ +[ + { + "type": "text", + "text": "VideoMAC: Video Masked Autoencoders Meet ConvNets", + "text_level": 1, + "bbox": [ + 196, + 130, + 774, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gensheng Pei\\*, Tao Chen\\*, Xiruo Jiang, Huafeng Liu, Zeren Sun,\\* Yazhou Yao\\* \nSchool of Computer Science and Engineering, Nanjing University of Science and Technology \nhttps://github.com/NUST-Machine-Intelligence-Laboratory/VideoMAC", + "bbox": [ + 181, + 180, + 792, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the advancement of self-supervised learning techniques, like masked autoencoders (MAE), has greatly influenced visual representation learning for images and videos. Nevertheless, it is worth noting that the predominant approaches in existing masked image / video modeling rely excessively on resource-intensive vision transformers (ViTs) as the feature encoder. In this paper, we propose a new approach termed as VideoMAC, which combines video masked autoencoders with resource-friendly ConvNets. Specifically, VideoMAC employs symmetric masking on randomly sampled pairs of video frames. To prevent the issue of mask pattern dissipation, we utilize ConvNets which are implemented with sparse convolutional operators as encoders. Simultaneously, we present a simple yet effective masked video modeling (MVM) approach, a dual encoder architecture comprising an online encoder and an exponential moving average target encoder, aimed to facilitate inter-frame reconstruction consistency in videos. Additionally, we demonstrate that VideoMAC, empowering classical (ResNet) / modern (ConvNeXt) convolutional encoders to harness the benefits of MVM, outperforms ViT-based approaches on downstream tasks, including video object segmentation (+5.2% / 6.4% J&F), body part propagation (+6.3% / 3.1% mIoU), and human pose tracking (+10.2% / 11.1% PCK@0.1).", + "bbox": [ + 76, + 301, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spurred by the success of masked language modeling in natural language processing [3, 10, 12], transformer-based visual models, notably vision transformers (ViT) [14] adopt masked image modeling (MIM) are attracting attention [1, 2, 15, 23, 34, 67]. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations.", + "bbox": [ + 75, + 738, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca4cde1d18bf259bdb656eb6c157807b92cbcd0f569194a2f8bd19d1734e5573.jpg", + "image_caption": [ + "Figure 1. State-of-the-art MAE architectures (e.g., MAE-ST [16] and SiamMAE [20]) in masked video modeling commonly employ ViT-based encoders. We propose VideoMAC, a new video masked autoencoder built using pure ConvNets. In this study, we model VideoMAC with two of the most emblematic families of ConvNets, namely ConvNeXtv2 [64] and ResNet [21]. Notably, VideoMAC exhibits superior performance on a range of downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26], compared to ViT-based methods [16, 20]." + ], + "image_footnote": [], + "bbox": [ + 501, + 266, + 890, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the successful application of MIM, some pioneering works [16, 56] have shifted the focus of video representation learning to masked video modeling (MVM), yielding promising results in downstream video tasks. However, features learned by MVM are designed explicitly for pixel reconstruction tasks. They perform well when fine-tuned for downstream video tasks but exhibit subpar performance without specific fine-tuning [20]. Their performance is even inferior to MIM methods [23, 30], and it lags significantly behind supervised methods [21, 36].", + "bbox": [ + 496, + 595, + 893, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We strive to discern the factors contributing to this issue. To the best of our knowledge, most current MVM methods are based on the isotropic design of ViT [16, 20, 56, 65]. The reason is that early MIM / MVM methods are explicitly developed for ViT, as they allow the use of only visible patches as the sole input to the encoder, which aligns well with the ViT architecture. For instance, the representative approach, masked autoencoders (MAE) [23], enhances the efficiency and effectiveness of model pre-training by focusing predominantly on the unmasked regions. However,", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22733", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9dd50d2de52ff30ad19264f6594a7ddac788c602a2a31c43d24a041d7ea95ab6.jpg", + "image_caption": [ + "(a)", + "Input" + ], + "image_footnote": [], + "bbox": [ + 112, + 98, + 194, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b906e3d8a645712c24a9cdbb5f24f335869a51f2ce49c52c500e828aa989f48a.jpg", + "image_caption": [ + "Masked" + ], + "image_footnote": [], + "bbox": [ + 199, + 99, + 281, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5d91cf721317ff40cad8b85c77c3131f313349c43d3546d585cb74dd3945a40d.jpg", + "image_caption": [ + "Reconstructed" + ], + "image_footnote": [], + "bbox": [ + 285, + 99, + 366, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d6dae044ebd5e2836c4c184376f9ae28cec74de1da4a4a676d23fbf699acd5cb.jpg", + "image_caption": [ + "Heatmap" + ], + "image_footnote": [], + "bbox": [ + 370, + 99, + 452, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b434930b33d94696817c659d12af41a5de19fd5199265af242645b34307bea6f.jpg", + "image_caption": [ + "(b)", + "Figure 2. Visualization of the heatmap during the reconstruction of masked patches in a random frame pair. It is evident that our approach highlights similar regions for both (a) past and (b) current frames, proficiently reconstructing colors and contours." + ], + "image_footnote": [], + "bbox": [ + 112, + 166, + 194, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/472226c5798a3a7ed092aec83705bde8c203ebbce9f7ef47c68ffa4fab33049a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 166, + 281, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c1999b4597d7fa62c72dbb2107d52fcf31a4f4ad78448beed7b1bbed1622bd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 166, + 366, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f91653fc48b39211fc0570d2fdaa2375e0060af10dd9bc2808bba78420ae9e3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 166, + 452, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "previous MVM approaches [16, 56], confined by the plain vision transformer design devoid of a hierarchical structure tend to result in inadequate spatial modeling capabilities for local features [17, 35, 49, 62, 70]. Such capabilities are crucial for dense downstream tasks like segmentation and detection. Moreover, the uniform reliance on ViT in existing MVM methods has resulted in significant computational resource consumption [16, 56, 58]. One might first consider ConvNets [21, 28, 29, 36, 50, 53], characterized by their low resource consumption and built-in hierarchical structure, as an ideal solution to tackle the aforementioned issues. Nevertheless, ConvNets appears understudied in MVM. Consequently, it raises the question: can ConvNets, endowed with a pyramid vision architecture, guide MVM towards the era of hierarchical pre-training?", + "bbox": [ + 75, + 296, + 467, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this concern, we propose a simple yet potent MVM approach, referred to as Video Masked Autoencoders with ConvNets (VideoMAC). Uniquely, our method represents an initial endeavor to supplement video pre-training by integrating hierarchical ConvNets. However, applying ConvNets to MVM proves more complex than utilizing ViTs. Typical ConvNets might be incompatible with MVM due to the employment of dense sliding windows with overlap, resulting in information leakage within masked regions. To efficiently incorporate mask modeling into ConvNets, we opt for sparse convolution over dense convolution to restructure the original ConvNets, an approach validated in recent studies [54, 64] to effectively maintain the structure of the masked region when utilizing ConvNets. Another hurdle for MVM is the effective incorporation of temporal data. Current MVM pre-training methods [16, 56] involving multi-frame inputs are primarily used in downstream resource-intensive tasks such as video action recognition. Although methods using two-frame inputs for pre-training lower resource usage [20, 65], they typically employ the Siamese architecture, necessitating robust skills for designing reconstruction targets. Importantly, all aforementioned methods utilize computationally intensive, non-hierarchical ViT [14] as their encoders. To this end, we utilize pairs of frames as inputs and introduce the reconstruction consis", + "bbox": [ + 75, + 523, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tency loss, a streamlined extension of MAE to accommodate spatio-temporal data. We formulate online networks through gradient optimization and establish target networks by updating parameters via an exponential moving average (EMA). This allows for individual recovery of masked frame pairs and computation of the reconstruction consistency loss between them. As illustrated in Fig. 2, our approach can focus on similar regions between two frames and reconstruct the masked portions efficiently.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notably, our approach transcends previous state-of-the-art techniques in three downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]. Moreover, we present promising results utilizing VideoMAC in fine-tuning for image recognition on ImageNet1K [11]. To summarize, our main contributions of this work include:", + "bbox": [ + 496, + 227, + 890, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a simple yet potent solution of masked video autoencoders, marking the initial endeavor to successfully implement MVM using pure ConvNets.", + "- We introduce the reconstruction consistency loss to enable elegant modeling of spatio-temporal data.", + "- Our approach significantly outperforms existing ViT-based MVM methods in various downstream tasks." + ], + "bbox": [ + 517, + 333, + 890, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the focus of the computer vision community is increasingly shifting towards vision transformers, we trust that VideoMAC will rekindle interest in exploring ConvNet-based MVM, thereby catalyzing the discovery of new potentials across a variety of video tasks.", + "bbox": [ + 496, + 439, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 527, + 638, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image Modeling. Self-supervised learning (SSL) has been rapidly developing in computer vision, demonstrating substantial potential for deriving valuable representations from extensive, unlabeled datasets. A key advantage of SSL [2, 5-8, 18, 22, 48, 57, 69, 72] is its independence from costly human annotations, thereby providing robust prior knowledge across various visual tasks. The success of masked language modeling in NLP [3, 10, 12, 13, 47] has catalyzed a rising interest in masked image modeling (MIM) for visual pre-training. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations [2, 23, 67]. Notably, pioneering works [54, 64] have recently applied ConvNets to MIM pre-training, paving the way for masking modeling methods based on ConvNets.", + "bbox": [ + 496, + 551, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Video Modeling. Recently, researchers have embarked on exploring MVM approaches [16, 19, 65, 66] to comprehend unsupervised video representations, exhibiting effectiveness across a multitude of downstream tasks [4, 25, 31, 32, 40-44, 60, 61, 71, 74, 75]. However, current MVM methodologies [20, 56, 58, 59] primarily depend on the isotropic ViT, a visual backbone compatible with mask", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22734", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing modeling. Furthermore, these ViT-based MVM strategies may encounter challenges with inadequate local feature modeling when dealing with dense tasks such as segmentation and detection due to their non-hierarchical structure [35, 37, 62, 70]. To the best of our knowledge, no MVM approaches are explicitly designed for hierarchical network-based architectures, especially ConvNets.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that while the isotropic ViT is well suited for the implementation of MVM methods [16, 19, 56], this is exceptionally difficult for ConvNets. In this paper, our VideoMAC is committed to developing an efficient video pre-training tool that can equip ConvNets with superior, or at least comparable, performance to ViTs.", + "bbox": [ + 75, + 196, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 301, + 184, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our VideoMAC builds upon MAE [23] and, pioneeringly, facilitates masked video modeling with a purely convolutional architecture. As depicted in Fig. 3, VideoMAC is designed to reconstruct masked patches, which are randomly and symmetrically masked on frame pairs at a high masking ratio (0.75 as used in this study). The framework achieves such reconstructions by exclusively delivering visible patches to both the online and target encoders. In this section, we describe its main components in detail.", + "bbox": [ + 75, + 328, + 467, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 474, + 217, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, masked autoencoders (MAE) [23] introduces a novel methodology to the domain of self-supervised visual representation learning. The primary process involves training a model to reconstruct masked areas in images whose contents are partially occluded. The fundamental architecture of a standard MAE model is composed of three core elements: the masking design, an encoder, and a decoder. In mathematical terms, given an image $\\pmb{I} \\in 3 \\times H \\times W$ , the MAE model patchifies $\\pmb{I}$ into $N$ distinct, non-overlapping patches, represented by $\\hat{\\pmb{I}} \\in N \\times (P^2 \\times 3)$ , each with a size of $P$ . Masks are randomly applied to a subset of these patches, denoted by the subset $\\mathcal{M}$ .", + "bbox": [ + 75, + 498, + 467, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder $f_{\\mathcal{E}}$ processes the visible patches $\\mathcal{V}$ , producing latent representations that are represented as $z = f_{\\mathcal{E}}(\\mathcal{V})$ . Then, the decoder $f_{\\mathcal{D}}$ attempts to restore the masked subset $\\mathcal{M}$ utilizing these latent representations, generating $O = f_{\\mathcal{D}}(z; \\mathcal{M})$ , where $O \\in N \\times (P^2 \\times 3)$ depicts the reconstructed patches. The MAE model utilizes the mean-squared error (MSE) to compute the reconstruction loss $\\mathcal{L}_{rec}$ , which is solely calculated based on the masked patches and is formulated as $\\mathcal{L}_{rec} = \\frac{1}{|\\mathcal{M}|} \\sum_{i \\in \\mathcal{M}} || \\hat{I}_i - O_i ||_2^2$ .", + "bbox": [ + 75, + 681, + 467, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video Masked Autoencoder with ConvNets", + "text_level": 1, + "bbox": [ + 76, + 830, + 441, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masking. In this work, we randomly select two consecutive frames, $\\pmb{I}^{(1)},\\pmb{I}^{(2)}\\in 3\\times H\\times W$ and divide them into $N$ nonoverlapping patches, represented as $\\hat{I}^{(1)},\\hat{I}^{(2)}\\in N\\times (P^2\\times 3)$", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6cfdd4f66713a37f9c4aa63786bda92e18bec84f821946013e1b007172e4bf2.jpg", + "image_caption": [ + "Figure 3. An illustration of VideoMAC for ConvNet-based MVM. During pre-training, we mask $75\\%$ of symmetric patches from two frames randomly. In our VideoMAC, the MVM of frame pairs is achieved by an online network optimized by gradients $(\\square)$ , online loss $\\mathcal{L}_o$ and a target network updated by EMA $(\\square)$ , target loss $\\mathcal{L}_t$ . $\\mathcal{L}_c$ is computed as the reconstruction consistency loss between reconstructed patches of frame pairs." + ], + "image_footnote": [], + "bbox": [ + 498, + 85, + 890, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regarding the mask configuration, our approach adopts symmetric masking (Table 3d) to guarantee consistency in the positions of the masked areas between frame pairs. Given the consecutive frames, $\\mathcal{M}^{(1)}$ and $\\mathcal{M}^{(2)}$ constitute the subsets of masked patches, whereas $\\mathcal{V}^{(1)}$ and $\\mathcal{V}^{(2)}$ correspond to the visible patches.", + "bbox": [ + 496, + 325, + 890, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is inherent that images and videos possess a greater level of information redundancy [16, 22] compared to languages. Accordingly, we employ a masking ratio of 0.75 (Table 4f) in our methodology. Subsequent empirical analyses (Sec. 4.2) have demonstrated an improvement in performance with higher masking ratios. Nevertheless, employing excessively high masking ratios may induce the risk of training collapse, thereby leading to sudden and substantial decreases in model performance.", + "bbox": [ + 496, + 417, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAC Encoder. In ViT-based models, the prevention of information leakage from masked regions is comparatively straightforward [64, 70]. These models solely employ the visible patches as the exclusive input to the encoder, thereby allowing ViT-based approaches [16, 20, 56, 65] to amplify the efficiency and efficacy of model pre-training by concentrating dedicatedly on unmasked regions during this phase. In contrast, achieving a similar task with ConvNets poses a more intricate dilemma, necessitating the maintenance of the 2D image structure. There are two naive solutions: (i) the instigation of learnable masked markers at the input stage, and (ii) the transformation of the hierarchical convolutional network into an isotropic structure [36] that mirrors the ViT-style technique (Table 2). Regardless, both strategies may decrease pre-training efficiency and possibly engender inconsistencies between training and testing.", + "bbox": [ + 496, + 553, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper is at the forefront of proposing an entirely ConvNet-based MVM approach to address this challenge. Drawing inspiration from [54, 64], we utilize sparse convolutional networks to develop a masked autoencoder, which usurps traditional dense convolution. This strategy aims to meet three primary goals: (i) maintaining the structural integrity of the image during feature encoding, (ii) preventing", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22735", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mask dissipation in the course of hierarchical convolution, and (iii) markedly amplifying the computational efficiency of MVM. Specifically, our VideoMAC uses a fully convolutional architecture. Conventionally, ConvNets are structured with four stages, where the resolution of features is halved after each stage, representing a patch size of 32. For an input frame size of $224 \\times 224$ , the features are sequentially downsampled at every stage to achieve resolutions of $56 \\times 56$ , $28 \\times 28$ , $14 \\times 14$ , and $7 \\times 7$ , respectively. Subsequently, masks are produced and recursively upsampled to the highest resolution (i.e., $56 \\times 56$ ) in the final stage.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current MVMs typically necessitate significant computational resources, posing a challenge for practical applications. Our strategy, which substitutes ConvNets for ViTs, does mitigate computational complexity. Furthermore, studying existing MVM methods, we observe that to capture temporal signals within videos, most methods either utilize multi-frame data input or opt for a siamese encoder structure. We diverge from these widespread strategies and instead adopt a more resource-efficient method, namely the online-target encoder architecture (Fig. 3), commonly applied in self-supervised image representation learning. The target encoder's weights are updated by the online encoder via EMA. This architectural integration into our MVM encoding process results in three-fold advantages: (i) considerable reduction of computational complexity, (ii) joint optimization of frame-to-feature representation by the online and target encoders, and (iii) simplicity in incorporating existing ConvNets into our framework, thus improving the performance of downstream tasks. Consequently, visible patches $\\mathcal{V}^{(1)}$ and $\\mathcal{V}^{(2)}$ of consecutive frames are processed by the online encoder $f_{\\mathcal{E}}^{o}$ and the target encoder $f_{\\mathcal{E}}^{t}$ to generate latent representations $z^{(1)} = f_{\\mathcal{E}}^{o}(\\mathcal{V}^{(1)})$ and $z^{(2)} = f_{\\mathcal{E}}^{t}(\\mathcal{V}^{(2)})$ , respectively.", + "bbox": [ + 75, + 260, + 472, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MAC Decoder. Unlike previous methods that utilize hierarchical decoders [54, 55] or transformer decoders[16, 33, 56], our VideoMAC implements a lightweight decoder derived from the ConvNeXt block [36]. This simplifies the design and bolsters the efficiency of the pre-training process. Analogous to our encoder, we construct an online decoder $f_{\\mathcal{D}}^{o}$ and a target decoder $f_{\\mathcal{D}}^{t}$ , further employing EMA for parameter updates (Fig. 3). Given the latent representations $z^{(1)}$ and $z^{(2)}$ from two successive frames, decoders $f_{\\mathcal{D}}^{o}$ and $f_{\\mathcal{D}}^{t}$ reconstruct masked patches $\\mathcal{M}^{(1)}$ and $\\mathcal{M}^{(2)}$ to yield $O^{(1)} = f_{\\mathcal{D}}^{o}(z^{(1)};\\mathcal{M}^{(1)})$ and $O^{(2)} = f_{\\mathcal{D}}^{t}(z^{(2)};\\mathcal{M}^{(2)})$ , respectively. In this paper, we employ a single-block design for the decoder, with the decoder feature projection dimension set to 512. Detailed ablation analysis can be found in Sec. 4.2, and Tables 3b and 3c.", + "bbox": [ + 75, + 609, + 472, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Consistency. In line with the majority of MAE-based methods, we compute the Mean Squared Error (MSE) between the reconstituted and original images to assess the reconstruction loss. We elect to aim for each patch", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of normalized pixels, executing the loss calculation solely for masked patches. As a result, we synchronize the values of normalized pixels within each image patch, applying the loss function exclusively to those patches subject to masking. Significantly, our VideoMAC yields two frames of reconstructed results from both the online and target encoders, necessitating individual loss computations:", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\frac {1}{| \\mathcal {M} ^ {(1)} |} \\sum_ {i \\in \\mathcal {M} ^ {(1)}} \\| \\hat {\\boldsymbol {I}} _ {i} ^ {(1)} - \\boldsymbol {O} _ {i} ^ {(1)} \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 207, + 893, + 253 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\frac {1}{| \\mathcal {M} ^ {(2)} |} \\sum_ {i \\in \\mathcal {M} ^ {(2)}} | | \\hat {\\boldsymbol {I}} _ {i} ^ {(2)} - \\boldsymbol {O} _ {i} ^ {(2)} | | _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 248, + 823, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_o$ and $\\mathcal{L}_t$ denote online and target losses correspondingly. Due to the correlation between consecutive video frames, we pursue temporal consistency in addition to the reconstruction loss. Here, we introduce an intuitive assumption: if the pixel deviation between the two raw frames tends to a particular value, denoted as $\\epsilon$ , it is inferred that the divergence between the reconstructed frames should also trend towards $\\epsilon$ . Based on this assumption, we can estimate the reconstruction consistency between two frames and use it as a proxy for temporal consistency. Hence, the reconstruction consistency loss $\\mathcal{L}_c$ is thus expressed as:", + "bbox": [ + 496, + 295, + 893, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{| \\mathcal {M} ^ {(1)} |} \\sum_ {i \\in \\mathcal {M} ^ {(1)}, j \\in \\mathcal {M} ^ {(2)}} \\| \\boldsymbol {O} _ {i} ^ {(1)} - \\boldsymbol {O} _ {j} ^ {(2)} \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 472, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|\\mathcal{M}^{(1)}|$ can be substituted by $|\\mathcal{M}^{(2)}|$ , courtesy of our utilization of symmetric masking. Ultimately, our approach culminates in a total loss expressed as:", + "bbox": [ + 496, + 523, + 893, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {o} + \\mathcal {L} _ {t} + \\gamma \\mathcal {L} _ {c}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 582, + 890, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\gamma$ symbolizing the weight factor of the consistency loss (see Table 4e and Fig. 7 for detailed analysis). Accordingly, our approach, VideoMAC, accomplishes the realization of a fully convolutional network MVM. Moreover, the reconstruction consistency loss introduced in this paper provides a promising solution for modeling spatio-temporal data.", + "bbox": [ + 496, + 609, + 893, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Implementation", + "text_level": 1, + "bbox": [ + 500, + 710, + 660, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. Our VideoMAC possesses the adaptability to employ any convolutional network the an encoder. To permit a fair comparison, we choose two representative convolutional networks for our experiments: ResNet (RN) [21] and ConvNeXtv2 (CNXv2) [64]. Although RN50 (26M) and CNXv2-S (50M) serve as the primary experimental models, we further investigate models of diverse sizes (Tables 2, 3a and 5). In this paper, the decoder component involves a readily available ConvNeXt block [36], thereby dispensing with the necessity for intricate network design. During pre-training, we employ an encoder implemented", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22736", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/09fd96e655a074a1b80e1801d0ca34c30a248ae590aa3e900edcad14b932122d.jpg", + "table_caption": [], + "table_footnote": [ + "SL: Supervised Learning. CL: Contrastive Learning. IN1K: ImageNet1K [11]. K400: Kinetics-400 [27]. YT18: YouTube-VOS 2018 [68]. Table 1. Comparing VideoMAC with previous supervised and self-supervised approaches (SL, CL, MIM, and MVM) on downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]." + ], + "table_body": "
MethodTypeBackboneTraining datasetEpochDAVIS17 [46]VIP [73] mIoUJHMDB [26] PCK@0.1PCK@0.2
J&FmJmFm
Supervised [21]CVPR16SLRN50IN1K (1.28M, -)10066.063.768.439.559.278.3
DenseCL [63]CVPR21CLRN50IN1K (1.28M, -)20061.159.862.632.556.178.0
DINO [6]ICCV21CLViT-S/16IN1K (1.28M, -)80061.860.263.436.245.675.0
ODIN² [24]ECCV22CLRN50IN1K (1.28M, -)100054.154.353.9---
CrOC [52]CVPR23CLViT-S/16IN1K (1.28M, -)30044.743.545.926.1--
MAE [23]CVPR22MIMViT-B/16IN1K (1.28M, -)160053.552.155.028.144.673.4
FCMAE [64]CVPR23MIMCNXv2-BIN1K (1.28M, -)160043.741.945.524.951.774.2
RC-MAE [30]ICLR23MIMViT-S/16IN1K (1.28M, -)160049.248.950.529.743.272.3
SparK [54]ICLR23MIMRN50IN1K (1.28M, -)160055.656.354.932.852.975.2
MAE-ST [16]NeurIPS22MVMViT-L/16K400 (-, 800 hours)160054.655.553.633.244.472.5
VideoMAE [56]NeurIPS22MVMViT-S/16K400 (-, 800 hours)160039.339.738.922.341.067.9
DropMAE [65]CVPR23MVMViT-B/16K400 (-, 800 hours)160053.451.855.031.142.369.2
SiamMAE [20]NeurIPS23MVMViT-S/16K400 (-, 800 hours)200062.060.363.737.347.076.1
Ours
VideoMACMVMRN50YT18 (-, 5.58 hours)10067.264.969.543.657.279.2
VideoMACMVMCNXv2-SYT18 (-, 5.58 hours)10068.465.371.440.458.180.0
", + "bbox": [ + 81, + 88, + 893, + 316 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/685467a23b25b3992db8e9c661272b37d73616b1bc913058d4c71433bd301b97.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparisons with isotropic (iso.) and hierarchical (hie.) based MVM methods. ‘†’ denotes our modified ios. version." + ], + "table_body": "
MVM MethodResolution#Param.FLOPsJ&Fm
Iso. ViT Backbone
MAE-ST [16]14 × 14304M61.6G54.6
VideoMAE [56]14 × 1422M4.6G39.3
DropMAE [65]14 × 1487M17.6G53.4
SiamMAE [20]14 × 1422M4.6G62.0
Iso. ConvNet Backbone
VideoMAC (CNXv2-S†)14 × 1422M4.3G53.8
Hie. ConvNet Backbone
VideoMAC (RN18)7 × 712M1.8G64.7
VideoMAC (RN50)7 × 726M4.1G67.2
VideoMAC (CNXv2-T)7 × 729M4.5G67.5
VideoMAC (CNXv2-S)7 × 750M8.7G68.4
", + "bbox": [ + 81, + 363, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the MinkowskiEngine sparse convolution library [9]. It's noteworthy that the sparse convolution layer can be seamlessly converted back to a standard convolution layer in downstream tasks without requiring additional processes.", + "bbox": [ + 75, + 566, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training. Different from previous methods that pretrain on the large-scale video dataset K400 [27], VideoMAC pre-trains on YouTube-VOS [68] for 100 epochs, leading to a significant reduction in pre-training time by over 140 times. During VideoMAC pre-training, the standard input comprises two frames with a spatial size of $224 \\times 224$ pixels. The frame interval for this study is determined to be 1 (Table 4c), implying consecutive frames are chosen. We employ minimal data augmentation, encompassing only random resized cropping and horizontal flipping. A masking ratio of 0.75 is utilized to pre-train our models. We conduct training using an AdamW optimizer [39] with a batch size of 512 and a learning rate of 2e-3. Cosine decay [38] is applied to adjust the learning rate.", + "bbox": [ + 75, + 628, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream Tasks. To gauge the efficacy of VideoMAC at video representation learning, we undertake comprehensive experiments on three video downstream tasks, namely video object segmentation, body part propagation, and hu", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "man pose tracking (Table 1). We also investigate the performance of image recognition through video pre-training. Specifically, our backbone is pre-trained on video data and then fine-tuned on ImageNet1K [11] (Table 5).", + "bbox": [ + 496, + 363, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 436, + 633, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section begins with assessing VideoMAC's performance in three downstream tasks, comparing it to previous state-of-the-art methods. Subsequently, we conduct extensive ablation studies to elucidate the effectiveness of our proposed approach. The final part of this section explores VideoMAC's ability for image recognition, accompanied by a visualization of the reconstruction results.", + "bbox": [ + 496, + 460, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparison with Previous methods", + "text_level": 1, + "bbox": [ + 498, + 575, + 808, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Object Segmentation. We evaluate VideoMAC on the popular DAVIS17 [46] dataset, which includes a validation set comprising 30 videos. Our evaluation methodology aligns with [6, 25], and we report the typical evaluation metrics [45], including the mean region similarity $\\mathcal{I}_m$ , the mean boundary accuracy $\\mathcal{F}_m$ , and their average values $\\mathcal{J} \\& \\mathcal{F}_m$ . As shown in Table 1, our VideoMAC achieves state-of-the-art results for video object segmentation. With a smaller video training dataset and fewer training epochs, both our VideoMAC models (i.e., CNXv2-S and RN50) surpass the SiamMAE [20] by $6.4\\%$ and $5.2\\%$ , respectively, in terms of $\\mathcal{J} \\& \\mathcal{F}_m$ . It's worth noting that our approach achieves superior performance to supervised method [21] based on MVM, with improvements of $2.4\\%$ and $1.4\\%$ , respectively. Owing to the hierarchical feature extraction capabilities inherent in ConvNets, our approach demonstrates superior proficiency in spatial information extraction in contrast to ViT-based MVM methods [16, 20, 56, 65]. Concurrently, the introduced reconstruction consistency framework strengthens the acquisition of temporal correlations.", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22737", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc779ede0ca14ccce373a4492c3fa80742e7aa07fd5fa91a0f0c7998d394cb40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 179, + 130 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1da5c536d33763451793841a7c88469e86b8055691da3e7dfb2556c6b86d2643.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 130, + 179, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59c4dc3fa0f7ce2379bd8ea90f51cc07f7bccf20dd8f78bfa84d403ab57233a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 171, + 179, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/beabd69a3a3bbc4e4e0aa2904e73160f8ed59d97015b78f31f3415062bfa6fc7.jpg", + "image_caption": [ + "Figure 4. Qualitative results of our VideoMAC (using CNXv2-S) for three video downstream tasks: (a) video object segmentation on DAVIS17 [46]), (b) body part propagation on VIP [73], and (c) human pose tracking on JHMDB [26]." + ], + "image_footnote": [], + "bbox": [ + 80, + 210, + 178, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39687ffb3312446beca58fb38acc3a0a9030db6a2d18559414163b847b87400b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 88, + 279, + 130 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7af476a6a50434a352f3b2d59aa13f561965a7386e62ddafbedcbba801bf9c9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 130, + 279, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e40adcfdd3b59ddcb68fa2a1882d7ff108a65a131d56af2ff1646273560adf21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 171, + 279, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cba9e282e25f99cc7c37a6577d419a4b5c039f717feca5d0c5e016557bf3cc8a.jpg", + "image_caption": [ + "(a) Video Object Segmentation" + ], + "image_footnote": [], + "bbox": [ + 179, + 210, + 279, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ee736ea6f99104038b48ab1192b6046d3869fb26b322f0319cb12e5663480b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 88, + 380, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/41c047bd1e556a3366499c82d0d823ed48cb3e8f909efa76b3aa6fa898c27c87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 130, + 380, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e81fb94ed7e1a937521e26acca5cac8e9e95f110629865cc1a8f7471a84a227.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 171, + 380, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/672ce70fa1eb31ffc003eed41139d947c52a9b9b664e3edd3f0edd160e4a0dce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 210, + 380, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dc185c900fa52fcb7ce480ce0de5a78e34871720679ab4b29620baff721392ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 88, + 482, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b7dc9ba9f5052d1bf22ad81136461583dba047d56476410f4f99e36c05c2d11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 130, + 482, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/030178ddea3e3964ceddaf96234691394f8c0713a51607b50a677aae2d476b2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 171, + 482, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8abbb16234408df96cac0a82d55e1bac421d74db6f0d2c5062325f7f67387ca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 210, + 482, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ee3b50f026dfec11e4fd7f62a93fc6090ad68689053047d89c3d4e7398967fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 586, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0dcefeb0c94cb09100575611095050d64ae9e5b73ce8a9003c3e09e289b98367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 128, + 586, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/780f8b318bd819a397bb8f5bf4f0928acc4a7318b50580525c4fc136bb72ec04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 88, + 687, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffe6d6cc59693db6108be3e0bb934a428b0dbc1c10dd1c0a5b74b6ead7e06acb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 128, + 687, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/45f1fbf905528e79e1f2dec0ed36d682bbf52dd7a929c2e1da31d677246a1af6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 88, + 789, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6aad074add475eff0ac939641540b24868a7327dbea84e0cbac933dd1b733868.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 128, + 787, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7bedc0740b8045a63ee17f8a8802aa38b82d457d5da3c6357e11d03e5c7685fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 88, + 890, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fd26f5d405364568ed7ece610516e363157e5817e212676eaa9352486beeee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 126, + 890, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc4fa73f645e81f89611fb321b357b3d7cb19070b3e5d559bab70fc658d8f69d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 176, + 586, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95eeaf3f6df19bf8954565979705f4b938ff7ef3fc51d8eb7806f391f0a33c99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 214, + 586, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2f107ce900efc6209ee6e07d5b9769078dee1ca0eeff81d739e4655103a1989.jpg", + "image_caption": [ + "(b) Body Part Propagation" + ], + "image_footnote": [], + "bbox": [ + 589, + 176, + 687, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71686b941ede7f1925c26f130ac37b92fcc0be581bf05c1a3cfe8f5c0b4e5344.jpg", + "image_caption": [ + "(c) Human Pose Tracking" + ], + "image_footnote": [], + "bbox": [ + 589, + 214, + 687, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3197cc538b71bd80bc20001bf8ab42572a4c5f4aa7d11496527579b7b688ff12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 176, + 789, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2074be0313e82ba1108d27f74e24085bf7ade27605fcd5b5dfd6f8cb1a23f5a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 214, + 787, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59ab462eeac5542ee870142213ceb9290a1c1f62a493b38391f06b6653a8d2df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 176, + 890, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb2d86d842d5a702a28d3e4eddf353a49c36ff68b25283b862123797373b8f74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 214, + 890, + 251 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6fe44acf70f5e15e84497a10d2738650e50582b1e9fc52ac61c9e965eefc6f7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneJ&FmJmFm
R1864.762.866.5
R5067.264.969.5
CNXv2-T67.564.470.6
CNXv2-S68.465.371.4
", + "bbox": [ + 94, + 306, + 328, + 381 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6d122d00ab9f69d10e65bd5c596e9bcc32302b5a500a5890c78160864799b739.jpg", + "table_caption": [ + "(a) Encoder type. VideoMAC is compatible with both classical and modern ConvNets." + ], + "table_footnote": [], + "table_body": "
Dim.J&FmJmFm
12867.964.871.0
25668.265.371.1
51268.465.371.4
102468.165.071.2
", + "bbox": [ + 94, + 407, + 328, + 483 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4252ccd3096b72799006f999bbe6279ead357fcd8452ee7a70de1e1c84f9072.jpg", + "table_caption": [ + "(c) Decoder width. A decoder width of 512 delivers the best performance." + ], + "table_footnote": [], + "table_body": "
BlocksJ&FmJmFm
168.465.371.4
268.265.271.2
468.365.471.2
868.165.271.0
", + "bbox": [ + 357, + 306, + 591, + 381 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e0b092ae55a01f8ce68eee868aed1b2a7a92632e88555317a461916d8be95ea.jpg", + "table_caption": [ + "(b) Decoder depth. One block achieves competitive performance and efficiency." + ], + "table_footnote": [], + "table_body": "
MaskingJ&FmJmFm
Asym.66.563.469.6
Sym.68.465.371.4
", + "bbox": [ + 357, + 407, + 591, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(d) Masking design. Symmetric (sym.) masking outperforms asymmetric (asym.) one.", + "bbox": [ + 348, + 484, + 593, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af566f9822535df270ac0e7184972c9b55ad6e8e0256d024df65fbd315924569.jpg", + "image_caption": [ + "Figure 5. For masked modeling, dense convolution usually results in the dissipation of mask structures. The deployment of sparse convolution proves to be an effective solution, enabling ConvNet-based MIM / MVM." + ], + "image_footnote": [], + "bbox": [ + 622, + 305, + 887, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. VideoMAC encoder and decoder ablation experiments on DAVIS17 [46]. We report ablation results for two of the most emblematic families of ConvNets. Unless specified otherwise, the encoder is CNXv2-S in default. 'Dim' indicates the number of decoder dimensions. Default settings employed in this paper are marked in blue.", + "bbox": [ + 76, + 511, + 602, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Body Part Propagation. We evaluate VideoMAC on the Video Instance Parsing (VIP) dataset [73], and the quantitative results, measured by mIoU, are detailed in Table 1. This validation set comprises of 50 videos that concentrate on the propagation of 19 body parts, such as arms and legs, thereby necessitating a heightened level of precision in matching as compared to video object segmentation. We adhere to the same parameters as [25], adjusting the video frames to $560 \\times 560$ dimensions. In the case of body part propagation, two variants of our VideoMAC demonstrate improvements by $3.1\\%$ (CNXv2-S) and $6.3\\%$ (RN50), respectively, in comparison with the top-performing SiamMAE [20] in MVM. Significantly, our RN50-based model outperforms CNXv2-S on the VIP dataset [73] due to the higher dimensionality of RN50 features (2048 vs. 768), which benefits complex tasks and improves accuracy in mask propagation.", + "bbox": [ + 75, + 582, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human Pose Tracking. We conduct the human pose tracking task on the validation set of JHMDB [26], which includes 268 videos and entails detecting 15 human keypoints. Conforming to the protocol delineated in [25], we resize video frames to a resolution of $320 \\times 320$ , and employ", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Possibility of Correct Keypoints (PCK) metric [51] for quantitative evaluation. As shown in Table 1, both variants (i.e., RN50 and CNXv2-S) of our VideoMAC outperforms all MVM methods in terms of PCK@0.1 and PCK@0.2.", + "bbox": [ + 496, + 582, + 890, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. As depicted in Fig. 4, we present the qualitative results of mask propagation achieved by VideoMAC on three downstream tasks. Our VideoMAC yields notable performance across a spectrum of complex videos.", + "bbox": [ + 496, + 642, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 710, + 663, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct extensive ablation studies to analyze the component design of our VideoMAC.", + "bbox": [ + 496, + 734, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder design: hierarchical vs. isotropic. In this ablation study, we scrutinize diverse encoder architectures to our VideoMAC, i.e., hierarchical and isotropic designs. Precisely, we adopted the ViT [14] design principles and restructured the CNXv2-S backbone from having a hierarchical structure into an isotropic one. This implies the creation of an architecture devoid of downsampling layers, ensuring consistent feature resolution (e.g., $14 \\times 14$ ) throughout all depths. As shown in Table 2, the performance dif", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22738", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac213057ede8f573f92fd865c69a6d724721510ad05b28650ffe88deb498b601.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSizeJ&FmJmFm
YT18 [68]5.58 hours68.465.371.4
K4001 [27]800 hours67.865.270.3
", + "bbox": [ + 88, + 88, + 403, + 133 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3fe293c003bebd83de49d5ed613a8387899fa6e9d2492f21f26ca85b054a992.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AugmentationJ&FmJmFm
spatial68.465.371.4
color64.961.868.0
spatial + color66.162.969.3
", + "bbox": [ + 419, + 88, + 676, + 146 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) Data augmentation. Minimal data augmentation achieves the best performance.", + "bbox": [ + 410, + 148, + 676, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74aefb25d9cb38918231068ba717a7b87cd5073f02eace53330a36052dd27b57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GapJ&FmJmFm
168.465.371.4
566.463.569.2
1063.260.465.9
", + "bbox": [ + 692, + 88, + 888, + 146 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(c) Frame gap. Video pairs with one frame gap perform the best results.", + "bbox": [ + 681, + 148, + 890, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d8ddfb775507dfe71f5f0f3e9b5b236263782db486cabeef8ff99061232d765.jpg", + "table_caption": [ + "(a) Training Dataset. Our VideoMAC pre-trained on YT18 yields better performance and efficiency compared to K400." + ], + "table_footnote": [], + "table_body": "
\\( {\\mathcal{L}}_{c} \\)\\( {\\mathcal{L}}_{t} \\)\\( \\mathcal{J}\\&{\\mathcal{F}}_{m} \\)\\( {\\mathcal{J}}_{m} \\)\\( {\\mathcal{F}}_{m} \\)
✘✘51.152.250.0
✘✓60.656.364.9
✓✘63.463.468.7
✓✓68.465.371.4
", + "bbox": [ + 89, + 175, + 333, + 248 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ef0bdd05e0688e52be411d0c71be1d20557547519b17fa609c0c2c490b9b64b.jpg", + "table_caption": [ + "(d) Model loss. VideoMAC utilizes consistency loss $\\mathcal{L}_c$ and target loss $\\mathcal{L}_t$ for the best." + ], + "table_footnote": [], + "table_body": "
ConsistencyJ&FmJmFm
γ=0.060.656.364.9
γ=0.164.862.067.6
γ=0.565.863.268.5
γ=1.068.465.371.4
", + "bbox": [ + 359, + 175, + 602, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/656416b244da2aecf2459b140fc78d96d65e55edf3197c0ca47481c94dd1a8ce.jpg", + "table_caption": [ + "(e) Reconstruction consistency. When $\\gamma$ is set to 1.0, the consistency loss works the best." + ], + "table_footnote": [], + "table_body": "
Masking ratioJ&FmJmFm
0.6566.563.869.1
0.7568.465.371.4
0.8568.365.371.3
0.9560.359.760.9
", + "bbox": [ + 627, + 175, + 887, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(f) Masking ratio. Our VideoMAC uses a masking ratio of 0.75 for the highest performance.", + "bbox": [ + 617, + 250, + 888, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. VideoMAC ablations on data, loss, and masking ratio configuration. We report two official metrics ${\\mathcal{I}}_{m}$ and ${\\mathcal{F}}_{m}$ on DAVIS17 [46].", + "bbox": [ + 76, + 276, + 890, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b27b70dea0c2e01c788fb99b392cc02eb752a1699c3ea50d8dd8e813dc78a963.jpg", + "image_caption": [ + "Figure 6. Analysis of individual loss components after Video-MAC pre-training 100 epochs on YouTube-VOS 2018 [68]." + ], + "image_footnote": [], + "bbox": [ + 76, + 299, + 271, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cd9e0c1d30254190a9cb34d2128c6d67750472d7b0cb8763719acc7dc19a00b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 299, + 468, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ference between our VideoMAC (iso.) and the ViT-based MVM methods (e.g., VideoMAE [56] and DropMAE [65]) is marginal. By maintaining the hierarchical structure of ConvNets within VideoMAC models (hie.), we improve the performance significantly, even though their feature resolution (i.e., $7 \\times 7$ ) is merely half that of the ViT-based models [16, 20, 56, 65]. This can be ascribed to ConvNets' capacity to preserve multi-scale spatial information.", + "bbox": [ + 75, + 455, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Encoder backbone: classical and modern. For our study, we incorporate two emblematic ConvNet families, i.e., ResNets [21] and ConvNeXts [36, 64]. As shown in Table 3a, we substitute dense convolution with sparse convolution in the baseline ConvNets. This adaptation seamlessly dovetails into our VideoMAC framework, thereby facilitating a ConvNet-based MVM. In addition, Fig. 5 highlights the influence of employing dense and sparse convolution on masks. It is apparent that sparse convolution effectively safeguards the structural integrity of the mask, thus enabling the integration of MAE pre-training within ConvNets.", + "bbox": [ + 75, + 579, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoder design: blocks and dimensions. Adhering to a simplistic design principle, our VideoMAC adopts the decoder from the purely convolutional network structure deployed in ConvNeXt [36]. We conduct ablation experiments focusing on the depth and width of the decoder, with the results detailed in Tables 3b and 3c. The decoder exhibits competitive results with a single block and 512 dimensions, striking a trade-off between performance and efficiency.", + "bbox": [ + 75, + 747, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Masking design: asymmetric vs. symmetric. In Table 3d, we investigate the impact of two masking strategies, namely", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb391479d6808a393acae730f483edd47360905efa464d34859cb3bcf326c937.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 299, + 694, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb4cc285ae593a038d53d28330309bfbf1163c8bf56f7226a961c3037f80f4d8.jpg", + "image_caption": [ + "Figure 7. Consistency loss in relation to $\\gamma$ (left: identical encoder, CNXv2-S) and encoders (right: identical weight factor, $\\gamma = 1.0$ )." + ], + "image_footnote": [], + "bbox": [ + 697, + 299, + 890, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "asymmetric (same masking ratio but different locations) and symmetric (same masking ratio and locations), on performance. Compared to asymmetric masking, symmetric masking simplifies the calculation of the reconstruction loss, as it covers the entire masked area. Moreover, it facilitates the inclusion of more reconstruction regions in the loss calculation, resulting in faster model convergence.", + "bbox": [ + 496, + 454, + 890, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data setting: dataset, augmentation, and gap. To assess the influence of different video datasets, Table 4a presents the results of pre-training our VideoMAC on both YT18 [68] and $\\mathrm{K400^1}$ [27]. The ablation results of VideoMAC on the two datasets differ by just $0.6\\%$ , demonstrating the robustness and adaptability of our method to different datasets. For data augmentation, as shown in Table 4b, our VideoMAC achieves the best results using only spatial strategy, i.e., random resized cropping and horizontal flipping. Additionally, we conduct an ablation study on the frame gap, and the optimal results are obtained with consecutive frames (i.e., the gap is 1), as indicated in Table 4c. The contribution of $\\mathcal{L}_c$ and $\\mathcal{L}_t$ . Previous MVM methods predominantly rely on a single loss, i.e., the reconstruction loss, necessitating multiple frames to be simultaneously processed by the model. In contrast, our study utilizes a more computation-friendly strategy by inputting frame pairs into two separate encoders, followed by parameter updates using EMA. Consequently, we introduce two distinct loss terms: online $\\mathcal{L}_o$ and target $\\mathcal{L}_t$ reconstruction losses. Furthermore, to integrate temporal information, we", + "bbox": [ + 496, + 561, + 892, + 877 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{1}$ K400 is sampled to match YT18's training data volume.", + "bbox": [ + 517, + 886, + 820, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22739", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/393a600959fe9b404075ae11d41b624fb408fcf8dd9c9cd2d9a889f058a3f415.jpg", + "image_caption": [ + "Figure 8. Visualization of reconstruction examples using a pre-trained VideoMAC (CNXv2-S) model with 0.75 masking ratio. These frames are randomly selected from the validation set of YT18 [68] and masked $60\\%$ of patches." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 888, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compute the loss between the reconstructed results of frame pairs, colloquially known as the reconstruction consistency loss $\\mathcal{L}_c$ . As shown in Table 4d, $\\mathcal{L}_t$ and $\\mathcal{L}_c$ improve $9.5\\%$ and $12.3\\%$ in terms of $\\mathcal{J} \\& \\mathcal{F}_m$ , respectively, with the full VideoMAC model yielding the best results. The diverse loss components during pre-training are illustrated in Fig. 6.", + "bbox": [ + 75, + 239, + 467, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of weight factor $\\gamma$ . To investigate the impact of the weight factor on reconstruction consistency loss $\\mathcal{L}_c$ , we conduct experiments with different factor settings, i.e., $\\gamma \\in \\{0.0, 0.1, 0.5, 1.0\\}$ . As illustrated in Table 4e, superior performance is attained when $\\gamma$ approximates to 1. When $\\gamma$ equals 0, which signifies the absence of the reconstruction consistency loss term, a detrimental effect on performance is noticeable. Simultaneously, we record $\\mathcal{L}_c$ for various weight factors during pre-training, as depicted in Fig. 7. It becomes evident that an elevated value of $\\gamma$ results in a lesser loss. Ultimately, both the R50 and CNXv2-S models converge to comparable final losses. It is noteworthy that $\\mathcal{L}_c$ presents an initial decrease, succeeded by an increase, and ultimately reaches convergence, which aligns with the description in Eq. (2), implying that the ultimate reconstruction results between two frames converge to $\\epsilon$ .", + "bbox": [ + 75, + 330, + 467, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of masking ratio. An ablation analysis of the masking ratio is conducted, and the results for our Video-MAC models (CNXv2-S) are documented in Table 4f. Ablation results portray a performance pattern that first ascends and subsequently descends with increasing masking ratios. Noticeably, the performance of our approach experiences a drastic decline at a 0.95 masking ratio caused by overfitting to the self-supervised pretext task.", + "bbox": [ + 75, + 574, + 467, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Additional Experiments", + "text_level": 1, + "bbox": [ + 76, + 709, + 297, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VideoMAC for Image Recognition. As shown in Table 5, utilizing CNXv1-T [36] / CNXv2-T [64] pre-trained by our VideoMAC on YT18 [68], we achieve $82.3\\%$ / $82.9\\%$ accuracy on IN1K [11] image classification. In comparison to CNXv1-T and CNXv2-T trained from scratch, our approach improves the image recognition results by $0.2\\%$ and $0.4\\%$ . Remarkably, the performance garnered by VideoMAC (MVM) is proximate to or commensurate with the state-of-the-art ConvNet-based MIM methods [54, 64]. Given the pre-training cost (100 epochs vs. 1600 epochs) and the domain gap (video vs. image), we regard this im", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d8f9dcb3617c5b4273b415dc016acd3e1f77e0090418bafbd5f3f6dd3b3e0eb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePT epochPT dataFT acc.
Supervised [36]CNXv1-T--82.1
FCMAE [64]CNXv1-T1600IN1K82.3 (+0.2)
Spark [54]CNXv1-T1600IN1K82.4 (+0.3)
VideoMACCNXv1-T100YT1882.3 (+0.2)
Supervised [64]CNXv2-T--82.5
FCMAE [64]CNXv2-T1600IN1K83.0 (+0.5)
VideoMACCNXv2-T100YT1882.9 (+0.4)
", + "bbox": [ + 504, + 243, + 890, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparing VideoMAC with previous ConvNet-based MIM approaches (e.g., Spark [54] and FCMAE [64]) on the image downstream task, i.e., fine-tuning recognition accuracy (acc.) on IN1K [11]. 'PT' and 'FT' indicate pre-training and fine-tuning.", + "bbox": [ + 498, + 351, + 890, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provement of our VideoMAC as noteworthy and promising compared to training from scratch and pre-training by MIM. Visualization. We visualize reconstruction results from the YT18 [68] validation set to analyze performance in MVM pre-training. In Fig. 8, VideoMAC makes reasonable predictions about the masked regions, e.g., color and contour.", + "bbox": [ + 498, + 412, + 890, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 520, + 617, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we present VideoMAC, an MVM pre-training framework constructed fully using ConvNets. Our masked modeling for frame pairs is executed by introducing an online network (encoder and decoder) optimized by gradients, as well as the target network that updates parameters via EMA. Concurrently, for reconstruction results derived from both online and target networks, our proposed loss of reconstruction consistency between frame pairs paves the way for temporal modeling. Our VideoMAC empowers a multitude of ConvNets to reap the performance gains for downstream tasks (e.g., video but also image) by MVM pre-training.", + "bbox": [ + 496, + 546, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. Existing ConvNet-based architectures typically use a fixed downsampling ratio, which in turn leads to a relatively fixed patch size. For future work, we will investigate ConvNets with adjustable patch size, allowing for higher feature resolution like ViT.", + "bbox": [ + 496, + 713, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No. 62102182, 62202227, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20220934, BK20220938, BK20220936), China Postdoctoral Science Foundation (No. 2022M721626, 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No.30923010303).", + "bbox": [ + 496, + 789, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22740", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022. 1", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 2", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1, 2", + "[4] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 2", + "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912–9924, 2020. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020.", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2", + "[9] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 5", + "[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 1, 2", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 5, 8", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, , and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 1, 2", + "[13] Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language understanding and generation. In NeurIPS, 2019. 2", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 6", + "[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 1", + "[16] Christoph Feichtenhofer, haoqi fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, pages 35946-35958, 2022. 1, 2, 3, 4, 5, 7", + "[17] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. NeurIPS, 35:35632-35644, 2022. 2", + "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 2", + "[19] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. In ICLR, 2023. 2, 3", + "[20] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 4, 5, 7", + "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5", + "[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In ECCV, pages 123-143, 2022. 5", + "[25] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, pages 19545-19560, 2020. 2, 5, 6", + "[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, pages 3192-3199, 2013. 1, 2, 5, 6", + "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5, 7", + "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2", + "[29] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2", + "[30] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. In ICLR, 2023. 1, 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22741", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Lulei Li, Tianfei Zhou, Wenguan Wang, Lu Yang, Jianwu Li, and Yi Yang. Locality-aware inter-and intra-video reconstruction for self-supervised correspondence learning. In CVPR, pages 8719-8730, 2022. 2", + "[32] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In CVPR, pages 18706-18716, 2023. 2", + "[33] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In CVPR, pages 6252-6261, 2023. 4", + "[34] Yuan Liu, Songyang Zhang, Jiacheng Chen, Zhaohui Yu, Kai Chen, and Dahua Lin. Improving pixel-based mim by reducing wasted modeling capability. In ICCV, pages 5361-5372, 2023. 1", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3", + "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 2, 3, 4, 7, 8", + "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3", + "[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5", + "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5", + "[40] Xiankai Lu, Wenguan Wang, Martin Danelljan, Tianfei Zhou, Jianbing Shen, and Luc Van Gool. Video object segmentation with episodic graph memory networks. In ECCV, pages 661-679, 2020. 2", + "[41] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In CVPR, pages 8960-8970, 2020.", + "[42] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, pages 596-613, 2022.", + "[43] Gensheng Pei, Fumin Shen, Yazhou Yao, Tao Chen, Xian-Sheng Hua, and Heng-Tao Shen. Hierarchical graph pattern understanding for zero-shot video object segmentation. TIP, 32:5909-5920, 2023.", + "[44] Gensheng Pei, Yazhou Yao, Fumin Shen, Dan Huang, Xingguo Huang, and Heng-Tao Shen. Hierarchical co-attention propagation network for zero-shot video object segmentation. TIP, pages 2348–2359, 2023. 2", + "[45] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5", + "[46] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2", + "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Malik, Yanghao Li, and Christoph Feichtenhofer. Hiera: A hierarchical vision transformer without the bells-and-whistles. ICML, 2023. 2", + "[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 2", + "[51] Jie Song, Limin Wang, Luc Van Gool, and Otmar Hilliges. Thin-slicing network: A deep structured model for pose estimation in videos. In CVPR, pages 4220-4229, 2017. 6", + "[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In CVPR, pages 7000–7009, 2023. 5", + "[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 2", + "[54] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. In ICLR, 2023. 2, 3, 4, 5, 8", + "[55] Yunjie Tian, Lingxi Xie, Zhaozhi Wang, Longhui Wei, Xiaopeng Zhang, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Integrally pre-trained transformer pyramid networks. In CVPR, pages 18610-18620, 2023. 4", + "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 1, 2, 3, 4, 5, 7", + "[57] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 2", + "[58] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2", + "[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2", + "[60] Wenguan Wang, Xiankai Lu, Jianbing Shen, David J Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, pages 9236-9245, 2019. 2", + "[61] Wenguan Wang, Hongmei Song, Shuyang Zhao, Jianbing Shen, Sanyuan Zhao, Steven CH Hoi, and Haibin Ling." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22742", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning unsupervised video object segmentation through visual attention. In CVPR, pages 3064-3074, 2019. 2", + "[62] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 2, 3", + "[63] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 5", + "[64] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In CVPR, pages 16133-16142, 2023. 1, 2, 3, 4, 5, 7, 8", + "[65] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In CVPR, pages 14561–14571, 2023. 1, 2, 3, 5, 7", + "[66] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, pages 13879-13889, 2023. 2", + "[67] Z. Xie, Z. Zhang, Y. Cao, Y. Lin, J. Bao, and Z. Yao. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 1, 2", + "[68] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, pages 585-601, 2018. 5, 7, 8", + "[69] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, pages 2623-2632, 2021. 2", + "[70] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 2, 3", + "[71] Yurong Zhang, Lielei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In CVPR, pages 2246-2256, 2023. 2", + "[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 2", + "[73] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In ACMMM, pages 1527-1535, 2018. 1, 2, 5, 6", + "[74] Tianfei Zhou, Fatih Porikli, David J Crandall, Luc Van Gool, and Wenguan Wang. A survey on deep learning technique for video segmentation. TPAMI, 45(6):7099-7122, 2023. 2", + "[75] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, pages 408-417, 2017. 2" + ], + "bbox": [ + 78, + 92, + 467, + 869 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22743", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_model.json b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b3fdfd23568517a3d7dce2f9bbbb8fdfad43ac4 --- /dev/null +++ b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_model.json @@ -0,0 +1,3005 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.776, + 0.151 + ], + "angle": 0, + "content": "VideoMAC: Video Masked Autoencoders Meet ConvNets" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.181, + 0.794, + 0.235 + ], + "angle": 0, + "content": "Gensheng Pei\\*, Tao Chen\\*, Xiruo Jiang, Huafeng Liu, Zeren Sun,\\* Yazhou Yao\\* \nSchool of Computer Science and Engineering, Nanjing University of Science and Technology \nhttps://github.com/NUST-Machine-Intelligence-Laboratory/VideoMAC" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.681 + ], + "angle": 0, + "content": "Recently, the advancement of self-supervised learning techniques, like masked autoencoders (MAE), has greatly influenced visual representation learning for images and videos. Nevertheless, it is worth noting that the predominant approaches in existing masked image / video modeling rely excessively on resource-intensive vision transformers (ViTs) as the feature encoder. In this paper, we propose a new approach termed as VideoMAC, which combines video masked autoencoders with resource-friendly ConvNets. Specifically, VideoMAC employs symmetric masking on randomly sampled pairs of video frames. To prevent the issue of mask pattern dissipation, we utilize ConvNets which are implemented with sparse convolutional operators as encoders. Simultaneously, we present a simple yet effective masked video modeling (MVM) approach, a dual encoder architecture comprising an online encoder and an exponential moving average target encoder, aimed to facilitate inter-frame reconstruction consistency in videos. Additionally, we demonstrate that VideoMAC, empowering classical (ResNet) / modern (ConvNeXt) convolutional encoders to harness the benefits of MVM, outperforms ViT-based approaches on downstream tasks, including video object segmentation (+5.2% / 6.4% J&F), body part propagation (+6.3% / 3.1% mIoU), and human pose tracking (+10.2% / 11.1% PCK@0.1)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Spurred by the success of masked language modeling in natural language processing [3, 10, 12], transformer-based visual models, notably vision transformers (ViT) [14] adopt masked image modeling (MIM) are attracting attention [1, 2, 15, 23, 34, 67]. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.449, + 0.895, + 0.59 + ], + "angle": 0, + "content": "Figure 1. State-of-the-art MAE architectures (e.g., MAE-ST [16] and SiamMAE [20]) in masked video modeling commonly employ ViT-based encoders. We propose VideoMAC, a new video masked autoencoder built using pure ConvNets. In this study, we model VideoMAC with two of the most emblematic families of ConvNets, namely ConvNeXtv2 [64] and ResNet [21]. Notably, VideoMAC exhibits superior performance on a range of downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26], compared to ViT-based methods [16, 20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.748 + ], + "angle": 0, + "content": "With the successful application of MIM, some pioneering works [16, 56] have shifted the focus of video representation learning to masked video modeling (MVM), yielding promising results in downstream video tasks. However, features learned by MVM are designed explicitly for pixel reconstruction tasks. They perform well when fine-tuned for downstream video tasks but exhibit subpar performance without specific fine-tuning [20]. Their performance is even inferior to MIM methods [23, 30], and it lags significantly behind supervised methods [21, 36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.896, + 0.903 + ], + "angle": 0, + "content": "We strive to discern the factors contributing to this issue. To the best of our knowledge, most current MVM methods are based on the isotropic design of ViT [16, 20, 56, 65]. The reason is that early MIM / MVM methods are explicitly developed for ViT, as they allow the use of only visible patches as the sole input to the encoder, which aligns well with the ViT architecture. For instance, the representative approach, masked autoencoders (MAE) [23], enhances the efficiency and effectiveness of model pre-training by focusing predominantly on the unmasked regions. However," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22733" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.127, + 0.11, + 0.14 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.089, + 0.171, + 0.099 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.099, + 0.196, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.089, + 0.263, + 0.098 + ], + "angle": 0, + "content": "Masked" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.101, + 0.282, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.089, + 0.365, + 0.099 + ], + "angle": 0, + "content": "Reconstructed" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.101, + 0.367, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.089, + 0.435, + 0.099 + ], + "angle": 0, + "content": "Heatmap" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.101, + 0.453, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.193, + 0.11, + 0.204 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.167, + 0.196, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.167, + 0.282, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.167, + 0.367, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.167, + 0.453, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Visualization of the heatmap during the reconstruction of masked patches in a random frame pair. It is evident that our approach highlights similar regions for both (a) past and (b) current frames, proficiently reconstructing colors and contours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.468, + 0.522 + ], + "angle": 0, + "content": "previous MVM approaches [16, 56], confined by the plain vision transformer design devoid of a hierarchical structure tend to result in inadequate spatial modeling capabilities for local features [17, 35, 49, 62, 70]. Such capabilities are crucial for dense downstream tasks like segmentation and detection. Moreover, the uniform reliance on ViT in existing MVM methods has resulted in significant computational resource consumption [16, 56, 58]. One might first consider ConvNets [21, 28, 29, 36, 50, 53], characterized by their low resource consumption and built-in hierarchical structure, as an ideal solution to tackle the aforementioned issues. Nevertheless, ConvNets appears understudied in MVM. Consequently, it raises the question: can ConvNets, endowed with a pyramid vision architecture, guide MVM towards the era of hierarchical pre-training?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To address this concern, we propose a simple yet potent MVM approach, referred to as Video Masked Autoencoders with ConvNets (VideoMAC). Uniquely, our method represents an initial endeavor to supplement video pre-training by integrating hierarchical ConvNets. However, applying ConvNets to MVM proves more complex than utilizing ViTs. Typical ConvNets might be incompatible with MVM due to the employment of dense sliding windows with overlap, resulting in information leakage within masked regions. To efficiently incorporate mask modeling into ConvNets, we opt for sparse convolution over dense convolution to restructure the original ConvNets, an approach validated in recent studies [54, 64] to effectively maintain the structure of the masked region when utilizing ConvNets. Another hurdle for MVM is the effective incorporation of temporal data. Current MVM pre-training methods [16, 56] involving multi-frame inputs are primarily used in downstream resource-intensive tasks such as video action recognition. Although methods using two-frame inputs for pre-training lower resource usage [20, 65], they typically employ the Siamese architecture, necessitating robust skills for designing reconstruction targets. Importantly, all aforementioned methods utilize computationally intensive, non-hierarchical ViT [14] as their encoders. To this end, we utilize pairs of frames as inputs and introduce the reconstruction consis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "tency loss, a streamlined extension of MAE to accommodate spatio-temporal data. We formulate online networks through gradient optimization and establish target networks by updating parameters via an exponential moving average (EMA). This allows for individual recovery of masked frame pairs and computation of the reconstruction consistency loss between them. As illustrated in Fig. 2, our approach can focus on similar regions between two frames and reconstruct the masked portions efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Notably, our approach transcends previous state-of-the-art techniques in three downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]. Moreover, we present promising results utilizing VideoMAC in fine-tuning for image recognition on ImageNet1K [11]. To summarize, our main contributions of this work include:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.334, + 0.892, + 0.378 + ], + "angle": 0, + "content": "- We propose a simple yet potent solution of masked video autoencoders, marking the initial endeavor to successfully implement MVM using pure ConvNets." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.379, + 0.891, + 0.409 + ], + "angle": 0, + "content": "- We introduce the reconstruction consistency loss to enable elegant modeling of spatio-temporal data." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.41, + 0.891, + 0.438 + ], + "angle": 0, + "content": "- Our approach significantly outperforms existing ViT-based MVM methods in various downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.334, + 0.892, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.514 + ], + "angle": 0, + "content": "While the focus of the computer vision community is increasingly shifting towards vision transformers, we trust that VideoMAC will rekindle interest in exploring ConvNet-based MVM, thereby catalyzing the discovery of new potentials across a variety of video tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.64, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Masked Image Modeling. Self-supervised learning (SSL) has been rapidly developing in computer vision, demonstrating substantial potential for deriving valuable representations from extensive, unlabeled datasets. A key advantage of SSL [2, 5-8, 18, 22, 48, 57, 69, 72] is its independence from costly human annotations, thereby providing robust prior knowledge across various visual tasks. The success of masked language modeling in NLP [3, 10, 12, 13, 47] has catalyzed a rising interest in masked image modeling (MIM) for visual pre-training. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations [2, 23, 67]. Notably, pioneering works [54, 64] have recently applied ConvNets to MIM pre-training, paving the way for masking modeling methods based on ConvNets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Masked Video Modeling. Recently, researchers have embarked on exploring MVM approaches [16, 19, 65, 66] to comprehend unsupervised video representations, exhibiting effectiveness across a multitude of downstream tasks [4, 25, 31, 32, 40-44, 60, 61, 71, 74, 75]. However, current MVM methodologies [20, 56, 58, 59] primarily depend on the isotropic ViT, a visual backbone compatible with mask" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22734" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "ing modeling. Furthermore, these ViT-based MVM strategies may encounter challenges with inadequate local feature modeling when dealing with dense tasks such as segmentation and detection due to their non-hierarchical structure [35, 37, 62, 70]. To the best of our knowledge, no MVM approaches are explicitly designed for hierarchical network-based architectures, especially ConvNets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.288 + ], + "angle": 0, + "content": "It is worth noting that while the isotropic ViT is well suited for the implementation of MVM methods [16, 19, 56], this is exceptionally difficult for ConvNets. In this paper, our VideoMAC is committed to developing an efficient video pre-training tool that can equip ConvNets with superior, or at least comparable, performance to ViTs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.303, + 0.185, + 0.32 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.468, + 0.464 + ], + "angle": 0, + "content": "Our VideoMAC builds upon MAE [23] and, pioneeringly, facilitates masked video modeling with a purely convolutional architecture. As depicted in Fig. 3, VideoMAC is designed to reconstruct masked patches, which are randomly and symmetrically masked on frame pairs at a high masking ratio (0.75 as used in this study). The framework achieves such reconstructions by exclusively delivering visible patches to both the online and target encoders. In this section, we describe its main components in detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.476, + 0.218, + 0.49 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.468, + 0.68 + ], + "angle": 0, + "content": "Recently, masked autoencoders (MAE) [23] introduces a novel methodology to the domain of self-supervised visual representation learning. The primary process involves training a model to reconstruct masked areas in images whose contents are partially occluded. The fundamental architecture of a standard MAE model is composed of three core elements: the masking design, an encoder, and a decoder. In mathematical terms, given an image \\( \\pmb{I} \\in 3 \\times H \\times W \\), the MAE model patchifies \\( \\pmb{I} \\) into \\( N \\) distinct, non-overlapping patches, represented by \\( \\hat{\\pmb{I}} \\in N \\times (P^2 \\times 3) \\), each with a size of \\( P \\). Masks are randomly applied to a subset of these patches, denoted by the subset \\( \\mathcal{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.468, + 0.822 + ], + "angle": 0, + "content": "The encoder \\( f_{\\mathcal{E}} \\) processes the visible patches \\( \\mathcal{V} \\), producing latent representations that are represented as \\( z = f_{\\mathcal{E}}(\\mathcal{V}) \\). Then, the decoder \\( f_{\\mathcal{D}} \\) attempts to restore the masked subset \\( \\mathcal{M} \\) utilizing these latent representations, generating \\( O = f_{\\mathcal{D}}(z; \\mathcal{M}) \\), where \\( O \\in N \\times (P^2 \\times 3) \\) depicts the reconstructed patches. The MAE model utilizes the mean-squared error (MSE) to compute the reconstruction loss \\( \\mathcal{L}_{rec} \\), which is solely calculated based on the masked patches and is formulated as \\( \\mathcal{L}_{rec} = \\frac{1}{|\\mathcal{M}|} \\sum_{i \\in \\mathcal{M}} || \\hat{I}_i - O_i ||_2^2 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.442, + 0.846 + ], + "angle": 0, + "content": "3.2. Video Masked Autoencoder with ConvNets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Masking. In this work, we randomly select two consecutive frames, \\(\\pmb{I}^{(1)},\\pmb{I}^{(2)}\\in 3\\times H\\times W\\) and divide them into \\(N\\) nonoverlapping patches, represented as \\(\\hat{I}^{(1)},\\hat{I}^{(2)}\\in N\\times (P^2\\times 3)\\)" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.087, + 0.892, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Figure 3. An illustration of VideoMAC for ConvNet-based MVM. During pre-training, we mask \\(75\\%\\) of symmetric patches from two frames randomly. In our VideoMAC, the MVM of frame pairs is achieved by an online network optimized by gradients \\((\\square)\\), online loss \\(\\mathcal{L}_o\\) and a target network updated by EMA \\((\\square)\\), target loss \\(\\mathcal{L}_t\\). \\(\\mathcal{L}_c\\) is computed as the reconstruction consistency loss between reconstructed patches of frame pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.891, + 0.417 + ], + "angle": 0, + "content": "Regarding the mask configuration, our approach adopts symmetric masking (Table 3d) to guarantee consistency in the positions of the masked areas between frame pairs. Given the consecutive frames, \\(\\mathcal{M}^{(1)}\\) and \\(\\mathcal{M}^{(2)}\\) constitute the subsets of masked patches, whereas \\(\\mathcal{V}^{(1)}\\) and \\(\\mathcal{V}^{(2)}\\) correspond to the visible patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.553 + ], + "angle": 0, + "content": "It is inherent that images and videos possess a greater level of information redundancy [16, 22] compared to languages. Accordingly, we employ a masking ratio of 0.75 (Table 4f) in our methodology. Subsequent empirical analyses (Sec. 4.2) have demonstrated an improvement in performance with higher masking ratios. Nevertheless, employing excessively high masking ratios may induce the risk of training collapse, thereby leading to sudden and substantial decreases in model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.795 + ], + "angle": 0, + "content": "MAC Encoder. In ViT-based models, the prevention of information leakage from masked regions is comparatively straightforward [64, 70]. These models solely employ the visible patches as the exclusive input to the encoder, thereby allowing ViT-based approaches [16, 20, 56, 65] to amplify the efficiency and efficacy of model pre-training by concentrating dedicatedly on unmasked regions during this phase. In contrast, achieving a similar task with ConvNets poses a more intricate dilemma, necessitating the maintenance of the 2D image structure. There are two naive solutions: (i) the instigation of learnable masked markers at the input stage, and (ii) the transformation of the hierarchical convolutional network into an isotropic structure [36] that mirrors the ViT-style technique (Table 2). Regardless, both strategies may decrease pre-training efficiency and possibly engender inconsistencies between training and testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper is at the forefront of proposing an entirely ConvNet-based MVM approach to address this challenge. Drawing inspiration from [54, 64], we utilize sparse convolutional networks to develop a masked autoencoder, which usurps traditional dense convolution. This strategy aims to meet three primary goals: (i) maintaining the structural integrity of the image during feature encoding, (ii) preventing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22735" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "mask dissipation in the course of hierarchical convolution, and (iii) markedly amplifying the computational efficiency of MVM. Specifically, our VideoMAC uses a fully convolutional architecture. Conventionally, ConvNets are structured with four stages, where the resolution of features is halved after each stage, representing a patch size of 32. For an input frame size of \\(224 \\times 224\\), the features are sequentially downsampled at every stage to achieve resolutions of \\(56 \\times 56\\), \\(28 \\times 28\\), \\(14 \\times 14\\), and \\(7 \\times 7\\), respectively. Subsequently, masks are produced and recursively upsampled to the highest resolution (i.e., \\(56 \\times 56\\)) in the final stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.609 + ], + "angle": 0, + "content": "Current MVMs typically necessitate significant computational resources, posing a challenge for practical applications. Our strategy, which substitutes ConvNets for ViTs, does mitigate computational complexity. Furthermore, studying existing MVM methods, we observe that to capture temporal signals within videos, most methods either utilize multi-frame data input or opt for a siamese encoder structure. We diverge from these widespread strategies and instead adopt a more resource-efficient method, namely the online-target encoder architecture (Fig. 3), commonly applied in self-supervised image representation learning. The target encoder's weights are updated by the online encoder via EMA. This architectural integration into our MVM encoding process results in three-fold advantages: (i) considerable reduction of computational complexity, (ii) joint optimization of frame-to-feature representation by the online and target encoders, and (iii) simplicity in incorporating existing ConvNets into our framework, thus improving the performance of downstream tasks. Consequently, visible patches \\(\\mathcal{V}^{(1)}\\) and \\(\\mathcal{V}^{(2)}\\) of consecutive frames are processed by the online encoder \\(f_{\\mathcal{E}}^{o}\\) and the target encoder \\(f_{\\mathcal{E}}^{t}\\) to generate latent representations \\(z^{(1)} = f_{\\mathcal{E}}^{o}(\\mathcal{V}^{(1)})\\) and \\(z^{(2)} = f_{\\mathcal{E}}^{t}(\\mathcal{V}^{(2)})\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.473, + 0.838 + ], + "angle": 0, + "content": "MAC Decoder. Unlike previous methods that utilize hierarchical decoders [54, 55] or transformer decoders[16, 33, 56], our VideoMAC implements a lightweight decoder derived from the ConvNeXt block [36]. This simplifies the design and bolsters the efficiency of the pre-training process. Analogous to our encoder, we construct an online decoder \\( f_{\\mathcal{D}}^{o} \\) and a target decoder \\( f_{\\mathcal{D}}^{t} \\), further employing EMA for parameter updates (Fig. 3). Given the latent representations \\( z^{(1)} \\) and \\( z^{(2)} \\) from two successive frames, decoders \\( f_{\\mathcal{D}}^{o} \\) and \\( f_{\\mathcal{D}}^{t} \\) reconstruct masked patches \\( \\mathcal{M}^{(1)} \\) and \\( \\mathcal{M}^{(2)} \\) to yield \\( O^{(1)} = f_{\\mathcal{D}}^{o}(z^{(1)};\\mathcal{M}^{(1)}) \\) and \\( O^{(2)} = f_{\\mathcal{D}}^{t}(z^{(2)};\\mathcal{M}^{(2)}) \\), respectively. In this paper, we employ a single-block design for the decoder, with the decoder feature projection dimension set to 512. Detailed ablation analysis can be found in Sec. 4.2, and Tables 3b and 3c." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Reconstruction Consistency. In line with the majority of MAE-based methods, we compute the Mean Squared Error (MSE) between the reconstituted and original images to assess the reconstruction loss. We elect to aim for each patch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "of normalized pixels, executing the loss calculation solely for masked patches. As a result, we synchronize the values of normalized pixels within each image patch, applying the loss function exclusively to those patches subject to masking. Significantly, our VideoMAC yields two frames of reconstructed results from both the online and target encoders, necessitating individual loss computations:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.208, + 0.895, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\frac {1}{| \\mathcal {M} ^ {(1)} |} \\sum_ {i \\in \\mathcal {M} ^ {(1)}} \\| \\hat {\\boldsymbol {I}} _ {i} ^ {(1)} - \\boldsymbol {O} _ {i} ^ {(1)} \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.249, + 0.824, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\frac {1}{| \\mathcal {M} ^ {(2)} |} \\sum_ {i \\in \\mathcal {M} ^ {(2)}} | | \\hat {\\boldsymbol {I}} _ {i} ^ {(2)} - \\boldsymbol {O} _ {i} ^ {(2)} | | _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.895, + 0.465 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_o\\) and \\(\\mathcal{L}_t\\) denote online and target losses correspondingly. Due to the correlation between consecutive video frames, we pursue temporal consistency in addition to the reconstruction loss. Here, we introduce an intuitive assumption: if the pixel deviation between the two raw frames tends to a particular value, denoted as \\(\\epsilon\\), it is inferred that the divergence between the reconstructed frames should also trend towards \\(\\epsilon\\). Based on this assumption, we can estimate the reconstruction consistency between two frames and use it as a proxy for temporal consistency. Hence, the reconstruction consistency loss \\(\\mathcal{L}_c\\) is thus expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.473, + 0.892, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{| \\mathcal {M} ^ {(1)} |} \\sum_ {i \\in \\mathcal {M} ^ {(1)}, j \\in \\mathcal {M} ^ {(2)}} \\| \\boldsymbol {O} _ {i} ^ {(1)} - \\boldsymbol {O} _ {j} ^ {(2)} \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.571 + ], + "angle": 0, + "content": "where \\(|\\mathcal{M}^{(1)}|\\) can be substituted by \\(|\\mathcal{M}^{(2)}|\\), courtesy of our utilization of symmetric masking. Ultimately, our approach culminates in a total loss expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.583, + 0.892, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {o} + \\mathcal {L} _ {t} + \\gamma \\mathcal {L} _ {c}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.895, + 0.704 + ], + "angle": 0, + "content": "with \\(\\gamma\\) symbolizing the weight factor of the consistency loss (see Table 4e and Fig. 7 for detailed analysis). Accordingly, our approach, VideoMAC, accomplishes the realization of a fully convolutional network MVM. Moreover, the reconstruction consistency loss introduced in this paper provides a promising solution for modeling spatio-temporal data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.661, + 0.728 + ], + "angle": 0, + "content": "3.3. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Architecture. Our VideoMAC possesses the adaptability to employ any convolutional network the an encoder. To permit a fair comparison, we choose two representative convolutional networks for our experiments: ResNet (RN) [21] and ConvNeXtv2 (CNXv2) [64]. Although RN50 (26M) and CNXv2-S (50M) serve as the primary experimental models, we further investigate models of diverse sizes (Tables 2, 3a and 5). In this paper, the decoder component involves a readily available ConvNeXt block [36], thereby dispensing with the necessity for intricate network design. During pre-training, we employ an encoder implemented" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22736" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.318 + ], + "angle": 0, + "content": "
MethodTypeBackboneTraining datasetEpochDAVIS17 [46]VIP [73] mIoUJHMDB [26] PCK@0.1PCK@0.2
J&FmJmFm
Supervised [21]CVPR16SLRN50IN1K (1.28M, -)10066.063.768.439.559.278.3
DenseCL [63]CVPR21CLRN50IN1K (1.28M, -)20061.159.862.632.556.178.0
DINO [6]ICCV21CLViT-S/16IN1K (1.28M, -)80061.860.263.436.245.675.0
ODIN² [24]ECCV22CLRN50IN1K (1.28M, -)100054.154.353.9---
CrOC [52]CVPR23CLViT-S/16IN1K (1.28M, -)30044.743.545.926.1--
MAE [23]CVPR22MIMViT-B/16IN1K (1.28M, -)160053.552.155.028.144.673.4
FCMAE [64]CVPR23MIMCNXv2-BIN1K (1.28M, -)160043.741.945.524.951.774.2
RC-MAE [30]ICLR23MIMViT-S/16IN1K (1.28M, -)160049.248.950.529.743.272.3
SparK [54]ICLR23MIMRN50IN1K (1.28M, -)160055.656.354.932.852.975.2
MAE-ST [16]NeurIPS22MVMViT-L/16K400 (-, 800 hours)160054.655.553.633.244.472.5
VideoMAE [56]NeurIPS22MVMViT-S/16K400 (-, 800 hours)160039.339.738.922.341.067.9
DropMAE [65]CVPR23MVMViT-B/16K400 (-, 800 hours)160053.451.855.031.142.369.2
SiamMAE [20]NeurIPS23MVMViT-S/16K400 (-, 800 hours)200062.060.363.737.347.076.1
Ours
VideoMACMVMRN50YT18 (-, 5.58 hours)10067.264.969.543.657.279.2
VideoMACMVMCNXv2-SYT18 (-, 5.58 hours)10068.465.371.440.458.180.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.318, + 0.894, + 0.358 + ], + "angle": 0, + "content": "SL: Supervised Learning. CL: Contrastive Learning. IN1K: ImageNet1K [11]. K400: Kinetics-400 [27]. YT18: YouTube-VOS 2018 [68]. Table 1. Comparing VideoMAC with previous supervised and self-supervised approaches (SL, CL, MIM, and MVM) on downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.364, + 0.47, + 0.535 + ], + "angle": 0, + "content": "
MVM MethodResolution#Param.FLOPsJ&Fm
Iso. ViT Backbone
MAE-ST [16]14 × 14304M61.6G54.6
VideoMAE [56]14 × 1422M4.6G39.3
DropMAE [65]14 × 1487M17.6G53.4
SiamMAE [20]14 × 1422M4.6G62.0
Iso. ConvNet Backbone
VideoMAC (CNXv2-S†)14 × 1422M4.3G53.8
Hie. ConvNet Backbone
VideoMAC (RN18)7 × 712M1.8G64.7
VideoMAC (RN50)7 × 726M4.1G67.2
VideoMAC (CNXv2-T)7 × 729M4.5G67.5
VideoMAC (CNXv2-S)7 × 750M8.7G68.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.536, + 0.47, + 0.563 + ], + "angle": 0, + "content": "Table 2. Comparisons with isotropic (iso.) and hierarchical (hie.) based MVM methods. ‘†’ denotes our modified ios. version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.63 + ], + "angle": 0, + "content": "with the MinkowskiEngine sparse convolution library [9]. It's noteworthy that the sparse convolution layer can be seamlessly converted back to a standard convolution layer in downstream tasks without requiring additional processes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Pre-training. Different from previous methods that pretrain on the large-scale video dataset K400 [27], VideoMAC pre-trains on YouTube-VOS [68] for 100 epochs, leading to a significant reduction in pre-training time by over 140 times. During VideoMAC pre-training, the standard input comprises two frames with a spatial size of \\(224 \\times 224\\) pixels. The frame interval for this study is determined to be 1 (Table 4c), implying consecutive frames are chosen. We employ minimal data augmentation, encompassing only random resized cropping and horizontal flipping. A masking ratio of 0.75 is utilized to pre-train our models. We conduct training using an AdamW optimizer [39] with a batch size of 512 and a learning rate of 2e-3. Cosine decay [38] is applied to adjust the learning rate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Downstream Tasks. To gauge the efficacy of VideoMAC at video representation learning, we undertake comprehensive experiments on three video downstream tasks, namely video object segmentation, body part propagation, and hu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.424 + ], + "angle": 0, + "content": "man pose tracking (Table 1). We also investigate the performance of image recognition through video pre-training. Specifically, our backbone is pre-trained on video data and then fine-tuned on ImageNet1K [11] (Table 5)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.437, + 0.634, + 0.454 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.567 + ], + "angle": 0, + "content": "This section begins with assessing VideoMAC's performance in three downstream tasks, comparing it to previous state-of-the-art methods. Subsequently, we conduct extensive ablation studies to elucidate the effectiveness of our proposed approach. The final part of this section explores VideoMAC's ability for image recognition, accompanied by a visualization of the reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.576, + 0.81, + 0.592 + ], + "angle": 0, + "content": "4.1. Comparison with Previous methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video Object Segmentation. We evaluate VideoMAC on the popular DAVIS17 [46] dataset, which includes a validation set comprising 30 videos. Our evaluation methodology aligns with [6, 25], and we report the typical evaluation metrics [45], including the mean region similarity \\(\\mathcal{I}_m\\), the mean boundary accuracy \\(\\mathcal{F}_m\\), and their average values \\(\\mathcal{J} \\& \\mathcal{F}_m\\). As shown in Table 1, our VideoMAC achieves state-of-the-art results for video object segmentation. With a smaller video training dataset and fewer training epochs, both our VideoMAC models (i.e., CNXv2-S and RN50) surpass the SiamMAE [20] by \\(6.4\\%\\) and \\(5.2\\%\\), respectively, in terms of \\(\\mathcal{J} \\& \\mathcal{F}_m\\). It's worth noting that our approach achieves superior performance to supervised method [21] based on MVM, with improvements of \\(2.4\\%\\) and \\(1.4\\%\\), respectively. Owing to the hierarchical feature extraction capabilities inherent in ConvNets, our approach demonstrates superior proficiency in spatial information extraction in contrast to ViT-based MVM methods [16, 20, 56, 65]. Concurrently, the introduced reconstruction consistency framework strengthens the acquisition of temporal correlations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.181, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.131, + 0.18, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.172, + 0.18, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.212, + 0.179, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.089, + 0.28, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.131, + 0.28, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.172, + 0.28, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.212, + 0.28, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.089, + 0.381, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.131, + 0.381, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.172, + 0.381, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.212, + 0.381, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.089, + 0.483, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.131, + 0.483, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.172, + 0.483, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.212, + 0.483, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.089, + 0.588, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.13, + 0.588, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.089, + 0.689, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.13, + 0.689, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.089, + 0.79, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.13, + 0.789, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.089, + 0.891, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.127, + 0.891, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.165, + 0.76, + 0.177 + ], + "angle": 0, + "content": "(b) Body Part Propagation" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.177, + 0.588, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.215, + 0.587, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.177, + 0.688, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.215, + 0.688, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.177, + 0.79, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.215, + 0.788, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.177, + 0.891, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.215, + 0.891, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.254, + 0.365, + 0.266 + ], + "angle": 0, + "content": "(a) Video Object Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.255, + 0.758, + 0.266 + ], + "angle": 0, + "content": "(c) Human Pose Tracking" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of our VideoMAC (using CNXv2-S) for three video downstream tasks: (a) video object segmentation on DAVIS17 [46]), (b) body part propagation on VIP [73], and (c) human pose tracking on JHMDB [26]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.307, + 0.33, + 0.382 + ], + "angle": 0, + "content": "
BackboneJ&FmJmFm
R1864.762.866.5
R5067.264.969.5
CNXv2-T67.564.470.6
CNXv2-S68.465.371.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.384, + 0.33, + 0.407 + ], + "angle": 0, + "content": "(a) Encoder type. VideoMAC is compatible with both classical and modern ConvNets." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.409, + 0.33, + 0.484 + ], + "angle": 0, + "content": "
Dim.J&FmJmFm
12867.964.871.0
25668.265.371.1
51268.465.371.4
102468.165.071.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.485, + 0.331, + 0.509 + ], + "angle": 0, + "content": "(c) Decoder width. A decoder width of 512 delivers the best performance." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.307, + 0.593, + 0.382 + ], + "angle": 0, + "content": "
BlocksJ&FmJmFm
168.465.371.4
268.265.271.2
468.365.471.2
868.165.271.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.35, + 0.384, + 0.593, + 0.408 + ], + "angle": 0, + "content": "(b) Decoder depth. One block achieves competitive performance and efficiency." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.409, + 0.593, + 0.455 + ], + "angle": 0, + "content": "
MaskingJ&FmJmFm
Asym.66.563.469.6
Sym.68.465.371.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.349, + 0.485, + 0.594, + 0.51 + ], + "angle": 0, + "content": "(d) Masking design. Symmetric (sym.) masking outperforms asymmetric (asym.) one." + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.306, + 0.888, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.5, + 0.888, + 0.569 + ], + "angle": 0, + "content": "Figure 5. For masked modeling, dense convolution usually results in the dissipation of mask structures. The deployment of sparse convolution proves to be an effective solution, enabling ConvNet-based MIM / MVM." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.512, + 0.603, + 0.569 + ], + "angle": 0, + "content": "Table 3. VideoMAC encoder and decoder ablation experiments on DAVIS17 [46]. We report ablation results for two of the most emblematic families of ConvNets. Unless specified otherwise, the encoder is CNXv2-S in default. 'Dim' indicates the number of decoder dimensions. Default settings employed in this paper are marked in blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Body Part Propagation. We evaluate VideoMAC on the Video Instance Parsing (VIP) dataset [73], and the quantitative results, measured by mIoU, are detailed in Table 1. This validation set comprises of 50 videos that concentrate on the propagation of 19 body parts, such as arms and legs, thereby necessitating a heightened level of precision in matching as compared to video object segmentation. We adhere to the same parameters as [25], adjusting the video frames to \\(560 \\times 560\\) dimensions. In the case of body part propagation, two variants of our VideoMAC demonstrate improvements by \\(3.1\\%\\) (CNXv2-S) and \\(6.3\\%\\) (RN50), respectively, in comparison with the top-performing SiamMAE [20] in MVM. Significantly, our RN50-based model outperforms CNXv2-S on the VIP dataset [73] due to the higher dimensionality of RN50 features (2048 vs. 768), which benefits complex tasks and improves accuracy in mask propagation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Human Pose Tracking. We conduct the human pose tracking task on the validation set of JHMDB [26], which includes 268 videos and entails detecting 15 human keypoints. Conforming to the protocol delineated in [25], we resize video frames to a resolution of \\(320 \\times 320\\), and employ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.642 + ], + "angle": 0, + "content": "the Possibility of Correct Keypoints (PCK) metric [51] for quantitative evaluation. As shown in Table 1, both variants (i.e., RN50 and CNXv2-S) of our VideoMAC outperforms all MVM methods in terms of PCK@0.1 and PCK@0.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Qualitative results. As depicted in Fig. 4, we present the qualitative results of mask propagation achieved by VideoMAC on three downstream tasks. Our VideoMAC yields notable performance across a spectrum of complex videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.664, + 0.726 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.891, + 0.764 + ], + "angle": 0, + "content": "In this section, we conduct extensive ablation studies to analyze the component design of our VideoMAC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Encoder design: hierarchical vs. isotropic. In this ablation study, we scrutinize diverse encoder architectures to our VideoMAC, i.e., hierarchical and isotropic designs. Precisely, we adopted the ViT [14] design principles and restructured the CNXv2-S backbone from having a hierarchical structure into an isotropic one. This implies the creation of an architecture devoid of downsampling layers, ensuring consistent feature resolution (e.g., \\(14 \\times 14\\)) throughout all depths. As shown in Table 2, the performance dif" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22738" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.405, + 0.134 + ], + "angle": 0, + "content": "
DatasetSizeJ&FmJmFm
YT18 [68]5.58 hours68.465.371.4
K4001 [27]800 hours67.865.270.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.149, + 0.405, + 0.173 + ], + "angle": 0, + "content": "(a) Training Dataset. Our VideoMAC pre-trained on YT18 yields better performance and efficiency compared to K400." + }, + { + "type": "table", + "bbox": [ + 0.421, + 0.089, + 0.677, + 0.147 + ], + "angle": 0, + "content": "
AugmentationJ&FmJmFm
spatial68.465.371.4
color64.961.868.0
spatial + color66.162.969.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.411, + 0.149, + 0.677, + 0.173 + ], + "angle": 0, + "content": "(b) Data augmentation. Minimal data augmentation achieves the best performance." + }, + { + "type": "table", + "bbox": [ + 0.693, + 0.089, + 0.89, + 0.147 + ], + "angle": 0, + "content": "
GapJ&FmJmFm
168.465.371.4
566.463.569.2
1063.260.465.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.683, + 0.15, + 0.891, + 0.173 + ], + "angle": 0, + "content": "(c) Frame gap. Video pairs with one frame gap perform the best results." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.176, + 0.334, + 0.249 + ], + "angle": 0, + "content": "
\\( {\\mathcal{L}}_{c} \\)\\( {\\mathcal{L}}_{t} \\)\\( \\mathcal{J}\\&{\\mathcal{F}}_{m} \\)\\( {\\mathcal{J}}_{m} \\)\\( {\\mathcal{F}}_{m} \\)
✘✘51.152.250.0
✘✓60.656.364.9
✓✘63.463.468.7
✓✓68.465.371.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.251, + 0.335, + 0.275 + ], + "angle": 0, + "content": "(d) Model loss. VideoMAC utilizes consistency loss \\(\\mathcal{L}_c\\) and target loss \\(\\mathcal{L}_t\\) for the best." + }, + { + "type": "table", + "bbox": [ + 0.36, + 0.176, + 0.603, + 0.25 + ], + "angle": 0, + "content": "
ConsistencyJ&FmJmFm
γ=0.060.656.364.9
γ=0.164.862.067.6
γ=0.565.863.268.5
γ=1.068.465.371.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.349, + 0.251, + 0.604, + 0.275 + ], + "angle": 0, + "content": "(e) Reconstruction consistency. When \\(\\gamma\\) is set to 1.0, the consistency loss works the best." + }, + { + "type": "table", + "bbox": [ + 0.629, + 0.176, + 0.888, + 0.249 + ], + "angle": 0, + "content": "
Masking ratioJ&FmJmFm
0.6566.563.869.1
0.7568.465.371.4
0.8568.365.371.3
0.9560.359.760.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.251, + 0.889, + 0.275 + ], + "angle": 0, + "content": "(f) Masking ratio. Our VideoMAC uses a masking ratio of 0.75 for the highest performance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.277, + 0.891, + 0.291 + ], + "angle": 0, + "content": "Table 4. VideoMAC ablations on data, loss, and masking ratio configuration. We report two official metrics \\( {\\mathcal{I}}_{m} \\) and \\( {\\mathcal{F}}_{m} \\) on DAVIS17 [46]." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.3, + 0.272, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.3, + 0.469, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.423, + 0.469, + 0.45 + ], + "angle": 0, + "content": "Figure 6. Analysis of individual loss components after Video-MAC pre-training 100 epochs on YouTube-VOS 2018 [68]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.577 + ], + "angle": 0, + "content": "ference between our VideoMAC (iso.) and the ViT-based MVM methods (e.g., VideoMAE [56] and DropMAE [65]) is marginal. By maintaining the hierarchical structure of ConvNets within VideoMAC models (hie.), we improve the performance significantly, even though their feature resolution (i.e., \\(7 \\times 7\\)) is merely half that of the ViT-based models [16, 20, 56, 65]. This can be ascribed to ConvNets' capacity to preserve multi-scale spatial information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.746 + ], + "angle": 0, + "content": "Encoder backbone: classical and modern. For our study, we incorporate two emblematic ConvNet families, i.e., ResNets [21] and ConvNeXts [36, 64]. As shown in Table 3a, we substitute dense convolution with sparse convolution in the baseline ConvNets. This adaptation seamlessly dovetails into our VideoMAC framework, thereby facilitating a ConvNet-based MVM. In addition, Fig. 5 highlights the influence of employing dense and sparse convolution on masks. It is apparent that sparse convolution effectively safeguards the structural integrity of the mask, thus enabling the integration of MAE pre-training within ConvNets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Decoder design: blocks and dimensions. Adhering to a simplistic design principle, our VideoMAC adopts the decoder from the purely convolutional network structure deployed in ConvNeXt [36]. We conduct ablation experiments focusing on the depth and width of the decoder, with the results detailed in Tables 3b and 3c. The decoder exhibits competitive results with a single block and 512 dimensions, striking a trade-off between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Masking design: asymmetric vs. symmetric. In Table 3d, we investigate the impact of two masking strategies, namely" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.3, + 0.695, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.3, + 0.892, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.423, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Figure 7. Consistency loss in relation to \\(\\gamma\\) (left: identical encoder, CNXv2-S) and encoders (right: identical weight factor, \\(\\gamma = 1.0\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.561 + ], + "angle": 0, + "content": "asymmetric (same masking ratio but different locations) and symmetric (same masking ratio and locations), on performance. Compared to asymmetric masking, symmetric masking simplifies the calculation of the reconstruction loss, as it covers the entire masked area. Moreover, it facilitates the inclusion of more reconstruction regions in the loss calculation, resulting in faster model convergence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.878 + ], + "angle": 0, + "content": "Data setting: dataset, augmentation, and gap. To assess the influence of different video datasets, Table 4a presents the results of pre-training our VideoMAC on both YT18 [68] and \\(\\mathrm{K400^1}\\) [27]. The ablation results of VideoMAC on the two datasets differ by just \\(0.6\\%\\) , demonstrating the robustness and adaptability of our method to different datasets. For data augmentation, as shown in Table 4b, our VideoMAC achieves the best results using only spatial strategy, i.e., random resized cropping and horizontal flipping. Additionally, we conduct an ablation study on the frame gap, and the optimal results are obtained with consecutive frames (i.e., the gap is 1), as indicated in Table 4c. The contribution of \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_t\\) . Previous MVM methods predominantly rely on a single loss, i.e., the reconstruction loss, necessitating multiple frames to be simultaneously processed by the model. In contrast, our study utilizes a more computation-friendly strategy by inputting frame pairs into two separate encoders, followed by parameter updates using EMA. Consequently, we introduce two distinct loss terms: online \\(\\mathcal{L}_o\\) and target \\(\\mathcal{L}_t\\) reconstruction losses. Furthermore, to integrate temporal information, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.821, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)K400 is sampled to match YT18's training data volume." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22739" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.89, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.203, + 0.893, + 0.231 + ], + "angle": 0, + "content": "Figure 8. Visualization of reconstruction examples using a pre-trained VideoMAC (CNXv2-S) model with 0.75 masking ratio. These frames are randomly selected from the validation set of YT18 [68] and masked \\(60\\%\\) of patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.468, + 0.331 + ], + "angle": 0, + "content": "compute the loss between the reconstructed results of frame pairs, colloquially known as the reconstruction consistency loss \\(\\mathcal{L}_c\\). As shown in Table 4d, \\(\\mathcal{L}_t\\) and \\(\\mathcal{L}_c\\) improve \\(9.5\\%\\) and \\(12.3\\%\\) in terms of \\(\\mathcal{J} \\& \\mathcal{F}_m\\), respectively, with the full VideoMAC model yielding the best results. The diverse loss components during pre-training are illustrated in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.468, + 0.573 + ], + "angle": 0, + "content": "The effect of weight factor \\(\\gamma\\). To investigate the impact of the weight factor on reconstruction consistency loss \\(\\mathcal{L}_c\\), we conduct experiments with different factor settings, i.e., \\(\\gamma \\in \\{0.0, 0.1, 0.5, 1.0\\}\\). As illustrated in Table 4e, superior performance is attained when \\(\\gamma\\) approximates to 1. When \\(\\gamma\\) equals 0, which signifies the absence of the reconstruction consistency loss term, a detrimental effect on performance is noticeable. Simultaneously, we record \\(\\mathcal{L}_c\\) for various weight factors during pre-training, as depicted in Fig. 7. It becomes evident that an elevated value of \\(\\gamma\\) results in a lesser loss. Ultimately, both the R50 and CNXv2-S models converge to comparable final losses. It is noteworthy that \\(\\mathcal{L}_c\\) presents an initial decrease, succeeded by an increase, and ultimately reaches convergence, which aligns with the description in Eq. (2), implying that the ultimate reconstruction results between two frames converge to \\(\\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.468, + 0.696 + ], + "angle": 0, + "content": "The impact of masking ratio. An ablation analysis of the masking ratio is conducted, and the results for our Video-MAC models (CNXv2-S) are documented in Table 4f. Ablation results portray a performance pattern that first ascends and subsequently descends with increasing masking ratios. Noticeably, the performance of our approach experiences a drastic decline at a 0.95 masking ratio caused by overfitting to the self-supervised pretext task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.299, + 0.727 + ], + "angle": 0, + "content": "4.3. Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "VideoMAC for Image Recognition. As shown in Table 5, utilizing CNXv1-T [36] / CNXv2-T [64] pre-trained by our VideoMAC on YT18 [68], we achieve \\(82.3\\%\\) / \\(82.9\\%\\) accuracy on IN1K [11] image classification. In comparison to CNXv1-T and CNXv2-T trained from scratch, our approach improves the image recognition results by \\(0.2\\%\\) and \\(0.4\\%\\). Remarkably, the performance garnered by VideoMAC (MVM) is proximate to or commensurate with the state-of-the-art ConvNet-based MIM methods [54, 64]. Given the pre-training cost (100 epochs vs. 1600 epochs) and the domain gap (video vs. image), we regard this im" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.244, + 0.892, + 0.35 + ], + "angle": 0, + "content": "
MethodBackbonePT epochPT dataFT acc.
Supervised [36]CNXv1-T--82.1
FCMAE [64]CNXv1-T1600IN1K82.3 (+0.2)
Spark [54]CNXv1-T1600IN1K82.4 (+0.3)
VideoMACCNXv1-T100YT1882.3 (+0.2)
Supervised [64]CNXv2-T--82.5
FCMAE [64]CNXv2-T1600IN1K83.0 (+0.5)
VideoMACCNXv2-T100YT1882.9 (+0.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.352, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Table 5. Comparing VideoMAC with previous ConvNet-based MIM approaches (e.g., Spark [54] and FCMAE [64]) on the image downstream task, i.e., fine-tuning recognition accuracy (acc.) on IN1K [11]. 'PT' and 'FT' indicate pre-training and fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.506 + ], + "angle": 0, + "content": "provement of our VideoMAC as noteworthy and promising compared to training from scratch and pre-training by MIM. Visualization. We visualize reconstruction results from the YT18 [68] validation set to analyze performance in MVM pre-training. In Fig. 8, VideoMAC makes reasonable predictions about the masked regions, e.g., color and contour." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.521, + 0.619, + 0.537 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.713 + ], + "angle": 0, + "content": "In this study, we present VideoMAC, an MVM pre-training framework constructed fully using ConvNets. Our masked modeling for frame pairs is executed by introducing an online network (encoder and decoder) optimized by gradients, as well as the target network that updates parameters via EMA. Concurrently, for reconstruction results derived from both online and target networks, our proposed loss of reconstruction consistency between frame pairs paves the way for temporal modeling. Our VideoMAC empowers a multitude of ConvNets to reap the performance gains for downstream tasks (e.g., video but also image) by MVM pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Limitation and Future Work. Existing ConvNet-based architectures typically use a fixed downsampling ratio, which in turn leads to a relatively fixed patch size. For future work, we will investigate ConvNets with adjustable patch size, allowing for higher feature resolution like ViT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No. 62102182, 62202227, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20220934, BK20220938, BK20220936), China Postdoctoral Science Foundation (No. 2022M721626, 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No.30923010303)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.47, + 0.298 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.3, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[4] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912–9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[9] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, , and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[13] Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language understanding and generation. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.869 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, haoqi fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, pages 35946-35958, 2022. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[17] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. NeurIPS, 35:35632-35644, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[19] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[20] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In ECCV, pages 123-143, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[25] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, pages 19545-19560, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, pages 3192-3199, 2013. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[29] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[30] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. In ICLR, 2023. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Lulei Li, Tianfei Zhou, Wenguan Wang, Lu Yang, Jianwu Li, and Yi Yang. Locality-aware inter-and intra-video reconstruction for self-supervised correspondence learning. In CVPR, pages 8719-8730, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[32] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In CVPR, pages 18706-18716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[33] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In CVPR, pages 6252-6261, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[34] Yuan Liu, Songyang Zhang, Jiacheng Chen, Zhaohui Yu, Kai Chen, and Dahua Lin. Improving pixel-based mim by reducing wasted modeling capability. In ICCV, pages 5361-5372, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.567 + ], + "angle": 0, + "content": "[40] Xiankai Lu, Wenguan Wang, Martin Danelljan, Tianfei Zhou, Jianbing Shen, and Luc Van Gool. Video object segmentation with episodic graph memory networks. In ECCV, pages 661-679, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[41] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In CVPR, pages 8960-8970, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[42] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, pages 596-613, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[43] Gensheng Pei, Fumin Shen, Yazhou Yao, Tao Chen, Xian-Sheng Hua, and Heng-Tao Shen. Hierarchical graph pattern understanding for zero-shot video object segmentation. TIP, 32:5909-5920, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[44] Gensheng Pei, Yazhou Yao, Fumin Shen, Dan Huang, Xingguo Huang, and Heng-Tao Shen. Hierarchical co-attention propagation network for zero-shot video object segmentation. TIP, pages 2348–2359, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[45] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[46] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Malik, Yanghao Li, and Christoph Feichtenhofer. Hiera: A hierarchical vision transformer without the bells-and-whistles. ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[51] Jie Song, Limin Wang, Luc Van Gool, and Otmar Hilliges. Thin-slicing network: A deep structured model for pose estimation in videos. In CVPR, pages 4220-4229, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In CVPR, pages 7000–7009, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[54] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. In ICLR, 2023. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[55] Yunjie Tian, Lingxi Xie, Zhaozhi Wang, Longhui Wei, Xiaopeng Zhang, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Integrally pre-trained transformer pyramid networks. In CVPR, pages 18610-18620, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[57] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[58] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[60] Wenguan Wang, Xiankai Lu, Jianbing Shen, David J Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, pages 9236-9245, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[61] Wenguan Wang, Hongmei Song, Shuyang Zhao, Jianbing Shen, Sanyuan Zhao, Steven CH Hoi, and Haibin Ling." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22742" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Learning unsupervised video object segmentation through visual attention. In CVPR, pages 3064-3074, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[62] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[63] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[64] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In CVPR, pages 16133-16142, 2023. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[65] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In CVPR, pages 14561–14571, 2023. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[66] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, pages 13879-13889, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[67] Z. Xie, Z. Zhang, Y. Cao, Y. Lin, J. Bao, and Z. Yao. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.468, + 0.516 + ], + "angle": 0, + "content": "[68] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, pages 585-601, 2018. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[69] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, pages 2623-2632, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[70] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.468, + 0.686 + ], + "angle": 0, + "content": "[71] Yurong Zhang, Lielei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In CVPR, pages 2246-2256, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[73] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In ACMMM, pages 1527-1535, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[74] Tianfei Zhou, Fatih Porikli, David J Crandall, Luc Van Gool, and Wenguan Wang. A survey on deep learning technique for video segmentation. TPAMI, 45(6):7099-7122, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[75] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, pages 408-417, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22743" + } + ] +] \ No newline at end of file diff --git a/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_origin.pdf b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c6c3f3ee099b21039189315d46ecd7744d389c43 --- /dev/null +++ b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/d723e353-3b17-4db6-af6c-92f3b43a3ed7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da8c2f20c97e0684a278b7d3856ecc308ebbc72772748fbc1acf847c76e46fff +size 2875981 diff --git a/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/full.md b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bbe5833dad331c62eec19e617458ab6f81016e86 --- /dev/null +++ b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/full.md @@ -0,0 +1,414 @@ +# VideoMAC: Video Masked Autoencoders Meet ConvNets + +Gensheng Pei\*, Tao Chen\*, Xiruo Jiang, Huafeng Liu, Zeren Sun,\* Yazhou Yao\* +School of Computer Science and Engineering, Nanjing University of Science and Technology +https://github.com/NUST-Machine-Intelligence-Laboratory/VideoMAC + +# Abstract + +Recently, the advancement of self-supervised learning techniques, like masked autoencoders (MAE), has greatly influenced visual representation learning for images and videos. Nevertheless, it is worth noting that the predominant approaches in existing masked image / video modeling rely excessively on resource-intensive vision transformers (ViTs) as the feature encoder. In this paper, we propose a new approach termed as VideoMAC, which combines video masked autoencoders with resource-friendly ConvNets. Specifically, VideoMAC employs symmetric masking on randomly sampled pairs of video frames. To prevent the issue of mask pattern dissipation, we utilize ConvNets which are implemented with sparse convolutional operators as encoders. Simultaneously, we present a simple yet effective masked video modeling (MVM) approach, a dual encoder architecture comprising an online encoder and an exponential moving average target encoder, aimed to facilitate inter-frame reconstruction consistency in videos. Additionally, we demonstrate that VideoMAC, empowering classical (ResNet) / modern (ConvNeXt) convolutional encoders to harness the benefits of MVM, outperforms ViT-based approaches on downstream tasks, including video object segmentation (+5.2% / 6.4% J&F), body part propagation (+6.3% / 3.1% mIoU), and human pose tracking (+10.2% / 11.1% PCK@0.1). + +# 1. Introduction + +Spurred by the success of masked language modeling in natural language processing [3, 10, 12], transformer-based visual models, notably vision transformers (ViT) [14] adopt masked image modeling (MIM) are attracting attention [1, 2, 15, 23, 34, 67]. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations. + +![](images/ca4cde1d18bf259bdb656eb6c157807b92cbcd0f569194a2f8bd19d1734e5573.jpg) +Figure 1. State-of-the-art MAE architectures (e.g., MAE-ST [16] and SiamMAE [20]) in masked video modeling commonly employ ViT-based encoders. We propose VideoMAC, a new video masked autoencoder built using pure ConvNets. In this study, we model VideoMAC with two of the most emblematic families of ConvNets, namely ConvNeXtv2 [64] and ResNet [21]. Notably, VideoMAC exhibits superior performance on a range of downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26], compared to ViT-based methods [16, 20]. + +With the successful application of MIM, some pioneering works [16, 56] have shifted the focus of video representation learning to masked video modeling (MVM), yielding promising results in downstream video tasks. However, features learned by MVM are designed explicitly for pixel reconstruction tasks. They perform well when fine-tuned for downstream video tasks but exhibit subpar performance without specific fine-tuning [20]. Their performance is even inferior to MIM methods [23, 30], and it lags significantly behind supervised methods [21, 36]. + +We strive to discern the factors contributing to this issue. To the best of our knowledge, most current MVM methods are based on the isotropic design of ViT [16, 20, 56, 65]. The reason is that early MIM / MVM methods are explicitly developed for ViT, as they allow the use of only visible patches as the sole input to the encoder, which aligns well with the ViT architecture. For instance, the representative approach, masked autoencoders (MAE) [23], enhances the efficiency and effectiveness of model pre-training by focusing predominantly on the unmasked regions. However, + +![](images/9dd50d2de52ff30ad19264f6594a7ddac788c602a2a31c43d24a041d7ea95ab6.jpg) +(a) +Input + +![](images/b906e3d8a645712c24a9cdbb5f24f335869a51f2ce49c52c500e828aa989f48a.jpg) +Masked + +![](images/5d91cf721317ff40cad8b85c77c3131f313349c43d3546d585cb74dd3945a40d.jpg) +Reconstructed + +![](images/d6dae044ebd5e2836c4c184376f9ae28cec74de1da4a4a676d23fbf699acd5cb.jpg) +Heatmap + +![](images/b434930b33d94696817c659d12af41a5de19fd5199265af242645b34307bea6f.jpg) +(b) +Figure 2. Visualization of the heatmap during the reconstruction of masked patches in a random frame pair. It is evident that our approach highlights similar regions for both (a) past and (b) current frames, proficiently reconstructing colors and contours. + +![](images/472226c5798a3a7ed092aec83705bde8c203ebbce9f7ef47c68ffa4fab33049a.jpg) + +![](images/5c1999b4597d7fa62c72dbb2107d52fcf31a4f4ad78448beed7b1bbed1622bd3.jpg) + +![](images/f91653fc48b39211fc0570d2fdaa2375e0060af10dd9bc2808bba78420ae9e3e.jpg) + +previous MVM approaches [16, 56], confined by the plain vision transformer design devoid of a hierarchical structure tend to result in inadequate spatial modeling capabilities for local features [17, 35, 49, 62, 70]. Such capabilities are crucial for dense downstream tasks like segmentation and detection. Moreover, the uniform reliance on ViT in existing MVM methods has resulted in significant computational resource consumption [16, 56, 58]. One might first consider ConvNets [21, 28, 29, 36, 50, 53], characterized by their low resource consumption and built-in hierarchical structure, as an ideal solution to tackle the aforementioned issues. Nevertheless, ConvNets appears understudied in MVM. Consequently, it raises the question: can ConvNets, endowed with a pyramid vision architecture, guide MVM towards the era of hierarchical pre-training? + +To address this concern, we propose a simple yet potent MVM approach, referred to as Video Masked Autoencoders with ConvNets (VideoMAC). Uniquely, our method represents an initial endeavor to supplement video pre-training by integrating hierarchical ConvNets. However, applying ConvNets to MVM proves more complex than utilizing ViTs. Typical ConvNets might be incompatible with MVM due to the employment of dense sliding windows with overlap, resulting in information leakage within masked regions. To efficiently incorporate mask modeling into ConvNets, we opt for sparse convolution over dense convolution to restructure the original ConvNets, an approach validated in recent studies [54, 64] to effectively maintain the structure of the masked region when utilizing ConvNets. Another hurdle for MVM is the effective incorporation of temporal data. Current MVM pre-training methods [16, 56] involving multi-frame inputs are primarily used in downstream resource-intensive tasks such as video action recognition. Although methods using two-frame inputs for pre-training lower resource usage [20, 65], they typically employ the Siamese architecture, necessitating robust skills for designing reconstruction targets. Importantly, all aforementioned methods utilize computationally intensive, non-hierarchical ViT [14] as their encoders. To this end, we utilize pairs of frames as inputs and introduce the reconstruction consis + +tency loss, a streamlined extension of MAE to accommodate spatio-temporal data. We formulate online networks through gradient optimization and establish target networks by updating parameters via an exponential moving average (EMA). This allows for individual recovery of masked frame pairs and computation of the reconstruction consistency loss between them. As illustrated in Fig. 2, our approach can focus on similar regions between two frames and reconstruct the masked portions efficiently. + +Notably, our approach transcends previous state-of-the-art techniques in three downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]. Moreover, we present promising results utilizing VideoMAC in fine-tuning for image recognition on ImageNet1K [11]. To summarize, our main contributions of this work include: + +- We propose a simple yet potent solution of masked video autoencoders, marking the initial endeavor to successfully implement MVM using pure ConvNets. +- We introduce the reconstruction consistency loss to enable elegant modeling of spatio-temporal data. +- Our approach significantly outperforms existing ViT-based MVM methods in various downstream tasks. + +While the focus of the computer vision community is increasingly shifting towards vision transformers, we trust that VideoMAC will rekindle interest in exploring ConvNet-based MVM, thereby catalyzing the discovery of new potentials across a variety of video tasks. + +# 2. Related Work + +Masked Image Modeling. Self-supervised learning (SSL) has been rapidly developing in computer vision, demonstrating substantial potential for deriving valuable representations from extensive, unlabeled datasets. A key advantage of SSL [2, 5-8, 18, 22, 48, 57, 69, 72] is its independence from costly human annotations, thereby providing robust prior knowledge across various visual tasks. The success of masked language modeling in NLP [3, 10, 12, 13, 47] has catalyzed a rising interest in masked image modeling (MIM) for visual pre-training. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations [2, 23, 67]. Notably, pioneering works [54, 64] have recently applied ConvNets to MIM pre-training, paving the way for masking modeling methods based on ConvNets. + +Masked Video Modeling. Recently, researchers have embarked on exploring MVM approaches [16, 19, 65, 66] to comprehend unsupervised video representations, exhibiting effectiveness across a multitude of downstream tasks [4, 25, 31, 32, 40-44, 60, 61, 71, 74, 75]. However, current MVM methodologies [20, 56, 58, 59] primarily depend on the isotropic ViT, a visual backbone compatible with mask + +ing modeling. Furthermore, these ViT-based MVM strategies may encounter challenges with inadequate local feature modeling when dealing with dense tasks such as segmentation and detection due to their non-hierarchical structure [35, 37, 62, 70]. To the best of our knowledge, no MVM approaches are explicitly designed for hierarchical network-based architectures, especially ConvNets. + +It is worth noting that while the isotropic ViT is well suited for the implementation of MVM methods [16, 19, 56], this is exceptionally difficult for ConvNets. In this paper, our VideoMAC is committed to developing an efficient video pre-training tool that can equip ConvNets with superior, or at least comparable, performance to ViTs. + +# 3. Approach + +Our VideoMAC builds upon MAE [23] and, pioneeringly, facilitates masked video modeling with a purely convolutional architecture. As depicted in Fig. 3, VideoMAC is designed to reconstruct masked patches, which are randomly and symmetrically masked on frame pairs at a high masking ratio (0.75 as used in this study). The framework achieves such reconstructions by exclusively delivering visible patches to both the online and target encoders. In this section, we describe its main components in detail. + +# 3.1. Preliminaries + +Recently, masked autoencoders (MAE) [23] introduces a novel methodology to the domain of self-supervised visual representation learning. The primary process involves training a model to reconstruct masked areas in images whose contents are partially occluded. The fundamental architecture of a standard MAE model is composed of three core elements: the masking design, an encoder, and a decoder. In mathematical terms, given an image $\pmb{I} \in 3 \times H \times W$ , the MAE model patchifies $\pmb{I}$ into $N$ distinct, non-overlapping patches, represented by $\hat{\pmb{I}} \in N \times (P^2 \times 3)$ , each with a size of $P$ . Masks are randomly applied to a subset of these patches, denoted by the subset $\mathcal{M}$ . + +The encoder $f_{\mathcal{E}}$ processes the visible patches $\mathcal{V}$ , producing latent representations that are represented as $z = f_{\mathcal{E}}(\mathcal{V})$ . Then, the decoder $f_{\mathcal{D}}$ attempts to restore the masked subset $\mathcal{M}$ utilizing these latent representations, generating $O = f_{\mathcal{D}}(z; \mathcal{M})$ , where $O \in N \times (P^2 \times 3)$ depicts the reconstructed patches. The MAE model utilizes the mean-squared error (MSE) to compute the reconstruction loss $\mathcal{L}_{rec}$ , which is solely calculated based on the masked patches and is formulated as $\mathcal{L}_{rec} = \frac{1}{|\mathcal{M}|} \sum_{i \in \mathcal{M}} || \hat{I}_i - O_i ||_2^2$ . + +# 3.2. Video Masked Autoencoder with ConvNets + +Masking. In this work, we randomly select two consecutive frames, $\pmb{I}^{(1)},\pmb{I}^{(2)}\in 3\times H\times W$ and divide them into $N$ nonoverlapping patches, represented as $\hat{I}^{(1)},\hat{I}^{(2)}\in N\times (P^2\times 3)$ + +![](images/a6cfdd4f66713a37f9c4aa63786bda92e18bec84f821946013e1b007172e4bf2.jpg) +Figure 3. An illustration of VideoMAC for ConvNet-based MVM. During pre-training, we mask $75\%$ of symmetric patches from two frames randomly. In our VideoMAC, the MVM of frame pairs is achieved by an online network optimized by gradients $(\square)$ , online loss $\mathcal{L}_o$ and a target network updated by EMA $(\square)$ , target loss $\mathcal{L}_t$ . $\mathcal{L}_c$ is computed as the reconstruction consistency loss between reconstructed patches of frame pairs. + +Regarding the mask configuration, our approach adopts symmetric masking (Table 3d) to guarantee consistency in the positions of the masked areas between frame pairs. Given the consecutive frames, $\mathcal{M}^{(1)}$ and $\mathcal{M}^{(2)}$ constitute the subsets of masked patches, whereas $\mathcal{V}^{(1)}$ and $\mathcal{V}^{(2)}$ correspond to the visible patches. + +It is inherent that images and videos possess a greater level of information redundancy [16, 22] compared to languages. Accordingly, we employ a masking ratio of 0.75 (Table 4f) in our methodology. Subsequent empirical analyses (Sec. 4.2) have demonstrated an improvement in performance with higher masking ratios. Nevertheless, employing excessively high masking ratios may induce the risk of training collapse, thereby leading to sudden and substantial decreases in model performance. + +MAC Encoder. In ViT-based models, the prevention of information leakage from masked regions is comparatively straightforward [64, 70]. These models solely employ the visible patches as the exclusive input to the encoder, thereby allowing ViT-based approaches [16, 20, 56, 65] to amplify the efficiency and efficacy of model pre-training by concentrating dedicatedly on unmasked regions during this phase. In contrast, achieving a similar task with ConvNets poses a more intricate dilemma, necessitating the maintenance of the 2D image structure. There are two naive solutions: (i) the instigation of learnable masked markers at the input stage, and (ii) the transformation of the hierarchical convolutional network into an isotropic structure [36] that mirrors the ViT-style technique (Table 2). Regardless, both strategies may decrease pre-training efficiency and possibly engender inconsistencies between training and testing. + +This paper is at the forefront of proposing an entirely ConvNet-based MVM approach to address this challenge. Drawing inspiration from [54, 64], we utilize sparse convolutional networks to develop a masked autoencoder, which usurps traditional dense convolution. This strategy aims to meet three primary goals: (i) maintaining the structural integrity of the image during feature encoding, (ii) preventing + +mask dissipation in the course of hierarchical convolution, and (iii) markedly amplifying the computational efficiency of MVM. Specifically, our VideoMAC uses a fully convolutional architecture. Conventionally, ConvNets are structured with four stages, where the resolution of features is halved after each stage, representing a patch size of 32. For an input frame size of $224 \times 224$ , the features are sequentially downsampled at every stage to achieve resolutions of $56 \times 56$ , $28 \times 28$ , $14 \times 14$ , and $7 \times 7$ , respectively. Subsequently, masks are produced and recursively upsampled to the highest resolution (i.e., $56 \times 56$ ) in the final stage. + +Current MVMs typically necessitate significant computational resources, posing a challenge for practical applications. Our strategy, which substitutes ConvNets for ViTs, does mitigate computational complexity. Furthermore, studying existing MVM methods, we observe that to capture temporal signals within videos, most methods either utilize multi-frame data input or opt for a siamese encoder structure. We diverge from these widespread strategies and instead adopt a more resource-efficient method, namely the online-target encoder architecture (Fig. 3), commonly applied in self-supervised image representation learning. The target encoder's weights are updated by the online encoder via EMA. This architectural integration into our MVM encoding process results in three-fold advantages: (i) considerable reduction of computational complexity, (ii) joint optimization of frame-to-feature representation by the online and target encoders, and (iii) simplicity in incorporating existing ConvNets into our framework, thus improving the performance of downstream tasks. Consequently, visible patches $\mathcal{V}^{(1)}$ and $\mathcal{V}^{(2)}$ of consecutive frames are processed by the online encoder $f_{\mathcal{E}}^{o}$ and the target encoder $f_{\mathcal{E}}^{t}$ to generate latent representations $z^{(1)} = f_{\mathcal{E}}^{o}(\mathcal{V}^{(1)})$ and $z^{(2)} = f_{\mathcal{E}}^{t}(\mathcal{V}^{(2)})$ , respectively. + +MAC Decoder. Unlike previous methods that utilize hierarchical decoders [54, 55] or transformer decoders[16, 33, 56], our VideoMAC implements a lightweight decoder derived from the ConvNeXt block [36]. This simplifies the design and bolsters the efficiency of the pre-training process. Analogous to our encoder, we construct an online decoder $f_{\mathcal{D}}^{o}$ and a target decoder $f_{\mathcal{D}}^{t}$ , further employing EMA for parameter updates (Fig. 3). Given the latent representations $z^{(1)}$ and $z^{(2)}$ from two successive frames, decoders $f_{\mathcal{D}}^{o}$ and $f_{\mathcal{D}}^{t}$ reconstruct masked patches $\mathcal{M}^{(1)}$ and $\mathcal{M}^{(2)}$ to yield $O^{(1)} = f_{\mathcal{D}}^{o}(z^{(1)};\mathcal{M}^{(1)})$ and $O^{(2)} = f_{\mathcal{D}}^{t}(z^{(2)};\mathcal{M}^{(2)})$ , respectively. In this paper, we employ a single-block design for the decoder, with the decoder feature projection dimension set to 512. Detailed ablation analysis can be found in Sec. 4.2, and Tables 3b and 3c. + +Reconstruction Consistency. In line with the majority of MAE-based methods, we compute the Mean Squared Error (MSE) between the reconstituted and original images to assess the reconstruction loss. We elect to aim for each patch + +of normalized pixels, executing the loss calculation solely for masked patches. As a result, we synchronize the values of normalized pixels within each image patch, applying the loss function exclusively to those patches subject to masking. Significantly, our VideoMAC yields two frames of reconstructed results from both the online and target encoders, necessitating individual loss computations: + +$$ +\mathcal {L} _ {o} = \frac {1}{| \mathcal {M} ^ {(1)} |} \sum_ {i \in \mathcal {M} ^ {(1)}} \| \hat {\boldsymbol {I}} _ {i} ^ {(1)} - \boldsymbol {O} _ {i} ^ {(1)} \| _ {2} ^ {2}, \tag {1} +$$ + +$$ +\mathcal {L} _ {t} = \frac {1}{| \mathcal {M} ^ {(2)} |} \sum_ {i \in \mathcal {M} ^ {(2)}} | | \hat {\boldsymbol {I}} _ {i} ^ {(2)} - \boldsymbol {O} _ {i} ^ {(2)} | | _ {2} ^ {2}, +$$ + +where $\mathcal{L}_o$ and $\mathcal{L}_t$ denote online and target losses correspondingly. Due to the correlation between consecutive video frames, we pursue temporal consistency in addition to the reconstruction loss. Here, we introduce an intuitive assumption: if the pixel deviation between the two raw frames tends to a particular value, denoted as $\epsilon$ , it is inferred that the divergence between the reconstructed frames should also trend towards $\epsilon$ . Based on this assumption, we can estimate the reconstruction consistency between two frames and use it as a proxy for temporal consistency. Hence, the reconstruction consistency loss $\mathcal{L}_c$ is thus expressed as: + +$$ +\mathcal {L} _ {c} = \frac {1}{| \mathcal {M} ^ {(1)} |} \sum_ {i \in \mathcal {M} ^ {(1)}, j \in \mathcal {M} ^ {(2)}} \| \boldsymbol {O} _ {i} ^ {(1)} - \boldsymbol {O} _ {j} ^ {(2)} \| _ {2} ^ {2}, \tag {2} +$$ + +where $|\mathcal{M}^{(1)}|$ can be substituted by $|\mathcal{M}^{(2)}|$ , courtesy of our utilization of symmetric masking. Ultimately, our approach culminates in a total loss expressed as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {o} + \mathcal {L} _ {t} + \gamma \mathcal {L} _ {c}, \tag {3} +$$ + +with $\gamma$ symbolizing the weight factor of the consistency loss (see Table 4e and Fig. 7 for detailed analysis). Accordingly, our approach, VideoMAC, accomplishes the realization of a fully convolutional network MVM. Moreover, the reconstruction consistency loss introduced in this paper provides a promising solution for modeling spatio-temporal data. + +# 3.3. Implementation + +Architecture. Our VideoMAC possesses the adaptability to employ any convolutional network the an encoder. To permit a fair comparison, we choose two representative convolutional networks for our experiments: ResNet (RN) [21] and ConvNeXtv2 (CNXv2) [64]. Although RN50 (26M) and CNXv2-S (50M) serve as the primary experimental models, we further investigate models of diverse sizes (Tables 2, 3a and 5). In this paper, the decoder component involves a readily available ConvNeXt block [36], thereby dispensing with the necessity for intricate network design. During pre-training, we employ an encoder implemented + +
MethodTypeBackboneTraining datasetEpochDAVIS17 [46]VIP [73] mIoUJHMDB [26] PCK@0.1PCK@0.2
J&FmJmFm
Supervised [21]CVPR16SLRN50IN1K (1.28M, -)10066.063.768.439.559.278.3
DenseCL [63]CVPR21CLRN50IN1K (1.28M, -)20061.159.862.632.556.178.0
DINO [6]ICCV21CLViT-S/16IN1K (1.28M, -)80061.860.263.436.245.675.0
ODIN² [24]ECCV22CLRN50IN1K (1.28M, -)100054.154.353.9---
CrOC [52]CVPR23CLViT-S/16IN1K (1.28M, -)30044.743.545.926.1--
MAE [23]CVPR22MIMViT-B/16IN1K (1.28M, -)160053.552.155.028.144.673.4
FCMAE [64]CVPR23MIMCNXv2-BIN1K (1.28M, -)160043.741.945.524.951.774.2
RC-MAE [30]ICLR23MIMViT-S/16IN1K (1.28M, -)160049.248.950.529.743.272.3
SparK [54]ICLR23MIMRN50IN1K (1.28M, -)160055.656.354.932.852.975.2
MAE-ST [16]NeurIPS22MVMViT-L/16K400 (-, 800 hours)160054.655.553.633.244.472.5
VideoMAE [56]NeurIPS22MVMViT-S/16K400 (-, 800 hours)160039.339.738.922.341.067.9
DropMAE [65]CVPR23MVMViT-B/16K400 (-, 800 hours)160053.451.855.031.142.369.2
SiamMAE [20]NeurIPS23MVMViT-S/16K400 (-, 800 hours)200062.060.363.737.347.076.1
Ours
VideoMACMVMRN50YT18 (-, 5.58 hours)10067.264.969.543.657.279.2
VideoMACMVMCNXv2-SYT18 (-, 5.58 hours)10068.465.371.440.458.180.0
+ +SL: Supervised Learning. CL: Contrastive Learning. IN1K: ImageNet1K [11]. K400: Kinetics-400 [27]. YT18: YouTube-VOS 2018 [68]. Table 1. Comparing VideoMAC with previous supervised and self-supervised approaches (SL, CL, MIM, and MVM) on downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]. + +
MVM MethodResolution#Param.FLOPsJ&Fm
Iso. ViT Backbone
MAE-ST [16]14 × 14304M61.6G54.6
VideoMAE [56]14 × 1422M4.6G39.3
DropMAE [65]14 × 1487M17.6G53.4
SiamMAE [20]14 × 1422M4.6G62.0
Iso. ConvNet Backbone
VideoMAC (CNXv2-S†)14 × 1422M4.3G53.8
Hie. ConvNet Backbone
VideoMAC (RN18)7 × 712M1.8G64.7
VideoMAC (RN50)7 × 726M4.1G67.2
VideoMAC (CNXv2-T)7 × 729M4.5G67.5
VideoMAC (CNXv2-S)7 × 750M8.7G68.4
+ +Table 2. Comparisons with isotropic (iso.) and hierarchical (hie.) based MVM methods. ‘†’ denotes our modified ios. version. + +with the MinkowskiEngine sparse convolution library [9]. It's noteworthy that the sparse convolution layer can be seamlessly converted back to a standard convolution layer in downstream tasks without requiring additional processes. + +Pre-training. Different from previous methods that pretrain on the large-scale video dataset K400 [27], VideoMAC pre-trains on YouTube-VOS [68] for 100 epochs, leading to a significant reduction in pre-training time by over 140 times. During VideoMAC pre-training, the standard input comprises two frames with a spatial size of $224 \times 224$ pixels. The frame interval for this study is determined to be 1 (Table 4c), implying consecutive frames are chosen. We employ minimal data augmentation, encompassing only random resized cropping and horizontal flipping. A masking ratio of 0.75 is utilized to pre-train our models. We conduct training using an AdamW optimizer [39] with a batch size of 512 and a learning rate of 2e-3. Cosine decay [38] is applied to adjust the learning rate. + +Downstream Tasks. To gauge the efficacy of VideoMAC at video representation learning, we undertake comprehensive experiments on three video downstream tasks, namely video object segmentation, body part propagation, and hu + +man pose tracking (Table 1). We also investigate the performance of image recognition through video pre-training. Specifically, our backbone is pre-trained on video data and then fine-tuned on ImageNet1K [11] (Table 5). + +# 4. Experiments + +This section begins with assessing VideoMAC's performance in three downstream tasks, comparing it to previous state-of-the-art methods. Subsequently, we conduct extensive ablation studies to elucidate the effectiveness of our proposed approach. The final part of this section explores VideoMAC's ability for image recognition, accompanied by a visualization of the reconstruction results. + +# 4.1. Comparison with Previous methods + +Video Object Segmentation. We evaluate VideoMAC on the popular DAVIS17 [46] dataset, which includes a validation set comprising 30 videos. Our evaluation methodology aligns with [6, 25], and we report the typical evaluation metrics [45], including the mean region similarity $\mathcal{I}_m$ , the mean boundary accuracy $\mathcal{F}_m$ , and their average values $\mathcal{J} \& \mathcal{F}_m$ . As shown in Table 1, our VideoMAC achieves state-of-the-art results for video object segmentation. With a smaller video training dataset and fewer training epochs, both our VideoMAC models (i.e., CNXv2-S and RN50) surpass the SiamMAE [20] by $6.4\%$ and $5.2\%$ , respectively, in terms of $\mathcal{J} \& \mathcal{F}_m$ . It's worth noting that our approach achieves superior performance to supervised method [21] based on MVM, with improvements of $2.4\%$ and $1.4\%$ , respectively. Owing to the hierarchical feature extraction capabilities inherent in ConvNets, our approach demonstrates superior proficiency in spatial information extraction in contrast to ViT-based MVM methods [16, 20, 56, 65]. Concurrently, the introduced reconstruction consistency framework strengthens the acquisition of temporal correlations. + +![](images/cc779ede0ca14ccce373a4492c3fa80742e7aa07fd5fa91a0f0c7998d394cb40.jpg) + +![](images/1da5c536d33763451793841a7c88469e86b8055691da3e7dfb2556c6b86d2643.jpg) + +![](images/59c4dc3fa0f7ce2379bd8ea90f51cc07f7bccf20dd8f78bfa84d403ab57233a2.jpg) + +![](images/beabd69a3a3bbc4e4e0aa2904e73160f8ed59d97015b78f31f3415062bfa6fc7.jpg) +Figure 4. Qualitative results of our VideoMAC (using CNXv2-S) for three video downstream tasks: (a) video object segmentation on DAVIS17 [46]), (b) body part propagation on VIP [73], and (c) human pose tracking on JHMDB [26]. + +![](images/39687ffb3312446beca58fb38acc3a0a9030db6a2d18559414163b847b87400b.jpg) + +![](images/7af476a6a50434a352f3b2d59aa13f561965a7386e62ddafbedcbba801bf9c9e.jpg) + +![](images/e40adcfdd3b59ddcb68fa2a1882d7ff108a65a131d56af2ff1646273560adf21.jpg) + +![](images/cba9e282e25f99cc7c37a6577d419a4b5c039f717feca5d0c5e016557bf3cc8a.jpg) +(a) Video Object Segmentation + +![](images/4ee736ea6f99104038b48ab1192b6046d3869fb26b322f0319cb12e5663480b0.jpg) + +![](images/41c047bd1e556a3366499c82d0d823ed48cb3e8f909efa76b3aa6fa898c27c87.jpg) + +![](images/9e81fb94ed7e1a937521e26acca5cac8e9e95f110629865cc1a8f7471a84a227.jpg) + +![](images/672ce70fa1eb31ffc003eed41139d947c52a9b9b664e3edd3f0edd160e4a0dce.jpg) + +![](images/dc185c900fa52fcb7ce480ce0de5a78e34871720679ab4b29620baff721392ee.jpg) + +![](images/8b7dc9ba9f5052d1bf22ad81136461583dba047d56476410f4f99e36c05c2d11.jpg) + +![](images/030178ddea3e3964ceddaf96234691394f8c0713a51607b50a677aae2d476b2f.jpg) + +![](images/8abbb16234408df96cac0a82d55e1bac421d74db6f0d2c5062325f7f67387ca1.jpg) + +![](images/1ee3b50f026dfec11e4fd7f62a93fc6090ad68689053047d89c3d4e7398967fb.jpg) + +![](images/0dcefeb0c94cb09100575611095050d64ae9e5b73ce8a9003c3e09e289b98367.jpg) + +![](images/780f8b318bd819a397bb8f5bf4f0928acc4a7318b50580525c4fc136bb72ec04.jpg) + +![](images/ffe6d6cc59693db6108be3e0bb934a428b0dbc1c10dd1c0a5b74b6ead7e06acb.jpg) + +![](images/45f1fbf905528e79e1f2dec0ed36d682bbf52dd7a929c2e1da31d677246a1af6.jpg) + +![](images/6aad074add475eff0ac939641540b24868a7327dbea84e0cbac933dd1b733868.jpg) + +![](images/7bedc0740b8045a63ee17f8a8802aa38b82d457d5da3c6357e11d03e5c7685fd.jpg) + +![](images/1fd26f5d405364568ed7ece610516e363157e5817e212676eaa9352486beeee7.jpg) + +![](images/bc4fa73f645e81f89611fb321b357b3d7cb19070b3e5d559bab70fc658d8f69d.jpg) + +![](images/95eeaf3f6df19bf8954565979705f4b938ff7ef3fc51d8eb7806f391f0a33c99.jpg) + +![](images/a2f107ce900efc6209ee6e07d5b9769078dee1ca0eeff81d739e4655103a1989.jpg) +(b) Body Part Propagation + +![](images/71686b941ede7f1925c26f130ac37b92fcc0be581bf05c1a3cfe8f5c0b4e5344.jpg) +(c) Human Pose Tracking + +![](images/3197cc538b71bd80bc20001bf8ab42572a4c5f4aa7d11496527579b7b688ff12.jpg) + +![](images/2074be0313e82ba1108d27f74e24085bf7ade27605fcd5b5dfd6f8cb1a23f5a1.jpg) + +![](images/59ab462eeac5542ee870142213ceb9290a1c1f62a493b38391f06b6653a8d2df.jpg) + +![](images/eb2d86d842d5a702a28d3e4eddf353a49c36ff68b25283b862123797373b8f74.jpg) + +
BackboneJ&FmJmFm
R1864.762.866.5
R5067.264.969.5
CNXv2-T67.564.470.6
CNXv2-S68.465.371.4
+ +(a) Encoder type. VideoMAC is compatible with both classical and modern ConvNets. + +
Dim.J&FmJmFm
12867.964.871.0
25668.265.371.1
51268.465.371.4
102468.165.071.2
+ +(c) Decoder width. A decoder width of 512 delivers the best performance. + +
BlocksJ&FmJmFm
168.465.371.4
268.265.271.2
468.365.471.2
868.165.271.0
+ +(b) Decoder depth. One block achieves competitive performance and efficiency. + +
MaskingJ&FmJmFm
Asym.66.563.469.6
Sym.68.465.371.4
+ +(d) Masking design. Symmetric (sym.) masking outperforms asymmetric (asym.) one. + +![](images/af566f9822535df270ac0e7184972c9b55ad6e8e0256d024df65fbd315924569.jpg) +Figure 5. For masked modeling, dense convolution usually results in the dissipation of mask structures. The deployment of sparse convolution proves to be an effective solution, enabling ConvNet-based MIM / MVM. + +Table 3. VideoMAC encoder and decoder ablation experiments on DAVIS17 [46]. We report ablation results for two of the most emblematic families of ConvNets. Unless specified otherwise, the encoder is CNXv2-S in default. 'Dim' indicates the number of decoder dimensions. Default settings employed in this paper are marked in blue. + +Body Part Propagation. We evaluate VideoMAC on the Video Instance Parsing (VIP) dataset [73], and the quantitative results, measured by mIoU, are detailed in Table 1. This validation set comprises of 50 videos that concentrate on the propagation of 19 body parts, such as arms and legs, thereby necessitating a heightened level of precision in matching as compared to video object segmentation. We adhere to the same parameters as [25], adjusting the video frames to $560 \times 560$ dimensions. In the case of body part propagation, two variants of our VideoMAC demonstrate improvements by $3.1\%$ (CNXv2-S) and $6.3\%$ (RN50), respectively, in comparison with the top-performing SiamMAE [20] in MVM. Significantly, our RN50-based model outperforms CNXv2-S on the VIP dataset [73] due to the higher dimensionality of RN50 features (2048 vs. 768), which benefits complex tasks and improves accuracy in mask propagation. + +Human Pose Tracking. We conduct the human pose tracking task on the validation set of JHMDB [26], which includes 268 videos and entails detecting 15 human keypoints. Conforming to the protocol delineated in [25], we resize video frames to a resolution of $320 \times 320$ , and employ + +the Possibility of Correct Keypoints (PCK) metric [51] for quantitative evaluation. As shown in Table 1, both variants (i.e., RN50 and CNXv2-S) of our VideoMAC outperforms all MVM methods in terms of PCK@0.1 and PCK@0.2. + +Qualitative results. As depicted in Fig. 4, we present the qualitative results of mask propagation achieved by VideoMAC on three downstream tasks. Our VideoMAC yields notable performance across a spectrum of complex videos. + +# 4.2. Ablation Studies + +In this section, we conduct extensive ablation studies to analyze the component design of our VideoMAC. + +Encoder design: hierarchical vs. isotropic. In this ablation study, we scrutinize diverse encoder architectures to our VideoMAC, i.e., hierarchical and isotropic designs. Precisely, we adopted the ViT [14] design principles and restructured the CNXv2-S backbone from having a hierarchical structure into an isotropic one. This implies the creation of an architecture devoid of downsampling layers, ensuring consistent feature resolution (e.g., $14 \times 14$ ) throughout all depths. As shown in Table 2, the performance dif + +
DatasetSizeJ&FmJmFm
YT18 [68]5.58 hours68.465.371.4
K4001 [27]800 hours67.865.270.3
+ +
AugmentationJ&FmJmFm
spatial68.465.371.4
color64.961.868.0
spatial + color66.162.969.3
+ +(b) Data augmentation. Minimal data augmentation achieves the best performance. + +
GapJ&FmJmFm
168.465.371.4
566.463.569.2
1063.260.465.9
+ +(c) Frame gap. Video pairs with one frame gap perform the best results. + +(a) Training Dataset. Our VideoMAC pre-trained on YT18 yields better performance and efficiency compared to K400. + +
\( {\mathcal{L}}_{c} \)\( {\mathcal{L}}_{t} \)\( \mathcal{J}\&{\mathcal{F}}_{m} \)\( {\mathcal{J}}_{m} \)\( {\mathcal{F}}_{m} \)
✘✘51.152.250.0
✘✓60.656.364.9
✓✘63.463.468.7
✓✓68.465.371.4
+ +(d) Model loss. VideoMAC utilizes consistency loss $\mathcal{L}_c$ and target loss $\mathcal{L}_t$ for the best. + +
ConsistencyJ&FmJmFm
γ=0.060.656.364.9
γ=0.164.862.067.6
γ=0.565.863.268.5
γ=1.068.465.371.4
+ +(e) Reconstruction consistency. When $\gamma$ is set to 1.0, the consistency loss works the best. + +
Masking ratioJ&FmJmFm
0.6566.563.869.1
0.7568.465.371.4
0.8568.365.371.3
0.9560.359.760.9
+ +(f) Masking ratio. Our VideoMAC uses a masking ratio of 0.75 for the highest performance. + +Table 4. VideoMAC ablations on data, loss, and masking ratio configuration. We report two official metrics ${\mathcal{I}}_{m}$ and ${\mathcal{F}}_{m}$ on DAVIS17 [46]. + +![](images/b27b70dea0c2e01c788fb99b392cc02eb752a1699c3ea50d8dd8e813dc78a963.jpg) +Figure 6. Analysis of individual loss components after Video-MAC pre-training 100 epochs on YouTube-VOS 2018 [68]. + +![](images/7cd9e0c1d30254190a9cb34d2128c6d67750472d7b0cb8763719acc7dc19a00b.jpg) + +ference between our VideoMAC (iso.) and the ViT-based MVM methods (e.g., VideoMAE [56] and DropMAE [65]) is marginal. By maintaining the hierarchical structure of ConvNets within VideoMAC models (hie.), we improve the performance significantly, even though their feature resolution (i.e., $7 \times 7$ ) is merely half that of the ViT-based models [16, 20, 56, 65]. This can be ascribed to ConvNets' capacity to preserve multi-scale spatial information. + +Encoder backbone: classical and modern. For our study, we incorporate two emblematic ConvNet families, i.e., ResNets [21] and ConvNeXts [36, 64]. As shown in Table 3a, we substitute dense convolution with sparse convolution in the baseline ConvNets. This adaptation seamlessly dovetails into our VideoMAC framework, thereby facilitating a ConvNet-based MVM. In addition, Fig. 5 highlights the influence of employing dense and sparse convolution on masks. It is apparent that sparse convolution effectively safeguards the structural integrity of the mask, thus enabling the integration of MAE pre-training within ConvNets. + +Decoder design: blocks and dimensions. Adhering to a simplistic design principle, our VideoMAC adopts the decoder from the purely convolutional network structure deployed in ConvNeXt [36]. We conduct ablation experiments focusing on the depth and width of the decoder, with the results detailed in Tables 3b and 3c. The decoder exhibits competitive results with a single block and 512 dimensions, striking a trade-off between performance and efficiency. + +Masking design: asymmetric vs. symmetric. In Table 3d, we investigate the impact of two masking strategies, namely + +![](images/fb391479d6808a393acae730f483edd47360905efa464d34859cb3bcf326c937.jpg) + +![](images/fb4cc285ae593a038d53d28330309bfbf1163c8bf56f7226a961c3037f80f4d8.jpg) +Figure 7. Consistency loss in relation to $\gamma$ (left: identical encoder, CNXv2-S) and encoders (right: identical weight factor, $\gamma = 1.0$ ). + +asymmetric (same masking ratio but different locations) and symmetric (same masking ratio and locations), on performance. Compared to asymmetric masking, symmetric masking simplifies the calculation of the reconstruction loss, as it covers the entire masked area. Moreover, it facilitates the inclusion of more reconstruction regions in the loss calculation, resulting in faster model convergence. + +Data setting: dataset, augmentation, and gap. To assess the influence of different video datasets, Table 4a presents the results of pre-training our VideoMAC on both YT18 [68] and $\mathrm{K400^1}$ [27]. The ablation results of VideoMAC on the two datasets differ by just $0.6\%$ , demonstrating the robustness and adaptability of our method to different datasets. For data augmentation, as shown in Table 4b, our VideoMAC achieves the best results using only spatial strategy, i.e., random resized cropping and horizontal flipping. Additionally, we conduct an ablation study on the frame gap, and the optimal results are obtained with consecutive frames (i.e., the gap is 1), as indicated in Table 4c. The contribution of $\mathcal{L}_c$ and $\mathcal{L}_t$ . Previous MVM methods predominantly rely on a single loss, i.e., the reconstruction loss, necessitating multiple frames to be simultaneously processed by the model. In contrast, our study utilizes a more computation-friendly strategy by inputting frame pairs into two separate encoders, followed by parameter updates using EMA. Consequently, we introduce two distinct loss terms: online $\mathcal{L}_o$ and target $\mathcal{L}_t$ reconstruction losses. Furthermore, to integrate temporal information, we + +![](images/393a600959fe9b404075ae11d41b624fb408fcf8dd9c9cd2d9a889f058a3f415.jpg) +Figure 8. Visualization of reconstruction examples using a pre-trained VideoMAC (CNXv2-S) model with 0.75 masking ratio. These frames are randomly selected from the validation set of YT18 [68] and masked $60\%$ of patches. + +compute the loss between the reconstructed results of frame pairs, colloquially known as the reconstruction consistency loss $\mathcal{L}_c$ . As shown in Table 4d, $\mathcal{L}_t$ and $\mathcal{L}_c$ improve $9.5\%$ and $12.3\%$ in terms of $\mathcal{J} \& \mathcal{F}_m$ , respectively, with the full VideoMAC model yielding the best results. The diverse loss components during pre-training are illustrated in Fig. 6. + +The effect of weight factor $\gamma$ . To investigate the impact of the weight factor on reconstruction consistency loss $\mathcal{L}_c$ , we conduct experiments with different factor settings, i.e., $\gamma \in \{0.0, 0.1, 0.5, 1.0\}$ . As illustrated in Table 4e, superior performance is attained when $\gamma$ approximates to 1. When $\gamma$ equals 0, which signifies the absence of the reconstruction consistency loss term, a detrimental effect on performance is noticeable. Simultaneously, we record $\mathcal{L}_c$ for various weight factors during pre-training, as depicted in Fig. 7. It becomes evident that an elevated value of $\gamma$ results in a lesser loss. Ultimately, both the R50 and CNXv2-S models converge to comparable final losses. It is noteworthy that $\mathcal{L}_c$ presents an initial decrease, succeeded by an increase, and ultimately reaches convergence, which aligns with the description in Eq. (2), implying that the ultimate reconstruction results between two frames converge to $\epsilon$ . + +The impact of masking ratio. An ablation analysis of the masking ratio is conducted, and the results for our Video-MAC models (CNXv2-S) are documented in Table 4f. Ablation results portray a performance pattern that first ascends and subsequently descends with increasing masking ratios. Noticeably, the performance of our approach experiences a drastic decline at a 0.95 masking ratio caused by overfitting to the self-supervised pretext task. + +# 4.3. Additional Experiments + +VideoMAC for Image Recognition. As shown in Table 5, utilizing CNXv1-T [36] / CNXv2-T [64] pre-trained by our VideoMAC on YT18 [68], we achieve $82.3\%$ / $82.9\%$ accuracy on IN1K [11] image classification. In comparison to CNXv1-T and CNXv2-T trained from scratch, our approach improves the image recognition results by $0.2\%$ and $0.4\%$ . Remarkably, the performance garnered by VideoMAC (MVM) is proximate to or commensurate with the state-of-the-art ConvNet-based MIM methods [54, 64]. Given the pre-training cost (100 epochs vs. 1600 epochs) and the domain gap (video vs. image), we regard this im + +
MethodBackbonePT epochPT dataFT acc.
Supervised [36]CNXv1-T--82.1
FCMAE [64]CNXv1-T1600IN1K82.3 (+0.2)
Spark [54]CNXv1-T1600IN1K82.4 (+0.3)
VideoMACCNXv1-T100YT1882.3 (+0.2)
Supervised [64]CNXv2-T--82.5
FCMAE [64]CNXv2-T1600IN1K83.0 (+0.5)
VideoMACCNXv2-T100YT1882.9 (+0.4)
+ +Table 5. Comparing VideoMAC with previous ConvNet-based MIM approaches (e.g., Spark [54] and FCMAE [64]) on the image downstream task, i.e., fine-tuning recognition accuracy (acc.) on IN1K [11]. 'PT' and 'FT' indicate pre-training and fine-tuning. + +provement of our VideoMAC as noteworthy and promising compared to training from scratch and pre-training by MIM. Visualization. We visualize reconstruction results from the YT18 [68] validation set to analyze performance in MVM pre-training. In Fig. 8, VideoMAC makes reasonable predictions about the masked regions, e.g., color and contour. + +# 5. Conclusion + +In this study, we present VideoMAC, an MVM pre-training framework constructed fully using ConvNets. Our masked modeling for frame pairs is executed by introducing an online network (encoder and decoder) optimized by gradients, as well as the target network that updates parameters via EMA. Concurrently, for reconstruction results derived from both online and target networks, our proposed loss of reconstruction consistency between frame pairs paves the way for temporal modeling. Our VideoMAC empowers a multitude of ConvNets to reap the performance gains for downstream tasks (e.g., video but also image) by MVM pre-training. + +Limitation and Future Work. Existing ConvNet-based architectures typically use a fixed downsampling ratio, which in turn leads to a relatively fixed patch size. For future work, we will investigate ConvNets with adjustable patch size, allowing for higher feature resolution like ViT. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China (No. 62102182, 62202227, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20220934, BK20220938, BK20220936), China Postdoctoral Science Foundation (No. 2022M721626, 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No.30923010303). + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022. 1 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 2 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1, 2 +[4] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 2 +[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912–9924, 2020. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2 +[9] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 5 +[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 1, 2 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 5, 8 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, , and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 1, 2 +[13] Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language understanding and generation. In NeurIPS, 2019. 2 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 6 +[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue + +Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 1 +[16] Christoph Feichtenhofer, haoqi fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, pages 35946-35958, 2022. 1, 2, 3, 4, 5, 7 +[17] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. NeurIPS, 35:35632-35644, 2022. 2 +[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 2 +[19] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. In ICLR, 2023. 2, 3 +[20] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 4, 5, 7 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5 +[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In ECCV, pages 123-143, 2022. 5 +[25] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, pages 19545-19560, 2020. 2, 5, 6 +[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, pages 3192-3199, 2013. 1, 2, 5, 6 +[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5, 7 +[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2 +[29] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2 +[30] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. In ICLR, 2023. 1, 5 + +[31] Lulei Li, Tianfei Zhou, Wenguan Wang, Lu Yang, Jianwu Li, and Yi Yang. Locality-aware inter-and intra-video reconstruction for self-supervised correspondence learning. In CVPR, pages 8719-8730, 2022. 2 +[32] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In CVPR, pages 18706-18716, 2023. 2 +[33] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In CVPR, pages 6252-6261, 2023. 4 +[34] Yuan Liu, Songyang Zhang, Jiacheng Chen, Zhaohui Yu, Kai Chen, and Dahua Lin. Improving pixel-based mim by reducing wasted modeling capability. In ICCV, pages 5361-5372, 2023. 1 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3 +[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 2, 3, 4, 7, 8 +[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3 +[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5 +[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 +[40] Xiankai Lu, Wenguan Wang, Martin Danelljan, Tianfei Zhou, Jianbing Shen, and Luc Van Gool. Video object segmentation with episodic graph memory networks. In ECCV, pages 661-679, 2020. 2 +[41] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In CVPR, pages 8960-8970, 2020. +[42] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, pages 596-613, 2022. +[43] Gensheng Pei, Fumin Shen, Yazhou Yao, Tao Chen, Xian-Sheng Hua, and Heng-Tao Shen. Hierarchical graph pattern understanding for zero-shot video object segmentation. TIP, 32:5909-5920, 2023. +[44] Gensheng Pei, Yazhou Yao, Fumin Shen, Dan Huang, Xingguo Huang, and Heng-Tao Shen. Hierarchical co-attention propagation network for zero-shot video object segmentation. TIP, pages 2348–2359, 2023. 2 +[45] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5 +[46] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2, 5, 6, 7 + +[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2 +[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Malik, Yanghao Li, and Christoph Feichtenhofer. Hiera: A hierarchical vision transformer without the bells-and-whistles. ICML, 2023. 2 +[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 2 +[51] Jie Song, Limin Wang, Luc Van Gool, and Otmar Hilliges. Thin-slicing network: A deep structured model for pose estimation in videos. In CVPR, pages 4220-4229, 2017. 6 +[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In CVPR, pages 7000–7009, 2023. 5 +[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 2 +[54] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. In ICLR, 2023. 2, 3, 4, 5, 8 +[55] Yunjie Tian, Lingxi Xie, Zhaozhi Wang, Longhui Wei, Xiaopeng Zhang, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Integrally pre-trained transformer pyramid networks. In CVPR, pages 18610-18620, 2023. 4 +[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 1, 2, 3, 4, 5, 7 +[57] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 2 +[58] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2 +[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2 +[60] Wenguan Wang, Xiankai Lu, Jianbing Shen, David J Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, pages 9236-9245, 2019. 2 +[61] Wenguan Wang, Hongmei Song, Shuyang Zhao, Jianbing Shen, Sanyuan Zhao, Steven CH Hoi, and Haibin Ling. + +Learning unsupervised video object segmentation through visual attention. In CVPR, pages 3064-3074, 2019. 2 +[62] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 2, 3 +[63] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 5 +[64] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In CVPR, pages 16133-16142, 2023. 1, 2, 3, 4, 5, 7, 8 +[65] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In CVPR, pages 14561–14571, 2023. 1, 2, 3, 5, 7 +[66] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, pages 13879-13889, 2023. 2 +[67] Z. Xie, Z. Zhang, Y. Cao, Y. Lin, J. Bao, and Z. Yao. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 1, 2 +[68] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, pages 585-601, 2018. 5, 7, 8 +[69] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, pages 2623-2632, 2021. 2 +[70] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 2, 3 +[71] Yurong Zhang, Lielei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In CVPR, pages 2246-2256, 2023. 2 +[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 2 +[73] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In ACMMM, pages 1527-1535, 2018. 1, 2, 5, 6 +[74] Tianfei Zhou, Fatih Porikli, David J Crandall, Luc Van Gool, and Wenguan Wang. A survey on deep learning technique for video segmentation. TPAMI, 45(6):7099-7122, 2023. 2 +[75] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, pages 408-417, 2017. 2 \ No newline at end of file diff --git a/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/images.zip b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae7079a5e7bd0a39c0ec1ce425fc8ea2cb202017 --- /dev/null +++ b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15dcca5a515b5de7396423f11b7585021db58d68535eab994d6545081c94f49d +size 754852 diff --git a/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/layout.json b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f3bb103718cd068ad7280e60d15fa3aa4a9dbe7 --- /dev/null +++ b/2024/VideoMAC_ Video Masked Autoencoders Meet ConvNets/layout.json @@ -0,0 +1,11792 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 474, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 474, + 119 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 474, + 119 + ], + "type": "text", + "content": "VideoMAC: Video Masked Autoencoders Meet ConvNets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 143, + 485, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 143, + 485, + 186 + ], + "spans": [ + { + "bbox": [ + 111, + 143, + 485, + 186 + ], + "type": "text", + "content": "Gensheng Pei\\*, Tao Chen\\*, Xiruo Jiang, Huafeng Liu, Zeren Sun,\\* Yazhou Yao\\* \nSchool of Computer Science and Engineering, Nanjing University of Science and Technology \nhttps://github.com/NUST-Machine-Intelligence-Laboratory/VideoMAC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 539 + ], + "type": "text", + "content": "Recently, the advancement of self-supervised learning techniques, like masked autoencoders (MAE), has greatly influenced visual representation learning for images and videos. Nevertheless, it is worth noting that the predominant approaches in existing masked image / video modeling rely excessively on resource-intensive vision transformers (ViTs) as the feature encoder. In this paper, we propose a new approach termed as VideoMAC, which combines video masked autoencoders with resource-friendly ConvNets. Specifically, VideoMAC employs symmetric masking on randomly sampled pairs of video frames. To prevent the issue of mask pattern dissipation, we utilize ConvNets which are implemented with sparse convolutional operators as encoders. Simultaneously, we present a simple yet effective masked video modeling (MVM) approach, a dual encoder architecture comprising an online encoder and an exponential moving average target encoder, aimed to facilitate inter-frame reconstruction consistency in videos. Additionally, we demonstrate that VideoMAC, empowering classical (ResNet) / modern (ConvNeXt) convolutional encoders to harness the benefits of MVM, outperforms ViT-based approaches on downstream tasks, including video object segmentation (+5.2% / 6.4% J&F), body part propagation (+6.3% / 3.1% mIoU), and human pose tracking (+10.2% / 11.1% PCK@0.1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 585, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 682 + ], + "type": "text", + "content": "Spurred by the success of masked language modeling in natural language processing [3, 10, 12], transformer-based visual models, notably vision transformers (ViT) [14] adopt masked image modeling (MIM) are attracting attention [1, 2, 15, 23, 34, 67]. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 211, + 545, + 352 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 352 + ], + "type": "image", + "image_path": "ca4cde1d18bf259bdb656eb6c157807b92cbcd0f569194a2f8bd19d1734e5573.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 355, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 467 + ], + "type": "text", + "content": "Figure 1. State-of-the-art MAE architectures (e.g., MAE-ST [16] and SiamMAE [20]) in masked video modeling commonly employ ViT-based encoders. We propose VideoMAC, a new video masked autoencoder built using pure ConvNets. In this study, we model VideoMAC with two of the most emblematic families of ConvNets, namely ConvNeXtv2 [64] and ResNet [21]. Notably, VideoMAC exhibits superior performance on a range of downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26], compared to ViT-based methods [16, 20]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 592 + ], + "type": "text", + "content": "With the successful application of MIM, some pioneering works [16, 56] have shifted the focus of video representation learning to masked video modeling (MVM), yielding promising results in downstream video tasks. However, features learned by MVM are designed explicitly for pixel reconstruction tasks. They perform well when fine-tuned for downstream video tasks but exhibit subpar performance without specific fine-tuning [20]. Their performance is even inferior to MIM methods [23, 30], and it lags significantly behind supervised methods [21, 36]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "We strive to discern the factors contributing to this issue. To the best of our knowledge, most current MVM methods are based on the isotropic design of ViT [16, 20, 56, 65]. The reason is that early MIM / MVM methods are explicitly developed for ViT, as they allow the use of only visible patches as the sole input to the encoder, which aligns well with the ViT architecture. For instance, the representative approach, masked autoencoders (MAE) [23], enhances the efficiency and effectiveness of model pre-training by focusing predominantly on the unmasked regions. However," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22733" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 78, + 119, + 129 + ], + "blocks": [ + { + "bbox": [ + 56, + 100, + 67, + 110 + ], + "lines": [ + { + "bbox": [ + 56, + 100, + 67, + 110 + ], + "spans": [ + { + "bbox": [ + 56, + 100, + 67, + 110 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 70, + 104, + 78 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 104, + 78 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 104, + 78 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 78, + 119, + 129 + ], + "lines": [ + { + "bbox": [ + 69, + 78, + 119, + 129 + ], + "spans": [ + { + "bbox": [ + 69, + 78, + 119, + 129 + ], + "type": "image", + "image_path": "9dd50d2de52ff30ad19264f6594a7ddac788c602a2a31c43d24a041d7ea95ab6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 122, + 79, + 172, + 129 + ], + "blocks": [ + { + "bbox": [ + 133, + 70, + 160, + 77 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 160, + 77 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 160, + 77 + ], + "type": "text", + "content": "Masked" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 122, + 79, + 172, + 129 + ], + "lines": [ + { + "bbox": [ + 122, + 79, + 172, + 129 + ], + "spans": [ + { + "bbox": [ + 122, + 79, + 172, + 129 + ], + "type": "image", + "image_path": "b906e3d8a645712c24a9cdbb5f24f335869a51f2ce49c52c500e828aa989f48a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 79, + 224, + 129 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 223, + 78 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 223, + 78 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 223, + 78 + ], + "type": "text", + "content": "Reconstructed" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 79, + 224, + 129 + ], + "lines": [ + { + "bbox": [ + 175, + 79, + 224, + 129 + ], + "spans": [ + { + "bbox": [ + 175, + 79, + 224, + 129 + ], + "type": "image", + "image_path": "5d91cf721317ff40cad8b85c77c3131f313349c43d3546d585cb74dd3945a40d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 227, + 79, + 277, + 129 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 266, + 78 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 266, + 78 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 266, + 78 + ], + "type": "text", + "content": "Heatmap" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 79, + 277, + 129 + ], + "lines": [ + { + "bbox": [ + 227, + 79, + 277, + 129 + ], + "spans": [ + { + "bbox": [ + 227, + 79, + 277, + 129 + ], + "type": "image", + "image_path": "d6dae044ebd5e2836c4c184376f9ae28cec74de1da4a4a676d23fbf699acd5cb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 69, + 132, + 119, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 152, + 67, + 161 + ], + "lines": [ + { + "bbox": [ + 56, + 152, + 67, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 67, + 161 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 132, + 119, + 182 + ], + "lines": [ + { + "bbox": [ + 69, + 132, + 119, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 132, + 119, + 182 + ], + "type": "image", + "image_path": "b434930b33d94696817c659d12af41a5de19fd5199265af242645b34307bea6f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 229 + ], + "type": "text", + "content": "Figure 2. Visualization of the heatmap during the reconstruction of masked patches in a random frame pair. It is evident that our approach highlights similar regions for both (a) past and (b) current frames, proficiently reconstructing colors and contours." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 121, + 132, + 172, + 182 + ], + "blocks": [ + { + "bbox": [ + 121, + 132, + 172, + 182 + ], + "lines": [ + { + "bbox": [ + 121, + 132, + 172, + 182 + ], + "spans": [ + { + "bbox": [ + 121, + 132, + 172, + 182 + ], + "type": "image", + "image_path": "472226c5798a3a7ed092aec83705bde8c203ebbce9f7ef47c68ffa4fab33049a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 175, + 132, + 224, + 182 + ], + "blocks": [ + { + "bbox": [ + 175, + 132, + 224, + 182 + ], + "lines": [ + { + "bbox": [ + 175, + 132, + 224, + 182 + ], + "spans": [ + { + "bbox": [ + 175, + 132, + 224, + 182 + ], + "type": "image", + "image_path": "5c1999b4597d7fa62c72dbb2107d52fcf31a4f4ad78448beed7b1bbed1622bd3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 227, + 132, + 277, + 182 + ], + "blocks": [ + { + "bbox": [ + 227, + 132, + 277, + 182 + ], + "lines": [ + { + "bbox": [ + 227, + 132, + 277, + 182 + ], + "spans": [ + { + "bbox": [ + 227, + 132, + 277, + 182 + ], + "type": "image", + "image_path": "f91653fc48b39211fc0570d2fdaa2375e0060af10dd9bc2808bba78420ae9e3e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 235, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 286, + 413 + ], + "type": "text", + "content": "previous MVM approaches [16, 56], confined by the plain vision transformer design devoid of a hierarchical structure tend to result in inadequate spatial modeling capabilities for local features [17, 35, 49, 62, 70]. Such capabilities are crucial for dense downstream tasks like segmentation and detection. Moreover, the uniform reliance on ViT in existing MVM methods has resulted in significant computational resource consumption [16, 56, 58]. One might first consider ConvNets [21, 28, 29, 36, 50, 53], characterized by their low resource consumption and built-in hierarchical structure, as an ideal solution to tackle the aforementioned issues. Nevertheless, ConvNets appears understudied in MVM. Consequently, it raises the question: can ConvNets, endowed with a pyramid vision architecture, guide MVM towards the era of hierarchical pre-training?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 415, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 286, + 713 + ], + "type": "text", + "content": "To address this concern, we propose a simple yet potent MVM approach, referred to as Video Masked Autoencoders with ConvNets (VideoMAC). Uniquely, our method represents an initial endeavor to supplement video pre-training by integrating hierarchical ConvNets. However, applying ConvNets to MVM proves more complex than utilizing ViTs. Typical ConvNets might be incompatible with MVM due to the employment of dense sliding windows with overlap, resulting in information leakage within masked regions. To efficiently incorporate mask modeling into ConvNets, we opt for sparse convolution over dense convolution to restructure the original ConvNets, an approach validated in recent studies [54, 64] to effectively maintain the structure of the masked region when utilizing ConvNets. Another hurdle for MVM is the effective incorporation of temporal data. Current MVM pre-training methods [16, 56] involving multi-frame inputs are primarily used in downstream resource-intensive tasks such as video action recognition. Although methods using two-frame inputs for pre-training lower resource usage [20, 65], they typically employ the Siamese architecture, necessitating robust skills for designing reconstruction targets. Importantly, all aforementioned methods utilize computationally intensive, non-hierarchical ViT [14] as their encoders. To this end, we utilize pairs of frames as inputs and introduce the reconstruction consis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "tency loss, a streamlined extension of MAE to accommodate spatio-temporal data. We formulate online networks through gradient optimization and establish target networks by updating parameters via an exponential moving average (EMA). This allows for individual recovery of masked frame pairs and computation of the reconstruction consistency loss between them. As illustrated in Fig. 2, our approach can focus on similar regions between two frames and reconstruct the masked portions efficiently." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 180, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 262 + ], + "type": "text", + "content": "Notably, our approach transcends previous state-of-the-art techniques in three downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]. Moreover, we present promising results utilizing VideoMAC in fine-tuning for image recognition on ImageNet1K [11]. To summarize, our main contributions of this work include:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 264, + 545, + 346 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 317, + 264, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 545, + 299 + ], + "type": "text", + "content": "- We propose a simple yet potent solution of masked video autoencoders, marking the initial endeavor to successfully implement MVM using pure ConvNets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 300, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 545, + 323 + ], + "type": "text", + "content": "- We introduce the reconstruction consistency loss to enable elegant modeling of spatio-temporal data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 324, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 324, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 324, + 545, + 346 + ], + "type": "text", + "content": "- Our approach significantly outperforms existing ViT-based MVM methods in various downstream tasks." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 348, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 407 + ], + "type": "text", + "content": "While the focus of the computer vision community is increasingly shifting towards vision transformers, we trust that VideoMAC will rekindle interest in exploring ConvNet-based MVM, thereby catalyzing the discovery of new potentials across a variety of video tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 418, + 391, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 391, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 391, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "content": "Masked Image Modeling. Self-supervised learning (SSL) has been rapidly developing in computer vision, demonstrating substantial potential for deriving valuable representations from extensive, unlabeled datasets. A key advantage of SSL [2, 5-8, 18, 22, 48, 57, 69, 72] is its independence from costly human annotations, thereby providing robust prior knowledge across various visual tasks. The success of masked language modeling in NLP [3, 10, 12, 13, 47] has catalyzed a rising interest in masked image modeling (MIM) for visual pre-training. This strategy involves masking or corrupting a portion of information within an image and then requiring the model to recover the masked portion, encouraging the model to learn useful image representations [2, 23, 67]. Notably, pioneering works [54, 64] have recently applied ConvNets to MIM pre-training, paving the way for masking modeling methods based on ConvNets." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Masked Video Modeling. Recently, researchers have embarked on exploring MVM approaches [16, 19, 65, 66] to comprehend unsupervised video representations, exhibiting effectiveness across a multitude of downstream tasks [4, 25, 31, 32, 40-44, 60, 61, 71, 74, 75]. However, current MVM methodologies [20, 56, 58, 59] primarily depend on the isotropic ViT, a visual backbone compatible with mask" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22734" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "ing modeling. Furthermore, these ViT-based MVM strategies may encounter challenges with inadequate local feature modeling when dealing with dense tasks such as segmentation and detection due to their non-hierarchical structure [35, 37, 62, 70]. To the best of our knowledge, no MVM approaches are explicitly designed for hierarchical network-based architectures, especially ConvNets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": "It is worth noting that while the isotropic ViT is well suited for the implementation of MVM methods [16, 19, 56], this is exceptionally difficult for ConvNets. In this paper, our VideoMAC is committed to developing an efficient video pre-training tool that can equip ConvNets with superior, or at least comparable, performance to ViTs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 239, + 113, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 113, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 113, + 253 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 260, + 286, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 286, + 367 + ], + "type": "text", + "content": "Our VideoMAC builds upon MAE [23] and, pioneeringly, facilitates masked video modeling with a purely convolutional architecture. As depicted in Fig. 3, VideoMAC is designed to reconstruct masked patches, which are randomly and symmetrically masked on frame pairs at a high masking ratio (0.75 as used in this study). The framework achieves such reconstructions by exclusively delivering visible patches to both the online and target encoders. In this section, we describe its main components in detail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 133, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 133, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 133, + 388 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": "Recently, masked autoencoders (MAE) [23] introduces a novel methodology to the domain of self-supervised visual representation learning. The primary process involves training a model to reconstruct masked areas in images whose contents are partially occluded. The fundamental architecture of a standard MAE model is composed of three core elements: the masking design, an encoder, and a decoder. In mathematical terms, given an image " + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\pmb{I} \\in 3 \\times H \\times W" + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": ", the MAE model patchifies " + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": " distinct, non-overlapping patches, represented by " + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{I}} \\in N \\times (P^2 \\times 3)" + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": ", each with a size of " + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": ". Masks are randomly applied to a subset of these patches, denoted by the subset " + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 395, + 286, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": " processes the visible patches " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": ", producing latent representations that are represented as " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "z = f_{\\mathcal{E}}(\\mathcal{V})" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": ". Then, the decoder " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": " attempts to restore the masked subset " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": " utilizing these latent representations, generating " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "O = f_{\\mathcal{D}}(z; \\mathcal{M})" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "O \\in N \\times (P^2 \\times 3)" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": " depicts the reconstructed patches. The MAE model utilizes the mean-squared error (MSE) to compute the reconstruction loss " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": ", which is solely calculated based on the masked patches and is formulated as " + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec} = \\frac{1}{|\\mathcal{M}|} \\sum_{i \\in \\mathcal{M}} || \\hat{I}_i - O_i ||_2^2" + }, + { + "bbox": [ + 46, + 540, + 286, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 270, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 270, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 270, + 670 + ], + "type": "text", + "content": "3.2. Video Masked Autoencoder with ConvNets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Masking. In this work, we randomly select two consecutive frames, " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{I}^{(1)},\\pmb{I}^{(2)}\\in 3\\times H\\times W" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " and divide them into " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " nonoverlapping patches, represented as " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}^{(1)},\\hat{I}^{(2)}\\in N\\times (P^2\\times 3)" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 545, + 173 + ], + "type": "image", + "image_path": "a6cfdd4f66713a37f9c4aa63786bda92e18bec84f821946013e1b007172e4bf2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": "Figure 3. An illustration of VideoMAC for ConvNet-based MVM. During pre-training, we mask " + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": " of symmetric patches from two frames randomly. In our VideoMAC, the MVM of frame pairs is achieved by an online network optimized by gradients " + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": ", online loss " + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_o" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": " and a target network updated by EMA " + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": ", target loss " + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 174, + 545, + 251 + ], + "type": "text", + "content": " is computed as the reconstruction consistency loss between reconstructed patches of frame pairs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": "Regarding the mask configuration, our approach adopts symmetric masking (Table 3d) to guarantee consistency in the positions of the masked areas between frame pairs. Given the consecutive frames, " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(1)}" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(2)}" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": " constitute the subsets of masked patches, whereas " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(1)}" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(2)}" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": " correspond to the visible patches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 331, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 437 + ], + "type": "text", + "content": "It is inherent that images and videos possess a greater level of information redundancy [16, 22] compared to languages. Accordingly, we employ a masking ratio of 0.75 (Table 4f) in our methodology. Subsequent empirical analyses (Sec. 4.2) have demonstrated an improvement in performance with higher masking ratios. Nevertheless, employing excessively high masking ratios may induce the risk of training collapse, thereby leading to sudden and substantial decreases in model performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 438, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 629 + ], + "type": "text", + "content": "MAC Encoder. In ViT-based models, the prevention of information leakage from masked regions is comparatively straightforward [64, 70]. These models solely employ the visible patches as the exclusive input to the encoder, thereby allowing ViT-based approaches [16, 20, 56, 65] to amplify the efficiency and efficacy of model pre-training by concentrating dedicatedly on unmasked regions during this phase. In contrast, achieving a similar task with ConvNets poses a more intricate dilemma, necessitating the maintenance of the 2D image structure. There are two naive solutions: (i) the instigation of learnable masked markers at the input stage, and (ii) the transformation of the hierarchical convolutional network into an isotropic structure [36] that mirrors the ViT-style technique (Table 2). Regardless, both strategies may decrease pre-training efficiency and possibly engender inconsistencies between training and testing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "This paper is at the forefront of proposing an entirely ConvNet-based MVM approach to address this challenge. Drawing inspiration from [54, 64], we utilize sparse convolutional networks to develop a masked autoencoder, which usurps traditional dense convolution. This strategy aims to meet three primary goals: (i) maintaining the structural integrity of the image during feature encoding, (ii) preventing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22735" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "mask dissipation in the course of hierarchical convolution, and (iii) markedly amplifying the computational efficiency of MVM. Specifically, our VideoMAC uses a fully convolutional architecture. Conventionally, ConvNets are structured with four stages, where the resolution of features is halved after each stage, representing a patch size of 32. For an input frame size of " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", the features are sequentially downsampled at every stage to achieve resolutions of " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "56 \\times 56" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", respectively. Subsequently, masks are produced and recursively upsampled to the highest resolution (i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "56 \\times 56" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ") in the final stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": "Current MVMs typically necessitate significant computational resources, posing a challenge for practical applications. Our strategy, which substitutes ConvNets for ViTs, does mitigate computational complexity. Furthermore, studying existing MVM methods, we observe that to capture temporal signals within videos, most methods either utilize multi-frame data input or opt for a siamese encoder structure. We diverge from these widespread strategies and instead adopt a more resource-efficient method, namely the online-target encoder architecture (Fig. 3), commonly applied in self-supervised image representation learning. The target encoder's weights are updated by the online encoder via EMA. This architectural integration into our MVM encoding process results in three-fold advantages: (i) considerable reduction of computational complexity, (ii) joint optimization of frame-to-feature representation by the online and target encoders, and (iii) simplicity in incorporating existing ConvNets into our framework, thus improving the performance of downstream tasks. Consequently, visible patches " + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(1)}" + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(2)}" + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": " of consecutive frames are processed by the online encoder " + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{E}}^{o}" + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": " and the target encoder " + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{E}}^{t}" + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": " to generate latent representations " + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "inline_equation", + "content": "z^{(1)} = f_{\\mathcal{E}}^{o}(\\mathcal{V}^{(1)})" + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "inline_equation", + "content": "z^{(2)} = f_{\\mathcal{E}}^{t}(\\mathcal{V}^{(2)})" + }, + { + "bbox": [ + 46, + 206, + 289, + 482 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": "MAC Decoder. Unlike previous methods that utilize hierarchical decoders [54, 55] or transformer decoders[16, 33, 56], our VideoMAC implements a lightweight decoder derived from the ConvNeXt block [36]. This simplifies the design and bolsters the efficiency of the pre-training process. Analogous to our encoder, we construct an online decoder " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}^{o}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " and a target decoder " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}^{t}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": ", further employing EMA for parameter updates (Fig. 3). Given the latent representations " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "z^{(1)}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "z^{(2)}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " from two successive frames, decoders " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}^{o}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{D}}^{t}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " reconstruct masked patches " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(1)}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(2)}" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " to yield " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "O^{(1)} = f_{\\mathcal{D}}^{o}(z^{(1)};\\mathcal{M}^{(1)})" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "inline_equation", + "content": "O^{(2)} = f_{\\mathcal{D}}^{t}(z^{(2)};\\mathcal{M}^{(2)})" + }, + { + "bbox": [ + 46, + 483, + 289, + 663 + ], + "type": "text", + "content": ", respectively. In this paper, we employ a single-block design for the decoder, with the decoder feature projection dimension set to 512. Detailed ablation analysis can be found in Sec. 4.2, and Tables 3b and 3c." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Reconstruction Consistency. In line with the majority of MAE-based methods, we compute the Mean Squared Error (MSE) between the reconstituted and original images to assess the reconstruction loss. We elect to aim for each patch" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "of normalized pixels, executing the loss calculation solely for masked patches. As a result, we synchronize the values of normalized pixels within each image patch, applying the loss function exclusively to those patches subject to masking. Significantly, our VideoMAC yields two frames of reconstructed results from both the online and target encoders, necessitating individual loss computations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 347, + 164, + 547, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 164, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 347, + 164, + 547, + 201 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\frac {1}{| \\mathcal {M} ^ {(1)} |} \\sum_ {i \\in \\mathcal {M} ^ {(1)}} \\| \\hat {\\boldsymbol {I}} _ {i} ^ {(1)} - \\boldsymbol {O} _ {i} ^ {(1)} \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "2625a02546396ae45ece3c8bf3f10dcf2bb21b741422ab7eac536d8294e97843.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 197, + 504, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 197, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 350, + 197, + 504, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\frac {1}{| \\mathcal {M} ^ {(2)} |} \\sum_ {i \\in \\mathcal {M} ^ {(2)}} | | \\hat {\\boldsymbol {I}} _ {i} ^ {(2)} - \\boldsymbol {O} _ {i} ^ {(2)} | | _ {2} ^ {2},", + "image_path": "2ea6557239e30756fcee803182978af94c9a86586c4b8618673a4af1a53565db.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_o" + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "content": " denote online and target losses correspondingly. Due to the correlation between consecutive video frames, we pursue temporal consistency in addition to the reconstruction loss. Here, we introduce an intuitive assumption: if the pixel deviation between the two raw frames tends to a particular value, denoted as " + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "content": ", it is inferred that the divergence between the reconstructed frames should also trend towards " + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "content": ". Based on this assumption, we can estimate the reconstruction consistency between two frames and use it as a proxy for temporal consistency. Hence, the reconstruction consistency loss " + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 234, + 547, + 368 + ], + "type": "text", + "content": " is thus expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 330, + 374, + 545, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 374, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 330, + 374, + 545, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{| \\mathcal {M} ^ {(1)} |} \\sum_ {i \\in \\mathcal {M} ^ {(1)}, j \\in \\mathcal {M} ^ {(2)}} \\| \\boldsymbol {O} _ {i} ^ {(1)} - \\boldsymbol {O} _ {j} ^ {(2)} \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "e97f246f1d7b0d1083c8919473d44c5af4789e3029901fc1d9ff995ac4243b35.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "type": "inline_equation", + "content": "|\\mathcal{M}^{(1)}|" + }, + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "type": "text", + "content": " can be substituted by " + }, + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "type": "inline_equation", + "content": "|\\mathcal{M}^{(2)}|" + }, + { + "bbox": [ + 304, + 415, + 547, + 452 + ], + "type": "text", + "content": ", courtesy of our utilization of symmetric masking. Ultimately, our approach culminates in a total loss expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 461, + 545, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 461, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 372, + 461, + 545, + 474 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {o} + \\mathcal {L} _ {t} + \\gamma \\mathcal {L} _ {c}, \\tag {3}", + "image_path": "0149c0af423c0c1d0db02cea7afd3112af4e990ddbf410074fa563cba994e953.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 557 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 483, + 547, + 557 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 483, + 547, + 557 + ], + "type": "text", + "content": " symbolizing the weight factor of the consistency loss (see Table 4e and Fig. 7 for detailed analysis). Accordingly, our approach, VideoMAC, accomplishes the realization of a fully convolutional network MVM. Moreover, the reconstruction consistency loss introduced in this paper provides a promising solution for modeling spatio-temporal data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 563, + 404, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 404, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 404, + 576 + ], + "type": "text", + "content": "3.3. Implementation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Architecture. Our VideoMAC possesses the adaptability to employ any convolutional network the an encoder. To permit a fair comparison, we choose two representative convolutional networks for our experiments: ResNet (RN) [21] and ConvNeXtv2 (CNXv2) [64]. Although RN50 (26M) and CNXv2-S (50M) serve as the primary experimental models, we further investigate models of diverse sizes (Tables 2, 3a and 5). In this paper, the decoder component involves a readily available ConvNeXt block [36], thereby dispensing with the necessity for intricate network design. During pre-training, we employ an encoder implemented" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22736" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 251 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 251 + ], + "type": "table", + "html": "
MethodTypeBackboneTraining datasetEpochDAVIS17 [46]VIP [73] mIoUJHMDB [26] PCK@0.1PCK@0.2
J&FmJmFm
Supervised [21]CVPR16SLRN50IN1K (1.28M, -)10066.063.768.439.559.278.3
DenseCL [63]CVPR21CLRN50IN1K (1.28M, -)20061.159.862.632.556.178.0
DINO [6]ICCV21CLViT-S/16IN1K (1.28M, -)80061.860.263.436.245.675.0
ODIN² [24]ECCV22CLRN50IN1K (1.28M, -)100054.154.353.9---
CrOC [52]CVPR23CLViT-S/16IN1K (1.28M, -)30044.743.545.926.1--
MAE [23]CVPR22MIMViT-B/16IN1K (1.28M, -)160053.552.155.028.144.673.4
FCMAE [64]CVPR23MIMCNXv2-BIN1K (1.28M, -)160043.741.945.524.951.774.2
RC-MAE [30]ICLR23MIMViT-S/16IN1K (1.28M, -)160049.248.950.529.743.272.3
SparK [54]ICLR23MIMRN50IN1K (1.28M, -)160055.656.354.932.852.975.2
MAE-ST [16]NeurIPS22MVMViT-L/16K400 (-, 800 hours)160054.655.553.633.244.472.5
VideoMAE [56]NeurIPS22MVMViT-S/16K400 (-, 800 hours)160039.339.738.922.341.067.9
DropMAE [65]CVPR23MVMViT-B/16K400 (-, 800 hours)160053.451.855.031.142.369.2
SiamMAE [20]NeurIPS23MVMViT-S/16K400 (-, 800 hours)200062.060.363.737.347.076.1
Ours
VideoMACMVMRN50YT18 (-, 5.58 hours)10067.264.969.543.657.279.2
VideoMACMVMCNXv2-SYT18 (-, 5.58 hours)10068.465.371.440.458.180.0
", + "image_path": "09fd96e655a074a1b80e1801d0ca34c30a248ae590aa3e900edcad14b932122d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 251, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 48, + 251, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 547, + 283 + ], + "type": "text", + "content": "SL: Supervised Learning. CL: Contrastive Learning. IN1K: ImageNet1K [11]. K400: Kinetics-400 [27]. YT18: YouTube-VOS 2018 [68]. Table 1. Comparing VideoMAC with previous supervised and self-supervised approaches (SL, CL, MIM, and MVM) on downstream tasks, e.g., video object segmentation on DAVIS17 [46], body part propagation on VIP [73], and human pose tracking on JHMDB [26]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 288, + 287, + 423 + ], + "blocks": [ + { + "bbox": [ + 50, + 288, + 287, + 423 + ], + "lines": [ + { + "bbox": [ + 50, + 288, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 50, + 288, + 287, + 423 + ], + "type": "table", + "html": "
MVM MethodResolution#Param.FLOPsJ&Fm
Iso. ViT Backbone
MAE-ST [16]14 × 14304M61.6G54.6
VideoMAE [56]14 × 1422M4.6G39.3
DropMAE [65]14 × 1487M17.6G53.4
SiamMAE [20]14 × 1422M4.6G62.0
Iso. ConvNet Backbone
VideoMAC (CNXv2-S†)14 × 1422M4.3G53.8
Hie. ConvNet Backbone
VideoMAC (RN18)7 × 712M1.8G64.7
VideoMAC (RN50)7 × 726M4.1G67.2
VideoMAC (CNXv2-T)7 × 729M4.5G67.5
VideoMAC (CNXv2-S)7 × 750M8.7G68.4
", + "image_path": "685467a23b25b3992db8e9c661272b37d73616b1bc913058d4c71433bd301b97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 424, + 287, + 445 + ], + "lines": [ + { + "bbox": [ + 47, + 424, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 287, + 445 + ], + "type": "text", + "content": "Table 2. Comparisons with isotropic (iso.) and hierarchical (hie.) based MVM methods. ‘†’ denotes our modified ios. version." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "type": "text", + "content": "with the MinkowskiEngine sparse convolution library [9]. It's noteworthy that the sparse convolution layer can be seamlessly converted back to a standard convolution layer in downstream tasks without requiring additional processes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": "Pre-training. Different from previous methods that pretrain on the large-scale video dataset K400 [27], VideoMAC pre-trains on YouTube-VOS [68] for 100 epochs, leading to a significant reduction in pre-training time by over 140 times. During VideoMAC pre-training, the standard input comprises two frames with a spatial size of " + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": " pixels. The frame interval for this study is determined to be 1 (Table 4c), implying consecutive frames are chosen. We employ minimal data augmentation, encompassing only random resized cropping and horizontal flipping. A masking ratio of 0.75 is utilized to pre-train our models. We conduct training using an AdamW optimizer [39] with a batch size of 512 and a learning rate of 2e-3. Cosine decay [38] is applied to adjust the learning rate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Downstream Tasks. To gauge the efficacy of VideoMAC at video representation learning, we undertake comprehensive experiments on three video downstream tasks, namely video object segmentation, body part propagation, and hu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": "man pose tracking (Table 1). We also investigate the performance of image recognition through video pre-training. Specifically, our backbone is pre-trained on video data and then fine-tuned on ImageNet1K [11] (Table 5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 346, + 388, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 388, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 388, + 359 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 449 + ], + "type": "text", + "content": "This section begins with assessing VideoMAC's performance in three downstream tasks, comparing it to previous state-of-the-art methods. Subsequently, we conduct extensive ablation studies to elucidate the effectiveness of our proposed approach. The final part of this section explores VideoMAC's ability for image recognition, accompanied by a visualization of the reconstruction results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 456, + 495, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 495, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 495, + 468 + ], + "type": "text", + "content": "4.1. Comparison with Previous methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "Video Object Segmentation. We evaluate VideoMAC on the popular DAVIS17 [46] dataset, which includes a validation set comprising 30 videos. Our evaluation methodology aligns with [6, 25], and we report the typical evaluation metrics [45], including the mean region similarity " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_m" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ", the mean boundary accuracy " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ", and their average values " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}_m" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ". As shown in Table 1, our VideoMAC achieves state-of-the-art results for video object segmentation. With a smaller video training dataset and fewer training epochs, both our VideoMAC models (i.e., CNXv2-S and RN50) surpass the SiamMAE [20] by " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ", respectively, in terms of " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}_m" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ". It's worth noting that our approach achieves superior performance to supervised method [21] based on MVM, with improvements of " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ", respectively. Owing to the hierarchical feature extraction capabilities inherent in ConvNets, our approach demonstrates superior proficiency in spatial information extraction in contrast to ViT-based MVM methods [16, 20, 56, 65]. Concurrently, the introduced reconstruction consistency framework strengthens the acquisition of temporal correlations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22737" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 110, + 103 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 110, + 103 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 110, + 103 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 110, + 103 + ], + "type": "image", + "image_path": "cc779ede0ca14ccce373a4492c3fa80742e7aa07fd5fa91a0f0c7998d394cb40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 103, + 110, + 136 + ], + "blocks": [ + { + "bbox": [ + 49, + 103, + 110, + 136 + ], + "lines": [ + { + "bbox": [ + 49, + 103, + 110, + 136 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 110, + 136 + ], + "type": "image", + "image_path": "1da5c536d33763451793841a7c88469e86b8055691da3e7dfb2556c6b86d2643.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 136, + 110, + 167 + ], + "blocks": [ + { + "bbox": [ + 49, + 136, + 110, + 167 + ], + "lines": [ + { + "bbox": [ + 49, + 136, + 110, + 167 + ], + "spans": [ + { + "bbox": [ + 49, + 136, + 110, + 167 + ], + "type": "image", + "image_path": "59c4dc3fa0f7ce2379bd8ea90f51cc07f7bccf20dd8f78bfa84d403ab57233a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 167, + 109, + 199 + ], + "blocks": [ + { + "bbox": [ + 49, + 167, + 109, + 199 + ], + "lines": [ + { + "bbox": [ + 49, + 167, + 109, + 199 + ], + "spans": [ + { + "bbox": [ + 49, + 167, + 109, + 199 + ], + "type": "image", + "image_path": "beabd69a3a3bbc4e4e0aa2904e73160f8ed59d97015b78f31f3415062bfa6fc7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 545, + 232 + ], + "type": "text", + "content": "Figure 4. Qualitative results of our VideoMAC (using CNXv2-S) for three video downstream tasks: (a) video object segmentation on DAVIS17 [46]), (b) body part propagation on VIP [73], and (c) human pose tracking on JHMDB [26]." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 110, + 70, + 171, + 103 + ], + "blocks": [ + { + "bbox": [ + 110, + 70, + 171, + 103 + ], + "lines": [ + { + "bbox": [ + 110, + 70, + 171, + 103 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 171, + 103 + ], + "type": "image", + "image_path": "39687ffb3312446beca58fb38acc3a0a9030db6a2d18559414163b847b87400b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 110, + 103, + 171, + 136 + ], + "blocks": [ + { + "bbox": [ + 110, + 103, + 171, + 136 + ], + "lines": [ + { + "bbox": [ + 110, + 103, + 171, + 136 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 171, + 136 + ], + "type": "image", + "image_path": "7af476a6a50434a352f3b2d59aa13f561965a7386e62ddafbedcbba801bf9c9e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 110, + 136, + 171, + 167 + ], + "blocks": [ + { + "bbox": [ + 110, + 136, + 171, + 167 + ], + "lines": [ + { + "bbox": [ + 110, + 136, + 171, + 167 + ], + "spans": [ + { + "bbox": [ + 110, + 136, + 171, + 167 + ], + "type": "image", + "image_path": "e40adcfdd3b59ddcb68fa2a1882d7ff108a65a131d56af2ff1646273560adf21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 110, + 167, + 171, + 199 + ], + "blocks": [ + { + "bbox": [ + 110, + 167, + 171, + 199 + ], + "lines": [ + { + "bbox": [ + 110, + 167, + 171, + 199 + ], + "spans": [ + { + "bbox": [ + 110, + 167, + 171, + 199 + ], + "type": "image", + "image_path": "cba9e282e25f99cc7c37a6577d419a4b5c039f717feca5d0c5e016557bf3cc8a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 201, + 223, + 210 + ], + "lines": [ + { + "bbox": [ + 121, + 201, + 223, + 210 + ], + "spans": [ + { + "bbox": [ + 121, + 201, + 223, + 210 + ], + "type": "text", + "content": "(a) Video Object Segmentation" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 172, + 70, + 233, + 102 + ], + "blocks": [ + { + "bbox": [ + 172, + 70, + 233, + 102 + ], + "lines": [ + { + "bbox": [ + 172, + 70, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 70, + 233, + 102 + ], + "type": "image", + "image_path": "4ee736ea6f99104038b48ab1192b6046d3869fb26b322f0319cb12e5663480b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 172, + 103, + 233, + 136 + ], + "blocks": [ + { + "bbox": [ + 172, + 103, + 233, + 136 + ], + "lines": [ + { + "bbox": [ + 172, + 103, + 233, + 136 + ], + "spans": [ + { + "bbox": [ + 172, + 103, + 233, + 136 + ], + "type": "image", + "image_path": "41c047bd1e556a3366499c82d0d823ed48cb3e8f909efa76b3aa6fa898c27c87.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 172, + 136, + 233, + 167 + ], + "blocks": [ + { + "bbox": [ + 172, + 136, + 233, + 167 + ], + "lines": [ + { + "bbox": [ + 172, + 136, + 233, + 167 + ], + "spans": [ + { + "bbox": [ + 172, + 136, + 233, + 167 + ], + "type": "image", + "image_path": "9e81fb94ed7e1a937521e26acca5cac8e9e95f110629865cc1a8f7471a84a227.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 172, + 167, + 233, + 199 + ], + "blocks": [ + { + "bbox": [ + 172, + 167, + 233, + 199 + ], + "lines": [ + { + "bbox": [ + 172, + 167, + 233, + 199 + ], + "spans": [ + { + "bbox": [ + 172, + 167, + 233, + 199 + ], + "type": "image", + "image_path": "672ce70fa1eb31ffc003eed41139d947c52a9b9b664e3edd3f0edd160e4a0dce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 295, + 102 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 295, + 102 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 295, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 295, + 102 + ], + "type": "image", + "image_path": "dc185c900fa52fcb7ce480ce0de5a78e34871720679ab4b29620baff721392ee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 235, + 103, + 295, + 136 + ], + "blocks": [ + { + "bbox": [ + 235, + 103, + 295, + 136 + ], + "lines": [ + { + "bbox": [ + 235, + 103, + 295, + 136 + ], + "spans": [ + { + "bbox": [ + 235, + 103, + 295, + 136 + ], + "type": "image", + "image_path": "8b7dc9ba9f5052d1bf22ad81136461583dba047d56476410f4f99e36c05c2d11.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 235, + 136, + 295, + 167 + ], + "blocks": [ + { + "bbox": [ + 235, + 136, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 235, + 136, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 235, + 136, + 295, + 167 + ], + "type": "image", + "image_path": "030178ddea3e3964ceddaf96234691394f8c0713a51607b50a677aae2d476b2f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 235, + 167, + 295, + 199 + ], + "blocks": [ + { + "bbox": [ + 235, + 167, + 295, + 199 + ], + "lines": [ + { + "bbox": [ + 235, + 167, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 235, + 167, + 295, + 199 + ], + "type": "image", + "image_path": "8abbb16234408df96cac0a82d55e1bac421d74db6f0d2c5062325f7f67387ca1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 359, + 102 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 359, + 102 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 359, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 359, + 102 + ], + "type": "image", + "image_path": "1ee3b50f026dfec11e4fd7f62a93fc6090ad68689053047d89c3d4e7398967fb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 298, + 102, + 359, + 129 + ], + "blocks": [ + { + "bbox": [ + 298, + 102, + 359, + 129 + ], + "lines": [ + { + "bbox": [ + 298, + 102, + 359, + 129 + ], + "spans": [ + { + "bbox": [ + 298, + 102, + 359, + 129 + ], + "type": "image", + "image_path": "0dcefeb0c94cb09100575611095050d64ae9e5b73ce8a9003c3e09e289b98367.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 361, + 70, + 421, + 102 + ], + "blocks": [ + { + "bbox": [ + 361, + 70, + 421, + 102 + ], + "lines": [ + { + "bbox": [ + 361, + 70, + 421, + 102 + ], + "spans": [ + { + "bbox": [ + 361, + 70, + 421, + 102 + ], + "type": "image", + "image_path": "780f8b318bd819a397bb8f5bf4f0928acc4a7318b50580525c4fc136bb72ec04.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 361, + 102, + 421, + 129 + ], + "blocks": [ + { + "bbox": [ + 361, + 102, + 421, + 129 + ], + "lines": [ + { + "bbox": [ + 361, + 102, + 421, + 129 + ], + "spans": [ + { + "bbox": [ + 361, + 102, + 421, + 129 + ], + "type": "image", + "image_path": "ffe6d6cc59693db6108be3e0bb934a428b0dbc1c10dd1c0a5b74b6ead7e06acb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 422, + 70, + 483, + 102 + ], + "blocks": [ + { + "bbox": [ + 422, + 70, + 483, + 102 + ], + "lines": [ + { + "bbox": [ + 422, + 70, + 483, + 102 + ], + "spans": [ + { + "bbox": [ + 422, + 70, + 483, + 102 + ], + "type": "image", + "image_path": "45f1fbf905528e79e1f2dec0ed36d682bbf52dd7a929c2e1da31d677246a1af6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 422, + 102, + 482, + 129 + ], + "blocks": [ + { + "bbox": [ + 422, + 102, + 482, + 129 + ], + "lines": [ + { + "bbox": [ + 422, + 102, + 482, + 129 + ], + "spans": [ + { + "bbox": [ + 422, + 102, + 482, + 129 + ], + "type": "image", + "image_path": "6aad074add475eff0ac939641540b24868a7327dbea84e0cbac933dd1b733868.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 484, + 70, + 545, + 100 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 545, + 100 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 545, + 100 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 545, + 100 + ], + "type": "image", + "image_path": "7bedc0740b8045a63ee17f8a8802aa38b82d457d5da3c6357e11d03e5c7685fd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 484, + 100, + 545, + 129 + ], + "blocks": [ + { + "bbox": [ + 484, + 100, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 484, + 100, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 484, + 100, + 545, + 129 + ], + "type": "image", + "image_path": "1fd26f5d405364568ed7ece610516e363157e5817e212676eaa9352486beeee7.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 298, + 140, + 359, + 170 + ], + "blocks": [ + { + "bbox": [ + 298, + 140, + 359, + 170 + ], + "lines": [ + { + "bbox": [ + 298, + 140, + 359, + 170 + ], + "spans": [ + { + "bbox": [ + 298, + 140, + 359, + 170 + ], + "type": "image", + "image_path": "bc4fa73f645e81f89611fb321b357b3d7cb19070b3e5d559bab70fc658d8f69d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 298, + 170, + 359, + 199 + ], + "blocks": [ + { + "bbox": [ + 298, + 170, + 359, + 199 + ], + "lines": [ + { + "bbox": [ + 298, + 170, + 359, + 199 + ], + "spans": [ + { + "bbox": [ + 298, + 170, + 359, + 199 + ], + "type": "image", + "image_path": "95eeaf3f6df19bf8954565979705f4b938ff7ef3fc51d8eb7806f391f0a33c99.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 361, + 140, + 421, + 170 + ], + "blocks": [ + { + "bbox": [ + 378, + 130, + 465, + 140 + ], + "lines": [ + { + "bbox": [ + 378, + 130, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 378, + 130, + 465, + 140 + ], + "type": "text", + "content": "(b) Body Part Propagation" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 140, + 421, + 170 + ], + "lines": [ + { + "bbox": [ + 361, + 140, + 421, + 170 + ], + "spans": [ + { + "bbox": [ + 361, + 140, + 421, + 170 + ], + "type": "image", + "image_path": "a2f107ce900efc6209ee6e07d5b9769078dee1ca0eeff81d739e4655103a1989.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 361, + 170, + 421, + 199 + ], + "blocks": [ + { + "bbox": [ + 361, + 170, + 421, + 199 + ], + "lines": [ + { + "bbox": [ + 361, + 170, + 421, + 199 + ], + "spans": [ + { + "bbox": [ + 361, + 170, + 421, + 199 + ], + "type": "image", + "image_path": "71686b941ede7f1925c26f130ac37b92fcc0be581bf05c1a3cfe8f5c0b4e5344.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 201, + 463, + 210 + ], + "lines": [ + { + "bbox": [ + 379, + 201, + 463, + 210 + ], + "spans": [ + { + "bbox": [ + 379, + 201, + 463, + 210 + ], + "type": "text", + "content": "(c) Human Pose Tracking" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 421, + 140, + 483, + 170 + ], + "blocks": [ + { + "bbox": [ + 421, + 140, + 483, + 170 + ], + "lines": [ + { + "bbox": [ + 421, + 140, + 483, + 170 + ], + "spans": [ + { + "bbox": [ + 421, + 140, + 483, + 170 + ], + "type": "image", + "image_path": "3197cc538b71bd80bc20001bf8ab42572a4c5f4aa7d11496527579b7b688ff12.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 421, + 170, + 482, + 199 + ], + "blocks": [ + { + "bbox": [ + 421, + 170, + 482, + 199 + ], + "lines": [ + { + "bbox": [ + 421, + 170, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 421, + 170, + 482, + 199 + ], + "type": "image", + "image_path": "2074be0313e82ba1108d27f74e24085bf7ade27605fcd5b5dfd6f8cb1a23f5a1.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 484, + 140, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 484, + 140, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 484, + 140, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 484, + 140, + 545, + 170 + ], + "type": "image", + "image_path": "59ab462eeac5542ee870142213ceb9290a1c1f62a493b38391f06b6653a8d2df.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 484, + 170, + 545, + 199 + ], + "blocks": [ + { + "bbox": [ + 484, + 170, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 484, + 170, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 484, + 170, + 545, + 199 + ], + "type": "image", + "image_path": "eb2d86d842d5a702a28d3e4eddf353a49c36ff68b25283b862123797373b8f74.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "table", + "bbox": [ + 58, + 243, + 201, + 302 + ], + "blocks": [ + { + "bbox": [ + 58, + 243, + 201, + 302 + ], + "lines": [ + { + "bbox": [ + 58, + 243, + 201, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 201, + 302 + ], + "type": "table", + "html": "
BackboneJ&FmJmFm
R1864.762.866.5
R5067.264.969.5
CNXv2-T67.564.470.6
CNXv2-S68.465.371.4
", + "image_path": "6fe44acf70f5e15e84497a10d2738650e50582b1e9fc52ac61c9e965eefc6f7b.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_body" + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 58, + 323, + 201, + 383 + ], + "blocks": [ + { + "bbox": [ + 52, + 304, + 201, + 322 + ], + "lines": [ + { + "bbox": [ + 52, + 304, + 201, + 322 + ], + "spans": [ + { + "bbox": [ + 52, + 304, + 201, + 322 + ], + "type": "text", + "content": "(a) Encoder type. VideoMAC is compatible with both classical and modern ConvNets." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 323, + 201, + 383 + ], + "lines": [ + { + "bbox": [ + 58, + 323, + 201, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 323, + 201, + 383 + ], + "type": "table", + "html": "
Dim.J&FmJmFm
12867.964.871.0
25668.265.371.1
51268.465.371.4
102468.165.071.2
", + "image_path": "6d122d00ab9f69d10e65bd5c596e9bcc32302b5a500a5890c78160864799b739.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 219, + 243, + 362, + 302 + ], + "blocks": [ + { + "bbox": [ + 52, + 384, + 202, + 403 + ], + "lines": [ + { + "bbox": [ + 52, + 384, + 202, + 403 + ], + "spans": [ + { + "bbox": [ + 52, + 384, + 202, + 403 + ], + "type": "text", + "content": "(c) Decoder width. A decoder width of 512 delivers the best performance." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 243, + 362, + 302 + ], + "lines": [ + { + "bbox": [ + 219, + 243, + 362, + 302 + ], + "spans": [ + { + "bbox": [ + 219, + 243, + 362, + 302 + ], + "type": "table", + "html": "
BlocksJ&FmJmFm
168.465.371.4
268.265.271.2
468.365.471.2
868.165.271.0
", + "image_path": "a4252ccd3096b72799006f999bbe6279ead357fcd8452ee7a70de1e1c84f9072.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_body" + } + ], + "index": 40 + }, + { + "type": "table", + "bbox": [ + 219, + 323, + 362, + 360 + ], + "blocks": [ + { + "bbox": [ + 214, + 304, + 362, + 323 + ], + "lines": [ + { + "bbox": [ + 214, + 304, + 362, + 323 + ], + "spans": [ + { + "bbox": [ + 214, + 304, + 362, + 323 + ], + "type": "text", + "content": "(b) Decoder depth. One block achieves competitive performance and efficiency." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 323, + 362, + 360 + ], + "lines": [ + { + "bbox": [ + 219, + 323, + 362, + 360 + ], + "spans": [ + { + "bbox": [ + 219, + 323, + 362, + 360 + ], + "type": "table", + "html": "
MaskingJ&FmJmFm
Asym.66.563.469.6
Sym.68.465.371.4
", + "image_path": "1e0b092ae55a01f8ce68eee868aed1b2a7a92632e88555317a461916d8be95ea.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 213, + 384, + 363, + 403 + ], + "lines": [ + { + "bbox": [ + 213, + 384, + 363, + 403 + ], + "spans": [ + { + "bbox": [ + 213, + 384, + 363, + 403 + ], + "type": "text", + "content": "(d) Masking design. Symmetric (sym.) masking outperforms asymmetric (asym.) one." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 381, + 242, + 543, + 388 + ], + "blocks": [ + { + "bbox": [ + 381, + 242, + 543, + 388 + ], + "lines": [ + { + "bbox": [ + 381, + 242, + 543, + 388 + ], + "spans": [ + { + "bbox": [ + 381, + 242, + 543, + 388 + ], + "type": "image", + "image_path": "af566f9822535df270ac0e7184972c9b55ad6e8e0256d024df65fbd315924569.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 396, + 543, + 450 + ], + "lines": [ + { + "bbox": [ + 380, + 396, + 543, + 450 + ], + "spans": [ + { + "bbox": [ + 380, + 396, + 543, + 450 + ], + "type": "text", + "content": "Figure 5. For masked modeling, dense convolution usually results in the dissipation of mask structures. The deployment of sparse convolution proves to be an effective solution, enabling ConvNet-based MIM / MVM." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 47, + 405, + 369, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 369, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 369, + 450 + ], + "type": "text", + "content": "Table 3. VideoMAC encoder and decoder ablation experiments on DAVIS17 [46]. We report ablation results for two of the most emblematic families of ConvNets. Unless specified otherwise, the encoder is CNXv2-S in default. 'Dim' indicates the number of decoder dimensions. Default settings employed in this paper are marked in blue." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": "Body Part Propagation. We evaluate VideoMAC on the Video Instance Parsing (VIP) dataset [73], and the quantitative results, measured by mIoU, are detailed in Table 1. This validation set comprises of 50 videos that concentrate on the propagation of 19 body parts, such as arms and legs, thereby necessitating a heightened level of precision in matching as compared to video object segmentation. We adhere to the same parameters as [25], adjusting the video frames to " + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "inline_equation", + "content": "560 \\times 560" + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": " dimensions. In the case of body part propagation, two variants of our VideoMAC demonstrate improvements by " + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": " (CNXv2-S) and " + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": " (RN50), respectively, in comparison with the top-performing SiamMAE [20] in MVM. Significantly, our RN50-based model outperforms CNXv2-S on the VIP dataset [73] due to the higher dimensionality of RN50 features (2048 vs. 768), which benefits complex tasks and improves accuracy in mask propagation." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Human Pose Tracking. We conduct the human pose tracking task on the validation set of JHMDB [26], which includes 268 videos and entails detecting 15 human keypoints. Conforming to the protocol delineated in [25], we resize video frames to a resolution of " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": ", and employ" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 461, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 508 + ], + "type": "text", + "content": "the Possibility of Correct Keypoints (PCK) metric [51] for quantitative evaluation. As shown in Table 1, both variants (i.e., RN50 and CNXv2-S) of our VideoMAC outperforms all MVM methods in terms of PCK@0.1 and PCK@0.2." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 509, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 557 + ], + "type": "text", + "content": "Qualitative results. As depicted in Fig. 4, we present the qualitative results of mask propagation achieved by VideoMAC on three downstream tasks. Our VideoMAC yields notable performance across a spectrum of complex videos." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 582, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 605 + ], + "type": "text", + "content": "In this section, we conduct extensive ablation studies to analyze the component design of our VideoMAC." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Encoder design: hierarchical vs. isotropic. In this ablation study, we scrutinize diverse encoder architectures to our VideoMAC, i.e., hierarchical and isotropic designs. Precisely, we adopted the ViT [14] design principles and restructured the CNXv2-S backbone from having a hierarchical structure into an isotropic one. This implies the creation of an architecture devoid of downsampling layers, ensuring consistent feature resolution (e.g., " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ") throughout all depths. As shown in Table 2, the performance dif" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22738" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 247, + 106 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 247, + 106 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 247, + 106 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 247, + 106 + ], + "type": "table", + "html": "
DatasetSizeJ&FmJmFm
YT18 [68]5.58 hours68.465.371.4
K4001 [27]800 hours67.865.270.3
", + "image_path": "ac213057ede8f573f92fd865c69a6d724721510ad05b28650ffe88deb498b601.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 257, + 70, + 414, + 116 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 414, + 116 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 414, + 116 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 414, + 116 + ], + "type": "table", + "html": "
AugmentationJ&FmJmFm
spatial68.465.371.4
color64.961.868.0
spatial + color66.162.969.3
", + "image_path": "b3fe293c003bebd83de49d5ed613a8387899fa6e9d2492f21f26ca85b054a992.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 251, + 118, + 414, + 137 + ], + "lines": [ + { + "bbox": [ + 251, + 118, + 414, + 137 + ], + "spans": [ + { + "bbox": [ + 251, + 118, + 414, + 137 + ], + "type": "text", + "content": "(b) Data augmentation. Minimal data augmentation achieves the best performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 424, + 70, + 544, + 116 + ], + "blocks": [ + { + "bbox": [ + 424, + 70, + 544, + 116 + ], + "lines": [ + { + "bbox": [ + 424, + 70, + 544, + 116 + ], + "spans": [ + { + "bbox": [ + 424, + 70, + 544, + 116 + ], + "type": "table", + "html": "
GapJ&FmJmFm
168.465.371.4
566.463.569.2
1063.260.465.9
", + "image_path": "74aefb25d9cb38918231068ba717a7b87cd5073f02eace53330a36052dd27b57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 417, + 118, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 417, + 118, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 417, + 118, + 545, + 137 + ], + "type": "text", + "content": "(c) Frame gap. Video pairs with one frame gap perform the best results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 55, + 139, + 204, + 197 + ], + "blocks": [ + { + "bbox": [ + 47, + 118, + 247, + 137 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 247, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 247, + 137 + ], + "type": "text", + "content": "(a) Training Dataset. Our VideoMAC pre-trained on YT18 yields better performance and efficiency compared to K400." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 139, + 204, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 139, + 204, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 204, + 197 + ], + "type": "table", + "html": "
\\( {\\mathcal{L}}_{c} \\)\\( {\\mathcal{L}}_{t} \\)\\( \\mathcal{J}\\&{\\mathcal{F}}_{m} \\)\\( {\\mathcal{J}}_{m} \\)\\( {\\mathcal{F}}_{m} \\)
✘✘51.152.250.0
✘✓60.656.364.9
✓✘63.463.468.7
✓✓68.465.371.4
", + "image_path": "1d8ddfb775507dfe71f5f0f3e9b5b236263782db486cabeef8ff99061232d765.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 220, + 139, + 369, + 198 + ], + "blocks": [ + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "type": "text", + "content": "(d) Model loss. VideoMAC utilizes consistency loss " + }, + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "type": "text", + "content": " and target loss " + }, + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 48, + 198, + 205, + 217 + ], + "type": "text", + "content": " for the best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 220, + 139, + 369, + 198 + ], + "lines": [ + { + "bbox": [ + 220, + 139, + 369, + 198 + ], + "spans": [ + { + "bbox": [ + 220, + 139, + 369, + 198 + ], + "type": "table", + "html": "
ConsistencyJ&FmJmFm
γ=0.060.656.364.9
γ=0.164.862.067.6
γ=0.565.863.268.5
γ=1.068.465.371.4
", + "image_path": "6ef0bdd05e0688e52be411d0c71be1d20557547519b17fa609c0c2c490b9b64b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 384, + 139, + 543, + 197 + ], + "blocks": [ + { + "bbox": [ + 213, + 198, + 369, + 217 + ], + "lines": [ + { + "bbox": [ + 213, + 198, + 369, + 217 + ], + "spans": [ + { + "bbox": [ + 213, + 198, + 369, + 217 + ], + "type": "text", + "content": "(e) Reconstruction consistency. When " + }, + { + "bbox": [ + 213, + 198, + 369, + 217 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 213, + 198, + 369, + 217 + ], + "type": "text", + "content": " is set to 1.0, the consistency loss works the best." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 384, + 139, + 543, + 197 + ], + "lines": [ + { + "bbox": [ + 384, + 139, + 543, + 197 + ], + "spans": [ + { + "bbox": [ + 384, + 139, + 543, + 197 + ], + "type": "table", + "html": "
Masking ratioJ&FmJmFm
0.6566.563.869.1
0.7568.465.371.4
0.8568.365.371.3
0.9560.359.760.9
", + "image_path": "656416b244da2aecf2459b140fc78d96d65e55edf3197c0ca47481c94dd1a8ce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 198, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 378, + 198, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 378, + 198, + 544, + 217 + ], + "type": "text", + "content": "(f) Masking ratio. Our VideoMAC uses a masking ratio of 0.75 for the highest performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "type": "text", + "content": "Table 4. VideoMAC ablations on data, loss, and masking ratio configuration. We report two official metrics " + }, + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "type": "inline_equation", + "content": "{\\mathcal{I}}_{m}" + }, + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "type": "inline_equation", + "content": "{\\mathcal{F}}_{m}" + }, + { + "bbox": [ + 47, + 219, + 545, + 230 + ], + "type": "text", + "content": " on DAVIS17 [46]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 47, + 237, + 166, + 335 + ], + "blocks": [ + { + "bbox": [ + 47, + 237, + 166, + 335 + ], + "lines": [ + { + "bbox": [ + 47, + 237, + 166, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 166, + 335 + ], + "type": "image", + "image_path": "b27b70dea0c2e01c788fb99b392cc02eb752a1699c3ea50d8dd8e813dc78a963.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 335, + 287, + 356 + ], + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 356 + ], + "type": "text", + "content": "Figure 6. Analysis of individual loss components after Video-MAC pre-training 100 epochs on YouTube-VOS 2018 [68]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 169, + 237, + 287, + 335 + ], + "blocks": [ + { + "bbox": [ + 169, + 237, + 287, + 335 + ], + "lines": [ + { + "bbox": [ + 169, + 237, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 169, + 237, + 287, + 335 + ], + "type": "image", + "image_path": "7cd9e0c1d30254190a9cb34d2128c6d67750472d7b0cb8763719acc7dc19a00b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "type": "text", + "content": "ference between our VideoMAC (iso.) and the ViT-based MVM methods (e.g., VideoMAE [56] and DropMAE [65]) is marginal. By maintaining the hierarchical structure of ConvNets within VideoMAC models (hie.), we improve the performance significantly, even though their feature resolution (i.e., " + }, + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "type": "text", + "content": ") is merely half that of the ViT-based models [16, 20, 56, 65]. This can be ascribed to ConvNets' capacity to preserve multi-scale spatial information." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 459, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 590 + ], + "type": "text", + "content": "Encoder backbone: classical and modern. For our study, we incorporate two emblematic ConvNet families, i.e., ResNets [21] and ConvNeXts [36, 64]. As shown in Table 3a, we substitute dense convolution with sparse convolution in the baseline ConvNets. This adaptation seamlessly dovetails into our VideoMAC framework, thereby facilitating a ConvNet-based MVM. In addition, Fig. 5 highlights the influence of employing dense and sparse convolution on masks. It is apparent that sparse convolution effectively safeguards the structural integrity of the mask, thus enabling the integration of MAE pre-training within ConvNets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 592, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 687 + ], + "type": "text", + "content": "Decoder design: blocks and dimensions. Adhering to a simplistic design principle, our VideoMAC adopts the decoder from the purely convolutional network structure deployed in ConvNeXt [36]. We conduct ablation experiments focusing on the depth and width of the decoder, with the results detailed in Tables 3b and 3c. The decoder exhibits competitive results with a single block and 512 dimensions, striking a trade-off between performance and efficiency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Masking design: asymmetric vs. symmetric. In Table 3d, we investigate the impact of two masking strategies, namely" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 306, + 237, + 425, + 334 + ], + "blocks": [ + { + "bbox": [ + 306, + 237, + 425, + 334 + ], + "lines": [ + { + "bbox": [ + 306, + 237, + 425, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 425, + 334 + ], + "type": "image", + "image_path": "fb391479d6808a393acae730f483edd47360905efa464d34859cb3bcf326c937.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 427, + 237, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 427, + 237, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 427, + 237, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 427, + 237, + 545, + 335 + ], + "type": "image", + "image_path": "fb4cc285ae593a038d53d28330309bfbf1163c8bf56f7226a961c3037f80f4d8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "type": "text", + "content": "Figure 7. Consistency loss in relation to " + }, + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "type": "text", + "content": " (left: identical encoder, CNXv2-S) and encoders (right: identical weight factor, " + }, + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 305, + 335, + 545, + 355 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 360, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 444 + ], + "type": "text", + "content": "asymmetric (same masking ratio but different locations) and symmetric (same masking ratio and locations), on performance. Compared to asymmetric masking, symmetric masking simplifies the calculation of the reconstruction loss, as it covers the entire masked area. Moreover, it facilitates the inclusion of more reconstruction regions in the loss calculation, resulting in faster model convergence." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": "Data setting: dataset, augmentation, and gap. To assess the influence of different video datasets, Table 4a presents the results of pre-training our VideoMAC on both YT18 [68] and " + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "inline_equation", + "content": "\\mathrm{K400^1}" + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": " [27]. The ablation results of VideoMAC on the two datasets differ by just " + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": " , demonstrating the robustness and adaptability of our method to different datasets. For data augmentation, as shown in Table 4b, our VideoMAC achieves the best results using only spatial strategy, i.e., random resized cropping and horizontal flipping. Additionally, we conduct an ablation study on the frame gap, and the optimal results are obtained with consecutive frames (i.e., the gap is 1), as indicated in Table 4c. The contribution of " + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": " . Previous MVM methods predominantly rely on a single loss, i.e., the reconstruction loss, necessitating multiple frames to be simultaneously processed by the model. In contrast, our study utilizes a more computation-friendly strategy by inputting frame pairs into two separate encoders, followed by parameter updates using EMA. Consequently, we introduce two distinct loss terms: online " + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_o" + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 304, + 445, + 546, + 695 + ], + "type": "text", + "content": " reconstruction losses. Furthermore, to integrate temporal information, we" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 502, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 502, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 502, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 702, + 502, + 712 + ], + "type": "text", + "content": "K400 is sampled to match YT18's training data volume." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22739" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 544, + 159 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 544, + 159 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 544, + 159 + ], + "type": "image", + "image_path": "393a600959fe9b404075ae11d41b624fb408fcf8dd9c9cd2d9a889f058a3f415.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 160, + 546, + 182 + ], + "lines": [ + { + "bbox": [ + 47, + 160, + 546, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 546, + 182 + ], + "type": "text", + "content": "Figure 8. Visualization of reconstruction examples using a pre-trained VideoMAC (CNXv2-S) model with 0.75 masking ratio. These frames are randomly selected from the validation set of YT18 [68] and masked " + }, + { + "bbox": [ + 47, + 160, + 546, + 182 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 47, + 160, + 546, + 182 + ], + "type": "text", + "content": " of patches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": "compute the loss between the reconstructed results of frame pairs, colloquially known as the reconstruction consistency loss " + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": ". As shown in Table 4d, " + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": " improve " + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}_m" + }, + { + "bbox": [ + 46, + 190, + 286, + 262 + ], + "type": "text", + "content": ", respectively, with the full VideoMAC model yielding the best results. The diverse loss components during pre-training are illustrated in Fig. 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": "The effect of weight factor " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": ". To investigate the impact of the weight factor on reconstruction consistency loss " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": ", we conduct experiments with different factor settings, i.e., " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\{0.0, 0.1, 0.5, 1.0\\}" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": ". As illustrated in Table 4e, superior performance is attained when " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": " approximates to 1. When " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": " equals 0, which signifies the absence of the reconstruction consistency loss term, a detrimental effect on performance is noticeable. Simultaneously, we record " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": " for various weight factors during pre-training, as depicted in Fig. 7. It becomes evident that an elevated value of " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": " results in a lesser loss. Ultimately, both the R50 and CNXv2-S models converge to comparable final losses. It is noteworthy that " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": " presents an initial decrease, succeeded by an increase, and ultimately reaches convergence, which aligns with the description in Eq. (2), implying that the ultimate reconstruction results between two frames converge to " + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 262, + 286, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 455, + 286, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 286, + 551 + ], + "type": "text", + "content": "The impact of masking ratio. An ablation analysis of the masking ratio is conducted, and the results for our Video-MAC models (CNXv2-S) are documented in Table 4f. Ablation results portray a performance pattern that first ascends and subsequently descends with increasing masking ratios. Noticeably, the performance of our approach experiences a drastic decline at a 0.95 masking ratio caused by overfitting to the self-supervised pretext task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 182, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 182, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 182, + 575 + ], + "type": "text", + "content": "4.3. Additional Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "VideoMAC for Image Recognition. As shown in Table 5, utilizing CNXv1-T [36] / CNXv2-T [64] pre-trained by our VideoMAC on YT18 [68], we achieve " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "82.3\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "82.9\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " accuracy on IN1K [11] image classification. In comparison to CNXv1-T and CNXv2-T trained from scratch, our approach improves the image recognition results by " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": ". Remarkably, the performance garnered by VideoMAC (MVM) is proximate to or commensurate with the state-of-the-art ConvNet-based MIM methods [54, 64]. Given the pre-training cost (100 epochs vs. 1600 epochs) and the domain gap (video vs. image), we regard this im" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 193, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 309, + 193, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 309, + 193, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 309, + 193, + 545, + 277 + ], + "type": "table", + "html": "
MethodBackbonePT epochPT dataFT acc.
Supervised [36]CNXv1-T--82.1
FCMAE [64]CNXv1-T1600IN1K82.3 (+0.2)
Spark [54]CNXv1-T1600IN1K82.4 (+0.3)
VideoMACCNXv1-T100YT1882.3 (+0.2)
Supervised [64]CNXv2-T--82.5
FCMAE [64]CNXv2-T1600IN1K83.0 (+0.5)
VideoMACCNXv2-T100YT1882.9 (+0.4)
", + "image_path": "d8f9dcb3617c5b4273b415dc016acd3e1f77e0090418bafbd5f3f6dd3b3e0eb2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 278, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 545, + 322 + ], + "type": "text", + "content": "Table 5. Comparing VideoMAC with previous ConvNet-based MIM approaches (e.g., Spark [54] and FCMAE [64]) on the image downstream task, i.e., fine-tuning recognition accuracy (acc.) on IN1K [11]. 'PT' and 'FT' indicate pre-training and fine-tuning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 327, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 400 + ], + "type": "text", + "content": "provement of our VideoMAC as noteworthy and promising compared to training from scratch and pre-training by MIM. Visualization. We visualize reconstruction results from the YT18 [68] validation set to analyze performance in MVM pre-training. In Fig. 8, VideoMAC makes reasonable predictions about the masked regions, e.g., color and contour." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 412, + 378, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 412, + 378, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 412, + 378, + 425 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 433, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 564 + ], + "type": "text", + "content": "In this study, we present VideoMAC, an MVM pre-training framework constructed fully using ConvNets. Our masked modeling for frame pairs is executed by introducing an online network (encoder and decoder) optimized by gradients, as well as the target network that updates parameters via EMA. Concurrently, for reconstruction results derived from both online and target networks, our proposed loss of reconstruction consistency between frame pairs paves the way for temporal modeling. Our VideoMAC empowers a multitude of ConvNets to reap the performance gains for downstream tasks (e.g., video but also image) by MVM pre-training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": "Limitation and Future Work. Existing ConvNet-based architectures typically use a fixed downsampling ratio, which in turn leads to a relatively fixed patch size. For future work, we will investigate ConvNets with adjustable patch size, allowing for higher feature resolution like ViT." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 625, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No. 62102182, 62202227, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20220934, BK20220938, BK20220936), China Postdoctoral Science Foundation (No. 2022M721626, 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No.30923010303)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22740" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 178 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 236 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 237, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 237, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 237, + 287, + 281 + ], + "type": "text", + "content": "[4] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 282, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 327 + ], + "type": "text", + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912–9924, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 372 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 407 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 430 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 463 + ], + "type": "text", + "content": "[9] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 498 + ], + "type": "text", + "content": "[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 533 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, , and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "text", + "content": "[13] Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language understanding and generation. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 688 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, haoqi fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, pages 35946-35958, 2022. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "text", + "content": "[17] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. NeurIPS, 35:35632-35644, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 173, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 237 + ], + "type": "text", + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[19] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "type": "text", + "content": "[20] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 328, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 369 + ], + "type": "text", + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 371, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 414 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In ECCV, pages 123-143, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[25] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, pages 19545-19560, 2020. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, pages 3192-3199, 2013. 1, 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 537, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 592 + ], + "type": "text", + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 667 + ], + "type": "text", + "content": "[29] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[30] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. In ICLR, 2023. 1, 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22741" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Lulei Li, Tianfei Zhou, Wenguan Wang, Lu Yang, Jianwu Li, and Yi Yang. Locality-aware inter-and intra-video reconstruction for self-supervised correspondence learning. In CVPR, pages 8719-8730, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[32] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In CVPR, pages 18706-18716, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[33] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In CVPR, pages 6252-6261, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "text", + "content": "[34] Yuan Liu, Songyang Zhang, Jiacheng Chen, Zhaohui Yu, Kai Chen, and Dahua Lin. Improving pixel-based mim by reducing wasted modeling capability. In ICCV, pages 5361-5372, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 249, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 288, + 293 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 326 + ], + "type": "text", + "content": "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "type": "text", + "content": "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 381 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 403 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "text", + "content": "[40] Xiankai Lu, Wenguan Wang, Martin Danelljan, Tianfei Zhou, Jianbing Shen, and Luc Van Gool. Video object segmentation with episodic graph memory networks. In ECCV, pages 661-679, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "text", + "content": "[41] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In CVPR, pages 8960-8970, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "type": "text", + "content": "[42] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, pages 596-613, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 537, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 580 + ], + "type": "text", + "content": "[43] Gensheng Pei, Fumin Shen, Yazhou Yao, Tao Chen, Xian-Sheng Hua, and Heng-Tao Shen. Hierarchical graph pattern understanding for zero-shot video object segmentation. TIP, 32:5909-5920, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[44] Gensheng Pei, Yazhou Yao, Fumin Shen, Dan Huang, Xingguo Huang, and Heng-Tao Shen. Hierarchical co-attention propagation network for zero-shot video object segmentation. TIP, pages 2348–2359, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[45] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[46] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 162, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 226 + ], + "type": "text", + "content": "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Malik, Yanghao Li, and Christoph Feichtenhofer. Hiera: A hierarchical vision transformer without the bells-and-whistles. ICML, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 258 + ], + "type": "text", + "content": "[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 260, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 293 + ], + "type": "text", + "content": "[51] Jie Song, Limin Wang, Luc Van Gool, and Otmar Hilliges. Thin-slicing network: A deep structured model for pose estimation in videos. In CVPR, pages 4220-4229, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "type": "text", + "content": "[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In CVPR, pages 7000–7009, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 415 + ], + "type": "text", + "content": "[54] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. In ICLR, 2023. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[55] Yunjie Tian, Lingxi Xie, Zhaozhi Wang, Longhui Wei, Xiaopeng Zhang, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Integrally pre-trained transformer pyramid networks. In CVPR, pages 18610-18620, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "type": "text", + "content": "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "text", + "content": "[57] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[58] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 592, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 647 + ], + "type": "text", + "content": "[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[60] Wenguan Wang, Xiankai Lu, Jianbing Shen, David J Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, pages 9236-9245, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 690, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 712 + ], + "type": "text", + "content": "[61] Wenguan Wang, Hongmei Song, Shuyang Zhao, Jianbing Shen, Sanyuan Zhao, Steven CH Hoi, and Haibin Ling." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22742" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 689 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "Learning unsupervised video object segmentation through visual attention. In CVPR, pages 3064-3074, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "type": "text", + "content": "[62] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 185 + ], + "type": "text", + "content": "[63] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "type": "text", + "content": "[64] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In CVPR, pages 16133-16142, 2023. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 286, + 285 + ], + "type": "text", + "content": "[65] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In CVPR, pages 14561–14571, 2023. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 286, + 319 + ], + "type": "text", + "content": "[66] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, pages 13879-13889, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 286, + 352 + ], + "type": "text", + "content": "[67] Z. Xie, Z. Zhang, Y. Cao, Y. Lin, J. Bao, and Z. Yao. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 286, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 286, + 408 + ], + "type": "text", + "content": "[68] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, pages 585-601, 2018. 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 410, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 286, + 453 + ], + "type": "text", + "content": "[69] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, pages 2623-2632, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 455, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 497 + ], + "type": "text", + "content": "[70] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 286, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 286, + 543 + ], + "type": "text", + "content": "[71] Yurong Zhang, Lielei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In CVPR, pages 2246-2256, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "text", + "content": "[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "text", + "content": "[73] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In ACMMM, pages 1527-1535, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 623, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 286, + 655 + ], + "type": "text", + "content": "[74] Tianfei Zhou, Fatih Porikli, David J Crandall, Luc Van Gool, and Wenguan Wang. A survey on deep learning technique for video segmentation. TPAMI, 45(6):7099-7122, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "text", + "content": "[75] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, pages 408-417, 2017. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22743" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_content_list.json b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..289710e5832ab9b174cefe7286d42ea39daaafb0 --- /dev/null +++ b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams", + "text_level": 1, + "bbox": [ + 99, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liao Wang $^{1,3*}$", + "bbox": [ + 107, + 179, + 223, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaixin Yao $^{1,3*}$", + "bbox": [ + 263, + 179, + 379, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengcheng Guo1", + "bbox": [ + 419, + 179, + 565, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhirui Zhang", + "bbox": [ + 604, + 180, + 720, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiang Hu4#", + "bbox": [ + 761, + 180, + 859, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyi Yu", + "bbox": [ + 272, + 205, + 357, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\operatorname{Lan} \\mathrm{Xu}^{1 \\dagger}$", + "bbox": [ + 437, + 205, + 514, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minye $\\mathsf{W}\\mathsf{u}^{2\\dagger}$", + "bbox": [ + 594, + 205, + 694, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 ShanghaiTech University", + "bbox": [ + 94, + 223, + 307, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 KU Leuven", + "bbox": [ + 348, + 224, + 454, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 NeuDim", + "bbox": [ + 495, + 224, + 578, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Shanghai Jiao Tong University", + "bbox": [ + 620, + 224, + 874, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfc08507d0ad659b88a7fdb9ff37678645d1074ea704b2df3fc367fb30c3cf71.jpg", + "image_caption": [ + "Figure 1. Our proposed VideoRF views dynamic radiance field as 2D feature video streams combined with deferred rendering. This technique facilitates hardware video codec and shader-based rendering, enabling smooth high-quality rendering across diverse devices." + ], + "image_footnote": [], + "bbox": [ + 84, + 250, + 385, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04801f56eef367f1f5977e4e9b1c0eba39dcbce44ec1c072b06903c8e51b4a61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 273, + 687, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc371545d012708588b29090605eb4a5be5625bee93bab7cfb7e03f278d94969.jpg", + "image_caption": [ + "iPad Pro:Apple M2 chip" + ], + "image_footnote": [], + "bbox": [ + 696, + 250, + 879, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1258810ec2881b0cd27c587a13f777e3f4c11ae60915e2e86ca5b76d802cf496.jpg", + "image_caption": [ + "1920x1080, 40FPS, 0.71MB/frame", + "Dell: i5-1135G7 CPU, Integrated GPU", + "1200x800, 49FPS, 0.65MB/frame" + ], + "image_footnote": [], + "bbox": [ + 696, + 349, + 877, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 508, + 312, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) excel in photorealistically rendering static scenes. However, rendering dynamic, long-duration radiance fields on ubiquitous devices remains challenging, due to data storage and computational constraints. In this paper, we introduce VideoRF, the first approach to enable real-time streaming and rendering of dynamic human-centric radiance fields on mobile platforms. At the core is a serialized 2D feature image stream representing the 4D radiance field all in one. We introduce a tailored training scheme directly applied to this 2D domain to impose the temporal and spatial redundancy of the feature image stream. By leveraging the redundancy, we show that the feature image stream can be efficiently compressed by 2D video CODECs, which allows us to exploit video hardware accelerators to achieve real-time decoding. On the other hand, based on the feature image stream, we propose a novel rendering pipeline for VideoRF, which has specialized space mappings to query radiance properties efficiently. Paired with a deferred shading model, VideoRF has the capability of real-time rendering on mobile devices", + "bbox": [ + 73, + 541, + 470, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thanks to its efficiency. We have developed a real-time interactive player that enables online streaming and rendering of dynamic scenes, offering a seamless and immersive free-viewpoint experience across a range of devices, from desktops to mobile phones. Our project page is available at https://aoliaol2138.github.io/VideoRF/.", + "bbox": [ + 496, + 496, + 890, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 617, + 630, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photorealistic Free-Viewpoint Video (FVV) of dynamic scenes offers an immersive experience in virtual reality and telepresence. Work involving Neural Radiance Fields (NeRFs) has shown great potential in creating photorealistic Free-Viewpoint Videos (FVVs). However, there are still challenges in smoothly delivering and rendering FVVs using NeRFs on commonly used devices, similar to the ease of watching online videos. The difficulty lies in reducing the data capacity for transmitting and storing long sequences and ensuring a low, mobile-compatible computational load.", + "bbox": [ + 496, + 642, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) [39] surpasses traditional 3D reconstruction methods in photorealistic novel view synthesis. Several works extend NeRF to dynamic scenes by maintaining a canonical space and matching it implicitly [10, 42, 48] or explicitly [34] to align with each frame's live space. However, their dependence on canonical space limits effectiveness in sequences with large motions or topo", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Authors contributed equally to this work. † The corresponding authors are Minye Wu (minye.wu@kuleuven.be) and Lan Xu (xulan1@shanghaiitech.edu.cn). # Work done while at ShanghaiTech University.", + "bbox": [ + 75, + 851, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "470", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "logical changes. Other methods introduce new representations like 4D feature grids [18] or temporal voxel features [11], achieving impressive results on the scenes with topological transformations. Yet, [11] struggles with representing longer sequences owing to model capacity constraints, while [18] encounters streaming challenges due to large storage needs. Recent efforts [26, 54, 65] focus on compressing dynamic frames for streaming, but their computational intensity limits mobile applicability. Concurrently, on mobile devices, real-time rendering of static scenes has been achieved by baking NeRF into mesh templates [1, 52, 75] or texture assets [50]. However, these techniques fall short for dynamic scenes as their per-frame representation becomes too bulky for real-time loading. Moreover, while NeRF compression methods [51, 56, 65] can be employed, they introduce decoding or rendering overheads unsuitable for mobile platforms. Despite the existence of NeRF solutions for dynamic scenes and mobile optimization, a cohesive approach that effectively addresses both still poses a difficulty.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose VideoRF - a novel neural modeling approach that enables real-time streaming and rendering of human-centric dynamic radiance fields on common mobile devices (see Fig. 1). Our key idea is to view the 4D feature volumes reconstructed from a dynamic scene as a 2D feature image stream, which is friendly to video codec. Each feature image records the densities and appearance features of one frame. We hence propose a rendering pipeline for this representation. In this rendering pipeline, VideoRF uses a set of mapping tables to connect the 3D space with the 2D feature images. This allows for O(1) density and feature retrieval from the images for every 3D sample point in the space. To reduce the computational complexity of rendering, we adopt the deferred shading model [50] with a global tiny MLP to decode the integrated feature vectors into pixel colors. All rendering operations here are low-cost and compatible with the fragment Shader model, making it possible to implement this rendering pipeline on various devices that have GPUs.", + "bbox": [ + 75, + 397, + 472, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we present a sequential training scheme to effectively generate the 2D feature image stream. Specifically, VideoRF adaptively groups frames by analyzing motion from the sequential data to ensure temporal stability for mapping table generation. We deploy 3D and 2D Morton sorting techniques to improve spatial consistency in the mapping. VideoRF contains spatial and temporal continuity regularizers that are applied on the mapped 2D feature images using the mapping table. This training strategy enforces the temporal and spatial redundancy to the 2D feature image stream. We show that our feature image stream with spatiotemporal sparsity can be efficiently compressed by off-the-shelf video codecs and reaches high compression rates.", + "bbox": [ + 75, + 688, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, based on these findings, we build a cross-platform player based on VideoRF that can play FVVs in real-time, supported by video hardware accelerators. In this way, users can interactively drag, rotate, pause, play, fast-forward, rewind, and jump to specific frames, providing a viewing experience as seamless as watching online videos on various devices, including smartphones, tablets, laptops, and desktops which was unseen before.", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions include:", + "bbox": [ + 517, + 212, + 792, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose VideoRF, a novel approach to enable real-time dynamic radiance field decoding, streaming and rendering on mobile devices.", + "- We present an efficient and compact representation, which represents 4D radiance field into 2D feature stream with low rendering complexity to support hardware video codec and Shader rendering.", + "- We introduce a training scheme to directly impose spatial-temporal consistency on our 2D feature stream for efficient compression." + ], + "bbox": [ + 500, + 227, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 392, + 637, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis for Dynamic Scenes. Dynamic scenes pose a greater challenge in achieving realistic view synthesis results due to moving objects. One approach is to reconstruct the dynamic scene and render the geometry from new viewpoints. The conventional RGB [7, 23, 30, 36, 37, 49, 80] or RGB-D [9, 19-21, 41, 70, 71] solutions have been extensively investigated. Mesh-based neural network [5, 61, 68] techniques are effective for compact data storage and can record view-dependent texture [5, 68], but such methods heavily rely on geometry, especially perform poorly for topologically complex scenarios.", + "bbox": [ + 496, + 416, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many methods extend NeRF into the dynamic view synthesis settings. Some methods [10, 15, 29, 31-33, 43, 48, 63, 69, 77, 79] handle spatial change directly conditions on time and [14, 45, 46, 73, 82] use feature latent code to represent time information. However, they do not support the streaming of long sequences due to the limited representation. Others learn spatial offsets from the live scene to the canonical radiance field by using explicit voxel matching [34], skeletal poses [25, 35, 44], deformed hashtable [24] and deformed volume [22, 59, 60, 66, 72, 81]. While these methods can successfully render dynamic scenes with photo-realistic quality, their heavy reliance on the canonical space makes these methods vulnerable to long sequences, large motions, and topological alterations. Recently, several methods using 4D planes [3, 13, 18, 53, 74], voxel grid [11], Fourier representation [64], residual layers [38] or dynamic MLP maps [47] to represent dynamic scene. However, the inference computational complexity makes it difficult for them to implement streaming and decoding on mobile devices. Additionally, several methods such as [26, 54, 65] have managed to facilitate the streaming of dynamic radi", + "bbox": [ + 496, + 584, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "471", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1375062f170c12c0227f12a79b1f8217deb280cc4b8ffd71fbf45552243974c8.jpg", + "image_caption": [ + "Figure 2. Demonstration of our rendering. For each 3D sample point, its density $\\sigma$ and feature $\\mathbf{f}$ are fetched from the 2D feature image through the mapping table $\\mathbf{M}$ . Each point feature is first volumetrically accumulated to get the ray feature $\\tilde{\\mathbf{f}}$ and pass MLP $\\Phi$ to decode the ray color." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ance fields by reducing the capacity of each frame, but the lack of general hardware acceleration makes it challenging to represent dynamic scenes in real-time on mobile phones.", + "bbox": [ + 75, + 342, + 468, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross Device Neural Radiance Field Rendering. Recent works have demonstrated that static neural radiance fields can be rendered on mobile devices. They achieve this by directly converting the neural radiance field to mesh [58] or using traditional texture to represent the scene. [1, 75, 78] have endeavored to augment the capabilities of neural rendering by employing mesh templates and feature texture, [2, 52] facilitating the rendering process through fragment Shader to achieve surface-like rendering with similar ideas. [62] represents neural radiance features encoded on a two-layer duplex mesh for better rendering quality.", + "bbox": [ + 75, + 407, + 468, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF Acceleration and Compression. NeRF demonstrates exceptional performance in generating images from arbitrary viewpoints, but it suffers from slow rendering efficiency. Some methods focus on integrating a compact structure with a simplified MLP, which reduces the complexity of MLP calculations in traditional NeRFs. Key strategies have been explored, including the employment of voxel grid [26, 55], octrees [12, 64, 76], tri-planes [4], hashing encoding [40], codebook [27, 56], tensor decomposition [6, 57] to accomplish this. Using explicit structure makes training and rendering faster, but it also means that the 3D structure takes up more storage space.", + "bbox": [ + 75, + 579, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, while methods such as CP-decomposition [6], rank reduction [57], and vector quantization [56] manage to attain modest levels of data compression, their application remains confined to static scenes. Additionally, some approaches [8, 51, 65] opt for post-processing techniques, initially utilizing baseline methods to train a decent NeRF representation, then special encoding and decoding schemes are employed to reduce the storage space. However, these kinds of methods are not hardware-friendly, the inference", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "process is computationally demanding and cannot support real-time streaming on mobile devices. In contrast, our VideoRF uses codec tailored training, hardware-friendly decoding and shader rendering, which enables real-time dynamic radiance field for long sequences with large motion.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VideoRF Representation", + "text_level": 1, + "bbox": [ + 498, + 181, + 730, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate dynamic radiance field rendering on mobile devices, we adopt a representation that aligns with the video codec format and maintains low computational complexity for shader-based rendering. We propose turning the 3D volume representation into a 2D formulation, coupled with a highly efficient rendering pipeline.", + "bbox": [ + 496, + 208, + 890, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our approach, each frame of the radiance field is represented as a feature image $\\mathbf{I}$ where the first channel stores density and the remaining $h$ channels store feature. As depicted in Fig. 2, given a 3D vertex position $\\mathbf{x}$ , we retrieve its density $\\sigma$ and feature $\\mathbf{f}$ using the equation:", + "bbox": [ + 496, + 300, + 890, + 376 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma , \\mathbf {f} = \\mathbf {I} [ \\mathbf {M} (\\mathbf {x}) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 388, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ is the 3D-to-2D mapping table from Sec. 4.1. This mapping table not only effectively excludes empty space to reduce storage but also specifics mapping from each non-empty 3D vertex to a corresponding 2D pixel. The lookup operation is highly efficient, with a time complexity of $O(1)$ , facilitating rapid and convenient queries. It's worth noting that our sequence of feature images is in a 2D format, which is friendly to the video codec hardware.", + "bbox": [ + 496, + 417, + 890, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rendering, inspired by [17, 50], we use a deferred rendering model. We first accumulate the features along the ray:", + "bbox": [ + 496, + 539, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {f}} (\\mathbf {r}) = \\sum_ {k = 1} ^ {n _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) \\mathbf {f} _ {k}, \\\\ T _ {k} = \\exp \\left(- \\sum_ {j = 1} ^ {k - 1} \\sigma_ {j} \\delta_ {j}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 598, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n_{s}$ is the number of sample points along the ray $\\mathbf{r}$ . $\\sigma_{k}$ , $\\mathbf{f}_{k}$ , $\\delta_{k}$ denotes the density, feature of the samples, and the interval between adjacent samples respectively. The view-dependent color of the ray is then computed using a tiny global MLP $\\Phi$ shared across the frames as:", + "bbox": [ + 496, + 700, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {C} (\\mathbf {r}) = \\operatorname {s i g m o i d} \\left(\\Phi (\\tilde {\\mathbf {f}} (\\mathbf {r}), \\mathbf {d})\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 787, + 890, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{d}$ is the view direction of the ray after positional encoding [39]. In this way, the computational burden is significantly reduced as each ray requires only a single MLP decoding which can be implemented in a Shader for real-time rendering on mobile devices.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "472", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/876dabac283503e3ed63fa3d8ca088da565c86e5d9d56dcb85b74472b4140246.jpg", + "image_caption": [ + "Figure 3. Overview of our video codec-friendly training. First, we apply our grid-based coarse training [55] to generate per-frame occupancy grid $\\mathbf{O}^t$ . Then, during baking, we adaptively group each frame and create a mapping table $\\mathbf{M}$ for each group. Next, we sequentially train each feature image $\\mathbf{I}^t$ through our spatial, temporal and photometric loss. Finally, feature images are compressed into the feature video streaming to the player." + ], + "image_footnote": [], + "bbox": [ + 81, + 70, + 893, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Video Codec-friendly Training", + "text_level": 1, + "bbox": [ + 76, + 366, + 359, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we propose a training scheme, as depicted in Fig. 3, to achieve a high compression rate by maintaining spatial and temporal consistency. For spatial aspect, we incorporate 3D-2D Morton sorting to preserve 3D continuity and apply a spatial consistency loss directly on 2D feature images to enforce spatial coherence. On the temporal front, we employ adaptive grouping, which allows frames within a group to share mappings, thereby reducing temporal disruptions, and temporal consistency loss, which serves to further reinforce temporal coherence. These consistencies are crucial which provide a stable and predictable layout for video codecs.", + "bbox": [ + 75, + 392, + 473, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Baking and Map Generation", + "text_level": 1, + "bbox": [ + 76, + 584, + 333, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse stage pre-training. Given multiview images of each frame, we first adopt an off-the-shelf approach [55] to generate the explicit density grid $\\mathbf{V}_{\\sigma}^{t}$ for each frame $t$ independently. We then create a per-frame occupancy grid $\\mathbf{O}^t$ by masking in voxels whose opacity exceeds a certain threshold $\\gamma$ . This coarse stage sets the foundation for our subsequent adaptive grouping in the baking stage.", + "bbox": [ + 75, + 607, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive group. Generating an independent mapping table for each frame $t$ could disrupt temporal continuity, leading to increased storage requirements after video codec compression. This issue arises because the same 3D point in adjacent frames might map to vastly different 2D positions on the feature image. On the other hand, uniformly applying a single mapping table across all frames could introduce significant spatial redundancy. Considering that a 3D point may only be occupied at sparse intervals, it could remain underutilized for most of the duration, resulting in inefficiency. To balance these factors, we divide the sequence into Groups of Frames (GOFs), maintaining a fixed resolu", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion for our 2D feature image and adaptively determining the number of frames in each group. For a set of consecutive frames $\\{i,i + 1,\\dots ,i + n\\}$ , we identify the maximum frame number $\\alpha$ such that the number of occupied voxels in the union from $i$ to $\\alpha$ does not exceed our pixel limit $\\theta$ :", + "bbox": [ + 496, + 367, + 893, + 444 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a r g m a x} _ {\\alpha} \\mathrm {g} \\left(\\bigcup_ {j = i} ^ {\\alpha} \\mathbf {O} ^ {j}\\right) \\leqslant \\theta , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 455, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{g}()$ means the number of occupied grids in the union occupancy grid. Then, we will set the frame $i$ to frame $\\alpha$ as a GOF and frame $\\alpha + 1$ to be the start frame of a new GOF. In this way, all the frames in the group share the same mapping. The same 3D position within the group will be mapped to the same 2D pixel location which keeps temporal consistency and saves the storage.", + "bbox": [ + 496, + 516, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our grouping strategy, unlike the fixed feature grid grouping in ReRF [65], is adaptive, aligning with our objectives. This flexibility allows frames within a group to share a mapping table, optimized in size for a perfect balance between storage efficiency and maintaining temporal continuity. On the other hand, the primary aim of ReRF is to facilitate fast-seeking. Furthermore, our approach diverges from the per-frame occupancy grid of MLP-Maps [47]. We employ a union occupancy grid to calculate the effective number of pixels for adaptive grouping, rather than for speeding up rendering.", + "bbox": [ + 496, + 623, + 893, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Morton sorting. After the union occupancy grid is achieved, we apply Morton sorting to record its 3D spatial continuity. Morton ordering [67], or Z-ordering, interleaves the binary representations of spatial coordinates, ensuring that spatially proximate entities remain adjacent in linear space. As shown in Fig. 4(a), we apply 3D Morton sorting to the vertices based on their position coordinates of the", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "473", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7c3df9f49c7f60e3955d88805590c6e82c9e7953b2859097fce2fcfb1f55f0bb.jpg", + "image_caption": [ + "(a)", + "(b)", + "(c)", + "Figure 4. Illustration of our mapping table generation. We first perform 3D Morton sorting (a) on each nonempty vertex and group it into chunks $C_i$ (b). Next, we lay out each chunk into each block $B_i$ of the feature image, arranged in 2D Morton order (c) within it." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 450, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "union occupancy grid. Vertices with a density below the threshold $\\gamma$ are excluded. This process effectively maintains 3D spatial consistency in the ordering.", + "bbox": [ + 75, + 319, + 468, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2D block partitioning and 2D Morton sorting. To preserve the 3D spatial continuity within a 2D framework, we employ 2D Morton sorting and partition the feature images into blocks. This approach aligns with the block-wise compression of frames in video CODECs, where blocks with local smoothness lead to more efficient storage. Specifically, we first divide the feature image into $N \\times 8$ blocks, denoted as $B_{i}$ , and correspondingly group the sorted vertices into $N \\times 8$ chunks, denoted as $C_i$ , as illustrated in Fig. 4(b). For each pixel $\\mathbf{p}$ in a block $B_{i}$ , we sort its relative position $(u,v)$ in 2D Morton order. As shown in Fig. 4(c), each chunk $C_i$ is then mapped to a block $B_{i}$ , arranged in 2D Morton order within the block, to form the mapping table M. This 2D Morton ordering ensures that sorted values in the 3D Morton ordering are positioned closely within each feature image block, facilitating efficient compression during the transformation process.", + "bbox": [ + 75, + 369, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Fine-grained Sequential Training", + "text_level": 1, + "bbox": [ + 76, + 643, + 369, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the aid of our mapping table, we can sequentially train our feature images within each group through spatial consistency loss applied to 2D feature images and temporal consistency loss between frames.", + "bbox": [ + 75, + 667, + 468, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial consistency loss. In video encoding, regions with homogeneous characteristics demonstrate high compression efficiency. This efficiency stems from the minimal variation in pixel values within these areas, leading to a significant reduction in high-frequency components post Discrete Cosine Transform (DCT) within the video codec. Consequently, these regions retain a higher proportion of low-frequency components. Furthermore, the quantized coefficients in such homogeneous regions are more likely to contain a greater number of zero values, facilitating a more compact data representation and reducing the overall data", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "volume. In order to enhance the homogeneity of our 2D feature image, we introduce a total variance loss, $\\mathcal{L}_{\\mathrm{spatial}}$ during our fine-grained training stage. For each channel of the feature image $\\mathbf{I}$ , we enforce its local smoothness by:", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {\\mathbf {p} \\in \\mathcal {V}} \\left(\\Delta_ {u} (\\mathbf {p}) + \\Delta_ {v} (\\mathbf {p})\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 161, + 890, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ is the pixels of the feature image, $\\Delta_u(\\mathbf{p})$ shorthand for Manhattan distance between the feature value at pixel $\\mathbf{p} := (u,v)$ and the feature value at pixel $(u + 1,v)$ normalized by the resolution, and analogously for $\\Delta_v(\\mathbf{p})$ . By increasing the spatial sparsity, the storage of feature videos after video encoding is decreased at the same quality.", + "bbox": [ + 498, + 209, + 890, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal consistency loss. A naive per-frame training scheme will disrupt temporal continuities by failing to incorporate inter-frame information and resulting in a high bitrate. This is because the residuals between frames are stored after entropy encoding. To optimize storage efficiency, we focus on minimizing the differences in the feature space between the adjacent frames to reduce the entropy. During our sequential training, we enhance interframe similarities by regularizing the current feature image with its predecessor, except for the initial frame of each adaptive group. This is achieved by applying", + "bbox": [ + 498, + 306, + 890, + 473 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p o r a l}} = \\left\\| \\mathbf {I} ^ {t} - \\mathbf {I} ^ {t - 1} \\right\\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 482, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for each frame $t$ in the group, ensuring small residuals between consecutive feature images. By employing temporal smoothness in this manner, we can further mitigate entropy throughout the video codec process, thereby facilitating storage conservation.", + "bbox": [ + 498, + 511, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training objective. Our total loss function is formulated as:", + "bbox": [ + 500, + 592, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r g b}} + \\lambda_ {s} \\mathcal {L} _ {\\text {s p a t i a l}} + \\lambda_ {t} \\mathcal {L} _ {\\text {t e m p o r a l}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 633, + 890, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{s}$ and $\\lambda_{t}$ are the weights for our regular terms and $\\mathcal{L}_{\\mathrm{rgb}}$ is the photometric loss,", + "bbox": [ + 500, + 662, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\mathbf {c} (\\mathbf {r}) - \\hat {\\mathbf {c}} (\\mathbf {r}) \\| ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 703, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the set of training pixel rays; $\\mathbf{c}(\\mathbf{r})$ and $\\hat{\\mathbf{c}} (\\mathbf{r})$ are the ground truth color and predicted color of a ray $\\mathbf{r}$ respectively.", + "bbox": [ + 498, + 747, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. VideoRF Player", + "text_level": 1, + "bbox": [ + 500, + 801, + 658, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we implement a companion VideoRF player to stream and render dynamic radiance fields on mobile devices. Given our feature streams, we quantize them into uint8 format and use H.264 for video codec. Different from [50], we save the mapping table in a 2D-to-3D format as an", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "474", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65122e77c6a6d2ded84b7ed638cda4df2dd222774685de2084abef1cd65eaab6.jpg", + "image_caption": [ + "Figure 5. Our VideoRF method generates results for inward-facing, $360^{\\circ}$ video sequences featuring human-object interactions with large motion. The images in the last row illustrate our ability to implement variable bitrate in these sequences." + ], + "image_footnote": [], + "bbox": [ + 78, + 55, + 883, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RGB image to conserve bitrate during streaming. Then, to enable rapid rendering, we first use this mapping table to recover a 3D volume. We employ a compute shader for efficient processing, segmenting the $512 \\times 512$ mapping table into $16 \\times 16$ workgroups, each handling a $32 \\times 32$ pixel section. The highly parallel architecture of compute shaders enables us to efficiently convert 2D features back into a 3D volume. For the rendering part, we implement it via a fragment shader. For fast raymarching, we employ a multiresolution hierarchy of occupancy grids for each group to skip empty space at different levels. We leverage matrix multiplication in the shader to simulate our tiny MLP calculation. These techniques increase the overall speed while ensuring compact storage.", + "bbox": [ + 75, + 518, + 472, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The VideoRF player marks a significant milestone, as it first enables users to experience real-time rendering of dynamic radiance fields of any length. Within this player, users can drag, rotate, pause, play, fast forward/backward, seek dynamic scenes, or switch between different resolutions like watching online video, offering an extraordinary high-quality free-viewpoint viewing experience. This capability also extends across a wide range of devices, from smartphones and tablets to laptops and desktops, broadening the accessibility and applicability of dynamic radiance fields.", + "bbox": [ + 75, + 734, + 472, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 517, + 705, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our VideoRF on a variety of challenging scenarios. We use the PyTorch Framework to train the model on a single NVIDIA GeForce RTX3090. Our new captured dynamic datasets contain around 80 views at the resolution of $1920 \\times 1080$ at 30 fps. To ensure the robustness of the algorithm, we also use the ReRF dataset and the HumanRF dataset for result demonstration in Fig. 5. We can render images for immersive, $360^{\\circ}$ video sequences that capture human interactions with objects, especially for large motion and long duration sequences. Please refer to the supplementary video for more video results.", + "bbox": [ + 496, + 544, + 893, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison", + "text_level": 1, + "bbox": [ + 500, + 739, + 632, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To the best of our knowledge, our approach is the first real-time dynamic radiance field approach that can decode and render on mobile devices. Therefore, we compare to existing dynamic neural rendering methods, including ReRF [65], HumanRF [18], TiNeuVox [11] and per-frame static reconstruction methods like MeRF [50]. We use ReRF [65] dataset and HumanRF [18] Actor3 sequence 1 for a fair comparison, both in storage memory and image quality. As shown in Fig. 6, though TiNeuVox [11] has small storage,", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "475", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a8073b1fdf35231adfe127754526320994b62dfd6ddcaf3f408c96e26066a5a.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods." + ], + "image_footnote": [], + "bbox": [ + 89, + 74, + 883, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80fafe97f806c06cb36e5cb1cca94d08d1361a7aae3888e7c80f99d8b9f91bff.jpg", + "table_caption": [ + "best second-best" + ], + "table_footnote": [], + "table_body": "
DatasetMethodPSNR↑SSIM↑LPIPS ↓Size(MB)↓
ReRFReRF [65]31.840.9740.0420.645
MeRF [50]31.120.9750.0307.529
HumanRF [18]28.820.9000.0692.800
TiNeuVox[11]22.700.9230.0830.337
Ours32.010.9760.0230.658
Actors-HQReRF [65]28.330.8360.2960.554
MeRF [50]27.220.8070.2717.164
HumanRF [18]28.980.8880.1512.286
TiNeuVox [11]22.980.7520.4300.447
Ours28.460.8380.2780.426
", + "bbox": [ + 78, + 347, + 483, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods.", + "bbox": [ + 76, + 503, + 470, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "it faces the challenge of intensifying blurring effects as the frame count rises. MeRF [50] suffers from the large storage which is not suitable for dynamic scenes when streaming and decoding for playing. HumanRF [18] though capable of representing dynamic scenes, also faces difficulties in streaming and rendering on mobiles due to its computation and storage load.", + "bbox": [ + 75, + 549, + 468, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct a quantitative comparison using metrics such as the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) as metrics shown in Tab. 1. For a fair comparison in our experiments, in the ReRF dataset, we test the first 200 frames on the scene Kpop. We use 6 and 39 as test views and the others as training views. As for the HumanRF dataset, we apply the test methods as referenced in HumanRF on the first 250 frames on Actor3, sequence 1. We utilize the values reported for HumanRF and TiNeuVox methods directly from the HumanRF study. In the ReRF dataset, which includes large motion scenarios, we match the rendering quality of the standard ReRF, while also maintaining a compact data storage size. Meanwhile, in the HumanRF dataset with relatively small motion, we achieve a rendering quality that is second only", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12c4303ccda2249dafd84251befd6522c39d49094c03f1941d40e2630f332443.jpg", + "image_caption": [ + "Figure 7. Rate distortion curve. The rate distortion curve illustrates the efficiency of various components within our system. We use different quantization factors to obtain the average rendering quality at different capacities. Our full model stands out as the most compact, allowing for flexible bitrate adjustments to meet diverse storage needs." + ], + "image_footnote": [], + "bbox": [ + 506, + 333, + 888, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to HumanRF, accomplished with the most efficient storage utilization. Moreover, as indicated in Tab. 2, our method significantly surpasses ReRF in rendering speed when rendering the Kpop scene at an i7-12700F CPU and NVIDIA RTX3090 GPU at the resolution of $1920 \\times 1080$ , even on a tablet (iPad Pro with an M2 chip). It's worth noting that, our method is the only method that can provide both mobile and dynamic rendering tasks. The existing dynamic reconstruction methods like TiNeuVox [11], HumanRF [18] and ReRF [65] all cannot effectively utilize hardware encoding and decoding techniques, preventing them from being displayed on mobile devices, especially for the dynamic scene of a long sequence.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "476", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3129818a6b6fc71192d6c4d25a58775283e2a37968c10317243ad02f6fcf162.jpg", + "image_caption": [ + "Figure 8. Qualitative evaluation of different variations in our method at 600KB." + ], + "image_footnote": [], + "bbox": [ + 88, + 89, + 883, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 261, + 196, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We analyze the impact of spatial consistency, temporal consistency, Morton sorting, block storage, and adaptive grouping on compression and rendering. In the case of models that do not employ 3D and 2D Morton sorting, spatial point data is sorted sequentially in row-major order while employing 2D block-wise storage. For models not utilizing block storage, a 3D spatial Morton sorting is applied, followed by storage in a row-major format within the 2D feature space. For models without adaptive grouping, the mapping table for each frame is calculated only based on the occupancy grid of the current frame. As illustrated in Fig. 7, our full model exhibits the best performance in terms of rendering quality and storage efficiency. By utilizing our modules, the quantization of residual frequency coefficients becomes more precise, retaining more information at the same bitrate. This results in less performance loss when compressed to the same size compared to the variations without them. While ReRF [65] exhibits slightly better quality in the storage range of 700-800KB, our method consistently demonstrates higher PSNR across most capacity ranges and supports mobile rendering. Fig. 8 shows that under a 600KB storage limit, our complete model yields more realistic results with less compression blur.", + "bbox": [ + 76, + 290, + 472, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross device runtime analysis. We evaluate the runtime breakdown analysis of VideoRF as detailed in Tab. 2. Our experimental setup includes a Desktop with an i7-12700F CPU and NVIDIA RTX3090 GPU, a Laptop with an i5-1135G7 CPU and Integrated GPU, a Tablet (iPad Pro) with an M2 chip, and a phone (iPhone 14 Pro) with an A16 Bionic chip. For desktops and Laptops, we employ a Python code base along with ModernGL. Meanwhile, for Tablet and Smartphone, the VideoRF player is developed using Swift and Metal. It's worth noting that the video decoding part primarily relies on CPU performance, while the rendering part mainly depends on GPU performance. These two parts operate asynchronously and simultaneously. The other part mainly covers operations such as data conversion between the GPU and CPU. Our results demonstrate that VideoRF allows users to enjoy free-view videos at high", + "bbox": [ + 76, + 659, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/50497defc30fd94f2f6793b569b39366cf573910b1ca403735318a6a2e1b9405.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDeviceFPSDecodingRenderingOthers
OursDesktop1165.548 ms3.602 ms3.072 ms
OursLaptop2510.54 ms29.31 ms11.69 ms
OursTablet403.822 ms23.86 ms1.150 ms
OursPhone2313.38 ms40.82 ms3.003 ms
ReRFDesktop445.74 ms194.7 ms—
", + "bbox": [ + 504, + 258, + 885, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Runtime analysis across different devices when processing the same HD image with a resolution of $1920 \\times 1080$ .", + "bbox": [ + 500, + 363, + 890, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frame rates on multiple devices, providing an experience that rivals the smoothness of watching 2D videos on platforms such as YouTube.", + "bbox": [ + 498, + 405, + 890, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 467, + 617, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented VideoRF, a novel approach enabling real-time streaming and rendering of dynamic radiance fields on mobile devices. Our VideoRF innovatively processes feature volumes as 2D feature streams and adopts deferred rendering to effectively leverage classical video codecs and rasterization pipelines. Our video codec friendly training scheme is implemented on 2D feature space to enhance spatial-temporal consistency for compactness. Additionally, our tailored player supports seamless streaming and rendering of dynamic radiance fields across a range of devices, from desktops to smartphones. Our experiments demonstrate its capability for compact and effective dynamic scene modeling. With the unique ability of real-time rendering of dynamic radiance fields on mobile devices, we believe that our approach marks a significant step forward in neural scene modeling and immersive VR/AR applications.", + "bbox": [ + 496, + 492, + 890, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 750, + 687, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Key R&D Program of China (2022YFF0902301), Shanghai Local college capacity building program (22010502800). We also acknowledge support from Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai). This work was supported by the programs of NSFC (61976138 and 61977047) and STCSM (2015F0203-000-06). We thank Siyan Zhuo for her assistance in the figures.", + "bbox": [ + 496, + 773, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "477", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures, author=Zhiqin Chen and Thomas Funkhouser and Peter Hedman and Andrea Tagliasacchi. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[2] Aljaž Božić, Denis Gladkov, Luke Doukakis, and Christoph Lassner. Neural assets: Volumetric object capture and rendering for interactive environments, 2022. 3", + "[3] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 2", + "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3", + "[5] Anpei Chen, Minye Wu, Yingliang Zhang, Nianyi Li, Jie Lu, Shenghua Gao, and Jingyi Yu. Deep surface light fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 1(1):1-17, 2018. 2", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 3", + "[7] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2", + "[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing Explicit Voxel Grid Representations: Fast NeRFs Become Also Small. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1236-1245, 2023. 3", + "[9] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6): 246:1-246:16, 2017. 2", + "[10] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2", + "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2, 6, 7", + "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3", + "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes:" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2", + "[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, 2021. 2", + "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 2", + "[16] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 2", + "[17] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855-5864, 2021. 3", + "[18] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. HumanRF: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 6, 7", + "[19] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6155-6165, 2022. 2", + "[20] Yuheng Jiang, Kaixin Yao, Zhuo Su, Zhehao Shen, Haimin Luo, and Lan Xu. Instant-nvr: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 595-605, 2023.", + "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8320-8329, 2018. 2", + "[22] Moritz Kappel, Vladislav Golyanik, Susana Castillo, Christian Theobalt, and Marcus Magnor. Fast non-rigid radiance fields from monocularized data, 2023. 2", + "[23] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3- d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2", + "[24] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2", + "[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis, 2023. 2", + "[26] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthe" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "478", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3", + "[27] Lingzhi Li, Zhongshu Wang, Zhen Shen, Li Shen, and Ping Tan. Compact real-time radiance fields with neural codebook, 2023. 3", + "[28] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 2", + "[29] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2", + "[30] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Learning the depths of moving people by watching frozen people. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4516-4525, 2019. 2", + "[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6498-6508, 2021. 2", + "[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering, 2023.", + "[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2", + "[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In Advances in Neural Information Processing Systems, 2022. 1, 2", + "[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), 2022. 2", + "[36] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (TOG), 39(4):71-1, 2020. 2", + "[37] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2", + "[38] Marko Mihajlovic, Sergey Prokudin, Marc Pollefeys, and Siyu Tang. Resfields: Residual neural fields for spatiotemporal signals. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3", + "[41] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2", + "[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1", + "[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph., 40(6), 2021. 2", + "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2", + "[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2", + "[46] Sida Peng, Chen Geng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Implicit neural representations with structured latent codes for human body modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9895-9907, 2023. 2", + "[47] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In CVPR, 2023. 2, 4", + "[48] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2", + "[49] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2", + "[50] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. SIGGRAPH, 2023. 2, 3, 5, 6, 7", + "[51] Daniel Rho, Byeonghyeon Lee, Seungtae Nam, Joo Chan Lee, Jong Hwan Ko, and Eunbyung Park. Masked wavelet representation for compact neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20680-20690, 2023. 2, 3", + "[52] Sara Rojas, Jesus Zarzar, Juan C. Pérez, Artsiom Sanakoyeu, Ali Thabet, Albert Pumarola, and Bernard Ghanem. Re" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "479", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ReND: Real-Time Rendering of NeRFs across Devices. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3632-3641, 2023. 2, 3", + "[53] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering, 2023. 2", + "[54] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. NeRFPlayer: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields, year=2023. IEEE Transactions on Visualization and Computer Graphics, 29 (5):2732-2742. 2", + "[55] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5459, 2022. 3, 4, 1", + "[56] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3", + "[57] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via Rank-residual Decomposition. In Advances in Neural Information Processing Systems, 2022. 3", + "[58] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate Textured Mesh Recovery from NeRF via Adaptive Surface Refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17739-17749, 2023. 3", + "[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2", + "[60] Edith Tretschk, Vladislav Golyanik, Michael Zollhöfer, Aljaz Bozic, Christoph Lassner, and Christian Theobalt. SceNeRFlow: Time-Consistent Reconstruction of General Dynamic. In International Conference on 3D Vision (3DV), 2024. 2", + "[61] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 2", + "[62] Ziyu Wan, Christian Richardt, Aljaž Božić, Chao Li, Vijay Rengarajan, Seonghyeon Nam, Xiaoyu Xiang, Tuotuo Li, Bo Zhu, Rakesh Ranjan, and Jing Liao. Learning neural duplex radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8307-8316, 2023. 3", + "[63] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2, 3", + "[65] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 76-87, 2023. 2, 3, 4, 6, 7, 8, 1", + "[66] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[67] Wikipedia. Z-order curve, 2024. 4", + "[68] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 287-296, 2000. 2", + "[69] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2", + "[70] Lan Xu, Zhuo Su, Lei Han, Tao Yu, Yebin Liu, and Lu Fang. Unstructuredfusion: Realtime 4d geometry and texture reconstruction using commercial rgbd cameras. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2508-2522, 2020. 2", + "[71] Lan Xu, Wei Cheng, Kaiwen Guo, Lei Han, Yebin Liu, and Lu Fang. Flyfusion: Realtime dynamic scene reconstruction using a flying depth camera. IEEE Transactions on Visualization and Computer Graphics, 27(1):68-82, 2021. 2", + "[72] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[73] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[74] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. 2023. 2", + "[75] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3", + "[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "480", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 5752-5761, 2021. 3", + "[77] Heng Yu, Joel Julin, Zoltán Á. Milacski, Koichiro Niinuma, and László A. Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12397-12406, 2023. 2", + "[78] Anlan Zhang, Chendong Wang, Bo Han, and Feng qian. Yuzu:neural-enhanced volumetric video streaming. 2022. 3", + "[79] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2", + "[80] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2022. 2", + "[81] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2", + "[82] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Transactions on Graphics (TOG), 42(4), 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 470 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "481", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_model.json b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a03fcdb5d2ad88efafd0a90942117fba9f98679c --- /dev/null +++ b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_model.json @@ -0,0 +1,2567 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.18, + 0.224, + 0.2 + ], + "angle": 0, + "content": "Liao Wang\\(^{1,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.18, + 0.38, + 0.198 + ], + "angle": 0, + "content": "Kaixin Yao\\(^{1,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.18, + 0.566, + 0.2 + ], + "angle": 0, + "content": "Chengcheng Guo1" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.181, + 0.721, + 0.2 + ], + "angle": 0, + "content": "Zhirui Zhang" + }, + { + "type": "text", + "bbox": [ + 0.762, + 0.181, + 0.86, + 0.2 + ], + "angle": 0, + "content": "Qiang Hu4#" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.206, + 0.358, + 0.224 + ], + "angle": 0, + "content": "Jingyi Yu" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.206, + 0.515, + 0.222 + ], + "angle": 0, + "content": "\\(\\operatorname{Lan} \\mathrm{Xu}^{1 \\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.596, + 0.206, + 0.696, + 0.224 + ], + "angle": 0, + "content": "Minye \\(\\mathsf{W}\\mathsf{u}^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.224, + 0.308, + 0.242 + ], + "angle": 0, + "content": "1 ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.225, + 0.455, + 0.24 + ], + "angle": 0, + "content": "2 KU Leuven" + }, + { + "type": "text", + "bbox": [ + 0.496, + 0.225, + 0.579, + 0.24 + ], + "angle": 0, + "content": "3 NeuDim" + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.225, + 0.875, + 0.242 + ], + "angle": 0, + "content": "\\(^{4}\\) Shanghai Jiao Tong University" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.25, + 0.387, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.274, + 0.688, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.251, + 0.88, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.327, + 0.845, + 0.337 + ], + "angle": 0, + "content": "iPad Pro:Apple M2 chip" + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.337, + 0.865, + 0.345 + ], + "angle": 0, + "content": "1920x1080, 40FPS, 0.71MB/frame" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.351, + 0.879, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.428, + 0.867, + 0.437 + ], + "angle": 0, + "content": "Dell: i5-1135G7 CPU, Integrated GPU" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.437, + 0.865, + 0.446 + ], + "angle": 0, + "content": "1200x800, 49FPS, 0.65MB/frame" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.454, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Figure 1. Our proposed VideoRF views dynamic radiance field as 2D feature video streams combined with deferred rendering. This technique facilitates hardware video codec and shader-based rendering, enabling smooth high-quality rendering across diverse devices." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.51, + 0.313, + 0.526 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.542, + 0.471, + 0.844 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) excel in photorealistically rendering static scenes. However, rendering dynamic, long-duration radiance fields on ubiquitous devices remains challenging, due to data storage and computational constraints. In this paper, we introduce VideoRF, the first approach to enable real-time streaming and rendering of dynamic human-centric radiance fields on mobile platforms. At the core is a serialized 2D feature image stream representing the 4D radiance field all in one. We introduce a tailored training scheme directly applied to this 2D domain to impose the temporal and spatial redundancy of the feature image stream. By leveraging the redundancy, we show that the feature image stream can be efficiently compressed by 2D video CODECs, which allows us to exploit video hardware accelerators to achieve real-time decoding. On the other hand, based on the feature image stream, we propose a novel rendering pipeline for VideoRF, which has specialized space mappings to query radiance properties efficiently. Paired with a deferred shading model, VideoRF has the capability of real-time rendering on mobile devices" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.47, + 0.902 + ], + "angle": 0, + "content": "* Authors contributed equally to this work. † The corresponding authors are Minye Wu (minye.wu@kuleuven.be) and Lan Xu (xulan1@shanghaiitech.edu.cn). # Work done while at ShanghaiTech University." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.587 + ], + "angle": 0, + "content": "thanks to its efficiency. We have developed a real-time interactive player that enables online streaming and rendering of dynamic scenes, offering a seamless and immersive free-viewpoint experience across a range of devices, from desktops to mobile phones. Our project page is available at https://aoliaol2138.github.io/VideoRF/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.618, + 0.631, + 0.633 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Photorealistic Free-Viewpoint Video (FVV) of dynamic scenes offers an immersive experience in virtual reality and telepresence. Work involving Neural Radiance Fields (NeRFs) has shown great potential in creating photorealistic Free-Viewpoint Videos (FVVs). However, there are still challenges in smoothly delivering and rendering FVVs using NeRFs on commonly used devices, similar to the ease of watching online videos. The difficulty lies in reducing the data capacity for transmitting and storing long sequences and ensuring a low, mobile-compatible computational load." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) [39] surpasses traditional 3D reconstruction methods in photorealistic novel view synthesis. Several works extend NeRF to dynamic scenes by maintaining a canonical space and matching it implicitly [10, 42, 48] or explicitly [34] to align with each frame's live space. However, their dependence on canonical space limits effectiveness in sequences with large motions or topo" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.395 + ], + "angle": 0, + "content": "logical changes. Other methods introduce new representations like 4D feature grids [18] or temporal voxel features [11], achieving impressive results on the scenes with topological transformations. Yet, [11] struggles with representing longer sequences owing to model capacity constraints, while [18] encounters streaming challenges due to large storage needs. Recent efforts [26, 54, 65] focus on compressing dynamic frames for streaming, but their computational intensity limits mobile applicability. Concurrently, on mobile devices, real-time rendering of static scenes has been achieved by baking NeRF into mesh templates [1, 52, 75] or texture assets [50]. However, these techniques fall short for dynamic scenes as their per-frame representation becomes too bulky for real-time loading. Moreover, while NeRF compression methods [51, 56, 65] can be employed, they introduce decoding or rendering overheads unsuitable for mobile platforms. Despite the existence of NeRF solutions for dynamic scenes and mobile optimization, a cohesive approach that effectively addresses both still poses a difficulty." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.686 + ], + "angle": 0, + "content": "In this paper, we propose VideoRF - a novel neural modeling approach that enables real-time streaming and rendering of human-centric dynamic radiance fields on common mobile devices (see Fig. 1). Our key idea is to view the 4D feature volumes reconstructed from a dynamic scene as a 2D feature image stream, which is friendly to video codec. Each feature image records the densities and appearance features of one frame. We hence propose a rendering pipeline for this representation. In this rendering pipeline, VideoRF uses a set of mapping tables to connect the 3D space with the 2D feature images. This allows for O(1) density and feature retrieval from the images for every 3D sample point in the space. To reduce the computational complexity of rendering, we adopt the deferred shading model [50] with a global tiny MLP to decode the integrated feature vectors into pixel colors. All rendering operations here are low-cost and compatible with the fragment Shader model, making it possible to implement this rendering pipeline on various devices that have GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Second, we present a sequential training scheme to effectively generate the 2D feature image stream. Specifically, VideoRF adaptively groups frames by analyzing motion from the sequential data to ensure temporal stability for mapping table generation. We deploy 3D and 2D Morton sorting techniques to improve spatial consistency in the mapping. VideoRF contains spatial and temporal continuity regularizers that are applied on the mapped 2D feature images using the mapping table. This training strategy enforces the temporal and spatial redundancy to the 2D feature image stream. We show that our feature image stream with spatiotemporal sparsity can be efficiently compressed by off-the-shelf video codecs and reaches high compression rates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.212 + ], + "angle": 0, + "content": "Moreover, based on these findings, we build a cross-platform player based on VideoRF that can play FVVs in real-time, supported by video hardware accelerators. In this way, users can interactively drag, rotate, pause, play, fast-forward, rewind, and jump to specific frames, providing a viewing experience as seamless as watching online videos on various devices, including smartphones, tablets, laptops, and desktops which was unseen before." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.213, + 0.794, + 0.227 + ], + "angle": 0, + "content": "To summarize, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.272 + ], + "angle": 0, + "content": "- We propose VideoRF, a novel approach to enable real-time dynamic radiance field decoding, streaming and rendering on mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.273, + 0.892, + 0.332 + ], + "angle": 0, + "content": "- We present an efficient and compact representation, which represents 4D radiance field into 2D feature stream with low rendering complexity to support hardware video codec and Shader rendering." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.333, + 0.892, + 0.379 + ], + "angle": 0, + "content": "- We introduce a training scheme to directly impose spatial-temporal consistency on our 2D feature stream for efficient compression." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.638, + 0.408 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Novel View Synthesis for Dynamic Scenes. Dynamic scenes pose a greater challenge in achieving realistic view synthesis results due to moving objects. One approach is to reconstruct the dynamic scene and render the geometry from new viewpoints. The conventional RGB [7, 23, 30, 36, 37, 49, 80] or RGB-D [9, 19-21, 41, 70, 71] solutions have been extensively investigated. Mesh-based neural network [5, 61, 68] techniques are effective for compact data storage and can record view-dependent texture [5, 68], but such methods heavily rely on geometry, especially perform poorly for topologically complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Many methods extend NeRF into the dynamic view synthesis settings. Some methods [10, 15, 29, 31-33, 43, 48, 63, 69, 77, 79] handle spatial change directly conditions on time and [14, 45, 46, 73, 82] use feature latent code to represent time information. However, they do not support the streaming of long sequences due to the limited representation. Others learn spatial offsets from the live scene to the canonical radiance field by using explicit voxel matching [34], skeletal poses [25, 35, 44], deformed hashtable [24] and deformed volume [22, 59, 60, 66, 72, 81]. While these methods can successfully render dynamic scenes with photo-realistic quality, their heavy reliance on the canonical space makes these methods vulnerable to long sequences, large motions, and topological alterations. Recently, several methods using 4D planes [3, 13, 18, 53, 74], voxel grid [11], Fourier representation [64], residual layers [38] or dynamic MLP maps [47] to represent dynamic scene. However, the inference computational complexity makes it difficult for them to implement streaming and decoding on mobile devices. Additionally, several methods such as [26, 54, 65] have managed to facilitate the streaming of dynamic radi" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.472, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Demonstration of our rendering. For each 3D sample point, its density \\(\\sigma\\) and feature \\(\\mathbf{f}\\) are fetched from the 2D feature image through the mapping table \\(\\mathbf{M}\\). Each point feature is first volumetrically accumulated to get the ray feature \\(\\tilde{\\mathbf{f}}\\) and pass MLP \\(\\Phi\\) to decode the ray color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.389 + ], + "angle": 0, + "content": "ance fields by reducing the capacity of each frame, but the lack of general hardware acceleration makes it challenging to represent dynamic scenes in real-time on mobile phones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Cross Device Neural Radiance Field Rendering. Recent works have demonstrated that static neural radiance fields can be rendered on mobile devices. They achieve this by directly converting the neural radiance field to mesh [58] or using traditional texture to represent the scene. [1, 75, 78] have endeavored to augment the capabilities of neural rendering by employing mesh templates and feature texture, [2, 52] facilitating the rendering process through fragment Shader to achieve surface-like rendering with similar ideas. [62] represents neural radiance features encoded on a two-layer duplex mesh for better rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.763 + ], + "angle": 0, + "content": "NeRF Acceleration and Compression. NeRF demonstrates exceptional performance in generating images from arbitrary viewpoints, but it suffers from slow rendering efficiency. Some methods focus on integrating a compact structure with a simplified MLP, which reduces the complexity of MLP calculations in traditional NeRFs. Key strategies have been explored, including the employment of voxel grid [26, 55], octrees [12, 64, 76], tri-planes [4], hashing encoding [40], codebook [27, 56], tensor decomposition [6, 57] to accomplish this. Using explicit structure makes training and rendering faster, but it also means that the 3D structure takes up more storage space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Therefore, while methods such as CP-decomposition [6], rank reduction [57], and vector quantization [56] manage to attain modest levels of data compression, their application remains confined to static scenes. Additionally, some approaches [8, 51, 65] opt for post-processing techniques, initially utilizing baseline methods to train a decent NeRF representation, then special encoding and decoding schemes are employed to reduce the storage space. However, these kinds of methods are not hardware-friendly, the inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "process is computationally demanding and cannot support real-time streaming on mobile devices. In contrast, our VideoRF uses codec tailored training, hardware-friendly decoding and shader rendering, which enables real-time dynamic radiance field for long sequences with large motion." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.183, + 0.731, + 0.201 + ], + "angle": 0, + "content": "3. VideoRF Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.3 + ], + "angle": 0, + "content": "To facilitate dynamic radiance field rendering on mobile devices, we adopt a representation that aligns with the video codec format and maintains low computational complexity for shader-based rendering. We propose turning the 3D volume representation into a 2D formulation, coupled with a highly efficient rendering pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.377 + ], + "angle": 0, + "content": "In our approach, each frame of the radiance field is represented as a feature image \\(\\mathbf{I}\\) where the first channel stores density and the remaining \\(h\\) channels store feature. As depicted in Fig. 2, given a 3D vertex position \\(\\mathbf{x}\\), we retrieve its density \\(\\sigma\\) and feature \\(\\mathbf{f}\\) using the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.389, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\sigma , \\mathbf {f} = \\mathbf {I} [ \\mathbf {M} (\\mathbf {x}) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.54 + ], + "angle": 0, + "content": "where \\( \\mathbf{M} \\) is the 3D-to-2D mapping table from Sec. 4.1. This mapping table not only effectively excludes empty space to reduce storage but also specifics mapping from each non-empty 3D vertex to a corresponding 2D pixel. The lookup operation is highly efficient, with a time complexity of \\( O(1) \\), facilitating rapid and convenient queries. It's worth noting that our sequence of feature images is in a 2D format, which is friendly to the video codec hardware." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.586 + ], + "angle": 0, + "content": "For rendering, inspired by [17, 50], we use a deferred rendering model. We first accumulate the features along the ray:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.599, + 0.892, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {f}} (\\mathbf {r}) = \\sum_ {k = 1} ^ {n _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) \\mathbf {f} _ {k}, \\\\ T _ {k} = \\exp \\left(- \\sum_ {j = 1} ^ {k - 1} \\sigma_ {j} \\delta_ {j}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.777 + ], + "angle": 0, + "content": "where \\(n_{s}\\) is the number of sample points along the ray \\(\\mathbf{r}\\). \\(\\sigma_{k}\\), \\(\\mathbf{f}_{k}\\), \\(\\delta_{k}\\) denotes the density, feature of the samples, and the interval between adjacent samples respectively. The view-dependent color of the ray is then computed using a tiny global MLP \\(\\Phi\\) shared across the frames as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.788, + 0.892, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\tilde {C} (\\mathbf {r}) = \\operatorname {s i g m o i d} \\left(\\Phi (\\tilde {\\mathbf {f}} (\\mathbf {r}), \\mathbf {d})\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}\\) is the view direction of the ray after positional encoding [39]. In this way, the computational burden is significantly reduced as each ray requires only a single MLP decoding which can be implemented in a Shader for real-time rendering on mobile devices." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.071, + 0.895, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Overview of our video codec-friendly training. First, we apply our grid-based coarse training [55] to generate per-frame occupancy grid \\(\\mathbf{O}^t\\). Then, during baking, we adaptively group each frame and create a mapping table \\(\\mathbf{M}\\) for each group. Next, we sequentially train each feature image \\(\\mathbf{I}^t\\) through our spatial, temporal and photometric loss. Finally, feature images are compressed into the feature video streaming to the player." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.367, + 0.36, + 0.385 + ], + "angle": 0, + "content": "4. Video Codec-friendly Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.474, + 0.573 + ], + "angle": 0, + "content": "In this section, we propose a training scheme, as depicted in Fig. 3, to achieve a high compression rate by maintaining spatial and temporal consistency. For spatial aspect, we incorporate 3D-2D Morton sorting to preserve 3D continuity and apply a spatial consistency loss directly on 2D feature images to enforce spatial coherence. On the temporal front, we employ adaptive grouping, which allows frames within a group to share mappings, thereby reducing temporal disruptions, and temporal consistency loss, which serves to further reinforce temporal coherence. These consistencies are crucial which provide a stable and predictable layout for video codecs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.334, + 0.601 + ], + "angle": 0, + "content": "4.1. Baking and Map Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Coarse stage pre-training. Given multiview images of each frame, we first adopt an off-the-shelf approach [55] to generate the explicit density grid \\(\\mathbf{V}_{\\sigma}^{t}\\) for each frame \\(t\\) independently. We then create a per-frame occupancy grid \\(\\mathbf{O}^t\\) by masking in voxels whose opacity exceeds a certain threshold \\(\\gamma\\). This coarse stage sets the foundation for our subsequent adaptive grouping in the baking stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Adaptive group. Generating an independent mapping table for each frame \\( t \\) could disrupt temporal continuity, leading to increased storage requirements after video codec compression. This issue arises because the same 3D point in adjacent frames might map to vastly different 2D positions on the feature image. On the other hand, uniformly applying a single mapping table across all frames could introduce significant spatial redundancy. Considering that a 3D point may only be occupied at sparse intervals, it could remain underutilized for most of the duration, resulting in inefficiency. To balance these factors, we divide the sequence into Groups of Frames (GOFs), maintaining a fixed resolu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.445 + ], + "angle": 0, + "content": "tion for our 2D feature image and adaptively determining the number of frames in each group. For a set of consecutive frames \\(\\{i,i + 1,\\dots ,i + n\\}\\), we identify the maximum frame number \\(\\alpha\\) such that the number of occupied voxels in the union from \\(i\\) to \\(\\alpha\\) does not exceed our pixel limit \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.456, + 0.892, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a r g m a x} _ {\\alpha} \\mathrm {g} \\left(\\bigcup_ {j = i} ^ {\\alpha} \\mathbf {O} ^ {j}\\right) \\leqslant \\theta , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.623 + ], + "angle": 0, + "content": "where \\( \\mathrm{g}() \\) means the number of occupied grids in the union occupancy grid. Then, we will set the frame \\( i \\) to frame \\( \\alpha \\) as a GOF and frame \\( \\alpha + 1 \\) to be the start frame of a new GOF. In this way, all the frames in the group share the same mapping. The same 3D position within the group will be mapped to the same 2D pixel location which keeps temporal consistency and saves the storage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.79 + ], + "angle": 0, + "content": "Our grouping strategy, unlike the fixed feature grid grouping in ReRF [65], is adaptive, aligning with our objectives. This flexibility allows frames within a group to share a mapping table, optimized in size for a perfect balance between storage efficiency and maintaining temporal continuity. On the other hand, the primary aim of ReRF is to facilitate fast-seeking. Furthermore, our approach diverges from the per-frame occupancy grid of MLP-Maps [47]. We employ a union occupancy grid to calculate the effective number of pixels for adaptive grouping, rather than for speeding up rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3D Morton sorting. After the union occupancy grid is achieved, we apply Morton sorting to record its 3D spatial continuity. Morton ordering [67], or Z-ordering, interleaves the binary representations of spatial coordinates, ensuring that spatially proximate entities remain adjacent in linear space. As shown in Fig. 4(a), we apply 3D Morton sorting to the vertices based on their position coordinates of the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "473" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.452, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.211, + 0.151, + 0.225 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.211, + 0.294, + 0.224 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.212, + 0.401, + 0.224 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Illustration of our mapping table generation. We first perform 3D Morton sorting (a) on each nonempty vertex and group it into chunks \\( C_i \\) (b). Next, we lay out each chunk into each block \\( B_i \\) of the feature image, arranged in 2D Morton order (c) within it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.469, + 0.367 + ], + "angle": 0, + "content": "union occupancy grid. Vertices with a density below the threshold \\(\\gamma\\) are excluded. This process effectively maintains 3D spatial consistency in the ordering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.629 + ], + "angle": 0, + "content": "2D block partitioning and 2D Morton sorting. To preserve the 3D spatial continuity within a 2D framework, we employ 2D Morton sorting and partition the feature images into blocks. This approach aligns with the block-wise compression of frames in video CODECs, where blocks with local smoothness lead to more efficient storage. Specifically, we first divide the feature image into \\( N \\times 8 \\) blocks, denoted as \\( B_{i} \\), and correspondingly group the sorted vertices into \\( N \\times 8 \\) chunks, denoted as \\( C_i \\), as illustrated in Fig. 4(b). For each pixel \\( \\mathbf{p} \\) in a block \\( B_{i} \\), we sort its relative position \\( (u,v) \\) in 2D Morton order. As shown in Fig. 4(c), each chunk \\( C_i \\) is then mapped to a block \\( B_{i} \\), arranged in 2D Morton order within the block, to form the mapping table M. This 2D Morton ordering ensures that sorted values in the 3D Morton ordering are positioned closely within each feature image block, facilitating efficient compression during the transformation process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.644, + 0.37, + 0.66 + ], + "angle": 0, + "content": "4.2. Fine-grained Sequential Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.469, + 0.729 + ], + "angle": 0, + "content": "With the aid of our mapping table, we can sequentially train our feature images within each group through spatial consistency loss applied to 2D feature images and temporal consistency loss between frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Spatial consistency loss. In video encoding, regions with homogeneous characteristics demonstrate high compression efficiency. This efficiency stems from the minimal variation in pixel values within these areas, leading to a significant reduction in high-frequency components post Discrete Cosine Transform (DCT) within the video codec. Consequently, these regions retain a higher proportion of low-frequency components. Furthermore, the quantized coefficients in such homogeneous regions are more likely to contain a greater number of zero values, facilitating a more compact data representation and reducing the overall data" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "volume. In order to enhance the homogeneity of our 2D feature image, we introduce a total variance loss, \\(\\mathcal{L}_{\\mathrm{spatial}}\\) during our fine-grained training stage. For each channel of the feature image \\(\\mathbf{I}\\), we enforce its local smoothness by:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.162, + 0.892, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {\\mathbf {p} \\in \\mathcal {V}} \\left(\\Delta_ {u} (\\mathbf {p}) + \\Delta_ {v} (\\mathbf {p})\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.301 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) is the pixels of the feature image, \\(\\Delta_u(\\mathbf{p})\\) shorthand for Manhattan distance between the feature value at pixel \\(\\mathbf{p} := (u,v)\\) and the feature value at pixel \\((u + 1,v)\\) normalized by the resolution, and analogously for \\(\\Delta_v(\\mathbf{p})\\). By increasing the spatial sparsity, the storage of feature videos after video encoding is decreased at the same quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Temporal consistency loss. A naive per-frame training scheme will disrupt temporal continuities by failing to incorporate inter-frame information and resulting in a high bitrate. This is because the residuals between frames are stored after entropy encoding. To optimize storage efficiency, we focus on minimizing the differences in the feature space between the adjacent frames to reduce the entropy. During our sequential training, we enhance interframe similarities by regularizing the current feature image with its predecessor, except for the initial frame of each adaptive group. This is achieved by applying" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.483, + 0.891, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p o r a l}} = \\left\\| \\mathbf {I} ^ {t} - \\mathbf {I} ^ {t - 1} \\right\\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.512, + 0.891, + 0.587 + ], + "angle": 0, + "content": "for each frame \\( t \\) in the group, ensuring small residuals between consecutive feature images. By employing temporal smoothness in this manner, we can further mitigate entropy throughout the video codec process, thereby facilitating storage conservation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.593, + 0.891, + 0.621 + ], + "angle": 0, + "content": "Training objective. Our total loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.635, + 0.891, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r g b}} + \\lambda_ {s} \\mathcal {L} _ {\\text {s p a t i a l}} + \\lambda_ {t} \\mathcal {L} _ {\\text {t e m p o r a l}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.663, + 0.891, + 0.693 + ], + "angle": 0, + "content": "where \\(\\lambda_{s}\\) and \\(\\lambda_{t}\\) are the weights for our regular terms and \\(\\mathcal{L}_{\\mathrm{rgb}}\\) is the photometric loss," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.704, + 0.891, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\mathbf {c} (\\mathbf {r}) - \\hat {\\mathbf {c}} (\\mathbf {r}) \\| ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.748, + 0.891, + 0.793 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the set of training pixel rays; \\(\\mathbf{c}(\\mathbf{r})\\) and \\(\\hat{\\mathbf{c}} (\\mathbf{r})\\) are the ground truth color and predicted color of a ray \\(\\mathbf{r}\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.659, + 0.818 + ], + "angle": 0, + "content": "4.3. VideoRF Player" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Finally, we implement a companion VideoRF player to stream and render dynamic radiance fields on mobile devices. Given our feature streams, we quantize them into uint8 format and use H.264 for video codec. Different from [50], we save the mapping table in a 2D-to-3D format as an" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "474" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.056, + 0.885, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Figure 5. Our VideoRF method generates results for inward-facing, \\(360^{\\circ}\\) video sequences featuring human-object interactions with large motion. The images in the last row illustrate our ability to implement variable bitrate in these sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.732 + ], + "angle": 0, + "content": "RGB image to conserve bitrate during streaming. Then, to enable rapid rendering, we first use this mapping table to recover a 3D volume. We employ a compute shader for efficient processing, segmenting the \\(512 \\times 512\\) mapping table into \\(16 \\times 16\\) workgroups, each handling a \\(32 \\times 32\\) pixel section. The highly parallel architecture of compute shaders enables us to efficiently convert 2D features back into a 3D volume. For the rendering part, we implement it via a fragment shader. For fast raymarching, we employ a multiresolution hierarchy of occupancy grids for each group to skip empty space at different levels. We leverage matrix multiplication in the shader to simulate our tiny MLP calculation. These techniques increase the overall speed while ensuring compact storage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.9 + ], + "angle": 0, + "content": "The VideoRF player marks a significant milestone, as it first enables users to experience real-time rendering of dynamic radiance fields of any length. Within this player, users can drag, rotate, pause, play, fast forward/backward, seek dynamic scenes, or switch between different resolutions like watching online video, offering an extraordinary high-quality free-viewpoint viewing experience. This capability also extends across a wide range of devices, from smartphones and tablets to laptops and desktops, broadening the accessibility and applicability of dynamic radiance fields." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.518, + 0.707, + 0.535 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.895, + 0.726 + ], + "angle": 0, + "content": "In this section, we evaluate our VideoRF on a variety of challenging scenarios. We use the PyTorch Framework to train the model on a single NVIDIA GeForce RTX3090. Our new captured dynamic datasets contain around 80 views at the resolution of \\(1920 \\times 1080\\) at 30 fps. To ensure the robustness of the algorithm, we also use the ReRF dataset and the HumanRF dataset for result demonstration in Fig. 5. We can render images for immersive, \\(360^{\\circ}\\) video sequences that capture human interactions with objects, especially for large motion and long duration sequences. Please refer to the supplementary video for more video results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.633, + 0.757 + ], + "angle": 0, + "content": "5.1. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To the best of our knowledge, our approach is the first real-time dynamic radiance field approach that can decode and render on mobile devices. Therefore, we compare to existing dynamic neural rendering methods, including ReRF [65], HumanRF [18], TiNeuVox [11] and per-frame static reconstruction methods like MeRF [50]. We use ReRF [65] dataset and HumanRF [18] Actor3 sequence 1 for a fair comparison, both in storage memory and image quality. As shown in Fig. 6, though TiNeuVox [11] has small storage," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "475" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.075, + 0.885, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.311, + 0.855, + 0.326 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.334, + 0.332, + 0.346 + ], + "angle": 0, + "content": "best second-best" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.348, + 0.484, + 0.499 + ], + "angle": 0, + "content": "
DatasetMethodPSNR↑SSIM↑LPIPS ↓Size(MB)↓
ReRFReRF [65]31.840.9740.0420.645
MeRF [50]31.120.9750.0307.529
HumanRF [18]28.820.9000.0692.800
TiNeuVox[11]22.700.9230.0830.337
Ours32.010.9760.0230.658
Actors-HQReRF [65]28.330.8360.2960.554
MeRF [50]27.220.8070.2717.164
HumanRF [18]28.980.8880.1512.286
TiNeuVox [11]22.980.7520.4300.447
Ours28.460.8380.2780.426
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.505, + 0.471, + 0.534 + ], + "angle": 0, + "content": "Table 1. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.656 + ], + "angle": 0, + "content": "it faces the challenge of intensifying blurring effects as the frame count rises. MeRF [50] suffers from the large storage which is not suitable for dynamic scenes when streaming and decoding for playing. HumanRF [18] though capable of representing dynamic scenes, also faces difficulties in streaming and rendering on mobiles due to its computation and storage load." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We also conduct a quantitative comparison using metrics such as the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) as metrics shown in Tab. 1. For a fair comparison in our experiments, in the ReRF dataset, we test the first 200 frames on the scene Kpop. We use 6 and 39 as test views and the others as training views. As for the HumanRF dataset, we apply the test methods as referenced in HumanRF on the first 250 frames on Actor3, sequence 1. We utilize the values reported for HumanRF and TiNeuVox methods directly from the HumanRF study. In the ReRF dataset, which includes large motion scenarios, we match the rendering quality of the standard ReRF, while also maintaining a compact data storage size. Meanwhile, in the HumanRF dataset with relatively small motion, we achieve a rendering quality that is second only" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.334, + 0.889, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.59, + 0.894, + 0.675 + ], + "angle": 0, + "content": "Figure 7. Rate distortion curve. The rate distortion curve illustrates the efficiency of various components within our system. We use different quantization factors to obtain the average rendering quality at different capacities. Our full model stands out as the most compact, allowing for flexible bitrate adjustments to meet diverse storage needs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "to HumanRF, accomplished with the most efficient storage utilization. Moreover, as indicated in Tab. 2, our method significantly surpasses ReRF in rendering speed when rendering the Kpop scene at an i7-12700F CPU and NVIDIA RTX3090 GPU at the resolution of \\(1920 \\times 1080\\), even on a tablet (iPad Pro with an M2 chip). It's worth noting that, our method is the only method that can provide both mobile and dynamic rendering tasks. The existing dynamic reconstruction methods like TiNeuVox [11], HumanRF [18] and ReRF [65] all cannot effectively utilize hardware encoding and decoding techniques, preventing them from being displayed on mobile devices, especially for the dynamic scene of a long sequence." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "476" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.09, + 0.885, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.233, + 0.721, + 0.247 + ], + "angle": 0, + "content": "Figure 8. Qualitative evaluation of different variations in our method at 600KB." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.262, + 0.197, + 0.276 + ], + "angle": 0, + "content": "5.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.291, + 0.473, + 0.653 + ], + "angle": 0, + "content": "Ablation study. We analyze the impact of spatial consistency, temporal consistency, Morton sorting, block storage, and adaptive grouping on compression and rendering. In the case of models that do not employ 3D and 2D Morton sorting, spatial point data is sorted sequentially in row-major order while employing 2D block-wise storage. For models not utilizing block storage, a 3D spatial Morton sorting is applied, followed by storage in a row-major format within the 2D feature space. For models without adaptive grouping, the mapping table for each frame is calculated only based on the occupancy grid of the current frame. As illustrated in Fig. 7, our full model exhibits the best performance in terms of rendering quality and storage efficiency. By utilizing our modules, the quantization of residual frequency coefficients becomes more precise, retaining more information at the same bitrate. This results in less performance loss when compressed to the same size compared to the variations without them. While ReRF [65] exhibits slightly better quality in the storage range of 700-800KB, our method consistently demonstrates higher PSNR across most capacity ranges and supports mobile rendering. Fig. 8 shows that under a 600KB storage limit, our complete model yields more realistic results with less compression blur." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Cross device runtime analysis. We evaluate the runtime breakdown analysis of VideoRF as detailed in Tab. 2. Our experimental setup includes a Desktop with an i7-12700F CPU and NVIDIA RTX3090 GPU, a Laptop with an i5-1135G7 CPU and Integrated GPU, a Tablet (iPad Pro) with an M2 chip, and a phone (iPhone 14 Pro) with an A16 Bionic chip. For desktops and Laptops, we employ a Python code base along with ModernGL. Meanwhile, for Tablet and Smartphone, the VideoRF player is developed using Swift and Metal. It's worth noting that the video decoding part primarily relies on CPU performance, while the rendering part mainly depends on GPU performance. These two parts operate asynchronously and simultaneously. The other part mainly covers operations such as data conversion between the GPU and CPU. Our results demonstrate that VideoRF allows users to enjoy free-view videos at high" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.26, + 0.887, + 0.358 + ], + "angle": 0, + "content": "
MethodDeviceFPSDecodingRenderingOthers
OursDesktop1165.548 ms3.602 ms3.072 ms
OursLaptop2510.54 ms29.31 ms11.69 ms
OursTablet403.822 ms23.86 ms1.150 ms
OursPhone2313.38 ms40.82 ms3.003 ms
ReRFDesktop445.74 ms194.7 ms—
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.364, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Table 2. Runtime analysis across different devices when processing the same HD image with a resolution of \\(1920 \\times 1080\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.451 + ], + "angle": 0, + "content": "frame rates on multiple devices, providing an experience that rivals the smoothness of watching 2D videos on platforms such as YouTube." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.619, + 0.483 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.735 + ], + "angle": 0, + "content": "We have presented VideoRF, a novel approach enabling real-time streaming and rendering of dynamic radiance fields on mobile devices. Our VideoRF innovatively processes feature volumes as 2D feature streams and adopts deferred rendering to effectively leverage classical video codecs and rasterization pipelines. Our video codec friendly training scheme is implemented on 2D feature space to enhance spatial-temporal consistency for compactness. Additionally, our tailored player supports seamless streaming and rendering of dynamic radiance fields across a range of devices, from desktops to smartphones. Our experiments demonstrate its capability for compact and effective dynamic scene modeling. With the unique ability of real-time rendering of dynamic radiance fields on mobile devices, we believe that our approach marks a significant step forward in neural scene modeling and immersive VR/AR applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.751, + 0.688, + 0.767 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.888 + ], + "angle": 0, + "content": "This work was supported by National Key R&D Program of China (2022YFF0902301), Shanghai Local college capacity building program (22010502800). We also acknowledge support from Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai). This work was supported by the programs of NSFC (61976138 and 61977047) and STCSM (2015F0203-000-06). We thank Siyan Zhuo for her assistance in the figures." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "477" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures, author=Zhiqin Chen and Thomas Funkhouser and Peter Hedman and Andrea Tagliasacchi. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[2] Aljaž Božić, Denis Gladkov, Luke Doukakis, and Christoph Lassner. Neural assets: Volumetric object capture and rendering for interactive environments, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Anpei Chen, Minye Wu, Yingliang Zhang, Nianyi Li, Jie Lu, Shenghua Gao, and Jingyi Yu. Deep surface light fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 1(1):1-17, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.458 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[7] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing Explicit Voxel Grid Representations: Fast NeRFs Become Also Small. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1236-1245, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[9] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6): 246:1-246:16, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[10] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.744, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes:" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[16] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[17] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855-5864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[18] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. HumanRF: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[19] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6155-6165, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[20] Yuheng Jiang, Kaixin Yao, Zhuo Su, Zhehao Shen, Haimin Luo, and Lan Xu. Instant-nvr: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 595-605, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8320-8329, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[22] Moritz Kappel, Vladislav Golyanik, Susana Castillo, Christian Theobalt, and Marcus Magnor. Fast non-rigid radiance fields from monocularized data, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[23] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3- d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[24] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthe" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "sis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[27] Lingzhi Li, Zhongshu Wang, Zhen Shen, Li Shen, and Ping Tan. Compact real-time radiance fields with neural codebook, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[28] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[29] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[30] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Learning the depths of moving people by watching frozen people. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4516-4525, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6498-6508, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.483 + ], + "angle": 0, + "content": "[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.608 + ], + "angle": 0, + "content": "[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In Advances in Neural Information Processing Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[36] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (TOG), 39(4):71-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[37] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[38] Marko Mihajlovic, Sergey Prokudin, Marc Pollefeys, and Siyu Tang. Resfields: Residual neural fields for spatiotemporal signals. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[41] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph., 40(6), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[46] Sida Peng, Chen Geng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Implicit neural representations with structured latent codes for human body modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9895-9907, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[47] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In CVPR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[48] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[49] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[50] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. SIGGRAPH, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[51] Daniel Rho, Byeonghyeon Lee, Seungtae Nam, Joo Chan Lee, Jong Hwan Ko, and Eunbyung Park. Masked wavelet representation for compact neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20680-20690, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Sara Rojas, Jesus Zarzar, Juan C. Pérez, Artsiom Sanakoyeu, Ali Thabet, Albert Pumarola, and Bernard Ghanem. Re" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "479" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "ReND: Real-Time Rendering of NeRFs across Devices. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3632-3641, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[53] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[54] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. NeRFPlayer: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields, year=2023. IEEE Transactions on Visualization and Computer Graphics, 29 (5):2732-2742. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[55] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5459, 2022. 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[56] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[57] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via Rank-residual Decomposition. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[58] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate Textured Mesh Recovery from NeRF via Adaptive Surface Refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17739-17749, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[60] Edith Tretschk, Vladislav Golyanik, Michael Zollhöfer, Aljaz Bozic, Christoph Lassner, and Christian Theobalt. SceNeRFlow: Time-Consistent Reconstruction of General Dynamic. In International Conference on 3D Vision (3DV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.692, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[61] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[62] Ziyu Wan, Christian Richardt, Aljaž Božić, Chao Li, Vijay Rengarajan, Seonghyeon Nam, Xiaoyu Xiang, Tuotuo Li, Bo Zhu, Rakesh Ranjan, and Jing Liao. Learning neural duplex radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8307-8316, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[63] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[64] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[65] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 76-87, 2023. 2, 3, 4, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[66] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.741, + 0.347 + ], + "angle": 0, + "content": "[67] Wikipedia. Z-order curve, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[68] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 287-296, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[69] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.891, + 0.544 + ], + "angle": 0, + "content": "[70] Lan Xu, Zhuo Su, Lei Han, Tao Yu, Yebin Liu, and Lu Fang. Unstructuredfusion: Realtime 4d geometry and texture reconstruction using commercial rgbd cameras. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2508-2522, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[71] Lan Xu, Wei Cheng, Kaiwen Guo, Lei Han, Yebin Liu, and Lu Fang. Flyfusion: Realtime dynamic scene reconstruction using a flying depth camera. IEEE Transactions on Visualization and Computer Graphics, 27(1):68-82, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[72] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[73] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[74] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[75] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "480" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 5752-5761, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[77] Heng Yu, Joel Julin, Zoltán Á. Milacski, Koichiro Niinuma, and László A. Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12397-12406, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.22 + ], + "angle": 0, + "content": "[78] Anlan Zhang, Chendong Wang, Bo Han, and Feng qian. Yuzu:neural-enhanced volumetric video streaming. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[79] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[80] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[81] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[82] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Transactions on Graphics (TOG), 42(4), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.956 + ], + "angle": 0, + "content": "481" + } + ] +] \ No newline at end of file diff --git a/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_origin.pdf b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..912d8a3d631c8fd0efa8901881efa6a493f6279e --- /dev/null +++ b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/12d72448-e14e-4deb-b4a1-e20c3934e649_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10b0a07b096794b05750ba455d5d385b5cf426dc4ad427e7af8d628cff34b13d +size 2298690 diff --git a/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/full.md b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d0ed202d83818048127dd99df60c48ca2227b91 --- /dev/null +++ b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/full.md @@ -0,0 +1,335 @@ +# VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams + +Liao Wang $^{1,3*}$ + +Kaixin Yao $^{1,3*}$ + +Chengcheng Guo1 + +Zhirui Zhang + +Qiang Hu4# + +Jingyi Yu + +$\operatorname{Lan} \mathrm{Xu}^{1 \dagger}$ + +Minye $\mathsf{W}\mathsf{u}^{2\dagger}$ + +1 ShanghaiTech University + +2 KU Leuven + +3 NeuDim + +$^{4}$ Shanghai Jiao Tong University + +![](images/cfc08507d0ad659b88a7fdb9ff37678645d1074ea704b2df3fc367fb30c3cf71.jpg) +Figure 1. Our proposed VideoRF views dynamic radiance field as 2D feature video streams combined with deferred rendering. This technique facilitates hardware video codec and shader-based rendering, enabling smooth high-quality rendering across diverse devices. + +![](images/04801f56eef367f1f5977e4e9b1c0eba39dcbce44ec1c072b06903c8e51b4a61.jpg) + +![](images/bc371545d012708588b29090605eb4a5be5625bee93bab7cfb7e03f278d94969.jpg) +iPad Pro:Apple M2 chip + +![](images/1258810ec2881b0cd27c587a13f777e3f4c11ae60915e2e86ca5b76d802cf496.jpg) +1920x1080, 40FPS, 0.71MB/frame +Dell: i5-1135G7 CPU, Integrated GPU +1200x800, 49FPS, 0.65MB/frame + +# Abstract + +Neural Radiance Fields (NeRFs) excel in photorealistically rendering static scenes. However, rendering dynamic, long-duration radiance fields on ubiquitous devices remains challenging, due to data storage and computational constraints. In this paper, we introduce VideoRF, the first approach to enable real-time streaming and rendering of dynamic human-centric radiance fields on mobile platforms. At the core is a serialized 2D feature image stream representing the 4D radiance field all in one. We introduce a tailored training scheme directly applied to this 2D domain to impose the temporal and spatial redundancy of the feature image stream. By leveraging the redundancy, we show that the feature image stream can be efficiently compressed by 2D video CODECs, which allows us to exploit video hardware accelerators to achieve real-time decoding. On the other hand, based on the feature image stream, we propose a novel rendering pipeline for VideoRF, which has specialized space mappings to query radiance properties efficiently. Paired with a deferred shading model, VideoRF has the capability of real-time rendering on mobile devices + +thanks to its efficiency. We have developed a real-time interactive player that enables online streaming and rendering of dynamic scenes, offering a seamless and immersive free-viewpoint experience across a range of devices, from desktops to mobile phones. Our project page is available at https://aoliaol2138.github.io/VideoRF/. + +# 1. Introduction + +Photorealistic Free-Viewpoint Video (FVV) of dynamic scenes offers an immersive experience in virtual reality and telepresence. Work involving Neural Radiance Fields (NeRFs) has shown great potential in creating photorealistic Free-Viewpoint Videos (FVVs). However, there are still challenges in smoothly delivering and rendering FVVs using NeRFs on commonly used devices, similar to the ease of watching online videos. The difficulty lies in reducing the data capacity for transmitting and storing long sequences and ensuring a low, mobile-compatible computational load. + +Neural Radiance Field (NeRF) [39] surpasses traditional 3D reconstruction methods in photorealistic novel view synthesis. Several works extend NeRF to dynamic scenes by maintaining a canonical space and matching it implicitly [10, 42, 48] or explicitly [34] to align with each frame's live space. However, their dependence on canonical space limits effectiveness in sequences with large motions or topo + +logical changes. Other methods introduce new representations like 4D feature grids [18] or temporal voxel features [11], achieving impressive results on the scenes with topological transformations. Yet, [11] struggles with representing longer sequences owing to model capacity constraints, while [18] encounters streaming challenges due to large storage needs. Recent efforts [26, 54, 65] focus on compressing dynamic frames for streaming, but their computational intensity limits mobile applicability. Concurrently, on mobile devices, real-time rendering of static scenes has been achieved by baking NeRF into mesh templates [1, 52, 75] or texture assets [50]. However, these techniques fall short for dynamic scenes as their per-frame representation becomes too bulky for real-time loading. Moreover, while NeRF compression methods [51, 56, 65] can be employed, they introduce decoding or rendering overheads unsuitable for mobile platforms. Despite the existence of NeRF solutions for dynamic scenes and mobile optimization, a cohesive approach that effectively addresses both still poses a difficulty. + +In this paper, we propose VideoRF - a novel neural modeling approach that enables real-time streaming and rendering of human-centric dynamic radiance fields on common mobile devices (see Fig. 1). Our key idea is to view the 4D feature volumes reconstructed from a dynamic scene as a 2D feature image stream, which is friendly to video codec. Each feature image records the densities and appearance features of one frame. We hence propose a rendering pipeline for this representation. In this rendering pipeline, VideoRF uses a set of mapping tables to connect the 3D space with the 2D feature images. This allows for O(1) density and feature retrieval from the images for every 3D sample point in the space. To reduce the computational complexity of rendering, we adopt the deferred shading model [50] with a global tiny MLP to decode the integrated feature vectors into pixel colors. All rendering operations here are low-cost and compatible with the fragment Shader model, making it possible to implement this rendering pipeline on various devices that have GPUs. + +Second, we present a sequential training scheme to effectively generate the 2D feature image stream. Specifically, VideoRF adaptively groups frames by analyzing motion from the sequential data to ensure temporal stability for mapping table generation. We deploy 3D and 2D Morton sorting techniques to improve spatial consistency in the mapping. VideoRF contains spatial and temporal continuity regularizers that are applied on the mapped 2D feature images using the mapping table. This training strategy enforces the temporal and spatial redundancy to the 2D feature image stream. We show that our feature image stream with spatiotemporal sparsity can be efficiently compressed by off-the-shelf video codecs and reaches high compression rates. + +Moreover, based on these findings, we build a cross-platform player based on VideoRF that can play FVVs in real-time, supported by video hardware accelerators. In this way, users can interactively drag, rotate, pause, play, fast-forward, rewind, and jump to specific frames, providing a viewing experience as seamless as watching online videos on various devices, including smartphones, tablets, laptops, and desktops which was unseen before. + +To summarize, our contributions include: + +- We propose VideoRF, a novel approach to enable real-time dynamic radiance field decoding, streaming and rendering on mobile devices. +- We present an efficient and compact representation, which represents 4D radiance field into 2D feature stream with low rendering complexity to support hardware video codec and Shader rendering. +- We introduce a training scheme to directly impose spatial-temporal consistency on our 2D feature stream for efficient compression. + +# 2. Related work + +Novel View Synthesis for Dynamic Scenes. Dynamic scenes pose a greater challenge in achieving realistic view synthesis results due to moving objects. One approach is to reconstruct the dynamic scene and render the geometry from new viewpoints. The conventional RGB [7, 23, 30, 36, 37, 49, 80] or RGB-D [9, 19-21, 41, 70, 71] solutions have been extensively investigated. Mesh-based neural network [5, 61, 68] techniques are effective for compact data storage and can record view-dependent texture [5, 68], but such methods heavily rely on geometry, especially perform poorly for topologically complex scenarios. + +Many methods extend NeRF into the dynamic view synthesis settings. Some methods [10, 15, 29, 31-33, 43, 48, 63, 69, 77, 79] handle spatial change directly conditions on time and [14, 45, 46, 73, 82] use feature latent code to represent time information. However, they do not support the streaming of long sequences due to the limited representation. Others learn spatial offsets from the live scene to the canonical radiance field by using explicit voxel matching [34], skeletal poses [25, 35, 44], deformed hashtable [24] and deformed volume [22, 59, 60, 66, 72, 81]. While these methods can successfully render dynamic scenes with photo-realistic quality, their heavy reliance on the canonical space makes these methods vulnerable to long sequences, large motions, and topological alterations. Recently, several methods using 4D planes [3, 13, 18, 53, 74], voxel grid [11], Fourier representation [64], residual layers [38] or dynamic MLP maps [47] to represent dynamic scene. However, the inference computational complexity makes it difficult for them to implement streaming and decoding on mobile devices. Additionally, several methods such as [26, 54, 65] have managed to facilitate the streaming of dynamic radi + +![](images/1375062f170c12c0227f12a79b1f8217deb280cc4b8ffd71fbf45552243974c8.jpg) +Figure 2. Demonstration of our rendering. For each 3D sample point, its density $\sigma$ and feature $\mathbf{f}$ are fetched from the 2D feature image through the mapping table $\mathbf{M}$ . Each point feature is first volumetrically accumulated to get the ray feature $\tilde{\mathbf{f}}$ and pass MLP $\Phi$ to decode the ray color. + +ance fields by reducing the capacity of each frame, but the lack of general hardware acceleration makes it challenging to represent dynamic scenes in real-time on mobile phones. + +Cross Device Neural Radiance Field Rendering. Recent works have demonstrated that static neural radiance fields can be rendered on mobile devices. They achieve this by directly converting the neural radiance field to mesh [58] or using traditional texture to represent the scene. [1, 75, 78] have endeavored to augment the capabilities of neural rendering by employing mesh templates and feature texture, [2, 52] facilitating the rendering process through fragment Shader to achieve surface-like rendering with similar ideas. [62] represents neural radiance features encoded on a two-layer duplex mesh for better rendering quality. + +NeRF Acceleration and Compression. NeRF demonstrates exceptional performance in generating images from arbitrary viewpoints, but it suffers from slow rendering efficiency. Some methods focus on integrating a compact structure with a simplified MLP, which reduces the complexity of MLP calculations in traditional NeRFs. Key strategies have been explored, including the employment of voxel grid [26, 55], octrees [12, 64, 76], tri-planes [4], hashing encoding [40], codebook [27, 56], tensor decomposition [6, 57] to accomplish this. Using explicit structure makes training and rendering faster, but it also means that the 3D structure takes up more storage space. + +Therefore, while methods such as CP-decomposition [6], rank reduction [57], and vector quantization [56] manage to attain modest levels of data compression, their application remains confined to static scenes. Additionally, some approaches [8, 51, 65] opt for post-processing techniques, initially utilizing baseline methods to train a decent NeRF representation, then special encoding and decoding schemes are employed to reduce the storage space. However, these kinds of methods are not hardware-friendly, the inference + +process is computationally demanding and cannot support real-time streaming on mobile devices. In contrast, our VideoRF uses codec tailored training, hardware-friendly decoding and shader rendering, which enables real-time dynamic radiance field for long sequences with large motion. + +# 3. VideoRF Representation + +To facilitate dynamic radiance field rendering on mobile devices, we adopt a representation that aligns with the video codec format and maintains low computational complexity for shader-based rendering. We propose turning the 3D volume representation into a 2D formulation, coupled with a highly efficient rendering pipeline. + +In our approach, each frame of the radiance field is represented as a feature image $\mathbf{I}$ where the first channel stores density and the remaining $h$ channels store feature. As depicted in Fig. 2, given a 3D vertex position $\mathbf{x}$ , we retrieve its density $\sigma$ and feature $\mathbf{f}$ using the equation: + +$$ +\sigma , \mathbf {f} = \mathbf {I} [ \mathbf {M} (\mathbf {x}) ], \tag {1} +$$ + +where $\mathbf{M}$ is the 3D-to-2D mapping table from Sec. 4.1. This mapping table not only effectively excludes empty space to reduce storage but also specifics mapping from each non-empty 3D vertex to a corresponding 2D pixel. The lookup operation is highly efficient, with a time complexity of $O(1)$ , facilitating rapid and convenient queries. It's worth noting that our sequence of feature images is in a 2D format, which is friendly to the video codec hardware. + +For rendering, inspired by [17, 50], we use a deferred rendering model. We first accumulate the features along the ray: + +$$ +\begin{array}{l} \tilde {\mathbf {f}} (\mathbf {r}) = \sum_ {k = 1} ^ {n _ {s}} T _ {k} \left(1 - \exp \left(- \sigma_ {k} \delta_ {k}\right)\right) \mathbf {f} _ {k}, \\ T _ {k} = \exp \left(- \sum_ {j = 1} ^ {k - 1} \sigma_ {j} \delta_ {j}\right), \tag {2} \\ \end{array} +$$ + +where $n_{s}$ is the number of sample points along the ray $\mathbf{r}$ . $\sigma_{k}$ , $\mathbf{f}_{k}$ , $\delta_{k}$ denotes the density, feature of the samples, and the interval between adjacent samples respectively. The view-dependent color of the ray is then computed using a tiny global MLP $\Phi$ shared across the frames as: + +$$ +\tilde {C} (\mathbf {r}) = \operatorname {s i g m o i d} \left(\Phi (\tilde {\mathbf {f}} (\mathbf {r}), \mathbf {d})\right), \tag {3} +$$ + +where $\mathbf{d}$ is the view direction of the ray after positional encoding [39]. In this way, the computational burden is significantly reduced as each ray requires only a single MLP decoding which can be implemented in a Shader for real-time rendering on mobile devices. + +![](images/876dabac283503e3ed63fa3d8ca088da565c86e5d9d56dcb85b74472b4140246.jpg) +Figure 3. Overview of our video codec-friendly training. First, we apply our grid-based coarse training [55] to generate per-frame occupancy grid $\mathbf{O}^t$ . Then, during baking, we adaptively group each frame and create a mapping table $\mathbf{M}$ for each group. Next, we sequentially train each feature image $\mathbf{I}^t$ through our spatial, temporal and photometric loss. Finally, feature images are compressed into the feature video streaming to the player. + +# 4. Video Codec-friendly Training + +In this section, we propose a training scheme, as depicted in Fig. 3, to achieve a high compression rate by maintaining spatial and temporal consistency. For spatial aspect, we incorporate 3D-2D Morton sorting to preserve 3D continuity and apply a spatial consistency loss directly on 2D feature images to enforce spatial coherence. On the temporal front, we employ adaptive grouping, which allows frames within a group to share mappings, thereby reducing temporal disruptions, and temporal consistency loss, which serves to further reinforce temporal coherence. These consistencies are crucial which provide a stable and predictable layout for video codecs. + +# 4.1. Baking and Map Generation + +Coarse stage pre-training. Given multiview images of each frame, we first adopt an off-the-shelf approach [55] to generate the explicit density grid $\mathbf{V}_{\sigma}^{t}$ for each frame $t$ independently. We then create a per-frame occupancy grid $\mathbf{O}^t$ by masking in voxels whose opacity exceeds a certain threshold $\gamma$ . This coarse stage sets the foundation for our subsequent adaptive grouping in the baking stage. + +Adaptive group. Generating an independent mapping table for each frame $t$ could disrupt temporal continuity, leading to increased storage requirements after video codec compression. This issue arises because the same 3D point in adjacent frames might map to vastly different 2D positions on the feature image. On the other hand, uniformly applying a single mapping table across all frames could introduce significant spatial redundancy. Considering that a 3D point may only be occupied at sparse intervals, it could remain underutilized for most of the duration, resulting in inefficiency. To balance these factors, we divide the sequence into Groups of Frames (GOFs), maintaining a fixed resolu + +tion for our 2D feature image and adaptively determining the number of frames in each group. For a set of consecutive frames $\{i,i + 1,\dots ,i + n\}$ , we identify the maximum frame number $\alpha$ such that the number of occupied voxels in the union from $i$ to $\alpha$ does not exceed our pixel limit $\theta$ : + +$$ +\operatorname {a r g m a x} _ {\alpha} \mathrm {g} \left(\bigcup_ {j = i} ^ {\alpha} \mathbf {O} ^ {j}\right) \leqslant \theta , \tag {4} +$$ + +where $\mathrm{g}()$ means the number of occupied grids in the union occupancy grid. Then, we will set the frame $i$ to frame $\alpha$ as a GOF and frame $\alpha + 1$ to be the start frame of a new GOF. In this way, all the frames in the group share the same mapping. The same 3D position within the group will be mapped to the same 2D pixel location which keeps temporal consistency and saves the storage. + +Our grouping strategy, unlike the fixed feature grid grouping in ReRF [65], is adaptive, aligning with our objectives. This flexibility allows frames within a group to share a mapping table, optimized in size for a perfect balance between storage efficiency and maintaining temporal continuity. On the other hand, the primary aim of ReRF is to facilitate fast-seeking. Furthermore, our approach diverges from the per-frame occupancy grid of MLP-Maps [47]. We employ a union occupancy grid to calculate the effective number of pixels for adaptive grouping, rather than for speeding up rendering. + +3D Morton sorting. After the union occupancy grid is achieved, we apply Morton sorting to record its 3D spatial continuity. Morton ordering [67], or Z-ordering, interleaves the binary representations of spatial coordinates, ensuring that spatially proximate entities remain adjacent in linear space. As shown in Fig. 4(a), we apply 3D Morton sorting to the vertices based on their position coordinates of the + +![](images/7c3df9f49c7f60e3955d88805590c6e82c9e7953b2859097fce2fcfb1f55f0bb.jpg) +(a) +(b) +(c) +Figure 4. Illustration of our mapping table generation. We first perform 3D Morton sorting (a) on each nonempty vertex and group it into chunks $C_i$ (b). Next, we lay out each chunk into each block $B_i$ of the feature image, arranged in 2D Morton order (c) within it. + +union occupancy grid. Vertices with a density below the threshold $\gamma$ are excluded. This process effectively maintains 3D spatial consistency in the ordering. + +2D block partitioning and 2D Morton sorting. To preserve the 3D spatial continuity within a 2D framework, we employ 2D Morton sorting and partition the feature images into blocks. This approach aligns with the block-wise compression of frames in video CODECs, where blocks with local smoothness lead to more efficient storage. Specifically, we first divide the feature image into $N \times 8$ blocks, denoted as $B_{i}$ , and correspondingly group the sorted vertices into $N \times 8$ chunks, denoted as $C_i$ , as illustrated in Fig. 4(b). For each pixel $\mathbf{p}$ in a block $B_{i}$ , we sort its relative position $(u,v)$ in 2D Morton order. As shown in Fig. 4(c), each chunk $C_i$ is then mapped to a block $B_{i}$ , arranged in 2D Morton order within the block, to form the mapping table M. This 2D Morton ordering ensures that sorted values in the 3D Morton ordering are positioned closely within each feature image block, facilitating efficient compression during the transformation process. + +# 4.2. Fine-grained Sequential Training + +With the aid of our mapping table, we can sequentially train our feature images within each group through spatial consistency loss applied to 2D feature images and temporal consistency loss between frames. + +Spatial consistency loss. In video encoding, regions with homogeneous characteristics demonstrate high compression efficiency. This efficiency stems from the minimal variation in pixel values within these areas, leading to a significant reduction in high-frequency components post Discrete Cosine Transform (DCT) within the video codec. Consequently, these regions retain a higher proportion of low-frequency components. Furthermore, the quantized coefficients in such homogeneous regions are more likely to contain a greater number of zero values, facilitating a more compact data representation and reducing the overall data + +volume. In order to enhance the homogeneity of our 2D feature image, we introduce a total variance loss, $\mathcal{L}_{\mathrm{spatial}}$ during our fine-grained training stage. For each channel of the feature image $\mathbf{I}$ , we enforce its local smoothness by: + +$$ +\mathcal {L} _ {\text {s p a t i a l}} = \frac {1}{| \mathcal {P} |} \sum_ {\mathbf {p} \in \mathcal {V}} \left(\Delta_ {u} (\mathbf {p}) + \Delta_ {v} (\mathbf {p})\right), \tag {5} +$$ + +where $\mathcal{P}$ is the pixels of the feature image, $\Delta_u(\mathbf{p})$ shorthand for Manhattan distance between the feature value at pixel $\mathbf{p} := (u,v)$ and the feature value at pixel $(u + 1,v)$ normalized by the resolution, and analogously for $\Delta_v(\mathbf{p})$ . By increasing the spatial sparsity, the storage of feature videos after video encoding is decreased at the same quality. + +Temporal consistency loss. A naive per-frame training scheme will disrupt temporal continuities by failing to incorporate inter-frame information and resulting in a high bitrate. This is because the residuals between frames are stored after entropy encoding. To optimize storage efficiency, we focus on minimizing the differences in the feature space between the adjacent frames to reduce the entropy. During our sequential training, we enhance interframe similarities by regularizing the current feature image with its predecessor, except for the initial frame of each adaptive group. This is achieved by applying + +$$ +\mathcal {L} _ {\text {t e m p o r a l}} = \left\| \mathbf {I} ^ {t} - \mathbf {I} ^ {t - 1} \right\| _ {1}, \tag {6} +$$ + +for each frame $t$ in the group, ensuring small residuals between consecutive feature images. By employing temporal smoothness in this manner, we can further mitigate entropy throughout the video codec process, thereby facilitating storage conservation. + +Training objective. Our total loss function is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r g b}} + \lambda_ {s} \mathcal {L} _ {\text {s p a t i a l}} + \lambda_ {t} \mathcal {L} _ {\text {t e m p o r a l}}, \tag {7} +$$ + +where $\lambda_{s}$ and $\lambda_{t}$ are the weights for our regular terms and $\mathcal{L}_{\mathrm{rgb}}$ is the photometric loss, + +$$ +\mathcal {L} _ {\mathrm {r g b}} = \sum_ {\mathbf {r} \in \mathcal {R}} \| \mathbf {c} (\mathbf {r}) - \hat {\mathbf {c}} (\mathbf {r}) \| ^ {2}, \tag {8} +$$ + +where $\mathcal{R}$ is the set of training pixel rays; $\mathbf{c}(\mathbf{r})$ and $\hat{\mathbf{c}} (\mathbf{r})$ are the ground truth color and predicted color of a ray $\mathbf{r}$ respectively. + +# 4.3. VideoRF Player + +Finally, we implement a companion VideoRF player to stream and render dynamic radiance fields on mobile devices. Given our feature streams, we quantize them into uint8 format and use H.264 for video codec. Different from [50], we save the mapping table in a 2D-to-3D format as an + +![](images/65122e77c6a6d2ded84b7ed638cda4df2dd222774685de2084abef1cd65eaab6.jpg) +Figure 5. Our VideoRF method generates results for inward-facing, $360^{\circ}$ video sequences featuring human-object interactions with large motion. The images in the last row illustrate our ability to implement variable bitrate in these sequences. + +RGB image to conserve bitrate during streaming. Then, to enable rapid rendering, we first use this mapping table to recover a 3D volume. We employ a compute shader for efficient processing, segmenting the $512 \times 512$ mapping table into $16 \times 16$ workgroups, each handling a $32 \times 32$ pixel section. The highly parallel architecture of compute shaders enables us to efficiently convert 2D features back into a 3D volume. For the rendering part, we implement it via a fragment shader. For fast raymarching, we employ a multiresolution hierarchy of occupancy grids for each group to skip empty space at different levels. We leverage matrix multiplication in the shader to simulate our tiny MLP calculation. These techniques increase the overall speed while ensuring compact storage. + +The VideoRF player marks a significant milestone, as it first enables users to experience real-time rendering of dynamic radiance fields of any length. Within this player, users can drag, rotate, pause, play, fast forward/backward, seek dynamic scenes, or switch between different resolutions like watching online video, offering an extraordinary high-quality free-viewpoint viewing experience. This capability also extends across a wide range of devices, from smartphones and tablets to laptops and desktops, broadening the accessibility and applicability of dynamic radiance fields. + +# 5. Experimental Results + +In this section, we evaluate our VideoRF on a variety of challenging scenarios. We use the PyTorch Framework to train the model on a single NVIDIA GeForce RTX3090. Our new captured dynamic datasets contain around 80 views at the resolution of $1920 \times 1080$ at 30 fps. To ensure the robustness of the algorithm, we also use the ReRF dataset and the HumanRF dataset for result demonstration in Fig. 5. We can render images for immersive, $360^{\circ}$ video sequences that capture human interactions with objects, especially for large motion and long duration sequences. Please refer to the supplementary video for more video results. + +# 5.1. Comparison + +To the best of our knowledge, our approach is the first real-time dynamic radiance field approach that can decode and render on mobile devices. Therefore, we compare to existing dynamic neural rendering methods, including ReRF [65], HumanRF [18], TiNeuVox [11] and per-frame static reconstruction methods like MeRF [50]. We use ReRF [65] dataset and HumanRF [18] Actor3 sequence 1 for a fair comparison, both in storage memory and image quality. As shown in Fig. 6, though TiNeuVox [11] has small storage, + +![](images/8a8073b1fdf35231adfe127754526320994b62dfd6ddcaf3f408c96e26066a5a.jpg) +Figure 6. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods. + +best second-best + +
DatasetMethodPSNR↑SSIM↑LPIPS ↓Size(MB)↓
ReRFReRF [65]31.840.9740.0420.645
MeRF [50]31.120.9750.0307.529
HumanRF [18]28.820.9000.0692.800
TiNeuVox[11]22.700.9230.0830.337
Ours32.010.9760.0230.658
Actors-HQReRF [65]28.330.8360.2960.554
MeRF [50]27.220.8070.2717.164
HumanRF [18]28.980.8880.1512.286
TiNeuVox [11]22.980.7520.4300.447
Ours28.460.8380.2780.426
+ +Table 1. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods. + +it faces the challenge of intensifying blurring effects as the frame count rises. MeRF [50] suffers from the large storage which is not suitable for dynamic scenes when streaming and decoding for playing. HumanRF [18] though capable of representing dynamic scenes, also faces difficulties in streaming and rendering on mobiles due to its computation and storage load. + +We also conduct a quantitative comparison using metrics such as the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) as metrics shown in Tab. 1. For a fair comparison in our experiments, in the ReRF dataset, we test the first 200 frames on the scene Kpop. We use 6 and 39 as test views and the others as training views. As for the HumanRF dataset, we apply the test methods as referenced in HumanRF on the first 250 frames on Actor3, sequence 1. We utilize the values reported for HumanRF and TiNeuVox methods directly from the HumanRF study. In the ReRF dataset, which includes large motion scenarios, we match the rendering quality of the standard ReRF, while also maintaining a compact data storage size. Meanwhile, in the HumanRF dataset with relatively small motion, we achieve a rendering quality that is second only + +![](images/12c4303ccda2249dafd84251befd6522c39d49094c03f1941d40e2630f332443.jpg) +Figure 7. Rate distortion curve. The rate distortion curve illustrates the efficiency of various components within our system. We use different quantization factors to obtain the average rendering quality at different capacities. Our full model stands out as the most compact, allowing for flexible bitrate adjustments to meet diverse storage needs. + +to HumanRF, accomplished with the most efficient storage utilization. Moreover, as indicated in Tab. 2, our method significantly surpasses ReRF in rendering speed when rendering the Kpop scene at an i7-12700F CPU and NVIDIA RTX3090 GPU at the resolution of $1920 \times 1080$ , even on a tablet (iPad Pro with an M2 chip). It's worth noting that, our method is the only method that can provide both mobile and dynamic rendering tasks. The existing dynamic reconstruction methods like TiNeuVox [11], HumanRF [18] and ReRF [65] all cannot effectively utilize hardware encoding and decoding techniques, preventing them from being displayed on mobile devices, especially for the dynamic scene of a long sequence. + +![](images/c3129818a6b6fc71192d6c4d25a58775283e2a37968c10317243ad02f6fcf162.jpg) +Figure 8. Qualitative evaluation of different variations in our method at 600KB. + +# 5.2. Evaluation + +Ablation study. We analyze the impact of spatial consistency, temporal consistency, Morton sorting, block storage, and adaptive grouping on compression and rendering. In the case of models that do not employ 3D and 2D Morton sorting, spatial point data is sorted sequentially in row-major order while employing 2D block-wise storage. For models not utilizing block storage, a 3D spatial Morton sorting is applied, followed by storage in a row-major format within the 2D feature space. For models without adaptive grouping, the mapping table for each frame is calculated only based on the occupancy grid of the current frame. As illustrated in Fig. 7, our full model exhibits the best performance in terms of rendering quality and storage efficiency. By utilizing our modules, the quantization of residual frequency coefficients becomes more precise, retaining more information at the same bitrate. This results in less performance loss when compressed to the same size compared to the variations without them. While ReRF [65] exhibits slightly better quality in the storage range of 700-800KB, our method consistently demonstrates higher PSNR across most capacity ranges and supports mobile rendering. Fig. 8 shows that under a 600KB storage limit, our complete model yields more realistic results with less compression blur. + +Cross device runtime analysis. We evaluate the runtime breakdown analysis of VideoRF as detailed in Tab. 2. Our experimental setup includes a Desktop with an i7-12700F CPU and NVIDIA RTX3090 GPU, a Laptop with an i5-1135G7 CPU and Integrated GPU, a Tablet (iPad Pro) with an M2 chip, and a phone (iPhone 14 Pro) with an A16 Bionic chip. For desktops and Laptops, we employ a Python code base along with ModernGL. Meanwhile, for Tablet and Smartphone, the VideoRF player is developed using Swift and Metal. It's worth noting that the video decoding part primarily relies on CPU performance, while the rendering part mainly depends on GPU performance. These two parts operate asynchronously and simultaneously. The other part mainly covers operations such as data conversion between the GPU and CPU. Our results demonstrate that VideoRF allows users to enjoy free-view videos at high + +
MethodDeviceFPSDecodingRenderingOthers
OursDesktop1165.548 ms3.602 ms3.072 ms
OursLaptop2510.54 ms29.31 ms11.69 ms
OursTablet403.822 ms23.86 ms1.150 ms
OursPhone2313.38 ms40.82 ms3.003 ms
ReRFDesktop445.74 ms194.7 ms—
+ +Table 2. Runtime analysis across different devices when processing the same HD image with a resolution of $1920 \times 1080$ . + +frame rates on multiple devices, providing an experience that rivals the smoothness of watching 2D videos on platforms such as YouTube. + +# 6. Conclusion + +We have presented VideoRF, a novel approach enabling real-time streaming and rendering of dynamic radiance fields on mobile devices. Our VideoRF innovatively processes feature volumes as 2D feature streams and adopts deferred rendering to effectively leverage classical video codecs and rasterization pipelines. Our video codec friendly training scheme is implemented on 2D feature space to enhance spatial-temporal consistency for compactness. Additionally, our tailored player supports seamless streaming and rendering of dynamic radiance fields across a range of devices, from desktops to smartphones. Our experiments demonstrate its capability for compact and effective dynamic scene modeling. With the unique ability of real-time rendering of dynamic radiance fields on mobile devices, we believe that our approach marks a significant step forward in neural scene modeling and immersive VR/AR applications. + +# 7. Acknowledgements + +This work was supported by National Key R&D Program of China (2022YFF0902301), Shanghai Local college capacity building program (22010502800). We also acknowledge support from Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai). This work was supported by the programs of NSFC (61976138 and 61977047) and STCSM (2015F0203-000-06). We thank Siyan Zhuo for her assistance in the figures. + +# References + +[1] MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures, author=Zhiqin Chen and Thomas Funkhouser and Peter Hedman and Andrea Tagliasacchi. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[2] Aljaž Božić, Denis Gladkov, Luke Doukakis, and Christoph Lassner. Neural assets: Volumetric object capture and rendering for interactive environments, 2022. 3 +[3] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 2 +[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3 +[5] Anpei Chen, Minye Wu, Yingliang Zhang, Nianyi Li, Jie Lu, Shenghua Gao, and Jingyi Yu. Deep surface light fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 1(1):1-17, 2018. 2 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 3 +[7] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2 +[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing Explicit Voxel Grid Representations: Fast NeRFs Become Also Small. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1236-1245, 2023. 3 +[9] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6): 246:1-246:16, 2017. 2 +[10] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2 +[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2, 6, 7 +[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3 +[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: + +Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2 +[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, 2021. 2 +[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 2 +[16] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 2 +[17] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855-5864, 2021. 3 +[18] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. HumanRF: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 6, 7 +[19] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6155-6165, 2022. 2 +[20] Yuheng Jiang, Kaixin Yao, Zhuo Su, Zhehao Shen, Haimin Luo, and Lan Xu. Instant-nvr: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 595-605, 2023. +[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8320-8329, 2018. 2 +[22] Moritz Kappel, Vladislav Golyanik, Susana Castillo, Christian Theobalt, and Marcus Magnor. Fast non-rigid radiance fields from monocularized data, 2023. 2 +[23] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3- d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2 +[24] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2 +[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis, 2023. 2 +[26] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthe + +sis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3 +[27] Lingzhi Li, Zhongshu Wang, Zhen Shen, Li Shen, and Ping Tan. Compact real-time radiance fields with neural codebook, 2023. 3 +[28] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 2 +[29] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2 +[30] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Learning the depths of moving people by watching frozen people. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4516-4525, 2019. 2 +[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6498-6508, 2021. 2 +[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering, 2023. +[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2 +[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In Advances in Neural Information Processing Systems, 2022. 1, 2 +[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), 2022. 2 +[36] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (TOG), 39(4):71-1, 2020. 2 +[37] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2 +[38] Marko Mihajlovic, Sergey Prokudin, Marc Pollefeys, and Siyu Tang. Resfields: Residual neural fields for spatiotemporal signals. In The Twelfth International Conference on Learning Representations, 2024. 2 +[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3 + +[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3 +[41] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2 +[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1 +[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph., 40(6), 2021. 2 +[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2 +[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2 +[46] Sida Peng, Chen Geng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Implicit neural representations with structured latent codes for human body modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9895-9907, 2023. 2 +[47] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In CVPR, 2023. 2, 4 +[48] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2 +[49] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2 +[50] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. SIGGRAPH, 2023. 2, 3, 5, 6, 7 +[51] Daniel Rho, Byeonghyeon Lee, Seungtae Nam, Joo Chan Lee, Jong Hwan Ko, and Eunbyung Park. Masked wavelet representation for compact neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20680-20690, 2023. 2, 3 +[52] Sara Rojas, Jesus Zarzar, Juan C. Pérez, Artsiom Sanakoyeu, Ali Thabet, Albert Pumarola, and Bernard Ghanem. Re + +ReND: Real-Time Rendering of NeRFs across Devices. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3632-3641, 2023. 2, 3 +[53] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering, 2023. 2 +[54] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. NeRFPlayer: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields, year=2023. IEEE Transactions on Visualization and Computer Graphics, 29 (5):2732-2742. 2 +[55] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5459, 2022. 3, 4, 1 +[56] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3 +[57] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via Rank-residual Decomposition. In Advances in Neural Information Processing Systems, 2022. 3 +[58] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate Textured Mesh Recovery from NeRF via Adaptive Surface Refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17739-17749, 2023. 3 +[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2 +[60] Edith Tretschk, Vladislav Golyanik, Michael Zollhöfer, Aljaz Bozic, Christoph Lassner, and Christian Theobalt. SceNeRFlow: Time-Consistent Reconstruction of General Dynamic. In International Conference on 3D Vision (3DV), 2024. 2 +[61] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 2 +[62] Ziyu Wan, Christian Richardt, Aljaž Božić, Chao Li, Vijay Rengarajan, Seonghyeon Nam, Xiaoyu Xiang, Tuotuo Li, Bo Zhu, Rakesh Ranjan, and Jing Liao. Learning neural duplex radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8307-8316, 2023. 3 +[63] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2 + +[64] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2, 3 +[65] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 76-87, 2023. 2, 3, 4, 6, 7, 8, 1 +[66] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[67] Wikipedia. Z-order curve, 2024. 4 +[68] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 287-296, 2000. 2 +[69] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2 +[70] Lan Xu, Zhuo Su, Lei Han, Tao Yu, Yebin Liu, and Lu Fang. Unstructuredfusion: Realtime 4d geometry and texture reconstruction using commercial rgbd cameras. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2508-2522, 2020. 2 +[71] Lan Xu, Wei Cheng, Kaiwen Guo, Lei Han, Yebin Liu, and Lu Fang. Flyfusion: Realtime dynamic scene reconstruction using a flying depth camera. IEEE Transactions on Visualization and Computer Graphics, 27(1):68-82, 2021. 2 +[72] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[73] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[74] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. 2023. 2 +[75] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3 +[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF + +International Conference on Computer Vision, pages 5752-5761, 2021. 3 +[77] Heng Yu, Joel Julin, Zoltán Á. Milacski, Koichiro Niinuma, and László A. Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12397-12406, 2023. 2 +[78] Anlan Zhang, Chendong Wang, Bo Han, and Feng qian. Yuzu:neural-enhanced volumetric video streaming. 2022. 3 +[79] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2 +[80] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2022. 2 +[81] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2 +[82] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Transactions on Graphics (TOG), 42(4), 2023. 2 \ No newline at end of file diff --git a/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/images.zip b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ddaf0f68fc8333888cb0e5e99085fde3327dfd5 --- /dev/null +++ b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:564c24b5321eead4727b046b33050f110709ee2488e3618cc09eb9173b2450a9 +size 644570 diff --git a/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/layout.json b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..80520c5f65a2c3b4aa10b17266ec38973632ff49 --- /dev/null +++ b/2024/VideoRF_ Rendering Dynamic Radiance Fields as 2D Feature Video Streams/layout.json @@ -0,0 +1,9235 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "text", + "content": "VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 142, + 137, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 142, + 137, + 158 + ], + "spans": [ + { + "bbox": [ + 66, + 142, + 137, + 158 + ], + "type": "text", + "content": "Liao Wang" + }, + { + "bbox": [ + 66, + 142, + 137, + 158 + ], + "type": "inline_equation", + "content": "^{1,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 142, + 232, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 142, + 232, + 156 + ], + "spans": [ + { + "bbox": [ + 161, + 142, + 232, + 156 + ], + "type": "text", + "content": "Kaixin Yao" + }, + { + "bbox": [ + 161, + 142, + 232, + 156 + ], + "type": "inline_equation", + "content": "^{1,3*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 142, + 346, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 142, + 346, + 158 + ], + "spans": [ + { + "bbox": [ + 257, + 142, + 346, + 158 + ], + "type": "text", + "content": "Chengcheng Guo1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 370, + 143, + 441, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 143, + 441, + 158 + ], + "spans": [ + { + "bbox": [ + 370, + 143, + 441, + 158 + ], + "type": "text", + "content": "Zhirui Zhang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 466, + 143, + 526, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 143, + 526, + 158 + ], + "spans": [ + { + "bbox": [ + 466, + 143, + 526, + 158 + ], + "type": "text", + "content": "Qiang Hu4#" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 163, + 219, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 163, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 167, + 163, + 219, + 177 + ], + "type": "text", + "content": "Jingyi Yu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 163, + 315, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 163, + 315, + 175 + ], + "spans": [ + { + "bbox": [ + 268, + 163, + 315, + 175 + ], + "type": "inline_equation", + "content": "\\operatorname{Lan} \\mathrm{Xu}^{1 \\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 163, + 425, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 163, + 425, + 177 + ], + "spans": [ + { + "bbox": [ + 364, + 163, + 425, + 177 + ], + "type": "text", + "content": "Minye " + }, + { + "bbox": [ + 364, + 163, + 425, + 177 + ], + "type": "inline_equation", + "content": "\\mathsf{W}\\mathsf{u}^{2\\dagger}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 177, + 188, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 177, + 188, + 191 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 188, + 191 + ], + "type": "text", + "content": "1 ShanghaiTech University" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 213, + 178, + 278, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 178, + 278, + 190 + ], + "spans": [ + { + "bbox": [ + 213, + 178, + 278, + 190 + ], + "type": "text", + "content": "2 KU Leuven" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 178, + 354, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 178, + 354, + 190 + ], + "spans": [ + { + "bbox": [ + 303, + 178, + 354, + 190 + ], + "type": "text", + "content": "3 NeuDim" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 178, + 535, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 178, + 535, + 191 + ], + "spans": [ + { + "bbox": [ + 380, + 178, + 535, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 380, + 178, + 535, + 191 + ], + "type": "text", + "content": " Shanghai Jiao Tong University" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 52, + 198, + 236, + 352 + ], + "blocks": [ + { + "bbox": [ + 52, + 198, + 236, + 352 + ], + "lines": [ + { + "bbox": [ + 52, + 198, + 236, + 352 + ], + "spans": [ + { + "bbox": [ + 52, + 198, + 236, + 352 + ], + "type": "image", + "image_path": "cfc08507d0ad659b88a7fdb9ff37678645d1074ea704b2df3fc367fb30c3cf71.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 359, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 359, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 545, + 381 + ], + "type": "text", + "content": "Figure 1. Our proposed VideoRF views dynamic radiance field as 2D feature video streams combined with deferred rendering. This technique facilitates hardware video codec and shader-based rendering, enabling smooth high-quality rendering across diverse devices." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 238, + 217, + 421, + 349 + ], + "blocks": [ + { + "bbox": [ + 238, + 217, + 421, + 349 + ], + "lines": [ + { + "bbox": [ + 238, + 217, + 421, + 349 + ], + "spans": [ + { + "bbox": [ + 238, + 217, + 421, + 349 + ], + "type": "image", + "image_path": "04801f56eef367f1f5977e4e9b1c0eba39dcbce44ec1c072b06903c8e51b4a61.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 426, + 198, + 538, + 257 + ], + "blocks": [ + { + "bbox": [ + 426, + 198, + 538, + 257 + ], + "lines": [ + { + "bbox": [ + 426, + 198, + 538, + 257 + ], + "spans": [ + { + "bbox": [ + 426, + 198, + 538, + 257 + ], + "type": "image", + "image_path": "bc371545d012708588b29090605eb4a5be5625bee93bab7cfb7e03f278d94969.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 258, + 517, + 266 + ], + "lines": [ + { + "bbox": [ + 447, + 258, + 517, + 266 + ], + "spans": [ + { + "bbox": [ + 447, + 258, + 517, + 266 + ], + "type": "text", + "content": "iPad Pro:Apple M2 chip" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 426, + 277, + 537, + 337 + ], + "blocks": [ + { + "bbox": [ + 435, + 266, + 529, + 273 + ], + "lines": [ + { + "bbox": [ + 435, + 266, + 529, + 273 + ], + "spans": [ + { + "bbox": [ + 435, + 266, + 529, + 273 + ], + "type": "text", + "content": "1920x1080, 40FPS, 0.71MB/frame" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 277, + 537, + 337 + ], + "lines": [ + { + "bbox": [ + 426, + 277, + 537, + 337 + ], + "spans": [ + { + "bbox": [ + 426, + 277, + 537, + 337 + ], + "type": "image", + "image_path": "1258810ec2881b0cd27c587a13f777e3f4c11ae60915e2e86ca5b76d802cf496.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 338, + 530, + 346 + ], + "lines": [ + { + "bbox": [ + 433, + 338, + 530, + 346 + ], + "spans": [ + { + "bbox": [ + 433, + 338, + 530, + 346 + ], + "type": "text", + "content": "Dell: i5-1135G7 CPU, Integrated GPU" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 346, + 529, + 353 + ], + "lines": [ + { + "bbox": [ + 437, + 346, + 529, + 353 + ], + "spans": [ + { + "bbox": [ + 437, + 346, + 529, + 353 + ], + "type": "text", + "content": "1200x800, 49FPS, 0.65MB/frame" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 403, + 191, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 403, + 191, + 416 + ], + "spans": [ + { + "bbox": [ + 143, + 403, + 191, + 416 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 45, + 429, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 429, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 45, + 429, + 288, + 668 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) excel in photorealistically rendering static scenes. However, rendering dynamic, long-duration radiance fields on ubiquitous devices remains challenging, due to data storage and computational constraints. In this paper, we introduce VideoRF, the first approach to enable real-time streaming and rendering of dynamic human-centric radiance fields on mobile platforms. At the core is a serialized 2D feature image stream representing the 4D radiance field all in one. We introduce a tailored training scheme directly applied to this 2D domain to impose the temporal and spatial redundancy of the feature image stream. By leveraging the redundancy, we show that the feature image stream can be efficiently compressed by 2D video CODECs, which allows us to exploit video hardware accelerators to achieve real-time decoding. On the other hand, based on the feature image stream, we propose a novel rendering pipeline for VideoRF, which has specialized space mappings to query radiance properties efficiently. Paired with a deferred shading model, VideoRF has the capability of real-time rendering on mobile devices" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 393, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 464 + ], + "type": "text", + "content": "thanks to its efficiency. We have developed a real-time interactive player that enables online streaming and rendering of dynamic scenes, offering a seamless and immersive free-viewpoint experience across a range of devices, from desktops to mobile phones. Our project page is available at https://aoliaol2138.github.io/VideoRF/." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 489, + 386, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 386, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 386, + 501 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "content": "Photorealistic Free-Viewpoint Video (FVV) of dynamic scenes offers an immersive experience in virtual reality and telepresence. Work involving Neural Radiance Fields (NeRFs) has shown great potential in creating photorealistic Free-Viewpoint Videos (FVVs). However, there are still challenges in smoothly delivering and rendering FVVs using NeRFs on commonly used devices, similar to the ease of watching online videos. The difficulty lies in reducing the data capacity for transmitting and storing long sequences and ensuring a low, mobile-compatible computational load." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) [39] surpasses traditional 3D reconstruction methods in photorealistic novel view synthesis. Several works extend NeRF to dynamic scenes by maintaining a canonical space and matching it implicitly [10, 42, 48] or explicitly [34] to align with each frame's live space. However, their dependence on canonical space limits effectiveness in sequences with large motions or topo" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 674, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 714 + ], + "type": "text", + "content": "* Authors contributed equally to this work. † The corresponding authors are Minye Wu (minye.wu@kuleuven.be) and Lan Xu (xulan1@shanghaiitech.edu.cn). # Work done while at ShanghaiTech University." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "470" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "logical changes. Other methods introduce new representations like 4D feature grids [18] or temporal voxel features [11], achieving impressive results on the scenes with topological transformations. Yet, [11] struggles with representing longer sequences owing to model capacity constraints, while [18] encounters streaming challenges due to large storage needs. Recent efforts [26, 54, 65] focus on compressing dynamic frames for streaming, but their computational intensity limits mobile applicability. Concurrently, on mobile devices, real-time rendering of static scenes has been achieved by baking NeRF into mesh templates [1, 52, 75] or texture assets [50]. However, these techniques fall short for dynamic scenes as their per-frame representation becomes too bulky for real-time loading. Moreover, while NeRF compression methods [51, 56, 65] can be employed, they introduce decoding or rendering overheads unsuitable for mobile platforms. Despite the existence of NeRF solutions for dynamic scenes and mobile optimization, a cohesive approach that effectively addresses both still poses a difficulty." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 543 + ], + "type": "text", + "content": "In this paper, we propose VideoRF - a novel neural modeling approach that enables real-time streaming and rendering of human-centric dynamic radiance fields on common mobile devices (see Fig. 1). Our key idea is to view the 4D feature volumes reconstructed from a dynamic scene as a 2D feature image stream, which is friendly to video codec. Each feature image records the densities and appearance features of one frame. We hence propose a rendering pipeline for this representation. In this rendering pipeline, VideoRF uses a set of mapping tables to connect the 3D space with the 2D feature images. This allows for O(1) density and feature retrieval from the images for every 3D sample point in the space. To reduce the computational complexity of rendering, we adopt the deferred shading model [50] with a global tiny MLP to decode the integrated feature vectors into pixel colors. All rendering operations here are low-cost and compatible with the fragment Shader model, making it possible to implement this rendering pipeline on various devices that have GPUs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": "Second, we present a sequential training scheme to effectively generate the 2D feature image stream. Specifically, VideoRF adaptively groups frames by analyzing motion from the sequential data to ensure temporal stability for mapping table generation. We deploy 3D and 2D Morton sorting techniques to improve spatial consistency in the mapping. VideoRF contains spatial and temporal continuity regularizers that are applied on the mapped 2D feature images using the mapping table. This training strategy enforces the temporal and spatial redundancy to the 2D feature image stream. We show that our feature image stream with spatiotemporal sparsity can be efficiently compressed by off-the-shelf video codecs and reaches high compression rates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "Moreover, based on these findings, we build a cross-platform player based on VideoRF that can play FVVs in real-time, supported by video hardware accelerators. In this way, users can interactively drag, rotate, pause, play, fast-forward, rewind, and jump to specific frames, providing a viewing experience as seamless as watching online videos on various devices, including smartphones, tablets, laptops, and desktops which was unseen before." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 168, + 485, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 168, + 485, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 485, + 179 + ], + "type": "text", + "content": "To summarize, our contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 180, + 545, + 300 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 180, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 215 + ], + "type": "text", + "content": "- We propose VideoRF, a novel approach to enable real-time dynamic radiance field decoding, streaming and rendering on mobile devices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 216, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 216, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 216, + 545, + 262 + ], + "type": "text", + "content": "- We present an efficient and compact representation, which represents 4D radiance field into 2D feature stream with low rendering complexity to support hardware video codec and Shader rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 263, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 545, + 300 + ], + "type": "text", + "content": "- We introduce a training scheme to directly impose spatial-temporal consistency on our 2D feature stream for efficient compression." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 311, + 390, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 390, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 390, + 323 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 330, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 462 + ], + "type": "text", + "content": "Novel View Synthesis for Dynamic Scenes. Dynamic scenes pose a greater challenge in achieving realistic view synthesis results due to moving objects. One approach is to reconstruct the dynamic scene and render the geometry from new viewpoints. The conventional RGB [7, 23, 30, 36, 37, 49, 80] or RGB-D [9, 19-21, 41, 70, 71] solutions have been extensively investigated. Mesh-based neural network [5, 61, 68] techniques are effective for compact data storage and can record view-dependent texture [5, 68], but such methods heavily rely on geometry, especially perform poorly for topologically complex scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "content": "Many methods extend NeRF into the dynamic view synthesis settings. Some methods [10, 15, 29, 31-33, 43, 48, 63, 69, 77, 79] handle spatial change directly conditions on time and [14, 45, 46, 73, 82] use feature latent code to represent time information. However, they do not support the streaming of long sequences due to the limited representation. Others learn spatial offsets from the live scene to the canonical radiance field by using explicit voxel matching [34], skeletal poses [25, 35, 44], deformed hashtable [24] and deformed volume [22, 59, 60, 66, 72, 81]. While these methods can successfully render dynamic scenes with photo-realistic quality, their heavy reliance on the canonical space makes these methods vulnerable to long sequences, large motions, and topological alterations. Recently, several methods using 4D planes [3, 13, 18, 53, 74], voxel grid [11], Fourier representation [64], residual layers [38] or dynamic MLP maps [47] to represent dynamic scene. However, the inference computational complexity makes it difficult for them to implement streaming and decoding on mobile devices. Additionally, several methods such as [26, 54, 65] have managed to facilitate the streaming of dynamic radi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "471" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 199 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 199 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 199 + ], + "type": "image", + "image_path": "1375062f170c12c0227f12a79b1f8217deb280cc4b8ffd71fbf45552243974c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "text", + "content": "Figure 2. Demonstration of our rendering. For each 3D sample point, its density " + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "text", + "content": " are fetched from the 2D feature image through the mapping table " + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "text", + "content": ". Each point feature is first volumetrically accumulated to get the ray feature " + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}" + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "text", + "content": " and pass MLP " + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 203, + 288, + 258 + ], + "type": "text", + "content": " to decode the ray color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 308 + ], + "type": "text", + "content": "ance fields by reducing the capacity of each frame, but the lack of general hardware acceleration makes it challenging to represent dynamic scenes in real-time on mobile phones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": "Cross Device Neural Radiance Field Rendering. Recent works have demonstrated that static neural radiance fields can be rendered on mobile devices. They achieve this by directly converting the neural radiance field to mesh [58] or using traditional texture to represent the scene. [1, 75, 78] have endeavored to augment the capabilities of neural rendering by employing mesh templates and feature texture, [2, 52] facilitating the rendering process through fragment Shader to achieve surface-like rendering with similar ideas. [62] represents neural radiance features encoded on a two-layer duplex mesh for better rendering quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 604 + ], + "type": "text", + "content": "NeRF Acceleration and Compression. NeRF demonstrates exceptional performance in generating images from arbitrary viewpoints, but it suffers from slow rendering efficiency. Some methods focus on integrating a compact structure with a simplified MLP, which reduces the complexity of MLP calculations in traditional NeRFs. Key strategies have been explored, including the employment of voxel grid [26, 55], octrees [12, 64, 76], tri-planes [4], hashing encoding [40], codebook [27, 56], tensor decomposition [6, 57] to accomplish this. Using explicit structure makes training and rendering faster, but it also means that the 3D structure takes up more storage space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Therefore, while methods such as CP-decomposition [6], rank reduction [57], and vector quantization [56] manage to attain modest levels of data compression, their application remains confined to static scenes. Additionally, some approaches [8, 51, 65] opt for post-processing techniques, initially utilizing baseline methods to train a decent NeRF representation, then special encoding and decoding schemes are employed to reduce the storage space. However, these kinds of methods are not hardware-friendly, the inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "process is computationally demanding and cannot support real-time streaming on mobile devices. In contrast, our VideoRF uses codec tailored training, hardware-friendly decoding and shader rendering, which enables real-time dynamic radiance field for long sequences with large motion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 144, + 447, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 447, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 447, + 159 + ], + "type": "text", + "content": "3. VideoRF Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 165, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 237 + ], + "type": "text", + "content": "To facilitate dynamic radiance field rendering on mobile devices, we adopt a representation that aligns with the video codec format and maintains low computational complexity for shader-based rendering. We propose turning the 3D volume representation into a 2D formulation, coupled with a highly efficient rendering pipeline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "content": "In our approach, each frame of the radiance field is represented as a feature image " + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "content": " where the first channel stores density and the remaining " + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "content": " channels store feature. As depicted in Fig. 2, given a 3D vertex position " + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "content": ", we retrieve its density " + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 238, + 545, + 298 + ], + "type": "text", + "content": " using the equation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 391, + 308, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 308, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 391, + 308, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\sigma , \\mathbf {f} = \\mathbf {I} [ \\mathbf {M} (\\mathbf {x}) ], \\tag {1}", + "image_path": "32845acc39356382653db15e916b8542984c458fe98d0dbbe3e1b91699345c05.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "text", + "content": " is the 3D-to-2D mapping table from Sec. 4.1. This mapping table not only effectively excludes empty space to reduce storage but also specifics mapping from each non-empty 3D vertex to a corresponding 2D pixel. The lookup operation is highly efficient, with a time complexity of " + }, + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "text", + "content": ", facilitating rapid and convenient queries. It's worth noting that our sequence of feature images is in a 2D format, which is friendly to the video codec hardware." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 464 + ], + "type": "text", + "content": "For rendering, inspired by [17, 50], we use a deferred rendering model. We first accumulate the features along the ray:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 474, + 545, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 474, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 351, + 474, + 545, + 545 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {f}} (\\mathbf {r}) = \\sum_ {k = 1} ^ {n _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) \\mathbf {f} _ {k}, \\\\ T _ {k} = \\exp \\left(- \\sum_ {j = 1} ^ {k - 1} \\sigma_ {j} \\delta_ {j}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "20d636f59126317ae1be054af5e6348d132b25f77c32fff22d6c29fbe55a2926.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "inline_equation", + "content": "n_{s}" + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": " is the number of sample points along the ray " + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{k}" + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": " denotes the density, feature of the samples, and the interval between adjacent samples respectively. The view-dependent color of the ray is then computed using a tiny global MLP " + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": " shared across the frames as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 357, + 624, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 624, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 357, + 624, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\tilde {C} (\\mathbf {r}) = \\operatorname {s i g m o i d} \\left(\\Phi (\\tilde {\\mathbf {f}} (\\mathbf {r}), \\mathbf {d})\\right), \\tag {3}", + "image_path": "3aeeb580ae8c13a751c6a160ae98adbce1e12523304e5de5af3e2d5501ec7297.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " is the view direction of the ray after positional encoding [39]. In this way, the computational burden is significantly reduced as each ray requires only a single MLP decoding which can be implemented in a Shader for real-time rendering on mobile devices." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "472" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 56, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 50, + 56, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 50, + 56, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 50, + 56, + 547, + 230 + ], + "type": "image", + "image_path": "876dabac283503e3ed63fa3d8ca088da565c86e5d9d56dcb85b74472b4140246.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "text", + "content": "Figure 3. Overview of our video codec-friendly training. First, we apply our grid-based coarse training [55] to generate per-frame occupancy grid " + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^t" + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "text", + "content": ". Then, during baking, we adaptively group each frame and create a mapping table " + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "text", + "content": " for each group. Next, we sequentially train each feature image " + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 46, + 234, + 547, + 281 + ], + "type": "text", + "content": " through our spatial, temporal and photometric loss. Finally, feature images are compressed into the feature video streaming to the player." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 290, + 220, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 290, + 220, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 220, + 304 + ], + "type": "text", + "content": "4. Video Codec-friendly Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 290, + 453 + ], + "type": "text", + "content": "In this section, we propose a training scheme, as depicted in Fig. 3, to achieve a high compression rate by maintaining spatial and temporal consistency. For spatial aspect, we incorporate 3D-2D Morton sorting to preserve 3D continuity and apply a spatial consistency loss directly on 2D feature images to enforce spatial coherence. On the temporal front, we employ adaptive grouping, which allows frames within a group to share mappings, thereby reducing temporal disruptions, and temporal consistency loss, which serves to further reinforce temporal coherence. These consistencies are crucial which provide a stable and predictable layout for video codecs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 463, + 204, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 204, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 204, + 475 + ], + "type": "text", + "content": "4.1. Baking and Map Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "text", + "content": "Coarse stage pre-training. Given multiview images of each frame, we first adopt an off-the-shelf approach [55] to generate the explicit density grid " + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\sigma}^{t}" + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "text", + "content": " independently. We then create a per-frame occupancy grid " + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^t" + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "text", + "content": " by masking in voxels whose opacity exceeds a certain threshold " + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 481, + 287, + 567 + ], + "type": "text", + "content": ". This coarse stage sets the foundation for our subsequent adaptive grouping in the baking stage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Adaptive group. Generating an independent mapping table for each frame " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " could disrupt temporal continuity, leading to increased storage requirements after video codec compression. This issue arises because the same 3D point in adjacent frames might map to vastly different 2D positions on the feature image. On the other hand, uniformly applying a single mapping table across all frames could introduce significant spatial redundancy. Considering that a 3D point may only be occupied at sparse intervals, it could remain underutilized for most of the duration, resulting in inefficiency. To balance these factors, we divide the sequence into Groups of Frames (GOFs), maintaining a fixed resolu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "content": "tion for our 2D feature image and adaptively determining the number of frames in each group. For a set of consecutive frames " + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{i,i + 1,\\dots ,i + n\\}" + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "content": ", we identify the maximum frame number " + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "content": " such that the number of occupied voxels in the union from " + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "content": " does not exceed our pixel limit " + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 291, + 547, + 352 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 361, + 545, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 361, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 367, + 361, + 545, + 399 + ], + "type": "interline_equation", + "content": "\\operatorname {a r g m a x} _ {\\alpha} \\mathrm {g} \\left(\\bigcup_ {j = i} ^ {\\alpha} \\mathbf {O} ^ {j}\\right) \\leqslant \\theta , \\tag {4}", + "image_path": "bb32836085bed677ae428b48ffea74923a8e63b2a1c8d20c915356c86a678a5d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{g}()" + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "text", + "content": " means the number of occupied grids in the union occupancy grid. Then, we will set the frame " + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "text", + "content": " as a GOF and frame " + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\alpha + 1" + }, + { + "bbox": [ + 304, + 409, + 545, + 493 + ], + "type": "text", + "content": " to be the start frame of a new GOF. In this way, all the frames in the group share the same mapping. The same 3D position within the group will be mapped to the same 2D pixel location which keeps temporal consistency and saves the storage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 494, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 625 + ], + "type": "text", + "content": "Our grouping strategy, unlike the fixed feature grid grouping in ReRF [65], is adaptive, aligning with our objectives. This flexibility allows frames within a group to share a mapping table, optimized in size for a perfect balance between storage efficiency and maintaining temporal continuity. On the other hand, the primary aim of ReRF is to facilitate fast-seeking. Furthermore, our approach diverges from the per-frame occupancy grid of MLP-Maps [47]. We employ a union occupancy grid to calculate the effective number of pixels for adaptive grouping, rather than for speeding up rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "3D Morton sorting. After the union occupancy grid is achieved, we apply Morton sorting to record its 3D spatial continuity. Morton ordering [67], or Z-ordering, interleaves the binary representations of spatial coordinates, ensuring that spatially proximate entities remain adjacent in linear space. As shown in Fig. 4(a), we apply 3D Morton sorting to the vertices based on their position coordinates of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "473" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 276, + 162 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 276, + 162 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 276, + 162 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 276, + 162 + ], + "type": "image", + "image_path": "7c3df9f49c7f60e3955d88805590c6e82c9e7953b2859097fce2fcfb1f55f0bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 167, + 92, + 178 + ], + "lines": [ + { + "bbox": [ + 80, + 167, + 92, + 178 + ], + "spans": [ + { + "bbox": [ + 80, + 167, + 92, + 178 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 167, + 179, + 177 + ], + "lines": [ + { + "bbox": [ + 167, + 167, + 179, + 177 + ], + "spans": [ + { + "bbox": [ + 167, + 167, + 179, + 177 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 167, + 245, + 177 + ], + "lines": [ + { + "bbox": [ + 233, + 167, + 245, + 177 + ], + "spans": [ + { + "bbox": [ + 233, + 167, + 245, + 177 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "type": "text", + "content": "Figure 4. Illustration of our mapping table generation. We first perform 3D Morton sorting (a) on each nonempty vertex and group it into chunks " + }, + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "type": "text", + "content": " (b). Next, we lay out each chunk into each block " + }, + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 46, + 182, + 287, + 237 + ], + "type": "text", + "content": " of the feature image, arranged in 2D Morton order (c) within it." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "content": "union occupancy grid. Vertices with a density below the threshold " + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "content": " are excluded. This process effectively maintains 3D spatial consistency in the ordering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": "2D block partitioning and 2D Morton sorting. To preserve the 3D spatial continuity within a 2D framework, we employ 2D Morton sorting and partition the feature images into blocks. This approach aligns with the block-wise compression of frames in video CODECs, where blocks with local smoothness lead to more efficient storage. Specifically, we first divide the feature image into " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "N \\times 8" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": " blocks, denoted as " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": ", and correspondingly group the sorted vertices into " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "N \\times 8" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": " chunks, denoted as " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": ", as illustrated in Fig. 4(b). For each pixel " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": " in a block " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": ", we sort its relative position " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": " in 2D Morton order. As shown in Fig. 4(c), each chunk " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": " is then mapped to a block " + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 498 + ], + "type": "text", + "content": ", arranged in 2D Morton order within the block, to form the mapping table M. This 2D Morton ordering ensures that sorted values in the 3D Morton ordering are positioned closely within each feature image block, facilitating efficient compression during the transformation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 510, + 226, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 226, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 226, + 522 + ], + "type": "text", + "content": "4.2. Fine-grained Sequential Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 529, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 577 + ], + "type": "text", + "content": "With the aid of our mapping table, we can sequentially train our feature images within each group through spatial consistency loss applied to 2D feature images and temporal consistency loss between frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Spatial consistency loss. In video encoding, regions with homogeneous characteristics demonstrate high compression efficiency. This efficiency stems from the minimal variation in pixel values within these areas, leading to a significant reduction in high-frequency components post Discrete Cosine Transform (DCT) within the video codec. Consequently, these regions retain a higher proportion of low-frequency components. Furthermore, the quantized coefficients in such homogeneous regions are more likely to contain a greater number of zero values, facilitating a more compact data representation and reducing the overall data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "volume. In order to enhance the homogeneity of our 2D feature image, we introduce a total variance loss, " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{spatial}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " during our fine-grained training stage. For each channel of the feature image " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ", we enforce its local smoothness by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 128, + 545, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 128, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 345, + 128, + 545, + 158 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a t i a l}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {\\mathbf {p} \\in \\mathcal {V}} \\left(\\Delta_ {u} (\\mathbf {p}) + \\Delta_ {v} (\\mathbf {p})\\right), \\tag {5}", + "image_path": "eb5c3d17ad00501c987ec1adc46e38b45440ecdfd9064452b99906bef7f312e6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "content": " is the pixels of the feature image, " + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\Delta_u(\\mathbf{p})" + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "content": " shorthand for Manhattan distance between the feature value at pixel " + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{p} := (u,v)" + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "content": " and the feature value at pixel " + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "inline_equation", + "content": "(u + 1,v)" + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "content": " normalized by the resolution, and analogously for " + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\Delta_v(\\mathbf{p})" + }, + { + "bbox": [ + 305, + 166, + 545, + 238 + ], + "type": "text", + "content": ". By increasing the spatial sparsity, the storage of feature videos after video encoding is decreased at the same quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 243, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 375 + ], + "type": "text", + "content": "Temporal consistency loss. A naive per-frame training scheme will disrupt temporal continuities by failing to incorporate inter-frame information and resulting in a high bitrate. This is because the residuals between frames are stored after entropy encoding. To optimize storage efficiency, we focus on minimizing the differences in the feature space between the adjacent frames to reduce the entropy. During our sequential training, we enhance interframe similarities by regularizing the current feature image with its predecessor, except for the initial frame of each adaptive group. This is achieved by applying" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 382, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 382, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 370, + 382, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p o r a l}} = \\left\\| \\mathbf {I} ^ {t} - \\mathbf {I} ^ {t - 1} \\right\\| _ {1}, \\tag {6}", + "image_path": "b70a983105c83a4fd7916fe08fa11d9e71a63824c39c44d307ddaa6fdbbd3bea.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 405, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 545, + 464 + ], + "type": "text", + "content": "for each frame " + }, + { + "bbox": [ + 305, + 405, + 545, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 405, + 545, + 464 + ], + "type": "text", + "content": " in the group, ensuring small residuals between consecutive feature images. By employing temporal smoothness in this manner, we can further mitigate entropy throughout the video codec process, thereby facilitating storage conservation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 469, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 545, + 491 + ], + "type": "text", + "content": "Training objective. Our total loss function is formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 502, + 545, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 502, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 339, + 502, + 545, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r g b}} + \\lambda_ {s} \\mathcal {L} _ {\\text {s p a t i a l}} + \\lambda_ {t} \\mathcal {L} _ {\\text {t e m p o r a l}}, \\tag {7}", + "image_path": "65ba2d2bf87f2f6e06efa93b8dadcb2dc5cd98a88fb902a1db02e951e68696f1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "text", + "content": " are the weights for our regular terms and " + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 306, + 525, + 545, + 548 + ], + "type": "text", + "content": " is the photometric loss," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 557, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 557, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 367, + 557, + 545, + 582 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\mathbf {c} (\\mathbf {r}) - \\hat {\\mathbf {c}} (\\mathbf {r}) \\| ^ {2}, \\tag {8}", + "image_path": "144b08b5152ed0b45919b1b239ed9890bf07a6e1e92ec1b21557198510ec772b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "content": " is the set of training pixel rays; " + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\mathbf{r})" + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}} (\\mathbf{r})" + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "content": " are the ground truth color and predicted color of a ray " + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 635, + 403, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 403, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 403, + 647 + ], + "type": "text", + "content": "4.3. VideoRF Player" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Finally, we implement a companion VideoRF player to stream and render dynamic radiance fields on mobile devices. Given our feature streams, we quantize them into uint8 format and use H.264 for video codec. Different from [50], we save the mapping table in a 2D-to-3D format as an" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "474" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 44, + 541, + 373 + ], + "blocks": [ + { + "bbox": [ + 48, + 44, + 541, + 373 + ], + "lines": [ + { + "bbox": [ + 48, + 44, + 541, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 44, + 541, + 373 + ], + "type": "image", + "image_path": "65122e77c6a6d2ded84b7ed638cda4df2dd222774685de2084abef1cd65eaab6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "type": "text", + "content": "Figure 5. Our VideoRF method generates results for inward-facing, " + }, + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "type": "text", + "content": " video sequences featuring human-object interactions with large motion. The images in the last row illustrate our ability to implement variable bitrate in these sequences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "content": "RGB image to conserve bitrate during streaming. Then, to enable rapid rendering, we first use this mapping table to recover a 3D volume. We employ a compute shader for efficient processing, segmenting the " + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "content": " mapping table into " + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "content": " workgroups, each handling a " + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "content": " pixel section. The highly parallel architecture of compute shaders enables us to efficiently convert 2D features back into a 3D volume. For the rendering part, we implement it via a fragment shader. For fast raymarching, we employ a multiresolution hierarchy of occupancy grids for each group to skip empty space at different levels. We leverage matrix multiplication in the shader to simulate our tiny MLP calculation. These techniques increase the overall speed while ensuring compact storage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 712 + ], + "type": "text", + "content": "The VideoRF player marks a significant milestone, as it first enables users to experience real-time rendering of dynamic radiance fields of any length. Within this player, users can drag, rotate, pause, play, fast forward/backward, seek dynamic scenes, or switch between different resolutions like watching online video, offering an extraordinary high-quality free-viewpoint viewing experience. This capability also extends across a wide range of devices, from smartphones and tablets to laptops and desktops, broadening the accessibility and applicability of dynamic radiance fields." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 410, + 432, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 432, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 432, + 423 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "type": "text", + "content": "In this section, we evaluate our VideoRF on a variety of challenging scenarios. We use the PyTorch Framework to train the model on a single NVIDIA GeForce RTX3090. Our new captured dynamic datasets contain around 80 views at the resolution of " + }, + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "type": "text", + "content": " at 30 fps. To ensure the robustness of the algorithm, we also use the ReRF dataset and the HumanRF dataset for result demonstration in Fig. 5. We can render images for immersive, " + }, + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 431, + 547, + 574 + ], + "type": "text", + "content": " video sequences that capture human interactions with objects, especially for large motion and long duration sequences. Please refer to the supplementary video for more video results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 586, + 387, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 387, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 387, + 599 + ], + "type": "text", + "content": "5.1. Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "To the best of our knowledge, our approach is the first real-time dynamic radiance field approach that can decode and render on mobile devices. Therefore, we compare to existing dynamic neural rendering methods, including ReRF [65], HumanRF [18], TiNeuVox [11] and per-frame static reconstruction methods like MeRF [50]. We use ReRF [65] dataset and HumanRF [18] Actor3 sequence 1 for a fair comparison, both in storage memory and image quality. As shown in Fig. 6, though TiNeuVox [11] has small storage," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "475" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 59, + 541, + 244 + ], + "blocks": [ + { + "bbox": [ + 55, + 59, + 541, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 59, + 541, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 59, + 541, + 244 + ], + "type": "image", + "image_path": "8a8073b1fdf35231adfe127754526320994b62dfd6ddcaf3f408c96e26066a5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 246, + 523, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 246, + 523, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 246, + 523, + 258 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 275, + 296, + 395 + ], + "blocks": [ + { + "bbox": [ + 139, + 264, + 203, + 274 + ], + "lines": [ + { + "bbox": [ + 139, + 264, + 203, + 274 + ], + "spans": [ + { + "bbox": [ + 139, + 264, + 203, + 274 + ], + "type": "text", + "content": "best second-best" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 275, + 296, + 395 + ], + "lines": [ + { + "bbox": [ + 48, + 275, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 296, + 395 + ], + "type": "table", + "html": "
DatasetMethodPSNR↑SSIM↑LPIPS ↓Size(MB)↓
ReRFReRF [65]31.840.9740.0420.645
MeRF [50]31.120.9750.0307.529
HumanRF [18]28.820.9000.0692.800
TiNeuVox[11]22.700.9230.0830.337
Ours32.010.9760.0230.658
Actors-HQReRF [65]28.330.8360.2960.554
MeRF [50]27.220.8070.2717.164
HumanRF [18]28.980.8880.1512.286
TiNeuVox [11]22.980.7520.4300.447
Ours28.460.8380.2780.426
", + "image_path": "80fafe97f806c06cb36e5cb1cca94d08d1361a7aae3888e7c80f99d8b9f91bff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 399, + 288, + 422 + ], + "lines": [ + { + "bbox": [ + 47, + 399, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 288, + 422 + ], + "type": "text", + "content": "Table 1. Qualitative comparison against dynamic scene reconstruction methods and per frame static reconstruction methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "content": "it faces the challenge of intensifying blurring effects as the frame count rises. MeRF [50] suffers from the large storage which is not suitable for dynamic scenes when streaming and decoding for playing. HumanRF [18] though capable of representing dynamic scenes, also faces difficulties in streaming and rendering on mobiles due to its computation and storage load." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "We also conduct a quantitative comparison using metrics such as the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) as metrics shown in Tab. 1. For a fair comparison in our experiments, in the ReRF dataset, we test the first 200 frames on the scene Kpop. We use 6 and 39 as test views and the others as training views. As for the HumanRF dataset, we apply the test methods as referenced in HumanRF on the first 250 frames on Actor3, sequence 1. We utilize the values reported for HumanRF and TiNeuVox methods directly from the HumanRF study. In the ReRF dataset, which includes large motion scenarios, we match the rendering quality of the standard ReRF, while also maintaining a compact data storage size. Meanwhile, in the HumanRF dataset with relatively small motion, we achieve a rendering quality that is second only" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 264, + 544, + 449 + ], + "blocks": [ + { + "bbox": [ + 310, + 264, + 544, + 449 + ], + "lines": [ + { + "bbox": [ + 310, + 264, + 544, + 449 + ], + "spans": [ + { + "bbox": [ + 310, + 264, + 544, + 449 + ], + "type": "image", + "image_path": "12c4303ccda2249dafd84251befd6522c39d49094c03f1941d40e2630f332443.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 467, + 547, + 534 + ], + "lines": [ + { + "bbox": [ + 304, + 467, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 547, + 534 + ], + "type": "text", + "content": "Figure 7. Rate distortion curve. The rate distortion curve illustrates the efficiency of various components within our system. We use different quantization factors to obtain the average rendering quality at different capacities. Our full model stands out as the most compact, allowing for flexible bitrate adjustments to meet diverse storage needs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "to HumanRF, accomplished with the most efficient storage utilization. Moreover, as indicated in Tab. 2, our method significantly surpasses ReRF in rendering speed when rendering the Kpop scene at an i7-12700F CPU and NVIDIA RTX3090 GPU at the resolution of " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ", even on a tablet (iPad Pro with an M2 chip). It's worth noting that, our method is the only method that can provide both mobile and dynamic rendering tasks. The existing dynamic reconstruction methods like TiNeuVox [11], HumanRF [18] and ReRF [65] all cannot effectively utilize hardware encoding and decoding techniques, preventing them from being displayed on mobile devices, especially for the dynamic scene of a long sequence." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "476" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 71, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 54, + 71, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 54, + 71, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 54, + 71, + 541, + 178 + ], + "type": "image", + "image_path": "c3129818a6b6fc71192d6c4d25a58775283e2a37968c10317243ad02f6fcf162.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 184, + 441, + 195 + ], + "lines": [ + { + "bbox": [ + 151, + 184, + 441, + 195 + ], + "spans": [ + { + "bbox": [ + 151, + 184, + 441, + 195 + ], + "type": "text", + "content": "Figure 8. Qualitative evaluation of different variations in our method at 600KB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 207, + 120, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 120, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 120, + 218 + ], + "type": "text", + "content": "5.2. Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 230, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 289, + 517 + ], + "type": "text", + "content": "Ablation study. We analyze the impact of spatial consistency, temporal consistency, Morton sorting, block storage, and adaptive grouping on compression and rendering. In the case of models that do not employ 3D and 2D Morton sorting, spatial point data is sorted sequentially in row-major order while employing 2D block-wise storage. For models not utilizing block storage, a 3D spatial Morton sorting is applied, followed by storage in a row-major format within the 2D feature space. For models without adaptive grouping, the mapping table for each frame is calculated only based on the occupancy grid of the current frame. As illustrated in Fig. 7, our full model exhibits the best performance in terms of rendering quality and storage efficiency. By utilizing our modules, the quantization of residual frequency coefficients becomes more precise, retaining more information at the same bitrate. This results in less performance loss when compressed to the same size compared to the variations without them. While ReRF [65] exhibits slightly better quality in the storage range of 700-800KB, our method consistently demonstrates higher PSNR across most capacity ranges and supports mobile rendering. Fig. 8 shows that under a 600KB storage limit, our complete model yields more realistic results with less compression blur." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 289, + 714 + ], + "type": "text", + "content": "Cross device runtime analysis. We evaluate the runtime breakdown analysis of VideoRF as detailed in Tab. 2. Our experimental setup includes a Desktop with an i7-12700F CPU and NVIDIA RTX3090 GPU, a Laptop with an i5-1135G7 CPU and Integrated GPU, a Tablet (iPad Pro) with an M2 chip, and a phone (iPhone 14 Pro) with an A16 Bionic chip. For desktops and Laptops, we employ a Python code base along with ModernGL. Meanwhile, for Tablet and Smartphone, the VideoRF player is developed using Swift and Metal. It's worth noting that the video decoding part primarily relies on CPU performance, while the rendering part mainly depends on GPU performance. These two parts operate asynchronously and simultaneously. The other part mainly covers operations such as data conversion between the GPU and CPU. Our results demonstrate that VideoRF allows users to enjoy free-view videos at high" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 205, + 542, + 283 + ], + "blocks": [ + { + "bbox": [ + 309, + 205, + 542, + 283 + ], + "lines": [ + { + "bbox": [ + 309, + 205, + 542, + 283 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 542, + 283 + ], + "type": "table", + "html": "
MethodDeviceFPSDecodingRenderingOthers
OursDesktop1165.548 ms3.602 ms3.072 ms
OursLaptop2510.54 ms29.31 ms11.69 ms
OursTablet403.822 ms23.86 ms1.150 ms
OursPhone2313.38 ms40.82 ms3.003 ms
ReRFDesktop445.74 ms194.7 ms—
", + "image_path": "50497defc30fd94f2f6793b569b39366cf573910b1ca403735318a6a2e1b9405.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 288, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 310 + ], + "type": "text", + "content": "Table 2. Runtime analysis across different devices when processing the same HD image with a resolution of " + }, + { + "bbox": [ + 306, + 288, + 545, + 310 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 306, + 288, + 545, + 310 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 321, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 357 + ], + "type": "text", + "content": "frame rates on multiple devices, providing an experience that rivals the smoothness of watching 2D videos on platforms such as YouTube." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 370, + 378, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 378, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 378, + 382 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 390, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 582 + ], + "type": "text", + "content": "We have presented VideoRF, a novel approach enabling real-time streaming and rendering of dynamic radiance fields on mobile devices. Our VideoRF innovatively processes feature volumes as 2D feature streams and adopts deferred rendering to effectively leverage classical video codecs and rasterization pipelines. Our video codec friendly training scheme is implemented on 2D feature space to enhance spatial-temporal consistency for compactness. Additionally, our tailored player supports seamless streaming and rendering of dynamic radiance fields across a range of devices, from desktops to smartphones. Our experiments demonstrate its capability for compact and effective dynamic scene modeling. With the unique ability of real-time rendering of dynamic radiance fields on mobile devices, we believe that our approach marks a significant step forward in neural scene modeling and immersive VR/AR applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 594, + 421, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 421, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 421, + 607 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 613, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 703 + ], + "type": "text", + "content": "This work was supported by National Key R&D Program of China (2022YFF0902301), Shanghai Local college capacity building program (22010502800). We also acknowledge support from Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai). This work was supported by the programs of NSFC (61976138 and 61977047) and STCSM (2015F0203-000-06). We thank Siyan Zhuo for her assistance in the figures." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "477" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures, author=Zhiqin Chen and Thomas Funkhouser and Peter Hedman and Andrea Tagliasacchi. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 191 + ], + "type": "text", + "content": "[2] Aljaž Božić, Denis Gladkov, Luke Doukakis, and Christoph Lassner. Neural assets: Volumetric object capture and rendering for interactive environments, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 215 + ], + "type": "text", + "content": "[3] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 282 + ], + "type": "text", + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Anpei Chen, Minye Wu, Yingliang Zhang, Nianyi Li, Jie Lu, Shenghua Gao, and Jingyi Yu. Deep surface light fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 1(1):1-17, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 362 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 418 + ], + "type": "text", + "content": "[7] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 474 + ], + "type": "text", + "content": "[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing Explicit Voxel Grid Representations: Fast NeRFs Become Also Small. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1236-1245, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 531 + ], + "type": "text", + "content": "[9] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6): 246:1-246:16, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "text", + "content": "[10] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "type": "text", + "content": "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "type": "text", + "content": "[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "text", + "content": "[16] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 307 + ], + "type": "text", + "content": "[17] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855-5864, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "text", + "content": "[18] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. HumanRF: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "text", + "content": "[19] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6155-6165, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 432, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 487 + ], + "type": "text", + "content": "[20] Yuheng Jiang, Kaixin Yao, Zhuo Su, Zhehao Shen, Haimin Luo, and Lan Xu. Instant-nvr: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 595-605, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "text", + "content": "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8320-8329, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 533, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 565 + ], + "type": "text", + "content": "[22] Moritz Kappel, Vladislav Golyanik, Susana Castillo, Christian Theobalt, and Marcus Magnor. Fast non-rigid radiance fields from monocularized data, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "text", + "content": "[23] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3- d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "text", + "content": "[24] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[26] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthe" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "478" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "sis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[27] Lingzhi Li, Zhongshu Wang, Zhen Shen, Li Shen, and Ping Tan. Compact real-time radiance fields with neural codebook, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[28] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 239 + ], + "type": "text", + "content": "[29] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[30] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Learning the depths of moving people by watching frozen people. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4516-4525, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "text", + "content": "[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6498-6508, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "type": "text", + "content": "[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 481 + ], + "type": "text", + "content": "[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In Advances in Neural Information Processing Systems, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "text", + "content": "[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "type": "text", + "content": "[36] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (TOG), 39(4):71-1, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[37] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[38] Marko Mihajlovic, Sergey Prokudin, Marc Pollefeys, and Siyu Tang. Resfields: Residual neural fields for spatiotemporal signals. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[41] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "text", + "content": "[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "text", + "content": "[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields. ACM Trans. Graph., 40(6), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "text", + "content": "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 437 + ], + "type": "text", + "content": "[46] Sida Peng, Chen Geng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Implicit neural representations with structured latent codes for human body modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9895-9907, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "text", + "content": "[47] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In CVPR, 2023. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "text", + "content": "[48] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "text", + "content": "[49] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "type": "text", + "content": "[50] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. SIGGRAPH, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 689 + ], + "type": "text", + "content": "[51] Daniel Rho, Byeonghyeon Lee, Seungtae Nam, Joo Chan Lee, Jong Hwan Ko, and Eunbyung Park. Masked wavelet representation for compact neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20680-20690, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[52] Sara Rojas, Jesus Zarzar, Juan C. Pérez, Artsiom Sanakoyeu, Ali Thabet, Albert Pumarola, and Bernard Ghanem. Re" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "479" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "ReND: Real-Time Rendering of NeRFs across Devices. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3632-3641, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 149 + ], + "type": "text", + "content": "[53] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 215 + ], + "type": "text", + "content": "[54] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. NeRFPlayer: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields, year=2023. IEEE Transactions on Visualization and Computer Graphics, 29 (5):2732-2742. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 270 + ], + "type": "text", + "content": "[55] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5459, 2022. 3, 4, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "type": "text", + "content": "[56] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "type": "text", + "content": "[57] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via Rank-residual Decomposition. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 361, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 425 + ], + "type": "text", + "content": "[58] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate Textured Mesh Recovery from NeRF via Adaptive Surface Refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17739-17749, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 426, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 491 + ], + "type": "text", + "content": "[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 493, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 287, + 546 + ], + "type": "text", + "content": "[60] Edith Tretschk, Vladislav Golyanik, Michael Zollhöfer, Aljaz Bozic, Christoph Lassner, and Christian Theobalt. SceNeRFlow: Time-Consistent Reconstruction of General Dynamic. In International Conference on 3D Vision (3DV), 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "type": "text", + "content": "[61] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 592, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 287, + 658 + ], + "type": "text", + "content": "[62] Ziyu Wan, Christian Richardt, Aljaž Božić, Chao Li, Vijay Rengarajan, Seonghyeon Nam, Xiaoyu Xiang, Tuotuo Li, Bo Zhu, Rakesh Ranjan, and Jing Liao. Learning neural duplex radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8307-8316, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "[63] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[64] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "text", + "content": "[65] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 76-87, 2023. 2, 3, 4, 6, 7, 8, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "type": "text", + "content": "[66] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 263, + 453, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 453, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 453, + 274 + ], + "type": "text", + "content": "[67] Wikipedia. Z-order curve, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[68] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 287-296, 2000. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 332, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 385 + ], + "type": "text", + "content": "[69] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "text", + "content": "[70] Lan Xu, Zhuo Su, Lei Han, Tao Yu, Yebin Liu, and Lu Fang. Unstructuredfusion: Realtime 4d geometry and texture reconstruction using commercial rgbd cameras. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2508-2522, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "text", + "content": "[71] Lan Xu, Wei Cheng, Kaiwen Guo, Lei Han, Yebin Liu, and Lu Fang. Flyfusion: Realtime dynamic scene reconstruction using a flying depth camera. IEEE Transactions on Visualization and Computer Graphics, 27(1):68-82, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[72] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "type": "text", + "content": "[73] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 612 + ], + "type": "text", + "content": "[74] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "text", + "content": "[75] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[76] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "480" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 373 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 5752-5761, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[77] Heng Yu, Joel Julin, Zoltán Á. Milacski, Koichiro Niinuma, and László A. Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12397-12406, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 174 + ], + "type": "text", + "content": "[78] Anlan Zhang, Chendong Wang, Bo Han, and Feng qian. Yuzu:neural-enhanced volumetric video streaming. 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[79] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[80] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[81] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "text", + "content": "[82] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Transactions on Graphics (TOG), 42(4), 2023. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "481" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_content_list.json b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce5d1a0e033fd57d62688f1fb9f5f625c1453ab6 --- /dev/null +++ b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_content_list.json @@ -0,0 +1,1715 @@ +[ + { + "type": "text", + "text": "VideoSwap: Customized Video Subject Swapping with Interactive Semantic Point Correspondence", + "text_level": 1, + "bbox": [ + 209, + 104, + 759, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchao Gu $^{1,2}$ , Yipin Zhou $^{2}$ , Bichen Wu $^{2}$ , Licheng Yu $^{2}$ , Jia-Wei Liu $^{1}$ , Rui Zhao $^{1}$ , Jay Zhangjie Wu $^{1}$ , David Junhao Zhang $^{1}$ , Mike Zheng Shou $^{1*}$ , Kevin Tang $^{2}$", + "bbox": [ + 86, + 176, + 906, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Show Lab, National University of Singapore 2GenAI, Meta", + "bbox": [ + 253, + 229, + 745, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://videoswap.github.io/", + "bbox": [ + 330, + 250, + 661, + 266 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27055b80c9a28a4f6cdc776294013d9cc4ea1b701cdd58b23ac3b51726cf07c9.jpg", + "image_caption": [ + "One-to-One Point Correspondence", + "(b) Results of Various Correspondence Used in Previous Video Editing", + "Figure 1. Customized video subject swapping results with VideoSwap. VideoSwap supports shape change in the swapped results while aligning with the source motion trajectory. The swapped target can be either a predefined concept from a pretrained model (e.g., helicopter) or a customized concept (denoted by $V^{*}$ ). Previous methods based on implicit motion encoding and dense correspondence do not perform well in subject swapping with shape changes. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + ], + "image_footnote": [], + "bbox": [ + 125, + 292, + 848, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 685, + 313, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current diffusion-based video editing primarily focuses on structure-preserved editing by utilizing various dense correspondences to ensure temporal consistency and motion alignment. However, these approaches are often ineffective when the target edit involves a shape change. To embark on video editing with shape change, we explore customized video subject swapping in this work, where we aim to replace the main subject in a source video with a target subject having a distinct identity and potentially different shape. In contrast to previous methods that rely on dense", + "bbox": [ + 75, + 719, + 470, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "correspondences, we introduce the VideoSwap framework that exploits semantic point correspondences, inspired by our observation that only a small number of semantic points are necessary to align the subject's motion trajectory and modify its shape. We also introduce various user-point interactions (e.g., removing points and dragging points) to address various semantic point correspondence. Extensive experiments demonstrate state-of-the-art video subject swapping results across a variety of real-world videos.", + "bbox": [ + 500, + 672, + 893, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 843, + 630, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based video editing [5, 10, 22, 25, 31, 43, 44, 49] is an emerging field that harnesses the capabilities of pre", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c00273f80ad5a159aead8e8f5637d61c155ce2ced2bc23724fec96641fcb14c.jpg", + "image_caption": [ + "Figure 2. Customized video subject swapping results of VideoSwap on various concepts. The swapping target can either be a predefined concept in the pretrained model (e.g., helicopter) or a customized concept created by ED-LoRA [13] (denoted as $V^{*}$ ). We encourage readers to click and play the video clips in this figure using Adobe Acrobat. For legal issues, we cannot display the human swap results." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 846, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained text-to-image/video diffusion models [14, 15, 32, 37] to facilitate a range of video editing tasks, including style change and subject/background swapping. The main challenge in video editing lies in how to extract motion from the source video and transfer it to the edited video while ensuring temporal consistency. Pioneer Tune-A-Video [43] implicitly encodes source motion in the diffusion model weights by tuning from the source video. While it demonstrates versatile applications for video editing, the", + "bbox": [ + 75, + 763, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "temporal consistency is far from satisfactory. Subsequent works make use of various dense correspondences extracted from the source video, including attention maps [23, 31], edge/depth maps [10, 22, 46, 49], optical flows [46], and deformation fields [5, 6] for video editing. While achieving better temporal consistency, dense correspondence imposes strict shape constraints on the target edit, which makes it ineffective for video editing with shape changes.", + "bbox": [ + 496, + 763, + 893, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To embark on video editing with shape change, we delve", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into a challenging task: customized video subject swapping. Unlike the conventional video subject swapping addressed in previous works [5, 46, 49], where the swapped subject conforms to the shape of source subject, customized subjects have a clearly defined identity in terms of both appearance and shape. These distinctive characteristics should be preserved in the target edit. Therefore, previous structure-preserved video editing methods are often ineffective for this problem, as shown in Fig. 1(b).", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address customized video subject swapping, our primary insight is that the subject's motion trajectory can be effectively described using a small number of semantic points. As shown in Fig. 1(a), the motion trajectory of an airplane can be depicted by semantic points located at its wings, nose, and tail. This insight naturally leads us to the following question: Can we utilize these semantic points as correspondences to align the motion trajectory while relaxing the shape constraints in video editing? To answer this question, we conduct a toy experiments and observe that it is possible to learn semantic point correspondence for a specific video subject using just a small number of source video frames. Users can interact with learned semantic points to generate unseen poses or modify the shape of the video subject. These observations suggest the potential for integrating semantic point correspondence into video editing, provided that we can obtain an accurate semantic point sequence for the target edit.", + "bbox": [ + 75, + 228, + 472, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To unleash the potential of semantic point correspondence, we introduce the VideoSwap framework for customized video subject swapping, which comprises the following primary designs: 1) Integrating the motion layer into the image diffusion model to ensure essential temporal consistency. 2) Registering semantic points on the source video and utilizing them to transfer the motion trajectory of source subject to the target edit. 3) Introducing user-point interactions (e.g., removing or dragging points) for various semantic point correspondence.", + "bbox": [ + 75, + 500, + 472, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 652, + 398, + 666 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Empirical observations that reveal the potential of semantic point correspondence for aligning motion trajectories and changing shapes in video editing.", + "- The VideoSwap framework, which minimizes user intervention while unleashing the potential of semantic point correspondence in customized video subject swapping.", + "- State-of-the-art results in customized video subject swapping, as demonstrated in Fig. 2." + ], + "bbox": [ + 76, + 667, + 468, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 805, + 218, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Diffusion-Based Video Editing", + "text_level": 1, + "bbox": [ + 76, + 830, + 346, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structure-Preserved Video Editing. FateZero [31] and Video-P2P [23] extract cross- and self-attention from the source video to control spatial layout. To achieve stricter", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "alignment of temporal consistency with the source video, Rerender-A-Video [46], Gen-1 [10], ControlVideo [49], and TokenFlow [12] extract and align optical flow, depth/edge maps, and nearest-neighbour field from the source video respectively, resulting in improved temporal consistency. StableVideo [5], VidEdit [6], and CoDEF [28] learn the canonical space for editing following the Layered Neural Atlas [19] or the deformation field in Dynamic Nerf [24, 30]. While achieving promising results in structure-preserved video editing, the above methods based on various dense correspondence are not suitable for handling subject swapping involving shape changes.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Editing with Shape Change. Tune-A-Video (TAV) [43] and FateZero [31] with TAV checkpoint can be utilized for video editing with shape change, as they implicitly encode the motion in model weights through tuning on the source video. However, they suffer from structure and appearance leakage due to model tuning. Shape-aware editing [21] is built on the Layered Neural Atlas [19], employing semantic correspondences to estimate shape deformation and warp the atlas. Nevertheless, texture warping will lead to unrealistic textures.", + "bbox": [ + 496, + 272, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Diffusion Models. Previous video editing primarily relies on the text-to-image diffusion model [32]. However, recent advancements are occurring in video foundation models [3, 14, 41, 48]. In this work, we add a motion layer [14] to the image diffusion model to provide essential temporal consistency for video editing, and we focus on exploiting semantic point correspondence to align the motion trajectory of the video subject.", + "bbox": [ + 496, + 422, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Point Correspondence", + "text_level": 1, + "bbox": [ + 500, + 555, + 705, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Correspondence in Diffusion Models. DIFT [39] initially uncovers robust semantic point correspondences in diffusion models. Building upon the observation in DIFT, subsequent works, DragDiffusion [36] and DragonDiffusion [26] extend DragGAN [29] to support interactive point-based image editing in diffusion models. However, point correspondences and interactive drag-based editing are seldom investigated for video editing.", + "bbox": [ + 496, + 579, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tracking Any Point in Video. TAP-Vid [8] first introduces the problem of tracking any point (TAP). Unlike optical flow estimation, TAP requires the establishment of long-range correspondence for all points in a video. In our work, we use TAP to reduce human intervention in annotating subject keypoints and acquire long-range motion estimation. Although several works [8, 9, 18, 42] address the TAP problem, we choose to employ Co-Tracker [18], which is the most efficient solution available.", + "bbox": [ + 496, + 700, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Concept Customization", + "text_level": 1, + "bbox": [ + 500, + 845, + 715, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept customization is mainly categorized into tuning-based approaches [11, 13, 20, 33, 40] and tuning-free solu", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9799fea7388a2a67a05c446022f2c8e5e87835a2e45ae85b080ec617b34b952d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 263, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b76bdf232898b4929a1ddaa69f63f3086d127d293af831685beb9ae092d7818.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 90, + 442, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ee7814d547abc27da13f7c4bf9d37ad48e4867287555019ac71695bab5244196.jpg", + "image_caption": [ + "Figure 3. Toy experiment exploring semantic point correspondence. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + ], + "image_footnote": [], + "bbox": [ + 96, + 244, + 269, + 412 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cba686cda9108fd02c0a052ba52acb5397c9511a3d670c2fbe0c5671faf9768f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 244, + 442, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions [17, 34, 35, 45, 47]. Tuning-free solutions are fast but typically adhere closely to the provided reference image and lack variation. On the other hand, tuning-based solutions can leverage multi-view images to ensure variation in the given concepts and maintain the same inference behavior to pretrained diffusion models. In this paper, we employ the tuning-based ED-LoRA [13] for encoding subject identity.", + "bbox": [ + 75, + 465, + 468, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. VideoSwap", + "text_level": 1, + "bbox": [ + 76, + 599, + 194, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we start by presenting a toy experiment to illustrate our motivation to explore semantic point correspondence in Sec. 3.1. Subsequently, we offer an overview of the VideoSwap pipeline for customized video subject swapping in Sec. 3.2. Following that, we explain the process of injecting semantic point correspondence in Sec. 3.3 to Sec. 3.5.", + "bbox": [ + 75, + 626, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 726, + 197, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dense correspondences explored in previous video editing methods [5, 10, 12, 28, 46] restrict the subject's shape change in the edited video. Therefore, our goal is to find a more flexible correspondence that can transfer the source subject's motion trajectory without imposing strict shape constraints. Motivated by this, we investigate sparse semantic points as correspondences. Unlike dense correspondences such as depth, edge, and optical flow, which are low-level cues shared across all video subjects, semantic points vary with different open-world concepts. Therefore, it is not", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f20b35764241a52a40230eb245fbb36ed8843baaeb09b6d9eda963d64235adc8.jpg", + "image_caption": [ + "Figure 4. Overview of the VideoSwap pipeline for customized video subject swapping." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 867, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feasible to train a general condition model for injecting semantic point correspondence. Instead, the question we aim to address in this section is, is it possible to learn semantic point control for a specific source video subject using only a small number of source video frames?", + "bbox": [ + 496, + 314, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Toy Experiment Setting. To address the above question, we perform a toy experiment. Firstly, for a given video, we manually define a set of semantic points. Next, we annotate the same set of points on eight frames of this video. Finally, we train a T2I-Adapter [27] on these data pairs, as illustrated in Fig. 3(a, b), to determine whether these semantic points can be used to control the source video subject.", + "bbox": [ + 496, + 390, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 1: Semantic points optimized on source video frames have the potential to align the subject's motion trajectory and change the subject's shape. As shown in Fig. 3(c, left), we drag the points on the cat's face. Despite this edited point map is not in the training data, the resulting image closely follows the adjusted semantic points, effectively generating the unseen pose of the subject. As shown in Fig. 3(c, right), when we drag the boundary semantic points of the car, the edited subject will also reshape to align with the semantic point. This suggests the potential of utilizing semantic points to align the motion trajectory or change the shape when a sequence of dragged points for all video frames is accessible.", + "bbox": [ + 496, + 496, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 2: Semantic points optimized on source video frames can transfer across semantic and low-level changes. As demonstrated in Fig. 3(d), when we replace the source subject with different semantics or modify the low-level information in the prompts, the optimized semantic point can also control the pose or shape of the target concept. This suggests that the semantic point can be transferred across both semantic and low-level changes.", + "bbox": [ + 496, + 691, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview of VideoSwap", + "text_level": 1, + "bbox": [ + 500, + 821, + 718, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Task Formulation", + "text_level": 1, + "bbox": [ + 500, + 845, + 678, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we focus on customized video subject swapping, with the goal of subject replacement and background", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3fc1e12ad5ab97690038e32acda1d2f6506bf862cd97cda5fd3c72db5d4113be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 89, + 472, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/89e0878ed99e66531a16a72d8ec4c1c65df6ba329b75895dea476e7365f22ebe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 90, + 846, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/908dab3b6c39ed46ddcd6d6e15772104e4bb7ff0f0bc8d4d8813031705af5ff8.jpg", + "image_caption": [ + "Figure 5. Pipelines for semantic point extraction (Sec. 3.3) and semantic point registration (Sec. 3.4) in VideoSwap. In semantic point extraction, users define semantic points at a keyframe. We then extract the trajectory and embedding of those semantic points from the video. In semantic point registration, the semantic point embedding is projected by multiple 2-layer learnable MLPs, placed in empty features based on their coordinates, and then added element-wise to the diffusion model as motion guidance." + ], + "image_footnote": [], + "bbox": [ + 117, + 195, + 377, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0915901e804140fd8ecd5901e011a5b0fc9d40795ac510a902f7920173fc7b13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 196, + 841, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "preservation. Subject replacement requires preserving the identity of the target subject in the swapped results, encompassing both its appearance and shape. Simultaneously, background preservation requires the unedited background area to remain the same with the source video. The primary challenge of this task lies in aligning the motion trajectory of the source subject while preserving the identity of the target concept, particularly its shape.", + "bbox": [ + 73, + 387, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Overall Pipeline", + "text_level": 1, + "bbox": [ + 75, + 527, + 243, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The VideoSwap pipeline is illustrated in Fig. 4. Following the latent diffusion model [32], we encode the source video with a VAE encoder to obtain the latent space representation $z_0$ . Subsequently, DDIM inversion [7, 38] is applied to transform the clean latent $z_0$ back to the noisy latent $z_T$ . After obtaining the DDIM inverted noise $z_T$ , we replace the source subject in the text prompt with the target subject and denoise it using the DDIM scheduler [38]. In this denoising process, we introduce semantic point correspondence to guide the subject's motion trajectory, as detailed in Sec. 3.3-Sec. 3.5. To preserve the unedited background, we leverage the concept of latent blending [1, 2], further explained in the supplementary material. Additionally, we incorporate the following designs:", + "bbox": [ + 75, + 551, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adopting Motion Layers. We integrate the motion layers [3, 14] into the image diffusion model to ensure essential temporal consistency for video editing.", + "bbox": [ + 75, + 763, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supporting Predefined and Customized Concepts. We support both predefined concepts from the pretrained model and customized concepts. To create customized concepts, we train ED-LoRA [13] on a set of representative images to encode their identity. After training, these concept ED-LoRAs can be used at the inference time.", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Semantic Point Extraction", + "text_level": 1, + "bbox": [ + 500, + 386, + 736, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first extract the trajectories of semantic points and their associated semantic embeddings from the source video.", + "bbox": [ + 498, + 410, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Trajectory Extraction. As depicted in Fig. 5, for a video containing $N$ frames, users specify $K$ semantic points at a keyframe $i$ . These user-defined semantic points are then propagated to the remaining $N - 1$ frames using a point tracker [18] or detector [4]. Subsequently, the motion trajectory of all semantic points in the entire video is obtained and represented as $\\mathbf{P}_{coord} = \\{Tra(k)|k = 1\\dots K\\}$ , where $Tra(k) = \\{(x_n^k,y_n^k,n)|n = 1\\dots N\\}$ represents the motion trajectory of semantic point $k$ across all $N$ frames.", + "bbox": [ + 498, + 441, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Embedding Extraction. To leverage semantic point correspondence, it is crucial to associate each point with its semantics. Specifically, we extract the DIFT [39] feature $\\mathbf{D}_n$ for each frame $n$ . Subsequently, the point embedding for semantic point $k$ at frame $n$ is obtained as $\\mathbf{v}_n^k = \\mathbf{D}_n(x_n^k, y_n^k)$ , where $(x_n^k, y_n^k)$ is retrieved from $\\mathbf{P}_{coord}$ . Following this, we aggregate the point embeddings acquired from all $N$ frames to obtain the final embedding for each semantic point $k$ by $\\mathbf{v}^k = \\frac{1}{N} \\sum_{n=1}^{N} \\mathbf{v}_n^k$ . Finally, we obtain the semantic embedding for all semantic points, succinctly represented as $\\mathbf{P}_{emb} = \\{\\mathbf{v}^k | k = 1 \\dots K\\}$ .", + "bbox": [ + 498, + 578, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Semantic Point Registration", + "text_level": 1, + "bbox": [ + 500, + 755, + 750, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After acquiring the motion trajectory $\\mathbf{P}_{\\text {coord }}$ and the embedding $\\mathbf{P}_{emb}$ for semantic points, we register these semantic points on the source video to enable them to offer motion guidance for the video subject.", + "bbox": [ + 498, + 779, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparse Motion Feature Creation. To utilize semantic points as guidance, we generate sparse motion features infused with semantic point embeddings, making them compatible with the Unet encoder. We denote $\\mathbf{F}_{enc} =$", + "bbox": [ + 498, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a78a191d8f4d99ed3f4cb95f39aabdeffe36dcacbceac825233386b4e79be1f3.jpg", + "image_caption": [ + "Figure 6. Point displacement propagation based on layered neural atlas (LNA) [16, 19]. Once a trained LNA is provided, users can drag a semantic point at the keyframe, and this displacement is consistently propagated to every frame through the canonical space of the LNA." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 509, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ad7dfb90deafd4f94ea2cf0cefcd132987354a888bd515eb7009d02cbce633a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 90, + 866, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\{\\mathbf{F}_{enc}^{1}, \\mathbf{F}_{enc}^{2}, \\mathbf{F}_{enc}^{3}, \\mathbf{F}_{enc}^{4}\\}$ as the multi-scale intermediate feature of the Unet encoder. For an input VAE-encoded latent with spatial-temporal size $(H, W, N)$ , the feature size of $\\mathbf{F}_{enc}^{l}$ for each Unet stage $l \\in [1, 4]$ can be computed as $(H / 2^l, W / 2^l, N)$ .", + "bbox": [ + 76, + 276, + 468, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 5 and detailed in Algorithm 1, we create a series of multi-scale conditional features $\\mathbf{F}_c = \\{\\mathbf{F}_c^1,\\mathbf{F}_c^2,\\mathbf{F}_c^3,\\mathbf{F}_c^4\\}$ . Notably, $\\mathbf{F}_c$ shares the same size as $\\mathbf{F}_{enc}$ and is initialized with zero vectors. And we introduce a series of learnable MLPs $\\phi = \\{\\phi^l | l \\in \\{1,2,3,4\\}\\}$ , each of $\\phi^l$ project the point embedding to match the feature dimension of the corresponding $\\mathbf{F}_c^l$ . Then, for each point $(x_n^k,y_n^k,n) \\in \\mathbf{P}_{coord}$ , we compute its corresponding spatial-temporal coordinate $(x,y,n)$ at $l$ -th Unet stage and assign the projected embedding based on the coordinate by $\\mathbf{F}_c^l(x,y,n) = \\phi^l(\\mathbf{P}_{emb}(k))$ .", + "bbox": [ + 76, + 354, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is crucial to emphasize that the motion feature $\\mathbf{F}_c$ demonstrates high sparsity, with only the semantic point trajectories containing the feature embeddings. Finally, $\\mathbf{F}_c$ are added element-wise into the intermediate feature $\\mathbf{F}_{enc}$ of Unet encoder as motion guidance:", + "bbox": [ + 75, + 522, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {e n c} ^ {l} = \\mathbf {F} _ {e n c} ^ {l} + \\mathbf {F} _ {c} ^ {l}, l \\in \\{1, 2, 3, 4 \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 598, + 468, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1: Sparse Motion Feature Creation", + "text_level": 1, + "bbox": [ + 86, + 623, + 395, + 638 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "1 Input: \n2 1. Point Trajectory $\\mathbf{P}_{\\text {coord }}$ \n3 2. Point Embedding $\\mathbf{P}_{emb}$ \n4 3. Learnable MLPs $\\phi = \\{\\phi^l | l \\in \\{1,2,3,4\\}\\}$ \n5 Output: \n6 Motion Feature $\\mathbf{F}_c = \\{\\mathbf{F}_c^l | l \\in \\{1,2,3,4\\}\\}$ \n7 Initialize: $\\mathbf{F}_c = \\{\\mathbf{F}_c^l = \\mathbf{0} | l \\in \\{1,2,3,4\\}\\}$ ; \n8 for $l \\gets 1$ to 4 do \n9 foreach $(x_n^k, y_n^k, n)$ in $\\mathbf{P}_{\\text {coord }}$ do \n10 $x, y = \\text{round}(x_n^k / 2^l), \\text{round}(y_n^k / 2^l)$ ; \n11 $\\mathbf{F}_c^l(x, y, n) = \\phi^l(\\mathbf{P}_{emb}(k))$ \n12 end \n13 end", + "bbox": [ + 81, + 642, + 405, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Point Registration on Source Video. Our objective is to optimize the projection MLPs $(\\phi)$ to facilitate better motion guidance from semantic points. This opti", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mization objective is defined as", + "bbox": [ + 500, + 277, + 710, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\phi} E _ {\\epsilon \\sim N (0, I), t \\sim U \\left(T _ {m i n}, T\\right)} | | [ \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, p, \\phi (\\mathbf {P} _ {e m b})\\right) ] \\odot \\Omega (\\mathbf {P} _ {c o o r d}) | | _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 299, + 910, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p$ represents the embedding for the text prompt, and $\\Omega(\\mathbf{P}_{\\text {coord }})$ denotes the binary mask that only turns on around the semantic point.", + "bbox": [ + 496, + 332, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt two techniques to enhance the learning of semantic point correspondences in Eq. (2). The first technique is semantic-enhanced schedule, controlled by $T_{min}$ . We set $T_{min} = T / 2$ to enhance the learning at the higher timestep, which prevents overfitting to low-level details and improves semantic point alignment. The second technique, point patch loss, constrains the computation of the loss to a local patch near each semantic point, which reduces structure leakage into the target swap. This is implemented by the a loss mask $\\Omega$ in Eq. (2).", + "bbox": [ + 496, + 378, + 892, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. User-Point Interaction at Inference Time", + "text_level": 1, + "bbox": [ + 500, + 542, + 848, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sec. 3.3, users define semantic points at the source video keyframe, and we subsequently extract their trajectory and semantic embedding in the video. To utilize these semantic points as correspondences, we register them on the source video in Sec. 3.4. Following these two steps, these semantic points become applicable for controlling the motion of the target object. In this section, we introduce user-point interaction to address various semantic point correspondence.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adopting Source Point Sequence. If there exist one-to-one semantic point correspondence between the source subject and the target subject, such as swapping the dog with $V^{catA}$ as illustrated in Fig. 1, we directly use the source point sequence as the motion guidance.", + "bbox": [ + 496, + 688, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Removing Parts of Semantic Point. When there only exists partial semantic point correspondence between the source subject and the target subject, we can remove redundant points to loosen shape constraints. For example, in scenarios such as swapping an airplane for a helicopter, as depicted in Fig. 1, we remove semantic points associated with the airplane's wings, given that helicopters typically lack wings. The remaining semantic points on the airplane's nose and tail are retained as motion guidance.", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7626", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8501526b5dbd604dac0b1f2f84c124cc3e784e7c4523d1a1cf1e17f333d03137.jpg", + "image_caption": [ + "Concept Images For V/" + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 212, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b715c40586cba0e742b131356cb39c5d6ddc8cbc96265bfae0f17eda5eb2aa00.jpg", + "image_caption": [ + "Source Video" + ], + "image_footnote": [], + "bbox": [ + 223, + 104, + 344, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e113eacf247babb83253dd1c6a5ab044e80d43fdc5a139523d448e547a667d7.jpg", + "image_caption": [ + "VideoSwap (Ours)" + ], + "image_footnote": [], + "bbox": [ + 346, + 93, + 468, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b964b2528271b7131fbaaa44b8ecb0c3a36873374ea62d00f32bb09053ed8915.jpg", + "image_caption": [ + "DDIM" + ], + "image_footnote": [], + "bbox": [ + 470, + 93, + 591, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b27df3855c97a54682f5bfdfd4e8c1f49c6c757ca0b8170dc11a70686d65b78.jpg", + "image_caption": [ + "DDIM + Tune-A-Video" + ], + "image_footnote": [], + "bbox": [ + 596, + 93, + 712, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12f0d99f86b55332f96fa51a9961933acf4a9ee2733d1b459bb07a5c1cf0868d.jpg", + "image_caption": [ + "DDIM + T2I-Adapter" + ], + "image_footnote": [], + "bbox": [ + 718, + 104, + 839, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3837e7d0b2a2a594fb6b7de3a57316329beeabb12912e8e0c2906c217366bd11.jpg", + "image_caption": [ + "Concept Images For VocA" + ], + "image_footnote": [], + "bbox": [ + 122, + 172, + 214, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0922f350a9d82b3199ee8cf1fdc6f909777f6a7188d648aeac7d87216f68d369.jpg", + "image_caption": [ + "Source Prompt: A dog sitting on the side of a car window looking out the window.", + "Source Video" + ], + "image_footnote": [], + "bbox": [ + 223, + 181, + 344, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5af481676066f9c7fbc9e6dfb165c6ab357633ad933196dfa6c9a926d5a6fa6.jpg", + "image_caption": [ + "VideoSwap (Ours)", + "Figure 7. Comparison of VideoSwap with several baselines built upon the same foundational model. The only difference lies in adopting different motion guidance. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + ], + "image_footnote": [], + "bbox": [ + 346, + 181, + 468, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7375fff4493371f23f8f45765ad77852f8bd8d61dc14ddec8b4e165af0834aaa.jpg", + "image_caption": [ + "DDIM" + ], + "image_footnote": [], + "bbox": [ + 470, + 183, + 593, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7c4e97af5bbcfc2b9f74aed6191ed66db33a4844108e12f35c46fd7bfc31858.jpg", + "image_caption": [ + "Target Edit: dog -> VcatA", + "DDIM + Tune-A-Video" + ], + "image_footnote": [], + "bbox": [ + 594, + 183, + 717, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3763ffe3cf74c12835c7b83482679e7cf69bf3d8d1d25bcc3e45470b31c86194.jpg", + "image_caption": [ + "DDIM + T2I-Adapter" + ], + "image_footnote": [], + "bbox": [ + 720, + 183, + 841, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Dragging of Semantic Point.** In situations where there exists semantic point correspondence between the source and target subjects, but misalignment occurs due to shape morphing, we provide users the option to manually drag the semantic points on a keyframe for better alignment of the shape changes. For instance, when swapping a jeep (tall and narrow) for a $V_{carA}$ (sports car, low and wide) as illustrated in Fig. 1, users can drag the semantic points to accurately reflect the shape change.", + "bbox": [ + 75, + 287, + 468, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Editing the shape change by dragging semantic points on a single frame is straightforward. However, consistently propagating these semantic point displacement over time is non-trivial, mainly because of the complex camera and subject motion in the video. Therefore, we introduce point displacement propagation to solve this problem.", + "bbox": [ + 75, + 434, + 468, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 6, we follow the Layered Neural Atlas (LNA) [16, 19] to learn the canonical space, as detailed in the supplementary material. Following LNA [16, 19], we establish a forward coordinate mapping from the video to the canonical space, denoted as $M \\colon (x,y,f) \\to (u,v)$ , along with its corresponding backward mapping $B \\colon (u,v,f) \\to (x,y)$ .", + "bbox": [ + 76, + 537, + 468, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a semantic point, with coordinates at the keyframe $f_{key}$ represented as $(x,y,f_{key})$ , its trajectory over time can be expressed as a function of time $f\\colon (x(f),y(f)) = P(f)$ . Suppose a user drag it to a new position at $(x + dx,y + dy,f_{key})$ , we aim to estimate the edited trajectory $P^{\\prime}(f)$ for $f = \\{0,\\dots,N\\}$ . We resort to LNA's representation, and first compute a linearized estimation of its shifted position on the canonical coordinate:", + "bbox": [ + 76, + 654, + 468, + 772 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n[ d u, d v ] ^ {T} = J _ {M} (x, y, f _ {\\text {k e y}}) [ d x, d y ] ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 773, + 467, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $J_{M}$ denote the Jacobian matrix with respect to $(x,y)$ . Next, at a given time $f$ , we estimate the edited coordinate in the pixel space as", + "bbox": [ + 75, + 792, + 467, + 835 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {\\prime} (f) = P (f) + J _ {B} (u, v, f) [ d u, d v ] ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 837, + 467, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $(u,v) = B(x,y,f_{key})$ and $J_B$ denote the Jacobian matrix with respect to $(u,v)$ . We then use this edited trajectory $P^{\\prime}(f)$ for the dragged semantic point during inference.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef8354adb0354824894ab5a3ce4c7791c24b0a57c7e825d53189256442a54197.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Human Evaluation on Video Subject Swapping Results." + ], + "table_body": "
Methods/MetricsSubject IdentityMotion AlignmentTemporal ConsistencyOverall Preference
VideoSwap v.s.Compare to Previous Video Editing Methods
Tune-A-Video80% v.s. 20%72% v.s. 28%80% v.s. 20%78% v.s. 22%
FateZero74% v.s. 26%67% v.s. 33%73% v.s. 27%70% v.s. 30%
Text2Video-Zero76% v.s. 24%71% v.s. 29%80% v.s. 20%80% v.s. 20%
Rerender-A-Video81% v.s. 19%72% v.s. 28%84% v.s. 16%84% v.s. 16%
Compare to Baselines onAnimateDiff
w/ DDIM70% v.s. 30%74% v.s. 26%69% v.s. 31%73% v.s. 27%
w/ DDIM+TAV68% v.s. 32%60% v.s. 40%66% v.s. 34%69% v.s. 31%
w/ DDIM+T2I-Adapter66% v.s. 34%59% v.s. 41%65% v.s. 35%66% v.s. 34%
", + "bbox": [ + 501, + 285, + 890, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 405, + 630, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement our method using the Latent Diffusion Model [32] and adopt the motion layer inAnimateDiff [14] as the foundational model. All videos consist of 16 frames. Additional implementation details are included in the supplementary material.", + "bbox": [ + 496, + 430, + 890, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 500, + 513, + 720, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with the State-of-the-Art. We qualitatively compare to Tune-A-Video [43], FateZero [31], Rerender-A-Video [46], TokenFlow [12] and StableVideo [5] in Fig. 1. Previous methods are less effective in revealing the correct shape of the target subject. Compared to them, VideoSwap can achieve a significant shape change while aligning the source motion trajectory.", + "bbox": [ + 496, + 537, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Baselines onAnimateDiff. As most state-of-the-art methods are based on image diffusion models, we also compare VideoSwap to several baselines on theAnimateDiff. The only distinctions from VideoSwap lie in different motion guidance, as shown in the results in Fig. 7:", + "bbox": [ + 496, + 643, + 890, + 719 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DDIM: The DDIM sampling without other motion guidance cannot produce the correct motion trajectory.", + "- DDIM+Tune-A-Video: If we tune the model as [43] to inject source motion, it achieves correct motion but suffers from severe structure and appearance leakage.", + "- DDIM+T2I-Adapter: If we add spatial controls [27], such as depth, to control the editing, we observe that 1) the shape is restricted by the source, and 2) the deformable motion cannot follow the source video." + ], + "bbox": [ + 500, + 719, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to all constructed baselines, our VideoSwap with semantic point correspondence can effectively align the motion trajectory while preserving the target concept's identity.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35efebd215b608ff1d1cbf58949505a61f7939ad704f7a8f27204289c90e3c13.jpg", + "image_caption": [ + "Figure 8. Ablation Study of our VideoSwap. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 823, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 375, + 308, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct both automatic and human evaluations to quantitatively compare VideoSwap with previous state-of-the-art methods and several baselines onAnimateDiff. Detailed evaluation settings and automatic evaluation results are provided in the supplementary material. For human evaluation, we distribute 1000 questionnaires on Amazon Mturk to assess various criteria in customized video subject swapping. From the human evaluation results in Table. 1, our method achieves a clear advantage over the compared methods.", + "bbox": [ + 75, + 398, + 470, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 544, + 230, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse Motion Feature. There are several variants for encoding semantic points to generate motion guidance. The most straightforward approach is to encode the point map of semantic points using T2I-Adapter [27]. However, this method leads to severe overfitting, as illustrated in Fig. 8(a). The issue arises because the encoded feature added to the diffusion model is non-sparse, and the background also contains features that may overfit to the source video. We use MLPs to encode point embeddings, positioning them in the empty feature to create sparse motion features for guidance without compromising video quality. Regarding point embeddings, we opt for DIFT embeddings, which inherently carry robust semantics. Compared to randomly initialized learnable embeddings, our approach achieves similar results with $3 \\times$ less time for point registration step.", + "bbox": [ + 75, + 568, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point Patch Loss. In the semantic point registration, we employ a point patch loss to reconstruct the local patch surrounding each semantic point. If we omit the point patch loss and opt to directly reconstruct the entire video, we observe that the source structure leaks into the target swapped results and thus produce artifacts, as depicted in Fig. 8(b).", + "bbox": [ + 75, + 795, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic-Enhanced Schedule. Our goal is to employ se", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mantic points to transfer motion trajectories, acting as a linkage between the source and target subjects. Therefore, we aim for the semantic points to emphasize high-level semantic alignment without transferring low-level details. This objective is achieved by registering points only on the early sampling steps, i.e., $T_{\\min} = \\frac{T}{2}$ in Eq. (2). As shown in Fig. 8(c), this technique prevents the model from learning excessive low-level details and enhances point alignment.", + "bbox": [ + 496, + 376, + 890, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Drag-based Point Control. As illustrated in Fig. 8(d), our goal is to swap the black swan to the duck. If we directly use the source point sequence as guidance, the duck's neck conforms to the shape of the black swan, resulting in an inferior identity. However, by employing the proposed point displacement propagation, we can drag the semantic point at the keyframe, ensuring a consistent motion trajectory after dragging. Utilizing the dragged point trajectory as motion guidance results in the correct identity of the duck.", + "bbox": [ + 496, + 496, + 892, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 645, + 617, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper uncovers the potential of semantic point correspondence in aligning motion trajectories and altering the subject's identity in video editing. From there, we present VideoSwap, a framework that minimizes human intervention while utilizing semantic point correspondences for customized video subject swapping. Through user-point interactions like point removal or dragging, we address various semantic point correspondence. VideoSwap facilitates shape changes in the target swap while aligning the motion trajectory with the source subject, demonstrating state-of-the-art results in customized video subject swapping.", + "bbox": [ + 496, + 671, + 890, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008.", + "bbox": [ + 498, + 854, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7628", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5", + "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), 2023. 5", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3, 5", + "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 5", + "[5] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 1, 2, 3, 4, 7", + "[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. Videdit: Zero-shot and spatially aware text-driven video editing. arXiv preprint arXiv:2306.08707, 2023. 2, 3", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5", + "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 3", + "[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 3", + "[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2, 3, 4", + "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3", + "[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3, 4, 7", + "[13] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292, 2023. 2, 3, 4, 5" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3, 5, 7", + "[15] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[16] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 6, 7", + "[17] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 4", + "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 3, 5", + "[19] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3, 6, 7", + "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3", + "[21] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14317-14326, 2023. 3", + "[22] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 1, 2", + "[23] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 2, 3", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1", + "[26] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 3", + "[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3, 4", + "[29] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3", + "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3", + "[31] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 1, 2, 3, 7", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 7", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3", + "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 4", + "[35] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 4", + "[36] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3", + "[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[39] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023. 3, 5", + "[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $p+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3", + "[42] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3", + "[43] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1, 2, 3, 7", + "[44] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 1", + "[45] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 4", + "[46] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 2, 3, 4, 7", + "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4", + "[48] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 3", + "[49] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023. 1, 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 627 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7630", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_model.json b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d535f5389144f4792af86f6675164d90fc61315 --- /dev/null +++ b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.105, + 0.761, + 0.15 + ], + "angle": 0, + "content": "VideoSwap: Customized Video Subject Swapping with Interactive Semantic Point Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.177, + 0.907, + 0.216 + ], + "angle": 0, + "content": "Yuchao Gu\\(^{1,2}\\), Yipin Zhou\\(^{2}\\), Bichen Wu\\(^{2}\\), Licheng Yu\\(^{2}\\), Jia-Wei Liu\\(^{1}\\), Rui Zhao\\(^{1}\\), Jay Zhangjie Wu\\(^{1}\\), David Junhao Zhang\\(^{1}\\), Mike Zheng Shou\\(^{1*}\\), Kevin Tang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.231, + 0.746, + 0.251 + ], + "angle": 0, + "content": "1Show Lab, National University of Singapore 2GenAI, Meta" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.251, + 0.662, + 0.267 + ], + "angle": 0, + "content": "https://videoswap.github.io/" + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.279, + 0.339, + 0.291 + ], + "angle": 0, + "content": "One-to-One Point Correspondence" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.293, + 0.849, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.581, + 0.674, + 0.594 + ], + "angle": 0, + "content": "(b) Results of Various Correspondence Used in Previous Video Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.597, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Figure 1. Customized video subject swapping results with VideoSwap. VideoSwap supports shape change in the swapped results while aligning with the source motion trajectory. The swapped target can be either a predefined concept from a pretrained model (e.g., helicopter) or a customized concept (denoted by \\( V^{*} \\)). Previous methods based on implicit motion encoding and dense correspondence do not perform well in subject swapping with shape changes. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.686, + 0.314, + 0.703 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.871 + ], + "angle": 0, + "content": "Current diffusion-based video editing primarily focuses on structure-preserved editing by utilizing various dense correspondences to ensure temporal consistency and motion alignment. However, these approaches are often ineffective when the target edit involves a shape change. To embark on video editing with shape change, we explore customized video subject swapping in this work, where we aim to replace the main subject in a source video with a target subject having a distinct identity and potentially different shape. In contrast to previous methods that rely on dense" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.673, + 0.894, + 0.81 + ], + "angle": 0, + "content": "correspondences, we introduce the VideoSwap framework that exploits semantic point correspondences, inspired by our observation that only a small number of semantic points are necessary to align the subject's motion trajectory and modify its shape. We also introduce various user-point interactions (e.g., removing points and dragging points) to address various semantic point correspondence. Extensive experiments demonstrate state-of-the-art video subject swapping results across a variety of real-world videos." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.844, + 0.632, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Diffusion-based video editing [5, 10, 22, 25, 31, 43, 44, 49] is an emerging field that harnesses the capabilities of pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7621" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.848, + 0.705 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.707, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Figure 2. Customized video subject swapping results of VideoSwap on various concepts. The swapping target can either be a predefined concept in the pretrained model (e.g., helicopter) or a customized concept created by ED-LoRA [13] (denoted as \\( V^{*} \\)). We encourage readers to click and play the video clips in this figure using Adobe Acrobat. For legal issues, we cannot display the human swap results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.473, + 0.901 + ], + "angle": 0, + "content": "trained text-to-image/video diffusion models [14, 15, 32, 37] to facilitate a range of video editing tasks, including style change and subject/background swapping. The main challenge in video editing lies in how to extract motion from the source video and transfer it to the edited video while ensuring temporal consistency. Pioneer Tune-A-Video [43] implicitly encodes source motion in the diffusion model weights by tuning from the source video. While it demonstrates versatile applications for video editing, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.884 + ], + "angle": 0, + "content": "temporal consistency is far from satisfactory. Subsequent works make use of various dense correspondences extracted from the source video, including attention maps [23, 31], edge/depth maps [10, 22, 46, 49], optical flows [46], and deformation fields [5, 6] for video editing. While achieving better temporal consistency, dense correspondence imposes strict shape constraints on the target edit, which makes it ineffective for video editing with shape changes." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To embark on video editing with shape change, we delve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7622" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "into a challenging task: customized video subject swapping. Unlike the conventional video subject swapping addressed in previous works [5, 46, 49], where the swapped subject conforms to the shape of source subject, customized subjects have a clearly defined identity in terms of both appearance and shape. These distinctive characteristics should be preserved in the target edit. Therefore, previous structure-preserved video editing methods are often ineffective for this problem, as shown in Fig. 1(b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.473, + 0.501 + ], + "angle": 0, + "content": "To address customized video subject swapping, our primary insight is that the subject's motion trajectory can be effectively described using a small number of semantic points. As shown in Fig. 1(a), the motion trajectory of an airplane can be depicted by semantic points located at its wings, nose, and tail. This insight naturally leads us to the following question: Can we utilize these semantic points as correspondences to align the motion trajectory while relaxing the shape constraints in video editing? To answer this question, we conduct a toy experiments and observe that it is possible to learn semantic point correspondence for a specific video subject using just a small number of source video frames. Users can interact with learned semantic points to generate unseen poses or modify the shape of the video subject. These observations suggest the potential for integrating semantic point correspondence into video editing, provided that we can obtain an accurate semantic point sequence for the target edit." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.652 + ], + "angle": 0, + "content": "To unleash the potential of semantic point correspondence, we introduce the VideoSwap framework for customized video subject swapping, which comprises the following primary designs: 1) Integrating the motion layer into the image diffusion model to ensure essential temporal consistency. 2) Registering semantic points on the source video and utilizing them to transfer the motion trajectory of source subject to the target edit. 3) Introducing user-point interactions (e.g., removing or dragging points) for various semantic point correspondence." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.653, + 0.4, + 0.667 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.669, + 0.47, + 0.715 + ], + "angle": 0, + "content": "- Empirical observations that reveal the potential of semantic point correspondence for aligning motion trajectories and changing shapes in video editing." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.715, + 0.47, + 0.76 + ], + "angle": 0, + "content": "- The VideoSwap framework, which minimizes user intervention while unleashing the potential of semantic point correspondence in customized video subject swapping." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.76, + 0.469, + 0.79 + ], + "angle": 0, + "content": "- State-of-the-art results in customized video subject swapping, as demonstrated in Fig. 2." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.669, + 0.47, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.22, + 0.821 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.348, + 0.848 + ], + "angle": 0, + "content": "2.1. Diffusion-Based Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Structure-Preserved Video Editing. FateZero [31] and Video-P2P [23] extract cross- and self-attention from the source video to control spatial layout. To achieve stricter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "alignment of temporal consistency with the source video, Rerender-A-Video [46], Gen-1 [10], ControlVideo [49], and TokenFlow [12] extract and align optical flow, depth/edge maps, and nearest-neighbour field from the source video respectively, resulting in improved temporal consistency. StableVideo [5], VidEdit [6], and CoDEF [28] learn the canonical space for editing following the Layered Neural Atlas [19] or the deformation field in Dynamic Nerf [24, 30]. While achieving promising results in structure-preserved video editing, the above methods based on various dense correspondence are not suitable for handling subject swapping involving shape changes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Video Editing with Shape Change. Tune-A-Video (TAV) [43] and FateZero [31] with TAV checkpoint can be utilized for video editing with shape change, as they implicitly encode the motion in model weights through tuning on the source video. However, they suffer from structure and appearance leakage due to model tuning. Shape-aware editing [21] is built on the Layered Neural Atlas [19], employing semantic correspondences to estimate shape deformation and warp the atlas. Nevertheless, texture warping will lead to unrealistic textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Video Diffusion Models. Previous video editing primarily relies on the text-to-image diffusion model [32]. However, recent advancements are occurring in video foundation models [3, 14, 41, 48]. In this work, we add a motion layer [14] to the image diffusion model to provide essential temporal consistency for video editing, and we focus on exploiting semantic point correspondence to align the motion trajectory of the video subject." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.707, + 0.571 + ], + "angle": 0, + "content": "2.2. Point Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Point Correspondence in Diffusion Models. DIFT [39] initially uncovers robust semantic point correspondences in diffusion models. Building upon the observation in DIFT, subsequent works, DragDiffusion [36] and DragonDiffusion [26] extend DragGAN [29] to support interactive point-based image editing in diffusion models. However, point correspondences and interactive drag-based editing are seldom investigated for video editing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.836 + ], + "angle": 0, + "content": "Tracking Any Point in Video. TAP-Vid [8] first introduces the problem of tracking any point (TAP). Unlike optical flow estimation, TAP requires the establishment of long-range correspondence for all points in a video. In our work, we use TAP to reduce human intervention in annotating subject keypoints and acquire long-range motion estimation. Although several works [8, 9, 18, 42] address the TAP problem, we choose to employ Co-Tracker [18], which is the most efficient solution available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.717, + 0.862 + ], + "angle": 0, + "content": "2.3. Concept Customization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Concept customization is mainly categorized into tuning-based approaches [11, 13, 20, 33, 40] and tuning-free solu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.264, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.091, + 0.444, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.246, + 0.27, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.246, + 0.444, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.459 + ], + "angle": 0, + "content": "Figure 3. Toy experiment exploring semantic point correspondence. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.574 + ], + "angle": 0, + "content": "tions [17, 34, 35, 45, 47]. Tuning-free solutions are fast but typically adhere closely to the provided reference image and lack variation. On the other hand, tuning-based solutions can leverage multi-view images to ensure variation in the given concepts and maintain the same inference behavior to pretrained diffusion models. In this paper, we employ the tuning-based ED-LoRA [13] for encoding subject identity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.196, + 0.619 + ], + "angle": 0, + "content": "3. VideoSwap" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.469, + 0.718 + ], + "angle": 0, + "content": "In this section, we start by presenting a toy experiment to illustrate our motivation to explore semantic point correspondence in Sec. 3.1. Subsequently, we offer an overview of the VideoSwap pipeline for customized video subject swapping in Sec. 3.2. Following that, we explain the process of injecting semantic point correspondence in Sec. 3.3 to Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.199, + 0.741 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Dense correspondences explored in previous video editing methods [5, 10, 12, 28, 46] restrict the subject's shape change in the edited video. Therefore, our goal is to find a more flexible correspondence that can transfer the source subject's motion trajectory without imposing strict shape constraints. Motivated by this, we investigate sparse semantic points as correspondences. Unlike dense correspondences such as depth, edge, and optical flow, which are low-level cues shared across all video subjects, semantic points vary with different open-world concepts. Therefore, it is not" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.869, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Figure 4. Overview of the VideoSwap pipeline for customized video subject swapping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.39 + ], + "angle": 0, + "content": "feasible to train a general condition model for injecting semantic point correspondence. Instead, the question we aim to address in this section is, is it possible to learn semantic point control for a specific source video subject using only a small number of source video frames?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Toy Experiment Setting. To address the above question, we perform a toy experiment. Firstly, for a given video, we manually define a set of semantic points. Next, we annotate the same set of points on eight frames of this video. Finally, we train a T2I-Adapter [27] on these data pairs, as illustrated in Fig. 3(a, b), to determine whether these semantic points can be used to control the source video subject." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Observation 1: Semantic points optimized on source video frames have the potential to align the subject's motion trajectory and change the subject's shape. As shown in Fig. 3(c, left), we drag the points on the cat's face. Despite this edited point map is not in the training data, the resulting image closely follows the adjusted semantic points, effectively generating the unseen pose of the subject. As shown in Fig. 3(c, right), when we drag the boundary semantic points of the car, the edited subject will also reshape to align with the semantic point. This suggests the potential of utilizing semantic points to align the motion trajectory or change the shape when a sequence of dragged points for all video frames is accessible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Observation 2: Semantic points optimized on source video frames can transfer across semantic and low-level changes. As demonstrated in Fig. 3(d), when we replace the source subject with different semantics or modify the low-level information in the prompts, the optimized semantic point can also control the pose or shape of the target concept. This suggests that the semantic point can be transferred across both semantic and low-level changes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.72, + 0.839 + ], + "angle": 0, + "content": "3.2. Overview of VideoSwap" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.679, + 0.86 + ], + "angle": 0, + "content": "3.2.1 Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we focus on customized video subject swapping, with the goal of subject replacement and background" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.473, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.091, + 0.847, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.196, + 0.379, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.197, + 0.842, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.324, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Figure 5. Pipelines for semantic point extraction (Sec. 3.3) and semantic point registration (Sec. 3.4) in VideoSwap. In semantic point extraction, users define semantic points at a keyframe. We then extract the trajectory and embedding of those semantic points from the video. In semantic point registration, the semantic point embedding is projected by multiple 2-layer learnable MLPs, placed in empty features based on their coordinates, and then added element-wise to the diffusion model as motion guidance." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.388, + 0.47, + 0.509 + ], + "angle": 0, + "content": "preservation. Subject replacement requires preserving the identity of the target subject in the swapped results, encompassing both its appearance and shape. Simultaneously, background preservation requires the unedited background area to remain the same with the source video. The primary challenge of this task lies in aligning the motion trajectory of the source subject while preserving the identity of the target concept, particularly its shape." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.529, + 0.245, + 0.543 + ], + "angle": 0, + "content": "3.2.2 Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.764 + ], + "angle": 0, + "content": "The VideoSwap pipeline is illustrated in Fig. 4. Following the latent diffusion model [32], we encode the source video with a VAE encoder to obtain the latent space representation \\( z_0 \\). Subsequently, DDIM inversion [7, 38] is applied to transform the clean latent \\( z_0 \\) back to the noisy latent \\( z_T \\). After obtaining the DDIM inverted noise \\( z_T \\), we replace the source subject in the text prompt with the target subject and denoise it using the DDIM scheduler [38]. In this denoising process, we introduce semantic point correspondence to guide the subject's motion trajectory, as detailed in Sec. 3.3-Sec. 3.5. To preserve the unedited background, we leverage the concept of latent blending [1, 2], further explained in the supplementary material. Additionally, we incorporate the following designs:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Adopting Motion Layers. We integrate the motion layers [3, 14] into the image diffusion model to ensure essential temporal consistency for video editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Supporting Predefined and Customized Concepts. We support both predefined concepts from the pretrained model and customized concepts. To create customized concepts, we train ED-LoRA [13] on a set of representative images to encode their identity. After training, these concept ED-LoRAs can be used at the inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.738, + 0.401 + ], + "angle": 0, + "content": "3.3. Semantic Point Extraction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.891, + 0.44 + ], + "angle": 0, + "content": "We first extract the trajectories of semantic points and their associated semantic embeddings from the source video." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.442, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Point Trajectory Extraction. As depicted in Fig. 5, for a video containing \\(N\\) frames, users specify \\(K\\) semantic points at a keyframe \\(i\\). These user-defined semantic points are then propagated to the remaining \\(N - 1\\) frames using a point tracker [18] or detector [4]. Subsequently, the motion trajectory of all semantic points in the entire video is obtained and represented as \\(\\mathbf{P}_{coord} = \\{Tra(k)|k = 1\\dots K\\}\\), where \\(Tra(k) = \\{(x_n^k,y_n^k,n)|n = 1\\dots N\\}\\) represents the motion trajectory of semantic point \\(k\\) across all \\(N\\) frames." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.579, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Point Embedding Extraction. To leverage semantic point correspondence, it is crucial to associate each point with its semantics. Specifically, we extract the DIFT [39] feature \\(\\mathbf{D}_n\\) for each frame \\(n\\). Subsequently, the point embedding for semantic point \\(k\\) at frame \\(n\\) is obtained as \\(\\mathbf{v}_n^k = \\mathbf{D}_n(x_n^k, y_n^k)\\), where \\((x_n^k, y_n^k)\\) is retrieved from \\(\\mathbf{P}_{coord}\\). Following this, we aggregate the point embeddings acquired from all \\(N\\) frames to obtain the final embedding for each semantic point \\(k\\) by \\(\\mathbf{v}^k = \\frac{1}{N} \\sum_{n=1}^{N} \\mathbf{v}_n^k\\). Finally, we obtain the semantic embedding for all semantic points, succinctly represented as \\(\\mathbf{P}_{emb} = \\{\\mathbf{v}^k | k = 1 \\dots K\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.75, + 0.772 + ], + "angle": 0, + "content": "3.4. Semantic Point Registration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.891, + 0.84 + ], + "angle": 0, + "content": "After acquiring the motion trajectory \\(\\mathbf{P}_{\\text {coord }}\\) and the embedding \\(\\mathbf{P}_{emb}\\) for semantic points, we register these semantic points on the source video to enable them to offer motion guidance for the video subject." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sparse Motion Feature Creation. To utilize semantic points as guidance, we generate sparse motion features infused with semantic point embeddings, making them compatible with the Unet encoder. We denote \\(\\mathbf{F}_{enc} =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7625" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.51, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.092, + 0.867, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Figure 6. Point displacement propagation based on layered neural atlas (LNA) [16, 19]. Once a trained LNA is provided, users can drag a semantic point at the keyframe, and this displacement is consistently propagated to every frame through the canonical space of the LNA." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.277, + 0.47, + 0.354 + ], + "angle": 0, + "content": "\\(\\{\\mathbf{F}_{enc}^{1}, \\mathbf{F}_{enc}^{2}, \\mathbf{F}_{enc}^{3}, \\mathbf{F}_{enc}^{4}\\}\\) as the multi-scale intermediate feature of the Unet encoder. For an input VAE-encoded latent with spatial-temporal size \\((H, W, N)\\), the feature size of \\(\\mathbf{F}_{enc}^{l}\\) for each Unet stage \\(l \\in [1, 4]\\) can be computed as \\((H / 2^l, W / 2^l, N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.47, + 0.522 + ], + "angle": 0, + "content": "As depicted in Fig. 5 and detailed in Algorithm 1, we create a series of multi-scale conditional features \\(\\mathbf{F}_c = \\{\\mathbf{F}_c^1,\\mathbf{F}_c^2,\\mathbf{F}_c^3,\\mathbf{F}_c^4\\}\\). Notably, \\(\\mathbf{F}_c\\) shares the same size as \\(\\mathbf{F}_{enc}\\) and is initialized with zero vectors. And we introduce a series of learnable MLPs \\(\\phi = \\{\\phi^l | l \\in \\{1,2,3,4\\}\\}\\), each of \\(\\phi^l\\) project the point embedding to match the feature dimension of the corresponding \\(\\mathbf{F}_c^l\\). Then, for each point \\((x_n^k,y_n^k,n) \\in \\mathbf{P}_{coord}\\), we compute its corresponding spatial-temporal coordinate \\((x,y,n)\\) at \\(l\\)-th Unet stage and assign the projected embedding based on the coordinate by \\(\\mathbf{F}_c^l(x,y,n) = \\phi^l(\\mathbf{P}_{emb}(k))\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.599 + ], + "angle": 0, + "content": "It is crucial to emphasize that the motion feature \\(\\mathbf{F}_c\\) demonstrates high sparsity, with only the semantic point trajectories containing the feature embeddings. Finally, \\(\\mathbf{F}_c\\) are added element-wise into the intermediate feature \\(\\mathbf{F}_{enc}\\) of Unet encoder as motion guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.599, + 0.469, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {e n c} ^ {l} = \\mathbf {F} _ {e n c} ^ {l} + \\mathbf {F} _ {c} ^ {l}, l \\in \\{1, 2, 3, 4 \\}. \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.624, + 0.396, + 0.639 + ], + "angle": 0, + "content": "Algorithm 1: Sparse Motion Feature Creation" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.643, + 0.406, + 0.846 + ], + "angle": 0, + "content": "1 Input: \n2 1. Point Trajectory \\(\\mathbf{P}_{\\text {coord }}\\) \n3 2. Point Embedding \\(\\mathbf{P}_{emb}\\) \n4 3. Learnable MLPs \\(\\phi = \\{\\phi^l | l \\in \\{1,2,3,4\\}\\}\\) \n5 Output: \n6 Motion Feature \\(\\mathbf{F}_c = \\{\\mathbf{F}_c^l | l \\in \\{1,2,3,4\\}\\}\\) \n7 Initialize: \\(\\mathbf{F}_c = \\{\\mathbf{F}_c^l = \\mathbf{0} | l \\in \\{1,2,3,4\\}\\}\\); \n8 for \\(l \\gets 1\\) to 4 do \n9 foreach \\((x_n^k, y_n^k, n)\\) in \\(\\mathbf{P}_{\\text {coord }}\\) do \n10 \\(x, y = \\text{round}(x_n^k / 2^l), \\text{round}(y_n^k / 2^l)\\); \n11 \\(\\mathbf{F}_c^l(x, y, n) = \\phi^l(\\mathbf{P}_{emb}(k))\\) \n12 end \n13 end" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Semantic Point Registration on Source Video. Our objective is to optimize the projection MLPs \\((\\phi)\\) to facilitate better motion guidance from semantic points. This opti" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.278, + 0.712, + 0.293 + ], + "angle": 0, + "content": "mization objective is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.3, + 0.911, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\phi} E _ {\\epsilon \\sim N (0, I), t \\sim U \\left(T _ {m i n}, T\\right)} | | [ \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, p, \\phi (\\mathbf {P} _ {e m b})\\right) ] \\odot \\Omega (\\mathbf {P} _ {c o o r d}) | | _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.378 + ], + "angle": 0, + "content": "where \\( p \\) represents the embedding for the text prompt, and \\( \\Omega(\\mathbf{P}_{\\text {coord }}) \\) denotes the binary mask that only turns on around the semantic point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.893, + 0.53 + ], + "angle": 0, + "content": "We adopt two techniques to enhance the learning of semantic point correspondences in Eq. (2). The first technique is semantic-enhanced schedule, controlled by \\( T_{min} \\). We set \\( T_{min} = T / 2 \\) to enhance the learning at the higher timestep, which prevents overfitting to low-level details and improves semantic point alignment. The second technique, point patch loss, constrains the computation of the loss to a local patch near each semantic point, which reduces structure leakage into the target swap. This is implemented by the a loss mask \\( \\Omega \\) in Eq. (2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.849, + 0.558 + ], + "angle": 0, + "content": "3.5. User-Point Interaction at Inference Time" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.688 + ], + "angle": 0, + "content": "In Sec. 3.3, users define semantic points at the source video keyframe, and we subsequently extract their trajectory and semantic embedding in the video. To utilize these semantic points as correspondences, we register them on the source video in Sec. 3.4. Following these two steps, these semantic points become applicable for controlling the motion of the target object. In this section, we introduce user-point interaction to address various semantic point correspondence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Adopting Source Point Sequence. If there exist one-to-one semantic point correspondence between the source subject and the target subject, such as swapping the dog with \\( V^{catA} \\) as illustrated in Fig. 1, we directly use the source point sequence as the motion guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Removing Parts of Semantic Point. When there only exists partial semantic point correspondence between the source subject and the target subject, we can remove redundant points to loosen shape constraints. For example, in scenarios such as swapping an airplane for a helicopter, as depicted in Fig. 1, we remove semantic points associated with the airplane's wings, given that helicopters typically lack wings. The remaining semantic points on the airplane's nose and tail are retained as motion guidance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.091, + 0.214, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.161, + 0.223, + 0.17 + ], + "angle": 0, + "content": "Concept Images For V/" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.105, + 0.346, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.162, + 0.313, + 0.169 + ], + "angle": 0, + "content": "Source Video" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.094, + 0.469, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.162, + 0.447, + 0.169 + ], + "angle": 0, + "content": "VideoSwap (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.094, + 0.593, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.162, + 0.547, + 0.169 + ], + "angle": 0, + "content": "DDIM" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.094, + 0.714, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.162, + 0.709, + 0.169 + ], + "angle": 0, + "content": "DDIM + Tune-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.105, + 0.841, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.162, + 0.828, + 0.17 + ], + "angle": 0, + "content": "DDIM + T2I-Adapter" + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.173, + 0.628, + 0.182 + ], + "angle": 0, + "content": "Source Prompt: A dog sitting on the side of a car window looking out the window." + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.174, + 0.741, + 0.182 + ], + "angle": 0, + "content": "Target Edit: dog -> VcatA" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.174, + 0.215, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.24, + 0.226, + 0.248 + ], + "angle": 0, + "content": "Concept Images For VocA" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.183, + 0.346, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.24, + 0.313, + 0.247 + ], + "angle": 0, + "content": "Source Video" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.183, + 0.47, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.24, + 0.447, + 0.248 + ], + "angle": 0, + "content": "VideoSwap (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.184, + 0.594, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.24, + 0.547, + 0.247 + ], + "angle": 0, + "content": "DDIM" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.184, + 0.718, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.24, + 0.708, + 0.248 + ], + "angle": 0, + "content": "DDIM + Tune-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.184, + 0.842, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.24, + 0.828, + 0.248 + ], + "angle": 0, + "content": "DDIM + T2I-Adapter" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Figure 7. Comparison of VideoSwap with several baselines built upon the same foundational model. The only difference lies in adopting different motion guidance. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.469, + 0.424 + ], + "angle": 0, + "content": "**Dragging of Semantic Point.** In situations where there exists semantic point correspondence between the source and target subjects, but misalignment occurs due to shape morphing, we provide users the option to manually drag the semantic points on a keyframe for better alignment of the shape changes. For instance, when swapping a jeep (tall and narrow) for a \\( V_{carA} \\) (sports car, low and wide) as illustrated in Fig. 1, users can drag the semantic points to accurately reflect the shape change." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.469, + 0.526 + ], + "angle": 0, + "content": "Editing the shape change by dragging semantic points on a single frame is straightforward. However, consistently propagating these semantic point displacement over time is non-trivial, mainly because of the complex camera and subject motion in the video. Therefore, we introduce point displacement propagation to solve this problem." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.469, + 0.643 + ], + "angle": 0, + "content": "As illustrated in Fig. 6, we follow the Layered Neural Atlas (LNA) [16, 19] to learn the canonical space, as detailed in the supplementary material. Following LNA [16, 19], we establish a forward coordinate mapping from the video to the canonical space, denoted as \\( M \\colon (x,y,f) \\to (u,v) \\), along with its corresponding backward mapping \\( B \\colon (u,v,f) \\to (x,y) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.655, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Given a semantic point, with coordinates at the keyframe \\( f_{key} \\) represented as \\( (x,y,f_{key}) \\), its trajectory over time can be expressed as a function of time \\( f\\colon (x(f),y(f)) = P(f) \\). Suppose a user drag it to a new position at \\( (x + dx,y + dy,f_{key}) \\), we aim to estimate the edited trajectory \\( P^{\\prime}(f) \\) for \\( f = \\{0,\\dots,N\\} \\). We resort to LNA's representation, and first compute a linearized estimation of its shifted position on the canonical coordinate:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.775, + 0.468, + 0.793 + ], + "angle": 0, + "content": "\\[\n[ d u, d v ] ^ {T} = J _ {M} (x, y, f _ {\\text {k e y}}) [ d x, d y ] ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.468, + 0.837 + ], + "angle": 0, + "content": "where \\(J_{M}\\) denote the Jacobian matrix with respect to \\((x,y)\\). Next, at a given time \\(f\\), we estimate the edited coordinate in the pixel space as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.838, + 0.468, + 0.855 + ], + "angle": 0, + "content": "\\[\nP ^ {\\prime} (f) = P (f) + J _ {B} (u, v, f) [ d u, d v ] ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\((u,v) = B(x,y,f_{key})\\) and \\(J_B\\) denote the Jacobian matrix with respect to \\((u,v)\\). We then use this edited trajectory \\(P^{\\prime}(f)\\) for the dragged semantic point during inference." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.286, + 0.892, + 0.389 + ], + "angle": 0, + "content": "
Methods/MetricsSubject IdentityMotion AlignmentTemporal ConsistencyOverall Preference
VideoSwap v.s.Compare to Previous Video Editing Methods
Tune-A-Video80% v.s. 20%72% v.s. 28%80% v.s. 20%78% v.s. 22%
FateZero74% v.s. 26%67% v.s. 33%73% v.s. 27%70% v.s. 30%
Text2Video-Zero76% v.s. 24%71% v.s. 29%80% v.s. 20%80% v.s. 20%
Rerender-A-Video81% v.s. 19%72% v.s. 28%84% v.s. 16%84% v.s. 16%
Compare to Baselines onAnimateDiff
w/ DDIM70% v.s. 30%74% v.s. 26%69% v.s. 31%73% v.s. 27%
w/ DDIM+TAV68% v.s. 32%60% v.s. 40%66% v.s. 34%69% v.s. 31%
w/ DDIM+T2I-Adapter66% v.s. 34%59% v.s. 41%65% v.s. 35%66% v.s. 34%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.503, + 0.39, + 0.886, + 0.403 + ], + "angle": 0, + "content": "Table 1. Human Evaluation on Video Subject Swapping Results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.406, + 0.632, + 0.422 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.506 + ], + "angle": 0, + "content": "We implement our method using the Latent Diffusion Model [32] and adopt the motion layer inAnimateDiff [14] as the foundational model. All videos consist of 16 frames. Additional implementation details are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.722, + 0.531 + ], + "angle": 0, + "content": "4.1. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Comparison with the State-of-the-Art. We qualitatively compare to Tune-A-Video [43], FateZero [31], Rerender-A-Video [46], TokenFlow [12] and StableVideo [5] in Fig. 1. Previous methods are less effective in revealing the correct shape of the target subject. Compared to them, VideoSwap can achieve a significant shape change while aligning the source motion trajectory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.72 + ], + "angle": 0, + "content": "Comparison with Baselines onAnimateDiff. As most state-of-the-art methods are based on image diffusion models, we also compare VideoSwap to several baselines on theAnimateDiff. The only distinctions from VideoSwap lie in different motion guidance, as shown in the results in Fig. 7:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.72, + 0.89, + 0.75 + ], + "angle": 0, + "content": "- DDIM: The DDIM sampling without other motion guidance cannot produce the correct motion trajectory." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.75, + 0.89, + 0.796 + ], + "angle": 0, + "content": "- DDIM+Tune-A-Video: If we tune the model as [43] to inject source motion, it achieves correct motion but suffers from severe structure and appearance leakage." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.796, + 0.891, + 0.854 + ], + "angle": 0, + "content": "- DDIM+T2I-Adapter: If we add spatial controls [27], such as depth, to control the editing, we observe that 1) the shape is restricted by the source, and 2) the deformable motion cannot follow the source video." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.72, + 0.891, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Compared to all constructed baselines, our VideoSwap with semantic point correspondence can effectively align the motion trajectory while preserving the target concept's identity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.825, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.354, + 0.885, + 0.368 + ], + "angle": 0, + "content": "Figure 8. Ablation Study of our VideoSwap. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.31, + 0.391 + ], + "angle": 0, + "content": "4.2. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.535 + ], + "angle": 0, + "content": "We conduct both automatic and human evaluations to quantitatively compare VideoSwap with previous state-of-the-art methods and several baselines onAnimateDiff. Detailed evaluation settings and automatic evaluation results are provided in the supplementary material. For human evaluation, we distribute 1000 questionnaires on Amazon Mturk to assess various criteria in customized video subject swapping. From the human evaluation results in Table. 1, our method achieves a clear advantage over the compared methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.231, + 0.56 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Sparse Motion Feature. There are several variants for encoding semantic points to generate motion guidance. The most straightforward approach is to encode the point map of semantic points using T2I-Adapter [27]. However, this method leads to severe overfitting, as illustrated in Fig. 8(a). The issue arises because the encoded feature added to the diffusion model is non-sparse, and the background also contains features that may overfit to the source video. We use MLPs to encode point embeddings, positioning them in the empty feature to create sparse motion features for guidance without compromising video quality. Regarding point embeddings, we opt for DIFT embeddings, which inherently carry robust semantics. Compared to randomly initialized learnable embeddings, our approach achieves similar results with \\(3 \\times\\) less time for point registration step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Point Patch Loss. In the semantic point registration, we employ a point patch loss to reconstruct the local patch surrounding each semantic point. If we omit the point patch loss and opt to directly reconstruct the entire video, we observe that the source structure leaks into the target swapped results and thus produce artifacts, as depicted in Fig. 8(b)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Semantic-Enhanced Schedule. Our goal is to employ se" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.497 + ], + "angle": 0, + "content": "mantic points to transfer motion trajectories, acting as a linkage between the source and target subjects. Therefore, we aim for the semantic points to emphasize high-level semantic alignment without transferring low-level details. This objective is achieved by registering points only on the early sampling steps, i.e., \\( T_{\\min} = \\frac{T}{2} \\) in Eq. (2). As shown in Fig. 8(c), this technique prevents the model from learning excessive low-level details and enhances point alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.893, + 0.633 + ], + "angle": 0, + "content": "Drag-based Point Control. As illustrated in Fig. 8(d), our goal is to swap the black swan to the duck. If we directly use the source point sequence as guidance, the duck's neck conforms to the shape of the black swan, resulting in an inferior identity. However, by employing the proposed point displacement propagation, we can drag the semantic point at the keyframe, ensuring a consistent motion trajectory after dragging. Utilizing the dragged point trajectory as motion guidance results in the correct identity of the duck." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.619, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.838 + ], + "angle": 0, + "content": "This paper uncovers the potential of semantic point correspondence in aligning motion trajectories and altering the subject's identity in video editing. From there, we present VideoSwap, a framework that minimizes human intervention while utilizing semantic point correspondences for customized video subject swapping. Through user-point interactions like point removal or dragging, we address various semantic point correspondence. VideoSwap facilitates shape changes in the target swap while aligning the motion trajectory with the source subject, demonstrating state-of-the-art results in customized video subject swapping." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7628" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.299 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[5] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. Videdit: Zero-shot and spatially aware text-driven video editing. arXiv preprint arXiv:2306.08707, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.501 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292, 2023. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[15] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[16] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[17] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[19] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[21] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14317-14326, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[23] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7629" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[29] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.47, + 0.364 + ], + "angle": 0, + "content": "[31] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.47, + 0.494 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[35] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.47, + 0.668 + ], + "angle": 0, + "content": "[36] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.47, + 0.739 + ], + "angle": 0, + "content": "[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.841 + ], + "angle": 0, + "content": "[39] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. \\(p+\\): Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[43] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[44] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[45] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[46] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[48] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[49] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7630" + } + ] +] \ No newline at end of file diff --git a/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_origin.pdf b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..33ebc6594a0b2767e47f98e2d103991ef29be4fb --- /dev/null +++ b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/f25f0e8f-6c9a-4388-bda9-449fe7becf31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3cd2f44997a7398fa0f4c406f7bb7599e8b419a8dd6a85909c2559b7234a387 +size 3951488 diff --git a/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/full.md b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65a23de9b2e71de994734a5473c5d666dd31d103 --- /dev/null +++ b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/full.md @@ -0,0 +1,344 @@ +# VideoSwap: Customized Video Subject Swapping with Interactive Semantic Point Correspondence + +Yuchao Gu $^{1,2}$ , Yipin Zhou $^{2}$ , Bichen Wu $^{2}$ , Licheng Yu $^{2}$ , Jia-Wei Liu $^{1}$ , Rui Zhao $^{1}$ , Jay Zhangjie Wu $^{1}$ , David Junhao Zhang $^{1}$ , Mike Zheng Shou $^{1*}$ , Kevin Tang $^{2}$ + +1Show Lab, National University of Singapore 2GenAI, Meta + +https://videoswap.github.io/ + +![](images/27055b80c9a28a4f6cdc776294013d9cc4ea1b701cdd58b23ac3b51726cf07c9.jpg) +One-to-One Point Correspondence +(b) Results of Various Correspondence Used in Previous Video Editing +Figure 1. Customized video subject swapping results with VideoSwap. VideoSwap supports shape change in the swapped results while aligning with the source motion trajectory. The swapped target can be either a predefined concept from a pretrained model (e.g., helicopter) or a customized concept (denoted by $V^{*}$ ). Previous methods based on implicit motion encoding and dense correspondence do not perform well in subject swapping with shape changes. We encourage readers to click and play the video clips in this figure using Adobe Acrobat. + +# Abstract + +Current diffusion-based video editing primarily focuses on structure-preserved editing by utilizing various dense correspondences to ensure temporal consistency and motion alignment. However, these approaches are often ineffective when the target edit involves a shape change. To embark on video editing with shape change, we explore customized video subject swapping in this work, where we aim to replace the main subject in a source video with a target subject having a distinct identity and potentially different shape. In contrast to previous methods that rely on dense + +correspondences, we introduce the VideoSwap framework that exploits semantic point correspondences, inspired by our observation that only a small number of semantic points are necessary to align the subject's motion trajectory and modify its shape. We also introduce various user-point interactions (e.g., removing points and dragging points) to address various semantic point correspondence. Extensive experiments demonstrate state-of-the-art video subject swapping results across a variety of real-world videos. + +# 1. Introduction + +Diffusion-based video editing [5, 10, 22, 25, 31, 43, 44, 49] is an emerging field that harnesses the capabilities of pre + +![](images/8c00273f80ad5a159aead8e8f5637d61c155ce2ced2bc23724fec96641fcb14c.jpg) +Figure 2. Customized video subject swapping results of VideoSwap on various concepts. The swapping target can either be a predefined concept in the pretrained model (e.g., helicopter) or a customized concept created by ED-LoRA [13] (denoted as $V^{*}$ ). We encourage readers to click and play the video clips in this figure using Adobe Acrobat. For legal issues, we cannot display the human swap results. + +trained text-to-image/video diffusion models [14, 15, 32, 37] to facilitate a range of video editing tasks, including style change and subject/background swapping. The main challenge in video editing lies in how to extract motion from the source video and transfer it to the edited video while ensuring temporal consistency. Pioneer Tune-A-Video [43] implicitly encodes source motion in the diffusion model weights by tuning from the source video. While it demonstrates versatile applications for video editing, the + +temporal consistency is far from satisfactory. Subsequent works make use of various dense correspondences extracted from the source video, including attention maps [23, 31], edge/depth maps [10, 22, 46, 49], optical flows [46], and deformation fields [5, 6] for video editing. While achieving better temporal consistency, dense correspondence imposes strict shape constraints on the target edit, which makes it ineffective for video editing with shape changes. + +To embark on video editing with shape change, we delve + +into a challenging task: customized video subject swapping. Unlike the conventional video subject swapping addressed in previous works [5, 46, 49], where the swapped subject conforms to the shape of source subject, customized subjects have a clearly defined identity in terms of both appearance and shape. These distinctive characteristics should be preserved in the target edit. Therefore, previous structure-preserved video editing methods are often ineffective for this problem, as shown in Fig. 1(b). + +To address customized video subject swapping, our primary insight is that the subject's motion trajectory can be effectively described using a small number of semantic points. As shown in Fig. 1(a), the motion trajectory of an airplane can be depicted by semantic points located at its wings, nose, and tail. This insight naturally leads us to the following question: Can we utilize these semantic points as correspondences to align the motion trajectory while relaxing the shape constraints in video editing? To answer this question, we conduct a toy experiments and observe that it is possible to learn semantic point correspondence for a specific video subject using just a small number of source video frames. Users can interact with learned semantic points to generate unseen poses or modify the shape of the video subject. These observations suggest the potential for integrating semantic point correspondence into video editing, provided that we can obtain an accurate semantic point sequence for the target edit. + +To unleash the potential of semantic point correspondence, we introduce the VideoSwap framework for customized video subject swapping, which comprises the following primary designs: 1) Integrating the motion layer into the image diffusion model to ensure essential temporal consistency. 2) Registering semantic points on the source video and utilizing them to transfer the motion trajectory of source subject to the target edit. 3) Introducing user-point interactions (e.g., removing or dragging points) for various semantic point correspondence. + +Our contributions are summarized as follows: + +- Empirical observations that reveal the potential of semantic point correspondence for aligning motion trajectories and changing shapes in video editing. +- The VideoSwap framework, which minimizes user intervention while unleashing the potential of semantic point correspondence in customized video subject swapping. +- State-of-the-art results in customized video subject swapping, as demonstrated in Fig. 2. + +# 2. Related Work + +# 2.1. Diffusion-Based Video Editing + +Structure-Preserved Video Editing. FateZero [31] and Video-P2P [23] extract cross- and self-attention from the source video to control spatial layout. To achieve stricter + +alignment of temporal consistency with the source video, Rerender-A-Video [46], Gen-1 [10], ControlVideo [49], and TokenFlow [12] extract and align optical flow, depth/edge maps, and nearest-neighbour field from the source video respectively, resulting in improved temporal consistency. StableVideo [5], VidEdit [6], and CoDEF [28] learn the canonical space for editing following the Layered Neural Atlas [19] or the deformation field in Dynamic Nerf [24, 30]. While achieving promising results in structure-preserved video editing, the above methods based on various dense correspondence are not suitable for handling subject swapping involving shape changes. + +Video Editing with Shape Change. Tune-A-Video (TAV) [43] and FateZero [31] with TAV checkpoint can be utilized for video editing with shape change, as they implicitly encode the motion in model weights through tuning on the source video. However, they suffer from structure and appearance leakage due to model tuning. Shape-aware editing [21] is built on the Layered Neural Atlas [19], employing semantic correspondences to estimate shape deformation and warp the atlas. Nevertheless, texture warping will lead to unrealistic textures. + +Video Diffusion Models. Previous video editing primarily relies on the text-to-image diffusion model [32]. However, recent advancements are occurring in video foundation models [3, 14, 41, 48]. In this work, we add a motion layer [14] to the image diffusion model to provide essential temporal consistency for video editing, and we focus on exploiting semantic point correspondence to align the motion trajectory of the video subject. + +# 2.2. Point Correspondence + +Point Correspondence in Diffusion Models. DIFT [39] initially uncovers robust semantic point correspondences in diffusion models. Building upon the observation in DIFT, subsequent works, DragDiffusion [36] and DragonDiffusion [26] extend DragGAN [29] to support interactive point-based image editing in diffusion models. However, point correspondences and interactive drag-based editing are seldom investigated for video editing. + +Tracking Any Point in Video. TAP-Vid [8] first introduces the problem of tracking any point (TAP). Unlike optical flow estimation, TAP requires the establishment of long-range correspondence for all points in a video. In our work, we use TAP to reduce human intervention in annotating subject keypoints and acquire long-range motion estimation. Although several works [8, 9, 18, 42] address the TAP problem, we choose to employ Co-Tracker [18], which is the most efficient solution available. + +# 2.3. Concept Customization + +Concept customization is mainly categorized into tuning-based approaches [11, 13, 20, 33, 40] and tuning-free solu + +![](images/9799fea7388a2a67a05c446022f2c8e5e87835a2e45ae85b080ec617b34b952d.jpg) + +![](images/2b76bdf232898b4929a1ddaa69f63f3086d127d293af831685beb9ae092d7818.jpg) + +![](images/ee7814d547abc27da13f7c4bf9d37ad48e4867287555019ac71695bab5244196.jpg) +Figure 3. Toy experiment exploring semantic point correspondence. We encourage readers to click and play the video clips in this figure using Adobe Acrobat. + +![](images/cba686cda9108fd02c0a052ba52acb5397c9511a3d670c2fbe0c5671faf9768f.jpg) + +tions [17, 34, 35, 45, 47]. Tuning-free solutions are fast but typically adhere closely to the provided reference image and lack variation. On the other hand, tuning-based solutions can leverage multi-view images to ensure variation in the given concepts and maintain the same inference behavior to pretrained diffusion models. In this paper, we employ the tuning-based ED-LoRA [13] for encoding subject identity. + +# 3. VideoSwap + +In this section, we start by presenting a toy experiment to illustrate our motivation to explore semantic point correspondence in Sec. 3.1. Subsequently, we offer an overview of the VideoSwap pipeline for customized video subject swapping in Sec. 3.2. Following that, we explain the process of injecting semantic point correspondence in Sec. 3.3 to Sec. 3.5. + +# 3.1. Motivation + +Dense correspondences explored in previous video editing methods [5, 10, 12, 28, 46] restrict the subject's shape change in the edited video. Therefore, our goal is to find a more flexible correspondence that can transfer the source subject's motion trajectory without imposing strict shape constraints. Motivated by this, we investigate sparse semantic points as correspondences. Unlike dense correspondences such as depth, edge, and optical flow, which are low-level cues shared across all video subjects, semantic points vary with different open-world concepts. Therefore, it is not + +![](images/f20b35764241a52a40230eb245fbb36ed8843baaeb09b6d9eda963d64235adc8.jpg) +Figure 4. Overview of the VideoSwap pipeline for customized video subject swapping. + +feasible to train a general condition model for injecting semantic point correspondence. Instead, the question we aim to address in this section is, is it possible to learn semantic point control for a specific source video subject using only a small number of source video frames? + +Toy Experiment Setting. To address the above question, we perform a toy experiment. Firstly, for a given video, we manually define a set of semantic points. Next, we annotate the same set of points on eight frames of this video. Finally, we train a T2I-Adapter [27] on these data pairs, as illustrated in Fig. 3(a, b), to determine whether these semantic points can be used to control the source video subject. + +Observation 1: Semantic points optimized on source video frames have the potential to align the subject's motion trajectory and change the subject's shape. As shown in Fig. 3(c, left), we drag the points on the cat's face. Despite this edited point map is not in the training data, the resulting image closely follows the adjusted semantic points, effectively generating the unseen pose of the subject. As shown in Fig. 3(c, right), when we drag the boundary semantic points of the car, the edited subject will also reshape to align with the semantic point. This suggests the potential of utilizing semantic points to align the motion trajectory or change the shape when a sequence of dragged points for all video frames is accessible. + +Observation 2: Semantic points optimized on source video frames can transfer across semantic and low-level changes. As demonstrated in Fig. 3(d), when we replace the source subject with different semantics or modify the low-level information in the prompts, the optimized semantic point can also control the pose or shape of the target concept. This suggests that the semantic point can be transferred across both semantic and low-level changes. + +# 3.2. Overview of VideoSwap + +# 3.2.1 Task Formulation + +In this paper, we focus on customized video subject swapping, with the goal of subject replacement and background + +![](images/3fc1e12ad5ab97690038e32acda1d2f6506bf862cd97cda5fd3c72db5d4113be.jpg) + +![](images/89e0878ed99e66531a16a72d8ec4c1c65df6ba329b75895dea476e7365f22ebe.jpg) + +![](images/908dab3b6c39ed46ddcd6d6e15772104e4bb7ff0f0bc8d4d8813031705af5ff8.jpg) +Figure 5. Pipelines for semantic point extraction (Sec. 3.3) and semantic point registration (Sec. 3.4) in VideoSwap. In semantic point extraction, users define semantic points at a keyframe. We then extract the trajectory and embedding of those semantic points from the video. In semantic point registration, the semantic point embedding is projected by multiple 2-layer learnable MLPs, placed in empty features based on their coordinates, and then added element-wise to the diffusion model as motion guidance. + +![](images/0915901e804140fd8ecd5901e011a5b0fc9d40795ac510a902f7920173fc7b13.jpg) + +preservation. Subject replacement requires preserving the identity of the target subject in the swapped results, encompassing both its appearance and shape. Simultaneously, background preservation requires the unedited background area to remain the same with the source video. The primary challenge of this task lies in aligning the motion trajectory of the source subject while preserving the identity of the target concept, particularly its shape. + +# 3.2.2 Overall Pipeline + +The VideoSwap pipeline is illustrated in Fig. 4. Following the latent diffusion model [32], we encode the source video with a VAE encoder to obtain the latent space representation $z_0$ . Subsequently, DDIM inversion [7, 38] is applied to transform the clean latent $z_0$ back to the noisy latent $z_T$ . After obtaining the DDIM inverted noise $z_T$ , we replace the source subject in the text prompt with the target subject and denoise it using the DDIM scheduler [38]. In this denoising process, we introduce semantic point correspondence to guide the subject's motion trajectory, as detailed in Sec. 3.3-Sec. 3.5. To preserve the unedited background, we leverage the concept of latent blending [1, 2], further explained in the supplementary material. Additionally, we incorporate the following designs: + +Adopting Motion Layers. We integrate the motion layers [3, 14] into the image diffusion model to ensure essential temporal consistency for video editing. + +Supporting Predefined and Customized Concepts. We support both predefined concepts from the pretrained model and customized concepts. To create customized concepts, we train ED-LoRA [13] on a set of representative images to encode their identity. After training, these concept ED-LoRAs can be used at the inference time. + +# 3.3. Semantic Point Extraction + +We first extract the trajectories of semantic points and their associated semantic embeddings from the source video. + +Point Trajectory Extraction. As depicted in Fig. 5, for a video containing $N$ frames, users specify $K$ semantic points at a keyframe $i$ . These user-defined semantic points are then propagated to the remaining $N - 1$ frames using a point tracker [18] or detector [4]. Subsequently, the motion trajectory of all semantic points in the entire video is obtained and represented as $\mathbf{P}_{coord} = \{Tra(k)|k = 1\dots K\}$ , where $Tra(k) = \{(x_n^k,y_n^k,n)|n = 1\dots N\}$ represents the motion trajectory of semantic point $k$ across all $N$ frames. + +Point Embedding Extraction. To leverage semantic point correspondence, it is crucial to associate each point with its semantics. Specifically, we extract the DIFT [39] feature $\mathbf{D}_n$ for each frame $n$ . Subsequently, the point embedding for semantic point $k$ at frame $n$ is obtained as $\mathbf{v}_n^k = \mathbf{D}_n(x_n^k, y_n^k)$ , where $(x_n^k, y_n^k)$ is retrieved from $\mathbf{P}_{coord}$ . Following this, we aggregate the point embeddings acquired from all $N$ frames to obtain the final embedding for each semantic point $k$ by $\mathbf{v}^k = \frac{1}{N} \sum_{n=1}^{N} \mathbf{v}_n^k$ . Finally, we obtain the semantic embedding for all semantic points, succinctly represented as $\mathbf{P}_{emb} = \{\mathbf{v}^k | k = 1 \dots K\}$ . + +# 3.4. Semantic Point Registration + +After acquiring the motion trajectory $\mathbf{P}_{\text {coord }}$ and the embedding $\mathbf{P}_{emb}$ for semantic points, we register these semantic points on the source video to enable them to offer motion guidance for the video subject. + +Sparse Motion Feature Creation. To utilize semantic points as guidance, we generate sparse motion features infused with semantic point embeddings, making them compatible with the Unet encoder. We denote $\mathbf{F}_{enc} =$ + +![](images/a78a191d8f4d99ed3f4cb95f39aabdeffe36dcacbceac825233386b4e79be1f3.jpg) +Figure 6. Point displacement propagation based on layered neural atlas (LNA) [16, 19]. Once a trained LNA is provided, users can drag a semantic point at the keyframe, and this displacement is consistently propagated to every frame through the canonical space of the LNA. + +![](images/8ad7dfb90deafd4f94ea2cf0cefcd132987354a888bd515eb7009d02cbce633a.jpg) + +$\{\mathbf{F}_{enc}^{1}, \mathbf{F}_{enc}^{2}, \mathbf{F}_{enc}^{3}, \mathbf{F}_{enc}^{4}\}$ as the multi-scale intermediate feature of the Unet encoder. For an input VAE-encoded latent with spatial-temporal size $(H, W, N)$ , the feature size of $\mathbf{F}_{enc}^{l}$ for each Unet stage $l \in [1, 4]$ can be computed as $(H / 2^l, W / 2^l, N)$ . + +As depicted in Fig. 5 and detailed in Algorithm 1, we create a series of multi-scale conditional features $\mathbf{F}_c = \{\mathbf{F}_c^1,\mathbf{F}_c^2,\mathbf{F}_c^3,\mathbf{F}_c^4\}$ . Notably, $\mathbf{F}_c$ shares the same size as $\mathbf{F}_{enc}$ and is initialized with zero vectors. And we introduce a series of learnable MLPs $\phi = \{\phi^l | l \in \{1,2,3,4\}\}$ , each of $\phi^l$ project the point embedding to match the feature dimension of the corresponding $\mathbf{F}_c^l$ . Then, for each point $(x_n^k,y_n^k,n) \in \mathbf{P}_{coord}$ , we compute its corresponding spatial-temporal coordinate $(x,y,n)$ at $l$ -th Unet stage and assign the projected embedding based on the coordinate by $\mathbf{F}_c^l(x,y,n) = \phi^l(\mathbf{P}_{emb}(k))$ . + +It is crucial to emphasize that the motion feature $\mathbf{F}_c$ demonstrates high sparsity, with only the semantic point trajectories containing the feature embeddings. Finally, $\mathbf{F}_c$ are added element-wise into the intermediate feature $\mathbf{F}_{enc}$ of Unet encoder as motion guidance: + +$$ +\mathbf {F} _ {e n c} ^ {l} = \mathbf {F} _ {e n c} ^ {l} + \mathbf {F} _ {c} ^ {l}, l \in \{1, 2, 3, 4 \}. \tag {1} +$$ + +# Algorithm 1: Sparse Motion Feature Creation + +1 Input: +2 1. Point Trajectory $\mathbf{P}_{\text {coord }}$ +3 2. Point Embedding $\mathbf{P}_{emb}$ +4 3. Learnable MLPs $\phi = \{\phi^l | l \in \{1,2,3,4\}\}$ +5 Output: +6 Motion Feature $\mathbf{F}_c = \{\mathbf{F}_c^l | l \in \{1,2,3,4\}\}$ +7 Initialize: $\mathbf{F}_c = \{\mathbf{F}_c^l = \mathbf{0} | l \in \{1,2,3,4\}\}$ ; +8 for $l \gets 1$ to 4 do +9 foreach $(x_n^k, y_n^k, n)$ in $\mathbf{P}_{\text {coord }}$ do +10 $x, y = \text{round}(x_n^k / 2^l), \text{round}(y_n^k / 2^l)$ ; +11 $\mathbf{F}_c^l(x, y, n) = \phi^l(\mathbf{P}_{emb}(k))$ +12 end +13 end + +Semantic Point Registration on Source Video. Our objective is to optimize the projection MLPs $(\phi)$ to facilitate better motion guidance from semantic points. This opti + +mization objective is defined as + +$$ +\min _ {\phi} E _ {\epsilon \sim N (0, I), t \sim U \left(T _ {m i n}, T\right)} | | [ \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, p, \phi (\mathbf {P} _ {e m b})\right) ] \odot \Omega (\mathbf {P} _ {c o o r d}) | | _ {2} ^ {2}, \tag {2} +$$ + +where $p$ represents the embedding for the text prompt, and $\Omega(\mathbf{P}_{\text {coord }})$ denotes the binary mask that only turns on around the semantic point. + +We adopt two techniques to enhance the learning of semantic point correspondences in Eq. (2). The first technique is semantic-enhanced schedule, controlled by $T_{min}$ . We set $T_{min} = T / 2$ to enhance the learning at the higher timestep, which prevents overfitting to low-level details and improves semantic point alignment. The second technique, point patch loss, constrains the computation of the loss to a local patch near each semantic point, which reduces structure leakage into the target swap. This is implemented by the a loss mask $\Omega$ in Eq. (2). + +# 3.5. User-Point Interaction at Inference Time + +In Sec. 3.3, users define semantic points at the source video keyframe, and we subsequently extract their trajectory and semantic embedding in the video. To utilize these semantic points as correspondences, we register them on the source video in Sec. 3.4. Following these two steps, these semantic points become applicable for controlling the motion of the target object. In this section, we introduce user-point interaction to address various semantic point correspondence. + +Adopting Source Point Sequence. If there exist one-to-one semantic point correspondence between the source subject and the target subject, such as swapping the dog with $V^{catA}$ as illustrated in Fig. 1, we directly use the source point sequence as the motion guidance. + +Removing Parts of Semantic Point. When there only exists partial semantic point correspondence between the source subject and the target subject, we can remove redundant points to loosen shape constraints. For example, in scenarios such as swapping an airplane for a helicopter, as depicted in Fig. 1, we remove semantic points associated with the airplane's wings, given that helicopters typically lack wings. The remaining semantic points on the airplane's nose and tail are retained as motion guidance. + +![](images/8501526b5dbd604dac0b1f2f84c124cc3e784e7c4523d1a1cf1e17f333d03137.jpg) +Concept Images For V/ + +![](images/b715c40586cba0e742b131356cb39c5d6ddc8cbc96265bfae0f17eda5eb2aa00.jpg) +Source Video + +![](images/5e113eacf247babb83253dd1c6a5ab044e80d43fdc5a139523d448e547a667d7.jpg) +VideoSwap (Ours) + +![](images/b964b2528271b7131fbaaa44b8ecb0c3a36873374ea62d00f32bb09053ed8915.jpg) +DDIM + +![](images/6b27df3855c97a54682f5bfdfd4e8c1f49c6c757ca0b8170dc11a70686d65b78.jpg) +DDIM + Tune-A-Video + +![](images/12f0d99f86b55332f96fa51a9961933acf4a9ee2733d1b459bb07a5c1cf0868d.jpg) +DDIM + T2I-Adapter + +![](images/3837e7d0b2a2a594fb6b7de3a57316329beeabb12912e8e0c2906c217366bd11.jpg) +Concept Images For VocA + +![](images/0922f350a9d82b3199ee8cf1fdc6f909777f6a7188d648aeac7d87216f68d369.jpg) +Source Prompt: A dog sitting on the side of a car window looking out the window. +Source Video + +![](images/a5af481676066f9c7fbc9e6dfb165c6ab357633ad933196dfa6c9a926d5a6fa6.jpg) +VideoSwap (Ours) +Figure 7. Comparison of VideoSwap with several baselines built upon the same foundational model. The only difference lies in adopting different motion guidance. We encourage readers to click and play the video clips in this figure using Adobe Acrobat. + +![](images/7375fff4493371f23f8f45765ad77852f8bd8d61dc14ddec8b4e165af0834aaa.jpg) +DDIM + +![](images/e7c4e97af5bbcfc2b9f74aed6191ed66db33a4844108e12f35c46fd7bfc31858.jpg) +Target Edit: dog -> VcatA +DDIM + Tune-A-Video + +![](images/3763ffe3cf74c12835c7b83482679e7cf69bf3d8d1d25bcc3e45470b31c86194.jpg) +DDIM + T2I-Adapter + +**Dragging of Semantic Point.** In situations where there exists semantic point correspondence between the source and target subjects, but misalignment occurs due to shape morphing, we provide users the option to manually drag the semantic points on a keyframe for better alignment of the shape changes. For instance, when swapping a jeep (tall and narrow) for a $V_{carA}$ (sports car, low and wide) as illustrated in Fig. 1, users can drag the semantic points to accurately reflect the shape change. + +Editing the shape change by dragging semantic points on a single frame is straightforward. However, consistently propagating these semantic point displacement over time is non-trivial, mainly because of the complex camera and subject motion in the video. Therefore, we introduce point displacement propagation to solve this problem. + +As illustrated in Fig. 6, we follow the Layered Neural Atlas (LNA) [16, 19] to learn the canonical space, as detailed in the supplementary material. Following LNA [16, 19], we establish a forward coordinate mapping from the video to the canonical space, denoted as $M \colon (x,y,f) \to (u,v)$ , along with its corresponding backward mapping $B \colon (u,v,f) \to (x,y)$ . + +Given a semantic point, with coordinates at the keyframe $f_{key}$ represented as $(x,y,f_{key})$ , its trajectory over time can be expressed as a function of time $f\colon (x(f),y(f)) = P(f)$ . Suppose a user drag it to a new position at $(x + dx,y + dy,f_{key})$ , we aim to estimate the edited trajectory $P^{\prime}(f)$ for $f = \{0,\dots,N\}$ . We resort to LNA's representation, and first compute a linearized estimation of its shifted position on the canonical coordinate: + +$$ +[ d u, d v ] ^ {T} = J _ {M} (x, y, f _ {\text {k e y}}) [ d x, d y ] ^ {T}, \tag {3} +$$ + +where $J_{M}$ denote the Jacobian matrix with respect to $(x,y)$ . Next, at a given time $f$ , we estimate the edited coordinate in the pixel space as + +$$ +P ^ {\prime} (f) = P (f) + J _ {B} (u, v, f) [ d u, d v ] ^ {T}, \tag {4} +$$ + +where $(u,v) = B(x,y,f_{key})$ and $J_B$ denote the Jacobian matrix with respect to $(u,v)$ . We then use this edited trajectory $P^{\prime}(f)$ for the dragged semantic point during inference. + +
Methods/MetricsSubject IdentityMotion AlignmentTemporal ConsistencyOverall Preference
VideoSwap v.s.Compare to Previous Video Editing Methods
Tune-A-Video80% v.s. 20%72% v.s. 28%80% v.s. 20%78% v.s. 22%
FateZero74% v.s. 26%67% v.s. 33%73% v.s. 27%70% v.s. 30%
Text2Video-Zero76% v.s. 24%71% v.s. 29%80% v.s. 20%80% v.s. 20%
Rerender-A-Video81% v.s. 19%72% v.s. 28%84% v.s. 16%84% v.s. 16%
Compare to Baselines onAnimateDiff
w/ DDIM70% v.s. 30%74% v.s. 26%69% v.s. 31%73% v.s. 27%
w/ DDIM+TAV68% v.s. 32%60% v.s. 40%66% v.s. 34%69% v.s. 31%
w/ DDIM+T2I-Adapter66% v.s. 34%59% v.s. 41%65% v.s. 35%66% v.s. 34%
+ +Table 1. Human Evaluation on Video Subject Swapping Results. + +# 4. Experiments + +We implement our method using the Latent Diffusion Model [32] and adopt the motion layer inAnimateDiff [14] as the foundational model. All videos consist of 16 frames. Additional implementation details are included in the supplementary material. + +# 4.1. Qualitative Comparison + +Comparison with the State-of-the-Art. We qualitatively compare to Tune-A-Video [43], FateZero [31], Rerender-A-Video [46], TokenFlow [12] and StableVideo [5] in Fig. 1. Previous methods are less effective in revealing the correct shape of the target subject. Compared to them, VideoSwap can achieve a significant shape change while aligning the source motion trajectory. + +Comparison with Baselines onAnimateDiff. As most state-of-the-art methods are based on image diffusion models, we also compare VideoSwap to several baselines on theAnimateDiff. The only distinctions from VideoSwap lie in different motion guidance, as shown in the results in Fig. 7: + +- DDIM: The DDIM sampling without other motion guidance cannot produce the correct motion trajectory. +- DDIM+Tune-A-Video: If we tune the model as [43] to inject source motion, it achieves correct motion but suffers from severe structure and appearance leakage. +- DDIM+T2I-Adapter: If we add spatial controls [27], such as depth, to control the editing, we observe that 1) the shape is restricted by the source, and 2) the deformable motion cannot follow the source video. + +Compared to all constructed baselines, our VideoSwap with semantic point correspondence can effectively align the motion trajectory while preserving the target concept's identity. + +![](images/35efebd215b608ff1d1cbf58949505a61f7939ad704f7a8f27204289c90e3c13.jpg) +Figure 8. Ablation Study of our VideoSwap. We encourage readers to click and play the video clips in this figure using Adobe Acrobat. + +# 4.2. Quantitative Comparison + +We conduct both automatic and human evaluations to quantitatively compare VideoSwap with previous state-of-the-art methods and several baselines onAnimateDiff. Detailed evaluation settings and automatic evaluation results are provided in the supplementary material. For human evaluation, we distribute 1000 questionnaires on Amazon Mturk to assess various criteria in customized video subject swapping. From the human evaluation results in Table. 1, our method achieves a clear advantage over the compared methods. + +# 4.3. Ablation Study + +Sparse Motion Feature. There are several variants for encoding semantic points to generate motion guidance. The most straightforward approach is to encode the point map of semantic points using T2I-Adapter [27]. However, this method leads to severe overfitting, as illustrated in Fig. 8(a). The issue arises because the encoded feature added to the diffusion model is non-sparse, and the background also contains features that may overfit to the source video. We use MLPs to encode point embeddings, positioning them in the empty feature to create sparse motion features for guidance without compromising video quality. Regarding point embeddings, we opt for DIFT embeddings, which inherently carry robust semantics. Compared to randomly initialized learnable embeddings, our approach achieves similar results with $3 \times$ less time for point registration step. + +Point Patch Loss. In the semantic point registration, we employ a point patch loss to reconstruct the local patch surrounding each semantic point. If we omit the point patch loss and opt to directly reconstruct the entire video, we observe that the source structure leaks into the target swapped results and thus produce artifacts, as depicted in Fig. 8(b). + +Semantic-Enhanced Schedule. Our goal is to employ se + +mantic points to transfer motion trajectories, acting as a linkage between the source and target subjects. Therefore, we aim for the semantic points to emphasize high-level semantic alignment without transferring low-level details. This objective is achieved by registering points only on the early sampling steps, i.e., $T_{\min} = \frac{T}{2}$ in Eq. (2). As shown in Fig. 8(c), this technique prevents the model from learning excessive low-level details and enhances point alignment. + +Drag-based Point Control. As illustrated in Fig. 8(d), our goal is to swap the black swan to the duck. If we directly use the source point sequence as guidance, the duck's neck conforms to the shape of the black swan, resulting in an inferior identity. However, by employing the proposed point displacement propagation, we can drag the semantic point at the keyframe, ensuring a consistent motion trajectory after dragging. Utilizing the dragged point trajectory as motion guidance results in the correct identity of the duck. + +# 5. Conclusion + +This paper uncovers the potential of semantic point correspondence in aligning motion trajectories and altering the subject's identity in video editing. From there, we present VideoSwap, a framework that minimizes human intervention while utilizing semantic point correspondences for customized video subject swapping. Through user-point interactions like point removal or dragging, we address various semantic point correspondence. VideoSwap facilitates shape changes in the target swap while aligning the motion trajectory with the source subject, demonstrating state-of-the-art results in customized video subject swapping. + +Acknowledgement This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), 2023. 5 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3, 5 +[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 5 +[5] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 1, 2, 3, 4, 7 +[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. Videdit: Zero-shot and spatially aware text-driven video editing. arXiv preprint arXiv:2306.08707, 2023. 2, 3 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5 +[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 3 +[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 3 +[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2, 3, 4 +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3 +[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3, 4, 7 +[13] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292, 2023. 2, 3, 4, 5 + +[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3, 5, 7 +[15] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[16] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 6, 7 +[17] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 4 +[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 3, 5 +[19] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3, 6, 7 +[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3 +[21] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14317-14326, 2023. 3 +[22] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 1, 2 +[23] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 2, 3 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1 +[26] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 3 +[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4, 7, 8 + +[28] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3, 4 +[29] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3 +[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3 +[31] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 1, 2, 3, 7 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 7 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3 +[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 4 +[35] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 4 +[36] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3 +[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[39] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023. 3, 5 +[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $p+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 + +[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3 +[42] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3 +[43] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1, 2, 3, 7 +[44] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 1 +[45] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 4 +[46] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 2, 3, 4, 7 +[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4 +[48] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 3 +[49] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023. 1, 2, 3 \ No newline at end of file diff --git a/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/images.zip b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94f58b98f922c3203c5c8bcfb2512563b916d811 --- /dev/null +++ b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eb44dfa690dbc9a212074446a1344344c844e04cfe10aaf16c51cd3d3b88ff6 +size 915298 diff --git a/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/layout.json b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..246898ffe7662cd0e3e40a3a70b2c3a987d6fc72 --- /dev/null +++ b/2024/VideoSwap_ Customized Video Subject Swapping with Interactive Semantic Point Correspondence/layout.json @@ -0,0 +1,9348 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 83, + 465, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 83, + 465, + 118 + ], + "spans": [ + { + "bbox": [ + 128, + 83, + 465, + 118 + ], + "type": "text", + "content": "VideoSwap: Customized Video Subject Swapping with Interactive Semantic Point Correspondence" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": "Yuchao Gu" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Yipin Zhou" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Bichen Wu" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Licheng Yu" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Jia-Wei Liu" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Rui Zhao" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Jay Zhangjie Wu" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", David Junhao Zhang" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Mike Zheng Shou" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "text", + "content": ", Kevin Tang" + }, + { + "bbox": [ + 53, + 140, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 182, + 456, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 182, + 456, + 198 + ], + "spans": [ + { + "bbox": [ + 155, + 182, + 456, + 198 + ], + "type": "text", + "content": "1Show Lab, National University of Singapore 2GenAI, Meta" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 202, + 198, + 405, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 198, + 405, + 211 + ], + "spans": [ + { + "bbox": [ + 202, + 198, + 405, + 211 + ], + "type": "text", + "content": "https://videoswap.github.io/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 77, + 232, + 519, + 459 + ], + "blocks": [ + { + "bbox": [ + 94, + 220, + 207, + 230 + ], + "lines": [ + { + "bbox": [ + 94, + 220, + 207, + 230 + ], + "spans": [ + { + "bbox": [ + 94, + 220, + 207, + 230 + ], + "type": "text", + "content": "One-to-One Point Correspondence" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 232, + 519, + 459 + ], + "lines": [ + { + "bbox": [ + 77, + 232, + 519, + 459 + ], + "spans": [ + { + "bbox": [ + 77, + 232, + 519, + 459 + ], + "type": "image", + "image_path": "27055b80c9a28a4f6cdc776294013d9cc4ea1b701cdd58b23ac3b51726cf07c9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 460, + 412, + 470 + ], + "lines": [ + { + "bbox": [ + 187, + 460, + 412, + 470 + ], + "spans": [ + { + "bbox": [ + 187, + 460, + 412, + 470 + ], + "type": "text", + "content": "(b) Results of Various Correspondence Used in Previous Video Editing" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 472, + 546, + 517 + ], + "lines": [ + { + "bbox": [ + 45, + 472, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 45, + 472, + 546, + 517 + ], + "type": "text", + "content": "Figure 1. Customized video subject swapping results with VideoSwap. VideoSwap supports shape change in the swapped results while aligning with the source motion trajectory. The swapped target can be either a predefined concept from a pretrained model (e.g., helicopter) or a customized concept (denoted by " + }, + { + "bbox": [ + 45, + 472, + 546, + 517 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 45, + 472, + 546, + 517 + ], + "type": "text", + "content": "). Previous methods based on implicit motion encoding and dense correspondence do not perform well in subject swapping with shape changes. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 543, + 192, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 543, + 192, + 556 + ], + "spans": [ + { + "bbox": [ + 143, + 543, + 192, + 556 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 570, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 689 + ], + "type": "text", + "content": "Current diffusion-based video editing primarily focuses on structure-preserved editing by utilizing various dense correspondences to ensure temporal consistency and motion alignment. However, these approaches are often ineffective when the target edit involves a shape change. To embark on video editing with shape change, we explore customized video subject swapping in this work, where we aim to replace the main subject in a source video with a target subject having a distinct identity and potentially different shape. In contrast to previous methods that rely on dense" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 533, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 547, + 641 + ], + "type": "text", + "content": "correspondences, we introduce the VideoSwap framework that exploits semantic point correspondences, inspired by our observation that only a small number of semantic points are necessary to align the subject's motion trajectory and modify its shape. We also introduce various user-point interactions (e.g., removing points and dragging points) to address various semantic point correspondence. Extensive experiments demonstrate state-of-the-art video subject swapping results across a variety of real-world videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 668, + 386, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 386, + 680 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 386, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "Diffusion-based video editing [5, 10, 22, 25, 31, 43, 44, 49] is an emerging field that harnesses the capabilities of pre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7621" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 518, + 558 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 518, + 558 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 518, + 558 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 518, + 558 + ], + "type": "image", + "image_path": "8c00273f80ad5a159aead8e8f5637d61c155ce2ced2bc23724fec96641fcb14c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "type": "text", + "content": "Figure 2. Customized video subject swapping results of VideoSwap on various concepts. The swapping target can either be a predefined concept in the pretrained model (e.g., helicopter) or a customized concept created by ED-LoRA [13] (denoted as " + }, + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "type": "text", + "content": "). We encourage readers to click and play the video clips in this figure using Adobe Acrobat. For legal issues, we cannot display the human swap results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "type": "text", + "content": "trained text-to-image/video diffusion models [14, 15, 32, 37] to facilitate a range of video editing tasks, including style change and subject/background swapping. The main challenge in video editing lies in how to extract motion from the source video and transfer it to the edited video while ensuring temporal consistency. Pioneer Tune-A-Video [43] implicitly encodes source motion in the diffusion model weights by tuning from the source video. While it demonstrates versatile applications for video editing, the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 605, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 700 + ], + "type": "text", + "content": "temporal consistency is far from satisfactory. Subsequent works make use of various dense correspondences extracted from the source video, including attention maps [23, 31], edge/depth maps [10, 22, 46, 49], optical flows [46], and deformation fields [5, 6] for video editing. While achieving better temporal consistency, dense correspondence imposes strict shape constraints on the target edit, which makes it ineffective for video editing with shape changes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To embark on video editing with shape change, we delve" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7622" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "into a challenging task: customized video subject swapping. Unlike the conventional video subject swapping addressed in previous works [5, 46, 49], where the swapped subject conforms to the shape of source subject, customized subjects have a clearly defined identity in terms of both appearance and shape. These distinctive characteristics should be preserved in the target edit. Therefore, previous structure-preserved video editing methods are often ineffective for this problem, as shown in Fig. 1(b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 289, + 396 + ], + "type": "text", + "content": "To address customized video subject swapping, our primary insight is that the subject's motion trajectory can be effectively described using a small number of semantic points. As shown in Fig. 1(a), the motion trajectory of an airplane can be depicted by semantic points located at its wings, nose, and tail. This insight naturally leads us to the following question: Can we utilize these semantic points as correspondences to align the motion trajectory while relaxing the shape constraints in video editing? To answer this question, we conduct a toy experiments and observe that it is possible to learn semantic point correspondence for a specific video subject using just a small number of source video frames. Users can interact with learned semantic points to generate unseen poses or modify the shape of the video subject. These observations suggest the potential for integrating semantic point correspondence into video editing, provided that we can obtain an accurate semantic point sequence for the target edit." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 289, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 516 + ], + "type": "text", + "content": "To unleash the potential of semantic point correspondence, we introduce the VideoSwap framework for customized video subject swapping, which comprises the following primary designs: 1) Integrating the motion layer into the image diffusion model to ensure essential temporal consistency. 2) Registering semantic points on the source video and utilizing them to transfer the motion trajectory of source subject to the target edit. 3) Introducing user-point interactions (e.g., removing or dragging points) for various semantic point correspondence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 517, + 244, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 244, + 528 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 244, + 528 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 529, + 287, + 625 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 529, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 566 + ], + "type": "text", + "content": "- Empirical observations that reveal the potential of semantic point correspondence for aligning motion trajectories and changing shapes in video editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 566, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 287, + 601 + ], + "type": "text", + "content": "- The VideoSwap framework, which minimizes user intervention while unleashing the potential of semantic point correspondence in customized video subject swapping." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 287, + 625 + ], + "type": "text", + "content": "- State-of-the-art results in customized video subject swapping, as demonstrated in Fig. 2." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 638, + 134, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 134, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 134, + 650 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 212, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 212, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 212, + 671 + ], + "type": "text", + "content": "2.1. Diffusion-Based Video Editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Structure-Preserved Video Editing. FateZero [31] and Video-P2P [23] extract cross- and self-attention from the source video to control spatial layout. To achieve stricter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "alignment of temporal consistency with the source video, Rerender-A-Video [46], Gen-1 [10], ControlVideo [49], and TokenFlow [12] extract and align optical flow, depth/edge maps, and nearest-neighbour field from the source video respectively, resulting in improved temporal consistency. StableVideo [5], VidEdit [6], and CoDEF [28] learn the canonical space for editing following the Layered Neural Atlas [19] or the deformation field in Dynamic Nerf [24, 30]. While achieving promising results in structure-preserved video editing, the above methods based on various dense correspondence are not suitable for handling subject swapping involving shape changes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "content": "Video Editing with Shape Change. Tune-A-Video (TAV) [43] and FateZero [31] with TAV checkpoint can be utilized for video editing with shape change, as they implicitly encode the motion in model weights through tuning on the source video. However, they suffer from structure and appearance leakage due to model tuning. Shape-aware editing [21] is built on the Layered Neural Atlas [19], employing semantic correspondences to estimate shape deformation and warp the atlas. Nevertheless, texture warping will lead to unrealistic textures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": "Video Diffusion Models. Previous video editing primarily relies on the text-to-image diffusion model [32]. However, recent advancements are occurring in video foundation models [3, 14, 41, 48]. In this work, we add a motion layer [14] to the image diffusion model to provide essential temporal consistency for video editing, and we focus on exploiting semantic point correspondence to align the motion trajectory of the video subject." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 440, + 432, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 432, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 432, + 452 + ], + "type": "text", + "content": "2.2. Point Correspondence" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "content": "Point Correspondence in Diffusion Models. DIFT [39] initially uncovers robust semantic point correspondences in diffusion models. Building upon the observation in DIFT, subsequent works, DragDiffusion [36] and DragonDiffusion [26] extend DragGAN [29] to support interactive point-based image editing in diffusion models. However, point correspondences and interactive drag-based editing are seldom investigated for video editing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 555, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 662 + ], + "type": "text", + "content": "Tracking Any Point in Video. TAP-Vid [8] first introduces the problem of tracking any point (TAP). Unlike optical flow estimation, TAP requires the establishment of long-range correspondence for all points in a video. In our work, we use TAP to reduce human intervention in annotating subject keypoints and acquire long-range motion estimation. Although several works [8, 9, 18, 42] address the TAP problem, we choose to employ Co-Tracker [18], which is the most efficient solution available." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 670, + 438, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 438, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 438, + 682 + ], + "type": "text", + "content": "2.3. Concept Customization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Concept customization is mainly categorized into tuning-based approaches [11, 13, 20, 33, 40] and tuning-free solu" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7623" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 161, + 192 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 161, + 192 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 161, + 192 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 161, + 192 + ], + "type": "image", + "image_path": "9799fea7388a2a67a05c446022f2c8e5e87835a2e45ae85b080ec617b34b952d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 72, + 271, + 191 + ], + "blocks": [ + { + "bbox": [ + 164, + 72, + 271, + 191 + ], + "lines": [ + { + "bbox": [ + 164, + 72, + 271, + 191 + ], + "spans": [ + { + "bbox": [ + 164, + 72, + 271, + 191 + ], + "type": "image", + "image_path": "2b76bdf232898b4929a1ddaa69f63f3086d127d293af831685beb9ae092d7818.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 194, + 165, + 327 + ], + "blocks": [ + { + "bbox": [ + 59, + 194, + 165, + 327 + ], + "lines": [ + { + "bbox": [ + 59, + 194, + 165, + 327 + ], + "spans": [ + { + "bbox": [ + 59, + 194, + 165, + 327 + ], + "type": "image", + "image_path": "ee7814d547abc27da13f7c4bf9d37ad48e4867287555019ac71695bab5244196.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 330, + 287, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 363 + ], + "type": "text", + "content": "Figure 3. Toy experiment exploring semantic point correspondence. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 166, + 194, + 271, + 327 + ], + "blocks": [ + { + "bbox": [ + 166, + 194, + 271, + 327 + ], + "lines": [ + { + "bbox": [ + 166, + 194, + 271, + 327 + ], + "spans": [ + { + "bbox": [ + 166, + 194, + 271, + 327 + ], + "type": "image", + "image_path": "cba686cda9108fd02c0a052ba52acb5397c9511a3d670c2fbe0c5671faf9768f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 454 + ], + "type": "text", + "content": "tions [17, 34, 35, 45, 47]. Tuning-free solutions are fast but typically adhere closely to the provided reference image and lack variation. On the other hand, tuning-based solutions can leverage multi-view images to ensure variation in the given concepts and maintain the same inference behavior to pretrained diffusion models. In this paper, we employ the tuning-based ED-LoRA [13] for encoding subject identity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 119, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 119, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 119, + 490 + ], + "type": "text", + "content": "3. VideoSwap" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 568 + ], + "type": "text", + "content": "In this section, we start by presenting a toy experiment to illustrate our motivation to explore semantic point correspondence in Sec. 3.1. Subsequently, we offer an overview of the VideoSwap pipeline for customized video subject swapping in Sec. 3.2. Following that, we explain the process of injecting semantic point correspondence in Sec. 3.3 to Sec. 3.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 121, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 121, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 121, + 586 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Dense correspondences explored in previous video editing methods [5, 10, 12, 28, 46] restrict the subject's shape change in the edited video. Therefore, our goal is to find a more flexible correspondence that can transfer the source subject's motion trajectory without imposing strict shape constraints. Motivated by this, we investigate sparse semantic points as correspondences. Unlike dense correspondences such as depth, edge, and optical flow, which are low-level cues shared across all video subjects, semantic points vary with different open-world concepts. Therefore, it is not" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 531, + 221 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 531, + 221 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 531, + 221 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 531, + 221 + ], + "type": "image", + "image_path": "f20b35764241a52a40230eb245fbb36ed8843baaeb09b6d9eda963d64235adc8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "type": "text", + "content": "Figure 4. Overview of the VideoSwap pipeline for customized video subject swapping." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 249, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 308 + ], + "type": "text", + "content": "feasible to train a general condition model for injecting semantic point correspondence. Instead, the question we aim to address in this section is, is it possible to learn semantic point control for a specific source video subject using only a small number of source video frames?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": "Toy Experiment Setting. To address the above question, we perform a toy experiment. Firstly, for a given video, we manually define a set of semantic points. Next, we annotate the same set of points on eight frames of this video. Finally, we train a T2I-Adapter [27] on these data pairs, as illustrated in Fig. 3(a, b), to determine whether these semantic points can be used to control the source video subject." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 393, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 548 + ], + "type": "text", + "content": "Observation 1: Semantic points optimized on source video frames have the potential to align the subject's motion trajectory and change the subject's shape. As shown in Fig. 3(c, left), we drag the points on the cat's face. Despite this edited point map is not in the training data, the resulting image closely follows the adjusted semantic points, effectively generating the unseen pose of the subject. As shown in Fig. 3(c, right), when we drag the boundary semantic points of the car, the edited subject will also reshape to align with the semantic point. This suggests the potential of utilizing semantic points to align the motion trajectory or change the shape when a sequence of dragged points for all video frames is accessible." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 548, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 644 + ], + "type": "text", + "content": "Observation 2: Semantic points optimized on source video frames can transfer across semantic and low-level changes. As demonstrated in Fig. 3(d), when we replace the source subject with different semantics or modify the low-level information in the prompts, the optimized semantic point can also control the pose or shape of the target concept. This suggests that the semantic point can be transferred across both semantic and low-level changes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 651, + 440, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 440, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 440, + 664 + ], + "type": "text", + "content": "3.2. Overview of VideoSwap" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 670, + 415, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 415, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 415, + 681 + ], + "type": "text", + "content": "3.2.1 Task Formulation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we focus on customized video subject swapping, with the goal of subject replacement and background" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7624" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 289, + 155 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 289, + 155 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 289, + 155 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 289, + 155 + ], + "type": "image", + "image_path": "3fc1e12ad5ab97690038e32acda1d2f6506bf862cd97cda5fd3c72db5d4113be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 292, + 72, + 518, + 155 + ], + "blocks": [ + { + "bbox": [ + 292, + 72, + 518, + 155 + ], + "lines": [ + { + "bbox": [ + 292, + 72, + 518, + 155 + ], + "spans": [ + { + "bbox": [ + 292, + 72, + 518, + 155 + ], + "type": "image", + "image_path": "89e0878ed99e66531a16a72d8ec4c1c65df6ba329b75895dea476e7365f22ebe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 155, + 231, + 256 + ], + "blocks": [ + { + "bbox": [ + 72, + 155, + 231, + 256 + ], + "lines": [ + { + "bbox": [ + 72, + 155, + 231, + 256 + ], + "spans": [ + { + "bbox": [ + 72, + 155, + 231, + 256 + ], + "type": "image", + "image_path": "908dab3b6c39ed46ddcd6d6e15772104e4bb7ff0f0bc8d4d8813031705af5ff8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 256, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 45, + 256, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 45, + 256, + 545, + 300 + ], + "type": "text", + "content": "Figure 5. Pipelines for semantic point extraction (Sec. 3.3) and semantic point registration (Sec. 3.4) in VideoSwap. In semantic point extraction, users define semantic points at a keyframe. We then extract the trajectory and embedding of those semantic points from the video. In semantic point registration, the semantic point embedding is projected by multiple 2-layer learnable MLPs, placed in empty features based on their coordinates, and then added element-wise to the diffusion model as motion guidance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 156, + 515, + 256 + ], + "blocks": [ + { + "bbox": [ + 235, + 156, + 515, + 256 + ], + "lines": [ + { + "bbox": [ + 235, + 156, + 515, + 256 + ], + "spans": [ + { + "bbox": [ + 235, + 156, + 515, + 256 + ], + "type": "image", + "image_path": "0915901e804140fd8ecd5901e011a5b0fc9d40795ac510a902f7920173fc7b13.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 307, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 307, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 45, + 307, + 287, + 403 + ], + "type": "text", + "content": "preservation. Subject replacement requires preserving the identity of the target subject in the swapped results, encompassing both its appearance and shape. Simultaneously, background preservation requires the unedited background area to remain the same with the source video. The primary challenge of this task lies in aligning the motion trajectory of the source subject while preserving the identity of the target concept, particularly its shape." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 418, + 149, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 149, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 149, + 430 + ], + "type": "text", + "content": "3.2.2 Overall Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "text", + "content": "The VideoSwap pipeline is illustrated in Fig. 4. Following the latent diffusion model [32], we encode the source video with a VAE encoder to obtain the latent space representation " + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "text", + "content": ". Subsequently, DDIM inversion [7, 38] is applied to transform the clean latent " + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "text", + "content": " back to the noisy latent " + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "text", + "content": ". After obtaining the DDIM inverted noise " + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 46, + 437, + 287, + 605 + ], + "type": "text", + "content": ", we replace the source subject in the text prompt with the target subject and denoise it using the DDIM scheduler [38]. In this denoising process, we introduce semantic point correspondence to guide the subject's motion trajectory, as detailed in Sec. 3.3-Sec. 3.5. To preserve the unedited background, we leverage the concept of latent blending [1, 2], further explained in the supplementary material. Additionally, we incorporate the following designs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 641 + ], + "type": "text", + "content": "Adopting Motion Layers. We integrate the motion layers [3, 14] into the image diffusion model to ensure essential temporal consistency for video editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "Supporting Predefined and Customized Concepts. We support both predefined concepts from the pretrained model and customized concepts. To create customized concepts, we train ED-LoRA [13] on a set of representative images to encode their identity. After training, these concept ED-LoRAs can be used at the inference time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 306, + 451, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 451, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 451, + 317 + ], + "type": "text", + "content": "3.3. Semantic Point Extraction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 325, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 348 + ], + "type": "text", + "content": "We first extract the trajectories of semantic points and their associated semantic embeddings from the source video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": "Point Trajectory Extraction. As depicted in Fig. 5, for a video containing " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": " frames, users specify " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": " semantic points at a keyframe " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": ". These user-defined semantic points are then propagated to the remaining " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": " frames using a point tracker [18] or detector [4]. Subsequently, the motion trajectory of all semantic points in the entire video is obtained and represented as " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{coord} = \\{Tra(k)|k = 1\\dots K\\}" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "Tra(k) = \\{(x_n^k,y_n^k,n)|n = 1\\dots N\\}" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": " represents the motion trajectory of semantic point " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": " across all " + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 350, + 545, + 456 + ], + "type": "text", + "content": " frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": "Point Embedding Extraction. To leverage semantic point correspondence, it is crucial to associate each point with its semantics. Specifically, we extract the DIFT [39] feature " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_n" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": ". Subsequently, the point embedding for semantic point " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": " is obtained as " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n^k = \\mathbf{D}_n(x_n^k, y_n^k)" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "(x_n^k, y_n^k)" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": " is retrieved from " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{coord}" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": ". Following this, we aggregate the point embeddings acquired from all " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": " frames to obtain the final embedding for each semantic point " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^k = \\frac{1}{N} \\sum_{n=1}^{N} \\mathbf{v}_n^k" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": ". Finally, we obtain the semantic embedding for all semantic points, succinctly represented as " + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{emb} = \\{\\mathbf{v}^k | k = 1 \\dots K\\}" + }, + { + "bbox": [ + 305, + 458, + 545, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 598, + 459, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 459, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 459, + 611 + ], + "type": "text", + "content": "3.4. Semantic Point Registration" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "type": "text", + "content": "After acquiring the motion trajectory " + }, + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text {coord }}" + }, + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "type": "text", + "content": " and the embedding " + }, + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{emb}" + }, + { + "bbox": [ + 305, + 617, + 545, + 665 + ], + "type": "text", + "content": " for semantic points, we register these semantic points on the source video to enable them to offer motion guidance for the video subject." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": "Sparse Motion Feature Creation. To utilize semantic points as guidance, we generate sparse motion features infused with semantic point embeddings, making them compatible with the Unet encoder. We denote " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{enc} =" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7625" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 312, + 191 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 312, + 191 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 312, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 312, + 191 + ], + "type": "image", + "image_path": "a78a191d8f4d99ed3f4cb95f39aabdeffe36dcacbceac825233386b4e79be1f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 547, + 216 + ], + "type": "text", + "content": "Figure 6. Point displacement propagation based on layered neural atlas (LNA) [16, 19]. Once a trained LNA is provided, users can drag a semantic point at the keyframe, and this displacement is consistently propagated to every frame through the canonical space of the LNA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 530, + 191 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 530, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 530, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 530, + 191 + ], + "type": "image", + "image_path": "8ad7dfb90deafd4f94ea2cf0cefcd132987354a888bd515eb7009d02cbce633a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{enc}^{1}, \\mathbf{F}_{enc}^{2}, \\mathbf{F}_{enc}^{3}, \\mathbf{F}_{enc}^{4}\\}" + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "text", + "content": " as the multi-scale intermediate feature of the Unet encoder. For an input VAE-encoded latent with spatial-temporal size " + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "inline_equation", + "content": "(H, W, N)" + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "text", + "content": ", the feature size of " + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{enc}^{l}" + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "text", + "content": " for each Unet stage " + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "inline_equation", + "content": "l \\in [1, 4]" + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "text", + "content": " can be computed as " + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "inline_equation", + "content": "(H / 2^l, W / 2^l, N)" + }, + { + "bbox": [ + 47, + 219, + 287, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": "As depicted in Fig. 5 and detailed in Algorithm 1, we create a series of multi-scale conditional features " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c = \\{\\mathbf{F}_c^1,\\mathbf{F}_c^2,\\mathbf{F}_c^3,\\mathbf{F}_c^4\\}" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": ". Notably, " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": " shares the same size as " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{enc}" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": " and is initialized with zero vectors. And we introduce a series of learnable MLPs " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\phi^l | l \\in \\{1,2,3,4\\}\\}" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": ", each of " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\phi^l" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": " project the point embedding to match the feature dimension of the corresponding " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c^l" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": ". Then, for each point " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "(x_n^k,y_n^k,n) \\in \\mathbf{P}_{coord}" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": ", we compute its corresponding spatial-temporal coordinate " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "(x,y,n)" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": "-th Unet stage and assign the projected embedding based on the coordinate by " + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c^l(x,y,n) = \\phi^l(\\mathbf{P}_{emb}(k))" + }, + { + "bbox": [ + 47, + 281, + 287, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "text", + "content": "It is crucial to emphasize that the motion feature " + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "text", + "content": " demonstrates high sparsity, with only the semantic point trajectories containing the feature embeddings. Finally, " + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "text", + "content": " are added element-wise into the intermediate feature " + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{enc}" + }, + { + "bbox": [ + 46, + 414, + 287, + 474 + ], + "type": "text", + "content": " of Unet encoder as motion guidance:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 474, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 474, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 94, + 474, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {e n c} ^ {l} = \\mathbf {F} _ {e n c} ^ {l} + \\mathbf {F} _ {c} ^ {l}, l \\in \\{1, 2, 3, 4 \\}. \\tag {1}", + "image_path": "ab199addd310c57accf88060fac957bae7cb5893dcef5c727c3cab9e5a3f62d3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 494, + 242, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 494, + 242, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 494, + 242, + 506 + ], + "type": "text", + "content": "Algorithm 1: Sparse Motion Feature Creation" + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 50, + 509, + 248, + 670 + ], + "blocks": [ + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "lines": [ + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "spans": [ + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": "1 Input: \n2 1. Point Trajectory " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text {coord }}" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " \n3 2. Point Embedding " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{emb}" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " \n4 3. Learnable MLPs " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\phi^l | l \\in \\{1,2,3,4\\}\\}" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " \n5 Output: \n6 Motion Feature " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c = \\{\\mathbf{F}_c^l | l \\in \\{1,2,3,4\\}\\}" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " \n7 Initialize: " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c = \\{\\mathbf{F}_c^l = \\mathbf{0} | l \\in \\{1,2,3,4\\}\\}" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": "; \n8 for " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "l \\gets 1" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " to 4 do \n9 foreach " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "(x_n^k, y_n^k, n)" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text {coord }}" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " do \n10 " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "x, y = \\text{round}(x_n^k / 2^l), \\text{round}(y_n^k / 2^l)" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": "; \n11 " + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c^l(x, y, n) = \\phi^l(\\mathbf{P}_{emb}(k))" + }, + { + "bbox": [ + 50, + 509, + 248, + 670 + ], + "type": "text", + "content": " \n12 end \n13 end" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Semantic Point Registration on Source Video. Our objective is to optimize the projection MLPs " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "(\\phi)" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " to facilitate better motion guidance from semantic points. This opti" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 220, + 435, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 435, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 435, + 232 + ], + "type": "text", + "content": "mization objective is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 237, + 557, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 237, + 557, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 557, + 262 + ], + "type": "interline_equation", + "content": "\\min _ {\\phi} E _ {\\epsilon \\sim N (0, I), t \\sim U \\left(T _ {m i n}, T\\right)} | | [ \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, p, \\phi (\\mathbf {P} _ {e m b})\\right) ] \\odot \\Omega (\\mathbf {P} _ {c o o r d}) | | _ {2} ^ {2}, \\tag {2}", + "image_path": "05644e9d7907ee99a3b5adc5b8ffb7e0569d390dcfcc22e673059f23b2a80b9a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "type": "text", + "content": " represents the embedding for the text prompt, and " + }, + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\Omega(\\mathbf{P}_{\\text {coord }})" + }, + { + "bbox": [ + 304, + 263, + 545, + 299 + ], + "type": "text", + "content": " denotes the binary mask that only turns on around the semantic point." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "text", + "content": "We adopt two techniques to enhance the learning of semantic point correspondences in Eq. (2). The first technique is semantic-enhanced schedule, controlled by " + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "inline_equation", + "content": "T_{min}" + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "inline_equation", + "content": "T_{min} = T / 2" + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "text", + "content": " to enhance the learning at the higher timestep, which prevents overfitting to low-level details and improves semantic point alignment. The second technique, point patch loss, constrains the computation of the loss to a local patch near each semantic point, which reduces structure leakage into the target swap. This is implemented by the a loss mask " + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 300, + 546, + 419 + ], + "type": "text", + "content": " in Eq. (2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 430, + 519, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 519, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 519, + 441 + ], + "type": "text", + "content": "3.5. User-Point Interaction at Inference Time" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "In Sec. 3.3, users define semantic points at the source video keyframe, and we subsequently extract their trajectory and semantic embedding in the video. To utilize these semantic points as correspondences, we register them on the source video in Sec. 3.4. Following these two steps, these semantic points become applicable for controlling the motion of the target object. In this section, we introduce user-point interaction to address various semantic point correspondence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "content": "Adopting Source Point Sequence. If there exist one-to-one semantic point correspondence between the source subject and the target subject, such as swapping the dog with " + }, + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "V^{catA}" + }, + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "content": " as illustrated in Fig. 1, we directly use the source point sequence as the motion guidance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Removing Parts of Semantic Point. When there only exists partial semantic point correspondence between the source subject and the target subject, we can remove redundant points to loosen shape constraints. For example, in scenarios such as swapping an airplane for a helicopter, as depicted in Fig. 1, we remove semantic points associated with the airplane's wings, given that helicopters typically lack wings. The remaining semantic points on the airplane's nose and tail are retained as motion guidance." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7626" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 130, + 126 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 130, + 126 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 130, + 126 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 130, + 126 + ], + "type": "image", + "image_path": "8501526b5dbd604dac0b1f2f84c124cc3e784e7c4523d1a1cf1e17f333d03137.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 127, + 136, + 134 + ], + "lines": [ + { + "bbox": [ + 75, + 127, + 136, + 134 + ], + "spans": [ + { + "bbox": [ + 75, + 127, + 136, + 134 + ], + "type": "text", + "content": "Concept Images For V/" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 83, + 211, + 126 + ], + "blocks": [ + { + "bbox": [ + 137, + 83, + 211, + 126 + ], + "lines": [ + { + "bbox": [ + 137, + 83, + 211, + 126 + ], + "spans": [ + { + "bbox": [ + 137, + 83, + 211, + 126 + ], + "type": "image", + "image_path": "b715c40586cba0e742b131356cb39c5d6ddc8cbc96265bfae0f17eda5eb2aa00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 128, + 191, + 133 + ], + "lines": [ + { + "bbox": [ + 157, + 128, + 191, + 133 + ], + "spans": [ + { + "bbox": [ + 157, + 128, + 191, + 133 + ], + "type": "text", + "content": "Source Video" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 74, + 287, + 126 + ], + "blocks": [ + { + "bbox": [ + 212, + 74, + 287, + 126 + ], + "lines": [ + { + "bbox": [ + 212, + 74, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 212, + 74, + 287, + 126 + ], + "type": "image", + "image_path": "5e113eacf247babb83253dd1c6a5ab044e80d43fdc5a139523d448e547a667d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 128, + 273, + 133 + ], + "lines": [ + { + "bbox": [ + 227, + 128, + 273, + 133 + ], + "spans": [ + { + "bbox": [ + 227, + 128, + 273, + 133 + ], + "type": "text", + "content": "VideoSwap (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 288, + 74, + 362, + 126 + ], + "blocks": [ + { + "bbox": [ + 288, + 74, + 362, + 126 + ], + "lines": [ + { + "bbox": [ + 288, + 74, + 362, + 126 + ], + "spans": [ + { + "bbox": [ + 288, + 74, + 362, + 126 + ], + "type": "image", + "image_path": "b964b2528271b7131fbaaa44b8ecb0c3a36873374ea62d00f32bb09053ed8915.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 128, + 334, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 128, + 334, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 334, + 133 + ], + "type": "text", + "content": "DDIM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 365, + 74, + 436, + 126 + ], + "blocks": [ + { + "bbox": [ + 365, + 74, + 436, + 126 + ], + "lines": [ + { + "bbox": [ + 365, + 74, + 436, + 126 + ], + "spans": [ + { + "bbox": [ + 365, + 74, + 436, + 126 + ], + "type": "image", + "image_path": "6b27df3855c97a54682f5bfdfd4e8c1f49c6c757ca0b8170dc11a70686d65b78.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 128, + 433, + 133 + ], + "lines": [ + { + "bbox": [ + 375, + 128, + 433, + 133 + ], + "spans": [ + { + "bbox": [ + 375, + 128, + 433, + 133 + ], + "type": "text", + "content": "DDIM + Tune-A-Video" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 83, + 514, + 126 + ], + "blocks": [ + { + "bbox": [ + 440, + 83, + 514, + 126 + ], + "lines": [ + { + "bbox": [ + 440, + 83, + 514, + 126 + ], + "spans": [ + { + "bbox": [ + 440, + 83, + 514, + 126 + ], + "type": "image", + "image_path": "12f0d99f86b55332f96fa51a9961933acf4a9ee2733d1b459bb07a5c1cf0868d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 128, + 506, + 134 + ], + "lines": [ + { + "bbox": [ + 452, + 128, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 452, + 128, + 506, + 134 + ], + "type": "text", + "content": "DDIM + T2I-Adapter" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 75, + 137, + 131, + 189 + ], + "blocks": [ + { + "bbox": [ + 75, + 137, + 131, + 189 + ], + "lines": [ + { + "bbox": [ + 75, + 137, + 131, + 189 + ], + "spans": [ + { + "bbox": [ + 75, + 137, + 131, + 189 + ], + "type": "image", + "image_path": "3837e7d0b2a2a594fb6b7de3a57316329beeabb12912e8e0c2906c217366bd11.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 190, + 138, + 196 + ], + "lines": [ + { + "bbox": [ + 74, + 190, + 138, + 196 + ], + "spans": [ + { + "bbox": [ + 74, + 190, + 138, + 196 + ], + "type": "text", + "content": "Concept Images For VocA" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 137, + 144, + 211, + 189 + ], + "blocks": [ + { + "bbox": [ + 200, + 137, + 384, + 144 + ], + "lines": [ + { + "bbox": [ + 200, + 137, + 384, + 144 + ], + "spans": [ + { + "bbox": [ + 200, + 137, + 384, + 144 + ], + "type": "text", + "content": "Source Prompt: A dog sitting on the side of a car window looking out the window." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 144, + 211, + 189 + ], + "lines": [ + { + "bbox": [ + 137, + 144, + 211, + 189 + ], + "spans": [ + { + "bbox": [ + 137, + 144, + 211, + 189 + ], + "type": "image", + "image_path": "0922f350a9d82b3199ee8cf1fdc6f909777f6a7188d648aeac7d87216f68d369.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 190, + 191, + 195 + ], + "lines": [ + { + "bbox": [ + 157, + 190, + 191, + 195 + ], + "spans": [ + { + "bbox": [ + 157, + 190, + 191, + 195 + ], + "type": "text", + "content": "Source Video" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 212, + 144, + 287, + 189 + ], + "blocks": [ + { + "bbox": [ + 212, + 144, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 212, + 144, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 212, + 144, + 287, + 189 + ], + "type": "image", + "image_path": "a5af481676066f9c7fbc9e6dfb165c6ab357633ad933196dfa6c9a926d5a6fa6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 190, + 273, + 196 + ], + "lines": [ + { + "bbox": [ + 227, + 190, + 273, + 196 + ], + "spans": [ + { + "bbox": [ + 227, + 190, + 273, + 196 + ], + "type": "text", + "content": "VideoSwap (Ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 199, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 545, + 220 + ], + "type": "text", + "content": "Figure 7. Comparison of VideoSwap with several baselines built upon the same foundational model. The only difference lies in adopting different motion guidance. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 288, + 145, + 363, + 189 + ], + "blocks": [ + { + "bbox": [ + 288, + 145, + 363, + 189 + ], + "lines": [ + { + "bbox": [ + 288, + 145, + 363, + 189 + ], + "spans": [ + { + "bbox": [ + 288, + 145, + 363, + 189 + ], + "type": "image", + "image_path": "7375fff4493371f23f8f45765ad77852f8bd8d61dc14ddec8b4e165af0834aaa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 190, + 334, + 195 + ], + "lines": [ + { + "bbox": [ + 317, + 190, + 334, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 190, + 334, + 195 + ], + "type": "text", + "content": "DDIM" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 364, + 145, + 439, + 189 + ], + "blocks": [ + { + "bbox": [ + 391, + 137, + 453, + 144 + ], + "lines": [ + { + "bbox": [ + 391, + 137, + 453, + 144 + ], + "spans": [ + { + "bbox": [ + 391, + 137, + 453, + 144 + ], + "type": "text", + "content": "Target Edit: dog -> VcatA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 145, + 439, + 189 + ], + "lines": [ + { + "bbox": [ + 364, + 145, + 439, + 189 + ], + "spans": [ + { + "bbox": [ + 364, + 145, + 439, + 189 + ], + "type": "image", + "image_path": "e7c4e97af5bbcfc2b9f74aed6191ed66db33a4844108e12f35c46fd7bfc31858.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 190, + 433, + 196 + ], + "lines": [ + { + "bbox": [ + 375, + 190, + 433, + 196 + ], + "spans": [ + { + "bbox": [ + 375, + 190, + 433, + 196 + ], + "type": "text", + "content": "DDIM + Tune-A-Video" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 441, + 145, + 515, + 189 + ], + "blocks": [ + { + "bbox": [ + 441, + 145, + 515, + 189 + ], + "lines": [ + { + "bbox": [ + 441, + 145, + 515, + 189 + ], + "spans": [ + { + "bbox": [ + 441, + 145, + 515, + 189 + ], + "type": "image", + "image_path": "3763ffe3cf74c12835c7b83482679e7cf69bf3d8d1d25bcc3e45470b31c86194.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 190, + 506, + 196 + ], + "lines": [ + { + "bbox": [ + 452, + 190, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 452, + 190, + 506, + 196 + ], + "type": "text", + "content": "DDIM + T2I-Adapter" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": "**Dragging of Semantic Point.** In situations where there exists semantic point correspondence between the source and target subjects, but misalignment occurs due to shape morphing, we provide users the option to manually drag the semantic points on a keyframe for better alignment of the shape changes. For instance, when swapping a jeep (tall and narrow) for a " + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "inline_equation", + "content": "V_{carA}" + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": " (sports car, low and wide) as illustrated in Fig. 1, users can drag the semantic points to accurately reflect the shape change." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 344, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 416 + ], + "type": "text", + "content": "Editing the shape change by dragging semantic points on a single frame is straightforward. However, consistently propagating these semantic point displacement over time is non-trivial, mainly because of the complex camera and subject motion in the video. Therefore, we introduce point displacement propagation to solve this problem." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "type": "text", + "content": "As illustrated in Fig. 6, we follow the Layered Neural Atlas (LNA) [16, 19] to learn the canonical space, as detailed in the supplementary material. Following LNA [16, 19], we establish a forward coordinate mapping from the video to the canonical space, denoted as " + }, + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "type": "inline_equation", + "content": "M \\colon (x,y,f) \\to (u,v)" + }, + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "type": "text", + "content": ", along with its corresponding backward mapping " + }, + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "type": "inline_equation", + "content": "B \\colon (u,v,f) \\to (x,y)" + }, + { + "bbox": [ + 47, + 426, + 287, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": "Given a semantic point, with coordinates at the keyframe " + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "inline_equation", + "content": "f_{key}" + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": " represented as " + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "inline_equation", + "content": "(x,y,f_{key})" + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": ", its trajectory over time can be expressed as a function of time " + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "inline_equation", + "content": "f\\colon (x(f),y(f)) = P(f)" + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": ". Suppose a user drag it to a new position at " + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "inline_equation", + "content": "(x + dx,y + dy,f_{key})" + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": ", we aim to estimate the edited trajectory " + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "inline_equation", + "content": "P^{\\prime}(f)" + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "inline_equation", + "content": "f = \\{0,\\dots,N\\}" + }, + { + "bbox": [ + 47, + 518, + 287, + 612 + ], + "type": "text", + "content": ". We resort to LNA's representation, and first compute a linearized estimation of its shifted position on the canonical coordinate:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 90, + 613, + 286, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 613, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 90, + 613, + 286, + 628 + ], + "type": "interline_equation", + "content": "[ d u, d v ] ^ {T} = J _ {M} (x, y, f _ {\\text {k e y}}) [ d x, d y ] ^ {T}, \\tag {3}", + "image_path": "b433fe08fc749f8d05de3d46aa03645489fbf2a88dc94a62c982d97db8a9351c.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "inline_equation", + "content": "J_{M}" + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "text", + "content": " denote the Jacobian matrix with respect to " + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "text", + "content": ". Next, at a given time " + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 628, + 286, + 662 + ], + "type": "text", + "content": ", we estimate the edited coordinate in the pixel space as" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 85, + 663, + 286, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 663, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 85, + 663, + 286, + 677 + ], + "type": "interline_equation", + "content": "P ^ {\\prime} (f) = P (f) + J _ {B} (u, v, f) [ d u, d v ] ^ {T}, \\tag {4}", + "image_path": "1bbd7721013dc8419184464f81e4e375bd518fcff1f2820df460ee0eec09b943.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "(u,v) = B(x,y,f_{key})" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "J_B" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " denote the Jacobian matrix with respect to " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": ". We then use this edited trajectory " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "P^{\\prime}(f)" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " for the dragged semantic point during inference." + } + ] + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 307, + 226, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 307, + 226, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 307, + 226, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 545, + 308 + ], + "type": "table", + "html": "
Methods/MetricsSubject IdentityMotion AlignmentTemporal ConsistencyOverall Preference
VideoSwap v.s.Compare to Previous Video Editing Methods
Tune-A-Video80% v.s. 20%72% v.s. 28%80% v.s. 20%78% v.s. 22%
FateZero74% v.s. 26%67% v.s. 33%73% v.s. 27%70% v.s. 30%
Text2Video-Zero76% v.s. 24%71% v.s. 29%80% v.s. 20%80% v.s. 20%
Rerender-A-Video81% v.s. 19%72% v.s. 28%84% v.s. 16%84% v.s. 16%
Compare to Baselines onAnimateDiff
w/ DDIM70% v.s. 30%74% v.s. 26%69% v.s. 31%73% v.s. 27%
w/ DDIM+TAV68% v.s. 32%60% v.s. 40%66% v.s. 34%69% v.s. 31%
w/ DDIM+T2I-Adapter66% v.s. 34%59% v.s. 41%65% v.s. 35%66% v.s. 34%
", + "image_path": "ef8354adb0354824894ab5a3ce4c7791c24b0a57c7e825d53189256442a54197.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 307, + 308, + 542, + 319 + ], + "lines": [ + { + "bbox": [ + 307, + 308, + 542, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 542, + 319 + ], + "type": "text", + "content": "Table 1. Human Evaluation on Video Subject Swapping Results." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 321, + 386, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 321, + 386, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 386, + 334 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 341, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 400 + ], + "type": "text", + "content": "We implement our method using the Latent Diffusion Model [32] and adopt the motion layer inAnimateDiff [14] as the foundational model. All videos consist of 16 frames. Additional implementation details are included in the supplementary material." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 407, + 441, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 441, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 441, + 420 + ], + "type": "text", + "content": "4.1. Qualitative Comparison" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": "Comparison with the State-of-the-Art. We qualitatively compare to Tune-A-Video [43], FateZero [31], Rerender-A-Video [46], TokenFlow [12] and StableVideo [5] in Fig. 1. Previous methods are less effective in revealing the correct shape of the target subject. Compared to them, VideoSwap can achieve a significant shape change while aligning the source motion trajectory." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 510, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 570 + ], + "type": "text", + "content": "Comparison with Baselines onAnimateDiff. As most state-of-the-art methods are based on image diffusion models, we also compare VideoSwap to several baselines on theAnimateDiff. The only distinctions from VideoSwap lie in different motion guidance, as shown in the results in Fig. 7:" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 570, + 545, + 676 + ], + "type": "list", + "angle": 0, + "index": 45, + "blocks": [ + { + "bbox": [ + 306, + 570, + 544, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 544, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 544, + 594 + ], + "type": "text", + "content": "- DDIM: The DDIM sampling without other motion guidance cannot produce the correct motion trajectory." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 594, + 544, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 544, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 544, + 630 + ], + "type": "text", + "content": "- DDIM+Tune-A-Video: If we tune the model as [43] to inject source motion, it achieves correct motion but suffers from severe structure and appearance leakage." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 630, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 630, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 630, + 545, + 676 + ], + "type": "text", + "content": "- DDIM+T2I-Adapter: If we add spatial controls [27], such as depth, to control the editing, we observe that 1) the shape is restricted by the source, and 2) the deformable motion cannot follow the source video." + } + ] + } + ], + "index": 44 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Compared to all constructed baselines, our VideoSwap with semantic point correspondence can effectively align the motion trajectory while preserving the target concept's identity." + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7627" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 504, + 280 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 504, + 280 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 504, + 280 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 504, + 280 + ], + "type": "image", + "image_path": "35efebd215b608ff1d1cbf58949505a61f7939ad704f7a8f27204289c90e3c13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 280, + 541, + 291 + ], + "lines": [ + { + "bbox": [ + 51, + 280, + 541, + 291 + ], + "spans": [ + { + "bbox": [ + 51, + 280, + 541, + 291 + ], + "type": "text", + "content": "Figure 8. Ablation Study of our VideoSwap. We encourage readers to click and play the video clips in this figure using Adobe Acrobat." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 297, + 189, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 189, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 189, + 309 + ], + "type": "text", + "content": "4.2. Quantitative Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 288, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 423 + ], + "type": "text", + "content": "We conduct both automatic and human evaluations to quantitatively compare VideoSwap with previous state-of-the-art methods and several baselines onAnimateDiff. Detailed evaluation settings and automatic evaluation results are provided in the supplementary material. For human evaluation, we distribute 1000 questionnaires on Amazon Mturk to assess various criteria in customized video subject swapping. From the human evaluation results in Table. 1, our method achieves a clear advantage over the compared methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 141, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 141, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 141, + 443 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 628 + ], + "type": "text", + "content": "Sparse Motion Feature. There are several variants for encoding semantic points to generate motion guidance. The most straightforward approach is to encode the point map of semantic points using T2I-Adapter [27]. However, this method leads to severe overfitting, as illustrated in Fig. 8(a). The issue arises because the encoded feature added to the diffusion model is non-sparse, and the background also contains features that may overfit to the source video. We use MLPs to encode point embeddings, positioning them in the empty feature to create sparse motion features for guidance without compromising video quality. Regarding point embeddings, we opt for DIFT embeddings, which inherently carry robust semantics. Compared to randomly initialized learnable embeddings, our approach achieves similar results with " + }, + { + "bbox": [ + 46, + 450, + 287, + 628 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 450, + 287, + 628 + ], + "type": "text", + "content": " less time for point registration step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "content": "Point Patch Loss. In the semantic point registration, we employ a point patch loss to reconstruct the local patch surrounding each semantic point. If we omit the point patch loss and opt to directly reconstruct the entire video, we observe that the source structure leaks into the target swapped results and thus produce artifacts, as depicted in Fig. 8(b)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Semantic-Enhanced Schedule. Our goal is to employ se" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 298, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 393 + ], + "type": "text", + "content": "mantic points to transfer motion trajectories, acting as a linkage between the source and target subjects. Therefore, we aim for the semantic points to emphasize high-level semantic alignment without transferring low-level details. This objective is achieved by registering points only on the early sampling steps, i.e., " + }, + { + "bbox": [ + 304, + 298, + 545, + 393 + ], + "type": "inline_equation", + "content": "T_{\\min} = \\frac{T}{2}" + }, + { + "bbox": [ + 304, + 298, + 545, + 393 + ], + "type": "text", + "content": " in Eq. (2). As shown in Fig. 8(c), this technique prevents the model from learning excessive low-level details and enhances point alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 393, + 546, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 546, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 546, + 501 + ], + "type": "text", + "content": "Drag-based Point Control. As illustrated in Fig. 8(d), our goal is to swap the black swan to the duck. If we directly use the source point sequence as guidance, the duck's neck conforms to the shape of the black swan, resulting in an inferior identity. However, by employing the proposed point displacement propagation, we can drag the semantic point at the keyframe, ensuring a consistent motion trajectory after dragging. Utilizing the dragged point trajectory as motion guidance results in the correct identity of the duck." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 532, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 663 + ], + "type": "text", + "content": "This paper uncovers the potential of semantic point correspondence in aligning motion trajectories and altering the subject's identity in video editing. From there, we present VideoSwap, a framework that minimizes human intervention while utilizing semantic point correspondences for customized video subject swapping. Through user-point interactions like point removal or dragging, we address various semantic point correspondence. VideoSwap facilitates shape changes in the target swap while aligning the motion trajectory with the source subject, demonstrating state-of-the-art results in customized video subject swapping." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgement This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7628" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 287, + 169 + ], + "type": "text", + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 236 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "type": "text", + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 317 + ], + "type": "text", + "content": "[5] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "text", + "content": "[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. Videdit: Zero-shot and spatially aware text-driven video editing. arXiv preprint arXiv:2306.08707, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "type": "text", + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "text", + "content": "[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "text", + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3, 4, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[15] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[16] Jiahui Huang, Leonid Sigal, Kwang Moo Yi, Oliver Wang, and Joon-Young Lee. Inve: Interactive neural video editing. arXiv preprint arXiv:2307.07663, 2023. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 260 + ], + "type": "text", + "content": "[17] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "type": "text", + "content": "[19] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "text", + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 436 + ], + "type": "text", + "content": "[21] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14317-14326, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[22] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "text", + "content": "[23] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "text", + "content": "[25] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "text", + "content": "[26] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7629" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "text", + "content": "[29] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "text", + "content": "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "text", + "content": "[31] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 290, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 288, + 345 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 347, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 391 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "text", + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "text", + "content": "[35] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "type": "text", + "content": "[36] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "type": "text", + "content": "[37] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "text", + "content": "[39] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. " + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "inline_equation", + "content": "p+" + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": ": Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 497 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[42] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "text", + "content": "[43] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[44] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[45] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 350 + ], + "type": "text", + "content": "[46] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 398, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 453 + ], + "type": "text", + "content": "[48] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "text", + "content": "[49] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7630" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_content_list.json b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b86f494726f59975bb5e49d5f03d578d2a9defb --- /dev/null +++ b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_content_list.json @@ -0,0 +1,1495 @@ +[ + { + "type": "text", + "text": "View From Above: Orthogonal-View aware Cross-view Localization", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shan Wang1,2 Chuong Nguyen1 Jiawei Liu2 Yanhao Zhang2 Sundaram Muthu1 Fahira Afzal Make1 Kaihao Zhang2 Hongdong Li2 1Data61, CSIRO 2Australian National University", + "bbox": [ + 174, + 179, + 792, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents a novel aerial-to-ground feature aggregation strategy, tailored for the task of cross-view image-based geo-localization. Conventional vision-based methods heavily rely on matching ground-view image features with a pre-recorded image database, often through establishing planar homography correspondences via a planar ground assumption. As such, they tend to ignore features that are off-ground and not suited for handling visual occlusions, leading to unreliable localization in challenging scenarios. We propose a Top-to-Ground Aggregation (T2GA) module that capitalizes aerial orthographic views to aggregate features down to the ground level, leveraging reliable off-ground information to improve feature alignment. Furthermore, we introduce a Cycle Domain Adaptation (CycDA) loss that ensures feature extraction robustness across domain changes. Additionally, an Equidistant Re-projection (ERP) loss is introduced to equalize the impact of all keypoints on orientation error, leading to a more extended distribution of keypoints which benefits orientation estimation. On both KITTI and Ford Multi-AV datasets, our method consistently achieves the lowest mean longitudinal and lateral translations across different settings and obtains the smallest orientation error when the initial pose is less accurate, a more challenging setting. Further, it can complete an entire route through continual vehicle pose estimation with initial vehicle pose given only at the starting point.", + "bbox": [ + 75, + 300, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 720, + 209, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual-based cross-view localization aims to locate query images taken from street-level cameras (referred to as ground view) within a satellite or aerial view map. Platforms like the Google Map API [6] have made satellite images accessible, spurring the development of cross-view localization techniques. Notable studies [4, 7, 10, 15, 20, 25, 26, 30] have focused on using satellite images for this task. Despite this, accurate localization remains challenging due", + "bbox": [ + 75, + 744, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c2274fc7cca192d79b8542da93b13793098cf7c5e1606d7a9a9a9740916040c.jpg", + "image_caption": [ + "Ground View" + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 754, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48b30efe7caab7dabfe78e7bcbcfdca01a7a7924e63a57ea63fca10d5b3bfd1d.jpg", + "image_caption": [ + "Aerial View", + "Figure 1. Efficacy of T2GA. T2GA aggregates off-ground features (e.g., the streetlight within the green square) and addresses occlusions (e.g., the tree within the red square). Unlike conventional methods that often overlook such elements, T2GA integrates these features, thus improving appearance matching for ground-level aligned pixels across views using ground plane homography. Insets are magnified for clarity, with dotted outlines indicating appearance misalignment and solid outlines indicating appearance alignment with the corresponding aerial view features." + ], + "image_footnote": [], + "bbox": [ + 756, + 268, + 890, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the significant viewpoint differences between ground and aerial images [14, 31]. These viewpoint differences result in a domain gap, adversely impacting feature alignment, thereby compromising the overall accuracy of localization.", + "bbox": [ + 496, + 534, + 893, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research has explored two main approaches to bridge the domain gap in image-based localization: generative-based and geometry-alignment-based methods. Generative-based methods, such as those utilizing GANs [19] and diffusion models [29], reduce the domain gap by transforming view styles from one view to another. However, the generated features for matching can lead to ambiguities in pose estimation. In contrast, geometry-alignment-based methods, employing techniques like polar transformations [15, 20] or homography [14, 23, 24, 26], focus on establishing correspondences for on-ground pixels. This often results in the neglect of off-ground features, such as streetlights, and difficulty in handling visual occlusions, such as the obscuring of road details by treetops in aerial views, both illustrated in Fig. 1. This neglect fails to leverage important geographic landmarks on the road and results in a lack of robustness to issues like road mark degradation (e.g. fading and damaged paintings). To address these limitations, we introduce the Top-to-Ground Aggregation (T2GA) module. T2GA employs top-down feature aggre", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at https://github.com/ShanWang-Shan/ViewFromAbove.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14843", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gation to enrich on-ground points with an above-view perspective appearance, significantly improving feature alignment and localization accuracy. Furthermore, we recognize that pixel positions above ground points may not necessarily indicate higher elevations; instead, they could represent objects located further along the camera's line of sight. To resolve this ambiguity, we integrate a transformer mechanism that assesses whether these pixels belong to the same object and determines occlusion precedence, such as shadows on the road, which suggest likely occlusions in the aerial view.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to viewpoint differences, ground and aerial views differ in cameras types, lighting conditions, tones, and resolutions. We introduce the Cycle Domain Adaptation (CycDA) loss function to address these variations. CycDA enables bidirectional feature generation between ground and aerial views. By minimizing the discrepancy between domain-adapted features and their target counterparts, our approach ensures that features extracted from one domain are effectively translatable to the other, fostering the extraction of features that are invariant across domains.", + "bbox": [ + 75, + 243, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further introduce an Equidistant Re-projection Loss to address a common bias in keypoint-synchronous detection localization methods, which tend to favor closer keypoints due to their reduced orientation errors. Our loss function mitigates this issue by applying a distance-weighted approach, ensuring that orientation errors are independent of keypoint distance. Consequently, keypoints are more uniformly distributed across various distances, leading to a more equitable and precise estimation of direction.", + "bbox": [ + 75, + 396, + 472, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 96, + 534, + 388, + 547 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A top-to-ground feature aggregation module that effectively bridges the domain gap between ground and aerial views by utilizing off-road cues and handling occlusions.", + "- A cycle domain adaptation loss function that promotes domain-invariant feature extraction, enhancing the robustness of cross-view localization methods.", + "- An equidistant re-projection loss function ensures orientation errors are consistent regardless of keypoint distance, leading to a more extended distribution of keypoints and more accurate orientation estimation." + ], + "bbox": [ + 76, + 551, + 468, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The evaluations are conducted on KITTI and Ford Multi-AV datasets while leveraging the Google Maps as corresponding satellite images. Experimental results demonstrate that our method can achieve consistent vehicle pose estimation under various challenging situations. Our method achieves the lowest mean longitudinal and lateral translation errors on both KITTI and Ford Multi-AV datasets under different settings. We also obtain the smallest orientation error when the initial pose is less accurate, demonstrating the robustness of our method. Further, with initial vehicle pose only at the starting point, our method can complete the route through continuous pose estimation, demonstrating its generalisation ability and potential for practical deployment.", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-level cross-view localization. Early visual-only cross-view localization methods [7, 10, 15, 16, 20, 30] approach the task as an image retrieval problem, focusing on coarse localization through image-to-image matching. To bridge the domain gap between ground and aerial views, various techniques have been proposed to facilitate cross-view feature matching. [10] incorporated per-pixel orientation information, while [15] and [20] utilized a predefined polar transform to align aerial-view images with ground views. Additionally, [19] and [29] employed GAN-based and diffusion model-based style transformations, respectively. While efforts have been made to minimize the domain gap, these methods rely on image-level feature matching, restricting their localization accuracy, often falling short of the precision achieved by commercial GPS systems in open areas [21].", + "bbox": [ + 496, + 114, + 890, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Patch-wise cross-view localization. To improve accuracy, several methods employ patch-wise feature matching, especially effective in aerial views with wide fields of view and high resolutions. For instance, [31] employs transformers to emphasize informative patches, [27, 28] computes dense spatial distributions using patch attention, and [9] introduces the 'slice-to-sector match' and 'Cross-view attention' to calculate similarity between aerial sectors and ground view slices. However, their attention queries are derived from aerial images, posing reliability issues when the pose is unknown. OrienterNet [13] transforms ground view images into Bird's Eye View (BEV) grids for matching with OpenStreetMap data. Despite these advancements, localization accuracy remains limited by the patch (grid) size.", + "bbox": [ + 496, + 357, + 890, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pixel-wise cross-view localization. Pixel-wise feature matching has been explored in various methods [4, 14, 24, 25] for precise localization, yet they struggle with the inherent domain gap between views. Our baseline, PureACL[24], and other geometry-alignment-based methods [14, 26] prioritize on-ground pixel correspondences but overlook off-ground features and occlusions, leading to suboptimal performance. CVGL [4] transforms the ground view image into a BEV for aerial matching. Their BEV transformation employs aerial coordinates as queries, introducing a high degree of freedom and increasing the complexity of matching. In the BoostAcc [17] framework, homography transformation is used on ground views to source query pixel data for the pixel-to-slice attention mechanism. This transformation introduces distortions from non-ground pixels, potentially exacerbating errors in subsequent processing stages. Moreover, the keys and values span entire and adjacent columns, risking compounded ambiguities in both longitudinal and lateral estimation. Our approach aims to address these challenges, proposing a novel pixel-to-overhead-column attention mechanism for improved cross-view localization accuracy.", + "bbox": [ + 496, + 569, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14844", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8c62cdf43ff5f11886d258e92cc1ca6325f8d87ff7c015e68783cb5c251e609.jpg", + "image_caption": [ + "Figure 2. Our method adopts PureACL [24] as the baseline model and introduces three novel components (highlighted in red): (1) T2GA (Sec.4.1): aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them to alleviate the representation gap of the same object across different views; (2) CycDA Loss (Sec.4.2): explicitly enforces a view-invariant representation for the same object; and (3) ERP Loss (Sec.4.3): allows the model to leverage key points that are farther away from the vehicle while allocating more emphasis on correcting the vehicle orientation estimation." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 291, + 202, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Settings. The Fine-Grained Cross-View Localization (FGCVL) task aims to estimate the accurate 3-DoF pose of a vehicle, including longitudinal and lateral translations and orientation on a satellite map $I^{s}$ , given an initial coarse pose $\\mathbf{P}_{init}$ . It also assumes that the vehicle is equipped with on-board camera(s) $^2$ , and the image captured by these cameras, denoted as the ground image $I^{g}$ , is accessible.", + "bbox": [ + 75, + 316, + 470, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baseline Model. PureACL [24] is adopted as the baseline model for our approach. It employs a share-weight U-Net to extract ground/satellite feature maps $F = \\{F^g, F^s\\}$ and ground/satellite confidence maps $C = \\{C^g, C^s\\}$ from the ground/satellite images $I = \\{I^g, I^s\\}$ . The ground confidence map $C^g$ is tasked to select the top-N most confident pixel-level features from $F^g$ at the positions $p^g = \\{(u_i^g, v_i^g)\\}_{i=1}^N$ . These features are then: (i) Projected from on-board camera coordinates to the satellite coordinates using $p_i^s(\\mathbf{P}) = K_s \\mathbf{P} K_g^{-1} p_i^{g,4}$ , where $K_s$ and $K_g$ are the intrinsic matrices of satellite and on-board cameras, respectively. $\\mathbf{P}$ is the initial pose $\\mathbf{P}_{init}$ at the first iteration and the estimated pose $\\mathbf{P}_{pred}$ in the subsequent iterations. (ii) A weighted residual $w(\\mathbf{P}) \\times r(\\mathbf{P})$ is computed between the top-N selected features from the ground feature map and the corresponding features from the satellite feature map as:", + "bbox": [ + 75, + 422, + 468, + 666 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw (\\mathbf {P}) = C ^ {s} \\left[ p ^ {s} (\\mathbf {P}) \\right] \\times C ^ {g} \\left[ p ^ {g} \\right], r (\\mathbf {P}) = F ^ {s} \\left[ p ^ {s} (\\mathbf {P}) \\right] - F ^ {g} \\left[ p ^ {g} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 674, + 468, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\cdot ]$ is a lookup operation in a feature/confidence map with sub-pixel interpolation. (iii) The Levenberg-Marquardt (LM) algorithm [11] takes the computed weighted residual $w(\\mathbf{P})\\times r(\\mathbf{P})$ as input and outputs the predicted vehicle pose $\\mathbf{P}_{pred}^{m}$ at the $m^{\\mathrm{th}}$ iteration, with $m\\in$ $\\{1,\\ldots ,M\\}$ . The process of (i) - (iii) repeats for $M = 20$", + "bbox": [ + 76, + 700, + 470, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "iterations to produce the final vehicle pose $\\mathbf{P}_{pred}$ .", + "bbox": [ + 500, + 292, + 828, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PureACL [24] employs two loss functions: a Triplet Loss [12] and a Re-projection Loss. The former supervises the weighted residual, enforcing that the encoder extracts pose-sensitive features from both ground and satellite images. The later utilises the projection error w.r.t the ground truth pose $\\mathbf{P}_{gt}$ to penalize incorrectly predicted vehicle pose $\\mathbf{P}_{pred}$ (Refer to Supplementary Sec.D for more details).", + "bbox": [ + 496, + 309, + 890, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 430, + 589, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel View From Above (VFA) method to tackle the FGCVL task by aligning feature representations across the orthogonal viewpoints, the ground view and the satellite views. As illustrated in Fig. 2, our method introduces three innovative components onto the baseline model: (1) Top-to-Ground Aggregation Module (T2GA) (Sec. 4.1); (2) Cycle Domain Adaptation Loss (CycDA) (Sec. 4.2); and (3) Equidistant Re-Projection Loss (Sec. 4.3).", + "bbox": [ + 496, + 455, + 890, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Top to Ground Aggregation", + "text_level": 1, + "bbox": [ + 500, + 589, + 750, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The appearance discrepancy between the same objects viewed from ground and satellite perspectives presents a significant challenge in achieving high precision in the FGCVL task [14, 17, 24, 26]. This issue is often encountered on tall structures, e.g. street light, and the occlusion from aerial objects, e.g. tree branches, where their near(on)-ground appearance, accessible to only on-board vehicle camera, is significantly different from their top appearances which are only registered by the satellite camera (see Fig. 1). Such discrepancies are common on roads and can lead to inaccuracies in cross-view localization. To address this challenge, we propose aligning the representations between ground and satellite views, focusing specifically on the on-ground pixels, where geometry alignment across the two views hold due to the ground plane homography. Specifically, we present a T2GA module that aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them, as illustrated in the 'T2GA' red box in Fig. 2. For an on-ground pixel:", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2While our method is compatible with multiple cameras, our primary experiments focus on one front camera and four surrounding cameras.", + "bbox": [ + 75, + 801, + 467, + 825 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3We simply refer to the fused confidence map, rather than distinguishing between view-consistent and on-ground confidence maps in PureACL.", + "bbox": [ + 76, + 825, + 467, + 851 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "PureACL obtains 3D world coordinates of the vehicle from homogeneous coordinates $K_{g}^{-1}p^{g}$ by utilizing on-ground characteristics. For simplicity in this context, we note homogeneous coordinates as directly corresponding to the 3D world coordinates.", + "bbox": [ + 76, + 851, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$p\\in \\{(u,v)\\}_{u = 0,v = \\tau}^{W^g -1,H^g -1}$ , and where $\\tau$ is a threshold corresponding to the height of horizon in the ground view image, and $W^{g}$ and $H^{g}$ are the width and height of the ground view image, respectively. The corresponding elevated pixels are $q\\in \\{(u_p,v)\\}_{v = 0}^{v_p}$ , where $(u_{p},v_{p})$ are the coordinates of the on-ground pixel $p$ . The aggregation is achieved via the attention mechanism [22], defined as:", + "bbox": [ + 75, + 87, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {a t t} [ p ] = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {T}\\right) \\mathbf {V}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 208, + 468, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F^{att}[p]$ represents the feature vector at a pixel $p$ on the feature map $F^{att}$ . In this context, $p$ and $q$ denote an on-ground pixel and its corresponding elevated pixels, respectively. The query is formed by taking the features of on-ground pixels $\\mathbf{Q} = \\mathcal{M}(F^g[p])$ where $\\mathcal{M}(\\cdot)$ represents a non-linear mapping consisting of an MLP layer followed by an activation function. The value is derived from the features of elevated pixels corresponding to the on-ground pixel $\\mathbf{V} = F^g[q]$ , and the key is then obtained by applying the same non-linear mapping to $\\mathbf{V}$ , resulting in $\\mathbf{K} = \\mathcal{M}(\\mathbf{V})$ . This non-linear mapping ensures that the query and key are mapped to the same feature space before computing their matrix product. $\\mathbf{Q}\\mathbf{K}^T$ . Notably, value $\\mathbf{V}$ is used directly in its original form for the computation, facilitating a straightforward fusion (for the same object) or replacement (in cases of occlusion) at ground level.", + "bbox": [ + 75, + 238, + 472, + 479 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/21d5ec7a225c4edfc15de4afa07b9b409317055b7e36f3ec5aa957ce9a2935a7.jpg", + "image_caption": [ + "Figure 3. Attention between the on-ground pixels and their corresponding elevated pixels is displayed column-wise. (Left example): The attention between the base and top of the streetlight is high despite their distinct appearances. This allows the resultant aggregated ground feature to be aligned with the satellite feature corresponding to the matching geographic location. (Right example): In the absence of occlusion from above, ground pixels maintain minimal attention with their elevated pixels. This avoids unnecessary dilution of the ground features which are already well aligned with their satellite counterparts." + ], + "image_footnote": [], + "bbox": [ + 81, + 494, + 477, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The computed attention map $\\operatorname{Softmax}(\\mathbf{Q}\\mathbf{K}^T)$ can accurately reflect the connection between the on-ground pixels and their corresponding elevated pixels. As illustrated in Fig. 3, high attention values are assigned to the base and top", + "bbox": [ + 75, + 775, + 470, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the streetlight despite their different appearances. This provides important cues to alleviate the representation gap between the ground feature and satellite features. On the other hand, in the absence of occlusion, e.g. tall structure and tree branch, thus the attention between on-ground and elevated pixels is minimal, avoiding diluting the ground features that are well aligned with their satellite counterparts.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The feature map $F^{att}$ after attention calculation is vertically stacked with the upper part of the original $F^g$ to generate the aggregated feature map $F^a$ :", + "bbox": [ + 496, + 196, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {a} = F _ {0: H ^ {g} - \\tau} ^ {g} \\oplus F ^ {a t t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 253, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes vertically stacking and $F_{0:H^g -\\tau}^g$ indicated the slicing of $F^{g}$ from row 0 to $H^{g} - \\tau$ . Subsequently, this aggregated feature map $F^a$ replaces $F^g$ in the residual calculation as per Eq.1, resulting in an updated formula: $r(\\mathbf{P}) = F^{s}[p^{s}(\\mathbf{P})] - F^{a}[p^{g}]$", + "bbox": [ + 496, + 282, + 890, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The advantage of the aggregated feature map is also indirectly reflected by the ground confidence map of the baseline model as shown in Fig. 4. Without the T2GA module, the confidence map only highlights pixels corresponding to road marks and curds, resulting in a concentrated sampling of key points for subsequent estimation of vehicle pose. This has the following drawbacks: (1) fail to leverage road landmarks e.g. traffic signal poles, that encode important geographic location; and (2) not robust to road mark degradation problems, e.g. fading and damaged paintings, and visibility issues, e.g. glow on road marks. T2GA alleviates the feature discrepancy between ground and satellite views. As a result, the traffic signal pole receives higher confidence, allowing it to be subsequently sampled as key points to estimate more precise vehicle pose.", + "bbox": [ + 496, + 359, + 892, + 585 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a68c878c1dd72a93fadf4187eeeb97cdb6f7b5ab34270c603c6b8b7618100bf.jpg", + "image_caption": [ + "Figure 4. Illustration on the effect of T2GA on the confidence map of the baseline model. The confidence map without T2GA (Left) predominantly highlights road marks and curbs, resulting in subsequent keypoint sampling missing important road landmarks, e.g. traffic signal poles, that provide important cues to vehicle pose estimation; The confidence map with T2GA (Right) has high-confidence values distributed across various road marks and traffic poles. With more geographic cues provided by multiple sources, the resultant pose prediction becomes more precise and robust." + ], + "image_footnote": [], + "bbox": [ + 504, + 598, + 700, + 648 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b0f56c8b54e0dc05d3122eb5250d2d7ca74407891b3a3902469405b916a275d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 598, + 898, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Cycle Domain Adaptation Loss", + "text_level": 1, + "bbox": [ + 498, + 801, + 774, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are multiple sources responsible for the representation gap between ground and satellite features. In addition to the varying appearances of the same object across different views, camera specifications, e.g. tone, hue, intensity, brightness and resolution, and temporal changes of", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "In processing all query pixels, we first compute the matrix multiplication $\\mathbf{Q}\\mathbf{K}^T$ using all column pixels $q\\in \\{(u_p,v)\\}_{v = 0}^{H^g -1}$ . We then eliminate the attention values corresponding to pixels below each query pixel $\\{(u_p,v)\\}_{v = v_p + 1}^{H^g -1}$ . After this elimination, the Softmax function is applied.", + "bbox": [ + 76, + 845, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14846", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "varying spans, can also lead to inconsistent representations for the same geographic location. Such issue is in general overlooked by the existing FGCVL methods [14, 28]. Despite introducing a triplet loss to further distinguish feature representations based on different poses, our baseline model, PureACL [24], falls short in enforcing invariant feature representation at the corresponding geographic locations across different views.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a Cycle Domain Adaptation (CycDA) loss to explicitly enforce view-invariant representations that consists of three L2-loss between ground and satellite feature representations in three different feature spaces. The representation loss in the ground feature space is defined as:", + "bbox": [ + 75, + 210, + 470, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| F ^ {a} \\left[ p _ {i} ^ {g} \\right] - \\mathcal {D} \\left(\\mathcal {E} \\left(F ^ {s} \\left[ p _ {i} ^ {s} \\left(\\mathbf {P} _ {g t}\\right) \\right] \\odot c _ {s}\\right) \\odot c _ {g}\\right) \\right\\| _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 308, + 468, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}(\\cdot)$ is a projection from the pixel feature space to a latent feature space, $\\mathcal{D}(\\cdot)$ is a projection from the latent feature space to the pixel feature space, $c_{s}$ and $c_{g}$ represent the focal length of satellite camera and vehicle camera respectively which are introduced to guide the projection functions, and $\\odot$ is a concatenation acting on the feature channel dimension. Similarly, the representation loss in the satellite feature space is:", + "bbox": [ + 75, + 353, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| F ^ {s} \\left[ p _ {i} ^ {s} \\left(\\mathbf {P} _ {g t}\\right) \\right] - \\mathcal {D} \\left(\\mathcal {E} \\left(F ^ {a} \\left[ p _ {i} ^ {g} \\right] \\odot c _ {g}\\right) \\odot c _ {s}\\right) \\right\\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 483, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also enforce the satellite feature to be aligned with the ground feature corresponding to the same geographic location in a latent feature space through:", + "bbox": [ + 75, + 527, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| \\mathcal {E} \\left(F ^ {a} [ p _ {i} ^ {g} ] @ c _ {g}\\right) - \\mathcal {E} \\left(F ^ {s} [ p _ {i} ^ {s} (\\mathbf {P} _ {g t}) ] @ c _ {s}\\right) \\right\\| _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 582, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall CycDA loss can then be defined as:", + "bbox": [ + 76, + 626, + 390, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C y c D A} = \\mathcal {L} _ {g} + \\mathcal {L} _ {s} + \\mathcal {L} _ {m} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 652, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Equidistant Re-projection Loss", + "text_level": 1, + "bbox": [ + 76, + 679, + 352, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method tackles a challenging setting of simultaneously detecting key points from the ground image and predicting the vehicle pose using the detected key points. Our baseline model introduces a re-projection loss computed over the selected top-N key points defined as:", + "bbox": [ + 75, + 702, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| p _ {i} ^ {s} (\\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\\mathbf {P} _ {g t}) \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 787, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i^s (\\mathbf{P}_{pred})$ and $p_i^s (\\mathbf{P}_{gt})$ are the projected coordinates on the satellite image from the selected top-N ground key points with the predicted vehicle pose and the groundtruth vehicle pose respectively, and $\\| \\cdot \\|_2$ denotes the L2-distance.", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a278da9d73fe47a641ce4bc30610a218894337cb1901b491a0a06e10daa1113.jpg", + "image_caption": [ + "Figure 5. Comparison of detected keypoints with and without ERP Loss. (Left) Without the ERP Loss, keypoints are predominantly located in close proximity to the vehicle. (Right) With the ERP Loss, there is a more dispersed distribution of keypoints." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 699, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d7b72283fb9b53d1a00eeefca7e1fbadd8b63ecf1e517fef4169d3bd492e30cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 89, + 898, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite being effective in pipelines where simultaneous detection of key points is not required [3, 18, 25], the reprojection loss tends to overly penalize the orientation errors of distant points. This forces the detected key points being closer to the vehicle, as illustrated in Fig. 5, and less accurate vehicle orientation estimation (see Sec. 5.2).", + "bbox": [ + 498, + 212, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the aforementioned issue, we propose an Equidistant Re-Projection (ERP) loss that is defined as:", + "bbox": [ + 500, + 303, + 890, + 333 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {E R P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\frac {p _ {i} ^ {s} (\\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\\mathbf {P} _ {g t})}{\\mathbf {D} _ {p _ {i} ^ {s}}} \\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 343, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{D}_{p_i^s} = \\| p_i^s (\\mathbf{P}_{gt}) - p_{vehicle}^s (\\mathbf{P}_{gt})\\|_2$ is the L2-distance between the satellite image coordinates of the $i^{\\mathrm{th}}$ keypoint and the vehicle. The design prevents the orientation error from scaling with the L2-distance between key points and vehicle on the satellite image, which dominates over the orientation error on distant key points. As shown in Fig. 5, our method can leverage key points that are farther from the vehicle than our baseline model.", + "bbox": [ + 498, + 393, + 892, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 529, + 632, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 553, + 715, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To evaluate our proposed approach, we conduct experiments on two well-established autonomous driving datasets: the Ford Multi-AV Seasonal dataset (FMAVS) [1] and the KITTI dataset [5]. Consistent with established practices [14, 24, 25], our primary focus is on images from the front left camera as query images. Additionally, following the methodology of PureACL [24], we broadened our analysis to include four camera perspectives: front left, rear right, side left, and side right. For data partitioning in the KITTI-CVL dataset, we utilized an approach in line with HighlyAcc [14], which involves two test splits. The first, 'Same,' includes images from the same trajectories as the training dataset, while the second, 'Cross,' comprises images from distinct trajectories. Regarding the FMAVS dataset, we follow the partitioning strategy proposed by PureACL [24], using all images from the same trajectories but different traversals. It is noteworthy that there are minor route variations within these same trajectories6", + "bbox": [ + 498, + 577, + 890, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We adopt evaluation metrics in line with HighlyAcc [14], SliceMatch [9], SIBCL [25], and PureACL [24].", + "bbox": [ + 500, + 848, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "An example is provided in Supplementary Fig. 1.", + "bbox": [ + 514, + 886, + 785, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c9b8e82a8d50285e0b04b770506da0258ede8cf9c324fe832dc567fd26b6f061.jpg", + "table_caption": [ + "Table 1. Comparison on the KITTI-CVL Dataset Under Initial Noise Conditions (±10°,±20m). Results marked with $\\star$ are sourced directly from the respective original papers. Those denoted by $\\diamond$ indicate retraining of methods for consistent evaluation criteria alignment. - indicates that the corresponding data were not provided." + ], + "table_footnote": [], + "table_body": "
ModelLocation(m)↓Lateral(%)↑Longit(%)↑Orient(°)↓Orient(%)↑
MeanMedianr@1mr@3mr@5mr@1mr@3mr@5mMeanMedianr@1°r@3°r@5°
Same★DSM[15]--10.1230.6748.244.0812.0120.14--3.5813.8124.44
★HighlyAcc[14]12.0811.4235.5470.7780.365.2215.8826.133.722.8319.6451.7671.72
★SliceMatch[9]7.964.3949.09-98.5215.19-57.354.123.6513.41-64.17
★BoostAcc[17]10.015.1976.4496.3498.8923.5450.5762.180.550.4299.10100.00100.00
★CCVPE [28]1.220.6297.3598.6599.7177.1396.0897.160.670.5477.3999.4799.95
★HC-Net [26]0.800.5099.01-99.7392.20-99.250.450.3391.35-99.84
◇PureACL [24]2.420.4291.9593.4094.2891.8692.1292.383.971.7732.7153.8971.66
Ours0.200.1799.9799.9799.9799.9799.9799.971.530.9350.9583.9795.45
Cross★DSM[15]--10.7731.3748.243.8711.7319.50--3.5314.0923.95
★HighlyAcc[14]12.5812.1127.8259.7972.895.7516.3626.483.953.0318.4249.7271.00
★SliceMatch[9]13.509.7732.43-86.448.30-35.574.206.6146.82-46.82
★BoostAcc[17]13.019.0657.7286.7791.1614.1534.5945.000.560.4398.98100.00100.00
★CCVPE [28]9.163.3344.0681.7290.2323.0852.8564.311.550.8457.7292.3496.19
★HC-Net [26]8.474.5775.00-97.7658.93-76.463.221.6333.58-83.78
◇PureACL [24]6.200.6167.2487.2692.7664.8173.6389.694.262.4823.4548.6959.24
Ours0.210.1899.9299.9699.9799.9199.9199.912.041.3838.3180.1292.48
", + "bbox": [ + 76, + 132, + 893, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our precision assessment includes median and mean errors for overall, lateral, and longitudinal translations, as well as orientation accuracy. Additionally, our analysis covers lateral and longitudinal translations and localization recall at various distances (0.25m, 0.5m, 1m, 3m, and 5m), and orientation recall within a range of $1^{\\circ}$ to $5^{\\circ}$ .", + "bbox": [ + 75, + 415, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We apply two sets of criteria: (1) Following HighlyAcc [14], ground images are processed at a resolution of $256 \\times 1024$ , and satellite images at $512 \\times 512$ , with initial pose noise $\\pm 10^{\\circ}$ for orientation and $\\pm 20\\mathrm{m}$ for translations. (2) In accordance with PureACL [24], we process FMAVS images at $432 \\times 816$ , KITTI images at $384 \\times 1248$ , and satellite images at $1,280 \\times 1,280$ . To better simulate turning scenarios, we adopt an expanded initial pose noise range of $\\pm 45^{\\circ}$ for orientation and $\\pm 20\\mathrm{m}$ for translations. For training, we utilize an NVIDIA RTX 3090 GPU with a batch size of 3, employing the Adam optimizer [8] with a learning rate of $10^{-5}$ . Feature extractor weights are adapted from PureACL [24], while other components are initialized randomly. Training iterations average around $285\\mathrm{ms}$ , including $200\\mathrm{ms}$ dedicated to optimization. The inference speed, subject to initial pose variability, averages at $222\\mathrm{ms}$ .", + "bbox": [ + 75, + 505, + 470, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 76, + 756, + 387, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method against recent visual-only approaches on the KITTI-CVL dataset following the metrics of HighlyAcc [14]. The results shown in Tab. 1 clearly indicate the superiority of our approach in spatial precision, consistently maintaining poses within a $1\\mathrm{m}$ radius in both 'Same' and 'Cross' areas with over $99.9 + \\%$ probability. While our method may not lead in orientation accuracy under an initial pose range of $\\pm 10^{\\circ}$ , it demonstrates", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "superior performance when the range is extended to $\\pm 45^{\\circ}$ , as detailed in Tab. 2. This underlines the robustness of our approach. Our approach leverages pixel-wise localization, offering finer granularity compared to the patchwise localization of SliceMatch[9], CCVPE [28], and the image-level localization of DSM [15]. Although HighlyAcc [14] and HC-Net [26] are pixel-wise localization methods, their homography-based mechanisms fail to incorporate off-road cues. In contrast, our approach effectively utilizes these cues, resulting in enhanced performance. Additionally, our longitudinal estimation significantly surpasses that of BoostAcc [17]. This improvement is likely attributable to our method's strategy of not aggregating information from pixels below, thereby reducing longitudinal ambiguity. In comparison to the baseline method PureACL [24], our approach achieves significant enhancements in both orientation and spatial accuracy, especially in mean error reduction. Specifically, we observe a $91^{+}\\%$ reduction in spatial mean error and a $52^{+}\\%$ reduction in orientation mean error. The difference between mean and median errors for PureACL indicates potential convergence issues in high-noise environments. By integrating the T2GA and CycDA modules, we bridge the domain gap and ensure convergence even under challenging conditions.", + "bbox": [ + 496, + 415, + 893, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Challenge Initial Pose and Stringent Metrics", + "text_level": 1, + "bbox": [ + 498, + 786, + 877, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To rigorously evaluate our method, we adopt the stringent metrics of PureACL [24], which are more demanding than those used by HighlyAcc[14]. We benchmark our method against baseline methods SIBCL [25] and PureACL [24], as well as SOTA methods CCVPE [28] and BoostAcc [17]. The results in Tab. 2 reaffirm the advantages of our", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14848", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5e71760d777453cd03eabd6dce9024eed14eb0489e49ad2585969206f55301e5.jpg", + "table_caption": [ + "Table 2. Comparison with Initial Noise Conditions $(\\pm 45^{\\circ},\\pm 20\\mathrm{m})$ .For the KITTI-CVL dataset, evaluations are conducted in 'K-Same' and 'K-Cross' areas. On the Ford-CVL dataset, the Log4' trajectory, as used in SIBCL [25], is chosen for its optimal satellite view alignment. Additional log evaluations are detailed in the supplementary material. F-1C' and F-4C' represent assessments on the Ford-CVL dataset using either a single front-facing camera or four surrounding cameras. Note: $\\star$ indicates that SIBCL [25] is a hybrid LiDAR-visual method." + ], + "table_footnote": [], + "table_body": "
ModelLateral(m)↓Mean MedianLateral(%)↑r@0.25mr@0.5mr@1mLongit(m)↓Mean MedianLongit(%)↑r@0.25mr@0.5mr@1mOrient(°)↓Mean MedianOrient(%)↑r@1°r@2°r@4°
K-SameCCVPE [28]2.321.2211.0522.0542.096.493.357.4513.3624.973.431.5334.6660.7284.03
BoostAcc[17]1.190.6321.9241.5167.7210.015.425.179.3816.543.882.9818.4535.2562.20
★ SIBCL[25]1.230.7017.0133.3461.883.140.8620.6831.4055.588.901.6935.5055.0471.06
PureACL[24]2.940.2358.7566.3470.603.150.3745.6568.2374.878.132.8037.9952.5764.13
Ours0.170.1476.2097.7399.970.080.0798.1899.8999.972.191.0748.9766.2990.54
K-CrossCCVPE [28]4.562.794.9610.0538.1212.068.163.136.1412.4519.4615.393.3810.2417.41
BoostAcc[17]2.650.9314.8829.2052.6410.468.923.385.5611.735.614.2612.4624.8747.26
★ SIBCL[25]2.720.7116.9032.6258.245.481.0519.5830.7449.259.671.9625.0250.1860.16
PureACL [24]4.240.2550.2562.6364.563.890.447.4162.0464.2310.963.1619.9536.2556.49
Ours0.170.1575.9697.4799.970.090.0797.4999.8799.972.881.4436.1063.0886.18
F-1CCCVPE [28]4.622.456.4812.5424.4610.218.413.228.2413.5320.4815.073.146.1416.14
BoostAcc[17]2.731.479.0518.1835.5910.516.144.218.1315.566.824.7911.6122.6442.63
★ SIBCL[25]2.590.7120.9241.5360.435.381.1812.5923.9943.876.341.5735.2158.4170.65
PureACL[24]2.941.5411.3421.1437.764.751.7410.4620.0737.647.383.3315.1230.4257.01
Ours0.490.4131.7759.1789.230.360.3238.4375.7297.462.191.0847.3172.7190.10
F-4CCVGL[4]1.280.5021.8150.0586.841.530.7019.3637.2665.51-----
PureACL[24]1.690.6721.0739.5365.273.381.1812.5923.9943.863.071.1943.8473.2782.41
Ours0.130.1088.2598.53100.000.200.1868.6796.68100.002.001.4935.7262.7091.83
", + "bbox": [ + 76, + 146, + 893, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/803aa4bd5d4ce7de98c86077c936849f854378747adffb0d7364370b7415f09b.jpg", + "image_caption": [ + "(a) Location" + ], + "image_footnote": [], + "bbox": [ + 81, + 483, + 272, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61468ab18c74c8c4334bbaa0adf3608827ee0a5aababe9613a7eb7987fa6f1ae.jpg", + "image_caption": [ + "(b) Orientation", + "Figure 6. Comparison of the baseline method PureACL and our proposed method under varying initial noise ranges in the 'cross' area of the KITTI-CVL dataset. The blue bar indicates the median value, while the red bar shows the difference between the median and mean values. Notably, our method demonstrates robust convergence, outperforming PureACL across all noise ranges, especially in larger noise scenarios where PureACL shows significant discrepancies between median and mean values." + ], + "image_footnote": [], + "bbox": [ + 282, + 483, + 465, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method, particularly in terms of reduced mean error and improved longitudinal estimation. Notably, under conditions of minimal initial pose noise $(\\pm 15^{\\circ}$ and $\\pm 5\\mathrm{m})$ , PureACL [24] outperforms SIBCL [25]. However, when faced with more challenging conditions $(\\pm 45^{\\circ}$ and $\\pm 20\\mathrm{m})$ , PureACL [24] encounters convergence challenges, unlike our method, which demonstrates consistent superiority. This highlights the effectiveness of our proposed modifications in feature alignment. Additionally, we assess our method's convergence range against PureACL [24] in the 'Cross' area of the KITTI-CVL dataset, as depicted in Fig. 6. The red bar in the figure highlights a significant reduction in the gap between median and mean errors. This observation con", + "bbox": [ + 75, + 704, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "firms our method's enhanced ability to converge effectively across a wider range of noise levels.", + "bbox": [ + 500, + 469, + 890, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results with Continual Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 512, + 838, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address GPS signal loss, we adopt an accumulated pose estimation strategy that leverages initial poses based on the vehicle's previous pose estimates $^{7}$ . We use the model trained with initial noise allowances of $\\pm 45^{\\circ}$ and $\\pm 20\\mathrm{m}$ . Our method's robustness is evidenced by the continuous running distance percentage presented in Tab. 3, showing successful pose chaining throughout all evaluate routes. In contrast, BoostAcc [17] and the single-camera PureACL [24] exhibit significant drift, leading to errors beyond the satellite map's coverage and causing evaluations to cease after covering less than $11\\%$ of the distance. Performance comparisons on the KITTI-CVL dataset are shown in Fig. 7 $^{8}$ . Our approach consistently achieves accurate pose estimation or experiences only minor drifts in challenging scenarios characterized by limited localization cues or severe occlusions. In contrast, BoostAcc [17] exhibits significant drift when a moving vehicle passes by, likely due to incorrect query data from projected vehicle pixels. Furthermore, PureACL [24] appears to struggle with incorrect orientation, resulting in reversed pose estimations. These out", + "bbox": [ + 496, + 536, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "This pose estimation is purely frame-based and does not involve any sequence-based filtering techniques.", + "bbox": [ + 500, + 849, + 890, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "8Performance comparisons on the Ford-CVL dataset is shown in Fig. 1 and Fig. 2 of supplementary material.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14849", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dba6508b74ea847d65b37c6f59c14ef1ebfa503e182090c33be623c1b44d26b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 99, + 271, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d50ca463c12e7424e05b9a54a4ad8730d11f9bb30130cb4e506abef533043bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 99, + 467, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63220183812e94646ecfc564e94df1d84abe5069491dc07688399c88daf2fba2.jpg", + "image_caption": [ + "Figure 7. Accumulated pose estimation performance on a KITTI-CVL dataset trajectory. (Top-left) Predicted poses by our method (blue) closely match the ground truth (red) over the 3676-meter route, outperforming PureACL (green) and BoostAcc (cyan), which exhibit severe drift. (Top-right) Our approach demonstrates rapid recovery upon encountering clear localization cues. (Bottom-left) (Bottom-right) BoostAcc demonstrates substantial drift when encountering moving vehicles." + ], + "image_footnote": [], + "bbox": [ + 80, + 212, + 272, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83fa2f54b3f2ceb0d3a63586dde6dd8fc200e2b21f8265a85929c88f7afe519a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 212, + 467, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comes highlight the importance of robustness in real-world applications, as drifts that extend beyond the coverage of satellite imagery can compromise all further pose estimation processes.", + "bbox": [ + 75, + 439, + 468, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e3f55b804473dcceb6deb9bb5ef9f99aa114ff75d6643e6f020c3e0369d8ee17.jpg", + "table_caption": [ + "Table 3. Comparison on Accumulated Pose Estimation." + ], + "table_footnote": [ + "- indicates that BoostAcc does not support 4-camera setting." + ], + "table_body": "
ModelRunning percentage(%)↑
K-CrossF-1CF-4C
BoostAcc[17]4.884.89-
PureACL[24]10.315.01100.00
Ours100.00100.00100.00
", + "bbox": [ + 89, + 525, + 457, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 633, + 228, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed components, we conduct ablation studies in the 'cross' area of the KITTI-CVL dataset. Our method's performance was compared across various configurations, including the presence and absence of CycDA and Equidistant Re-projection (ERP) Loss functions, and the T2GA module. The results, as detailed in Tab. 4, underscore the significant role each component plays in improving our method's performance.", + "bbox": [ + 75, + 659, + 467, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The T2GA module aggregates information from top to ground based on the assumption that the ground and satellite views are orthogonal. However, this assumption might not be accurate in scenarios involving uphill/downhill or turns. To test our method's robustness under these conditions, we introduced affine warping to the ground view images with a shear range of $\\pm 15^{\\circ}$ . The outcomes, labeled 'w/ Affine' in Tab. 4, indicate our method's efficacy in handling such", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scenarios. This success is primarily due to two factors: (1) the use of a U-Net for feature extraction allows each pixel in the feature map to encapsulate information from its surrounding area, enabling the handling of a degree of angular variation. (2) the satellite view from Google, which is not strictly orthogonal to the ground view, suggests that our method is already adapted to these types of challenges.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e46f206c991a394e82477bc75c5dcabdcfd40f39cf6e3fd5cce45370165ff43c.jpg", + "table_caption": [ + "Table 4. Ablation study on the KITTI-CVL dataset 'Cross' area with initial noise $(\\pm 45^{\\circ},\\pm 20\\mathrm{m})$" + ], + "table_footnote": [], + "table_body": "
T2GA CycDA ERPLateral(m)↓Mean MedianLongit(m)↓Mean MedianOrient(°)↓Mean Median
4.240.253.890.4410.963.16
✓1.290.191.270.174.112.13
✓✓0.210.180.120.103.381.89
✓✓✓0.170.150.090.072.881.44
w/ Affine0.180.150.090.072.881.52
night0.390.270.390.306.642.55
", + "bbox": [ + 503, + 233, + 885, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate our method on an artificial dataset derived from the KITTI test sets, created using the cross-domain deep network proposed by Arruda et al. [2]. The outcomes, labeled as 'night' in Tab. 4 show our method's stable performance in nighttime scenarios, highlighting its resilience to variations in lighting and time of day. For further details, please refer to Supplementary Sec. C.", + "bbox": [ + 496, + 380, + 890, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 500, + 617, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a novel top-to-ground feature aggregation for enhancing cross-view image-based geolocation. Our method overcomes limitations by incorporating aerial perspectives and utilizing a cycle domain adaptation loss for consistent feature extraction despite visual disparities. The introduction of the equidistant re-projection loss balances keypoint impact, promoting wider distribution and thus enhancing orientation accuracy. Our method excels in vehicle pose estimation across challenging scenarios, achieving the lowest translation errors in KITTI and Ford datasets, and minimal orientation error with less accurate initial poses. Crucially, by relying solely on the initial vehicle pose at the start, our method successfully completes routes via continuous pose estimation. This paves the way for real-world applications like autonomous driving and outdoor robotics. Future work will focus on integrating this method into SLAM systems to remove loop closure dependency and facilitate high-definition map generation.", + "bbox": [ + 496, + 526, + 890, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project is supported by CSIRO Julius Career Award, CSIRO Early Research Career Fellowship, CSIRO Future Digital Manufacturing Fund Postdoc Fellowship, CSIRO Data61 PhD Scholarship to SW, and partially by Australian Research Council Discovery Grant (DP220100800) to HL.", + "bbox": [ + 496, + 816, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14850", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 119, + 173, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Siddharth Agarwal, Ankit Vora, Gaurav Pandey, Wayne Williams, Helen Kourous, and James McBride. Ford multi-AV seasonal dataset. The International Journal of Robotics Research, 39(12):1367-1376, 2020. 5", + "[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 8", + "[3] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 5", + "[4] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1, 2, 7", + "[5] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5", + "[6] Google. Maps static api, 2023. 2023. 1", + "[7] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 1, 2", + "[8] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[9] Ted Lentsch, Zimin Xia, Holger Caesar, and Julian FP Kooij. Slicematch: Geometry-guided aggregation for cross-view pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17225-17234, 2023. 2, 5, 6", + "[10] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2", + "[11] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 3", + "[12] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 3", + "[13] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios" + ], + "bbox": [ + 78, + 146, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Balntas. OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In CVPR, 2023. 2", + "[14] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6", + "[15] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation estimation by cross-view matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064-4072, 2020. 1, 2, 6", + "[16] Yujiao Shi, Xin Yu, Liu Liu, Tong Zhang, and Hongdong Li. Optimal feature transport for cross-view image geolocation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11990-11997, 2020. 2", + "[17] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer, 2023. 2, 3, 6, 7, 8", + "[18] Shinya Sumikura, Mikiya Shibuya, and Ken Sakurada. Openvslam: A versatile visual slam framework. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2292-2295, 2019. 5", + "[19] Xiaoyang Tian, Jie Shao, Deqiang Ouyang, and Heng Tao Shen. Uav-satellite view synthesis for cross-view geolocation. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4804-4815, 2021. 1, 2", + "[20] Aysim Token, Qunjie Zhou, Maxim Maximov, and Laura Leal-Taixe. Coming down to earth: Satellite-to-street view synthesis for geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2021. 1, 2", + "[21] Frank Van Diggelen and Per Enge. The world's first gps mooc and worldwide laboratory using smartphones. In Proceedings of the 28th international technical meeting of the satellite division of the institute of navigation (ION GNSS+ 2015), pages 361-369, 2015. 2", + "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[23] Shan Wang, Chuong Nguyen, Jiawei Liu, Kaihao Zhang, Wenhan Luo, Yanhao Zhang, Sundaram Muthu, Fahira Afzal Maken, and Hongdong Li. Homography guided temporal fusion for road line and marking segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1075-1085, 2023. 1", + "[24] Shan Wang, Yanhao Zhang, Akhil Perincherry, Ankit Vora, and Hongdong Li. View consistent purification for accurate cross-view localization, 2023. 1, 2, 3, 5, 6, 7, 8", + "[25] Shan Wang, Yanhao Zhang, Ankit Vora, Akhil Perincherry, and Hengdong Li. Satellite image based cross-view localization for autonomous vehicle. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3592-3599. IEEE, 2023. 1, 2, 5, 6, 7", + "[26] Xiaolong Wang, Runsen Xu, Zuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14851", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a correlation-aware homography estimator. arXiv preprint arXiv:2308.16906, 2023. 1, 2, 3, 6", + "[27] Zimin Xia, Olaf Booij, Marco Manfredi, and Julian FP Kooij. Visual cross-view metric localization with dense uncertainty estimates. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 90-106. Springer, 2022. 2", + "[28] Zimin Xia, Olaf Booij, and Julian FP Kooij. Convolutional cross-view pose estimation. arXiv preprint arXiv:2303.05915, 2023. 2, 5, 6, 7", + "[29] Zelong Zeng, Zheng Wang, Fan Yang, and Shin'ichi Satoh. Geo-localization via ground-to-satellite cross-view image retrieval. IEEE Transactions on Multimedia, 25:2176-2188, 2023. 1, 2", + "[30] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021. 1, 2", + "[31] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1162-1171, 2022. 1, 2" + ], + "bbox": [ + 78, + 92, + 468, + 428 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14852", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_model.json b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe44e7ca70115c65c812ea4b696bb6c81cd12959 --- /dev/null +++ b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "View From Above: Orthogonal-View aware Cross-view Localization" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.18, + 0.793, + 0.234 + ], + "angle": 0, + "content": "Shan Wang1,2 Chuong Nguyen1 Jiawei Liu2 Yanhao Zhang2 Sundaram Muthu1 Fahira Afzal Make1 Kaihao Zhang2 Hongdong Li2 1Data61, CSIRO 2Australian National University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.695 + ], + "angle": 0, + "content": "This paper presents a novel aerial-to-ground feature aggregation strategy, tailored for the task of cross-view image-based geo-localization. Conventional vision-based methods heavily rely on matching ground-view image features with a pre-recorded image database, often through establishing planar homography correspondences via a planar ground assumption. As such, they tend to ignore features that are off-ground and not suited for handling visual occlusions, leading to unreliable localization in challenging scenarios. We propose a Top-to-Ground Aggregation (T2GA) module that capitalizes aerial orthographic views to aggregate features down to the ground level, leveraging reliable off-ground information to improve feature alignment. Furthermore, we introduce a Cycle Domain Adaptation (CycDA) loss that ensures feature extraction robustness across domain changes. Additionally, an Equidistant Re-projection (ERP) loss is introduced to equalize the impact of all keypoints on orientation error, leading to a more extended distribution of keypoints which benefits orientation estimation. On both KITTI and Ford Multi-AV datasets, our method consistently achieves the lowest mean longitudinal and lateral translations across different settings and obtains the smallest orientation error when the initial pose is less accurate, a more challenging setting. Further, it can complete an entire route through continual vehicle pose estimation with initial vehicle pose given only at the starting point." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.21, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Visual-based cross-view localization aims to locate query images taken from street-level cameras (referred to as ground view) within a satellite or aerial view map. Platforms like the Google Map API [6] have made satellite images accessible, spurring the development of cross-view localization techniques. Notable studies [4, 7, 10, 15, 20, 25, 26, 30] have focused on using satellite images for this task. Despite this, accurate localization remains challenging due" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.756, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.374, + 0.671, + 0.385 + ], + "angle": 0, + "content": "Ground View" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.269, + 0.892, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.374, + 0.86, + 0.385 + ], + "angle": 0, + "content": "Aerial View" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Efficacy of T2GA. T2GA aggregates off-ground features (e.g., the streetlight within the green square) and addresses occlusions (e.g., the tree within the red square). Unlike conventional methods that often overlook such elements, T2GA integrates these features, thus improving appearance matching for ground-level aligned pixels across views using ground plane homography. Insets are magnified for clarity, with dotted outlines indicating appearance misalignment and solid outlines indicating appearance alignment with the corresponding aerial view features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.894, + 0.596 + ], + "angle": 0, + "content": "to the significant viewpoint differences between ground and aerial images [14, 31]. These viewpoint differences result in a domain gap, adversely impacting feature alignment, thereby compromising the overall accuracy of localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recent research has explored two main approaches to bridge the domain gap in image-based localization: generative-based and geometry-alignment-based methods. Generative-based methods, such as those utilizing GANs [19] and diffusion models [29], reduce the domain gap by transforming view styles from one view to another. However, the generated features for matching can lead to ambiguities in pose estimation. In contrast, geometry-alignment-based methods, employing techniques like polar transformations [15, 20] or homography [14, 23, 24, 26], focus on establishing correspondences for on-ground pixels. This often results in the neglect of off-ground features, such as streetlights, and difficulty in handling visual occlusions, such as the obscuring of road details by treetops in aerial views, both illustrated in Fig. 1. This neglect fails to leverage important geographic landmarks on the road and results in a lack of robustness to issues like road mark degradation (e.g. fading and damaged paintings). To address these limitations, we introduce the Top-to-Ground Aggregation (T2GA) module. T2GA employs top-down feature aggre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.899 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at https://github.com/ShanWang-Shan/ViewFromAbove." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14843" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "gation to enrich on-ground points with an above-view perspective appearance, significantly improving feature alignment and localization accuracy. Furthermore, we recognize that pixel positions above ground points may not necessarily indicate higher elevations; instead, they could represent objects located further along the camera's line of sight. To resolve this ambiguity, we integrate a transformer mechanism that assesses whether these pixels belong to the same object and determines occlusion precedence, such as shadows on the road, which suggest likely occlusions in the aerial view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.395 + ], + "angle": 0, + "content": "In addition to viewpoint differences, ground and aerial views differ in cameras types, lighting conditions, tones, and resolutions. We introduce the Cycle Domain Adaptation (CycDA) loss function to address these variations. CycDA enables bidirectional feature generation between ground and aerial views. By minimizing the discrepancy between domain-adapted features and their target counterparts, our approach ensures that features extracted from one domain are effectively translatable to the other, fostering the extraction of features that are invariant across domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.473, + 0.533 + ], + "angle": 0, + "content": "We further introduce an Equidistant Re-projection Loss to address a common bias in keypoint-synchronous detection localization methods, which tend to favor closer keypoints due to their reduced orientation errors. Our loss function mitigates this issue by applying a distance-weighted approach, ensuring that orientation errors are independent of keypoint distance. Consequently, keypoints are more uniformly distributed across various distances, leading to a more equitable and precise estimation of direction." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.535, + 0.39, + 0.549 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.552, + 0.469, + 0.597 + ], + "angle": 0, + "content": "- A top-to-ground feature aggregation module that effectively bridges the domain gap between ground and aerial views by utilizing off-road cues and handling occlusions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.598, + 0.469, + 0.641 + ], + "angle": 0, + "content": "- A cycle domain adaptation loss function that promotes domain-invariant feature extraction, enhancing the robustness of cross-view localization methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.643, + 0.469, + 0.702 + ], + "angle": 0, + "content": "- An equidistant re-projection loss function ensures orientation errors are consistent regardless of keypoint distance, leading to a more extended distribution of keypoints and more accurate orientation estimation." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.552, + 0.469, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The evaluations are conducted on KITTI and Ford Multi-AV datasets while leveraging the Google Maps as corresponding satellite images. Experimental results demonstrate that our method can achieve consistent vehicle pose estimation under various challenging situations. Our method achieves the lowest mean longitudinal and lateral translation errors on both KITTI and Ford Multi-AV datasets under different settings. We also obtain the smallest orientation error when the initial pose is less accurate, demonstrating the robustness of our method. Further, with initial vehicle pose only at the starting point, our method can complete the route through continuous pose estimation, demonstrating its generalisation ability and potential for practical deployment." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Image-level cross-view localization. Early visual-only cross-view localization methods [7, 10, 15, 16, 20, 30] approach the task as an image retrieval problem, focusing on coarse localization through image-to-image matching. To bridge the domain gap between ground and aerial views, various techniques have been proposed to facilitate cross-view feature matching. [10] incorporated per-pixel orientation information, while [15] and [20] utilized a predefined polar transform to align aerial-view images with ground views. Additionally, [19] and [29] employed GAN-based and diffusion model-based style transformations, respectively. While efforts have been made to minimize the domain gap, these methods rely on image-level feature matching, restricting their localization accuracy, often falling short of the precision achieved by commercial GPS systems in open areas [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Patch-wise cross-view localization. To improve accuracy, several methods employ patch-wise feature matching, especially effective in aerial views with wide fields of view and high resolutions. For instance, [31] employs transformers to emphasize informative patches, [27, 28] computes dense spatial distributions using patch attention, and [9] introduces the 'slice-to-sector match' and 'Cross-view attention' to calculate similarity between aerial sectors and ground view slices. However, their attention queries are derived from aerial images, posing reliability issues when the pose is unknown. OrienterNet [13] transforms ground view images into Bird's Eye View (BEV) grids for matching with OpenStreetMap data. Despite these advancements, localization accuracy remains limited by the patch (grid) size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Pixel-wise cross-view localization. Pixel-wise feature matching has been explored in various methods [4, 14, 24, 25] for precise localization, yet they struggle with the inherent domain gap between views. Our baseline, PureACL[24], and other geometry-alignment-based methods [14, 26] prioritize on-ground pixel correspondences but overlook off-ground features and occlusions, leading to suboptimal performance. CVGL [4] transforms the ground view image into a BEV for aerial matching. Their BEV transformation employs aerial coordinates as queries, introducing a high degree of freedom and increasing the complexity of matching. In the BoostAcc [17] framework, homography transformation is used on ground views to source query pixel data for the pixel-to-slice attention mechanism. This transformation introduces distortions from non-ground pixels, potentially exacerbating errors in subsequent processing stages. Moreover, the keys and values span entire and adjacent columns, risking compounded ambiguities in both longitudinal and lateral estimation. Our approach aims to address these challenges, proposing a novel pixel-to-overhead-column attention mechanism for improved cross-view localization accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Figure 2. Our method adopts PureACL [24] as the baseline model and introduces three novel components (highlighted in red): (1) T2GA (Sec.4.1): aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them to alleviate the representation gap of the same object across different views; (2) CycDA Loss (Sec.4.2): explicitly enforces a view-invariant representation for the same object; and (3) ERP Loss (Sec.4.3): allows the model to leverage key points that are farther away from the vehicle while allocating more emphasis on correcting the vehicle orientation estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.292, + 0.204, + 0.31 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.422 + ], + "angle": 0, + "content": "Task Settings. The Fine-Grained Cross-View Localization (FGCVL) task aims to estimate the accurate 3-DoF pose of a vehicle, including longitudinal and lateral translations and orientation on a satellite map \\( I^{s} \\), given an initial coarse pose \\( \\mathbf{P}_{init} \\). It also assumes that the vehicle is equipped with on-board camera(s) \\( ^2 \\), and the image captured by these cameras, denoted as the ground image \\( I^{g} \\), is accessible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.667 + ], + "angle": 0, + "content": "Baseline Model. PureACL [24] is adopted as the baseline model for our approach. It employs a share-weight U-Net to extract ground/satellite feature maps \\( F = \\{F^g, F^s\\} \\) and ground/satellite confidence maps \\( C = \\{C^g, C^s\\} \\) from the ground/satellite images \\( I = \\{I^g, I^s\\} \\). The ground confidence map \\( C^g \\) is tasked to select the top-N most confident pixel-level features from \\( F^g \\) at the positions \\( p^g = \\{(u_i^g, v_i^g)\\}_{i=1}^N \\). These features are then: (i) Projected from on-board camera coordinates to the satellite coordinates using \\( p_i^s(\\mathbf{P}) = K_s \\mathbf{P} K_g^{-1} p_i^{g,4} \\), where \\( K_s \\) and \\( K_g \\) are the intrinsic matrices of satellite and on-board cameras, respectively. \\( \\mathbf{P} \\) is the initial pose \\( \\mathbf{P}_{init} \\) at the first iteration and the estimated pose \\( \\mathbf{P}_{pred} \\) in the subsequent iterations. (ii) A weighted residual \\( w(\\mathbf{P}) \\times r(\\mathbf{P}) \\) is computed between the top-N selected features from the ground feature map and the corresponding features from the satellite feature map as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.675, + 0.47, + 0.693 + ], + "angle": 0, + "content": "\\[\nw (\\mathbf {P}) = C ^ {s} \\left[ p ^ {s} (\\mathbf {P}) \\right] \\times C ^ {g} \\left[ p ^ {g} \\right], r (\\mathbf {P}) = F ^ {s} \\left[ p ^ {s} (\\mathbf {P}) \\right] - F ^ {g} \\left[ p ^ {g} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.701, + 0.471, + 0.795 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\) is a lookup operation in a feature/confidence map with sub-pixel interpolation. (iii) The Levenberg-Marquardt (LM) algorithm [11] takes the computed weighted residual \\(w(\\mathbf{P})\\times r(\\mathbf{P})\\) as input and outputs the predicted vehicle pose \\(\\mathbf{P}_{pred}^{m}\\) at the \\(m^{\\mathrm{th}}\\) iteration, with \\(m\\in\\) \\(\\{1,\\ldots ,M\\}\\). The process of (i) - (iii) repeats for \\(M = 20\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.803, + 0.468, + 0.827 + ], + "angle": 0, + "content": "2While our method is compatible with multiple cameras, our primary experiments focus on one front camera and four surrounding cameras." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.827, + 0.468, + 0.852 + ], + "angle": 0, + "content": "3We simply refer to the fused confidence map, rather than distinguishing between view-consistent and on-ground confidence maps in PureACL." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.852, + 0.469, + 0.9 + ], + "angle": 0, + "content": "PureACL obtains 3D world coordinates of the vehicle from homogeneous coordinates \\( K_{g}^{-1}p^{g} \\) by utilizing on-ground characteristics. For simplicity in this context, we note homogeneous coordinates as directly corresponding to the 3D world coordinates." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.803, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.294, + 0.829, + 0.309 + ], + "angle": 0, + "content": "iterations to produce the final vehicle pose \\(\\mathbf{P}_{pred}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.416 + ], + "angle": 0, + "content": "PureACL [24] employs two loss functions: a Triplet Loss [12] and a Re-projection Loss. The former supervises the weighted residual, enforcing that the encoder extracts pose-sensitive features from both ground and satellite images. The later utilises the projection error w.r.t the ground truth pose \\(\\mathbf{P}_{gt}\\) to penalize incorrectly predicted vehicle pose \\(\\mathbf{P}_{pred}\\) (Refer to Supplementary Sec.D for more details)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.591, + 0.447 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.579 + ], + "angle": 0, + "content": "We propose a novel View From Above (VFA) method to tackle the FGCVL task by aligning feature representations across the orthogonal viewpoints, the ground view and the satellite views. As illustrated in Fig. 2, our method introduces three innovative components onto the baseline model: (1) Top-to-Ground Aggregation Module (T2GA) (Sec. 4.1); (2) Cycle Domain Adaptation Loss (CycDA) (Sec. 4.2); and (3) Equidistant Re-Projection Loss (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.75, + 0.607 + ], + "angle": 0, + "content": "4.1. Top to Ground Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The appearance discrepancy between the same objects viewed from ground and satellite perspectives presents a significant challenge in achieving high precision in the FGCVL task [14, 17, 24, 26]. This issue is often encountered on tall structures, e.g. street light, and the occlusion from aerial objects, e.g. tree branches, where their near(on)-ground appearance, accessible to only on-board vehicle camera, is significantly different from their top appearances which are only registered by the satellite camera (see Fig. 1). Such discrepancies are common on roads and can lead to inaccuracies in cross-view localization. To address this challenge, we propose aligning the representations between ground and satellite views, focusing specifically on the on-ground pixels, where geometry alignment across the two views hold due to the ground plane homography. Specifically, we present a T2GA module that aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them, as illustrated in the 'T2GA' red box in Fig. 2. For an on-ground pixel:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14845" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.088, + 0.473, + 0.198 + ], + "angle": 0, + "content": "\\(p\\in \\{(u,v)\\}_{u = 0,v = \\tau}^{W^g -1,H^g -1}\\) , and where \\(\\tau\\) is a threshold corresponding to the height of horizon in the ground view image, and \\(W^{g}\\) and \\(H^{g}\\) are the width and height of the ground view image, respectively. The corresponding elevated pixels are \\(q\\in \\{(u_p,v)\\}_{v = 0}^{v_p}\\) , where \\((u_{p},v_{p})\\) are the coordinates of the on-ground pixel \\(p\\) . The aggregation is achieved via the attention mechanism [22], defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.209, + 0.47, + 0.227 + ], + "angle": 0, + "content": "\\[\nF ^ {a t t} [ p ] = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {T}\\right) \\mathbf {V}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.473, + 0.481 + ], + "angle": 0, + "content": "where \\( F^{att}[p] \\) represents the feature vector at a pixel \\( p \\) on the feature map \\( F^{att} \\). In this context, \\( p \\) and \\( q \\) denote an on-ground pixel and its corresponding elevated pixels, respectively. The query is formed by taking the features of on-ground pixels \\( \\mathbf{Q} = \\mathcal{M}(F^g[p]) \\) where \\( \\mathcal{M}(\\cdot) \\) represents a non-linear mapping consisting of an MLP layer followed by an activation function. The value is derived from the features of elevated pixels corresponding to the on-ground pixel \\( \\mathbf{V} = F^g[q] \\), and the key is then obtained by applying the same non-linear mapping to \\( \\mathbf{V} \\), resulting in \\( \\mathbf{K} = \\mathcal{M}(\\mathbf{V}) \\). This non-linear mapping ensures that the query and key are mapped to the same feature space before computing their matrix product. \\( \\mathbf{Q}\\mathbf{K}^T \\). Notably, value \\( \\mathbf{V} \\) is used directly in its original form for the computation, facilitating a straightforward fusion (for the same object) or replacement (in cases of occlusion) at ground level." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.495, + 0.478, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.766 + ], + "angle": 0, + "content": "Figure 3. Attention between the on-ground pixels and their corresponding elevated pixels is displayed column-wise. (Left example): The attention between the base and top of the streetlight is high despite their distinct appearances. This allows the resultant aggregated ground feature to be aligned with the satellite feature corresponding to the matching geographic location. (Right example): In the absence of occlusion from above, ground pixels maintain minimal attention with their elevated pixels. This avoids unnecessary dilution of the ground features which are already well aligned with their satellite counterparts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.471, + 0.839 + ], + "angle": 0, + "content": "The computed attention map \\(\\operatorname{Softmax}(\\mathbf{Q}\\mathbf{K}^T)\\) can accurately reflect the connection between the on-ground pixels and their corresponding elevated pixels. As illustrated in Fig. 3, high attention values are assigned to the base and top" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.847, + 0.47, + 0.904 + ], + "angle": 0, + "content": "In processing all query pixels, we first compute the matrix multiplication \\(\\mathbf{Q}\\mathbf{K}^T\\) using all column pixels \\(q\\in \\{(u_p,v)\\}_{v = 0}^{H^g -1}\\). We then eliminate the attention values corresponding to pixels below each query pixel \\(\\{(u_p,v)\\}_{v = v_p + 1}^{H^g -1}\\). After this elimination, the Softmax function is applied." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "of the streetlight despite their different appearances. This provides important cues to alleviate the representation gap between the ground feature and satellite features. On the other hand, in the absence of occlusion, e.g. tall structure and tree branch, thus the attention between on-ground and elevated pixels is minimal, avoiding diluting the ground features that are well aligned with their satellite counterparts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.243 + ], + "angle": 0, + "content": "The feature map \\( F^{att} \\) after attention calculation is vertically stacked with the upper part of the original \\( F^g \\) to generate the aggregated feature map \\( F^a \\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.255, + 0.892, + 0.273 + ], + "angle": 0, + "content": "\\[\nF ^ {a} = F _ {0: H ^ {g} - \\tau} ^ {g} \\oplus F ^ {a t t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.892, + 0.359 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes vertically stacking and \\(F_{0:H^g -\\tau}^g\\) indicated the slicing of \\(F^{g}\\) from row 0 to \\(H^{g} - \\tau\\) . Subsequently, this aggregated feature map \\(F^a\\) replaces \\(F^g\\) in the residual calculation as per Eq.1, resulting in an updated formula: \\(r(\\mathbf{P}) = F^{s}[p^{s}(\\mathbf{P})] - F^{a}[p^{g}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.893, + 0.587 + ], + "angle": 0, + "content": "The advantage of the aggregated feature map is also indirectly reflected by the ground confidence map of the baseline model as shown in Fig. 4. Without the T2GA module, the confidence map only highlights pixels corresponding to road marks and curds, resulting in a concentrated sampling of key points for subsequent estimation of vehicle pose. This has the following drawbacks: (1) fail to leverage road landmarks e.g. traffic signal poles, that encode important geographic location; and (2) not robust to road mark degradation problems, e.g. fading and damaged paintings, and visibility issues, e.g. glow on road marks. T2GA alleviates the feature discrepancy between ground and satellite views. As a result, the traffic signal pole receives higher confidence, allowing it to be subsequently sampled as key points to estimate more precise vehicle pose." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.599, + 0.702, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.599, + 0.899, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Figure 4. Illustration on the effect of T2GA on the confidence map of the baseline model. The confidence map without T2GA (Left) predominantly highlights road marks and curbs, resulting in subsequent keypoint sampling missing important road landmarks, e.g. traffic signal poles, that provide important cues to vehicle pose estimation; The confidence map with T2GA (Right) has high-confidence values distributed across various road marks and traffic poles. With more geographic cues provided by multiple sources, the resultant pose prediction becomes more precise and robust." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.776, + 0.818 + ], + "angle": 0, + "content": "4.2. Cycle Domain Adaptation Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "There are multiple sources responsible for the representation gap between ground and satellite features. In addition to the varying appearances of the same object across different views, camera specifications, e.g. tone, hue, intensity, brightness and resolution, and temporal changes of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14846" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.211 + ], + "angle": 0, + "content": "varying spans, can also lead to inconsistent representations for the same geographic location. Such issue is in general overlooked by the existing FGCVL methods [14, 28]. Despite introducing a triplet loss to further distinguish feature representations based on different poses, our baseline model, PureACL [24], falls short in enforcing invariant feature representation at the corresponding geographic locations across different views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.288 + ], + "angle": 0, + "content": "We propose a Cycle Domain Adaptation (CycDA) loss to explicitly enforce view-invariant representations that consists of three L2-loss between ground and satellite feature representations in three different feature spaces. The representation loss in the ground feature space is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.309, + 0.47, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| F ^ {a} \\left[ p _ {i} ^ {g} \\right] - \\mathcal {D} \\left(\\mathcal {E} \\left(F ^ {s} \\left[ p _ {i} ^ {s} \\left(\\mathbf {P} _ {g t}\\right) \\right] \\odot c _ {s}\\right) \\odot c _ {g}\\right) \\right\\| _ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.476 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}(\\cdot)\\) is a projection from the pixel feature space to a latent feature space, \\(\\mathcal{D}(\\cdot)\\) is a projection from the latent feature space to the pixel feature space, \\(c_{s}\\) and \\(c_{g}\\) represent the focal length of satellite camera and vehicle camera respectively which are introduced to guide the projection functions, and \\(\\odot\\) is a concatenation acting on the feature channel dimension. Similarly, the representation loss in the satellite feature space is:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.484, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| F ^ {s} \\left[ p _ {i} ^ {s} \\left(\\mathbf {P} _ {g t}\\right) \\right] - \\mathcal {D} \\left(\\mathcal {E} \\left(F ^ {a} \\left[ p _ {i} ^ {g} \\right] \\odot c _ {g}\\right) \\odot c _ {s}\\right) \\right\\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.575 + ], + "angle": 0, + "content": "We also enforce the satellite feature to be aligned with the ground feature corresponding to the same geographic location in a latent feature space through:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.583, + 0.47, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| \\mathcal {E} \\left(F ^ {a} [ p _ {i} ^ {g} ] @ c _ {g}\\right) - \\mathcal {E} \\left(F ^ {s} [ p _ {i} ^ {s} (\\mathbf {P} _ {g t}) ] @ c _ {s}\\right) \\right\\| _ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.627, + 0.392, + 0.643 + ], + "angle": 0, + "content": "The overall CycDA loss can then be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.654, + 0.469, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C y c D A} = \\mathcal {L} _ {g} + \\mathcal {L} _ {s} + \\mathcal {L} _ {m} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.354, + 0.697 + ], + "angle": 0, + "content": "4.3. Equidistant Re-projection Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Our method tackles a challenging setting of simultaneously detecting key points from the ground image and predicting the vehicle pose using the detected key points. Our baseline model introduces a re-projection loss computed over the selected top-N key points defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.789, + 0.47, + 0.831 + ], + "angle": 0, + "content": "\\[\nL _ {R P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| p _ {i} ^ {s} (\\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\\mathbf {P} _ {g t}) \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( p_i^s (\\mathbf{P}_{pred}) \\) and \\( p_i^s (\\mathbf{P}_{gt}) \\) are the projected coordinates on the satellite image from the selected top-N ground key points with the predicted vehicle pose and the groundtruth vehicle pose respectively, and \\( \\| \\cdot \\|_2 \\) denotes the L2-distance." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.7, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.09, + 0.899, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.15, + 0.894, + 0.206 + ], + "angle": 0, + "content": "Figure 5. Comparison of detected keypoints with and without ERP Loss. (Left) Without the ERP Loss, keypoints are predominantly located in close proximity to the vehicle. (Right) With the ERP Loss, there is a more dispersed distribution of keypoints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Despite being effective in pipelines where simultaneous detection of key points is not required [3, 18, 25], the reprojection loss tends to overly penalize the orientation errors of distant points. This forces the detected key points being closer to the vehicle, as illustrated in Fig. 5, and less accurate vehicle orientation estimation (see Sec. 5.2)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.334 + ], + "angle": 0, + "content": "To mitigate the aforementioned issue, we propose an Equidistant Re-Projection (ERP) loss that is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "\\[\nL _ {E R P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\frac {p _ {i} ^ {s} (\\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\\mathbf {P} _ {g t})}{\\mathbf {D} _ {p _ {i} ^ {s}}} \\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.893, + 0.516 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}_{p_i^s} = \\| p_i^s (\\mathbf{P}_{gt}) - p_{vehicle}^s (\\mathbf{P}_{gt})\\|_2\\) is the L2-distance between the satellite image coordinates of the \\(i^{\\mathrm{th}}\\) keypoint and the vehicle. The design prevents the orientation error from scaling with the L2-distance between key points and vehicle on the satellite image, which dominates over the orientation error on distant key points. As shown in Fig. 5, our method can leverage key points that are farther from the vehicle than our baseline model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.633, + 0.547 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.554, + 0.717, + 0.57 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Datasets. To evaluate our proposed approach, we conduct experiments on two well-established autonomous driving datasets: the Ford Multi-AV Seasonal dataset (FMAVS) [1] and the KITTI dataset [5]. Consistent with established practices [14, 24, 25], our primary focus is on images from the front left camera as query images. Additionally, following the methodology of PureACL [24], we broadened our analysis to include four camera perspectives: front left, rear right, side left, and side right. For data partitioning in the KITTI-CVL dataset, we utilized an approach in line with HighlyAcc [14], which involves two test splits. The first, 'Same,' includes images from the same trajectories as the training dataset, while the second, 'Cross,' comprises images from distinct trajectories. Regarding the FMAVS dataset, we follow the partitioning strategy proposed by PureACL [24], using all images from the same trajectories but different traversals. It is noteworthy that there are minor route variations within these same trajectories6" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.849, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Metrics. We adopt evaluation metrics in line with HighlyAcc [14], SliceMatch [9], SIBCL [25], and PureACL [24]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.786, + 0.901 + ], + "angle": 0, + "content": "An example is provided in Supplementary Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14847" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 1. Comparison on the KITTI-CVL Dataset Under Initial Noise Conditions (±10°,±20m). Results marked with \\(\\star\\) are sourced directly from the respective original papers. Those denoted by \\(\\diamond\\) indicate retraining of methods for consistent evaluation criteria alignment. - indicates that the corresponding data were not provided." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.133, + 0.895, + 0.409 + ], + "angle": 0, + "content": "
ModelLocation(m)↓Lateral(%)↑Longit(%)↑Orient(°)↓Orient(%)↑
MeanMedianr@1mr@3mr@5mr@1mr@3mr@5mMeanMedianr@1°r@3°r@5°
Same★DSM[15]--10.1230.6748.244.0812.0120.14--3.5813.8124.44
★HighlyAcc[14]12.0811.4235.5470.7780.365.2215.8826.133.722.8319.6451.7671.72
★SliceMatch[9]7.964.3949.09-98.5215.19-57.354.123.6513.41-64.17
★BoostAcc[17]10.015.1976.4496.3498.8923.5450.5762.180.550.4299.10100.00100.00
★CCVPE [28]1.220.6297.3598.6599.7177.1396.0897.160.670.5477.3999.4799.95
★HC-Net [26]0.800.5099.01-99.7392.20-99.250.450.3391.35-99.84
◇PureACL [24]2.420.4291.9593.4094.2891.8692.1292.383.971.7732.7153.8971.66
Ours0.200.1799.9799.9799.9799.9799.9799.971.530.9350.9583.9795.45
Cross★DSM[15]--10.7731.3748.243.8711.7319.50--3.5314.0923.95
★HighlyAcc[14]12.5812.1127.8259.7972.895.7516.3626.483.953.0318.4249.7271.00
★SliceMatch[9]13.509.7732.43-86.448.30-35.574.206.6146.82-46.82
★BoostAcc[17]13.019.0657.7286.7791.1614.1534.5945.000.560.4398.98100.00100.00
★CCVPE [28]9.163.3344.0681.7290.2323.0852.8564.311.550.8457.7292.3496.19
★HC-Net [26]8.474.5775.00-97.7658.93-76.463.221.6333.58-83.78
◇PureACL [24]6.200.6167.2487.2692.7664.8173.6389.694.262.4823.4548.6959.24
Ours0.210.1899.9299.9699.9799.9199.9199.912.041.3838.3180.1292.48
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Our precision assessment includes median and mean errors for overall, lateral, and longitudinal translations, as well as orientation accuracy. Additionally, our analysis covers lateral and longitudinal translations and localization recall at various distances (0.25m, 0.5m, 1m, 3m, and 5m), and orientation recall within a range of \\(1^{\\circ}\\) to \\(5^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Training Details. We apply two sets of criteria: (1) Following HighlyAcc [14], ground images are processed at a resolution of \\(256 \\times 1024\\), and satellite images at \\(512 \\times 512\\), with initial pose noise \\(\\pm 10^{\\circ}\\) for orientation and \\(\\pm 20\\mathrm{m}\\) for translations. (2) In accordance with PureACL [24], we process FMAVS images at \\(432 \\times 816\\), KITTI images at \\(384 \\times 1248\\), and satellite images at \\(1,280 \\times 1,280\\). To better simulate turning scenarios, we adopt an expanded initial pose noise range of \\(\\pm 45^{\\circ}\\) for orientation and \\(\\pm 20\\mathrm{m}\\) for translations. For training, we utilize an NVIDIA RTX 3090 GPU with a batch size of 3, employing the Adam optimizer [8] with a learning rate of \\(10^{-5}\\). Feature extractor weights are adapted from PureACL [24], while other components are initialized randomly. Training iterations average around \\(285\\mathrm{ms}\\), including \\(200\\mathrm{ms}\\) dedicated to optimization. The inference speed, subject to initial pose variability, averages at \\(222\\mathrm{ms}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.388, + 0.773 + ], + "angle": 0, + "content": "5.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method against recent visual-only approaches on the KITTI-CVL dataset following the metrics of HighlyAcc [14]. The results shown in Tab. 1 clearly indicate the superiority of our approach in spatial precision, consistently maintaining poses within a \\(1\\mathrm{m}\\) radius in both 'Same' and 'Cross' areas with over \\(99.9 + \\%\\) probability. While our method may not lead in orientation accuracy under an initial pose range of \\(\\pm 10^{\\circ}\\), it demonstrates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.78 + ], + "angle": 0, + "content": "superior performance when the range is extended to \\(\\pm 45^{\\circ}\\), as detailed in Tab. 2. This underlines the robustness of our approach. Our approach leverages pixel-wise localization, offering finer granularity compared to the patchwise localization of SliceMatch[9], CCVPE [28], and the image-level localization of DSM [15]. Although HighlyAcc [14] and HC-Net [26] are pixel-wise localization methods, their homography-based mechanisms fail to incorporate off-road cues. In contrast, our approach effectively utilizes these cues, resulting in enhanced performance. Additionally, our longitudinal estimation significantly surpasses that of BoostAcc [17]. This improvement is likely attributable to our method's strategy of not aggregating information from pixels below, thereby reducing longitudinal ambiguity. In comparison to the baseline method PureACL [24], our approach achieves significant enhancements in both orientation and spatial accuracy, especially in mean error reduction. Specifically, we observe a \\(91^{+}\\%\\) reduction in spatial mean error and a \\(52^{+}\\%\\) reduction in orientation mean error. The difference between mean and median errors for PureACL indicates potential convergence issues in high-noise environments. By integrating the T2GA and CycDA modules, we bridge the domain gap and ensure convergence even under challenging conditions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.878, + 0.803 + ], + "angle": 0, + "content": "5.3. Challenge Initial Pose and Stringent Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To rigorously evaluate our method, we adopt the stringent metrics of PureACL [24], which are more demanding than those used by HighlyAcc[14]. We benchmark our method against baseline methods SIBCL [25] and PureACL [24], as well as SOTA methods CCVPE [28] and BoostAcc [17]. The results in Tab. 2 reaffirm the advantages of our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14848" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 2. Comparison with Initial Noise Conditions \\((\\pm 45^{\\circ},\\pm 20\\mathrm{m})\\) .For the KITTI-CVL dataset, evaluations are conducted in 'K-Same' and 'K-Cross' areas. On the Ford-CVL dataset, the Log4' trajectory, as used in SIBCL [25], is chosen for its optimal satellite view alignment. Additional log evaluations are detailed in the supplementary material. F-1C' and F-4C' represent assessments on the Ford-CVL dataset using either a single front-facing camera or four surrounding cameras. Note: \\(\\star\\) indicates that SIBCL [25] is a hybrid LiDAR-visual method." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.147, + 0.895, + 0.465 + ], + "angle": 0, + "content": "
ModelLateral(m)↓Mean MedianLateral(%)↑r@0.25mr@0.5mr@1mLongit(m)↓Mean MedianLongit(%)↑r@0.25mr@0.5mr@1mOrient(°)↓Mean MedianOrient(%)↑r@1°r@2°r@4°
K-SameCCVPE [28]2.321.2211.0522.0542.096.493.357.4513.3624.973.431.5334.6660.7284.03
BoostAcc[17]1.190.6321.9241.5167.7210.015.425.179.3816.543.882.9818.4535.2562.20
★ SIBCL[25]1.230.7017.0133.3461.883.140.8620.6831.4055.588.901.6935.5055.0471.06
PureACL[24]2.940.2358.7566.3470.603.150.3745.6568.2374.878.132.8037.9952.5764.13
Ours0.170.1476.2097.7399.970.080.0798.1899.8999.972.191.0748.9766.2990.54
K-CrossCCVPE [28]4.562.794.9610.0538.1212.068.163.136.1412.4519.4615.393.3810.2417.41
BoostAcc[17]2.650.9314.8829.2052.6410.468.923.385.5611.735.614.2612.4624.8747.26
★ SIBCL[25]2.720.7116.9032.6258.245.481.0519.5830.7449.259.671.9625.0250.1860.16
PureACL [24]4.240.2550.2562.6364.563.890.447.4162.0464.2310.963.1619.9536.2556.49
Ours0.170.1575.9697.4799.970.090.0797.4999.8799.972.881.4436.1063.0886.18
F-1CCCVPE [28]4.622.456.4812.5424.4610.218.413.228.2413.5320.4815.073.146.1416.14
BoostAcc[17]2.731.479.0518.1835.5910.516.144.218.1315.566.824.7911.6122.6442.63
★ SIBCL[25]2.590.7120.9241.5360.435.381.1812.5923.9943.876.341.5735.2158.4170.65
PureACL[24]2.941.5411.3421.1437.764.751.7410.4620.0737.647.383.3315.1230.4257.01
Ours0.490.4131.7759.1789.230.360.3238.4375.7297.462.191.0847.3172.7190.10
F-4CCVGL[4]1.280.5021.8150.0586.841.530.7019.3637.2665.51-----
PureACL[24]1.690.6721.0739.5365.273.381.1812.5923.9943.863.071.1943.8473.2782.41
Ours0.130.1088.2598.53100.000.200.1868.6796.68100.002.001.4935.7262.7091.83
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.484, + 0.273, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.569, + 0.211, + 0.58 + ], + "angle": 0, + "content": "(a) Location" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.484, + 0.467, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.568, + 0.412, + 0.58 + ], + "angle": 0, + "content": "(b) Orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Figure 6. Comparison of the baseline method PureACL and our proposed method under varying initial noise ranges in the 'cross' area of the KITTI-CVL dataset. The blue bar indicates the median value, while the red bar shows the difference between the median and mean values. Notably, our method demonstrates robust convergence, outperforming PureACL across all noise ranges, especially in larger noise scenarios where PureACL shows significant discrepancies between median and mean values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.901 + ], + "angle": 0, + "content": "method, particularly in terms of reduced mean error and improved longitudinal estimation. Notably, under conditions of minimal initial pose noise \\((\\pm 15^{\\circ}\\) and \\(\\pm 5\\mathrm{m})\\), PureACL [24] outperforms SIBCL [25]. However, when faced with more challenging conditions \\((\\pm 45^{\\circ}\\) and \\(\\pm 20\\mathrm{m})\\), PureACL [24] encounters convergence challenges, unlike our method, which demonstrates consistent superiority. This highlights the effectiveness of our proposed modifications in feature alignment. Additionally, we assess our method's convergence range against PureACL [24] in the 'Cross' area of the KITTI-CVL dataset, as depicted in Fig. 6. The red bar in the figure highlights a significant reduction in the gap between median and mean errors. This observation con" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.47, + 0.892, + 0.501 + ], + "angle": 0, + "content": "firms our method's enhanced ability to converge effectively across a wider range of noise levels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.839, + 0.528 + ], + "angle": 0, + "content": "5.4. Results with Continual Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.84 + ], + "angle": 0, + "content": "To address GPS signal loss, we adopt an accumulated pose estimation strategy that leverages initial poses based on the vehicle's previous pose estimates \\(^{7}\\). We use the model trained with initial noise allowances of \\(\\pm 45^{\\circ}\\) and \\(\\pm 20\\mathrm{m}\\). Our method's robustness is evidenced by the continuous running distance percentage presented in Tab. 3, showing successful pose chaining throughout all evaluate routes. In contrast, BoostAcc [17] and the single-camera PureACL [24] exhibit significant drift, leading to errors beyond the satellite map's coverage and causing evaluations to cease after covering less than \\(11\\%\\) of the distance. Performance comparisons on the KITTI-CVL dataset are shown in Fig. 7 \\(^{8}\\). Our approach consistently achieves accurate pose estimation or experiences only minor drifts in challenging scenarios characterized by limited localization cues or severe occlusions. In contrast, BoostAcc [17] exhibits significant drift when a moving vehicle passes by, likely due to incorrect query data from projected vehicle pixels. Furthermore, PureACL [24] appears to struggle with incorrect orientation, resulting in reversed pose estimations. These out" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.875 + ], + "angle": 0, + "content": "This pose estimation is purely frame-based and does not involve any sequence-based filtering techniques." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.891, + 0.901 + ], + "angle": 0, + "content": "8Performance comparisons on the Ford-CVL dataset is shown in Fig. 1 and Fig. 2 of supplementary material." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.101, + 0.272, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.101, + 0.468, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.213, + 0.273, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.213, + 0.468, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.436 + ], + "angle": 0, + "content": "Figure 7. Accumulated pose estimation performance on a KITTI-CVL dataset trajectory. (Top-left) Predicted poses by our method (blue) closely match the ground truth (red) over the 3676-meter route, outperforming PureACL (green) and BoostAcc (cyan), which exhibit severe drift. (Top-right) Our approach demonstrates rapid recovery upon encountering clear localization cues. (Bottom-left) (Bottom-right) BoostAcc demonstrates substantial drift when encountering moving vehicles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.501 + ], + "angle": 0, + "content": "comes highlight the importance of robustness in real-world applications, as drifts that extend beyond the coverage of satellite imagery can compromise all further pose estimation processes." + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.51, + 0.438, + 0.524 + ], + "angle": 0, + "content": "Table 3. Comparison on Accumulated Pose Estimation." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.526, + 0.458, + 0.611 + ], + "angle": 0, + "content": "
ModelRunning percentage(%)↑
K-CrossF-1CF-4C
BoostAcc[17]4.884.89-
PureACL[24]10.315.01100.00
Ours100.00100.00100.00
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.099, + 0.614, + 0.458, + 0.628 + ], + "angle": 0, + "content": "- indicates that BoostAcc does not support 4-camera setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.229, + 0.651 + ], + "angle": 0, + "content": "6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.78 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed components, we conduct ablation studies in the 'cross' area of the KITTI-CVL dataset. Our method's performance was compared across various configurations, including the presence and absence of CycDA and Equidistant Re-projection (ERP) Loss functions, and the T2GA module. The results, as detailed in Tab. 4, underscore the significant role each component plays in improving our method's performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The T2GA module aggregates information from top to ground based on the assumption that the ground and satellite views are orthogonal. However, this assumption might not be accurate in scenarios involving uphill/downhill or turns. To test our method's robustness under these conditions, we introduced affine warping to the ground view images with a shear range of \\(\\pm 15^{\\circ}\\). The outcomes, labeled 'w/ Affine' in Tab. 4, indicate our method's efficacy in handling such" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "scenarios. This success is primarily due to two factors: (1) the use of a U-Net for feature extraction allows each pixel in the feature map to encapsulate information from its surrounding area, enabling the handling of a degree of angular variation. (2) the satellite view from Google, which is not strictly orthogonal to the ground view, suggests that our method is already adapted to these types of challenges." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.206, + 0.892, + 0.233 + ], + "angle": 0, + "content": "Table 4. Ablation study on the KITTI-CVL dataset 'Cross' area with initial noise \\((\\pm 45^{\\circ},\\pm 20\\mathrm{m})\\)" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.234, + 0.887, + 0.37 + ], + "angle": 0, + "content": "
T2GA CycDA ERPLateral(m)↓Mean MedianLongit(m)↓Mean MedianOrient(°)↓Mean Median
4.240.253.890.4410.963.16
✓1.290.191.270.174.112.13
✓✓0.210.180.120.103.381.89
✓✓✓0.170.150.090.072.881.44
w/ Affine0.180.150.090.072.881.52
night0.390.270.390.306.642.55
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.487 + ], + "angle": 0, + "content": "We further evaluate our method on an artificial dataset derived from the KITTI test sets, created using the cross-domain deep network proposed by Arruda et al. [2]. The outcomes, labeled as 'night' in Tab. 4 show our method's stable performance in nighttime scenarios, highlighting its resilience to variations in lighting and time of day. For further details, please refer to Supplementary Sec. C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.501, + 0.619, + 0.517 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.8 + ], + "angle": 0, + "content": "We have presented a novel top-to-ground feature aggregation for enhancing cross-view image-based geolocation. Our method overcomes limitations by incorporating aerial perspectives and utilizing a cycle domain adaptation loss for consistent feature extraction despite visual disparities. The introduction of the equidistant re-projection loss balances keypoint impact, promoting wider distribution and thus enhancing orientation accuracy. Our method excels in vehicle pose estimation across challenging scenarios, achieving the lowest translation errors in KITTI and Ford datasets, and minimal orientation error with less accurate initial poses. Crucially, by relying solely on the initial vehicle pose at the start, our method successfully completes routes via continuous pose estimation. This paves the way for real-world applications like autonomous driving and outdoor robotics. Future work will focus on integrating this method into SLAM systems to remove loop closure dependency and facilitate high-definition map generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project is supported by CSIRO Julius Career Award, CSIRO Early Research Career Fellowship, CSIRO Future Digital Manufacturing Fund Postdoc Fellowship, CSIRO Data61 PhD Scholarship to SW, and partially by Australian Research Council Discovery Grant (DP220100800) to HL." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14850" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.121, + 0.174, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[1] Siddharth Agarwal, Ankit Vora, Gaurav Pandey, Wayne Williams, Helen Kourous, and James McBride. Ford multi-AV seasonal dataset. The International Journal of Robotics Research, 39(12):1367-1376, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.472, + 0.344 + ], + "angle": 0, + "content": "[3] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[4] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[5] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.348, + 0.502 + ], + "angle": 0, + "content": "[6] Google. Maps static api, 2023. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[7] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.472, + 0.614 + ], + "angle": 0, + "content": "[8] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[9] Ted Lentsch, Zimin Xia, Holger Caesar, and Julian FP Kooij. Slicematch: Geometry-guided aggregation for cross-view pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17225-17234, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.688, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[10] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.746, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[11] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[12] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.147, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Balntas. OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[14] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[15] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation estimation by cross-view matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064-4072, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[16] Yujiao Shi, Xin Yu, Liu Liu, Tong Zhang, and Hongdong Li. Optimal feature transport for cross-view image geolocation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11990-11997, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.894, + 0.361 + ], + "angle": 0, + "content": "[17] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[18] Shinya Sumikura, Mikiya Shibuya, and Ken Sakurada. Openvslam: A versatile visual slam framework. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2292-2295, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[19] Xiaoyang Tian, Jie Shao, Deqiang Ouyang, and Heng Tao Shen. Uav-satellite view synthesis for cross-view geolocation. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4804-4815, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.894, + 0.546 + ], + "angle": 0, + "content": "[20] Aysim Token, Qunjie Zhou, Maxim Maximov, and Laura Leal-Taixe. Coming down to earth: Satellite-to-street view synthesis for geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[21] Frank Van Diggelen and Per Enge. The world's first gps mooc and worldwide laboratory using smartphones. In Proceedings of the 28th international technical meeting of the satellite division of the institute of navigation (ION GNSS+ 2015), pages 361-369, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.894, + 0.673 + ], + "angle": 0, + "content": "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[23] Shan Wang, Chuong Nguyen, Jiawei Liu, Kaihao Zhang, Wenhan Luo, Yanhao Zhang, Sundaram Muthu, Fahira Afzal Maken, and Hongdong Li. Homography guided temporal fusion for road line and marking segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1075-1085, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[24] Shan Wang, Yanhao Zhang, Akhil Perincherry, Ankit Vora, and Hongdong Li. View consistent purification for accurate cross-view localization, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[25] Shan Wang, Yanhao Zhang, Ankit Vora, Akhil Perincherry, and Hengdong Li. Satellite image based cross-view localization for autonomous vehicle. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3592-3599. IEEE, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Xiaolong Wang, Runsen Xu, Zuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14851" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "a correlation-aware homography estimator. arXiv preprint arXiv:2308.16906, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[27] Zimin Xia, Olaf Booij, Marco Manfredi, and Julian FP Kooij. Visual cross-view metric localization with dense uncertainty estimates. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 90-106. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[28] Zimin Xia, Olaf Booij, and Julian FP Kooij. Convolutional cross-view pose estimation. arXiv preprint arXiv:2303.05915, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[29] Zelong Zeng, Zheng Wang, Fan Yang, and Shin'ichi Satoh. Geo-localization via ground-to-satellite cross-view image retrieval. IEEE Transactions on Multimedia, 25:2176-2188, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[30] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[31] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1162-1171, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14852" + } + ] +] \ No newline at end of file diff --git a/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_origin.pdf b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0bf15edaf1fed0ab01bc67be525b3d5a3a068092 --- /dev/null +++ b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/7abae4da-a8d5-4e84-bac8-8b520ca7f32d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5370426ad5c763400ffa5cd543861d5f22c3dfbd34e2d3b30736152f4bd41eef +size 937745 diff --git a/2024/View From Above_ Orthogonal-View aware Cross-view Localization/full.md b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e53eca0c87ac86732c1e6577cebb0714617e8641 --- /dev/null +++ b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/full.md @@ -0,0 +1,276 @@ +# View From Above: Orthogonal-View aware Cross-view Localization + +Shan Wang1,2 Chuong Nguyen1 Jiawei Liu2 Yanhao Zhang2 Sundaram Muthu1 Fahira Afzal Make1 Kaihao Zhang2 Hongdong Li2 1Data61, CSIRO 2Australian National University + +# Abstract + +This paper presents a novel aerial-to-ground feature aggregation strategy, tailored for the task of cross-view image-based geo-localization. Conventional vision-based methods heavily rely on matching ground-view image features with a pre-recorded image database, often through establishing planar homography correspondences via a planar ground assumption. As such, they tend to ignore features that are off-ground and not suited for handling visual occlusions, leading to unreliable localization in challenging scenarios. We propose a Top-to-Ground Aggregation (T2GA) module that capitalizes aerial orthographic views to aggregate features down to the ground level, leveraging reliable off-ground information to improve feature alignment. Furthermore, we introduce a Cycle Domain Adaptation (CycDA) loss that ensures feature extraction robustness across domain changes. Additionally, an Equidistant Re-projection (ERP) loss is introduced to equalize the impact of all keypoints on orientation error, leading to a more extended distribution of keypoints which benefits orientation estimation. On both KITTI and Ford Multi-AV datasets, our method consistently achieves the lowest mean longitudinal and lateral translations across different settings and obtains the smallest orientation error when the initial pose is less accurate, a more challenging setting. Further, it can complete an entire route through continual vehicle pose estimation with initial vehicle pose given only at the starting point. + +# 1. Introduction + +Visual-based cross-view localization aims to locate query images taken from street-level cameras (referred to as ground view) within a satellite or aerial view map. Platforms like the Google Map API [6] have made satellite images accessible, spurring the development of cross-view localization techniques. Notable studies [4, 7, 10, 15, 20, 25, 26, 30] have focused on using satellite images for this task. Despite this, accurate localization remains challenging due + +![](images/4c2274fc7cca192d79b8542da93b13793098cf7c5e1606d7a9a9a9740916040c.jpg) +Ground View + +![](images/48b30efe7caab7dabfe78e7bcbcfdca01a7a7924e63a57ea63fca10d5b3bfd1d.jpg) +Aerial View +Figure 1. Efficacy of T2GA. T2GA aggregates off-ground features (e.g., the streetlight within the green square) and addresses occlusions (e.g., the tree within the red square). Unlike conventional methods that often overlook such elements, T2GA integrates these features, thus improving appearance matching for ground-level aligned pixels across views using ground plane homography. Insets are magnified for clarity, with dotted outlines indicating appearance misalignment and solid outlines indicating appearance alignment with the corresponding aerial view features. + +to the significant viewpoint differences between ground and aerial images [14, 31]. These viewpoint differences result in a domain gap, adversely impacting feature alignment, thereby compromising the overall accuracy of localization. + +Recent research has explored two main approaches to bridge the domain gap in image-based localization: generative-based and geometry-alignment-based methods. Generative-based methods, such as those utilizing GANs [19] and diffusion models [29], reduce the domain gap by transforming view styles from one view to another. However, the generated features for matching can lead to ambiguities in pose estimation. In contrast, geometry-alignment-based methods, employing techniques like polar transformations [15, 20] or homography [14, 23, 24, 26], focus on establishing correspondences for on-ground pixels. This often results in the neglect of off-ground features, such as streetlights, and difficulty in handling visual occlusions, such as the obscuring of road details by treetops in aerial views, both illustrated in Fig. 1. This neglect fails to leverage important geographic landmarks on the road and results in a lack of robustness to issues like road mark degradation (e.g. fading and damaged paintings). To address these limitations, we introduce the Top-to-Ground Aggregation (T2GA) module. T2GA employs top-down feature aggre + +gation to enrich on-ground points with an above-view perspective appearance, significantly improving feature alignment and localization accuracy. Furthermore, we recognize that pixel positions above ground points may not necessarily indicate higher elevations; instead, they could represent objects located further along the camera's line of sight. To resolve this ambiguity, we integrate a transformer mechanism that assesses whether these pixels belong to the same object and determines occlusion precedence, such as shadows on the road, which suggest likely occlusions in the aerial view. + +In addition to viewpoint differences, ground and aerial views differ in cameras types, lighting conditions, tones, and resolutions. We introduce the Cycle Domain Adaptation (CycDA) loss function to address these variations. CycDA enables bidirectional feature generation between ground and aerial views. By minimizing the discrepancy between domain-adapted features and their target counterparts, our approach ensures that features extracted from one domain are effectively translatable to the other, fostering the extraction of features that are invariant across domains. + +We further introduce an Equidistant Re-projection Loss to address a common bias in keypoint-synchronous detection localization methods, which tend to favor closer keypoints due to their reduced orientation errors. Our loss function mitigates this issue by applying a distance-weighted approach, ensuring that orientation errors are independent of keypoint distance. Consequently, keypoints are more uniformly distributed across various distances, leading to a more equitable and precise estimation of direction. + +We summarize our contributions as follows: + +- A top-to-ground feature aggregation module that effectively bridges the domain gap between ground and aerial views by utilizing off-road cues and handling occlusions. +- A cycle domain adaptation loss function that promotes domain-invariant feature extraction, enhancing the robustness of cross-view localization methods. +- An equidistant re-projection loss function ensures orientation errors are consistent regardless of keypoint distance, leading to a more extended distribution of keypoints and more accurate orientation estimation. + +The evaluations are conducted on KITTI and Ford Multi-AV datasets while leveraging the Google Maps as corresponding satellite images. Experimental results demonstrate that our method can achieve consistent vehicle pose estimation under various challenging situations. Our method achieves the lowest mean longitudinal and lateral translation errors on both KITTI and Ford Multi-AV datasets under different settings. We also obtain the smallest orientation error when the initial pose is less accurate, demonstrating the robustness of our method. Further, with initial vehicle pose only at the starting point, our method can complete the route through continuous pose estimation, demonstrating its generalisation ability and potential for practical deployment. + +# 2. Related Work + +Image-level cross-view localization. Early visual-only cross-view localization methods [7, 10, 15, 16, 20, 30] approach the task as an image retrieval problem, focusing on coarse localization through image-to-image matching. To bridge the domain gap between ground and aerial views, various techniques have been proposed to facilitate cross-view feature matching. [10] incorporated per-pixel orientation information, while [15] and [20] utilized a predefined polar transform to align aerial-view images with ground views. Additionally, [19] and [29] employed GAN-based and diffusion model-based style transformations, respectively. While efforts have been made to minimize the domain gap, these methods rely on image-level feature matching, restricting their localization accuracy, often falling short of the precision achieved by commercial GPS systems in open areas [21]. + +Patch-wise cross-view localization. To improve accuracy, several methods employ patch-wise feature matching, especially effective in aerial views with wide fields of view and high resolutions. For instance, [31] employs transformers to emphasize informative patches, [27, 28] computes dense spatial distributions using patch attention, and [9] introduces the 'slice-to-sector match' and 'Cross-view attention' to calculate similarity between aerial sectors and ground view slices. However, their attention queries are derived from aerial images, posing reliability issues when the pose is unknown. OrienterNet [13] transforms ground view images into Bird's Eye View (BEV) grids for matching with OpenStreetMap data. Despite these advancements, localization accuracy remains limited by the patch (grid) size. + +Pixel-wise cross-view localization. Pixel-wise feature matching has been explored in various methods [4, 14, 24, 25] for precise localization, yet they struggle with the inherent domain gap between views. Our baseline, PureACL[24], and other geometry-alignment-based methods [14, 26] prioritize on-ground pixel correspondences but overlook off-ground features and occlusions, leading to suboptimal performance. CVGL [4] transforms the ground view image into a BEV for aerial matching. Their BEV transformation employs aerial coordinates as queries, introducing a high degree of freedom and increasing the complexity of matching. In the BoostAcc [17] framework, homography transformation is used on ground views to source query pixel data for the pixel-to-slice attention mechanism. This transformation introduces distortions from non-ground pixels, potentially exacerbating errors in subsequent processing stages. Moreover, the keys and values span entire and adjacent columns, risking compounded ambiguities in both longitudinal and lateral estimation. Our approach aims to address these challenges, proposing a novel pixel-to-overhead-column attention mechanism for improved cross-view localization accuracy. + +![](images/b8c62cdf43ff5f11886d258e92cc1ca6325f8d87ff7c015e68783cb5c251e609.jpg) +Figure 2. Our method adopts PureACL [24] as the baseline model and introduces three novel components (highlighted in red): (1) T2GA (Sec.4.1): aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them to alleviate the representation gap of the same object across different views; (2) CycDA Loss (Sec.4.2): explicitly enforces a view-invariant representation for the same object; and (3) ERP Loss (Sec.4.3): allows the model to leverage key points that are farther away from the vehicle while allocating more emphasis on correcting the vehicle orientation estimation. + +# 3. Preliminary + +Task Settings. The Fine-Grained Cross-View Localization (FGCVL) task aims to estimate the accurate 3-DoF pose of a vehicle, including longitudinal and lateral translations and orientation on a satellite map $I^{s}$ , given an initial coarse pose $\mathbf{P}_{init}$ . It also assumes that the vehicle is equipped with on-board camera(s) $^2$ , and the image captured by these cameras, denoted as the ground image $I^{g}$ , is accessible. + +Baseline Model. PureACL [24] is adopted as the baseline model for our approach. It employs a share-weight U-Net to extract ground/satellite feature maps $F = \{F^g, F^s\}$ and ground/satellite confidence maps $C = \{C^g, C^s\}$ from the ground/satellite images $I = \{I^g, I^s\}$ . The ground confidence map $C^g$ is tasked to select the top-N most confident pixel-level features from $F^g$ at the positions $p^g = \{(u_i^g, v_i^g)\}_{i=1}^N$ . These features are then: (i) Projected from on-board camera coordinates to the satellite coordinates using $p_i^s(\mathbf{P}) = K_s \mathbf{P} K_g^{-1} p_i^{g,4}$ , where $K_s$ and $K_g$ are the intrinsic matrices of satellite and on-board cameras, respectively. $\mathbf{P}$ is the initial pose $\mathbf{P}_{init}$ at the first iteration and the estimated pose $\mathbf{P}_{pred}$ in the subsequent iterations. (ii) A weighted residual $w(\mathbf{P}) \times r(\mathbf{P})$ is computed between the top-N selected features from the ground feature map and the corresponding features from the satellite feature map as: + +$$ +w (\mathbf {P}) = C ^ {s} \left[ p ^ {s} (\mathbf {P}) \right] \times C ^ {g} \left[ p ^ {g} \right], r (\mathbf {P}) = F ^ {s} \left[ p ^ {s} (\mathbf {P}) \right] - F ^ {g} \left[ p ^ {g} \right], \tag {1} +$$ + +where $[\cdot ]$ is a lookup operation in a feature/confidence map with sub-pixel interpolation. (iii) The Levenberg-Marquardt (LM) algorithm [11] takes the computed weighted residual $w(\mathbf{P})\times r(\mathbf{P})$ as input and outputs the predicted vehicle pose $\mathbf{P}_{pred}^{m}$ at the $m^{\mathrm{th}}$ iteration, with $m\in$ $\{1,\ldots ,M\}$ . The process of (i) - (iii) repeats for $M = 20$ + +iterations to produce the final vehicle pose $\mathbf{P}_{pred}$ . + +PureACL [24] employs two loss functions: a Triplet Loss [12] and a Re-projection Loss. The former supervises the weighted residual, enforcing that the encoder extracts pose-sensitive features from both ground and satellite images. The later utilises the projection error w.r.t the ground truth pose $\mathbf{P}_{gt}$ to penalize incorrectly predicted vehicle pose $\mathbf{P}_{pred}$ (Refer to Supplementary Sec.D for more details). + +# 4. Method + +We propose a novel View From Above (VFA) method to tackle the FGCVL task by aligning feature representations across the orthogonal viewpoints, the ground view and the satellite views. As illustrated in Fig. 2, our method introduces three innovative components onto the baseline model: (1) Top-to-Ground Aggregation Module (T2GA) (Sec. 4.1); (2) Cycle Domain Adaptation Loss (CycDA) (Sec. 4.2); and (3) Equidistant Re-Projection Loss (Sec. 4.3). + +# 4.1. Top to Ground Aggregation + +The appearance discrepancy between the same objects viewed from ground and satellite perspectives presents a significant challenge in achieving high precision in the FGCVL task [14, 17, 24, 26]. This issue is often encountered on tall structures, e.g. street light, and the occlusion from aerial objects, e.g. tree branches, where their near(on)-ground appearance, accessible to only on-board vehicle camera, is significantly different from their top appearances which are only registered by the satellite camera (see Fig. 1). Such discrepancies are common on roads and can lead to inaccuracies in cross-view localization. To address this challenge, we propose aligning the representations between ground and satellite views, focusing specifically on the on-ground pixels, where geometry alignment across the two views hold due to the ground plane homography. Specifically, we present a T2GA module that aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them, as illustrated in the 'T2GA' red box in Fig. 2. For an on-ground pixel: + +$p\in \{(u,v)\}_{u = 0,v = \tau}^{W^g -1,H^g -1}$ , and where $\tau$ is a threshold corresponding to the height of horizon in the ground view image, and $W^{g}$ and $H^{g}$ are the width and height of the ground view image, respectively. The corresponding elevated pixels are $q\in \{(u_p,v)\}_{v = 0}^{v_p}$ , where $(u_{p},v_{p})$ are the coordinates of the on-ground pixel $p$ . The aggregation is achieved via the attention mechanism [22], defined as: + +$$ +F ^ {a t t} [ p ] = \operatorname {S o f t m a x} \left(\mathbf {Q K} ^ {T}\right) \mathbf {V}, \tag {2} +$$ + +where $F^{att}[p]$ represents the feature vector at a pixel $p$ on the feature map $F^{att}$ . In this context, $p$ and $q$ denote an on-ground pixel and its corresponding elevated pixels, respectively. The query is formed by taking the features of on-ground pixels $\mathbf{Q} = \mathcal{M}(F^g[p])$ where $\mathcal{M}(\cdot)$ represents a non-linear mapping consisting of an MLP layer followed by an activation function. The value is derived from the features of elevated pixels corresponding to the on-ground pixel $\mathbf{V} = F^g[q]$ , and the key is then obtained by applying the same non-linear mapping to $\mathbf{V}$ , resulting in $\mathbf{K} = \mathcal{M}(\mathbf{V})$ . This non-linear mapping ensures that the query and key are mapped to the same feature space before computing their matrix product. $\mathbf{Q}\mathbf{K}^T$ . Notably, value $\mathbf{V}$ is used directly in its original form for the computation, facilitating a straightforward fusion (for the same object) or replacement (in cases of occlusion) at ground level. + +![](images/21d5ec7a225c4edfc15de4afa07b9b409317055b7e36f3ec5aa957ce9a2935a7.jpg) +Figure 3. Attention between the on-ground pixels and their corresponding elevated pixels is displayed column-wise. (Left example): The attention between the base and top of the streetlight is high despite their distinct appearances. This allows the resultant aggregated ground feature to be aligned with the satellite feature corresponding to the matching geographic location. (Right example): In the absence of occlusion from above, ground pixels maintain minimal attention with their elevated pixels. This avoids unnecessary dilution of the ground features which are already well aligned with their satellite counterparts. + +The computed attention map $\operatorname{Softmax}(\mathbf{Q}\mathbf{K}^T)$ can accurately reflect the connection between the on-ground pixels and their corresponding elevated pixels. As illustrated in Fig. 3, high attention values are assigned to the base and top + +of the streetlight despite their different appearances. This provides important cues to alleviate the representation gap between the ground feature and satellite features. On the other hand, in the absence of occlusion, e.g. tall structure and tree branch, thus the attention between on-ground and elevated pixels is minimal, avoiding diluting the ground features that are well aligned with their satellite counterparts. + +The feature map $F^{att}$ after attention calculation is vertically stacked with the upper part of the original $F^g$ to generate the aggregated feature map $F^a$ : + +$$ +F ^ {a} = F _ {0: H ^ {g} - \tau} ^ {g} \oplus F ^ {a t t}, \tag {3} +$$ + +where $\oplus$ denotes vertically stacking and $F_{0:H^g -\tau}^g$ indicated the slicing of $F^{g}$ from row 0 to $H^{g} - \tau$ . Subsequently, this aggregated feature map $F^a$ replaces $F^g$ in the residual calculation as per Eq.1, resulting in an updated formula: $r(\mathbf{P}) = F^{s}[p^{s}(\mathbf{P})] - F^{a}[p^{g}]$ + +The advantage of the aggregated feature map is also indirectly reflected by the ground confidence map of the baseline model as shown in Fig. 4. Without the T2GA module, the confidence map only highlights pixels corresponding to road marks and curds, resulting in a concentrated sampling of key points for subsequent estimation of vehicle pose. This has the following drawbacks: (1) fail to leverage road landmarks e.g. traffic signal poles, that encode important geographic location; and (2) not robust to road mark degradation problems, e.g. fading and damaged paintings, and visibility issues, e.g. glow on road marks. T2GA alleviates the feature discrepancy between ground and satellite views. As a result, the traffic signal pole receives higher confidence, allowing it to be subsequently sampled as key points to estimate more precise vehicle pose. + +![](images/0a68c878c1dd72a93fadf4187eeeb97cdb6f7b5ab34270c603c6b8b7618100bf.jpg) +Figure 4. Illustration on the effect of T2GA on the confidence map of the baseline model. The confidence map without T2GA (Left) predominantly highlights road marks and curbs, resulting in subsequent keypoint sampling missing important road landmarks, e.g. traffic signal poles, that provide important cues to vehicle pose estimation; The confidence map with T2GA (Right) has high-confidence values distributed across various road marks and traffic poles. With more geographic cues provided by multiple sources, the resultant pose prediction becomes more precise and robust. + +![](images/b0f56c8b54e0dc05d3122eb5250d2d7ca74407891b3a3902469405b916a275d2.jpg) + +# 4.2. Cycle Domain Adaptation Loss + +There are multiple sources responsible for the representation gap between ground and satellite features. In addition to the varying appearances of the same object across different views, camera specifications, e.g. tone, hue, intensity, brightness and resolution, and temporal changes of + +varying spans, can also lead to inconsistent representations for the same geographic location. Such issue is in general overlooked by the existing FGCVL methods [14, 28]. Despite introducing a triplet loss to further distinguish feature representations based on different poses, our baseline model, PureACL [24], falls short in enforcing invariant feature representation at the corresponding geographic locations across different views. + +We propose a Cycle Domain Adaptation (CycDA) loss to explicitly enforce view-invariant representations that consists of three L2-loss between ground and satellite feature representations in three different feature spaces. The representation loss in the ground feature space is defined as: + +$$ +\mathcal {L} _ {g} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\| F ^ {a} \left[ p _ {i} ^ {g} \right] - \mathcal {D} \left(\mathcal {E} \left(F ^ {s} \left[ p _ {i} ^ {s} \left(\mathbf {P} _ {g t}\right) \right] \odot c _ {s}\right) \odot c _ {g}\right) \right\| _ {2}, \tag {4} +$$ + +where $\mathcal{E}(\cdot)$ is a projection from the pixel feature space to a latent feature space, $\mathcal{D}(\cdot)$ is a projection from the latent feature space to the pixel feature space, $c_{s}$ and $c_{g}$ represent the focal length of satellite camera and vehicle camera respectively which are introduced to guide the projection functions, and $\odot$ is a concatenation acting on the feature channel dimension. Similarly, the representation loss in the satellite feature space is: + +$$ +\mathcal {L} _ {s} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\| F ^ {s} \left[ p _ {i} ^ {s} \left(\mathbf {P} _ {g t}\right) \right] - \mathcal {D} \left(\mathcal {E} \left(F ^ {a} \left[ p _ {i} ^ {g} \right] \odot c _ {g}\right) \odot c _ {s}\right) \right\| _ {2}, \tag {5} +$$ + +We also enforce the satellite feature to be aligned with the ground feature corresponding to the same geographic location in a latent feature space through: + +$$ +\mathcal {L} _ {m} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\| \mathcal {E} \left(F ^ {a} [ p _ {i} ^ {g} ] @ c _ {g}\right) - \mathcal {E} \left(F ^ {s} [ p _ {i} ^ {s} (\mathbf {P} _ {g t}) ] @ c _ {s}\right) \right\| _ {2}. \tag {6} +$$ + +The overall CycDA loss can then be defined as: + +$$ +\mathcal {L} _ {C y c D A} = \mathcal {L} _ {g} + \mathcal {L} _ {s} + \mathcal {L} _ {m} \tag {7} +$$ + +# 4.3. Equidistant Re-projection Loss + +Our method tackles a challenging setting of simultaneously detecting key points from the ground image and predicting the vehicle pose using the detected key points. Our baseline model introduces a re-projection loss computed over the selected top-N key points defined as: + +$$ +L _ {R P} = \frac {1}{N} \sum_ {i = 1} ^ {N} \| p _ {i} ^ {s} (\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\mathbf {P} _ {g t}) \| _ {2} ^ {2}, \tag {8} +$$ + +where $p_i^s (\mathbf{P}_{pred})$ and $p_i^s (\mathbf{P}_{gt})$ are the projected coordinates on the satellite image from the selected top-N ground key points with the predicted vehicle pose and the groundtruth vehicle pose respectively, and $\| \cdot \|_2$ denotes the L2-distance. + +![](images/9a278da9d73fe47a641ce4bc30610a218894337cb1901b491a0a06e10daa1113.jpg) +Figure 5. Comparison of detected keypoints with and without ERP Loss. (Left) Without the ERP Loss, keypoints are predominantly located in close proximity to the vehicle. (Right) With the ERP Loss, there is a more dispersed distribution of keypoints. + +![](images/d7b72283fb9b53d1a00eeefca7e1fbadd8b63ecf1e517fef4169d3bd492e30cc.jpg) + +Despite being effective in pipelines where simultaneous detection of key points is not required [3, 18, 25], the reprojection loss tends to overly penalize the orientation errors of distant points. This forces the detected key points being closer to the vehicle, as illustrated in Fig. 5, and less accurate vehicle orientation estimation (see Sec. 5.2). + +To mitigate the aforementioned issue, we propose an Equidistant Re-Projection (ERP) loss that is defined as: + +$$ +L _ {E R P} = \frac {1}{N} \sum_ {i = 1} ^ {N} \| \frac {p _ {i} ^ {s} (\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\mathbf {P} _ {g t})}{\mathbf {D} _ {p _ {i} ^ {s}}} \| _ {2} ^ {2}, \tag {9} +$$ + +where $\mathbf{D}_{p_i^s} = \| p_i^s (\mathbf{P}_{gt}) - p_{vehicle}^s (\mathbf{P}_{gt})\|_2$ is the L2-distance between the satellite image coordinates of the $i^{\mathrm{th}}$ keypoint and the vehicle. The design prevents the orientation error from scaling with the L2-distance between key points and vehicle on the satellite image, which dominates over the orientation error on distant key points. As shown in Fig. 5, our method can leverage key points that are farther from the vehicle than our baseline model. + +# 5. Experiments + +# 5.1. Implementation Details + +Datasets. To evaluate our proposed approach, we conduct experiments on two well-established autonomous driving datasets: the Ford Multi-AV Seasonal dataset (FMAVS) [1] and the KITTI dataset [5]. Consistent with established practices [14, 24, 25], our primary focus is on images from the front left camera as query images. Additionally, following the methodology of PureACL [24], we broadened our analysis to include four camera perspectives: front left, rear right, side left, and side right. For data partitioning in the KITTI-CVL dataset, we utilized an approach in line with HighlyAcc [14], which involves two test splits. The first, 'Same,' includes images from the same trajectories as the training dataset, while the second, 'Cross,' comprises images from distinct trajectories. Regarding the FMAVS dataset, we follow the partitioning strategy proposed by PureACL [24], using all images from the same trajectories but different traversals. It is noteworthy that there are minor route variations within these same trajectories6 + +Metrics. We adopt evaluation metrics in line with HighlyAcc [14], SliceMatch [9], SIBCL [25], and PureACL [24]. + +Table 1. Comparison on the KITTI-CVL Dataset Under Initial Noise Conditions (±10°,±20m). Results marked with $\star$ are sourced directly from the respective original papers. Those denoted by $\diamond$ indicate retraining of methods for consistent evaluation criteria alignment. - indicates that the corresponding data were not provided. + +
ModelLocation(m)↓Lateral(%)↑Longit(%)↑Orient(°)↓Orient(%)↑
MeanMedianr@1mr@3mr@5mr@1mr@3mr@5mMeanMedianr@1°r@3°r@5°
Same★DSM[15]--10.1230.6748.244.0812.0120.14--3.5813.8124.44
★HighlyAcc[14]12.0811.4235.5470.7780.365.2215.8826.133.722.8319.6451.7671.72
★SliceMatch[9]7.964.3949.09-98.5215.19-57.354.123.6513.41-64.17
★BoostAcc[17]10.015.1976.4496.3498.8923.5450.5762.180.550.4299.10100.00100.00
★CCVPE [28]1.220.6297.3598.6599.7177.1396.0897.160.670.5477.3999.4799.95
★HC-Net [26]0.800.5099.01-99.7392.20-99.250.450.3391.35-99.84
◇PureACL [24]2.420.4291.9593.4094.2891.8692.1292.383.971.7732.7153.8971.66
Ours0.200.1799.9799.9799.9799.9799.9799.971.530.9350.9583.9795.45
Cross★DSM[15]--10.7731.3748.243.8711.7319.50--3.5314.0923.95
★HighlyAcc[14]12.5812.1127.8259.7972.895.7516.3626.483.953.0318.4249.7271.00
★SliceMatch[9]13.509.7732.43-86.448.30-35.574.206.6146.82-46.82
★BoostAcc[17]13.019.0657.7286.7791.1614.1534.5945.000.560.4398.98100.00100.00
★CCVPE [28]9.163.3344.0681.7290.2323.0852.8564.311.550.8457.7292.3496.19
★HC-Net [26]8.474.5775.00-97.7658.93-76.463.221.6333.58-83.78
◇PureACL [24]6.200.6167.2487.2692.7664.8173.6389.694.262.4823.4548.6959.24
Ours0.210.1899.9299.9699.9799.9199.9199.912.041.3838.3180.1292.48
+ +Our precision assessment includes median and mean errors for overall, lateral, and longitudinal translations, as well as orientation accuracy. Additionally, our analysis covers lateral and longitudinal translations and localization recall at various distances (0.25m, 0.5m, 1m, 3m, and 5m), and orientation recall within a range of $1^{\circ}$ to $5^{\circ}$ . + +Training Details. We apply two sets of criteria: (1) Following HighlyAcc [14], ground images are processed at a resolution of $256 \times 1024$ , and satellite images at $512 \times 512$ , with initial pose noise $\pm 10^{\circ}$ for orientation and $\pm 20\mathrm{m}$ for translations. (2) In accordance with PureACL [24], we process FMAVS images at $432 \times 816$ , KITTI images at $384 \times 1248$ , and satellite images at $1,280 \times 1,280$ . To better simulate turning scenarios, we adopt an expanded initial pose noise range of $\pm 45^{\circ}$ for orientation and $\pm 20\mathrm{m}$ for translations. For training, we utilize an NVIDIA RTX 3090 GPU with a batch size of 3, employing the Adam optimizer [8] with a learning rate of $10^{-5}$ . Feature extractor weights are adapted from PureACL [24], while other components are initialized randomly. Training iterations average around $285\mathrm{ms}$ , including $200\mathrm{ms}$ dedicated to optimization. The inference speed, subject to initial pose variability, averages at $222\mathrm{ms}$ . + +# 5.2. Comparison with Existing Methods + +We evaluate our method against recent visual-only approaches on the KITTI-CVL dataset following the metrics of HighlyAcc [14]. The results shown in Tab. 1 clearly indicate the superiority of our approach in spatial precision, consistently maintaining poses within a $1\mathrm{m}$ radius in both 'Same' and 'Cross' areas with over $99.9 + \%$ probability. While our method may not lead in orientation accuracy under an initial pose range of $\pm 10^{\circ}$ , it demonstrates + +superior performance when the range is extended to $\pm 45^{\circ}$ , as detailed in Tab. 2. This underlines the robustness of our approach. Our approach leverages pixel-wise localization, offering finer granularity compared to the patchwise localization of SliceMatch[9], CCVPE [28], and the image-level localization of DSM [15]. Although HighlyAcc [14] and HC-Net [26] are pixel-wise localization methods, their homography-based mechanisms fail to incorporate off-road cues. In contrast, our approach effectively utilizes these cues, resulting in enhanced performance. Additionally, our longitudinal estimation significantly surpasses that of BoostAcc [17]. This improvement is likely attributable to our method's strategy of not aggregating information from pixels below, thereby reducing longitudinal ambiguity. In comparison to the baseline method PureACL [24], our approach achieves significant enhancements in both orientation and spatial accuracy, especially in mean error reduction. Specifically, we observe a $91^{+}\%$ reduction in spatial mean error and a $52^{+}\%$ reduction in orientation mean error. The difference between mean and median errors for PureACL indicates potential convergence issues in high-noise environments. By integrating the T2GA and CycDA modules, we bridge the domain gap and ensure convergence even under challenging conditions. + +# 5.3. Challenge Initial Pose and Stringent Metrics + +To rigorously evaluate our method, we adopt the stringent metrics of PureACL [24], which are more demanding than those used by HighlyAcc[14]. We benchmark our method against baseline methods SIBCL [25] and PureACL [24], as well as SOTA methods CCVPE [28] and BoostAcc [17]. The results in Tab. 2 reaffirm the advantages of our + +Table 2. Comparison with Initial Noise Conditions $(\pm 45^{\circ},\pm 20\mathrm{m})$ .For the KITTI-CVL dataset, evaluations are conducted in 'K-Same' and 'K-Cross' areas. On the Ford-CVL dataset, the Log4' trajectory, as used in SIBCL [25], is chosen for its optimal satellite view alignment. Additional log evaluations are detailed in the supplementary material. F-1C' and F-4C' represent assessments on the Ford-CVL dataset using either a single front-facing camera or four surrounding cameras. Note: $\star$ indicates that SIBCL [25] is a hybrid LiDAR-visual method. + +
ModelLateral(m)↓Mean MedianLateral(%)↑r@0.25mr@0.5mr@1mLongit(m)↓Mean MedianLongit(%)↑r@0.25mr@0.5mr@1mOrient(°)↓Mean MedianOrient(%)↑r@1°r@2°r@4°
K-SameCCVPE [28]2.321.2211.0522.0542.096.493.357.4513.3624.973.431.5334.6660.7284.03
BoostAcc[17]1.190.6321.9241.5167.7210.015.425.179.3816.543.882.9818.4535.2562.20
★ SIBCL[25]1.230.7017.0133.3461.883.140.8620.6831.4055.588.901.6935.5055.0471.06
PureACL[24]2.940.2358.7566.3470.603.150.3745.6568.2374.878.132.8037.9952.5764.13
Ours0.170.1476.2097.7399.970.080.0798.1899.8999.972.191.0748.9766.2990.54
K-CrossCCVPE [28]4.562.794.9610.0538.1212.068.163.136.1412.4519.4615.393.3810.2417.41
BoostAcc[17]2.650.9314.8829.2052.6410.468.923.385.5611.735.614.2612.4624.8747.26
★ SIBCL[25]2.720.7116.9032.6258.245.481.0519.5830.7449.259.671.9625.0250.1860.16
PureACL [24]4.240.2550.2562.6364.563.890.447.4162.0464.2310.963.1619.9536.2556.49
Ours0.170.1575.9697.4799.970.090.0797.4999.8799.972.881.4436.1063.0886.18
F-1CCCVPE [28]4.622.456.4812.5424.4610.218.413.228.2413.5320.4815.073.146.1416.14
BoostAcc[17]2.731.479.0518.1835.5910.516.144.218.1315.566.824.7911.6122.6442.63
★ SIBCL[25]2.590.7120.9241.5360.435.381.1812.5923.9943.876.341.5735.2158.4170.65
PureACL[24]2.941.5411.3421.1437.764.751.7410.4620.0737.647.383.3315.1230.4257.01
Ours0.490.4131.7759.1789.230.360.3238.4375.7297.462.191.0847.3172.7190.10
F-4CCVGL[4]1.280.5021.8150.0586.841.530.7019.3637.2665.51-----
PureACL[24]1.690.6721.0739.5365.273.381.1812.5923.9943.863.071.1943.8473.2782.41
Ours0.130.1088.2598.53100.000.200.1868.6796.68100.002.001.4935.7262.7091.83
+ +![](images/803aa4bd5d4ce7de98c86077c936849f854378747adffb0d7364370b7415f09b.jpg) +(a) Location + +![](images/61468ab18c74c8c4334bbaa0adf3608827ee0a5aababe9613a7eb7987fa6f1ae.jpg) +(b) Orientation +Figure 6. Comparison of the baseline method PureACL and our proposed method under varying initial noise ranges in the 'cross' area of the KITTI-CVL dataset. The blue bar indicates the median value, while the red bar shows the difference between the median and mean values. Notably, our method demonstrates robust convergence, outperforming PureACL across all noise ranges, especially in larger noise scenarios where PureACL shows significant discrepancies between median and mean values. + +method, particularly in terms of reduced mean error and improved longitudinal estimation. Notably, under conditions of minimal initial pose noise $(\pm 15^{\circ}$ and $\pm 5\mathrm{m})$ , PureACL [24] outperforms SIBCL [25]. However, when faced with more challenging conditions $(\pm 45^{\circ}$ and $\pm 20\mathrm{m})$ , PureACL [24] encounters convergence challenges, unlike our method, which demonstrates consistent superiority. This highlights the effectiveness of our proposed modifications in feature alignment. Additionally, we assess our method's convergence range against PureACL [24] in the 'Cross' area of the KITTI-CVL dataset, as depicted in Fig. 6. The red bar in the figure highlights a significant reduction in the gap between median and mean errors. This observation con + +firms our method's enhanced ability to converge effectively across a wider range of noise levels. + +# 5.4. Results with Continual Pose Estimation + +To address GPS signal loss, we adopt an accumulated pose estimation strategy that leverages initial poses based on the vehicle's previous pose estimates $^{7}$ . We use the model trained with initial noise allowances of $\pm 45^{\circ}$ and $\pm 20\mathrm{m}$ . Our method's robustness is evidenced by the continuous running distance percentage presented in Tab. 3, showing successful pose chaining throughout all evaluate routes. In contrast, BoostAcc [17] and the single-camera PureACL [24] exhibit significant drift, leading to errors beyond the satellite map's coverage and causing evaluations to cease after covering less than $11\%$ of the distance. Performance comparisons on the KITTI-CVL dataset are shown in Fig. 7 $^{8}$ . Our approach consistently achieves accurate pose estimation or experiences only minor drifts in challenging scenarios characterized by limited localization cues or severe occlusions. In contrast, BoostAcc [17] exhibits significant drift when a moving vehicle passes by, likely due to incorrect query data from projected vehicle pixels. Furthermore, PureACL [24] appears to struggle with incorrect orientation, resulting in reversed pose estimations. These out + +![](images/dba6508b74ea847d65b37c6f59c14ef1ebfa503e182090c33be623c1b44d26b6.jpg) + +![](images/d50ca463c12e7424e05b9a54a4ad8730d11f9bb30130cb4e506abef533043bd7.jpg) + +![](images/63220183812e94646ecfc564e94df1d84abe5069491dc07688399c88daf2fba2.jpg) +Figure 7. Accumulated pose estimation performance on a KITTI-CVL dataset trajectory. (Top-left) Predicted poses by our method (blue) closely match the ground truth (red) over the 3676-meter route, outperforming PureACL (green) and BoostAcc (cyan), which exhibit severe drift. (Top-right) Our approach demonstrates rapid recovery upon encountering clear localization cues. (Bottom-left) (Bottom-right) BoostAcc demonstrates substantial drift when encountering moving vehicles. + +![](images/83fa2f54b3f2ceb0d3a63586dde6dd8fc200e2b21f8265a85929c88f7afe519a.jpg) + +comes highlight the importance of robustness in real-world applications, as drifts that extend beyond the coverage of satellite imagery can compromise all further pose estimation processes. + +Table 3. Comparison on Accumulated Pose Estimation. + +
ModelRunning percentage(%)↑
K-CrossF-1CF-4C
BoostAcc[17]4.884.89-
PureACL[24]10.315.01100.00
Ours100.00100.00100.00
+ +- indicates that BoostAcc does not support 4-camera setting. + +# 6. Ablation Study + +To evaluate the effectiveness of our proposed components, we conduct ablation studies in the 'cross' area of the KITTI-CVL dataset. Our method's performance was compared across various configurations, including the presence and absence of CycDA and Equidistant Re-projection (ERP) Loss functions, and the T2GA module. The results, as detailed in Tab. 4, underscore the significant role each component plays in improving our method's performance. + +The T2GA module aggregates information from top to ground based on the assumption that the ground and satellite views are orthogonal. However, this assumption might not be accurate in scenarios involving uphill/downhill or turns. To test our method's robustness under these conditions, we introduced affine warping to the ground view images with a shear range of $\pm 15^{\circ}$ . The outcomes, labeled 'w/ Affine' in Tab. 4, indicate our method's efficacy in handling such + +scenarios. This success is primarily due to two factors: (1) the use of a U-Net for feature extraction allows each pixel in the feature map to encapsulate information from its surrounding area, enabling the handling of a degree of angular variation. (2) the satellite view from Google, which is not strictly orthogonal to the ground view, suggests that our method is already adapted to these types of challenges. + +Table 4. Ablation study on the KITTI-CVL dataset 'Cross' area with initial noise $(\pm 45^{\circ},\pm 20\mathrm{m})$ + +
T2GA CycDA ERPLateral(m)↓Mean MedianLongit(m)↓Mean MedianOrient(°)↓Mean Median
4.240.253.890.4410.963.16
✓1.290.191.270.174.112.13
✓✓0.210.180.120.103.381.89
✓✓✓0.170.150.090.072.881.44
w/ Affine0.180.150.090.072.881.52
night0.390.270.390.306.642.55
+ +We further evaluate our method on an artificial dataset derived from the KITTI test sets, created using the cross-domain deep network proposed by Arruda et al. [2]. The outcomes, labeled as 'night' in Tab. 4 show our method's stable performance in nighttime scenarios, highlighting its resilience to variations in lighting and time of day. For further details, please refer to Supplementary Sec. C. + +# 7. Conclusion + +We have presented a novel top-to-ground feature aggregation for enhancing cross-view image-based geolocation. Our method overcomes limitations by incorporating aerial perspectives and utilizing a cycle domain adaptation loss for consistent feature extraction despite visual disparities. The introduction of the equidistant re-projection loss balances keypoint impact, promoting wider distribution and thus enhancing orientation accuracy. Our method excels in vehicle pose estimation across challenging scenarios, achieving the lowest translation errors in KITTI and Ford datasets, and minimal orientation error with less accurate initial poses. Crucially, by relying solely on the initial vehicle pose at the start, our method successfully completes routes via continuous pose estimation. This paves the way for real-world applications like autonomous driving and outdoor robotics. Future work will focus on integrating this method into SLAM systems to remove loop closure dependency and facilitate high-definition map generation. + +Acknowledgements. This project is supported by CSIRO Julius Career Award, CSIRO Early Research Career Fellowship, CSIRO Future Digital Manufacturing Fund Postdoc Fellowship, CSIRO Data61 PhD Scholarship to SW, and partially by Australian Research Council Discovery Grant (DP220100800) to HL. + +# References + +[1] Siddharth Agarwal, Ankit Vora, Gaurav Pandey, Wayne Williams, Helen Kourous, and James McBride. Ford multi-AV seasonal dataset. The International Journal of Robotics Research, 39(12):1367-1376, 2020. 5 +[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 8 +[3] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 5 +[4] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1, 2, 7 +[5] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5 +[6] Google. Maps static api, 2023. 2023. 1 +[7] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 1, 2 +[8] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[9] Ted Lentsch, Zimin Xia, Holger Caesar, and Julian FP Kooij. Slicematch: Geometry-guided aggregation for cross-view pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17225-17234, 2023. 2, 5, 6 +[10] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2 +[11] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 3 +[12] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 3 +[13] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios + +Balntas. OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In CVPR, 2023. 2 +[14] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6 +[15] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation estimation by cross-view matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064-4072, 2020. 1, 2, 6 +[16] Yujiao Shi, Xin Yu, Liu Liu, Tong Zhang, and Hongdong Li. Optimal feature transport for cross-view image geolocation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11990-11997, 2020. 2 +[17] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer, 2023. 2, 3, 6, 7, 8 +[18] Shinya Sumikura, Mikiya Shibuya, and Ken Sakurada. Openvslam: A versatile visual slam framework. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2292-2295, 2019. 5 +[19] Xiaoyang Tian, Jie Shao, Deqiang Ouyang, and Heng Tao Shen. Uav-satellite view synthesis for cross-view geolocation. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4804-4815, 2021. 1, 2 +[20] Aysim Token, Qunjie Zhou, Maxim Maximov, and Laura Leal-Taixe. Coming down to earth: Satellite-to-street view synthesis for geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2021. 1, 2 +[21] Frank Van Diggelen and Per Enge. The world's first gps mooc and worldwide laboratory using smartphones. In Proceedings of the 28th international technical meeting of the satellite division of the institute of navigation (ION GNSS+ 2015), pages 361-369, 2015. 2 +[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[23] Shan Wang, Chuong Nguyen, Jiawei Liu, Kaihao Zhang, Wenhan Luo, Yanhao Zhang, Sundaram Muthu, Fahira Afzal Maken, and Hongdong Li. Homography guided temporal fusion for road line and marking segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1075-1085, 2023. 1 +[24] Shan Wang, Yanhao Zhang, Akhil Perincherry, Ankit Vora, and Hongdong Li. View consistent purification for accurate cross-view localization, 2023. 1, 2, 3, 5, 6, 7, 8 +[25] Shan Wang, Yanhao Zhang, Ankit Vora, Akhil Perincherry, and Hengdong Li. Satellite image based cross-view localization for autonomous vehicle. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3592-3599. IEEE, 2023. 1, 2, 5, 6, 7 +[26] Xiaolong Wang, Runsen Xu, Zuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using + +a correlation-aware homography estimator. arXiv preprint arXiv:2308.16906, 2023. 1, 2, 3, 6 +[27] Zimin Xia, Olaf Booij, Marco Manfredi, and Julian FP Kooij. Visual cross-view metric localization with dense uncertainty estimates. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 90-106. Springer, 2022. 2 +[28] Zimin Xia, Olaf Booij, and Julian FP Kooij. Convolutional cross-view pose estimation. arXiv preprint arXiv:2303.05915, 2023. 2, 5, 6, 7 +[29] Zelong Zeng, Zheng Wang, Fan Yang, and Shin'ichi Satoh. Geo-localization via ground-to-satellite cross-view image retrieval. IEEE Transactions on Multimedia, 25:2176-2188, 2023. 1, 2 +[30] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021. 1, 2 +[31] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1162-1171, 2022. 1, 2 \ No newline at end of file diff --git a/2024/View From Above_ Orthogonal-View aware Cross-view Localization/images.zip b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..331a77ead84e8a8c68551f14095c1e0fe8d7fbd6 --- /dev/null +++ b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ad50477283ebd94aeb2be90ae7a7ea26ede50a99a3c6c86cd15f7c09a4f612b +size 686155 diff --git a/2024/View From Above_ Orthogonal-View aware Cross-view Localization/layout.json b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2bab56a687dcc36e515dcff31c8cbcea98fe8c7 --- /dev/null +++ b/2024/View From Above_ Orthogonal-View aware Cross-view Localization/layout.json @@ -0,0 +1,8144 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 120 + ], + "type": "text", + "content": "View From Above: Orthogonal-View aware Cross-view Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 485, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 485, + 185 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 485, + 185 + ], + "type": "text", + "content": "Shan Wang1,2 Chuong Nguyen1 Jiawei Liu2 Yanhao Zhang2 Sundaram Muthu1 Fahira Afzal Make1 Kaihao Zhang2 Hongdong Li2 1Data61, CSIRO 2Australian National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 550 + ], + "type": "text", + "content": "This paper presents a novel aerial-to-ground feature aggregation strategy, tailored for the task of cross-view image-based geo-localization. Conventional vision-based methods heavily rely on matching ground-view image features with a pre-recorded image database, often through establishing planar homography correspondences via a planar ground assumption. As such, they tend to ignore features that are off-ground and not suited for handling visual occlusions, leading to unreliable localization in challenging scenarios. We propose a Top-to-Ground Aggregation (T2GA) module that capitalizes aerial orthographic views to aggregate features down to the ground level, leveraging reliable off-ground information to improve feature alignment. Furthermore, we introduce a Cycle Domain Adaptation (CycDA) loss that ensures feature extraction robustness across domain changes. Additionally, an Equidistant Re-projection (ERP) loss is introduced to equalize the impact of all keypoints on orientation error, leading to a more extended distribution of keypoints which benefits orientation estimation. On both KITTI and Ford Multi-AV datasets, our method consistently achieves the lowest mean longitudinal and lateral translations across different settings and obtains the smallest orientation error when the initial pose is less accurate, a more challenging setting. Further, it can complete an entire route through continual vehicle pose estimation with initial vehicle pose given only at the starting point." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 687 + ], + "type": "text", + "content": "Visual-based cross-view localization aims to locate query images taken from street-level cameras (referred to as ground view) within a satellite or aerial view map. Platforms like the Google Map API [6] have made satellite images accessible, spurring the development of cross-view localization techniques. Notable studies [4, 7, 10, 15, 20, 25, 26, 30] have focused on using satellite images for this task. Despite this, accurate localization remains challenging due" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 462, + 293 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 462, + 293 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 462, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 462, + 293 + ], + "type": "image", + "image_path": "4c2274fc7cca192d79b8542da93b13793098cf7c5e1606d7a9a9a9740916040c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 296, + 410, + 304 + ], + "lines": [ + { + "bbox": [ + 360, + 296, + 410, + 304 + ], + "spans": [ + { + "bbox": [ + 360, + 296, + 410, + 304 + ], + "type": "text", + "content": "Ground View" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 463, + 213, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 463, + 213, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 463, + 213, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 463, + 213, + 545, + 293 + ], + "type": "image", + "image_path": "48b30efe7caab7dabfe78e7bcbcfdca01a7a7924e63a57ea63fca10d5b3bfd1d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 296, + 526, + 304 + ], + "lines": [ + { + "bbox": [ + 482, + 296, + 526, + 304 + ], + "spans": [ + { + "bbox": [ + 482, + 296, + 526, + 304 + ], + "type": "text", + "content": "Aerial View" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 308, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 407 + ], + "type": "text", + "content": "Figure 1. Efficacy of T2GA. T2GA aggregates off-ground features (e.g., the streetlight within the green square) and addresses occlusions (e.g., the tree within the red square). Unlike conventional methods that often overlook such elements, T2GA integrates these features, thus improving appearance matching for ground-level aligned pixels across views using ground plane homography. Insets are magnified for clarity, with dotted outlines indicating appearance misalignment and solid outlines indicating appearance alignment with the corresponding aerial view features." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 423, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 472 + ], + "type": "text", + "content": "to the significant viewpoint differences between ground and aerial images [14, 31]. These viewpoint differences result in a domain gap, adversely impacting feature alignment, thereby compromising the overall accuracy of localization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "Recent research has explored two main approaches to bridge the domain gap in image-based localization: generative-based and geometry-alignment-based methods. Generative-based methods, such as those utilizing GANs [19] and diffusion models [29], reduce the domain gap by transforming view styles from one view to another. However, the generated features for matching can lead to ambiguities in pose estimation. In contrast, geometry-alignment-based methods, employing techniques like polar transformations [15, 20] or homography [14, 23, 24, 26], focus on establishing correspondences for on-ground pixels. This often results in the neglect of off-ground features, such as streetlights, and difficulty in handling visual occlusions, such as the obscuring of road details by treetops in aerial views, both illustrated in Fig. 1. This neglect fails to leverage important geographic landmarks on the road and results in a lack of robustness to issues like road mark degradation (e.g. fading and damaged paintings). To address these limitations, we introduce the Top-to-Ground Aggregation (T2GA) module. T2GA employs top-down feature aggre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "Code is available at https://github.com/ShanWang-Shan/ViewFromAbove." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14843" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "gation to enrich on-ground points with an above-view perspective appearance, significantly improving feature alignment and localization accuracy. Furthermore, we recognize that pixel positions above ground points may not necessarily indicate higher elevations; instead, they could represent objects located further along the camera's line of sight. To resolve this ambiguity, we integrate a transformer mechanism that assesses whether these pixels belong to the same object and determines occlusion precedence, such as shadows on the road, which suggest likely occlusions in the aerial view." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 312 + ], + "type": "text", + "content": "In addition to viewpoint differences, ground and aerial views differ in cameras types, lighting conditions, tones, and resolutions. We introduce the Cycle Domain Adaptation (CycDA) loss function to address these variations. CycDA enables bidirectional feature generation between ground and aerial views. By minimizing the discrepancy between domain-adapted features and their target counterparts, our approach ensures that features extracted from one domain are effectively translatable to the other, fostering the extraction of features that are invariant across domains." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 314, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 289, + 422 + ], + "type": "text", + "content": "We further introduce an Equidistant Re-projection Loss to address a common bias in keypoint-synchronous detection localization methods, which tend to favor closer keypoints due to their reduced orientation errors. Our loss function mitigates this issue by applying a distance-weighted approach, ensuring that orientation errors are independent of keypoint distance. Consequently, keypoints are more uniformly distributed across various distances, leading to a more equitable and precise estimation of direction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 423, + 238, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 423, + 238, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 423, + 238, + 434 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 287, + 555 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "type": "text", + "content": "- A top-to-ground feature aggregation module that effectively bridges the domain gap between ground and aerial views by utilizing off-road cues and handling occlusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 473, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 287, + 507 + ], + "type": "text", + "content": "- A cycle domain adaptation loss function that promotes domain-invariant feature extraction, enhancing the robustness of cross-view localization methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "type": "text", + "content": "- An equidistant re-projection loss function ensures orientation errors are consistent regardless of keypoint distance, leading to a more extended distribution of keypoints and more accurate orientation estimation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "The evaluations are conducted on KITTI and Ford Multi-AV datasets while leveraging the Google Maps as corresponding satellite images. Experimental results demonstrate that our method can achieve consistent vehicle pose estimation under various challenging situations. Our method achieves the lowest mean longitudinal and lateral translation errors on both KITTI and Ford Multi-AV datasets under different settings. We also obtain the smallest orientation error when the initial pose is less accurate, demonstrating the robustness of our method. Further, with initial vehicle pose only at the starting point, our method can complete the route through continuous pose estimation, demonstrating its generalisation ability and potential for practical deployment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 281 + ], + "type": "text", + "content": "Image-level cross-view localization. Early visual-only cross-view localization methods [7, 10, 15, 16, 20, 30] approach the task as an image retrieval problem, focusing on coarse localization through image-to-image matching. To bridge the domain gap between ground and aerial views, various techniques have been proposed to facilitate cross-view feature matching. [10] incorporated per-pixel orientation information, while [15] and [20] utilized a predefined polar transform to align aerial-view images with ground views. Additionally, [19] and [29] employed GAN-based and diffusion model-based style transformations, respectively. While efforts have been made to minimize the domain gap, these methods rely on image-level feature matching, restricting their localization accuracy, often falling short of the precision achieved by commercial GPS systems in open areas [21]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 283, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 450 + ], + "type": "text", + "content": "Patch-wise cross-view localization. To improve accuracy, several methods employ patch-wise feature matching, especially effective in aerial views with wide fields of view and high resolutions. For instance, [31] employs transformers to emphasize informative patches, [27, 28] computes dense spatial distributions using patch attention, and [9] introduces the 'slice-to-sector match' and 'Cross-view attention' to calculate similarity between aerial sectors and ground view slices. However, their attention queries are derived from aerial images, posing reliability issues when the pose is unknown. OrienterNet [13] transforms ground view images into Bird's Eye View (BEV) grids for matching with OpenStreetMap data. Despite these advancements, localization accuracy remains limited by the patch (grid) size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": "Pixel-wise cross-view localization. Pixel-wise feature matching has been explored in various methods [4, 14, 24, 25] for precise localization, yet they struggle with the inherent domain gap between views. Our baseline, PureACL[24], and other geometry-alignment-based methods [14, 26] prioritize on-ground pixel correspondences but overlook off-ground features and occlusions, leading to suboptimal performance. CVGL [4] transforms the ground view image into a BEV for aerial matching. Their BEV transformation employs aerial coordinates as queries, introducing a high degree of freedom and increasing the complexity of matching. In the BoostAcc [17] framework, homography transformation is used on ground views to source query pixel data for the pixel-to-slice attention mechanism. This transformation introduces distortions from non-ground pixels, potentially exacerbating errors in subsequent processing stages. Moreover, the keys and values span entire and adjacent columns, risking compounded ambiguities in both longitudinal and lateral estimation. Our approach aims to address these challenges, proposing a novel pixel-to-overhead-column attention mechanism for improved cross-view localization accuracy." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14844" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 165 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 165 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 165 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 165 + ], + "type": "image", + "image_path": "b8c62cdf43ff5f11886d258e92cc1ca6325f8d87ff7c015e68783cb5c251e609.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 228 + ], + "type": "text", + "content": "Figure 2. Our method adopts PureACL [24] as the baseline model and introduces three novel components (highlighted in red): (1) T2GA (Sec.4.1): aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them to alleviate the representation gap of the same object across different views; (2) CycDA Loss (Sec.4.2): explicitly enforces a view-invariant representation for the same object; and (3) ERP Loss (Sec.4.3): allows the model to leverage key points that are farther away from the vehicle while allocating more emphasis on correcting the vehicle orientation estimation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 231, + 124, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 124, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 124, + 245 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "content": "Task Settings. The Fine-Grained Cross-View Localization (FGCVL) task aims to estimate the accurate 3-DoF pose of a vehicle, including longitudinal and lateral translations and orientation on a satellite map " + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "inline_equation", + "content": "I^{s}" + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "content": ", given an initial coarse pose " + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{init}" + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "content": ". It also assumes that the vehicle is equipped with on-board camera(s) " + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "content": ", and the image captured by these cameras, denoted as the ground image " + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "inline_equation", + "content": "I^{g}" + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "content": ", is accessible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": "Baseline Model. PureACL [24] is adopted as the baseline model for our approach. It employs a share-weight U-Net to extract ground/satellite feature maps " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "F = \\{F^g, F^s\\}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " and ground/satellite confidence maps " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "C = \\{C^g, C^s\\}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " from the ground/satellite images " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "I = \\{I^g, I^s\\}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": ". The ground confidence map " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "C^g" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " is tasked to select the top-N most confident pixel-level features from " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "F^g" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " at the positions " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "p^g = \\{(u_i^g, v_i^g)\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": ". These features are then: (i) Projected from on-board camera coordinates to the satellite coordinates using " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "p_i^s(\\mathbf{P}) = K_s \\mathbf{P} K_g^{-1} p_i^{g,4}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "K_s" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "K_g" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " are the intrinsic matrices of satellite and on-board cameras, respectively. " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " is the initial pose " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{init}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " at the first iteration and the estimated pose " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " in the subsequent iterations. (ii) A weighted residual " + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "inline_equation", + "content": "w(\\mathbf{P}) \\times r(\\mathbf{P})" + }, + { + "bbox": [ + 46, + 335, + 287, + 528 + ], + "type": "text", + "content": " is computed between the top-N selected features from the ground feature map and the corresponding features from the satellite feature map as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 534, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 534, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 52, + 534, + 287, + 548 + ], + "type": "interline_equation", + "content": "w (\\mathbf {P}) = C ^ {s} \\left[ p ^ {s} (\\mathbf {P}) \\right] \\times C ^ {g} \\left[ p ^ {g} \\right], r (\\mathbf {P}) = F ^ {s} \\left[ p ^ {s} (\\mathbf {P}) \\right] - F ^ {g} \\left[ p ^ {g} \\right], \\tag {1}", + "image_path": "442d52eafae41ca4f151c23acddb07aa9a363c164320235b0f581e828a0385cb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "content": " is a lookup operation in a feature/confidence map with sub-pixel interpolation. (iii) The Levenberg-Marquardt (LM) algorithm [11] takes the computed weighted residual " + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "w(\\mathbf{P})\\times r(\\mathbf{P})" + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "content": " as input and outputs the predicted vehicle pose " + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}^{m}" + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{th}}" + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "content": " iteration, with " + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "m\\in" + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\{1,\\ldots ,M\\}" + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "text", + "content": ". The process of (i) - (iii) repeats for " + }, + { + "bbox": [ + 47, + 555, + 288, + 629 + ], + "type": "inline_equation", + "content": "M = 20" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 232, + 507, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 507, + 244 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 507, + 244 + ], + "type": "text", + "content": "iterations to produce the final vehicle pose " + }, + { + "bbox": [ + 306, + 232, + 507, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}" + }, + { + "bbox": [ + 306, + 232, + 507, + 244 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "type": "text", + "content": "PureACL [24] employs two loss functions: a Triplet Loss [12] and a Re-projection Loss. The former supervises the weighted residual, enforcing that the encoder extracts pose-sensitive features from both ground and satellite images. The later utilises the projection error w.r.t the ground truth pose " + }, + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{gt}" + }, + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "type": "text", + "content": " to penalize incorrectly predicted vehicle pose " + }, + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}" + }, + { + "bbox": [ + 304, + 245, + 545, + 329 + ], + "type": "text", + "content": " (Refer to Supplementary Sec.D for more details)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 341, + 361, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 361, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 361, + 354 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 458 + ], + "type": "text", + "content": "We propose a novel View From Above (VFA) method to tackle the FGCVL task by aligning feature representations across the orthogonal viewpoints, the ground view and the satellite views. As illustrated in Fig. 2, our method introduces three innovative components onto the baseline model: (1) Top-to-Ground Aggregation Module (T2GA) (Sec. 4.1); (2) Cycle Domain Adaptation Loss (CycDA) (Sec. 4.2); and (3) Equidistant Re-Projection Loss (Sec. 4.3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 467, + 459, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 459, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 459, + 480 + ], + "type": "text", + "content": "4.1. Top to Ground Aggregation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "The appearance discrepancy between the same objects viewed from ground and satellite perspectives presents a significant challenge in achieving high precision in the FGCVL task [14, 17, 24, 26]. This issue is often encountered on tall structures, e.g. street light, and the occlusion from aerial objects, e.g. tree branches, where their near(on)-ground appearance, accessible to only on-board vehicle camera, is significantly different from their top appearances which are only registered by the satellite camera (see Fig. 1). Such discrepancies are common on roads and can lead to inaccuracies in cross-view localization. To address this challenge, we propose aligning the representations between ground and satellite views, focusing specifically on the on-ground pixels, where geometry alignment across the two views hold due to the ground plane homography. Specifically, we present a T2GA module that aggregates the features of elevated pixels onto the feature of the on-ground pixel that is directly beneath them, as illustrated in the 'T2GA' red box in Fig. 2. For an on-ground pixel:" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 635, + 286, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 286, + 654 + ], + "type": "text", + "content": "2While our method is compatible with multiple cameras, our primary experiments focus on one front camera and four surrounding cameras." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 654, + 286, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 674 + ], + "type": "text", + "content": "3We simply refer to the fused confidence map, rather than distinguishing between view-consistent and on-ground confidence maps in PureACL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "text", + "content": "PureACL obtains 3D world coordinates of the vehicle from homogeneous coordinates " + }, + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "inline_equation", + "content": "K_{g}^{-1}p^{g}" + }, + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "text", + "content": " by utilizing on-ground characteristics. For simplicity in this context, we note homogeneous coordinates as directly corresponding to the 3D world coordinates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "14845" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "p\\in \\{(u,v)\\}_{u = 0,v = \\tau}^{W^g -1,H^g -1}" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " , and where " + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " is a threshold corresponding to the height of horizon in the ground view image, and " + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "W^{g}" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "H^{g}" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " are the width and height of the ground view image, respectively. The corresponding elevated pixels are " + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "q\\in \\{(u_p,v)\\}_{v = 0}^{v_p}" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "(u_{p},v_{p})" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " are the coordinates of the on-ground pixel " + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 69, + 289, + 156 + ], + "type": "text", + "content": " . The aggregation is achieved via the attention mechanism [22], defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 165, + 287, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 165, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 287, + 179 + ], + "type": "interline_equation", + "content": "F ^ {a t t} [ p ] = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {T}\\right) \\mathbf {V}, \\tag {2}", + "image_path": "f4470a43898ab0d16a3a85abd45541fe1743c96a0b38e8601a11ec2d277591fd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "F^{att}[p]" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " represents the feature vector at a pixel " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " on the feature map " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "F^{att}" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": ". In this context, " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " denote an on-ground pixel and its corresponding elevated pixels, respectively. The query is formed by taking the features of on-ground pixels " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathcal{M}(F^g[p])" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " represents a non-linear mapping consisting of an MLP layer followed by an activation function. The value is derived from the features of elevated pixels corresponding to the on-ground pixel " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = F^g[q]" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": ", and the key is then obtained by applying the same non-linear mapping to " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathcal{M}(\\mathbf{V})" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": ". This non-linear mapping ensures that the query and key are mapped to the same feature space before computing their matrix product. " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{K}^T" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": ". Notably, value " + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 189, + 289, + 380 + ], + "type": "text", + "content": " is used directly in its original form for the computation, facilitating a straightforward fusion (for the same object) or replacement (in cases of occlusion) at ground level." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 392, + 292, + 494 + ], + "blocks": [ + { + "bbox": [ + 50, + 392, + 292, + 494 + ], + "lines": [ + { + "bbox": [ + 50, + 392, + 292, + 494 + ], + "spans": [ + { + "bbox": [ + 50, + 392, + 292, + 494 + ], + "type": "image", + "image_path": "21d5ec7a225c4edfc15de4afa07b9b409317055b7e36f3ec5aa957ce9a2935a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 497, + 287, + 606 + ], + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 606 + ], + "type": "text", + "content": "Figure 3. Attention between the on-ground pixels and their corresponding elevated pixels is displayed column-wise. (Left example): The attention between the base and top of the streetlight is high despite their distinct appearances. This allows the resultant aggregated ground feature to be aligned with the satellite feature corresponding to the matching geographic location. (Right example): In the absence of occlusion from above, ground pixels maintain minimal attention with their elevated pixels. This avoids unnecessary dilution of the ground features which are already well aligned with their satellite counterparts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 614, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 288, + 664 + ], + "type": "text", + "content": "The computed attention map " + }, + { + "bbox": [ + 46, + 614, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\operatorname{Softmax}(\\mathbf{Q}\\mathbf{K}^T)" + }, + { + "bbox": [ + 46, + 614, + 288, + 664 + ], + "type": "text", + "content": " can accurately reflect the connection between the on-ground pixels and their corresponding elevated pixels. As illustrated in Fig. 3, high attention values are assigned to the base and top" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "of the streetlight despite their different appearances. This provides important cues to alleviate the representation gap between the ground feature and satellite features. On the other hand, in the absence of occlusion, e.g. tall structure and tree branch, thus the attention between on-ground and elevated pixels is minimal, avoiding diluting the ground features that are well aligned with their satellite counterparts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "content": "The feature map " + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "inline_equation", + "content": "F^{att}" + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "content": " after attention calculation is vertically stacked with the upper part of the original " + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "inline_equation", + "content": "F^g" + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "content": " to generate the aggregated feature map " + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "inline_equation", + "content": "F^a" + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 201, + 545, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 201, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 376, + 201, + 545, + 216 + ], + "type": "interline_equation", + "content": "F ^ {a} = F _ {0: H ^ {g} - \\tau} ^ {g} \\oplus F ^ {a t t}, \\tag {3}", + "image_path": "89f65ebd5b3a5d3f2e58fc17d229022024c96382a435347a6015969e95053ac1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": " denotes vertically stacking and " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "F_{0:H^g -\\tau}^g" + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": " indicated the slicing of " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "F^{g}" + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": " from row 0 to " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "H^{g} - \\tau" + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": " . Subsequently, this aggregated feature map " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "F^a" + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": " replaces " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "F^g" + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "text", + "content": " in the residual calculation as per Eq.1, resulting in an updated formula: " + }, + { + "bbox": [ + 304, + 224, + 545, + 284 + ], + "type": "inline_equation", + "content": "r(\\mathbf{P}) = F^{s}[p^{s}(\\mathbf{P})] - F^{a}[p^{g}]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 285, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 546, + 464 + ], + "type": "text", + "content": "The advantage of the aggregated feature map is also indirectly reflected by the ground confidence map of the baseline model as shown in Fig. 4. Without the T2GA module, the confidence map only highlights pixels corresponding to road marks and curds, resulting in a concentrated sampling of key points for subsequent estimation of vehicle pose. This has the following drawbacks: (1) fail to leverage road landmarks e.g. traffic signal poles, that encode important geographic location; and (2) not robust to road mark degradation problems, e.g. fading and damaged paintings, and visibility issues, e.g. glow on road marks. T2GA alleviates the feature discrepancy between ground and satellite views. As a result, the traffic signal pole receives higher confidence, allowing it to be subsequently sampled as key points to estimate more precise vehicle pose." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 474, + 429, + 514 + ], + "blocks": [ + { + "bbox": [ + 309, + 474, + 429, + 514 + ], + "lines": [ + { + "bbox": [ + 309, + 474, + 429, + 514 + ], + "spans": [ + { + "bbox": [ + 309, + 474, + 429, + 514 + ], + "type": "image", + "image_path": "0a68c878c1dd72a93fadf4187eeeb97cdb6f7b5ab34270c603c6b8b7618100bf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 525, + 545, + 624 + ], + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 624 + ], + "type": "text", + "content": "Figure 4. Illustration on the effect of T2GA on the confidence map of the baseline model. The confidence map without T2GA (Left) predominantly highlights road marks and curbs, resulting in subsequent keypoint sampling missing important road landmarks, e.g. traffic signal poles, that provide important cues to vehicle pose estimation; The confidence map with T2GA (Right) has high-confidence values distributed across various road marks and traffic poles. With more geographic cues provided by multiple sources, the resultant pose prediction becomes more precise and robust." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 474, + 550, + 513 + ], + "blocks": [ + { + "bbox": [ + 430, + 474, + 550, + 513 + ], + "lines": [ + { + "bbox": [ + 430, + 474, + 550, + 513 + ], + "spans": [ + { + "bbox": [ + 430, + 474, + 550, + 513 + ], + "type": "image", + "image_path": "b0f56c8b54e0dc05d3122eb5250d2d7ca74407891b3a3902469405b916a275d2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 635, + 474, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 474, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 474, + 647 + ], + "type": "text", + "content": "4.2. Cycle Domain Adaptation Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "There are multiple sources responsible for the representation gap between ground and satellite features. In addition to the varying appearances of the same object across different views, camera specifications, e.g. tone, hue, intensity, brightness and resolution, and temporal changes of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "text", + "content": "In processing all query pixels, we first compute the matrix multiplication " + }, + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{K}^T" + }, + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "text", + "content": " using all column pixels " + }, + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "inline_equation", + "content": "q\\in \\{(u_p,v)\\}_{v = 0}^{H^g -1}" + }, + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "text", + "content": ". We then eliminate the attention values corresponding to pixels below each query pixel " + }, + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{(u_p,v)\\}_{v = v_p + 1}^{H^g -1}" + }, + { + "bbox": [ + 47, + 670, + 287, + 715 + ], + "type": "text", + "content": ". After this elimination, the Softmax function is applied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14846" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "varying spans, can also lead to inconsistent representations for the same geographic location. Such issue is in general overlooked by the existing FGCVL methods [14, 28]. Despite introducing a triplet loss to further distinguish feature representations based on different poses, our baseline model, PureACL [24], falls short in enforcing invariant feature representation at the corresponding geographic locations across different views." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 228 + ], + "type": "text", + "content": "We propose a Cycle Domain Adaptation (CycDA) loss to explicitly enforce view-invariant representations that consists of three L2-loss between ground and satellite feature representations in three different feature spaces. The representation loss in the ground feature space is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 244, + 287, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 244, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 52, + 244, + 287, + 273 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| F ^ {a} \\left[ p _ {i} ^ {g} \\right] - \\mathcal {D} \\left(\\mathcal {E} \\left(F ^ {s} \\left[ p _ {i} ^ {s} \\left(\\mathbf {P} _ {g t}\\right) \\right] \\odot c _ {s}\\right) \\odot c _ {g}\\right) \\right\\| _ {2}, \\tag {4}", + "image_path": "213be681f907a535dccd4f6ac42886ee443205eec9c2bf485f4cdde1792a4834.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "content": " is a projection from the pixel feature space to a latent feature space, " + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "content": " is a projection from the latent feature space to the pixel feature space, " + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "inline_equation", + "content": "c_{g}" + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "content": " represent the focal length of satellite camera and vehicle camera respectively which are introduced to guide the projection functions, and " + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 280, + 287, + 376 + ], + "type": "text", + "content": " is a concatenation acting on the feature channel dimension. Similarly, the representation loss in the satellite feature space is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 383, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 383, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 52, + 383, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| F ^ {s} \\left[ p _ {i} ^ {s} \\left(\\mathbf {P} _ {g t}\\right) \\right] - \\mathcal {D} \\left(\\mathcal {E} \\left(F ^ {a} \\left[ p _ {i} ^ {g} \\right] \\odot c _ {g}\\right) \\odot c _ {s}\\right) \\right\\| _ {2}, \\tag {5}", + "image_path": "77af4b21b0fe587d5f81ea426b595ae37c314e55ace26622217c3ce5030a7839.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 455 + ], + "type": "text", + "content": "We also enforce the satellite feature to be aligned with the ground feature corresponding to the same geographic location in a latent feature space through:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 461, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 461, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 461, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| \\mathcal {E} \\left(F ^ {a} [ p _ {i} ^ {g} ] @ c _ {g}\\right) - \\mathcal {E} \\left(F ^ {s} [ p _ {i} ^ {s} (\\mathbf {P} _ {g t}) ] @ c _ {s}\\right) \\right\\| _ {2}. \\tag {6}", + "image_path": "5b07cc6ee3b164052e2a93afd7bc43e116b8e89790c125bb3d805b6a09115865.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 496, + 239, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 239, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 239, + 509 + ], + "type": "text", + "content": "The overall CycDA loss can then be defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 517, + 287, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 517, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 111, + 517, + 287, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C y c D A} = \\mathcal {L} _ {g} + \\mathcal {L} _ {s} + \\mathcal {L} _ {m} \\tag {7}", + "image_path": "920d1c90ea9d572464e66185cb3dba34adc583be74aac14a8e4545e5cad84fdf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 538, + 216, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 216, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 216, + 552 + ], + "type": "text", + "content": "4.3. Equidistant Re-projection Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": "Our method tackles a challenging setting of simultaneously detecting key points from the ground image and predicting the vehicle pose using the detected key points. Our baseline model introduces a re-projection loss computed over the selected top-N key points defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 624, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 624, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 83, + 624, + 287, + 658 + ], + "type": "interline_equation", + "content": "L _ {R P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| p _ {i} ^ {s} (\\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\\mathbf {P} _ {g t}) \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "1175bc95fed255ea7ab1aefae6a31b8989270a221c6ffcb5704ac377cb7bb1c9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_i^s (\\mathbf{P}_{pred})" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_i^s (\\mathbf{P}_{gt})" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " are the projected coordinates on the satellite image from the selected top-N ground key points with the predicted vehicle pose and the groundtruth vehicle pose respectively, and " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " denotes the L2-distance." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 428, + 109 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 428, + 109 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 428, + 109 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 428, + 109 + ], + "type": "image", + "image_path": "9a278da9d73fe47a641ce4bc30610a218894337cb1901b491a0a06e10daa1113.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 118, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 118, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 547, + 163 + ], + "type": "text", + "content": "Figure 5. Comparison of detected keypoints with and without ERP Loss. (Left) Without the ERP Loss, keypoints are predominantly located in close proximity to the vehicle. (Right) With the ERP Loss, there is a more dispersed distribution of keypoints." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 429, + 71, + 550, + 109 + ], + "blocks": [ + { + "bbox": [ + 429, + 71, + 550, + 109 + ], + "lines": [ + { + "bbox": [ + 429, + 71, + 550, + 109 + ], + "spans": [ + { + "bbox": [ + 429, + 71, + 550, + 109 + ], + "type": "image", + "image_path": "d7b72283fb9b53d1a00eeefca7e1fbadd8b63ecf1e517fef4169d3bd492e30cc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 168, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 239 + ], + "type": "text", + "content": "Despite being effective in pipelines where simultaneous detection of key points is not required [3, 18, 25], the reprojection loss tends to overly penalize the orientation errors of distant points. This forces the detected key points being closer to the vehicle, as illustrated in Fig. 5, and less accurate vehicle orientation estimation (see Sec. 5.2)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 240, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 264 + ], + "type": "text", + "content": "To mitigate the aforementioned issue, we propose an Equidistant Re-Projection (ERP) loss that is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 337, + 272, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 337, + 272, + 545, + 304 + ], + "type": "interline_equation", + "content": "L _ {E R P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\frac {p _ {i} ^ {s} (\\mathbf {P} _ {p r e d}) - p _ {i} ^ {s} (\\mathbf {P} _ {g t})}{\\mathbf {D} _ {p _ {i} ^ {s}}} \\| _ {2} ^ {2}, \\tag {9}", + "image_path": "53bababf905540d93fae00d3254e114af7a87de7228f932b9fed2b183cbff4e1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{p_i^s} = \\| p_i^s (\\mathbf{P}_{gt}) - p_{vehicle}^s (\\mathbf{P}_{gt})\\|_2" + }, + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "type": "text", + "content": " is the L2-distance between the satellite image coordinates of the " + }, + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 312, + 546, + 408 + ], + "type": "text", + "content": " keypoint and the vehicle. The design prevents the orientation error from scaling with the L2-distance between key points and vehicle on the satellite image, which dominates over the orientation error on distant key points. As shown in Fig. 5, our method can leverage key points that are farther from the vehicle than our baseline model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 419, + 387, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 387, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 387, + 433 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 438, + 438, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 438, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 438, + 451 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 457, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 672 + ], + "type": "text", + "content": "Datasets. To evaluate our proposed approach, we conduct experiments on two well-established autonomous driving datasets: the Ford Multi-AV Seasonal dataset (FMAVS) [1] and the KITTI dataset [5]. Consistent with established practices [14, 24, 25], our primary focus is on images from the front left camera as query images. Additionally, following the methodology of PureACL [24], we broadened our analysis to include four camera perspectives: front left, rear right, side left, and side right. For data partitioning in the KITTI-CVL dataset, we utilized an approach in line with HighlyAcc [14], which involves two test splits. The first, 'Same,' includes images from the same trajectories as the training dataset, while the second, 'Cross,' comprises images from distinct trajectories. Regarding the FMAVS dataset, we follow the partitioning strategy proposed by PureACL [24], using all images from the same trajectories but different traversals. It is noteworthy that there are minor route variations within these same trajectories6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 672, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 672, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 306, + 672, + 545, + 696 + ], + "type": "text", + "content": "Metrics. We adopt evaluation metrics in line with HighlyAcc [14], SliceMatch [9], SIBCL [25], and PureACL [24]." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 481, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 481, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 481, + 713 + ], + "type": "text", + "content": "An example is provided in Supplementary Fig. 1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14847" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 105, + 547, + 323 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "text", + "content": "Table 1. Comparison on the KITTI-CVL Dataset Under Initial Noise Conditions (±10°,±20m). Results marked with " + }, + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "text", + "content": " are sourced directly from the respective original papers. Those denoted by " + }, + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "text", + "content": " indicate retraining of methods for consistent evaluation criteria alignment. - indicates that the corresponding data were not provided." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 105, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 47, + 105, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 547, + 323 + ], + "type": "table", + "html": "
ModelLocation(m)↓Lateral(%)↑Longit(%)↑Orient(°)↓Orient(%)↑
MeanMedianr@1mr@3mr@5mr@1mr@3mr@5mMeanMedianr@1°r@3°r@5°
Same★DSM[15]--10.1230.6748.244.0812.0120.14--3.5813.8124.44
★HighlyAcc[14]12.0811.4235.5470.7780.365.2215.8826.133.722.8319.6451.7671.72
★SliceMatch[9]7.964.3949.09-98.5215.19-57.354.123.6513.41-64.17
★BoostAcc[17]10.015.1976.4496.3498.8923.5450.5762.180.550.4299.10100.00100.00
★CCVPE [28]1.220.6297.3598.6599.7177.1396.0897.160.670.5477.3999.4799.95
★HC-Net [26]0.800.5099.01-99.7392.20-99.250.450.3391.35-99.84
◇PureACL [24]2.420.4291.9593.4094.2891.8692.1292.383.971.7732.7153.8971.66
Ours0.200.1799.9799.9799.9799.9799.9799.971.530.9350.9583.9795.45
Cross★DSM[15]--10.7731.3748.243.8711.7319.50--3.5314.0923.95
★HighlyAcc[14]12.5812.1127.8259.7972.895.7516.3626.483.953.0318.4249.7271.00
★SliceMatch[9]13.509.7732.43-86.448.30-35.574.206.6146.82-46.82
★BoostAcc[17]13.019.0657.7286.7791.1614.1534.5945.000.560.4398.98100.00100.00
★CCVPE [28]9.163.3344.0681.7290.2323.0852.8564.311.550.8457.7292.3496.19
★HC-Net [26]8.474.5775.00-97.7658.93-76.463.221.6333.58-83.78
◇PureACL [24]6.200.6167.2487.2692.7664.8173.6389.694.262.4823.4548.6959.24
Ours0.210.1899.9299.9699.9799.9199.9199.912.041.3838.3180.1292.48
", + "image_path": "c9b8e82a8d50285e0b04b770506da0258ede8cf9c324fe832dc567fd26b6f061.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": "Our precision assessment includes median and mean errors for overall, lateral, and longitudinal translations, as well as orientation accuracy. Additionally, our analysis covers lateral and longitudinal translations and localization recall at various distances (0.25m, 0.5m, 1m, 3m, and 5m), and orientation recall within a range of " + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "inline_equation", + "content": "1^{\\circ}" + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": "Training Details. We apply two sets of criteria: (1) Following HighlyAcc [14], ground images are processed at a resolution of " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "256 \\times 1024" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ", and satellite images at " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ", with initial pose noise " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\pm 10^{\\circ}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": " for orientation and " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\pm 20\\mathrm{m}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": " for translations. (2) In accordance with PureACL [24], we process FMAVS images at " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "432 \\times 816" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ", KITTI images at " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "384 \\times 1248" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ", and satellite images at " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "1,280 \\times 1,280" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ". To better simulate turning scenarios, we adopt an expanded initial pose noise range of " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\pm 45^{\\circ}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": " for orientation and " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\pm 20\\mathrm{m}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": " for translations. For training, we utilize an NVIDIA RTX 3090 GPU with a batch size of 3, employing the Adam optimizer [8] with a learning rate of " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ". Feature extractor weights are adapted from PureACL [24], while other components are initialized randomly. Training iterations average around " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "285\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": ", including " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "200\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": " dedicated to optimization. The inference speed, subject to initial pose variability, averages at " + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "inline_equation", + "content": "222\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 400, + 288, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 599, + 237, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 237, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 237, + 612 + ], + "type": "text", + "content": "5.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "We evaluate our method against recent visual-only approaches on the KITTI-CVL dataset following the metrics of HighlyAcc [14]. The results shown in Tab. 1 clearly indicate the superiority of our approach in spatial precision, consistently maintaining poses within a " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " radius in both 'Same' and 'Cross' areas with over " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "99.9 + \\%" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " probability. While our method may not lead in orientation accuracy under an initial pose range of " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pm 10^{\\circ}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": ", it demonstrates" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "text", + "content": "superior performance when the range is extended to " + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\pm 45^{\\circ}" + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "text", + "content": ", as detailed in Tab. 2. This underlines the robustness of our approach. Our approach leverages pixel-wise localization, offering finer granularity compared to the patchwise localization of SliceMatch[9], CCVPE [28], and the image-level localization of DSM [15]. Although HighlyAcc [14] and HC-Net [26] are pixel-wise localization methods, their homography-based mechanisms fail to incorporate off-road cues. In contrast, our approach effectively utilizes these cues, resulting in enhanced performance. Additionally, our longitudinal estimation significantly surpasses that of BoostAcc [17]. This improvement is likely attributable to our method's strategy of not aggregating information from pixels below, thereby reducing longitudinal ambiguity. In comparison to the baseline method PureACL [24], our approach achieves significant enhancements in both orientation and spatial accuracy, especially in mean error reduction. Specifically, we observe a " + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "inline_equation", + "content": "91^{+}\\%" + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "text", + "content": " reduction in spatial mean error and a " + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "inline_equation", + "content": "52^{+}\\%" + }, + { + "bbox": [ + 304, + 329, + 547, + 617 + ], + "type": "text", + "content": " reduction in orientation mean error. The difference between mean and median errors for PureACL indicates potential convergence issues in high-noise environments. By integrating the T2GA and CycDA modules, we bridge the domain gap and ensure convergence even under challenging conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 623, + 537, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 537, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 537, + 635 + ], + "type": "text", + "content": "5.3. Challenge Initial Pose and Stringent Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "To rigorously evaluate our method, we adopt the stringent metrics of PureACL [24], which are more demanding than those used by HighlyAcc[14]. We benchmark our method against baseline methods SIBCL [25] and PureACL [24], as well as SOTA methods CCVPE [28] and BoostAcc [17]. The results in Tab. 2 reaffirm the advantages of our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14848" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 116, + 547, + 368 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 2. Comparison with Initial Noise Conditions " + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "(\\pm 45^{\\circ},\\pm 20\\mathrm{m})" + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "text", + "content": " .For the KITTI-CVL dataset, evaluations are conducted in 'K-Same' and 'K-Cross' areas. On the Ford-CVL dataset, the Log4' trajectory, as used in SIBCL [25], is chosen for its optimal satellite view alignment. Additional log evaluations are detailed in the supplementary material. F-1C' and F-4C' represent assessments on the Ford-CVL dataset using either a single front-facing camera or four surrounding cameras. Note: " + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "text", + "content": " indicates that SIBCL [25] is a hybrid LiDAR-visual method." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 116, + 547, + 368 + ], + "lines": [ + { + "bbox": [ + 47, + 116, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 547, + 368 + ], + "type": "table", + "html": "
ModelLateral(m)↓Mean MedianLateral(%)↑r@0.25mr@0.5mr@1mLongit(m)↓Mean MedianLongit(%)↑r@0.25mr@0.5mr@1mOrient(°)↓Mean MedianOrient(%)↑r@1°r@2°r@4°
K-SameCCVPE [28]2.321.2211.0522.0542.096.493.357.4513.3624.973.431.5334.6660.7284.03
BoostAcc[17]1.190.6321.9241.5167.7210.015.425.179.3816.543.882.9818.4535.2562.20
★ SIBCL[25]1.230.7017.0133.3461.883.140.8620.6831.4055.588.901.6935.5055.0471.06
PureACL[24]2.940.2358.7566.3470.603.150.3745.6568.2374.878.132.8037.9952.5764.13
Ours0.170.1476.2097.7399.970.080.0798.1899.8999.972.191.0748.9766.2990.54
K-CrossCCVPE [28]4.562.794.9610.0538.1212.068.163.136.1412.4519.4615.393.3810.2417.41
BoostAcc[17]2.650.9314.8829.2052.6410.468.923.385.5611.735.614.2612.4624.8747.26
★ SIBCL[25]2.720.7116.9032.6258.245.481.0519.5830.7449.259.671.9625.0250.1860.16
PureACL [24]4.240.2550.2562.6364.563.890.447.4162.0464.2310.963.1619.9536.2556.49
Ours0.170.1575.9697.4799.970.090.0797.4999.8799.972.881.4436.1063.0886.18
F-1CCCVPE [28]4.622.456.4812.5424.4610.218.413.228.2413.5320.4815.073.146.1416.14
BoostAcc[17]2.731.479.0518.1835.5910.516.144.218.1315.566.824.7911.6122.6442.63
★ SIBCL[25]2.590.7120.9241.5360.435.381.1812.5923.9943.876.341.5735.2158.4170.65
PureACL[24]2.941.5411.3421.1437.764.751.7410.4620.0737.647.383.3315.1230.4257.01
Ours0.490.4131.7759.1789.230.360.3238.4375.7297.462.191.0847.3172.7190.10
F-4CCVGL[4]1.280.5021.8150.0586.841.530.7019.3637.2665.51-----
PureACL[24]1.690.6721.0739.5365.273.381.1812.5923.9943.863.071.1943.8473.2782.41
Ours0.130.1088.2598.53100.000.200.1868.6796.68100.002.001.4935.7262.7091.83
", + "image_path": "5e71760d777453cd03eabd6dce9024eed14eb0489e49ad2585969206f55301e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 383, + 167, + 445 + ], + "blocks": [ + { + "bbox": [ + 50, + 383, + 167, + 445 + ], + "lines": [ + { + "bbox": [ + 50, + 383, + 167, + 445 + ], + "spans": [ + { + "bbox": [ + 50, + 383, + 167, + 445 + ], + "type": "image", + "image_path": "803aa4bd5d4ce7de98c86077c936849f854378747adffb0d7364370b7415f09b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 450, + 129, + 459 + ], + "lines": [ + { + "bbox": [ + 88, + 450, + 129, + 459 + ], + "spans": [ + { + "bbox": [ + 88, + 450, + 129, + 459 + ], + "type": "text", + "content": "(a) Location" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 383, + 285, + 445 + ], + "blocks": [ + { + "bbox": [ + 173, + 383, + 285, + 445 + ], + "lines": [ + { + "bbox": [ + 173, + 383, + 285, + 445 + ], + "spans": [ + { + "bbox": [ + 173, + 383, + 285, + 445 + ], + "type": "image", + "image_path": "61468ab18c74c8c4334bbaa0adf3608827ee0a5aababe9613a7eb7987fa6f1ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 449, + 252, + 459 + ], + "lines": [ + { + "bbox": [ + 203, + 449, + 252, + 459 + ], + "spans": [ + { + "bbox": [ + 203, + 449, + 252, + 459 + ], + "type": "text", + "content": "(b) Orientation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 459, + 287, + 547 + ], + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 547 + ], + "type": "text", + "content": "Figure 6. Comparison of the baseline method PureACL and our proposed method under varying initial noise ranges in the 'cross' area of the KITTI-CVL dataset. The blue bar indicates the median value, while the red bar shows the difference between the median and mean values. Notably, our method demonstrates robust convergence, outperforming PureACL across all noise ranges, especially in larger noise scenarios where PureACL shows significant discrepancies between median and mean values." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": "method, particularly in terms of reduced mean error and improved longitudinal estimation. Notably, under conditions of minimal initial pose noise " + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "inline_equation", + "content": "(\\pm 15^{\\circ}" + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\pm 5\\mathrm{m})" + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": ", PureACL [24] outperforms SIBCL [25]. However, when faced with more challenging conditions " + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "inline_equation", + "content": "(\\pm 45^{\\circ}" + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\pm 20\\mathrm{m})" + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": ", PureACL [24] encounters convergence challenges, unlike our method, which demonstrates consistent superiority. This highlights the effectiveness of our proposed modifications in feature alignment. Additionally, we assess our method's convergence range against PureACL [24] in the 'Cross' area of the KITTI-CVL dataset, as depicted in Fig. 6. The red bar in the figure highlights a significant reduction in the gap between median and mean errors. This observation con" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 372, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 545, + 396 + ], + "type": "text", + "content": "firms our method's enhanced ability to converge effectively across a wider range of noise levels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 406, + 513, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 513, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 513, + 418 + ], + "type": "text", + "content": "5.4. Results with Continual Pose Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "content": "To address GPS signal loss, we adopt an accumulated pose estimation strategy that leverages initial poses based on the vehicle's previous pose estimates " + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "content": ". We use the model trained with initial noise allowances of " + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\pm 45^{\\circ}" + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\pm 20\\mathrm{m}" + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "content": ". Our method's robustness is evidenced by the continuous running distance percentage presented in Tab. 3, showing successful pose chaining throughout all evaluate routes. In contrast, BoostAcc [17] and the single-camera PureACL [24] exhibit significant drift, leading to errors beyond the satellite map's coverage and causing evaluations to cease after covering less than " + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "content": " of the distance. Performance comparisons on the KITTI-CVL dataset are shown in Fig. 7 " + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 304, + 425, + 545, + 665 + ], + "type": "text", + "content": ". Our approach consistently achieves accurate pose estimation or experiences only minor drifts in challenging scenarios characterized by limited localization cues or severe occlusions. In contrast, BoostAcc [17] exhibits significant drift when a moving vehicle passes by, likely due to incorrect query data from projected vehicle pixels. Furthermore, PureACL [24] appears to struggle with incorrect orientation, resulting in reversed pose estimations. These out" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "type": "text", + "content": "This pose estimation is purely frame-based and does not involve any sequence-based filtering techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "8Performance comparisons on the Ford-CVL dataset is shown in Fig. 1 and Fig. 2 of supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14849" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 79, + 166, + 167 + ], + "blocks": [ + { + "bbox": [ + 49, + 79, + 166, + 167 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 166, + 167 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 166, + 167 + ], + "type": "image", + "image_path": "dba6508b74ea847d65b37c6f59c14ef1ebfa503e182090c33be623c1b44d26b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 79, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 167, + 79, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 167, + 79, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 167, + 79, + 286, + 167 + ], + "type": "image", + "image_path": "d50ca463c12e7424e05b9a54a4ad8730d11f9bb30130cb4e506abef533043bd7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 168, + 167, + 253 + ], + "blocks": [ + { + "bbox": [ + 49, + 168, + 167, + 253 + ], + "lines": [ + { + "bbox": [ + 49, + 168, + 167, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 168, + 167, + 253 + ], + "type": "image", + "image_path": "63220183812e94646ecfc564e94df1d84abe5069491dc07688399c88daf2fba2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 345 + ], + "type": "text", + "content": "Figure 7. Accumulated pose estimation performance on a KITTI-CVL dataset trajectory. (Top-left) Predicted poses by our method (blue) closely match the ground truth (red) over the 3676-meter route, outperforming PureACL (green) and BoostAcc (cyan), which exhibit severe drift. (Top-right) Our approach demonstrates rapid recovery upon encountering clear localization cues. (Bottom-left) (Bottom-right) BoostAcc demonstrates substantial drift when encountering moving vehicles." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 168, + 286, + 253 + ], + "blocks": [ + { + "bbox": [ + 167, + 168, + 286, + 253 + ], + "lines": [ + { + "bbox": [ + 167, + 168, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 167, + 168, + 286, + 253 + ], + "type": "image", + "image_path": "83fa2f54b3f2ceb0d3a63586dde6dd8fc200e2b21f8265a85929c88f7afe519a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "type": "text", + "content": "comes highlight the importance of robustness in real-world applications, as drifts that extend beyond the coverage of satellite imagery can compromise all further pose estimation processes." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 55, + 416, + 280, + 483 + ], + "blocks": [ + { + "bbox": [ + 66, + 403, + 268, + 415 + ], + "lines": [ + { + "bbox": [ + 66, + 403, + 268, + 415 + ], + "spans": [ + { + "bbox": [ + 66, + 403, + 268, + 415 + ], + "type": "text", + "content": "Table 3. Comparison on Accumulated Pose Estimation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 416, + 280, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 416, + 280, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 280, + 483 + ], + "type": "table", + "html": "
ModelRunning percentage(%)↑
K-CrossF-1CF-4C
BoostAcc[17]4.884.89-
PureACL[24]10.315.01100.00
Ours100.00100.00100.00
", + "image_path": "e3f55b804473dcceb6deb9bb5ef9f99aa114ff75d6643e6f020c3e0369d8ee17.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 60, + 486, + 280, + 497 + ], + "lines": [ + { + "bbox": [ + 60, + 486, + 280, + 497 + ], + "spans": [ + { + "bbox": [ + 60, + 486, + 280, + 497 + ], + "type": "text", + "content": "- indicates that BoostAcc does not support 4-camera setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 502, + 140, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 140, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 140, + 515 + ], + "type": "text", + "content": "6. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 522, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 617 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed components, we conduct ablation studies in the 'cross' area of the KITTI-CVL dataset. Our method's performance was compared across various configurations, including the presence and absence of CycDA and Equidistant Re-projection (ERP) Loss functions, and the T2GA module. The results, as detailed in Tab. 4, underscore the significant role each component plays in improving our method's performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "The T2GA module aggregates information from top to ground based on the assumption that the ground and satellite views are orthogonal. However, this assumption might not be accurate in scenarios involving uphill/downhill or turns. To test our method's robustness under these conditions, we introduced affine warping to the ground view images with a shear range of " + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\pm 15^{\\circ}" + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": ". The outcomes, labeled 'w/ Affine' in Tab. 4, indicate our method's efficacy in handling such" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "scenarios. This success is primarily due to two factors: (1) the use of a U-Net for feature extraction allows each pixel in the feature map to encapsulate information from its surrounding area, enabling the handling of a degree of angular variation. (2) the satellite view from Google, which is not strictly orthogonal to the ground view, suggests that our method is already adapted to these types of challenges." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 185, + 542, + 293 + ], + "blocks": [ + { + "bbox": [ + 305, + 163, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 545, + 184 + ], + "type": "text", + "content": "Table 4. Ablation study on the KITTI-CVL dataset 'Cross' area with initial noise " + }, + { + "bbox": [ + 305, + 163, + 545, + 184 + ], + "type": "inline_equation", + "content": "(\\pm 45^{\\circ},\\pm 20\\mathrm{m})" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 185, + 542, + 293 + ], + "lines": [ + { + "bbox": [ + 308, + 185, + 542, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 542, + 293 + ], + "type": "table", + "html": "
T2GA CycDA ERPLateral(m)↓Mean MedianLongit(m)↓Mean MedianOrient(°)↓Mean Median
4.240.253.890.4410.963.16
✓1.290.191.270.174.112.13
✓✓0.210.180.120.103.381.89
✓✓✓0.170.150.090.072.881.44
w/ Affine0.180.150.090.072.881.52
night0.390.270.390.306.642.55
", + "image_path": "e46f206c991a394e82477bc75c5dcabdcfd40f39cf6e3fd5cce45370165ff43c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 301, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 385 + ], + "type": "text", + "content": "We further evaluate our method on an artificial dataset derived from the KITTI test sets, created using the cross-domain deep network proposed by Arruda et al. [2]. The outcomes, labeled as 'night' in Tab. 4 show our method's stable performance in nighttime scenarios, highlighting its resilience to variations in lighting and time of day. For further details, please refer to Supplementary Sec. C." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 396, + 378, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 378, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 378, + 409 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 417, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 633 + ], + "type": "text", + "content": "We have presented a novel top-to-ground feature aggregation for enhancing cross-view image-based geolocation. Our method overcomes limitations by incorporating aerial perspectives and utilizing a cycle domain adaptation loss for consistent feature extraction despite visual disparities. The introduction of the equidistant re-projection loss balances keypoint impact, promoting wider distribution and thus enhancing orientation accuracy. Our method excels in vehicle pose estimation across challenging scenarios, achieving the lowest translation errors in KITTI and Ford datasets, and minimal orientation error with less accurate initial poses. Crucially, by relying solely on the initial vehicle pose at the start, our method successfully completes routes via continuous pose estimation. This paves the way for real-world applications like autonomous driving and outdoor robotics. Future work will focus on integrating this method into SLAM systems to remove loop closure dependency and facilitate high-definition map generation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project is supported by CSIRO Julius Career Award, CSIRO Early Research Career Fellowship, CSIRO Future Digital Manufacturing Fund Postdoc Fellowship, CSIRO Data61 PhD Scholarship to SW, and partially by Australian Research Council Discovery Grant (DP220100800) to HL." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14850" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 116, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 287, + 159 + ], + "type": "text", + "content": "[1] Siddharth Agarwal, Ankit Vora, Gaurav Pandey, Wayne Williams, Helen Kourous, and James McBride. Ford multi-AV seasonal dataset. The International Journal of Robotics Research, 39(12):1367-1376, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 161, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 288, + 228 + ], + "type": "text", + "content": "[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 228, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 272 + ], + "type": "text", + "content": "[3] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 274, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 339 + ], + "type": "text", + "content": "[4] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "text", + "content": "[5] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 386, + 212, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 212, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 212, + 397 + ], + "type": "text", + "content": "[6] Google. Maps static api, 2023. 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "type": "text", + "content": "[7] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 486 + ], + "type": "text", + "content": "[8] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 488, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 288, + 543 + ], + "type": "text", + "content": "[9] Ted Lentsch, Zimin Xia, Holger Caesar, and Julian FP Kooij. Slicematch: Geometry-guided aggregation for cross-view pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17225-17234, 2023. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "text", + "content": "[10] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "type": "text", + "content": "[11] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[12] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[13] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Balntas. OrienterNet: Visual Localization in 2D Public Maps with Neural Matching. In CVPR, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "text", + "content": "[14] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "text", + "content": "[15] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation estimation by cross-view matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064-4072, 2020. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 547, + 241 + ], + "type": "text", + "content": "[16] Yujiao Shi, Xin Yu, Liu Liu, Tong Zhang, and Hongdong Li. Optimal feature transport for cross-view image geolocation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11990-11997, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 547, + 285 + ], + "type": "text", + "content": "[17] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 331 + ], + "type": "text", + "content": "[18] Shinya Sumikura, Mikiya Shibuya, and Ken Sakurada. Openvslam: A versatile visual slam framework. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2292-2295, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 333, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 547, + 376 + ], + "type": "text", + "content": "[19] Xiaoyang Tian, Jie Shao, Deqiang Ouyang, and Heng Tao Shen. Uav-satellite view synthesis for cross-view geolocation. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4804-4815, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 547, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 547, + 432 + ], + "type": "text", + "content": "[20] Aysim Token, Qunjie Zhou, Maxim Maximov, and Laura Leal-Taixe. Coming down to earth: Satellite-to-street view synthesis for geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 434, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 547, + 488 + ], + "type": "text", + "content": "[21] Frank Van Diggelen and Per Enge. The world's first gps mooc and worldwide laboratory using smartphones. In Proceedings of the 28th international technical meeting of the satellite division of the institute of navigation (ION GNSS+ 2015), pages 361-369, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 490, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 547, + 533 + ], + "type": "text", + "content": "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 535, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 547, + 600 + ], + "type": "text", + "content": "[23] Shan Wang, Chuong Nguyen, Jiawei Liu, Kaihao Zhang, Wenhan Luo, Yanhao Zhang, Sundaram Muthu, Fahira Afzal Maken, and Hongdong Li. Homography guided temporal fusion for road line and marking segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1075-1085, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 547, + 634 + ], + "type": "text", + "content": "[24] Shan Wang, Yanhao Zhang, Akhil Perincherry, Ankit Vora, and Hongdong Li. View consistent purification for accurate cross-view localization, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 690 + ], + "type": "text", + "content": "[25] Shan Wang, Yanhao Zhang, Ankit Vora, Akhil Perincherry, and Hengdong Li. Satellite image based cross-view localization for autonomous vehicle. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3592-3599. IEEE, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 547, + 714 + ], + "type": "text", + "content": "[26] Xiaolong Wang, Runsen Xu, Zuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14851" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 339 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "a correlation-aware homography estimator. arXiv preprint arXiv:2308.16906, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[27] Zimin Xia, Olaf Booij, Marco Manfredi, and Julian FP Kooij. Visual cross-view metric localization with dense uncertainty estimates. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 90-106. Springer, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[28] Zimin Xia, Olaf Booij, and Julian FP Kooij. Convolutional cross-view pose estimation. arXiv preprint arXiv:2303.05915, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "text", + "content": "[29] Zelong Zeng, Zheng Wang, Fan Yang, and Shin'ichi Satoh. Geo-localization via ground-to-satellite cross-view image retrieval. IEEE Transactions on Multimedia, 25:2176-2188, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "text", + "content": "[30] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 339 + ], + "type": "text", + "content": "[31] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1162-1171, 2022. 1, 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14852" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_content_list.json b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1554d3a521835e40df7ca10ba86b5470a198628e --- /dev/null +++ b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_content_list.json @@ -0,0 +1,1700 @@ +[ + { + "type": "text", + "text": "View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shilong Ou, Zhe Xue*, Yawen Li, Meiyu Liang, Yuanqiang Cai, Junjiang Wu \nBeijing University of Posts and Telecommunications, China", + "bbox": [ + 117, + 203, + 843, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{os1, xuezhe, meiyul210, caiyuanqiang, wujunjiang}@bupt.edu.cn, warmly0716@126.com", + "bbox": [ + 117, + 241, + 843, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a problem often encountered in real-world scenarios, multi-view multi-label learning has attracted considerable research attention. However, due to oversights in data collection and uncertainties in manual annotation, real-world data often suffer from incompleteness. Regrettably, most existing multi-view multi-label learning methods sidestep missing views and labels. Furthermore, they often neglect the potential of harnessing complementary information between views and labels, thus constraining their classification capabilities. To address these challenges, we propose a view-category interactive sharing transformer tailored for incomplete multi-view multi-label learning. Within this network, we incorporate a two-layer transformer module to characterize the interplay between views and labels. Additionally, to address view incompleteness, a KNN-style missing view generation module is employed. Finally, we introduce a view-category consistency guided embedding enhancement module to align different views and improve the discriminating power of the embeddings. Collectively, these modules synergistically integrate to classify the incomplete multi-view multi-label data effectively. Extensive experiments substantiate that our approach outperforms the existing state-of-the-art methods.", + "bbox": [ + 76, + 323, + 472, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The proliferation of data sources, coupled with advancements in data collection techniques, enables the extraction of diverse types of features, extending beyond a single perspective. This leads to an intriguing research question: how to effectively integrate multi-view features and extract meaningful features from them. Simultaneously, it is observed that single-label data often falls short in encapsulating the complexities of real-world scenarios. For example, an image of a flower may also warrant labels like 'leaf' or 'garden', illustrating that multi-labels provide a richer cate", + "bbox": [ + 76, + 724, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gory feature space and more accurately preserve inter-label relationships [8]. Given its practical relevance, the field of multi-view multi-label learning emerges as a domain of substantial research interest.", + "bbox": [ + 500, + 292, + 890, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of multi-view learning, multi-view multi-label classification (MVMLC) poses a significant challenge. Considering the consistency across views and the complexity of labels, it is essential for methods to simultaneously align views and extract diverse features from an integrated representation. In recent years, numerous researchers have made strides towards this objective. One such effort is the introduction of a multi-view label embedding model [30], which employs the Hilbert-Schmidt Independence Criterion [4] to establish a link between view feature and category feature spaces. Concurrently, deep neural networks have been employed in this realm. A noteworthy development is the SIMM neural network [3], which utilizes adversarial and label losses to identify shared features while implementing regularization to capture view-specific details.", + "bbox": [ + 496, + 354, + 892, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While MVMLC has attracted considerable research attention, it often operates under the presumption that all views and labels are complete. This assumption does not consistently align with real-world scenarios. For instance, a video might lack audio or text content. Multimedia data shared on social media networks might lack comprehensive annotations from users. Such instances culminate in the pervasive challenge of missing multi-view features and incomplete category information within real-world datasets. Some methods aim to address this problem by either masking or restoring the missing views [9, 20, 22, 25]. Although these approaches for handling incomplete multi-view or multi-label learning have achieved notable results, they often fall short in simultaneously addressing both types of incompleteness. To handle the double incomplete multi-view multi-label learning problem (DIMVMLC), several deep contrastive networks are proposed [10, 16]. Furthermore, an incomplete multi-view multi-label learning method based on transformers have been proposed [11], showing adaptability to multi-view and multi-label datasets.", + "bbox": [ + 496, + 580, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, most existing methods for double incomplete", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27467", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multi-view multi-label classification overlook two critical aspects. Firstly, while some methods recognize the importance of complementary information across multiple views, they often underestimate the complementarity between multi-view features and categories. Efficiently facilitating interaction between views and categories can dramatically enhance the ability to counteract the simultaneous absence of both multi-view feature and category information. Secondly, these methods frequently overlook the issue of missing multi-view features and do not consider completing missing multi-view features. This negligence makes them susceptible to vulnerabilities in insufficient view features, thereby limiting their classification performance.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the prevailing challenges in double incomplete multi-view multi-label learning, we propose the View-category Interactive Sharing Transformer (VIST), as shown in Fig. 1. VIST is an architecturally advanced framework specifically designed to exploit the synergy between multiple views and categories and adeptly generate missing views. The dual capability of this method not only amplifies the effectiveness of multi-view embeddings through high-level semantics of categories but also significantly enhances the precision of multi-label classification by utilizing these enriched embeddings. This advancement inherently fortifies the method's resilience, particularly in scenarios characterized by the lack of specific views and labels, thereby ensuring robust and reliable performance. VIST integrates three key components: view-category interactive sharing transformer, missing view generation, and view-category consistency guided embedding enhancement. The view-category interactive sharing transformer, a compact two-layer network with four transformer blocks, fosters an interaction between multi-view and category information, uncovering their complementary potential. The missing view generation, adopting a KNN-style approach, adeptly completes missing views in line with each sample's distribution. The view-category consistency guided embedding enhancement module utilizes contrastive learning strategy to sharpen the discriminative strength of embeddings and boost classification accuracy. Extensive experiments demonstrate the excellent performance of VIST. In summary, this paper makes the following contributions:", + "bbox": [ + 75, + 289, + 472, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel view-category interactive sharing transformer for incomplete multi-view multi-label learning. Our method can effectively exploit the complementary information between views and categories through the interaction between multi-view feature and category information, thereby effectively counteracting the challenges posed by incomplete views and labels.", + "- We propose a missing view generation method specifically designed for the incomplete multi-view multi-label learning task. This method combines the KNN strategy with a multivariate Gaussian distribution, ensuring a sta" + ], + "bbox": [ + 76, + 734, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tistically sound imputation of missing views that are cohesively aligned with the corresponding samples.", + "bbox": [ + 511, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To bolster the discriminating power of the multi-view embeddings, we develop a view-category consistency guided embedding enhancement module. It adopts contrastive learning to align embeddings across different views and leverages category information to guide embedding learning process, significantly boosting the effectiveness of the resulting embeddings and the performance of multi-label classification.", + "bbox": [ + 500, + 121, + 893, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many methods have been proposed for the MVMLC task. IrMMC [12] ensures the common subspace remains low rank, making it compatible with matrix completion, while also learning combination weights to tap into the distinct strengths of each view. Matrix factorization is utilized to draw out complementary data across different views [27]. Despite their strengths, these models struggles with handling datasets that have missing views or labels. To solve this, MVL-IV [23] delves into incomplete multi-view learning by leveraging the relationships between views. The iMSF model is another incomplete multi-view single-label learning approach. It ingeniously splits the incomplete multi-view classification responsibilities into several complete subtasks [26]. However, both MVL-IV and iMSF have a common shortcoming: they only address view incompleteness. Conversely, MvEL targets the extraction of context and neighborhood consistency but is tailored only for the incomplete multi-label scenario [29].", + "bbox": [ + 496, + 280, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, for double incomplete multi-view multi-label learning task, iMvWL [14] and NAIML [9], have ventured into addressing the incompleteness in both views and labels. iMvWL aligns multi-view features and multi-label data into a shared subspace, enriched by a correlation matrix that bolsters the projection from label space to the embedding subspace. NAIML addresses this dual incompleteness by leaning on the low-rank assumption of the sub-label matrix, subtly harnessing sub-class correlations. In recent years, deep learning models have been employed to address this issue. DDINet [21], DICNet [10] and LMVCAT [11] are deep learning models for DIMVMLC. DDINet is the first deep learning model which contains a view-specific decoder network, successfully preserving the key information of raw views. DICNet is the first deep learning model for DIMVMLC which employs contrastive learning at the instance-level, guiding model to extract deep cross-view features. LMVCAT is a transformer-based double incomplete multi-view multi-label learning framework, consisting of two transformers specifically tailored for views and categories, and constructs a graph on incomplete labels to guide the encoder in extracting view-specific features.", + "bbox": [ + 496, + 553, + 892, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27468", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 89, + 210, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 114, + 259, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $n$ multi-view data points from $V$ views, they can be represented as a set of matrices $\\{\\mathbf{X}_v\\}_{v=1}^V$ , with $\\mathbf{X}_v \\in \\mathbb{R}^{n \\times d_v}$ . Here, $d_v$ denotes the dimension of the $v$ -th view. The label matrix for all samples is denoted by $\\mathbf{Y} \\in \\{0,1\\}^{n \\times L}$ , where $L$ is the number of categories. If $\\mathbf{Y}_{i,j} = 1$ , it indicates that the $i$ -th sample belongs to category $j$ ; if $\\mathbf{Y}_{i,j} = 0$ , it indicates non-membership in this category. To denote multi-view data with missing features, we introduce the missing-view indicator matrix $\\mathbf{W} \\in \\{0,1\\}^{n \\times V}$ . $\\mathbf{W}_{i,j} = 1$ means that the $j$ -th view of the $i$ -th sample is available, whereas $\\mathbf{W}_{i,j} = 0$ indicates unavailable. Similarly, we define the missing-label indicator matrix $\\mathbf{U} \\in \\{0,1\\}^{n \\times L}$ , where $\\mathbf{U}_{i,j} = 1$ indicates that the $j$ -th category of the $i$ -th sample is available, and $\\mathbf{U}_{i,j} = 0$ signifies the opposite. During the data pre-processing phase, all missing information for view $\\mathbf{X}_v$ and label $\\mathbf{Y}$ is assigned the value 0. The objective of our double incomplete multi-view multi-label learning task is to train a model capable of accurately predicting multiple categories for each sample. In subsequent sections, the detailed mechanisms of the proposed VIST are presented.", + "bbox": [ + 76, + 138, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. View-Category Interactive Sharing Transformer", + "text_level": 1, + "bbox": [ + 76, + 467, + 468, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different views harbor complementary information, necessitating their interaction for effective utilization. Furthermore, category information embodies the data's high-level semantics, serving to augment the discriminative capability of multi-view representations. Inversely, multi-view features can remedy inaccuracies in label information comprehension due to incomplete labels. To this end, we introduce a view-category interactive sharing transformer. This novel transformer is designed to enhance information exchange between views and categories, consequently boosting the discriminative strength of multi-view representations and refining the precision of multi-label classification.", + "bbox": [ + 75, + 507, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The view-category interactive sharing transformer comprises two key layers. The first layer focuses on extracting shared features that captures correlations among multiple views and categories. An essential preliminary step involves embedding the multi-view data $\\mathbf{X}_v\\in \\mathbb{R}^{n\\times d_v}$ . This embedding process ensures the homogenization of feature dimensions across views. For a given sample $\\mathbf{x}_v\\in \\mathbb{R}^{d_v}$ from $\\mathbf{X}_v$ , the embedding vector $\\mathbf{e}_v\\in \\mathbb{R}^{d_e}$ can be denoted as $\\mathbf{e}_v = Embedding(\\mathbf{x}_v)$ , where $d_{e}$ denotes the dimension of embedding space, and $Embedding(\\cdot)$ is a fully connected neural network. Then we stack the embedding vectors to obtain an original multi-view embedding sequence $\\mathbf{E} = [\\mathbf{e}_1,\\mathbf{e}_2,\\dots ,\\mathbf{e}_V]\\in \\mathbb{R}^{V\\times d_e}$ , which is further used as input vectors for the transformer. Noted that for the incom", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plete multi-view data, we adopt missing view generation method (as detailed in Sec. 3.3) to produce embeddings of missing views, ensuring that $\\mathbf{E}$ is complete across all views.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The structure of our transformer is the encoder module of the classical transformer [18]. A transformer block is denoted by $\\text{Transformer}(\\Phi)$ , and the input of the transformer is $\\Phi \\in \\mathbb{R}^{t \\times d_t}$ , where $t$ and $d_t$ denote the number of tokens and the dimension of tokens. The first layer consists of two transformer blocks, denoted as $\\text{Transformer}_{sv}$ and $\\text{Transformer}_{sc}$ . By discerning correlations among original views, $\\text{Transformer}_{sv}$ serves to complement information across original views and get multi-view embedding $\\tilde{\\mathbf{E}}$ , which is the enhanced representation of original views. It can be illustrated as follows:", + "bbox": [ + 496, + 136, + 892, + 301 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {E}} = \\text {T r a n s f o r m e r} _ {s v} (\\mathbf {E}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 311, + 890, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{E}} = [\\tilde{\\mathbf{e}}_1,\\tilde{\\mathbf{e}}_2,\\dots ,\\tilde{\\mathbf{e}}_V]\\in \\mathbb{R}^{V\\times d_e}$ . Here, an adaptive fusion layer is introduced to fuse the information from multiple views into a shared view feature $s_v$ . The fusion process is formulated as follows:", + "bbox": [ + 496, + 338, + 890, + 400 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {v} = \\sum_ {v = 1} ^ {V} \\frac {e ^ {\\theta_ {v} ^ {\\epsilon}} \\tilde {\\mathbf {e}} _ {v}}{\\sum_ {j} e ^ {\\theta_ {j} ^ {\\epsilon}}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 407, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{e}}_v\\in \\mathbb{R}^{d_e}$ is the $v$ -th row of $\\tilde{\\mathbf{E}}$ $\\theta_v$ denotes the learnable weight and $\\epsilon$ is a adjustment factor. By interacting with the shared view feature $\\mathbf{s}_v$ and discerning correlations among original categories, $T r a n s f o r m e r_{s c}$ serves to complement information across original categories and get category vectors $\\tilde{\\mathbf{C}}$ , which can be illustrated as follows:", + "bbox": [ + 496, + 460, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n[ \\tilde {\\mathbf {s}} _ {v}, \\tilde {\\mathbf {C}} ] = T r a n s f o r m e r _ {s c} ([ \\mathbf {s} _ {v}, \\mathbf {C} ]) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 561, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ denotes the concatenation operation and $\\mathbf{C} \\in \\mathbb{R}^{L \\times d_e}$ denotes original category vectors which are randomly initialized. Subsequently, the output feature of $Transformer_{sc}$ , $\\tilde{\\mathbf{s}}_v$ , is propagated into the second layer.", + "bbox": [ + 496, + 590, + 890, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second layer is designed to extract advanced shared features, which is achieved by promoting the interaction and fusion between view and category information based on the features extracted from the first layer, thereby obtaining a more discriminative multi-view data representation. This layer incorporates two transformer blocks, denoted as $\\text{Transformer}_{av}$ and $\\text{Transformer}_{ac}$ . $\\text{Transformer}_{av}$ is employed to complement information across views and extract advanced multi-view embedding $\\bar{\\mathbf{E}}$ , which is the enhanced representation of views, by interacting with the shared category feature $\\mathbf{s}_c$ and discerning view correlations, as represented below:", + "bbox": [ + 496, + 651, + 892, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {c} = \\operatorname {P r o j} _ {c v} (\\tilde {\\mathbf {s}} _ {v}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 843, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n[ \\bar {\\mathbf {s}} _ {c}, \\bar {\\mathbf {E}} ] = T r a n s f o r m e r _ {a v} ([ \\mathbf {s} _ {c}, \\tilde {\\mathbf {E}} ]) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65e8fe268821bf0f354bf6ea3ad07c8eaf196ca5d3e799eaae48106d254d06ba.jpg", + "image_caption": [ + "Figure 1. An overview of our method VIST. It comprises three modules, namely View-Category Interactive Sharing Transformer Module, Missing View Generation Module, and View-Category Consistency Guided Embedding Enhancement Module. Different colors represent" + ], + "image_footnote": [], + "bbox": [ + 158, + 93, + 810, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "different samples, while different shapes signify different features. where $Proj_{cv}(\\cdot)$ is a linear layer as a projection function designed to map vectors from the category feature space to the view feature space, $\\bar{\\mathbf{E}} = [\\bar{\\mathbf{e}}_1,\\bar{\\mathbf{e}}_2,\\dots ,\\bar{\\mathbf{e}}_V]\\in \\mathbb{R}^{V\\times d_e}$ . Correspondingly, $Transformer_{ac}$ is employed to complement information across categories and extract advanced category vectors $\\bar{\\mathbf{C}}$ by leveraging the advanced shared feature $\\mathbf{p}_v$ and discerning category correlations, as illustrated below:", + "bbox": [ + 75, + 393, + 470, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {v} = \\operatorname {P r o j} _ {v c} (\\overline {{\\mathbf {s}}} _ {c}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 523, + 468, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\bar {\\mathbf {p}} _ {v}, \\bar {\\mathbf {C}} ] = T r a n s f o r m e r _ {a c} ([ \\mathbf {p} _ {v}, \\bar {\\mathbf {C}} ]) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 559, + 468, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Proj_{vc}(\\cdot)$ is a linear layer as a projection function to map vectors from the view feature space to the category feature space. Through the propagation of vectors $\\mathbf{s}_v$ , $\\mathbf{s}_c$ , and $\\mathbf{p}_v$ among the transformer blocks, we facilitate the sharing of information between the view and category feature spaces, thereby extracting more refined and effective features of views and categories.", + "bbox": [ + 75, + 582, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Missing View Generation", + "text_level": 1, + "bbox": [ + 76, + 696, + 307, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the absence of views in the dataset, an intuitive consideration is that the proper completion of these missing views can enhance the model's performance. Through the contrastive learning process mentioned in Sec. 3.4, the distance between views of a sample tends to be relatively proximate. Thus, generating missing views utilizing the existing ones is a plausible approach. We adopt a method analogous to $k$ -nearest neighbors, leveraging original category vectors $\\mathbf{C} = [\\mathbf{c}_1, \\mathbf{c}_2, \\dots, \\mathbf{c}_L]$ to assist in generating the missing views. Specifically, for a certain sample $i$ , let $\\mathcal{E} = \\{v | \\mathbf{W}_{i,v} = 1\\}$ denote the index of existing views, and $\\mathcal{M} = \\{v | \\mathbf{W}_{i,v} = 0\\}$ denote the index of missing views. To", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "complete the missing features of sample $i$ with its original multi-view embedding $\\mathbf{E}$ , we first find the $k$ nearest neighbors in the projected category feature space. The set of the neighbors $\\mathcal{D}$ is constructed as follows:", + "bbox": [ + 496, + 404, + 890, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = \\left\\{d \\mid T o p K \\left(\\sum_ {j \\in \\mathcal {E}} \\| \\mathbf {e} _ {j} - \\mathbf {c} _ {d} \\| _ {2}, d \\in \\{1, 2, 3, \\dots , L \\}\\right) \\right\\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 473, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $TopK(\\cdot)$ is a function designed to identify the indices of the top $k$ categories, based on the smallest distance between embedding vectors and category vectors. Then, we employ a statistical method to describe the distribution of the missing views. We hypothesize that these missing views $\\{\\mathbf{e}_m\\}_{m\\in \\mathcal{M}}$ adhere to a multivariate Gaussian distribution $\\mathcal{N}(\\mu ,\\Sigma)$ , with a mean vector and covariance matrix denoted as:", + "bbox": [ + 496, + 518, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\frac {\\sum_ {d \\in \\mathcal {D}} \\mathbf {c} _ {d}}{| \\mathcal {D} |} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 636, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\frac {1}{| \\mathcal {D} | - 1} \\sum_ {d \\in \\mathcal {D}} \\left(\\mathbf {c} _ {d} - \\mu\\right) \\left(\\mathbf {c} _ {d} - \\mu\\right) ^ {T} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 676, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the missing views, we sample from this distribution $|\\mathcal{M}|$ times and substitute the missing views with the sampled results. Consequently, we can obtain the complete embeddings for the incomplete multi-view data. By reconstructing the absent multi-view data, our method further enhances its performance in category predictions.", + "bbox": [ + 496, + 717, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. View-Category Consistency Guided Embedding Enhancement", + "text_level": 1, + "bbox": [ + 498, + 816, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the aforementioned view-category interactive sharing transformer, we extract three multi-view embeddings $\\mathbf{E}$ , $\\tilde{\\mathbf{E}}$ and $\\bar{\\mathbf{E}}$ from different layers. In order to obtain", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27470", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "better classification performance, enhancing the discriminating power and effectiveness of these embeddings is of great importance. Specifically, predicated on the view consistency assumption [10, 24], the embedding $\\mathbf{E}$ of a sample on different views should be aligned. Moreover, to improve the discriminating power of $\\tilde{\\mathbf{E}}$ and $\\bar{\\mathbf{E}}$ , the consistency between category correlations and feature correlations of multi-view data can be leveraged. In light of these factors, we introduce view-category consistency guided embedding enhancement.", + "bbox": [ + 75, + 90, + 472, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To learn more effective embedding $\\mathbf{E}$ , we use contrastive learning to align the embeddings of the same sample across different views. According to [13, 17], the loss function of contrastive learning can be formulated as follows:", + "bbox": [ + 75, + 242, + 470, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} - \\log \\frac {e ^ {\\operatorname {s i m} (x , x ^ {+})}}{e ^ {\\operatorname {s i m} (x , x ^ {+})} + \\sum_ {x ^ {-} \\in \\mathcal {F} (x)} e ^ {\\operatorname {s i m} (x , x ^ {-})}} \\tag {11} \\\\ \\approx - s i m (x, x ^ {+}) + \\log \\sum_ {x ^ {-} \\in \\mathcal {F} (x)} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 311, + 468, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x, x^{+}, x^{-}$ denotes the anchor sample, its positive sample and negative sample. $\\mathcal{F}(x)$ denotes the set of negative pairs. $sim(\\cdot, \\cdot)$ calculates the similarity between two samples.", + "bbox": [ + 75, + 397, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evidently, our objective is to maximize the similarity $sim(x,x^{+})$ while simultaneously minimizing the similarity $\\sum_{x^{-}\\in \\mathcal{F}(x)}e^{sim(x,x^{-})}$ . We denote $\\mathbf{e}_{i,v}$ as the embedding of the $v$ -th view for the $i$ -th sample, then $\\mathcal{F}(x)$ can be rewritten as $\\mathcal{F}(i,v) = \\{j,w|\\sum^L\\mathbf{y}_i\\circ \\mathbf{u}_i\\circ \\mathbf{y}_j\\circ \\mathbf{u}_j = 0,\\mathbf{U}_{j,w} = 1\\}$ , where $\\circ$ denotes Hadamard product, $\\mathbf{u}_i$ denotes the missing vector of the $i$ -th sample from $U$ . As for maximizing $e^{sim(x,x^{+})}$ , we use the following loss function:", + "bbox": [ + 76, + 458, + 472, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p o s} = \\sum_ {v _ {1} = 1} ^ {V} \\sum_ {v _ {2} = 1, v _ {2} \\neq v _ {1}} ^ {V} \\frac {1}{d _ {e}} \\| n \\left(\\mathbf {e} _ {i, v _ {1}}\\right) - n \\left(\\mathbf {e} _ {i, v _ {2}}\\right) \\| _ {2} ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 593, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n(\\cdot)$ denotes $L_{2}$ normalization function. To minimize $\\sum_{x^{-}\\in \\mathcal{F}(x)}e^{sim(x,x^{-})}$ , the following loss function can be adopted:", + "bbox": [ + 76, + 646, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n e g} = \\sum_ {v _ {1} = 1} ^ {V} \\sum_ {j, v _ {2} \\in \\mathcal {F} (i, v _ {1})} \\frac {- 1}{d _ {e}} \\| n \\left(\\mathbf {e} _ {i, v _ {1}}\\right) - n \\left(\\mathbf {e} _ {j, v _ {2}}\\right) \\| _ {2} ^ {2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 702, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further improve the effectiveness of $\\tilde{\\mathbf{E}}$ and $\\bar{\\mathbf{E}}$ , we utilize the category information to guide the learning process of these embeddings. We enhance the consistency of data correlations in the view feature space and the category feature space as detailed below:", + "bbox": [ + 75, + 757, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {v} ^ {1} (i, j) = \\left(n \\left(\\tilde {\\mathbf {e}} _ {i, v}\\right) \\left(n \\left(\\tilde {\\mathbf {e}} _ {j, v}\\right)\\right) ^ {T} + 1\\right) / 2 \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 842, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {v} ^ {2} (i, j) = \\left(n \\left(\\bar {\\mathbf {e}} _ {i, v}\\right) \\left(n \\left(\\bar {\\mathbf {e}} _ {j, v}\\right)\\right) ^ {T} + 1\\right) / 2 \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\left(\\left(\\mathbf {Y} \\circ \\mathbf {U}\\right) \\left(\\mathbf {Y} \\circ \\mathbf {U}\\right) ^ {T}\\right). / \\left(\\mathbf {U U} ^ {T}\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 104, + 890, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_v^1$ , $\\mathbf{G}_v^2$ and $\\mathbf{H}$ are the correlation matrices of $\\tilde{\\mathbf{E}}$ , $\\bar{\\mathbf{E}}$ and labels respectively. $\\mathbf{e}_{i,v}$ and $\\bar{\\mathbf{e}}_{i,v}$ represent the feature of the $v$ -th view of the $i$ -th sample from $\\tilde{\\mathbf{E}}$ and $\\bar{\\mathbf{E}}$ respectively. Through the cross-entropy loss function, we can align the data correlations across view feature space and category feature space so that more discriminative embedding can be obtained:", + "bbox": [ + 498, + 126, + 893, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a u x} = \\frac {- 1}{2 n V} \\sum_ {v = 1} ^ {V} \\sum_ {i = 1} ^ {n} \\sum_ {j \\neq i} ^ {n} \\left(\\mathbf {H} _ {i, j} \\left(\\log \\mathbf {G} _ {v} ^ {1} (i, j) + \\log \\mathbf {G} _ {v} ^ {2} (i, j)\\right) \\right. \\\\ + \\left(\\mathbf {U} _ {i, j} - \\mathbf {H} _ {i, j}\\right) \\left(\\log \\left(1 - \\mathbf {G} _ {v} ^ {1} (i, j)\\right) + \\log \\left(1 - \\mathbf {G} _ {v} ^ {2} (i, j)\\right)\\right)) \\tag {17} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 242, + 901, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the above factors, the loss function for view-category consistency guided embedding enhancement can be summarized as:", + "bbox": [ + 498, + 319, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\mathcal {L} _ {\\text {p o s}} + \\mathcal {L} _ {\\text {n e g}} + \\alpha \\mathcal {L} _ {\\text {a u x}} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 375, + 890, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a weight parameter. Through optimizing the above loss function, the learned data embeddings become more discriminative, laying a solid foundation for subsequent multi-label classification.", + "bbox": [ + 498, + 400, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Overall Loss Function", + "text_level": 1, + "bbox": [ + 500, + 468, + 705, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve multi-label classification results, we partition the output $[\\bar{\\mathbf{p}}_v,\\bar{\\mathbf{C}} ]$ of $Transforer_{ac}$ into $\\bar{\\mathbf{p}}_v$ and $\\bar{\\mathbf{C}}$ . Specifically, for sample $i$ , $\\bar{\\mathbf{p}}_v^i$ passes through a linear layer to yield $\\mathbf{p}_f^i\\in [0,1]^L$ , representing the final complementary information between views and categories. $\\bar{\\mathbf{C}}^i$ is processed through a set of category-aware linear layers, resulting in $\\mathbf{z}^i\\in [0,1]^L$ , which signifies category-aware predictions. With the assistance of $\\mathbf{p}_f^i$ , we get the final prediction $\\hat{\\mathbf{y}}^i$ :", + "bbox": [ + 498, + 491, + 890, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} ^ {i} = \\operatorname {c l a s s i f i e r} \\left(\\mathbf {p} _ {f} ^ {i} + \\mathbf {z} ^ {i}\\right) \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 628, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where classifier $(\\cdot)$ is a linear classifier. We utilize the masked cross-entropy function to compute the multi-label classification loss:", + "bbox": [ + 498, + 651, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\frac {\\sum_ {i = 1} ^ {n} \\sum_ {l = 1} ^ {L} \\left(\\mathbf {Y} _ {i , l} \\log \\hat {\\mathbf {y}} _ {l} ^ {i} + \\left(1 - \\mathbf {Y} _ {i , l}\\right) \\log \\left(1 - \\hat {\\mathbf {y}} _ {l} ^ {i}\\right)\\right) \\mathbf {U} _ {i , l}}{- \\sum_ {i , l} \\mathbf {U} _ {i , l}} \\tag{20}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 703, + 911, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that class imbalance is an inherent characteristic of multi-label data, impacting the generalization performance of multi-label prediction [15, 28]. Typically, for certain class labels, the instances of positive label assignments in training data are significantly fewer than those of negative assignments. To address this class-imbalance issue, we introduce a masked asymmetric loss function:", + "bbox": [ + 496, + 753, + 890, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {l} ^ {i} = \\left(\\max \\left(\\hat {\\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\\right)\\right) ^ {2} \\log \\left(1 - \\max \\left(\\hat {\\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\\right)\\right) \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a} = \\frac {\\sum_ {i = 1} ^ {n} \\sum_ {l = 1} ^ {L} \\left(\\mathbf {Y} _ {i , l} \\log \\hat {\\mathbf {y}} _ {l} ^ {i} + \\left(1 - \\mathbf {Y} _ {i , l}\\right) \\mathbf {L} _ {l} ^ {i}\\right) \\mathbf {U} _ {i , l}}{- \\sum_ {i , l} \\mathbf {U} _ {i , l}} \\tag{22}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 114, + 468, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ultimately, we derive the overall loss function of the model as follows:", + "bbox": [ + 76, + 162, + 468, + 190 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\mathcal {L} _ {m} + \\beta \\mathcal {L} _ {a} + \\gamma \\mathcal {L} _ {c} \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 191, + 468, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta, \\gamma$ denotes different weight parameters. By optimizing $\\mathcal{L}_o$ , our method proficiently learns the embeddings for incomplete multi-view data and consequently produces accurate multi-label classification results.", + "bbox": [ + 76, + 215, + 468, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 289, + 209, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 313, + 277, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Datasets", + "text_level": 1, + "bbox": [ + 76, + 335, + 191, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow [9, 11, 14] and evaluate the performance of the proposed VIST on five datasets. These datasets encompass:", + "bbox": [ + 76, + 359, + 468, + 390 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Corel 5k [1]: The Corel5k dataset is a standard dataset frequently utilized in image annotation and retrieval research, which has 5000 images with 260 types of tags.", + "- Pascal07 [2]: Pascal07, also known as PASCAL VOC 2007, is a renowned dataset for visual object classes, instrumental in object recognition and classification. It consists of 9963 images and 20 types of tags.", + "- Espgame [19]: The Espgame dataset is derived from an online game where participants generate labels for images, which contains 20770 samples and 268 types of tags.", + "- IAPRTC12 [5]: IAPRTC12 is an international image annotation dataset, comprising approximately 20,000 images and 290 categories. These images cover a wide range of subjects and scenes, each accompanied by detailed descriptive text.", + "- Mirflickr [7]: The Mirflickr dataset consists of 25,000 images and 38 types of tags sourced from Flickr. Selected from the public domain, these images include a variety of themes and scenes." + ], + "bbox": [ + 76, + 391, + 468, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the multi-view feature extraction method outlined in [6]. The images in these datasets undergo a comprehensive preprocessing procedure that transforms them into six distinct views. These views are GIST, HSV, HUE, LAB, RGB, and SIFT. Each of these views offers a unique perspective on the image data, capturing different aspects of visual information.", + "bbox": [ + 75, + 691, + 468, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 816, + 294, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For datasets, in line with [14], we process the five datasets to emulate an incomplete scenario as described below: For every view, we randomly exclude $50\\%$ of the samples, ensuring that each sample retains at least one view. Within", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each category, we arbitrarily designate $50\\%$ of both positive and negative tags as missing labels. For parameter settings, we choose $d_{e} = 512$ , $\\alpha = 0.1$ , $\\beta = 0.1$ , $\\gamma = 0.1$ and $k = 20$ for all the experiments. The model are trained for 500 epochs with a batch-size of 128 via SGD optimizer on an NVIDIA RTX A6000 GPU. To ensure the credibility of the results, we repeat the experiments 10 times and record the mean and variance of the outcomes.", + "bbox": [ + 498, + 90, + 890, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3 Compared Methods", + "text_level": 1, + "bbox": [ + 500, + 232, + 692, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose ten representative methods for comparison. Out of these, eight methods - lrMMC, MVL-IV, iMSF, iMvWL, NAIML, DDINet, DICNet and LMVCAT - are detailed in Sec. 2. Additionally, we incorporate GLOCAL [31], to broaden the assessment spectrum. Because iMvWL and NAIML cater to datasets with missing views and labels, we have to modify other methods for consistency. Drawing inspiration from [14] and [9], we input average values from accessible views into lrMMC for the non-operational ones. For MVL-IV and iMSF, absent tags are treated as negative. GLOCAL are executed individually for each view, and we highlight their optimal outcomes. We adhere to the recommended settings for these comparative techniques, as mentioned in their respective publications or code repositories, to maintain an impartial evaluation.", + "bbox": [ + 498, + 257, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.4 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 503, + 687, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to [11], we employ Average Precision (AP), Ranking Loss (RL) and Area Under Curve (AUC). For a convenient comparison, we adopt 1-RL for evaluation. A larger value indicates superior model performance.", + "bbox": [ + 498, + 521, + 890, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 589, + 700, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we show the experimental results of all the methods on five datasets with $70\\%$ training samples, $50\\%$ missing views and missing labels. Based on the results from Tab. 1, our method exhibits a distinct advantage across most metrics. For instance, on the AP metric for dataset Corel5k, our method surpasses the second-best method, LMVCAT, by $3\\%$ . Besides, the performance of DDINet, DICNet and LMVCAT surpasses that of the first six methods. This can be attributed to their applicability specifically to the DIMVMLC issue. Additionally, they are both deep learning models, their superior outcomes underscore the vast potential deep learning holds for addressing this particular challenge.", + "bbox": [ + 498, + 613, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While our method primarily addresses the DIMVMLC issue, we also evaluate the model's performance in the full-view and full-label context. The results are presented in Tab. 2. It can be observed that the model still exhibits superior performance in this scenario, suggesting that our method can be effectively applied in MVMLC.", + "bbox": [ + 498, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27472", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/44e1a92e774908ba32babc90820e4274e37f26b853db67c58ee4f7ccef515661.jpg", + "table_caption": [ + "Table 1. Experimental results of different methods on the five datasets with $70\\%$ training samples, $50\\%$ missing instances and missing labels. The best results are marked in bold. For simplicity, percent sign is omitted in the following section." + ], + "table_footnote": [], + "table_body": "
DatasetMetriclrMMCMVL-IViMSFGLOCALiMvWLNAIMLDDINetDICNetLMVCATVIST
Corel5kAP24.0±0.224.0±0.118.9±0.228.5±0.428.3±0.730.9±0.436.4±0.138.1±0.438.4±0.441.5±0.2
1-RL76.2±0.275.6±0.170.9±0.580.4±0.386.5±0.387.8±0.287.1±0.088.2±0.488.0±0.290.2±0.2
AUC76.3±0.276.2±0.166.3±0.584.3±0.386.8±0.388.1±0.287.5±0.188.4±0.488.3±0.290.6±0.1
Pascal07AP42.5±0.343.3±0.232.5±0.049.6±0.444.1±1.748.8±0.353.6±0.250.5±1.251.9±0.553.9±0.3
1-RL69.8±0.370.2±0.156.8±0.076.7±0.473.7±0.978.3±0.180.7±0.178.3±0.881.1±0.485.6±0.1
AUC72.8±0.273.0±0.168.6±0.562.0±0.178.6±0.376.7±1.282.7±0.087.6±0.283.4±0.488.3±0.2
EspgameAP18.8±0.018.9±0.010.8±0.022.1±0.224.2±0.324.6±0.228.3±0.129.7±0.229.4±0.430.7±0.3
1-RL77.7±0.177.8±0.072.2±0.278.0±0.480.7±0.181.8±0.281.5±0.083.2±0.182.8±0.284.4±0.0
AUC78.3±0.178.4±0.167.4±0.378.4±0.481.3±0.282.4±0.282.0±0.183.2±0.182.8±0.285.0±0.1
IAPRTC12AP19.7±0.019.8±0.010.1±0.025.6±0.223.5±0.426.1±0.130.3±0.232.3±0.131.7±0.333.9±0.2
1-RL80.1±0.079.9±0.163.1±0.082.5±0.283.3±0.384.8±0.185.3±0.187.3±0.187.0±0.188.4±0.2
AUC80.5±0.080.4±0.166.5±0.183.0±0.183.6±0.285.0±0.185.4±0.087.4±0.187.2±0.188.6±0.1
MirflickrAP44.1±0.144.9±0.132.3±0.053.7±0.249.5±1.255.1±0.259.8±0.258.9±0.559.4±0.560.4±0.3
1-RL80.5±0.080.4±0.166.5±0.183.2±0.183.6±0.285.0±0.186.3±0.086.3±0.486.5±0.387.9±0.1
AUC80.6±0.180.7±0.076.1±0.182.8±0.179.4±1.583.7±0.185.2±0.184.9±0.485.3±0.386.9±0.1
", + "bbox": [ + 78, + 121, + 901, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bcd553afe8d16fb14bc1d2e4975501d0179296bad9b7d0614d924f8801bd5ef0.jpg", + "table_caption": [ + "Table 2. Experimental results of different methods on the five datasets with $70\\%$ training samples, full views and labels. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
DatasetMetricNAIMLDICNetLMVCATVIST
Corel5kAP32.750.952.155.4
1-RL89.092.992.893.5
AUC89.393.193.094.2
Pascal07AP49.660.862.965.5
1-RL79.585.987.889.4
AUC82.287.689.291.2
EspgameAP25.139.138.540.1
1-RL82.587.187.687.7
AUC83.087.488.089.6
IAPRTC12AP26.741.843.645.7
1-RL82.591.291.892.7
AUC83.091.191.892.0
MirflickrAP55.565.968.470.4
1-RL84.789.590.591.3
AUC83.987.688.989.3
", + "bbox": [ + 80, + 397, + 465, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the performance of the model under various missing ratios. It is interesting that while both incomplete views and incomplete labels detrimentally affect the model's performance, the latter has a more pronounced impact. Moreover, despite our efforts in view imputation, the model's accuracy noticeably diminishes as the view missing ratio escalates. This is primarily due to our reliance on existing views for imputing the missing views. As the number of missing views increases, the imputation algorithm tends to degrade towards random generation.", + "bbox": [ + 75, + 645, + 468, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3 depicts the performance of the model under various training sample ratios. While the ratio's impact on the model is not markedly significant, a general trend is observed: as the ratio increases, there is a corresponding improvement in the model's performance. Thus, selecting a $70\\%$ training sample ratio for our experiments is a proper choice.", + "bbox": [ + 75, + 805, + 470, + 910 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/743b3a995735b83a9b070b7405cdca4e85d2949459ef9d0e39c2b499110dc572.jpg", + "image_caption": [ + "Figure 2. The results on the Corel5k dataset are presented with (a) different missing-view ratios accompanied by a $50\\%$ missing-label ratio and (b) a consistent $50\\%$ missing-view ratio paired with different missing-label ratios." + ], + "image_footnote": [], + "bbox": [ + 522, + 352, + 694, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87dda4ae5840060ddb0ef295a426347545b21999827977a57a5a9c34990328ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 352, + 870, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a59077620e17ccf34ce506b6f34efefd411a9e2d9503a8e9813d8bebf3bc7fd2.jpg", + "image_caption": [ + "Figure 3. The results of different training sample ratios on (a) Corel5k dataset and (b) Espgame dataset with $50\\%$ missing-view ratio and $50\\%$ missing-label ratio." + ], + "image_footnote": [], + "bbox": [ + 521, + 556, + 694, + 670 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fa6b0455a34a6ed75e589f067ad832f65441ce4fed584165a32b8e9c52c10a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 556, + 870, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Parameter Analysis", + "text_level": 1, + "bbox": [ + 500, + 738, + 687, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate four hyperparameters in our model, i.e., $\\alpha$ , $\\beta$ , $\\gamma$ and $k$ . To assess the influence of these hyperparameters on model performance, we adjust their values within their respective ranges and test their combinations based on the AP metric on different datasets with $50\\%$ missing views, $50\\%$ missing labels and $70\\%$ training sample ratio. Note that $k$ is inherently related to the number of labels, thus, the actual value of $\\hat{k}$ utilized in computation should be determined by $\\hat{k}_j = \\frac{k_i}{L_i} L_j$ , where $i$ and $j$ represents the $i$ -th", + "bbox": [ + 500, + 763, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae6e17a7a8302f0d96fd850bf4d75ebd85658d09174966ff9ba53d608416fdf.jpg", + "image_caption": [ + "(a) Corel5k" + ], + "image_footnote": [], + "bbox": [ + 94, + 102, + 263, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a508277c886370beb211dd7665003fd294ab114ca79916706d87d9e72873339.jpg", + "image_caption": [ + "(b) Pascal07" + ], + "image_footnote": [], + "bbox": [ + 272, + 104, + 439, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aeee90621fe8e357a42eebe6af10e753a24da45a8dc3cda77240745deea76700.jpg", + "image_caption": [ + "(c) Espgame", + "Figure 4. The results of different value combinations of hyperparameter $\\beta$ and $\\gamma$ on different datasets. $\\alpha$ and $k$ are respectively set to default values of 0.1 and 20." + ], + "image_footnote": [], + "bbox": [ + 94, + 241, + 261, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fafa396ed5683bac6c1aad87237864870c668cce7831ecd6d2f8f8385e9a46a8.jpg", + "image_caption": [ + "(d) IAPRTC12" + ], + "image_footnote": [], + "bbox": [ + 272, + 247, + 437, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and $j$ -th dataset. Based on Fig. 4, $\\beta$ and $\\gamma$ exhibit comparatively favorable performances within the ranges of [5e-2, 5e-1] and [5e-2, 1e0] respectively on dataset Corel5k. This indicates that the values of $\\beta$ and $\\gamma$ should neither be excessively large nor too small, and the optimal value of $\\beta$ approximated around 0.1, the optimal value of $\\gamma$ should also be selected in the vicinity of 0.1.", + "bbox": [ + 75, + 429, + 468, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As for $\\alpha$ and $k$ , we assess the impact of varying $\\alpha$ and $k$ on the model under fixed conditions of $\\beta = 0.1$ and $\\gamma = 0.1$ by measuring metrics AP and AUC. As illustrated in Fig. 5, on datasets Corel5k and Espgame, both the AP and AUC metrics remain within a narrow range, indicating that the model is not highly sensitive to $\\alpha$ and $k$ . This might be because they operate on the deep-layer mechanism, variations within certain ranges do not significantly impact the model's performance.", + "bbox": [ + 75, + 535, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 681, + 228, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the ablation study, we evaluate the performance of each module of VIST on three representative datasets, the results are shown in Tab. 3. We can observe that view-category interactive sharing transformer module plays a pivotal role in enhancing the model's performance. Furthermore, the view-category consistency guided embedding enhancement module is able to improve the discriminating power of embeddings and boost the classification performance. The missing view generation module also contributes to the model's overall efficacy. Additionally, upon the removal of $\\mathcal{L}_a$ , there is a noticeable decline in the model's metrics, particularly in the AUC. This suggests that the masked asymmetric loss indeed addresses issues associated with imbal", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65dbfc9f2a5aed684e65c1b14af1a4133500f3463ff4c9e7423c87bc1479dbf0.jpg", + "image_caption": [ + "(a) $\\alpha$ on Corel5k" + ], + "image_footnote": [], + "bbox": [ + 519, + 97, + 681, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6b4f0a6c98659c14df6879be6b55c10622d7e49989558d4f689c2577a4dbcf1.jpg", + "image_caption": [ + "(b) $\\alpha$ on Espgame" + ], + "image_footnote": [], + "bbox": [ + 697, + 98, + 857, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c7ff54188740fd592c41db2a017fd3146d3e28d92a51bf79a6d3f1353e0f042.jpg", + "image_caption": [ + "(c) $k$ on Corel5k", + "Figure 5. The results of different values of hyperparameter $\\alpha$ and $k$ on different datasets. $\\beta$ and $\\gamma$ are respectively set to default values of 0.1 and 0.1." + ], + "image_footnote": [], + "bbox": [ + 517, + 215, + 679, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0cffde94fa81c2904f5d2ad7934c4bfb7b0833a01218b46e87971e31a2f0edc0.jpg", + "image_caption": [ + "(d) $k$ on Espgame" + ], + "image_footnote": [], + "bbox": [ + 696, + 215, + 856, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "anced data. The ablation study demonstrates the effectiveness and reasonableness of each proposed module.", + "bbox": [ + 498, + 378, + 890, + 410 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/233c17f7029dc754f313501a6537ac6401d560c0816d26f27d48ca8549114c72.jpg", + "table_caption": [ + "Table 3. Results of ablation experiments on the three representative datasets. Vanilla Transformer denotes the encoders of four vanilla transformers, $\\mathbf{M}_1$ represents base view-category interactive sharing transformer module, $\\mathbf{M}_2$ signifies the view-category consistency guided embedding enhancement module and $\\mathbf{M}_3$ denotes missing view generation module." + ], + "table_footnote": [], + "table_body": "
MethodCorel5kEspgameIAPRTC12
APAUCAPAUCAPAUC
Vanilla Transformer34.987.827.683.428.187.2
M139.589.629.484.131.687.8
M1 + M240.490.129.684.732.788.3
M1 + M2 + M341.590.230.785.033.988.6
w/o Lα40.689.829.684.532.488.2
", + "bbox": [ + 511, + 507, + 874, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 625, + 617, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel view-category interactive sharing transformer VIST for addressing the DIMVMLC challenge. Our method leverages a two-layer view-category transformer architecture to extract deep representations of both views and labels. Additionally, we employ a KNN-style approach with the multivariate Gaussian distribution to complete the missing views. To further enhance the efficacy of our model, a view-category consistency guided embedding enhancement module is integrated, significantly boosting the embedding quality and classification accuracy. Our comprehensive experimental analysis convincingly demonstrates that our method outperforms existing state-of-the-art approaches.", + "bbox": [ + 496, + 650, + 890, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 830, + 638, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (No.62272058, No.62192784, No.62172056), and Beijing Natural Science Foundation (No.4242027, No.4232025).", + "bbox": [ + 498, + 844, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27474", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] P. Duygulu, K. Barnard, J. F. G. de Freitas, and D. A. Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision — ECCV 2002, pages 97–112, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 6", + "[2] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6", + "[3] Zheng Fang and Zhongfei Zhang. Simultaneously combining multi-view multi-label learning with maximum margin classification. In 2012 IEEE 12th International Conference on Data Mining, pages 864-869. IEEE, 2012. 1", + "[4] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 1", + "[5] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc12 benchmark: A new evaluation resource for visual information systems. Workshop Ontoimage, 2006. 6", + "[6] Matthieu Guillaumin, Jakob Verbeek, and Cordelia Schmid. Multimodal semi-supervised learning for image classification. In 2010 IEEE Computer society conference on computer vision and pattern recognition, pages 902-909. IEEE, 2010. 6", + "[7] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6", + "[8] Junlong Li, Peipei Li, Xuegang Hu, and Kui Yu. Learning common and label-specific features for multi-label classification with correlation information. Pattern Recognition, 121:108259, 2022. 1", + "[9] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2022. 1, 2, 6", + "[10] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8807-8815, 2023. 1, 2, 5", + "[11] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8816–8824, 2023. 1, 2, 6", + "[12] Meng Liu, Yong Luo, Dacheng Tao, Chao Xu, and Yonggang Wen. Low-rank multi-view learning in matrix completion for multi-label image classification. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2", + "[13] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893, 2018. 5" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 2703–2709. International Joint Conferences on Artificial Intelligence Organization, 2018. 2, 6", + "[15] Adane Nega Tarekegn, Mario Giacobini, and Krzysztof Michalak. A review of methods for imbalanced multi-label classification. Pattern Recognition, 118:107965, 2021. 5", + "[16] Daniel J Trosten, Sigurd Lokse, Robert Jenssen, and Michael Kampffmeyer. Reconsidering representation alignment for multi-view clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1255-1265, 2021. 1", + "[17] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5", + "[18] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2017. 3", + "[19] Luis von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 319-326, New York, NY, USA, 2004. Association for Computing Machinery. 6", + "[20] Jie Wen, Zheng Zhang, Lunke Fei, Bob Zhang, Yong Xu, Zhao Zhang, and Jinxing Li. A survey on incomplete multiview clustering. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 53(2):1136-1149, 2022. 1", + "[21] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[22] Mengyao Xie, Zongbo Han, Changqing Zhang, Yichen Bai, and Qinghua Hu. Exploring and exploiting uncertainty for incomplete multi-view classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19873-19882, 2023. 1", + "[23] Chang Xu, Dacheng Tao, and Chao Xu. Multi-view learning with incomplete views. IEEE Transactions on Image Processing, 24(12):5812-5825, 2015. 2", + "[24] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16051-16060, 2022. 5", + "[25] Nan Xu, Yanqing Guo, Xin Zheng, Qianyu Wang, and Xi-angyang Luo. Partial multi-view subspace clustering. In Proceedings of the 26th ACM International conference on multimedia, pages 1794–1801, 2018. 1", + "[26] Lei Yuan, Yalin Wang, Paul M Thompson, Vaibhav A Narayan, Jieping Ye, Alzheimer's Disease Neuroimaging Initiative, et al. Multi-source feature learning for joint analysis of incomplete multiple heterogeneous neuroimaging data. NeuroImage, 61(3):622-632, 2012. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Changqing Zhang, Ziwei Yu, Qinghua Hu, Pengfei Zhu, Xinwang Liu, and Xiaobo Wang. Latent semantic aware multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2", + "[28] Min-Ling Zhang, Yu-Kun Li, Hao Yang, and Xu-Ying Liu. Towards class-imbalance aware multi-label learning. IEEE Transactions on Cybernetics, 52(6):4459-4471, 2022. 5", + "[29] Wei Zhang, Ke Zhang, Pan Gu, and Xiangyang Xue. Multiview embedding learning for incompletely labeled data. In Twenty-Third International Joint Conference on Artificial Intelligence, 2013. 2", + "[30] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern recognition, 84:126-135, 2018. 1", + "[31] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 6" + ], + "bbox": [ + 78, + 90, + 468, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27476", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_model.json b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c2f3b530fe94a2505b7694c9aa7d1ae425579166 --- /dev/null +++ b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.204, + 0.844, + 0.239 + ], + "angle": 0, + "content": "Shilong Ou, Zhe Xue*, Yawen Li, Meiyu Liang, Yuanqiang Cai, Junjiang Wu \nBeijing University of Posts and Telecommunications, China" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.242, + 0.844, + 0.258 + ], + "angle": 0, + "content": "{os1, xuezhe, meiyul210, caiyuanqiang, wujunjiang}@bupt.edu.cn, warmly0716@126.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.473, + 0.673 + ], + "angle": 0, + "content": "As a problem often encountered in real-world scenarios, multi-view multi-label learning has attracted considerable research attention. However, due to oversights in data collection and uncertainties in manual annotation, real-world data often suffer from incompleteness. Regrettably, most existing multi-view multi-label learning methods sidestep missing views and labels. Furthermore, they often neglect the potential of harnessing complementary information between views and labels, thus constraining their classification capabilities. To address these challenges, we propose a view-category interactive sharing transformer tailored for incomplete multi-view multi-label learning. Within this network, we incorporate a two-layer transformer module to characterize the interplay between views and labels. Additionally, to address view incompleteness, a KNN-style missing view generation module is employed. Finally, we introduce a view-category consistency guided embedding enhancement module to align different views and improve the discriminating power of the embeddings. Collectively, these modules synergistically integrate to classify the incomplete multi-view multi-label data effectively. Extensive experiments substantiate that our approach outperforms the existing state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.701, + 0.21, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.47, + 0.879 + ], + "angle": 0, + "content": "The proliferation of data sources, coupled with advancements in data collection techniques, enables the extraction of diverse types of features, extending beyond a single perspective. This leads to an intriguing research question: how to effectively integrate multi-view features and extract meaningful features from them. Simultaneously, it is observed that single-label data often falls short in encapsulating the complexities of real-world scenarios. For example, an image of a flower may also warrant labels like 'leaf' or 'garden', illustrating that multi-labels provide a richer cate" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.892, + 0.354 + ], + "angle": 0, + "content": "gory feature space and more accurately preserve inter-label relationships [8]. Given its practical relevance, the field of multi-view multi-label learning emerges as a domain of substantial research interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.893, + 0.581 + ], + "angle": 0, + "content": "In the field of multi-view learning, multi-view multi-label classification (MVMLC) poses a significant challenge. Considering the consistency across views and the complexity of labels, it is essential for methods to simultaneously align views and extract diverse features from an integrated representation. In recent years, numerous researchers have made strides towards this objective. One such effort is the introduction of a multi-view label embedding model [30], which employs the Hilbert-Schmidt Independence Criterion [4] to establish a link between view feature and category feature spaces. Concurrently, deep neural networks have been employed in this realm. A noteworthy development is the SIMM neural network [3], which utilizes adversarial and label losses to identify shared features while implementing regularization to capture view-specific details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.884 + ], + "angle": 0, + "content": "While MVMLC has attracted considerable research attention, it often operates under the presumption that all views and labels are complete. This assumption does not consistently align with real-world scenarios. For instance, a video might lack audio or text content. Multimedia data shared on social media networks might lack comprehensive annotations from users. Such instances culminate in the pervasive challenge of missing multi-view features and incomplete category information within real-world datasets. Some methods aim to address this problem by either masking or restoring the missing views [9, 20, 22, 25]. Although these approaches for handling incomplete multi-view or multi-label learning have achieved notable results, they often fall short in simultaneously addressing both types of incompleteness. To handle the double incomplete multi-view multi-label learning problem (DIMVMLC), several deep contrastive networks are proposed [10, 16]. Furthermore, an incomplete multi-view multi-label learning method based on transformers have been proposed [11], showing adaptability to multi-view and multi-label datasets." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "However, most existing methods for double incomplete" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "multi-view multi-label classification overlook two critical aspects. Firstly, while some methods recognize the importance of complementary information across multiple views, they often underestimate the complementarity between multi-view features and categories. Efficiently facilitating interaction between views and categories can dramatically enhance the ability to counteract the simultaneous absence of both multi-view feature and category information. Secondly, these methods frequently overlook the issue of missing multi-view features and do not consider completing missing multi-view features. This negligence makes them susceptible to vulnerabilities in insufficient view features, thereby limiting their classification performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.73 + ], + "angle": 0, + "content": "To overcome the prevailing challenges in double incomplete multi-view multi-label learning, we propose the View-category Interactive Sharing Transformer (VIST), as shown in Fig. 1. VIST is an architecturally advanced framework specifically designed to exploit the synergy between multiple views and categories and adeptly generate missing views. The dual capability of this method not only amplifies the effectiveness of multi-view embeddings through high-level semantics of categories but also significantly enhances the precision of multi-label classification by utilizing these enriched embeddings. This advancement inherently fortifies the method's resilience, particularly in scenarios characterized by the lack of specific views and labels, thereby ensuring robust and reliable performance. VIST integrates three key components: view-category interactive sharing transformer, missing view generation, and view-category consistency guided embedding enhancement. The view-category interactive sharing transformer, a compact two-layer network with four transformer blocks, fosters an interaction between multi-view and category information, uncovering their complementary potential. The missing view generation, adopting a KNN-style approach, adeptly completes missing views in line with each sample's distribution. The view-category consistency guided embedding enhancement module utilizes contrastive learning strategy to sharpen the discriminative strength of embeddings and boost classification accuracy. Extensive experiments demonstrate the excellent performance of VIST. In summary, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.84 + ], + "angle": 0, + "content": "- We introduce a novel view-category interactive sharing transformer for incomplete multi-view multi-label learning. Our method can effectively exploit the complementary information between views and categories through the interaction between multi-view feature and category information, thereby effectively counteracting the challenges posed by incomplete views and labels." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We propose a missing view generation method specifically designed for the incomplete multi-view multi-label learning task. This method combines the KNN strategy with a multivariate Gaussian distribution, ensuring a sta" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "tistically sound imputation of missing views that are cohesively aligned with the corresponding samples." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.895, + 0.242 + ], + "angle": 0, + "content": "- To bolster the discriminating power of the multi-view embeddings, we develop a view-category consistency guided embedding enhancement module. It adopts contrastive learning to align embeddings across different views and leverages category information to guide embedding learning process, significantly boosting the effectiveness of the resulting embeddings and the performance of multi-label classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Many methods have been proposed for the MVMLC task. IrMMC [12] ensures the common subspace remains low rank, making it compatible with matrix completion, while also learning combination weights to tap into the distinct strengths of each view. Matrix factorization is utilized to draw out complementary data across different views [27]. Despite their strengths, these models struggles with handling datasets that have missing views or labels. To solve this, MVL-IV [23] delves into incomplete multi-view learning by leveraging the relationships between views. The iMSF model is another incomplete multi-view single-label learning approach. It ingeniously splits the incomplete multi-view classification responsibilities into several complete subtasks [26]. However, both MVL-IV and iMSF have a common shortcoming: they only address view incompleteness. Conversely, MvEL targets the extraction of context and neighborhood consistency but is tailored only for the incomplete multi-label scenario [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.902 + ], + "angle": 0, + "content": "On the other hand, for double incomplete multi-view multi-label learning task, iMvWL [14] and NAIML [9], have ventured into addressing the incompleteness in both views and labels. iMvWL aligns multi-view features and multi-label data into a shared subspace, enriched by a correlation matrix that bolsters the projection from label space to the embedding subspace. NAIML addresses this dual incompleteness by leaning on the low-rank assumption of the sub-label matrix, subtly harnessing sub-class correlations. In recent years, deep learning models have been employed to address this issue. DDINet [21], DICNet [10] and LMVCAT [11] are deep learning models for DIMVMLC. DDINet is the first deep learning model which contains a view-specific decoder network, successfully preserving the key information of raw views. DICNet is the first deep learning model for DIMVMLC which employs contrastive learning at the instance-level, guiding model to extract deep cross-view features. LMVCAT is a transformer-based double incomplete multi-view multi-label learning framework, consisting of two transformers specifically tailored for views and categories, and constructs a graph on incomplete labels to guide the encoder in extracting view-specific features." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27468" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.212, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.26, + 0.131 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.473, + 0.457 + ], + "angle": 0, + "content": "Given \\( n \\) multi-view data points from \\( V \\) views, they can be represented as a set of matrices \\( \\{\\mathbf{X}_v\\}_{v=1}^V \\), with \\( \\mathbf{X}_v \\in \\mathbb{R}^{n \\times d_v} \\). Here, \\( d_v \\) denotes the dimension of the \\( v \\)-th view. The label matrix for all samples is denoted by \\( \\mathbf{Y} \\in \\{0,1\\}^{n \\times L} \\), where \\( L \\) is the number of categories. If \\( \\mathbf{Y}_{i,j} = 1 \\), it indicates that the \\( i \\)-th sample belongs to category \\( j \\); if \\( \\mathbf{Y}_{i,j} = 0 \\), it indicates non-membership in this category. To denote multi-view data with missing features, we introduce the missing-view indicator matrix \\( \\mathbf{W} \\in \\{0,1\\}^{n \\times V} \\). \\( \\mathbf{W}_{i,j} = 1 \\) means that the \\( j \\)-th view of the \\( i \\)-th sample is available, whereas \\( \\mathbf{W}_{i,j} = 0 \\) indicates unavailable. Similarly, we define the missing-label indicator matrix \\( \\mathbf{U} \\in \\{0,1\\}^{n \\times L} \\), where \\( \\mathbf{U}_{i,j} = 1 \\) indicates that the \\( j \\)-th category of the \\( i \\)-th sample is available, and \\( \\mathbf{U}_{i,j} = 0 \\) signifies the opposite. During the data pre-processing phase, all missing information for view \\( \\mathbf{X}_v \\) and label \\( \\mathbf{Y} \\) is assigned the value 0. The objective of our double incomplete multi-view multi-label learning task is to train a model capable of accurately predicting multiple categories for each sample. In subsequent sections, the detailed mechanisms of the proposed VIST are presented." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.47, + 0.498 + ], + "angle": 0, + "content": "3.2. View-Category Interactive Sharing Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Different views harbor complementary information, necessitating their interaction for effective utilization. Furthermore, category information embodies the data's high-level semantics, serving to augment the discriminative capability of multi-view representations. Inversely, multi-view features can remedy inaccuracies in label information comprehension due to incomplete labels. To this end, we introduce a view-category interactive sharing transformer. This novel transformer is designed to enhance information exchange between views and categories, consequently boosting the discriminative strength of multi-view representations and refining the precision of multi-label classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The view-category interactive sharing transformer comprises two key layers. The first layer focuses on extracting shared features that captures correlations among multiple views and categories. An essential preliminary step involves embedding the multi-view data \\(\\mathbf{X}_v\\in \\mathbb{R}^{n\\times d_v}\\). This embedding process ensures the homogenization of feature dimensions across views. For a given sample \\(\\mathbf{x}_v\\in \\mathbb{R}^{d_v}\\) from \\(\\mathbf{X}_v\\), the embedding vector \\(\\mathbf{e}_v\\in \\mathbb{R}^{d_e}\\) can be denoted as \\(\\mathbf{e}_v = Embedding(\\mathbf{x}_v)\\), where \\(d_{e}\\) denotes the dimension of embedding space, and \\(Embedding(\\cdot)\\) is a fully connected neural network. Then we stack the embedding vectors to obtain an original multi-view embedding sequence \\(\\mathbf{E} = [\\mathbf{e}_1,\\mathbf{e}_2,\\dots ,\\mathbf{e}_V]\\in \\mathbb{R}^{V\\times d_e}\\), which is further used as input vectors for the transformer. Noted that for the incom" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "plete multi-view data, we adopt missing view generation method (as detailed in Sec. 3.3) to produce embeddings of missing views, ensuring that \\(\\mathbf{E}\\) is complete across all views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.302 + ], + "angle": 0, + "content": "The structure of our transformer is the encoder module of the classical transformer [18]. A transformer block is denoted by \\( \\text{Transformer}(\\Phi) \\), and the input of the transformer is \\( \\Phi \\in \\mathbb{R}^{t \\times d_t} \\), where \\( t \\) and \\( d_t \\) denote the number of tokens and the dimension of tokens. The first layer consists of two transformer blocks, denoted as \\( \\text{Transformer}_{sv} \\) and \\( \\text{Transformer}_{sc} \\). By discerning correlations among original views, \\( \\text{Transformer}_{sv} \\) serves to complement information across original views and get multi-view embedding \\( \\tilde{\\mathbf{E}} \\), which is the enhanced representation of original views. It can be illustrated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.312, + 0.892, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {E}} = \\text {T r a n s f o r m e r} _ {s v} (\\mathbf {E}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.401 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{E}} = [\\tilde{\\mathbf{e}}_1,\\tilde{\\mathbf{e}}_2,\\dots ,\\tilde{\\mathbf{e}}_V]\\in \\mathbb{R}^{V\\times d_e}\\). Here, an adaptive fusion layer is introduced to fuse the information from multiple views into a shared view feature \\(s_v\\). The fusion process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.409, + 0.892, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {v} = \\sum_ {v = 1} ^ {V} \\frac {e ^ {\\theta_ {v} ^ {\\epsilon}} \\tilde {\\mathbf {e}} _ {v}}{\\sum_ {j} e ^ {\\theta_ {j} ^ {\\epsilon}}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.553 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{e}}_v\\in \\mathbb{R}^{d_e}\\) is the \\(v\\) -th row of \\(\\tilde{\\mathbf{E}}\\) \\(\\theta_v\\) denotes the learnable weight and \\(\\epsilon\\) is a adjustment factor. By interacting with the shared view feature \\(\\mathbf{s}_v\\) and discerning correlations among original categories, \\(T r a n s f o r m e r_{s c}\\) serves to complement information across original categories and get category vectors \\(\\tilde{\\mathbf{C}}\\) , which can be illustrated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.563, + 0.892, + 0.582 + ], + "angle": 0, + "content": "\\[\n[ \\tilde {\\mathbf {s}} _ {v}, \\tilde {\\mathbf {C}} ] = T r a n s f o r m e r _ {s c} ([ \\mathbf {s} _ {v}, \\mathbf {C} ]) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.652 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) denotes the concatenation operation and \\(\\mathbf{C} \\in \\mathbb{R}^{L \\times d_e}\\) denotes original category vectors which are randomly initialized. Subsequently, the output feature of \\(Transformer_{sc}\\), \\(\\tilde{\\mathbf{s}}_v\\), is propagated into the second layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.893, + 0.833 + ], + "angle": 0, + "content": "The second layer is designed to extract advanced shared features, which is achieved by promoting the interaction and fusion between view and category information based on the features extracted from the first layer, thereby obtaining a more discriminative multi-view data representation. This layer incorporates two transformer blocks, denoted as \\( \\text{Transformer}_{av} \\) and \\( \\text{Transformer}_{ac} \\). \\( \\text{Transformer}_{av} \\) is employed to complement information across views and extract advanced multi-view embedding \\( \\bar{\\mathbf{E}} \\), which is the enhanced representation of views, by interacting with the shared category feature \\( \\mathbf{s}_c \\) and discerning view correlations, as represented below:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.844, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {c} = \\operatorname {P r o j} _ {c v} (\\tilde {\\mathbf {s}} _ {v}) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n[ \\bar {\\mathbf {s}} _ {c}, \\bar {\\mathbf {E}} ] = T r a n s f o r m e r _ {a v} ([ \\mathbf {s} _ {c}, \\tilde {\\mathbf {E}} ]) \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27469" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.094, + 0.812, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.894, + 0.393 + ], + "angle": 0, + "content": "Figure 1. An overview of our method VIST. It comprises three modules, namely View-Category Interactive Sharing Transformer Module, Missing View Generation Module, and View-Category Consistency Guided Embedding Enhancement Module. Different colors represent" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.472, + 0.523 + ], + "angle": 0, + "content": "different samples, while different shapes signify different features. where \\(Proj_{cv}(\\cdot)\\) is a linear layer as a projection function designed to map vectors from the category feature space to the view feature space, \\(\\bar{\\mathbf{E}} = [\\bar{\\mathbf{e}}_1,\\bar{\\mathbf{e}}_2,\\dots ,\\bar{\\mathbf{e}}_V]\\in \\mathbb{R}^{V\\times d_e}\\). Correspondingly, \\(Transformer_{ac}\\) is employed to complement information across categories and extract advanced category vectors \\(\\bar{\\mathbf{C}}\\) by leveraging the advanced shared feature \\(\\mathbf{p}_v\\) and discerning category correlations, as illustrated below:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.525, + 0.47, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {v} = \\operatorname {P r o j} _ {v c} (\\overline {{\\mathbf {s}}} _ {c}) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.56, + 0.47, + 0.578 + ], + "angle": 0, + "content": "\\[\n[ \\bar {\\mathbf {p}} _ {v}, \\bar {\\mathbf {C}} ] = T r a n s f o r m e r _ {a c} ([ \\mathbf {p} _ {v}, \\bar {\\mathbf {C}} ]) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.689 + ], + "angle": 0, + "content": "where \\(Proj_{vc}(\\cdot)\\) is a linear layer as a projection function to map vectors from the view feature space to the category feature space. Through the propagation of vectors \\(\\mathbf{s}_v\\), \\(\\mathbf{s}_c\\), and \\(\\mathbf{p}_v\\) among the transformer blocks, we facilitate the sharing of information between the view and category feature spaces, thereby extracting more refined and effective features of views and categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.308, + 0.712 + ], + "angle": 0, + "content": "3.3. Missing View Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Considering the absence of views in the dataset, an intuitive consideration is that the proper completion of these missing views can enhance the model's performance. Through the contrastive learning process mentioned in Sec. 3.4, the distance between views of a sample tends to be relatively proximate. Thus, generating missing views utilizing the existing ones is a plausible approach. We adopt a method analogous to \\( k \\)-nearest neighbors, leveraging original category vectors \\( \\mathbf{C} = [\\mathbf{c}_1, \\mathbf{c}_2, \\dots, \\mathbf{c}_L] \\) to assist in generating the missing views. Specifically, for a certain sample \\( i \\), let \\( \\mathcal{E} = \\{v | \\mathbf{W}_{i,v} = 1\\} \\) denote the index of existing views, and \\( \\mathcal{M} = \\{v | \\mathbf{W}_{i,v} = 0\\} \\) denote the index of missing views. To" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.465 + ], + "angle": 0, + "content": "complete the missing features of sample \\(i\\) with its original multi-view embedding \\(\\mathbf{E}\\), we first find the \\(k\\) nearest neighbors in the projected category feature space. The set of the neighbors \\(\\mathcal{D}\\) is constructed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.474, + 0.892, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = \\left\\{d \\mid T o p K \\left(\\sum_ {j \\in \\mathcal {E}} \\| \\mathbf {e} _ {j} - \\mathbf {c} _ {d} \\| _ {2}, d \\in \\{1, 2, 3, \\dots , L \\}\\right) \\right\\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.638 + ], + "angle": 0, + "content": "where \\(TopK(\\cdot)\\) is a function designed to identify the indices of the top \\(k\\) categories, based on the smallest distance between embedding vectors and category vectors. Then, we employ a statistical method to describe the distribution of the missing views. We hypothesize that these missing views \\(\\{\\mathbf{e}_m\\}_{m\\in \\mathcal{M}}\\) adhere to a multivariate Gaussian distribution \\(\\mathcal{N}(\\mu ,\\Sigma)\\), with a mean vector and covariance matrix denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.637, + 0.892, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\mu = \\frac {\\sum_ {d \\in \\mathcal {D}} \\mathbf {c} _ {d}}{| \\mathcal {D} |} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.678, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\frac {1}{| \\mathcal {D} | - 1} \\sum_ {d \\in \\mathcal {D}} \\left(\\mathbf {c} _ {d} - \\mu\\right) \\left(\\mathbf {c} _ {d} - \\mu\\right) ^ {T} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.809 + ], + "angle": 0, + "content": "For the missing views, we sample from this distribution \\( |\\mathcal{M}| \\) times and substitute the missing views with the sampled results. Consequently, we can obtain the complete embeddings for the incomplete multi-view data. By reconstructing the absent multi-view data, our method further enhances its performance in category predictions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.817, + 0.892, + 0.849 + ], + "angle": 0, + "content": "3.4. View-Category Consistency Guided Embedding Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Through the aforementioned view-category interactive sharing transformer, we extract three multi-view embeddings \\(\\mathbf{E}\\), \\(\\tilde{\\mathbf{E}}\\) and \\(\\bar{\\mathbf{E}}\\) from different layers. In order to obtain" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.241 + ], + "angle": 0, + "content": "better classification performance, enhancing the discriminating power and effectiveness of these embeddings is of great importance. Specifically, predicated on the view consistency assumption [10, 24], the embedding \\(\\mathbf{E}\\) of a sample on different views should be aligned. Moreover, to improve the discriminating power of \\(\\tilde{\\mathbf{E}}\\) and \\(\\bar{\\mathbf{E}}\\), the consistency between category correlations and feature correlations of multi-view data can be leveraged. In light of these factors, we introduce view-category consistency guided embedding enhancement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.472, + 0.303 + ], + "angle": 0, + "content": "To learn more effective embedding \\(\\mathbf{E}\\), we use contrastive learning to align the embeddings of the same sample across different views. According to [13, 17], the loss function of contrastive learning can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.312, + 0.469, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} - \\log \\frac {e ^ {\\operatorname {s i m} (x , x ^ {+})}}{e ^ {\\operatorname {s i m} (x , x ^ {+})} + \\sum_ {x ^ {-} \\in \\mathcal {F} (x)} e ^ {\\operatorname {s i m} (x , x ^ {-})}} \\tag {11} \\\\ \\approx - s i m (x, x ^ {+}) + \\log \\sum_ {x ^ {-} \\in \\mathcal {F} (x)} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.457 + ], + "angle": 0, + "content": "where \\(x, x^{+}, x^{-}\\) denotes the anchor sample, its positive sample and negative sample. \\(\\mathcal{F}(x)\\) denotes the set of negative pairs. \\(sim(\\cdot, \\cdot)\\) calculates the similarity between two samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.473, + 0.587 + ], + "angle": 0, + "content": "Evidently, our objective is to maximize the similarity \\(sim(x,x^{+})\\) while simultaneously minimizing the similarity \\(\\sum_{x^{-}\\in \\mathcal{F}(x)}e^{sim(x,x^{-})}\\). We denote \\(\\mathbf{e}_{i,v}\\) as the embedding of the \\(v\\)-th view for the \\(i\\)-th sample, then \\(\\mathcal{F}(x)\\) can be rewritten as \\(\\mathcal{F}(i,v) = \\{j,w|\\sum^L\\mathbf{y}_i\\circ \\mathbf{u}_i\\circ \\mathbf{y}_j\\circ \\mathbf{u}_j = 0,\\mathbf{U}_{j,w} = 1\\}\\), where \\(\\circ\\) denotes Hadamard product, \\(\\mathbf{u}_i\\) denotes the missing vector of the \\(i\\)-th sample from \\(U\\). As for maximizing \\(e^{sim(x,x^{+})}\\), we use the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.594, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p o s} = \\sum_ {v _ {1} = 1} ^ {V} \\sum_ {v _ {2} = 1, v _ {2} \\neq v _ {1}} ^ {V} \\frac {1}{d _ {e}} \\| n \\left(\\mathbf {e} _ {i, v _ {1}}\\right) - n \\left(\\mathbf {e} _ {i, v _ {2}}\\right) \\| _ {2} ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.647, + 0.47, + 0.695 + ], + "angle": 0, + "content": "where \\( n(\\cdot) \\) denotes \\( L_{2} \\) normalization function. To minimize \\( \\sum_{x^{-}\\in \\mathcal{F}(x)}e^{sim(x,x^{-})} \\), the following loss function can be adopted:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.703, + 0.47, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n e g} = \\sum_ {v _ {1} = 1} ^ {V} \\sum_ {j, v _ {2} \\in \\mathcal {F} (i, v _ {1})} \\frac {- 1}{d _ {e}} \\| n \\left(\\mathbf {e} _ {i, v _ {1}}\\right) - n \\left(\\mathbf {e} _ {j, v _ {2}}\\right) \\| _ {2} ^ {2} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.835 + ], + "angle": 0, + "content": "To further improve the effectiveness of \\(\\tilde{\\mathbf{E}}\\) and \\(\\bar{\\mathbf{E}}\\), we utilize the category information to guide the learning process of these embeddings. We enhance the consistency of data correlations in the view feature space and the category feature space as detailed below:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.843, + 0.47, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {v} ^ {1} (i, j) = \\left(n \\left(\\tilde {\\mathbf {e}} _ {i, v}\\right) \\left(n \\left(\\tilde {\\mathbf {e}} _ {j, v}\\right)\\right) ^ {T} + 1\\right) / 2 \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.884, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {v} ^ {2} (i, j) = \\left(n \\left(\\bar {\\mathbf {e}} _ {i, v}\\right) \\left(n \\left(\\bar {\\mathbf {e}} _ {j, v}\\right)\\right) ^ {T} + 1\\right) / 2 \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.105, + 0.892, + 0.123 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\left(\\left(\\mathbf {Y} \\circ \\mathbf {U}\\right) \\left(\\mathbf {Y} \\circ \\mathbf {U}\\right) ^ {T}\\right). / \\left(\\mathbf {U U} ^ {T}\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.894, + 0.235 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_v^1\\), \\(\\mathbf{G}_v^2\\) and \\(\\mathbf{H}\\) are the correlation matrices of \\(\\tilde{\\mathbf{E}}\\), \\(\\bar{\\mathbf{E}}\\) and labels respectively. \\(\\mathbf{e}_{i,v}\\) and \\(\\bar{\\mathbf{e}}_{i,v}\\) represent the feature of the \\(v\\)-th view of the \\(i\\)-th sample from \\(\\tilde{\\mathbf{E}}\\) and \\(\\bar{\\mathbf{E}}\\) respectively. Through the cross-entropy loss function, we can align the data correlations across view feature space and category feature space so that more discriminative embedding can be obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.243, + 0.902, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a u x} = \\frac {- 1}{2 n V} \\sum_ {v = 1} ^ {V} \\sum_ {i = 1} ^ {n} \\sum_ {j \\neq i} ^ {n} \\left(\\mathbf {H} _ {i, j} \\left(\\log \\mathbf {G} _ {v} ^ {1} (i, j) + \\log \\mathbf {G} _ {v} ^ {2} (i, j)\\right) \\right. \\\\ + \\left(\\mathbf {U} _ {i, j} - \\mathbf {H} _ {i, j}\\right) \\left(\\log \\left(1 - \\mathbf {G} _ {v} ^ {1} (i, j)\\right) + \\log \\left(1 - \\mathbf {G} _ {v} ^ {2} (i, j)\\right)\\right)) \\tag {17} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.32, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Considering the above factors, the loss function for view-category consistency guided embedding enhancement can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.375, + 0.892, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\mathcal {L} _ {\\text {p o s}} + \\mathcal {L} _ {\\text {n e g}} + \\alpha \\mathcal {L} _ {\\text {a u x}} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.401, + 0.892, + 0.461 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a weight parameter. Through optimizing the above loss function, the learned data embeddings become more discriminative, laying a solid foundation for subsequent multi-label classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.707, + 0.485 + ], + "angle": 0, + "content": "3.5. Overall Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.618 + ], + "angle": 0, + "content": "To achieve multi-label classification results, we partition the output \\([\\bar{\\mathbf{p}}_v,\\bar{\\mathbf{C}} ]\\) of \\(Transforer_{ac}\\) into \\(\\bar{\\mathbf{p}}_v\\) and \\(\\bar{\\mathbf{C}}\\). Specifically, for sample \\(i\\), \\(\\bar{\\mathbf{p}}_v^i\\) passes through a linear layer to yield \\(\\mathbf{p}_f^i\\in [0,1]^L\\), representing the final complementary information between views and categories. \\(\\bar{\\mathbf{C}}^i\\) is processed through a set of category-aware linear layers, resulting in \\(\\mathbf{z}^i\\in [0,1]^L\\), which signifies category-aware predictions. With the assistance of \\(\\mathbf{p}_f^i\\), we get the final prediction \\(\\hat{\\mathbf{y}}^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.629, + 0.892, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} ^ {i} = \\operatorname {c l a s s i f i e r} \\left(\\mathbf {p} _ {f} ^ {i} + \\mathbf {z} ^ {i}\\right) \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.652, + 0.892, + 0.697 + ], + "angle": 0, + "content": "where classifier \\((\\cdot)\\) is a linear classifier. We utilize the masked cross-entropy function to compute the multi-label classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.704, + 0.912, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\frac {\\sum_ {i = 1} ^ {n} \\sum_ {l = 1} ^ {L} \\left(\\mathbf {Y} _ {i , l} \\log \\hat {\\mathbf {y}} _ {l} ^ {i} + \\left(1 - \\mathbf {Y} _ {i , l}\\right) \\log \\left(1 - \\hat {\\mathbf {y}} _ {l} ^ {i}\\right)\\right) \\mathbf {U} _ {i , l}}{- \\sum_ {i , l} \\mathbf {U} _ {i , l}} \\tag{20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.892, + 0.874 + ], + "angle": 0, + "content": "It is important to note that class imbalance is an inherent characteristic of multi-label data, impacting the generalization performance of multi-label prediction [15, 28]. Typically, for certain class labels, the instances of positive label assignments in training data are significantly fewer than those of negative assignments. To address this class-imbalance issue, we introduce a masked asymmetric loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {l} ^ {i} = \\left(\\max \\left(\\hat {\\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\\right)\\right) ^ {2} \\log \\left(1 - \\max \\left(\\hat {\\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\\right)\\right) \\tag {21}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27471" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a} = \\frac {\\sum_ {i = 1} ^ {n} \\sum_ {l = 1} ^ {L} \\left(\\mathbf {Y} _ {i , l} \\log \\hat {\\mathbf {y}} _ {l} ^ {i} + \\left(1 - \\mathbf {Y} _ {i , l}\\right) \\mathbf {L} _ {l} ^ {i}\\right) \\mathbf {U} _ {i , l}}{- \\sum_ {i , l} \\mathbf {U} _ {i , l}} \\tag{22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.163, + 0.47, + 0.191 + ], + "angle": 0, + "content": "Ultimately, we derive the overall loss function of the model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.193, + 0.469, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\mathcal {L} _ {m} + \\beta \\mathcal {L} _ {a} + \\gamma \\mathcal {L} _ {c} \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.216, + 0.469, + 0.275 + ], + "angle": 0, + "content": "where \\(\\beta, \\gamma\\) denotes different weight parameters. By optimizing \\(\\mathcal{L}_o\\), our method proficiently learns the embeddings for incomplete multi-view data and consequently produces accurate multi-label classification results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.29, + 0.21, + 0.307 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.278, + 0.33 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.337, + 0.192, + 0.351 + ], + "angle": 0, + "content": "4.1.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.361, + 0.469, + 0.391 + ], + "angle": 0, + "content": "We follow [9, 11, 14] and evaluate the performance of the proposed VIST on five datasets. These datasets encompass:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.392, + 0.469, + 0.436 + ], + "angle": 0, + "content": "- Corel 5k [1]: The Corel5k dataset is a standard dataset frequently utilized in image annotation and retrieval research, which has 5000 images with 260 types of tags." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.437, + 0.469, + 0.496 + ], + "angle": 0, + "content": "- Pascal07 [2]: Pascal07, also known as PASCAL VOC 2007, is a renowned dataset for visual object classes, instrumental in object recognition and classification. It consists of 9963 images and 20 types of tags." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.497, + 0.469, + 0.557 + ], + "angle": 0, + "content": "- Espgame [19]: The Espgame dataset is derived from an online game where participants generate labels for images, which contains 20770 samples and 268 types of tags." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.558, + 0.469, + 0.632 + ], + "angle": 0, + "content": "- IAPRTC12 [5]: IAPRTC12 is an international image annotation dataset, comprising approximately 20,000 images and 290 categories. These images cover a wide range of subjects and scenes, each accompanied by detailed descriptive text." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.633, + 0.469, + 0.691 + ], + "angle": 0, + "content": "- Mirflickr [7]: The Mirflickr dataset consists of 25,000 images and 38 types of tags sourced from Flickr. Selected from the public domain, these images include a variety of themes and scenes." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.392, + 0.469, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.469, + 0.798 + ], + "angle": 0, + "content": "We adopt the multi-view feature extraction method outlined in [6]. The images in these datasets undergo a comprehensive preprocessing procedure that transforms them into six distinct views. These views are GIST, HSV, HUE, LAB, RGB, and SIFT. Each of these views offers a unique perspective on the image data, capturing different aspects of visual information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.295, + 0.832 + ], + "angle": 0, + "content": "4.1.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For datasets, in line with [14], we process the five datasets to emulate an incomplete scenario as described below: For every view, we randomly exclude \\(50\\%\\) of the samples, ensuring that each sample retains at least one view. Within" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "each category, we arbitrarily designate \\(50\\%\\) of both positive and negative tags as missing labels. For parameter settings, we choose \\(d_{e} = 512\\), \\(\\alpha = 0.1\\), \\(\\beta = 0.1\\), \\(\\gamma = 0.1\\) and \\(k = 20\\) for all the experiments. The model are trained for 500 epochs with a batch-size of 128 via SGD optimizer on an NVIDIA RTX A6000 GPU. To ensure the credibility of the results, we repeat the experiments 10 times and record the mean and variance of the outcomes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.233, + 0.694, + 0.248 + ], + "angle": 0, + "content": "4.1.3 Compared Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.258, + 0.892, + 0.485 + ], + "angle": 0, + "content": "We choose ten representative methods for comparison. Out of these, eight methods - lrMMC, MVL-IV, iMSF, iMvWL, NAIML, DDINet, DICNet and LMVCAT - are detailed in Sec. 2. Additionally, we incorporate GLOCAL [31], to broaden the assessment spectrum. Because iMvWL and NAIML cater to datasets with missing views and labels, we have to modify other methods for consistency. Drawing inspiration from [14] and [9], we input average values from accessible views into lrMMC for the non-operational ones. For MVL-IV and iMSF, absent tags are treated as negative. GLOCAL are executed individually for each view, and we highlight their optimal outcomes. We adhere to the recommended settings for these comparative techniques, as mentioned in their respective publications or code repositories, to maintain an impartial evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.688, + 0.519 + ], + "angle": 0, + "content": "4.1.4 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Similar to [11], we employ Average Precision (AP), Ranking Loss (RL) and Area Under Curve (AUC). For a convenient comparison, we adopt 1-RL for evaluation. A larger value indicates superior model performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.702, + 0.606 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.809 + ], + "angle": 0, + "content": "In Tab. 1, we show the experimental results of all the methods on five datasets with \\(70\\%\\) training samples, \\(50\\%\\) missing views and missing labels. Based on the results from Tab. 1, our method exhibits a distinct advantage across most metrics. For instance, on the AP metric for dataset Corel5k, our method surpasses the second-best method, LMVCAT, by \\(3\\%\\). Besides, the performance of DDINet, DICNet and LMVCAT surpasses that of the first six methods. This can be attributed to their applicability specifically to the DIMVMLC issue. Additionally, they are both deep learning models, their superior outcomes underscore the vast potential deep learning holds for addressing this particular challenge." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While our method primarily addresses the DIMVMLC issue, we also evaluate the model's performance in the full-view and full-label context. The results are presented in Tab. 2. It can be observed that the model still exhibits superior performance in this scenario, suggesting that our method can be effectively applied in MVMLC." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27472" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Experimental results of different methods on the five datasets with \\(70\\%\\) training samples, \\(50\\%\\) missing instances and missing labels. The best results are marked in bold. For simplicity, percent sign is omitted in the following section." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.122, + 0.902, + 0.327 + ], + "angle": 0, + "content": "
DatasetMetriclrMMCMVL-IViMSFGLOCALiMvWLNAIMLDDINetDICNetLMVCATVIST
Corel5kAP24.0±0.224.0±0.118.9±0.228.5±0.428.3±0.730.9±0.436.4±0.138.1±0.438.4±0.441.5±0.2
1-RL76.2±0.275.6±0.170.9±0.580.4±0.386.5±0.387.8±0.287.1±0.088.2±0.488.0±0.290.2±0.2
AUC76.3±0.276.2±0.166.3±0.584.3±0.386.8±0.388.1±0.287.5±0.188.4±0.488.3±0.290.6±0.1
Pascal07AP42.5±0.343.3±0.232.5±0.049.6±0.444.1±1.748.8±0.353.6±0.250.5±1.251.9±0.553.9±0.3
1-RL69.8±0.370.2±0.156.8±0.076.7±0.473.7±0.978.3±0.180.7±0.178.3±0.881.1±0.485.6±0.1
AUC72.8±0.273.0±0.168.6±0.562.0±0.178.6±0.376.7±1.282.7±0.087.6±0.283.4±0.488.3±0.2
EspgameAP18.8±0.018.9±0.010.8±0.022.1±0.224.2±0.324.6±0.228.3±0.129.7±0.229.4±0.430.7±0.3
1-RL77.7±0.177.8±0.072.2±0.278.0±0.480.7±0.181.8±0.281.5±0.083.2±0.182.8±0.284.4±0.0
AUC78.3±0.178.4±0.167.4±0.378.4±0.481.3±0.282.4±0.282.0±0.183.2±0.182.8±0.285.0±0.1
IAPRTC12AP19.7±0.019.8±0.010.1±0.025.6±0.223.5±0.426.1±0.130.3±0.232.3±0.131.7±0.333.9±0.2
1-RL80.1±0.079.9±0.163.1±0.082.5±0.283.3±0.384.8±0.185.3±0.187.3±0.187.0±0.188.4±0.2
AUC80.5±0.080.4±0.166.5±0.183.0±0.183.6±0.285.0±0.185.4±0.087.4±0.187.2±0.188.6±0.1
MirflickrAP44.1±0.144.9±0.132.3±0.053.7±0.249.5±1.255.1±0.259.8±0.258.9±0.559.4±0.560.4±0.3
1-RL80.5±0.080.4±0.166.5±0.183.2±0.183.6±0.285.0±0.186.3±0.086.3±0.486.5±0.387.9±0.1
AUC80.6±0.180.7±0.076.1±0.182.8±0.179.4±1.583.7±0.185.2±0.184.9±0.485.3±0.386.9±0.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.351, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Table 2. Experimental results of different methods on the five datasets with \\(70\\%\\) training samples, full views and labels. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.398, + 0.467, + 0.602 + ], + "angle": 0, + "content": "
DatasetMetricNAIMLDICNetLMVCATVIST
Corel5kAP32.750.952.155.4
1-RL89.092.992.893.5
AUC89.393.193.094.2
Pascal07AP49.660.862.965.5
1-RL79.585.987.889.4
AUC82.287.689.291.2
EspgameAP25.139.138.540.1
1-RL82.587.187.687.7
AUC83.087.488.089.6
IAPRTC12AP26.741.843.645.7
1-RL82.591.291.892.7
AUC83.091.191.892.0
MirflickrAP55.565.968.470.4
1-RL84.789.590.591.3
AUC83.987.688.989.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.798 + ], + "angle": 0, + "content": "Fig. 2 illustrates the performance of the model under various missing ratios. It is interesting that while both incomplete views and incomplete labels detrimentally affect the model's performance, the latter has a more pronounced impact. Moreover, despite our efforts in view imputation, the model's accuracy noticeably diminishes as the view missing ratio escalates. This is primarily due to our reliance on existing views for imputing the missing views. As the number of missing views increases, the imputation algorithm tends to degrade towards random generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.806, + 0.471, + 0.911 + ], + "angle": 0, + "content": "Fig. 3 depicts the performance of the model under various training sample ratios. While the ratio's impact on the model is not markedly significant, a general trend is observed: as the ratio increases, there is a corresponding improvement in the model's performance. Thus, selecting a \\(70\\%\\) training sample ratio for our experiments is a proper choice." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.353, + 0.695, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.353, + 0.871, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.48, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Figure 2. The results on the Corel5k dataset are presented with (a) different missing-view ratios accompanied by a \\(50\\%\\) missing-label ratio and (b) a consistent \\(50\\%\\) missing-view ratio paired with different missing-label ratios." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.557, + 0.695, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.557, + 0.871, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.685, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Figure 3. The results of different training sample ratios on (a) Corel5k dataset and (b) Espgame dataset with \\(50\\%\\) missing-view ratio and \\(50\\%\\) missing-label ratio." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.688, + 0.756 + ], + "angle": 0, + "content": "4.3. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.904 + ], + "angle": 0, + "content": "We evaluate four hyperparameters in our model, i.e., \\(\\alpha\\), \\(\\beta\\), \\(\\gamma\\) and \\(k\\). To assess the influence of these hyperparameters on model performance, we adjust their values within their respective ranges and test their combinations based on the AP metric on different datasets with \\(50\\%\\) missing views, \\(50\\%\\) missing labels and \\(70\\%\\) training sample ratio. Note that \\(k\\) is inherently related to the number of labels, thus, the actual value of \\(\\hat{k}\\) utilized in computation should be determined by \\(\\hat{k}_j = \\frac{k_i}{L_i} L_j\\), where \\(i\\) and \\(j\\) represents the \\(i\\)-th" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27473" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.103, + 0.264, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.216, + 0.216, + 0.227 + ], + "angle": 0, + "content": "(a) Corel5k" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.106, + 0.44, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.216, + 0.396, + 0.228 + ], + "angle": 0, + "content": "(b) Pascal07" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.242, + 0.262, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.355, + 0.219, + 0.367 + ], + "angle": 0, + "content": "(c) Espgame" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.248, + 0.439, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.355, + 0.403, + 0.367 + ], + "angle": 0, + "content": "(d) IAPRTC12" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.421 + ], + "angle": 0, + "content": "Figure 4. The results of different value combinations of hyperparameter \\(\\beta\\) and \\(\\gamma\\) on different datasets. \\(\\alpha\\) and \\(k\\) are respectively set to default values of 0.1 and 20." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.469, + 0.536 + ], + "angle": 0, + "content": "and \\(j\\)-th dataset. Based on Fig. 4, \\(\\beta\\) and \\(\\gamma\\) exhibit comparatively favorable performances within the ranges of [5e-2, 5e-1] and [5e-2, 1e0] respectively on dataset Corel5k. This indicates that the values of \\(\\beta\\) and \\(\\gamma\\) should neither be excessively large nor too small, and the optimal value of \\(\\beta\\) approximated around 0.1, the optimal value of \\(\\gamma\\) should also be selected in the vicinity of 0.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.673 + ], + "angle": 0, + "content": "As for \\(\\alpha\\) and \\(k\\), we assess the impact of varying \\(\\alpha\\) and \\(k\\) on the model under fixed conditions of \\(\\beta = 0.1\\) and \\(\\gamma = 0.1\\) by measuring metrics AP and AUC. As illustrated in Fig. 5, on datasets Corel5k and Espgame, both the AP and AUC metrics remain within a narrow range, indicating that the model is not highly sensitive to \\(\\alpha\\) and \\(k\\). This might be because they operate on the deep-layer mechanism, variations within certain ranges do not significantly impact the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.23, + 0.699 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In the ablation study, we evaluate the performance of each module of VIST on three representative datasets, the results are shown in Tab. 3. We can observe that view-category interactive sharing transformer module plays a pivotal role in enhancing the model's performance. Furthermore, the view-category consistency guided embedding enhancement module is able to improve the discriminating power of embeddings and boost the classification performance. The missing view generation module also contributes to the model's overall efficacy. Additionally, upon the removal of \\(\\mathcal{L}_a\\), there is a noticeable decline in the model's metrics, particularly in the AUC. This suggests that the masked asymmetric loss indeed addresses issues associated with imbal" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.098, + 0.682, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.194, + 0.653, + 0.206 + ], + "angle": 0, + "content": "(a) \\(\\alpha\\) on Corel5k" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.099, + 0.859, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.194, + 0.834, + 0.207 + ], + "angle": 0, + "content": "(b) \\(\\alpha\\) on Espgame" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.216, + 0.68, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.312, + 0.65, + 0.323 + ], + "angle": 0, + "content": "(c) \\(k\\) on Corel5k" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.216, + 0.857, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.312, + 0.832, + 0.324 + ], + "angle": 0, + "content": "(d) \\(k\\) on Espgame" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Figure 5. The results of different values of hyperparameter \\(\\alpha\\) and \\(k\\) on different datasets. \\(\\beta\\) and \\(\\gamma\\) are respectively set to default values of 0.1 and 0.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.38, + 0.891, + 0.411 + ], + "angle": 0, + "content": "anced data. The ablation study demonstrates the effectiveness and reasonableness of each proposed module." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Table 3. Results of ablation experiments on the three representative datasets. Vanilla Transformer denotes the encoders of four vanilla transformers, \\(\\mathbf{M}_1\\) represents base view-category interactive sharing transformer module, \\(\\mathbf{M}_2\\) signifies the view-category consistency guided embedding enhancement module and \\(\\mathbf{M}_3\\) denotes missing view generation module." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.508, + 0.875, + 0.607 + ], + "angle": 0, + "content": "
MethodCorel5kEspgameIAPRTC12
APAUCAPAUCAPAUC
Vanilla Transformer34.987.827.683.428.187.2
M139.589.629.484.131.687.8
M1 + M240.490.129.684.732.788.3
M1 + M2 + M341.590.230.785.033.988.6
w/o Lα40.689.829.684.532.488.2
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.619, + 0.642 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.831 + ], + "angle": 0, + "content": "In this paper, we propose a novel view-category interactive sharing transformer VIST for addressing the DIMVMLC challenge. Our method leverages a two-layer view-category transformer architecture to extract deep representations of both views and labels. Additionally, we employ a KNN-style approach with the multivariate Gaussian distribution to complete the missing views. To further enhance the efficacy of our model, a view-category consistency guided embedding enhancement module is integrated, significantly boosting the embedding quality and classification accuracy. Our comprehensive experimental analysis convincingly demonstrates that our method outperforms existing state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.64, + 0.845 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (No.62272058, No.62192784, No.62172056), and Beijing Natural Science Foundation (No.4242027, No.4232025)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27474" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] P. Duygulu, K. Barnard, J. F. G. de Freitas, and D. A. Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision — ECCV 2002, pages 97–112, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Zheng Fang and Zhongfei Zhang. Simultaneously combining multi-view multi-label learning with maximum margin classification. In 2012 IEEE 12th International Conference on Data Mining, pages 864-869. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.471, + 0.414 + ], + "angle": 0, + "content": "[5] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc12 benchmark: A new evaluation resource for visual information systems. Workshop Ontoimage, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[6] Matthieu Guillaumin, Jakob Verbeek, and Cordelia Schmid. Multimodal semi-supervised learning for image classification. In 2010 IEEE Computer society conference on computer vision and pattern recognition, pages 902-909. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.471, + 0.542 + ], + "angle": 0, + "content": "[7] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[8] Junlong Li, Peipei Li, Xuegang Hu, and Kui Yu. Learning common and label-specific features for multi-label classification with correlation information. Pattern Recognition, 121:108259, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[10] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8807-8815, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8816–8824, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Meng Liu, Yong Luo, Dacheng Tao, Chao Xu, and Yonggang Wen. Low-rank multi-view learning in matrix completion for multi-label image classification. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[14] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 2703–2709. International Joint Conferences on Artificial Intelligence Organization, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[15] Adane Nega Tarekegn, Mario Giacobini, and Krzysztof Michalak. A review of methods for imbalanced multi-label classification. Pattern Recognition, 118:107965, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.894, + 0.291 + ], + "angle": 0, + "content": "[16] Daniel J Trosten, Sigurd Lokse, Robert Jenssen, and Michael Kampffmeyer. Reconsidering representation alignment for multi-view clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1255-1265, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.893, + 0.334 + ], + "angle": 0, + "content": "[17] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.893, + 0.38 + ], + "angle": 0, + "content": "[18] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.893, + 0.453 + ], + "angle": 0, + "content": "[19] Luis von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 319-326, New York, NY, USA, 2004. Association for Computing Machinery. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.894, + 0.511 + ], + "angle": 0, + "content": "[20] Jie Wen, Zheng Zhang, Lunke Fei, Bob Zhang, Yong Xu, Zhao Zhang, and Jinxing Li. A survey on incomplete multiview clustering. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 53(2):1136-1149, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.894, + 0.582 + ], + "angle": 0, + "content": "[21] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.894, + 0.655 + ], + "angle": 0, + "content": "[22] Mengyao Xie, Zongbo Han, Changqing Zhang, Yichen Bai, and Qinghua Hu. Exploring and exploiting uncertainty for incomplete multi-view classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19873-19882, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.893, + 0.699 + ], + "angle": 0, + "content": "[23] Chang Xu, Dacheng Tao, and Chao Xu. Multi-view learning with incomplete views. IEEE Transactions on Image Processing, 24(12):5812-5825, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.893, + 0.77 + ], + "angle": 0, + "content": "[24] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16051-16060, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.829 + ], + "angle": 0, + "content": "[25] Nan Xu, Yanqing Guo, Xin Zheng, Qianyu Wang, and Xi-angyang Luo. Partial multi-view subspace clustering. In Proceedings of the 26th ACM International conference on multimedia, pages 1794–1801, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[26] Lei Yuan, Yalin Wang, Paul M Thompson, Vaibhav A Narayan, Jieping Ye, Alzheimer's Disease Neuroimaging Initiative, et al. Multi-source feature learning for joint analysis of incomplete multiple heterogeneous neuroimaging data. NeuroImage, 61(3):622-632, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Changqing Zhang, Ziwei Yu, Qinghua Hu, Pengfei Zhu, Xinwang Liu, and Xiaobo Wang. Latent semantic aware multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[28] Min-Ling Zhang, Yu-Kun Li, Hao Yang, and Xu-Ying Liu. Towards class-imbalance aware multi-label learning. IEEE Transactions on Cybernetics, 52(6):4459-4471, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[29] Wei Zhang, Ke Zhang, Pan Gu, and Xiangyang Xue. Multiview embedding learning for incompletely labeled data. In Twenty-Third International Joint Conference on Artificial Intelligence, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[30] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern recognition, 84:126-135, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[31] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "27476" + } + ] +] \ No newline at end of file diff --git a/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_origin.pdf b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5293e0a50d5a51c0cd8db8b84e5ee2b6f758da9f --- /dev/null +++ b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/18ba094d-30eb-4383-97ef-520cbe2b100a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f199494d7e2922d9be57555294751ede45e97fd3cab864d6a103ce8d18161b6 +size 2028651 diff --git a/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/full.md b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0101c686c40a7e1f6e32b68346331bd039d4a284 --- /dev/null +++ b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/full.md @@ -0,0 +1,352 @@ +# View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning + +Shilong Ou, Zhe Xue*, Yawen Li, Meiyu Liang, Yuanqiang Cai, Junjiang Wu +Beijing University of Posts and Telecommunications, China + +{os1, xuezhe, meiyul210, caiyuanqiang, wujunjiang}@bupt.edu.cn, warmly0716@126.com + +# Abstract + +As a problem often encountered in real-world scenarios, multi-view multi-label learning has attracted considerable research attention. However, due to oversights in data collection and uncertainties in manual annotation, real-world data often suffer from incompleteness. Regrettably, most existing multi-view multi-label learning methods sidestep missing views and labels. Furthermore, they often neglect the potential of harnessing complementary information between views and labels, thus constraining their classification capabilities. To address these challenges, we propose a view-category interactive sharing transformer tailored for incomplete multi-view multi-label learning. Within this network, we incorporate a two-layer transformer module to characterize the interplay between views and labels. Additionally, to address view incompleteness, a KNN-style missing view generation module is employed. Finally, we introduce a view-category consistency guided embedding enhancement module to align different views and improve the discriminating power of the embeddings. Collectively, these modules synergistically integrate to classify the incomplete multi-view multi-label data effectively. Extensive experiments substantiate that our approach outperforms the existing state-of-the-art methods. + +# 1. Introduction + +The proliferation of data sources, coupled with advancements in data collection techniques, enables the extraction of diverse types of features, extending beyond a single perspective. This leads to an intriguing research question: how to effectively integrate multi-view features and extract meaningful features from them. Simultaneously, it is observed that single-label data often falls short in encapsulating the complexities of real-world scenarios. For example, an image of a flower may also warrant labels like 'leaf' or 'garden', illustrating that multi-labels provide a richer cate + +gory feature space and more accurately preserve inter-label relationships [8]. Given its practical relevance, the field of multi-view multi-label learning emerges as a domain of substantial research interest. + +In the field of multi-view learning, multi-view multi-label classification (MVMLC) poses a significant challenge. Considering the consistency across views and the complexity of labels, it is essential for methods to simultaneously align views and extract diverse features from an integrated representation. In recent years, numerous researchers have made strides towards this objective. One such effort is the introduction of a multi-view label embedding model [30], which employs the Hilbert-Schmidt Independence Criterion [4] to establish a link between view feature and category feature spaces. Concurrently, deep neural networks have been employed in this realm. A noteworthy development is the SIMM neural network [3], which utilizes adversarial and label losses to identify shared features while implementing regularization to capture view-specific details. + +While MVMLC has attracted considerable research attention, it often operates under the presumption that all views and labels are complete. This assumption does not consistently align with real-world scenarios. For instance, a video might lack audio or text content. Multimedia data shared on social media networks might lack comprehensive annotations from users. Such instances culminate in the pervasive challenge of missing multi-view features and incomplete category information within real-world datasets. Some methods aim to address this problem by either masking or restoring the missing views [9, 20, 22, 25]. Although these approaches for handling incomplete multi-view or multi-label learning have achieved notable results, they often fall short in simultaneously addressing both types of incompleteness. To handle the double incomplete multi-view multi-label learning problem (DIMVMLC), several deep contrastive networks are proposed [10, 16]. Furthermore, an incomplete multi-view multi-label learning method based on transformers have been proposed [11], showing adaptability to multi-view and multi-label datasets. + +However, most existing methods for double incomplete + +multi-view multi-label classification overlook two critical aspects. Firstly, while some methods recognize the importance of complementary information across multiple views, they often underestimate the complementarity between multi-view features and categories. Efficiently facilitating interaction between views and categories can dramatically enhance the ability to counteract the simultaneous absence of both multi-view feature and category information. Secondly, these methods frequently overlook the issue of missing multi-view features and do not consider completing missing multi-view features. This negligence makes them susceptible to vulnerabilities in insufficient view features, thereby limiting their classification performance. + +To overcome the prevailing challenges in double incomplete multi-view multi-label learning, we propose the View-category Interactive Sharing Transformer (VIST), as shown in Fig. 1. VIST is an architecturally advanced framework specifically designed to exploit the synergy between multiple views and categories and adeptly generate missing views. The dual capability of this method not only amplifies the effectiveness of multi-view embeddings through high-level semantics of categories but also significantly enhances the precision of multi-label classification by utilizing these enriched embeddings. This advancement inherently fortifies the method's resilience, particularly in scenarios characterized by the lack of specific views and labels, thereby ensuring robust and reliable performance. VIST integrates three key components: view-category interactive sharing transformer, missing view generation, and view-category consistency guided embedding enhancement. The view-category interactive sharing transformer, a compact two-layer network with four transformer blocks, fosters an interaction between multi-view and category information, uncovering their complementary potential. The missing view generation, adopting a KNN-style approach, adeptly completes missing views in line with each sample's distribution. The view-category consistency guided embedding enhancement module utilizes contrastive learning strategy to sharpen the discriminative strength of embeddings and boost classification accuracy. Extensive experiments demonstrate the excellent performance of VIST. In summary, this paper makes the following contributions: + +- We introduce a novel view-category interactive sharing transformer for incomplete multi-view multi-label learning. Our method can effectively exploit the complementary information between views and categories through the interaction between multi-view feature and category information, thereby effectively counteracting the challenges posed by incomplete views and labels. +- We propose a missing view generation method specifically designed for the incomplete multi-view multi-label learning task. This method combines the KNN strategy with a multivariate Gaussian distribution, ensuring a sta + +tistically sound imputation of missing views that are cohesively aligned with the corresponding samples. + +- To bolster the discriminating power of the multi-view embeddings, we develop a view-category consistency guided embedding enhancement module. It adopts contrastive learning to align embeddings across different views and leverages category information to guide embedding learning process, significantly boosting the effectiveness of the resulting embeddings and the performance of multi-label classification. + +# 2. Related Work + +Many methods have been proposed for the MVMLC task. IrMMC [12] ensures the common subspace remains low rank, making it compatible with matrix completion, while also learning combination weights to tap into the distinct strengths of each view. Matrix factorization is utilized to draw out complementary data across different views [27]. Despite their strengths, these models struggles with handling datasets that have missing views or labels. To solve this, MVL-IV [23] delves into incomplete multi-view learning by leveraging the relationships between views. The iMSF model is another incomplete multi-view single-label learning approach. It ingeniously splits the incomplete multi-view classification responsibilities into several complete subtasks [26]. However, both MVL-IV and iMSF have a common shortcoming: they only address view incompleteness. Conversely, MvEL targets the extraction of context and neighborhood consistency but is tailored only for the incomplete multi-label scenario [29]. + +On the other hand, for double incomplete multi-view multi-label learning task, iMvWL [14] and NAIML [9], have ventured into addressing the incompleteness in both views and labels. iMvWL aligns multi-view features and multi-label data into a shared subspace, enriched by a correlation matrix that bolsters the projection from label space to the embedding subspace. NAIML addresses this dual incompleteness by leaning on the low-rank assumption of the sub-label matrix, subtly harnessing sub-class correlations. In recent years, deep learning models have been employed to address this issue. DDINet [21], DICNet [10] and LMVCAT [11] are deep learning models for DIMVMLC. DDINet is the first deep learning model which contains a view-specific decoder network, successfully preserving the key information of raw views. DICNet is the first deep learning model for DIMVMLC which employs contrastive learning at the instance-level, guiding model to extract deep cross-view features. LMVCAT is a transformer-based double incomplete multi-view multi-label learning framework, consisting of two transformers specifically tailored for views and categories, and constructs a graph on incomplete labels to guide the encoder in extracting view-specific features. + +# 3. Methodology + +# 3.1. Problem Definition + +Given $n$ multi-view data points from $V$ views, they can be represented as a set of matrices $\{\mathbf{X}_v\}_{v=1}^V$ , with $\mathbf{X}_v \in \mathbb{R}^{n \times d_v}$ . Here, $d_v$ denotes the dimension of the $v$ -th view. The label matrix for all samples is denoted by $\mathbf{Y} \in \{0,1\}^{n \times L}$ , where $L$ is the number of categories. If $\mathbf{Y}_{i,j} = 1$ , it indicates that the $i$ -th sample belongs to category $j$ ; if $\mathbf{Y}_{i,j} = 0$ , it indicates non-membership in this category. To denote multi-view data with missing features, we introduce the missing-view indicator matrix $\mathbf{W} \in \{0,1\}^{n \times V}$ . $\mathbf{W}_{i,j} = 1$ means that the $j$ -th view of the $i$ -th sample is available, whereas $\mathbf{W}_{i,j} = 0$ indicates unavailable. Similarly, we define the missing-label indicator matrix $\mathbf{U} \in \{0,1\}^{n \times L}$ , where $\mathbf{U}_{i,j} = 1$ indicates that the $j$ -th category of the $i$ -th sample is available, and $\mathbf{U}_{i,j} = 0$ signifies the opposite. During the data pre-processing phase, all missing information for view $\mathbf{X}_v$ and label $\mathbf{Y}$ is assigned the value 0. The objective of our double incomplete multi-view multi-label learning task is to train a model capable of accurately predicting multiple categories for each sample. In subsequent sections, the detailed mechanisms of the proposed VIST are presented. + +# 3.2. View-Category Interactive Sharing Transformer + +Different views harbor complementary information, necessitating their interaction for effective utilization. Furthermore, category information embodies the data's high-level semantics, serving to augment the discriminative capability of multi-view representations. Inversely, multi-view features can remedy inaccuracies in label information comprehension due to incomplete labels. To this end, we introduce a view-category interactive sharing transformer. This novel transformer is designed to enhance information exchange between views and categories, consequently boosting the discriminative strength of multi-view representations and refining the precision of multi-label classification. + +The view-category interactive sharing transformer comprises two key layers. The first layer focuses on extracting shared features that captures correlations among multiple views and categories. An essential preliminary step involves embedding the multi-view data $\mathbf{X}_v\in \mathbb{R}^{n\times d_v}$ . This embedding process ensures the homogenization of feature dimensions across views. For a given sample $\mathbf{x}_v\in \mathbb{R}^{d_v}$ from $\mathbf{X}_v$ , the embedding vector $\mathbf{e}_v\in \mathbb{R}^{d_e}$ can be denoted as $\mathbf{e}_v = Embedding(\mathbf{x}_v)$ , where $d_{e}$ denotes the dimension of embedding space, and $Embedding(\cdot)$ is a fully connected neural network. Then we stack the embedding vectors to obtain an original multi-view embedding sequence $\mathbf{E} = [\mathbf{e}_1,\mathbf{e}_2,\dots ,\mathbf{e}_V]\in \mathbb{R}^{V\times d_e}$ , which is further used as input vectors for the transformer. Noted that for the incom + +plete multi-view data, we adopt missing view generation method (as detailed in Sec. 3.3) to produce embeddings of missing views, ensuring that $\mathbf{E}$ is complete across all views. + +The structure of our transformer is the encoder module of the classical transformer [18]. A transformer block is denoted by $\text{Transformer}(\Phi)$ , and the input of the transformer is $\Phi \in \mathbb{R}^{t \times d_t}$ , where $t$ and $d_t$ denote the number of tokens and the dimension of tokens. The first layer consists of two transformer blocks, denoted as $\text{Transformer}_{sv}$ and $\text{Transformer}_{sc}$ . By discerning correlations among original views, $\text{Transformer}_{sv}$ serves to complement information across original views and get multi-view embedding $\tilde{\mathbf{E}}$ , which is the enhanced representation of original views. It can be illustrated as follows: + +$$ +\tilde {\mathbf {E}} = \text {T r a n s f o r m e r} _ {s v} (\mathbf {E}) \tag {1} +$$ + +where $\tilde{\mathbf{E}} = [\tilde{\mathbf{e}}_1,\tilde{\mathbf{e}}_2,\dots ,\tilde{\mathbf{e}}_V]\in \mathbb{R}^{V\times d_e}$ . Here, an adaptive fusion layer is introduced to fuse the information from multiple views into a shared view feature $s_v$ . The fusion process is formulated as follows: + +$$ +\mathbf {s} _ {v} = \sum_ {v = 1} ^ {V} \frac {e ^ {\theta_ {v} ^ {\epsilon}} \tilde {\mathbf {e}} _ {v}}{\sum_ {j} e ^ {\theta_ {j} ^ {\epsilon}}} \tag {2} +$$ + +where $\tilde{\mathbf{e}}_v\in \mathbb{R}^{d_e}$ is the $v$ -th row of $\tilde{\mathbf{E}}$ $\theta_v$ denotes the learnable weight and $\epsilon$ is a adjustment factor. By interacting with the shared view feature $\mathbf{s}_v$ and discerning correlations among original categories, $T r a n s f o r m e r_{s c}$ serves to complement information across original categories and get category vectors $\tilde{\mathbf{C}}$ , which can be illustrated as follows: + +$$ +[ \tilde {\mathbf {s}} _ {v}, \tilde {\mathbf {C}} ] = T r a n s f o r m e r _ {s c} ([ \mathbf {s} _ {v}, \mathbf {C} ]) \tag {3} +$$ + +where $[\cdot, \cdot]$ denotes the concatenation operation and $\mathbf{C} \in \mathbb{R}^{L \times d_e}$ denotes original category vectors which are randomly initialized. Subsequently, the output feature of $Transformer_{sc}$ , $\tilde{\mathbf{s}}_v$ , is propagated into the second layer. + +The second layer is designed to extract advanced shared features, which is achieved by promoting the interaction and fusion between view and category information based on the features extracted from the first layer, thereby obtaining a more discriminative multi-view data representation. This layer incorporates two transformer blocks, denoted as $\text{Transformer}_{av}$ and $\text{Transformer}_{ac}$ . $\text{Transformer}_{av}$ is employed to complement information across views and extract advanced multi-view embedding $\bar{\mathbf{E}}$ , which is the enhanced representation of views, by interacting with the shared category feature $\mathbf{s}_c$ and discerning view correlations, as represented below: + +$$ +\mathbf {s} _ {c} = \operatorname {P r o j} _ {c v} (\tilde {\mathbf {s}} _ {v}) \tag {4} +$$ + +$$ +[ \bar {\mathbf {s}} _ {c}, \bar {\mathbf {E}} ] = T r a n s f o r m e r _ {a v} ([ \mathbf {s} _ {c}, \tilde {\mathbf {E}} ]) \tag {5} +$$ + +![](images/65e8fe268821bf0f354bf6ea3ad07c8eaf196ca5d3e799eaae48106d254d06ba.jpg) +Figure 1. An overview of our method VIST. It comprises three modules, namely View-Category Interactive Sharing Transformer Module, Missing View Generation Module, and View-Category Consistency Guided Embedding Enhancement Module. Different colors represent + +different samples, while different shapes signify different features. where $Proj_{cv}(\cdot)$ is a linear layer as a projection function designed to map vectors from the category feature space to the view feature space, $\bar{\mathbf{E}} = [\bar{\mathbf{e}}_1,\bar{\mathbf{e}}_2,\dots ,\bar{\mathbf{e}}_V]\in \mathbb{R}^{V\times d_e}$ . Correspondingly, $Transformer_{ac}$ is employed to complement information across categories and extract advanced category vectors $\bar{\mathbf{C}}$ by leveraging the advanced shared feature $\mathbf{p}_v$ and discerning category correlations, as illustrated below: + +$$ +\mathbf {p} _ {v} = \operatorname {P r o j} _ {v c} (\overline {{\mathbf {s}}} _ {c}) \tag {6} +$$ + +$$ +[ \bar {\mathbf {p}} _ {v}, \bar {\mathbf {C}} ] = T r a n s f o r m e r _ {a c} ([ \mathbf {p} _ {v}, \bar {\mathbf {C}} ]) \tag {7} +$$ + +where $Proj_{vc}(\cdot)$ is a linear layer as a projection function to map vectors from the view feature space to the category feature space. Through the propagation of vectors $\mathbf{s}_v$ , $\mathbf{s}_c$ , and $\mathbf{p}_v$ among the transformer blocks, we facilitate the sharing of information between the view and category feature spaces, thereby extracting more refined and effective features of views and categories. + +# 3.3. Missing View Generation + +Considering the absence of views in the dataset, an intuitive consideration is that the proper completion of these missing views can enhance the model's performance. Through the contrastive learning process mentioned in Sec. 3.4, the distance between views of a sample tends to be relatively proximate. Thus, generating missing views utilizing the existing ones is a plausible approach. We adopt a method analogous to $k$ -nearest neighbors, leveraging original category vectors $\mathbf{C} = [\mathbf{c}_1, \mathbf{c}_2, \dots, \mathbf{c}_L]$ to assist in generating the missing views. Specifically, for a certain sample $i$ , let $\mathcal{E} = \{v | \mathbf{W}_{i,v} = 1\}$ denote the index of existing views, and $\mathcal{M} = \{v | \mathbf{W}_{i,v} = 0\}$ denote the index of missing views. To + +complete the missing features of sample $i$ with its original multi-view embedding $\mathbf{E}$ , we first find the $k$ nearest neighbors in the projected category feature space. The set of the neighbors $\mathcal{D}$ is constructed as follows: + +$$ +\mathcal {D} = \left\{d \mid T o p K \left(\sum_ {j \in \mathcal {E}} \| \mathbf {e} _ {j} - \mathbf {c} _ {d} \| _ {2}, d \in \{1, 2, 3, \dots , L \}\right) \right\} \tag {8} +$$ + +where $TopK(\cdot)$ is a function designed to identify the indices of the top $k$ categories, based on the smallest distance between embedding vectors and category vectors. Then, we employ a statistical method to describe the distribution of the missing views. We hypothesize that these missing views $\{\mathbf{e}_m\}_{m\in \mathcal{M}}$ adhere to a multivariate Gaussian distribution $\mathcal{N}(\mu ,\Sigma)$ , with a mean vector and covariance matrix denoted as: + +$$ +\mu = \frac {\sum_ {d \in \mathcal {D}} \mathbf {c} _ {d}}{| \mathcal {D} |} \tag {9} +$$ + +$$ +\boldsymbol {\Sigma} = \frac {1}{| \mathcal {D} | - 1} \sum_ {d \in \mathcal {D}} \left(\mathbf {c} _ {d} - \mu\right) \left(\mathbf {c} _ {d} - \mu\right) ^ {T} \tag {10} +$$ + +For the missing views, we sample from this distribution $|\mathcal{M}|$ times and substitute the missing views with the sampled results. Consequently, we can obtain the complete embeddings for the incomplete multi-view data. By reconstructing the absent multi-view data, our method further enhances its performance in category predictions. + +# 3.4. View-Category Consistency Guided Embedding Enhancement + +Through the aforementioned view-category interactive sharing transformer, we extract three multi-view embeddings $\mathbf{E}$ , $\tilde{\mathbf{E}}$ and $\bar{\mathbf{E}}$ from different layers. In order to obtain + +better classification performance, enhancing the discriminating power and effectiveness of these embeddings is of great importance. Specifically, predicated on the view consistency assumption [10, 24], the embedding $\mathbf{E}$ of a sample on different views should be aligned. Moreover, to improve the discriminating power of $\tilde{\mathbf{E}}$ and $\bar{\mathbf{E}}$ , the consistency between category correlations and feature correlations of multi-view data can be leveraged. In light of these factors, we introduce view-category consistency guided embedding enhancement. + +To learn more effective embedding $\mathbf{E}$ , we use contrastive learning to align the embeddings of the same sample across different views. According to [13, 17], the loss function of contrastive learning can be formulated as follows: + +$$ +\begin{array}{l} - \log \frac {e ^ {\operatorname {s i m} (x , x ^ {+})}}{e ^ {\operatorname {s i m} (x , x ^ {+})} + \sum_ {x ^ {-} \in \mathcal {F} (x)} e ^ {\operatorname {s i m} (x , x ^ {-})}} \tag {11} \\ \approx - s i m (x, x ^ {+}) + \log \sum_ {x ^ {-} \in \mathcal {F} (x)} \\ \end{array} +$$ + +where $x, x^{+}, x^{-}$ denotes the anchor sample, its positive sample and negative sample. $\mathcal{F}(x)$ denotes the set of negative pairs. $sim(\cdot, \cdot)$ calculates the similarity between two samples. + +Evidently, our objective is to maximize the similarity $sim(x,x^{+})$ while simultaneously minimizing the similarity $\sum_{x^{-}\in \mathcal{F}(x)}e^{sim(x,x^{-})}$ . We denote $\mathbf{e}_{i,v}$ as the embedding of the $v$ -th view for the $i$ -th sample, then $\mathcal{F}(x)$ can be rewritten as $\mathcal{F}(i,v) = \{j,w|\sum^L\mathbf{y}_i\circ \mathbf{u}_i\circ \mathbf{y}_j\circ \mathbf{u}_j = 0,\mathbf{U}_{j,w} = 1\}$ , where $\circ$ denotes Hadamard product, $\mathbf{u}_i$ denotes the missing vector of the $i$ -th sample from $U$ . As for maximizing $e^{sim(x,x^{+})}$ , we use the following loss function: + +$$ +\mathcal {L} _ {p o s} = \sum_ {v _ {1} = 1} ^ {V} \sum_ {v _ {2} = 1, v _ {2} \neq v _ {1}} ^ {V} \frac {1}{d _ {e}} \| n \left(\mathbf {e} _ {i, v _ {1}}\right) - n \left(\mathbf {e} _ {i, v _ {2}}\right) \| _ {2} ^ {2} \tag {12} +$$ + +where $n(\cdot)$ denotes $L_{2}$ normalization function. To minimize $\sum_{x^{-}\in \mathcal{F}(x)}e^{sim(x,x^{-})}$ , the following loss function can be adopted: + +$$ +\mathcal {L} _ {n e g} = \sum_ {v _ {1} = 1} ^ {V} \sum_ {j, v _ {2} \in \mathcal {F} (i, v _ {1})} \frac {- 1}{d _ {e}} \| n \left(\mathbf {e} _ {i, v _ {1}}\right) - n \left(\mathbf {e} _ {j, v _ {2}}\right) \| _ {2} ^ {2} \tag {13} +$$ + +To further improve the effectiveness of $\tilde{\mathbf{E}}$ and $\bar{\mathbf{E}}$ , we utilize the category information to guide the learning process of these embeddings. We enhance the consistency of data correlations in the view feature space and the category feature space as detailed below: + +$$ +\mathbf {G} _ {v} ^ {1} (i, j) = \left(n \left(\tilde {\mathbf {e}} _ {i, v}\right) \left(n \left(\tilde {\mathbf {e}} _ {j, v}\right)\right) ^ {T} + 1\right) / 2 \tag {14} +$$ + +$$ +\mathbf {G} _ {v} ^ {2} (i, j) = \left(n \left(\bar {\mathbf {e}} _ {i, v}\right) \left(n \left(\bar {\mathbf {e}} _ {j, v}\right)\right) ^ {T} + 1\right) / 2 \tag {15} +$$ + +$$ +\mathbf {H} = \left(\left(\mathbf {Y} \circ \mathbf {U}\right) \left(\mathbf {Y} \circ \mathbf {U}\right) ^ {T}\right). / \left(\mathbf {U U} ^ {T}\right) \tag {16} +$$ + +where $\mathbf{G}_v^1$ , $\mathbf{G}_v^2$ and $\mathbf{H}$ are the correlation matrices of $\tilde{\mathbf{E}}$ , $\bar{\mathbf{E}}$ and labels respectively. $\mathbf{e}_{i,v}$ and $\bar{\mathbf{e}}_{i,v}$ represent the feature of the $v$ -th view of the $i$ -th sample from $\tilde{\mathbf{E}}$ and $\bar{\mathbf{E}}$ respectively. Through the cross-entropy loss function, we can align the data correlations across view feature space and category feature space so that more discriminative embedding can be obtained: + +$$ +\begin{array}{l} \mathcal {L} _ {a u x} = \frac {- 1}{2 n V} \sum_ {v = 1} ^ {V} \sum_ {i = 1} ^ {n} \sum_ {j \neq i} ^ {n} \left(\mathbf {H} _ {i, j} \left(\log \mathbf {G} _ {v} ^ {1} (i, j) + \log \mathbf {G} _ {v} ^ {2} (i, j)\right) \right. \\ + \left(\mathbf {U} _ {i, j} - \mathbf {H} _ {i, j}\right) \left(\log \left(1 - \mathbf {G} _ {v} ^ {1} (i, j)\right) + \log \left(1 - \mathbf {G} _ {v} ^ {2} (i, j)\right)\right)) \tag {17} \\ \end{array} +$$ + +Considering the above factors, the loss function for view-category consistency guided embedding enhancement can be summarized as: + +$$ +\mathcal {L} _ {c} = \mathcal {L} _ {\text {p o s}} + \mathcal {L} _ {\text {n e g}} + \alpha \mathcal {L} _ {\text {a u x}} \tag {18} +$$ + +where $\alpha$ is a weight parameter. Through optimizing the above loss function, the learned data embeddings become more discriminative, laying a solid foundation for subsequent multi-label classification. + +# 3.5. Overall Loss Function + +To achieve multi-label classification results, we partition the output $[\bar{\mathbf{p}}_v,\bar{\mathbf{C}} ]$ of $Transforer_{ac}$ into $\bar{\mathbf{p}}_v$ and $\bar{\mathbf{C}}$ . Specifically, for sample $i$ , $\bar{\mathbf{p}}_v^i$ passes through a linear layer to yield $\mathbf{p}_f^i\in [0,1]^L$ , representing the final complementary information between views and categories. $\bar{\mathbf{C}}^i$ is processed through a set of category-aware linear layers, resulting in $\mathbf{z}^i\in [0,1]^L$ , which signifies category-aware predictions. With the assistance of $\mathbf{p}_f^i$ , we get the final prediction $\hat{\mathbf{y}}^i$ : + +$$ +\hat {\mathbf {y}} ^ {i} = \operatorname {c l a s s i f i e r} \left(\mathbf {p} _ {f} ^ {i} + \mathbf {z} ^ {i}\right) \tag {19} +$$ + +where classifier $(\cdot)$ is a linear classifier. We utilize the masked cross-entropy function to compute the multi-label classification loss: + +$$ +\mathcal {L} _ {m} = \frac {\sum_ {i = 1} ^ {n} \sum_ {l = 1} ^ {L} \left(\mathbf {Y} _ {i , l} \log \hat {\mathbf {y}} _ {l} ^ {i} + \left(1 - \mathbf {Y} _ {i , l}\right) \log \left(1 - \hat {\mathbf {y}} _ {l} ^ {i}\right)\right) \mathbf {U} _ {i , l}}{- \sum_ {i , l} \mathbf {U} _ {i , l}} \tag{20} +$$ + +It is important to note that class imbalance is an inherent characteristic of multi-label data, impacting the generalization performance of multi-label prediction [15, 28]. Typically, for certain class labels, the instances of positive label assignments in training data are significantly fewer than those of negative assignments. To address this class-imbalance issue, we introduce a masked asymmetric loss function: + +$$ +\mathbf {L} _ {l} ^ {i} = \left(\max \left(\hat {\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\right)\right) ^ {2} \log \left(1 - \max \left(\hat {\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\right)\right) \tag {21} +$$ + +$$ +\mathcal {L} _ {a} = \frac {\sum_ {i = 1} ^ {n} \sum_ {l = 1} ^ {L} \left(\mathbf {Y} _ {i , l} \log \hat {\mathbf {y}} _ {l} ^ {i} + \left(1 - \mathbf {Y} _ {i , l}\right) \mathbf {L} _ {l} ^ {i}\right) \mathbf {U} _ {i , l}}{- \sum_ {i , l} \mathbf {U} _ {i , l}} \tag{22} +$$ + +Ultimately, we derive the overall loss function of the model as follows: + +$$ +\mathcal {L} _ {o} = \mathcal {L} _ {m} + \beta \mathcal {L} _ {a} + \gamma \mathcal {L} _ {c} \tag {23} +$$ + +where $\beta, \gamma$ denotes different weight parameters. By optimizing $\mathcal{L}_o$ , our method proficiently learns the embeddings for incomplete multi-view data and consequently produces accurate multi-label classification results. + +# 4. Experiments + +# 4.1. Experimental Setting + +# 4.1.1 Datasets + +We follow [9, 11, 14] and evaluate the performance of the proposed VIST on five datasets. These datasets encompass: + +- Corel 5k [1]: The Corel5k dataset is a standard dataset frequently utilized in image annotation and retrieval research, which has 5000 images with 260 types of tags. +- Pascal07 [2]: Pascal07, also known as PASCAL VOC 2007, is a renowned dataset for visual object classes, instrumental in object recognition and classification. It consists of 9963 images and 20 types of tags. +- Espgame [19]: The Espgame dataset is derived from an online game where participants generate labels for images, which contains 20770 samples and 268 types of tags. +- IAPRTC12 [5]: IAPRTC12 is an international image annotation dataset, comprising approximately 20,000 images and 290 categories. These images cover a wide range of subjects and scenes, each accompanied by detailed descriptive text. +- Mirflickr [7]: The Mirflickr dataset consists of 25,000 images and 38 types of tags sourced from Flickr. Selected from the public domain, these images include a variety of themes and scenes. + +We adopt the multi-view feature extraction method outlined in [6]. The images in these datasets undergo a comprehensive preprocessing procedure that transforms them into six distinct views. These views are GIST, HSV, HUE, LAB, RGB, and SIFT. Each of these views offers a unique perspective on the image data, capturing different aspects of visual information. + +# 4.1.2 Implementation Details + +For datasets, in line with [14], we process the five datasets to emulate an incomplete scenario as described below: For every view, we randomly exclude $50\%$ of the samples, ensuring that each sample retains at least one view. Within + +each category, we arbitrarily designate $50\%$ of both positive and negative tags as missing labels. For parameter settings, we choose $d_{e} = 512$ , $\alpha = 0.1$ , $\beta = 0.1$ , $\gamma = 0.1$ and $k = 20$ for all the experiments. The model are trained for 500 epochs with a batch-size of 128 via SGD optimizer on an NVIDIA RTX A6000 GPU. To ensure the credibility of the results, we repeat the experiments 10 times and record the mean and variance of the outcomes. + +# 4.1.3 Compared Methods + +We choose ten representative methods for comparison. Out of these, eight methods - lrMMC, MVL-IV, iMSF, iMvWL, NAIML, DDINet, DICNet and LMVCAT - are detailed in Sec. 2. Additionally, we incorporate GLOCAL [31], to broaden the assessment spectrum. Because iMvWL and NAIML cater to datasets with missing views and labels, we have to modify other methods for consistency. Drawing inspiration from [14] and [9], we input average values from accessible views into lrMMC for the non-operational ones. For MVL-IV and iMSF, absent tags are treated as negative. GLOCAL are executed individually for each view, and we highlight their optimal outcomes. We adhere to the recommended settings for these comparative techniques, as mentioned in their respective publications or code repositories, to maintain an impartial evaluation. + +# 4.1.4 Evaluation Metrics + +Similar to [11], we employ Average Precision (AP), Ranking Loss (RL) and Area Under Curve (AUC). For a convenient comparison, we adopt 1-RL for evaluation. A larger value indicates superior model performance. + +# 4.2. Experimental Results + +In Tab. 1, we show the experimental results of all the methods on five datasets with $70\%$ training samples, $50\%$ missing views and missing labels. Based on the results from Tab. 1, our method exhibits a distinct advantage across most metrics. For instance, on the AP metric for dataset Corel5k, our method surpasses the second-best method, LMVCAT, by $3\%$ . Besides, the performance of DDINet, DICNet and LMVCAT surpasses that of the first six methods. This can be attributed to their applicability specifically to the DIMVMLC issue. Additionally, they are both deep learning models, their superior outcomes underscore the vast potential deep learning holds for addressing this particular challenge. + +While our method primarily addresses the DIMVMLC issue, we also evaluate the model's performance in the full-view and full-label context. The results are presented in Tab. 2. It can be observed that the model still exhibits superior performance in this scenario, suggesting that our method can be effectively applied in MVMLC. + +Table 1. Experimental results of different methods on the five datasets with $70\%$ training samples, $50\%$ missing instances and missing labels. The best results are marked in bold. For simplicity, percent sign is omitted in the following section. + +
DatasetMetriclrMMCMVL-IViMSFGLOCALiMvWLNAIMLDDINetDICNetLMVCATVIST
Corel5kAP24.0±0.224.0±0.118.9±0.228.5±0.428.3±0.730.9±0.436.4±0.138.1±0.438.4±0.441.5±0.2
1-RL76.2±0.275.6±0.170.9±0.580.4±0.386.5±0.387.8±0.287.1±0.088.2±0.488.0±0.290.2±0.2
AUC76.3±0.276.2±0.166.3±0.584.3±0.386.8±0.388.1±0.287.5±0.188.4±0.488.3±0.290.6±0.1
Pascal07AP42.5±0.343.3±0.232.5±0.049.6±0.444.1±1.748.8±0.353.6±0.250.5±1.251.9±0.553.9±0.3
1-RL69.8±0.370.2±0.156.8±0.076.7±0.473.7±0.978.3±0.180.7±0.178.3±0.881.1±0.485.6±0.1
AUC72.8±0.273.0±0.168.6±0.562.0±0.178.6±0.376.7±1.282.7±0.087.6±0.283.4±0.488.3±0.2
EspgameAP18.8±0.018.9±0.010.8±0.022.1±0.224.2±0.324.6±0.228.3±0.129.7±0.229.4±0.430.7±0.3
1-RL77.7±0.177.8±0.072.2±0.278.0±0.480.7±0.181.8±0.281.5±0.083.2±0.182.8±0.284.4±0.0
AUC78.3±0.178.4±0.167.4±0.378.4±0.481.3±0.282.4±0.282.0±0.183.2±0.182.8±0.285.0±0.1
IAPRTC12AP19.7±0.019.8±0.010.1±0.025.6±0.223.5±0.426.1±0.130.3±0.232.3±0.131.7±0.333.9±0.2
1-RL80.1±0.079.9±0.163.1±0.082.5±0.283.3±0.384.8±0.185.3±0.187.3±0.187.0±0.188.4±0.2
AUC80.5±0.080.4±0.166.5±0.183.0±0.183.6±0.285.0±0.185.4±0.087.4±0.187.2±0.188.6±0.1
MirflickrAP44.1±0.144.9±0.132.3±0.053.7±0.249.5±1.255.1±0.259.8±0.258.9±0.559.4±0.560.4±0.3
1-RL80.5±0.080.4±0.166.5±0.183.2±0.183.6±0.285.0±0.186.3±0.086.3±0.486.5±0.387.9±0.1
AUC80.6±0.180.7±0.076.1±0.182.8±0.179.4±1.583.7±0.185.2±0.184.9±0.485.3±0.386.9±0.1
+ +Table 2. Experimental results of different methods on the five datasets with $70\%$ training samples, full views and labels. The best results are marked in bold. + +
DatasetMetricNAIMLDICNetLMVCATVIST
Corel5kAP32.750.952.155.4
1-RL89.092.992.893.5
AUC89.393.193.094.2
Pascal07AP49.660.862.965.5
1-RL79.585.987.889.4
AUC82.287.689.291.2
EspgameAP25.139.138.540.1
1-RL82.587.187.687.7
AUC83.087.488.089.6
IAPRTC12AP26.741.843.645.7
1-RL82.591.291.892.7
AUC83.091.191.892.0
MirflickrAP55.565.968.470.4
1-RL84.789.590.591.3
AUC83.987.688.989.3
+ +Fig. 2 illustrates the performance of the model under various missing ratios. It is interesting that while both incomplete views and incomplete labels detrimentally affect the model's performance, the latter has a more pronounced impact. Moreover, despite our efforts in view imputation, the model's accuracy noticeably diminishes as the view missing ratio escalates. This is primarily due to our reliance on existing views for imputing the missing views. As the number of missing views increases, the imputation algorithm tends to degrade towards random generation. + +Fig. 3 depicts the performance of the model under various training sample ratios. While the ratio's impact on the model is not markedly significant, a general trend is observed: as the ratio increases, there is a corresponding improvement in the model's performance. Thus, selecting a $70\%$ training sample ratio for our experiments is a proper choice. + +![](images/743b3a995735b83a9b070b7405cdca4e85d2949459ef9d0e39c2b499110dc572.jpg) +Figure 2. The results on the Corel5k dataset are presented with (a) different missing-view ratios accompanied by a $50\%$ missing-label ratio and (b) a consistent $50\%$ missing-view ratio paired with different missing-label ratios. + +![](images/87dda4ae5840060ddb0ef295a426347545b21999827977a57a5a9c34990328ea.jpg) + +![](images/a59077620e17ccf34ce506b6f34efefd411a9e2d9503a8e9813d8bebf3bc7fd2.jpg) +Figure 3. The results of different training sample ratios on (a) Corel5k dataset and (b) Espgame dataset with $50\%$ missing-view ratio and $50\%$ missing-label ratio. + +![](images/9fa6b0455a34a6ed75e589f067ad832f65441ce4fed584165a32b8e9c52c10a8.jpg) + +# 4.3. Parameter Analysis + +We evaluate four hyperparameters in our model, i.e., $\alpha$ , $\beta$ , $\gamma$ and $k$ . To assess the influence of these hyperparameters on model performance, we adjust their values within their respective ranges and test their combinations based on the AP metric on different datasets with $50\%$ missing views, $50\%$ missing labels and $70\%$ training sample ratio. Note that $k$ is inherently related to the number of labels, thus, the actual value of $\hat{k}$ utilized in computation should be determined by $\hat{k}_j = \frac{k_i}{L_i} L_j$ , where $i$ and $j$ represents the $i$ -th + +![](images/bae6e17a7a8302f0d96fd850bf4d75ebd85658d09174966ff9ba53d608416fdf.jpg) +(a) Corel5k + +![](images/6a508277c886370beb211dd7665003fd294ab114ca79916706d87d9e72873339.jpg) +(b) Pascal07 + +![](images/aeee90621fe8e357a42eebe6af10e753a24da45a8dc3cda77240745deea76700.jpg) +(c) Espgame +Figure 4. The results of different value combinations of hyperparameter $\beta$ and $\gamma$ on different datasets. $\alpha$ and $k$ are respectively set to default values of 0.1 and 20. + +![](images/fafa396ed5683bac6c1aad87237864870c668cce7831ecd6d2f8f8385e9a46a8.jpg) +(d) IAPRTC12 + +and $j$ -th dataset. Based on Fig. 4, $\beta$ and $\gamma$ exhibit comparatively favorable performances within the ranges of [5e-2, 5e-1] and [5e-2, 1e0] respectively on dataset Corel5k. This indicates that the values of $\beta$ and $\gamma$ should neither be excessively large nor too small, and the optimal value of $\beta$ approximated around 0.1, the optimal value of $\gamma$ should also be selected in the vicinity of 0.1. + +As for $\alpha$ and $k$ , we assess the impact of varying $\alpha$ and $k$ on the model under fixed conditions of $\beta = 0.1$ and $\gamma = 0.1$ by measuring metrics AP and AUC. As illustrated in Fig. 5, on datasets Corel5k and Espgame, both the AP and AUC metrics remain within a narrow range, indicating that the model is not highly sensitive to $\alpha$ and $k$ . This might be because they operate on the deep-layer mechanism, variations within certain ranges do not significantly impact the model's performance. + +# 4.4. Ablation Study + +In the ablation study, we evaluate the performance of each module of VIST on three representative datasets, the results are shown in Tab. 3. We can observe that view-category interactive sharing transformer module plays a pivotal role in enhancing the model's performance. Furthermore, the view-category consistency guided embedding enhancement module is able to improve the discriminating power of embeddings and boost the classification performance. The missing view generation module also contributes to the model's overall efficacy. Additionally, upon the removal of $\mathcal{L}_a$ , there is a noticeable decline in the model's metrics, particularly in the AUC. This suggests that the masked asymmetric loss indeed addresses issues associated with imbal + +![](images/65dbfc9f2a5aed684e65c1b14af1a4133500f3463ff4c9e7423c87bc1479dbf0.jpg) +(a) $\alpha$ on Corel5k + +![](images/d6b4f0a6c98659c14df6879be6b55c10622d7e49989558d4f689c2577a4dbcf1.jpg) +(b) $\alpha$ on Espgame + +![](images/3c7ff54188740fd592c41db2a017fd3146d3e28d92a51bf79a6d3f1353e0f042.jpg) +(c) $k$ on Corel5k +Figure 5. The results of different values of hyperparameter $\alpha$ and $k$ on different datasets. $\beta$ and $\gamma$ are respectively set to default values of 0.1 and 0.1. + +![](images/0cffde94fa81c2904f5d2ad7934c4bfb7b0833a01218b46e87971e31a2f0edc0.jpg) +(d) $k$ on Espgame + +anced data. The ablation study demonstrates the effectiveness and reasonableness of each proposed module. + +Table 3. Results of ablation experiments on the three representative datasets. Vanilla Transformer denotes the encoders of four vanilla transformers, $\mathbf{M}_1$ represents base view-category interactive sharing transformer module, $\mathbf{M}_2$ signifies the view-category consistency guided embedding enhancement module and $\mathbf{M}_3$ denotes missing view generation module. + +
MethodCorel5kEspgameIAPRTC12
APAUCAPAUCAPAUC
Vanilla Transformer34.987.827.683.428.187.2
M139.589.629.484.131.687.8
M1 + M240.490.129.684.732.788.3
M1 + M2 + M341.590.230.785.033.988.6
w/o Lα40.689.829.684.532.488.2
+ +# 5. Conclusion + +In this paper, we propose a novel view-category interactive sharing transformer VIST for addressing the DIMVMLC challenge. Our method leverages a two-layer view-category transformer architecture to extract deep representations of both views and labels. Additionally, we employ a KNN-style approach with the multivariate Gaussian distribution to complete the missing views. To further enhance the efficacy of our model, a view-category consistency guided embedding enhancement module is integrated, significantly boosting the embedding quality and classification accuracy. Our comprehensive experimental analysis convincingly demonstrates that our method outperforms existing state-of-the-art approaches. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China (No.62272058, No.62192784, No.62172056), and Beijing Natural Science Foundation (No.4242027, No.4232025). + +# References + +[1] P. Duygulu, K. Barnard, J. F. G. de Freitas, and D. A. Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision — ECCV 2002, pages 97–112, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 6 +[2] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[3] Zheng Fang and Zhongfei Zhang. Simultaneously combining multi-view multi-label learning with maximum margin classification. In 2012 IEEE 12th International Conference on Data Mining, pages 864-869. IEEE, 2012. 1 +[4] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 1 +[5] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc12 benchmark: A new evaluation resource for visual information systems. Workshop Ontoimage, 2006. 6 +[6] Matthieu Guillaumin, Jakob Verbeek, and Cordelia Schmid. Multimodal semi-supervised learning for image classification. In 2010 IEEE Computer society conference on computer vision and pattern recognition, pages 902-909. IEEE, 2010. 6 +[7] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6 +[8] Junlong Li, Peipei Li, Xuegang Hu, and Kui Yu. Learning common and label-specific features for multi-label classification with correlation information. Pattern Recognition, 121:108259, 2022. 1 +[9] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2022. 1, 2, 6 +[10] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8807-8815, 2023. 1, 2, 5 +[11] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8816–8824, 2023. 1, 2, 6 +[12] Meng Liu, Yong Luo, Dacheng Tao, Chao Xu, and Yonggang Wen. Low-rank multi-view learning in matrix completion for multi-label image classification. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2 +[13] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893, 2018. 5 + +[14] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 2703–2709. International Joint Conferences on Artificial Intelligence Organization, 2018. 2, 6 +[15] Adane Nega Tarekegn, Mario Giacobini, and Krzysztof Michalak. A review of methods for imbalanced multi-label classification. Pattern Recognition, 118:107965, 2021. 5 +[16] Daniel J Trosten, Sigurd Lokse, Robert Jenssen, and Michael Kampffmeyer. Reconsidering representation alignment for multi-view clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1255-1265, 2021. 1 +[17] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5 +[18] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2017. 3 +[19] Luis von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 319-326, New York, NY, USA, 2004. Association for Computing Machinery. 6 +[20] Jie Wen, Zheng Zhang, Lunke Fei, Bob Zhang, Yong Xu, Zhao Zhang, and Jinxing Li. A survey on incomplete multiview clustering. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 53(2):1136-1149, 2022. 1 +[21] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[22] Mengyao Xie, Zongbo Han, Changqing Zhang, Yichen Bai, and Qinghua Hu. Exploring and exploiting uncertainty for incomplete multi-view classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19873-19882, 2023. 1 +[23] Chang Xu, Dacheng Tao, and Chao Xu. Multi-view learning with incomplete views. IEEE Transactions on Image Processing, 24(12):5812-5825, 2015. 2 +[24] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16051-16060, 2022. 5 +[25] Nan Xu, Yanqing Guo, Xin Zheng, Qianyu Wang, and Xi-angyang Luo. Partial multi-view subspace clustering. In Proceedings of the 26th ACM International conference on multimedia, pages 1794–1801, 2018. 1 +[26] Lei Yuan, Yalin Wang, Paul M Thompson, Vaibhav A Narayan, Jieping Ye, Alzheimer's Disease Neuroimaging Initiative, et al. Multi-source feature learning for joint analysis of incomplete multiple heterogeneous neuroimaging data. NeuroImage, 61(3):622-632, 2012. 2 + +[27] Changqing Zhang, Ziwei Yu, Qinghua Hu, Pengfei Zhu, Xinwang Liu, and Xiaobo Wang. Latent semantic aware multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2 +[28] Min-Ling Zhang, Yu-Kun Li, Hao Yang, and Xu-Ying Liu. Towards class-imbalance aware multi-label learning. IEEE Transactions on Cybernetics, 52(6):4459-4471, 2022. 5 +[29] Wei Zhang, Ke Zhang, Pan Gu, and Xiangyang Xue. Multiview embedding learning for incompletely labeled data. In Twenty-Third International Joint Conference on Artificial Intelligence, 2013. 2 +[30] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern recognition, 84:126-135, 2018. 1 +[31] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 6 \ No newline at end of file diff --git a/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/images.zip b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..923a662c1daeece2f26d62e9239d52a4537c636f --- /dev/null +++ b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23c588e582c4f0bfbbeec7bd5ff980d4d207f700c4382ccb9ca2d68a20b5b49c +size 665102 diff --git a/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/layout.json b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9bfeb0b20281209a9bdb545ad03845ec7158db1f --- /dev/null +++ b/2024/View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning/layout.json @@ -0,0 +1,10591 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "text", + "content": "View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 161, + 516, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 161, + 516, + 189 + ], + "spans": [ + { + "bbox": [ + 72, + 161, + 516, + 189 + ], + "type": "text", + "content": "Shilong Ou, Zhe Xue*, Yawen Li, Meiyu Liang, Yuanqiang Cai, Junjiang Wu \nBeijing University of Posts and Telecommunications, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 191, + 516, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 191, + 516, + 204 + ], + "spans": [ + { + "bbox": [ + 72, + 191, + 516, + 204 + ], + "type": "text", + "content": "{os1, xuezhe, meiyul210, caiyuanqiang, wujunjiang}@bupt.edu.cn, warmly0716@126.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 256, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 289, + 533 + ], + "type": "text", + "content": "As a problem often encountered in real-world scenarios, multi-view multi-label learning has attracted considerable research attention. However, due to oversights in data collection and uncertainties in manual annotation, real-world data often suffer from incompleteness. Regrettably, most existing multi-view multi-label learning methods sidestep missing views and labels. Furthermore, they often neglect the potential of harnessing complementary information between views and labels, thus constraining their classification capabilities. To address these challenges, we propose a view-category interactive sharing transformer tailored for incomplete multi-view multi-label learning. Within this network, we incorporate a two-layer transformer module to characterize the interplay between views and labels. Additionally, to address view incompleteness, a KNN-style missing view generation module is employed. Finally, we introduce a view-category consistency guided embedding enhancement module to align different views and improve the discriminating power of the embeddings. Collectively, these modules synergistically integrate to classify the incomplete multi-view multi-label data effectively. Extensive experiments substantiate that our approach outperforms the existing state-of-the-art methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 696 + ], + "type": "text", + "content": "The proliferation of data sources, coupled with advancements in data collection techniques, enables the extraction of diverse types of features, extending beyond a single perspective. This leads to an intriguing research question: how to effectively integrate multi-view features and extract meaningful features from them. Simultaneously, it is observed that single-label data often falls short in encapsulating the complexities of real-world scenarios. For example, an image of a flower may also warrant labels like 'leaf' or 'garden', illustrating that multi-labels provide a richer cate" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 232, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 280 + ], + "type": "text", + "content": "gory feature space and more accurately preserve inter-label relationships [8]. Given its practical relevance, the field of multi-view multi-label learning emerges as a domain of substantial research interest." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 546, + 460 + ], + "type": "text", + "content": "In the field of multi-view learning, multi-view multi-label classification (MVMLC) poses a significant challenge. Considering the consistency across views and the complexity of labels, it is essential for methods to simultaneously align views and extract diverse features from an integrated representation. In recent years, numerous researchers have made strides towards this objective. One such effort is the introduction of a multi-view label embedding model [30], which employs the Hilbert-Schmidt Independence Criterion [4] to establish a link between view feature and category feature spaces. Concurrently, deep neural networks have been employed in this realm. A noteworthy development is the SIMM neural network [3], which utilizes adversarial and label losses to identify shared features while implementing regularization to capture view-specific details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 700 + ], + "type": "text", + "content": "While MVMLC has attracted considerable research attention, it often operates under the presumption that all views and labels are complete. This assumption does not consistently align with real-world scenarios. For instance, a video might lack audio or text content. Multimedia data shared on social media networks might lack comprehensive annotations from users. Such instances culminate in the pervasive challenge of missing multi-view features and incomplete category information within real-world datasets. Some methods aim to address this problem by either masking or restoring the missing views [9, 20, 22, 25]. Although these approaches for handling incomplete multi-view or multi-label learning have achieved notable results, they often fall short in simultaneously addressing both types of incompleteness. To handle the double incomplete multi-view multi-label learning problem (DIMVMLC), several deep contrastive networks are proposed [10, 16]. Furthermore, an incomplete multi-view multi-label learning method based on transformers have been proposed [11], showing adaptability to multi-view and multi-label datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "However, most existing methods for double incomplete" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27467" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "multi-view multi-label classification overlook two critical aspects. Firstly, while some methods recognize the importance of complementary information across multiple views, they often underestimate the complementarity between multi-view features and categories. Efficiently facilitating interaction between views and categories can dramatically enhance the ability to counteract the simultaneous absence of both multi-view feature and category information. Secondly, these methods frequently overlook the issue of missing multi-view features and do not consider completing missing multi-view features. This negligence makes them susceptible to vulnerabilities in insufficient view features, thereby limiting their classification performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 578 + ], + "type": "text", + "content": "To overcome the prevailing challenges in double incomplete multi-view multi-label learning, we propose the View-category Interactive Sharing Transformer (VIST), as shown in Fig. 1. VIST is an architecturally advanced framework specifically designed to exploit the synergy between multiple views and categories and adeptly generate missing views. The dual capability of this method not only amplifies the effectiveness of multi-view embeddings through high-level semantics of categories but also significantly enhances the precision of multi-label classification by utilizing these enriched embeddings. This advancement inherently fortifies the method's resilience, particularly in scenarios characterized by the lack of specific views and labels, thereby ensuring robust and reliable performance. VIST integrates three key components: view-category interactive sharing transformer, missing view generation, and view-category consistency guided embedding enhancement. The view-category interactive sharing transformer, a compact two-layer network with four transformer blocks, fosters an interaction between multi-view and category information, uncovering their complementary potential. The missing view generation, adopting a KNN-style approach, adeptly completes missing views in line with each sample's distribution. The view-category consistency guided embedding enhancement module utilizes contrastive learning strategy to sharpen the discriminative strength of embeddings and boost classification accuracy. Extensive experiments demonstrate the excellent performance of VIST. In summary, this paper makes the following contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": "- We introduce a novel view-category interactive sharing transformer for incomplete multi-view multi-label learning. Our method can effectively exploit the complementary information between views and categories through the interaction between multi-view feature and category information, thereby effectively counteracting the challenges posed by incomplete views and labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "- We propose a missing view generation method specifically designed for the incomplete multi-view multi-label learning task. This method combines the KNN strategy with a multivariate Gaussian distribution, ensuring a sta" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "type": "text", + "content": "tistically sound imputation of missing views that are cohesively aligned with the corresponding samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 96, + 547, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 191 + ], + "type": "text", + "content": "- To bolster the discriminating power of the multi-view embeddings, we develop a view-category consistency guided embedding enhancement module. It adopts contrastive learning to align embeddings across different views and leverages category information to guide embedding learning process, significantly boosting the effectiveness of the resulting embeddings and the performance of multi-label classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 222, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 437 + ], + "type": "text", + "content": "Many methods have been proposed for the MVMLC task. IrMMC [12] ensures the common subspace remains low rank, making it compatible with matrix completion, while also learning combination weights to tap into the distinct strengths of each view. Matrix factorization is utilized to draw out complementary data across different views [27]. Despite their strengths, these models struggles with handling datasets that have missing views or labels. To solve this, MVL-IV [23] delves into incomplete multi-view learning by leveraging the relationships between views. The iMSF model is another incomplete multi-view single-label learning approach. It ingeniously splits the incomplete multi-view classification responsibilities into several complete subtasks [26]. However, both MVL-IV and iMSF have a common shortcoming: they only address view incompleteness. Conversely, MvEL targets the extraction of context and neighborhood consistency but is tailored only for the incomplete multi-label scenario [29]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 714 + ], + "type": "text", + "content": "On the other hand, for double incomplete multi-view multi-label learning task, iMvWL [14] and NAIML [9], have ventured into addressing the incompleteness in both views and labels. iMvWL aligns multi-view features and multi-label data into a shared subspace, enriched by a correlation matrix that bolsters the projection from label space to the embedding subspace. NAIML addresses this dual incompleteness by leaning on the low-rank assumption of the sub-label matrix, subtly harnessing sub-class correlations. In recent years, deep learning models have been employed to address this issue. DDINet [21], DICNet [10] and LMVCAT [11] are deep learning models for DIMVMLC. DDINet is the first deep learning model which contains a view-specific decoder network, successfully preserving the key information of raw views. DICNet is the first deep learning model for DIMVMLC which employs contrastive learning at the instance-level, guiding model to extract deep cross-view features. LMVCAT is a transformer-based double incomplete multi-view multi-label learning framework, consisting of two transformers specifically tailored for views and categories, and constructs a graph on incomplete labels to guide the encoder in extracting view-specific features." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27468" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 159, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 159, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 159, + 103 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " multi-view data points from " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " views, they can be represented as a set of matrices " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_v\\}_{v=1}^V" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v \\in \\mathbb{R}^{n \\times d_v}" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " denotes the dimension of the " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "-th view. The label matrix for all samples is denoted by " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\{0,1\\}^{n \\times L}" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " is the number of categories. If " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{i,j} = 1" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ", it indicates that the " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "-th sample belongs to category " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "; if " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{i,j} = 0" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ", it indicates non-membership in this category. To denote multi-view data with missing features, we introduce the missing-view indicator matrix " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\{0,1\\}^{n \\times V}" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{i,j} = 1" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " means that the " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "-th view of the " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "-th sample is available, whereas " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{i,j} = 0" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " indicates unavailable. Similarly, we define the missing-label indicator matrix " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\{0,1\\}^{n \\times L}" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{i,j} = 1" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " indicates that the " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "-th category of the " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": "-th sample is available, and " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{i,j} = 0" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " signifies the opposite. During the data pre-processing phase, all missing information for view " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " and label " + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 47, + 110, + 289, + 361 + ], + "type": "text", + "content": " is assigned the value 0. The objective of our double incomplete multi-view multi-label learning task is to train a model capable of accurately predicting multiple categories for each sample. In subsequent sections, the detailed mechanisms of the proposed VIST are presented." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 287, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 394 + ], + "type": "text", + "content": "3.2. View-Category Interactive Sharing Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 544 + ], + "type": "text", + "content": "Different views harbor complementary information, necessitating their interaction for effective utilization. Furthermore, category information embodies the data's high-level semantics, serving to augment the discriminative capability of multi-view representations. Inversely, multi-view features can remedy inaccuracies in label information comprehension due to incomplete labels. To this end, we introduce a view-category interactive sharing transformer. This novel transformer is designed to enhance information exchange between views and categories, consequently boosting the discriminative strength of multi-view representations and refining the precision of multi-label classification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "The view-category interactive sharing transformer comprises two key layers. The first layer focuses on extracting shared features that captures correlations among multiple views and categories. An essential preliminary step involves embedding the multi-view data " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v\\in \\mathbb{R}^{n\\times d_v}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". This embedding process ensures the homogenization of feature dimensions across views. For a given sample " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_v\\in \\mathbb{R}^{d_v}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", the embedding vector " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_v\\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " can be denoted as " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_v = Embedding(\\mathbf{x}_v)" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "d_{e}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " denotes the dimension of embedding space, and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "Embedding(\\cdot)" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " is a fully connected neural network. Then we stack the embedding vectors to obtain an original multi-view embedding sequence " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = [\\mathbf{e}_1,\\mathbf{e}_2,\\dots ,\\mathbf{e}_V]\\in \\mathbb{R}^{V\\times d_e}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", which is further used as input vectors for the transformer. Noted that for the incom" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "plete multi-view data, we adopt missing view generation method (as detailed in Sec. 3.3) to produce embeddings of missing views, ensuring that " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " is complete across all views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": "The structure of our transformer is the encoder module of the classical transformer [18]. A transformer block is denoted by " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\text{Transformer}(\\Phi)" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": ", and the input of the transformer is " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{t \\times d_t}" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": " denote the number of tokens and the dimension of tokens. The first layer consists of two transformer blocks, denoted as " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\text{Transformer}_{sv}" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\text{Transformer}_{sc}" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": ". By discerning correlations among original views, " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\text{Transformer}_{sv}" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": " serves to complement information across original views and get multi-view embedding " + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 304, + 108, + 546, + 239 + ], + "type": "text", + "content": ", which is the enhanced representation of original views. It can be illustrated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 370, + 247, + 545, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 247, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 370, + 247, + 545, + 262 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {E}} = \\text {T r a n s f o r m e r} _ {s v} (\\mathbf {E}) \\tag {1}", + "image_path": "12d12510f2a63bee6bc51e7524be518262f77249a792448686fee3237760e4e8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}} = [\\tilde{\\mathbf{e}}_1,\\tilde{\\mathbf{e}}_2,\\dots ,\\tilde{\\mathbf{e}}_V]\\in \\mathbb{R}^{V\\times d_e}" + }, + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "text", + "content": ". Here, an adaptive fusion layer is introduced to fuse the information from multiple views into a shared view feature " + }, + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "text", + "content": ". The fusion process is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 388, + 323, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 323, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 388, + 323, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {v} = \\sum_ {v = 1} ^ {V} \\frac {e ^ {\\theta_ {v} ^ {\\epsilon}} \\tilde {\\mathbf {e}} _ {v}}{\\sum_ {j} e ^ {\\theta_ {j} ^ {\\epsilon}}} \\tag {2}", + "image_path": "a20c0ae8a77b18e15e3cec07e434d874567dafcb73e416c49be89b07cb8dad0b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{e}}_v\\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " -th row of " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " denotes the learnable weight and " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " is a adjustment factor. By interacting with the shared view feature " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_v" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " and discerning correlations among original categories, " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "T r a n s f o r m e r_{s c}" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " serves to complement information across original categories and get category vectors " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{C}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " , which can be illustrated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 445, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 445, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 351, + 445, + 545, + 460 + ], + "type": "interline_equation", + "content": "[ \\tilde {\\mathbf {s}} _ {v}, \\tilde {\\mathbf {C}} ] = T r a n s f o r m e r _ {s c} ([ \\mathbf {s} _ {v}, \\mathbf {C} ]) \\tag {3}", + "image_path": "d1cff6475035ad4c9639a4f071f31f2dadb503a4ac81bc4a9bd7a57deaf03fa4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "text", + "content": " denotes the concatenation operation and " + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{L \\times d_e}" + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "text", + "content": " denotes original category vectors which are randomly initialized. Subsequently, the output feature of " + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "inline_equation", + "content": "Transformer_{sc}" + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{s}}_v" + }, + { + "bbox": [ + 304, + 468, + 545, + 516 + ], + "type": "text", + "content": ", is propagated into the second layer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "content": "The second layer is designed to extract advanced shared features, which is achieved by promoting the interaction and fusion between view and category information based on the features extracted from the first layer, thereby obtaining a more discriminative multi-view data representation. This layer incorporates two transformer blocks, denoted as " + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "inline_equation", + "content": "\\text{Transformer}_{av}" + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "inline_equation", + "content": "\\text{Transformer}_{ac}" + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "inline_equation", + "content": "\\text{Transformer}_{av}" + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "content": " is employed to complement information across views and extract advanced multi-view embedding " + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}}" + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "content": ", which is the enhanced representation of views, by interacting with the shared category feature " + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c" + }, + { + "bbox": [ + 304, + 516, + 546, + 659 + ], + "type": "text", + "content": " and discerning view correlations, as represented below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 668, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 668, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 389, + 668, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {c} = \\operatorname {P r o j} _ {c v} (\\tilde {\\mathbf {s}} _ {v}) \\tag {4}", + "image_path": "44bc788cfd6ac06d201997bf050c46ca58a49e640033df33b4a2ab441cc17f3f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 352, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 352, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "[ \\bar {\\mathbf {s}} _ {c}, \\bar {\\mathbf {E}} ] = T r a n s f o r m e r _ {a v} ([ \\mathbf {s} _ {c}, \\tilde {\\mathbf {E}} ]) \\tag {5}", + "image_path": "327f32ab9fa9758e597ad5c7ed544ed9dd438a6926ce54984086ee2dcae9de6b.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27469" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 74, + 496, + 280 + ], + "blocks": [ + { + "bbox": [ + 97, + 74, + 496, + 280 + ], + "lines": [ + { + "bbox": [ + 97, + 74, + 496, + 280 + ], + "spans": [ + { + "bbox": [ + 97, + 74, + 496, + 280 + ], + "type": "image", + "image_path": "65e8fe268821bf0f354bf6ea3ad07c8eaf196ca5d3e799eaae48106d254d06ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 547, + 311 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 547, + 311 + ], + "type": "text", + "content": "Figure 1. An overview of our method VIST. It comprises three modules, namely View-Category Interactive Sharing Transformer Module, Missing View Generation Module, and View-Category Consistency Guided Embedding Enhancement Module. Different colors represent" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "content": "different samples, while different shapes signify different features. where " + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "inline_equation", + "content": "Proj_{cv}(\\cdot)" + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "content": " is a linear layer as a projection function designed to map vectors from the category feature space to the view feature space, " + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}} = [\\bar{\\mathbf{e}}_1,\\bar{\\mathbf{e}}_2,\\dots ,\\bar{\\mathbf{e}}_V]\\in \\mathbb{R}^{V\\times d_e}" + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "content": ". Correspondingly, " + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "inline_equation", + "content": "Transformer_{ac}" + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "content": " is employed to complement information across categories and extract advanced category vectors " + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{C}}" + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "content": " by leveraging the advanced shared feature " + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_v" + }, + { + "bbox": [ + 46, + 312, + 288, + 414 + ], + "type": "text", + "content": " and discerning category correlations, as illustrated below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 415, + 287, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 415, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 129, + 415, + 287, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {v} = \\operatorname {P r o j} _ {v c} (\\overline {{\\mathbf {s}}} _ {c}) \\tag {6}", + "image_path": "600489ac9122b22e1fce4fab8aaa8e351bdf934d94bf83cd3fe81b0a0a22175e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 443, + 287, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 443, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 91, + 443, + 287, + 457 + ], + "type": "interline_equation", + "content": "[ \\bar {\\mathbf {p}} _ {v}, \\bar {\\mathbf {C}} ] = T r a n s f o r m e r _ {a c} ([ \\mathbf {p} _ {v}, \\bar {\\mathbf {C}} ]) \\tag {7}", + "image_path": "3f5af892884764fd6f4d8540faa6fa85df2c5dbdeac45a420ef338b5ef3f85d6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "inline_equation", + "content": "Proj_{vc}(\\cdot)" + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "text", + "content": " is a linear layer as a projection function to map vectors from the view feature space to the category feature space. Through the propagation of vectors " + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_v" + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c" + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_v" + }, + { + "bbox": [ + 46, + 461, + 287, + 545 + ], + "type": "text", + "content": " among the transformer blocks, we facilitate the sharing of information between the view and category feature spaces, thereby extracting more refined and effective features of views and categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 188, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 188, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 188, + 563 + ], + "type": "text", + "content": "3.3. Missing View Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "Considering the absence of views in the dataset, an intuitive consideration is that the proper completion of these missing views can enhance the model's performance. Through the contrastive learning process mentioned in Sec. 3.4, the distance between views of a sample tends to be relatively proximate. Thus, generating missing views utilizing the existing ones is a plausible approach. We adopt a method analogous to " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "-nearest neighbors, leveraging original category vectors " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = [\\mathbf{c}_1, \\mathbf{c}_2, \\dots, \\mathbf{c}_L]" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " to assist in generating the missing views. Specifically, for a certain sample " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{v | \\mathbf{W}_{i,v} = 1\\}" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " denote the index of existing views, and " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{v | \\mathbf{W}_{i,v} = 0\\}" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " denote the index of missing views. To" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "text", + "content": "complete the missing features of sample " + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "text", + "content": " with its original multi-view embedding " + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "text", + "content": ", we first find the " + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "text", + "content": " nearest neighbors in the projected category feature space. The set of the neighbors " + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 320, + 545, + 368 + ], + "type": "text", + "content": " is constructed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 375, + 545, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 375, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 312, + 375, + 545, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = \\left\\{d \\mid T o p K \\left(\\sum_ {j \\in \\mathcal {E}} \\| \\mathbf {e} _ {j} - \\mathbf {c} _ {d} \\| _ {2}, d \\in \\{1, 2, 3, \\dots , L \\}\\right) \\right\\} \\tag {8}", + "image_path": "14dda254a8cc9b04480674404c0c2d097d96e0ead3b7c55859f3adb5474ae53f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "inline_equation", + "content": "TopK(\\cdot)" + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "text", + "content": " is a function designed to identify the indices of the top " + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "text", + "content": " categories, based on the smallest distance between embedding vectors and category vectors. Then, we employ a statistical method to describe the distribution of the missing views. We hypothesize that these missing views " + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_m\\}_{m\\in \\mathcal{M}}" + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "text", + "content": " adhere to a multivariate Gaussian distribution " + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu ,\\Sigma)" + }, + { + "bbox": [ + 304, + 411, + 545, + 505 + ], + "type": "text", + "content": ", with a mean vector and covariance matrix denoted as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 393, + 504, + 545, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 504, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 393, + 504, + 545, + 530 + ], + "type": "interline_equation", + "content": "\\mu = \\frac {\\sum_ {d \\in \\mathcal {D}} \\mathbf {c} _ {d}}{| \\mathcal {D} |} \\tag {9}", + "image_path": "657fe7face6574b3f062e35eb5429560072d7a9addd2ac5259ffa53cacd41e01.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 536, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 536, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 348, + 536, + 545, + 564 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\frac {1}{| \\mathcal {D} | - 1} \\sum_ {d \\in \\mathcal {D}} \\left(\\mathbf {c} _ {d} - \\mu\\right) \\left(\\mathbf {c} _ {d} - \\mu\\right) ^ {T} \\tag {10}", + "image_path": "e60c481f5045ceee93346bbc16a96518f80b5835005ff6c37d7452a9062925bf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": "For the missing views, we sample from this distribution " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "|\\mathcal{M}|" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " times and substitute the missing views with the sampled results. Consequently, we can obtain the complete embeddings for the incomplete multi-view data. By reconstructing the absent multi-view data, our method further enhances its performance in category predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 647, + 545, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 545, + 672 + ], + "type": "text", + "content": "3.4. View-Category Consistency Guided Embedding Enhancement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Through the aforementioned view-category interactive sharing transformer, we extract three multi-view embeddings " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " from different layers. In order to obtain" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27470" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "text", + "content": "better classification performance, enhancing the discriminating power and effectiveness of these embeddings is of great importance. Specifically, predicated on the view consistency assumption [10, 24], the embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "text", + "content": " of a sample on different views should be aligned. Moreover, to improve the discriminating power of " + }, + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 190 + ], + "type": "text", + "content": ", the consistency between category correlations and feature correlations of multi-view data can be leveraged. In light of these factors, we introduce view-category consistency guided embedding enhancement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "content": "To learn more effective embedding " + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "content": ", we use contrastive learning to align the embeddings of the same sample across different views. According to [13, 17], the loss function of contrastive learning can be formulated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 247, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 247, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 81, + 247, + 287, + 306 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} - \\log \\frac {e ^ {\\operatorname {s i m} (x , x ^ {+})}}{e ^ {\\operatorname {s i m} (x , x ^ {+})} + \\sum_ {x ^ {-} \\in \\mathcal {F} (x)} e ^ {\\operatorname {s i m} (x , x ^ {-})}} \\tag {11} \\\\ \\approx - s i m (x, x ^ {+}) + \\log \\sum_ {x ^ {-} \\in \\mathcal {F} (x)} \\\\ \\end{array}", + "image_path": "da4ee007c6a67e20e05051e1bb7487b18db05ceb560dd687cc42a8361d176e02.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "inline_equation", + "content": "x, x^{+}, x^{-}" + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "content": " denotes the anchor sample, its positive sample and negative sample. " + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)" + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "content": " denotes the set of negative pairs. " + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "inline_equation", + "content": "sim(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "content": " calculates the similarity between two samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": "Evidently, our objective is to maximize the similarity " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "sim(x,x^{+})" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": " while simultaneously minimizing the similarity " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\sum_{x^{-}\\in \\mathcal{F}(x)}e^{sim(x,x^{-})}" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{i,v}" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": " as the embedding of the " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": "-th view for the " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": "-th sample, then " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": " can be rewritten as " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(i,v) = \\{j,w|\\sum^L\\mathbf{y}_i\\circ \\mathbf{u}_i\\circ \\mathbf{y}_j\\circ \\mathbf{u}_j = 0,\\mathbf{U}_{j,w} = 1\\}" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": " denotes Hadamard product, " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": " denotes the missing vector of the " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": "-th sample from " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": ". As for maximizing " + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "inline_equation", + "content": "e^{sim(x,x^{+})}" + }, + { + "bbox": [ + 47, + 363, + 289, + 464 + ], + "type": "text", + "content": ", we use the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 470, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 470, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 470, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p o s} = \\sum_ {v _ {1} = 1} ^ {V} \\sum_ {v _ {2} = 1, v _ {2} \\neq v _ {1}} ^ {V} \\frac {1}{d _ {e}} \\| n \\left(\\mathbf {e} _ {i, v _ {1}}\\right) - n \\left(\\mathbf {e} _ {i, v _ {2}}\\right) \\| _ {2} ^ {2} \\tag {12}", + "image_path": "717cbe3b17b8504f6b4114a02b7f485f36b845da015f3dc82231817984a27ca6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "inline_equation", + "content": "n(\\cdot)" + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "text", + "content": " normalization function. To minimize " + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\sum_{x^{-}\\in \\mathcal{F}(x)}e^{sim(x,x^{-})}" + }, + { + "bbox": [ + 47, + 512, + 287, + 550 + ], + "type": "text", + "content": ", the following loss function can be adopted:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 556, + 287, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 287, + 593 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n e g} = \\sum_ {v _ {1} = 1} ^ {V} \\sum_ {j, v _ {2} \\in \\mathcal {F} (i, v _ {1})} \\frac {- 1}{d _ {e}} \\| n \\left(\\mathbf {e} _ {i, v _ {1}}\\right) - n \\left(\\mathbf {e} _ {j, v _ {2}}\\right) \\| _ {2} ^ {2} \\tag {13}", + "image_path": "e862deb044f3bd3ad3f1af41ea68851731a91cb2b0d98fe4a201637c3ee305e1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "content": "To further improve the effectiveness of " + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}}" + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "content": ", we utilize the category information to guide the learning process of these embeddings. We enhance the consistency of data correlations in the view feature space and the category feature space as detailed below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 667, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 667, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 88, + 667, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {v} ^ {1} (i, j) = \\left(n \\left(\\tilde {\\mathbf {e}} _ {i, v}\\right) \\left(n \\left(\\tilde {\\mathbf {e}} _ {j, v}\\right)\\right) ^ {T} + 1\\right) / 2 \\tag {14}", + "image_path": "da2602d4799f724fca1d87024224c9587d5982f02a646c8a3e24a364cffe7eca.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 88, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {v} ^ {2} (i, j) = \\left(n \\left(\\bar {\\mathbf {e}} _ {i, v}\\right) \\left(n \\left(\\bar {\\mathbf {e}} _ {j, v}\\right)\\right) ^ {T} + 1\\right) / 2 \\tag {15}", + "image_path": "cf69ea851768350a3012d38cfc97cff8a87b3ccc1aaa3a9f561eb41e0a1b3540.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 83, + 545, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 83, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 351, + 83, + 545, + 97 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\left(\\left(\\mathbf {Y} \\circ \\mathbf {U}\\right) \\left(\\mathbf {Y} \\circ \\mathbf {U}\\right) ^ {T}\\right). / \\left(\\mathbf {U U} ^ {T}\\right) \\tag {16}", + "image_path": "c446915f78b21e6f96ffe2e1b7176f1165f7c0bb5a9faeb0b68c08a10729b497.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_v^1" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_v^2" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " are the correlation matrices of " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " and labels respectively. " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{i,v}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{e}}_{i,v}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " represent the feature of the " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": "-th view of the " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": "-th sample from " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{E}}" + }, + { + "bbox": [ + 305, + 100, + 547, + 186 + ], + "type": "text", + "content": " respectively. Through the cross-entropy loss function, we can align the data correlations across view feature space and category feature space so that more discriminative embedding can be obtained:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 192, + 552, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 552, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 552, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a u x} = \\frac {- 1}{2 n V} \\sum_ {v = 1} ^ {V} \\sum_ {i = 1} ^ {n} \\sum_ {j \\neq i} ^ {n} \\left(\\mathbf {H} _ {i, j} \\left(\\log \\mathbf {G} _ {v} ^ {1} (i, j) + \\log \\mathbf {G} _ {v} ^ {2} (i, j)\\right) \\right. \\\\ + \\left(\\mathbf {U} _ {i, j} - \\mathbf {H} _ {i, j}\\right) \\left(\\log \\left(1 - \\mathbf {G} _ {v} ^ {1} (i, j)\\right) + \\log \\left(1 - \\mathbf {G} _ {v} ^ {2} (i, j)\\right)\\right)) \\tag {17} \\\\ \\end{array}", + "image_path": "5935d6130d777135ebf701a7a0a7dc9f587aa9fac91352e773d335236bca7156.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 253, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 253, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 545, + 287 + ], + "type": "text", + "content": "Considering the above factors, the loss function for view-category consistency guided embedding enhancement can be summarized as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 366, + 297, + 545, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 297, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 366, + 297, + 545, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\mathcal {L} _ {\\text {p o s}} + \\mathcal {L} _ {\\text {n e g}} + \\alpha \\mathcal {L} _ {\\text {a u x}} \\tag {18}", + "image_path": "30275609b17b3087a042f46f89002eb160da404a42587e327c9fd297c82ac00a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 317, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 317, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 545, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 317, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 317, + 545, + 365 + ], + "type": "text", + "content": " is a weight parameter. Through optimizing the above loss function, the learned data embeddings become more discriminative, laying a solid foundation for subsequent multi-label classification." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 371, + 432, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 432, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 432, + 384 + ], + "type": "text", + "content": "3.5. Overall Loss Function" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": "To achieve multi-label classification results, we partition the output " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "[\\bar{\\mathbf{p}}_v,\\bar{\\mathbf{C}} ]" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "Transforer_{ac}" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}_v" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{C}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": ". Specifically, for sample " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}_v^i" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": " passes through a linear layer to yield " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_f^i\\in [0,1]^L" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": ", representing the final complementary information between views and categories. " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{C}}^i" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": " is processed through a set of category-aware linear layers, resulting in " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^i\\in [0,1]^L" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": ", which signifies category-aware predictions. With the assistance of " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_f^i" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": ", we get the final prediction " + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^i" + }, + { + "bbox": [ + 305, + 389, + 545, + 489 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 370, + 498, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 498, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 370, + 498, + 545, + 513 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} ^ {i} = \\operatorname {c l a s s i f i e r} \\left(\\mathbf {p} _ {f} ^ {i} + \\mathbf {z} ^ {i}\\right) \\tag {19}", + "image_path": "16a31e38dabf0dbcfeafcb484274479be9d3649bedc3b61068433a564ea0b7e3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 516, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 516, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 516, + 545, + 552 + ], + "type": "text", + "content": "where classifier " + }, + { + "bbox": [ + 305, + 516, + 545, + 552 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 305, + 516, + 545, + 552 + ], + "type": "text", + "content": " is a linear classifier. We utilize the masked cross-entropy function to compute the multi-label classification loss:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 557, + 558, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 558, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 558, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\frac {\\sum_ {i = 1} ^ {n} \\sum_ {l = 1} ^ {L} \\left(\\mathbf {Y} _ {i , l} \\log \\hat {\\mathbf {y}} _ {l} ^ {i} + \\left(1 - \\mathbf {Y} _ {i , l}\\right) \\log \\left(1 - \\hat {\\mathbf {y}} _ {l} ^ {i}\\right)\\right) \\mathbf {U} _ {i , l}}{- \\sum_ {i , l} \\mathbf {U} _ {i , l}} \\tag{20}", + "image_path": "5a571d4f4a57c882a136727a418041a0c6e89b35f6611b1564fadbef91db0a93.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 597, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 692 + ], + "type": "text", + "content": "It is important to note that class imbalance is an inherent characteristic of multi-label data, impacting the generalization performance of multi-label prediction [15, 28]. Typically, for certain class labels, the instances of positive label assignments in training data are significantly fewer than those of negative assignments. To address this class-imbalance issue, we introduce a masked asymmetric loss function:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 310, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {l} ^ {i} = \\left(\\max \\left(\\hat {\\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\\right)\\right) ^ {2} \\log \\left(1 - \\max \\left(\\hat {\\mathbf {y}} _ {l} ^ {i} - 0. 5, 0\\right)\\right) \\tag {21}", + "image_path": "939906c40a0298ba86d059c3a0347d39536137d86fe6ce504f69c49612e2fd23.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27471" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a} = \\frac {\\sum_ {i = 1} ^ {n} \\sum_ {l = 1} ^ {L} \\left(\\mathbf {Y} _ {i , l} \\log \\hat {\\mathbf {y}} _ {l} ^ {i} + \\left(1 - \\mathbf {Y} _ {i , l}\\right) \\mathbf {L} _ {l} ^ {i}\\right) \\mathbf {U} _ {i , l}}{- \\sum_ {i , l} \\mathbf {U} _ {i , l}} \\tag{22}", + "image_path": "6bea9671482b4334f0b748426946513c903a845992aa90c942aff4accc3a7b3c.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 287, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 287, + 151 + ], + "type": "text", + "content": "Ultimately, we derive the overall loss function of the model as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 152, + 287, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 152, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 287, + 165 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\mathcal {L} _ {m} + \\beta \\mathcal {L} _ {a} + \\gamma \\mathcal {L} _ {c} \\tag {23}", + "image_path": "255f942115eed63d34e1ec7cd8fd4755e353630f9ee1ecde646e32ee12f2b2a2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma" + }, + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "type": "text", + "content": " denotes different weight parameters. By optimizing " + }, + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_o" + }, + { + "bbox": [ + 47, + 171, + 287, + 217 + ], + "type": "text", + "content": ", our method proficiently learns the embeddings for incomplete multi-view data and consequently produces accurate multi-label classification results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 229, + 128, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 128, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 128, + 243 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 248, + 170, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 170, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 170, + 261 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 266, + 117, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 117, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 117, + 277 + ], + "type": "text", + "content": "4.1.1 Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "text", + "content": "We follow [9, 11, 14] and evaluate the performance of the proposed VIST on five datasets. These datasets encompass:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 310, + 287, + 547 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 310, + 287, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 345 + ], + "type": "text", + "content": "- Corel 5k [1]: The Corel5k dataset is a standard dataset frequently utilized in image annotation and retrieval research, which has 5000 images with 260 types of tags." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 346, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 392 + ], + "type": "text", + "content": "- Pascal07 [2]: Pascal07, also known as PASCAL VOC 2007, is a renowned dataset for visual object classes, instrumental in object recognition and classification. It consists of 9963 images and 20 types of tags." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 393, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 287, + 441 + ], + "type": "text", + "content": "- Espgame [19]: The Espgame dataset is derived from an online game where participants generate labels for images, which contains 20770 samples and 268 types of tags." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 441, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 500 + ], + "type": "text", + "content": "- IAPRTC12 [5]: IAPRTC12 is an international image annotation dataset, comprising approximately 20,000 images and 290 categories. These images cover a wide range of subjects and scenes, each accompanied by detailed descriptive text." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 501, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 287, + 547 + ], + "type": "text", + "content": "- Mirflickr [7]: The Mirflickr dataset consists of 25,000 images and 38 types of tags sourced from Flickr. Selected from the public domain, these images include a variety of themes and scenes." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 548, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 632 + ], + "type": "text", + "content": "We adopt the multi-view feature extraction method outlined in [6]. The images in these datasets undergo a comprehensive preprocessing procedure that transforms them into six distinct views. These views are GIST, HSV, HUE, LAB, RGB, and SIFT. Each of these views offers a unique perspective on the image data, capturing different aspects of visual information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 647, + 180, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 180, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 180, + 658 + ], + "type": "text", + "content": "4.1.2 Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "For datasets, in line with [14], we process the five datasets to emulate an incomplete scenario as described below: For every view, we randomly exclude " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " of the samples, ensuring that each sample retains at least one view. Within" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": "each category, we arbitrarily designate " + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": " of both positive and negative tags as missing labels. For parameter settings, we choose " + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "d_{e} = 512" + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 305, + 72, + 545, + 167 + ], + "type": "text", + "content": " for all the experiments. The model are trained for 500 epochs with a batch-size of 128 via SGD optimizer on an NVIDIA RTX A6000 GPU. To ensure the credibility of the results, we repeat the experiments 10 times and record the mean and variance of the outcomes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 184, + 424, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 184, + 424, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 424, + 196 + ], + "type": "text", + "content": "4.1.3 Compared Methods" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 204, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 545, + 384 + ], + "type": "text", + "content": "We choose ten representative methods for comparison. Out of these, eight methods - lrMMC, MVL-IV, iMSF, iMvWL, NAIML, DDINet, DICNet and LMVCAT - are detailed in Sec. 2. Additionally, we incorporate GLOCAL [31], to broaden the assessment spectrum. Because iMvWL and NAIML cater to datasets with missing views and labels, we have to modify other methods for consistency. Drawing inspiration from [14] and [9], we input average values from accessible views into lrMMC for the non-operational ones. For MVL-IV and iMSF, absent tags are treated as negative. GLOCAL are executed individually for each view, and we highlight their optimal outcomes. We adhere to the recommended settings for these comparative techniques, as mentioned in their respective publications or code repositories, to maintain an impartial evaluation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 399, + 421, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 421, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 421, + 411 + ], + "type": "text", + "content": "4.1.4 Evaluation Metrics" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 413, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 461 + ], + "type": "text", + "content": "Similar to [11], we employ Average Precision (AP), Ranking Loss (RL) and Area Under Curve (AUC). For a convenient comparison, we adopt 1-RL for evaluation. A larger value indicates superior model performance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 467, + 429, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 429, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 429, + 479 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "text", + "content": "In Tab. 1, we show the experimental results of all the methods on five datasets with " + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "text", + "content": " training samples, " + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "text", + "content": " missing views and missing labels. Based on the results from Tab. 1, our method exhibits a distinct advantage across most metrics. For instance, on the AP metric for dataset Corel5k, our method surpasses the second-best method, LMVCAT, by " + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 305, + 486, + 545, + 640 + ], + "type": "text", + "content": ". Besides, the performance of DDINet, DICNet and LMVCAT surpasses that of the first six methods. This can be attributed to their applicability specifically to the DIMVMLC issue. Additionally, they are both deep learning models, their superior outcomes underscore the vast potential deep learning holds for addressing this particular challenge." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "content": "While our method primarily addresses the DIMVMLC issue, we also evaluate the model's performance in the full-view and full-label context. The results are presented in Tab. 2. It can be observed that the model still exhibits superior performance in this scenario, suggesting that our method can be effectively applied in MVMLC." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27472" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 96, + 552, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Experimental results of different methods on the five datasets with " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " training samples, " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " missing instances and missing labels. The best results are marked in bold. For simplicity, percent sign is omitted in the following section." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 96, + 552, + 258 + ], + "lines": [ + { + "bbox": [ + 48, + 96, + 552, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 552, + 258 + ], + "type": "table", + "html": "
DatasetMetriclrMMCMVL-IViMSFGLOCALiMvWLNAIMLDDINetDICNetLMVCATVIST
Corel5kAP24.0±0.224.0±0.118.9±0.228.5±0.428.3±0.730.9±0.436.4±0.138.1±0.438.4±0.441.5±0.2
1-RL76.2±0.275.6±0.170.9±0.580.4±0.386.5±0.387.8±0.287.1±0.088.2±0.488.0±0.290.2±0.2
AUC76.3±0.276.2±0.166.3±0.584.3±0.386.8±0.388.1±0.287.5±0.188.4±0.488.3±0.290.6±0.1
Pascal07AP42.5±0.343.3±0.232.5±0.049.6±0.444.1±1.748.8±0.353.6±0.250.5±1.251.9±0.553.9±0.3
1-RL69.8±0.370.2±0.156.8±0.076.7±0.473.7±0.978.3±0.180.7±0.178.3±0.881.1±0.485.6±0.1
AUC72.8±0.273.0±0.168.6±0.562.0±0.178.6±0.376.7±1.282.7±0.087.6±0.283.4±0.488.3±0.2
EspgameAP18.8±0.018.9±0.010.8±0.022.1±0.224.2±0.324.6±0.228.3±0.129.7±0.229.4±0.430.7±0.3
1-RL77.7±0.177.8±0.072.2±0.278.0±0.480.7±0.181.8±0.281.5±0.083.2±0.182.8±0.284.4±0.0
AUC78.3±0.178.4±0.167.4±0.378.4±0.481.3±0.282.4±0.282.0±0.183.2±0.182.8±0.285.0±0.1
IAPRTC12AP19.7±0.019.8±0.010.1±0.025.6±0.223.5±0.426.1±0.130.3±0.232.3±0.131.7±0.333.9±0.2
1-RL80.1±0.079.9±0.163.1±0.082.5±0.283.3±0.384.8±0.185.3±0.187.3±0.187.0±0.188.4±0.2
AUC80.5±0.080.4±0.166.5±0.183.0±0.183.6±0.285.0±0.185.4±0.087.4±0.187.2±0.188.6±0.1
MirflickrAP44.1±0.144.9±0.132.3±0.053.7±0.249.5±1.255.1±0.259.8±0.258.9±0.559.4±0.560.4±0.3
1-RL80.5±0.080.4±0.166.5±0.183.2±0.183.6±0.285.0±0.186.3±0.086.3±0.486.5±0.387.9±0.1
AUC80.6±0.180.7±0.076.1±0.182.8±0.179.4±1.583.7±0.185.2±0.184.9±0.485.3±0.386.9±0.1
", + "image_path": "44e1a92e774908ba32babc90820e4274e37f26b853db67c58ee4f7ccef515661.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 315, + 285, + 476 + ], + "blocks": [ + { + "bbox": [ + 47, + 277, + 287, + 310 + ], + "lines": [ + { + "bbox": [ + 47, + 277, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 287, + 310 + ], + "type": "text", + "content": "Table 2. Experimental results of different methods on the five datasets with " + }, + { + "bbox": [ + 47, + 277, + 287, + 310 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 47, + 277, + 287, + 310 + ], + "type": "text", + "content": " training samples, full views and labels. The best results are marked in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 315, + 285, + 476 + ], + "lines": [ + { + "bbox": [ + 49, + 315, + 285, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 315, + 285, + 476 + ], + "type": "table", + "html": "
DatasetMetricNAIMLDICNetLMVCATVIST
Corel5kAP32.750.952.155.4
1-RL89.092.992.893.5
AUC89.393.193.094.2
Pascal07AP49.660.862.965.5
1-RL79.585.987.889.4
AUC82.287.689.291.2
EspgameAP25.139.138.540.1
1-RL82.587.187.687.7
AUC83.087.488.089.6
IAPRTC12AP26.741.843.645.7
1-RL82.591.291.892.7
AUC83.091.191.892.0
MirflickrAP55.565.968.470.4
1-RL84.789.590.591.3
AUC83.987.688.989.3
", + "image_path": "bcd553afe8d16fb14bc1d2e4975501d0179296bad9b7d0614d924f8801bd5ef0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 511, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 632 + ], + "type": "text", + "content": "Fig. 2 illustrates the performance of the model under various missing ratios. It is interesting that while both incomplete views and incomplete labels detrimentally affect the model's performance, the latter has a more pronounced impact. Moreover, despite our efforts in view imputation, the model's accuracy noticeably diminishes as the view missing ratio escalates. This is primarily due to our reliance on existing views for imputing the missing views. As the number of missing views increases, the imputation algorithm tends to degrade towards random generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 638, + 288, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 638, + 288, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 638, + 288, + 721 + ], + "type": "text", + "content": "Fig. 3 depicts the performance of the model under various training sample ratios. While the ratio's impact on the model is not markedly significant, a general trend is observed: as the ratio increases, there is a corresponding improvement in the model's performance. Thus, selecting a " + }, + { + "bbox": [ + 46, + 638, + 288, + 721 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 638, + 288, + 721 + ], + "type": "text", + "content": " training sample ratio for our experiments is a proper choice." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 279, + 425, + 371 + ], + "blocks": [ + { + "bbox": [ + 320, + 279, + 425, + 371 + ], + "lines": [ + { + "bbox": [ + 320, + 279, + 425, + 371 + ], + "spans": [ + { + "bbox": [ + 320, + 279, + 425, + 371 + ], + "type": "image", + "image_path": "743b3a995735b83a9b070b7405cdca4e85d2949459ef9d0e39c2b499110dc572.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "type": "text", + "content": "Figure 2. The results on the Corel5k dataset are presented with (a) different missing-view ratios accompanied by a " + }, + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "type": "text", + "content": " missing-label ratio and (b) a consistent " + }, + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 380, + 545, + 424 + ], + "type": "text", + "content": " missing-view ratio paired with different missing-label ratios." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 428, + 279, + 533, + 371 + ], + "blocks": [ + { + "bbox": [ + 428, + 279, + 533, + 371 + ], + "lines": [ + { + "bbox": [ + 428, + 279, + 533, + 371 + ], + "spans": [ + { + "bbox": [ + 428, + 279, + 533, + 371 + ], + "type": "image", + "image_path": "87dda4ae5840060ddb0ef295a426347545b21999827977a57a5a9c34990328ea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 441, + 425, + 531 + ], + "blocks": [ + { + "bbox": [ + 319, + 441, + 425, + 531 + ], + "lines": [ + { + "bbox": [ + 319, + 441, + 425, + 531 + ], + "spans": [ + { + "bbox": [ + 319, + 441, + 425, + 531 + ], + "type": "image", + "image_path": "a59077620e17ccf34ce506b6f34efefd411a9e2d9503a8e9813d8bebf3bc7fd2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "lines": [ + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "type": "text", + "content": "Figure 3. The results of different training sample ratios on (a) Corel5k dataset and (b) Espgame dataset with " + }, + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "type": "text", + "content": " missing-view ratio and " + }, + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 542, + 545, + 575 + ], + "type": "text", + "content": " missing-label ratio." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 428, + 441, + 533, + 533 + ], + "blocks": [ + { + "bbox": [ + 428, + 441, + 533, + 533 + ], + "lines": [ + { + "bbox": [ + 428, + 441, + 533, + 533 + ], + "spans": [ + { + "bbox": [ + 428, + 441, + 533, + 533 + ], + "type": "image", + "image_path": "9fa6b0455a34a6ed75e589f067ad832f65441ce4fed584165a32b8e9c52c10a8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 585, + 421, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 421, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 421, + 598 + ], + "type": "text", + "content": "4.3. Parameter Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": "We evaluate four hyperparameters in our model, i.e., " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": ". To assess the influence of these hyperparameters on model performance, we adjust their values within their respective ranges and test their combinations based on the AP metric on different datasets with " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " missing views, " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " missing labels and " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " training sample ratio. Note that " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " is inherently related to the number of labels, thus, the actual value of " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{k}" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " utilized in computation should be determined by " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{k}_j = \\frac{k_i}{L_i} L_j" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 605, + 545, + 715 + ], + "type": "text", + "content": "-th" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27473" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 81, + 161, + 168 + ], + "blocks": [ + { + "bbox": [ + 58, + 81, + 161, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 81, + 161, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 81, + 161, + 168 + ], + "type": "image", + "image_path": "bae6e17a7a8302f0d96fd850bf4d75ebd85658d09174966ff9ba53d608416fdf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 171, + 132, + 179 + ], + "lines": [ + { + "bbox": [ + 94, + 171, + 132, + 179 + ], + "spans": [ + { + "bbox": [ + 94, + 171, + 132, + 179 + ], + "type": "text", + "content": "(a) Corel5k" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 83, + 269, + 168 + ], + "blocks": [ + { + "bbox": [ + 167, + 83, + 269, + 168 + ], + "lines": [ + { + "bbox": [ + 167, + 83, + 269, + 168 + ], + "spans": [ + { + "bbox": [ + 167, + 83, + 269, + 168 + ], + "type": "image", + "image_path": "6a508277c886370beb211dd7665003fd294ab114ca79916706d87d9e72873339.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 171, + 242, + 180 + ], + "lines": [ + { + "bbox": [ + 201, + 171, + 242, + 180 + ], + "spans": [ + { + "bbox": [ + 201, + 171, + 242, + 180 + ], + "type": "text", + "content": "(b) Pascal07" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 191, + 160, + 279 + ], + "blocks": [ + { + "bbox": [ + 58, + 191, + 160, + 279 + ], + "lines": [ + { + "bbox": [ + 58, + 191, + 160, + 279 + ], + "spans": [ + { + "bbox": [ + 58, + 191, + 160, + 279 + ], + "type": "image", + "image_path": "aeee90621fe8e357a42eebe6af10e753a24da45a8dc3cda77240745deea76700.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 281, + 134, + 290 + ], + "lines": [ + { + "bbox": [ + 93, + 281, + 134, + 290 + ], + "spans": [ + { + "bbox": [ + 93, + 281, + 134, + 290 + ], + "type": "text", + "content": "(c) Espgame" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "text", + "content": "Figure 4. The results of different value combinations of hyperparameter " + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "text", + "content": " on different datasets. " + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 300, + 287, + 333 + ], + "type": "text", + "content": " are respectively set to default values of 0.1 and 20." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 167, + 196, + 268, + 279 + ], + "blocks": [ + { + "bbox": [ + 167, + 196, + 268, + 279 + ], + "lines": [ + { + "bbox": [ + 167, + 196, + 268, + 279 + ], + "spans": [ + { + "bbox": [ + 167, + 196, + 268, + 279 + ], + "type": "image", + "image_path": "fafa396ed5683bac6c1aad87237864870c668cce7831ecd6d2f8f8385e9a46a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 281, + 246, + 290 + ], + "lines": [ + { + "bbox": [ + 198, + 281, + 246, + 290 + ], + "spans": [ + { + "bbox": [ + 198, + 281, + 246, + 290 + ], + "type": "text", + "content": "(d) IAPRTC12" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": "-th dataset. Based on Fig. 4, " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": " exhibit comparatively favorable performances within the ranges of [5e-2, 5e-1] and [5e-2, 1e0] respectively on dataset Corel5k. This indicates that the values of " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": " should neither be excessively large nor too small, and the optimal value of " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": " approximated around 0.1, the optimal value of " + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": " should also be selected in the vicinity of 0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": "As for " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": ", we assess the impact of varying " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": " on the model under fixed conditions of " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": " by measuring metrics AP and AUC. As illustrated in Fig. 5, on datasets Corel5k and Espgame, both the AP and AUC metrics remain within a narrow range, indicating that the model is not highly sensitive to " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 424, + 287, + 533 + ], + "type": "text", + "content": ". This might be because they operate on the deep-layer mechanism, variations within certain ranges do not significantly impact the model's performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 540, + 140, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 140, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 140, + 553 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "In the ablation study, we evaluate the performance of each module of VIST on three representative datasets, the results are shown in Tab. 3. We can observe that view-category interactive sharing transformer module plays a pivotal role in enhancing the model's performance. Furthermore, the view-category consistency guided embedding enhancement module is able to improve the discriminating power of embeddings and boost the classification performance. The missing view generation module also contributes to the model's overall efficacy. Additionally, upon the removal of " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", there is a noticeable decline in the model's metrics, particularly in the AUC. This suggests that the masked asymmetric loss indeed addresses issues associated with imbal" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 77, + 417, + 151 + ], + "blocks": [ + { + "bbox": [ + 318, + 77, + 417, + 151 + ], + "lines": [ + { + "bbox": [ + 318, + 77, + 417, + 151 + ], + "spans": [ + { + "bbox": [ + 318, + 77, + 417, + 151 + ], + "type": "image", + "image_path": "65dbfc9f2a5aed684e65c1b14af1a4133500f3463ff4c9e7423c87bc1479dbf0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 153, + 399, + 163 + ], + "lines": [ + { + "bbox": [ + 344, + 153, + 399, + 163 + ], + "spans": [ + { + "bbox": [ + 344, + 153, + 399, + 163 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 344, + 153, + 399, + 163 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 344, + 153, + 399, + 163 + ], + "type": "text", + "content": " on Corel5k" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 427, + 78, + 525, + 151 + ], + "blocks": [ + { + "bbox": [ + 427, + 78, + 525, + 151 + ], + "lines": [ + { + "bbox": [ + 427, + 78, + 525, + 151 + ], + "spans": [ + { + "bbox": [ + 427, + 78, + 525, + 151 + ], + "type": "image", + "image_path": "d6b4f0a6c98659c14df6879be6b55c10622d7e49989558d4f689c2577a4dbcf1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 153, + 510, + 163 + ], + "lines": [ + { + "bbox": [ + 451, + 153, + 510, + 163 + ], + "spans": [ + { + "bbox": [ + 451, + 153, + 510, + 163 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 451, + 153, + 510, + 163 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 451, + 153, + 510, + 163 + ], + "type": "text", + "content": " on Espgame" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 171, + 416, + 243 + ], + "blocks": [ + { + "bbox": [ + 317, + 171, + 416, + 243 + ], + "lines": [ + { + "bbox": [ + 317, + 171, + 416, + 243 + ], + "spans": [ + { + "bbox": [ + 317, + 171, + 416, + 243 + ], + "type": "image", + "image_path": "3c7ff54188740fd592c41db2a017fd3146d3e28d92a51bf79a6d3f1353e0f042.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 247, + 397, + 255 + ], + "lines": [ + { + "bbox": [ + 343, + 247, + 397, + 255 + ], + "spans": [ + { + "bbox": [ + 343, + 247, + 397, + 255 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 343, + 247, + 397, + 255 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 343, + 247, + 397, + 255 + ], + "type": "text", + "content": " on Corel5k" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": "Figure 5. The results of different values of hyperparameter " + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": " on different datasets. " + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": " are respectively set to default values of 0.1 and 0.1." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 426, + 171, + 524, + 243 + ], + "blocks": [ + { + "bbox": [ + 426, + 171, + 524, + 243 + ], + "lines": [ + { + "bbox": [ + 426, + 171, + 524, + 243 + ], + "spans": [ + { + "bbox": [ + 426, + 171, + 524, + 243 + ], + "type": "image", + "image_path": "0cffde94fa81c2904f5d2ad7934c4bfb7b0833a01218b46e87971e31a2f0edc0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 247, + 509, + 256 + ], + "lines": [ + { + "bbox": [ + 450, + 247, + 509, + 256 + ], + "spans": [ + { + "bbox": [ + 450, + 247, + 509, + 256 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 450, + 247, + 509, + 256 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 450, + 247, + 509, + 256 + ], + "type": "text", + "content": " on Espgame" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 300, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 325 + ], + "type": "text", + "content": "anced data. The ablation study demonstrates the effectiveness and reasonableness of each proposed module." + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 313, + 402, + 535, + 480 + ], + "blocks": [ + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "text", + "content": "Table 3. Results of ablation experiments on the three representative datasets. Vanilla Transformer denotes the encoders of four vanilla transformers, " + }, + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_1" + }, + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "text", + "content": " represents base view-category interactive sharing transformer module, " + }, + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_2" + }, + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "text", + "content": " signifies the view-category consistency guided embedding enhancement module and " + }, + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_3" + }, + { + "bbox": [ + 305, + 332, + 545, + 399 + ], + "type": "text", + "content": " denotes missing view generation module." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 402, + 535, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 535, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 535, + 480 + ], + "type": "table", + "html": "
MethodCorel5kEspgameIAPRTC12
APAUCAPAUCAPAUC
Vanilla Transformer34.987.827.683.428.187.2
M139.589.629.484.131.687.8
M1 + M240.490.129.684.732.788.3
M1 + M2 + M341.590.230.785.033.988.6
w/o Lα40.689.829.684.532.488.2
", + "image_path": "233c17f7029dc754f313501a6537ac6401d560c0816d26f27d48ca8549114c72.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 495, + 378, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 378, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 378, + 508 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": "In this paper, we propose a novel view-category interactive sharing transformer VIST for addressing the DIMVMLC challenge. Our method leverages a two-layer view-category transformer architecture to extract deep representations of both views and labels. Additionally, we employ a KNN-style approach with the multivariate Gaussian distribution to complete the missing views. To further enhance the efficacy of our model, a view-category consistency guided embedding enhancement module is integrated, significantly boosting the embedding quality and classification accuracy. Our comprehensive experimental analysis convincingly demonstrates that our method outperforms existing state-of-the-art approaches." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 658, + 391, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 391, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 391, + 669 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 669, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 545, + 712 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (No.62272058, No.62192784, No.62172056), and Beijing Natural Science Foundation (No.4242027, No.4232025)." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27474" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] P. Duygulu, K. Barnard, J. F. G. de Freitas, and D. A. Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision — ECCV 2002, pages 97–112, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 237 + ], + "type": "text", + "content": "[3] Zheng Fang and Zhongfei Zhang. Simultaneously combining multi-view multi-label learning with maximum margin classification. In 2012 IEEE 12th International Conference on Data Mining, pages 864-869. IEEE, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "type": "text", + "content": "[4] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "text", + "content": "[5] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc12 benchmark: A new evaluation resource for visual information systems. Workshop Ontoimage, 2006. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "text", + "content": "[6] Matthieu Guillaumin, Jakob Verbeek, and Cordelia Schmid. Multimodal semi-supervised learning for image classification. In 2010 IEEE Computer society conference on computer vision and pattern recognition, pages 902-909. IEEE, 2010. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 385, + 288, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 288, + 429 + ], + "type": "text", + "content": "[7] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 288, + 474 + ], + "type": "text", + "content": "[8] Junlong Li, Peipei Li, Xuegang Hu, and Kui Yu. Learning common and label-specific features for multi-label classification with correlation information. Pattern Recognition, 121:108259, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "type": "text", + "content": "[9] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2022. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "text", + "content": "[10] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8807-8815, 2023. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "text", + "content": "[11] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8816–8824, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Meng Liu, Yong Luo, Dacheng Tao, Chao Xu, and Yonggang Wen. Low-rank multi-view learning in matrix completion for multi-label image classification. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893, 2018. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[14] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 2703–2709. International Joint Conferences on Artificial Intelligence Organization, 2018. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "text", + "content": "[15] Adane Nega Tarekegn, Mario Giacobini, and Krzysztof Michalak. A review of methods for imbalanced multi-label classification. Pattern Recognition, 118:107965, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 176, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 547, + 230 + ], + "type": "text", + "content": "[16] Daniel J Trosten, Sigurd Lokse, Robert Jenssen, and Michael Kampffmeyer. Reconsidering representation alignment for multi-view clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1255-1265, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 233, + 546, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 546, + 264 + ], + "type": "text", + "content": "[17] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 268, + 546, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 546, + 300 + ], + "type": "text", + "content": "[18] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 303, + 546, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 546, + 358 + ], + "type": "text", + "content": "[19] Luis von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 319-326, New York, NY, USA, 2004. Association for Computing Machinery. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 547, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 547, + 404 + ], + "type": "text", + "content": "[20] Jie Wen, Zheng Zhang, Lunke Fei, Bob Zhang, Yong Xu, Zhao Zhang, and Jinxing Li. A survey on incomplete multiview clustering. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 53(2):1136-1149, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 406, + 547, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 547, + 460 + ], + "type": "text", + "content": "[21] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 463, + 547, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 547, + 518 + ], + "type": "text", + "content": "[22] Mengyao Xie, Zongbo Han, Changqing Zhang, Yichen Bai, and Qinghua Hu. Exploring and exploiting uncertainty for incomplete multi-view classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19873-19882, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 520, + 546, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 546, + 553 + ], + "type": "text", + "content": "[23] Chang Xu, Dacheng Tao, and Chao Xu. Multi-view learning with incomplete views. IEEE Transactions on Image Processing, 24(12):5812-5825, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "type": "text", + "content": "[24] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16051-16060, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 546, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 546, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 546, + 656 + ], + "type": "text", + "content": "[25] Nan Xu, Yanqing Guo, Xin Zheng, Qianyu Wang, and Xi-angyang Luo. Partial multi-view subspace clustering. In Proceedings of the 26th ACM International conference on multimedia, pages 1794–1801, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "text", + "content": "[26] Lei Yuan, Yalin Wang, Paul M Thompson, Vaibhav A Narayan, Jieping Ye, Alzheimer's Disease Neuroimaging Initiative, et al. Multi-source feature learning for joint analysis of incomplete multiple heterogeneous neuroimaging data. NeuroImage, 61(3):622-632, 2012. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27475" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Changqing Zhang, Ziwei Yu, Qinghua Hu, Pengfei Zhu, Xinwang Liu, and Xiaobo Wang. Latent semantic aware multi-view multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[28] Min-Ling Zhang, Yu-Kun Li, Hao Yang, and Xu-Ying Liu. Towards class-imbalance aware multi-label learning. IEEE Transactions on Cybernetics, 52(6):4459-4471, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "text", + "content": "[29] Wei Zhang, Ke Zhang, Pan Gu, and Xiangyang Xue. Multiview embedding learning for incompletely labeled data. In Twenty-Third International Joint Conference on Artificial Intelligence, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 228 + ], + "type": "text", + "content": "[30] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern recognition, 84:126-135, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "text", + "content": "[31] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27476" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_content_list.json b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e30bc09ea549e812fd7a344f754aabd0a54c2628 --- /dev/null +++ b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network", + "text_level": 1, + "bbox": [ + 194, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quan Zhang $^{1,5}$ Lei Wang $^{1}$ Vishal M. Patel $^{5}$ Xiaohua Xie $^{1,2,3,4}$ Jianhuang Lai $^{1,2,3,4*}$ $^{1}$ School of Computer Science and Engineering, Sun Yat-Sen University, China \n $^{2}$ Pazhou Lab (HuangPu), Guangdong, China", + "bbox": [ + 91, + 200, + 883, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China \n4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China \n5Department of Electrical and Computer Engineering, Johns Hopkins University, USA \n{zhangq48, wanglei75}@mail2.sysu.edu.cn, {stsljh, xiexiaoh6}@mail.sysu.edu.cn, vpatel36@jhu.edu", + "bbox": [ + 96, + 256, + 874, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 378, + 313, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing person re-identification methods have achieved remarkable advances in appearance-based identity association across homogeneous cameras, such as ground-ground matching. However, as a more practical scenario, aerial-ground person re-identification (AGPReID) among heterogeneous cameras has received minimal attention. To alleviate the disruption of discriminative identity representation by dramatic view discrepancy as the most significant challenge in AGPReID, the view-decoupled transformer (VDT) is proposed as a simple yet effective framework. Two major components are designed in VDT to decouple view-related and view-unrelated features, namely hierarchical subtractive separation and orthogonal loss, where the former separates these two features inside the VDT, and the latter constrains these two to be independent. In addition, we contribute a large-scale AGPReID dataset called CARGO, consisting of five/eight aerial/ground cameras, 5,000 identities, and 108,563 images. Experiments on two datasets show that VDT is a feasible and effective solution for AGPReID, surpassing the previous method on mAP/Rank1 by up to $5.0\\% /2.7\\%$ on CARGO and $3.7\\% /5.2\\%$ on AG-ReID, keeping the same magnitude of computational complexity. Our project is available at https://github.com/LinlyAC/VDT-AGPReID.", + "bbox": [ + 76, + 412, + 473, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 804, + 209, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) aims to associate interested person images based on cross-camera identity similarity, which plays a significant and positive contribution to the se", + "bbox": [ + 76, + 830, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e500db2cf3afb8a177a92c40844fda3cf765402699f37bfe6096491cb8b918cd.jpg", + "image_caption": [ + "Figure 1. View-homogeneous vs. view-heterogeneous ReID, where the former focuses on ground-only or aerial-only camera networks, and the latter considers the aerial-ground mixed camera network. Thus, view-heterogeneous ReID considers aerial-aerial, ground-ground, and aerial-ground matching, which is more challenging and practical than the existing view-homogeneous ReID." + ], + "image_footnote": [], + "bbox": [ + 522, + 378, + 870, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "curity and safety of society and citizens [1-5]. Traditional ReID has achieved remarkable progress in the deep learning era [6-16], but it is far from real-world scenarios for the following reasons. (1) Most existing ReID datasets are collected from homogeneous camera networks, which comprises of the same type cameras, such as ground-only [17] or aerial-only [18] camera networks. However, as can be seen from Fig. 1, real-world surveillance systems tend to be deployed as heterogeneous camera networks comprising", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 886, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22000", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aerial and ground cameras instead of a single type alone. Ground cameras cover well-developed areas (city centers), and aerial cameras cover poorly-developed areas (suburbs) due to their broad view range. The complementary nature of the two types would maximize the effectiveness of ReID. (2) Existing ReID methods mainly consider homogeneous matching (ground-ground [19-21] and aerial-aerial [9, 22, 23]), which is ineffective in dealing with the dramatic view discrepancy among heterogeneous matching (aerial-ground). Despite its significance, related research is extremely scarce.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we consider the novel and practical view-heterogeneous ReID problem, specifically ReID under the aerial-ground camera network (AGPReID). We propose a view-decoupled transformer (VDT) to specifically tackle the dramatic view discrepancy, which serves as a significant challenge within AGPReID to hinder homogeneous and heterogeneous matchings. Motivated by the fact that view-related features are useless for discriminative identity representations, VDT aims to decouple view-related and view-unrelated components, further facilitating discriminative identity learning from the remaining view-unrelated features. There are two key parts within VDT to achieve view decoupling, namely (a) hierarchically subtractive separation and (b) orthogonal loss. Specifically, after tokenizing the input image to a series of patch tokens, VDT appends two extra tokens (the meta token and view token) to feed them into a transformer stacked by multiple VDT blocks. The meta token captures global representation in the image, and the view token aims to extract view-related features. In each VDT block, (a) is achieved by the subtractive values of the meta token and view token after self-attention operations as the updated meta token for the next block. This means that VDT hierarchically separates view-related features from global features to facilitate identity learning from remaining view-unrelated features. When the meta and view tokens are obtained after the last block, (b) is designed to constrain the identity features to be independent of the view features, ultimately achieving the orthogonal decoupling of view-related and view-unrelated components. Meanwhile, the meta and view tokens will be supervised by identity and view labels, respectively.", + "bbox": [ + 75, + 261, + 472, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the scarcity [22] and privacy [5, 24] of the AGPReID datasets, we contribute a large-scale Civic AeRial-GrOund (CARGO) dataset, which is collected from a synthetic city scenario including five aerial and eight ground cameras. CARGO has totally collected 5,000 identities and 108,563 images, which contain not only significant view discrepancy but also plenty of variations of resolution, illumination, occlusion, etc. Regarding evaluation, CARGO considers multiple patterns (aerial-aerial, ground-ground, aerial-ground matchings), which could serve as a comprehensive benchmark. Contributions of this paper can", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "be summarized as follows:", + "bbox": [ + 498, + 90, + 679, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We focus on ReID under the aerial-ground camera network (AGPReID) scenario and propose a view-decoupled transformer (VDT) specifically for the dramatic view discrepancy within AGPReID, which achieves view-related and view-unrelated features decoupling by hierarchical subtractive separation and orthogonal loss.", + "- We contribute a large-scale synthetic dataset, called CARGO, to advance the AGPReID as a benchmark, which contains five/eight aerial/ground cameras, 5,000 identities, 108,563 images, and complex challenges.", + "- Experiments demonstrate the superiority of VDT on two datasets, which shows that VDT surpasses the previous method on mAP/Rank1 by up to $5.0\\% / 2.7\\%$ on CARGO and $3.7\\% / 5.2\\%$ on AG-ReID, respectively, maintaining the same magnitude of computational complexity." + ], + "bbox": [ + 500, + 107, + 893, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 347, + 640, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. View-homogeneous ReID", + "text_level": 1, + "bbox": [ + 500, + 373, + 730, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "View-homogeneous ReID deals with images from ground-only or aerial-only camera networks, where the ground-only camera network receives the broadest attention in the ReID task because it is the most common use case. Many ground-only datasets have been contributed in the literature, such as Market1501 [17], MSMT17 [25], etc. Accordingly, lots of methods have been proposed to bring rapid development to ReID, including handcrafted feature-based methods [26, 27], CNN-based methods [19-21, 28], and transformer-based methods [29]. Regarding aerial-only research, only some pioneering efforts have contributed substantial datasets and methods [18, 30, 31] because it has not received as much attention as ground-only ReID.", + "bbox": [ + 496, + 397, + 890, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is difficult to directly transfer the above methods to the view-heterogeneous ReID because these methods ignore the significant view differences between aerial and ground cameras, leading to weak performances.", + "bbox": [ + 496, + 593, + 890, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. View-heterogeneous ReID", + "text_level": 1, + "bbox": [ + 500, + 664, + 736, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to the abundant research in Sec. 2.1, studies on view-heterogeneous ReID are almost nonexistent. In this paper, heterogeneous views refer to AGPReID, and the fundamental challenge is the dramatic view discrepancy. To the best of our knowledge, one recent work [22] has formally attempted to address this task. They collected an outdoor scene dataset, AG-ReID, with both identity and attribute labeling of pedestrians. Moreover, they proposed an explainable model by attribute information to guide model training.", + "bbox": [ + 496, + 688, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While pioneering and critical to AGPReID, this work is still imperfect for the following reasons. (1) Dataset. Compared to [22], our CARGO dataset comprehensively focuses on multiple matching patterns that may appear in heterogeneous views rather than a single aerial-ground matching in", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22001", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4fe5321429469995b02bae8316f7a9290adee4613e4f0ab4a79e6a757f85c5a2.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed VDT framework, which consists of $N$ VDT blocks and three-part loss functions. Meta and view tokens capture global and view-related features in images, respectively. Each VDT block (light blue module) consists of a standard self-attention encoder layer and an inner feature subtraction operation, achieving layer-by-layer decoupling of view-related and view-unrelated features. Orthogonal loss constrains the above two features to be further independent." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 872, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AG-ReID. (2) Method. Compared to explainable methods in [22], our VDT relies on a lesser amount of priori labeling and thus has a stronger generalization. The experiments also show that our VDT achieves stronger performances.", + "bbox": [ + 75, + 438, + 470, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Synthetic ReID Dataset", + "text_level": 1, + "bbox": [ + 76, + 508, + 294, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data privacy is an inevitable issue for the ReID datasets, and synthetic data provides a feasible solution because it not only achieves privacy protection for real-world pedestrians but also has lower collection costs. Many synthetic ReID datasets have been published in the literature, such as PersonX [14], UnrealPerson [24], WePerson [2], etc.", + "bbox": [ + 75, + 532, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, existing synthetic datasets mainly focus on simulating pedestrian data under view-homogeneous camera networks, and the simulation of aerial-ground views is poorly studied. To this end, we contribute the CARGO dataset as the first large-scale synthetic dataset for AGPReID.", + "bbox": [ + 75, + 625, + 470, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 729, + 168, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulation and Overview", + "text_level": 1, + "bbox": [ + 76, + 756, + 321, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formulation. An AGPReLU dataset $\\mathcal{D} = \\{(x_i, y_i, v_i)\\}_{i=1}^{|\\mathcal{D}|}$ consists of a training set $\\mathcal{D}^{tr}$ and a test set $\\mathcal{D}^{te}$ , where $x_i$ represents the $i$ -th person image, and $y_i$ and $v_i$ stand for the corresponding identity and view labels, respectively. Note that $v_i \\in \\{v^a, v^g\\}$ is readily acquired by the known camera labels in $\\mathcal{D}$ to easily distinguish whether $x_i$ belongs to an aerial $v^a$ or ground $v^g$ view. Significant view discrepancy between $v^a$ and $v^g$ leads to a view-biased feature", + "bbox": [ + 75, + 777, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space, which behaves as low intra-identity and high interidentity similarity. Therefore, the goal of AGPReID methods is to design a model $\\mathcal{F}(\\cdot;\\theta_{\\mathcal{F}})$ with learnable parameter $\\theta_{\\mathcal{F}}$ against the view bias, which can be written as:", + "bbox": [ + 498, + 438, + 892, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\sum_ {i} \\left[ \\begin{array}{c} \\left\\| \\mathcal {F} \\left(x _ {i}; \\theta_ {\\mathcal {F}}\\right) - \\mathcal {F} \\left(x _ {i} ^ {+}; \\theta_ {\\mathcal {F}}\\right) \\right\\| _ {2} ^ {2} - \\\\ \\left\\| \\mathcal {F} \\left(x _ {i}; \\theta_ {\\mathcal {F}}\\right) - \\mathcal {F} \\left(x _ {i} ^ {-}; \\theta_ {\\mathcal {F}}\\right) \\right\\| _ {2} ^ {2} \\end{array} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 508, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _2$ denotes the L2 distance, $x_{i}^{+} / x_{i}^{-}$ denotes the person image that has the same/different identity with $x_{i}$ .", + "bbox": [ + 496, + 563, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. As shown in Fig. 2, we design a simple yet effective framework called view-decoupled transformer (VDT), to tackle the view discrepancy challenge in AGPReID. For a batched data $B$ that contains both $v^a$ and $v^g$ view, weckenize them as a series of image tokens (green squares), then append a meta and view token (blue and yellow squares) to them, which serve as input to VDT. The VDT network consists of $N$ blocks, where each block first performs the standard self-attention encoding and then performs a subtraction operation between meta and view tokens to explicitly separate the view-related feature from the global feature. Finally, the identity and view classifier will supervise the updated meta and view tokens outputted from VDT. Besides, an orthogonal loss is proposed that makes these two tokens irrelevant, thus achieving the complete decoupling of view-related and view-unrelated features.", + "bbox": [ + 496, + 595, + 892, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. View-decoupled Transformer", + "text_level": 1, + "bbox": [ + 500, + 845, + 761, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed VDT framework is based on Vit-Base [34]. For each image $x_{i}$ in the batched data $B$ , VDT first", + "bbox": [ + 498, + 869, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22002", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4a2f776f38c01bdfc28de98ab086798141ebddac73ad90e487d637d7296d6df3.jpg", + "table_caption": [ + "Table 1. Statistical comparisons of existing datasets, including view-homogeneous (ground or aerial) and view-heterogeneous (ground and aerial) ReID. “ $\\star$ ” denotes City1M is used for group ReID, where its images correspond to the number of group images." + ], + "table_footnote": [], + "table_body": "
DatasetViewData#PersonID#Camera#Image#Height
Market1501 [17]GroundReal1,501632,668< 10m
PersonX [14]GroundSynthetic1,2666273,456-
RandPerson [32]GroundSynthetic8,00019228,655-
UnrealPerson [24]GroundSynthetic3,00034120,000-
WePerson [2]GroundSynthetic1,500404,000,000-
ClonedPerson [33]GroundSynthetic5,62124887,766-
City1M* [5]GroundSynthetic45,00081,840,000-
PRAI1581 [18]AerialReal1,581239,46120 ~ 60m
UAVHuman [30]AerialReal1,144141,2902 ~ 8m
AG-ReID [22]Aerial-GroundReal3882 (1A+1G)21,89315 ~ 45m
CARGO (Ours)Aerial-GroundSynthetic5,00013 (5A+8G)108,5635 ~ 75m
", + "bbox": [ + 158, + 128, + 810, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "divides $x_{i}$ evenly and non-overlappingly into $M$ image patches, which are further tokenized into $M$ patch embeddings by a $1 \\times 1$ convolution and can be denoted as $[t_{p,1}; t_{p,2}; \\dots; t_{p,M}]$ . After that, we append two additional learnable tokens, meta token $t_{m}$ and view token $t_{v}$ , in the patch tokens, where $t_{m}$ aims to capture the global image representation, $t_{v}$ focuses only on the view-related feature. Next, we serialize all the tokens by assigning them the corresponding positional embedding, which can be written as:", + "bbox": [ + 75, + 338, + 472, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (x _ {i}) \\triangleq \\left[ t _ {m}; t _ {p, 1}; t _ {p, 2}; \\dots ; t _ {p, M}; t _ {v} \\right] + \\mathbf {E} _ {p o s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 488, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{S}(\\cdot)$ is defined as the serialization operation for $x_{i}$ , $t_m, t_{p,i}, t_v \\in \\mathbb{R}^d$ , and $\\mathbf{E}_{pos} \\in \\mathbb{R}^{(M + 2) \\times d}$ .", + "bbox": [ + 75, + 512, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, VDT $\\mathcal{F}(\\cdot ;\\theta_{\\mathcal{F}})$ is stacked with $N$ VDT blocks $\\mathcal{F}_j(\\cdot ;\\theta_{\\mathcal{F}}^j),1\\leq j\\leq N.$ Each $\\mathcal{F}_j(\\cdot ;\\theta_{\\mathcal{F}}^j)$ contains two operations, self-attention encoding and subtractive separation, which are described as:", + "bbox": [ + 75, + 542, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right); \\dots ; t _ {v} ^ {(j + 1)} \\left(x _ {i}\\right) \\right] = \\theta_ {\\mathcal {F}} ^ {(j)} \\left(\\left[ t _ {m} ^ {(j)} \\left(x _ {i}\\right); \\dots ; t _ {v} ^ {(j)} \\left(x _ {i}\\right) \\right]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 614, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {m} ^ {(j + 1)} \\left(x _ {i}\\right) \\leftarrow t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right) - t _ {v} ^ {(j + 1)} \\left(x _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 660, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathcal{F}}^{(j)}$ represents the self-attention parameters in $\\mathcal{F}_j$ . If $j = 1$ , the input of $\\mathcal{F}_1$ becomes $\\mathcal{S}(x_i)$ . Note that $t_m^{(j + 1)}(x_i)$ and $t_v^{(j + 1)}(x_i)$ in Eq. (3) represent the updated meta and view token after the self-attention operation in $\\mathcal{F}_j$ with patch tokens of $x_i$ . Eq. (4) explicitly guides that $t_m^{(j + 1)}$ and $t_v^{(j + 1)}$ are distinguishable after Eq. (3), where $t_v^{(j + 1)}$ captures more view-related features and $t_m^{(j + 1)}$ can capture more view-unrelated features by removing $t_v^{(j + 1)}$ from the global (meta) features. Eq. (4) facilitates the learning of identity features via $t_m^{(j + 1)}$ , alleviating the disturbing of view bias. VDT decouples the view-related and -unrelated features layer by layer until the updated $t_m^{(N + 1)}$ and $t_v^{(N + 1)}$ are obtained, which serves as the well-decoupled identity", + "bbox": [ + 75, + 686, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and view features, respectively. During inference, we only use $t_m^{(N + 1)}$ as the whole retrieval evidence.", + "bbox": [ + 496, + 338, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, the proposed VDT maintains the same computational complexity as baseline (ViT). Given that the input to ViT has $T$ tokens and the dimensions of the tokens are $d$ , then the computational complexity of ViT is $\\mathcal{O}\\big(4Td^2 + 2T^2d\\big)$ [35]. Our VDT adds an additional $t_v$ token based on the baseline and adds an Eq. (4) operation in each VDT block. Therefore, the total computational complexity can be written as:", + "bbox": [ + 496, + 369, + 892, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {O} \\left(4 (T + 1) d ^ {2} + 2 (T + 1) ^ {2} d + N\\right) \\\\ \\Leftrightarrow \\mathcal {O} \\left(\\underbrace {4 T d ^ {2} + 2 T ^ {2} d} _ {\\text {b a s e l i n e c o m p l e x i t y}} + \\underbrace {4 d ^ {2} + 2 d + N} _ {\\text {c o n s t a n t c o m p l e x i t y}}\\right) \\tag {5} \\\\ \\Leftrightarrow \\mathcal {O} \\left(4 T d ^ {2} + 2 T ^ {2} d\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 506, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ counts all Eq. (4) in $N$ VDT blocks. Eq. (5) shows that the extra operations VDT introduces depend on only the network hyperparameters, which belong to constant complexity and can be omitted. Therefore, VDT maintains the same magnitude of complexity as that of the baseline.", + "bbox": [ + 496, + 614, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Optimization", + "text_level": 1, + "bbox": [ + 500, + 699, + 638, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To meet the task goal (Eq. (1)), the objective of $\\mathcal{F}(\\cdot, \\theta_{\\mathcal{F}})$ consists of three components: identity classifier, view classifier, and orthogonal loss.", + "bbox": [ + 496, + 722, + 890, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each $x_{i} \\in B$ , the identity classifier utilizes both the cross-entropy loss and the triplet loss to supervise $t_m^{N + 1}(x_i)$ , which can be written as:", + "bbox": [ + 496, + 768, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {c} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} y _ {i} \\log (\\hat {y} _ {i}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 823, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|B|$ represents the batch size, $y_{i}$ represents the identity label, and $\\hat{y}_i$ denotes the identity prediction by", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22003", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c66d843dc5e9a3cdded5ccb45c918918a0cc7a8b34ffbd02e4e7d44138c5c161.jpg", + "image_caption": [ + "(a) 3D pedestrian models." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 264, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/099d26879c3c1fbd25b09cf92162eda3e37660735b02d8ba6f33ed913a23e95a.jpg", + "image_caption": [ + "(b) Aerial-ground camera network deployment." + ], + "image_footnote": [], + "bbox": [ + 267, + 92, + 715, + 281 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/043a0a48f6b815686e3bac3f1b6ac828f467c54e999663cbffbb73da92dfc402.jpg", + "image_caption": [ + "(c) Diversity challenges.", + "Figure 3. Fig. $3\\mathrm{a}\\sim$ Fig. 3c shows the pedestrian models, camera deployment, and challenges during the CARGO construction, respectively. In Fig. 3b, \"A-Cam\" and \"G-Cam\" represent the aerial and ground cameras, where the yellow sectors represent the view range of the ground cameras, and the green arrows represent the motion strategy of the aerial cameras. The challenges displayed in Fig. 3c are view variation, illumination variation, occlusion, and resolution variation from top to bottom." + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 870, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$t_m^{N + 1}(x_i)$ . The triplet loss is defined as follows:", + "bbox": [ + 76, + 388, + 393, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {t} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} \\left[ \\begin{array}{c} \\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) - t _ {m} ^ {N + 1} \\left(x _ {i} ^ {+}\\right) \\right\\| _ {2} ^ {2} - \\\\ \\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) - t _ {m} ^ {N + 1} \\left(x _ {i} ^ {-}\\right) \\right\\| _ {2} ^ {2} + \\alpha \\end{array} \\right] _ {+}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 415, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{i}^{+} / x_{i}^{-}$ denotes the hard positive/negative sample of $x_{i}$ in the current $B$ , $[\\cdot]_{+}$ stands for $\\max(\\cdot, 0)$ , $\\alpha$ is the hyperparameter of the margin.", + "bbox": [ + 76, + 465, + 468, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "View classifier supervises $t_v^{N + 1}(x_i)$ with only the cross-entropy loss, which can be written as:", + "bbox": [ + 76, + 510, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v} ^ {c} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} v _ {i} \\log (\\hat {v} _ {i}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 551, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{i}$ represents the view label of $x_{i}$ , and $\\hat{v}_{i}$ denotes the view prediction by $t_v^{N + 1}(x_i)$ .", + "bbox": [ + 76, + 598, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable $t_m^{N + 1}(x_i)$ and $t_v^{N + 1}(x_i)$ to be truly decoupled, we impose an orthogonality loss that makes them less similar to each other, which can be written as:", + "bbox": [ + 76, + 628, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} \\frac {\\left| \\left\\langle t _ {m} ^ {N + 1} \\left(x _ {i}\\right) , t _ {v} ^ {N + 1} \\left(x _ {i}\\right) \\right\\rangle \\right|}{\\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) \\right\\| _ {2} \\cdot \\left\\| t _ {v} ^ {N + 1} \\left(x _ {i}\\right) \\right\\| _ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 681, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\langle \\cdot ,\\cdot \\rangle |$ represents the absolute value after the dot product of two token embeddings. We also provide a clear illustration of Eq. (9) in Fig. 2 (the gray rectangle in the middle on the right). Please note that the dashed blue and the yellow vector represent the meta token and view token outputted by Eq. (3) in $\\mathcal{F}_N$ respectively, and the blue vector represents the result of Eq. (4), which is the updated $t_m^{N + 1}(x_i)$ outputted by $\\mathcal{F}_N$ . It is shown that achieving decoupling requires a two-part collaboration, where Eq. (4)", + "bbox": [ + 76, + 729, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "separates view-related features from global features, and Eq. (9) constrains the remaining view-unrelated features to be orthogonally independent of view-related features. Overall, the total objective of VDT can be written as:", + "bbox": [ + 498, + 390, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {i} ^ {c} + \\mathcal {L} _ {i} ^ {t} + \\lambda \\left(\\mathcal {L} _ {v} ^ {c} + \\mathcal {L} _ {o}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 460, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter to balance multiple objectives.", + "bbox": [ + 500, + 489, + 885, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset: CARGO", + "text_level": 1, + "bbox": [ + 500, + 517, + 671, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Motivation", + "text_level": 1, + "bbox": [ + 500, + 542, + 620, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 1 compares the statistics and several other details of various ReID datasets, demonstrating the gap between the homogeneous and heterogeneous views. First, the dataset scale of the homogeneous view has far exceeded that of the heterogeneous view. For example, the identities in the latest AG-ReID [22] are about a quarter of that of the early Market1501 [17]. Second, synthetic data has been well-studied in the homogeneous view, and benefiting from low construction costs, synthetic datasets tend to contain more identities and images. However, the synthetic data study is absent in the view-heterogeneous ReID. The above limitations motivate us to contribute a large-scale synthetic dataset in the heterogeneous view. As can be seen from Tab. 1, the identities and images of CARGO are 12.8 and 4.9 times larger than that of AG-ReID, which brings the view-heterogeneous ReID to a comparable scale with the view-homogeneous ReID.", + "bbox": [ + 496, + 566, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.Description", + "text_level": 1, + "bbox": [ + 500, + 832, + 625, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The construction of CARGO is summarized as the construction of human models, the deployment of the camera network, and the data collection, shown in Fig. 3.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1The output of $|\\cdot |$ is the count of elements when the input is a set, and the output is the absolute value when the input is a number.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22004", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/15f0773b5e37ded495831b38b4aee4c04dffc1cd2511bfea9b1695febfb1ceec.jpg", + "table_caption": [ + "Table 2. Performance comparison of the mainstream methods under four settings of the proposed CARGO dataset. \"ALL\" denotes the overall retrieval performance of each method. \"G ↔ G\", \"A ↔ A\", and \"A ↔ G\" represent the performance of each model in several specific retrieval patterns. Rank1, mAP, and mINP are reported (\\%). The best performance is shown in **bold**." + ], + "table_footnote": [], + "table_body": "
MethodProtocol 1: ALLProtocol 2: G↔GProtocol 3: A↔AProtocol 4: A↔G
Rank1mAPmINPRank1mAPmINPRank1mAPmINPRank1mAPmINP
SBS [36]50.3243.0929.7672.3162.9948.2467.5049.7329.3231.2529.0018.71
PCB [37]51.0044.5032.2074.1067.6055.1055.0044.6027.0034.4030.4020.10
BoT [38]54.8146.4932.4077.6866.4751.3465.0049.7929.8236.2532.5621.46
MGN [39]54.8149.0836.5283.9371.0555.2065.0052.9636.7831.8733.4724.64
VV [40, 41]45.8338.8439.5772.3162.9948.2467.5049.7329.3231.2529.0018.71
AGW [21]60.2653.4440.2281.2571.6658.0967.5056.4840.4043.5740.9029.39
ViT [34]61.5453.5439.6282.1471.3457.5580.0064.4747.0743.1340.1128.20
VDT (Ours)64.1055.2041.1382.1471.5958.3982.5066.8350.2248.1242.7629.95
", + "bbox": [ + 78, + 142, + 893, + 327 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e80969a70fc9047e008b1165959131037647cd5db5e676a283d824c10ead5a3.jpg", + "table_caption": [ + "Table 3. Performance comparison of the mainstream methods under two settings of AG-ReID dataset. \"A $\\leftrightarrow$ G\", and \"G $\\leftrightarrow$ A\" represent the performance in two specific patterns. Rank1, mAP, and mINP are reported (\\%). The best performance is shown in bold." + ], + "table_footnote": [], + "table_body": "
MethodProtocol 1: A→GProtocol 2: G→A
Rank1mAPmINPRank1mAPmINP
SBS [36]73.5459.77-73.7062.27-
BoT [38]70.0155.47-71.2058.83-
OSNet [42]72.5958.32-74.2260.99-
VV [40, 41]77.2267.2341.4379.7369.8342.37
ViT [34]81.2872.38-82.6473.35-
Explain [22]81.4772.61-82.8573.39-
VDT (Ours)82.9174.4451.0686.5978.5752.87
", + "bbox": [ + 81, + 417, + 465, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, MakeHuman [43] is adopted to create person models with different identities by assigning randomly different values to predefined attributes, such as gender, height, age, body shape, skin color, hair, clothing, etc. We have created 5,000 pedestrians for subsequent operations, and some models have been shown in Fig. 3a.", + "bbox": [ + 75, + 582, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, Unity3D [44] is utilized to simulate a real-world streetscape. As shown in Fig. 3b, we pick three blocks to deploy the camera network, where a ground camera network (\"G-Cam\" area) serves relatively convenient areas such as downtown, and two aerial camera networks (\"A-Cam\" area) serve the suburbs of a city. In the \"G-Cam\" area, we deploy eight fixed cameras at each street corner. In the \"A-Cam\" area, we design different drone roaming strategies based on the size of the surveillance area. For the small area (left area), we deploy one drone with a $90^{\\circ}$ overhead view, allowing it to move counterclockwise around each street. For a large area (right area), we deploy individual drones on each of the four streets with a $45^{\\circ} \\sim 60^{\\circ}$ tilt view, allowing them to move back and forth on corresponding streets.", + "bbox": [ + 75, + 672, + 470, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we import human models into the streetscape", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and further move them under the ground and aerial networks. We allow multiple pedestrians to move simultaneously to simulate the random occlusion among pedestrians in the real world. Meanwhile, we collect valid images of the target persons if they appear in cameras without severe occlusion (train $\\leq 70\\%$ , and test $\\leq 75\\%$ ). In total, CARGO contains 108,563 person images from 13 cameras.", + "bbox": [ + 496, + 352, + 893, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Challenge", + "text_level": 1, + "bbox": [ + 500, + 468, + 614, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3c illustrates abundant challenges prevalent in real-world scenarios in the CARGO dataset. The most significant challenge is the view discrepancy from different types of cameras, which produces a vast variation and rotation of person postures. Besides, Fig. 3c and Tab. 1 show another change in the camera height, from 5 meters for the lowest ground camera to 75 meters for the highest $90^{\\circ}$ view drone, resulting in a huge resolution variation. The maximum resolution in CARGO is $1009 \\times 539$ , while the minimum is only $22 \\times 7$ . Moreover, CARGO also contains other complex illumination and occlusion variations, etc. Overall, CARGO simulates and covers major real-world scenarios and challenges and can be considered a large-scale benchmark.", + "bbox": [ + 496, + 492, + 893, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 500, + 703, + 624, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, please refer to the supplementary material for detailed experiment settings, visualization of retrieval and feature, cross-dataset evaluation, and other discussions.", + "bbox": [ + 496, + 729, + 893, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset and Metric", + "text_level": 1, + "bbox": [ + 500, + 786, + 684, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Two datasets are adopted for evaluation, CARGO and AG-ReID [22]. Some statistics regarding these datasets are shown in Tab. 1. As for CARGO, we split it into the train (51,451 images with 2500 IDs) and test sets (51,024 images with the remaining 2500 IDs) with an almost 1:1 ratio. In the training set, each person has an average of 10 images", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22005", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ea3f714c80118a1e4f945357c564964808e00914a4845bc2fe68e868f0d39b5.jpg", + "image_caption": [ + "(a) CARGO under Protocol 1." + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 331, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1d113337a88105d93a838a3bea75403120d242727524fa11ff3589eedf91e5a.jpg", + "image_caption": [ + "(b) AG-ReID under Protocol 1.", + "Figure 4. Fig. $4\\mathrm{a}\\sim$ Fig. 4c show the ablation experiments about orthogonal decoupling of view-related and view-unrelated features in the VDT on two dataset, which consists of two important parts, Eq. (4) and Eq. (9). Rank1, mAP, and mINP are reported $(\\%)$ ." + ], + "image_footnote": [], + "bbox": [ + 356, + 95, + 602, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dca13d0f43f36b7b417a67eb6e7656ea334e13fe8ead45f7575ae7d2266a6700.jpg", + "image_caption": [ + "(c) AG-RelD under Protocol 2." + ], + "image_footnote": [], + "bbox": [ + 627, + 95, + 872, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and on average moves across 6.47 ground and 3.61 aerial cameras. In the test set, we select 149 IDs as the query set, and all the remaining IDs are placed in the gallery as distractors. For each identity selected as a query, we keep only its images from two random cameras and use the image from one camera as the query and the image from the other as the gallery. On this basis, we design four protocols (ALL, $\\mathrm{A}\\leftrightarrow \\mathrm{A}$ , $\\mathrm{G}\\leftrightarrow \\mathrm{G}$ , and $\\mathrm{A}\\leftrightarrow \\mathrm{G}$ ) to adequately evaluate the model performances, where \"ALL\" focuses on the comprehensive retrieval performance, and the latter three focus on some specific retrieval patterns. $\\mathrm{A}\\leftrightarrow \\mathrm{A}$ only retains the data under the aerial camera in the test set for evaluation (60 query IDs with 134 images, 2404 gallery IDs with 18,444 images), same for $\\mathrm{G}\\leftrightarrow \\mathrm{G}$ (89 query IDs with 178 images, 2447 gallery IDs with 32,268 images). $\\mathrm{A}\\leftrightarrow \\mathrm{G}$ relabels the original test set into two domains (aerial and ground domain) based on the view label. The training set of all testing protocols retains same.", + "bbox": [ + 75, + 325, + 468, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for AG-ReID, there are 11,554 images with 199 IDs for training and 12,464 images with 189 IDs for testing following [22]. The test set contains 2,033 query images and 10,429 gallery images. There are two protocols in AG-ReID, denoted as $\\mathrm{A}\\rightarrow \\mathrm{G}$ and $\\mathrm{G}\\rightarrow \\mathrm{A}$ , where the former contains 1,701 aerial query images and 3,331 ground gallery images for 189 identities, and the latter contains 962 ground query images and 7,204 aerial gallery images for 189 identities. Optionally, AG-ReID also provides detailed person attribute labeling as additional prior information.", + "bbox": [ + 75, + 598, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric. The cumulative matching characteristic at Rank1, mean Average Precision (mAP), and mean Inverse Negative Penalty (mINP) [21] are adopted as the evaluation metrics.", + "bbox": [ + 76, + 750, + 468, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Performance", + "text_level": 1, + "bbox": [ + 76, + 801, + 212, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Accuracy. We compare the proposed VDT with other competitive methods on CARGO in Tab. 2, including CNN-based (BoT [38], SBS [36], VV [40, 41], MGN [39], AGW [21]) and transformer-based methods (ViT [34]), where VV means a vehicle ReID baseline with vanilla batch-based", + "bbox": [ + 76, + 824, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "triplet loss and diverse viewpoint batch selection. Similar experimental results on AG-ReID are reported in Tab. 3, where Explain [22] relies on pedestrian attribute priors and therefore does not generalize well to CARGO. Several performances in Tab. 3 follow the results reported in [22]. We can draw the following conclusions: (1) The proposed VDT achieves the state-of-the-art performances that are clearly improved over the baseline, especially in heterogeneous matching. For example, VDT surpasses the mAP/Rank1/mINP baseline by $4.99\\% / 2.65\\% / 1.75\\%$ on the $\\mathrm{A} \\leftrightarrow \\mathrm{G}$ of CARGO. Besides, VDT also brings different degree of benefits to other CARGO protocols. VDT also surpasses Explain on mAP/Rank1 by $3.74\\% / 5.18\\%$ on the $\\mathrm{G} \\rightarrow \\mathrm{A}$ of AG-ReID, which strongly demonstrates the importance of view decoupled operations in AGPReID to mitigate the disruption of identity representation by view bias whether in homogeneous or heterogeneous matching. (2) Previous view-homogeneous methods show varying degrees of performance degradation in AGPReID, especially on view-heterogeneous protocols, indicating that view bias causes the identity feature under different views to behave with poor cohesion. However, the above methods ignore this issue. (3) The superior performance of VDT does not come from a strong baseline but from the proposed method itself. On the CARGO dataset, the baseline does not consistently outperform the CNN-based AGW on all metrics; on the AGPReID dataset, the baseline is also weaker than the previous Explain. However, VDT achieves the SOTA performance on two datasets, especially on AG-ReID, using fewer prior labels and achieving better performance.", + "bbox": [ + 496, + 325, + 890, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speed. We report the inference time for baseline and VDT for retrieving a single image, which is 2.27ms and 2.28ms on the CARGO dataset and 1.44ms and 1.46ms on the AGReID dataset, respectively. These results validate the previous complexity analysis in Sec. 3.2 and show that VDT does not add additional time consumption. Slight time fluctuations tend to be acceptable, due to hardware devices.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22006", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78e5419c735f6b59836922ac26d9b3d964ee1d5cce75a1bdbff0f4014036bc6a.jpg", + "image_caption": [ + "(a) CARGO under Protocol 1." + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 349, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d900f7e1adbb94863037a0b3432ed0550fe6f2e9ba74aab45e5ec9b42358c353.jpg", + "image_caption": [ + "(b) AG-ReID under Protocol 1.", + "Figure 5. Fig. $5\\mathrm{a}\\sim$ Fig. 5c show the relationship between $\\lambda$ and performance on two datasets. For simplicity, only protocol 1 is shown on the CARGO dataset. Rank1, mAP, and mINP are reported (\\%). Avg represents the average performance of Rank1, mAP, and mINP." + ], + "image_footnote": [], + "bbox": [ + 357, + 95, + 620, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3be247d4965a2a69b5b0e74e30708154466d6aa0e4ca9937be193a74ac7d917.jpg", + "image_caption": [ + "(c) AG-ReID under Protocol 2." + ], + "image_footnote": [], + "bbox": [ + 630, + 95, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 325, + 228, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As described in Sec. 3.2, the core contribution of VDT lies in the decoupling of identity and view features by Eq. (4) and Eq. (9). Therefore, we aim to demonstrate the contribution of each part to VDT, which has been shown in Fig. 4. Similar trends in the two datasets demonstrate that the absence of each part leads to a degradation of the model's performance, and the absence of both simultaneously leads to a more severe performance degradation.", + "bbox": [ + 75, + 349, + 468, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By comparing Row1 and Row2, the absence of Eq. (4) leads to the degradation of $t_m$ and $t_v$ to be equivalent, focusing on the identity and view features in the image respectively, and constraining them to be independent through Eq. (9). However, lacking the guidance of explicit separation within VDT, the decoupling goal cannot be effectively achieved through Eq. (9) alone. Similarly, by comparing Row1 and Row3, the identity and view features outputted by VDT are not guaranteed to be orthogonally separated from each other if Eq. (9) is missing, and thus $t_m$ containing a minor view bias is still not strongly discriminative.", + "bbox": [ + 75, + 472, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Parameter Analysis", + "text_level": 1, + "bbox": [ + 76, + 648, + 264, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As described in Sec. 3.2, VDT introduces only one hyperparameter $\\lambda$ in Eq. (10) that needs to be manually tuned for multi-task balance. We select six magnitudes of $\\lambda$ from $1e - 4$ to 10 and the corresponding results are shown in Fig. 5. To show a clear tendency, we show the average performance of Rank1, mAP, and mINP at each $\\lambda$ . Although the performance shows the same trend of increasing and then decreasing on both datasets as $\\lambda$ increases, the peak performance is achieved when $\\lambda = 0.001$ on CARGO (Fig. 5a) and $\\lambda = 1.0$ on AG-ReID (Fig. 5b and Fig. 5c).", + "bbox": [ + 75, + 672, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The key of this phenomenon is the synchronous updating between identity and view classifier, because the view classifier is a simple binary classification, but the identity classifier on the two datasets shows different difficulties depending on the number of identities. As shown in Tab. 1,", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CARGO owns a thousand-level identities, so the identity classification is relatively complicated, and in order to keep tasks in Eq. (10) able to update synchronously, a small $\\lambda$ is more appropriate for view-related objectives. Similarly, a large $\\lambda$ is selected because of a hundred-level identities in AGPReID. We also provide an empirical $\\lambda$ selection guideline to avoid exhaustive search based on above analysis: $\\lambda$ tends to be smaller for datasets with larger identity scale.", + "bbox": [ + 496, + 325, + 890, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 498, + 460, + 767, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper focuses on ReID under a view-heterogeneous scenario, aerial-ground ReID (AGPReID). First, the view-decoupled transformer (VDT) is proposed specifically for the dramatic view differences in AGPReID, which decouples view-related and view-unrelated features by two major operations: hierarchically subtractive separation and orthogonal loss. Second, we contribute a large-scale dataset called CARGO, in which the identities and images are 12.8 and 4.9 times larger than the previous dataset. Experiments on two datasets demonstrate the superiority of VDT and the necessity of view decoupling in AGPReID.", + "bbox": [ + 496, + 486, + 890, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although view discrepancy is the most significant factor in AGPReID, it is not the only source of disruptions. Therefore, further exploration of discriminative identity representations under multiple disturbances in the aerial-ground camera network is a promising avenue of future research.", + "bbox": [ + 496, + 652, + 890, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 741, + 658, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project was supported in part by the NSFC (U22A2095, 62076258), in part by the Key-Area Research and Development Program of Guangzhou (202206030003), in part by Guangdong Project (No. 2020B1515120085), and in part by International Program Fund for Young Talent Scientific Research People, Sun Yat-Sen University.", + "bbox": [ + 496, + 767, + 890, + 858 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Geon Lee, Sanghoon Lee, Dohyung Kim, Younghoon Shin, Yongsang Yoon, and Bumsub Ham. Camera-driven representation learning for unsupervised domain adaptive person re-identification. In Int. Conf. Comput. Vis., pages 11453–11462, 2023. 1", + "[2] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In ACM Int. Conf. Multimedia, page 3115–3123, 2021. 3, 4", + "[3] Quan Zhang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling with second-order transformer for group re-identification. In AAAI, volume 36, pages 3318–3325, 2022.", + "[4] Quan Zhang, Jianhuang Lai, Xiaohua Xie, Xiaofeng Jin, and Sien Huang. Separable spatial-temporal residual graph for cloth-changing group re-identification. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-16, 2024.", + "[5] Quan Zhang, Kaiheng Dang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Modeling 3d layout for group re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7512-7520, 2022. 1, 2, 4", + "[6] Jianxiong Tang, Jianhuang Lai, Xiaohua Xie, Lingxiao Yang, and Wei-Shi Zheng. Ac2as: Activation consistency coupled ann-snn framework for fast and memory-efficient snn training. Pattern Recognition, 2023. 1", + "[7] Qi Chen, Yun Chen, Yuheng Huang, Xiaohua Xie, and Lingxiao Yang. Region-based online selective examination for weakly supervised semantic segmentation. Information Fusion, page 102311, 2024.", + "[8] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7713-7722, June 2022.", + "[9] Junyang Qiu, Zhanxiang Feng, Lei Wang, and Jianhuang Lai. Salient part-aligned and keypoint disentangling transformer for person re-identification in aerial imagery. In Int. Conf. Multimedia and Expo, 2024. 2", + "[10] Pengze Zhang, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Pose guided person image generation via dual-task correlation and affinity learning. IEEE Trans. Vis. Comput. Graph., pages 1-18, 2023.", + "[11] Jianxiong Tang, Jian-Huang Lai, Xiaohua Xie, and Lingxiao Yang. Spike count maximization for neuromorphic vision recognition. In *IJCAI*, 2023.", + "[12] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Adv. Neural Inform. Process. Syst., 2023.", + "[13] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4288-4298, 2022.", + "[14] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4", + "[15] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling for group re-identification. Int. J. Comput. Vis., 2024." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Lingxiao Yang, Ru-Yuan Zhang, Lida Li, and Xiaohua Xie. Simam: A simple, parameter-free attention module for convolutional neural networks. pages 11863-11874. PMLR, 2021. 1", + "[17] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Int. Conf. Comput. Vis., pages 1116–1124, 2015. 1, 2, 4, 5", + "[18] Shizhou Zhang, Qi Zhang, Yifei Yang, Xing Wei, Peng Wang, Bingliang Jiao, and Yanning Zhang. Person re-identification in aerial imagery. IEEE Trans. Multimedia, 23:281-291, 2021. 1, 2, 4", + "[19] Quan Zhang, Jianhuang Lai, and Xiaohua Xie. Learning modal-invariant angular metric by cyclic projection network for vis-nir person re-identification. IEEE Trans. Image Process., 30:8019-8033, 2021. 2", + "[20] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Seeing like a human: Asynchronous learning with dynamic progressive refinement for person re-identification. IEEE Trans. Image Process., 31:352-365, 2022.", + "[21] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44(6):2872-2893, 2022. 2, 6, 7, 1", + "[22] H. Nguyen, K. Nguyen, S. Sridharan, and C. Fookes. Aerial-ground person re-id. In Int. Conf. Multimedia and Expo, pages 2585-2590, 2023. 2, 3, 4, 5, 6, 7", + "[23] Lei Wang, Quan Zhang, Junyang Qiu, and Jianhuang Lai. Rotation exploration transformer for aerial person re-identification. In Int. Conf. Multimedia and Expo, 2024. 2", + "[24] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11506–11515, 2021. 2, 3, 4", + "[25] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 79–88, 2018. 2", + "[26] Fei Xiong, Mengran Gou, Octavia Camps, and Mario Sznaier. Person re-identification using kernel-based metric learning methods. In Eur. Conf. Comput. Vis., pages 1–16, 2014. 2", + "[27] Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. Person re-identification by local maximal occurrence representation and metric learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2197–2206, 2015. 2", + "[28] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In Eur. Conf. Comput. Vis., pages 501-518, 2018. 2", + "[29] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Int. Conf. Comput. Vis., pages 15013-15022, 2021. 2", + "[30] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22008", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In IEEE Conf. Comput. Vis. Pattern Recog., pages 16266-16275, 2021. 2, 4", + "[31] Shuoyi Chen, Mang Ye, and Bo Du. Rotation invariant transformer for recognizing object in uavs. In ACM Int. Conf. Multimedia, page 2565-2574, 2022. 2", + "[32] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In ACM Int. Conf. Multimedia, page 3422–3430, 2020. 4", + "[33] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4890–4899, 2022. 4", + "[34] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2021. 3, 6, 7, 1", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 4", + "[36] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. In ACM Int. Conf. Multimedia, page 9664–9667, 2023. 6, 7, 1", + "[37] Yifan Sun, Liang Zheng, Yali Li, Yi Yang, Qi Tian, and Shengjin Wang. Learning part-based convolutional features for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 43:902–917, 2021. 6", + "[38] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2019. 6, 7", + "[39] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM Int. Conf. Multimedia, pages 274–282, 2018. 6, 7", + "[40] Ratnesh Kuma, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. Vehicle re-identification: an efficient baseline using triplet embedding. In International Joint Conference on Neural Networks, pages 1–9, 2019. 6, 7", + "[41] Ratnesh Kumar, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. A strong and efficient baseline for vehicle re-identification using deep triplet embedding. Journal of Artificial Intelligence and Soft Computing Research, 10(1):27-45, 2020. 6, 7", + "[42] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Learning generalisable omni-scale representations for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):5056–5069, 2022. 6", + "[43] Leyde Briceno and Gunther Paul. Makehuman: a review of the modelling framework. In Congress of the International Ergonomics Association, pages 224-232, 2018. 6", + "[44] Unity Technologies. Unity3D: Cross-platform 3D engine, 2021. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248–255, 2009. 1", + "[46] Léon Bottou. Stochastic gradient descent tricks. Neural Networks: Tricks of the Trade: Second Edition, pages 421-436, 2012. 1", + "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., pages 8024-8035, 2019. 1" + ], + "bbox": [ + 501, + 92, + 890, + 314 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22009", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_model.json b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4749267cafb1585db76d3fe43f29914c761d0668 --- /dev/null +++ b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.884, + 0.257 + ], + "angle": 0, + "content": "Quan Zhang\\(^{1,5}\\) Lei Wang\\(^{1}\\) Vishal M. Patel\\(^{5}\\) Xiaohua Xie\\(^{1,2,3,4}\\) Jianhuang Lai\\(^{1,2,3,4*}\\) \n\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-Sen University, China \n\\(^{2}\\)Pazhou Lab (HuangPu), Guangdong, China" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.257, + 0.875, + 0.345 + ], + "angle": 0, + "content": "3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China \n4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China \n5Department of Electrical and Computer Engineering, Johns Hopkins University, USA \n{zhangq48, wanglei75}@mail2.sysu.edu.cn, {stsljh, xiexiaoh6}@mail.sysu.edu.cn, vpatel36@jhu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.38, + 0.314, + 0.396 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.474, + 0.775 + ], + "angle": 0, + "content": "Existing person re-identification methods have achieved remarkable advances in appearance-based identity association across homogeneous cameras, such as ground-ground matching. However, as a more practical scenario, aerial-ground person re-identification (AGPReID) among heterogeneous cameras has received minimal attention. To alleviate the disruption of discriminative identity representation by dramatic view discrepancy as the most significant challenge in AGPReID, the view-decoupled transformer (VDT) is proposed as a simple yet effective framework. Two major components are designed in VDT to decouple view-related and view-unrelated features, namely hierarchical subtractive separation and orthogonal loss, where the former separates these two features inside the VDT, and the latter constrains these two to be independent. In addition, we contribute a large-scale AGPReID dataset called CARGO, consisting of five/eight aerial/ground cameras, 5,000 identities, and 108,563 images. Experiments on two datasets show that VDT is a feasible and effective solution for AGPReID, surpassing the previous method on mAP/Rank1 by up to \\(5.0\\% /2.7\\%\\) on CARGO and \\(3.7\\% /5.2\\%\\) on AG-ReID, keeping the same magnitude of computational complexity. Our project is available at https://github.com/LinlyAC/VDT-AGPReID." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.805, + 0.21, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.831, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Person re-identification (ReID) aims to associate interested person images based on cross-camera identity similarity, which plays a significant and positive contribution to the se" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.379, + 0.871, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.643, + 0.894, + 0.728 + ], + "angle": 0, + "content": "Figure 1. View-homogeneous vs. view-heterogeneous ReID, where the former focuses on ground-only or aerial-only camera networks, and the latter considers the aerial-ground mixed camera network. Thus, view-heterogeneous ReID considers aerial-aerial, ground-ground, and aerial-ground matching, which is more challenging and practical than the existing view-homogeneous ReID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "curity and safety of society and citizens [1-5]. Traditional ReID has achieved remarkable progress in the deep learning era [6-16], but it is far from real-world scenarios for the following reasons. (1) Most existing ReID datasets are collected from homogeneous camera networks, which comprises of the same type cameras, such as ground-only [17] or aerial-only [18] camera networks. However, as can be seen from Fig. 1, real-world surveillance systems tend to be deployed as heterogeneous camera networks comprising" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22000" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "aerial and ground cameras instead of a single type alone. Ground cameras cover well-developed areas (city centers), and aerial cameras cover poorly-developed areas (suburbs) due to their broad view range. The complementary nature of the two types would maximize the effectiveness of ReID. (2) Existing ReID methods mainly consider homogeneous matching (ground-ground [19-21] and aerial-aerial [9, 22, 23]), which is ineffective in dealing with the dramatic view discrepancy among heterogeneous matching (aerial-ground). Despite its significance, related research is extremely scarce." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.732 + ], + "angle": 0, + "content": "In this paper, we consider the novel and practical view-heterogeneous ReID problem, specifically ReID under the aerial-ground camera network (AGPReID). We propose a view-decoupled transformer (VDT) to specifically tackle the dramatic view discrepancy, which serves as a significant challenge within AGPReID to hinder homogeneous and heterogeneous matchings. Motivated by the fact that view-related features are useless for discriminative identity representations, VDT aims to decouple view-related and view-unrelated components, further facilitating discriminative identity learning from the remaining view-unrelated features. There are two key parts within VDT to achieve view decoupling, namely (a) hierarchically subtractive separation and (b) orthogonal loss. Specifically, after tokenizing the input image to a series of patch tokens, VDT appends two extra tokens (the meta token and view token) to feed them into a transformer stacked by multiple VDT blocks. The meta token captures global representation in the image, and the view token aims to extract view-related features. In each VDT block, (a) is achieved by the subtractive values of the meta token and view token after self-attention operations as the updated meta token for the next block. This means that VDT hierarchically separates view-related features from global features to facilitate identity learning from remaining view-unrelated features. When the meta and view tokens are obtained after the last block, (b) is designed to constrain the identity features to be independent of the view features, ultimately achieving the orthogonal decoupling of view-related and view-unrelated components. Meanwhile, the meta and view tokens will be supervised by identity and view labels, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Considering the scarcity [22] and privacy [5, 24] of the AGPReID datasets, we contribute a large-scale Civic AeRial-GrOund (CARGO) dataset, which is collected from a synthetic city scenario including five aerial and eight ground cameras. CARGO has totally collected 5,000 identities and 108,563 images, which contain not only significant view discrepancy but also plenty of variations of resolution, illumination, occlusion, etc. Regarding evaluation, CARGO considers multiple patterns (aerial-aerial, ground-ground, aerial-ground matchings), which could serve as a comprehensive benchmark. Contributions of this paper can" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.68, + 0.106 + ], + "angle": 0, + "content": "be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.108, + 0.894, + 0.198 + ], + "angle": 0, + "content": "- We focus on ReID under the aerial-ground camera network (AGPReID) scenario and propose a view-decoupled transformer (VDT) specifically for the dramatic view discrepancy within AGPReID, which achieves view-related and view-unrelated features decoupling by hierarchical subtractive separation and orthogonal loss." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.894, + 0.258 + ], + "angle": 0, + "content": "- We contribute a large-scale synthetic dataset, called CARGO, to advance the AGPReID as a benchmark, which contains five/eight aerial/ground cameras, 5,000 identities, 108,563 images, and complex challenges." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.258, + 0.894, + 0.335 + ], + "angle": 0, + "content": "- Experiments demonstrate the superiority of VDT on two datasets, which shows that VDT surpasses the previous method on mAP/Rank1 by up to \\(5.0\\% / 2.7\\%\\) on CARGO and \\(3.7\\% / 5.2\\%\\) on AG-ReID, respectively, maintaining the same magnitude of computational complexity." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.108, + 0.894, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.348, + 0.642, + 0.365 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.374, + 0.731, + 0.39 + ], + "angle": 0, + "content": "2.1. View-homogeneous ReID" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.593 + ], + "angle": 0, + "content": "View-homogeneous ReID deals with images from ground-only or aerial-only camera networks, where the ground-only camera network receives the broadest attention in the ReID task because it is the most common use case. Many ground-only datasets have been contributed in the literature, such as Market1501 [17], MSMT17 [25], etc. Accordingly, lots of methods have been proposed to bring rapid development to ReID, including handcrafted feature-based methods [26, 27], CNN-based methods [19-21, 28], and transformer-based methods [29]. Regarding aerial-only research, only some pioneering efforts have contributed substantial datasets and methods [18, 30, 31] because it has not received as much attention as ground-only ReID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.656 + ], + "angle": 0, + "content": "However, it is difficult to directly transfer the above methods to the view-heterogeneous ReID because these methods ignore the significant view differences between aerial and ground cameras, leading to weak performances." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.738, + 0.682 + ], + "angle": 0, + "content": "2.2. View-heterogeneous ReID" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Compared to the abundant research in Sec. 2.1, studies on view-heterogeneous ReID are almost nonexistent. In this paper, heterogeneous views refer to AGPReID, and the fundamental challenge is the dramatic view discrepancy. To the best of our knowledge, one recent work [22] has formally attempted to address this task. They collected an outdoor scene dataset, AG-ReID, with both identity and attribute labeling of pedestrians. Moreover, they proposed an explainable model by attribute information to guide model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "While pioneering and critical to AGPReID, this work is still imperfect for the following reasons. (1) Dataset. Compared to [22], our CARGO dataset comprehensively focuses on multiple matching patterns that may appear in heterogeneous views rather than a single aerial-ground matching in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22001" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.873, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.895, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed VDT framework, which consists of \\( N \\) VDT blocks and three-part loss functions. Meta and view tokens capture global and view-related features in images, respectively. Each VDT block (light blue module) consists of a standard self-attention encoder layer and an inner feature subtraction operation, achieving layer-by-layer decoupling of view-related and view-unrelated features. Orthogonal loss constrains the above two features to be further independent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.471, + 0.501 + ], + "angle": 0, + "content": "AG-ReID. (2) Method. Compared to explainable methods in [22], our VDT relies on a lesser amount of priori labeling and thus has a stronger generalization. The experiments also show that our VDT achieves stronger performances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.295, + 0.526 + ], + "angle": 0, + "content": "2.3. Synthetic ReID Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Data privacy is an inevitable issue for the ReID datasets, and synthetic data provides a feasible solution because it not only achieves privacy protection for real-world pedestrians but also has lower collection costs. Many synthetic ReID datasets have been published in the literature, such as PersonX [14], UnrealPerson [24], WePerson [2], etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.471, + 0.716 + ], + "angle": 0, + "content": "However, existing synthetic datasets mainly focus on simulating pedestrian data under view-homogeneous camera networks, and the simulation of aerial-ground views is poorly studied. To this end, we contribute the CARGO dataset as the first large-scale synthetic dataset for AGPReID." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.169, + 0.747 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.322, + 0.772 + ], + "angle": 0, + "content": "3.1. Formulation and Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Formulation. An AGPReLU dataset \\(\\mathcal{D} = \\{(x_i, y_i, v_i)\\}_{i=1}^{|\\mathcal{D}|}\\) consists of a training set \\(\\mathcal{D}^{tr}\\) and a test set \\(\\mathcal{D}^{te}\\), where \\(x_i\\) represents the \\(i\\)-th person image, and \\(y_i\\) and \\(v_i\\) stand for the corresponding identity and view labels, respectively. Note that \\(v_i \\in \\{v^a, v^g\\}\\) is readily acquired by the known camera labels in \\(\\mathcal{D}\\) to easily distinguish whether \\(x_i\\) belongs to an aerial \\(v^a\\) or ground \\(v^g\\) view. Significant view discrepancy between \\(v^a\\) and \\(v^g\\) leads to a view-biased feature" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.893, + 0.5 + ], + "angle": 0, + "content": "space, which behaves as low intra-identity and high interidentity similarity. Therefore, the goal of AGPReID methods is to design a model \\(\\mathcal{F}(\\cdot;\\theta_{\\mathcal{F}})\\) with learnable parameter \\(\\theta_{\\mathcal{F}}\\) against the view bias, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.51, + 0.892, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\sum_ {i} \\left[ \\begin{array}{c} \\left\\| \\mathcal {F} \\left(x _ {i}; \\theta_ {\\mathcal {F}}\\right) - \\mathcal {F} \\left(x _ {i} ^ {+}; \\theta_ {\\mathcal {F}}\\right) \\right\\| _ {2} ^ {2} - \\\\ \\left\\| \\mathcal {F} \\left(x _ {i}; \\theta_ {\\mathcal {F}}\\right) - \\mathcal {F} \\left(x _ {i} ^ {-}; \\theta_ {\\mathcal {F}}\\right) \\right\\| _ {2} ^ {2} \\end{array} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.595 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _2\\) denotes the L2 distance, \\(x_{i}^{+} / x_{i}^{-}\\) denotes the person image that has the same/different identity with \\(x_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.893, + 0.837 + ], + "angle": 0, + "content": "Overview. As shown in Fig. 2, we design a simple yet effective framework called view-decoupled transformer (VDT), to tackle the view discrepancy challenge in AGPReID. For a batched data \\( B \\) that contains both \\( v^a \\) and \\( v^g \\) view, weckenize them as a series of image tokens (green squares), then append a meta and view token (blue and yellow squares) to them, which serve as input to VDT. The VDT network consists of \\( N \\) blocks, where each block first performs the standard self-attention encoding and then performs a subtraction operation between meta and view tokens to explicitly separate the view-related feature from the global feature. Finally, the identity and view classifier will supervise the updated meta and view tokens outputted from VDT. Besides, an orthogonal loss is proposed that makes these two tokens irrelevant, thus achieving the complete decoupling of view-related and view-unrelated features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.762, + 0.864 + ], + "angle": 0, + "content": "3.2. View-decoupled Transformer" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The proposed VDT framework is based on Vit-Base [34]. For each image \\( x_{i} \\) in the batched data \\( B \\), VDT first" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22002" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.12 + ], + "angle": 0, + "content": "Table 1. Statistical comparisons of existing datasets, including view-homogeneous (ground or aerial) and view-heterogeneous (ground and aerial) ReID. “\\(\\star\\)” denotes City1M is used for group ReID, where its images correspond to the number of group images." + }, + { + "type": "table", + "bbox": [ + 0.16, + 0.129, + 0.812, + 0.314 + ], + "angle": 0, + "content": "
DatasetViewData#PersonID#Camera#Image#Height
Market1501 [17]GroundReal1,501632,668< 10m
PersonX [14]GroundSynthetic1,2666273,456-
RandPerson [32]GroundSynthetic8,00019228,655-
UnrealPerson [24]GroundSynthetic3,00034120,000-
WePerson [2]GroundSynthetic1,500404,000,000-
ClonedPerson [33]GroundSynthetic5,62124887,766-
City1M* [5]GroundSynthetic45,00081,840,000-
PRAI1581 [18]AerialReal1,581239,46120 ~ 60m
UAVHuman [30]AerialReal1,144141,2902 ~ 8m
AG-ReID [22]Aerial-GroundReal3882 (1A+1G)21,89315 ~ 45m
CARGO (Ours)Aerial-GroundSynthetic5,00013 (5A+8G)108,5635 ~ 75m
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.476 + ], + "angle": 0, + "content": "divides \\( x_{i} \\) evenly and non-overlappingly into \\( M \\) image patches, which are further tokenized into \\( M \\) patch embeddings by a \\( 1 \\times 1 \\) convolution and can be denoted as \\( [t_{p,1}; t_{p,2}; \\dots; t_{p,M}] \\). After that, we append two additional learnable tokens, meta token \\( t_{m} \\) and view token \\( t_{v} \\), in the patch tokens, where \\( t_{m} \\) aims to capture the global image representation, \\( t_{v} \\) focuses only on the view-related feature. Next, we serialize all the tokens by assigning them the corresponding positional embedding, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.489, + 0.47, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (x _ {i}) \\triangleq \\left[ t _ {m}; t _ {p, 1}; t _ {p, 2}; \\dots ; t _ {p, M}; t _ {v} \\right] + \\mathbf {E} _ {p o s}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.544 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}(\\cdot)\\) is defined as the serialization operation for \\(x_{i}\\), \\(t_m, t_{p,i}, t_v \\in \\mathbb{R}^d\\), and \\(\\mathbf{E}_{pos} \\in \\mathbb{R}^{(M + 2) \\times d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.603 + ], + "angle": 0, + "content": "As shown in Fig. 2, VDT \\(\\mathcal{F}(\\cdot ;\\theta_{\\mathcal{F}})\\) is stacked with \\(N\\) VDT blocks \\(\\mathcal{F}_j(\\cdot ;\\theta_{\\mathcal{F}}^j),1\\leq j\\leq N.\\) Each \\(\\mathcal{F}_j(\\cdot ;\\theta_{\\mathcal{F}}^j)\\) contains two operations, self-attention encoding and subtractive separation, which are described as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.615, + 0.47, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\left[ t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right); \\dots ; t _ {v} ^ {(j + 1)} \\left(x _ {i}\\right) \\right] = \\theta_ {\\mathcal {F}} ^ {(j)} \\left(\\left[ t _ {m} ^ {(j)} \\left(x _ {i}\\right); \\dots ; t _ {v} ^ {(j)} \\left(x _ {i}\\right) \\right]\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.661, + 0.469, + 0.681 + ], + "angle": 0, + "content": "\\[\nt _ {m} ^ {(j + 1)} \\left(x _ {i}\\right) \\leftarrow t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right) - t _ {v} ^ {(j + 1)} \\left(x _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathcal{F}}^{(j)}\\) represents the self-attention parameters in \\(\\mathcal{F}_j\\). If \\(j = 1\\), the input of \\(\\mathcal{F}_1\\) becomes \\(\\mathcal{S}(x_i)\\). Note that \\(t_m^{(j + 1)}(x_i)\\) and \\(t_v^{(j + 1)}(x_i)\\) in Eq. (3) represent the updated meta and view token after the self-attention operation in \\(\\mathcal{F}_j\\) with patch tokens of \\(x_i\\). Eq. (4) explicitly guides that \\(t_m^{(j + 1)}\\) and \\(t_v^{(j + 1)}\\) are distinguishable after Eq. (3), where \\(t_v^{(j + 1)}\\) captures more view-related features and \\(t_m^{(j + 1)}\\) can capture more view-unrelated features by removing \\(t_v^{(j + 1)}\\) from the global (meta) features. Eq. (4) facilitates the learning of identity features via \\(t_m^{(j + 1)}\\), alleviating the disturbing of view bias. VDT decouples the view-related and -unrelated features layer by layer until the updated \\(t_m^{(N + 1)}\\) and \\(t_v^{(N + 1)}\\) are obtained, which serves as the well-decoupled identity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.371 + ], + "angle": 0, + "content": "and view features, respectively. During inference, we only use \\( t_m^{(N + 1)} \\) as the whole retrieval evidence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.893, + 0.492 + ], + "angle": 0, + "content": "In addition, the proposed VDT maintains the same computational complexity as baseline (ViT). Given that the input to ViT has \\( T \\) tokens and the dimensions of the tokens are \\( d \\), then the computational complexity of ViT is \\( \\mathcal{O}\\big(4Td^2 + 2T^2d\\big) \\) [35]. Our VDT adds an additional \\( t_v \\) token based on the baseline and adds an Eq. (4) operation in each VDT block. Therefore, the total computational complexity can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.507, + 0.892, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {O} \\left(4 (T + 1) d ^ {2} + 2 (T + 1) ^ {2} d + N\\right) \\\\ \\Leftrightarrow \\mathcal {O} \\left(\\underbrace {4 T d ^ {2} + 2 T ^ {2} d} _ {\\text {b a s e l i n e c o m p l e x i t y}} + \\underbrace {4 d ^ {2} + 2 d + N} _ {\\text {c o n s t a n t c o m p l e x i t y}}\\right) \\tag {5} \\\\ \\Leftrightarrow \\mathcal {O} \\left(4 T d ^ {2} + 2 T ^ {2} d\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.692 + ], + "angle": 0, + "content": "where \\(N\\) counts all Eq. (4) in \\(N\\) VDT blocks. Eq. (5) shows that the extra operations VDT introduces depend on only the network hyperparameters, which belong to constant complexity and can be omitted. Therefore, VDT maintains the same magnitude of complexity as that of the baseline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.7, + 0.64, + 0.717 + ], + "angle": 0, + "content": "3.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.768 + ], + "angle": 0, + "content": "To meet the task goal (Eq. (1)), the objective of \\(\\mathcal{F}(\\cdot, \\theta_{\\mathcal{F}})\\) consists of three components: identity classifier, view classifier, and orthogonal loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.769, + 0.892, + 0.816 + ], + "angle": 0, + "content": "For each \\( x_{i} \\in B \\), the identity classifier utilizes both the cross-entropy loss and the triplet loss to supervise \\( t_m^{N + 1}(x_i) \\), which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.824, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {c} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} y _ {i} \\log (\\hat {y} _ {i}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(|B|\\) represents the batch size, \\(y_{i}\\) represents the identity label, and \\(\\hat{y}_i\\) denotes the identity prediction by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22003" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.266, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.284, + 0.251, + 0.296 + ], + "angle": 0, + "content": "(a) 3D pedestrian models." + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.093, + 0.716, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.284, + 0.615, + 0.297 + ], + "angle": 0, + "content": "(b) Aerial-ground camera network deployment." + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.091, + 0.871, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.284, + 0.858, + 0.297 + ], + "angle": 0, + "content": "(c) Diversity challenges." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.894, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Fig. \\(3\\mathrm{a}\\sim\\) Fig. 3c shows the pedestrian models, camera deployment, and challenges during the CARGO construction, respectively. In Fig. 3b, \"A-Cam\" and \"G-Cam\" represent the aerial and ground cameras, where the yellow sectors represent the view range of the ground cameras, and the green arrows represent the motion strategy of the aerial cameras. The challenges displayed in Fig. 3c are view variation, illumination variation, occlusion, and resolution variation from top to bottom." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.39, + 0.395, + 0.407 + ], + "angle": 0, + "content": "\\(t_m^{N + 1}(x_i)\\). The triplet loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.416, + 0.47, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {t} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} \\left[ \\begin{array}{c} \\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) - t _ {m} ^ {N + 1} \\left(x _ {i} ^ {+}\\right) \\right\\| _ {2} ^ {2} - \\\\ \\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) - t _ {m} ^ {N + 1} \\left(x _ {i} ^ {-}\\right) \\right\\| _ {2} ^ {2} + \\alpha \\end{array} \\right] _ {+}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.47, + 0.512 + ], + "angle": 0, + "content": "where \\( x_{i}^{+} / x_{i}^{-} \\) denotes the hard positive/negative sample of \\( x_{i} \\) in the current \\( B \\), \\( [\\cdot]_{+} \\) stands for \\( \\max(\\cdot, 0) \\), \\( \\alpha \\) is the hyperparameter of the margin." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.511, + 0.469, + 0.543 + ], + "angle": 0, + "content": "View classifier supervises \\(t_v^{N + 1}(x_i)\\) with only the cross-entropy loss, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.552, + 0.47, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v} ^ {c} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} v _ {i} \\log (\\hat {v} _ {i}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.469, + 0.629 + ], + "angle": 0, + "content": "where \\( v_{i} \\) represents the view label of \\( x_{i} \\), and \\( \\hat{v}_{i} \\) denotes the view prediction by \\( t_v^{N + 1}(x_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.469, + 0.673 + ], + "angle": 0, + "content": "To enable \\( t_m^{N + 1}(x_i) \\) and \\( t_v^{N + 1}(x_i) \\) to be truly decoupled, we impose an orthogonality loss that makes them less similar to each other, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.683, + 0.47, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} \\frac {\\left| \\left\\langle t _ {m} ^ {N + 1} \\left(x _ {i}\\right) , t _ {v} ^ {N + 1} \\left(x _ {i}\\right) \\right\\rangle \\right|}{\\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) \\right\\| _ {2} \\cdot \\left\\| t _ {v} ^ {N + 1} \\left(x _ {i}\\right) \\right\\| _ {2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.47, + 0.868 + ], + "angle": 0, + "content": "where \\(|\\langle \\cdot ,\\cdot \\rangle |\\) represents the absolute value after the dot product of two token embeddings. We also provide a clear illustration of Eq. (9) in Fig. 2 (the gray rectangle in the middle on the right). Please note that the dashed blue and the yellow vector represent the meta token and view token outputted by Eq. (3) in \\(\\mathcal{F}_N\\) respectively, and the blue vector represents the result of Eq. (4), which is the updated \\(t_m^{N + 1}(x_i)\\) outputted by \\(\\mathcal{F}_N\\). It is shown that achieving decoupling requires a two-part collaboration, where Eq. (4)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1The output of \\(|\\cdot |\\) is the count of elements when the input is a set, and the output is the absolute value when the input is a number." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.391, + 0.892, + 0.452 + ], + "angle": 0, + "content": "separates view-related features from global features, and Eq. (9) constrains the remaining view-unrelated features to be orthogonally independent of view-related features. Overall, the total objective of VDT can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.462, + 0.892, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {i} ^ {c} + \\mathcal {L} _ {i} ^ {t} + \\lambda \\left(\\mathcal {L} _ {v} ^ {c} + \\mathcal {L} _ {o}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.491, + 0.886, + 0.507 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter to balance multiple objectives." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.673, + 0.535 + ], + "angle": 0, + "content": "4. Dataset: CARGO" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.622, + 0.558 + ], + "angle": 0, + "content": "4.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Tab. 1 compares the statistics and several other details of various ReID datasets, demonstrating the gap between the homogeneous and heterogeneous views. First, the dataset scale of the homogeneous view has far exceeded that of the heterogeneous view. For example, the identities in the latest AG-ReID [22] are about a quarter of that of the early Market1501 [17]. Second, synthetic data has been well-studied in the homogeneous view, and benefiting from low construction costs, synthetic datasets tend to contain more identities and images. However, the synthetic data study is absent in the view-heterogeneous ReID. The above limitations motivate us to contribute a large-scale synthetic dataset in the heterogeneous view. As can be seen from Tab. 1, the identities and images of CARGO are 12.8 and 4.9 times larger than that of AG-ReID, which brings the view-heterogeneous ReID to a comparable scale with the view-homogeneous ReID." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.627, + 0.849 + ], + "angle": 0, + "content": "4.2.Description" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The construction of CARGO is summarized as the construction of human models, the deployment of the camera network, and the data collection, shown in Fig. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22004" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Performance comparison of the mainstream methods under four settings of the proposed CARGO dataset. \"ALL\" denotes the overall retrieval performance of each method. \"G ↔ G\", \"A ↔ A\", and \"A ↔ G\" represent the performance of each model in several specific retrieval patterns. Rank1, mAP, and mINP are reported (\\%). The best performance is shown in **bold**." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.143, + 0.894, + 0.328 + ], + "angle": 0, + "content": "
MethodProtocol 1: ALLProtocol 2: G↔GProtocol 3: A↔AProtocol 4: A↔G
Rank1mAPmINPRank1mAPmINPRank1mAPmINPRank1mAPmINP
SBS [36]50.3243.0929.7672.3162.9948.2467.5049.7329.3231.2529.0018.71
PCB [37]51.0044.5032.2074.1067.6055.1055.0044.6027.0034.4030.4020.10
BoT [38]54.8146.4932.4077.6866.4751.3465.0049.7929.8236.2532.5621.46
MGN [39]54.8149.0836.5283.9371.0555.2065.0052.9636.7831.8733.4724.64
VV [40, 41]45.8338.8439.5772.3162.9948.2467.5049.7329.3231.2529.0018.71
AGW [21]60.2653.4440.2281.2571.6658.0967.5056.4840.4043.5740.9029.39
ViT [34]61.5453.5439.6282.1471.3457.5580.0064.4747.0743.1340.1128.20
VDT (Ours)64.1055.2041.1382.1471.5958.3982.5066.8350.2248.1242.7629.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.351, + 0.472, + 0.408 + ], + "angle": 0, + "content": "Table 3. Performance comparison of the mainstream methods under two settings of AG-ReID dataset. \"A \\(\\leftrightarrow\\) G\", and \"G \\(\\leftrightarrow\\) A\" represent the performance in two specific patterns. Rank1, mAP, and mINP are reported (\\%). The best performance is shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.418, + 0.466, + 0.557 + ], + "angle": 0, + "content": "
MethodProtocol 1: A→GProtocol 2: G→A
Rank1mAPmINPRank1mAPmINP
SBS [36]73.5459.77-73.7062.27-
BoT [38]70.0155.47-71.2058.83-
OSNet [42]72.5958.32-74.2260.99-
VV [40, 41]77.2267.2341.4379.7369.8342.37
ViT [34]81.2872.38-82.6473.35-
Explain [22]81.4772.61-82.8573.39-
VDT (Ours)82.9174.4451.0686.5978.5752.87
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.674 + ], + "angle": 0, + "content": "First, MakeHuman [43] is adopted to create person models with different identities by assigning randomly different values to predefined attributes, such as gender, height, age, body shape, skin color, hair, clothing, etc. We have created 5,000 pedestrians for subsequent operations, and some models have been shown in Fig. 3a." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.886 + ], + "angle": 0, + "content": "Second, Unity3D [44] is utilized to simulate a real-world streetscape. As shown in Fig. 3b, we pick three blocks to deploy the camera network, where a ground camera network (\"G-Cam\" area) serves relatively convenient areas such as downtown, and two aerial camera networks (\"A-Cam\" area) serve the suburbs of a city. In the \"G-Cam\" area, we deploy eight fixed cameras at each street corner. In the \"A-Cam\" area, we design different drone roaming strategies based on the size of the surveillance area. For the small area (left area), we deploy one drone with a \\(90^{\\circ}\\) overhead view, allowing it to move counterclockwise around each street. For a large area (right area), we deploy individual drones on each of the four streets with a \\(45^{\\circ} \\sim 60^{\\circ}\\) tilt view, allowing them to move back and forth on corresponding streets." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Finally, we import human models into the streetscape" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.895, + 0.459 + ], + "angle": 0, + "content": "and further move them under the ground and aerial networks. We allow multiple pedestrians to move simultaneously to simulate the random occlusion among pedestrians in the real world. Meanwhile, we collect valid images of the target persons if they appear in cameras without severe occlusion (train \\(\\leq 70\\%\\), and test \\(\\leq 75\\%\\)). In total, CARGO contains 108,563 person images from 13 cameras." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.615, + 0.485 + ], + "angle": 0, + "content": "4.3. Challenge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.691 + ], + "angle": 0, + "content": "Fig. 3c illustrates abundant challenges prevalent in real-world scenarios in the CARGO dataset. The most significant challenge is the view discrepancy from different types of cameras, which produces a vast variation and rotation of person postures. Besides, Fig. 3c and Tab. 1 show another change in the camera height, from 5 meters for the lowest ground camera to 75 meters for the highest \\(90^{\\circ}\\) view drone, resulting in a huge resolution variation. The maximum resolution in CARGO is \\(1009 \\times 539\\), while the minimum is only \\(22 \\times 7\\). Moreover, CARGO also contains other complex illumination and occlusion variations, etc. Overall, CARGO simulates and covers major real-world scenarios and challenges and can be considered a large-scale benchmark." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.704, + 0.625, + 0.722 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.895, + 0.777 + ], + "angle": 0, + "content": "Importantly, please refer to the supplementary material for detailed experiment settings, visualization of retrieval and feature, cross-dataset evaluation, and other discussions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.686, + 0.802 + ], + "angle": 0, + "content": "5.1. Dataset and Metric" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dataset. Two datasets are adopted for evaluation, CARGO and AG-ReID [22]. Some statistics regarding these datasets are shown in Tab. 1. As for CARGO, we split it into the train (51,451 images with 2500 IDs) and test sets (51,024 images with the remaining 2500 IDs) with an almost 1:1 ratio. In the training set, each person has an average of 10 images" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22005" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.095, + 0.332, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.248, + 0.294, + 0.26 + ], + "angle": 0, + "content": "(a) CARGO under Protocol 1." + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.096, + 0.603, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.248, + 0.568, + 0.26 + ], + "angle": 0, + "content": "(b) AG-ReID under Protocol 1." + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.096, + 0.874, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.248, + 0.838, + 0.26 + ], + "angle": 0, + "content": "(c) AG-RelD under Protocol 2." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Fig. \\(4\\mathrm{a}\\sim\\) Fig. 4c show the ablation experiments about orthogonal decoupling of view-related and view-unrelated features in the VDT on two dataset, which consists of two important parts, Eq. (4) and Eq. (9). Rank1, mAP, and mINP are reported \\((\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.599 + ], + "angle": 0, + "content": "and on average moves across 6.47 ground and 3.61 aerial cameras. In the test set, we select 149 IDs as the query set, and all the remaining IDs are placed in the gallery as distractors. For each identity selected as a query, we keep only its images from two random cameras and use the image from one camera as the query and the image from the other as the gallery. On this basis, we design four protocols (ALL, \\(\\mathrm{A}\\leftrightarrow \\mathrm{A}\\), \\(\\mathrm{G}\\leftrightarrow \\mathrm{G}\\), and \\(\\mathrm{A}\\leftrightarrow \\mathrm{G}\\)) to adequately evaluate the model performances, where \"ALL\" focuses on the comprehensive retrieval performance, and the latter three focus on some specific retrieval patterns. \\(\\mathrm{A}\\leftrightarrow \\mathrm{A}\\) only retains the data under the aerial camera in the test set for evaluation (60 query IDs with 134 images, 2404 gallery IDs with 18,444 images), same for \\(\\mathrm{G}\\leftrightarrow \\mathrm{G}\\) (89 query IDs with 178 images, 2447 gallery IDs with 32,268 images). \\(\\mathrm{A}\\leftrightarrow \\mathrm{G}\\) relabels the original test set into two domains (aerial and ground domain) based on the view label. The training set of all testing protocols retains same." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.75 + ], + "angle": 0, + "content": "As for AG-ReID, there are 11,554 images with 199 IDs for training and 12,464 images with 189 IDs for testing following [22]. The test set contains 2,033 query images and 10,429 gallery images. There are two protocols in AG-ReID, denoted as \\(\\mathrm{A}\\rightarrow \\mathrm{G}\\) and \\(\\mathrm{G}\\rightarrow \\mathrm{A}\\), where the former contains 1,701 aerial query images and 3,331 ground gallery images for 189 identities, and the latter contains 962 ground query images and 7,204 aerial gallery images for 189 identities. Optionally, AG-ReID also provides detailed person attribute labeling as additional prior information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Metric. The cumulative matching characteristic at Rank1, mean Average Precision (mAP), and mean Inverse Negative Penalty (mINP) [21] are adopted as the evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.214, + 0.817 + ], + "angle": 0, + "content": "5.2. Performance" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Accuracy. We compare the proposed VDT with other competitive methods on CARGO in Tab. 2, including CNN-based (BoT [38], SBS [36], VV [40, 41], MGN [39], AGW [21]) and transformer-based methods (ViT [34]), where VV means a vehicle ReID baseline with vanilla batch-based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.782 + ], + "angle": 0, + "content": "triplet loss and diverse viewpoint batch selection. Similar experimental results on AG-ReID are reported in Tab. 3, where Explain [22] relies on pedestrian attribute priors and therefore does not generalize well to CARGO. Several performances in Tab. 3 follow the results reported in [22]. We can draw the following conclusions: (1) The proposed VDT achieves the state-of-the-art performances that are clearly improved over the baseline, especially in heterogeneous matching. For example, VDT surpasses the mAP/Rank1/mINP baseline by \\(4.99\\% / 2.65\\% / 1.75\\%\\) on the \\(\\mathrm{A} \\leftrightarrow \\mathrm{G}\\) of CARGO. Besides, VDT also brings different degree of benefits to other CARGO protocols. VDT also surpasses Explain on mAP/Rank1 by \\(3.74\\% / 5.18\\%\\) on the \\(\\mathrm{G} \\rightarrow \\mathrm{A}\\) of AG-ReID, which strongly demonstrates the importance of view decoupled operations in AGPReID to mitigate the disruption of identity representation by view bias whether in homogeneous or heterogeneous matching. (2) Previous view-homogeneous methods show varying degrees of performance degradation in AGPReID, especially on view-heterogeneous protocols, indicating that view bias causes the identity feature under different views to behave with poor cohesion. However, the above methods ignore this issue. (3) The superior performance of VDT does not come from a strong baseline but from the proposed method itself. On the CARGO dataset, the baseline does not consistently outperform the CNN-based AGW on all metrics; on the AGPReID dataset, the baseline is also weaker than the previous Explain. However, VDT achieves the SOTA performance on two datasets, especially on AG-ReID, using fewer prior labels and achieving better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Speed. We report the inference time for baseline and VDT for retrieving a single image, which is 2.27ms and 2.28ms on the CARGO dataset and 1.44ms and 1.46ms on the AGReID dataset, respectively. These results validate the previous complexity analysis in Sec. 3.2 and show that VDT does not add additional time consumption. Slight time fluctuations tend to be acceptable, due to hardware devices." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22006" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.095, + 0.35, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.248, + 0.294, + 0.26 + ], + "angle": 0, + "content": "(a) CARGO under Protocol 1." + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.096, + 0.621, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.248, + 0.568, + 0.26 + ], + "angle": 0, + "content": "(b) AG-ReID under Protocol 1." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.096, + 0.892, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.248, + 0.838, + 0.26 + ], + "angle": 0, + "content": "(c) AG-ReID under Protocol 2." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Fig. \\(5\\mathrm{a}\\sim\\) Fig. 5c show the relationship between \\(\\lambda\\) and performance on two datasets. For simplicity, only protocol 1 is shown on the CARGO dataset. Rank1, mAP, and mINP are reported (\\%). Avg represents the average performance of Rank1, mAP, and mINP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.327, + 0.23, + 0.343 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.472 + ], + "angle": 0, + "content": "As described in Sec. 3.2, the core contribution of VDT lies in the decoupling of identity and view features by Eq. (4) and Eq. (9). Therefore, we aim to demonstrate the contribution of each part to VDT, which has been shown in Fig. 4. Similar trends in the two datasets demonstrate that the absence of each part leads to a degradation of the model's performance, and the absence of both simultaneously leads to a more severe performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.639 + ], + "angle": 0, + "content": "By comparing Row1 and Row2, the absence of Eq. (4) leads to the degradation of \\( t_m \\) and \\( t_v \\) to be equivalent, focusing on the identity and view features in the image respectively, and constraining them to be independent through Eq. (9). However, lacking the guidance of explicit separation within VDT, the decoupling goal cannot be effectively achieved through Eq. (9) alone. Similarly, by comparing Row1 and Row3, the identity and view features outputted by VDT are not guaranteed to be orthogonally separated from each other if Eq. (9) is missing, and thus \\( t_m \\) containing a minor view bias is still not strongly discriminative." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.265, + 0.667 + ], + "angle": 0, + "content": "5.4. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.469, + 0.825 + ], + "angle": 0, + "content": "As described in Sec. 3.2, VDT introduces only one hyperparameter \\(\\lambda\\) in Eq. (10) that needs to be manually tuned for multi-task balance. We select six magnitudes of \\(\\lambda\\) from \\(1e - 4\\) to 10 and the corresponding results are shown in Fig. 5. To show a clear tendency, we show the average performance of Rank1, mAP, and mINP at each \\(\\lambda\\). Although the performance shows the same trend of increasing and then decreasing on both datasets as \\(\\lambda\\) increases, the peak performance is achieved when \\(\\lambda = 0.001\\) on CARGO (Fig. 5a) and \\(\\lambda = 1.0\\) on AG-ReID (Fig. 5b and Fig. 5c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The key of this phenomenon is the synchronous updating between identity and view classifier, because the view classifier is a simple binary classification, but the identity classifier on the two datasets shows different difficulties depending on the number of identities. As shown in Tab. 1," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.449 + ], + "angle": 0, + "content": "CARGO owns a thousand-level identities, so the identity classification is relatively complicated, and in order to keep tasks in Eq. (10) able to update synchronously, a small \\(\\lambda\\) is more appropriate for view-related objectives. Similarly, a large \\(\\lambda\\) is selected because of a hundred-level identities in AGPReID. We also provide an empirical \\(\\lambda\\) selection guideline to avoid exhaustive search based on above analysis: \\(\\lambda\\) tends to be smaller for datasets with larger identity scale." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.462, + 0.769, + 0.478 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.653 + ], + "angle": 0, + "content": "This paper focuses on ReID under a view-heterogeneous scenario, aerial-ground ReID (AGPReID). First, the view-decoupled transformer (VDT) is proposed specifically for the dramatic view differences in AGPReID, which decouples view-related and view-unrelated features by two major operations: hierarchically subtractive separation and orthogonal loss. Second, we contribute a large-scale dataset called CARGO, in which the identities and images are 12.8 and 4.9 times larger than the previous dataset. Experiments on two datasets demonstrate the superiority of VDT and the necessity of view decoupling in AGPReID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Although view discrepancy is the most significant factor in AGPReID, it is not the only source of disruptions. Therefore, further exploration of discriminative identity representations under multiple disturbances in the aerial-ground camera network is a promising avenue of future research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.659, + 0.759 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.768, + 0.892, + 0.859 + ], + "angle": 0, + "content": "This project was supported in part by the NSFC (U22A2095, 62076258), in part by the Key-Area Research and Development Program of Guangzhou (202206030003), in part by Guangdong Project (No. 2020B1515120085), and in part by International Program Fund for Young Talent Scientific Research People, Sun Yat-Sen University." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22007" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Geon Lee, Sanghoon Lee, Dohyung Kim, Younghoon Shin, Yongsang Yoon, and Bumsub Ham. Camera-driven representation learning for unsupervised domain adaptive person re-identification. In Int. Conf. Comput. Vis., pages 11453–11462, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In ACM Int. Conf. Multimedia, page 3115–3123, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Quan Zhang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling with second-order transformer for group re-identification. In AAAI, volume 36, pages 3318–3325, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] Quan Zhang, Jianhuang Lai, Xiaohua Xie, Xiaofeng Jin, and Sien Huang. Separable spatial-temporal residual graph for cloth-changing group re-identification. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-16, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Quan Zhang, Kaiheng Dang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Modeling 3d layout for group re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7512-7520, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[6] Jianxiong Tang, Jianhuang Lai, Xiaohua Xie, Lingxiao Yang, and Wei-Shi Zheng. Ac2as: Activation consistency coupled ann-snn framework for fast and memory-efficient snn training. Pattern Recognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Qi Chen, Yun Chen, Yuheng Huang, Xiaohua Xie, and Lingxiao Yang. Region-based online selective examination for weakly supervised semantic segmentation. Information Fusion, page 102311, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7713-7722, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Junyang Qiu, Zhanxiang Feng, Lei Wang, and Jianhuang Lai. Salient part-aligned and keypoint disentangling transformer for person re-identification in aerial imagery. In Int. Conf. Multimedia and Expo, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[10] Pengze Zhang, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Pose guided person image generation via dual-task correlation and affinity learning. IEEE Trans. Vis. Comput. Graph., pages 1-18, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[11] Jianxiong Tang, Jian-Huang Lai, Xiaohua Xie, and Lingxiao Yang. Spike count maximization for neuromorphic vision recognition. In *IJCAI*, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[12] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Adv. Neural Inform. Process. Syst., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.817 + ], + "angle": 0, + "content": "[13] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4288-4298, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[14] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling for group re-identification. Int. J. Comput. Vis., 2024." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Lingxiao Yang, Ru-Yuan Zhang, Lida Li, and Xiaohua Xie. Simam: A simple, parameter-free attention module for convolutional neural networks. pages 11863-11874. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[17] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Int. Conf. Comput. Vis., pages 1116–1124, 2015. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[18] Shizhou Zhang, Qi Zhang, Yifei Yang, Xing Wei, Peng Wang, Bingliang Jiao, and Yanning Zhang. Person re-identification in aerial imagery. IEEE Trans. Multimedia, 23:281-291, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[19] Quan Zhang, Jianhuang Lai, and Xiaohua Xie. Learning modal-invariant angular metric by cyclic projection network for vis-nir person re-identification. IEEE Trans. Image Process., 30:8019-8033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[20] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Seeing like a human: Asynchronous learning with dynamic progressive refinement for person re-identification. IEEE Trans. Image Process., 31:352-365, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[21] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44(6):2872-2893, 2022. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[22] H. Nguyen, K. Nguyen, S. Sridharan, and C. Fookes. Aerial-ground person re-id. In Int. Conf. Multimedia and Expo, pages 2585-2590, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[23] Lei Wang, Quan Zhang, Junyang Qiu, and Jianhuang Lai. Rotation exploration transformer for aerial person re-identification. In Int. Conf. Multimedia and Expo, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[24] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11506–11515, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[25] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 79–88, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[26] Fei Xiong, Mengran Gou, Octavia Camps, and Mario Sznaier. Person re-identification using kernel-based metric learning methods. In Eur. Conf. Comput. Vis., pages 1–16, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[27] Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. Person re-identification by local maximal occurrence representation and metric learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2197–2206, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[28] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In Eur. Conf. Comput. Vis., pages 501-518, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[29] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Int. Conf. Comput. Vis., pages 15013-15022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22008" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "In IEEE Conf. Comput. Vis. Pattern Recog., pages 16266-16275, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[31] Shuoyi Chen, Mang Ye, and Bo Du. Rotation invariant transformer for recognizing object in uavs. In ACM Int. Conf. Multimedia, page 2565-2574, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[32] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In ACM Int. Conf. Multimedia, page 3422–3430, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[33] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4890–4899, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[34] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2021. 3, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[36] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. In ACM Int. Conf. Multimedia, page 9664–9667, 2023. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[37] Yifan Sun, Liang Zheng, Yali Li, Yi Yang, Qi Tian, and Shengjin Wang. Learning part-based convolutional features for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 43:902–917, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[38] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[39] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM Int. Conf. Multimedia, pages 274–282, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[40] Ratnesh Kuma, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. Vehicle re-identification: an efficient baseline using triplet embedding. In International Joint Conference on Neural Networks, pages 1–9, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[41] Ratnesh Kumar, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. A strong and efficient baseline for vehicle re-identification using deep triplet embedding. Journal of Artificial Intelligence and Soft Computing Research, 10(1):27-45, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[42] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Learning generalisable omni-scale representations for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):5056–5069, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[43] Leyde Briceno and Gunther Paul. Makehuman: a review of the modelling framework. In Congress of the International Ergonomics Association, pages 224-232, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[44] Unity Technologies. Unity3D: Cross-platform 3D engine, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[45] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248–255, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[46] Léon Bottou. Stochastic gradient descent tricks. Neural Networks: Tricks of the Trade: Second Edition, pages 421-436, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., pages 8024-8035, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22009" + } + ] +] \ No newline at end of file diff --git a/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_origin.pdf b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..68f63c6b3abed89b47507b892227759fe5de6542 --- /dev/null +++ b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/66ee0e69-7b19-441f-b54a-195f4cb4500a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:939819251e7bf2e4db1178927b1e780a38d59024c0b434eea2ff29a921b38e1c +size 1174527 diff --git a/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/full.md b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b2af5f23d2361d0f9500cfef17f3fdd96c79193 --- /dev/null +++ b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/full.md @@ -0,0 +1,317 @@ +# View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network + +Quan Zhang $^{1,5}$ Lei Wang $^{1}$ Vishal M. Patel $^{5}$ Xiaohua Xie $^{1,2,3,4}$ Jianhuang Lai $^{1,2,3,4*}$ $^{1}$ School of Computer Science and Engineering, Sun Yat-Sen University, China + $^{2}$ Pazhou Lab (HuangPu), Guangdong, China + +3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China +4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China +5Department of Electrical and Computer Engineering, Johns Hopkins University, USA +{zhangq48, wanglei75}@mail2.sysu.edu.cn, {stsljh, xiexiaoh6}@mail.sysu.edu.cn, vpatel36@jhu.edu + +# Abstract + +Existing person re-identification methods have achieved remarkable advances in appearance-based identity association across homogeneous cameras, such as ground-ground matching. However, as a more practical scenario, aerial-ground person re-identification (AGPReID) among heterogeneous cameras has received minimal attention. To alleviate the disruption of discriminative identity representation by dramatic view discrepancy as the most significant challenge in AGPReID, the view-decoupled transformer (VDT) is proposed as a simple yet effective framework. Two major components are designed in VDT to decouple view-related and view-unrelated features, namely hierarchical subtractive separation and orthogonal loss, where the former separates these two features inside the VDT, and the latter constrains these two to be independent. In addition, we contribute a large-scale AGPReID dataset called CARGO, consisting of five/eight aerial/ground cameras, 5,000 identities, and 108,563 images. Experiments on two datasets show that VDT is a feasible and effective solution for AGPReID, surpassing the previous method on mAP/Rank1 by up to $5.0\% /2.7\%$ on CARGO and $3.7\% /5.2\%$ on AG-ReID, keeping the same magnitude of computational complexity. Our project is available at https://github.com/LinlyAC/VDT-AGPReID. + +# 1. Introduction + +Person re-identification (ReID) aims to associate interested person images based on cross-camera identity similarity, which plays a significant and positive contribution to the se + +![](images/e500db2cf3afb8a177a92c40844fda3cf765402699f37bfe6096491cb8b918cd.jpg) +Figure 1. View-homogeneous vs. view-heterogeneous ReID, where the former focuses on ground-only or aerial-only camera networks, and the latter considers the aerial-ground mixed camera network. Thus, view-heterogeneous ReID considers aerial-aerial, ground-ground, and aerial-ground matching, which is more challenging and practical than the existing view-homogeneous ReID. + +curity and safety of society and citizens [1-5]. Traditional ReID has achieved remarkable progress in the deep learning era [6-16], but it is far from real-world scenarios for the following reasons. (1) Most existing ReID datasets are collected from homogeneous camera networks, which comprises of the same type cameras, such as ground-only [17] or aerial-only [18] camera networks. However, as can be seen from Fig. 1, real-world surveillance systems tend to be deployed as heterogeneous camera networks comprising + +aerial and ground cameras instead of a single type alone. Ground cameras cover well-developed areas (city centers), and aerial cameras cover poorly-developed areas (suburbs) due to their broad view range. The complementary nature of the two types would maximize the effectiveness of ReID. (2) Existing ReID methods mainly consider homogeneous matching (ground-ground [19-21] and aerial-aerial [9, 22, 23]), which is ineffective in dealing with the dramatic view discrepancy among heterogeneous matching (aerial-ground). Despite its significance, related research is extremely scarce. + +In this paper, we consider the novel and practical view-heterogeneous ReID problem, specifically ReID under the aerial-ground camera network (AGPReID). We propose a view-decoupled transformer (VDT) to specifically tackle the dramatic view discrepancy, which serves as a significant challenge within AGPReID to hinder homogeneous and heterogeneous matchings. Motivated by the fact that view-related features are useless for discriminative identity representations, VDT aims to decouple view-related and view-unrelated components, further facilitating discriminative identity learning from the remaining view-unrelated features. There are two key parts within VDT to achieve view decoupling, namely (a) hierarchically subtractive separation and (b) orthogonal loss. Specifically, after tokenizing the input image to a series of patch tokens, VDT appends two extra tokens (the meta token and view token) to feed them into a transformer stacked by multiple VDT blocks. The meta token captures global representation in the image, and the view token aims to extract view-related features. In each VDT block, (a) is achieved by the subtractive values of the meta token and view token after self-attention operations as the updated meta token for the next block. This means that VDT hierarchically separates view-related features from global features to facilitate identity learning from remaining view-unrelated features. When the meta and view tokens are obtained after the last block, (b) is designed to constrain the identity features to be independent of the view features, ultimately achieving the orthogonal decoupling of view-related and view-unrelated components. Meanwhile, the meta and view tokens will be supervised by identity and view labels, respectively. + +Considering the scarcity [22] and privacy [5, 24] of the AGPReID datasets, we contribute a large-scale Civic AeRial-GrOund (CARGO) dataset, which is collected from a synthetic city scenario including five aerial and eight ground cameras. CARGO has totally collected 5,000 identities and 108,563 images, which contain not only significant view discrepancy but also plenty of variations of resolution, illumination, occlusion, etc. Regarding evaluation, CARGO considers multiple patterns (aerial-aerial, ground-ground, aerial-ground matchings), which could serve as a comprehensive benchmark. Contributions of this paper can + +be summarized as follows: + +- We focus on ReID under the aerial-ground camera network (AGPReID) scenario and propose a view-decoupled transformer (VDT) specifically for the dramatic view discrepancy within AGPReID, which achieves view-related and view-unrelated features decoupling by hierarchical subtractive separation and orthogonal loss. +- We contribute a large-scale synthetic dataset, called CARGO, to advance the AGPReID as a benchmark, which contains five/eight aerial/ground cameras, 5,000 identities, 108,563 images, and complex challenges. +- Experiments demonstrate the superiority of VDT on two datasets, which shows that VDT surpasses the previous method on mAP/Rank1 by up to $5.0\% / 2.7\%$ on CARGO and $3.7\% / 5.2\%$ on AG-ReID, respectively, maintaining the same magnitude of computational complexity. + +# 2. Related Work + +# 2.1. View-homogeneous ReID + +View-homogeneous ReID deals with images from ground-only or aerial-only camera networks, where the ground-only camera network receives the broadest attention in the ReID task because it is the most common use case. Many ground-only datasets have been contributed in the literature, such as Market1501 [17], MSMT17 [25], etc. Accordingly, lots of methods have been proposed to bring rapid development to ReID, including handcrafted feature-based methods [26, 27], CNN-based methods [19-21, 28], and transformer-based methods [29]. Regarding aerial-only research, only some pioneering efforts have contributed substantial datasets and methods [18, 30, 31] because it has not received as much attention as ground-only ReID. + +However, it is difficult to directly transfer the above methods to the view-heterogeneous ReID because these methods ignore the significant view differences between aerial and ground cameras, leading to weak performances. + +# 2.2. View-heterogeneous ReID + +Compared to the abundant research in Sec. 2.1, studies on view-heterogeneous ReID are almost nonexistent. In this paper, heterogeneous views refer to AGPReID, and the fundamental challenge is the dramatic view discrepancy. To the best of our knowledge, one recent work [22] has formally attempted to address this task. They collected an outdoor scene dataset, AG-ReID, with both identity and attribute labeling of pedestrians. Moreover, they proposed an explainable model by attribute information to guide model training. + +While pioneering and critical to AGPReID, this work is still imperfect for the following reasons. (1) Dataset. Compared to [22], our CARGO dataset comprehensively focuses on multiple matching patterns that may appear in heterogeneous views rather than a single aerial-ground matching in + +![](images/4fe5321429469995b02bae8316f7a9290adee4613e4f0ab4a79e6a757f85c5a2.jpg) +Figure 2. Illustration of the proposed VDT framework, which consists of $N$ VDT blocks and three-part loss functions. Meta and view tokens capture global and view-related features in images, respectively. Each VDT block (light blue module) consists of a standard self-attention encoder layer and an inner feature subtraction operation, achieving layer-by-layer decoupling of view-related and view-unrelated features. Orthogonal loss constrains the above two features to be further independent. + +AG-ReID. (2) Method. Compared to explainable methods in [22], our VDT relies on a lesser amount of priori labeling and thus has a stronger generalization. The experiments also show that our VDT achieves stronger performances. + +# 2.3. Synthetic ReID Dataset + +Data privacy is an inevitable issue for the ReID datasets, and synthetic data provides a feasible solution because it not only achieves privacy protection for real-world pedestrians but also has lower collection costs. Many synthetic ReID datasets have been published in the literature, such as PersonX [14], UnrealPerson [24], WePerson [2], etc. + +However, existing synthetic datasets mainly focus on simulating pedestrian data under view-homogeneous camera networks, and the simulation of aerial-ground views is poorly studied. To this end, we contribute the CARGO dataset as the first large-scale synthetic dataset for AGPReID. + +# 3. Method + +# 3.1. Formulation and Overview + +Formulation. An AGPReLU dataset $\mathcal{D} = \{(x_i, y_i, v_i)\}_{i=1}^{|\mathcal{D}|}$ consists of a training set $\mathcal{D}^{tr}$ and a test set $\mathcal{D}^{te}$ , where $x_i$ represents the $i$ -th person image, and $y_i$ and $v_i$ stand for the corresponding identity and view labels, respectively. Note that $v_i \in \{v^a, v^g\}$ is readily acquired by the known camera labels in $\mathcal{D}$ to easily distinguish whether $x_i$ belongs to an aerial $v^a$ or ground $v^g$ view. Significant view discrepancy between $v^a$ and $v^g$ leads to a view-biased feature + +space, which behaves as low intra-identity and high interidentity similarity. Therefore, the goal of AGPReID methods is to design a model $\mathcal{F}(\cdot;\theta_{\mathcal{F}})$ with learnable parameter $\theta_{\mathcal{F}}$ against the view bias, which can be written as: + +$$ +\min _ {\theta} \sum_ {i} \left[ \begin{array}{c} \left\| \mathcal {F} \left(x _ {i}; \theta_ {\mathcal {F}}\right) - \mathcal {F} \left(x _ {i} ^ {+}; \theta_ {\mathcal {F}}\right) \right\| _ {2} ^ {2} - \\ \left\| \mathcal {F} \left(x _ {i}; \theta_ {\mathcal {F}}\right) - \mathcal {F} \left(x _ {i} ^ {-}; \theta_ {\mathcal {F}}\right) \right\| _ {2} ^ {2} \end{array} \right], \tag {1} +$$ + +where $\| \cdot \| _2$ denotes the L2 distance, $x_{i}^{+} / x_{i}^{-}$ denotes the person image that has the same/different identity with $x_{i}$ . + +Overview. As shown in Fig. 2, we design a simple yet effective framework called view-decoupled transformer (VDT), to tackle the view discrepancy challenge in AGPReID. For a batched data $B$ that contains both $v^a$ and $v^g$ view, weckenize them as a series of image tokens (green squares), then append a meta and view token (blue and yellow squares) to them, which serve as input to VDT. The VDT network consists of $N$ blocks, where each block first performs the standard self-attention encoding and then performs a subtraction operation between meta and view tokens to explicitly separate the view-related feature from the global feature. Finally, the identity and view classifier will supervise the updated meta and view tokens outputted from VDT. Besides, an orthogonal loss is proposed that makes these two tokens irrelevant, thus achieving the complete decoupling of view-related and view-unrelated features. + +# 3.2. View-decoupled Transformer + +The proposed VDT framework is based on Vit-Base [34]. For each image $x_{i}$ in the batched data $B$ , VDT first + +Table 1. Statistical comparisons of existing datasets, including view-homogeneous (ground or aerial) and view-heterogeneous (ground and aerial) ReID. “ $\star$ ” denotes City1M is used for group ReID, where its images correspond to the number of group images. + +
DatasetViewData#PersonID#Camera#Image#Height
Market1501 [17]GroundReal1,501632,668< 10m
PersonX [14]GroundSynthetic1,2666273,456-
RandPerson [32]GroundSynthetic8,00019228,655-
UnrealPerson [24]GroundSynthetic3,00034120,000-
WePerson [2]GroundSynthetic1,500404,000,000-
ClonedPerson [33]GroundSynthetic5,62124887,766-
City1M* [5]GroundSynthetic45,00081,840,000-
PRAI1581 [18]AerialReal1,581239,46120 ~ 60m
UAVHuman [30]AerialReal1,144141,2902 ~ 8m
AG-ReID [22]Aerial-GroundReal3882 (1A+1G)21,89315 ~ 45m
CARGO (Ours)Aerial-GroundSynthetic5,00013 (5A+8G)108,5635 ~ 75m
+ +divides $x_{i}$ evenly and non-overlappingly into $M$ image patches, which are further tokenized into $M$ patch embeddings by a $1 \times 1$ convolution and can be denoted as $[t_{p,1}; t_{p,2}; \dots; t_{p,M}]$ . After that, we append two additional learnable tokens, meta token $t_{m}$ and view token $t_{v}$ , in the patch tokens, where $t_{m}$ aims to capture the global image representation, $t_{v}$ focuses only on the view-related feature. Next, we serialize all the tokens by assigning them the corresponding positional embedding, which can be written as: + +$$ +\mathcal {S} (x _ {i}) \triangleq \left[ t _ {m}; t _ {p, 1}; t _ {p, 2}; \dots ; t _ {p, M}; t _ {v} \right] + \mathbf {E} _ {p o s}, \tag {2} +$$ + +where $\mathcal{S}(\cdot)$ is defined as the serialization operation for $x_{i}$ , $t_m, t_{p,i}, t_v \in \mathbb{R}^d$ , and $\mathbf{E}_{pos} \in \mathbb{R}^{(M + 2) \times d}$ . + +As shown in Fig. 2, VDT $\mathcal{F}(\cdot ;\theta_{\mathcal{F}})$ is stacked with $N$ VDT blocks $\mathcal{F}_j(\cdot ;\theta_{\mathcal{F}}^j),1\leq j\leq N.$ Each $\mathcal{F}_j(\cdot ;\theta_{\mathcal{F}}^j)$ contains two operations, self-attention encoding and subtractive separation, which are described as: + +$$ +\left[ t _ {m} ^ {(j + 1)} \left(x _ {i}\right); \dots ; t _ {v} ^ {(j + 1)} \left(x _ {i}\right) \right] = \theta_ {\mathcal {F}} ^ {(j)} \left(\left[ t _ {m} ^ {(j)} \left(x _ {i}\right); \dots ; t _ {v} ^ {(j)} \left(x _ {i}\right) \right]\right), \tag {3} +$$ + +$$ +t _ {m} ^ {(j + 1)} \left(x _ {i}\right) \leftarrow t _ {m} ^ {(j + 1)} \left(x _ {i}\right) - t _ {v} ^ {(j + 1)} \left(x _ {i}\right), \tag {4} +$$ + +where $\theta_{\mathcal{F}}^{(j)}$ represents the self-attention parameters in $\mathcal{F}_j$ . If $j = 1$ , the input of $\mathcal{F}_1$ becomes $\mathcal{S}(x_i)$ . Note that $t_m^{(j + 1)}(x_i)$ and $t_v^{(j + 1)}(x_i)$ in Eq. (3) represent the updated meta and view token after the self-attention operation in $\mathcal{F}_j$ with patch tokens of $x_i$ . Eq. (4) explicitly guides that $t_m^{(j + 1)}$ and $t_v^{(j + 1)}$ are distinguishable after Eq. (3), where $t_v^{(j + 1)}$ captures more view-related features and $t_m^{(j + 1)}$ can capture more view-unrelated features by removing $t_v^{(j + 1)}$ from the global (meta) features. Eq. (4) facilitates the learning of identity features via $t_m^{(j + 1)}$ , alleviating the disturbing of view bias. VDT decouples the view-related and -unrelated features layer by layer until the updated $t_m^{(N + 1)}$ and $t_v^{(N + 1)}$ are obtained, which serves as the well-decoupled identity + +and view features, respectively. During inference, we only use $t_m^{(N + 1)}$ as the whole retrieval evidence. + +In addition, the proposed VDT maintains the same computational complexity as baseline (ViT). Given that the input to ViT has $T$ tokens and the dimensions of the tokens are $d$ , then the computational complexity of ViT is $\mathcal{O}\big(4Td^2 + 2T^2d\big)$ [35]. Our VDT adds an additional $t_v$ token based on the baseline and adds an Eq. (4) operation in each VDT block. Therefore, the total computational complexity can be written as: + +$$ +\begin{array}{l} \mathcal {O} \left(4 (T + 1) d ^ {2} + 2 (T + 1) ^ {2} d + N\right) \\ \Leftrightarrow \mathcal {O} \left(\underbrace {4 T d ^ {2} + 2 T ^ {2} d} _ {\text {b a s e l i n e c o m p l e x i t y}} + \underbrace {4 d ^ {2} + 2 d + N} _ {\text {c o n s t a n t c o m p l e x i t y}}\right) \tag {5} \\ \Leftrightarrow \mathcal {O} \left(4 T d ^ {2} + 2 T ^ {2} d\right), \\ \end{array} +$$ + +where $N$ counts all Eq. (4) in $N$ VDT blocks. Eq. (5) shows that the extra operations VDT introduces depend on only the network hyperparameters, which belong to constant complexity and can be omitted. Therefore, VDT maintains the same magnitude of complexity as that of the baseline. + +# 3.3. Optimization + +To meet the task goal (Eq. (1)), the objective of $\mathcal{F}(\cdot, \theta_{\mathcal{F}})$ consists of three components: identity classifier, view classifier, and orthogonal loss. + +For each $x_{i} \in B$ , the identity classifier utilizes both the cross-entropy loss and the triplet loss to supervise $t_m^{N + 1}(x_i)$ , which can be written as: + +$$ +\mathcal {L} _ {i} ^ {c} = \frac {1}{| B |} \sum_ {i = 1} ^ {| B |} y _ {i} \log (\hat {y} _ {i}), \tag {6} +$$ + +where $|B|$ represents the batch size, $y_{i}$ represents the identity label, and $\hat{y}_i$ denotes the identity prediction by + +![](images/c66d843dc5e9a3cdded5ccb45c918918a0cc7a8b34ffbd02e4e7d44138c5c161.jpg) +(a) 3D pedestrian models. + +![](images/099d26879c3c1fbd25b09cf92162eda3e37660735b02d8ba6f33ed913a23e95a.jpg) +(b) Aerial-ground camera network deployment. + +![](images/043a0a48f6b815686e3bac3f1b6ac828f467c54e999663cbffbb73da92dfc402.jpg) +(c) Diversity challenges. +Figure 3. Fig. $3\mathrm{a}\sim$ Fig. 3c shows the pedestrian models, camera deployment, and challenges during the CARGO construction, respectively. In Fig. 3b, "A-Cam" and "G-Cam" represent the aerial and ground cameras, where the yellow sectors represent the view range of the ground cameras, and the green arrows represent the motion strategy of the aerial cameras. The challenges displayed in Fig. 3c are view variation, illumination variation, occlusion, and resolution variation from top to bottom. + +$t_m^{N + 1}(x_i)$ . The triplet loss is defined as follows: + +$$ +\mathcal {L} _ {i} ^ {t} = \frac {1}{| B |} \sum_ {i = 1} ^ {| B |} \left[ \begin{array}{c} \left\| t _ {m} ^ {N + 1} \left(x _ {i}\right) - t _ {m} ^ {N + 1} \left(x _ {i} ^ {+}\right) \right\| _ {2} ^ {2} - \\ \left\| t _ {m} ^ {N + 1} \left(x _ {i}\right) - t _ {m} ^ {N + 1} \left(x _ {i} ^ {-}\right) \right\| _ {2} ^ {2} + \alpha \end{array} \right] _ {+}, \tag {7} +$$ + +where $x_{i}^{+} / x_{i}^{-}$ denotes the hard positive/negative sample of $x_{i}$ in the current $B$ , $[\cdot]_{+}$ stands for $\max(\cdot, 0)$ , $\alpha$ is the hyperparameter of the margin. + +View classifier supervises $t_v^{N + 1}(x_i)$ with only the cross-entropy loss, which can be written as: + +$$ +\mathcal {L} _ {v} ^ {c} = \frac {1}{| B |} \sum_ {i = 1} ^ {| B |} v _ {i} \log (\hat {v} _ {i}), \tag {8} +$$ + +where $v_{i}$ represents the view label of $x_{i}$ , and $\hat{v}_{i}$ denotes the view prediction by $t_v^{N + 1}(x_i)$ . + +To enable $t_m^{N + 1}(x_i)$ and $t_v^{N + 1}(x_i)$ to be truly decoupled, we impose an orthogonality loss that makes them less similar to each other, which can be written as: + +$$ +\mathcal {L} _ {o} = \frac {1}{| B |} \sum_ {i = 1} ^ {| B |} \frac {\left| \left\langle t _ {m} ^ {N + 1} \left(x _ {i}\right) , t _ {v} ^ {N + 1} \left(x _ {i}\right) \right\rangle \right|}{\left\| t _ {m} ^ {N + 1} \left(x _ {i}\right) \right\| _ {2} \cdot \left\| t _ {v} ^ {N + 1} \left(x _ {i}\right) \right\| _ {2}}, \tag {9} +$$ + +where $|\langle \cdot ,\cdot \rangle |$ represents the absolute value after the dot product of two token embeddings. We also provide a clear illustration of Eq. (9) in Fig. 2 (the gray rectangle in the middle on the right). Please note that the dashed blue and the yellow vector represent the meta token and view token outputted by Eq. (3) in $\mathcal{F}_N$ respectively, and the blue vector represents the result of Eq. (4), which is the updated $t_m^{N + 1}(x_i)$ outputted by $\mathcal{F}_N$ . It is shown that achieving decoupling requires a two-part collaboration, where Eq. (4) + +separates view-related features from global features, and Eq. (9) constrains the remaining view-unrelated features to be orthogonally independent of view-related features. Overall, the total objective of VDT can be written as: + +$$ +\mathcal {L} = \mathcal {L} _ {i} ^ {c} + \mathcal {L} _ {i} ^ {t} + \lambda \left(\mathcal {L} _ {v} ^ {c} + \mathcal {L} _ {o}\right), \tag {10} +$$ + +where $\lambda$ is a hyperparameter to balance multiple objectives. + +# 4. Dataset: CARGO + +# 4.1. Motivation + +Tab. 1 compares the statistics and several other details of various ReID datasets, demonstrating the gap between the homogeneous and heterogeneous views. First, the dataset scale of the homogeneous view has far exceeded that of the heterogeneous view. For example, the identities in the latest AG-ReID [22] are about a quarter of that of the early Market1501 [17]. Second, synthetic data has been well-studied in the homogeneous view, and benefiting from low construction costs, synthetic datasets tend to contain more identities and images. However, the synthetic data study is absent in the view-heterogeneous ReID. The above limitations motivate us to contribute a large-scale synthetic dataset in the heterogeneous view. As can be seen from Tab. 1, the identities and images of CARGO are 12.8 and 4.9 times larger than that of AG-ReID, which brings the view-heterogeneous ReID to a comparable scale with the view-homogeneous ReID. + +# 4.2.Description + +The construction of CARGO is summarized as the construction of human models, the deployment of the camera network, and the data collection, shown in Fig. 3. + +Table 2. Performance comparison of the mainstream methods under four settings of the proposed CARGO dataset. "ALL" denotes the overall retrieval performance of each method. "G ↔ G", "A ↔ A", and "A ↔ G" represent the performance of each model in several specific retrieval patterns. Rank1, mAP, and mINP are reported (\%). The best performance is shown in **bold**. + +
MethodProtocol 1: ALLProtocol 2: G↔GProtocol 3: A↔AProtocol 4: A↔G
Rank1mAPmINPRank1mAPmINPRank1mAPmINPRank1mAPmINP
SBS [36]50.3243.0929.7672.3162.9948.2467.5049.7329.3231.2529.0018.71
PCB [37]51.0044.5032.2074.1067.6055.1055.0044.6027.0034.4030.4020.10
BoT [38]54.8146.4932.4077.6866.4751.3465.0049.7929.8236.2532.5621.46
MGN [39]54.8149.0836.5283.9371.0555.2065.0052.9636.7831.8733.4724.64
VV [40, 41]45.8338.8439.5772.3162.9948.2467.5049.7329.3231.2529.0018.71
AGW [21]60.2653.4440.2281.2571.6658.0967.5056.4840.4043.5740.9029.39
ViT [34]61.5453.5439.6282.1471.3457.5580.0064.4747.0743.1340.1128.20
VDT (Ours)64.1055.2041.1382.1471.5958.3982.5066.8350.2248.1242.7629.95
+ +Table 3. Performance comparison of the mainstream methods under two settings of AG-ReID dataset. "A $\leftrightarrow$ G", and "G $\leftrightarrow$ A" represent the performance in two specific patterns. Rank1, mAP, and mINP are reported (\%). The best performance is shown in bold. + +
MethodProtocol 1: A→GProtocol 2: G→A
Rank1mAPmINPRank1mAPmINP
SBS [36]73.5459.77-73.7062.27-
BoT [38]70.0155.47-71.2058.83-
OSNet [42]72.5958.32-74.2260.99-
VV [40, 41]77.2267.2341.4379.7369.8342.37
ViT [34]81.2872.38-82.6473.35-
Explain [22]81.4772.61-82.8573.39-
VDT (Ours)82.9174.4451.0686.5978.5752.87
+ +First, MakeHuman [43] is adopted to create person models with different identities by assigning randomly different values to predefined attributes, such as gender, height, age, body shape, skin color, hair, clothing, etc. We have created 5,000 pedestrians for subsequent operations, and some models have been shown in Fig. 3a. + +Second, Unity3D [44] is utilized to simulate a real-world streetscape. As shown in Fig. 3b, we pick three blocks to deploy the camera network, where a ground camera network ("G-Cam" area) serves relatively convenient areas such as downtown, and two aerial camera networks ("A-Cam" area) serve the suburbs of a city. In the "G-Cam" area, we deploy eight fixed cameras at each street corner. In the "A-Cam" area, we design different drone roaming strategies based on the size of the surveillance area. For the small area (left area), we deploy one drone with a $90^{\circ}$ overhead view, allowing it to move counterclockwise around each street. For a large area (right area), we deploy individual drones on each of the four streets with a $45^{\circ} \sim 60^{\circ}$ tilt view, allowing them to move back and forth on corresponding streets. + +Finally, we import human models into the streetscape + +and further move them under the ground and aerial networks. We allow multiple pedestrians to move simultaneously to simulate the random occlusion among pedestrians in the real world. Meanwhile, we collect valid images of the target persons if they appear in cameras without severe occlusion (train $\leq 70\%$ , and test $\leq 75\%$ ). In total, CARGO contains 108,563 person images from 13 cameras. + +# 4.3. Challenge + +Fig. 3c illustrates abundant challenges prevalent in real-world scenarios in the CARGO dataset. The most significant challenge is the view discrepancy from different types of cameras, which produces a vast variation and rotation of person postures. Besides, Fig. 3c and Tab. 1 show another change in the camera height, from 5 meters for the lowest ground camera to 75 meters for the highest $90^{\circ}$ view drone, resulting in a huge resolution variation. The maximum resolution in CARGO is $1009 \times 539$ , while the minimum is only $22 \times 7$ . Moreover, CARGO also contains other complex illumination and occlusion variations, etc. Overall, CARGO simulates and covers major real-world scenarios and challenges and can be considered a large-scale benchmark. + +# 5. Experiment + +Importantly, please refer to the supplementary material for detailed experiment settings, visualization of retrieval and feature, cross-dataset evaluation, and other discussions. + +# 5.1. Dataset and Metric + +Dataset. Two datasets are adopted for evaluation, CARGO and AG-ReID [22]. Some statistics regarding these datasets are shown in Tab. 1. As for CARGO, we split it into the train (51,451 images with 2500 IDs) and test sets (51,024 images with the remaining 2500 IDs) with an almost 1:1 ratio. In the training set, each person has an average of 10 images + +![](images/9ea3f714c80118a1e4f945357c564964808e00914a4845bc2fe68e868f0d39b5.jpg) +(a) CARGO under Protocol 1. + +![](images/f1d113337a88105d93a838a3bea75403120d242727524fa11ff3589eedf91e5a.jpg) +(b) AG-ReID under Protocol 1. +Figure 4. Fig. $4\mathrm{a}\sim$ Fig. 4c show the ablation experiments about orthogonal decoupling of view-related and view-unrelated features in the VDT on two dataset, which consists of two important parts, Eq. (4) and Eq. (9). Rank1, mAP, and mINP are reported $(\%)$ . + +![](images/dca13d0f43f36b7b417a67eb6e7656ea334e13fe8ead45f7575ae7d2266a6700.jpg) +(c) AG-RelD under Protocol 2. + +and on average moves across 6.47 ground and 3.61 aerial cameras. In the test set, we select 149 IDs as the query set, and all the remaining IDs are placed in the gallery as distractors. For each identity selected as a query, we keep only its images from two random cameras and use the image from one camera as the query and the image from the other as the gallery. On this basis, we design four protocols (ALL, $\mathrm{A}\leftrightarrow \mathrm{A}$ , $\mathrm{G}\leftrightarrow \mathrm{G}$ , and $\mathrm{A}\leftrightarrow \mathrm{G}$ ) to adequately evaluate the model performances, where "ALL" focuses on the comprehensive retrieval performance, and the latter three focus on some specific retrieval patterns. $\mathrm{A}\leftrightarrow \mathrm{A}$ only retains the data under the aerial camera in the test set for evaluation (60 query IDs with 134 images, 2404 gallery IDs with 18,444 images), same for $\mathrm{G}\leftrightarrow \mathrm{G}$ (89 query IDs with 178 images, 2447 gallery IDs with 32,268 images). $\mathrm{A}\leftrightarrow \mathrm{G}$ relabels the original test set into two domains (aerial and ground domain) based on the view label. The training set of all testing protocols retains same. + +As for AG-ReID, there are 11,554 images with 199 IDs for training and 12,464 images with 189 IDs for testing following [22]. The test set contains 2,033 query images and 10,429 gallery images. There are two protocols in AG-ReID, denoted as $\mathrm{A}\rightarrow \mathrm{G}$ and $\mathrm{G}\rightarrow \mathrm{A}$ , where the former contains 1,701 aerial query images and 3,331 ground gallery images for 189 identities, and the latter contains 962 ground query images and 7,204 aerial gallery images for 189 identities. Optionally, AG-ReID also provides detailed person attribute labeling as additional prior information. + +Metric. The cumulative matching characteristic at Rank1, mean Average Precision (mAP), and mean Inverse Negative Penalty (mINP) [21] are adopted as the evaluation metrics. + +# 5.2. Performance + +Accuracy. We compare the proposed VDT with other competitive methods on CARGO in Tab. 2, including CNN-based (BoT [38], SBS [36], VV [40, 41], MGN [39], AGW [21]) and transformer-based methods (ViT [34]), where VV means a vehicle ReID baseline with vanilla batch-based + +triplet loss and diverse viewpoint batch selection. Similar experimental results on AG-ReID are reported in Tab. 3, where Explain [22] relies on pedestrian attribute priors and therefore does not generalize well to CARGO. Several performances in Tab. 3 follow the results reported in [22]. We can draw the following conclusions: (1) The proposed VDT achieves the state-of-the-art performances that are clearly improved over the baseline, especially in heterogeneous matching. For example, VDT surpasses the mAP/Rank1/mINP baseline by $4.99\% / 2.65\% / 1.75\%$ on the $\mathrm{A} \leftrightarrow \mathrm{G}$ of CARGO. Besides, VDT also brings different degree of benefits to other CARGO protocols. VDT also surpasses Explain on mAP/Rank1 by $3.74\% / 5.18\%$ on the $\mathrm{G} \rightarrow \mathrm{A}$ of AG-ReID, which strongly demonstrates the importance of view decoupled operations in AGPReID to mitigate the disruption of identity representation by view bias whether in homogeneous or heterogeneous matching. (2) Previous view-homogeneous methods show varying degrees of performance degradation in AGPReID, especially on view-heterogeneous protocols, indicating that view bias causes the identity feature under different views to behave with poor cohesion. However, the above methods ignore this issue. (3) The superior performance of VDT does not come from a strong baseline but from the proposed method itself. On the CARGO dataset, the baseline does not consistently outperform the CNN-based AGW on all metrics; on the AGPReID dataset, the baseline is also weaker than the previous Explain. However, VDT achieves the SOTA performance on two datasets, especially on AG-ReID, using fewer prior labels and achieving better performance. + +Speed. We report the inference time for baseline and VDT for retrieving a single image, which is 2.27ms and 2.28ms on the CARGO dataset and 1.44ms and 1.46ms on the AGReID dataset, respectively. These results validate the previous complexity analysis in Sec. 3.2 and show that VDT does not add additional time consumption. Slight time fluctuations tend to be acceptable, due to hardware devices. + +![](images/78e5419c735f6b59836922ac26d9b3d964ee1d5cce75a1bdbff0f4014036bc6a.jpg) +(a) CARGO under Protocol 1. + +![](images/d900f7e1adbb94863037a0b3432ed0550fe6f2e9ba74aab45e5ec9b42358c353.jpg) +(b) AG-ReID under Protocol 1. +Figure 5. Fig. $5\mathrm{a}\sim$ Fig. 5c show the relationship between $\lambda$ and performance on two datasets. For simplicity, only protocol 1 is shown on the CARGO dataset. Rank1, mAP, and mINP are reported (\%). Avg represents the average performance of Rank1, mAP, and mINP. + +![](images/c3be247d4965a2a69b5b0e74e30708154466d6aa0e4ca9937be193a74ac7d917.jpg) +(c) AG-ReID under Protocol 2. + +# 5.3. Ablation Study + +As described in Sec. 3.2, the core contribution of VDT lies in the decoupling of identity and view features by Eq. (4) and Eq. (9). Therefore, we aim to demonstrate the contribution of each part to VDT, which has been shown in Fig. 4. Similar trends in the two datasets demonstrate that the absence of each part leads to a degradation of the model's performance, and the absence of both simultaneously leads to a more severe performance degradation. + +By comparing Row1 and Row2, the absence of Eq. (4) leads to the degradation of $t_m$ and $t_v$ to be equivalent, focusing on the identity and view features in the image respectively, and constraining them to be independent through Eq. (9). However, lacking the guidance of explicit separation within VDT, the decoupling goal cannot be effectively achieved through Eq. (9) alone. Similarly, by comparing Row1 and Row3, the identity and view features outputted by VDT are not guaranteed to be orthogonally separated from each other if Eq. (9) is missing, and thus $t_m$ containing a minor view bias is still not strongly discriminative. + +# 5.4. Parameter Analysis + +As described in Sec. 3.2, VDT introduces only one hyperparameter $\lambda$ in Eq. (10) that needs to be manually tuned for multi-task balance. We select six magnitudes of $\lambda$ from $1e - 4$ to 10 and the corresponding results are shown in Fig. 5. To show a clear tendency, we show the average performance of Rank1, mAP, and mINP at each $\lambda$ . Although the performance shows the same trend of increasing and then decreasing on both datasets as $\lambda$ increases, the peak performance is achieved when $\lambda = 0.001$ on CARGO (Fig. 5a) and $\lambda = 1.0$ on AG-ReID (Fig. 5b and Fig. 5c). + +The key of this phenomenon is the synchronous updating between identity and view classifier, because the view classifier is a simple binary classification, but the identity classifier on the two datasets shows different difficulties depending on the number of identities. As shown in Tab. 1, + +CARGO owns a thousand-level identities, so the identity classification is relatively complicated, and in order to keep tasks in Eq. (10) able to update synchronously, a small $\lambda$ is more appropriate for view-related objectives. Similarly, a large $\lambda$ is selected because of a hundred-level identities in AGPReID. We also provide an empirical $\lambda$ selection guideline to avoid exhaustive search based on above analysis: $\lambda$ tends to be smaller for datasets with larger identity scale. + +# 6. Conclusion and Future Work + +This paper focuses on ReID under a view-heterogeneous scenario, aerial-ground ReID (AGPReID). First, the view-decoupled transformer (VDT) is proposed specifically for the dramatic view differences in AGPReID, which decouples view-related and view-unrelated features by two major operations: hierarchically subtractive separation and orthogonal loss. Second, we contribute a large-scale dataset called CARGO, in which the identities and images are 12.8 and 4.9 times larger than the previous dataset. Experiments on two datasets demonstrate the superiority of VDT and the necessity of view decoupling in AGPReID. + +Although view discrepancy is the most significant factor in AGPReID, it is not the only source of disruptions. Therefore, further exploration of discriminative identity representations under multiple disturbances in the aerial-ground camera network is a promising avenue of future research. + +# Acknowledgments + +This project was supported in part by the NSFC (U22A2095, 62076258), in part by the Key-Area Research and Development Program of Guangzhou (202206030003), in part by Guangdong Project (No. 2020B1515120085), and in part by International Program Fund for Young Talent Scientific Research People, Sun Yat-Sen University. + +# References + +[1] Geon Lee, Sanghoon Lee, Dohyung Kim, Younghoon Shin, Yongsang Yoon, and Bumsub Ham. Camera-driven representation learning for unsupervised domain adaptive person re-identification. In Int. Conf. Comput. Vis., pages 11453–11462, 2023. 1 +[2] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In ACM Int. Conf. Multimedia, page 3115–3123, 2021. 3, 4 +[3] Quan Zhang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling with second-order transformer for group re-identification. In AAAI, volume 36, pages 3318–3325, 2022. +[4] Quan Zhang, Jianhuang Lai, Xiaohua Xie, Xiaofeng Jin, and Sien Huang. Separable spatial-temporal residual graph for cloth-changing group re-identification. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-16, 2024. +[5] Quan Zhang, Kaiheng Dang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Modeling 3d layout for group re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7512-7520, 2022. 1, 2, 4 +[6] Jianxiong Tang, Jianhuang Lai, Xiaohua Xie, Lingxiao Yang, and Wei-Shi Zheng. Ac2as: Activation consistency coupled ann-snn framework for fast and memory-efficient snn training. Pattern Recognition, 2023. 1 +[7] Qi Chen, Yun Chen, Yuheng Huang, Xiaohua Xie, and Lingxiao Yang. Region-based online selective examination for weakly supervised semantic segmentation. Information Fusion, page 102311, 2024. +[8] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7713-7722, June 2022. +[9] Junyang Qiu, Zhanxiang Feng, Lei Wang, and Jianhuang Lai. Salient part-aligned and keypoint disentangling transformer for person re-identification in aerial imagery. In Int. Conf. Multimedia and Expo, 2024. 2 +[10] Pengze Zhang, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Pose guided person image generation via dual-task correlation and affinity learning. IEEE Trans. Vis. Comput. Graph., pages 1-18, 2023. +[11] Jianxiong Tang, Jian-Huang Lai, Xiaohua Xie, and Lingxiao Yang. Spike count maximization for neuromorphic vision recognition. In *IJCAI*, 2023. +[12] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Adv. Neural Inform. Process. Syst., 2023. +[13] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4288-4298, 2022. +[14] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4 +[15] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling for group re-identification. Int. J. Comput. Vis., 2024. + +[16] Lingxiao Yang, Ru-Yuan Zhang, Lida Li, and Xiaohua Xie. Simam: A simple, parameter-free attention module for convolutional neural networks. pages 11863-11874. PMLR, 2021. 1 +[17] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Int. Conf. Comput. Vis., pages 1116–1124, 2015. 1, 2, 4, 5 +[18] Shizhou Zhang, Qi Zhang, Yifei Yang, Xing Wei, Peng Wang, Bingliang Jiao, and Yanning Zhang. Person re-identification in aerial imagery. IEEE Trans. Multimedia, 23:281-291, 2021. 1, 2, 4 +[19] Quan Zhang, Jianhuang Lai, and Xiaohua Xie. Learning modal-invariant angular metric by cyclic projection network for vis-nir person re-identification. IEEE Trans. Image Process., 30:8019-8033, 2021. 2 +[20] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Seeing like a human: Asynchronous learning with dynamic progressive refinement for person re-identification. IEEE Trans. Image Process., 31:352-365, 2022. +[21] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44(6):2872-2893, 2022. 2, 6, 7, 1 +[22] H. Nguyen, K. Nguyen, S. Sridharan, and C. Fookes. Aerial-ground person re-id. In Int. Conf. Multimedia and Expo, pages 2585-2590, 2023. 2, 3, 4, 5, 6, 7 +[23] Lei Wang, Quan Zhang, Junyang Qiu, and Jianhuang Lai. Rotation exploration transformer for aerial person re-identification. In Int. Conf. Multimedia and Expo, 2024. 2 +[24] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11506–11515, 2021. 2, 3, 4 +[25] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 79–88, 2018. 2 +[26] Fei Xiong, Mengran Gou, Octavia Camps, and Mario Sznaier. Person re-identification using kernel-based metric learning methods. In Eur. Conf. Comput. Vis., pages 1–16, 2014. 2 +[27] Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. Person re-identification by local maximal occurrence representation and metric learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2197–2206, 2015. 2 +[28] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In Eur. Conf. Comput. Vis., pages 501-518, 2018. 2 +[29] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Int. Conf. Comput. Vis., pages 15013-15022, 2021. 2 +[30] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles. + +In IEEE Conf. Comput. Vis. Pattern Recog., pages 16266-16275, 2021. 2, 4 +[31] Shuoyi Chen, Mang Ye, and Bo Du. Rotation invariant transformer for recognizing object in uavs. In ACM Int. Conf. Multimedia, page 2565-2574, 2022. 2 +[32] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In ACM Int. Conf. Multimedia, page 3422–3430, 2020. 4 +[33] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4890–4899, 2022. 4 +[34] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2021. 3, 6, 7, 1 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 4 +[36] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. In ACM Int. Conf. Multimedia, page 9664–9667, 2023. 6, 7, 1 +[37] Yifan Sun, Liang Zheng, Yali Li, Yi Yang, Qi Tian, and Shengjin Wang. Learning part-based convolutional features for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 43:902–917, 2021. 6 +[38] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2019. 6, 7 +[39] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM Int. Conf. Multimedia, pages 274–282, 2018. 6, 7 +[40] Ratnesh Kuma, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. Vehicle re-identification: an efficient baseline using triplet embedding. In International Joint Conference on Neural Networks, pages 1–9, 2019. 6, 7 +[41] Ratnesh Kumar, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. A strong and efficient baseline for vehicle re-identification using deep triplet embedding. Journal of Artificial Intelligence and Soft Computing Research, 10(1):27-45, 2020. 6, 7 +[42] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Learning generalisable omni-scale representations for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):5056–5069, 2022. 6 +[43] Leyde Briceno and Gunther Paul. Makehuman: a review of the modelling framework. In Congress of the International Ergonomics Association, pages 224-232, 2018. 6 +[44] Unity Technologies. Unity3D: Cross-platform 3D engine, 2021. 6 + +[45] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248–255, 2009. 1 +[46] Léon Bottou. Stochastic gradient descent tricks. Neural Networks: Tricks of the Trade: Second Edition, pages 421-436, 2012. 1 +[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., pages 8024-8035, 2019. 1 \ No newline at end of file diff --git a/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/images.zip b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8df88d81d95e722195c51451d2de22f64e40b9d3 --- /dev/null +++ b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ea8cdf8d23e79c9086d05feeaaedb5357501ca50a7c91b67a5086ab06bd2eb7 +size 682232 diff --git a/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/layout.json b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f167dd9e7982b52e87b6529518d5d2d450c08743 --- /dev/null +++ b/2024/View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network/layout.json @@ -0,0 +1,9550 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "type": "text", + "content": "View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": "Quan Zhang" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": " Lei Wang" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": " Vishal M. Patel" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": " Xiaohua Xie" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,3,4}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": " Jianhuang Lai" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,3,4*}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-Sen University, China \n" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 159, + 541, + 203 + ], + "type": "text", + "content": "Pazhou Lab (HuangPu), Guangdong, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 203, + 535, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 203, + 535, + 273 + ], + "spans": [ + { + "bbox": [ + 59, + 203, + 535, + 273 + ], + "type": "text", + "content": "3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China \n4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China \n5Department of Electrical and Computer Engineering, Johns Hopkins University, USA \n{zhangq48, wanglei75}@mail2.sysu.edu.cn, {stsljh, xiexiaoh6}@mail.sysu.edu.cn, vpatel36@jhu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 300, + 192, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 300, + 192, + 313 + ], + "spans": [ + { + "bbox": [ + 143, + 300, + 192, + 313 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "type": "text", + "content": "Existing person re-identification methods have achieved remarkable advances in appearance-based identity association across homogeneous cameras, such as ground-ground matching. However, as a more practical scenario, aerial-ground person re-identification (AGPReID) among heterogeneous cameras has received minimal attention. To alleviate the disruption of discriminative identity representation by dramatic view discrepancy as the most significant challenge in AGPReID, the view-decoupled transformer (VDT) is proposed as a simple yet effective framework. Two major components are designed in VDT to decouple view-related and view-unrelated features, namely hierarchical subtractive separation and orthogonal loss, where the former separates these two features inside the VDT, and the latter constrains these two to be independent. In addition, we contribute a large-scale AGPReID dataset called CARGO, consisting of five/eight aerial/ground cameras, 5,000 identities, and 108,563 images. Experiments on two datasets show that VDT is a feasible and effective solution for AGPReID, surpassing the previous method on mAP/Rank1 by up to " + }, + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "type": "inline_equation", + "content": "5.0\\% /2.7\\%" + }, + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "type": "text", + "content": " on CARGO and " + }, + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "type": "inline_equation", + "content": "3.7\\% /5.2\\%" + }, + { + "bbox": [ + 47, + 327, + 290, + 613 + ], + "type": "text", + "content": " on AG-ReID, keeping the same magnitude of computational complexity. Our project is available at https://github.com/LinlyAC/VDT-AGPReID." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "content": "Person re-identification (ReID) aims to associate interested person images based on cross-camera identity similarity, which plays a significant and positive contribution to the se" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 300, + 533, + 501 + ], + "blocks": [ + { + "bbox": [ + 320, + 300, + 533, + 501 + ], + "lines": [ + { + "bbox": [ + 320, + 300, + 533, + 501 + ], + "spans": [ + { + "bbox": [ + 320, + 300, + 533, + 501 + ], + "type": "image", + "image_path": "e500db2cf3afb8a177a92c40844fda3cf765402699f37bfe6096491cb8b918cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 509, + 547, + 576 + ], + "lines": [ + { + "bbox": [ + 305, + 509, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 547, + 576 + ], + "type": "text", + "content": "Figure 1. View-homogeneous vs. view-heterogeneous ReID, where the former focuses on ground-only or aerial-only camera networks, and the latter considers the aerial-ground mixed camera network. Thus, view-heterogeneous ReID considers aerial-aerial, ground-ground, and aerial-ground matching, which is more challenging and practical than the existing view-homogeneous ReID." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "curity and safety of society and citizens [1-5]. Traditional ReID has achieved remarkable progress in the deep learning era [6-16], but it is far from real-world scenarios for the following reasons. (1) Most existing ReID datasets are collected from homogeneous camera networks, which comprises of the same type cameras, such as ground-only [17] or aerial-only [18] camera networks. However, as can be seen from Fig. 1, real-world surveillance systems tend to be deployed as heterogeneous camera networks comprising" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22000" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "aerial and ground cameras instead of a single type alone. Ground cameras cover well-developed areas (city centers), and aerial cameras cover poorly-developed areas (suburbs) due to their broad view range. The complementary nature of the two types would maximize the effectiveness of ReID. (2) Existing ReID methods mainly consider homogeneous matching (ground-ground [19-21] and aerial-aerial [9, 22, 23]), which is ineffective in dealing with the dramatic view discrepancy among heterogeneous matching (aerial-ground). Despite its significance, related research is extremely scarce." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 579 + ], + "type": "text", + "content": "In this paper, we consider the novel and practical view-heterogeneous ReID problem, specifically ReID under the aerial-ground camera network (AGPReID). We propose a view-decoupled transformer (VDT) to specifically tackle the dramatic view discrepancy, which serves as a significant challenge within AGPReID to hinder homogeneous and heterogeneous matchings. Motivated by the fact that view-related features are useless for discriminative identity representations, VDT aims to decouple view-related and view-unrelated components, further facilitating discriminative identity learning from the remaining view-unrelated features. There are two key parts within VDT to achieve view decoupling, namely (a) hierarchically subtractive separation and (b) orthogonal loss. Specifically, after tokenizing the input image to a series of patch tokens, VDT appends two extra tokens (the meta token and view token) to feed them into a transformer stacked by multiple VDT blocks. The meta token captures global representation in the image, and the view token aims to extract view-related features. In each VDT block, (a) is achieved by the subtractive values of the meta token and view token after self-attention operations as the updated meta token for the next block. This means that VDT hierarchically separates view-related features from global features to facilitate identity learning from remaining view-unrelated features. When the meta and view tokens are obtained after the last block, (b) is designed to constrain the identity features to be independent of the view features, ultimately achieving the orthogonal decoupling of view-related and view-unrelated components. Meanwhile, the meta and view tokens will be supervised by identity and view labels, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Considering the scarcity [22] and privacy [5, 24] of the AGPReID datasets, we contribute a large-scale Civic AeRial-GrOund (CARGO) dataset, which is collected from a synthetic city scenario including five aerial and eight ground cameras. CARGO has totally collected 5,000 identities and 108,563 images, which contain not only significant view discrepancy but also plenty of variations of resolution, illumination, occlusion, etc. Regarding evaluation, CARGO considers multiple patterns (aerial-aerial, ground-ground, aerial-ground matchings), which could serve as a comprehensive benchmark. Contributions of this paper can" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 416, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 416, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 416, + 83 + ], + "type": "text", + "content": "be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 85, + 547, + 265 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 85, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 85, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 547, + 156 + ], + "type": "text", + "content": "- We focus on ReID under the aerial-ground camera network (AGPReID) scenario and propose a view-decoupled transformer (VDT) specifically for the dramatic view discrepancy within AGPReID, which achieves view-related and view-unrelated features decoupling by hierarchical subtractive separation and orthogonal loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 156, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 547, + 204 + ], + "type": "text", + "content": "- We contribute a large-scale synthetic dataset, called CARGO, to advance the AGPReID as a benchmark, which contains five/eight aerial/ground cameras, 5,000 identities, 108,563 images, and complex challenges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "type": "text", + "content": "- Experiments demonstrate the superiority of VDT on two datasets, which shows that VDT surpasses the previous method on mAP/Rank1 by up to " + }, + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "type": "inline_equation", + "content": "5.0\\% / 2.7\\%" + }, + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "type": "text", + "content": " on CARGO and " + }, + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "type": "inline_equation", + "content": "3.7\\% / 5.2\\%" + }, + { + "bbox": [ + 306, + 204, + 547, + 265 + ], + "type": "text", + "content": " on AG-ReID, respectively, maintaining the same magnitude of computational complexity." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 275, + 392, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 275, + 392, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 275, + 392, + 289 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 296, + 447, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 447, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 447, + 308 + ], + "type": "text", + "content": "2.1. View-homogeneous ReID" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 315, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 469 + ], + "type": "text", + "content": "View-homogeneous ReID deals with images from ground-only or aerial-only camera networks, where the ground-only camera network receives the broadest attention in the ReID task because it is the most common use case. Many ground-only datasets have been contributed in the literature, such as Market1501 [17], MSMT17 [25], etc. Accordingly, lots of methods have been proposed to bring rapid development to ReID, including handcrafted feature-based methods [26, 27], CNN-based methods [19-21, 28], and transformer-based methods [29]. Regarding aerial-only research, only some pioneering efforts have contributed substantial datasets and methods [18, 30, 31] because it has not received as much attention as ground-only ReID." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 470, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 519 + ], + "type": "text", + "content": "However, it is difficult to directly transfer the above methods to the view-heterogeneous ReID because these methods ignore the significant view differences between aerial and ground cameras, leading to weak performances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 526, + 451, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 451, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 451, + 540 + ], + "type": "text", + "content": "2.2. View-heterogeneous ReID" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": "Compared to the abundant research in Sec. 2.1, studies on view-heterogeneous ReID are almost nonexistent. In this paper, heterogeneous views refer to AGPReID, and the fundamental challenge is the dramatic view discrepancy. To the best of our knowledge, one recent work [22] has formally attempted to address this task. They collected an outdoor scene dataset, AG-ReID, with both identity and attribute labeling of pedestrians. Moreover, they proposed an explainable model by attribute information to guide model training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "While pioneering and critical to AGPReID, this work is still imperfect for the following reasons. (1) Dataset. Compared to [22], our CARGO dataset comprehensively focuses on multiple matching patterns that may appear in heterogeneous views rather than a single aerial-ground matching in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22001" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 534, + 276 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 534, + 276 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 534, + 276 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 534, + 276 + ], + "type": "image", + "image_path": "4fe5321429469995b02bae8316f7a9290adee4613e4f0ab4a79e6a757f85c5a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed VDT framework, which consists of " + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": " VDT blocks and three-part loss functions. Meta and view tokens capture global and view-related features in images, respectively. Each VDT block (light blue module) consists of a standard self-attention encoder layer and an inner feature subtraction operation, achieving layer-by-layer decoupling of view-related and view-unrelated features. Orthogonal loss constrains the above two features to be further independent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 288, + 396 + ], + "type": "text", + "content": "AG-ReID. (2) Method. Compared to explainable methods in [22], our VDT relies on a lesser amount of priori labeling and thus has a stronger generalization. The experiments also show that our VDT achieves stronger performances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 180, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 180, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 180, + 416 + ], + "type": "text", + "content": "2.3. Synthetic ReID Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "text", + "content": "Data privacy is an inevitable issue for the ReID datasets, and synthetic data provides a feasible solution because it not only achieves privacy protection for real-world pedestrians but also has lower collection costs. Many synthetic ReID datasets have been published in the literature, such as PersonX [14], UnrealPerson [24], WePerson [2], etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 288, + 567 + ], + "type": "text", + "content": "However, existing synthetic datasets mainly focus on simulating pedestrian data under view-homogeneous camera networks, and the simulation of aerial-ground views is poorly studied. To this end, we contribute the CARGO dataset as the first large-scale synthetic dataset for AGPReID." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 578, + 103, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 103, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 103, + 591 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 197, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 197, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 197, + 611 + ], + "type": "text", + "content": "3.1. Formulation and Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": "Formulation. An AGPReLU dataset " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i, v_i)\\}_{i=1}^{|\\mathcal{D}|}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " consists of a training set " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{tr}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " and a test set " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{te}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": "-th person image, and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " stand for the corresponding identity and view labels, respectively. Note that " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_i \\in \\{v^a, v^g\\}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is readily acquired by the known camera labels in " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " to easily distinguish whether " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " belongs to an aerial " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "v^a" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " or ground " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "v^g" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " view. Significant view discrepancy between " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "v^a" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "v^g" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " leads to a view-biased feature" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "type": "text", + "content": "space, which behaves as low intra-identity and high interidentity similarity. Therefore, the goal of AGPReID methods is to design a model " + }, + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot;\\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "type": "text", + "content": " with learnable parameter " + }, + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}" + }, + { + "bbox": [ + 305, + 347, + 546, + 396 + ], + "type": "text", + "content": " against the view bias, which can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 403, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 403, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 331, + 403, + 545, + 437 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\sum_ {i} \\left[ \\begin{array}{c} \\left\\| \\mathcal {F} \\left(x _ {i}; \\theta_ {\\mathcal {F}}\\right) - \\mathcal {F} \\left(x _ {i} ^ {+}; \\theta_ {\\mathcal {F}}\\right) \\right\\| _ {2} ^ {2} - \\\\ \\left\\| \\mathcal {F} \\left(x _ {i}; \\theta_ {\\mathcal {F}}\\right) - \\mathcal {F} \\left(x _ {i} ^ {-}; \\theta_ {\\mathcal {F}}\\right) \\right\\| _ {2} ^ {2} \\end{array} \\right], \\tag {1}", + "image_path": "65fa9ad030bb8984ffd8b2d03ecde71bdfc5540578dad9444a622e1702020fc1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "text", + "content": " denotes the L2 distance, " + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "inline_equation", + "content": "x_{i}^{+} / x_{i}^{-}" + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "text", + "content": " denotes the person image that has the same/different identity with " + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 446, + 545, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "text", + "content": "Overview. As shown in Fig. 2, we design a simple yet effective framework called view-decoupled transformer (VDT), to tackle the view discrepancy challenge in AGPReID. For a batched data " + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "text", + "content": " that contains both " + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "inline_equation", + "content": "v^a" + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "inline_equation", + "content": "v^g" + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "text", + "content": " view, weckenize them as a series of image tokens (green squares), then append a meta and view token (blue and yellow squares) to them, which serve as input to VDT. The VDT network consists of " + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 472, + 546, + 662 + ], + "type": "text", + "content": " blocks, where each block first performs the standard self-attention encoding and then performs a subtraction operation between meta and view tokens to explicitly separate the view-related feature from the global feature. Finally, the identity and view classifier will supervise the updated meta and view tokens outputted from VDT. Besides, an orthogonal loss is proposed that makes these two tokens irrelevant, thus achieving the complete decoupling of view-related and view-unrelated features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 670, + 466, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 466, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 466, + 684 + ], + "type": "text", + "content": "3.2. View-decoupled Transformer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": "The proposed VDT framework is based on Vit-Base [34]. For each image " + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": " in the batched data " + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": ", VDT first" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22002" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 102, + 496, + 248 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 1. Statistical comparisons of existing datasets, including view-homogeneous (ground or aerial) and view-heterogeneous (ground and aerial) ReID. “" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "” denotes City1M is used for group ReID, where its images correspond to the number of group images." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 102, + 496, + 248 + ], + "lines": [ + { + "bbox": [ + 97, + 102, + 496, + 248 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 496, + 248 + ], + "type": "table", + "html": "
DatasetViewData#PersonID#Camera#Image#Height
Market1501 [17]GroundReal1,501632,668< 10m
PersonX [14]GroundSynthetic1,2666273,456-
RandPerson [32]GroundSynthetic8,00019228,655-
UnrealPerson [24]GroundSynthetic3,00034120,000-
WePerson [2]GroundSynthetic1,500404,000,000-
ClonedPerson [33]GroundSynthetic5,62124887,766-
City1M* [5]GroundSynthetic45,00081,840,000-
PRAI1581 [18]AerialReal1,581239,46120 ~ 60m
UAVHuman [30]AerialReal1,144141,2902 ~ 8m
AG-ReID [22]Aerial-GroundReal3882 (1A+1G)21,89315 ~ 45m
CARGO (Ours)Aerial-GroundSynthetic5,00013 (5A+8G)108,5635 ~ 75m
", + "image_path": "4a2f776f38c01bdfc28de98ab086798141ebddac73ad90e487d637d7296d6df3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": "divides " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " evenly and non-overlappingly into " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " image patches, which are further tokenized into " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " patch embeddings by a " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " convolution and can be denoted as " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "[t_{p,1}; t_{p,2}; \\dots; t_{p,M}]" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": ". After that, we append two additional learnable tokens, meta token " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "t_{m}" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " and view token " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "t_{v}" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": ", in the patch tokens, where " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "t_{m}" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " aims to capture the global image representation, " + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "inline_equation", + "content": "t_{v}" + }, + { + "bbox": [ + 46, + 268, + 289, + 376 + ], + "type": "text", + "content": " focuses only on the view-related feature. Next, we serialize all the tokens by assigning them the corresponding positional embedding, which can be written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 387, + 287, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 387, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 74, + 387, + 287, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (x _ {i}) \\triangleq \\left[ t _ {m}; t _ {p, 1}; t _ {p, 2}; \\dots ; t _ {p, M}; t _ {v} \\right] + \\mathbf {E} _ {p o s}, \\tag {2}", + "image_path": "83321240a7e5aba9932f6587d913a8d882bd57b3b3b44f6b13b744ac962916bd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(\\cdot)" + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": " is defined as the serialization operation for " + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "inline_equation", + "content": "t_m, t_{p,i}, t_v \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{pos} \\in \\mathbb{R}^{(M + 2) \\times d}" + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "text", + "content": "As shown in Fig. 2, VDT " + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot ;\\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "text", + "content": " is stacked with " + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "text", + "content": " VDT blocks " + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_j(\\cdot ;\\theta_{\\mathcal{F}}^j),1\\leq j\\leq N." + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "text", + "content": " Each " + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_j(\\cdot ;\\theta_{\\mathcal{F}}^j)" + }, + { + "bbox": [ + 46, + 430, + 287, + 477 + ], + "type": "text", + "content": " contains two operations, self-attention encoding and subtractive separation, which are described as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 487, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 487, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 51, + 487, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\left[ t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right); \\dots ; t _ {v} ^ {(j + 1)} \\left(x _ {i}\\right) \\right] = \\theta_ {\\mathcal {F}} ^ {(j)} \\left(\\left[ t _ {m} ^ {(j)} \\left(x _ {i}\\right); \\dots ; t _ {v} ^ {(j)} \\left(x _ {i}\\right) \\right]\\right), \\tag {3}", + "image_path": "512bdd9df6fd45cd8599af0e41d60cffa422e1da55ee391dd647a35ec8337bd8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 523, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 523, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 88, + 523, + 287, + 539 + ], + "type": "interline_equation", + "content": "t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right) \\leftarrow t _ {m} ^ {(j + 1)} \\left(x _ {i}\\right) - t _ {v} ^ {(j + 1)} \\left(x _ {i}\\right), \\tag {4}", + "image_path": "6a8ea083cce367d645327e5ca0560b7bb173c5ff712820e1f8f6faa7820de64d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}^{(j)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " represents the self-attention parameters in " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_j" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "j = 1" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": ", the input of " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_1" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " becomes " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(x_i)" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_m^{(j + 1)}(x_i)" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_v^{(j + 1)}(x_i)" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " in Eq. (3) represent the updated meta and view token after the self-attention operation in " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_j" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " with patch tokens of " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": ". Eq. (4) explicitly guides that " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_m^{(j + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_v^{(j + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " are distinguishable after Eq. (3), where " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_v^{(j + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " captures more view-related features and " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_m^{(j + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " can capture more view-unrelated features by removing " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_v^{(j + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " from the global (meta) features. Eq. (4) facilitates the learning of identity features via " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_m^{(j + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": ", alleviating the disturbing of view bias. VDT decouples the view-related and -unrelated features layer by layer until the updated " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_m^{(N + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_v^{(N + 1)}" + }, + { + "bbox": [ + 46, + 544, + 289, + 714 + ], + "type": "text", + "content": " are obtained, which serves as the well-decoupled identity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 293 + ], + "type": "text", + "content": "and view features, respectively. During inference, we only use " + }, + { + "bbox": [ + 304, + 268, + 545, + 293 + ], + "type": "inline_equation", + "content": "t_m^{(N + 1)}" + }, + { + "bbox": [ + 304, + 268, + 545, + 293 + ], + "type": "text", + "content": " as the whole retrieval evidence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "text", + "content": "In addition, the proposed VDT maintains the same computational complexity as baseline (ViT). Given that the input to ViT has " + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "text", + "content": " tokens and the dimensions of the tokens are " + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "text", + "content": ", then the computational complexity of ViT is " + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\big(4Td^2 + 2T^2d\\big)" + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "text", + "content": " [35]. Our VDT adds an additional " + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "inline_equation", + "content": "t_v" + }, + { + "bbox": [ + 304, + 293, + 546, + 389 + ], + "type": "text", + "content": " token based on the baseline and adds an Eq. (4) operation in each VDT block. Therefore, the total computational complexity can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 401, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 401, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 319, + 401, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {O} \\left(4 (T + 1) d ^ {2} + 2 (T + 1) ^ {2} d + N\\right) \\\\ \\Leftrightarrow \\mathcal {O} \\left(\\underbrace {4 T d ^ {2} + 2 T ^ {2} d} _ {\\text {b a s e l i n e c o m p l e x i t y}} + \\underbrace {4 d ^ {2} + 2 d + N} _ {\\text {c o n s t a n t c o m p l e x i t y}}\\right) \\tag {5} \\\\ \\Leftrightarrow \\mathcal {O} \\left(4 T d ^ {2} + 2 T ^ {2} d\\right), \\\\ \\end{array}", + "image_path": "66fe20612f3cdef4653984227138e1c0a1cadb56e846e1331c4d3029f68a02b5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": " counts all Eq. (4) in " + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": " VDT blocks. Eq. (5) shows that the extra operations VDT introduces depend on only the network hyperparameters, which belong to constant complexity and can be omitted. Therefore, VDT maintains the same magnitude of complexity as that of the baseline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 554, + 391, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 391, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 391, + 567 + ], + "type": "text", + "content": "3.3. Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 572, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 608 + ], + "type": "text", + "content": "To meet the task goal (Eq. (1)), the objective of " + }, + { + "bbox": [ + 304, + 572, + 545, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot, \\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 304, + 572, + 545, + 608 + ], + "type": "text", + "content": " consists of three components: identity classifier, view classifier, and orthogonal loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "type": "text", + "content": "For each " + }, + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "type": "inline_equation", + "content": "x_{i} \\in B" + }, + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "type": "text", + "content": ", the identity classifier utilizes both the cross-entropy loss and the triplet loss to supervise " + }, + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "type": "inline_equation", + "content": "t_m^{N + 1}(x_i)" + }, + { + "bbox": [ + 304, + 609, + 545, + 646 + ], + "type": "text", + "content": ", which can be written as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 652, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 652, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 376, + 652, + 545, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {c} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} y _ {i} \\log (\\hat {y} _ {i}), \\tag {6}", + "image_path": "256126abcbba9b0b57efb45e405a30dfce047fd2e3dc018a69fd7367049e32f2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "|B|" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " represents the batch size, " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " represents the identity label, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " denotes the identity prediction by" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22003" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 162, + 222 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 162, + 222 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 162, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 162, + 222 + ], + "type": "image", + "image_path": "c66d843dc5e9a3cdded5ccb45c918918a0cc7a8b34ffbd02e4e7d44138c5c161.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 224, + 153, + 234 + ], + "lines": [ + { + "bbox": [ + 69, + 224, + 153, + 234 + ], + "spans": [ + { + "bbox": [ + 69, + 224, + 153, + 234 + ], + "type": "text", + "content": "(a) 3D pedestrian models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 73, + 438, + 223 + ], + "blocks": [ + { + "bbox": [ + 164, + 73, + 438, + 223 + ], + "lines": [ + { + "bbox": [ + 164, + 73, + 438, + 223 + ], + "spans": [ + { + "bbox": [ + 164, + 73, + 438, + 223 + ], + "type": "image", + "image_path": "099d26879c3c1fbd25b09cf92162eda3e37660735b02d8ba6f33ed913a23e95a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 224, + 376, + 235 + ], + "lines": [ + { + "bbox": [ + 223, + 224, + 376, + 235 + ], + "spans": [ + { + "bbox": [ + 223, + 224, + 376, + 235 + ], + "type": "text", + "content": "(b) Aerial-ground camera network deployment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 533, + 222 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 533, + 222 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 533, + 222 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 533, + 222 + ], + "type": "image", + "image_path": "043a0a48f6b815686e3bac3f1b6ac828f467c54e999663cbffbb73da92dfc402.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 224, + 525, + 235 + ], + "lines": [ + { + "bbox": [ + 446, + 224, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 446, + 224, + 525, + 235 + ], + "type": "text", + "content": "(c) Diversity challenges." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "type": "text", + "content": "Figure 3. Fig. " + }, + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "type": "inline_equation", + "content": "3\\mathrm{a}\\sim" + }, + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "type": "text", + "content": " Fig. 3c shows the pedestrian models, camera deployment, and challenges during the CARGO construction, respectively. In Fig. 3b, \"A-Cam\" and \"G-Cam\" represent the aerial and ground cameras, where the yellow sectors represent the view range of the ground cameras, and the green arrows represent the motion strategy of the aerial cameras. The challenges displayed in Fig. 3c are view variation, illumination variation, occlusion, and resolution variation from top to bottom." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 308, + 241, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 241, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 241, + 322 + ], + "type": "inline_equation", + "content": "t_m^{N + 1}(x_i)" + }, + { + "bbox": [ + 47, + 308, + 241, + 322 + ], + "type": "text", + "content": ". The triplet loss is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 329, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 329, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 60, + 329, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {t} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} \\left[ \\begin{array}{c} \\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) - t _ {m} ^ {N + 1} \\left(x _ {i} ^ {+}\\right) \\right\\| _ {2} ^ {2} - \\\\ \\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) - t _ {m} ^ {N + 1} \\left(x _ {i} ^ {-}\\right) \\right\\| _ {2} ^ {2} + \\alpha \\end{array} \\right] _ {+}, \\tag {7}", + "image_path": "fb1d9f23dbefd7460299af198ce61e27ab26780bac296578da866029751a0ab7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "x_{i}^{+} / x_{i}^{-}" + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": " denotes the hard positive/negative sample of " + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": " in the current " + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "[\\cdot]_{+}" + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": " stands for " + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\max(\\cdot, 0)" + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 369, + 287, + 405 + ], + "type": "text", + "content": " is the hyperparameter of the margin." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 404, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 287, + 430 + ], + "type": "text", + "content": "View classifier supervises " + }, + { + "bbox": [ + 47, + 404, + 287, + 430 + ], + "type": "inline_equation", + "content": "t_v^{N + 1}(x_i)" + }, + { + "bbox": [ + 47, + 404, + 287, + 430 + ], + "type": "text", + "content": " with only the cross-entropy loss, which can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 437, + 287, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 437, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 118, + 437, + 287, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v} ^ {c} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} v _ {i} \\log (\\hat {v} _ {i}), \\tag {8}", + "image_path": "0db2ebf52430311c56e897929ab0a3ee115909de345c7f02fbbd2de1592ad2b9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "text", + "content": " represents the view label of " + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{v}_{i}" + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "text", + "content": " denotes the view prediction by " + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "inline_equation", + "content": "t_v^{N + 1}(x_i)" + }, + { + "bbox": [ + 47, + 474, + 287, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "type": "text", + "content": "To enable " + }, + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "type": "inline_equation", + "content": "t_m^{N + 1}(x_i)" + }, + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "type": "inline_equation", + "content": "t_v^{N + 1}(x_i)" + }, + { + "bbox": [ + 47, + 498, + 287, + 533 + ], + "type": "text", + "content": " to be truly decoupled, we impose an orthogonality loss that makes them less similar to each other, which can be written as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 83, + 540, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 540, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 83, + 540, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\frac {1}{| B |} \\sum_ {i = 1} ^ {| B |} \\frac {\\left| \\left\\langle t _ {m} ^ {N + 1} \\left(x _ {i}\\right) , t _ {v} ^ {N + 1} \\left(x _ {i}\\right) \\right\\rangle \\right|}{\\left\\| t _ {m} ^ {N + 1} \\left(x _ {i}\\right) \\right\\| _ {2} \\cdot \\left\\| t _ {v} ^ {N + 1} \\left(x _ {i}\\right) \\right\\| _ {2}}, \\tag {9}", + "image_path": "cad510b84e12ee46584bbb54d5bbbe7ac0202768392844758951856f4969633d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "inline_equation", + "content": "|\\langle \\cdot ,\\cdot \\rangle |" + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "text", + "content": " represents the absolute value after the dot product of two token embeddings. We also provide a clear illustration of Eq. (9) in Fig. 2 (the gray rectangle in the middle on the right). Please note that the dashed blue and the yellow vector represent the meta token and view token outputted by Eq. (3) in " + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_N" + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "text", + "content": " respectively, and the blue vector represents the result of Eq. (4), which is the updated " + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "inline_equation", + "content": "t_m^{N + 1}(x_i)" + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "text", + "content": " outputted by " + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_N" + }, + { + "bbox": [ + 47, + 578, + 287, + 687 + ], + "type": "text", + "content": ". It is shown that achieving decoupling requires a two-part collaboration, where Eq. (4)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 309, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 309, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 545, + 357 + ], + "type": "text", + "content": "separates view-related features from global features, and Eq. (9) constrains the remaining view-unrelated features to be orthogonally independent of view-related features. Overall, the total objective of VDT can be written as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 365, + 545, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 365, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 365, + 365, + 545, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {i} ^ {c} + \\mathcal {L} _ {i} ^ {t} + \\lambda \\left(\\mathcal {L} _ {v} ^ {c} + \\mathcal {L} _ {o}\\right), \\tag {10}", + "image_path": "4ea0918d760b24848bf0d1cacfc9c5605082aba17e3e7f44908e9fe1c08cb85d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 388, + 542, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 542, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 542, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 388, + 542, + 401 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 388, + 542, + 401 + ], + "type": "text", + "content": " is a hyperparameter to balance multiple objectives." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 410, + 411, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 411, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 411, + 423 + ], + "type": "text", + "content": "4. Dataset: CARGO" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 430, + 380, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 380, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 380, + 441 + ], + "type": "text", + "content": "4.1. Motivation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 449, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 652 + ], + "type": "text", + "content": "Tab. 1 compares the statistics and several other details of various ReID datasets, demonstrating the gap between the homogeneous and heterogeneous views. First, the dataset scale of the homogeneous view has far exceeded that of the heterogeneous view. For example, the identities in the latest AG-ReID [22] are about a quarter of that of the early Market1501 [17]. Second, synthetic data has been well-studied in the homogeneous view, and benefiting from low construction costs, synthetic datasets tend to contain more identities and images. However, the synthetic data study is absent in the view-heterogeneous ReID. The above limitations motivate us to contribute a large-scale synthetic dataset in the heterogeneous view. As can be seen from Tab. 1, the identities and images of CARGO are 12.8 and 4.9 times larger than that of AG-ReID, which brings the view-heterogeneous ReID to a comparable scale with the view-homogeneous ReID." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 659, + 383, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 383, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 383, + 672 + ], + "type": "text", + "content": "4.2.Description" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "The construction of CARGO is summarized as the construction of human models, the deployment of the camera network, and the data collection, shown in Fig. 3." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "1The output of " + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": " is the count of elements when the input is a set, and the output is the absolute value when the input is a number." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22004" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 113, + 547, + 259 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Performance comparison of the mainstream methods under four settings of the proposed CARGO dataset. \"ALL\" denotes the overall retrieval performance of each method. \"G ↔ G\", \"A ↔ A\", and \"A ↔ G\" represent the performance of each model in several specific retrieval patterns. Rank1, mAP, and mINP are reported (\\%). The best performance is shown in **bold**." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 113, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 113, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 113, + 547, + 259 + ], + "type": "table", + "html": "
MethodProtocol 1: ALLProtocol 2: G↔GProtocol 3: A↔AProtocol 4: A↔G
Rank1mAPmINPRank1mAPmINPRank1mAPmINPRank1mAPmINP
SBS [36]50.3243.0929.7672.3162.9948.2467.5049.7329.3231.2529.0018.71
PCB [37]51.0044.5032.2074.1067.6055.1055.0044.6027.0034.4030.4020.10
BoT [38]54.8146.4932.4077.6866.4751.3465.0049.7929.8236.2532.5621.46
MGN [39]54.8149.0836.5283.9371.0555.2065.0052.9636.7831.8733.4724.64
VV [40, 41]45.8338.8439.5772.3162.9948.2467.5049.7329.3231.2529.0018.71
AGW [21]60.2653.4440.2281.2571.6658.0967.5056.4840.4043.5740.9029.39
ViT [34]61.5453.5439.6282.1471.3457.5580.0064.4747.0743.1340.1128.20
VDT (Ours)64.1055.2041.1382.1471.5958.3982.5066.8350.2248.1242.7629.95
", + "image_path": "15f0773b5e37ded495831b38b4aee4c04dffc1cd2511bfea9b1695febfb1ceec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 331, + 285, + 441 + ], + "blocks": [ + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "type": "text", + "content": "Table 3. Performance comparison of the mainstream methods under two settings of AG-ReID dataset. \"A " + }, + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "type": "text", + "content": " G\", and \"G " + }, + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 46, + 277, + 288, + 323 + ], + "type": "text", + "content": " A\" represent the performance in two specific patterns. Rank1, mAP, and mINP are reported (\\%). The best performance is shown in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 331, + 285, + 441 + ], + "lines": [ + { + "bbox": [ + 50, + 331, + 285, + 441 + ], + "spans": [ + { + "bbox": [ + 50, + 331, + 285, + 441 + ], + "type": "table", + "html": "
MethodProtocol 1: A→GProtocol 2: G→A
Rank1mAPmINPRank1mAPmINP
SBS [36]73.5459.77-73.7062.27-
BoT [38]70.0155.47-71.2058.83-
OSNet [42]72.5958.32-74.2260.99-
VV [40, 41]77.2267.2341.4379.7369.8342.37
ViT [34]81.2872.38-82.6473.35-
Explain [22]81.4772.61-82.8573.39-
VDT (Ours)82.9174.4451.0686.5978.5752.87
", + "image_path": "6e80969a70fc9047e008b1165959131037647cd5db5e676a283d824c10ead5a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 533 + ], + "type": "text", + "content": "First, MakeHuman [43] is adopted to create person models with different identities by assigning randomly different values to predefined attributes, such as gender, height, age, body shape, skin color, hair, clothing, etc. We have created 5,000 pedestrians for subsequent operations, and some models have been shown in Fig. 3a." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": "Second, Unity3D [44] is utilized to simulate a real-world streetscape. As shown in Fig. 3b, we pick three blocks to deploy the camera network, where a ground camera network (\"G-Cam\" area) serves relatively convenient areas such as downtown, and two aerial camera networks (\"A-Cam\" area) serve the suburbs of a city. In the \"G-Cam\" area, we deploy eight fixed cameras at each street corner. In the \"A-Cam\" area, we design different drone roaming strategies based on the size of the surveillance area. For the small area (left area), we deploy one drone with a " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " overhead view, allowing it to move counterclockwise around each street. For a large area (right area), we deploy individual drones on each of the four streets with a " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "45^{\\circ} \\sim 60^{\\circ}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " tilt view, allowing them to move back and forth on corresponding streets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Finally, we import human models into the streetscape" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "type": "text", + "content": "and further move them under the ground and aerial networks. We allow multiple pedestrians to move simultaneously to simulate the random occlusion among pedestrians in the real world. Meanwhile, we collect valid images of the target persons if they appear in cameras without severe occlusion (train " + }, + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "type": "inline_equation", + "content": "\\leq 70\\%" + }, + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "type": "text", + "content": ", and test " + }, + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "type": "inline_equation", + "content": "\\leq 75\\%" + }, + { + "bbox": [ + 304, + 279, + 547, + 363 + ], + "type": "text", + "content": "). In total, CARGO contains 108,563 person images from 13 cameras." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 371, + 376, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 376, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 376, + 384 + ], + "type": "text", + "content": "4.3. Challenge" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "text", + "content": "Fig. 3c illustrates abundant challenges prevalent in real-world scenarios in the CARGO dataset. The most significant challenge is the view discrepancy from different types of cameras, which produces a vast variation and rotation of person postures. Besides, Fig. 3c and Tab. 1 show another change in the camera height, from 5 meters for the lowest ground camera to 75 meters for the highest " + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "text", + "content": " view drone, resulting in a huge resolution variation. The maximum resolution in CARGO is " + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "inline_equation", + "content": "1009 \\times 539" + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "text", + "content": ", while the minimum is only " + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "inline_equation", + "content": "22 \\times 7" + }, + { + "bbox": [ + 304, + 390, + 547, + 547 + ], + "type": "text", + "content": ". Moreover, CARGO also contains other complex illumination and occlusion variations, etc. Overall, CARGO simulates and covers major real-world scenarios and challenges and can be considered a large-scale benchmark." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 557, + 382, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 382, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 382, + 571 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "text", + "content": "Importantly, please refer to the supplementary material for detailed experiment settings, visualization of retrieval and feature, cross-dataset evaluation, and other discussions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 419, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 419, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 419, + 635 + ], + "type": "text", + "content": "5.1. Dataset and Metric" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Dataset. Two datasets are adopted for evaluation, CARGO and AG-ReID [22]. Some statistics regarding these datasets are shown in Tab. 1. As for CARGO, we split it into the train (51,451 images with 2500 IDs) and test sets (51,024 images with the remaining 2500 IDs) with an almost 1:1 ratio. In the training set, each person has an average of 10 images" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22005" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 203, + 192 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 203, + 192 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 203, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 203, + 192 + ], + "type": "image", + "image_path": "9ea3f714c80118a1e4f945357c564964808e00914a4845bc2fe68e868f0d39b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 196, + 179, + 205 + ], + "lines": [ + { + "bbox": [ + 82, + 196, + 179, + 205 + ], + "spans": [ + { + "bbox": [ + 82, + 196, + 179, + 205 + ], + "type": "text", + "content": "(a) CARGO under Protocol 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 218, + 76, + 369, + 191 + ], + "blocks": [ + { + "bbox": [ + 218, + 76, + 369, + 191 + ], + "lines": [ + { + "bbox": [ + 218, + 76, + 369, + 191 + ], + "spans": [ + { + "bbox": [ + 218, + 76, + 369, + 191 + ], + "type": "image", + "image_path": "f1d113337a88105d93a838a3bea75403120d242727524fa11ff3589eedf91e5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 196, + 347, + 205 + ], + "lines": [ + { + "bbox": [ + 246, + 196, + 347, + 205 + ], + "spans": [ + { + "bbox": [ + 246, + 196, + 347, + 205 + ], + "type": "text", + "content": "(b) AG-ReID under Protocol 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": "Figure 4. Fig. " + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "inline_equation", + "content": "4\\mathrm{a}\\sim" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": " Fig. 4c show the ablation experiments about orthogonal decoupling of view-related and view-unrelated features in the VDT on two dataset, which consists of two important parts, Eq. (4) and Eq. (9). Rank1, mAP, and mINP are reported " + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 384, + 76, + 534, + 191 + ], + "blocks": [ + { + "bbox": [ + 384, + 76, + 534, + 191 + ], + "lines": [ + { + "bbox": [ + 384, + 76, + 534, + 191 + ], + "spans": [ + { + "bbox": [ + 384, + 76, + 534, + 191 + ], + "type": "image", + "image_path": "dca13d0f43f36b7b417a67eb6e7656ea334e13fe8ead45f7575ae7d2266a6700.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 196, + 512, + 205 + ], + "lines": [ + { + "bbox": [ + 412, + 196, + 512, + 205 + ], + "spans": [ + { + "bbox": [ + 412, + 196, + 512, + 205 + ], + "type": "text", + "content": "(c) AG-RelD under Protocol 2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": "and on average moves across 6.47 ground and 3.61 aerial cameras. In the test set, we select 149 IDs as the query set, and all the remaining IDs are placed in the gallery as distractors. For each identity selected as a query, we keep only its images from two random cameras and use the image from one camera as the query and the image from the other as the gallery. On this basis, we design four protocols (ALL, " + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\leftrightarrow \\mathrm{A}" + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{G}\\leftrightarrow \\mathrm{G}" + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\leftrightarrow \\mathrm{G}" + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": ") to adequately evaluate the model performances, where \"ALL\" focuses on the comprehensive retrieval performance, and the latter three focus on some specific retrieval patterns. " + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\leftrightarrow \\mathrm{A}" + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": " only retains the data under the aerial camera in the test set for evaluation (60 query IDs with 134 images, 2404 gallery IDs with 18,444 images), same for " + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{G}\\leftrightarrow \\mathrm{G}" + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": " (89 query IDs with 178 images, 2447 gallery IDs with 32,268 images). " + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\leftrightarrow \\mathrm{G}" + }, + { + "bbox": [ + 46, + 258, + 287, + 474 + ], + "type": "text", + "content": " relabels the original test set into two domains (aerial and ground domain) based on the view label. The training set of all testing protocols retains same." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "type": "text", + "content": "As for AG-ReID, there are 11,554 images with 199 IDs for training and 12,464 images with 189 IDs for testing following [22]. The test set contains 2,033 query images and 10,429 gallery images. There are two protocols in AG-ReID, denoted as " + }, + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\rightarrow \\mathrm{G}" + }, + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{G}\\rightarrow \\mathrm{A}" + }, + { + "bbox": [ + 46, + 474, + 287, + 594 + ], + "type": "text", + "content": ", where the former contains 1,701 aerial query images and 3,331 ground gallery images for 189 identities, and the latter contains 962 ground query images and 7,204 aerial gallery images for 189 identities. Optionally, AG-ReID also provides detailed person attribute labeling as additional prior information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "type": "text", + "content": "Metric. The cumulative matching characteristic at Rank1, mean Average Precision (mAP), and mean Inverse Negative Penalty (mINP) [21] are adopted as the evaluation metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 130, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 130, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 130, + 647 + ], + "type": "text", + "content": "5.2. Performance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": "Accuracy. We compare the proposed VDT with other competitive methods on CARGO in Tab. 2, including CNN-based (BoT [38], SBS [36], VV [40, 41], MGN [39], AGW [21]) and transformer-based methods (ViT [34]), where VV means a vehicle ReID baseline with vanilla batch-based" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "text", + "content": "triplet loss and diverse viewpoint batch selection. Similar experimental results on AG-ReID are reported in Tab. 3, where Explain [22] relies on pedestrian attribute priors and therefore does not generalize well to CARGO. Several performances in Tab. 3 follow the results reported in [22]. We can draw the following conclusions: (1) The proposed VDT achieves the state-of-the-art performances that are clearly improved over the baseline, especially in heterogeneous matching. For example, VDT surpasses the mAP/Rank1/mINP baseline by " + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "inline_equation", + "content": "4.99\\% / 2.65\\% / 1.75\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\leftrightarrow \\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "text", + "content": " of CARGO. Besides, VDT also brings different degree of benefits to other CARGO protocols. VDT also surpasses Explain on mAP/Rank1 by " + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "inline_equation", + "content": "3.74\\% / 5.18\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{G} \\rightarrow \\mathrm{A}" + }, + { + "bbox": [ + 304, + 258, + 545, + 619 + ], + "type": "text", + "content": " of AG-ReID, which strongly demonstrates the importance of view decoupled operations in AGPReID to mitigate the disruption of identity representation by view bias whether in homogeneous or heterogeneous matching. (2) Previous view-homogeneous methods show varying degrees of performance degradation in AGPReID, especially on view-heterogeneous protocols, indicating that view bias causes the identity feature under different views to behave with poor cohesion. However, the above methods ignore this issue. (3) The superior performance of VDT does not come from a strong baseline but from the proposed method itself. On the CARGO dataset, the baseline does not consistently outperform the CNN-based AGW on all metrics; on the AGPReID dataset, the baseline is also weaker than the previous Explain. However, VDT achieves the SOTA performance on two datasets, especially on AG-ReID, using fewer prior labels and achieving better performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Speed. We report the inference time for baseline and VDT for retrieving a single image, which is 2.27ms and 2.28ms on the CARGO dataset and 1.44ms and 1.46ms on the AGReID dataset, respectively. These results validate the previous complexity analysis in Sec. 3.2 and show that VDT does not add additional time consumption. Slight time fluctuations tend to be acceptable, due to hardware devices." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22006" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 214, + 194 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 214, + 194 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 214, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 214, + 194 + ], + "type": "image", + "image_path": "78e5419c735f6b59836922ac26d9b3d964ee1d5cce75a1bdbff0f4014036bc6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 196, + 179, + 205 + ], + "lines": [ + { + "bbox": [ + 82, + 196, + 179, + 205 + ], + "spans": [ + { + "bbox": [ + 82, + 196, + 179, + 205 + ], + "type": "text", + "content": "(a) CARGO under Protocol 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 76, + 380, + 194 + ], + "blocks": [ + { + "bbox": [ + 219, + 76, + 380, + 194 + ], + "lines": [ + { + "bbox": [ + 219, + 76, + 380, + 194 + ], + "spans": [ + { + "bbox": [ + 219, + 76, + 380, + 194 + ], + "type": "image", + "image_path": "d900f7e1adbb94863037a0b3432ed0550fe6f2e9ba74aab45e5ec9b42358c353.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 196, + 347, + 205 + ], + "lines": [ + { + "bbox": [ + 246, + 196, + 347, + 205 + ], + "spans": [ + { + "bbox": [ + 246, + 196, + 347, + 205 + ], + "type": "text", + "content": "(b) AG-ReID under Protocol 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": "Figure 5. Fig. " + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "inline_equation", + "content": "5\\mathrm{a}\\sim" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": " Fig. 5c show the relationship between " + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": " and performance on two datasets. For simplicity, only protocol 1 is shown on the CARGO dataset. Rank1, mAP, and mINP are reported (\\%). Avg represents the average performance of Rank1, mAP, and mINP." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 386, + 76, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 386, + 76, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 386, + 76, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 386, + 76, + 545, + 194 + ], + "type": "image", + "image_path": "c3be247d4965a2a69b5b0e74e30708154466d6aa0e4ca9937be193a74ac7d917.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 196, + 512, + 205 + ], + "lines": [ + { + "bbox": [ + 412, + 196, + 512, + 205 + ], + "spans": [ + { + "bbox": [ + 412, + 196, + 512, + 205 + ], + "type": "text", + "content": "(c) AG-ReID under Protocol 2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 258, + 140, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 140, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 140, + 271 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 277, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 373 + ], + "type": "text", + "content": "As described in Sec. 3.2, the core contribution of VDT lies in the decoupling of identity and view features by Eq. (4) and Eq. (9). Therefore, we aim to demonstrate the contribution of each part to VDT, which has been shown in Fig. 4. Similar trends in the two datasets demonstrate that the absence of each part leads to a degradation of the model's performance, and the absence of both simultaneously leads to a more severe performance degradation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "content": "By comparing Row1 and Row2, the absence of Eq. (4) leads to the degradation of " + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "inline_equation", + "content": "t_v" + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "content": " to be equivalent, focusing on the identity and view features in the image respectively, and constraining them to be independent through Eq. (9). However, lacking the guidance of explicit separation within VDT, the decoupling goal cannot be effectively achieved through Eq. (9) alone. Similarly, by comparing Row1 and Row3, the identity and view features outputted by VDT are not guaranteed to be orthogonally separated from each other if Eq. (9) is missing, and thus " + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "content": " containing a minor view bias is still not strongly discriminative." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 514, + 162, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 162, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 162, + 528 + ], + "type": "text", + "content": "5.4. Parameter Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": "As described in Sec. 3.2, VDT introduces only one hyperparameter " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " in Eq. (10) that needs to be manually tuned for multi-task balance. We select six magnitudes of " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " to 10 and the corresponding results are shown in Fig. 5. To show a clear tendency, we show the average performance of Rank1, mAP, and mINP at each " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": ". Although the performance shows the same trend of increasing and then decreasing on both datasets as " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " increases, the peak performance is achieved when " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\lambda = 0.001" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " on CARGO (Fig. 5a) and " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " on AG-ReID (Fig. 5b and Fig. 5c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "The key of this phenomenon is the synchronous updating between identity and view classifier, because the view classifier is a simple binary classification, but the identity classifier on the two datasets shows different difficulties depending on the number of identities. As shown in Tab. 1," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "content": "CARGO owns a thousand-level identities, so the identity classification is relatively complicated, and in order to keep tasks in Eq. (10) able to update synchronously, a small " + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "content": " is more appropriate for view-related objectives. Similarly, a large " + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "content": " is selected because of a hundred-level identities in AGPReID. We also provide an empirical " + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "content": " selection guideline to avoid exhaustive search based on above analysis: " + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "content": " tends to be smaller for datasets with larger identity scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 365, + 470, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 470, + 378 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": "This paper focuses on ReID under a view-heterogeneous scenario, aerial-ground ReID (AGPReID). First, the view-decoupled transformer (VDT) is proposed specifically for the dramatic view differences in AGPReID, which decouples view-related and view-unrelated features by two major operations: hierarchically subtractive separation and orthogonal loss. Second, we contribute a large-scale dataset called CARGO, in which the identities and images are 12.8 and 4.9 times larger than the previous dataset. Experiments on two datasets demonstrate the superiority of VDT and the necessity of view decoupling in AGPReID." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 517, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 577 + ], + "type": "text", + "content": "Although view discrepancy is the most significant factor in AGPReID, it is not the only source of disruptions. Therefore, further exploration of discriminative identity representations under multiple disturbances in the aerial-ground camera network is a promising avenue of future research." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 587, + 403, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 403, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 403, + 601 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 608, + 545, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 545, + 680 + ], + "type": "text", + "content": "This project was supported in part by the NSFC (U22A2095, 62076258), in part by the Key-Area Research and Development Program of Guangzhou (202206030003), in part by Guangdong Project (No. 2020B1515120085), and in part by International Program Fund for Young Talent Scientific Research People, Sun Yat-Sen University." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22007" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Geon Lee, Sanghoon Lee, Dohyung Kim, Younghoon Shin, Yongsang Yoon, and Bumsub Ham. Camera-driven representation learning for unsupervised domain adaptive person re-identification. In Int. Conf. Comput. Vis., pages 11453–11462, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 179 + ], + "type": "text", + "content": "[2] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In ACM Int. Conf. Multimedia, page 3115–3123, 2021. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "text", + "content": "[3] Quan Zhang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling with second-order transformer for group re-identification. In AAAI, volume 36, pages 3318–3325, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "text", + "content": "[4] Quan Zhang, Jianhuang Lai, Xiaohua Xie, Xiaofeng Jin, and Sien Huang. Separable spatial-temporal residual graph for cloth-changing group re-identification. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-16, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[5] Quan Zhang, Kaiheng Dang, Jian-Huang Lai, Zhanxiang Feng, and Xiaohua Xie. Modeling 3d layout for group re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7512-7520, 2022. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "text", + "content": "[6] Jianxiong Tang, Jianhuang Lai, Xiaohua Xie, Lingxiao Yang, and Wei-Shi Zheng. Ac2as: Activation consistency coupled ann-snn framework for fast and memory-efficient snn training. Pattern Recognition, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[7] Qi Chen, Yun Chen, Yuheng Huang, Xiaohua Xie, and Lingxiao Yang. Region-based online selective examination for weakly supervised semantic segmentation. Information Fusion, page 102311, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[8] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7713-7722, June 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Junyang Qiu, Zhanxiang Feng, Lei Wang, and Jianhuang Lai. Salient part-aligned and keypoint disentangling transformer for person re-identification in aerial imagery. In Int. Conf. Multimedia and Expo, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[10] Pengze Zhang, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Pose guided person image generation via dual-task correlation and affinity learning. IEEE Trans. Vis. Comput. Graph., pages 1-18, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "text", + "content": "[11] Jianxiong Tang, Jian-Huang Lai, Xiaohua Xie, and Lingxiao Yang. Spike count maximization for neuromorphic vision recognition. In *IJCAI*, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 602 + ], + "type": "text", + "content": "[12] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Adv. Neural Inform. Process. Syst., 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 647 + ], + "type": "text", + "content": "[13] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4288-4298, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[14] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[15] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Uncertainty modeling for group re-identification. Int. J. Comput. Vis., 2024." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Lingxiao Yang, Ru-Yuan Zhang, Lida Li, and Xiaohua Xie. Simam: A simple, parameter-free attention module for convolutional neural networks. pages 11863-11874. PMLR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[17] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Int. Conf. Comput. Vis., pages 1116–1124, 2015. 1, 2, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[18] Shizhou Zhang, Qi Zhang, Yifei Yang, Xing Wei, Peng Wang, Bingliang Jiao, and Yanning Zhang. Person re-identification in aerial imagery. IEEE Trans. Multimedia, 23:281-291, 2021. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "text", + "content": "[19] Quan Zhang, Jianhuang Lai, and Xiaohua Xie. Learning modal-invariant angular metric by cyclic projection network for vis-nir person re-identification. IEEE Trans. Image Process., 30:8019-8033, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "text", + "content": "[20] Quan Zhang, Jianhuang Lai, Zhanxiang Feng, and Xiaohua Xie. Seeing like a human: Asynchronous learning with dynamic progressive refinement for person re-identification. IEEE Trans. Image Process., 31:352-365, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "type": "text", + "content": "[21] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44(6):2872-2893, 2022. 2, 6, 7, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 371 + ], + "type": "text", + "content": "[22] H. Nguyen, K. Nguyen, S. Sridharan, and C. Fookes. Aerial-ground person re-id. In Int. Conf. Multimedia and Expo, pages 2585-2590, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 372, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 405 + ], + "type": "text", + "content": "[23] Lei Wang, Quan Zhang, Junyang Qiu, and Jianhuang Lai. Rotation exploration transformer for aerial person re-identification. In Int. Conf. Multimedia and Expo, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 405, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 459 + ], + "type": "text", + "content": "[24] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11506–11515, 2021. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 459, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 503 + ], + "type": "text", + "content": "[25] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 79–88, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 503, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 547 + ], + "type": "text", + "content": "[26] Fei Xiong, Mengran Gou, Octavia Camps, and Mario Sznaier. Person re-identification using kernel-based metric learning methods. In Eur. Conf. Comput. Vis., pages 1–16, 2014. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[27] Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. Person re-identification by local maximal occurrence representation and metric learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2197–2206, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "text", + "content": "[28] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In Eur. Conf. Comput. Vis., pages 501-518, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[29] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Int. Conf. Comput. Vis., pages 15013-15022, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22008" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "In IEEE Conf. Comput. Vis. Pattern Recog., pages 16266-16275, 2021. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[31] Shuoyi Chen, Mang Ye, and Bo Du. Rotation invariant transformer for recognizing object in uavs. In ACM Int. Conf. Multimedia, page 2565-2574, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[32] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In ACM Int. Conf. Multimedia, page 3422–3430, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[33] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4890–4899, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "text", + "content": "[34] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2021. 3, 6, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "type": "text", + "content": "[36] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. In ACM Int. Conf. Multimedia, page 9664–9667, 2023. 6, 7, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "text", + "content": "[37] Yifan Sun, Liang Zheng, Yali Li, Yi Yang, Qi Tian, and Shengjin Wang. Learning part-based convolutional features for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 43:902–917, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[38] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2019. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[39] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM Int. Conf. Multimedia, pages 274–282, 2018. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[40] Ratnesh Kuma, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. Vehicle re-identification: an efficient baseline using triplet embedding. In International Joint Conference on Neural Networks, pages 1–9, 2019. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "type": "text", + "content": "[41] Ratnesh Kumar, Edwin Weill, Farzin Aghdasi, and Parthasarathy Sriram. A strong and efficient baseline for vehicle re-identification using deep triplet embedding. Journal of Artificial Intelligence and Soft Computing Research, 10(1):27-45, 2020. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "type": "text", + "content": "[42] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Learning generalisable omni-scale representations for person re-identification. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):5056–5069, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[43] Leyde Briceno and Gunther Paul. Makehuman: a review of the modelling framework. In Congress of the International Ergonomics Association, pages 224-232, 2018. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[44] Unity Technologies. Unity3D: Cross-platform 3D engine, 2021. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 249 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[45] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248–255, 2009. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[46] Léon Bottou. Stochastic gradient descent tricks. Neural Networks: Tricks of the Trade: Second Edition, pages 421-436, 2012. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 249 + ], + "type": "text", + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., pages 8024-8035, 2019. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22009" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_content_list.json b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43893be612adef4cedf0e29869598a7db192d3bd --- /dev/null +++ b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_content_list.json @@ -0,0 +1,1337 @@ +[ + { + "type": "text", + "text": "ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lukas Hollein $^{1,2}$ Aljaž Božić $^{2}$ Norman Müller $^{2}$ David Novotny $^{2}$ Hung-Yu Tseng $^{2}$ Christian Richardt $^{2}$ Michael Zollhöfer $^{2}$ Matthias Nießner $^{1}$", + "bbox": [ + 122, + 178, + 844, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich $^{2}$ Meta https://lukashoel.github.io/ViewDiff/", + "bbox": [ + 263, + 223, + 702, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2d37ef423678c9073328eb73feb890f65acb4e8a171a86435fa42a5d35a9e53.jpg", + "image_caption": [ + "Figure 1. Multi-view consistent image generation. Our method takes as input a text description, or any number of posed input images, and generates high-quality, multi-view consistent images of a real-world 3D object in authentic surroundings from any desired camera poses." + ], + "image_footnote": [], + "bbox": [ + 76, + 280, + 895, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 627, + 313, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D asset generation is getting massive amounts of attention, inspired by the recent success of text-guided 2D content creation. Existing text-to-3D methods use pretrained text-to-image diffusion models in an optimization problem or fine-tune them on synthetic data, which often results in nonphotorealistic 3D objects without backgrounds. In this paper, we present a method that leverages pretrained text-to-image models as a prior, and learn to generate multi-view images in a single denoising process from real-world data. Concretely, we propose to integrate 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network of the text-to-image model. Moreover, we design an autoregressive generation that renders more 3D-consistent images at any viewpoint. We train our model on real-world datasets of objects and showcase its capabilities to generate instances with a variety of high-quality shapes and textures", + "bbox": [ + 73, + 659, + 473, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in authentic surroundings. Compared to the existing methods, the results generated by our method are consistent, and have favorable visual quality $(-30\\%$ FID, $-37\\%$ KID).", + "bbox": [ + 496, + 603, + 893, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 662, + 632, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, text-to-image (T2I) diffusion models [29, 33] have emerged as cutting-edge technologies, revolutionizing high-quality and imaginative 2D content creation guided by text descriptions. These frameworks have found widespread applications, including extensions such as ControlNet [53] and DreamBooth [32], showcasing their versatility and potential. An intriguing direction in this domain is to use T2I models as powerful 2D priors for generating three-dimensional (3D) assets. How can we effectively use these models to create photo-realistic and diverse 3D assets?", + "bbox": [ + 496, + 688, + 893, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods like DreamFusion [27], Fantasy3D [5], and ProlificDreamer [47] have demonstrated exciting results by optimizing a 3D representation through score distillation sampling [27] from pretrained T2I diffusion models. The", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5043", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D assets generated by these methods exhibit compelling diversity. However, their visual quality is not consistently as high as that of the images generated by T2I models. A key step to obtaining 3D assets is the ability to generate consistent multi-view images of the desired objects and their surroundings. These images can then be fitted to 3D representations like NeRF [25] or NeuS [45]. HoloDiffusion [19] and ViewsetDiffusion [38] train a diffusion model from scratch using multi-view images and output 3D-consistent images. GeNVS [4] and DFM [41] additionally produce object surroundings, thereby increasing the realism of the generation. These methods ensure (photo)-realistic results by training on real-world 3D datasets [31, 56]. However, these datasets are orders of magnitude smaller than the 2D dataset used to train T2I diffusion models. As a result, these approaches produce realistic but less-diverse 3D assets. Alternatively, recent works like Zero-1-to-3 [23] and One-2-3-45 [22] leverage a pretrained T2I model and fine-tune it for 3D consistency. These methods successfully preserve the diversity of generated results by training on a large synthetic 3D dataset [7]. Nonetheless, the produced objects can be less photo-realistic and are without surroundings.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a method that leverages the 2D priors of the pretrained T2I diffusion models to produce photo-realistic and 3D-consistent 3D asset renderings. As shown in the first two rows of Fig. 1, the input is a text description or an image of an object, along with the camera poses of the desired rendered images. The proposed approach produces multiple images of the same object in a single forward pass. Moreover, we design an autoregressive generation scheme that allows to render more images at any novel viewpoint (Fig. 1, third row). Concretely, we introduce projection and cross-frame-attention layers, that are strategically placed into the existing U-Net architecture, to encode explicit 3D knowledge about the generated object (see Fig. 2). By doing so, our approach paves the way to fine-tune T2I models on real-world 3D datasets, such as CO3D [31], while benefiting from the large 2D prior encoded in the pretrained weights. Our generated images are consistent, diverse, and realistic renderings of objects.", + "bbox": [ + 76, + 425, + 472, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "text_level": 1, + "bbox": [ + 76, + 699, + 323, + 714 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a method that utilizes the pretrained 2D prior of text-to-image models and turns them into 3D-consistent image generators. We train our approach on real-world multi-view datasets, allowing us to produce realistic and high-quality images of objects and their surroundings (Sec. 3.1).", + "- a novel U-Net architecture that combines commonly used 2D layers with 3D-aware layers. Our projection and cross-frame-attention layers encode explicit 3D knowledge into each block of the U-Net architecture (Sec. 3.2).", + "- an autoregressive generation scheme that renders images of a 3D object from any desired viewpoint directly with our model in a 3D-consistent way (Sec. 3.3)." + ], + "bbox": [ + 76, + 715, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-To-2D. Denoising diffusion probabilistic models (DDPM) [14] model a data distribution by learning to invert a Gaussian noisin process with a deep network. Recently, DDPMs were shown to be superior to generative adversarial networks [8], becoming the state-of-the-art framework for image generation. Soon after, large text-conditioned models trained on billion-scale data were proposed inImagen [33] or Dall-E 2 [29]. While [8] achieved conditional generation via guidance with a classifier, [13] proposed classifier-free guidance. ControlNet [53] proposed a way to tune the diffusion outputs by conditioning on various modalities, such as image segmentation or normal maps. Similar to ControlNet, our method builds on the strong 2D prior of a pretrained text-to-image (T2I) model. We further demonstrate how to adjust this prior to generate 3D-consistent images of objects.", + "bbox": [ + 496, + 117, + 895, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-To-3D. 2D DDPMs were applied to the generation of 3D shapes [28, 34, 39, 42, 44, 50, 58] or scenes [10, 15, 40] from text descriptions. DreamFusion [27] proposed score distillation sampling (SDS) which optimizes a 3D shape whose renders match the belief of the DDPM. Improved sample quality was achieved by a second-stage mesh optimization [5, 21], and smoother SDS convergence [35, 47]. Several methods use 3D data to train a novel-view synthesis model whose multi-view samples can be later converted to 3D, e.g. conditioning a 2D DDPM on an image and a relative camera motion to generate novel views [23, 48]. However, due to no explicit modelling of geometry, the outputs are view-inconsistent. Consistency can be improved with epipolar attention [43, 57], or optimizing a 3D shape from multi-view proposals [22]. Our work fine-tunes a 2D T2I model to generate renders of a 3D object; however, we propose explicit 3D unprojection and rendering operators to improve view-consistency. Concurrently, SyncDreamer [24] also add 3D layers in their 2D DDPM. We differ by training on real data with backgrounds and by showing that autoregressive generation is sufficient to generate consistent images, making the second 3D reconstruction stage expendable.", + "bbox": [ + 496, + 358, + 895, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion on 3D Representations. Several works model the distribution of 3D representations. While DiffRF [26] leverages ground-truth 3D shapes, HoloDiffusion [19] is supervised only with 2D images. HoloFusion [18] extends this work with a 2D diffusion render post-processor. Images can also be denoised by rendering a reconstructing 3D shape [1, 38]. Unfortunately, the limited scale of existing 3D datasets prevents these 3D diffusion models from extrapolating beyond the training distribution. Instead, we exploit a large 2D pretrained DDPM and add 3D components that are tuned on smaller-scale multi-view data. This leads to improved multi-view consistency while maintaining the expressivity brought by pretraining on billion-scale image data.", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5044", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/205bf188fa66e3171a5254ac3f40c63e7803ea1955b2a89fa4645c637176a595.jpg", + "image_caption": [ + "Figure 2. Method Overview. We augment the U-Net architecture of pretrained text-to-image models with new layers in every U-Net block. These layers facilitate communication between multi-view images in a batch, resulting in a denoising process that jointly produces 3D-consistent images. First, we replace self-attention with cross-frame-attention (yellow) which compares the spatial features of all views. We condition all attention layers on pose $(RT)$ , intrinsics $(K)$ , and intensity $(I)$ of each image. Second, we add a projection layer (green) into the inner blocks of the U-Net. It creates a 3D representation from multi-view features and renders them into 3D-consistent features. We fine-tune the U-Net using the diffusion denoising objective (Eq. 3) at timestep $t$ , supervised from captioned multi-view images." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 893, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 401, + 171, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a method that produces 3D-consistent images from a given text or posed image input (see Fig. 1 top/mid). Concretely, given desired output poses, we jointly generate all images corresponding to the condition. We leverage pretrained text-to-image (T2I) models [29, 33] and fine-tune them on multi-view data [31]. We propose to augment the existing U-Net architecture by adding new layers into each block (see Fig. 2). At test time, we can condition our method on multiple images (see Fig. 1 bottom), which allows us to autoregressively render the same object from any viewpoint directly with the diffusion model (see Sec. 3.3).", + "bbox": [ + 75, + 426, + 473, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D-Consistent Diffusion", + "text_level": 1, + "bbox": [ + 76, + 603, + 299, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [14, 36] are a class of generative models that learn the probability distribution $p_{\\theta}(x_0) = \\int p_{\\theta}(x_{0:T}) dx_{1:T}$ over data $x_0 \\sim q(x_0)$ and latent variables $x_{1:T} = x_1, \\ldots, x_T$ . Our method is based on pretrained text-to-image models, which are diffusion models $p_{\\theta}(x_0 \\mid c)$ with an additional text condition $c$ . For clarity, we drop the condition $c$ for the remainder of this section.", + "bbox": [ + 75, + 626, + 470, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To produce multiple images $x_0^{0:N}$ at once, which are 3D-consistent with each other, we seek to model their joint probability distribution $p_{\\theta}(x_0^{0:N}) = \\int p_{\\theta}(x_0^{0:N}) dx_{1:T}^{0:N}$ . Similarly to concurrent work by Liu et al. [24], we generate one set of images $p_{\\theta}(x_0^{0:N})$ by adapting the reverse process of DDPMs [14] as a Markov chain over all images jointly:", + "bbox": [ + 75, + 732, + 470, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0: T} ^ {0: N}\\right) := p \\left(x _ {T} ^ {0: N}\\right) \\prod_ {t = 1} ^ {T} \\prod_ {n = 0} ^ {N} p _ {\\theta} \\left(x _ {t - 1} ^ {n} \\mid x _ {t} ^ {0: N}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 833, + 470, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we start the generation from Gaussian noise sam", + "bbox": [ + 76, + 885, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pled separately per image $p(x_{T}^{n}) = \\mathcal{N}(x_{T}^{n};\\mathbf{0},\\mathbf{I}),\\forall n\\in [0,N]$ . We gradually denoise samples $p_{\\theta}(x_{t - 1}^{n}\\mid x_{t}^{0:N}) = \\mathcal{N}(x_{t - 1};\\mu_{\\theta}^{n}(x_{t}^{0:N},t),\\sigma_{t}^{2}\\mathbf{I})$ by predicting the per-image mean $\\mu_{\\theta}^{n}(x_{t}^{0:N},t)$ through a neural network $\\mu_{\\theta}$ that is shared between all images. Importantly, at each step, the model uses the previous states $x_{t}^{0:N}$ of all images, i.e., there is communication between images during the model prediction. We refer to Sec. 3.2 for details on how this is implemented. To train $\\mu_{\\theta}$ , we define the forward process as a Markov chain:", + "bbox": [ + 496, + 402, + 893, + 540 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {1: T} ^ {0: N} \\mid x _ {0} ^ {0: N}\\right) = \\prod_ {t = 1} ^ {T} \\prod_ {n = 0} ^ {N} q \\left(x _ {t} ^ {n} \\mid x _ {t - 1} ^ {n}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 542, + 892, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q(x_{t}^{n} \\mid x_{t - 1}^{n}) = \\mathcal{N}(x_{t}^{n}; \\sqrt{1 - \\beta_{t}} x_{t - 1}^{n}, \\beta_{t}\\mathbf{I})$ and $\\beta_{1}, \\ldots, \\beta_{T}$ define a constant variance schedule, i.e., we apply separate noise per image to produce training samples.", + "bbox": [ + 496, + 593, + 893, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow Ho et al. [14] by learning a noise predictor $\\epsilon_{\\theta}$ instead of $\\mu_{\\theta}$ . This allows to train $\\epsilon_{\\theta}$ with an $L2$ loss:", + "bbox": [ + 496, + 640, + 893, + 670 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {x _ {0} ^ {0: N}, \\epsilon^ {0: N}} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), n \\left[ \\left\\| \\epsilon^ {n} - \\epsilon_ {\\theta} ^ {n} \\left(x _ {t} ^ {0: N}, t\\right) \\right\\| ^ {2} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 676, + 892, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Augmentation of the U-Net architecture", + "text_level": 1, + "bbox": [ + 500, + 710, + 841, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To model a 3D-consistent denoising process over all images, we predict per-image noise $\\epsilon_{\\theta}^{n}(x_{t}^{0:N}, t)$ through a neural network $\\epsilon_{\\theta}$ . This neural network is initialized from the pretrained weights of existing text-to-image models, and is usually defined as a U-Net architecture [29, 33]. We seek to leverage the previous states $x_{t}^{0:N}$ of all images to arrive at a 3D-consistent denoising step. To this end, we propose to add two layers into the U-Net architecture, namely a cross-frame-attention layer and a projection layer. We note that the predicted per-image noise needs to be image specific, since all images are generated starting from separate Gaussian", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "noise. It is therefore important to keep around 2D layers that act separately on each image, which we achieve by finetuning the existing ResNet [11] and ViT [9] blocks. We summarize our architecture in Fig. 2. In the following, we discuss our two proposed layers in more detail.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Frame Attention. Inspired by video diffusion [49, 51], we add cross-frame-attention layers into the U-Net architecture. Concretely, we modify the existing self-attention layers to calculate $CFAtn(Q,K,V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right)V$ with", + "bbox": [ + 75, + 170, + 470, + 238 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = W ^ {Q} h _ {i}, K = W ^ {K} \\left[ h _ {j} \\right] _ {j \\neq i}, V = W ^ {V} \\left[ h _ {j} \\right] _ {j \\neq i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 256, + 470, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W^{Q}, W^{K}, W^{V}$ are the pretrained weights for feature projection, and $h_{i} \\in \\mathbb{R}^{C \\times H \\times W}$ is the input spatial feature of each image $i \\in [1, N]$ . Intuitively, this matches features across all frames, which allows generating the same global style.", + "bbox": [ + 75, + 284, + 468, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we add a conditioning vector to all cross-frame and cross-attention layers to inform the network about the viewpoint of each image. First, we add pose information by encoding each image's camera matrix $p \\in \\mathbb{R}^{4 \\times 4}$ into an embedding $z_1 \\in \\mathbb{R}^4$ similar to Zero-1-to-3 [23]. Additionally, we concatenate the focal length and principal point of each camera into an embedding $z_2 \\in \\mathbb{R}^4$ . Finally, we provide an intensity encoding $z_3 \\in \\mathbb{R}^2$ , which stores the mean and variance of the image RGB values. At training time, we set $z_3$ to the true values of each input image, and at test time, we set $z_3 = [0.5, 0]$ for all images. This helps to reduce the view-dependent lighting differences contained in the dataset (e.g., due to different camera exposure). We construct the conditioning vector as $z = [z_1, z_2, z_3]$ , and add it through a LoRA-linear-layer [16] $W'^Q$ to the feature projection matrix $Q$ . Concretely, we compute the projected features as:", + "bbox": [ + 75, + 345, + 470, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = W ^ {Q} h _ {i} + s \\cdot W ^ {\\prime Q} [ h _ {i}; z ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 595, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we set $s = 1$ . Similarly, we add the condition via $W^{\\prime K}$ to $K$ , and $W^{\\prime V}$ to $V$ .", + "bbox": [ + 75, + 625, + 467, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Projection Layer. Cross-frame attention layers are helpful to produce globally 3D-consistent images. However, the objects do not precisely follow the specified poses, which leads to view-inconsistencies (see Fig. 5 and Tab. 3). To this end, we add a projection layer into the U-Net architecture (Fig. 3). The idea of this layer is to create 3D-consistent features that are then further processed by the next U-Net layers (e.g. ResNet blocks). By repeating this layer across all stages of the U-Net, we ensure that the per-image features are in a 3D-consistent space. We do not add the projection layer to the first and last U-Net blocks, as we saw no benefit from them at these locations. We reason that the network processes image-specific information at those stages and thus does not need a 3D-consistent feature space.", + "bbox": [ + 75, + 659, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by multi-view stereo literature [3, 17, 37], we create a 3D feature voxel grid from all input spatial features", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/56b3342c3f17dc641e6334c87d8732ad9f2e90747fb268bda54a236c8601c825.jpg", + "image_caption": [ + "Figure 3. Architecture of the projection layer. We produce 3D-consistent output features from posed input features. First, we unproject the compressed image features into 3D and aggregate them into a joint voxel grid with an MLP. Then we refine the voxel grid with a 3D CNN. A volume renderer similar to NeRF [25] renders 3D-consistent features from the grid. Finally, we apply a learned scale function and expand the feature dimension." + ], + "image_footnote": [], + "bbox": [ + 509, + 87, + 893, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$h_{\\mathrm{in}}^{0:N} \\in \\mathbb{R}^{C \\times H \\times W}$ by projecting each voxel into each image plane. First, we compress $h_{\\mathrm{in}}^{0:N}$ with a $1 \\times 1$ convolution to a reduced feature dimension $C' = 16$ . We then take the bilinearly interpolated feature at the image plane location and place it into the voxel. This way, we create a separate voxel grid per view, and merge them into a single grid through an aggregator MLP. Inspired by IBRNet [46], the MLP predicts per-view weights followed by a weighted feature average. We then run a small 3D CNN on the voxel grid to refine the 3D feature space. Afterwards, we render the voxel grid into output features $h_{\\mathrm{out}}^{0:N} \\in \\mathbb{R}^{C' \\times H \\times W}$ with volumetric rendering similar to NeRF [25]. We dedicate half of the voxel grid to foreground and half to background and apply the background model from MERF [30] during ray-marching.", + "bbox": [ + 496, + 454, + 893, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We found it is necessary to add a scale function after the volume rendering output. The volume renderer typically uses a sigmoid activation function as the final layer during ray-marching [25]. However, the input features are defined in an arbitrary floating-point range. To convert $h_{\\mathrm{out}}^{0:N}$ back into the same range, we non-linearly scale the features with $1 \\times 1$ convolutions and $ReLU$ activations. Finally, we expand $h_{\\mathrm{out}}^{0:N}$ to the input feature dimension $C$ . We refer to the supplemental material for details about each component's architecture.", + "bbox": [ + 496, + 667, + 893, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Autoregressive Generation", + "text_level": 1, + "bbox": [ + 500, + 816, + 741, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method takes as input multiple samples $x_{t}^{0:N}$ at once and denoises them 3D-consistently. During training, we set $N = 5$ , but can increase it at inference time up to memory constraints, e.g., $N = 30$ . However, we want to render an", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5046", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "object from any possible viewpoint directly with our network. To this end, we propose an autoregressive image generation scheme, i.e., we condition the generation of next viewpoints on previously generated images. We provide the timesteps $t^{0:N}$ of each image as input to the U-Net. By varying $t^{0:N}$ , we can achieve different types of conditioning.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unconditional Generation. All samples are initialized to Gaussian noise and are denoised jointly. The timesteps $t^{0:N}$ are kept identical for all samples throughout the reverse process. We provide different cameras per image and a single text prompt. The generated images are 3D-consistent, showing the object from the desired viewpoints (Figs. 4 and 5).", + "bbox": [ + 75, + 186, + 470, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-Conditional Generation. We divide the total number of samples $N = n_{\\mathrm{c}} + n_{\\mathrm{g}}$ into a conditional part $n_{\\mathrm{c}}$ and generative part $n_{\\mathrm{g}}$ . The first $n_{\\mathrm{c}}$ samples correspond to images and cameras that are provided as input. The other $n_{\\mathrm{g}}$ samples should generate novel views that are similar to the conditioning images. We start the generation from Gaussian noise for the $n_{\\mathrm{g}}$ samples and provide the un-noised images for the other samples. Similarly, we set $t^{0:n_{\\mathrm{c}}} = 0$ for all denoising steps, while gradually decreasing $t^{n_{\\mathrm{g}}:N}$ .", + "bbox": [ + 75, + 285, + 470, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When $n_{\\mathrm{c}} = 1$ , our method performs single-image reconstruction (Fig. 6). Setting $n_{\\mathrm{c}} > 1$ allows to autoregressively generate novel views from previous images (Fig. 1 bottom). In practice, we first generate one batch of images unconditionally and then condition the next batches on a subset of previous images. This allows us to render smooth trajectories around 3D objects (see the supplemental material).", + "bbox": [ + 75, + 421, + 470, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 537, + 294, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We train our method on the large-scale CO3Dv2 [31] dataset, which consists of posed multi-view images of real-world objects. Concretely, we choose the categories Teddybear, Hydrant, Apple, and Donut. Per category, we train on 500-1000 objects with each 200 images at resolution $256 \\times 256$ . We generate text captions with the BLIP-2 model [20] and sample one of 5 proposals per object.", + "bbox": [ + 75, + 561, + 470, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We base our model on a pretrained latent-diffusion text-to-image model. We only fine-tune the U-Net and keep the VAE encoder and decoder frozen. In each iteration, we select $N = 5$ images and their poses. We sample one denoising timestep $t \\sim [0,1000]$ , add noise to the images according to Eq. 2, and compute the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, weskip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of unconditional image generation. We report average FID [12] and KID [2] per category and improve by a significant margin. This signals that our images are more similar to the distribution of real images in the dataset. We mask away the background for our method and the real images to ensure comparability of numbers with the baselines.", + "bbox": [ + 496, + 89, + 893, + 172 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/343b0bcaf9eeb4428f993b80887b92fdf83b7b790a1ffea54ccce150324f21d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryHF [18]VD [38]Ours
FID↓KID↓FID↓KID↓FID↓KID↓
Teddybear81.930.072201.710.16949.390.036
Hydrant61.190.042138.450.11846.450.033
Donut105.970.091199.140.13668.860.054
Apple62.190.056183.670.14956.850.043
", + "bbox": [ + 504, + 183, + 885, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We fine-tune the model on $2 \\times$ A100 GPUs for 60K iterations (7 days) with a total batch size of 64. We set the learning rate for the volume renderer to 0.005 and for all other layers to $5 \\times 10^{-5}$ , and use the AdamW optimizer [33]. During inference, we can increase $N$ and generate up to 30 images/batch on an RTX 3090 GPU. We use the UniPC [55] sampler with 10 denoising steps, which takes 15 seconds.", + "bbox": [ + 496, + 321, + 893, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 443, + 586, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare against recent state-of-the-art works for 3D generative modeling. Our goal is to create multi-view consistent images from real-world, realistic objects with authentic surroundings. Therefore, we consider methods that are trained on real-world datasets and select HoloFusion (HF) [18], ViewsetDiffusion (VD) [38], and DFM [41]. We show results on two tasks: unconditional generation (Sec. 4.1) and single-image reconstruction (Sec. 4.2).", + "bbox": [ + 496, + 468, + 893, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We report FID [12] and KID [2] as common metrics for 2D/3D generation and measure the multi-view consistency of generated images with peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and LPIPS [54]. To ensure comparability, we evaluate all metrics on images without backgrounds, as not every baseline models them.", + "bbox": [ + 496, + 594, + 893, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Unconditional Generation", + "text_level": 1, + "bbox": [ + 500, + 695, + 736, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method can be used to generate 3D-consistent views of an object from any pose with only text as input by using our autoregressive generation (Sec. 3.3). Concretely, we sample an (unobserved) image caption from the test set for the first batch and generate $N = 10$ images with a guidance scale [13] of $\\lambda_{\\mathrm{cfg}} = 7.5$ . We then set $\\lambda_{\\mathrm{cfg}} = 0$ for subsequent batches, and create a total of 100 images per object.", + "bbox": [ + 496, + 719, + 893, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate against HoloFusion (HF) [18] and Viewset-Diffusion (VD) [38]. We report quantitative results in Tab. 1 and qualitative results in Figs. 4 and 5. HF [18] creates diverse images that sometimes show view-dependent floating artifacts (see Fig. 5). VD [38] creates consistent but blurry", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5aa021e3a6fe15e71389213ce20c4d047b73d8a93acf00d9dbdefe44c7795f3b.jpg", + "image_caption": [ + "Figure 4. Unconditional image generation of our method and baselines. We show renderings from different viewpoints for multiple objects and categories. Our method produces consistent objects and backgrounds. Our textures are sharper in comparison to baselines. Please see the supplemental material for more examples and animations." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 911, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. In contrast, our method produces images with backgrounds and higher-resolution object details.", + "bbox": [ + 75, + 765, + 470, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Single-Image Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 814, + 334, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method can be conditioned on multiple images in order to render any novel view in an autoregressive fashion (Sec. 3.3). To measure the 3D-consistency of our generated images, we compare single-image reconstruction against", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ViewsetDiffusion (VD) [38] and DFM [41]. Concretely, we sample one image from the dataset and generate 20 images at novel views also sampled from the dataset. We follow Szymanowicz et al. [38] and report the per-view maximum PSNR/SSIM and average LPIPS across multiple objects and viewpoints for all methods. We report quantitative results in Tab. 2 and show qualitative results in Fig. 6. VD [38] creates plausible results without backgrounds. DFM [41] creates", + "bbox": [ + 496, + 765, + 890, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5ea05c6650f41156fffecf58689e42a1bbebdc3e84ca3fd9144bf80cb33f2d7.jpg", + "image_caption": [ + "Figure 5. Multi-view consistency of unconditional image generation. HoloFusion (HF) [18] has view-dependent floating artifacts (the base in first row). ViewsetDiffusion (VD) [38] has blurrier renderings (second row). Without the projection layer, our method has no precise control over viewpoints (third row). Without cross-frame-attention, our method suffers from identity changes of the object (fourth row). Our full method produces detailed images that are 3D-consistent (fifth row)." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 911, + 452 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c4236fc099827a5eb22c476fd12737db107919ef5d1aee5f5842d969c3e3756d.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of single-image reconstruction. Given a single image as input, we measure the quality of novel views through average PSNR, SSIM, and LPIPS [54] per category. We mask away the generated backgrounds to ensure comparability across all methods. We improve over VD [38] while being on-par with DFM [41]." + ], + "table_footnote": [], + "table_body": "
MethodTeddybearHydrantDonutApple
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
VD [38]19.680.700.3022.360.800.1918.270.680.1419.540.640.31
DFM [41]21.810.820.1622.670.830.1223.910.860.1025.790.910.07
Ours21.980.840.1322.490.850.1121.500.850.1825.940.910.11
", + "bbox": [ + 91, + 585, + 874, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "consistent results with backgrounds at a lower image resolution $(128 \\times 128)$ . Our method produces higher resolution images with similar reconstruction results and backgrounds.", + "bbox": [ + 75, + 709, + 470, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 75, + 765, + 189, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The key ingredients of our method are the cross-frame-attention and projection layers that we add to the U-Net (Sec. 3.2). We highlight their importance in Tab. 3 and Fig. 5.", + "bbox": [ + 75, + 789, + 470, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How important are the projection layers? They are necessary to allow precise control over the image viewpoints (e.g., Fig. 5 row 3 does not follow the specified rotation). Our goal is to generate a consistent set of images from any", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "viewpoint directly with our model (Sec. 3.3). Being able to control the pose of the object is therefore an essential part of our contribution. The projection layers build up a 3D representation of the object that is explicitly rendered into 3D-consistent features through volume rendering. This allows us to achieve viewpoint consistency, as also demonstrated through single-image reconstruction (Tab. 3).", + "bbox": [ + 496, + 709, + 893, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How important are cross-frame-attention layers? They are necessary to create images of the same object. Without them, the teddybear in Fig. 5 (row 4) has the same general color scheme and follows the specified poses. However, differences in shape and texture lead to an inconsistent set of", + "bbox": [ + 496, + 824, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5049", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/690d863621e40efdb937860b154165fad4af9996f68fd48ac3b8fdd1cea3e0d3.jpg", + "image_caption": [ + "Input VD [38] DFM [41] Ours Real Image" + ], + "image_footnote": [], + "bbox": [ + 80, + 102, + 485, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/284a0f9f5d00ff3ca9ac962edc64882e09415e50894d94c9287ad6ac811f0282.jpg", + "image_caption": [ + "Figure 6. Single-image reconstruction of our method and baselines. Given one image/pose as input, our method produces plausible novel views that are consistent with the real shape and texture. We can also produce detailed backgrounds that match the input." + ], + "image_footnote": [], + "bbox": [ + 80, + 166, + 485, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52a9f32e95e4b50b008dd4be85b353ab8479d5c08108858c25235119714b0a6e.jpg", + "image_caption": [ + "[C] wearing a green hat", + "[C] sits on a green blanket" + ], + "image_footnote": [], + "bbox": [ + 78, + 310, + 488, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ac938690c40fc05dcab0f0e164f8bd350f9158c82ae26715e70207a667b3082.jpg", + "image_caption": [ + "rusty $[C]$ in the grass", + "yellow and white [C]", + "Figure 7. Diversity of generated results. We condition our method on text input, which allows to create objects in a desired style. We show samples for hand-crafted text descriptions that combine attributes (e.g., color, shape, background) in a novel way. Each row shows a different generation proposal from our method and we denote the object category (Teddybear, Hydrant) as [C]. This showcases the diversity of generated results, i.e., multiple different objects are generated for the same description." + ], + "image_footnote": [], + "bbox": [ + 78, + 476, + 488, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 292, + 476, + 488, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images. We reason that the cross-frame-attention layers are essential for defining a consistent object identity.", + "bbox": [ + 75, + 787, + 468, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does the 2D prior help? We utilize a 2D prior in form of the pretrained text-to-image model that we fine-tune in a 3D-consistent fashion (Sec. 3.1). This enables our method to produce sharp and detailed images of objects from different viewpoints. Also, we train our method on captioned images", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/41f52d429d6ca99f058b3d6ff5034cb1b0623d98c84386379b5752d24b2f9d1b.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of our method and ablations. We report average PSNR, SSIM, LPIPS [54], FID [12], KID [2] over Teddybear and Hydrant categories. We compare against dropping the projection layer (\"no proj\") and cross-frame-attention (\"no cfa\") from the U-Net (see Sec. 3.2). While still producing high-quality images with similar FID/KID, this shows that our proposed layers are necessary to obtain 3D-consistent images." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓FID↓KID↓
Ours (no proj)16.550.710.2947.950.034
Ours (no cfa)18.150.760.2547.930.034
Ours22.240.840.1147.920.034
", + "bbox": [ + 506, + 191, + 883, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to retain the controllable generation through text descriptions (Sec. 3.4). We show the diversity and controllability of our generations in Fig. 7 with hand-crafted text prompts. This highlights that, after finetuning, our model is still faithful to text input, and can combine attributes in a novel way, i.e., our model learns to extrapolate from the training set.", + "bbox": [ + 496, + 294, + 892, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 500, + 395, + 625, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method generates 3D-consistent, high-quality images of diverse objects according to text descriptions or input images. Nevertheless, there are several limitations. First, our method sometimes produces images with slight inconsistency, as shown in the supplement. Since the model is fine-tuned on a real-world dataset consisting of view-dependent effects (e.g., exposure changes), our framework learns to generate such variations across different viewpoints. A potential solution is to add lighting condition through a ControlNet [53]. Second, our work focuses on objects, but similarly scene-scale generation on large datasets [6, 52] can be explored.", + "bbox": [ + 496, + 417, + 893, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 599, + 617, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented ViewDiff, a method that, given text or image input, generates 3D-consistent images of real-world objects placed in authentic surroundings. Our method leverages the expressivity of large 2D text-to-image models and fine-tunes this 2D prior on real-world 3D datasets to produce diverse multi-view images in a joint denoising process. The core insight of our work are two novel layers, namely cross-frame-attention and the projection layer (Sec. 3.2). Our autoregressive generation scheme (Sec. 3.3) allows to directly render high-quality novel views of a generated 3D object.", + "bbox": [ + 496, + 625, + 893, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 791, + 687, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was done during Lukas' internship at Meta Reality Labs Zurich as well as at TU Munich, funded by a Meta sponsored research agreement. Matthias Nießner was also supported by the ERC Starting Grant Scan2CAD (804724). We also thank Angela Dai for the video voiceover.", + "bbox": [ + 496, + 816, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5050", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2", + "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 5, 8", + "[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. TransformerFusion: Monocular RGB scene reconstruction using transformers. In NeurIPS, 2021. 4", + "[4] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3D-aware diffusion models. arXiv:2304.02602, 2023. 2", + "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 1, 2", + "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 8", + "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 2", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \\times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 4", + "[10] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-NeRF2NeRF: Editing 3D scenes with instructions. In ICCV, 2023. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 5, 8", + "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 2, 5", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3", + "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3D meshes from 2D text-to-image models. In ICCV, 2023. 2", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Atlas: Few-shot learning with retrieval augmented language models. JMLR, 24(251):1-43, 2023. 4", + "[18] Animesh Karnewar, Niloy J. Mitra, Andrea Vedaldi, and David Novotny. HoloFusion: Towards photo-realistic 3D generative modeling. In ICCV, 2023. 2, 5, 6, 7", + "[19] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J. Mitra. HoloDiffusion: Training a 3D diffusion model using 2D images. In CVPR, 2023. 2", + "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 5", + "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. In CVPR, 2023. 2", + "[22] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. arXiv:2306.16928, 2023. 2", + "[23] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2, 4", + "[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 4", + "[26] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3D radiance field diffusion. In CVPR, 2023. 2", + "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 1, 2", + "[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Ali-aksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv:2306.17843, 2023. 2", + "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv:2204.06125, 2022. 1, 2, 3", + "[30] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5051", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 4", + "[31] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2, 3, 5", + "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5", + "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2, 3, 5", + "[34] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2D diffusion model know 3D-consistency for robust text-to-3D generation. arXiv:2303.07937, 2023. 2", + "[35] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. arXiv:2308.16512, 2023. 2", + "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3", + "[37] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In CVPR, 2021. 4", + "[38] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. ViewSet diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 2, 5, 6, 7, 8", + "[39] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3D: High-fidelity 3D creation from a single image with diffusion prior. In ICCV, 2023. 2", + "[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. MVDiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 2", + "[41] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. In NeurIPS, 2023. 2, 5, 6, 7, 8", + "[42] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. TextMesh: Generation of realistic 3D meshes from text prompts. In 3DV, 2024. 2", + "[43] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2", + "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian chaining: Lifting pretrained 2D diffusion models for 3D generation. In CVPR, 2023. 2" + ], + "bbox": [ + 78, + 92, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2", + "[46] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 4", + "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. arXiv:2305.16213, 2023. 1, 2", + "[48] Daniel Watson, William Chan, Ricardo Martin Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2", + "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 4", + "[50] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3D-aware image generation using 2D diffusion models. In ICCV, 2023. 2", + "[51] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 4", + "[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 8", + "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 8", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 7, 8", + "[55] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. UniPC: A unified predictor-corrector framework for fast sampling of diffusion models. In NeurIPS, 2023. 5", + "[56] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Transactions on Graphics, 37(4):65:1-12, 2018. 2", + "[57] Zhizhuo Zhou and Shubham Tulsiani. SparseFusion: Distilling view-conditioned diffusion for 3D reconstruction. In CVPR, 2023. 2", + "[58] Joseph Zhu and Peiye Zhuang. HiFA: High-fidelity text-to-3D with advanced diffusion guidance. arXiv:2305.18766, 2023.2" + ], + "bbox": [ + 501, + 92, + 893, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5052", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_model.json b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..18ea77ab8f3b0e4629fdfda96b0b17aaea78be07 --- /dev/null +++ b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.843, + 0.154 + ], + "angle": 0, + "content": "ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.179, + 0.846, + 0.217 + ], + "angle": 0, + "content": "Lukas Hollein\\(^{1,2}\\) Aljaž Božić\\(^{2}\\) Norman Müller\\(^{2}\\) David Novotny\\(^{2}\\) Hung-Yu Tseng\\(^{2}\\) Christian Richardt\\(^{2}\\) Michael Zollhöfer\\(^{2}\\) Matthias Nießner\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.224, + 0.704, + 0.26 + ], + "angle": 0, + "content": "\\(^{1}\\)Technical University of Munich \\(^{2}\\)Meta https://lukashoel.github.io/ViewDiff/" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.281, + 0.896, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.567, + 0.893, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Multi-view consistent image generation. Our method takes as input a text description, or any number of posed input images, and generates high-quality, multi-view consistent images of a real-world 3D object in authentic surroundings from any desired camera poses." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.628, + 0.314, + 0.643 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.66, + 0.474, + 0.902 + ], + "angle": 0, + "content": "3D asset generation is getting massive amounts of attention, inspired by the recent success of text-guided 2D content creation. Existing text-to-3D methods use pretrained text-to-image diffusion models in an optimization problem or fine-tune them on synthetic data, which often results in nonphotorealistic 3D objects without backgrounds. In this paper, we present a method that leverages pretrained text-to-image models as a prior, and learn to generate multi-view images in a single denoising process from real-world data. Concretely, we propose to integrate 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network of the text-to-image model. Moreover, we design an autoregressive generation that renders more 3D-consistent images at any viewpoint. We train our model on real-world datasets of objects and showcase its capabilities to generate instances with a variety of high-quality shapes and textures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.651 + ], + "angle": 0, + "content": "in authentic surroundings. Compared to the existing methods, the results generated by our method are consistent, and have favorable visual quality \\((-30\\%\\) FID, \\(-37\\%\\) KID)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.664, + 0.633, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.84 + ], + "angle": 0, + "content": "In recent years, text-to-image (T2I) diffusion models [29, 33] have emerged as cutting-edge technologies, revolutionizing high-quality and imaginative 2D content creation guided by text descriptions. These frameworks have found widespread applications, including extensions such as ControlNet [53] and DreamBooth [32], showcasing their versatility and potential. An intriguing direction in this domain is to use T2I models as powerful 2D priors for generating three-dimensional (3D) assets. How can we effectively use these models to create photo-realistic and diverse 3D assets?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Existing methods like DreamFusion [27], Fantasy3D [5], and ProlificDreamer [47] have demonstrated exciting results by optimizing a 3D representation through score distillation sampling [27] from pretrained T2I diffusion models. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "3D assets generated by these methods exhibit compelling diversity. However, their visual quality is not consistently as high as that of the images generated by T2I models. A key step to obtaining 3D assets is the ability to generate consistent multi-view images of the desired objects and their surroundings. These images can then be fitted to 3D representations like NeRF [25] or NeuS [45]. HoloDiffusion [19] and ViewsetDiffusion [38] train a diffusion model from scratch using multi-view images and output 3D-consistent images. GeNVS [4] and DFM [41] additionally produce object surroundings, thereby increasing the realism of the generation. These methods ensure (photo)-realistic results by training on real-world 3D datasets [31, 56]. However, these datasets are orders of magnitude smaller than the 2D dataset used to train T2I diffusion models. As a result, these approaches produce realistic but less-diverse 3D assets. Alternatively, recent works like Zero-1-to-3 [23] and One-2-3-45 [22] leverage a pretrained T2I model and fine-tune it for 3D consistency. These methods successfully preserve the diversity of generated results by training on a large synthetic 3D dataset [7]. Nonetheless, the produced objects can be less photo-realistic and are without surroundings." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.473, + 0.699 + ], + "angle": 0, + "content": "In this paper, we propose a method that leverages the 2D priors of the pretrained T2I diffusion models to produce photo-realistic and 3D-consistent 3D asset renderings. As shown in the first two rows of Fig. 1, the input is a text description or an image of an object, along with the camera poses of the desired rendered images. The proposed approach produces multiple images of the same object in a single forward pass. Moreover, we design an autoregressive generation scheme that allows to render more images at any novel viewpoint (Fig. 1, third row). Concretely, we introduce projection and cross-frame-attention layers, that are strategically placed into the existing U-Net architecture, to encode explicit 3D knowledge about the generated object (see Fig. 2). By doing so, our approach paves the way to fine-tune T2I models on real-world 3D datasets, such as CO3D [31], while benefiting from the large 2D prior encoded in the pretrained weights. Our generated images are consistent, diverse, and realistic renderings of objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.324, + 0.715 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.717, + 0.472, + 0.793 + ], + "angle": 0, + "content": "- a method that utilizes the pretrained 2D prior of text-to-image models and turns them into 3D-consistent image generators. We train our approach on real-world multi-view datasets, allowing us to produce realistic and high-quality images of objects and their surroundings (Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.794, + 0.472, + 0.853 + ], + "angle": 0, + "content": "- a novel U-Net architecture that combines commonly used 2D layers with 3D-aware layers. Our projection and cross-frame-attention layers encode explicit 3D knowledge into each block of the U-Net architecture (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.854, + 0.472, + 0.899 + ], + "angle": 0, + "content": "- an autoregressive generation scheme that renders images of a 3D object from any desired viewpoint directly with our model in a 3D-consistent way (Sec. 3.3)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.717, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.118, + 0.896, + 0.346 + ], + "angle": 0, + "content": "Text-To-2D. Denoising diffusion probabilistic models (DDPM) [14] model a data distribution by learning to invert a Gaussian noisin process with a deep network. Recently, DDPMs were shown to be superior to generative adversarial networks [8], becoming the state-of-the-art framework for image generation. Soon after, large text-conditioned models trained on billion-scale data were proposed inImagen [33] or Dall-E 2 [29]. While [8] achieved conditional generation via guidance with a classifier, [13] proposed classifier-free guidance. ControlNet [53] proposed a way to tune the diffusion outputs by conditioning on various modalities, such as image segmentation or normal maps. Similar to ControlNet, our method builds on the strong 2D prior of a pretrained text-to-image (T2I) model. We further demonstrate how to adjust this prior to generate 3D-consistent images of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.897, + 0.692 + ], + "angle": 0, + "content": "Text-To-3D. 2D DDPMs were applied to the generation of 3D shapes [28, 34, 39, 42, 44, 50, 58] or scenes [10, 15, 40] from text descriptions. DreamFusion [27] proposed score distillation sampling (SDS) which optimizes a 3D shape whose renders match the belief of the DDPM. Improved sample quality was achieved by a second-stage mesh optimization [5, 21], and smoother SDS convergence [35, 47]. Several methods use 3D data to train a novel-view synthesis model whose multi-view samples can be later converted to 3D, e.g. conditioning a 2D DDPM on an image and a relative camera motion to generate novel views [23, 48]. However, due to no explicit modelling of geometry, the outputs are view-inconsistent. Consistency can be improved with epipolar attention [43, 57], or optimizing a 3D shape from multi-view proposals [22]. Our work fine-tunes a 2D T2I model to generate renders of a 3D object; however, we propose explicit 3D unprojection and rendering operators to improve view-consistency. Concurrently, SyncDreamer [24] also add 3D layers in their 2D DDPM. We differ by training on real data with backgrounds and by showing that autoregressive generation is sufficient to generate consistent images, making the second 3D reconstruction stage expendable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Diffusion on 3D Representations. Several works model the distribution of 3D representations. While DiffRF [26] leverages ground-truth 3D shapes, HoloDiffusion [19] is supervised only with 2D images. HoloFusion [18] extends this work with a 2D diffusion render post-processor. Images can also be denoised by rendering a reconstructing 3D shape [1, 38]. Unfortunately, the limited scale of existing 3D datasets prevents these 3D diffusion models from extrapolating beyond the training distribution. Instead, we exploit a large 2D pretrained DDPM and add 3D components that are tuned on smaller-scale multi-view data. This leads to improved multi-view consistency while maintaining the expressivity brought by pretraining on billion-scale image data." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.895, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.897, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Method Overview. We augment the U-Net architecture of pretrained text-to-image models with new layers in every U-Net block. These layers facilitate communication between multi-view images in a batch, resulting in a denoising process that jointly produces 3D-consistent images. First, we replace self-attention with cross-frame-attention (yellow) which compares the spatial features of all views. We condition all attention layers on pose \\((RT)\\), intrinsics \\((K)\\), and intensity \\((I)\\) of each image. Second, we add a projection layer (green) into the inner blocks of the U-Net. It creates a 3D representation from multi-view features and renders them into 3D-consistent features. We fine-tune the U-Net using the diffusion denoising objective (Eq. 3) at timestep \\(t\\), supervised from captioned multi-view images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.402, + 0.172, + 0.418 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.474, + 0.595 + ], + "angle": 0, + "content": "We propose a method that produces 3D-consistent images from a given text or posed image input (see Fig. 1 top/mid). Concretely, given desired output poses, we jointly generate all images corresponding to the condition. We leverage pretrained text-to-image (T2I) models [29, 33] and fine-tune them on multi-view data [31]. We propose to augment the existing U-Net architecture by adding new layers into each block (see Fig. 2). At test time, we can condition our method on multiple images (see Fig. 1 bottom), which allows us to autoregressively render the same object from any viewpoint directly with the diffusion model (see Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.3, + 0.619 + ], + "angle": 0, + "content": "3.1. 3D-Consistent Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.472, + 0.732 + ], + "angle": 0, + "content": "Diffusion models [14, 36] are a class of generative models that learn the probability distribution \\( p_{\\theta}(x_0) = \\int p_{\\theta}(x_{0:T}) dx_{1:T} \\) over data \\( x_0 \\sim q(x_0) \\) and latent variables \\( x_{1:T} = x_1, \\ldots, x_T \\). Our method is based on pretrained text-to-image models, which are diffusion models \\( p_{\\theta}(x_0 \\mid c) \\) with an additional text condition \\( c \\). For clarity, we drop the condition \\( c \\) for the remainder of this section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.472, + 0.825 + ], + "angle": 0, + "content": "To produce multiple images \\( x_0^{0:N} \\) at once, which are 3D-consistent with each other, we seek to model their joint probability distribution \\( p_{\\theta}(x_0^{0:N}) = \\int p_{\\theta}(x_0^{0:N}) dx_{1:T}^{0:N} \\). Similarly to concurrent work by Liu et al. [24], we generate one set of images \\( p_{\\theta}(x_0^{0:N}) \\) by adapting the reverse process of DDPMs [14] as a Markov chain over all images jointly:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.834, + 0.472, + 0.878 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0: T} ^ {0: N}\\right) := p \\left(x _ {T} ^ {0: N}\\right) \\prod_ {t = 1} ^ {T} \\prod_ {n = 0} ^ {N} p _ {\\theta} \\left(x _ {t - 1} ^ {n} \\mid x _ {t} ^ {0: N}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where we start the generation from Gaussian noise sam" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.895, + 0.541 + ], + "angle": 0, + "content": "pled separately per image \\(p(x_{T}^{n}) = \\mathcal{N}(x_{T}^{n};\\mathbf{0},\\mathbf{I}),\\forall n\\in [0,N]\\). We gradually denoise samples \\(p_{\\theta}(x_{t - 1}^{n}\\mid x_{t}^{0:N}) = \\mathcal{N}(x_{t - 1};\\mu_{\\theta}^{n}(x_{t}^{0:N},t),\\sigma_{t}^{2}\\mathbf{I})\\) by predicting the per-image mean \\(\\mu_{\\theta}^{n}(x_{t}^{0:N},t)\\) through a neural network \\(\\mu_{\\theta}\\) that is shared between all images. Importantly, at each step, the model uses the previous states \\(x_{t}^{0:N}\\) of all images, i.e., there is communication between images during the model prediction. We refer to Sec. 3.2 for details on how this is implemented. To train \\(\\mu_{\\theta}\\), we define the forward process as a Markov chain:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.544, + 0.893, + 0.584 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {1: T} ^ {0: N} \\mid x _ {0} ^ {0: N}\\right) = \\prod_ {t = 1} ^ {T} \\prod_ {n = 0} ^ {N} q \\left(x _ {t} ^ {n} \\mid x _ {t - 1} ^ {n}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.895, + 0.64 + ], + "angle": 0, + "content": "where \\( q(x_{t}^{n} \\mid x_{t - 1}^{n}) = \\mathcal{N}(x_{t}^{n}; \\sqrt{1 - \\beta_{t}} x_{t - 1}^{n}, \\beta_{t}\\mathbf{I}) \\) and \\( \\beta_{1}, \\ldots, \\beta_{T} \\) define a constant variance schedule, i.e., we apply separate noise per image to produce training samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.671 + ], + "angle": 0, + "content": "We follow Ho et al. [14] by learning a noise predictor \\(\\epsilon_{\\theta}\\) instead of \\(\\mu_{\\theta}\\). This allows to train \\(\\epsilon_{\\theta}\\) with an \\(L2\\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.678, + 0.893, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {x _ {0} ^ {0: N}, \\epsilon^ {0: N}} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), n \\left[ \\left\\| \\epsilon^ {n} - \\epsilon_ {\\theta} ^ {n} \\left(x _ {t} ^ {0: N}, t\\right) \\right\\| ^ {2} \\right]. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.842, + 0.728 + ], + "angle": 0, + "content": "3.2. Augmentation of the U-Net architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To model a 3D-consistent denoising process over all images, we predict per-image noise \\(\\epsilon_{\\theta}^{n}(x_{t}^{0:N}, t)\\) through a neural network \\(\\epsilon_{\\theta}\\). This neural network is initialized from the pretrained weights of existing text-to-image models, and is usually defined as a U-Net architecture [29, 33]. We seek to leverage the previous states \\(x_{t}^{0:N}\\) of all images to arrive at a 3D-consistent denoising step. To this end, we propose to add two layers into the U-Net architecture, namely a cross-frame-attention layer and a projection layer. We note that the predicted per-image noise needs to be image specific, since all images are generated starting from separate Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "noise. It is therefore important to keep around 2D layers that act separately on each image, which we achieve by finetuning the existing ResNet [11] and ViT [9] blocks. We summarize our architecture in Fig. 2. In the following, we discuss our two proposed layers in more detail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.472, + 0.239 + ], + "angle": 0, + "content": "Cross-Frame Attention. Inspired by video diffusion [49, 51], we add cross-frame-attention layers into the U-Net architecture. Concretely, we modify the existing self-attention layers to calculate \\( CFAtn(Q,K,V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right)V \\) with" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.257, + 0.471, + 0.277 + ], + "angle": 0, + "content": "\\[\nQ = W ^ {Q} h _ {i}, K = W ^ {K} \\left[ h _ {j} \\right] _ {j \\neq i}, V = W ^ {V} \\left[ h _ {j} \\right] _ {j \\neq i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.346 + ], + "angle": 0, + "content": "where \\( W^{Q}, W^{K}, W^{V} \\) are the pretrained weights for feature projection, and \\( h_{i} \\in \\mathbb{R}^{C \\times H \\times W} \\) is the input spatial feature of each image \\( i \\in [1, N] \\). Intuitively, this matches features across all frames, which allows generating the same global style." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.472, + 0.59 + ], + "angle": 0, + "content": "Additionally, we add a conditioning vector to all cross-frame and cross-attention layers to inform the network about the viewpoint of each image. First, we add pose information by encoding each image's camera matrix \\( p \\in \\mathbb{R}^{4 \\times 4} \\) into an embedding \\( z_1 \\in \\mathbb{R}^4 \\) similar to Zero-1-to-3 [23]. Additionally, we concatenate the focal length and principal point of each camera into an embedding \\( z_2 \\in \\mathbb{R}^4 \\). Finally, we provide an intensity encoding \\( z_3 \\in \\mathbb{R}^2 \\), which stores the mean and variance of the image RGB values. At training time, we set \\( z_3 \\) to the true values of each input image, and at test time, we set \\( z_3 = [0.5, 0] \\) for all images. This helps to reduce the view-dependent lighting differences contained in the dataset (e.g., due to different camera exposure). We construct the conditioning vector as \\( z = [z_1, z_2, z_3] \\), and add it through a LoRA-linear-layer [16] \\( W'^Q \\) to the feature projection matrix \\( Q \\). Concretely, we compute the projected features as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.597, + 0.47, + 0.615 + ], + "angle": 0, + "content": "\\[\nQ = W ^ {Q} h _ {i} + s \\cdot W ^ {\\prime Q} [ h _ {i}; z ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.468, + 0.655 + ], + "angle": 0, + "content": "where we set \\(s = 1\\). Similarly, we add the condition via \\(W^{\\prime K}\\) to \\(K\\), and \\(W^{\\prime V}\\) to \\(V\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Projection Layer. Cross-frame attention layers are helpful to produce globally 3D-consistent images. However, the objects do not precisely follow the specified poses, which leads to view-inconsistencies (see Fig. 5 and Tab. 3). To this end, we add a projection layer into the U-Net architecture (Fig. 3). The idea of this layer is to create 3D-consistent features that are then further processed by the next U-Net layers (e.g. ResNet blocks). By repeating this layer across all stages of the U-Net, we ensure that the per-image features are in a 3D-consistent space. We do not add the projection layer to the first and last U-Net blocks, as we saw no benefit from them at these locations. We reason that the network processes image-specific information at those stages and thus does not need a 3D-consistent feature space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Inspired by multi-view stereo literature [3, 17, 37], we create a 3D feature voxel grid from all input spatial features" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.088, + 0.895, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.331, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Architecture of the projection layer. We produce 3D-consistent output features from posed input features. First, we unproject the compressed image features into 3D and aggregate them into a joint voxel grid with an MLP. Then we refine the voxel grid with a 3D CNN. A volume renderer similar to NeRF [25] renders 3D-consistent features from the grid. Finally, we apply a learned scale function and expand the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.669 + ], + "angle": 0, + "content": "\\(h_{\\mathrm{in}}^{0:N} \\in \\mathbb{R}^{C \\times H \\times W}\\) by projecting each voxel into each image plane. First, we compress \\(h_{\\mathrm{in}}^{0:N}\\) with a \\(1 \\times 1\\) convolution to a reduced feature dimension \\(C' = 16\\). We then take the bilinearly interpolated feature at the image plane location and place it into the voxel. This way, we create a separate voxel grid per view, and merge them into a single grid through an aggregator MLP. Inspired by IBRNet [46], the MLP predicts per-view weights followed by a weighted feature average. We then run a small 3D CNN on the voxel grid to refine the 3D feature space. Afterwards, we render the voxel grid into output features \\(h_{\\mathrm{out}}^{0:N} \\in \\mathbb{R}^{C' \\times H \\times W}\\) with volumetric rendering similar to NeRF [25]. We dedicate half of the voxel grid to foreground and half to background and apply the background model from MERF [30] during ray-marching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.895, + 0.805 + ], + "angle": 0, + "content": "We found it is necessary to add a scale function after the volume rendering output. The volume renderer typically uses a sigmoid activation function as the final layer during ray-marching [25]. However, the input features are defined in an arbitrary floating-point range. To convert \\( h_{\\mathrm{out}}^{0:N} \\) back into the same range, we non-linearly scale the features with \\( 1 \\times 1 \\) convolutions and \\( ReLU \\) activations. Finally, we expand \\( h_{\\mathrm{out}}^{0:N} \\) to the input feature dimension \\( C \\). We refer to the supplemental material for details about each component's architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.742, + 0.833 + ], + "angle": 0, + "content": "3.3. Autoregressive Generation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our method takes as input multiple samples \\( x_{t}^{0:N} \\) at once and denoises them 3D-consistently. During training, we set \\( N = 5 \\), but can increase it at inference time up to memory constraints, e.g., \\( N = 30 \\). However, we want to render an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5046" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.183 + ], + "angle": 0, + "content": "object from any possible viewpoint directly with our network. To this end, we propose an autoregressive image generation scheme, i.e., we condition the generation of next viewpoints on previously generated images. We provide the timesteps \\(t^{0:N}\\) of each image as input to the U-Net. By varying \\(t^{0:N}\\), we can achieve different types of conditioning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.188, + 0.472, + 0.281 + ], + "angle": 0, + "content": "Unconditional Generation. All samples are initialized to Gaussian noise and are denoised jointly. The timesteps \\(t^{0:N}\\) are kept identical for all samples throughout the reverse process. We provide different cameras per image and a single text prompt. The generated images are 3D-consistent, showing the object from the desired viewpoints (Figs. 4 and 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.472, + 0.42 + ], + "angle": 0, + "content": "Image-Conditional Generation. We divide the total number of samples \\( N = n_{\\mathrm{c}} + n_{\\mathrm{g}} \\) into a conditional part \\( n_{\\mathrm{c}} \\) and generative part \\( n_{\\mathrm{g}} \\). The first \\( n_{\\mathrm{c}} \\) samples correspond to images and cameras that are provided as input. The other \\( n_{\\mathrm{g}} \\) samples should generate novel views that are similar to the conditioning images. We start the generation from Gaussian noise for the \\( n_{\\mathrm{g}} \\) samples and provide the un-noised images for the other samples. Similarly, we set \\( t^{0:n_{\\mathrm{c}}} = 0 \\) for all denoising steps, while gradually decreasing \\( t^{n_{\\mathrm{g}}:N} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.472, + 0.527 + ], + "angle": 0, + "content": "When \\( n_{\\mathrm{c}} = 1 \\), our method performs single-image reconstruction (Fig. 6). Setting \\( n_{\\mathrm{c}} > 1 \\) allows to autoregressively generate novel views from previous images (Fig. 1 bottom). In practice, we first generate one batch of images unconditionally and then condition the next batches on a subset of previous images. This allows us to render smooth trajectories around 3D objects (see the supplemental material)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.539, + 0.295, + 0.555 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.472, + 0.669 + ], + "angle": 0, + "content": "Dataset. We train our method on the large-scale CO3Dv2 [31] dataset, which consists of posed multi-view images of real-world objects. Concretely, we choose the categories Teddybear, Hydrant, Apple, and Donut. Per category, we train on 500-1000 objects with each 200 images at resolution \\(256 \\times 256\\). We generate text captions with the BLIP-2 model [20] and sample one of 5 proposals per object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Training. We base our model on a pretrained latent-diffusion text-to-image model. We only fine-tune the U-Net and keep the VAE encoder and decoder frozen. In each iteration, we select \\( N = 5 \\) images and their poses. We sample one denoising timestep \\( t \\sim [0,1000] \\), add noise to the images according to Eq. 2, and compute the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, weskip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.895, + 0.173 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of unconditional image generation. We report average FID [12] and KID [2] per category and improve by a significant margin. This signals that our images are more similar to the distribution of real images in the dataset. We mask away the background for our method and the real images to ensure comparability of numbers with the baselines." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.184, + 0.887, + 0.299 + ], + "angle": 0, + "content": "
CategoryHF [18]VD [38]Ours
FID↓KID↓FID↓KID↓FID↓KID↓
Teddybear81.930.072201.710.16949.390.036
Hydrant61.190.042138.450.11846.450.033
Donut105.970.091199.140.13668.860.054
Apple62.190.056183.670.14956.850.043
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.43 + ], + "angle": 0, + "content": "We fine-tune the model on \\(2 \\times\\) A100 GPUs for 60K iterations (7 days) with a total batch size of 64. We set the learning rate for the volume renderer to 0.005 and for all other layers to \\(5 \\times 10^{-5}\\), and use the AdamW optimizer [33]. During inference, we can increase \\(N\\) and generate up to 30 images/batch on an RTX 3090 GPU. We use the UniPC [55] sampler with 10 denoising steps, which takes 15 seconds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.587, + 0.459 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.895, + 0.59 + ], + "angle": 0, + "content": "Baselines. We compare against recent state-of-the-art works for 3D generative modeling. Our goal is to create multi-view consistent images from real-world, realistic objects with authentic surroundings. Therefore, we consider methods that are trained on real-world datasets and select HoloFusion (HF) [18], ViewsetDiffusion (VD) [38], and DFM [41]. We show results on two tasks: unconditional generation (Sec. 4.1) and single-image reconstruction (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.687 + ], + "angle": 0, + "content": "Metrics. We report FID [12] and KID [2] as common metrics for 2D/3D generation and measure the multi-view consistency of generated images with peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and LPIPS [54]. To ensure comparability, we evaluate all metrics on images without backgrounds, as not every baseline models them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.738, + 0.712 + ], + "angle": 0, + "content": "4.1. Unconditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.825 + ], + "angle": 0, + "content": "Our method can be used to generate 3D-consistent views of an object from any pose with only text as input by using our autoregressive generation (Sec. 3.3). Concretely, we sample an (unobserved) image caption from the test set for the first batch and generate \\( N = 10 \\) images with a guidance scale [13] of \\( \\lambda_{\\mathrm{cfg}} = 7.5 \\). We then set \\( \\lambda_{\\mathrm{cfg}} = 0 \\) for subsequent batches, and create a total of 100 images per object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We evaluate against HoloFusion (HF) [18] and Viewset-Diffusion (VD) [38]. We report quantitative results in Tab. 1 and qualitative results in Figs. 4 and 5. HF [18] creates diverse images that sometimes show view-dependent floating artifacts (see Fig. 5). VD [38] creates consistent but blurry" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5047" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.913, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.697, + 0.895, + 0.74 + ], + "angle": 0, + "content": "Figure 4. Unconditional image generation of our method and baselines. We show renderings from different viewpoints for multiple objects and categories. Our method produces consistent objects and backgrounds. Our textures are sharper in comparison to baselines. Please see the supplemental material for more examples and animations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.471, + 0.797 + ], + "angle": 0, + "content": "images. In contrast, our method produces images with backgrounds and higher-resolution object details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.336, + 0.831 + ], + "angle": 0, + "content": "4.2. Single-Image Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our method can be conditioned on multiple images in order to render any novel view in an autoregressive fashion (Sec. 3.3). To measure the 3D-consistency of our generated images, we compare single-image reconstruction against" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.888 + ], + "angle": 0, + "content": "ViewsetDiffusion (VD) [38] and DFM [41]. Concretely, we sample one image from the dataset and generate 20 images at novel views also sampled from the dataset. We follow Szymanowicz et al. [38] and report the per-view maximum PSNR/SSIM and average LPIPS across multiple objects and viewpoints for all methods. We report quantitative results in Tab. 2 and show qualitative results in Fig. 6. VD [38] creates plausible results without backgrounds. DFM [41] creates" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.913, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.462, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Figure 5. Multi-view consistency of unconditional image generation. HoloFusion (HF) [18] has view-dependent floating artifacts (the base in first row). ViewsetDiffusion (VD) [38] has blurrier renderings (second row). Without the projection layer, our method has no precise control over viewpoints (third row). Without cross-frame-attention, our method suffers from identity changes of the object (fourth row). Our full method produces detailed images that are 3D-consistent (fifth row)." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.533, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of single-image reconstruction. Given a single image as input, we measure the quality of novel views through average PSNR, SSIM, and LPIPS [54] per category. We mask away the generated backgrounds to ensure comparability across all methods. We improve over VD [38] while being on-par with DFM [41]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.586, + 0.875, + 0.685 + ], + "angle": 0, + "content": "
MethodTeddybearHydrantDonutApple
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
VD [38]19.680.700.3022.360.800.1918.270.680.1419.540.640.31
DFM [41]21.810.820.1622.670.830.1223.910.860.1025.790.910.07
Ours21.980.840.1322.490.850.1121.500.850.1825.940.910.11
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.471, + 0.756 + ], + "angle": 0, + "content": "consistent results with backgrounds at a lower image resolution \\((128 \\times 128)\\). Our method produces higher resolution images with similar reconstruction results and backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.766, + 0.191, + 0.781 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.472, + 0.835 + ], + "angle": 0, + "content": "The key ingredients of our method are the cross-frame-attention and projection layers that we add to the U-Net (Sec. 3.2). We highlight their importance in Tab. 3 and Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "How important are the projection layers? They are necessary to allow precise control over the image viewpoints (e.g., Fig. 5 row 3 does not follow the specified rotation). Our goal is to generate a consistent set of images from any" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.817 + ], + "angle": 0, + "content": "viewpoint directly with our model (Sec. 3.3). Being able to control the pose of the object is therefore an essential part of our contribution. The projection layers build up a 3D representation of the object that is explicitly rendered into 3D-consistent features through volume rendering. This allows us to achieve viewpoint consistency, as also demonstrated through single-image reconstruction (Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.903 + ], + "angle": 0, + "content": "How important are cross-frame-attention layers? They are necessary to create images of the same object. Without them, the teddybear in Fig. 5 (row 4) has the same general color scheme and follows the specified poses. However, differences in shape and texture lead to an inconsistent set of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5049" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.089, + 0.484, + 0.102 + ], + "angle": 0, + "content": "Input VD [38] DFM [41] Ours Real Image" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.103, + 0.486, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.167, + 0.486, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.24, + 0.471, + 0.296 + ], + "angle": 0, + "content": "Figure 6. Single-image reconstruction of our method and baselines. Given one image/pose as input, our method produces plausible novel views that are consistent with the real shape and texture. We can also produce detailed backgrounds that match the input." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.311, + 0.49, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.464, + 0.279, + 0.476 + ], + "angle": 0, + "content": "[C] wearing a green hat" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.464, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[C] sits on a green blanket" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.477, + 0.49, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.629, + 0.269, + 0.642 + ], + "angle": 0, + "content": "rusty \\([C]\\) in the grass" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.477, + 0.489, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.629, + 0.462, + 0.641 + ], + "angle": 0, + "content": "yellow and white [C]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Figure 7. Diversity of generated results. We condition our method on text input, which allows to create objects in a desired style. We show samples for hand-crafted text descriptions that combine attributes (e.g., color, shape, background) in a novel way. Each row shows a different generation proposal from our method and we denote the object category (Teddybear, Hydrant) as [C]. This showcases the diversity of generated results, i.e., multiple different objects are generated for the same description." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.469, + 0.818 + ], + "angle": 0, + "content": "images. We reason that the cross-frame-attention layers are essential for defining a consistent object identity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Does the 2D prior help? We utilize a 2D prior in form of the pretrained text-to-image model that we fine-tune in a 3D-consistent fashion (Sec. 3.1). This enables our method to produce sharp and detailed images of objects from different viewpoints. Also, we train our method on captioned images" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.894, + 0.187 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of our method and ablations. We report average PSNR, SSIM, LPIPS [54], FID [12], KID [2] over Teddybear and Hydrant categories. We compare against dropping the projection layer (\"no proj\") and cross-frame-attention (\"no cfa\") from the U-Net (see Sec. 3.2). While still producing high-quality images with similar FID/KID, this shows that our proposed layers are necessary to obtain 3D-consistent images." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.192, + 0.884, + 0.269 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓FID↓KID↓
Ours (no proj)16.550.710.2947.950.034
Ours (no cfa)18.150.760.2547.930.034
Ours22.240.840.1147.920.034
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.386 + ], + "angle": 0, + "content": "to retain the controllable generation through text descriptions (Sec. 3.4). We show the diversity and controllability of our generations in Fig. 7 with hand-crafted text prompts. This highlights that, after finetuning, our model is still faithful to text input, and can combine attributes in a novel way, i.e., our model learns to extrapolate from the training set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.396, + 0.627, + 0.41 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.894, + 0.586 + ], + "angle": 0, + "content": "Our method generates 3D-consistent, high-quality images of diverse objects according to text descriptions or input images. Nevertheless, there are several limitations. First, our method sometimes produces images with slight inconsistency, as shown in the supplement. Since the model is fine-tuned on a real-world dataset consisting of view-dependent effects (e.g., exposure changes), our framework learns to generate such variations across different viewpoints. A potential solution is to add lighting condition through a ControlNet [53]. Second, our work focuses on objects, but similarly scene-scale generation on large datasets [6, 52] can be explored." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.6, + 0.619, + 0.616 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.779 + ], + "angle": 0, + "content": "We presented ViewDiff, a method that, given text or image input, generates 3D-consistent images of real-world objects placed in authentic surroundings. Our method leverages the expressivity of large 2D text-to-image models and fine-tunes this 2D prior on real-world 3D datasets to produce diverse multi-view images in a joint denoising process. The core insight of our work are two novel layers, namely cross-frame-attention and the projection layer (Sec. 3.2). Our autoregressive generation scheme (Sec. 3.3) allows to directly render high-quality novel views of a generated 3D object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.792, + 0.688, + 0.809 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.894, + 0.9 + ], + "angle": 0, + "content": "This work was done during Lukas' internship at Meta Reality Labs Zurich as well as at TU Munich, funded by a Meta sponsored research agreement. Matthias Nießner was also supported by the ERC Starting Grant Scan2CAD (804724). We also thank Angela Dai for the video voiceover." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5050" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.17 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. TransformerFusion: Monocular RGB scene reconstruction using transformers. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[4] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3D-aware diffusion models. arXiv:2304.02602, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.472, + 0.427 + ], + "angle": 0, + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.472, + 0.598 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth \\(16 \\times 16\\) words: Transformers for image recognition at scale. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[10] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-NeRF2NeRF: Editing 3D scenes with instructions. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.472, + 0.683 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3D meshes from 2D text-to-image models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[17] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Atlas: Few-shot learning with retrieval augmented language models. JMLR, 24(251):1-43, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[18] Animesh Karnewar, Niloy J. Mitra, Andrea Vedaldi, and David Novotny. HoloFusion: Towards photo-realistic 3D generative modeling. In ICCV, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[19] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J. Mitra. HoloDiffusion: Training a 3D diffusion model using 2D images. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[22] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. arXiv:2306.16928, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.893, + 0.488 + ], + "angle": 0, + "content": "[23] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.894, + 0.628 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.894, + 0.686 + ], + "angle": 0, + "content": "[26] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3D radiance field diffusion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Ali-aksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv:2306.17843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.856 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv:2204.06125, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[30] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5051" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[31] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.472, + 0.384 + ], + "angle": 0, + "content": "[34] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2D diffusion model know 3D-consistency for robust text-to-3D generation. arXiv:2303.07937, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[35] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. arXiv:2308.16512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[37] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[38] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. ViewSet diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[39] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3D: High-fidelity 3D creation from a single image with diffusion prior. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.472, + 0.677 + ], + "angle": 0, + "content": "[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. MVDiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.472, + 0.747 + ], + "angle": 0, + "content": "[41] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. In NeurIPS, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[42] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. TextMesh: Generation of realistic 3D meshes from text prompts. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[43] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian chaining: Lifting pretrained 2D diffusion models for 3D generation. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[45] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[46] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. arXiv:2305.16213, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[48] Daniel Watson, William Chan, Ricardo Martin Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.894, + 0.401 + ], + "angle": 0, + "content": "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.894, + 0.444 + ], + "angle": 0, + "content": "[50] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3D-aware image generation using 2D diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[51] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.894, + 0.572 + ], + "angle": 0, + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[55] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. UniPC: A unified predictor-corrector framework for fast sampling of diffusion models. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.894, + 0.715 + ], + "angle": 0, + "content": "[56] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Transactions on Graphics, 37(4):65:1-12, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.894, + 0.757 + ], + "angle": 0, + "content": "[57] Zhizhuo Zhou and Shubham Tulsiani. SparseFusion: Distilling view-conditioned diffusion for 3D reconstruction. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.894, + 0.799 + ], + "angle": 0, + "content": "[58] Joseph Zhu and Peiye Zhuang. HiFA: High-fidelity text-to-3D with advanced diffusion guidance. arXiv:2305.18766, 2023.2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5052" + } + ] +] \ No newline at end of file diff --git a/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_origin.pdf b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..99c2ee9427d0cacba6eb4cb48521f4fcc4def659 --- /dev/null +++ b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/ca4acaac-6100-4888-9856-8b0c1fe30b3f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e53f5706c0f437ceb45dabf0575626951798539caac4d6a061eeea9fd3d94aa3 +size 9037029 diff --git a/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/full.md b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1195202d1dd7d936e3c4e784203052ef33081a5c --- /dev/null +++ b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/full.md @@ -0,0 +1,272 @@ +# ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models + +Lukas Hollein $^{1,2}$ Aljaž Božić $^{2}$ Norman Müller $^{2}$ David Novotny $^{2}$ Hung-Yu Tseng $^{2}$ Christian Richardt $^{2}$ Michael Zollhöfer $^{2}$ Matthias Nießner $^{1}$ + +$^{1}$ Technical University of Munich $^{2}$ Meta https://lukashoel.github.io/ViewDiff/ + +![](images/a2d37ef423678c9073328eb73feb890f65acb4e8a171a86435fa42a5d35a9e53.jpg) +Figure 1. Multi-view consistent image generation. Our method takes as input a text description, or any number of posed input images, and generates high-quality, multi-view consistent images of a real-world 3D object in authentic surroundings from any desired camera poses. + +# Abstract + +3D asset generation is getting massive amounts of attention, inspired by the recent success of text-guided 2D content creation. Existing text-to-3D methods use pretrained text-to-image diffusion models in an optimization problem or fine-tune them on synthetic data, which often results in nonphotorealistic 3D objects without backgrounds. In this paper, we present a method that leverages pretrained text-to-image models as a prior, and learn to generate multi-view images in a single denoising process from real-world data. Concretely, we propose to integrate 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network of the text-to-image model. Moreover, we design an autoregressive generation that renders more 3D-consistent images at any viewpoint. We train our model on real-world datasets of objects and showcase its capabilities to generate instances with a variety of high-quality shapes and textures + +in authentic surroundings. Compared to the existing methods, the results generated by our method are consistent, and have favorable visual quality $(-30\%$ FID, $-37\%$ KID). + +# 1. Introduction + +In recent years, text-to-image (T2I) diffusion models [29, 33] have emerged as cutting-edge technologies, revolutionizing high-quality and imaginative 2D content creation guided by text descriptions. These frameworks have found widespread applications, including extensions such as ControlNet [53] and DreamBooth [32], showcasing their versatility and potential. An intriguing direction in this domain is to use T2I models as powerful 2D priors for generating three-dimensional (3D) assets. How can we effectively use these models to create photo-realistic and diverse 3D assets? + +Existing methods like DreamFusion [27], Fantasy3D [5], and ProlificDreamer [47] have demonstrated exciting results by optimizing a 3D representation through score distillation sampling [27] from pretrained T2I diffusion models. The + +3D assets generated by these methods exhibit compelling diversity. However, their visual quality is not consistently as high as that of the images generated by T2I models. A key step to obtaining 3D assets is the ability to generate consistent multi-view images of the desired objects and their surroundings. These images can then be fitted to 3D representations like NeRF [25] or NeuS [45]. HoloDiffusion [19] and ViewsetDiffusion [38] train a diffusion model from scratch using multi-view images and output 3D-consistent images. GeNVS [4] and DFM [41] additionally produce object surroundings, thereby increasing the realism of the generation. These methods ensure (photo)-realistic results by training on real-world 3D datasets [31, 56]. However, these datasets are orders of magnitude smaller than the 2D dataset used to train T2I diffusion models. As a result, these approaches produce realistic but less-diverse 3D assets. Alternatively, recent works like Zero-1-to-3 [23] and One-2-3-45 [22] leverage a pretrained T2I model and fine-tune it for 3D consistency. These methods successfully preserve the diversity of generated results by training on a large synthetic 3D dataset [7]. Nonetheless, the produced objects can be less photo-realistic and are without surroundings. + +In this paper, we propose a method that leverages the 2D priors of the pretrained T2I diffusion models to produce photo-realistic and 3D-consistent 3D asset renderings. As shown in the first two rows of Fig. 1, the input is a text description or an image of an object, along with the camera poses of the desired rendered images. The proposed approach produces multiple images of the same object in a single forward pass. Moreover, we design an autoregressive generation scheme that allows to render more images at any novel viewpoint (Fig. 1, third row). Concretely, we introduce projection and cross-frame-attention layers, that are strategically placed into the existing U-Net architecture, to encode explicit 3D knowledge about the generated object (see Fig. 2). By doing so, our approach paves the way to fine-tune T2I models on real-world 3D datasets, such as CO3D [31], while benefiting from the large 2D prior encoded in the pretrained weights. Our generated images are consistent, diverse, and realistic renderings of objects. + +# To summarize, our contributions are: + +- a method that utilizes the pretrained 2D prior of text-to-image models and turns them into 3D-consistent image generators. We train our approach on real-world multi-view datasets, allowing us to produce realistic and high-quality images of objects and their surroundings (Sec. 3.1). +- a novel U-Net architecture that combines commonly used 2D layers with 3D-aware layers. Our projection and cross-frame-attention layers encode explicit 3D knowledge into each block of the U-Net architecture (Sec. 3.2). +- an autoregressive generation scheme that renders images of a 3D object from any desired viewpoint directly with our model in a 3D-consistent way (Sec. 3.3). + +# 2. Related Work + +Text-To-2D. Denoising diffusion probabilistic models (DDPM) [14] model a data distribution by learning to invert a Gaussian noisin process with a deep network. Recently, DDPMs were shown to be superior to generative adversarial networks [8], becoming the state-of-the-art framework for image generation. Soon after, large text-conditioned models trained on billion-scale data were proposed inImagen [33] or Dall-E 2 [29]. While [8] achieved conditional generation via guidance with a classifier, [13] proposed classifier-free guidance. ControlNet [53] proposed a way to tune the diffusion outputs by conditioning on various modalities, such as image segmentation or normal maps. Similar to ControlNet, our method builds on the strong 2D prior of a pretrained text-to-image (T2I) model. We further demonstrate how to adjust this prior to generate 3D-consistent images of objects. + +Text-To-3D. 2D DDPMs were applied to the generation of 3D shapes [28, 34, 39, 42, 44, 50, 58] or scenes [10, 15, 40] from text descriptions. DreamFusion [27] proposed score distillation sampling (SDS) which optimizes a 3D shape whose renders match the belief of the DDPM. Improved sample quality was achieved by a second-stage mesh optimization [5, 21], and smoother SDS convergence [35, 47]. Several methods use 3D data to train a novel-view synthesis model whose multi-view samples can be later converted to 3D, e.g. conditioning a 2D DDPM on an image and a relative camera motion to generate novel views [23, 48]. However, due to no explicit modelling of geometry, the outputs are view-inconsistent. Consistency can be improved with epipolar attention [43, 57], or optimizing a 3D shape from multi-view proposals [22]. Our work fine-tunes a 2D T2I model to generate renders of a 3D object; however, we propose explicit 3D unprojection and rendering operators to improve view-consistency. Concurrently, SyncDreamer [24] also add 3D layers in their 2D DDPM. We differ by training on real data with backgrounds and by showing that autoregressive generation is sufficient to generate consistent images, making the second 3D reconstruction stage expendable. + +Diffusion on 3D Representations. Several works model the distribution of 3D representations. While DiffRF [26] leverages ground-truth 3D shapes, HoloDiffusion [19] is supervised only with 2D images. HoloFusion [18] extends this work with a 2D diffusion render post-processor. Images can also be denoised by rendering a reconstructing 3D shape [1, 38]. Unfortunately, the limited scale of existing 3D datasets prevents these 3D diffusion models from extrapolating beyond the training distribution. Instead, we exploit a large 2D pretrained DDPM and add 3D components that are tuned on smaller-scale multi-view data. This leads to improved multi-view consistency while maintaining the expressivity brought by pretraining on billion-scale image data. + +![](images/205bf188fa66e3171a5254ac3f40c63e7803ea1955b2a89fa4645c637176a595.jpg) +Figure 2. Method Overview. We augment the U-Net architecture of pretrained text-to-image models with new layers in every U-Net block. These layers facilitate communication between multi-view images in a batch, resulting in a denoising process that jointly produces 3D-consistent images. First, we replace self-attention with cross-frame-attention (yellow) which compares the spatial features of all views. We condition all attention layers on pose $(RT)$ , intrinsics $(K)$ , and intensity $(I)$ of each image. Second, we add a projection layer (green) into the inner blocks of the U-Net. It creates a 3D representation from multi-view features and renders them into 3D-consistent features. We fine-tune the U-Net using the diffusion denoising objective (Eq. 3) at timestep $t$ , supervised from captioned multi-view images. + +# 3. Method + +We propose a method that produces 3D-consistent images from a given text or posed image input (see Fig. 1 top/mid). Concretely, given desired output poses, we jointly generate all images corresponding to the condition. We leverage pretrained text-to-image (T2I) models [29, 33] and fine-tune them on multi-view data [31]. We propose to augment the existing U-Net architecture by adding new layers into each block (see Fig. 2). At test time, we can condition our method on multiple images (see Fig. 1 bottom), which allows us to autoregressively render the same object from any viewpoint directly with the diffusion model (see Sec. 3.3). + +# 3.1. 3D-Consistent Diffusion + +Diffusion models [14, 36] are a class of generative models that learn the probability distribution $p_{\theta}(x_0) = \int p_{\theta}(x_{0:T}) dx_{1:T}$ over data $x_0 \sim q(x_0)$ and latent variables $x_{1:T} = x_1, \ldots, x_T$ . Our method is based on pretrained text-to-image models, which are diffusion models $p_{\theta}(x_0 \mid c)$ with an additional text condition $c$ . For clarity, we drop the condition $c$ for the remainder of this section. + +To produce multiple images $x_0^{0:N}$ at once, which are 3D-consistent with each other, we seek to model their joint probability distribution $p_{\theta}(x_0^{0:N}) = \int p_{\theta}(x_0^{0:N}) dx_{1:T}^{0:N}$ . Similarly to concurrent work by Liu et al. [24], we generate one set of images $p_{\theta}(x_0^{0:N})$ by adapting the reverse process of DDPMs [14] as a Markov chain over all images jointly: + +$$ +p _ {\theta} \left(x _ {0: T} ^ {0: N}\right) := p \left(x _ {T} ^ {0: N}\right) \prod_ {t = 1} ^ {T} \prod_ {n = 0} ^ {N} p _ {\theta} \left(x _ {t - 1} ^ {n} \mid x _ {t} ^ {0: N}\right), \tag {1} +$$ + +where we start the generation from Gaussian noise sam + +pled separately per image $p(x_{T}^{n}) = \mathcal{N}(x_{T}^{n};\mathbf{0},\mathbf{I}),\forall n\in [0,N]$ . We gradually denoise samples $p_{\theta}(x_{t - 1}^{n}\mid x_{t}^{0:N}) = \mathcal{N}(x_{t - 1};\mu_{\theta}^{n}(x_{t}^{0:N},t),\sigma_{t}^{2}\mathbf{I})$ by predicting the per-image mean $\mu_{\theta}^{n}(x_{t}^{0:N},t)$ through a neural network $\mu_{\theta}$ that is shared between all images. Importantly, at each step, the model uses the previous states $x_{t}^{0:N}$ of all images, i.e., there is communication between images during the model prediction. We refer to Sec. 3.2 for details on how this is implemented. To train $\mu_{\theta}$ , we define the forward process as a Markov chain: + +$$ +q \left(x _ {1: T} ^ {0: N} \mid x _ {0} ^ {0: N}\right) = \prod_ {t = 1} ^ {T} \prod_ {n = 0} ^ {N} q \left(x _ {t} ^ {n} \mid x _ {t - 1} ^ {n}\right), \tag {2} +$$ + +where $q(x_{t}^{n} \mid x_{t - 1}^{n}) = \mathcal{N}(x_{t}^{n}; \sqrt{1 - \beta_{t}} x_{t - 1}^{n}, \beta_{t}\mathbf{I})$ and $\beta_{1}, \ldots, \beta_{T}$ define a constant variance schedule, i.e., we apply separate noise per image to produce training samples. + +We follow Ho et al. [14] by learning a noise predictor $\epsilon_{\theta}$ instead of $\mu_{\theta}$ . This allows to train $\epsilon_{\theta}$ with an $L2$ loss: + +$$ +\mathbb {E} _ {x _ {0} ^ {0: N}, \epsilon^ {0: N}} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), n \left[ \left\| \epsilon^ {n} - \epsilon_ {\theta} ^ {n} \left(x _ {t} ^ {0: N}, t\right) \right\| ^ {2} \right]. \tag {3} +$$ + +# 3.2. Augmentation of the U-Net architecture + +To model a 3D-consistent denoising process over all images, we predict per-image noise $\epsilon_{\theta}^{n}(x_{t}^{0:N}, t)$ through a neural network $\epsilon_{\theta}$ . This neural network is initialized from the pretrained weights of existing text-to-image models, and is usually defined as a U-Net architecture [29, 33]. We seek to leverage the previous states $x_{t}^{0:N}$ of all images to arrive at a 3D-consistent denoising step. To this end, we propose to add two layers into the U-Net architecture, namely a cross-frame-attention layer and a projection layer. We note that the predicted per-image noise needs to be image specific, since all images are generated starting from separate Gaussian + +noise. It is therefore important to keep around 2D layers that act separately on each image, which we achieve by finetuning the existing ResNet [11] and ViT [9] blocks. We summarize our architecture in Fig. 2. In the following, we discuss our two proposed layers in more detail. + +Cross-Frame Attention. Inspired by video diffusion [49, 51], we add cross-frame-attention layers into the U-Net architecture. Concretely, we modify the existing self-attention layers to calculate $CFAtn(Q,K,V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d}}\right)V$ with + +$$ +Q = W ^ {Q} h _ {i}, K = W ^ {K} \left[ h _ {j} \right] _ {j \neq i}, V = W ^ {V} \left[ h _ {j} \right] _ {j \neq i}, \tag {4} +$$ + +where $W^{Q}, W^{K}, W^{V}$ are the pretrained weights for feature projection, and $h_{i} \in \mathbb{R}^{C \times H \times W}$ is the input spatial feature of each image $i \in [1, N]$ . Intuitively, this matches features across all frames, which allows generating the same global style. + +Additionally, we add a conditioning vector to all cross-frame and cross-attention layers to inform the network about the viewpoint of each image. First, we add pose information by encoding each image's camera matrix $p \in \mathbb{R}^{4 \times 4}$ into an embedding $z_1 \in \mathbb{R}^4$ similar to Zero-1-to-3 [23]. Additionally, we concatenate the focal length and principal point of each camera into an embedding $z_2 \in \mathbb{R}^4$ . Finally, we provide an intensity encoding $z_3 \in \mathbb{R}^2$ , which stores the mean and variance of the image RGB values. At training time, we set $z_3$ to the true values of each input image, and at test time, we set $z_3 = [0.5, 0]$ for all images. This helps to reduce the view-dependent lighting differences contained in the dataset (e.g., due to different camera exposure). We construct the conditioning vector as $z = [z_1, z_2, z_3]$ , and add it through a LoRA-linear-layer [16] $W'^Q$ to the feature projection matrix $Q$ . Concretely, we compute the projected features as: + +$$ +Q = W ^ {Q} h _ {i} + s \cdot W ^ {\prime Q} [ h _ {i}; z ], \tag {5} +$$ + +where we set $s = 1$ . Similarly, we add the condition via $W^{\prime K}$ to $K$ , and $W^{\prime V}$ to $V$ . + +Projection Layer. Cross-frame attention layers are helpful to produce globally 3D-consistent images. However, the objects do not precisely follow the specified poses, which leads to view-inconsistencies (see Fig. 5 and Tab. 3). To this end, we add a projection layer into the U-Net architecture (Fig. 3). The idea of this layer is to create 3D-consistent features that are then further processed by the next U-Net layers (e.g. ResNet blocks). By repeating this layer across all stages of the U-Net, we ensure that the per-image features are in a 3D-consistent space. We do not add the projection layer to the first and last U-Net blocks, as we saw no benefit from them at these locations. We reason that the network processes image-specific information at those stages and thus does not need a 3D-consistent feature space. + +Inspired by multi-view stereo literature [3, 17, 37], we create a 3D feature voxel grid from all input spatial features + +![](images/56b3342c3f17dc641e6334c87d8732ad9f2e90747fb268bda54a236c8601c825.jpg) +Figure 3. Architecture of the projection layer. We produce 3D-consistent output features from posed input features. First, we unproject the compressed image features into 3D and aggregate them into a joint voxel grid with an MLP. Then we refine the voxel grid with a 3D CNN. A volume renderer similar to NeRF [25] renders 3D-consistent features from the grid. Finally, we apply a learned scale function and expand the feature dimension. + +$h_{\mathrm{in}}^{0:N} \in \mathbb{R}^{C \times H \times W}$ by projecting each voxel into each image plane. First, we compress $h_{\mathrm{in}}^{0:N}$ with a $1 \times 1$ convolution to a reduced feature dimension $C' = 16$ . We then take the bilinearly interpolated feature at the image plane location and place it into the voxel. This way, we create a separate voxel grid per view, and merge them into a single grid through an aggregator MLP. Inspired by IBRNet [46], the MLP predicts per-view weights followed by a weighted feature average. We then run a small 3D CNN on the voxel grid to refine the 3D feature space. Afterwards, we render the voxel grid into output features $h_{\mathrm{out}}^{0:N} \in \mathbb{R}^{C' \times H \times W}$ with volumetric rendering similar to NeRF [25]. We dedicate half of the voxel grid to foreground and half to background and apply the background model from MERF [30] during ray-marching. + +We found it is necessary to add a scale function after the volume rendering output. The volume renderer typically uses a sigmoid activation function as the final layer during ray-marching [25]. However, the input features are defined in an arbitrary floating-point range. To convert $h_{\mathrm{out}}^{0:N}$ back into the same range, we non-linearly scale the features with $1 \times 1$ convolutions and $ReLU$ activations. Finally, we expand $h_{\mathrm{out}}^{0:N}$ to the input feature dimension $C$ . We refer to the supplemental material for details about each component's architecture. + +# 3.3. Autoregressive Generation + +Our method takes as input multiple samples $x_{t}^{0:N}$ at once and denoises them 3D-consistently. During training, we set $N = 5$ , but can increase it at inference time up to memory constraints, e.g., $N = 30$ . However, we want to render an + +object from any possible viewpoint directly with our network. To this end, we propose an autoregressive image generation scheme, i.e., we condition the generation of next viewpoints on previously generated images. We provide the timesteps $t^{0:N}$ of each image as input to the U-Net. By varying $t^{0:N}$ , we can achieve different types of conditioning. + +Unconditional Generation. All samples are initialized to Gaussian noise and are denoised jointly. The timesteps $t^{0:N}$ are kept identical for all samples throughout the reverse process. We provide different cameras per image and a single text prompt. The generated images are 3D-consistent, showing the object from the desired viewpoints (Figs. 4 and 5). + +Image-Conditional Generation. We divide the total number of samples $N = n_{\mathrm{c}} + n_{\mathrm{g}}$ into a conditional part $n_{\mathrm{c}}$ and generative part $n_{\mathrm{g}}$ . The first $n_{\mathrm{c}}$ samples correspond to images and cameras that are provided as input. The other $n_{\mathrm{g}}$ samples should generate novel views that are similar to the conditioning images. We start the generation from Gaussian noise for the $n_{\mathrm{g}}$ samples and provide the un-noised images for the other samples. Similarly, we set $t^{0:n_{\mathrm{c}}} = 0$ for all denoising steps, while gradually decreasing $t^{n_{\mathrm{g}}:N}$ . + +When $n_{\mathrm{c}} = 1$ , our method performs single-image reconstruction (Fig. 6). Setting $n_{\mathrm{c}} > 1$ allows to autoregressively generate novel views from previous images (Fig. 1 bottom). In practice, we first generate one batch of images unconditionally and then condition the next batches on a subset of previous images. This allows us to render smooth trajectories around 3D objects (see the supplemental material). + +# 3.4. Implementation Details + +Dataset. We train our method on the large-scale CO3Dv2 [31] dataset, which consists of posed multi-view images of real-world objects. Concretely, we choose the categories Teddybear, Hydrant, Apple, and Donut. Per category, we train on 500-1000 objects with each 200 images at resolution $256 \times 256$ . We generate text captions with the BLIP-2 model [20] and sample one of 5 proposals per object. + +Training. We base our model on a pretrained latent-diffusion text-to-image model. We only fine-tune the U-Net and keep the VAE encoder and decoder frozen. In each iteration, we select $N = 5$ images and their poses. We sample one denoising timestep $t \sim [0,1000]$ , add noise to the images according to Eq. 2, and compute the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, weskip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according + +Table 1. Quantitative comparison of unconditional image generation. We report average FID [12] and KID [2] per category and improve by a significant margin. This signals that our images are more similar to the distribution of real images in the dataset. We mask away the background for our method and the real images to ensure comparability of numbers with the baselines. + +
CategoryHF [18]VD [38]Ours
FID↓KID↓FID↓KID↓FID↓KID↓
Teddybear81.930.072201.710.16949.390.036
Hydrant61.190.042138.450.11846.450.033
Donut105.970.091199.140.13668.860.054
Apple62.190.056183.670.14956.850.043
+ +We fine-tune the model on $2 \times$ A100 GPUs for 60K iterations (7 days) with a total batch size of 64. We set the learning rate for the volume renderer to 0.005 and for all other layers to $5 \times 10^{-5}$ , and use the AdamW optimizer [33]. During inference, we can increase $N$ and generate up to 30 images/batch on an RTX 3090 GPU. We use the UniPC [55] sampler with 10 denoising steps, which takes 15 seconds. + +# 4. Results + +Baselines. We compare against recent state-of-the-art works for 3D generative modeling. Our goal is to create multi-view consistent images from real-world, realistic objects with authentic surroundings. Therefore, we consider methods that are trained on real-world datasets and select HoloFusion (HF) [18], ViewsetDiffusion (VD) [38], and DFM [41]. We show results on two tasks: unconditional generation (Sec. 4.1) and single-image reconstruction (Sec. 4.2). + +Metrics. We report FID [12] and KID [2] as common metrics for 2D/3D generation and measure the multi-view consistency of generated images with peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and LPIPS [54]. To ensure comparability, we evaluate all metrics on images without backgrounds, as not every baseline models them. + +# 4.1. Unconditional Generation + +Our method can be used to generate 3D-consistent views of an object from any pose with only text as input by using our autoregressive generation (Sec. 3.3). Concretely, we sample an (unobserved) image caption from the test set for the first batch and generate $N = 10$ images with a guidance scale [13] of $\lambda_{\mathrm{cfg}} = 7.5$ . We then set $\lambda_{\mathrm{cfg}} = 0$ for subsequent batches, and create a total of 100 images per object. + +We evaluate against HoloFusion (HF) [18] and Viewset-Diffusion (VD) [38]. We report quantitative results in Tab. 1 and qualitative results in Figs. 4 and 5. HF [18] creates diverse images that sometimes show view-dependent floating artifacts (see Fig. 5). VD [38] creates consistent but blurry + +![](images/5aa021e3a6fe15e71389213ce20c4d047b73d8a93acf00d9dbdefe44c7795f3b.jpg) +Figure 4. Unconditional image generation of our method and baselines. We show renderings from different viewpoints for multiple objects and categories. Our method produces consistent objects and backgrounds. Our textures are sharper in comparison to baselines. Please see the supplemental material for more examples and animations. + +images. In contrast, our method produces images with backgrounds and higher-resolution object details. + +# 4.2. Single-Image Reconstruction + +Our method can be conditioned on multiple images in order to render any novel view in an autoregressive fashion (Sec. 3.3). To measure the 3D-consistency of our generated images, we compare single-image reconstruction against + +ViewsetDiffusion (VD) [38] and DFM [41]. Concretely, we sample one image from the dataset and generate 20 images at novel views also sampled from the dataset. We follow Szymanowicz et al. [38] and report the per-view maximum PSNR/SSIM and average LPIPS across multiple objects and viewpoints for all methods. We report quantitative results in Tab. 2 and show qualitative results in Fig. 6. VD [38] creates plausible results without backgrounds. DFM [41] creates + +![](images/f5ea05c6650f41156fffecf58689e42a1bbebdc3e84ca3fd9144bf80cb33f2d7.jpg) +Figure 5. Multi-view consistency of unconditional image generation. HoloFusion (HF) [18] has view-dependent floating artifacts (the base in first row). ViewsetDiffusion (VD) [38] has blurrier renderings (second row). Without the projection layer, our method has no precise control over viewpoints (third row). Without cross-frame-attention, our method suffers from identity changes of the object (fourth row). Our full method produces detailed images that are 3D-consistent (fifth row). + +Table 2. Quantitative comparison of single-image reconstruction. Given a single image as input, we measure the quality of novel views through average PSNR, SSIM, and LPIPS [54] per category. We mask away the generated backgrounds to ensure comparability across all methods. We improve over VD [38] while being on-par with DFM [41]. + +
MethodTeddybearHydrantDonutApple
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
VD [38]19.680.700.3022.360.800.1918.270.680.1419.540.640.31
DFM [41]21.810.820.1622.670.830.1223.910.860.1025.790.910.07
Ours21.980.840.1322.490.850.1121.500.850.1825.940.910.11
+ +consistent results with backgrounds at a lower image resolution $(128 \times 128)$ . Our method produces higher resolution images with similar reconstruction results and backgrounds. + +# 4.3. Ablations + +The key ingredients of our method are the cross-frame-attention and projection layers that we add to the U-Net (Sec. 3.2). We highlight their importance in Tab. 3 and Fig. 5. + +How important are the projection layers? They are necessary to allow precise control over the image viewpoints (e.g., Fig. 5 row 3 does not follow the specified rotation). Our goal is to generate a consistent set of images from any + +viewpoint directly with our model (Sec. 3.3). Being able to control the pose of the object is therefore an essential part of our contribution. The projection layers build up a 3D representation of the object that is explicitly rendered into 3D-consistent features through volume rendering. This allows us to achieve viewpoint consistency, as also demonstrated through single-image reconstruction (Tab. 3). + +How important are cross-frame-attention layers? They are necessary to create images of the same object. Without them, the teddybear in Fig. 5 (row 4) has the same general color scheme and follows the specified poses. However, differences in shape and texture lead to an inconsistent set of + +![](images/690d863621e40efdb937860b154165fad4af9996f68fd48ac3b8fdd1cea3e0d3.jpg) +Input VD [38] DFM [41] Ours Real Image + +![](images/284a0f9f5d00ff3ca9ac962edc64882e09415e50894d94c9287ad6ac811f0282.jpg) +Figure 6. Single-image reconstruction of our method and baselines. Given one image/pose as input, our method produces plausible novel views that are consistent with the real shape and texture. We can also produce detailed backgrounds that match the input. + +![](images/52a9f32e95e4b50b008dd4be85b353ab8479d5c08108858c25235119714b0a6e.jpg) +[C] wearing a green hat +[C] sits on a green blanket + +![](images/5ac938690c40fc05dcab0f0e164f8bd350f9158c82ae26715e70207a667b3082.jpg) +rusty $[C]$ in the grass +yellow and white [C] +Figure 7. Diversity of generated results. We condition our method on text input, which allows to create objects in a desired style. We show samples for hand-crafted text descriptions that combine attributes (e.g., color, shape, background) in a novel way. Each row shows a different generation proposal from our method and we denote the object category (Teddybear, Hydrant) as [C]. This showcases the diversity of generated results, i.e., multiple different objects are generated for the same description. + +images. We reason that the cross-frame-attention layers are essential for defining a consistent object identity. + +Does the 2D prior help? We utilize a 2D prior in form of the pretrained text-to-image model that we fine-tune in a 3D-consistent fashion (Sec. 3.1). This enables our method to produce sharp and detailed images of objects from different viewpoints. Also, we train our method on captioned images + +Table 3. Quantitative comparison of our method and ablations. We report average PSNR, SSIM, LPIPS [54], FID [12], KID [2] over Teddybear and Hydrant categories. We compare against dropping the projection layer ("no proj") and cross-frame-attention ("no cfa") from the U-Net (see Sec. 3.2). While still producing high-quality images with similar FID/KID, this shows that our proposed layers are necessary to obtain 3D-consistent images. + +
MethodPSNR↑SSIM↑LPIPS↓FID↓KID↓
Ours (no proj)16.550.710.2947.950.034
Ours (no cfa)18.150.760.2547.930.034
Ours22.240.840.1147.920.034
+ +to retain the controllable generation through text descriptions (Sec. 3.4). We show the diversity and controllability of our generations in Fig. 7 with hand-crafted text prompts. This highlights that, after finetuning, our model is still faithful to text input, and can combine attributes in a novel way, i.e., our model learns to extrapolate from the training set. + +# 4.4. Limitations + +Our method generates 3D-consistent, high-quality images of diverse objects according to text descriptions or input images. Nevertheless, there are several limitations. First, our method sometimes produces images with slight inconsistency, as shown in the supplement. Since the model is fine-tuned on a real-world dataset consisting of view-dependent effects (e.g., exposure changes), our framework learns to generate such variations across different viewpoints. A potential solution is to add lighting condition through a ControlNet [53]. Second, our work focuses on objects, but similarly scene-scale generation on large datasets [6, 52] can be explored. + +# 5. Conclusion + +We presented ViewDiff, a method that, given text or image input, generates 3D-consistent images of real-world objects placed in authentic surroundings. Our method leverages the expressivity of large 2D text-to-image models and fine-tunes this 2D prior on real-world 3D datasets to produce diverse multi-view images in a joint denoising process. The core insight of our work are two novel layers, namely cross-frame-attention and the projection layer (Sec. 3.2). Our autoregressive generation scheme (Sec. 3.3) allows to directly render high-quality novel views of a generated 3D object. + +# 6. Acknowledgements + +This work was done during Lukas' internship at Meta Reality Labs Zurich as well as at TU Munich, funded by a Meta sponsored research agreement. Matthias Nießner was also supported by the ERC Starting Grant Scan2CAD (804724). We also thank Angela Dai for the video voiceover. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2 +[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 5, 8 +[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. TransformerFusion: Monocular RGB scene reconstruction using transformers. In NeurIPS, 2021. 4 +[4] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3D-aware diffusion models. arXiv:2304.02602, 2023. 2 +[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 1, 2 +[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 8 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 2 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 4 +[10] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-NeRF2NeRF: Editing 3D scenes with instructions. In ICCV, 2023. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 5, 8 +[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 2, 5 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3 +[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3D meshes from 2D text-to-image models. In ICCV, 2023. 2 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4 + +[17] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Atlas: Few-shot learning with retrieval augmented language models. JMLR, 24(251):1-43, 2023. 4 +[18] Animesh Karnewar, Niloy J. Mitra, Andrea Vedaldi, and David Novotny. HoloFusion: Towards photo-realistic 3D generative modeling. In ICCV, 2023. 2, 5, 6, 7 +[19] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J. Mitra. HoloDiffusion: Training a 3D diffusion model using 2D images. In CVPR, 2023. 2 +[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 5 +[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. In CVPR, 2023. 2 +[22] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. arXiv:2306.16928, 2023. 2 +[23] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2, 4 +[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 4 +[26] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3D radiance field diffusion. In CVPR, 2023. 2 +[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 1, 2 +[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Ali-aksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv:2306.17843, 2023. 2 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv:2204.06125, 2022. 1, 2, 3 +[30] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time + +view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 4 +[31] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2, 3, 5 +[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5 +[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2, 3, 5 +[34] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2D diffusion model know 3D-consistency for robust text-to-3D generation. arXiv:2303.07937, 2023. 2 +[35] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. arXiv:2308.16512, 2023. 2 +[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3 +[37] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In CVPR, 2021. 4 +[38] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. ViewSet diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 2, 5, 6, 7, 8 +[39] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3D: High-fidelity 3D creation from a single image with diffusion prior. In ICCV, 2023. 2 +[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. MVDiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 2 +[41] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. In NeurIPS, 2023. 2, 5, 6, 7, 8 +[42] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. TextMesh: Generation of realistic 3D meshes from text prompts. In 3DV, 2024. 2 +[43] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2 +[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian chaining: Lifting pretrained 2D diffusion models for 3D generation. In CVPR, 2023. 2 + +[45] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2 +[46] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 4 +[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. arXiv:2305.16213, 2023. 1, 2 +[48] Daniel Watson, William Chan, Ricardo Martin Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2 +[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 4 +[50] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3D-aware image generation using 2D diffusion models. In ICCV, 2023. 2 +[51] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 4 +[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 8 +[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 8 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 7, 8 +[55] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. UniPC: A unified predictor-corrector framework for fast sampling of diffusion models. In NeurIPS, 2023. 5 +[56] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Transactions on Graphics, 37(4):65:1-12, 2018. 2 +[57] Zhizhuo Zhou and Shubham Tulsiani. SparseFusion: Distilling view-conditioned diffusion for 3D reconstruction. In CVPR, 2023. 2 +[58] Joseph Zhu and Peiye Zhuang. HiFA: High-fidelity text-to-3D with advanced diffusion guidance. arXiv:2305.18766, 2023.2 \ No newline at end of file diff --git a/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/images.zip b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5e9af3fe98165183e5b14107f4ac8bc3a029414f --- /dev/null +++ b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:889daec82b9eadd90da38aa401e1a11e8cff6ec42ef4b353e47c312e3a4a1dab +size 800729 diff --git a/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/layout.json b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c86345e753458c8a784a87e0a2efaf79c3470be4 --- /dev/null +++ b/2024/ViewDiff_ 3D-Consistent Image Generation with Text-to-Image Models/layout.json @@ -0,0 +1,8135 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": "Lukas Hollein" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " Aljaž Božić" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " Norman Müller" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " David Novotny" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " Hung-Yu Tseng" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " Christian Richardt" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " Michael Zollhöfer" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "text", + "content": " Matthias Nießner" + }, + { + "bbox": [ + 75, + 141, + 517, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 177, + 430, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 177, + 430, + 205 + ], + "spans": [ + { + "bbox": [ + 161, + 177, + 430, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 161, + 177, + 430, + 205 + ], + "type": "text", + "content": "Technical University of Munich " + }, + { + "bbox": [ + 161, + 177, + 430, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 177, + 430, + 205 + ], + "type": "text", + "content": "Meta https://lukashoel.github.io/ViewDiff/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 222, + 548, + 445 + ], + "blocks": [ + { + "bbox": [ + 47, + 222, + 548, + 445 + ], + "lines": [ + { + "bbox": [ + 47, + 222, + 548, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 548, + 445 + ], + "type": "image", + "image_path": "a2d37ef423678c9073328eb73feb890f65acb4e8a171a86435fa42a5d35a9e53.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 449, + 546, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 546, + 472 + ], + "type": "text", + "content": "Figure 1. Multi-view consistent image generation. Our method takes as input a text description, or any number of posed input images, and generates high-quality, multi-view consistent images of a real-world 3D object in authentic surroundings from any desired camera poses." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 497, + 192, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 497, + 192, + 509 + ], + "spans": [ + { + "bbox": [ + 143, + 497, + 192, + 509 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 522, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 522, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 522, + 290, + 714 + ], + "type": "text", + "content": "3D asset generation is getting massive amounts of attention, inspired by the recent success of text-guided 2D content creation. Existing text-to-3D methods use pretrained text-to-image diffusion models in an optimization problem or fine-tune them on synthetic data, which often results in nonphotorealistic 3D objects without backgrounds. In this paper, we present a method that leverages pretrained text-to-image models as a prior, and learn to generate multi-view images in a single denoising process from real-world data. Concretely, we propose to integrate 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network of the text-to-image model. Moreover, we design an autoregressive generation that renders more 3D-consistent images at any viewpoint. We train our model on real-world datasets of objects and showcase its capabilities to generate instances with a variety of high-quality shapes and textures" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "type": "text", + "content": "in authentic surroundings. Compared to the existing methods, the results generated by our method are consistent, and have favorable visual quality " + }, + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "type": "inline_equation", + "content": "(-30\\%" + }, + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "type": "text", + "content": " FID, " + }, + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "type": "inline_equation", + "content": "-37\\%" + }, + { + "bbox": [ + 304, + 478, + 547, + 515 + ], + "type": "text", + "content": " KID)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 525, + 387, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 387, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 387, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 665 + ], + "type": "text", + "content": "In recent years, text-to-image (T2I) diffusion models [29, 33] have emerged as cutting-edge technologies, revolutionizing high-quality and imaginative 2D content creation guided by text descriptions. These frameworks have found widespread applications, including extensions such as ControlNet [53] and DreamBooth [32], showcasing their versatility and potential. An intriguing direction in this domain is to use T2I models as powerful 2D priors for generating three-dimensional (3D) assets. How can we effectively use these models to create photo-realistic and diverse 3D assets?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Existing methods like DreamFusion [27], Fantasy3D [5], and ProlificDreamer [47] have demonstrated exciting results by optimizing a 3D representation through score distillation sampling [27] from pretrained T2I diffusion models. The" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5043" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "3D assets generated by these methods exhibit compelling diversity. However, their visual quality is not consistently as high as that of the images generated by T2I models. A key step to obtaining 3D assets is the ability to generate consistent multi-view images of the desired objects and their surroundings. These images can then be fitted to 3D representations like NeRF [25] or NeuS [45]. HoloDiffusion [19] and ViewsetDiffusion [38] train a diffusion model from scratch using multi-view images and output 3D-consistent images. GeNVS [4] and DFM [41] additionally produce object surroundings, thereby increasing the realism of the generation. These methods ensure (photo)-realistic results by training on real-world 3D datasets [31, 56]. However, these datasets are orders of magnitude smaller than the 2D dataset used to train T2I diffusion models. As a result, these approaches produce realistic but less-diverse 3D assets. Alternatively, recent works like Zero-1-to-3 [23] and One-2-3-45 [22] leverage a pretrained T2I model and fine-tune it for 3D consistency. These methods successfully preserve the diversity of generated results by training on a large synthetic 3D dataset [7]. Nonetheless, the produced objects can be less photo-realistic and are without surroundings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 337, + 289, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 289, + 553 + ], + "type": "text", + "content": "In this paper, we propose a method that leverages the 2D priors of the pretrained T2I diffusion models to produce photo-realistic and 3D-consistent 3D asset renderings. As shown in the first two rows of Fig. 1, the input is a text description or an image of an object, along with the camera poses of the desired rendered images. The proposed approach produces multiple images of the same object in a single forward pass. Moreover, we design an autoregressive generation scheme that allows to render more images at any novel viewpoint (Fig. 1, third row). Concretely, we introduce projection and cross-frame-attention layers, that are strategically placed into the existing U-Net architecture, to encode explicit 3D knowledge about the generated object (see Fig. 2). By doing so, our approach paves the way to fine-tune T2I models on real-world 3D datasets, such as CO3D [31], while benefiting from the large 2D prior encoded in the pretrained weights. Our generated images are consistent, diverse, and realistic renderings of objects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 554, + 198, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 198, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 198, + 566 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 567, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 567, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 288, + 628 + ], + "type": "text", + "content": "- a method that utilizes the pretrained 2D prior of text-to-image models and turns them into 3D-consistent image generators. We train our approach on real-world multi-view datasets, allowing us to produce realistic and high-quality images of objects and their surroundings (Sec. 3.1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 628, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 288, + 675 + ], + "type": "text", + "content": "- a novel U-Net architecture that combines commonly used 2D layers with 3D-aware layers. Our projection and cross-frame-attention layers encode explicit 3D knowledge into each block of the U-Net architecture (Sec. 3.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 676, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 288, + 712 + ], + "type": "text", + "content": "- an autoregressive generation scheme that renders images of a 3D object from any desired viewpoint directly with our model in a 3D-consistent way (Sec. 3.3)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 93, + 548, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 93, + 548, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 93, + 548, + 274 + ], + "type": "text", + "content": "Text-To-2D. Denoising diffusion probabilistic models (DDPM) [14] model a data distribution by learning to invert a Gaussian noisin process with a deep network. Recently, DDPMs were shown to be superior to generative adversarial networks [8], becoming the state-of-the-art framework for image generation. Soon after, large text-conditioned models trained on billion-scale data were proposed inImagen [33] or Dall-E 2 [29]. While [8] achieved conditional generation via guidance with a classifier, [13] proposed classifier-free guidance. ControlNet [53] proposed a way to tune the diffusion outputs by conditioning on various modalities, such as image segmentation or normal maps. Similar to ControlNet, our method builds on the strong 2D prior of a pretrained text-to-image (T2I) model. We further demonstrate how to adjust this prior to generate 3D-consistent images of objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 284, + 548, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 548, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 548, + 548 + ], + "type": "text", + "content": "Text-To-3D. 2D DDPMs were applied to the generation of 3D shapes [28, 34, 39, 42, 44, 50, 58] or scenes [10, 15, 40] from text descriptions. DreamFusion [27] proposed score distillation sampling (SDS) which optimizes a 3D shape whose renders match the belief of the DDPM. Improved sample quality was achieved by a second-stage mesh optimization [5, 21], and smoother SDS convergence [35, 47]. Several methods use 3D data to train a novel-view synthesis model whose multi-view samples can be later converted to 3D, e.g. conditioning a 2D DDPM on an image and a relative camera motion to generate novel views [23, 48]. However, due to no explicit modelling of geometry, the outputs are view-inconsistent. Consistency can be improved with epipolar attention [43, 57], or optimizing a 3D shape from multi-view proposals [22]. Our work fine-tunes a 2D T2I model to generate renders of a 3D object; however, we propose explicit 3D unprojection and rendering operators to improve view-consistency. Concurrently, SyncDreamer [24] also add 3D layers in their 2D DDPM. We differ by training on real data with backgrounds and by showing that autoregressive generation is sufficient to generate consistent images, making the second 3D reconstruction stage expendable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "Diffusion on 3D Representations. Several works model the distribution of 3D representations. While DiffRF [26] leverages ground-truth 3D shapes, HoloDiffusion [19] is supervised only with 2D images. HoloFusion [18] extends this work with a 2D diffusion render post-processor. Images can also be denoised by rendering a reconstructing 3D shape [1, 38]. Unfortunately, the limited scale of existing 3D datasets prevents these 3D diffusion models from extrapolating beyond the training distribution. Instead, we exploit a large 2D pretrained DDPM and add 3D components that are tuned on smaller-scale multi-view data. This leads to improved multi-view consistency while maintaining the expressivity brought by pretraining on billion-scale image data." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5044" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 547, + 225 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 225 + ], + "type": "image", + "image_path": "205bf188fa66e3171a5254ac3f40c63e7803ea1955b2a89fa4645c637176a595.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": "Figure 2. Method Overview. We augment the U-Net architecture of pretrained text-to-image models with new layers in every U-Net block. These layers facilitate communication between multi-view images in a batch, resulting in a denoising process that jointly produces 3D-consistent images. First, we replace self-attention with cross-frame-attention (yellow) which compares the spatial features of all views. We condition all attention layers on pose " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "(RT)" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": ", intrinsics " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": ", and intensity " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " of each image. Second, we add a projection layer (green) into the inner blocks of the U-Net. It creates a 3D representation from multi-view features and renders them into 3D-consistent features. We fine-tune the U-Net using the diffusion denoising objective (Eq. 3) at timestep " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": ", supervised from captioned multi-view images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 318, + 105, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 105, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 105, + 331 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 290, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 290, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 290, + 471 + ], + "type": "text", + "content": "We propose a method that produces 3D-consistent images from a given text or posed image input (see Fig. 1 top/mid). Concretely, given desired output poses, we jointly generate all images corresponding to the condition. We leverage pretrained text-to-image (T2I) models [29, 33] and fine-tune them on multi-view data [31]. We propose to augment the existing U-Net architecture by adding new layers into each block (see Fig. 2). At test time, we can condition our method on multiple images (see Fig. 1 bottom), which allows us to autoregressively render the same object from any viewpoint directly with the diffusion model (see Sec. 3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 478, + 183, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 183, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 183, + 490 + ], + "type": "text", + "content": "3.1. 3D-Consistent Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": "Diffusion models [14, 36] are a class of generative models that learn the probability distribution " + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0) = \\int p_{\\theta}(x_{0:T}) dx_{1:T}" + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": " over data " + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "inline_equation", + "content": "x_0 \\sim q(x_0)" + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": " and latent variables " + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "inline_equation", + "content": "x_{1:T} = x_1, \\ldots, x_T" + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": ". Our method is based on pretrained text-to-image models, which are diffusion models " + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0 \\mid c)" + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": " with an additional text condition " + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": ". For clarity, we drop the condition " + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 496, + 288, + 579 + ], + "type": "text", + "content": " for the remainder of this section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "text", + "content": "To produce multiple images " + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "inline_equation", + "content": "x_0^{0:N}" + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "text", + "content": " at once, which are 3D-consistent with each other, we seek to model their joint probability distribution " + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0^{0:N}) = \\int p_{\\theta}(x_0^{0:N}) dx_{1:T}^{0:N}" + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "text", + "content": ". Similarly to concurrent work by Liu et al. [24], we generate one set of images " + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0^{0:N})" + }, + { + "bbox": [ + 46, + 580, + 288, + 653 + ], + "type": "text", + "content": " by adapting the reverse process of DDPMs [14] as a Markov chain over all images jointly:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 660, + 288, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 660, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 71, + 660, + 288, + 695 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0: T} ^ {0: N}\\right) := p \\left(x _ {T} ^ {0: N}\\right) \\prod_ {t = 1} ^ {T} \\prod_ {n = 0} ^ {N} p _ {\\theta} \\left(x _ {t - 1} ^ {n} \\mid x _ {t} ^ {0: N}\\right), \\tag {1}", + "image_path": "e755906af8698462f26cd0b733a66de9440d83026c54bb566f46726962df965c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 289, + 714 + ], + "type": "text", + "content": "where we start the generation from Gaussian noise sam" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": "pled separately per image " + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "inline_equation", + "content": "p(x_{T}^{n}) = \\mathcal{N}(x_{T}^{n};\\mathbf{0},\\mathbf{I}),\\forall n\\in [0,N]" + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": ". We gradually denoise samples " + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}^{n}\\mid x_{t}^{0:N}) = \\mathcal{N}(x_{t - 1};\\mu_{\\theta}^{n}(x_{t}^{0:N},t),\\sigma_{t}^{2}\\mathbf{I})" + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": " by predicting the per-image mean " + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{n}(x_{t}^{0:N},t)" + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": " through a neural network " + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": " that is shared between all images. Importantly, at each step, the model uses the previous states " + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "inline_equation", + "content": "x_{t}^{0:N}" + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": " of all images, i.e., there is communication between images during the model prediction. We refer to Sec. 3.2 for details on how this is implemented. To train " + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 304, + 319, + 547, + 428 + ], + "type": "text", + "content": ", we define the forward process as a Markov chain:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 430, + 546, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 430, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 344, + 430, + 546, + 462 + ], + "type": "interline_equation", + "content": "q \\left(x _ {1: T} ^ {0: N} \\mid x _ {0} ^ {0: N}\\right) = \\prod_ {t = 1} ^ {T} \\prod_ {n = 0} ^ {N} q \\left(x _ {t} ^ {n} \\mid x _ {t - 1} ^ {n}\\right), \\tag {2}", + "image_path": "29708fdeaceec88398dac268544fa1e3ee7d5f7b426e1e9fadb0e08fdeaf922e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "inline_equation", + "content": "q(x_{t}^{n} \\mid x_{t - 1}^{n}) = \\mathcal{N}(x_{t}^{n}; \\sqrt{1 - \\beta_{t}} x_{t - 1}^{n}, \\beta_{t}\\mathbf{I})" + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\beta_{1}, \\ldots, \\beta_{T}" + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": " define a constant variance schedule, i.e., we apply separate noise per image to produce training samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "text", + "content": "We follow Ho et al. [14] by learning a noise predictor " + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "text", + "content": ". This allows to train " + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "text", + "content": " with an " + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 304, + 507, + 547, + 531 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 536, + 546, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 536, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 337, + 536, + 546, + 559 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {x _ {0} ^ {0: N}, \\epsilon^ {0: N}} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), n \\left[ \\left\\| \\epsilon^ {n} - \\epsilon_ {\\theta} ^ {n} \\left(x _ {t} ^ {0: N}, t\\right) \\right\\| ^ {2} \\right]. \\tag {3}", + "image_path": "ba3df8ea81927340b89593472a58f3ba4cabc648be7a6ac8e0f4a6a09da3a1d5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 563, + 515, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 515, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 515, + 576 + ], + "type": "text", + "content": "3.2. Augmentation of the U-Net architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "To model a 3D-consistent denoising process over all images, we predict per-image noise " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{n}(x_{t}^{0:N}, t)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " through a neural network " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ". This neural network is initialized from the pretrained weights of existing text-to-image models, and is usually defined as a U-Net architecture [29, 33]. We seek to leverage the previous states " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_{t}^{0:N}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " of all images to arrive at a 3D-consistent denoising step. To this end, we propose to add two layers into the U-Net architecture, namely a cross-frame-attention layer and a projection layer. We note that the predicted per-image noise needs to be image specific, since all images are generated starting from separate Gaussian" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5045" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "noise. It is therefore important to keep around 2D layers that act separately on each image, which we achieve by finetuning the existing ResNet [11] and ViT [9] blocks. We summarize our architecture in Fig. 2. In the following, we discuss our two proposed layers in more detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 135, + 288, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 288, + 189 + ], + "type": "text", + "content": "Cross-Frame Attention. Inspired by video diffusion [49, 51], we add cross-frame-attention layers into the U-Net architecture. Concretely, we modify the existing self-attention layers to calculate " + }, + { + "bbox": [ + 46, + 135, + 288, + 189 + ], + "type": "inline_equation", + "content": "CFAtn(Q,K,V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right)V" + }, + { + "bbox": [ + 46, + 135, + 288, + 189 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 203, + 288, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 203, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 58, + 203, + 288, + 219 + ], + "type": "interline_equation", + "content": "Q = W ^ {Q} h _ {i}, K = W ^ {K} \\left[ h _ {j} \\right] _ {j \\neq i}, V = W ^ {V} \\left[ h _ {j} \\right] _ {j \\neq i}, \\tag {4}", + "image_path": "aa49c5b9b289c548c3fb6d928e517c1669a2cffc9b5cd220373d8593c76c135d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "inline_equation", + "content": "W^{Q}, W^{K}, W^{V}" + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "text", + "content": " are the pretrained weights for feature projection, and " + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "inline_equation", + "content": "h_{i} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "text", + "content": " is the input spatial feature of each image " + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "inline_equation", + "content": "i \\in [1, N]" + }, + { + "bbox": [ + 46, + 225, + 287, + 274 + ], + "type": "text", + "content": ". Intuitively, this matches features across all frames, which allows generating the same global style." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": "Additionally, we add a conditioning vector to all cross-frame and cross-attention layers to inform the network about the viewpoint of each image. First, we add pose information by encoding each image's camera matrix " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": " into an embedding " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "z_1 \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": " similar to Zero-1-to-3 [23]. Additionally, we concatenate the focal length and principal point of each camera into an embedding " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "z_2 \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": ". Finally, we provide an intensity encoding " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "z_3 \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": ", which stores the mean and variance of the image RGB values. At training time, we set " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "z_3" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": " to the true values of each input image, and at test time, we set " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "z_3 = [0.5, 0]" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": " for all images. This helps to reduce the view-dependent lighting differences contained in the dataset (e.g., due to different camera exposure). We construct the conditioning vector as " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "z = [z_1, z_2, z_3]" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": ", and add it through a LoRA-linear-layer [16] " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "W'^Q" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": " to the feature projection matrix " + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 274, + 288, + 467 + ], + "type": "text", + "content": ". Concretely, we compute the projected features as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 472, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 472, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 287, + 487 + ], + "type": "interline_equation", + "content": "Q = W ^ {Q} h _ {i} + s \\cdot W ^ {\\prime Q} [ h _ {i}; z ], \\tag {5}", + "image_path": "959faac36916b0eb860931c284f4a96d11eb6d371046288475713546c44d4c72.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "content": ". Similarly, we add the condition via " + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "inline_equation", + "content": "W^{\\prime K}" + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "inline_equation", + "content": "W^{\\prime V}" + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 495, + 286, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 689 + ], + "type": "text", + "content": "Projection Layer. Cross-frame attention layers are helpful to produce globally 3D-consistent images. However, the objects do not precisely follow the specified poses, which leads to view-inconsistencies (see Fig. 5 and Tab. 3). To this end, we add a projection layer into the U-Net architecture (Fig. 3). The idea of this layer is to create 3D-consistent features that are then further processed by the next U-Net layers (e.g. ResNet blocks). By repeating this layer across all stages of the U-Net, we ensure that the per-image features are in a 3D-consistent space. We do not add the projection layer to the first and last U-Net blocks, as we saw no benefit from them at these locations. We reason that the network processes image-specific information at those stages and thus does not need a 3D-consistent feature space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Inspired by multi-view stereo literature [3, 17, 37], we create a 3D feature voxel grid from all input spatial features" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 69, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 312, + 69, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 312, + 69, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 69, + 547, + 255 + ], + "type": "image", + "image_path": "56b3342c3f17dc641e6334c87d8732ad9f2e90747fb268bda54a236c8601c825.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 262, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 547, + 340 + ], + "type": "text", + "content": "Figure 3. Architecture of the projection layer. We produce 3D-consistent output features from posed input features. First, we unproject the compressed image features into 3D and aggregate them into a joint voxel grid with an MLP. Then we refine the voxel grid with a 3D CNN. A volume renderer similar to NeRF [25] renders 3D-consistent features from the grid. Finally, we apply a learned scale function and expand the feature dimension." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{in}}^{0:N} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "text", + "content": " by projecting each voxel into each image plane. First, we compress " + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{in}}^{0:N}" + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "text", + "content": " convolution to a reduced feature dimension " + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "inline_equation", + "content": "C' = 16" + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "text", + "content": ". We then take the bilinearly interpolated feature at the image plane location and place it into the voxel. This way, we create a separate voxel grid per view, and merge them into a single grid through an aggregator MLP. Inspired by IBRNet [46], the MLP predicts per-view weights followed by a weighted feature average. We then run a small 3D CNN on the voxel grid to refine the 3D feature space. Afterwards, we render the voxel grid into output features " + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{out}}^{0:N} \\in \\mathbb{R}^{C' \\times H \\times W}" + }, + { + "bbox": [ + 304, + 360, + 547, + 529 + ], + "type": "text", + "content": " with volumetric rendering similar to NeRF [25]. We dedicate half of the voxel grid to foreground and half to background and apply the background model from MERF [30] during ray-marching." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": "We found it is necessary to add a scale function after the volume rendering output. The volume renderer typically uses a sigmoid activation function as the final layer during ray-marching [25]. However, the input features are defined in an arbitrary floating-point range. To convert " + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{out}}^{0:N}" + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": " back into the same range, we non-linearly scale the features with " + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": " convolutions and " + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "inline_equation", + "content": "ReLU" + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": " activations. Finally, we expand " + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{out}}^{0:N}" + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": " to the input feature dimension " + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": ". We refer to the supplemental material for details about each component's architecture." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 647, + 454, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 454, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 454, + 659 + ], + "type": "text", + "content": "3.3. Autoregressive Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "Our method takes as input multiple samples " + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{t}^{0:N}" + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": " at once and denoises them 3D-consistently. During training, we set " + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": ", but can increase it at inference time up to memory constraints, e.g., " + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "N = 30" + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": ". However, we want to render an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5046" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "object from any possible viewpoint directly with our network. To this end, we propose an autoregressive image generation scheme, i.e., we condition the generation of next viewpoints on previously generated images. We provide the timesteps " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "t^{0:N}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " of each image as input to the U-Net. By varying " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "t^{0:N}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ", we can achieve different types of conditioning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 148, + 288, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 288, + 222 + ], + "type": "text", + "content": "Unconditional Generation. All samples are initialized to Gaussian noise and are denoised jointly. The timesteps " + }, + { + "bbox": [ + 46, + 148, + 288, + 222 + ], + "type": "inline_equation", + "content": "t^{0:N}" + }, + { + "bbox": [ + 46, + 148, + 288, + 222 + ], + "type": "text", + "content": " are kept identical for all samples throughout the reverse process. We provide different cameras per image and a single text prompt. The generated images are 3D-consistent, showing the object from the desired viewpoints (Figs. 4 and 5)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": "Image-Conditional Generation. We divide the total number of samples " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "N = n_{\\mathrm{c}} + n_{\\mathrm{g}}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": " into a conditional part " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": " and generative part " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{g}}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": ". The first " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": " samples correspond to images and cameras that are provided as input. The other " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{g}}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": " samples should generate novel views that are similar to the conditioning images. We start the generation from Gaussian noise for the " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{g}}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": " samples and provide the un-noised images for the other samples. Similarly, we set " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "t^{0:n_{\\mathrm{c}}} = 0" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": " for all denoising steps, while gradually decreasing " + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "inline_equation", + "content": "t^{n_{\\mathrm{g}}:N}" + }, + { + "bbox": [ + 46, + 226, + 288, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{c}} = 1" + }, + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "type": "text", + "content": ", our method performs single-image reconstruction (Fig. 6). Setting " + }, + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{c}} > 1" + }, + { + "bbox": [ + 46, + 334, + 288, + 417 + ], + "type": "text", + "content": " allows to autoregressively generate novel views from previous images (Fig. 1 bottom). In practice, we first generate one batch of images unconditionally and then condition the next batches on a subset of previous images. This allows us to render smooth trajectories around 3D objects (see the supplemental material)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 426, + 180, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 180, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 180, + 439 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 529 + ], + "type": "text", + "content": "Dataset. We train our method on the large-scale CO3Dv2 [31] dataset, which consists of posed multi-view images of real-world objects. Concretely, we choose the categories Teddybear, Hydrant, Apple, and Donut. Per category, we train on 500-1000 objects with each 200 images at resolution " + }, + { + "bbox": [ + 46, + 445, + 288, + 529 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 445, + 288, + 529 + ], + "type": "text", + "content": ". We generate text captions with the BLIP-2 model [20] and sample one of 5 proposals per object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Training. We base our model on a pretrained latent-diffusion text-to-image model. We only fine-tune the U-Net and keep the VAE encoder and decoder frozen. In each iteration, we select " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " images and their poses. We sample one denoising timestep " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "t \\sim [0,1000]" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", add noise to the images according to Eq. 2, and compute the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, weskip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according to Eq. 3. In the projection layers, we skip the last image when building the voxel grid, which enforces the loss according" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 71, + 547, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 547, + 137 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of unconditional image generation. We report average FID [12] and KID [2] per category and improve by a significant margin. This signals that our images are more similar to the distribution of real images in the dataset. We mask away the background for our method and the real images to ensure comparability of numbers with the baselines." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 145, + 542, + 236 + ], + "blocks": [ + { + "bbox": [ + 309, + 145, + 542, + 236 + ], + "lines": [ + { + "bbox": [ + 309, + 145, + 542, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 145, + 542, + 236 + ], + "type": "table", + "html": "
CategoryHF [18]VD [38]Ours
FID↓KID↓FID↓KID↓FID↓KID↓
Teddybear81.930.072201.710.16949.390.036
Hydrant61.190.042138.450.11846.450.033
Donut105.970.091199.140.13668.860.054
Apple62.190.056183.670.14956.850.043
", + "image_path": "343b0bcaf9eeb4428f993b80887b92fdf83b7b790a1ffea54ccce150324f21d8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "text", + "content": "We fine-tune the model on " + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "text", + "content": " A100 GPUs for 60K iterations (7 days) with a total batch size of 64. We set the learning rate for the volume renderer to 0.005 and for all other layers to " + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "text", + "content": ", and use the AdamW optimizer [33]. During inference, we can increase " + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 255, + 547, + 340 + ], + "type": "text", + "content": " and generate up to 30 images/batch on an RTX 3090 GPU. We use the UniPC [55] sampler with 10 denoising steps, which takes 15 seconds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 351, + 359, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 359, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 359, + 363 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 371, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 547, + 467 + ], + "type": "text", + "content": "Baselines. We compare against recent state-of-the-art works for 3D generative modeling. Our goal is to create multi-view consistent images from real-world, realistic objects with authentic surroundings. Therefore, we consider methods that are trained on real-world datasets and select HoloFusion (HF) [18], ViewsetDiffusion (VD) [38], and DFM [41]. We show results on two tasks: unconditional generation (Sec. 4.1) and single-image reconstruction (Sec. 4.2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 544 + ], + "type": "text", + "content": "Metrics. We report FID [12] and KID [2] as common metrics for 2D/3D generation and measure the multi-view consistency of generated images with peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and LPIPS [54]. To ensure comparability, we evaluate all metrics on images without backgrounds, as not every baseline models them." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 551, + 451, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 451, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 451, + 563 + ], + "type": "text", + "content": "4.1. Unconditional Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "content": "Our method can be used to generate 3D-consistent views of an object from any pose with only text as input by using our autoregressive generation (Sec. 3.3). Concretely, we sample an (unobserved) image caption from the test set for the first batch and generate " + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "content": " images with a guidance scale [13] of " + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cfg}} = 7.5" + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "content": ". We then set " + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cfg}} = 0" + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "content": " for subsequent batches, and create a total of 100 images per object." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We evaluate against HoloFusion (HF) [18] and Viewset-Diffusion (VD) [38]. We report quantitative results in Tab. 1 and qualitative results in Figs. 4 and 5. HF [18] creates diverse images that sometimes show view-dependent floating artifacts (see Fig. 5). VD [38] creates consistent but blurry" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5047" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 558, + 543 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 558, + 543 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 558, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 558, + 543 + ], + "type": "image", + "image_path": "5aa021e3a6fe15e71389213ce20c4d047b73d8a93acf00d9dbdefe44c7795f3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 552, + 547, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 547, + 586 + ], + "type": "text", + "content": "Figure 4. Unconditional image generation of our method and baselines. We show renderings from different viewpoints for multiple objects and categories. Our method produces consistent objects and backgrounds. Our textures are sharper in comparison to baselines. Please see the supplemental material for more examples and animations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 606, + 288, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 288, + 631 + ], + "type": "text", + "content": "images. In contrast, our method produces images with backgrounds and higher-resolution object details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 645, + 205, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 205, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 205, + 658 + ], + "type": "text", + "content": "4.2. Single-Image Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Our method can be conditioned on multiple images in order to render any novel view in an autoregressive fashion (Sec. 3.3). To measure the 3D-consistency of our generated images, we compare single-image reconstruction against" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 606, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 703 + ], + "type": "text", + "content": "ViewsetDiffusion (VD) [38] and DFM [41]. Concretely, we sample one image from the dataset and generate 20 images at novel views also sampled from the dataset. We follow Szymanowicz et al. [38] and report the per-view maximum PSNR/SSIM and average LPIPS across multiple objects and viewpoints for all methods. We report quantitative results in Tab. 2 and show qualitative results in Fig. 6. VD [38] creates plausible results without backgrounds. DFM [41] creates" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5048" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 558, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 558, + 358 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 558, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 558, + 358 + ], + "type": "image", + "image_path": "f5ea05c6650f41156fffecf58689e42a1bbebdc3e84ca3fd9144bf80cb33f2d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 365, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 45, + 365, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 547, + 411 + ], + "type": "text", + "content": "Figure 5. Multi-view consistency of unconditional image generation. HoloFusion (HF) [18] has view-dependent floating artifacts (the base in first row). ViewsetDiffusion (VD) [38] has blurrier renderings (second row). Without the projection layer, our method has no precise control over viewpoints (third row). Without cross-frame-attention, our method suffers from identity changes of the object (fourth row). Our full method produces detailed images that are 3D-consistent (fifth row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 464, + 535, + 542 + ], + "blocks": [ + { + "bbox": [ + 45, + 422, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 45, + 422, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 422, + 547, + 456 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of single-image reconstruction. Given a single image as input, we measure the quality of novel views through average PSNR, SSIM, and LPIPS [54] per category. We mask away the generated backgrounds to ensure comparability across all methods. We improve over VD [38] while being on-par with DFM [41]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 464, + 535, + 542 + ], + "lines": [ + { + "bbox": [ + 56, + 464, + 535, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 535, + 542 + ], + "type": "table", + "html": "
MethodTeddybearHydrantDonutApple
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
VD [38]19.680.700.3022.360.800.1918.270.680.1419.540.640.31
DFM [41]21.810.820.1622.670.830.1223.910.860.1025.790.910.07
Ours21.980.840.1322.490.850.1121.500.850.1825.940.910.11
", + "image_path": "c4236fc099827a5eb22c476fd12737db107919ef5d1aee5f5842d969c3e3756d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 562, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 288, + 598 + ], + "type": "text", + "content": "consistent results with backgrounds at a lower image resolution " + }, + { + "bbox": [ + 46, + 562, + 288, + 598 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 46, + 562, + 288, + 598 + ], + "type": "text", + "content": ". Our method produces higher resolution images with similar reconstruction results and backgrounds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 606, + 116, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 116, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 116, + 618 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 625, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 288, + 661 + ], + "type": "text", + "content": "The key ingredients of our method are the cross-frame-attention and projection layers that we add to the U-Net (Sec. 3.2). We highlight their importance in Tab. 3 and Fig. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "How important are the projection layers? They are necessary to allow precise control over the image viewpoints (e.g., Fig. 5 row 3 does not follow the specified rotation). Our goal is to generate a consistent set of images from any" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 562, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 647 + ], + "type": "text", + "content": "viewpoint directly with our model (Sec. 3.3). Being able to control the pose of the object is therefore an essential part of our contribution. The projection layers build up a 3D representation of the object that is explicitly rendered into 3D-consistent features through volume rendering. This allows us to achieve viewpoint consistency, as also demonstrated through single-image reconstruction (Tab. 3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "content": "How important are cross-frame-attention layers? They are necessary to create images of the same object. Without them, the teddybear in Fig. 5 (row 4) has the same general color scheme and follows the specified poses. However, differences in shape and texture lead to an inconsistent set of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5049" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 81, + 297, + 131 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 296, + 80 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 296, + 80 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 296, + 80 + ], + "type": "text", + "content": "Input VD [38] DFM [41] Ours Real Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 81, + 297, + 131 + ], + "lines": [ + { + "bbox": [ + 49, + 81, + 297, + 131 + ], + "spans": [ + { + "bbox": [ + 49, + 81, + 297, + 131 + ], + "type": "image", + "image_path": "690d863621e40efdb937860b154165fad4af9996f68fd48ac3b8fdd1cea3e0d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 132, + 297, + 183 + ], + "blocks": [ + { + "bbox": [ + 49, + 132, + 297, + 183 + ], + "lines": [ + { + "bbox": [ + 49, + 132, + 297, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 132, + 297, + 183 + ], + "type": "image", + "image_path": "284a0f9f5d00ff3ca9ac962edc64882e09415e50894d94c9287ad6ac811f0282.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 190, + 288, + 234 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 288, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 288, + 234 + ], + "type": "text", + "content": "Figure 6. Single-image reconstruction of our method and baselines. Given one image/pose as input, our method produces plausible novel views that are consistent with the real shape and texture. We can also produce detailed backgrounds that match the input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 246, + 299, + 367 + ], + "blocks": [ + { + "bbox": [ + 48, + 246, + 299, + 367 + ], + "lines": [ + { + "bbox": [ + 48, + 246, + 299, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 299, + 367 + ], + "type": "image", + "image_path": "52a9f32e95e4b50b008dd4be85b353ab8479d5c08108858c25235119714b0a6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 367, + 170, + 376 + ], + "lines": [ + { + "bbox": [ + 71, + 367, + 170, + 376 + ], + "spans": [ + { + "bbox": [ + 71, + 367, + 170, + 376 + ], + "type": "text", + "content": "[C] wearing a green hat" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 367, + 295, + 376 + ], + "lines": [ + { + "bbox": [ + 180, + 367, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 180, + 367, + 295, + 376 + ], + "type": "text", + "content": "[C] sits on a green blanket" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 377, + 299, + 497 + ], + "blocks": [ + { + "bbox": [ + 48, + 377, + 299, + 497 + ], + "lines": [ + { + "bbox": [ + 48, + 377, + 299, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 299, + 497 + ], + "type": "image", + "image_path": "5ac938690c40fc05dcab0f0e164f8bd350f9158c82ae26715e70207a667b3082.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 498, + 164, + 508 + ], + "lines": [ + { + "bbox": [ + 74, + 498, + 164, + 508 + ], + "spans": [ + { + "bbox": [ + 74, + 498, + 164, + 508 + ], + "type": "text", + "content": "rusty " + }, + { + "bbox": [ + 74, + 498, + 164, + 508 + ], + "type": "inline_equation", + "content": "[C]" + }, + { + "bbox": [ + 74, + 498, + 164, + 508 + ], + "type": "text", + "content": " in the grass" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 193, + 498, + 282, + 507 + ], + "lines": [ + { + "bbox": [ + 193, + 498, + 282, + 507 + ], + "spans": [ + { + "bbox": [ + 193, + 498, + 282, + 507 + ], + "type": "text", + "content": "yellow and white [C]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 513, + 287, + 601 + ], + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 601 + ], + "type": "text", + "content": "Figure 7. Diversity of generated results. We condition our method on text input, which allows to create objects in a desired style. We show samples for hand-crafted text descriptions that combine attributes (e.g., color, shape, background) in a novel way. Each row shows a different generation proposal from our method and we denote the object category (Teddybear, Hydrant) as [C]. This showcases the diversity of generated results, i.e., multiple different objects are generated for the same description." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 179, + 377, + 299, + 497 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 377, + 299, + 497 + ], + "spans": [ + { + "bbox": [ + 179, + 377, + 299, + 497 + ], + "type": "image", + "image_path": "1c0b8d70f6a2333844dfa9c8e796e74838fa1a7e2baf439324036f78cd2006a7.jpg" + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "bbox": [ + 46, + 624, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 647 + ], + "type": "text", + "content": "images. We reason that the cross-frame-attention layers are essential for defining a consistent object identity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Does the 2D prior help? We utilize a 2D prior in form of the pretrained text-to-image model that we fine-tune in a 3D-consistent fashion (Sec. 3.1). This enables our method to produce sharp and detailed images of objects from different viewpoints. Also, we train our method on captioned images" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 310, + 152, + 541, + 213 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 148 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of our method and ablations. We report average PSNR, SSIM, LPIPS [54], FID [12], KID [2] over Teddybear and Hydrant categories. We compare against dropping the projection layer (\"no proj\") and cross-frame-attention (\"no cfa\") from the U-Net (see Sec. 3.2). While still producing high-quality images with similar FID/KID, this shows that our proposed layers are necessary to obtain 3D-consistent images." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 152, + 541, + 213 + ], + "lines": [ + { + "bbox": [ + 310, + 152, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 310, + 152, + 541, + 213 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓FID↓KID↓
Ours (no proj)16.550.710.2947.950.034
Ours (no cfa)18.150.760.2547.930.034
Ours22.240.840.1147.920.034
", + "image_path": "41f52d429d6ca99f058b3d6ff5034cb1b0623d98c84386379b5752d24b2f9d1b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 233, + 546, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 305 + ], + "type": "text", + "content": "to retain the controllable generation through text descriptions (Sec. 3.4). We show the diversity and controllability of our generations in Fig. 7 with hand-crafted text prompts. This highlights that, after finetuning, our model is still faithful to text input, and can combine attributes in a novel way, i.e., our model learns to extrapolate from the training set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 313, + 383, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 313, + 383, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 383, + 324 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 331, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 464 + ], + "type": "text", + "content": "Our method generates 3D-consistent, high-quality images of diverse objects according to text descriptions or input images. Nevertheless, there are several limitations. First, our method sometimes produces images with slight inconsistency, as shown in the supplement. Since the model is fine-tuned on a real-world dataset consisting of view-dependent effects (e.g., exposure changes), our framework learns to generate such variations across different viewpoints. A potential solution is to add lighting condition through a ControlNet [53]. Second, our work focuses on objects, but similarly scene-scale generation on large datasets [6, 52] can be explored." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 475, + 378, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 378, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 378, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "content": "We presented ViewDiff, a method that, given text or image input, generates 3D-consistent images of real-world objects placed in authentic surroundings. Our method leverages the expressivity of large 2D text-to-image models and fine-tunes this 2D prior on real-world 3D datasets to produce diverse multi-view images in a joint denoising process. The core insight of our work are two novel layers, namely cross-frame-attention and the projection layer (Sec. 3.2). Our autoregressive generation scheme (Sec. 3.3) allows to directly render high-quality novel views of a generated 3D object." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 627, + 421, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 421, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 421, + 640 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 647, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 547, + 712 + ], + "type": "text", + "content": "This work was done during Lukas' internship at Meta Reality Labs Zurich as well as at TU Munich, funded by a Meta sponsored research agreement. Matthias Nießner was also supported by the ERC Starting Grant Scan2CAD (804724). We also thank Angela Dai for the video voiceover." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5050" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "type": "text", + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In ICLR, 2018. 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 203 + ], + "type": "text", + "content": "[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. TransformerFusion: Monocular RGB scene reconstruction using transformers. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "text", + "content": "[4] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3D-aware diffusion models. arXiv:2304.02602, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 260, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 293 + ], + "type": "text", + "content": "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 294, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 338 + ], + "type": "text", + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 339, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 384 + ], + "type": "text", + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 385, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 288, + 407 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 407, + 288, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 473 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth " + }, + { + "bbox": [ + 53, + 407, + 288, + 473 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 53, + 407, + 288, + 473 + ], + "type": "text", + "content": " words: Transformers for image recognition at scale. In ICLR, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "type": "text", + "content": "[10] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-NeRF2NeRF: Editing 3D scenes with instructions. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 288, + 610 + ], + "type": "text", + "content": "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 634 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "type": "text", + "content": "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3D meshes from 2D text-to-image models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[17] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Atlas: Few-shot learning with retrieval augmented language models. JMLR, 24(251):1-43, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 129, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 162 + ], + "type": "text", + "content": "[18] Animesh Karnewar, Niloy J. Mitra, Andrea Vedaldi, and David Novotny. HoloFusion: Towards photo-realistic 3D generative modeling. In ICCV, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 196 + ], + "type": "text", + "content": "[19] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J. Mitra. HoloDiffusion: Training a 3D diffusion model using 2D images. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "text", + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "text", + "content": "[22] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. arXiv:2306.16928, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 331, + 546, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 386 + ], + "type": "text", + "content": "[23] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 442 + ], + "type": "text", + "content": "[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 547, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 547, + 497 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 499, + 547, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 547, + 543 + ], + "type": "text", + "content": "[26] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3D radiance field diffusion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 544, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 547, + 578 + ], + "type": "text", + "content": "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "text", + "content": "[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Ali-aksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. arXiv:2306.17843, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 547, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 677 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv:2204.06125, 2022. 1, 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 679, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 547, + 713 + ], + "type": "text", + "content": "[30] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5051" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 139 + ], + "type": "text", + "content": "[31] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 249 + ], + "type": "text", + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "text", + "content": "[34] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2D diffusion model know 3D-consistency for robust text-to-3D generation. arXiv:2303.07937, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 337 + ], + "type": "text", + "content": "[35] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. arXiv:2308.16512, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "text", + "content": "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "type": "text", + "content": "[37] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In CVPR, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 446 + ], + "type": "text", + "content": "[38] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. ViewSet diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "text", + "content": "[39] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3D: High-fidelity 3D creation from a single image with diffusion prior. In ICCV, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "type": "text", + "content": "[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. MVDiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "text", + "content": "[41] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. In NeurIPS, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 634 + ], + "type": "text", + "content": "[42] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. TextMesh: Generation of realistic 3D meshes from text prompts. In 3DV, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "text", + "content": "[43] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian chaining: Lifting pretrained 2D diffusion models for 3D generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 632 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[45] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[46] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 217 + ], + "type": "text", + "content": "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. arXiv:2305.16213, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 218, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 547, + 261 + ], + "type": "text", + "content": "[48] Daniel Watson, William Chan, Ricardo Martin Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 262, + 547, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 547, + 317 + ], + "type": "text", + "content": "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 318, + 547, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 547, + 351 + ], + "type": "text", + "content": "[50] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3D-aware image generation using 2D diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 353, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 547, + 385 + ], + "type": "text", + "content": "[51] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 387, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 547, + 419 + ], + "type": "text", + "content": "[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 420, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 547, + 453 + ], + "type": "text", + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 454, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 547, + 487 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 488, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 547, + 521 + ], + "type": "text", + "content": "[55] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. UniPC: A unified predictor-corrector framework for fast sampling of diffusion models. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 522, + 547, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 547, + 566 + ], + "type": "text", + "content": "[56] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Transactions on Graphics, 37(4):65:1-12, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 547, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 547, + 599 + ], + "type": "text", + "content": "[57] Zhizhuo Zhou and Shubham Tulsiani. SparseFusion: Distilling view-conditioned diffusion for 3D reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 600, + 547, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 547, + 632 + ], + "type": "text", + "content": "[58] Joseph Zhu and Peiye Zhuang. HiFA: High-fidelity text-to-3D with advanced diffusion guidance. arXiv:2305.18766, 2023.2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5052" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_content_list.json b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..844607f4191ca40ac5a92fcbc209637e78a2c18e --- /dev/null +++ b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "ViewFusion: Towards Multi-View Consistency via Interpolated Denoising", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianghui Yang $^{1,2}$ , Yan Zuo $^{1}$ , Sameera Ramasinghe $^{1}$ , Loris Bazzani $^{1}$ , Gil Avraham $^{1}$ , Anton van den Hengel $^{1,3}$ , ${}^{1}$ Amazon, ${}^{2}$ The University of Sydney, ${}^{3}$ The University of Adelaide", + "bbox": [ + 86, + 179, + 942, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 250, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel-view synthesis through diffusion models has demonstrated remarkable potential for generating diverse and high-quality images. Yet, the independent process of image generation in these prevailing methods leads to challenges in maintaining multiple view consistency. To address this, we introduce ViewFusion, a novel, training-free algorithm that can be seamlessly integrated into existing pre-trained diffusion models. Our approach adopts an auto-regressive method that implicitly leverages previously generated views as context for next view generation, ensuring robust multi-view consistency during the novel-view generation process. Through a diffusion process that fuses known-view information via interpolated denoising, our framework successfully extends single-view conditioned models to work in multiple-view conditional settings without any additional fine-tuning. Extensive experimental results demonstrate the effectiveness of ViewFusion in generating consistent and detailed novel views.", + "bbox": [ + 73, + 282, + 473, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans have a remarkable capacity for visualizing unseen perspectives from just a single image view - an intuitive process that remains complex to model. Such an ability is known as Novel View Synthesis (NVS) and necessitates robust geometric priors to accurately infer three-dimensional details from flat imagery; lifting from a two-dimensional projection to a three-dimensional form involves assumptions and knowledge about the nature of the object and space. Recently, significant advancements in NVS have been brought forward by neural networks [14, 15, 30, 41, 60, 64, 65, 72, 75, 76], where novel view generation for downstream reconstruction shows promising potential [34, 69].", + "bbox": [ + 75, + 607, + 468, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, diffusion models [19, 51] and their ability to generate high-quality 2D images have garnered significant attention in the 3D domain, where pre-trained, text-conditioned 2D diffusion models have been re-purposed for 3D applications via distillation [4, 31, 40, 47, 49, 57, 63,", + "bbox": [ + 75, + 803, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a02c626c539bcb87a527e3783ab55f7cc4db15ad6a8ba4f6dad639fdaf6b8a26.jpg", + "image_caption": [ + "(a) Inconsistent generation." + ], + "image_footnote": [], + "bbox": [ + 504, + 247, + 694, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55e4996694bcb11496ee64ff3477afc4d0537aa530919a5023cae23851235651.jpg", + "image_caption": [ + "(b) Consistent generation (ours).", + "Figure 1. The cause of multi-view inconsistency in diffusion-based novel-view synthesis models. (a) Diffusion models incorporate randomness for diversity and better distribution modeling; this independent generation process produces realistic views under specific instances but may produce different plausible views for various instances, lacking alignment across adjacent views. (b) In contrast, ViewFusion incorporates an auto-regressive process to reduce uncertainty and achieve multi-view consistency, by ensuring a correlated denoising process that ends at the same high-density area, fostering consistency across views." + ], + "image_footnote": [], + "bbox": [ + 696, + 250, + 888, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "68, 73]. Follow-up approaches [34, 69] remove the requirement of text conditioning and instead take an image and target pose as conditions for NVS. However, distillation [63] is still required as the diffusion model cannot produce the multi-view consistent outputs that are appropriate for certain downstream tasks (e.g., optimizing Neural Radiance Fields (NeRFs) [41]).", + "bbox": [ + 496, + 551, + 890, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Under the single-view setting, maintaining multi-view consistency remains particularly challenging since there may exist several plausible outputs for a novel view that are aligned with the given input image. For diffusion-based approaches which generate novel views in an independent manner [34, 69], this results in synthesized views containing artifacts of multi-view inconsistency (Fig. 1a). Previous work [33, 36, 38, 53, 74, 77] focuses on improving the robustness of the downstream reconstruction to address the inconsistency issue, including feature projection layers in the NeRF [33] or utilising three-dimensional priors to constrain NeRF optimization [36, 77], yet these techniques require training or fine-tuning to align additional modules to the original diffusion models.", + "bbox": [ + 496, + 657, + 892, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we address the multi-view inconsistency that arises during the process of view synthesis. Rather than", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at Amazon", + "bbox": [ + 94, + 887, + 315, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9870", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "independently synthesizing views conditioned only on the initial reference image, we develop a novel approach where each subsequently generated view is also conditioned on the entire set of previously generated views. Specifically, our method incorporates an auto-regressive process into the diffusion process to model the joint distribution of views, guiding our novel-view synthesis by maintaining the denoising direction towards the same high density area of already generated views (Fig. 1b).", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework, named ViewFusion, relaxes the single-view conditioning requirement of typical diffusion models through an interpolated denoising process. ViewFusion offers several additional advantages: 1) it can utilize all available views as guidance, thereby enhancing the quality of generated images by incorporating more information; 2) it does not require any additional fine-tuning, effortlessly converting pre-trained single-view conditioned diffusion models into multi-view conditioned diffusion models; 3) it provides greater flexibility in setting adaptive weights for condition images based on their relative view distance to the target view.", + "bbox": [ + 75, + 227, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are the following:", + "bbox": [ + 96, + 409, + 421, + 425 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a training-free algorithm which can be directly applied to pre-trained diffusion models to improve multi-view consistency of synthesized views and supports multiple conditional inputs.", + "- Our method utilizes a novel, auto-regressive approach which we call Interpolated Denoising, that implicitly addresses key limitations of previous auto-regressive approaches for view synthesis.", + "- Extensive empirical analysis on ABO [6] and GSO [9] show that our method is able to achieve better 3D consistency in image generation, leading to significant improvements in novel view synthesis and 3D reconstruction of shapes under single-view and multi-view image settings over other baseline methods." + ], + "bbox": [ + 76, + 428, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 654, + 218, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D-adapted Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 680, + 339, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models have excelled in image generation using conditional inputs [20, 46, 50, 84] and given this success in the 2D domain, recent works have tried to extend diffusion models to 3D content generation [1, 3, 5, 10, 16, 18, 22, 24, 25, 27, 37, 43-45, 66, 82, 83] - although the scarcity of 3D data presents a significant challenge to directly train these diffusion models. Nonetheless, pioneer works such as DreamFusion [47] and Score Jacobian Chaining [63] leverage pre-trained text-conditioned diffusion models to craft 3D models via distillation. Follow-up approaches [4, 31, 57, 68] improve this distillation in terms of speed, resolution and shape quality. Approaches such as [40, 49, 57, 73] extend upon this to support image condi", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions through the use of captions with limited success due to the non-trivial nature of textual inversion [13].", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Novel View Synthesis Diffusion Models", + "text_level": 1, + "bbox": [ + 498, + 133, + 834, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of research [2, 8, 17, 28, 35, 56, 58, 61, 62, 69, 71, 78, 80, 86] directly applies 2D diffusion models to generate multi-view images for shape reconstruction. To circumvent the weakness of text-conditioned diffusion models, novel-view synthesis diffusion models [34, 69] have also been explored, which take an image and target pose as conditions to generate novel views. However, for these approaches, recovering a 3D consistent shape is still a key challenge. To mitigate 3D inconsistency, Liu et al. [33] suggests training a Neural Radiance Field (NeRF) with feature projection layers. Concurrently, other works [36, 38, 70, 74, 77] add modules to original diffusion models for multi-view consistency, including epipolar attention [77], synchronized multi-view noise predictor [36] and cross-view attention [38, 70]; although these methods require fine-tuning an already pre-trained model. We adopt a different paradigm, instead of extending a single-view diffusion model with additional trainable models that incorporate multi-view conditions, our training-free method enables pre-trained diffusion models to incorporate previously generated views via the denoising step and holistically extends these models into multi-view settings.", + "bbox": [ + 496, + 159, + 890, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Other Single-view Reconstruction Methods", + "text_level": 1, + "bbox": [ + 498, + 505, + 867, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before the prosperity of generative models used in 3D reconstruction, many works [11, 12, 14, 15, 26, 29, 30, 59, 64, 75] reconstructed 3D shapes from single-view images using regression [14, 15, 29, 64, 75] or retrieval [59], both of which face difficulties in generalizing to real data or new categories. Methods based on Neural Radiance Fields (NeRFs) [41] have found success in novel-view synthesis, but these approaches typically depend on densely captured images with accurately calibrated camera positions. Currently, several studies are investigating the adaptation of NeRF to single-view settings [21, 32, 52, 79]; although, reconstructing arbitrary objects from single-view images is still a challenging problem.", + "bbox": [ + 496, + 530, + 890, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 743, + 591, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Denoising Diffusion Probabilistic Models", + "text_level": 1, + "bbox": [ + 498, + 770, + 849, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising diffusion probabilistic models (DDPM) [19, 54] are a class of generative models that model the real data distribution $q(x_0)$ with a tractable model distribution $p_{\\theta}(x_0)$ by learning to iteratively denoise samples. It learns a probability model $p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T}$ to convert unstructured noise $\\mathbf{x}_T$ to real samples $\\mathbf{x}_0$ in the form of a Markov chain, with Gaussian transitions. The Gaussian transition is", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9871", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/528efee45774322d21d034cc41d8286bfdb4fd5f9e973e81c37b10e572449e0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 80, + 885, + 176 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09a29e0c94767c8285e2f8fb88c6e6a41161cdf4c3bf0a2c04e7980e011aeb2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 178, + 885, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f65e89dee7baac846a779ab1e994dee7b4eabdd8ebb56394b0c62a31b118d7d9.jpg", + "image_caption": [ + "中", + "Figure 2. Illustration of the Auto-Regressive Generation Process. In our approach, we extend a pre-trained diffusion model from single-stage to multi-stage generation and we maintain a view set that contains all generated views. For each stage, we construct $N$ reverse diffusion processes and sharing a common starting noise. At each time step within this generation stage, the diffusion model predicts $N$ noises individually. These $N$ noises are then subjected to weighted interpolation through the Noise Interpolation Module, concluding the denoising step with the a shared interpolated noise for subsequent denoising steps." + ], + "image_footnote": [], + "bbox": [ + 78, + 271, + 885, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defined as:", + "bbox": [ + 76, + 469, + 151, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 488, + 485, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}, t \\in \\{1, \\dots, T\\}$ are the variance schedule parameter and timestep in the denoising process respectively. The reverse denoising process starts from a noise sampled from a Gaussian distribution $q(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ and is constructed as:", + "bbox": [ + 76, + 542, + 468, + 616 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {T}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 621, + 486, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the variance $\\sigma_t^2$ is a time-dependent constant [19], and $\\mu_{\\theta}(\\mathbf{x}_t,t)$ is the mean from the learned noise predictor $\\epsilon_{\\theta}$ :", + "bbox": [ + 76, + 675, + 468, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} (\\mathbf {x} _ {t}, t) = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right). \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 724, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\alpha_{t}$ and $\\bar{\\alpha}_{t}$ are constants derived from $\\beta_{t}$ . The objective of noise predictor $\\epsilon_{\\theta}$ is simplified to:", + "bbox": [ + 76, + 763, + 468, + 795 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell = \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| _ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 800, + 468, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a random variable sampled from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ [19].", + "bbox": [ + 76, + 824, + 452, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pose-Conditional Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 847, + 380, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to other generative models [42, 55], diffusion models inherently possess the capability to model conditional", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distributions of the form $p_{\\theta}(x_{t - 1}|x_t,y)$ where $y$ is the condition. We employ a conditional denoising autoencoder, denoted as $\\epsilon_{\\theta}(\\mathbf{x}_t,t,y)$ which enables controlling the synthesis process through a variety of input modalities, including textual descriptions [50], semantic maps [20, 46], or other image-to-image translation tasks [20]. In the following, we present a range of approaches to novel-view synthesis, exploring how various works, including our own, approach the concept of a single reverse diffusion step. Through this comparison, we clarify and establish the underlying relationships between these different methodologies. The notation will follow that bottom subscript $(\\cdot)_t$ indicates the diffusion step and upper subscript $(\\cdot)^i$ relates to the view index. Subsequently, the $i$ -th condition image and its relative pose to the target view are defined as $\\mathbf{y}^i$ and $\\pi^i$ , respectively, and the noisy image to be denoised at timestep $t$ is defined as $\\mathbf{x}_t$ .", + "bbox": [ + 496, + 468, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direct condition was applied by Zero 1-to-3 [34] to the reverse process when given a single input image and target pose $\\mathbf{y}^1$ , $\\pi^1$ :", + "bbox": [ + 496, + 744, + 890, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {y} ^ {1}, \\pi^ {1}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 790, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stochastic conditioning was formulated by [69] which can leverage multiple views sampled from a collection of views $p_{\\mathbf{y},\\pi}(\\mathcal{Y},\\pi)$ :", + "bbox": [ + 496, + 825, + 890, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {y} ^ {i}, \\pi^ {i}\\right), \\left\\{\\mathbf {y} ^ {i}, \\pi^ {i} \\right\\} \\sim p _ {\\mathbf {y}, \\pi} (\\mathcal {Y}, \\pi), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9872", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the sampling of image and pose happens at each diffusion step $t$ .", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint output distribution was shown in SyncDreamer[36] which learns a joint distribution of many views given an image condition $y^{1}$ :", + "bbox": [ + 76, + 157, + 468, + 203 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {t - 1} ^ {1: N} \\mid \\mathbf {x} _ {t} ^ {1: N}, \\mathbf {y} ^ {1}, e ^ {1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 213, + 468, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of generated novel views and $e$ is the elevation condition (partial pose information). We note that in this formulation the target poses are not fully specified as part of the condition allowing for diverse pose generation of outputs.", + "bbox": [ + 76, + 244, + 468, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Auto-regressive distribution is an auto-regressive distribution setting which can generate an arbitrary number of views given a single or multiple condition images and poses contained in the set of $\\mathbf{y}^{1:N - 1}$ , $\\pi^{1:N - 1}$ :", + "bbox": [ + 76, + 356, + 468, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}, \\pi^ {1: N - 1}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 426, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach falls in the auto-regressive category and for the remainder of this section we detail the implementation to achieve this sampling strategy.", + "bbox": [ + 76, + 458, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Interpolated Denoising", + "text_level": 1, + "bbox": [ + 76, + 513, + 289, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The standard DDPM model has been adapted for novel-view image synthesis by using an image and target pose (i.e., rotation and translation offsets) as conditional inputs [69]. Following training on a large-scale dataset, this approach has demonstrated the capability for zero-shot reconstruction [34]. To address the challenge of maintaining multi-view consistency, we employ an auto-regressive approach for generating sequential frames (See Fig. 2). Instead of independently producing each frame from just the input images – a process prone to significant variations between adjacent images – we integrate an auto-regressive algorithm into the diffusion process. This integration enables us to model a conditional joint distribution, ensuring smoother and more consistent transitions between frames.", + "bbox": [ + 76, + 537, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guide the synthesis of novel views using images under different views, we design an interpolated denoising process. For the purpose of this derivation, we assume access to an image set containing $N - 1$ images denoted as $\\{\\mathbf{y}^1,\\dots,\\mathbf{y}^{N - 1}\\}$ . We want to model the distribution of the $N$ -th view image conditioned on these $N - 1$ views $q(\\mathbf{x}_{1:T}^N|\\mathbf{y}^{1:N - 1})$ , where the relative pose offsets $\\pi^i,i\\in$ $\\{1,N - 1\\}$ between the condition images $\\{\\mathbf{y}^1,\\dots,\\mathbf{y}^{N - 1}\\}$ and target image $\\mathbf{x}_0^N$ are omitted for simplicity. The forward process of the multi-view conditioned diffusion model", + "bbox": [ + 76, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a direct extension of the vanilla DDPM in Eq. 1, where noises are added to every view independently by", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} ^ {N} \\mid \\mathbf {y} ^ {1: N}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} ^ {N} \\mid \\mathbf {x} _ {t - 1} ^ {N}, \\mathbf {y} ^ {1: N}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 128, + 890, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q(\\mathbf{x}_t^N | \\mathbf{x}_{t-1}^N, \\mathbf{y}^{1:N}) = \\mathcal{N}(\\mathbf{x}_t^N; \\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}^N, \\beta_t \\mathbf{I})$ . The initial is defined as $\\mathbf{x}_0^N := \\mathbf{y}^N$ . Similarly, following Eq. 2, the log reverse process is constructed as", + "bbox": [ + 498, + 176, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {N} \\mid \\mathbf {x} _ {T} ^ {N}, \\mathbf {y} ^ {1: N - 1}\\right) = \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}\\right) \\\\ \\approx_ {(1)} ^ {\\sum} \\log \\prod_ {n = 1} ^ {N - 1} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {n}\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N - 1} \\log \\mathcal {N} \\left(\\mathbf {x} _ {t - 1} ^ {N}; \\mu_ {\\theta} ^ {n} \\left(\\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {n}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t - 1} ^ {N}; \\bar {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}, t\\right), \\bar {\\sigma} _ {t} ^ {2} \\mathbf {I}\\right). \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 232, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\bar{\\mu}_{\\theta}, \\bar{\\sigma}_{t}^{2}$ are taken as the mean and variance of the summation of $N - 1$ log-normal distributions. A note on subscript (1) in Eq.10; to avoid cluttering the derivation, we assume $N - 1$ independent inferences of the same random variable $\\mathbf{x}_{t-1}^{N}$ using a different $\\mathbf{y}^{n}$ that results in $N - 1$ independent normal distributions, which would require an additional subscript that we omitted for clarity.", + "bbox": [ + 498, + 417, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Single and Multi-view Denoising", + "text_level": 1, + "bbox": [ + 500, + 531, + 784, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, however, we may not have all $N - 1$ views but a single view or a handful of views. For the reminder of this section, we treat an estimated view as $\\mathbf{x}_0^n$ , to be the $n$ -th view $\\mathbf{y}^n$ after a full reverse diffusion process. We use $\\bar{\\mu}_{\\theta}(\\mathbf{x}_t,\\mathbf{y}^{1:N - 1},t)$ as the weighted average of $\\mu_{\\theta}^{n}(\\mathbf{x}_{t},\\mathbf{y}^{n},t)$ . For computing $\\bar{\\mu}_{\\theta}$ using both given views and estimated views we adopt an approach where different views contribute differently to the target view, and we assign the weight $\\omega_{n}$ for the $n$ -th view in practice while satisfying the constraint $\\sum_{n = 1}^{N - 1}w_n = 1$ . The Noise Interpolation Module in Fig. 2 is modeled as:", + "bbox": [ + 498, + 554, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {1: N - 1}, t\\right) = \\sum_ {n = 1} ^ {N - 1} \\omega_ {n} \\mu_ {\\theta} ^ {n} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right) \\\\ = \\sum_ {n = 1} ^ {N - 1} \\omega_ {n} \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right)\\right) \\\\ = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\sum_ {n = 1} ^ {N} \\omega_ {n} \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right)\\right). \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 729, + 900, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, as the full view set is not given to us, we approximate this process by an auto-regressive way and", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9873", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "grow the condition set during the generation. We define the weight parameter $\\omega_{n}$ based on the angle offset, i.e., azimuth $(\\Delta_{a}^{n})$ , elevation $(\\Delta_{e}^{n})$ , and distance $(\\Delta_{d}^{n})$ , between the target view and the $n - th$ condition view. The core idea is to assign higher importance to near-view images during the denoising process while ensuring that the weight for the initial condition image does not diminish too rapidly, even when the target view is positioned at a nearly opposite angle. We use an exponential decay weight function for the initial condition image, defined as $\\omega_{n} = e^{-\\frac{\\Delta^{n}}{\\tau_{c}}}$ . Here, $\\tau_{c}$ is the temperature parameter that regulates the decay speed, and $\\Delta^n$ is the sum of the absolute relative azimuth $(\\Delta_{a}^{n})$ , elevation $(\\Delta_{e}^{n})$ , and distance $(\\Delta_{d}^{n})$ between the target and condition poses. We calculate $\\Delta^n$ as $\\Delta^n = |\\Delta_a^n| / \\pi +|\\Delta_e^n| / \\pi +|\\Delta_d^n|$ .", + "bbox": [ + 75, + 90, + 472, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the weights of the remaining images denoted as $\\{x_0^2,\\dots,x_0^N\\}$ , all generated from the initial condition image $y^{1}:= x_{0}^{1}$ , we use a softmax function to define the weights $\\omega_{n}$ :", + "bbox": [ + 76, + 306, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {n} = \\operatorname {S o f t m a x} \\left(\\frac {e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}{\\sum_ {n = 2} ^ {N} e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}\\right), n = 2, \\dots , N \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 378, + 468, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, $\\Delta_{n}$ represents the relative pose offset between target view and the $n$ -th generated view, and $\\tau_{g}$ represents the temperature parameter for generated views. As an example, in the single-view case, the weights are expressed as follows,", + "bbox": [ + 75, + 434, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {n} = \\left\\{ \\begin{array}{l l} e x p \\left(- \\frac {\\Delta^ {n}}{\\tau_ {c}}\\right), & n = 1 \\\\ (1 - \\omega_ {1}) \\operatorname {S o f t m a x} \\left(\\frac {e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}{\\sum_ {n = 2} ^ {N} e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}\\right), & n \\neq 1 \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 520, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we apply the term $1 - \\omega_{1}$ on the generated image weights to ensure the requirement of $\\sum_{n = 1}^{N - 1}w_n = 1$ will be met. In practice, Eq.13 is generalised to allow the condition set can be larger than 1, i.e., multi-view generation (see supplementary).", + "bbox": [ + 75, + 609, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Step-by-step Generation", + "text_level": 1, + "bbox": [ + 76, + 696, + 302, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single image generation. When applying the autoregressive approach to image generation, we have devised a generation trajectory, as illustrated in Fig. 3a. We uniformly sample views along this trajectory in sequence. Each previously generated view image on this trajectory is incorporated into the condition set, providing guidance for the subsequent denoising process via our interpolated denoising method. To determine the number of steps, denoted as $S$ , needed for this trajectory, we use the following formula:", + "bbox": [ + 75, + 720, + 468, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = \\max \\left(\\lceil \\frac {\\Delta_ {a} ^ {N}}{\\delta} \\rceil , \\lceil \\frac {\\Delta_ {e} ^ {N}}{\\delta} \\rceil\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 869, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76a32f7bd0c47340cfd1ac3d1c3fae945a5e8ae9fd01e5e586c6f6cbd17d840e.jpg", + "image_caption": [ + "(a) Single image generation." + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 679, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/150217d51a3426f2d2d7b72a52d4648341a3bb8a84f1b5c79978fb099df608be.jpg", + "image_caption": [ + "(b) Spin video generation.", + "Figure 3. Illustration of Step-by-step Generation. (a) we uniformly sample views along this trajectory in sequence to generate a novel-view image; (b) we sample views from nearest to furthest views according to to view distance to generate a $360^{\\circ}$ spin video." + ], + "image_footnote": [], + "bbox": [ + 705, + 85, + 887, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we set the maximum offset per step $\\delta$ to determine the step count $S$ , also based on the target view offsets $\\Delta_{a}^{N}$ and $\\Delta_{e}^{N}$ . We then proceed to sample the $n$ -th view using the following equation:", + "bbox": [ + 498, + 316, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\Delta_ {a} ^ {n}, \\Delta_ {e} ^ {n}, \\Delta_ {d} ^ {n}\\right) = \\left(\\frac {\\Delta_ {a} ^ {N}}{S} * n, \\frac {\\Delta_ {e} ^ {N}}{S} * n, \\frac {\\Delta_ {d} ^ {N}}{S} * n\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 390, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spin videos generation. In contrast to generating a single target image, the process of spin video generation begins from an initial image and concludes at the same position. To achieve this, we need to modify the generation order to leverage the broad range of rotation images, rather than simply following the rotation degree range of $[0^{\\circ}, 360^{\\circ}]$ in sequence. This is because, at $\\Delta_{a} = \\pi$ , the view is opposite to the conditioning view, marking the end of the generation process. To establish the generation order for spin video generation, we introduce the minimum azimuth offset, denoted as $\\delta$ , and employ a skip trajectory with the following order: $\\{\\delta, -\\delta, 2\\delta, -2\\delta\\dots, N\\delta\\}$ , shown in Fig. 3b. For simplicity, we only consider rotation along the azimuth dimension in this context.", + "bbox": [ + 498, + 436, + 892, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 665, + 633, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method and compare to baselines on the ABO [6] and GSO [9] datasets. These datasets are out-of-the-distribution as all baselines are trained on the Objaverse [7]. We also provide qualitative results on real images to showcase performance of our method on in-the-wild images in the supplementary. For additional results, please refer to the videos contained in the supplementary.", + "bbox": [ + 498, + 691, + 890, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We assess our novel-view synthesis on three main criteria:", + "bbox": [ + 500, + 823, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Image Quality: LPIPS [85], PSNR, and SSIM [67] metrics to help gauge the similarity between synthesized and ground-truth views.", + "bbox": [ + 500, + 854, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9874", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7ce42425512954afc043f1e4bb7fa203db6dc13c8bed435103fa7095531dda58.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative results on ABO and GSO datasets with arbitrary (left) and discrete (right) rotation and translation. Free renderings are a set of arbitrary rotation and translation as target generation view, while SyncDreamer renderings are a fixed set of 16 views with discrete azimuth, fixed elevation and distance, i.e., azimuth $\\in$ $\\{0^{\\circ},22.5^{\\circ},45^{\\circ},\\dots,315^{\\circ},337.5^{\\circ}\\}$ , elevation $= 30^{\\circ}$ . Note that SyncDreamer [36] cannot generate images under arbitrary views apart from the predefined 16 camera positions." + ], + "table_body": "
DatasetMethodFree RenderingsSyncDreamer Renderings
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
ABOZero123 [34]0.879621.330.09610.782218.270.1999
SyncDre. [36]0.771213.430.21820.803119.070.1816
Ours0.884821.430.09230.798318.750.1985
GSOZero123 [34]0.871020.330.10290.792518.060.1714
SyncDre. [36]0.802314.420.18330.802418.200.1647
Ours0.882020.730.09580.807618.400.1703
", + "bbox": [ + 80, + 88, + 467, + 183 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Multi-View Consistency: Using SIFT [39], LPIPS [85] and CLIP [48], we measure the uniformity of images across various perspectives.", + "3. 3D Reconstruction: Chamfer distances and F-score between ground-truth and reconstructed shapes determine geometrical consistency." + ], + "bbox": [ + 76, + 321, + 467, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Novel-view synthesis", + "text_level": 1, + "bbox": [ + 76, + 421, + 272, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we show quantitative results for novel-view synthesis under arbitrary and fixed-view settings. The fixed-view setting uses the rendering set of [36] and ensures a fair comparison to [36] which is limited to this fixed-view generation setting. As shown in Tab. 1, our method can produce comparable results to [36], without any fine-tuning on the rendering set of [36]. Under the arbitrary-view setting, we sample the nearest view from the rendering set of [36] to the designated target view for generation; [36] underperforms both [34] and our approach. Overall, regardless of the evaluation setting, our method consistently outperforms [34] and even outperforms [36] on the GSO dataset under the favourable fixed-view setting for [36].", + "bbox": [ + 75, + 444, + 467, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The task of novel view synthesis serves as a precursor for enabling more significant downstream applications, such as 3D reconstruction, by generating requisite input views. Thus, in addition to image quality, thorough evaluation needs to encompass the multi-view consistency of the synthesized perspectives. This additional criterion ensures that the generated imagery not only appears visually compelling but also aligns geometrically across different viewpoints. In the following sections, we evaluate against baselines on 3D consistency and show that the significant improvements our approach offers.", + "bbox": [ + 75, + 641, + 467, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-view Consistency. For measuring multi-view consistency of synthesized views, quantitative results are shown in Tab. 2, while qualitative comparisons are shown in Figs. 4 and 5. Here, we can see when compared to the baselines of [34] and [36], our approach excels in generating images", + "bbox": [ + 75, + 824, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9366e1829717b98dbd79e98adadb87459242e34f5e09f9986bd809919deea84.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantitative results for multi-view consistency. We report the SIFT matching point number, LPIPS and CLIP similarity between adjacent frames to evaluate the multi-view consistency. Note that SyncDreamer can only generate 16 view images for a spin video due to the constraints imposed by its training." + ], + "table_body": "
DatasetMethodViewSIFT↑LPIPS↓CLIP↑
ABOZero123 [34]13.380.17820.9604
SyncDreamer [36]1612.360.18950.9584
Ours13.510.16020.9664
Zero123 [34]17.030.12310.9725
Ours3618.010.09660.9812
GSOZero123 [34]12.580.14110.9482
SyncDreamer [36]1613.240.13150.9532
Ours13.830.11870.9601
Zero123 [34]15.200.10560.9592
Ours3617.950.06760.9773
", + "bbox": [ + 531, + 88, + 857, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/560c0d82367b56e044d421ea2c5c95e41483286281615da872e76beceadcbb3f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Quantitative results for multi-conditioned generation. Our approach outperforms the Zero-1-to-3 baseline [34] by extending its single-view reconstruction framework to a multi-view reconstruction framework. Note that 3 condition views are removed from test set, and thus the results are slightly different to Tab. 1." + ], + "table_body": "
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 [34] (1 view)0.882021.510.0945
Ours (1 view)0.887021.610.0904
Ours (2 views)0.891321.920.0887
Ours (3 views)0.899522.740.0815
GSOZero123 [34] (1 view)0.872120.420.1017
Ours (1 view)0.883020.870.0948
Ours (2 views)0.890121.250.0893
Ours (3 views)0.897921.950.0792
", + "bbox": [ + 529, + 306, + 857, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that are both semantically consistent with the input image and maintains multi-view consistency in terms of colors and geometry under arbitrary-view settings.", + "bbox": [ + 496, + 536, + 890, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Multi-view conditional setting for NVS.", + "text_level": 1, + "bbox": [ + 498, + 601, + 836, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, our approach allows for the extension of single-view conditioned models into multi-view conditioned models which can take multiple conditional images as input. The results presented in Tab. 3 showcase the advantages our method offers, as novel-view synthesis quality improves with an increasing number of conditional input views. This is a significant improvement over the Zero-1-to-3 baseline and demonstrates the efficacy of our proposed method in the multi-view setting.", + "bbox": [ + 496, + 628, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. 3D Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 782, + 679, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For shape reconstruction, we present our results both quantitatively (Tab. 4) and qualitatively (Fig. 6). We compare against baselines which synthesize novel views as well as direct image-to-shape approaches [23, 44]. For the former approach, instead of using distillation [63], we train NeuS [65] on synthesized images to recover a shape.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9875", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0fa386518989aafa33eda739ac2cfba214f2b67bda5dbad0a554e72140b8cd47.jpg", + "image_caption": [ + "Figure 4. Qualitative results for $360^{\\circ}$ Spin Video Generation. Note the additional consistency in generated views our approach offers over the competing baselines shown in the bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 101, + 77, + 867, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54e7d11b2170e71469b442caff833d81bbe54f0a89fa1ee14e4e40f8c7ac3aa5.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison for Motion Smoothness. We visualize the output videos using space-time Y-t slices through frames of the generated spin video (along the scanline shown in the condition)." + ], + "image_footnote": [], + "bbox": [ + 86, + 354, + 467, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-view Reconstruction. We first compare our proposed method with several baselines under the single-view reconstruction setting, including Point-E [44] and Shap-E [23], Zero-1-to-3 [34], One-2-3-45 [33] and SyncDreamer [36]. Qualitative results are shown in Fig. 6; Point-E and Shap-E tend to generate incomplete shapes due to the single-view setting. Multi-view alignment methods such as Syncdreamer [36] and One-2-3-45 [33] capture the general geometry but tend to lose fine details. In comparison, our proposed method achieves the highest reconstruction quality amongst all approaches, where we can generate smooth surfaces and capture detailed geometry with precision.", + "bbox": [ + 75, + 623, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-view Reconstruction Given our approach can also synthesis novel views under the multi-view setting, we also show results where we use 3 conditional input views to synthesize 36 views for shape reconstruction. Compared with existing multi-view reconstruction frameworks such as", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c62dc23edf7d8123a418b0c2563a251d04f4589bc92c0849cdbe1011da827fcb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cond. viewsMethodGen. viewsABOGSO
CD↓F-score↑CD↓F-score↑
1Point-E [44]N/A0.04280.71440.06720.6340
Shap-E [23]N/A0.04660.73640.03840.7313
One-2-3-45 [33]320.04190.66650.04080.6490
SyncDreamer [36]160.01600.81870.02290.7767
Zero123 [34]160.01470.82260.02060.8045
Zero123 [34]360.01390.82470.02070.8078
Ours160.01330.84230.01770.8274
Ours360.01260.84720.01640.8436
3MonoSDF [81]N/A0.10200.39630.08300.4581
Ours360.01150.85870.01240.8628
", + "bbox": [ + 503, + 354, + 890, + 489 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f85e0f8a3ecf6929b70da0fce80db99b3bb92d0997eef93531b8d8801bf40cfb.jpg", + "table_caption": [ + "Table 4. Quantitative results on reconstructing 3D Shapes using the generated images." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 (no interp.)0.879621.330.0961
Standard auto-regression0.801016.770.1854
Interpolated conditions0.724313.260.3770
Interpolated outputs0.8925*21.95*0.1246
Stochastic conditioning [69]0.869920.640.1106
Interpolated denoising0.884821.430.0923
GSOZero123 (no interp.)0.871020.330.1029
Standard auto-regression0.809416.300.1801
Interpolated conditions0.766114.380.3427
Interpolated outputs0.879920.440.1659
Stochastic conditioning [69]0.865819.970.1119
Interpolated denoising0.882020.730.0958
", + "bbox": [ + 519, + 523, + 867, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison between our method to different auto-regressive variants. Note that the Interpolated outputs variant achieves the highest SSIM and PSNR values on the ABO dataset, but generates blurred images.", + "bbox": [ + 496, + 679, + 890, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MonoSDF [81], our approach captures more details and generates smoother surfaces (note that [81] also relies on additional depth and normal estimations for its reconstruction whereas ours does not).", + "bbox": [ + 496, + 746, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 500, + 816, + 609, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we study the effectiveness of our proposed interpolated denoising process by exploring various autoregressive generation variants (in Tab. 5 and Fig. 7). Specifically, we investigate four variants of denoising: stan-", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9876", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e0bf45faf0c6ab9f136010eaf1f0f6dbada123e15c66b9e53d9eda58ef68c08.jpg", + "image_caption": [ + "Figure 6. Qualitative results for single-view (1 condition image) and multi-view (3 condition images) 3D Reconstruction. Our method offers the most consistent results across a variety of different reconstructed shapes." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 883, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/494e0ead145369685abfd915e3e2c9b9b2cb47fc3480f78c88f58dffd3336812.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison for different auto-regressive generations." + ], + "image_footnote": [], + "bbox": [ + 86, + 354, + 464, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dard auto-regression, interpolated conditions, interpolated outputs, and stochastic conditioning which was proposed in [69].", + "bbox": [ + 75, + 542, + 468, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Standard Auto-regression. One initial approach to auto-regression involves using the last generated view as the subsequent conditioning. However, this method exhibits bad generation quality, due to the accumulation of errors during the sequential generation process. As each subsequent view relies on the accuracy of the previous one, any inaccuracies or imperfections in earlier stages can compound, leading to a degradation in overall image quality. This limitation highlights the need for more sophisticated auto-regressive strategies to address the issue of error propagation and enhance the quality of generated views.", + "bbox": [ + 75, + 608, + 468, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interpolated Conditions and Interpolated Outputs. Interpolated Conditions and Interpolated Outputs are two straightforward approaches to introduce auto-regressive generation into an existing diffusion model. The former method involves interpolating feature embeddings from condition images and poses, while the latter interpolates the final image feature maps produced by the model. Despite", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SSIM and PSNR metrics showing favorable results for Interpolated Outputs over others in Tab. 5, as illustrated in visual comparisons in Fig. 7 shows that it leads to blurring of the output views and this is corroborated by larger LPIPS distance.", + "bbox": [ + 496, + 357, + 892, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stochastic Conditioning. We also explore the application of the Stochastic Conditioning Sampler proposed by [69] to the Zero-1-to-3 model. We observe that Stochastic Conditioning fails to deliver the desired auto-regressive generation results; this may be attributed to the specific category on which the diffusion model used by Watson et al. [69] was trained, allowing it to handle plausible condition images more effectively. By contrast, Zero-1-to-3 [34] was trained on a cross-category dataset and designed for zero-shot reconstruction. Additionally, our evaluation data contains out-of-the-distribution data (i.e., ABO [6] and GSO [9]).", + "bbox": [ + 496, + 472, + 892, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have developed ViewFusion, a novel algorithm that addresses the challenge of multi-view consistency in novel-view synthesis with diffusion models. Our approach circumvents the need for fine-tuning or additional modules by integrating an auto-regressive mechanism that incrementally refines view synthesis, utilizing the entire history of previously generated views. Our proposed diffusion interpolation technique extends the denoising process in pre-trained diffusion models from a single-view setting to a multi-view setting without training requirements. Empirical evidence underscores ViewFusion's capability to produce consistently high-quality views, and we achieve significant steps forward in novel view synthesis and 3D reconstruction applications.", + "bbox": [ + 496, + 691, + 892, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9877", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In CVPR, 2023. 2", + "[2] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 2", + "[3] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 2", + "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873, 2023. 1, 2", + "[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2", + "[6] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 2, 5, 8", + "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 5", + "[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2", + "[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, 2022. 2, 5, 8", + "[10] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. arXiv preprint arXiv:2303.17015, 2023. 2", + "[11] George Fahim, Khalid Amin, and Sameh Zarif. Single-view 3d reconstruction: A survey of deep learning methods. Computers & Graphics, 94:164-190, 2021. 2", + "[12] Kui Fu, Jiansheng Peng, Qiwen He, and Hanxiao Zhang. Single image 3d object reconstruction based on deep learning: A review. Multimedia Tools and Applications, 80:463-498, 2021. 2", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Georgia Gkioxari, Justin Johnson, and Jitendra Malik. Mesh r-cnn. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2", + "[15] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mache approach to learning 3d surface generation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2", + "[16] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images. arXiv preprint arXiv:2304.06700, 2023. 2", + "[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In ICML, 2023. 2", + "[18] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3", + "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 2, 3", + "[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2", + "[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2", + "[23] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 6, 7", + "[24] Animesh Karnewar, Niloy J Mitra, Andrea Vedaldi, and David Novotny. Holofusion: Towards photo-realistic 3d generative modeling. In ICCV, 2023. 2", + "[25] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In CVPR, 2023. 2", + "[26] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2", + "[27] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In CVPR, 2023. 2", + "[28] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. In CVPR, 2022. 2", + "[29] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2", + "[30] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdfsrn: Learning signed distance 3d object reconstruction from static images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9878", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2", + "[32] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2", + "[33] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 1, 2, 7", + "[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 1, 2, 3, 4, 6, 7, 8", + "[35] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2", + "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 2, 4, 6, 7", + "[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. In ICLR, 2023. 2", + "[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 1, 2", + "[39] D.G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 1150-1157 vol.2, 1999. 6", + "[40] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 1, 2", + "[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[42] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. CoRR, abs/1411.1784, 2014. 3", + "[43] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2", + "[44] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 6, 7", + "[45] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc Van Gool, and Sergey Tulyakov." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Autodecoding latent 3d diffusion models. arXiv preprint arXiv:2307.05445, 2023. 2", + "[46] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3", + "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[49] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 1, 2", + "[50] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2, 3", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1", + "[52] Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, and Deqing Sun. Vq3d: Learning a 3d-aware generative model on image-net. arXiv preprint arXiv:2302.06833, 2023. 2", + "[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2023. 1", + "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2", + "[55] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Neural Information Processing Systems, 2015. 3", + "[56] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2", + "[57] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 1, 2", + "[58] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2", + "[59] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9879", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] A. Tewari, O. Fried, J. Thies, V. Sitzmann, S. Lombardi, K. Sunkavalli, R. Martin-Brualla, T. Simon, J. Saragih, M. Nießner, R. Pandey, S. Fanello, G. Wetzstein, J.-Y. Zhu, C. Theobalt, M. Agrawala, E. Shechtman, D. B Goldman, and M. Zollhöfer. State of the art on neural rendering. Computer Graphics Forum, 39(2):701-727, 2020. 1", + "[61] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B Tenenbaum, Frédo Durand, William T Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. arXiv preprint arXiv:2306.11719, 2023. 2", + "[62] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2", + "[63] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 1, 2, 6", + "[64] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 1, 2", + "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 1, 6", + "[66] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2", + "[67] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5", + "[68] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2", + "[69] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 1, 2, 3, 4, 7, 8", + "[70] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 2", + "[71] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2", + "[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1", + "[73] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints, pages arXiv-2211, 2022. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion, 2023. 1, 2", + "[75] Xianghui Yang, Guosheng Lin, and Luping Zhou. Single-view 3d mesh reconstruction for seen and unseen categories. IEEE Transactions on Image Processing, pages 1-1, 2023. 1, 2", + "[76] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 1", + "[77] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models, 2023. 1, 2", + "[78] Paul Yoo, Jiaxian Guo, Yutaka Matsuo, and Shixiang Shane Gu. Dreamsparse: Escaping from Plato's cave with 2d frozen diffusion model given sparse views. CoRR, 2023. 2", + "[79] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[80] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2", + "[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2022. 7", + "[82] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 2", + "[83] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. In SIGGRAPH, 2023. 2", + "[84] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2", + "[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6", + "[86] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 742 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9880", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_model.json b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87a144bc3ed3e316801cf436f6c761133a431acd --- /dev/null +++ b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "ViewFusion: Towards Multi-View Consistency via Interpolated Denoising" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.18, + 0.944, + 0.218 + ], + "angle": 0, + "content": "Xianghui Yang\\(^{1,2}\\), Yan Zuo\\(^{1}\\), Sameera Ramasinghe\\(^{1}\\), Loris Bazzani\\(^{1}\\), Gil Avraham\\(^{1}\\), Anton van den Hengel\\(^{1,3}\\), \\({}^{1}\\)Amazon, \\({}^{2}\\)The University of Sydney, \\({}^{3}\\)The University of Adelaide" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.251, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.283, + 0.474, + 0.555 + ], + "angle": 0, + "content": "Novel-view synthesis through diffusion models has demonstrated remarkable potential for generating diverse and high-quality images. Yet, the independent process of image generation in these prevailing methods leads to challenges in maintaining multiple view consistency. To address this, we introduce ViewFusion, a novel, training-free algorithm that can be seamlessly integrated into existing pre-trained diffusion models. Our approach adopts an auto-regressive method that implicitly leverages previously generated views as context for next view generation, ensuring robust multi-view consistency during the novel-view generation process. Through a diffusion process that fuses known-view information via interpolated denoising, our framework successfully extends single-view conditioned models to work in multiple-view conditional settings without any additional fine-tuning. Extensive experimental results demonstrate the effectiveness of ViewFusion in generating consistent and detailed novel views." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.582, + 0.21, + 0.599 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Humans have a remarkable capacity for visualizing unseen perspectives from just a single image view - an intuitive process that remains complex to model. Such an ability is known as Novel View Synthesis (NVS) and necessitates robust geometric priors to accurately infer three-dimensional details from flat imagery; lifting from a two-dimensional projection to a three-dimensional form involves assumptions and knowledge about the nature of the object and space. Recently, significant advancements in NVS have been brought forward by neural networks [14, 15, 30, 41, 60, 64, 65, 72, 75, 76], where novel view generation for downstream reconstruction shows promising potential [34, 69]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Specifically, diffusion models [19, 51] and their ability to generate high-quality 2D images have garnered significant attention in the 3D domain, where pre-trained, text-conditioned 2D diffusion models have been re-purposed for 3D applications via distillation [4, 31, 40, 47, 49, 57, 63," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.248, + 0.695, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.382, + 0.671, + 0.394 + ], + "angle": 0, + "content": "(a) Inconsistent generation." + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.251, + 0.889, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.382, + 0.877, + 0.394 + ], + "angle": 0, + "content": "(b) Consistent generation (ours)." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.54 + ], + "angle": 0, + "content": "Figure 1. The cause of multi-view inconsistency in diffusion-based novel-view synthesis models. (a) Diffusion models incorporate randomness for diversity and better distribution modeling; this independent generation process produces realistic views under specific instances but may produce different plausible views for various instances, lacking alignment across adjacent views. (b) In contrast, ViewFusion incorporates an auto-regressive process to reduce uncertainty and achieve multi-view consistency, by ensuring a correlated denoising process that ends at the same high-density area, fostering consistency across views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.658 + ], + "angle": 0, + "content": "68, 73]. Follow-up approaches [34, 69] remove the requirement of text conditioning and instead take an image and target pose as conditions for NVS. However, distillation [63] is still required as the diffusion model cannot produce the multi-view consistent outputs that are appropriate for certain downstream tasks (e.g., optimizing Neural Radiance Fields (NeRFs) [41])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Under the single-view setting, maintaining multi-view consistency remains particularly challenging since there may exist several plausible outputs for a novel view that are aligned with the given input image. For diffusion-based approaches which generate novel views in an independent manner [34, 69], this results in synthesized views containing artifacts of multi-view inconsistency (Fig. 1a). Previous work [33, 36, 38, 53, 74, 77] focuses on improving the robustness of the downstream reconstruction to address the inconsistency issue, including feature projection layers in the NeRF [33] or utilising three-dimensional priors to constrain NeRF optimization [36, 77], yet these techniques require training or fine-tuning to align additional modules to the original diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we address the multi-view inconsistency that arises during the process of view synthesis. Rather than" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.316, + 0.9 + ], + "angle": 0, + "content": "*Work done during internship at Amazon" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "independently synthesizing views conditioned only on the initial reference image, we develop a novel approach where each subsequently generated view is also conditioned on the entire set of previously generated views. Specifically, our method incorporates an auto-regressive process into the diffusion process to model the joint distribution of views, guiding our novel-view synthesis by maintaining the denoising direction towards the same high density area of already generated views (Fig. 1b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Our framework, named ViewFusion, relaxes the single-view conditioning requirement of typical diffusion models through an interpolated denoising process. ViewFusion offers several additional advantages: 1) it can utilize all available views as guidance, thereby enhancing the quality of generated images by incorporating more information; 2) it does not require any additional fine-tuning, effortlessly converting pre-trained single-view conditioned diffusion models into multi-view conditioned diffusion models; 3) it provides greater flexibility in setting adaptive weights for condition images based on their relative view distance to the target view." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.41, + 0.423, + 0.426 + ], + "angle": 0, + "content": "The contributions of this paper are the following:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.429, + 0.47, + 0.489 + ], + "angle": 0, + "content": "- We propose a training-free algorithm which can be directly applied to pre-trained diffusion models to improve multi-view consistency of synthesized views and supports multiple conditional inputs." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.489, + 0.47, + 0.549 + ], + "angle": 0, + "content": "- Our method utilizes a novel, auto-regressive approach which we call Interpolated Denoising, that implicitly addresses key limitations of previous auto-regressive approaches for view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.55, + 0.47, + 0.639 + ], + "angle": 0, + "content": "- Extensive empirical analysis on ABO [6] and GSO [9] show that our method is able to achieve better 3D consistency in image generation, leading to significant improvements in novel view synthesis and 3D reconstruction of shapes under single-view and multi-view image settings over other baseline methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.429, + 0.47, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.22, + 0.672 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.34, + 0.697 + ], + "angle": 0, + "content": "2.1. 3D-adapted Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Diffusion models have excelled in image generation using conditional inputs [20, 46, 50, 84] and given this success in the 2D domain, recent works have tried to extend diffusion models to 3D content generation [1, 3, 5, 10, 16, 18, 22, 24, 25, 27, 37, 43-45, 66, 82, 83] - although the scarcity of 3D data presents a significant challenge to directly train these diffusion models. Nonetheless, pioneer works such as DreamFusion [47] and Score Jacobian Chaining [63] leverage pre-trained text-conditioned diffusion models to craft 3D models via distillation. Follow-up approaches [4, 31, 57, 68] improve this distillation in terms of speed, resolution and shape quality. Approaches such as [40, 49, 57, 73] extend upon this to support image condi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "tions through the use of captions with limited success due to the non-trivial nature of textual inversion [13]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.135, + 0.836, + 0.151 + ], + "angle": 0, + "content": "2.2. Novel View Synthesis Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Another line of research [2, 8, 17, 28, 35, 56, 58, 61, 62, 69, 71, 78, 80, 86] directly applies 2D diffusion models to generate multi-view images for shape reconstruction. To circumvent the weakness of text-conditioned diffusion models, novel-view synthesis diffusion models [34, 69] have also been explored, which take an image and target pose as conditions to generate novel views. However, for these approaches, recovering a 3D consistent shape is still a key challenge. To mitigate 3D inconsistency, Liu et al. [33] suggests training a Neural Radiance Field (NeRF) with feature projection layers. Concurrently, other works [36, 38, 70, 74, 77] add modules to original diffusion models for multi-view consistency, including epipolar attention [77], synchronized multi-view noise predictor [36] and cross-view attention [38, 70]; although these methods require fine-tuning an already pre-trained model. We adopt a different paradigm, instead of extending a single-view diffusion model with additional trainable models that incorporate multi-view conditions, our training-free method enables pre-trained diffusion models to incorporate previously generated views via the denoising step and holistically extends these models into multi-view settings." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.506, + 0.868, + 0.521 + ], + "angle": 0, + "content": "2.3. Other Single-view Reconstruction Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Before the prosperity of generative models used in 3D reconstruction, many works [11, 12, 14, 15, 26, 29, 30, 59, 64, 75] reconstructed 3D shapes from single-view images using regression [14, 15, 29, 64, 75] or retrieval [59], both of which face difficulties in generalizing to real data or new categories. Methods based on Neural Radiance Fields (NeRFs) [41] have found success in novel-view synthesis, but these approaches typically depend on densely captured images with accurately calibrated camera positions. Currently, several studies are investigating the adaptation of NeRF to single-view settings [21, 32, 52, 79]; although, reconstructing arbitrary objects from single-view images is still a challenging problem." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.744, + 0.592, + 0.76 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.85, + 0.787 + ], + "angle": 0, + "content": "3.1. Denoising Diffusion Probabilistic Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Denoising diffusion probabilistic models (DDPM) [19, 54] are a class of generative models that model the real data distribution \\( q(x_0) \\) with a tractable model distribution \\( p_{\\theta}(x_0) \\) by learning to iteratively denoise samples. It learns a probability model \\( p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T} \\) to convert unstructured noise \\( \\mathbf{x}_T \\) to real samples \\( \\mathbf{x}_0 \\) in the form of a Markov chain, with Gaussian transitions. The Gaussian transition is" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.081, + 0.887, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.179, + 0.887, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.272, + 0.887, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.363, + 0.488, + 0.378 + ], + "angle": 0, + "content": "中" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Auto-Regressive Generation Process. In our approach, we extend a pre-trained diffusion model from single-stage to multi-stage generation and we maintain a view set that contains all generated views. For each stage, we construct \\(N\\) reverse diffusion processes and sharing a common starting noise. At each time step within this generation stage, the diffusion model predicts \\(N\\) noises individually. These \\(N\\) noises are then subjected to weighted interpolation through the Noise Interpolation Module, concluding the denoising step with the a shared interpolated noise for subsequent denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.153, + 0.483 + ], + "angle": 0, + "content": "defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.489, + 0.486, + 0.542 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.543, + 0.47, + 0.617 + ], + "angle": 0, + "content": "where \\(\\beta_{t}, t \\in \\{1, \\dots, T\\}\\) are the variance schedule parameter and timestep in the denoising process respectively. The reverse denoising process starts from a noise sampled from a Gaussian distribution \\(q(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) and is constructed as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.622, + 0.487, + 0.675 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {T}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.676, + 0.47, + 0.72 + ], + "angle": 0, + "content": "where the variance \\(\\sigma_t^2\\) is a time-dependent constant [19], and \\(\\mu_{\\theta}(\\mathbf{x}_t,t)\\) is the mean from the learned noise predictor \\(\\epsilon_{\\theta}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.726, + 0.47, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} (\\mathbf {x} _ {t}, t) = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right). \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Here, \\(\\alpha_{t}\\) and \\(\\bar{\\alpha}_{t}\\) are constants derived from \\(\\beta_{t}\\). The objective of noise predictor \\(\\epsilon_{\\theta}\\) is simplified to:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.801, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\ell = \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| _ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.454, + 0.841 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a random variable sampled from \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) [19]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.382, + 0.863 + ], + "angle": 0, + "content": "3.2. Pose-Conditional Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Similar to other generative models [42, 55], diffusion models inherently possess the capability to model conditional" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.728 + ], + "angle": 0, + "content": "distributions of the form \\( p_{\\theta}(x_{t - 1}|x_t,y) \\) where \\( y \\) is the condition. We employ a conditional denoising autoencoder, denoted as \\( \\epsilon_{\\theta}(\\mathbf{x}_t,t,y) \\) which enables controlling the synthesis process through a variety of input modalities, including textual descriptions [50], semantic maps [20, 46], or other image-to-image translation tasks [20]. In the following, we present a range of approaches to novel-view synthesis, exploring how various works, including our own, approach the concept of a single reverse diffusion step. Through this comparison, we clarify and establish the underlying relationships between these different methodologies. The notation will follow that bottom subscript \\( (\\cdot)_t \\) indicates the diffusion step and upper subscript \\( (\\cdot)^i \\) relates to the view index. Subsequently, the \\( i \\)-th condition image and its relative pose to the target view are defined as \\( \\mathbf{y}^i \\) and \\( \\pi^i \\), respectively, and the noisy image to be denoised at timestep \\( t \\) is defined as \\( \\mathbf{x}_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Direct condition was applied by Zero 1-to-3 [34] to the reverse process when given a single input image and target pose \\(\\mathbf{y}^1\\), \\(\\pi^1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.791, + 0.892, + 0.81 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {y} ^ {1}, \\pi^ {1}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.874 + ], + "angle": 0, + "content": "Stochastic conditioning was formulated by [69] which can leverage multiple views sampled from a collection of views \\( p_{\\mathbf{y},\\pi}(\\mathcal{Y},\\pi) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.885, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {y} ^ {i}, \\pi^ {i}\\right), \\left\\{\\mathbf {y} ^ {i}, \\pi^ {i} \\right\\} \\sim p _ {\\mathbf {y}, \\pi} (\\mathcal {Y}, \\pi), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "where the sampling of image and pose happens at each diffusion step \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.158, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Joint output distribution was shown in SyncDreamer[36] which learns a joint distribution of many views given an image condition \\( y^{1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.214, + 0.469, + 0.233 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {t - 1} ^ {1: N} \\mid \\mathbf {x} _ {t} ^ {1: N}, \\mathbf {y} ^ {1}, e ^ {1}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.245, + 0.469, + 0.321 + ], + "angle": 0, + "content": "where \\(N\\) is the number of generated novel views and \\(e\\) is the elevation condition (partial pose information). We note that in this formulation the target poses are not fully specified as part of the condition allowing for diverse pose generation of outputs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.469, + 0.416 + ], + "angle": 0, + "content": "Auto-regressive distribution is an auto-regressive distribution setting which can generate an arbitrary number of views given a single or multiple condition images and poses contained in the set of \\(\\mathbf{y}^{1:N - 1}\\), \\(\\pi^{1:N - 1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.428, + 0.469, + 0.447 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}, \\pi^ {1: N - 1}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.469, + 0.504 + ], + "angle": 0, + "content": "Our approach falls in the auto-regressive category and for the remainder of this section we detail the implementation to achieve this sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.515, + 0.29, + 0.531 + ], + "angle": 0, + "content": "3.3. Interpolated Denoising" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.469, + 0.749 + ], + "angle": 0, + "content": "The standard DDPM model has been adapted for novel-view image synthesis by using an image and target pose (i.e., rotation and translation offsets) as conditional inputs [69]. Following training on a large-scale dataset, this approach has demonstrated the capability for zero-shot reconstruction [34]. To address the challenge of maintaining multi-view consistency, we employ an auto-regressive approach for generating sequential frames (See Fig. 2). Instead of independently producing each frame from just the input images – a process prone to significant variations between adjacent images – we integrate an auto-regressive algorithm into the diffusion process. This integration enables us to model a conditional joint distribution, ensuring smoother and more consistent transitions between frames." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To guide the synthesis of novel views using images under different views, we design an interpolated denoising process. For the purpose of this derivation, we assume access to an image set containing \\(N - 1\\) images denoted as \\(\\{\\mathbf{y}^1,\\dots,\\mathbf{y}^{N - 1}\\}\\). We want to model the distribution of the \\(N\\)-th view image conditioned on these \\(N - 1\\) views \\(q(\\mathbf{x}_{1:T}^N|\\mathbf{y}^{1:N - 1})\\), where the relative pose offsets \\(\\pi^i,i\\in\\) \\(\\{1,N - 1\\}\\) between the condition images \\(\\{\\mathbf{y}^1,\\dots,\\mathbf{y}^{N - 1}\\}\\) and target image \\(\\mathbf{x}_0^N\\) are omitted for simplicity. The forward process of the multi-view conditioned diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "is a direct extension of the vanilla DDPM in Eq. 1, where noises are added to every view independently by" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.129, + 0.892, + 0.17 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} ^ {N} \\mid \\mathbf {y} ^ {1: N}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} ^ {N} \\mid \\mathbf {x} _ {t - 1} ^ {N}, \\mathbf {y} ^ {1: N}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.178, + 0.892, + 0.226 + ], + "angle": 0, + "content": "where \\( q(\\mathbf{x}_t^N | \\mathbf{x}_{t-1}^N, \\mathbf{y}^{1:N}) = \\mathcal{N}(\\mathbf{x}_t^N; \\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}^N, \\beta_t \\mathbf{I}) \\). The initial is defined as \\( \\mathbf{x}_0^N := \\mathbf{y}^N \\). Similarly, following Eq. 2, the log reverse process is constructed as" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.233, + 0.891, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {N} \\mid \\mathbf {x} _ {T} ^ {N}, \\mathbf {y} ^ {1: N - 1}\\right) = \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}\\right) \\\\ \\approx_ {(1)} ^ {\\sum} \\log \\prod_ {n = 1} ^ {N - 1} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {n}\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N - 1} \\log \\mathcal {N} \\left(\\mathbf {x} _ {t - 1} ^ {N}; \\mu_ {\\theta} ^ {n} \\left(\\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {n}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t - 1} ^ {N}; \\bar {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}, t\\right), \\bar {\\sigma} _ {t} ^ {2} \\mathbf {I}\\right). \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.419, + 0.892, + 0.524 + ], + "angle": 0, + "content": "Where \\(\\bar{\\mu}_{\\theta}, \\bar{\\sigma}_{t}^{2}\\) are taken as the mean and variance of the summation of \\(N - 1\\) log-normal distributions. A note on subscript (1) in Eq.10; to avoid cluttering the derivation, we assume \\(N - 1\\) independent inferences of the same random variable \\(\\mathbf{x}_{t-1}^{N}\\) using a different \\(\\mathbf{y}^{n}\\) that results in \\(N - 1\\) independent normal distributions, which would require an additional subscript that we omitted for clarity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.785, + 0.548 + ], + "angle": 0, + "content": "3.4. Single and Multi-view Denoising" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.721 + ], + "angle": 0, + "content": "In practice, however, we may not have all \\(N - 1\\) views but a single view or a handful of views. For the reminder of this section, we treat an estimated view as \\(\\mathbf{x}_0^n\\), to be the \\(n\\)-th view \\(\\mathbf{y}^n\\) after a full reverse diffusion process. We use \\(\\bar{\\mu}_{\\theta}(\\mathbf{x}_t,\\mathbf{y}^{1:N - 1},t)\\) as the weighted average of \\(\\mu_{\\theta}^{n}(\\mathbf{x}_{t},\\mathbf{y}^{n},t)\\). For computing \\(\\bar{\\mu}_{\\theta}\\) using both given views and estimated views we adopt an approach where different views contribute differently to the target view, and we assign the weight \\(\\omega_{n}\\) for the \\(n\\)-th view in practice while satisfying the constraint \\(\\sum_{n = 1}^{N - 1}w_n = 1\\). The Noise Interpolation Module in Fig. 2 is modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.73, + 0.901, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {1: N - 1}, t\\right) = \\sum_ {n = 1} ^ {N - 1} \\omega_ {n} \\mu_ {\\theta} ^ {n} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right) \\\\ = \\sum_ {n = 1} ^ {N - 1} \\omega_ {n} \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right)\\right) \\\\ = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\sum_ {n = 1} ^ {N} \\omega_ {n} \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right)\\right). \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our approach, as the full view set is not given to us, we approximate this process by an auto-regressive way and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9873" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.307 + ], + "angle": 0, + "content": "grow the condition set during the generation. We define the weight parameter \\(\\omega_{n}\\) based on the angle offset, i.e., azimuth \\((\\Delta_{a}^{n})\\), elevation \\((\\Delta_{e}^{n})\\), and distance \\((\\Delta_{d}^{n})\\), between the target view and the \\(n - th\\) condition view. The core idea is to assign higher importance to near-view images during the denoising process while ensuring that the weight for the initial condition image does not diminish too rapidly, even when the target view is positioned at a nearly opposite angle. We use an exponential decay weight function for the initial condition image, defined as \\(\\omega_{n} = e^{-\\frac{\\Delta^{n}}{\\tau_{c}}}\\). Here, \\(\\tau_{c}\\) is the temperature parameter that regulates the decay speed, and \\(\\Delta^n\\) is the sum of the absolute relative azimuth \\((\\Delta_{a}^{n})\\), elevation \\((\\Delta_{e}^{n})\\), and distance \\((\\Delta_{d}^{n})\\) between the target and condition poses. We calculate \\(\\Delta^n\\) as \\(\\Delta^n = |\\Delta_a^n| / \\pi +|\\Delta_e^n| / \\pi +|\\Delta_d^n|\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.308, + 0.47, + 0.368 + ], + "angle": 0, + "content": "For the weights of the remaining images denoted as \\(\\{x_0^2,\\dots,x_0^N\\}\\), all generated from the initial condition image \\(y^{1}:= x_{0}^{1}\\), we use a softmax function to define the weights \\(\\omega_{n}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.379, + 0.47, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\omega_ {n} = \\operatorname {S o f t m a x} \\left(\\frac {e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}{\\sum_ {n = 2} ^ {N} e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}\\right), n = 2, \\dots , N \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Similarly, \\(\\Delta_{n}\\) represents the relative pose offset between target view and the \\(n\\)-th generated view, and \\(\\tau_{g}\\) represents the temperature parameter for generated views. As an example, in the single-view case, the weights are expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.521, + 0.47, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\omega_ {n} = \\left\\{ \\begin{array}{l l} e x p \\left(- \\frac {\\Delta^ {n}}{\\tau_ {c}}\\right), & n = 1 \\\\ (1 - \\omega_ {1}) \\operatorname {S o f t m a x} \\left(\\frac {e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}{\\sum_ {n = 2} ^ {N} e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}\\right), & n \\neq 1 \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.687 + ], + "angle": 0, + "content": "we apply the term \\(1 - \\omega_{1}\\) on the generated image weights to ensure the requirement of \\(\\sum_{n = 1}^{N - 1}w_n = 1\\) will be met. In practice, Eq.13 is generalised to allow the condition set can be larger than 1, i.e., multi-view generation (see supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.303, + 0.714 + ], + "angle": 0, + "content": "3.5. Step-by-step Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Single image generation. When applying the autoregressive approach to image generation, we have devised a generation trajectory, as illustrated in Fig. 3a. We uniformly sample views along this trajectory in sequence. Each previously generated view image on this trajectory is incorporated into the condition set, providing guidance for the subsequent denoising process via our interpolated denoising method. To determine the number of steps, denoted as \\( S \\), needed for this trajectory, we use the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.87, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\nS = \\max \\left(\\lceil \\frac {\\Delta_ {a} ^ {N}}{\\delta} \\rceil , \\lceil \\frac {\\Delta_ {e} ^ {N}}{\\delta} \\rceil\\right). \\tag {14}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.681, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.228, + 0.666, + 0.241 + ], + "angle": 0, + "content": "(a) Single image generation." + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.087, + 0.888, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.228, + 0.868, + 0.241 + ], + "angle": 0, + "content": "(b) Spin video generation." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 3. Illustration of Step-by-step Generation. (a) we uniformly sample views along this trajectory in sequence to generate a novel-view image; (b) we sample views from nearest to furthest views according to to view distance to generate a \\(360^{\\circ}\\) spin video." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Here, we set the maximum offset per step \\(\\delta\\) to determine the step count \\(S\\), also based on the target view offsets \\(\\Delta_{a}^{N}\\) and \\(\\Delta_{e}^{N}\\). We then proceed to sample the \\(n\\)-th view using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.391, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\left(\\Delta_ {a} ^ {n}, \\Delta_ {e} ^ {n}, \\Delta_ {d} ^ {n}\\right) = \\left(\\frac {\\Delta_ {a} ^ {N}}{S} * n, \\frac {\\Delta_ {e} ^ {N}}{S} * n, \\frac {\\Delta_ {d} ^ {N}}{S} * n\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.893, + 0.648 + ], + "angle": 0, + "content": "Spin videos generation. In contrast to generating a single target image, the process of spin video generation begins from an initial image and concludes at the same position. To achieve this, we need to modify the generation order to leverage the broad range of rotation images, rather than simply following the rotation degree range of \\([0^{\\circ}, 360^{\\circ}]\\) in sequence. This is because, at \\(\\Delta_{a} = \\pi\\), the view is opposite to the conditioning view, marking the end of the generation process. To establish the generation order for spin video generation, we introduce the minimum azimuth offset, denoted as \\(\\delta\\), and employ a skip trajectory with the following order: \\(\\{\\delta, -\\delta, 2\\delta, -2\\delta\\dots, N\\delta\\}\\), shown in Fig. 3b. For simplicity, we only consider rotation along the azimuth dimension in this context." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.634, + 0.684 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.693, + 0.892, + 0.8 + ], + "angle": 0, + "content": "Datasets. We evaluate our method and compare to baselines on the ABO [6] and GSO [9] datasets. These datasets are out-of-the-distribution as all baselines are trained on the Objaverse [7]. We also provide qualitative results on real images to showcase performance of our method on in-the-wild images in the supplementary. For additional results, please refer to the videos contained in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.824, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Metrics. We assess our novel-view synthesis on three main criteria:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1. Image Quality: LPIPS [85], PSNR, and SSIM [67] metrics to help gauge the similarity between synthesized and ground-truth views." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9874" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.184 + ], + "angle": 0, + "content": "
DatasetMethodFree RenderingsSyncDreamer Renderings
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
ABOZero123 [34]0.879621.330.09610.782218.270.1999
SyncDre. [36]0.771213.430.21820.803119.070.1816
Ours0.884821.430.09230.798318.750.1985
GSOZero123 [34]0.871020.330.10290.792518.060.1714
SyncDre. [36]0.802314.420.18330.802418.200.1647
Ours0.882020.730.09580.807618.400.1703
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.298 + ], + "angle": 0, + "content": "Table 1. Quantitative results on ABO and GSO datasets with arbitrary (left) and discrete (right) rotation and translation. Free renderings are a set of arbitrary rotation and translation as target generation view, while SyncDreamer renderings are a fixed set of 16 views with discrete azimuth, fixed elevation and distance, i.e., azimuth \\(\\in\\) \\(\\{0^{\\circ},22.5^{\\circ},45^{\\circ},\\dots,315^{\\circ},337.5^{\\circ}\\}\\), elevation \\(= 30^{\\circ}\\). Note that SyncDreamer [36] cannot generate images under arbitrary views apart from the predefined 16 camera positions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.323, + 0.468, + 0.367 + ], + "angle": 0, + "content": "2. Multi-View Consistency: Using SIFT [39], LPIPS [85] and CLIP [48], we measure the uniformity of images across various perspectives." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.368, + 0.468, + 0.413 + ], + "angle": 0, + "content": "3. 3D Reconstruction: Chamfer distances and F-score between ground-truth and reconstructed shapes determine geometrical consistency." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.323, + 0.468, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.273, + 0.437 + ], + "angle": 0, + "content": "4.1. Novel-view synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.468, + 0.641 + ], + "angle": 0, + "content": "In Tab. 1, we show quantitative results for novel-view synthesis under arbitrary and fixed-view settings. The fixed-view setting uses the rendering set of [36] and ensures a fair comparison to [36] which is limited to this fixed-view generation setting. As shown in Tab. 1, our method can produce comparable results to [36], without any fine-tuning on the rendering set of [36]. Under the arbitrary-view setting, we sample the nearest view from the rendering set of [36] to the designated target view for generation; [36] underperforms both [34] and our approach. Overall, regardless of the evaluation setting, our method consistently outperforms [34] and even outperforms [36] on the GSO dataset under the favourable fixed-view setting for [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.807 + ], + "angle": 0, + "content": "The task of novel view synthesis serves as a precursor for enabling more significant downstream applications, such as 3D reconstruction, by generating requisite input views. Thus, in addition to image quality, thorough evaluation needs to encompass the multi-view consistency of the synthesized perspectives. This additional criterion ensures that the generated imagery not only appears visually compelling but also aligns geometrically across different viewpoints. In the following sections, we evaluate against baselines on 3D consistency and show that the significant improvements our approach offers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Multi-view Consistency. For measuring multi-view consistency of synthesized views, quantitative results are shown in Tab. 2, while qualitative comparisons are shown in Figs. 4 and 5. Here, we can see when compared to the baselines of [34] and [36], our approach excels in generating images" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.089, + 0.859, + 0.226 + ], + "angle": 0, + "content": "
DatasetMethodViewSIFT↑LPIPS↓CLIP↑
ABOZero123 [34]13.380.17820.9604
SyncDreamer [36]1612.360.18950.9584
Ours13.510.16020.9664
Zero123 [34]17.030.12310.9725
Ours3618.010.09660.9812
GSOZero123 [34]12.580.14110.9482
SyncDreamer [36]1613.240.13150.9532
Ours13.830.11870.9601
Zero123 [34]15.200.10560.9592
Ours3617.950.06760.9773
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.227, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Table 2. Quantitative results for multi-view consistency. We report the SIFT matching point number, LPIPS and CLIP similarity between adjacent frames to evaluate the multi-view consistency. Note that SyncDreamer can only generate 16 view images for a spin video due to the constraints imposed by its training." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.307, + 0.859, + 0.43 + ], + "angle": 0, + "content": "
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 [34] (1 view)0.882021.510.0945
Ours (1 view)0.887021.610.0904
Ours (2 views)0.891321.920.0887
Ours (3 views)0.899522.740.0815
GSOZero123 [34] (1 view)0.872120.420.1017
Ours (1 view)0.883020.870.0948
Ours (2 views)0.890121.250.0893
Ours (3 views)0.897921.950.0792
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.513 + ], + "angle": 0, + "content": "Table 3. Quantitative results for multi-conditioned generation. Our approach outperforms the Zero-1-to-3 baseline [34] by extending its single-view reconstruction framework to a multi-view reconstruction framework. Note that 3 condition views are removed from test set, and thus the results are slightly different to Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.583 + ], + "angle": 0, + "content": "that are both semantically consistent with the input image and maintains multi-view consistency in terms of colors and geometry under arbitrary-view settings." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.602, + 0.838, + 0.618 + ], + "angle": 0, + "content": "4.2. Multi-view conditional setting for NVS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Additionally, our approach allows for the extension of single-view conditioned models into multi-view conditioned models which can take multiple conditional images as input. The results presented in Tab. 3 showcase the advantages our method offers, as novel-view synthesis quality improves with an increasing number of conditional input views. This is a significant improvement over the Zero-1-to-3 baseline and demonstrates the efficacy of our proposed method in the multi-view setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.681, + 0.799 + ], + "angle": 0, + "content": "4.3. 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For shape reconstruction, we present our results both quantitatively (Tab. 4) and qualitatively (Fig. 6). We compare against baselines which synthesize novel views as well as direct image-to-shape approaches [23, 44]. For the former approach, instead of using distillation [63], we train NeuS [65] on synthesized images to recover a shape." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9875" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.078, + 0.868, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for \\(360^{\\circ}\\) Spin Video Generation. Note the additional consistency in generated views our approach offers over the competing baselines shown in the bounding boxes." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.356, + 0.468, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.551, + 0.472, + 0.607 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison for Motion Smoothness. We visualize the output videos using space-time Y-t slices through frames of the generated spin video (along the scanline shown in the condition)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Single-view Reconstruction. We first compare our proposed method with several baselines under the single-view reconstruction setting, including Point-E [44] and Shap-E [23], Zero-1-to-3 [34], One-2-3-45 [33] and SyncDreamer [36]. Qualitative results are shown in Fig. 6; Point-E and Shap-E tend to generate incomplete shapes due to the single-view setting. Multi-view alignment methods such as Syncdreamer [36] and One-2-3-45 [33] capture the general geometry but tend to lose fine details. In comparison, our proposed method achieves the highest reconstruction quality amongst all approaches, where we can generate smooth surfaces and capture detailed geometry with precision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Multi-view Reconstruction Given our approach can also synthesis novel views under the multi-view setting, we also show results where we use 3 conditional input views to synthesize 36 views for shape reconstruction. Compared with existing multi-view reconstruction frameworks such as" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.356, + 0.892, + 0.49 + ], + "angle": 0, + "content": "
Cond. viewsMethodGen. viewsABOGSO
CD↓F-score↑CD↓F-score↑
1Point-E [44]N/A0.04280.71440.06720.6340
Shap-E [23]N/A0.04660.73640.03840.7313
One-2-3-45 [33]320.04190.66650.04080.6490
SyncDreamer [36]160.01600.81870.02290.7767
Zero123 [34]160.01470.82260.02060.8045
Zero123 [34]360.01390.82470.02070.8078
Ours160.01330.84230.01770.8274
Ours360.01260.84720.01640.8436
3MonoSDF [81]N/A0.10200.39630.08300.4581
Ours360.01150.85870.01240.8628
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.492, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Table 4. Quantitative results on reconstructing 3D Shapes using the generated images." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.524, + 0.868, + 0.678 + ], + "angle": 0, + "content": "
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 (no interp.)0.879621.330.0961
Standard auto-regression0.801016.770.1854
Interpolated conditions0.724313.260.3770
Interpolated outputs0.8925*21.95*0.1246
Stochastic conditioning [69]0.869920.640.1106
Interpolated denoising0.884821.430.0923
GSOZero123 (no interp.)0.871020.330.1029
Standard auto-regression0.809416.300.1801
Interpolated conditions0.766114.380.3427
Interpolated outputs0.879920.440.1659
Stochastic conditioning [69]0.865819.970.1119
Interpolated denoising0.882020.730.0958
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison between our method to different auto-regressive variants. Note that the Interpolated outputs variant achieves the highest SSIM and PSNR values on the ABO dataset, but generates blurred images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.809 + ], + "angle": 0, + "content": "MonoSDF [81], our approach captures more details and generates smoother surfaces (note that [81] also relies on additional depth and normal estimations for its reconstruction whereas ours does not)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.611, + 0.832 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Here, we study the effectiveness of our proposed interpolated denoising process by exploring various autoregressive generation variants (in Tab. 5 and Fig. 7). Specifically, we investigate four variants of denoising: stan-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.885, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Qualitative results for single-view (1 condition image) and multi-view (3 condition images) 3D Reconstruction. Our method offers the most consistent results across a variety of different reconstructed shapes." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.355, + 0.465, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.527 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison for different auto-regressive generations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.588 + ], + "angle": 0, + "content": "dard auto-regression, interpolated conditions, interpolated outputs, and stochastic conditioning which was proposed in [69]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Standard Auto-regression. One initial approach to auto-regression involves using the last generated view as the subsequent conditioning. However, this method exhibits bad generation quality, due to the accumulation of errors during the sequential generation process. As each subsequent view relies on the accuracy of the previous one, any inaccuracies or imperfections in earlier stages can compound, leading to a degradation in overall image quality. This limitation highlights the need for more sophisticated auto-regressive strategies to address the issue of error propagation and enhance the quality of generated views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Interpolated Conditions and Interpolated Outputs. Interpolated Conditions and Interpolated Outputs are two straightforward approaches to introduce auto-regressive generation into an existing diffusion model. The former method involves interpolating feature embeddings from condition images and poses, while the latter interpolates the final image feature maps produced by the model. Despite" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.893, + 0.435 + ], + "angle": 0, + "content": "SSIM and PSNR metrics showing favorable results for Interpolated Outputs over others in Tab. 5, as illustrated in visual comparisons in Fig. 7 shows that it leads to blurring of the output views and this is corroborated by larger LPIPS distance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.64 + ], + "angle": 0, + "content": "Stochastic Conditioning. We also explore the application of the Stochastic Conditioning Sampler proposed by [69] to the Zero-1-to-3 model. We observe that Stochastic Conditioning fails to deliver the desired auto-regressive generation results; this may be attributed to the specific category on which the diffusion model used by Watson et al. [69] was trained, allowing it to handle plausible condition images more effectively. By contrast, Zero-1-to-3 [34] was trained on a cross-category dataset and designed for zero-shot reconstruction. Additionally, our evaluation data contains out-of-the-distribution data (i.e., ABO [6] and GSO [9])." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.619, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.893, + 0.889 + ], + "angle": 0, + "content": "In this work, we have developed ViewFusion, a novel algorithm that addresses the challenge of multi-view consistency in novel-view synthesis with diffusion models. Our approach circumvents the need for fine-tuning or additional modules by integrating an auto-regressive mechanism that incrementally refines view synthesis, utilizing the entire history of previously generated views. Our proposed diffusion interpolation technique extends the denoising process in pre-trained diffusion models from a single-view setting to a multi-view setting without training requirements. Empirical evidence underscores ViewFusion's capability to produce consistently high-quality views, and we achieve significant steps forward in novel view synthesis and 3D reconstruction applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9877" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[10] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. arXiv preprint arXiv:2303.17015, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[11] George Fahim, Khalid Amin, and Sameh Zarif. Single-view 3d reconstruction: A survey of deep learning methods. Computers & Graphics, 94:164-190, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[12] Kui Fu, Jiansheng Peng, Qiwen He, and Hanxiao Zhang. Single image 3d object reconstruction based on deep learning: A review. Multimedia Tools and Applications, 80:463-498, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Georgia Gkioxari, Justin Johnson, and Jitendra Malik. Mesh r-cnn. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mache approach to learning 3d surface generation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images. arXiv preprint arXiv:2304.06700, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[18] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[23] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[24] Animesh Karnewar, Niloy J Mitra, Andrea Vedaldi, and David Novotny. Holofusion: Towards photo-realistic 3d generative modeling. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[25] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[26] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[27] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[28] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[29] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdfsrn: Learning signed distance 3d object reconstruction from static images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[32] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[33] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[35] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[39] D.G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 1150-1157 vol.2, 1999. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[40] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[42] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. CoRR, abs/1411.1784, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[43] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[44] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc Van Gool, and Sergey Tulyakov." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Autodecoding latent 3d diffusion models. arXiv preprint arXiv:2307.05445, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[46] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[49] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[50] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.425, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[52] Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, and Deqing Sun. Vq3d: Learning a 3d-aware generative model on image-net. arXiv preprint arXiv:2302.06833, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[55] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Neural Information Processing Systems, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[56] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[57] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[58] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[59] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9879" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[60] A. Tewari, O. Fried, J. Thies, V. Sitzmann, S. Lombardi, K. Sunkavalli, R. Martin-Brualla, T. Simon, J. Saragih, M. Nießner, R. Pandey, S. Fanello, G. Wetzstein, J.-Y. Zhu, C. Theobalt, M. Agrawala, E. Shechtman, D. B Goldman, and M. Zollhöfer. State of the art on neural rendering. Computer Graphics Forum, 39(2):701-727, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[61] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B Tenenbaum, Frédo Durand, William T Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. arXiv preprint arXiv:2306.11719, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[62] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[63] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[64] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[66] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[67] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[68] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[69] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[70] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[71] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[73] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints, pages arXiv-2211, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[74] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[75] Xianghui Yang, Guosheng Lin, and Luping Zhou. Single-view 3d mesh reconstruction for seen and unseen categories. IEEE Transactions on Image Processing, pages 1-1, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[76] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[77] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[78] Paul Yoo, Jiaxian Guo, Yutaka Matsuo, and Shixiang Shane Gu. Dreamsparse: Escaping from Plato's cave with 2d frozen diffusion model given sparse views. CoRR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[79] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[80] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[82] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[83] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. In SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[84] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[86] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9880" + } + ] +] \ No newline at end of file diff --git a/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_origin.pdf b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f86a9cd57bea41ab10b0e46cbf77a738f98e53b9 --- /dev/null +++ b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/cae62f84-c32b-4a99-a5f6-784c0ce32aad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:944fd4b19452d703b59de234f455f7fb2b1f08dde30f8b7af89a7b80205960ab +size 8026391 diff --git a/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/full.md b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6778a94ce25a2649205509abc3fd4eb8673681d --- /dev/null +++ b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/full.md @@ -0,0 +1,377 @@ +# ViewFusion: Towards Multi-View Consistency via Interpolated Denoising + +Xianghui Yang $^{1,2}$ , Yan Zuo $^{1}$ , Sameera Ramasinghe $^{1}$ , Loris Bazzani $^{1}$ , Gil Avraham $^{1}$ , Anton van den Hengel $^{1,3}$ , ${}^{1}$ Amazon, ${}^{2}$ The University of Sydney, ${}^{3}$ The University of Adelaide + +# Abstract + +Novel-view synthesis through diffusion models has demonstrated remarkable potential for generating diverse and high-quality images. Yet, the independent process of image generation in these prevailing methods leads to challenges in maintaining multiple view consistency. To address this, we introduce ViewFusion, a novel, training-free algorithm that can be seamlessly integrated into existing pre-trained diffusion models. Our approach adopts an auto-regressive method that implicitly leverages previously generated views as context for next view generation, ensuring robust multi-view consistency during the novel-view generation process. Through a diffusion process that fuses known-view information via interpolated denoising, our framework successfully extends single-view conditioned models to work in multiple-view conditional settings without any additional fine-tuning. Extensive experimental results demonstrate the effectiveness of ViewFusion in generating consistent and detailed novel views. + +# 1. Introduction + +Humans have a remarkable capacity for visualizing unseen perspectives from just a single image view - an intuitive process that remains complex to model. Such an ability is known as Novel View Synthesis (NVS) and necessitates robust geometric priors to accurately infer three-dimensional details from flat imagery; lifting from a two-dimensional projection to a three-dimensional form involves assumptions and knowledge about the nature of the object and space. Recently, significant advancements in NVS have been brought forward by neural networks [14, 15, 30, 41, 60, 64, 65, 72, 75, 76], where novel view generation for downstream reconstruction shows promising potential [34, 69]. + +Specifically, diffusion models [19, 51] and their ability to generate high-quality 2D images have garnered significant attention in the 3D domain, where pre-trained, text-conditioned 2D diffusion models have been re-purposed for 3D applications via distillation [4, 31, 40, 47, 49, 57, 63, + +![](images/a02c626c539bcb87a527e3783ab55f7cc4db15ad6a8ba4f6dad639fdaf6b8a26.jpg) +(a) Inconsistent generation. + +![](images/55e4996694bcb11496ee64ff3477afc4d0537aa530919a5023cae23851235651.jpg) +(b) Consistent generation (ours). +Figure 1. The cause of multi-view inconsistency in diffusion-based novel-view synthesis models. (a) Diffusion models incorporate randomness for diversity and better distribution modeling; this independent generation process produces realistic views under specific instances but may produce different plausible views for various instances, lacking alignment across adjacent views. (b) In contrast, ViewFusion incorporates an auto-regressive process to reduce uncertainty and achieve multi-view consistency, by ensuring a correlated denoising process that ends at the same high-density area, fostering consistency across views. + +68, 73]. Follow-up approaches [34, 69] remove the requirement of text conditioning and instead take an image and target pose as conditions for NVS. However, distillation [63] is still required as the diffusion model cannot produce the multi-view consistent outputs that are appropriate for certain downstream tasks (e.g., optimizing Neural Radiance Fields (NeRFs) [41]). + +Under the single-view setting, maintaining multi-view consistency remains particularly challenging since there may exist several plausible outputs for a novel view that are aligned with the given input image. For diffusion-based approaches which generate novel views in an independent manner [34, 69], this results in synthesized views containing artifacts of multi-view inconsistency (Fig. 1a). Previous work [33, 36, 38, 53, 74, 77] focuses on improving the robustness of the downstream reconstruction to address the inconsistency issue, including feature projection layers in the NeRF [33] or utilising three-dimensional priors to constrain NeRF optimization [36, 77], yet these techniques require training or fine-tuning to align additional modules to the original diffusion models. + +In this work, we address the multi-view inconsistency that arises during the process of view synthesis. Rather than + +independently synthesizing views conditioned only on the initial reference image, we develop a novel approach where each subsequently generated view is also conditioned on the entire set of previously generated views. Specifically, our method incorporates an auto-regressive process into the diffusion process to model the joint distribution of views, guiding our novel-view synthesis by maintaining the denoising direction towards the same high density area of already generated views (Fig. 1b). + +Our framework, named ViewFusion, relaxes the single-view conditioning requirement of typical diffusion models through an interpolated denoising process. ViewFusion offers several additional advantages: 1) it can utilize all available views as guidance, thereby enhancing the quality of generated images by incorporating more information; 2) it does not require any additional fine-tuning, effortlessly converting pre-trained single-view conditioned diffusion models into multi-view conditioned diffusion models; 3) it provides greater flexibility in setting adaptive weights for condition images based on their relative view distance to the target view. + +The contributions of this paper are the following: + +- We propose a training-free algorithm which can be directly applied to pre-trained diffusion models to improve multi-view consistency of synthesized views and supports multiple conditional inputs. +- Our method utilizes a novel, auto-regressive approach which we call Interpolated Denoising, that implicitly addresses key limitations of previous auto-regressive approaches for view synthesis. +- Extensive empirical analysis on ABO [6] and GSO [9] show that our method is able to achieve better 3D consistency in image generation, leading to significant improvements in novel view synthesis and 3D reconstruction of shapes under single-view and multi-view image settings over other baseline methods. + +# 2. Related Work + +# 2.1. 3D-adapted Diffusion Models + +Diffusion models have excelled in image generation using conditional inputs [20, 46, 50, 84] and given this success in the 2D domain, recent works have tried to extend diffusion models to 3D content generation [1, 3, 5, 10, 16, 18, 22, 24, 25, 27, 37, 43-45, 66, 82, 83] - although the scarcity of 3D data presents a significant challenge to directly train these diffusion models. Nonetheless, pioneer works such as DreamFusion [47] and Score Jacobian Chaining [63] leverage pre-trained text-conditioned diffusion models to craft 3D models via distillation. Follow-up approaches [4, 31, 57, 68] improve this distillation in terms of speed, resolution and shape quality. Approaches such as [40, 49, 57, 73] extend upon this to support image condi + +tions through the use of captions with limited success due to the non-trivial nature of textual inversion [13]. + +# 2.2. Novel View Synthesis Diffusion Models + +Another line of research [2, 8, 17, 28, 35, 56, 58, 61, 62, 69, 71, 78, 80, 86] directly applies 2D diffusion models to generate multi-view images for shape reconstruction. To circumvent the weakness of text-conditioned diffusion models, novel-view synthesis diffusion models [34, 69] have also been explored, which take an image and target pose as conditions to generate novel views. However, for these approaches, recovering a 3D consistent shape is still a key challenge. To mitigate 3D inconsistency, Liu et al. [33] suggests training a Neural Radiance Field (NeRF) with feature projection layers. Concurrently, other works [36, 38, 70, 74, 77] add modules to original diffusion models for multi-view consistency, including epipolar attention [77], synchronized multi-view noise predictor [36] and cross-view attention [38, 70]; although these methods require fine-tuning an already pre-trained model. We adopt a different paradigm, instead of extending a single-view diffusion model with additional trainable models that incorporate multi-view conditions, our training-free method enables pre-trained diffusion models to incorporate previously generated views via the denoising step and holistically extends these models into multi-view settings. + +# 2.3. Other Single-view Reconstruction Methods + +Before the prosperity of generative models used in 3D reconstruction, many works [11, 12, 14, 15, 26, 29, 30, 59, 64, 75] reconstructed 3D shapes from single-view images using regression [14, 15, 29, 64, 75] or retrieval [59], both of which face difficulties in generalizing to real data or new categories. Methods based on Neural Radiance Fields (NeRFs) [41] have found success in novel-view synthesis, but these approaches typically depend on densely captured images with accurately calibrated camera positions. Currently, several studies are investigating the adaptation of NeRF to single-view settings [21, 32, 52, 79]; although, reconstructing arbitrary objects from single-view images is still a challenging problem. + +# 3. Method + +# 3.1. Denoising Diffusion Probabilistic Models + +Denoising diffusion probabilistic models (DDPM) [19, 54] are a class of generative models that model the real data distribution $q(x_0)$ with a tractable model distribution $p_{\theta}(x_0)$ by learning to iteratively denoise samples. It learns a probability model $p_{\theta}(\mathbf{x}_0) = \int p_{\theta}(\mathbf{x}_{0:T}) d\mathbf{x}_{1:T}$ to convert unstructured noise $\mathbf{x}_T$ to real samples $\mathbf{x}_0$ in the form of a Markov chain, with Gaussian transitions. The Gaussian transition is + +![](images/528efee45774322d21d034cc41d8286bfdb4fd5f9e973e81c37b10e572449e0e.jpg) + +![](images/09a29e0c94767c8285e2f8fb88c6e6a41161cdf4c3bf0a2c04e7980e011aeb2f.jpg) + +![](images/f65e89dee7baac846a779ab1e994dee7b4eabdd8ebb56394b0c62a31b118d7d9.jpg) +中 +Figure 2. Illustration of the Auto-Regressive Generation Process. In our approach, we extend a pre-trained diffusion model from single-stage to multi-stage generation and we maintain a view set that contains all generated views. For each stage, we construct $N$ reverse diffusion processes and sharing a common starting noise. At each time step within this generation stage, the diffusion model predicts $N$ noises individually. These $N$ noises are then subjected to weighted interpolation through the Noise Interpolation Module, concluding the denoising step with the a shared interpolated noise for subsequent denoising steps. + +defined as: + +$$ +q \left(\mathbf {x} _ {T} \mid \mathbf {x} _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \prod_ {t = 1} ^ {T} \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{t}, t \in \{1, \dots, T\}$ are the variance schedule parameter and timestep in the denoising process respectively. The reverse denoising process starts from a noise sampled from a Gaussian distribution $q(\mathbf{x}_T) = \mathcal{N}(\mathbf{0}, \mathbf{I})$ and is constructed as: + +$$ +p _ {\theta} \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {T}\right) = \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) = \prod_ {t = 1} ^ {T} \mathcal {N} \left(\mathbf {x} _ {t - 1}; \mu_ {\theta} \left(\mathbf {x} _ {t}, t\right), \sigma_ {t} ^ {2} \mathbf {I}\right), \tag {2} +$$ + +where the variance $\sigma_t^2$ is a time-dependent constant [19], and $\mu_{\theta}(\mathbf{x}_t,t)$ is the mean from the learned noise predictor $\epsilon_{\theta}$ : + +$$ +\mu_ {\theta} (\mathbf {x} _ {t}, t) = \frac {1}{\sqrt {\alpha} _ {t}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (\mathbf {x} _ {t}, t)\right). \quad (3) +$$ + +Here, $\alpha_{t}$ and $\bar{\alpha}_{t}$ are constants derived from $\beta_{t}$ . The objective of noise predictor $\epsilon_{\theta}$ is simplified to: + +$$ +\ell = \mathbb {E} _ {t, \mathbf {x} _ {0}, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \| _ {2} \right], \tag {4} +$$ + +where $\epsilon$ is a random variable sampled from $\mathcal{N}(\mathbf{0},\mathbf{I})$ [19]. + +# 3.2. Pose-Conditional Diffusion Models + +Similar to other generative models [42, 55], diffusion models inherently possess the capability to model conditional + +distributions of the form $p_{\theta}(x_{t - 1}|x_t,y)$ where $y$ is the condition. We employ a conditional denoising autoencoder, denoted as $\epsilon_{\theta}(\mathbf{x}_t,t,y)$ which enables controlling the synthesis process through a variety of input modalities, including textual descriptions [50], semantic maps [20, 46], or other image-to-image translation tasks [20]. In the following, we present a range of approaches to novel-view synthesis, exploring how various works, including our own, approach the concept of a single reverse diffusion step. Through this comparison, we clarify and establish the underlying relationships between these different methodologies. The notation will follow that bottom subscript $(\cdot)_t$ indicates the diffusion step and upper subscript $(\cdot)^i$ relates to the view index. Subsequently, the $i$ -th condition image and its relative pose to the target view are defined as $\mathbf{y}^i$ and $\pi^i$ , respectively, and the noisy image to be denoised at timestep $t$ is defined as $\mathbf{x}_t$ . + +Direct condition was applied by Zero 1-to-3 [34] to the reverse process when given a single input image and target pose $\mathbf{y}^1$ , $\pi^1$ : + +$$ +p \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \mathbf {y} ^ {1}, \pi^ {1}\right). \tag {5} +$$ + +Stochastic conditioning was formulated by [69] which can leverage multiple views sampled from a collection of views $p_{\mathbf{y},\pi}(\mathcal{Y},\pi)$ : + +$$ +p \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \mathbf {y} ^ {i}, \pi^ {i}\right), \left\{\mathbf {y} ^ {i}, \pi^ {i} \right\} \sim p _ {\mathbf {y}, \pi} (\mathcal {Y}, \pi), \tag {6} +$$ + +where the sampling of image and pose happens at each diffusion step $t$ . + +Joint output distribution was shown in SyncDreamer[36] which learns a joint distribution of many views given an image condition $y^{1}$ : + +$$ +p \left(\mathbf {x} _ {t - 1} ^ {1: N} \mid \mathbf {x} _ {t} ^ {1: N}, \mathbf {y} ^ {1}, e ^ {1}\right), \tag {7} +$$ + +where $N$ is the number of generated novel views and $e$ is the elevation condition (partial pose information). We note that in this formulation the target poses are not fully specified as part of the condition allowing for diverse pose generation of outputs. + +Auto-regressive distribution is an auto-regressive distribution setting which can generate an arbitrary number of views given a single or multiple condition images and poses contained in the set of $\mathbf{y}^{1:N - 1}$ , $\pi^{1:N - 1}$ : + +$$ +p \left(\mathbf {x} _ {t - 1} ^ {N} \mid \mathbf {x} _ {t} ^ {N}, \mathbf {y} ^ {1: N - 1}, \pi^ {1: N - 1}\right). \tag {8} +$$ + +Our approach falls in the auto-regressive category and for the remainder of this section we detail the implementation to achieve this sampling strategy. + +# 3.3. Interpolated Denoising + +The standard DDPM model has been adapted for novel-view image synthesis by using an image and target pose (i.e., rotation and translation offsets) as conditional inputs [69]. Following training on a large-scale dataset, this approach has demonstrated the capability for zero-shot reconstruction [34]. To address the challenge of maintaining multi-view consistency, we employ an auto-regressive approach for generating sequential frames (See Fig. 2). Instead of independently producing each frame from just the input images – a process prone to significant variations between adjacent images – we integrate an auto-regressive algorithm into the diffusion process. This integration enables us to model a conditional joint distribution, ensuring smoother and more consistent transitions between frames. + +To guide the synthesis of novel views using images under different views, we design an interpolated denoising process. For the purpose of this derivation, we assume access to an image set containing $N - 1$ images denoted as $\{\mathbf{y}^1,\dots,\mathbf{y}^{N - 1}\}$ . We want to model the distribution of the $N$ -th view image conditioned on these $N - 1$ views $q(\mathbf{x}_{1:T}^N|\mathbf{y}^{1:N - 1})$ , where the relative pose offsets $\pi^i,i\in$ $\{1,N - 1\}$ between the condition images $\{\mathbf{y}^1,\dots,\mathbf{y}^{N - 1}\}$ and target image $\mathbf{x}_0^N$ are omitted for simplicity. The forward process of the multi-view conditioned diffusion model + +is a direct extension of the vanilla DDPM in Eq. 1, where noises are added to every view independently by + +$$ +q \left(\mathbf {x} _ {1: T} ^ {N} \mid \mathbf {y} ^ {1: N}\right) = \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} ^ {N} \mid \mathbf {x} _ {t - 1} ^ {N}, \mathbf {y} ^ {1: N}\right) \tag {9} +$$ + +where $q(\mathbf{x}_t^N | \mathbf{x}_{t-1}^N, \mathbf{y}^{1:N}) = \mathcal{N}(\mathbf{x}_t^N; \sqrt{1 - \beta_t} \mathbf{x}_{t-1}^N, \beta_t \mathbf{I})$ . The initial is defined as $\mathbf{x}_0^N := \mathbf{y}^N$ . Similarly, following Eq. 2, the log reverse process is constructed as + +$$ +\begin{array}{l} \log p _ {\theta} \left(\mathbf {x} _ {0} ^ {N} \mid \mathbf {x} _ {T} ^ {N}, \mathbf {y} ^ {1: N - 1}\right) = \sum_ {t = 1} ^ {T} \log p _ {\theta} \left(\mathbf {x} _ {t - 1} ^ {N} \mid \mathbf {x} _ {t} ^ {N}, \mathbf {y} ^ {1: N - 1}\right) \\ \approx_ {(1)} ^ {\sum} \log \prod_ {n = 1} ^ {N - 1} p _ {\theta} \left(\mathbf {x} _ {t - 1} ^ {N} \mid \mathbf {x} _ {t} ^ {N}, \mathbf {y} ^ {n}\right) \\ = \sum_ {t = 1} ^ {T} \sum_ {n = 1} ^ {N - 1} \log \mathcal {N} \left(\mathbf {x} _ {t - 1} ^ {N}; \mu_ {\theta} ^ {n} \left(\mathbf {x} _ {t} ^ {N}, \mathbf {y} ^ {n}, t\right), \sigma_ {t} ^ {2} \mathbf {I}\right) \\ = \sum_ {t = 1} ^ {T} \mathcal {N} \left(\mathbf {x} _ {t - 1} ^ {N}; \bar {\mu} _ {\theta} \left(\mathbf {x} _ {t} ^ {N}, \mathbf {y} ^ {1: N - 1}, t\right), \bar {\sigma} _ {t} ^ {2} \mathbf {I}\right). \tag {10} \\ \end{array} +$$ + +Where $\bar{\mu}_{\theta}, \bar{\sigma}_{t}^{2}$ are taken as the mean and variance of the summation of $N - 1$ log-normal distributions. A note on subscript (1) in Eq.10; to avoid cluttering the derivation, we assume $N - 1$ independent inferences of the same random variable $\mathbf{x}_{t-1}^{N}$ using a different $\mathbf{y}^{n}$ that results in $N - 1$ independent normal distributions, which would require an additional subscript that we omitted for clarity. + +# 3.4. Single and Multi-view Denoising + +In practice, however, we may not have all $N - 1$ views but a single view or a handful of views. For the reminder of this section, we treat an estimated view as $\mathbf{x}_0^n$ , to be the $n$ -th view $\mathbf{y}^n$ after a full reverse diffusion process. We use $\bar{\mu}_{\theta}(\mathbf{x}_t,\mathbf{y}^{1:N - 1},t)$ as the weighted average of $\mu_{\theta}^{n}(\mathbf{x}_{t},\mathbf{y}^{n},t)$ . For computing $\bar{\mu}_{\theta}$ using both given views and estimated views we adopt an approach where different views contribute differently to the target view, and we assign the weight $\omega_{n}$ for the $n$ -th view in practice while satisfying the constraint $\sum_{n = 1}^{N - 1}w_n = 1$ . The Noise Interpolation Module in Fig. 2 is modeled as: + +$$ +\begin{array}{l} \bar {\mu} _ {\theta} \left(\mathbf {x} _ {t}, \mathbf {y} ^ {1: N - 1}, t\right) = \sum_ {n = 1} ^ {N - 1} \omega_ {n} \mu_ {\theta} ^ {n} \left(\mathbf {x} _ {t}, \mathbf {y} ^ {n}, t\right) \\ = \sum_ {n = 1} ^ {N - 1} \omega_ {n} \frac {1}{\sqrt {\alpha} _ {t}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(\mathbf {x} _ {t}, \mathbf {y} ^ {n}, t\right)\right) \\ = \frac {1}{\sqrt {\alpha} _ {t}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \sum_ {n = 1} ^ {N} \omega_ {n} \epsilon_ {\theta} \left(\mathbf {x} _ {t}, \mathbf {y} ^ {n}, t\right)\right). \tag {11} \\ \end{array} +$$ + +In our approach, as the full view set is not given to us, we approximate this process by an auto-regressive way and + +grow the condition set during the generation. We define the weight parameter $\omega_{n}$ based on the angle offset, i.e., azimuth $(\Delta_{a}^{n})$ , elevation $(\Delta_{e}^{n})$ , and distance $(\Delta_{d}^{n})$ , between the target view and the $n - th$ condition view. The core idea is to assign higher importance to near-view images during the denoising process while ensuring that the weight for the initial condition image does not diminish too rapidly, even when the target view is positioned at a nearly opposite angle. We use an exponential decay weight function for the initial condition image, defined as $\omega_{n} = e^{-\frac{\Delta^{n}}{\tau_{c}}}$ . Here, $\tau_{c}$ is the temperature parameter that regulates the decay speed, and $\Delta^n$ is the sum of the absolute relative azimuth $(\Delta_{a}^{n})$ , elevation $(\Delta_{e}^{n})$ , and distance $(\Delta_{d}^{n})$ between the target and condition poses. We calculate $\Delta^n$ as $\Delta^n = |\Delta_a^n| / \pi +|\Delta_e^n| / \pi +|\Delta_d^n|$ . + +For the weights of the remaining images denoted as $\{x_0^2,\dots,x_0^N\}$ , all generated from the initial condition image $y^{1}:= x_{0}^{1}$ , we use a softmax function to define the weights $\omega_{n}$ : + +$$ +\omega_ {n} = \operatorname {S o f t m a x} \left(\frac {e ^ {- \frac {\Delta^ {n}}{\tau_ {g}}}}{\sum_ {n = 2} ^ {N} e ^ {- \frac {\Delta^ {n}}{\tau_ {g}}}}\right), n = 2, \dots , N \tag {12} +$$ + +Similarly, $\Delta_{n}$ represents the relative pose offset between target view and the $n$ -th generated view, and $\tau_{g}$ represents the temperature parameter for generated views. As an example, in the single-view case, the weights are expressed as follows, + +$$ +\omega_ {n} = \left\{ \begin{array}{l l} e x p \left(- \frac {\Delta^ {n}}{\tau_ {c}}\right), & n = 1 \\ (1 - \omega_ {1}) \operatorname {S o f t m a x} \left(\frac {e ^ {- \frac {\Delta^ {n}}{\tau_ {g}}}}{\sum_ {n = 2} ^ {N} e ^ {- \frac {\Delta^ {n}}{\tau_ {g}}}}\right), & n \neq 1 \end{array} \right. \tag {13} +$$ + +we apply the term $1 - \omega_{1}$ on the generated image weights to ensure the requirement of $\sum_{n = 1}^{N - 1}w_n = 1$ will be met. In practice, Eq.13 is generalised to allow the condition set can be larger than 1, i.e., multi-view generation (see supplementary). + +# 3.5. Step-by-step Generation + +Single image generation. When applying the autoregressive approach to image generation, we have devised a generation trajectory, as illustrated in Fig. 3a. We uniformly sample views along this trajectory in sequence. Each previously generated view image on this trajectory is incorporated into the condition set, providing guidance for the subsequent denoising process via our interpolated denoising method. To determine the number of steps, denoted as $S$ , needed for this trajectory, we use the following formula: + +$$ +S = \max \left(\lceil \frac {\Delta_ {a} ^ {N}}{\delta} \rceil , \lceil \frac {\Delta_ {e} ^ {N}}{\delta} \rceil\right). \tag {14} +$$ + +![](images/76a32f7bd0c47340cfd1ac3d1c3fae945a5e8ae9fd01e5e586c6f6cbd17d840e.jpg) +(a) Single image generation. + +![](images/150217d51a3426f2d2d7b72a52d4648341a3bb8a84f1b5c79978fb099df608be.jpg) +(b) Spin video generation. +Figure 3. Illustration of Step-by-step Generation. (a) we uniformly sample views along this trajectory in sequence to generate a novel-view image; (b) we sample views from nearest to furthest views according to to view distance to generate a $360^{\circ}$ spin video. + +Here, we set the maximum offset per step $\delta$ to determine the step count $S$ , also based on the target view offsets $\Delta_{a}^{N}$ and $\Delta_{e}^{N}$ . We then proceed to sample the $n$ -th view using the following equation: + +$$ +\left(\Delta_ {a} ^ {n}, \Delta_ {e} ^ {n}, \Delta_ {d} ^ {n}\right) = \left(\frac {\Delta_ {a} ^ {N}}{S} * n, \frac {\Delta_ {e} ^ {N}}{S} * n, \frac {\Delta_ {d} ^ {N}}{S} * n\right) \tag {15} +$$ + +Spin videos generation. In contrast to generating a single target image, the process of spin video generation begins from an initial image and concludes at the same position. To achieve this, we need to modify the generation order to leverage the broad range of rotation images, rather than simply following the rotation degree range of $[0^{\circ}, 360^{\circ}]$ in sequence. This is because, at $\Delta_{a} = \pi$ , the view is opposite to the conditioning view, marking the end of the generation process. To establish the generation order for spin video generation, we introduce the minimum azimuth offset, denoted as $\delta$ , and employ a skip trajectory with the following order: $\{\delta, -\delta, 2\delta, -2\delta\dots, N\delta\}$ , shown in Fig. 3b. For simplicity, we only consider rotation along the azimuth dimension in this context. + +# 4. Experiments + +Datasets. We evaluate our method and compare to baselines on the ABO [6] and GSO [9] datasets. These datasets are out-of-the-distribution as all baselines are trained on the Objaverse [7]. We also provide qualitative results on real images to showcase performance of our method on in-the-wild images in the supplementary. For additional results, please refer to the videos contained in the supplementary. + +Metrics. We assess our novel-view synthesis on three main criteria: + +1. Image Quality: LPIPS [85], PSNR, and SSIM [67] metrics to help gauge the similarity between synthesized and ground-truth views. + +
DatasetMethodFree RenderingsSyncDreamer Renderings
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
ABOZero123 [34]0.879621.330.09610.782218.270.1999
SyncDre. [36]0.771213.430.21820.803119.070.1816
Ours0.884821.430.09230.798318.750.1985
GSOZero123 [34]0.871020.330.10290.792518.060.1714
SyncDre. [36]0.802314.420.18330.802418.200.1647
Ours0.882020.730.09580.807618.400.1703
+ +Table 1. Quantitative results on ABO and GSO datasets with arbitrary (left) and discrete (right) rotation and translation. Free renderings are a set of arbitrary rotation and translation as target generation view, while SyncDreamer renderings are a fixed set of 16 views with discrete azimuth, fixed elevation and distance, i.e., azimuth $\in$ $\{0^{\circ},22.5^{\circ},45^{\circ},\dots,315^{\circ},337.5^{\circ}\}$ , elevation $= 30^{\circ}$ . Note that SyncDreamer [36] cannot generate images under arbitrary views apart from the predefined 16 camera positions. + +2. Multi-View Consistency: Using SIFT [39], LPIPS [85] and CLIP [48], we measure the uniformity of images across various perspectives. +3. 3D Reconstruction: Chamfer distances and F-score between ground-truth and reconstructed shapes determine geometrical consistency. + +# 4.1. Novel-view synthesis + +In Tab. 1, we show quantitative results for novel-view synthesis under arbitrary and fixed-view settings. The fixed-view setting uses the rendering set of [36] and ensures a fair comparison to [36] which is limited to this fixed-view generation setting. As shown in Tab. 1, our method can produce comparable results to [36], without any fine-tuning on the rendering set of [36]. Under the arbitrary-view setting, we sample the nearest view from the rendering set of [36] to the designated target view for generation; [36] underperforms both [34] and our approach. Overall, regardless of the evaluation setting, our method consistently outperforms [34] and even outperforms [36] on the GSO dataset under the favourable fixed-view setting for [36]. + +The task of novel view synthesis serves as a precursor for enabling more significant downstream applications, such as 3D reconstruction, by generating requisite input views. Thus, in addition to image quality, thorough evaluation needs to encompass the multi-view consistency of the synthesized perspectives. This additional criterion ensures that the generated imagery not only appears visually compelling but also aligns geometrically across different viewpoints. In the following sections, we evaluate against baselines on 3D consistency and show that the significant improvements our approach offers. + +Multi-view Consistency. For measuring multi-view consistency of synthesized views, quantitative results are shown in Tab. 2, while qualitative comparisons are shown in Figs. 4 and 5. Here, we can see when compared to the baselines of [34] and [36], our approach excels in generating images + +
DatasetMethodViewSIFT↑LPIPS↓CLIP↑
ABOZero123 [34]13.380.17820.9604
SyncDreamer [36]1612.360.18950.9584
Ours13.510.16020.9664
Zero123 [34]17.030.12310.9725
Ours3618.010.09660.9812
GSOZero123 [34]12.580.14110.9482
SyncDreamer [36]1613.240.13150.9532
Ours13.830.11870.9601
Zero123 [34]15.200.10560.9592
Ours3617.950.06760.9773
+ +Table 2. Quantitative results for multi-view consistency. We report the SIFT matching point number, LPIPS and CLIP similarity between adjacent frames to evaluate the multi-view consistency. Note that SyncDreamer can only generate 16 view images for a spin video due to the constraints imposed by its training. + +
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 [34] (1 view)0.882021.510.0945
Ours (1 view)0.887021.610.0904
Ours (2 views)0.891321.920.0887
Ours (3 views)0.899522.740.0815
GSOZero123 [34] (1 view)0.872120.420.1017
Ours (1 view)0.883020.870.0948
Ours (2 views)0.890121.250.0893
Ours (3 views)0.897921.950.0792
+ +Table 3. Quantitative results for multi-conditioned generation. Our approach outperforms the Zero-1-to-3 baseline [34] by extending its single-view reconstruction framework to a multi-view reconstruction framework. Note that 3 condition views are removed from test set, and thus the results are slightly different to Tab. 1. + +that are both semantically consistent with the input image and maintains multi-view consistency in terms of colors and geometry under arbitrary-view settings. + +# 4.2. Multi-view conditional setting for NVS. + +Additionally, our approach allows for the extension of single-view conditioned models into multi-view conditioned models which can take multiple conditional images as input. The results presented in Tab. 3 showcase the advantages our method offers, as novel-view synthesis quality improves with an increasing number of conditional input views. This is a significant improvement over the Zero-1-to-3 baseline and demonstrates the efficacy of our proposed method in the multi-view setting. + +# 4.3. 3D Reconstruction + +For shape reconstruction, we present our results both quantitatively (Tab. 4) and qualitatively (Fig. 6). We compare against baselines which synthesize novel views as well as direct image-to-shape approaches [23, 44]. For the former approach, instead of using distillation [63], we train NeuS [65] on synthesized images to recover a shape. + +![](images/0fa386518989aafa33eda739ac2cfba214f2b67bda5dbad0a554e72140b8cd47.jpg) +Figure 4. Qualitative results for $360^{\circ}$ Spin Video Generation. Note the additional consistency in generated views our approach offers over the competing baselines shown in the bounding boxes. + +![](images/54e7d11b2170e71469b442caff833d81bbe54f0a89fa1ee14e4e40f8c7ac3aa5.jpg) +Figure 5. Qualitative comparison for Motion Smoothness. We visualize the output videos using space-time Y-t slices through frames of the generated spin video (along the scanline shown in the condition). + +Single-view Reconstruction. We first compare our proposed method with several baselines under the single-view reconstruction setting, including Point-E [44] and Shap-E [23], Zero-1-to-3 [34], One-2-3-45 [33] and SyncDreamer [36]. Qualitative results are shown in Fig. 6; Point-E and Shap-E tend to generate incomplete shapes due to the single-view setting. Multi-view alignment methods such as Syncdreamer [36] and One-2-3-45 [33] capture the general geometry but tend to lose fine details. In comparison, our proposed method achieves the highest reconstruction quality amongst all approaches, where we can generate smooth surfaces and capture detailed geometry with precision. + +Multi-view Reconstruction Given our approach can also synthesis novel views under the multi-view setting, we also show results where we use 3 conditional input views to synthesize 36 views for shape reconstruction. Compared with existing multi-view reconstruction frameworks such as + +
Cond. viewsMethodGen. viewsABOGSO
CD↓F-score↑CD↓F-score↑
1Point-E [44]N/A0.04280.71440.06720.6340
Shap-E [23]N/A0.04660.73640.03840.7313
One-2-3-45 [33]320.04190.66650.04080.6490
SyncDreamer [36]160.01600.81870.02290.7767
Zero123 [34]160.01470.82260.02060.8045
Zero123 [34]360.01390.82470.02070.8078
Ours160.01330.84230.01770.8274
Ours360.01260.84720.01640.8436
3MonoSDF [81]N/A0.10200.39630.08300.4581
Ours360.01150.85870.01240.8628
+ +Table 4. Quantitative results on reconstructing 3D Shapes using the generated images. + +
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 (no interp.)0.879621.330.0961
Standard auto-regression0.801016.770.1854
Interpolated conditions0.724313.260.3770
Interpolated outputs0.8925*21.95*0.1246
Stochastic conditioning [69]0.869920.640.1106
Interpolated denoising0.884821.430.0923
GSOZero123 (no interp.)0.871020.330.1029
Standard auto-regression0.809416.300.1801
Interpolated conditions0.766114.380.3427
Interpolated outputs0.879920.440.1659
Stochastic conditioning [69]0.865819.970.1119
Interpolated denoising0.882020.730.0958
+ +Table 5. Quantitative comparison between our method to different auto-regressive variants. Note that the Interpolated outputs variant achieves the highest SSIM and PSNR values on the ABO dataset, but generates blurred images. + +MonoSDF [81], our approach captures more details and generates smoother surfaces (note that [81] also relies on additional depth and normal estimations for its reconstruction whereas ours does not). + +# 4.4. Ablations + +Here, we study the effectiveness of our proposed interpolated denoising process by exploring various autoregressive generation variants (in Tab. 5 and Fig. 7). Specifically, we investigate four variants of denoising: stan- + +![](images/3e0bf45faf0c6ab9f136010eaf1f0f6dbada123e15c66b9e53d9eda58ef68c08.jpg) +Figure 6. Qualitative results for single-view (1 condition image) and multi-view (3 condition images) 3D Reconstruction. Our method offers the most consistent results across a variety of different reconstructed shapes. + +![](images/494e0ead145369685abfd915e3e2c9b9b2cb47fc3480f78c88f58dffd3336812.jpg) +Figure 7. Qualitative comparison for different auto-regressive generations. + +dard auto-regression, interpolated conditions, interpolated outputs, and stochastic conditioning which was proposed in [69]. + +Standard Auto-regression. One initial approach to auto-regression involves using the last generated view as the subsequent conditioning. However, this method exhibits bad generation quality, due to the accumulation of errors during the sequential generation process. As each subsequent view relies on the accuracy of the previous one, any inaccuracies or imperfections in earlier stages can compound, leading to a degradation in overall image quality. This limitation highlights the need for more sophisticated auto-regressive strategies to address the issue of error propagation and enhance the quality of generated views. + +Interpolated Conditions and Interpolated Outputs. Interpolated Conditions and Interpolated Outputs are two straightforward approaches to introduce auto-regressive generation into an existing diffusion model. The former method involves interpolating feature embeddings from condition images and poses, while the latter interpolates the final image feature maps produced by the model. Despite + +SSIM and PSNR metrics showing favorable results for Interpolated Outputs over others in Tab. 5, as illustrated in visual comparisons in Fig. 7 shows that it leads to blurring of the output views and this is corroborated by larger LPIPS distance. + +Stochastic Conditioning. We also explore the application of the Stochastic Conditioning Sampler proposed by [69] to the Zero-1-to-3 model. We observe that Stochastic Conditioning fails to deliver the desired auto-regressive generation results; this may be attributed to the specific category on which the diffusion model used by Watson et al. [69] was trained, allowing it to handle plausible condition images more effectively. By contrast, Zero-1-to-3 [34] was trained on a cross-category dataset and designed for zero-shot reconstruction. Additionally, our evaluation data contains out-of-the-distribution data (i.e., ABO [6] and GSO [9]). + +# 5. Conclusion + +In this work, we have developed ViewFusion, a novel algorithm that addresses the challenge of multi-view consistency in novel-view synthesis with diffusion models. Our approach circumvents the need for fine-tuning or additional modules by integrating an auto-regressive mechanism that incrementally refines view synthesis, utilizing the entire history of previously generated views. Our proposed diffusion interpolation technique extends the denoising process in pre-trained diffusion models from a single-view setting to a multi-view setting without training requirements. Empirical evidence underscores ViewFusion's capability to produce consistently high-quality views, and we achieve significant steps forward in novel view synthesis and 3D reconstruction applications. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In CVPR, 2023. 2 +[2] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 2 +[3] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 2 +[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873, 2023. 1, 2 +[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2 +[6] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 2, 5, 8 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 5 +[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2 +[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, 2022. 2, 5, 8 +[10] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. arXiv preprint arXiv:2303.17015, 2023. 2 +[11] George Fahim, Khalid Amin, and Sameh Zarif. Single-view 3d reconstruction: A survey of deep learning methods. Computers & Graphics, 94:164-190, 2021. 2 +[12] Kui Fu, Jiansheng Peng, Qiwen He, and Hanxiao Zhang. Single image 3d object reconstruction based on deep learning: A review. Multimedia Tools and Applications, 80:463-498, 2021. 2 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 + +[14] Georgia Gkioxari, Justin Johnson, and Jitendra Malik. Mesh r-cnn. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2 +[15] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mache approach to learning 3d surface generation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2 +[16] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images. arXiv preprint arXiv:2304.06700, 2023. 2 +[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In ICML, 2023. 2 +[18] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3 +[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 2, 3 +[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2 +[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2 +[23] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 6, 7 +[24] Animesh Karnewar, Niloy J Mitra, Andrea Vedaldi, and David Novotny. Holofusion: Towards photo-realistic 3d generative modeling. In ICCV, 2023. 2 +[25] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In CVPR, 2023. 2 +[26] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2 +[27] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In CVPR, 2023. 2 +[28] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. In CVPR, 2022. 2 +[29] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2 +[30] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdfsrn: Learning signed distance 3d object reconstruction from static images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2 + +[31] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2 +[32] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2 +[33] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 1, 2, 7 +[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 1, 2, 3, 4, 6, 7, 8 +[35] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2 +[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 2, 4, 6, 7 +[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. In ICLR, 2023. 2 +[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 1, 2 +[39] D.G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 1150-1157 vol.2, 1999. 6 +[40] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 1, 2 +[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[42] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. CoRR, abs/1411.1784, 2014. 3 +[43] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2 +[44] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 6, 7 +[45] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc Van Gool, and Sergey Tulyakov. + +Autodecoding latent 3d diffusion models. arXiv preprint arXiv:2307.05445, 2023. 2 +[46] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3 +[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[49] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 1, 2 +[50] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2, 3 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1 +[52] Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, and Deqing Sun. Vq3d: Learning a 3d-aware generative model on image-net. arXiv preprint arXiv:2302.06833, 2023. 2 +[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2023. 1 +[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2 +[55] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Neural Information Processing Systems, 2015. 3 +[56] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2 +[57] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 1, 2 +[58] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2 +[59] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2 + +[60] A. Tewari, O. Fried, J. Thies, V. Sitzmann, S. Lombardi, K. Sunkavalli, R. Martin-Brualla, T. Simon, J. Saragih, M. Nießner, R. Pandey, S. Fanello, G. Wetzstein, J.-Y. Zhu, C. Theobalt, M. Agrawala, E. Shechtman, D. B Goldman, and M. Zollhöfer. State of the art on neural rendering. Computer Graphics Forum, 39(2):701-727, 2020. 1 +[61] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B Tenenbaum, Frédo Durand, William T Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. arXiv preprint arXiv:2306.11719, 2023. 2 +[62] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2 +[63] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 1, 2, 6 +[64] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 1, 2 +[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 1, 6 +[66] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2 +[67] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[68] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2 +[69] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 1, 2, 3, 4, 7, 8 +[70] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 2 +[71] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2 +[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1 +[73] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints, pages arXiv-2211, 2022. 1, 2 + +[74] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion, 2023. 1, 2 +[75] Xianghui Yang, Guosheng Lin, and Luping Zhou. Single-view 3d mesh reconstruction for seen and unseen categories. IEEE Transactions on Image Processing, pages 1-1, 2023. 1, 2 +[76] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 1 +[77] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models, 2023. 1, 2 +[78] Paul Yoo, Jiaxian Guo, Yutaka Matsuo, and Shixiang Shane Gu. Dreamsparse: Escaping from Plato's cave with 2d frozen diffusion model given sparse views. CoRR, 2023. 2 +[79] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[80] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2 +[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2022. 7 +[82] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 2 +[83] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. In SIGGRAPH, 2023. 2 +[84] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2 +[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6 +[86] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2 \ No newline at end of file diff --git a/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/images.zip b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c68f0841fd20c6d3e1803a43deda72b0212a02cd --- /dev/null +++ b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16aa9cdb65463e4366244b511dc5fd6d19d009d01fdfec7d4ba76c251ea8450f +size 623850 diff --git a/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/layout.json b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c4bf5ad13ff5764ac89b3f86e051c19ff13d3e4d --- /dev/null +++ b/2024/ViewFusion_ Towards Multi-View Consistency via Interpolated Denoising/layout.json @@ -0,0 +1,10419 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "text", + "content": "ViewFusion: Towards Multi-View Consistency via Interpolated Denoising" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": "Xianghui Yang" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": ", Yan Zuo" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": ", Sameera Ramasinghe" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": ", Loris Bazzani" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": ", Gil Avraham" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": ", Anton van den Hengel" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": "Amazon, " + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": "The University of Sydney, " + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 53, + 142, + 577, + 172 + ], + "type": "text", + "content": "The University of Adelaide" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 198, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 198, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 224, + 290, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 224, + 290, + 439 + ], + "spans": [ + { + "bbox": [ + 45, + 224, + 290, + 439 + ], + "type": "text", + "content": "Novel-view synthesis through diffusion models has demonstrated remarkable potential for generating diverse and high-quality images. Yet, the independent process of image generation in these prevailing methods leads to challenges in maintaining multiple view consistency. To address this, we introduce ViewFusion, a novel, training-free algorithm that can be seamlessly integrated into existing pre-trained diffusion models. Our approach adopts an auto-regressive method that implicitly leverages previously generated views as context for next view generation, ensuring robust multi-view consistency during the novel-view generation process. Through a diffusion process that fuses known-view information via interpolated denoising, our framework successfully extends single-view conditioned models to work in multiple-view conditional settings without any additional fine-tuning. Extensive experimental results demonstrate the effectiveness of ViewFusion in generating consistent and detailed novel views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 481, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 635 + ], + "type": "text", + "content": "Humans have a remarkable capacity for visualizing unseen perspectives from just a single image view - an intuitive process that remains complex to model. Such an ability is known as Novel View Synthesis (NVS) and necessitates robust geometric priors to accurately infer three-dimensional details from flat imagery; lifting from a two-dimensional projection to a three-dimensional form involves assumptions and knowledge about the nature of the object and space. Recently, significant advancements in NVS have been brought forward by neural networks [14, 15, 30, 41, 60, 64, 65, 72, 75, 76], where novel view generation for downstream reconstruction shows promising potential [34, 69]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "content": "Specifically, diffusion models [19, 51] and their ability to generate high-quality 2D images have garnered significant attention in the 3D domain, where pre-trained, text-conditioned 2D diffusion models have been re-purposed for 3D applications via distillation [4, 31, 40, 47, 49, 57, 63," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 196, + 425, + 299 + ], + "blocks": [ + { + "bbox": [ + 309, + 196, + 425, + 299 + ], + "lines": [ + { + "bbox": [ + 309, + 196, + 425, + 299 + ], + "spans": [ + { + "bbox": [ + 309, + 196, + 425, + 299 + ], + "type": "image", + "image_path": "a02c626c539bcb87a527e3783ab55f7cc4db15ad6a8ba4f6dad639fdaf6b8a26.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 302, + 410, + 312 + ], + "lines": [ + { + "bbox": [ + 321, + 302, + 410, + 312 + ], + "spans": [ + { + "bbox": [ + 321, + 302, + 410, + 312 + ], + "type": "text", + "content": "(a) Inconsistent generation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 198, + 544, + 300 + ], + "blocks": [ + { + "bbox": [ + 426, + 198, + 544, + 300 + ], + "lines": [ + { + "bbox": [ + 426, + 198, + 544, + 300 + ], + "spans": [ + { + "bbox": [ + 426, + 198, + 544, + 300 + ], + "type": "image", + "image_path": "55e4996694bcb11496ee64ff3477afc4d0537aa530919a5023cae23851235651.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 302, + 536, + 312 + ], + "lines": [ + { + "bbox": [ + 431, + 302, + 536, + 312 + ], + "spans": [ + { + "bbox": [ + 431, + 302, + 536, + 312 + ], + "type": "text", + "content": "(b) Consistent generation (ours)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 316, + 547, + 427 + ], + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 427 + ], + "type": "text", + "content": "Figure 1. The cause of multi-view inconsistency in diffusion-based novel-view synthesis models. (a) Diffusion models incorporate randomness for diversity and better distribution modeling; this independent generation process produces realistic views under specific instances but may produce different plausible views for various instances, lacking alignment across adjacent views. (b) In contrast, ViewFusion incorporates an auto-regressive process to reduce uncertainty and achieve multi-view consistency, by ensuring a correlated denoising process that ends at the same high-density area, fostering consistency across views." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": "68, 73]. Follow-up approaches [34, 69] remove the requirement of text conditioning and instead take an image and target pose as conditions for NVS. However, distillation [63] is still required as the diffusion model cannot produce the multi-view consistent outputs that are appropriate for certain downstream tasks (e.g., optimizing Neural Radiance Fields (NeRFs) [41])." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": "Under the single-view setting, maintaining multi-view consistency remains particularly challenging since there may exist several plausible outputs for a novel view that are aligned with the given input image. For diffusion-based approaches which generate novel views in an independent manner [34, 69], this results in synthesized views containing artifacts of multi-view inconsistency (Fig. 1a). Previous work [33, 36, 38, 53, 74, 77] focuses on improving the robustness of the downstream reconstruction to address the inconsistency issue, including feature projection layers in the NeRF [33] or utilising three-dimensional priors to constrain NeRF optimization [36, 77], yet these techniques require training or fine-tuning to align additional modules to the original diffusion models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this work, we address the multi-view inconsistency that arises during the process of view synthesis. Rather than" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 193, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 193, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 193, + 712 + ], + "type": "text", + "content": "*Work done during internship at Amazon" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9870" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "independently synthesizing views conditioned only on the initial reference image, we develop a novel approach where each subsequently generated view is also conditioned on the entire set of previously generated views. Specifically, our method incorporates an auto-regressive process into the diffusion process to model the joint distribution of views, guiding our novel-view synthesis by maintaining the denoising direction towards the same high density area of already generated views (Fig. 1b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 324 + ], + "type": "text", + "content": "Our framework, named ViewFusion, relaxes the single-view conditioning requirement of typical diffusion models through an interpolated denoising process. ViewFusion offers several additional advantages: 1) it can utilize all available views as guidance, thereby enhancing the quality of generated images by incorporating more information; 2) it does not require any additional fine-tuning, effortlessly converting pre-trained single-view conditioned diffusion models into multi-view conditioned diffusion models; 3) it provides greater flexibility in setting adaptive weights for condition images based on their relative view distance to the target view." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 324, + 258, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 324, + 258, + 337 + ], + "spans": [ + { + "bbox": [ + 59, + 324, + 258, + 337 + ], + "type": "text", + "content": "The contributions of this paper are the following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 339, + 287, + 506 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 339, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 287, + 387 + ], + "type": "text", + "content": "- We propose a training-free algorithm which can be directly applied to pre-trained diffusion models to improve multi-view consistency of synthesized views and supports multiple conditional inputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 387, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 287, + 434 + ], + "type": "text", + "content": "- Our method utilizes a novel, auto-regressive approach which we call Interpolated Denoising, that implicitly addresses key limitations of previous auto-regressive approaches for view synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 435, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 287, + 506 + ], + "type": "text", + "content": "- Extensive empirical analysis on ABO [6] and GSO [9] show that our method is able to achieve better 3D consistency in image generation, leading to significant improvements in novel view synthesis and 3D reconstruction of shapes under single-view and multi-view image settings over other baseline methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 518, + 134, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 134, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 134, + 532 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 539, + 208, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 208, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 208, + 552 + ], + "type": "text", + "content": "2.1. 3D-adapted Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Diffusion models have excelled in image generation using conditional inputs [20, 46, 50, 84] and given this success in the 2D domain, recent works have tried to extend diffusion models to 3D content generation [1, 3, 5, 10, 16, 18, 22, 24, 25, 27, 37, 43-45, 66, 82, 83] - although the scarcity of 3D data presents a significant challenge to directly train these diffusion models. Nonetheless, pioneer works such as DreamFusion [47] and Score Jacobian Chaining [63] leverage pre-trained text-conditioned diffusion models to craft 3D models via distillation. Follow-up approaches [4, 31, 57, 68] improve this distillation in terms of speed, resolution and shape quality. Approaches such as [40, 49, 57, 73] extend upon this to support image condi" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "tions through the use of captions with limited success due to the non-trivial nature of textual inversion [13]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 106, + 511, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 106, + 511, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 106, + 511, + 119 + ], + "type": "text", + "content": "2.2. Novel View Synthesis Diffusion Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 126, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 545, + 390 + ], + "type": "text", + "content": "Another line of research [2, 8, 17, 28, 35, 56, 58, 61, 62, 69, 71, 78, 80, 86] directly applies 2D diffusion models to generate multi-view images for shape reconstruction. To circumvent the weakness of text-conditioned diffusion models, novel-view synthesis diffusion models [34, 69] have also been explored, which take an image and target pose as conditions to generate novel views. However, for these approaches, recovering a 3D consistent shape is still a key challenge. To mitigate 3D inconsistency, Liu et al. [33] suggests training a Neural Radiance Field (NeRF) with feature projection layers. Concurrently, other works [36, 38, 70, 74, 77] add modules to original diffusion models for multi-view consistency, including epipolar attention [77], synchronized multi-view noise predictor [36] and cross-view attention [38, 70]; although these methods require fine-tuning an already pre-trained model. We adopt a different paradigm, instead of extending a single-view diffusion model with additional trainable models that incorporate multi-view conditions, our training-free method enables pre-trained diffusion models to incorporate previously generated views via the denoising step and holistically extends these models into multi-view settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 400, + 531, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 531, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 531, + 412 + ], + "type": "text", + "content": "2.3. Other Single-view Reconstruction Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 420, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 576 + ], + "type": "text", + "content": "Before the prosperity of generative models used in 3D reconstruction, many works [11, 12, 14, 15, 26, 29, 30, 59, 64, 75] reconstructed 3D shapes from single-view images using regression [14, 15, 29, 64, 75] or retrieval [59], both of which face difficulties in generalizing to real data or new categories. Methods based on Neural Radiance Fields (NeRFs) [41] have found success in novel-view synthesis, but these approaches typically depend on densely captured images with accurately calibrated camera positions. Currently, several studies are investigating the adaptation of NeRF to single-view settings [21, 32, 52, 79]; although, reconstructing arbitrary objects from single-view images is still a challenging problem." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 589, + 362, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 589, + 362, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 589, + 362, + 601 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 610, + 520, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 520, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 520, + 623 + ], + "type": "text", + "content": "3.1. Denoising Diffusion Probabilistic Models" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Denoising diffusion probabilistic models (DDPM) [19, 54] are a class of generative models that model the real data distribution " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " with a tractable model distribution " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " by learning to iteratively denoise samples. It learns a probability model " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " to convert unstructured noise " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " to real samples " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " in the form of a Markov chain, with Gaussian transitions. The Gaussian transition is" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9871" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 64, + 542, + 140 + ], + "blocks": [ + { + "bbox": [ + 51, + 64, + 542, + 140 + ], + "lines": [ + { + "bbox": [ + 51, + 64, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 51, + 64, + 542, + 140 + ], + "type": "image", + "image_path": "528efee45774322d21d034cc41d8286bfdb4fd5f9e973e81c37b10e572449e0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 141, + 542, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 141, + 542, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 141, + 542, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 542, + 213 + ], + "type": "image", + "image_path": "09a29e0c94767c8285e2f8fb88c6e6a41161cdf4c3bf0a2c04e7980e011aeb2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 215, + 542, + 285 + ], + "blocks": [ + { + "bbox": [ + 48, + 215, + 542, + 285 + ], + "lines": [ + { + "bbox": [ + 48, + 215, + 542, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 215, + 542, + 285 + ], + "type": "image", + "image_path": "f65e89dee7baac846a779ab1e994dee7b4eabdd8ebb56394b0c62a31b118d7d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 287, + 298, + 299 + ], + "lines": [ + { + "bbox": [ + 292, + 287, + 298, + 299 + ], + "spans": [ + { + "bbox": [ + 292, + 287, + 298, + 299 + ], + "type": "text", + "content": "中" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "text", + "content": "Figure 2. Illustration of the Auto-Regressive Generation Process. In our approach, we extend a pre-trained diffusion model from single-stage to multi-stage generation and we maintain a view set that contains all generated views. For each stage, we construct " + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "text", + "content": " reverse diffusion processes and sharing a common starting noise. At each time step within this generation stage, the diffusion model predicts " + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "text", + "content": " noises individually. These " + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 300, + 545, + 357 + ], + "type": "text", + "content": " noises are then subjected to weighted interpolation through the Noise Interpolation Module, concluding the denoising step with the a shared interpolated noise for subsequent denoising steps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 93, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 93, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 93, + 382 + ], + "type": "text", + "content": "defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 387, + 297, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 297, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 297, + 429 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "073782d68b452f3c38e1fd04a20473ef385c0e839ded85ed0bd2ed77bb29e97e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\beta_{t}, t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "text", + "content": " are the variance schedule parameter and timestep in the denoising process respectively. The reverse denoising process starts from a noise sampled from a Gaussian distribution " + }, + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "text", + "content": " and is constructed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 492, + 298, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 298, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 298, + 534 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {T}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}", + "image_path": "f7d71a2ea4c74633a11ea6bd488f5c5b17378429d91d753bdcc23f2b2705db00.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": "where the variance " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": " is a time-dependent constant [19], and " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": " is the mean from the learned noise predictor " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 73, + 574, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 574, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 73, + 574, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} (\\mathbf {x} _ {t}, t) = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right). \\quad (3)", + "image_path": "ffb529edd6e37f2b71a0c4b516d03c7543a255b334d78101ffef48f8e43bbcb4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "content": " are constants derived from " + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "content": ". The objective of noise predictor " + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "content": " is simplified to:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 634, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 634, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 62, + 634, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\ell = \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| _ {2} \\right], \\tag {4}", + "image_path": "e55d7cc23a5469827015470c0f864789e914355929e9fad9cf7e9f7a5743e432.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "type": "text", + "content": " is a random variable sampled from " + }, + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 653, + 277, + 666 + ], + "type": "text", + "content": " [19]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 671, + 233, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 233, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 233, + 683 + ], + "type": "text", + "content": "3.2. Pose-Conditional Diffusion Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Similar to other generative models [42, 55], diffusion models inherently possess the capability to model conditional" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": "distributions of the form " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,y)" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " is the condition. We employ a conditional denoising autoencoder, denoted as " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,y)" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " which enables controlling the synthesis process through a variety of input modalities, including textual descriptions [50], semantic maps [20, 46], or other image-to-image translation tasks [20]. In the following, we present a range of approaches to novel-view synthesis, exploring how various works, including our own, approach the concept of a single reverse diffusion step. Through this comparison, we clarify and establish the underlying relationships between these different methodologies. The notation will follow that bottom subscript " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "(\\cdot)_t" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " indicates the diffusion step and upper subscript " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "(\\cdot)^i" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " relates to the view index. Subsequently, the " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": "-th condition image and its relative pose to the target view are defined as " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^i" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\pi^i" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": ", respectively, and the noisy image to be denoised at timestep " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 371, + 545, + 576 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "text", + "content": "Direct condition was applied by Zero 1-to-3 [34] to the reverse process when given a single input image and target pose " + }, + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^1" + }, + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\pi^1" + }, + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 384, + 626, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 626, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 384, + 626, + 545, + 641 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {y} ^ {1}, \\pi^ {1}\\right). \\tag {5}", + "image_path": "37fdf6d62b06b1b109c695792454b9553cd3dc70a8f4c303b4cae7aa9f1e9679.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 692 + ], + "type": "text", + "content": "Stochastic conditioning was formulated by [69] which can leverage multiple views sampled from a collection of views " + }, + { + "bbox": [ + 304, + 654, + 545, + 692 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{y},\\pi}(\\mathcal{Y},\\pi)" + }, + { + "bbox": [ + 304, + 654, + 545, + 692 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 338, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 338, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {y} ^ {i}, \\pi^ {i}\\right), \\left\\{\\mathbf {y} ^ {i}, \\pi^ {i} \\right\\} \\sim p _ {\\mathbf {y}, \\pi} (\\mathcal {Y}, \\pi), \\tag {6}", + "image_path": "d89c5709d39ac34cbec031eecc560533bf3fe3105d20dfceccd0c1701d9f0f68.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9872" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "where the sampling of image and pose happens at each diffusion step " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 125, + 287, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 125, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 287, + 161 + ], + "type": "text", + "content": "Joint output distribution was shown in SyncDreamer[36] which learns a joint distribution of many views given an image condition " + }, + { + "bbox": [ + 47, + 125, + 287, + 161 + ], + "type": "inline_equation", + "content": "y^{1}" + }, + { + "bbox": [ + 47, + 125, + 287, + 161 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 121, + 169, + 287, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 169, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 121, + 169, + 287, + 184 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {t - 1} ^ {1: N} \\mid \\mathbf {x} _ {t} ^ {1: N}, \\mathbf {y} ^ {1}, e ^ {1}\\right), \\tag {7}", + "image_path": "00a37d90ed7669fd387757a0f2efc71888bc8f4ad611c343b5b05df06f12ab1b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "type": "text", + "content": " is the number of generated novel views and " + }, + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 47, + 194, + 287, + 254 + ], + "type": "text", + "content": " is the elevation condition (partial pose information). We note that in this formulation the target poses are not fully specified as part of the condition allowing for diverse pose generation of outputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "type": "text", + "content": "Auto-regressive distribution is an auto-regressive distribution setting which can generate an arbitrary number of views given a single or multiple condition images and poses contained in the set of " + }, + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{1:N - 1}" + }, + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\pi^{1:N - 1}" + }, + { + "bbox": [ + 47, + 282, + 287, + 329 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 338, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 338, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 338, + 287, + 354 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}, \\pi^ {1: N - 1}\\right). \\tag {8}", + "image_path": "e594eaf9a3703eb45170e8187f59982fa5aa90632e7073cd7e0dec5af754ec7b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 363, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 287, + 399 + ], + "type": "text", + "content": "Our approach falls in the auto-regressive category and for the remainder of this section we detail the implementation to achieve this sampling strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 407, + 177, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 177, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 177, + 420 + ], + "type": "text", + "content": "3.3. Interpolated Denoising" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 426, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 593 + ], + "type": "text", + "content": "The standard DDPM model has been adapted for novel-view image synthesis by using an image and target pose (i.e., rotation and translation offsets) as conditional inputs [69]. Following training on a large-scale dataset, this approach has demonstrated the capability for zero-shot reconstruction [34]. To address the challenge of maintaining multi-view consistency, we employ an auto-regressive approach for generating sequential frames (See Fig. 2). Instead of independently producing each frame from just the input images – a process prone to significant variations between adjacent images – we integrate an auto-regressive algorithm into the diffusion process. This integration enables us to model a conditional joint distribution, ensuring smoother and more consistent transitions between frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": "To guide the synthesis of novel views using images under different views, we design an interpolated denoising process. For the purpose of this derivation, we assume access to an image set containing " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " images denoted as " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}^1,\\dots,\\mathbf{y}^{N - 1}\\}" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ". We want to model the distribution of the " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": "-th view image conditioned on these " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " views " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T}^N|\\mathbf{y}^{1:N - 1})" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", where the relative pose offsets " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pi^i,i\\in" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{1,N - 1\\}" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " between the condition images " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}^1,\\dots,\\mathbf{y}^{N - 1}\\}" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^N" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " are omitted for simplicity. The forward process of the multi-view conditioned diffusion model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "is a direct extension of the vanilla DDPM in Eq. 1, where noises are added to every view independently by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 102, + 545, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 102, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 346, + 102, + 545, + 134 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} ^ {N} \\mid \\mathbf {y} ^ {1: N}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} ^ {N} \\mid \\mathbf {x} _ {t - 1} ^ {N}, \\mathbf {y} ^ {1: N}\\right) \\tag {9}", + "image_path": "945b271d2176d3a00d3b9f7cf42b2be5d8d6361d04a21be407246c537ee4154b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t^N | \\mathbf{x}_{t-1}^N, \\mathbf{y}^{1:N}) = \\mathcal{N}(\\mathbf{x}_t^N; \\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}^N, \\beta_t \\mathbf{I})" + }, + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "type": "text", + "content": ". The initial is defined as " + }, + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^N := \\mathbf{y}^N" + }, + { + "bbox": [ + 305, + 140, + 545, + 178 + ], + "type": "text", + "content": ". Similarly, following Eq. 2, the log reverse process is constructed as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 184, + 545, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 184, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 309, + 184, + 545, + 330 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {N} \\mid \\mathbf {x} _ {T} ^ {N}, \\mathbf {y} ^ {1: N - 1}\\right) = \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}\\right) \\\\ \\approx_ {(1)} ^ {\\sum} \\log \\prod_ {n = 1} ^ {N - 1} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {N} \\mid \\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {n}\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N - 1} \\log \\mathcal {N} \\left(\\mathbf {x} _ {t - 1} ^ {N}; \\mu_ {\\theta} ^ {n} \\left(\\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {n}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t - 1} ^ {N}; \\bar {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {N}, \\mathbf {y} ^ {1: N - 1}, t\\right), \\bar {\\sigma} _ {t} ^ {2} \\mathbf {I}\\right). \\tag {10} \\\\ \\end{array}", + "image_path": "e127687a899dd8cf9d06d822e8c1ea4dde0db2ccb32a2d154fb912db067dc445.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\mu}_{\\theta}, \\bar{\\sigma}_{t}^{2}" + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": " are taken as the mean and variance of the summation of " + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": " log-normal distributions. A note on subscript (1) in Eq.10; to avoid cluttering the derivation, we assume " + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": " independent inferences of the same random variable " + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}^{N}" + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": " using a different " + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{n}" + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": " that results in " + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 305, + 331, + 545, + 415 + ], + "type": "text", + "content": " independent normal distributions, which would require an additional subscript that we omitted for clarity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 421, + 480, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 480, + 434 + ], + "type": "text", + "content": "3.4. Single and Multi-view Denoising" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": "In practice, however, we may not have all " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": " views but a single view or a handful of views. For the reminder of this section, we treat an estimated view as " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^n" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": ", to be the " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": "-th view " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^n" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": " after a full reverse diffusion process. We use " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\bar{\\mu}_{\\theta}(\\mathbf{x}_t,\\mathbf{y}^{1:N - 1},t)" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": " as the weighted average of " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{n}(\\mathbf{x}_{t},\\mathbf{y}^{n},t)" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": ". For computing " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\bar{\\mu}_{\\theta}" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": " using both given views and estimated views we adopt an approach where different views contribute differently to the target view, and we assign the weight " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\omega_{n}" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": "-th view in practice while satisfying the constraint " + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\sum_{n = 1}^{N - 1}w_n = 1" + }, + { + "bbox": [ + 305, + 439, + 545, + 571 + ], + "type": "text", + "content": ". The Noise Interpolation Module in Fig. 2 is modeled as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 578, + 551, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 551, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 551, + 689 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {1: N - 1}, t\\right) = \\sum_ {n = 1} ^ {N - 1} \\omega_ {n} \\mu_ {\\theta} ^ {n} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right) \\\\ = \\sum_ {n = 1} ^ {N - 1} \\omega_ {n} \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right)\\right) \\\\ = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\sum_ {n = 1} ^ {N} \\omega_ {n} \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, \\mathbf {y} ^ {n}, t\\right)\\right). \\tag {11} \\\\ \\end{array}", + "image_path": "a09318c47512fedf2b788f2047131f6ee8f12bef46549421c4566735aa5fed68.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In our approach, as the full view set is not given to us, we approximate this process by an auto-regressive way and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9873" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": "grow the condition set during the generation. We define the weight parameter " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\omega_{n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": " based on the angle offset, i.e., azimuth " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "(\\Delta_{a}^{n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": ", elevation " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "(\\Delta_{e}^{n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": ", and distance " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "(\\Delta_{d}^{n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": ", between the target view and the " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "n - th" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": " condition view. The core idea is to assign higher importance to near-view images during the denoising process while ensuring that the weight for the initial condition image does not diminish too rapidly, even when the target view is positioned at a nearly opposite angle. We use an exponential decay weight function for the initial condition image, defined as " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\omega_{n} = e^{-\\frac{\\Delta^{n}}{\\tau_{c}}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": " is the temperature parameter that regulates the decay speed, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\Delta^n" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": " is the sum of the absolute relative azimuth " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "(\\Delta_{a}^{n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": ", elevation " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "(\\Delta_{e}^{n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": ", and distance " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "(\\Delta_{d}^{n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": " between the target and condition poses. We calculate " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\Delta^n" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\Delta^n = |\\Delta_a^n| / \\pi +|\\Delta_e^n| / \\pi +|\\Delta_d^n|" + }, + { + "bbox": [ + 46, + 72, + 289, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "text", + "content": "For the weights of the remaining images denoted as " + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\{x_0^2,\\dots,x_0^N\\}" + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "text", + "content": ", all generated from the initial condition image " + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "inline_equation", + "content": "y^{1}:= x_{0}^{1}" + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "text", + "content": ", we use a softmax function to define the weights " + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\omega_{n}" + }, + { + "bbox": [ + 47, + 243, + 287, + 291 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 300, + 287, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 300, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 68, + 300, + 287, + 336 + ], + "type": "interline_equation", + "content": "\\omega_ {n} = \\operatorname {S o f t m a x} \\left(\\frac {e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}{\\sum_ {n = 2} ^ {N} e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}\\right), n = 2, \\dots , N \\tag {12}", + "image_path": "64ff12f27e366206066fcdf6eb55a202b2c88bc9fb8d40dc97d64d030db0ffef.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\Delta_{n}" + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "text", + "content": " represents the relative pose offset between target view and the " + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "text", + "content": "-th generated view, and " + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 46, + 344, + 287, + 403 + ], + "type": "text", + "content": " represents the temperature parameter for generated views. As an example, in the single-view case, the weights are expressed as follows," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 287, + 475 + ], + "type": "interline_equation", + "content": "\\omega_ {n} = \\left\\{ \\begin{array}{l l} e x p \\left(- \\frac {\\Delta^ {n}}{\\tau_ {c}}\\right), & n = 1 \\\\ (1 - \\omega_ {1}) \\operatorname {S o f t m a x} \\left(\\frac {e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}{\\sum_ {n = 2} ^ {N} e ^ {- \\frac {\\Delta^ {n}}{\\tau_ {g}}}}\\right), & n \\neq 1 \\end{array} \\right. \\tag {13}", + "image_path": "84fa77be64cc0d31dda17d32aaed51246a91d9454f483609ebda5b6856fcf43f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "type": "text", + "content": "we apply the term " + }, + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "1 - \\omega_{1}" + }, + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "type": "text", + "content": " on the generated image weights to ensure the requirement of " + }, + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\sum_{n = 1}^{N - 1}w_n = 1" + }, + { + "bbox": [ + 46, + 483, + 287, + 544 + ], + "type": "text", + "content": " will be met. In practice, Eq.13 is generalised to allow the condition set can be larger than 1, i.e., multi-view generation (see supplementary)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 185, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 185, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 185, + 565 + ], + "type": "text", + "content": "3.5. Step-by-step Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 571, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 679 + ], + "type": "text", + "content": "Single image generation. When applying the autoregressive approach to image generation, we have devised a generation trajectory, as illustrated in Fig. 3a. We uniformly sample views along this trajectory in sequence. Each previously generated view image on this trajectory is incorporated into the condition set, providing guidance for the subsequent denoising process via our interpolated denoising method. To determine the number of steps, denoted as " + }, + { + "bbox": [ + 46, + 571, + 287, + 679 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 571, + 287, + 679 + ], + "type": "text", + "content": ", needed for this trajectory, we use the following formula:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 689, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 689, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 287, + 717 + ], + "type": "interline_equation", + "content": "S = \\max \\left(\\lceil \\frac {\\Delta_ {a} ^ {N}}{\\delta} \\rceil , \\lceil \\frac {\\Delta_ {e} ^ {N}}{\\delta} \\rceil\\right). \\tag {14}", + "image_path": "160e94f531d033546fbb9dc5e0ca0ae33bcc6752df691c0927553812788680a9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 416, + 177 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 416, + 177 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 416, + 177 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 416, + 177 + ], + "type": "image", + "image_path": "76a32f7bd0c47340cfd1ac3d1c3fae945a5e8ae9fd01e5e586c6f6cbd17d840e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 180, + 407, + 190 + ], + "lines": [ + { + "bbox": [ + 315, + 180, + 407, + 190 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 407, + 190 + ], + "type": "text", + "content": "(a) Single image generation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 68, + 543, + 177 + ], + "blocks": [ + { + "bbox": [ + 432, + 68, + 543, + 177 + ], + "lines": [ + { + "bbox": [ + 432, + 68, + 543, + 177 + ], + "spans": [ + { + "bbox": [ + 432, + 68, + 543, + 177 + ], + "type": "image", + "image_path": "150217d51a3426f2d2d7b72a52d4648341a3bb8a84f1b5c79978fb099df608be.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 180, + 531, + 190 + ], + "lines": [ + { + "bbox": [ + 446, + 180, + 531, + 190 + ], + "spans": [ + { + "bbox": [ + 446, + 180, + 531, + 190 + ], + "type": "text", + "content": "(b) Spin video generation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 192, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 237 + ], + "type": "text", + "content": "Figure 3. Illustration of Step-by-step Generation. (a) we uniformly sample views along this trajectory in sequence to generate a novel-view image; (b) we sample views from nearest to furthest views according to to view distance to generate a " + }, + { + "bbox": [ + 305, + 192, + 545, + 237 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 305, + 192, + 545, + 237 + ], + "type": "text", + "content": " spin video." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "content": "Here, we set the maximum offset per step " + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "content": " to determine the step count " + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "content": ", also based on the target view offsets " + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\Delta_{a}^{N}" + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\Delta_{e}^{N}" + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "content": ". We then proceed to sample the " + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 251, + 545, + 300 + ], + "type": "text", + "content": "-th view using the following equation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 309, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 309, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 321, + 309, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\left(\\Delta_ {a} ^ {n}, \\Delta_ {e} ^ {n}, \\Delta_ {d} ^ {n}\\right) = \\left(\\frac {\\Delta_ {a} ^ {N}}{S} * n, \\frac {\\Delta_ {e} ^ {N}}{S} * n, \\frac {\\Delta_ {d} ^ {N}}{S} * n\\right) \\tag {15}", + "image_path": "33c85e6dbc187afa69ea7cd2de0adf2ceb315c136a7412b2d917e87cec388889.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "text", + "content": "Spin videos generation. In contrast to generating a single target image, the process of spin video generation begins from an initial image and concludes at the same position. To achieve this, we need to modify the generation order to leverage the broad range of rotation images, rather than simply following the rotation degree range of " + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "inline_equation", + "content": "[0^{\\circ}, 360^{\\circ}]" + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "text", + "content": " in sequence. This is because, at " + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "inline_equation", + "content": "\\Delta_{a} = \\pi" + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "text", + "content": ", the view is opposite to the conditioning view, marking the end of the generation process. To establish the generation order for spin video generation, we introduce the minimum azimuth offset, denoted as " + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "text", + "content": ", and employ a skip trajectory with the following order: " + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "inline_equation", + "content": "\\{\\delta, -\\delta, 2\\delta, -2\\delta\\dots, N\\delta\\}" + }, + { + "bbox": [ + 305, + 346, + 546, + 513 + ], + "type": "text", + "content": ", shown in Fig. 3b. For simplicity, we only consider rotation along the azimuth dimension in this context." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 527, + 388, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 388, + 541 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 388, + 541 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 548, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 545, + 633 + ], + "type": "text", + "content": "Datasets. We evaluate our method and compare to baselines on the ABO [6] and GSO [9] datasets. These datasets are out-of-the-distribution as all baselines are trained on the Objaverse [7]. We also provide qualitative results on real images to showcase performance of our method on in-the-wild images in the supplementary. For additional results, please refer to the videos contained in the supplementary." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 652, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 652, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 652, + 545, + 676 + ], + "type": "text", + "content": "Metrics. We assess our novel-view synthesis on three main criteria:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 547, + 713 + ], + "type": "text", + "content": "1. Image Quality: LPIPS [85], PSNR, and SSIM [67] metrics to help gauge the similarity between synthesized and ground-truth views." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9874" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 145 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 145 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 145 + ], + "type": "table", + "html": "
DatasetMethodFree RenderingsSyncDreamer Renderings
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
ABOZero123 [34]0.879621.330.09610.782218.270.1999
SyncDre. [36]0.771213.430.21820.803119.070.1816
Ours0.884821.430.09230.798318.750.1985
GSOZero123 [34]0.871020.330.10290.792518.060.1714
SyncDre. [36]0.802314.420.18330.802418.200.1647
Ours0.882020.730.09580.807618.400.1703
", + "image_path": "7ce42425512954afc043f1e4bb7fa203db6dc13c8bed435103fa7095531dda58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "type": "text", + "content": "Table 1. Quantitative results on ABO and GSO datasets with arbitrary (left) and discrete (right) rotation and translation. Free renderings are a set of arbitrary rotation and translation as target generation view, while SyncDreamer renderings are a fixed set of 16 views with discrete azimuth, fixed elevation and distance, i.e., azimuth " + }, + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\{0^{\\circ},22.5^{\\circ},45^{\\circ},\\dots,315^{\\circ},337.5^{\\circ}\\}" + }, + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "type": "text", + "content": ", elevation " + }, + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "type": "inline_equation", + "content": "= 30^{\\circ}" + }, + { + "bbox": [ + 46, + 147, + 287, + 236 + ], + "type": "text", + "content": ". Note that SyncDreamer [36] cannot generate images under arbitrary views apart from the predefined 16 camera positions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 255, + 286, + 327 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 255, + 286, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 286, + 290 + ], + "type": "text", + "content": "2. Multi-View Consistency: Using SIFT [39], LPIPS [85] and CLIP [48], we measure the uniformity of images across various perspectives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 291, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 286, + 327 + ], + "type": "text", + "content": "3. 3D Reconstruction: Chamfer distances and F-score between ground-truth and reconstructed shapes determine geometrical consistency." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 334, + 167, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 167, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 167, + 346 + ], + "type": "text", + "content": "4.1. Novel-view synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 352, + 286, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 286, + 507 + ], + "type": "text", + "content": "In Tab. 1, we show quantitative results for novel-view synthesis under arbitrary and fixed-view settings. The fixed-view setting uses the rendering set of [36] and ensures a fair comparison to [36] which is limited to this fixed-view generation setting. As shown in Tab. 1, our method can produce comparable results to [36], without any fine-tuning on the rendering set of [36]. Under the arbitrary-view setting, we sample the nearest view from the rendering set of [36] to the designated target view for generation; [36] underperforms both [34] and our approach. Overall, regardless of the evaluation setting, our method consistently outperforms [34] and even outperforms [36] on the GSO dataset under the favourable fixed-view setting for [36]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "content": "The task of novel view synthesis serves as a precursor for enabling more significant downstream applications, such as 3D reconstruction, by generating requisite input views. Thus, in addition to image quality, thorough evaluation needs to encompass the multi-view consistency of the synthesized perspectives. This additional criterion ensures that the generated imagery not only appears visually compelling but also aligns geometrically across different viewpoints. In the following sections, we evaluate against baselines on 3D consistency and show that the significant improvements our approach offers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 714 + ], + "type": "text", + "content": "Multi-view Consistency. For measuring multi-view consistency of synthesized views, quantitative results are shown in Tab. 2, while qualitative comparisons are shown in Figs. 4 and 5. Here, we can see when compared to the baselines of [34] and [36], our approach excels in generating images" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 70, + 525, + 178 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 525, + 178 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 525, + 178 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 525, + 178 + ], + "type": "table", + "html": "
DatasetMethodViewSIFT↑LPIPS↓CLIP↑
ABOZero123 [34]13.380.17820.9604
SyncDreamer [36]1612.360.18950.9584
Ours13.510.16020.9664
Zero123 [34]17.030.12310.9725
Ours3618.010.09660.9812
GSOZero123 [34]12.580.14110.9482
SyncDreamer [36]1613.240.13150.9532
Ours13.830.11870.9601
Zero123 [34]15.200.10560.9592
Ours3617.950.06760.9773
", + "image_path": "a9366e1829717b98dbd79e98adadb87459242e34f5e09f9986bd809919deea84.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 179, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 304, + 179, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 545, + 235 + ], + "type": "text", + "content": "Table 2. Quantitative results for multi-view consistency. We report the SIFT matching point number, LPIPS and CLIP similarity between adjacent frames to evaluate the multi-view consistency. Note that SyncDreamer can only generate 16 view images for a spin video due to the constraints imposed by its training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 324, + 243, + 525, + 340 + ], + "blocks": [ + { + "bbox": [ + 324, + 243, + 525, + 340 + ], + "lines": [ + { + "bbox": [ + 324, + 243, + 525, + 340 + ], + "spans": [ + { + "bbox": [ + 324, + 243, + 525, + 340 + ], + "type": "table", + "html": "
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 [34] (1 view)0.882021.510.0945
Ours (1 view)0.887021.610.0904
Ours (2 views)0.891321.920.0887
Ours (3 views)0.899522.740.0815
GSOZero123 [34] (1 view)0.872120.420.1017
Ours (1 view)0.883020.870.0948
Ours (2 views)0.890121.250.0893
Ours (3 views)0.897921.950.0792
", + "image_path": "560c0d82367b56e044d421ea2c5c95e41483286281615da872e76beceadcbb3f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 341, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 406 + ], + "type": "text", + "content": "Table 3. Quantitative results for multi-conditioned generation. Our approach outperforms the Zero-1-to-3 baseline [34] by extending its single-view reconstruction framework to a multi-view reconstruction framework. Note that 3 condition views are removed from test set, and thus the results are slightly different to Tab. 1." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 425, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 461 + ], + "type": "text", + "content": "that are both semantically consistent with the input image and maintains multi-view consistency in terms of colors and geometry under arbitrary-view settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 476, + 512, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 476, + 512, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 476, + 512, + 489 + ], + "type": "text", + "content": "4.2. Multi-view conditional setting for NVS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": "Additionally, our approach allows for the extension of single-view conditioned models into multi-view conditioned models which can take multiple conditional images as input. The results presented in Tab. 3 showcase the advantages our method offers, as novel-view synthesis quality improves with an increasing number of conditional input views. This is a significant improvement over the Zero-1-to-3 baseline and demonstrates the efficacy of our proposed method in the multi-view setting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 620, + 416, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 416, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 416, + 632 + ], + "type": "text", + "content": "4.3. 3D Reconstruction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "For shape reconstruction, we present our results both quantitatively (Tab. 4) and qualitatively (Fig. 6). We compare against baselines which synthesize novel views as well as direct image-to-shape approaches [23, 44]. For the former approach, instead of using distillation [63], we train NeuS [65] on synthesized images to recover a shape." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9875" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 61, + 531, + 248 + ], + "blocks": [ + { + "bbox": [ + 62, + 61, + 531, + 248 + ], + "lines": [ + { + "bbox": [ + 62, + 61, + 531, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 61, + 531, + 248 + ], + "type": "image", + "image_path": "0fa386518989aafa33eda739ac2cfba214f2b67bda5dbad0a554e72140b8cd47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 547, + 273 + ], + "type": "text", + "content": "Figure 4. Qualitative results for " + }, + { + "bbox": [ + 46, + 249, + 547, + 273 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 249, + 547, + 273 + ], + "type": "text", + "content": " Spin Video Generation. Note the additional consistency in generated views our approach offers over the competing baselines shown in the bounding boxes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 281, + 286, + 434 + ], + "blocks": [ + { + "bbox": [ + 53, + 281, + 286, + 434 + ], + "lines": [ + { + "bbox": [ + 53, + 281, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 286, + 434 + ], + "type": "image", + "image_path": "54e7d11b2170e71469b442caff833d81bbe54f0a89fa1ee14e4e40f8c7ac3aa5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 436, + 288, + 480 + ], + "lines": [ + { + "bbox": [ + 46, + 436, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 288, + 480 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison for Motion Smoothness. We visualize the output videos using space-time Y-t slices through frames of the generated spin video (along the scanline shown in the condition)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 494, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 639 + ], + "type": "text", + "content": "Single-view Reconstruction. We first compare our proposed method with several baselines under the single-view reconstruction setting, including Point-E [44] and Shap-E [23], Zero-1-to-3 [34], One-2-3-45 [33] and SyncDreamer [36]. Qualitative results are shown in Fig. 6; Point-E and Shap-E tend to generate incomplete shapes due to the single-view setting. Multi-view alignment methods such as Syncdreamer [36] and One-2-3-45 [33] capture the general geometry but tend to lose fine details. In comparison, our proposed method achieves the highest reconstruction quality amongst all approaches, where we can generate smooth surfaces and capture detailed geometry with precision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Multi-view Reconstruction Given our approach can also synthesis novel views under the multi-view setting, we also show results where we use 3 conditional input views to synthesize 36 views for shape reconstruction. Compared with existing multi-view reconstruction frameworks such as" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 281, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 308, + 281, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 388 + ], + "type": "table", + "html": "
Cond. viewsMethodGen. viewsABOGSO
CD↓F-score↑CD↓F-score↑
1Point-E [44]N/A0.04280.71440.06720.6340
Shap-E [23]N/A0.04660.73640.03840.7313
One-2-3-45 [33]320.04190.66650.04080.6490
SyncDreamer [36]160.01600.81870.02290.7767
Zero123 [34]160.01470.82260.02060.8045
Zero123 [34]360.01390.82470.02070.8078
Ours160.01330.84230.01770.8274
Ours360.01260.84720.01640.8436
3MonoSDF [81]N/A0.10200.39630.08300.4581
Ours360.01150.85870.01240.8628
", + "image_path": "c62dc23edf7d8123a418b0c2563a251d04f4589bc92c0849cdbe1011da827fcb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 415, + 531, + 536 + ], + "blocks": [ + { + "bbox": [ + 306, + 389, + 545, + 412 + ], + "lines": [ + { + "bbox": [ + 306, + 389, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 545, + 412 + ], + "type": "text", + "content": "Table 4. Quantitative results on reconstructing 3D Shapes using the generated images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 415, + 531, + 536 + ], + "lines": [ + { + "bbox": [ + 318, + 415, + 531, + 536 + ], + "spans": [ + { + "bbox": [ + 318, + 415, + 531, + 536 + ], + "type": "table", + "html": "
DatasetMethodSSIM↑PSNR↑LPIPS↓
ABOZero123 (no interp.)0.879621.330.0961
Standard auto-regression0.801016.770.1854
Interpolated conditions0.724313.260.3770
Interpolated outputs0.8925*21.95*0.1246
Stochastic conditioning [69]0.869920.640.1106
Interpolated denoising0.884821.430.0923
GSOZero123 (no interp.)0.871020.330.1029
Standard auto-regression0.809416.300.1801
Interpolated conditions0.766114.380.3427
Interpolated outputs0.879920.440.1659
Stochastic conditioning [69]0.865819.970.1119
Interpolated denoising0.882020.730.0958
", + "image_path": "f85e0f8a3ecf6929b70da0fce80db99b3bb92d0997eef93531b8d8801bf40cfb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 538, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 583 + ], + "type": "text", + "content": "Table 5. Quantitative comparison between our method to different auto-regressive variants. Note that the Interpolated outputs variant achieves the highest SSIM and PSNR values on the ABO dataset, but generates blurred images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 591, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 640 + ], + "type": "text", + "content": "MonoSDF [81], our approach captures more details and generates smoother surfaces (note that [81] also relies on additional depth and normal estimations for its reconstruction whereas ours does not)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 647, + 373, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 373, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 373, + 658 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Here, we study the effectiveness of our proposed interpolated denoising process by exploring various autoregressive generation variants (in Tab. 5 and Fig. 7). Specifically, we investigate four variants of denoising: stan-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9876" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 541, + 247 + ], + "type": "image", + "image_path": "3e0bf45faf0c6ab9f136010eaf1f0f6dbada123e15c66b9e53d9eda58ef68c08.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 546, + 270 + ], + "type": "text", + "content": "Figure 6. Qualitative results for single-view (1 condition image) and multi-view (3 condition images) 3D Reconstruction. Our method offers the most consistent results across a variety of different reconstructed shapes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 281, + 284, + 395 + ], + "blocks": [ + { + "bbox": [ + 53, + 281, + 284, + 395 + ], + "lines": [ + { + "bbox": [ + 53, + 281, + 284, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 284, + 395 + ], + "type": "image", + "image_path": "494e0ead145369685abfd915e3e2c9b9b2cb47fc3480f78c88f58dffd3336812.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 287, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 417 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison for different auto-regressive generations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "content": "dard auto-regression, interpolated conditions, interpolated outputs, and stochastic conditioning which was proposed in [69]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 482, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 613 + ], + "type": "text", + "content": "Standard Auto-regression. One initial approach to auto-regression involves using the last generated view as the subsequent conditioning. However, this method exhibits bad generation quality, due to the accumulation of errors during the sequential generation process. As each subsequent view relies on the accuracy of the previous one, any inaccuracies or imperfections in earlier stages can compound, leading to a degradation in overall image quality. This limitation highlights the need for more sophisticated auto-regressive strategies to address the issue of error propagation and enhance the quality of generated views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Interpolated Conditions and Interpolated Outputs. Interpolated Conditions and Interpolated Outputs are two straightforward approaches to introduce auto-regressive generation into an existing diffusion model. The former method involves interpolating feature embeddings from condition images and poses, while the latter interpolates the final image feature maps produced by the model. Despite" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 283, + 546, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 546, + 344 + ], + "type": "text", + "content": "SSIM and PSNR metrics showing favorable results for Interpolated Outputs over others in Tab. 5, as illustrated in visual comparisons in Fig. 7 shows that it leads to blurring of the output views and this is corroborated by larger LPIPS distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 374, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 506 + ], + "type": "text", + "content": "Stochastic Conditioning. We also explore the application of the Stochastic Conditioning Sampler proposed by [69] to the Zero-1-to-3 model. We observe that Stochastic Conditioning fails to deliver the desired auto-regressive generation results; this may be attributed to the specific category on which the diffusion model used by Watson et al. [69] was trained, allowing it to handle plausible condition images more effectively. By contrast, Zero-1-to-3 [34] was trained on a cross-category dataset and designed for zero-shot reconstruction. Additionally, our evaluation data contains out-of-the-distribution data (i.e., ABO [6] and GSO [9])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 548, + 546, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 546, + 704 + ], + "type": "text", + "content": "In this work, we have developed ViewFusion, a novel algorithm that addresses the challenge of multi-view consistency in novel-view synthesis with diffusion models. Our approach circumvents the need for fine-tuning or additional modules by integrating an auto-regressive mechanism that incrementally refines view synthesis, utilizing the entire history of previously generated views. Our proposed diffusion interpolation technique extends the denoising process in pre-trained diffusion models from a single-view setting to a multi-view setting without training requirements. Empirical evidence underscores ViewFusion's capability to produce consistently high-quality views, and we achieve significant steps forward in novel view synthesis and 3D reconstruction applications." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9877" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "type": "text", + "content": "[3] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "text", + "content": "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "type": "text", + "content": "[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 390 + ], + "type": "text", + "content": "[6] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 2, 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 435 + ], + "type": "text", + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "text", + "content": "[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "text", + "content": "[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, 2022. 2, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[10] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. arXiv preprint arXiv:2303.17015, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "text", + "content": "[11] George Fahim, Khalid Amin, and Sameh Zarif. Single-view 3d reconstruction: A survey of deep learning methods. Computers & Graphics, 94:164-190, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[12] Kui Fu, Jiansheng Peng, Qiwen He, and Hanxiao Zhang. Single image 3d object reconstruction based on deep learning: A review. Multimedia Tools and Applications, 80:463-498, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Georgia Gkioxari, Justin Johnson, and Jitendra Malik. Mesh r-cnn. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[15] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mache approach to learning 3d surface generation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[16] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images. arXiv preprint arXiv:2304.06700, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In ICML, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 251, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 285 + ], + "type": "text", + "content": "[18] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 286, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 308 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 387 + ], + "type": "text", + "content": "[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 443 + ], + "type": "text", + "content": "[23] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "type": "text", + "content": "[24] Animesh Karnewar, Niloy J Mitra, Andrea Vedaldi, and David Novotny. Holofusion: Towards photo-realistic 3d generative modeling. In ICCV, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 511 + ], + "type": "text", + "content": "[25] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In CVPR, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 512, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 533 + ], + "type": "text", + "content": "[26] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "type": "text", + "content": "[27] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "text", + "content": "[28] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[29] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[30] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdfsrn: Learning signed distance 3d object reconstruction from static images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9878" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[32] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "text", + "content": "[33] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "text", + "content": "[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[35] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "type": "text", + "content": "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. In ICLR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "text", + "content": "[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "text", + "content": "[39] D.G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 1150-1157 vol.2, 1999. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[40] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 579 + ], + "type": "text", + "content": "[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 601 + ], + "type": "text", + "content": "[42] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. CoRR, abs/1411.1784, 2014. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "type": "text", + "content": "[43] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[44] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[45] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc Van Gool, and Sergey Tulyakov." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Autodecoding latent 3d diffusion models. arXiv preprint arXiv:2307.05445, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "text", + "content": "[46] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "text", + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 177, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 231 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 234, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 288 + ], + "type": "text", + "content": "[49] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "type": "text", + "content": "[50] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 336, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 336, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 336, + 545, + 369 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[52] Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, and Deqing Sun. Vq3d: Learning a 3d-aware generative model on image-net. arXiv preprint arXiv:2302.06833, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 417, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 449 + ], + "type": "text", + "content": "[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 452, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 484 + ], + "type": "text", + "content": "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "type": "text", + "content": "[55] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Neural Information Processing Systems, 2015. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 532, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 574 + ], + "type": "text", + "content": "[56] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 578, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 620 + ], + "type": "text", + "content": "[57] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[58] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[59] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9879" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[60] A. Tewari, O. Fried, J. Thies, V. Sitzmann, S. Lombardi, K. Sunkavalli, R. Martin-Brualla, T. Simon, J. Saragih, M. Nießner, R. Pandey, S. Fanello, G. Wetzstein, J.-Y. Zhu, C. Theobalt, M. Agrawala, E. Shechtman, D. B Goldman, and M. Zollhöfer. State of the art on neural rendering. Computer Graphics Forum, 39(2):701-727, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[61] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B Tenenbaum, Frédo Durand, William T Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. arXiv preprint arXiv:2306.11719, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "text", + "content": "[62] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "text", + "content": "[63] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[64] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "text", + "content": "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 402 + ], + "type": "text", + "content": "[66] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 447 + ], + "type": "text", + "content": "[67] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[68] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "text", + "content": "[69] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "text", + "content": "[70] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "text", + "content": "[71] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 615, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 667 + ], + "type": "text", + "content": "[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[73] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints, pages arXiv-2211, 2022. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 588 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[74] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[75] Xianghui Yang, Guosheng Lin, and Luping Zhou. Single-view 3d mesh reconstruction for seen and unseen categories. IEEE Transactions on Image Processing, pages 1-1, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[76] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[77] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[78] Paul Yoo, Jiaxian Guo, Yutaka Matsuo, and Shixiang Shane Gu. Dreamsparse: Escaping from Plato's cave with 2d frozen diffusion model given sparse views. CoRR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[79] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "text", + "content": "[80] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 343, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 397 + ], + "type": "text", + "content": "[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "type": "text", + "content": "[82] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 444, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 486 + ], + "type": "text", + "content": "[83] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. In SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "text", + "content": "[84] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 588 + ], + "type": "text", + "content": "[86] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9880" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_content_list.json b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e880f93e10119283093434e79a58d08c835ec2f5 --- /dev/null +++ b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_content_list.json @@ -0,0 +1,1526 @@ +[ + { + "type": "text", + "text": "Viewpoint-Aware Visual Grounding in 3D Scenes", + "text_level": 1, + "bbox": [ + 236, + 130, + 733, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangxi Shi \nOregon State University \nshixia@oregonstate.edu", + "bbox": [ + 130, + 181, + 326, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhonghua Wu SenseTime Research wuzhonghua@sensetime.com", + "bbox": [ + 370, + 181, + 584, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefan Lee \nOregon State University \nleestef@oregonstate.edu", + "bbox": [ + 627, + 181, + 833, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring expressions for visual objects often include descriptions of relative spatial arrangements to other objects - e.g. \"to the right of\" - that depend on the point of view of the speaker. In 2D referring expression tasks, this viewpoint is captured unambiguously in the image. However, grounding expressions with such spatial language in 3D without viewpoint annotations can be ambiguous. In this paper, we investigate the significance of viewpoint information in 3D visual grounding - introducing a model that explicitly predicts the speaker's viewpoint based on the referring expression and scene. We pretrain this model on a synthetically generated dataset that provides viewpoint annotations and then finetune on 3D referring expression datasets. Further, we introduce an auxiliary uniform object representation loss to encourage viewpoint invariance in learned object representations. We find that our proposed ViewPoint Prediction Network (VPP-Net) achieves state-of-the-art performance on ScanRefer, SR3D, and NR3D - improving Accuracy@0.25IoU by $1.06\\%$ , $0.60\\%$ , and $2.00\\%$ respectively compared to prior work.", + "bbox": [ + 75, + 300, + 473, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 632, + 209, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual grounding of referring expressions requires algorithms to reason about natural language object descriptions to identify object referents in visual scenes. Naturally, describing relative spatial relations between objects is a common strategy when a speaker produces a referring expression – e.g., “The chair to the right of the couch.” As such, many methods for visual grounding put algorithmic emphasis on handling these spatial relations [7, 18, 24, 34, 36].", + "bbox": [ + 75, + 657, + 468, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, these sorts of spatial relations depend upon the point of view of the speaker. For instance, Fig. 1 depicts two different viewpoints of the same 3D scene and notes that these views produce contradictory referring expressions. In 2D visual grounding tasks, the point of view of the speaker is typically also that of the image – introducing no ambiguity into spatial reasoning. However, viewpoint annotations are not typically captured for 3D visual grounding tasks that", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf6c096bcbbc885a1f52dc7855c0fbf3a77f727705e99c500519eeaa452d9108.jpg", + "image_caption": [ + "Figure 1. Expressions referring to the same object from different viewpoints can lead to contradictory spatial relations." + ], + "image_footnote": [], + "bbox": [ + 517, + 270, + 689, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ff6caed88ec5c442b930ac9232976e4f38a69100be06fd70279376febcce266.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 268, + 874, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reason about whole scenes. As a result, identifying referent objects may be ambiguous if multiple object instances align with the referring expression in differing views. Moreover, the lack of reliable grounding for spatial language may simply make the reasoning task more difficult to learn even when the referent is still uniquely identifiable.", + "bbox": [ + 496, + 506, + 892, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite this, most of the recent work in 3D visual grounding has not addressed the role of viewpoint in the learning process [12, 17, 26, 35, 43, 46]. Those methods that have attempted to consider the influence of viewpoint have been hindered by the lack of viewpoint supervision in existing datasets – instead relying on rendering multiple views of a 3D scene and using attention mechanisms to softly select between them [13, 16] in an end-to-end model.", + "bbox": [ + 496, + 597, + 892, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we develop a straightforward approach to viewpoint-aware visual grounding in 3D scenes – explicitly predicting the speaker's viewpoint and then transforming the input accordingly. To support this direction, we create a synthetic visual grounding dataset that provides annotations to directly supervise viewpoint prediction. We propose a novel framework called ViewPoint Prediction Net (VPP-Net) to estimate the viewpoints of expression-pointcloud pairs and rotate the 3D scenes accordingly before performing expression grounding. This model is first pre-trained on the synthetic dataset and then fine-tuned on a combination of the synthetic and target datasets. To our knowledge, this", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14056", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is the first work to introduce a supervised viewpoint estimation model for 3D visual grounding tasks. In experiments on three common 3D visual grounding datasets, this approach paired with viewpoint-related auxiliary losses leads to improved performance over prior art.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our contributions can be summarized as:", + "bbox": [ + 76, + 167, + 455, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce viewpoint prediction as an auxiliary task in the 3D visual grounding task and propose synthetic data generation pipeline to provide viewpoint supervision.", + "- We propose VPP-Net - a novel model that learns viewpoint prediction to transform 3D scenes to reduce ambiguity in spatial-relation grounding.", + "- We design a Uniform Object Representation auxiliary loss that promotes viewpoint invariant feature learning.", + "- Combing these techniques, our proposed approach achieves state-of-the-art performance on the ScanRefer [5], SR3D [2], and NR3D [2] datasets." + ], + "bbox": [ + 76, + 184, + 467, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 361, + 217, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relation-aware 2D Visual Grounding. The visual grounding task has been extensively explored in 2D image scenarios. The goal of a 2D visual grounding model is to locate an object according to a given expression in a 2D image. In recent work, the relationships between objects have been identified as a crucial cue in complex environments. Several works [7, 23, 24, 34, 42] have parsed referring expressions into phrases for better text understanding. Wang et al. [34] proposed to align the generated visual graph with the self-attend object and relation features using a graph attention module. Liu [24] proposed to capture the visual context from a set of coarse-level object proposals by reconstructing the corresponding relations captured from the text. Chen [7] et al. initialize a visual graph with proposed node and edge transformers to learn the representations of objects and relations. Besides explicitly encoding visual relations in the model, relationships can also benefit models by constructing pseudo data [18, 36]. Wu et al. [36] proposed to detect mismatched relations from synthetic data to evaluate the ability of relation understanding of the model. Jiang et.al. [18] introduce an automatic way to generate pseudo visual grounding data for supervised training.", + "bbox": [ + 75, + 386, + 467, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Visual Grounding. The first dataset for visual grounding in 3D was introduced by Chen et al. [5] - presenting the ScanRefer dataset consisting of 3D scenes and natural expression. Achlioptas et al. [2] propose two datasets - the synthetic dataset SR3D and the human-annotated NR3D. Further, they show that template-based synthetic data can benefit the task when models are jointly trained with naturally collected data. In standard approaches, a 3D grounding model first encodes instructions and 3D scene with pretrained text [10, 22, 31] and 3D visual [19, 27-30, 40, 41] models. Then a multimodal module is proposed to fuse the encoded features to estimate the location and", + "bbox": [ + 75, + 719, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "size of the grounded object. 3D-SPS [26] proposed a one-stage method that estimates the location of objects from the point cloud directly. In other works, 3D object detections [11, 19, 25, 33] are leveraged to provide an object-focused candidate pool consisting of object-level features and/or bounding boxes. BUTD-DETR [17] encodes object bounding boxes as a reference and jointly trains on object detection datasets as an auxiliary task. 2D images [20, 21, 37-39] are also utilized to boost the performance of 3D visual grounding. The model proposed in [5] encodes 2D images together with 3D scenes for a better 3D understanding. Likewise, Yang et al. [43] leverage 2D semantics to assist 3D representation learning.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The viewpoint of the annotator is also a significant cue for a better alignment between 3D scenes and referring expressions. Currently, only a few works focused on this aspect have been proposed. MVT [16] and ViewRefer [13] rotate the 3D scenes in different angles and encode them to multi-view representations without explicit supervision of which viewpoint matches to the annotator. Compared with previous work, we first provide a template-based synthetic visual grounding dataset including location and perspective supervision and then propose a supervised training method for viewpoint estimation. We demonstrate this is be beneficial for visual grounding performance.", + "bbox": [ + 496, + 289, + 890, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 488, + 596, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a straight-forward approach for learning to reason about viewpoint in 3D referring expression grounding – directly predicting a rotation and translation of the 3D environment that aligns it with the assumed viewpoint of the observer providing the referring expression. To accomplish this, we introduce a ViewPoint Prediction Network (VPP-Net) – a model that performs 3D referring expression grounding while explicitly estimating the location and heading of the observer. As existing datasets lack observer viewpoint annotations, we develop a synthetic data generation pipeline to provide appropriate training supervision for viewpoint-aware pretraining and then transfer the pretrained model to downstream tasks. Further, we introduce a Uniform Object Representation (UOR) auxiliary loss to ensure object instances are represented consistently regardless of viewpoint. We describe each component below.", + "bbox": [ + 496, + 513, + 890, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Generating Synthetic Viewpoint Supervision", + "text_level": 1, + "bbox": [ + 498, + 770, + 875, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our synthetic data generation pipeline operates in four stages - for a given environment, we (1) sample a location and heading for a virtual observer, (2) identify spatial relationships between annotated objects in view, (3) generate templated referring expressions accordingly, and then (4) find other nearby viewpoints that could also have produced this reference. The result is a tuple of the 3D environment,", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14057", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "referring expression, and valid viewpoint map which we can use for training viewpoint prediction.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input Scenes. We utilize 3D scenes from the ScanNet [9] dataset which are annotated with detected object bounding boxes. For a given scene, we denote bounding box as a set $B$ , their centroids as a set of homogenous XYZ-coordinates $\\mathbf{C} = \\{c_0, c_1, \\dots, c_n \\mid c_i = (x_i, y_i, z_i, 1)\\}$ , and the corresponding object category labels $\\mathcal{Y}_o$ .", + "bbox": [ + 76, + 126, + 467, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Viewpoint Sampling. Regardless of scene size, we split the horizontal (XY) plane into a $10 \\times 10$ grid and uniformly sample a 2D $(x,y)$ translation from the grid. Likewise, we sample a rotation $\\theta$ about the vertical axis by randomly selecting a $10^{\\circ}$ increment between $0^{\\circ}$ and $360^{\\circ}$ .", + "bbox": [ + 75, + 220, + 467, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identifying Spatial Relationships. We apply the sampled translation and rotation to the scene by computing updated bounding box centers. With the affine transformation", + "bbox": [ + 75, + 301, + 467, + 347 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM (\\theta , x, y) = \\left( \\begin{array}{c c c c} \\cos \\theta & - \\sin \\theta & 0 & - x \\\\ \\sin \\theta & \\cos \\theta & 0 & - y \\\\ 0 & 0 & 1 & 0 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 357, + 468, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we update the bounding box centroids as $c_i' = c_i M(\\theta, x, y)^T$ . The effect of this operation is to align the observer's relative spatial terms (e.g. left/right) to the XYZ axes. To detect potential spatial relations between a pair of objects $o_i$ and $o_j$ , we compute the vector $r_{ij} = c_i' - c_j'$ to represent their relative position. We take a set of four unit direction vectors as $\\{d_{front}, d_{back}, d_{left}, d_{right}\\}$ and compute the cosine similarities between $d_k$ and $r_{ij}$", + "bbox": [ + 75, + 430, + 467, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j, k} = \\frac {r _ {i j} d _ {k}}{\\left| \\left| r _ {i j} \\right| \\right| \\times \\left| \\left| d _ {k} \\right| \\right|}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 561, + 468, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a spatial relation $k$ to be valid if $s_{i,j,k} > 0.3$ . For above and below relations, we instead consider the heights and overlap of the bounding boxes. From all valid tuples, we select a random subset as the relation tuple set $\\mathbf{R}$ . We note that this process does not consider the \"visibility\" of objects from the observer location.", + "bbox": [ + 75, + 603, + 467, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expression Generation. Given R and object category annotations, we can produce a set of object referring expressions. For each object mentioned as part of a relation in R, we randomly sample up to four valid relations and convert each into simple templated text. For example, the tuple (\"chair\", \"left\", \"bed\") is converted to a string like \"The chair is to the left of the bed.\" These are then concatenated to form a multi-sentence referring expression for the object.", + "bbox": [ + 75, + 699, + 467, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Viewpoint Supervision. For a viewpoint $(x,y,\\theta)$ and referring expression $T_{o}$ , we provide viewpoint supervision decomposed into location and heading prediction. In most cases, the randomly selected viewpoint is not the only one for which the referring expression would be valid - small", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shifts in either location or perspective may not change the relationship between observed objects. For perspective, we produce a binary vector $\\mathcal{V}_{Per}$ with 36 entries where $\\mathcal{V}_{Per}[k]$ is 1 if the referring expression $T_{o}$ is valid from viewpoint $(x,y,k*10^{\\circ})$ . For location, we produce a $10\\times 10$ binary matrix $\\mathcal{V}_{Loc}$ where $\\mathcal{V}_{Per}[i,j]$ is 1 if the referring expression $T_{o}$ is valid from viewpoint $(i,j,\\theta)$ . For both, we check validity by repeating the spatial relationship tests described earlier from the new viewpoint. We note that this approach considers only a subset of all possible viewpoints and assumes that perspectives valid at $(x,y)$ are also valid for any point $(i,j)$ for which the original perspective was valid.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Viewpoint-Aware 3D Grounding Network", + "text_level": 1, + "bbox": [ + 498, + 280, + 854, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As outlined in Fig. 2, our ViewPoint Prediction Network (VPP-Net) approach consists of a multimodal encoder-decoder architecture. After encoding the referring expression and scene, we predict the location and perspective of the observer. The encoded features and transformed candidate bounding boxes and XYZ point positions are then passed to a decoding model which ranks the candidates.", + "bbox": [ + 496, + 303, + 890, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VPP-Net Backbone Architecture. Our method is based on BUTD-DETR [17]. We review its construction here.", + "bbox": [ + 498, + 414, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input Encoding. An RGBXYZ point cloud representing the 3D scene is encoded with a pretrained PointNet++ [29], producing $n_p$ point features $\\mathbf{e}_{\\text{vision}} \\in \\mathbb{R}^{n_p \\times d}$ and the corresponding XYZ points $\\mathbf{p} \\in \\mathbb{R}^{n_p \\times 3}$ . Referring expressions are encoded with a pre-trained RoBerta [22] model into a sequence of text embeddings $\\mathbf{e}_{\\text{text}} \\in \\mathbb{R}^{|T| \\times d}$ , where $|T|$ is the token length.", + "bbox": [ + 496, + 449, + 890, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Encoder. The multimodal encoder consists of $N$ layers of attention-based modules. Initializing visual features $f_{v}^{(0)}$ and textual features $f_{t}^{(0)}$ with $e_{vision}$ and $e_{text}$ respectively, each layer $i$ computes update representations by applying per-modality self attention followed by cross-modality attention as below:", + "bbox": [ + 496, + 560, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {v} ^ {(i + 1)}, \\mathbf {f} _ {t} ^ {(i + 1)} = f _ {\\text {C r o s s - A t t}} \\left(f _ {\\text {S e l f - A t t}} \\left(\\mathbf {f} _ {v} ^ {(i)}\\right), f _ {\\text {S e l f - A t t}} \\left(\\mathbf {f} _ {t} ^ {(i)}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 659, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the final layer outputs as $\\mathbf{f}_t^{(N)}$ and $\\mathbf{f}_v^{(N)}$ .", + "bbox": [ + 500, + 688, + 834, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Decoder. Between the encoder and decoder, the visual features $\\mathbf{f}_v^{(N)}$ are passed through as small MLP to produce scores for each visual input that are then passed through a softmax. The top-K scoring transformed features are retained as object queries $\\hat{\\mathbf{f}}$ . The object decoder is an N-layer transformer-based module. The decoder takes $\\hat{\\mathbf{f}}$ , $\\mathbf{f}_t^{(N)}$ , $\\mathbf{f}_v^{(N)}$ , hierarchical points $\\mathbf{c}_v$ and detected bounding boxes $\\mathbf{B}_o$ as input to predict $\\mathbf{r}$ , $\\mathbf{f}_o$ , $\\mathbf{f}_l$ , $\\mathbf{f}_s$ :", + "bbox": [ + 496, + 710, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r}, \\mathbf {f} _ {o}, \\mathbf {f} _ {l}, \\mathbf {f} _ {s} = f _ {\\mathrm {D e c}} \\left(\\hat {\\mathbf {f}}, \\mathbf {f} _ {t}, \\mathbf {f} _ {v}, \\mathbf {c} _ {v}, \\mathbf {B} _ {o}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 844, + 890, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{r} \\in \\mathbb{R}^{n_p \\times 1}$ is the object-expression matching score; $\\mathbf{f}_o \\in \\mathbb{R}^{n_p \\times D}$ are object features; and $\\mathbf{f}_l, \\mathbf{f}_s \\in \\mathbb{R}^{n_p \\times 3}$ are", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c329894d27a521b1efb9162494bcc562e7e89069af3231c8742914e1a09183a2.jpg", + "image_caption": [ + "Figure 2. An overview of our ViewPoint Predictor Network (VPP-Net) approach. (Left) Our model builds on the encoder-decoder structure of [17], but modifies the encoder to explicitly predict location and perspective of the observer. The decoder then processes appropriately transformed inputs to rank candidate bounding boxes. (Right) Our encoder model updates features of location and perspective at each layer that are supervised to predict viewpoints where the referring expression is valid." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 895, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the location and size of objects. The object with the highest $r_i$ is chosen and its corresponding $\\mathbf{f}_{l,i}$ and $\\mathbf{f}_{s,i}$ are used to compute the bounding box for the object reference.", + "bbox": [ + 76, + 424, + 468, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adding Viewpoint Prediction. As shown in Fig. 2, our approach extends this architecture in two ways – by modifying the encoder to enable location/perspective prediction and by using these predictions to transform bounding boxes $B$ and point coordinates $p$ prior to invoking the decoder.", + "bbox": [ + 76, + 487, + 468, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We add two transformer-based predictors into each multimodal encoder layer shown in Fig. 2 (right). As both perspective and location features are updated in the same fashion, we write the update steps below for location only:", + "bbox": [ + 75, + 563, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {l o c, v} ^ {(i)} = f _ {\\text {A t t n}} \\left(\\mathbf {f} _ {l o c, t} ^ {(i - 1)}, \\mathbf {f} _ {v} ^ {(i)}, \\mathbf {f} _ {v} ^ {(i)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 628, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {l o c, t} ^ {(i)} = f _ {\\text {A t t n}} \\left(\\mathbf {f} _ {l o c, v} ^ {(i)}, \\mathbf {f} _ {t} ^ {(i)}, \\mathbf {f} _ {t} ^ {(i)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 652, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{Attn}}(\\cdot ,\\cdot ,\\cdot)$ denotes multi-head attention with key, query, and value arguments. The location feature is updated each layer by querying visual features and then querying textual features. We initialize the perspective and location features $\\mathbf{f}_{per,t}^{(0)}$ with learnable embeddings $e_{per}$ and $\\mathbf{f}_{loc,t}^{(0)}$ with $e_{loc}$ - a linear mapping of viewpoint $(x,y)$ . Denote the final representations as $\\mathbf{f}_{per,t}^{(N)}$ and $\\mathbf{f}_{loc,t}^{(N)}$ .", + "bbox": [ + 75, + 678, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We predict a distribution over locations and perspective from each layer's location and perspective features respectively using a learned linear layer. We denote our predictions over locations and perspective respectively as", + "bbox": [ + 75, + 789, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (l o c \\mid \\mathbf {f} _ {l o c, m} ^ {(i)}) = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {l o c} \\mathbf {f} _ {l o c, m} ^ {(i)} + \\mathbf {b} _ {l o c}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 854, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (p e r \\mid \\mathbf {f} _ {p e r, m} ^ {(i)}) = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {l o c} \\mathbf {f} _ {p e r, m} ^ {(i)} + \\mathbf {b} _ {p e r}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 878, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m \\in \\{\\mathrm{v},\\mathrm{t}\\}$ . We supervise each of these predictions with synthetic viewpoint supervision using an averaged cross-entropy loss - denoting it as", + "bbox": [ + 498, + 424, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a, m} = C E \\left(\\mathcal {Y} _ {a}, P _ {i} (a | m)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 483, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a \\in \\{per, loc\\}$ and $\\mathcal{Y}_a$ as defined in Sec. 3.1.", + "bbox": [ + 498, + 511, + 841, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Decoder with Affine Transformation. After encoding the input, we take the highest probability location $(\\hat{x},\\hat{y})$ and perspective $\\hat{\\theta}$ predictions generated from $f_{loc,t}^{(N)}$ and $f_{per,t}^{(N)}$ as an estimate of the observers viewpoint. Given these, we apply the corresponding affine transform $M(\\hat{\\theta},\\hat{x},\\hat{y})$ to the point locations $p$ and bounding boxes $B-$ denoting the transformed versions as $p^{\\prime}$ and $B^{\\prime}$ respectively. These transformed versions along with the other features are then passed to the decoder as in Eq. 4. After the decoder predicts a bounding box location $\\mathbf{f}_l^{\\prime}$ and size $\\mathbf{f}_s^{\\prime}$ in this transformed coordinate system, we transform the prediction back to the original scene using the inverse of $\\mathbf{M}(\\hat{\\theta},\\hat{x},\\hat{y})$", + "bbox": [ + 496, + 532, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Encouraging Uniform Object Representation", + "text_level": 1, + "bbox": [ + 498, + 726, + 880, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An object can be described differently in different viewpoints - potentially causing our model's representation of the same object to be quite different from different viewpoints. To encourage our model to be more robust to these difference, we add an auxiliary loss that encourages the decoder output $f_{o}$ for object $o$ to be consistent across viewpoints. To do this, we introduce a set of instance-wise learnable vectors $\\mathbf{U} = \\{u_0, u_1, \\ldots\\}$ . We refer to these as uniform object representations. For a training sample whose target is $o$ , we concatenate the predicted object features $f_{o}$", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14059", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and final perspective embeddings $f_{per,t}^{(N)}$ to estimate the uniform object representation $u_{o}$ with a linear layer. The loss is computed between the predicted representation and $u_{o}$ :", + "bbox": [ + 76, + 88, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l 1} = | | \\tanh (\\mathbf {W} _ {f u s e} [ f _ {o}, f _ {p e r, t} ^ {(N)} ] + \\mathbf {b} _ {f u s e}) - u _ {o} | | _ {1} \\quad (1 0)\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 145, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This encourages representation to be predictable from object features and viewpoint information – intuitively, predicting an object instance's identity from its observation at a specific viewpoint. Note that these learnable vectors are not used at inference time.", + "bbox": [ + 75, + 174, + 468, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, we also use $f_{o}$ and $u_{o}$ to classify the object's class to encourage semantic consistency. For both $f_{o}$ and $u_{o}$ , we make a prediction over the object class and compute a Cross-Entropy loss:", + "bbox": [ + 76, + 250, + 468, + 310 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c l s} = C E \\left(\\mathcal {Y} _ {o}, \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {s e m} \\phi + \\mathbf {b} _ {s e m}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 319, + 468, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi \\in \\{f_o,u_o\\}$ and $\\mathcal{V}_o$ is the object's class.", + "bbox": [ + 76, + 345, + 393, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 76, + 368, + 235, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training loss is a weight sum of our location and perspective supervision loss $L_{i,a,m}$ , the uniform object representation loss $L_{l1}$ , the object classification loss $L_{cls}$ and the grounding loss $L_{gr}$ proposed in BUTD-DETR [17]:", + "bbox": [ + 76, + 392, + 468, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\alpha_ {1} \\sum_ {a, m} L _ {a, m} + \\alpha_ {2} L _ {l 1} + \\alpha_ {3} L _ {c l s} + \\alpha_ {4} L _ {g r} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 462, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the coefficients $\\alpha$ 's are hyperparameters. Additionally, we find two additional techniques to be useful.", + "bbox": [ + 76, + 503, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic Training Curriculum. While training on the synthetic data for viewpoint supervision, we apply a filtering mechanism on visual features $\\mathbf{f}_v^{(0)}$ that reduces the task complexity of early training examples. We retain all points (and associated features) that occur within the bounding box of an object mentioned in the referring expression. For points outside these boxes, we drop the point with probability $p_{drop}$ . At the start of training, we set $p_{drop} = 1 - \\text{effectively removing the background and other objects from the scene. As training progresses, we anneal } p_{drop} \\text{ to 0 such that whole scenes are observed.}$", + "bbox": [ + 75, + 537, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Viewpoint Data Augmentation. During both synthetic training and finetuning, we apply a viewpoint-based augmentation scheme. After running an example through our model, we retain the predicted affine transformation $M_0$ . We then rotate the scene's pointcloud and supervision by $M_0$ to create a viewpoint augmented sample. We then run the model again with this sample.", + "bbox": [ + 76, + 710, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 830, + 310, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To study the effectiveness of our proposed VPP-Net, we conduct experiments on three existing 3D visual grounding datasets - ScanRefer [5], SR3D [2], and NR3D [2].", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 90, + 808, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate on ScanRefer, SR3D, and NR3D which we describe below: ScanRefer. The ScanRefer dataset is built on 800 3D scenes collected from ScanNet. The 3D scenes are labeled with about 51,000 natural expressions and 11,000 objects. The labeled natural expressions have different types of phrases, including intra/inter-class spatial relations and comparatives/superlatives. The dataset also provides real images of the scene, which are used to enhance the grounding performance in some methods. We use the official splits provided by ScanRefer to train and evaluate our model.", + "bbox": [ + 496, + 118, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SR3D and NR3D. SR3D and NR3D are two datasets based on the ScanNet scenes proposed by Achlioptas et al. [2]. SR3D consists of 83K expressions generated by a template-based text generator. The expressions contain five types of spatial object-to-object relations: Horizontal/Vertical Proximity, Between, Allocentric, and Support, taking up $81.37\\%$ , $3.80\\%$ , $1.79\\%$ , $4.50\\%$ , and $8.54\\%$ separately of all expressions. NR3D is a 3D visual grounding dataset with human-annotated expressions similar to ScanRefer. It consists of 41.5K human-annotated expressions and distractors collected through a \"player-listener\" game. Note that both SR3D and NR3D provide ground-truth bounding boxes for all candidate objects, unlike ScanRefer.", + "bbox": [ + 496, + 295, + 890, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Settings. We evaluate our proposed VPP-Net on these datasets and report results in Tables 1 and 2. For ScanRefer, we report the top-1 accuracy when IoU is larger than 0.25 and 0.5 in \"Unique\", \"Multi\" and \"Overall\" cases. \"Unique\" vs \"Multi\" indicate whether distractors exist. \"Modality\" shows whether a method gets benefit from 2D images, with \"3D\" meaning the model only takes the 3D point cloud and \"3D+2D\" indicating that 2D images are also used. For SR3D and NR3D, following the experiment setting of EDA, we report Acc@0.25IoU.", + "bbox": [ + 496, + 497, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 659, + 715, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic data. We select 557 and 140 scenes from ScanNet to build the train and validation set of synthetic datasets. With the proposed synthetic viewpoint supervision mechanism, we repeat the generation process 75 times for each scene to generate the synthetic dataset – resulting in 36117 training and 9167 validation samples.", + "bbox": [ + 496, + 683, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model configuration. VPP-Net contains a 6-layer encoder and decoder. The learning rate of for the PointNet++ and Roberta is $2e^{-3}$ . The learning rate of other modules is $2e^{-4}$ . We first pre-train the model on the synthetic dataset to learn the viewpoint estimation for 150 epochs. Then we fine-tune using a combination of synthetic and the downstream dataset (Scanrefer/NR3D/SR3D) for another 150 epochs. We apply the point dropout mechanism on the synthetic", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14060", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/79d420a0eaca8ef18c1cc14a8d7a84d1f82289e1437311dac5ef06469b9359d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalityUnique(%)Multi(%)Overall(%)
Acc@0.25Acc@0.50Acc@0.25Acc@0.50Acc@0.25Acc@0.50
ScanRefer [5]2+3D76.3353.5132.7321.1141.1927.40
ReferIt3D [2]3D53.837.521.012.826.416.9
TGNN [15]3D68.6156.8029.8423.1837.3729.70
InstanceRefer [44]3D77.4566.8331.2724.7740.2332.93
SAT [43]2+3D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]3D78.8067.9435.1925.7041.3334.01
3DVG-Transformer [46]2+3D81.9360.6439.3028.4247.5734.67
3D-SPS [26]2+3D84.1266.7240.3229.8248.8236.98
3DJCG [4]2+3D83.4764.3441.3930.8249.5637.33
D3Net [6]3D+2D-70.35-30.50-37.87
UniT3D [8]3D+2D82.7573.1436.3631.0545.2739.14
BUTD-BasedBUTD-DETR [17]3D84.266.346.635.152.239.8
EDA [35]3D85.7668.5749.1337.6454.5942.26
Multi-ViewM3DRef-CLIP [45]3D+2D85.377.243.836.851.944.7
MVT [16]3D+2D--31.4624.8539.9532.28
ViewRefer [13]3D+2D--33.0826.5041.3033.66
VPP-Net (Ours)3D86.0567.0950.3239.0355.6543.29
", + "bbox": [ + 81, + 88, + 893, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. 3D visual grounding results on ScanRefer. Acc@0.25/0.5 means top-1 accuracy when IoU is larger than 0.25/0.5 - higher is better. We group BUTD-Based and Multi-View methods for ease of comparison. BUTD-based use the same backbone of our model. The Multi-View methods all explored notions of viewpoint or 2D image rendering in 3D visual grounding. The state-of-the-art methods are bold and the second best performance is underlined in the table.", + "bbox": [ + 75, + 421, + 892, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data in both the pre-training and finetuning processes. The dropout rate $p_{drop}$ is kept at 1 in the first 40 epochs then decreases to 0 linearly in the following 70 epochs.", + "bbox": [ + 75, + 503, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bounding boxes. Bounding boxes $\\mathbf{B}$ are a required input for the decoder. In ScanRefer, no ground-truth object boundary box are provided. Thus we use GroupFree [25] to detect a set of bounding boxes $\\mathbf{B}_d$ , which $\\mathbf{B}_d \\in \\mathbb{R}^{m \\times 8 \\times 3}$ . Following the experiment setting of EDA [35] and BUTDDETR [17], in SR3D and NR3D, the model takes the ground-truth boundary boxes to replace the detected ones as the input of the decoder.", + "bbox": [ + 75, + 555, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. 3D Visual Grounding Results", + "text_level": 1, + "bbox": [ + 76, + 691, + 336, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of ScanRefer Results. Our experimental results on ScanRefer are shown in Table 1. VPP-Net achieves state-of-the-art performance in terms of Acc@0.25 of \"Overall\" and \"Unique\", surpassing EDA by $1.06\\%$ and $0.29\\%$ respectively. An improvement of $1.19\\%$ and $1.29\\%$ can be observed at \"Multi\" cases to $50.32\\%$ and $39.03\\%$ separately. The \"Multi\" cases are more challenging than \"Unique\" cases generally because of the existence of distractors. This suggests that the 3D scene transformed with the predicted viewpoint provides less ambiguous spatial relations with which to reason about distractors.", + "bbox": [ + 75, + 715, + 468, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Among the listed methods, we draw attention to \"BUTD", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based\" and \"Multi-View\" models in the table - BUTD-DETR [17], EDA [35], M3DRef-CLIP [45], MVT [16], and ViewRefer [13]. BUTD-based models leverage the same BUTD-DETR backbone we build on. Compared with BUTD-DETR, our proposed VVP-Net has a large margin of improvement across all evaluation metrics. Compared to EDA, we can still observe an improvement of around $1.04\\%$ and $1.39\\%$ in \"Overall\" and \"Multi\". Multi-View models involve multi-view scenes or images that are related to our method. VPP-Net outperforms MVT and ViewRefer across all metrics. Compared with M3DRef-CLIP, our model achieves comparable but less competitive performance in Acc@0.5 of \"Unique\" and Acc@0.5 of \"Overall\", while achieving better performance in all other cases, especially, the Acc@0.5 of \"Multi\" which is more challenging. We hypothesize this is due to the multi-view images rendered of object candidates used in M3DRef-CLIP. Multi-view rendered images provide detailed information about an object, allowing the model localize the object more accurately. The multi-view images focus on single objects and do not provide much benefit in the \"Multi\" case, since the model can be fooled by multiple objects with similar 2D visual attributes in a scene. Thus, we can observe a larger margin of Acc@0.5 of \"Multi\" between C3DRef-CLIP and ours (improved by $+2.23\\%$ ).", + "bbox": [ + 496, + 503, + 890, + 881 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14061", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/228438ec7006f885d5df16de09a7edef91f4a60b5b5d766715586b9988454db4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalitySR3DNR3D
ReferIt3D [2]3D39.835.6
TGNN [15]3D45.037.3
TransRefer3D [14]3D57.442.1
InstanceRefer [44]3D48.038.8
3DVG-Transfor [46]3D51.440.8
FFL-3DOG [12]3D-41.7
SAT [43]3D+2D57.949.2
3DReferTrans [1]3D47.039.0
LanguageRefer [32]3D56.043.9
3D-SPS [26]3D+2D62.651.5
BUTD-basedEDA [35]3D68.152.1
BUTD-DETR [17]3D67.154.9
Multi ViewM3DRef-CLIP [45]3D+2D-49.4
LAR [3]3D59.648.9
VPP-Net(Ours)3D68.756.9
", + "bbox": [ + 81, + 87, + 472, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of SR3D & NR3D Results. We report results for SR3D and NR3D datasets in Table 2. Compared with existing methods, VPP-Net achieves state-of-the-art on both. Our observed improvement on NR3D (+2.00%) is larger than we see in SR3D (0.6%). We argue that this may be because there is a large domain gap between SR3D and our synthetic dataset. To show the domain gap, we count the frequency of spatial-relation words in SR3D, NR3D, ScanRefer, and our synthetic dataset in Table 4. We find an obvious gap: proximity relations - \"Closest/Farthest\" - dominate in SR3D, while relative spatial relations - \"Front/Behind\" and \"Left/Right\" - are more prominent in the human-annotated NR3D and ScanRefer datasets as well as our synthetic dataset. Despite this, VPP-Net still boosts BUTD-DETR performance on SR3D (+1.6%).", + "bbox": [ + 75, + 469, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explicitly evaluate the efficiency of VPP-Net on viewpoint-dependent samples, we report results on view dependent / independent (dep./indep.) subsets of NR3D in Table 3. We see large gains over our backbone Butd-Detr for view dep. cases (46 vs. $52.4\\%$ ) while achieving a similar result in view indep. case (58 vs. $58.6\\%$ ). In contrast, EDA's improvement for view dependent cases is smaller and results in performance loss on view independent cases.", + "bbox": [ + 75, + 696, + 468, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 830, + 232, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report an ablation study of our method on ScanRefer in Table 5. We remove different modules, auxiliary losses, or training procedures from our approach to show their contri", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74b7aee89b1ef4426f9df46e49bdaf4ddc089f5ed9605ca43032baa65b6c1fb6.jpg", + "table_caption": [ + "Table 2. Result of SR3D and NR3D datasets. We report Acc@0.25IoU as the evaluation metric. Best and 2nd best results are denoted in bold or underlined." + ], + "table_footnote": [], + "table_body": "
ModelView Dep.View Indep.
Butd-Detr [17]46.0%58.0%
EDA [31]50.2 %53.1%
VPPNet (ours)52.4 %58.6%
", + "bbox": [ + 519, + 88, + 870, + 165 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1a448812f42e1a5626c3d0bcd16fe32c1a71d8f8fe4c52f6233e4c262a6416d0.jpg", + "table_caption": [ + "Table 3. Results of Acc@0.25 on View Dep/Indep-endent cases of NR3D dataset." + ], + "table_footnote": [], + "table_body": "
RelationSynthetic(%)Natural(%)
SR3DOursNR3DScanRefer
Left/Right8.9244.5836.0055.66
Front/Behind2.5750.0819.3920.11
Above/Under3.815.3414.1913.09
Between8.540.003.216.77
Closest/Farthest81.380.0031.214.38
", + "bbox": [ + 504, + 224, + 893, + 333 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/934afe934a6b6f24e76e37832e3efbda2465e90c042ca0ec005e24ed2aebc3e6.jpg", + "table_caption": [ + "Table 4. Normalized frequency of spatial relation words across datasets. We find SR3D has significantly fewer relative spatial relations than our synthetic datasets or either human-collected datasets on which we evaluate." + ], + "table_footnote": [], + "table_body": "
Synthetic PretrainingViewpoint PredictionUniform Obj. Rep.Curriculum FilteringViewpoint Data Aug.Acc@0.25Acc@0.50
0*-----52.139.8
1--✓✓-51.738.9
2✓✓-✓✓53.7838.98
3✓✓✓-✓51.5238.45
4✓✓✓✓-53.7840.41
5-✓✓✓✓49.2333.67
6✓✓✓✓✓55.6543.29
", + "bbox": [ + 506, + 428, + 890, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation of our model components, auxiliary losses, and training methods on the ScanRefer dataset - checkmarks denote the corresponding item is active. Rows are numbered.", + "bbox": [ + 498, + 520, + 892, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bution. Note that row \"0*\" without any of our modifications is equivalent to BUTD-DETR. When viewpoint prediction is disabled (row 1), we also disable the modules that rely on it – pretraining and viewpoint data augmentation.", + "bbox": [ + 496, + 604, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We find that removing the viewpoint prediction and data augmentation (row 1), or the curriculum filtering during pretraining (row 3) results in performance similar to the baseline. These correspond to models that either do not use viewpoint prediction (row 1) or learn viewpoint prediction poorly (row 3) – suggesting the importance of viewpoint in 3D visual grounding. We also note that directly jointly training VPP-Net without first developing a strong viewpoint prediction capability during pretraining (row 5) yields worse results than not considering viewpoint at all. Removing either the uniform object representation loss (row 2) or the viewpoint data augmentation (row 4) yields degraded performance compared with the whole model (row 6). Both are designed to improve learned representations and we see they have a positive effect in these experiments.", + "bbox": [ + 496, + 674, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14062", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39a706bbac73097ff7e6a588be3bbfed38be5544414382f3c7a0537d05566c88.jpg", + "image_caption": [ + "Figure 3. Example VPP-Net results. The first row shows successful examples where both viewpoint prediction and visual grounding are correct. The second row shows failed samples. The first two predict incorrect viewpoints but still succeed in visual grounding. The predicted observer position (red dot) and facing direction (facing away from the green arrow) are shown in 3D and a top-down view (top right corner). We zoom up on the target region in example (7) within the red circle to provide better visualization." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 893, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Examples", + "text_level": 1, + "bbox": [ + 76, + 497, + 279, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We depict the quantitative results of VPP-Net in Fig. 3. In the figure, we list 8 examples of which four are successful (top row) and four are not (bottom row). In each example, the red dot represents the predicted location and the green and blue arrows represent the \"behind\" and \"right\" direction predicted in the model such that the predicted observer is facing away from the green arrow. The ground truth bounding boxes and target words are noted with green and the mentioned objects are noted with red. We also provide the predicted object bounding box in the image, shown in blue. The spatial relations are noted with blue in the text. The successful examples show that with an accurate viewpoint prediction, the expression can better match the 3D scene, resulting in more accurate groundings. Failed examples (5) and (6) are successful at grounding despite having failed to predict valid viewpoints. Both contain ambiguous elements in text or image, i.e. the expression \"walk thru the door on the right\" in example (5) and the distorted door (blob on bottom left of the scene) in example (6). Examples (7) and (8) predict valid viewpoints but produce incorrect visual groundings. The predicted bounding box in example (7) covers a certain region of the ground truth and involved patch contains similar visual attributes (white). In example (8), both ground truth and prediction are on the left of the viewpoint, matching the expression. However,", + "bbox": [ + 75, + 523, + 473, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model failed to understand the horizontal proximity (farthest), leading to a incorrect grounding.", + "bbox": [ + 498, + 497, + 890, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 568, + 619, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we have proposed a Viewpoint Prediction Network (VPP-Net) for 3D visual grounding tasks. Our model addresses the critical challenge of viewpoint-dependent ambiguity in 3D visual grounding. Specifically, we first introduce a synthetic dataset and then use it to train a 3D visual grounding model that can better disambiguate referring expressions by explicitly predicting potential viewpoints. Further, we design a uniform object representation auxiliary loss and viewpoint data augmentation scheme that further improve the performance of our model. Experiments on the ScanRefer, SR3D, and NR3D demonstrate the effectiveness of our proposed methods.", + "bbox": [ + 496, + 602, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The synthetic datasets we generate are highly dependent on the templated referring expressions, which limits their diversity and realism. Future works should include strategies to generate more diverse datasets or collect human-generated referring expressions with annotated viewpoints to further improve.", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14063", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dreftransformer: Fine-grained object identification in realworld scenes using natural language. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 7", + "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 422-440. Springer, 2020. 2, 5, 6, 7", + "[3] Eslam Bakri, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 7", + "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 6", + "[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 5, 6", + "[6] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. 2021. 6", + "[7] Sijia Chen and Baochun Li. Multi-modal dynamic graph transformer for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15534-15543, 2022. 1, 2", + "[8] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 6", + "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 3", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[11] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 202-210. Springer, 2019. 2", + "[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1, 6, 7", + "[13] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 1, 2, 6", + "[14] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 7", + "[15] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 6, 7", + "[16] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1, 2, 6", + "[17] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, pages 417-433. Springer, 2022. 1, 2, 3, 4, 5, 6, 7", + "[18] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1, 2", + "[19] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2", + "[20] Weide Liu, Zhonghua Wu, Henghui Ding, Fayao Liu, Jie Lin, and Guosheng Lin. Few-shot segmentation with global and local contrastive learning. arXiv preprint arXiv:2108.05293, 2021. 2", + "[21] Weide Liu, Zhonghua Wu, Yang Zhao, Yuming Fang, Chuan-Sheng Foo, Jun Cheng, and Guosheng Lin. Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation. International Journal of Computer Vision, pages 1-15, 2023. 2", + "[22] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3", + "[23] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 2", + "[24] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14064", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5612-5621, 2021. 1, 2", + "[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2, 6", + "[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 6, 7", + "[27] Zhipeng Luo, Gongjie Zhang, Changqing Zhou, Zhonghua Wu, Qingyi Tao, Lewei Lu, and Shijian Lu. Modeling continuous motion for 3d point cloud object tracking. arXiv preprint arXiv:2303.07605, 2023. 2", + "[28] Hui En Pang, Zhongang Cai, Lei Yang, Qingyi Tao, Zhonghua Wu, Tianwei Zhang, and Ziwei Liu. Towards robust and expressive whole-body human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024.", + "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[30] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[32] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7", + "[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 2", + "[34] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 1, 2", + "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 1, 6, 7", + "[36] Yu Wu, Yana Wei, Haozhe Wang, Yongfei Liu, Sibei Yang, and Xuming He. Grounded image text matching with mismatched relation reasoning. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 2976-2987, 2023. 1, 2", + "[37] Zhonghua Wu, Guosheng Lin, and Jianfei Cai. Keypoint based weakly supervised human parsing. Image and Vision Computing, 91:103801, 2019. 2", + "[38] Zhonghua Wu, Qingyi Tao, Guosheng Lin, and Jianfei Cai. Exploring bottom-up and top-down cues with attentive learning for webly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12936-12945, 2020.", + "[39] Zhonghua Wu, Xiangxi Shi, Guosheng Lin, and Jianfei Cai. Learning meta-class memory for few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 517-526, 2021. 2", + "[40] Zhonghua Wu, Yicheng Wu, Guosheng Lin, Jianfei Cai, and Chen Qian. Dual adaptive transformations for weakly supervised point cloud segmentation. In European conference on computer vision, pages 78-96. Springer Nature Switzerland Cham, 2022. 2", + "[41] Zhonghua Wu, Yicheng Wu, Guosheng Lin, and Jianfei Cai. Reliability-adaptive consistency regularization for weakly-supervised point cloud segmentation. International Journal of Computer Vision, pages 1-14, 2024. 2", + "[42] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 2", + "[43] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2, 6, 7", + "[44] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 6, 7", + "[45] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 6, 7", + "[46] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 1, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 767 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_model.json b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4cf169dc83520a8e46a28067face5337d5ccaa70 --- /dev/null +++ b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.238, + 0.131, + 0.734, + 0.154 + ], + "angle": 0, + "content": "Viewpoint-Aware Visual Grounding in 3D Scenes" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.182, + 0.327, + 0.234 + ], + "angle": 0, + "content": "Xiangxi Shi \nOregon State University \nshixia@oregonstate.edu" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.182, + 0.586, + 0.234 + ], + "angle": 0, + "content": "Zhonghua Wu SenseTime Research wuzhonghua@sensetime.com" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.182, + 0.834, + 0.234 + ], + "angle": 0, + "content": "Stefan Lee \nOregon State University \nleestef@oregonstate.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.604 + ], + "angle": 0, + "content": "Referring expressions for visual objects often include descriptions of relative spatial arrangements to other objects - e.g. \"to the right of\" - that depend on the point of view of the speaker. In 2D referring expression tasks, this viewpoint is captured unambiguously in the image. However, grounding expressions with such spatial language in 3D without viewpoint annotations can be ambiguous. In this paper, we investigate the significance of viewpoint information in 3D visual grounding - introducing a model that explicitly predicts the speaker's viewpoint based on the referring expression and scene. We pretrain this model on a synthetically generated dataset that provides viewpoint annotations and then finetune on 3D referring expression datasets. Further, we introduce an auxiliary uniform object representation loss to encourage viewpoint invariance in learned object representations. We find that our proposed ViewPoint Prediction Network (VPP-Net) achieves state-of-the-art performance on ScanRefer, SR3D, and NR3D - improving Accuracy@0.25IoU by \\(1.06\\%\\), \\(0.60\\%\\), and \\(2.00\\%\\) respectively compared to prior work." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.633, + 0.21, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Visual grounding of referring expressions requires algorithms to reason about natural language object descriptions to identify object referents in visual scenes. Naturally, describing relative spatial relations between objects is a common strategy when a speaker produces a referring expression – e.g., “The chair to the right of the couch.” As such, many methods for visual grounding put algorithmic emphasis on handling these spatial relations [7, 18, 24, 34, 36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "However, these sorts of spatial relations depend upon the point of view of the speaker. For instance, Fig. 1 depicts two different viewpoints of the same 3D scene and notes that these views produce contradictory referring expressions. In 2D visual grounding tasks, the point of view of the speaker is typically also that of the image – introducing no ambiguity into spatial reasoning. However, viewpoint annotations are not typically captured for 3D visual grounding tasks that" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.271, + 0.691, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.27, + 0.875, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.449, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Expressions referring to the same object from different viewpoints can lead to contradictory spatial relations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.597 + ], + "angle": 0, + "content": "reason about whole scenes. As a result, identifying referent objects may be ambiguous if multiple object instances align with the referring expression in differing views. Moreover, the lack of reliable grounding for spatial language may simply make the reasoning task more difficult to learn even when the referent is still uniquely identifiable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.718 + ], + "angle": 0, + "content": "Despite this, most of the recent work in 3D visual grounding has not addressed the role of viewpoint in the learning process [12, 17, 26, 35, 43, 46]. Those methods that have attempted to consider the influence of viewpoint have been hindered by the lack of viewpoint supervision in existing datasets – instead relying on rendering multiple views of a 3D scene and using attention mechanisms to softly select between them [13, 16] in an end-to-end model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this work, we develop a straightforward approach to viewpoint-aware visual grounding in 3D scenes – explicitly predicting the speaker's viewpoint and then transforming the input accordingly. To support this direction, we create a synthetic visual grounding dataset that provides annotations to directly supervise viewpoint prediction. We propose a novel framework called ViewPoint Prediction Net (VPP-Net) to estimate the viewpoints of expression-pointcloud pairs and rotate the 3D scenes accordingly before performing expression grounding. This model is first pre-trained on the synthetic dataset and then fine-tuned on a combination of the synthetic and target datasets. To our knowledge, this" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14056" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "is the first work to introduce a supervised viewpoint estimation model for 3D visual grounding tasks. In experiments on three common 3D visual grounding datasets, this approach paired with viewpoint-related auxiliary losses leads to improved performance over prior art." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.168, + 0.457, + 0.182 + ], + "angle": 0, + "content": "Contributions. Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.185, + 0.468, + 0.228 + ], + "angle": 0, + "content": "- We introduce viewpoint prediction as an auxiliary task in the 3D visual grounding task and propose synthetic data generation pipeline to provide viewpoint supervision." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.23, + 0.468, + 0.274 + ], + "angle": 0, + "content": "- We propose VPP-Net - a novel model that learns viewpoint prediction to transform 3D scenes to reduce ambiguity in spatial-relation grounding." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.275, + 0.468, + 0.304 + ], + "angle": 0, + "content": "- We design a Uniform Object Representation auxiliary loss that promotes viewpoint invariant feature learning." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.305, + 0.468, + 0.349 + ], + "angle": 0, + "content": "- Combing these techniques, our proposed approach achieves state-of-the-art performance on the ScanRefer [5], SR3D [2], and NR3D [2] datasets." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.185, + 0.468, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.362, + 0.218, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.468, + 0.719 + ], + "angle": 0, + "content": "Relation-aware 2D Visual Grounding. The visual grounding task has been extensively explored in 2D image scenarios. The goal of a 2D visual grounding model is to locate an object according to a given expression in a 2D image. In recent work, the relationships between objects have been identified as a crucial cue in complex environments. Several works [7, 23, 24, 34, 42] have parsed referring expressions into phrases for better text understanding. Wang et al. [34] proposed to align the generated visual graph with the self-attend object and relation features using a graph attention module. Liu [24] proposed to capture the visual context from a set of coarse-level object proposals by reconstructing the corresponding relations captured from the text. Chen [7] et al. initialize a visual graph with proposed node and edge transformers to learn the representations of objects and relations. Besides explicitly encoding visual relations in the model, relationships can also benefit models by constructing pseudo data [18, 36]. Wu et al. [36] proposed to detect mismatched relations from synthetic data to evaluate the ability of relation understanding of the model. Jiang et.al. [18] introduce an automatic way to generate pseudo visual grounding data for supervised training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "3D Visual Grounding. The first dataset for visual grounding in 3D was introduced by Chen et al. [5] - presenting the ScanRefer dataset consisting of 3D scenes and natural expression. Achlioptas et al. [2] propose two datasets - the synthetic dataset SR3D and the human-annotated NR3D. Further, they show that template-based synthetic data can benefit the task when models are jointly trained with naturally collected data. In standard approaches, a 3D grounding model first encodes instructions and 3D scene with pretrained text [10, 22, 31] and 3D visual [19, 27-30, 40, 41] models. Then a multimodal module is proposed to fuse the encoded features to estimate the location and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "size of the grounded object. 3D-SPS [26] proposed a one-stage method that estimates the location of objects from the point cloud directly. In other works, 3D object detections [11, 19, 25, 33] are leveraged to provide an object-focused candidate pool consisting of object-level features and/or bounding boxes. BUTD-DETR [17] encodes object bounding boxes as a reference and jointly trains on object detection datasets as an auxiliary task. 2D images [20, 21, 37-39] are also utilized to boost the performance of 3D visual grounding. The model proposed in [5] encodes 2D images together with 3D scenes for a better 3D understanding. Likewise, Yang et al. [43] leverage 2D semantics to assist 3D representation learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.471 + ], + "angle": 0, + "content": "The viewpoint of the annotator is also a significant cue for a better alignment between 3D scenes and referring expressions. Currently, only a few works focused on this aspect have been proposed. MVT [16] and ViewRefer [13] rotate the 3D scenes in different angles and encode them to multi-view representations without explicit supervision of which viewpoint matches to the annotator. Compared with previous work, we first provide a template-based synthetic visual grounding dataset including location and perspective supervision and then propose a supervised training method for viewpoint estimation. We demonstrate this is be beneficial for visual grounding performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.598, + 0.505 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.757 + ], + "angle": 0, + "content": "We propose a straight-forward approach for learning to reason about viewpoint in 3D referring expression grounding – directly predicting a rotation and translation of the 3D environment that aligns it with the assumed viewpoint of the observer providing the referring expression. To accomplish this, we introduce a ViewPoint Prediction Network (VPP-Net) – a model that performs 3D referring expression grounding while explicitly estimating the location and heading of the observer. As existing datasets lack observer viewpoint annotations, we develop a synthetic data generation pipeline to provide appropriate training supervision for viewpoint-aware pretraining and then transfer the pretrained model to downstream tasks. Further, we introduce a Uniform Object Representation (UOR) auxiliary loss to ensure object instances are represented consistently regardless of viewpoint. We describe each component below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.877, + 0.788 + ], + "angle": 0, + "content": "3.1. Generating Synthetic Viewpoint Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our synthetic data generation pipeline operates in four stages - for a given environment, we (1) sample a location and heading for a virtual observer, (2) identify spatial relationships between annotated objects in view, (3) generate templated referring expressions accordingly, and then (4) find other nearby viewpoints that could also have produced this reference. The result is a tuple of the 3D environment," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14057" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "referring expression, and valid viewpoint map which we can use for training viewpoint prediction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.468, + 0.218 + ], + "angle": 0, + "content": "Input Scenes. We utilize 3D scenes from the ScanNet [9] dataset which are annotated with detected object bounding boxes. For a given scene, we denote bounding box as a set \\(B\\), their centroids as a set of homogenous XYZ-coordinates \\(\\mathbf{C} = \\{c_0, c_1, \\dots, c_n \\mid c_i = (x_i, y_i, z_i, 1)\\}\\), and the corresponding object category labels \\(\\mathcal{Y}_o\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.468, + 0.298 + ], + "angle": 0, + "content": "Viewpoint Sampling. Regardless of scene size, we split the horizontal (XY) plane into a \\(10 \\times 10\\) grid and uniformly sample a 2D \\((x,y)\\) translation from the grid. Likewise, we sample a rotation \\(\\theta\\) about the vertical axis by randomly selecting a \\(10^{\\circ}\\) increment between \\(0^{\\circ}\\) and \\(360^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.348 + ], + "angle": 0, + "content": "Identifying Spatial Relationships. We apply the sampled translation and rotation to the scene by computing updated bounding box centers. With the affine transformation" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.358, + 0.469, + 0.422 + ], + "angle": 0, + "content": "\\[\nM (\\theta , x, y) = \\left( \\begin{array}{c c c c} \\cos \\theta & - \\sin \\theta & 0 & - x \\\\ \\sin \\theta & \\cos \\theta & 0 & - y \\\\ 0 & 0 & 1 & 0 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.468, + 0.552 + ], + "angle": 0, + "content": "we update the bounding box centroids as \\( c_i' = c_i M(\\theta, x, y)^T \\). The effect of this operation is to align the observer's relative spatial terms (e.g. left/right) to the XYZ axes. To detect potential spatial relations between a pair of objects \\( o_i \\) and \\( o_j \\), we compute the vector \\( r_{ij} = c_i' - c_j' \\) to represent their relative position. We take a set of four unit direction vectors as \\( \\{d_{front}, d_{back}, d_{left}, d_{right}\\} \\) and compute the cosine similarities between \\( d_k \\) and \\( r_{ij} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.562, + 0.469, + 0.595 + ], + "angle": 0, + "content": "\\[\ns _ {i, j, k} = \\frac {r _ {i j} d _ {k}}{\\left| \\left| r _ {i j} \\right| \\right| \\times \\left| \\left| d _ {k} \\right| \\right|}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.468, + 0.694 + ], + "angle": 0, + "content": "We consider a spatial relation \\(k\\) to be valid if \\(s_{i,j,k} > 0.3\\). For above and below relations, we instead consider the heights and overlap of the bounding boxes. From all valid tuples, we select a random subset as the relation tuple set \\(\\mathbf{R}\\). We note that this process does not consider the \"visibility\" of objects from the observer location." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.468, + 0.821 + ], + "angle": 0, + "content": "Expression Generation. Given R and object category annotations, we can produce a set of object referring expressions. For each object mentioned as part of a relation in R, we randomly sample up to four valid relations and convert each into simple templated text. For example, the tuple (\"chair\", \"left\", \"bed\") is converted to a string like \"The chair is to the left of the bed.\" These are then concatenated to form a multi-sentence referring expression for the object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Viewpoint Supervision. For a viewpoint \\((x,y,\\theta)\\) and referring expression \\(T_{o}\\), we provide viewpoint supervision decomposed into location and heading prediction. In most cases, the randomly selected viewpoint is not the only one for which the referring expression would be valid - small" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "shifts in either location or perspective may not change the relationship between observed objects. For perspective, we produce a binary vector \\(\\mathcal{V}_{Per}\\) with 36 entries where \\(\\mathcal{V}_{Per}[k]\\) is 1 if the referring expression \\(T_{o}\\) is valid from viewpoint \\((x,y,k*10^{\\circ})\\). For location, we produce a \\(10\\times 10\\) binary matrix \\(\\mathcal{V}_{Loc}\\) where \\(\\mathcal{V}_{Per}[i,j]\\) is 1 if the referring expression \\(T_{o}\\) is valid from viewpoint \\((i,j,\\theta)\\). For both, we check validity by repeating the spatial relationship tests described earlier from the new viewpoint. We note that this approach considers only a subset of all possible viewpoints and assumes that perspectives valid at \\((x,y)\\) are also valid for any point \\((i,j)\\) for which the original perspective was valid." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.281, + 0.855, + 0.297 + ], + "angle": 0, + "content": "3.2. Viewpoint-Aware 3D Grounding Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.41 + ], + "angle": 0, + "content": "As outlined in Fig. 2, our ViewPoint Prediction Network (VPP-Net) approach consists of a multimodal encoder-decoder architecture. After encoding the referring expression and scene, we predict the location and perspective of the observer. The encoded features and transformed candidate bounding boxes and XYZ point positions are then passed to a decoding model which ranks the candidates." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.415, + 0.892, + 0.445 + ], + "angle": 0, + "content": "VPP-Net Backbone Architecture. Our method is based on BUTD-DETR [17]. We review its construction here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Input Encoding. An RGBXYZ point cloud representing the 3D scene is encoded with a pretrained PointNet++ [29], producing \\( n_p \\) point features \\( \\mathbf{e}_{\\text{vision}} \\in \\mathbb{R}^{n_p \\times d} \\) and the corresponding XYZ points \\( \\mathbf{p} \\in \\mathbb{R}^{n_p \\times 3} \\). Referring expressions are encoded with a pre-trained RoBerta [22] model into a sequence of text embeddings \\( \\mathbf{e}_{\\text{text}} \\in \\mathbb{R}^{|T| \\times d} \\), where \\( |T| \\) is the token length." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Multimodal Encoder. The multimodal encoder consists of \\(N\\) layers of attention-based modules. Initializing visual features \\(f_{v}^{(0)}\\) and textual features \\(f_{t}^{(0)}\\) with \\(e_{vision}\\) and \\(e_{text}\\) respectively, each layer \\(i\\) computes update representations by applying per-modality self attention followed by cross-modality attention as below:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.66, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {v} ^ {(i + 1)}, \\mathbf {f} _ {t} ^ {(i + 1)} = f _ {\\text {C r o s s - A t t}} \\left(f _ {\\text {S e l f - A t t}} \\left(\\mathbf {f} _ {v} ^ {(i)}\\right), f _ {\\text {S e l f - A t t}} \\left(\\mathbf {f} _ {t} ^ {(i)}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.689, + 0.836, + 0.708 + ], + "angle": 0, + "content": "We denote the final layer outputs as \\(\\mathbf{f}_t^{(N)}\\) and \\(\\mathbf{f}_v^{(N)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Object Decoder. Between the encoder and decoder, the visual features \\(\\mathbf{f}_v^{(N)}\\) are passed through as small MLP to produce scores for each visual input that are then passed through a softmax. The top-K scoring transformed features are retained as object queries \\(\\hat{\\mathbf{f}}\\). The object decoder is an N-layer transformer-based module. The decoder takes \\(\\hat{\\mathbf{f}}\\), \\(\\mathbf{f}_t^{(N)}\\), \\(\\mathbf{f}_v^{(N)}\\), hierarchical points \\(\\mathbf{c}_v\\) and detected bounding boxes \\(\\mathbf{B}_o\\) as input to predict \\(\\mathbf{r}\\), \\(\\mathbf{f}_o\\), \\(\\mathbf{f}_l\\), \\(\\mathbf{f}_s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.845, + 0.891, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r}, \\mathbf {f} _ {o}, \\mathbf {f} _ {l}, \\mathbf {f} _ {s} = f _ {\\mathrm {D e c}} \\left(\\hat {\\mathbf {f}}, \\mathbf {f} _ {t}, \\mathbf {f} _ {v}, \\mathbf {c} _ {v}, \\mathbf {B} _ {o}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{r} \\in \\mathbb{R}^{n_p \\times 1}\\) is the object-expression matching score; \\(\\mathbf{f}_o \\in \\mathbb{R}^{n_p \\times D}\\) are object features; and \\(\\mathbf{f}_l, \\mathbf{f}_s \\in \\mathbb{R}^{n_p \\times 3}\\) are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14058" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.096, + 0.896, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.894, + 0.4 + ], + "angle": 0, + "content": "Figure 2. An overview of our ViewPoint Predictor Network (VPP-Net) approach. (Left) Our model builds on the encoder-decoder structure of [17], but modifies the encoder to explicitly predict location and perspective of the observer. The decoder then processes appropriately transformed inputs to rank candidate bounding boxes. (Right) Our encoder model updates features of location and perspective at each layer that are supervised to predict viewpoints where the referring expression is valid." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.425, + 0.47, + 0.471 + ], + "angle": 0, + "content": "the location and size of objects. The object with the highest \\( r_i \\) is chosen and its corresponding \\( \\mathbf{f}_{l,i} \\) and \\( \\mathbf{f}_{s,i} \\) are used to compute the bounding box for the object reference." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.47, + 0.564 + ], + "angle": 0, + "content": "Adding Viewpoint Prediction. As shown in Fig. 2, our approach extends this architecture in two ways – by modifying the encoder to enable location/perspective prediction and by using these predictions to transform bounding boxes \\( B \\) and point coordinates \\( p \\) prior to invoking the decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.625 + ], + "angle": 0, + "content": "We add two transformer-based predictors into each multimodal encoder layer shown in Fig. 2 (right). As both perspective and location features are updated in the same fashion, we write the update steps below for location only:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.629, + 0.469, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {l o c, v} ^ {(i)} = f _ {\\text {A t t n}} \\left(\\mathbf {f} _ {l o c, t} ^ {(i - 1)}, \\mathbf {f} _ {v} ^ {(i)}, \\mathbf {f} _ {v} ^ {(i)}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.653, + 0.469, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {l o c, t} ^ {(i)} = f _ {\\text {A t t n}} \\left(\\mathbf {f} _ {l o c, v} ^ {(i)}, \\mathbf {f} _ {t} ^ {(i)}, \\mathbf {f} _ {t} ^ {(i)}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.469, + 0.789 + ], + "angle": 0, + "content": "where \\(f_{\\mathrm{Attn}}(\\cdot ,\\cdot ,\\cdot)\\) denotes multi-head attention with key, query, and value arguments. The location feature is updated each layer by querying visual features and then querying textual features. We initialize the perspective and location features \\(\\mathbf{f}_{per,t}^{(0)}\\) with learnable embeddings \\(e_{per}\\) and \\(\\mathbf{f}_{loc,t}^{(0)}\\) with \\(e_{loc}\\) - a linear mapping of viewpoint \\((x,y)\\). Denote the final representations as \\(\\mathbf{f}_{per,t}^{(N)}\\) and \\(\\mathbf{f}_{loc,t}^{(N)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.469, + 0.85 + ], + "angle": 0, + "content": "We predict a distribution over locations and perspective from each layer's location and perspective features respectively using a learned linear layer. We denote our predictions over locations and perspective respectively as" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.855, + 0.469, + 0.876 + ], + "angle": 0, + "content": "\\[\nP (l o c \\mid \\mathbf {f} _ {l o c, m} ^ {(i)}) = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {l o c} \\mathbf {f} _ {l o c, m} ^ {(i)} + \\mathbf {b} _ {l o c}\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.879, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\nP (p e r \\mid \\mathbf {f} _ {p e r, m} ^ {(i)}) = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {l o c} \\mathbf {f} _ {p e r, m} ^ {(i)} + \\mathbf {b} _ {p e r}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.471 + ], + "angle": 0, + "content": "where \\(m \\in \\{\\mathrm{v},\\mathrm{t}\\}\\). We supervise each of these predictions with synthetic viewpoint supervision using an averaged cross-entropy loss - denoting it as" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.484, + 0.892, + 0.5 + ], + "angle": 0, + "content": "\\[\nL _ {a, m} = C E \\left(\\mathcal {Y} _ {a}, P _ {i} (a | m)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.512, + 0.842, + 0.528 + ], + "angle": 0, + "content": "where \\(a \\in \\{per, loc\\}\\) and \\(\\mathcal{Y}_a\\) as defined in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Object Decoder with Affine Transformation. After encoding the input, we take the highest probability location \\((\\hat{x},\\hat{y})\\) and perspective \\(\\hat{\\theta}\\) predictions generated from \\(f_{loc,t}^{(N)}\\) and \\(f_{per,t}^{(N)}\\) as an estimate of the observers viewpoint. Given these, we apply the corresponding affine transform \\(M(\\hat{\\theta},\\hat{x},\\hat{y})\\) to the point locations \\(p\\) and bounding boxes \\(B-\\) denoting the transformed versions as \\(p^{\\prime}\\) and \\(B^{\\prime}\\) respectively. These transformed versions along with the other features are then passed to the decoder as in Eq. 4. After the decoder predicts a bounding box location \\(\\mathbf{f}_l^{\\prime}\\) and size \\(\\mathbf{f}_s^{\\prime}\\) in this transformed coordinate system, we transform the prediction back to the original scene using the inverse of \\(\\mathbf{M}(\\hat{\\theta},\\hat{x},\\hat{y})\\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.727, + 0.882, + 0.743 + ], + "angle": 0, + "content": "3.3. Encouraging Uniform Object Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "An object can be described differently in different viewpoints - potentially causing our model's representation of the same object to be quite different from different viewpoints. To encourage our model to be more robust to these difference, we add an auxiliary loss that encourages the decoder output \\( f_{o} \\) for object \\( o \\) to be consistent across viewpoints. To do this, we introduce a set of instance-wise learnable vectors \\( \\mathbf{U} = \\{u_0, u_1, \\ldots\\} \\). We refer to these as uniform object representations. For a training sample whose target is \\( o \\), we concatenate the predicted object features \\( f_{o} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14059" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.138 + ], + "angle": 0, + "content": "and final perspective embeddings \\( f_{per,t}^{(N)} \\) to estimate the uniform object representation \\( u_{o} \\) with a linear layer. The loss is computed between the predicted representation and \\( u_{o} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.146, + 0.469, + 0.168 + ], + "angle": 0, + "content": "\\[\nL _ {l 1} = | | \\tanh (\\mathbf {W} _ {f u s e} [ f _ {o}, f _ {p e r, t} ^ {(N)} ] + \\mathbf {b} _ {f u s e}) - u _ {o} | | _ {1} \\quad (1 0)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.175, + 0.469, + 0.25 + ], + "angle": 0, + "content": "This encourages representation to be predictable from object features and viewpoint information – intuitively, predicting an object instance's identity from its observation at a specific viewpoint. Note that these learnable vectors are not used at inference time." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.469, + 0.311 + ], + "angle": 0, + "content": "Further, we also use \\( f_{o} \\) and \\( u_{o} \\) to classify the object's class to encourage semantic consistency. For both \\( f_{o} \\) and \\( u_{o} \\), we make a prediction over the object class and compute a Cross-Entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.32, + 0.469, + 0.339 + ], + "angle": 0, + "content": "\\[\nL _ {c l s} = C E \\left(\\mathcal {Y} _ {o}, \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {s e m} \\phi + \\mathbf {b} _ {s e m}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.346, + 0.395, + 0.362 + ], + "angle": 0, + "content": "where \\(\\phi \\in \\{f_o,u_o\\}\\) and \\(\\mathcal{V}_o\\) is the object's class." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.369, + 0.236, + 0.386 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.393, + 0.469, + 0.455 + ], + "angle": 0, + "content": "The overall training loss is a weight sum of our location and perspective supervision loss \\( L_{i,a,m} \\), the uniform object representation loss \\( L_{l1} \\), the object classification loss \\( L_{cls} \\) and the grounding loss \\( L_{gr} \\) proposed in BUTD-DETR [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.463, + 0.469, + 0.495 + ], + "angle": 0, + "content": "\\[\nL = \\alpha_ {1} \\sum_ {a, m} L _ {a, m} + \\alpha_ {2} L _ {l 1} + \\alpha_ {3} L _ {c l s} + \\alpha_ {4} L _ {g r} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.504, + 0.469, + 0.534 + ], + "angle": 0, + "content": "where the coefficients \\(\\alpha\\)'s are hyperparameters. Additionally, we find two additional techniques to be useful." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.706 + ], + "angle": 0, + "content": "Synthetic Training Curriculum. While training on the synthetic data for viewpoint supervision, we apply a filtering mechanism on visual features \\(\\mathbf{f}_v^{(0)}\\) that reduces the task complexity of early training examples. We retain all points (and associated features) that occur within the bounding box of an object mentioned in the referring expression. For points outside these boxes, we drop the point with probability \\(p_{drop}\\). At the start of training, we set \\(p_{drop} = 1 - \\text{effectively removing the background and other objects from the scene. As training progresses, we anneal } p_{drop} \\text{ to 0 such that whole scenes are observed.}\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.469, + 0.818 + ], + "angle": 0, + "content": "Viewpoint Data Augmentation. During both synthetic training and finetuning, we apply a viewpoint-based augmentation scheme. After running an example through our model, we retain the predicted affine transformation \\( M_0 \\). We then rotate the scene's pointcloud and supervision by \\( M_0 \\) to create a viewpoint augmented sample. We then run the model again with this sample." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.312, + 0.848 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "To study the effectiveness of our proposed VPP-Net, we conduct experiments on three existing 3D visual grounding datasets - ScanRefer [5], SR3D [2], and NR3D [2]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.809, + 0.108 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.119, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Datasets. We evaluate on ScanRefer, SR3D, and NR3D which we describe below: ScanRefer. The ScanRefer dataset is built on 800 3D scenes collected from ScanNet. The 3D scenes are labeled with about 51,000 natural expressions and 11,000 objects. The labeled natural expressions have different types of phrases, including intra/inter-class spatial relations and comparatives/superlatives. The dataset also provides real images of the scene, which are used to enhance the grounding performance in some methods. We use the official splits provided by ScanRefer to train and evaluate our model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.493 + ], + "angle": 0, + "content": "SR3D and NR3D. SR3D and NR3D are two datasets based on the ScanNet scenes proposed by Achlioptas et al. [2]. SR3D consists of 83K expressions generated by a template-based text generator. The expressions contain five types of spatial object-to-object relations: Horizontal/Vertical Proximity, Between, Allocentric, and Support, taking up \\(81.37\\%\\), \\(3.80\\%\\), \\(1.79\\%\\), \\(4.50\\%\\), and \\(8.54\\%\\) separately of all expressions. NR3D is a 3D visual grounding dataset with human-annotated expressions similar to ScanRefer. It consists of 41.5K human-annotated expressions and distractors collected through a \"player-listener\" game. Note that both SR3D and NR3D provide ground-truth bounding boxes for all candidate objects, unlike ScanRefer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.65 + ], + "angle": 0, + "content": "Experimental Settings. We evaluate our proposed VPP-Net on these datasets and report results in Tables 1 and 2. For ScanRefer, we report the top-1 accuracy when IoU is larger than 0.25 and 0.5 in \"Unique\", \"Multi\" and \"Overall\" cases. \"Unique\" vs \"Multi\" indicate whether distractors exist. \"Modality\" shows whether a method gets benefit from 2D images, with \"3D\" meaning the model only takes the 3D point cloud and \"3D+2D\" indicating that 2D images are also used. For SR3D and NR3D, following the experiment setting of EDA, we report Acc@0.25IoU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.717, + 0.675 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Synthetic data. We select 557 and 140 scenes from ScanNet to build the train and validation set of synthetic datasets. With the proposed synthetic viewpoint supervision mechanism, we repeat the generation process 75 times for each scene to generate the synthetic dataset – resulting in 36117 training and 9167 validation samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Model configuration. VPP-Net contains a 6-layer encoder and decoder. The learning rate of for the PointNet++ and Roberta is \\(2e^{-3}\\). The learning rate of other modules is \\(2e^{-4}\\). We first pre-train the model on the synthetic dataset to learn the viewpoint estimation for 150 epochs. Then we fine-tune using a combination of synthetic and the downstream dataset (Scanrefer/NR3D/SR3D) for another 150 epochs. We apply the point dropout mechanism on the synthetic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14060" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.411 + ], + "angle": 0, + "content": "
MethodModalityUnique(%)Multi(%)Overall(%)
Acc@0.25Acc@0.50Acc@0.25Acc@0.50Acc@0.25Acc@0.50
ScanRefer [5]2+3D76.3353.5132.7321.1141.1927.40
ReferIt3D [2]3D53.837.521.012.826.416.9
TGNN [15]3D68.6156.8029.8423.1837.3729.70
InstanceRefer [44]3D77.4566.8331.2724.7740.2332.93
SAT [43]2+3D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]3D78.8067.9435.1925.7041.3334.01
3DVG-Transformer [46]2+3D81.9360.6439.3028.4247.5734.67
3D-SPS [26]2+3D84.1266.7240.3229.8248.8236.98
3DJCG [4]2+3D83.4764.3441.3930.8249.5637.33
D3Net [6]3D+2D-70.35-30.50-37.87
UniT3D [8]3D+2D82.7573.1436.3631.0545.2739.14
BUTD-BasedBUTD-DETR [17]3D84.266.346.635.152.239.8
EDA [35]3D85.7668.5749.1337.6454.5942.26
Multi-ViewM3DRef-CLIP [45]3D+2D85.377.243.836.851.944.7
MVT [16]3D+2D--31.4624.8539.9532.28
ViewRefer [13]3D+2D--33.0826.5041.3033.66
VPP-Net (Ours)3D86.0567.0950.3239.0355.6543.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.422, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Table 1. 3D visual grounding results on ScanRefer. Acc@0.25/0.5 means top-1 accuracy when IoU is larger than 0.25/0.5 - higher is better. We group BUTD-Based and Multi-View methods for ease of comparison. BUTD-based use the same backbone of our model. The Multi-View methods all explored notions of viewpoint or 2D image rendering in 3D visual grounding. The state-of-the-art methods are bold and the second best performance is underlined in the table." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.55 + ], + "angle": 0, + "content": "data in both the pre-training and finetuning processes. The dropout rate \\( p_{drop} \\) is kept at 1 in the first 40 epochs then decreases to 0 linearly in the following 70 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Bounding boxes. Bounding boxes \\(\\mathbf{B}\\) are a required input for the decoder. In ScanRefer, no ground-truth object boundary box are provided. Thus we use GroupFree [25] to detect a set of bounding boxes \\(\\mathbf{B}_d\\), which \\(\\mathbf{B}_d \\in \\mathbb{R}^{m \\times 8 \\times 3}\\). Following the experiment setting of EDA [35] and BUTDDETR [17], in SR3D and NR3D, the model takes the ground-truth boundary boxes to replace the detected ones as the input of the decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.338, + 0.709 + ], + "angle": 0, + "content": "4.3. 3D Visual Grounding Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Analysis of ScanRefer Results. Our experimental results on ScanRefer are shown in Table 1. VPP-Net achieves state-of-the-art performance in terms of Acc@0.25 of \"Overall\" and \"Unique\", surpassing EDA by \\(1.06\\%\\) and \\(0.29\\%\\) respectively. An improvement of \\(1.19\\%\\) and \\(1.29\\%\\) can be observed at \"Multi\" cases to \\(50.32\\%\\) and \\(39.03\\%\\) separately. The \"Multi\" cases are more challenging than \"Unique\" cases generally because of the existence of distractors. This suggests that the 3D scene transformed with the predicted viewpoint provides less ambiguous spatial relations with which to reason about distractors." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Among the listed methods, we draw attention to \"BUTD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.882 + ], + "angle": 0, + "content": "Based\" and \"Multi-View\" models in the table - BUTD-DETR [17], EDA [35], M3DRef-CLIP [45], MVT [16], and ViewRefer [13]. BUTD-based models leverage the same BUTD-DETR backbone we build on. Compared with BUTD-DETR, our proposed VVP-Net has a large margin of improvement across all evaluation metrics. Compared to EDA, we can still observe an improvement of around \\(1.04\\%\\) and \\(1.39\\%\\) in \"Overall\" and \"Multi\". Multi-View models involve multi-view scenes or images that are related to our method. VPP-Net outperforms MVT and ViewRefer across all metrics. Compared with M3DRef-CLIP, our model achieves comparable but less competitive performance in Acc@0.5 of \"Unique\" and Acc@0.5 of \"Overall\", while achieving better performance in all other cases, especially, the Acc@0.5 of \"Multi\" which is more challenging. We hypothesize this is due to the multi-view images rendered of object candidates used in M3DRef-CLIP. Multi-view rendered images provide detailed information about an object, allowing the model localize the object more accurately. The multi-view images focus on single objects and do not provide much benefit in the \"Multi\" case, since the model can be fooled by multiple objects with similar 2D visual attributes in a scene. Thus, we can observe a larger margin of Acc@0.5 of \"Multi\" between C3DRef-CLIP and ours (improved by \\(+2.23\\%\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14061" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.391 + ], + "angle": 0, + "content": "
MethodModalitySR3DNR3D
ReferIt3D [2]3D39.835.6
TGNN [15]3D45.037.3
TransRefer3D [14]3D57.442.1
InstanceRefer [44]3D48.038.8
3DVG-Transfor [46]3D51.440.8
FFL-3DOG [12]3D-41.7
SAT [43]3D+2D57.949.2
3DReferTrans [1]3D47.039.0
LanguageRefer [32]3D56.043.9
3D-SPS [26]3D+2D62.651.5
BUTD-basedEDA [35]3D68.152.1
BUTD-DETR [17]3D67.154.9
Multi ViewM3DRef-CLIP [45]3D+2D-49.4
LAR [3]3D59.648.9
VPP-Net(Ours)3D68.756.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.4, + 0.472, + 0.444 + ], + "angle": 0, + "content": "Table 2. Result of SR3D and NR3D datasets. We report Acc@0.25IoU as the evaluation metric. Best and 2nd best results are denoted in bold or underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Analysis of SR3D & NR3D Results. We report results for SR3D and NR3D datasets in Table 2. Compared with existing methods, VPP-Net achieves state-of-the-art on both. Our observed improvement on NR3D (+2.00%) is larger than we see in SR3D (0.6%). We argue that this may be because there is a large domain gap between SR3D and our synthetic dataset. To show the domain gap, we count the frequency of spatial-relation words in SR3D, NR3D, ScanRefer, and our synthetic dataset in Table 4. We find an obvious gap: proximity relations - \"Closest/Farthest\" - dominate in SR3D, while relative spatial relations - \"Front/Behind\" and \"Left/Right\" - are more prominent in the human-annotated NR3D and ScanRefer datasets as well as our synthetic dataset. Despite this, VPP-Net still boosts BUTD-DETR performance on SR3D (+1.6%)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.82 + ], + "angle": 0, + "content": "To explicitly evaluate the efficiency of VPP-Net on viewpoint-dependent samples, we report results on view dependent / independent (dep./indep.) subsets of NR3D in Table 3. We see large gains over our backbone Butd-Detr for view dep. cases (46 vs. \\(52.4\\%\\) ) while achieving a similar result in view indep. case (58 vs. \\(58.6\\%\\)). In contrast, EDA's improvement for view dependent cases is smaller and results in performance loss on view independent cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.233, + 0.849 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We report an ablation study of our method on ScanRefer in Table 5. We remove different modules, auxiliary losses, or training procedures from our approach to show their contri" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.871, + 0.166 + ], + "angle": 0, + "content": "
ModelView Dep.View Indep.
Butd-Detr [17]46.0%58.0%
EDA [31]50.2 %53.1%
VPPNet (ours)52.4 %58.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.168, + 0.894, + 0.195 + ], + "angle": 0, + "content": "Table 3. Results of Acc@0.25 on View Dep/Indep-endent cases of NR3D dataset." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.225, + 0.894, + 0.334 + ], + "angle": 0, + "content": "
RelationSynthetic(%)Natural(%)
SR3DOursNR3DScanRefer
Left/Right8.9244.5836.0055.66
Front/Behind2.5750.0819.3920.11
Above/Under3.815.3414.1913.09
Between8.540.003.216.77
Closest/Farthest81.380.0031.214.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.343, + 0.894, + 0.4 + ], + "angle": 0, + "content": "Table 4. Normalized frequency of spatial relation words across datasets. We find SR3D has significantly fewer relative spatial relations than our synthetic datasets or either human-collected datasets on which we evaluate." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.429, + 0.891, + 0.512 + ], + "angle": 0, + "content": "
Synthetic PretrainingViewpoint PredictionUniform Obj. Rep.Curriculum FilteringViewpoint Data Aug.Acc@0.25Acc@0.50
0*-----52.139.8
1--✓✓-51.738.9
2✓✓-✓✓53.7838.98
3✓✓✓-✓51.5238.45
4✓✓✓✓-53.7840.41
5-✓✓✓✓49.2333.67
6✓✓✓✓✓55.6543.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.521, + 0.893, + 0.564 + ], + "angle": 0, + "content": "Table 5. Ablation of our model components, auxiliary losses, and training methods on the ScanRefer dataset - checkmarks denote the corresponding item is active. Rows are numbered." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.669 + ], + "angle": 0, + "content": "bution. Note that row \"0*\" without any of our modifications is equivalent to BUTD-DETR. When viewpoint prediction is disabled (row 1), we also disable the modules that rely on it – pretraining and viewpoint data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We find that removing the viewpoint prediction and data augmentation (row 1), or the curriculum filtering during pretraining (row 3) results in performance similar to the baseline. These correspond to models that either do not use viewpoint prediction (row 1) or learn viewpoint prediction poorly (row 3) – suggesting the importance of viewpoint in 3D visual grounding. We also note that directly jointly training VPP-Net without first developing a strong viewpoint prediction capability during pretraining (row 5) yields worse results than not considering viewpoint at all. Removing either the uniform object representation loss (row 2) or the viewpoint data augmentation (row 4) yields degraded performance compared with the whole model (row 6). Both are designed to improve learned representations and we see they have a positive effect in these experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.088, + 0.895, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.416, + 0.895, + 0.475 + ], + "angle": 0, + "content": "Figure 3. Example VPP-Net results. The first row shows successful examples where both viewpoint prediction and visual grounding are correct. The second row shows failed samples. The first two predict incorrect viewpoints but still succeed in visual grounding. The predicted observer position (red dot) and facing direction (facing away from the green arrow) are shown in 3D and a top-down view (top right corner). We zoom up on the target region in example (7) within the red circle to provide better visualization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.281, + 0.515 + ], + "angle": 0, + "content": "4.5. Qualitative Examples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We depict the quantitative results of VPP-Net in Fig. 3. In the figure, we list 8 examples of which four are successful (top row) and four are not (bottom row). In each example, the red dot represents the predicted location and the green and blue arrows represent the \"behind\" and \"right\" direction predicted in the model such that the predicted observer is facing away from the green arrow. The ground truth bounding boxes and target words are noted with green and the mentioned objects are noted with red. We also provide the predicted object bounding box in the image, shown in blue. The spatial relations are noted with blue in the text. The successful examples show that with an accurate viewpoint prediction, the expression can better match the 3D scene, resulting in more accurate groundings. Failed examples (5) and (6) are successful at grounding despite having failed to predict valid viewpoints. Both contain ambiguous elements in text or image, i.e. the expression \"walk thru the door on the right\" in example (5) and the distorted door (blob on bottom left of the scene) in example (6). Examples (7) and (8) predict valid viewpoints but produce incorrect visual groundings. The predicted bounding box in example (7) covers a certain region of the ground truth and involved patch contains similar visual attributes (white). In example (8), both ground truth and prediction are on the left of the viewpoint, matching the expression. However," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.531 + ], + "angle": 0, + "content": "the model failed to understand the horizontal proximity (farthest), leading to a incorrect grounding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.569, + 0.62, + 0.585 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.786 + ], + "angle": 0, + "content": "In conclusion, we have proposed a Viewpoint Prediction Network (VPP-Net) for 3D visual grounding tasks. Our model addresses the critical challenge of viewpoint-dependent ambiguity in 3D visual grounding. Specifically, we first introduce a synthetic dataset and then use it to train a 3D visual grounding model that can better disambiguate referring expressions by explicitly predicting potential viewpoints. Further, we design a uniform object representation auxiliary loss and viewpoint data augmentation scheme that further improve the performance of our model. Experiments on the ScanRefer, SR3D, and NR3D demonstrate the effectiveness of our proposed methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Limitations. The synthetic datasets we generate are highly dependent on the templated referring expressions, which limits their diversity and realism. Future works should include strategies to generate more diverse datasets or collect human-generated referring expressions with annotated viewpoints to further improve." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14063" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dreftransformer: Fine-grained object identification in realworld scenes using natural language. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 422-440. Springer, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[3] Eslam Bakri, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.409 + ], + "angle": 0, + "content": "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.52 + ], + "angle": 0, + "content": "[6] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[7] Sijia Chen and Baochun Li. Multi-modal dynamic graph transformer for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15534-15543, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.468, + 0.717 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[11] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 202-210. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[13] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[14] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[15] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[16] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[17] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, pages 417-433. Springer, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[18] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[19] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[20] Weide Liu, Zhonghua Wu, Henghui Ding, Fayao Liu, Jie Lin, and Guosheng Lin. Few-shot segmentation with global and local contrastive learning. arXiv preprint arXiv:2108.05293, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[21] Weide Liu, Zhonghua Wu, Yang Zhao, Yuming Fang, Chuan-Sheng Foo, Jun Cheng, and Guosheng Lin. Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation. International Journal of Computer Vision, pages 1-15, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[22] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[23] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14064" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5612-5621, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[27] Zhipeng Luo, Gongjie Zhang, Changqing Zhou, Zhonghua Wu, Qingyi Tao, Lewei Lu, and Shijian Lu. Modeling continuous motion for 3d point cloud object tracking. arXiv preprint arXiv:2303.07605, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[28] Hui En Pang, Zhongang Cai, Lei Yang, Qingyi Tao, Zhonghua Wu, Tianwei Zhang, and Ziwei Liu. Towards robust and expressive whole-body human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[30] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[32] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[34] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Yu Wu, Yana Wei, Haozhe Wang, Yongfei Liu, Sibei Yang, and Xuming He. Grounded image text matching with mismatched relation reasoning. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 2976-2987, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[37] Zhonghua Wu, Guosheng Lin, and Jianfei Cai. Keypoint based weakly supervised human parsing. Image and Vision Computing, 91:103801, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[38] Zhonghua Wu, Qingyi Tao, Guosheng Lin, and Jianfei Cai. Exploring bottom-up and top-down cues with attentive learning for webly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12936-12945, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[39] Zhonghua Wu, Xiangxi Shi, Guosheng Lin, and Jianfei Cai. Learning meta-class memory for few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 517-526, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[40] Zhonghua Wu, Yicheng Wu, Guosheng Lin, Jianfei Cai, and Chen Qian. Dual adaptive transformations for weakly supervised point cloud segmentation. In European conference on computer vision, pages 78-96. Springer Nature Switzerland Cham, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[41] Zhonghua Wu, Yicheng Wu, Guosheng Lin, and Jianfei Cai. Reliability-adaptive consistency regularization for weakly-supervised point cloud segmentation. International Journal of Computer Vision, pages 1-14, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[42] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[43] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[44] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[45] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[46] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14065" + } + ] +] \ No newline at end of file diff --git a/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_origin.pdf b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06b6ab9dc71daf31aef24418d4e7fd623e803bfb --- /dev/null +++ b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/8d04b1f0-f2a1-4a61-aa7a-713e99eb821c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe99a8bdc48f65b9590cdad2428c0f123f4198b633bfd479b54711b9db42eeb5 +size 6624218 diff --git a/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/full.md b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9068ef2bafe48b628bbbb034add4da616187337d --- /dev/null +++ b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/full.md @@ -0,0 +1,314 @@ +# Viewpoint-Aware Visual Grounding in 3D Scenes + +Xiangxi Shi +Oregon State University +shixia@oregonstate.edu + +Zhonghua Wu SenseTime Research wuzhonghua@sensetime.com + +Stefan Lee +Oregon State University +leestef@oregonstate.edu + +# Abstract + +Referring expressions for visual objects often include descriptions of relative spatial arrangements to other objects - e.g. "to the right of" - that depend on the point of view of the speaker. In 2D referring expression tasks, this viewpoint is captured unambiguously in the image. However, grounding expressions with such spatial language in 3D without viewpoint annotations can be ambiguous. In this paper, we investigate the significance of viewpoint information in 3D visual grounding - introducing a model that explicitly predicts the speaker's viewpoint based on the referring expression and scene. We pretrain this model on a synthetically generated dataset that provides viewpoint annotations and then finetune on 3D referring expression datasets. Further, we introduce an auxiliary uniform object representation loss to encourage viewpoint invariance in learned object representations. We find that our proposed ViewPoint Prediction Network (VPP-Net) achieves state-of-the-art performance on ScanRefer, SR3D, and NR3D - improving Accuracy@0.25IoU by $1.06\%$ , $0.60\%$ , and $2.00\%$ respectively compared to prior work. + +# 1. Introduction + +Visual grounding of referring expressions requires algorithms to reason about natural language object descriptions to identify object referents in visual scenes. Naturally, describing relative spatial relations between objects is a common strategy when a speaker produces a referring expression – e.g., “The chair to the right of the couch.” As such, many methods for visual grounding put algorithmic emphasis on handling these spatial relations [7, 18, 24, 34, 36]. + +However, these sorts of spatial relations depend upon the point of view of the speaker. For instance, Fig. 1 depicts two different viewpoints of the same 3D scene and notes that these views produce contradictory referring expressions. In 2D visual grounding tasks, the point of view of the speaker is typically also that of the image – introducing no ambiguity into spatial reasoning. However, viewpoint annotations are not typically captured for 3D visual grounding tasks that + +![](images/bf6c096bcbbc885a1f52dc7855c0fbf3a77f727705e99c500519eeaa452d9108.jpg) +Figure 1. Expressions referring to the same object from different viewpoints can lead to contradictory spatial relations. + +![](images/7ff6caed88ec5c442b930ac9232976e4f38a69100be06fd70279376febcce266.jpg) + +reason about whole scenes. As a result, identifying referent objects may be ambiguous if multiple object instances align with the referring expression in differing views. Moreover, the lack of reliable grounding for spatial language may simply make the reasoning task more difficult to learn even when the referent is still uniquely identifiable. + +Despite this, most of the recent work in 3D visual grounding has not addressed the role of viewpoint in the learning process [12, 17, 26, 35, 43, 46]. Those methods that have attempted to consider the influence of viewpoint have been hindered by the lack of viewpoint supervision in existing datasets – instead relying on rendering multiple views of a 3D scene and using attention mechanisms to softly select between them [13, 16] in an end-to-end model. + +In this work, we develop a straightforward approach to viewpoint-aware visual grounding in 3D scenes – explicitly predicting the speaker's viewpoint and then transforming the input accordingly. To support this direction, we create a synthetic visual grounding dataset that provides annotations to directly supervise viewpoint prediction. We propose a novel framework called ViewPoint Prediction Net (VPP-Net) to estimate the viewpoints of expression-pointcloud pairs and rotate the 3D scenes accordingly before performing expression grounding. This model is first pre-trained on the synthetic dataset and then fine-tuned on a combination of the synthetic and target datasets. To our knowledge, this + +is the first work to introduce a supervised viewpoint estimation model for 3D visual grounding tasks. In experiments on three common 3D visual grounding datasets, this approach paired with viewpoint-related auxiliary losses leads to improved performance over prior art. + +Contributions. Our contributions can be summarized as: + +- We introduce viewpoint prediction as an auxiliary task in the 3D visual grounding task and propose synthetic data generation pipeline to provide viewpoint supervision. +- We propose VPP-Net - a novel model that learns viewpoint prediction to transform 3D scenes to reduce ambiguity in spatial-relation grounding. +- We design a Uniform Object Representation auxiliary loss that promotes viewpoint invariant feature learning. +- Combing these techniques, our proposed approach achieves state-of-the-art performance on the ScanRefer [5], SR3D [2], and NR3D [2] datasets. + +# 2. Related Work + +Relation-aware 2D Visual Grounding. The visual grounding task has been extensively explored in 2D image scenarios. The goal of a 2D visual grounding model is to locate an object according to a given expression in a 2D image. In recent work, the relationships between objects have been identified as a crucial cue in complex environments. Several works [7, 23, 24, 34, 42] have parsed referring expressions into phrases for better text understanding. Wang et al. [34] proposed to align the generated visual graph with the self-attend object and relation features using a graph attention module. Liu [24] proposed to capture the visual context from a set of coarse-level object proposals by reconstructing the corresponding relations captured from the text. Chen [7] et al. initialize a visual graph with proposed node and edge transformers to learn the representations of objects and relations. Besides explicitly encoding visual relations in the model, relationships can also benefit models by constructing pseudo data [18, 36]. Wu et al. [36] proposed to detect mismatched relations from synthetic data to evaluate the ability of relation understanding of the model. Jiang et.al. [18] introduce an automatic way to generate pseudo visual grounding data for supervised training. + +3D Visual Grounding. The first dataset for visual grounding in 3D was introduced by Chen et al. [5] - presenting the ScanRefer dataset consisting of 3D scenes and natural expression. Achlioptas et al. [2] propose two datasets - the synthetic dataset SR3D and the human-annotated NR3D. Further, they show that template-based synthetic data can benefit the task when models are jointly trained with naturally collected data. In standard approaches, a 3D grounding model first encodes instructions and 3D scene with pretrained text [10, 22, 31] and 3D visual [19, 27-30, 40, 41] models. Then a multimodal module is proposed to fuse the encoded features to estimate the location and + +size of the grounded object. 3D-SPS [26] proposed a one-stage method that estimates the location of objects from the point cloud directly. In other works, 3D object detections [11, 19, 25, 33] are leveraged to provide an object-focused candidate pool consisting of object-level features and/or bounding boxes. BUTD-DETR [17] encodes object bounding boxes as a reference and jointly trains on object detection datasets as an auxiliary task. 2D images [20, 21, 37-39] are also utilized to boost the performance of 3D visual grounding. The model proposed in [5] encodes 2D images together with 3D scenes for a better 3D understanding. Likewise, Yang et al. [43] leverage 2D semantics to assist 3D representation learning. + +The viewpoint of the annotator is also a significant cue for a better alignment between 3D scenes and referring expressions. Currently, only a few works focused on this aspect have been proposed. MVT [16] and ViewRefer [13] rotate the 3D scenes in different angles and encode them to multi-view representations without explicit supervision of which viewpoint matches to the annotator. Compared with previous work, we first provide a template-based synthetic visual grounding dataset including location and perspective supervision and then propose a supervised training method for viewpoint estimation. We demonstrate this is be beneficial for visual grounding performance. + +# 3. Methods + +We propose a straight-forward approach for learning to reason about viewpoint in 3D referring expression grounding – directly predicting a rotation and translation of the 3D environment that aligns it with the assumed viewpoint of the observer providing the referring expression. To accomplish this, we introduce a ViewPoint Prediction Network (VPP-Net) – a model that performs 3D referring expression grounding while explicitly estimating the location and heading of the observer. As existing datasets lack observer viewpoint annotations, we develop a synthetic data generation pipeline to provide appropriate training supervision for viewpoint-aware pretraining and then transfer the pretrained model to downstream tasks. Further, we introduce a Uniform Object Representation (UOR) auxiliary loss to ensure object instances are represented consistently regardless of viewpoint. We describe each component below. + +# 3.1. Generating Synthetic Viewpoint Supervision + +Our synthetic data generation pipeline operates in four stages - for a given environment, we (1) sample a location and heading for a virtual observer, (2) identify spatial relationships between annotated objects in view, (3) generate templated referring expressions accordingly, and then (4) find other nearby viewpoints that could also have produced this reference. The result is a tuple of the 3D environment, + +referring expression, and valid viewpoint map which we can use for training viewpoint prediction. + +Input Scenes. We utilize 3D scenes from the ScanNet [9] dataset which are annotated with detected object bounding boxes. For a given scene, we denote bounding box as a set $B$ , their centroids as a set of homogenous XYZ-coordinates $\mathbf{C} = \{c_0, c_1, \dots, c_n \mid c_i = (x_i, y_i, z_i, 1)\}$ , and the corresponding object category labels $\mathcal{Y}_o$ . + +Viewpoint Sampling. Regardless of scene size, we split the horizontal (XY) plane into a $10 \times 10$ grid and uniformly sample a 2D $(x,y)$ translation from the grid. Likewise, we sample a rotation $\theta$ about the vertical axis by randomly selecting a $10^{\circ}$ increment between $0^{\circ}$ and $360^{\circ}$ . + +Identifying Spatial Relationships. We apply the sampled translation and rotation to the scene by computing updated bounding box centers. With the affine transformation + +$$ +M (\theta , x, y) = \left( \begin{array}{c c c c} \cos \theta & - \sin \theta & 0 & - x \\ \sin \theta & \cos \theta & 0 & - y \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{array} \right), \tag {1} +$$ + +we update the bounding box centroids as $c_i' = c_i M(\theta, x, y)^T$ . The effect of this operation is to align the observer's relative spatial terms (e.g. left/right) to the XYZ axes. To detect potential spatial relations between a pair of objects $o_i$ and $o_j$ , we compute the vector $r_{ij} = c_i' - c_j'$ to represent their relative position. We take a set of four unit direction vectors as $\{d_{front}, d_{back}, d_{left}, d_{right}\}$ and compute the cosine similarities between $d_k$ and $r_{ij}$ + +$$ +s _ {i, j, k} = \frac {r _ {i j} d _ {k}}{\left| \left| r _ {i j} \right| \right| \times \left| \left| d _ {k} \right| \right|}. \tag {2} +$$ + +We consider a spatial relation $k$ to be valid if $s_{i,j,k} > 0.3$ . For above and below relations, we instead consider the heights and overlap of the bounding boxes. From all valid tuples, we select a random subset as the relation tuple set $\mathbf{R}$ . We note that this process does not consider the "visibility" of objects from the observer location. + +Expression Generation. Given R and object category annotations, we can produce a set of object referring expressions. For each object mentioned as part of a relation in R, we randomly sample up to four valid relations and convert each into simple templated text. For example, the tuple ("chair", "left", "bed") is converted to a string like "The chair is to the left of the bed." These are then concatenated to form a multi-sentence referring expression for the object. + +Viewpoint Supervision. For a viewpoint $(x,y,\theta)$ and referring expression $T_{o}$ , we provide viewpoint supervision decomposed into location and heading prediction. In most cases, the randomly selected viewpoint is not the only one for which the referring expression would be valid - small + +shifts in either location or perspective may not change the relationship between observed objects. For perspective, we produce a binary vector $\mathcal{V}_{Per}$ with 36 entries where $\mathcal{V}_{Per}[k]$ is 1 if the referring expression $T_{o}$ is valid from viewpoint $(x,y,k*10^{\circ})$ . For location, we produce a $10\times 10$ binary matrix $\mathcal{V}_{Loc}$ where $\mathcal{V}_{Per}[i,j]$ is 1 if the referring expression $T_{o}$ is valid from viewpoint $(i,j,\theta)$ . For both, we check validity by repeating the spatial relationship tests described earlier from the new viewpoint. We note that this approach considers only a subset of all possible viewpoints and assumes that perspectives valid at $(x,y)$ are also valid for any point $(i,j)$ for which the original perspective was valid. + +# 3.2. Viewpoint-Aware 3D Grounding Network + +As outlined in Fig. 2, our ViewPoint Prediction Network (VPP-Net) approach consists of a multimodal encoder-decoder architecture. After encoding the referring expression and scene, we predict the location and perspective of the observer. The encoded features and transformed candidate bounding boxes and XYZ point positions are then passed to a decoding model which ranks the candidates. + +VPP-Net Backbone Architecture. Our method is based on BUTD-DETR [17]. We review its construction here. + +Input Encoding. An RGBXYZ point cloud representing the 3D scene is encoded with a pretrained PointNet++ [29], producing $n_p$ point features $\mathbf{e}_{\text{vision}} \in \mathbb{R}^{n_p \times d}$ and the corresponding XYZ points $\mathbf{p} \in \mathbb{R}^{n_p \times 3}$ . Referring expressions are encoded with a pre-trained RoBerta [22] model into a sequence of text embeddings $\mathbf{e}_{\text{text}} \in \mathbb{R}^{|T| \times d}$ , where $|T|$ is the token length. + +Multimodal Encoder. The multimodal encoder consists of $N$ layers of attention-based modules. Initializing visual features $f_{v}^{(0)}$ and textual features $f_{t}^{(0)}$ with $e_{vision}$ and $e_{text}$ respectively, each layer $i$ computes update representations by applying per-modality self attention followed by cross-modality attention as below: + +$$ +\mathbf {f} _ {v} ^ {(i + 1)}, \mathbf {f} _ {t} ^ {(i + 1)} = f _ {\text {C r o s s - A t t}} \left(f _ {\text {S e l f - A t t}} \left(\mathbf {f} _ {v} ^ {(i)}\right), f _ {\text {S e l f - A t t}} \left(\mathbf {f} _ {t} ^ {(i)}\right)\right). \tag {3} +$$ + +We denote the final layer outputs as $\mathbf{f}_t^{(N)}$ and $\mathbf{f}_v^{(N)}$ . + +Object Decoder. Between the encoder and decoder, the visual features $\mathbf{f}_v^{(N)}$ are passed through as small MLP to produce scores for each visual input that are then passed through a softmax. The top-K scoring transformed features are retained as object queries $\hat{\mathbf{f}}$ . The object decoder is an N-layer transformer-based module. The decoder takes $\hat{\mathbf{f}}$ , $\mathbf{f}_t^{(N)}$ , $\mathbf{f}_v^{(N)}$ , hierarchical points $\mathbf{c}_v$ and detected bounding boxes $\mathbf{B}_o$ as input to predict $\mathbf{r}$ , $\mathbf{f}_o$ , $\mathbf{f}_l$ , $\mathbf{f}_s$ : + +$$ +\mathbf {r}, \mathbf {f} _ {o}, \mathbf {f} _ {l}, \mathbf {f} _ {s} = f _ {\mathrm {D e c}} \left(\hat {\mathbf {f}}, \mathbf {f} _ {t}, \mathbf {f} _ {v}, \mathbf {c} _ {v}, \mathbf {B} _ {o}\right), \tag {4} +$$ + +where $\mathbf{r} \in \mathbb{R}^{n_p \times 1}$ is the object-expression matching score; $\mathbf{f}_o \in \mathbb{R}^{n_p \times D}$ are object features; and $\mathbf{f}_l, \mathbf{f}_s \in \mathbb{R}^{n_p \times 3}$ are + +![](images/c329894d27a521b1efb9162494bcc562e7e89069af3231c8742914e1a09183a2.jpg) +Figure 2. An overview of our ViewPoint Predictor Network (VPP-Net) approach. (Left) Our model builds on the encoder-decoder structure of [17], but modifies the encoder to explicitly predict location and perspective of the observer. The decoder then processes appropriately transformed inputs to rank candidate bounding boxes. (Right) Our encoder model updates features of location and perspective at each layer that are supervised to predict viewpoints where the referring expression is valid. + +the location and size of objects. The object with the highest $r_i$ is chosen and its corresponding $\mathbf{f}_{l,i}$ and $\mathbf{f}_{s,i}$ are used to compute the bounding box for the object reference. + +Adding Viewpoint Prediction. As shown in Fig. 2, our approach extends this architecture in two ways – by modifying the encoder to enable location/perspective prediction and by using these predictions to transform bounding boxes $B$ and point coordinates $p$ prior to invoking the decoder. + +We add two transformer-based predictors into each multimodal encoder layer shown in Fig. 2 (right). As both perspective and location features are updated in the same fashion, we write the update steps below for location only: + +$$ +\mathbf {f} _ {l o c, v} ^ {(i)} = f _ {\text {A t t n}} \left(\mathbf {f} _ {l o c, t} ^ {(i - 1)}, \mathbf {f} _ {v} ^ {(i)}, \mathbf {f} _ {v} ^ {(i)}\right), \tag {5} +$$ + +$$ +\mathbf {f} _ {l o c, t} ^ {(i)} = f _ {\text {A t t n}} \left(\mathbf {f} _ {l o c, v} ^ {(i)}, \mathbf {f} _ {t} ^ {(i)}, \mathbf {f} _ {t} ^ {(i)}\right), \tag {6} +$$ + +where $f_{\mathrm{Attn}}(\cdot ,\cdot ,\cdot)$ denotes multi-head attention with key, query, and value arguments. The location feature is updated each layer by querying visual features and then querying textual features. We initialize the perspective and location features $\mathbf{f}_{per,t}^{(0)}$ with learnable embeddings $e_{per}$ and $\mathbf{f}_{loc,t}^{(0)}$ with $e_{loc}$ - a linear mapping of viewpoint $(x,y)$ . Denote the final representations as $\mathbf{f}_{per,t}^{(N)}$ and $\mathbf{f}_{loc,t}^{(N)}$ . + +We predict a distribution over locations and perspective from each layer's location and perspective features respectively using a learned linear layer. We denote our predictions over locations and perspective respectively as + +$$ +P (l o c \mid \mathbf {f} _ {l o c, m} ^ {(i)}) = \operatorname {S o f t m a x} \left(\mathbf {W} _ {l o c} \mathbf {f} _ {l o c, m} ^ {(i)} + \mathbf {b} _ {l o c}\right) \tag {7} +$$ + +$$ +P (p e r \mid \mathbf {f} _ {p e r, m} ^ {(i)}) = \operatorname {S o f t m a x} \left(\mathbf {W} _ {l o c} \mathbf {f} _ {p e r, m} ^ {(i)} + \mathbf {b} _ {p e r}\right) \tag {8} +$$ + +where $m \in \{\mathrm{v},\mathrm{t}\}$ . We supervise each of these predictions with synthetic viewpoint supervision using an averaged cross-entropy loss - denoting it as + +$$ +L _ {a, m} = C E \left(\mathcal {Y} _ {a}, P _ {i} (a | m)\right) \tag {9} +$$ + +where $a \in \{per, loc\}$ and $\mathcal{Y}_a$ as defined in Sec. 3.1. + +Object Decoder with Affine Transformation. After encoding the input, we take the highest probability location $(\hat{x},\hat{y})$ and perspective $\hat{\theta}$ predictions generated from $f_{loc,t}^{(N)}$ and $f_{per,t}^{(N)}$ as an estimate of the observers viewpoint. Given these, we apply the corresponding affine transform $M(\hat{\theta},\hat{x},\hat{y})$ to the point locations $p$ and bounding boxes $B-$ denoting the transformed versions as $p^{\prime}$ and $B^{\prime}$ respectively. These transformed versions along with the other features are then passed to the decoder as in Eq. 4. After the decoder predicts a bounding box location $\mathbf{f}_l^{\prime}$ and size $\mathbf{f}_s^{\prime}$ in this transformed coordinate system, we transform the prediction back to the original scene using the inverse of $\mathbf{M}(\hat{\theta},\hat{x},\hat{y})$ + +# 3.3. Encouraging Uniform Object Representation + +An object can be described differently in different viewpoints - potentially causing our model's representation of the same object to be quite different from different viewpoints. To encourage our model to be more robust to these difference, we add an auxiliary loss that encourages the decoder output $f_{o}$ for object $o$ to be consistent across viewpoints. To do this, we introduce a set of instance-wise learnable vectors $\mathbf{U} = \{u_0, u_1, \ldots\}$ . We refer to these as uniform object representations. For a training sample whose target is $o$ , we concatenate the predicted object features $f_{o}$ + +and final perspective embeddings $f_{per,t}^{(N)}$ to estimate the uniform object representation $u_{o}$ with a linear layer. The loss is computed between the predicted representation and $u_{o}$ : + +$$ +L _ {l 1} = | | \tanh (\mathbf {W} _ {f u s e} [ f _ {o}, f _ {p e r, t} ^ {(N)} ] + \mathbf {b} _ {f u s e}) - u _ {o} | | _ {1} \quad (1 0) +$$ + +This encourages representation to be predictable from object features and viewpoint information – intuitively, predicting an object instance's identity from its observation at a specific viewpoint. Note that these learnable vectors are not used at inference time. + +Further, we also use $f_{o}$ and $u_{o}$ to classify the object's class to encourage semantic consistency. For both $f_{o}$ and $u_{o}$ , we make a prediction over the object class and compute a Cross-Entropy loss: + +$$ +L _ {c l s} = C E \left(\mathcal {Y} _ {o}, \operatorname {S o f t m a x} \left(\mathbf {W} _ {s e m} \phi + \mathbf {b} _ {s e m}\right)\right), \tag {11} +$$ + +where $\phi \in \{f_o,u_o\}$ and $\mathcal{V}_o$ is the object's class. + +# 3.4. Model Training + +The overall training loss is a weight sum of our location and perspective supervision loss $L_{i,a,m}$ , the uniform object representation loss $L_{l1}$ , the object classification loss $L_{cls}$ and the grounding loss $L_{gr}$ proposed in BUTD-DETR [17]: + +$$ +L = \alpha_ {1} \sum_ {a, m} L _ {a, m} + \alpha_ {2} L _ {l 1} + \alpha_ {3} L _ {c l s} + \alpha_ {4} L _ {g r} \tag {12} +$$ + +where the coefficients $\alpha$ 's are hyperparameters. Additionally, we find two additional techniques to be useful. + +Synthetic Training Curriculum. While training on the synthetic data for viewpoint supervision, we apply a filtering mechanism on visual features $\mathbf{f}_v^{(0)}$ that reduces the task complexity of early training examples. We retain all points (and associated features) that occur within the bounding box of an object mentioned in the referring expression. For points outside these boxes, we drop the point with probability $p_{drop}$ . At the start of training, we set $p_{drop} = 1 - \text{effectively removing the background and other objects from the scene. As training progresses, we anneal } p_{drop} \text{ to 0 such that whole scenes are observed.}$ + +Viewpoint Data Augmentation. During both synthetic training and finetuning, we apply a viewpoint-based augmentation scheme. After running an example through our model, we retain the predicted affine transformation $M_0$ . We then rotate the scene's pointcloud and supervision by $M_0$ to create a viewpoint augmented sample. We then run the model again with this sample. + +# 4. Experiments and Results + +To study the effectiveness of our proposed VPP-Net, we conduct experiments on three existing 3D visual grounding datasets - ScanRefer [5], SR3D [2], and NR3D [2]. + +# 4.1. Datasets and Experimental Settings + +Datasets. We evaluate on ScanRefer, SR3D, and NR3D which we describe below: ScanRefer. The ScanRefer dataset is built on 800 3D scenes collected from ScanNet. The 3D scenes are labeled with about 51,000 natural expressions and 11,000 objects. The labeled natural expressions have different types of phrases, including intra/inter-class spatial relations and comparatives/superlatives. The dataset also provides real images of the scene, which are used to enhance the grounding performance in some methods. We use the official splits provided by ScanRefer to train and evaluate our model. + +SR3D and NR3D. SR3D and NR3D are two datasets based on the ScanNet scenes proposed by Achlioptas et al. [2]. SR3D consists of 83K expressions generated by a template-based text generator. The expressions contain five types of spatial object-to-object relations: Horizontal/Vertical Proximity, Between, Allocentric, and Support, taking up $81.37\%$ , $3.80\%$ , $1.79\%$ , $4.50\%$ , and $8.54\%$ separately of all expressions. NR3D is a 3D visual grounding dataset with human-annotated expressions similar to ScanRefer. It consists of 41.5K human-annotated expressions and distractors collected through a "player-listener" game. Note that both SR3D and NR3D provide ground-truth bounding boxes for all candidate objects, unlike ScanRefer. + +Experimental Settings. We evaluate our proposed VPP-Net on these datasets and report results in Tables 1 and 2. For ScanRefer, we report the top-1 accuracy when IoU is larger than 0.25 and 0.5 in "Unique", "Multi" and "Overall" cases. "Unique" vs "Multi" indicate whether distractors exist. "Modality" shows whether a method gets benefit from 2D images, with "3D" meaning the model only takes the 3D point cloud and "3D+2D" indicating that 2D images are also used. For SR3D and NR3D, following the experiment setting of EDA, we report Acc@0.25IoU. + +# 4.2. Implementation Details + +Synthetic data. We select 557 and 140 scenes from ScanNet to build the train and validation set of synthetic datasets. With the proposed synthetic viewpoint supervision mechanism, we repeat the generation process 75 times for each scene to generate the synthetic dataset – resulting in 36117 training and 9167 validation samples. + +Model configuration. VPP-Net contains a 6-layer encoder and decoder. The learning rate of for the PointNet++ and Roberta is $2e^{-3}$ . The learning rate of other modules is $2e^{-4}$ . We first pre-train the model on the synthetic dataset to learn the viewpoint estimation for 150 epochs. Then we fine-tune using a combination of synthetic and the downstream dataset (Scanrefer/NR3D/SR3D) for another 150 epochs. We apply the point dropout mechanism on the synthetic + +
MethodModalityUnique(%)Multi(%)Overall(%)
Acc@0.25Acc@0.50Acc@0.25Acc@0.50Acc@0.25Acc@0.50
ScanRefer [5]2+3D76.3353.5132.7321.1141.1927.40
ReferIt3D [2]3D53.837.521.012.826.416.9
TGNN [15]3D68.6156.8029.8423.1837.3729.70
InstanceRefer [44]3D77.4566.8331.2724.7740.2332.93
SAT [43]2+3D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]3D78.8067.9435.1925.7041.3334.01
3DVG-Transformer [46]2+3D81.9360.6439.3028.4247.5734.67
3D-SPS [26]2+3D84.1266.7240.3229.8248.8236.98
3DJCG [4]2+3D83.4764.3441.3930.8249.5637.33
D3Net [6]3D+2D-70.35-30.50-37.87
UniT3D [8]3D+2D82.7573.1436.3631.0545.2739.14
BUTD-BasedBUTD-DETR [17]3D84.266.346.635.152.239.8
EDA [35]3D85.7668.5749.1337.6454.5942.26
Multi-ViewM3DRef-CLIP [45]3D+2D85.377.243.836.851.944.7
MVT [16]3D+2D--31.4624.8539.9532.28
ViewRefer [13]3D+2D--33.0826.5041.3033.66
VPP-Net (Ours)3D86.0567.0950.3239.0355.6543.29
+ +Table 1. 3D visual grounding results on ScanRefer. Acc@0.25/0.5 means top-1 accuracy when IoU is larger than 0.25/0.5 - higher is better. We group BUTD-Based and Multi-View methods for ease of comparison. BUTD-based use the same backbone of our model. The Multi-View methods all explored notions of viewpoint or 2D image rendering in 3D visual grounding. The state-of-the-art methods are bold and the second best performance is underlined in the table. + +data in both the pre-training and finetuning processes. The dropout rate $p_{drop}$ is kept at 1 in the first 40 epochs then decreases to 0 linearly in the following 70 epochs. + +Bounding boxes. Bounding boxes $\mathbf{B}$ are a required input for the decoder. In ScanRefer, no ground-truth object boundary box are provided. Thus we use GroupFree [25] to detect a set of bounding boxes $\mathbf{B}_d$ , which $\mathbf{B}_d \in \mathbb{R}^{m \times 8 \times 3}$ . Following the experiment setting of EDA [35] and BUTDDETR [17], in SR3D and NR3D, the model takes the ground-truth boundary boxes to replace the detected ones as the input of the decoder. + +# 4.3. 3D Visual Grounding Results + +Analysis of ScanRefer Results. Our experimental results on ScanRefer are shown in Table 1. VPP-Net achieves state-of-the-art performance in terms of Acc@0.25 of "Overall" and "Unique", surpassing EDA by $1.06\%$ and $0.29\%$ respectively. An improvement of $1.19\%$ and $1.29\%$ can be observed at "Multi" cases to $50.32\%$ and $39.03\%$ separately. The "Multi" cases are more challenging than "Unique" cases generally because of the existence of distractors. This suggests that the 3D scene transformed with the predicted viewpoint provides less ambiguous spatial relations with which to reason about distractors. + +Among the listed methods, we draw attention to "BUTD + +Based" and "Multi-View" models in the table - BUTD-DETR [17], EDA [35], M3DRef-CLIP [45], MVT [16], and ViewRefer [13]. BUTD-based models leverage the same BUTD-DETR backbone we build on. Compared with BUTD-DETR, our proposed VVP-Net has a large margin of improvement across all evaluation metrics. Compared to EDA, we can still observe an improvement of around $1.04\%$ and $1.39\%$ in "Overall" and "Multi". Multi-View models involve multi-view scenes or images that are related to our method. VPP-Net outperforms MVT and ViewRefer across all metrics. Compared with M3DRef-CLIP, our model achieves comparable but less competitive performance in Acc@0.5 of "Unique" and Acc@0.5 of "Overall", while achieving better performance in all other cases, especially, the Acc@0.5 of "Multi" which is more challenging. We hypothesize this is due to the multi-view images rendered of object candidates used in M3DRef-CLIP. Multi-view rendered images provide detailed information about an object, allowing the model localize the object more accurately. The multi-view images focus on single objects and do not provide much benefit in the "Multi" case, since the model can be fooled by multiple objects with similar 2D visual attributes in a scene. Thus, we can observe a larger margin of Acc@0.5 of "Multi" between C3DRef-CLIP and ours (improved by $+2.23\%$ ). + +
MethodModalitySR3DNR3D
ReferIt3D [2]3D39.835.6
TGNN [15]3D45.037.3
TransRefer3D [14]3D57.442.1
InstanceRefer [44]3D48.038.8
3DVG-Transfor [46]3D51.440.8
FFL-3DOG [12]3D-41.7
SAT [43]3D+2D57.949.2
3DReferTrans [1]3D47.039.0
LanguageRefer [32]3D56.043.9
3D-SPS [26]3D+2D62.651.5
BUTD-basedEDA [35]3D68.152.1
BUTD-DETR [17]3D67.154.9
Multi ViewM3DRef-CLIP [45]3D+2D-49.4
LAR [3]3D59.648.9
VPP-Net(Ours)3D68.756.9
+ +Analysis of SR3D & NR3D Results. We report results for SR3D and NR3D datasets in Table 2. Compared with existing methods, VPP-Net achieves state-of-the-art on both. Our observed improvement on NR3D (+2.00%) is larger than we see in SR3D (0.6%). We argue that this may be because there is a large domain gap between SR3D and our synthetic dataset. To show the domain gap, we count the frequency of spatial-relation words in SR3D, NR3D, ScanRefer, and our synthetic dataset in Table 4. We find an obvious gap: proximity relations - "Closest/Farthest" - dominate in SR3D, while relative spatial relations - "Front/Behind" and "Left/Right" - are more prominent in the human-annotated NR3D and ScanRefer datasets as well as our synthetic dataset. Despite this, VPP-Net still boosts BUTD-DETR performance on SR3D (+1.6%). + +To explicitly evaluate the efficiency of VPP-Net on viewpoint-dependent samples, we report results on view dependent / independent (dep./indep.) subsets of NR3D in Table 3. We see large gains over our backbone Butd-Detr for view dep. cases (46 vs. $52.4\%$ ) while achieving a similar result in view indep. case (58 vs. $58.6\%$ ). In contrast, EDA's improvement for view dependent cases is smaller and results in performance loss on view independent cases. + +# 4.4. Ablation Study + +We report an ablation study of our method on ScanRefer in Table 5. We remove different modules, auxiliary losses, or training procedures from our approach to show their contri + +Table 2. Result of SR3D and NR3D datasets. We report Acc@0.25IoU as the evaluation metric. Best and 2nd best results are denoted in bold or underlined. + +
ModelView Dep.View Indep.
Butd-Detr [17]46.0%58.0%
EDA [31]50.2 %53.1%
VPPNet (ours)52.4 %58.6%
+ +Table 3. Results of Acc@0.25 on View Dep/Indep-endent cases of NR3D dataset. + +
RelationSynthetic(%)Natural(%)
SR3DOursNR3DScanRefer
Left/Right8.9244.5836.0055.66
Front/Behind2.5750.0819.3920.11
Above/Under3.815.3414.1913.09
Between8.540.003.216.77
Closest/Farthest81.380.0031.214.38
+ +Table 4. Normalized frequency of spatial relation words across datasets. We find SR3D has significantly fewer relative spatial relations than our synthetic datasets or either human-collected datasets on which we evaluate. + +
Synthetic PretrainingViewpoint PredictionUniform Obj. Rep.Curriculum FilteringViewpoint Data Aug.Acc@0.25Acc@0.50
0*-----52.139.8
1--✓✓-51.738.9
2✓✓-✓✓53.7838.98
3✓✓✓-✓51.5238.45
4✓✓✓✓-53.7840.41
5-✓✓✓✓49.2333.67
6✓✓✓✓✓55.6543.29
+ +Table 5. Ablation of our model components, auxiliary losses, and training methods on the ScanRefer dataset - checkmarks denote the corresponding item is active. Rows are numbered. + +bution. Note that row "0*" without any of our modifications is equivalent to BUTD-DETR. When viewpoint prediction is disabled (row 1), we also disable the modules that rely on it – pretraining and viewpoint data augmentation. + +We find that removing the viewpoint prediction and data augmentation (row 1), or the curriculum filtering during pretraining (row 3) results in performance similar to the baseline. These correspond to models that either do not use viewpoint prediction (row 1) or learn viewpoint prediction poorly (row 3) – suggesting the importance of viewpoint in 3D visual grounding. We also note that directly jointly training VPP-Net without first developing a strong viewpoint prediction capability during pretraining (row 5) yields worse results than not considering viewpoint at all. Removing either the uniform object representation loss (row 2) or the viewpoint data augmentation (row 4) yields degraded performance compared with the whole model (row 6). Both are designed to improve learned representations and we see they have a positive effect in these experiments. + +![](images/39a706bbac73097ff7e6a588be3bbfed38be5544414382f3c7a0537d05566c88.jpg) +Figure 3. Example VPP-Net results. The first row shows successful examples where both viewpoint prediction and visual grounding are correct. The second row shows failed samples. The first two predict incorrect viewpoints but still succeed in visual grounding. The predicted observer position (red dot) and facing direction (facing away from the green arrow) are shown in 3D and a top-down view (top right corner). We zoom up on the target region in example (7) within the red circle to provide better visualization. + +# 4.5. Qualitative Examples + +We depict the quantitative results of VPP-Net in Fig. 3. In the figure, we list 8 examples of which four are successful (top row) and four are not (bottom row). In each example, the red dot represents the predicted location and the green and blue arrows represent the "behind" and "right" direction predicted in the model such that the predicted observer is facing away from the green arrow. The ground truth bounding boxes and target words are noted with green and the mentioned objects are noted with red. We also provide the predicted object bounding box in the image, shown in blue. The spatial relations are noted with blue in the text. The successful examples show that with an accurate viewpoint prediction, the expression can better match the 3D scene, resulting in more accurate groundings. Failed examples (5) and (6) are successful at grounding despite having failed to predict valid viewpoints. Both contain ambiguous elements in text or image, i.e. the expression "walk thru the door on the right" in example (5) and the distorted door (blob on bottom left of the scene) in example (6). Examples (7) and (8) predict valid viewpoints but produce incorrect visual groundings. The predicted bounding box in example (7) covers a certain region of the ground truth and involved patch contains similar visual attributes (white). In example (8), both ground truth and prediction are on the left of the viewpoint, matching the expression. However, + +the model failed to understand the horizontal proximity (farthest), leading to a incorrect grounding. + +# 5. Conclusion + +In conclusion, we have proposed a Viewpoint Prediction Network (VPP-Net) for 3D visual grounding tasks. Our model addresses the critical challenge of viewpoint-dependent ambiguity in 3D visual grounding. Specifically, we first introduce a synthetic dataset and then use it to train a 3D visual grounding model that can better disambiguate referring expressions by explicitly predicting potential viewpoints. Further, we design a uniform object representation auxiliary loss and viewpoint data augmentation scheme that further improve the performance of our model. Experiments on the ScanRefer, SR3D, and NR3D demonstrate the effectiveness of our proposed methods. + +Limitations. The synthetic datasets we generate are highly dependent on the templated referring expressions, which limits their diversity and realism. Future works should include strategies to generate more diverse datasets or collect human-generated referring expressions with annotated viewpoints to further improve. + +# References + +[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dreftransformer: Fine-grained object identification in realworld scenes using natural language. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 7 +[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 422-440. Springer, 2020. 2, 5, 6, 7 +[3] Eslam Bakri, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 7 +[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 6 +[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 5, 6 +[6] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. 2021. 6 +[7] Sijia Chen and Baochun Li. Multi-modal dynamic graph transformer for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15534-15543, 2022. 1, 2 +[8] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 6 +[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 3 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[11] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 202-210. Springer, 2019. 2 +[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net + +work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1, 6, 7 +[13] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 1, 2, 6 +[14] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 7 +[15] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 6, 7 +[16] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1, 2, 6 +[17] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, pages 417-433. Springer, 2022. 1, 2, 3, 4, 5, 6, 7 +[18] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1, 2 +[19] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2 +[20] Weide Liu, Zhonghua Wu, Henghui Ding, Fayao Liu, Jie Lin, and Guosheng Lin. Few-shot segmentation with global and local contrastive learning. arXiv preprint arXiv:2108.05293, 2021. 2 +[21] Weide Liu, Zhonghua Wu, Yang Zhao, Yuming Fang, Chuan-Sheng Foo, Jun Cheng, and Guosheng Lin. Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation. International Journal of Computer Vision, pages 1-15, 2023. 2 +[22] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3 +[23] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 2 +[24] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual + +grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5612-5621, 2021. 1, 2 +[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2, 6 +[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 6, 7 +[27] Zhipeng Luo, Gongjie Zhang, Changqing Zhou, Zhonghua Wu, Qingyi Tao, Lewei Lu, and Shijian Lu. Modeling continuous motion for 3d point cloud object tracking. arXiv preprint arXiv:2303.07605, 2023. 2 +[28] Hui En Pang, Zhongang Cai, Lei Yang, Qingyi Tao, Zhonghua Wu, Tianwei Zhang, and Ziwei Liu. Towards robust and expressive whole-body human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. +[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[30] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[32] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7 +[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 2 +[34] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 1, 2 +[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 1, 6, 7 +[36] Yu Wu, Yana Wei, Haozhe Wang, Yongfei Liu, Sibei Yang, and Xuming He. Grounded image text matching with mismatched relation reasoning. In Proceedings of the IEEE/CVF + +International Conference on Computer Vision, pages 2976-2987, 2023. 1, 2 +[37] Zhonghua Wu, Guosheng Lin, and Jianfei Cai. Keypoint based weakly supervised human parsing. Image and Vision Computing, 91:103801, 2019. 2 +[38] Zhonghua Wu, Qingyi Tao, Guosheng Lin, and Jianfei Cai. Exploring bottom-up and top-down cues with attentive learning for webly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12936-12945, 2020. +[39] Zhonghua Wu, Xiangxi Shi, Guosheng Lin, and Jianfei Cai. Learning meta-class memory for few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 517-526, 2021. 2 +[40] Zhonghua Wu, Yicheng Wu, Guosheng Lin, Jianfei Cai, and Chen Qian. Dual adaptive transformations for weakly supervised point cloud segmentation. In European conference on computer vision, pages 78-96. Springer Nature Switzerland Cham, 2022. 2 +[41] Zhonghua Wu, Yicheng Wu, Guosheng Lin, and Jianfei Cai. Reliability-adaptive consistency regularization for weakly-supervised point cloud segmentation. International Journal of Computer Vision, pages 1-14, 2024. 2 +[42] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 2 +[43] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2, 6, 7 +[44] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 6, 7 +[45] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 6, 7 +[46] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 1, 6, 7 \ No newline at end of file diff --git a/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/images.zip b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..927e928c63b932d3a5faca193ba6f8e95ff2026a --- /dev/null +++ b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db8e1d13ad6315a675abca29158a5d0e5923660762b74ea8600a851b96e60764 +size 591248 diff --git a/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/layout.json b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e557f8941b24f4377f7d1df6901079ba111d6f5 --- /dev/null +++ b/2024/Viewpoint-Aware Visual Grounding in 3D Scenes/layout.json @@ -0,0 +1,8964 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 449, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 449, + 121 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 449, + 121 + ], + "type": "text", + "content": "Viewpoint-Aware Visual Grounding in 3D Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 144, + 200, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 144, + 200, + 185 + ], + "spans": [ + { + "bbox": [ + 80, + 144, + 200, + 185 + ], + "type": "text", + "content": "Xiangxi Shi \nOregon State University \nshixia@oregonstate.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 144, + 358, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 144, + 358, + 185 + ], + "spans": [ + { + "bbox": [ + 227, + 144, + 358, + 185 + ], + "type": "text", + "content": "Zhonghua Wu SenseTime Research wuzhonghua@sensetime.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 384, + 144, + 510, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 144, + 510, + 185 + ], + "spans": [ + { + "bbox": [ + 384, + 144, + 510, + 185 + ], + "type": "text", + "content": "Stefan Lee \nOregon State University \nleestef@oregonstate.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": "Referring expressions for visual objects often include descriptions of relative spatial arrangements to other objects - e.g. \"to the right of\" - that depend on the point of view of the speaker. In 2D referring expression tasks, this viewpoint is captured unambiguously in the image. However, grounding expressions with such spatial language in 3D without viewpoint annotations can be ambiguous. In this paper, we investigate the significance of viewpoint information in 3D visual grounding - introducing a model that explicitly predicts the speaker's viewpoint based on the referring expression and scene. We pretrain this model on a synthetically generated dataset that provides viewpoint annotations and then finetune on 3D referring expression datasets. Further, we introduce an auxiliary uniform object representation loss to encourage viewpoint invariance in learned object representations. We find that our proposed ViewPoint Prediction Network (VPP-Net) achieves state-of-the-art performance on ScanRefer, SR3D, and NR3D - improving Accuracy@0.25IoU by " + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "inline_equation", + "content": "1.06\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "inline_equation", + "content": "0.60\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "inline_equation", + "content": "2.00\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": " respectively compared to prior work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 501, + 128, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 128, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 128, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 617 + ], + "type": "text", + "content": "Visual grounding of referring expressions requires algorithms to reason about natural language object descriptions to identify object referents in visual scenes. Naturally, describing relative spatial relations between objects is a common strategy when a speaker produces a referring expression – e.g., “The chair to the right of the couch.” As such, many methods for visual grounding put algorithmic emphasis on handling these spatial relations [7, 18, 24, 34, 36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "However, these sorts of spatial relations depend upon the point of view of the speaker. For instance, Fig. 1 depicts two different viewpoints of the same 3D scene and notes that these views produce contradictory referring expressions. In 2D visual grounding tasks, the point of view of the speaker is typically also that of the image – introducing no ambiguity into spatial reasoning. However, viewpoint annotations are not typically captured for 3D visual grounding tasks that" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 214, + 422, + 346 + ], + "blocks": [ + { + "bbox": [ + 317, + 214, + 422, + 346 + ], + "lines": [ + { + "bbox": [ + 317, + 214, + 422, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 214, + 422, + 346 + ], + "type": "image", + "image_path": "bf6c096bcbbc885a1f52dc7855c0fbf3a77f727705e99c500519eeaa452d9108.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 355, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 305, + 355, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 546, + 378 + ], + "type": "text", + "content": "Figure 1. Expressions referring to the same object from different viewpoints can lead to contradictory spatial relations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 424, + 213, + 535, + 346 + ], + "blocks": [ + { + "bbox": [ + 424, + 213, + 535, + 346 + ], + "lines": [ + { + "bbox": [ + 424, + 213, + 535, + 346 + ], + "spans": [ + { + "bbox": [ + 424, + 213, + 535, + 346 + ], + "type": "image", + "image_path": "7ff6caed88ec5c442b930ac9232976e4f38a69100be06fd70279376febcce266.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 401, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 472 + ], + "type": "text", + "content": "reason about whole scenes. As a result, identifying referent objects may be ambiguous if multiple object instances align with the referring expression in differing views. Moreover, the lack of reliable grounding for spatial language may simply make the reasoning task more difficult to learn even when the referent is still uniquely identifiable." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 473, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 568 + ], + "type": "text", + "content": "Despite this, most of the recent work in 3D visual grounding has not addressed the role of viewpoint in the learning process [12, 17, 26, 35, 43, 46]. Those methods that have attempted to consider the influence of viewpoint have been hindered by the lack of viewpoint supervision in existing datasets – instead relying on rendering multiple views of a 3D scene and using attention mechanisms to softly select between them [13, 16] in an end-to-end model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In this work, we develop a straightforward approach to viewpoint-aware visual grounding in 3D scenes – explicitly predicting the speaker's viewpoint and then transforming the input accordingly. To support this direction, we create a synthetic visual grounding dataset that provides annotations to directly supervise viewpoint prediction. We propose a novel framework called ViewPoint Prediction Net (VPP-Net) to estimate the viewpoints of expression-pointcloud pairs and rotate the 3D scenes accordingly before performing expression grounding. This model is first pre-trained on the synthetic dataset and then fine-tuned on a combination of the synthetic and target datasets. To our knowledge, this" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14056" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "is the first work to introduce a supervised viewpoint estimation model for 3D visual grounding tasks. In experiments on three common 3D visual grounding datasets, this approach paired with viewpoint-related auxiliary losses leads to improved performance over prior art." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 279, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 279, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 279, + 144 + ], + "type": "text", + "content": "Contributions. Our contributions can be summarized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 146, + 286, + 276 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 146, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 286, + 180 + ], + "type": "text", + "content": "- We introduce viewpoint prediction as an auxiliary task in the 3D visual grounding task and propose synthetic data generation pipeline to provide viewpoint supervision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 182, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 217 + ], + "type": "text", + "content": "- We propose VPP-Net - a novel model that learns viewpoint prediction to transform 3D scenes to reduce ambiguity in spatial-relation grounding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "content": "- We design a Uniform Object Representation auxiliary loss that promotes viewpoint invariant feature learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 241, + 286, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 286, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 286, + 276 + ], + "type": "text", + "content": "- Combing these techniques, our proposed approach achieves state-of-the-art performance on the ScanRefer [5], SR3D [2], and NR3D [2] datasets." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 286, + 133, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 133, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 133, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 306, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 286, + 569 + ], + "type": "text", + "content": "Relation-aware 2D Visual Grounding. The visual grounding task has been extensively explored in 2D image scenarios. The goal of a 2D visual grounding model is to locate an object according to a given expression in a 2D image. In recent work, the relationships between objects have been identified as a crucial cue in complex environments. Several works [7, 23, 24, 34, 42] have parsed referring expressions into phrases for better text understanding. Wang et al. [34] proposed to align the generated visual graph with the self-attend object and relation features using a graph attention module. Liu [24] proposed to capture the visual context from a set of coarse-level object proposals by reconstructing the corresponding relations captured from the text. Chen [7] et al. initialize a visual graph with proposed node and edge transformers to learn the representations of objects and relations. Besides explicitly encoding visual relations in the model, relationships can also benefit models by constructing pseudo data [18, 36]. Wu et al. [36] proposed to detect mismatched relations from synthetic data to evaluate the ability of relation understanding of the model. Jiang et.al. [18] introduce an automatic way to generate pseudo visual grounding data for supervised training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": "3D Visual Grounding. The first dataset for visual grounding in 3D was introduced by Chen et al. [5] - presenting the ScanRefer dataset consisting of 3D scenes and natural expression. Achlioptas et al. [2] propose two datasets - the synthetic dataset SR3D and the human-annotated NR3D. Further, they show that template-based synthetic data can benefit the task when models are jointly trained with naturally collected data. In standard approaches, a 3D grounding model first encodes instructions and 3D scene with pretrained text [10, 22, 31] and 3D visual [19, 27-30, 40, 41] models. Then a multimodal module is proposed to fuse the encoded features to estimate the location and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "size of the grounded object. 3D-SPS [26] proposed a one-stage method that estimates the location of objects from the point cloud directly. In other works, 3D object detections [11, 19, 25, 33] are leveraged to provide an object-focused candidate pool consisting of object-level features and/or bounding boxes. BUTD-DETR [17] encodes object bounding boxes as a reference and jointly trains on object detection datasets as an auxiliary task. 2D images [20, 21, 37-39] are also utilized to boost the performance of 3D visual grounding. The model proposed in [5] encodes 2D images together with 3D scenes for a better 3D understanding. Likewise, Yang et al. [43] leverage 2D semantics to assist 3D representation learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 229, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 373 + ], + "type": "text", + "content": "The viewpoint of the annotator is also a significant cue for a better alignment between 3D scenes and referring expressions. Currently, only a few works focused on this aspect have been proposed. MVT [16] and ViewRefer [13] rotate the 3D scenes in different angles and encode them to multi-view representations without explicit supervision of which viewpoint matches to the annotator. Compared with previous work, we first provide a template-based synthetic visual grounding dataset including location and perspective supervision and then propose a supervised training method for viewpoint estimation. We demonstrate this is be beneficial for visual grounding performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 387, + 365, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 365, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 365, + 399 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 599 + ], + "type": "text", + "content": "We propose a straight-forward approach for learning to reason about viewpoint in 3D referring expression grounding – directly predicting a rotation and translation of the 3D environment that aligns it with the assumed viewpoint of the observer providing the referring expression. To accomplish this, we introduce a ViewPoint Prediction Network (VPP-Net) – a model that performs 3D referring expression grounding while explicitly estimating the location and heading of the observer. As existing datasets lack observer viewpoint annotations, we develop a synthetic data generation pipeline to provide appropriate training supervision for viewpoint-aware pretraining and then transfer the pretrained model to downstream tasks. Further, we introduce a Uniform Object Representation (UOR) auxiliary loss to ensure object instances are represented consistently regardless of viewpoint. We describe each component below." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 610, + 536, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 536, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 536, + 624 + ], + "type": "text", + "content": "3.1. Generating Synthetic Viewpoint Supervision" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Our synthetic data generation pipeline operates in four stages - for a given environment, we (1) sample a location and heading for a virtual observer, (2) identify spatial relationships between annotated objects in view, (3) generate templated referring expressions accordingly, and then (4) find other nearby viewpoints that could also have produced this reference. The result is a tuple of the 3D environment," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14057" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "referring expression, and valid viewpoint map which we can use for training viewpoint prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "text", + "content": "Input Scenes. We utilize 3D scenes from the ScanNet [9] dataset which are annotated with detected object bounding boxes. For a given scene, we denote bounding box as a set " + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "text", + "content": ", their centroids as a set of homogenous XYZ-coordinates " + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{c_0, c_1, \\dots, c_n \\mid c_i = (x_i, y_i, z_i, 1)\\}" + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "text", + "content": ", and the corresponding object category labels " + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_o" + }, + { + "bbox": [ + 47, + 100, + 286, + 172 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": "Viewpoint Sampling. Regardless of scene size, we split the horizontal (XY) plane into a " + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "inline_equation", + "content": "10 \\times 10" + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": " grid and uniformly sample a 2D " + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": " translation from the grid. Likewise, we sample a rotation " + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": " about the vertical axis by randomly selecting a " + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": " increment between " + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 175, + 286, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 286, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 275 + ], + "type": "text", + "content": "Identifying Spatial Relationships. We apply the sampled translation and rotation to the scene by computing updated bounding box centers. With the affine transformation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 283, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 283, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 72, + 283, + 287, + 334 + ], + "type": "interline_equation", + "content": "M (\\theta , x, y) = \\left( \\begin{array}{c c c c} \\cos \\theta & - \\sin \\theta & 0 & - x \\\\ \\sin \\theta & \\cos \\theta & 0 & - y \\\\ 0 & 0 & 1 & 0 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right), \\tag {1}", + "image_path": "98a86119d65cd2962ccd188fcd70041d68c110994dea3e0f279a38d527506d4e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": "we update the bounding box centroids as " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "c_i' = c_i M(\\theta, x, y)^T" + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": ". The effect of this operation is to align the observer's relative spatial terms (e.g. left/right) to the XYZ axes. To detect potential spatial relations between a pair of objects " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": ", we compute the vector " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "r_{ij} = c_i' - c_j'" + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": " to represent their relative position. We take a set of four unit direction vectors as " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "\\{d_{front}, d_{back}, d_{left}, d_{right}\\}" + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": " and compute the cosine similarities between " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 341, + 286, + 437 + ], + "type": "inline_equation", + "content": "r_{ij}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 445, + 287, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 445, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 117, + 445, + 287, + 471 + ], + "type": "interline_equation", + "content": "s _ {i, j, k} = \\frac {r _ {i j} d _ {k}}{\\left| \\left| r _ {i j} \\right| \\right| \\times \\left| \\left| d _ {k} \\right| \\right|}. \\tag {2}", + "image_path": "cccdf0d546c788c815cc0800fc54cf38e8f30cf69432b8ab5d286c70a1388ccf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "text", + "content": "We consider a spatial relation " + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "text", + "content": " to be valid if " + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "inline_equation", + "content": "s_{i,j,k} > 0.3" + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "text", + "content": ". For above and below relations, we instead consider the heights and overlap of the bounding boxes. From all valid tuples, we select a random subset as the relation tuple set " + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 478, + 286, + 549 + ], + "type": "text", + "content": ". We note that this process does not consider the \"visibility\" of objects from the observer location." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 554, + 286, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 286, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 286, + 650 + ], + "type": "text", + "content": "Expression Generation. Given R and object category annotations, we can produce a set of object referring expressions. For each object mentioned as part of a relation in R, we randomly sample up to four valid relations and convert each into simple templated text. For example, the tuple (\"chair\", \"left\", \"bed\") is converted to a string like \"The chair is to the left of the bed.\" These are then concatenated to form a multi-sentence referring expression for the object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Viewpoint Supervision. For a viewpoint " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "(x,y,\\theta)" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " and referring expression " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": ", we provide viewpoint supervision decomposed into location and heading prediction. In most cases, the randomly selected viewpoint is not the only one for which the referring expression would be valid - small" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "shifts in either location or perspective may not change the relationship between observed objects. For perspective, we produce a binary vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{Per}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " with 36 entries where " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{Per}[k]" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " is 1 if the referring expression " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " is valid from viewpoint " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "(x,y,k*10^{\\circ})" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". For location, we produce a " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "10\\times 10" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " binary matrix " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{Loc}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{Per}[i,j]" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " is 1 if the referring expression " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " is valid from viewpoint " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "(i,j,\\theta)" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". For both, we check validity by repeating the spatial relationship tests described earlier from the new viewpoint. We note that this approach considers only a subset of all possible viewpoints and assumes that perspectives valid at " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " are also valid for any point " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " for which the original perspective was valid." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 222, + 523, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 523, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 523, + 235 + ], + "type": "text", + "content": "3.2. Viewpoint-Aware 3D Grounding Network" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "type": "text", + "content": "As outlined in Fig. 2, our ViewPoint Prediction Network (VPP-Net) approach consists of a multimodal encoder-decoder architecture. After encoding the referring expression and scene, we predict the location and perspective of the observer. The encoded features and transformed candidate bounding boxes and XYZ point positions are then passed to a decoding model which ranks the candidates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 328, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 545, + 352 + ], + "type": "text", + "content": "VPP-Net Backbone Architecture. Our method is based on BUTD-DETR [17]. We review its construction here." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": "Input Encoding. An RGBXYZ point cloud representing the 3D scene is encoded with a pretrained PointNet++ [29], producing " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": " point features " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\text{vision}} \\in \\mathbb{R}^{n_p \\times d}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": " and the corresponding XYZ points " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{n_p \\times 3}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": ". Referring expressions are encoded with a pre-trained RoBerta [22] model into a sequence of text embeddings " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\text{text}} \\in \\mathbb{R}^{|T| \\times d}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "|T|" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": " is the token length." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": "Multimodal Encoder. The multimodal encoder consists of " + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": " layers of attention-based modules. Initializing visual features " + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "inline_equation", + "content": "f_{v}^{(0)}" + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": " and textual features " + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "inline_equation", + "content": "f_{t}^{(0)}" + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "inline_equation", + "content": "e_{vision}" + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "inline_equation", + "content": "e_{text}" + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": " respectively, each layer " + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 444, + 545, + 517 + ], + "type": "text", + "content": " computes update representations by applying per-modality self attention followed by cross-modality attention as below:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 522, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {v} ^ {(i + 1)}, \\mathbf {f} _ {t} ^ {(i + 1)} = f _ {\\text {C r o s s - A t t}} \\left(f _ {\\text {S e l f - A t t}} \\left(\\mathbf {f} _ {v} ^ {(i)}\\right), f _ {\\text {S e l f - A t t}} \\left(\\mathbf {f} _ {t} ^ {(i)}\\right)\\right). \\tag {3}", + "image_path": "e21afa4d75848b6ab38aab852349367f6cbb82f59a21b83e864a94144d9f8592.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "type": "text", + "content": "We denote the final layer outputs as " + }, + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{(N)}" + }, + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v^{(N)}" + }, + { + "bbox": [ + 306, + 545, + 511, + 560 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": "Object Decoder. Between the encoder and decoder, the visual features " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v^{(N)}" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": " are passed through as small MLP to produce scores for each visual input that are then passed through a softmax. The top-K scoring transformed features are retained as object queries " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ". The object decoder is an N-layer transformer-based module. The decoder takes " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{(N)}" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v^{(N)}" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ", hierarchical points " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_v" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": " and detected bounding boxes " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_o" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": " as input to predict " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 304, + 563, + 545, + 663 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 669, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 669, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 353, + 669, + 545, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {r}, \\mathbf {f} _ {o}, \\mathbf {f} _ {l}, \\mathbf {f} _ {s} = f _ {\\mathrm {D e c}} \\left(\\hat {\\mathbf {f}}, \\mathbf {f} _ {t}, \\mathbf {f} _ {v}, \\mathbf {c} _ {v}, \\mathbf {B} _ {o}\\right), \\tag {4}", + "image_path": "3fb7a8e0f682f174f3002bd8f9c11bd645bc8e2c2544a232b808f5521daeee65.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^{n_p \\times 1}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the object-expression matching score; " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o \\in \\mathbb{R}^{n_p \\times D}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " are object features; and " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l, \\mathbf{f}_s \\in \\mathbb{R}^{n_p \\times 3}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14058" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 548, + 262 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 548, + 262 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 548, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 548, + 262 + ], + "type": "image", + "image_path": "c329894d27a521b1efb9162494bcc562e7e89069af3231c8742914e1a09183a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": "Figure 2. An overview of our ViewPoint Predictor Network (VPP-Net) approach. (Left) Our model builds on the encoder-decoder structure of [17], but modifies the encoder to explicitly predict location and perspective of the observer. The decoder then processes appropriately transformed inputs to rank candidate bounding boxes. (Right) Our encoder model updates features of location and perspective at each layer that are supervised to predict viewpoints where the referring expression is valid." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "text", + "content": "the location and size of objects. The object with the highest " + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "text", + "content": " is chosen and its corresponding " + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{l,i}" + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{s,i}" + }, + { + "bbox": [ + 47, + 336, + 287, + 373 + ], + "type": "text", + "content": " are used to compute the bounding box for the object reference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "type": "text", + "content": "Adding Viewpoint Prediction. As shown in Fig. 2, our approach extends this architecture in two ways – by modifying the encoder to enable location/perspective prediction and by using these predictions to transform bounding boxes " + }, + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "type": "text", + "content": " and point coordinates " + }, + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 386, + 287, + 446 + ], + "type": "text", + "content": " prior to invoking the decoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "content": "We add two transformer-based predictors into each multimodal encoder layer shown in Fig. 2 (right). As both perspective and location features are updated in the same fashion, we write the update steps below for location only:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 498, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 498, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 102, + 498, + 287, + 515 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {l o c, v} ^ {(i)} = f _ {\\text {A t t n}} \\left(\\mathbf {f} _ {l o c, t} ^ {(i - 1)}, \\mathbf {f} _ {v} ^ {(i)}, \\mathbf {f} _ {v} ^ {(i)}\\right), \\tag {5}", + "image_path": "015ab58aba0752d9dc28f08229da87238a66baeb6c5e8993142049f83d355eb1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 517, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 517, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {l o c, t} ^ {(i)} = f _ {\\text {A t t n}} \\left(\\mathbf {f} _ {l o c, v} ^ {(i)}, \\mathbf {f} _ {t} ^ {(i)}, \\mathbf {f} _ {t} ^ {(i)}\\right), \\tag {6}", + "image_path": "3ca62e0d8baf0a9b68027bf58205f3368ac3e990d0a42b99420cf836d0fcdede.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Attn}}(\\cdot ,\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": " denotes multi-head attention with key, query, and value arguments. The location feature is updated each layer by querying visual features and then querying textual features. We initialize the perspective and location features " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{per,t}^{(0)}" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": " with learnable embeddings " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "e_{per}" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{loc,t}^{(0)}" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "e_{loc}" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": " - a linear mapping of viewpoint " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": ". Denote the final representations as " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{per,t}^{(N)}" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{loc,t}^{(N)}" + }, + { + "bbox": [ + 46, + 537, + 287, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "text", + "content": "We predict a distribution over locations and perspective from each layer's location and perspective features respectively using a learned linear layer. We denote our predictions over locations and perspective respectively as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 677, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 677, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 677, + 287, + 693 + ], + "type": "interline_equation", + "content": "P (l o c \\mid \\mathbf {f} _ {l o c, m} ^ {(i)}) = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {l o c} \\mathbf {f} _ {l o c, m} ^ {(i)} + \\mathbf {b} _ {l o c}\\right) \\tag {7}", + "image_path": "c18031825a0e261cc2056a5f812f65575c7b05644dbbd9a0aabdb009e82cb32f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 696, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 696, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 63, + 696, + 287, + 712 + ], + "type": "interline_equation", + "content": "P (p e r \\mid \\mathbf {f} _ {p e r, m} ^ {(i)}) = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {l o c} \\mathbf {f} _ {p e r, m} ^ {(i)} + \\mathbf {b} _ {p e r}\\right) \\tag {8}", + "image_path": "6f44e96278db87fdb94da1bc5ffed3d648e8949d2e2581e8692bfa927c14aa1f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 336, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 336, + 545, + 373 + ], + "type": "inline_equation", + "content": "m \\in \\{\\mathrm{v},\\mathrm{t}\\}" + }, + { + "bbox": [ + 305, + 336, + 545, + 373 + ], + "type": "text", + "content": ". We supervise each of these predictions with synthetic viewpoint supervision using an averaged cross-entropy loss - denoting it as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 383, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 383, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 370, + 383, + 545, + 396 + ], + "type": "interline_equation", + "content": "L _ {a, m} = C E \\left(\\mathcal {Y} _ {a}, P _ {i} (a | m)\\right) \\tag {9}", + "image_path": "a7565586c0cbb33fb4e7576b7f03c21c2ea792eec4f7ced600dcf94b5f72a971.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "type": "inline_equation", + "content": "a \\in \\{per, loc\\}" + }, + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_a" + }, + { + "bbox": [ + 305, + 405, + 515, + 418 + ], + "type": "text", + "content": " as defined in Sec. 3.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": "Object Decoder with Affine Transformation. After encoding the input, we take the highest probability location " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "(\\hat{x},\\hat{y})" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " and perspective " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " predictions generated from " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "f_{loc,t}^{(N)}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "f_{per,t}^{(N)}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " as an estimate of the observers viewpoint. Given these, we apply the corresponding affine transform " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "M(\\hat{\\theta},\\hat{x},\\hat{y})" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " to the point locations " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " and bounding boxes " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "B-" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " denoting the transformed versions as " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "p^{\\prime}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "B^{\\prime}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " respectively. These transformed versions along with the other features are then passed to the decoder as in Eq. 4. After the decoder predicts a bounding box location " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^{\\prime}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " and size " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s^{\\prime}" + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "text", + "content": " in this transformed coordinate system, we transform the prediction back to the original scene using the inverse of " + }, + { + "bbox": [ + 304, + 422, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\hat{\\theta},\\hat{x},\\hat{y})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 575, + 539, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 539, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 539, + 588 + ], + "type": "text", + "content": "3.3. Encouraging Uniform Object Representation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "An object can be described differently in different viewpoints - potentially causing our model's representation of the same object to be quite different from different viewpoints. To encourage our model to be more robust to these difference, we add an auxiliary loss that encourages the decoder output " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " for object " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " to be consistent across viewpoints. To do this, we introduce a set of instance-wise learnable vectors " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = \\{u_0, u_1, \\ldots\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": ". We refer to these as uniform object representations. For a training sample whose target is " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": ", we concatenate the predicted object features " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{o}" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14059" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "text", + "content": "and final perspective embeddings " + }, + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "inline_equation", + "content": "f_{per,t}^{(N)}" + }, + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "text", + "content": " to estimate the uniform object representation " + }, + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "inline_equation", + "content": "u_{o}" + }, + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "text", + "content": " with a linear layer. The loss is computed between the predicted representation and " + }, + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "inline_equation", + "content": "u_{o}" + }, + { + "bbox": [ + 47, + 70, + 287, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 115, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 115, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 115, + 287, + 133 + ], + "type": "interline_equation", + "content": "L _ {l 1} = | | \\tanh (\\mathbf {W} _ {f u s e} [ f _ {o}, f _ {p e r, t} ^ {(N)} ] + \\mathbf {b} _ {f u s e}) - u _ {o} | | _ {1} \\quad (1 0)", + "image_path": "46429c5c9132c5296c66b710df931c75be54a664823c626a75399a114222c99b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 138, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 198 + ], + "type": "text", + "content": "This encourages representation to be predictable from object features and viewpoint information – intuitively, predicting an object instance's identity from its observation at a specific viewpoint. Note that these learnable vectors are not used at inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "text", + "content": "Further, we also use " + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "inline_equation", + "content": "u_{o}" + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "text", + "content": " to classify the object's class to encourage semantic consistency. For both " + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "inline_equation", + "content": "u_{o}" + }, + { + "bbox": [ + 47, + 198, + 287, + 246 + ], + "type": "text", + "content": ", we make a prediction over the object class and compute a Cross-Entropy loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 253, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 253, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 67, + 253, + 287, + 268 + ], + "type": "interline_equation", + "content": "L _ {c l s} = C E \\left(\\mathcal {Y} _ {o}, \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {s e m} \\phi + \\mathbf {b} _ {s e m}\\right)\\right), \\tag {11}", + "image_path": "6127ceecf5f306ef19f042946e460584ba35984d84978bed5aaff5bd669e3b13.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "type": "inline_equation", + "content": "\\phi \\in \\{f_o,u_o\\}" + }, + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_o" + }, + { + "bbox": [ + 47, + 274, + 241, + 286 + ], + "type": "text", + "content": " is the object's class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 292, + 144, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 144, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 144, + 305 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "text", + "content": "The overall training loss is a weight sum of our location and perspective supervision loss " + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "inline_equation", + "content": "L_{i,a,m}" + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "text", + "content": ", the uniform object representation loss " + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "inline_equation", + "content": "L_{l1}" + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "text", + "content": ", the object classification loss " + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "text", + "content": " and the grounding loss " + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "inline_equation", + "content": "L_{gr}" + }, + { + "bbox": [ + 47, + 311, + 287, + 360 + ], + "type": "text", + "content": " proposed in BUTD-DETR [17]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 366, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 366, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 64, + 366, + 287, + 392 + ], + "type": "interline_equation", + "content": "L = \\alpha_ {1} \\sum_ {a, m} L _ {a, m} + \\alpha_ {2} L _ {l 1} + \\alpha_ {3} L _ {c l s} + \\alpha_ {4} L _ {g r} \\tag {12}", + "image_path": "97ad1c26b5e6254097b39b5aebba7229d0b7a01d7dfa88d1980db9f6cca35dda.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "type": "text", + "content": "where the coefficients " + }, + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "type": "text", + "content": "'s are hyperparameters. Additionally, we find two additional techniques to be useful." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "content": "Synthetic Training Curriculum. While training on the synthetic data for viewpoint supervision, we apply a filtering mechanism on visual features " + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v^{(0)}" + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "content": " that reduces the task complexity of early training examples. We retain all points (and associated features) that occur within the bounding box of an object mentioned in the referring expression. For points outside these boxes, we drop the point with probability " + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "inline_equation", + "content": "p_{drop}" + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "content": ". At the start of training, we set " + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "inline_equation", + "content": "p_{drop} = 1 - \\text{effectively removing the background and other objects from the scene. As training progresses, we anneal } p_{drop} \\text{ to 0 such that whole scenes are observed.}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "type": "text", + "content": "Viewpoint Data Augmentation. During both synthetic training and finetuning, we apply a viewpoint-based augmentation scheme. After running an example through our model, we retain the predicted affine transformation " + }, + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "type": "text", + "content": ". We then rotate the scene's pointcloud and supervision by " + }, + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 47, + 563, + 287, + 647 + ], + "type": "text", + "content": " to create a viewpoint augmented sample. We then run the model again with this sample." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 658, + 190, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 190, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 190, + 671 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "To study the effectiveness of our proposed VPP-Net, we conduct experiments on three existing 3D visual grounding datasets - ScanRefer [5], SR3D [2], and NR3D [2]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 495, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 495, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 495, + 85 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 94, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 94, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 94, + 545, + 230 + ], + "type": "text", + "content": "Datasets. We evaluate on ScanRefer, SR3D, and NR3D which we describe below: ScanRefer. The ScanRefer dataset is built on 800 3D scenes collected from ScanNet. The 3D scenes are labeled with about 51,000 natural expressions and 11,000 objects. The labeled natural expressions have different types of phrases, including intra/inter-class spatial relations and comparatives/superlatives. The dataset also provides real images of the scene, which are used to enhance the grounding performance in some methods. We use the official splits provided by ScanRefer to train and evaluate our model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "content": "SR3D and NR3D. SR3D and NR3D are two datasets based on the ScanNet scenes proposed by Achlioptas et al. [2]. SR3D consists of 83K expressions generated by a template-based text generator. The expressions contain five types of spatial object-to-object relations: Horizontal/Vertical Proximity, Between, Allocentric, and Support, taking up " + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "inline_equation", + "content": "81.37\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "inline_equation", + "content": "3.80\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "inline_equation", + "content": "4.50\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "inline_equation", + "content": "8.54\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 390 + ], + "type": "text", + "content": " separately of all expressions. NR3D is a 3D visual grounding dataset with human-annotated expressions similar to ScanRefer. It consists of 41.5K human-annotated expressions and distractors collected through a \"player-listener\" game. Note that both SR3D and NR3D provide ground-truth bounding boxes for all candidate objects, unlike ScanRefer." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "type": "text", + "content": "Experimental Settings. We evaluate our proposed VPP-Net on these datasets and report results in Tables 1 and 2. For ScanRefer, we report the top-1 accuracy when IoU is larger than 0.25 and 0.5 in \"Unique\", \"Multi\" and \"Overall\" cases. \"Unique\" vs \"Multi\" indicate whether distractors exist. \"Modality\" shows whether a method gets benefit from 2D images, with \"3D\" meaning the model only takes the 3D point cloud and \"3D+2D\" indicating that 2D images are also used. For SR3D and NR3D, following the experiment setting of EDA, we report Acc@0.25IoU." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 522, + 438, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 438, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 438, + 534 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "type": "text", + "content": "Synthetic data. We select 557 and 140 scenes from ScanNet to build the train and validation set of synthetic datasets. With the proposed synthetic viewpoint supervision mechanism, we repeat the generation process 75 times for each scene to generate the synthetic dataset – resulting in 36117 training and 9167 validation samples." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Model configuration. VPP-Net contains a 6-layer encoder and decoder. The learning rate of for the PointNet++ and Roberta is " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "2e^{-3}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ". The learning rate of other modules is " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ". We first pre-train the model on the synthetic dataset to learn the viewpoint estimation for 150 epochs. Then we fine-tune using a combination of synthetic and the downstream dataset (Scanrefer/NR3D/SR3D) for another 150 epochs. We apply the point dropout mechanism on the synthetic" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14060" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 325 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 325 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 325 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 325 + ], + "type": "table", + "html": "
MethodModalityUnique(%)Multi(%)Overall(%)
Acc@0.25Acc@0.50Acc@0.25Acc@0.50Acc@0.25Acc@0.50
ScanRefer [5]2+3D76.3353.5132.7321.1141.1927.40
ReferIt3D [2]3D53.837.521.012.826.416.9
TGNN [15]3D68.6156.8029.8423.1837.3729.70
InstanceRefer [44]3D77.4566.8331.2724.7740.2332.93
SAT [43]2+3D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]3D78.8067.9435.1925.7041.3334.01
3DVG-Transformer [46]2+3D81.9360.6439.3028.4247.5734.67
3D-SPS [26]2+3D84.1266.7240.3229.8248.8236.98
3DJCG [4]2+3D83.4764.3441.3930.8249.5637.33
D3Net [6]3D+2D-70.35-30.50-37.87
UniT3D [8]3D+2D82.7573.1436.3631.0545.2739.14
BUTD-BasedBUTD-DETR [17]3D84.266.346.635.152.239.8
EDA [35]3D85.7668.5749.1337.6454.5942.26
Multi-ViewM3DRef-CLIP [45]3D+2D85.377.243.836.851.944.7
MVT [16]3D+2D--31.4624.8539.9532.28
ViewRefer [13]3D+2D--33.0826.5041.3033.66
VPP-Net (Ours)3D86.0567.0950.3239.0355.6543.29
", + "image_path": "79d420a0eaca8ef18c1cc14a8d7a84d1f82289e1437311dac5ef06469b9359d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 334, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 546, + 377 + ], + "type": "text", + "content": "Table 1. 3D visual grounding results on ScanRefer. Acc@0.25/0.5 means top-1 accuracy when IoU is larger than 0.25/0.5 - higher is better. We group BUTD-Based and Multi-View methods for ease of comparison. BUTD-based use the same backbone of our model. The Multi-View methods all explored notions of viewpoint or 2D image rendering in 3D visual grounding. The state-of-the-art methods are bold and the second best performance is underlined in the table." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 399, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 435 + ], + "type": "text", + "content": "data in both the pre-training and finetuning processes. The dropout rate " + }, + { + "bbox": [ + 46, + 399, + 287, + 435 + ], + "type": "inline_equation", + "content": "p_{drop}" + }, + { + "bbox": [ + 46, + 399, + 287, + 435 + ], + "type": "text", + "content": " is kept at 1 in the first 40 epochs then decreases to 0 linearly in the following 70 epochs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "text", + "content": "Bounding boxes. Bounding boxes " + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "text", + "content": " are a required input for the decoder. In ScanRefer, no ground-truth object boundary box are provided. Thus we use GroupFree [25] to detect a set of bounding boxes " + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_d" + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "text", + "content": ", which " + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_d \\in \\mathbb{R}^{m \\times 8 \\times 3}" + }, + { + "bbox": [ + 46, + 440, + 287, + 536 + ], + "type": "text", + "content": ". Following the experiment setting of EDA [35] and BUTDDETR [17], in SR3D and NR3D, the model takes the ground-truth boundary boxes to replace the detected ones as the input of the decoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 548, + 206, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 206, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 206, + 561 + ], + "type": "text", + "content": "4.3. 3D Visual Grounding Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": "Analysis of ScanRefer Results. Our experimental results on ScanRefer are shown in Table 1. VPP-Net achieves state-of-the-art performance in terms of Acc@0.25 of \"Overall\" and \"Unique\", surpassing EDA by " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "1.06\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "0.29\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " respectively. An improvement of " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "1.29\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " can be observed at \"Multi\" cases to " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "50.32\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "39.03\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " separately. The \"Multi\" cases are more challenging than \"Unique\" cases generally because of the existence of distractors. This suggests that the 3D scene transformed with the predicted viewpoint provides less ambiguous spatial relations with which to reason about distractors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Among the listed methods, we draw attention to \"BUTD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "text", + "content": "Based\" and \"Multi-View\" models in the table - BUTD-DETR [17], EDA [35], M3DRef-CLIP [45], MVT [16], and ViewRefer [13]. BUTD-based models leverage the same BUTD-DETR backbone we build on. Compared with BUTD-DETR, our proposed VVP-Net has a large margin of improvement across all evaluation metrics. Compared to EDA, we can still observe an improvement of around " + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "inline_equation", + "content": "1.04\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "inline_equation", + "content": "1.39\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "text", + "content": " in \"Overall\" and \"Multi\". Multi-View models involve multi-view scenes or images that are related to our method. VPP-Net outperforms MVT and ViewRefer across all metrics. Compared with M3DRef-CLIP, our model achieves comparable but less competitive performance in Acc@0.5 of \"Unique\" and Acc@0.5 of \"Overall\", while achieving better performance in all other cases, especially, the Acc@0.5 of \"Multi\" which is more challenging. We hypothesize this is due to the multi-view images rendered of object candidates used in M3DRef-CLIP. Multi-view rendered images provide detailed information about an object, allowing the model localize the object more accurately. The multi-view images focus on single objects and do not provide much benefit in the \"Multi\" case, since the model can be fooled by multiple objects with similar 2D visual attributes in a scene. Thus, we can observe a larger margin of Acc@0.5 of \"Multi\" between C3DRef-CLIP and ours (improved by " + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "inline_equation", + "content": "+2.23\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 698 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14061" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 289, + 309 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 309 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 309 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 309 + ], + "type": "table", + "html": "
MethodModalitySR3DNR3D
ReferIt3D [2]3D39.835.6
TGNN [15]3D45.037.3
TransRefer3D [14]3D57.442.1
InstanceRefer [44]3D48.038.8
3DVG-Transfor [46]3D51.440.8
FFL-3DOG [12]3D-41.7
SAT [43]3D+2D57.949.2
3DReferTrans [1]3D47.039.0
LanguageRefer [32]3D56.043.9
3D-SPS [26]3D+2D62.651.5
BUTD-basedEDA [35]3D68.152.1
BUTD-DETR [17]3D67.154.9
Multi ViewM3DRef-CLIP [45]3D+2D-49.4
LAR [3]3D59.648.9
VPP-Net(Ours)3D68.756.9
", + "image_path": "228438ec7006f885d5df16de09a7edef91f4a60b5b5d766715586b9988454db4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 372, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 552 + ], + "type": "text", + "content": "Analysis of SR3D & NR3D Results. We report results for SR3D and NR3D datasets in Table 2. Compared with existing methods, VPP-Net achieves state-of-the-art on both. Our observed improvement on NR3D (+2.00%) is larger than we see in SR3D (0.6%). We argue that this may be because there is a large domain gap between SR3D and our synthetic dataset. To show the domain gap, we count the frequency of spatial-relation words in SR3D, NR3D, ScanRefer, and our synthetic dataset in Table 4. We find an obvious gap: proximity relations - \"Closest/Farthest\" - dominate in SR3D, while relative spatial relations - \"Front/Behind\" and \"Left/Right\" - are more prominent in the human-annotated NR3D and ScanRefer datasets as well as our synthetic dataset. Despite this, VPP-Net still boosts BUTD-DETR performance on SR3D (+1.6%)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "text", + "content": "To explicitly evaluate the efficiency of VPP-Net on viewpoint-dependent samples, we report results on view dependent / independent (dep./indep.) subsets of NR3D in Table 3. We see large gains over our backbone Butd-Detr for view dep. cases (46 vs. " + }, + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "inline_equation", + "content": "52.4\\%" + }, + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "text", + "content": " ) while achieving a similar result in view indep. case (58 vs. " + }, + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "inline_equation", + "content": "58.6\\%" + }, + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "text", + "content": "). In contrast, EDA's improvement for view dependent cases is smaller and results in performance loss on view independent cases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 658, + 142, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 142, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 142, + 672 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We report an ablation study of our method on ScanRefer in Table 5. We remove different modules, auxiliary losses, or training procedures from our approach to show their contri" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 533, + 131 + ], + "blocks": [ + { + "bbox": [ + 47, + 316, + 288, + 351 + ], + "lines": [ + { + "bbox": [ + 47, + 316, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 288, + 351 + ], + "type": "text", + "content": "Table 2. Result of SR3D and NR3D datasets. We report Acc@0.25IoU as the evaluation metric. Best and 2nd best results are denoted in bold or underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 533, + 131 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 533, + 131 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 533, + 131 + ], + "type": "table", + "html": "
ModelView Dep.View Indep.
Butd-Detr [17]46.0%58.0%
EDA [31]50.2 %53.1%
VPPNet (ours)52.4 %58.6%
", + "image_path": "74b7aee89b1ef4426f9df46e49bdaf4ddc089f5ed9605ca43032baa65b6c1fb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 178, + 547, + 264 + ], + "blocks": [ + { + "bbox": [ + 306, + 133, + 547, + 154 + ], + "lines": [ + { + "bbox": [ + 306, + 133, + 547, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 547, + 154 + ], + "type": "text", + "content": "Table 3. Results of Acc@0.25 on View Dep/Indep-endent cases of NR3D dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 178, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 547, + 264 + ], + "type": "table", + "html": "
RelationSynthetic(%)Natural(%)
SR3DOursNR3DScanRefer
Left/Right8.9244.5836.0055.66
Front/Behind2.5750.0819.3920.11
Above/Under3.815.3414.1913.09
Between8.540.003.216.77
Closest/Farthest81.380.0031.214.38
", + "image_path": "1a448812f42e1a5626c3d0bcd16fe32c1a71d8f8fe4c52f6233e4c262a6416d0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 339, + 545, + 405 + ], + "blocks": [ + { + "bbox": [ + 305, + 271, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 547, + 316 + ], + "type": "text", + "content": "Table 4. Normalized frequency of spatial relation words across datasets. We find SR3D has significantly fewer relative spatial relations than our synthetic datasets or either human-collected datasets on which we evaluate." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 339, + 545, + 405 + ], + "lines": [ + { + "bbox": [ + 310, + 339, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 310, + 339, + 545, + 405 + ], + "type": "table", + "html": "
Synthetic PretrainingViewpoint PredictionUniform Obj. Rep.Curriculum FilteringViewpoint Data Aug.Acc@0.25Acc@0.50
0*-----52.139.8
1--✓✓-51.738.9
2✓✓-✓✓53.7838.98
3✓✓✓-✓51.5238.45
4✓✓✓✓-53.7840.41
5-✓✓✓✓49.2333.67
6✓✓✓✓✓55.6543.29
", + "image_path": "934afe934a6b6f24e76e37832e3efbda2465e90c042ca0ec005e24ed2aebc3e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 412, + 546, + 446 + ], + "lines": [ + { + "bbox": [ + 305, + 412, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 546, + 446 + ], + "type": "text", + "content": "Table 5. Ablation of our model components, auxiliary losses, and training methods on the ScanRefer dataset - checkmarks denote the corresponding item is active. Rows are numbered." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 479, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 529 + ], + "type": "text", + "content": "bution. Note that row \"0*\" without any of our modifications is equivalent to BUTD-DETR. When viewpoint prediction is disabled (row 1), we also disable the modules that rely on it – pretraining and viewpoint data augmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": "We find that removing the viewpoint prediction and data augmentation (row 1), or the curriculum filtering during pretraining (row 3) results in performance similar to the baseline. These correspond to models that either do not use viewpoint prediction (row 1) or learn viewpoint prediction poorly (row 3) – suggesting the importance of viewpoint in 3D visual grounding. We also note that directly jointly training VPP-Net without first developing a strong viewpoint prediction capability during pretraining (row 5) yields worse results than not considering viewpoint at all. Removing either the uniform object representation loss (row 2) or the viewpoint data augmentation (row 4) yields degraded performance compared with the whole model (row 6). Both are designed to improve learned representations and we see they have a positive effect in these experiments." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14062" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 547, + 315 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 315 + ], + "type": "image", + "image_path": "39a706bbac73097ff7e6a588be3bbfed38be5544414382f3c7a0537d05566c88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 329, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 329, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 547, + 376 + ], + "type": "text", + "content": "Figure 3. Example VPP-Net results. The first row shows successful examples where both viewpoint prediction and visual grounding are correct. The second row shows failed samples. The first two predict incorrect viewpoints but still succeed in visual grounding. The predicted observer position (red dot) and facing direction (facing away from the green arrow) are shown in 3D and a top-down view (top right corner). We zoom up on the target region in example (7) within the red circle to provide better visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 394, + 171, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 171, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 171, + 407 + ], + "type": "text", + "content": "4.5. Qualitative Examples" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 290, + 715 + ], + "type": "text", + "content": "We depict the quantitative results of VPP-Net in Fig. 3. In the figure, we list 8 examples of which four are successful (top row) and four are not (bottom row). In each example, the red dot represents the predicted location and the green and blue arrows represent the \"behind\" and \"right\" direction predicted in the model such that the predicted observer is facing away from the green arrow. The ground truth bounding boxes and target words are noted with green and the mentioned objects are noted with red. We also provide the predicted object bounding box in the image, shown in blue. The spatial relations are noted with blue in the text. The successful examples show that with an accurate viewpoint prediction, the expression can better match the 3D scene, resulting in more accurate groundings. Failed examples (5) and (6) are successful at grounding despite having failed to predict valid viewpoints. Both contain ambiguous elements in text or image, i.e. the expression \"walk thru the door on the right\" in example (5) and the distorted door (blob on bottom left of the scene) in example (6). Examples (7) and (8) predict valid viewpoints but produce incorrect visual groundings. The predicted bounding box in example (7) covers a certain region of the ground truth and involved patch contains similar visual attributes (white). In example (8), both ground truth and prediction are on the left of the viewpoint, matching the expression. However," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 394, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 420 + ], + "type": "text", + "content": "the model failed to understand the horizontal proximity (farthest), leading to a incorrect grounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 450, + 379, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 379, + 463 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 379, + 463 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 477, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 622 + ], + "type": "text", + "content": "In conclusion, we have proposed a Viewpoint Prediction Network (VPP-Net) for 3D visual grounding tasks. Our model addresses the critical challenge of viewpoint-dependent ambiguity in 3D visual grounding. Specifically, we first introduce a synthetic dataset and then use it to train a 3D visual grounding model that can better disambiguate referring expressions by explicitly predicting potential viewpoints. Further, we design a uniform object representation auxiliary loss and viewpoint data augmentation scheme that further improve the performance of our model. Experiments on the ScanRefer, SR3D, and NR3D demonstrate the effectiveness of our proposed methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Limitations. The synthetic datasets we generate are highly dependent on the templated referring expressions, which limits their diversity and realism. Future works should include strategies to generate more diverse datasets or collect human-generated referring expressions with annotated viewpoints to further improve." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14063" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dreftransformer: Fine-grained object identification in realworld scenes using natural language. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 223 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 422-440. Springer, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[3] Eslam Bakri, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "text", + "content": "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "type": "text", + "content": "[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "text", + "content": "[6] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: a speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "type": "text", + "content": "[7] Sijia Chen and Baochun Li. Multi-modal dynamic graph transformer for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15534-15543, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 513 + ], + "type": "text", + "content": "[8] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 286, + 567 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 679 + ], + "type": "text", + "content": "[11] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 202-210. Springer, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "text", + "content": "[13] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 218 + ], + "type": "text", + "content": "[14] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[15] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[16] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "text", + "content": "[17] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, pages 417-433. Springer, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "text", + "content": "[18] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "text", + "content": "[19] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 489, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 532 + ], + "type": "text", + "content": "[20] Weide Liu, Zhonghua Wu, Henghui Ding, Fayao Liu, Jie Lin, and Guosheng Lin. Few-shot segmentation with global and local contrastive learning. arXiv preprint arXiv:2108.05293, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "type": "text", + "content": "[21] Weide Liu, Zhonghua Wu, Yang Zhao, Yuming Fang, Chuan-Sheng Foo, Jun Cheng, and Guosheng Lin. Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation. International Journal of Computer Vision, pages 1-15, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "text", + "content": "[22] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "type": "text", + "content": "[23] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[24] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14064" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5612-5621, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "type": "text", + "content": "[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[27] Zhipeng Luo, Gongjie Zhang, Changqing Zhou, Zhonghua Wu, Qingyi Tao, Lewei Lu, and Shijian Lu. Modeling continuous motion for 3d point cloud object tracking. arXiv preprint arXiv:2303.07605, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "text", + "content": "[28] Hui En Pang, Zhongang Cai, Lei Yang, Qingyi Tao, Zhonghua Wu, Tianwei Zhang, and Ziwei Liu. Towards robust and expressive whole-body human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "text", + "content": "[30] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 469 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[32] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "text", + "content": "[34] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 680 + ], + "type": "text", + "content": "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 1, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[36] Yu Wu, Yana Wei, Haozhe Wang, Yongfei Liu, Sibei Yang, and Xuming He. Grounded image text matching with mismatched relation reasoning. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 608 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 2976-2987, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[37] Zhonghua Wu, Guosheng Lin, and Jianfei Cai. Keypoint based weakly supervised human parsing. Image and Vision Computing, 91:103801, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "text", + "content": "[38] Zhonghua Wu, Qingyi Tao, Guosheng Lin, and Jianfei Cai. Exploring bottom-up and top-down cues with attentive learning for webly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12936-12945, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[39] Zhonghua Wu, Xiangxi Shi, Guosheng Lin, and Jianfei Cai. Learning meta-class memory for few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 517-526, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "text", + "content": "[40] Zhonghua Wu, Yicheng Wu, Guosheng Lin, Jianfei Cai, and Chen Qian. Dual adaptive transformations for weakly supervised point cloud segmentation. In European conference on computer vision, pages 78-96. Springer Nature Switzerland Cham, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[41] Zhonghua Wu, Yicheng Wu, Guosheng Lin, and Jianfei Cai. Reliability-adaptive consistency regularization for weakly-supervised point cloud segmentation. International Journal of Computer Vision, pages 1-14, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "text", + "content": "[42] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "text", + "content": "[43] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 432, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 497 + ], + "type": "text", + "content": "[44] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 551 + ], + "type": "text", + "content": "[45] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 554, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 608 + ], + "type": "text", + "content": "[46] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 1, 6, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14065" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_content_list.json b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..59deb2ac88334615f663b99a9c921b7107d81a84 --- /dev/null +++ b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_content_list.json @@ -0,0 +1,1498 @@ +[ + { + "type": "text", + "text": "Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahan Li $^{1}$ , Jiuyang Dong $^{1}$ , Shenjin Huang $^{1}$ , Xi Li $^{3}$ , Junjun Jiang $^{1}$ , Xiaopeng Fan $^{1*}$ , Yongbing Zhang $^{2*}$", + "bbox": [ + 148, + 202, + 823, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Harbin Institute of Technology, $^{2}$ Harbin Institute of Technology, Shenzhen Department of Gastroenterology, Peking University Shenzhen Hospital", + "bbox": [ + 187, + 238, + 781, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jiahan.li,jiuyang.dong，shenjinhuang}@stu.hit.edu.cn,", + "bbox": [ + 240, + 277, + 751, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junjunjiang, fxp, ybzhang08}@hit.edu.cn, lixi122188@sina.com", + "bbox": [ + 214, + 295, + 751, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, virtual staining technology has greatly promoted the advancement of histopathology. Despite the practical successes achieved, the outstanding performance of most virtual staining methods relies on hard-to-obtain paired images in training. In this paper, we propose a method for virtual immunohistochemistry (IHC) staining, named confusion-GAN, which does not require paired images and can achieve comparable performance to supervised algorithms. Specifically, we propose a multi-branch discriminator, which judges if the features of generated images can be embedded into the feature pool of target domain images, to improve the visual quality of generated images. Meanwhile, we also propose a novel patch-level pathology information extractor, which is assisted by multiple instance learning, to ensure pathological consistency during virtual staining. Extensive experiments were conducted on three types of IHC images, including a high-resolution hepatocellular carcinoma immunohistochemical dataset proposed by us. The results demonstrated that our proposed confusionGAN can generate highly realistic images that are capable of deceiving even experienced pathologists. Furthermore, compared to using H&E images directly, the downstream diagnosis achieved higher accuracy when using images generated by confusion-GAN. Our dataset and codes will be available at https://github.com/jiahanli2022/confusionGAN.", + "bbox": [ + 76, + 377, + 472, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 801, + 209, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Histopathological staining is a key step in clinical pathological analysis [2, 33]. As the most common histopathological staining method, H&E staining can clearly delineate", + "bbox": [ + 76, + 827, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the cellular structures in tissues. At the same time, immunohistochemistry (IHC) staining is a functional staining method used for evaluation of protein-specific expression, which greatly assists in tumor diagnosis and cancer prognosis [1]. For example, breast cancer with overexpression of human epidermal growth factor receptor 2 (HER2) tends to have aggressive clinical behaviour [24], while breast cancer with low expression of estrogen receptor (ER) often accompanies poor response to endocrine therapy [18]. Besides, glypican-3 (GPC3) is a recently discovered immunohistochemical protein that shows promising potential for indicating the presence of hepatocellular carcinoma.", + "bbox": [ + 500, + 345, + 893, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, the process of IHC staining is laborious, time-consuming, and expensive [42]. Therefore, many researches have focused on utilizing virtual staining techniques to convert H&E images into IHC images [38, 41, 43]. The existing virtual IHC staining methods can be roughly divided into two categories: supervised learning-based methods and unsupervised learning-based methods. Supervised learning-based methods require a large amount of paired images for training. These H&E-IHC pairs typically come from adjacent layers of same tissues, accompanied by complex data processing and alignment. What's more, these methods cannot be used in scenarios where adjacent layers are not available. On the other hand, unsupervised methods often focus on the style of the generated images, ignoring the transmission of pathological information.", + "bbox": [ + 496, + 532, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these aforementioned problems, we propose a virtual IHC staining framework called confusion-GAN, which utilizes weakly supervised learning to facilitate the unsupervised staining process of unpaired images. To explore the pathological information of H&E patches, a novel patch-level pathology information extractor which is assisted by multiple instance learning (MIL) is proposed. This method utilizes a dual-spherical loss proposed by us to push positive patches away from negative patches in the feature", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Xiaopeng Fan, Yongbing Zhang", + "bbox": [ + 94, + 886, + 397, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11259", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "space, aiming to achieve accurate acquisition of pathological information. When training the confusion-GAN, we constrain the pathological consistency between the staining results and the input images. In addition, we propose a multi-branch discriminator, which contains a confusion discriminator as well as a common single-image discriminator, aiming to enhance the fidelity of the staining results. Specifically, we first combine a generated image with multiple real images to create an image pool in each iteration. The confusion discriminator aims to distinguish the unique generated image from this pool. Simultaneously, the generator tries to hinder this judgment of the confusion discriminator.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed approach, we conducted extensive experiments on three datasets: 1) converting H&E staining images to HER2 staining images over the BCI dataset [24], 2) converting H&E staining images to ER staining images over the MIST dataset [18], and 3) convert H&E staining images to GPC3 staining images over our collected hepatocellular carcinoma immunohistochemical (HCI) dataset. In terms of image quality and pathological information transmission, our confusion-GAN can achieve superior performance compared to existing state-of-the-art methods and deceive human pathologists to some extent. Additionally, we compared the classification accuracy of ABMIL [13], a widely used algorithm for pathological image diagnosis, by using GPC3 images generated from our confusion-GAN and the original H&E images as the input data. The experimental results clearly indicate that the utilization of generated GPC3 images can significantly improve the diagnosis performance.", + "bbox": [ + 75, + 272, + 472, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions are summarized as follows:", + "bbox": [ + 94, + 559, + 398, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an algorithm that utilizes weakly supervised learning to assist unsupervised virtual staining for the first time, which generates satisfactory images on three virtual IHC staining tasks.", + "- We propose a novel patch-level pathology information extractor (PPIE) for H&E image classification. As a plug-and-play method, PPIE is demonstrated to significantly improve the instance-level classification performance when combined with existing MIL methods.", + "- We propose a multi-branch discriminator (MBD) that can significantly enhance the fidelity of the staining results compared to only using a single-image discriminator.", + "- We collected a high-resolution hepatocellular carcinoma immunohistochemical dataset, which is contains both H&E and GPC3 staining images for the first time." + ], + "bbox": [ + 76, + 574, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 821, + 218, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Virtual IHC Staining Techniques", + "text_level": 1, + "bbox": [ + 76, + 845, + 366, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing virtual IHC staining methods can be broadly divided into two categories: supervised and unsupervised", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning based methods. Supervised learning based methods require a large number of H&E-IHC pairs as training images [8, 18, 24]. There are two ways of obtaining these paired images: (1) Carrying out axially consecutive scanning of tissues, actually staining H&E and IHC on adjacent slices, then performing alignment between H&E and IHC slides through complex data processing and registration to obtain the paired images for training. For example, Liu et al. [24] train their proposed PyramidP2P network over a registered H&E-HER2 dataset. However, the tissue contents from different layers are inherently difficult to be consistent, making pixel-level alignment impossible. Therefore, the inability to fully align will inevitably result in performance limitation. In addition, such supervised methods are unlikely to be applicable when adjacent layer data is unavailable. (2) Using images of a third modality other than H&E and IHC, such as microscopic fluorescence data, as a bridge, to generate pixel-wise aligned H&E and IHC training data [8]. Unfortunately, access to the third modality data will also add additional overhead.", + "bbox": [ + 496, + 90, + 890, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The vast majority of existing unsupervised virtual IHC staining methods are based on image translation networks [5, 9, 21, 30, 33, 35, 40]. Although these networks can convert H&E-style images into IHC-style images, it does not mean that the Pos/Neg properties of the translated results will be consistent with that of their inputs. To keep the aforementioned pathological consistency, Liu et al. [23] and Boyd et al. [3] introduce expert annotations, which is an expensive solution.", + "bbox": [ + 496, + 393, + 890, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Patch-level Classification Assisted by MIL", + "text_level": 1, + "bbox": [ + 500, + 540, + 857, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, MIL are used for patch-level classification [4, 13, 16, 19, 32]. In the case of pathological images, each WSI is regarded as a bag, and each patch within the WSI is treated as an instance. A bag label is assigned to each bag, which is classified as positive if the bag contains at least one positive instance. These bag labels serve as the ground truths for training MIL networks in an end-to-end manner. Existing methods determine whether a patch is positive or not by comparing the attention scores (or instance-level predictions) of all patches [6, 20, 22, 29, 31, 34]. Unfortunately, these methods can only ensure accurate patch-level classification within a limited region of high attention scores. Therefore, we propose a novel patch-level pathology information extractor to compensate for the shortcomings of these methods.", + "bbox": [ + 496, + 564, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 806, + 671, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Framework", + "text_level": 1, + "bbox": [ + 500, + 830, + 627, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One significant difference between the virtual IHC staining and the image translation of natural images is that generated IHC images must maintain pathological consistency with", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11260", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a5cb8d4624bffc597ea3e917395960d082fe807f267f124806fadccf2ddd86c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 88, + 844, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23ee533a12ea7a3d7cdc944a5e3c9f7dcce8944daa6c79495a48887faae56f31.jpg", + "image_caption": [ + "Figure 1. Structure of the proposed confusion-GAN. (a) The confusion-GAN contains two important modules: patch-level pathology information extractor (PPIE) and multi-branch discriminator (MBD). (b) The training process of PPIE. (c) The structure of the MBD." + ], + "image_footnote": [], + "bbox": [ + 122, + 275, + 488, + 448 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/861113c4b9c2670c4ef619fd59ffa13db5345f439378e9eaea1f65d14073e5a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 276, + 844, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "their inputs while preserving texture consistency. From the perspectives of improving the quality of generated images and accurately transmitting pathological information, we propose confusion-GAN, which is illustrated in Fig. 1. The confusion-GAN contains two important modules, patch-level pathology information extraction module and multibranch discriminator, which are employed to maintain the pathological consistency and generate more realistic images, respectively.", + "bbox": [ + 75, + 515, + 468, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the real H&E image $x_{i}$ is first passed through the generator $G_{A}$ to obtain the IHC-style image $\\hat{y}_i$ . Then, $\\hat{y}_i$ is further remapped into the H&E-style image $\\tilde{x}_i$ using the generator $G_{B}$ . Similarly, we can obtain $\\hat{x}_i$ and $\\tilde{y}_i$ from real IHC image $y_{i}$ .", + "bbox": [ + 75, + 655, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the process of generating IHC image from the H&E image, we devise a Pathology Loss to ensure the accurate transmission of pathological information. IHC classifier $Cls_{ihc}$ and a PPIE module are designed to obtain the pathological information of the H&E image $x_{i}$ and the generated IHC image $\\hat{y}_{i}$ , respectively. To pretrain a reliable IHC classifier, we simulate the annotation process of pathologists through color analysis to obtain patch-level pos/neg labels for actually captured IHC images: patches with positive exhibit protein-specific expression within more than 1% of the area (corresponding to yellow-brown coloration", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the image). A light-weight image classifier $Cls_{ihc}$ is then designed utilizing the pos/neg labels obtained by the previously described process, which consists of 6 convolutional layers activated by ReLU and a fully connected layer (FCN) for the final classification. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1. Similarly, the pathological information of the H&E image is obtained utilizing the PPIE module, which will be elaborated in Sec. 3.2. Finally, we adopt the Pathology Loss to constrain the consistency of pathological information between $x_{i}$ and $\\hat{y}_{i}$ :", + "bbox": [ + 496, + 515, + 892, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {p a t h}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(P P I E \\left(x _ {i}\\right) \\log \\left(C l s ^ {i h c} \\left(\\hat {y} _ {i}\\right)\\right) \\right. \\tag {1} \\\\ + (1 - P P I E \\left(x _ {i}\\right)) \\log \\left(1 - C l s ^ {i h c} \\left(\\hat {y} _ {i}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 686, + 890, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $PPI(\\cdot)$ represents the operation of PPI module.", + "bbox": [ + 498, + 752, + 883, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, to enhance the quality of the generated images, we propose a multi-branch discriminator for both H&E-to-IHC and IHC-to-H&E process in the confusionGAN. Detailed description of the multi-branch discriminator can be found in Sec. 3.3.", + "bbox": [ + 496, + 768, + 890, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, the hybrid loss for the generator is as follows:", + "bbox": [ + 517, + 843, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {G} = \\alpha * L _ {\\text {p a t h}} + \\beta * L _ {G} ^ {C} + \\lambda * L _ {G} ^ {A d v} \\\\ + \\eta * L _ {G} ^ {C y c l e} + \\iota * L _ {G} ^ {\\text {I d e n t i t y}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 863, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11261", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{G}^{Adv}$ , $L_{G}^{Cycle}$ , and $L_{G}^{Identity}$ represent the adversarial loss, cycled loss, and identity loss similar to CycleGAN [44], while $\\alpha$ , $\\beta$ , $\\lambda$ , $\\eta$ , and $\\iota$ are hyper-parameters controlling the weighting coefficients. $L_{G}^{C}$ is the loss function used to guide the generator optimization with the confusion discriminator, which will be described in detail in Sec. 3.3.", + "bbox": [ + 76, + 88, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Patch-level Pathology Information Extraction", + "text_level": 1, + "bbox": [ + 76, + 193, + 464, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to memory limitations, virtual staining can only be performed at the patch level. Unfortunately, only WSI-level labels are available for majority cases, since one WSI usually contains gigapixel, impeding the patch-wise annotation in clinical. Therefore, it is necessary to extract patch-level pathological information (i.e., the pos/neg category of each patch) from the H&E WSI to ensure accurate virtual staining.", + "bbox": [ + 75, + 217, + 468, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For H&E images that only have WSI level labels (lacking patch-level annotations), the classification of each patch is typically completed through attention scores or instance-level classification predictions within the MIL framework. In each positive WSI, there is a high likelihood that patches with the highest attention scores or instance-level classification scores will be positive. We refer to the action of selecting the top $k$ patches of each positive WSI based on the highest attention scores (or instance-level prediction scores) as \"top-k selection\". When $k$ is relatively small, it is ensured that the selected patches are all positive; when $k$ is large, the selected collections may include many negative patches [26]. Therefore, the current methods have not been able to accurately classify all patches.", + "bbox": [ + 75, + 339, + 472, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this problem, building upon the top-k selection, we propose a plug-and-play PPIE module that can improve the classification performance of H&E patches. Specifically, we first perform top-k selection to obtain $X_{pos}^{he}$ from positive WSIs in the training set. Next, we crop a set of negative patches $X_{neg}^{he}$ from negative WSIs. It should be noted all patches in a negative WSI are negative, unlike the cases of positive WSIs where both positive and negative patches may coexist. Besides, we utilized a state-of-the-art diffusion model BBDM [17] to learn how to generate a distribution similar to that of $X_{pos}^{he}$ using $X_{neg}^{he}$ . The set of images generated by BBDM is referred to as $X_{pos}^{he}$ .", + "bbox": [ + 75, + 550, + 472, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On this basis, we borrow the idea of anomaly detection to tackle the classification problem of patch-level H&E images [10]. In each iteration of training the PPIE, we sample $b$ images from $X_{\\text{neg}}^{\\text{he}}$ , $X_{\\text{pos}}^{\\text{he}}$ , and $X_{\\text{pos}}^{\\text{he}}$ respectively, and feed them into the encoder to obtain the feature sets $C_1 = \\{c_{1i}\\}_{i=1}^b$ , $C_2 = \\{c_{2i}\\}_{i=1}^b$ , and $C_3 = \\{c_{3i}\\}_{i=1}^b$ , as shown in Fig. 1(b). Here, $b$ represents the batch size during training. The encoder consists of 6 convolutional layers activated by ReLU and a fully connected layer with an output dimension of 512. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1.", + "bbox": [ + 75, + 734, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As is well known, the success of the anomaly detection depends on two aspects: 1) the samples in the in-distribution (ID) class should be as close as possible in the feature space; 2) the samples in the out-of-distribution (OOD) class should be as far as possible from samples in the ID class [39]. We use the loss function $L_{ID}$ to force all the samples in the ID class to be distributed as close to the center as possible:", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {I D} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\left\\| c _ {1 i} - O \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 207, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $O$ represents the average features of all the ID data $X_{neg}^{he}$ . It should be noted that $O$ will be cumulatively updated during training. Besides, we propose a double-spherical loss $L_{OOD}$ to ensure that the overall distribution of $C_3$ has a greater distance from $O$ compared to the overall distribution of $C_2$ :", + "bbox": [ + 496, + 257, + 893, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {O O D} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\max \\left(R _ {1} - \\left\\| c _ {2 i} - O \\right\\| ^ {2}, 0\\right) \\tag {4} \\\\ + \\frac {1}{b} \\sum_ {i = 1} ^ {b} m a x (R _ {2} - | | c _ {3 i} - O | | ^ {2}, 0). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 358, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We set the hyperparameter $R_{2}$ to control the distance and ensure that $R_{2}$ is greater than $R_{1}$ , as shown in Fig. 1(b). Employing $L_{ID}$ and $L_{OOD}$ , we can ensure that $X_{neg}^{he}$ and $X_{pos}^{he}$ have strong separability in the feature space. Finally, we treat $X_{neg}^{he}$ and $X_{pos}^{he}$ as ID and OOD classes respectively, and train the patch-level classifier $Cls^{he}$ by cross entropy loss $L_{cls}$ :", + "bbox": [ + 496, + 452, + 893, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {c l s} = - \\frac {1}{n} \\sum_ {i = 1} ^ {s} \\left(l _ {i} \\log \\left(C l s ^ {h e} \\left(x _ {i} ^ {h e}\\right)\\right) \\right. \\tag {5} \\\\ + \\left(1 - l _ {i}\\right) \\log \\left(1 - C l s ^ {h e} \\left(x _ {i} ^ {h e}\\right)\\right)). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 575, + 890, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When using PPIE for inference, negative patches are considered as ID class, while positive patches are considered as OOD class.", + "bbox": [ + 496, + 643, + 890, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, the hybrid loss function $L_{PPIE}$ for training the PPIE module is as follows:", + "bbox": [ + 496, + 690, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {P P I E} = \\mu * L _ {I D} + \\tau * L _ {O O D} + v * L _ {c l s}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 733, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu, \\tau,$ and $v$ are hyper-parameters to control the weighting coefficients.", + "bbox": [ + 496, + 762, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-branch Discriminator", + "text_level": 1, + "bbox": [ + 500, + 801, + 754, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve more realistic image generation, we propose a multi-branch discriminator, as shown in Fig. 1(c). Taking generating an IHC-style image from a H&E image as an example, we send the result $\\hat{y}_i$ obtained by the generator $G_{A}$ together with a reference pool $P_i^{ihc}$ consisting of", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11262", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$N - 1$ real images from the same domain to the confusion discriminator $C_B$ . $C_B$ shares the feature extraction layer $D_B^S$ with the single-image discriminator $D_B$ , which helps eliminate the interference of the content of the reference pool images on the confusion discriminator. This one-to-many discrimination paradigm requires the generator's results to possess more common attributes of the current domain images, thereby forcing the generated distribution to approximate the real distribution more closely.", + "bbox": [ + 75, + 90, + 468, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training of the $C_B$ , we aim to accurately distinguish between the generated result and real images. Then we randomly insert $\\hat{y}_i$ into a position in the list of $P_i^{ihc}$ , and finally concatenate them in the channel dimension to obtain the input $m_i^{ihc}$ for $C_B$ . Correspondingly, we construct an $N$ -dimensional vector $l_i^{ihc}$ to represent the labels of each element in $m_i^{ihc}$ , with 0 indicating real image and 1 indicating generated image. Next, we pass $m_i^{ihc}$ into $D_B^S$ to obtain a feature pool, which is fed into an MLP layer of $C_B$ for liveness verification. $C_B$ is then updated by minimizing:", + "bbox": [ + 75, + 227, + 470, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C _ {B}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {B} \\left(m _ {i} ^ {i h c}\\right) - l _ {i} ^ {i h c} \\right| \\right| ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 388, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, during the process of IHC-to-H&E image translation, the loss function $L_{C_A}$ of the confusion discriminator $C_A$ is defined as follows:", + "bbox": [ + 76, + 436, + 468, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C _ {A}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {A} \\left(m _ {i} ^ {h e}\\right) - l _ {i} ^ {h e} \\right| \\right| ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 491, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m_{i}^{he}$ represents the concatenation result of the generated H&E image and $N - 1$ real H&E images, and $l_{i}^{he}$ represents the corresponding labels of $m_{i}^{he}$ . The proposed confusion discriminators are trained using the following formula:", + "bbox": [ + 76, + 541, + 468, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C} = L _ {C _ {B}} + L _ {C _ {A}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 613, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function used for training $D_B$ and $D_A$ is consistent with CycleGAN.", + "bbox": [ + 76, + 638, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When training the generator, we obtain $m_i^{ihc}$ and $m_i^{he}$ in the same manner. However, in this stage, we want the confusion discriminator unable to identify the generator's results from these $N$ images. Therefore, all labels in $l_i^{ihc}$ and $l_i^{he}$ are set to 0. The generators are updated based on the following loss functions:", + "bbox": [ + 75, + 669, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {G} ^ {C} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {B} \\left(m _ {i} ^ {i h c}\\right) - l _ {i} ^ {i h c} \\right| \\right| ^ {2} \\tag {10} \\\\ + \\left\\| C _ {A} \\left(m _ {i} ^ {h e}\\right) - l _ {i} ^ {h e} \\right\\| ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 771, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In each iteration, we train the generators $G_{A}$ and $G_{B}$ , $D_{A}$ and $D_{B}$ , and our proposed $C_{A}$ and $C_{B}$ in an alternating manner, following a sequential order. This process is similar to what most generative adversarial networks do. Here", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$G_{A}$ and $G_{B}$ employ the unet256 network used in CycleGAN and the structures of $D_{A}$ and $D_{B}$ are also consistent with those in CycleGAN.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 151, + 630, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 176, + 581, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. GPC3, discovered as a tumor marker in recent years, shows great potential for specific expression in Hepatocellular carcinoma (HCC). We have collected a HCC Immunohistochemical Image dataset named HCI, which includes both the H&E and GPC3 images. All data comes from Peking University Shenzhen Hospital. We selected 20 H&E-IHC image pairs from a pool of WSIs for testing and validation, and perform registration on them using the DeepHistReg [37]. Next, we conducted separate screenings for a total of 30 H&E images and 30 IHC images, and these images are used for training. Furthermore, we segment these 100 WSIs into approximately 1.4 million patches sized of $256 \\times 256$ in a non-overlapping manner at 20x magnification. All H&E images have undergone staining normalization. We also verified our method on public datasets: BCI [24] and MIST [18].", + "bbox": [ + 496, + 199, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluated the image quality of the staining results using objective metrics such as Frechet Inception Distance (FID) [27], Learnable Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index (SSIM) [11]. Additionally, we invited three pathologists to provide subjective evaluations of the staining results. In assessing the patch-level classification performance of the aforementioned PPIE, we utilized metrics such as area under curve (AUC) [25], False Positive Rate (FPR) [28], True Positive Rate (TPR), and Geometric Mean (G-mean) [7].", + "bbox": [ + 496, + 441, + 890, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We trained the PPIE using the Adam optimizer with a learning rate of 0.001. The dual-sphere distances $R_{1}$ and $R_{2}$ in Eq. (4) are set to 150 and 262.5 respectively. The hyper-parameters $\\mu$ , $\\tau$ , and $\\upsilon$ in Eq. (6) are all set to 1. When training the virtual staining network, $\\alpha$ , $\\beta$ , $\\lambda$ , $\\eta$ and $\\iota$ in Eq. (2) are set to 1, 1, 1, 10, and 5 respectively. We set the reference pool size $N$ as 32 and kept the other configuration of the staining network consistent with CycleGAN. In the downstream experiments, we trained the ABMIL model on a re-partitioned dataset for a maximum of 15 epochs. For all experiments, we trained and tested our confusion-GAN on patches of size $256 \\times 256$ .", + "bbox": [ + 496, + 594, + 890, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Staining Performance", + "text_level": 1, + "bbox": [ + 500, + 786, + 702, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HCI. As shown in Fig. 2, we compared our method with three unsupervised image translation methods, namely CycleGAN [44], UGATIT [15], and DCD [12], on the HCI dataset. The first column in Fig. 2 displays the input H&E images, while the last column shows actually stained IHC (GPC3) images for reference. These images are from corre", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11263", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13b54aa30a1deb61bdbb5a9935b55b4971c4f8c473ad5d90e3620c0c26a651a1.jpg", + "image_caption": [ + "Figure 2. Comparison of the abilities of different unsupervised algorithms to preserve pathological consistency on HCI. The red dashed box indicates the area where the generated image cannot maintain consistent pathological information with the input H&E image." + ], + "image_footnote": [], + "bbox": [ + 166, + 90, + 803, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59b48f485350d25439c25c8b3a0e5efb8769c30cc4d750fb61824d00fd1207a8.jpg", + "image_caption": [ + "Figure 3. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on BCI. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors." + ], + "image_footnote": [], + "bbox": [ + 166, + 364, + 805, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sponding regions of adjacent slices in the same tissue. We marked in red dashed lines the regions where the pathological information is inconsistent with the input. Obviously, although the competing methods CycleGAN, UGATIT, and DCD can maintain the texture details of the input images well, there are significant mistaining areas since the consistency of pathological information can not be ensured before and after staining. This phenomenon demonstrates the effectiveness of the proposed PPIE module. Besides, we use three objective metrics, FID, LPIPS, and SSIM to evaluate these virtual staining methods. As shown in Tab. 1, our method achieves the best performance in all the three metrics, indicating that the output distribution of the confusion-GAN is closest to the distribution of the real images.", + "bbox": [ + 75, + 661, + 472, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BCI and MIST. BCI and MIST datasets provide paired", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/87891f4796c23d207a45feba9f1f5fb5527fa7bd8ab27f0d781ab5a61206f9a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓LPIPS↓SSIM↑
CycleGAN(unet)105.40.6360.1870
CycleGAN(resnet)90.90.6220.2086
UGATIT95.30.6020.2223
DCD91.30.6120.2266
Ours77.30.5700.2300
", + "bbox": [ + 553, + 657, + 844, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The evaluation of different algorithms on HCI.", + "bbox": [ + 529, + 750, + 859, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "H&E-IHC patch images, where the acquisition of these patches relies on alignment with the sliced images from adjacent layers within the same tissue. However, the tissue contents from different adjacent layers are inherently impossible to be completely consistent, making pixel-level alignment impossible. We compared our method with four fully supervised methods (Pix2pix [14], Pix2pixHD [36],", + "bbox": [ + 498, + 794, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11264", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6bca552fd26dc44d1aa5c1686e70b84f3953e065b421bb93c3466e5a2a0b4c5d.jpg", + "image_caption": [ + "Figure 4. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on MIST. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors." + ], + "image_footnote": [], + "bbox": [ + 158, + 90, + 810, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PyramidP2P [24], ASP [18]) on these two datasets, as shown in Fig. 3 and Fig. 4. We also marked the areas with pathological information errors in the staining results with dashed lines. Additionally, we marked the areas with content errors in the staining results with solid lines. From Fig. 3 and Fig. 4, it can be observed that Pix2pix, Pix2pixHD, and PyramidP2P methods are unable to maintain the textures of the input images, often resulting in significant presence of artifacts in the generated results. This phenomenon is attributed to the significant discrepancy in content between the input H&E images and the reference IHC images used as the ground truth during the training process. The ASP based on contrastive learning demonstrates good performance in preserving content. However, the misalignment of supervised information can still introduce some errors in the pathological information of the staining results. In contrast, our proposed confusion-GAN has achieved better results than these supervised learning methods.", + "bbox": [ + 75, + 393, + 472, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We used FID, LPIPS, and SSIM for objective evaluation. As can be seen from Tab. 2, our confusion-GAN achieved the best FID and LPIPS scores on both datasets, which indicating the effectiveness of our algorithm. Regarding the SSIM evaluation metric, which is directly calculated in the pixel space, our method still achieved the highest score on MIST dataset. Meanwhile, our method can achieve the best visual effects on both BCI and MIST, especially considering the fidelity of textures in the generated results. It is worth mentioning that, due to the lack of complete WSIs in MIST and BCI, we assigned labels to IHC images using the color analysis mentioned in Sec. 3.1, and then transfer the labels of IHC images to the corresponding H&E images. Regarding to BCI, we directly utilized the provided patch-level la", + "bbox": [ + 75, + 689, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bels, which guides the training of confusion-GAN.", + "bbox": [ + 500, + 393, + 836, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c28a841ccbc0e83b4009037e34c964de420350311c33653dfeb8acf59ebfb777.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBCIMIST
FID↓LPIPS↓SSIM↑FID↓LPIPS↓SSIM↑
Pix2pix110.70.4800.495780.90.5290.2127
Pix2pixHD106.80.5050.4766131.00.5640.2129
PyramidP2P99.60.5050.466397.50.5380.2061
ASP54.30.5030.492341.20.5320.2062
Ours49.20.4780.457939.80.5160.2174
", + "bbox": [ + 511, + 419, + 885, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 568, + 653, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We explored the working mechanism of PPIE and analyzed how PPIE and the MBD affect the staining results.", + "bbox": [ + 498, + 590, + 890, + 622 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/552e471f0896a369562074772bd955e5477f31bfb47857ea62ce4d9ec0b9ac54.jpg", + "table_caption": [ + "Table 2. Comparing the performance of our method and supervised algorithms on BCI and MIST." + ], + "table_footnote": [], + "table_body": "
MethodkAUC↑FPR↓TPR↑G-mean↑
ABMIL100.7130.2770.6410.680
300.7820.2260.7260.750
1000.6370.4080.5890.591
wo0.7320.3230.6860.681
Maxpool100.6820.3240.6660.671
300.7310.2570.6890.715
1000.7060.2670.6670.699
wo0.6800.3280.6130.680
ClamMb100.7660.2380.6910.726
300.7700.2190.7080.744
1000.7320.2860.6270.669
wo0.7260.3380.6870.675
Meanpool100.7450.2530.6710.708
300.7680.2390.7280.744
1000.7090.2540.6190.679
wo0.7580.3240.7100.693
", + "bbox": [ + 550, + 632, + 846, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Patch-level classification results obtained from different top-k patches on HCI.", + "bbox": [ + 498, + 842, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of $k$ in the top-k selection. We selected four clas", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sic MIL methods (ABMIL [13], MaxPool, CLAM-MB[26], and MeanPool) for top-k selection in PPIE training. As shown in Tab. 3, we evaluate the classification performance of PPIE using AUC, FPR, TPR, and G-mean. We found that the classification performance is best when $k$ is set to 30. This is because when $k$ is too small, the PPIE cannot obtain sufficient training data. On the other hand, when $k$ is too large, the top-k selection results will contain a large number of negative patches, which hinders the effectiveness of the double-spherical loss. Among these four MIL methods, we obtained similar conclusions and improved the patch-level classification performance, achieving gains of $5.0\\%$ , $5.1\\%$ , $4.4\\%$ , and $1.0\\%$ respectively.", + "bbox": [ + 75, + 90, + 472, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of PPIE and MBD. To investigate how PPIE and MBD impact the performance, we conducted an ablation experiment on HCI, as shown in Fig. 5. By comparing the second, third, and fourth columns in Fig. 5, it can be observed that the PPIE plays a crucial role in accurately transmitting pathological information. On the other hand, the MBD can enhance the clarity of cell nuclei and surrounding tissues, affecting the overall style of the generated results. What's more, based on the qualitative analysis of the indicators in Tab. 4, we found that the method combining PPIE and the MBD achieved the best results.", + "bbox": [ + 75, + 287, + 473, + 454 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa0447c4b606af7719a02527cdd47b44920d9a6b9051db67914c1897e22e6bfa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PPIEXX✓✓
MBDX✓X✓
FID↓105.482.4103.377.3
LPIPS↓0.6360.5880.5740.570
SSIM↑0.18700.22230.22940.2300
", + "bbox": [ + 132, + 463, + 415, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Down-stream Classification Performance", + "text_level": 1, + "bbox": [ + 76, + 585, + 428, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted downstream experiments related to the HCC diagnosis, and compared the classification performance achieved by using the H&E and generated GPC3 slides as inputs of the ABMIL model, respectively. When using H&E images as input achieves an AUC of $55.6\\%$ , while using the generated images as input achieves an AUC of $88.9\\%$ . This indicates that compared to H&E images, our generated GPC3 images can improve the classification accuracy of HCC.", + "bbox": [ + 75, + 609, + 470, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Subjective Evaluation by Pathologists", + "text_level": 1, + "bbox": [ + 76, + 755, + 401, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We invited three pathologists to provide subjective evaluations on two aspects: transmission of pathological information and the quality of the images. We randomly selected 50 pairs of generated GPC3 images along with their corresponding adjacent tissues and then asked pathologists to evaluate whether the protein expression levels were consistent between the two. As shown in the first row of Tab. 5, in most cases, pathologists believed that the protein expression", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03a34eb11974582e2152893ccdf8f5a60ca643cb2de74790604f262e3df2c700.jpg", + "image_caption": [ + "Figure 5. Ablation study: effect of PPIE and MBD on HCI. The red box indicates regions with pathological information errors." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 882, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "levels were consistent between them. This indicates that our confusion-GAN can effectively maintain pathological consistency. Afterwards, we randomly selected 50 GPC3 images and 50 generated GPC3 images for pathologists to distinguish which are generated images. From the second row of the Tab. 5, it can be seen that doctors cannot distinguish the generated images well, indicating that our confusion-GAN can generate images comparable to real images.", + "bbox": [ + 496, + 306, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2c3d0383bdcd67c88f23c385d1efc7b8d835900cf4958b101ab92b9e7378375.jpg", + "table_caption": [ + "Table 4. Exploring the impact of PPIE and MBD on HCI." + ], + "table_footnote": [], + "table_body": "
Accuracypathologist1pathologist2pathologist3
Neg/Pos0.860.800.84
Fake/True0.620.480.52
", + "bbox": [ + 531, + 439, + 864, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Subjective evaluations of pathologists on HCI.", + "bbox": [ + 531, + 493, + 859, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 536, + 617, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a method called confusion-GAN, which is the first to utilize weakly-supervised learning to assist in virtual IHC staining of unpaired pathological images. Extensive experiments have demonstrated that utilizing patch-level pathological information to guide the training of staining network is crucial for achieving accurate virtual IHC staining. Furthermore, we propose a multi-branch discriminator for achieving high-fidelity image generation. Our approach achieved the state-of-the-art results on multiple datasets. To advance the diagnosis of HCC, we have also proposed the first GPC3 dataset and demonstrated that the images generated by confusion-GAN are more helpful in HCC detection compared to H&E images. We believe that unsupervised methods like confusion-GAN, which do not require fine annotation or complex preprocessing, will be crucial in virtual staining research in the future.", + "bbox": [ + 496, + 561, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62031023&62331011), in part by the Shenzhen Science and Technology Project (GXWD20220818170353009), and in part by the Fundamental Research Funds for the Central Universities (Grant No. HIT.OCEF.2023050).", + "bbox": [ + 496, + 803, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11266", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bijie Bai, Xilin Yang, Yuzhu Li, Yijie Zhang, Nir Pillar, and Aydogan Ozcan. Deep learning-enabled virtual histological staining of biological samples. Light: Science & Applications, 12(1):57, 2023. 1", + "[2] Marian Boktor, Benjamin R Ecclestone, Vlad Pekar, Deepak Dinakaran, John R Mackey, Paul Fieguth, and Parsin Haji Reza. Virtual histological staining of label-free total absorption photoacoustic remote sensing (ta-pars). Scientific Reports, 12(1):10296, 2022. 1", + "[3] Joseph Boyd, Irène Villa, Marie-Christine Mathieu, Eric Deutsch, Nikos Paragios, Maria Vakalopoulou, and Stergios Christodoulidis. Region-guided cyclegans for stain transfer in whole slide images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 356-365. Springer, 2022. 2", + "[4] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2", + "[5] Tanujit Chakraborty, Shraddha M Naik, Madhurima Panja, Bayapureddy Manvitha, et al. Ten years of generative adversarial nets (gans): A survey of the state-of-the-art. arXiv preprint arXiv:2308.16316, 2023. 2", + "[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 2", + "[7] Yanyan Chen, Kuaini Wang, and Ping Zhong. One-class support tensor machine. Knowledge-Based Systems, 96:14-28, 2016. 5", + "[8] Kevin de Haan, Yijie Zhang, Jonathan E Zuckerman, Tairan Liu, Anthony E Sisk, Miguel FP Diaz, Kuang-Yu Jen, Alexander Nobori, Sofia Liou, Sarah Zhang, et al. Deep learning-based transformation of h&e stained tissues into special stains. Nature communications, 12(1):4884, 2021. 2", + "[9] Michael Gadermayr, Maximilian Tschuchnig, Lea Maria Stangassinger, Christina Kreutzer, Sebastien Couillard-Despres, Gertie Janneke Oostingh, and Anton Hittmair. Improving automated thyroid cancer classification of frozen sections by the aid of virtual image translation and stain normalization. Computer Methods and Programs in Biomedicine Update, 3:100092, 2023. 2", + "[10] Xiaoyuan Guo, Judy W Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Margin-aware intraclass novelty identification for medical images. Journal of Medical Imaging, 9(1): 014004-014004, 2022. 4", + "[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5", + "[12] Tie Hu, Mingbao Lin, Lizhou You, Fei Chao, and Rongrong Ji. Discriminator-cooperated feature map distillation for gan" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20351-20360, 2023. 5", + "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 8", + "[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 6", + "[15] Junho Kim, Minjae Kim, Hyeonwoo Kang, and Kwanghee Lee. U-gat-it: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830, 2019. 5", + "[16] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 2", + "[17] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 4", + "[18] Fangda Li, Zhiqiang Hu, Wen Chen, and Avinash Kak. Adaptive supervised patchnce loss for learning h&e-to-ihc stain translation with inconsistent groundtruth image pairs. arXiv preprint arXiv:2303.06193, 2023. 1, 2, 5, 7", + "[19] Weijian Li, Viet-Duy Nguyen, Haofu Liao, Matt Wilder, Ke Cheng, and Jiebo Luo. Patch transformer for multi-tagging whole slide histopathology images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 532-540. Springer, 2019. 2", + "[20] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 2", + "[21] Kechun Liu, Beibin Li, Wenjun Wu, Caitlin May, Oliver Chang, Stevan Knezevich, Lisa Reisch, Joann Elmore, and Linda Shapiro. Vsgd-net: Virtual staining guided melanocyte detection on histopathological images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1918-1927, 2023. 2", + "[22] Kangning Liu, Weicheng Zhu, Yiqiu Shen, Sheng Liu, Narges Razavian, Krzysztof J Geras, and Carlos Fernandez-Granda. Multiple instance learning via iterative self-paced supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355–3365, 2023. 2", + "[23] Shuting Liu, Baochang Zhang, Yiqing Liu, Anjia Han, Huijuan Shi, Tian Guan, and Yonghong He. Unpaired stain transfer using pathology-consistent constrained generative" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11267", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial networks. IEEE Transactions on Medical Imaging, 40(8):1977-1989, 2021. 2", + "[24] Shengjie Liu, Chuang Zhu, Feng Xu, Xinyu Jia, Zhongyue Shi, and Mulan Jin. Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1815-1824, 2022. 1, 2, 5, 7", + "[25] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5", + "[26] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 4, 8", + "[27] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 5", + "[28] Alvin I Mushlin, Ruth W Kouides, and David E Shapiro. Estimating the accuracy of screening mammography: a meta-analysis. American journal of preventive medicine, 14(2): 143-153, 1998. 5", + "[29] Linhao Qu, Manning Wang, Zhijian Song, et al. Bidirectional weakly supervised knowledge distillation for whole slide image classification. Advances in Neural Information Processing Systems, 35:15368-15381, 2022. 2", + "[30] Jesus Salido, Noelia Valdez, Lucia González-López, Oscar Deniz, and Gloria Bueno. Comparison of deep learning models for digital h&E staining from unpaired label-free multispectral microscopy images. Computer Methods and Programs in Biomedicine, 235:107528, 2023. 2", + "[31] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2", + "[32] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2", + "[33] Lulin Shi, Ivy HM Wong, Claudia TK Lo, Lauren WK Tsui, and Terence TW Wong. Unsupervised multiple virtual histological staining from label-free autofluorescence images. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 1, 2", + "[34] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 2", + "[35] Elena Y Trizna, Aleksandr M Sinitca, Asya I Lyanova, Diana R Baidamshina, Pavel V Zelenikhin, Dmitrii I Kaplun," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Airat R Kayumov, and Mikhail I Bogachev. Brightfield vs fluorescent staining dataset-a test bed image set for machine learning based virtual staining. Scientific Data, 10(1):160, 2023. 2", + "[36] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 6", + "[37] Marek Wodzinski and Henning Müller. Deephistreg: Unsupervised deep learning registration framework for differently stained histology samples. Computer methods and programs in biomedicine, 198:105799, 2021. 5", + "[38] Georg Wölfein, In Hwa Um, David J Harrison, and Ognjen Arandjelović. Hoechstgan: Virtual lymphocyte staining using generative adversarial networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4997-5007, 2023. 1", + "[39] Xuan Xia, Xizhou Pan, Nan Li, Xing He, Lin Ma, Xiaoguang Zhang, and Ning Ding. Gan-based anomaly detection: A review. Neurocomputing, 493:497-535, 2022. 4", + "[40] Zhaoyang Xu, Xingru Huang, Carlos Fernández Moro, Béla Bozóky, and Qianni Zhang. Gan-based virtual re-staining: a promising solution for whole slide image analysis. arXiv preprint arXiv:1901.04059, 2019. 2", + "[41] Bowei Zeng, Yiyang Lin, Yifeng Wang, Yang Chen, Jiuyang Dong, Xi Li, and Yongbing Zhang. Semi-supervised pr virtual staining for breast histopathological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 232-241. Springer, 2022. 1", + "[42] Ranran Zhang, Yankun Cao, Yujun Li, Zhi Liu, Jianye Wang, Jiahuan He, Chenyang Zhang, Xiaoyu Sui, Pengfei Zhang, Lizhen Cui, et al. Mvfstain: multiple virtual functional stain histopathology images generation based on specific domain mapping. Medical Image Analysis, 80:102520, 2022. 1", + "[43] Yijie Zhang, Luzhe Huang, Tairan Liu, Keyi Cheng, Kevin de Haan, Yuzhu Li, Bijie Bai, and Aydogan Ozcan. Virtual staining of defocused autofluorescence images of unlabeled tissue using deep neural networks. Intelligent Computing, 2022. 1", + "[44] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 738 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11268", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_model.json b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87de708c41702886c37d978e87dadeaec51b9877 --- /dev/null +++ b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.178 + ], + "angle": 0, + "content": "Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.203, + 0.825, + 0.24 + ], + "angle": 0, + "content": "Jiahan Li\\(^{1}\\), Jiuyang Dong\\(^{1}\\), Shenjin Huang\\(^{1}\\), Xi Li\\(^{3}\\), Junjun Jiang\\(^{1}\\), Xiaopeng Fan\\(^{1*}\\), Yongbing Zhang\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.239, + 0.782, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\)Harbin Institute of Technology, \\(^{2}\\)Harbin Institute of Technology, Shenzhen Department of Gastroenterology, Peking University Shenzhen Hospital" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.278, + 0.752, + 0.293 + ], + "angle": 0, + "content": "{jiahan.li,jiuyang.dong，shenjinhuang}@stu.hit.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.296, + 0.753, + 0.31 + ], + "angle": 0, + "content": "{junjunjiang, fxp, ybzhang08}@hit.edu.cn, lixi122188@sina.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.378, + 0.473, + 0.768 + ], + "angle": 0, + "content": "Recently, virtual staining technology has greatly promoted the advancement of histopathology. Despite the practical successes achieved, the outstanding performance of most virtual staining methods relies on hard-to-obtain paired images in training. In this paper, we propose a method for virtual immunohistochemistry (IHC) staining, named confusion-GAN, which does not require paired images and can achieve comparable performance to supervised algorithms. Specifically, we propose a multi-branch discriminator, which judges if the features of generated images can be embedded into the feature pool of target domain images, to improve the visual quality of generated images. Meanwhile, we also propose a novel patch-level pathology information extractor, which is assisted by multiple instance learning, to ensure pathological consistency during virtual staining. Extensive experiments were conducted on three types of IHC images, including a high-resolution hepatocellular carcinoma immunohistochemical dataset proposed by us. The results demonstrated that our proposed confusionGAN can generate highly realistic images that are capable of deceiving even experienced pathologists. Furthermore, compared to using H&E images directly, the downstream diagnosis achieved higher accuracy when using images generated by confusion-GAN. Our dataset and codes will be available at https://github.com/jiahanli2022/confusionGAN." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.21, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.828, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Histopathological staining is a key step in clinical pathological analysis [2, 33]. As the most common histopathological staining method, H&E staining can clearly delineate" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.346, + 0.895, + 0.528 + ], + "angle": 0, + "content": "the cellular structures in tissues. At the same time, immunohistochemistry (IHC) staining is a functional staining method used for evaluation of protein-specific expression, which greatly assists in tumor diagnosis and cancer prognosis [1]. For example, breast cancer with overexpression of human epidermal growth factor receptor 2 (HER2) tends to have aggressive clinical behaviour [24], while breast cancer with low expression of estrogen receptor (ER) often accompanies poor response to endocrine therapy [18]. Besides, glypican-3 (GPC3) is a recently discovered immunohistochemical protein that shows promising potential for indicating the presence of hepatocellular carcinoma." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Unfortunately, the process of IHC staining is laborious, time-consuming, and expensive [42]. Therefore, many researches have focused on utilizing virtual staining techniques to convert H&E images into IHC images [38, 41, 43]. The existing virtual IHC staining methods can be roughly divided into two categories: supervised learning-based methods and unsupervised learning-based methods. Supervised learning-based methods require a large amount of paired images for training. These H&E-IHC pairs typically come from adjacent layers of same tissues, accompanied by complex data processing and alignment. What's more, these methods cannot be used in scenarios where adjacent layers are not available. On the other hand, unsupervised methods often focus on the style of the generated images, ignoring the transmission of pathological information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "To address these aforementioned problems, we propose a virtual IHC staining framework called confusion-GAN, which utilizes weakly supervised learning to facilitate the unsupervised staining process of unpaired images. To explore the pathological information of H&E patches, a novel patch-level pathology information extractor which is assisted by multiple instance learning (MIL) is proposed. This method utilizes a dual-spherical loss proposed by us to push positive patches away from negative patches in the feature" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.398, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author: Xiaopeng Fan, Yongbing Zhang" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11259" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "space, aiming to achieve accurate acquisition of pathological information. When training the confusion-GAN, we constrain the pathological consistency between the staining results and the input images. In addition, we propose a multi-branch discriminator, which contains a confusion discriminator as well as a common single-image discriminator, aiming to enhance the fidelity of the staining results. Specifically, we first combine a generated image with multiple real images to create an image pool in each iteration. The confusion discriminator aims to distinguish the unique generated image from this pool. Simultaneously, the generator tries to hinder this judgment of the confusion discriminator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.558 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed approach, we conducted extensive experiments on three datasets: 1) converting H&E staining images to HER2 staining images over the BCI dataset [24], 2) converting H&E staining images to ER staining images over the MIST dataset [18], and 3) convert H&E staining images to GPC3 staining images over our collected hepatocellular carcinoma immunohistochemical (HCI) dataset. In terms of image quality and pathological information transmission, our confusion-GAN can achieve superior performance compared to existing state-of-the-art methods and deceive human pathologists to some extent. Additionally, we compared the classification accuracy of ABMIL [13], a widely used algorithm for pathological image diagnosis, by using GPC3 images generated from our confusion-GAN and the original H&E images as the input data. The experimental results clearly indicate that the utilization of generated GPC3 images can significantly improve the diagnosis performance." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.56, + 0.4, + 0.574 + ], + "angle": 0, + "content": "The contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.575, + 0.47, + 0.635 + ], + "angle": 0, + "content": "- We propose an algorithm that utilizes weakly supervised learning to assist unsupervised virtual staining for the first time, which generates satisfactory images on three virtual IHC staining tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.638, + 0.47, + 0.713 + ], + "angle": 0, + "content": "- We propose a novel patch-level pathology information extractor (PPIE) for H&E image classification. As a plug-and-play method, PPIE is demonstrated to significantly improve the instance-level classification performance when combined with existing MIL methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.716, + 0.469, + 0.761 + ], + "angle": 0, + "content": "- We propose a multi-branch discriminator (MBD) that can significantly enhance the fidelity of the staining results compared to only using a single-image discriminator." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.764, + 0.469, + 0.809 + ], + "angle": 0, + "content": "- We collected a high-resolution hepatocellular carcinoma immunohistochemical dataset, which is contains both H&E and GPC3 staining images for the first time." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.575, + 0.47, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.22, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.367, + 0.864 + ], + "angle": 0, + "content": "2.1. Virtual IHC Staining Techniques" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Existing virtual IHC staining methods can be broadly divided into two categories: supervised and unsupervised" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.393 + ], + "angle": 0, + "content": "learning based methods. Supervised learning based methods require a large number of H&E-IHC pairs as training images [8, 18, 24]. There are two ways of obtaining these paired images: (1) Carrying out axially consecutive scanning of tissues, actually staining H&E and IHC on adjacent slices, then performing alignment between H&E and IHC slides through complex data processing and registration to obtain the paired images for training. For example, Liu et al. [24] train their proposed PyramidP2P network over a registered H&E-HER2 dataset. However, the tissue contents from different layers are inherently difficult to be consistent, making pixel-level alignment impossible. Therefore, the inability to fully align will inevitably result in performance limitation. In addition, such supervised methods are unlikely to be applicable when adjacent layer data is unavailable. (2) Using images of a third modality other than H&E and IHC, such as microscopic fluorescence data, as a bridge, to generate pixel-wise aligned H&E and IHC training data [8]. Unfortunately, access to the third modality data will also add additional overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.53 + ], + "angle": 0, + "content": "The vast majority of existing unsupervised virtual IHC staining methods are based on image translation networks [5, 9, 21, 30, 33, 35, 40]. Although these networks can convert H&E-style images into IHC-style images, it does not mean that the Pos/Neg properties of the translated results will be consistent with that of their inputs. To keep the aforementioned pathological consistency, Liu et al. [23] and Boyd et al. [3] introduce expert annotations, which is an expensive solution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.859, + 0.557 + ], + "angle": 0, + "content": "2.2. Patch-level Classification Assisted by MIL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Currently, MIL are used for patch-level classification [4, 13, 16, 19, 32]. In the case of pathological images, each WSI is regarded as a bag, and each patch within the WSI is treated as an instance. A bag label is assigned to each bag, which is classified as positive if the bag contains at least one positive instance. These bag labels serve as the ground truths for training MIL networks in an end-to-end manner. Existing methods determine whether a patch is positive or not by comparing the attention scores (or instance-level predictions) of all patches [6, 20, 22, 29, 31, 34]. Unfortunately, these methods can only ensure accurate patch-level classification within a limited region of high attention scores. Therefore, we propose a novel patch-level pathology information extractor to compensate for the shortcomings of these methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.673, + 0.823 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.628, + 0.846 + ], + "angle": 0, + "content": "3.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "One significant difference between the virtual IHC staining and the image translation of natural images is that generated IHC images must maintain pathological consistency with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11260" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.845, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.276, + 0.49, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.277, + 0.846, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.893, + 0.491 + ], + "angle": 0, + "content": "Figure 1. Structure of the proposed confusion-GAN. (a) The confusion-GAN contains two important modules: patch-level pathology information extractor (PPIE) and multi-branch discriminator (MBD). (b) The training process of PPIE. (c) The structure of the MBD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.653 + ], + "angle": 0, + "content": "their inputs while preserving texture consistency. From the perspectives of improving the quality of generated images and accurately transmitting pathological information, we propose confusion-GAN, which is illustrated in Fig. 1. The confusion-GAN contains two important modules, patch-level pathology information extraction module and multibranch discriminator, which are employed to maintain the pathological consistency and generate more realistic images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.732 + ], + "angle": 0, + "content": "As shown in Fig. 1, the real H&E image \\( x_{i} \\) is first passed through the generator \\( G_{A} \\) to obtain the IHC-style image \\( \\hat{y}_i \\). Then, \\( \\hat{y}_i \\) is further remapped into the H&E-style image \\( \\tilde{x}_i \\) using the generator \\( G_{B} \\). Similarly, we can obtain \\( \\hat{x}_i \\) and \\( \\tilde{y}_i \\) from real IHC image \\( y_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "During the process of generating IHC image from the H&E image, we devise a Pathology Loss to ensure the accurate transmission of pathological information. IHC classifier \\( Cls_{ihc} \\) and a PPIE module are designed to obtain the pathological information of the H&E image \\( x_{i} \\) and the generated IHC image \\( \\hat{y}_{i} \\), respectively. To pretrain a reliable IHC classifier, we simulate the annotation process of pathologists through color analysis to obtain patch-level pos/neg labels for actually captured IHC images: patches with positive exhibit protein-specific expression within more than 1% of the area (corresponding to yellow-brown coloration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.893, + 0.683 + ], + "angle": 0, + "content": "in the image). A light-weight image classifier \\( Cls_{ihc} \\) is then designed utilizing the pos/neg labels obtained by the previously described process, which consists of 6 convolutional layers activated by ReLU and a fully connected layer (FCN) for the final classification. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1. Similarly, the pathological information of the H&E image is obtained utilizing the PPIE module, which will be elaborated in Sec. 3.2. Finally, we adopt the Pathology Loss to constrain the consistency of pathological information between \\( x_{i} \\) and \\( \\hat{y}_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.687, + 0.891, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {p a t h}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(P P I E \\left(x _ {i}\\right) \\log \\left(C l s ^ {i h c} \\left(\\hat {y} _ {i}\\right)\\right) \\right. \\tag {1} \\\\ + (1 - P P I E \\left(x _ {i}\\right)) \\log \\left(1 - C l s ^ {i h c} \\left(\\hat {y} _ {i}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.753, + 0.885, + 0.768 + ], + "angle": 0, + "content": "where \\( PPI(\\cdot) \\) represents the operation of PPI module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.769, + 0.892, + 0.843 + ], + "angle": 0, + "content": "In addition, to enhance the quality of the generated images, we propose a multi-branch discriminator for both H&E-to-IHC and IHC-to-H&E process in the confusionGAN. Detailed description of the multi-branch discriminator can be found in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.844, + 0.892, + 0.86 + ], + "angle": 0, + "content": "In general, the hybrid loss for the generator is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.864, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {G} = \\alpha * L _ {\\text {p a t h}} + \\beta * L _ {G} ^ {C} + \\lambda * L _ {G} ^ {A d v} \\\\ + \\eta * L _ {G} ^ {C y c l e} + \\iota * L _ {G} ^ {\\text {I d e n t i t y}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.182 + ], + "angle": 0, + "content": "where \\(L_{G}^{Adv}\\), \\(L_{G}^{Cycle}\\), and \\(L_{G}^{Identity}\\) represent the adversarial loss, cycled loss, and identity loss similar to CycleGAN [44], while \\(\\alpha\\), \\(\\beta\\), \\(\\lambda\\), \\(\\eta\\), and \\(\\iota\\) are hyper-parameters controlling the weighting coefficients. \\(L_{G}^{C}\\) is the loss function used to guide the generator optimization with the confusion discriminator, which will be described in detail in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.194, + 0.465, + 0.21 + ], + "angle": 0, + "content": "3.2. Patch-level Pathology Information Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.338 + ], + "angle": 0, + "content": "Due to memory limitations, virtual staining can only be performed at the patch level. Unfortunately, only WSI-level labels are available for majority cases, since one WSI usually contains gigapixel, impeding the patch-wise annotation in clinical. Therefore, it is necessary to extract patch-level pathological information (i.e., the pos/neg category of each patch) from the H&E WSI to ensure accurate virtual staining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.55 + ], + "angle": 0, + "content": "For H&E images that only have WSI level labels (lacking patch-level annotations), the classification of each patch is typically completed through attention scores or instance-level classification predictions within the MIL framework. In each positive WSI, there is a high likelihood that patches with the highest attention scores or instance-level classification scores will be positive. We refer to the action of selecting the top \\( k \\) patches of each positive WSI based on the highest attention scores (or instance-level prediction scores) as \"top-k selection\". When \\( k \\) is relatively small, it is ensured that the selected patches are all positive; when \\( k \\) is large, the selected collections may include many negative patches [26]. Therefore, the current methods have not been able to accurately classify all patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.736 + ], + "angle": 0, + "content": "To address this problem, building upon the top-k selection, we propose a plug-and-play PPIE module that can improve the classification performance of H&E patches. Specifically, we first perform top-k selection to obtain \\( X_{pos}^{he} \\) from positive WSIs in the training set. Next, we crop a set of negative patches \\( X_{neg}^{he} \\) from negative WSIs. It should be noted all patches in a negative WSI are negative, unlike the cases of positive WSIs where both positive and negative patches may coexist. Besides, we utilized a state-of-the-art diffusion model BBDM [17] to learn how to generate a distribution similar to that of \\( X_{pos}^{he} \\) using \\( X_{neg}^{he} \\). The set of images generated by BBDM is referred to as \\( X_{pos}^{he} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.473, + 0.902 + ], + "angle": 0, + "content": "On this basis, we borrow the idea of anomaly detection to tackle the classification problem of patch-level H&E images [10]. In each iteration of training the PPIE, we sample \\( b \\) images from \\( X_{\\text{neg}}^{\\text{he}} \\), \\( X_{\\text{pos}}^{\\text{he}} \\), and \\( X_{\\text{pos}}^{\\text{he}} \\) respectively, and feed them into the encoder to obtain the feature sets \\( C_1 = \\{c_{1i}\\}_{i=1}^b \\), \\( C_2 = \\{c_{2i}\\}_{i=1}^b \\), and \\( C_3 = \\{c_{3i}\\}_{i=1}^b \\), as shown in Fig. 1(b). Here, \\( b \\) represents the batch size during training. The encoder consists of 6 convolutional layers activated by ReLU and a fully connected layer with an output dimension of 512. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "As is well known, the success of the anomaly detection depends on two aspects: 1) the samples in the in-distribution (ID) class should be as close as possible in the feature space; 2) the samples in the out-of-distribution (OOD) class should be as far as possible from samples in the ID class [39]. We use the loss function \\( L_{ID} \\) to force all the samples in the ID class to be distributed as close to the center as possible:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.208, + 0.892, + 0.25 + ], + "angle": 0, + "content": "\\[\nL _ {I D} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\left\\| c _ {1 i} - O \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.895, + 0.349 + ], + "angle": 0, + "content": "where \\( O \\) represents the average features of all the ID data \\( X_{neg}^{he} \\). It should be noted that \\( O \\) will be cumulatively updated during training. Besides, we propose a double-spherical loss \\( L_{OOD} \\) to ensure that the overall distribution of \\( C_3 \\) has a greater distance from \\( O \\) compared to the overall distribution of \\( C_2 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.359, + 0.892, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {O O D} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\max \\left(R _ {1} - \\left\\| c _ {2 i} - O \\right\\| ^ {2}, 0\\right) \\tag {4} \\\\ + \\frac {1}{b} \\sum_ {i = 1} ^ {b} m a x (R _ {2} - | | c _ {3 i} - O | | ^ {2}, 0). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.895, + 0.567 + ], + "angle": 0, + "content": "We set the hyperparameter \\( R_{2} \\) to control the distance and ensure that \\( R_{2} \\) is greater than \\( R_{1} \\), as shown in Fig. 1(b). Employing \\( L_{ID} \\) and \\( L_{OOD} \\), we can ensure that \\( X_{neg}^{he} \\) and \\( X_{pos}^{he} \\) have strong separability in the feature space. Finally, we treat \\( X_{neg}^{he} \\) and \\( X_{pos}^{he} \\) as ID and OOD classes respectively, and train the patch-level classifier \\( Cls^{he} \\) by cross entropy loss \\( L_{cls} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.577, + 0.892, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {c l s} = - \\frac {1}{n} \\sum_ {i = 1} ^ {s} \\left(l _ {i} \\log \\left(C l s ^ {h e} \\left(x _ {i} ^ {h e}\\right)\\right) \\right. \\tag {5} \\\\ + \\left(1 - l _ {i}\\right) \\log \\left(1 - C l s ^ {h e} \\left(x _ {i} ^ {h e}\\right)\\right)). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.69 + ], + "angle": 0, + "content": "When using PPIE for inference, negative patches are considered as ID class, while positive patches are considered as OOD class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.721 + ], + "angle": 0, + "content": "Overall, the hybrid loss function \\(L_{PPIE}\\) for training the PPIE module is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.734, + 0.892, + 0.751 + ], + "angle": 0, + "content": "\\[\nL _ {P P I E} = \\mu * L _ {I D} + \\tau * L _ {O O D} + v * L _ {c l s}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.794 + ], + "angle": 0, + "content": "where \\(\\mu, \\tau,\\) and \\(v\\) are hyper-parameters to control the weighting coefficients." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.755, + 0.817 + ], + "angle": 0, + "content": "3.3. Multi-branch Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To achieve more realistic image generation, we propose a multi-branch discriminator, as shown in Fig. 1(c). Taking generating an IHC-style image from a H&E image as an example, we send the result \\(\\hat{y}_i\\) obtained by the generator \\(G_{A}\\) together with a reference pool \\(P_i^{ihc}\\) consisting of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11262" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.227 + ], + "angle": 0, + "content": "\\(N - 1\\) real images from the same domain to the confusion discriminator \\(C_B\\). \\(C_B\\) shares the feature extraction layer \\(D_B^S\\) with the single-image discriminator \\(D_B\\), which helps eliminate the interference of the content of the reference pool images on the confusion discriminator. This one-to-many discrimination paradigm requires the generator's results to possess more common attributes of the current domain images, thereby forcing the generated distribution to approximate the real distribution more closely." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.471, + 0.38 + ], + "angle": 0, + "content": "During the training of the \\( C_B \\), we aim to accurately distinguish between the generated result and real images. Then we randomly insert \\( \\hat{y}_i \\) into a position in the list of \\( P_i^{ihc} \\), and finally concatenate them in the channel dimension to obtain the input \\( m_i^{ihc} \\) for \\( C_B \\). Correspondingly, we construct an \\( N \\)-dimensional vector \\( l_i^{ihc} \\) to represent the labels of each element in \\( m_i^{ihc} \\), with 0 indicating real image and 1 indicating generated image. Next, we pass \\( m_i^{ihc} \\) into \\( D_B^S \\) to obtain a feature pool, which is fed into an MLP layer of \\( C_B \\) for liveness verification. \\( C_B \\) is then updated by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.389, + 0.47, + 0.428 + ], + "angle": 0, + "content": "\\[\nL _ {C _ {B}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {B} \\left(m _ {i} ^ {i h c}\\right) - l _ {i} ^ {i h c} \\right| \\right| ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.47, + 0.482 + ], + "angle": 0, + "content": "Similarly, during the process of IHC-to-H&E image translation, the loss function \\( L_{C_A} \\) of the confusion discriminator \\( C_A \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.492, + 0.47, + 0.531 + ], + "angle": 0, + "content": "\\[\nL _ {C _ {A}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {A} \\left(m _ {i} ^ {h e}\\right) - l _ {i} ^ {h e} \\right| \\right| ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.542, + 0.469, + 0.603 + ], + "angle": 0, + "content": "where \\( m_{i}^{he} \\) represents the concatenation result of the generated H&E image and \\( N - 1 \\) real H&E images, and \\( l_{i}^{he} \\) represents the corresponding labels of \\( m_{i}^{he} \\). The proposed confusion discriminators are trained using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.614, + 0.47, + 0.631 + ], + "angle": 0, + "content": "\\[\nL _ {C} = L _ {C _ {B}} + L _ {C _ {A}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.469, + 0.669 + ], + "angle": 0, + "content": "The loss function used for training \\(D_B\\) and \\(D_A\\) is consistent with CycleGAN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.761 + ], + "angle": 0, + "content": "When training the generator, we obtain \\( m_i^{ihc} \\) and \\( m_i^{he} \\) in the same manner. However, in this stage, we want the confusion discriminator unable to identify the generator's results from these \\( N \\) images. Therefore, all labels in \\( l_i^{ihc} \\) and \\( l_i^{he} \\) are set to 0. The generators are updated based on the following loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.772, + 0.469, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {G} ^ {C} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {B} \\left(m _ {i} ^ {i h c}\\right) - l _ {i} ^ {i h c} \\right| \\right| ^ {2} \\tag {10} \\\\ + \\left\\| C _ {A} \\left(m _ {i} ^ {h e}\\right) - l _ {i} ^ {h e} \\right\\| ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In each iteration, we train the generators \\( G_{A} \\) and \\( G_{B} \\), \\( D_{A} \\) and \\( D_{B} \\), and our proposed \\( C_{A} \\) and \\( C_{B} \\) in an alternating manner, following a sequential order. This process is similar to what most generative adversarial networks do. Here" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "\\(G_{A}\\) and \\(G_{B}\\) employ the unet256 network used in CycleGAN and the structures of \\(D_{A}\\) and \\(D_{B}\\) are also consistent with those in CycleGAN." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.152, + 0.632, + 0.169 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.177, + 0.582, + 0.194 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Datasets. GPC3, discovered as a tumor marker in recent years, shows great potential for specific expression in Hepatocellular carcinoma (HCC). We have collected a HCC Immunohistochemical Image dataset named HCI, which includes both the H&E and GPC3 images. All data comes from Peking University Shenzhen Hospital. We selected 20 H&E-IHC image pairs from a pool of WSIs for testing and validation, and perform registration on them using the DeepHistReg [37]. Next, we conducted separate screenings for a total of 30 H&E images and 30 IHC images, and these images are used for training. Furthermore, we segment these 100 WSIs into approximately 1.4 million patches sized of \\(256 \\times 256\\) in a non-overlapping manner at 20x magnification. All H&E images have undergone staining normalization. We also verified our method on public datasets: BCI [24] and MIST [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Metrics. We evaluated the image quality of the staining results using objective metrics such as Frechet Inception Distance (FID) [27], Learnable Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index (SSIM) [11]. Additionally, we invited three pathologists to provide subjective evaluations of the staining results. In assessing the patch-level classification performance of the aforementioned PPIE, we utilized metrics such as area under curve (AUC) [25], False Positive Rate (FPR) [28], True Positive Rate (TPR), and Geometric Mean (G-mean) [7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.776 + ], + "angle": 0, + "content": "Implementation Details. We trained the PPIE using the Adam optimizer with a learning rate of 0.001. The dual-sphere distances \\( R_{1} \\) and \\( R_{2} \\) in Eq. (4) are set to 150 and 262.5 respectively. The hyper-parameters \\( \\mu \\), \\( \\tau \\), and \\( \\upsilon \\) in Eq. (6) are all set to 1. When training the virtual staining network, \\( \\alpha \\), \\( \\beta \\), \\( \\lambda \\), \\( \\eta \\) and \\( \\iota \\) in Eq. (2) are set to 1, 1, 1, 10, and 5 respectively. We set the reference pool size \\( N \\) as 32 and kept the other configuration of the staining network consistent with CycleGAN. In the downstream experiments, we trained the ABMIL model on a re-partitioned dataset for a maximum of 15 epochs. For all experiments, we trained and tested our confusion-GAN on patches of size \\( 256 \\times 256 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.704, + 0.803 + ], + "angle": 0, + "content": "4.2. Staining Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "HCI. As shown in Fig. 2, we compared our method with three unsupervised image translation methods, namely CycleGAN [44], UGATIT [15], and DCD [12], on the HCI dataset. The first column in Fig. 2 displays the input H&E images, while the last column shows actually stained IHC (GPC3) images for reference. These images are from corre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11263" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.091, + 0.805, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Comparison of the abilities of different unsupervised algorithms to preserve pathological consistency on HCI. The red dashed box indicates the area where the generated image cannot maintain consistent pathological information with the input H&E image." + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.366, + 0.806, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.593, + 0.895, + 0.637 + ], + "angle": 0, + "content": "Figure 3. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on BCI. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.473, + 0.875 + ], + "angle": 0, + "content": "sponding regions of adjacent slices in the same tissue. We marked in red dashed lines the regions where the pathological information is inconsistent with the input. Obviously, although the competing methods CycleGAN, UGATIT, and DCD can maintain the texture details of the input images well, there are significant mistaining areas since the consistency of pathological information can not be ensured before and after staining. This phenomenon demonstrates the effectiveness of the proposed PPIE module. Besides, we use three objective metrics, FID, LPIPS, and SSIM to evaluate these virtual staining methods. As shown in Tab. 1, our method achieves the best performance in all the three metrics, indicating that the output distribution of the confusion-GAN is closest to the distribution of the real images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "BCI and MIST. BCI and MIST datasets provide paired" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.658, + 0.845, + 0.741 + ], + "angle": 0, + "content": "
MethodFID↓LPIPS↓SSIM↑
CycleGAN(unet)105.40.6360.1870
CycleGAN(resnet)90.90.6220.2086
UGATIT95.30.6020.2223
DCD91.30.6120.2266
Ours77.30.5700.2300
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.75, + 0.86, + 0.765 + ], + "angle": 0, + "content": "Table 1. The evaluation of different algorithms on HCI." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "H&E-IHC patch images, where the acquisition of these patches relies on alignment with the sliced images from adjacent layers within the same tissue. However, the tissue contents from different adjacent layers are inherently impossible to be completely consistent, making pixel-level alignment impossible. We compared our method with four fully supervised methods (Pix2pix [14], Pix2pixHD [36]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11264" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.091, + 0.812, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.895, + 0.369 + ], + "angle": 0, + "content": "Figure 4. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on MIST. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.681 + ], + "angle": 0, + "content": "PyramidP2P [24], ASP [18]) on these two datasets, as shown in Fig. 3 and Fig. 4. We also marked the areas with pathological information errors in the staining results with dashed lines. Additionally, we marked the areas with content errors in the staining results with solid lines. From Fig. 3 and Fig. 4, it can be observed that Pix2pix, Pix2pixHD, and PyramidP2P methods are unable to maintain the textures of the input images, often resulting in significant presence of artifacts in the generated results. This phenomenon is attributed to the significant discrepancy in content between the input H&E images and the reference IHC images used as the ground truth during the training process. The ASP based on contrastive learning demonstrates good performance in preserving content. However, the misalignment of supervised information can still introduce some errors in the pathological information of the staining results. In contrast, our proposed confusion-GAN has achieved better results than these supervised learning methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We used FID, LPIPS, and SSIM for objective evaluation. As can be seen from Tab. 2, our confusion-GAN achieved the best FID and LPIPS scores on both datasets, which indicating the effectiveness of our algorithm. Regarding the SSIM evaluation metric, which is directly calculated in the pixel space, our method still achieved the highest score on MIST dataset. Meanwhile, our method can achieve the best visual effects on both BCI and MIST, especially considering the fidelity of textures in the generated results. It is worth mentioning that, due to the lack of complete WSIs in MIST and BCI, we assigned labels to IHC images using the color analysis mentioned in Sec. 3.1, and then transfer the labels of IHC images to the corresponding H&E images. Regarding to BCI, we directly utilized the provided patch-level la" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.394, + 0.837, + 0.41 + ], + "angle": 0, + "content": "bels, which guides the training of confusion-GAN." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.42, + 0.887, + 0.507 + ], + "angle": 0, + "content": "
MethodsBCIMIST
FID↓LPIPS↓SSIM↑FID↓LPIPS↓SSIM↑
Pix2pix110.70.4800.495780.90.5290.2127
Pix2pixHD106.80.5050.4766131.00.5640.2129
PyramidP2P99.60.5050.466397.50.5380.2061
ASP54.30.5030.492341.20.5320.2062
Ours49.20.4780.457939.80.5160.2174
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Table 2. Comparing the performance of our method and supervised algorithms on BCI and MIST." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.655, + 0.585 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.623 + ], + "angle": 0, + "content": "We explored the working mechanism of PPIE and analyzed how PPIE and the MBD affect the staining results." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.633, + 0.847, + 0.832 + ], + "angle": 0, + "content": "
MethodkAUC↑FPR↓TPR↑G-mean↑
ABMIL100.7130.2770.6410.680
300.7820.2260.7260.750
1000.6370.4080.5890.591
wo0.7320.3230.6860.681
Maxpool100.6820.3240.6660.671
300.7310.2570.6890.715
1000.7060.2670.6670.699
wo0.6800.3280.6130.680
ClamMb100.7660.2380.6910.726
300.7700.2190.7080.744
1000.7320.2860.6270.669
wo0.7260.3380.6870.675
Meanpool100.7450.2530.6710.708
300.7680.2390.7280.744
1000.7090.2540.6190.679
wo0.7580.3240.7100.693
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.843, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Table 3. Patch-level classification results obtained from different top-k patches on HCI." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of \\(k\\) in the top-k selection. We selected four clas" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.287 + ], + "angle": 0, + "content": "sic MIL methods (ABMIL [13], MaxPool, CLAM-MB[26], and MeanPool) for top-k selection in PPIE training. As shown in Tab. 3, we evaluate the classification performance of PPIE using AUC, FPR, TPR, and G-mean. We found that the classification performance is best when \\( k \\) is set to 30. This is because when \\( k \\) is too small, the PPIE cannot obtain sufficient training data. On the other hand, when \\( k \\) is too large, the top-k selection results will contain a large number of negative patches, which hinders the effectiveness of the double-spherical loss. Among these four MIL methods, we obtained similar conclusions and improved the patch-level classification performance, achieving gains of \\( 5.0\\% \\), \\( 5.1\\% \\), \\( 4.4\\% \\), and \\( 1.0\\% \\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.474, + 0.455 + ], + "angle": 0, + "content": "Effect of PPIE and MBD. To investigate how PPIE and MBD impact the performance, we conducted an ablation experiment on HCI, as shown in Fig. 5. By comparing the second, third, and fourth columns in Fig. 5, it can be observed that the PPIE plays a crucial role in accurately transmitting pathological information. On the other hand, the MBD can enhance the clarity of cell nuclei and surrounding tissues, affecting the overall style of the generated results. What's more, based on the qualitative analysis of the indicators in Tab. 4, we found that the method combining PPIE and the MBD achieved the best results." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.464, + 0.416, + 0.535 + ], + "angle": 0, + "content": "
PPIEXX✓✓
MBDX✓X✓
FID↓105.482.4103.377.3
LPIPS↓0.6360.5880.5740.570
SSIM↑0.18700.22230.22940.2300
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.543, + 0.444, + 0.558 + ], + "angle": 0, + "content": "Table 4. Exploring the impact of PPIE and MBD on HCI." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.429, + 0.603 + ], + "angle": 0, + "content": "4.4. Down-stream Classification Performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.471, + 0.748 + ], + "angle": 0, + "content": "We conducted downstream experiments related to the HCC diagnosis, and compared the classification performance achieved by using the H&E and generated GPC3 slides as inputs of the ABMIL model, respectively. When using H&E images as input achieves an AUC of \\(55.6\\%\\), while using the generated images as input achieves an AUC of \\(88.9\\%\\). This indicates that compared to H&E images, our generated GPC3 images can improve the classification accuracy of HCC." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.403, + 0.773 + ], + "angle": 0, + "content": "4.5. Subjective Evaluation by Pathologists" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We invited three pathologists to provide subjective evaluations on two aspects: transmission of pathological information and the quality of the images. We randomly selected 50 pairs of generated GPC3 images along with their corresponding adjacent tissues and then asked pathologists to evaluate whether the protein expression levels were consistent between the two. As shown in the first row of Tab. 5, in most cases, pathologists believed that the protein expression" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.092, + 0.883, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Figure 5. Ablation study: effect of PPIE and MBD on HCI. The red box indicates regions with pathological information errors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.43 + ], + "angle": 0, + "content": "levels were consistent between them. This indicates that our confusion-GAN can effectively maintain pathological consistency. Afterwards, we randomly selected 50 GPC3 images and 50 generated GPC3 images for pathologists to distinguish which are generated images. From the second row of the Tab. 5, it can be seen that doctors cannot distinguish the generated images well, indicating that our confusion-GAN can generate images comparable to real images." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.44, + 0.866, + 0.485 + ], + "angle": 0, + "content": "
Accuracypathologist1pathologist2pathologist3
Neg/Pos0.860.800.84
Fake/True0.620.480.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.494, + 0.86, + 0.508 + ], + "angle": 0, + "content": "Table 5. Subjective evaluations of pathologists on HCI." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.619, + 0.552 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.803 + ], + "angle": 0, + "content": "In this paper, we propose a method called confusion-GAN, which is the first to utilize weakly-supervised learning to assist in virtual IHC staining of unpaired pathological images. Extensive experiments have demonstrated that utilizing patch-level pathological information to guide the training of staining network is crucial for achieving accurate virtual IHC staining. Furthermore, we propose a multi-branch discriminator for achieving high-fidelity image generation. Our approach achieved the state-of-the-art results on multiple datasets. To advance the diagnosis of HCC, we have also proposed the first GPC3 dataset and demonstrated that the images generated by confusion-GAN are more helpful in HCC detection compared to H&E images. We believe that unsupervised methods like confusion-GAN, which do not require fine annotation or complex preprocessing, will be crucial in virtual staining research in the future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.892, + 0.888 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62031023&62331011), in part by the Shenzhen Science and Technology Project (GXWD20220818170353009), and in part by the Fundamental Research Funds for the Central Universities (Grant No. HIT.OCEF.2023050)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11266" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Bijie Bai, Xilin Yang, Yuzhu Li, Yijie Zhang, Nir Pillar, and Aydogan Ozcan. Deep learning-enabled virtual histological staining of biological samples. Light: Science & Applications, 12(1):57, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Marian Boktor, Benjamin R Ecclestone, Vlad Pekar, Deepak Dinakaran, John R Mackey, Paul Fieguth, and Parsin Haji Reza. Virtual histological staining of label-free total absorption photoacoustic remote sensing (ta-pars). Scientific Reports, 12(1):10296, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Joseph Boyd, Irène Villa, Marie-Christine Mathieu, Eric Deutsch, Nikos Paragios, Maria Vakalopoulou, and Stergios Christodoulidis. Region-guided cyclegans for stain transfer in whole slide images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 356-365. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.408 + ], + "angle": 0, + "content": "[4] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.464 + ], + "angle": 0, + "content": "[5] Tanujit Chakraborty, Shraddha M Naik, Madhurima Panja, Bayapureddy Manvitha, et al. Ten years of generative adversarial nets (gans): A survey of the state-of-the-art. arXiv preprint arXiv:2308.16316, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.472, + 0.549 + ], + "angle": 0, + "content": "[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[7] Yanyan Chen, Kuaini Wang, and Ping Zhong. One-class support tensor machine. Knowledge-Based Systems, 96:14-28, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.471, + 0.674 + ], + "angle": 0, + "content": "[8] Kevin de Haan, Yijie Zhang, Jonathan E Zuckerman, Tairan Liu, Anthony E Sisk, Miguel FP Diaz, Kuang-Yu Jen, Alexander Nobori, Sofia Liou, Sarah Zhang, et al. Deep learning-based transformation of h&e stained tissues into special stains. Nature communications, 12(1):4884, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[9] Michael Gadermayr, Maximilian Tschuchnig, Lea Maria Stangassinger, Christina Kreutzer, Sebastien Couillard-Despres, Gertie Janneke Oostingh, and Anton Hittmair. Improving automated thyroid cancer classification of frozen sections by the aid of virtual image translation and stain normalization. Computer Methods and Programs in Biomedicine Update, 3:100092, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[10] Xiaoyuan Guo, Judy W Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Margin-aware intraclass novelty identification for medical images. Journal of Medical Imaging, 9(1): 014004-014004, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Tie Hu, Mingbao Lin, Lizhou You, Fei Chao, and Rongrong Ji. Discriminator-cooperated feature map distillation for gan" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20351-20360, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[15] Junho Kim, Minjae Kim, Hyeonwoo Kang, and Kwanghee Lee. U-gat-it: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[16] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[17] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[18] Fangda Li, Zhiqiang Hu, Wen Chen, and Avinash Kak. Adaptive supervised patchnce loss for learning h&e-to-ihc stain translation with inconsistent groundtruth image pairs. arXiv preprint arXiv:2303.06193, 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[19] Weijian Li, Viet-Duy Nguyen, Haofu Liao, Matt Wilder, Ke Cheng, and Jiebo Luo. Patch transformer for multi-tagging whole slide histopathology images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 532-540. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[20] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[21] Kechun Liu, Beibin Li, Wenjun Wu, Caitlin May, Oliver Chang, Stevan Knezevich, Lisa Reisch, Joann Elmore, and Linda Shapiro. Vsgd-net: Virtual staining guided melanocyte detection on histopathological images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1918-1927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[22] Kangning Liu, Weicheng Zhu, Yiqiu Shen, Sheng Liu, Narges Razavian, Krzysztof J Geras, and Carlos Fernandez-Granda. Multiple instance learning via iterative self-paced supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355–3365, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[23] Shuting Liu, Baochang Zhang, Yiqing Liu, Anjia Han, Huijuan Shi, Tian Guan, and Yonghong He. Unpaired stain transfer using pathology-consistent constrained generative" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11267" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "adversarial networks. IEEE Transactions on Medical Imaging, 40(8):1977-1989, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[24] Shengjie Liu, Chuang Zhu, Feng Xu, Xinyu Jia, Zhongyue Shi, and Mulan Jin. Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1815-1824, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[25] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[26] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[27] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.468, + 0.433 + ], + "angle": 0, + "content": "[28] Alvin I Mushlin, Ruth W Kouides, and David E Shapiro. Estimating the accuracy of screening mammography: a meta-analysis. American journal of preventive medicine, 14(2): 143-153, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.468, + 0.49 + ], + "angle": 0, + "content": "[29] Linhao Qu, Manning Wang, Zhijian Song, et al. Bidirectional weakly supervised knowledge distillation for whole slide image classification. Advances in Neural Information Processing Systems, 35:15368-15381, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[30] Jesus Salido, Noelia Valdez, Lucia González-López, Oscar Deniz, and Gloria Bueno. Comparison of deep learning models for digital h&E staining from unpaired label-free multispectral microscopy images. Computer Methods and Programs in Biomedicine, 235:107528, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.468, + 0.631 + ], + "angle": 0, + "content": "[31] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[32] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.787 + ], + "angle": 0, + "content": "[33] Lulin Shi, Ivy HM Wong, Claudia TK Lo, Lauren WK Tsui, and Terence TW Wong. Unsupervised multiple virtual histological staining from label-free autofluorescence images. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[34] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[35] Elena Y Trizna, Aleksandr M Sinitca, Asya I Lyanova, Diana R Baidamshina, Pavel V Zelenikhin, Dmitrii I Kaplun," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Airat R Kayumov, and Mikhail I Bogachev. Brightfield vs fluorescent staining dataset-a test bed image set for machine learning based virtual staining. Scientific Data, 10(1):160, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[36] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[37] Marek Wodzinski and Henning Müller. Deephistreg: Unsupervised deep learning registration framework for differently stained histology samples. Computer methods and programs in biomedicine, 198:105799, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[38] Georg Wölfein, In Hwa Um, David J Harrison, and Ognjen Arandjelović. Hoechstgan: Virtual lymphocyte staining using generative adversarial networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4997-5007, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[39] Xuan Xia, Xizhou Pan, Nan Li, Xing He, Lin Ma, Xiaoguang Zhang, and Ning Ding. Gan-based anomaly detection: A review. Neurocomputing, 493:497-535, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[40] Zhaoyang Xu, Xingru Huang, Carlos Fernández Moro, Béla Bozóky, and Qianni Zhang. Gan-based virtual re-staining: a promising solution for whole slide image analysis. arXiv preprint arXiv:1901.04059, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[41] Bowei Zeng, Yiyang Lin, Yifeng Wang, Yang Chen, Jiuyang Dong, Xi Li, and Yongbing Zhang. Semi-supervised pr virtual staining for breast histopathological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 232-241. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[42] Ranran Zhang, Yankun Cao, Yujun Li, Zhi Liu, Jianye Wang, Jiahuan He, Chenyang Zhang, Xiaoyu Sui, Pengfei Zhang, Lizhen Cui, et al. Mvfstain: multiple virtual functional stain histopathology images generation based on specific domain mapping. Medical Image Analysis, 80:102520, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[43] Yijie Zhang, Luzhe Huang, Tairan Liu, Keyi Cheng, Kevin de Haan, Yuzhu Li, Bijie Bai, and Aydogan Ozcan. Virtual staining of defocused autofluorescence images of unlabeled tissue using deep neural networks. Intelligent Computing, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[44] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11268" + } + ] +] \ No newline at end of file diff --git a/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_origin.pdf b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5249934501c20db5430225536460ec8e843ea379 --- /dev/null +++ b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/0012216a-0bd1-4dfe-bd44-a5f2cb587723_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b8bfa04bc50c7ce16e7eb14a941d9932516040051a14ec56f85d4719abb5d16 +size 6795713 diff --git a/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/full.md b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d1417086f7f51821d1f21a94d91aa5ff70ae297 --- /dev/null +++ b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/full.md @@ -0,0 +1,296 @@ +# Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning + +Jiahan Li $^{1}$ , Jiuyang Dong $^{1}$ , Shenjin Huang $^{1}$ , Xi Li $^{3}$ , Junjun Jiang $^{1}$ , Xiaopeng Fan $^{1*}$ , Yongbing Zhang $^{2*}$ + +$^{1}$ Harbin Institute of Technology, $^{2}$ Harbin Institute of Technology, Shenzhen Department of Gastroenterology, Peking University Shenzhen Hospital + +{jiahan.li,jiuyang.dong，shenjinhuang}@stu.hit.edu.cn, + +{junjunjiang, fxp, ybzhang08}@hit.edu.cn, lixi122188@sina.com + +# Abstract + +Recently, virtual staining technology has greatly promoted the advancement of histopathology. Despite the practical successes achieved, the outstanding performance of most virtual staining methods relies on hard-to-obtain paired images in training. In this paper, we propose a method for virtual immunohistochemistry (IHC) staining, named confusion-GAN, which does not require paired images and can achieve comparable performance to supervised algorithms. Specifically, we propose a multi-branch discriminator, which judges if the features of generated images can be embedded into the feature pool of target domain images, to improve the visual quality of generated images. Meanwhile, we also propose a novel patch-level pathology information extractor, which is assisted by multiple instance learning, to ensure pathological consistency during virtual staining. Extensive experiments were conducted on three types of IHC images, including a high-resolution hepatocellular carcinoma immunohistochemical dataset proposed by us. The results demonstrated that our proposed confusionGAN can generate highly realistic images that are capable of deceiving even experienced pathologists. Furthermore, compared to using H&E images directly, the downstream diagnosis achieved higher accuracy when using images generated by confusion-GAN. Our dataset and codes will be available at https://github.com/jiahanli2022/confusionGAN. + +# 1. Introduction + +Histopathological staining is a key step in clinical pathological analysis [2, 33]. As the most common histopathological staining method, H&E staining can clearly delineate + +the cellular structures in tissues. At the same time, immunohistochemistry (IHC) staining is a functional staining method used for evaluation of protein-specific expression, which greatly assists in tumor diagnosis and cancer prognosis [1]. For example, breast cancer with overexpression of human epidermal growth factor receptor 2 (HER2) tends to have aggressive clinical behaviour [24], while breast cancer with low expression of estrogen receptor (ER) often accompanies poor response to endocrine therapy [18]. Besides, glypican-3 (GPC3) is a recently discovered immunohistochemical protein that shows promising potential for indicating the presence of hepatocellular carcinoma. + +Unfortunately, the process of IHC staining is laborious, time-consuming, and expensive [42]. Therefore, many researches have focused on utilizing virtual staining techniques to convert H&E images into IHC images [38, 41, 43]. The existing virtual IHC staining methods can be roughly divided into two categories: supervised learning-based methods and unsupervised learning-based methods. Supervised learning-based methods require a large amount of paired images for training. These H&E-IHC pairs typically come from adjacent layers of same tissues, accompanied by complex data processing and alignment. What's more, these methods cannot be used in scenarios where adjacent layers are not available. On the other hand, unsupervised methods often focus on the style of the generated images, ignoring the transmission of pathological information. + +To address these aforementioned problems, we propose a virtual IHC staining framework called confusion-GAN, which utilizes weakly supervised learning to facilitate the unsupervised staining process of unpaired images. To explore the pathological information of H&E patches, a novel patch-level pathology information extractor which is assisted by multiple instance learning (MIL) is proposed. This method utilizes a dual-spherical loss proposed by us to push positive patches away from negative patches in the feature + +space, aiming to achieve accurate acquisition of pathological information. When training the confusion-GAN, we constrain the pathological consistency between the staining results and the input images. In addition, we propose a multi-branch discriminator, which contains a confusion discriminator as well as a common single-image discriminator, aiming to enhance the fidelity of the staining results. Specifically, we first combine a generated image with multiple real images to create an image pool in each iteration. The confusion discriminator aims to distinguish the unique generated image from this pool. Simultaneously, the generator tries to hinder this judgment of the confusion discriminator. + +To demonstrate the effectiveness of our proposed approach, we conducted extensive experiments on three datasets: 1) converting H&E staining images to HER2 staining images over the BCI dataset [24], 2) converting H&E staining images to ER staining images over the MIST dataset [18], and 3) convert H&E staining images to GPC3 staining images over our collected hepatocellular carcinoma immunohistochemical (HCI) dataset. In terms of image quality and pathological information transmission, our confusion-GAN can achieve superior performance compared to existing state-of-the-art methods and deceive human pathologists to some extent. Additionally, we compared the classification accuracy of ABMIL [13], a widely used algorithm for pathological image diagnosis, by using GPC3 images generated from our confusion-GAN and the original H&E images as the input data. The experimental results clearly indicate that the utilization of generated GPC3 images can significantly improve the diagnosis performance. + +The contributions are summarized as follows: + +- We propose an algorithm that utilizes weakly supervised learning to assist unsupervised virtual staining for the first time, which generates satisfactory images on three virtual IHC staining tasks. +- We propose a novel patch-level pathology information extractor (PPIE) for H&E image classification. As a plug-and-play method, PPIE is demonstrated to significantly improve the instance-level classification performance when combined with existing MIL methods. +- We propose a multi-branch discriminator (MBD) that can significantly enhance the fidelity of the staining results compared to only using a single-image discriminator. +- We collected a high-resolution hepatocellular carcinoma immunohistochemical dataset, which is contains both H&E and GPC3 staining images for the first time. + +# 2. Related Work + +# 2.1. Virtual IHC Staining Techniques + +Existing virtual IHC staining methods can be broadly divided into two categories: supervised and unsupervised + +learning based methods. Supervised learning based methods require a large number of H&E-IHC pairs as training images [8, 18, 24]. There are two ways of obtaining these paired images: (1) Carrying out axially consecutive scanning of tissues, actually staining H&E and IHC on adjacent slices, then performing alignment between H&E and IHC slides through complex data processing and registration to obtain the paired images for training. For example, Liu et al. [24] train their proposed PyramidP2P network over a registered H&E-HER2 dataset. However, the tissue contents from different layers are inherently difficult to be consistent, making pixel-level alignment impossible. Therefore, the inability to fully align will inevitably result in performance limitation. In addition, such supervised methods are unlikely to be applicable when adjacent layer data is unavailable. (2) Using images of a third modality other than H&E and IHC, such as microscopic fluorescence data, as a bridge, to generate pixel-wise aligned H&E and IHC training data [8]. Unfortunately, access to the third modality data will also add additional overhead. + +The vast majority of existing unsupervised virtual IHC staining methods are based on image translation networks [5, 9, 21, 30, 33, 35, 40]. Although these networks can convert H&E-style images into IHC-style images, it does not mean that the Pos/Neg properties of the translated results will be consistent with that of their inputs. To keep the aforementioned pathological consistency, Liu et al. [23] and Boyd et al. [3] introduce expert annotations, which is an expensive solution. + +# 2.2. Patch-level Classification Assisted by MIL + +Currently, MIL are used for patch-level classification [4, 13, 16, 19, 32]. In the case of pathological images, each WSI is regarded as a bag, and each patch within the WSI is treated as an instance. A bag label is assigned to each bag, which is classified as positive if the bag contains at least one positive instance. These bag labels serve as the ground truths for training MIL networks in an end-to-end manner. Existing methods determine whether a patch is positive or not by comparing the attention scores (or instance-level predictions) of all patches [6, 20, 22, 29, 31, 34]. Unfortunately, these methods can only ensure accurate patch-level classification within a limited region of high attention scores. Therefore, we propose a novel patch-level pathology information extractor to compensate for the shortcomings of these methods. + +# 3. Proposed Method + +# 3.1. Framework + +One significant difference between the virtual IHC staining and the image translation of natural images is that generated IHC images must maintain pathological consistency with + +![](images/7a5cb8d4624bffc597ea3e917395960d082fe807f267f124806fadccf2ddd86c.jpg) + +![](images/23ee533a12ea7a3d7cdc944a5e3c9f7dcce8944daa6c79495a48887faae56f31.jpg) +Figure 1. Structure of the proposed confusion-GAN. (a) The confusion-GAN contains two important modules: patch-level pathology information extractor (PPIE) and multi-branch discriminator (MBD). (b) The training process of PPIE. (c) The structure of the MBD. + +![](images/861113c4b9c2670c4ef619fd59ffa13db5345f439378e9eaea1f65d14073e5a6.jpg) + +their inputs while preserving texture consistency. From the perspectives of improving the quality of generated images and accurately transmitting pathological information, we propose confusion-GAN, which is illustrated in Fig. 1. The confusion-GAN contains two important modules, patch-level pathology information extraction module and multibranch discriminator, which are employed to maintain the pathological consistency and generate more realistic images, respectively. + +As shown in Fig. 1, the real H&E image $x_{i}$ is first passed through the generator $G_{A}$ to obtain the IHC-style image $\hat{y}_i$ . Then, $\hat{y}_i$ is further remapped into the H&E-style image $\tilde{x}_i$ using the generator $G_{B}$ . Similarly, we can obtain $\hat{x}_i$ and $\tilde{y}_i$ from real IHC image $y_{i}$ . + +During the process of generating IHC image from the H&E image, we devise a Pathology Loss to ensure the accurate transmission of pathological information. IHC classifier $Cls_{ihc}$ and a PPIE module are designed to obtain the pathological information of the H&E image $x_{i}$ and the generated IHC image $\hat{y}_{i}$ , respectively. To pretrain a reliable IHC classifier, we simulate the annotation process of pathologists through color analysis to obtain patch-level pos/neg labels for actually captured IHC images: patches with positive exhibit protein-specific expression within more than 1% of the area (corresponding to yellow-brown coloration + +in the image). A light-weight image classifier $Cls_{ihc}$ is then designed utilizing the pos/neg labels obtained by the previously described process, which consists of 6 convolutional layers activated by ReLU and a fully connected layer (FCN) for the final classification. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1. Similarly, the pathological information of the H&E image is obtained utilizing the PPIE module, which will be elaborated in Sec. 3.2. Finally, we adopt the Pathology Loss to constrain the consistency of pathological information between $x_{i}$ and $\hat{y}_{i}$ : + +$$ +\begin{array}{l} L _ {\text {p a t h}} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \left(P P I E \left(x _ {i}\right) \log \left(C l s ^ {i h c} \left(\hat {y} _ {i}\right)\right) \right. \tag {1} \\ + (1 - P P I E \left(x _ {i}\right)) \log \left(1 - C l s ^ {i h c} \left(\hat {y} _ {i}\right)\right), \\ \end{array} +$$ + +where $PPI(\cdot)$ represents the operation of PPI module. + +In addition, to enhance the quality of the generated images, we propose a multi-branch discriminator for both H&E-to-IHC and IHC-to-H&E process in the confusionGAN. Detailed description of the multi-branch discriminator can be found in Sec. 3.3. + +In general, the hybrid loss for the generator is as follows: + +$$ +\begin{array}{l} L _ {G} = \alpha * L _ {\text {p a t h}} + \beta * L _ {G} ^ {C} + \lambda * L _ {G} ^ {A d v} \\ + \eta * L _ {G} ^ {C y c l e} + \iota * L _ {G} ^ {\text {I d e n t i t y}}, \\ \end{array} +$$ + +where $L_{G}^{Adv}$ , $L_{G}^{Cycle}$ , and $L_{G}^{Identity}$ represent the adversarial loss, cycled loss, and identity loss similar to CycleGAN [44], while $\alpha$ , $\beta$ , $\lambda$ , $\eta$ , and $\iota$ are hyper-parameters controlling the weighting coefficients. $L_{G}^{C}$ is the loss function used to guide the generator optimization with the confusion discriminator, which will be described in detail in Sec. 3.3. + +# 3.2. Patch-level Pathology Information Extraction + +Due to memory limitations, virtual staining can only be performed at the patch level. Unfortunately, only WSI-level labels are available for majority cases, since one WSI usually contains gigapixel, impeding the patch-wise annotation in clinical. Therefore, it is necessary to extract patch-level pathological information (i.e., the pos/neg category of each patch) from the H&E WSI to ensure accurate virtual staining. + +For H&E images that only have WSI level labels (lacking patch-level annotations), the classification of each patch is typically completed through attention scores or instance-level classification predictions within the MIL framework. In each positive WSI, there is a high likelihood that patches with the highest attention scores or instance-level classification scores will be positive. We refer to the action of selecting the top $k$ patches of each positive WSI based on the highest attention scores (or instance-level prediction scores) as "top-k selection". When $k$ is relatively small, it is ensured that the selected patches are all positive; when $k$ is large, the selected collections may include many negative patches [26]. Therefore, the current methods have not been able to accurately classify all patches. + +To address this problem, building upon the top-k selection, we propose a plug-and-play PPIE module that can improve the classification performance of H&E patches. Specifically, we first perform top-k selection to obtain $X_{pos}^{he}$ from positive WSIs in the training set. Next, we crop a set of negative patches $X_{neg}^{he}$ from negative WSIs. It should be noted all patches in a negative WSI are negative, unlike the cases of positive WSIs where both positive and negative patches may coexist. Besides, we utilized a state-of-the-art diffusion model BBDM [17] to learn how to generate a distribution similar to that of $X_{pos}^{he}$ using $X_{neg}^{he}$ . The set of images generated by BBDM is referred to as $X_{pos}^{he}$ . + +On this basis, we borrow the idea of anomaly detection to tackle the classification problem of patch-level H&E images [10]. In each iteration of training the PPIE, we sample $b$ images from $X_{\text{neg}}^{\text{he}}$ , $X_{\text{pos}}^{\text{he}}$ , and $X_{\text{pos}}^{\text{he}}$ respectively, and feed them into the encoder to obtain the feature sets $C_1 = \{c_{1i}\}_{i=1}^b$ , $C_2 = \{c_{2i}\}_{i=1}^b$ , and $C_3 = \{c_{3i}\}_{i=1}^b$ , as shown in Fig. 1(b). Here, $b$ represents the batch size during training. The encoder consists of 6 convolutional layers activated by ReLU and a fully connected layer with an output dimension of 512. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1. + +As is well known, the success of the anomaly detection depends on two aspects: 1) the samples in the in-distribution (ID) class should be as close as possible in the feature space; 2) the samples in the out-of-distribution (OOD) class should be as far as possible from samples in the ID class [39]. We use the loss function $L_{ID}$ to force all the samples in the ID class to be distributed as close to the center as possible: + +$$ +L _ {I D} = \frac {1}{b} \sum_ {i = 1} ^ {b} \left\| c _ {1 i} - O \right\| ^ {2}, \tag {3} +$$ + +where $O$ represents the average features of all the ID data $X_{neg}^{he}$ . It should be noted that $O$ will be cumulatively updated during training. Besides, we propose a double-spherical loss $L_{OOD}$ to ensure that the overall distribution of $C_3$ has a greater distance from $O$ compared to the overall distribution of $C_2$ : + +$$ +\begin{array}{l} L _ {O O D} = \frac {1}{b} \sum_ {i = 1} ^ {b} \max \left(R _ {1} - \left\| c _ {2 i} - O \right\| ^ {2}, 0\right) \tag {4} \\ + \frac {1}{b} \sum_ {i = 1} ^ {b} m a x (R _ {2} - | | c _ {3 i} - O | | ^ {2}, 0). \\ \end{array} +$$ + +We set the hyperparameter $R_{2}$ to control the distance and ensure that $R_{2}$ is greater than $R_{1}$ , as shown in Fig. 1(b). Employing $L_{ID}$ and $L_{OOD}$ , we can ensure that $X_{neg}^{he}$ and $X_{pos}^{he}$ have strong separability in the feature space. Finally, we treat $X_{neg}^{he}$ and $X_{pos}^{he}$ as ID and OOD classes respectively, and train the patch-level classifier $Cls^{he}$ by cross entropy loss $L_{cls}$ : + +$$ +\begin{array}{l} L _ {c l s} = - \frac {1}{n} \sum_ {i = 1} ^ {s} \left(l _ {i} \log \left(C l s ^ {h e} \left(x _ {i} ^ {h e}\right)\right) \right. \tag {5} \\ + \left(1 - l _ {i}\right) \log \left(1 - C l s ^ {h e} \left(x _ {i} ^ {h e}\right)\right)). \\ \end{array} +$$ + +When using PPIE for inference, negative patches are considered as ID class, while positive patches are considered as OOD class. + +Overall, the hybrid loss function $L_{PPIE}$ for training the PPIE module is as follows: + +$$ +L _ {P P I E} = \mu * L _ {I D} + \tau * L _ {O O D} + v * L _ {c l s}, \tag {6} +$$ + +where $\mu, \tau,$ and $v$ are hyper-parameters to control the weighting coefficients. + +# 3.3. Multi-branch Discriminator + +To achieve more realistic image generation, we propose a multi-branch discriminator, as shown in Fig. 1(c). Taking generating an IHC-style image from a H&E image as an example, we send the result $\hat{y}_i$ obtained by the generator $G_{A}$ together with a reference pool $P_i^{ihc}$ consisting of + +$N - 1$ real images from the same domain to the confusion discriminator $C_B$ . $C_B$ shares the feature extraction layer $D_B^S$ with the single-image discriminator $D_B$ , which helps eliminate the interference of the content of the reference pool images on the confusion discriminator. This one-to-many discrimination paradigm requires the generator's results to possess more common attributes of the current domain images, thereby forcing the generated distribution to approximate the real distribution more closely. + +During the training of the $C_B$ , we aim to accurately distinguish between the generated result and real images. Then we randomly insert $\hat{y}_i$ into a position in the list of $P_i^{ihc}$ , and finally concatenate them in the channel dimension to obtain the input $m_i^{ihc}$ for $C_B$ . Correspondingly, we construct an $N$ -dimensional vector $l_i^{ihc}$ to represent the labels of each element in $m_i^{ihc}$ , with 0 indicating real image and 1 indicating generated image. Next, we pass $m_i^{ihc}$ into $D_B^S$ to obtain a feature pool, which is fed into an MLP layer of $C_B$ for liveness verification. $C_B$ is then updated by minimizing: + +$$ +L _ {C _ {B}} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \left| C _ {B} \left(m _ {i} ^ {i h c}\right) - l _ {i} ^ {i h c} \right| \right| ^ {2}. \tag {7} +$$ + +Similarly, during the process of IHC-to-H&E image translation, the loss function $L_{C_A}$ of the confusion discriminator $C_A$ is defined as follows: + +$$ +L _ {C _ {A}} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \left| C _ {A} \left(m _ {i} ^ {h e}\right) - l _ {i} ^ {h e} \right| \right| ^ {2}, \tag {8} +$$ + +where $m_{i}^{he}$ represents the concatenation result of the generated H&E image and $N - 1$ real H&E images, and $l_{i}^{he}$ represents the corresponding labels of $m_{i}^{he}$ . The proposed confusion discriminators are trained using the following formula: + +$$ +L _ {C} = L _ {C _ {B}} + L _ {C _ {A}}. \tag {9} +$$ + +The loss function used for training $D_B$ and $D_A$ is consistent with CycleGAN. + +When training the generator, we obtain $m_i^{ihc}$ and $m_i^{he}$ in the same manner. However, in this stage, we want the confusion discriminator unable to identify the generator's results from these $N$ images. Therefore, all labels in $l_i^{ihc}$ and $l_i^{he}$ are set to 0. The generators are updated based on the following loss functions: + +$$ +\begin{array}{l} L _ {G} ^ {C} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \left| C _ {B} \left(m _ {i} ^ {i h c}\right) - l _ {i} ^ {i h c} \right| \right| ^ {2} \tag {10} \\ + \left\| C _ {A} \left(m _ {i} ^ {h e}\right) - l _ {i} ^ {h e} \right\| ^ {2}. \\ \end{array} +$$ + +In each iteration, we train the generators $G_{A}$ and $G_{B}$ , $D_{A}$ and $D_{B}$ , and our proposed $C_{A}$ and $C_{B}$ in an alternating manner, following a sequential order. This process is similar to what most generative adversarial networks do. Here + +$G_{A}$ and $G_{B}$ employ the unet256 network used in CycleGAN and the structures of $D_{A}$ and $D_{B}$ are also consistent with those in CycleGAN. + +# 4. Experiments + +# 4.1. Setup + +Datasets. GPC3, discovered as a tumor marker in recent years, shows great potential for specific expression in Hepatocellular carcinoma (HCC). We have collected a HCC Immunohistochemical Image dataset named HCI, which includes both the H&E and GPC3 images. All data comes from Peking University Shenzhen Hospital. We selected 20 H&E-IHC image pairs from a pool of WSIs for testing and validation, and perform registration on them using the DeepHistReg [37]. Next, we conducted separate screenings for a total of 30 H&E images and 30 IHC images, and these images are used for training. Furthermore, we segment these 100 WSIs into approximately 1.4 million patches sized of $256 \times 256$ in a non-overlapping manner at 20x magnification. All H&E images have undergone staining normalization. We also verified our method on public datasets: BCI [24] and MIST [18]. + +Metrics. We evaluated the image quality of the staining results using objective metrics such as Frechet Inception Distance (FID) [27], Learnable Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index (SSIM) [11]. Additionally, we invited three pathologists to provide subjective evaluations of the staining results. In assessing the patch-level classification performance of the aforementioned PPIE, we utilized metrics such as area under curve (AUC) [25], False Positive Rate (FPR) [28], True Positive Rate (TPR), and Geometric Mean (G-mean) [7]. + +Implementation Details. We trained the PPIE using the Adam optimizer with a learning rate of 0.001. The dual-sphere distances $R_{1}$ and $R_{2}$ in Eq. (4) are set to 150 and 262.5 respectively. The hyper-parameters $\mu$ , $\tau$ , and $\upsilon$ in Eq. (6) are all set to 1. When training the virtual staining network, $\alpha$ , $\beta$ , $\lambda$ , $\eta$ and $\iota$ in Eq. (2) are set to 1, 1, 1, 10, and 5 respectively. We set the reference pool size $N$ as 32 and kept the other configuration of the staining network consistent with CycleGAN. In the downstream experiments, we trained the ABMIL model on a re-partitioned dataset for a maximum of 15 epochs. For all experiments, we trained and tested our confusion-GAN on patches of size $256 \times 256$ . + +# 4.2. Staining Performance + +HCI. As shown in Fig. 2, we compared our method with three unsupervised image translation methods, namely CycleGAN [44], UGATIT [15], and DCD [12], on the HCI dataset. The first column in Fig. 2 displays the input H&E images, while the last column shows actually stained IHC (GPC3) images for reference. These images are from corre + +![](images/13b54aa30a1deb61bdbb5a9935b55b4971c4f8c473ad5d90e3620c0c26a651a1.jpg) +Figure 2. Comparison of the abilities of different unsupervised algorithms to preserve pathological consistency on HCI. The red dashed box indicates the area where the generated image cannot maintain consistent pathological information with the input H&E image. + +![](images/59b48f485350d25439c25c8b3a0e5efb8769c30cc4d750fb61824d00fd1207a8.jpg) +Figure 3. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on BCI. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors. + +sponding regions of adjacent slices in the same tissue. We marked in red dashed lines the regions where the pathological information is inconsistent with the input. Obviously, although the competing methods CycleGAN, UGATIT, and DCD can maintain the texture details of the input images well, there are significant mistaining areas since the consistency of pathological information can not be ensured before and after staining. This phenomenon demonstrates the effectiveness of the proposed PPIE module. Besides, we use three objective metrics, FID, LPIPS, and SSIM to evaluate these virtual staining methods. As shown in Tab. 1, our method achieves the best performance in all the three metrics, indicating that the output distribution of the confusion-GAN is closest to the distribution of the real images. + +BCI and MIST. BCI and MIST datasets provide paired + +
MethodFID↓LPIPS↓SSIM↑
CycleGAN(unet)105.40.6360.1870
CycleGAN(resnet)90.90.6220.2086
UGATIT95.30.6020.2223
DCD91.30.6120.2266
Ours77.30.5700.2300
+ +Table 1. The evaluation of different algorithms on HCI. + +H&E-IHC patch images, where the acquisition of these patches relies on alignment with the sliced images from adjacent layers within the same tissue. However, the tissue contents from different adjacent layers are inherently impossible to be completely consistent, making pixel-level alignment impossible. We compared our method with four fully supervised methods (Pix2pix [14], Pix2pixHD [36], + +![](images/6bca552fd26dc44d1aa5c1686e70b84f3953e065b421bb93c3466e5a2a0b4c5d.jpg) +Figure 4. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on MIST. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors. + +PyramidP2P [24], ASP [18]) on these two datasets, as shown in Fig. 3 and Fig. 4. We also marked the areas with pathological information errors in the staining results with dashed lines. Additionally, we marked the areas with content errors in the staining results with solid lines. From Fig. 3 and Fig. 4, it can be observed that Pix2pix, Pix2pixHD, and PyramidP2P methods are unable to maintain the textures of the input images, often resulting in significant presence of artifacts in the generated results. This phenomenon is attributed to the significant discrepancy in content between the input H&E images and the reference IHC images used as the ground truth during the training process. The ASP based on contrastive learning demonstrates good performance in preserving content. However, the misalignment of supervised information can still introduce some errors in the pathological information of the staining results. In contrast, our proposed confusion-GAN has achieved better results than these supervised learning methods. + +We used FID, LPIPS, and SSIM for objective evaluation. As can be seen from Tab. 2, our confusion-GAN achieved the best FID and LPIPS scores on both datasets, which indicating the effectiveness of our algorithm. Regarding the SSIM evaluation metric, which is directly calculated in the pixel space, our method still achieved the highest score on MIST dataset. Meanwhile, our method can achieve the best visual effects on both BCI and MIST, especially considering the fidelity of textures in the generated results. It is worth mentioning that, due to the lack of complete WSIs in MIST and BCI, we assigned labels to IHC images using the color analysis mentioned in Sec. 3.1, and then transfer the labels of IHC images to the corresponding H&E images. Regarding to BCI, we directly utilized the provided patch-level la + +bels, which guides the training of confusion-GAN. + +
MethodsBCIMIST
FID↓LPIPS↓SSIM↑FID↓LPIPS↓SSIM↑
Pix2pix110.70.4800.495780.90.5290.2127
Pix2pixHD106.80.5050.4766131.00.5640.2129
PyramidP2P99.60.5050.466397.50.5380.2061
ASP54.30.5030.492341.20.5320.2062
Ours49.20.4780.457939.80.5160.2174
+ +# 4.3. Ablation Study + +We explored the working mechanism of PPIE and analyzed how PPIE and the MBD affect the staining results. + +Table 2. Comparing the performance of our method and supervised algorithms on BCI and MIST. + +
MethodkAUC↑FPR↓TPR↑G-mean↑
ABMIL100.7130.2770.6410.680
300.7820.2260.7260.750
1000.6370.4080.5890.591
wo0.7320.3230.6860.681
Maxpool100.6820.3240.6660.671
300.7310.2570.6890.715
1000.7060.2670.6670.699
wo0.6800.3280.6130.680
ClamMb100.7660.2380.6910.726
300.7700.2190.7080.744
1000.7320.2860.6270.669
wo0.7260.3380.6870.675
Meanpool100.7450.2530.6710.708
300.7680.2390.7280.744
1000.7090.2540.6190.679
wo0.7580.3240.7100.693
+ +Table 3. Patch-level classification results obtained from different top-k patches on HCI. + +Effect of $k$ in the top-k selection. We selected four clas + +sic MIL methods (ABMIL [13], MaxPool, CLAM-MB[26], and MeanPool) for top-k selection in PPIE training. As shown in Tab. 3, we evaluate the classification performance of PPIE using AUC, FPR, TPR, and G-mean. We found that the classification performance is best when $k$ is set to 30. This is because when $k$ is too small, the PPIE cannot obtain sufficient training data. On the other hand, when $k$ is too large, the top-k selection results will contain a large number of negative patches, which hinders the effectiveness of the double-spherical loss. Among these four MIL methods, we obtained similar conclusions and improved the patch-level classification performance, achieving gains of $5.0\%$ , $5.1\%$ , $4.4\%$ , and $1.0\%$ respectively. + +Effect of PPIE and MBD. To investigate how PPIE and MBD impact the performance, we conducted an ablation experiment on HCI, as shown in Fig. 5. By comparing the second, third, and fourth columns in Fig. 5, it can be observed that the PPIE plays a crucial role in accurately transmitting pathological information. On the other hand, the MBD can enhance the clarity of cell nuclei and surrounding tissues, affecting the overall style of the generated results. What's more, based on the qualitative analysis of the indicators in Tab. 4, we found that the method combining PPIE and the MBD achieved the best results. + +
PPIEXX✓✓
MBDX✓X✓
FID↓105.482.4103.377.3
LPIPS↓0.6360.5880.5740.570
SSIM↑0.18700.22230.22940.2300
+ +# 4.4. Down-stream Classification Performance + +We conducted downstream experiments related to the HCC diagnosis, and compared the classification performance achieved by using the H&E and generated GPC3 slides as inputs of the ABMIL model, respectively. When using H&E images as input achieves an AUC of $55.6\%$ , while using the generated images as input achieves an AUC of $88.9\%$ . This indicates that compared to H&E images, our generated GPC3 images can improve the classification accuracy of HCC. + +# 4.5. Subjective Evaluation by Pathologists + +We invited three pathologists to provide subjective evaluations on two aspects: transmission of pathological information and the quality of the images. We randomly selected 50 pairs of generated GPC3 images along with their corresponding adjacent tissues and then asked pathologists to evaluate whether the protein expression levels were consistent between the two. As shown in the first row of Tab. 5, in most cases, pathologists believed that the protein expression + +![](images/03a34eb11974582e2152893ccdf8f5a60ca643cb2de74790604f262e3df2c700.jpg) +Figure 5. Ablation study: effect of PPIE and MBD on HCI. The red box indicates regions with pathological information errors. + +levels were consistent between them. This indicates that our confusion-GAN can effectively maintain pathological consistency. Afterwards, we randomly selected 50 GPC3 images and 50 generated GPC3 images for pathologists to distinguish which are generated images. From the second row of the Tab. 5, it can be seen that doctors cannot distinguish the generated images well, indicating that our confusion-GAN can generate images comparable to real images. + +Table 4. Exploring the impact of PPIE and MBD on HCI. + +
Accuracypathologist1pathologist2pathologist3
Neg/Pos0.860.800.84
Fake/True0.620.480.52
+ +Table 5. Subjective evaluations of pathologists on HCI. + +# 5. Conclusion + +In this paper, we propose a method called confusion-GAN, which is the first to utilize weakly-supervised learning to assist in virtual IHC staining of unpaired pathological images. Extensive experiments have demonstrated that utilizing patch-level pathological information to guide the training of staining network is crucial for achieving accurate virtual IHC staining. Furthermore, we propose a multi-branch discriminator for achieving high-fidelity image generation. Our approach achieved the state-of-the-art results on multiple datasets. To advance the diagnosis of HCC, we have also proposed the first GPC3 dataset and demonstrated that the images generated by confusion-GAN are more helpful in HCC detection compared to H&E images. We believe that unsupervised methods like confusion-GAN, which do not require fine annotation or complex preprocessing, will be crucial in virtual staining research in the future. + +Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62031023&62331011), in part by the Shenzhen Science and Technology Project (GXWD20220818170353009), and in part by the Fundamental Research Funds for the Central Universities (Grant No. HIT.OCEF.2023050). + +# References + +[1] Bijie Bai, Xilin Yang, Yuzhu Li, Yijie Zhang, Nir Pillar, and Aydogan Ozcan. Deep learning-enabled virtual histological staining of biological samples. Light: Science & Applications, 12(1):57, 2023. 1 +[2] Marian Boktor, Benjamin R Ecclestone, Vlad Pekar, Deepak Dinakaran, John R Mackey, Paul Fieguth, and Parsin Haji Reza. Virtual histological staining of label-free total absorption photoacoustic remote sensing (ta-pars). Scientific Reports, 12(1):10296, 2022. 1 +[3] Joseph Boyd, Irène Villa, Marie-Christine Mathieu, Eric Deutsch, Nikos Paragios, Maria Vakalopoulou, and Stergios Christodoulidis. Region-guided cyclegans for stain transfer in whole slide images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 356-365. Springer, 2022. 2 +[4] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2 +[5] Tanujit Chakraborty, Shraddha M Naik, Madhurima Panja, Bayapureddy Manvitha, et al. Ten years of generative adversarial nets (gans): A survey of the state-of-the-art. arXiv preprint arXiv:2308.16316, 2023. 2 +[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 2 +[7] Yanyan Chen, Kuaini Wang, and Ping Zhong. One-class support tensor machine. Knowledge-Based Systems, 96:14-28, 2016. 5 +[8] Kevin de Haan, Yijie Zhang, Jonathan E Zuckerman, Tairan Liu, Anthony E Sisk, Miguel FP Diaz, Kuang-Yu Jen, Alexander Nobori, Sofia Liou, Sarah Zhang, et al. Deep learning-based transformation of h&e stained tissues into special stains. Nature communications, 12(1):4884, 2021. 2 +[9] Michael Gadermayr, Maximilian Tschuchnig, Lea Maria Stangassinger, Christina Kreutzer, Sebastien Couillard-Despres, Gertie Janneke Oostingh, and Anton Hittmair. Improving automated thyroid cancer classification of frozen sections by the aid of virtual image translation and stain normalization. Computer Methods and Programs in Biomedicine Update, 3:100092, 2023. 2 +[10] Xiaoyuan Guo, Judy W Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Margin-aware intraclass novelty identification for medical images. Journal of Medical Imaging, 9(1): 014004-014004, 2022. 4 +[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5 +[12] Tie Hu, Mingbao Lin, Lizhou You, Fei Chao, and Rongrong Ji. Discriminator-cooperated feature map distillation for gan + +compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20351-20360, 2023. 5 +[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 8 +[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 6 +[15] Junho Kim, Minjae Kim, Hyeonwoo Kang, and Kwanghee Lee. U-gat-it: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830, 2019. 5 +[16] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 2 +[17] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 4 +[18] Fangda Li, Zhiqiang Hu, Wen Chen, and Avinash Kak. Adaptive supervised patchnce loss for learning h&e-to-ihc stain translation with inconsistent groundtruth image pairs. arXiv preprint arXiv:2303.06193, 2023. 1, 2, 5, 7 +[19] Weijian Li, Viet-Duy Nguyen, Haofu Liao, Matt Wilder, Ke Cheng, and Jiebo Luo. Patch transformer for multi-tagging whole slide histopathology images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 532-540. Springer, 2019. 2 +[20] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 2 +[21] Kechun Liu, Beibin Li, Wenjun Wu, Caitlin May, Oliver Chang, Stevan Knezevich, Lisa Reisch, Joann Elmore, and Linda Shapiro. Vsgd-net: Virtual staining guided melanocyte detection on histopathological images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1918-1927, 2023. 2 +[22] Kangning Liu, Weicheng Zhu, Yiqiu Shen, Sheng Liu, Narges Razavian, Krzysztof J Geras, and Carlos Fernandez-Granda. Multiple instance learning via iterative self-paced supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355–3365, 2023. 2 +[23] Shuting Liu, Baochang Zhang, Yiqing Liu, Anjia Han, Huijuan Shi, Tian Guan, and Yonghong He. Unpaired stain transfer using pathology-consistent constrained generative + +adversarial networks. IEEE Transactions on Medical Imaging, 40(8):1977-1989, 2021. 2 +[24] Shengjie Liu, Chuang Zhu, Feng Xu, Xinyu Jia, Zhongyue Shi, and Mulan Jin. Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1815-1824, 2022. 1, 2, 5, 7 +[25] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5 +[26] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 4, 8 +[27] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 5 +[28] Alvin I Mushlin, Ruth W Kouides, and David E Shapiro. Estimating the accuracy of screening mammography: a meta-analysis. American journal of preventive medicine, 14(2): 143-153, 1998. 5 +[29] Linhao Qu, Manning Wang, Zhijian Song, et al. Bidirectional weakly supervised knowledge distillation for whole slide image classification. Advances in Neural Information Processing Systems, 35:15368-15381, 2022. 2 +[30] Jesus Salido, Noelia Valdez, Lucia González-López, Oscar Deniz, and Gloria Bueno. Comparison of deep learning models for digital h&E staining from unpaired label-free multispectral microscopy images. Computer Methods and Programs in Biomedicine, 235:107528, 2023. 2 +[31] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2 +[32] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2 +[33] Lulin Shi, Ivy HM Wong, Claudia TK Lo, Lauren WK Tsui, and Terence TW Wong. Unsupervised multiple virtual histological staining from label-free autofluorescence images. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 1, 2 +[34] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 2 +[35] Elena Y Trizna, Aleksandr M Sinitca, Asya I Lyanova, Diana R Baidamshina, Pavel V Zelenikhin, Dmitrii I Kaplun, + +Airat R Kayumov, and Mikhail I Bogachev. Brightfield vs fluorescent staining dataset-a test bed image set for machine learning based virtual staining. Scientific Data, 10(1):160, 2023. 2 +[36] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 6 +[37] Marek Wodzinski and Henning Müller. Deephistreg: Unsupervised deep learning registration framework for differently stained histology samples. Computer methods and programs in biomedicine, 198:105799, 2021. 5 +[38] Georg Wölfein, In Hwa Um, David J Harrison, and Ognjen Arandjelović. Hoechstgan: Virtual lymphocyte staining using generative adversarial networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4997-5007, 2023. 1 +[39] Xuan Xia, Xizhou Pan, Nan Li, Xing He, Lin Ma, Xiaoguang Zhang, and Ning Ding. Gan-based anomaly detection: A review. Neurocomputing, 493:497-535, 2022. 4 +[40] Zhaoyang Xu, Xingru Huang, Carlos Fernández Moro, Béla Bozóky, and Qianni Zhang. Gan-based virtual re-staining: a promising solution for whole slide image analysis. arXiv preprint arXiv:1901.04059, 2019. 2 +[41] Bowei Zeng, Yiyang Lin, Yifeng Wang, Yang Chen, Jiuyang Dong, Xi Li, and Yongbing Zhang. Semi-supervised pr virtual staining for breast histopathological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 232-241. Springer, 2022. 1 +[42] Ranran Zhang, Yankun Cao, Yujun Li, Zhi Liu, Jianye Wang, Jiahuan He, Chenyang Zhang, Xiaoyu Sui, Pengfei Zhang, Lizhen Cui, et al. Mvfstain: multiple virtual functional stain histopathology images generation based on specific domain mapping. Medical Image Analysis, 80:102520, 2022. 1 +[43] Yijie Zhang, Luzhe Huang, Tairan Liu, Keyi Cheng, Kevin de Haan, Yuzhu Li, Bijie Bai, and Aydogan Ozcan. Virtual staining of defocused autofluorescence images of unlabeled tissue using deep neural networks. Intelligent Computing, 2022. 1 +[44] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 4, 5 \ No newline at end of file diff --git a/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/images.zip b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf22ee4c4a88558b75671dea71cd427ca7909b44 --- /dev/null +++ b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1518037884b8078abd4d89061fab83962e7173dae33216319e74cff74960fa3a +size 716779 diff --git a/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/layout.json b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f373bd6de866d75215d8132fcb0ee5f63d25ad18 --- /dev/null +++ b/2024/Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning/layout.json @@ -0,0 +1,8906 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "text", + "content": "Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": "Jiahan Li" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": ", Jiuyang Dong" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": ", Shenjin Huang" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": ", Xi Li" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": ", Junjun Jiang" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": ", Xiaopeng Fan" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "text", + "content": ", Yongbing Zhang" + }, + { + "bbox": [ + 91, + 160, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 189, + 478, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 189, + 478, + 218 + ], + "spans": [ + { + "bbox": [ + 115, + 189, + 478, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 189, + 478, + 218 + ], + "type": "text", + "content": "Harbin Institute of Technology, " + }, + { + "bbox": [ + 115, + 189, + 478, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 189, + 478, + 218 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen Department of Gastroenterology, Peking University Shenzhen Hospital" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 220, + 460, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 220, + 460, + 232 + ], + "spans": [ + { + "bbox": [ + 147, + 220, + 460, + 232 + ], + "type": "text", + "content": "{jiahan.li,jiuyang.dong，shenjinhuang}@stu.hit.edu.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 234, + 460, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 234, + 460, + 245 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 460, + 245 + ], + "type": "text", + "content": "{junjunjiang, fxp, ybzhang08}@hit.edu.cn, lixi122188@sina.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 299, + 289, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 289, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 289, + 608 + ], + "type": "text", + "content": "Recently, virtual staining technology has greatly promoted the advancement of histopathology. Despite the practical successes achieved, the outstanding performance of most virtual staining methods relies on hard-to-obtain paired images in training. In this paper, we propose a method for virtual immunohistochemistry (IHC) staining, named confusion-GAN, which does not require paired images and can achieve comparable performance to supervised algorithms. Specifically, we propose a multi-branch discriminator, which judges if the features of generated images can be embedded into the feature pool of target domain images, to improve the visual quality of generated images. Meanwhile, we also propose a novel patch-level pathology information extractor, which is assisted by multiple instance learning, to ensure pathological consistency during virtual staining. Extensive experiments were conducted on three types of IHC images, including a high-resolution hepatocellular carcinoma immunohistochemical dataset proposed by us. The results demonstrated that our proposed confusionGAN can generate highly realistic images that are capable of deceiving even experienced pathologists. Furthermore, compared to using H&E images directly, the downstream diagnosis achieved higher accuracy when using images generated by confusion-GAN. Our dataset and codes will be available at https://github.com/jiahanli2022/confusionGAN." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 655, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 288, + 693 + ], + "type": "text", + "content": "Histopathological staining is a key step in clinical pathological analysis [2, 33]. As the most common histopathological staining method, H&E staining can clearly delineate" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 274, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 547, + 418 + ], + "type": "text", + "content": "the cellular structures in tissues. At the same time, immunohistochemistry (IHC) staining is a functional staining method used for evaluation of protein-specific expression, which greatly assists in tumor diagnosis and cancer prognosis [1]. For example, breast cancer with overexpression of human epidermal growth factor receptor 2 (HER2) tends to have aggressive clinical behaviour [24], while breast cancer with low expression of estrogen receptor (ER) often accompanies poor response to endocrine therapy [18]. Besides, glypican-3 (GPC3) is a recently discovered immunohistochemical protein that shows promising potential for indicating the presence of hepatocellular carcinoma." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 602 + ], + "type": "text", + "content": "Unfortunately, the process of IHC staining is laborious, time-consuming, and expensive [42]. Therefore, many researches have focused on utilizing virtual staining techniques to convert H&E images into IHC images [38, 41, 43]. The existing virtual IHC staining methods can be roughly divided into two categories: supervised learning-based methods and unsupervised learning-based methods. Supervised learning-based methods require a large amount of paired images for training. These H&E-IHC pairs typically come from adjacent layers of same tissues, accompanied by complex data processing and alignment. What's more, these methods cannot be used in scenarios where adjacent layers are not available. On the other hand, unsupervised methods often focus on the style of the generated images, ignoring the transmission of pathological information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "To address these aforementioned problems, we propose a virtual IHC staining framework called confusion-GAN, which utilizes weakly supervised learning to facilitate the unsupervised staining process of unpaired images. To explore the pathological information of H&E patches, a novel patch-level pathology information extractor which is assisted by multiple instance learning (MIL) is proposed. This method utilizes a dual-spherical loss proposed by us to push positive patches away from negative patches in the feature" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 243, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 243, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 243, + 714 + ], + "type": "text", + "content": "*Corresponding author: Xiaopeng Fan, Yongbing Zhang" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11259" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "space, aiming to achieve accurate acquisition of pathological information. When training the confusion-GAN, we constrain the pathological consistency between the staining results and the input images. In addition, we propose a multi-branch discriminator, which contains a confusion discriminator as well as a common single-image discriminator, aiming to enhance the fidelity of the staining results. Specifically, we first combine a generated image with multiple real images to create an image pool in each iteration. The confusion discriminator aims to distinguish the unique generated image from this pool. Simultaneously, the generator tries to hinder this judgment of the confusion discriminator." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 441 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed approach, we conducted extensive experiments on three datasets: 1) converting H&E staining images to HER2 staining images over the BCI dataset [24], 2) converting H&E staining images to ER staining images over the MIST dataset [18], and 3) convert H&E staining images to GPC3 staining images over our collected hepatocellular carcinoma immunohistochemical (HCI) dataset. In terms of image quality and pathological information transmission, our confusion-GAN can achieve superior performance compared to existing state-of-the-art methods and deceive human pathologists to some extent. Additionally, we compared the classification accuracy of ABMIL [13], a widely used algorithm for pathological image diagnosis, by using GPC3 images generated from our confusion-GAN and the original H&E images as the input data. The experimental results clearly indicate that the utilization of generated GPC3 images can significantly improve the diagnosis performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 443, + 244, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 443, + 244, + 454 + ], + "spans": [ + { + "bbox": [ + 58, + 443, + 244, + 454 + ], + "type": "text", + "content": "The contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 455, + 287, + 640 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 455, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 502 + ], + "type": "text", + "content": "- We propose an algorithm that utilizes weakly supervised learning to assist unsupervised virtual staining for the first time, which generates satisfactory images on three virtual IHC staining tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 505, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 287, + 564 + ], + "type": "text", + "content": "- We propose a novel patch-level pathology information extractor (PPIE) for H&E image classification. As a plug-and-play method, PPIE is demonstrated to significantly improve the instance-level classification performance when combined with existing MIL methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 567, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 287, + 602 + ], + "type": "text", + "content": "- We propose a multi-branch discriminator (MBD) that can significantly enhance the fidelity of the staining results compared to only using a single-image discriminator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 605, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 640 + ], + "type": "text", + "content": "- We collected a high-resolution hepatocellular carcinoma immunohistochemical dataset, which is contains both H&E and GPC3 staining images for the first time." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 224, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 224, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 224, + 684 + ], + "type": "text", + "content": "2.1. Virtual IHC Staining Techniques" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Existing virtual IHC staining methods can be broadly divided into two categories: supervised and unsupervised" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 311 + ], + "type": "text", + "content": "learning based methods. Supervised learning based methods require a large number of H&E-IHC pairs as training images [8, 18, 24]. There are two ways of obtaining these paired images: (1) Carrying out axially consecutive scanning of tissues, actually staining H&E and IHC on adjacent slices, then performing alignment between H&E and IHC slides through complex data processing and registration to obtain the paired images for training. For example, Liu et al. [24] train their proposed PyramidP2P network over a registered H&E-HER2 dataset. However, the tissue contents from different layers are inherently difficult to be consistent, making pixel-level alignment impossible. Therefore, the inability to fully align will inevitably result in performance limitation. In addition, such supervised methods are unlikely to be applicable when adjacent layer data is unavailable. (2) Using images of a third modality other than H&E and IHC, such as microscopic fluorescence data, as a bridge, to generate pixel-wise aligned H&E and IHC training data [8]. Unfortunately, access to the third modality data will also add additional overhead." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": "The vast majority of existing unsupervised virtual IHC staining methods are based on image translation networks [5, 9, 21, 30, 33, 35, 40]. Although these networks can convert H&E-style images into IHC-style images, it does not mean that the Pos/Neg properties of the translated results will be consistent with that of their inputs. To keep the aforementioned pathological consistency, Liu et al. [23] and Boyd et al. [3] introduce expert annotations, which is an expensive solution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 428, + 525, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 525, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 525, + 441 + ], + "type": "text", + "content": "2.2. Patch-level Classification Assisted by MIL" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": "Currently, MIL are used for patch-level classification [4, 13, 16, 19, 32]. In the case of pathological images, each WSI is regarded as a bag, and each patch within the WSI is treated as an instance. A bag label is assigned to each bag, which is classified as positive if the bag contains at least one positive instance. These bag labels serve as the ground truths for training MIL networks in an end-to-end manner. Existing methods determine whether a patch is positive or not by comparing the attention scores (or instance-level predictions) of all patches [6, 20, 22, 29, 31, 34]. Unfortunately, these methods can only ensure accurate patch-level classification within a limited region of high attention scores. Therefore, we propose a novel patch-level pathology information extractor to compensate for the shortcomings of these methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 639, + 411, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 411, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 411, + 651 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 658, + 384, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 384, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 384, + 670 + ], + "type": "text", + "content": "3.1. Framework" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "One significant difference between the virtual IHC staining and the image translation of natural images is that generated IHC images must maintain pathological consistency with" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11260" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 517, + 213 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 517, + 213 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 517, + 213 + ], + "type": "image", + "image_path": "7a5cb8d4624bffc597ea3e917395960d082fe807f267f124806fadccf2ddd86c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 218, + 299, + 355 + ], + "blocks": [ + { + "bbox": [ + 75, + 218, + 299, + 355 + ], + "lines": [ + { + "bbox": [ + 75, + 218, + 299, + 355 + ], + "spans": [ + { + "bbox": [ + 75, + 218, + 299, + 355 + ], + "type": "image", + "image_path": "23ee533a12ea7a3d7cdc944a5e3c9f7dcce8944daa6c79495a48887faae56f31.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 546, + 388 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 546, + 388 + ], + "type": "text", + "content": "Figure 1. Structure of the proposed confusion-GAN. (a) The confusion-GAN contains two important modules: patch-level pathology information extractor (PPIE) and multi-branch discriminator (MBD). (b) The training process of PPIE. (c) The structure of the MBD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 219, + 517, + 355 + ], + "blocks": [ + { + "bbox": [ + 301, + 219, + 517, + 355 + ], + "lines": [ + { + "bbox": [ + 301, + 219, + 517, + 355 + ], + "spans": [ + { + "bbox": [ + 301, + 219, + 517, + 355 + ], + "type": "image", + "image_path": "861113c4b9c2670c4ef619fd59ffa13db5345f439378e9eaea1f65d14073e5a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 517 + ], + "type": "text", + "content": "their inputs while preserving texture consistency. From the perspectives of improving the quality of generated images and accurately transmitting pathological information, we propose confusion-GAN, which is illustrated in Fig. 1. The confusion-GAN contains two important modules, patch-level pathology information extraction module and multibranch discriminator, which are employed to maintain the pathological consistency and generate more realistic images, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": "As shown in Fig. 1, the real H&E image " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": " is first passed through the generator " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "G_{A}" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": " to obtain the IHC-style image " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": " is further remapped into the H&E-style image " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": " using the generator " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "G_{B}" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": ". Similarly, we can obtain " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": " from real IHC image " + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "During the process of generating IHC image from the H&E image, we devise a Pathology Loss to ensure the accurate transmission of pathological information. IHC classifier " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "Cls_{ihc}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and a PPIE module are designed to obtain the pathological information of the H&E image " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and the generated IHC image " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ", respectively. To pretrain a reliable IHC classifier, we simulate the annotation process of pathologists through color analysis to obtain patch-level pos/neg labels for actually captured IHC images: patches with positive exhibit protein-specific expression within more than 1% of the area (corresponding to yellow-brown coloration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "text", + "content": "in the image). A light-weight image classifier " + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "inline_equation", + "content": "Cls_{ihc}" + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "text", + "content": " is then designed utilizing the pos/neg labels obtained by the previously described process, which consists of 6 convolutional layers activated by ReLU and a fully connected layer (FCN) for the final classification. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1. Similarly, the pathological information of the H&E image is obtained utilizing the PPIE module, which will be elaborated in Sec. 3.2. Finally, we adopt the Pathology Loss to constrain the consistency of pathological information between " + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i}" + }, + { + "bbox": [ + 304, + 408, + 546, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 332, + 544, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 544, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 332, + 544, + 545, + 591 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {p a t h}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(P P I E \\left(x _ {i}\\right) \\log \\left(C l s ^ {i h c} \\left(\\hat {y} _ {i}\\right)\\right) \\right. \\tag {1} \\\\ + (1 - P P I E \\left(x _ {i}\\right)) \\log \\left(1 - C l s ^ {i h c} \\left(\\hat {y} _ {i}\\right)\\right), \\\\ \\end{array}", + "image_path": "3da735fa270b50f0414e9d56c2fc54cc009eb060a8e7208cfefc9fd0748c7ec4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 596, + 541, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 596, + 541, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 596, + 541, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 596, + 541, + 608 + ], + "type": "inline_equation", + "content": "PPI(\\cdot)" + }, + { + "bbox": [ + 305, + 596, + 541, + 608 + ], + "type": "text", + "content": " represents the operation of PPI module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 609, + 545, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 667 + ], + "type": "text", + "content": "In addition, to enhance the quality of the generated images, we propose a multi-branch discriminator for both H&E-to-IHC and IHC-to-H&E process in the confusionGAN. Detailed description of the multi-branch discriminator can be found in Sec. 3.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 668, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 668, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 317, + 668, + 545, + 681 + ], + "type": "text", + "content": "In general, the hybrid loss for the generator is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 684, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 684, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 347, + 684, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {G} = \\alpha * L _ {\\text {p a t h}} + \\beta * L _ {G} ^ {C} + \\lambda * L _ {G} ^ {A d v} \\\\ + \\eta * L _ {G} ^ {C y c l e} + \\iota * L _ {G} ^ {\\text {I d e n t i t y}}, \\\\ \\end{array}", + "image_path": "844ff2299b765673b0db5970fff9faaa56f65f98df5fb19abb3d5fba7868092a.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11261" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "L_{G}^{Adv}" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "L_{G}^{Cycle}" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "L_{G}^{Identity}" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": " represent the adversarial loss, cycled loss, and identity loss similar to CycleGAN [44], while " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": " are hyper-parameters controlling the weighting coefficients. " + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "inline_equation", + "content": "L_{G}^{C}" + }, + { + "bbox": [ + 47, + 70, + 289, + 144 + ], + "type": "text", + "content": " is the loss function used to guide the generator optimization with the confusion discriminator, which will be described in detail in Sec. 3.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 153, + 284, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 284, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 284, + 166 + ], + "type": "text", + "content": "3.2. Patch-level Pathology Information Extraction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 172, + 287, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 267 + ], + "type": "text", + "content": "Due to memory limitations, virtual staining can only be performed at the patch level. Unfortunately, only WSI-level labels are available for majority cases, since one WSI usually contains gigapixel, impeding the patch-wise annotation in clinical. Therefore, it is necessary to extract patch-level pathological information (i.e., the pos/neg category of each patch) from the H&E WSI to ensure accurate virtual staining." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "text", + "content": "For H&E images that only have WSI level labels (lacking patch-level annotations), the classification of each patch is typically completed through attention scores or instance-level classification predictions within the MIL framework. In each positive WSI, there is a high likelihood that patches with the highest attention scores or instance-level classification scores will be positive. We refer to the action of selecting the top " + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "text", + "content": " patches of each positive WSI based on the highest attention scores (or instance-level prediction scores) as \"top-k selection\". When " + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "text", + "content": " is relatively small, it is ensured that the selected patches are all positive; when " + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 269, + 289, + 435 + ], + "type": "text", + "content": " is large, the selected collections may include many negative patches [26]. Therefore, the current methods have not been able to accurately classify all patches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "content": "To address this problem, building upon the top-k selection, we propose a plug-and-play PPIE module that can improve the classification performance of H&E patches. Specifically, we first perform top-k selection to obtain " + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "inline_equation", + "content": "X_{pos}^{he}" + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "content": " from positive WSIs in the training set. Next, we crop a set of negative patches " + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "inline_equation", + "content": "X_{neg}^{he}" + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "content": " from negative WSIs. It should be noted all patches in a negative WSI are negative, unlike the cases of positive WSIs where both positive and negative patches may coexist. Besides, we utilized a state-of-the-art diffusion model BBDM [17] to learn how to generate a distribution similar to that of " + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "inline_equation", + "content": "X_{pos}^{he}" + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "inline_equation", + "content": "X_{neg}^{he}" + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "content": ". The set of images generated by BBDM is referred to as " + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "inline_equation", + "content": "X_{pos}^{he}" + }, + { + "bbox": [ + 46, + 436, + 289, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": "On this basis, we borrow the idea of anomaly detection to tackle the classification problem of patch-level H&E images [10]. In each iteration of training the PPIE, we sample " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": " images from " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "X_{\\text{neg}}^{\\text{he}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "X_{\\text{pos}}^{\\text{he}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "X_{\\text{pos}}^{\\text{he}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": " respectively, and feed them into the encoder to obtain the feature sets " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "C_1 = \\{c_{1i}\\}_{i=1}^b" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "C_2 = \\{c_{2i}\\}_{i=1}^b" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "C_3 = \\{c_{3i}\\}_{i=1}^b" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ", as shown in Fig. 1(b). Here, " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": " represents the batch size during training. The encoder consists of 6 convolutional layers activated by ReLU and a fully connected layer with an output dimension of 512. The kernel size for the convolutional layers is 4, with a stride of 2 and padding of 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "As is well known, the success of the anomaly detection depends on two aspects: 1) the samples in the in-distribution (ID) class should be as close as possible in the feature space; 2) the samples in the out-of-distribution (OOD) class should be as far as possible from samples in the ID class [39]. We use the loss function " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "L_{ID}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " to force all the samples in the ID class to be distributed as close to the center as possible:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 370, + 164, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 164, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 370, + 164, + 545, + 198 + ], + "type": "interline_equation", + "content": "L _ {I D} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\left\\| c _ {1 i} - O \\right\\| ^ {2}, \\tag {3}", + "image_path": "8ae8d66888a13a7bc37b65524d493576a843d1a303e230e56a9e141897bd54c0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": " represents the average features of all the ID data " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "X_{neg}^{he}" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": ". It should be noted that " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": " will be cumulatively updated during training. Besides, we propose a double-spherical loss " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "L_{OOD}" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": " to ensure that the overall distribution of " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": " has a greater distance from " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": " compared to the overall distribution of " + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 304, + 204, + 547, + 276 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 336, + 284, + 545, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 284, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 336, + 284, + 545, + 352 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {O O D} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\max \\left(R _ {1} - \\left\\| c _ {2 i} - O \\right\\| ^ {2}, 0\\right) \\tag {4} \\\\ + \\frac {1}{b} \\sum_ {i = 1} ^ {b} m a x (R _ {2} - | | c _ {3 i} - O | | ^ {2}, 0). \\\\ \\end{array}", + "image_path": "16afcf68f6f4ab9e99bccc693a22268716f4a03dd6dbff14541ec74a2c7b7b35.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": "We set the hyperparameter " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " to control the distance and ensure that " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": ", as shown in Fig. 1(b). Employing " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "L_{ID}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "L_{OOD}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": ", we can ensure that " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "X_{neg}^{he}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "X_{pos}^{he}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " have strong separability in the feature space. Finally, we treat " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "X_{neg}^{he}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "X_{pos}^{he}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " as ID and OOD classes respectively, and train the patch-level classifier " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "Cls^{he}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": " by cross entropy loss " + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 304, + 358, + 547, + 449 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 456, + 545, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 456, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 347, + 456, + 545, + 504 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {c l s} = - \\frac {1}{n} \\sum_ {i = 1} ^ {s} \\left(l _ {i} \\log \\left(C l s ^ {h e} \\left(x _ {i} ^ {h e}\\right)\\right) \\right. \\tag {5} \\\\ + \\left(1 - l _ {i}\\right) \\log \\left(1 - C l s ^ {h e} \\left(x _ {i} ^ {h e}\\right)\\right)). \\\\ \\end{array}", + "image_path": "081699ab07aeea667bcc99a538a1075b5b350051edb9dfad860b4f1270c58325.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "content": "When using PPIE for inference, negative patches are considered as ID class, while positive patches are considered as OOD class." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 547, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 571 + ], + "type": "text", + "content": "Overall, the hybrid loss function " + }, + { + "bbox": [ + 304, + 547, + 545, + 571 + ], + "type": "inline_equation", + "content": "L_{PPIE}" + }, + { + "bbox": [ + 304, + 547, + 545, + 571 + ], + "type": "text", + "content": " for training the PPIE module is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 581, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 581, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 335, + 581, + 545, + 594 + ], + "type": "interline_equation", + "content": "L _ {P P I E} = \\mu * L _ {I D} + \\tau * L _ {O O D} + v * L _ {c l s}, \\tag {6}", + "image_path": "a9db7bf4d0b870c5c3aea30322f5475139ea05d44c28fd328bee99948bbfb1a4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mu, \\tau," + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": " are hyper-parameters to control the weighting coefficients." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 635, + 462, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 462, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 462, + 647 + ], + "type": "text", + "content": "3.3. Multi-branch Discriminator" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "To achieve more realistic image generation, we propose a multi-branch discriminator, as shown in Fig. 1(c). Taking generating an IHC-style image from a H&E image as an example, we send the result " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " obtained by the generator " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "G_{A}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " together with a reference pool " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "P_i^{ihc}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " consisting of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11262" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " real images from the same domain to the confusion discriminator " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " shares the feature extraction layer " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "D_B^S" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " with the single-image discriminator " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "D_B" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": ", which helps eliminate the interference of the content of the reference pool images on the confusion discriminator. This one-to-many discrimination paradigm requires the generator's results to possess more common attributes of the current domain images, thereby forcing the generated distribution to approximate the real distribution more closely." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": "During the training of the " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": ", we aim to accurately distinguish between the generated result and real images. Then we randomly insert " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " into a position in the list of " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "P_i^{ihc}" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": ", and finally concatenate them in the channel dimension to obtain the input " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "m_i^{ihc}" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": ". Correspondingly, we construct an " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": "-dimensional vector " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "l_i^{ihc}" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " to represent the labels of each element in " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "m_i^{ihc}" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": ", with 0 indicating real image and 1 indicating generated image. Next, we pass " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "m_i^{ihc}" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "D_B^S" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " to obtain a feature pool, which is fed into an MLP layer of " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " for liveness verification. " + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 46, + 180, + 288, + 300 + ], + "type": "text", + "content": " is then updated by minimizing:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 92, + 308, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 308, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 92, + 308, + 287, + 338 + ], + "type": "interline_equation", + "content": "L _ {C _ {B}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {B} \\left(m _ {i} ^ {i h c}\\right) - l _ {i} ^ {i h c} \\right| \\right| ^ {2}. \\tag {7}", + "image_path": "7061d8a70cadbc2aec1a62636e463492ecf96e99306252ddbcb6a831e405208d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "text", + "content": "Similarly, during the process of IHC-to-H&E image translation, the loss function " + }, + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "inline_equation", + "content": "L_{C_A}" + }, + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "text", + "content": " of the confusion discriminator " + }, + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "inline_equation", + "content": "C_A" + }, + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 389, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 389, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 95, + 389, + 287, + 420 + ], + "type": "interline_equation", + "content": "L _ {C _ {A}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {A} \\left(m _ {i} ^ {h e}\\right) - l _ {i} ^ {h e} \\right| \\right| ^ {2}, \\tag {8}", + "image_path": "b3aed0821c1cfb029822c732e0e94631d1f4be32ab66f121f0c881b510b1491e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "inline_equation", + "content": "m_{i}^{he}" + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "text", + "content": " represents the concatenation result of the generated H&E image and " + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "text", + "content": " real H&E images, and " + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "inline_equation", + "content": "l_{i}^{he}" + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "text", + "content": " represents the corresponding labels of " + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "inline_equation", + "content": "m_{i}^{he}" + }, + { + "bbox": [ + 47, + 429, + 287, + 477 + ], + "type": "text", + "content": ". The proposed confusion discriminators are trained using the following formula:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 486, + 287, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 486, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 125, + 486, + 287, + 499 + ], + "type": "interline_equation", + "content": "L _ {C} = L _ {C _ {B}} + L _ {C _ {A}}. \\tag {9}", + "image_path": "f35a9207f07fad8b5447781938d17430b96c8173893f3499daaf2e34115b904a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "type": "text", + "content": "The loss function used for training " + }, + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "type": "inline_equation", + "content": "D_B" + }, + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "type": "inline_equation", + "content": "D_A" + }, + { + "bbox": [ + 47, + 506, + 287, + 529 + ], + "type": "text", + "content": " is consistent with CycleGAN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "content": "When training the generator, we obtain " + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "m_i^{ihc}" + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "m_i^{he}" + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "content": " in the same manner. However, in this stage, we want the confusion discriminator unable to identify the generator's results from these " + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "content": " images. Therefore, all labels in " + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "l_i^{ihc}" + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "l_i^{he}" + }, + { + "bbox": [ + 46, + 530, + 287, + 602 + ], + "type": "text", + "content": " are set to 0. The generators are updated based on the following loss functions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 611, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 611, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 97, + 611, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {G} ^ {C} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| C _ {B} \\left(m _ {i} ^ {i h c}\\right) - l _ {i} ^ {i h c} \\right| \\right| ^ {2} \\tag {10} \\\\ + \\left\\| C _ {A} \\left(m _ {i} ^ {h e}\\right) - l _ {i} ^ {h e} \\right\\| ^ {2}. \\\\ \\end{array}", + "image_path": "dd3cd2631fef097ca90cd37259178b9b5484606dfec0ff9836d7f5a99312af20.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "In each iteration, we train the generators " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "G_{A}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "G_{B}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{A}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{B}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": ", and our proposed " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_{A}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_{B}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " in an alternating manner, following a sequential order. This process is similar to what most generative adversarial networks do. Here" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "G_{A}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "G_{B}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " employ the unet256 network used in CycleGAN and the structures of " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "D_{A}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "D_{B}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " are also consistent with those in CycleGAN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 120, + 386, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 386, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 386, + 133 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 140, + 356, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 356, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 356, + 153 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 158, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 350 + ], + "type": "text", + "content": "Datasets. GPC3, discovered as a tumor marker in recent years, shows great potential for specific expression in Hepatocellular carcinoma (HCC). We have collected a HCC Immunohistochemical Image dataset named HCI, which includes both the H&E and GPC3 images. All data comes from Peking University Shenzhen Hospital. We selected 20 H&E-IHC image pairs from a pool of WSIs for testing and validation, and perform registration on them using the DeepHistReg [37]. Next, we conducted separate screenings for a total of 30 H&E images and 30 IHC images, and these images are used for training. Furthermore, we segment these 100 WSIs into approximately 1.4 million patches sized of " + }, + { + "bbox": [ + 304, + 158, + 545, + 350 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 158, + 545, + 350 + ], + "type": "text", + "content": " in a non-overlapping manner at 20x magnification. All H&E images have undergone staining normalization. We also verified our method on public datasets: BCI [24] and MIST [18]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": "Metrics. We evaluated the image quality of the staining results using objective metrics such as Frechet Inception Distance (FID) [27], Learnable Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index (SSIM) [11]. Additionally, we invited three pathologists to provide subjective evaluations of the staining results. In assessing the patch-level classification performance of the aforementioned PPIE, we utilized metrics such as area under curve (AUC) [25], False Positive Rate (FPR) [28], True Positive Rate (TPR), and Geometric Mean (G-mean) [7]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": "Implementation Details. We trained the PPIE using the Adam optimizer with a learning rate of 0.001. The dual-sphere distances " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": " in Eq. (4) are set to 150 and 262.5 respectively. The hyper-parameters " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\upsilon" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": " in Eq. (6) are all set to 1. When training the virtual staining network, " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": " in Eq. (2) are set to 1, 1, 1, 10, and 5 respectively. We set the reference pool size " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": " as 32 and kept the other configuration of the staining network consistent with CycleGAN. In the downstream experiments, we trained the ABMIL model on a re-partitioned dataset for a maximum of 15 epochs. For all experiments, we trained and tested our confusion-GAN on patches of size " + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 471, + 545, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 623, + 430, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 430, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 430, + 635 + ], + "type": "text", + "content": "4.2. Staining Performance" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "HCI. As shown in Fig. 2, we compared our method with three unsupervised image translation methods, namely CycleGAN [44], UGATIT [15], and DCD [12], on the HCI dataset. The first column in Fig. 2 displays the input H&E images, while the last column shows actually stained IHC (GPC3) images for reference. These images are from corre" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11263" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 72, + 492, + 246 + ], + "blocks": [ + { + "bbox": [ + 102, + 72, + 492, + 246 + ], + "lines": [ + { + "bbox": [ + 102, + 72, + 492, + 246 + ], + "spans": [ + { + "bbox": [ + 102, + 72, + 492, + 246 + ], + "type": "image", + "image_path": "13b54aa30a1deb61bdbb5a9935b55b4971c4f8c473ad5d90e3620c0c26a651a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. Comparison of the abilities of different unsupervised algorithms to preserve pathological consistency on HCI. The red dashed box indicates the area where the generated image cannot maintain consistent pathological information with the input H&E image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 102, + 289, + 493, + 462 + ], + "blocks": [ + { + "bbox": [ + 102, + 289, + 493, + 462 + ], + "lines": [ + { + "bbox": [ + 102, + 289, + 493, + 462 + ], + "spans": [ + { + "bbox": [ + 102, + 289, + 493, + 462 + ], + "type": "image", + "image_path": "59b48f485350d25439c25c8b3a0e5efb8769c30cc4d750fb61824d00fd1207a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 469, + 547, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 469, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 547, + 504 + ], + "type": "text", + "content": "Figure 3. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on BCI. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 524, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 289, + 693 + ], + "type": "text", + "content": "sponding regions of adjacent slices in the same tissue. We marked in red dashed lines the regions where the pathological information is inconsistent with the input. Obviously, although the competing methods CycleGAN, UGATIT, and DCD can maintain the texture details of the input images well, there are significant mistaining areas since the consistency of pathological information can not be ensured before and after staining. This phenomenon demonstrates the effectiveness of the proposed PPIE module. Besides, we use three objective metrics, FID, LPIPS, and SSIM to evaluate these virtual staining methods. As shown in Tab. 1, our method achieves the best performance in all the three metrics, indicating that the output distribution of the confusion-GAN is closest to the distribution of the real images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "BCI and MIST. BCI and MIST datasets provide paired" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 339, + 521, + 517, + 586 + ], + "blocks": [ + { + "bbox": [ + 339, + 521, + 517, + 586 + ], + "lines": [ + { + "bbox": [ + 339, + 521, + 517, + 586 + ], + "spans": [ + { + "bbox": [ + 339, + 521, + 517, + 586 + ], + "type": "table", + "html": "
MethodFID↓LPIPS↓SSIM↑
CycleGAN(unet)105.40.6360.1870
CycleGAN(resnet)90.90.6220.2086
UGATIT95.30.6020.2223
DCD91.30.6120.2266
Ours77.30.5700.2300
", + "image_path": "87891f4796c23d207a45feba9f1f5fb5527fa7bd8ab27f0d781ab5a61206f9a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 594, + 526, + 605 + ], + "lines": [ + { + "bbox": [ + 324, + 594, + 526, + 605 + ], + "spans": [ + { + "bbox": [ + 324, + 594, + 526, + 605 + ], + "type": "text", + "content": "Table 1. The evaluation of different algorithms on HCI." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 715 + ], + "type": "text", + "content": "H&E-IHC patch images, where the acquisition of these patches relies on alignment with the sliced images from adjacent layers within the same tissue. However, the tissue contents from different adjacent layers are inherently impossible to be completely consistent, making pixel-level alignment impossible. We compared our method with four fully supervised methods (Pix2pix [14], Pix2pixHD [36]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11264" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 72, + 496, + 248 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 496, + 248 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 496, + 248 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 496, + 248 + ], + "type": "image", + "image_path": "6bca552fd26dc44d1aa5c1686e70b84f3953e065b421bb93c3466e5a2a0b4c5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 547, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 547, + 292 + ], + "type": "text", + "content": "Figure 4. Comparing the ability of the state-of-the-art supervised methods and ours to simultaneously preserve the texture and pathological information of the input image on MIST. The red solid rectangles indicate regions with texture errors, while the red dashed rectangles represent regions with pathological information errors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 539 + ], + "type": "text", + "content": "PyramidP2P [24], ASP [18]) on these two datasets, as shown in Fig. 3 and Fig. 4. We also marked the areas with pathological information errors in the staining results with dashed lines. Additionally, we marked the areas with content errors in the staining results with solid lines. From Fig. 3 and Fig. 4, it can be observed that Pix2pix, Pix2pixHD, and PyramidP2P methods are unable to maintain the textures of the input images, often resulting in significant presence of artifacts in the generated results. This phenomenon is attributed to the significant discrepancy in content between the input H&E images and the reference IHC images used as the ground truth during the training process. The ASP based on contrastive learning demonstrates good performance in preserving content. However, the misalignment of supervised information can still introduce some errors in the pathological information of the staining results. In contrast, our proposed confusion-GAN has achieved better results than these supervised learning methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 546, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 715 + ], + "type": "text", + "content": "We used FID, LPIPS, and SSIM for objective evaluation. As can be seen from Tab. 2, our confusion-GAN achieved the best FID and LPIPS scores on both datasets, which indicating the effectiveness of our algorithm. Regarding the SSIM evaluation metric, which is directly calculated in the pixel space, our method still achieved the highest score on MIST dataset. Meanwhile, our method can achieve the best visual effects on both BCI and MIST, especially considering the fidelity of textures in the generated results. It is worth mentioning that, due to the lack of complete WSIs in MIST and BCI, we assigned labels to IHC images using the color analysis mentioned in Sec. 3.1, and then transfer the labels of IHC images to the corresponding H&E images. Regarding to BCI, we directly utilized the provided patch-level la" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 312, + 512, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 512, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 512, + 324 + ], + "type": "text", + "content": "bels, which guides the training of confusion-GAN." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 313, + 332, + 542, + 401 + ], + "blocks": [ + { + "bbox": [ + 313, + 332, + 542, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 332, + 542, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 542, + 401 + ], + "type": "table", + "html": "
MethodsBCIMIST
FID↓LPIPS↓SSIM↑FID↓LPIPS↓SSIM↑
Pix2pix110.70.4800.495780.90.5290.2127
Pix2pixHD106.80.5050.4766131.00.5640.2129
PyramidP2P99.60.5050.466397.50.5380.2061
ASP54.30.5030.492341.20.5320.2062
Ours49.20.4780.457939.80.5160.2174
", + "image_path": "c28a841ccbc0e83b4009037e34c964de420350311c33653dfeb8acf59ebfb777.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 450, + 400, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 400, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 400, + 463 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 468, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 493 + ], + "type": "text", + "content": "We explored the working mechanism of PPIE and analyzed how PPIE and the MBD affect the staining results." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 337, + 501, + 518, + 658 + ], + "blocks": [ + { + "bbox": [ + 305, + 408, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 431 + ], + "type": "text", + "content": "Table 2. Comparing the performance of our method and supervised algorithms on BCI and MIST." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 501, + 518, + 658 + ], + "lines": [ + { + "bbox": [ + 337, + 501, + 518, + 658 + ], + "spans": [ + { + "bbox": [ + 337, + 501, + 518, + 658 + ], + "type": "table", + "html": "
MethodkAUC↑FPR↓TPR↑G-mean↑
ABMIL100.7130.2770.6410.680
300.7820.2260.7260.750
1000.6370.4080.5890.591
wo0.7320.3230.6860.681
Maxpool100.6820.3240.6660.671
300.7310.2570.6890.715
1000.7060.2670.6670.699
wo0.6800.3280.6130.680
ClamMb100.7660.2380.6910.726
300.7700.2190.7080.744
1000.7320.2860.6270.669
wo0.7260.3380.6870.675
Meanpool100.7450.2530.6710.708
300.7680.2390.7280.744
1000.7090.2540.6190.679
wo0.7580.3240.7100.693
", + "image_path": "552e471f0896a369562074772bd955e5477f31bfb47857ea62ce4d9ec0b9ac54.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 667, + 545, + 689 + ], + "lines": [ + { + "bbox": [ + 305, + 667, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 667, + 545, + 689 + ], + "type": "text", + "content": "Table 3. Patch-level classification results obtained from different top-k patches on HCI." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " in the top-k selection. We selected four clas" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11265" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "sic MIL methods (ABMIL [13], MaxPool, CLAM-MB[26], and MeanPool) for top-k selection in PPIE training. As shown in Tab. 3, we evaluate the classification performance of PPIE using AUC, FPR, TPR, and G-mean. We found that the classification performance is best when " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " is set to 30. This is because when " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " is too small, the PPIE cannot obtain sufficient training data. On the other hand, when " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " is too large, the top-k selection results will contain a large number of negative patches, which hinders the effectiveness of the double-spherical loss. Among these four MIL methods, we obtained similar conclusions and improved the patch-level classification performance, achieving gains of " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 290, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 360 + ], + "type": "text", + "content": "Effect of PPIE and MBD. To investigate how PPIE and MBD impact the performance, we conducted an ablation experiment on HCI, as shown in Fig. 5. By comparing the second, third, and fourth columns in Fig. 5, it can be observed that the PPIE plays a crucial role in accurately transmitting pathological information. On the other hand, the MBD can enhance the clarity of cell nuclei and surrounding tissues, affecting the overall style of the generated results. What's more, based on the qualitative analysis of the indicators in Tab. 4, we found that the method combining PPIE and the MBD achieved the best results." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 81, + 367, + 254, + 423 + ], + "blocks": [ + { + "bbox": [ + 81, + 367, + 254, + 423 + ], + "lines": [ + { + "bbox": [ + 81, + 367, + 254, + 423 + ], + "spans": [ + { + "bbox": [ + 81, + 367, + 254, + 423 + ], + "type": "table", + "html": "
PPIEXX✓✓
MBDX✓X✓
FID↓105.482.4103.377.3
LPIPS↓0.6360.5880.5740.570
SSIM↑0.18700.22230.22940.2300
", + "image_path": "aa0447c4b606af7719a02527cdd47b44920d9a6b9051db67914c1897e22e6bfa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 262, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 262, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 262, + 477 + ], + "type": "text", + "content": "4.4. Down-stream Classification Performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "text", + "content": "We conducted downstream experiments related to the HCC diagnosis, and compared the classification performance achieved by using the H&E and generated GPC3 slides as inputs of the ABMIL model, respectively. When using H&E images as input achieves an AUC of " + }, + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "inline_equation", + "content": "55.6\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "text", + "content": ", while using the generated images as input achieves an AUC of " + }, + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "text", + "content": ". This indicates that compared to H&E images, our generated GPC3 images can improve the classification accuracy of HCC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 246, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 246, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 246, + 612 + ], + "type": "text", + "content": "4.5. Subjective Evaluation by Pathologists" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "We invited three pathologists to provide subjective evaluations on two aspects: transmission of pathological information and the quality of the images. We randomly selected 50 pairs of generated GPC3 images along with their corresponding adjacent tissues and then asked pathologists to evaluate whether the protein expression levels were consistent between the two. As shown in the first row of Tab. 5, in most cases, pathologists believed that the protein expression" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 540, + 194 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 540, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 540, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 540, + 194 + ], + "type": "image", + "image_path": "03a34eb11974582e2152893ccdf8f5a60ca643cb2de74790604f262e3df2c700.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "content": "Figure 5. Ablation study: effect of PPIE and MBD on HCI. The red box indicates regions with pathological information errors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 243, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 340 + ], + "type": "text", + "content": "levels were consistent between them. This indicates that our confusion-GAN can effectively maintain pathological consistency. Afterwards, we randomly selected 50 GPC3 images and 50 generated GPC3 images for pathologists to distinguish which are generated images. From the second row of the Tab. 5, it can be seen that doctors cannot distinguish the generated images well, indicating that our confusion-GAN can generate images comparable to real images." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 325, + 348, + 529, + 384 + ], + "blocks": [ + { + "bbox": [ + 62, + 430, + 271, + 441 + ], + "lines": [ + { + "bbox": [ + 62, + 430, + 271, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 271, + 441 + ], + "type": "text", + "content": "Table 4. Exploring the impact of PPIE and MBD on HCI." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 348, + 529, + 384 + ], + "lines": [ + { + "bbox": [ + 325, + 348, + 529, + 384 + ], + "spans": [ + { + "bbox": [ + 325, + 348, + 529, + 384 + ], + "type": "table", + "html": "
Accuracypathologist1pathologist2pathologist3
Neg/Pos0.860.800.84
Fake/True0.620.480.52
", + "image_path": "c2c3d0383bdcd67c88f23c385d1efc7b8d835900cf4958b101ab92b9e7378375.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 391, + 526, + 402 + ], + "lines": [ + { + "bbox": [ + 325, + 391, + 526, + 402 + ], + "spans": [ + { + "bbox": [ + 325, + 391, + 526, + 402 + ], + "type": "text", + "content": "Table 5. Subjective evaluations of pathologists on HCI." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 445, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 635 + ], + "type": "text", + "content": "In this paper, we propose a method called confusion-GAN, which is the first to utilize weakly-supervised learning to assist in virtual IHC staining of unpaired pathological images. Extensive experiments have demonstrated that utilizing patch-level pathological information to guide the training of staining network is crucial for achieving accurate virtual IHC staining. Furthermore, we propose a multi-branch discriminator for achieving high-fidelity image generation. Our approach achieved the state-of-the-art results on multiple datasets. To advance the diagnosis of HCC, we have also proposed the first GPC3 dataset and demonstrated that the images generated by confusion-GAN are more helpful in HCC detection compared to H&E images. We believe that unsupervised methods like confusion-GAN, which do not require fine annotation or complex preprocessing, will be crucial in virtual staining research in the future." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 636, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 545, + 703 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62031023&62331011), in part by the Shenzhen Science and Technology Project (GXWD20220818170353009), and in part by the Fundamental Research Funds for the Central Universities (Grant No. HIT.OCEF.2023050)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11266" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Bijie Bai, Xilin Yang, Yuzhu Li, Yijie Zhang, Nir Pillar, and Aydogan Ozcan. Deep learning-enabled virtual histological staining of biological samples. Light: Science & Applications, 12(1):57, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Marian Boktor, Benjamin R Ecclestone, Vlad Pekar, Deepak Dinakaran, John R Mackey, Paul Fieguth, and Parsin Haji Reza. Virtual histological staining of label-free total absorption photoacoustic remote sensing (ta-pars). Scientific Reports, 12(1):10296, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "text", + "content": "[3] Joseph Boyd, Irène Villa, Marie-Christine Mathieu, Eric Deutsch, Nikos Paragios, Maria Vakalopoulou, and Stergios Christodoulidis. Region-guided cyclegans for stain transfer in whole slide images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 356-365. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "text", + "content": "[4] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "type": "text", + "content": "[5] Tanujit Chakraborty, Shraddha M Naik, Madhurima Panja, Bayapureddy Manvitha, et al. Ten years of generative adversarial nets (gans): A survey of the state-of-the-art. arXiv preprint arXiv:2308.16316, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 434 + ], + "type": "text", + "content": "[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 467 + ], + "type": "text", + "content": "[7] Yanyan Chen, Kuaini Wang, and Ping Zhong. One-class support tensor machine. Knowledge-Based Systems, 96:14-28, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 533 + ], + "type": "text", + "content": "[8] Kevin de Haan, Yijie Zhang, Jonathan E Zuckerman, Tairan Liu, Anthony E Sisk, Miguel FP Diaz, Kuang-Yu Jen, Alexander Nobori, Sofia Liou, Sarah Zhang, et al. Deep learning-based transformation of h&e stained tissues into special stains. Nature communications, 12(1):4884, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 535, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 288, + 612 + ], + "type": "text", + "content": "[9] Michael Gadermayr, Maximilian Tschuchnig, Lea Maria Stangassinger, Christina Kreutzer, Sebastien Couillard-Despres, Gertie Janneke Oostingh, and Anton Hittmair. Improving automated thyroid cancer classification of frozen sections by the aid of virtual image translation and stain normalization. Computer Methods and Programs in Biomedicine Update, 3:100092, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[10] Xiaoyuan Guo, Judy W Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Margin-aware intraclass novelty identification for medical images. Journal of Medical Imaging, 9(1): 014004-014004, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[12] Tie Hu, Mingbao Lin, Lizhou You, Fei Chao, and Rongrong Ji. Discriminator-cooperated feature map distillation for gan" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20351-20360, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[15] Junho Kim, Minjae Kim, Hyeonwoo Kang, and Kwanghee Lee. U-gat-it: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[16] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "type": "text", + "content": "[17] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 366, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 410 + ], + "type": "text", + "content": "[18] Fangda Li, Zhiqiang Hu, Wen Chen, and Avinash Kak. Adaptive supervised patchnce loss for learning h&e-to-ihc stain translation with inconsistent groundtruth image pairs. arXiv preprint arXiv:2303.06193, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 488 + ], + "type": "text", + "content": "[19] Weijian Li, Viet-Duy Nguyen, Haofu Liao, Matt Wilder, Ke Cheng, and Jiebo Luo. Patch transformer for multi-tagging whole slide histopathology images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 532-540. Springer, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[20] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 545, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 612 + ], + "type": "text", + "content": "[21] Kechun Liu, Beibin Li, Wenjun Wu, Caitlin May, Oliver Chang, Stevan Knezevich, Lisa Reisch, Joann Elmore, and Linda Shapiro. Vsgd-net: Virtual staining guided melanocyte detection on histopathological images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1918-1927, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "type": "text", + "content": "[22] Kangning Liu, Weicheng Zhu, Yiqiu Shen, Sheng Liu, Narges Razavian, Krzysztof J Geras, and Carlos Fernandez-Granda. Multiple instance learning via iterative self-paced supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355–3365, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[23] Shuting Liu, Baochang Zhang, Yiqing Liu, Anjia Han, Huijuan Shi, Tian Guan, and Yonghong He. Unpaired stain transfer using pathology-consistent constrained generative" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11267" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "adversarial networks. IEEE Transactions on Medical Imaging, 40(8):1977-1989, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[24] Shengjie Liu, Chuang Zhu, Feng Xu, Xinyu Jia, Zhongyue Shi, and Mulan Jin. Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1815-1824, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[25] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 286, + 251 + ], + "type": "text", + "content": "[26] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "type": "text", + "content": "[27] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "type": "text", + "content": "[28] Alvin I Mushlin, Ruth W Kouides, and David E Shapiro. Estimating the accuracy of screening mammography: a meta-analysis. American journal of preventive medicine, 14(2): 143-153, 1998. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 286, + 388 + ], + "type": "text", + "content": "[29] Linhao Qu, Manning Wang, Zhijian Song, et al. Bidirectional weakly supervised knowledge distillation for whole slide image classification. Advances in Neural Information Processing Systems, 35:15368-15381, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 389, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 286, + 443 + ], + "type": "text", + "content": "[30] Jesus Salido, Noelia Valdez, Lucia González-López, Oscar Deniz, and Gloria Bueno. Comparison of deep learning models for digital h&E staining from unpaired label-free multispectral microscopy images. Computer Methods and Programs in Biomedicine, 235:107528, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 286, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 286, + 499 + ], + "type": "text", + "content": "[31] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 286, + 567 + ], + "type": "text", + "content": "[32] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 286, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 623 + ], + "type": "text", + "content": "[33] Lulin Shi, Ivy HM Wong, Claudia TK Lo, Lauren WK Tsui, and Terence TW Wong. Unsupervised multiple virtual histological staining from label-free autofluorescence images. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "text", + "content": "[34] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[35] Elena Y Trizna, Aleksandr M Sinitca, Asya I Lyanova, Diana R Baidamshina, Pavel V Zelenikhin, Dmitrii I Kaplun," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 585 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Airat R Kayumov, and Mikhail I Bogachev. Brightfield vs fluorescent staining dataset-a test bed image set for machine learning based virtual staining. Scientific Data, 10(1):160, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[36] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[37] Marek Wodzinski and Henning Müller. Deephistreg: Unsupervised deep learning registration framework for differently stained histology samples. Computer methods and programs in biomedicine, 198:105799, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "text", + "content": "[38] Georg Wölfein, In Hwa Um, David J Harrison, and Ognjen Arandjelović. Hoechstgan: Virtual lymphocyte staining using generative adversarial networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4997-5007, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "text", + "content": "[39] Xuan Xia, Xizhou Pan, Nan Li, Xing He, Lin Ma, Xiaoguang Zhang, and Ning Ding. Gan-based anomaly detection: A review. Neurocomputing, 493:497-535, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[40] Zhaoyang Xu, Xingru Huang, Carlos Fernández Moro, Béla Bozóky, and Qianni Zhang. Gan-based virtual re-staining: a promising solution for whole slide image analysis. arXiv preprint arXiv:1901.04059, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 418 + ], + "type": "text", + "content": "[41] Bowei Zeng, Yiyang Lin, Yifeng Wang, Yang Chen, Jiuyang Dong, Xi Li, and Yongbing Zhang. Semi-supervised pr virtual staining for breast histopathological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 232-241. Springer, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 420, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 474 + ], + "type": "text", + "content": "[42] Ranran Zhang, Yankun Cao, Yujun Li, Zhi Liu, Jianye Wang, Jiahuan He, Chenyang Zhang, Xiaoyu Sui, Pengfei Zhang, Lizhen Cui, et al. Mvfstain: multiple virtual functional stain histopathology images generation based on specific domain mapping. Medical Image Analysis, 80:102520, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "text", + "content": "[43] Yijie Zhang, Luzhe Huang, Tairan Liu, Keyi Cheng, Kevin de Haan, Yuzhu Li, Bijie Bai, and Aydogan Ozcan. Virtual staining of defocused autofluorescence images of unlabeled tissue using deep neural networks. Intelligent Computing, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "type": "text", + "content": "[44] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 4, 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11268" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_content_list.json b/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..80697d767e0f399a09f36aaccf8a9a9fc9190768 --- /dev/null +++ b/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_content_list.json @@ -0,0 +1,2006 @@ +[ + { + "type": "text", + "text": "Vision-and-Language Navigation via Causal Learning", + "text_level": 1, + "bbox": [ + 212, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu*, Qijun Chen* \nSchool of Electronic and Information Engineering, Tongji University, Shanghai, China", + "bbox": [ + 138, + 180, + 828, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wly, xingchen327, dangronghao, liuchengju, qjchen}@tongji.edu.cn, mojo.shum@outlook.com", + "bbox": [ + 91, + 219, + 870, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the pursuit of robust and generalizable environment perception and language understanding, the ubiquitous challenge of dataset bias continues to plague vision-and-language navigation (VLN) agents, hindering their performance in unseen environments. This paper introduces the generalized cross-modal causal transformer (GOAT), a pioneering solution rooted in the paradigm of causal inference. By delving into both observable and unobservable confounders within vision, language, and history, we propose the back-door and front-door adjustment causal learning (BACL and FACL) modules to promote unbiased learning by comprehensively mitigating potential spurious correlations. Additionally, to capture global confounder features, we propose a cross-modal feature pooling (CFP) module supervised by contrastive learning, which is also shown to be effective in improving cross-modal representations during pre-training. Extensive experiments across multiple VLN datasets (R2R, REVERIE, RxR, and SOON) underscore the superiority of our proposed method over previous state-of-the-art approaches. Code is available at https://github.com/CrystalSixone/VLN-GOAT.", + "bbox": [ + 75, + 301, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 650, + 209, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Effective environment perception, language understanding, and historical utilization are at the core of vision-and-language navigation (VLN) [4]. Despite significant progress, deploying VLN in the real world remains a huge challenge, primarily due to diversities and uncertainties in environments and instructions. A key hindrance is dataset bias [79, 82], e.g., agents may overfit to familiar visual environments, resulting in diminished performance in environments with diverse appearances and layouts [24]. This over-reliance on specific patterns, like biased structural trajectories and repeated entity components, raises concerns about the robustness and generalizability of VLN systems.", + "bbox": [ + 75, + 676, + 468, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One way to mitigate dataset bias in VLN is to build", + "bbox": [ + 94, + 859, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8705d332ccafaf18cd2753b9850bba1e2643f8af6881359d216b300f587b5720.jpg", + "image_caption": [ + "Figure 1. In response to language instructions, the VLN agent is required to navigate to the target based on visual cues. This paper introduces a causal learning pipeline using the do-operator to reduce bias from confounders in VLN action prediction." + ], + "image_footnote": [], + "bbox": [ + 506, + 268, + 887, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "broader and more diverse datasets, which is what numerous recent studies have focused on. These include using speaker models to generate pseudo instructions [14, 16, 61, 62, 64, 65], synthesizing cross-environmental trajectories [26, 39], transferring image styles [30, 34], collecting data from the web [18, 36, 44, 70], and labeling more fine-grained, entity-aligned instructions [10, 21, 29]. However, achieving a perfectly balanced dataset devoid of bias is nearly impossible. Consequently, we often find ourselves caught in a cycle of \"creating a dataset\" - \"identifying bias\" - \"creating a new dataset\". Therefore, we are prompted to move from diagnosis to treatment [73], transitioning from the continuous collection of new datasets to the development of unbiased models that can confront and mitigate bias.", + "bbox": [ + 496, + 479, + 893, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing methods that focus on model designing mainly concentrate on introducing more types of inputs (e.g., objects [2, 12, 13, 23, 35, 63] and depth [3, 20, 40, 77]), or constructing the global graph [8, 9, 46] to represent environments. These efforts, although valuable, often overlook underlying dataset biases and the essential causal logic behind the task. In fact, the reason why humans can well execute various instructions and navigate in unknown environments is that we can learn the inherent causality of events beyond biased observation, achieving good analogical association capability. Therefore, for the first time, we propose to use causal inference [51] to equip VLN agents with similar cognitive abilities that we have, and then allow", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13139", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "them to make more reasoned decisions.", + "bbox": [ + 76, + 90, + 339, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then, how to develop such a causal inference capability? Although there is no single answer, we propose to exploit the concepts of intervention [52] - technique that uses a do-operator to alleviate the negative effects raised by confounders. Here, confounders are variables that influence both inputs and outcomes, creating spurious correlations and biases. Intervention empowers researchers to mitigate the impact of confounders, enabling the model to grasp the causation of events during data fitting. However, given the fact that VLN is such a complex task that involves cross-modal inputs and a long-term decision-making process, it is challenging to identify underlying confounders and apply intervention to debias through network learning.", + "bbox": [ + 75, + 107, + 467, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we propose a generalized cross-modal causal transformer (GOAT) approach that enables the VLN model to alleviate the negative effects raised by confounders, thus achieving causal decision-making (Fig. 1). Firstly, we propose a unified structural causal model to describe the VLN system, involving two distinct categories of confounders: observable and unobservable. Observable confounders are content-related and easily identifiable (e.g., the keywords in instructions and the room references in environments), whereas unobservable confounders refer to intricate stylistic nuances that are harder to discern but can impact the overall system (e.g., decoration styles in vision, sentence patterns in language, and trajectory trends in history). Then, we propose to address these confounders via two causal learning modules that are based on back-door and front-door adjustments [52] (namely BACL and FACL), respectively. Furthermore, to build global dictionaries for representing confounders, we devise a cross-modal feature pooling (CFP) module to effectively aggregate long-sequential features. Contrastive learning [56] is adopted to optimize CFP, serving as an additional auxiliary task during pre-training. As demonstrated by thorough experiments, our findings reveal the impact of integrating causal learning to deconfound biases on cross-modal inputs, offering valuable insights for enhancing generalization in similar tasks across diverse scenarios.", + "bbox": [ + 75, + 305, + 467, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 96, + 699, + 447, + 713 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified structural causal model for VLN by comprehensively considering the observable and unobservable confounders hidden in different modalities.", + "- we propose BACL and FACL, using the back-door and front-door adjustments to allow end-to-end unbiased cross-modal intervention and decision-making.", + "- We propose CFP, a cross-modal feature pooling module designed to aggregate sequence features for semantic alignment and confounder dictionaries construction.", + "- Our GOAT model demonstrates exceptional generalization across multiple VLN datasets (R2R [4], RxR [29], REVERIE [53], and SOON [81]), outperforming existing" + ], + "bbox": [ + 76, + 719, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "state-of-the-art methods. A comprehensive causal learning pipeline is presented to inspire future research.", + "bbox": [ + 511, + 90, + 888, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 133, + 640, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation (VLN) requires agents to navigate to specific locations [4, 29] or find target objects [53, 81] in real visual environments based on natural instructions. Its practicability has led to significant interest, showing potential in fundamental embodied AI skills. Initial models relied on recurrent neural networks [2, 11, 21, 68]. Transformer-based models [8, 9, 19, 22, 63] brought substantial progress due to their powerful long-distance encoding. However, small-scale datasets in VLN were found to cause bias, leading to serious overfitting. Consequently, several approaches were devised to tackle this challenge. Speaker-follower frameworks [16, 27, 58, 62, 64] used the speaker model to generate pseudo instructions. VLN-BERT [44], AirBERT [18], and Lily [36] collected trajectory-instruction pairs from diverse sources. REM [39] and EnvEdit [34] created new environments by editing existing environments. Although these methods improve generalization, they cannot completely eliminate inherent dataset biases. Therefore, we propose to develop an unbiased model by equipping the agent with the causal inference capability to learn cause-effect relations during data fitting, enabling them to adapt adeptly to diverse situations.", + "bbox": [ + 496, + 159, + 890, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Causal Inference is an emerging technique exploring task causality [52], leading to a surge in efforts to integrate it with deep learning, in tasks like image recognition [67, 69, 76], image captioning [38, 73], and visual question answering [33, 49]. One popular way is to use the adjustment technique to alleviate the negative effects caused by confounders, and some studies exploring the use of counterfactuals [1, 49, 50]. This paper emphasizes the adjustment method due to its practicality. However, most of the existing causal learning tasks are simple without considering more challenging tasks like VLN. Additionally, current methods apply back-door [38, 66, 75, 78] or front-door [42, 73, 74] adjustments separately across modalities, lacking comprehensive confounder assumptions and complete bias corrections. In this paper, we propose to simultaneously tackle both observable and unobservable confounders in vision, language, and history. This approach significantly reduces overall bias, enhancing the generalization capabilities of embodied VLN agents.", + "bbox": [ + 496, + 492, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 792, + 624, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Task Formulation", + "text_level": 1, + "bbox": [ + 500, + 816, + 671, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The VLN task [4] involves an embodied agent following natural language instructions to navigate real indoor environments. Matterport3D simulator [7] is used to allow interaction, where the environment is provided as graphs", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13140", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c4d24ae1254ed429bf53c78f1708a40b95a7a6c4f5935dee0fc2107b6c8e846b.jpg", + "image_caption": [ + "Figure 2. Illustration of the structural causal model of VLN." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 429, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with connected navigable nodes. The agent receives natural language instructions $\\mathcal{I} = \\{w_1, w_2, \\dots, w_L\\}$ with $L$ words, and the current panorama separated into 36 subimages $\\mathcal{V} = \\{v_1, v_2, \\dots, v_{36}\\}$ [16]. The agent also knows its current heading $\\theta$ and elevation $\\phi$ . During navigation, the agent needs to select the next point from nearby candidates or predicts the stop signal based on visual cues. Success is defined when the stop location is within 3 meters of the ground-truth position. For the goal-oriented task, REVERIE [53] and SOON [81] additionally require locating the target object at the final destination.", + "bbox": [ + 75, + 266, + 472, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Structural Causal Model of VLN", + "text_level": 1, + "bbox": [ + 76, + 449, + 367, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, we construct a structural causal model capturing the relationships among the key variables in VLN: visual observation $\\mathcal{V}$ , linguistic instruction $\\mathcal{I}$ , decision history $\\mathcal{H}$ , and action prediction $\\mathcal{A}$ . To clarify, we use $\\mathcal{X}$ to denote inputs $(\\mathcal{X} = \\{\\mathcal{V}, \\mathcal{I}, \\mathcal{H}\\})$ and $\\mathcal{Y}$ for output $(\\mathcal{Y} = \\mathcal{A})$ . In this directed acyclic graph, the starting and ending points represent the cause and effect, respectively. Traditional VLN methods focus on learning the observational association $P(\\mathcal{Y}|\\mathcal{X})$ , overlooking the ambiguity introduced by confounders $\\mathcal{Z}$ in the back-door path $\\mathcal{X} \\gets \\mathcal{Z} \\to \\mathcal{Y}$ . Here, confounders are extraneous variables that influence both causes and effects, e.g., frequently occurring content or specific attributes. $\\mathcal{Z} \\rightarrow \\mathcal{X}$ arises since the combined probability of input data is inevitably affected by the limited resources available in the real world when collection and simulation. Additionally, $\\mathcal{Z} \\rightarrow \\mathcal{Y}$ exists because collected environments, labeled instructions, or sampled trajectories also affect the probability of action distributions. These confounder links enable spurious shortcuts during training but can be detrimental in new situations.", + "bbox": [ + 75, + 474, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose to distinguish hidden confounders from different modalities into observable and unobservable categories, enhancing our prior knowledge integration and the rationality of assumption. Concretely, observable confounders encompass instances that can be recognized (e.g., room references $z_v^o$ and guiding keywords $z_i^o$ ). In contrast, unobservable confounders consist of intricate patterns and style-related elements that are challenging to qualitatively", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "describe (e.g., decoration style in vision $z_v^u$ , sentence pattern in language $z_i^u$ , and trajectory trend in history $z_h^u$ ). Since we cannot explicitly model unobservable confounders $\\mathcal{Z}^u$ , the additional mediators $\\mathcal{M}$ are inserted between $\\mathcal{X}$ and $\\mathcal{Y}$ to establish front-door paths $\\mathcal{X} \\to \\mathcal{M} \\to \\mathcal{Y}$ . Detailed adjustment methods are introduced in subsequent sections.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 498, + 196, + 635, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of the GOAT model is shown in Fig. 3. The proposed back-door and front-door adjustment causal learning modules are detailed in Sec. 4.1 and Sec. 4.2, respectively. The cross-modal feature pooling method is subsequently introduced in Sec. 4.3. Finally, a practical causal learning pipeline is presented in Sec. 4.4.", + "bbox": [ + 496, + 220, + 890, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Observable Causal Inference", + "text_level": 1, + "bbox": [ + 498, + 321, + 759, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Back-door Adjustment Causal Learning (BACL). Based on Bayes's theorem, the typical observational likelihood is as $P(\\mathcal{Y}|\\mathcal{X}) = \\sum_{z}P(\\mathcal{Y}|\\mathcal{X},z)\\underline{P(z|\\mathcal{X})}$ , where $P(z|\\mathcal{X})$ could bring biased weights. Do-operator [52] provides scientifically sound methods for determining causal effects by severing the back-door link between $\\mathcal{Z}$ and $\\mathcal{X}$ . According to the invariance and independence rules [17], we have:", + "bbox": [ + 496, + 345, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathcal {Y} | d o (\\mathcal {X})) = \\sum_ {z} P (\\mathcal {Y} | d o (\\mathcal {X}), z) P (z | d o (\\mathcal {X})) (1) \\\\ = \\sum_ {z} P (\\mathcal {Y} | \\mathcal {X}, z) \\underline {{P (z)}} (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 452, + 890, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In such case, the intervention is achieved by blocking the back-door path $\\mathcal{Z} \\to \\mathcal{X}$ , making $\\mathcal{X}$ have a fair opportunity to incorporate causality-related factors for prediction. Previous methods [42, 73, 78] used the NWGM approximation [6] to directly pursue causal learning to the final outputs. However, these methods limit the intervention only to the network's final Softmax layer, overlooking possible biased features in shallow layers. Since the conditional probability is implicit in pattern recognition made by the trained neural network [48], we release the target effect of causal hypothesis to learned features rather than merely outputs. Obtaining unbiased features leads to unbiased predictions. Consequently, the specific network module is formulated as $f(x, z)$ , and the implementation of Eq. (2) becomes:", + "bbox": [ + 496, + 516, + 890, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} (\\boldsymbol {x}, \\boldsymbol {z}) = \\mathbb {E} _ {z} [ f (\\boldsymbol {x}, \\boldsymbol {z}) ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 729, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The linear function is used as $f(\\pmb{x}, \\pmb{z}) = f_{x}(\\pmb{x}) + f_{z}(\\pmb{z})$ . Then Eq. (3) becomes $f_{x}(\\pmb{x}) + \\mathbb{E}_{z}[f_{z}(\\pmb{z})]$ . To obtain $\\mathbb{E}_{z}[f_{z}(\\pmb{z})]$ , there are two prevalent approaches: statistic-based [42, 66] and attention-based methods [38, 73]:", + "bbox": [ + 496, + 746, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S t a t i s t i c}: \\mathbb {E} _ {z} [ f _ {z} (\\boldsymbol {z}) ] = \\sum_ {i} \\frac {\\left| z _ {i} \\right|}{\\sum_ {j} \\left| z _ {j} \\right|} f _ {z} \\left(\\boldsymbol {z} _ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 808, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A t t e n t i o n}: \\mathbb {E} _ {z} [ f _ {z} (z) ] = \\sum_ {i} \\frac {\\exp \\left(h z _ {i} ^ {T}\\right)}{\\sum_ {j} \\exp \\left(h z _ {j} ^ {T}\\right)} f _ {z} (z _ {i}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 845, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $|z_{i}|$ denotes the number of $z$ belonging to the $i$ -th", + "bbox": [ + 504, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13141", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/80ffc652faa72e6959d9435f02acb9c430c4b3d49abf57a9dc5dd3bb25c1b670.jpg", + "image_caption": [ + "Figure 3. Framework of GOAT, built on the foundation of the dual-scale graph transformer [9]. Back-door and front-door adjustment causal learning mechanisms are used for mitigating spurious correlations, enabling unbiased feature learning and decision-making." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3e102ecfb78a6d6c64bc12365dccbe70f7904291db8ecde04268a23e05085d8.jpg", + "image_caption": [ + "(a) Back-door Adjustment Causal Learning (BACL)" + ], + "image_footnote": [], + "bbox": [ + 81, + 301, + 467, + 385 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7b42454595816a909fcf839037ac9659893f9dbeaa94fa97b0753d094f8bc40.jpg", + "image_caption": [ + "(b) Front-door Adjustment Causal Learning (FACL)", + "Figure 4. Illustration of BACL and FACL." + ], + "image_footnote": [], + "bbox": [ + 81, + 397, + 467, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "category in the training set, and $h$ means hidden features. The illustration is shown in Fig. 4(a). Our experiments in Sec. 5.3 reveal modality-specific calculation preferences.", + "bbox": [ + 75, + 511, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BACL in Text Content. In VLN instructions like \"Exit the office and turn right into the kitchen,\" essential guiding elements such as directions (e.g., \"exit\" and \"right\") and landmarks (e.g., \"office\" and \"kitchen\") play significant roles. These keywords which are common causes of instruction construction and action distribution, serve as observable confounders. Firstly, we build the text keyword dictionary $\\mathcal{D}_I^O = [z_{i,1}^o,z_{i,2}^o,\\dots,z_{i,K}^o ]$ with $K$ classes to store confounder features. Direction-and-landmark keywords are extracted based on their part-of-speech tags [63]. We use the pre-trained RoBERTa [41] to obtain feature representations for each extracted token $\\pmb{f}_{i}^{o}$ . Since the same word can have different features across sentences, we calculate the average feature for each keyword: $\\pmb{z}_{i,n}^{o} = \\frac{1}{|z_{i,n}^{o}|}\\sum_{j}\\pmb{f}_{i,n,j}^{o}$ . Subsequently, the text content causal representation $G_{s}$ is calculated as follows:", + "bbox": [ + 75, + 558, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI = \\operatorname {R o B E R T a} \\left(\\psi_ {t} (\\mathcal {I}) + \\psi_ {t} (\\mathcal {P})\\right), Z _ {k} = \\mathrm {L N} \\left(\\phi_ {k} \\left(\\mathcal {D} _ {I} ^ {O}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 803, + 467, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {s} = \\operatorname {L N} \\left[ \\phi_ {i} \\left(\\mathcal {B} \\left(I, Z _ {k}\\right)\\right) \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 823, + 467, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot)$ and $\\phi (\\cdot)$ denote the learnable embedding layer and the full-connection layer, respectively. The absolute positional encoding $\\mathcal{P}$ [60] is added to present the position information, and the layer normalization LN [5] is employed", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for stabilizing hidden states during training.", + "bbox": [ + 498, + 292, + 790, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BACL in Vision Content. For each step, the panorama $\\mathcal{V}$ is divided into 36 sub-images. Since existing VLN datasets primarily involve indoor room navigation, visual room references are treated as observable confounders. CLIP [56] is used to extract image features. Since room labels are not directly provided, we employ BLIP [32], a pre-trained VQA model to capture the room information for each image, by querying the model with a fixed prompt \"what kind of room is this?\" The average value of each room reference type is calculated, forming a visual room reference dictionary $\\mathcal{D}_V^O = [z_{v,1}^o,z_{v,2}^o,\\dots,z_{v,M}^o]$ , where $M$ is the number of room types. Additionally, the matrix $\\gamma = \\left[(\\sin \\theta_i,\\cos \\theta_i,\\sin \\eta_i,\\cos \\eta_i)_{i = 1}^{36}\\right]$ is used to present the direction of each image's shift relative to the agent, where $\\theta$ and $\\eta$ denote the heading and elevation direction. If there are additional object features (for goal-oriented tasks), they are concatenated with image features. Subsequently, a 2-layer transformer encoder is used to capture spatial dependencies. The above process is formulated as follows:", + "bbox": [ + 496, + 308, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = \\operatorname {C L I P} (\\mathcal {V}), Z _ {r} = \\ln \\left(\\phi_ {r} \\left(\\mathcal {D} _ {V} ^ {O}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 595, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {v} = \\operatorname {L N} \\left[ \\phi_ {v} \\left(\\mathcal {B} (V, Z _ {r})\\right) \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 614, + 890, + 631 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {v} = \\operatorname {T r a n s} \\left(V _ {v} + \\psi_ {d} (\\gamma)\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 633, + 890, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Unobservable Causal Inference", + "text_level": 1, + "bbox": [ + 500, + 664, + 777, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Front-door Adjustment Causal Learning (FACL). In the previous section, we employed the back-door adjustment technique to handle bias caused by observable confounders. However, there are additional unobservable confounders that cannot be explicitly captured and modeled in advance. To address this, we introduce another technique - front-door adjustment [17]. As shown in Fig. 4(b), an additional mediator $\\mathcal{M}$ is inserted between inputs and outcomes to construct the front door path $\\mathcal{X} \\rightarrow \\mathcal{M} \\rightarrow \\mathcal{Y}$ . In VLN, an attention-based model $P(\\mathcal{Y}|\\mathcal{X}) = \\sum_{m} P(\\mathcal{Y}|m)P(m|\\mathcal{X})$ will select key regions $\\mathcal{M}$ from inputs $\\mathcal{X}$ for action prediction $\\mathcal{Y}$ . Therefore, the model inference can be represented by two parts: the feature selector $\\mathcal{X} \\rightarrow \\mathcal{M}$ which selects suitable knowledge $\\mathcal{M}$ from $\\mathcal{X}$ , and the action predictor", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13142", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{M} \\to \\mathcal{V}$ which exploits $\\mathcal{M}$ to predict $\\mathcal{V}$ . To eliminate spurious correlation brought by unobservable confounder $\\mathcal{Z}$ , we simultaneously deploy $do$ -operator to $\\mathcal{X}$ and $\\mathcal{M}$ :", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathcal {Y} | d o (\\mathcal {X})) = \\sum_ {m} P (\\mathcal {Y} | d o (m)) P (m | d o (\\mathcal {X})) (11) \\\\ = \\sum_ {x ^ {\\prime}} P \\left(x ^ {\\prime}\\right) \\sum_ {m} P (\\mathcal {Y} | m, x ^ {\\prime}) P (m | \\mathcal {X}) (12) \\\\ = \\mathbb {E} _ {x ^ {\\prime}} \\mathbb {E} _ {m | x} [ P (\\mathcal {Y} | x ^ {\\prime}, m) ] (13) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 136, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x'$ denotes potential input samples of the whole representation space, different from current inputs $\\mathcal{X} = x$ . We use the bold symbol $\\pmb{m}$ to denote the in-sampling features obtained by the feature extractor acting on the current input, and $\\pmb{x}'$ to mean the cross-sampling features randomly sampled by the K-means-based feature selector from the entire training samples. Based on the linear mapping model, Eq. (13) becomes $\\mathbb{E}_{m|x}[m] + \\mathbb{E}_{x'}[x']$ . As it is intractable to get a closed-form solution of expectations involving the complex representation space, the estimation is achieved by the query mechanism. Two embedding functions [42, 73] are used to transmit input $\\pmb{x}$ into two query sets $g_{1} = q_{1}(\\pmb{x})$ and $g_{2} = q_{2}(\\pmb{x})$ . Then, the front-door adjustment is approximated as follows:", + "bbox": [ + 76, + 222, + 468, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {x} ^ {\\prime}} \\left[ \\boldsymbol {x} ^ {\\prime} \\right] \\approx \\sum_ {\\boldsymbol {x} ^ {\\prime}} P \\left(\\boldsymbol {x} ^ {\\prime} \\mid \\boldsymbol {g} _ {1}\\right) \\boldsymbol {x} ^ {\\prime} = \\sum_ {i} \\frac {\\exp \\left(\\boldsymbol {g} _ {1} \\boldsymbol {x} _ {i} ^ {\\prime T}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {g} _ {1} \\boldsymbol {x} _ {j} ^ {\\prime T}\\right)} \\boldsymbol {x} _ {i} ^ {\\prime} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 422, + 468, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {m \\mid x} [ \\boldsymbol {m} ] \\approx \\sum_ {m} P (\\boldsymbol {m} | \\boldsymbol {g} _ {2}) \\boldsymbol {m} = \\sum_ {i} \\frac {\\exp \\left(\\boldsymbol {g} _ {2} \\boldsymbol {m} _ {i} ^ {T}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {g} _ {2} \\boldsymbol {m} _ {j} ^ {T}\\right)} \\boldsymbol {m} _ {i} (15) \\\\ \\mathcal {F} \\left(\\boldsymbol {x}, \\boldsymbol {x} ^ {\\prime}\\right) = \\mathbb {E} _ {\\boldsymbol {x} ^ {\\prime}} \\left[ \\boldsymbol {x} ^ {\\prime} \\right] + \\mathbb {E} _ {m | x} [ \\boldsymbol {m} ] (16) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 460, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above process can be efficiently implemented using the multi-head attention [60], enabling seamless integration of causal adjustments into existing transformer-based frameworks with minimal modifications.", + "bbox": [ + 76, + 525, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FACL in Text, Vision, and History. Considering the characteristics of VLN, we propose to eliminate unobservable confounders from three kinds of inputs in VLN, i.e., vision, language, and history. First, following previous graph-based methods [9, 63], we construct the vision sequence $G_{v} = \\{[\\mathrm{STOP}], [\\mathrm{MEM}], Q_{v}\\}$ and the history sequence $G_{h} = \\{[\\mathrm{STOP}], [\\mathrm{MEM}], \\{\\bar{Q}_{t}\\}_{t=1}^{T}\\}$ by adding the additional token for presenting stop and recurrent memory states, respectively. $\\bar{Q}_{t}$ means the learned weight sum of panoramic features for the $t$ -th step. To condense the lengthy sequence of features and generate global features $x'$ for cross-sampling, we devise the CFP module (as described in Sec. 4.3) with the attentive pooling mechanism to construct confounder dictionaries for vision, history, and instruction, denoted as $D_{v}^{U}, D_{h}^{U}$ , and $D_{s}^{U}$ , respectively. Then the causality-enhanced features $R_{v}, R_{h}$ and $R_{s}$ are calculated based on Eq. (16):", + "bbox": [ + 76, + 584, + 468, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {v} = \\mathcal {F} \\left(G _ {v}, D _ {v} ^ {U}\\right), R _ {h} = \\mathcal {F} \\left(G _ {h}, D _ {h} ^ {U}\\right), R _ {s} = \\mathcal {F} \\left(G _ {s}, D _ {s} ^ {U}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 829, + 468, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we introduce an adaptive gate fusion (AGF) method to enhance the stability of learning by integrating", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c384ce9f20c3f5aa2f54af95e127afce576f5343224f778a1cfcf400a372ad89.jpg", + "image_caption": [ + "Figure 5. Illustration of the cross-modal feature pooling (CFP)." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "causality-enhanced features with the original context features for each modality:", + "bbox": [ + 498, + 226, + 888, + 255 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {x} = \\delta \\left(R _ {x} W _ {x} + G _ {x} W _ {g} + b\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 257, + 888, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {x} = \\omega_ {x} \\odot R _ {x} + (1 - \\omega_ {x}) \\odot G _ {x} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 276, + 888, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ and $\\odot$ mean the Sigmoid function and elementwise multiplication. Suppose $R_{x}$ and $G_{x} \\in \\mathbb{R}^{L_{x} \\times d_{h}}$ , then $W_{x/g} \\in \\mathbb{R}^{d_{h} \\times 1}$ and $b \\in \\mathbb{R}^{L_{x} \\times 1}$ are learnable parameters. Next, the cross-modal fused local features $F_{l}$ and global features $F_{g}$ are obtained by the cross-attention encoders $\\mathcal{C}$ from METER [15]. The dynamic fusion $\\mathcal{DF}$ [9] followed by Softmax $\\mathcal{SF}$ is applied for action prediction:", + "bbox": [ + 498, + 292, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l} = \\mathcal {C} \\left(U _ {v}, U _ {s}, U _ {s}\\right), F _ {g} = \\mathcal {C} \\left(U _ {h}, U _ {s}, U _ {s}\\right) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 398, + 888, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {f} = \\mathcal {D F} \\left(F _ {l}, F _ {g}\\right), y _ {t} = \\mathcal {S F} \\left(F _ {f}\\right) \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 417, + 888, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cross-entropy loss is used to optimize the network:", + "bbox": [ + 517, + 435, + 888, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = \\sum_ {t = 1} ^ {T} - \\log P \\left(y _ {t} ^ {*} \\mid \\mathcal {I}, \\mathcal {V} _ {t}, \\mathcal {H} _ {1: t - 1}\\right) \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 462, + 888, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Cross-modal Feature Pooling", + "text_level": 1, + "bbox": [ + 500, + 508, + 759, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One challenge of implementing the front-door adjustment in VLN is constructing efficient dictionaries for global features from long sequences. This requires compressing sequential features of varying lengths into a unified feature space to represent each sample effectively. Let $H \\in \\mathbb{R}^{L \\times d_h}$ be the sequential features, the following attentive pooling is used to effectively compress the sequence length:", + "bbox": [ + 496, + 532, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\mathcal {T} (H), \\alpha = \\mathcal {S F} \\left(A W _ {a}\\right), \\bar {H} = \\mathcal {T} \\left(\\alpha^ {T} H\\right) \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 638, + 888, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}$ denote the Tanh activation, $W_{a} \\in \\mathbb{R}^{d_{h} \\times 1}$ is the learnable attention matrix, and $\\bar{H} \\in \\mathbb{R}^{1 \\times d_{h}}$ . As shown in Fig. 5, for vision, history, local-global fusion, and text features, we use one transformer layer as the encoder followed by the attentive pooling to obtain flattened features $\\bar{C}^{l}, \\bar{C}^{g}, \\bar{C}^{f}$ , and $\\bar{C}^{s}$ , respectively. Then, we adopt contrastive learning [35, 57, 62] to optimize this cross-modal feature pooling (CFP) module, meanwhile improving semantic alignments for different modalities. The contrastive loss $\\mathcal{L}_{ls}$ is constructed as:", + "bbox": [ + 496, + 657, + 890, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {l s} = - \\frac {1}{2 B} \\sum_ {j = 1} ^ {B} \\log \\frac {\\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {j} ^ {s} \\right\\rangle / t\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)} \\tag {24} \\\\ - \\frac {1}{2 B} \\sum_ {k = 1} ^ {B} \\log \\frac {\\exp \\left(\\left\\langle \\bar {C} _ {k} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 816, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ and $t$ mean the batch size and temperature, respectively. Similarly, contrastive losses $\\mathcal{L}_{gs}$ and $\\mathcal{L}_{fs}$ are calculated by replacing $\\bar{C}^l$ with $\\bar{C}^g$ and $\\bar{C}^f$ . The overall CFP loss is the sum of these losses $\\mathcal{L}_{\\mathrm{CFP}} = \\mathcal{L}_{ls} + \\mathcal{L}_{gs} + \\mathcal{L}_{fs}$ .", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable the network more adaptive to characteristics for VLN and thus facilitate the building of the front-door confounder dictionaries for samples, we train the CFP module alongside other auxiliary tasks [9] during pre-training. Subsequently, the trained attentive pooling modules are used to extract global features for different modalities. In the fine-tuning stage, we employ the BACL and FACL with established dictionaries for intervention. The CFP offers dual benefits: it aligns different modalities more effectively during pre-training and provides a systematic approach for extracting coherent representations from sequence inputs.", + "bbox": [ + 75, + 152, + 472, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Causal Learning Pipeline", + "text_level": 1, + "bbox": [ + 76, + 329, + 307, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 6, we summarize a causal learning pipeline that serves as a blueprint for similar learning-based methods. First, it begins with task formulation, where the specific task and its objectives are precisely defined. Next, the observable and unobservable confounders are explicitly assumed. Both back-door and front-door adjustment strategies are employed to tackle these confounders, either simultaneously or sequentially, contingent on task specifics. Then, the pipeline proceeds", + "bbox": [ + 75, + 353, + 279, + 625 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/275afebef1bbe4e948e4a9c031740d79c720e486ebf20ca5a47b6ae66964b3e0.jpg", + "image_caption": [ + "Figure 6. Pipeline of the causal learning." + ], + "image_footnote": [], + "bbox": [ + 310, + 368, + 468, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to model calculation and result prediction. Throughout network optimization, both network parameters and confounder features are continuously updated. Ultimately, this iterative process leads us to the development of a robust causal model capable of generating unbiased features, thereby advancing the generalizability of AI systems.", + "bbox": [ + 75, + 625, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 755, + 284, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Datasets. We verify GOAT on two kinds of VLN benchmarks: fine-grained datasets (R2R [4] and RxR-English [29]), which provide long step-by-step navigation instructions, and goal-oriented datasets (REVERIE [53] and SOON [81]), which additionally requires for the target object. Formally, the datasets are partitioned into four splits: training, validation seen (sharing the same environments with the training set), validation unseen (having different", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "environments from the training set), and test unseen sets (reported by the online leaderboard for fair comparison).", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) Evaluation Metrics. In R2R, key metrics include Navigation Error (NE), Success Rate (SR), Oracle SR (OSR), and SR Weighted by Path Length (SPL). RxR adds Normalized Dynamic Time Warping (nDTW) and SR Weighted by Dynamic Time Warping (sDTW). REVERIE and SOON introduce Remote Grounding Success Rate (RGS) and RGS Weighted by Path Length (RGSPL).", + "3) Implementation Details. Our model consists of 6 transformer layers for text, 2 for panorama, and 3 for cross-modal encoding. We use CLIP-B/16 [56] for image feature extraction and initialize network weights with METER [15]. In pre-training, MLM [28], SAP [8], and the proposed CFP are for R2R and RxR. OG [37] is added for REVERIE and SOON. EnvEdit [34] is employed for feature augmentation. The synthetic extended datasets [19, 62, 65] are used for R2R, REVERIE, and RxR, respectively. Pretraining is done on a single Tesla V100 GPU for a maximum of 300K iterations by the AdamW [43] optimizer, with a batch size of 48 and a learning rate of $5 \\times 10^{-5}$ . The numbers of classes of keywords and rooms are 74 and 50, and the temperature $t$ is set to 1. In fine-tuning, the front-door dictionaries are randomly sampled from the K-Means clustering features. As the text transformer RoBERTa [41] is involved in end-to-end training, the textual keywords dictionary is also iteratively updated. Speaker models with the environmental dropout [58, 62, 64] are used to provide dynamic pseudo labels. We employ batch size 12 for R2R, REVERIE, and 5 for RxR and SOON, with a learning rate of $2 \\times 10^{-5}$ and a maximum of 100K iterations." + ], + "bbox": [ + 496, + 121, + 893, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 498, + 570, + 812, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, 2, 3, 4, we compare GOAT with the previous state-of-the-art (SoTA) methods on the R2R, REVERIE, RxR-English, and SOON datasets, respectively. On all these four datasets, our approach exhibits superior navigation performance, precise instruction-following alignment, and accurate object grounding across both seen and unseen environments. For instance, in R2R, GOAT achieves remarkable improvements in SPL compared to BEVBert [3], with relative increases of $7.41\\%$ , $6.45\\%$ , and $4.74\\%$ on three subsets. In REVERIE, GOAT shows substantial relative enhancements in RGSPL by $11.83\\%$ , $6.55\\%$ , and $20.87\\%$ on three subsets. In the challenging SOON and RxR tasks, GOAT also exhibits significant improvements in performance metrics, highlighting its robustness and superior generalization capabilities over previous methods.", + "bbox": [ + 496, + 594, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.Quantitative Analysis", + "text_level": 1, + "bbox": [ + 500, + 830, + 702, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Effect of Causal Inference. Fig. 7 verifies the impact of causal inference on GOAT across four diverse VLN datasets in unseen environments. \"W/o intervention\" signifies the", + "bbox": [ + 500, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13144", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0312d508b57cf705b0c29ddb4353bda04b6f44dd9143bc1afb4dea0791df113a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑
HAMT [8]76722.518266613.297365603.9372
DUET [9]79732.288672603.318169593.6576
TD-STP [80]77732.348370633.227667613.7372
GeoVLN [25]79762.22-68633.35-65613.95-
DSRG [63]81762.238873623.008172613.3378
GridMM [71]----75642.83-73623.35-
GELA [10]76732.39-71653.11-67623.59-
EnvEdit [34]77742.32-69643.24-68643.59-
BEVBert [3]81742.178875642.818473623.1381
GOAT (Ours)83.7479.481.7988.6477.8268.132.4084.7274.5764.943.0480.35
", + "bbox": [ + 119, + 88, + 851, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/151ac0f56fd36ec88f940dc3de6f750d6c4ca981619f12ad98b87a9dfb9f4b00.jpg", + "table_caption": [ + "Table 1. Comparison with other state-of-the-art methods on the R2R dataset [4]. '-': unavailable statistics." + ], + "table_footnote": [], + "table_body": "
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑
HAMT [8]43.2940.1927.2025.1832.9530.2018.9217.2830.4026.6714.8813.08
HOP+ [54]55.8749.5540.7636.2236.0731.1322.4919.3333.8228.2420.2016.86
DUET [9]71.7563.9457.4151.1446.9833.7332.1523.0352.5136.0631.8822.06
DSRG [63]75.6968.0961.0754.7247.8334.0232.6923.3754.0437.0932.4922.18
GridMM [71]----51.3736.4734.5724.5653.1336.6034.8723.45
BEVBert [3]73.7265.3257.7051.7351.7836.3734.7124.4452.8136.4132.0622.09
GOAT (Ours)78.6471.4063.7457.8553.3736.7038.4326.0957.7240.5338.3226.70
", + "bbox": [ + 86, + 282, + 879, + 411 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b86268a41c7bc8f58b50f5ee8f8c9a4982dcaa7502c200d46c27e546a06316e.jpg", + "table_caption": [ + "Table 2. Comparison with other state-of-the-art methods on the REVERIE dataset [53]. '-': unavailable statistics." + ], + "table_footnote": [], + "table_body": "
MethodValidation SeenValidation Unseen
SR↑SPL↑nDTW↑sDTW↑SR↑SPL↑nDTW↑sDTW↑
Syntax [31]48.144.058.040.039.235.052.032.0
SOAT [45]----44.2-54.836.4
HOP+ [54]53.647.959.043.045.738.452.036.0
FOAM [14]----42.838.754.135.6
ADAPT [35]50.344.656.340.646.940.254.137.7
\\( \\mathsf{MAR}_{\\mathsf{M-MP}} \\) [27]----50.2-60.343.9
VLN-PETL [55]60.556.865.751.757.954.264.949.7
GOAT (Ours)74.168.171.061.468.261.767.156.6
", + "bbox": [ + 78, + 450, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2bbfc5d31cd83483d7a4a76f819f3b8b456f0af6c0215402e843bdbb0fc35ef.jpg", + "table_caption": [ + "Table 3. Comparison on the RxR-English dataset [29]." + ], + "table_footnote": [], + "table_body": "
MethodValidation UnseenTest Unseen
OSR↑SR↑SPL↑RGSPL↑OSR↑SR↑SPL↑RGSPL↑
GBE [81]28.5419.5213.341.1621.4512.909.230.45
DUET [9]50.9136.2822.583.7543.0033.4421.424.17
GridMM [71]53.3937.4624.813.9148.0236.2721.254.15
GOAT (Ours)54.6940.3528.055.8550.6340.5025.186.10
", + "bbox": [ + 78, + 612, + 468, + 694 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4f989acaa141cb83a38070786678f8a1a76c54893358743b8d9d931f1eadec2.jpg", + "image_caption": [ + "Figure 7. Effect of the intervention on various VLN datasets." + ], + "image_footnote": [], + "bbox": [ + 78, + 729, + 465, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exclusion of the proposed BACL and FACL interventions across all modalities. On each of the datasets examined, the", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9aa8a1c2c5ba774097b713babb8fd89ed2bdc4b814d4d90b3b2fa459b157ac0.jpg", + "table_caption": [ + "Table 4. Comparison on the SOON dataset [81]." + ], + "table_footnote": [], + "table_body": "
IdBACLFACLSR↑SPL↑NE↓OSR↑
1XX75.2764.692.7283.14
2✓X76.3767.132.6083.99
3X✓76.5066.922.5384.21
4✓✓77.8268.132.4084.72
", + "bbox": [ + 519, + 450, + 874, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Effect of back-door and front-door adjustments.", + "bbox": [ + 526, + 529, + 862, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "integration of causal inference leads to significant enhancements in the model's performance. This strongly demonstrates causal learning's considerable popularization potential in enhancing learning-based model generalization.", + "bbox": [ + 496, + 556, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) Effect of BACL and FACL. Tab. 5 analyzes the effects of the proposed BACL and FACL on the R2R val-unseen subset. Compared to the baseline (#1), individual application of either BACL (#2) or FACL (#3) leads to performance improvements. Concurrent use of BACL and FACL (#4) leads to further performance enhancements. These findings underscore our assumption about the presence of both observable and unobservable confounders. Integrating both back-door and front-door adjustments is crucial to comprehensively addressing dataset biases, and enhancing the model's robustness and generalization.", + "3) Effect of CFP. In Tab. 6, we assess the efficacy of the proposed CFP on the R2R val-unseen subset. During pretraining (PT), incorporating CFP as an additional auxiliary task (CFP-P) enhances training performance, improving SR and SPL by $3.11\\%$ and $3.02\\%$ (#A1), respectively. In the fine-tuning (FT) stage, we compare the performance with and without the use of trained attention modules from CFP" + ], + "bbox": [ + 496, + 622, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13145", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bd97183c570501cdf3a7e14a584e7771775712d3fe49bb88e512d3b74ad61b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StageIdMethodSR↑SPL↑NE↓OSR↑
PTA0w/o CFP-P40.3637.886.5248.87
A1w/ CFP-P43.4740.906.0653.13
FTB0A0 w/o CFP-F75.5665.902.6382.42
B1A1 w/o CFP-F76.6366.172.6384.67
B2A1 w/ CFP-F77.8268.132.4084.72
", + "bbox": [ + 80, + 88, + 468, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4365e56e31cff8d249a49777282cef7ef194f746777f579ad5f33a855c808aeb.jpg", + "table_caption": [ + "Table 6. Effect of CFP in pre-training and fine-tuning." + ], + "table_footnote": [], + "table_body": "
IdTextVisionSR↑SPL↑NE↓OSR↑
1StatsStats75.2264.782.7183.65
2StatsAttn76.5965.392.5685.31
3AttnStats77.8268.132.4084.72
4AttnAttn75.9565.832.6483.91
", + "bbox": [ + 99, + 213, + 447, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Effect of Statistic and Attention methods in BACL.", + "bbox": [ + 94, + 291, + 450, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47a6a564e7234786ba2b7ba4463bd1a41ee5cc62862ef899ec9eb2809a49fb1e.jpg", + "image_caption": [ + "Figure 8. Comparison of the distribution of SR and SPL of instructions involving specific terms on the R2R val-unseen split." + ], + "image_footnote": [], + "bbox": [ + 78, + 321, + 468, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to extract global features for front-door dictionaries (CFP-F). \"W/o CFP-F\" signifies the use of simple average pooling to compress features from the pre-trained model. #B2 shows that the CFP provides more reliable confounder representations for causal learning (SPL $\\uparrow 1.96\\%$ ).", + "bbox": [ + 75, + 484, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4) Effect of Different BACL in Different Modalities. Tab. 7 investigates the effect of various combinations of statistic-based and attention-based methods for text and vision in BACL on the R2R val-unseen subset. The results indicate that employing the attention method for text and the statistic method for vision yields the best performance (#3). Intuitively, this can be explained by the structured nature of textual information and the involvement of RoBERTa's end-to-end training, enabling the attention method to effectively capture contextual nuances. Conversely, images lack explicit causality, and CLIP, the image extractor, isn't trained directly for efficiency reasons. Consequently, the statistic method ensures a stable causal learning process, preserving the integrity of vision-related features.", + "bbox": [ + 75, + 561, + 468, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 785, + 267, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1) Bias Elimination Effect. In Fig. 8, the compactness of the boxes represents concentrated data distribution and reduced variability, while a median line closer to the center signifies even data distribution. It shows that GOAT obtains narrower boxes and more central midlines across diverse objects. This finding showcases that with the integration of", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b891ed21cd13a524a64835a274d050eef13f0d10953fd8fee0fccd7a64edc064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 92, + 529, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Starting point", + "bbox": [ + 531, + 93, + 601, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b76c982cc7e2998589bc6f30992839df3bccb0424ac074f8dd852fcac86d9bc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 93, + 633, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Destination", + "bbox": [ + 635, + 93, + 689, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdc97a6da8f62ed7905910767cf90479b83fa0595f9c90896f98a25bed6f87ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 93, + 728, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sited point", + "bbox": [ + 730, + 93, + 782, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac5893ffc1d6d8a8a2080b95c6c45e3e79dee430a2ca9c97049eaaff89b0863c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 93, + 813, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "visited point", + "bbox": [ + 825, + 93, + 882, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/adb9e57de36b26ee94ec1eafaa75f51e3c88f4498fff801771719b7c12ed4526.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 107, + 529, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground-truth", + "bbox": [ + 531, + 107, + 598, + 118 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5110e0d59417ee3c68d119ed32530a3ae9227ecdd506b1f29d190786e8d5553a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 107, + 633, + 118 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DSRG", + "bbox": [ + 635, + 107, + 661, + 118 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ab80d19cffa611cda098bd44745bba0d2ef722bf0d7ad3a4548ab15b651b3f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 107, + 683, + 118 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GOAT (Ours)", + "bbox": [ + 684, + 107, + 754, + 118 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Instruction:", + "text_level": 1, + "bbox": [ + 511, + 125, + 598, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Walk out of the room and go past the stairs and into the bedroom that is straight ahead, go to the bathroom doorway and wait there.", + "bbox": [ + 511, + 136, + 705, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/558c752f0ac4f1db01869ae62f6f41379dd5aa3eddb12983f197cc3594dac444.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 118, + 888, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Instruction:", + "text_level": 1, + "bbox": [ + 511, + 191, + 599, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Slightly move over the green chair closest to the corner of the reception desk in the lounge.", + "bbox": [ + 511, + 204, + 656, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6f1139e53657d981f57a2a45b4c77dc05c0e975f6313697287caa05faf5105d.jpg", + "image_caption": [ + "Figure 9. Predicted trajectories in unseen environments." + ], + "image_footnote": [], + "bbox": [ + 668, + 186, + 888, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "causal intervention, GOAT significantly reduces prediction bias, thereby enhancing its generalization capability in previously unseen environments.", + "bbox": [ + 496, + 287, + 890, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2) Visualized Trajectories. In Fig. 9, we visualize some predicted trajectories in unseen environments, comparing them with DSRG on R2R and REVERIE datasets. Notably, GOAT precisely captures directional cues like \"straight ahead\" and nuanced instructions like \"closest to the corner\", enabling accurate predictions. These instances highlight the intricate causal connections in VLN tasks, where specific instructions prompt corresponding actions. GOAT's enhanced causal inference capability enables it to generate more reasoned responses aligned with the provided instructions, underscoring the significance of robust causal inference in VLN systems. Please refer to our supplementary material for more detailed discussions and visualizations.", + "bbox": [ + 496, + 334, + 892, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 545, + 617, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work presents GOAT, a novel approach that addresses the dataset bias in VLN from the perspective of causal learning. The back-door and front-door adjustment causal learning (BACL and FACL) mechanisms are proposed to adjust for observable and unobservable confounders, respectively. The cross-modal feature pooling (CFP) module is adopted to promote feature learning and extraction through contrastive learning. The practical causal learning pipeline is presented to illuminate other similar learning-based methods. Experiments on R2R, REVERIE, RxR, and SOON datasets show that GOAT can reasonably discover sequence visual-linguistic causal structures and significantly improve performance. Beyond VLN, the underlying confounder assumption and causal inference principles are generalizable to other similar fields.", + "bbox": [ + 496, + 571, + 890, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 814, + 658, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper is supported by the National Natural Science Foundation of China under Grants (62233013, 62073245, 62173248). Shanghai Science and Technology Innovation Action Plan (22511104900).", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13146", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10044-10054, 2020. 2", + "[2] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5101-5109, 2021. 1, 2", + "[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 6, 7", + "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 6, 7, 14", + "[5] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. ArXiv, abs/1607.06450, 2016. 4", + "[6] Pierre Baldi and Peter Sadowski. The dropout learning algorithm. Artificial intelligence, 210:78-122, 2014. 3", + "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2", + "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7", + "[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 5, 6, 7, 14", + "[10] Yibo Cui, Liang Xie, Yakun Zhang, Meishan Zhang, Ye Yan, and Erwei Yin. Grounded entity-landmark adaptive pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12043–12053, 2023. 1, 7", + "[11] Ronghao Dang, Zhuofan Shi, Liuyi Wang, Zongtao He, Chengju Liu, and Qijun Chen. Unbiased directed object attention graph for object navigation. In Proceedings of the 30th ACM International Conference on Multimedia, page 3617-3627, New York, NY, USA, 2022. Association for Computing Machinery. 2", + "[12] Ronghao Dang, Lu Chen, Liuyi Wang, He Zongtao, Chengju Liu, and Qijun Chen. Multiple thinking achieving meta-ability decoupling for object navigation. In International Conference on Machine Learning (ICML), 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ronghao Dang, Liuyi Wang, Zongtao He, Shuai Su, Jiagui Tang, Chengju Liu, and Qijun Chen. Search for or navigate to? dual adaptive thinking for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8250-8259, 2023. 1", + "[14] Zi-Yi Dou and Nanyun Peng. Foam: A follower-aware speaker model for vision-and-language navigation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4332-4340, 2022. 1, 7", + "[15] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 5, 6", + "[16] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. Advances in Neural Information Processing Systems, 31, 2018. 1, 2, 3", + "[17] Madelyn Glymour, Judea Pearl, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016. 3, 4", + "[18] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 1, 2", + "[19] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13137-13146, 2020. 2, 6", + "[20] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1", + "[21] Zongtao He, Liuyi Wang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Mlanet: Multi-level attention network with sub-instruction for continuous vision-and-language navigation. arXiv preprint arXiv:2303.01396, 2023. 1, 2", + "[22] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2", + "[23] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3055-3067, 2023. 1", + "[24] Ronghang Hu, Daniel Fried, Anna Rohrbach, Dan Klein, Trevor Darrell, and Kate Saenko. Are you looking? grounding to multiple modalities in vision-and-language navigation." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13147", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6551-6557, 2019. 1", + "[25] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23212-23221, 2023. 7", + "[26] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1862-1872, 2019. 1", + "[27] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 7", + "[28] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 6", + "[29] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392-4412, 2020. 1, 2, 6, 7, 14", + "[30] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 1", + "[31] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1041–1050, 2021. 7", + "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4", + "[33] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21273–21282, 2022. 2", + "[34] Jialu Li, Hao Tan, and Mohit Bansal. Envideit: Environment editing for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15407-15417, 2022. 1, 2, 6, 7", + "[35] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In Pro" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2022. 1, 5, 7", + "[36] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8317-8326, 2023. 1, 2", + "[37] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7036-7045, 2021. 6", + "[38] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18041-18050, 2022. 2, 3", + "[39] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1644-1654, 2021. 1, 2", + "[40] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1", + "[41] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6", + "[42] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3, 5", + "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6", + "[44] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In European Conference on Computer Vision, pages 259-274. Springer, 2020. 1, 2", + "[45] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34:7357-7367, 2021. 7", + "[46] Laksh Nanwani, Anmol Agarwal, Kanishk Jain, Raghav Prabhakar, Aaron Monis, Aditya Mathur, Krishna Murthy, Abdul Hafez, Vineet Gandhi, and K Madhava Krishna. Instance-level semantic maps for vision language navigation. arXiv preprint arXiv:2305.12363, 2023. 1", + "[47] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13148", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Natural Language Processing (EMNLP-IJCNLP), pages 684-695, 2019. 16", + "[48] Siqi Nie, Meng Zheng, and Qiang Ji. The deep regression bayesian network and its applications: Probabilistic deep learning for computer vision. IEEE Signal Processing Magazine, 35(1):101-111, 2018. 3", + "[49] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2", + "[50] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Javen Qinfeng Shi, and Anton Van den Hengel. Counterfactual vision-and-language navigation: Unravelling the unseen. Advances in Neural Information Processing Systems, 33:5296-5307, 2020. 2", + "[51] Judea Pearl. Causality. Cambridge university press, 2009. 1", + "[52] Judea Pearl and Dana Mackenzie. The book of why: the new science of cause and effect. Basic books, 2018. 2, 3, 13", + "[53] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 3, 6, 7, 14", + "[54] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7", + "[55] Yanyuan Qiao, Zheng Yu, and Qi Wu. Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 7", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6", + "[57] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? In International Conference on Learning Representations. 5", + "[58] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, 2019. 2, 6", + "[59] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning, pages 394-406. PMLR, 2020. 16", + "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "information processing systems, pages 5998-6008, 2017. 4, 5", + "[61] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15471-15481, 2022. 1", + "[62] Liuyi Wang, Zongtao He, Ronghao Dang, Huiyi Chen, Chengju Liu, and Qijun Chen. Res-sts: Referring expression speaker via self-training with scorer for goal-oriented vision-language navigation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 2, 5, 6", + "[63] Liuyi Wang, Zongtao He, jiagui Tang, Ronghao Dang, naijia Wang, Chengju Liu, and Qijun Chen. A dual semantic-aware recurrent global-adaptive network for vision-and-language navigation. In International Joint Conferences on Artificial Intelligence (IJCAI), 2023. 1, 2, 4, 5, 7", + "[64] Liuyi Wang, Chengju Liu, Zongtao He, Shu Li, Qingqing Yan, Huiyi Chen, and Qijun Chen. Pasts: Progress-aware spatio-temporal transformer speaker for vision-and-language navigation. arXiv preprint arXiv:2305.11918, 2023. 1, 2, 6", + "[65] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 1, 6", + "[66] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10760-10770, 2020. 2, 3", + "[67] Tan Wang, Chang Zhou, Qianru Sun, and Hanwang Zhang. Causal attention for unbiased visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3091-3100, 2021. 2", + "[68] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Wang, and Lei Zhang. Vision-language navigation policy learning and adaptation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2", + "[69] Yuqing Wang, Xiangxian Li, Zhuang Qi, Jingyu Li, Xuelong Li, Xiangxu Meng, and Lei Meng. Meta-causal feature learning for out-of-distribution generalization. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 530-545. Springer, 2023. 2", + "[70] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12009-12020, 2023. 1", + "[71] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15625-15636, 2023. 7" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13149", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 16", + "[73] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 3, 5", + "[74] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 2", + "[75] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. Advances in neural information processing systems, 33:2734-2746, 2020. 2", + "[76] Hua Zhang, Liqiang Xiao, Xiaochun Cao, and Hassan Foroosh. Multiple adverse weather conditions adaptation for object detection via causal intervention. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2", + "[77] Jiazhao Zhang, Liu Dai, Fanpeng Meng, Qingnan Fan, Xuelin Chen, Kai Xu, and He Wang. 3d-aware object goal navigation via simultaneous exploration and identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6672-6682, 2023. 1", + "[78] Shengyu Zhang, Tan Jiang, Tan Wang, Kun Kuang, Zhou Zhao, Jianke Zhu, Jin Yu, Hongxia Yang, and Fei Wu. Devlbert: Learning deconfounded visio-linguistic representations. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4373–4382, 2020. 2, 3", + "[79] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, 2021. 1", + "[80] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4194-4203, 2022. 7", + "[81] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12699, 2021. 2, 3, 6, 7, 14", + "[82] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1" + ], + "bbox": [ + 78, + 90, + 468, + 824 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13150", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_model.json b/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d0c9f6a8539873d0b4674875030fb692532585fc --- /dev/null +++ b/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_model.json @@ -0,0 +1,2996 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.758, + 0.154 + ], + "angle": 0, + "content": "Vision-and-Language Navigation via Causal Learning" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.181, + 0.829, + 0.218 + ], + "angle": 0, + "content": "Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu*, Qijun Chen* \nSchool of Electronic and Information Engineering, Tongji University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.22, + 0.872, + 0.235 + ], + "angle": 0, + "content": "{wly, xingchen327, dangronghao, liuchengju, qjchen}@tongji.edu.cn, mojo.shum@outlook.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.621 + ], + "angle": 0, + "content": "In the pursuit of robust and generalizable environment perception and language understanding, the ubiquitous challenge of dataset bias continues to plague vision-and-language navigation (VLN) agents, hindering their performance in unseen environments. This paper introduces the generalized cross-modal causal transformer (GOAT), a pioneering solution rooted in the paradigm of causal inference. By delving into both observable and unobservable confounders within vision, language, and history, we propose the back-door and front-door adjustment causal learning (BACL and FACL) modules to promote unbiased learning by comprehensively mitigating potential spurious correlations. Additionally, to capture global confounder features, we propose a cross-modal feature pooling (CFP) module supervised by contrastive learning, which is also shown to be effective in improving cross-modal representations during pre-training. Extensive experiments across multiple VLN datasets (R2R, REVERIE, RxR, and SOON) underscore the superiority of our proposed method over previous state-of-the-art approaches. Code is available at https://github.com/CrystalSixone/VLN-GOAT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.21, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Effective environment perception, language understanding, and historical utilization are at the core of vision-and-language navigation (VLN) [4]. Despite significant progress, deploying VLN in the real world remains a huge challenge, primarily due to diversities and uncertainties in environments and instructions. A key hindrance is dataset bias [79, 82], e.g., agents may overfit to familiar visual environments, resulting in diminished performance in environments with diverse appearances and layouts [24]. This over-reliance on specific patterns, like biased structural trajectories and repeated entity components, raises concerns about the robustness and generalizability of VLN systems." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.86, + 0.47, + 0.875 + ], + "angle": 0, + "content": "One way to mitigate dataset bias in VLN is to build" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.269, + 0.888, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.41, + 0.895, + 0.466 + ], + "angle": 0, + "content": "Figure 1. In response to language instructions, the VLN agent is required to navigate to the target based on visual cues. This paper introduces a causal learning pipeline using the do-operator to reduce bias from confounders in VLN action prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.895, + 0.691 + ], + "angle": 0, + "content": "broader and more diverse datasets, which is what numerous recent studies have focused on. These include using speaker models to generate pseudo instructions [14, 16, 61, 62, 64, 65], synthesizing cross-environmental trajectories [26, 39], transferring image styles [30, 34], collecting data from the web [18, 36, 44, 70], and labeling more fine-grained, entity-aligned instructions [10, 21, 29]. However, achieving a perfectly balanced dataset devoid of bias is nearly impossible. Consequently, we often find ourselves caught in a cycle of \"creating a dataset\" - \"identifying bias\" - \"creating a new dataset\". Therefore, we are prompted to move from diagnosis to treatment [73], transitioning from the continuous collection of new datasets to the development of unbiased models that can confront and mitigate bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.896, + 0.903 + ], + "angle": 0, + "content": "However, existing methods that focus on model designing mainly concentrate on introducing more types of inputs (e.g., objects [2, 12, 13, 23, 35, 63] and depth [3, 20, 40, 77]), or constructing the global graph [8, 9, 46] to represent environments. These efforts, although valuable, often overlook underlying dataset biases and the essential causal logic behind the task. In fact, the reason why humans can well execute various instructions and navigate in unknown environments is that we can learn the inherent causality of events beyond biased observation, achieving good analogical association capability. Therefore, for the first time, we propose to use causal inference [51] to equip VLN agents with similar cognitive abilities that we have, and then allow" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13139" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.34, + 0.106 + ], + "angle": 0, + "content": "them to make more reasoned decisions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Then, how to develop such a causal inference capability? Although there is no single answer, we propose to exploit the concepts of intervention [52] - technique that uses a do-operator to alleviate the negative effects raised by confounders. Here, confounders are variables that influence both inputs and outcomes, creating spurious correlations and biases. Intervention empowers researchers to mitigate the impact of confounders, enabling the model to grasp the causation of events during data fitting. However, given the fact that VLN is such a complex task that involves cross-modal inputs and a long-term decision-making process, it is challenging to identify underlying confounders and apply intervention to debias through network learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.468, + 0.697 + ], + "angle": 0, + "content": "To address the above challenges, we propose a generalized cross-modal causal transformer (GOAT) approach that enables the VLN model to alleviate the negative effects raised by confounders, thus achieving causal decision-making (Fig. 1). Firstly, we propose a unified structural causal model to describe the VLN system, involving two distinct categories of confounders: observable and unobservable. Observable confounders are content-related and easily identifiable (e.g., the keywords in instructions and the room references in environments), whereas unobservable confounders refer to intricate stylistic nuances that are harder to discern but can impact the overall system (e.g., decoration styles in vision, sentence patterns in language, and trajectory trends in history). Then, we propose to address these confounders via two causal learning modules that are based on back-door and front-door adjustments [52] (namely BACL and FACL), respectively. Furthermore, to build global dictionaries for representing confounders, we devise a cross-modal feature pooling (CFP) module to effectively aggregate long-sequential features. Contrastive learning [56] is adopted to optimize CFP, serving as an additional auxiliary task during pre-training. As demonstrated by thorough experiments, our findings reveal the impact of integrating causal learning to deconfound biases on cross-modal inputs, offering valuable insights for enhancing generalization in similar tasks across diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.7, + 0.449, + 0.714 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.72, + 0.468, + 0.763 + ], + "angle": 0, + "content": "- We propose a unified structural causal model for VLN by comprehensively considering the observable and unobservable confounders hidden in different modalities." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.766, + 0.468, + 0.809 + ], + "angle": 0, + "content": "- we propose BACL and FACL, using the back-door and front-door adjustments to allow end-to-end unbiased cross-modal intervention and decision-making." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.853 + ], + "angle": 0, + "content": "- We propose CFP, a cross-modal feature pooling module designed to aggregate sequence features for semantic alignment and confounder dictionaries construction." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Our GOAT model demonstrates exceptional generalization across multiple VLN datasets (R2R [4], RxR [29], REVERIE [53], and SOON [81]), outperforming existing" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "state-of-the-art methods. A comprehensive causal learning pipeline is presented to inspire future research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.135, + 0.642, + 0.15 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Vision-and-Language Navigation (VLN) requires agents to navigate to specific locations [4, 29] or find target objects [53, 81] in real visual environments based on natural instructions. Its practicability has led to significant interest, showing potential in fundamental embodied AI skills. Initial models relied on recurrent neural networks [2, 11, 21, 68]. Transformer-based models [8, 9, 19, 22, 63] brought substantial progress due to their powerful long-distance encoding. However, small-scale datasets in VLN were found to cause bias, leading to serious overfitting. Consequently, several approaches were devised to tackle this challenge. Speaker-follower frameworks [16, 27, 58, 62, 64] used the speaker model to generate pseudo instructions. VLN-BERT [44], AirBERT [18], and Lily [36] collected trajectory-instruction pairs from diverse sources. REM [39] and EnvEdit [34] created new environments by editing existing environments. Although these methods improve generalization, they cannot completely eliminate inherent dataset biases. Therefore, we propose to develop an unbiased model by equipping the agent with the causal inference capability to learn cause-effect relations during data fitting, enabling them to adapt adeptly to diverse situations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Causal Inference is an emerging technique exploring task causality [52], leading to a surge in efforts to integrate it with deep learning, in tasks like image recognition [67, 69, 76], image captioning [38, 73], and visual question answering [33, 49]. One popular way is to use the adjustment technique to alleviate the negative effects caused by confounders, and some studies exploring the use of counterfactuals [1, 49, 50]. This paper emphasizes the adjustment method due to its practicality. However, most of the existing causal learning tasks are simple without considering more challenging tasks like VLN. Additionally, current methods apply back-door [38, 66, 75, 78] or front-door [42, 73, 74] adjustments separately across modalities, lacking comprehensive confounder assumptions and complete bias corrections. In this paper, we propose to simultaneously tackle both observable and unobservable confounders in vision, language, and history. This approach significantly reduces overall bias, enhancing the generalization capabilities of embodied VLN agents." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.625, + 0.809 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.673, + 0.832 + ], + "angle": 0, + "content": "3.1. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The VLN task [4] involves an embodied agent following natural language instructions to navigate real indoor environments. Matterport3D simulator [7] is used to allow interaction, where the environment is provided as graphs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13140" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.431, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.239, + 0.454, + 0.253 + ], + "angle": 0, + "content": "Figure 2. Illustration of the structural causal model of VLN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.473, + 0.435 + ], + "angle": 0, + "content": "with connected navigable nodes. The agent receives natural language instructions \\(\\mathcal{I} = \\{w_1, w_2, \\dots, w_L\\}\\) with \\(L\\) words, and the current panorama separated into 36 subimages \\(\\mathcal{V} = \\{v_1, v_2, \\dots, v_{36}\\}\\) [16]. The agent also knows its current heading \\(\\theta\\) and elevation \\(\\phi\\). During navigation, the agent needs to select the next point from nearby candidates or predicts the stop signal based on visual cues. Success is defined when the stop location is within 3 meters of the ground-truth position. For the goal-oriented task, REVERIE [53] and SOON [81] additionally require locating the target object at the final destination." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.369, + 0.466 + ], + "angle": 0, + "content": "3.2. Structural Causal Model of VLN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.779 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, we construct a structural causal model capturing the relationships among the key variables in VLN: visual observation \\(\\mathcal{V}\\), linguistic instruction \\(\\mathcal{I}\\), decision history \\(\\mathcal{H}\\), and action prediction \\(\\mathcal{A}\\). To clarify, we use \\(\\mathcal{X}\\) to denote inputs \\((\\mathcal{X} = \\{\\mathcal{V}, \\mathcal{I}, \\mathcal{H}\\})\\) and \\(\\mathcal{Y}\\) for output \\((\\mathcal{Y} = \\mathcal{A})\\). In this directed acyclic graph, the starting and ending points represent the cause and effect, respectively. Traditional VLN methods focus on learning the observational association \\(P(\\mathcal{Y}|\\mathcal{X})\\), overlooking the ambiguity introduced by confounders \\(\\mathcal{Z}\\) in the back-door path \\(\\mathcal{X} \\gets \\mathcal{Z} \\to \\mathcal{Y}\\). Here, confounders are extraneous variables that influence both causes and effects, e.g., frequently occurring content or specific attributes. \\(\\mathcal{Z} \\rightarrow \\mathcal{X}\\) arises since the combined probability of input data is inevitably affected by the limited resources available in the real world when collection and simulation. Additionally, \\(\\mathcal{Z} \\rightarrow \\mathcal{Y}\\) exists because collected environments, labeled instructions, or sampled trajectories also affect the probability of action distributions. These confounder links enable spurious shortcuts during training but can be detrimental in new situations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We propose to distinguish hidden confounders from different modalities into observable and unobservable categories, enhancing our prior knowledge integration and the rationality of assumption. Concretely, observable confounders encompass instances that can be recognized (e.g., room references \\( z_v^o \\) and guiding keywords \\( z_i^o \\)). In contrast, unobservable confounders consist of intricate patterns and style-related elements that are challenging to qualitatively" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "describe (e.g., decoration style in vision \\( z_v^u \\), sentence pattern in language \\( z_i^u \\), and trajectory trend in history \\( z_h^u \\)). Since we cannot explicitly model unobservable confounders \\( \\mathcal{Z}^u \\), the additional mediators \\( \\mathcal{M} \\) are inserted between \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\) to establish front-door paths \\( \\mathcal{X} \\to \\mathcal{M} \\to \\mathcal{Y} \\). Detailed adjustment methods are introduced in subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.197, + 0.637, + 0.214 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.892, + 0.314 + ], + "angle": 0, + "content": "The overview of the GOAT model is shown in Fig. 3. The proposed back-door and front-door adjustment causal learning modules are detailed in Sec. 4.1 and Sec. 4.2, respectively. The cross-modal feature pooling method is subsequently introduced in Sec. 4.3. Finally, a practical causal learning pipeline is presented in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.323, + 0.76, + 0.339 + ], + "angle": 0, + "content": "4.1. Observable Causal Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.452 + ], + "angle": 0, + "content": "Back-door Adjustment Causal Learning (BACL). Based on Bayes's theorem, the typical observational likelihood is as \\( P(\\mathcal{Y}|\\mathcal{X}) = \\sum_{z}P(\\mathcal{Y}|\\mathcal{X},z)\\underline{P(z|\\mathcal{X})} \\), where \\( P(z|\\mathcal{X}) \\) could bring biased weights. Do-operator [52] provides scientifically sound methods for determining causal effects by severing the back-door link between \\( \\mathcal{Z} \\) and \\( \\mathcal{X} \\). According to the invariance and independence rules [17], we have:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.453, + 0.892, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathcal {Y} | d o (\\mathcal {X})) = \\sum_ {z} P (\\mathcal {Y} | d o (\\mathcal {X}), z) P (z | d o (\\mathcal {X})) (1) \\\\ = \\sum_ {z} P (\\mathcal {Y} | \\mathcal {X}, z) \\underline {{P (z)}} (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.729 + ], + "angle": 0, + "content": "In such case, the intervention is achieved by blocking the back-door path \\(\\mathcal{Z} \\to \\mathcal{X}\\), making \\(\\mathcal{X}\\) have a fair opportunity to incorporate causality-related factors for prediction. Previous methods [42, 73, 78] used the NWGM approximation [6] to directly pursue causal learning to the final outputs. However, these methods limit the intervention only to the network's final Softmax layer, overlooking possible biased features in shallow layers. Since the conditional probability is implicit in pattern recognition made by the trained neural network [48], we release the target effect of causal hypothesis to learned features rather than merely outputs. Obtaining unbiased features leads to unbiased predictions. Consequently, the specific network module is formulated as \\(f(x, z)\\), and the implementation of Eq. (2) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.73, + 0.891, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} (\\boldsymbol {x}, \\boldsymbol {z}) = \\mathbb {E} _ {z} [ f (\\boldsymbol {x}, \\boldsymbol {z}) ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.808 + ], + "angle": 0, + "content": "The linear function is used as \\( f(\\pmb{x}, \\pmb{z}) = f_{x}(\\pmb{x}) + f_{z}(\\pmb{z}) \\). Then Eq. (3) becomes \\( f_{x}(\\pmb{x}) + \\mathbb{E}_{z}[f_{z}(\\pmb{z})] \\). To obtain \\( \\mathbb{E}_{z}[f_{z}(\\pmb{z})] \\), there are two prevalent approaches: statistic-based [42, 66] and attention-based methods [38, 73]:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.809, + 0.892, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\text {S t a t i s t i c}: \\mathbb {E} _ {z} [ f _ {z} (\\boldsymbol {z}) ] = \\sum_ {i} \\frac {\\left| z _ {i} \\right|}{\\sum_ {j} \\left| z _ {j} \\right|} f _ {z} \\left(\\boldsymbol {z} _ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.847, + 0.892, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\text {A t t e n t i o n}: \\mathbb {E} _ {z} [ f _ {z} (z) ] = \\sum_ {i} \\frac {\\exp \\left(h z _ {i} ^ {T}\\right)}{\\sum_ {j} \\exp \\left(h z _ {j} ^ {T}\\right)} f _ {z} (z _ {i}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(|z_{i}|\\) denotes the number of \\(z\\) belonging to the \\(i\\)-th" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13141" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.885, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Figure 3. Framework of GOAT, built on the foundation of the dual-scale graph transformer [9]. Back-door and front-door adjustment causal learning mechanisms are used for mitigating spurious correlations, enabling unbiased feature learning and decision-making." + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.291, + 0.353, + 0.303 + ], + "angle": 0, + "content": "(a) Back-door Adjustment Causal Learning (BACL)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.302, + 0.468, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.387, + 0.356, + 0.398 + ], + "angle": 0, + "content": "(b) Front-door Adjustment Causal Learning (FACL)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.398, + 0.468, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.484, + 0.399, + 0.497 + ], + "angle": 0, + "content": "Figure 4. Illustration of BACL and FACL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.557 + ], + "angle": 0, + "content": "category in the training set, and \\( h \\) means hidden features. The illustration is shown in Fig. 4(a). Our experiments in Sec. 5.3 reveal modality-specific calculation preferences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.802 + ], + "angle": 0, + "content": "BACL in Text Content. In VLN instructions like \"Exit the office and turn right into the kitchen,\" essential guiding elements such as directions (e.g., \"exit\" and \"right\") and landmarks (e.g., \"office\" and \"kitchen\") play significant roles. These keywords which are common causes of instruction construction and action distribution, serve as observable confounders. Firstly, we build the text keyword dictionary \\(\\mathcal{D}_I^O = [z_{i,1}^o,z_{i,2}^o,\\dots,z_{i,K}^o ]\\) with \\(K\\) classes to store confounder features. Direction-and-landmark keywords are extracted based on their part-of-speech tags [63]. We use the pre-trained RoBERTa [41] to obtain feature representations for each extracted token \\(\\pmb{f}_{i}^{o}\\). Since the same word can have different features across sentences, we calculate the average feature for each keyword: \\(\\pmb{z}_{i,n}^{o} = \\frac{1}{|z_{i,n}^{o}|}\\sum_{j}\\pmb{f}_{i,n,j}^{o}\\). Subsequently, the text content causal representation \\(G_{s}\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.804, + 0.468, + 0.822 + ], + "angle": 0, + "content": "\\[\nI = \\operatorname {R o B E R T a} \\left(\\psi_ {t} (\\mathcal {I}) + \\psi_ {t} (\\mathcal {P})\\right), Z _ {k} = \\mathrm {L N} \\left(\\phi_ {k} \\left(\\mathcal {D} _ {I} ^ {O}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.824, + 0.468, + 0.84 + ], + "angle": 0, + "content": "\\[\nG _ {s} = \\operatorname {L N} \\left[ \\phi_ {i} \\left(\\mathcal {B} \\left(I, Z _ {k}\\right)\\right) \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot)\\) and \\(\\phi (\\cdot)\\) denote the learnable embedding layer and the full-connection layer, respectively. The absolute positional encoding \\(\\mathcal{P}\\) [60] is added to present the position information, and the layer normalization LN [5] is employed" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.791, + 0.308 + ], + "angle": 0, + "content": "for stabilizing hidden states during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.595 + ], + "angle": 0, + "content": "BACL in Vision Content. For each step, the panorama \\(\\mathcal{V}\\) is divided into 36 sub-images. Since existing VLN datasets primarily involve indoor room navigation, visual room references are treated as observable confounders. CLIP [56] is used to extract image features. Since room labels are not directly provided, we employ BLIP [32], a pre-trained VQA model to capture the room information for each image, by querying the model with a fixed prompt \"what kind of room is this?\" The average value of each room reference type is calculated, forming a visual room reference dictionary \\(\\mathcal{D}_V^O = [z_{v,1}^o,z_{v,2}^o,\\dots,z_{v,M}^o]\\), where \\(M\\) is the number of room types. Additionally, the matrix \\(\\gamma = \\left[(\\sin \\theta_i,\\cos \\theta_i,\\sin \\eta_i,\\cos \\eta_i)_{i = 1}^{36}\\right]\\) is used to present the direction of each image's shift relative to the agent, where \\(\\theta\\) and \\(\\eta\\) denote the heading and elevation direction. If there are additional object features (for goal-oriented tasks), they are concatenated with image features. Subsequently, a 2-layer transformer encoder is used to capture spatial dependencies. The above process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.596, + 0.891, + 0.614 + ], + "angle": 0, + "content": "\\[\nV = \\operatorname {C L I P} (\\mathcal {V}), Z _ {r} = \\ln \\left(\\phi_ {r} \\left(\\mathcal {D} _ {V} ^ {O}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.616, + 0.891, + 0.632 + ], + "angle": 0, + "content": "\\[\nV _ {v} = \\operatorname {L N} \\left[ \\phi_ {v} \\left(\\mathcal {B} (V, Z _ {r})\\right) \\right] \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.635, + 0.891, + 0.652 + ], + "angle": 0, + "content": "\\[\nQ _ {v} = \\operatorname {T r a n s} \\left(V _ {v} + \\psi_ {d} (\\gamma)\\right) \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.779, + 0.68 + ], + "angle": 0, + "content": "4.2. Unobservable Causal Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Front-door Adjustment Causal Learning (FACL). In the previous section, we employed the back-door adjustment technique to handle bias caused by observable confounders. However, there are additional unobservable confounders that cannot be explicitly captured and modeled in advance. To address this, we introduce another technique - front-door adjustment [17]. As shown in Fig. 4(b), an additional mediator \\(\\mathcal{M}\\) is inserted between inputs and outcomes to construct the front door path \\(\\mathcal{X} \\rightarrow \\mathcal{M} \\rightarrow \\mathcal{Y}\\). In VLN, an attention-based model \\(P(\\mathcal{Y}|\\mathcal{X}) = \\sum_{m} P(\\mathcal{Y}|m)P(m|\\mathcal{X})\\) will select key regions \\(\\mathcal{M}\\) from inputs \\(\\mathcal{X}\\) for action prediction \\(\\mathcal{Y}\\). Therefore, the model inference can be represented by two parts: the feature selector \\(\\mathcal{X} \\rightarrow \\mathcal{M}\\) which selects suitable knowledge \\(\\mathcal{M}\\) from \\(\\mathcal{X}\\), and the action predictor" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "\\(\\mathcal{M} \\to \\mathcal{V}\\) which exploits \\(\\mathcal{M}\\) to predict \\(\\mathcal{V}\\). To eliminate spurious correlation brought by unobservable confounder \\(\\mathcal{Z}\\), we simultaneously deploy \\(do\\)-operator to \\(\\mathcal{X}\\) and \\(\\mathcal{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.137, + 0.47, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathcal {Y} | d o (\\mathcal {X})) = \\sum_ {m} P (\\mathcal {Y} | d o (m)) P (m | d o (\\mathcal {X})) (11) \\\\ = \\sum_ {x ^ {\\prime}} P \\left(x ^ {\\prime}\\right) \\sum_ {m} P (\\mathcal {Y} | m, x ^ {\\prime}) P (m | \\mathcal {X}) (12) \\\\ = \\mathbb {E} _ {x ^ {\\prime}} \\mathbb {E} _ {m | x} [ P (\\mathcal {Y} | x ^ {\\prime}, m) ] (13) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.223, + 0.469, + 0.417 + ], + "angle": 0, + "content": "where \\( x' \\) denotes potential input samples of the whole representation space, different from current inputs \\( \\mathcal{X} = x \\). We use the bold symbol \\( \\pmb{m} \\) to denote the in-sampling features obtained by the feature extractor acting on the current input, and \\( \\pmb{x}' \\) to mean the cross-sampling features randomly sampled by the K-means-based feature selector from the entire training samples. Based on the linear mapping model, Eq. (13) becomes \\( \\mathbb{E}_{m|x}[m] + \\mathbb{E}_{x'}[x'] \\). As it is intractable to get a closed-form solution of expectations involving the complex representation space, the estimation is achieved by the query mechanism. Two embedding functions [42, 73] are used to transmit input \\( \\pmb{x} \\) into two query sets \\( g_{1} = q_{1}(\\pmb{x}) \\) and \\( g_{2} = q_{2}(\\pmb{x}) \\). Then, the front-door adjustment is approximated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.424, + 0.469, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {x} ^ {\\prime}} \\left[ \\boldsymbol {x} ^ {\\prime} \\right] \\approx \\sum_ {\\boldsymbol {x} ^ {\\prime}} P \\left(\\boldsymbol {x} ^ {\\prime} \\mid \\boldsymbol {g} _ {1}\\right) \\boldsymbol {x} ^ {\\prime} = \\sum_ {i} \\frac {\\exp \\left(\\boldsymbol {g} _ {1} \\boldsymbol {x} _ {i} ^ {\\prime T}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {g} _ {1} \\boldsymbol {x} _ {j} ^ {\\prime T}\\right)} \\boldsymbol {x} _ {i} ^ {\\prime} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.462, + 0.469, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {m \\mid x} [ \\boldsymbol {m} ] \\approx \\sum_ {m} P (\\boldsymbol {m} | \\boldsymbol {g} _ {2}) \\boldsymbol {m} = \\sum_ {i} \\frac {\\exp \\left(\\boldsymbol {g} _ {2} \\boldsymbol {m} _ {i} ^ {T}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {g} _ {2} \\boldsymbol {m} _ {j} ^ {T}\\right)} \\boldsymbol {m} _ {i} (15) \\\\ \\mathcal {F} \\left(\\boldsymbol {x}, \\boldsymbol {x} ^ {\\prime}\\right) = \\mathbb {E} _ {\\boldsymbol {x} ^ {\\prime}} \\left[ \\boldsymbol {x} ^ {\\prime} \\right] + \\mathbb {E} _ {m | x} [ \\boldsymbol {m} ] (16) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.526, + 0.47, + 0.584 + ], + "angle": 0, + "content": "The above process can be efficiently implemented using the multi-head attention [60], enabling seamless integration of causal adjustments into existing transformer-based frameworks with minimal modifications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.585, + 0.469, + 0.822 + ], + "angle": 0, + "content": "FACL in Text, Vision, and History. Considering the characteristics of VLN, we propose to eliminate unobservable confounders from three kinds of inputs in VLN, i.e., vision, language, and history. First, following previous graph-based methods [9, 63], we construct the vision sequence \\( G_{v} = \\{[\\mathrm{STOP}], [\\mathrm{MEM}], Q_{v}\\} \\) and the history sequence \\( G_{h} = \\{[\\mathrm{STOP}], [\\mathrm{MEM}], \\{\\bar{Q}_{t}\\}_{t=1}^{T}\\} \\) by adding the additional token for presenting stop and recurrent memory states, respectively. \\( \\bar{Q}_{t} \\) means the learned weight sum of panoramic features for the \\( t \\)-th step. To condense the lengthy sequence of features and generate global features \\( x' \\) for cross-sampling, we devise the CFP module (as described in Sec. 4.3) with the attentive pooling mechanism to construct confounder dictionaries for vision, history, and instruction, denoted as \\( D_{v}^{U}, D_{h}^{U} \\), and \\( D_{s}^{U} \\), respectively. Then the causality-enhanced features \\( R_{v}, R_{h} \\) and \\( R_{s} \\) are calculated based on Eq. (16):" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.83, + 0.469, + 0.86 + ], + "angle": 0, + "content": "\\[\nR _ {v} = \\mathcal {F} \\left(G _ {v}, D _ {v} ^ {U}\\right), R _ {h} = \\mathcal {F} \\left(G _ {h}, D _ {h} ^ {U}\\right), R _ {s} = \\mathcal {F} \\left(G _ {s}, D _ {s} ^ {U}\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we introduce an adaptive gate fusion (AGF) method to enhance the stability of learning by integrating" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.198, + 0.882, + 0.213 + ], + "angle": 0, + "content": "Figure 5. Illustration of the cross-modal feature pooling (CFP)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.227, + 0.89, + 0.256 + ], + "angle": 0, + "content": "causality-enhanced features with the original context features for each modality:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.258, + 0.89, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\omega_ {x} = \\delta \\left(R _ {x} W _ {x} + G _ {x} W _ {g} + b\\right) \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.277, + 0.89, + 0.293 + ], + "angle": 0, + "content": "\\[\nU _ {x} = \\omega_ {x} \\odot R _ {x} + (1 - \\omega_ {x}) \\odot G _ {x} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.399 + ], + "angle": 0, + "content": "where \\(\\delta\\) and \\(\\odot\\) mean the Sigmoid function and elementwise multiplication. Suppose \\(R_{x}\\) and \\(G_{x} \\in \\mathbb{R}^{L_{x} \\times d_{h}}\\), then \\(W_{x/g} \\in \\mathbb{R}^{d_{h} \\times 1}\\) and \\(b \\in \\mathbb{R}^{L_{x} \\times 1}\\) are learnable parameters. Next, the cross-modal fused local features \\(F_{l}\\) and global features \\(F_{g}\\) are obtained by the cross-attention encoders \\(\\mathcal{C}\\) from METER [15]. The dynamic fusion \\(\\mathcal{DF}\\) [9] followed by Softmax \\(\\mathcal{SF}\\) is applied for action prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.4, + 0.89, + 0.417 + ], + "angle": 0, + "content": "\\[\nF _ {l} = \\mathcal {C} \\left(U _ {v}, U _ {s}, U _ {s}\\right), F _ {g} = \\mathcal {C} \\left(U _ {h}, U _ {s}, U _ {s}\\right) \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.419, + 0.89, + 0.435 + ], + "angle": 0, + "content": "\\[\nF _ {f} = \\mathcal {D F} \\left(F _ {l}, F _ {g}\\right), y _ {t} = \\mathcal {S F} \\left(F _ {f}\\right) \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.436, + 0.89, + 0.451 + ], + "angle": 0, + "content": "The cross-entropy loss is used to optimize the network:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.463, + 0.89, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = \\sum_ {t = 1} ^ {T} - \\log P \\left(y _ {t} ^ {*} \\mid \\mathcal {I}, \\mathcal {V} _ {t}, \\mathcal {H} _ {1: t - 1}\\right) \\tag {22}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.76, + 0.527 + ], + "angle": 0, + "content": "4.3. Cross-modal Feature Pooling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.639 + ], + "angle": 0, + "content": "One challenge of implementing the front-door adjustment in VLN is constructing efficient dictionaries for global features from long sequences. This requires compressing sequential features of varying lengths into a unified feature space to represent each sample effectively. Let \\( H \\in \\mathbb{R}^{L \\times d_h} \\) be the sequential features, the following attentive pooling is used to effectively compress the sequence length:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.64, + 0.89, + 0.657 + ], + "angle": 0, + "content": "\\[\nA = \\mathcal {T} (H), \\alpha = \\mathcal {S F} \\left(A W _ {a}\\right), \\bar {H} = \\mathcal {T} \\left(\\alpha^ {T} H\\right) \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.811 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}\\) denote the Tanh activation, \\(W_{a} \\in \\mathbb{R}^{d_{h} \\times 1}\\) is the learnable attention matrix, and \\(\\bar{H} \\in \\mathbb{R}^{1 \\times d_{h}}\\). As shown in Fig. 5, for vision, history, local-global fusion, and text features, we use one transformer layer as the encoder followed by the attentive pooling to obtain flattened features \\(\\bar{C}^{l}, \\bar{C}^{g}, \\bar{C}^{f}\\), and \\(\\bar{C}^{s}\\), respectively. Then, we adopt contrastive learning [35, 57, 62] to optimize this cross-modal feature pooling (CFP) module, meanwhile improving semantic alignments for different modalities. The contrastive loss \\(\\mathcal{L}_{ls}\\) is constructed as:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.817, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {l s} = - \\frac {1}{2 B} \\sum_ {j = 1} ^ {B} \\log \\frac {\\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {j} ^ {s} \\right\\rangle / t\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)} \\tag {24} \\\\ - \\frac {1}{2 B} \\sum_ {k = 1} ^ {B} \\log \\frac {\\exp \\left(\\left\\langle \\bar {C} _ {k} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "where \\(B\\) and \\(t\\) mean the batch size and temperature, respectively. Similarly, contrastive losses \\(\\mathcal{L}_{gs}\\) and \\(\\mathcal{L}_{fs}\\) are calculated by replacing \\(\\bar{C}^l\\) with \\(\\bar{C}^g\\) and \\(\\bar{C}^f\\). The overall CFP loss is the sum of these losses \\(\\mathcal{L}_{\\mathrm{CFP}} = \\mathcal{L}_{ls} + \\mathcal{L}_{gs} + \\mathcal{L}_{fs}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.32 + ], + "angle": 0, + "content": "To enable the network more adaptive to characteristics for VLN and thus facilitate the building of the front-door confounder dictionaries for samples, we train the CFP module alongside other auxiliary tasks [9] during pre-training. Subsequently, the trained attentive pooling modules are used to extract global features for different modalities. In the fine-tuning stage, we employ the BACL and FACL with established dictionaries for intervention. The CFP offers dual benefits: it aligns different modalities more effectively during pre-training and provides a systematic approach for extracting coherent representations from sequence inputs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.33, + 0.308, + 0.346 + ], + "angle": 0, + "content": "4.4. Causal Learning Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.281, + 0.625 + ], + "angle": 0, + "content": "As shown in Fig. 6, we summarize a causal learning pipeline that serves as a blueprint for similar learning-based methods. First, it begins with task formulation, where the specific task and its objectives are precisely defined. Next, the observable and unobservable confounders are explicitly assumed. Both back-door and front-door adjustment strategies are employed to tackle these confounders, either simultaneously or sequentially, contingent on task specifics. Then, the pipeline proceeds" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.369, + 0.47, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.584, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Figure 6. Pipeline of the causal learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.716 + ], + "angle": 0, + "content": "to model calculation and result prediction. Throughout network optimization, both network parameters and confounder features are continuously updated. Ultimately, this iterative process leads us to the development of a robust causal model capable of generating unbiased features, thereby advancing the generalizability of AI systems." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.21, + 0.749 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.285, + 0.773 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "1) Datasets. We verify GOAT on two kinds of VLN benchmarks: fine-grained datasets (R2R [4] and RxR-English [29]), which provide long step-by-step navigation instructions, and goal-oriented datasets (REVERIE [53] and SOON [81]), which additionally requires for the target object. Formally, the datasets are partitioned into four splits: training, validation seen (sharing the same environments with the training set), validation unseen (having different" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "environments from the training set), and test unseen sets (reported by the online leaderboard for fair comparison)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.228 + ], + "angle": 0, + "content": "2) Evaluation Metrics. In R2R, key metrics include Navigation Error (NE), Success Rate (SR), Oracle SR (OSR), and SR Weighted by Path Length (SPL). RxR adds Normalized Dynamic Time Warping (nDTW) and SR Weighted by Dynamic Time Warping (sDTW). REVERIE and SOON introduce Remote Grounding Success Rate (RGS) and RGS Weighted by Path Length (RGSPL)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.56 + ], + "angle": 0, + "content": "3) Implementation Details. Our model consists of 6 transformer layers for text, 2 for panorama, and 3 for cross-modal encoding. We use CLIP-B/16 [56] for image feature extraction and initialize network weights with METER [15]. In pre-training, MLM [28], SAP [8], and the proposed CFP are for R2R and RxR. OG [37] is added for REVERIE and SOON. EnvEdit [34] is employed for feature augmentation. The synthetic extended datasets [19, 62, 65] are used for R2R, REVERIE, and RxR, respectively. Pretraining is done on a single Tesla V100 GPU for a maximum of 300K iterations by the AdamW [43] optimizer, with a batch size of 48 and a learning rate of \\(5 \\times 10^{-5}\\). The numbers of classes of keywords and rooms are 74 and 50, and the temperature \\(t\\) is set to 1. In fine-tuning, the front-door dictionaries are randomly sampled from the K-Means clustering features. As the text transformer RoBERTa [41] is involved in end-to-end training, the textual keywords dictionary is also iteratively updated. Speaker models with the environmental dropout [58, 62, 64] are used to provide dynamic pseudo labels. We employ batch size 12 for R2R, REVERIE, and 5 for RxR and SOON, with a learning rate of \\(2 \\times 10^{-5}\\) and a maximum of 100K iterations." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.571, + 0.813, + 0.587 + ], + "angle": 0, + "content": "5.2. Comparisons with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.822 + ], + "angle": 0, + "content": "In Tab. 1, 2, 3, 4, we compare GOAT with the previous state-of-the-art (SoTA) methods on the R2R, REVERIE, RxR-English, and SOON datasets, respectively. On all these four datasets, our approach exhibits superior navigation performance, precise instruction-following alignment, and accurate object grounding across both seen and unseen environments. For instance, in R2R, GOAT achieves remarkable improvements in SPL compared to BEVBert [3], with relative increases of \\(7.41\\%\\), \\(6.45\\%\\), and \\(4.74\\%\\) on three subsets. In REVERIE, GOAT shows substantial relative enhancements in RGSPL by \\(11.83\\%\\), \\(6.55\\%\\), and \\(20.87\\%\\) on three subsets. In the challenging SOON and RxR tasks, GOAT also exhibits significant improvements in performance metrics, highlighting its robustness and superior generalization capabilities over previous methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.704, + 0.848 + ], + "angle": 0, + "content": "5.3.Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1) Effect of Causal Inference. Fig. 7 verifies the impact of causal inference on GOAT across four diverse VLN datasets in unseen environments. \"W/o intervention\" signifies the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13144" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.852, + 0.254 + ], + "angle": 0, + "content": "
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑
HAMT [8]76722.518266613.297365603.9372
DUET [9]79732.288672603.318169593.6576
TD-STP [80]77732.348370633.227667613.7372
GeoVLN [25]79762.22-68633.35-65613.95-
DSRG [63]81762.238873623.008172613.3378
GridMM [71]----75642.83-73623.35-
GELA [10]76732.39-71653.11-67623.59-
EnvEdit [34]77742.32-69643.24-68643.59-
BEVBert [3]81742.178875642.818473623.1381
GOAT (Ours)83.7479.481.7988.6477.8268.132.4084.7274.5764.943.0480.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.169, + 0.255, + 0.798, + 0.269 + ], + "angle": 0, + "content": "Table 1. Comparison with other state-of-the-art methods on the R2R dataset [4]. '-': unavailable statistics." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.283, + 0.88, + 0.412 + ], + "angle": 0, + "content": "
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑
HAMT [8]43.2940.1927.2025.1832.9530.2018.9217.2830.4026.6714.8813.08
HOP+ [54]55.8749.5540.7636.2236.0731.1322.4919.3333.8228.2420.2016.86
DUET [9]71.7563.9457.4151.1446.9833.7332.1523.0352.5136.0631.8822.06
DSRG [63]75.6968.0961.0754.7247.8334.0232.6923.3754.0437.0932.4922.18
GridMM [71]----51.3736.4734.5724.5653.1336.6034.8723.45
BEVBert [3]73.7265.3257.7051.7351.7836.3734.7124.4452.8136.4132.0622.09
GOAT (Ours)78.6471.4063.7457.8553.3736.7038.4326.0957.7240.5338.3226.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.414, + 0.819, + 0.428 + ], + "angle": 0, + "content": "Table 2. Comparison with other state-of-the-art methods on the REVERIE dataset [53]. '-': unavailable statistics." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.452, + 0.469, + 0.577 + ], + "angle": 0, + "content": "
MethodValidation SeenValidation Unseen
SR↑SPL↑nDTW↑sDTW↑SR↑SPL↑nDTW↑sDTW↑
Syntax [31]48.144.058.040.039.235.052.032.0
SOAT [45]----44.2-54.836.4
HOP+ [54]53.647.959.043.045.738.452.036.0
FOAM [14]----42.838.754.135.6
ADAPT [35]50.344.656.340.646.940.254.137.7
\\( \\mathsf{MAR}_{\\mathsf{M-MP}} \\) [27]----50.2-60.343.9
VLN-PETL [55]60.556.865.751.757.954.264.949.7
GOAT (Ours)74.168.171.061.468.261.767.156.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.579, + 0.434, + 0.593 + ], + "angle": 0, + "content": "Table 3. Comparison on the RxR-English dataset [29]." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.695 + ], + "angle": 0, + "content": "
MethodValidation UnseenTest Unseen
OSR↑SR↑SPL↑RGSPL↑OSR↑SR↑SPL↑RGSPL↑
GBE [81]28.5419.5213.341.1621.4512.909.230.45
DUET [9]50.9136.2822.583.7543.0033.4421.424.17
GridMM [71]53.3937.4624.813.9148.0236.2721.254.15
GOAT (Ours)54.6940.3528.055.8550.6340.5025.186.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.697, + 0.416, + 0.712 + ], + "angle": 0, + "content": "Table 4. Comparison on the SOON dataset [81]." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.73, + 0.467, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.842, + 0.453, + 0.856 + ], + "angle": 0, + "content": "Figure 7. Effect of the intervention on various VLN datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "exclusion of the proposed BACL and FACL interventions across all modalities. On each of the datasets examined, the" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.452, + 0.875, + 0.527 + ], + "angle": 0, + "content": "
IdBACLFACLSR↑SPL↑NE↓OSR↑
1XX75.2764.692.7283.14
2✓X76.3767.132.6083.99
3X✓76.5066.922.5384.21
4✓✓77.8268.132.4084.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.53, + 0.863, + 0.544 + ], + "angle": 0, + "content": "Table 5. Effect of back-door and front-door adjustments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.619 + ], + "angle": 0, + "content": "integration of causal inference leads to significant enhancements in the model's performance. This strongly demonstrates causal learning's considerable popularization potential in enhancing learning-based model generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.79 + ], + "angle": 0, + "content": "2) Effect of BACL and FACL. Tab. 5 analyzes the effects of the proposed BACL and FACL on the R2R val-unseen subset. Compared to the baseline (#1), individual application of either BACL (#2) or FACL (#3) leads to performance improvements. Concurrent use of BACL and FACL (#4) leads to further performance enhancements. These findings underscore our assumption about the presence of both observable and unobservable confounders. Integrating both back-door and front-door adjustments is crucial to comprehensively addressing dataset biases, and enhancing the model's robustness and generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.9 + ], + "angle": 0, + "content": "3) Effect of CFP. In Tab. 6, we assess the efficacy of the proposed CFP on the R2R val-unseen subset. During pretraining (PT), incorporating CFP as an additional auxiliary task (CFP-P) enhances training performance, improving SR and SPL by \\(3.11\\%\\) and \\(3.02\\%\\) (#A1), respectively. In the fine-tuning (FT) stage, we compare the performance with and without the use of trained attention modules from CFP" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13145" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.183 + ], + "angle": 0, + "content": "
StageIdMethodSR↑SPL↑NE↓OSR↑
PTA0w/o CFP-P40.3637.886.5248.87
A1w/ CFP-P43.4740.906.0653.13
FTB0A0 w/o CFP-F75.5665.902.6382.42
B1A1 w/o CFP-F76.6366.172.6384.67
B2A1 w/ CFP-F77.8268.132.4084.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.185, + 0.434, + 0.199 + ], + "angle": 0, + "content": "Table 6. Effect of CFP in pre-training and fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.214, + 0.449, + 0.29 + ], + "angle": 0, + "content": "
IdTextVisionSR↑SPL↑NE↓OSR↑
1StatsStats75.2264.782.7183.65
2StatsAttn76.5965.392.5685.31
3AttnStats77.8268.132.4084.72
4AttnAttn75.9565.832.6483.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.292, + 0.451, + 0.306 + ], + "angle": 0, + "content": "Table 7. Effect of Statistic and Attention methods in BACL." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.322, + 0.47, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.444, + 0.47, + 0.471 + ], + "angle": 0, + "content": "Figure 8. Comparison of the distribution of SR and SPL of instructions involving specific terms on the R2R val-unseen split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.469, + 0.56 + ], + "angle": 0, + "content": "to extract global features for front-door dictionaries (CFP-F). \"W/o CFP-F\" signifies the use of simple average pooling to compress features from the pre-trained model. #B2 shows that the CFP provides more reliable confounder representations for causal learning (SPL \\(\\uparrow 1.96\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.773 + ], + "angle": 0, + "content": "4) Effect of Different BACL in Different Modalities. Tab. 7 investigates the effect of various combinations of statistic-based and attention-based methods for text and vision in BACL on the R2R val-unseen subset. The results indicate that employing the attention method for text and the statistic method for vision yields the best performance (#3). Intuitively, this can be explained by the structured nature of textual information and the involvement of RoBERTa's end-to-end training, enabling the attention method to effectively capture contextual nuances. Conversely, images lack explicit causality, and CLIP, the image extractor, isn't trained directly for efficiency reasons. Consequently, the statistic method ensures a stable causal learning process, preserving the integrity of vision-related features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.269, + 0.803 + ], + "angle": 0, + "content": "5.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1) Bias Elimination Effect. In Fig. 8, the compactness of the boxes represents concentrated data distribution and reduced variability, while a median line closer to the center signifies even data distribution. It shows that GOAT obtains narrower boxes and more central midlines across diverse objects. This finding showcases that with the integration of" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.093, + 0.53, + 0.107 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.094, + 0.602, + 0.105 + ], + "angle": 0, + "content": "Starting point" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.094, + 0.635, + 0.107 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.094, + 0.69, + 0.105 + ], + "angle": 0, + "content": "Destination" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.094, + 0.729, + 0.107 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.732, + 0.094, + 0.784, + 0.105 + ], + "angle": 0, + "content": "sited point" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.094, + 0.814, + 0.105 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.826, + 0.094, + 0.883, + 0.105 + ], + "angle": 0, + "content": "visited point" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.108, + 0.53, + 0.122 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.108, + 0.599, + 0.119 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.108, + 0.635, + 0.119 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.108, + 0.662, + 0.119 + ], + "angle": 0, + "content": "DSRG" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.108, + 0.684, + 0.119 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.108, + 0.756, + 0.119 + ], + "angle": 0, + "content": "GOAT (Ours)" + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.125, + 0.599, + 0.136 + ], + "angle": 0, + "content": "- Instruction:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.706, + 0.184 + ], + "angle": 0, + "content": "Walk out of the room and go past the stairs and into the bedroom that is straight ahead, go to the bathroom doorway and wait there." + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.119, + 0.89, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.193, + 0.6, + 0.204 + ], + "angle": 0, + "content": "- Instruction:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.658, + 0.251 + ], + "angle": 0, + "content": "Slightly move over the green chair closest to the corner of the reception desk in the lounge." + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.188, + 0.889, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.26, + 0.862, + 0.275 + ], + "angle": 0, + "content": "Figure 9. Predicted trajectories in unseen environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.333 + ], + "angle": 0, + "content": "causal intervention, GOAT significantly reduces prediction bias, thereby enhancing its generalization capability in previously unseen environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.893, + 0.53 + ], + "angle": 0, + "content": "2) Visualized Trajectories. In Fig. 9, we visualize some predicted trajectories in unseen environments, comparing them with DSRG on R2R and REVERIE datasets. Notably, GOAT precisely captures directional cues like \"straight ahead\" and nuanced instructions like \"closest to the corner\", enabling accurate predictions. These instances highlight the intricate causal connections in VLN tasks, where specific instructions prompt corresponding actions. GOAT's enhanced causal inference capability enables it to generate more reasoned responses aligned with the provided instructions, underscoring the significance of robust causal inference in VLN systems. Please refer to our supplementary material for more detailed discussions and visualizations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.546, + 0.619, + 0.562 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.798 + ], + "angle": 0, + "content": "Our work presents GOAT, a novel approach that addresses the dataset bias in VLN from the perspective of causal learning. The back-door and front-door adjustment causal learning (BACL and FACL) mechanisms are proposed to adjust for observable and unobservable confounders, respectively. The cross-modal feature pooling (CFP) module is adopted to promote feature learning and extraction through contrastive learning. The practical causal learning pipeline is presented to illuminate other similar learning-based methods. Experiments on R2R, REVERIE, RxR, and SOON datasets show that GOAT can reasonably discover sequence visual-linguistic causal structures and significantly improve performance. Beyond VLN, the underlying confounder assumption and causal inference principles are generalizable to other similar fields." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.659, + 0.831 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper is supported by the National Natural Science Foundation of China under Grants (62233013, 62073245, 62173248). Shanghai Science and Technology Innovation Action Plan (22511104900)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13146" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10044-10054, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5101-5109, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[5] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. ArXiv, abs/1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[6] Pierre Baldi and Peter Sadowski. The dropout learning algorithm. Artificial intelligence, 210:78-122, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 5, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[10] Yibo Cui, Liang Xie, Yakun Zhang, Meishan Zhang, Ye Yan, and Erwei Yin. Grounded entity-landmark adaptive pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12043–12053, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[11] Ronghao Dang, Zhuofan Shi, Liuyi Wang, Zongtao He, Chengju Liu, and Qijun Chen. Unbiased directed object attention graph for object navigation. In Proceedings of the 30th ACM International Conference on Multimedia, page 3617-3627, New York, NY, USA, 2022. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Ronghao Dang, Lu Chen, Liuyi Wang, He Zongtao, Chengju Liu, and Qijun Chen. Multiple thinking achieving meta-ability decoupling for object navigation. In International Conference on Machine Learning (ICML), 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[13] Ronghao Dang, Liuyi Wang, Zongtao He, Shuai Su, Jiagui Tang, Chengju Liu, and Qijun Chen. Search for or navigate to? dual adaptive thinking for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8250-8259, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.244 + ], + "angle": 0, + "content": "[14] Zi-Yi Dou and Nanyun Peng. Foam: A follower-aware speaker model for vision-and-language navigation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4332-4340, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[15] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[16] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. Advances in Neural Information Processing Systems, 31, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[17] Madelyn Glymour, Judea Pearl, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[18] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[19] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13137-13146, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[20] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[21] Zongtao He, Liuyi Wang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Mlanet: Multi-level attention network with sub-instruction for continuous vision-and-language navigation. arXiv preprint arXiv:2303.01396, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[22] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[23] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3055-3067, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Ronghang Hu, Daniel Fried, Anna Rohrbach, Dan Klein, Trevor Darrell, and Kate Saenko. Are you looking? grounding to multiple modalities in vision-and-language navigation." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13147" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6551-6557, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[25] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23212-23221, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[26] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1862-1872, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[27] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[28] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[29] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392-4412, 2020. 1, 2, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[30] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[31] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1041–1050, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[33] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21273–21282, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[34] Jialu Li, Hao Tan, and Mohit Bansal. Envideit: Environment editing for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15407-15417, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2022. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[36] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8317-8326, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[37] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7036-7045, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[38] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18041-18050, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[39] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1644-1654, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[40] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[41] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[42] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[44] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In European Conference on Computer Vision, pages 259-274. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[45] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34:7357-7367, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[46] Laksh Nanwani, Anmol Agarwal, Kanishk Jain, Raghav Prabhakar, Aaron Monis, Aditya Mathur, Krishna Murthy, Abdul Hafez, Vineet Gandhi, and K Madhava Krishna. Instance-level semantic maps for vision language navigation. arXiv preprint arXiv:2305.12363, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[47] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13148" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "on Natural Language Processing (EMNLP-IJCNLP), pages 684-695, 2019. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[48] Siqi Nie, Meng Zheng, and Qiang Ji. The deep regression bayesian network and its applications: Probabilistic deep learning for computer vision. IEEE Signal Processing Magazine, 35(1):101-111, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[49] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[50] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Javen Qinfeng Shi, and Anton Van den Hengel. Counterfactual vision-and-language navigation: Unravelling the unseen. Advances in Neural Information Processing Systems, 33:5296-5307, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.334 + ], + "angle": 0, + "content": "[51] Judea Pearl. Causality. Cambridge university press, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.336, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[52] Judea Pearl and Dana Mackenzie. The book of why: the new science of cause and effect. Basic books, 2018. 2, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[53] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 3, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.45, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[54] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[55] Yanyuan Qiao, Zheng Yu, and Qi Wu. Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[57] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? In International Conference on Learning Representations. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[58] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[59] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning, pages 394-406. PMLR, 2020. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.118 + ], + "angle": 0, + "content": "information processing systems, pages 5998-6008, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[61] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15471-15481, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[62] Liuyi Wang, Zongtao He, Ronghao Dang, Huiyi Chen, Chengju Liu, and Qijun Chen. Res-sts: Referring expression speaker via self-training with scorer for goal-oriented vision-language navigation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[63] Liuyi Wang, Zongtao He, jiagui Tang, Ronghao Dang, naijia Wang, Chengju Liu, and Qijun Chen. A dual semantic-aware recurrent global-adaptive network for vision-and-language navigation. In International Joint Conferences on Artificial Intelligence (IJCAI), 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[64] Liuyi Wang, Chengju Liu, Zongtao He, Shu Li, Qingqing Yan, Huiyi Chen, and Qijun Chen. Pasts: Progress-aware spatio-temporal transformer speaker for vision-and-language navigation. arXiv preprint arXiv:2305.11918, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[65] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[66] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10760-10770, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[67] Tan Wang, Chang Zhou, Qianru Sun, and Hanwang Zhang. Causal attention for unbiased visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3091-3100, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[68] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Wang, and Lei Zhang. Vision-language navigation policy learning and adaptation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[69] Yuqing Wang, Xiangxian Li, Zhuang Qi, Jingyu Li, Xuelong Li, Xiangxu Meng, and Lei Meng. Meta-causal feature learning for out-of-distribution generalization. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 530-545. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[70] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12009-12020, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[71] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15625-15636, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13149" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[73] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[74] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[75] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. Advances in neural information processing systems, 33:2734-2746, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.373 + ], + "angle": 0, + "content": "[76] Hua Zhang, Liqiang Xiao, Xiaochun Cao, and Hassan Foroosh. Multiple adverse weather conditions adaptation for object detection via causal intervention. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[77] Jiazhao Zhang, Liu Dai, Fanpeng Meng, Qingnan Fan, Xuelin Chen, Kai Xu, and He Wang. 3d-aware object goal navigation via simultaneous exploration and identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6672-6682, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[78] Shengyu Zhang, Tan Jiang, Tan Wang, Kun Kuang, Zhou Zhao, Jianke Zhu, Jin Yu, Hongxia Yang, and Fei Wu. Devlbert: Learning deconfounded visio-linguistic representations. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4373–4382, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[79] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[80] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4194-4203, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[81] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12699, 2021. 2, 3, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.469, + 0.825 + ], + "angle": 0, + "content": "[82] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13150" + } + ] +] \ No newline at end of file diff --git a/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_origin.pdf b/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0fb1afd2ea1fc461d1a661dc17ed52f4d7fcd32 --- /dev/null +++ b/2024/Vision-and-Language Navigation via Causal Learning/8ed24918-937e-4d68-a4ff-fac9ff1bc6a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f500ee8fc7cdd486ae23de327d90ca6622f983ac132fb18c6591fc9aeea79d56 +size 2606083 diff --git a/2024/Vision-and-Language Navigation via Causal Learning/full.md b/2024/Vision-and-Language Navigation via Causal Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..15f4bb8987027b7f43ceff1701d47772d1b04f73 --- /dev/null +++ b/2024/Vision-and-Language Navigation via Causal Learning/full.md @@ -0,0 +1,435 @@ +# Vision-and-Language Navigation via Causal Learning + +Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu*, Qijun Chen* +School of Electronic and Information Engineering, Tongji University, Shanghai, China + +{wly, xingchen327, dangronghao, liuchengju, qjchen}@tongji.edu.cn, mojo.shum@outlook.com + +# Abstract + +In the pursuit of robust and generalizable environment perception and language understanding, the ubiquitous challenge of dataset bias continues to plague vision-and-language navigation (VLN) agents, hindering their performance in unseen environments. This paper introduces the generalized cross-modal causal transformer (GOAT), a pioneering solution rooted in the paradigm of causal inference. By delving into both observable and unobservable confounders within vision, language, and history, we propose the back-door and front-door adjustment causal learning (BACL and FACL) modules to promote unbiased learning by comprehensively mitigating potential spurious correlations. Additionally, to capture global confounder features, we propose a cross-modal feature pooling (CFP) module supervised by contrastive learning, which is also shown to be effective in improving cross-modal representations during pre-training. Extensive experiments across multiple VLN datasets (R2R, REVERIE, RxR, and SOON) underscore the superiority of our proposed method over previous state-of-the-art approaches. Code is available at https://github.com/CrystalSixone/VLN-GOAT. + +# 1. Introduction + +Effective environment perception, language understanding, and historical utilization are at the core of vision-and-language navigation (VLN) [4]. Despite significant progress, deploying VLN in the real world remains a huge challenge, primarily due to diversities and uncertainties in environments and instructions. A key hindrance is dataset bias [79, 82], e.g., agents may overfit to familiar visual environments, resulting in diminished performance in environments with diverse appearances and layouts [24]. This over-reliance on specific patterns, like biased structural trajectories and repeated entity components, raises concerns about the robustness and generalizability of VLN systems. + +One way to mitigate dataset bias in VLN is to build + +![](images/8705d332ccafaf18cd2753b9850bba1e2643f8af6881359d216b300f587b5720.jpg) +Figure 1. In response to language instructions, the VLN agent is required to navigate to the target based on visual cues. This paper introduces a causal learning pipeline using the do-operator to reduce bias from confounders in VLN action prediction. + +broader and more diverse datasets, which is what numerous recent studies have focused on. These include using speaker models to generate pseudo instructions [14, 16, 61, 62, 64, 65], synthesizing cross-environmental trajectories [26, 39], transferring image styles [30, 34], collecting data from the web [18, 36, 44, 70], and labeling more fine-grained, entity-aligned instructions [10, 21, 29]. However, achieving a perfectly balanced dataset devoid of bias is nearly impossible. Consequently, we often find ourselves caught in a cycle of "creating a dataset" - "identifying bias" - "creating a new dataset". Therefore, we are prompted to move from diagnosis to treatment [73], transitioning from the continuous collection of new datasets to the development of unbiased models that can confront and mitigate bias. + +However, existing methods that focus on model designing mainly concentrate on introducing more types of inputs (e.g., objects [2, 12, 13, 23, 35, 63] and depth [3, 20, 40, 77]), or constructing the global graph [8, 9, 46] to represent environments. These efforts, although valuable, often overlook underlying dataset biases and the essential causal logic behind the task. In fact, the reason why humans can well execute various instructions and navigate in unknown environments is that we can learn the inherent causality of events beyond biased observation, achieving good analogical association capability. Therefore, for the first time, we propose to use causal inference [51] to equip VLN agents with similar cognitive abilities that we have, and then allow + +them to make more reasoned decisions. + +Then, how to develop such a causal inference capability? Although there is no single answer, we propose to exploit the concepts of intervention [52] - technique that uses a do-operator to alleviate the negative effects raised by confounders. Here, confounders are variables that influence both inputs and outcomes, creating spurious correlations and biases. Intervention empowers researchers to mitigate the impact of confounders, enabling the model to grasp the causation of events during data fitting. However, given the fact that VLN is such a complex task that involves cross-modal inputs and a long-term decision-making process, it is challenging to identify underlying confounders and apply intervention to debias through network learning. + +To address the above challenges, we propose a generalized cross-modal causal transformer (GOAT) approach that enables the VLN model to alleviate the negative effects raised by confounders, thus achieving causal decision-making (Fig. 1). Firstly, we propose a unified structural causal model to describe the VLN system, involving two distinct categories of confounders: observable and unobservable. Observable confounders are content-related and easily identifiable (e.g., the keywords in instructions and the room references in environments), whereas unobservable confounders refer to intricate stylistic nuances that are harder to discern but can impact the overall system (e.g., decoration styles in vision, sentence patterns in language, and trajectory trends in history). Then, we propose to address these confounders via two causal learning modules that are based on back-door and front-door adjustments [52] (namely BACL and FACL), respectively. Furthermore, to build global dictionaries for representing confounders, we devise a cross-modal feature pooling (CFP) module to effectively aggregate long-sequential features. Contrastive learning [56] is adopted to optimize CFP, serving as an additional auxiliary task during pre-training. As demonstrated by thorough experiments, our findings reveal the impact of integrating causal learning to deconfound biases on cross-modal inputs, offering valuable insights for enhancing generalization in similar tasks across diverse scenarios. + +To summarize, our main contributions are as follows: + +- We propose a unified structural causal model for VLN by comprehensively considering the observable and unobservable confounders hidden in different modalities. +- we propose BACL and FACL, using the back-door and front-door adjustments to allow end-to-end unbiased cross-modal intervention and decision-making. +- We propose CFP, a cross-modal feature pooling module designed to aggregate sequence features for semantic alignment and confounder dictionaries construction. +- Our GOAT model demonstrates exceptional generalization across multiple VLN datasets (R2R [4], RxR [29], REVERIE [53], and SOON [81]), outperforming existing + +state-of-the-art methods. A comprehensive causal learning pipeline is presented to inspire future research. + +# 2. Related Work + +Vision-and-Language Navigation (VLN) requires agents to navigate to specific locations [4, 29] or find target objects [53, 81] in real visual environments based on natural instructions. Its practicability has led to significant interest, showing potential in fundamental embodied AI skills. Initial models relied on recurrent neural networks [2, 11, 21, 68]. Transformer-based models [8, 9, 19, 22, 63] brought substantial progress due to their powerful long-distance encoding. However, small-scale datasets in VLN were found to cause bias, leading to serious overfitting. Consequently, several approaches were devised to tackle this challenge. Speaker-follower frameworks [16, 27, 58, 62, 64] used the speaker model to generate pseudo instructions. VLN-BERT [44], AirBERT [18], and Lily [36] collected trajectory-instruction pairs from diverse sources. REM [39] and EnvEdit [34] created new environments by editing existing environments. Although these methods improve generalization, they cannot completely eliminate inherent dataset biases. Therefore, we propose to develop an unbiased model by equipping the agent with the causal inference capability to learn cause-effect relations during data fitting, enabling them to adapt adeptly to diverse situations. + +Causal Inference is an emerging technique exploring task causality [52], leading to a surge in efforts to integrate it with deep learning, in tasks like image recognition [67, 69, 76], image captioning [38, 73], and visual question answering [33, 49]. One popular way is to use the adjustment technique to alleviate the negative effects caused by confounders, and some studies exploring the use of counterfactuals [1, 49, 50]. This paper emphasizes the adjustment method due to its practicality. However, most of the existing causal learning tasks are simple without considering more challenging tasks like VLN. Additionally, current methods apply back-door [38, 66, 75, 78] or front-door [42, 73, 74] adjustments separately across modalities, lacking comprehensive confounder assumptions and complete bias corrections. In this paper, we propose to simultaneously tackle both observable and unobservable confounders in vision, language, and history. This approach significantly reduces overall bias, enhancing the generalization capabilities of embodied VLN agents. + +# 3. Preliminary + +# 3.1. Task Formulation + +The VLN task [4] involves an embodied agent following natural language instructions to navigate real indoor environments. Matterport3D simulator [7] is used to allow interaction, where the environment is provided as graphs + +![](images/c4d24ae1254ed429bf53c78f1708a40b95a7a6c4f5935dee0fc2107b6c8e846b.jpg) +Figure 2. Illustration of the structural causal model of VLN. + +with connected navigable nodes. The agent receives natural language instructions $\mathcal{I} = \{w_1, w_2, \dots, w_L\}$ with $L$ words, and the current panorama separated into 36 subimages $\mathcal{V} = \{v_1, v_2, \dots, v_{36}\}$ [16]. The agent also knows its current heading $\theta$ and elevation $\phi$ . During navigation, the agent needs to select the next point from nearby candidates or predicts the stop signal based on visual cues. Success is defined when the stop location is within 3 meters of the ground-truth position. For the goal-oriented task, REVERIE [53] and SOON [81] additionally require locating the target object at the final destination. + +# 3.2. Structural Causal Model of VLN + +As illustrated in Fig. 2, we construct a structural causal model capturing the relationships among the key variables in VLN: visual observation $\mathcal{V}$ , linguistic instruction $\mathcal{I}$ , decision history $\mathcal{H}$ , and action prediction $\mathcal{A}$ . To clarify, we use $\mathcal{X}$ to denote inputs $(\mathcal{X} = \{\mathcal{V}, \mathcal{I}, \mathcal{H}\})$ and $\mathcal{Y}$ for output $(\mathcal{Y} = \mathcal{A})$ . In this directed acyclic graph, the starting and ending points represent the cause and effect, respectively. Traditional VLN methods focus on learning the observational association $P(\mathcal{Y}|\mathcal{X})$ , overlooking the ambiguity introduced by confounders $\mathcal{Z}$ in the back-door path $\mathcal{X} \gets \mathcal{Z} \to \mathcal{Y}$ . Here, confounders are extraneous variables that influence both causes and effects, e.g., frequently occurring content or specific attributes. $\mathcal{Z} \rightarrow \mathcal{X}$ arises since the combined probability of input data is inevitably affected by the limited resources available in the real world when collection and simulation. Additionally, $\mathcal{Z} \rightarrow \mathcal{Y}$ exists because collected environments, labeled instructions, or sampled trajectories also affect the probability of action distributions. These confounder links enable spurious shortcuts during training but can be detrimental in new situations. + +We propose to distinguish hidden confounders from different modalities into observable and unobservable categories, enhancing our prior knowledge integration and the rationality of assumption. Concretely, observable confounders encompass instances that can be recognized (e.g., room references $z_v^o$ and guiding keywords $z_i^o$ ). In contrast, unobservable confounders consist of intricate patterns and style-related elements that are challenging to qualitatively + +describe (e.g., decoration style in vision $z_v^u$ , sentence pattern in language $z_i^u$ , and trajectory trend in history $z_h^u$ ). Since we cannot explicitly model unobservable confounders $\mathcal{Z}^u$ , the additional mediators $\mathcal{M}$ are inserted between $\mathcal{X}$ and $\mathcal{Y}$ to establish front-door paths $\mathcal{X} \to \mathcal{M} \to \mathcal{Y}$ . Detailed adjustment methods are introduced in subsequent sections. + +# 4. Methodology + +The overview of the GOAT model is shown in Fig. 3. The proposed back-door and front-door adjustment causal learning modules are detailed in Sec. 4.1 and Sec. 4.2, respectively. The cross-modal feature pooling method is subsequently introduced in Sec. 4.3. Finally, a practical causal learning pipeline is presented in Sec. 4.4. + +# 4.1. Observable Causal Inference + +Back-door Adjustment Causal Learning (BACL). Based on Bayes's theorem, the typical observational likelihood is as $P(\mathcal{Y}|\mathcal{X}) = \sum_{z}P(\mathcal{Y}|\mathcal{X},z)\underline{P(z|\mathcal{X})}$ , where $P(z|\mathcal{X})$ could bring biased weights. Do-operator [52] provides scientifically sound methods for determining causal effects by severing the back-door link between $\mathcal{Z}$ and $\mathcal{X}$ . According to the invariance and independence rules [17], we have: + +$$ +\begin{array}{l} P (\mathcal {Y} | d o (\mathcal {X})) = \sum_ {z} P (\mathcal {Y} | d o (\mathcal {X}), z) P (z | d o (\mathcal {X})) (1) \\ = \sum_ {z} P (\mathcal {Y} | \mathcal {X}, z) \underline {{P (z)}} (2) \\ \end{array} +$$ + +In such case, the intervention is achieved by blocking the back-door path $\mathcal{Z} \to \mathcal{X}$ , making $\mathcal{X}$ have a fair opportunity to incorporate causality-related factors for prediction. Previous methods [42, 73, 78] used the NWGM approximation [6] to directly pursue causal learning to the final outputs. However, these methods limit the intervention only to the network's final Softmax layer, overlooking possible biased features in shallow layers. Since the conditional probability is implicit in pattern recognition made by the trained neural network [48], we release the target effect of causal hypothesis to learned features rather than merely outputs. Obtaining unbiased features leads to unbiased predictions. Consequently, the specific network module is formulated as $f(x, z)$ , and the implementation of Eq. (2) becomes: + +$$ +\mathcal {B} (\boldsymbol {x}, \boldsymbol {z}) = \mathbb {E} _ {z} [ f (\boldsymbol {x}, \boldsymbol {z}) ] \tag {3} +$$ + +The linear function is used as $f(\pmb{x}, \pmb{z}) = f_{x}(\pmb{x}) + f_{z}(\pmb{z})$ . Then Eq. (3) becomes $f_{x}(\pmb{x}) + \mathbb{E}_{z}[f_{z}(\pmb{z})]$ . To obtain $\mathbb{E}_{z}[f_{z}(\pmb{z})]$ , there are two prevalent approaches: statistic-based [42, 66] and attention-based methods [38, 73]: + +$$ +\text {S t a t i s t i c}: \mathbb {E} _ {z} [ f _ {z} (\boldsymbol {z}) ] = \sum_ {i} \frac {\left| z _ {i} \right|}{\sum_ {j} \left| z _ {j} \right|} f _ {z} \left(\boldsymbol {z} _ {i}\right) \tag {4} +$$ + +$$ +\text {A t t e n t i o n}: \mathbb {E} _ {z} [ f _ {z} (z) ] = \sum_ {i} \frac {\exp \left(h z _ {i} ^ {T}\right)}{\sum_ {j} \exp \left(h z _ {j} ^ {T}\right)} f _ {z} (z _ {i}) \tag {5} +$$ + +where $|z_{i}|$ denotes the number of $z$ belonging to the $i$ -th + +![](images/80ffc652faa72e6959d9435f02acb9c430c4b3d49abf57a9dc5dd3bb25c1b670.jpg) +Figure 3. Framework of GOAT, built on the foundation of the dual-scale graph transformer [9]. Back-door and front-door adjustment causal learning mechanisms are used for mitigating spurious correlations, enabling unbiased feature learning and decision-making. + +![](images/f3e102ecfb78a6d6c64bc12365dccbe70f7904291db8ecde04268a23e05085d8.jpg) +(a) Back-door Adjustment Causal Learning (BACL) + +![](images/d7b42454595816a909fcf839037ac9659893f9dbeaa94fa97b0753d094f8bc40.jpg) +(b) Front-door Adjustment Causal Learning (FACL) +Figure 4. Illustration of BACL and FACL. + +category in the training set, and $h$ means hidden features. The illustration is shown in Fig. 4(a). Our experiments in Sec. 5.3 reveal modality-specific calculation preferences. + +BACL in Text Content. In VLN instructions like "Exit the office and turn right into the kitchen," essential guiding elements such as directions (e.g., "exit" and "right") and landmarks (e.g., "office" and "kitchen") play significant roles. These keywords which are common causes of instruction construction and action distribution, serve as observable confounders. Firstly, we build the text keyword dictionary $\mathcal{D}_I^O = [z_{i,1}^o,z_{i,2}^o,\dots,z_{i,K}^o ]$ with $K$ classes to store confounder features. Direction-and-landmark keywords are extracted based on their part-of-speech tags [63]. We use the pre-trained RoBERTa [41] to obtain feature representations for each extracted token $\pmb{f}_{i}^{o}$ . Since the same word can have different features across sentences, we calculate the average feature for each keyword: $\pmb{z}_{i,n}^{o} = \frac{1}{|z_{i,n}^{o}|}\sum_{j}\pmb{f}_{i,n,j}^{o}$ . Subsequently, the text content causal representation $G_{s}$ is calculated as follows: + +$$ +I = \operatorname {R o B E R T a} \left(\psi_ {t} (\mathcal {I}) + \psi_ {t} (\mathcal {P})\right), Z _ {k} = \mathrm {L N} \left(\phi_ {k} \left(\mathcal {D} _ {I} ^ {O}\right)\right) \tag {6} +$$ + +$$ +G _ {s} = \operatorname {L N} \left[ \phi_ {i} \left(\mathcal {B} \left(I, Z _ {k}\right)\right) \right] \tag {7} +$$ + +where $\psi (\cdot)$ and $\phi (\cdot)$ denote the learnable embedding layer and the full-connection layer, respectively. The absolute positional encoding $\mathcal{P}$ [60] is added to present the position information, and the layer normalization LN [5] is employed + +for stabilizing hidden states during training. + +BACL in Vision Content. For each step, the panorama $\mathcal{V}$ is divided into 36 sub-images. Since existing VLN datasets primarily involve indoor room navigation, visual room references are treated as observable confounders. CLIP [56] is used to extract image features. Since room labels are not directly provided, we employ BLIP [32], a pre-trained VQA model to capture the room information for each image, by querying the model with a fixed prompt "what kind of room is this?" The average value of each room reference type is calculated, forming a visual room reference dictionary $\mathcal{D}_V^O = [z_{v,1}^o,z_{v,2}^o,\dots,z_{v,M}^o]$ , where $M$ is the number of room types. Additionally, the matrix $\gamma = \left[(\sin \theta_i,\cos \theta_i,\sin \eta_i,\cos \eta_i)_{i = 1}^{36}\right]$ is used to present the direction of each image's shift relative to the agent, where $\theta$ and $\eta$ denote the heading and elevation direction. If there are additional object features (for goal-oriented tasks), they are concatenated with image features. Subsequently, a 2-layer transformer encoder is used to capture spatial dependencies. The above process is formulated as follows: + +$$ +V = \operatorname {C L I P} (\mathcal {V}), Z _ {r} = \ln \left(\phi_ {r} \left(\mathcal {D} _ {V} ^ {O}\right)\right) \tag {8} +$$ + +$$ +V _ {v} = \operatorname {L N} \left[ \phi_ {v} \left(\mathcal {B} (V, Z _ {r})\right) \right] \tag {9} +$$ + +$$ +Q _ {v} = \operatorname {T r a n s} \left(V _ {v} + \psi_ {d} (\gamma)\right) \tag {10} +$$ + +# 4.2. Unobservable Causal Inference + +Front-door Adjustment Causal Learning (FACL). In the previous section, we employed the back-door adjustment technique to handle bias caused by observable confounders. However, there are additional unobservable confounders that cannot be explicitly captured and modeled in advance. To address this, we introduce another technique - front-door adjustment [17]. As shown in Fig. 4(b), an additional mediator $\mathcal{M}$ is inserted between inputs and outcomes to construct the front door path $\mathcal{X} \rightarrow \mathcal{M} \rightarrow \mathcal{Y}$ . In VLN, an attention-based model $P(\mathcal{Y}|\mathcal{X}) = \sum_{m} P(\mathcal{Y}|m)P(m|\mathcal{X})$ will select key regions $\mathcal{M}$ from inputs $\mathcal{X}$ for action prediction $\mathcal{Y}$ . Therefore, the model inference can be represented by two parts: the feature selector $\mathcal{X} \rightarrow \mathcal{M}$ which selects suitable knowledge $\mathcal{M}$ from $\mathcal{X}$ , and the action predictor + +$\mathcal{M} \to \mathcal{V}$ which exploits $\mathcal{M}$ to predict $\mathcal{V}$ . To eliminate spurious correlation brought by unobservable confounder $\mathcal{Z}$ , we simultaneously deploy $do$ -operator to $\mathcal{X}$ and $\mathcal{M}$ : + +$$ +\begin{array}{l} P (\mathcal {Y} | d o (\mathcal {X})) = \sum_ {m} P (\mathcal {Y} | d o (m)) P (m | d o (\mathcal {X})) (11) \\ = \sum_ {x ^ {\prime}} P \left(x ^ {\prime}\right) \sum_ {m} P (\mathcal {Y} | m, x ^ {\prime}) P (m | \mathcal {X}) (12) \\ = \mathbb {E} _ {x ^ {\prime}} \mathbb {E} _ {m | x} [ P (\mathcal {Y} | x ^ {\prime}, m) ] (13) \\ \end{array} +$$ + +where $x'$ denotes potential input samples of the whole representation space, different from current inputs $\mathcal{X} = x$ . We use the bold symbol $\pmb{m}$ to denote the in-sampling features obtained by the feature extractor acting on the current input, and $\pmb{x}'$ to mean the cross-sampling features randomly sampled by the K-means-based feature selector from the entire training samples. Based on the linear mapping model, Eq. (13) becomes $\mathbb{E}_{m|x}[m] + \mathbb{E}_{x'}[x']$ . As it is intractable to get a closed-form solution of expectations involving the complex representation space, the estimation is achieved by the query mechanism. Two embedding functions [42, 73] are used to transmit input $\pmb{x}$ into two query sets $g_{1} = q_{1}(\pmb{x})$ and $g_{2} = q_{2}(\pmb{x})$ . Then, the front-door adjustment is approximated as follows: + +$$ +\mathbb {E} _ {\boldsymbol {x} ^ {\prime}} \left[ \boldsymbol {x} ^ {\prime} \right] \approx \sum_ {\boldsymbol {x} ^ {\prime}} P \left(\boldsymbol {x} ^ {\prime} \mid \boldsymbol {g} _ {1}\right) \boldsymbol {x} ^ {\prime} = \sum_ {i} \frac {\exp \left(\boldsymbol {g} _ {1} \boldsymbol {x} _ {i} ^ {\prime T}\right)}{\sum_ {j} \exp \left(\boldsymbol {g} _ {1} \boldsymbol {x} _ {j} ^ {\prime T}\right)} \boldsymbol {x} _ {i} ^ {\prime} \tag {14} +$$ + +$$ +\begin{array}{l} \mathbb {E} _ {m \mid x} [ \boldsymbol {m} ] \approx \sum_ {m} P (\boldsymbol {m} | \boldsymbol {g} _ {2}) \boldsymbol {m} = \sum_ {i} \frac {\exp \left(\boldsymbol {g} _ {2} \boldsymbol {m} _ {i} ^ {T}\right)}{\sum_ {j} \exp \left(\boldsymbol {g} _ {2} \boldsymbol {m} _ {j} ^ {T}\right)} \boldsymbol {m} _ {i} (15) \\ \mathcal {F} \left(\boldsymbol {x}, \boldsymbol {x} ^ {\prime}\right) = \mathbb {E} _ {\boldsymbol {x} ^ {\prime}} \left[ \boldsymbol {x} ^ {\prime} \right] + \mathbb {E} _ {m | x} [ \boldsymbol {m} ] (16) \\ \end{array} +$$ + +The above process can be efficiently implemented using the multi-head attention [60], enabling seamless integration of causal adjustments into existing transformer-based frameworks with minimal modifications. + +FACL in Text, Vision, and History. Considering the characteristics of VLN, we propose to eliminate unobservable confounders from three kinds of inputs in VLN, i.e., vision, language, and history. First, following previous graph-based methods [9, 63], we construct the vision sequence $G_{v} = \{[\mathrm{STOP}], [\mathrm{MEM}], Q_{v}\}$ and the history sequence $G_{h} = \{[\mathrm{STOP}], [\mathrm{MEM}], \{\bar{Q}_{t}\}_{t=1}^{T}\}$ by adding the additional token for presenting stop and recurrent memory states, respectively. $\bar{Q}_{t}$ means the learned weight sum of panoramic features for the $t$ -th step. To condense the lengthy sequence of features and generate global features $x'$ for cross-sampling, we devise the CFP module (as described in Sec. 4.3) with the attentive pooling mechanism to construct confounder dictionaries for vision, history, and instruction, denoted as $D_{v}^{U}, D_{h}^{U}$ , and $D_{s}^{U}$ , respectively. Then the causality-enhanced features $R_{v}, R_{h}$ and $R_{s}$ are calculated based on Eq. (16): + +$$ +R _ {v} = \mathcal {F} \left(G _ {v}, D _ {v} ^ {U}\right), R _ {h} = \mathcal {F} \left(G _ {h}, D _ {h} ^ {U}\right), R _ {s} = \mathcal {F} \left(G _ {s}, D _ {s} ^ {U}\right) \tag {17} +$$ + +Furthermore, we introduce an adaptive gate fusion (AGF) method to enhance the stability of learning by integrating + +![](images/c384ce9f20c3f5aa2f54af95e127afce576f5343224f778a1cfcf400a372ad89.jpg) +Figure 5. Illustration of the cross-modal feature pooling (CFP). + +causality-enhanced features with the original context features for each modality: + +$$ +\omega_ {x} = \delta \left(R _ {x} W _ {x} + G _ {x} W _ {g} + b\right) \tag {18} +$$ + +$$ +U _ {x} = \omega_ {x} \odot R _ {x} + (1 - \omega_ {x}) \odot G _ {x} \tag {19} +$$ + +where $\delta$ and $\odot$ mean the Sigmoid function and elementwise multiplication. Suppose $R_{x}$ and $G_{x} \in \mathbb{R}^{L_{x} \times d_{h}}$ , then $W_{x/g} \in \mathbb{R}^{d_{h} \times 1}$ and $b \in \mathbb{R}^{L_{x} \times 1}$ are learnable parameters. Next, the cross-modal fused local features $F_{l}$ and global features $F_{g}$ are obtained by the cross-attention encoders $\mathcal{C}$ from METER [15]. The dynamic fusion $\mathcal{DF}$ [9] followed by Softmax $\mathcal{SF}$ is applied for action prediction: + +$$ +F _ {l} = \mathcal {C} \left(U _ {v}, U _ {s}, U _ {s}\right), F _ {g} = \mathcal {C} \left(U _ {h}, U _ {s}, U _ {s}\right) \tag {20} +$$ + +$$ +F _ {f} = \mathcal {D F} \left(F _ {l}, F _ {g}\right), y _ {t} = \mathcal {S F} \left(F _ {f}\right) \tag {21} +$$ + +The cross-entropy loss is used to optimize the network: + +$$ +\mathcal {L} _ {c e} = \sum_ {t = 1} ^ {T} - \log P \left(y _ {t} ^ {*} \mid \mathcal {I}, \mathcal {V} _ {t}, \mathcal {H} _ {1: t - 1}\right) \tag {22} +$$ + +# 4.3. Cross-modal Feature Pooling + +One challenge of implementing the front-door adjustment in VLN is constructing efficient dictionaries for global features from long sequences. This requires compressing sequential features of varying lengths into a unified feature space to represent each sample effectively. Let $H \in \mathbb{R}^{L \times d_h}$ be the sequential features, the following attentive pooling is used to effectively compress the sequence length: + +$$ +A = \mathcal {T} (H), \alpha = \mathcal {S F} \left(A W _ {a}\right), \bar {H} = \mathcal {T} \left(\alpha^ {T} H\right) \tag {23} +$$ + +where $\mathcal{T}$ denote the Tanh activation, $W_{a} \in \mathbb{R}^{d_{h} \times 1}$ is the learnable attention matrix, and $\bar{H} \in \mathbb{R}^{1 \times d_{h}}$ . As shown in Fig. 5, for vision, history, local-global fusion, and text features, we use one transformer layer as the encoder followed by the attentive pooling to obtain flattened features $\bar{C}^{l}, \bar{C}^{g}, \bar{C}^{f}$ , and $\bar{C}^{s}$ , respectively. Then, we adopt contrastive learning [35, 57, 62] to optimize this cross-modal feature pooling (CFP) module, meanwhile improving semantic alignments for different modalities. The contrastive loss $\mathcal{L}_{ls}$ is constructed as: + +$$ +\begin{array}{l} \mathcal {L} _ {l s} = - \frac {1}{2 B} \sum_ {j = 1} ^ {B} \log \frac {\exp \left(\left\langle \bar {C} _ {j} ^ {l} , \bar {C} _ {j} ^ {s} \right\rangle / t\right)}{\sum_ {k = 1} ^ {B} \exp \left(\left\langle \bar {C} _ {j} ^ {l} , \bar {C} _ {k} ^ {s} \right\rangle / t\right)} \tag {24} \\ - \frac {1}{2 B} \sum_ {k = 1} ^ {B} \log \frac {\exp \left(\left\langle \bar {C} _ {k} ^ {l} , \bar {C} _ {k} ^ {s} \right\rangle / t\right)}{\sum_ {j = 1} ^ {B} \exp \left(\left\langle \bar {C} _ {j} ^ {l} , \bar {C} _ {k} ^ {s} \right\rangle / t\right)} \\ \end{array} +$$ + +where $B$ and $t$ mean the batch size and temperature, respectively. Similarly, contrastive losses $\mathcal{L}_{gs}$ and $\mathcal{L}_{fs}$ are calculated by replacing $\bar{C}^l$ with $\bar{C}^g$ and $\bar{C}^f$ . The overall CFP loss is the sum of these losses $\mathcal{L}_{\mathrm{CFP}} = \mathcal{L}_{ls} + \mathcal{L}_{gs} + \mathcal{L}_{fs}$ . + +To enable the network more adaptive to characteristics for VLN and thus facilitate the building of the front-door confounder dictionaries for samples, we train the CFP module alongside other auxiliary tasks [9] during pre-training. Subsequently, the trained attentive pooling modules are used to extract global features for different modalities. In the fine-tuning stage, we employ the BACL and FACL with established dictionaries for intervention. The CFP offers dual benefits: it aligns different modalities more effectively during pre-training and provides a systematic approach for extracting coherent representations from sequence inputs. + +# 4.4. Causal Learning Pipeline + +As shown in Fig. 6, we summarize a causal learning pipeline that serves as a blueprint for similar learning-based methods. First, it begins with task formulation, where the specific task and its objectives are precisely defined. Next, the observable and unobservable confounders are explicitly assumed. Both back-door and front-door adjustment strategies are employed to tackle these confounders, either simultaneously or sequentially, contingent on task specifics. Then, the pipeline proceeds + +![](images/275afebef1bbe4e948e4a9c031740d79c720e486ebf20ca5a47b6ae66964b3e0.jpg) +Figure 6. Pipeline of the causal learning. + +to model calculation and result prediction. Throughout network optimization, both network parameters and confounder features are continuously updated. Ultimately, this iterative process leads us to the development of a robust causal model capable of generating unbiased features, thereby advancing the generalizability of AI systems. + +# 5. Experiments + +# 5.1. Experimental Settings + +1) Datasets. We verify GOAT on two kinds of VLN benchmarks: fine-grained datasets (R2R [4] and RxR-English [29]), which provide long step-by-step navigation instructions, and goal-oriented datasets (REVERIE [53] and SOON [81]), which additionally requires for the target object. Formally, the datasets are partitioned into four splits: training, validation seen (sharing the same environments with the training set), validation unseen (having different + +environments from the training set), and test unseen sets (reported by the online leaderboard for fair comparison). + +2) Evaluation Metrics. In R2R, key metrics include Navigation Error (NE), Success Rate (SR), Oracle SR (OSR), and SR Weighted by Path Length (SPL). RxR adds Normalized Dynamic Time Warping (nDTW) and SR Weighted by Dynamic Time Warping (sDTW). REVERIE and SOON introduce Remote Grounding Success Rate (RGS) and RGS Weighted by Path Length (RGSPL). +3) Implementation Details. Our model consists of 6 transformer layers for text, 2 for panorama, and 3 for cross-modal encoding. We use CLIP-B/16 [56] for image feature extraction and initialize network weights with METER [15]. In pre-training, MLM [28], SAP [8], and the proposed CFP are for R2R and RxR. OG [37] is added for REVERIE and SOON. EnvEdit [34] is employed for feature augmentation. The synthetic extended datasets [19, 62, 65] are used for R2R, REVERIE, and RxR, respectively. Pretraining is done on a single Tesla V100 GPU for a maximum of 300K iterations by the AdamW [43] optimizer, with a batch size of 48 and a learning rate of $5 \times 10^{-5}$ . The numbers of classes of keywords and rooms are 74 and 50, and the temperature $t$ is set to 1. In fine-tuning, the front-door dictionaries are randomly sampled from the K-Means clustering features. As the text transformer RoBERTa [41] is involved in end-to-end training, the textual keywords dictionary is also iteratively updated. Speaker models with the environmental dropout [58, 62, 64] are used to provide dynamic pseudo labels. We employ batch size 12 for R2R, REVERIE, and 5 for RxR and SOON, with a learning rate of $2 \times 10^{-5}$ and a maximum of 100K iterations. + +# 5.2. Comparisons with State-of-the-Arts + +In Tab. 1, 2, 3, 4, we compare GOAT with the previous state-of-the-art (SoTA) methods on the R2R, REVERIE, RxR-English, and SOON datasets, respectively. On all these four datasets, our approach exhibits superior navigation performance, precise instruction-following alignment, and accurate object grounding across both seen and unseen environments. For instance, in R2R, GOAT achieves remarkable improvements in SPL compared to BEVBert [3], with relative increases of $7.41\%$ , $6.45\%$ , and $4.74\%$ on three subsets. In REVERIE, GOAT shows substantial relative enhancements in RGSPL by $11.83\%$ , $6.55\%$ , and $20.87\%$ on three subsets. In the challenging SOON and RxR tasks, GOAT also exhibits significant improvements in performance metrics, highlighting its robustness and superior generalization capabilities over previous methods. + +# 5.3.Quantitative Analysis + +1) Effect of Causal Inference. Fig. 7 verifies the impact of causal inference on GOAT across four diverse VLN datasets in unseen environments. "W/o intervention" signifies the + +
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑
HAMT [8]76722.518266613.297365603.9372
DUET [9]79732.288672603.318169593.6576
TD-STP [80]77732.348370633.227667613.7372
GeoVLN [25]79762.22-68633.35-65613.95-
DSRG [63]81762.238873623.008172613.3378
GridMM [71]----75642.83-73623.35-
GELA [10]76732.39-71653.11-67623.59-
EnvEdit [34]77742.32-69643.24-68643.59-
BEVBert [3]81742.178875642.818473623.1381
GOAT (Ours)83.7479.481.7988.6477.8268.132.4084.7274.5764.943.0480.35
+ +Table 1. Comparison with other state-of-the-art methods on the R2R dataset [4]. '-': unavailable statistics. + +
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑
HAMT [8]43.2940.1927.2025.1832.9530.2018.9217.2830.4026.6714.8813.08
HOP+ [54]55.8749.5540.7636.2236.0731.1322.4919.3333.8228.2420.2016.86
DUET [9]71.7563.9457.4151.1446.9833.7332.1523.0352.5136.0631.8822.06
DSRG [63]75.6968.0961.0754.7247.8334.0232.6923.3754.0437.0932.4922.18
GridMM [71]----51.3736.4734.5724.5653.1336.6034.8723.45
BEVBert [3]73.7265.3257.7051.7351.7836.3734.7124.4452.8136.4132.0622.09
GOAT (Ours)78.6471.4063.7457.8553.3736.7038.4326.0957.7240.5338.3226.70
+ +Table 2. Comparison with other state-of-the-art methods on the REVERIE dataset [53]. '-': unavailable statistics. + +
MethodValidation SeenValidation Unseen
SR↑SPL↑nDTW↑sDTW↑SR↑SPL↑nDTW↑sDTW↑
Syntax [31]48.144.058.040.039.235.052.032.0
SOAT [45]----44.2-54.836.4
HOP+ [54]53.647.959.043.045.738.452.036.0
FOAM [14]----42.838.754.135.6
ADAPT [35]50.344.656.340.646.940.254.137.7
\( \mathsf{MAR}_{\mathsf{M-MP}} \) [27]----50.2-60.343.9
VLN-PETL [55]60.556.865.751.757.954.264.949.7
GOAT (Ours)74.168.171.061.468.261.767.156.6
+ +Table 3. Comparison on the RxR-English dataset [29]. + +
MethodValidation UnseenTest Unseen
OSR↑SR↑SPL↑RGSPL↑OSR↑SR↑SPL↑RGSPL↑
GBE [81]28.5419.5213.341.1621.4512.909.230.45
DUET [9]50.9136.2822.583.7543.0033.4421.424.17
GridMM [71]53.3937.4624.813.9148.0236.2721.254.15
GOAT (Ours)54.6940.3528.055.8550.6340.5025.186.10
+ +![](images/c4f989acaa141cb83a38070786678f8a1a76c54893358743b8d9d931f1eadec2.jpg) +Figure 7. Effect of the intervention on various VLN datasets. + +exclusion of the proposed BACL and FACL interventions across all modalities. On each of the datasets examined, the + +Table 4. Comparison on the SOON dataset [81]. + +
IdBACLFACLSR↑SPL↑NE↓OSR↑
1XX75.2764.692.7283.14
2✓X76.3767.132.6083.99
3X✓76.5066.922.5384.21
4✓✓77.8268.132.4084.72
+ +Table 5. Effect of back-door and front-door adjustments. + +integration of causal inference leads to significant enhancements in the model's performance. This strongly demonstrates causal learning's considerable popularization potential in enhancing learning-based model generalization. + +2) Effect of BACL and FACL. Tab. 5 analyzes the effects of the proposed BACL and FACL on the R2R val-unseen subset. Compared to the baseline (#1), individual application of either BACL (#2) or FACL (#3) leads to performance improvements. Concurrent use of BACL and FACL (#4) leads to further performance enhancements. These findings underscore our assumption about the presence of both observable and unobservable confounders. Integrating both back-door and front-door adjustments is crucial to comprehensively addressing dataset biases, and enhancing the model's robustness and generalization. +3) Effect of CFP. In Tab. 6, we assess the efficacy of the proposed CFP on the R2R val-unseen subset. During pretraining (PT), incorporating CFP as an additional auxiliary task (CFP-P) enhances training performance, improving SR and SPL by $3.11\%$ and $3.02\%$ (#A1), respectively. In the fine-tuning (FT) stage, we compare the performance with and without the use of trained attention modules from CFP + +
StageIdMethodSR↑SPL↑NE↓OSR↑
PTA0w/o CFP-P40.3637.886.5248.87
A1w/ CFP-P43.4740.906.0653.13
FTB0A0 w/o CFP-F75.5665.902.6382.42
B1A1 w/o CFP-F76.6366.172.6384.67
B2A1 w/ CFP-F77.8268.132.4084.72
+ +Table 6. Effect of CFP in pre-training and fine-tuning. + +
IdTextVisionSR↑SPL↑NE↓OSR↑
1StatsStats75.2264.782.7183.65
2StatsAttn76.5965.392.5685.31
3AttnStats77.8268.132.4084.72
4AttnAttn75.9565.832.6483.91
+ +Table 7. Effect of Statistic and Attention methods in BACL. + +![](images/47a6a564e7234786ba2b7ba4463bd1a41ee5cc62862ef899ec9eb2809a49fb1e.jpg) +Figure 8. Comparison of the distribution of SR and SPL of instructions involving specific terms on the R2R val-unseen split. + +to extract global features for front-door dictionaries (CFP-F). "W/o CFP-F" signifies the use of simple average pooling to compress features from the pre-trained model. #B2 shows that the CFP provides more reliable confounder representations for causal learning (SPL $\uparrow 1.96\%$ ). + +4) Effect of Different BACL in Different Modalities. Tab. 7 investigates the effect of various combinations of statistic-based and attention-based methods for text and vision in BACL on the R2R val-unseen subset. The results indicate that employing the attention method for text and the statistic method for vision yields the best performance (#3). Intuitively, this can be explained by the structured nature of textual information and the involvement of RoBERTa's end-to-end training, enabling the attention method to effectively capture contextual nuances. Conversely, images lack explicit causality, and CLIP, the image extractor, isn't trained directly for efficiency reasons. Consequently, the statistic method ensures a stable causal learning process, preserving the integrity of vision-related features. + +# 5.4. Qualitative Analysis + +1) Bias Elimination Effect. In Fig. 8, the compactness of the boxes represents concentrated data distribution and reduced variability, while a median line closer to the center signifies even data distribution. It shows that GOAT obtains narrower boxes and more central midlines across diverse objects. This finding showcases that with the integration of + +![](images/b891ed21cd13a524a64835a274d050eef13f0d10953fd8fee0fccd7a64edc064.jpg) + +Starting point + +![](images/b76c982cc7e2998589bc6f30992839df3bccb0424ac074f8dd852fcac86d9bc2.jpg) + +Destination + +![](images/cdc97a6da8f62ed7905910767cf90479b83fa0595f9c90896f98a25bed6f87ff.jpg) + +sited point + +![](images/ac5893ffc1d6d8a8a2080b95c6c45e3e79dee430a2ca9c97049eaaff89b0863c.jpg) + +visited point + +![](images/adb9e57de36b26ee94ec1eafaa75f51e3c88f4498fff801771719b7c12ed4526.jpg) + +Ground-truth + +![](images/5110e0d59417ee3c68d119ed32530a3ae9227ecdd506b1f29d190786e8d5553a.jpg) + +DSRG + +![](images/3ab80d19cffa611cda098bd44745bba0d2ef722bf0d7ad3a4548ab15b651b3f2.jpg) + +GOAT (Ours) + +# - Instruction: + +Walk out of the room and go past the stairs and into the bedroom that is straight ahead, go to the bathroom doorway and wait there. + +![](images/558c752f0ac4f1db01869ae62f6f41379dd5aa3eddb12983f197cc3594dac444.jpg) + +# - Instruction: + +Slightly move over the green chair closest to the corner of the reception desk in the lounge. + +![](images/a6f1139e53657d981f57a2a45b4c77dc05c0e975f6313697287caa05faf5105d.jpg) +Figure 9. Predicted trajectories in unseen environments. + +causal intervention, GOAT significantly reduces prediction bias, thereby enhancing its generalization capability in previously unseen environments. + +2) Visualized Trajectories. In Fig. 9, we visualize some predicted trajectories in unseen environments, comparing them with DSRG on R2R and REVERIE datasets. Notably, GOAT precisely captures directional cues like "straight ahead" and nuanced instructions like "closest to the corner", enabling accurate predictions. These instances highlight the intricate causal connections in VLN tasks, where specific instructions prompt corresponding actions. GOAT's enhanced causal inference capability enables it to generate more reasoned responses aligned with the provided instructions, underscoring the significance of robust causal inference in VLN systems. Please refer to our supplementary material for more detailed discussions and visualizations. + +# 6. Conclusion + +Our work presents GOAT, a novel approach that addresses the dataset bias in VLN from the perspective of causal learning. The back-door and front-door adjustment causal learning (BACL and FACL) mechanisms are proposed to adjust for observable and unobservable confounders, respectively. The cross-modal feature pooling (CFP) module is adopted to promote feature learning and extraction through contrastive learning. The practical causal learning pipeline is presented to illuminate other similar learning-based methods. Experiments on R2R, REVERIE, RxR, and SOON datasets show that GOAT can reasonably discover sequence visual-linguistic causal structures and significantly improve performance. Beyond VLN, the underlying confounder assumption and causal inference principles are generalizable to other similar fields. + +# Acknowledgments + +This paper is supported by the National Natural Science Foundation of China under Grants (62233013, 62073245, 62173248). Shanghai Science and Technology Innovation Action Plan (22511104900). + +# References + +[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10044-10054, 2020. 2 +[2] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5101-5109, 2021. 1, 2 +[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 6, 7 +[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 6, 7, 14 +[5] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. ArXiv, abs/1607.06450, 2016. 4 +[6] Pierre Baldi and Peter Sadowski. The dropout learning algorithm. Artificial intelligence, 210:78-122, 2014. 3 +[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2 +[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7 +[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 5, 6, 7, 14 +[10] Yibo Cui, Liang Xie, Yakun Zhang, Meishan Zhang, Ye Yan, and Erwei Yin. Grounded entity-landmark adaptive pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12043–12053, 2023. 1, 7 +[11] Ronghao Dang, Zhuofan Shi, Liuyi Wang, Zongtao He, Chengju Liu, and Qijun Chen. Unbiased directed object attention graph for object navigation. In Proceedings of the 30th ACM International Conference on Multimedia, page 3617-3627, New York, NY, USA, 2022. Association for Computing Machinery. 2 +[12] Ronghao Dang, Lu Chen, Liuyi Wang, He Zongtao, Chengju Liu, and Qijun Chen. Multiple thinking achieving meta-ability decoupling for object navigation. In International Conference on Machine Learning (ICML), 2023. 1 + +[13] Ronghao Dang, Liuyi Wang, Zongtao He, Shuai Su, Jiagui Tang, Chengju Liu, and Qijun Chen. Search for or navigate to? dual adaptive thinking for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8250-8259, 2023. 1 +[14] Zi-Yi Dou and Nanyun Peng. Foam: A follower-aware speaker model for vision-and-language navigation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4332-4340, 2022. 1, 7 +[15] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 5, 6 +[16] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. Advances in Neural Information Processing Systems, 31, 2018. 1, 2, 3 +[17] Madelyn Glymour, Judea Pearl, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016. 3, 4 +[18] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 1, 2 +[19] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13137-13146, 2020. 2, 6 +[20] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1 +[21] Zongtao He, Liuyi Wang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Mlanet: Multi-level attention network with sub-instruction for continuous vision-and-language navigation. arXiv preprint arXiv:2303.01396, 2023. 1, 2 +[22] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2 +[23] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3055-3067, 2023. 1 +[24] Ronghang Hu, Daniel Fried, Anna Rohrbach, Dan Klein, Trevor Darrell, and Kate Saenko. Are you looking? grounding to multiple modalities in vision-and-language navigation. + +In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6551-6557, 2019. 1 +[25] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23212-23221, 2023. 7 +[26] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1862-1872, 2019. 1 +[27] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 7 +[28] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 6 +[29] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392-4412, 2020. 1, 2, 6, 7, 14 +[30] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 1 +[31] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1041–1050, 2021. 7 +[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4 +[33] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21273–21282, 2022. 2 +[34] Jialu Li, Hao Tan, and Mohit Bansal. Envideit: Environment editing for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15407-15417, 2022. 1, 2, 6, 7 +[35] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2022. 1, 5, 7 +[36] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8317-8326, 2023. 1, 2 +[37] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7036-7045, 2021. 6 +[38] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18041-18050, 2022. 2, 3 +[39] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1644-1654, 2021. 1, 2 +[40] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1 +[41] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6 +[42] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3, 5 +[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6 +[44] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In European Conference on Computer Vision, pages 259-274. Springer, 2020. 1, 2 +[45] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34:7357-7367, 2021. 7 +[46] Laksh Nanwani, Anmol Agarwal, Kanishk Jain, Raghav Prabhakar, Aaron Monis, Aditya Mathur, Krishna Murthy, Abdul Hafez, Vineet Gandhi, and K Madhava Krishna. Instance-level semantic maps for vision language navigation. arXiv preprint arXiv:2305.12363, 2023. 1 +[47] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference + +on Natural Language Processing (EMNLP-IJCNLP), pages 684-695, 2019. 16 +[48] Siqi Nie, Meng Zheng, and Qiang Ji. The deep regression bayesian network and its applications: Probabilistic deep learning for computer vision. IEEE Signal Processing Magazine, 35(1):101-111, 2018. 3 +[49] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2 +[50] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Javen Qinfeng Shi, and Anton Van den Hengel. Counterfactual vision-and-language navigation: Unravelling the unseen. Advances in Neural Information Processing Systems, 33:5296-5307, 2020. 2 +[51] Judea Pearl. Causality. Cambridge university press, 2009. 1 +[52] Judea Pearl and Dana Mackenzie. The book of why: the new science of cause and effect. Basic books, 2018. 2, 3, 13 +[53] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 3, 6, 7, 14 +[54] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7 +[55] Yanyuan Qiao, Zheng Yu, and Qi Wu. Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 7 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6 +[57] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? In International Conference on Learning Representations. 5 +[58] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, 2019. 2, 6 +[59] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning, pages 394-406. PMLR, 2020. 16 +[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural + +information processing systems, pages 5998-6008, 2017. 4, 5 +[61] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15471-15481, 2022. 1 +[62] Liuyi Wang, Zongtao He, Ronghao Dang, Huiyi Chen, Chengju Liu, and Qijun Chen. Res-sts: Referring expression speaker via self-training with scorer for goal-oriented vision-language navigation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 2, 5, 6 +[63] Liuyi Wang, Zongtao He, jiagui Tang, Ronghao Dang, naijia Wang, Chengju Liu, and Qijun Chen. A dual semantic-aware recurrent global-adaptive network for vision-and-language navigation. In International Joint Conferences on Artificial Intelligence (IJCAI), 2023. 1, 2, 4, 5, 7 +[64] Liuyi Wang, Chengju Liu, Zongtao He, Shu Li, Qingqing Yan, Huiyi Chen, and Qijun Chen. Pasts: Progress-aware spatio-temporal transformer speaker for vision-and-language navigation. arXiv preprint arXiv:2305.11918, 2023. 1, 2, 6 +[65] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 1, 6 +[66] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10760-10770, 2020. 2, 3 +[67] Tan Wang, Chang Zhou, Qianru Sun, and Hanwang Zhang. Causal attention for unbiased visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3091-3100, 2021. 2 +[68] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Wang, and Lei Zhang. Vision-language navigation policy learning and adaptation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2 +[69] Yuqing Wang, Xiangxian Li, Zhuang Qi, Jingyu Li, Xuelong Li, Xiangxu Meng, and Lei Meng. Meta-causal feature learning for out-of-distribution generalization. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 530-545. Springer, 2023. 2 +[70] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12009-12020, 2023. 1 +[71] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15625-15636, 2023. 7 + +[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 16 +[73] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 3, 5 +[74] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 2 +[75] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. Advances in neural information processing systems, 33:2734-2746, 2020. 2 +[76] Hua Zhang, Liqiang Xiao, Xiaochun Cao, and Hassan Foroosh. Multiple adverse weather conditions adaptation for object detection via causal intervention. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2 +[77] Jiazhao Zhang, Liu Dai, Fanpeng Meng, Qingnan Fan, Xuelin Chen, Kai Xu, and He Wang. 3d-aware object goal navigation via simultaneous exploration and identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6672-6682, 2023. 1 +[78] Shengyu Zhang, Tan Jiang, Tan Wang, Kun Kuang, Zhou Zhao, Jianke Zhu, Jin Yu, Hongxia Yang, and Fei Wu. Devlbert: Learning deconfounded visio-linguistic representations. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4373–4382, 2020. 2, 3 +[79] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, 2021. 1 +[80] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4194-4203, 2022. 7 +[81] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12699, 2021. 2, 3, 6, 7, 14 +[82] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1 \ No newline at end of file diff --git a/2024/Vision-and-Language Navigation via Causal Learning/images.zip b/2024/Vision-and-Language Navigation via Causal Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e6362eb34003b3ec81f74aae95187608a003043 --- /dev/null +++ b/2024/Vision-and-Language Navigation via Causal Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e94c19b7af50e7a5aff27a474d3b1d8c2ab2f9cb8e8633299083ea1b4b787a86 +size 797355 diff --git a/2024/Vision-and-Language Navigation via Causal Learning/layout.json b/2024/Vision-and-Language Navigation via Causal Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c384243d0a6779bae52f46bfd1a28dcf3240045b --- /dev/null +++ b/2024/Vision-and-Language Navigation via Causal Learning/layout.json @@ -0,0 +1,11715 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 463, + 121 + ], + "type": "text", + "content": "Vision-and-Language Navigation via Causal Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 143, + 507, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 143, + 507, + 172 + ], + "spans": [ + { + "bbox": [ + 85, + 143, + 507, + 172 + ], + "type": "text", + "content": "Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu*, Qijun Chen* \nSchool of Electronic and Information Engineering, Tongji University, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 174, + 533, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 533, + 186 + ], + "type": "text", + "content": "{wly, xingchen327, dangronghao, liuchengju, qjchen}@tongji.edu.cn, mojo.shum@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 491 + ], + "type": "text", + "content": "In the pursuit of robust and generalizable environment perception and language understanding, the ubiquitous challenge of dataset bias continues to plague vision-and-language navigation (VLN) agents, hindering their performance in unseen environments. This paper introduces the generalized cross-modal causal transformer (GOAT), a pioneering solution rooted in the paradigm of causal inference. By delving into both observable and unobservable confounders within vision, language, and history, we propose the back-door and front-door adjustment causal learning (BACL and FACL) modules to promote unbiased learning by comprehensively mitigating potential spurious correlations. Additionally, to capture global confounder features, we propose a cross-modal feature pooling (CFP) module supervised by contrastive learning, which is also shown to be effective in improving cross-modal representations during pre-training. Extensive experiments across multiple VLN datasets (R2R, REVERIE, RxR, and SOON) underscore the superiority of our proposed method over previous state-of-the-art approaches. Code is available at https://github.com/CrystalSixone/VLN-GOAT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 536, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 679 + ], + "type": "text", + "content": "Effective environment perception, language understanding, and historical utilization are at the core of vision-and-language navigation (VLN) [4]. Despite significant progress, deploying VLN in the real world remains a huge challenge, primarily due to diversities and uncertainties in environments and instructions. A key hindrance is dataset bias [79, 82], e.g., agents may overfit to familiar visual environments, resulting in diminished performance in environments with diverse appearances and layouts [24]. This over-reliance on specific patterns, like biased structural trajectories and repeated entity components, raises concerns about the robustness and generalizability of VLN systems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "type": "text", + "content": "One way to mitigate dataset bias in VLN is to build" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 213, + 543, + 322 + ], + "blocks": [ + { + "bbox": [ + 310, + 213, + 543, + 322 + ], + "lines": [ + { + "bbox": [ + 310, + 213, + 543, + 322 + ], + "spans": [ + { + "bbox": [ + 310, + 213, + 543, + 322 + ], + "type": "image", + "image_path": "8705d332ccafaf18cd2753b9850bba1e2643f8af6881359d216b300f587b5720.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 324, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 305, + 324, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 547, + 369 + ], + "type": "text", + "content": "Figure 1. In response to language instructions, the VLN agent is required to navigate to the target based on visual cues. This paper introduces a causal learning pipeline using the do-operator to reduce bias from confounders in VLN action prediction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 547 + ], + "type": "text", + "content": "broader and more diverse datasets, which is what numerous recent studies have focused on. These include using speaker models to generate pseudo instructions [14, 16, 61, 62, 64, 65], synthesizing cross-environmental trajectories [26, 39], transferring image styles [30, 34], collecting data from the web [18, 36, 44, 70], and labeling more fine-grained, entity-aligned instructions [10, 21, 29]. However, achieving a perfectly balanced dataset devoid of bias is nearly impossible. Consequently, we often find ourselves caught in a cycle of \"creating a dataset\" - \"identifying bias\" - \"creating a new dataset\". Therefore, we are prompted to move from diagnosis to treatment [73], transitioning from the continuous collection of new datasets to the development of unbiased models that can confront and mitigate bias." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "However, existing methods that focus on model designing mainly concentrate on introducing more types of inputs (e.g., objects [2, 12, 13, 23, 35, 63] and depth [3, 20, 40, 77]), or constructing the global graph [8, 9, 46] to represent environments. These efforts, although valuable, often overlook underlying dataset biases and the essential causal logic behind the task. In fact, the reason why humans can well execute various instructions and navigate in unknown environments is that we can learn the inherent causality of events beyond biased observation, achieving good analogical association capability. Therefore, for the first time, we propose to use causal inference [51] to equip VLN agents with similar cognitive abilities that we have, and then allow" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13139" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 208, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 208, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 208, + 83 + ], + "type": "text", + "content": "them to make more reasoned decisions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 240 + ], + "type": "text", + "content": "Then, how to develop such a causal inference capability? Although there is no single answer, we propose to exploit the concepts of intervention [52] - technique that uses a do-operator to alleviate the negative effects raised by confounders. Here, confounders are variables that influence both inputs and outcomes, creating spurious correlations and biases. Intervention empowers researchers to mitigate the impact of confounders, enabling the model to grasp the causation of events during data fitting. However, given the fact that VLN is such a complex task that involves cross-modal inputs and a long-term decision-making process, it is challenging to identify underlying confounders and apply intervention to debias through network learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 242, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 286, + 552 + ], + "type": "text", + "content": "To address the above challenges, we propose a generalized cross-modal causal transformer (GOAT) approach that enables the VLN model to alleviate the negative effects raised by confounders, thus achieving causal decision-making (Fig. 1). Firstly, we propose a unified structural causal model to describe the VLN system, involving two distinct categories of confounders: observable and unobservable. Observable confounders are content-related and easily identifiable (e.g., the keywords in instructions and the room references in environments), whereas unobservable confounders refer to intricate stylistic nuances that are harder to discern but can impact the overall system (e.g., decoration styles in vision, sentence patterns in language, and trajectory trends in history). Then, we propose to address these confounders via two causal learning modules that are based on back-door and front-door adjustments [52] (namely BACL and FACL), respectively. Furthermore, to build global dictionaries for representing confounders, we devise a cross-modal feature pooling (CFP) module to effectively aggregate long-sequential features. Contrastive learning [56] is adopted to optimize CFP, serving as an additional auxiliary task during pre-training. As demonstrated by thorough experiments, our findings reveal the impact of integrating causal learning to deconfound biases on cross-modal inputs, offering valuable insights for enhancing generalization in similar tasks across diverse scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 554, + 274, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 554, + 274, + 565 + ], + "spans": [ + { + "bbox": [ + 59, + 554, + 274, + 565 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 570, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 570, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 286, + 604 + ], + "type": "text", + "content": "- We propose a unified structural causal model for VLN by comprehensively considering the observable and unobservable confounders hidden in different modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 606, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 286, + 640 + ], + "type": "text", + "content": "- we propose BACL and FACL, using the back-door and front-door adjustments to allow end-to-end unbiased cross-modal intervention and decision-making." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 642, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 675 + ], + "type": "text", + "content": "- We propose CFP, a cross-modal feature pooling module designed to aggregate sequence features for semantic alignment and confounder dictionaries construction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "- Our GOAT model demonstrates exceptional generalization across multiple VLN datasets (R2R [4], RxR [29], REVERIE [53], and SOON [81]), outperforming existing" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 96 + ], + "type": "text", + "content": "state-of-the-art methods. A comprehensive causal learning pipeline is presented to inspire future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 106, + 392, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 392, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 392, + 118 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "type": "text", + "content": "Vision-and-Language Navigation (VLN) requires agents to navigate to specific locations [4, 29] or find target objects [53, 81] in real visual environments based on natural instructions. Its practicability has led to significant interest, showing potential in fundamental embodied AI skills. Initial models relied on recurrent neural networks [2, 11, 21, 68]. Transformer-based models [8, 9, 19, 22, 63] brought substantial progress due to their powerful long-distance encoding. However, small-scale datasets in VLN were found to cause bias, leading to serious overfitting. Consequently, several approaches were devised to tackle this challenge. Speaker-follower frameworks [16, 27, 58, 62, 64] used the speaker model to generate pseudo instructions. VLN-BERT [44], AirBERT [18], and Lily [36] collected trajectory-instruction pairs from diverse sources. REM [39] and EnvEdit [34] created new environments by editing existing environments. Although these methods improve generalization, they cannot completely eliminate inherent dataset biases. Therefore, we propose to develop an unbiased model by equipping the agent with the causal inference capability to learn cause-effect relations during data fitting, enabling them to adapt adeptly to diverse situations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 617 + ], + "type": "text", + "content": "Causal Inference is an emerging technique exploring task causality [52], leading to a surge in efforts to integrate it with deep learning, in tasks like image recognition [67, 69, 76], image captioning [38, 73], and visual question answering [33, 49]. One popular way is to use the adjustment technique to alleviate the negative effects caused by confounders, and some studies exploring the use of counterfactuals [1, 49, 50]. This paper emphasizes the adjustment method due to its practicality. However, most of the existing causal learning tasks are simple without considering more challenging tasks like VLN. Additionally, current methods apply back-door [38, 66, 75, 78] or front-door [42, 73, 74] adjustments separately across modalities, lacking comprehensive confounder assumptions and complete bias corrections. In this paper, we propose to simultaneously tackle both observable and unobservable confounders in vision, language, and history. This approach significantly reduces overall bias, enhancing the generalization capabilities of embodied VLN agents." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 628, + 382, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 382, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 382, + 640 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 647, + 411, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 411, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 411, + 658 + ], + "type": "text", + "content": "3.1. Task Formulation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The VLN task [4] involves an embodied agent following natural language instructions to navigate real indoor environments. Matterport3D simulator [7] is used to allow interaction, where the environment is provided as graphs" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13140" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 263, + 187 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 263, + 187 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 263, + 187 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 263, + 187 + ], + "type": "image", + "image_path": "c4d24ae1254ed429bf53c78f1708a40b95a7a6c4f5935dee0fc2107b6c8e846b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 189, + 277, + 200 + ], + "lines": [ + { + "bbox": [ + 57, + 189, + 277, + 200 + ], + "spans": [ + { + "bbox": [ + 57, + 189, + 277, + 200 + ], + "type": "text", + "content": "Figure 2. Illustration of the structural causal model of VLN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": "with connected navigable nodes. The agent receives natural language instructions " + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{w_1, w_2, \\dots, w_L\\}" + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": " words, and the current panorama separated into 36 subimages " + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_1, v_2, \\dots, v_{36}\\}" + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": " [16]. The agent also knows its current heading " + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": " and elevation " + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": ". During navigation, the agent needs to select the next point from nearby candidates or predicts the stop signal based on visual cues. Success is defined when the stop location is within 3 meters of the ground-truth position. For the goal-oriented task, REVERIE [53] and SOON [81] additionally require locating the target object at the final destination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 356, + 225, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 225, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 225, + 369 + ], + "type": "text", + "content": "3.2. Structural Causal Model of VLN" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": "As illustrated in Fig. 2, we construct a structural causal model capturing the relationships among the key variables in VLN: visual observation " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ", linguistic instruction " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ", decision history " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ", and action prediction " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ". To clarify, we use " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": " to denote inputs " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "(\\mathcal{X} = \\{\\mathcal{V}, \\mathcal{I}, \\mathcal{H}\\})" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": " for output " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "(\\mathcal{Y} = \\mathcal{A})" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ". In this directed acyclic graph, the starting and ending points represent the cause and effect, respectively. Traditional VLN methods focus on learning the observational association " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}|\\mathcal{X})" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ", overlooking the ambiguity introduced by confounders " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": " in the back-door path " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\gets \\mathcal{Z} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": ". Here, confounders are extraneous variables that influence both causes and effects, e.g., frequently occurring content or specific attributes. " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\rightarrow \\mathcal{X}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": " arises since the combined probability of input data is inevitably affected by the limited resources available in the real world when collection and simulation. Additionally, " + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\rightarrow \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": " exists because collected environments, labeled instructions, or sampled trajectories also affect the probability of action distributions. These confounder links enable spurious shortcuts during training but can be detrimental in new situations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "We propose to distinguish hidden confounders from different modalities into observable and unobservable categories, enhancing our prior knowledge integration and the rationality of assumption. Concretely, observable confounders encompass instances that can be recognized (e.g., room references " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_v^o" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and guiding keywords " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_i^o" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "). In contrast, unobservable confounders consist of intricate patterns and style-related elements that are challenging to qualitatively" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "describe (e.g., decoration style in vision " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "z_v^u" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", sentence pattern in language " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "z_i^u" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", and trajectory trend in history " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "z_h^u" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "). Since we cannot explicitly model unobservable confounders " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^u" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", the additional mediators " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " are inserted between " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " to establish front-door paths " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\to \\mathcal{M} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ". Detailed adjustment methods are introduced in subsequent sections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 156, + 389, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 389, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 389, + 169 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 175, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 248 + ], + "type": "text", + "content": "The overview of the GOAT model is shown in Fig. 3. The proposed back-door and front-door adjustment causal learning modules are detailed in Sec. 4.1 and Sec. 4.2, respectively. The cross-modal feature pooling method is subsequently introduced in Sec. 4.3. Finally, a practical causal learning pipeline is presented in Sec. 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 255, + 465, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 465, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 465, + 268 + ], + "type": "text", + "content": "4.1. Observable Causal Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": "Back-door Adjustment Causal Learning (BACL). Based on Bayes's theorem, the typical observational likelihood is as " + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}|\\mathcal{X}) = \\sum_{z}P(\\mathcal{Y}|\\mathcal{X},z)\\underline{P(z|\\mathcal{X})}" + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "inline_equation", + "content": "P(z|\\mathcal{X})" + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": " could bring biased weights. Do-operator [52] provides scientifically sound methods for determining causal effects by severing the back-door link between " + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": ". According to the invariance and independence rules [17], we have:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 358, + 545, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 358, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 322, + 358, + 545, + 408 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathcal {Y} | d o (\\mathcal {X})) = \\sum_ {z} P (\\mathcal {Y} | d o (\\mathcal {X}), z) P (z | d o (\\mathcal {X})) (1) \\\\ = \\sum_ {z} P (\\mathcal {Y} | \\mathcal {X}, z) \\underline {{P (z)}} (2) \\\\ \\end{array}", + "image_path": "4b6717639008a936073b632c42552aed60b25315038342197c058244f241899a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "text", + "content": "In such case, the intervention is achieved by blocking the back-door path " + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\to \\mathcal{X}" + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "text", + "content": ", making " + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "text", + "content": " have a fair opportunity to incorporate causality-related factors for prediction. Previous methods [42, 73, 78] used the NWGM approximation [6] to directly pursue causal learning to the final outputs. However, these methods limit the intervention only to the network's final Softmax layer, overlooking possible biased features in shallow layers. Since the conditional probability is implicit in pattern recognition made by the trained neural network [48], we release the target effect of causal hypothesis to learned features rather than merely outputs. Obtaining unbiased features leads to unbiased predictions. Consequently, the specific network module is formulated as " + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "inline_equation", + "content": "f(x, z)" + }, + { + "bbox": [ + 304, + 409, + 545, + 577 + ], + "type": "text", + "content": ", and the implementation of Eq. (2) becomes:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 578, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 578, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 377, + 578, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {B} (\\boldsymbol {x}, \\boldsymbol {z}) = \\mathbb {E} _ {z} [ f (\\boldsymbol {x}, \\boldsymbol {z}) ] \\tag {3}", + "image_path": "c5d27d46f578e14a2d62993f94ba570797617512ee64b39d935ed7abfa700921.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "content": "The linear function is used as " + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}, \\pmb{z}) = f_{x}(\\pmb{x}) + f_{z}(\\pmb{z})" + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "content": ". Then Eq. (3) becomes " + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "inline_equation", + "content": "f_{x}(\\pmb{x}) + \\mathbb{E}_{z}[f_{z}(\\pmb{z})]" + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "content": ". To obtain " + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{z}[f_{z}(\\pmb{z})]" + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "content": ", there are two prevalent approaches: statistic-based [42, 66] and attention-based methods [38, 73]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 640, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 640, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 315, + 640, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\text {S t a t i s t i c}: \\mathbb {E} _ {z} [ f _ {z} (\\boldsymbol {z}) ] = \\sum_ {i} \\frac {\\left| z _ {i} \\right|}{\\sum_ {j} \\left| z _ {j} \\right|} f _ {z} \\left(\\boldsymbol {z} _ {i}\\right) \\tag {4}", + "image_path": "277ebf3e3a386e6c98be448aa69f86b80488b2e3de8044acbb499967e36a84b5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 670, + 545, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 670, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 315, + 670, + 545, + 700 + ], + "type": "interline_equation", + "content": "\\text {A t t e n t i o n}: \\mathbb {E} _ {z} [ f _ {z} (z) ] = \\sum_ {i} \\frac {\\exp \\left(h z _ {i} ^ {T}\\right)}{\\sum_ {j} \\exp \\left(h z _ {j} ^ {T}\\right)} f _ {z} (z _ {i}) \\tag {5}", + "image_path": "27b907f08a28cb601acb4c4c7e7aa00fac9fe5aa28b311c462c36d7492622179.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "|z_{i}|" + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "text", + "content": " denotes the number of " + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "text", + "content": " belonging to the " + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 309, + 701, + 545, + 714 + ], + "type": "text", + "content": "-th" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13141" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 186 + ], + "type": "image", + "image_path": "80ffc652faa72e6959d9435f02acb9c430c4b3d49abf57a9dc5dd3bb25c1b670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "type": "text", + "content": "Figure 3. Framework of GOAT, built on the foundation of the dual-scale graph transformer [9]. Back-door and front-door adjustment causal learning mechanisms are used for mitigating spurious correlations, enabling unbiased feature learning and decision-making." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 239, + 286, + 305 + ], + "blocks": [ + { + "bbox": [ + 50, + 230, + 216, + 239 + ], + "lines": [ + { + "bbox": [ + 50, + 230, + 216, + 239 + ], + "spans": [ + { + "bbox": [ + 50, + 230, + 216, + 239 + ], + "type": "text", + "content": "(a) Back-door Adjustment Causal Learning (BACL)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 239, + 286, + 305 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 286, + 305 + ], + "type": "image", + "image_path": "f3e102ecfb78a6d6c64bc12365dccbe70f7904291db8ecde04268a23e05085d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 315, + 286, + 380 + ], + "blocks": [ + { + "bbox": [ + 50, + 306, + 217, + 315 + ], + "lines": [ + { + "bbox": [ + 50, + 306, + 217, + 315 + ], + "spans": [ + { + "bbox": [ + 50, + 306, + 217, + 315 + ], + "type": "text", + "content": "(b) Front-door Adjustment Causal Learning (FACL)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 315, + 286, + 380 + ], + "lines": [ + { + "bbox": [ + 50, + 315, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 50, + 315, + 286, + 380 + ], + "type": "image", + "image_path": "d7b42454595816a909fcf839037ac9659893f9dbeaa94fa97b0753d094f8bc40.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 383, + 244, + 393 + ], + "lines": [ + { + "bbox": [ + 89, + 383, + 244, + 393 + ], + "spans": [ + { + "bbox": [ + 89, + 383, + 244, + 393 + ], + "type": "text", + "content": "Figure 4. Illustration of BACL and FACL." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "content": "category in the training set, and " + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "content": " means hidden features. The illustration is shown in Fig. 4(a). Our experiments in Sec. 5.3 reveal modality-specific calculation preferences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "content": "BACL in Text Content. In VLN instructions like \"Exit the office and turn right into the kitchen,\" essential guiding elements such as directions (e.g., \"exit\" and \"right\") and landmarks (e.g., \"office\" and \"kitchen\") play significant roles. These keywords which are common causes of instruction construction and action distribution, serve as observable confounders. Firstly, we build the text keyword dictionary " + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_I^O = [z_{i,1}^o,z_{i,2}^o,\\dots,z_{i,K}^o ]" + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "content": " classes to store confounder features. Direction-and-landmark keywords are extracted based on their part-of-speech tags [63]. We use the pre-trained RoBERTa [41] to obtain feature representations for each extracted token " + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{i}^{o}" + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "content": ". Since the same word can have different features across sentences, we calculate the average feature for each keyword: " + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{i,n}^{o} = \\frac{1}{|z_{i,n}^{o}|}\\sum_{j}\\pmb{f}_{i,n,j}^{o}" + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "content": ". Subsequently, the text content causal representation " + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 46, + 442, + 287, + 635 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 636, + 286, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 636, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 52, + 636, + 286, + 651 + ], + "type": "interline_equation", + "content": "I = \\operatorname {R o B E R T a} \\left(\\psi_ {t} (\\mathcal {I}) + \\psi_ {t} (\\mathcal {P})\\right), Z _ {k} = \\mathrm {L N} \\left(\\phi_ {k} \\left(\\mathcal {D} _ {I} ^ {O}\\right)\\right) \\tag {6}", + "image_path": "f82ab36bf666f22771bd25816cbe3ac4acc6dcca955e0acc0e241477c6f59c8f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 652, + 286, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 652, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 117, + 652, + 286, + 665 + ], + "type": "interline_equation", + "content": "G _ {s} = \\operatorname {L N} \\left[ \\phi_ {i} \\left(\\mathcal {B} \\left(I, Z _ {k}\\right)\\right) \\right] \\tag {7}", + "image_path": "415f4d9b70dbf99a9e337146cfbf8a5efec9b323712bd96608440461b9895ba4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " denote the learnable embedding layer and the full-connection layer, respectively. The absolute positional encoding " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " [60] is added to present the position information, and the layer normalization LN [5] is employed" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 232, + 484, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 484, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 484, + 243 + ], + "type": "text", + "content": "for stabilizing hidden states during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": "BACL in Vision Content. For each step, the panorama " + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": " is divided into 36 sub-images. Since existing VLN datasets primarily involve indoor room navigation, visual room references are treated as observable confounders. CLIP [56] is used to extract image features. Since room labels are not directly provided, we employ BLIP [32], a pre-trained VQA model to capture the room information for each image, by querying the model with a fixed prompt \"what kind of room is this?\" The average value of each room reference type is calculated, forming a visual room reference dictionary " + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_V^O = [z_{v,1}^o,z_{v,2}^o,\\dots,z_{v,M}^o]" + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": " is the number of room types. Additionally, the matrix " + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\gamma = \\left[(\\sin \\theta_i,\\cos \\theta_i,\\sin \\eta_i,\\cos \\eta_i)_{i = 1}^{36}\\right]" + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": " is used to present the direction of each image's shift relative to the agent, where " + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 244, + 545, + 471 + ], + "type": "text", + "content": " denote the heading and elevation direction. If there are additional object features (for goal-oriented tasks), they are concatenated with image features. Subsequently, a 2-layer transformer encoder is used to capture spatial dependencies. The above process is formulated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 472, + 545, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 472, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 353, + 472, + 545, + 486 + ], + "type": "interline_equation", + "content": "V = \\operatorname {C L I P} (\\mathcal {V}), Z _ {r} = \\ln \\left(\\phi_ {r} \\left(\\mathcal {D} _ {V} ^ {O}\\right)\\right) \\tag {8}", + "image_path": "d4d1ae4aff56fb61c8aa86421dd5a8031745bb5bcf6b7384eb11a71a12c88980.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 487, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 487, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 376, + 487, + 545, + 500 + ], + "type": "interline_equation", + "content": "V _ {v} = \\operatorname {L N} \\left[ \\phi_ {v} \\left(\\mathcal {B} (V, Z _ {r})\\right) \\right] \\tag {9}", + "image_path": "951d62e8e86a25f20dc04865954405a7b241dbb46aaef411a941d5cee80d16aa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 502, + 545, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 502, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 372, + 502, + 545, + 516 + ], + "type": "interline_equation", + "content": "Q _ {v} = \\operatorname {T r a n s} \\left(V _ {v} + \\psi_ {d} (\\gamma)\\right) \\tag {10}", + "image_path": "77d611c754e4e3be7689edcce4927c38957d6879627bda9792bd4d6726639a7e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 526, + 476, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 476, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 476, + 538 + ], + "type": "text", + "content": "4.2. Unobservable Causal Inference" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Front-door Adjustment Causal Learning (FACL). In the previous section, we employed the back-door adjustment technique to handle bias caused by observable confounders. However, there are additional unobservable confounders that cannot be explicitly captured and modeled in advance. To address this, we introduce another technique - front-door adjustment [17]. As shown in Fig. 4(b), an additional mediator " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " is inserted between inputs and outcomes to construct the front door path " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\rightarrow \\mathcal{M} \\rightarrow \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ". In VLN, an attention-based model " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}|\\mathcal{X}) = \\sum_{m} P(\\mathcal{Y}|m)P(m|\\mathcal{X})" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " will select key regions " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " from inputs " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " for action prediction " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ". Therefore, the model inference can be represented by two parts: the feature selector " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\rightarrow \\mathcal{M}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " which selects suitable knowledge " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", and the action predictor" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "13142" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\to \\mathcal{V}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": " which exploits " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": ". To eliminate spurious correlation brought by unobservable confounder " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": ", we simultaneously deploy " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "do" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "-operator to " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 108, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 108, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 108, + 287, + 175 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathcal {Y} | d o (\\mathcal {X})) = \\sum_ {m} P (\\mathcal {Y} | d o (m)) P (m | d o (\\mathcal {X})) (11) \\\\ = \\sum_ {x ^ {\\prime}} P \\left(x ^ {\\prime}\\right) \\sum_ {m} P (\\mathcal {Y} | m, x ^ {\\prime}) P (m | \\mathcal {X}) (12) \\\\ = \\mathbb {E} _ {x ^ {\\prime}} \\mathbb {E} _ {m | x} [ P (\\mathcal {Y} | x ^ {\\prime}, m) ] (13) \\\\ \\end{array}", + "image_path": "53d22f215dccd63edd7304a9e2cb852db7c57bcea4bba3051baba11447185dd5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": " denotes potential input samples of the whole representation space, different from current inputs " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = x" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": ". We use the bold symbol " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\pmb{m}" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": " to denote the in-sampling features obtained by the feature extractor acting on the current input, and " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\pmb{x}'" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": " to mean the cross-sampling features randomly sampled by the K-means-based feature selector from the entire training samples. Based on the linear mapping model, Eq. (13) becomes " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{m|x}[m] + \\mathbb{E}_{x'}[x']" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": ". As it is intractable to get a closed-form solution of expectations involving the complex representation space, the estimation is achieved by the query mechanism. Two embedding functions [42, 73] are used to transmit input " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": " into two query sets " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "g_{1} = q_{1}(\\pmb{x})" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "inline_equation", + "content": "g_{2} = q_{2}(\\pmb{x})" + }, + { + "bbox": [ + 47, + 176, + 287, + 330 + ], + "type": "text", + "content": ". Then, the front-door adjustment is approximated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 335, + 287, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 335, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 287, + 363 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {x} ^ {\\prime}} \\left[ \\boldsymbol {x} ^ {\\prime} \\right] \\approx \\sum_ {\\boldsymbol {x} ^ {\\prime}} P \\left(\\boldsymbol {x} ^ {\\prime} \\mid \\boldsymbol {g} _ {1}\\right) \\boldsymbol {x} ^ {\\prime} = \\sum_ {i} \\frac {\\exp \\left(\\boldsymbol {g} _ {1} \\boldsymbol {x} _ {i} ^ {\\prime T}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {g} _ {1} \\boldsymbol {x} _ {j} ^ {\\prime T}\\right)} \\boldsymbol {x} _ {i} ^ {\\prime} \\tag {14}", + "image_path": "8b829421f357102f969b478408a14888ca850b3f6326bef70bbe6e58608e072e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 365, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {m \\mid x} [ \\boldsymbol {m} ] \\approx \\sum_ {m} P (\\boldsymbol {m} | \\boldsymbol {g} _ {2}) \\boldsymbol {m} = \\sum_ {i} \\frac {\\exp \\left(\\boldsymbol {g} _ {2} \\boldsymbol {m} _ {i} ^ {T}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {g} _ {2} \\boldsymbol {m} _ {j} ^ {T}\\right)} \\boldsymbol {m} _ {i} (15) \\\\ \\mathcal {F} \\left(\\boldsymbol {x}, \\boldsymbol {x} ^ {\\prime}\\right) = \\mathbb {E} _ {\\boldsymbol {x} ^ {\\prime}} \\left[ \\boldsymbol {x} ^ {\\prime} \\right] + \\mathbb {E} _ {m | x} [ \\boldsymbol {m} ] (16) \\\\ \\end{array}", + "image_path": "404a68fb6c4758b28e70a59543e16a18fe4637eedfc98226028346abae46d98f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 416, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 287, + 462 + ], + "type": "text", + "content": "The above process can be efficiently implemented using the multi-head attention [60], enabling seamless integration of causal adjustments into existing transformer-based frameworks with minimal modifications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": "FACL in Text, Vision, and History. Considering the characteristics of VLN, we propose to eliminate unobservable confounders from three kinds of inputs in VLN, i.e., vision, language, and history. First, following previous graph-based methods [9, 63], we construct the vision sequence " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "G_{v} = \\{[\\mathrm{STOP}], [\\mathrm{MEM}], Q_{v}\\}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": " and the history sequence " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "G_{h} = \\{[\\mathrm{STOP}], [\\mathrm{MEM}], \\{\\bar{Q}_{t}\\}_{t=1}^{T}\\}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": " by adding the additional token for presenting stop and recurrent memory states, respectively. " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\bar{Q}_{t}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": " means the learned weight sum of panoramic features for the " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": "-th step. To condense the lengthy sequence of features and generate global features " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": " for cross-sampling, we devise the CFP module (as described in Sec. 4.3) with the attentive pooling mechanism to construct confounder dictionaries for vision, history, and instruction, denoted as " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "D_{v}^{U}, D_{h}^{U}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "D_{s}^{U}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": ", respectively. Then the causality-enhanced features " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "R_{v}, R_{h}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 47, + 463, + 287, + 651 + ], + "type": "text", + "content": " are calculated based on Eq. (16):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 657, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 657, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 59, + 657, + 287, + 681 + ], + "type": "interline_equation", + "content": "R _ {v} = \\mathcal {F} \\left(G _ {v}, D _ {v} ^ {U}\\right), R _ {h} = \\mathcal {F} \\left(G _ {h}, D _ {h} ^ {U}\\right), R _ {s} = \\mathcal {F} \\left(G _ {s}, D _ {s} ^ {U}\\right) \\tag {17}", + "image_path": "31cbfca9140e5a0070acd97b6812396da75f51e70654441ceb75e4241eaebf57.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Furthermore, we introduce an adaptive gate fusion (AGF) method to enhance the stability of learning by integrating" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 153 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 153 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 153 + ], + "type": "image", + "image_path": "c384ce9f20c3f5aa2f54af95e127afce576f5343224f778a1cfcf400a372ad89.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 156, + 539, + 168 + ], + "lines": [ + { + "bbox": [ + 310, + 156, + 539, + 168 + ], + "spans": [ + { + "bbox": [ + 310, + 156, + 539, + 168 + ], + "type": "text", + "content": "Figure 5. Illustration of the cross-modal feature pooling (CFP)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 179, + 544, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 179, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 544, + 202 + ], + "type": "text", + "content": "causality-enhanced features with the original context features for each modality:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 204, + 544, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 204, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 364, + 204, + 544, + 217 + ], + "type": "interline_equation", + "content": "\\omega_ {x} = \\delta \\left(R _ {x} W _ {x} + G _ {x} W _ {g} + b\\right) \\tag {18}", + "image_path": "227c03f08f313ccc95cb5365c296b9e3be7c39d1f1820186b84f45313f4cf944.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 219, + 544, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 219, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 357, + 219, + 544, + 232 + ], + "type": "interline_equation", + "content": "U _ {x} = \\omega_ {x} \\odot R _ {x} + (1 - \\omega_ {x}) \\odot G _ {x} \\tag {19}", + "image_path": "45955088e086558f2b113ae30c0344afe687ae4e434ac3564d16228a60450e5e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " mean the Sigmoid function and elementwise multiplication. Suppose " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "R_{x}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "G_{x} \\in \\mathbb{R}^{L_{x} \\times d_{h}}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "W_{x/g} \\in \\mathbb{R}^{d_{h} \\times 1}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}^{L_{x} \\times 1}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " are learnable parameters. Next, the cross-modal fused local features " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " and global features " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " are obtained by the cross-attention encoders " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " from METER [15]. The dynamic fusion " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{DF}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " [9] followed by Softmax " + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{SF}" + }, + { + "bbox": [ + 305, + 232, + 545, + 316 + ], + "type": "text", + "content": " is applied for action prediction:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 316, + 544, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 316, + 544, + 330 + ], + "spans": [ + { + "bbox": [ + 341, + 316, + 544, + 330 + ], + "type": "interline_equation", + "content": "F _ {l} = \\mathcal {C} \\left(U _ {v}, U _ {s}, U _ {s}\\right), F _ {g} = \\mathcal {C} \\left(U _ {h}, U _ {s}, U _ {s}\\right) \\tag {20}", + "image_path": "9b9445737a2cec1109121fc2a8ea549464cf58d05737da38343fd57a0656229a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 331, + 544, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 331, + 544, + 344 + ], + "spans": [ + { + "bbox": [ + 356, + 331, + 544, + 344 + ], + "type": "interline_equation", + "content": "F _ {f} = \\mathcal {D F} \\left(F _ {l}, F _ {g}\\right), y _ {t} = \\mathcal {S F} \\left(F _ {f}\\right) \\tag {21}", + "image_path": "1ae888d5ad049cc763e95911499c24ba2e9ba34237057d2f6cc4d8f79e0890f7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 345, + 544, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 544, + 357 + ], + "type": "text", + "content": "The cross-entropy loss is used to optimize the network:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 349, + 366, + 544, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 366, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 349, + 366, + 544, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = \\sum_ {t = 1} ^ {T} - \\log P \\left(y _ {t} ^ {*} \\mid \\mathcal {I}, \\mathcal {V} _ {t}, \\mathcal {H} _ {1: t - 1}\\right) \\tag {22}", + "image_path": "dd7764d59eb74ea0bf492a65047cd4f576fd41e4825a5a57d6c9e4d78a336f9e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 403, + 465, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 465, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 465, + 417 + ], + "type": "text", + "content": "4.3. Cross-modal Feature Pooling" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 422, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 506 + ], + "type": "text", + "content": "One challenge of implementing the front-door adjustment in VLN is constructing efficient dictionaries for global features from long sequences. This requires compressing sequential features of varying lengths into a unified feature space to represent each sample effectively. Let " + }, + { + "bbox": [ + 304, + 422, + 545, + 506 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^{L \\times d_h}" + }, + { + "bbox": [ + 304, + 422, + 545, + 506 + ], + "type": "text", + "content": " be the sequential features, the following attentive pooling is used to effectively compress the sequence length:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 325, + 506, + 544, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 506, + 544, + 520 + ], + "spans": [ + { + "bbox": [ + 325, + 506, + 544, + 520 + ], + "type": "interline_equation", + "content": "A = \\mathcal {T} (H), \\alpha = \\mathcal {S F} \\left(A W _ {a}\\right), \\bar {H} = \\mathcal {T} \\left(\\alpha^ {T} H\\right) \\tag {23}", + "image_path": "db8fb0a87fed6f3d9697768879b2bff9b2d9671a7ada011f65a1722961073cc7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": " denote the Tanh activation, " + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "inline_equation", + "content": "W_{a} \\in \\mathbb{R}^{d_{h} \\times 1}" + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": " is the learnable attention matrix, and " + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\bar{H} \\in \\mathbb{R}^{1 \\times d_{h}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": ". As shown in Fig. 5, for vision, history, local-global fusion, and text features, we use one transformer layer as the encoder followed by the attentive pooling to obtain flattened features " + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\bar{C}^{l}, \\bar{C}^{g}, \\bar{C}^{f}" + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\bar{C}^{s}" + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": ", respectively. Then, we adopt contrastive learning [35, 57, 62] to optimize this cross-modal feature pooling (CFP) module, meanwhile improving semantic alignments for different modalities. The contrastive loss " + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ls}" + }, + { + "bbox": [ + 304, + 521, + 545, + 642 + ], + "type": "text", + "content": " is constructed as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 324, + 647, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 647, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 324, + 647, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {l s} = - \\frac {1}{2 B} \\sum_ {j = 1} ^ {B} \\log \\frac {\\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {j} ^ {s} \\right\\rangle / t\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)} \\tag {24} \\\\ - \\frac {1}{2 B} \\sum_ {k = 1} ^ {B} \\log \\frac {\\exp \\left(\\left\\langle \\bar {C} _ {k} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left\\langle \\bar {C} _ {j} ^ {l} , \\bar {C} _ {k} ^ {s} \\right\\rangle / t\\right)} \\\\ \\end{array}", + "image_path": "99c58d9b43a28de892e3c25d0cdb0d8bb85977ddb8106b9e2b9dca3cb228cf0b.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13143" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " mean the batch size and temperature, respectively. Similarly, contrastive losses " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gs}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fs}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " are calculated by replacing " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\bar{C}^l" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\bar{C}^g" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\bar{C}^f" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". The overall CFP loss is the sum of these losses " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CFP}} = \\mathcal{L}_{ls} + \\mathcal{L}_{gs} + \\mathcal{L}_{fs}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 289, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 253 + ], + "type": "text", + "content": "To enable the network more adaptive to characteristics for VLN and thus facilitate the building of the front-door confounder dictionaries for samples, we train the CFP module alongside other auxiliary tasks [9] during pre-training. Subsequently, the trained attentive pooling modules are used to extract global features for different modalities. In the fine-tuning stage, we employ the BACL and FACL with established dictionaries for intervention. The CFP offers dual benefits: it aligns different modalities more effectively during pre-training and provides a systematic approach for extracting coherent representations from sequence inputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 261, + 188, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 188, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 188, + 274 + ], + "type": "text", + "content": "4.4. Causal Learning Pipeline" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 171, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 171, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 171, + 495 + ], + "type": "text", + "content": "As shown in Fig. 6, we summarize a causal learning pipeline that serves as a blueprint for similar learning-based methods. First, it begins with task formulation, where the specific task and its objectives are precisely defined. Next, the observable and unobservable confounders are explicitly assumed. Both back-door and front-door adjustment strategies are employed to tackle these confounders, either simultaneously or sequentially, contingent on task specifics. Then, the pipeline proceeds" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 190, + 292, + 287, + 459 + ], + "blocks": [ + { + "bbox": [ + 190, + 292, + 287, + 459 + ], + "lines": [ + { + "bbox": [ + 190, + 292, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 190, + 292, + 287, + 459 + ], + "type": "image", + "image_path": "275afebef1bbe4e948e4a9c031740d79c720e486ebf20ca5a47b6ae66964b3e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 462, + 287, + 484 + ], + "lines": [ + { + "bbox": [ + 188, + 462, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 188, + 462, + 287, + 484 + ], + "type": "text", + "content": "Figure 6. Pipeline of the causal learning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": "to model calculation and result prediction. Throughout network optimization, both network parameters and confounder features are continuously updated. Ultimately, this iterative process leads us to the development of a robust causal model capable of generating unbiased features, thereby advancing the generalizability of AI systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 128, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 593 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 598, + 174, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 174, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 174, + 612 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "1) Datasets. We verify GOAT on two kinds of VLN benchmarks: fine-grained datasets (R2R [4] and RxR-English [29]), which provide long step-by-step navigation instructions, and goal-oriented datasets (REVERIE [53] and SOON [81]), which additionally requires for the target object. Formally, the datasets are partitioned into four splits: training, validation seen (sharing the same environments with the training set), validation unseen (having different" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "environments from the training set), and test unseen sets (reported by the online leaderboard for fair comparison)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 96, + 547, + 443 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 304, + 96, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 180 + ], + "type": "text", + "content": "2) Evaluation Metrics. In R2R, key metrics include Navigation Error (NE), Success Rate (SR), Oracle SR (OSR), and SR Weighted by Path Length (SPL). RxR adds Normalized Dynamic Time Warping (nDTW) and SR Weighted by Dynamic Time Warping (sDTW). REVERIE and SOON introduce Remote Grounding Success Rate (RGS) and RGS Weighted by Path Length (RGSPL)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "text", + "content": "3) Implementation Details. Our model consists of 6 transformer layers for text, 2 for panorama, and 3 for cross-modal encoding. We use CLIP-B/16 [56] for image feature extraction and initialize network weights with METER [15]. In pre-training, MLM [28], SAP [8], and the proposed CFP are for R2R and RxR. OG [37] is added for REVERIE and SOON. EnvEdit [34] is employed for feature augmentation. The synthetic extended datasets [19, 62, 65] are used for R2R, REVERIE, and RxR, respectively. Pretraining is done on a single Tesla V100 GPU for a maximum of 300K iterations by the AdamW [43] optimizer, with a batch size of 48 and a learning rate of " + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "text", + "content": ". The numbers of classes of keywords and rooms are 74 and 50, and the temperature " + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "text", + "content": " is set to 1. In fine-tuning, the front-door dictionaries are randomly sampled from the K-Means clustering features. As the text transformer RoBERTa [41] is involved in end-to-end training, the textual keywords dictionary is also iteratively updated. Speaker models with the environmental dropout [58, 62, 64] are used to provide dynamic pseudo labels. We employ batch size 12 for R2R, REVERIE, and 5 for RxR and SOON, with a learning rate of " + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 180, + 546, + 443 + ], + "type": "text", + "content": " and a maximum of 100K iterations." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 452, + 497, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 497, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 497, + 464 + ], + "type": "text", + "content": "5.2. Comparisons with State-of-the-Arts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": "In Tab. 1, 2, 3, 4, we compare GOAT with the previous state-of-the-art (SoTA) methods on the R2R, REVERIE, RxR-English, and SOON datasets, respectively. On all these four datasets, our approach exhibits superior navigation performance, precise instruction-following alignment, and accurate object grounding across both seen and unseen environments. For instance, in R2R, GOAT achieves remarkable improvements in SPL compared to BEVBert [3], with relative increases of " + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "inline_equation", + "content": "7.41\\%" + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "inline_equation", + "content": "6.45\\%" + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "inline_equation", + "content": "4.74\\%" + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": " on three subsets. In REVERIE, GOAT shows substantial relative enhancements in RGSPL by " + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "inline_equation", + "content": "11.83\\%" + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "inline_equation", + "content": "6.55\\%" + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "inline_equation", + "content": "20.87\\%" + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": " on three subsets. In the challenging SOON and RxR tasks, GOAT also exhibits significant improvements in performance metrics, highlighting its robustness and superior generalization capabilities over previous methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 658, + 430, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 430, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 430, + 671 + ], + "type": "text", + "content": "5.3.Quantitative Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 547, + 713 + ], + "type": "text", + "content": "1) Effect of Causal Inference. Fig. 7 verifies the impact of causal inference on GOAT across four diverse VLN datasets in unseen environments. \"W/o intervention\" signifies the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13144" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 521, + 201 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 521, + 201 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 521, + 201 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 521, + 201 + ], + "type": "table", + "html": "
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑SR↑SPL↑NE↓OSR↑
HAMT [8]76722.518266613.297365603.9372
DUET [9]79732.288672603.318169593.6576
TD-STP [80]77732.348370633.227667613.7372
GeoVLN [25]79762.22-68633.35-65613.95-
DSRG [63]81762.238873623.008172613.3378
GridMM [71]----75642.83-73623.35-
GELA [10]76732.39-71653.11-67623.59-
EnvEdit [34]77742.32-69643.24-68643.59-
BEVBert [3]81742.178875642.818473623.1381
GOAT (Ours)83.7479.481.7988.6477.8268.132.4084.7274.5764.943.0480.35
", + "image_path": "0312d508b57cf705b0c29ddb4353bda04b6f44dd9143bc1afb4dea0791df113a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 224, + 538, + 326 + ], + "blocks": [ + { + "bbox": [ + 103, + 201, + 488, + 213 + ], + "lines": [ + { + "bbox": [ + 103, + 201, + 488, + 213 + ], + "spans": [ + { + "bbox": [ + 103, + 201, + 488, + 213 + ], + "type": "text", + "content": "Table 1. Comparison with other state-of-the-art methods on the R2R dataset [4]. '-': unavailable statistics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 224, + 538, + 326 + ], + "lines": [ + { + "bbox": [ + 53, + 224, + 538, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 538, + 326 + ], + "type": "table", + "html": "
MethodValidation SeenValidation UnseenTest Unseen
SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑SR↑SPL↑RGS↑RGSPL↑
HAMT [8]43.2940.1927.2025.1832.9530.2018.9217.2830.4026.6714.8813.08
HOP+ [54]55.8749.5540.7636.2236.0731.1322.4919.3333.8228.2420.2016.86
DUET [9]71.7563.9457.4151.1446.9833.7332.1523.0352.5136.0631.8822.06
DSRG [63]75.6968.0961.0754.7247.8334.0232.6923.3754.0437.0932.4922.18
GridMM [71]----51.3736.4734.5724.5653.1336.6034.8723.45
BEVBert [3]73.7265.3257.7051.7351.7836.3734.7124.4452.8136.4132.0622.09
GOAT (Ours)78.6471.4063.7457.8553.3736.7038.4326.0957.7240.5338.3226.70
", + "image_path": "151ac0f56fd36ec88f940dc3de6f750d6c4ca981619f12ad98b87a9dfb9f4b00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 357, + 287, + 456 + ], + "blocks": [ + { + "bbox": [ + 90, + 327, + 501, + 338 + ], + "lines": [ + { + "bbox": [ + 90, + 327, + 501, + 338 + ], + "spans": [ + { + "bbox": [ + 90, + 327, + 501, + 338 + ], + "type": "text", + "content": "Table 2. Comparison with other state-of-the-art methods on the REVERIE dataset [53]. '-': unavailable statistics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 357, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 48, + 357, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 357, + 287, + 456 + ], + "type": "table", + "html": "
MethodValidation SeenValidation Unseen
SR↑SPL↑nDTW↑sDTW↑SR↑SPL↑nDTW↑sDTW↑
Syntax [31]48.144.058.040.039.235.052.032.0
SOAT [45]----44.2-54.836.4
HOP+ [54]53.647.959.043.045.738.452.036.0
FOAM [14]----42.838.754.135.6
ADAPT [35]50.344.656.340.646.940.254.137.7
\\( \\mathsf{MAR}_{\\mathsf{M-MP}} \\) [27]----50.2-60.343.9
VLN-PETL [55]60.556.865.751.757.954.264.949.7
GOAT (Ours)74.168.171.061.468.261.767.156.6
", + "image_path": "7b86268a41c7bc8f58b50f5ee8f8c9a4982dcaa7502c200d46c27e546a06316e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 485, + 287, + 550 + ], + "blocks": [ + { + "bbox": [ + 67, + 458, + 265, + 469 + ], + "lines": [ + { + "bbox": [ + 67, + 458, + 265, + 469 + ], + "spans": [ + { + "bbox": [ + 67, + 458, + 265, + 469 + ], + "type": "text", + "content": "Table 3. Comparison on the RxR-English dataset [29]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 485, + 287, + 550 + ], + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 550 + ], + "type": "table", + "html": "
MethodValidation UnseenTest Unseen
OSR↑SR↑SPL↑RGSPL↑OSR↑SR↑SPL↑RGSPL↑
GBE [81]28.5419.5213.341.1621.4512.909.230.45
DUET [9]50.9136.2822.583.7543.0033.4421.424.17
GridMM [71]53.3937.4624.813.9148.0236.2721.254.15
GOAT (Ours)54.6940.3528.055.8550.6340.5025.186.10
", + "image_path": "c2bbfc5d31cd83483d7a4a76f819f3b8b456f0af6c0215402e843bdbb0fc35ef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 578, + 285, + 665 + ], + "blocks": [ + { + "bbox": [ + 48, + 578, + 285, + 665 + ], + "lines": [ + { + "bbox": [ + 48, + 578, + 285, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 285, + 665 + ], + "type": "image", + "image_path": "c4f989acaa141cb83a38070786678f8a1a76c54893358743b8d9d931f1eadec2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 666, + 277, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 666, + 277, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 277, + 677 + ], + "type": "text", + "content": "Figure 7. Effect of the intervention on various VLN datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "exclusion of the proposed BACL and FACL interventions across all modalities. On each of the datasets examined, the" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 357, + 535, + 417 + ], + "blocks": [ + { + "bbox": [ + 79, + 552, + 254, + 563 + ], + "lines": [ + { + "bbox": [ + 79, + 552, + 254, + 563 + ], + "spans": [ + { + "bbox": [ + 79, + 552, + 254, + 563 + ], + "type": "text", + "content": "Table 4. Comparison on the SOON dataset [81]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 357, + 535, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 357, + 535, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 357, + 535, + 417 + ], + "type": "table", + "html": "
IdBACLFACLSR↑SPL↑NE↓OSR↑
1XX75.2764.692.7283.14
2✓X76.3767.132.6083.99
3X✓76.5066.922.5384.21
4✓✓77.8268.132.4084.72
", + "image_path": "a9aa8a1c2c5ba774097b713babb8fd89ed2bdc4b814d4d90b3b2fa459b157ac0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 419, + 528, + 430 + ], + "lines": [ + { + "bbox": [ + 322, + 419, + 528, + 430 + ], + "spans": [ + { + "bbox": [ + 322, + 419, + 528, + 430 + ], + "type": "text", + "content": "Table 5. Effect of back-door and front-door adjustments." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 441, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 490 + ], + "type": "text", + "content": "integration of causal inference leads to significant enhancements in the model's performance. This strongly demonstrates causal learning's considerable popularization potential in enhancing learning-based model generalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 493, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 493, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 625 + ], + "type": "text", + "content": "2) Effect of BACL and FACL. Tab. 5 analyzes the effects of the proposed BACL and FACL on the R2R val-unseen subset. Compared to the baseline (#1), individual application of either BACL (#2) or FACL (#3) leads to performance improvements. Concurrent use of BACL and FACL (#4) leads to further performance enhancements. These findings underscore our assumption about the presence of both observable and unobservable confounders. Integrating both back-door and front-door adjustments is crucial to comprehensively addressing dataset biases, and enhancing the model's robustness and generalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "content": "3) Effect of CFP. In Tab. 6, we assess the efficacy of the proposed CFP on the R2R val-unseen subset. During pretraining (PT), incorporating CFP as an additional auxiliary task (CFP-P) enhances training performance, improving SR and SPL by " + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "3.11\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "3.02\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "content": " (#A1), respectively. In the fine-tuning (FT) stage, we compare the performance with and without the use of trained attention modules from CFP" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13145" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 287, + 144 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 144 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 144 + ], + "type": "table", + "html": "
StageIdMethodSR↑SPL↑NE↓OSR↑
PTA0w/o CFP-P40.3637.886.5248.87
A1w/ CFP-P43.4740.906.0653.13
FTB0A0 w/o CFP-F75.5665.902.6382.42
B1A1 w/o CFP-F76.6366.172.6384.67
B2A1 w/ CFP-F77.8268.132.4084.72
", + "image_path": "3bd97183c570501cdf3a7e14a584e7771775712d3fe49bb88e512d3b74ad61b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 169, + 274, + 229 + ], + "blocks": [ + { + "bbox": [ + 69, + 146, + 265, + 157 + ], + "lines": [ + { + "bbox": [ + 69, + 146, + 265, + 157 + ], + "spans": [ + { + "bbox": [ + 69, + 146, + 265, + 157 + ], + "type": "text", + "content": "Table 6. Effect of CFP in pre-training and fine-tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 169, + 274, + 229 + ], + "lines": [ + { + "bbox": [ + 61, + 169, + 274, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 274, + 229 + ], + "type": "table", + "html": "
IdTextVisionSR↑SPL↑NE↓OSR↑
1StatsStats75.2264.782.7183.65
2StatsAttn76.5965.392.5685.31
3AttnStats77.8268.132.4084.72
4AttnAttn75.9565.832.6483.91
", + "image_path": "4365e56e31cff8d249a49777282cef7ef194f746777f579ad5f33a855c808aeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 231, + 276, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 231, + 276, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 231, + 276, + 242 + ], + "type": "text", + "content": "Table 7. Effect of Statistic and Attention methods in BACL." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 255, + 287, + 350 + ], + "blocks": [ + { + "bbox": [ + 48, + 255, + 287, + 350 + ], + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 350 + ], + "type": "image", + "image_path": "47a6a564e7234786ba2b7ba4463bd1a41ee5cc62862ef899ec9eb2809a49fb1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 351, + 287, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 351, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 287, + 373 + ], + "type": "text", + "content": "Figure 8. Comparison of the distribution of SR and SPL of instructions involving specific terms on the R2R val-unseen split." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "content": "to extract global features for front-door dictionaries (CFP-F). \"W/o CFP-F\" signifies the use of simple average pooling to compress features from the pre-trained model. #B2 shows that the CFP provides more reliable confounder representations for causal learning (SPL " + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\uparrow 1.96\\%" + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 445, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 612 + ], + "type": "text", + "content": "4) Effect of Different BACL in Different Modalities. Tab. 7 investigates the effect of various combinations of statistic-based and attention-based methods for text and vision in BACL on the R2R val-unseen subset. The results indicate that employing the attention method for text and the statistic method for vision yields the best performance (#3). Intuitively, this can be explained by the structured nature of textual information and the involvement of RoBERTa's end-to-end training, enabling the attention method to effectively capture contextual nuances. Conversely, images lack explicit causality, and CLIP, the image extractor, isn't trained directly for efficiency reasons. Consequently, the statistic method ensures a stable causal learning process, preserving the integrity of vision-related features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 622, + 164, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 164, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 164, + 635 + ], + "type": "text", + "content": "5.4. Qualitative Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "1) Bias Elimination Effect. In Fig. 8, the compactness of the boxes represents concentrated data distribution and reduced variability, while a median line closer to the center signifies even data distribution. It shows that GOAT obtains narrower boxes and more central midlines across diverse objects. This finding showcases that with the integration of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 73, + 324, + 84 + ], + "blocks": [ + { + "bbox": [ + 312, + 73, + 324, + 84 + ], + "lines": [ + { + "bbox": [ + 312, + 73, + 324, + 84 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 324, + 84 + ], + "type": "image", + "image_path": "b891ed21cd13a524a64835a274d050eef13f0d10953fd8fee0fccd7a64edc064.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 74, + 368, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 74, + 368, + 83 + ], + "spans": [ + { + "bbox": [ + 325, + 74, + 368, + 83 + ], + "type": "text", + "content": "Starting point" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 375, + 74, + 388, + 84 + ], + "blocks": [ + { + "bbox": [ + 375, + 74, + 388, + 84 + ], + "lines": [ + { + "bbox": [ + 375, + 74, + 388, + 84 + ], + "spans": [ + { + "bbox": [ + 375, + 74, + 388, + 84 + ], + "type": "image", + "image_path": "b76c982cc7e2998589bc6f30992839df3bccb0424ac074f8dd852fcac86d9bc2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 389, + 74, + 422, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 74, + 422, + 83 + ], + "spans": [ + { + "bbox": [ + 389, + 74, + 422, + 83 + ], + "type": "text", + "content": "Destination" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 74, + 446, + 84 + ], + "blocks": [ + { + "bbox": [ + 432, + 74, + 446, + 84 + ], + "lines": [ + { + "bbox": [ + 432, + 74, + 446, + 84 + ], + "spans": [ + { + "bbox": [ + 432, + 74, + 446, + 84 + ], + "type": "image", + "image_path": "cdc97a6da8f62ed7905910767cf90479b83fa0595f9c90896f98a25bed6f87ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 447, + 74, + 479, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 447, + 74, + 479, + 83 + ], + "spans": [ + { + "bbox": [ + 447, + 74, + 479, + 83 + ], + "type": "text", + "content": "sited point" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 486, + 74, + 498, + 83 + ], + "blocks": [ + { + "bbox": [ + 486, + 74, + 498, + 83 + ], + "lines": [ + { + "bbox": [ + 486, + 74, + 498, + 83 + ], + "spans": [ + { + "bbox": [ + 486, + 74, + 498, + 83 + ], + "type": "image", + "image_path": "ac5893ffc1d6d8a8a2080b95c6c45e3e79dee430a2ca9c97049eaaff89b0863c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 505, + 74, + 540, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 505, + 74, + 540, + 83 + ], + "spans": [ + { + "bbox": [ + 505, + 74, + 540, + 83 + ], + "type": "text", + "content": "visited point" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 312, + 85, + 324, + 96 + ], + "blocks": [ + { + "bbox": [ + 312, + 85, + 324, + 96 + ], + "lines": [ + { + "bbox": [ + 312, + 85, + 324, + 96 + ], + "spans": [ + { + "bbox": [ + 312, + 85, + 324, + 96 + ], + "type": "image", + "image_path": "adb9e57de36b26ee94ec1eafaa75f51e3c88f4498fff801771719b7c12ed4526.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 85, + 366, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 85, + 366, + 94 + ], + "spans": [ + { + "bbox": [ + 325, + 85, + 366, + 94 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 376, + 85, + 388, + 94 + ], + "blocks": [ + { + "bbox": [ + 376, + 85, + 388, + 94 + ], + "lines": [ + { + "bbox": [ + 376, + 85, + 388, + 94 + ], + "spans": [ + { + "bbox": [ + 376, + 85, + 388, + 94 + ], + "type": "image", + "image_path": "5110e0d59417ee3c68d119ed32530a3ae9227ecdd506b1f29d190786e8d5553a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 389, + 85, + 405, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 85, + 405, + 94 + ], + "spans": [ + { + "bbox": [ + 389, + 85, + 405, + 94 + ], + "type": "text", + "content": "DSRG" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 412, + 85, + 418, + 94 + ], + "blocks": [ + { + "bbox": [ + 412, + 85, + 418, + 94 + ], + "lines": [ + { + "bbox": [ + 412, + 85, + 418, + 94 + ], + "spans": [ + { + "bbox": [ + 412, + 85, + 418, + 94 + ], + "type": "image", + "image_path": "3ab80d19cffa611cda098bd44745bba0d2ef722bf0d7ad3a4548ab15b651b3f2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 419, + 85, + 462, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 85, + 462, + 94 + ], + "spans": [ + { + "bbox": [ + 419, + 85, + 462, + 94 + ], + "type": "text", + "content": "GOAT (Ours)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 99, + 366, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 366, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 366, + 107 + ], + "type": "text", + "content": "- Instruction:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 108, + 432, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 432, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 432, + 145 + ], + "type": "text", + "content": "Walk out of the room and go past the stairs and into the bedroom that is straight ahead, go to the bathroom doorway and wait there." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 438, + 94, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 438, + 94, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 438, + 94, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 438, + 94, + 544, + 148 + ], + "type": "image", + "image_path": "558c752f0ac4f1db01869ae62f6f41379dd5aa3eddb12983f197cc3594dac444.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 152, + 367, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 367, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 367, + 161 + ], + "type": "text", + "content": "- Instruction:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 162, + 402, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 402, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 402, + 198 + ], + "type": "text", + "content": "Slightly move over the green chair closest to the corner of the reception desk in the lounge." + } + ] + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 409, + 148, + 544, + 202 + ], + "blocks": [ + { + "bbox": [ + 409, + 148, + 544, + 202 + ], + "lines": [ + { + "bbox": [ + 409, + 148, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 409, + 148, + 544, + 202 + ], + "type": "image", + "image_path": "a6f1139e53657d981f57a2a45b4c77dc05c0e975f6313697287caa05faf5105d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 205, + 527, + 217 + ], + "lines": [ + { + "bbox": [ + 323, + 205, + 527, + 217 + ], + "spans": [ + { + "bbox": [ + 323, + 205, + 527, + 217 + ], + "type": "text", + "content": "Figure 9. Predicted trajectories in unseen environments." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 228, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 263 + ], + "type": "text", + "content": "causal intervention, GOAT significantly reduces prediction bias, thereby enhancing its generalization capability in previously unseen environments." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 265, + 546, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 546, + 419 + ], + "type": "text", + "content": "2) Visualized Trajectories. In Fig. 9, we visualize some predicted trajectories in unseen environments, comparing them with DSRG on R2R and REVERIE datasets. Notably, GOAT precisely captures directional cues like \"straight ahead\" and nuanced instructions like \"closest to the corner\", enabling accurate predictions. These instances highlight the intricate causal connections in VLN tasks, where specific instructions prompt corresponding actions. GOAT's enhanced causal inference capability enables it to generate more reasoned responses aligned with the provided instructions, underscoring the significance of robust causal inference in VLN systems. Please refer to our supplementary material for more detailed discussions and visualizations." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 432, + 378, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 378, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 378, + 445 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 453, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 632 + ], + "type": "text", + "content": "Our work presents GOAT, a novel approach that addresses the dataset bias in VLN from the perspective of causal learning. The back-door and front-door adjustment causal learning (BACL and FACL) mechanisms are proposed to adjust for observable and unobservable confounders, respectively. The cross-modal feature pooling (CFP) module is adopted to promote feature learning and extraction through contrastive learning. The practical causal learning pipeline is presented to illuminate other similar learning-based methods. Experiments on R2R, REVERIE, RxR, and SOON datasets show that GOAT can reasonably discover sequence visual-linguistic causal structures and significantly improve performance. Beyond VLN, the underlying confounder assumption and causal inference principles are generalizable to other similar fields." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This paper is supported by the National Natural Science Foundation of China under Grants (62233013, 62073245, 62173248). Shanghai Science and Technology Innovation Action Plan (22511104900)." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13146" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10044-10054, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5101-5109, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 334 + ], + "type": "text", + "content": "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 6, 7, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 335, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 357 + ], + "type": "text", + "content": "[5] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. ArXiv, abs/1607.06450, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 379 + ], + "type": "text", + "content": "[6] Pierre Baldi and Peter Sadowski. The dropout learning algorithm. Artificial intelligence, 210:78-122, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "text", + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "text", + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 545 + ], + "type": "text", + "content": "[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 5, 6, 7, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[10] Yibo Cui, Liang Xie, Yakun Zhang, Meishan Zhang, Ye Yan, and Erwei Yin. Grounded entity-landmark adaptive pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12043–12053, 2023. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 669 + ], + "type": "text", + "content": "[11] Ronghao Dang, Zhuofan Shi, Liuyi Wang, Zongtao He, Chengju Liu, and Qijun Chen. Unbiased directed object attention graph for object navigation. In Proceedings of the 30th ACM International Conference on Multimedia, page 3617-3627, New York, NY, USA, 2022. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Ronghao Dang, Lu Chen, Liuyi Wang, He Zongtao, Chengju Liu, and Qijun Chen. Multiple thinking achieving meta-ability decoupling for object navigation. In International Conference on Machine Learning (ICML), 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[13] Ronghao Dang, Liuyi Wang, Zongtao He, Shuai Su, Jiagui Tang, Chengju Liu, and Qijun Chen. Search for or navigate to? dual adaptive thinking for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8250-8259, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 193 + ], + "type": "text", + "content": "[14] Zi-Yi Dou and Nanyun Peng. Foam: A follower-aware speaker model for vision-and-language navigation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4332-4340, 2022. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 547, + 260 + ], + "type": "text", + "content": "[15] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 261, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 326 + ], + "type": "text", + "content": "[16] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. Advances in Neural Information Processing Systems, 31, 2018. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "type": "text", + "content": "[17] Madelyn Glymour, Judea Pearl, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 360, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 547, + 415 + ], + "type": "text", + "content": "[18] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 415, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 547, + 470 + ], + "type": "text", + "content": "[19] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13137-13146, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "text", + "content": "[20] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 526, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 570 + ], + "type": "text", + "content": "[21] Zongtao He, Liuyi Wang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Mlanet: Multi-level attention network with sub-instruction for continuous vision-and-language navigation. arXiv preprint arXiv:2303.01396, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[22] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 680 + ], + "type": "text", + "content": "[23] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3055-3067, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[24] Ronghang Hu, Daniel Fried, Anna Rohrbach, Dan Klein, Trevor Darrell, and Kate Saenko. Are you looking? grounding to multiple modalities in vision-and-language navigation." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13147" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6551-6557, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[25] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23212-23221, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 219 + ], + "type": "text", + "content": "[26] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1862-1872, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 297 + ], + "type": "text", + "content": "[27] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "text", + "content": "[28] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 408 + ], + "type": "text", + "content": "[29] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392-4412, 2020. 1, 2, 6, 7, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 454 + ], + "type": "text", + "content": "[30] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "type": "text", + "content": "[31] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1041–1050, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "text", + "content": "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "type": "text", + "content": "[33] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21273–21282, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[34] Jialu Li, Hao Tan, and Mohit Bansal. Envideit: Environment editing for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15407-15417, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[35] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In Pro" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2022. 1, 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "type": "text", + "content": "[36] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8317-8326, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[37] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7036-7045, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "text", + "content": "[38] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18041-18050, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[39] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1644-1654, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "text", + "content": "[40] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "text", + "content": "[41] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[42] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 490, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 545 + ], + "type": "text", + "content": "[44] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In European Conference on Computer Vision, pages 259-274. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "type": "text", + "content": "[45] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. Advances in Neural Information Processing Systems, 34:7357-7367, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "type": "text", + "content": "[46] Laksh Nanwani, Anmol Agarwal, Kanishk Jain, Raghav Prabhakar, Aaron Monis, Aditya Mathur, Krishna Murthy, Abdul Hafez, Vineet Gandhi, and K Madhava Krishna. Instance-level semantic maps for vision language navigation. arXiv preprint arXiv:2305.12363, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[47] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13148" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "on Natural Language Processing (EMNLP-IJCNLP), pages 684-695, 2019. 16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[48] Siqi Nie, Meng Zheng, and Qiang Ji. The deep regression bayesian network and its applications: Probabilistic deep learning for computer vision. IEEE Signal Processing Magazine, 35(1):101-111, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[49] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "text", + "content": "[50] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Javen Qinfeng Shi, and Anton Van den Hengel. Counterfactual vision-and-language navigation: Unravelling the unseen. Advances in Neural Information Processing Systems, 33:5296-5307, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 286, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 264 + ], + "type": "text", + "content": "[51] Judea Pearl. Causality. Cambridge university press, 2009. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 287 + ], + "type": "text", + "content": "[52] Judea Pearl and Dana Mackenzie. The book of why: the new science of cause and effect. Basic books, 2018. 2, 3, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 289, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 354 + ], + "type": "text", + "content": "[53] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 3, 6, 7, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 356, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 356, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 356, + 287, + 409 + ], + "type": "text", + "content": "[54] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[55] Yanyuan Qiao, Zheng Yu, and Qi Wu. Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[57] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? In International Conference on Learning Representations. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 645 + ], + "type": "text", + "content": "[58] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, 2019. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[59] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning, pages 394-406. PMLR, 2020. 16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 93 + ], + "type": "text", + "content": "information processing systems, pages 5998-6008, 2017. 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "text", + "content": "[61] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15471-15481, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "type": "text", + "content": "[62] Liuyi Wang, Zongtao He, Ronghao Dang, Huiyi Chen, Chengju Liu, and Qijun Chen. Res-sts: Referring expression speaker via self-training with scorer for goal-oriented vision-language navigation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[63] Liuyi Wang, Zongtao He, jiagui Tang, Ronghao Dang, naijia Wang, Chengju Liu, and Qijun Chen. A dual semantic-aware recurrent global-adaptive network for vision-and-language navigation. In International Joint Conferences on Artificial Intelligence (IJCAI), 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 273, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 316 + ], + "type": "text", + "content": "[64] Liuyi Wang, Chengju Liu, Zongtao He, Shu Li, Qingqing Yan, Huiyi Chen, and Qijun Chen. Pasts: Progress-aware spatio-temporal transformer speaker for vision-and-language navigation. arXiv preprint arXiv:2305.11918, 2023. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 317, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 392 + ], + "type": "text", + "content": "[65] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[66] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10760-10770, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "type": "text", + "content": "[67] Tan Wang, Chang Zhou, Qianru Sun, and Hanwang Zhang. Causal attention for unbiased visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3091-3100, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 483, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 536 + ], + "type": "text", + "content": "[68] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Wang, and Lei Zhang. Vision-language navigation policy learning and adaptation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "text", + "content": "[69] Yuqing Wang, Xiangxian Li, Zhuang Qi, Jingyu Li, Xuelong Li, Xiangxu Meng, and Lei Meng. Meta-causal feature learning for out-of-distribution generalization. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 530-545. Springer, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[70] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12009-12020, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[71] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15625-15636, 2023. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13149" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 653 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[73] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[74] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "type": "text", + "content": "[75] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. Advances in neural information processing systems, 33:2734-2746, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 295 + ], + "type": "text", + "content": "[76] Hua Zhang, Liqiang Xiao, Xiaochun Cao, and Hassan Foroosh. Multiple adverse weather conditions adaptation for object detection via causal intervention. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[77] Jiazhao Zhang, Liu Dai, Fanpeng Meng, Qingnan Fan, Xuelin Chen, Kai Xu, and He Wang. 3d-aware object goal navigation via simultaneous exploration and identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6672-6682, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "text", + "content": "[78] Shengyu Zhang, Tan Jiang, Tan Wang, Kun Kuang, Zhou Zhao, Jianke Zhu, Jin Yu, Hongxia Yang, and Fei Wu. Devlbert: Learning deconfounded visio-linguistic representations. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4373–4382, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[79] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 508 + ], + "type": "text", + "content": "[80] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4194-4203, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 573 + ], + "type": "text", + "content": "[81] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12699, 2021. 2, 3, 6, 7, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 576, + 287, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 287, + 653 + ], + "type": "text", + "content": "[82] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13150" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_content_list.json b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ac876d8c54dda6e45515650f973b650e467bf18 --- /dev/null +++ b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_content_list.json @@ -0,0 +1,2444 @@ +[ + { + "type": "text", + "text": "Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models", + "text_level": 1, + "bbox": [ + 272, + 130, + 697, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Geng", + "bbox": [ + 282, + 203, + 387, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inbum Park", + "bbox": [ + 426, + 204, + 521, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Owens", + "bbox": [ + 560, + 204, + 686, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Michigan", + "bbox": [ + 392, + 222, + 576, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://dangeng.github.io/visual_anagrams/", + "bbox": [ + 233, + 239, + 728, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/987e86a9986e16caf859d08ef52f5288cce7239213cb341ce13a1756cd6d684e.jpg", + "image_caption": [ + "Flips", + "eupued pe e joo bututed e", + "a painting of kitchenware" + ], + "image_footnote": [], + "bbox": [ + 91, + 314, + 173, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ef23218fd0b728272b1004bdff997e429dc204dc0a90fdcf6c7c1680c97aaba.jpg", + "image_caption": [ + "xnp 0j uedabouy", + "a lithograph of a fish", + "\"Sky and Water.\" M. C. Escher, 1938." + ], + "image_footnote": [], + "bbox": [ + 176, + 314, + 258, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9866bc691897863eed0186177327c89f907dff90b5f42b699d99224eac040e1c.jpg", + "image_caption": [ + "Bn 1000", + "an oil painting of people at a campfire" + ], + "image_footnote": [], + "bbox": [ + 261, + 313, + 344, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5002171d43b964390e5b4e8d5536f69d7ae0de236088a8e8c42228dd75c6518.jpg", + "image_caption": [ + "e e e", + "a painting of vases" + ], + "image_footnote": [], + "bbox": [ + 349, + 313, + 429, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa560a81240e262e00a002d483d702f6c0814fbfc7caf12565d1d3e2920416ed.jpg", + "image_caption": [ + "unbued", + "e go butmexp e", + "a drawing of a giraffe" + ], + "image_footnote": [], + "bbox": [ + 89, + 436, + 171, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17b26d7646fc32b6f3f6e973689e99a21fab5854e48c20f541913c62779d9acd.jpg", + "image_caption": [ + "pqrjgo", + "buTuTeD TIO ue", + "an oil paintingof a ship" + ], + "image_footnote": [], + "bbox": [ + 176, + 436, + 258, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f456a772756a1f7c605e6a28e7258e91abc13d52a400a0a76a2435cb609da2de.jpg", + "image_caption": [ + "ueWOM PRO UP", + "0000", + "a photo of a young lady", + "f. My Wife and My Mother-in-Law.\"1888," + ], + "image_footnote": [], + "bbox": [ + 261, + 436, + 344, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfe296c804a04f820d1ded5cd8f930242d4f6b738755209214359697f6f2be26.jpg", + "image_caption": [ + "e", + "to butuared e", + "a painting of wine and cheese" + ], + "image_footnote": [], + "bbox": [ + 346, + 436, + 429, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/814d18dc060c9849f68422ce09e319896cee2d5748752b1ee7fcd13e78a9f597.jpg", + "image_caption": [ + "Three Views", + "1JQREIO", + "putaired R", + "a painting of", + "a teddy bear", + "STATE JOBUSTED TO" + ], + "image_footnote": [], + "bbox": [ + 88, + 568, + 187, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f406ae660298293948f2795a3d68208f454f82f1111a69578881394acbdb628.jpg", + "image_caption": [ + "a painting", + "of einstein" + ], + "image_footnote": [], + "bbox": [ + 189, + 569, + 271, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f3612421973e08ad23070c591a14346668c9ff18f251c09b96f783e5d2950d6.jpg", + "image_caption": [ + "Four Views", + "an oil painting", + "of a rabbit" + ], + "image_footnote": [], + "bbox": [ + 313, + 556, + 431, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab5446bbb66a3eaaea6b9ef7dcf3e936070ec8d9f39c7784130cdaaec7e924e8.jpg", + "image_caption": [ + "\"Polymorphic\" Jigsaw Puzzles", + "an oil painting of a fruit bowl" + ], + "image_footnote": [], + "bbox": [ + 450, + 285, + 532, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b05af9a20c9095c3272e31b24ce458c6a3851db8537d2aeb118275dade15c495.jpg", + "image_caption": [ + "cf. \"Reversible Head with Basket of Fruit.\" Giuseppe Arcimboldo, 1950." + ], + "image_footnote": [], + "bbox": [ + 534, + 285, + 666, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ab94f84e4472e9192962bfb56e301da8ebe55cbc3ca5f8eb7a311df638806e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 310, + 689, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25664e8a4b8663c777e78e78dcd23b4b30df5dad23e07c1389db1bbc968e7ca1.jpg", + "image_caption": [ + "an oil painting of a monkey" + ], + "image_footnote": [], + "bbox": [ + 692, + 289, + 772, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca4af49d55157db81c431f3d909ac5c65a386223850acb5b715c39d7ce239d5d.jpg", + "image_caption": [ + "a watercolor of a kitten" + ], + "image_footnote": [], + "bbox": [ + 452, + 372, + 534, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b324c3d93dea8478143771c28d2981dc493f74f3345ab4b3de840f2767b1622.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 372, + 689, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f32c78681f796c55340ff92bfd5d13da89a91bbdf898b3c261b3c3af4bec3c70.jpg", + "image_caption": [ + "a watercolor of a puppy" + ], + "image_footnote": [], + "bbox": [ + 692, + 372, + 774, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48895abba167250c0c1120004e8662599da7d4bf6e112e22ef21eb121cc9423f.jpg", + "image_caption": [ + "a painting of houseplants" + ], + "image_footnote": [], + "bbox": [ + 452, + 457, + 534, + 520 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c80a59a8268180de20ecc68d6f660ab405f01ff88c7db1f679482956692380bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 455, + 678, + 536 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05de3f3cdc2c3da0ee416b080e0ffb2df3902cebfcdcb509fbf0a59cc01b1152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 478, + 689, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3898c7aea2239ea57a07b4c83808a89e1e373bbff8e6564004a3779379ee12d9.jpg", + "image_caption": [ + "a painting of", + "marilyn monroe", + "m111yHnH", + "Aude Oliva et al. 2006." + ], + "image_footnote": [], + "bbox": [ + 692, + 458, + 772, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/934841861477715a388eb484315e5741c424e207ae77287fe76462967ada9877.jpg", + "image_caption": [ + "Inner Rotations", + "a pop art of albert einstel" + ], + "image_footnote": [], + "bbox": [ + 799, + 304, + 880, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a22f77bf05a1a396c3124f3191bd1eb9223e3a0de6c5e0997d022d1113a28c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 371, + 880, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13416d1ab527a0400c0154000a9b2ee990a27a5d052afacc3b890b0fe6256518.jpg", + "image_caption": [ + "a pop art of", + "marilyn monroe", + "cf. \"Hybrid Images.\"" + ], + "image_footnote": [], + "bbox": [ + 797, + 440, + 880, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/074baec6063ef9c797d9147e38bcaeca649852d95ae0e4b90869204234cf5548.jpg", + "image_caption": [ + "Negatives", + "a lithograph of a teddy bear", + "Figure 1. Generating Multi-View Illusions. We propose a method for generating optical illusions from an off-the-shelf text-to-image diffusion model. We create images that match different prompts after undergoing a transformation. Our approach supports a variety of transformations, including flips, rotations, skews, color inversions, and jigsaw rearrangements. All images are hand selected. For random samples, please see Fig. 8 and Appendix D. For easier viewing, please see our webpage for animated versions of these illusions." + ], + "image_footnote": [], + "bbox": [ + 96, + 667, + 181, + 729 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af6f890fab785cbf670fb63a7dc9acffab7aa9a82c641981e68dfd1737235ba7.jpg", + "image_caption": [ + "a lithograph of a rabbit" + ], + "image_footnote": [], + "bbox": [ + 189, + 667, + 259, + 729 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed2a4e30afafb61febb75a05648eeffa19ae870cea53073c3c323903d85d0b58.jpg", + "image_caption": [ + "a lithograph of a landscape" + ], + "image_footnote": [], + "bbox": [ + 284, + 667, + 366, + 729 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6454857f19ee4e6c2c329140bad8d1468cfcbf9bcc4e9feac5a93356071201f0.jpg", + "image_caption": [ + "a lithograph of houseplants" + ], + "image_footnote": [], + "bbox": [ + 367, + 667, + 447, + 729 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f77fb1c778a0974cd974b2b8bdbf4be85b0d05f146794ac3f8e2aa2ebef26ac.jpg", + "image_caption": [ + "Skews", + "an oil painting ofa tudor portrait", + "cf. \"The Ambassadors.\" Hans Holbein, 1533." + ], + "image_footnote": [], + "bbox": [ + 464, + 560, + 547, + 625 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/160dee21df8225e42a92d803bf76065cf631ea6c5f6a7e8f2e89fb80542cb4d7.jpg", + "image_caption": [ + "an oil painting", + "of a skull" + ], + "image_footnote": [], + "bbox": [ + 547, + 560, + 653, + 625 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c3ad80605fecab77af38c4afb6172e05ab2efa04195e404793e67db4e4c9087.jpg", + "image_caption": [ + "an oil painting", + "of a soldier", + "" + ], + "image_footnote": [], + "bbox": [ + 681, + 560, + 766, + 625 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e7e6f14479d7436c07de49868eb1b48a9980f6af3142f8d71818a2c7f927a85.jpg", + "image_caption": [ + "an oil painting", + "of a houseplant" + ], + "image_footnote": [], + "bbox": [ + 792, + 560, + 874, + 625 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/940f09fb808c4d5569cf14b6e31fc4c27c4f4f5bb2d35f1b687bc3a86b0e7598.jpg", + "image_caption": [ + "Ambigrams", + "the word \"happy\"", + "cursive writing", + "$180^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 676, + 667, + 759, + 729 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05608620aa038aba6a1fdf954101c9b21f4dca55f4cd718625091c90c4fb977f.jpg", + "image_caption": [ + "the word \"holiday\".", + "cursive writing" + ], + "image_footnote": [], + "bbox": [ + 785, + 667, + 867, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 818, + 312, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the problem of synthesizing multi-view optical illusions: images that change appearance upon a transformation, such as a flip or rotation. We propose a simple,", + "bbox": [ + 73, + 856, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zero-shot method for obtaining these illusions from off-the-shelf text-to-image diffusion models. During the reverse diffusion process, we estimate the noise from different views of a noisy image, and then combine these noise estimates together and denoise the image. A theoretical analysis sug-", + "bbox": [ + 496, + 819, + 892, + 895 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 808, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 17, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24154", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gests that this method works precisely for views that can be written as orthogonal transformations, of which permutations are a subset. This leads to the idea of a visual anagram—an image that changes appearance under some rearrangement of pixels. This includes rotations and flips, but also more exotic pixel permutations such as a jigsaw rearrangement. Our approach also naturally extends to illusions with more than two views. We provide both qualitative and quantitative results demonstrating the effectiveness and flexibility of our method. Please see our project webpage for additional visualizations and results: https://dangeng.github.io/visual_anagrams/", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 286, + 209, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Images that change their appearance under a transformation, such as a rotation or a flip, have long fascinated students of perception, from Salvador Dalí to M. C. Escher. The appeal of these multi-view optical illusions lies partly in the challenge of arranging visual elements such that they may be understood in multiple different ways. Creating these illusions requires accurately modeling—and then subverting—visual perception.", + "bbox": [ + 75, + 311, + 468, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple, zero-shot method for creating multi-view illusions with off-the-shelf text-to-image diffusion models. In contrast to most previous work on computationally generating optical illusions [3-5, 10, 12, 15, 18, 20, 28, 31, 32, 38], our method does not require an explicit model of human perception. Rather, our approach builds on work that suggests generative models may process optical illusions in a way similar to humans [14, 23, 29]. In this way, our method is similar to recent work that uses diffusion models to create optical illusions by Burgert et al. [2] and Tancik [42].", + "bbox": [ + 75, + 433, + 468, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method can generate many types of classic illusions, such as images that change appearance when flipped or rotated (Fig. 1), as well as a new class of illusions which we term visual anagrams. These are images that change appearance under a permutation of their pixels. Image flips and rotations are a subset of these, as they can both be expressed as a permutation of pixels, but we also consider more exotic permutations. For example, we generate jigsaw puzzles that can be solved in two different ways, which we call \"polymorphic jigsaws.\" In addition, we successfully apply our approach to generating illusions with three and four views (Fig. 1).", + "bbox": [ + 75, + 599, + 468, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method works by using a diffusion model to denoise an image from multiple views, obtaining multiple noise estimates. These noise estimates are then combined to form a single noise estimate which is used to perform a step in the reverse diffusion process. However, we show that care must be taken in choosing these views. For one, the transformation must keep the statistics of the noise intact, as the diffusion model is trained under the assumption of i.i.d. Gaus", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sian noise. We provide an analysis of these conditions and give an exact specification of the class of transformations supported. Our contributions are as follows:", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a simple yet effective method for generating multi-view optical illusions using diffusion models.", + "- We derive a precise description of the set of views that our method supports and provide empirical evidence that these views work.", + "- We consider practical design decisions, crucial to optimizing the quality of generated illusions, and report ablations on our choices.", + "- We provide quantitative and qualitative results, showcasing both the efficacy and flexibility of our method." + ], + "bbox": [ + 500, + 138, + 890, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 305, + 640, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models [6, 17, 22, 35-37, 39-41] are a class of powerful generative models that iteratively convert a sample from a noise distribution to a sample from some data distribution. These models work by estimating the noise in a noisy sample, and removing the estimated noise following some update rule such as DDPM [22] or DDIM [40]. A prominent application of diffusion models has been text-conditioned image synthesis [24, 30, 36, 37]. In addition to a noisy image and a timestep, these models take a language model embedding of a text prompt as conditioning. Our approach is closely related to recent works that experiment with composing energy-based models and diffusion models [7-9, 13, 26, 27]. These approaches [9, 27] have shown that noise estimates from multiple conditional distributions can be combined together to obtain samples from compositions of the learned distributions. Our method uses a similar approach, and we apply it to the problem of multi-view illusion generation.", + "bbox": [ + 496, + 330, + 890, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computational Optical Illusions. Optical illusions serve as a testbed for understanding both human and machine perception [14, 19, 23, 29, 45]. We focus on generating illusions computationally, an area which has primarily relied on models of how our brains process external stimuli. Freeman et al. [12] create the illusion of constant motion in a desired direction by locally applying a filter with continuously shifting phase, relying on the observation that local phaseshifts are interpreted as global movement. Oliva et al. [31] propose a method to make \"hybrid images,\" which change appearance depending on the distance they are viewed from. This method takes advantage of the multiscale nature of human perception by blending high frequencies of one image with low frequencies from another. Chu et al. [5] camouflage objects in a scene through re-texturing, with additional constraints on luminance as to preserve salient features of the object, and other work camouflages objects from multiple viewpoints in 3D scenes [18, 32]. Recently, Chandra et al. [3] design color-constancy, size constancy, and face", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24155", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perception illusions by differentiating through a Bayesian model of human vision. Our method likewise generates illusions, but does not depend on an explicit model of human perception. Instead, our method works by leveraging visual priors in diffusion models learned implicitly through data. This aligns with observations [14, 23, 29] that generative models process illusions similarly to humans, and predict the same ambiguities. From this perspective, we can view our method as leveraging generative, rather than discriminative, models to synthesize adversarial examples [16] against humans [11].", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Illusions with Diffusion Models. Very recently, artists and researchers have taken steps that show the potential of using diffusion models to create illusions. An artist under the pseudonym MrUgleh [43] repurposed a model finetuned for generating QR codes [25, 47] to create images whose global structure subtly matches a given template image. In contrast, we study multi-view illusions that can be created zero-shot from off-the-shelf diffusion models, and our illusions are specified via text rather than images. Burgert et al. [2] use score distillation sampling (SDS) [33, 44] to create images that align with different prompts from different views. While in principle this approach supports a superset of our views, the use of SDS results in significantly lower quality results, and the need for explicit optimization leads to long sampling times. Our method is most similar to a proof-of-concept by Tancik [42], which creates rotation illusions by sampling from a latent diffusion model [36] while alternating noise estimates between different views and prompts. While our technical approach is similar, by contrast we systematically study multi-view illusions, both by experimentally evaluating many different types of illusions and by providing a theoretical analysis of which views are (and are not) supported. In doing so, we go beyond just rotation views. We also make a number of improvements that result in qualitatively and quantitatively better illusions, such as by identifying a source of artifacts from latent diffusion, and by adding support for an arbitrary number of views. To our knowledge, we are the first to systematically evaluate illusions generated by these approaches.", + "bbox": [ + 75, + 266, + 472, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 719, + 169, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to produce multi-view optical illusions using a pretrained diffusion model. That is, we seek to synthesize images that change appearance or identity when transformed, such as when flipped or rotated.", + "bbox": [ + 75, + 744, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text-conditioned Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 816, + 379, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [22, 39, 41] take i.i.d. Gaussian noise, $\\mathbf{x}_T$ , and iteratively denoise it to produce a sample, $\\mathbf{x}_0$ , from some data distribution. These models are parameterized by a neural network which estimates the noise in", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/489bdfa0d17e3a87fa56c8242c836b46eaab7b811772e79f5d1ceeb26146a288.jpg", + "image_caption": [ + "Figure 2. Algorithm Overview. Our method works by simultaneously denoising multiple views of an image. Given a noisy image $\\mathbf{x}_t$ , we compute noise estimates, $\\epsilon_t^i$ , conditioned on different prompts, after applying views $v_i$ . We then apply the inverse view $v_i^{-1}$ to align estimates, average the estimates, and perform a reverse diffusion step. The final output is an optical illusion." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "some intermediate, partially denoised data point $\\mathbf{x}_t$ , denoted as $\\epsilon_{\\theta}(\\mathbf{x}_t,y,t)$ , where $y$ is some conditioning such as text prompts and $t$ is the timestep in the diffusion process. The estimated noise is then used in an update rule [22, 40], from which $\\mathbf{x}_{t - 1}$ is computed from $\\mathbf{x}_t$ .", + "bbox": [ + 496, + 320, + 890, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To condition the diffusion model on another input, such as a text prompt, a common approach is to use classifier-free guidance [21]. With this method, unconditional noise estimates (usually obtained by passing the null text prompt as conditioning) and conditional noise estimates are combined together:", + "bbox": [ + 496, + 396, + 890, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} ^ {\\mathrm {C F G}} = \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\varnothing) + \\gamma \\left(\\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, y\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\varnothing\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 496, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\varnothing$ denotes the embedding of the empty string and $\\gamma$ is a parameter that controls the strength of the guidance. Classifier-free guidance acts to sharpen the distribution of generated images to produce higher quality results. It also enables negative prompting [1], in which the empty text prompt embedding, $\\varnothing$ , is replaced by a text prompt that we would like to discourage the model from generating.", + "bbox": [ + 496, + 523, + 893, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Parallel Denoising", + "text_level": 1, + "bbox": [ + 500, + 638, + 678, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We produce multi-view illusions by using a diffusion model to simultaneously denoise multiple views of an image. Concretely, we take a set of $N$ prompts, $y_{i}$ , each associated with a view function $v_{i}(\\cdot)$ , which applies a transformation to an image. These transformations may be, for example, the identity function, an image flip, or a permutation of pixels. Then given a diffusion model, $\\epsilon_{\\theta}(\\cdot)$ , and a partially denoised image, $\\mathbf{x}_t$ , we combine noise estimates from different views into a single noise estimate by averaging:", + "bbox": [ + 496, + 662, + 890, + 799 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {t} = \\frac {1}{N} \\sum_ {i} v _ {i} ^ {- 1} \\left(\\epsilon_ {\\theta} \\left(v _ {i} \\left(\\mathbf {x} _ {t}\\right), y _ {i}, t\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 801, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Effectively, we use each view $v_{i}$ to transform the noisy image $\\mathbf{x}_t$ , estimate the noise in the transformed images, and then apply $v_{i}^{-1}$ to the estimates in order to transform them back to the original view. Taking an average of these noise", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24156", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "estimates gives us our combined noise estimate, which we can then use with our choice of diffusion sampler. We note that this technique of combining noise estimates is similar to previous work on compositionality [7-9, 13, 26, 27], where the idea is studied in further detail. In order to incorporate classifier-free guidance we simply replace the estimates $\\epsilon_{\\theta}(v_i(\\mathbf{x}_t),y_i,t)$ with their classifier-free estimates, $\\epsilon_t^{\\mathrm{CFG}}$", + "bbox": [ + 75, + 90, + 472, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditions on Views", + "text_level": 1, + "bbox": [ + 76, + 207, + 272, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One straightforward condition for the views is that they must be invertible. But diffusion models also implicitly impose other conditions on the views $v_{i}(\\cdot)$ . We describe two such conditions below. We find that if these conditions are not satisfied, the denoising process produces poor results.", + "bbox": [ + 75, + 231, + 468, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Linearity. The diffusion model, $\\epsilon_{\\theta}$ , acts on noisy images, $\\mathbf{x}_t$ . That is, specifically images of the form:", + "bbox": [ + 75, + 316, + 468, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = w _ {t} ^ {\\text {s i g n a l}} \\underbrace {\\mathbf {x} _ {0}} _ {\\text {s i g n a l}} + w _ {t} ^ {\\text {n o i s e}} \\underbrace {\\epsilon} _ {\\text {n o i s e}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 349, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The exact values of $w_{t}^{\\mathrm{signal}}$ and $w_{t}^{\\mathrm{noise}}$ depend on model implementation details such as the variance schedule, but are unimportant for our work, so we omit them for clarity. What is important is that $\\mathbf{x}_t$ is a linear combination of pure signal, $\\mathbf{x}_0$ , and pure noise, $\\epsilon$ , for some specific $w_{t}^{\\mathrm{signal}}$ and $w_{t}^{\\mathrm{noise}}$ . Therefore our view $v_{i}$ must take a noisy image $\\mathbf{x}_t$ and transform it into a new noisy image $v_{i}(\\mathbf{x}_{t})$ that is also a linear combination of pure signal and pure noise with the same weighting. This can be achieved by requiring $v_{i}$ to be a linear transformation, of the form", + "bbox": [ + 75, + 392, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} \\left(\\mathbf {x} _ {t}\\right) = \\mathbf {A} _ {i} \\mathbf {x} _ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 549, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for some matrix $\\mathbf{A}_i$ , and some flattened noisy image $\\mathbf{x}_t$ . By linearity, we are effectively applying the view $v_i$ to the signal and the noise separately:", + "bbox": [ + 75, + 569, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} v _ {i} \\left(\\mathbf {x} _ {t}\\right) = \\mathbf {A} _ {i} \\left(w _ {t} ^ {\\text {s i g n a l}} \\mathbf {x} _ {0} + w _ {t} ^ {\\text {n o i s e}} \\epsilon\\right) (5) \\\\ = w _ {t} ^ {\\text {s i g n a l}} \\underbrace {\\mathbf {A} _ {i} \\mathbf {x} _ {0}} _ {\\text {n e w s i g n a l}} + w _ {t} ^ {\\text {n o i s e}} \\underbrace {\\mathbf {A} _ {i} \\epsilon} _ {\\text {n e w n o i s e}}. (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 618, + 468, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This results in a linear combination of transformed signal, $\\mathbf{A}_i\\mathbf{x}_0$ , and transformed noise, $\\mathbf{A}_i\\epsilon$ , weighted with the correct scaling factors. For further discussion, please see Appendix H.", + "bbox": [ + 76, + 678, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Statistical Consistency. In addition to expecting a linear combination of signal and noise at a specific weighting, the diffusion model also expects the noise to have a precise distribution. In particular, most diffusion networks are trained with $\\epsilon \\sim \\mathcal{N}(0,I)$ . Therefore, we must ensure that our transformed noise, $\\mathbf{A}_i\\epsilon$ , is likewise drawn from $\\mathcal{N}(0,I)$ . This is true if and only if $\\mathbf{A}_i$ is an orthogonal matrix. We provide a proof in Appendix I, but intuitively this fact reflects the spherical symmetry of the standard Gaussian density. Orthogonal transformations, being generalizations of rota", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions and flips to higher dimensions, preserve this spherically symmetric density. Note that these are rotations in the pixel values as opposed to spatial rotations.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Views Considered", + "text_level": 1, + "bbox": [ + 498, + 150, + 671, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The vast majority of orthogonal transformations applied to an image will not correspond to an intuitive image transformation. However, a number of these transformations do. Below, we enumerate the orthogonal transformations which we consider, all of which can be seen in the illusions in Fig. 1 unless otherwise specified.", + "bbox": [ + 496, + 175, + 890, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identity. The simplest transformation we consider is the identity transformation. Using this view allows us to optimize the untransformed image to align with a chosen prompt.", + "bbox": [ + 496, + 282, + 890, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Standard Image Manipulations. We also consider spatial rotations of an image, which can be viewed as permutations of pixels. This works because permutations are in turn orthogonal. However, caution must be exercised when applying a rotation view, as common anti-aliasing operations such as bilinear sampling will modify the statistics of the noise. We discuss this further in Sec. 4.4. Spatial reflections are also permutations of pixels. As such we can use these views to generate illusions. Finally, we implement an approximation to skewing by rolling columns of pixels by different displacements.", + "bbox": [ + 496, + 359, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General Permutations. We have already considered the special cases of spatial rotation, reflection, and skews but we can also consider other permutations. For example, we can divide an image into jigsaw pieces and rearrange these pieces to generate jigsaw puzzles with two solutions—what we call \"polymorphic\" jigsaw puzzles. Implementation details can be found in Appendix F.", + "bbox": [ + 496, + 541, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also consider the extreme case of sampling a completely random permutation of pixels and treating it as our view. Additionally, we can reduce the complexity of this by considering permutations of square patches, rather than pixels. Examples of these illusions can be found in Fig. 6 and are discussed in Sec. 4.3.", + "bbox": [ + 496, + 648, + 890, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we consider rotating a circle within an image while leaving the rest of the image stationary, which we term inner rotations. Note that the permutations we consider are certainly not exhaustive, and many clever transformations exist which we do not study.", + "bbox": [ + 496, + 741, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Color Inversion. Negation is an orthogonal transform; it is intuitively a 180 degree rotation generalized to higher dimensions. This allows us to generate illusions that change appearance upon color inversion, assuming pixel values are centered at 0 (e.g., in the range [-1, 1]).", + "bbox": [ + 496, + 833, + 890, + 909 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24157", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/227c327efbba6a7c244bb09a675c60ed147651e3088d6c1d4909d16a84a40076.jpg", + "image_caption": [ + "Figure 3. Latent-Based Artifacts. Manipulating the location of latent codes does not change the orientation of the blocks for which they encode. Therefore, when using latent diffusion models we see artifacts as shown above, in which straight lines are thatched under a rotation." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 467, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Arbitrary Orthogonal Transformations. An arbitrary rotation of an image in pixel space is uninterpretable. Nevertheless, we demonstrate that our method works for these transformations as well. While these \"illusions\" are inscrutable to the human eye, they serve as confirmation that any orthogonal transformation works as a view with our method. These can be found in Fig. 7 and are discussed in Sec. 4.3.", + "bbox": [ + 75, + 299, + 467, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Design Decisions", + "text_level": 1, + "bbox": [ + 76, + 428, + 241, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Beyond the core method, we also consider various design decisions aimed at maximizing illusion quality.", + "bbox": [ + 75, + 450, + 468, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel Diffusion Model. Previous work [42] performed multi-view denoising using Stable Diffusion [36], a latent diffusion model. However, the latent representation effectively encodes patches of pixels. This leads to artifacts under rotations or flips, where the location of latents change, but the content and orientation of these blocks do not. We show a qualitative example of this in Fig. 3, in which the model is forced to generate thatched lines to produce straight lines under a $90^{\\circ}$ rotation.", + "bbox": [ + 75, + 488, + 468, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ameliorate this issue, we implement our method using a pixel-based diffusion model, DeepFloyd IF [24]. DeepFloyd denoises directly on pixels, effectively side-stepping the problem of orientation in latent code blocks.", + "bbox": [ + 75, + 625, + 468, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining Noise Estimates. In addition to taking a mean of noise estimates from different views, we also consider alternating through them by timestep, using the estimate", + "bbox": [ + 75, + 691, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {t} = v _ {t \\bmod N} ^ {- 1} \\left(\\epsilon_ {\\theta} \\left(v _ {t \\bmod N} (\\mathbf {x} _ {t}), t, y)\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 750, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This is the reduction strategy used by [42], but we show in ablations in Sec. 4.2 that it performs worse than averaging.", + "bbox": [ + 75, + 772, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Negative Prompting. We experiment with negative prompting [1] in the 2-view case by using one view's prompt as a negative for the other view, and vice versa. This encourages the model to hide the other view's prompt for a given view. For a discussion, please see the ablations in Sec. 4.2.", + "bbox": [ + 75, + 809, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0abf702e141ec0a5049dd73d7975e42144a74f37ff24f7c2b4c7b5a87dd832fc.jpg", + "table_caption": [ + "Table 1. Quantitative Results. We report the alignment score, $\\mathcal{A}$ , and the concealment score, $\\mathcal{C}$ , as well as quantiles of these scores. For a discussion, please see Sec. 4.1." + ], + "table_footnote": [], + "table_body": "
Prompt PairMethodA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
CIFARBurgert et al. [2]0.2250.2530.2600.5010.5260.537
Tancik [42]0.2780.3100.3160.5950.6920.712
Ours0.2870.3210.3270.6240.7170.739
OursBurgert et al. [2]0.2330.2700.2830.5010.5260.538
Tancik [42]0.2560.2940.3090.5450.6210.655
Ours0.2750.3150.3260.5740.6680.694
", + "bbox": [ + 503, + 133, + 888, + 229 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/58d89858e28ccbf4a978b78b13ff9e220f6dacab9edd16be99d7437aafc6ec02.jpg", + "table_caption": [ + "Table 2. Ablations. We ablate negative prompting, reduction methods, and guidance scales on our dataset." + ], + "table_footnote": [], + "table_body": "
AblationA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
Negative Prompting0.240.270.2760.5760.6590.683
No Negative Prompting0.2550.2850.2950.5670.6430.679
Alternating Reduction0.2520.2860.2920.5600.6390.664
Mean Reduction0.2550.2850.2950.5670.6430.679
γ = 3.00.2390.2710.2850.5370.6100.629
γ = 7.00.2550.2850.2950.5670.6430.679
γ = 10.00.2590.2900.2970.5760.6640.702
", + "bbox": [ + 503, + 272, + 888, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 415, + 586, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide quantitative and qualitative results, and quantitative ablations. If not specified, qualitative results have been hand picked for quality. For random samples please see Fig. 8 and Appendix D. All implementation details can be found in Appendix A.", + "bbox": [ + 496, + 441, + 890, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 498, + 527, + 692, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We use CLIP [34] to measure how well views align with the desired prompts. We consider two metrics derived from a score matrix $\\mathbf{S} \\in \\mathbb{R}^{N \\times N}$ , defined as", + "bbox": [ + 496, + 553, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i j} = \\phi_ {\\mathrm {i m g}} \\left(v _ {i} (\\mathbf {x})\\right) ^ {T} \\phi_ {\\text {t e x t}} \\left(p _ {j}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 602, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{\\mathrm{img}}$ and $\\phi_{\\mathrm{text}}$ are the CLIP visual and textual encoders respectively, returning a unit-norm vector embedding. $\\mathbf{x}$ is our generated illusion, and $v_{i}$ are our views with associated prompts $p_i$ . A higher dot product indicates higher similarity between the image and text.", + "bbox": [ + 496, + 625, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first metric we consider is $\\min \\operatorname{diag}(\\mathbf{S})$ , which intuitively measures the worst alignment of all the views. We term this metric $\\mathcal{A}$ , the alignment score. However, this metric does not account for the possibility of seeing prompt $p_i$ in view $v_j$ for $i \\neq j$ . This is an occasional failure case of our method and to quantify this we propose a second derived metric which we term $\\mathcal{C}$ , the concealment score, computed as", + "bbox": [ + 496, + 699, + 890, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N} \\operatorname {t r} (\\operatorname {s o f t m a x} (S / \\tau)), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 820, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter of CLIP. In computing this metric we average both directions of the softmax, so that this metric measures how well CLIP can classify a view", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24158", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6fc4757101a20b40ff7974893e548082e7e5c5f95a5eeea1d98597a3118d1b29.jpg", + "image_caption": [ + "Figure 4. Flip View CLIP Score Distribution. We visualize trade-offs between flipped and unflipped views by plotting the distribution of CLIP scores on the datasets. Note that the quality of the flipped image is as good as the unflipped image, with parity indicated by the dashed line." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 467, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as one of the $N$ prompts and vice versa.", + "bbox": [ + 76, + 353, + 341, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To evaluate our method and baselines we compile two datasets of prompt pairs for 2-view illusions. One dataset uses the 10 classes from CIFAR-10 and contains a prompt per pair of classes, for a total of 45 prompt pairs. We refer to this as CIFAR. The other dataset we compile by hand, with the process documented in Appendix B. This dataset consists of 50 prompt pairs, which we refer to as Ours.", + "bbox": [ + 75, + 377, + 468, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We use two baselines that generate illusions using off-the-shelf diffusion models. One, which we denote \"Burgert et al. [2],\" uses Score Distillation Sampling. The other, which we denote \"Tancik [42],\" is an earlier version of our method, with differences discussed in more detail in Sec. 2", + "bbox": [ + 75, + 507, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We show results comparing our method to baselines on both datasets in Tab. 1 using vertical flips. We use vertical flips because it is a transformation supported by our method as well as the baselines. We use 10 samples per prompt, for a total of 450 and 500 samples for the CIFAR dataset and our dataset respectively. It is hard to perform a fair comparison with more samples because the Burgert et al. method uses SDS, which is quite slow1. Because we are particularly interested in the \"best-case\" performance, we also report quantiles of metrics, which we denote as $\\mathcal{A}_{0.9}$ for the 90th percentile, for example. As can be seen, our method performs consistently better than the baselines, in both the alignment score and the concealment score.", + "bbox": [ + 75, + 607, + 468, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to give a clearer understanding of trade-offs when optimizing two views, we show density plots which plot the CLIP scores of each of the two views of an illusion in Fig. 4. As can be seen, we do better than the baselines", + "bbox": [ + 75, + 804, + 470, + 864 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b503c0fc497282bbeadba423851a5a5ac4e0f229f0bbd9af5784d8508090082.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparisons. We compare illusions generated by baselines to our illusions. We show examples from both our prompt dataset and the CIFAR prompt dataset." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 888, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on average and in the best-case. Moreover, flipping during denoising does not hurt performance. The quality of the flipped images is as high as the unflipped images.", + "bbox": [ + 498, + 497, + 890, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablations", + "text_level": 1, + "bbox": [ + 500, + 554, + 609, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We ablate out the noise estimate reduction strategy, negative prompting, and the guidance scale in Tab. 2. We use our dataset, with 10 samples for each prompt for a total of 500 illusions.", + "bbox": [ + 498, + 577, + 890, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reduction Strategy. We find that mean reduction does better than alternating. Our hypothesis is that alternating the noise estimates results in \"thrashing,\" causing poor convergence. Moreover, we find that the alternating strategy gives poor results on illusions with more than 2 views, as each view has fewer denoising steps. Qualitative examples of this can be found in Appendix G.", + "bbox": [ + 496, + 648, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Negative Prompting. When using negative prompting, care must be taken to omit any overlap between the negative and positive prompt. For example, given the two prompts \"oil painting of a dog\" and \"oil painting of a cat\", using one prompt as the negative for the other would simultaneously encourage and discourage the style \"oil painting\". Rather, the negative prompts should be \"a cat\" and \"a dog\" respectively. We find that negative prompting can improve the concealment score, indi", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "Sampling just 10 images per prompt already takes more than a week of GPU-hours.", + "bbox": [ + 75, + 875, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24159", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abd14ef2b8aaffdeb18e4e2b2abe31c8bb3049e1a80febde2978136c9bc0378c.jpg", + "image_caption": [ + "8×8", + "a pencil sket of a lemur", + "h a pencil sketch of a kangaroo" + ], + "image_footnote": [], + "bbox": [ + 86, + 102, + 166, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60e13cf12c3f825e12d1f3e1aa39cd51d5186bf306fcb538da68bc03a54404b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 102, + 264, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebc9417e6357030c8d8d6cbc36c352109321d79e31e5ea1d6c345af1311393a3.jpg", + "image_caption": [ + "a watercolor of a rabbit" + ], + "image_footnote": [], + "bbox": [ + 282, + 102, + 361, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbb70ef7122727c40692d583b525d34c3e4d24e158ab9ccbee8883c8a07d71de.jpg", + "image_caption": [ + "a watercolor of a duck" + ], + "image_footnote": [], + "bbox": [ + 361, + 102, + 462, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62985ce8ec139ba1c9ebbee338e48eccc0028d07a88f0f523cfd1a7116892138.jpg", + "image_caption": [ + "[16×16]", + "an oil painting of a young man" + ], + "image_footnote": [], + "bbox": [ + 137, + 194, + 215, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/831727bd920fa7d3762c66bda5f78631933079b8b99f7f936346f9feb90f9368.jpg", + "image_caption": [ + "cf. \"Kanichen und Ende.\" Fliegende Blätter, 1892." + ], + "image_footnote": [], + "bbox": [ + 227, + 193, + 320, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75b032df22ddbc67ddfdfe6abea5b07081a84d38f3c33e5be0a4a01fd6cfcb69.jpg", + "image_caption": [ + "an oil painting of an old man" + ], + "image_footnote": [], + "bbox": [ + 333, + 194, + 411, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ae74fd3489ddf4d5c824aa3a324827e8daf3ccb3020cb313fc1eaf29a394413.jpg", + "image_caption": [ + "32×32" + ], + "image_footnote": [], + "bbox": [ + 84, + 291, + 163, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddfd37af6f4c49f3d525a16387cc64a3b969a7b7a4d388b7f124363f3e6f08cb.jpg", + "image_caption": [ + "an oil painting of a young man", + "ng an oil paintingin of an old man ,", + "Figure 6. Permutation Illusions. We synthesize images whose appearance changes upon permutation of patches. Even in the difficult case of a $64 \\times 64$ grid of patches, in which every pixel is effectively shuffled, we are able to generate meaningful images." + ], + "image_footnote": [], + "bbox": [ + 163, + 291, + 267, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69abf4c288cc173d6e075a7862e1c00400346fbb6d1a793753a09d465e2eadfa.jpg", + "image_caption": [ + "64×64", + "a mosaic of a duck" + ], + "image_footnote": [], + "bbox": [ + 279, + 291, + 359, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac663315c9966c922d45d6c64a28d582a7a68b3ea671b85d46840ae8a61712d0.jpg", + "image_caption": [ + "a mosaic of a rabbit" + ], + "image_footnote": [], + "bbox": [ + 359, + 291, + 465, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cating that it is working as intended. But this comes at the cost of worse alignment score. This is because the negative and positive prompt may have fundamental similarities. For example, using \"a cat\" as the negative prompt for the prompt \"an oil painting of a dog\" may discourage the model from synthesizing anything remotely cat-like—such as fur, four legs, or a tail—even if it helps in producing a dog. For this reason we opt not to use negative prompting with our method.", + "bbox": [ + 75, + 441, + 468, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Guidance Scale. We also ablate out various guidance scales, $\\gamma$ , for our method. We find that a higher guidance scale tends to do better. This is presumably because a higher guidance scale results in a sharper sampling distribution.", + "bbox": [ + 75, + 585, + 468, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 657, + 259, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show qualitative results in Fig. 1, Fig. 5, Fig. 6, and Fig. 7. Again, random samples may be found in Fig. 8 and Appendix D. Additional qualitative samples can be found in Appendix C. Overall, we find that our method can produce very high quality optical illusions for a wide range of views. Interestingly, our method often finds clever ways of reusing elements from one view for another, such as in the \"waterfalls\"/\"rabbit\"/\"teddy bear\" three-view illusion in Fig. 1, in which the nose of the teddy bear is the eye of the rabbit, and a rock on the waterfalls.", + "bbox": [ + 75, + 680, + 470, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We provide qualitative comparisons of our method to baselines in Fig. 5, where we pick the best images out of 100 samples for each method. As can be seen,", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6780f9bb96e032916b5336a8617d9b81193e4d9df5cc3166f4ef6346bb31ff81.jpg", + "image_caption": [ + "an oil painting of an old woman", + "Figure 7. Orthogonal Illusions. We show that our method works, even when the view is a randomly sampled orthogonal transformation $\\mathbf{A}$ . While these \"illusions\" are incomprehensible to human perception, they serve as a confirmation for our mathematical analysis." + ], + "image_footnote": [], + "bbox": [ + 501, + 104, + 586, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dfa5c16125480acdd32228222eeb38fa866f901767a7b32e08355c6d61a21e6.jpg", + "image_caption": [ + "an oil painting of a young lady" + ], + "image_footnote": [], + "bbox": [ + 589, + 104, + 692, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c13a2c339044b8068911dd4afa40ddb3abdc55d05a1949fdc16a904b9284df7.jpg", + "image_caption": [ + "a photo of a duck" + ], + "image_footnote": [], + "bbox": [ + 699, + 104, + 785, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/612aa866ee39112ad67565a5a49a38d7576333f7c035cf85bbce3d8ca7236025.jpg", + "image_caption": [ + "a photo of a rabbit" + ], + "image_footnote": [], + "bbox": [ + 787, + 104, + 890, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images generated using our method match the prompts in both views equally well and are of higher quality.", + "bbox": [ + 498, + 253, + 890, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Permutations. Pixel and patch permutations, being a subset of orthogonal transformations, should work with our method. We show that this is indeed the case in Fig. 6, where we have results on patch grids of various sizes under randomly sampled permutations. The $64 \\times 64$ case is quite hard, yet our method is able to generate images that satisfy the constraint, albeit at lower quality.", + "bbox": [ + 498, + 291, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Arbitrary Orthogonal Transformations. As discussed in Sec. 3.3, our method works for any orthogonal transformation. So far, we have shown illusions based on a subset of orthogonal views that correspond to intuitive image transformations. In Fig. 7, we show \"illusions\" using an arbitrary orthogonal transformation as a view. We use Stable Diffusion [36] and sample a random orthogonal matrix $\\mathbf{A} \\in \\mathbb{R}^{16384 \\times 16384}$ by projecting an i.i.d. random Gaussian matrix with an SVD. These dimensions correspond to the size of the Stable Diffusion latent space. We note that this is an incredibly hard and unnatural transformation of an image, and results are accordingly of lower quality, but our method is still able to produce reasonable images.", + "bbox": [ + 498, + 406, + 892, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Random Samples. We show random samples for selected prompts in Fig. 8. As can be seen, these random samples, while not as good as those in Fig. 1, are still very high quality. Some failure cases can be seen where the model prefers one prompt over another. We add further discussion and present more random samples in Appendix D.", + "bbox": [ + 498, + 611, + 890, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Failures", + "text_level": 1, + "bbox": [ + 500, + 712, + 598, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We highlight three interesting failure cases of our method in Fig. 9.", + "bbox": [ + 498, + 734, + 890, + 766 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Independent Synthesis. The first of these cases involves the model synthesizing prompts separately, without combining elements of the two to form an illusion. Empirically, this happens surprisingly rarely, especially given that it seems to be such an easy shortcut solution. We hypothesize that this is because the diffusion model is biased toward centering its content, resulting in far more images with content that is integrated and centered as opposed to separate and off-center.", + "bbox": [ + 496, + 775, + 890, + 910 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24160", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a15a49f015b1205e9f1ca11d1635a681b97bf6857b0c8dcdd8536a087c150025.jpg", + "image_caption": [ + "Figure 8. Random Samples. We show random samples, along with their corresponding view, for selected prompts. For more random samples please see Appendix D. For best quality, view digitally and zoom-in." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Noise Shift. Using views that preserve noise statistics is critical to our method's success. For example, we attempted to recreate the \"Dress\" illusion [46], in which a dress can be seen as either \"blue and black\" or \"white and gold.\" We used simple white balancing as our view, in which pixel values were scaled by a constant factor. While this transformation is linear, it does not preserve the statistics of Gaussian noise. As a result, we see artifacts in the forms of spots, which we hypothesize is the result of the model interpreting the scaled Gaussian noise as signal and actively denoising peaks in the scaled noise.", + "bbox": [ + 75, + 425, + 472, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Correlated Noise. While our method supports rotations as transformations, as demonstrated with the \"3-view,\" \"4-view,\" and \"Inner Rotation\" illusions in Fig. 1, care must be taken that the rotation does not introduce correlations in the noise, such as through anti-aliasing. For example, bilinear sampling introduces significant correlations in the noise, as it is a linear combination of four adjacent pixels. Therefore, seemingly innocuous rotations may result in divergent samples if transformations are not carefully kept correlation free, as shown with the 45 degree bilinear rotation in Fig. 9.", + "bbox": [ + 75, + 602, + 468, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Conclusions", + "text_level": 1, + "bbox": [ + 76, + 768, + 341, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method to produce compelling and diverse optical illusions. Our method is simple and straightforward to implement, and additionally amenable to theoretical analysis. We prove that our method works for a broad set of transformations, and qualitatively show that it can generate a wide array of optical illusions. However, at the same time many possible illusions and transformations are still", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d49d0720d5e01c1700bf17fed202ce36077ef1225d768aa5dc67a4a1daf839c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 422, + 604, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3e25a3767a1636fbd3ad4548a33c27ee05256106c2a1129c3634eea9b1c9e7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 503, + 604, + 584 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/463e60f322959c2406320ce19530d2cca2c11e2cdc3690651a36d06b9d582367.jpg", + "image_caption": [ + "Figure 9. Failures. We highlight three interesting failure cases, which are discussed in Sec. 4.4." + ], + "image_footnote": [], + "bbox": [ + 501, + 585, + 604, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure: Independent Synthesis", + "bbox": [ + 612, + 434, + 795, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View: Vertical Flip", + "bbox": [ + 612, + 450, + 718, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "an oil painting of a bowl of fruit \nan oil painting of a monkey", + "bbox": [ + 612, + 464, + 874, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure: Noise Shift", + "bbox": [ + 612, + 516, + 728, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View: White Balancing", + "bbox": [ + 612, + 530, + 745, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a photo of a black and blue dressa photo of a white and gold dress", + "bbox": [ + 612, + 546, + 867, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure: Correlated Noise", + "bbox": [ + 612, + 597, + 761, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View: Rotate Inner Circle 45 Degrees (Bilinear)", + "bbox": [ + 612, + 612, + 877, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a sketch of an elephant a sketch of a mouse", + "bbox": [ + 612, + 628, + 790, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not possible using our method, such as color constancy illusions, homographies, stretches, and more generally nonvolume-preserving deformations. We leave implementation of these views for future work. Moreover, our method does not consistently produce perfect illusions. This may be a symptom of the difficulty of producing good illusions, but may indicate future work to be done to improve consistency.", + "bbox": [ + 496, + 703, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We thank William Henning, Trenton Chang, Kimball Strong, Jeongsoo Park, Patrick Chao, Kurtland Chua, and Mohamed El Banani for their feedback on early drafts. Daniel is supported by the National Science Foundation Graduate Research Fellowship under Grant No. 1841052.", + "bbox": [ + 496, + 816, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24161", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 94, + 173, + 109 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] AUTOMATIC1111. Negative prompt. https : // github . com / AUTOMATIC1111 / stable - diffusion - webui / wiki / Negative - prompt, 2022. Accessed: November 7, 2023. 3, 5", + "[2] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael Ryoo. Diffusion illusions: Hiding images in plain sight. https://ryanndagreat.github.io/Diffusion-Illusions, 2023. 2, 3, 5, 6", + "[3] Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum, and Jonathan Ragan-Kelley. Designing perceptual puzzles by differentiating probabilistic programs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2", + "[4] Ming-Te Chi, Chih-Yuan Yao, Eugene Zhang, and Tong-Yee Lee. Optical illusion shape texturing using repeated asymmetric patterns. The Visual Computer, 30:809–819, 2014.", + "[5] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 2", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 2, 4", + "[8] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020.", + "[9] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 2, 4", + "[10] Werner Ehm. A variational approach to geometric-optical illusions modeling. Proceedings of Fechner Day, 27(1):41-46, 2011. 2", + "[11] Gamaleldin Elsayed, Shreya Shankar, Brian Cheung, Nicolas Papernot, Alexey Kurakin, Ian Goodfellow, and Jascha Sohl-Dickstein. Adversarial examples that fool both computer vision and time-limited humans. Advances in neural information processing systems, 31, 2018. 3", + "[12] William T Freeman, Edward H Adelson, and David J Heeger. Motion without movement. ACM Siggraph Computer Graphics, 25(4):27-30, 1991. 2", + "[13] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023. 2, 4", + "[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 3", + "[15] Alex Gomez-Villa, Adrián Martín, Javier Vazquez-Corral, Marcelo Bertalmío, and Jesús Malo. On the synthesis of" + ], + "bbox": [ + 78, + 119, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 2", + "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3", + "[17] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35:14715-14728, 2022. 2", + "[18] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[19] Aaron Hertzmann. Visual indeterminacy in gan art. In ACM SIGGRAPH 2020 Art Gallery, pages 424-428. 2020. 2", + "[20] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2, 3", + "[23] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023. 2, 3", + "[24] Mikhail Konstantinov, Alex Shonenkov, Daria Bakshandaeva, and Ksenia Ivanova. If by deepfloyd lab at stabilityai, 2023. GitHub repository. 2, 5, 11", + "[25] Monster Labs. Controlnet qr code monster v2 for sd-1.5, 2023.3", + "[26] Nan Liu, Shuang Li, Yilun Du, Josh Tenenbaum, and Antonio Torralba. Learning to compose visual relations. Advances in Neural Information Processing Systems, 34: 23166-23178, 2021. 2, 4", + "[27] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4", + "[28] Dominique Makowski, Zen J Lau, Tam Pham, W Paul Boyce, and SH Annabel Chen. A parametric framework to generate visual illusions using python. Perception, 50(11):950-965, 2021. 2", + "[29] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? 2023. 2, 3", + "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2", + "[31] Aude Oliva, Antonio Torralba, and Philippe G. Schyns. Hybrid images. ACM Trans. Graph., 25(3):527-532, 2006. 2", + "[32] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. 2014. 2", + "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3" + ], + "bbox": [ + 503, + 93, + 892, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24162", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5", + "[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2", + "[38] Troy Shinbrot, Miguel Vivar Lazo, and Theo Siu. Network simulations of optical illusions. International Journal of Modern Physics C, 28(02):1750018, 2017. 2", + "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2, 3", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 3", + "[42] Matthew Tancik. Illusion diffusion. https://github.com/tancik/Illusion-Diffusion, 2023.2, 3, 5, 6, 12", + "[43] Ugleh. Spiral town - different approach to qr monster. https://www.reddit.com/r/StableDiffusion/comments/16ew9fz/spiral_town_differentapproach_to_qr_monster/, 2023.3", + "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3", + "[45] Xi Wang, Zoya Bylinskii, Aaron Hertzmann, and Robert Pepperell. Toward quantifying ambiguities in artistic images. ACM Transactions on Applied Perception (TAP), 17(4):1-10, 2020. 2", + "[46] Wikipedia contributors. The dress. https://en." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wikipedia.org/wiki/The_dress. Accessed: November 9, 2023. 8", + "[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 148 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24163", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_model.json b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8339efb09a493396debf0e73129bb492d21cbdcf --- /dev/null +++ b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_model.json @@ -0,0 +1,3795 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.018, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.273, + 0.131, + 0.699, + 0.177 + ], + "angle": 0, + "content": "Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.204, + 0.388, + 0.222 + ], + "angle": 0, + "content": "Daniel Geng" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.205, + 0.522, + 0.221 + ], + "angle": 0, + "content": "Inbum Park" + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.205, + 0.687, + 0.221 + ], + "angle": 0, + "content": "Andrew Owens" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.223, + 0.578, + 0.24 + ], + "angle": 0, + "content": "University of Michigan" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.24, + 0.73, + 0.258 + ], + "angle": 0, + "content": "https://dangeng.github.io/visual_anagrams/" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.273, + 0.129, + 0.284 + ], + "angle": 0, + "content": "Flips" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.299, + 0.162, + 0.313 + ], + "angle": 0, + "content": "eupued pe e joo bututed e" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.315, + 0.174, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.38, + 0.162, + 0.393 + ], + "angle": 0, + "content": "a painting of kitchenware" + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.299, + 0.245, + 0.313 + ], + "angle": 0, + "content": "xnp 0j uedabouy" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.315, + 0.259, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.379, + 0.246, + 0.393 + ], + "angle": 0, + "content": "a lithograph of a fish" + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.394, + 0.248, + 0.408 + ], + "angle": 0, + "content": "\"Sky and Water.\" M. C. Escher, 1938." + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.299, + 0.338, + 0.313 + ], + "angle": 0, + "content": "Bn 1000" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.314, + 0.345, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.379, + 0.338, + 0.401 + ], + "angle": 0, + "content": "an oil painting of people at a campfire" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.299, + 0.413, + 0.313 + ], + "angle": 0, + "content": "e e e" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.314, + 0.431, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.379, + 0.412, + 0.393 + ], + "angle": 0, + "content": "a painting of vases" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.421, + 0.15, + 0.428 + ], + "angle": 0, + "content": "unbued" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.429, + 0.167, + 0.436 + ], + "angle": 0, + "content": "e go butmexp e" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.437, + 0.173, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.502, + 0.165, + 0.516 + ], + "angle": 0, + "content": "a drawing of a giraffe" + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.422, + 0.245, + 0.429 + ], + "angle": 0, + "content": "pqrjgo" + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.43, + 0.255, + 0.436 + ], + "angle": 0, + "content": "buTuTeD TIO ue" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.437, + 0.259, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.502, + 0.253, + 0.516 + ], + "angle": 0, + "content": "an oil paintingof a ship" + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.422, + 0.331, + 0.429 + ], + "angle": 0, + "content": "ueWOM PRO UP" + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.43, + 0.328, + 0.436 + ], + "angle": 0, + "content": "0000" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.437, + 0.345, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.503, + 0.333, + 0.517 + ], + "angle": 0, + "content": "a photo of a young lady" + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.518, + 0.334, + 0.531 + ], + "angle": 0, + "content": "f. My Wife and My Mother-in-Law.\"1888," + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.422, + 0.41, + 0.429 + ], + "angle": 0, + "content": "e" + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.429, + 0.42, + 0.436 + ], + "angle": 0, + "content": "to butuared e" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.437, + 0.43, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.502, + 0.423, + 0.516 + ], + "angle": 0, + "content": "a painting of wine and cheese" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.544, + 0.16, + 0.554 + ], + "angle": 0, + "content": "Three Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.555, + 0.172, + 0.561 + ], + "angle": 0, + "content": "1JQREIO" + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.562, + 0.172, + 0.567 + ], + "angle": 0, + "content": "putaired R" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.569, + 0.188, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.637, + 0.177, + 0.643 + ], + "angle": 0, + "content": "a painting of" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.644, + 0.177, + 0.65 + ], + "angle": 0, + "content": "a teddy bear" + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.555, + 0.254, + 0.569 + ], + "angle": 0, + "content": "STATE JOBUSTED TO" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.57, + 0.272, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.637, + 0.256, + 0.643 + ], + "angle": 0, + "content": "a painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.644, + 0.256, + 0.649 + ], + "angle": 0, + "content": "of einstein" + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.545, + 0.379, + 0.554 + ], + "angle": 0, + "content": "Four Views" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.557, + 0.433, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.635, + 0.405, + 0.64 + ], + "angle": 0, + "content": "an oil painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.641, + 0.398, + 0.647 + ], + "angle": 0, + "content": "of a rabbit" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.273, + 0.612, + 0.284 + ], + "angle": 0, + "content": "\"Polymorphic\" Jigsaw Puzzles" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.286, + 0.534, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.355, + 0.531, + 0.37 + ], + "angle": 0, + "content": "an oil painting of a fruit bowl" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.286, + 0.668, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.356, + 0.675, + 0.369 + ], + "angle": 0, + "content": "cf. \"Reversible Head with Basket of Fruit.\" Giuseppe Arcimboldo, 1950." + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.311, + 0.691, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.29, + 0.774, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.355, + 0.77, + 0.37 + ], + "angle": 0, + "content": "an oil painting of a monkey" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.373, + 0.535, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.438, + 0.524, + 0.452 + ], + "angle": 0, + "content": "a watercolor of a kitten" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.373, + 0.69, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.373, + 0.775, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.438, + 0.764, + 0.453 + ], + "angle": 0, + "content": "a watercolor of a puppy" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.458, + 0.535, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.523, + 0.526, + 0.538 + ], + "angle": 0, + "content": "a painting of houseplants" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.457, + 0.679, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.479, + 0.691, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.459, + 0.774, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.523, + 0.768, + 0.53 + ], + "angle": 0, + "content": "a painting of" + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.53, + 0.768, + 0.535 + ], + "angle": 0, + "content": "marilyn monroe" + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.535, + 0.768, + 0.54 + ], + "angle": 0, + "content": "m111yHnH" + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.274, + 0.877, + 0.283 + ], + "angle": 0, + "content": "Inner Rotations" + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.288, + 0.868, + 0.302 + ], + "angle": 0, + "content": "a pop art of albert einstel" + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.305, + 0.882, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.372, + 0.882, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.441, + 0.882, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.507, + 0.872, + 0.513 + ], + "angle": 0, + "content": "a pop art of" + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.514, + 0.872, + 0.52 + ], + "angle": 0, + "content": "marilyn monroe" + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.523, + 0.868, + 0.529 + ], + "angle": 0, + "content": "cf. \"Hybrid Images.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.53, + 0.872, + 0.536 + ], + "angle": 0, + "content": "Aude Oliva et al. 2006." + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.655, + 0.152, + 0.665 + ], + "angle": 0, + "content": "Negatives" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.668, + 0.182, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.733, + 0.17, + 0.746 + ], + "angle": 0, + "content": "a lithograph of a teddy bear" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.668, + 0.261, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.733, + 0.248, + 0.746 + ], + "angle": 0, + "content": "a lithograph of a rabbit" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.669, + 0.367, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.733, + 0.359, + 0.746 + ], + "angle": 0, + "content": "a lithograph of a landscape" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.669, + 0.449, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.733, + 0.443, + 0.746 + ], + "angle": 0, + "content": "a lithograph of houseplants" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.545, + 0.501, + 0.554 + ], + "angle": 0, + "content": "Skews" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.561, + 0.548, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.626, + 0.547, + 0.639 + ], + "angle": 0, + "content": "an oil painting ofa tudor portrait" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.561, + 0.655, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.626, + 0.645, + 0.632 + ], + "angle": 0, + "content": "an oil painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.633, + 0.637, + 0.639 + ], + "angle": 0, + "content": "of a skull" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.643, + 0.622, + 0.65 + ], + "angle": 0, + "content": "cf. \"The Ambassadors.\" Hans Holbein, 1533." + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.561, + 0.767, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.626, + 0.761, + 0.632 + ], + "angle": 0, + "content": "an oil painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.633, + 0.753, + 0.639 + ], + "angle": 0, + "content": "of a soldier" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.643, + 0.791, + 0.649 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.561, + 0.875, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.626, + 0.868, + 0.632 + ], + "angle": 0, + "content": "an oil painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.633, + 0.869, + 0.639 + ], + "angle": 0, + "content": "of a houseplant" + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.656, + 0.735, + 0.665 + ], + "angle": 0, + "content": "Ambigrams" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.668, + 0.761, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.732, + 0.753, + 0.738 + ], + "angle": 0, + "content": "the word \"happy\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.739, + 0.753, + 0.745 + ], + "angle": 0, + "content": "cursive writing" + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.684, + 0.784, + 0.692 + ], + "angle": 0, + "content": "\\(180^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.669, + 0.868, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.732, + 0.868, + 0.738 + ], + "angle": 0, + "content": "the word \"holiday\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.739, + 0.86, + 0.745 + ], + "angle": 0, + "content": "cursive writing" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.752, + 0.894, + 0.81 + ], + "angle": 0, + "content": "Figure 1. Generating Multi-View Illusions. We propose a method for generating optical illusions from an off-the-shelf text-to-image diffusion model. We create images that match different prompts after undergoing a transformation. Our approach supports a variety of transformations, including flips, rotations, skews, color inversions, and jigsaw rearrangements. All images are hand selected. For random samples, please see Fig. 8 and Appendix D. For easier viewing, please see our webpage for animated versions of these illusions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.819, + 0.313, + 0.834 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.857, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We address the problem of synthesizing multi-view optical illusions: images that change appearance upon a transformation, such as a flip or rotation. We propose a simple," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.893, + 0.896 + ], + "angle": 0, + "content": "zero-shot method for obtaining these illusions from off-the-shelf text-to-image diffusion models. During the reverse diffusion process, we estimate the noise from different views of a noisy image, and then combine these noise estimates together and denoise the image. A theoretical analysis sug-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24154" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "gests that this method works precisely for views that can be written as orthogonal transformations, of which permutations are a subset. This leads to the idea of a visual anagram—an image that changes appearance under some rearrangement of pixels. This includes rotations and flips, but also more exotic pixel permutations such as a jigsaw rearrangement. Our approach also naturally extends to illusions with more than two views. We provide both qualitative and quantitative results demonstrating the effectiveness and flexibility of our method. Please see our project webpage for additional visualizations and results: https://dangeng.github.io/visual_anagrams/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.287, + 0.21, + 0.303 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.47, + 0.433 + ], + "angle": 0, + "content": "Images that change their appearance under a transformation, such as a rotation or a flip, have long fascinated students of perception, from Salvador Dalí to M. C. Escher. The appeal of these multi-view optical illusions lies partly in the challenge of arranging visual elements such that they may be understood in multiple different ways. Creating these illusions requires accurately modeling—and then subverting—visual perception." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.599 + ], + "angle": 0, + "content": "In this paper, we propose a simple, zero-shot method for creating multi-view illusions with off-the-shelf text-to-image diffusion models. In contrast to most previous work on computationally generating optical illusions [3-5, 10, 12, 15, 18, 20, 28, 31, 32, 38], our method does not require an explicit model of human perception. Rather, our approach builds on work that suggests generative models may process optical illusions in a way similar to humans [14, 23, 29]. In this way, our method is similar to recent work that uses diffusion models to create optical illusions by Burgert et al. [2] and Tancik [42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Our method can generate many types of classic illusions, such as images that change appearance when flipped or rotated (Fig. 1), as well as a new class of illusions which we term visual anagrams. These are images that change appearance under a permutation of their pixels. Image flips and rotations are a subset of these, as they can both be expressed as a permutation of pixels, but we also consider more exotic permutations. For example, we generate jigsaw puzzles that can be solved in two different ways, which we call \"polymorphic jigsaws.\" In addition, we successfully apply our approach to generating illusions with three and four views (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our method works by using a diffusion model to denoise an image from multiple views, obtaining multiple noise estimates. These noise estimates are then combined to form a single noise estimate which is used to perform a step in the reverse diffusion process. However, we show that care must be taken in choosing these views. For one, the transformation must keep the statistics of the noise intact, as the diffusion model is trained under the assumption of i.i.d. Gaus" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "sian noise. We provide an analysis of these conditions and give an exact specification of the class of transformations supported. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.14, + 0.892, + 0.17 + ], + "angle": 0, + "content": "- We present a simple yet effective method for generating multi-view optical illusions using diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.171, + 0.892, + 0.214 + ], + "angle": 0, + "content": "- We derive a precise description of the set of views that our method supports and provide empirical evidence that these views work." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.216, + 0.892, + 0.259 + ], + "angle": 0, + "content": "- We consider practical design decisions, crucial to optimizing the quality of generated illusions, and report ablations on our choices." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.262, + 0.892, + 0.291 + ], + "angle": 0, + "content": "- We provide quantitative and qualitative results, showcasing both the efficacy and flexibility of our method." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.14, + 0.892, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.642, + 0.322 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models [6, 17, 22, 35-37, 39-41] are a class of powerful generative models that iteratively convert a sample from a noise distribution to a sample from some data distribution. These models work by estimating the noise in a noisy sample, and removing the estimated noise following some update rule such as DDPM [22] or DDIM [40]. A prominent application of diffusion models has been text-conditioned image synthesis [24, 30, 36, 37]. In addition to a noisy image and a timestep, these models take a language model embedding of a text prompt as conditioning. Our approach is closely related to recent works that experiment with composing energy-based models and diffusion models [7-9, 13, 26, 27]. These approaches [9, 27] have shown that noise estimates from multiple conditional distributions can be combined together to obtain samples from compositions of the learned distributions. Our method uses a similar approach, and we apply it to the problem of multi-view illusion generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Computational Optical Illusions. Optical illusions serve as a testbed for understanding both human and machine perception [14, 19, 23, 29, 45]. We focus on generating illusions computationally, an area which has primarily relied on models of how our brains process external stimuli. Freeman et al. [12] create the illusion of constant motion in a desired direction by locally applying a filter with continuously shifting phase, relying on the observation that local phaseshifts are interpreted as global movement. Oliva et al. [31] propose a method to make \"hybrid images,\" which change appearance depending on the distance they are viewed from. This method takes advantage of the multiscale nature of human perception by blending high frequencies of one image with low frequencies from another. Chu et al. [5] camouflage objects in a scene through re-texturing, with additional constraints on luminance as to preserve salient features of the object, and other work camouflages objects from multiple viewpoints in 3D scenes [18, 32]. Recently, Chandra et al. [3] design color-constancy, size constancy, and face" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "perception illusions by differentiating through a Bayesian model of human vision. Our method likewise generates illusions, but does not depend on an explicit model of human perception. Instead, our method works by leveraging visual priors in diffusion models learned implicitly through data. This aligns with observations [14, 23, 29] that generative models process illusions similarly to humans, and predict the same ambiguities. From this perspective, we can view our method as leveraging generative, rather than discriminative, models to synthesize adversarial examples [16] against humans [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.473, + 0.707 + ], + "angle": 0, + "content": "Illusions with Diffusion Models. Very recently, artists and researchers have taken steps that show the potential of using diffusion models to create illusions. An artist under the pseudonym MrUgleh [43] repurposed a model finetuned for generating QR codes [25, 47] to create images whose global structure subtly matches a given template image. In contrast, we study multi-view illusions that can be created zero-shot from off-the-shelf diffusion models, and our illusions are specified via text rather than images. Burgert et al. [2] use score distillation sampling (SDS) [33, 44] to create images that align with different prompts from different views. While in principle this approach supports a superset of our views, the use of SDS results in significantly lower quality results, and the need for explicit optimization leads to long sampling times. Our method is most similar to a proof-of-concept by Tancik [42], which creates rotation illusions by sampling from a latent diffusion model [36] while alternating noise estimates between different views and prompts. While our technical approach is similar, by contrast we systematically study multi-view illusions, both by experimentally evaluating many different types of illusions and by providing a theoretical analysis of which views are (and are not) supported. In doing so, we go beyond just rotation views. We also make a number of improvements that result in qualitatively and quantitatively better illusions, such as by identifying a source of artifacts from latent diffusion, and by adding support for an arbitrary number of views. To our knowledge, we are the first to systematically evaluate illusions generated by these approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.17, + 0.736 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Our goal is to produce multi-view optical illusions using a pretrained diffusion model. That is, we seek to synthesize images that change appearance or identity when transformed, such as when flipped or rotated." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.38, + 0.832 + ], + "angle": 0, + "content": "3.1. Text-conditioned Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [22, 39, 41] take i.i.d. Gaussian noise, \\(\\mathbf{x}_T\\), and iteratively denoise it to produce a sample, \\(\\mathbf{x}_0\\), from some data distribution. These models are parameterized by a neural network which estimates the noise in" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.218, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Figure 2. Algorithm Overview. Our method works by simultaneously denoising multiple views of an image. Given a noisy image \\(\\mathbf{x}_t\\), we compute noise estimates, \\(\\epsilon_t^i\\), conditioned on different prompts, after applying views \\(v_i\\). We then apply the inverse view \\(v_i^{-1}\\) to align estimates, average the estimates, and perform a reverse diffusion step. The final output is an optical illusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.396 + ], + "angle": 0, + "content": "some intermediate, partially denoised data point \\(\\mathbf{x}_t\\), denoted as \\(\\epsilon_{\\theta}(\\mathbf{x}_t,y,t)\\), where \\(y\\) is some conditioning such as text prompts and \\(t\\) is the timestep in the diffusion process. The estimated noise is then used in an update rule [22, 40], from which \\(\\mathbf{x}_{t - 1}\\) is computed from \\(\\mathbf{x}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.487 + ], + "angle": 0, + "content": "To condition the diffusion model on another input, such as a text prompt, a common approach is to use classifier-free guidance [21]. With this method, unconditional noise estimates (usually obtained by passing the null text prompt as conditioning) and conditional noise estimates are combined together:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.497, + 0.892, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} ^ {\\mathrm {C F G}} = \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\varnothing) + \\gamma \\left(\\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, y\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\varnothing\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.894, + 0.632 + ], + "angle": 0, + "content": "Here, \\(\\varnothing\\) denotes the embedding of the empty string and \\(\\gamma\\) is a parameter that controls the strength of the guidance. Classifier-free guidance acts to sharpen the distribution of generated images to produce higher quality results. It also enables negative prompting [1], in which the empty text prompt embedding, \\(\\varnothing\\), is replaced by a text prompt that we would like to discourage the model from generating." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.679, + 0.656 + ], + "angle": 0, + "content": "3.2. Parallel Denoising" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.8 + ], + "angle": 0, + "content": "We produce multi-view illusions by using a diffusion model to simultaneously denoise multiple views of an image. Concretely, we take a set of \\(N\\) prompts, \\(y_{i}\\), each associated with a view function \\(v_{i}(\\cdot)\\), which applies a transformation to an image. These transformations may be, for example, the identity function, an image flip, or a permutation of pixels. Then given a diffusion model, \\(\\epsilon_{\\theta}(\\cdot)\\), and a partially denoised image, \\(\\mathbf{x}_t\\), we combine noise estimates from different views into a single noise estimate by averaging:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.803, + 0.892, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {t} = \\frac {1}{N} \\sum_ {i} v _ {i} ^ {- 1} \\left(\\epsilon_ {\\theta} \\left(v _ {i} \\left(\\mathbf {x} _ {t}\\right), y _ {i}, t\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Effectively, we use each view \\( v_{i} \\) to transform the noisy image \\( \\mathbf{x}_t \\), estimate the noise in the transformed images, and then apply \\( v_{i}^{-1} \\) to the estimates in order to transform them back to the original view. Taking an average of these noise" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.199 + ], + "angle": 0, + "content": "estimates gives us our combined noise estimate, which we can then use with our choice of diffusion sampler. We note that this technique of combining noise estimates is similar to previous work on compositionality [7-9, 13, 26, 27], where the idea is studied in further detail. In order to incorporate classifier-free guidance we simply replace the estimates \\(\\epsilon_{\\theta}(v_i(\\mathbf{x}_t),y_i,t)\\) with their classifier-free estimates, \\(\\epsilon_t^{\\mathrm{CFG}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.208, + 0.273, + 0.223 + ], + "angle": 0, + "content": "3.3. Conditions on Views" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.309 + ], + "angle": 0, + "content": "One straightforward condition for the views is that they must be invertible. But diffusion models also implicitly impose other conditions on the views \\( v_{i}(\\cdot) \\). We describe two such conditions below. We find that if these conditions are not satisfied, the denoising process produces poor results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.469, + 0.348 + ], + "angle": 0, + "content": "Linearity. The diffusion model, \\(\\epsilon_{\\theta}\\), acts on noisy images, \\(\\mathbf{x}_t\\). That is, specifically images of the form:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.351, + 0.469, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = w _ {t} ^ {\\text {s i g n a l}} \\underbrace {\\mathbf {x} _ {0}} _ {\\text {s i g n a l}} + w _ {t} ^ {\\text {n o i s e}} \\underbrace {\\epsilon} _ {\\text {n o i s e}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.545 + ], + "angle": 0, + "content": "The exact values of \\( w_{t}^{\\mathrm{signal}} \\) and \\( w_{t}^{\\mathrm{noise}} \\) depend on model implementation details such as the variance schedule, but are unimportant for our work, so we omit them for clarity. What is important is that \\( \\mathbf{x}_t \\) is a linear combination of pure signal, \\( \\mathbf{x}_0 \\), and pure noise, \\( \\epsilon \\), for some specific \\( w_{t}^{\\mathrm{signal}} \\) and \\( w_{t}^{\\mathrm{noise}} \\). Therefore our view \\( v_{i} \\) must take a noisy image \\( \\mathbf{x}_t \\) and transform it into a new noisy image \\( v_{i}(\\mathbf{x}_{t}) \\) that is also a linear combination of pure signal and pure noise with the same weighting. This can be achieved by requiring \\( v_{i} \\) to be a linear transformation, of the form" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.55, + 0.469, + 0.567 + ], + "angle": 0, + "content": "\\[\nv _ {i} \\left(\\mathbf {x} _ {t}\\right) = \\mathbf {A} _ {i} \\mathbf {x} _ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.469, + 0.616 + ], + "angle": 0, + "content": "for some matrix \\(\\mathbf{A}_i\\), and some flattened noisy image \\(\\mathbf{x}_t\\). By linearity, we are effectively applying the view \\(v_i\\) to the signal and the noise separately:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.619, + 0.469, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} v _ {i} \\left(\\mathbf {x} _ {t}\\right) = \\mathbf {A} _ {i} \\left(w _ {t} ^ {\\text {s i g n a l}} \\mathbf {x} _ {0} + w _ {t} ^ {\\text {n o i s e}} \\epsilon\\right) (5) \\\\ = w _ {t} ^ {\\text {s i g n a l}} \\underbrace {\\mathbf {A} _ {i} \\mathbf {x} _ {0}} _ {\\text {n e w s i g n a l}} + w _ {t} ^ {\\text {n o i s e}} \\underbrace {\\mathbf {A} _ {i} \\epsilon} _ {\\text {n e w n o i s e}}. (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.679, + 0.469, + 0.74 + ], + "angle": 0, + "content": "This results in a linear combination of transformed signal, \\(\\mathbf{A}_i\\mathbf{x}_0\\), and transformed noise, \\(\\mathbf{A}_i\\epsilon\\), weighted with the correct scaling factors. For further discussion, please see Appendix H." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Statistical Consistency. In addition to expecting a linear combination of signal and noise at a specific weighting, the diffusion model also expects the noise to have a precise distribution. In particular, most diffusion networks are trained with \\(\\epsilon \\sim \\mathcal{N}(0,I)\\). Therefore, we must ensure that our transformed noise, \\(\\mathbf{A}_i\\epsilon\\), is likewise drawn from \\(\\mathcal{N}(0,I)\\). This is true if and only if \\(\\mathbf{A}_i\\) is an orthogonal matrix. We provide a proof in Appendix I, but intuitively this fact reflects the spherical symmetry of the standard Gaussian density. Orthogonal transformations, being generalizations of rota" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "tions and flips to higher dimensions, preserve this spherically symmetric density. Note that these are rotations in the pixel values as opposed to spatial rotations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.151, + 0.673, + 0.166 + ], + "angle": 0, + "content": "3.4. Views Considered" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.176, + 0.892, + 0.268 + ], + "angle": 0, + "content": "The vast majority of orthogonal transformations applied to an image will not correspond to an intuitive image transformation. However, a number of these transformations do. Below, we enumerate the orthogonal transformations which we consider, all of which can be seen in the illusions in Fig. 1 unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.892, + 0.344 + ], + "angle": 0, + "content": "Identity. The simplest transformation we consider is the identity transformation. Using this view allows us to optimize the untransformed image to align with a chosen prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Standard Image Manipulations. We also consider spatial rotations of an image, which can be viewed as permutations of pixels. This works because permutations are in turn orthogonal. However, caution must be exercised when applying a rotation view, as common anti-aliasing operations such as bilinear sampling will modify the statistics of the noise. We discuss this further in Sec. 4.4. Spatial reflections are also permutations of pixels. As such we can use these views to generate illusions. Finally, we implement an approximation to skewing by rolling columns of pixels by different displacements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.648 + ], + "angle": 0, + "content": "General Permutations. We have already considered the special cases of spatial rotation, reflection, and skews but we can also consider other permutations. For example, we can divide an image into jigsaw pieces and rearrange these pieces to generate jigsaw puzzles with two solutions—what we call \"polymorphic\" jigsaw puzzles. Implementation details can be found in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.739 + ], + "angle": 0, + "content": "We also consider the extreme case of sampling a completely random permutation of pixels and treating it as our view. Additionally, we can reduce the complexity of this by considering permutations of square patches, rather than pixels. Examples of these illusions can be found in Fig. 6 and are discussed in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Finally, we consider rotating a circle within an image while leaving the rest of the image stationary, which we term inner rotations. Note that the permutations we consider are certainly not exhaustive, and many clever transformations exist which we do not study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.834, + 0.892, + 0.91 + ], + "angle": 0, + "content": "Color Inversion. Negation is an orthogonal transform; it is intuitively a 180 degree rotation generalized to higher dimensions. This allows us to generate illusions that change appearance upon color inversion, assuming pixel values are centered at 0 (e.g., in the range [-1, 1])." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.468, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.219, + 0.47, + 0.287 + ], + "angle": 0, + "content": "Figure 3. Latent-Based Artifacts. Manipulating the location of latent codes does not change the orientation of the blocks for which they encode. Therefore, when using latent diffusion models we see artifacts as shown above, in which straight lines are thatched under a rotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.468, + 0.419 + ], + "angle": 0, + "content": "Arbitrary Orthogonal Transformations. An arbitrary rotation of an image in pixel space is uninterpretable. Nevertheless, we demonstrate that our method works for these transformations as well. While these \"illusions\" are inscrutable to the human eye, they serve as confirmation that any orthogonal transformation works as a view with our method. These can be found in Fig. 7 and are discussed in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.429, + 0.243, + 0.444 + ], + "angle": 0, + "content": "3.5. Design Decisions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.469, + 0.482 + ], + "angle": 0, + "content": "Beyond the core method, we also consider various design decisions aimed at maximizing illusion quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.469, + 0.624 + ], + "angle": 0, + "content": "Pixel Diffusion Model. Previous work [42] performed multi-view denoising using Stable Diffusion [36], a latent diffusion model. However, the latent representation effectively encodes patches of pixels. This leads to artifacts under rotations or flips, where the location of latents change, but the content and orientation of these blocks do not. We show a qualitative example of this in Fig. 3, in which the model is forced to generate thatched lines to produce straight lines under a \\(90^{\\circ}\\) rotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.469, + 0.684 + ], + "angle": 0, + "content": "To ameliorate this issue, we implement our method using a pixel-based diffusion model, DeepFloyd IF [24]. DeepFloyd denoises directly on pixels, effectively side-stepping the problem of orientation in latent code blocks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Combining Noise Estimates. In addition to taking a mean of noise estimates from different views, we also consider alternating through them by timestep, using the estimate" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.751, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {t} = v _ {t \\bmod N} ^ {- 1} \\left(\\epsilon_ {\\theta} \\left(v _ {t \\bmod N} (\\mathbf {x} _ {t}), t, y)\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.469, + 0.803 + ], + "angle": 0, + "content": "This is the reduction strategy used by [42], but we show in ablations in Sec. 4.2 that it performs worse than averaging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Negative Prompting. We experiment with negative prompting [1] in the 2-view case by using one view's prompt as a negative for the other view, and vice versa. This encourages the model to hide the other view's prompt for a given view. For a discussion, please see the ablations in Sec. 4.2." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative Results. We report the alignment score, \\(\\mathcal{A}\\), and the concealment score, \\(\\mathcal{C}\\), as well as quantiles of these scores. For a discussion, please see Sec. 4.1." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.134, + 0.89, + 0.23 + ], + "angle": 0, + "content": "
Prompt PairMethodA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
CIFARBurgert et al. [2]0.2250.2530.2600.5010.5260.537
Tancik [42]0.2780.3100.3160.5950.6920.712
Ours0.2870.3210.3270.6240.7170.739
OursBurgert et al. [2]0.2330.2700.2830.5010.5260.538
Tancik [42]0.2560.2940.3090.5450.6210.655
Ours0.2750.3150.3260.5740.6680.694
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Table 2. Ablations. We ablate negative prompting, reduction methods, and guidance scales on our dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.273, + 0.889, + 0.394 + ], + "angle": 0, + "content": "
AblationA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
Negative Prompting0.240.270.2760.5760.6590.683
No Negative Prompting0.2550.2850.2950.5670.6430.679
Alternating Reduction0.2520.2860.2920.5600.6390.664
Mean Reduction0.2550.2850.2950.5670.6430.679
γ = 3.00.2390.2710.2850.5370.6100.629
γ = 7.00.2550.2850.2950.5670.6430.679
γ = 10.00.2590.2900.2970.5760.6640.702
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.587, + 0.431 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.517 + ], + "angle": 0, + "content": "We provide quantitative and qualitative results, and quantitative ablations. If not specified, qualitative results have been hand picked for quality. For random samples please see Fig. 8 and Appendix D. All implementation details can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.529, + 0.694, + 0.544 + ], + "angle": 0, + "content": "4.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Metrics. We use CLIP [34] to measure how well views align with the desired prompts. We consider two metrics derived from a score matrix \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times N}\\), defined as" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.603, + 0.891, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i j} = \\phi_ {\\mathrm {i m g}} \\left(v _ {i} (\\mathbf {x})\\right) ^ {T} \\phi_ {\\text {t e x t}} \\left(p _ {j}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.699 + ], + "angle": 0, + "content": "where \\(\\phi_{\\mathrm{img}}\\) and \\(\\phi_{\\mathrm{text}}\\) are the CLIP visual and textual encoders respectively, returning a unit-norm vector embedding. \\(\\mathbf{x}\\) is our generated illusion, and \\(v_{i}\\) are our views with associated prompts \\(p_i\\). A higher dot product indicates higher similarity between the image and text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.82 + ], + "angle": 0, + "content": "The first metric we consider is \\(\\min \\operatorname{diag}(\\mathbf{S})\\), which intuitively measures the worst alignment of all the views. We term this metric \\(\\mathcal{A}\\), the alignment score. However, this metric does not account for the possibility of seeing prompt \\(p_i\\) in view \\(v_j\\) for \\(i \\neq j\\). This is an occasional failure case of our method and to quantify this we propose a second derived metric which we term \\(\\mathcal{C}\\), the concealment score, computed as" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.821, + 0.891, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N} \\operatorname {t r} (\\operatorname {s o f t m a x} (S / \\tau)), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter of CLIP. In computing this metric we average both directions of the softmax, so that this metric measures how well CLIP can classify a view" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24158" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.472, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Flip View CLIP Score Distribution. We visualize trade-offs between flipped and unflipped views by plotting the distribution of CLIP scores on the datasets. Note that the quality of the flipped image is as good as the unflipped image, with parity indicated by the dashed line." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.354, + 0.343, + 0.37 + ], + "angle": 0, + "content": "as one of the \\(N\\) prompts and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.498 + ], + "angle": 0, + "content": "Dataset. To evaluate our method and baselines we compile two datasets of prompt pairs for 2-view illusions. One dataset uses the 10 classes from CIFAR-10 and contains a prompt per pair of classes, for a total of 45 prompt pairs. We refer to this as CIFAR. The other dataset we compile by hand, with the process documented in Appendix B. This dataset consists of 50 prompt pairs, which we refer to as Ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Baselines. We use two baselines that generate illusions using off-the-shelf diffusion models. One, which we denote \"Burgert et al. [2],\" uses Score Distillation Sampling. The other, which we denote \"Tancik [42],\" is an earlier version of our method, with differences discussed in more detail in Sec. 2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Results. We show results comparing our method to baselines on both datasets in Tab. 1 using vertical flips. We use vertical flips because it is a transformation supported by our method as well as the baselines. We use 10 samples per prompt, for a total of 450 and 500 samples for the CIFAR dataset and our dataset respectively. It is hard to perform a fair comparison with more samples because the Burgert et al. method uses SDS, which is quite slow1. Because we are particularly interested in the \"best-case\" performance, we also report quantiles of metrics, which we denote as \\(\\mathcal{A}_{0.9}\\) for the 90th percentile, for example. As can be seen, our method performs consistently better than the baselines, in both the alignment score and the concealment score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.471, + 0.866 + ], + "angle": 0, + "content": "In order to give a clearer understanding of trade-offs when optimizing two views, we show density plots which plot the CLIP scores of each of the two views of an illusion in Fig. 4. As can be seen, we do better than the baselines" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.889, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.441, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparisons. We compare illusions generated by baselines to our illusions. We show examples from both our prompt dataset and the CIFAR prompt dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.545 + ], + "angle": 0, + "content": "on average and in the best-case. Moreover, flipping during denoising does not hurt performance. The quality of the flipped images is as high as the unflipped images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.611, + 0.569 + ], + "angle": 0, + "content": "4.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.638 + ], + "angle": 0, + "content": "We ablate out the noise estimate reduction strategy, negative prompting, and the guidance scale in Tab. 2. We use our dataset, with 10 samples for each prompt for a total of 500 illusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Reduction Strategy. We find that mean reduction does better than alternating. Our hypothesis is that alternating the noise estimates results in \"thrashing,\" causing poor convergence. Moreover, we find that the alternating strategy gives poor results on illusions with more than 2 views, as each view has fewer denoising steps. Qualitative examples of this can be found in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Negative Prompting. When using negative prompting, care must be taken to omit any overlap between the negative and positive prompt. For example, given the two prompts \"oil painting of a dog\" and \"oil painting of a cat\", using one prompt as the negative for the other would simultaneously encourage and discourage the style \"oil painting\". Rather, the negative prompts should be \"a cat\" and \"a dog\" respectively. We find that negative prompting can improve the concealment score, indi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Sampling just 10 images per prompt already takes more than a week of GPU-hours." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24159" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.09, + 0.132, + 0.101 + ], + "angle": 0, + "content": "8×8" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.103, + 0.167, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.164, + 0.155, + 0.178 + ], + "angle": 0, + "content": "a pencil sket of a lemur" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.103, + 0.265, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.164, + 0.266, + 0.179 + ], + "angle": 0, + "content": "h a pencil sketch of a kangaroo" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.103, + 0.362, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.164, + 0.355, + 0.179 + ], + "angle": 0, + "content": "a watercolor of a rabbit" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.103, + 0.463, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.165, + 0.454, + 0.179 + ], + "angle": 0, + "content": "a watercolor of a duck" + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.184, + 0.455, + 0.192 + ], + "angle": 0, + "content": "cf. \"Kanichen und Ende.\" Fliegende Blätter, 1892." + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.189, + 0.137, + 0.197 + ], + "angle": 0, + "content": "[16×16]" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.195, + 0.216, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.257, + 0.216, + 0.273 + ], + "angle": 0, + "content": "an oil painting of a young man" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.194, + 0.321, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.195, + 0.412, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.257, + 0.413, + 0.273 + ], + "angle": 0, + "content": "an oil painting of an old man" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.278, + 0.139, + 0.288 + ], + "angle": 0, + "content": "32×32" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.292, + 0.164, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.354, + 0.168, + 0.369 + ], + "angle": 0, + "content": "an oil painting of a young man" + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.292, + 0.268, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.354, + 0.266, + 0.369 + ], + "angle": 0, + "content": "ng an oil paintingin of an old man ," + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.279, + 0.337, + 0.289 + ], + "angle": 0, + "content": "64×64" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.292, + 0.36, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.354, + 0.351, + 0.368 + ], + "angle": 0, + "content": "a mosaic of a duck" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.292, + 0.466, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.354, + 0.454, + 0.369 + ], + "angle": 0, + "content": "a mosaic of a rabbit" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.374, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Figure 6. Permutation Illusions. We synthesize images whose appearance changes upon permutation of patches. Even in the difficult case of a \\(64 \\times 64\\) grid of patches, in which every pixel is effectively shuffled, we are able to generate meaningful images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.579 + ], + "angle": 0, + "content": "cating that it is working as intended. But this comes at the cost of worse alignment score. This is because the negative and positive prompt may have fundamental similarities. For example, using \"a cat\" as the negative prompt for the prompt \"an oil painting of a dog\" may discourage the model from synthesizing anything remotely cat-like—such as fur, four legs, or a tail—even if it helps in producing a dog. For this reason we opt not to use negative prompting with our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.649 + ], + "angle": 0, + "content": "Guidance Scale. We also ablate out various guidance scales, \\(\\gamma\\), for our method. We find that a higher guidance scale tends to do better. This is presumably because a higher guidance scale results in a sharper sampling distribution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.658, + 0.26, + 0.673 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.471, + 0.846 + ], + "angle": 0, + "content": "We show qualitative results in Fig. 1, Fig. 5, Fig. 6, and Fig. 7. Again, random samples may be found in Fig. 8 and Appendix D. Additional qualitative samples can be found in Appendix C. Overall, we find that our method can produce very high quality optical illusions for a wide range of views. Interestingly, our method often finds clever ways of reusing elements from one view for another, such as in the \"waterfalls\"/\"rabbit\"/\"teddy bear\" three-view illusion in Fig. 1, in which the nose of the teddy bear is the eye of the rabbit, and a rock on the waterfalls." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Baselines. We provide qualitative comparisons of our method to baselines in Fig. 5, where we pick the best images out of 100 samples for each method. As can be seen," + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.091, + 0.584, + 0.105 + ], + "angle": 0, + "content": "an oil painting of an old woman" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.091, + 0.687, + 0.106 + ], + "angle": 0, + "content": "an oil painting of a young lady" + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.095, + 0.786, + 0.103 + ], + "angle": 0, + "content": "a photo of a duck" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.095, + 0.891, + 0.102 + ], + "angle": 0, + "content": "a photo of a rabbit" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.106, + 0.588, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.106, + 0.693, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.106, + 0.787, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.106, + 0.891, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.175, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Figure 7. Orthogonal Illusions. We show that our method works, even when the view is a randomly sampled orthogonal transformation \\(\\mathbf{A}\\). While these \"illusions\" are incomprehensible to human perception, they serve as a confirmation for our mathematical analysis." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.284 + ], + "angle": 0, + "content": "images generated using our method match the prompts in both views equally well and are of higher quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Permutations. Pixel and patch permutations, being a subset of orthogonal transformations, should work with our method. We show that this is indeed the case in Fig. 6, where we have results on patch grids of various sizes under randomly sampled permutations. The \\(64 \\times 64\\) case is quite hard, yet our method is able to generate images that satisfy the constraint, albeit at lower quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.407, + 0.893, + 0.604 + ], + "angle": 0, + "content": "Arbitrary Orthogonal Transformations. As discussed in Sec. 3.3, our method works for any orthogonal transformation. So far, we have shown illusions based on a subset of orthogonal views that correspond to intuitive image transformations. In Fig. 7, we show \"illusions\" using an arbitrary orthogonal transformation as a view. We use Stable Diffusion [36] and sample a random orthogonal matrix \\(\\mathbf{A} \\in \\mathbb{R}^{16384 \\times 16384}\\) by projecting an i.i.d. random Gaussian matrix with an SVD. These dimensions correspond to the size of the Stable Diffusion latent space. We note that this is an incredibly hard and unnatural transformation of an image, and results are accordingly of lower quality, but our method is still able to produce reasonable images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.612, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Random Samples. We show random samples for selected prompts in Fig. 8. As can be seen, these random samples, while not as good as those in Fig. 1, are still very high quality. Some failure cases can be seen where the model prefers one prompt over another. We add further discussion and present more random samples in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.713, + 0.599, + 0.728 + ], + "angle": 0, + "content": "4.4. Failures" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.736, + 0.892, + 0.767 + ], + "angle": 0, + "content": "We highlight three interesting failure cases of our method in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.911 + ], + "angle": 0, + "content": "Independent Synthesis. The first of these cases involves the model synthesizing prompts separately, without combining elements of the two to form an illusion. Empirically, this happens surprisingly rarely, especially given that it seems to be such an easy shortcut solution. We hypothesize that this is because the diffusion model is biased toward centering its content, resulting in far more images with content that is integrated and centered as opposed to separate and off-center." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24160" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.893, + 0.412 + ], + "angle": 0, + "content": "Figure 8. Random Samples. We show random samples, along with their corresponding view, for selected prompts. For more random samples please see Appendix D. For best quality, view digitally and zoom-in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.592 + ], + "angle": 0, + "content": "Noise Shift. Using views that preserve noise statistics is critical to our method's success. For example, we attempted to recreate the \"Dress\" illusion [46], in which a dress can be seen as either \"blue and black\" or \"white and gold.\" We used simple white balancing as our view, in which pixel values were scaled by a constant factor. While this transformation is linear, it does not preserve the statistics of Gaussian noise. As a result, we see artifacts in the forms of spots, which we hypothesize is the result of the model interpreting the scaled Gaussian noise as signal and actively denoising peaks in the scaled noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.47, + 0.755 + ], + "angle": 0, + "content": "Correlated Noise. While our method supports rotations as transformations, as demonstrated with the \"3-view,\" \"4-view,\" and \"Inner Rotation\" illusions in Fig. 1, care must be taken that the rotation does not introduce correlations in the noise, such as through anti-aliasing. For example, bilinear sampling introduces significant correlations in the noise, as it is a linear combination of four adjacent pixels. Therefore, seemingly innocuous rotations may result in divergent samples if transformations are not carefully kept correlation free, as shown with the 45 degree bilinear rotation in Fig. 9." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.342, + 0.786 + ], + "angle": 0, + "content": "5. Limitations and Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We present a method to produce compelling and diverse optical illusions. Our method is simple and straightforward to implement, and additionally amenable to theoretical analysis. We prove that our method works for a broad set of transformations, and qualitatively show that it can generate a wide array of optical illusions. However, at the same time many possible illusions and transformations are still" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.424, + 0.606, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.504, + 0.606, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.587, + 0.606, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.435, + 0.796, + 0.449 + ], + "angle": 0, + "content": "Failure: Independent Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.451, + 0.72, + 0.463 + ], + "angle": 0, + "content": "View: Vertical Flip" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.465, + 0.875, + 0.493 + ], + "angle": 0, + "content": "an oil painting of a bowl of fruit \nan oil painting of a monkey" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.517, + 0.729, + 0.529 + ], + "angle": 0, + "content": "Failure: Noise Shift" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.531, + 0.746, + 0.545 + ], + "angle": 0, + "content": "View: White Balancing" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.547, + 0.868, + 0.574 + ], + "angle": 0, + "content": "a photo of a black and blue dressa photo of a white and gold dress" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.598, + 0.762, + 0.61 + ], + "angle": 0, + "content": "Failure: Correlated Noise" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.613, + 0.878, + 0.627 + ], + "angle": 0, + "content": "View: Rotate Inner Circle 45 Degrees (Bilinear)" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.629, + 0.792, + 0.655 + ], + "angle": 0, + "content": "a sketch of an elephant a sketch of a mouse" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.667, + 0.89, + 0.694 + ], + "angle": 0, + "content": "Figure 9. Failures. We highlight three interesting failure cases, which are discussed in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.81 + ], + "angle": 0, + "content": "not possible using our method, such as color constancy illusions, homographies, stretches, and more generally nonvolume-preserving deformations. We leave implementation of these views for future work. Moreover, our method does not consistently produce perfect illusions. This may be a symptom of the difficulty of producing good illusions, but may indicate future work to be done to improve consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements We thank William Henning, Trenton Chang, Kimball Strong, Jeongsoo Park, Patrick Chao, Kurtland Chua, and Mohamed El Banani for their feedback on early drafts. Daniel is supported by the National Science Foundation Graduate Research Fellowship under Grant No. 1841052." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24161" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.095, + 0.174, + 0.111 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[1] AUTOMATIC1111. Negative prompt. https : // github . com / AUTOMATIC1111 / stable - diffusion - webui / wiki / Negative - prompt, 2022. Accessed: November 7, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[2] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael Ryoo. Diffusion illusions: Hiding images in plain sight. https://ryanndagreat.github.io/Diffusion-Illusions, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[3] Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum, and Jonathan Ragan-Kelley. Designing perceptual puzzles by differentiating probabilistic programs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[4] Ming-Te Chi, Chih-Yuan Yao, Eugene Zhang, and Tong-Yee Lee. Optical illusion shape texturing using repeated asymmetric patterns. The Visual Computer, 30:809–819, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[5] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.42, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.463, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[8] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.506, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[9] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[10] Werner Ehm. A variational approach to geometric-optical illusions modeling. Proceedings of Fechner Day, 27(1):41-46, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[11] Gamaleldin Elsayed, Shreya Shankar, Brian Cheung, Nicolas Papernot, Alexey Kurakin, Ian Goodfellow, and Jascha Sohl-Dickstein. Adversarial examples that fool both computer vision and time-limited humans. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[12] William T Freeman, Edward H Adelson, and David J Heeger. Motion without movement. ACM Siggraph Computer Graphics, 25(4):27-30, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[13] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Alex Gomez-Villa, Adrián Martín, Javier Vazquez-Corral, Marcelo Bertalmío, and Jesús Malo. On the synthesis of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.893, + 0.12 + ], + "angle": 0, + "content": "visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.893, + 0.163 + ], + "angle": 0, + "content": "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[17] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35:14715-14728, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[18] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[19] Aaron Hertzmann. Visual indeterminacy in gan art. In ACM SIGGRAPH 2020 Art Gallery, pages 424-428. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[23] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[24] Mikhail Konstantinov, Alex Shonenkov, Daria Bakshandaeva, and Ksenia Ivanova. If by deepfloyd lab at stabilityai, 2023. GitHub repository. 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[25] Monster Labs. Controlnet qr code monster v2 for sd-1.5, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[26] Nan Liu, Shuang Li, Yilun Du, Josh Tenenbaum, and Antonio Torralba. Learning to compose visual relations. Advances in Neural Information Processing Systems, 34: 23166-23178, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[27] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[28] Dominique Makowski, Zen J Lau, Tam Pham, W Paul Boyce, and SH Annabel Chen. A parametric framework to generate visual illusions using python. Perception, 50(11):950-965, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[29] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[31] Aude Oliva, Antonio Torralba, and Philippe G. Schyns. Hybrid images. ACM Trans. Graph., 25(3):527-532, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[32] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.893, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24162" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.191, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.416 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[38] Troy Shinbrot, Miguel Vivar Lazo, and Theo Siu. Network simulations of optical illusions. International Journal of Modern Physics C, 28(02):1750018, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[42] Matthew Tancik. Illusion diffusion. https://github.com/tancik/Illusion-Diffusion, 2023.2, 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[43] Ugleh. Spiral town - different approach to qr monster. https://www.reddit.com/r/StableDiffusion/comments/16ew9fz/spiral_town_differentapproach_to_qr_monster/, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.884 + ], + "angle": 0, + "content": "[45] Xi Wang, Zoya Bylinskii, Aaron Hertzmann, and Robert Pepperell. Toward quantifying ambiguities in artistic images. ACM Transactions on Applied Perception (TAP), 17(4):1-10, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[46] Wikipedia contributors. The dress. https://en." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "wikipedia.org/wiki/The_dress. Accessed: November 9, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24163" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_origin.pdf b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e182201f205ffc42bd3bc1821497cd223e1b5a94 --- /dev/null +++ b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/f2a047a9-fca8-496c-9ac0-94e4fff7168a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc2521fb6c1248a7fb2a8a372549bc9316eaaa8c743ac0f74d2072969fde83ca +size 6528379 diff --git a/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/full.md b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83199edca6abd0aaa1ec7bd36f284fcc50df6566 --- /dev/null +++ b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/full.md @@ -0,0 +1,514 @@ +# Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models + +Daniel Geng + +Inbum Park + +Andrew Owens + +University of Michigan + +https://dangeng.github.io/visual_anagrams/ + +![](images/987e86a9986e16caf859d08ef52f5288cce7239213cb341ce13a1756cd6d684e.jpg) +Flips +eupued pe e joo bututed e +a painting of kitchenware + +![](images/2ef23218fd0b728272b1004bdff997e429dc204dc0a90fdcf6c7c1680c97aaba.jpg) +xnp 0j uedabouy +a lithograph of a fish +"Sky and Water." M. C. Escher, 1938. + +![](images/9866bc691897863eed0186177327c89f907dff90b5f42b699d99224eac040e1c.jpg) +Bn 1000 +an oil painting of people at a campfire + +![](images/a5002171d43b964390e5b4e8d5536f69d7ae0de236088a8e8c42228dd75c6518.jpg) +e e e +a painting of vases + +![](images/fa560a81240e262e00a002d483d702f6c0814fbfc7caf12565d1d3e2920416ed.jpg) +unbued +e go butmexp e +a drawing of a giraffe + +![](images/17b26d7646fc32b6f3f6e973689e99a21fab5854e48c20f541913c62779d9acd.jpg) +pqrjgo +buTuTeD TIO ue +an oil paintingof a ship + +![](images/f456a772756a1f7c605e6a28e7258e91abc13d52a400a0a76a2435cb609da2de.jpg) +ueWOM PRO UP +0000 +a photo of a young lady +f. My Wife and My Mother-in-Law."1888, + +![](images/cfe296c804a04f820d1ded5cd8f930242d4f6b738755209214359697f6f2be26.jpg) +e +to butuared e +a painting of wine and cheese + +![](images/814d18dc060c9849f68422ce09e319896cee2d5748752b1ee7fcd13e78a9f597.jpg) +Three Views +1JQREIO +putaired R +a painting of +a teddy bear +STATE JOBUSTED TO + +![](images/3f406ae660298293948f2795a3d68208f454f82f1111a69578881394acbdb628.jpg) +a painting +of einstein + +![](images/7f3612421973e08ad23070c591a14346668c9ff18f251c09b96f783e5d2950d6.jpg) +Four Views +an oil painting +of a rabbit + +![](images/ab5446bbb66a3eaaea6b9ef7dcf3e936070ec8d9f39c7784130cdaaec7e924e8.jpg) +"Polymorphic" Jigsaw Puzzles +an oil painting of a fruit bowl + +![](images/b05af9a20c9095c3272e31b24ce458c6a3851db8537d2aeb118275dade15c495.jpg) +cf. "Reversible Head with Basket of Fruit." Giuseppe Arcimboldo, 1950. + +![](images/3ab94f84e4472e9192962bfb56e301da8ebe55cbc3ca5f8eb7a311df638806e9.jpg) + +![](images/25664e8a4b8663c777e78e78dcd23b4b30df5dad23e07c1389db1bbc968e7ca1.jpg) +an oil painting of a monkey + +![](images/ca4af49d55157db81c431f3d909ac5c65a386223850acb5b715c39d7ce239d5d.jpg) +a watercolor of a kitten + +![](images/0b324c3d93dea8478143771c28d2981dc493f74f3345ab4b3de840f2767b1622.jpg) + +![](images/f32c78681f796c55340ff92bfd5d13da89a91bbdf898b3c261b3c3af4bec3c70.jpg) +a watercolor of a puppy + +![](images/48895abba167250c0c1120004e8662599da7d4bf6e112e22ef21eb121cc9423f.jpg) +a painting of houseplants + +![](images/c80a59a8268180de20ecc68d6f660ab405f01ff88c7db1f679482956692380bb.jpg) + +![](images/05de3f3cdc2c3da0ee416b080e0ffb2df3902cebfcdcb509fbf0a59cc01b1152.jpg) + +![](images/3898c7aea2239ea57a07b4c83808a89e1e373bbff8e6564004a3779379ee12d9.jpg) +a painting of +marilyn monroe +m111yHnH +Aude Oliva et al. 2006. + +![](images/934841861477715a388eb484315e5741c424e207ae77287fe76462967ada9877.jpg) +Inner Rotations +a pop art of albert einstel + +![](images/a22f77bf05a1a396c3124f3191bd1eb9223e3a0de6c5e0997d022d1113a28c7b.jpg) + +![](images/13416d1ab527a0400c0154000a9b2ee990a27a5d052afacc3b890b0fe6256518.jpg) +a pop art of +marilyn monroe +cf. "Hybrid Images." + +![](images/074baec6063ef9c797d9147e38bcaeca649852d95ae0e4b90869204234cf5548.jpg) +Negatives +a lithograph of a teddy bear +Figure 1. Generating Multi-View Illusions. We propose a method for generating optical illusions from an off-the-shelf text-to-image diffusion model. We create images that match different prompts after undergoing a transformation. Our approach supports a variety of transformations, including flips, rotations, skews, color inversions, and jigsaw rearrangements. All images are hand selected. For random samples, please see Fig. 8 and Appendix D. For easier viewing, please see our webpage for animated versions of these illusions. + +![](images/af6f890fab785cbf670fb63a7dc9acffab7aa9a82c641981e68dfd1737235ba7.jpg) +a lithograph of a rabbit + +![](images/ed2a4e30afafb61febb75a05648eeffa19ae870cea53073c3c323903d85d0b58.jpg) +a lithograph of a landscape + +![](images/6454857f19ee4e6c2c329140bad8d1468cfcbf9bcc4e9feac5a93356071201f0.jpg) +a lithograph of houseplants + +![](images/6f77fb1c778a0974cd974b2b8bdbf4be85b0d05f146794ac3f8e2aa2ebef26ac.jpg) +Skews +an oil painting ofa tudor portrait +cf. "The Ambassadors." Hans Holbein, 1533. + +![](images/160dee21df8225e42a92d803bf76065cf631ea6c5f6a7e8f2e89fb80542cb4d7.jpg) +an oil painting +of a skull + +![](images/3c3ad80605fecab77af38c4afb6172e05ab2efa04195e404793e67db4e4c9087.jpg) +an oil painting +of a soldier + +![](images/3e7e6f14479d7436c07de49868eb1b48a9980f6af3142f8d71818a2c7f927a85.jpg) +an oil painting +of a houseplant + +![](images/940f09fb808c4d5569cf14b6e31fc4c27c4f4f5bb2d35f1b687bc3a86b0e7598.jpg) +Ambigrams +the word "happy" +cursive writing +$180^{\circ}$ + +![](images/05608620aa038aba6a1fdf954101c9b21f4dca55f4cd718625091c90c4fb977f.jpg) +the word "holiday". +cursive writing + +# Abstract + +We address the problem of synthesizing multi-view optical illusions: images that change appearance upon a transformation, such as a flip or rotation. We propose a simple, + +zero-shot method for obtaining these illusions from off-the-shelf text-to-image diffusion models. During the reverse diffusion process, we estimate the noise from different views of a noisy image, and then combine these noise estimates together and denoise the image. A theoretical analysis sug- + +gests that this method works precisely for views that can be written as orthogonal transformations, of which permutations are a subset. This leads to the idea of a visual anagram—an image that changes appearance under some rearrangement of pixels. This includes rotations and flips, but also more exotic pixel permutations such as a jigsaw rearrangement. Our approach also naturally extends to illusions with more than two views. We provide both qualitative and quantitative results demonstrating the effectiveness and flexibility of our method. Please see our project webpage for additional visualizations and results: https://dangeng.github.io/visual_anagrams/ + +# 1. Introduction + +Images that change their appearance under a transformation, such as a rotation or a flip, have long fascinated students of perception, from Salvador Dalí to M. C. Escher. The appeal of these multi-view optical illusions lies partly in the challenge of arranging visual elements such that they may be understood in multiple different ways. Creating these illusions requires accurately modeling—and then subverting—visual perception. + +In this paper, we propose a simple, zero-shot method for creating multi-view illusions with off-the-shelf text-to-image diffusion models. In contrast to most previous work on computationally generating optical illusions [3-5, 10, 12, 15, 18, 20, 28, 31, 32, 38], our method does not require an explicit model of human perception. Rather, our approach builds on work that suggests generative models may process optical illusions in a way similar to humans [14, 23, 29]. In this way, our method is similar to recent work that uses diffusion models to create optical illusions by Burgert et al. [2] and Tancik [42]. + +Our method can generate many types of classic illusions, such as images that change appearance when flipped or rotated (Fig. 1), as well as a new class of illusions which we term visual anagrams. These are images that change appearance under a permutation of their pixels. Image flips and rotations are a subset of these, as they can both be expressed as a permutation of pixels, but we also consider more exotic permutations. For example, we generate jigsaw puzzles that can be solved in two different ways, which we call "polymorphic jigsaws." In addition, we successfully apply our approach to generating illusions with three and four views (Fig. 1). + +Our method works by using a diffusion model to denoise an image from multiple views, obtaining multiple noise estimates. These noise estimates are then combined to form a single noise estimate which is used to perform a step in the reverse diffusion process. However, we show that care must be taken in choosing these views. For one, the transformation must keep the statistics of the noise intact, as the diffusion model is trained under the assumption of i.i.d. Gaus + +sian noise. We provide an analysis of these conditions and give an exact specification of the class of transformations supported. Our contributions are as follows: + +- We present a simple yet effective method for generating multi-view optical illusions using diffusion models. +- We derive a precise description of the set of views that our method supports and provide empirical evidence that these views work. +- We consider practical design decisions, crucial to optimizing the quality of generated illusions, and report ablations on our choices. +- We provide quantitative and qualitative results, showcasing both the efficacy and flexibility of our method. + +# 2. Related Work + +Diffusion Models. Diffusion models [6, 17, 22, 35-37, 39-41] are a class of powerful generative models that iteratively convert a sample from a noise distribution to a sample from some data distribution. These models work by estimating the noise in a noisy sample, and removing the estimated noise following some update rule such as DDPM [22] or DDIM [40]. A prominent application of diffusion models has been text-conditioned image synthesis [24, 30, 36, 37]. In addition to a noisy image and a timestep, these models take a language model embedding of a text prompt as conditioning. Our approach is closely related to recent works that experiment with composing energy-based models and diffusion models [7-9, 13, 26, 27]. These approaches [9, 27] have shown that noise estimates from multiple conditional distributions can be combined together to obtain samples from compositions of the learned distributions. Our method uses a similar approach, and we apply it to the problem of multi-view illusion generation. + +Computational Optical Illusions. Optical illusions serve as a testbed for understanding both human and machine perception [14, 19, 23, 29, 45]. We focus on generating illusions computationally, an area which has primarily relied on models of how our brains process external stimuli. Freeman et al. [12] create the illusion of constant motion in a desired direction by locally applying a filter with continuously shifting phase, relying on the observation that local phaseshifts are interpreted as global movement. Oliva et al. [31] propose a method to make "hybrid images," which change appearance depending on the distance they are viewed from. This method takes advantage of the multiscale nature of human perception by blending high frequencies of one image with low frequencies from another. Chu et al. [5] camouflage objects in a scene through re-texturing, with additional constraints on luminance as to preserve salient features of the object, and other work camouflages objects from multiple viewpoints in 3D scenes [18, 32]. Recently, Chandra et al. [3] design color-constancy, size constancy, and face + +perception illusions by differentiating through a Bayesian model of human vision. Our method likewise generates illusions, but does not depend on an explicit model of human perception. Instead, our method works by leveraging visual priors in diffusion models learned implicitly through data. This aligns with observations [14, 23, 29] that generative models process illusions similarly to humans, and predict the same ambiguities. From this perspective, we can view our method as leveraging generative, rather than discriminative, models to synthesize adversarial examples [16] against humans [11]. + +Illusions with Diffusion Models. Very recently, artists and researchers have taken steps that show the potential of using diffusion models to create illusions. An artist under the pseudonym MrUgleh [43] repurposed a model finetuned for generating QR codes [25, 47] to create images whose global structure subtly matches a given template image. In contrast, we study multi-view illusions that can be created zero-shot from off-the-shelf diffusion models, and our illusions are specified via text rather than images. Burgert et al. [2] use score distillation sampling (SDS) [33, 44] to create images that align with different prompts from different views. While in principle this approach supports a superset of our views, the use of SDS results in significantly lower quality results, and the need for explicit optimization leads to long sampling times. Our method is most similar to a proof-of-concept by Tancik [42], which creates rotation illusions by sampling from a latent diffusion model [36] while alternating noise estimates between different views and prompts. While our technical approach is similar, by contrast we systematically study multi-view illusions, both by experimentally evaluating many different types of illusions and by providing a theoretical analysis of which views are (and are not) supported. In doing so, we go beyond just rotation views. We also make a number of improvements that result in qualitatively and quantitatively better illusions, such as by identifying a source of artifacts from latent diffusion, and by adding support for an arbitrary number of views. To our knowledge, we are the first to systematically evaluate illusions generated by these approaches. + +# 3. Method + +Our goal is to produce multi-view optical illusions using a pretrained diffusion model. That is, we seek to synthesize images that change appearance or identity when transformed, such as when flipped or rotated. + +# 3.1. Text-conditioned Diffusion Models + +Diffusion models [22, 39, 41] take i.i.d. Gaussian noise, $\mathbf{x}_T$ , and iteratively denoise it to produce a sample, $\mathbf{x}_0$ , from some data distribution. These models are parameterized by a neural network which estimates the noise in + +![](images/489bdfa0d17e3a87fa56c8242c836b46eaab7b811772e79f5d1ceeb26146a288.jpg) +Figure 2. Algorithm Overview. Our method works by simultaneously denoising multiple views of an image. Given a noisy image $\mathbf{x}_t$ , we compute noise estimates, $\epsilon_t^i$ , conditioned on different prompts, after applying views $v_i$ . We then apply the inverse view $v_i^{-1}$ to align estimates, average the estimates, and perform a reverse diffusion step. The final output is an optical illusion. + +some intermediate, partially denoised data point $\mathbf{x}_t$ , denoted as $\epsilon_{\theta}(\mathbf{x}_t,y,t)$ , where $y$ is some conditioning such as text prompts and $t$ is the timestep in the diffusion process. The estimated noise is then used in an update rule [22, 40], from which $\mathbf{x}_{t - 1}$ is computed from $\mathbf{x}_t$ . + +To condition the diffusion model on another input, such as a text prompt, a common approach is to use classifier-free guidance [21]. With this method, unconditional noise estimates (usually obtained by passing the null text prompt as conditioning) and conditional noise estimates are combined together: + +$$ +\epsilon_ {t} ^ {\mathrm {C F G}} = \epsilon_ {\theta} (\mathbf {x} _ {t}, t, \varnothing) + \gamma \left(\epsilon_ {\theta} \left(\mathbf {x} _ {t}, t, y\right) - \epsilon_ {\theta} \left(\mathbf {x} _ {t}, t, \varnothing\right)\right). \tag {1} +$$ + +Here, $\varnothing$ denotes the embedding of the empty string and $\gamma$ is a parameter that controls the strength of the guidance. Classifier-free guidance acts to sharpen the distribution of generated images to produce higher quality results. It also enables negative prompting [1], in which the empty text prompt embedding, $\varnothing$ , is replaced by a text prompt that we would like to discourage the model from generating. + +# 3.2. Parallel Denoising + +We produce multi-view illusions by using a diffusion model to simultaneously denoise multiple views of an image. Concretely, we take a set of $N$ prompts, $y_{i}$ , each associated with a view function $v_{i}(\cdot)$ , which applies a transformation to an image. These transformations may be, for example, the identity function, an image flip, or a permutation of pixels. Then given a diffusion model, $\epsilon_{\theta}(\cdot)$ , and a partially denoised image, $\mathbf{x}_t$ , we combine noise estimates from different views into a single noise estimate by averaging: + +$$ +\tilde {\epsilon} _ {t} = \frac {1}{N} \sum_ {i} v _ {i} ^ {- 1} \left(\epsilon_ {\theta} \left(v _ {i} \left(\mathbf {x} _ {t}\right), y _ {i}, t\right)\right). \tag {2} +$$ + +Effectively, we use each view $v_{i}$ to transform the noisy image $\mathbf{x}_t$ , estimate the noise in the transformed images, and then apply $v_{i}^{-1}$ to the estimates in order to transform them back to the original view. Taking an average of these noise + +estimates gives us our combined noise estimate, which we can then use with our choice of diffusion sampler. We note that this technique of combining noise estimates is similar to previous work on compositionality [7-9, 13, 26, 27], where the idea is studied in further detail. In order to incorporate classifier-free guidance we simply replace the estimates $\epsilon_{\theta}(v_i(\mathbf{x}_t),y_i,t)$ with their classifier-free estimates, $\epsilon_t^{\mathrm{CFG}}$ + +# 3.3. Conditions on Views + +One straightforward condition for the views is that they must be invertible. But diffusion models also implicitly impose other conditions on the views $v_{i}(\cdot)$ . We describe two such conditions below. We find that if these conditions are not satisfied, the denoising process produces poor results. + +Linearity. The diffusion model, $\epsilon_{\theta}$ , acts on noisy images, $\mathbf{x}_t$ . That is, specifically images of the form: + +$$ +\mathbf {x} _ {t} = w _ {t} ^ {\text {s i g n a l}} \underbrace {\mathbf {x} _ {0}} _ {\text {s i g n a l}} + w _ {t} ^ {\text {n o i s e}} \underbrace {\epsilon} _ {\text {n o i s e}}. \tag {3} +$$ + +The exact values of $w_{t}^{\mathrm{signal}}$ and $w_{t}^{\mathrm{noise}}$ depend on model implementation details such as the variance schedule, but are unimportant for our work, so we omit them for clarity. What is important is that $\mathbf{x}_t$ is a linear combination of pure signal, $\mathbf{x}_0$ , and pure noise, $\epsilon$ , for some specific $w_{t}^{\mathrm{signal}}$ and $w_{t}^{\mathrm{noise}}$ . Therefore our view $v_{i}$ must take a noisy image $\mathbf{x}_t$ and transform it into a new noisy image $v_{i}(\mathbf{x}_{t})$ that is also a linear combination of pure signal and pure noise with the same weighting. This can be achieved by requiring $v_{i}$ to be a linear transformation, of the form + +$$ +v _ {i} \left(\mathbf {x} _ {t}\right) = \mathbf {A} _ {i} \mathbf {x} _ {t}, \tag {4} +$$ + +for some matrix $\mathbf{A}_i$ , and some flattened noisy image $\mathbf{x}_t$ . By linearity, we are effectively applying the view $v_i$ to the signal and the noise separately: + +$$ +\begin{array}{l} v _ {i} \left(\mathbf {x} _ {t}\right) = \mathbf {A} _ {i} \left(w _ {t} ^ {\text {s i g n a l}} \mathbf {x} _ {0} + w _ {t} ^ {\text {n o i s e}} \epsilon\right) (5) \\ = w _ {t} ^ {\text {s i g n a l}} \underbrace {\mathbf {A} _ {i} \mathbf {x} _ {0}} _ {\text {n e w s i g n a l}} + w _ {t} ^ {\text {n o i s e}} \underbrace {\mathbf {A} _ {i} \epsilon} _ {\text {n e w n o i s e}}. (6) \\ \end{array} +$$ + +This results in a linear combination of transformed signal, $\mathbf{A}_i\mathbf{x}_0$ , and transformed noise, $\mathbf{A}_i\epsilon$ , weighted with the correct scaling factors. For further discussion, please see Appendix H. + +Statistical Consistency. In addition to expecting a linear combination of signal and noise at a specific weighting, the diffusion model also expects the noise to have a precise distribution. In particular, most diffusion networks are trained with $\epsilon \sim \mathcal{N}(0,I)$ . Therefore, we must ensure that our transformed noise, $\mathbf{A}_i\epsilon$ , is likewise drawn from $\mathcal{N}(0,I)$ . This is true if and only if $\mathbf{A}_i$ is an orthogonal matrix. We provide a proof in Appendix I, but intuitively this fact reflects the spherical symmetry of the standard Gaussian density. Orthogonal transformations, being generalizations of rota + +tions and flips to higher dimensions, preserve this spherically symmetric density. Note that these are rotations in the pixel values as opposed to spatial rotations. + +# 3.4. Views Considered + +The vast majority of orthogonal transformations applied to an image will not correspond to an intuitive image transformation. However, a number of these transformations do. Below, we enumerate the orthogonal transformations which we consider, all of which can be seen in the illusions in Fig. 1 unless otherwise specified. + +Identity. The simplest transformation we consider is the identity transformation. Using this view allows us to optimize the untransformed image to align with a chosen prompt. + +Standard Image Manipulations. We also consider spatial rotations of an image, which can be viewed as permutations of pixels. This works because permutations are in turn orthogonal. However, caution must be exercised when applying a rotation view, as common anti-aliasing operations such as bilinear sampling will modify the statistics of the noise. We discuss this further in Sec. 4.4. Spatial reflections are also permutations of pixels. As such we can use these views to generate illusions. Finally, we implement an approximation to skewing by rolling columns of pixels by different displacements. + +General Permutations. We have already considered the special cases of spatial rotation, reflection, and skews but we can also consider other permutations. For example, we can divide an image into jigsaw pieces and rearrange these pieces to generate jigsaw puzzles with two solutions—what we call "polymorphic" jigsaw puzzles. Implementation details can be found in Appendix F. + +We also consider the extreme case of sampling a completely random permutation of pixels and treating it as our view. Additionally, we can reduce the complexity of this by considering permutations of square patches, rather than pixels. Examples of these illusions can be found in Fig. 6 and are discussed in Sec. 4.3. + +Finally, we consider rotating a circle within an image while leaving the rest of the image stationary, which we term inner rotations. Note that the permutations we consider are certainly not exhaustive, and many clever transformations exist which we do not study. + +Color Inversion. Negation is an orthogonal transform; it is intuitively a 180 degree rotation generalized to higher dimensions. This allows us to generate illusions that change appearance upon color inversion, assuming pixel values are centered at 0 (e.g., in the range [-1, 1]). + +![](images/227c327efbba6a7c244bb09a675c60ed147651e3088d6c1d4909d16a84a40076.jpg) +Figure 3. Latent-Based Artifacts. Manipulating the location of latent codes does not change the orientation of the blocks for which they encode. Therefore, when using latent diffusion models we see artifacts as shown above, in which straight lines are thatched under a rotation. + +Arbitrary Orthogonal Transformations. An arbitrary rotation of an image in pixel space is uninterpretable. Nevertheless, we demonstrate that our method works for these transformations as well. While these "illusions" are inscrutable to the human eye, they serve as confirmation that any orthogonal transformation works as a view with our method. These can be found in Fig. 7 and are discussed in Sec. 4.3. + +# 3.5. Design Decisions + +Beyond the core method, we also consider various design decisions aimed at maximizing illusion quality. + +Pixel Diffusion Model. Previous work [42] performed multi-view denoising using Stable Diffusion [36], a latent diffusion model. However, the latent representation effectively encodes patches of pixels. This leads to artifacts under rotations or flips, where the location of latents change, but the content and orientation of these blocks do not. We show a qualitative example of this in Fig. 3, in which the model is forced to generate thatched lines to produce straight lines under a $90^{\circ}$ rotation. + +To ameliorate this issue, we implement our method using a pixel-based diffusion model, DeepFloyd IF [24]. DeepFloyd denoises directly on pixels, effectively side-stepping the problem of orientation in latent code blocks. + +Combining Noise Estimates. In addition to taking a mean of noise estimates from different views, we also consider alternating through them by timestep, using the estimate + +$$ +\tilde {\epsilon} _ {t} = v _ {t \bmod N} ^ {- 1} \left(\epsilon_ {\theta} \left(v _ {t \bmod N} (\mathbf {x} _ {t}), t, y)\right)\right). \tag {7} +$$ + +This is the reduction strategy used by [42], but we show in ablations in Sec. 4.2 that it performs worse than averaging. + +Negative Prompting. We experiment with negative prompting [1] in the 2-view case by using one view's prompt as a negative for the other view, and vice versa. This encourages the model to hide the other view's prompt for a given view. For a discussion, please see the ablations in Sec. 4.2. + +Table 1. Quantitative Results. We report the alignment score, $\mathcal{A}$ , and the concealment score, $\mathcal{C}$ , as well as quantiles of these scores. For a discussion, please see Sec. 4.1. + +
Prompt PairMethodA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
CIFARBurgert et al. [2]0.2250.2530.2600.5010.5260.537
Tancik [42]0.2780.3100.3160.5950.6920.712
Ours0.2870.3210.3270.6240.7170.739
OursBurgert et al. [2]0.2330.2700.2830.5010.5260.538
Tancik [42]0.2560.2940.3090.5450.6210.655
Ours0.2750.3150.3260.5740.6680.694
+ +Table 2. Ablations. We ablate negative prompting, reduction methods, and guidance scales on our dataset. + +
AblationA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
Negative Prompting0.240.270.2760.5760.6590.683
No Negative Prompting0.2550.2850.2950.5670.6430.679
Alternating Reduction0.2520.2860.2920.5600.6390.664
Mean Reduction0.2550.2850.2950.5670.6430.679
γ = 3.00.2390.2710.2850.5370.6100.629
γ = 7.00.2550.2850.2950.5670.6430.679
γ = 10.00.2590.2900.2970.5760.6640.702
+ +# 4. Results + +We provide quantitative and qualitative results, and quantitative ablations. If not specified, qualitative results have been hand picked for quality. For random samples please see Fig. 8 and Appendix D. All implementation details can be found in Appendix A. + +# 4.1. Quantitative Results + +Metrics. We use CLIP [34] to measure how well views align with the desired prompts. We consider two metrics derived from a score matrix $\mathbf{S} \in \mathbb{R}^{N \times N}$ , defined as + +$$ +\mathbf {S} _ {i j} = \phi_ {\mathrm {i m g}} \left(v _ {i} (\mathbf {x})\right) ^ {T} \phi_ {\text {t e x t}} \left(p _ {j}\right), \tag {8} +$$ + +where $\phi_{\mathrm{img}}$ and $\phi_{\mathrm{text}}$ are the CLIP visual and textual encoders respectively, returning a unit-norm vector embedding. $\mathbf{x}$ is our generated illusion, and $v_{i}$ are our views with associated prompts $p_i$ . A higher dot product indicates higher similarity between the image and text. + +The first metric we consider is $\min \operatorname{diag}(\mathbf{S})$ , which intuitively measures the worst alignment of all the views. We term this metric $\mathcal{A}$ , the alignment score. However, this metric does not account for the possibility of seeing prompt $p_i$ in view $v_j$ for $i \neq j$ . This is an occasional failure case of our method and to quantify this we propose a second derived metric which we term $\mathcal{C}$ , the concealment score, computed as + +$$ +\frac {1}{N} \operatorname {t r} (\operatorname {s o f t m a x} (S / \tau)), \tag {9} +$$ + +where $\tau$ is the temperature parameter of CLIP. In computing this metric we average both directions of the softmax, so that this metric measures how well CLIP can classify a view + +![](images/6fc4757101a20b40ff7974893e548082e7e5c5f95a5eeea1d98597a3118d1b29.jpg) +Figure 4. Flip View CLIP Score Distribution. We visualize trade-offs between flipped and unflipped views by plotting the distribution of CLIP scores on the datasets. Note that the quality of the flipped image is as good as the unflipped image, with parity indicated by the dashed line. + +as one of the $N$ prompts and vice versa. + +Dataset. To evaluate our method and baselines we compile two datasets of prompt pairs for 2-view illusions. One dataset uses the 10 classes from CIFAR-10 and contains a prompt per pair of classes, for a total of 45 prompt pairs. We refer to this as CIFAR. The other dataset we compile by hand, with the process documented in Appendix B. This dataset consists of 50 prompt pairs, which we refer to as Ours. + +Baselines. We use two baselines that generate illusions using off-the-shelf diffusion models. One, which we denote "Burgert et al. [2]," uses Score Distillation Sampling. The other, which we denote "Tancik [42]," is an earlier version of our method, with differences discussed in more detail in Sec. 2 + +Results. We show results comparing our method to baselines on both datasets in Tab. 1 using vertical flips. We use vertical flips because it is a transformation supported by our method as well as the baselines. We use 10 samples per prompt, for a total of 450 and 500 samples for the CIFAR dataset and our dataset respectively. It is hard to perform a fair comparison with more samples because the Burgert et al. method uses SDS, which is quite slow1. Because we are particularly interested in the "best-case" performance, we also report quantiles of metrics, which we denote as $\mathcal{A}_{0.9}$ for the 90th percentile, for example. As can be seen, our method performs consistently better than the baselines, in both the alignment score and the concealment score. + +In order to give a clearer understanding of trade-offs when optimizing two views, we show density plots which plot the CLIP scores of each of the two views of an illusion in Fig. 4. As can be seen, we do better than the baselines + +![](images/6b503c0fc497282bbeadba423851a5a5ac4e0f229f0bbd9af5784d8508090082.jpg) +Figure 5. Qualitative Comparisons. We compare illusions generated by baselines to our illusions. We show examples from both our prompt dataset and the CIFAR prompt dataset. + +on average and in the best-case. Moreover, flipping during denoising does not hurt performance. The quality of the flipped images is as high as the unflipped images. + +# 4.2. Ablations + +We ablate out the noise estimate reduction strategy, negative prompting, and the guidance scale in Tab. 2. We use our dataset, with 10 samples for each prompt for a total of 500 illusions. + +Reduction Strategy. We find that mean reduction does better than alternating. Our hypothesis is that alternating the noise estimates results in "thrashing," causing poor convergence. Moreover, we find that the alternating strategy gives poor results on illusions with more than 2 views, as each view has fewer denoising steps. Qualitative examples of this can be found in Appendix G. + +Negative Prompting. When using negative prompting, care must be taken to omit any overlap between the negative and positive prompt. For example, given the two prompts "oil painting of a dog" and "oil painting of a cat", using one prompt as the negative for the other would simultaneously encourage and discourage the style "oil painting". Rather, the negative prompts should be "a cat" and "a dog" respectively. We find that negative prompting can improve the concealment score, indi + +![](images/abd14ef2b8aaffdeb18e4e2b2abe31c8bb3049e1a80febde2978136c9bc0378c.jpg) +8×8 +a pencil sket of a lemur +h a pencil sketch of a kangaroo + +![](images/60e13cf12c3f825e12d1f3e1aa39cd51d5186bf306fcb538da68bc03a54404b6.jpg) + +![](images/ebc9417e6357030c8d8d6cbc36c352109321d79e31e5ea1d6c345af1311393a3.jpg) +a watercolor of a rabbit + +![](images/dbb70ef7122727c40692d583b525d34c3e4d24e158ab9ccbee8883c8a07d71de.jpg) +a watercolor of a duck + +![](images/62985ce8ec139ba1c9ebbee338e48eccc0028d07a88f0f523cfd1a7116892138.jpg) +[16×16] +an oil painting of a young man + +![](images/831727bd920fa7d3762c66bda5f78631933079b8b99f7f936346f9feb90f9368.jpg) +cf. "Kanichen und Ende." Fliegende Blätter, 1892. + +![](images/75b032df22ddbc67ddfdfe6abea5b07081a84d38f3c33e5be0a4a01fd6cfcb69.jpg) +an oil painting of an old man + +![](images/6ae74fd3489ddf4d5c824aa3a324827e8daf3ccb3020cb313fc1eaf29a394413.jpg) +32×32 + +![](images/ddfd37af6f4c49f3d525a16387cc64a3b969a7b7a4d388b7f124363f3e6f08cb.jpg) +an oil painting of a young man +ng an oil paintingin of an old man , +Figure 6. Permutation Illusions. We synthesize images whose appearance changes upon permutation of patches. Even in the difficult case of a $64 \times 64$ grid of patches, in which every pixel is effectively shuffled, we are able to generate meaningful images. + +![](images/69abf4c288cc173d6e075a7862e1c00400346fbb6d1a793753a09d465e2eadfa.jpg) +64×64 +a mosaic of a duck + +![](images/ac663315c9966c922d45d6c64a28d582a7a68b3ea671b85d46840ae8a61712d0.jpg) +a mosaic of a rabbit + +cating that it is working as intended. But this comes at the cost of worse alignment score. This is because the negative and positive prompt may have fundamental similarities. For example, using "a cat" as the negative prompt for the prompt "an oil painting of a dog" may discourage the model from synthesizing anything remotely cat-like—such as fur, four legs, or a tail—even if it helps in producing a dog. For this reason we opt not to use negative prompting with our method. + +Guidance Scale. We also ablate out various guidance scales, $\gamma$ , for our method. We find that a higher guidance scale tends to do better. This is presumably because a higher guidance scale results in a sharper sampling distribution. + +# 4.3. Qualitative Results + +We show qualitative results in Fig. 1, Fig. 5, Fig. 6, and Fig. 7. Again, random samples may be found in Fig. 8 and Appendix D. Additional qualitative samples can be found in Appendix C. Overall, we find that our method can produce very high quality optical illusions for a wide range of views. Interestingly, our method often finds clever ways of reusing elements from one view for another, such as in the "waterfalls"/"rabbit"/"teddy bear" three-view illusion in Fig. 1, in which the nose of the teddy bear is the eye of the rabbit, and a rock on the waterfalls. + +Baselines. We provide qualitative comparisons of our method to baselines in Fig. 5, where we pick the best images out of 100 samples for each method. As can be seen, + +![](images/6780f9bb96e032916b5336a8617d9b81193e4d9df5cc3166f4ef6346bb31ff81.jpg) +an oil painting of an old woman +Figure 7. Orthogonal Illusions. We show that our method works, even when the view is a randomly sampled orthogonal transformation $\mathbf{A}$ . While these "illusions" are incomprehensible to human perception, they serve as a confirmation for our mathematical analysis. + +![](images/8dfa5c16125480acdd32228222eeb38fa866f901767a7b32e08355c6d61a21e6.jpg) +an oil painting of a young lady + +![](images/0c13a2c339044b8068911dd4afa40ddb3abdc55d05a1949fdc16a904b9284df7.jpg) +a photo of a duck + +![](images/612aa866ee39112ad67565a5a49a38d7576333f7c035cf85bbce3d8ca7236025.jpg) +a photo of a rabbit + +images generated using our method match the prompts in both views equally well and are of higher quality. + +Permutations. Pixel and patch permutations, being a subset of orthogonal transformations, should work with our method. We show that this is indeed the case in Fig. 6, where we have results on patch grids of various sizes under randomly sampled permutations. The $64 \times 64$ case is quite hard, yet our method is able to generate images that satisfy the constraint, albeit at lower quality. + +Arbitrary Orthogonal Transformations. As discussed in Sec. 3.3, our method works for any orthogonal transformation. So far, we have shown illusions based on a subset of orthogonal views that correspond to intuitive image transformations. In Fig. 7, we show "illusions" using an arbitrary orthogonal transformation as a view. We use Stable Diffusion [36] and sample a random orthogonal matrix $\mathbf{A} \in \mathbb{R}^{16384 \times 16384}$ by projecting an i.i.d. random Gaussian matrix with an SVD. These dimensions correspond to the size of the Stable Diffusion latent space. We note that this is an incredibly hard and unnatural transformation of an image, and results are accordingly of lower quality, but our method is still able to produce reasonable images. + +Random Samples. We show random samples for selected prompts in Fig. 8. As can be seen, these random samples, while not as good as those in Fig. 1, are still very high quality. Some failure cases can be seen where the model prefers one prompt over another. We add further discussion and present more random samples in Appendix D. + +# 4.4. Failures + +We highlight three interesting failure cases of our method in Fig. 9. + +Independent Synthesis. The first of these cases involves the model synthesizing prompts separately, without combining elements of the two to form an illusion. Empirically, this happens surprisingly rarely, especially given that it seems to be such an easy shortcut solution. We hypothesize that this is because the diffusion model is biased toward centering its content, resulting in far more images with content that is integrated and centered as opposed to separate and off-center. + +![](images/a15a49f015b1205e9f1ca11d1635a681b97bf6857b0c8dcdd8536a087c150025.jpg) +Figure 8. Random Samples. We show random samples, along with their corresponding view, for selected prompts. For more random samples please see Appendix D. For best quality, view digitally and zoom-in. + +Noise Shift. Using views that preserve noise statistics is critical to our method's success. For example, we attempted to recreate the "Dress" illusion [46], in which a dress can be seen as either "blue and black" or "white and gold." We used simple white balancing as our view, in which pixel values were scaled by a constant factor. While this transformation is linear, it does not preserve the statistics of Gaussian noise. As a result, we see artifacts in the forms of spots, which we hypothesize is the result of the model interpreting the scaled Gaussian noise as signal and actively denoising peaks in the scaled noise. + +Correlated Noise. While our method supports rotations as transformations, as demonstrated with the "3-view," "4-view," and "Inner Rotation" illusions in Fig. 1, care must be taken that the rotation does not introduce correlations in the noise, such as through anti-aliasing. For example, bilinear sampling introduces significant correlations in the noise, as it is a linear combination of four adjacent pixels. Therefore, seemingly innocuous rotations may result in divergent samples if transformations are not carefully kept correlation free, as shown with the 45 degree bilinear rotation in Fig. 9. + +# 5. Limitations and Conclusions + +We present a method to produce compelling and diverse optical illusions. Our method is simple and straightforward to implement, and additionally amenable to theoretical analysis. We prove that our method works for a broad set of transformations, and qualitatively show that it can generate a wide array of optical illusions. However, at the same time many possible illusions and transformations are still + +![](images/d49d0720d5e01c1700bf17fed202ce36077ef1225d768aa5dc67a4a1daf839c9.jpg) + +![](images/f3e25a3767a1636fbd3ad4548a33c27ee05256106c2a1129c3634eea9b1c9e7a.jpg) + +![](images/463e60f322959c2406320ce19530d2cca2c11e2cdc3690651a36d06b9d582367.jpg) +Figure 9. Failures. We highlight three interesting failure cases, which are discussed in Sec. 4.4. + +Failure: Independent Synthesis + +View: Vertical Flip + +an oil painting of a bowl of fruit +an oil painting of a monkey + +Failure: Noise Shift + +View: White Balancing + +a photo of a black and blue dressa photo of a white and gold dress + +Failure: Correlated Noise + +View: Rotate Inner Circle 45 Degrees (Bilinear) + +a sketch of an elephant a sketch of a mouse + +not possible using our method, such as color constancy illusions, homographies, stretches, and more generally nonvolume-preserving deformations. We leave implementation of these views for future work. Moreover, our method does not consistently produce perfect illusions. This may be a symptom of the difficulty of producing good illusions, but may indicate future work to be done to improve consistency. + +Acknowledgements We thank William Henning, Trenton Chang, Kimball Strong, Jeongsoo Park, Patrick Chao, Kurtland Chua, and Mohamed El Banani for their feedback on early drafts. Daniel is supported by the National Science Foundation Graduate Research Fellowship under Grant No. 1841052. + +# References + +[1] AUTOMATIC1111. Negative prompt. https : // github . com / AUTOMATIC1111 / stable - diffusion - webui / wiki / Negative - prompt, 2022. Accessed: November 7, 2023. 3, 5 +[2] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael Ryoo. Diffusion illusions: Hiding images in plain sight. https://ryanndagreat.github.io/Diffusion-Illusions, 2023. 2, 3, 5, 6 +[3] Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum, and Jonathan Ragan-Kelley. Designing perceptual puzzles by differentiating probabilistic programs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2 +[4] Ming-Te Chi, Chih-Yuan Yao, Eugene Zhang, and Tong-Yee Lee. Optical illusion shape texturing using repeated asymmetric patterns. The Visual Computer, 30:809–819, 2014. +[5] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 2 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 2, 4 +[8] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020. +[9] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 2, 4 +[10] Werner Ehm. A variational approach to geometric-optical illusions modeling. Proceedings of Fechner Day, 27(1):41-46, 2011. 2 +[11] Gamaleldin Elsayed, Shreya Shankar, Brian Cheung, Nicolas Papernot, Alexey Kurakin, Ian Goodfellow, and Jascha Sohl-Dickstein. Adversarial examples that fool both computer vision and time-limited humans. Advances in neural information processing systems, 31, 2018. 3 +[12] William T Freeman, Edward H Adelson, and David J Heeger. Motion without movement. ACM Siggraph Computer Graphics, 25(4):27-30, 1991. 2 +[13] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023. 2, 4 +[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 3 +[15] Alex Gomez-Villa, Adrián Martín, Javier Vazquez-Corral, Marcelo Bertalmío, and Jesús Malo. On the synthesis of + +visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 2 +[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3 +[17] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35:14715-14728, 2022. 2 +[18] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[19] Aaron Hertzmann. Visual indeterminacy in gan art. In ACM SIGGRAPH 2020 Art Gallery, pages 424-428. 2020. 2 +[20] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2, 3 +[23] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023. 2, 3 +[24] Mikhail Konstantinov, Alex Shonenkov, Daria Bakshandaeva, and Ksenia Ivanova. If by deepfloyd lab at stabilityai, 2023. GitHub repository. 2, 5, 11 +[25] Monster Labs. Controlnet qr code monster v2 for sd-1.5, 2023.3 +[26] Nan Liu, Shuang Li, Yilun Du, Josh Tenenbaum, and Antonio Torralba. Learning to compose visual relations. Advances in Neural Information Processing Systems, 34: 23166-23178, 2021. 2, 4 +[27] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4 +[28] Dominique Makowski, Zen J Lau, Tam Pham, W Paul Boyce, and SH Annabel Chen. A parametric framework to generate visual illusions using python. Perception, 50(11):950-965, 2021. 2 +[29] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? 2023. 2, 3 +[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2 +[31] Aude Oliva, Antonio Torralba, and Philippe G. Schyns. Hybrid images. ACM Trans. Graph., 25(3):527-532, 2006. 2 +[32] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. 2014. 2 +[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3 + +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5 +[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2 +[38] Troy Shinbrot, Miguel Vivar Lazo, and Theo Siu. Network simulations of optical illusions. International Journal of Modern Physics C, 28(02):1750018, 2017. 2 +[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2, 3 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 3 +[42] Matthew Tancik. Illusion diffusion. https://github.com/tancik/Illusion-Diffusion, 2023.2, 3, 5, 6, 12 +[43] Ugleh. Spiral town - different approach to qr monster. https://www.reddit.com/r/StableDiffusion/comments/16ew9fz/spiral_town_differentapproach_to_qr_monster/, 2023.3 +[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3 +[45] Xi Wang, Zoya Bylinskii, Aaron Hertzmann, and Robert Pepperell. Toward quantifying ambiguities in artistic images. ACM Transactions on Applied Perception (TAP), 17(4):1-10, 2020. 2 +[46] Wikipedia contributors. The dress. https://en. + +wikipedia.org/wiki/The_dress. Accessed: November 9, 2023. 8 +[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3 \ No newline at end of file diff --git a/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/images.zip b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1cd7283ee59754752aefb2059bda044e75b853c --- /dev/null +++ b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51e583ab8103ff37d2d0a138cac6f53327939a2526fbe248b0c4b0bbf7427b11 +size 836663 diff --git a/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/layout.json b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..728ba9fd739f0f458be74a446e58c31532a2e68c --- /dev/null +++ b/2024/Visual Anagrams_ Generating Multi-View Optical Illusions with Diffusion Models/layout.json @@ -0,0 +1,13480 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 103, + 427, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 103, + 427, + 140 + ], + "spans": [ + { + "bbox": [ + 167, + 103, + 427, + 140 + ], + "type": "text", + "content": "Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 173, + 161, + 237, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 161, + 237, + 175 + ], + "spans": [ + { + "bbox": [ + 173, + 161, + 237, + 175 + ], + "type": "text", + "content": "Daniel Geng" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 162, + 319, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 162, + 319, + 175 + ], + "spans": [ + { + "bbox": [ + 261, + 162, + 319, + 175 + ], + "type": "text", + "content": "Inbum Park" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 343, + 162, + 420, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 162, + 420, + 175 + ], + "spans": [ + { + "bbox": [ + 343, + 162, + 420, + 175 + ], + "type": "text", + "content": "Andrew Owens" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 240, + 176, + 353, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 176, + 353, + 190 + ], + "spans": [ + { + "bbox": [ + 240, + 176, + 353, + 190 + ], + "type": "text", + "content": "University of Michigan" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 190, + 446, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 190, + 446, + 204 + ], + "spans": [ + { + "bbox": [ + 143, + 190, + 446, + 204 + ], + "type": "text", + "content": "https://dangeng.github.io/visual_anagrams/" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 56, + 249, + 106, + 299 + ], + "blocks": [ + { + "bbox": [ + 62, + 216, + 78, + 224 + ], + "lines": [ + { + "bbox": [ + 62, + 216, + 78, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 78, + 224 + ], + "type": "text", + "content": "Flips" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 236, + 99, + 247 + ], + "lines": [ + { + "bbox": [ + 63, + 236, + 99, + 247 + ], + "spans": [ + { + "bbox": [ + 63, + 236, + 99, + 247 + ], + "type": "text", + "content": "eupued pe e joo bututed e" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 249, + 106, + 299 + ], + "lines": [ + { + "bbox": [ + 56, + 249, + 106, + 299 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 106, + 299 + ], + "type": "image", + "image_path": "987e86a9986e16caf859d08ef52f5288cce7239213cb341ce13a1756cd6d684e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 300, + 99, + 311 + ], + "lines": [ + { + "bbox": [ + 62, + 300, + 99, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 99, + 311 + ], + "type": "text", + "content": "a painting of kitchenware" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 108, + 249, + 158, + 299 + ], + "blocks": [ + { + "bbox": [ + 117, + 236, + 149, + 247 + ], + "lines": [ + { + "bbox": [ + 117, + 236, + 149, + 247 + ], + "spans": [ + { + "bbox": [ + 117, + 236, + 149, + 247 + ], + "type": "text", + "content": "xnp 0j uedabouy" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 249, + 158, + 299 + ], + "lines": [ + { + "bbox": [ + 108, + 249, + 158, + 299 + ], + "spans": [ + { + "bbox": [ + 108, + 249, + 158, + 299 + ], + "type": "image", + "image_path": "2ef23218fd0b728272b1004bdff997e429dc204dc0a90fdcf6c7c1680c97aaba.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 300, + 150, + 311 + ], + "lines": [ + { + "bbox": [ + 115, + 300, + 150, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 300, + 150, + 311 + ], + "type": "text", + "content": "a lithograph of a fish" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 113, + 312, + 151, + 323 + ], + "lines": [ + { + "bbox": [ + 113, + 312, + 151, + 323 + ], + "spans": [ + { + "bbox": [ + 113, + 312, + 151, + 323 + ], + "type": "text", + "content": "\"Sky and Water.\" M. C. Escher, 1938." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 160, + 248, + 211, + 299 + ], + "blocks": [ + { + "bbox": [ + 165, + 236, + 206, + 247 + ], + "lines": [ + { + "bbox": [ + 165, + 236, + 206, + 247 + ], + "spans": [ + { + "bbox": [ + 165, + 236, + 206, + 247 + ], + "type": "text", + "content": "Bn 1000" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 248, + 211, + 299 + ], + "lines": [ + { + "bbox": [ + 160, + 248, + 211, + 299 + ], + "spans": [ + { + "bbox": [ + 160, + 248, + 211, + 299 + ], + "type": "image", + "image_path": "9866bc691897863eed0186177327c89f907dff90b5f42b699d99224eac040e1c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 300, + 206, + 317 + ], + "lines": [ + { + "bbox": [ + 164, + 300, + 206, + 317 + ], + "spans": [ + { + "bbox": [ + 164, + 300, + 206, + 317 + ], + "type": "text", + "content": "an oil painting of people at a campfire" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 214, + 248, + 263, + 299 + ], + "blocks": [ + { + "bbox": [ + 223, + 236, + 252, + 247 + ], + "lines": [ + { + "bbox": [ + 223, + 236, + 252, + 247 + ], + "spans": [ + { + "bbox": [ + 223, + 236, + 252, + 247 + ], + "type": "text", + "content": "e e e" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 248, + 263, + 299 + ], + "lines": [ + { + "bbox": [ + 214, + 248, + 263, + 299 + ], + "spans": [ + { + "bbox": [ + 214, + 248, + 263, + 299 + ], + "type": "image", + "image_path": "a5002171d43b964390e5b4e8d5536f69d7ae0de236088a8e8c42228dd75c6518.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 300, + 252, + 311 + ], + "lines": [ + { + "bbox": [ + 222, + 300, + 252, + 311 + ], + "spans": [ + { + "bbox": [ + 222, + 300, + 252, + 311 + ], + "type": "text", + "content": "a painting of vases" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 55, + 346, + 105, + 396 + ], + "blocks": [ + { + "bbox": [ + 69, + 333, + 91, + 338 + ], + "lines": [ + { + "bbox": [ + 69, + 333, + 91, + 338 + ], + "spans": [ + { + "bbox": [ + 69, + 333, + 91, + 338 + ], + "type": "text", + "content": "unbued" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 339, + 102, + 345 + ], + "lines": [ + { + "bbox": [ + 59, + 339, + 102, + 345 + ], + "spans": [ + { + "bbox": [ + 59, + 339, + 102, + 345 + ], + "type": "text", + "content": "e go butmexp e" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 346, + 105, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 105, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 105, + 396 + ], + "type": "image", + "image_path": "fa560a81240e262e00a002d483d702f6c0814fbfc7caf12565d1d3e2920416ed.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 397, + 100, + 408 + ], + "lines": [ + { + "bbox": [ + 60, + 397, + 100, + 408 + ], + "spans": [ + { + "bbox": [ + 60, + 397, + 100, + 408 + ], + "type": "text", + "content": "a drawing of a giraffe" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 108, + 346, + 158, + 396 + ], + "blocks": [ + { + "bbox": [ + 119, + 334, + 149, + 339 + ], + "lines": [ + { + "bbox": [ + 119, + 334, + 149, + 339 + ], + "spans": [ + { + "bbox": [ + 119, + 334, + 149, + 339 + ], + "type": "text", + "content": "pqrjgo" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 340, + 156, + 345 + ], + "lines": [ + { + "bbox": [ + 108, + 340, + 156, + 345 + ], + "spans": [ + { + "bbox": [ + 108, + 340, + 156, + 345 + ], + "type": "text", + "content": "buTuTeD TIO ue" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 346, + 158, + 396 + ], + "lines": [ + { + "bbox": [ + 108, + 346, + 158, + 396 + ], + "spans": [ + { + "bbox": [ + 108, + 346, + 158, + 396 + ], + "type": "image", + "image_path": "17b26d7646fc32b6f3f6e973689e99a21fab5854e48c20f541913c62779d9acd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 397, + 154, + 408 + ], + "lines": [ + { + "bbox": [ + 110, + 397, + 154, + 408 + ], + "spans": [ + { + "bbox": [ + 110, + 397, + 154, + 408 + ], + "type": "text", + "content": "an oil paintingof a ship" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 160, + 346, + 211, + 396 + ], + "blocks": [ + { + "bbox": [ + 168, + 334, + 202, + 339 + ], + "lines": [ + { + "bbox": [ + 168, + 334, + 202, + 339 + ], + "spans": [ + { + "bbox": [ + 168, + 334, + 202, + 339 + ], + "type": "text", + "content": "ueWOM PRO UP" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 340, + 200, + 345 + ], + "lines": [ + { + "bbox": [ + 166, + 340, + 200, + 345 + ], + "spans": [ + { + "bbox": [ + 166, + 340, + 200, + 345 + ], + "type": "text", + "content": "0000" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 346, + 211, + 396 + ], + "lines": [ + { + "bbox": [ + 160, + 346, + 211, + 396 + ], + "spans": [ + { + "bbox": [ + 160, + 346, + 211, + 396 + ], + "type": "image", + "image_path": "f456a772756a1f7c605e6a28e7258e91abc13d52a400a0a76a2435cb609da2de.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 398, + 203, + 409 + ], + "lines": [ + { + "bbox": [ + 167, + 398, + 203, + 409 + ], + "spans": [ + { + "bbox": [ + 167, + 398, + 203, + 409 + ], + "type": "text", + "content": "a photo of a young lady" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 165, + 410, + 204, + 420 + ], + "lines": [ + { + "bbox": [ + 165, + 410, + 204, + 420 + ], + "spans": [ + { + "bbox": [ + 165, + 410, + 204, + 420 + ], + "type": "text", + "content": "f. My Wife and My Mother-in-Law.\"1888," + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 212, + 346, + 263, + 396 + ], + "blocks": [ + { + "bbox": [ + 225, + 334, + 250, + 339 + ], + "lines": [ + { + "bbox": [ + 225, + 334, + 250, + 339 + ], + "spans": [ + { + "bbox": [ + 225, + 334, + 250, + 339 + ], + "type": "text", + "content": "e" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 339, + 257, + 345 + ], + "lines": [ + { + "bbox": [ + 222, + 339, + 257, + 345 + ], + "spans": [ + { + "bbox": [ + 222, + 339, + 257, + 345 + ], + "type": "text", + "content": "to butuared e" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 346, + 263, + 396 + ], + "lines": [ + { + "bbox": [ + 212, + 346, + 263, + 396 + ], + "spans": [ + { + "bbox": [ + 212, + 346, + 263, + 396 + ], + "type": "image", + "image_path": "cfe296c804a04f820d1ded5cd8f930242d4f6b738755209214359697f6f2be26.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 397, + 258, + 408 + ], + "lines": [ + { + "bbox": [ + 216, + 397, + 258, + 408 + ], + "spans": [ + { + "bbox": [ + 216, + 397, + 258, + 408 + ], + "type": "text", + "content": "a painting of wine and cheese" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 54, + 450, + 115, + 502 + ], + "blocks": [ + { + "bbox": [ + 61, + 430, + 97, + 438 + ], + "lines": [ + { + "bbox": [ + 61, + 430, + 97, + 438 + ], + "spans": [ + { + "bbox": [ + 61, + 430, + 97, + 438 + ], + "type": "text", + "content": "Three Views" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 439, + 105, + 444 + ], + "lines": [ + { + "bbox": [ + 75, + 439, + 105, + 444 + ], + "spans": [ + { + "bbox": [ + 75, + 439, + 105, + 444 + ], + "type": "text", + "content": "1JQREIO" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 445, + 105, + 449 + ], + "lines": [ + { + "bbox": [ + 75, + 445, + 105, + 449 + ], + "spans": [ + { + "bbox": [ + 75, + 445, + 105, + 449 + ], + "type": "text", + "content": "putaired R" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 450, + 115, + 502 + ], + "lines": [ + { + "bbox": [ + 54, + 450, + 115, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 115, + 502 + ], + "type": "image", + "image_path": "814d18dc060c9849f68422ce09e319896cee2d5748752b1ee7fcd13e78a9f597.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 504, + 108, + 509 + ], + "lines": [ + { + "bbox": [ + 71, + 504, + 108, + 509 + ], + "spans": [ + { + "bbox": [ + 71, + 504, + 108, + 509 + ], + "type": "text", + "content": "a painting of" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 510, + 108, + 514 + ], + "lines": [ + { + "bbox": [ + 71, + 510, + 108, + 514 + ], + "spans": [ + { + "bbox": [ + 71, + 510, + 108, + 514 + ], + "type": "text", + "content": "a teddy bear" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 439, + 155, + 450 + ], + "lines": [ + { + "bbox": [ + 116, + 439, + 155, + 450 + ], + "spans": [ + { + "bbox": [ + 116, + 439, + 155, + 450 + ], + "type": "text", + "content": "STATE JOBUSTED TO" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 116, + 451, + 166, + 502 + ], + "blocks": [ + { + "bbox": [ + 116, + 451, + 166, + 502 + ], + "lines": [ + { + "bbox": [ + 116, + 451, + 166, + 502 + ], + "spans": [ + { + "bbox": [ + 116, + 451, + 166, + 502 + ], + "type": "image", + "image_path": "3f406ae660298293948f2795a3d68208f454f82f1111a69578881394acbdb628.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 504, + 156, + 509 + ], + "lines": [ + { + "bbox": [ + 126, + 504, + 156, + 509 + ], + "spans": [ + { + "bbox": [ + 126, + 504, + 156, + 509 + ], + "type": "text", + "content": "a painting" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 510, + 156, + 514 + ], + "lines": [ + { + "bbox": [ + 126, + 510, + 156, + 514 + ], + "spans": [ + { + "bbox": [ + 126, + 510, + 156, + 514 + ], + "type": "text", + "content": "of einstein" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 192, + 441, + 264, + 502 + ], + "blocks": [ + { + "bbox": [ + 198, + 431, + 231, + 438 + ], + "lines": [ + { + "bbox": [ + 198, + 431, + 231, + 438 + ], + "spans": [ + { + "bbox": [ + 198, + 431, + 231, + 438 + ], + "type": "text", + "content": "Four Views" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 441, + 264, + 502 + ], + "lines": [ + { + "bbox": [ + 192, + 441, + 264, + 502 + ], + "spans": [ + { + "bbox": [ + 192, + 441, + 264, + 502 + ], + "type": "image", + "image_path": "7f3612421973e08ad23070c591a14346668c9ff18f251c09b96f783e5d2950d6.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 502, + 247, + 506 + ], + "lines": [ + { + "bbox": [ + 205, + 502, + 247, + 506 + ], + "spans": [ + { + "bbox": [ + 205, + 502, + 247, + 506 + ], + "type": "text", + "content": "an oil painting" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 507, + 243, + 512 + ], + "lines": [ + { + "bbox": [ + 211, + 507, + 243, + 512 + ], + "spans": [ + { + "bbox": [ + 211, + 507, + 243, + 512 + ], + "type": "text", + "content": "of a rabbit" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 276, + 226, + 326, + 280 + ], + "blocks": [ + { + "bbox": [ + 286, + 216, + 374, + 224 + ], + "lines": [ + { + "bbox": [ + 286, + 216, + 374, + 224 + ], + "spans": [ + { + "bbox": [ + 286, + 216, + 374, + 224 + ], + "type": "text", + "content": "\"Polymorphic\" Jigsaw Puzzles" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 276, + 226, + 326, + 280 + ], + "lines": [ + { + "bbox": [ + 276, + 226, + 326, + 280 + ], + "spans": [ + { + "bbox": [ + 276, + 226, + 326, + 280 + ], + "type": "image", + "image_path": "ab5446bbb66a3eaaea6b9ef7dcf3e936070ec8d9f39c7784130cdaaec7e924e8.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 281, + 324, + 293 + ], + "lines": [ + { + "bbox": [ + 278, + 281, + 324, + 293 + ], + "spans": [ + { + "bbox": [ + 278, + 281, + 324, + 293 + ], + "type": "text", + "content": "an oil painting of a fruit bowl" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 327, + 226, + 408, + 281 + ], + "blocks": [ + { + "bbox": [ + 327, + 226, + 408, + 281 + ], + "lines": [ + { + "bbox": [ + 327, + 226, + 408, + 281 + ], + "spans": [ + { + "bbox": [ + 327, + 226, + 408, + 281 + ], + "type": "image", + "image_path": "b05af9a20c9095c3272e31b24ce458c6a3851db8537d2aeb118275dade15c495.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 281, + 413, + 292 + ], + "lines": [ + { + "bbox": [ + 340, + 281, + 413, + 292 + ], + "spans": [ + { + "bbox": [ + 340, + 281, + 413, + 292 + ], + "type": "text", + "content": "cf. \"Reversible Head with Basket of Fruit.\" Giuseppe Arcimboldo, 1950." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 410, + 246, + 422, + 264 + ], + "blocks": [ + { + "bbox": [ + 410, + 246, + 422, + 264 + ], + "lines": [ + { + "bbox": [ + 410, + 246, + 422, + 264 + ], + "spans": [ + { + "bbox": [ + 410, + 246, + 422, + 264 + ], + "type": "image", + "image_path": "3ab94f84e4472e9192962bfb56e301da8ebe55cbc3ca5f8eb7a311df638806e9.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 424, + 229, + 473, + 280 + ], + "blocks": [ + { + "bbox": [ + 424, + 229, + 473, + 280 + ], + "lines": [ + { + "bbox": [ + 424, + 229, + 473, + 280 + ], + "spans": [ + { + "bbox": [ + 424, + 229, + 473, + 280 + ], + "type": "image", + "image_path": "25664e8a4b8663c777e78e78dcd23b4b30df5dad23e07c1389db1bbc968e7ca1.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 281, + 471, + 293 + ], + "lines": [ + { + "bbox": [ + 425, + 281, + 471, + 293 + ], + "spans": [ + { + "bbox": [ + 425, + 281, + 471, + 293 + ], + "type": "text", + "content": "an oil painting of a monkey" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 277, + 295, + 327, + 345 + ], + "blocks": [ + { + "bbox": [ + 277, + 295, + 327, + 345 + ], + "lines": [ + { + "bbox": [ + 277, + 295, + 327, + 345 + ], + "spans": [ + { + "bbox": [ + 277, + 295, + 327, + 345 + ], + "type": "image", + "image_path": "ca4af49d55157db81c431f3d909ac5c65a386223850acb5b715c39d7ce239d5d.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 346, + 320, + 357 + ], + "lines": [ + { + "bbox": [ + 284, + 346, + 320, + 357 + ], + "spans": [ + { + "bbox": [ + 284, + 346, + 320, + 357 + ], + "type": "text", + "content": "a watercolor of a kitten" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 328, + 295, + 422, + 353 + ], + "blocks": [ + { + "bbox": [ + 328, + 295, + 422, + 353 + ], + "lines": [ + { + "bbox": [ + 328, + 295, + 422, + 353 + ], + "spans": [ + { + "bbox": [ + 328, + 295, + 422, + 353 + ], + "type": "image", + "image_path": "0b324c3d93dea8478143771c28d2981dc493f74f3345ab4b3de840f2767b1622.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 424, + 295, + 474, + 345 + ], + "blocks": [ + { + "bbox": [ + 424, + 295, + 474, + 345 + ], + "lines": [ + { + "bbox": [ + 424, + 295, + 474, + 345 + ], + "spans": [ + { + "bbox": [ + 424, + 295, + 474, + 345 + ], + "type": "image", + "image_path": "f32c78681f796c55340ff92bfd5d13da89a91bbdf898b3c261b3c3af4bec3c70.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 346, + 467, + 358 + ], + "lines": [ + { + "bbox": [ + 430, + 346, + 467, + 358 + ], + "spans": [ + { + "bbox": [ + 430, + 346, + 467, + 358 + ], + "type": "text", + "content": "a watercolor of a puppy" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 277, + 362, + 327, + 412 + ], + "blocks": [ + { + "bbox": [ + 277, + 362, + 327, + 412 + ], + "lines": [ + { + "bbox": [ + 277, + 362, + 327, + 412 + ], + "spans": [ + { + "bbox": [ + 277, + 362, + 327, + 412 + ], + "type": "image", + "image_path": "48895abba167250c0c1120004e8662599da7d4bf6e112e22ef21eb121cc9423f.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 414, + 321, + 426 + ], + "lines": [ + { + "bbox": [ + 284, + 414, + 321, + 426 + ], + "spans": [ + { + "bbox": [ + 284, + 414, + 321, + 426 + ], + "type": "text", + "content": "a painting of houseplants" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 328, + 361, + 415, + 425 + ], + "blocks": [ + { + "bbox": [ + 328, + 361, + 415, + 425 + ], + "lines": [ + { + "bbox": [ + 328, + 361, + 415, + 425 + ], + "spans": [ + { + "bbox": [ + 328, + 361, + 415, + 425 + ], + "type": "image", + "image_path": "c80a59a8268180de20ecc68d6f660ab405f01ff88c7db1f679482956692380bb.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 411, + 379, + 422, + 397 + ], + "blocks": [ + { + "bbox": [ + 411, + 379, + 422, + 397 + ], + "lines": [ + { + "bbox": [ + 411, + 379, + 422, + 397 + ], + "spans": [ + { + "bbox": [ + 411, + 379, + 422, + 397 + ], + "type": "image", + "image_path": "05de3f3cdc2c3da0ee416b080e0ffb2df3902cebfcdcb509fbf0a59cc01b1152.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 424, + 363, + 473, + 413 + ], + "blocks": [ + { + "bbox": [ + 424, + 363, + 473, + 413 + ], + "lines": [ + { + "bbox": [ + 424, + 363, + 473, + 413 + ], + "spans": [ + { + "bbox": [ + 424, + 363, + 473, + 413 + ], + "type": "image", + "image_path": "3898c7aea2239ea57a07b4c83808a89e1e373bbff8e6564004a3779379ee12d9.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 414, + 470, + 419 + ], + "lines": [ + { + "bbox": [ + 426, + 414, + 470, + 419 + ], + "spans": [ + { + "bbox": [ + 426, + 414, + 470, + 419 + ], + "type": "text", + "content": "a painting of" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 419, + 470, + 423 + ], + "lines": [ + { + "bbox": [ + 426, + 419, + 470, + 423 + ], + "spans": [ + { + "bbox": [ + 426, + 419, + 470, + 423 + ], + "type": "text", + "content": "marilyn monroe" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 423, + 470, + 427 + ], + "lines": [ + { + "bbox": [ + 426, + 423, + 470, + 427 + ], + "spans": [ + { + "bbox": [ + 426, + 423, + 470, + 427 + ], + "type": "text", + "content": "m111yHnH" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 489, + 419, + 533, + 424 + ], + "lines": [ + { + "bbox": [ + 489, + 419, + 533, + 424 + ], + "spans": [ + { + "bbox": [ + 489, + 419, + 533, + 424 + ], + "type": "text", + "content": "Aude Oliva et al. 2006." + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 489, + 241, + 539, + 292 + ], + "blocks": [ + { + "bbox": [ + 490, + 217, + 536, + 224 + ], + "lines": [ + { + "bbox": [ + 490, + 217, + 536, + 224 + ], + "spans": [ + { + "bbox": [ + 490, + 217, + 536, + 224 + ], + "type": "text", + "content": "Inner Rotations" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 228, + 531, + 239 + ], + "lines": [ + { + "bbox": [ + 493, + 228, + 531, + 239 + ], + "spans": [ + { + "bbox": [ + 493, + 228, + 531, + 239 + ], + "type": "text", + "content": "a pop art of albert einstel" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 489, + 241, + 539, + 292 + ], + "lines": [ + { + "bbox": [ + 489, + 241, + 539, + 292 + ], + "spans": [ + { + "bbox": [ + 489, + 241, + 539, + 292 + ], + "type": "image", + "image_path": "934841861477715a388eb484315e5741c424e207ae77287fe76462967ada9877.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + } + ], + "index": 78 + }, + { + "type": "image", + "bbox": [ + 488, + 294, + 539, + 346 + ], + "blocks": [ + { + "bbox": [ + 488, + 294, + 539, + 346 + ], + "lines": [ + { + "bbox": [ + 488, + 294, + 539, + 346 + ], + "spans": [ + { + "bbox": [ + 488, + 294, + 539, + 346 + ], + "type": "image", + "image_path": "a22f77bf05a1a396c3124f3191bd1eb9223e3a0de6c5e0997d022d1113a28c7b.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 488, + 349, + 539, + 399 + ], + "blocks": [ + { + "bbox": [ + 488, + 349, + 539, + 399 + ], + "lines": [ + { + "bbox": [ + 488, + 349, + 539, + 399 + ], + "spans": [ + { + "bbox": [ + 488, + 349, + 539, + 399 + ], + "type": "image", + "image_path": "13416d1ab527a0400c0154000a9b2ee990a27a5d052afacc3b890b0fe6256518.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 401, + 533, + 406 + ], + "lines": [ + { + "bbox": [ + 493, + 401, + 533, + 406 + ], + "spans": [ + { + "bbox": [ + 493, + 401, + 533, + 406 + ], + "type": "text", + "content": "a pop art of" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 407, + 533, + 411 + ], + "lines": [ + { + "bbox": [ + 493, + 407, + 533, + 411 + ], + "spans": [ + { + "bbox": [ + 493, + 407, + 533, + 411 + ], + "type": "text", + "content": "marilyn monroe" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 494, + 414, + 531, + 418 + ], + "lines": [ + { + "bbox": [ + 494, + 414, + 531, + 418 + ], + "spans": [ + { + "bbox": [ + 494, + 414, + 531, + 418 + ], + "type": "text", + "content": "cf. \"Hybrid Images.\"" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 59, + 529, + 111, + 578 + ], + "blocks": [ + { + "bbox": [ + 61, + 518, + 93, + 526 + ], + "lines": [ + { + "bbox": [ + 61, + 518, + 93, + 526 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 93, + 526 + ], + "type": "text", + "content": "Negatives" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 529, + 111, + 578 + ], + "lines": [ + { + "bbox": [ + 59, + 529, + 111, + 578 + ], + "spans": [ + { + "bbox": [ + 59, + 529, + 111, + 578 + ], + "type": "image", + "image_path": "074baec6063ef9c797d9147e38bcaeca649852d95ae0e4b90869204234cf5548.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 580, + 104, + 590 + ], + "lines": [ + { + "bbox": [ + 61, + 580, + 104, + 590 + ], + "spans": [ + { + "bbox": [ + 61, + 580, + 104, + 590 + ], + "type": "text", + "content": "a lithograph of a teddy bear" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 595, + 547, + 641 + ], + "lines": [ + { + "bbox": [ + 46, + 595, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 547, + 641 + ], + "type": "text", + "content": "Figure 1. Generating Multi-View Illusions. We propose a method for generating optical illusions from an off-the-shelf text-to-image diffusion model. We create images that match different prompts after undergoing a transformation. Our approach supports a variety of transformations, including flips, rotations, skews, color inversions, and jigsaw rearrangements. All images are hand selected. For random samples, please see Fig. 8 and Appendix D. For easier viewing, please see our webpage for animated versions of these illusions." + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_caption" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 116, + 529, + 159, + 578 + ], + "blocks": [ + { + "bbox": [ + 116, + 529, + 159, + 578 + ], + "lines": [ + { + "bbox": [ + 116, + 529, + 159, + 578 + ], + "spans": [ + { + "bbox": [ + 116, + 529, + 159, + 578 + ], + "type": "image", + "image_path": "af6f890fab785cbf670fb63a7dc9acffab7aa9a82c641981e68dfd1737235ba7.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 580, + 151, + 590 + ], + "lines": [ + { + "bbox": [ + 117, + 580, + 151, + 590 + ], + "spans": [ + { + "bbox": [ + 117, + 580, + 151, + 590 + ], + "type": "text", + "content": "a lithograph of a rabbit" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_caption" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 174, + 529, + 224, + 578 + ], + "blocks": [ + { + "bbox": [ + 174, + 529, + 224, + 578 + ], + "lines": [ + { + "bbox": [ + 174, + 529, + 224, + 578 + ], + "spans": [ + { + "bbox": [ + 174, + 529, + 224, + 578 + ], + "type": "image", + "image_path": "ed2a4e30afafb61febb75a05648eeffa19ae870cea53073c3c323903d85d0b58.jpg" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 580, + 219, + 590 + ], + "lines": [ + { + "bbox": [ + 179, + 580, + 219, + 590 + ], + "spans": [ + { + "bbox": [ + 179, + 580, + 219, + 590 + ], + "type": "text", + "content": "a lithograph of a landscape" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_caption" + } + ], + "index": 90 + }, + { + "type": "image", + "bbox": [ + 225, + 529, + 274, + 578 + ], + "blocks": [ + { + "bbox": [ + 225, + 529, + 274, + 578 + ], + "lines": [ + { + "bbox": [ + 225, + 529, + 274, + 578 + ], + "spans": [ + { + "bbox": [ + 225, + 529, + 274, + 578 + ], + "type": "image", + "image_path": "6454857f19ee4e6c2c329140bad8d1468cfcbf9bcc4e9feac5a93356071201f0.jpg" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 580, + 271, + 590 + ], + "lines": [ + { + "bbox": [ + 228, + 580, + 271, + 590 + ], + "spans": [ + { + "bbox": [ + 228, + 580, + 271, + 590 + ], + "type": "text", + "content": "a lithograph of houseplants" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_caption" + } + ], + "index": 92 + }, + { + "type": "image", + "bbox": [ + 284, + 444, + 335, + 495 + ], + "blocks": [ + { + "bbox": [ + 286, + 431, + 306, + 438 + ], + "lines": [ + { + "bbox": [ + 286, + 431, + 306, + 438 + ], + "spans": [ + { + "bbox": [ + 286, + 431, + 306, + 438 + ], + "type": "text", + "content": "Skews" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 284, + 444, + 335, + 495 + ], + "lines": [ + { + "bbox": [ + 284, + 444, + 335, + 495 + ], + "spans": [ + { + "bbox": [ + 284, + 444, + 335, + 495 + ], + "type": "image", + "image_path": "6f77fb1c778a0974cd974b2b8bdbf4be85b0d05f146794ac3f8e2aa2ebef26ac.jpg" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 495, + 334, + 506 + ], + "lines": [ + { + "bbox": [ + 285, + 495, + 334, + 506 + ], + "spans": [ + { + "bbox": [ + 285, + 495, + 334, + 506 + ], + "type": "text", + "content": "an oil painting ofa tudor portrait" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 509, + 380, + 514 + ], + "lines": [ + { + "bbox": [ + 302, + 509, + 380, + 514 + ], + "spans": [ + { + "bbox": [ + 302, + 509, + 380, + 514 + ], + "type": "text", + "content": "cf. \"The Ambassadors.\" Hans Holbein, 1533." + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_caption" + } + ], + "index": 95 + }, + { + "type": "image", + "bbox": [ + 335, + 444, + 400, + 495 + ], + "blocks": [ + { + "bbox": [ + 335, + 444, + 400, + 495 + ], + "lines": [ + { + "bbox": [ + 335, + 444, + 400, + 495 + ], + "spans": [ + { + "bbox": [ + 335, + 444, + 400, + 495 + ], + "type": "image", + "image_path": "160dee21df8225e42a92d803bf76065cf631ea6c5f6a7e8f2e89fb80542cb4d7.jpg" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 495, + 394, + 500 + ], + "lines": [ + { + "bbox": [ + 353, + 495, + 394, + 500 + ], + "spans": [ + { + "bbox": [ + 353, + 495, + 394, + 500 + ], + "type": "text", + "content": "an oil painting" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 501, + 389, + 506 + ], + "lines": [ + { + "bbox": [ + 353, + 501, + 389, + 506 + ], + "spans": [ + { + "bbox": [ + 353, + 501, + 389, + 506 + ], + "type": "text", + "content": "of a skull" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_caption" + } + ], + "index": 97 + }, + { + "type": "image", + "bbox": [ + 417, + 444, + 469, + 495 + ], + "blocks": [ + { + "bbox": [ + 417, + 444, + 469, + 495 + ], + "lines": [ + { + "bbox": [ + 417, + 444, + 469, + 495 + ], + "spans": [ + { + "bbox": [ + 417, + 444, + 469, + 495 + ], + "type": "image", + "image_path": "3c3ad80605fecab77af38c4afb6172e05ab2efa04195e404793e67db4e4c9087.jpg" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 495, + 465, + 500 + ], + "lines": [ + { + "bbox": [ + 423, + 495, + 465, + 500 + ], + "spans": [ + { + "bbox": [ + 423, + 495, + 465, + 500 + ], + "type": "text", + "content": "an oil painting" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 501, + 460, + 506 + ], + "lines": [ + { + "bbox": [ + 426, + 501, + 460, + 506 + ], + "spans": [ + { + "bbox": [ + 426, + 501, + 460, + 506 + ], + "type": "text", + "content": "of a soldier" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 509, + 484, + 514 + ], + "lines": [ + { + "bbox": [ + 469, + 509, + 484, + 514 + ], + "spans": [ + { + "bbox": [ + 469, + 509, + 484, + 514 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_caption" + } + ], + "index": 101 + }, + { + "type": "image", + "bbox": [ + 485, + 444, + 535, + 495 + ], + "blocks": [ + { + "bbox": [ + 485, + 444, + 535, + 495 + ], + "lines": [ + { + "bbox": [ + 485, + 444, + 535, + 495 + ], + "spans": [ + { + "bbox": [ + 485, + 444, + 535, + 495 + ], + "type": "image", + "image_path": "3e7e6f14479d7436c07de49868eb1b48a9980f6af3142f8d71818a2c7f927a85.jpg" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 495, + 531, + 500 + ], + "lines": [ + { + "bbox": [ + 488, + 495, + 531, + 500 + ], + "spans": [ + { + "bbox": [ + 488, + 495, + 531, + 500 + ], + "type": "text", + "content": "an oil painting" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 501, + 531, + 506 + ], + "lines": [ + { + "bbox": [ + 488, + 501, + 531, + 506 + ], + "spans": [ + { + "bbox": [ + 488, + 501, + 531, + 506 + ], + "type": "text", + "content": "of a houseplant" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_caption" + } + ], + "index": 105 + }, + { + "type": "image", + "bbox": [ + 414, + 529, + 465, + 578 + ], + "blocks": [ + { + "bbox": [ + 415, + 519, + 449, + 526 + ], + "lines": [ + { + "bbox": [ + 415, + 519, + 449, + 526 + ], + "spans": [ + { + "bbox": [ + 415, + 519, + 449, + 526 + ], + "type": "text", + "content": "Ambigrams" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 529, + 465, + 578 + ], + "lines": [ + { + "bbox": [ + 414, + 529, + 465, + 578 + ], + "spans": [ + { + "bbox": [ + 414, + 529, + 465, + 578 + ], + "type": "image", + "image_path": "940f09fb808c4d5569cf14b6e31fc4c27c4f4f5bb2d35f1b687bc3a86b0e7598.jpg" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 579, + 460, + 584 + ], + "lines": [ + { + "bbox": [ + 416, + 579, + 460, + 584 + ], + "spans": [ + { + "bbox": [ + 416, + 579, + 460, + 584 + ], + "type": "text", + "content": "the word \"happy\"" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 585, + 460, + 590 + ], + "lines": [ + { + "bbox": [ + 417, + 585, + 460, + 590 + ], + "spans": [ + { + "bbox": [ + 417, + 585, + 460, + 590 + ], + "type": "text", + "content": "cursive writing" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 467, + 541, + 479, + 548 + ], + "lines": [ + { + "bbox": [ + 467, + 541, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 467, + 541, + 479, + 548 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "image_caption" + } + ], + "index": 109 + }, + { + "type": "image", + "bbox": [ + 481, + 529, + 531, + 578 + ], + "blocks": [ + { + "bbox": [ + 481, + 529, + 531, + 578 + ], + "lines": [ + { + "bbox": [ + 481, + 529, + 531, + 578 + ], + "spans": [ + { + "bbox": [ + 481, + 529, + 531, + 578 + ], + "type": "image", + "image_path": "05608620aa038aba6a1fdf954101c9b21f4dca55f4cd718625091c90c4fb977f.jpg" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 579, + 531, + 584 + ], + "lines": [ + { + "bbox": [ + 479, + 579, + 531, + 584 + ], + "spans": [ + { + "bbox": [ + 479, + 579, + 531, + 584 + ], + "type": "text", + "content": "the word \"holiday\"." + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 585, + 526, + 590 + ], + "lines": [ + { + "bbox": [ + 484, + 585, + 526, + 590 + ], + "spans": [ + { + "bbox": [ + 484, + 585, + 526, + 590 + ], + "type": "text", + "content": "cursive writing" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_caption" + } + ], + "index": 113 + }, + { + "bbox": [ + 143, + 648, + 191, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 648, + 191, + 660 + ], + "spans": [ + { + "bbox": [ + 143, + 648, + 191, + 660 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 117 + }, + { + "bbox": [ + 45, + 678, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 678, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 678, + 287, + 715 + ], + "type": "text", + "content": "We address the problem of synthesizing multi-view optical illusions: images that change appearance upon a transformation, such as a flip or rotation. We propose a simple," + } + ] + } + ], + "index": 118 + }, + { + "bbox": [ + 304, + 649, + 546, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 546, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 546, + 709 + ], + "type": "text", + "content": "zero-shot method for obtaining these illusions from off-the-shelf text-to-image diffusion models. During the reverse diffusion process, we estimate the noise from different views of a noisy image, and then combine these noise estimates together and denoise the image. A theoretical analysis sug-" + } + ] + } + ], + "index": 119 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24154" + } + ] + } + ], + "index": 120 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "gests that this method works precisely for views that can be written as orthogonal transformations, of which permutations are a subset. This leads to the idea of a visual anagram—an image that changes appearance under some rearrangement of pixels. This includes rotations and flips, but also more exotic pixel permutations such as a jigsaw rearrangement. Our approach also naturally extends to illusions with more than two views. We provide both qualitative and quantitative results demonstrating the effectiveness and flexibility of our method. Please see our project webpage for additional visualizations and results: https://dangeng.github.io/visual_anagrams/" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 227, + 128, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 128, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 128, + 239 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 342 + ], + "type": "text", + "content": "Images that change their appearance under a transformation, such as a rotation or a flip, have long fascinated students of perception, from Salvador Dalí to M. C. Escher. The appeal of these multi-view optical illusions lies partly in the challenge of arranging visual elements such that they may be understood in multiple different ways. Creating these illusions requires accurately modeling—and then subverting—visual perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 343, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 474 + ], + "type": "text", + "content": "In this paper, we propose a simple, zero-shot method for creating multi-view illusions with off-the-shelf text-to-image diffusion models. In contrast to most previous work on computationally generating optical illusions [3-5, 10, 12, 15, 18, 20, 28, 31, 32, 38], our method does not require an explicit model of human perception. Rather, our approach builds on work that suggests generative models may process optical illusions in a way similar to humans [14, 23, 29]. In this way, our method is similar to recent work that uses diffusion models to create optical illusions by Burgert et al. [2] and Tancik [42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 475, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 617 + ], + "type": "text", + "content": "Our method can generate many types of classic illusions, such as images that change appearance when flipped or rotated (Fig. 1), as well as a new class of illusions which we term visual anagrams. These are images that change appearance under a permutation of their pixels. Image flips and rotations are a subset of these, as they can both be expressed as a permutation of pixels, but we also consider more exotic permutations. For example, we generate jigsaw puzzles that can be solved in two different ways, which we call \"polymorphic jigsaws.\" In addition, we successfully apply our approach to generating illusions with three and four views (Fig. 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Our method works by using a diffusion model to denoise an image from multiple views, obtaining multiple noise estimates. These noise estimates are then combined to form a single noise estimate which is used to perform a step in the reverse diffusion process. However, we show that care must be taken in choosing these views. For one, the transformation must keep the statistics of the noise intact, as the diffusion model is trained under the assumption of i.i.d. Gaus" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "sian noise. We provide an analysis of these conditions and give an exact specification of the class of transformations supported. Our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 110, + 545, + 230 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 110, + 545, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 545, + 134 + ], + "type": "text", + "content": "- We present a simple yet effective method for generating multi-view optical illusions using diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 135, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 135, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 135, + 545, + 169 + ], + "type": "text", + "content": "- We derive a precise description of the set of views that our method supports and provide empirical evidence that these views work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 171, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 545, + 205 + ], + "type": "text", + "content": "- We consider practical design decisions, crucial to optimizing the quality of generated illusions, and report ablations on our choices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 207, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 545, + 230 + ], + "type": "text", + "content": "- We provide quantitative and qualitative results, showcasing both the efficacy and flexibility of our method." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 242, + 392, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 392, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 392, + 255 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 262, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 478 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models [6, 17, 22, 35-37, 39-41] are a class of powerful generative models that iteratively convert a sample from a noise distribution to a sample from some data distribution. These models work by estimating the noise in a noisy sample, and removing the estimated noise following some update rule such as DDPM [22] or DDIM [40]. A prominent application of diffusion models has been text-conditioned image synthesis [24, 30, 36, 37]. In addition to a noisy image and a timestep, these models take a language model embedding of a text prompt as conditioning. Our approach is closely related to recent works that experiment with composing energy-based models and diffusion models [7-9, 13, 26, 27]. These approaches [9, 27] have shown that noise estimates from multiple conditional distributions can be combined together to obtain samples from compositions of the learned distributions. Our method uses a similar approach, and we apply it to the problem of multi-view illusion generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Computational Optical Illusions. Optical illusions serve as a testbed for understanding both human and machine perception [14, 19, 23, 29, 45]. We focus on generating illusions computationally, an area which has primarily relied on models of how our brains process external stimuli. Freeman et al. [12] create the illusion of constant motion in a desired direction by locally applying a filter with continuously shifting phase, relying on the observation that local phaseshifts are interpreted as global movement. Oliva et al. [31] propose a method to make \"hybrid images,\" which change appearance depending on the distance they are viewed from. This method takes advantage of the multiscale nature of human perception by blending high frequencies of one image with low frequencies from another. Chu et al. [5] camouflage objects in a scene through re-texturing, with additional constraints on luminance as to preserve salient features of the object, and other work camouflages objects from multiple viewpoints in 3D scenes [18, 32]. Recently, Chandra et al. [3] design color-constancy, size constancy, and face" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24155" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "perception illusions by differentiating through a Bayesian model of human vision. Our method likewise generates illusions, but does not depend on an explicit model of human perception. Instead, our method works by leveraging visual priors in diffusion models learned implicitly through data. This aligns with observations [14, 23, 29] that generative models process illusions similarly to humans, and predict the same ambiguities. From this perspective, we can view our method as leveraging generative, rather than discriminative, models to synthesize adversarial examples [16] against humans [11]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 211, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 289, + 559 + ], + "type": "text", + "content": "Illusions with Diffusion Models. Very recently, artists and researchers have taken steps that show the potential of using diffusion models to create illusions. An artist under the pseudonym MrUgleh [43] repurposed a model finetuned for generating QR codes [25, 47] to create images whose global structure subtly matches a given template image. In contrast, we study multi-view illusions that can be created zero-shot from off-the-shelf diffusion models, and our illusions are specified via text rather than images. Burgert et al. [2] use score distillation sampling (SDS) [33, 44] to create images that align with different prompts from different views. While in principle this approach supports a superset of our views, the use of SDS results in significantly lower quality results, and the need for explicit optimization leads to long sampling times. Our method is most similar to a proof-of-concept by Tancik [42], which creates rotation illusions by sampling from a latent diffusion model [36] while alternating noise estimates between different views and prompts. While our technical approach is similar, by contrast we systematically study multi-view illusions, both by experimentally evaluating many different types of illusions and by providing a theoretical analysis of which views are (and are not) supported. In doing so, we go beyond just rotation views. We also make a number of improvements that result in qualitatively and quantitatively better illusions, such as by identifying a source of artifacts from latent diffusion, and by adding support for an arbitrary number of views. To our knowledge, we are the first to systematically evaluate illusions generated by these approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 570, + 104, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 104, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 104, + 582 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "Our goal is to produce multi-view optical illusions using a pretrained diffusion model. That is, we seek to synthesize images that change appearance or identity when transformed, such as when flipped or rotated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 647, + 232, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 232, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 232, + 658 + ], + "type": "text", + "content": "3.1. Text-conditioned Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Diffusion models [22, 39, 41] take i.i.d. Gaussian noise, " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", and iteratively denoise it to produce a sample, " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", from some data distribution. These models are parameterized by a neural network which estimates the noise in" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 170 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 170 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 170 + ], + "type": "image", + "image_path": "489bdfa0d17e3a87fa56c8242c836b46eaab7b811772e79f5d1ceeb26146a288.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "text", + "content": "Figure 2. Algorithm Overview. Our method works by simultaneously denoising multiple views of an image. Given a noisy image " + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "text", + "content": ", we compute noise estimates, " + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "inline_equation", + "content": "\\epsilon_t^i" + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "text", + "content": ", conditioned on different prompts, after applying views " + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "text", + "content": ". We then apply the inverse view " + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "inline_equation", + "content": "v_i^{-1}" + }, + { + "bbox": [ + 304, + 172, + 547, + 239 + ], + "type": "text", + "content": " to align estimates, average the estimates, and perform a reverse diffusion step. The final output is an optical illusion." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": "some intermediate, partially denoised data point " + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,y,t)" + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": " is some conditioning such as text prompts and " + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": " is the timestep in the diffusion process. The estimated noise is then used in an update rule [22, 40], from which " + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": " is computed from " + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 254, + 545, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 314, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 385 + ], + "type": "text", + "content": "To condition the diffusion model on another input, such as a text prompt, a common approach is to use classifier-free guidance [21]. With this method, unconditional noise estimates (usually obtained by passing the null text prompt as conditioning) and conditional noise estimates are combined together:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 393, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 393, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 393, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} ^ {\\mathrm {C F G}} = \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\varnothing) + \\gamma \\left(\\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, y\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\varnothing\\right)\\right). \\tag {1}", + "image_path": "e98065e1ff081c70f8c9f440740f4d9922b146ccc4a1919015768933a0f1a99d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "text", + "content": " denotes the embedding of the empty string and " + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "text", + "content": " is a parameter that controls the strength of the guidance. Classifier-free guidance acts to sharpen the distribution of generated images to produce higher quality results. It also enables negative prompting [1], in which the empty text prompt embedding, " + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 304, + 415, + 547, + 500 + ], + "type": "text", + "content": ", is replaced by a text prompt that we would like to discourage the model from generating." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 506, + 415, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 415, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 415, + 519 + ], + "type": "text", + "content": "3.2. Parallel Denoising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "content": "We produce multi-view illusions by using a diffusion model to simultaneously denoise multiple views of an image. Concretely, we take a set of " + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "content": " prompts, " + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "content": ", each associated with a view function " + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "inline_equation", + "content": "v_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "content": ", which applies a transformation to an image. These transformations may be, for example, the identity function, an image flip, or a permutation of pixels. Then given a diffusion model, " + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "content": ", and a partially denoised image, " + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 525, + 545, + 633 + ], + "type": "text", + "content": ", we combine noise estimates from different views into a single noise estimate by averaging:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 635, + 545, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 635, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 352, + 635, + 545, + 663 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {t} = \\frac {1}{N} \\sum_ {i} v _ {i} ^ {- 1} \\left(\\epsilon_ {\\theta} \\left(v _ {i} \\left(\\mathbf {x} _ {t}\\right), y _ {i}, t\\right)\\right). \\tag {2}", + "image_path": "cad81df50ab7b162d8e8a10881f6f098d0b3d506fd8943e66a83f877a0097aef.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "Effectively, we use each view " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " to transform the noisy image " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": ", estimate the noise in the transformed images, and then apply " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "v_{i}^{-1}" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " to the estimates in order to transform them back to the original view. Taking an average of these noise" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24156" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": "estimates gives us our combined noise estimate, which we can then use with our choice of diffusion sampler. We note that this technique of combining noise estimates is similar to previous work on compositionality [7-9, 13, 26, 27], where the idea is studied in further detail. In order to incorporate classifier-free guidance we simply replace the estimates " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(v_i(\\mathbf{x}_t),y_i,t)" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " with their classifier-free estimates, " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{\\mathrm{CFG}}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 164, + 167, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 167, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 167, + 176 + ], + "type": "text", + "content": "3.3. Conditions on Views" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 287, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 244 + ], + "type": "text", + "content": "One straightforward condition for the views is that they must be invertible. But diffusion models also implicitly impose other conditions on the views " + }, + { + "bbox": [ + 46, + 183, + 287, + 244 + ], + "type": "inline_equation", + "content": "v_{i}(\\cdot)" + }, + { + "bbox": [ + 46, + 183, + 287, + 244 + ], + "type": "text", + "content": ". We describe two such conditions below. We find that if these conditions are not satisfied, the denoising process produces poor results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "content": "Linearity. The diffusion model, " + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "content": ", acts on noisy images, " + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "content": ". That is, specifically images of the form:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 277, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 277, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 103, + 277, + 287, + 306 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = w _ {t} ^ {\\text {s i g n a l}} \\underbrace {\\mathbf {x} _ {0}} _ {\\text {s i g n a l}} + w _ {t} ^ {\\text {n o i s e}} \\underbrace {\\epsilon} _ {\\text {n o i s e}}. \\tag {3}", + "image_path": "349be2b4ec7dc65cebf313102578a92797a2bab71862ada975e94ebb3927db70.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": "The exact values of " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "w_{t}^{\\mathrm{signal}}" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "w_{t}^{\\mathrm{noise}}" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " depend on model implementation details such as the variance schedule, but are unimportant for our work, so we omit them for clarity. What is important is that " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " is a linear combination of pure signal, " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": ", and pure noise, " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": ", for some specific " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "w_{t}^{\\mathrm{signal}}" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "w_{t}^{\\mathrm{noise}}" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": ". Therefore our view " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " must take a noisy image " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " and transform it into a new noisy image " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "v_{i}(\\mathbf{x}_{t})" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " that is also a linear combination of pure signal and pure noise with the same weighting. This can be achieved by requiring " + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 311, + 287, + 431 + ], + "type": "text", + "content": " to be a linear transformation, of the form" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 435, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 435, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 134, + 435, + 287, + 449 + ], + "type": "interline_equation", + "content": "v _ {i} \\left(\\mathbf {x} _ {t}\\right) = \\mathbf {A} _ {i} \\mathbf {x} _ {t}, \\tag {4}", + "image_path": "a7e8540ac8cf0346a82456ddc5b5277f3a4c078fbe6092da91964f7c7f026070.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "text", + "content": "for some matrix " + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "text", + "content": ", and some flattened noisy image " + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "text", + "content": ". By linearity, we are effectively applying the view " + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 46, + 451, + 287, + 487 + ], + "type": "text", + "content": " to the signal and the noise separately:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 490, + 287, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 490, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 84, + 490, + 287, + 535 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} v _ {i} \\left(\\mathbf {x} _ {t}\\right) = \\mathbf {A} _ {i} \\left(w _ {t} ^ {\\text {s i g n a l}} \\mathbf {x} _ {0} + w _ {t} ^ {\\text {n o i s e}} \\epsilon\\right) (5) \\\\ = w _ {t} ^ {\\text {s i g n a l}} \\underbrace {\\mathbf {A} _ {i} \\mathbf {x} _ {0}} _ {\\text {n e w s i g n a l}} + w _ {t} ^ {\\text {n o i s e}} \\underbrace {\\mathbf {A} _ {i} \\epsilon} _ {\\text {n e w n o i s e}}. (6) \\\\ \\end{array}", + "image_path": "37eac05e1271301d07d9066d489f2680d5f74c6e92d2bc98c8b7ffc931ec41a6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "type": "text", + "content": "This results in a linear combination of transformed signal, " + }, + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\mathbf{x}_0" + }, + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "type": "text", + "content": ", and transformed noise, " + }, + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\epsilon" + }, + { + "bbox": [ + 47, + 537, + 287, + 586 + ], + "type": "text", + "content": ", weighted with the correct scaling factors. For further discussion, please see Appendix H." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Statistical Consistency. In addition to expecting a linear combination of signal and noise at a specific weighting, the diffusion model also expects the noise to have a precise distribution. In particular, most diffusion networks are trained with " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". Therefore, we must ensure that our transformed noise, " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\epsilon" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", is likewise drawn from " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". This is true if and only if " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " is an orthogonal matrix. We provide a proof in Appendix I, but intuitively this fact reflects the spherical symmetry of the standard Gaussian density. Orthogonal transformations, being generalizations of rota" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "tions and flips to higher dimensions, preserve this spherically symmetric density. Note that these are rotations in the pixel values as opposed to spatial rotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 119, + 411, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 119, + 411, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 119, + 411, + 131 + ], + "type": "text", + "content": "3.4. Views Considered" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 139, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 545, + 212 + ], + "type": "text", + "content": "The vast majority of orthogonal transformations applied to an image will not correspond to an intuitive image transformation. However, a number of these transformations do. Below, we enumerate the orthogonal transformations which we consider, all of which can be seen in the illusions in Fig. 1 unless otherwise specified." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 224, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 272 + ], + "type": "text", + "content": "Identity. The simplest transformation we consider is the identity transformation. Using this view allows us to optimize the untransformed image to align with a chosen prompt." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 285, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 416 + ], + "type": "text", + "content": "Standard Image Manipulations. We also consider spatial rotations of an image, which can be viewed as permutations of pixels. This works because permutations are in turn orthogonal. However, caution must be exercised when applying a rotation view, as common anti-aliasing operations such as bilinear sampling will modify the statistics of the noise. We discuss this further in Sec. 4.4. Spatial reflections are also permutations of pixels. As such we can use these views to generate illusions. Finally, we implement an approximation to skewing by rolling columns of pixels by different displacements." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": "General Permutations. We have already considered the special cases of spatial rotation, reflection, and skews but we can also consider other permutations. For example, we can divide an image into jigsaw pieces and rearrange these pieces to generate jigsaw puzzles with two solutions—what we call \"polymorphic\" jigsaw puzzles. Implementation details can be found in Appendix F." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 514, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 585 + ], + "type": "text", + "content": "We also consider the extreme case of sampling a completely random permutation of pixels and treating it as our view. Additionally, we can reduce the complexity of this by considering permutations of square patches, rather than pixels. Examples of these illusions can be found in Fig. 6 and are discussed in Sec. 4.3." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 587, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 545, + 647 + ], + "type": "text", + "content": "Finally, we consider rotating a circle within an image while leaving the rest of the image stationary, which we term inner rotations. Note that the permutations we consider are certainly not exhaustive, and many clever transformations exist which we do not study." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 660, + 545, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 660, + 545, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 660, + 545, + 720 + ], + "type": "text", + "content": "Color Inversion. Negation is an orthogonal transform; it is intuitively a 180 degree rotation generalized to higher dimensions. This allows us to generate illusions that change appearance upon color inversion, assuming pixel values are centered at 0 (e.g., in the range [-1, 1])." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24157" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 286, + 170 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 286, + 170 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 286, + 170 + ], + "type": "image", + "image_path": "227c327efbba6a7c244bb09a675c60ed147651e3088d6c1d4909d16a84a40076.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 173, + 287, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 287, + 227 + ], + "type": "text", + "content": "Figure 3. Latent-Based Artifacts. Manipulating the location of latent codes does not change the orientation of the blocks for which they encode. Therefore, when using latent diffusion models we see artifacts as shown above, in which straight lines are thatched under a rotation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 286, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 286, + 331 + ], + "type": "text", + "content": "Arbitrary Orthogonal Transformations. An arbitrary rotation of an image in pixel space is uninterpretable. Nevertheless, we demonstrate that our method works for these transformations as well. While these \"illusions\" are inscrutable to the human eye, they serve as confirmation that any orthogonal transformation works as a view with our method. These can be found in Fig. 7 and are discussed in Sec. 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 339, + 148, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 148, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 148, + 351 + ], + "type": "text", + "content": "3.5. Design Decisions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 381 + ], + "type": "text", + "content": "Beyond the core method, we also consider various design decisions aimed at maximizing illusion quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 387, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 494 + ], + "type": "text", + "content": "Pixel Diffusion Model. Previous work [42] performed multi-view denoising using Stable Diffusion [36], a latent diffusion model. However, the latent representation effectively encodes patches of pixels. This leads to artifacts under rotations or flips, where the location of latents change, but the content and orientation of these blocks do not. We show a qualitative example of this in Fig. 3, in which the model is forced to generate thatched lines to produce straight lines under a " + }, + { + "bbox": [ + 46, + 387, + 287, + 494 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 387, + 287, + 494 + ], + "type": "text", + "content": " rotation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 541 + ], + "type": "text", + "content": "To ameliorate this issue, we implement our method using a pixel-based diffusion model, DeepFloyd IF [24]. DeepFloyd denoises directly on pixels, effectively side-stepping the problem of orientation in latent code blocks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 594 + ], + "type": "text", + "content": "Combining Noise Estimates. In addition to taking a mean of noise estimates from different views, we also consider alternating through them by timestep, using the estimate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 594, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 594, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 86, + 594, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {t} = v _ {t \\bmod N} ^ {- 1} \\left(\\epsilon_ {\\theta} \\left(v _ {t \\bmod N} (\\mathbf {x} _ {t}), t, y)\\right)\\right). \\tag {7}", + "image_path": "08d1b8072f4d2f2342941801f469c4691e4d17692f55e6beda9912565839f21d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 612, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 635 + ], + "type": "text", + "content": "This is the reduction strategy used by [42], but we show in ablations in Sec. 4.2 that it performs worse than averaging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "content": "Negative Prompting. We experiment with negative prompting [1] in the 2-view case by using one view's prompt as a negative for the other view, and vice versa. This encourages the model to hide the other view's prompt for a given view. For a discussion, please see the ablations in Sec. 4.2." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 106, + 544, + 182 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative Results. We report the alignment score, " + }, + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "text", + "content": ", and the concealment score, " + }, + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "text", + "content": ", as well as quantiles of these scores. For a discussion, please see Sec. 4.1." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 106, + 544, + 182 + ], + "lines": [ + { + "bbox": [ + 308, + 106, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 544, + 182 + ], + "type": "table", + "html": "
Prompt PairMethodA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
CIFARBurgert et al. [2]0.2250.2530.2600.5010.5260.537
Tancik [42]0.2780.3100.3160.5950.6920.712
Ours0.2870.3210.3270.6240.7170.739
OursBurgert et al. [2]0.2330.2700.2830.5010.5260.538
Tancik [42]0.2560.2940.3090.5450.6210.655
Ours0.2750.3150.3260.5740.6680.694
", + "image_path": "0abf702e141ec0a5049dd73d7975e42144a74f37ff24f7c2b4c7b5a87dd832fc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 216, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 305, + 191, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 213 + ], + "type": "text", + "content": "Table 2. Ablations. We ablate negative prompting, reduction methods, and guidance scales on our dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 216, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 216, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 216, + 544, + 312 + ], + "type": "table", + "html": "
AblationA↑A0.9↑A0.95↑C↑C0.9↑C0.95↑
Negative Prompting0.240.270.2760.5760.6590.683
No Negative Prompting0.2550.2850.2950.5670.6430.679
Alternating Reduction0.2520.2860.2920.5600.6390.664
Mean Reduction0.2550.2850.2950.5670.6430.679
γ = 3.00.2390.2710.2850.5370.6100.629
γ = 7.00.2550.2850.2950.5670.6430.679
γ = 10.00.2590.2900.2970.5760.6640.702
", + "image_path": "58d89858e28ccbf4a978b78b13ff9e220f6dacab9edd16be99d7437aafc6ec02.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 329, + 359, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 359, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 359, + 341 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "type": "text", + "content": "We provide quantitative and qualitative results, and quantitative ablations. If not specified, qualitative results have been hand picked for quality. For random samples please see Fig. 8 and Appendix D. All implementation details can be found in Appendix A." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 418, + 424, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 424, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 424, + 430 + ], + "type": "text", + "content": "4.1. Quantitative Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 438, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 474 + ], + "type": "text", + "content": "Metrics. We use CLIP [34] to measure how well views align with the desired prompts. We consider two metrics derived from a score matrix " + }, + { + "bbox": [ + 304, + 438, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 304, + 438, + 545, + 474 + ], + "type": "text", + "content": ", defined as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 477, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 477, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 365, + 477, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i j} = \\phi_ {\\mathrm {i m g}} \\left(v _ {i} (\\mathbf {x})\\right) ^ {T} \\phi_ {\\text {t e x t}} \\left(p _ {j}\\right), \\tag {8}", + "image_path": "a355d1319ee6ad7925500ffa27fccdc31f10289c2f453f6d4387b976e44b106e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{text}}" + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "content": " are the CLIP visual and textual encoders respectively, returning a unit-norm vector embedding. " + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "content": " is our generated illusion, and " + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "content": " are our views with associated prompts " + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 495, + 545, + 553 + ], + "type": "text", + "content": ". A higher dot product indicates higher similarity between the image and text." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": "The first metric we consider is " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\min \\operatorname{diag}(\\mathbf{S})" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": ", which intuitively measures the worst alignment of all the views. We term this metric " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": ", the alignment score. However, this metric does not account for the possibility of seeing prompt " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": " in view " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": ". This is an occasional failure case of our method and to quantify this we propose a second derived metric which we term " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": ", the concealment score, computed as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 379, + 650, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 650, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 379, + 650, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\frac {1}{N} \\operatorname {t r} (\\operatorname {s o f t m a x} (S / \\tau)), \\tag {9}", + "image_path": "d9a0289ccaf8303cf850c72c632ae8c966808a2c861b4d356100f3fd61e2fb48.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the temperature parameter of CLIP. In computing this metric we average both directions of the softmax, so that this metric measures how well CLIP can classify a view" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24158" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 286, + 201 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 201 + ], + "type": "image", + "image_path": "6fc4757101a20b40ff7974893e548082e7e5c5f95a5eeea1d98597a3118d1b29.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 288, + 258 + ], + "type": "text", + "content": "Figure 4. Flip View CLIP Score Distribution. We visualize trade-offs between flipped and unflipped views by plotting the distribution of CLIP scores on the datasets. Note that the quality of the flipped image is as good as the unflipped image, with parity indicated by the dashed line." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 280, + 209, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 209, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 209, + 293 + ], + "type": "text", + "content": "as one of the " + }, + { + "bbox": [ + 47, + 280, + 209, + 293 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 280, + 209, + 293 + ], + "type": "text", + "content": " prompts and vice versa." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 299, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 394 + ], + "type": "text", + "content": "Dataset. To evaluate our method and baselines we compile two datasets of prompt pairs for 2-view illusions. One dataset uses the 10 classes from CIFAR-10 and contains a prompt per pair of classes, for a total of 45 prompt pairs. We refer to this as CIFAR. The other dataset we compile by hand, with the process documented in Appendix B. This dataset consists of 50 prompt pairs, which we refer to as Ours." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": "Baselines. We use two baselines that generate illusions using off-the-shelf diffusion models. One, which we denote \"Burgert et al. [2],\" uses Score Distillation Sampling. The other, which we denote \"Tancik [42],\" is an earlier version of our method, with differences discussed in more detail in Sec. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 481, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 636 + ], + "type": "text", + "content": "Results. We show results comparing our method to baselines on both datasets in Tab. 1 using vertical flips. We use vertical flips because it is a transformation supported by our method as well as the baselines. We use 10 samples per prompt, for a total of 450 and 500 samples for the CIFAR dataset and our dataset respectively. It is hard to perform a fair comparison with more samples because the Burgert et al. method uses SDS, which is quite slow1. Because we are particularly interested in the \"best-case\" performance, we also report quantiles of metrics, which we denote as " + }, + { + "bbox": [ + 46, + 481, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{0.9}" + }, + { + "bbox": [ + 46, + 481, + 287, + 636 + ], + "type": "text", + "content": " for the 90th percentile, for example. As can be seen, our method performs consistently better than the baselines, in both the alignment score and the concealment score." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 637, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 288, + 685 + ], + "type": "text", + "content": "In order to give a clearer understanding of trade-offs when optimizing two views, we show density plots which plot the CLIP scores of each of the two views of an illusion in Fig. 4. As can be seen, we do better than the baselines" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 544, + 348 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 544, + 348 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 544, + 348 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 544, + 348 + ], + "type": "image", + "image_path": "6b503c0fc497282bbeadba423851a5a5ac4e0f229f0bbd9af5784d8508090082.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 349, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 305, + 349, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 545, + 383 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparisons. We compare illusions generated by baselines to our illusions. We show examples from both our prompt dataset and the CIFAR prompt dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 394, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 431 + ], + "type": "text", + "content": "on average and in the best-case. Moreover, flipping during denoising does not hurt performance. The quality of the flipped images is as high as the unflipped images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 439, + 373, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 373, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 373, + 450 + ], + "type": "text", + "content": "4.2. Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 457, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 505 + ], + "type": "text", + "content": "We ablate out the noise estimate reduction strategy, negative prompting, and the guidance scale in Tab. 2. We use our dataset, with 10 samples for each prompt for a total of 500 illusions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 514, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 598 + ], + "type": "text", + "content": "Reduction Strategy. We find that mean reduction does better than alternating. Our hypothesis is that alternating the noise estimates results in \"thrashing,\" causing poor convergence. Moreover, we find that the alternating strategy gives poor results on illusions with more than 2 views, as each view has fewer denoising steps. Qualitative examples of this can be found in Appendix G." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "Negative Prompting. When using negative prompting, care must be taken to omit any overlap between the negative and positive prompt. For example, given the two prompts \"oil painting of a dog\" and \"oil painting of a cat\", using one prompt as the negative for the other would simultaneously encourage and discourage the style \"oil painting\". Rather, the negative prompts should be \"a cat\" and \"a dog\" respectively. We find that negative prompting can improve the concealment score, indi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "text", + "content": "Sampling just 10 images per prompt already takes more than a week of GPU-hours." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24159" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 81, + 102, + 129 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 80, + 79 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 80, + 79 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 80, + 79 + ], + "type": "text", + "content": "8×8" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 81, + 102, + 129 + ], + "lines": [ + { + "bbox": [ + 53, + 81, + 102, + 129 + ], + "spans": [ + { + "bbox": [ + 53, + 81, + 102, + 129 + ], + "type": "image", + "image_path": "abd14ef2b8aaffdeb18e4e2b2abe31c8bb3049e1a80febde2978136c9bc0378c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 129, + 94, + 140 + ], + "lines": [ + { + "bbox": [ + 53, + 129, + 94, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 94, + 140 + ], + "type": "text", + "content": "a pencil sket of a lemur" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 129, + 162, + 141 + ], + "lines": [ + { + "bbox": [ + 99, + 129, + 162, + 141 + ], + "spans": [ + { + "bbox": [ + 99, + 129, + 162, + 141 + ], + "type": "text", + "content": "h a pencil sketch of a kangaroo" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 100, + 81, + 162, + 129 + ], + "blocks": [ + { + "bbox": [ + 100, + 81, + 162, + 129 + ], + "lines": [ + { + "bbox": [ + 100, + 81, + 162, + 129 + ], + "spans": [ + { + "bbox": [ + 100, + 81, + 162, + 129 + ], + "type": "image", + "image_path": "60e13cf12c3f825e12d1f3e1aa39cd51d5186bf306fcb538da68bc03a54404b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 81, + 221, + 129 + ], + "blocks": [ + { + "bbox": [ + 173, + 81, + 221, + 129 + ], + "lines": [ + { + "bbox": [ + 173, + 81, + 221, + 129 + ], + "spans": [ + { + "bbox": [ + 173, + 81, + 221, + 129 + ], + "type": "image", + "image_path": "ebc9417e6357030c8d8d6cbc36c352109321d79e31e5ea1d6c345af1311393a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 129, + 217, + 141 + ], + "lines": [ + { + "bbox": [ + 173, + 129, + 217, + 141 + ], + "spans": [ + { + "bbox": [ + 173, + 129, + 217, + 141 + ], + "type": "text", + "content": "a watercolor of a rabbit" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 81, + 283, + 129 + ], + "blocks": [ + { + "bbox": [ + 221, + 81, + 283, + 129 + ], + "lines": [ + { + "bbox": [ + 221, + 81, + 283, + 129 + ], + "spans": [ + { + "bbox": [ + 221, + 81, + 283, + 129 + ], + "type": "image", + "image_path": "dbb70ef7122727c40692d583b525d34c3e4d24e158ab9ccbee8883c8a07d71de.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 130, + 277, + 141 + ], + "lines": [ + { + "bbox": [ + 237, + 130, + 277, + 141 + ], + "spans": [ + { + "bbox": [ + 237, + 130, + 277, + 141 + ], + "type": "text", + "content": "a watercolor of a duck" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 84, + 154, + 132, + 201 + ], + "blocks": [ + { + "bbox": [ + 56, + 149, + 83, + 156 + ], + "lines": [ + { + "bbox": [ + 56, + 149, + 83, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 83, + 156 + ], + "type": "text", + "content": "[16×16]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 84, + 154, + 132, + 201 + ], + "lines": [ + { + "bbox": [ + 84, + 154, + 132, + 201 + ], + "spans": [ + { + "bbox": [ + 84, + 154, + 132, + 201 + ], + "type": "image", + "image_path": "62985ce8ec139ba1c9ebbee338e48eccc0028d07a88f0f523cfd1a7116892138.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 203, + 132, + 216 + ], + "lines": [ + { + "bbox": [ + 83, + 203, + 132, + 216 + ], + "spans": [ + { + "bbox": [ + 83, + 203, + 132, + 216 + ], + "type": "text", + "content": "an oil painting of a young man" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 139, + 153, + 196, + 209 + ], + "blocks": [ + { + "bbox": [ + 175, + 145, + 278, + 152 + ], + "lines": [ + { + "bbox": [ + 175, + 145, + 278, + 152 + ], + "spans": [ + { + "bbox": [ + 175, + 145, + 278, + 152 + ], + "type": "text", + "content": "cf. \"Kanichen und Ende.\" Fliegende Blätter, 1892." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 153, + 196, + 209 + ], + "lines": [ + { + "bbox": [ + 139, + 153, + 196, + 209 + ], + "spans": [ + { + "bbox": [ + 139, + 153, + 196, + 209 + ], + "type": "image", + "image_path": "831727bd920fa7d3762c66bda5f78631933079b8b99f7f936346f9feb90f9368.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 204, + 154, + 252, + 201 + ], + "blocks": [ + { + "bbox": [ + 204, + 154, + 252, + 201 + ], + "lines": [ + { + "bbox": [ + 204, + 154, + 252, + 201 + ], + "spans": [ + { + "bbox": [ + 204, + 154, + 252, + 201 + ], + "type": "image", + "image_path": "75b032df22ddbc67ddfdfe6abea5b07081a84d38f3c33e5be0a4a01fd6cfcb69.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 203, + 252, + 216 + ], + "lines": [ + { + "bbox": [ + 203, + 203, + 252, + 216 + ], + "spans": [ + { + "bbox": [ + 203, + 203, + 252, + 216 + ], + "type": "text", + "content": "an oil painting of an old man" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 52, + 231, + 100, + 278 + ], + "blocks": [ + { + "bbox": [ + 59, + 220, + 85, + 228 + ], + "lines": [ + { + "bbox": [ + 59, + 220, + 85, + 228 + ], + "spans": [ + { + "bbox": [ + 59, + 220, + 85, + 228 + ], + "type": "text", + "content": "32×32" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 231, + 100, + 278 + ], + "lines": [ + { + "bbox": [ + 52, + 231, + 100, + 278 + ], + "spans": [ + { + "bbox": [ + 52, + 231, + 100, + 278 + ], + "type": "image", + "image_path": "6ae74fd3489ddf4d5c824aa3a324827e8daf3ccb3020cb313fc1eaf29a394413.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 100, + 231, + 164, + 279 + ], + "blocks": [ + { + "bbox": [ + 53, + 280, + 102, + 292 + ], + "lines": [ + { + "bbox": [ + 53, + 280, + 102, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 102, + 292 + ], + "type": "text", + "content": "an oil painting of a young man" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 231, + 164, + 279 + ], + "lines": [ + { + "bbox": [ + 100, + 231, + 164, + 279 + ], + "spans": [ + { + "bbox": [ + 100, + 231, + 164, + 279 + ], + "type": "image", + "image_path": "ddfd37af6f4c49f3d525a16387cc64a3b969a7b7a4d388b7f124363f3e6f08cb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 280, + 162, + 292 + ], + "lines": [ + { + "bbox": [ + 97, + 280, + 162, + 292 + ], + "spans": [ + { + "bbox": [ + 97, + 280, + 162, + 292 + ], + "type": "text", + "content": "ng an oil paintingin of an old man ," + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 296, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 296, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 287, + 340 + ], + "type": "text", + "content": "Figure 6. Permutation Illusions. We synthesize images whose appearance changes upon permutation of patches. Even in the difficult case of a " + }, + { + "bbox": [ + 47, + 296, + 287, + 340 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 47, + 296, + 287, + 340 + ], + "type": "text", + "content": " grid of patches, in which every pixel is effectively shuffled, we are able to generate meaningful images." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 171, + 231, + 220, + 278 + ], + "blocks": [ + { + "bbox": [ + 176, + 220, + 206, + 228 + ], + "lines": [ + { + "bbox": [ + 176, + 220, + 206, + 228 + ], + "spans": [ + { + "bbox": [ + 176, + 220, + 206, + 228 + ], + "type": "text", + "content": "64×64" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 231, + 220, + 278 + ], + "lines": [ + { + "bbox": [ + 171, + 231, + 220, + 278 + ], + "spans": [ + { + "bbox": [ + 171, + 231, + 220, + 278 + ], + "type": "image", + "image_path": "69abf4c288cc173d6e075a7862e1c00400346fbb6d1a793753a09d465e2eadfa.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 280, + 214, + 291 + ], + "lines": [ + { + "bbox": [ + 177, + 280, + 214, + 291 + ], + "spans": [ + { + "bbox": [ + 177, + 280, + 214, + 291 + ], + "type": "text", + "content": "a mosaic of a duck" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 220, + 231, + 285, + 279 + ], + "blocks": [ + { + "bbox": [ + 220, + 231, + 285, + 279 + ], + "lines": [ + { + "bbox": [ + 220, + 231, + 285, + 279 + ], + "spans": [ + { + "bbox": [ + 220, + 231, + 285, + 279 + ], + "type": "image", + "image_path": "ac663315c9966c922d45d6c64a28d582a7a68b3ea671b85d46840ae8a61712d0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 280, + 277, + 292 + ], + "lines": [ + { + "bbox": [ + 241, + 280, + 277, + 292 + ], + "spans": [ + { + "bbox": [ + 241, + 280, + 277, + 292 + ], + "type": "text", + "content": "a mosaic of a rabbit" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 350, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 458 + ], + "type": "text", + "content": "cating that it is working as intended. But this comes at the cost of worse alignment score. This is because the negative and positive prompt may have fundamental similarities. For example, using \"a cat\" as the negative prompt for the prompt \"an oil painting of a dog\" may discourage the model from synthesizing anything remotely cat-like—such as fur, four legs, or a tail—even if it helps in producing a dog. For this reason we opt not to use negative prompting with our method." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": "Guidance Scale. We also ablate out various guidance scales, " + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": ", for our method. We find that a higher guidance scale tends to do better. This is presumably because a higher guidance scale results in a sharper sampling distribution." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 521, + 159, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 159, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 159, + 533 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 539, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 670 + ], + "type": "text", + "content": "We show qualitative results in Fig. 1, Fig. 5, Fig. 6, and Fig. 7. Again, random samples may be found in Fig. 8 and Appendix D. Additional qualitative samples can be found in Appendix C. Overall, we find that our method can produce very high quality optical illusions for a wide range of views. Interestingly, our method often finds clever ways of reusing elements from one view for another, such as in the \"waterfalls\"/\"rabbit\"/\"teddy bear\" three-view illusion in Fig. 1, in which the nose of the teddy bear is the eye of the rabbit, and a rock on the waterfalls." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Baselines. We provide qualitative comparisons of our method to baselines in Fig. 5, where we pick the best images out of 100 samples for each method. As can be seen," + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 307, + 83, + 359, + 137 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 357, + 83 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 357, + 83 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 357, + 83 + ], + "type": "text", + "content": "an oil painting of an old woman" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 83, + 359, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 359, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 359, + 137 + ], + "type": "image", + "image_path": "6780f9bb96e032916b5336a8617d9b81193e4d9df5cc3166f4ef6346bb31ff81.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 138, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 306, + 138, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 545, + 193 + ], + "type": "text", + "content": "Figure 7. Orthogonal Illusions. We show that our method works, even when the view is a randomly sampled orthogonal transformation " + }, + { + "bbox": [ + 306, + 138, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 306, + 138, + 545, + 193 + ], + "type": "text", + "content": ". While these \"illusions\" are incomprehensible to human perception, they serve as a confirmation for our mathematical analysis." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 361, + 83, + 424, + 137 + ], + "blocks": [ + { + "bbox": [ + 374, + 72, + 420, + 83 + ], + "lines": [ + { + "bbox": [ + 374, + 72, + 420, + 83 + ], + "spans": [ + { + "bbox": [ + 374, + 72, + 420, + 83 + ], + "type": "text", + "content": "an oil painting of a young lady" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 83, + 424, + 137 + ], + "lines": [ + { + "bbox": [ + 361, + 83, + 424, + 137 + ], + "spans": [ + { + "bbox": [ + 361, + 83, + 424, + 137 + ], + "type": "image", + "image_path": "8dfa5c16125480acdd32228222eeb38fa866f901767a7b32e08355c6d61a21e6.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 428, + 83, + 481, + 137 + ], + "blocks": [ + { + "bbox": [ + 429, + 75, + 481, + 81 + ], + "lines": [ + { + "bbox": [ + 429, + 75, + 481, + 81 + ], + "spans": [ + { + "bbox": [ + 429, + 75, + 481, + 81 + ], + "type": "text", + "content": "a photo of a duck" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 83, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 428, + 83, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 428, + 83, + 481, + 137 + ], + "type": "image", + "image_path": "0c13a2c339044b8068911dd4afa40ddb3abdc55d05a1949fdc16a904b9284df7.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 482, + 83, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 488, + 75, + 545, + 80 + ], + "lines": [ + { + "bbox": [ + 488, + 75, + 545, + 80 + ], + "spans": [ + { + "bbox": [ + 488, + 75, + 545, + 80 + ], + "type": "text", + "content": "a photo of a rabbit" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 482, + 83, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 482, + 83, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 482, + 83, + 545, + 137 + ], + "type": "image", + "image_path": "612aa866ee39112ad67565a5a49a38d7576333f7c035cf85bbce3d8ca7236025.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "content": "images generated using our method match the prompts in both views equally well and are of higher quality." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": "Permutations. Pixel and patch permutations, being a subset of orthogonal transformations, should work with our method. We show that this is indeed the case in Fig. 6, where we have results on patch grids of various sizes under randomly sampled permutations. The " + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": " case is quite hard, yet our method is able to generate images that satisfy the constraint, albeit at lower quality." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 305, + 322, + 546, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 546, + 478 + ], + "type": "text", + "content": "Arbitrary Orthogonal Transformations. As discussed in Sec. 3.3, our method works for any orthogonal transformation. So far, we have shown illusions based on a subset of orthogonal views that correspond to intuitive image transformations. In Fig. 7, we show \"illusions\" using an arbitrary orthogonal transformation as a view. We use Stable Diffusion [36] and sample a random orthogonal matrix " + }, + { + "bbox": [ + 305, + 322, + 546, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{16384 \\times 16384}" + }, + { + "bbox": [ + 305, + 322, + 546, + 478 + ], + "type": "text", + "content": " by projecting an i.i.d. random Gaussian matrix with an SVD. These dimensions correspond to the size of the Stable Diffusion latent space. We note that this is an incredibly hard and unnatural transformation of an image, and results are accordingly of lower quality, but our method is still able to produce reasonable images." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 305, + 484, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 484, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 484, + 545, + 557 + ], + "type": "text", + "content": "Random Samples. We show random samples for selected prompts in Fig. 8. As can be seen, these random samples, while not as good as those in Fig. 1, are still very high quality. Some failure cases can be seen where the model prefers one prompt over another. We add further discussion and present more random samples in Appendix D." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 564, + 366, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 366, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 366, + 576 + ], + "type": "text", + "content": "4.4. Failures" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "content": "We highlight three interesting failure cases of our method in Fig. 9." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 614, + 545, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 721 + ], + "type": "text", + "content": "Independent Synthesis. The first of these cases involves the model synthesizing prompts separately, without combining elements of the two to form an illusion. Empirically, this happens surprisingly rarely, especially given that it seems to be such an easy shortcut solution. We hypothesize that this is because the diffusion model is biased toward centering its content, resulting in far more images with content that is integrated and centered as opposed to separate and off-center." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24160" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 302 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 302 + ], + "type": "image", + "image_path": "a15a49f015b1205e9f1ca11d1635a681b97bf6857b0c8dcdd8536a087c150025.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 546, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 546, + 326 + ], + "type": "text", + "content": "Figure 8. Random Samples. We show random samples, along with their corresponding view, for selected prompts. For more random samples please see Appendix D. For best quality, view digitally and zoom-in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 337, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 468 + ], + "type": "text", + "content": "Noise Shift. Using views that preserve noise statistics is critical to our method's success. For example, we attempted to recreate the \"Dress\" illusion [46], in which a dress can be seen as either \"blue and black\" or \"white and gold.\" We used simple white balancing as our view, in which pixel values were scaled by a constant factor. While this transformation is linear, it does not preserve the statistics of Gaussian noise. As a result, we see artifacts in the forms of spots, which we hypothesize is the result of the model interpreting the scaled Gaussian noise as signal and actively denoising peaks in the scaled noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 477, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 597 + ], + "type": "text", + "content": "Correlated Noise. While our method supports rotations as transformations, as demonstrated with the \"3-view,\" \"4-view,\" and \"Inner Rotation\" illusions in Fig. 1, care must be taken that the rotation does not introduce correlations in the noise, such as through anti-aliasing. For example, bilinear sampling introduces significant correlations in the noise, as it is a linear combination of four adjacent pixels. Therefore, seemingly innocuous rotations may result in divergent samples if transformations are not carefully kept correlation free, as shown with the 45 degree bilinear rotation in Fig. 9." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 209, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 209, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 209, + 622 + ], + "type": "text", + "content": "5. Limitations and Conclusions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "We present a method to produce compelling and diverse optical illusions. Our method is simple and straightforward to implement, and additionally amenable to theoretical analysis. We prove that our method works for a broad set of transformations, and qualitatively show that it can generate a wide array of optical illusions. However, at the same time many possible illusions and transformations are still" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 335, + 370, + 398 + ], + "blocks": [ + { + "bbox": [ + 307, + 335, + 370, + 398 + ], + "lines": [ + { + "bbox": [ + 307, + 335, + 370, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 370, + 398 + ], + "type": "image", + "image_path": "d49d0720d5e01c1700bf17fed202ce36077ef1225d768aa5dc67a4a1daf839c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 399, + 370, + 463 + ], + "blocks": [ + { + "bbox": [ + 307, + 399, + 370, + 463 + ], + "lines": [ + { + "bbox": [ + 307, + 399, + 370, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 370, + 463 + ], + "type": "image", + "image_path": "f3e25a3767a1636fbd3ad4548a33c27ee05256106c2a1129c3634eea9b1c9e7a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 464, + 370, + 527 + ], + "blocks": [ + { + "bbox": [ + 307, + 464, + 370, + 527 + ], + "lines": [ + { + "bbox": [ + 307, + 464, + 370, + 527 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 370, + 527 + ], + "type": "image", + "image_path": "463e60f322959c2406320ce19530d2cca2c11e2cdc3690651a36d06b9d582367.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 528, + 544, + 549 + ], + "lines": [ + { + "bbox": [ + 306, + 528, + 544, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 544, + 549 + ], + "type": "text", + "content": "Figure 9. Failures. We highlight three interesting failure cases, which are discussed in Sec. 4.4." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 375, + 344, + 487, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 344, + 487, + 355 + ], + "spans": [ + { + "bbox": [ + 375, + 344, + 487, + 355 + ], + "type": "text", + "content": "Failure: Independent Synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 357, + 440, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 357, + 440, + 366 + ], + "spans": [ + { + "bbox": [ + 375, + 357, + 440, + 366 + ], + "type": "text", + "content": "View: Vertical Flip" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 368, + 535, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 368, + 535, + 390 + ], + "spans": [ + { + "bbox": [ + 375, + 368, + 535, + 390 + ], + "type": "text", + "content": "an oil painting of a bowl of fruit \nan oil painting of a monkey" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 409, + 446, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 409, + 446, + 418 + ], + "spans": [ + { + "bbox": [ + 375, + 409, + 446, + 418 + ], + "type": "text", + "content": "Failure: Noise Shift" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 420, + 456, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 420, + 456, + 431 + ], + "spans": [ + { + "bbox": [ + 375, + 420, + 456, + 431 + ], + "type": "text", + "content": "View: White Balancing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 433, + 531, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 433, + 531, + 454 + ], + "spans": [ + { + "bbox": [ + 375, + 433, + 531, + 454 + ], + "type": "text", + "content": "a photo of a black and blue dressa photo of a white and gold dress" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 473, + 466, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 473, + 466, + 483 + ], + "spans": [ + { + "bbox": [ + 375, + 473, + 466, + 483 + ], + "type": "text", + "content": "Failure: Correlated Noise" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 485, + 537, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 485, + 537, + 496 + ], + "spans": [ + { + "bbox": [ + 375, + 485, + 537, + 496 + ], + "type": "text", + "content": "View: Rotate Inner Circle 45 Degrees (Bilinear)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 498, + 484, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 498, + 484, + 518 + ], + "spans": [ + { + "bbox": [ + 375, + 498, + 484, + 518 + ], + "type": "text", + "content": "a sketch of an elephant a sketch of a mouse" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "type": "text", + "content": "not possible using our method, such as color constancy illusions, homographies, stretches, and more generally nonvolume-preserving deformations. We leave implementation of these views for future work. Moreover, our method does not consistently produce perfect illusions. This may be a symptom of the difficulty of producing good illusions, but may indicate future work to be done to improve consistency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements We thank William Henning, Trenton Chang, Kimball Strong, Jeongsoo Park, Patrick Chao, Kurtland Chua, and Mohamed El Banani for their feedback on early drafts. Daniel is supported by the National Science Foundation Graduate Research Fellowship under Grant No. 1841052." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24161" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 75, + 106, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 75, + 106, + 87 + ], + "spans": [ + { + "bbox": [ + 48, + 75, + 106, + 87 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 139 + ], + "type": "text", + "content": "[1] AUTOMATIC1111. Negative prompt. https : // github . com / AUTOMATIC1111 / stable - diffusion - webui / wiki / Negative - prompt, 2022. Accessed: November 7, 2023. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 287, + 183 + ], + "type": "text", + "content": "[2] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael Ryoo. Diffusion illusions: Hiding images in plain sight. https://ryanndagreat.github.io/Diffusion-Illusions, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 185, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 287, + 228 + ], + "type": "text", + "content": "[3] Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum, and Jonathan Ragan-Kelley. Designing perceptual puzzles by differentiating probabilistic programs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 287, + 262 + ], + "type": "text", + "content": "[4] Ming-Te Chi, Chih-Yuan Yao, Eugene Zhang, and Tong-Yee Lee. Optical illusion shape texturing using repeated asymmetric patterns. The Visual Computer, 30:809–819, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 297 + ], + "type": "text", + "content": "[5] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 287, + 331 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 332, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 287, + 364 + ], + "type": "text", + "content": "[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 366, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 366, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 287, + 399 + ], + "type": "text", + "content": "[8] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 400, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 400, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 400, + 287, + 465 + ], + "type": "text", + "content": "[9] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "text", + "content": "[10] Werner Ehm. A variational approach to geometric-optical illusions modeling. Proceedings of Fechner Day, 27(1):41-46, 2011. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[11] Gamaleldin Elsayed, Shreya Shankar, Brian Cheung, Nicolas Papernot, Alexey Kurakin, Ian Goodfellow, and Jascha Sohl-Dickstein. Adversarial examples that fool both computer vision and time-limited humans. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 556, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 589 + ], + "type": "text", + "content": "[12] William T Freeman, Edward H Adelson, and David J Heeger. Motion without movement. ACM Siggraph Computer Graphics, 25(4):27-30, 1991. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[13] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[14] Alexander Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Alex Gomez-Villa, Adrián Martín, Javier Vazquez-Corral, Marcelo Bertalmío, and Jesús Malo. On the synthesis of" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 327, + 74, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 546, + 95 + ], + "type": "text", + "content": "visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 546, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 546, + 129 + ], + "type": "text", + "content": "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 130, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 172 + ], + "type": "text", + "content": "[17] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35:14715-14728, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "text", + "content": "[18] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 219, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 240 + ], + "type": "text", + "content": "[19] Aaron Hertzmann. Visual indeterminacy in gan art. In ACM SIGGRAPH 2020 Art Gallery, pages 424-428. 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "type": "text", + "content": "[20] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 276, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 297 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 298, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 330 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "type": "text", + "content": "[23] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 366, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 398 + ], + "type": "text", + "content": "[24] Mikhail Konstantinov, Alex Shonenkov, Daria Bakshandaeva, and Ksenia Ivanova. If by deepfloyd lab at stabilityai, 2023. GitHub repository. 2, 5, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 399, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 420 + ], + "type": "text", + "content": "[25] Monster Labs. Controlnet qr code monster v2 for sd-1.5, 2023.3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[26] Nan Liu, Shuang Li, Yilun Du, Josh Tenenbaum, and Antonio Torralba. Learning to compose visual relations. Advances in Neural Information Processing Systems, 34: 23166-23178, 2021. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "text", + "content": "[27] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[28] Dominique Makowski, Zen J Lau, Tam Pham, W Paul Boyce, and SH Annabel Chen. A parametric framework to generate visual illusions using python. Perception, 50(11):950-965, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "type": "text", + "content": "[29] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? 2023. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "text", + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 624, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 645 + ], + "type": "text", + "content": "[31] Aude Oliva, Antonio Torralba, and Philippe G. Schyns. Hybrid images. ACM Trans. Graph., 25(3):527-532, 2006. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[32] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. 2014. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24162" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 149 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 207, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 287, + 262 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 363 + ], + "type": "text", + "content": "[38] Troy Shinbrot, Miguel Vivar Lazo, and Theo Siu. Network simulations of optical illusions. International Journal of Modern Physics C, 28(02):1750018, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "type": "text", + "content": "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 509 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "text", + "content": "[42] Matthew Tancik. Illusion diffusion. https://github.com/tancik/Illusion-Diffusion, 2023.2, 3, 5, 6, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[43] Ugleh. Spiral town - different approach to qr monster. https://www.reddit.com/r/StableDiffusion/comments/16ew9fz/spiral_town_differentapproach_to_qr_monster/, 2023.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 700 + ], + "type": "text", + "content": "[45] Xi Wang, Zoya Bylinskii, Aaron Hertzmann, and Robert Pepperell. Toward quantifying ambiguities in artistic images. ACM Transactions on Applied Perception (TAP), 17(4):1-10, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "text", + "content": "[46] Wikipedia contributors. The dress. https://en." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "wikipedia.org/wiki/The_dress. Accessed: November 9, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "text", + "content": "[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24163" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_content_list.json b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02bcaba1b2c88f6c639b3d3314d431ca8ab6cdfa --- /dev/null +++ b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_content_list.json @@ -0,0 +1,1559 @@ +[ + { + "type": "text", + "text": "Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models", + "text_level": 1, + "bbox": [ + 91, + 128, + 879, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthew Kowal1,3 Richard P. Wildes1,2 Konstantinos G. Derpanis1,2,3", + "bbox": [ + 186, + 202, + 781, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ York University, $^{2}$ Samsung AI Centre Toronto, $^{3}$ Vector Institute yorkucvil.github.io/VCC", + "bbox": [ + 269, + 222, + 699, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 294, + 313, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding what deep network models capture in their learned representations is a fundamental challenge in computer vision. We present a new methodology to understanding such vision models, the Visual Concept Connectome (VCC), which discovers human interpretable concepts and their interlayer connections in a fully unsupervised manner. Our approach simultaneously reveals fine-grained concepts at a layer, connection weightings across all layers and is amendable to global analysis of network structure (e.g. branching pattern of hierarchical concept assemblies). Previous work yielded ways to extract interpretable concepts from single layers and examine their impact on classification, but did not afford multilayer concept analysis across an entire network architecture. Quantitative and qualitative empirical results show the effectiveness of VCCs in the domain of image classification. Also, we leverage VCCs for the application of failure mode debugging to reveal where mistakes arise in deep networks.", + "bbox": [ + 75, + 325, + 473, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 609, + 209, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper focuses on interpreting the intermediate representations of deep networks for computer vision. The goal is understanding how various encoded concepts impact a model's prediction as well as concepts in other layers. We define concepts as abstractions that generalize from particular instances, including those defined locally (e.g. color and orientation), regionally (e.g. texture and shading) and from higher-level considerations (e.g. object parts, wholes and groupings). Human-interpretable concepts are of particular interest in increasing human understanding of models; however, extracting these concepts encoded in deep networks remains an open challenge in computer vision due to the complexity and opaque nature of these models.", + "bbox": [ + 75, + 628, + 470, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding what concepts are learned by deep models and how they are encoded is important for both science and applications. For science, understanding what information drives a model's encoding of different concepts may indicate directions to advance performance. For applications,", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "several negative consequences of deploying opaque vision models have been documented, e.g. [8, 24].", + "bbox": [ + 498, + 296, + 890, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous work has focused on interpreting models via feature attribution, which measures the contribution of individual inputs to a model's output [10, 46, 59]; so, explanations are of single pixels and may be difficult to interpret. Other work generates images that maximize activation of a model's features [19, 38, 54]. Like feature attribution, these approaches are qualitative and place most of the burden on the user to determine the concepts revealed. Concept-based interpretability, which identifies human-interpretable abstractions in a model's latent space [20, 23, 29, 32], yield quantitative contributions of a concept to the model's output and explanations on the class-level (vs. pixel-level). These approaches are easier to understand and validate; however, they have not been used to explore interlayer relationships.", + "bbox": [ + 496, + 327, + 892, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As it stands, no approaches can quantify the interlayer affect of a given distributed concept at one layer, $l$ , to another concept at a different layer, $l'$ (rather than the model output). Even though it is well established that deep networks learn to build concepts hierarchically as information flows through the network [39, 54], understanding the hierarchical representations has been under-researched. Indeed, little is known about the characteristics of this concept hierarchy for today's models. Questions abound: How many concepts exist in a network? What are the connections and weights between concepts? Does the model architecture impact the hierarchical structure of concept abstractions?", + "bbox": [ + 496, + 537, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to these questions, we take inspiration from the biological notion of a connectome [47], defined as “a comprehensive structural description of the network of elements and connections forming a brain.” Analogously, we present the Visual Concept Connectome (VCC), a comprehensive structural description of a deep pretrained network model in terms of human-interpretable concepts and their relationships that form the internal representation maintained by the model; Fig. 1 shows an example. Notably, VCC generation works in the open-world setting, i.e. the concepts and interlayer connections are discovered without the need for any predefined concept dictionary.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10895", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d5f6b084c6bcf97c82706cceb1a60f6fabd7772e1316cc51488b53170d1037b.jpg", + "image_caption": [ + "Figure 1. A Visual Concept Connectome (VCC). At each layer the visual concepts learned by a deep model for a given class are revealed as are the learned interlayer concept connections. For each concept, up to four exemplars are shown as unmasked regions in a $2 \\times 2$ image. Interlayer concept connections are shown as lines with darker lines indicating larger contributions. Shown is a VCC for every convolutional layer of a VGG16 model [49] trained on ImageNet [15] targeting recognition of class \"Tow Truck\". A closer visualization of VCC subgraphs reveals interesting compositions occurring at different levels of abstraction corresponding at different depths of the model. At early layers (bottom left), we observe oriented patterns ( $C_{1,1}$ ) and brown color ( $C_{1,2}$ ) composing the concept of green and brown orientation ( $C_{2,1}$ ). Middle layers (right) show the concept of 'wheel on the road' ( $C_{9,1}$ ) being composed of wheels ( $C_{8,1}$ ) and regions of asphalt ( $C_{8,2}$ ). The final layer concepts (top left) show that both foreground objects, e.g. tow trucks ( $C_{13,1}$ ), and background regions, e.g. road, trees, humans, or car being towed ( $C_{13,2}$ ), concepts highly influence the final category ( $C_{TowTruck}$ )." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 854, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related research", + "text_level": 1, + "bbox": [ + 76, + 517, + 241, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While a number of intrinsically interpretable networks have appeared (e.g. [6, 12, 31, 45, 56, 57]), we focus on work that, like ours, endeavours to interpret black-box models.", + "bbox": [ + 75, + 537, + 468, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept-based interpretability. Closed-world concept interpretability considers cases where a labelled dataset defines the concepts of interest for post-hoc analysis; approaches include Network Dissection [3] and Testing with Concept Activation Vectors (TCAV) [29]. However, a desirable property for concept-based interpretability is not to be restricted to a set of predefined classes (i.e. closed-world), but to support discovery of new and previously unlabelled concepts (i.e. open-world). Approaches to open-world concept interpretability include Automatic Concept Explanations (ACE) [23], SegDiscover [27], ConceptSHAP [53], Invertible Concept Explanations [58] and CRAFT [20]. These approaches are limited to single layer analysis and do not explain interlayer relationships.", + "bbox": [ + 75, + 583, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Activation maximization [19, 36, 38, 54] visualizes the input that most activates a model component (e.g. filter, layer or logit) or how same layer units combine information [40]. These approaches do not capture interlayer relations and place a heavy burden on the user to interpret what concepts appear in their output where there may be no clear resemblance to natural concepts. Class Activation Maps", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(CAMs) and extensions visualize an input image's local region that contributes to a model's output [9, 46, 59]. Layerwise Relevance Propagation [5] (LRP) generates pixel contribution heatmaps by assuming conservation of information is propagated through each neuron. These approaches interpret single images (not classes), are purely qualitative, and are sensitive to small input perturbations [22, 30], yet insensitive to model changes [2]. Less related are directions interpreting generative models, e.g. [4, 34].", + "bbox": [ + 496, + 518, + 890, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpretability via decomposition. Most closely related to our work are those that decompose a model's prediction into interpretable components. CRAFT [20] extracts concepts from the last layer of a convolutional neural network (CNN), but also searches earlier layers for sub-concepts for the maximally activating images of a given concept, but do not provide any way to quantify the contribution of sub-concepts to concepts. Another approach combined CAMs of multiple concepts to explain the final prediction [60]; however, it is limited to labelled data (so cannot discover concepts outside training) and only explains a final output, not intermediate relations. Other work interprets CNNs by distilling them into a graph [55]; however, it applies only to individual filters conceived as capturing object parts (i.e. not concepts other than objects and their parts, e.g. textures, colors or grouping), does not con", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10896", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sider the fact that representations are encoded via superposition [7, 17, 40] (i.e. more than one concept can be captured per-channel), does not yield meaningful edge weights and is only for CNNs. Yet other work extends Grad-CAM or LRP to perform intra-layer visualizations [1, 13]. While these approaches capture concepts other than object parts, they still ignore the core problem of superposition (and distributed representations) and do not yield interpretation for an entire class, just single images.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contributions. In the light of previous work, our contributions are fourfold. (i) We introduce and formalize the notion of a Visual Concept Connectome (VCC) for deep network models. The VCC reveals concepts represented at any given layer of a network as well as their interlayer connectivity. (ii) We present a method for extracting a VCC from any pretrained deep network in an unsupervised, open world setting, with a focus on classification. (iii) We validate our approach with quantitative and qualitative experiments wherein we examine various standard models to yield insights into model architectures and training tasks. (iv) We apply VCCs to explaining failure modes in deep models.", + "bbox": [ + 75, + 227, + 472, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Visual Concept Connectomes (VCCs)", + "text_level": 1, + "bbox": [ + 76, + 421, + 413, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A VCC is a directed acyclic graph, $\\mathbf{G}(\\mathbf{Q},\\mathbf{E})$ , created by distilling a pretrained deep network. The graph is topologically sorted based on the layer order found in the original network and has $n + 1$ layers, consisting of $n$ network layers to be analyzed (i.e. any subset of layers that a user selects, including all layers at the extreme) plus the final prediction (e.g. the object category for object recognition). The graph's nodes, $\\mathbf{Q}$ , are vectors (cluster centroids) representing interpretable concepts and its edges, $\\mathbf{E}$ , are scalars representing the contribution of one concept to the existence of another.", + "bbox": [ + 75, + 446, + 468, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To construct a VCC, three inputs are required: (i) a set of $I$ images representative of a given task (e.g. exemplar images of an object category for object recognition), $\\mathcal{I} = \\{\\mathcal{I}^1,\\dots ,\\mathcal{I}^I\\}$ , $\\mathcal{I}^i\\in \\mathbb{R}^{h\\times w\\times c}$ , with $h,w$ and $c$ the image height, width and channel dimension (e.g. three for RGB), respectively, (ii) an $N$ layer network, $F(\\cdot) = \\{f_{1}(\\cdot),\\ldots ,f_{N}(\\cdot)\\}$ , with $f_{j}$ denoting feature extraction at layer $j$ , and (iii) a set of $n$ selected layers (a subset of $F$ , possibly improper), which are to be studied.", + "bbox": [ + 75, + 598, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A VCC is constructed in three main steps. (i) Image segments are extracted in the model's feature space via divisive clustering to produce semantically meaningful image regions for each selected layer (Sec. 3.1). (ii) Layer-wise concepts, i.e. the nodes of the graph, are discovered in an open world fashion (i.e. no labelled data is required) via a second round of clustering over the dataset of image regions, independently for each layer (Sec. 3.2). (iii) Edges are calculated that indicate the contribution of concepts from earlier to deeper layers via an approach we introduce, Interlayer Testing with Concept Activation Vectors (ITCAVs)", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(Sec. 3.3). We use object recognition for the explanation of the approach; nevertheless, it could be extended to other tasks in a straightforward way (e.g. semantic segmentation). An overview of these steps is provided in Fig. 2.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature space image segments", + "text_level": 1, + "bbox": [ + 498, + 159, + 767, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each selected layer, we produce a set of image regions that plausibly belong to a concept encoded by the model at that layer. Previous work produces concept segments in RGB space, i.e. superpixels or random crops [20, 23]; however, segmentation based on features different from those over which concepts are to be discovered, i.e. the model's deep features, results in the support for the discovery process being divorced from the process that generated the support. Therefore, our segmentation uses the same deep features to be used subsequently for concept discovery. Our divisive clustering approach is presented visually in Fig. 2 (A). We pass the entire set of $I$ input images, $\\mathcal{I}$ , for a given class through the model, $F$ , to get features, $\\mathbf{z}_n^i \\in \\mathbb{R}^{h_n \\times w_n \\times c_n}$ , at each selected layer, $n$ according to", + "bbox": [ + 496, + 181, + 890, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {n} = \\left\\{f _ {n} \\left(\\mathcal {I} ^ {1}\\right), \\dots , f _ {n} \\left(\\mathcal {I} ^ {I}\\right) \\right\\} = \\left\\{\\mathbf {z} _ {n} ^ {1}, \\dots , \\mathbf {z} _ {n} ^ {I} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 398, + 890, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract image segments at layer $n$ for concept discovery (Sec. 3.2), we cluster the image activations, $\\mathbf{Z}_n$ , conditioned on the clusters (i.e. masks) generated at the next higher layer, $n + 1$ . Let $\\mathbf{p} = (x,y)$ index spatial coordinates and $\\mathbf{B}_n^i (\\mathbf{p};\\gamma)$ be a binary mask for cluster $\\gamma$ such that $\\mathbf{B}_n^i (\\mathbf{p};\\gamma) = 1 \\iff \\mathbf{z}_n^i (\\mathbf{p}) \\in \\gamma$ ; otherwise, $\\mathbf{B}_n^i (\\mathbf{p};\\gamma) = 0$ .", + "bbox": [ + 496, + 425, + 890, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We calculate a set of such masks by applying a clustering algorithm, $\\mathsf{C}_{\\Gamma}^{seg}(\\cdot)$ , that returns binary support masks for $\\Gamma$ clusters on its argument. Let $\\Gamma_{n+1}$ be the number of segments at layer $n+1$ and $1 \\leq g \\leq \\Gamma_{n+1}$ index a particular segment, $g$ . For each $g$ we calculate segments at layer $n$ as", + "bbox": [ + 496, + 516, + 890, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {B} _ {n} ^ {i} (\\mathbf {p}; \\gamma) \\right\\} _ {\\gamma = 1} ^ {\\Gamma} = \\mathrm {C} _ {\\Gamma} ^ {s e g} \\left(\\mathbf {z} _ {n} ^ {i} (\\mathbf {p}) \\odot \\widetilde {\\mathbf {B}} _ {n + 1} ^ {i} (\\mathbf {p}; g)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 595, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{B}}_{n + 1}^i$ is $\\mathbf{B}_{n + 1}^i$ upsampled to the resolution of layer $n$ and $\\odot$ indicates spatial element-wise masking applied to the individual image activations. The element-wise masking ensures clustering is done on the masked area of $\\mathbf{z}_n^i$ . Thus, the support of the concept discovery at layer $n$ comes from the support of concepts at layer $n + 1$ . Notably, we let $\\Gamma$ vary with $g$ ; so, different segments, $g$ , at layer $n + 1$ can yield a different number of segments, $\\Gamma$ , at layer $n$ .", + "bbox": [ + 496, + 628, + 890, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The top-down conditional clustering, (2), is repeated recursively for all image segments in all selected layers. To initiate the recursion, (2), at the top layer, $n_{top}$ , all the features are used by setting the number of clusters to be one, and having $\\widetilde{\\mathbf{B}}_{n_{top} + 1}^i (\\mathbf{p};1) = 1$ for all images.", + "bbox": [ + 496, + 753, + 890, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At each layer, $n$ , we construct a set of RGB segmentations, $\\mathbf{M}$ , for use in concept discovery (Sec. 3.2) by applying upsampled masks, $\\mathbf{B}_n^i (\\mathbf{p};\\gamma)$ , to a given image, $\\mathcal{I}^i$ , via", + "bbox": [ + 496, + 833, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {n} ^ {i} (\\mathbf {p}; \\gamma) = \\mathcal {I} ^ {i} (\\mathbf {p}) \\odot \\mathbf {B} _ {n} ^ {i} (\\mathbf {p}; \\gamma), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10897", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/457a2c85cd74cfe2897fc892925150a793916aa62f396961d5e3ba0e856b1358.jpg", + "image_caption": [ + "Figure 2. The three steps in building a Visual Concept Connectome (VCC). (A) For a given image, $\\mathcal{I}^i\\in \\mathcal{I}$ , model, $F$ , and layer, $n$ , we produce a set of image segments, $\\mathbf{M}_n^i\\in \\mathbf{M}_n$ , based on a recursive spatial clustering, $\\mathsf{C}^{seg}$ (2), of the features $\\mathbf{z}_n^i$ conditioned on the clusters from the layer above, $n + 1$ . We then use (3) to generate a set of masked RGB image segments for each layer, $\\mathbf{M}_j$ . (B) For a given layer, $j$ , we pass the image segments from all images, $\\mathbf{M}_j$ , through $f_{j}$ and cluster, $\\mathsf{C}^{con}$ (4), these features across the dataset to produce $m_j$ concept centroids, $\\{\\mathbf{q}_j^1,\\dots ,\\mathbf{q}_j^{m_j}\\}$ . (C) To measure the contribution of an earlier layer concept, $\\mathbf{q}_j^{m_j}$ , to a later layer concept, $\\mathbf{q}_l^{m_l}$ , we employ our Interlayer Testing with CAV (ITCAV) approach (Sec. 3.3), which uses the Concept Activation Vector (CAV) [29] of the earlier concept, $\\mathbf{V}_{\\mathfrak{q}_j^{m_j}}$ (that points away from random examples, $\\mathcal{I}_{rnd}$ , but toward concept exemplars, $\\mathbf{M}_{\\mathfrak{q}_j^{m_j}}$ ), and the deeper layer concept, $\\mathbf{q}_l^{m_l}$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 428, + 343 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5f2d71779ddf2a157a18eb67605389de8bb55a4b57b2b8f47ecafc83038d903c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 90, + 872, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_n^i (\\mathbf{p};\\gamma)\\in \\{0,\\dots,255\\}^{h_n\\times w_n}$ , with $\\{0,\\dots,255\\}$ specifying RGB value. We follow by defining $\\mathbf{M}_n$ as the set of all RGB image segments at layer $n$ (with each segment given in terms of (3)), and letting $\\mathbf{M} = \\{\\mathbf{M}_1,\\dots,\\mathbf{M}_n\\}$ .", + "bbox": [ + 75, + 450, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We instantiate the clustering, $C^{seg}$ , via maskSLIC [28], an extension of k-means, with an $l_{2}$ norm, that clusters features while respecting a mask and automate selection of the number of clusters via the silhouette method [43].", + "bbox": [ + 75, + 513, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Layer-wise concept discovery", + "text_level": 1, + "bbox": [ + 76, + 582, + 336, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the dataset of masked image segments for each layer, $\\mathbf{M} = \\{\\mathbf{M}_1,\\dots ,\\mathbf{M}_n\\}$ , we follow [23] and bilinearly interpolate the segments to the input image size and then pass the segments through the model to get pre-segmented activations $\\mathbf{Z}_{\\mathbf{M}_j} = f_j(\\mathbf{M}_j)$ from the layer where the segments were found in the divisive clustering step (Sec. 3.1). We pass the image segments (as opposed to using the image activations) through the model to remove any contextual information before we discover concepts (the segments are set to zero everywhere outside of the segment boundary). Otherwise, $\\mathbf{M}_j$ could mix information from outside the segment.", + "bbox": [ + 75, + 604, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The layer-wise concept discovery step is presented visually in Fig. 2 (B). With the pre-segmented activations, $\\mathbf{Z}_{\\mathbf{M}_j}$ , calculated for all layers, $j$ , we cluster over the dataset of segment features for each individual layer to produce cluster centroids (i.e. concepts) according to", + "bbox": [ + 75, + 771, + 468, + 847 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {j} = \\left\\{\\mathbf {q} _ {j} ^ {1}, \\dots , \\mathbf {q} _ {j} ^ {m _ {j}} \\right\\} = \\mathrm {C} _ {m _ {j}} ^ {\\text {c o n}} (\\operatorname {G A P} \\left(\\mathbf {Z} _ {\\mathbf {M} _ {j}}\\right)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 857, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_j$ is the number of concepts discovered at layer $j$ ,", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GAP is spatial global average pooling and $\\mathsf{C}_{m_j}^{con}$ is a clustering algorithm that returns $m_j$ cluster centroids. Following previous work [23], we instantiate $\\mathsf{C}_{m_j}^{con}$ in terms of standard k-means, using the $l_2$ norm, and then use a large number of clusters with subsequent pruning to remove noisy clusters; see the supplement for details. The output of this stage results in discovered concepts (i.e. cluster centroids and associated segments) for all layers as $\\mathbf{Q} = \\{\\mathbf{Q}_1, \\dots, \\mathbf{Q}_n\\}$ for the centroids and $\\mathbf{M}_{\\mathbf{q}_j^{m_j}}$ their associated RGB segments.", + "bbox": [ + 496, + 450, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Interlayer concept connectivity", + "text_level": 1, + "bbox": [ + 500, + 602, + 772, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach to quantifying the contribution of discovered concepts between two layers, Interlayer Testing with Concept Activation Vectors (ITCAV), generalizes TCAV [29]. Specifically, while TCAV calculates the contribution of a single mid-layer concept to the class logit using a dataset of labelled concept images, we generalize to operate between any two layers and without a dataset of labelled concept images. Without loss of generality, we consider a single pair of concepts at two layers, which need not be adjacent.", + "bbox": [ + 496, + 625, + 890, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ITCAV method is presented visually in Fig. 2 (C). We seek to measure the degree to which a concept at an earlier layer contributes to a deeper layer. To do so, we construct two vectors: (i) a vector that represents the concept at the earlier layer, $j$ , and (ii) a vector that represents a positive contribution (i.e. gradient) of a concept at a later layer, $l$ . The closer the alignment of these two vectors, the larger the contribution of the concept at layer $j$ to the concept at layer $l$ . Let $\\mathbf{q}_j^{m_j}$ and $\\mathbf{M}_{\\mathbf{q}_j^{m_j}}$ be the cluster centroid and as", + "bbox": [ + 496, + 762, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10898", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5881e4c22dd331a2fa5610ade43f18d73597099d7caabd81203b32ddb60b19c3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 261, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/92390b4b63fbd6b7de8210b4b6988778a1cbaa23eba2cb61d1b41d10143c7c06.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 295, + 97, + 483, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2efe2dfd7064929fbf47a9aa2501d613deb5aba2f5c35f32c81a1dcf570ab1b.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 491, + 90, + 687, + 204 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ad322a960174b95501f00da33b42799d03fe1798b0309ac4bb33968636edf53.jpg", + "image_caption": [ + "(d)", + "Figure 3. Validation of the three VCC method components. (a) Validation of segment proposals. The relative (Rel.) concept segment size compared to entire image for a given layer is plotted against the receptive field (RF) width/height of the same layer. (b) Validation of discovered concepts. For 50 randomly selected ImageNet classes, we discover concepts in four layers of the model. During inference, one randomly selected concept at each layer is suppressed by a factor of $\\epsilon$ . (c, d) Validation of interlayer concept weights. The unnormalized logit sum (LS) scores, (8), for the target class are plotted against the average path strength (APS) scores, (7). A positive correlation implies that the ITCAV edge weights connecting a concept to the class are predictive of the model output having a higher probability for that class." + ], + "image_footnote": [], + "bbox": [ + 691, + 90, + 883, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sociated RGB image segments in layer $j$ , and let $\\mathbf{q}_l^{m_l}$ and $\\mathbf{M}_{\\mathbf{q}_l^{m_l}}$ be the cluster centroid and associated RGB image segments in a deeper layer $l$ , i.e. $l > j$ .", + "bbox": [ + 75, + 311, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We proceed by constructing the vector that represents the concept in the feature space of the earlier layer, $j$ . To do so, we employ the Concept Activation Vector (CAV) [29] for the earlier concept, $\\mathbf{q}_j^{m_j}$ , by training a linear classifier, $h(\\cdot)$ on the features of layer $j$ between positive concept inputs, $f_{j}(\\mathbf{M}_{\\mathbf{q}_{j}^{m_{j}}})$ , and random images, $f_{j}(\\mathcal{I}_{rnd})$ . The orthogonal vector to the hyperplane defined by $h(\\cdot)$ , denoted as $\\mathbf{V}_{\\mathbf{q}_j^{m_j}}$ is the CAV for concept $\\mathbf{q}_j^{m_j}$ and points in the direction of concept $\\mathbf{q}_j^{m_j}$ in the feature space of layer $j$ .", + "bbox": [ + 75, + 359, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we construct the vector in the feature space of layer $j$ that points in the direction of positive contribution of concept $\\mathbf{q}_l^{m_l}$ . To do so, we calculate $\\nabla f_l(f_j(\\cdot))$ , i.e. the gradient of the deeper concept, $\\mathbf{q}_l^{m_l}$ , at layer $l$ , with respect to the earlier layer, $j$ . Our approach to measuring the sensitivity between the two concepts is then", + "bbox": [ + 75, + 507, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} (x) = \\nabla \\left(\\left\\| f _ {l} \\left(f _ {j} (x)\\right) - \\mathbf {q} _ {l} ^ {m _ {l}} \\right\\| _ {2}\\right) \\cdot \\mathbf {V} _ {\\mathbf {q} _ {j} ^ {m _ {j}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 599, + 468, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x \\in \\mathbf{M}_{\\mathbf{q}_l^{m_l}}$ . The sensitivity score should be positive because when the feature at layer $j$ is perturbed in the direction of concept $\\mathbf{q}_j^{m_j}$ , i.e. in the direction $\\mathbf{V}_{\\mathbf{q}_j^{m_j}}$ , then the feature vector in deeper layer, $l$ , will be pushed closer to concept $\\mathbf{q}_l^{m_l}$ . Following TCAV [29], the final ITCAV edge weight, $e_{\\mathbf{q}_j^{m_j},\\mathbf{q}_l^{m_l}} \\in \\mathbf{E}$ , is the ratio of positive sensitivities,", + "bbox": [ + 75, + 632, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} = | x \\in \\mathbf {M} _ {\\mathbf {q} _ {l} ^ {m _ {l}}}: S _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} (x) > 0 | / | \\mathbf {M} _ {\\mathbf {q} _ {l} ^ {m _ {l}}} |. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 734, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Key to our sensitivity score, (5), is the $l_{2}$ distance between the output, $f_{l}(f_{j}(x))$ , and cluster centroid, $\\mathbf{q}_l^{m_l}$ , before taking the gradient, cf. CAV [29]. This choice serves two goals: (i) The subset of dimensions of individual image features, $f_{j}(x)$ , that are not well aligned with the cluster centroid, $\\mathbf{q}_l^{m_l}$ , will be penalized. This penalty will suppress the impact of noisy dimensions from individual samples on ITCAV. We use the $l_{2}$ norm because the employed clustering approach used during concept discovery, k-means, uses", + "bbox": [ + 73, + 763, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$l_{2}$ ; see Sec. 3.2. (ii) Collapsing the tensor to a scalar (i.e. a concept similarity score) makes the gradient calculation computationally feasible compared to the full Jacobian.", + "bbox": [ + 496, + 311, + 892, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 367, + 633, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 498, + 391, + 714, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation we use ResNet50 [25], VGG16 [48], MobileNetv3 [26], ViT-b [51] and MViT [18] to sample a broad range of popular architectures. The exact layers selected are in the supplement, but always include the last feature layer before the fully connected layer to probe concepts closest to the model output. Following previous work [23, 29], we perform a two-sided t-test on ITCAV scores for multiple runs of the same concept vs. different random sets of images to remove statistically insignificant scores and randomly sample from the Broden dataset [3] to get the images, $\\mathcal{I}_{rnd}$ (Sec. 3.3), used in statistical testing and CAV training. On average, four and all-layer VCCs take 15 minutes and 36 hours, resp., to generate on an NVIDIA Quadro RTX 6000 GPU. The supplement has more details and VCCs for other layers, models and datasets [52].", + "bbox": [ + 496, + 414, + 893, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. VCC component validation", + "text_level": 1, + "bbox": [ + 498, + 650, + 746, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segment proposal validation. Our approach segments concepts based on the feature space of a given layer to allow us to capture concepts across multiple layers (Sec. 3.1). So, the spatial support of valid segments at a layer should follow the receptive field (RF) of filters at that layer, as the segments are constrained by the filters from which they are derived. For comparison, we consider ACE [23], which does segmentation with a layer independent feature (color), limited to the input image to discover concepts at a single layer for a single model. Being directly related to the features at a layer, our approach should produce concepts that better respect the RF at that layer compared to the baseline, whose features are layer independent (e.g. at CNN early layers the RF is small; so, intuitively the patch proposals should be smaller as well and conversely for the later layers).", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10899", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76f1e05871175636312fd26837cec36445a232f6a6369bbf3f7cbd8513d7d1e1.jpg", + "image_caption": [ + "Figure 4. A VCC for three selected layers of a GoogLeNet model [49] targeting recognition of class 'Jay'. Darker lines denote stronger connection weights." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 472, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 (a) shows the average concept size (in terms of the ratio of segment pixels to the entire image) vs. the RF at the corresponding layer for three different models: (i) ResNet50 [25], (ii) VGG16 [48] and (iii) ViT-b [16] (note we use mean attention distance instead of receptive field for ViT-b). While the baseline [23] produces segments with sizes invariant to the layer analyzed, our approach produces segments with image sizes that scale with the RF. Additional results with other models (including more transformers) are given in the supplement, which further support this observation. Overall, it is expected by design and we find that our segment proposals follow RFs at each layer.", + "bbox": [ + 75, + 476, + 472, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concept fidelity. Concept fidelity measures the meaningfulness of discovered concepts with respect to the target model. While other work focuses on single layer concept fidelity [20, 23, 29], we measure the effect on a model's output when suppressing the encoding of concepts as the information propagates through the model. The expectation is that suppressing the discovered concepts should result in a quicker decrease in performance compared to a non-concept direction (e.g. a randomly selected direction) as the amount of suppression gets larger. For a given model and category, we compute the concepts at various layers using our method (Sec. 3.2). We perform concept suppression for a given image, $\\mathcal{I}^i$ , at a given layer, $j$ , for concept $\\mathbf{q}_j$ , according to $\\mathbf{z}_j^i = \\mathbf{z}_j^i - \\epsilon \\cdot \\mathbf{q}_j / ||\\mathbf{q}_j||_2$ , where $\\epsilon$ controls the degree of perturbation. If the concepts discovered are meaningful, then the accuracy for the model's target class should decrease faster", + "bbox": [ + 75, + 657, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc44278740832248dd9bdd782de3e20e213c70c052d9e739efad4ffddfb990f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 92, + 702, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c30cc6e647fdc7cd8c27d2617d17a6aa47114bcae5e1cea57208ee5f651219d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 92, + 895, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cff0b91a0307d8f889552b5c476100f36baf0122cf0ac0bca4c45dd26b679c51.jpg", + "image_caption": [ + "Figure 5. Graph metrics on four layer VCCs comparing CNN vs. transformer architectures and training objectives." + ], + "image_footnote": [], + "bbox": [ + 511, + 169, + 699, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34ebd885aa6c5d2ee5c2b43d285ff062691e7f2c137d2376e9fb78d93dd3af9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 167, + 893, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compared to random perturbations, $\\mathbf{q}_{rnd}$ , as $\\epsilon$ increases.", + "bbox": [ + 498, + 280, + 870, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 (b) has concept suppression results averaged over 50 randomly selected ImageNet [15] classes (documented in the supplement), where a concept is randomly chosen to be suppressed at each layer. Note that the $\\epsilon$ values required to reduce the model accuracy to zero differs by model; so, we scale the $\\epsilon$ values to [0, 1] for visualization purposes. For all models, it is seen that perturbing in the opposite direction of a recovered concept more severely impacts the accuracy than a random perturbation.", + "bbox": [ + 496, + 295, + 890, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ITCAV validation. If the interlayer concept connection weights are meaningful, then the accumulated weights from earlier to later layer concepts should be correlated with their probability of predicting the target class. We use this intuition to validate our ITCAV approach to recovering interlayer concept weights as follows. We define the average path strength (APS), a scalar value between zero and one representing the average strength of the connection between a concept and the class logit. Given concept $\\mathbf{q}_j$ at layer $j$ , let there be $L$ layers in the VCC between the concept and the logit layer; thus, each path from the concept to the logit consists of $L$ edge weights. We consider all possible paths in the VCC from this concept to the class logit. Let $e_{\\mathbf{q}_j}(l,p)$ be the edge weight (i.e. ITCAV score) of the $p^{\\mathrm{th}}$ path (from concept $\\mathbf{q}_j$ to the class concept) for VCC layer $l$ . To calculate the APS score for a given concept, we average over all paths, $P$ , and edge weights in each path to get", + "bbox": [ + 496, + 431, + 892, + 690 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A P S} \\left(\\mathbf {q} _ {j}\\right) = \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\left[ \\frac {1}{L} \\sum_ {l = 1} ^ {L} e _ {\\mathbf {q} _ {j}} (l, p) \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 698, + 890, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we calculate the logit sum (LS) score for a given concept, $\\mathbf{q}_j$ , by passing all of the corresponding masked segments, $\\mathbf{M}_{\\mathbf{q}_j}$ , through the model, $F(\\cdot)$ . We then sum up the logit scores for the target class, $c$ , to yield", + "bbox": [ + 498, + 748, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L S} \\left(\\mathbf {q} _ {j}\\right) = \\sum_ {i = 1} ^ {\\left| \\mathbf {M} _ {\\mathbf {q} _ {j}} \\right|} F \\left(\\mathbf {M} _ {\\mathbf {q} _ {j}} ^ {i}\\right) | _ {c}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 811, + 890, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{\\mathbf{q}_j}^i$ denotes the $i^{\\mathrm{th}}$ segment mask, and $F(\\cdot)|_c$ denotes the $c^{\\mathrm{th}}$ logit score. If the ITCAV edge weights are", + "bbox": [ + 498, + 866, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10900", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f38a7386107a106becddbfd0c0f63f015fa3ad24608e06b0c3628e99fb464a75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 92, + 272, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fe9fecc4656ffa6232ffdcaa02dc4264a94987c188c093bae92ff55bbc1bf81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 90, + 467, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec38988f507cf770b2e0c96757209e02957d90efb49d6a1316a0342d4db47f2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 166, + 272, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/945936437e3d7a3789c998f383588d8241ed663aaba0068ef7f8ade02c5315b8.jpg", + "image_caption": [ + "Figure 6. Graph metrics of all layer VCCs for three diverse architectures. Layer number normalized to allow for comparison of models with different numbers of layers." + ], + "image_footnote": [], + "bbox": [ + 282, + 165, + 464, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "meaningful, then there should be positive correlation between the APS and LS scores, i.e. $\\mathrm{APS}(\\mathbf{q})\\propto \\mathrm{LS}(\\mathbf{q})$", + "bbox": [ + 75, + 292, + 468, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3 (c, d) shows the LS scores plotted vs. the APS scores for VGG16 [25] and ViT-b [51]. A positive correlation is found between the APS and LS scores for both architectures. These results suggest that the combination of ITCAV scores is predictive of whether a concept is representative of the target class. The supplement has more LS vs. APS plots for more models that support these findings.", + "bbox": [ + 75, + 324, + 468, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Understanding models with VCCs", + "text_level": 1, + "bbox": [ + 76, + 439, + 375, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate VCC's unique ability to interpret models at different resolutions, we present model analyses in three parts: (i) qualitative analyses of VCCs generated for a subset of layers as well as at all layers, (ii) quantitative analysis of VCCs generated for a subset of four layers, (iii) a broadened quantitative analysis to encompass all model layers.", + "bbox": [ + 75, + 462, + 468, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualizing concept hierarchies. Figure 4 shows a three layer VCC for GoogLeNet [49] targeting the class \"Jay\". Notice how the inception5b 'bird' concepts $(c_{51}, c_{52})$ form as selective weighting of inception4c concepts background $(c_{41})$ , bird part $(c_{42}, c_{44})$ and tree branch $(c_{43})$ , while the inception5b background concepts $(c_{53}, c_{54})$ form differently from weighting of solely inception4c background part concepts (e.g. tree branch $(c_{43})$ and green leaves $(c_{41})$ ). Notably, the network separates subspecies of Jay in the final layer (e.g. Blue Jay $(c_{52})$ and other types $(c_{51}))$ . The concepts found in inception4c are composed from varying combinations of colors and parts found in conv3 (e.g. various bird parts $(c_{31}, c_{33})$ contribute to the bird concepts at inception4c). In the end, both scene and object contribute with strong weights to the final category.", + "bbox": [ + 75, + 553, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An all layer VGG-16 VCC visualization was presented in Fig. 1. As discussed in the caption, hierarchical concept assemblies again are revealed, with both target object as well as its background contributing to the final classification. While both the few layer and all layer visualizations reveal the concept representations of the models under analysis, they afford different levels of granularity: Few layer visualization provides a concise summary with a focus on", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2d87502e4cf58d753bf76746ecc560d10081152709067273814440e0b0ca221.jpg", + "image_caption": [ + "Figure 7. VCCs for a CNN (ResNet50) and transformer (MViT) for the class 'Tow Truck'. Shown are only the last two layers with the class output." + ], + "image_footnote": [], + "bbox": [ + 501, + 85, + 723, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82c317552b64a101232d7d886207adb3fb4315b72a3489b98040814d2c5335ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 87, + 890, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "specific layers, while all layer provides very detailed study.", + "bbox": [ + 498, + 308, + 888, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subset VCC analysis. We begin our quantitative analysis on four-layer VCCs generated for the CNNs and transformers listed in Sec. 4.1. Layers are selected approximately uniformly (see supplement for exact layers) across a model to capture concepts at different abstractions. We use standard metrics for analyzing tree-like data structures in a per-layer fashion: branching factor (extent to which the representation is distributed), number of nodes (how many concepts the model uses to encode a particular class), edge weights (strength of connection between concepts) and edge weight variance (variability of connection strength). We calculate averages over 50 VCCs for 50 randomly selected ImageNet classes. Results are shown in Fig. 5.", + "bbox": [ + 496, + 324, + 890, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Patterns are apparent in concept numbers and edge weights: more, but weaker, concepts in early layers vs. fewer, but stronger, concepts in later layers. These patterns reflect the shared low-level features (e.g. colors, textures) across classes and more specific features near the end, which yield larger ITCAV values. Also, CNNs show a decreasing branching factor, while transformers maintain a consistent number until the final layer, where all models converge to about two concepts, typically of an ImageNet class's foreground-background structure. Transformers have higher final layer edge weight variance compared to CNNs, indicating their ability to better differentiate earlier concepts' importance in forming later concepts, potentially explaining their superior classification performance (i.e. all information is not equally valuable). We also compare models (ResNet50 and VGG16) when trained with self-supervision [11] (SimCLR) or for adversarial robustness, i.e. on Stylized ImageNet [21] (ADV). We observe that robust and self-supervised models have fewer low-level concepts and compositionality than the originals, likely as their training yields less reliance on texture (stylization perturbs texture) and color (SimCLR training jitters color).", + "bbox": [ + 496, + 522, + 892, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To examine these patterns further, VCC visualizations for a CNN and transformer are shown in Fig. 7. Here, we limit to two (diverse) models and two later layers with class", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2136961e6f5b39aad7f0423052f19666731f87c8ad96e1697cdb4aa1961128a9.jpg", + "image_caption": [ + "Figure 8. Debugging model failure modes with VCCs. We show an image of a tricycle incorrectly predicted by a ResNet50 as a rickshaw (middle) as well as the top-down segmentation of the image (Sec. 3.1). We also show the incorrect (left) and correct (right) VCCs. Following the hierarchy of concepts reveals that the model incorrectly focused heavily on the collapsible hood, starting at Layer 2." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 370, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b42ad8a22bfe6577980ecb7fc52666132f1ed6b77372bd56051da4ae76e3acde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 87, + 614, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5db879e63f4b8af358d713e5108927d602eb635369029d7736be6310fef398f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 88, + 874, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "output for space. The connection diversity in the last layer is indeed observed to be larger for the transformer vs. the CNN. Notably, over half of the concepts in the CNN capture background centric concepts, while the transformer has only a single background centric concept.", + "bbox": [ + 75, + 339, + 468, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All layer VCCs. We present quantitative analyses on all layer VCCs for three diverse models: (i) a standard CNN, VGG-16, (ii) an efficient model, MobileNetv3 and (iii) a transformer, MViT. Averages are taken over 10 VCCs for 10 random ImageNet classes. Results are shown in Fig. 6. Common trends appear in all models. Concept composition is non-linear across layers, with branching factor ranging from 5-15 and converging to approximately two near the last layers. The peak number of concepts, around 20, is consistently at $30 - 40\\%$ of network depth and, as in the four-layer analysis, also converges to two in the final layer.", + "bbox": [ + 75, + 415, + 468, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Edge weights and variances are in accord with the main findings of the four-layer analysis, but also reveal other insights into the compositional nature of the models. At fine grained, all layer analysis, each model more readily displays unique edge weight characteristics: VGG16's average weights decrease in later layers, MobileNetv3's drop greatly before the final layer and MViT maintains consistent values. Still, overall these results indicate that penultimate concepts differ between CNNs vs. transformers, as supported by our four-layer VCC analysis; see Fig. 5. Higher variances in initial layers suggest a diverse combination of concepts, whereas deeper layers indicate a more uniform composition. Transformers, however, show a variance increase in the final layer, indicating greater compositionality.", + "bbox": [ + 75, + 583, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the four and all-layer VCC analyses consistently highlight key aspects of deep networks: (i) Early and mid-layers use more concepts compared to deeper layers. (ii) Concept interactions, e.g. ITCAV values and variances, differ greatly between adjacent layers, as opposed to across multiple layers. (iii) CNNs and transformers compose concepts differently, especially at the final layers.", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Application: Diagnosing failure predictions", + "text_level": 1, + "bbox": [ + 498, + 338, + 885, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To show the VCC's practical utility, we consider application to model failure analysis. Contrasting failure analysis methods that do not explain how errors are made compositionally [46] or only offer explanations of single neuron heatmaps [13], VCCs provide insights on compositional concepts across layers and distributed representations.", + "bbox": [ + 496, + 363, + 890, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 8 shows a 'Tricycle' incorrectly classified as a 'Rickshaw' by a ResNet50 model, and the corresponding incorrect VCC ('Rickshaw', left) and correct VCC ('Tricycle', right). As the image is decomposed using our top-down segmentation (Sec. 3.1), it is revealed that the majority of pooled segments are closer, in terms of $l_{2}$ distance, to concepts in the Rickshaw VCC (red outlines) than the tricycle VCC (green outlines). While the model correctly encoded the wheel and handlebar portions of the images as tricycle concepts, the background and collapsible hood concepts are composed from layers two through four as rickshaw concepts, which may cause the error. We also note the lack of other tricycle-specific concepts (e.g. children).", + "bbox": [ + 496, + 454, + 890, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and conclusion", + "text_level": 1, + "bbox": [ + 500, + 659, + 741, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Visual Concept Connectome (VCC) is a method for discovering human interpretable concepts and their interlayer connections in a fully unsupervised way. VCCs allow for the interpretation of deep models at multiple resolutions, revealing interesting properties via application to a range of networks. When comparing architecture types, CNNs were found to rely heavily on final layer concepts for prediction, while transformers were found to have more diverse connection strengths in the final layer and a more complex assembly of later concepts from earlier ones. We also applied VCCs to model debugging and showed how the hierarchical concept representation can indicate how mistakes arise. Given VCCs generality, our methodology should be applicable to additional networks, tasks and applications.", + "bbox": [ + 496, + 684, + 890, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10902", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From “where” to “what”: Towards human-understandable explanations through concept relevance propagation. arXiv preprint arXiv:2206.03208, 2022. 3", + "[2] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, 2018. 2", + "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Conference on Computer Vision and Pattern Recognition, pages 6541-6549, 2017. 2, 5, 1, 3", + "[4] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In European Conference on Computer Vision, pages 351-369. Springer, 2020. 2", + "[5] Alexander Binder, Gregoire Montavon, Sebastian Lapischkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning, pages 63-71. Springer, 2016. 2", + "[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2", + "[7] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Christopher Olah. Towards monoseismicity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2023. https://transformer-circuits.pub/2023/monoseismicfeatures/index.html. 3", + "[8] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018. 1", + "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In Winter Conference on Applications of Computer Vision, pages 839-847, 2018. 2", + "[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. Nature Machine Intelligence, 2(12):772-782, 2020. 2", + "[13] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain CNN via hierarchical decomposition. International Journal of Computer Vision, 131:1091-1105, 2023. 3, 8", + "[14] European Commission. Laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts. European Commission, 2021. 5", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 6, 1, 4", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6, 4, 9", + "[17] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/toy_model/index.html. 3", + "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. International Conference on Computer Vision, 2021. 5, 2, 3, 4, 8", + "[19] Christoph Feichtenhofer, Axel Pinz, Richard P Wildes, and Andrew Zisserman. Deep insights into convolutional networks for video recognition. International Journal of Computer Vision, 128:420-437, 2020. 1, 2", + "[20] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. CRAFT: Concept recursive activation factorization for explainability. In Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2, 3, 6", + "[21] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, 2018. 7", + "[22] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI Conference on Artificial Intelligence, pages 3681-3688, 2019. 2", + "[23] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems, 2019. 1, 2, 3, 4, 5, 6", + "[24] Sven Ove Hansson, Matts-Åke Belin, and Björn Lundgren. Self-driving vehicles-An ethical overview. Philosophy & Technology, 34:1383-1408, 2021. 1", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6, 7, 2, 4, 9", + "[26] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, and Vijay Vasudevan. Searching for MobilNetv3. In International Conference on Computer Vision, pages 1314-1324, 2019. 5, 3, 4, 8", + "[27] Haiyang Huang, Zhi Chen, and Cynthia Rudin. SegDiscover: Visual concept discovery via unsupervised semantic segmentation. arXiv preprint arXiv:2204.10926, 2022. 2", + "[28] Benjamin Irving. MaskSLIC: Regional superpixel generation with application to local pathology characterisation in medical images. arXiv preprint arXiv:1606.09518, 2016. 4, 1, 3", + "[29] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, and Fernanda Viegas. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In International Conference on Machine Learning, pages 2668-2677, 2018. 1, 2, 4, 5, 6", + "[30] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schutt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267-280, 2019. 2", + "[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348, 2020. 2", + "[32] Matthew Kowal, Achal Dave, Rares Ambrus, Adrien Gaidon, Konstantinos G Derpanis, and Pavel Tokmakov. Understanding video transformers via universal concept discovery. arXiv preprint arXiv:2401.10831, 2024. 1", + "[33] Satyapriya Krishna, Tessa Han, Alex Gu, Javin Pombra, Shahin Jabbari, Steven Wu, and Himabindu Lakkaraju. The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602, 2022. 6", + "[34] Mengchen Liu, Jiaxin Shi, Kelei Cao, Jun Zhu, and Shixia Liu. Analyzing the training processes of deep generative models. Transactions on Visualization and Computer Graphics, 24(1):77-87, 2017. 2", + "[35] Stuart Lloyd. Least squares quantization in PCM. Transactions on information theory, 28(2):129-137, 1982. 2", + "[36] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Conference on Computer Vision and Pattern Recognition, pages 5188-5196, 2015. 2", + "[37] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54 (6):1-35, 2021. 5", + "[38] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/features-visualization.1,2", + "[39] Chris Olah, Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, and Alexander Mordvintsev. The building blocks of interpretability. Distill, 2018. https://distill.pub/2018/building-blocks.1" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 2020. https://distill.pub/2020/circuits/zoom-in. 2, 3", + "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, and Luca Antiga. PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 2, 4", + "[43] Peter J Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 4, 5", + "[44] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetv2: Inverted residuals and linear bottlenecks. In Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2", + "[45] Anirban Sarkar, Deepak Vijaykeerthy, Anindya Sarkar, and Vineeth N Balasubramanian. A framework for learning antehoc explainable models via concepts. In Conference on Computer Vision and Pattern Recognition, pages 10286-10295, 2022. 2", + "[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision, pages 618-626, 2017. 1, 2, 8", + "[47] S. Seung. Connectome: How the Brain's Wiring Makes Us Who We Are. Houghton Mifflin Harcourt, 2012. 1", + "[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations, 2014. 5, 6, 2, 4", + "[49] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2, 6, 7, 4", + "[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. Scikit-image: image processing in python. PeerJ, 2:e453, 2014. 3", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 5, 7, 2, 3, 4", + "[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech UCSD birds 2011 dataset. Technical" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10904", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 4, 9", + "[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565, 2020. 2", + "[54] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pages 818-833, 2014. 1, 2", + "[55] Quanshi Zhang, Ruiming Cao, Feng Shi, Ying Nian Wu, and Song-Chun Zhu. Interpreting CNN knowledge via an explanatory graph. In AAAI Conference on Artificial Intelligence, pages 4454-4463, 2018. 2", + "[56] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, pages 8827-8836, 2018. 2", + "[57] Quanshi Zhang, Yu Yang, Haotian Ma, and Ying Nian Wu. Interpreting CNNs via decision trees. In Conference on Computer Vision and Pattern Recognition, pages 6261-6270, 2019. 2", + "[58] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for CNN models with non-negative concept activation vectors. In AAAI Conference on Artificial Intelligence, pages 11682-11690, 2021. 2", + "[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Conference on Computer Vision and Pattern Recognition, pages 2921-2929, 2016. 1, 2", + "[60] Bolei Zhou, Yiyou Sun, David Bau, and Antonio Torralba. Interpretable basis decomposition for visual explanation. In European Conference on Computer Vision, pages 119-134, 2018. 2" + ], + "bbox": [ + 78, + 90, + 467, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_model.json b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72471aeb61f249cdc8c0b4e0929c56482ea0db37 --- /dev/null +++ b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.88, + 0.175 + ], + "angle": 0, + "content": "Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.203, + 0.782, + 0.221 + ], + "angle": 0, + "content": "Matthew Kowal1,3 Richard P. Wildes1,2 Konstantinos G. Derpanis1,2,3" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.223, + 0.7, + 0.256 + ], + "angle": 0, + "content": "\\(^{1}\\)York University, \\(^{2}\\)Samsung AI Centre Toronto, \\(^{3}\\)Vector Institute yorkucvil.github.io/VCC" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.295, + 0.314, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.474, + 0.6 + ], + "angle": 0, + "content": "Understanding what deep network models capture in their learned representations is a fundamental challenge in computer vision. We present a new methodology to understanding such vision models, the Visual Concept Connectome (VCC), which discovers human interpretable concepts and their interlayer connections in a fully unsupervised manner. Our approach simultaneously reveals fine-grained concepts at a layer, connection weightings across all layers and is amendable to global analysis of network structure (e.g. branching pattern of hierarchical concept assemblies). Previous work yielded ways to extract interpretable concepts from single layers and examine their impact on classification, but did not afford multilayer concept analysis across an entire network architecture. Quantitative and qualitative empirical results show the effectiveness of VCCs in the domain of image classification. Also, we leverage VCCs for the application of failure mode debugging to reveal where mistakes arise in deep networks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.611, + 0.21, + 0.627 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.825 + ], + "angle": 0, + "content": "This paper focuses on interpreting the intermediate representations of deep networks for computer vision. The goal is understanding how various encoded concepts impact a model's prediction as well as concepts in other layers. We define concepts as abstractions that generalize from particular instances, including those defined locally (e.g. color and orientation), regionally (e.g. texture and shading) and from higher-level considerations (e.g. object parts, wholes and groupings). Human-interpretable concepts are of particular interest in increasing human understanding of models; however, extracting these concepts encoded in deep networks remains an open challenge in computer vision due to the complexity and opaque nature of these models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Understanding what concepts are learned by deep models and how they are encoded is important for both science and applications. For science, understanding what information drives a model's encoding of different concepts may indicate directions to advance performance. For applications," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.297, + 0.892, + 0.327 + ], + "angle": 0, + "content": "several negative consequences of deploying opaque vision models have been documented, e.g. [8, 24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.893, + 0.538 + ], + "angle": 0, + "content": "Previous work has focused on interpreting models via feature attribution, which measures the contribution of individual inputs to a model's output [10, 46, 59]; so, explanations are of single pixels and may be difficult to interpret. Other work generates images that maximize activation of a model's features [19, 38, 54]. Like feature attribution, these approaches are qualitative and place most of the burden on the user to determine the concepts revealed. Concept-based interpretability, which identifies human-interpretable abstractions in a model's latent space [20, 23, 29, 32], yield quantitative contributions of a concept to the model's output and explanations on the class-level (vs. pixel-level). These approaches are easier to understand and validate; however, they have not been used to explore interlayer relationships." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.719 + ], + "angle": 0, + "content": "As it stands, no approaches can quantify the interlayer affect of a given distributed concept at one layer, \\( l \\), to another concept at a different layer, \\( l' \\) (rather than the model output). Even though it is well established that deep networks learn to build concepts hierarchically as information flows through the network [39, 54], understanding the hierarchical representations has been under-researched. Indeed, little is known about the characteristics of this concept hierarchy for today's models. Questions abound: How many concepts exist in a network? What are the connections and weights between concepts? Does the model architecture impact the hierarchical structure of concept abstractions?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In response to these questions, we take inspiration from the biological notion of a connectome [47], defined as “a comprehensive structural description of the network of elements and connections forming a brain.” Analogously, we present the Visual Concept Connectome (VCC), a comprehensive structural description of a deep pretrained network model in terms of human-interpretable concepts and their relationships that form the internal representation maintained by the model; Fig. 1 shows an example. Notably, VCC generation works in the open-world setting, i.e. the concepts and interlayer connections are discovered without the need for any predefined concept dictionary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10895" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.856, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 1. A Visual Concept Connectome (VCC). At each layer the visual concepts learned by a deep model for a given class are revealed as are the learned interlayer concept connections. For each concept, up to four exemplars are shown as unmasked regions in a \\(2 \\times 2\\) image. Interlayer concept connections are shown as lines with darker lines indicating larger contributions. Shown is a VCC for every convolutional layer of a VGG16 model [49] trained on ImageNet [15] targeting recognition of class \"Tow Truck\". A closer visualization of VCC subgraphs reveals interesting compositions occurring at different levels of abstraction corresponding at different depths of the model. At early layers (bottom left), we observe oriented patterns (\\(C_{1,1}\\)) and brown color (\\(C_{1,2}\\)) composing the concept of green and brown orientation (\\(C_{2,1}\\)). Middle layers (right) show the concept of 'wheel on the road' (\\(C_{9,1}\\)) being composed of wheels (\\(C_{8,1}\\)) and regions of asphalt (\\(C_{8,2}\\)). The final layer concepts (top left) show that both foreground objects, e.g. tow trucks (\\(C_{13,1}\\)), and background regions, e.g. road, trees, humans, or car being towed (\\(C_{13,2}\\)), concepts highly influence the final category (\\(C_{TowTruck}\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.518, + 0.242, + 0.533 + ], + "angle": 0, + "content": "2. Related research" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.583 + ], + "angle": 0, + "content": "While a number of intrinsically interpretable networks have appeared (e.g. [6, 12, 31, 45, 56, 57]), we focus on work that, like ours, endeavours to interpret black-box models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Concept-based interpretability. Closed-world concept interpretability considers cases where a labelled dataset defines the concepts of interest for post-hoc analysis; approaches include Network Dissection [3] and Testing with Concept Activation Vectors (TCAV) [29]. However, a desirable property for concept-based interpretability is not to be restricted to a set of predefined classes (i.e. closed-world), but to support discovery of new and previously unlabelled concepts (i.e. open-world). Approaches to open-world concept interpretability include Automatic Concept Explanations (ACE) [23], SegDiscover [27], ConceptSHAP [53], Invertible Concept Explanations [58] and CRAFT [20]. These approaches are limited to single layer analysis and do not explain interlayer relationships." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Activation maximization [19, 36, 38, 54] visualizes the input that most activates a model component (e.g. filter, layer or logit) or how same layer units combine information [40]. These approaches do not capture interlayer relations and place a heavy burden on the user to interpret what concepts appear in their output where there may be no clear resemblance to natural concepts. Class Activation Maps" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.656 + ], + "angle": 0, + "content": "(CAMs) and extensions visualize an input image's local region that contributes to a model's output [9, 46, 59]. Layerwise Relevance Propagation [5] (LRP) generates pixel contribution heatmaps by assuming conservation of information is propagated through each neuron. These approaches interpret single images (not classes), are purely qualitative, and are sensitive to small input perturbations [22, 30], yet insensitive to model changes [2]. Less related are directions interpreting generative models, e.g. [4, 34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Interpretability via decomposition. Most closely related to our work are those that decompose a model's prediction into interpretable components. CRAFT [20] extracts concepts from the last layer of a convolutional neural network (CNN), but also searches earlier layers for sub-concepts for the maximally activating images of a given concept, but do not provide any way to quantify the contribution of sub-concepts to concepts. Another approach combined CAMs of multiple concepts to explain the final prediction [60]; however, it is limited to labelled data (so cannot discover concepts outside training) and only explains a final output, not intermediate relations. Other work interprets CNNs by distilling them into a graph [55]; however, it applies only to individual filters conceived as capturing object parts (i.e. not concepts other than objects and their parts, e.g. textures, colors or grouping), does not con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "sider the fact that representations are encoded via superposition [7, 17, 40] (i.e. more than one concept can be captured per-channel), does not yield meaningful edge weights and is only for CNNs. Yet other work extends Grad-CAM or LRP to perform intra-layer visualizations [1, 13]. While these approaches capture concepts other than object parts, they still ignore the core problem of superposition (and distributed representations) and do not yield interpretation for an entire class, just single images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Contributions. In the light of previous work, our contributions are fourfold. (i) We introduce and formalize the notion of a Visual Concept Connectome (VCC) for deep network models. The VCC reveals concepts represented at any given layer of a network as well as their interlayer connectivity. (ii) We present a method for extracting a VCC from any pretrained deep network in an unsupervised, open world setting, with a focus on classification. (iii) We validate our approach with quantitative and qualitative experiments wherein we examine various standard models to yield insights into model architectures and training tasks. (iv) We apply VCCs to explaining failure modes in deep models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.415, + 0.44 + ], + "angle": 0, + "content": "3. Visual Concept Connectomes (VCCs)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.599 + ], + "angle": 0, + "content": "A VCC is a directed acyclic graph, \\(\\mathbf{G}(\\mathbf{Q},\\mathbf{E})\\), created by distilling a pretrained deep network. The graph is topologically sorted based on the layer order found in the original network and has \\(n + 1\\) layers, consisting of \\(n\\) network layers to be analyzed (i.e. any subset of layers that a user selects, including all layers at the extreme) plus the final prediction (e.g. the object category for object recognition). The graph's nodes, \\(\\mathbf{Q}\\), are vectors (cluster centroids) representing interpretable concepts and its edges, \\(\\mathbf{E}\\), are scalars representing the contribution of one concept to the existence of another." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.734 + ], + "angle": 0, + "content": "To construct a VCC, three inputs are required: (i) a set of \\(I\\) images representative of a given task (e.g. exemplar images of an object category for object recognition), \\(\\mathcal{I} = \\{\\mathcal{I}^1,\\dots ,\\mathcal{I}^I\\}\\), \\(\\mathcal{I}^i\\in \\mathbb{R}^{h\\times w\\times c}\\), with \\(h,w\\) and \\(c\\) the image height, width and channel dimension (e.g. three for RGB), respectively, (ii) an \\(N\\) layer network, \\(F(\\cdot) = \\{f_{1}(\\cdot),\\ldots ,f_{N}(\\cdot)\\}\\), with \\(f_{j}\\) denoting feature extraction at layer \\(j\\), and (iii) a set of \\(n\\) selected layers (a subset of \\(F\\), possibly improper), which are to be studied." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "A VCC is constructed in three main steps. (i) Image segments are extracted in the model's feature space via divisive clustering to produce semantically meaningful image regions for each selected layer (Sec. 3.1). (ii) Layer-wise concepts, i.e. the nodes of the graph, are discovered in an open world fashion (i.e. no labelled data is required) via a second round of clustering over the dataset of image regions, independently for each layer (Sec. 3.2). (iii) Edges are calculated that indicate the contribution of concepts from earlier to deeper layers via an approach we introduce, Interlayer Testing with Concept Activation Vectors (ITCAVs)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "(Sec. 3.3). We use object recognition for the explanation of the approach; nevertheless, it could be extended to other tasks in a straightforward way (e.g. semantic segmentation). An overview of these steps is provided in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.16, + 0.768, + 0.176 + ], + "angle": 0, + "content": "3.1. Feature space image segments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.395 + ], + "angle": 0, + "content": "For each selected layer, we produce a set of image regions that plausibly belong to a concept encoded by the model at that layer. Previous work produces concept segments in RGB space, i.e. superpixels or random crops [20, 23]; however, segmentation based on features different from those over which concepts are to be discovered, i.e. the model's deep features, results in the support for the discovery process being divorced from the process that generated the support. Therefore, our segmentation uses the same deep features to be used subsequently for concept discovery. Our divisive clustering approach is presented visually in Fig. 2 (A). We pass the entire set of \\(I\\) input images, \\(\\mathcal{I}\\), for a given class through the model, \\(F\\), to get features, \\(\\mathbf{z}_n^i \\in \\mathbb{R}^{h_n \\times w_n \\times c_n}\\), at each selected layer, \\(n\\) according to" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.4, + 0.892, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {n} = \\left\\{f _ {n} \\left(\\mathcal {I} ^ {1}\\right), \\dots , f _ {n} \\left(\\mathcal {I} ^ {I}\\right) \\right\\} = \\left\\{\\mathbf {z} _ {n} ^ {1}, \\dots , \\mathbf {z} _ {n} ^ {I} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.517 + ], + "angle": 0, + "content": "To extract image segments at layer \\( n \\) for concept discovery (Sec. 3.2), we cluster the image activations, \\( \\mathbf{Z}_n \\), conditioned on the clusters (i.e. masks) generated at the next higher layer, \\( n + 1 \\). Let \\( \\mathbf{p} = (x,y) \\) index spatial coordinates and \\( \\mathbf{B}_n^i (\\mathbf{p};\\gamma) \\) be a binary mask for cluster \\( \\gamma \\) such that \\( \\mathbf{B}_n^i (\\mathbf{p};\\gamma) = 1 \\iff \\mathbf{z}_n^i (\\mathbf{p}) \\in \\gamma \\); otherwise, \\( \\mathbf{B}_n^i (\\mathbf{p};\\gamma) = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.592 + ], + "angle": 0, + "content": "We calculate a set of such masks by applying a clustering algorithm, \\( \\mathsf{C}_{\\Gamma}^{seg}(\\cdot) \\), that returns binary support masks for \\( \\Gamma \\) clusters on its argument. Let \\( \\Gamma_{n+1} \\) be the number of segments at layer \\( n+1 \\) and \\( 1 \\leq g \\leq \\Gamma_{n+1} \\) index a particular segment, \\( g \\). For each \\( g \\) we calculate segments at layer \\( n \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.597, + 0.892, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {B} _ {n} ^ {i} (\\mathbf {p}; \\gamma) \\right\\} _ {\\gamma = 1} ^ {\\Gamma} = \\mathrm {C} _ {\\Gamma} ^ {s e g} \\left(\\mathbf {z} _ {n} ^ {i} (\\mathbf {p}) \\odot \\widetilde {\\mathbf {B}} _ {n + 1} ^ {i} (\\mathbf {p}; g)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.753 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{B}}_{n + 1}^i\\) is \\(\\mathbf{B}_{n + 1}^i\\) upsampled to the resolution of layer \\(n\\) and \\(\\odot\\) indicates spatial element-wise masking applied to the individual image activations. The element-wise masking ensures clustering is done on the masked area of \\(\\mathbf{z}_n^i\\). Thus, the support of the concept discovery at layer \\(n\\) comes from the support of concepts at layer \\(n + 1\\). Notably, we let \\(\\Gamma\\) vary with \\(g\\); so, different segments, \\(g\\), at layer \\(n + 1\\) can yield a different number of segments, \\(\\Gamma\\), at layer \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.892, + 0.834 + ], + "angle": 0, + "content": "The top-down conditional clustering, (2), is repeated recursively for all image segments in all selected layers. To initiate the recursion, (2), at the top layer, \\( n_{top} \\), all the features are used by setting the number of clusters to be one, and having \\( \\widetilde{\\mathbf{B}}_{n_{top} + 1}^i (\\mathbf{p};1) = 1 \\) for all images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.834, + 0.892, + 0.879 + ], + "angle": 0, + "content": "At each layer, \\(n\\), we construct a set of RGB segmentations, \\(\\mathbf{M}\\), for use in concept discovery (Sec. 3.2) by applying upsampled masks, \\(\\mathbf{B}_n^i (\\mathbf{p};\\gamma)\\), to a given image, \\(\\mathcal{I}^i\\), via" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {n} ^ {i} (\\mathbf {p}; \\gamma) = \\mathcal {I} ^ {i} (\\mathbf {p}) \\odot \\mathbf {B} _ {n} ^ {i} (\\mathbf {p}; \\gamma), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.092, + 0.429, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.091, + 0.874, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Figure 2. The three steps in building a Visual Concept Connectome (VCC). (A) For a given image, \\(\\mathcal{I}^i\\in \\mathcal{I}\\), model, \\(F\\), and layer, \\(n\\), we produce a set of image segments, \\(\\mathbf{M}_n^i\\in \\mathbf{M}_n\\), based on a recursive spatial clustering, \\(\\mathsf{C}^{seg}\\) (2), of the features \\(\\mathbf{z}_n^i\\) conditioned on the clusters from the layer above, \\(n + 1\\). We then use (3) to generate a set of masked RGB image segments for each layer, \\(\\mathbf{M}_j\\). (B) For a given layer, \\(j\\), we pass the image segments from all images, \\(\\mathbf{M}_j\\), through \\(f_{j}\\) and cluster, \\(\\mathsf{C}^{con}\\) (4), these features across the dataset to produce \\(m_j\\) concept centroids, \\(\\{\\mathbf{q}_j^1,\\dots ,\\mathbf{q}_j^{m_j}\\}\\). (C) To measure the contribution of an earlier layer concept, \\(\\mathbf{q}_j^{m_j}\\), to a later layer concept, \\(\\mathbf{q}_l^{m_l}\\), we employ our Interlayer Testing with CAV (ITCAV) approach (Sec. 3.3), which uses the Concept Activation Vector (CAV) [29] of the earlier concept, \\(\\mathbf{V}_{\\mathfrak{q}_j^{m_j}}\\) (that points away from random examples, \\(\\mathcal{I}_{rnd}\\), but toward concept exemplars, \\(\\mathbf{M}_{\\mathfrak{q}_j^{m_j}}\\)), and the deeper layer concept, \\(\\mathbf{q}_l^{m_l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.469, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_n^i (\\mathbf{p};\\gamma)\\in \\{0,\\dots,255\\}^{h_n\\times w_n}\\), with \\(\\{0,\\dots,255\\}\\) specifying RGB value. We follow by defining \\(\\mathbf{M}_n\\) as the set of all RGB image segments at layer \\(n\\) (with each segment given in terms of (3)), and letting \\(\\mathbf{M} = \\{\\mathbf{M}_1,\\dots,\\mathbf{M}_n\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.573 + ], + "angle": 0, + "content": "We instantiate the clustering, \\( C^{seg} \\), via maskSLIC [28], an extension of k-means, with an \\( l_{2} \\) norm, that clusters features while respecting a mask and automate selection of the number of clusters via the silhouette method [43]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.583, + 0.338, + 0.599 + ], + "angle": 0, + "content": "3.2. Layer-wise concept discovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.771 + ], + "angle": 0, + "content": "Given the dataset of masked image segments for each layer, \\(\\mathbf{M} = \\{\\mathbf{M}_1,\\dots ,\\mathbf{M}_n\\}\\), we follow [23] and bilinearly interpolate the segments to the input image size and then pass the segments through the model to get pre-segmented activations \\(\\mathbf{Z}_{\\mathbf{M}_j} = f_j(\\mathbf{M}_j)\\) from the layer where the segments were found in the divisive clustering step (Sec. 3.1). We pass the image segments (as opposed to using the image activations) through the model to remove any contextual information before we discover concepts (the segments are set to zero everywhere outside of the segment boundary). Otherwise, \\(\\mathbf{M}_j\\) could mix information from outside the segment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.469, + 0.848 + ], + "angle": 0, + "content": "The layer-wise concept discovery step is presented visually in Fig. 2 (B). With the pre-segmented activations, \\(\\mathbf{Z}_{\\mathbf{M}_j}\\), calculated for all layers, \\(j\\), we cluster over the dataset of segment features for each individual layer to produce cluster centroids (i.e. concepts) according to" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.858, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {j} = \\left\\{\\mathbf {q} _ {j} ^ {1}, \\dots , \\mathbf {q} _ {j} ^ {m _ {j}} \\right\\} = \\mathrm {C} _ {m _ {j}} ^ {\\text {c o n}} (\\operatorname {G A P} \\left(\\mathbf {Z} _ {\\mathbf {M} _ {j}}\\right)), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(m_j\\) is the number of concepts discovered at layer \\(j\\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.593 + ], + "angle": 0, + "content": "GAP is spatial global average pooling and \\( \\mathsf{C}_{m_j}^{con} \\) is a clustering algorithm that returns \\( m_j \\) cluster centroids. Following previous work [23], we instantiate \\( \\mathsf{C}_{m_j}^{con} \\) in terms of standard k-means, using the \\( l_2 \\) norm, and then use a large number of clusters with subsequent pruning to remove noisy clusters; see the supplement for details. The output of this stage results in discovered concepts (i.e. cluster centroids and associated segments) for all layers as \\( \\mathbf{Q} = \\{\\mathbf{Q}_1, \\dots, \\mathbf{Q}_n\\} \\) for the centroids and \\( \\mathbf{M}_{\\mathbf{q}_j^{m_j}} \\) their associated RGB segments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.774, + 0.619 + ], + "angle": 0, + "content": "3.3. Interlayer concept connectivity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.762 + ], + "angle": 0, + "content": "Our approach to quantifying the contribution of discovered concepts between two layers, Interlayer Testing with Concept Activation Vectors (ITCAV), generalizes TCAV [29]. Specifically, while TCAV calculates the contribution of a single mid-layer concept to the class logit using a dataset of labelled concept images, we generalize to operate between any two layers and without a dataset of labelled concept images. Without loss of generality, we consider a single pair of concepts at two layers, which need not be adjacent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.905 + ], + "angle": 0, + "content": "The ITCAV method is presented visually in Fig. 2 (C). We seek to measure the degree to which a concept at an earlier layer contributes to a deeper layer. To do so, we construct two vectors: (i) a vector that represents the concept at the earlier layer, \\( j \\), and (ii) a vector that represents a positive contribution (i.e. gradient) of a concept at a later layer, \\( l \\). The closer the alignment of these two vectors, the larger the contribution of the concept at layer \\( j \\) to the concept at layer \\( l \\). Let \\( \\mathbf{q}_j^{m_j} \\) and \\( \\mathbf{M}_{\\mathbf{q}_j^{m_j}} \\) be the cluster centroid and as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.262, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.208, + 0.186, + 0.218 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.098, + 0.484, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.208, + 0.402, + 0.219 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.688, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.208, + 0.6, + 0.218 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.092, + 0.885, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.208, + 0.798, + 0.218 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 3. Validation of the three VCC method components. (a) Validation of segment proposals. The relative (Rel.) concept segment size compared to entire image for a given layer is plotted against the receptive field (RF) width/height of the same layer. (b) Validation of discovered concepts. For 50 randomly selected ImageNet classes, we discover concepts in four layers of the model. During inference, one randomly selected concept at each layer is suppressed by a factor of \\(\\epsilon\\). (c, d) Validation of interlayer concept weights. The unnormalized logit sum (LS) scores, (8), for the target class are plotted against the average path strength (APS) scores, (7). A positive correlation implies that the ITCAV edge weights connecting a concept to the class are predictive of the model output having a higher probability for that class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.469, + 0.36 + ], + "angle": 0, + "content": "sociated RGB image segments in layer \\(j\\), and let \\(\\mathbf{q}_l^{m_l}\\) and \\(\\mathbf{M}_{\\mathbf{q}_l^{m_l}}\\) be the cluster centroid and associated RGB image segments in a deeper layer \\(l\\), i.e. \\(l > j\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.509 + ], + "angle": 0, + "content": "We proceed by constructing the vector that represents the concept in the feature space of the earlier layer, \\( j \\). To do so, we employ the Concept Activation Vector (CAV) [29] for the earlier concept, \\( \\mathbf{q}_j^{m_j} \\), by training a linear classifier, \\( h(\\cdot) \\) on the features of layer \\( j \\) between positive concept inputs, \\( f_{j}(\\mathbf{M}_{\\mathbf{q}_{j}^{m_{j}}}) \\), and random images, \\( f_{j}(\\mathcal{I}_{rnd}) \\). The orthogonal vector to the hyperplane defined by \\( h(\\cdot) \\), denoted as \\( \\mathbf{V}_{\\mathbf{q}_j^{m_j}} \\) is the CAV for concept \\( \\mathbf{q}_j^{m_j} \\) and points in the direction of concept \\( \\mathbf{q}_j^{m_j} \\) in the feature space of layer \\( j \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Next, we construct the vector in the feature space of layer \\(j\\) that points in the direction of positive contribution of concept \\(\\mathbf{q}_l^{m_l}\\). To do so, we calculate \\(\\nabla f_l(f_j(\\cdot))\\), i.e. the gradient of the deeper concept, \\(\\mathbf{q}_l^{m_l}\\), at layer \\(l\\), with respect to the earlier layer, \\(j\\). Our approach to measuring the sensitivity between the two concepts is then" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.601, + 0.47, + 0.628 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} (x) = \\nabla \\left(\\left\\| f _ {l} \\left(f _ {j} (x)\\right) - \\mathbf {q} _ {l} ^ {m _ {l}} \\right\\| _ {2}\\right) \\cdot \\mathbf {V} _ {\\mathbf {q} _ {j} ^ {m _ {j}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.469, + 0.734 + ], + "angle": 0, + "content": "where \\(x \\in \\mathbf{M}_{\\mathbf{q}_l^{m_l}}\\). The sensitivity score should be positive because when the feature at layer \\(j\\) is perturbed in the direction of concept \\(\\mathbf{q}_j^{m_j}\\), i.e. in the direction \\(\\mathbf{V}_{\\mathbf{q}_j^{m_j}}\\), then the feature vector in deeper layer, \\(l\\), will be pushed closer to concept \\(\\mathbf{q}_l^{m_l}\\). Following TCAV [29], the final ITCAV edge weight, \\(e_{\\mathbf{q}_j^{m_j},\\mathbf{q}_l^{m_l}} \\in \\mathbf{E}\\), is the ratio of positive sensitivities," + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.736, + 0.47, + 0.76 + ], + "angle": 0, + "content": "\\[\ne _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} = | x \\in \\mathbf {M} _ {\\mathbf {q} _ {l} ^ {m _ {l}}}: S _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} (x) > 0 | / | \\mathbf {M} _ {\\mathbf {q} _ {l} ^ {m _ {l}}} |. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Key to our sensitivity score, (5), is the \\(l_{2}\\) distance between the output, \\(f_{l}(f_{j}(x))\\), and cluster centroid, \\(\\mathbf{q}_l^{m_l}\\), before taking the gradient, cf. CAV [29]. This choice serves two goals: (i) The subset of dimensions of individual image features, \\(f_{j}(x)\\), that are not well aligned with the cluster centroid, \\(\\mathbf{q}_l^{m_l}\\), will be penalized. This penalty will suppress the impact of noisy dimensions from individual samples on ITCAV. We use the \\(l_{2}\\) norm because the employed clustering approach used during concept discovery, k-means, uses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.893, + 0.359 + ], + "angle": 0, + "content": "\\(l_{2}\\); see Sec. 3.2. (ii) Collapsing the tensor to a scalar (i.e. a concept similarity score) makes the gradient calculation computationally feasible compared to the full Jacobian." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.368, + 0.634, + 0.385 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.392, + 0.715, + 0.409 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.644 + ], + "angle": 0, + "content": "For evaluation we use ResNet50 [25], VGG16 [48], MobileNetv3 [26], ViT-b [51] and MViT [18] to sample a broad range of popular architectures. The exact layers selected are in the supplement, but always include the last feature layer before the fully connected layer to probe concepts closest to the model output. Following previous work [23, 29], we perform a two-sided t-test on ITCAV scores for multiple runs of the same concept vs. different random sets of images to remove statistically insignificant scores and randomly sample from the Broden dataset [3] to get the images, \\(\\mathcal{I}_{rnd}\\) (Sec. 3.3), used in statistical testing and CAV training. On average, four and all-layer VCCs take 15 minutes and 36 hours, resp., to generate on an NVIDIA Quadro RTX 6000 GPU. The supplement has more details and VCCs for other layers, models and datasets [52]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.747, + 0.667 + ], + "angle": 0, + "content": "4.2. VCC component validation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Segment proposal validation. Our approach segments concepts based on the feature space of a given layer to allow us to capture concepts across multiple layers (Sec. 3.1). So, the spatial support of valid segments at a layer should follow the receptive field (RF) of filters at that layer, as the segments are constrained by the filters from which they are derived. For comparison, we consider ACE [23], which does segmentation with a layer independent feature (color), limited to the input image to discover concepts at a single layer for a single model. Being directly related to the features at a layer, our approach should produce concepts that better respect the RF at that layer compared to the baseline, whose features are layer independent (e.g. at CNN early layers the RF is small; so, intuitively the patch proposals should be smaller as well and conversely for the later layers)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10899" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Figure 4. A VCC for three selected layers of a GoogLeNet model [49] targeting recognition of class 'Jay'. Darker lines denote stronger connection weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.657 + ], + "angle": 0, + "content": "Figure 3 (a) shows the average concept size (in terms of the ratio of segment pixels to the entire image) vs. the RF at the corresponding layer for three different models: (i) ResNet50 [25], (ii) VGG16 [48] and (iii) ViT-b [16] (note we use mean attention distance instead of receptive field for ViT-b). While the baseline [23] produces segments with sizes invariant to the layer analyzed, our approach produces segments with image sizes that scale with the RF. Additional results with other models (including more transformers) are given in the supplement, which further support this observation. Overall, it is expected by design and we find that our segment proposals follow RFs at each layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Concept fidelity. Concept fidelity measures the meaningfulness of discovered concepts with respect to the target model. While other work focuses on single layer concept fidelity [20, 23, 29], we measure the effect on a model's output when suppressing the encoding of concepts as the information propagates through the model. The expectation is that suppressing the discovered concepts should result in a quicker decrease in performance compared to a non-concept direction (e.g. a randomly selected direction) as the amount of suppression gets larger. For a given model and category, we compute the concepts at various layers using our method (Sec. 3.2). We perform concept suppression for a given image, \\(\\mathcal{I}^i\\), at a given layer, \\(j\\), for concept \\(\\mathbf{q}_j\\), according to \\(\\mathbf{z}_j^i = \\mathbf{z}_j^i - \\epsilon \\cdot \\mathbf{q}_j / ||\\mathbf{q}_j||_2\\), where \\(\\epsilon\\) controls the degree of perturbation. If the concepts discovered are meaningful, then the accuracy for the model's target class should decrease faster" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.093, + 0.703, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.093, + 0.896, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.17, + 0.7, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.168, + 0.894, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Graph metrics on four layer VCCs comparing CNN vs. transformer architectures and training objectives." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.281, + 0.871, + 0.296 + ], + "angle": 0, + "content": "compared to random perturbations, \\(\\mathbf{q}_{rnd}\\), as \\(\\epsilon\\) increases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Figure 3 (b) has concept suppression results averaged over 50 randomly selected ImageNet [15] classes (documented in the supplement), where a concept is randomly chosen to be suppressed at each layer. Note that the \\(\\epsilon\\) values required to reduce the model accuracy to zero differs by model; so, we scale the \\(\\epsilon\\) values to [0, 1] for visualization purposes. For all models, it is seen that perturbing in the opposite direction of a recovered concept more severely impacts the accuracy than a random perturbation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.691 + ], + "angle": 0, + "content": "ITCAV validation. If the interlayer concept connection weights are meaningful, then the accumulated weights from earlier to later layer concepts should be correlated with their probability of predicting the target class. We use this intuition to validate our ITCAV approach to recovering interlayer concept weights as follows. We define the average path strength (APS), a scalar value between zero and one representing the average strength of the connection between a concept and the class logit. Given concept \\(\\mathbf{q}_j\\) at layer \\(j\\), let there be \\(L\\) layers in the VCC between the concept and the logit layer; thus, each path from the concept to the logit consists of \\(L\\) edge weights. We consider all possible paths in the VCC from this concept to the class logit. Let \\(e_{\\mathbf{q}_j}(l,p)\\) be the edge weight (i.e. ITCAV score) of the \\(p^{\\mathrm{th}}\\) path (from concept \\(\\mathbf{q}_j\\) to the class concept) for VCC layer \\(l\\). To calculate the APS score for a given concept, we average over all paths, \\(P\\), and edge weights in each path to get" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.699, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A P S} \\left(\\mathbf {q} _ {j}\\right) = \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\left[ \\frac {1}{L} \\sum_ {l = 1} ^ {L} e _ {\\mathbf {q} _ {j}} (l, p) \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.749, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Next, we calculate the logit sum (LS) score for a given concept, \\(\\mathbf{q}_j\\), by passing all of the corresponding masked segments, \\(\\mathbf{M}_{\\mathbf{q}_j}\\), through the model, \\(F(\\cdot)\\). We then sum up the logit scores for the target class, \\(c\\), to yield" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.813, + 0.892, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L S} \\left(\\mathbf {q} _ {j}\\right) = \\sum_ {i = 1} ^ {\\left| \\mathbf {M} _ {\\mathbf {q} _ {j}} \\right|} F \\left(\\mathbf {M} _ {\\mathbf {q} _ {j}} ^ {i}\\right) | _ {c}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.867, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{\\mathbf{q}_j}^i\\) denotes the \\(i^{\\mathrm{th}}\\) segment mask, and \\(F(\\cdot)|_c\\) denotes the \\(c^{\\mathrm{th}}\\) logit score. If the ITCAV edge weights are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10900" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.093, + 0.273, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.092, + 0.468, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.167, + 0.273, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.166, + 0.465, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.47, + 0.289 + ], + "angle": 0, + "content": "Figure 6. Graph metrics of all layer VCCs for three diverse architectures. Layer number normalized to allow for comparison of models with different numbers of layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.469, + 0.325 + ], + "angle": 0, + "content": "meaningful, then there should be positive correlation between the APS and LS scores, i.e. \\(\\mathrm{APS}(\\mathbf{q})\\propto \\mathrm{LS}(\\mathbf{q})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.469, + 0.43 + ], + "angle": 0, + "content": "Figure 3 (c, d) shows the LS scores plotted vs. the APS scores for VGG16 [25] and ViT-b [51]. A positive correlation is found between the APS and LS scores for both architectures. These results suggest that the combination of ITCAV scores is predictive of whether a concept is representative of the target class. The supplement has more LS vs. APS plots for more models that support these findings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.377, + 0.456 + ], + "angle": 0, + "content": "4.3. Understanding models with VCCs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.553 + ], + "angle": 0, + "content": "To demonstrate VCC's unique ability to interpret models at different resolutions, we present model analyses in three parts: (i) qualitative analyses of VCCs generated for a subset of layers as well as at all layers, (ii) quantitative analysis of VCCs generated for a subset of four layers, (iii) a broadened quantitative analysis to encompass all model layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Visualizing concept hierarchies. Figure 4 shows a three layer VCC for GoogLeNet [49] targeting the class \"Jay\". Notice how the inception5b 'bird' concepts \\((c_{51}, c_{52})\\) form as selective weighting of inception4c concepts background \\((c_{41})\\), bird part \\((c_{42}, c_{44})\\) and tree branch \\((c_{43})\\), while the inception5b background concepts \\((c_{53}, c_{54})\\) form differently from weighting of solely inception4c background part concepts (e.g. tree branch \\((c_{43})\\) and green leaves \\((c_{41})\\)). Notably, the network separates subspecies of Jay in the final layer (e.g. Blue Jay \\((c_{52})\\) and other types \\((c_{51}))\\). The concepts found in inception4c are composed from varying combinations of colors and parts found in conv3 (e.g. various bird parts \\((c_{31}, c_{33})\\) contribute to the bird concepts at inception4c). In the end, both scene and object contribute with strong weights to the final category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "An all layer VGG-16 VCC visualization was presented in Fig. 1. As discussed in the caption, hierarchical concept assemblies again are revealed, with both target object as well as its background contributing to the final classification. While both the few layer and all layer visualizations reveal the concept representations of the models under analysis, they afford different levels of granularity: Few layer visualization provides a concise summary with a focus on" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.087, + 0.725, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.088, + 0.892, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Figure 7. VCCs for a CNN (ResNet50) and transformer (MViT) for the class 'Tow Truck'. Shown are only the last two layers with the class output." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.309, + 0.889, + 0.324 + ], + "angle": 0, + "content": "specific layers, while all layer provides very detailed study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Subset VCC analysis. We begin our quantitative analysis on four-layer VCCs generated for the CNNs and transformers listed in Sec. 4.1. Layers are selected approximately uniformly (see supplement for exact layers) across a model to capture concepts at different abstractions. We use standard metrics for analyzing tree-like data structures in a per-layer fashion: branching factor (extent to which the representation is distributed), number of nodes (how many concepts the model uses to encode a particular class), edge weights (strength of connection between concepts) and edge weight variance (variability of connection strength). We calculate averages over 50 VCCs for 50 randomly selected ImageNet classes. Results are shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Patterns are apparent in concept numbers and edge weights: more, but weaker, concepts in early layers vs. fewer, but stronger, concepts in later layers. These patterns reflect the shared low-level features (e.g. colors, textures) across classes and more specific features near the end, which yield larger ITCAV values. Also, CNNs show a decreasing branching factor, while transformers maintain a consistent number until the final layer, where all models converge to about two concepts, typically of an ImageNet class's foreground-background structure. Transformers have higher final layer edge weight variance compared to CNNs, indicating their ability to better differentiate earlier concepts' importance in forming later concepts, potentially explaining their superior classification performance (i.e. all information is not equally valuable). We also compare models (ResNet50 and VGG16) when trained with self-supervision [11] (SimCLR) or for adversarial robustness, i.e. on Stylized ImageNet [21] (ADV). We observe that robust and self-supervised models have fewer low-level concepts and compositionality than the originals, likely as their training yields less reliance on texture (stylization perturbs texture) and color (SimCLR training jitters color)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To examine these patterns further, VCC visualizations for a CNN and transformer are shown in Fig. 7. Here, we limit to two (diverse) models and two later layers with class" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.372, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.088, + 0.616, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.089, + 0.875, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Figure 8. Debugging model failure modes with VCCs. We show an image of a tricycle incorrectly predicted by a ResNet50 as a rickshaw (middle) as well as the top-down segmentation of the image (Sec. 3.1). We also show the incorrect (left) and correct (right) VCCs. Following the hierarchy of concepts reveals that the model incorrectly focused heavily on the collapsible hood, starting at Layer 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.416 + ], + "angle": 0, + "content": "output for space. The connection diversity in the last layer is indeed observed to be larger for the transformer vs. the CNN. Notably, over half of the concepts in the CNN capture background centric concepts, while the transformer has only a single background centric concept." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.583 + ], + "angle": 0, + "content": "All layer VCCs. We present quantitative analyses on all layer VCCs for three diverse models: (i) a standard CNN, VGG-16, (ii) an efficient model, MobileNetv3 and (iii) a transformer, MViT. Averages are taken over 10 VCCs for 10 random ImageNet classes. Results are shown in Fig. 6. Common trends appear in all models. Concept composition is non-linear across layers, with branching factor ranging from 5-15 and converging to approximately two near the last layers. The peak number of concepts, around 20, is consistently at \\(30 - 40\\%\\) of network depth and, as in the four-layer analysis, also converges to two in the final layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Edge weights and variances are in accord with the main findings of the four-layer analysis, but also reveal other insights into the compositional nature of the models. At fine grained, all layer analysis, each model more readily displays unique edge weight characteristics: VGG16's average weights decrease in later layers, MobileNetv3's drop greatly before the final layer and MViT maintains consistent values. Still, overall these results indicate that penultimate concepts differ between CNNs vs. transformers, as supported by our four-layer VCC analysis; see Fig. 5. Higher variances in initial layers suggest a diverse combination of concepts, whereas deeper layers indicate a more uniform composition. Transformers, however, show a variance increase in the final layer, indicating greater compositionality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In summary, the four and all-layer VCC analyses consistently highlight key aspects of deep networks: (i) Early and mid-layers use more concepts compared to deeper layers. (ii) Concept interactions, e.g. ITCAV values and variances, differ greatly between adjacent layers, as opposed to across multiple layers. (iii) CNNs and transformers compose concepts differently, especially at the final layers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.339, + 0.886, + 0.357 + ], + "angle": 0, + "content": "5. Application: Diagnosing failure predictions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.454 + ], + "angle": 0, + "content": "To show the VCC's practical utility, we consider application to model failure analysis. Contrasting failure analysis methods that do not explain how errors are made compositionally [46] or only offer explanations of single neuron heatmaps [13], VCCs provide insights on compositional concepts across layers and distributed representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 8 shows a 'Tricycle' incorrectly classified as a 'Rickshaw' by a ResNet50 model, and the corresponding incorrect VCC ('Rickshaw', left) and correct VCC ('Tricycle', right). As the image is decomposed using our top-down segmentation (Sec. 3.1), it is revealed that the majority of pooled segments are closer, in terms of \\( l_{2} \\) distance, to concepts in the Rickshaw VCC (red outlines) than the tricycle VCC (green outlines). While the model correctly encoded the wheel and handlebar portions of the images as tricycle concepts, the background and collapsible hood concepts are composed from layers two through four as rickshaw concepts, which may cause the error. We also note the lack of other tricycle-specific concepts (e.g. children)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.742, + 0.675 + ], + "angle": 0, + "content": "6. Discussion and conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.896 + ], + "angle": 0, + "content": "The Visual Concept Connectome (VCC) is a method for discovering human interpretable concepts and their interlayer connections in a fully unsupervised way. VCCs allow for the interpretation of deep models at multiple resolutions, revealing interesting properties via application to a range of networks. When comparing architecture types, CNNs were found to rely heavily on final layer concepts for prediction, while transformers were found to have more diverse connection strengths in the final layer and a more complex assembly of later concepts from earlier ones. We also applied VCCs to model debugging and showed how the hierarchical concept representation can indicate how mistakes arise. Given VCCs generality, our methodology should be applicable to additional networks, tasks and applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10902" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From “where” to “what”: Towards human-understandable explanations through concept relevance propagation. arXiv preprint arXiv:2206.03208, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Conference on Computer Vision and Pattern Recognition, pages 6541-6549, 2017. 2, 5, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[4] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In European Conference on Computer Vision, pages 351-369. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[5] Alexander Binder, Gregoire Montavon, Sebastian Lapischkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning, pages 63-71. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[7] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Christopher Olah. Towards monoseismicity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2023. https://transformer-circuits.pub/2023/monoseismicfeatures/index.html. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[8] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.719, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In Winter Conference on Applications of Computer Vision, pages 839-847, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[12] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. Nature Machine Intelligence, 2(12):772-782, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[13] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain CNN via hierarchical decomposition. International Journal of Computer Vision, 131:1091-1105, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[14] European Commission. Laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts. European Commission, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 6, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[17] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/toy_model/index.html. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. International Conference on Computer Vision, 2021. 5, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Axel Pinz, Richard P Wildes, and Andrew Zisserman. Deep insights into convolutional networks for video recognition. International Journal of Computer Vision, 128:420-437, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[20] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. CRAFT: Concept recursive activation factorization for explainability. In Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[21] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[22] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI Conference on Artificial Intelligence, pages 3681-3688, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[23] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems, 2019. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Sven Ove Hansson, Matts-Åke Belin, and Björn Lundgren. Self-driving vehicles-An ethical overview. Philosophy & Technology, 34:1383-1408, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10903" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6, 7, 2, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, and Vijay Vasudevan. Searching for MobilNetv3. In International Conference on Computer Vision, pages 1314-1324, 2019. 5, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[27] Haiyang Huang, Zhi Chen, and Cynthia Rudin. SegDiscover: Visual concept discovery via unsupervised semantic segmentation. arXiv preprint arXiv:2204.10926, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[28] Benjamin Irving. MaskSLIC: Regional superpixel generation with application to local pathology characterisation in medical images. arXiv preprint arXiv:1606.09518, 2016. 4, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[29] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, and Fernanda Viegas. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In International Conference on Machine Learning, pages 2668-2677, 2018. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[30] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schutt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267-280, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[32] Matthew Kowal, Achal Dave, Rares Ambrus, Adrien Gaidon, Konstantinos G Derpanis, and Pavel Tokmakov. Understanding video transformers via universal concept discovery. arXiv preprint arXiv:2401.10831, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[33] Satyapriya Krishna, Tessa Han, Alex Gu, Javin Pombra, Shahin Jabbari, Steven Wu, and Himabindu Lakkaraju. The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[34] Mengchen Liu, Jiaxin Shi, Kelei Cao, Jun Zhu, and Shixia Liu. Analyzing the training processes of deep generative models. Transactions on Visualization and Computer Graphics, 24(1):77-87, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[35] Stuart Lloyd. Least squares quantization in PCM. Transactions on information theory, 28(2):129-137, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[36] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Conference on Computer Vision and Pattern Recognition, pages 5188-5196, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[37] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54 (6):1-35, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[38] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/features-visualization.1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Chris Olah, Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, and Alexander Mordvintsev. The building blocks of interpretability. Distill, 2018. https://distill.pub/2018/building-blocks.1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[40] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 2020. https://distill.pub/2020/circuits/zoom-in. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, and Luca Antiga. PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[43] Peter J Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.389, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[44] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetv2: Inverted residuals and linear bottlenecks. In Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[45] Anirban Sarkar, Deepak Vijaykeerthy, Anindya Sarkar, and Vineeth N Balasubramanian. A framework for learning antehoc explainable models via concepts. In Conference on Computer Vision and Pattern Recognition, pages 10286-10295, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision, pages 618-626, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[47] S. Seung. Connectome: How the Brain's Wiring Makes Us Who We Are. Houghton Mifflin Harcourt, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations, 2014. 5, 6, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[49] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. Scikit-image: image processing in python. PeerJ, 2:e453, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 5, 7, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech UCSD birds 2011 dataset. Technical" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10904" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[54] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pages 818-833, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[55] Quanshi Zhang, Ruiming Cao, Feng Shi, Ying Nian Wu, and Song-Chun Zhu. Interpreting CNN knowledge via an explanatory graph. In AAAI Conference on Artificial Intelligence, pages 4454-4463, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[56] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, pages 8827-8836, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[57] Quanshi Zhang, Yu Yang, Haotian Ma, and Ying Nian Wu. Interpreting CNNs via decision trees. In Conference on Computer Vision and Pattern Recognition, pages 6261-6270, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[58] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for CNN models with non-negative concept activation vectors. In AAAI Conference on Artificial Intelligence, pages 11682-11690, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Conference on Computer Vision and Pattern Recognition, pages 2921-2929, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.468, + 0.585 + ], + "angle": 0, + "content": "[60] Bolei Zhou, Yiyou Sun, David Bau, and Antonio Torralba. Interpretable basis decomposition for visual explanation. In European Conference on Computer Vision, pages 119-134, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10905" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_origin.pdf b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6cab577a85d4dc4fc77b06a8013d710feeb50a51 --- /dev/null +++ b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/1a616d0d-3730-44ab-a630-ecdfca09f3c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a26b63fd0706c63f6d7f4abcbab78dc098a47cd326e3daa246f61289e8a1c961 +size 2242451 diff --git a/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/full.md b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b5f35ccf109e8686481099ce69968464bfbb546d --- /dev/null +++ b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/full.md @@ -0,0 +1,312 @@ +# Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models + +Matthew Kowal1,3 Richard P. Wildes1,2 Konstantinos G. Derpanis1,2,3 + +$^{1}$ York University, $^{2}$ Samsung AI Centre Toronto, $^{3}$ Vector Institute yorkucvil.github.io/VCC + +# Abstract + +Understanding what deep network models capture in their learned representations is a fundamental challenge in computer vision. We present a new methodology to understanding such vision models, the Visual Concept Connectome (VCC), which discovers human interpretable concepts and their interlayer connections in a fully unsupervised manner. Our approach simultaneously reveals fine-grained concepts at a layer, connection weightings across all layers and is amendable to global analysis of network structure (e.g. branching pattern of hierarchical concept assemblies). Previous work yielded ways to extract interpretable concepts from single layers and examine their impact on classification, but did not afford multilayer concept analysis across an entire network architecture. Quantitative and qualitative empirical results show the effectiveness of VCCs in the domain of image classification. Also, we leverage VCCs for the application of failure mode debugging to reveal where mistakes arise in deep networks. + +# 1. Introduction + +This paper focuses on interpreting the intermediate representations of deep networks for computer vision. The goal is understanding how various encoded concepts impact a model's prediction as well as concepts in other layers. We define concepts as abstractions that generalize from particular instances, including those defined locally (e.g. color and orientation), regionally (e.g. texture and shading) and from higher-level considerations (e.g. object parts, wholes and groupings). Human-interpretable concepts are of particular interest in increasing human understanding of models; however, extracting these concepts encoded in deep networks remains an open challenge in computer vision due to the complexity and opaque nature of these models. + +Understanding what concepts are learned by deep models and how they are encoded is important for both science and applications. For science, understanding what information drives a model's encoding of different concepts may indicate directions to advance performance. For applications, + +several negative consequences of deploying opaque vision models have been documented, e.g. [8, 24]. + +Previous work has focused on interpreting models via feature attribution, which measures the contribution of individual inputs to a model's output [10, 46, 59]; so, explanations are of single pixels and may be difficult to interpret. Other work generates images that maximize activation of a model's features [19, 38, 54]. Like feature attribution, these approaches are qualitative and place most of the burden on the user to determine the concepts revealed. Concept-based interpretability, which identifies human-interpretable abstractions in a model's latent space [20, 23, 29, 32], yield quantitative contributions of a concept to the model's output and explanations on the class-level (vs. pixel-level). These approaches are easier to understand and validate; however, they have not been used to explore interlayer relationships. + +As it stands, no approaches can quantify the interlayer affect of a given distributed concept at one layer, $l$ , to another concept at a different layer, $l'$ (rather than the model output). Even though it is well established that deep networks learn to build concepts hierarchically as information flows through the network [39, 54], understanding the hierarchical representations has been under-researched. Indeed, little is known about the characteristics of this concept hierarchy for today's models. Questions abound: How many concepts exist in a network? What are the connections and weights between concepts? Does the model architecture impact the hierarchical structure of concept abstractions? + +In response to these questions, we take inspiration from the biological notion of a connectome [47], defined as “a comprehensive structural description of the network of elements and connections forming a brain.” Analogously, we present the Visual Concept Connectome (VCC), a comprehensive structural description of a deep pretrained network model in terms of human-interpretable concepts and their relationships that form the internal representation maintained by the model; Fig. 1 shows an example. Notably, VCC generation works in the open-world setting, i.e. the concepts and interlayer connections are discovered without the need for any predefined concept dictionary. + +![](images/1d5f6b084c6bcf97c82706cceb1a60f6fabd7772e1316cc51488b53170d1037b.jpg) +Figure 1. A Visual Concept Connectome (VCC). At each layer the visual concepts learned by a deep model for a given class are revealed as are the learned interlayer concept connections. For each concept, up to four exemplars are shown as unmasked regions in a $2 \times 2$ image. Interlayer concept connections are shown as lines with darker lines indicating larger contributions. Shown is a VCC for every convolutional layer of a VGG16 model [49] trained on ImageNet [15] targeting recognition of class "Tow Truck". A closer visualization of VCC subgraphs reveals interesting compositions occurring at different levels of abstraction corresponding at different depths of the model. At early layers (bottom left), we observe oriented patterns ( $C_{1,1}$ ) and brown color ( $C_{1,2}$ ) composing the concept of green and brown orientation ( $C_{2,1}$ ). Middle layers (right) show the concept of 'wheel on the road' ( $C_{9,1}$ ) being composed of wheels ( $C_{8,1}$ ) and regions of asphalt ( $C_{8,2}$ ). The final layer concepts (top left) show that both foreground objects, e.g. tow trucks ( $C_{13,1}$ ), and background regions, e.g. road, trees, humans, or car being towed ( $C_{13,2}$ ), concepts highly influence the final category ( $C_{TowTruck}$ ). + +# 2. Related research + +While a number of intrinsically interpretable networks have appeared (e.g. [6, 12, 31, 45, 56, 57]), we focus on work that, like ours, endeavours to interpret black-box models. + +Concept-based interpretability. Closed-world concept interpretability considers cases where a labelled dataset defines the concepts of interest for post-hoc analysis; approaches include Network Dissection [3] and Testing with Concept Activation Vectors (TCAV) [29]. However, a desirable property for concept-based interpretability is not to be restricted to a set of predefined classes (i.e. closed-world), but to support discovery of new and previously unlabelled concepts (i.e. open-world). Approaches to open-world concept interpretability include Automatic Concept Explanations (ACE) [23], SegDiscover [27], ConceptSHAP [53], Invertible Concept Explanations [58] and CRAFT [20]. These approaches are limited to single layer analysis and do not explain interlayer relationships. + +Activation maximization [19, 36, 38, 54] visualizes the input that most activates a model component (e.g. filter, layer or logit) or how same layer units combine information [40]. These approaches do not capture interlayer relations and place a heavy burden on the user to interpret what concepts appear in their output where there may be no clear resemblance to natural concepts. Class Activation Maps + +(CAMs) and extensions visualize an input image's local region that contributes to a model's output [9, 46, 59]. Layerwise Relevance Propagation [5] (LRP) generates pixel contribution heatmaps by assuming conservation of information is propagated through each neuron. These approaches interpret single images (not classes), are purely qualitative, and are sensitive to small input perturbations [22, 30], yet insensitive to model changes [2]. Less related are directions interpreting generative models, e.g. [4, 34]. + +Interpretability via decomposition. Most closely related to our work are those that decompose a model's prediction into interpretable components. CRAFT [20] extracts concepts from the last layer of a convolutional neural network (CNN), but also searches earlier layers for sub-concepts for the maximally activating images of a given concept, but do not provide any way to quantify the contribution of sub-concepts to concepts. Another approach combined CAMs of multiple concepts to explain the final prediction [60]; however, it is limited to labelled data (so cannot discover concepts outside training) and only explains a final output, not intermediate relations. Other work interprets CNNs by distilling them into a graph [55]; however, it applies only to individual filters conceived as capturing object parts (i.e. not concepts other than objects and their parts, e.g. textures, colors or grouping), does not con + +sider the fact that representations are encoded via superposition [7, 17, 40] (i.e. more than one concept can be captured per-channel), does not yield meaningful edge weights and is only for CNNs. Yet other work extends Grad-CAM or LRP to perform intra-layer visualizations [1, 13]. While these approaches capture concepts other than object parts, they still ignore the core problem of superposition (and distributed representations) and do not yield interpretation for an entire class, just single images. + +Contributions. In the light of previous work, our contributions are fourfold. (i) We introduce and formalize the notion of a Visual Concept Connectome (VCC) for deep network models. The VCC reveals concepts represented at any given layer of a network as well as their interlayer connectivity. (ii) We present a method for extracting a VCC from any pretrained deep network in an unsupervised, open world setting, with a focus on classification. (iii) We validate our approach with quantitative and qualitative experiments wherein we examine various standard models to yield insights into model architectures and training tasks. (iv) We apply VCCs to explaining failure modes in deep models. + +# 3. Visual Concept Connectomes (VCCs) + +A VCC is a directed acyclic graph, $\mathbf{G}(\mathbf{Q},\mathbf{E})$ , created by distilling a pretrained deep network. The graph is topologically sorted based on the layer order found in the original network and has $n + 1$ layers, consisting of $n$ network layers to be analyzed (i.e. any subset of layers that a user selects, including all layers at the extreme) plus the final prediction (e.g. the object category for object recognition). The graph's nodes, $\mathbf{Q}$ , are vectors (cluster centroids) representing interpretable concepts and its edges, $\mathbf{E}$ , are scalars representing the contribution of one concept to the existence of another. + +To construct a VCC, three inputs are required: (i) a set of $I$ images representative of a given task (e.g. exemplar images of an object category for object recognition), $\mathcal{I} = \{\mathcal{I}^1,\dots ,\mathcal{I}^I\}$ , $\mathcal{I}^i\in \mathbb{R}^{h\times w\times c}$ , with $h,w$ and $c$ the image height, width and channel dimension (e.g. three for RGB), respectively, (ii) an $N$ layer network, $F(\cdot) = \{f_{1}(\cdot),\ldots ,f_{N}(\cdot)\}$ , with $f_{j}$ denoting feature extraction at layer $j$ , and (iii) a set of $n$ selected layers (a subset of $F$ , possibly improper), which are to be studied. + +A VCC is constructed in three main steps. (i) Image segments are extracted in the model's feature space via divisive clustering to produce semantically meaningful image regions for each selected layer (Sec. 3.1). (ii) Layer-wise concepts, i.e. the nodes of the graph, are discovered in an open world fashion (i.e. no labelled data is required) via a second round of clustering over the dataset of image regions, independently for each layer (Sec. 3.2). (iii) Edges are calculated that indicate the contribution of concepts from earlier to deeper layers via an approach we introduce, Interlayer Testing with Concept Activation Vectors (ITCAVs) + +(Sec. 3.3). We use object recognition for the explanation of the approach; nevertheless, it could be extended to other tasks in a straightforward way (e.g. semantic segmentation). An overview of these steps is provided in Fig. 2. + +# 3.1. Feature space image segments + +For each selected layer, we produce a set of image regions that plausibly belong to a concept encoded by the model at that layer. Previous work produces concept segments in RGB space, i.e. superpixels or random crops [20, 23]; however, segmentation based on features different from those over which concepts are to be discovered, i.e. the model's deep features, results in the support for the discovery process being divorced from the process that generated the support. Therefore, our segmentation uses the same deep features to be used subsequently for concept discovery. Our divisive clustering approach is presented visually in Fig. 2 (A). We pass the entire set of $I$ input images, $\mathcal{I}$ , for a given class through the model, $F$ , to get features, $\mathbf{z}_n^i \in \mathbb{R}^{h_n \times w_n \times c_n}$ , at each selected layer, $n$ according to + +$$ +\mathbf {Z} _ {n} = \left\{f _ {n} \left(\mathcal {I} ^ {1}\right), \dots , f _ {n} \left(\mathcal {I} ^ {I}\right) \right\} = \left\{\mathbf {z} _ {n} ^ {1}, \dots , \mathbf {z} _ {n} ^ {I} \right\}. \tag {1} +$$ + +To extract image segments at layer $n$ for concept discovery (Sec. 3.2), we cluster the image activations, $\mathbf{Z}_n$ , conditioned on the clusters (i.e. masks) generated at the next higher layer, $n + 1$ . Let $\mathbf{p} = (x,y)$ index spatial coordinates and $\mathbf{B}_n^i (\mathbf{p};\gamma)$ be a binary mask for cluster $\gamma$ such that $\mathbf{B}_n^i (\mathbf{p};\gamma) = 1 \iff \mathbf{z}_n^i (\mathbf{p}) \in \gamma$ ; otherwise, $\mathbf{B}_n^i (\mathbf{p};\gamma) = 0$ . + +We calculate a set of such masks by applying a clustering algorithm, $\mathsf{C}_{\Gamma}^{seg}(\cdot)$ , that returns binary support masks for $\Gamma$ clusters on its argument. Let $\Gamma_{n+1}$ be the number of segments at layer $n+1$ and $1 \leq g \leq \Gamma_{n+1}$ index a particular segment, $g$ . For each $g$ we calculate segments at layer $n$ as + +$$ +\left\{\mathbf {B} _ {n} ^ {i} (\mathbf {p}; \gamma) \right\} _ {\gamma = 1} ^ {\Gamma} = \mathrm {C} _ {\Gamma} ^ {s e g} \left(\mathbf {z} _ {n} ^ {i} (\mathbf {p}) \odot \widetilde {\mathbf {B}} _ {n + 1} ^ {i} (\mathbf {p}; g)\right), \tag {2} +$$ + +where $\widetilde{\mathbf{B}}_{n + 1}^i$ is $\mathbf{B}_{n + 1}^i$ upsampled to the resolution of layer $n$ and $\odot$ indicates spatial element-wise masking applied to the individual image activations. The element-wise masking ensures clustering is done on the masked area of $\mathbf{z}_n^i$ . Thus, the support of the concept discovery at layer $n$ comes from the support of concepts at layer $n + 1$ . Notably, we let $\Gamma$ vary with $g$ ; so, different segments, $g$ , at layer $n + 1$ can yield a different number of segments, $\Gamma$ , at layer $n$ . + +The top-down conditional clustering, (2), is repeated recursively for all image segments in all selected layers. To initiate the recursion, (2), at the top layer, $n_{top}$ , all the features are used by setting the number of clusters to be one, and having $\widetilde{\mathbf{B}}_{n_{top} + 1}^i (\mathbf{p};1) = 1$ for all images. + +At each layer, $n$ , we construct a set of RGB segmentations, $\mathbf{M}$ , for use in concept discovery (Sec. 3.2) by applying upsampled masks, $\mathbf{B}_n^i (\mathbf{p};\gamma)$ , to a given image, $\mathcal{I}^i$ , via + +$$ +\mathbf {M} _ {n} ^ {i} (\mathbf {p}; \gamma) = \mathcal {I} ^ {i} (\mathbf {p}) \odot \mathbf {B} _ {n} ^ {i} (\mathbf {p}; \gamma), \tag {3} +$$ + +![](images/457a2c85cd74cfe2897fc892925150a793916aa62f396961d5e3ba0e856b1358.jpg) +Figure 2. The three steps in building a Visual Concept Connectome (VCC). (A) For a given image, $\mathcal{I}^i\in \mathcal{I}$ , model, $F$ , and layer, $n$ , we produce a set of image segments, $\mathbf{M}_n^i\in \mathbf{M}_n$ , based on a recursive spatial clustering, $\mathsf{C}^{seg}$ (2), of the features $\mathbf{z}_n^i$ conditioned on the clusters from the layer above, $n + 1$ . We then use (3) to generate a set of masked RGB image segments for each layer, $\mathbf{M}_j$ . (B) For a given layer, $j$ , we pass the image segments from all images, $\mathbf{M}_j$ , through $f_{j}$ and cluster, $\mathsf{C}^{con}$ (4), these features across the dataset to produce $m_j$ concept centroids, $\{\mathbf{q}_j^1,\dots ,\mathbf{q}_j^{m_j}\}$ . (C) To measure the contribution of an earlier layer concept, $\mathbf{q}_j^{m_j}$ , to a later layer concept, $\mathbf{q}_l^{m_l}$ , we employ our Interlayer Testing with CAV (ITCAV) approach (Sec. 3.3), which uses the Concept Activation Vector (CAV) [29] of the earlier concept, $\mathbf{V}_{\mathfrak{q}_j^{m_j}}$ (that points away from random examples, $\mathcal{I}_{rnd}$ , but toward concept exemplars, $\mathbf{M}_{\mathfrak{q}_j^{m_j}}$ ), and the deeper layer concept, $\mathbf{q}_l^{m_l}$ . + +![](images/5f2d71779ddf2a157a18eb67605389de8bb55a4b57b2b8f47ecafc83038d903c.jpg) + +where $\mathbf{M}_n^i (\mathbf{p};\gamma)\in \{0,\dots,255\}^{h_n\times w_n}$ , with $\{0,\dots,255\}$ specifying RGB value. We follow by defining $\mathbf{M}_n$ as the set of all RGB image segments at layer $n$ (with each segment given in terms of (3)), and letting $\mathbf{M} = \{\mathbf{M}_1,\dots,\mathbf{M}_n\}$ . + +We instantiate the clustering, $C^{seg}$ , via maskSLIC [28], an extension of k-means, with an $l_{2}$ norm, that clusters features while respecting a mask and automate selection of the number of clusters via the silhouette method [43]. + +# 3.2. Layer-wise concept discovery + +Given the dataset of masked image segments for each layer, $\mathbf{M} = \{\mathbf{M}_1,\dots ,\mathbf{M}_n\}$ , we follow [23] and bilinearly interpolate the segments to the input image size and then pass the segments through the model to get pre-segmented activations $\mathbf{Z}_{\mathbf{M}_j} = f_j(\mathbf{M}_j)$ from the layer where the segments were found in the divisive clustering step (Sec. 3.1). We pass the image segments (as opposed to using the image activations) through the model to remove any contextual information before we discover concepts (the segments are set to zero everywhere outside of the segment boundary). Otherwise, $\mathbf{M}_j$ could mix information from outside the segment. + +The layer-wise concept discovery step is presented visually in Fig. 2 (B). With the pre-segmented activations, $\mathbf{Z}_{\mathbf{M}_j}$ , calculated for all layers, $j$ , we cluster over the dataset of segment features for each individual layer to produce cluster centroids (i.e. concepts) according to + +$$ +\mathbf {Q} _ {j} = \left\{\mathbf {q} _ {j} ^ {1}, \dots , \mathbf {q} _ {j} ^ {m _ {j}} \right\} = \mathrm {C} _ {m _ {j}} ^ {\text {c o n}} (\operatorname {G A P} \left(\mathbf {Z} _ {\mathbf {M} _ {j}}\right)), \tag {4} +$$ + +where $m_j$ is the number of concepts discovered at layer $j$ , + +GAP is spatial global average pooling and $\mathsf{C}_{m_j}^{con}$ is a clustering algorithm that returns $m_j$ cluster centroids. Following previous work [23], we instantiate $\mathsf{C}_{m_j}^{con}$ in terms of standard k-means, using the $l_2$ norm, and then use a large number of clusters with subsequent pruning to remove noisy clusters; see the supplement for details. The output of this stage results in discovered concepts (i.e. cluster centroids and associated segments) for all layers as $\mathbf{Q} = \{\mathbf{Q}_1, \dots, \mathbf{Q}_n\}$ for the centroids and $\mathbf{M}_{\mathbf{q}_j^{m_j}}$ their associated RGB segments. + +# 3.3. Interlayer concept connectivity + +Our approach to quantifying the contribution of discovered concepts between two layers, Interlayer Testing with Concept Activation Vectors (ITCAV), generalizes TCAV [29]. Specifically, while TCAV calculates the contribution of a single mid-layer concept to the class logit using a dataset of labelled concept images, we generalize to operate between any two layers and without a dataset of labelled concept images. Without loss of generality, we consider a single pair of concepts at two layers, which need not be adjacent. + +The ITCAV method is presented visually in Fig. 2 (C). We seek to measure the degree to which a concept at an earlier layer contributes to a deeper layer. To do so, we construct two vectors: (i) a vector that represents the concept at the earlier layer, $j$ , and (ii) a vector that represents a positive contribution (i.e. gradient) of a concept at a later layer, $l$ . The closer the alignment of these two vectors, the larger the contribution of the concept at layer $j$ to the concept at layer $l$ . Let $\mathbf{q}_j^{m_j}$ and $\mathbf{M}_{\mathbf{q}_j^{m_j}}$ be the cluster centroid and as + +![](images/5881e4c22dd331a2fa5610ade43f18d73597099d7caabd81203b32ddb60b19c3.jpg) +(a) + +![](images/92390b4b63fbd6b7de8210b4b6988778a1cbaa23eba2cb61d1b41d10143c7c06.jpg) +(b) + +![](images/f2efe2dfd7064929fbf47a9aa2501d613deb5aba2f5c35f32c81a1dcf570ab1b.jpg) +(c) + +![](images/6ad322a960174b95501f00da33b42799d03fe1798b0309ac4bb33968636edf53.jpg) +(d) +Figure 3. Validation of the three VCC method components. (a) Validation of segment proposals. The relative (Rel.) concept segment size compared to entire image for a given layer is plotted against the receptive field (RF) width/height of the same layer. (b) Validation of discovered concepts. For 50 randomly selected ImageNet classes, we discover concepts in four layers of the model. During inference, one randomly selected concept at each layer is suppressed by a factor of $\epsilon$ . (c, d) Validation of interlayer concept weights. The unnormalized logit sum (LS) scores, (8), for the target class are plotted against the average path strength (APS) scores, (7). A positive correlation implies that the ITCAV edge weights connecting a concept to the class are predictive of the model output having a higher probability for that class. + +sociated RGB image segments in layer $j$ , and let $\mathbf{q}_l^{m_l}$ and $\mathbf{M}_{\mathbf{q}_l^{m_l}}$ be the cluster centroid and associated RGB image segments in a deeper layer $l$ , i.e. $l > j$ . + +We proceed by constructing the vector that represents the concept in the feature space of the earlier layer, $j$ . To do so, we employ the Concept Activation Vector (CAV) [29] for the earlier concept, $\mathbf{q}_j^{m_j}$ , by training a linear classifier, $h(\cdot)$ on the features of layer $j$ between positive concept inputs, $f_{j}(\mathbf{M}_{\mathbf{q}_{j}^{m_{j}}})$ , and random images, $f_{j}(\mathcal{I}_{rnd})$ . The orthogonal vector to the hyperplane defined by $h(\cdot)$ , denoted as $\mathbf{V}_{\mathbf{q}_j^{m_j}}$ is the CAV for concept $\mathbf{q}_j^{m_j}$ and points in the direction of concept $\mathbf{q}_j^{m_j}$ in the feature space of layer $j$ . + +Next, we construct the vector in the feature space of layer $j$ that points in the direction of positive contribution of concept $\mathbf{q}_l^{m_l}$ . To do so, we calculate $\nabla f_l(f_j(\cdot))$ , i.e. the gradient of the deeper concept, $\mathbf{q}_l^{m_l}$ , at layer $l$ , with respect to the earlier layer, $j$ . Our approach to measuring the sensitivity between the two concepts is then + +$$ +S _ {\mathbf {q} _ {j} ^ {m _ {j}}, \mathbf {q} _ {l} ^ {m _ {l}}} (x) = \nabla \left(\left\| f _ {l} \left(f _ {j} (x)\right) - \mathbf {q} _ {l} ^ {m _ {l}} \right\| _ {2}\right) \cdot \mathbf {V} _ {\mathbf {q} _ {j} ^ {m _ {j}}}, \tag {5} +$$ + +where $x \in \mathbf{M}_{\mathbf{q}_l^{m_l}}$ . The sensitivity score should be positive because when the feature at layer $j$ is perturbed in the direction of concept $\mathbf{q}_j^{m_j}$ , i.e. in the direction $\mathbf{V}_{\mathbf{q}_j^{m_j}}$ , then the feature vector in deeper layer, $l$ , will be pushed closer to concept $\mathbf{q}_l^{m_l}$ . Following TCAV [29], the final ITCAV edge weight, $e_{\mathbf{q}_j^{m_j},\mathbf{q}_l^{m_l}} \in \mathbf{E}$ , is the ratio of positive sensitivities, + +$$ +e _ {\mathbf {q} _ {j} ^ {m _ {j}}, \mathbf {q} _ {l} ^ {m _ {l}}} = | x \in \mathbf {M} _ {\mathbf {q} _ {l} ^ {m _ {l}}}: S _ {\mathbf {q} _ {j} ^ {m _ {j}}, \mathbf {q} _ {l} ^ {m _ {l}}} (x) > 0 | / | \mathbf {M} _ {\mathbf {q} _ {l} ^ {m _ {l}}} |. \tag {6} +$$ + +Key to our sensitivity score, (5), is the $l_{2}$ distance between the output, $f_{l}(f_{j}(x))$ , and cluster centroid, $\mathbf{q}_l^{m_l}$ , before taking the gradient, cf. CAV [29]. This choice serves two goals: (i) The subset of dimensions of individual image features, $f_{j}(x)$ , that are not well aligned with the cluster centroid, $\mathbf{q}_l^{m_l}$ , will be penalized. This penalty will suppress the impact of noisy dimensions from individual samples on ITCAV. We use the $l_{2}$ norm because the employed clustering approach used during concept discovery, k-means, uses + +$l_{2}$ ; see Sec. 3.2. (ii) Collapsing the tensor to a scalar (i.e. a concept similarity score) makes the gradient calculation computationally feasible compared to the full Jacobian. + +# 4. Experiments + +# 4.1. Implementation details + +For evaluation we use ResNet50 [25], VGG16 [48], MobileNetv3 [26], ViT-b [51] and MViT [18] to sample a broad range of popular architectures. The exact layers selected are in the supplement, but always include the last feature layer before the fully connected layer to probe concepts closest to the model output. Following previous work [23, 29], we perform a two-sided t-test on ITCAV scores for multiple runs of the same concept vs. different random sets of images to remove statistically insignificant scores and randomly sample from the Broden dataset [3] to get the images, $\mathcal{I}_{rnd}$ (Sec. 3.3), used in statistical testing and CAV training. On average, four and all-layer VCCs take 15 minutes and 36 hours, resp., to generate on an NVIDIA Quadro RTX 6000 GPU. The supplement has more details and VCCs for other layers, models and datasets [52]. + +# 4.2. VCC component validation + +Segment proposal validation. Our approach segments concepts based on the feature space of a given layer to allow us to capture concepts across multiple layers (Sec. 3.1). So, the spatial support of valid segments at a layer should follow the receptive field (RF) of filters at that layer, as the segments are constrained by the filters from which they are derived. For comparison, we consider ACE [23], which does segmentation with a layer independent feature (color), limited to the input image to discover concepts at a single layer for a single model. Being directly related to the features at a layer, our approach should produce concepts that better respect the RF at that layer compared to the baseline, whose features are layer independent (e.g. at CNN early layers the RF is small; so, intuitively the patch proposals should be smaller as well and conversely for the later layers). + +![](images/76f1e05871175636312fd26837cec36445a232f6a6369bbf3f7cbd8513d7d1e1.jpg) +Figure 4. A VCC for three selected layers of a GoogLeNet model [49] targeting recognition of class 'Jay'. Darker lines denote stronger connection weights. + +Figure 3 (a) shows the average concept size (in terms of the ratio of segment pixels to the entire image) vs. the RF at the corresponding layer for three different models: (i) ResNet50 [25], (ii) VGG16 [48] and (iii) ViT-b [16] (note we use mean attention distance instead of receptive field for ViT-b). While the baseline [23] produces segments with sizes invariant to the layer analyzed, our approach produces segments with image sizes that scale with the RF. Additional results with other models (including more transformers) are given in the supplement, which further support this observation. Overall, it is expected by design and we find that our segment proposals follow RFs at each layer. + +Concept fidelity. Concept fidelity measures the meaningfulness of discovered concepts with respect to the target model. While other work focuses on single layer concept fidelity [20, 23, 29], we measure the effect on a model's output when suppressing the encoding of concepts as the information propagates through the model. The expectation is that suppressing the discovered concepts should result in a quicker decrease in performance compared to a non-concept direction (e.g. a randomly selected direction) as the amount of suppression gets larger. For a given model and category, we compute the concepts at various layers using our method (Sec. 3.2). We perform concept suppression for a given image, $\mathcal{I}^i$ , at a given layer, $j$ , for concept $\mathbf{q}_j$ , according to $\mathbf{z}_j^i = \mathbf{z}_j^i - \epsilon \cdot \mathbf{q}_j / ||\mathbf{q}_j||_2$ , where $\epsilon$ controls the degree of perturbation. If the concepts discovered are meaningful, then the accuracy for the model's target class should decrease faster + +![](images/bc44278740832248dd9bdd782de3e20e213c70c052d9e739efad4ffddfb990f4.jpg) + +![](images/c30cc6e647fdc7cd8c27d2617d17a6aa47114bcae5e1cea57208ee5f651219d8.jpg) + +![](images/cff0b91a0307d8f889552b5c476100f36baf0122cf0ac0bca4c45dd26b679c51.jpg) +Figure 5. Graph metrics on four layer VCCs comparing CNN vs. transformer architectures and training objectives. + +![](images/34ebd885aa6c5d2ee5c2b43d285ff062691e7f2c137d2376e9fb78d93dd3af9b.jpg) + +compared to random perturbations, $\mathbf{q}_{rnd}$ , as $\epsilon$ increases. + +Figure 3 (b) has concept suppression results averaged over 50 randomly selected ImageNet [15] classes (documented in the supplement), where a concept is randomly chosen to be suppressed at each layer. Note that the $\epsilon$ values required to reduce the model accuracy to zero differs by model; so, we scale the $\epsilon$ values to [0, 1] for visualization purposes. For all models, it is seen that perturbing in the opposite direction of a recovered concept more severely impacts the accuracy than a random perturbation. + +ITCAV validation. If the interlayer concept connection weights are meaningful, then the accumulated weights from earlier to later layer concepts should be correlated with their probability of predicting the target class. We use this intuition to validate our ITCAV approach to recovering interlayer concept weights as follows. We define the average path strength (APS), a scalar value between zero and one representing the average strength of the connection between a concept and the class logit. Given concept $\mathbf{q}_j$ at layer $j$ , let there be $L$ layers in the VCC between the concept and the logit layer; thus, each path from the concept to the logit consists of $L$ edge weights. We consider all possible paths in the VCC from this concept to the class logit. Let $e_{\mathbf{q}_j}(l,p)$ be the edge weight (i.e. ITCAV score) of the $p^{\mathrm{th}}$ path (from concept $\mathbf{q}_j$ to the class concept) for VCC layer $l$ . To calculate the APS score for a given concept, we average over all paths, $P$ , and edge weights in each path to get + +$$ +\operatorname {A P S} \left(\mathbf {q} _ {j}\right) = \frac {1}{P} \sum_ {p = 1} ^ {P} \left[ \frac {1}{L} \sum_ {l = 1} ^ {L} e _ {\mathbf {q} _ {j}} (l, p) \right]. \tag {7} +$$ + +Next, we calculate the logit sum (LS) score for a given concept, $\mathbf{q}_j$ , by passing all of the corresponding masked segments, $\mathbf{M}_{\mathbf{q}_j}$ , through the model, $F(\cdot)$ . We then sum up the logit scores for the target class, $c$ , to yield + +$$ +\operatorname {L S} \left(\mathbf {q} _ {j}\right) = \sum_ {i = 1} ^ {\left| \mathbf {M} _ {\mathbf {q} _ {j}} \right|} F \left(\mathbf {M} _ {\mathbf {q} _ {j}} ^ {i}\right) | _ {c}, \tag {8} +$$ + +where $\mathbf{M}_{\mathbf{q}_j}^i$ denotes the $i^{\mathrm{th}}$ segment mask, and $F(\cdot)|_c$ denotes the $c^{\mathrm{th}}$ logit score. If the ITCAV edge weights are + +![](images/f38a7386107a106becddbfd0c0f63f015fa3ad24608e06b0c3628e99fb464a75.jpg) + +![](images/3fe9fecc4656ffa6232ffdcaa02dc4264a94987c188c093bae92ff55bbc1bf81.jpg) + +![](images/ec38988f507cf770b2e0c96757209e02957d90efb49d6a1316a0342d4db47f2f.jpg) + +![](images/945936437e3d7a3789c998f383588d8241ed663aaba0068ef7f8ade02c5315b8.jpg) +Figure 6. Graph metrics of all layer VCCs for three diverse architectures. Layer number normalized to allow for comparison of models with different numbers of layers. + +meaningful, then there should be positive correlation between the APS and LS scores, i.e. $\mathrm{APS}(\mathbf{q})\propto \mathrm{LS}(\mathbf{q})$ + +Figure 3 (c, d) shows the LS scores plotted vs. the APS scores for VGG16 [25] and ViT-b [51]. A positive correlation is found between the APS and LS scores for both architectures. These results suggest that the combination of ITCAV scores is predictive of whether a concept is representative of the target class. The supplement has more LS vs. APS plots for more models that support these findings. + +# 4.3. Understanding models with VCCs + +To demonstrate VCC's unique ability to interpret models at different resolutions, we present model analyses in three parts: (i) qualitative analyses of VCCs generated for a subset of layers as well as at all layers, (ii) quantitative analysis of VCCs generated for a subset of four layers, (iii) a broadened quantitative analysis to encompass all model layers. + +Visualizing concept hierarchies. Figure 4 shows a three layer VCC for GoogLeNet [49] targeting the class "Jay". Notice how the inception5b 'bird' concepts $(c_{51}, c_{52})$ form as selective weighting of inception4c concepts background $(c_{41})$ , bird part $(c_{42}, c_{44})$ and tree branch $(c_{43})$ , while the inception5b background concepts $(c_{53}, c_{54})$ form differently from weighting of solely inception4c background part concepts (e.g. tree branch $(c_{43})$ and green leaves $(c_{41})$ ). Notably, the network separates subspecies of Jay in the final layer (e.g. Blue Jay $(c_{52})$ and other types $(c_{51}))$ . The concepts found in inception4c are composed from varying combinations of colors and parts found in conv3 (e.g. various bird parts $(c_{31}, c_{33})$ contribute to the bird concepts at inception4c). In the end, both scene and object contribute with strong weights to the final category. + +An all layer VGG-16 VCC visualization was presented in Fig. 1. As discussed in the caption, hierarchical concept assemblies again are revealed, with both target object as well as its background contributing to the final classification. While both the few layer and all layer visualizations reveal the concept representations of the models under analysis, they afford different levels of granularity: Few layer visualization provides a concise summary with a focus on + +![](images/a2d87502e4cf58d753bf76746ecc560d10081152709067273814440e0b0ca221.jpg) +Figure 7. VCCs for a CNN (ResNet50) and transformer (MViT) for the class 'Tow Truck'. Shown are only the last two layers with the class output. + +![](images/82c317552b64a101232d7d886207adb3fb4315b72a3489b98040814d2c5335ef.jpg) + +specific layers, while all layer provides very detailed study. + +Subset VCC analysis. We begin our quantitative analysis on four-layer VCCs generated for the CNNs and transformers listed in Sec. 4.1. Layers are selected approximately uniformly (see supplement for exact layers) across a model to capture concepts at different abstractions. We use standard metrics for analyzing tree-like data structures in a per-layer fashion: branching factor (extent to which the representation is distributed), number of nodes (how many concepts the model uses to encode a particular class), edge weights (strength of connection between concepts) and edge weight variance (variability of connection strength). We calculate averages over 50 VCCs for 50 randomly selected ImageNet classes. Results are shown in Fig. 5. + +Patterns are apparent in concept numbers and edge weights: more, but weaker, concepts in early layers vs. fewer, but stronger, concepts in later layers. These patterns reflect the shared low-level features (e.g. colors, textures) across classes and more specific features near the end, which yield larger ITCAV values. Also, CNNs show a decreasing branching factor, while transformers maintain a consistent number until the final layer, where all models converge to about two concepts, typically of an ImageNet class's foreground-background structure. Transformers have higher final layer edge weight variance compared to CNNs, indicating their ability to better differentiate earlier concepts' importance in forming later concepts, potentially explaining their superior classification performance (i.e. all information is not equally valuable). We also compare models (ResNet50 and VGG16) when trained with self-supervision [11] (SimCLR) or for adversarial robustness, i.e. on Stylized ImageNet [21] (ADV). We observe that robust and self-supervised models have fewer low-level concepts and compositionality than the originals, likely as their training yields less reliance on texture (stylization perturbs texture) and color (SimCLR training jitters color). + +To examine these patterns further, VCC visualizations for a CNN and transformer are shown in Fig. 7. Here, we limit to two (diverse) models and two later layers with class + +![](images/2136961e6f5b39aad7f0423052f19666731f87c8ad96e1697cdb4aa1961128a9.jpg) +Figure 8. Debugging model failure modes with VCCs. We show an image of a tricycle incorrectly predicted by a ResNet50 as a rickshaw (middle) as well as the top-down segmentation of the image (Sec. 3.1). We also show the incorrect (left) and correct (right) VCCs. Following the hierarchy of concepts reveals that the model incorrectly focused heavily on the collapsible hood, starting at Layer 2. + +![](images/b42ad8a22bfe6577980ecb7fc52666132f1ed6b77372bd56051da4ae76e3acde.jpg) + +![](images/5db879e63f4b8af358d713e5108927d602eb635369029d7736be6310fef398f7.jpg) + +output for space. The connection diversity in the last layer is indeed observed to be larger for the transformer vs. the CNN. Notably, over half of the concepts in the CNN capture background centric concepts, while the transformer has only a single background centric concept. + +All layer VCCs. We present quantitative analyses on all layer VCCs for three diverse models: (i) a standard CNN, VGG-16, (ii) an efficient model, MobileNetv3 and (iii) a transformer, MViT. Averages are taken over 10 VCCs for 10 random ImageNet classes. Results are shown in Fig. 6. Common trends appear in all models. Concept composition is non-linear across layers, with branching factor ranging from 5-15 and converging to approximately two near the last layers. The peak number of concepts, around 20, is consistently at $30 - 40\%$ of network depth and, as in the four-layer analysis, also converges to two in the final layer. + +Edge weights and variances are in accord with the main findings of the four-layer analysis, but also reveal other insights into the compositional nature of the models. At fine grained, all layer analysis, each model more readily displays unique edge weight characteristics: VGG16's average weights decrease in later layers, MobileNetv3's drop greatly before the final layer and MViT maintains consistent values. Still, overall these results indicate that penultimate concepts differ between CNNs vs. transformers, as supported by our four-layer VCC analysis; see Fig. 5. Higher variances in initial layers suggest a diverse combination of concepts, whereas deeper layers indicate a more uniform composition. Transformers, however, show a variance increase in the final layer, indicating greater compositionality. + +In summary, the four and all-layer VCC analyses consistently highlight key aspects of deep networks: (i) Early and mid-layers use more concepts compared to deeper layers. (ii) Concept interactions, e.g. ITCAV values and variances, differ greatly between adjacent layers, as opposed to across multiple layers. (iii) CNNs and transformers compose concepts differently, especially at the final layers. + +# 5. Application: Diagnosing failure predictions + +To show the VCC's practical utility, we consider application to model failure analysis. Contrasting failure analysis methods that do not explain how errors are made compositionally [46] or only offer explanations of single neuron heatmaps [13], VCCs provide insights on compositional concepts across layers and distributed representations. + +Figure 8 shows a 'Tricycle' incorrectly classified as a 'Rickshaw' by a ResNet50 model, and the corresponding incorrect VCC ('Rickshaw', left) and correct VCC ('Tricycle', right). As the image is decomposed using our top-down segmentation (Sec. 3.1), it is revealed that the majority of pooled segments are closer, in terms of $l_{2}$ distance, to concepts in the Rickshaw VCC (red outlines) than the tricycle VCC (green outlines). While the model correctly encoded the wheel and handlebar portions of the images as tricycle concepts, the background and collapsible hood concepts are composed from layers two through four as rickshaw concepts, which may cause the error. We also note the lack of other tricycle-specific concepts (e.g. children). + +# 6. Discussion and conclusion + +The Visual Concept Connectome (VCC) is a method for discovering human interpretable concepts and their interlayer connections in a fully unsupervised way. VCCs allow for the interpretation of deep models at multiple resolutions, revealing interesting properties via application to a range of networks. When comparing architecture types, CNNs were found to rely heavily on final layer concepts for prediction, while transformers were found to have more diverse connection strengths in the final layer and a more complex assembly of later concepts from earlier ones. We also applied VCCs to model debugging and showed how the hierarchical concept representation can indicate how mistakes arise. Given VCCs generality, our methodology should be applicable to additional networks, tasks and applications. + +# References + +[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From “where” to “what”: Towards human-understandable explanations through concept relevance propagation. arXiv preprint arXiv:2206.03208, 2022. 3 +[2] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, 2018. 2 +[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Conference on Computer Vision and Pattern Recognition, pages 6541-6549, 2017. 2, 5, 1, 3 +[4] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In European Conference on Computer Vision, pages 351-369. Springer, 2020. 2 +[5] Alexander Binder, Gregoire Montavon, Sebastian Lapischkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning, pages 63-71. Springer, 2016. 2 +[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2 +[7] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Christopher Olah. Towards monoseismicity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2023. https://transformer-circuits.pub/2023/monoseismicfeatures/index.html. 3 +[8] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018. 1 +[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In Winter Conference on Applications of Computer Vision, pages 839-847, 2018. 2 +[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7 + +[12] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. Nature Machine Intelligence, 2(12):772-782, 2020. 2 +[13] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain CNN via hierarchical decomposition. International Journal of Computer Vision, 131:1091-1105, 2023. 3, 8 +[14] European Commission. Laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts. European Commission, 2021. 5 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 6, 1, 4 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6, 4, 9 +[17] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/toy_model/index.html. 3 +[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. International Conference on Computer Vision, 2021. 5, 2, 3, 4, 8 +[19] Christoph Feichtenhofer, Axel Pinz, Richard P Wildes, and Andrew Zisserman. Deep insights into convolutional networks for video recognition. International Journal of Computer Vision, 128:420-437, 2020. 1, 2 +[20] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. CRAFT: Concept recursive activation factorization for explainability. In Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2, 3, 6 +[21] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, 2018. 7 +[22] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI Conference on Artificial Intelligence, pages 3681-3688, 2019. 2 +[23] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems, 2019. 1, 2, 3, 4, 5, 6 +[24] Sven Ove Hansson, Matts-Åke Belin, and Björn Lundgren. Self-driving vehicles-An ethical overview. Philosophy & Technology, 34:1383-1408, 2021. 1 +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference + +on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6, 7, 2, 4, 9 +[26] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, and Vijay Vasudevan. Searching for MobilNetv3. In International Conference on Computer Vision, pages 1314-1324, 2019. 5, 3, 4, 8 +[27] Haiyang Huang, Zhi Chen, and Cynthia Rudin. SegDiscover: Visual concept discovery via unsupervised semantic segmentation. arXiv preprint arXiv:2204.10926, 2022. 2 +[28] Benjamin Irving. MaskSLIC: Regional superpixel generation with application to local pathology characterisation in medical images. arXiv preprint arXiv:1606.09518, 2016. 4, 1, 3 +[29] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, and Fernanda Viegas. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In International Conference on Machine Learning, pages 2668-2677, 2018. 1, 2, 4, 5, 6 +[30] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schutt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267-280, 2019. 2 +[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348, 2020. 2 +[32] Matthew Kowal, Achal Dave, Rares Ambrus, Adrien Gaidon, Konstantinos G Derpanis, and Pavel Tokmakov. Understanding video transformers via universal concept discovery. arXiv preprint arXiv:2401.10831, 2024. 1 +[33] Satyapriya Krishna, Tessa Han, Alex Gu, Javin Pombra, Shahin Jabbari, Steven Wu, and Himabindu Lakkaraju. The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602, 2022. 6 +[34] Mengchen Liu, Jiaxin Shi, Kelei Cao, Jun Zhu, and Shixia Liu. Analyzing the training processes of deep generative models. Transactions on Visualization and Computer Graphics, 24(1):77-87, 2017. 2 +[35] Stuart Lloyd. Least squares quantization in PCM. Transactions on information theory, 28(2):129-137, 1982. 2 +[36] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Conference on Computer Vision and Pattern Recognition, pages 5188-5196, 2015. 2 +[37] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54 (6):1-35, 2021. 5 +[38] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/features-visualization.1,2 +[39] Chris Olah, Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, and Alexander Mordvintsev. The building blocks of interpretability. Distill, 2018. https://distill.pub/2018/building-blocks.1 + +[40] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 2020. https://distill.pub/2020/circuits/zoom-in. 2, 3 +[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, and Luca Antiga. PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 2, 4 +[43] Peter J Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 4, 5 +[44] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetv2: Inverted residuals and linear bottlenecks. In Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2 +[45] Anirban Sarkar, Deepak Vijaykeerthy, Anindya Sarkar, and Vineeth N Balasubramanian. A framework for learning antehoc explainable models via concepts. In Conference on Computer Vision and Pattern Recognition, pages 10286-10295, 2022. 2 +[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision, pages 618-626, 2017. 1, 2, 8 +[47] S. Seung. Connectome: How the Brain's Wiring Makes Us Who We Are. Houghton Mifflin Harcourt, 2012. 1 +[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations, 2014. 5, 6, 2, 4 +[49] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2, 6, 7, 4 +[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. Scikit-image: image processing in python. PeerJ, 2:e453, 2014. 3 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 5, 7, 2, 3, 4 +[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech UCSD birds 2011 dataset. Technical + +Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 4, 9 +[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565, 2020. 2 +[54] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pages 818-833, 2014. 1, 2 +[55] Quanshi Zhang, Ruiming Cao, Feng Shi, Ying Nian Wu, and Song-Chun Zhu. Interpreting CNN knowledge via an explanatory graph. In AAAI Conference on Artificial Intelligence, pages 4454-4463, 2018. 2 +[56] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, pages 8827-8836, 2018. 2 +[57] Quanshi Zhang, Yu Yang, Haotian Ma, and Ying Nian Wu. Interpreting CNNs via decision trees. In Conference on Computer Vision and Pattern Recognition, pages 6261-6270, 2019. 2 +[58] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for CNN models with non-negative concept activation vectors. In AAAI Conference on Artificial Intelligence, pages 11682-11690, 2021. 2 +[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Conference on Computer Vision and Pattern Recognition, pages 2921-2929, 2016. 1, 2 +[60] Bolei Zhou, Yiyou Sun, David Bau, and Antonio Torralba. Interpretable basis decomposition for visual explanation. In European Conference on Computer Vision, pages 119-134, 2018. 2 \ No newline at end of file diff --git a/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/images.zip b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e5d84a19857d3d0a2f0cec72c906b5685d297ba --- /dev/null +++ b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af3bed6b595ef7c3a6ba3ba8d5d493b95c90586327b7433380e1f8b5bf050025 +size 694532 diff --git a/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/layout.json b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c8198dad657bd80533d00e3e66c6abc16b09a1a0 --- /dev/null +++ b/2024/Visual Concept Connectome (VCC)_ Open World Concept Discovery and their Interlayer Connections in Deep Models/layout.json @@ -0,0 +1,10775 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "type": "text", + "content": "Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 160, + 478, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 160, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 114, + 160, + 478, + 175 + ], + "type": "text", + "content": "Matthew Kowal1,3 Richard P. Wildes1,2 Konstantinos G. Derpanis1,2,3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "spans": [ + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "text", + "content": "York University, " + }, + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "text", + "content": "Samsung AI Centre Toronto, " + }, + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 165, + 176, + 428, + 202 + ], + "type": "text", + "content": "Vector Institute yorkucvil.github.io/VCC" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 233, + 192, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 233, + 192, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 233, + 192, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 475 + ], + "type": "text", + "content": "Understanding what deep network models capture in their learned representations is a fundamental challenge in computer vision. We present a new methodology to understanding such vision models, the Visual Concept Connectome (VCC), which discovers human interpretable concepts and their interlayer connections in a fully unsupervised manner. Our approach simultaneously reveals fine-grained concepts at a layer, connection weightings across all layers and is amendable to global analysis of network structure (e.g. branching pattern of hierarchical concept assemblies). Previous work yielded ways to extract interpretable concepts from single layers and examine their impact on classification, but did not afford multilayer concept analysis across an entire network architecture. Quantitative and qualitative empirical results show the effectiveness of VCCs in the domain of image classification. Also, we leverage VCCs for the application of failure mode debugging to reveal where mistakes arise in deep networks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 483, + 128, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 128, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 128, + 496 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 653 + ], + "type": "text", + "content": "This paper focuses on interpreting the intermediate representations of deep networks for computer vision. The goal is understanding how various encoded concepts impact a model's prediction as well as concepts in other layers. We define concepts as abstractions that generalize from particular instances, including those defined locally (e.g. color and orientation), regionally (e.g. texture and shading) and from higher-level considerations (e.g. object parts, wholes and groupings). Human-interpretable concepts are of particular interest in increasing human understanding of models; however, extracting these concepts encoded in deep networks remains an open challenge in computer vision due to the complexity and opaque nature of these models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Understanding what concepts are learned by deep models and how they are encoded is important for both science and applications. For science, understanding what information drives a model's encoding of different concepts may indicate directions to advance performance. For applications," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 235, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 545, + 258 + ], + "type": "text", + "content": "several negative consequences of deploying opaque vision models have been documented, e.g. [8, 24]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 259, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 546, + 426 + ], + "type": "text", + "content": "Previous work has focused on interpreting models via feature attribution, which measures the contribution of individual inputs to a model's output [10, 46, 59]; so, explanations are of single pixels and may be difficult to interpret. Other work generates images that maximize activation of a model's features [19, 38, 54]. Like feature attribution, these approaches are qualitative and place most of the burden on the user to determine the concepts revealed. Concept-based interpretability, which identifies human-interpretable abstractions in a model's latent space [20, 23, 29, 32], yield quantitative contributions of a concept to the model's output and explanations on the class-level (vs. pixel-level). These approaches are easier to understand and validate; however, they have not been used to explore interlayer relationships." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": "As it stands, no approaches can quantify the interlayer affect of a given distributed concept at one layer, " + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": ", to another concept at a different layer, " + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "inline_equation", + "content": "l'" + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": " (rather than the model output). Even though it is well established that deep networks learn to build concepts hierarchically as information flows through the network [39, 54], understanding the hierarchical representations has been under-researched. Indeed, little is known about the characteristics of this concept hierarchy for today's models. Questions abound: How many concepts exist in a network? What are the connections and weights between concepts? Does the model architecture impact the hierarchical structure of concept abstractions?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "In response to these questions, we take inspiration from the biological notion of a connectome [47], defined as “a comprehensive structural description of the network of elements and connections forming a brain.” Analogously, we present the Visual Concept Connectome (VCC), a comprehensive structural description of a deep pretrained network model in terms of human-interpretable concepts and their relationships that form the internal representation maintained by the model; Fig. 1 shows an example. Notably, VCC generation works in the open-world setting, i.e. the concepts and interlayer connections are discovered without the need for any predefined concept dictionary." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10895" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 523, + 300 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 523, + 300 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 523, + 300 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 523, + 300 + ], + "type": "image", + "image_path": "1d5f6b084c6bcf97c82706cceb1a60f6fabd7772e1316cc51488b53170d1037b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": "Figure 1. A Visual Concept Connectome (VCC). At each layer the visual concepts learned by a deep model for a given class are revealed as are the learned interlayer concept connections. For each concept, up to four exemplars are shown as unmasked regions in a " + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": " image. Interlayer concept connections are shown as lines with darker lines indicating larger contributions. Shown is a VCC for every convolutional layer of a VGG16 model [49] trained on ImageNet [15] targeting recognition of class \"Tow Truck\". A closer visualization of VCC subgraphs reveals interesting compositions occurring at different levels of abstraction corresponding at different depths of the model. At early layers (bottom left), we observe oriented patterns (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{1,1}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": ") and brown color (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{1,2}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": ") composing the concept of green and brown orientation (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{2,1}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": "). Middle layers (right) show the concept of 'wheel on the road' (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{9,1}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": ") being composed of wheels (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{8,1}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": ") and regions of asphalt (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{8,2}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": "). The final layer concepts (top left) show that both foreground objects, e.g. tow trucks (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{13,1}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": "), and background regions, e.g. road, trees, humans, or car being towed (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{13,2}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": "), concepts highly influence the final category (" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "inline_equation", + "content": "C_{TowTruck}" + }, + { + "bbox": [ + 46, + 302, + 545, + 401 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 410, + 148, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 148, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 148, + 422 + ], + "type": "text", + "content": "2. Related research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "content": "While a number of intrinsically interpretable networks have appeared (e.g. [6, 12, 31, 45, 56, 57]), we focus on work that, like ours, endeavours to interpret black-box models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": "Concept-based interpretability. Closed-world concept interpretability considers cases where a labelled dataset defines the concepts of interest for post-hoc analysis; approaches include Network Dissection [3] and Testing with Concept Activation Vectors (TCAV) [29]. However, a desirable property for concept-based interpretability is not to be restricted to a set of predefined classes (i.e. closed-world), but to support discovery of new and previously unlabelled concepts (i.e. open-world). Approaches to open-world concept interpretability include Automatic Concept Explanations (ACE) [23], SegDiscover [27], ConceptSHAP [53], Invertible Concept Explanations [58] and CRAFT [20]. These approaches are limited to single layer analysis and do not explain interlayer relationships." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Activation maximization [19, 36, 38, 54] visualizes the input that most activates a model component (e.g. filter, layer or logit) or how same layer units combine information [40]. These approaches do not capture interlayer relations and place a heavy burden on the user to interpret what concepts appear in their output where there may be no clear resemblance to natural concepts. Class Activation Maps" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": "(CAMs) and extensions visualize an input image's local region that contributes to a model's output [9, 46, 59]. Layerwise Relevance Propagation [5] (LRP) generates pixel contribution heatmaps by assuming conservation of information is propagated through each neuron. These approaches interpret single images (not classes), are purely qualitative, and are sensitive to small input perturbations [22, 30], yet insensitive to model changes [2]. Less related are directions interpreting generative models, e.g. [4, 34]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Interpretability via decomposition. Most closely related to our work are those that decompose a model's prediction into interpretable components. CRAFT [20] extracts concepts from the last layer of a convolutional neural network (CNN), but also searches earlier layers for sub-concepts for the maximally activating images of a given concept, but do not provide any way to quantify the contribution of sub-concepts to concepts. Another approach combined CAMs of multiple concepts to explain the final prediction [60]; however, it is limited to labelled data (so cannot discover concepts outside training) and only explains a final output, not intermediate relations. Other work interprets CNNs by distilling them into a graph [55]; however, it applies only to individual filters conceived as capturing object parts (i.e. not concepts other than objects and their parts, e.g. textures, colors or grouping), does not con" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10896" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "sider the fact that representations are encoded via superposition [7, 17, 40] (i.e. more than one concept can be captured per-channel), does not yield meaningful edge weights and is only for CNNs. Yet other work extends Grad-CAM or LRP to perform intra-layer visualizations [1, 13]. While these approaches capture concepts other than object parts, they still ignore the core problem of superposition (and distributed representations) and do not yield interpretation for an entire class, just single images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 324 + ], + "type": "text", + "content": "Contributions. In the light of previous work, our contributions are fourfold. (i) We introduce and formalize the notion of a Visual Concept Connectome (VCC) for deep network models. The VCC reveals concepts represented at any given layer of a network as well as their interlayer connectivity. (ii) We present a method for extracting a VCC from any pretrained deep network in an unsupervised, open world setting, with a focus on classification. (iii) We validate our approach with quantitative and qualitative experiments wherein we examine various standard models to yield insights into model architectures and training tasks. (iv) We apply VCCs to explaining failure modes in deep models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 334, + 253, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 253, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 253, + 348 + ], + "type": "text", + "content": "3. Visual Concept Connectomes (VCCs)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": "A VCC is a directed acyclic graph, " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\mathbf{Q},\\mathbf{E})" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ", created by distilling a pretrained deep network. The graph is topologically sorted based on the layer order found in the original network and has " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": " layers, consisting of " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": " network layers to be analyzed (i.e. any subset of layers that a user selects, including all layers at the extreme) plus the final prediction (e.g. the object category for object recognition). The graph's nodes, " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ", are vectors (cluster centroids) representing interpretable concepts and its edges, " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ", are scalars representing the contribution of one concept to the existence of another." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": "To construct a VCC, three inputs are required: (i) a set of " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": " images representative of a given task (e.g. exemplar images of an object category for object recognition), " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathcal{I}^1,\\dots ,\\mathcal{I}^I\\}" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^i\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "h,w" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": " the image height, width and channel dimension (e.g. three for RGB), respectively, (ii) an " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": " layer network, " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "F(\\cdot) = \\{f_{1}(\\cdot),\\ldots ,f_{N}(\\cdot)\\}" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": " denoting feature extraction at layer " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": ", and (iii) a set of " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": " selected layers (a subset of " + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": ", possibly improper), which are to be studied." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "A VCC is constructed in three main steps. (i) Image segments are extracted in the model's feature space via divisive clustering to produce semantically meaningful image regions for each selected layer (Sec. 3.1). (ii) Layer-wise concepts, i.e. the nodes of the graph, are discovered in an open world fashion (i.e. no labelled data is required) via a second round of clustering over the dataset of image regions, independently for each layer (Sec. 3.2). (iii) Edges are calculated that indicate the contribution of concepts from earlier to deeper layers via an approach we introduce, Interlayer Testing with Concept Activation Vectors (ITCAVs)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "(Sec. 3.3). We use object recognition for the explanation of the approach; nevertheless, it could be extended to other tasks in a straightforward way (e.g. semantic segmentation). An overview of these steps is provided in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 126, + 470, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 126, + 470, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 470, + 139 + ], + "type": "text", + "content": "3.1. Feature space image segments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": "For each selected layer, we produce a set of image regions that plausibly belong to a concept encoded by the model at that layer. Previous work produces concept segments in RGB space, i.e. superpixels or random crops [20, 23]; however, segmentation based on features different from those over which concepts are to be discovered, i.e. the model's deep features, results in the support for the discovery process being divorced from the process that generated the support. Therefore, our segmentation uses the same deep features to be used subsequently for concept discovery. Our divisive clustering approach is presented visually in Fig. 2 (A). We pass the entire set of " + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": " input images, " + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": ", for a given class through the model, " + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": ", to get features, " + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n^i \\in \\mathbb{R}^{h_n \\times w_n \\times c_n}" + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": ", at each selected layer, " + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 144, + 545, + 312 + ], + "type": "text", + "content": " according to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 336, + 316, + 545, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 316, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 336, + 316, + 545, + 332 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {n} = \\left\\{f _ {n} \\left(\\mathcal {I} ^ {1}\\right), \\dots , f _ {n} \\left(\\mathcal {I} ^ {I}\\right) \\right\\} = \\left\\{\\mathbf {z} _ {n} ^ {1}, \\dots , \\mathbf {z} _ {n} ^ {I} \\right\\}. \\tag {1}", + "image_path": "86fb34d693e296e6e7183e98efece47cb103c1fa664b2536aee509443e48f5a0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": "To extract image segments at layer " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": " for concept discovery (Sec. 3.2), we cluster the image activations, " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_n" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": ", conditioned on the clusters (i.e. masks) generated at the next higher layer, " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (x,y)" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": " index spatial coordinates and " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_n^i (\\mathbf{p};\\gamma)" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": " be a binary mask for cluster " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_n^i (\\mathbf{p};\\gamma) = 1 \\iff \\mathbf{z}_n^i (\\mathbf{p}) \\in \\gamma" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": "; otherwise, " + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_n^i (\\mathbf{p};\\gamma) = 0" + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": "We calculate a set of such masks by applying a clustering algorithm, " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathsf{C}_{\\Gamma}^{seg}(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": ", that returns binary support masks for " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " clusters on its argument. Let " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\Gamma_{n+1}" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " be the number of segments at layer " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "n+1" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "1 \\leq g \\leq \\Gamma_{n+1}" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " index a particular segment, " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " we calculate segments at layer " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 472, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 472, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 323, + 472, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {B} _ {n} ^ {i} (\\mathbf {p}; \\gamma) \\right\\} _ {\\gamma = 1} ^ {\\Gamma} = \\mathrm {C} _ {\\Gamma} ^ {s e g} \\left(\\mathbf {z} _ {n} ^ {i} (\\mathbf {p}) \\odot \\widetilde {\\mathbf {B}} _ {n + 1} ^ {i} (\\mathbf {p}; g)\\right), \\tag {2}", + "image_path": "bf19ae8d494d2daeffbcffa79b3618ac97041cbeeacdb76ababa27fdf829f0cc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{B}}_{n + 1}^i" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{n + 1}^i" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " upsampled to the resolution of layer " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " indicates spatial element-wise masking applied to the individual image activations. The element-wise masking ensures clustering is done on the masked area of " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n^i" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": ". Thus, the support of the concept discovery at layer " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " comes from the support of concepts at layer " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": ". Notably, we let " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " vary with " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": "; so, different segments, " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": ", at layer " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": " can yield a different number of segments, " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": ", at layer " + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "type": "text", + "content": "The top-down conditional clustering, (2), is repeated recursively for all image segments in all selected layers. To initiate the recursion, (2), at the top layer, " + }, + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "type": "inline_equation", + "content": "n_{top}" + }, + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "type": "text", + "content": ", all the features are used by setting the number of clusters to be one, and having " + }, + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{B}}_{n_{top} + 1}^i (\\mathbf{p};1) = 1" + }, + { + "bbox": [ + 304, + 597, + 545, + 660 + ], + "type": "text", + "content": " for all images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "text", + "content": "At each layer, " + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "text", + "content": ", we construct a set of RGB segmentations, " + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "text", + "content": ", for use in concept discovery (Sec. 3.2) by applying upsampled masks, " + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_n^i (\\mathbf{p};\\gamma)" + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "text", + "content": ", to a given image, " + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^i" + }, + { + "bbox": [ + 304, + 660, + 545, + 696 + ], + "type": "text", + "content": ", via" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 359, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {n} ^ {i} (\\mathbf {p}; \\gamma) = \\mathcal {I} ^ {i} (\\mathbf {p}) \\odot \\mathbf {B} _ {n} ^ {i} (\\mathbf {p}; \\gamma), \\tag {3}", + "image_path": "af7d1d2d7091b4e57a045222ea5cae39bd59032dc714f4340dbdacaf5ac3b123.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10897" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 262, + 272 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 262, + 272 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 262, + 272 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 262, + 272 + ], + "type": "image", + "image_path": "457a2c85cd74cfe2897fc892925150a793916aa62f396961d5e3ba0e856b1358.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": "Figure 2. The three steps in building a Visual Concept Connectome (VCC). (A) For a given image, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^i\\in \\mathcal{I}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", model, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", and layer, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", we produce a set of image segments, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_n^i\\in \\mathbf{M}_n" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", based on a recursive spatial clustering, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathsf{C}^{seg}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": " (2), of the features " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n^i" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": " conditioned on the clusters from the layer above, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ". We then use (3) to generate a set of masked RGB image segments for each layer, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ". (B) For a given layer, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", we pass the image segments from all images, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", through " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": " and cluster, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathsf{C}^{con}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": " (4), these features across the dataset to produce " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": " concept centroids, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_j^1,\\dots ,\\mathbf{q}_j^{m_j}\\}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ". (C) To measure the contribution of an earlier layer concept, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j^{m_j}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", to a later layer concept, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", we employ our Interlayer Testing with CAV (ITCAV) approach (Sec. 3.3), which uses the Concept Activation Vector (CAV) [29] of the earlier concept, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathfrak{q}_j^{m_j}}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": " (that points away from random examples, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{rnd}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": ", but toward concept exemplars, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathfrak{q}_j^{m_j}}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": "), and the deeper layer concept, " + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 46, + 274, + 545, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 265, + 72, + 534, + 272 + ], + "blocks": [ + { + "bbox": [ + 265, + 72, + 534, + 272 + ], + "lines": [ + { + "bbox": [ + 265, + 72, + 534, + 272 + ], + "spans": [ + { + "bbox": [ + 265, + 72, + 534, + 272 + ], + "type": "image", + "image_path": "5f2d71779ddf2a157a18eb67605389de8bb55a4b57b2b8f47ecafc83038d903c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_n^i (\\mathbf{p};\\gamma)\\in \\{0,\\dots,255\\}^{h_n\\times w_n}" + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\{0,\\dots,255\\}" + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "content": " specifying RGB value. We follow by defining " + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_n" + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "content": " as the set of all RGB image segments at layer " + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "content": " (with each segment given in terms of (3)), and letting " + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{\\mathbf{M}_1,\\dots,\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 46, + 357, + 287, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "type": "text", + "content": "We instantiate the clustering, " + }, + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "type": "inline_equation", + "content": "C^{seg}" + }, + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "type": "text", + "content": ", via maskSLIC [28], an extension of k-means, with an " + }, + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 407, + 287, + 453 + ], + "type": "text", + "content": " norm, that clusters features while respecting a mask and automate selection of the number of clusters via the silhouette method [43]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 461, + 206, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 206, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 206, + 474 + ], + "type": "text", + "content": "3.2. Layer-wise concept discovery" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "text", + "content": "Given the dataset of masked image segments for each layer, " + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{\\mathbf{M}_1,\\dots ,\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "text", + "content": ", we follow [23] and bilinearly interpolate the segments to the input image size and then pass the segments through the model to get pre-segmented activations " + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{M}_j} = f_j(\\mathbf{M}_j)" + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "text", + "content": " from the layer where the segments were found in the divisive clustering step (Sec. 3.1). We pass the image segments (as opposed to using the image activations) through the model to remove any contextual information before we discover concepts (the segments are set to zero everywhere outside of the segment boundary). Otherwise, " + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 46, + 479, + 287, + 610 + ], + "type": "text", + "content": " could mix information from outside the segment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "type": "text", + "content": "The layer-wise concept discovery step is presented visually in Fig. 2 (B). With the pre-segmented activations, " + }, + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{M}_j}" + }, + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "type": "text", + "content": ", calculated for all layers, " + }, + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 611, + 287, + 671 + ], + "type": "text", + "content": ", we cluster over the dataset of segment features for each individual layer to produce cluster centroids (i.e. concepts) according to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 679, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 679, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 78, + 679, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {j} = \\left\\{\\mathbf {q} _ {j} ^ {1}, \\dots , \\mathbf {q} _ {j} ^ {m _ {j}} \\right\\} = \\mathrm {C} _ {m _ {j}} ^ {\\text {c o n}} (\\operatorname {G A P} \\left(\\mathbf {Z} _ {\\mathbf {M} _ {j}}\\right)), \\tag {4}", + "image_path": "f560ac90114ae8ba425a1882c2c5a520ed3f023acc3bafd2021802fa7c238476.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is the number of concepts discovered at layer " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": "GAP is spatial global average pooling and " + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathsf{C}_{m_j}^{con}" + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": " is a clustering algorithm that returns " + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": " cluster centroids. Following previous work [23], we instantiate " + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathsf{C}_{m_j}^{con}" + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": " in terms of standard k-means, using the " + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": " norm, and then use a large number of clusters with subsequent pruning to remove noisy clusters; see the supplement for details. The output of this stage results in discovered concepts (i.e. cluster centroids and associated segments) for all layers as " + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\{\\mathbf{Q}_1, \\dots, \\mathbf{Q}_n\\}" + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": " for the centroids and " + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{q}_j^{m_j}}" + }, + { + "bbox": [ + 304, + 357, + 545, + 469 + ], + "type": "text", + "content": " their associated RGB segments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 477, + 473, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 473, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 473, + 490 + ], + "type": "text", + "content": "3.3. Interlayer concept connectivity" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "Our approach to quantifying the contribution of discovered concepts between two layers, Interlayer Testing with Concept Activation Vectors (ITCAV), generalizes TCAV [29]. Specifically, while TCAV calculates the contribution of a single mid-layer concept to the class logit using a dataset of labelled concept images, we generalize to operate between any two layers and without a dataset of labelled concept images. Without loss of generality, we consider a single pair of concepts at two layers, which need not be adjacent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": "The ITCAV method is presented visually in Fig. 2 (C). We seek to measure the degree to which a concept at an earlier layer contributes to a deeper layer. To do so, we construct two vectors: (i) a vector that represents the concept at the earlier layer, " + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": ", and (ii) a vector that represents a positive contribution (i.e. gradient) of a concept at a later layer, " + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": ". The closer the alignment of these two vectors, the larger the contribution of the concept at layer " + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": " to the concept at layer " + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j^{m_j}" + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{q}_j^{m_j}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 716 + ], + "type": "text", + "content": " be the cluster centroid and as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10898" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 160, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 160, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 160, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 160, + 159 + ], + "type": "image", + "image_path": "5881e4c22dd331a2fa5610ade43f18d73597099d7caabd81203b32ddb60b19c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 164, + 113, + 172 + ], + "lines": [ + { + "bbox": [ + 103, + 164, + 113, + 172 + ], + "spans": [ + { + "bbox": [ + 103, + 164, + 113, + 172 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 77, + 296, + 159 + ], + "blocks": [ + { + "bbox": [ + 181, + 77, + 296, + 159 + ], + "lines": [ + { + "bbox": [ + 181, + 77, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 181, + 77, + 296, + 159 + ], + "type": "image", + "image_path": "92390b4b63fbd6b7de8210b4b6988778a1cbaa23eba2cb61d1b41d10143c7c06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 164, + 246, + 173 + ], + "lines": [ + { + "bbox": [ + 234, + 164, + 246, + 173 + ], + "spans": [ + { + "bbox": [ + 234, + 164, + 246, + 173 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 421, + 162 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 421, + 162 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 421, + 162 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 421, + 162 + ], + "type": "image", + "image_path": "f2efe2dfd7064929fbf47a9aa2501d613deb5aba2f5c35f32c81a1dcf570ab1b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 164, + 367, + 172 + ], + "lines": [ + { + "bbox": [ + 356, + 164, + 367, + 172 + ], + "spans": [ + { + "bbox": [ + 356, + 164, + 367, + 172 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 423, + 72, + 541, + 163 + ], + "blocks": [ + { + "bbox": [ + 423, + 72, + 541, + 163 + ], + "lines": [ + { + "bbox": [ + 423, + 72, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 423, + 72, + 541, + 163 + ], + "type": "image", + "image_path": "6ad322a960174b95501f00da33b42799d03fe1798b0309ac4bb33968636edf53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 164, + 488, + 172 + ], + "lines": [ + { + "bbox": [ + 477, + 164, + 488, + 172 + ], + "spans": [ + { + "bbox": [ + 477, + 164, + 488, + 172 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 172, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 547, + 239 + ], + "type": "text", + "content": "Figure 3. Validation of the three VCC method components. (a) Validation of segment proposals. The relative (Rel.) concept segment size compared to entire image for a given layer is plotted against the receptive field (RF) width/height of the same layer. (b) Validation of discovered concepts. For 50 randomly selected ImageNet classes, we discover concepts in four layers of the model. During inference, one randomly selected concept at each layer is suppressed by a factor of " + }, + { + "bbox": [ + 46, + 172, + 547, + 239 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 172, + 547, + 239 + ], + "type": "text", + "content": ". (c, d) Validation of interlayer concept weights. The unnormalized logit sum (LS) scores, (8), for the target class are plotted against the average path strength (APS) scores, (7). A positive correlation implies that the ITCAV edge weights connecting a concept to the class are predictive of the model output having a higher probability for that class." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "content": "sociated RGB image segments in layer " + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{q}_l^{m_l}}" + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "content": " be the cluster centroid and associated RGB image segments in a deeper layer " + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "inline_equation", + "content": "l > j" + }, + { + "bbox": [ + 46, + 247, + 287, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": "We proceed by constructing the vector that represents the concept in the feature space of the earlier layer, " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": ". To do so, we employ the Concept Activation Vector (CAV) [29] for the earlier concept, " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j^{m_j}" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": ", by training a linear classifier, " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": " on the features of layer " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": " between positive concept inputs, " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "f_{j}(\\mathbf{M}_{\\mathbf{q}_{j}^{m_{j}}})" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": ", and random images, " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "f_{j}(\\mathcal{I}_{rnd})" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": ". The orthogonal vector to the hyperplane defined by " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{q}_j^{m_j}}" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": " is the CAV for concept " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j^{m_j}" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": " and points in the direction of concept " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j^{m_j}" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": " in the feature space of layer " + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 285, + 287, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": "Next, we construct the vector in the feature space of layer " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": " that points in the direction of positive contribution of concept " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": ". To do so, we calculate " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\nabla f_l(f_j(\\cdot))" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": ", i.e. the gradient of the deeper concept, " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": ", at layer " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": ", with respect to the earlier layer, " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": ". Our approach to measuring the sensitivity between the two concepts is then" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 475, + 287, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 475, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 61, + 475, + 287, + 497 + ], + "type": "interline_equation", + "content": "S _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} (x) = \\nabla \\left(\\left\\| f _ {l} \\left(f _ {j} (x)\\right) - \\mathbf {q} _ {l} ^ {m _ {l}} \\right\\| _ {2}\\right) \\cdot \\mathbf {V} _ {\\mathbf {q} _ {j} ^ {m _ {j}}}, \\tag {5}", + "image_path": "45016d0d6067f6b7f6459a56708381f269d8f0b9f489a10956772ca78000bde2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{M}_{\\mathbf{q}_l^{m_l}}" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": ". The sensitivity score should be positive because when the feature at layer " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": " is perturbed in the direction of concept " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j^{m_j}" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": ", i.e. in the direction " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{q}_j^{m_j}}" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": ", then the feature vector in deeper layer, " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": ", will be pushed closer to concept " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": ". Following TCAV [29], the final ITCAV edge weight, " + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "inline_equation", + "content": "e_{\\mathbf{q}_j^{m_j},\\mathbf{q}_l^{m_l}} \\in \\mathbf{E}" + }, + { + "bbox": [ + 46, + 501, + 287, + 581 + ], + "type": "text", + "content": ", is the ratio of positive sensitivities," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 52, + 582, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 582, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 52, + 582, + 287, + 601 + ], + "type": "interline_equation", + "content": "e _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} = | x \\in \\mathbf {M} _ {\\mathbf {q} _ {l} ^ {m _ {l}}}: S _ {\\mathbf {q} _ {j} ^ {m _ {j}}, \\mathbf {q} _ {l} ^ {m _ {l}}} (x) > 0 | / | \\mathbf {M} _ {\\mathbf {q} _ {l} ^ {m _ {l}}} |. \\tag {6}", + "image_path": "44a9ed4c3958c8999113f7a778ac16e0e9aa0208c038abe411c84317381d38b1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": "Key to our sensitivity score, (5), is the " + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": " distance between the output, " + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{l}(f_{j}(x))" + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": ", and cluster centroid, " + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": ", before taking the gradient, cf. CAV [29]. This choice serves two goals: (i) The subset of dimensions of individual image features, " + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{j}(x)" + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": ", that are not well aligned with the cluster centroid, " + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_l^{m_l}" + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": ", will be penalized. This penalty will suppress the impact of noisy dimensions from individual samples on ITCAV. We use the " + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 45, + 605, + 288, + 714 + ], + "type": "text", + "content": " norm because the employed clustering approach used during concept discovery, k-means, uses" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 247, + 546, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 546, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 546, + 284 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 247, + 546, + 284 + ], + "type": "text", + "content": "; see Sec. 3.2. (ii) Collapsing the tensor to a scalar (i.e. a concept similarity score) makes the gradient calculation computationally feasible compared to the full Jacobian." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 291, + 388, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 388, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 388, + 304 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 310, + 437, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 437, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 437, + 323 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "type": "text", + "content": "For evaluation we use ResNet50 [25], VGG16 [48], MobileNetv3 [26], ViT-b [51] and MViT [18] to sample a broad range of popular architectures. The exact layers selected are in the supplement, but always include the last feature layer before the fully connected layer to probe concepts closest to the model output. Following previous work [23, 29], we perform a two-sided t-test on ITCAV scores for multiple runs of the same concept vs. different random sets of images to remove statistically insignificant scores and randomly sample from the Broden dataset [3] to get the images, " + }, + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{rnd}" + }, + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "type": "text", + "content": " (Sec. 3.3), used in statistical testing and CAV training. On average, four and all-layer VCCs take 15 minutes and 36 hours, resp., to generate on an NVIDIA Quadro RTX 6000 GPU. The supplement has more details and VCCs for other layers, models and datasets [52]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 515, + 457, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 457, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 457, + 528 + ], + "type": "text", + "content": "4.2. VCC component validation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Segment proposal validation. Our approach segments concepts based on the feature space of a given layer to allow us to capture concepts across multiple layers (Sec. 3.1). So, the spatial support of valid segments at a layer should follow the receptive field (RF) of filters at that layer, as the segments are constrained by the filters from which they are derived. For comparison, we consider ACE [23], which does segmentation with a layer independent feature (color), limited to the input image to discover concepts at a single layer for a single model. Being directly related to the features at a layer, our approach should produce concepts that better respect the RF at that layer compared to the baseline, whose features are layer independent (e.g. at CNN early layers the RF is small; so, intuitively the patch proposals should be smaller as well and conversely for the later layers)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10899" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 289, + 344 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 344 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 344 + ], + "type": "image", + "image_path": "76f1e05871175636312fd26837cec36445a232f6a6369bbf3f7cbd8513d7d1e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 376 + ], + "type": "text", + "content": "Figure 4. A VCC for three selected layers of a GoogLeNet model [49] targeting recognition of class 'Jay'. Darker lines denote stronger connection weights." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 377, + 289, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 520 + ], + "type": "text", + "content": "Figure 3 (a) shows the average concept size (in terms of the ratio of segment pixels to the entire image) vs. the RF at the corresponding layer for three different models: (i) ResNet50 [25], (ii) VGG16 [48] and (iii) ViT-b [16] (note we use mean attention distance instead of receptive field for ViT-b). While the baseline [23] produces segments with sizes invariant to the layer analyzed, our approach produces segments with image sizes that scale with the RF. Additional results with other models (including more transformers) are given in the supplement, which further support this observation. Overall, it is expected by design and we find that our segment proposals follow RFs at each layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": "Concept fidelity. Concept fidelity measures the meaningfulness of discovered concepts with respect to the target model. While other work focuses on single layer concept fidelity [20, 23, 29], we measure the effect on a model's output when suppressing the encoding of concepts as the information propagates through the model. The expectation is that suppressing the discovered concepts should result in a quicker decrease in performance compared to a non-concept direction (e.g. a randomly selected direction) as the amount of suppression gets larger. For a given model and category, we compute the concepts at various layers using our method (Sec. 3.2). We perform concept suppression for a given image, " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^i" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ", at a given layer, " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ", for concept " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ", according to " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j^i = \\mathbf{z}_j^i - \\epsilon \\cdot \\mathbf{q}_j / ||\\mathbf{q}_j||_2" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " controls the degree of perturbation. If the concepts discovered are meaningful, then the accuracy for the model's target class should decrease faster" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 73, + 430, + 131 + ], + "blocks": [ + { + "bbox": [ + 313, + 73, + 430, + 131 + ], + "lines": [ + { + "bbox": [ + 313, + 73, + 430, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 430, + 131 + ], + "type": "image", + "image_path": "bc44278740832248dd9bdd782de3e20e213c70c052d9e739efad4ffddfb990f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 73, + 548, + 132 + ], + "blocks": [ + { + "bbox": [ + 433, + 73, + 548, + 132 + ], + "lines": [ + { + "bbox": [ + 433, + 73, + 548, + 132 + ], + "spans": [ + { + "bbox": [ + 433, + 73, + 548, + 132 + ], + "type": "image", + "image_path": "c30cc6e647fdc7cd8c27d2617d17a6aa47114bcae5e1cea57208ee5f651219d8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 134, + 428, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 134, + 428, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 134, + 428, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 428, + 198 + ], + "type": "image", + "image_path": "cff0b91a0307d8f889552b5c476100f36baf0122cf0ac0bca4c45dd26b679c51.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "text", + "content": "Figure 5. Graph metrics on four layer VCCs comparing CNN vs. transformer architectures and training objectives." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 433, + 133, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 433, + 133, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 433, + 133, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 433, + 133, + 547, + 198 + ], + "type": "image", + "image_path": "34ebd885aa6c5d2ee5c2b43d285ff062691e7f2c137d2376e9fb78d93dd3af9b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "type": "text", + "content": "compared to random perturbations, " + }, + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{rnd}" + }, + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 222, + 533, + 234 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "type": "text", + "content": "Figure 3 (b) has concept suppression results averaged over 50 randomly selected ImageNet [15] classes (documented in the supplement), where a concept is randomly chosen to be suppressed at each layer. Note that the " + }, + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "type": "text", + "content": " values required to reduce the model accuracy to zero differs by model; so, we scale the " + }, + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 234, + 545, + 341 + ], + "type": "text", + "content": " values to [0, 1] for visualization purposes. For all models, it is seen that perturbing in the opposite direction of a recovered concept more severely impacts the accuracy than a random perturbation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": "ITCAV validation. If the interlayer concept connection weights are meaningful, then the accumulated weights from earlier to later layer concepts should be correlated with their probability of predicting the target class. We use this intuition to validate our ITCAV approach to recovering interlayer concept weights as follows. We define the average path strength (APS), a scalar value between zero and one representing the average strength of the connection between a concept and the class logit. Given concept " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": ", let there be " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": " layers in the VCC between the concept and the logit layer; thus, each path from the concept to the logit consists of " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": " edge weights. We consider all possible paths in the VCC from this concept to the class logit. Let " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "e_{\\mathbf{q}_j}(l,p)" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": " be the edge weight (i.e. ITCAV score) of the " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": " path (from concept " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": " to the class concept) for VCC layer " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": ". To calculate the APS score for a given concept, we average over all paths, " + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 342, + 546, + 547 + ], + "type": "text", + "content": ", and edge weights in each path to get" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 553, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 553, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 349, + 553, + 545, + 587 + ], + "type": "interline_equation", + "content": "\\operatorname {A P S} \\left(\\mathbf {q} _ {j}\\right) = \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\left[ \\frac {1}{L} \\sum_ {l = 1} ^ {L} e _ {\\mathbf {q} _ {j}} (l, p) \\right]. \\tag {7}", + "image_path": "c2fde735d6c7e1679dd56fb4268855c0fccc1444881530c0e7b9dccd71856770.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "text", + "content": "Next, we calculate the logit sum (LS) score for a given concept, " + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "text", + "content": ", by passing all of the corresponding masked segments, " + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{q}_j}" + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "text", + "content": ", through the model, " + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "text", + "content": ". We then sum up the logit scores for the target class, " + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 593, + 545, + 641 + ], + "type": "text", + "content": ", to yield" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 643, + 545, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 643, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 369, + 643, + 545, + 679 + ], + "type": "interline_equation", + "content": "\\operatorname {L S} \\left(\\mathbf {q} _ {j}\\right) = \\sum_ {i = 1} ^ {\\left| \\mathbf {M} _ {\\mathbf {q} _ {j}} \\right|} F \\left(\\mathbf {M} _ {\\mathbf {q} _ {j}} ^ {i}\\right) | _ {c}, \\tag {8}", + "image_path": "c813fe116f67e8a2477c5a7c6d2e963c3fbbea12209225157263983d0af26ba3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{q}_j}^i" + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "text", + "content": " segment mask, and " + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "inline_equation", + "content": "F(\\cdot)|_c" + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 686, + 545, + 714 + ], + "type": "text", + "content": " logit score. If the ITCAV edge weights are" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10900" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 167, + 130 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 167, + 130 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 167, + 130 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 167, + 130 + ], + "type": "image", + "image_path": "f38a7386107a106becddbfd0c0f63f015fa3ad24608e06b0c3628e99fb464a75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 286, + 129 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 286, + 129 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 286, + 129 + ], + "type": "image", + "image_path": "3fe9fecc4656ffa6232ffdcaa02dc4264a94987c188c093bae92ff55bbc1bf81.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 132, + 167, + 194 + ], + "blocks": [ + { + "bbox": [ + 54, + 132, + 167, + 194 + ], + "lines": [ + { + "bbox": [ + 54, + 132, + 167, + 194 + ], + "spans": [ + { + "bbox": [ + 54, + 132, + 167, + 194 + ], + "type": "image", + "image_path": "ec38988f507cf770b2e0c96757209e02957d90efb49d6a1316a0342d4db47f2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 131, + 284, + 195 + ], + "blocks": [ + { + "bbox": [ + 173, + 131, + 284, + 195 + ], + "lines": [ + { + "bbox": [ + 173, + 131, + 284, + 195 + ], + "spans": [ + { + "bbox": [ + 173, + 131, + 284, + 195 + ], + "type": "image", + "image_path": "945936437e3d7a3789c998f383588d8241ed663aaba0068ef7f8ade02c5315b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 287, + 228 + ], + "type": "text", + "content": "Figure 6. Graph metrics of all layer VCCs for three diverse architectures. Layer number normalized to allow for comparison of models with different numbers of layers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 232, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 257 + ], + "type": "text", + "content": "meaningful, then there should be positive correlation between the APS and LS scores, i.e. " + }, + { + "bbox": [ + 46, + 232, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{APS}(\\mathbf{q})\\propto \\mathrm{LS}(\\mathbf{q})" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 257, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 340 + ], + "type": "text", + "content": "Figure 3 (c, d) shows the LS scores plotted vs. the APS scores for VGG16 [25] and ViT-b [51]. A positive correlation is found between the APS and LS scores for both architectures. These results suggest that the combination of ITCAV scores is predictive of whether a concept is representative of the target class. The supplement has more LS vs. APS plots for more models that support these findings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 348, + 230, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 230, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 230, + 361 + ], + "type": "text", + "content": "4.3. Understanding models with VCCs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 366, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 437 + ], + "type": "text", + "content": "To demonstrate VCC's unique ability to interpret models at different resolutions, we present model analyses in three parts: (i) qualitative analyses of VCCs generated for a subset of layers as well as at all layers, (ii) quantitative analysis of VCCs generated for a subset of four layers, (iii) a broadened quantitative analysis to encompass all model layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": "Visualizing concept hierarchies. Figure 4 shows a three layer VCC for GoogLeNet [49] targeting the class \"Jay\". Notice how the inception5b 'bird' concepts " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{51}, c_{52})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " form as selective weighting of inception4c concepts background " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{41})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": ", bird part " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{42}, c_{44})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " and tree branch " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{43})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": ", while the inception5b background concepts " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{53}, c_{54})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " form differently from weighting of solely inception4c background part concepts (e.g. tree branch " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{43})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " and green leaves " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{41})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": "). Notably, the network separates subspecies of Jay in the final layer (e.g. Blue Jay " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{52})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " and other types " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{51}))" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": ". The concepts found in inception4c are composed from varying combinations of colors and parts found in conv3 (e.g. various bird parts " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "(c_{31}, c_{33})" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " contribute to the bird concepts at inception4c). In the end, both scene and object contribute with strong weights to the final category." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "An all layer VGG-16 VCC visualization was presented in Fig. 1. As discussed in the caption, hierarchical concept assemblies again are revealed, with both target object as well as its background contributing to the final classification. While both the few layer and all layer visualizations reveal the concept representations of the models under analysis, they afford different levels of granularity: Few layer visualization provides a concise summary with a focus on" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 443, + 207 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 443, + 207 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 443, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 443, + 207 + ], + "type": "image", + "image_path": "a2d87502e4cf58d753bf76746ecc560d10081152709067273814440e0b0ca221.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "text", + "content": "Figure 7. VCCs for a CNN (ResNet50) and transformer (MViT) for the class 'Tow Truck'. Shown are only the last two layers with the class output." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 449, + 69, + 545, + 207 + ], + "blocks": [ + { + "bbox": [ + 449, + 69, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 449, + 69, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 449, + 69, + 545, + 207 + ], + "type": "image", + "image_path": "82c317552b64a101232d7d886207adb3fb4315b72a3489b98040814d2c5335ef.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 244, + 544, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 544, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 544, + 256 + ], + "type": "text", + "content": "specific layers, while all layer provides very detailed study." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 257, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 412 + ], + "type": "text", + "content": "Subset VCC analysis. We begin our quantitative analysis on four-layer VCCs generated for the CNNs and transformers listed in Sec. 4.1. Layers are selected approximately uniformly (see supplement for exact layers) across a model to capture concepts at different abstractions. We use standard metrics for analyzing tree-like data structures in a per-layer fashion: branching factor (extent to which the representation is distributed), number of nodes (how many concepts the model uses to encode a particular class), edge weights (strength of connection between concepts) and edge weight variance (variability of connection strength). We calculate averages over 50 VCCs for 50 randomly selected ImageNet classes. Results are shown in Fig. 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 414, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 546, + 676 + ], + "type": "text", + "content": "Patterns are apparent in concept numbers and edge weights: more, but weaker, concepts in early layers vs. fewer, but stronger, concepts in later layers. These patterns reflect the shared low-level features (e.g. colors, textures) across classes and more specific features near the end, which yield larger ITCAV values. Also, CNNs show a decreasing branching factor, while transformers maintain a consistent number until the final layer, where all models converge to about two concepts, typically of an ImageNet class's foreground-background structure. Transformers have higher final layer edge weight variance compared to CNNs, indicating their ability to better differentiate earlier concepts' importance in forming later concepts, potentially explaining their superior classification performance (i.e. all information is not equally valuable). We also compare models (ResNet50 and VGG16) when trained with self-supervision [11] (SimCLR) or for adversarial robustness, i.e. on Stylized ImageNet [21] (ADV). We observe that robust and self-supervised models have fewer low-level concepts and compositionality than the originals, likely as their training yields less reliance on texture (stylization perturbs texture) and color (SimCLR training jitters color)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "To examine these patterns further, VCC visualizations for a CNN and transformer are shown in Fig. 7. Here, we limit to two (diverse) models and two later layers with class" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10901" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 227, + 228 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 227, + 228 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 227, + 228 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 227, + 228 + ], + "type": "image", + "image_path": "2136961e6f5b39aad7f0423052f19666731f87c8ad96e1697cdb4aa1961128a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 545, + 265 + ], + "type": "text", + "content": "Figure 8. Debugging model failure modes with VCCs. We show an image of a tricycle incorrectly predicted by a ResNet50 as a rickshaw (middle) as well as the top-down segmentation of the image (Sec. 3.1). We also show the incorrect (left) and correct (right) VCCs. Following the hierarchy of concepts reveals that the model incorrectly focused heavily on the collapsible hood, starting at Layer 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 69, + 376, + 227 + ], + "blocks": [ + { + "bbox": [ + 228, + 69, + 376, + 227 + ], + "lines": [ + { + "bbox": [ + 228, + 69, + 376, + 227 + ], + "spans": [ + { + "bbox": [ + 228, + 69, + 376, + 227 + ], + "type": "image", + "image_path": "b42ad8a22bfe6577980ecb7fc52666132f1ed6b77372bd56051da4ae76e3acde.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 377, + 70, + 535, + 227 + ], + "blocks": [ + { + "bbox": [ + 377, + 70, + 535, + 227 + ], + "lines": [ + { + "bbox": [ + 377, + 70, + 535, + 227 + ], + "spans": [ + { + "bbox": [ + 377, + 70, + 535, + 227 + ], + "type": "image", + "image_path": "5db879e63f4b8af358d713e5108927d602eb635369029d7736be6310fef398f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": "output for space. The connection diversity in the last layer is indeed observed to be larger for the transformer vs. the CNN. Notably, over half of the concepts in the CNN capture background centric concepts, while the transformer has only a single background centric concept." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 329, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 461 + ], + "type": "text", + "content": "All layer VCCs. We present quantitative analyses on all layer VCCs for three diverse models: (i) a standard CNN, VGG-16, (ii) an efficient model, MobileNetv3 and (iii) a transformer, MViT. Averages are taken over 10 VCCs for 10 random ImageNet classes. Results are shown in Fig. 6. Common trends appear in all models. Concept composition is non-linear across layers, with branching factor ranging from 5-15 and converging to approximately two near the last layers. The peak number of concepts, around 20, is consistently at " + }, + { + "bbox": [ + 46, + 329, + 287, + 461 + ], + "type": "inline_equation", + "content": "30 - 40\\%" + }, + { + "bbox": [ + 46, + 329, + 287, + 461 + ], + "type": "text", + "content": " of network depth and, as in the four-layer analysis, also converges to two in the final layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 629 + ], + "type": "text", + "content": "Edge weights and variances are in accord with the main findings of the four-layer analysis, but also reveal other insights into the compositional nature of the models. At fine grained, all layer analysis, each model more readily displays unique edge weight characteristics: VGG16's average weights decrease in later layers, MobileNetv3's drop greatly before the final layer and MViT maintains consistent values. Still, overall these results indicate that penultimate concepts differ between CNNs vs. transformers, as supported by our four-layer VCC analysis; see Fig. 5. Higher variances in initial layers suggest a diverse combination of concepts, whereas deeper layers indicate a more uniform composition. Transformers, however, show a variance increase in the final layer, indicating greater compositionality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "In summary, the four and all-layer VCC analyses consistently highlight key aspects of deep networks: (i) Early and mid-layers use more concepts compared to deeper layers. (ii) Concept interactions, e.g. ITCAV values and variances, differ greatly between adjacent layers, as opposed to across multiple layers. (iii) CNNs and transformers compose concepts differently, especially at the final layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 268, + 542, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 542, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 542, + 282 + ], + "type": "text", + "content": "5. Application: Diagnosing failure predictions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 288, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 359 + ], + "type": "text", + "content": "To show the VCC's practical utility, we consider application to model failure analysis. Contrasting failure analysis methods that do not explain how errors are made compositionally [46] or only offer explanations of single neuron heatmaps [13], VCCs provide insights on compositional concepts across layers and distributed representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "type": "text", + "content": "Figure 8 shows a 'Tricycle' incorrectly classified as a 'Rickshaw' by a ResNet50 model, and the corresponding incorrect VCC ('Rickshaw', left) and correct VCC ('Tricycle', right). As the image is decomposed using our top-down segmentation (Sec. 3.1), it is revealed that the majority of pooled segments are closer, in terms of " + }, + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "type": "text", + "content": " distance, to concepts in the Rickshaw VCC (red outlines) than the tricycle VCC (green outlines). While the model correctly encoded the wheel and handlebar portions of the images as tricycle concepts, the background and collapsible hood concepts are composed from layers two through four as rickshaw concepts, which may cause the error. We also note the lack of other tricycle-specific concepts (e.g. children)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 522, + 454, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 454, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 454, + 534 + ], + "type": "text", + "content": "6. Discussion and conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 542, + 545, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 709 + ], + "type": "text", + "content": "The Visual Concept Connectome (VCC) is a method for discovering human interpretable concepts and their interlayer connections in a fully unsupervised way. VCCs allow for the interpretation of deep models at multiple resolutions, revealing interesting properties via application to a range of networks. When comparing architecture types, CNNs were found to rely heavily on final layer concepts for prediction, while transformers were found to have more diverse connection strengths in the final layer and a more complex assembly of later concepts from earlier ones. We also applied VCCs to model debugging and showed how the hierarchical concept representation can indicate how mistakes arise. Given VCCs generality, our methodology should be applicable to additional networks, tasks and applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10902" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From “where” to “what”: Towards human-understandable explanations through concept relevance propagation. arXiv preprint arXiv:2206.03208, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 203, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 256 + ], + "type": "text", + "content": "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Conference on Computer Vision and Pattern Recognition, pages 6541-6549, 2017. 2, 5, 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[4] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In European Conference on Computer Vision, pages 351-369. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 303, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 357 + ], + "type": "text", + "content": "[5] Alexander Binder, Gregoire Montavon, Sebastian Lapischkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning, pages 63-71. Springer, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "text", + "content": "[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 522 + ], + "type": "text", + "content": "[7] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Christopher Olah. Towards monoseismicity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2023. https://transformer-circuits.pub/2023/monoseismicfeatures/index.html. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 524, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 287, + 568 + ], + "type": "text", + "content": "[8] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 569, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 569, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 569, + 288, + 624 + ], + "type": "text", + "content": "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In Winter Conference on Applications of Computer Vision, pages 839-847, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[12] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. Nature Machine Intelligence, 2(12):772-782, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[13] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain CNN via hierarchical decomposition. International Journal of Computer Vision, 131:1091-1105, 2023. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[14] European Commission. Laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts. European Commission, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 6, 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6, 4, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 287, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 373 + ], + "type": "text", + "content": "[17] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/toy_model/index.html. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. International Conference on Computer Vision, 2021. 5, 2, 3, 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Axel Pinz, Richard P Wildes, and Andrew Zisserman. Deep insights into convolutional networks for video recognition. International Journal of Computer Vision, 128:420-437, 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 466, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 521 + ], + "type": "text", + "content": "[20] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. CRAFT: Concept recursive activation factorization for explainability. In Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 522, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 577 + ], + "type": "text", + "content": "[21] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, 2018. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "text", + "content": "[22] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI Conference on Artificial Intelligence, pages 3681-3688, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "text", + "content": "[23] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems, 2019. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[24] Sven Ove Hansson, Matts-Åke Belin, and Björn Lundgren. Self-driving vehicles-An ethical overview. Philosophy & Technology, 34:1383-1408, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10903" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6, 7, 2, 4, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 150 + ], + "type": "text", + "content": "[26] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, and Vijay Vasudevan. Searching for MobilNetv3. In International Conference on Computer Vision, pages 1314-1324, 2019. 5, 3, 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 183 + ], + "type": "text", + "content": "[27] Haiyang Huang, Zhi Chen, and Cynthia Rudin. SegDiscover: Visual concept discovery via unsupervised semantic segmentation. arXiv preprint arXiv:2204.10926, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "text", + "content": "[28] Benjamin Irving. MaskSLIC: Regional superpixel generation with application to local pathology characterisation in medical images. arXiv preprint arXiv:1606.09518, 2016. 4, 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "text", + "content": "[29] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, and Fernanda Viegas. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In International Conference on Machine Learning, pages 2668-2677, 2018. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "type": "text", + "content": "[30] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schutt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267-280, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "text", + "content": "[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[32] Matthew Kowal, Achal Dave, Rares Ambrus, Adrien Gaidon, Konstantinos G Derpanis, and Pavel Tokmakov. Understanding video transformers via universal concept discovery. arXiv preprint arXiv:2401.10831, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[33] Satyapriya Krishna, Tessa Han, Alex Gu, Javin Pombra, Shahin Jabbari, Steven Wu, and Himabindu Lakkaraju. The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[34] Mengchen Liu, Jiaxin Shi, Kelei Cao, Jun Zhu, and Shixia Liu. Analyzing the training processes of deep generative models. Transactions on Visualization and Computer Graphics, 24(1):77-87, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "type": "text", + "content": "[35] Stuart Lloyd. Least squares quantization in PCM. Transactions on information theory, 28(2):129-137, 1982. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[36] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Conference on Computer Vision and Pattern Recognition, pages 5188-5196, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[37] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54 (6):1-35, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "text", + "content": "[38] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/features-visualization.1,2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[39] Chris Olah, Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, and Alexander Mordvintsev. The building blocks of interpretability. Distill, 2018. https://distill.pub/2018/building-blocks.1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[40] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 2020. https://distill.pub/2020/circuits/zoom-in. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "text", + "content": "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, and Luca Antiga. PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 261 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 262, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 306 + ], + "type": "text", + "content": "[43] Peter J Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "text", + "content": "[44] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetv2: Inverted residuals and linear bottlenecks. In Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 353, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 407 + ], + "type": "text", + "content": "[45] Anirban Sarkar, Deepak Vijaykeerthy, Anindya Sarkar, and Vineeth N Balasubramanian. A framework for learning antehoc explainable models via concepts. In Conference on Computer Vision and Pattern Recognition, pages 10286-10295, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 409, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 545, + 464 + ], + "type": "text", + "content": "[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision, pages 618-626, 2017. 1, 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 465, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 545, + 487 + ], + "type": "text", + "content": "[47] S. Seung. Connectome: How the Brain's Wiring Makes Us Who We Are. Houghton Mifflin Harcourt, 2012. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "text", + "content": "[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations, 2014. 5, 6, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 533, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 589 + ], + "type": "text", + "content": "[49] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2, 6, 7, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. Scikit-image: image processing in python. PeerJ, 2:e453, 2014. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 5, 7, 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "text", + "content": "[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech UCSD birds 2011 dataset. Technical" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10904" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 463 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 4, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "text", + "content": "[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "text", + "content": "[54] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pages 818-833, 2014. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 229 + ], + "type": "text", + "content": "[55] Quanshi Zhang, Ruiming Cao, Feng Shi, Ying Nian Wu, and Song-Chun Zhu. Interpreting CNN knowledge via an explanatory graph. In AAAI Conference on Artificial Intelligence, pages 4454-4463, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 286, + 274 + ], + "type": "text", + "content": "[56] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, pages 8827-8836, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 286, + 318 + ], + "type": "text", + "content": "[57] Quanshi Zhang, Yu Yang, Haotian Ma, and Ying Nian Wu. Interpreting CNNs via decision trees. In Conference on Computer Vision and Pattern Recognition, pages 6261-6270, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 319, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 286, + 374 + ], + "type": "text", + "content": "[58] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for CNN models with non-negative concept activation vectors. In AAAI Conference on Artificial Intelligence, pages 11682-11690, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 286, + 419 + ], + "type": "text", + "content": "[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Conference on Computer Vision and Pattern Recognition, pages 2921-2929, 2016. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 286, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 286, + 463 + ], + "type": "text", + "content": "[60] Bolei Zhou, Yiyou Sun, David Bau, and Antonio Torralba. Interpretable basis decomposition for visual explanation. In European Conference on Computer Vision, pages 119-134, 2018. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10905" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_content_list.json b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8f5b88afba4a663e835c771d5371df2562fd626 --- /dev/null +++ b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_content_list.json @@ -0,0 +1,1372 @@ +[ + { + "type": "text", + "text": "Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval", + "text_level": 1, + "bbox": [ + 204, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Young Kyun Jang\\*, Donghyun Kim\\*, Zihang Meng, Dat Huynh, and Ser-Nam Lim3 Meta AI, Korea University, University of Central Florida", + "bbox": [ + 142, + 202, + 826, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed Image Retrieval (CIR) is a task that retrieves images similar to a query, based on a provided textual modification. Current techniques rely on supervised learning for CIR models using labeled triplets of the . These specific triplets are not as commonly available as simple image-text pairs, limiting the widespread use of CIR and its scalability. On the other hand, zero-shot CIR can be relatively easily trained with image-caption pairs without considering the image-to-image relation, but this approach tends to yield lower accuracy. We propose a new semi-supervised CIR approach where we search for a reference and its related target images in auxiliary data and learn our large language model-based Visual Delta Generator (VDG) to generate text describing the visual difference (i.e., visual delta) between the two. VDG, equipped with fluent language knowledge and being model agnostic, can generate pseudo triplets to boost the performance of CIR models. Our approach significantly improves the existing supervised learning approaches and achieves state-of-the-art results on the CIR benchmarks.", + "bbox": [ + 75, + 306, + 473, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 638, + 209, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-to-image or text-to-image retrieval, where a query image/text is used to retrieve similar ones from a gallery, has grown into a pivotal research field with many practical applications [38]. However, relying solely on image queries is limiting, as they primarily retrieve similar images, making it challenging to understand the user's intent for modifications in the results. On the other hand, relying solely on text queries can also be restrictive, as it may not effectively convey the user's desired detailed visual contents. To address this, Composed Image Retrieval (CIR) was introduced [2, 36, 42, 52]. CIR seeks to retrieve images using a query that combines both an image and a textual description of the user's intent (referred to as the visual delta), which allows more flexible retrieval. Due to the convenience and diverse", + "bbox": [ + 75, + 665, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a50b27cee53aeea613fbe36603465063165538f1648a9eb6aa1cb4e1512ffc5.jpg", + "image_caption": [ + "Figure 1. An illustration of the data preparation process of (a) conventional supervised Composed Image Retrieval (CIR) vs. (b) our proposed semi-supervised CIR. While supervised CIR struggles to scale up due to high annotation costs, our semi-supervised method offers a cost-effective and scalable solution. It augments training samples efficiently by generating pseudo triplets through our Large Language Model (LLM)-based Visual Delta Generator." + ], + "image_footnote": [ + "(a) CIR triplet generation with human supervision (expensive).", + "(b) Visual Delta Generator for generating pseudo triplets." + ], + "bbox": [ + 506, + 271, + 888, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "applicability of CIR, it has attracted increased attention recently for a variety of real-world applications.", + "bbox": [ + 496, + 582, + 890, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing research on CIR has been developed under two major settings: (1) Supervised CIR: Learning with supervised triplets (i.e. ) [2, 11, 28, 35, 50] as shown in Fig. 1 (a), and (2) Zero-shot CIR: Learning with massive noisy pairs [3, 42, 49], without any CIR supervision. Supervised CIR would obviously yield much higher accuracy in retrieval but requires expensive two-stage data collection processes - collecting pairs of related reference and target images, and then annotating them with visual delta that depicts the difference between them. On the other hand, zero-shot CIR does not incur additional labeling costs and utilizes web-collected noisy image text caption pairs directly. However, it has a much lower performance bar compared to supervised approaches and lacks the ability to specialize in specific CIR domain tasks.", + "bbox": [ + 496, + 613, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we investigate a class of CIR called semi-supervised CIR, blending supervised and unsupervised samples to enhance generalization (Fig. 1(b)). This", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 94, + 886, + 253, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "method focuses on boosting CIR performance in specific retrieval domains by creating new triplets from unsupervised data. Building on this concept, we introduce a novel technique, Visual Delta Generator (VDG), designed to tap into the extensive natural language capabilities of Large Language Models (LLMs) [5, 21, 47, 48]. Our approach involves projecting reference and target images from supervised CIR triplets into the language embedding space, making them suitable inputs for the LLM. We then fine-tune the model by using prompts such as 'Describe the differences between images.' to induce it to yield human-like visual delta as illustrated in Fig. 2. Furthermore, we employ a parameter-efficient fine-tuning technique, LoRA [19], on the LLM. This choice of design effectively enhances the quality of visual deltas while also preserving the LLM's original capabilities, without harming its inherent knowledge.", + "bbox": [ + 75, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After the VDG is trained, it knows how to distinguish between a given reference and target image and produce visual delta as a textual response. If we forward two similar images with different compositions, we can thus obtain the corresponding visual delta easily with the VDG. This allows us to achieve two purposes. First, we can now augment existing CIR triplets by adding generated visual deltas to pairs of reference and target images from the training set. Second, we can also harvest new reference-target pairs from an unlabeled database based on visual similarity, after which we forward these images to VDG to configure new pseudo triplets for CIR training. Note that our VDG is model agnostic - it simply increases the number of triplet candidates for training any given supervised CIR baselines. This strategy strikes a balance between maintaining the integrity of supervisory concepts derived from a supervised dataset and the capacity for effortless expansion using new, unlabeled image samples. It's a cost-effective and scalable solution, ensuring uniformity in annotations across extensive datasets. By generating pseudo triplets with VDG, we significantly reduce annotation costs and enhance the performance of CIR models trained solely on supervised learning, as well as those trained without supervised triplets. Our approach leads to state-of-the-art results in CIR benchmarks.", + "bbox": [ + 75, + 349, + 472, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key advantages of our semi-supervised CIR include:", + "bbox": [ + 96, + 713, + 467, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to transfer knowledge from Large Language Models (LLMs) and connect it with semi-supervised Composed Image Retrieval (CIR).", + "- We propose a novel Visual Delta Generator (VDG) that generates synthetic visual deltas for augmenting the supervised dataset and allowing the integration of an auxiliary image gallery for CIR model training.", + "- Comprehensive experimental results confirm the effectiveness of our method, demonstrating state-of-the-art retrieval rankings and showcasing the potential of our work." + ], + "bbox": [ + 76, + 734, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13694db774af6b195b976f17288d5ec01aa18222a2b0d4febaca9dadd03cfb76.jpg", + "image_caption": [ + "Figure 2. An overview of the VDG tuning process. It includes (a) a vision projector and (b) a Large Language Model (LLM). The VDG is trained to produce visual delta that accurately describes the difference between a reference image and its corresponding target image." + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 874, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 398, + 640, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Composed Image Retrieval. The field of image retrieval has captured the interest of many researchers in our community [6, 38]. One notable area that has seen much progress recently is Composed Image Retrieval (CIR), a problem that focuses on retrieving images that best match a given pair of a query image and textual intent. Supervised CIR methods [2, 11, 35, 50] are trained on human-annotated triplets, consisting of a reference image, a target image, and their textual difference. On the other hand, zero-shot CIR [3, 9, 16, 42] operates without relying on human-guided descriptions of the differences between the two images. Instead, it uses noisy image-text pairs, aiming to find a function that can translate images into words. This approach, designed to discriminate subtle differences between images based solely on text captions, poses challenges, but also scales well, making it easy to add more data for CIR training. Addressing the limitations of both approaches, we explore the field of semi-supervised learning-based retrieval in this work, leveraging the strengths of both labeled and unlabeled data to enhance retrieval performance.", + "bbox": [ + 496, + 425, + 893, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised Learning. Semi-supervised learning has been an active research topic in visual recognition for a long time [4, 18, 20, 25-27, 39, 44, 46, 51]. Semi-supervised learning can be roughly categorized into two groups, consistency learning [25, 39, 46] and pseudolabeling based learning [27, 44]. Consistency-based methods, as the name implies, encourage consistency in the output of the model by adding noise to model weights or using Exponential Moving Average (EMA) from a teacher model to a student model. In pseudo-labeling based meth", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16806", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods, hard/soft pseudo-labels obtained from a pretrained model are assigned to unlabeled images [15, 44]. These pseudo-labels can be filtered using confidence thresholding or multi-view consistency. However, neither consistency-based nor pseudo labels methods are directly applicable to CIR. This is because the relative textual descriptions of the corresponding visual differences (visual delta) needed in CIR are not the intended outputs of these methods. In this work, we propose the Visual Delta Generator (VDG), a multi-modal pseudo-label generator. VDG processes two input images and generates text that describes their visual differences, making it an ideal candidate for constructing pseudo triplets for CIR.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal Models for Image-Text Retrieval. Models like CLIP [41] and BLIP [29] have showcased the advantages of training models on extensive image-text pairs, enabling precise alignment between language and vision representations that is crucial for image-text retrieval. Building on this, there have been significant advancements in utilizing Large Language Models (LLMs) for enhanced vision-language understanding [1, 10, 31, 33, 53]. Especially for CIR, Fromage [23] utilizes LLMs to directly produce embeddings for retrieval, allowing cross-modal compositional search with a single image and textual intent. CoVR [49] and SEARLE [3] apply LLMs to generate visual deltas using image captions without incorporating visual data, which limits the generation of accurate CIR training samples. Our proposed method overcomes these challenges by finetuning LLMs with the integration of a pretrained vision-language alignment module. This integration empowers LLMs to perceive and comprehend images, making them applicable even in scenarios with only image datasets. The method excels in generating accurate visual deltas, a key factor in training efficient external CIR models. These enhancements optimize the use of LLMs while ensuring computational efficiency, thereby expanding the versatility of LLMs in image-related tasks.", + "bbox": [ + 75, + 304, + 472, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 679, + 169, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Our goal is to establish a semi-supervised Composed Image Retrieval (CIR) system that merges image reference features with user textual descriptions to retrieve images from a large-scale database. We face a challenge in the limited availability of supervised triplets necessary for robust CIR model training. To overcome this limitation, we introduce a novel semi-supervised approach for CIR by leveraging an instruction-tuned Large Language Model (LLM), which we call Visual Delta Generator (VDG). The VDG learns to discriminate differences between two images and produces a textual response. This capability allows us to generate additional CIR training triplets, which in turn contributes to the development of more robust CIR models.", + "bbox": [ + 75, + 703, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sec. 3.1 provides detailed insights into the construction of the VDG. Sec. 3.2 describes the pseudo triplet generation process. Training of CIR models with pseudo triplets and our additional objective function for better optimization are described in Sec. 3.3.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visual Delta Generator Training", + "text_level": 1, + "bbox": [ + 498, + 174, + 785, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While semi-supervised learning methods have been actively developed for standard visual recognition tasks, these cannot be directly applied to CIR. In CIR, pseudo-label generation requires a detailed semantic understanding of two separate images such that their difference can be automatically expressed in the form of text. We leverage vision-language pretraining models and LLMs to achieve the requirements. With the huge success of LLMs, there are approaches that aim to utilize their understanding of the language domain for improving vision tasks. Particularly, LLaVA [33] and InstructBLIP [10] which are trained on top of the chat bot style instruction tuned LLM, Vicuna [7], have shown interesting results on vision-language tasks. Inspired by these, we propose VDG, which allows the LLM to take two images (reference, target) with similar contents and discriminate their difference in the form of text response (visual delta) as shown in Fig. 2.", + "bbox": [ + 496, + 196, + 890, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, to enable the LLM to interpret images, we use the Querying Transformer (Q-Former) motivated by InstructBLIP [10] to prepare images for the LLM input (i.e., Vision Projector (VP) in Fig. 2(a)). Q-Former, a transformer encoder [12, 21], processes a fixed set of 32 learnable query tokens (embeddings). These tokens are modified through self-attention layers and interact with image features of Vision Transformer (ViT) [12] through cross-attention layers. As a result, the query tokens are infused with the visual information from the provided image, making them suitable for LLM processing.", + "bbox": [ + 496, + 454, + 892, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage 1: Alignment. The outputs of the VP are inherently not aligned with the tokenized word embeddings of the inputs to the LLM, which we chose to be LLaMA2 [48] in this work. Alignment between the VP and LLM needs to be attained by fine-tuning trainable projection of VP in Fig. 2(a) as was done in LLaVA [33]. Specifically, we employ large-scale image-caption pairs to foster alignment between the image representations understood by the VP and the LLM. This step includes minimizing standard next token prediction loss which is generally used to train the decoder-based LLM as:", + "bbox": [ + 496, + 637, + 893, + 803 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L L M} = - \\sum_ {t = 1} ^ {T} \\log P \\left(\\mathrm {w} _ {t} \\mid x, I _ {\\text {i n s t}}, \\mathrm {w} _ {(1.., t - 1)}; \\theta_ {\\text {p r o j}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 823, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{w}_t$ denotes a token at time step $t$ , $T$ is the length of the sequence, $P(\\cdot)$ is the probability assigned by the model", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16807", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to the actual token $\\mathrm{w}_t$ given with the image $x$ , instruction $I_{inst}$ and previous tokens $\\mathrm{w}_{(1,\\dots,t - 1)}$ , and $\\theta_{proj}$ is parameters of projection layer. Following standard practice, the prediction for each token is computed using a softmax over the vocabulary of the LLM, and the cross-entropy loss is computed between the predicted probabilities and the textual token labels as a classification.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2: Instruction Tuning. In this stage, we conduct instruction tuning to equip our LLM with the ability to understand image pairs and produce visual delta, as shown in Fig. 2(b). Referring to Fig. 3, we train the LLM with an instruction (i.e., \"Request\") to generate the visual delta in textual format (i.e., \"Response\"). We forward two distinct images (i.e., \"Reference\" and \"Target\") into the LLM via the VP. We employ a parameter-efficient fine-tuning technique, LoRA [19], directly on the LLM. This fine-tuning process is designed to provide \"extra room\" for the LLM to undertake new tasks, all without compromising its foundational capabilities. In this stage, we introduce a specific prompt structure to the LLM as below:", + "bbox": [ + 75, + 215, + 472, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Request: Analyze given reference and target images and provide a description that transforms the reference to match the target.", + "bbox": [ + 94, + 429, + 444, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reference: ", + "bbox": [ + 98, + 455, + 346, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target: ", + "bbox": [ + 99, + 469, + 305, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Response: $<$ Visual delta (text label)>", + "bbox": [ + 99, + 481, + 313, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which guides the LLM to generate the corresponding visual delta. We leverage the same training loss from Eqn. 1 to train LoRA parameters.", + "bbox": [ + 75, + 536, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pseudo Triplet Generation for CIR", + "text_level": 1, + "bbox": [ + 76, + 590, + 382, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training VDG, we can generate visual delta of two images, which can be used to form pseudo triplets for CIR. However, it is important to choose two images that not only share certain attributes and similarities but also present other distinct attributes (i.e., visual delta). To gather suitable pairs of reference and target images, we utilize an image encoder to select them from an auxiliary image gallery, which we denote as $G'$ . Note that, we note upper strophe (') on samples and embeddings that are obtained from $G'$ , in the following.", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the strategy in CIRR [36], we start with an anchor image $x_{a}$ and retrieve the top 20 images from $G'$ using cosine similarity scores between ResNet 152 [17] embeddings, pretrained on ImageNet [24]. We exclude images with scores above 0.94 and sequentially add images to a subgroup of size 6, skipping those within a 0.002 score of the previously included image. As depicted in Fig. 4, we establish dense connections between all pairs (both consecutive, represented by the outer circle, and non-consecutive,", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b146655b3f6dbd991aec33738b9171248f9f038243e4aa288d6296248d943bac.jpg", + "image_caption": [ + "Figure 3. A template prompt for VDG instruction tuning.", + "Figure 4. The process of pseudo triplet generation. First, an image subgroup is constructed based on visual similarity (left). Then, paired reference and target images are fed into the VDG to generate the visual delta, completing the triplet formation (right)." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 875, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "represented by the dotted inner connections), while ensuring no overlaps. The arrow's starting point denotes the reference, while its endpoint indicates the target. Given images $x_{i}^{\\prime r}$ and $x_{j}^{\\prime t}$ , VDG produces $\\delta_{i,j}^{\\prime}$ . This allows us to formulate the pseudo triplet as $\\{x_{i}^{\\prime r}, x_{j}^{\\prime t}, \\delta_{i,j}^{\\prime}\\}$ .", + "bbox": [ + 496, + 301, + 890, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semi-supervised CIR Training", + "text_level": 1, + "bbox": [ + 500, + 388, + 769, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminaries. Suppose we have access to a CIR dataset of triplets: $D = \\{(x_i^r, x_j^t, \\delta_{n(i,j)})\\}_{n=1}^N$ where $x_i^r, x_j^t$ denote reference and target images, respectively, $\\delta_{n(i,j)}$ represents their visual delta, and $N$ denotes the total triplet counts.", + "bbox": [ + 496, + 411, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the pursuit of enhancing CIR through a semi-supervised approach, our method's strength lies in its model-agnosticism, allowing for seamless integration with a variety of CIR models. Following recent trends, we opt to use the encoders from vision-language pretraining models [10, 29, 31, 41], notably CLIP and BLIP, as our baseline backbones. These encoders are naturally equipped to understand and convert both visual and textual elements into a joint embedding space, making them suitable for CIR tasks.", + "bbox": [ + 496, + 472, + 892, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The image encoder takes patchified image token embeddings $x = [\\mathrm{x}_1, \\dots, \\mathrm{x}_{K_{img}}]$ with the learnable image $cls$ token embedding $[\\mathrm{x}_{cls}]$ , and outputs the visual feature embeddings $E_{img}(\\mathrm{x}_{cls}, x) = [\\hat{\\mathrm{x}}_{cls}, \\hat{\\mathrm{x}}_1, \\dots, \\hat{\\mathrm{x}}_{K_{img}}}$ where $\\hat{\\mathrm{x}} \\in R^{d_i}$ of $d_i$ dimension, and $K_{img}$ denotes the number of generated tokens from each image. The text encoder processes tokenizer output embedding of visual delta $\\delta = [\\mathrm{z}_1, \\dots, \\mathrm{z}_{K_{txt}}}$ with the learnable text $cls$ token embeddings $[\\mathrm{z}_{cls}]$ to produce its textual feature embeddings $E_{txt}(\\mathrm{z}_{cls}, \\delta) = [\\hat{\\mathrm{z}}_{cls}, \\hat{\\mathrm{z}}_1, \\dots, \\hat{\\mathrm{z}}_{K_{txt}}}$ , where $\\hat{\\mathrm{z}} \\in R^{d_t}$ of $d_t$ dimension, and $K_{txt}$ denotes the number of generated text tokens from each visual delta.", + "bbox": [ + 496, + 609, + 893, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Architecture. To carry out CIR, we establish a fusion function $f$ that takes a reference image $x^r$ and a visual delta $\\delta$ , producing a composed embedding $\\mathbf{c}$ as: $f(x^r, \\delta) = \\mathbf{c}$ , and notates its trainable components as $f_\\theta$ . We utilize two well-known backbones, CLIP and BLIP, to configure $f$ . In the case of CLIP, we employ the text encoder $E_{txt:\\theta}$ and an", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16808", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "additional Combiner module [2] $\\mathcal{C}_{\\theta}(\\hat{\\mathrm{x}}_{cls}^{r},\\hat{\\mathrm{z}}_{cls})$ that outputs $\\mathbf{c}$ to be the components of $f_{\\theta}$ (i.e., $f_{\\theta}^{\\mathrm{CLIP}} = \\{E_{txt:\\theta},\\mathcal{C}_{\\theta}\\}$ ). The Combiner is designed to optimally blend $\\hat{\\mathrm{x}}_{cls}$ and $\\hat{\\mathrm{z}}_{cls}$ carefully weighing their individual impacts while adeptly mixing them. In the BLIP case, we exclusively use BLIP's text encoder, grounded in the image, and designate it as the trainable $E_{txt:\\theta}$ (i.e., $f_{\\theta}^{\\mathrm{BLIP}} = E_{txt:\\theta}$ , and $\\mathbf{c} = \\hat{\\mathbf{z}}_{cls}$ ), without incorporating additional modules. BLIP's text encoder inherently fuses image and text signals in its cross-attention layer, eliminating the need for a separate combining module. Notably, we freeze all vision encoders in our setup to ensure compatibility with existing image retrieval galleries and to enhance training efficiency.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised / Pseudo Separated Contrastive Loss. The training objective of CIR is to achieve strong alignment between the target image's embedding $\\mathbf{x}$ (where $\\mathbf{x} = \\hat{\\mathbf{x}}_{cls}^{t}$ for simplicity), and the composed embedding $\\mathbf{c}$ . On this purpose, we utilize HN-NCE [40] loss for a given training batch $\\mathcal{B} \\sim D$ and $\\mathcal{B}' \\sim D'$ , where $D'$ contains pseudo triplets. Additionally, to mitigate the impact of noise in pseudo triplets and ensure consistent contributions from supervised triplets, we compute a target-composed contrastive loss (tcc) as:", + "bbox": [ + 75, + 305, + 470, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t c c} (\\mathcal {B}, \\mathcal {B} ^ {\\prime}) = \\mathcal {L} _ {c} (\\mathcal {B}; \\tau) + \\mathcal {L} _ {c} (\\mathcal {B} \\oplus \\mathcal {B} ^ {\\prime}; \\tau) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 470, + 468, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c$ is defined as:", + "bbox": [ + 76, + 494, + 233, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {c} (\\mathcal {B}) = - \\sum_ {i = 1} ^ {n} \\log \\left[ \\frac {e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau}}{\\alpha \\cdot e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau} + \\sum_ {j \\neq i} ^ {\\mathcal {B}} e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {j} / \\tau} \\cdot w _ {\\mathbf {x} _ {i} , \\mathbf {c} _ {j}}} \\right] \\\\ - \\sum_ {i = 1} ^ {n} \\log \\left[ \\frac {e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau}}{\\alpha \\cdot e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau} + \\sum_ {j \\neq i} ^ {\\mathcal {B}} e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {j} / \\tau} \\cdot w _ {\\mathbf {x} _ {j} , \\mathbf {c} _ {i}}} \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 531, + 473, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau, \\alpha$ denote hyper-parameters, and $\\oplus$ denotes concatenation along the batch axis, and $w_{\\mathbf{x}_i,\\mathbf{c}_j}, w_{\\mathbf{x}_j,\\mathbf{c}_i}$ are set as in [40]. This design facilitates the independent yet concurrent investigation of both supervised and semi-supervised CIR embedding spaces in a contrastive manner.", + "bbox": [ + 75, + 625, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target-Delta Matching Loss. In the case of BLIP's image-grounded text encoder structure, we introduce a new target-delta matching loss (tdm). Our insight stems from the observation that, while all reference image patch tokens are considered in the cross-attention layer of the BLIP text encoder, target image patch tokens are overlooked when training solely with contrastive learning betweencls token embeddings. As an example shown in Fig. 5, the visual delta can be seen as a weakly correlated caption to the target image. Thus, we aim to align the target image and visual delta to enable the BLIP text encoder to identify the image tokens related to textual input. The tdm loss is applied as:", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9acc4d73c9bb68c4f258615b7f3979d99bf5767bc72d0f8adbef05669e037d6.jpg", + "image_caption": [ + "Figure 5. Illustration of our proposed adaptation of the BLIP image-grounded text encoder for CIR. Both reference $(x^r)$ and target image $(x^t)$ patch tokens are processed by the text encoder $(f_{\\theta})$" + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 875, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t d m}} = \\mathbb {E} _ {(x ^ {t}, \\delta) \\sim (D, D ^ {\\prime})} \\mathrm {H} \\left(\\mathbf {y} ^ {t d m}, \\mathbf {p} ^ {t d m} \\left(x ^ {t}, \\delta\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 463, + 890, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{H}$ is cross entropy, $\\mathbf{y}^{tdm}$ is a 2-dimensional one-hot vector label obtained through the hard negative mining process proposed in [30], where a pair of image and text is configured as matched or not. $\\mathbf{p}^{tdm} \\in \\mathbb{R}^2$ generates both positive and negative target-delta matching scores and is computed as:", + "bbox": [ + 496, + 493, + 890, + 584 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {t d m} (x ^ {t}, \\delta) = \\sum_ {i = 1} ^ {K ^ {t x t}} p _ {\\theta} (\\hat {\\mathbf {z}} _ {i} ^ {t}) / K ^ {t x t},\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 593, + 820, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{z}_i^t$ denotes the $i$ -th token embedding in $f_{\\theta}(x^{t},\\delta)$ and $p_{\\theta}$ is a FC layer that outputs a 2-D vector. This loss ensures the text encoder fully processes the target image features, improving its understanding of the visual delta.", + "bbox": [ + 496, + 632, + 893, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 707, + 633, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 4.1 details the setup and VDG generation. Sec. 4.2 covers quality checks of visual deltas. Evaluations and baseline comparisons are in Sec. 4.3, and further analyses in Sec. 4.4.", + "bbox": [ + 496, + 731, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 786, + 581, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We utilize InstructBLIP's pretrained weights [10], which have been trained with both a ViT-G/14 [13] and the Q-Former based on the Vicuna-13B model [7], without using prompts for Q-Former. For the stage 1 training described in Sec. 3.1, we employ 595K filtered image-text pairs from CC3M [43] provided by [33] to", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16809", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "find alignment with our baseline LLM, LLaMA2-13B [48]. In Stage 2, we implement instruction tuning with LoRA parameters [19], following the fixed prompt as outlined in Fig. 3. Visual deltas are generated in an autoregressive manner, predicted based on the LLaMA2 vocabulary. For the CLIP-based CIR training, we select the ViT-L/14 model combined with a Combiner [2]. For the BLIP-based model [29], we use a dedicated BLIP text encoder for image-text matching, paired with the ViT-L/16 model. Additional details can be found in the appendix.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets for CIR Evaluation. There are two standard benchmarks in CIR, one is CIRR [36] which deals with natural images, and the other is FashionIQ [52] which focuses on fashion domain images. Each presents unique challenges and datasets that help researchers push the boundaries of what's possible in CIR. Following the protocols utilized in benchmarks [36, 52], we report the CIR results with recall scores at top K retrieval results (R@K), or results under collected subset $(\\mathrm{R}_s@\\mathrm{K})$ . Specifically, CIRR is configured with 4,351 subgroups (subsets), each containing six similar images, sourced from NLVR2 [45]. For experimental purposes these groups are distributed into train (3,345 subgroups of 16,742 images), validation (503 subgroups of 2,265 images), and test (503 subgroups of 2,178 images) sets. FashionIQ is divided into three categories of Dress, Shirt, and Toptee (Tops and Tees). The reference and target images are paired based on their category similarities. The 18,000 CIR triplets for training are pooled from 45,429 images of the training set, and 6,016 CIR triplets for the test are chosen from 15,415 images of the validation set.", + "bbox": [ + 75, + 268, + 472, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Delta Generation. To produce pseudo triplets, we expand our dataset sources to configure an auxiliary gallery $(G^{\\prime})$ which is built upon the grouping strategy introduced in Sec. 3.2, while excluding images that overlap with the benchmark sets. Once the subgroups are constructed, we further filter them to avoid heavy overlap. In total, we draw upon 42,390 unique subgroups from NLVR2 [45], and 79,427 from COCO [32]. Similarly, we build 27,957 individual subgroups from FashionIQ [52], and 30,880 from DeepFashion [34]. We mark upper strophe (') to these datasets. For the semi-supervised settings, we randomly select 3,345 groups from NLVR2' and COCO' for the CIRR case, as well as 3,600 groups from FashionIQ' and DeepFashion'. We denote these datasets used for semi-supervised CIR with the subscript $se$ , (e.g., NLVR2'se, COCO'se). Fig. 6 shows the comparison between human-annotated visual deltas and those generated by the VDG in both natural and fashion domains. We observe that the VDG is effective in generating high-quality visual deltas - additional results can be found in the appendix.", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/147fb22b2bec61b0e22255dd01a8ba62a80aa236fc914815918287634d2d9360.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison on visual deltas, human vs. VDG on CIRR and FashionIQ datasets. For both natural and fashion domain images, VDG can produce informative visual deltas." + ], + "image_footnote": [], + "bbox": [ + 521, + 88, + 870, + 268 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb3f4c256f41aa2cece4799bb99f759b6c85fd8551f8b7e6b8e4ca3ba265ac4f.jpg", + "table_caption": [ + "Table 1. Retrieval results on CIRR validation set. Human + VDG represents utilizing both human annotated and VDG-generated visual deltas when training the CIR model. $\\mathrm{BLIP}_{tdm}$ represents a model with target-delta matching loss. The best scores for each Val. set supervision are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Training Set SupervisionVal. Set SupervisionBaselineR@1R@5R@10R@50
(a) HumanHumanCombiner37.9871.4982.5295.29
BLIPtdm53.1782.0989.8197.54
(b) VDGHumanCombiner35.4768.2980.2994.31
BLIPtdm50.1680.0387.7896.75
(c) Human + VDGHumanCombiner39.1173.0284.4195.96
BLIPtdm53.6782.9989.9797.84
(d) HumanVDGCombiner38.9673.3184.3696.22
BLIPtdm51.6483.3391.3997.90
(e) VDGVDGCombiner41.5977.5787.3597.11
BLIPtdm52.6985.1792.3798.59
(f) Human + VDGVDGCombiner41.4777.9087.4297.30
BLIPtdm52.9585.2792.5498.61
", + "bbox": [ + 509, + 410, + 879, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quality Check of VDG Responses", + "text_level": 1, + "bbox": [ + 500, + 604, + 795, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the quality of VDG-generated visual deltas, we executed a series of experiments as outlined in Table 1. We use two backbones, Combiner, and BLIP with the tdm loss (denoted as $\\mathrm{BLIP}_{\\mathrm{tdm}}$ ), as our baselines for evaluation.", + "bbox": [ + 496, + 627, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initially, we compare the visual deltas generated by VDG with those annotated by humans. This is done by replacing the deltas in the training set with VDG-generated deltas for the same reference-target pairs. When comparing (a) and (b), we observe only a marginal drop in performance upon switching to VDG-generated deltas. This indicates that the deltas generated by VDG are as effective as those created by humans. More importantly, the improved performance observed when comparing (a) with (c) — which combines human-annotated and VDG-generated deltas — highlights the effectiveness of this hybrid approach as a robust data augmentation strategy to enhance CIR models.", + "bbox": [ + 496, + 688, + 892, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate VDG's performance on new image pairs, we conduct experiments by substituting the human-annotated", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16810", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bcbd7fd9c426ce0788ffda519723fa9cfd77dbffb43d111d9dba683eb21ba3a9.jpg", + "table_caption": [ + "Table 2. Retrieval results on CIRR test set. * denotes our baselines, † denotes VDG generated visual deltas are applied to augment original training set. We categorize into two distinct groups: one is Seen: Supervised / Supervised + External / Supervised + Aux. with VDG which utilize human-annotated visual delta for CIR model training, and the other is Unseen: Zero-shot / Aux. with VDG which does not utilize human-annotated visual delta for CIR model training. Within each category, best viewed with **bold**." + ], + "table_footnote": [], + "table_body": "
MethodR@1R@5R@10R@50Rs@1Rs@2Rs@3
(a) Supervised
ARTEMIS [11]16.9646.1061.3187.7339.9962.2075.67
CIRPLANT [35]19.5552.5568.3992.3839.2063.0379.49
Combiner [2]33.5965.3577.3595.2162.3981.8192.02
CASE [28]48.0079.1187.2597.5775.8890.5896.00
CoVR [49]48.8478.0586.1094.1975.7888.2292.80
Combiner34.3966.2276.5891.0468.5586.3693.98
\\( Combiner^† \\)36.9169.2179.5492.0470.0087.4594.39
\\( BLIP_{tdm} \\)48.9477.8386.1594.1775.7189.7195.81
\\( BLIP_{tdm}^† \\)49.0878.9886.8994.2476.1890.6295.86
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]49.3580.0288.7597.4776.4890.3795.71
CoVR + WebVid [49]49.6978.6086.7794.3175.0188.1293.16
(c) Supervised + Auxiliary Gallery with VDG
\\( Combiner^† + COCO_{se}^{\\prime} \\)38.7769.2579.2191.5271.2587.4994.34
\\( Combiner^† + NLVR2_{se}^{\\prime} \\)38.8969.8479.4191.1871.9287.8994.29
\\( BLIP_{tdm}^† + COCO_{se}^{\\prime} \\)49.3778.1285.5293.7476.6890.4696.05
\\( BLIP_{tdm}^† + NLVR2_{se}^{\\prime} \\)50.9680.1586.8694.4677.4590.6596.10
(d) Zero-shot
Pic2Word by CC3M [42]23.9051.7065.3087.80---
SEARLE by ImageNet [3]24.2252.4166.2988.6353.7174.6387.61
CASE by LasCo.Ca. [28]35.4065.7878.5394.6364.2982.6691.61
CoVR by WebVid [49]38.4866.7077.2591.4769.2883.7691.11
(e) Auxiliary Gallery with VDG
Combiner by COCO'26.8054.0565.3083.8867.2885.4292.99
Combiner by NLVR2'31.5761.3772.1088.9467.9886.1893.18
\\( BLIP_{tdm} \\)by COCO'43.4972.0781.5993.2172.3687.9894.72
\\( BLIP_{tdm} \\)by NLVR2'45.7475.0182.5293.1372.6087.9094.77
", + "bbox": [ + 76, + 204, + 480, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "deltas in the validation set with deltas generated by VDG. The comparison of (a) and (d) reveals comparable performance levels, suggesting that deltas generated by VDG are as effective as human annotations. When we incorporate VDG-generated deltas into the training set (cases (e) and (f)), we achieve a notable performance improvement over the human-annotated validation set across all metrics, with the exception of a small reduction in $\\mathbb{R}@\\mathbb{1}$ . For example, there is a $5.14\\%$ p increase in $\\mathbb{R}@\\mathbb{5}$ between (b) and (e). The improvement observed may stem from the varied text descriptions by different human annotators, leading to inconsistencies in style and detail that potentially affect the distribution of visual differences between the annotated triplets. VDG offers a more consistent solution in generating visual deltas, thereby reducing errors associated with differences in human annotation. Our comprehensive testing demonstrates VDG's reliability as an annotation tool for CIR.", + "bbox": [ + 75, + 573, + 472, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with Other Methods", + "text_level": 1, + "bbox": [ + 76, + 845, + 369, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tables 2 and 3 present comparative evaluations between our method and other CIR approaches on CIRR and Fash", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9388c2d17df4469751d9e58f2d7f20090b9757a4dd36f8a98ae25a0f1530df37.jpg", + "table_caption": [ + "Table 3. Retrieval results on FashionIQ validation set. We utilize the same notations and same categorizations as Table 2." + ], + "table_footnote": [], + "table_body": "
MethodDressShirtToptee
R@10R@50R@10R@50R@10R@50
(a) Supervised
ARTEMIS [11]27.1652.4021.7843.6429.2053.83
DCNet [22]28.9556.0723.9547.3030.4458.29
FashionVLP [14]32.4260.2931.8958.4438.5168.79
Combiner [2]31.6356.6736.3658.0038.1962.42
CoVR [49]43.5167.9448.2866.6851.5373.60
*Combiner31.9555.0539.2156.8238.5562.16
*Combiner†35.4059.9942.3061.6343.0966.96
*BILIPtdm44.8766.8349.6166.9350.5472.26
*BILIP†tdm46.9068.8650.2868.0452.7374.45
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]47.4469.3648.4870.2350.1872.24
CoVR + WebVid [49]44.5569.0348.4367.4252.6074.31
(c) Supervised + Auxiliary Gallery with VDG
Combiner† + DeepFashion′se35.5060.0942.5462.8643.2767.86
Combiner† + FashionIQ′se36.3060.1943.9862.2744.3368.06
BILIPtdm + DeepFashion′se47.1069.1049.9569.9653.9074.35
BILIP†tdm + FashionIQ′se47.8969.8151.3671.0853.2974.65
(d) Zero-shot
Pic2Word by CC3M [42]20.0040.2026.2043.6027.9047.40
SEARLE by ImageNet [3]20.3243.1827.4345.6829.3250.17
CoVR by WebVid [49]21.9539.0530.3746.1230.7848.73
(e) Auxiliary Gallery with VDG
Combiner by DeepFashion'23.3046.3630.8649.0231.8751.96
Combiner by FashionIQ'28.2651.4632.5851.2834.8855.79
BILIPtdm by DeepFashion'32.6754.3935.4855.0539.4759.92
BILIP†tdm by FashionIQ'37.4858.7037.2957.1142.1262.32
", + "bbox": [ + 500, + 119, + 901, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ionIQ evaluation protocols. We categorize the experiments into two distinct groups. The first group, Seen, includes scenarios where the CIR model is trained with human-annotated training triplets. In contrast, the second group, Unseen, comprises scenarios in which the CIR model is trained without human-annotated triplets. Our implementations of Combiner and BLIP are denoted by grey box. Specifically, methods under (a) are trained with supervised triplets, while those notated with the symbol $(\\dagger)$ are additionally augmented with VDG-generated visual deltas from supervised image pairs (i.e., the same setting as Human + VDG in Table 1). The methods in (b) have been pretrained on external datasets for CIR tasks, resulting in enhanced performances. In our semi-supervised setup (c), the models are trained with augmented supervised triplets and pseudo triplets sub-sampled from the auxiliary gallery as described in Secs. 3.2 and 4.1. Methods in (d) are trained with large-scale datasets with the intention of producing zero-shot performances. In (e), we only use pseudo triplets generated from the auxiliary gallery for training.", + "bbox": [ + 496, + 518, + 893, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results clearly demonstrate that our VDG implementation significantly improves the performance of both Combiner and $\\mathrm{BLIP}_{tdm}$ baselines. In (a), a detailed comparison shows that models augmented with VDG samples (with $\\dagger$ ) consistently outperform their counterparts (without $\\dagger$ ) in all", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16811", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53295b57ed81bf034a4429dbe88ed736c84dc39e29d12f743e4fbd1aa17ac7b6.jpg", + "table_caption": [ + "Table 4. Ablation results on $\\mathrm{BLIP}_{tdm}$ baseline for CIRR validation set with NLVR2' as auxiliary gallery. Best viewed with bold." + ], + "table_footnote": [], + "table_body": "
groupingconcat.tdm lossR@1R@5R@10R@50
52.4581.1888.62,96.88
✓55.6683.4591.2197.33
✓✓56.2884.3991.4197.92
✓✓✓57.8885.5893.2198.33
", + "bbox": [ + 112, + 114, + 424, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/612c8710939a4a85a99ea0a46c08a4506ddcf9a947748bb2a0738cb7373a7950.jpg", + "image_caption": [ + "(a) CIRR validation", + "Figure 7. Analysis on the scale for (c) Supervised + Auxiliary Gallery with $VDG$ of $\\mathrm{NLVR2}_{se}^{\\prime}$ and FashionIQ'se. The x-axis denotes the ratio of additional images to training images.", + "Table 5. Analysis on the scale for (e) Auxiliary Gallery with VDG. The dataset is scaled from one eighth (1/8) to the full set (1). Evaluated on CIRR Test set." + ], + "image_footnote": [], + "bbox": [ + 81, + 183, + 267, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02d04f82ef67ea5b2423f043e1f9a3a81527c27574530da36b44a178f40654af.jpg", + "image_caption": [ + "(b) FashionIQ-Shirt" + ], + "image_footnote": [], + "bbox": [ + 272, + 183, + 457, + 255 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b181be67b10a668cccf4b5f0d2d71870caa8be75417d8a96d4b0f67778f916e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
R@11/81/41/21R@101/81/41/21
COCO42.1542.8943.1043.49COCO79.8181.0181.1681.59
NLVR244.3144.9945.1345.74NLVR282.4882.7582.4382.52
", + "bbox": [ + 78, + 364, + 470, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "recall metrics. Moreover, in (c), when we enhance our CIR baselines with additional pseudo triplets, there is a notable performance improvement. Specifically, when $\\mathrm{BLIP}_{tdm}$ is combined with the auxiliary gallery, it surpasses the previous state-of-the-art results in most recall metrics, showcasing the advantages of VDG. Additionally, it's important to highlight that our semi-supervised CIR method achieves these improvements with considerably fewer images than methods like CASE or CoVR, demonstrating its efficiency.", + "bbox": [ + 75, + 412, + 468, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the unseen scenario, our implemented Combiner baseline not only outperforms Pic2Word, which uses the same CLIP backbone, but our $\\mathrm{BLIP}_{tdm}$ also significantly exceeds the former best-performing SEARLE or CoVR in most metrics. VDG's key advantage is its ability to generate visual deltas that align with the targeted domain, such as using COCO for natural images and DeepFashion for fashion images, by solely utilizing images without any accompanying captions. As a result, VDG significantly reduces the number of training samples required, yet achieves enhanced retrieval performances compared to zero-shot approaches like Pic2Word, CASE, or CoVR. This demonstrates that by focusing on similar domain, image-only datasets, we can substantially improve the efficacy of CIR model training.", + "bbox": [ + 75, + 550, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further Analyses", + "text_level": 1, + "bbox": [ + 76, + 771, + 246, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study on Each Component. To showcase the effectiveness of our proposed learning strategies, we conducted an ablation study and presented the findings in Table 4. We consider three scenarios: (1) without grouping, where we randomly sample images from the auxiliary gallery to create pairs; (2) without concat., where we remove the concatenation step used in Eqn. 2 and use pseudo triplets only;", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94c62f3e17c01e2d5119a6c86fea7816a6988019622ffb34814861089bf7b592.jpg", + "image_caption": [ + "Figure 8. Retrieval results on auxiliary galleries, COCO, and DeepFashion. Actual user intents are used as text queries." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 875, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and (3) without $tdm$ loss, where we exclude the tdm loss during training. The results indicate that each element sufficiently contributes to enhancing the performance.", + "bbox": [ + 496, + 260, + 890, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Generated Data Scale. Figs. 7a and 7b demonstrate the effect of the size of VDG-generated data. In a semi-supervised setup, we find that the performance peaks when the number of pseudo-selected images approximates the size of the training set, and then saturates. This saturation might arise because additional pseudo triplets fail to represent the test sample distribution, particularly as we rigorously remove overlapping images with the test set when forming $G'$ . For the unseen case as shown in Table 5, performance is enhanced with an increase in VDG-generated data. This suggests that even in the absence of supervised triplets for CIR model training, a larger set of pseudo triplets not only expands the model's comprehension of visual compositions but also aligns more closely with the distribution of supervised triplets. This alignment facilitates improved targeted domain retrieval performances, striking a balance between generalization and performance enhancement.", + "bbox": [ + 496, + 306, + 892, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Retrieval Results. We facilitate retrieval by providing user intents along with query images, drawing from auxiliary galleries. The retrieval results, as depicted in Fig. 8, demonstrate accurate and relevant image retrieval. Additional results are detailed in the appendix.", + "bbox": [ + 496, + 561, + 890, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 651, + 617, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we investigate a novel semi-supervised learning approach in the context of Composed Image Retrieval (CIR). Our findings reveal that integrating human-annotated data with pseudo triplets generated by the Visual Delta Generator (VDG) significantly enhances the generalization capacity of CIR models. The VDG approach not only streamlines the generation of visual deltas but also emerges as a cost-efficient and effective alternative to extensive human annotation. This work paves the way for future research in semi-supervised learning and showcases VDG as a promising direction for advancing CIR systems.", + "bbox": [ + 496, + 676, + 890, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Acknowledgment: This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University)).", + "bbox": [ + 500, + 851, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16812", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3", + "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 2, 5, 6, 7", + "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 7", + "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS, 32, 2019. 2", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2", + "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, pages 104-120. Springer, 2020. 2", + "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 3, 5", + "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2", + "[9] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In ECCV, pages 558-577. Springer, 2022. 2", + "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 4, 5", + "[11] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 2, 7", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3", + "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 5", + "[14] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, pages 14105-14115, 2022. 7", + "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3", + "[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4", + "[18] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised semantic segmentation. NeurIPS, 2015. 2", + "[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 4, 6", + "[20] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, pages 3420-3429, 2020. 2", + "[21] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2, 3", + "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 7", + "[23] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3", + "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 4", + "[25] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.2", + "[26] Semi-Supervised Learning. Semi-supervised learning. CSZ2006. html, 2006.", + "[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2", + "[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and early fusion for composed image retrieval. arXiv preprint arXiv:2303.09429, 2023. 1, 7", + "[29] Junnan Li and et al. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3, 4, 6", + "[30] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16813", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 5", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. 2023. 3, 4", + "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6", + "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 5, 1", + "[34] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6", + "[35] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 1, 2, 7", + "[36] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In CVPR, pages 2125–2134, 2021. 1, 4, 6", + "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 1", + "[38] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 129:23-79, 2021. 1, 2", + "[39] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 41(8):1979-1993, 2018. 2", + "[40] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, pages 6967-6977, 2023. 5", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 4", + "[42] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, pages 19305-19314, 2023. 1, 2, 7", + "[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 5", + "[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 33:596-608, 2020. 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 6", + "[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 30, 2017. 2", + "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6", + "[49] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 1, 3, 7", + "[50] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1, 2", + "[51] Haokun Wen, Xuemeng Song, Jianhua Yin, Jianlong Wu, Weili Guan, and Liqiang Nie. Self-training boosted multifaceted matching network for composed image retrieval. arXiv preprint arXiv:2305.09979, 2023. 2", + "[52] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 6", + "[53] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16814", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_model.json b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb635249de9ece8c21c9f6e4f8eea8da8650fb4a --- /dev/null +++ b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.828, + 0.24 + ], + "angle": 0, + "content": "Young Kyun Jang\\*, Donghyun Kim\\*, Zihang Meng, Dat Huynh, and Ser-Nam Lim3 Meta AI, Korea University, University of Central Florida" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.474, + 0.61 + ], + "angle": 0, + "content": "Composed Image Retrieval (CIR) is a task that retrieves images similar to a query, based on a provided textual modification. Current techniques rely on supervised learning for CIR models using labeled triplets of the . These specific triplets are not as commonly available as simple image-text pairs, limiting the widespread use of CIR and its scalability. On the other hand, zero-shot CIR can be relatively easily trained with image-caption pairs without considering the image-to-image relation, but this approach tends to yield lower accuracy. We propose a new semi-supervised CIR approach where we search for a reference and its related target images in auxiliary data and learn our large language model-based Visual Delta Generator (VDG) to generate text describing the visual difference (i.e., visual delta) between the two. VDG, equipped with fluent language knowledge and being model agnostic, can generate pseudo triplets to boost the performance of CIR models. Our approach significantly improves the existing supervised learning approaches and achieves state-of-the-art results on the CIR benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.64, + 0.21, + 0.655 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Image-to-image or text-to-image retrieval, where a query image/text is used to retrieve similar ones from a gallery, has grown into a pivotal research field with many practical applications [38]. However, relying solely on image queries is limiting, as they primarily retrieve similar images, making it challenging to understand the user's intent for modifications in the results. On the other hand, relying solely on text queries can also be restrictive, as it may not effectively convey the user's desired detailed visual contents. To address this, Composed Image Retrieval (CIR) was introduced [2, 36, 42, 52]. CIR seeks to retrieve images using a query that combines both an image and a textual description of the user's intent (referred to as the visual delta), which allows more flexible retrieval. Due to the convenience and diverse" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.272, + 0.889, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.506, + 0.421, + 0.699, + 0.445 + ], + "angle": 0, + "content": "(a) CIR triplet generation with human supervision (expensive)." + }, + { + "type": "image_footnote", + "bbox": [ + 0.701, + 0.421, + 0.887, + 0.445 + ], + "angle": 0, + "content": "(b) Visual Delta Generator for generating pseudo triplets." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Figure 1. An illustration of the data preparation process of (a) conventional supervised Composed Image Retrieval (CIR) vs. (b) our proposed semi-supervised CIR. While supervised CIR struggles to scale up due to high annotation costs, our semi-supervised method offers a cost-effective and scalable solution. It augments training samples efficiently by generating pseudo triplets through our Large Language Model (LLM)-based Visual Delta Generator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.891, + 0.613 + ], + "angle": 0, + "content": "applicability of CIR, it has attracted increased attention recently for a variety of real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Existing research on CIR has been developed under two major settings: (1) Supervised CIR: Learning with supervised triplets (i.e. ) [2, 11, 28, 35, 50] as shown in Fig. 1 (a), and (2) Zero-shot CIR: Learning with massive noisy pairs [3, 42, 49], without any CIR supervision. Supervised CIR would obviously yield much higher accuracy in retrieval but requires expensive two-stage data collection processes - collecting pairs of related reference and target images, and then annotating them with visual delta that depicts the difference between them. On the other hand, zero-shot CIR does not incur additional labeling costs and utilizes web-collected noisy image text caption pairs directly. However, it has a much lower performance bar compared to supervised approaches and lacks the ability to specialize in specific CIR domain tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this paper, we investigate a class of CIR called semi-supervised CIR, blending supervised and unsupervised samples to enhance generalization (Fig. 1(b)). This" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.254, + 0.901 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16805" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "method focuses on boosting CIR performance in specific retrieval domains by creating new triplets from unsupervised data. Building on this concept, we introduce a novel technique, Visual Delta Generator (VDG), designed to tap into the extensive natural language capabilities of Large Language Models (LLMs) [5, 21, 47, 48]. Our approach involves projecting reference and target images from supervised CIR triplets into the language embedding space, making them suitable inputs for the LLM. We then fine-tune the model by using prompts such as 'Describe the differences between images.' to induce it to yield human-like visual delta as illustrated in Fig. 2. Furthermore, we employ a parameter-efficient fine-tuning technique, LoRA [19], on the LLM. This choice of design effectively enhances the quality of visual deltas while also preserving the LLM's original capabilities, without harming its inherent knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.713 + ], + "angle": 0, + "content": "After the VDG is trained, it knows how to distinguish between a given reference and target image and produce visual delta as a textual response. If we forward two similar images with different compositions, we can thus obtain the corresponding visual delta easily with the VDG. This allows us to achieve two purposes. First, we can now augment existing CIR triplets by adding generated visual deltas to pairs of reference and target images from the training set. Second, we can also harvest new reference-target pairs from an unlabeled database based on visual similarity, after which we forward these images to VDG to configure new pseudo triplets for CIR training. Note that our VDG is model agnostic - it simply increases the number of triplet candidates for training any given supervised CIR baselines. This strategy strikes a balance between maintaining the integrity of supervisory concepts derived from a supervised dataset and the capacity for effortless expansion using new, unlabeled image samples. It's a cost-effective and scalable solution, ensuring uniformity in annotations across extensive datasets. By generating pseudo triplets with VDG, we significantly reduce annotation costs and enhance the performance of CIR models trained solely on supervised learning, as well as those trained without supervised triplets. Our approach leads to state-of-the-art results in CIR benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.714, + 0.468, + 0.73 + ], + "angle": 0, + "content": "The key advantages of our semi-supervised CIR include:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.795 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to transfer knowledge from Large Language Models (LLMs) and connect it with semi-supervised Composed Image Retrieval (CIR)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.47, + 0.856 + ], + "angle": 0, + "content": "- We propose a novel Visual Delta Generator (VDG) that generates synthetic visual deltas for augmenting the supervised dataset and allowing the integration of an auxiliary image gallery for CIR model training." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- Comprehensive experimental results confirm the effectiveness of our method, demonstrating state-of-the-art retrieval rankings and showcasing the potential of our work." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.875, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.304, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 2. An overview of the VDG tuning process. It includes (a) a vision projector and (b) a Large Language Model (LLM). The VDG is trained to produce visual delta that accurately describes the difference between a reference image and its corresponding target image." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.4, + 0.642, + 0.416 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.73 + ], + "angle": 0, + "content": "Composed Image Retrieval. The field of image retrieval has captured the interest of many researchers in our community [6, 38]. One notable area that has seen much progress recently is Composed Image Retrieval (CIR), a problem that focuses on retrieving images that best match a given pair of a query image and textual intent. Supervised CIR methods [2, 11, 35, 50] are trained on human-annotated triplets, consisting of a reference image, a target image, and their textual difference. On the other hand, zero-shot CIR [3, 9, 16, 42] operates without relying on human-guided descriptions of the differences between the two images. Instead, it uses noisy image-text pairs, aiming to find a function that can translate images into words. This approach, designed to discriminate subtle differences between images based solely on text captions, poses challenges, but also scales well, making it easy to add more data for CIR training. Addressing the limitations of both approaches, we explore the field of semi-supervised learning-based retrieval in this work, leveraging the strengths of both labeled and unlabeled data to enhance retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Semi-supervised Learning. Semi-supervised learning has been an active research topic in visual recognition for a long time [4, 18, 20, 25-27, 39, 44, 46, 51]. Semi-supervised learning can be roughly categorized into two groups, consistency learning [25, 39, 46] and pseudolabeling based learning [27, 44]. Consistency-based methods, as the name implies, encourage consistency in the output of the model by adding noise to model weights or using Exponential Moving Average (EMA) from a teacher model to a student model. In pseudo-labeling based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "ods, hard/soft pseudo-labels obtained from a pretrained model are assigned to unlabeled images [15, 44]. These pseudo-labels can be filtered using confidence thresholding or multi-view consistency. However, neither consistency-based nor pseudo labels methods are directly applicable to CIR. This is because the relative textual descriptions of the corresponding visual differences (visual delta) needed in CIR are not the intended outputs of these methods. In this work, we propose the Visual Delta Generator (VDG), a multi-modal pseudo-label generator. VDG processes two input images and generates text that describes their visual differences, making it an ideal candidate for constructing pseudo triplets for CIR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Multi-modal Models for Image-Text Retrieval. Models like CLIP [41] and BLIP [29] have showcased the advantages of training models on extensive image-text pairs, enabling precise alignment between language and vision representations that is crucial for image-text retrieval. Building on this, there have been significant advancements in utilizing Large Language Models (LLMs) for enhanced vision-language understanding [1, 10, 31, 33, 53]. Especially for CIR, Fromage [23] utilizes LLMs to directly produce embeddings for retrieval, allowing cross-modal compositional search with a single image and textual intent. CoVR [49] and SEARLE [3] apply LLMs to generate visual deltas using image captions without incorporating visual data, which limits the generation of accurate CIR training samples. Our proposed method overcomes these challenges by finetuning LLMs with the integration of a pretrained vision-language alignment module. This integration empowers LLMs to perceive and comprehend images, making them applicable even in scenarios with only image datasets. The method excels in generating accurate visual deltas, a key factor in training efficient external CIR models. These enhancements optimize the use of LLMs while ensuring computational efficiency, thereby expanding the versatility of LLMs in image-related tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.17, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Overview. Our goal is to establish a semi-supervised Composed Image Retrieval (CIR) system that merges image reference features with user textual descriptions to retrieve images from a large-scale database. We face a challenge in the limited availability of supervised triplets necessary for robust CIR model training. To overcome this limitation, we introduce a novel semi-supervised approach for CIR by leveraging an instruction-tuned Large Language Model (LLM), which we call Visual Delta Generator (VDG). The VDG learns to discriminate differences between two images and produces a textual response. This capability allows us to generate additional CIR training triplets, which in turn contributes to the development of more robust CIR models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "Sec. 3.1 provides detailed insights into the construction of the VDG. Sec. 3.2 describes the pseudo triplet generation process. Training of CIR models with pseudo triplets and our additional objective function for better optimization are described in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.175, + 0.787, + 0.192 + ], + "angle": 0, + "content": "3.1. Visual Delta Generator Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.454 + ], + "angle": 0, + "content": "While semi-supervised learning methods have been actively developed for standard visual recognition tasks, these cannot be directly applied to CIR. In CIR, pseudo-label generation requires a detailed semantic understanding of two separate images such that their difference can be automatically expressed in the form of text. We leverage vision-language pretraining models and LLMs to achieve the requirements. With the huge success of LLMs, there are approaches that aim to utilize their understanding of the language domain for improving vision tasks. Particularly, LLaVA [33] and InstructBLIP [10] which are trained on top of the chat bot style instruction tuned LLM, Vicuna [7], have shown interesting results on vision-language tasks. Inspired by these, we propose VDG, which allows the LLM to take two images (reference, target) with similar contents and discriminate their difference in the form of text response (visual delta) as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.621 + ], + "angle": 0, + "content": "First, to enable the LLM to interpret images, we use the Querying Transformer (Q-Former) motivated by InstructBLIP [10] to prepare images for the LLM input (i.e., Vision Projector (VP) in Fig. 2(a)). Q-Former, a transformer encoder [12, 21], processes a fixed set of 32 learnable query tokens (embeddings). These tokens are modified through self-attention layers and interact with image features of Vision Transformer (ViT) [12] through cross-attention layers. As a result, the query tokens are infused with the visual information from the provided image, making them suitable for LLM processing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.895, + 0.804 + ], + "angle": 0, + "content": "Stage 1: Alignment. The outputs of the VP are inherently not aligned with the tokenized word embeddings of the inputs to the LLM, which we chose to be LLaMA2 [48] in this work. Alignment between the VP and LLM needs to be attained by fine-tuning trainable projection of VP in Fig. 2(a) as was done in LLaVA [33]. Specifically, we employ large-scale image-caption pairs to foster alignment between the image representations understood by the VP and the LLM. This step includes minimizing standard next token prediction loss which is generally used to train the decoder-based LLM as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.824, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\nL _ {L L M} = - \\sum_ {t = 1} ^ {T} \\log P \\left(\\mathrm {w} _ {t} \\mid x, I _ {\\text {i n s t}}, \\mathrm {w} _ {(1.., t - 1)}; \\theta_ {\\text {p r o j}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathrm{w}_t\\) denotes a token at time step \\(t\\), \\(T\\) is the length of the sequence, \\(P(\\cdot)\\) is the probability assigned by the model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16807" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "to the actual token \\(\\mathrm{w}_t\\) given with the image \\(x\\), instruction \\(I_{inst}\\) and previous tokens \\(\\mathrm{w}_{(1,\\dots,t - 1)}\\), and \\(\\theta_{proj}\\) is parameters of projection layer. Following standard practice, the prediction for each token is computed using a softmax over the vocabulary of the LLM, and the cross-entropy loss is computed between the predicted probabilities and the textual token labels as a classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.473, + 0.413 + ], + "angle": 0, + "content": "Stage 2: Instruction Tuning. In this stage, we conduct instruction tuning to equip our LLM with the ability to understand image pairs and produce visual delta, as shown in Fig. 2(b). Referring to Fig. 3, we train the LLM with an instruction (i.e., \"Request\") to generate the visual delta in textual format (i.e., \"Response\"). We forward two distinct images (i.e., \"Reference\" and \"Target\") into the LLM via the VP. We employ a parameter-efficient fine-tuning technique, LoRA [19], directly on the LLM. This fine-tuning process is designed to provide \"extra room\" for the LLM to undertake new tasks, all without compromising its foundational capabilities. In this stage, we introduce a specific prompt structure to the LLM as below:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.43, + 0.446, + 0.457 + ], + "angle": 0, + "content": "Request: Analyze given reference and target images and provide a description that transforms the reference to match the target." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.457, + 0.348, + 0.47 + ], + "angle": 0, + "content": "Reference: " + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.47, + 0.307, + 0.481 + ], + "angle": 0, + "content": "Target: " + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.482, + 0.315, + 0.494 + ], + "angle": 0, + "content": "Response: \\(< \\) Visual delta (text label)>" + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.506, + 0.443, + 0.521 + ], + "angle": 0, + "content": "Figure 3. A template prompt for VDG instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.582 + ], + "angle": 0, + "content": "which guides the LLM to generate the corresponding visual delta. We leverage the same training loss from Eqn. 1 to train LoRA parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.383, + 0.607 + ], + "angle": 0, + "content": "3.2. Pseudo Triplet Generation for CIR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.765 + ], + "angle": 0, + "content": "After training VDG, we can generate visual delta of two images, which can be used to form pseudo triplets for CIR. However, it is important to choose two images that not only share certain attributes and similarities but also present other distinct attributes (i.e., visual delta). To gather suitable pairs of reference and target images, we utilize an image encoder to select them from an auxiliary image gallery, which we denote as \\( G' \\). Note that, we note upper strophe (') on samples and embeddings that are obtained from \\( G' \\), in the following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Following the strategy in CIRR [36], we start with an anchor image \\( x_{a} \\) and retrieve the top 20 images from \\( G' \\) using cosine similarity scores between ResNet 152 [17] embeddings, pretrained on ImageNet [24]. We exclude images with scores above 0.94 and sequentially add images to a subgroup of size 6, skipping those within a 0.002 score of the previously included image. As depicted in Fig. 4, we establish dense connections between all pairs (both consecutive, represented by the outer circle, and non-consecutive," + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.088, + 0.877, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.22, + 0.894, + 0.277 + ], + "angle": 0, + "content": "Figure 4. The process of pseudo triplet generation. First, an image subgroup is constructed based on visual similarity (left). Then, paired reference and target images are fed into the VDG to generate the visual delta, completing the triplet formation (right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.382 + ], + "angle": 0, + "content": "represented by the dotted inner connections), while ensuring no overlaps. The arrow's starting point denotes the reference, while its endpoint indicates the target. Given images \\( x_{i}^{\\prime r} \\) and \\( x_{j}^{\\prime t} \\), VDG produces \\( \\delta_{i,j}^{\\prime} \\). This allows us to formulate the pseudo triplet as \\( \\{x_{i}^{\\prime r}, x_{j}^{\\prime t}, \\delta_{i,j}^{\\prime}\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.771, + 0.405 + ], + "angle": 0, + "content": "3.3. Semi-supervised CIR Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Preliminaries. Suppose we have access to a CIR dataset of triplets: \\( D = \\{(x_i^r, x_j^t, \\delta_{n(i,j)})\\}_{n=1}^N \\) where \\( x_i^r, x_j^t \\) denote reference and target images, respectively, \\( \\delta_{n(i,j)} \\) represents their visual delta, and \\( N \\) denotes the total triplet counts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.609 + ], + "angle": 0, + "content": "In the pursuit of enhancing CIR through a semi-supervised approach, our method's strength lies in its model-agnosticism, allowing for seamless integration with a variety of CIR models. Following recent trends, we opt to use the encoders from vision-language pretraining models [10, 29, 31, 41], notably CLIP and BLIP, as our baseline backbones. These encoders are naturally equipped to understand and convert both visual and textual elements into a joint embedding space, making them suitable for CIR tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.894, + 0.79 + ], + "angle": 0, + "content": "The image encoder takes patchified image token embeddings \\( x = [\\mathrm{x}_1, \\dots, \\mathrm{x}_{K_{img}}] \\) with the learnable image \\( cls \\) token embedding \\( [\\mathrm{x}_{cls}] \\), and outputs the visual feature embeddings \\( E_{img}(\\mathrm{x}_{cls}, x) = [\\hat{\\mathrm{x}}_{cls}, \\hat{\\mathrm{x}}_1, \\dots, \\hat{\\mathrm{x}}_{K_{img}}} \\) where \\( \\hat{\\mathrm{x}} \\in R^{d_i} \\) of \\( d_i \\) dimension, and \\( K_{img} \\) denotes the number of generated tokens from each image. The text encoder processes tokenizer output embedding of visual delta \\( \\delta = [\\mathrm{z}_1, \\dots, \\mathrm{z}_{K_{txt}}} \\) with the learnable text \\( cls \\) token embeddings \\( [\\mathrm{z}_{cls}] \\) to produce its textual feature embeddings \\( E_{txt}(\\mathrm{z}_{cls}, \\delta) = [\\hat{\\mathrm{z}}_{cls}, \\hat{\\mathrm{z}}_1, \\dots, \\hat{\\mathrm{z}}_{K_{txt}}} \\), where \\( \\hat{\\mathrm{z}} \\in R^{d_t} \\) of \\( d_t \\) dimension, and \\( K_{txt} \\) denotes the number of generated text tokens from each visual delta." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Model Architecture. To carry out CIR, we establish a fusion function \\( f \\) that takes a reference image \\( x^r \\) and a visual delta \\( \\delta \\), producing a composed embedding \\( \\mathbf{c} \\) as: \\( f(x^r, \\delta) = \\mathbf{c} \\), and notates its trainable components as \\( f_\\theta \\). We utilize two well-known backbones, CLIP and BLIP, to configure \\( f \\). In the case of CLIP, we employ the text encoder \\( E_{txt:\\theta} \\) and an" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16808" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.289 + ], + "angle": 0, + "content": "additional Combiner module [2] \\(\\mathcal{C}_{\\theta}(\\hat{\\mathrm{x}}_{cls}^{r},\\hat{\\mathrm{z}}_{cls})\\) that outputs \\(\\mathbf{c}\\) to be the components of \\(f_{\\theta}\\) (i.e., \\(f_{\\theta}^{\\mathrm{CLIP}} = \\{E_{txt:\\theta},\\mathcal{C}_{\\theta}\\}\\)). The Combiner is designed to optimally blend \\(\\hat{\\mathrm{x}}_{cls}\\) and \\(\\hat{\\mathrm{z}}_{cls}\\) carefully weighing their individual impacts while adeptly mixing them. In the BLIP case, we exclusively use BLIP's text encoder, grounded in the image, and designate it as the trainable \\(E_{txt:\\theta}\\) (i.e., \\(f_{\\theta}^{\\mathrm{BLIP}} = E_{txt:\\theta}\\), and \\(\\mathbf{c} = \\hat{\\mathbf{z}}_{cls}\\)), without incorporating additional modules. BLIP's text encoder inherently fuses image and text signals in its cross-attention layer, eliminating the need for a separate combining module. Notably, we freeze all vision encoders in our setup to ensure compatibility with existing image retrieval galleries and to enhance training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.472, + 0.457 + ], + "angle": 0, + "content": "Supervised / Pseudo Separated Contrastive Loss. The training objective of CIR is to achieve strong alignment between the target image's embedding \\(\\mathbf{x}\\) (where \\(\\mathbf{x} = \\hat{\\mathbf{x}}_{cls}^{t}\\) for simplicity), and the composed embedding \\(\\mathbf{c}\\). On this purpose, we utilize HN-NCE [40] loss for a given training batch \\(\\mathcal{B} \\sim D\\) and \\(\\mathcal{B}' \\sim D'\\), where \\(D'\\) contains pseudo triplets. Additionally, to mitigate the impact of noise in pseudo triplets and ensure consistent contributions from supervised triplets, we compute a target-composed contrastive loss (tcc) as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.471, + 0.47, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t c c} (\\mathcal {B}, \\mathcal {B} ^ {\\prime}) = \\mathcal {L} _ {c} (\\mathcal {B}; \\tau) + \\mathcal {L} _ {c} (\\mathcal {B} \\oplus \\mathcal {B} ^ {\\prime}; \\tau) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.495, + 0.235, + 0.509 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.532, + 0.474, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {c} (\\mathcal {B}) = - \\sum_ {i = 1} ^ {n} \\log \\left[ \\frac {e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau}}{\\alpha \\cdot e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau} + \\sum_ {j \\neq i} ^ {\\mathcal {B}} e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {j} / \\tau} \\cdot w _ {\\mathbf {x} _ {i} , \\mathbf {c} _ {j}}} \\right] \\\\ - \\sum_ {i = 1} ^ {n} \\log \\left[ \\frac {e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau}}{\\alpha \\cdot e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau} + \\sum_ {j \\neq i} ^ {\\mathcal {B}} e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {j} / \\tau} \\cdot w _ {\\mathbf {x} _ {j} , \\mathbf {c} _ {i}}} \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.702 + ], + "angle": 0, + "content": "where \\(\\tau, \\alpha\\) denote hyper-parameters, and \\(\\oplus\\) denotes concatenation along the batch axis, and \\(w_{\\mathbf{x}_i,\\mathbf{c}_j}, w_{\\mathbf{x}_j,\\mathbf{c}_i}\\) are set as in [40]. This design facilitates the independent yet concurrent investigation of both supervised and semi-supervised CIR embedding spaces in a contrastive manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Target-Delta Matching Loss. In the case of BLIP's image-grounded text encoder structure, we introduce a new target-delta matching loss (tdm). Our insight stems from the observation that, while all reference image patch tokens are considered in the cross-attention layer of the BLIP text encoder, target image patch tokens are overlooked when training solely with contrastive learning betweencls token embeddings. As an example shown in Fig. 5, the visual delta can be seen as a weakly correlated caption to the target image. Thus, we aim to align the target image and visual delta to enable the BLIP text encoder to identify the image tokens related to textual input. The tdm loss is applied as:" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.877, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.369, + 0.895, + 0.413 + ], + "angle": 0, + "content": "Figure 5. Illustration of our proposed adaptation of the BLIP image-grounded text encoder for CIR. Both reference \\((x^r)\\) and target image \\((x^t)\\) patch tokens are processed by the text encoder \\((f_{\\theta})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.464, + 0.892, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t d m}} = \\mathbb {E} _ {(x ^ {t}, \\delta) \\sim (D, D ^ {\\prime})} \\mathrm {H} \\left(\\mathbf {y} ^ {t d m}, \\mathbf {p} ^ {t d m} \\left(x ^ {t}, \\delta\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.585 + ], + "angle": 0, + "content": "where \\( \\mathrm{H} \\) is cross entropy, \\( \\mathbf{y}^{tdm} \\) is a 2-dimensional one-hot vector label obtained through the hard negative mining process proposed in [30], where a pair of image and text is configured as matched or not. \\( \\mathbf{p}^{tdm} \\in \\mathbb{R}^2 \\) generates both positive and negative target-delta matching scores and is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.594, + 0.821, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {t d m} (x ^ {t}, \\delta) = \\sum_ {i = 1} ^ {K ^ {t x t}} p _ {\\theta} (\\hat {\\mathbf {z}} _ {i} ^ {t}) / K ^ {t x t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.894, + 0.695 + ], + "angle": 0, + "content": "where \\(\\hat{z}_i^t\\) denotes the \\(i\\)-th token embedding in \\(f_{\\theta}(x^{t},\\delta)\\) and \\(p_{\\theta}\\) is a FC layer that outputs a 2-D vector. This loss ensures the text encoder fully processes the target image features, improving its understanding of the visual delta." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.634, + 0.725 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.894, + 0.78 + ], + "angle": 0, + "content": "Sec. 4.1 details the setup and VDG generation. Sec. 4.2 covers quality checks of visual deltas. Evaluations and baseline comparisons are in Sec. 4.3, and further analyses in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.582, + 0.804 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We utilize InstructBLIP's pretrained weights [10], which have been trained with both a ViT-G/14 [13] and the Q-Former based on the Vicuna-13B model [7], without using prompts for Q-Former. For the stage 1 training described in Sec. 3.1, we employ 595K filtered image-text pairs from CC3M [43] provided by [33] to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.244 + ], + "angle": 0, + "content": "find alignment with our baseline LLM, LLaMA2-13B [48]. In Stage 2, we implement instruction tuning with LoRA parameters [19], following the fixed prompt as outlined in Fig. 3. Visual deltas are generated in an autoregressive manner, predicted based on the LLaMA2 vocabulary. For the CLIP-based CIR training, we select the ViT-L/14 model combined with a Combiner [2]. For the BLIP-based model [29], we use a dedicated BLIP text encoder for image-text matching, paired with the ViT-L/16 model. Additional details can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.573 + ], + "angle": 0, + "content": "Datasets for CIR Evaluation. There are two standard benchmarks in CIR, one is CIRR [36] which deals with natural images, and the other is FashionIQ [52] which focuses on fashion domain images. Each presents unique challenges and datasets that help researchers push the boundaries of what's possible in CIR. Following the protocols utilized in benchmarks [36, 52], we report the CIR results with recall scores at top K retrieval results (R@K), or results under collected subset \\((\\mathrm{R}_s@\\mathrm{K})\\). Specifically, CIRR is configured with 4,351 subgroups (subsets), each containing six similar images, sourced from NLVR2 [45]. For experimental purposes these groups are distributed into train (3,345 subgroups of 16,742 images), validation (503 subgroups of 2,265 images), and test (503 subgroups of 2,178 images) sets. FashionIQ is divided into three categories of Dress, Shirt, and Toptee (Tops and Tees). The reference and target images are paired based on their category similarities. The 18,000 CIR triplets for training are pooled from 45,429 images of the training set, and 6,016 CIR triplets for the test are chosen from 15,415 images of the validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Visual Delta Generation. To produce pseudo triplets, we expand our dataset sources to configure an auxiliary gallery \\((G^{\\prime})\\) which is built upon the grouping strategy introduced in Sec. 3.2, while excluding images that overlap with the benchmark sets. Once the subgroups are constructed, we further filter them to avoid heavy overlap. In total, we draw upon 42,390 unique subgroups from NLVR2 [45], and 79,427 from COCO [32]. Similarly, we build 27,957 individual subgroups from FashionIQ [52], and 30,880 from DeepFashion [34]. We mark upper strophe (') to these datasets. For the semi-supervised settings, we randomly select 3,345 groups from NLVR2' and COCO' for the CIRR case, as well as 3,600 groups from FashionIQ' and DeepFashion'. We denote these datasets used for semi-supervised CIR with the subscript \\(se\\), (e.g., NLVR2'se, COCO'se). Fig. 6 shows the comparison between human-annotated visual deltas and those generated by the VDG in both natural and fashion domains. We observe that the VDG is effective in generating high-quality visual deltas - additional results can be found in the appendix." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.089, + 0.872, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.279, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on visual deltas, human vs. VDG on CIRR and FashionIQ datasets. For both natural and fashion domain images, VDG can produce informative visual deltas." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.329, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Table 1. Retrieval results on CIRR validation set. Human + VDG represents utilizing both human annotated and VDG-generated visual deltas when training the CIR model. \\(\\mathrm{BLIP}_{tdm}\\) represents a model with target-delta matching loss. The best scores for each Val. set supervision are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.411, + 0.88, + 0.579 + ], + "angle": 0, + "content": "
Training Set SupervisionVal. Set SupervisionBaselineR@1R@5R@10R@50
(a) HumanHumanCombiner37.9871.4982.5295.29
BLIPtdm53.1782.0989.8197.54
(b) VDGHumanCombiner35.4768.2980.2994.31
BLIPtdm50.1680.0387.7896.75
(c) Human + VDGHumanCombiner39.1173.0284.4195.96
BLIPtdm53.6782.9989.9797.84
(d) HumanVDGCombiner38.9673.3184.3696.22
BLIPtdm51.6483.3391.3997.90
(e) VDGVDGCombiner41.5977.5787.3597.11
BLIPtdm52.6985.1792.3798.59
(f) Human + VDGVDGCombiner41.4777.9087.4297.30
BLIPtdm52.9585.2792.5498.61
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.796, + 0.623 + ], + "angle": 0, + "content": "4.2. Quality Check of VDG Responses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.688 + ], + "angle": 0, + "content": "To assess the quality of VDG-generated visual deltas, we executed a series of experiments as outlined in Table 1. We use two backbones, Combiner, and BLIP with the tdm loss (denoted as \\(\\mathrm{BLIP}_{\\mathrm{tdm}}\\)), as our baselines for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Initially, we compare the visual deltas generated by VDG with those annotated by humans. This is done by replacing the deltas in the training set with VDG-generated deltas for the same reference-target pairs. When comparing (a) and (b), we observe only a marginal drop in performance upon switching to VDG-generated deltas. This indicates that the deltas generated by VDG are as effective as those created by humans. More importantly, the improved performance observed when comparing (a) with (c) — which combines human-annotated and VDG-generated deltas — highlights the effectiveness of this hybrid approach as a robust data augmentation strategy to enhance CIR models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "To evaluate VDG's performance on new image pairs, we conduct experiments by substituting the human-annotated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16810" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.083, + 0.473, + 0.194 + ], + "angle": 0, + "content": "Table 2. Retrieval results on CIRR test set. * denotes our baselines, † denotes VDG generated visual deltas are applied to augment original training set. We categorize into two distinct groups: one is Seen: Supervised / Supervised + External / Supervised + Aux. with VDG which utilize human-annotated visual delta for CIR model training, and the other is Unseen: Zero-shot / Aux. with VDG which does not utilize human-annotated visual delta for CIR model training. Within each category, best viewed with **bold**." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.205, + 0.481, + 0.543 + ], + "angle": 0, + "content": "
MethodR@1R@5R@10R@50Rs@1Rs@2Rs@3
(a) Supervised
ARTEMIS [11]16.9646.1061.3187.7339.9962.2075.67
CIRPLANT [35]19.5552.5568.3992.3839.2063.0379.49
Combiner [2]33.5965.3577.3595.2162.3981.8192.02
CASE [28]48.0079.1187.2597.5775.8890.5896.00
CoVR [49]48.8478.0586.1094.1975.7888.2292.80
Combiner34.3966.2276.5891.0468.5586.3693.98
\\( Combiner^† \\)36.9169.2179.5492.0470.0087.4594.39
\\( BLIP_{tdm} \\)48.9477.8386.1594.1775.7189.7195.81
\\( BLIP_{tdm}^† \\)49.0878.9886.8994.2476.1890.6295.86
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]49.3580.0288.7597.4776.4890.3795.71
CoVR + WebVid [49]49.6978.6086.7794.3175.0188.1293.16
(c) Supervised + Auxiliary Gallery with VDG
\\( Combiner^† + COCO_{se}^{\\prime} \\)38.7769.2579.2191.5271.2587.4994.34
\\( Combiner^† + NLVR2_{se}^{\\prime} \\)38.8969.8479.4191.1871.9287.8994.29
\\( BLIP_{tdm}^† + COCO_{se}^{\\prime} \\)49.3778.1285.5293.7476.6890.4696.05
\\( BLIP_{tdm}^† + NLVR2_{se}^{\\prime} \\)50.9680.1586.8694.4677.4590.6596.10
(d) Zero-shot
Pic2Word by CC3M [42]23.9051.7065.3087.80---
SEARLE by ImageNet [3]24.2252.4166.2988.6353.7174.6387.61
CASE by LasCo.Ca. [28]35.4065.7878.5394.6364.2982.6691.61
CoVR by WebVid [49]38.4866.7077.2591.4769.2883.7691.11
(e) Auxiliary Gallery with VDG
Combiner by COCO'26.8054.0565.3083.8867.2885.4292.99
Combiner by NLVR2'31.5761.3772.1088.9467.9886.1893.18
\\( BLIP_{tdm} \\)by COCO'43.4972.0781.5993.2172.3687.9894.72
\\( BLIP_{tdm} \\)by NLVR2'45.7475.0182.5293.1372.6087.9094.77
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.473, + 0.831 + ], + "angle": 0, + "content": "deltas in the validation set with deltas generated by VDG. The comparison of (a) and (d) reveals comparable performance levels, suggesting that deltas generated by VDG are as effective as human annotations. When we incorporate VDG-generated deltas into the training set (cases (e) and (f)), we achieve a notable performance improvement over the human-annotated validation set across all metrics, with the exception of a small reduction in \\(\\mathbb{R}@\\mathbb{1}\\). For example, there is a \\(5.14\\%\\) p increase in \\(\\mathbb{R}@\\mathbb{5}\\) between (b) and (e). The improvement observed may stem from the varied text descriptions by different human annotators, leading to inconsistencies in style and detail that potentially affect the distribution of visual differences between the annotated triplets. VDG offers a more consistent solution in generating visual deltas, thereby reducing errors associated with differences in human annotation. Our comprehensive testing demonstrates VDG's reliability as an annotation tool for CIR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.37, + 0.863 + ], + "angle": 0, + "content": "4.3. Comparison with Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Tables 2 and 3 present comparative evaluations between our method and other CIR approaches on CIRR and Fash" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.083, + 0.892, + 0.11 + ], + "angle": 0, + "content": "Table 3. Retrieval results on FashionIQ validation set. We utilize the same notations and same categorizations as Table 2." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.121, + 0.902, + 0.486 + ], + "angle": 0, + "content": "
MethodDressShirtToptee
R@10R@50R@10R@50R@10R@50
(a) Supervised
ARTEMIS [11]27.1652.4021.7843.6429.2053.83
DCNet [22]28.9556.0723.9547.3030.4458.29
FashionVLP [14]32.4260.2931.8958.4438.5168.79
Combiner [2]31.6356.6736.3658.0038.1962.42
CoVR [49]43.5167.9448.2866.6851.5373.60
*Combiner31.9555.0539.2156.8238.5562.16
*Combiner†35.4059.9942.3061.6343.0966.96
*BILIPtdm44.8766.8349.6166.9350.5472.26
*BILIP†tdm46.9068.8650.2868.0452.7374.45
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]47.4469.3648.4870.2350.1872.24
CoVR + WebVid [49]44.5569.0348.4367.4252.6074.31
(c) Supervised + Auxiliary Gallery with VDG
Combiner† + DeepFashion′se35.5060.0942.5462.8643.2767.86
Combiner† + FashionIQ′se36.3060.1943.9862.2744.3368.06
BILIPtdm + DeepFashion′se47.1069.1049.9569.9653.9074.35
BILIP†tdm + FashionIQ′se47.8969.8151.3671.0853.2974.65
(d) Zero-shot
Pic2Word by CC3M [42]20.0040.2026.2043.6027.9047.40
SEARLE by ImageNet [3]20.3243.1827.4345.6829.3250.17
CoVR by WebVid [49]21.9539.0530.3746.1230.7848.73
(e) Auxiliary Gallery with VDG
Combiner by DeepFashion'23.3046.3630.8649.0231.8751.96
Combiner by FashionIQ'28.2651.4632.5851.2834.8855.79
BILIPtdm by DeepFashion'32.6754.3935.4855.0539.4759.92
BILIP†tdm by FashionIQ'37.4858.7037.2957.1142.1262.32
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.822 + ], + "angle": 0, + "content": "ionIQ evaluation protocols. We categorize the experiments into two distinct groups. The first group, Seen, includes scenarios where the CIR model is trained with human-annotated training triplets. In contrast, the second group, Unseen, comprises scenarios in which the CIR model is trained without human-annotated triplets. Our implementations of Combiner and BLIP are denoted by grey box. Specifically, methods under (a) are trained with supervised triplets, while those notated with the symbol \\((\\dagger)\\) are additionally augmented with VDG-generated visual deltas from supervised image pairs (i.e., the same setting as Human + VDG in Table 1). The methods in (b) have been pretrained on external datasets for CIR tasks, resulting in enhanced performances. In our semi-supervised setup (c), the models are trained with augmented supervised triplets and pseudo triplets sub-sampled from the auxiliary gallery as described in Secs. 3.2 and 4.1. Methods in (d) are trained with large-scale datasets with the intention of producing zero-shot performances. In (e), we only use pseudo triplets generated from the auxiliary gallery for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The results clearly demonstrate that our VDG implementation significantly improves the performance of both Combiner and \\(\\mathrm{BLIP}_{tdm}\\) baselines. In (a), a detailed comparison shows that models augmented with VDG samples (with \\(\\dagger\\)) consistently outperform their counterparts (without \\(\\dagger\\)) in all" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16811" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.083, + 0.47, + 0.11 + ], + "angle": 0, + "content": "Table 4. Ablation results on \\(\\mathrm{BLIP}_{tdm}\\) baseline for CIRR validation set with NLVR2' as auxiliary gallery. Best viewed with bold." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.115, + 0.426, + 0.178 + ], + "angle": 0, + "content": "
groupingconcat.tdm lossR@1R@5R@10R@50
52.4581.1888.62,96.88
✓55.6683.4591.2197.33
✓✓56.2884.3991.4197.92
✓✓✓57.8885.5893.2198.33
" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.184, + 0.269, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.26, + 0.229, + 0.272 + ], + "angle": 0, + "content": "(a) CIRR validation" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.184, + 0.459, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.26, + 0.419, + 0.272 + ], + "angle": 0, + "content": "(b) FashionIQ-Shirt" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 7. Analysis on the scale for (c) Supervised + Auxiliary Gallery with \\(VDG\\) of \\(\\mathrm{NLVR2}_{se}^{\\prime}\\) and FashionIQ'se. The x-axis denotes the ratio of additional images to training images." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.32, + 0.469, + 0.36 + ], + "angle": 0, + "content": "Table 5. Analysis on the scale for (e) Auxiliary Gallery with VDG. The dataset is scaled from one eighth (1/8) to the full set (1). Evaluated on CIRR Test set." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.366, + 0.472, + 0.407 + ], + "angle": 0, + "content": "
R@11/81/41/21R@101/81/41/21
COCO42.1542.8943.1043.49COCO79.8181.0181.1681.59
NLVR244.3144.9945.1345.74NLVR282.4882.7582.4382.52
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.55 + ], + "angle": 0, + "content": "recall metrics. Moreover, in (c), when we enhance our CIR baselines with additional pseudo triplets, there is a notable performance improvement. Specifically, when \\(\\mathrm{BLIP}_{tdm}\\) is combined with the auxiliary gallery, it surpasses the previous state-of-the-art results in most recall metrics, showcasing the advantages of VDG. Additionally, it's important to highlight that our semi-supervised CIR method achieves these improvements with considerably fewer images than methods like CASE or CoVR, demonstrating its efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In the unseen scenario, our implemented Combiner baseline not only outperforms Pic2Word, which uses the same CLIP backbone, but our \\(\\mathrm{BLIP}_{tdm}\\) also significantly exceeds the former best-performing SEARLE or CoVR in most metrics. VDG's key advantage is its ability to generate visual deltas that align with the targeted domain, such as using COCO for natural images and DeepFashion for fashion images, by solely utilizing images without any accompanying captions. As a result, VDG significantly reduces the number of training samples required, yet achieves enhanced retrieval performances compared to zero-shot approaches like Pic2Word, CASE, or CoVR. This demonstrates that by focusing on similar domain, image-only datasets, we can substantially improve the efficacy of CIR model training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.248, + 0.788 + ], + "angle": 0, + "content": "4.4. Further Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Ablation study on Each Component. To showcase the effectiveness of our proposed learning strategies, we conducted an ablation study and presented the findings in Table 4. We consider three scenarios: (1) without grouping, where we randomly sample images from the auxiliary gallery to create pairs; (2) without concat., where we remove the concatenation step used in Eqn. 2 and use pseudo triplets only;" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.877, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Figure 8. Retrieval results on auxiliary galleries, COCO, and DeepFashion. Actual user intents are used as text queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.306 + ], + "angle": 0, + "content": "and (3) without \\( tdm \\) loss, where we exclude the tdm loss during training. The results indicate that each element sufficiently contributes to enhancing the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Impact of Generated Data Scale. Figs. 7a and 7b demonstrate the effect of the size of VDG-generated data. In a semi-supervised setup, we find that the performance peaks when the number of pseudo-selected images approximates the size of the training set, and then saturates. This saturation might arise because additional pseudo triplets fail to represent the test sample distribution, particularly as we rigorously remove overlapping images with the test set when forming \\( G' \\). For the unseen case as shown in Table 5, performance is enhanced with an increase in VDG-generated data. This suggests that even in the absence of supervised triplets for CIR model training, a larger set of pseudo triplets not only expands the model's comprehension of visual compositions but also aligns more closely with the distribution of supervised triplets. This alignment facilitates improved targeted domain retrieval performances, striking a balance between generalization and performance enhancement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Qualitative Retrieval Results. We facilitate retrieval by providing user intents along with query images, drawing from auxiliary galleries. The retrieval results, as depicted in Fig. 8, demonstrate accurate and relevant image retrieval. Additional results are detailed in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.619, + 0.667 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.892, + 0.844 + ], + "angle": 0, + "content": "In this study, we investigate a novel semi-supervised learning approach in the context of Composed Image Retrieval (CIR). Our findings reveal that integrating human-annotated data with pseudo triplets generated by the Visual Delta Generator (VDG) significantly enhances the generalization capacity of CIR models. The VDG approach not only streamlines the generation of visual deltas but also emerges as a cost-efficient and effective alternative to extensive human annotation. This work paves the way for future research in semi-supervised learning and showcases VDG as a promising direction for advancing CIR systems." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.852, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment: This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University))." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16812" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.283 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.408 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, pages 104-120. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[9] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In ECCV, pages 558-577. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[11] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, pages 14105-14115, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.294, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.338, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[18] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised semantic segmentation. NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.381, + 0.892, + 0.423 + ], + "angle": 0, + "content": "[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.425, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[20] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, pages 3420-3429, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.468, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[21] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[23] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.655, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[25] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.698, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[26] Semi-Supervised Learning. Semi-supervised learning. CSZ2006. html, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.728, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and early fusion for composed image retrieval. arXiv preprint arXiv:2303.09429, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Junnan Li and et al. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[30] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[34] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[35] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[36] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In CVPR, pages 2125–2134, 2021. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[38] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 129:23-79, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[39] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 41(8):1979-1993, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[40] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, pages 6967-6977, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[42] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, pages 19305-19314, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 33:596-608, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[45] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[49] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[50] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[51] Haokun Wen, Xuemeng Song, Jianhua Yin, Jianlong Wu, Weili Guan, and Liqiang Nie. Self-training boosted multifaceted matching network for composed image retrieval. arXiv preprint arXiv:2305.09979, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[52] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[53] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16814" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_origin.pdf b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e46198cf7683064f0ebf73e4c667aad0b5bc7b1 --- /dev/null +++ b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/0e65e087-fb63-4098-9579-6b8185c1ed79_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6c260ee09c6080f8ef7210bc45f04b9304eea9b412cb2647cd03460135404cc +size 3400326 diff --git a/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/full.md b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6afee98b0eac7a20145a2ff8ce83fbe59a66ed2a --- /dev/null +++ b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/full.md @@ -0,0 +1,277 @@ +# Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval + +Young Kyun Jang\*, Donghyun Kim\*, Zihang Meng, Dat Huynh, and Ser-Nam Lim3 Meta AI, Korea University, University of Central Florida + +# Abstract + +Composed Image Retrieval (CIR) is a task that retrieves images similar to a query, based on a provided textual modification. Current techniques rely on supervised learning for CIR models using labeled triplets of the . These specific triplets are not as commonly available as simple image-text pairs, limiting the widespread use of CIR and its scalability. On the other hand, zero-shot CIR can be relatively easily trained with image-caption pairs without considering the image-to-image relation, but this approach tends to yield lower accuracy. We propose a new semi-supervised CIR approach where we search for a reference and its related target images in auxiliary data and learn our large language model-based Visual Delta Generator (VDG) to generate text describing the visual difference (i.e., visual delta) between the two. VDG, equipped with fluent language knowledge and being model agnostic, can generate pseudo triplets to boost the performance of CIR models. Our approach significantly improves the existing supervised learning approaches and achieves state-of-the-art results on the CIR benchmarks. + +# 1. Introduction + +Image-to-image or text-to-image retrieval, where a query image/text is used to retrieve similar ones from a gallery, has grown into a pivotal research field with many practical applications [38]. However, relying solely on image queries is limiting, as they primarily retrieve similar images, making it challenging to understand the user's intent for modifications in the results. On the other hand, relying solely on text queries can also be restrictive, as it may not effectively convey the user's desired detailed visual contents. To address this, Composed Image Retrieval (CIR) was introduced [2, 36, 42, 52]. CIR seeks to retrieve images using a query that combines both an image and a textual description of the user's intent (referred to as the visual delta), which allows more flexible retrieval. Due to the convenience and diverse + +Figure 1. An illustration of the data preparation process of (a) conventional supervised Composed Image Retrieval (CIR) vs. (b) our proposed semi-supervised CIR. While supervised CIR struggles to scale up due to high annotation costs, our semi-supervised method offers a cost-effective and scalable solution. It augments training samples efficiently by generating pseudo triplets through our Large Language Model (LLM)-based Visual Delta Generator. +![](images/6a50b27cee53aeea613fbe36603465063165538f1648a9eb6aa1cb4e1512ffc5.jpg) +(a) CIR triplet generation with human supervision (expensive). +(b) Visual Delta Generator for generating pseudo triplets. + +applicability of CIR, it has attracted increased attention recently for a variety of real-world applications. + +Existing research on CIR has been developed under two major settings: (1) Supervised CIR: Learning with supervised triplets (i.e. ) [2, 11, 28, 35, 50] as shown in Fig. 1 (a), and (2) Zero-shot CIR: Learning with massive noisy pairs [3, 42, 49], without any CIR supervision. Supervised CIR would obviously yield much higher accuracy in retrieval but requires expensive two-stage data collection processes - collecting pairs of related reference and target images, and then annotating them with visual delta that depicts the difference between them. On the other hand, zero-shot CIR does not incur additional labeling costs and utilizes web-collected noisy image text caption pairs directly. However, it has a much lower performance bar compared to supervised approaches and lacks the ability to specialize in specific CIR domain tasks. + +In this paper, we investigate a class of CIR called semi-supervised CIR, blending supervised and unsupervised samples to enhance generalization (Fig. 1(b)). This + +method focuses on boosting CIR performance in specific retrieval domains by creating new triplets from unsupervised data. Building on this concept, we introduce a novel technique, Visual Delta Generator (VDG), designed to tap into the extensive natural language capabilities of Large Language Models (LLMs) [5, 21, 47, 48]. Our approach involves projecting reference and target images from supervised CIR triplets into the language embedding space, making them suitable inputs for the LLM. We then fine-tune the model by using prompts such as 'Describe the differences between images.' to induce it to yield human-like visual delta as illustrated in Fig. 2. Furthermore, we employ a parameter-efficient fine-tuning technique, LoRA [19], on the LLM. This choice of design effectively enhances the quality of visual deltas while also preserving the LLM's original capabilities, without harming its inherent knowledge. + +After the VDG is trained, it knows how to distinguish between a given reference and target image and produce visual delta as a textual response. If we forward two similar images with different compositions, we can thus obtain the corresponding visual delta easily with the VDG. This allows us to achieve two purposes. First, we can now augment existing CIR triplets by adding generated visual deltas to pairs of reference and target images from the training set. Second, we can also harvest new reference-target pairs from an unlabeled database based on visual similarity, after which we forward these images to VDG to configure new pseudo triplets for CIR training. Note that our VDG is model agnostic - it simply increases the number of triplet candidates for training any given supervised CIR baselines. This strategy strikes a balance between maintaining the integrity of supervisory concepts derived from a supervised dataset and the capacity for effortless expansion using new, unlabeled image samples. It's a cost-effective and scalable solution, ensuring uniformity in annotations across extensive datasets. By generating pseudo triplets with VDG, we significantly reduce annotation costs and enhance the performance of CIR models trained solely on supervised learning, as well as those trained without supervised triplets. Our approach leads to state-of-the-art results in CIR benchmarks. + +The key advantages of our semi-supervised CIR include: + +- To the best of our knowledge, we are the first to transfer knowledge from Large Language Models (LLMs) and connect it with semi-supervised Composed Image Retrieval (CIR). +- We propose a novel Visual Delta Generator (VDG) that generates synthetic visual deltas for augmenting the supervised dataset and allowing the integration of an auxiliary image gallery for CIR model training. +- Comprehensive experimental results confirm the effectiveness of our method, demonstrating state-of-the-art retrieval rankings and showcasing the potential of our work. + +![](images/13694db774af6b195b976f17288d5ec01aa18222a2b0d4febaca9dadd03cfb76.jpg) +Figure 2. An overview of the VDG tuning process. It includes (a) a vision projector and (b) a Large Language Model (LLM). The VDG is trained to produce visual delta that accurately describes the difference between a reference image and its corresponding target image. + +# 2. Related Work + +Composed Image Retrieval. The field of image retrieval has captured the interest of many researchers in our community [6, 38]. One notable area that has seen much progress recently is Composed Image Retrieval (CIR), a problem that focuses on retrieving images that best match a given pair of a query image and textual intent. Supervised CIR methods [2, 11, 35, 50] are trained on human-annotated triplets, consisting of a reference image, a target image, and their textual difference. On the other hand, zero-shot CIR [3, 9, 16, 42] operates without relying on human-guided descriptions of the differences between the two images. Instead, it uses noisy image-text pairs, aiming to find a function that can translate images into words. This approach, designed to discriminate subtle differences between images based solely on text captions, poses challenges, but also scales well, making it easy to add more data for CIR training. Addressing the limitations of both approaches, we explore the field of semi-supervised learning-based retrieval in this work, leveraging the strengths of both labeled and unlabeled data to enhance retrieval performance. + +Semi-supervised Learning. Semi-supervised learning has been an active research topic in visual recognition for a long time [4, 18, 20, 25-27, 39, 44, 46, 51]. Semi-supervised learning can be roughly categorized into two groups, consistency learning [25, 39, 46] and pseudolabeling based learning [27, 44]. Consistency-based methods, as the name implies, encourage consistency in the output of the model by adding noise to model weights or using Exponential Moving Average (EMA) from a teacher model to a student model. In pseudo-labeling based meth + +ods, hard/soft pseudo-labels obtained from a pretrained model are assigned to unlabeled images [15, 44]. These pseudo-labels can be filtered using confidence thresholding or multi-view consistency. However, neither consistency-based nor pseudo labels methods are directly applicable to CIR. This is because the relative textual descriptions of the corresponding visual differences (visual delta) needed in CIR are not the intended outputs of these methods. In this work, we propose the Visual Delta Generator (VDG), a multi-modal pseudo-label generator. VDG processes two input images and generates text that describes their visual differences, making it an ideal candidate for constructing pseudo triplets for CIR. + +Multi-modal Models for Image-Text Retrieval. Models like CLIP [41] and BLIP [29] have showcased the advantages of training models on extensive image-text pairs, enabling precise alignment between language and vision representations that is crucial for image-text retrieval. Building on this, there have been significant advancements in utilizing Large Language Models (LLMs) for enhanced vision-language understanding [1, 10, 31, 33, 53]. Especially for CIR, Fromage [23] utilizes LLMs to directly produce embeddings for retrieval, allowing cross-modal compositional search with a single image and textual intent. CoVR [49] and SEARLE [3] apply LLMs to generate visual deltas using image captions without incorporating visual data, which limits the generation of accurate CIR training samples. Our proposed method overcomes these challenges by finetuning LLMs with the integration of a pretrained vision-language alignment module. This integration empowers LLMs to perceive and comprehend images, making them applicable even in scenarios with only image datasets. The method excels in generating accurate visual deltas, a key factor in training efficient external CIR models. These enhancements optimize the use of LLMs while ensuring computational efficiency, thereby expanding the versatility of LLMs in image-related tasks. + +# 3. Method + +Overview. Our goal is to establish a semi-supervised Composed Image Retrieval (CIR) system that merges image reference features with user textual descriptions to retrieve images from a large-scale database. We face a challenge in the limited availability of supervised triplets necessary for robust CIR model training. To overcome this limitation, we introduce a novel semi-supervised approach for CIR by leveraging an instruction-tuned Large Language Model (LLM), which we call Visual Delta Generator (VDG). The VDG learns to discriminate differences between two images and produces a textual response. This capability allows us to generate additional CIR training triplets, which in turn contributes to the development of more robust CIR models. + +Sec. 3.1 provides detailed insights into the construction of the VDG. Sec. 3.2 describes the pseudo triplet generation process. Training of CIR models with pseudo triplets and our additional objective function for better optimization are described in Sec. 3.3. + +# 3.1. Visual Delta Generator Training + +While semi-supervised learning methods have been actively developed for standard visual recognition tasks, these cannot be directly applied to CIR. In CIR, pseudo-label generation requires a detailed semantic understanding of two separate images such that their difference can be automatically expressed in the form of text. We leverage vision-language pretraining models and LLMs to achieve the requirements. With the huge success of LLMs, there are approaches that aim to utilize their understanding of the language domain for improving vision tasks. Particularly, LLaVA [33] and InstructBLIP [10] which are trained on top of the chat bot style instruction tuned LLM, Vicuna [7], have shown interesting results on vision-language tasks. Inspired by these, we propose VDG, which allows the LLM to take two images (reference, target) with similar contents and discriminate their difference in the form of text response (visual delta) as shown in Fig. 2. + +First, to enable the LLM to interpret images, we use the Querying Transformer (Q-Former) motivated by InstructBLIP [10] to prepare images for the LLM input (i.e., Vision Projector (VP) in Fig. 2(a)). Q-Former, a transformer encoder [12, 21], processes a fixed set of 32 learnable query tokens (embeddings). These tokens are modified through self-attention layers and interact with image features of Vision Transformer (ViT) [12] through cross-attention layers. As a result, the query tokens are infused with the visual information from the provided image, making them suitable for LLM processing. + +Stage 1: Alignment. The outputs of the VP are inherently not aligned with the tokenized word embeddings of the inputs to the LLM, which we chose to be LLaMA2 [48] in this work. Alignment between the VP and LLM needs to be attained by fine-tuning trainable projection of VP in Fig. 2(a) as was done in LLaVA [33]. Specifically, we employ large-scale image-caption pairs to foster alignment between the image representations understood by the VP and the LLM. This step includes minimizing standard next token prediction loss which is generally used to train the decoder-based LLM as: + +$$ +L _ {L L M} = - \sum_ {t = 1} ^ {T} \log P \left(\mathrm {w} _ {t} \mid x, I _ {\text {i n s t}}, \mathrm {w} _ {(1.., t - 1)}; \theta_ {\text {p r o j}}\right), \tag {1} +$$ + +where $\mathrm{w}_t$ denotes a token at time step $t$ , $T$ is the length of the sequence, $P(\cdot)$ is the probability assigned by the model + +to the actual token $\mathrm{w}_t$ given with the image $x$ , instruction $I_{inst}$ and previous tokens $\mathrm{w}_{(1,\dots,t - 1)}$ , and $\theta_{proj}$ is parameters of projection layer. Following standard practice, the prediction for each token is computed using a softmax over the vocabulary of the LLM, and the cross-entropy loss is computed between the predicted probabilities and the textual token labels as a classification. + +Stage 2: Instruction Tuning. In this stage, we conduct instruction tuning to equip our LLM with the ability to understand image pairs and produce visual delta, as shown in Fig. 2(b). Referring to Fig. 3, we train the LLM with an instruction (i.e., "Request") to generate the visual delta in textual format (i.e., "Response"). We forward two distinct images (i.e., "Reference" and "Target") into the LLM via the VP. We employ a parameter-efficient fine-tuning technique, LoRA [19], directly on the LLM. This fine-tuning process is designed to provide "extra room" for the LLM to undertake new tasks, all without compromising its foundational capabilities. In this stage, we introduce a specific prompt structure to the LLM as below: + +Request: Analyze given reference and target images and provide a description that transforms the reference to match the target. + +Reference: + +Target: + +Response: $<$ Visual delta (text label)> + +which guides the LLM to generate the corresponding visual delta. We leverage the same training loss from Eqn. 1 to train LoRA parameters. + +# 3.2. Pseudo Triplet Generation for CIR + +After training VDG, we can generate visual delta of two images, which can be used to form pseudo triplets for CIR. However, it is important to choose two images that not only share certain attributes and similarities but also present other distinct attributes (i.e., visual delta). To gather suitable pairs of reference and target images, we utilize an image encoder to select them from an auxiliary image gallery, which we denote as $G'$ . Note that, we note upper strophe (') on samples and embeddings that are obtained from $G'$ , in the following. + +Following the strategy in CIRR [36], we start with an anchor image $x_{a}$ and retrieve the top 20 images from $G'$ using cosine similarity scores between ResNet 152 [17] embeddings, pretrained on ImageNet [24]. We exclude images with scores above 0.94 and sequentially add images to a subgroup of size 6, skipping those within a 0.002 score of the previously included image. As depicted in Fig. 4, we establish dense connections between all pairs (both consecutive, represented by the outer circle, and non-consecutive, + +![](images/b146655b3f6dbd991aec33738b9171248f9f038243e4aa288d6296248d943bac.jpg) +Figure 3. A template prompt for VDG instruction tuning. +Figure 4. The process of pseudo triplet generation. First, an image subgroup is constructed based on visual similarity (left). Then, paired reference and target images are fed into the VDG to generate the visual delta, completing the triplet formation (right). + +represented by the dotted inner connections), while ensuring no overlaps. The arrow's starting point denotes the reference, while its endpoint indicates the target. Given images $x_{i}^{\prime r}$ and $x_{j}^{\prime t}$ , VDG produces $\delta_{i,j}^{\prime}$ . This allows us to formulate the pseudo triplet as $\{x_{i}^{\prime r}, x_{j}^{\prime t}, \delta_{i,j}^{\prime}\}$ . + +# 3.3. Semi-supervised CIR Training + +Preliminaries. Suppose we have access to a CIR dataset of triplets: $D = \{(x_i^r, x_j^t, \delta_{n(i,j)})\}_{n=1}^N$ where $x_i^r, x_j^t$ denote reference and target images, respectively, $\delta_{n(i,j)}$ represents their visual delta, and $N$ denotes the total triplet counts. + +In the pursuit of enhancing CIR through a semi-supervised approach, our method's strength lies in its model-agnosticism, allowing for seamless integration with a variety of CIR models. Following recent trends, we opt to use the encoders from vision-language pretraining models [10, 29, 31, 41], notably CLIP and BLIP, as our baseline backbones. These encoders are naturally equipped to understand and convert both visual and textual elements into a joint embedding space, making them suitable for CIR tasks. + +The image encoder takes patchified image token embeddings $x = [\mathrm{x}_1, \dots, \mathrm{x}_{K_{img}}]$ with the learnable image $cls$ token embedding $[\mathrm{x}_{cls}]$ , and outputs the visual feature embeddings $E_{img}(\mathrm{x}_{cls}, x) = [\hat{\mathrm{x}}_{cls}, \hat{\mathrm{x}}_1, \dots, \hat{\mathrm{x}}_{K_{img}}}$ where $\hat{\mathrm{x}} \in R^{d_i}$ of $d_i$ dimension, and $K_{img}$ denotes the number of generated tokens from each image. The text encoder processes tokenizer output embedding of visual delta $\delta = [\mathrm{z}_1, \dots, \mathrm{z}_{K_{txt}}}$ with the learnable text $cls$ token embeddings $[\mathrm{z}_{cls}]$ to produce its textual feature embeddings $E_{txt}(\mathrm{z}_{cls}, \delta) = [\hat{\mathrm{z}}_{cls}, \hat{\mathrm{z}}_1, \dots, \hat{\mathrm{z}}_{K_{txt}}}$ , where $\hat{\mathrm{z}} \in R^{d_t}$ of $d_t$ dimension, and $K_{txt}$ denotes the number of generated text tokens from each visual delta. + +Model Architecture. To carry out CIR, we establish a fusion function $f$ that takes a reference image $x^r$ and a visual delta $\delta$ , producing a composed embedding $\mathbf{c}$ as: $f(x^r, \delta) = \mathbf{c}$ , and notates its trainable components as $f_\theta$ . We utilize two well-known backbones, CLIP and BLIP, to configure $f$ . In the case of CLIP, we employ the text encoder $E_{txt:\theta}$ and an + +additional Combiner module [2] $\mathcal{C}_{\theta}(\hat{\mathrm{x}}_{cls}^{r},\hat{\mathrm{z}}_{cls})$ that outputs $\mathbf{c}$ to be the components of $f_{\theta}$ (i.e., $f_{\theta}^{\mathrm{CLIP}} = \{E_{txt:\theta},\mathcal{C}_{\theta}\}$ ). The Combiner is designed to optimally blend $\hat{\mathrm{x}}_{cls}$ and $\hat{\mathrm{z}}_{cls}$ carefully weighing their individual impacts while adeptly mixing them. In the BLIP case, we exclusively use BLIP's text encoder, grounded in the image, and designate it as the trainable $E_{txt:\theta}$ (i.e., $f_{\theta}^{\mathrm{BLIP}} = E_{txt:\theta}$ , and $\mathbf{c} = \hat{\mathbf{z}}_{cls}$ ), without incorporating additional modules. BLIP's text encoder inherently fuses image and text signals in its cross-attention layer, eliminating the need for a separate combining module. Notably, we freeze all vision encoders in our setup to ensure compatibility with existing image retrieval galleries and to enhance training efficiency. + +Supervised / Pseudo Separated Contrastive Loss. The training objective of CIR is to achieve strong alignment between the target image's embedding $\mathbf{x}$ (where $\mathbf{x} = \hat{\mathbf{x}}_{cls}^{t}$ for simplicity), and the composed embedding $\mathbf{c}$ . On this purpose, we utilize HN-NCE [40] loss for a given training batch $\mathcal{B} \sim D$ and $\mathcal{B}' \sim D'$ , where $D'$ contains pseudo triplets. Additionally, to mitigate the impact of noise in pseudo triplets and ensure consistent contributions from supervised triplets, we compute a target-composed contrastive loss (tcc) as: + +$$ +\mathcal {L} _ {t c c} (\mathcal {B}, \mathcal {B} ^ {\prime}) = \mathcal {L} _ {c} (\mathcal {B}; \tau) + \mathcal {L} _ {c} (\mathcal {B} \oplus \mathcal {B} ^ {\prime}; \tau) \tag {2} +$$ + +where $\mathcal{L}_c$ is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {c} (\mathcal {B}) = - \sum_ {i = 1} ^ {n} \log \left[ \frac {e ^ {\mathbf {x} _ {i} ^ {T} \mathbf {c} _ {i} / \tau}}{\alpha \cdot e ^ {\mathbf {x} _ {i} ^ {T} \mathbf {c} _ {i} / \tau} + \sum_ {j \neq i} ^ {\mathcal {B}} e ^ {\mathbf {x} _ {i} ^ {T} \mathbf {c} _ {j} / \tau} \cdot w _ {\mathbf {x} _ {i} , \mathbf {c} _ {j}}} \right] \\ - \sum_ {i = 1} ^ {n} \log \left[ \frac {e ^ {\mathbf {x} _ {i} ^ {T} \mathbf {c} _ {i} / \tau}}{\alpha \cdot e ^ {\mathbf {x} _ {i} ^ {T} \mathbf {c} _ {i} / \tau} + \sum_ {j \neq i} ^ {\mathcal {B}} e ^ {\mathbf {x} _ {i} ^ {T} \mathbf {c} _ {j} / \tau} \cdot w _ {\mathbf {x} _ {j} , \mathbf {c} _ {i}}} \right]. \\ \end{array} +$$ + +where $\tau, \alpha$ denote hyper-parameters, and $\oplus$ denotes concatenation along the batch axis, and $w_{\mathbf{x}_i,\mathbf{c}_j}, w_{\mathbf{x}_j,\mathbf{c}_i}$ are set as in [40]. This design facilitates the independent yet concurrent investigation of both supervised and semi-supervised CIR embedding spaces in a contrastive manner. + +Target-Delta Matching Loss. In the case of BLIP's image-grounded text encoder structure, we introduce a new target-delta matching loss (tdm). Our insight stems from the observation that, while all reference image patch tokens are considered in the cross-attention layer of the BLIP text encoder, target image patch tokens are overlooked when training solely with contrastive learning betweencls token embeddings. As an example shown in Fig. 5, the visual delta can be seen as a weakly correlated caption to the target image. Thus, we aim to align the target image and visual delta to enable the BLIP text encoder to identify the image tokens related to textual input. The tdm loss is applied as: + +![](images/d9acc4d73c9bb68c4f258615b7f3979d99bf5767bc72d0f8adbef05669e037d6.jpg) +Figure 5. Illustration of our proposed adaptation of the BLIP image-grounded text encoder for CIR. Both reference $(x^r)$ and target image $(x^t)$ patch tokens are processed by the text encoder $(f_{\theta})$ + +$$ +\mathcal {L} _ {\mathrm {t d m}} = \mathbb {E} _ {(x ^ {t}, \delta) \sim (D, D ^ {\prime})} \mathrm {H} \left(\mathbf {y} ^ {t d m}, \mathbf {p} ^ {t d m} \left(x ^ {t}, \delta\right)\right) \tag {3} +$$ + +where $\mathrm{H}$ is cross entropy, $\mathbf{y}^{tdm}$ is a 2-dimensional one-hot vector label obtained through the hard negative mining process proposed in [30], where a pair of image and text is configured as matched or not. $\mathbf{p}^{tdm} \in \mathbb{R}^2$ generates both positive and negative target-delta matching scores and is computed as: + +$$ +\mathbf {p} ^ {t d m} (x ^ {t}, \delta) = \sum_ {i = 1} ^ {K ^ {t x t}} p _ {\theta} (\hat {\mathbf {z}} _ {i} ^ {t}) / K ^ {t x t}, +$$ + +where $\hat{z}_i^t$ denotes the $i$ -th token embedding in $f_{\theta}(x^{t},\delta)$ and $p_{\theta}$ is a FC layer that outputs a 2-D vector. This loss ensures the text encoder fully processes the target image features, improving its understanding of the visual delta. + +# 4. Experiments + +Sec. 4.1 details the setup and VDG generation. Sec. 4.2 covers quality checks of visual deltas. Evaluations and baseline comparisons are in Sec. 4.3, and further analyses in Sec. 4.4. + +# 4.1. Setup + +Implementation Details. We utilize InstructBLIP's pretrained weights [10], which have been trained with both a ViT-G/14 [13] and the Q-Former based on the Vicuna-13B model [7], without using prompts for Q-Former. For the stage 1 training described in Sec. 3.1, we employ 595K filtered image-text pairs from CC3M [43] provided by [33] to + +find alignment with our baseline LLM, LLaMA2-13B [48]. In Stage 2, we implement instruction tuning with LoRA parameters [19], following the fixed prompt as outlined in Fig. 3. Visual deltas are generated in an autoregressive manner, predicted based on the LLaMA2 vocabulary. For the CLIP-based CIR training, we select the ViT-L/14 model combined with a Combiner [2]. For the BLIP-based model [29], we use a dedicated BLIP text encoder for image-text matching, paired with the ViT-L/16 model. Additional details can be found in the appendix. + +Datasets for CIR Evaluation. There are two standard benchmarks in CIR, one is CIRR [36] which deals with natural images, and the other is FashionIQ [52] which focuses on fashion domain images. Each presents unique challenges and datasets that help researchers push the boundaries of what's possible in CIR. Following the protocols utilized in benchmarks [36, 52], we report the CIR results with recall scores at top K retrieval results (R@K), or results under collected subset $(\mathrm{R}_s@\mathrm{K})$ . Specifically, CIRR is configured with 4,351 subgroups (subsets), each containing six similar images, sourced from NLVR2 [45]. For experimental purposes these groups are distributed into train (3,345 subgroups of 16,742 images), validation (503 subgroups of 2,265 images), and test (503 subgroups of 2,178 images) sets. FashionIQ is divided into three categories of Dress, Shirt, and Toptee (Tops and Tees). The reference and target images are paired based on their category similarities. The 18,000 CIR triplets for training are pooled from 45,429 images of the training set, and 6,016 CIR triplets for the test are chosen from 15,415 images of the validation set. + +Visual Delta Generation. To produce pseudo triplets, we expand our dataset sources to configure an auxiliary gallery $(G^{\prime})$ which is built upon the grouping strategy introduced in Sec. 3.2, while excluding images that overlap with the benchmark sets. Once the subgroups are constructed, we further filter them to avoid heavy overlap. In total, we draw upon 42,390 unique subgroups from NLVR2 [45], and 79,427 from COCO [32]. Similarly, we build 27,957 individual subgroups from FashionIQ [52], and 30,880 from DeepFashion [34]. We mark upper strophe (') to these datasets. For the semi-supervised settings, we randomly select 3,345 groups from NLVR2' and COCO' for the CIRR case, as well as 3,600 groups from FashionIQ' and DeepFashion'. We denote these datasets used for semi-supervised CIR with the subscript $se$ , (e.g., NLVR2'se, COCO'se). Fig. 6 shows the comparison between human-annotated visual deltas and those generated by the VDG in both natural and fashion domains. We observe that the VDG is effective in generating high-quality visual deltas - additional results can be found in the appendix. + +![](images/147fb22b2bec61b0e22255dd01a8ba62a80aa236fc914815918287634d2d9360.jpg) +Figure 6. Qualitative comparison on visual deltas, human vs. VDG on CIRR and FashionIQ datasets. For both natural and fashion domain images, VDG can produce informative visual deltas. + +Table 1. Retrieval results on CIRR validation set. Human + VDG represents utilizing both human annotated and VDG-generated visual deltas when training the CIR model. $\mathrm{BLIP}_{tdm}$ represents a model with target-delta matching loss. The best scores for each Val. set supervision are highlighted in bold. + +
Training Set SupervisionVal. Set SupervisionBaselineR@1R@5R@10R@50
(a) HumanHumanCombiner37.9871.4982.5295.29
BLIPtdm53.1782.0989.8197.54
(b) VDGHumanCombiner35.4768.2980.2994.31
BLIPtdm50.1680.0387.7896.75
(c) Human + VDGHumanCombiner39.1173.0284.4195.96
BLIPtdm53.6782.9989.9797.84
(d) HumanVDGCombiner38.9673.3184.3696.22
BLIPtdm51.6483.3391.3997.90
(e) VDGVDGCombiner41.5977.5787.3597.11
BLIPtdm52.6985.1792.3798.59
(f) Human + VDGVDGCombiner41.4777.9087.4297.30
BLIPtdm52.9585.2792.5498.61
+ +# 4.2. Quality Check of VDG Responses + +To assess the quality of VDG-generated visual deltas, we executed a series of experiments as outlined in Table 1. We use two backbones, Combiner, and BLIP with the tdm loss (denoted as $\mathrm{BLIP}_{\mathrm{tdm}}$ ), as our baselines for evaluation. + +Initially, we compare the visual deltas generated by VDG with those annotated by humans. This is done by replacing the deltas in the training set with VDG-generated deltas for the same reference-target pairs. When comparing (a) and (b), we observe only a marginal drop in performance upon switching to VDG-generated deltas. This indicates that the deltas generated by VDG are as effective as those created by humans. More importantly, the improved performance observed when comparing (a) with (c) — which combines human-annotated and VDG-generated deltas — highlights the effectiveness of this hybrid approach as a robust data augmentation strategy to enhance CIR models. + +To evaluate VDG's performance on new image pairs, we conduct experiments by substituting the human-annotated + +Table 2. Retrieval results on CIRR test set. * denotes our baselines, † denotes VDG generated visual deltas are applied to augment original training set. We categorize into two distinct groups: one is Seen: Supervised / Supervised + External / Supervised + Aux. with VDG which utilize human-annotated visual delta for CIR model training, and the other is Unseen: Zero-shot / Aux. with VDG which does not utilize human-annotated visual delta for CIR model training. Within each category, best viewed with **bold**. + +
MethodR@1R@5R@10R@50Rs@1Rs@2Rs@3
(a) Supervised
ARTEMIS [11]16.9646.1061.3187.7339.9962.2075.67
CIRPLANT [35]19.5552.5568.3992.3839.2063.0379.49
Combiner [2]33.5965.3577.3595.2162.3981.8192.02
CASE [28]48.0079.1187.2597.5775.8890.5896.00
CoVR [49]48.8478.0586.1094.1975.7888.2292.80
Combiner34.3966.2276.5891.0468.5586.3693.98
\( Combiner^† \)36.9169.2179.5492.0470.0087.4594.39
\( BLIP_{tdm} \)48.9477.8386.1594.1775.7189.7195.81
\( BLIP_{tdm}^† \)49.0878.9886.8994.2476.1890.6295.86
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]49.3580.0288.7597.4776.4890.3795.71
CoVR + WebVid [49]49.6978.6086.7794.3175.0188.1293.16
(c) Supervised + Auxiliary Gallery with VDG
\( Combiner^† + COCO_{se}^{\prime} \)38.7769.2579.2191.5271.2587.4994.34
\( Combiner^† + NLVR2_{se}^{\prime} \)38.8969.8479.4191.1871.9287.8994.29
\( BLIP_{tdm}^† + COCO_{se}^{\prime} \)49.3778.1285.5293.7476.6890.4696.05
\( BLIP_{tdm}^† + NLVR2_{se}^{\prime} \)50.9680.1586.8694.4677.4590.6596.10
(d) Zero-shot
Pic2Word by CC3M [42]23.9051.7065.3087.80---
SEARLE by ImageNet [3]24.2252.4166.2988.6353.7174.6387.61
CASE by LasCo.Ca. [28]35.4065.7878.5394.6364.2982.6691.61
CoVR by WebVid [49]38.4866.7077.2591.4769.2883.7691.11
(e) Auxiliary Gallery with VDG
Combiner by COCO'26.8054.0565.3083.8867.2885.4292.99
Combiner by NLVR2'31.5761.3772.1088.9467.9886.1893.18
\( BLIP_{tdm} \)by COCO'43.4972.0781.5993.2172.3687.9894.72
\( BLIP_{tdm} \)by NLVR2'45.7475.0182.5293.1372.6087.9094.77
+ +deltas in the validation set with deltas generated by VDG. The comparison of (a) and (d) reveals comparable performance levels, suggesting that deltas generated by VDG are as effective as human annotations. When we incorporate VDG-generated deltas into the training set (cases (e) and (f)), we achieve a notable performance improvement over the human-annotated validation set across all metrics, with the exception of a small reduction in $\mathbb{R}@\mathbb{1}$ . For example, there is a $5.14\%$ p increase in $\mathbb{R}@\mathbb{5}$ between (b) and (e). The improvement observed may stem from the varied text descriptions by different human annotators, leading to inconsistencies in style and detail that potentially affect the distribution of visual differences between the annotated triplets. VDG offers a more consistent solution in generating visual deltas, thereby reducing errors associated with differences in human annotation. Our comprehensive testing demonstrates VDG's reliability as an annotation tool for CIR. + +# 4.3. Comparison with Other Methods + +Tables 2 and 3 present comparative evaluations between our method and other CIR approaches on CIRR and Fash + +Table 3. Retrieval results on FashionIQ validation set. We utilize the same notations and same categorizations as Table 2. + +
MethodDressShirtToptee
R@10R@50R@10R@50R@10R@50
(a) Supervised
ARTEMIS [11]27.1652.4021.7843.6429.2053.83
DCNet [22]28.9556.0723.9547.3030.4458.29
FashionVLP [14]32.4260.2931.8958.4438.5168.79
Combiner [2]31.6356.6736.3658.0038.1962.42
CoVR [49]43.5167.9448.2866.6851.5373.60
*Combiner31.9555.0539.2156.8238.5562.16
*Combiner†35.4059.9942.3061.6343.0966.96
*BILIPtdm44.8766.8349.6166.9350.5472.26
*BILIP†tdm46.9068.8650.2868.0452.7374.45
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]47.4469.3648.4870.2350.1872.24
CoVR + WebVid [49]44.5569.0348.4367.4252.6074.31
(c) Supervised + Auxiliary Gallery with VDG
Combiner† + DeepFashion′se35.5060.0942.5462.8643.2767.86
Combiner† + FashionIQ′se36.3060.1943.9862.2744.3368.06
BILIPtdm + DeepFashion′se47.1069.1049.9569.9653.9074.35
BILIP†tdm + FashionIQ′se47.8969.8151.3671.0853.2974.65
(d) Zero-shot
Pic2Word by CC3M [42]20.0040.2026.2043.6027.9047.40
SEARLE by ImageNet [3]20.3243.1827.4345.6829.3250.17
CoVR by WebVid [49]21.9539.0530.3746.1230.7848.73
(e) Auxiliary Gallery with VDG
Combiner by DeepFashion'23.3046.3630.8649.0231.8751.96
Combiner by FashionIQ'28.2651.4632.5851.2834.8855.79
BILIPtdm by DeepFashion'32.6754.3935.4855.0539.4759.92
BILIP†tdm by FashionIQ'37.4858.7037.2957.1142.1262.32
+ +ionIQ evaluation protocols. We categorize the experiments into two distinct groups. The first group, Seen, includes scenarios where the CIR model is trained with human-annotated training triplets. In contrast, the second group, Unseen, comprises scenarios in which the CIR model is trained without human-annotated triplets. Our implementations of Combiner and BLIP are denoted by grey box. Specifically, methods under (a) are trained with supervised triplets, while those notated with the symbol $(\dagger)$ are additionally augmented with VDG-generated visual deltas from supervised image pairs (i.e., the same setting as Human + VDG in Table 1). The methods in (b) have been pretrained on external datasets for CIR tasks, resulting in enhanced performances. In our semi-supervised setup (c), the models are trained with augmented supervised triplets and pseudo triplets sub-sampled from the auxiliary gallery as described in Secs. 3.2 and 4.1. Methods in (d) are trained with large-scale datasets with the intention of producing zero-shot performances. In (e), we only use pseudo triplets generated from the auxiliary gallery for training. + +The results clearly demonstrate that our VDG implementation significantly improves the performance of both Combiner and $\mathrm{BLIP}_{tdm}$ baselines. In (a), a detailed comparison shows that models augmented with VDG samples (with $\dagger$ ) consistently outperform their counterparts (without $\dagger$ ) in all + +Table 4. Ablation results on $\mathrm{BLIP}_{tdm}$ baseline for CIRR validation set with NLVR2' as auxiliary gallery. Best viewed with bold. + +
groupingconcat.tdm lossR@1R@5R@10R@50
52.4581.1888.62,96.88
✓55.6683.4591.2197.33
✓✓56.2884.3991.4197.92
✓✓✓57.8885.5893.2198.33
+ +![](images/612c8710939a4a85a99ea0a46c08a4506ddcf9a947748bb2a0738cb7373a7950.jpg) +(a) CIRR validation +Figure 7. Analysis on the scale for (c) Supervised + Auxiliary Gallery with $VDG$ of $\mathrm{NLVR2}_{se}^{\prime}$ and FashionIQ'se. The x-axis denotes the ratio of additional images to training images. +Table 5. Analysis on the scale for (e) Auxiliary Gallery with VDG. The dataset is scaled from one eighth (1/8) to the full set (1). Evaluated on CIRR Test set. + +![](images/02d04f82ef67ea5b2423f043e1f9a3a81527c27574530da36b44a178f40654af.jpg) +(b) FashionIQ-Shirt + +
R@11/81/41/21R@101/81/41/21
COCO42.1542.8943.1043.49COCO79.8181.0181.1681.59
NLVR244.3144.9945.1345.74NLVR282.4882.7582.4382.52
+ +recall metrics. Moreover, in (c), when we enhance our CIR baselines with additional pseudo triplets, there is a notable performance improvement. Specifically, when $\mathrm{BLIP}_{tdm}$ is combined with the auxiliary gallery, it surpasses the previous state-of-the-art results in most recall metrics, showcasing the advantages of VDG. Additionally, it's important to highlight that our semi-supervised CIR method achieves these improvements with considerably fewer images than methods like CASE or CoVR, demonstrating its efficiency. + +In the unseen scenario, our implemented Combiner baseline not only outperforms Pic2Word, which uses the same CLIP backbone, but our $\mathrm{BLIP}_{tdm}$ also significantly exceeds the former best-performing SEARLE or CoVR in most metrics. VDG's key advantage is its ability to generate visual deltas that align with the targeted domain, such as using COCO for natural images and DeepFashion for fashion images, by solely utilizing images without any accompanying captions. As a result, VDG significantly reduces the number of training samples required, yet achieves enhanced retrieval performances compared to zero-shot approaches like Pic2Word, CASE, or CoVR. This demonstrates that by focusing on similar domain, image-only datasets, we can substantially improve the efficacy of CIR model training. + +# 4.4. Further Analyses + +Ablation study on Each Component. To showcase the effectiveness of our proposed learning strategies, we conducted an ablation study and presented the findings in Table 4. We consider three scenarios: (1) without grouping, where we randomly sample images from the auxiliary gallery to create pairs; (2) without concat., where we remove the concatenation step used in Eqn. 2 and use pseudo triplets only; + +![](images/94c62f3e17c01e2d5119a6c86fea7816a6988019622ffb34814861089bf7b592.jpg) +Figure 8. Retrieval results on auxiliary galleries, COCO, and DeepFashion. Actual user intents are used as text queries. + +and (3) without $tdm$ loss, where we exclude the tdm loss during training. The results indicate that each element sufficiently contributes to enhancing the performance. + +Impact of Generated Data Scale. Figs. 7a and 7b demonstrate the effect of the size of VDG-generated data. In a semi-supervised setup, we find that the performance peaks when the number of pseudo-selected images approximates the size of the training set, and then saturates. This saturation might arise because additional pseudo triplets fail to represent the test sample distribution, particularly as we rigorously remove overlapping images with the test set when forming $G'$ . For the unseen case as shown in Table 5, performance is enhanced with an increase in VDG-generated data. This suggests that even in the absence of supervised triplets for CIR model training, a larger set of pseudo triplets not only expands the model's comprehension of visual compositions but also aligns more closely with the distribution of supervised triplets. This alignment facilitates improved targeted domain retrieval performances, striking a balance between generalization and performance enhancement. + +Qualitative Retrieval Results. We facilitate retrieval by providing user intents along with query images, drawing from auxiliary galleries. The retrieval results, as depicted in Fig. 8, demonstrate accurate and relevant image retrieval. Additional results are detailed in the appendix. + +# 5. Conclusion + +In this study, we investigate a novel semi-supervised learning approach in the context of Composed Image Retrieval (CIR). Our findings reveal that integrating human-annotated data with pseudo triplets generated by the Visual Delta Generator (VDG) significantly enhances the generalization capacity of CIR models. The VDG approach not only streamlines the generation of visual deltas but also emerges as a cost-efficient and effective alternative to extensive human annotation. This work paves the way for future research in semi-supervised learning and showcases VDG as a promising direction for advancing CIR systems. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3 +[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 2, 5, 6, 7 +[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 7 +[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS, 32, 2019. 2 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2 +[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, pages 104-120. Springer, 2020. 2 +[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 3, 5 +[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2 +[9] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. "this is my unicorn, fluffy": Personalizing frozen vision-language representations. In ECCV, pages 558-577. Springer, 2022. 2 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 4, 5 +[11] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 2, 7 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3 +[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue + +Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 5 +[14] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, pages 14105-14115, 2022. 7 +[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3 +[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4 +[18] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised semantic segmentation. NeurIPS, 2015. 2 +[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 4, 6 +[20] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, pages 3420-3429, 2020. 2 +[21] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2, 3 +[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 7 +[23] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3 +[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 4 +[25] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.2 +[26] Semi-Supervised Learning. Semi-supervised learning. CSZ2006. html, 2006. +[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2 +[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and early fusion for composed image retrieval. arXiv preprint arXiv:2303.09429, 2023. 1, 7 +[29] Junnan Li and et al. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3, 4, 6 +[30] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. + +Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 5 +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. 2023. 3, 4 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 5, 1 +[34] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6 +[35] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 1, 2, 7 +[36] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In CVPR, pages 2125–2134, 2021. 1, 4, 6 +[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 1 +[38] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 129:23-79, 2021. 1, 2 +[39] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 41(8):1979-1993, 2018. 2 +[40] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, pages 6967-6977, 2023. 5 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 4 +[42] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, pages 19305-19314, 2023. 1, 2, 7 +[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 5 +[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 33:596-608, 2020. 2, 3 + +[45] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 6 +[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 30, 2017. 2 +[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6 +[49] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 1, 3, 7 +[50] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1, 2 +[51] Haokun Wen, Xuemeng Song, Jianhua Yin, Jianlong Wu, Weili Guan, and Liqiang Nie. Self-training boosted multifaceted matching network for composed image retrieval. arXiv preprint arXiv:2305.09979, 2023. 2 +[52] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 6 +[53] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3 \ No newline at end of file diff --git a/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/images.zip b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4309deec07e508fdefdbea7dddc12fe12059eb1c --- /dev/null +++ b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e25a3c2ecba7ec4439b9e6c2c60add041bed84cca061fa32db6df400afce28d5 +size 598085 diff --git a/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/layout.json b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8930c0c01246c930409ab2667909bcb1cec30163 --- /dev/null +++ b/2024/Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval/layout.json @@ -0,0 +1,8175 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 470, + 140 + ], + "type": "text", + "content": "Visual Delta Generator with Large Multi-modal Models for Semi-supervised Composed Image Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 506, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 506, + 190 + ], + "type": "text", + "content": "Young Kyun Jang\\*, Donghyun Kim\\*, Zihang Meng, Dat Huynh, and Ser-Nam Lim3 Meta AI, Korea University, University of Central Florida" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 243, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 290, + 483 + ], + "type": "text", + "content": "Composed Image Retrieval (CIR) is a task that retrieves images similar to a query, based on a provided textual modification. Current techniques rely on supervised learning for CIR models using labeled triplets of the . These specific triplets are not as commonly available as simple image-text pairs, limiting the widespread use of CIR and its scalability. On the other hand, zero-shot CIR can be relatively easily trained with image-caption pairs without considering the image-to-image relation, but this approach tends to yield lower accuracy. We propose a new semi-supervised CIR approach where we search for a reference and its related target images in auxiliary data and learn our large language model-based Visual Delta Generator (VDG) to generate text describing the visual difference (i.e., visual delta) between the two. VDG, equipped with fluent language knowledge and being model agnostic, can generate pseudo triplets to boost the performance of CIR models. Our approach significantly improves the existing supervised learning approaches and achieves state-of-the-art results on the CIR benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 527, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 694 + ], + "type": "text", + "content": "Image-to-image or text-to-image retrieval, where a query image/text is used to retrieve similar ones from a gallery, has grown into a pivotal research field with many practical applications [38]. However, relying solely on image queries is limiting, as they primarily retrieve similar images, making it challenging to understand the user's intent for modifications in the results. On the other hand, relying solely on text queries can also be restrictive, as it may not effectively convey the user's desired detailed visual contents. To address this, Composed Image Retrieval (CIR) was introduced [2, 36, 42, 52]. CIR seeks to retrieve images using a query that combines both an image and a textual description of the user's intent (referred to as the visual delta), which allows more flexible retrieval. Due to the convenience and diverse" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 544, + 331 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 544, + 331 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 544, + 331 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 544, + 331 + ], + "type": "image", + "image_path": "6a50b27cee53aeea613fbe36603465063165538f1648a9eb6aa1cb4e1512ffc5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 333, + 427, + 352 + ], + "lines": [ + { + "bbox": [ + 309, + 333, + 427, + 352 + ], + "spans": [ + { + "bbox": [ + 309, + 333, + 427, + 352 + ], + "type": "text", + "content": "(a) CIR triplet generation with human supervision (expensive)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 429, + 333, + 542, + 352 + ], + "lines": [ + { + "bbox": [ + 429, + 333, + 542, + 352 + ], + "spans": [ + { + "bbox": [ + 429, + 333, + 542, + 352 + ], + "type": "text", + "content": "(b) Visual Delta Generator for generating pseudo triplets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 304, + 361, + 545, + 438 + ], + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 438 + ], + "type": "text", + "content": "Figure 1. An illustration of the data preparation process of (a) conventional supervised Composed Image Retrieval (CIR) vs. (b) our proposed semi-supervised CIR. While supervised CIR struggles to scale up due to high annotation costs, our semi-supervised method offers a cost-effective and scalable solution. It augments training samples efficiently by generating pseudo triplets through our Large Language Model (LLM)-based Visual Delta Generator." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 485 + ], + "type": "text", + "content": "applicability of CIR, it has attracted increased attention recently for a variety of real-world applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 676 + ], + "type": "text", + "content": "Existing research on CIR has been developed under two major settings: (1) Supervised CIR: Learning with supervised triplets (i.e. ) [2, 11, 28, 35, 50] as shown in Fig. 1 (a), and (2) Zero-shot CIR: Learning with massive noisy pairs [3, 42, 49], without any CIR supervision. Supervised CIR would obviously yield much higher accuracy in retrieval but requires expensive two-stage data collection processes - collecting pairs of related reference and target images, and then annotating them with visual delta that depicts the difference between them. On the other hand, zero-shot CIR does not incur additional labeling costs and utilizes web-collected noisy image text caption pairs directly. However, it has a much lower performance bar compared to supervised approaches and lacks the ability to specialize in specific CIR domain tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we investigate a class of CIR called semi-supervised CIR, blending supervised and unsupervised samples to enhance generalization (Fig. 1(b)). This" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 155, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 155, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 155, + 713 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16805" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "content": "method focuses on boosting CIR performance in specific retrieval domains by creating new triplets from unsupervised data. Building on this concept, we introduce a novel technique, Visual Delta Generator (VDG), designed to tap into the extensive natural language capabilities of Large Language Models (LLMs) [5, 21, 47, 48]. Our approach involves projecting reference and target images from supervised CIR triplets into the language embedding space, making them suitable inputs for the LLM. We then fine-tune the model by using prompts such as 'Describe the differences between images.' to induce it to yield human-like visual delta as illustrated in Fig. 2. Furthermore, we employ a parameter-efficient fine-tuning technique, LoRA [19], on the LLM. This choice of design effectively enhances the quality of visual deltas while also preserving the LLM's original capabilities, without harming its inherent knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 564 + ], + "type": "text", + "content": "After the VDG is trained, it knows how to distinguish between a given reference and target image and produce visual delta as a textual response. If we forward two similar images with different compositions, we can thus obtain the corresponding visual delta easily with the VDG. This allows us to achieve two purposes. First, we can now augment existing CIR triplets by adding generated visual deltas to pairs of reference and target images from the training set. Second, we can also harvest new reference-target pairs from an unlabeled database based on visual similarity, after which we forward these images to VDG to configure new pseudo triplets for CIR training. Note that our VDG is model agnostic - it simply increases the number of triplet candidates for training any given supervised CIR baselines. This strategy strikes a balance between maintaining the integrity of supervisory concepts derived from a supervised dataset and the capacity for effortless expansion using new, unlabeled image samples. It's a cost-effective and scalable solution, ensuring uniformity in annotations across extensive datasets. By generating pseudo triplets with VDG, we significantly reduce annotation costs and enhance the performance of CIR models trained solely on supervised learning, as well as those trained without supervised triplets. Our approach leads to state-of-the-art results in CIR benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 565, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 565, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 59, + 565, + 286, + 578 + ], + "type": "text", + "content": "The key advantages of our semi-supervised CIR include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 582, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 629 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to transfer knowledge from Large Language Models (LLMs) and connect it with semi-supervised Composed Image Retrieval (CIR)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "content": "- We propose a novel Visual Delta Generator (VDG) that generates synthetic visual deltas for augmenting the supervised dataset and allowing the integration of an auxiliary image gallery for CIR model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "- Comprehensive experimental results confirm the effectiveness of our method, demonstrating state-of-the-art retrieval rankings and showcasing the potential of our work." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 535, + 232 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 535, + 232 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 535, + 232 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 535, + 232 + ], + "type": "image", + "image_path": "13694db774af6b195b976f17288d5ec01aa18222a2b0d4febaca9dadd03cfb76.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 240, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 547, + 297 + ], + "type": "text", + "content": "Figure 2. An overview of the VDG tuning process. It includes (a) a vision projector and (b) a Large Language Model (LLM). The VDG is trained to produce visual delta that accurately describes the difference between a reference image and its corresponding target image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 316, + 392, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 392, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 392, + 329 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 337, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 578 + ], + "type": "text", + "content": "Composed Image Retrieval. The field of image retrieval has captured the interest of many researchers in our community [6, 38]. One notable area that has seen much progress recently is Composed Image Retrieval (CIR), a problem that focuses on retrieving images that best match a given pair of a query image and textual intent. Supervised CIR methods [2, 11, 35, 50] are trained on human-annotated triplets, consisting of a reference image, a target image, and their textual difference. On the other hand, zero-shot CIR [3, 9, 16, 42] operates without relying on human-guided descriptions of the differences between the two images. Instead, it uses noisy image-text pairs, aiming to find a function that can translate images into words. This approach, designed to discriminate subtle differences between images based solely on text captions, poses challenges, but also scales well, making it easy to add more data for CIR training. Addressing the limitations of both approaches, we explore the field of semi-supervised learning-based retrieval in this work, leveraging the strengths of both labeled and unlabeled data to enhance retrieval performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Semi-supervised Learning. Semi-supervised learning has been an active research topic in visual recognition for a long time [4, 18, 20, 25-27, 39, 44, 46, 51]. Semi-supervised learning can be roughly categorized into two groups, consistency learning [25, 39, 46] and pseudolabeling based learning [27, 44]. Consistency-based methods, as the name implies, encourage consistency in the output of the model by adding noise to model weights or using Exponential Moving Average (EMA) from a teacher model to a student model. In pseudo-labeling based meth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16806" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "ods, hard/soft pseudo-labels obtained from a pretrained model are assigned to unlabeled images [15, 44]. These pseudo-labels can be filtered using confidence thresholding or multi-view consistency. However, neither consistency-based nor pseudo labels methods are directly applicable to CIR. This is because the relative textual descriptions of the corresponding visual differences (visual delta) needed in CIR are not the intended outputs of these methods. In this work, we propose the Visual Delta Generator (VDG), a multi-modal pseudo-label generator. VDG processes two input images and generates text that describes their visual differences, making it an ideal candidate for constructing pseudo triplets for CIR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 289, + 529 + ], + "type": "text", + "content": "Multi-modal Models for Image-Text Retrieval. Models like CLIP [41] and BLIP [29] have showcased the advantages of training models on extensive image-text pairs, enabling precise alignment between language and vision representations that is crucial for image-text retrieval. Building on this, there have been significant advancements in utilizing Large Language Models (LLMs) for enhanced vision-language understanding [1, 10, 31, 33, 53]. Especially for CIR, Fromage [23] utilizes LLMs to directly produce embeddings for retrieval, allowing cross-modal compositional search with a single image and textual intent. CoVR [49] and SEARLE [3] apply LLMs to generate visual deltas using image captions without incorporating visual data, which limits the generation of accurate CIR training samples. Our proposed method overcomes these challenges by finetuning LLMs with the integration of a pretrained vision-language alignment module. This integration empowers LLMs to perceive and comprehend images, making them applicable even in scenarios with only image datasets. The method excels in generating accurate visual deltas, a key factor in training efficient external CIR models. These enhancements optimize the use of LLMs while ensuring computational efficiency, thereby expanding the versatility of LLMs in image-related tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 538, + 104, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 104, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 104, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": "Overview. Our goal is to establish a semi-supervised Composed Image Retrieval (CIR) system that merges image reference features with user textual descriptions to retrieve images from a large-scale database. We face a challenge in the limited availability of supervised triplets necessary for robust CIR model training. To overcome this limitation, we introduce a novel semi-supervised approach for CIR by leveraging an instruction-tuned Large Language Model (LLM), which we call Visual Delta Generator (VDG). The VDG learns to discriminate differences between two images and produces a textual response. This capability allows us to generate additional CIR training triplets, which in turn contributes to the development of more robust CIR models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "Sec. 3.1 provides detailed insights into the construction of the VDG. Sec. 3.2 describes the pseudo triplet generation process. Training of CIR models with pseudo triplets and our additional objective function for better optimization are described in Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 138, + 481, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 138, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 138, + 481, + 152 + ], + "type": "text", + "content": "3.1. Visual Delta Generator Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 156, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 359 + ], + "type": "text", + "content": "While semi-supervised learning methods have been actively developed for standard visual recognition tasks, these cannot be directly applied to CIR. In CIR, pseudo-label generation requires a detailed semantic understanding of two separate images such that their difference can be automatically expressed in the form of text. We leverage vision-language pretraining models and LLMs to achieve the requirements. With the huge success of LLMs, there are approaches that aim to utilize their understanding of the language domain for improving vision tasks. Particularly, LLaVA [33] and InstructBLIP [10] which are trained on top of the chat bot style instruction tuned LLM, Vicuna [7], have shown interesting results on vision-language tasks. Inspired by these, we propose VDG, which allows the LLM to take two images (reference, target) with similar contents and discriminate their difference in the form of text response (visual delta) as shown in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 360, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 491 + ], + "type": "text", + "content": "First, to enable the LLM to interpret images, we use the Querying Transformer (Q-Former) motivated by InstructBLIP [10] to prepare images for the LLM input (i.e., Vision Projector (VP) in Fig. 2(a)). Q-Former, a transformer encoder [12, 21], processes a fixed set of 32 learnable query tokens (embeddings). These tokens are modified through self-attention layers and interact with image features of Vision Transformer (ViT) [12] through cross-attention layers. As a result, the query tokens are infused with the visual information from the provided image, making them suitable for LLM processing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 505, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 636 + ], + "type": "text", + "content": "Stage 1: Alignment. The outputs of the VP are inherently not aligned with the tokenized word embeddings of the inputs to the LLM, which we chose to be LLaMA2 [48] in this work. Alignment between the VP and LLM needs to be attained by fine-tuning trainable projection of VP in Fig. 2(a) as was done in LLaVA [33]. Specifically, we employ large-scale image-caption pairs to foster alignment between the image representations understood by the VP and the LLM. This step includes minimizing standard next token prediction loss which is generally used to train the decoder-based LLM as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 652, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 652, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 652, + 545, + 685 + ], + "type": "interline_equation", + "content": "L _ {L L M} = - \\sum_ {t = 1} ^ {T} \\log P \\left(\\mathrm {w} _ {t} \\mid x, I _ {\\text {i n s t}}, \\mathrm {w} _ {(1.., t - 1)}; \\theta_ {\\text {p r o j}}\\right), \\tag {1}", + "image_path": "f5c88c9cf62adb657d9575884d08e8b96ba0fe9fe10c512ab0bd32f4c14c4952.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{w}_t" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " denotes a token at time step " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " is the length of the sequence, " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " is the probability assigned by the model" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16807" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "to the actual token " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{w}_t" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " given with the image " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", instruction " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "I_{inst}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and previous tokens " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{w}_{(1,\\dots,t - 1)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\theta_{proj}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is parameters of projection layer. Following standard practice, the prediction for each token is computed using a softmax over the vocabulary of the LLM, and the cross-entropy loss is computed between the predicted probabilities and the textual token labels as a classification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 327 + ], + "type": "text", + "content": "Stage 2: Instruction Tuning. In this stage, we conduct instruction tuning to equip our LLM with the ability to understand image pairs and produce visual delta, as shown in Fig. 2(b). Referring to Fig. 3, we train the LLM with an instruction (i.e., \"Request\") to generate the visual delta in textual format (i.e., \"Response\"). We forward two distinct images (i.e., \"Reference\" and \"Target\") into the LLM via the VP. We employ a parameter-efficient fine-tuning technique, LoRA [19], directly on the LLM. This fine-tuning process is designed to provide \"extra room\" for the LLM to undertake new tasks, all without compromising its foundational capabilities. In this stage, we introduce a specific prompt structure to the LLM as below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 340, + 272, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 340, + 272, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 340, + 272, + 361 + ], + "type": "text", + "content": "Request: Analyze given reference and target images and provide a description that transforms the reference to match the target." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 361, + 212, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 361, + 212, + 372 + ], + "spans": [ + { + "bbox": [ + 60, + 361, + 212, + 372 + ], + "type": "text", + "content": "Reference: " + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 372, + 187, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 372, + 187, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 372, + 187, + 380 + ], + "type": "text", + "content": "Target: " + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 381, + 192, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 381, + 192, + 391 + ], + "spans": [ + { + "bbox": [ + 61, + 381, + 192, + 391 + ], + "type": "text", + "content": "Response: " + }, + { + "bbox": [ + 61, + 381, + 192, + 391 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 61, + 381, + 192, + 391 + ], + "type": "text", + "content": " Visual delta (text label)>" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "content": "which guides the LLM to generate the corresponding visual delta. We leverage the same training loss from Eqn. 1 to train LoRA parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 468, + 234, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 234, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 234, + 480 + ], + "type": "text", + "content": "3.2. Pseudo Triplet Generation for CIR" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "After training VDG, we can generate visual delta of two images, which can be used to form pseudo triplets for CIR. However, it is important to choose two images that not only share certain attributes and similarities but also present other distinct attributes (i.e., visual delta). To gather suitable pairs of reference and target images, we utilize an image encoder to select them from an auxiliary image gallery, which we denote as " + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": ". Note that, we note upper strophe (') on samples and embeddings that are obtained from " + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": ", in the following." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Following the strategy in CIRR [36], we start with an anchor image " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " and retrieve the top 20 images from " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " using cosine similarity scores between ResNet 152 [17] embeddings, pretrained on ImageNet [24]. We exclude images with scores above 0.94 and sequentially add images to a subgroup of size 6, skipping those within a 0.002 score of the previously included image. As depicted in Fig. 4, we establish dense connections between all pairs (both consecutive, represented by the outer circle, and non-consecutive," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 536, + 167 + ], + "blocks": [ + { + "bbox": [ + 63, + 400, + 271, + 412 + ], + "lines": [ + { + "bbox": [ + 63, + 400, + 271, + 412 + ], + "spans": [ + { + "bbox": [ + 63, + 400, + 271, + 412 + ], + "type": "text", + "content": "Figure 3. A template prompt for VDG instruction tuning." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 69, + 536, + 167 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 536, + 167 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 536, + 167 + ], + "type": "image", + "image_path": "b146655b3f6dbd991aec33738b9171248f9f038243e4aa288d6296248d943bac.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 174, + 547, + 219 + ], + "lines": [ + { + "bbox": [ + 304, + 174, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 547, + 219 + ], + "type": "text", + "content": "Figure 4. The process of pseudo triplet generation. First, an image subgroup is constructed based on visual similarity (left). Then, paired reference and target images are fed into the VDG to generate the visual delta, completing the triplet formation (right)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "text", + "content": "represented by the dotted inner connections), while ensuring no overlaps. The arrow's starting point denotes the reference, while its endpoint indicates the target. Given images " + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime r}" + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "inline_equation", + "content": "x_{j}^{\\prime t}" + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "text", + "content": ", VDG produces " + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\delta_{i,j}^{\\prime}" + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "text", + "content": ". This allows us to formulate the pseudo triplet as " + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\{x_{i}^{\\prime r}, x_{j}^{\\prime t}, \\delta_{i,j}^{\\prime}\\}" + }, + { + "bbox": [ + 304, + 239, + 545, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 308, + 471, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 471, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 471, + 320 + ], + "type": "text", + "content": "3.3. Semi-supervised CIR Training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "text", + "content": "Preliminaries. Suppose we have access to a CIR dataset of triplets: " + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "inline_equation", + "content": "D = \\{(x_i^r, x_j^t, \\delta_{n(i,j)})\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "inline_equation", + "content": "x_i^r, x_j^t" + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "text", + "content": " denote reference and target images, respectively, " + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\delta_{n(i,j)}" + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "text", + "content": " represents their visual delta, and " + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 326, + 545, + 374 + ], + "type": "text", + "content": " denotes the total triplet counts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 374, + 546, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 482 + ], + "type": "text", + "content": "In the pursuit of enhancing CIR through a semi-supervised approach, our method's strength lies in its model-agnosticism, allowing for seamless integration with a variety of CIR models. Following recent trends, we opt to use the encoders from vision-language pretraining models [10, 29, 31, 41], notably CLIP and BLIP, as our baseline backbones. These encoders are naturally equipped to understand and convert both visual and textual elements into a joint embedding space, making them suitable for CIR tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": "The image encoder takes patchified image token embeddings " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "x = [\\mathrm{x}_1, \\dots, \\mathrm{x}_{K_{img}}]" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " with the learnable image " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "cls" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " token embedding " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "[\\mathrm{x}_{cls}]" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": ", and outputs the visual feature embeddings " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "E_{img}(\\mathrm{x}_{cls}, x) = [\\hat{\\mathrm{x}}_{cls}, \\hat{\\mathrm{x}}_1, \\dots, \\hat{\\mathrm{x}}_{K_{img}}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{x}} \\in R^{d_i}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " dimension, and " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "K_{img}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " denotes the number of generated tokens from each image. The text encoder processes tokenizer output embedding of visual delta " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "\\delta = [\\mathrm{z}_1, \\dots, \\mathrm{z}_{K_{txt}}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " with the learnable text " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "cls" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " token embeddings " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "[\\mathrm{z}_{cls}]" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " to produce its textual feature embeddings " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "E_{txt}(\\mathrm{z}_{cls}, \\delta) = [\\hat{\\mathrm{z}}_{cls}, \\hat{\\mathrm{z}}_1, \\dots, \\hat{\\mathrm{z}}_{K_{txt}}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{z}} \\in R^{d_t}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " dimension, and " + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "inline_equation", + "content": "K_{txt}" + }, + { + "bbox": [ + 304, + 483, + 547, + 625 + ], + "type": "text", + "content": " denotes the number of generated text tokens from each visual delta." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Model Architecture. To carry out CIR, we establish a fusion function " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " that takes a reference image " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " and a visual delta " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": ", producing a composed embedding " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "f(x^r, \\delta) = \\mathbf{c}" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": ", and notates its trainable components as " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": ". We utilize two well-known backbones, CLIP and BLIP, to configure " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": ". In the case of CLIP, we employ the text encoder " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "E_{txt:\\theta}" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " and an" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16808" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "additional Combiner module [2] " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\theta}(\\hat{\\mathrm{x}}_{cls}^{r},\\hat{\\mathrm{z}}_{cls})" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " that outputs " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " to be the components of " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{\\mathrm{CLIP}} = \\{E_{txt:\\theta},\\mathcal{C}_{\\theta}\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "). The Combiner is designed to optimally blend " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{x}}_{cls}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{z}}_{cls}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " carefully weighing their individual impacts while adeptly mixing them. In the BLIP case, we exclusively use BLIP's text encoder, grounded in the image, and designate it as the trainable " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "E_{txt:\\theta}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{\\mathrm{BLIP}} = E_{txt:\\theta}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\hat{\\mathbf{z}}_{cls}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "), without incorporating additional modules. BLIP's text encoder inherently fuses image and text signals in its cross-attention layer, eliminating the need for a separate combining module. Notably, we freeze all vision encoders in our setup to ensure compatibility with existing image retrieval galleries and to enhance training efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": "Supervised / Pseudo Separated Contrastive Loss. The training objective of CIR is to achieve strong alignment between the target image's embedding " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\hat{\\mathbf{x}}_{cls}^{t}" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": " for simplicity), and the composed embedding " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": ". On this purpose, we utilize HN-NCE [40] loss for a given training batch " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\sim D" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{B}' \\sim D'" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": " contains pseudo triplets. Additionally, to mitigate the impact of noise in pseudo triplets and ensure consistent contributions from supervised triplets, we compute a target-composed contrastive loss (tcc) as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 373, + 287, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 373, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 83, + 373, + 287, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t c c} (\\mathcal {B}, \\mathcal {B} ^ {\\prime}) = \\mathcal {L} _ {c} (\\mathcal {B}; \\tau) + \\mathcal {L} _ {c} (\\mathcal {B} \\oplus \\mathcal {B} ^ {\\prime}; \\tau) \\tag {2}", + "image_path": "662c9f637742ee5f8a453e011e453b518bffce6a8ecf7e0b424f02d25849bed2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 143, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 143, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 143, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 392, + 143, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 47, + 392, + 143, + 403 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 421, + 290, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 290, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 290, + 489 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {c} (\\mathcal {B}) = - \\sum_ {i = 1} ^ {n} \\log \\left[ \\frac {e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau}}{\\alpha \\cdot e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau} + \\sum_ {j \\neq i} ^ {\\mathcal {B}} e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {j} / \\tau} \\cdot w _ {\\mathbf {x} _ {i} , \\mathbf {c} _ {j}}} \\right] \\\\ - \\sum_ {i = 1} ^ {n} \\log \\left[ \\frac {e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau}}{\\alpha \\cdot e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {i} / \\tau} + \\sum_ {j \\neq i} ^ {\\mathcal {B}} e ^ {\\mathbf {x} _ {i} ^ {T} \\mathbf {c} _ {j} / \\tau} \\cdot w _ {\\mathbf {x} _ {j} , \\mathbf {c} _ {i}}} \\right]. \\\\ \\end{array}", + "image_path": "dc22333efd69469883ecb7f1149da8b08a132574b5e5b33d2fe6b8f1f1578035.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\tau, \\alpha" + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": " denote hyper-parameters, and " + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": " denotes concatenation along the batch axis, and " + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{x}_i,\\mathbf{c}_j}, w_{\\mathbf{x}_j,\\mathbf{c}_i}" + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": " are set as in [40]. This design facilitates the independent yet concurrent investigation of both supervised and semi-supervised CIR embedding spaces in a contrastive manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Target-Delta Matching Loss. In the case of BLIP's image-grounded text encoder structure, we introduce a new target-delta matching loss (tdm). Our insight stems from the observation that, while all reference image patch tokens are considered in the cross-attention layer of the BLIP text encoder, target image patch tokens are overlooked when training solely with contrastive learning betweencls token embeddings. As an example shown in Fig. 5, the visual delta can be seen as a weakly correlated caption to the target image. Thus, we aim to align the target image and visual delta to enable the BLIP text encoder to identify the image tokens related to textual input. The tdm loss is applied as:" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 536, + 286 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 536, + 286 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 536, + 286 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 536, + 286 + ], + "type": "image", + "image_path": "d9acc4d73c9bb68c4f258615b7f3979d99bf5767bc72d0f8adbef05669e037d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "type": "text", + "content": "Figure 5. Illustration of our proposed adaptation of the BLIP image-grounded text encoder for CIR. Both reference " + }, + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "type": "inline_equation", + "content": "(x^r)" + }, + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "type": "inline_equation", + "content": "(x^t)" + }, + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "type": "text", + "content": " patch tokens are processed by the text encoder " + }, + { + "bbox": [ + 305, + 292, + 547, + 327 + ], + "type": "inline_equation", + "content": "(f_{\\theta})" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 335, + 367, + 545, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 367, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 335, + 367, + 545, + 383 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t d m}} = \\mathbb {E} _ {(x ^ {t}, \\delta) \\sim (D, D ^ {\\prime})} \\mathrm {H} \\left(\\mathbf {y} ^ {t d m}, \\mathbf {p} ^ {t d m} \\left(x ^ {t}, \\delta\\right)\\right) \\tag {3}", + "image_path": "df4d22f0c4c14525597baf03fd00e8e4cd9f623b69d014261b9b9406491e04c9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{H}" + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "content": " is cross entropy, " + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{tdm}" + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "content": " is a 2-dimensional one-hot vector label obtained through the hard negative mining process proposed in [30], where a pair of image and text is configured as matched or not. " + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{tdm} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "content": " generates both positive and negative target-delta matching scores and is computed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 470, + 502, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 470, + 502, + 494 + ], + "spans": [ + { + "bbox": [ + 349, + 470, + 502, + 494 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {t d m} (x ^ {t}, \\delta) = \\sum_ {i = 1} ^ {K ^ {t x t}} p _ {\\theta} (\\hat {\\mathbf {z}} _ {i} ^ {t}) / K ^ {t x t},", + "image_path": "c23ccef275b21f561088cfe0e4f5f58bbabfcefbfe585b7191de6f610003f863.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "inline_equation", + "content": "\\hat{z}_i^t" + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "text", + "content": "-th token embedding in " + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x^{t},\\delta)" + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 304, + 501, + 547, + 550 + ], + "type": "text", + "content": " is a FC layer that outputs a 2-D vector. This loss ensures the text encoder fully processes the target image features, improving its understanding of the visual delta." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 560, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 388, + 574 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 579, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 617 + ], + "type": "text", + "content": "Sec. 4.1 details the setup and VDG generation. Sec. 4.2 covers quality checks of visual deltas. Evaluations and baseline comparisons are in Sec. 4.3, and further analyses in Sec. 4.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 356, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 356, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 356, + 636 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Implementation Details. We utilize InstructBLIP's pretrained weights [10], which have been trained with both a ViT-G/14 [13] and the Q-Former based on the Vicuna-13B model [7], without using prompts for Q-Former. For the stage 1 training described in Sec. 3.1, we employ 595K filtered image-text pairs from CC3M [43] provided by [33] to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16809" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "find alignment with our baseline LLM, LLaMA2-13B [48]. In Stage 2, we implement instruction tuning with LoRA parameters [19], following the fixed prompt as outlined in Fig. 3. Visual deltas are generated in an autoregressive manner, predicted based on the LLaMA2 vocabulary. For the CLIP-based CIR training, we select the ViT-L/14 model combined with a Combiner [2]. For the BLIP-based model [29], we use a dedicated BLIP text encoder for image-text matching, paired with the ViT-L/16 model. Additional details can be found in the appendix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 453 + ], + "type": "text", + "content": "Datasets for CIR Evaluation. There are two standard benchmarks in CIR, one is CIRR [36] which deals with natural images, and the other is FashionIQ [52] which focuses on fashion domain images. Each presents unique challenges and datasets that help researchers push the boundaries of what's possible in CIR. Following the protocols utilized in benchmarks [36, 52], we report the CIR results with recall scores at top K retrieval results (R@K), or results under collected subset " + }, + { + "bbox": [ + 46, + 213, + 289, + 453 + ], + "type": "inline_equation", + "content": "(\\mathrm{R}_s@\\mathrm{K})" + }, + { + "bbox": [ + 46, + 213, + 289, + 453 + ], + "type": "text", + "content": ". Specifically, CIRR is configured with 4,351 subgroups (subsets), each containing six similar images, sourced from NLVR2 [45]. For experimental purposes these groups are distributed into train (3,345 subgroups of 16,742 images), validation (503 subgroups of 2,265 images), and test (503 subgroups of 2,178 images) sets. FashionIQ is divided into three categories of Dress, Shirt, and Toptee (Tops and Tees). The reference and target images are paired based on their category similarities. The 18,000 CIR triplets for training are pooled from 45,429 images of the training set, and 6,016 CIR triplets for the test are chosen from 15,415 images of the validation set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "Visual Delta Generation. To produce pseudo triplets, we expand our dataset sources to configure an auxiliary gallery " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "(G^{\\prime})" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": " which is built upon the grouping strategy introduced in Sec. 3.2, while excluding images that overlap with the benchmark sets. Once the subgroups are constructed, we further filter them to avoid heavy overlap. In total, we draw upon 42,390 unique subgroups from NLVR2 [45], and 79,427 from COCO [32]. Similarly, we build 27,957 individual subgroups from FashionIQ [52], and 30,880 from DeepFashion [34]. We mark upper strophe (') to these datasets. For the semi-supervised settings, we randomly select 3,345 groups from NLVR2' and COCO' for the CIRR case, as well as 3,600 groups from FashionIQ' and DeepFashion'. We denote these datasets used for semi-supervised CIR with the subscript " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "se" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": ", (e.g., NLVR2'se, COCO'se). Fig. 6 shows the comparison between human-annotated visual deltas and those generated by the VDG in both natural and fashion domains. We observe that the VDG is effective in generating high-quality visual deltas - additional results can be found in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 70, + 533, + 213 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 533, + 213 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 533, + 213 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 533, + 213 + ], + "type": "image", + "image_path": "147fb22b2bec61b0e22255dd01a8ba62a80aa236fc914815918287634d2d9360.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 220, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 255 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on visual deltas, human vs. VDG on CIRR and FashionIQ datasets. For both natural and fashion domain images, VDG can produce informative visual deltas." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 325, + 538, + 458 + ], + "blocks": [ + { + "bbox": [ + 304, + 260, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 304, + 260, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 547, + 316 + ], + "type": "text", + "content": "Table 1. Retrieval results on CIRR validation set. Human + VDG represents utilizing both human annotated and VDG-generated visual deltas when training the CIR model. " + }, + { + "bbox": [ + 304, + 260, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{tdm}" + }, + { + "bbox": [ + 304, + 260, + 547, + 316 + ], + "type": "text", + "content": " represents a model with target-delta matching loss. The best scores for each Val. set supervision are highlighted in bold." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 325, + 538, + 458 + ], + "lines": [ + { + "bbox": [ + 312, + 325, + 538, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 325, + 538, + 458 + ], + "type": "table", + "html": "
Training Set SupervisionVal. Set SupervisionBaselineR@1R@5R@10R@50
(a) HumanHumanCombiner37.9871.4982.5295.29
BLIPtdm53.1782.0989.8197.54
(b) VDGHumanCombiner35.4768.2980.2994.31
BLIPtdm50.1680.0387.7896.75
(c) Human + VDGHumanCombiner39.1173.0284.4195.96
BLIPtdm53.6782.9989.9797.84
(d) HumanVDGCombiner38.9673.3184.3696.22
BLIPtdm51.6483.3391.3997.90
(e) VDGVDGCombiner41.5977.5787.3597.11
BLIPtdm52.6985.1792.3798.59
(f) Human + VDGVDGCombiner41.4777.9087.4297.30
BLIPtdm52.9585.2792.5498.61
", + "image_path": "cb3f4c256f41aa2cece4799bb99f759b6c85fd8551f8b7e6b8e4ca3ba265ac4f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 479, + 487, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 487, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 487, + 493 + ], + "type": "text", + "content": "4.2. Quality Check of VDG Responses" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": "To assess the quality of VDG-generated visual deltas, we executed a series of experiments as outlined in Table 1. We use two backbones, Combiner, and BLIP with the tdm loss (denoted as " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{\\mathrm{tdm}}" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": "), as our baselines for evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "text", + "content": "Initially, we compare the visual deltas generated by VDG with those annotated by humans. This is done by replacing the deltas in the training set with VDG-generated deltas for the same reference-target pairs. When comparing (a) and (b), we observe only a marginal drop in performance upon switching to VDG-generated deltas. This indicates that the deltas generated by VDG are as effective as those created by humans. More importantly, the improved performance observed when comparing (a) with (c) — which combines human-annotated and VDG-generated deltas — highlights the effectiveness of this hybrid approach as a robust data augmentation strategy to enhance CIR models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "To evaluate VDG's performance on new image pairs, we conduct experiments by substituting the human-annotated" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16810" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 162, + 294, + 430 + ], + "blocks": [ + { + "bbox": [ + 46, + 65, + 289, + 153 + ], + "lines": [ + { + "bbox": [ + 46, + 65, + 289, + 153 + ], + "spans": [ + { + "bbox": [ + 46, + 65, + 289, + 153 + ], + "type": "text", + "content": "Table 2. Retrieval results on CIRR test set. * denotes our baselines, † denotes VDG generated visual deltas are applied to augment original training set. We categorize into two distinct groups: one is Seen: Supervised / Supervised + External / Supervised + Aux. with VDG which utilize human-annotated visual delta for CIR model training, and the other is Unseen: Zero-shot / Aux. with VDG which does not utilize human-annotated visual delta for CIR model training. Within each category, best viewed with **bold**." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 162, + 294, + 430 + ], + "lines": [ + { + "bbox": [ + 47, + 162, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 162, + 294, + 430 + ], + "type": "table", + "html": "
MethodR@1R@5R@10R@50Rs@1Rs@2Rs@3
(a) Supervised
ARTEMIS [11]16.9646.1061.3187.7339.9962.2075.67
CIRPLANT [35]19.5552.5568.3992.3839.2063.0379.49
Combiner [2]33.5965.3577.3595.2162.3981.8192.02
CASE [28]48.0079.1187.2597.5775.8890.5896.00
CoVR [49]48.8478.0586.1094.1975.7888.2292.80
Combiner34.3966.2276.5891.0468.5586.3693.98
\\( Combiner^† \\)36.9169.2179.5492.0470.0087.4594.39
\\( BLIP_{tdm} \\)48.9477.8386.1594.1775.7189.7195.81
\\( BLIP_{tdm}^† \\)49.0878.9886.8994.2476.1890.6295.86
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]49.3580.0288.7597.4776.4890.3795.71
CoVR + WebVid [49]49.6978.6086.7794.3175.0188.1293.16
(c) Supervised + Auxiliary Gallery with VDG
\\( Combiner^† + COCO_{se}^{\\prime} \\)38.7769.2579.2191.5271.2587.4994.34
\\( Combiner^† + NLVR2_{se}^{\\prime} \\)38.8969.8479.4191.1871.9287.8994.29
\\( BLIP_{tdm}^† + COCO_{se}^{\\prime} \\)49.3778.1285.5293.7476.6890.4696.05
\\( BLIP_{tdm}^† + NLVR2_{se}^{\\prime} \\)50.9680.1586.8694.4677.4590.6596.10
(d) Zero-shot
Pic2Word by CC3M [42]23.9051.7065.3087.80---
SEARLE by ImageNet [3]24.2252.4166.2988.6353.7174.6387.61
CASE by LasCo.Ca. [28]35.4065.7878.5394.6364.2982.6691.61
CoVR by WebVid [49]38.4866.7077.2591.4769.2883.7691.11
(e) Auxiliary Gallery with VDG
Combiner by COCO'26.8054.0565.3083.8867.2885.4292.99
Combiner by NLVR2'31.5761.3772.1088.9467.9886.1893.18
\\( BLIP_{tdm} \\)by COCO'43.4972.0781.5993.2172.3687.9894.72
\\( BLIP_{tdm} \\)by NLVR2'45.7475.0182.5293.1372.6087.9094.77
", + "image_path": "bcbd7fd9c426ce0788ffda519723fa9cfd77dbffb43d111d9dba683eb21ba3a9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "text", + "content": "deltas in the validation set with deltas generated by VDG. The comparison of (a) and (d) reveals comparable performance levels, suggesting that deltas generated by VDG are as effective as human annotations. When we incorporate VDG-generated deltas into the training set (cases (e) and (f)), we achieve a notable performance improvement over the human-annotated validation set across all metrics, with the exception of a small reduction in " + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "inline_equation", + "content": "\\mathbb{R}@\\mathbb{1}" + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "text", + "content": ". For example, there is a " + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "inline_equation", + "content": "5.14\\%" + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "text", + "content": " p increase in " + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "inline_equation", + "content": "\\mathbb{R}@\\mathbb{5}" + }, + { + "bbox": [ + 46, + 454, + 289, + 658 + ], + "type": "text", + "content": " between (b) and (e). The improvement observed may stem from the varied text descriptions by different human annotators, leading to inconsistencies in style and detail that potentially affect the distribution of visual differences between the annotated triplets. VDG offers a more consistent solution in generating visual deltas, thereby reducing errors associated with differences in human annotation. Our comprehensive testing demonstrates VDG's reliability as an annotation tool for CIR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 670, + 226, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 226, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 226, + 683 + ], + "type": "text", + "content": "4.3. Comparison with Other Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Tables 2 and 3 present comparative evaluations between our method and other CIR approaches on CIRR and Fash" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 95, + 552, + 384 + ], + "blocks": [ + { + "bbox": [ + 305, + 65, + 545, + 87 + ], + "lines": [ + { + "bbox": [ + 305, + 65, + 545, + 87 + ], + "spans": [ + { + "bbox": [ + 305, + 65, + 545, + 87 + ], + "type": "text", + "content": "Table 3. Retrieval results on FashionIQ validation set. We utilize the same notations and same categorizations as Table 2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 95, + 552, + 384 + ], + "lines": [ + { + "bbox": [ + 306, + 95, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 552, + 384 + ], + "type": "table", + "html": "
MethodDressShirtToptee
R@10R@50R@10R@50R@10R@50
(a) Supervised
ARTEMIS [11]27.1652.4021.7843.6429.2053.83
DCNet [22]28.9556.0723.9547.3030.4458.29
FashionVLP [14]32.4260.2931.8958.4438.5168.79
Combiner [2]31.6356.6736.3658.0038.1962.42
CoVR [49]43.5167.9448.2866.6851.5373.60
*Combiner31.9555.0539.2156.8238.5562.16
*Combiner†35.4059.9942.3061.6343.0966.96
*BILIPtdm44.8766.8349.6166.9350.5472.26
*BILIP†tdm46.9068.8650.2868.0452.7374.45
(b) Supervised + External Dataset for Pretraining
CASE + LasCo.Ca. [28]47.4469.3648.4870.2350.1872.24
CoVR + WebVid [49]44.5569.0348.4367.4252.6074.31
(c) Supervised + Auxiliary Gallery with VDG
Combiner† + DeepFashion′se35.5060.0942.5462.8643.2767.86
Combiner† + FashionIQ′se36.3060.1943.9862.2744.3368.06
BILIPtdm + DeepFashion′se47.1069.1049.9569.9653.9074.35
BILIP†tdm + FashionIQ′se47.8969.8151.3671.0853.2974.65
(d) Zero-shot
Pic2Word by CC3M [42]20.0040.2026.2043.6027.9047.40
SEARLE by ImageNet [3]20.3243.1827.4345.6829.3250.17
CoVR by WebVid [49]21.9539.0530.3746.1230.7848.73
(e) Auxiliary Gallery with VDG
Combiner by DeepFashion'23.3046.3630.8649.0231.8751.96
Combiner by FashionIQ'28.2651.4632.5851.2834.8855.79
BILIPtdm by DeepFashion'32.6754.3935.4855.0539.4759.92
BILIP†tdm by FashionIQ'37.4858.7037.2957.1142.1262.32
", + "image_path": "9388c2d17df4469751d9e58f2d7f20090b9757a4dd36f8a98ae25a0f1530df37.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 411, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 651 + ], + "type": "text", + "content": "ionIQ evaluation protocols. We categorize the experiments into two distinct groups. The first group, Seen, includes scenarios where the CIR model is trained with human-annotated training triplets. In contrast, the second group, Unseen, comprises scenarios in which the CIR model is trained without human-annotated triplets. Our implementations of Combiner and BLIP are denoted by grey box. Specifically, methods under (a) are trained with supervised triplets, while those notated with the symbol " + }, + { + "bbox": [ + 304, + 411, + 547, + 651 + ], + "type": "inline_equation", + "content": "(\\dagger)" + }, + { + "bbox": [ + 304, + 411, + 547, + 651 + ], + "type": "text", + "content": " are additionally augmented with VDG-generated visual deltas from supervised image pairs (i.e., the same setting as Human + VDG in Table 1). The methods in (b) have been pretrained on external datasets for CIR tasks, resulting in enhanced performances. In our semi-supervised setup (c), the models are trained with augmented supervised triplets and pseudo triplets sub-sampled from the auxiliary gallery as described in Secs. 3.2 and 4.1. Methods in (d) are trained with large-scale datasets with the intention of producing zero-shot performances. In (e), we only use pseudo triplets generated from the auxiliary gallery for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "The results clearly demonstrate that our VDG implementation significantly improves the performance of both Combiner and " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{tdm}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " baselines. In (a), a detailed comparison shows that models augmented with VDG samples (with " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ") consistently outperform their counterparts (without " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ") in all" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16811" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 91, + 260, + 140 + ], + "blocks": [ + { + "bbox": [ + 47, + 65, + 287, + 87 + ], + "lines": [ + { + "bbox": [ + 47, + 65, + 287, + 87 + ], + "spans": [ + { + "bbox": [ + 47, + 65, + 287, + 87 + ], + "type": "text", + "content": "Table 4. Ablation results on " + }, + { + "bbox": [ + 47, + 65, + 287, + 87 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{tdm}" + }, + { + "bbox": [ + 47, + 65, + 287, + 87 + ], + "type": "text", + "content": " baseline for CIRR validation set with NLVR2' as auxiliary gallery. Best viewed with bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 91, + 260, + 140 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 260, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 260, + 140 + ], + "type": "table", + "html": "
groupingconcat.tdm lossR@1R@5R@10R@50
52.4581.1888.62,96.88
✓55.6683.4591.2197.33
✓✓56.2884.3991.4197.92
✓✓✓57.8885.5893.2198.33
", + "image_path": "53295b57ed81bf034a4429dbe88ed736c84dc39e29d12f743e4fbd1aa17ac7b6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 145, + 164, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 145, + 164, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 145, + 164, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 145, + 164, + 202 + ], + "type": "image", + "image_path": "612c8710939a4a85a99ea0a46c08a4506ddcf9a947748bb2a0738cb7373a7950.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 205, + 140, + 215 + ], + "lines": [ + { + "bbox": [ + 75, + 205, + 140, + 215 + ], + "spans": [ + { + "bbox": [ + 75, + 205, + 140, + 215 + ], + "type": "text", + "content": "(a) CIRR validation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "type": "text", + "content": "Figure 7. Analysis on the scale for (c) Supervised + Auxiliary Gallery with " + }, + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "type": "inline_equation", + "content": "VDG" + }, + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{NLVR2}_{se}^{\\prime}" + }, + { + "bbox": [ + 46, + 219, + 287, + 251 + ], + "type": "text", + "content": " and FashionIQ'se. The x-axis denotes the ratio of additional images to training images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 253, + 287, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 253, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 287, + 285 + ], + "type": "text", + "content": "Table 5. Analysis on the scale for (e) Auxiliary Gallery with VDG. The dataset is scaled from one eighth (1/8) to the full set (1). Evaluated on CIRR Test set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 145, + 280, + 202 + ], + "blocks": [ + { + "bbox": [ + 167, + 145, + 280, + 202 + ], + "lines": [ + { + "bbox": [ + 167, + 145, + 280, + 202 + ], + "spans": [ + { + "bbox": [ + 167, + 145, + 280, + 202 + ], + "type": "image", + "image_path": "02d04f82ef67ea5b2423f043e1f9a3a81527c27574530da36b44a178f40654af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 205, + 256, + 215 + ], + "lines": [ + { + "bbox": [ + 190, + 205, + 256, + 215 + ], + "spans": [ + { + "bbox": [ + 190, + 205, + 256, + 215 + ], + "type": "text", + "content": "(b) FashionIQ-Shirt" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 289, + 288, + 322 + ], + "blocks": [ + { + "bbox": [ + 48, + 289, + 288, + 322 + ], + "lines": [ + { + "bbox": [ + 48, + 289, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 288, + 322 + ], + "type": "table", + "html": "
R@11/81/41/21R@101/81/41/21
COCO42.1542.8943.1043.49COCO79.8181.0181.1681.59
NLVR244.3144.9945.1345.74NLVR282.4882.7582.4382.52
", + "image_path": "b181be67b10a668cccf4b5f0d2d71870caa8be75417d8a96d4b0f67778f916e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": "recall metrics. Moreover, in (c), when we enhance our CIR baselines with additional pseudo triplets, there is a notable performance improvement. Specifically, when " + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{tdm}" + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": " is combined with the auxiliary gallery, it surpasses the previous state-of-the-art results in most recall metrics, showcasing the advantages of VDG. Additionally, it's important to highlight that our semi-supervised CIR method achieves these improvements with considerably fewer images than methods like CASE or CoVR, demonstrating its efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 436, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 604 + ], + "type": "text", + "content": "In the unseen scenario, our implemented Combiner baseline not only outperforms Pic2Word, which uses the same CLIP backbone, but our " + }, + { + "bbox": [ + 46, + 436, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{tdm}" + }, + { + "bbox": [ + 46, + 436, + 287, + 604 + ], + "type": "text", + "content": " also significantly exceeds the former best-performing SEARLE or CoVR in most metrics. VDG's key advantage is its ability to generate visual deltas that align with the targeted domain, such as using COCO for natural images and DeepFashion for fashion images, by solely utilizing images without any accompanying captions. As a result, VDG significantly reduces the number of training samples required, yet achieves enhanced retrieval performances compared to zero-shot approaches like Pic2Word, CASE, or CoVR. This demonstrates that by focusing on similar domain, image-only datasets, we can substantially improve the efficacy of CIR model training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 611, + 151, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 151, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 151, + 624 + ], + "type": "text", + "content": "4.4. Further Analyses" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "Ablation study on Each Component. To showcase the effectiveness of our proposed learning strategies, we conducted an ablation study and presented the findings in Table 4. We consider three scenarios: (1) without grouping, where we randomly sample images from the auxiliary gallery to create pairs; (2) without concat., where we remove the concatenation step used in Eqn. 2 and use pseudo triplets only;" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 536, + 175 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 536, + 175 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 536, + 175 + ], + "type": "image", + "image_path": "94c62f3e17c01e2d5119a6c86fea7816a6988019622ffb34814861089bf7b592.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "content": "Figure 8. Retrieval results on auxiliary galleries, COCO, and DeepFashion. Actual user intents are used as text queries." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "type": "text", + "content": "and (3) without " + }, + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "type": "inline_equation", + "content": "tdm" + }, + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "type": "text", + "content": " loss, where we exclude the tdm loss during training. The results indicate that each element sufficiently contributes to enhancing the performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 243, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 445 + ], + "type": "text", + "content": "Impact of Generated Data Scale. Figs. 7a and 7b demonstrate the effect of the size of VDG-generated data. In a semi-supervised setup, we find that the performance peaks when the number of pseudo-selected images approximates the size of the training set, and then saturates. This saturation might arise because additional pseudo triplets fail to represent the test sample distribution, particularly as we rigorously remove overlapping images with the test set when forming " + }, + { + "bbox": [ + 304, + 243, + 546, + 445 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 304, + 243, + 546, + 445 + ], + "type": "text", + "content": ". For the unseen case as shown in Table 5, performance is enhanced with an increase in VDG-generated data. This suggests that even in the absence of supervised triplets for CIR model training, a larger set of pseudo triplets not only expands the model's comprehension of visual compositions but also aligns more closely with the distribution of supervised triplets. This alignment facilitates improved targeted domain retrieval performances, striking a balance between generalization and performance enhancement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 445, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 505 + ], + "type": "text", + "content": "Qualitative Retrieval Results. We facilitate retrieval by providing user intents along with query images, drawing from auxiliary galleries. The retrieval results, as depicted in Fig. 8, demonstrate accurate and relevant image retrieval. Additional results are detailed in the appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 516, + 378, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 378, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 378, + 528 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 536, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 668 + ], + "type": "text", + "content": "In this study, we investigate a novel semi-supervised learning approach in the context of Composed Image Retrieval (CIR). Our findings reveal that integrating human-annotated data with pseudo triplets generated by the Visual Delta Generator (VDG) significantly enhances the generalization capacity of CIR models. The VDG approach not only streamlines the generation of visual deltas but also emerges as a cost-efficient and effective alternative to extensive human annotation. This work paves the way for future research in semi-supervised learning and showcases VDG as a promising direction for advancing CIR systems." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 674, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 674, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 674, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgment: This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University))." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16812" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "text", + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "type": "text", + "content": "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, pages 104-120. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "text", + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "text", + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "type": "text", + "content": "[9] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In ECCV, pages 558-577. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "text", + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "type": "text", + "content": "[11] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "text", + "content": "[14] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, pages 14105-14115, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 186 + ], + "type": "text", + "content": "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "type": "text", + "content": "[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 266 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 267, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 267, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 545, + 300 + ], + "type": "text", + "content": "[18] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised semantic segmentation. NeurIPS, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 301, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 301, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 301, + 545, + 335 + ], + "type": "text", + "content": "[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 336, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 336, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 336, + 545, + 369 + ], + "type": "text", + "content": "[20] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, pages 3420-3429, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 370, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 370, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 370, + 545, + 414 + ], + "type": "text", + "content": "[21] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186, 2019. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "text", + "content": "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 483 + ], + "type": "text", + "content": "[23] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 518, + 545, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 518, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 308, + 518, + 545, + 550 + ], + "type": "text", + "content": "[25] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 552, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 552, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 552, + 545, + 575 + ], + "type": "text", + "content": "[26] Semi-Supervised Learning. Semi-supervised learning. CSZ2006. html, 2006." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 576, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 576, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 576, + 545, + 620 + ], + "type": "text", + "content": "[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 622, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 545, + 655 + ], + "type": "text", + "content": "[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and early fusion for composed image retrieval. arXiv preprint arXiv:2303.09429, 2023. 1, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "text", + "content": "[29] Junnan Li and et al. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3, 4, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "text", + "content": "[30] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16813" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 5, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "text", + "content": "[34] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[35] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[36] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In CVPR, pages 2125–2134, 2021. 1, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 376 + ], + "type": "text", + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 378, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 410 + ], + "type": "text", + "content": "[38] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 129:23-79, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 412, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 455 + ], + "type": "text", + "content": "[39] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 41(8):1979-1993, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 457, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 511 + ], + "type": "text", + "content": "[40] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, pages 6967-6977, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[42] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, pages 19305-19314, 2023. 1, 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 33:596-608, 2020. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 464 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[45] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[49] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 1, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[50] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[51] Haokun Wen, Xuemeng Song, Jianhua Yin, Jianlong Wu, Weili Guan, and Liqiang Nie. Self-training boosted multifaceted matching network for composed image retrieval. arXiv preprint arXiv:2305.09979, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[52] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[53] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16814" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_content_list.json b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02f799dcbbd1be4b06c53205e8afb3db758af770 --- /dev/null +++ b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_content_list.json @@ -0,0 +1,1472 @@ +[ + { + "type": "text", + "text": "Visual Fact Checker: Enabling High-Fidelity Detailed Caption Generation", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunhao Ge Xiaohui Zeng Jacob Samuel Huffman Tsung-Yi Lin Ming-Yu Liu Yin Cui", + "bbox": [ + 107, + 180, + 859, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVIDIA", + "bbox": [ + 446, + 215, + 521, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://research.nvidia.com/labs/dir/vfc/", + "bbox": [ + 276, + 234, + 681, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47a603a269fbdc63a9cc62eca601ffcd59f133a164af4a91cb0930c938d43cab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 299, + 101, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GPT-4V", + "text_level": 1, + "bbox": [ + 104, + 301, + 148, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The image shows a young girl in a field with her arms joyfully raised. She is holding a bunch of pink balloons in her right hand. She's wearing a plaid shirt and jeans, and her hair is styled in what appears to be a ponytail. The girl is smiling widely, and her pose suggests she is in mid-celebration or play. The field is filled with wildflowers, and trees can be seen in the background under a blue sky with some clouds. The setting appears to be a sunny day, perfect for outdoor activities. The image captures a moment of carefree childhood happiness.", + "bbox": [ + 81, + 316, + 336, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c92ae57f7240e51237a6b26fc46628e52ce7f94ace1fa4f0eb70378e13772fda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 285, + 596, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6a1c0f8db182ea61e87b8d41e4d73c375dbbf3563866da63822f81ddea9fffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 297, + 655, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VFC", + "text_level": 1, + "bbox": [ + 658, + 297, + 683, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A happy little girl is standing in a green field, wearing a plaid shirt and holding onto a string of pink balloons. The balloons are floating in the air, creating a playful and joyful atmosphere. The girl appears to be enjoying the moment, with a smile on her face. In the background, there is a vast expanse of green grass, stretching out as far as the eye can see. The sky above is a bright blue, with a few white clouds scattered about. The overall mood of the scene is one of carefree happiness and joy, with the pink balloons adding a touch of whimsy and playfulness to the atmosphere.", + "bbox": [ + 630, + 316, + 887, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1894e6b7bc4c70aa520e357cff3ab5b98f24aa9f56cc4b8c7b6150b3ff83db83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 502, + 107, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cap3D", + "text_level": 1, + "bbox": [ + 112, + 505, + 148, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A 3D model of a small wooden tower with a blue roof.", + "bbox": [ + 89, + 527, + 334, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a46745bcb26d4b301c683abf169c5bdf48e792161b55cd4098f42f309579053e.jpg", + "image_caption": [ + "Figure 1. Comparison of VisualFactChecker (VFC) with GPT-4V and Cap3D. VFC can generate high-fidelity detailed captions that closely match GPT-4V's quality for 2D images and offer significantly more details for 3D objects than Cap3D. VFC used a pre-trained Llama-2 as the LLM when generating the caption for the above 2D image." + ], + "image_footnote": [], + "bbox": [ + 385, + 474, + 475, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/339025ec5a80eafef85cc97d595a97ae86ba26380006fef6a4ff7f7d3fdba461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 476, + 612, + 603 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbbd59de05aace3209690e0310e496ba880ea8365c98b37a600cdd7f6542c01a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 477, + 660, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VFC", + "text_level": 1, + "bbox": [ + 661, + 479, + 686, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A 3D model of a three-level, intricately designed wooden tower with a contrasting blue roof and door, standing on a wooden base. The tower, brown in color, resembles a fusion of a house, a tower, and a castle. At the very top of the tower, there is a crescent moon design. The overall design adds a touch of fantasy to the scene.", + "bbox": [ + 635, + 505, + 883, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 672, + 313, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing automatic captioning methods for visual content face challenges such as lack of detail, content hallucination, and poor instruction following. In this work, we propose VisualFactChecker (VFC), a flexible training-free pipeline that generates high-fidelity and detailed captions for both 2D images and 3D objects. VFC consists of three steps: 1) proposal, where image-to-text captioning models propose multiple initial captions; 2) verification, where a large language model (LLM) utilizes tools such as object detection and VQA models to fact-check proposed captions; 3) captioning, where an LLM generates the final caption by summarizing caption proposals and the fact check verifica", + "bbox": [ + 73, + 719, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion results. In this step, VFC can flexibly generate captions in various styles following complex instructions. We conduct comprehensive captioning evaluations using four metrics: 1) CLIP-Score for image-text similarity; 2) CLIP-Image-Score for measuring the image-image similarity between the original and the reconstructed image generated by a text-to-image model using the caption. 3) human study on Amazon Mechanical Turk; 4) GPT-4V for fine-grained evaluation. Evaluation results show that VFC outperforms state-of-the-art open-sourced captioning methods for 2D images on the COCO dataset and 3D assets on the Objverse dataset. Our study demonstrates that by combining open-source models into a pipeline, we can attain captioning capability comparable to proprietary models such as GPT-4V, despite being over $10 \\times$ smaller in model size.", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14033", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image captioning is a pivotal challenge in computer vision and natural language processing. Its central goal is to encapsulate visual data within a textual description, which requires a nuanced understanding of both modalities. The recent advent of multimodal large language models (MM-LLMs), such as GPT-4V [26], and text-to-image generation models, such as DALLE-3 [3], has marked significant progress in this field. These proprietary models could leverage expansive human-labeled data and enormous computing resources to learn to generate detailed and contextually appropriate image descriptions. On the other hand, existing open-sourced captioning methods in the community still face significant challenges. Methods such as BLIP-2 [17] and OFA [35] often yield overly succinct captions that neglect essential visual information. Conversely, systems like Mini-GPT4 [39], InstructBLIP [8], and LLaVA [20, 21] can suffer from hallucination, producing long descriptions that do not align with the actual content of the images.", + "bbox": [ + 80, + 114, + 467, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this, we propose VisualFactChecker (VFC), a flexible training-free pipeline designed to produce accurate and comprehensive captions for both 2D images and 3D objects. Fig. 1 shows examples of captions generated by VFC and their comparisons with captions generated by GPT-4V [26] and Cap3D [23]. Captions generated by VFC are faithful textural representations of the visual contents. This can also be verified by reconstructing images and 3d objects from captions using state-of-the-art text-to-image and text-to-3d models, as shown in Fig. 2.", + "bbox": [ + 80, + 387, + 467, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VFC focuses on tackling hallucinations and insufficient details in generated captions and is structured around three core components: Proposer, serving as the system's \"eye\", creating detailed caption proposals as preliminary captions by using image-to-text captioning models; Large Language Model, acting as the \"brain\", calling and summarizing information from other components, and leveraging its advanced generalization capabilities to steer the captioning process following specified captioning instructions; Detector and VQA models, functioning as \"tools\" utilized by the LLM for fact-checking caption proposals, ensuring the fidelity of the final generated caption. VFC is versatile and effectively handles captioning for both 2D images and 3D objects through a unified pipeline. Fig. 3 shows an overview of the pipeline. The details of each component and their interplay are explained in Sec. 3.", + "bbox": [ + 80, + 537, + 467, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To comprehensively evaluate the generated captions, other than leveraging the commonly used CLIP-Score that primarily gauges the image-caption similarity, we propose a new metric: the CLIP-Image-Score. This metric assesses the similarity between the input image and a reconstructed image created by a text-to-image model from the caption, offering a complementary measure. Furthermore, we conducted a human study on Amazon Mechanical Turk for cap", + "bbox": [ + 80, + 780, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion evaluation. Finally, we also performed a fine-grained evaluation by asking GPT-4V to compare and judge captions with detailed reasoning. The combination of CLIP-Score, CLIP-Image-Score, GPT-4V, and human study provides a more robust evaluation of captions.", + "bbox": [ + 503, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows: 1) We propose VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity detailed 2D and 3D captions, effectively mitigating the challenge of hallucination in long captions. (2) CLIP-Image-Score: A novel caption evaluation metric that measures the similarity between the input image and a reconstructed image from the caption. (3) Our evaluation shows that VisualFactChecker achieves state-of-the-art results in 2D and 3D captioning tasks compared with open-sourced models. (4) Our work shows that using an LLM to chain open-source models can achieve captioning capability on par with proprietary models such as GPT-4V.", + "bbox": [ + 503, + 167, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 363, + 638, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Captioning", + "text_level": 1, + "bbox": [ + 503, + 388, + 671, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image captioning has made significant progress with the advent of deep learning. Pioneering works [2, 10, 14] primarily focus on integrating deep neural networks for enhanced image understanding and language generation.", + "bbox": [ + 503, + 412, + 890, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent strides have been made with the introduction of Multimodal-Large Language Models (MM-LLMs), which are trained on extensive vision and language data. The general approach involves leveraging a pre-trained large language model (LLM) and a vision encoder with a projector to align with the LLM's embeddings, thus enhancing visual understanding. Several models have emerged as significant contributors in this domain. BLIP [16], BLIP2 [17], OFA [35], Flamingo [1], Kosmos-2 [27], MiniGPT4 [39], InstructBLIP [8], LLaVA [20, 21] have demonstrated impressive performance in single-view image captioning tasks. However, they exhibit varying limitations. For instance, BLIP-2 and OFA often generate overly concise captions, while others, like InstructBLIP, can produce detailed captions that often include inaccurate or hallucinatory content. Our method aims to address these limitations by combining different models into a pipeline via an LLM, striking a better balance between accuracy and detailedness in generated captions while mitigating hallucinations.", + "bbox": [ + 503, + 473, + 890, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Large Language Models for Captioning", + "text_level": 1, + "bbox": [ + 503, + 771, + 838, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in large language models (LLMs) like GPT-3 [5], LAMDA [30], PALM [7], Llama [32], GPT-4 [26] have demonstrated exceptional zero-shot capabilities in language analysis and summarization tasks. This proficiency has naturally extended to the multimodal domain, particularly in image-language contexts, where LLMs can summarize multimodal information in a zero-shot manner.", + "bbox": [ + 503, + 794, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14034", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43520a7ed5c9835236b65f482fe3cf6d16c53e5fad72b91f335b052ef502c1d9.jpg", + "image_caption": [ + "Figure 2. We use DALLE-3 [3] as a text-to-image model to reconstruct 2D images using generated captions from different captioning methods (BLIP-2, LLaVA-1.5 and ours). Similarly, we use MVDream [29] as a text-to-3D model to reconstruct 3D objects using different 3D captions (generated by Cap3D [23] and ours). From the results, we can see that the reconstructed images or 3D objects using BLIP-2 or Cap3D captions are less similar than the input ones, suggesting their captions may not contain sufficient information or incorrectly describe the visual contents; the reconstructed images using LLaVA-1.5 captions contain objects or scenes that are not present in the original images (top: people in the background, bottom: pedestrians and cars on the street), suggesting there might be hallucinations in LLaVA-1.5 captions. Images or 3D objects reconstructed using our captions are more similar to the inputs." + ], + "image_footnote": [], + "bbox": [ + 76, + 90, + 558, + 304 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8c65815c524bd46c9b82e60c2f682e30b4bdda81322e45cba8da9fcae37f06a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 90, + 880, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-blind LLMs are prominent in multimodal applications, often utilizing language-only prefixes generated by pre-trained tools. Clipcap [25] demonstrates this by using a continuous embedding as a prompt for a GPT-style language model, achieving notable performance in single-viewpoint image captioning. Similarly, Promptcap [13] and PNP-VQA [31] leverage natural language prompts with GPT models to excel in visual question answering.", + "bbox": [ + 75, + 441, + 468, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent methods have employed LLMs to generate image captions by summarizing initial captions or keywords from Vision-Language models. For instance, Socratic models [37] use a CLIP-based model to extract key tags from images, followed by GPT-3 with specialized prompts to create stylized captions. ChatCaptioner [38] builds upon this by integrating ChatGPT and BLIP-2 [17] in a conversational approach for question-answering about the image, and summarizing them into a caption. Visual Clues [36] uses similar tags to generate a paragraph-caption. IC3 [6] and LLM-Fusion [4] use LLMs to summarize captions from existing models augmented with temperature-based sampling. Cap3D [24] extends this concept to 3D object.", + "bbox": [ + 75, + 566, + 470, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method differentiates itself in two critical ways: First, we focus on reducing hallucinations in captions by employing visual grounding tools, such as object detection, to fact-check captions for enhanced accuracy. Second, our pipeline can be used for captioning both 2D images and 3D objects. Unlike previous methods that rely on a single captioning model, we integrate multiple captioning sources from different models, ensuring a more comprehensive coverage of visual content to generate captions.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Hallucination in MM-LLM", + "text_level": 1, + "bbox": [ + 498, + 441, + 746, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are two popular topics on the hallucination of MM-LLMs. (1) Hallucination evaluation: Detection approaches such as Gunjal et al. [11] train classification models to identify hallucination. They focus on distinguishing between accurate and hallucinated content. Ground truth comparison methods [18, 34] compare model outputs with ground truth data to detect hallucinations. These techniques assess the alignment of generated captions with actual image content. (2) Mitigation Strategies [22]: Data optimization methods such as Liu et al. [19] address hallucination by creating negative instances in training datasets to reduce model overconfidence. Iterative generation methods such as Wang et al. [33] adopt an iterative process for caption generation, where brief answers are generated in succession and amalgamated, aiming to improve accuracy and relevance.", + "bbox": [ + 496, + 465, + 890, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our VisualFactChecker is a training-free pipeline mitigating hallucination in image captioning. Our method utilizes visual grounding tools for improved accuracy, thereby actively reducing the hallucination and offering high-fidelity captions for both 2D images and 3D objects.", + "bbox": [ + 496, + 693, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Visual Fact Checker", + "text_level": 1, + "bbox": [ + 500, + 784, + 694, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces the key components of VisualFactChecker as shown in Fig. 3 in detail and explains their interplay in generating accurate and detailed captions. The following sections delve into specifics. First, we detail the pipeline for 2D image captioning (Sec. 3.1), with Fig. 3 (top) illustrating this process. Then, we explore how our", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14035", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3b92976ffcac2b8757203defdb16f022ca02943023ae194a39cd44644195c39b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 867, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e25a6dd7bf2f14bcd4d8c693e023bfe146fcd49e9581ff2ea924bb7ee09c90b2.jpg", + "image_caption": [ + "Figure 3. Pipeline of the VisualFactChecker for captioning 2D images (top) and 3D objects (bottom). The process begins with the input being captioned by two multimodal captioning models (Captioner-1 and Captioner-2) to generate preliminary captions. These captions are then verified using a Large Language Model (LLM) to call object detection (Detector) and VQA models for fact-checking the captions. Finally, the LLM incorporates all the results and summarizes the final caption by following instructions." + ], + "image_footnote": [], + "bbox": [ + 83, + 255, + 864, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "approach is adapted for 3D object captioning as shown in Fig. 3 (bottom), underscoring both shared methodologies and unique aspects relevant to 3D contexts (Sec. 3.2).", + "bbox": [ + 75, + 502, + 468, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2D Image Captioning", + "text_level": 1, + "bbox": [ + 76, + 559, + 277, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The caption generation takes three steps: 1) proposal, 2) verification, and 3) captioning. Each step is detailed below.", + "bbox": [ + 75, + 582, + 468, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Proposal:** The Proposal step serves as the cornerstone of the captioning process that generates initial captions. This is achieved through the utilization of advanced image-to-text models, specifically \"LLaVA\" and \"Kosmos2\". These models are trained on expansive datasets, enabling them to comprehend and interpret visual content effectively. By analyzing the input image, they suggest various preliminary captions, each reflecting different facets and interpretations of the image (Fig. 3 top). The rationale behind using multiple image-to-text multimodal LLMs lies in the complexity of adequately capturing an image's essence in a single attempt. Since an image can be accurately described in numerous ways, different models bring unique perspectives, thereby encompassing a broader range of information present in the image. Although the initial captions proposed may not possess perfect fidelity, the primary objective at this stage is to generate captions that are as comprehensive as possible. Fig. 3 displays the specific prompts we used for each step, with more details in Appendix A.", + "bbox": [ + 75, + 613, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Verification and Captioning: The goal of the verification step is to scrutinize and rectify any inaccuracies or hallucinations in the captions during the proposal step. It employs a combination of a Large Language Model (LLM) and grounding tools, including an open-vocabulary object detection model and/or a visual question answering (VQA) model. Here the LLM can be GPT-4 or Llama2. As shown in Fig. 3 (top), the process involves the following steps: Step 1: LLM first summarizes the initial detailed descriptions from different MM-LLMs into a single, detailed caption. While this caption is comprehensive, it may not always be accurate. Step 2: The LLM then analyzes this synthesized caption, identifying all objects that could be verified by object detection and summarizing an object checklist. In 2D image captioning, the focus is on eliminating hallucinations, particularly descriptions of non-existent objects in the image. Identifying these objects is crucial for the subsequent verification process. Step 3: Taking the object checklist as input, an open-vocabulary object detection model examines candidate objects in the checklist and determines their presence in the image. This step is pivotal in validating the existence of objects mentioned in the caption, thus supporting the fidelity of the caption.", + "bbox": [ + 496, + 502, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After verification, we go to the last captioning step: Based on the object detection results, the LLM revises the summarized single detailed caption. Each object described", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14036", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the caption is cross-checked; if detected, it remains unchanged, while undetected objects are considered potential hallucinations and are removed from the caption. This step results in a final caption that is both detailed and reliable. The underlying assumption is that the detection model, serving as an object grounding expert, provides more reliable results than a general image descriptor.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the verification and captioning steps, the LLM plays a pivotal role as a \"brain\". It starts by parsing the initial caption and identifying key objects for detailed examination. The LLM then meticulously assesses whether each object mentioned actually appears in the image based on detection results. Following this thorough analysis, it refines and revises the initial captions, transforming them into final versions that are both coherent and richly detailed. The LLM is instrumental in guaranteeing linguistic fluency, ensuring that the captions not only accurately represent the image but also maintain the necessary level of detail for high-fidelity captioning. Moreover, the LLM can follow complex instructions to write the captions in a specified style, such as a caption that only mentions the foreground objects without mentioning the background. Fig. 3 displays the specific prompts used for each step.", + "bbox": [ + 75, + 198, + 473, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. 3D Object Captioning", + "text_level": 1, + "bbox": [ + 76, + 453, + 284, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 3D object captioning process follows a similar structural pipeline to that of 2D images, with a few key distinctions in certain steps, as depicted in Fig. 3 (bottom). In 3D captioning, an object may present multiple views, each offering unique information. The comprehensive caption for a 3D object is derived by integrating the perspectives from all these views. For each view, VisualFactChecker is employed to create a detailed, high-fidelity description. Subsequently, the LLM (GPT-4 or Llama-2) is used to amalgamate the information from all views, producing a unified caption for the 3D object. In particular, for each view's captioning, we have the same three-step approach akin to 2D image captioning. In the proposal step, LLaVA-1.5 and InstructBLIP are utilized for generating initial detailed descriptions. We opt out of using Kosmos2 for single 3D objects due to its less effective performance in providing detailed descriptions, possibly linked to its reliance on an implicit detection model. Additionally, a slightly modified prompt is used (see Fig. 3 bottom), which incorporates 3D-specific considerations. In the verification and captioning step, we primarily address hallucinations related to the attributes of 3D objects, such as shape and color. To mitigate these inaccuracies, rather than enumerating potential objects, we employ the LLM to generate five critical questions that could influence a text-to-3D generation model in reconstructing the 3D model. Following this, we utilize VQA models (specifically LLaVA-1.5) to respond to these questions based on the input 3D object view image. Subsequently, the LLM amends", + "bbox": [ + 75, + 478, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9ef51dbb1039ec8cab64168025cbc7ce14c2160531aef1c3d93f35d131460dfe.jpg", + "image_caption": [ + "Figure 4. The CLIP-Image-Score pipeline evaluates caption accuracy by encoding an original image $X$ into a feature representation $I_{X}$ using a CLIP image encoder. A captioning model generates a caption that is then input into a text-to-image model to reconstruct an image $X'$ , which is encoded to $I_{X'}$ . The score is computed by assessing the cosine similarity between $I_{X}$ and $I_{X'}$ , providing a measure of the caption's fidelity and hallucination detection." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 883, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the initial caption in accordance with the answers provided by the VQA model. We operate under the assumption that answering targeted questions results in fewer hallucinations compared to generating a general description. Once the caption for each individual view is complete, the LLM synthesizes these multiple perspectives into a singular, comprehensive caption for the entire 3D object. The prompts used for the LLM at each stage are detailed in Appendix A.", + "bbox": [ + 496, + 328, + 892, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. CLIP-Image-Score", + "text_level": 1, + "bbox": [ + 500, + 465, + 684, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accurate evaluation of caption correctness and detailedness is paramount in determining the performance of an image captioning model. Traditional metrics like the CLIP-Score [12] have served as a standard for measuring the alignment between generated captions and their corresponding images. However, our CLIP-score may lack the sensitivity needed to detect the specific issue of hallucination within captions.", + "bbox": [ + 496, + 492, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the CLIP-Image-Score, an alternative metric specifically developed to reflect the subtleties of caption quality. This metric is different from CLIP-Score by introducing an additional reconstruction step. Specifically, the CLIP-Image-score evaluates the similarity between the original image and a reconstructed version of the image generated by a fixed text-to-image model using the caption as a prompt. By comparing the raw image to its reconstructed image, the metric is able to detect discrepancies indicative of hallucination, thus providing a different perspective of the caption quality assessment. The underlying principle of the CLIP-Image-Score is the recognition that multiple \"correct\" captions may exist for a single image. However, it's only when a caption is both \"detail\" and \"correct\" that the reconstructed image closely resembles the original. Moreover, any hallucinations present in the caption become evident in the reconstructed image. Fig. 2 presents examples of such reconstructions. For instance, consider the results from LLaVA-1.5 shown in the third column. The cap", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14037", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion generated for the first image falsely mentions \"several other people in the background\". This error is clearly reflected in the image reconstructed by the text-to-image generator. In essence, comparing the two images indirectly ensures alignment between the image and its caption, thereby providing a complementary method to assess the quality of the caption than directly comparing the image and caption.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The CLIP-Image-Score evaluation process is depicted in the following steps:", + "bbox": [ + 76, + 196, + 468, + 226 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Caption Generation: An original image $X$ is input into a captioning model, which generates a caption.", + "- Caption-to-Image Reconstruction: This generated caption is then used as input for a text-to-image model, which creates a reconstructed image $X'$ that visually represents the textual description.", + "- Raw Image Encoding: The original image $X$ is processed through a CLIP image encoder, translating the visual content into an encoded representation $I_{X}$ .", + "- Reconstructed Image Encoding: The reconstructed image is also processed through the CLIP image encoder to obtain its encoded representation $I_{X'}$ .", + "- Score Calculation: Finally, the encoded representations of the original and reconstructed images are compared to calculate the CLIP-Image-Score. The score is given by the cosine similarity, which assesses the congruence between $I_{X}$ and $I_{X'}$ :" + ], + "bbox": [ + 76, + 227, + 468, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {C L I P - I m a g e - S c o r e} = \\frac {I _ {X} \\cdot I _ {X ^ {\\prime}}}{\\| I _ {X} \\| \\times \\| I _ {X ^ {\\prime}} \\|} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 489, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Most notably, CLIP-Image-Score offers a sensitive measure for detecting hallucinations. In scenarios where the generated caption includes elements that are not in the original image, the reconstructed image will also likely contain these discrepancies. By comparing the original and reconstructed images, the CLIP-Image-Score can effectively highlight these differences, offering a clearer insight into the fidelity and accuracy of the generated caption.", + "bbox": [ + 75, + 530, + 468, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, CLIP-Image-Score turns a cross-modality comparison into a more intuitive comparison in the same image modality (as shown in Fig. 4). CLIP-Image-Score represents a new complementary perspective for image captioning evaluation. By leveraging the capabilities of text-to-image models and focusing on the congruence between the original and reconstructed images, it provides an accurate assessment of caption quality, particularly in identifying and measuring hallucinations, thereby enhancing the overall reliability of caption generation systems.", + "bbox": [ + 75, + 651, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 815, + 207, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section presents a thorough evaluation of captioning models across both 2D and 3D visual content, employing a variety of datasets and methodologies. Table 1 provides a summary of our comprehensive evaluation experiments.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/25026c76475dbd41210b740ea177cf6c854a45799a93813d513cf4c86faa448c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EvalInput pairs for evaluationMethodReference
2DRaw imageCaptionCLIP-ScoreTable 2
Human evaluationFig. 6
CPT4V evaluationFig. 7
Raw imageImage(recon)CLIP-Image-ScoreTable 2
3DMulti-view (raw)CaptionCLIP-ScoreTable 3
GPT4V evaluationFig. 7
Multi-view (raw)Multi-view (recon)CLIP-Image-ScoreTable 3
", + "bbox": [ + 504, + 88, + 888, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Overall: CLIP-Score and CLIP-Image-Score", + "text_level": 1, + "bbox": [ + 498, + 227, + 879, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2D image captioning. Dataset: Our evaluation utilized 5,000 COCO test images from the Karpathy split. Baseline methods: We benchmarked against state-of-the-art captioning models, including BLIP-2 [17], InstructBLIP [8], and LLaVA-1.5 [20]. The evaluation focused on each model's ability to produce accurate, detailed, and coherent captions that effectively encapsulate the essence of the images. Evaluation Metric: We employed two metrics: CLIP-Score [12] and CLIP-Image-Score (Sec. 4). The CLIP-Score, a prevalent metric in image caption quality assessment, involves processing the raw image through the CLIP image encoder and the caption through the CLIP text encoder. The resultant embeddings are then compared for cosine similarity, with a higher score indicating greater semantic resemblance between the image and the caption. For our analysis, we first calculated the CLIP-Score for each image_caption pair, then averaged these scores across all 50,000 text/image pairs, scaling the result by a factor of 100. Table 2 displays the comparative performance of various image captioning methods on the 5,000 COCO test set images. The results demonstrate that our VisualFactChecker surpasses all baseline methods in performance.", + "bbox": [ + 496, + 251, + 890, + 583 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ce006f99f9e04518dd5a4ad7a63fcfedfe482c4e79fc49cd4fc7ba61ed6d2df.jpg", + "table_caption": [ + "Table 1. Summary of evaluation methods and results." + ], + "table_footnote": [], + "table_body": "
Captioning MethodCLIP-Score (%) ↑CLIP-Image-Score (%) ↑
Human Label (COCO GT)30.36 (-2.54)71.21 (-2.40)
BLIP230.11 (-2.79)70.79 (-2.82)
InstructBLIP31.45 (-1.45)72.95 (-0.66)
LLaVA-1.532.08 (-0.82)73.24 (-0.37)
Kosmos-232.32 (-0.58)73.28 (-0.33)
VisualFactChecker (Ours)32.9073.61
", + "bbox": [ + 503, + 595, + 888, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Image captioning comparison with different metrics on 5000 COCO test set in Karpathy split, we use raw image and caption as input pairs for evaluation.", + "bbox": [ + 498, + 690, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As outlined in Sec. 4, the CLIP-Image-Score provides a complementary view to assess the quality of image captions. This metric is derived by comparing the cosine similarity between the CLIP embeddings of two images: the original image and a reconstructed image, which is generated using the provided caption through a text-to-image generation model. A higher CLIP-Image-Score signifies a more accurate and effective image caption. For this process, Stable Diffusion XL (SDXL) [28] is utilized as the designated text-to-image model to reconstruct images based on", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14038", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a43fcd50ba640e0ae28cf138ad87081ab69bd4f561da931ac2e875dda75fd5f8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. 3D object captioning comparison with different metrics on 1000 objects in Objaverse. For CLIP-Score, we use the average score of two views for evaluation. For CLIP-Image-Score, we use an off-the-shelf text-to-3D model, MVDream, to generate 3D models from 3D captions. We compare two views of the raw object and the same views of generated 3D object for evaluation." + ], + "table_body": "
Captioning MethodCLIP-Score (%)↑CLIP-Image-Score (%)↑
Cap3D33.44 (-0.57)79.88 (-0.44)
VisualFactChecker (Ours)34.0180.32
", + "bbox": [ + 78, + 89, + 467, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the generated captions. Table 2 presents the CLIP-Image-Scores obtained for the 5000 images in the COCO test set, where our method outperforms all baseline methods.", + "bbox": [ + 75, + 250, + 468, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D object captioning. Dataset: 1,000 3D objects sampled from Objaverse dataset [9]. Baseline methods: We use state-of-the-art 3D object captioning model Cap3D [23] as the baseline. Cap3D uses 8 view images to generate the final object caption, our VisualFactChecker uses only 2 views to generate the object caption. Evaluation Metric: CLIP-Score and CLIP-Image-Score on multiple views rendered from 3D objects. To evaluate the similarity of a 3D object and the generated caption, we evaluate the similarity of the caption with the multi-view images used to generate the caption. Specifically, we evaluate the similarity of the generated caption with the two views that were used to generate the caption and use the average score to represent the CLIP-Score. Table. 3 shows the performance of 3D object captioning methods on 1,000 3D objects from Objaverse dataset. VisualFactChecker outperforms Cap3D.", + "bbox": [ + 75, + 294, + 468, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also use CLIP-Image-Score to evaluate the 3D caption quality. CLIP-Image-Score needs reconstructed images to compare with the raw images. We treat the two views that were used to generate the 3D object caption as the raw image. To obtain the reconstructed image, we use an off-the-shelf text-to-3D generation model, MVDream, to generate a 3D object given the generated 3D object caption. We then render the same two views of images based on the generated 3D object, and we calculate the CLIP-Image-Score between the raw image and the rendered image. Table. 3 shows the CLIP-Image-Score on 1000 objects in Objaverse dataset.", + "bbox": [ + 75, + 536, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Per Image Evaluation: Wining Rate", + "text_level": 1, + "bbox": [ + 76, + 710, + 390, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CLIP-Score and CLIP-Image-Score indicate an overall performance comparison, which shows an average score among all 5000 images. The average score may be dominated by a small group of images that have extremely high or low scores. To zoom in and show a more detailed comparison, we try to answer the following question: Given an image, what is the probability that one method performs better than another method on caption generation? To answer this question, we need to go over each image and calculate the winning rate for a pair of methods.", + "bbox": [ + 75, + 734, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for each image, we compare the CLIP-", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a124c0833c55a0b0af162c9a7fe2e88e4508b82c4185a3077a20e79a7382f589.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 89, + 885, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d81d81730484ac0fabc42c59d14555aca14fedf801aa2c86d30fcbc152f653c.jpg", + "image_caption": [ + "Figure 5. 2D image captioning comparison with pair-wise winning rate. VisualFactChecker (VFC) outperforms all baseline methods on both CLIP-Score (top) and CLIP-Image-Score (bottom)." + ], + "image_footnote": [], + "bbox": [ + 504, + 186, + 885, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Score of our VisualFactChecker caption against the captions generated from different baselines respectively, and calculate the wining probability of our method and the baselines. Fig. 5 shows the results, for example, we can see that in the pair-wise comparison, our VisualFactChecker performs better (higher CLIP-Score) than LLaVA-1.5 on $64.4\\%$ of 5000 images (3220 images).", + "bbox": [ + 496, + 354, + 890, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calculating the winning rate over all images provides a more detailed analysis that zooms in on the comparison of each image, which shows a complementary view than overall average CLIP-Score.", + "bbox": [ + 496, + 460, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Fine-grained Evaluation: Human and GPT-4V", + "text_level": 1, + "bbox": [ + 498, + 530, + 890, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The CLIP-Score and CLIP-Image-Score offer a general comparison of overall performance. A pairwise per-image winning rate provides a more specific analysis, evaluating performance on individual images. However, the research highlighted in related studies [15] indicates that the CLIP-Score may not be ideally suited for image-to-image comparison tasks. Furthermore, relying on a single score fails to provide a nuanced comparison across criteria, such as accuracy and level of detail. We use Human evaluation and GPT-4V to provide a more fine-grained evaluation.", + "bbox": [ + 496, + 551, + 890, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human evaluation using Amazon Mechanical Turk (AMT). We employed a pairwise comparison strategy. From the COCO dataset, we randomly selected 100 images out of 5000. For each image, our caption was compared against 5 baseline captions respectively. To reduce variance, each comparison was done by 3 different AMT workers and we used their majority voting as the final selection. This resulted in a total of 1500 comparisons collected on AMT. AMT UI is shown in the appendix. The workers were presented with two competing captions — one from a baseline method and one from our VisualFactChecker, in randomized order. They were instructed to select the better caption describing the image based on 3 aspects: correctness, de", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14039", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tailness, and fluency. Results in Fig. 6 show our captions are more preferred by humans. The human evaluation instruction and web UI is shown in Appendix B.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4ad4af8a681f8313aac73b805ea572e75585a1e3de13cafb537c815675f431f.jpg", + "image_caption": [ + "Figure 6. Amazon Mechanical Turk human evaluation results." + ], + "image_footnote": [], + "bbox": [ + 81, + 150, + 464, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPT-4V evaluation. Our study applied GPT-4V for evaluating captions in a manner akin to the caption evaluation process used in DALLE-3. We use the same randomly selected 100 images from COCO as in Human evaluation. For each image, we considered the captions generated by 5 baseline methods alongside the caption produced by our VisualFactChecker. We then presented GPT-4V with the raw image, our reference caption, and the four baseline captions. Our designed prompt instructed GPT-4V to compare each baseline caption against our reference caption, focusing on two primary aspects: correctness and detail. GPT-4V was tasked with providing a pairwise, detailed comparison for each pair, including justifications for its assessments. Based on these comparative insights, GPT-4V classified each baseline method caption as either \"better\" or \"worse\" than our VisualFactChecker. Fig. 5 shows the comprehensive results. More details about the GPT-4V evaluation prompt and examples are shown in Appendix B.", + "bbox": [ + 76, + 282, + 472, + 556 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a7b94bc4a4cfa9b509422ca41fabccc66307afa9d8ab8bc6273f3f7be418a74.jpg", + "image_caption": [ + "Figure 7. GPT-4V evaluation results. Our captions are significantly better than baselines." + ], + "image_footnote": [], + "bbox": [ + 81, + 568, + 464, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 726, + 230, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our ablation study, we explore the impact of various components on performance. For 2D captioning tasks, we assess the efficacy of initial captioning models, LLaVA-1.5 and Kosmos-2, using the CLIP-Score metric for the captions they generate on the same 5000 COCO test images. Additionally, we ablate our method's performance in the absence of the verification (fact checker) step, which aims to mitigate hallucinations through detection grounding. Table 4 shows the detailed results. Likewise, in the context of 3D object captioning, we evaluate the individual contribu", + "bbox": [ + 76, + 750, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/17ba8fc4ca5ec2978648e38d168aeea0de131f09f2c14ba5ce22b7dbbf234bcf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods or StepsCLIP-Score
2DLLaVA-1.532.08 (-0.33)
Kosmos-232.32 (-0.09)
VisualFactChecker (w/o fact check)32.41
VisualFactChecker32.90 (+0.49)
3DLLaVA-1.532.05 (-0.66)
InstructBLIP32.51 (-0.20)
VisualFactChecker (w/o fact check)32.71
VisualFactChecker34.01 (+1.30)
", + "bbox": [ + 504, + 88, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on captioning 2D images (5000 COCO test dataset) and 3D objects (1000 Objaverse).", + "bbox": [ + 498, + 238, + 890, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions of initial captioners, namely LLaVA-1.5 and InstructBLIP on the same 1000 Objaverse 3D objects. We further investigate the performance of our methodology without the fact checker, which in this case operates by leveraging a VQA model to reduce hallucinations. Table 4 shows the detailed results. These results highlight the significance of fact checker in our approach.", + "bbox": [ + 496, + 292, + 890, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Qualitative Results and Prompt Following", + "text_level": 1, + "bbox": [ + 498, + 409, + 857, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other than quantitative evaluation results, we show more qualitative examples of VisualFactChecker for 2D and 3D captions in Appendix C.", + "bbox": [ + 496, + 431, + 890, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By leveraging an LLM, VisualFactChecker can follow complex instructions to write captions in various styles. Examples are shown in Appendix D.", + "bbox": [ + 496, + 477, + 890, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 536, + 617, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity and detailed captions. By utilizing an LLM to chain multimodal models and object detection and VQA models, VFC reduces hallucination in long captions. We conducted a comprehensive caption evaluation using different metrics, including 1) image-text similarity using CLIP-Score, 2) image-reconstructed image similarity using our proposed CLIP-Image-Score, 3) human study, and 4) fine-grained evaluation using GPT-4V. Compared with open-sourced captioning models, our method achieves state-of-the-art in both 2D and 3D captioning. Our work shows combining open-sourced models into a pipeline can significantly close the captioning performance gap with proprietary models like GPT-4V. In the future, we plan to improve our pipeline further by including more components for fact-checking and making it more automatic in deciding which components to use.", + "bbox": [ + 496, + 561, + 890, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments We would like to thank Siddharth Gururani for helping with our human evaluation using Amazon Mechanical Turk; Haochen Wang for his help in preprocessing 3D data. We also thank Qinsheng Zhang, Yogesh Balaji, and Yen-Chen Lin for their helpful discussion.", + "bbox": [ + 496, + 818, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14040", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2", + "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, 2018. 2", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. Technical report, OpenAI, 2023. 2, 3, 11", + "[4] Simone Bianco, Luigi Celona, Marco Donzella, and Paolo Napoletano. Improving image captioning descriptiveness by ranking and llm-based fusion. arXiv preprint arXiv:2306.11593, 2023. 3", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2", + "[6] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. Ic3: Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 3", + "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2", + "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 6", + "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 7", + "[10] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2", + "[11] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 3", + "[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6", + "[13] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In ICCV, 2019. 2", + "[15] Klemen Kotar, Stephen Tian, Hong-Xing Yu, Daniel LK Yamins, and Jiajun Wu. Are these the same apple? comparing images based on object intrinsics. arXiv preprint arXiv:2311.00750, 2023. 7", + "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2", + "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 6", + "[18] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3", + "[19] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3", + "[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 6", + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2", + "[22] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3", + "[23] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. In NeurIPS, 2023. 2, 3, 7, 17", + "[24] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3", + "[25] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3", + "[26] OpenAI. Gpt-4 technical report, 2023. 2", + "[27] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2", + "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6", + "[29] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 17", + "[30] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2", + "[31] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven CH Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773, 2022. 3", + "[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[33] Bin Wang, Fan Wu, Xiao Han, Jiahui Peng, Huaping Zhong, Pan Zhang, Xiaoyi Dong, Weijia Li, Wei Li, Jiaqi Wang, et al. Vigo: Visual instruction generation and correction. arXiv preprint arXiv:2308.12714, 2023. 3", + "[34] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 3", + "[35] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 2", + "[36] Yujia Xie, Luowei Zhou, Xiyang Dai, Lu Yuan, Nguyen Bach, Ce Liu, and Michael Zeng. Visual clues: Bridging vision and language foundations for image paragraph captioning. In NeurIPS, 2022. 3", + "[37] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3", + "[38] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 3", + "[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 712 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14042", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_model.json b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d709400408e69ce3c26cd4151a67e8f46f969510 --- /dev/null +++ b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Visual Fact Checker: Enabling High-Fidelity Detailed Caption Generation" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.181, + 0.861, + 0.201 + ], + "angle": 0, + "content": "Yunhao Ge Xiaohui Zeng Jacob Samuel Huffman Tsung-Yi Lin Ming-Yu Liu Yin Cui" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.217, + 0.522, + 0.232 + ], + "angle": 0, + "content": "NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.236, + 0.682, + 0.252 + ], + "angle": 0, + "content": "https://research.nvidia.com/labs/dir/vfc/" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.3, + 0.102, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.303, + 0.15, + 0.314 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.318, + 0.337, + 0.429 + ], + "angle": 0, + "content": "The image shows a young girl in a field with her arms joyfully raised. She is holding a bunch of pink balloons in her right hand. She's wearing a plaid shirt and jeans, and her hair is styled in what appears to be a ponytail. The girl is smiling widely, and her pose suggests she is in mid-celebration or play. The field is filled with wildflowers, and trees can be seen in the background under a blue sky with some clouds. The setting appears to be a sunny day, perfect for outdoor activities. The image captures a moment of carefree childhood happiness." + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.286, + 0.597, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.298, + 0.656, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.66, + 0.299, + 0.684, + 0.31 + ], + "angle": 0, + "content": "VFC" + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.318, + 0.888, + 0.437 + ], + "angle": 0, + "content": "A happy little girl is standing in a green field, wearing a plaid shirt and holding onto a string of pink balloons. The balloons are floating in the air, creating a playful and joyful atmosphere. The girl appears to be enjoying the moment, with a smile on her face. In the background, there is a vast expanse of green grass, stretching out as far as the eye can see. The sky above is a bright blue, with a few white clouds scattered about. The overall mood of the scene is one of carefree happiness and joy, with the pink balloons adding a touch of whimsy and playfulness to the atmosphere." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.503, + 0.108, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.506, + 0.149, + 0.517 + ], + "angle": 0, + "content": "Cap3D" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.336, + 0.538 + ], + "angle": 0, + "content": "A 3D model of a small wooden tower with a blue roof." + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.475, + 0.477, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.477, + 0.614, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.478, + 0.661, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.663, + 0.48, + 0.687, + 0.49 + ], + "angle": 0, + "content": "VFC" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.506, + 0.884, + 0.575 + ], + "angle": 0, + "content": "A 3D model of a three-level, intricately designed wooden tower with a contrasting blue roof and door, standing on a wooden base. The tower, brown in color, resembles a fusion of a house, a tower, and a castle. At the very top of the tower, there is a crescent moon design. The overall design adds a touch of fantasy to the scene." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.619, + 0.894, + 0.661 + ], + "angle": 0, + "content": "Figure 1. Comparison of VisualFactChecker (VFC) with GPT-4V and Cap3D. VFC can generate high-fidelity detailed captions that closely match GPT-4V's quality for 2D images and offer significantly more details for 3D objects than Cap3D. VFC used a pre-trained Llama-2 as the LLM when generating the caption for the above 2D image." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.674, + 0.314, + 0.691 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Existing automatic captioning methods for visual content face challenges such as lack of detail, content hallucination, and poor instruction following. In this work, we propose VisualFactChecker (VFC), a flexible training-free pipeline that generates high-fidelity and detailed captions for both 2D images and 3D objects. VFC consists of three steps: 1) proposal, where image-to-text captioning models propose multiple initial captions; 2) verification, where a large language model (LLM) utilizes tools such as object detection and VQA models to fact-check proposed captions; 3) captioning, where an LLM generates the final caption by summarizing caption proposals and the fact check verifica" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "tion results. In this step, VFC can flexibly generate captions in various styles following complex instructions. We conduct comprehensive captioning evaluations using four metrics: 1) CLIP-Score for image-text similarity; 2) CLIP-Image-Score for measuring the image-image similarity between the original and the reconstructed image generated by a text-to-image model using the caption. 3) human study on Amazon Mechanical Turk; 4) GPT-4V for fine-grained evaluation. Evaluation results show that VFC outperforms state-of-the-art open-sourced captioning methods for 2D images on the COCO dataset and 3D assets on the Objverse dataset. Our study demonstrates that by combining open-source models into a pipeline, we can attain captioning capability comparable to proprietary models such as GPT-4V, despite being over \\(10 \\times\\) smaller in model size." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14033" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.115, + 0.468, + 0.386 + ], + "angle": 0, + "content": "Image captioning is a pivotal challenge in computer vision and natural language processing. Its central goal is to encapsulate visual data within a textual description, which requires a nuanced understanding of both modalities. The recent advent of multimodal large language models (MM-LLMs), such as GPT-4V [26], and text-to-image generation models, such as DALLE-3 [3], has marked significant progress in this field. These proprietary models could leverage expansive human-labeled data and enormous computing resources to learn to generate detailed and contextually appropriate image descriptions. On the other hand, existing open-sourced captioning methods in the community still face significant challenges. Methods such as BLIP-2 [17] and OFA [35] often yield overly succinct captions that neglect essential visual information. Conversely, systems like Mini-GPT4 [39], InstructBLIP [8], and LLaVA [20, 21] can suffer from hallucination, producing long descriptions that do not align with the actual content of the images." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.388, + 0.468, + 0.537 + ], + "angle": 0, + "content": "In light of this, we propose VisualFactChecker (VFC), a flexible training-free pipeline designed to produce accurate and comprehensive captions for both 2D images and 3D objects. Fig. 1 shows examples of captions generated by VFC and their comparisons with captions generated by GPT-4V [26] and Cap3D [23]. Captions generated by VFC are faithful textural representations of the visual contents. This can also be verified by reconstructing images and 3d objects from captions using state-of-the-art text-to-image and text-to-3d models, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.539, + 0.468, + 0.779 + ], + "angle": 0, + "content": "VFC focuses on tackling hallucinations and insufficient details in generated captions and is structured around three core components: Proposer, serving as the system's \"eye\", creating detailed caption proposals as preliminary captions by using image-to-text captioning models; Large Language Model, acting as the \"brain\", calling and summarizing information from other components, and leveraging its advanced generalization capabilities to steer the captioning process following specified captioning instructions; Detector and VQA models, functioning as \"tools\" utilized by the LLM for fact-checking caption proposals, ensuring the fidelity of the final generated caption. VFC is versatile and effectively handles captioning for both 2D images and 3D objects through a unified pipeline. Fig. 3 shows an overview of the pipeline. The details of each component and their interplay are explained in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.781, + 0.468, + 0.9 + ], + "angle": 0, + "content": "To comprehensively evaluate the generated captions, other than leveraging the commonly used CLIP-Score that primarily gauges the image-caption similarity, we propose a new metric: the CLIP-Image-Score. This metric assesses the similarity between the input image and a reconstructed image created by a text-to-image model from the caption, offering a complementary measure. Furthermore, we conducted a human study on Amazon Mechanical Turk for cap" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.166 + ], + "angle": 0, + "content": "tion evaluation. Finally, we also performed a fine-grained evaluation by asking GPT-4V to compare and judge captions with detailed reasoning. The combination of CLIP-Score, CLIP-Image-Score, GPT-4V, and human study provides a more robust evaluation of captions." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.168, + 0.891, + 0.348 + ], + "angle": 0, + "content": "We summarize our main contributions as follows: 1) We propose VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity detailed 2D and 3D captions, effectively mitigating the challenge of hallucination in long captions. (2) CLIP-Image-Score: A novel caption evaluation metric that measures the similarity between the input image and a reconstructed image from the caption. (3) Our evaluation shows that VisualFactChecker achieves state-of-the-art results in 2D and 3D captioning tasks compared with open-sourced models. (4) Our work shows that using an LLM to chain open-source models can achieve captioning capability on par with proprietary models such as GPT-4V." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.364, + 0.64, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.389, + 0.673, + 0.405 + ], + "angle": 0, + "content": "2.1. Image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.413, + 0.891, + 0.473 + ], + "angle": 0, + "content": "Image captioning has made significant progress with the advent of deep learning. Pioneering works [2, 10, 14] primarily focus on integrating deep neural networks for enhanced image understanding and language generation." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.474, + 0.891, + 0.76 + ], + "angle": 0, + "content": "Recent strides have been made with the introduction of Multimodal-Large Language Models (MM-LLMs), which are trained on extensive vision and language data. The general approach involves leveraging a pre-trained large language model (LLM) and a vision encoder with a projector to align with the LLM's embeddings, thus enhancing visual understanding. Several models have emerged as significant contributors in this domain. BLIP [16], BLIP2 [17], OFA [35], Flamingo [1], Kosmos-2 [27], MiniGPT4 [39], InstructBLIP [8], LLaVA [20, 21] have demonstrated impressive performance in single-view image captioning tasks. However, they exhibit varying limitations. For instance, BLIP-2 and OFA often generate overly concise captions, while others, like InstructBLIP, can produce detailed captions that often include inaccurate or hallucinatory content. Our method aims to address these limitations by combining different models into a pipeline via an LLM, striking a better balance between accuracy and detailedness in generated captions while mitigating hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.772, + 0.839, + 0.788 + ], + "angle": 0, + "content": "2.2. Large Language Models for Captioning" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.795, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Recent advancements in large language models (LLMs) like GPT-3 [5], LAMDA [30], PALM [7], Llama [32], GPT-4 [26] have demonstrated exceptional zero-shot capabilities in language analysis and summarization tasks. This proficiency has naturally extended to the multimodal domain, particularly in image-language contexts, where LLMs can summarize multimodal information in a zero-shot manner." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14034" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.092, + 0.56, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.092, + 0.882, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.896, + 0.419 + ], + "angle": 0, + "content": "Figure 2. We use DALLE-3 [3] as a text-to-image model to reconstruct 2D images using generated captions from different captioning methods (BLIP-2, LLaVA-1.5 and ours). Similarly, we use MVDream [29] as a text-to-3D model to reconstruct 3D objects using different 3D captions (generated by Cap3D [23] and ours). From the results, we can see that the reconstructed images or 3D objects using BLIP-2 or Cap3D captions are less similar than the input ones, suggesting their captions may not contain sufficient information or incorrectly describe the visual contents; the reconstructed images using LLaVA-1.5 captions contain objects or scenes that are not present in the original images (top: people in the background, bottom: pedestrians and cars on the street), suggesting there might be hallucinations in LLaVA-1.5 captions. Images or 3D objects reconstructed using our captions are more similar to the inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Vision-blind LLMs are prominent in multimodal applications, often utilizing language-only prefixes generated by pre-trained tools. Clipcap [25] demonstrates this by using a continuous embedding as a prompt for a GPT-style language model, achieving notable performance in single-viewpoint image captioning. Similarly, Promptcap [13] and PNP-VQA [31] leverage natural language prompts with GPT models to excel in visual question answering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.764 + ], + "angle": 0, + "content": "Recent methods have employed LLMs to generate image captions by summarizing initial captions or keywords from Vision-Language models. For instance, Socratic models [37] use a CLIP-based model to extract key tags from images, followed by GPT-3 with specialized prompts to create stylized captions. ChatCaptioner [38] builds upon this by integrating ChatGPT and BLIP-2 [17] in a conversational approach for question-answering about the image, and summarizing them into a caption. Visual Clues [36] uses similar tags to generate a paragraph-caption. IC3 [6] and LLM-Fusion [4] use LLMs to summarize captions from existing models augmented with temperature-based sampling. Cap3D [24] extends this concept to 3D object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our method differentiates itself in two critical ways: First, we focus on reducing hallucinations in captions by employing visual grounding tools, such as object detection, to fact-check captions for enhanced accuracy. Second, our pipeline can be used for captioning both 2D images and 3D objects. Unlike previous methods that rely on a single captioning model, we integrate multiple captioning sources from different models, ensuring a more comprehensive coverage of visual content to generate captions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.443, + 0.747, + 0.457 + ], + "angle": 0, + "content": "2.3. Hallucination in MM-LLM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.693 + ], + "angle": 0, + "content": "There are two popular topics on the hallucination of MM-LLMs. (1) Hallucination evaluation: Detection approaches such as Gunjal et al. [11] train classification models to identify hallucination. They focus on distinguishing between accurate and hallucinated content. Ground truth comparison methods [18, 34] compare model outputs with ground truth data to detect hallucinations. These techniques assess the alignment of generated captions with actual image content. (2) Mitigation Strategies [22]: Data optimization methods such as Liu et al. [19] address hallucination by creating negative instances in training datasets to reduce model overconfidence. Iterative generation methods such as Wang et al. [33] adopt an iterative process for caption generation, where brief answers are generated in succession and amalgamated, aiming to improve accuracy and relevance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Our VisualFactChecker is a training-free pipeline mitigating hallucination in image captioning. Our method utilizes visual grounding tools for improved accuracy, thereby actively reducing the hallucination and offering high-fidelity captions for both 2D images and 3D objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.695, + 0.8 + ], + "angle": 0, + "content": "3. Visual Fact Checker" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This section introduces the key components of VisualFactChecker as shown in Fig. 3 in detail and explains their interplay in generating accurate and detailed captions. The following sections delve into specifics. First, we detail the pipeline for 2D image captioning (Sec. 3.1), with Fig. 3 (top) illustrating this process. Then, we explore how our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14035" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.868, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.256, + 0.866, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 3. Pipeline of the VisualFactChecker for captioning 2D images (top) and 3D objects (bottom). The process begins with the input being captioned by two multimodal captioning models (Captioner-1 and Captioner-2) to generate preliminary captions. These captions are then verified using a Large Language Model (LLM) to call object detection (Detector) and VQA models for fact-checking the captions. Finally, the LLM incorporates all the results and summarizes the final caption by following instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.549 + ], + "angle": 0, + "content": "approach is adapted for 3D object captioning as shown in Fig. 3 (bottom), underscoring both shared methodologies and unique aspects relevant to 3D contexts (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.279, + 0.575 + ], + "angle": 0, + "content": "3.1.2D Image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.614 + ], + "angle": 0, + "content": "The caption generation takes three steps: 1) proposal, 2) verification, and 3) captioning. Each step is detailed below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.901 + ], + "angle": 0, + "content": "**Proposal:** The Proposal step serves as the cornerstone of the captioning process that generates initial captions. This is achieved through the utilization of advanced image-to-text models, specifically \"LLaVA\" and \"Kosmos2\". These models are trained on expansive datasets, enabling them to comprehend and interpret visual content effectively. By analyzing the input image, they suggest various preliminary captions, each reflecting different facets and interpretations of the image (Fig. 3 top). The rationale behind using multiple image-to-text multimodal LLMs lies in the complexity of adequately capturing an image's essence in a single attempt. Since an image can be accurately described in numerous ways, different models bring unique perspectives, thereby encompassing a broader range of information present in the image. Although the initial captions proposed may not possess perfect fidelity, the primary objective at this stage is to generate captions that are as comprehensive as possible. Fig. 3 displays the specific prompts we used for each step, with more details in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Verification and Captioning: The goal of the verification step is to scrutinize and rectify any inaccuracies or hallucinations in the captions during the proposal step. It employs a combination of a Large Language Model (LLM) and grounding tools, including an open-vocabulary object detection model and/or a visual question answering (VQA) model. Here the LLM can be GPT-4 or Llama2. As shown in Fig. 3 (top), the process involves the following steps: Step 1: LLM first summarizes the initial detailed descriptions from different MM-LLMs into a single, detailed caption. While this caption is comprehensive, it may not always be accurate. Step 2: The LLM then analyzes this synthesized caption, identifying all objects that could be verified by object detection and summarizing an object checklist. In 2D image captioning, the focus is on eliminating hallucinations, particularly descriptions of non-existent objects in the image. Identifying these objects is crucial for the subsequent verification process. Step 3: Taking the object checklist as input, an open-vocabulary object detection model examines candidate objects in the checklist and determines their presence in the image. This step is pivotal in validating the existence of objects mentioned in the caption, thus supporting the fidelity of the caption." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "After verification, we go to the last captioning step: Based on the object detection results, the LLM revises the summarized single detailed caption. Each object described" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14036" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "in the caption is cross-checked; if detected, it remains unchanged, while undetected objects are considered potential hallucinations and are removed from the caption. This step results in a final caption that is both detailed and reliable. The underlying assumption is that the detection model, serving as an object grounding expert, provides more reliable results than a general image descriptor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.474, + 0.442 + ], + "angle": 0, + "content": "In the verification and captioning steps, the LLM plays a pivotal role as a \"brain\". It starts by parsing the initial caption and identifying key objects for detailed examination. The LLM then meticulously assesses whether each object mentioned actually appears in the image based on detection results. Following this thorough analysis, it refines and revises the initial captions, transforming them into final versions that are both coherent and richly detailed. The LLM is instrumental in guaranteeing linguistic fluency, ensuring that the captions not only accurately represent the image but also maintain the necessary level of detail for high-fidelity captioning. Moreover, the LLM can follow complex instructions to write the captions in a specified style, such as a caption that only mentions the foreground objects without mentioning the background. Fig. 3 displays the specific prompts used for each step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.285, + 0.47 + ], + "angle": 0, + "content": "3.2. 3D Object Captioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The 3D object captioning process follows a similar structural pipeline to that of 2D images, with a few key distinctions in certain steps, as depicted in Fig. 3 (bottom). In 3D captioning, an object may present multiple views, each offering unique information. The comprehensive caption for a 3D object is derived by integrating the perspectives from all these views. For each view, VisualFactChecker is employed to create a detailed, high-fidelity description. Subsequently, the LLM (GPT-4 or Llama-2) is used to amalgamate the information from all views, producing a unified caption for the 3D object. In particular, for each view's captioning, we have the same three-step approach akin to 2D image captioning. In the proposal step, LLaVA-1.5 and InstructBLIP are utilized for generating initial detailed descriptions. We opt out of using Kosmos2 for single 3D objects due to its less effective performance in providing detailed descriptions, possibly linked to its reliance on an implicit detection model. Additionally, a slightly modified prompt is used (see Fig. 3 bottom), which incorporates 3D-specific considerations. In the verification and captioning step, we primarily address hallucinations related to the attributes of 3D objects, such as shape and color. To mitigate these inaccuracies, rather than enumerating potential objects, we employ the LLM to generate five critical questions that could influence a text-to-3D generation model in reconstructing the 3D model. Following this, we utilize VQA models (specifically LLaVA-1.5) to respond to these questions based on the input 3D object view image. Subsequently, the LLM amends" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.885, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.203, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 4. The CLIP-Image-Score pipeline evaluates caption accuracy by encoding an original image \\( X \\) into a feature representation \\( I_{X} \\) using a CLIP image encoder. A captioning model generates a caption that is then input into a text-to-image model to reconstruct an image \\( X' \\), which is encoded to \\( I_{X'} \\). The score is computed by assessing the cosine similarity between \\( I_{X} \\) and \\( I_{X'} \\), providing a measure of the caption's fidelity and hallucination detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.893, + 0.452 + ], + "angle": 0, + "content": "the initial caption in accordance with the answers provided by the VQA model. We operate under the assumption that answering targeted questions results in fewer hallucinations compared to generating a general description. Once the caption for each individual view is complete, the LLM synthesizes these multiple perspectives into a singular, comprehensive caption for the entire 3D object. The prompts used for the LLM at each stage are detailed in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.685, + 0.484 + ], + "angle": 0, + "content": "4. CLIP-Image-Score" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Accurate evaluation of caption correctness and detailedness is paramount in determining the performance of an image captioning model. Traditional metrics like the CLIP-Score [12] have served as a standard for measuring the alignment between generated captions and their corresponding images. However, our CLIP-score may lack the sensitivity needed to detect the specific issue of hallucination within captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We present the CLIP-Image-Score, an alternative metric specifically developed to reflect the subtleties of caption quality. This metric is different from CLIP-Score by introducing an additional reconstruction step. Specifically, the CLIP-Image-score evaluates the similarity between the original image and a reconstructed version of the image generated by a fixed text-to-image model using the caption as a prompt. By comparing the raw image to its reconstructed image, the metric is able to detect discrepancies indicative of hallucination, thus providing a different perspective of the caption quality assessment. The underlying principle of the CLIP-Image-Score is the recognition that multiple \"correct\" captions may exist for a single image. However, it's only when a caption is both \"detail\" and \"correct\" that the reconstructed image closely resembles the original. Moreover, any hallucinations present in the caption become evident in the reconstructed image. Fig. 2 presents examples of such reconstructions. For instance, consider the results from LLaVA-1.5 shown in the third column. The cap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14037" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "tion generated for the first image falsely mentions \"several other people in the background\". This error is clearly reflected in the image reconstructed by the text-to-image generator. In essence, comparing the two images indirectly ensures alignment between the image and its caption, thereby providing a complementary method to assess the quality of the caption than directly comparing the image and caption." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.198, + 0.469, + 0.227 + ], + "angle": 0, + "content": "The CLIP-Image-Score evaluation process is depicted in the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.228, + 0.468, + 0.257 + ], + "angle": 0, + "content": "- Caption Generation: An original image \\( X \\) is input into a captioning model, which generates a caption." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.258, + 0.469, + 0.317 + ], + "angle": 0, + "content": "- Caption-to-Image Reconstruction: This generated caption is then used as input for a text-to-image model, which creates a reconstructed image \\( X' \\) that visually represents the textual description." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.318, + 0.468, + 0.362 + ], + "angle": 0, + "content": "- Raw Image Encoding: The original image \\( X \\) is processed through a CLIP image encoder, translating the visual content into an encoded representation \\( I_{X} \\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.364, + 0.468, + 0.408 + ], + "angle": 0, + "content": "- Reconstructed Image Encoding: The reconstructed image is also processed through the CLIP image encoder to obtain its encoded representation \\( I_{X'} \\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.409, + 0.468, + 0.483 + ], + "angle": 0, + "content": "- Score Calculation: Finally, the encoded representations of the original and reconstructed images are compared to calculate the CLIP-Image-Score. The score is given by the cosine similarity, which assesses the congruence between \\( I_{X} \\) and \\( I_{X'} \\):" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.228, + 0.469, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.491, + 0.469, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\text {C L I P - I m a g e - S c o r e} = \\frac {I _ {X} \\cdot I _ {X ^ {\\prime}}}{\\| I _ {X} \\| \\times \\| I _ {X ^ {\\prime}} \\|} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.469, + 0.652 + ], + "angle": 0, + "content": "Most notably, CLIP-Image-Score offers a sensitive measure for detecting hallucinations. In scenarios where the generated caption includes elements that are not in the original image, the reconstructed image will also likely contain these discrepancies. By comparing the original and reconstructed images, the CLIP-Image-Score can effectively highlight these differences, offering a clearer insight into the fidelity and accuracy of the generated caption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.469, + 0.803 + ], + "angle": 0, + "content": "Furthermore, CLIP-Image-Score turns a cross-modality comparison into a more intuitive comparison in the same image modality (as shown in Fig. 4). CLIP-Image-Score represents a new complementary perspective for image captioning evaluation. By leveraging the capabilities of text-to-image models and focusing on the congruence between the original and reconstructed images, it provides an accurate assessment of caption quality, particularly in identifying and measuring hallucinations, thereby enhancing the overall reliability of caption generation systems." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.209, + 0.832 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This section presents a thorough evaluation of captioning models across both 2D and 3D visual content, employing a variety of datasets and methodologies. Table 1 provides a summary of our comprehensive evaluation experiments." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.178 + ], + "angle": 0, + "content": "
EvalInput pairs for evaluationMethodReference
2DRaw imageCaptionCLIP-ScoreTable 2
Human evaluationFig. 6
CPT4V evaluationFig. 7
Raw imageImage(recon)CLIP-Image-ScoreTable 2
3DMulti-view (raw)CaptionCLIP-ScoreTable 3
GPT4V evaluationFig. 7
Multi-view (raw)Multi-view (recon)CLIP-Image-ScoreTable 3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.188, + 0.854, + 0.202 + ], + "angle": 0, + "content": "Table 1. Summary of evaluation methods and results." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.228, + 0.88, + 0.244 + ], + "angle": 0, + "content": "5.1. Overall: CLIP-Score and CLIP-Image-Score" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.892, + 0.584 + ], + "angle": 0, + "content": "2D image captioning. Dataset: Our evaluation utilized 5,000 COCO test images from the Karpathy split. Baseline methods: We benchmarked against state-of-the-art captioning models, including BLIP-2 [17], InstructBLIP [8], and LLaVA-1.5 [20]. The evaluation focused on each model's ability to produce accurate, detailed, and coherent captions that effectively encapsulate the essence of the images. Evaluation Metric: We employed two metrics: CLIP-Score [12] and CLIP-Image-Score (Sec. 4). The CLIP-Score, a prevalent metric in image caption quality assessment, involves processing the raw image through the CLIP image encoder and the caption through the CLIP text encoder. The resultant embeddings are then compared for cosine similarity, with a higher score indicating greater semantic resemblance between the image and the caption. For our analysis, we first calculated the CLIP-Score for each image_caption pair, then averaged these scores across all 50,000 text/image pairs, scaling the result by a factor of 100. Table 2 displays the comparative performance of various image captioning methods on the 5,000 COCO test set images. The results demonstrate that our VisualFactChecker surpasses all baseline methods in performance." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.597, + 0.89, + 0.681 + ], + "angle": 0, + "content": "
Captioning MethodCLIP-Score (%) ↑CLIP-Image-Score (%) ↑
Human Label (COCO GT)30.36 (-2.54)71.21 (-2.40)
BLIP230.11 (-2.79)70.79 (-2.82)
InstructBLIP31.45 (-1.45)72.95 (-0.66)
LLaVA-1.532.08 (-0.82)73.24 (-0.37)
Kosmos-232.32 (-0.58)73.28 (-0.33)
VisualFactChecker (Ours)32.9073.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.691, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Table 2. Image captioning comparison with different metrics on 5000 COCO test set in Karpathy split, we use raw image and caption as input pairs for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As outlined in Sec. 4, the CLIP-Image-Score provides a complementary view to assess the quality of image captions. This metric is derived by comparing the cosine similarity between the CLIP embeddings of two images: the original image and a reconstructed image, which is generated using the provided caption through a text-to-image generation model. A higher CLIP-Image-Score signifies a more accurate and effective image caption. For this process, Stable Diffusion XL (SDXL) [28] is utilized as the designated text-to-image model to reconstruct images based on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14038" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.468, + 0.129 + ], + "angle": 0, + "content": "
Captioning MethodCLIP-Score (%)↑CLIP-Image-Score (%)↑
Cap3D33.44 (-0.57)79.88 (-0.44)
VisualFactChecker (Ours)34.0180.32
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.14, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Table 3. 3D object captioning comparison with different metrics on 1000 objects in Objaverse. For CLIP-Score, we use the average score of two views for evaluation. For CLIP-Image-Score, we use an off-the-shelf text-to-3D model, MVDream, to generate 3D models from 3D captions. We compare two views of the raw object and the same views of generated 3D object for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.469, + 0.294 + ], + "angle": 0, + "content": "the generated captions. Table 2 presents the CLIP-Image-Scores obtained for the 5000 images in the COCO test set, where our method outperforms all baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.469, + 0.536 + ], + "angle": 0, + "content": "3D object captioning. Dataset: 1,000 3D objects sampled from Objaverse dataset [9]. Baseline methods: We use state-of-the-art 3D object captioning model Cap3D [23] as the baseline. Cap3D uses 8 view images to generate the final object caption, our VisualFactChecker uses only 2 views to generate the object caption. Evaluation Metric: CLIP-Score and CLIP-Image-Score on multiple views rendered from 3D objects. To evaluate the similarity of a 3D object and the generated caption, we evaluate the similarity of the caption with the multi-view images used to generate the caption. Specifically, we evaluate the similarity of the generated caption with the two views that were used to generate the caption and use the average score to represent the CLIP-Score. Table. 3 shows the performance of 3D object captioning methods on 1,000 3D objects from Objaverse dataset. VisualFactChecker outperforms Cap3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.703 + ], + "angle": 0, + "content": "We also use CLIP-Image-Score to evaluate the 3D caption quality. CLIP-Image-Score needs reconstructed images to compare with the raw images. We treat the two views that were used to generate the 3D object caption as the raw image. To obtain the reconstructed image, we use an off-the-shelf text-to-3D generation model, MVDream, to generate a 3D object given the generated 3D object caption. We then render the same two views of images based on the generated 3D object, and we calculate the CLIP-Image-Score between the raw image and the rendered image. Table. 3 shows the CLIP-Image-Score on 1000 objects in Objaverse dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.391, + 0.728 + ], + "angle": 0, + "content": "5.2. Per Image Evaluation: Wining Rate" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.885 + ], + "angle": 0, + "content": "CLIP-Score and CLIP-Image-Score indicate an overall performance comparison, which shows an average score among all 5000 images. The average score may be dominated by a small group of images that have extremely high or low scores. To zoom in and show a more detailed comparison, we try to answer the following question: Given an image, what is the probability that one method performs better than another method on caption generation? To answer this question, we need to go over each image and calculate the winning rate for a pair of methods." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Specifically, for each image, we compare the CLIP-" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.887, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.187, + 0.887, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Figure 5. 2D image captioning comparison with pair-wise winning rate. VisualFactChecker (VFC) outperforms all baseline methods on both CLIP-Score (top) and CLIP-Image-Score (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Score of our VisualFactChecker caption against the captions generated from different baselines respectively, and calculate the wining probability of our method and the baselines. Fig. 5 shows the results, for example, we can see that in the pair-wise comparison, our VisualFactChecker performs better (higher CLIP-Score) than LLaVA-1.5 on \\(64.4\\%\\) of 5000 images (3220 images)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Calculating the winning rate over all images provides a more detailed analysis that zooms in on the comparison of each image, which shows a complementary view than overall average CLIP-Score." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.891, + 0.547 + ], + "angle": 0, + "content": "5.3. Fine-grained Evaluation: Human and GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The CLIP-Score and CLIP-Image-Score offer a general comparison of overall performance. A pairwise per-image winning rate provides a more specific analysis, evaluating performance on individual images. However, the research highlighted in related studies [15] indicates that the CLIP-Score may not be ideally suited for image-to-image comparison tasks. Furthermore, relying on a single score fails to provide a nuanced comparison across criteria, such as accuracy and level of detail. We use Human evaluation and GPT-4V to provide a more fine-grained evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Human evaluation using Amazon Mechanical Turk (AMT). We employed a pairwise comparison strategy. From the COCO dataset, we randomly selected 100 images out of 5000. For each image, our caption was compared against 5 baseline captions respectively. To reduce variance, each comparison was done by 3 different AMT workers and we used their majority voting as the final selection. This resulted in a total of 1500 comparisons collected on AMT. AMT UI is shown in the appendix. The workers were presented with two competing captions — one from a baseline method and one from our VisualFactChecker, in randomized order. They were instructed to select the better caption describing the image based on 3 aspects: correctness, de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "tailness, and fluency. Results in Fig. 6 show our captions are more preferred by humans. The human evaluation instruction and web UI is shown in Appendix B." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.151, + 0.465, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.253, + 0.457, + 0.267 + ], + "angle": 0, + "content": "Figure 6. Amazon Mechanical Turk human evaluation results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.284, + 0.473, + 0.557 + ], + "angle": 0, + "content": "GPT-4V evaluation. Our study applied GPT-4V for evaluating captions in a manner akin to the caption evaluation process used in DALLE-3. We use the same randomly selected 100 images from COCO as in Human evaluation. For each image, we considered the captions generated by 5 baseline methods alongside the caption produced by our VisualFactChecker. We then presented GPT-4V with the raw image, our reference caption, and the four baseline captions. Our designed prompt instructed GPT-4V to compare each baseline caption against our reference caption, focusing on two primary aspects: correctness and detail. GPT-4V was tasked with providing a pairwise, detailed comparison for each pair, including justifications for its assessments. Based on these comparative insights, GPT-4V classified each baseline method caption as either \"better\" or \"worse\" than our VisualFactChecker. Fig. 5 shows the comprehensive results. More details about the GPT-4V evaluation prompt and examples are shown in Appendix B." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.569, + 0.465, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.672, + 0.468, + 0.701 + ], + "angle": 0, + "content": "Figure 7. GPT-4V evaluation results. Our captions are significantly better than baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.231, + 0.743 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In our ablation study, we explore the impact of various components on performance. For 2D captioning tasks, we assess the efficacy of initial captioning models, LLaVA-1.5 and Kosmos-2, using the CLIP-Score metric for the captions they generate on the same 5000 COCO test images. Additionally, we ablate our method's performance in the absence of the verification (fact checker) step, which aims to mitigate hallucinations through detection grounding. Table 4 shows the detailed results. Likewise, in the context of 3D object captioning, we evaluate the individual contribu" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.23 + ], + "angle": 0, + "content": "
Methods or StepsCLIP-Score
2DLLaVA-1.532.08 (-0.33)
Kosmos-232.32 (-0.09)
VisualFactChecker (w/o fact check)32.41
VisualFactChecker32.90 (+0.49)
3DLLaVA-1.532.05 (-0.66)
InstructBLIP32.51 (-0.20)
VisualFactChecker (w/o fact check)32.71
VisualFactChecker34.01 (+1.30)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Table 4. Ablation study on captioning 2D images (5000 COCO test dataset) and 3D objects (1000 Objaverse)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.401 + ], + "angle": 0, + "content": "tions of initial captioners, namely LLaVA-1.5 and InstructBLIP on the same 1000 Objaverse 3D objects. We further investigate the performance of our methodology without the fact checker, which in this case operates by leveraging a VQA model to reduce hallucinations. Table 4 shows the detailed results. These results highlight the significance of fact checker in our approach." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.41, + 0.859, + 0.427 + ], + "angle": 0, + "content": "5.5. Qualitative Results and Prompt Following" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Other than quantitative evaluation results, we show more qualitative examples of VisualFactChecker for 2D and 3D captions in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.525 + ], + "angle": 0, + "content": "By leveraging an LLM, VisualFactChecker can follow complex instructions to write captions in various styles. Examples are shown in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.619, + 0.553 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.819 + ], + "angle": 0, + "content": "We propose the VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity and detailed captions. By utilizing an LLM to chain multimodal models and object detection and VQA models, VFC reduces hallucination in long captions. We conducted a comprehensive caption evaluation using different metrics, including 1) image-text similarity using CLIP-Score, 2) image-reconstructed image similarity using our proposed CLIP-Image-Score, 3) human study, and 4) fine-grained evaluation using GPT-4V. Compared with open-sourced captioning models, our method achieves state-of-the-art in both 2D and 3D captioning. Our work shows combining open-sourced models into a pipeline can significantly close the captioning performance gap with proprietary models like GPT-4V. In the future, we plan to improve our pipeline further by including more components for fact-checking and making it more automatic in deciding which components to use." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.819, + 0.892, + 0.895 + ], + "angle": 0, + "content": "Acknowledgments We would like to thank Siddharth Gururani for helping with our human evaluation using Amazon Mechanical Turk; Haochen Wang for his help in preprocessing 3D data. We also thank Qinsheng Zhang, Yogesh Balaji, and Yen-Chen Lin for their helpful discussion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14040" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. Technical report, OpenAI, 2023. 2, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Simone Bianco, Luigi Celona, Marco Donzella, and Paolo Napoletano. Improving image captioning descriptiveness by ranking and llm-based fusion. arXiv preprint arXiv:2306.11593, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[6] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. Ic3: Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[10] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[11] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[14] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[15] Klemen Kotar, Stephen Tian, Hong-Xing Yu, Daniel LK Yamins, and Jiajun Wu. Are these the same apple? comparing images based on object intrinsics. arXiv preprint arXiv:2311.00750, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[18] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[19] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[22] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[23] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. In NeurIPS, 2023. 2, 3, 7, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[24] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[25] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.778, + 0.703 + ], + "angle": 0, + "content": "[26] OpenAI. Gpt-4 technical report, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[27] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[29] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[30] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14041" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[31] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven CH Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[33] Bin Wang, Fan Wu, Xiao Han, Jiahui Peng, Huaping Zhong, Pan Zhang, Xiaoyi Dong, Weijia Li, Wei Li, Jiaqi Wang, et al. Vigo: Visual instruction generation and correction. arXiv preprint arXiv:2308.12714, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[34] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.458 + ], + "angle": 0, + "content": "[35] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[36] Yujia Xie, Luowei Zhou, Xiyang Dai, Lu Yuan, Nguyen Bach, Ce Liu, and Michael Zeng. Visual clues: Bridging vision and language foundations for image paragraph captioning. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[37] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[38] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14042" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_origin.pdf b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56ec592d2d0491a0e618b62c57dcc32e9c7d5b96 --- /dev/null +++ b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/41baf37b-e81b-4a82-a385-a78a616c7868_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e536edeafea5b8dff88460a931e65c6362e980a97a4defd8c4b83584598f5ead +size 4106874 diff --git a/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/full.md b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8e71350ba42a2dbcc78904ea09edb594f7f12f6 --- /dev/null +++ b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/full.md @@ -0,0 +1,271 @@ +# Visual Fact Checker: Enabling High-Fidelity Detailed Caption Generation + +Yunhao Ge Xiaohui Zeng Jacob Samuel Huffman Tsung-Yi Lin Ming-Yu Liu Yin Cui + +NVIDIA + +https://research.nvidia.com/labs/dir/vfc/ + +![](images/47a603a269fbdc63a9cc62eca601ffcd59f133a164af4a91cb0930c938d43cab.jpg) + +# GPT-4V + +The image shows a young girl in a field with her arms joyfully raised. She is holding a bunch of pink balloons in her right hand. She's wearing a plaid shirt and jeans, and her hair is styled in what appears to be a ponytail. The girl is smiling widely, and her pose suggests she is in mid-celebration or play. The field is filled with wildflowers, and trees can be seen in the background under a blue sky with some clouds. The setting appears to be a sunny day, perfect for outdoor activities. The image captures a moment of carefree childhood happiness. + +![](images/c92ae57f7240e51237a6b26fc46628e52ce7f94ace1fa4f0eb70378e13772fda.jpg) + +![](images/b6a1c0f8db182ea61e87b8d41e4d73c375dbbf3563866da63822f81ddea9fffd.jpg) + +# VFC + +A happy little girl is standing in a green field, wearing a plaid shirt and holding onto a string of pink balloons. The balloons are floating in the air, creating a playful and joyful atmosphere. The girl appears to be enjoying the moment, with a smile on her face. In the background, there is a vast expanse of green grass, stretching out as far as the eye can see. The sky above is a bright blue, with a few white clouds scattered about. The overall mood of the scene is one of carefree happiness and joy, with the pink balloons adding a touch of whimsy and playfulness to the atmosphere. + +![](images/1894e6b7bc4c70aa520e357cff3ab5b98f24aa9f56cc4b8c7b6150b3ff83db83.jpg) + +# Cap3D + +A 3D model of a small wooden tower with a blue roof. + +![](images/a46745bcb26d4b301c683abf169c5bdf48e792161b55cd4098f42f309579053e.jpg) +Figure 1. Comparison of VisualFactChecker (VFC) with GPT-4V and Cap3D. VFC can generate high-fidelity detailed captions that closely match GPT-4V's quality for 2D images and offer significantly more details for 3D objects than Cap3D. VFC used a pre-trained Llama-2 as the LLM when generating the caption for the above 2D image. + +![](images/339025ec5a80eafef85cc97d595a97ae86ba26380006fef6a4ff7f7d3fdba461.jpg) + +![](images/cbbd59de05aace3209690e0310e496ba880ea8365c98b37a600cdd7f6542c01a.jpg) + +# VFC + +A 3D model of a three-level, intricately designed wooden tower with a contrasting blue roof and door, standing on a wooden base. The tower, brown in color, resembles a fusion of a house, a tower, and a castle. At the very top of the tower, there is a crescent moon design. The overall design adds a touch of fantasy to the scene. + +# Abstract + +Existing automatic captioning methods for visual content face challenges such as lack of detail, content hallucination, and poor instruction following. In this work, we propose VisualFactChecker (VFC), a flexible training-free pipeline that generates high-fidelity and detailed captions for both 2D images and 3D objects. VFC consists of three steps: 1) proposal, where image-to-text captioning models propose multiple initial captions; 2) verification, where a large language model (LLM) utilizes tools such as object detection and VQA models to fact-check proposed captions; 3) captioning, where an LLM generates the final caption by summarizing caption proposals and the fact check verifica + +tion results. In this step, VFC can flexibly generate captions in various styles following complex instructions. We conduct comprehensive captioning evaluations using four metrics: 1) CLIP-Score for image-text similarity; 2) CLIP-Image-Score for measuring the image-image similarity between the original and the reconstructed image generated by a text-to-image model using the caption. 3) human study on Amazon Mechanical Turk; 4) GPT-4V for fine-grained evaluation. Evaluation results show that VFC outperforms state-of-the-art open-sourced captioning methods for 2D images on the COCO dataset and 3D assets on the Objverse dataset. Our study demonstrates that by combining open-source models into a pipeline, we can attain captioning capability comparable to proprietary models such as GPT-4V, despite being over $10 \times$ smaller in model size. + +# 1. Introduction + +Image captioning is a pivotal challenge in computer vision and natural language processing. Its central goal is to encapsulate visual data within a textual description, which requires a nuanced understanding of both modalities. The recent advent of multimodal large language models (MM-LLMs), such as GPT-4V [26], and text-to-image generation models, such as DALLE-3 [3], has marked significant progress in this field. These proprietary models could leverage expansive human-labeled data and enormous computing resources to learn to generate detailed and contextually appropriate image descriptions. On the other hand, existing open-sourced captioning methods in the community still face significant challenges. Methods such as BLIP-2 [17] and OFA [35] often yield overly succinct captions that neglect essential visual information. Conversely, systems like Mini-GPT4 [39], InstructBLIP [8], and LLaVA [20, 21] can suffer from hallucination, producing long descriptions that do not align with the actual content of the images. + +In light of this, we propose VisualFactChecker (VFC), a flexible training-free pipeline designed to produce accurate and comprehensive captions for both 2D images and 3D objects. Fig. 1 shows examples of captions generated by VFC and their comparisons with captions generated by GPT-4V [26] and Cap3D [23]. Captions generated by VFC are faithful textural representations of the visual contents. This can also be verified by reconstructing images and 3d objects from captions using state-of-the-art text-to-image and text-to-3d models, as shown in Fig. 2. + +VFC focuses on tackling hallucinations and insufficient details in generated captions and is structured around three core components: Proposer, serving as the system's "eye", creating detailed caption proposals as preliminary captions by using image-to-text captioning models; Large Language Model, acting as the "brain", calling and summarizing information from other components, and leveraging its advanced generalization capabilities to steer the captioning process following specified captioning instructions; Detector and VQA models, functioning as "tools" utilized by the LLM for fact-checking caption proposals, ensuring the fidelity of the final generated caption. VFC is versatile and effectively handles captioning for both 2D images and 3D objects through a unified pipeline. Fig. 3 shows an overview of the pipeline. The details of each component and their interplay are explained in Sec. 3. + +To comprehensively evaluate the generated captions, other than leveraging the commonly used CLIP-Score that primarily gauges the image-caption similarity, we propose a new metric: the CLIP-Image-Score. This metric assesses the similarity between the input image and a reconstructed image created by a text-to-image model from the caption, offering a complementary measure. Furthermore, we conducted a human study on Amazon Mechanical Turk for cap + +tion evaluation. Finally, we also performed a fine-grained evaluation by asking GPT-4V to compare and judge captions with detailed reasoning. The combination of CLIP-Score, CLIP-Image-Score, GPT-4V, and human study provides a more robust evaluation of captions. + +We summarize our main contributions as follows: 1) We propose VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity detailed 2D and 3D captions, effectively mitigating the challenge of hallucination in long captions. (2) CLIP-Image-Score: A novel caption evaluation metric that measures the similarity between the input image and a reconstructed image from the caption. (3) Our evaluation shows that VisualFactChecker achieves state-of-the-art results in 2D and 3D captioning tasks compared with open-sourced models. (4) Our work shows that using an LLM to chain open-source models can achieve captioning capability on par with proprietary models such as GPT-4V. + +# 2. Related Work + +# 2.1. Image Captioning + +Image captioning has made significant progress with the advent of deep learning. Pioneering works [2, 10, 14] primarily focus on integrating deep neural networks for enhanced image understanding and language generation. + +Recent strides have been made with the introduction of Multimodal-Large Language Models (MM-LLMs), which are trained on extensive vision and language data. The general approach involves leveraging a pre-trained large language model (LLM) and a vision encoder with a projector to align with the LLM's embeddings, thus enhancing visual understanding. Several models have emerged as significant contributors in this domain. BLIP [16], BLIP2 [17], OFA [35], Flamingo [1], Kosmos-2 [27], MiniGPT4 [39], InstructBLIP [8], LLaVA [20, 21] have demonstrated impressive performance in single-view image captioning tasks. However, they exhibit varying limitations. For instance, BLIP-2 and OFA often generate overly concise captions, while others, like InstructBLIP, can produce detailed captions that often include inaccurate or hallucinatory content. Our method aims to address these limitations by combining different models into a pipeline via an LLM, striking a better balance between accuracy and detailedness in generated captions while mitigating hallucinations. + +# 2.2. Large Language Models for Captioning + +Recent advancements in large language models (LLMs) like GPT-3 [5], LAMDA [30], PALM [7], Llama [32], GPT-4 [26] have demonstrated exceptional zero-shot capabilities in language analysis and summarization tasks. This proficiency has naturally extended to the multimodal domain, particularly in image-language contexts, where LLMs can summarize multimodal information in a zero-shot manner. + +![](images/43520a7ed5c9835236b65f482fe3cf6d16c53e5fad72b91f335b052ef502c1d9.jpg) +Figure 2. We use DALLE-3 [3] as a text-to-image model to reconstruct 2D images using generated captions from different captioning methods (BLIP-2, LLaVA-1.5 and ours). Similarly, we use MVDream [29] as a text-to-3D model to reconstruct 3D objects using different 3D captions (generated by Cap3D [23] and ours). From the results, we can see that the reconstructed images or 3D objects using BLIP-2 or Cap3D captions are less similar than the input ones, suggesting their captions may not contain sufficient information or incorrectly describe the visual contents; the reconstructed images using LLaVA-1.5 captions contain objects or scenes that are not present in the original images (top: people in the background, bottom: pedestrians and cars on the street), suggesting there might be hallucinations in LLaVA-1.5 captions. Images or 3D objects reconstructed using our captions are more similar to the inputs. + +![](images/c8c65815c524bd46c9b82e60c2f682e30b4bdda81322e45cba8da9fcae37f06a.jpg) + +Vision-blind LLMs are prominent in multimodal applications, often utilizing language-only prefixes generated by pre-trained tools. Clipcap [25] demonstrates this by using a continuous embedding as a prompt for a GPT-style language model, achieving notable performance in single-viewpoint image captioning. Similarly, Promptcap [13] and PNP-VQA [31] leverage natural language prompts with GPT models to excel in visual question answering. + +Recent methods have employed LLMs to generate image captions by summarizing initial captions or keywords from Vision-Language models. For instance, Socratic models [37] use a CLIP-based model to extract key tags from images, followed by GPT-3 with specialized prompts to create stylized captions. ChatCaptioner [38] builds upon this by integrating ChatGPT and BLIP-2 [17] in a conversational approach for question-answering about the image, and summarizing them into a caption. Visual Clues [36] uses similar tags to generate a paragraph-caption. IC3 [6] and LLM-Fusion [4] use LLMs to summarize captions from existing models augmented with temperature-based sampling. Cap3D [24] extends this concept to 3D object. + +Our method differentiates itself in two critical ways: First, we focus on reducing hallucinations in captions by employing visual grounding tools, such as object detection, to fact-check captions for enhanced accuracy. Second, our pipeline can be used for captioning both 2D images and 3D objects. Unlike previous methods that rely on a single captioning model, we integrate multiple captioning sources from different models, ensuring a more comprehensive coverage of visual content to generate captions. + +# 2.3. Hallucination in MM-LLM + +There are two popular topics on the hallucination of MM-LLMs. (1) Hallucination evaluation: Detection approaches such as Gunjal et al. [11] train classification models to identify hallucination. They focus on distinguishing between accurate and hallucinated content. Ground truth comparison methods [18, 34] compare model outputs with ground truth data to detect hallucinations. These techniques assess the alignment of generated captions with actual image content. (2) Mitigation Strategies [22]: Data optimization methods such as Liu et al. [19] address hallucination by creating negative instances in training datasets to reduce model overconfidence. Iterative generation methods such as Wang et al. [33] adopt an iterative process for caption generation, where brief answers are generated in succession and amalgamated, aiming to improve accuracy and relevance. + +Our VisualFactChecker is a training-free pipeline mitigating hallucination in image captioning. Our method utilizes visual grounding tools for improved accuracy, thereby actively reducing the hallucination and offering high-fidelity captions for both 2D images and 3D objects. + +# 3. Visual Fact Checker + +This section introduces the key components of VisualFactChecker as shown in Fig. 3 in detail and explains their interplay in generating accurate and detailed captions. The following sections delve into specifics. First, we detail the pipeline for 2D image captioning (Sec. 3.1), with Fig. 3 (top) illustrating this process. Then, we explore how our + +![](images/3b92976ffcac2b8757203defdb16f022ca02943023ae194a39cd44644195c39b.jpg) + +![](images/e25a6dd7bf2f14bcd4d8c693e023bfe146fcd49e9581ff2ea924bb7ee09c90b2.jpg) +Figure 3. Pipeline of the VisualFactChecker for captioning 2D images (top) and 3D objects (bottom). The process begins with the input being captioned by two multimodal captioning models (Captioner-1 and Captioner-2) to generate preliminary captions. These captions are then verified using a Large Language Model (LLM) to call object detection (Detector) and VQA models for fact-checking the captions. Finally, the LLM incorporates all the results and summarizes the final caption by following instructions. + +approach is adapted for 3D object captioning as shown in Fig. 3 (bottom), underscoring both shared methodologies and unique aspects relevant to 3D contexts (Sec. 3.2). + +# 3.1.2D Image Captioning + +The caption generation takes three steps: 1) proposal, 2) verification, and 3) captioning. Each step is detailed below. + +**Proposal:** The Proposal step serves as the cornerstone of the captioning process that generates initial captions. This is achieved through the utilization of advanced image-to-text models, specifically "LLaVA" and "Kosmos2". These models are trained on expansive datasets, enabling them to comprehend and interpret visual content effectively. By analyzing the input image, they suggest various preliminary captions, each reflecting different facets and interpretations of the image (Fig. 3 top). The rationale behind using multiple image-to-text multimodal LLMs lies in the complexity of adequately capturing an image's essence in a single attempt. Since an image can be accurately described in numerous ways, different models bring unique perspectives, thereby encompassing a broader range of information present in the image. Although the initial captions proposed may not possess perfect fidelity, the primary objective at this stage is to generate captions that are as comprehensive as possible. Fig. 3 displays the specific prompts we used for each step, with more details in Appendix A. + +Verification and Captioning: The goal of the verification step is to scrutinize and rectify any inaccuracies or hallucinations in the captions during the proposal step. It employs a combination of a Large Language Model (LLM) and grounding tools, including an open-vocabulary object detection model and/or a visual question answering (VQA) model. Here the LLM can be GPT-4 or Llama2. As shown in Fig. 3 (top), the process involves the following steps: Step 1: LLM first summarizes the initial detailed descriptions from different MM-LLMs into a single, detailed caption. While this caption is comprehensive, it may not always be accurate. Step 2: The LLM then analyzes this synthesized caption, identifying all objects that could be verified by object detection and summarizing an object checklist. In 2D image captioning, the focus is on eliminating hallucinations, particularly descriptions of non-existent objects in the image. Identifying these objects is crucial for the subsequent verification process. Step 3: Taking the object checklist as input, an open-vocabulary object detection model examines candidate objects in the checklist and determines their presence in the image. This step is pivotal in validating the existence of objects mentioned in the caption, thus supporting the fidelity of the caption. + +After verification, we go to the last captioning step: Based on the object detection results, the LLM revises the summarized single detailed caption. Each object described + +in the caption is cross-checked; if detected, it remains unchanged, while undetected objects are considered potential hallucinations and are removed from the caption. This step results in a final caption that is both detailed and reliable. The underlying assumption is that the detection model, serving as an object grounding expert, provides more reliable results than a general image descriptor. + +In the verification and captioning steps, the LLM plays a pivotal role as a "brain". It starts by parsing the initial caption and identifying key objects for detailed examination. The LLM then meticulously assesses whether each object mentioned actually appears in the image based on detection results. Following this thorough analysis, it refines and revises the initial captions, transforming them into final versions that are both coherent and richly detailed. The LLM is instrumental in guaranteeing linguistic fluency, ensuring that the captions not only accurately represent the image but also maintain the necessary level of detail for high-fidelity captioning. Moreover, the LLM can follow complex instructions to write the captions in a specified style, such as a caption that only mentions the foreground objects without mentioning the background. Fig. 3 displays the specific prompts used for each step. + +# 3.2. 3D Object Captioning + +The 3D object captioning process follows a similar structural pipeline to that of 2D images, with a few key distinctions in certain steps, as depicted in Fig. 3 (bottom). In 3D captioning, an object may present multiple views, each offering unique information. The comprehensive caption for a 3D object is derived by integrating the perspectives from all these views. For each view, VisualFactChecker is employed to create a detailed, high-fidelity description. Subsequently, the LLM (GPT-4 or Llama-2) is used to amalgamate the information from all views, producing a unified caption for the 3D object. In particular, for each view's captioning, we have the same three-step approach akin to 2D image captioning. In the proposal step, LLaVA-1.5 and InstructBLIP are utilized for generating initial detailed descriptions. We opt out of using Kosmos2 for single 3D objects due to its less effective performance in providing detailed descriptions, possibly linked to its reliance on an implicit detection model. Additionally, a slightly modified prompt is used (see Fig. 3 bottom), which incorporates 3D-specific considerations. In the verification and captioning step, we primarily address hallucinations related to the attributes of 3D objects, such as shape and color. To mitigate these inaccuracies, rather than enumerating potential objects, we employ the LLM to generate five critical questions that could influence a text-to-3D generation model in reconstructing the 3D model. Following this, we utilize VQA models (specifically LLaVA-1.5) to respond to these questions based on the input 3D object view image. Subsequently, the LLM amends + +![](images/9ef51dbb1039ec8cab64168025cbc7ce14c2160531aef1c3d93f35d131460dfe.jpg) +Figure 4. The CLIP-Image-Score pipeline evaluates caption accuracy by encoding an original image $X$ into a feature representation $I_{X}$ using a CLIP image encoder. A captioning model generates a caption that is then input into a text-to-image model to reconstruct an image $X'$ , which is encoded to $I_{X'}$ . The score is computed by assessing the cosine similarity between $I_{X}$ and $I_{X'}$ , providing a measure of the caption's fidelity and hallucination detection. + +the initial caption in accordance with the answers provided by the VQA model. We operate under the assumption that answering targeted questions results in fewer hallucinations compared to generating a general description. Once the caption for each individual view is complete, the LLM synthesizes these multiple perspectives into a singular, comprehensive caption for the entire 3D object. The prompts used for the LLM at each stage are detailed in Appendix A. + +# 4. CLIP-Image-Score + +Accurate evaluation of caption correctness and detailedness is paramount in determining the performance of an image captioning model. Traditional metrics like the CLIP-Score [12] have served as a standard for measuring the alignment between generated captions and their corresponding images. However, our CLIP-score may lack the sensitivity needed to detect the specific issue of hallucination within captions. + +We present the CLIP-Image-Score, an alternative metric specifically developed to reflect the subtleties of caption quality. This metric is different from CLIP-Score by introducing an additional reconstruction step. Specifically, the CLIP-Image-score evaluates the similarity between the original image and a reconstructed version of the image generated by a fixed text-to-image model using the caption as a prompt. By comparing the raw image to its reconstructed image, the metric is able to detect discrepancies indicative of hallucination, thus providing a different perspective of the caption quality assessment. The underlying principle of the CLIP-Image-Score is the recognition that multiple "correct" captions may exist for a single image. However, it's only when a caption is both "detail" and "correct" that the reconstructed image closely resembles the original. Moreover, any hallucinations present in the caption become evident in the reconstructed image. Fig. 2 presents examples of such reconstructions. For instance, consider the results from LLaVA-1.5 shown in the third column. The cap + +tion generated for the first image falsely mentions "several other people in the background". This error is clearly reflected in the image reconstructed by the text-to-image generator. In essence, comparing the two images indirectly ensures alignment between the image and its caption, thereby providing a complementary method to assess the quality of the caption than directly comparing the image and caption. + +The CLIP-Image-Score evaluation process is depicted in the following steps: + +- Caption Generation: An original image $X$ is input into a captioning model, which generates a caption. +- Caption-to-Image Reconstruction: This generated caption is then used as input for a text-to-image model, which creates a reconstructed image $X'$ that visually represents the textual description. +- Raw Image Encoding: The original image $X$ is processed through a CLIP image encoder, translating the visual content into an encoded representation $I_{X}$ . +- Reconstructed Image Encoding: The reconstructed image is also processed through the CLIP image encoder to obtain its encoded representation $I_{X'}$ . +- Score Calculation: Finally, the encoded representations of the original and reconstructed images are compared to calculate the CLIP-Image-Score. The score is given by the cosine similarity, which assesses the congruence between $I_{X}$ and $I_{X'}$ : + +$$ +\text {C L I P - I m a g e - S c o r e} = \frac {I _ {X} \cdot I _ {X ^ {\prime}}}{\| I _ {X} \| \times \| I _ {X ^ {\prime}} \|} \tag {1} +$$ + +Most notably, CLIP-Image-Score offers a sensitive measure for detecting hallucinations. In scenarios where the generated caption includes elements that are not in the original image, the reconstructed image will also likely contain these discrepancies. By comparing the original and reconstructed images, the CLIP-Image-Score can effectively highlight these differences, offering a clearer insight into the fidelity and accuracy of the generated caption. + +Furthermore, CLIP-Image-Score turns a cross-modality comparison into a more intuitive comparison in the same image modality (as shown in Fig. 4). CLIP-Image-Score represents a new complementary perspective for image captioning evaluation. By leveraging the capabilities of text-to-image models and focusing on the congruence between the original and reconstructed images, it provides an accurate assessment of caption quality, particularly in identifying and measuring hallucinations, thereby enhancing the overall reliability of caption generation systems. + +# 5. Experiments + +This section presents a thorough evaluation of captioning models across both 2D and 3D visual content, employing a variety of datasets and methodologies. Table 1 provides a summary of our comprehensive evaluation experiments. + +
EvalInput pairs for evaluationMethodReference
2DRaw imageCaptionCLIP-ScoreTable 2
Human evaluationFig. 6
CPT4V evaluationFig. 7
Raw imageImage(recon)CLIP-Image-ScoreTable 2
3DMulti-view (raw)CaptionCLIP-ScoreTable 3
GPT4V evaluationFig. 7
Multi-view (raw)Multi-view (recon)CLIP-Image-ScoreTable 3
+ +# 5.1. Overall: CLIP-Score and CLIP-Image-Score + +2D image captioning. Dataset: Our evaluation utilized 5,000 COCO test images from the Karpathy split. Baseline methods: We benchmarked against state-of-the-art captioning models, including BLIP-2 [17], InstructBLIP [8], and LLaVA-1.5 [20]. The evaluation focused on each model's ability to produce accurate, detailed, and coherent captions that effectively encapsulate the essence of the images. Evaluation Metric: We employed two metrics: CLIP-Score [12] and CLIP-Image-Score (Sec. 4). The CLIP-Score, a prevalent metric in image caption quality assessment, involves processing the raw image through the CLIP image encoder and the caption through the CLIP text encoder. The resultant embeddings are then compared for cosine similarity, with a higher score indicating greater semantic resemblance between the image and the caption. For our analysis, we first calculated the CLIP-Score for each image_caption pair, then averaged these scores across all 50,000 text/image pairs, scaling the result by a factor of 100. Table 2 displays the comparative performance of various image captioning methods on the 5,000 COCO test set images. The results demonstrate that our VisualFactChecker surpasses all baseline methods in performance. + +Table 1. Summary of evaluation methods and results. + +
Captioning MethodCLIP-Score (%) ↑CLIP-Image-Score (%) ↑
Human Label (COCO GT)30.36 (-2.54)71.21 (-2.40)
BLIP230.11 (-2.79)70.79 (-2.82)
InstructBLIP31.45 (-1.45)72.95 (-0.66)
LLaVA-1.532.08 (-0.82)73.24 (-0.37)
Kosmos-232.32 (-0.58)73.28 (-0.33)
VisualFactChecker (Ours)32.9073.61
+ +Table 2. Image captioning comparison with different metrics on 5000 COCO test set in Karpathy split, we use raw image and caption as input pairs for evaluation. + +As outlined in Sec. 4, the CLIP-Image-Score provides a complementary view to assess the quality of image captions. This metric is derived by comparing the cosine similarity between the CLIP embeddings of two images: the original image and a reconstructed image, which is generated using the provided caption through a text-to-image generation model. A higher CLIP-Image-Score signifies a more accurate and effective image caption. For this process, Stable Diffusion XL (SDXL) [28] is utilized as the designated text-to-image model to reconstruct images based on + +
Captioning MethodCLIP-Score (%)↑CLIP-Image-Score (%)↑
Cap3D33.44 (-0.57)79.88 (-0.44)
VisualFactChecker (Ours)34.0180.32
+ +Table 3. 3D object captioning comparison with different metrics on 1000 objects in Objaverse. For CLIP-Score, we use the average score of two views for evaluation. For CLIP-Image-Score, we use an off-the-shelf text-to-3D model, MVDream, to generate 3D models from 3D captions. We compare two views of the raw object and the same views of generated 3D object for evaluation. + +the generated captions. Table 2 presents the CLIP-Image-Scores obtained for the 5000 images in the COCO test set, where our method outperforms all baseline methods. + +3D object captioning. Dataset: 1,000 3D objects sampled from Objaverse dataset [9]. Baseline methods: We use state-of-the-art 3D object captioning model Cap3D [23] as the baseline. Cap3D uses 8 view images to generate the final object caption, our VisualFactChecker uses only 2 views to generate the object caption. Evaluation Metric: CLIP-Score and CLIP-Image-Score on multiple views rendered from 3D objects. To evaluate the similarity of a 3D object and the generated caption, we evaluate the similarity of the caption with the multi-view images used to generate the caption. Specifically, we evaluate the similarity of the generated caption with the two views that were used to generate the caption and use the average score to represent the CLIP-Score. Table. 3 shows the performance of 3D object captioning methods on 1,000 3D objects from Objaverse dataset. VisualFactChecker outperforms Cap3D. + +We also use CLIP-Image-Score to evaluate the 3D caption quality. CLIP-Image-Score needs reconstructed images to compare with the raw images. We treat the two views that were used to generate the 3D object caption as the raw image. To obtain the reconstructed image, we use an off-the-shelf text-to-3D generation model, MVDream, to generate a 3D object given the generated 3D object caption. We then render the same two views of images based on the generated 3D object, and we calculate the CLIP-Image-Score between the raw image and the rendered image. Table. 3 shows the CLIP-Image-Score on 1000 objects in Objaverse dataset. + +# 5.2. Per Image Evaluation: Wining Rate + +CLIP-Score and CLIP-Image-Score indicate an overall performance comparison, which shows an average score among all 5000 images. The average score may be dominated by a small group of images that have extremely high or low scores. To zoom in and show a more detailed comparison, we try to answer the following question: Given an image, what is the probability that one method performs better than another method on caption generation? To answer this question, we need to go over each image and calculate the winning rate for a pair of methods. + +Specifically, for each image, we compare the CLIP- + +![](images/a124c0833c55a0b0af162c9a7fe2e88e4508b82c4185a3077a20e79a7382f589.jpg) + +![](images/7d81d81730484ac0fabc42c59d14555aca14fedf801aa2c86d30fcbc152f653c.jpg) +Figure 5. 2D image captioning comparison with pair-wise winning rate. VisualFactChecker (VFC) outperforms all baseline methods on both CLIP-Score (top) and CLIP-Image-Score (bottom). + +Score of our VisualFactChecker caption against the captions generated from different baselines respectively, and calculate the wining probability of our method and the baselines. Fig. 5 shows the results, for example, we can see that in the pair-wise comparison, our VisualFactChecker performs better (higher CLIP-Score) than LLaVA-1.5 on $64.4\%$ of 5000 images (3220 images). + +Calculating the winning rate over all images provides a more detailed analysis that zooms in on the comparison of each image, which shows a complementary view than overall average CLIP-Score. + +# 5.3. Fine-grained Evaluation: Human and GPT-4V + +The CLIP-Score and CLIP-Image-Score offer a general comparison of overall performance. A pairwise per-image winning rate provides a more specific analysis, evaluating performance on individual images. However, the research highlighted in related studies [15] indicates that the CLIP-Score may not be ideally suited for image-to-image comparison tasks. Furthermore, relying on a single score fails to provide a nuanced comparison across criteria, such as accuracy and level of detail. We use Human evaluation and GPT-4V to provide a more fine-grained evaluation. + +Human evaluation using Amazon Mechanical Turk (AMT). We employed a pairwise comparison strategy. From the COCO dataset, we randomly selected 100 images out of 5000. For each image, our caption was compared against 5 baseline captions respectively. To reduce variance, each comparison was done by 3 different AMT workers and we used their majority voting as the final selection. This resulted in a total of 1500 comparisons collected on AMT. AMT UI is shown in the appendix. The workers were presented with two competing captions — one from a baseline method and one from our VisualFactChecker, in randomized order. They were instructed to select the better caption describing the image based on 3 aspects: correctness, de + +tailness, and fluency. Results in Fig. 6 show our captions are more preferred by humans. The human evaluation instruction and web UI is shown in Appendix B. + +![](images/c4ad4af8a681f8313aac73b805ea572e75585a1e3de13cafb537c815675f431f.jpg) +Figure 6. Amazon Mechanical Turk human evaluation results. + +GPT-4V evaluation. Our study applied GPT-4V for evaluating captions in a manner akin to the caption evaluation process used in DALLE-3. We use the same randomly selected 100 images from COCO as in Human evaluation. For each image, we considered the captions generated by 5 baseline methods alongside the caption produced by our VisualFactChecker. We then presented GPT-4V with the raw image, our reference caption, and the four baseline captions. Our designed prompt instructed GPT-4V to compare each baseline caption against our reference caption, focusing on two primary aspects: correctness and detail. GPT-4V was tasked with providing a pairwise, detailed comparison for each pair, including justifications for its assessments. Based on these comparative insights, GPT-4V classified each baseline method caption as either "better" or "worse" than our VisualFactChecker. Fig. 5 shows the comprehensive results. More details about the GPT-4V evaluation prompt and examples are shown in Appendix B. + +![](images/0a7b94bc4a4cfa9b509422ca41fabccc66307afa9d8ab8bc6273f3f7be418a74.jpg) +Figure 7. GPT-4V evaluation results. Our captions are significantly better than baselines. + +# 5.4. Ablation Study + +In our ablation study, we explore the impact of various components on performance. For 2D captioning tasks, we assess the efficacy of initial captioning models, LLaVA-1.5 and Kosmos-2, using the CLIP-Score metric for the captions they generate on the same 5000 COCO test images. Additionally, we ablate our method's performance in the absence of the verification (fact checker) step, which aims to mitigate hallucinations through detection grounding. Table 4 shows the detailed results. Likewise, in the context of 3D object captioning, we evaluate the individual contribu + +
Methods or StepsCLIP-Score
2DLLaVA-1.532.08 (-0.33)
Kosmos-232.32 (-0.09)
VisualFactChecker (w/o fact check)32.41
VisualFactChecker32.90 (+0.49)
3DLLaVA-1.532.05 (-0.66)
InstructBLIP32.51 (-0.20)
VisualFactChecker (w/o fact check)32.71
VisualFactChecker34.01 (+1.30)
+ +Table 4. Ablation study on captioning 2D images (5000 COCO test dataset) and 3D objects (1000 Objaverse). + +tions of initial captioners, namely LLaVA-1.5 and InstructBLIP on the same 1000 Objaverse 3D objects. We further investigate the performance of our methodology without the fact checker, which in this case operates by leveraging a VQA model to reduce hallucinations. Table 4 shows the detailed results. These results highlight the significance of fact checker in our approach. + +# 5.5. Qualitative Results and Prompt Following + +Other than quantitative evaluation results, we show more qualitative examples of VisualFactChecker for 2D and 3D captions in Appendix C. + +By leveraging an LLM, VisualFactChecker can follow complex instructions to write captions in various styles. Examples are shown in Appendix D. + +# 6. Conclusion + +We propose the VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity and detailed captions. By utilizing an LLM to chain multimodal models and object detection and VQA models, VFC reduces hallucination in long captions. We conducted a comprehensive caption evaluation using different metrics, including 1) image-text similarity using CLIP-Score, 2) image-reconstructed image similarity using our proposed CLIP-Image-Score, 3) human study, and 4) fine-grained evaluation using GPT-4V. Compared with open-sourced captioning models, our method achieves state-of-the-art in both 2D and 3D captioning. Our work shows combining open-sourced models into a pipeline can significantly close the captioning performance gap with proprietary models like GPT-4V. In the future, we plan to improve our pipeline further by including more components for fact-checking and making it more automatic in deciding which components to use. + +Acknowledgments We would like to thank Siddharth Gururani for helping with our human evaluation using Amazon Mechanical Turk; Haochen Wang for his help in preprocessing 3D data. We also thank Qinsheng Zhang, Yogesh Balaji, and Yen-Chen Lin for their helpful discussion. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2 +[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, 2018. 2 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. Technical report, OpenAI, 2023. 2, 3, 11 +[4] Simone Bianco, Luigi Celona, Marco Donzella, and Paolo Napoletano. Improving image captioning descriptiveness by ranking and llm-based fusion. arXiv preprint arXiv:2306.11593, 2023. 3 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2 +[6] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. Ic3: Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 3 +[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2 +[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 6 +[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 7 +[10] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2 +[11] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 3 +[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6 +[13] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 3 + +[14] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In ICCV, 2019. 2 +[15] Klemen Kotar, Stephen Tian, Hong-Xing Yu, Daniel LK Yamins, and Jiajun Wu. Are these the same apple? comparing images based on object intrinsics. arXiv preprint arXiv:2311.00750, 2023. 7 +[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2 +[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 6 +[18] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3 +[19] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3 +[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 6 +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2 +[22] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3 +[23] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. In NeurIPS, 2023. 2, 3, 7, 17 +[24] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3 +[25] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3 +[26] OpenAI. Gpt-4 technical report, 2023. 2 +[27] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6 +[29] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 17 +[30] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia + +Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2 +[31] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven CH Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773, 2022. 3 +[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[33] Bin Wang, Fan Wu, Xiao Han, Jiahui Peng, Huaping Zhong, Pan Zhang, Xiaoyi Dong, Weijia Li, Wei Li, Jiaqi Wang, et al. Vigo: Visual instruction generation and correction. arXiv preprint arXiv:2308.12714, 2023. 3 +[34] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 3 +[35] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 2 +[36] Yujia Xie, Luowei Zhou, Xiyang Dai, Lu Yuan, Nguyen Bach, Ce Liu, and Michael Zeng. Visual clues: Bridging vision and language foundations for image paragraph captioning. In NeurIPS, 2022. 3 +[37] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3 +[38] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 3 +[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/images.zip b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3b7e3796d493a5c5896c0407e658ac1ae2bca8a --- /dev/null +++ b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3e88d65f2afdd718c956370258caa58c65d5c955cc04c0294c8c97ef44bbe7a +size 449544 diff --git a/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/layout.json b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5d9568d9d163841e87bab2d62c5e5fc044965a6c --- /dev/null +++ b/2024/Visual Fact Checker_ Enabling High-Fidelity Detailed Caption Generation/layout.json @@ -0,0 +1,6360 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "Visual Fact Checker: Enabling High-Fidelity Detailed Caption Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 143, + 526, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 143, + 526, + 159 + ], + "spans": [ + { + "bbox": [ + 66, + 143, + 526, + 159 + ], + "type": "text", + "content": "Yunhao Ge Xiaohui Zeng Jacob Samuel Huffman Tsung-Yi Lin Ming-Yu Liu Yin Cui" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 273, + 171, + 319, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 171, + 319, + 183 + ], + "spans": [ + { + "bbox": [ + 273, + 171, + 319, + 183 + ], + "type": "text", + "content": "NVIDIA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 186, + 417, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 186, + 417, + 199 + ], + "spans": [ + { + "bbox": [ + 169, + 186, + 417, + 199 + ], + "type": "text", + "content": "https://research.nvidia.com/labs/dir/vfc/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 237, + 62, + 251 + ], + "blocks": [ + { + "bbox": [ + 49, + 237, + 62, + 251 + ], + "lines": [ + { + "bbox": [ + 49, + 237, + 62, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 237, + 62, + 251 + ], + "type": "image", + "image_path": "47a603a269fbdc63a9cc62eca601ffcd59f133a164af4a91cb0930c938d43cab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 239, + 91, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 239, + 91, + 248 + ], + "spans": [ + { + "bbox": [ + 64, + 239, + 91, + 248 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 50, + 251, + 206, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 251, + 206, + 339 + ], + "spans": [ + { + "bbox": [ + 50, + 251, + 206, + 339 + ], + "type": "text", + "content": "The image shows a young girl in a field with her arms joyfully raised. She is holding a bunch of pink balloons in her right hand. She's wearing a plaid shirt and jeans, and her hair is styled in what appears to be a ponytail. The girl is smiling widely, and her pose suggests she is in mid-celebration or play. The field is filled with wildflowers, and trees can be seen in the background under a blue sky with some clouds. The setting appears to be a sunny day, perfect for outdoor activities. The image captures a moment of carefree childhood happiness." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 230, + 226, + 365, + 363 + ], + "blocks": [ + { + "bbox": [ + 230, + 226, + 365, + 363 + ], + "lines": [ + { + "bbox": [ + 230, + 226, + 365, + 363 + ], + "spans": [ + { + "bbox": [ + 230, + 226, + 365, + 363 + ], + "type": "image", + "image_path": "c92ae57f7240e51237a6b26fc46628e52ce7f94ace1fa4f0eb70378e13772fda.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 387, + 236, + 401, + 247 + ], + "blocks": [ + { + "bbox": [ + 387, + 236, + 401, + 247 + ], + "lines": [ + { + "bbox": [ + 387, + 236, + 401, + 247 + ], + "spans": [ + { + "bbox": [ + 387, + 236, + 401, + 247 + ], + "type": "image", + "image_path": "b6a1c0f8db182ea61e87b8d41e4d73c375dbbf3563866da63822f81ddea9fffd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 403, + 236, + 418, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 236, + 418, + 245 + ], + "spans": [ + { + "bbox": [ + 403, + 236, + 418, + 245 + ], + "type": "text", + "content": "VFC" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 386, + 251, + 543, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 251, + 543, + 346 + ], + "spans": [ + { + "bbox": [ + 386, + 251, + 543, + 346 + ], + "type": "text", + "content": "A happy little girl is standing in a green field, wearing a plaid shirt and holding onto a string of pink balloons. The balloons are floating in the air, creating a playful and joyful atmosphere. The girl appears to be enjoying the moment, with a smile on her face. In the background, there is a vast expanse of green grass, stretching out as far as the eye can see. The sky above is a bright blue, with a few white clouds scattered about. The overall mood of the scene is one of carefree happiness and joy, with the pink balloons adding a touch of whimsy and playfulness to the atmosphere." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 53, + 398, + 66, + 411 + ], + "blocks": [ + { + "bbox": [ + 53, + 398, + 66, + 411 + ], + "lines": [ + { + "bbox": [ + 53, + 398, + 66, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 66, + 411 + ], + "type": "image", + "image_path": "1894e6b7bc4c70aa520e357cff3ab5b98f24aa9f56cc4b8c7b6150b3ff83db83.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 400, + 91, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 400, + 91, + 409 + ], + "spans": [ + { + "bbox": [ + 69, + 400, + 91, + 409 + ], + "type": "text", + "content": "Cap3D" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 418, + 205, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 205, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 205, + 426 + ], + "type": "text", + "content": "A 3D model of a small wooden tower with a blue roof." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 236, + 376, + 291, + 474 + ], + "blocks": [ + { + "bbox": [ + 236, + 376, + 291, + 474 + ], + "lines": [ + { + "bbox": [ + 236, + 376, + 291, + 474 + ], + "spans": [ + { + "bbox": [ + 236, + 376, + 291, + 474 + ], + "type": "image", + "image_path": "a46745bcb26d4b301c683abf169c5bdf48e792161b55cd4098f42f309579053e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 490, + 547, + 523 + ], + "lines": [ + { + "bbox": [ + 45, + 490, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 45, + 490, + 547, + 523 + ], + "type": "text", + "content": "Figure 1. Comparison of VisualFactChecker (VFC) with GPT-4V and Cap3D. VFC can generate high-fidelity detailed captions that closely match GPT-4V's quality for 2D images and offer significantly more details for 3D objects than Cap3D. VFC used a pre-trained Llama-2 as the LLM when generating the caption for the above 2D image." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 296, + 377, + 375, + 478 + ], + "blocks": [ + { + "bbox": [ + 296, + 377, + 375, + 478 + ], + "lines": [ + { + "bbox": [ + 296, + 377, + 375, + 478 + ], + "spans": [ + { + "bbox": [ + 296, + 377, + 375, + 478 + ], + "type": "image", + "image_path": "339025ec5a80eafef85cc97d595a97ae86ba26380006fef6a4ff7f7d3fdba461.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 389, + 378, + 404, + 389 + ], + "blocks": [ + { + "bbox": [ + 389, + 378, + 404, + 389 + ], + "lines": [ + { + "bbox": [ + 389, + 378, + 404, + 389 + ], + "spans": [ + { + "bbox": [ + 389, + 378, + 404, + 389 + ], + "type": "image", + "image_path": "cbbd59de05aace3209690e0310e496ba880ea8365c98b37a600cdd7f6542c01a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 405, + 380, + 420, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 380, + 420, + 388 + ], + "spans": [ + { + "bbox": [ + 405, + 380, + 420, + 388 + ], + "type": "text", + "content": "VFC" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 389, + 400, + 541, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 400, + 541, + 455 + ], + "spans": [ + { + "bbox": [ + 389, + 400, + 541, + 455 + ], + "type": "text", + "content": "A 3D model of a three-level, intricately designed wooden tower with a contrasting blue roof and door, standing on a wooden base. The tower, brown in color, resembles a fusion of a house, a tower, and a castle. At the very top of the tower, there is a crescent moon design. The overall design adds a touch of fantasy to the scene." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 533, + 192, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 533, + 192, + 547 + ], + "spans": [ + { + "bbox": [ + 143, + 533, + 192, + 547 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 45, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 570, + 289, + 715 + ], + "type": "text", + "content": "Existing automatic captioning methods for visual content face challenges such as lack of detail, content hallucination, and poor instruction following. In this work, we propose VisualFactChecker (VFC), a flexible training-free pipeline that generates high-fidelity and detailed captions for both 2D images and 3D objects. VFC consists of three steps: 1) proposal, where image-to-text captioning models propose multiple initial captions; 2) verification, where a large language model (LLM) utilizes tools such as object detection and VQA models to fact-check proposed captions; 3) captioning, where an LLM generates the final caption by summarizing caption proposals and the fact check verifica" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "tion results. In this step, VFC can flexibly generate captions in various styles following complex instructions. We conduct comprehensive captioning evaluations using four metrics: 1) CLIP-Score for image-text similarity; 2) CLIP-Image-Score for measuring the image-image similarity between the original and the reconstructed image generated by a text-to-image model using the caption. 3) human study on Amazon Mechanical Turk; 4) GPT-4V for fine-grained evaluation. Evaluation results show that VFC outperforms state-of-the-art open-sourced captioning methods for 2D images on the COCO dataset and 3D assets on the Objverse dataset. Our study demonstrates that by combining open-source models into a pipeline, we can attain captioning capability comparable to proprietary models such as GPT-4V, despite being over " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " smaller in model size." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14033" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 305 + ], + "type": "text", + "content": "Image captioning is a pivotal challenge in computer vision and natural language processing. Its central goal is to encapsulate visual data within a textual description, which requires a nuanced understanding of both modalities. The recent advent of multimodal large language models (MM-LLMs), such as GPT-4V [26], and text-to-image generation models, such as DALLE-3 [3], has marked significant progress in this field. These proprietary models could leverage expansive human-labeled data and enormous computing resources to learn to generate detailed and contextually appropriate image descriptions. On the other hand, existing open-sourced captioning methods in the community still face significant challenges. Methods such as BLIP-2 [17] and OFA [35] often yield overly succinct captions that neglect essential visual information. Conversely, systems like Mini-GPT4 [39], InstructBLIP [8], and LLaVA [20, 21] can suffer from hallucination, producing long descriptions that do not align with the actual content of the images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 307, + 286, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 307, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 307, + 286, + 425 + ], + "type": "text", + "content": "In light of this, we propose VisualFactChecker (VFC), a flexible training-free pipeline designed to produce accurate and comprehensive captions for both 2D images and 3D objects. Fig. 1 shows examples of captions generated by VFC and their comparisons with captions generated by GPT-4V [26] and Cap3D [23]. Captions generated by VFC are faithful textural representations of the visual contents. This can also be verified by reconstructing images and 3d objects from captions using state-of-the-art text-to-image and text-to-3d models, as shown in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 426, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 286, + 616 + ], + "type": "text", + "content": "VFC focuses on tackling hallucinations and insufficient details in generated captions and is structured around three core components: Proposer, serving as the system's \"eye\", creating detailed caption proposals as preliminary captions by using image-to-text captioning models; Large Language Model, acting as the \"brain\", calling and summarizing information from other components, and leveraging its advanced generalization capabilities to steer the captioning process following specified captioning instructions; Detector and VQA models, functioning as \"tools\" utilized by the LLM for fact-checking caption proposals, ensuring the fidelity of the final generated caption. VFC is versatile and effectively handles captioning for both 2D images and 3D objects through a unified pipeline. Fig. 3 shows an overview of the pipeline. The details of each component and their interplay are explained in Sec. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 618, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 618, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 618, + 286, + 712 + ], + "type": "text", + "content": "To comprehensively evaluate the generated captions, other than leveraging the commonly used CLIP-Score that primarily gauges the image-caption similarity, we propose a new metric: the CLIP-Image-Score. This metric assesses the similarity between the input image and a reconstructed image created by a text-to-image model from the caption, offering a complementary measure. Furthermore, we conducted a human study on Amazon Mechanical Turk for cap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 131 + ], + "type": "text", + "content": "tion evaluation. Finally, we also performed a fine-grained evaluation by asking GPT-4V to compare and judge captions with detailed reasoning. The combination of CLIP-Score, CLIP-Image-Score, GPT-4V, and human study provides a more robust evaluation of captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "content": "We summarize our main contributions as follows: 1) We propose VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity detailed 2D and 3D captions, effectively mitigating the challenge of hallucination in long captions. (2) CLIP-Image-Score: A novel caption evaluation metric that measures the similarity between the input image and a reconstructed image from the caption. (3) Our evaluation shows that VisualFactChecker achieves state-of-the-art results in 2D and 3D captioning tasks compared with open-sourced models. (4) Our work shows that using an LLM to chain open-source models can achieve captioning capability on par with proprietary models such as GPT-4V." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 288, + 391, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 391, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 391, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 308, + 411, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 411, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 411, + 320 + ], + "type": "text", + "content": "2.1. Image Captioning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 327, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 374 + ], + "type": "text", + "content": "Image captioning has made significant progress with the advent of deep learning. Pioneering works [2, 10, 14] primarily focus on integrating deep neural networks for enhanced image understanding and language generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 375, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 601 + ], + "type": "text", + "content": "Recent strides have been made with the introduction of Multimodal-Large Language Models (MM-LLMs), which are trained on extensive vision and language data. The general approach involves leveraging a pre-trained large language model (LLM) and a vision encoder with a projector to align with the LLM's embeddings, thus enhancing visual understanding. Several models have emerged as significant contributors in this domain. BLIP [16], BLIP2 [17], OFA [35], Flamingo [1], Kosmos-2 [27], MiniGPT4 [39], InstructBLIP [8], LLaVA [20, 21] have demonstrated impressive performance in single-view image captioning tasks. However, they exhibit varying limitations. For instance, BLIP-2 and OFA often generate overly concise captions, while others, like InstructBLIP, can produce detailed captions that often include inaccurate or hallucinatory content. Our method aims to address these limitations by combining different models into a pipeline via an LLM, striking a better balance between accuracy and detailedness in generated captions while mitigating hallucinations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 611, + 513, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 513, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 513, + 624 + ], + "type": "text", + "content": "2.2. Large Language Models for Captioning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 629, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 629, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 629, + 545, + 712 + ], + "type": "text", + "content": "Recent advancements in large language models (LLMs) like GPT-3 [5], LAMDA [30], PALM [7], Llama [32], GPT-4 [26] have demonstrated exceptional zero-shot capabilities in language analysis and summarization tasks. This proficiency has naturally extended to the multimodal domain, particularly in image-language contexts, where LLMs can summarize multimodal information in a zero-shot manner." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "14034" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 72, + 342, + 241 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 342, + 241 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 342, + 241 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 342, + 241 + ], + "type": "image", + "image_path": "43520a7ed5c9835236b65f482fe3cf6d16c53e5fad72b91f335b052ef502c1d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 548, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 548, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 548, + 331 + ], + "type": "text", + "content": "Figure 2. We use DALLE-3 [3] as a text-to-image model to reconstruct 2D images using generated captions from different captioning methods (BLIP-2, LLaVA-1.5 and ours). Similarly, we use MVDream [29] as a text-to-3D model to reconstruct 3D objects using different 3D captions (generated by Cap3D [23] and ours). From the results, we can see that the reconstructed images or 3D objects using BLIP-2 or Cap3D captions are less similar than the input ones, suggesting their captions may not contain sufficient information or incorrectly describe the visual contents; the reconstructed images using LLaVA-1.5 captions contain objects or scenes that are not present in the original images (top: people in the background, bottom: pedestrians and cars on the street), suggesting there might be hallucinations in LLaVA-1.5 captions. Images or 3D objects reconstructed using our captions are more similar to the inputs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 350, + 72, + 539, + 242 + ], + "blocks": [ + { + "bbox": [ + 350, + 72, + 539, + 242 + ], + "lines": [ + { + "bbox": [ + 350, + 72, + 539, + 242 + ], + "spans": [ + { + "bbox": [ + 350, + 72, + 539, + 242 + ], + "type": "image", + "image_path": "c8c65815c524bd46c9b82e60c2f682e30b4bdda81322e45cba8da9fcae37f06a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 448 + ], + "type": "text", + "content": "Vision-blind LLMs are prominent in multimodal applications, often utilizing language-only prefixes generated by pre-trained tools. Clipcap [25] demonstrates this by using a continuous embedding as a prompt for a GPT-style language model, achieving notable performance in single-viewpoint image captioning. Similarly, Promptcap [13] and PNP-VQA [31] leverage natural language prompts with GPT models to excel in visual question answering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 605 + ], + "type": "text", + "content": "Recent methods have employed LLMs to generate image captions by summarizing initial captions or keywords from Vision-Language models. For instance, Socratic models [37] use a CLIP-based model to extract key tags from images, followed by GPT-3 with specialized prompts to create stylized captions. ChatCaptioner [38] builds upon this by integrating ChatGPT and BLIP-2 [17] in a conversational approach for question-answering about the image, and summarizing them into a caption. Visual Clues [36] uses similar tags to generate a paragraph-caption. IC3 [6] and LLM-Fusion [4] use LLMs to summarize captions from existing models augmented with temperature-based sampling. Cap3D [24] extends this concept to 3D object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Our method differentiates itself in two critical ways: First, we focus on reducing hallucinations in captions by employing visual grounding tools, such as object detection, to fact-check captions for enhanced accuracy. Second, our pipeline can be used for captioning both 2D images and 3D objects. Unlike previous methods that rely on a single captioning model, we integrate multiple captioning sources from different models, ensuring a more comprehensive coverage of visual content to generate captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 350, + 457, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 457, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 457, + 361 + ], + "type": "text", + "content": "2.3. Hallucination in MM-LLM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 369, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 548 + ], + "type": "text", + "content": "There are two popular topics on the hallucination of MM-LLMs. (1) Hallucination evaluation: Detection approaches such as Gunjal et al. [11] train classification models to identify hallucination. They focus on distinguishing between accurate and hallucinated content. Ground truth comparison methods [18, 34] compare model outputs with ground truth data to detect hallucinations. These techniques assess the alignment of generated captions with actual image content. (2) Mitigation Strategies [22]: Data optimization methods such as Liu et al. [19] address hallucination by creating negative instances in training datasets to reduce model overconfidence. Iterative generation methods such as Wang et al. [33] adopt an iterative process for caption generation, where brief answers are generated in succession and amalgamated, aiming to improve accuracy and relevance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": "Our VisualFactChecker is a training-free pipeline mitigating hallucination in image captioning. Our method utilizes visual grounding tools for improved accuracy, thereby actively reducing the hallucination and offering high-fidelity captions for both 2D images and 3D objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 621, + 425, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 425, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 425, + 633 + ], + "type": "text", + "content": "3. Visual Fact Checker" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "This section introduces the key components of VisualFactChecker as shown in Fig. 3 in detail and explains their interplay in generating accurate and detailed captions. The following sections delve into specifics. First, we detail the pipeline for 2D image captioning (Sec. 3.1), with Fig. 3 (top) illustrating this process. Then, we explore how our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14035" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 531, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 531, + 194 + ], + "type": "image", + "image_path": "3b92976ffcac2b8757203defdb16f022ca02943023ae194a39cd44644195c39b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 202, + 529, + 323 + ], + "blocks": [ + { + "bbox": [ + 51, + 202, + 529, + 323 + ], + "lines": [ + { + "bbox": [ + 51, + 202, + 529, + 323 + ], + "spans": [ + { + "bbox": [ + 51, + 202, + 529, + 323 + ], + "type": "image", + "image_path": "e25a6dd7bf2f14bcd4d8c693e023bfe146fcd49e9581ff2ea924bb7ee09c90b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 377 + ], + "type": "text", + "content": "Figure 3. Pipeline of the VisualFactChecker for captioning 2D images (top) and 3D objects (bottom). The process begins with the input being captioned by two multimodal captioning models (Captioner-1 and Captioner-2) to generate preliminary captions. These captions are then verified using a Large Language Model (LLM) to call object detection (Detector) and VQA models for fact-checking the captions. Finally, the LLM incorporates all the results and summarizes the final caption by following instructions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 398, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 434 + ], + "type": "text", + "content": "approach is adapted for 3D object captioning as shown in Fig. 3 (bottom), underscoring both shared methodologies and unique aspects relevant to 3D contexts (Sec. 3.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 170, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 170, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 170, + 455 + ], + "type": "text", + "content": "3.1.2D Image Captioning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "content": "The caption generation takes three steps: 1) proposal, 2) verification, and 3) captioning. Each step is detailed below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": "**Proposal:** The Proposal step serves as the cornerstone of the captioning process that generates initial captions. This is achieved through the utilization of advanced image-to-text models, specifically \"LLaVA\" and \"Kosmos2\". These models are trained on expansive datasets, enabling them to comprehend and interpret visual content effectively. By analyzing the input image, they suggest various preliminary captions, each reflecting different facets and interpretations of the image (Fig. 3 top). The rationale behind using multiple image-to-text multimodal LLMs lies in the complexity of adequately capturing an image's essence in a single attempt. Since an image can be accurately described in numerous ways, different models bring unique perspectives, thereby encompassing a broader range of information present in the image. Although the initial captions proposed may not possess perfect fidelity, the primary objective at this stage is to generate captions that are as comprehensive as possible. Fig. 3 displays the specific prompts we used for each step, with more details in Appendix A." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 398, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 673 + ], + "type": "text", + "content": "Verification and Captioning: The goal of the verification step is to scrutinize and rectify any inaccuracies or hallucinations in the captions during the proposal step. It employs a combination of a Large Language Model (LLM) and grounding tools, including an open-vocabulary object detection model and/or a visual question answering (VQA) model. Here the LLM can be GPT-4 or Llama2. As shown in Fig. 3 (top), the process involves the following steps: Step 1: LLM first summarizes the initial detailed descriptions from different MM-LLMs into a single, detailed caption. While this caption is comprehensive, it may not always be accurate. Step 2: The LLM then analyzes this synthesized caption, identifying all objects that could be verified by object detection and summarizing an object checklist. In 2D image captioning, the focus is on eliminating hallucinations, particularly descriptions of non-existent objects in the image. Identifying these objects is crucial for the subsequent verification process. Step 3: Taking the object checklist as input, an open-vocabulary object detection model examines candidate objects in the checklist and determines their presence in the image. This step is pivotal in validating the existence of objects mentioned in the caption, thus supporting the fidelity of the caption." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "After verification, we go to the last captioning step: Based on the object detection results, the LLM revises the summarized single detailed caption. Each object described" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14036" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "in the caption is cross-checked; if detected, it remains unchanged, while undetected objects are considered potential hallucinations and are removed from the caption. This step results in a final caption that is both detailed and reliable. The underlying assumption is that the detection model, serving as an object grounding expert, provides more reliable results than a general image descriptor." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 290, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 290, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 290, + 350 + ], + "type": "text", + "content": "In the verification and captioning steps, the LLM plays a pivotal role as a \"brain\". It starts by parsing the initial caption and identifying key objects for detailed examination. The LLM then meticulously assesses whether each object mentioned actually appears in the image based on detection results. Following this thorough analysis, it refines and revises the initial captions, transforming them into final versions that are both coherent and richly detailed. The LLM is instrumental in guaranteeing linguistic fluency, ensuring that the captions not only accurately represent the image but also maintain the necessary level of detail for high-fidelity captioning. Moreover, the LLM can follow complex instructions to write the captions in a specified style, such as a caption that only mentions the foreground objects without mentioning the background. Fig. 3 displays the specific prompts used for each step." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 359, + 174, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 174, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 174, + 372 + ], + "type": "text", + "content": "3.2. 3D Object Captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 715 + ], + "type": "text", + "content": "The 3D object captioning process follows a similar structural pipeline to that of 2D images, with a few key distinctions in certain steps, as depicted in Fig. 3 (bottom). In 3D captioning, an object may present multiple views, each offering unique information. The comprehensive caption for a 3D object is derived by integrating the perspectives from all these views. For each view, VisualFactChecker is employed to create a detailed, high-fidelity description. Subsequently, the LLM (GPT-4 or Llama-2) is used to amalgamate the information from all views, producing a unified caption for the 3D object. In particular, for each view's captioning, we have the same three-step approach akin to 2D image captioning. In the proposal step, LLaVA-1.5 and InstructBLIP are utilized for generating initial detailed descriptions. We opt out of using Kosmos2 for single 3D objects due to its less effective performance in providing detailed descriptions, possibly linked to its reliance on an implicit detection model. Additionally, a slightly modified prompt is used (see Fig. 3 bottom), which incorporates 3D-specific considerations. In the verification and captioning step, we primarily address hallucinations related to the attributes of 3D objects, such as shape and color. To mitigate these inaccuracies, rather than enumerating potential objects, we employ the LLM to generate five critical questions that could influence a text-to-3D generation model in reconstructing the 3D model. Following this, we utilize VQA models (specifically LLaVA-1.5) to respond to these questions based on the input 3D object view image. Subsequently, the LLM amends" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 541, + 153 + ], + "type": "image", + "image_path": "9ef51dbb1039ec8cab64168025cbc7ce14c2160531aef1c3d93f35d131460dfe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": "Figure 4. The CLIP-Image-Score pipeline evaluates caption accuracy by encoding an original image " + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": " into a feature representation " + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "inline_equation", + "content": "I_{X}" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": " using a CLIP image encoder. A captioning model generates a caption that is then input into a text-to-image model to reconstruct an image " + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": ", which is encoded to " + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "inline_equation", + "content": "I_{X'}" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": ". The score is computed by assessing the cosine similarity between " + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "inline_equation", + "content": "I_{X}" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "inline_equation", + "content": "I_{X'}" + }, + { + "bbox": [ + 305, + 160, + 547, + 239 + ], + "type": "text", + "content": ", providing a measure of the caption's fidelity and hallucination detection." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 260, + 546, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 546, + 357 + ], + "type": "text", + "content": "the initial caption in accordance with the answers provided by the VQA model. We operate under the assumption that answering targeted questions results in fewer hallucinations compared to generating a general description. Once the caption for each individual view is complete, the LLM synthesizes these multiple perspectives into a singular, comprehensive caption for the entire 3D object. The prompts used for the LLM at each stage are detailed in Appendix A." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 369, + 419, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 419, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 419, + 383 + ], + "type": "text", + "content": "4. CLIP-Image-Score" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "type": "text", + "content": "Accurate evaluation of caption correctness and detailedness is paramount in determining the performance of an image captioning model. Traditional metrics like the CLIP-Score [12] have served as a standard for measuring the alignment between generated captions and their corresponding images. However, our CLIP-score may lack the sensitivity needed to detect the specific issue of hallucination within captions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "We present the CLIP-Image-Score, an alternative metric specifically developed to reflect the subtleties of caption quality. This metric is different from CLIP-Score by introducing an additional reconstruction step. Specifically, the CLIP-Image-score evaluates the similarity between the original image and a reconstructed version of the image generated by a fixed text-to-image model using the caption as a prompt. By comparing the raw image to its reconstructed image, the metric is able to detect discrepancies indicative of hallucination, thus providing a different perspective of the caption quality assessment. The underlying principle of the CLIP-Image-Score is the recognition that multiple \"correct\" captions may exist for a single image. However, it's only when a caption is both \"detail\" and \"correct\" that the reconstructed image closely resembles the original. Moreover, any hallucinations present in the caption become evident in the reconstructed image. Fig. 2 presents examples of such reconstructions. For instance, consider the results from LLaVA-1.5 shown in the third column. The cap" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14037" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "tion generated for the first image falsely mentions \"several other people in the background\". This error is clearly reflected in the image reconstructed by the text-to-image generator. In essence, comparing the two images indirectly ensures alignment between the image and its caption, thereby providing a complementary method to assess the quality of the caption than directly comparing the image and caption." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 179 + ], + "type": "text", + "content": "The CLIP-Image-Score evaluation process is depicted in the following steps:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 180, + 287, + 382 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 180, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 286, + 203 + ], + "type": "text", + "content": "- Caption Generation: An original image " + }, + { + "bbox": [ + 47, + 180, + 286, + 203 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 180, + 286, + 203 + ], + "type": "text", + "content": " is input into a captioning model, which generates a caption." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 287, + 251 + ], + "type": "text", + "content": "- Caption-to-Image Reconstruction: This generated caption is then used as input for a text-to-image model, which creates a reconstructed image " + }, + { + "bbox": [ + 47, + 204, + 287, + 251 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 47, + 204, + 287, + 251 + ], + "type": "text", + "content": " that visually represents the textual description." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "type": "text", + "content": "- Raw Image Encoding: The original image " + }, + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "type": "text", + "content": " is processed through a CLIP image encoder, translating the visual content into an encoded representation " + }, + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "type": "inline_equation", + "content": "I_{X}" + }, + { + "bbox": [ + 47, + 251, + 286, + 286 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 288, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 286, + 323 + ], + "type": "text", + "content": "- Reconstructed Image Encoding: The reconstructed image is also processed through the CLIP image encoder to obtain its encoded representation " + }, + { + "bbox": [ + 47, + 288, + 286, + 323 + ], + "type": "inline_equation", + "content": "I_{X'}" + }, + { + "bbox": [ + 47, + 288, + 286, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "type": "text", + "content": "- Score Calculation: Finally, the encoded representations of the original and reconstructed images are compared to calculate the CLIP-Image-Score. The score is given by the cosine similarity, which assesses the congruence between " + }, + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "type": "inline_equation", + "content": "I_{X}" + }, + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "type": "inline_equation", + "content": "I_{X'}" + }, + { + "bbox": [ + 47, + 323, + 286, + 382 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 94, + 388, + 287, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 388, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 94, + 388, + 287, + 415 + ], + "type": "interline_equation", + "content": "\\text {C L I P - I m a g e - S c o r e} = \\frac {I _ {X} \\cdot I _ {X ^ {\\prime}}}{\\| I _ {X} \\| \\times \\| I _ {X ^ {\\prime}} \\|} \\tag {1}", + "image_path": "eb35b7e7bdde0906c6082d2cf07259989b89aa927998467450b4793a4a2793d0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 420, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 516 + ], + "type": "text", + "content": "Most notably, CLIP-Image-Score offers a sensitive measure for detecting hallucinations. In scenarios where the generated caption includes elements that are not in the original image, the reconstructed image will also likely contain these discrepancies. By comparing the original and reconstructed images, the CLIP-Image-Score can effectively highlight these differences, offering a clearer insight into the fidelity and accuracy of the generated caption." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 516, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 635 + ], + "type": "text", + "content": "Furthermore, CLIP-Image-Score turns a cross-modality comparison into a more intuitive comparison in the same image modality (as shown in Fig. 4). CLIP-Image-Score represents a new complementary perspective for image captioning evaluation. By leveraging the capabilities of text-to-image models and focusing on the congruence between the original and reconstructed images, it provides an accurate assessment of caption quality, particularly in identifying and measuring hallucinations, thereby enhancing the overall reliability of caption generation systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 646, + 127, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 127, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 127, + 658 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "This section presents a thorough evaluation of captioning models across both 2D and 3D visual content, employing a variety of datasets and methodologies. Table 1 provides a summary of our comprehensive evaluation experiments." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "type": "table", + "html": "
EvalInput pairs for evaluationMethodReference
2DRaw imageCaptionCLIP-ScoreTable 2
Human evaluationFig. 6
CPT4V evaluationFig. 7
Raw imageImage(recon)CLIP-Image-ScoreTable 2
3DMulti-view (raw)CaptionCLIP-ScoreTable 3
GPT4V evaluationFig. 7
Multi-view (raw)Multi-view (recon)CLIP-Image-ScoreTable 3
", + "image_path": "25026c76475dbd41210b740ea177cf6c854a45799a93813d513cf4c86faa448c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 180, + 538, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 538, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 538, + 193 + ], + "type": "text", + "content": "5.1. Overall: CLIP-Score and CLIP-Image-Score" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 199, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 545, + 462 + ], + "type": "text", + "content": "2D image captioning. Dataset: Our evaluation utilized 5,000 COCO test images from the Karpathy split. Baseline methods: We benchmarked against state-of-the-art captioning models, including BLIP-2 [17], InstructBLIP [8], and LLaVA-1.5 [20]. The evaluation focused on each model's ability to produce accurate, detailed, and coherent captions that effectively encapsulate the essence of the images. Evaluation Metric: We employed two metrics: CLIP-Score [12] and CLIP-Image-Score (Sec. 4). The CLIP-Score, a prevalent metric in image caption quality assessment, involves processing the raw image through the CLIP image encoder and the caption through the CLIP text encoder. The resultant embeddings are then compared for cosine similarity, with a higher score indicating greater semantic resemblance between the image and the caption. For our analysis, we first calculated the CLIP-Score for each image_caption pair, then averaged these scores across all 50,000 text/image pairs, scaling the result by a factor of 100. Table 2 displays the comparative performance of various image captioning methods on the 5,000 COCO test set images. The results demonstrate that our VisualFactChecker surpasses all baseline methods in performance." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 308, + 472, + 544, + 539 + ], + "blocks": [ + { + "bbox": [ + 328, + 148, + 522, + 159 + ], + "lines": [ + { + "bbox": [ + 328, + 148, + 522, + 159 + ], + "spans": [ + { + "bbox": [ + 328, + 148, + 522, + 159 + ], + "type": "text", + "content": "Table 1. Summary of evaluation methods and results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 472, + 544, + 539 + ], + "lines": [ + { + "bbox": [ + 308, + 472, + 544, + 539 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 544, + 539 + ], + "type": "table", + "html": "
Captioning MethodCLIP-Score (%) ↑CLIP-Image-Score (%) ↑
Human Label (COCO GT)30.36 (-2.54)71.21 (-2.40)
BLIP230.11 (-2.79)70.79 (-2.82)
InstructBLIP31.45 (-1.45)72.95 (-0.66)
LLaVA-1.532.08 (-0.82)73.24 (-0.37)
Kosmos-232.32 (-0.58)73.28 (-0.33)
VisualFactChecker (Ours)32.9073.61
", + "image_path": "7ce006f99f9e04518dd5a4ad7a63fcfedfe482c4e79fc49cd4fc7ba61ed6d2df.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 547, + 545, + 581 + ], + "lines": [ + { + "bbox": [ + 305, + 547, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 547, + 545, + 581 + ], + "type": "text", + "content": "Table 2. Image captioning comparison with different metrics on 5000 COCO test set in Karpathy split, we use raw image and caption as input pairs for evaluation." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "As outlined in Sec. 4, the CLIP-Image-Score provides a complementary view to assess the quality of image captions. This metric is derived by comparing the cosine similarity between the CLIP embeddings of two images: the original image and a reconstructed image, which is generated using the provided caption through a text-to-image generation model. A higher CLIP-Image-Score signifies a more accurate and effective image caption. For this process, Stable Diffusion XL (SDXL) [28] is utilized as the designated text-to-image model to reconstruct images based on" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14038" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 286, + 102 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 102 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 102 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 102 + ], + "type": "table", + "html": "
Captioning MethodCLIP-Score (%)↑CLIP-Image-Score (%)↑
Cap3D33.44 (-0.57)79.88 (-0.44)
VisualFactChecker (Ours)34.0180.32
", + "image_path": "a43fcd50ba640e0ae28cf138ad87081ab69bd4f561da931ac2e875dda75fd5f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 110, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 177 + ], + "type": "text", + "content": "Table 3. 3D object captioning comparison with different metrics on 1000 objects in Objaverse. For CLIP-Score, we use the average score of two views for evaluation. For CLIP-Image-Score, we use an off-the-shelf text-to-3D model, MVDream, to generate 3D models from 3D captions. We compare two views of the raw object and the same views of generated 3D object for evaluation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "type": "text", + "content": "the generated captions. Table 2 presents the CLIP-Image-Scores obtained for the 5000 images in the COCO test set, where our method outperforms all baseline methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 424 + ], + "type": "text", + "content": "3D object captioning. Dataset: 1,000 3D objects sampled from Objaverse dataset [9]. Baseline methods: We use state-of-the-art 3D object captioning model Cap3D [23] as the baseline. Cap3D uses 8 view images to generate the final object caption, our VisualFactChecker uses only 2 views to generate the object caption. Evaluation Metric: CLIP-Score and CLIP-Image-Score on multiple views rendered from 3D objects. To evaluate the similarity of a 3D object and the generated caption, we evaluate the similarity of the caption with the multi-view images used to generate the caption. Specifically, we evaluate the similarity of the generated caption with the two views that were used to generate the caption and use the average score to represent the CLIP-Score. Table. 3 shows the performance of 3D object captioning methods on 1,000 3D objects from Objaverse dataset. VisualFactChecker outperforms Cap3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": "We also use CLIP-Image-Score to evaluate the 3D caption quality. CLIP-Image-Score needs reconstructed images to compare with the raw images. We treat the two views that were used to generate the 3D object caption as the raw image. To obtain the reconstructed image, we use an off-the-shelf text-to-3D generation model, MVDream, to generate a 3D object given the generated 3D object caption. We then render the same two views of images based on the generated 3D object, and we calculate the CLIP-Image-Score between the raw image and the rendered image. Table. 3 shows the CLIP-Image-Score on 1000 objects in Objaverse dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 239, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 239, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 239, + 576 + ], + "type": "text", + "content": "5.2. Per Image Evaluation: Wining Rate" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 700 + ], + "type": "text", + "content": "CLIP-Score and CLIP-Image-Score indicate an overall performance comparison, which shows an average score among all 5000 images. The average score may be dominated by a small group of images that have extremely high or low scores. To zoom in and show a more detailed comparison, we try to answer the following question: Given an image, what is the probability that one method performs better than another method on caption generation? To answer this question, we need to go over each image and calculate the winning rate for a pair of methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Specifically, for each image, we compare the CLIP-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 542, + 141 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 141 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 141 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 141 + ], + "type": "image", + "image_path": "a124c0833c55a0b0af162c9a7fe2e88e4508b82c4185a3077a20e79a7382f589.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 148, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 309, + 148, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 309, + 148, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 309, + 148, + 542, + 217 + ], + "type": "image", + "image_path": "7d81d81730484ac0fabc42c59d14555aca14fedf801aa2c86d30fcbc152f653c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 228, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 262 + ], + "type": "text", + "content": "Figure 5. 2D image captioning comparison with pair-wise winning rate. VisualFactChecker (VFC) outperforms all baseline methods on both CLIP-Score (top) and CLIP-Image-Score (bottom)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": "Score of our VisualFactChecker caption against the captions generated from different baselines respectively, and calculate the wining probability of our method and the baselines. Fig. 5 shows the results, for example, we can see that in the pair-wise comparison, our VisualFactChecker performs better (higher CLIP-Score) than LLaVA-1.5 on " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "64.4\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": " of 5000 images (3220 images)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 413 + ], + "type": "text", + "content": "Calculating the winning rate over all images provides a more detailed analysis that zooms in on the comparison of each image, which shows a complementary view than overall average CLIP-Score." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 420, + 545, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 433 + ], + "type": "text", + "content": "5.3. Fine-grained Evaluation: Human and GPT-4V" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "content": "The CLIP-Score and CLIP-Image-Score offer a general comparison of overall performance. A pairwise per-image winning rate provides a more specific analysis, evaluating performance on individual images. However, the research highlighted in related studies [15] indicates that the CLIP-Score may not be ideally suited for image-to-image comparison tasks. Furthermore, relying on a single score fails to provide a nuanced comparison across criteria, such as accuracy and level of detail. We use Human evaluation and GPT-4V to provide a more fine-grained evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Human evaluation using Amazon Mechanical Turk (AMT). We employed a pairwise comparison strategy. From the COCO dataset, we randomly selected 100 images out of 5000. For each image, our caption was compared against 5 baseline captions respectively. To reduce variance, each comparison was done by 3 different AMT workers and we used their majority voting as the final selection. This resulted in a total of 1500 comparisons collected on AMT. AMT UI is shown in the appendix. The workers were presented with two competing captions — one from a baseline method and one from our VisualFactChecker, in randomized order. They were instructed to select the better caption describing the image based on 3 aspects: correctness, de" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14039" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "tailness, and fluency. Results in Fig. 6 show our captions are more preferred by humans. The human evaluation instruction and web UI is shown in Appendix B." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 119, + 284, + 190 + ], + "blocks": [ + { + "bbox": [ + 50, + 119, + 284, + 190 + ], + "lines": [ + { + "bbox": [ + 50, + 119, + 284, + 190 + ], + "spans": [ + { + "bbox": [ + 50, + 119, + 284, + 190 + ], + "type": "image", + "image_path": "c4ad4af8a681f8313aac73b805ea572e75585a1e3de13cafb537c815675f431f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 200, + 279, + 211 + ], + "lines": [ + { + "bbox": [ + 53, + 200, + 279, + 211 + ], + "spans": [ + { + "bbox": [ + 53, + 200, + 279, + 211 + ], + "type": "text", + "content": "Figure 6. Amazon Mechanical Turk human evaluation results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 224, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 289, + 441 + ], + "type": "text", + "content": "GPT-4V evaluation. Our study applied GPT-4V for evaluating captions in a manner akin to the caption evaluation process used in DALLE-3. We use the same randomly selected 100 images from COCO as in Human evaluation. For each image, we considered the captions generated by 5 baseline methods alongside the caption produced by our VisualFactChecker. We then presented GPT-4V with the raw image, our reference caption, and the four baseline captions. Our designed prompt instructed GPT-4V to compare each baseline caption against our reference caption, focusing on two primary aspects: correctness and detail. GPT-4V was tasked with providing a pairwise, detailed comparison for each pair, including justifications for its assessments. Based on these comparative insights, GPT-4V classified each baseline method caption as either \"better\" or \"worse\" than our VisualFactChecker. Fig. 5 shows the comprehensive results. More details about the GPT-4V evaluation prompt and examples are shown in Appendix B." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 450, + 284, + 521 + ], + "blocks": [ + { + "bbox": [ + 50, + 450, + 284, + 521 + ], + "lines": [ + { + "bbox": [ + 50, + 450, + 284, + 521 + ], + "spans": [ + { + "bbox": [ + 50, + 450, + 284, + 521 + ], + "type": "image", + "image_path": "0a7b94bc4a4cfa9b509422ca41fabccc66307afa9d8ab8bc6273f3f7be418a74.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "lines": [ + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "text", + "content": "Figure 7. GPT-4V evaluation results. Our captions are significantly better than baselines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 141, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 141, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 141, + 588 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 289, + 714 + ], + "type": "text", + "content": "In our ablation study, we explore the impact of various components on performance. For 2D captioning tasks, we assess the efficacy of initial captioning models, LLaVA-1.5 and Kosmos-2, using the CLIP-Score metric for the captions they generate on the same 5000 COCO test images. Additionally, we ablate our method's performance in the absence of the verification (fact checker) step, which aims to mitigate hallucinations through detection grounding. Table 4 shows the detailed results. Likewise, in the context of 3D object captioning, we evaluate the individual contribu" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 182 + ], + "type": "table", + "html": "
Methods or StepsCLIP-Score
2DLLaVA-1.532.08 (-0.33)
Kosmos-232.32 (-0.09)
VisualFactChecker (w/o fact check)32.41
VisualFactChecker32.90 (+0.49)
3DLLaVA-1.532.05 (-0.66)
InstructBLIP32.51 (-0.20)
VisualFactChecker (w/o fact check)32.71
VisualFactChecker34.01 (+1.30)
", + "image_path": "17ba8fc4ca5ec2978648e38d168aeea0de131f09f2c14ba5ce22b7dbbf234bcf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 189, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 212 + ], + "type": "text", + "content": "Table 4. Ablation study on captioning 2D images (5000 COCO test dataset) and 3D objects (1000 Objaverse)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 232, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 317 + ], + "type": "text", + "content": "tions of initial captioners, namely LLaVA-1.5 and InstructBLIP on the same 1000 Objaverse 3D objects. We further investigate the performance of our methodology without the fact checker, which in this case operates by leveraging a VQA model to reduce hallucinations. Table 4 shows the detailed results. These results highlight the significance of fact checker in our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 324, + 525, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 525, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 525, + 338 + ], + "type": "text", + "content": "5.5. Qualitative Results and Prompt Following" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 378 + ], + "type": "text", + "content": "Other than quantitative evaluation results, we show more qualitative examples of VisualFactChecker for 2D and 3D captions in Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 378, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 415 + ], + "type": "text", + "content": "By leveraging an LLM, VisualFactChecker can follow complex instructions to write captions in various styles. Examples are shown in Appendix D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 648 + ], + "type": "text", + "content": "We propose the VisualFactChecker (VFC), a training-free pipeline to generate high-fidelity and detailed captions. By utilizing an LLM to chain multimodal models and object detection and VQA models, VFC reduces hallucination in long captions. We conducted a comprehensive caption evaluation using different metrics, including 1) image-text similarity using CLIP-Score, 2) image-reconstructed image similarity using our proposed CLIP-Image-Score, 3) human study, and 4) fine-grained evaluation using GPT-4V. Compared with open-sourced captioning models, our method achieves state-of-the-art in both 2D and 3D captioning. Our work shows combining open-sourced models into a pipeline can significantly close the captioning performance gap with proprietary models like GPT-4V. In the future, we plan to improve our pipeline further by including more components for fact-checking and making it more automatic in deciding which components to use." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 648, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 648, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 648, + 545, + 708 + ], + "type": "text", + "content": "Acknowledgments We would like to thank Siddharth Gururani for helping with our human evaluation using Amazon Mechanical Turk; Haochen Wang for his help in preprocessing 3D data. We also thank Qinsheng Zhang, Yogesh Balaji, and Yen-Chen Lin for their helpful discussion." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14040" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 247 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. Technical report, OpenAI, 2023. 2, 3, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "text", + "content": "[4] Simone Bianco, Luigi Celona, Marco Donzella, and Paolo Napoletano. Improving image captioning descriptiveness by ranking and llm-based fusion. arXiv preprint arXiv:2306.11593, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "text", + "content": "[6] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. Ic3: Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "type": "text", + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "type": "text", + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 535 + ], + "type": "text", + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[10] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "text", + "content": "[11] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[13] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In ICCV, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[15] Klemen Kotar, Stephen Tian, Hong-Xing Yu, Daniel LK Yamins, and Jiajun Wu. Are these the same apple? comparing images based on object intrinsics. arXiv preprint arXiv:2311.00750, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "type": "text", + "content": "[18] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 285, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 330 + ], + "type": "text", + "content": "[19] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "text", + "content": "[20] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[22] Jiaying Lu, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Baochen Sun, Carl Yang, and Jie Yang. Evaluation and mitigation of agnosia in multimodal large language models. arXiv preprint arXiv:2309.04041, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 443, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 476 + ], + "type": "text", + "content": "[23] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. In NeurIPS, 2023. 2, 3, 7, 17" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[24] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "text", + "content": "[25] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 544, + 476, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 476, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 476, + 556 + ], + "type": "text", + "content": "[26] OpenAI. Gpt-4 technical report, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[27] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "type": "text", + "content": "[29] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 17" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[30] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14041" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 564 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[31] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven CH Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[33] Bin Wang, Fan Wu, Xiao Han, Jiahui Peng, Huaping Zhong, Pan Zhang, Xiaoyi Dong, Weijia Li, Wei Li, Jiaqi Wang, et al. Vigo: Visual instruction generation and correction. arXiv preprint arXiv:2308.12714, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "text", + "content": "[34] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "type": "text", + "content": "[35] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "text", + "content": "[36] Yujia Xie, Luowei Zhou, Xiyang Dai, Lu Yuan, Nguyen Bach, Ce Liu, and Michael Zeng. Visual clues: Bridging vision and language foundations for image paragraph captioning. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 464 + ], + "type": "text", + "content": "[37] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 465, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 518 + ], + "type": "text", + "content": "[38] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "type": "text", + "content": "[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14042" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_content_list.json b/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b31e0422be03201f788d89077c280e8bd62fe251 --- /dev/null +++ b/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_content_list.json @@ -0,0 +1,1410 @@ +[ + { + "type": "text", + "text": "Visual In-Context Prompting", + "text_level": 1, + "bbox": [ + 336, + 131, + 633, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng Li\\*, Qing Jiang#, Hao Zhang#, Tianhe Ren†, Shilong Liu#, Xueyan Zou#, Huaizhe Xu#, Hongyang Li#, Jianwei Yang‡1, Chunyuan Li#, Lei Zhang†2, Jianfeng Gao‡2", + "bbox": [ + 91, + 179, + 892, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\spadesuit$ HKUST $\\ddagger$ Microsoft Research, Redmond $\\dagger$ IDEA $\\sharp$ SCUT $\\sharp$ Tsinghua $\\S$ UW-Madison", + "bbox": [ + 228, + 224, + 735, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Project Lead 2. Equal Advisory Contribution", + "bbox": [ + 364, + 244, + 596, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59decd3dcd4ec115971bd04324fc4f2e92255dd3f0472c963e28a3be2da2ef44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 276, + 849, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Prompting Generic Segmentation", + "text_level": 1, + "bbox": [ + 321, + 375, + 643, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b0a5accbb7a6116badfa51cef6ddaeca37bb024a2f463b5f6bd14c33cc10722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 398, + 846, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Prompting Referring Segmentation", + "text_level": 1, + "bbox": [ + 313, + 488, + 651, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6729c19fb7a6a9022f4850f2626c775613f7a51ad499e2b1327bbef7baccb160.jpg", + "image_caption": [ + "Figure 1. Our model $DINOV$ supports generic and referring segmentation to associate multiple or single objects with the user input visual prompts. A user can input one or more in-context visual prompts (scribbles, masks, boxes, etc.) to improve the segmentation performance." + ], + "image_footnote": [], + "bbox": [ + 122, + 516, + 841, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 670, + 313, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In-context prompting in large language models (LLMs) has become a prevalent approach to improve zero-shot capabilities, but this idea is less explored in the vision domain. Existing visual prompting methods focus on referring segmentation to segment the most relevant object, falling short of addressing many generic vision tasks like open-set segmentation and detection. In this paper, we introduce a universal visual in-context prompting framework for both tasks, as shown in Fig. 1. In particular, we build on top of an encoder-decoder architecture, and develop a versatile prompt encoder to support a variety of prompts like strokes, boxes, and points. We further enhance it to take an arbitrary number of reference image", + "bbox": [ + 73, + 704, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segments as the context. Our extensive explorations show that the proposed visual in-context prompting elicits extraordinary referring and generic segmentation capabilities to refer and detect, yielding competitive performance to close-set in-domain datasets and showing promising results on many open-set segmentation datasets. By joint training on COCO and SA-1B, DINOv achieves 57.7 PQ on COCO and 23.2 PQ on ADE20K. Code will be available at https://github.com/UX-Decoder/DINOv", + "bbox": [ + 496, + 671, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 842, + 632, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent progress in large language models (LLMs) like GPT [1, 27] has shown promising results towards arti", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12861", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ficial general intelligence (AGI) by training unified models on large amounts of text data. These giant LLMs manifest themselves with intriguing emerging capabilities such as in-context learning. Nevertheless, similar paradigms have not yet succeeded in solving all vision tasks due to the diversity of scenarios in computer vision [15]. Some works [23, 49] have combined LLMs and vision models to tackle complex image understanding tasks with text outputs such as visual question answering, but challenges remain in fine-grained tasks that require pixel-level outputs, like instance masks, rather than just text.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, the community has observed a growing interest in the development of language-enhanced vision foundation models. These models demonstrate profound competencies in open-world visual understanding tasks using text prompts, encompassing areas like open-set detection [19, 24, 44] and segmentation [38, 42, 44, 50]. Visual prompt, a different prompting mechanism has been explored in some recent segmentation models [13, 17, 51]. In these works, different visual prompting formats (e.g., points, boxes and strokes, etc) have been explored to facilitate the segmentation of visual contents specified by users.", + "bbox": [ + 75, + 260, + 472, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another critical LLM technique, in-context learning, has been less explored. In-context learning specifies the new task instruction using examples, and allows models to adapt to new tasks or domains without explicit retraining by providing relevant examples. One pioneering work in this area is SegGPT [34], which demonstrates the ability to output an image mask based on visual examples. However, these works focus on associating a user visual prompt with one most relevant object and have the limited ability to identify multiple objects of the same semantic concept. Therefore, these approaches fall short of addressing many generic vision tasks like open-set object detection and segmentation, which often require the segmentation of multiple objects of a given concept. On the other hand, while text-prompted vision models [24, 50] do not align with in-context learning methodologies, they exhibit notable flexibility in managing both referring and generic tasks in detection or segmentation, but with language prompts. In this paper, we strive to develop a model that supports visual in-context prompting for all types of image segmentation tasks. A comparison between our work and previous work is shown in Fig. 2. Besides supporting both single-image and cross-image visual prompting, our model distinguishes itself by effectively handling both referring and generic segmentation problems.", + "bbox": [ + 75, + 429, + 472, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this goal, we build a model called DINOv to support versatile visual prompting capabilities, based on the unified detection and segmentation model MaskDINO [18]. DINOv follows the general encoder-decoder design with an extra prompt encoder to formulate and sample visual prompts. The decoder takes in segmentation queries and reference prompt queries to generate segmentation masks", + "bbox": [ + 75, + 795, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc9e0efdce78726f2416c72899fea74c1531f42c5e4f431a93b6fea73ba21bd0.jpg", + "image_caption": [ + "Figure 2. Comparison with related works. Generic: segment all objects of the same semantic concept that match the user prompt. Refer: segment a particular object with the user input visual prompts. Image prompt: crop the image regions as prompts. (single) Visual prompt: one image-prompt example to segment. In-context prompt: one or multiple image-prompt examples. We can do single-image and cross-image visual prompting tasks and support referring and generic segmentation." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 888, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and target visual prompts, and we associate the output segmentation masks with the target prompt queries for the final output. We can define the visual in-context samples with a set of reference image $(Q)$ - visual prompt $(A)$ pairs. The visual prompt can be in various types, including mask, scribble, box, etc. With the in-context examples, our model takes in a target image and outputs the masks. The creation of target visual prompts involves a prompt encoder that extracts reference visual prompts from a Q-A pair. This is followed by a decoder to get the target visual prompt by attending reference visual prompts to the target image. During training, to construct positive and negative samples for generic segmentation, we sample reference visual prompts in a batch across different images. To address task and data discrepancies, we formulate generic latent queries and point queries for generic and referring segmentation, respectively. By joint training on COCO [21] and SA-1B [13] for generic and referring segmentation, our model attains competitive performance on in-domain segmentation tasks compared with text-prompted models and shows promising generalization capability on a wide range of open-set segmentation benchmarks using purely visual prompts.", + "bbox": [ + 496, + 488, + 890, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are threefold:", + "bbox": [ + 517, + 821, + 826, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We are the first to extend visual in-context prompting to support generic vision tasks like open-set generic segmentation and detection, and achieve comparable performance with text prompt-based open-set models.", + "bbox": [ + 517, + 839, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12862", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6c209b8377017c0adc0031b895dcb1ddf31ece7417925a88f798fc4639b7d47c.jpg", + "image_caption": [ + "(a) General Framework" + ], + "image_footnote": [], + "bbox": [ + 83, + 95, + 457, + 318 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/913b8314b27070f521cadf0a1098ce61fa718bd2cadf94a8607f1077c1bb0e02.jpg", + "image_caption": [ + "(b) Visual Generic Segmentation", + "(c) Visual Referring Segmentation", + "Figure 3. DINov is a universal segmentation framework that can do generic segmentation and referring image segmentation. The vision encoder is used to extract image features.(b) An illustration of losses for visual generic segmentation. In the example, there are 6 visual prompts sampled from 6 masks from 3 categories. The visual prompts from the instances of the same class are averaged as the class embedding. Each column of the matching matrix is a 3-dimension one-hot vector which is a one-hot class label of the instance; (c) An illustration of losses for visual referring segmentation. Each visual prompt is classified to one of the 6 instances." + ], + "image_footnote": [], + "bbox": [ + 472, + 92, + 890, + 309 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We build $DINOV$ , a unified framework for referring segmentation and generic segmentation based on visual in-context prompting. This unification simplifies model design and allows our model to consume both semantically-labelled and unlabelled data for better performance.", + "- We conduct extensive experiments and visualizations to show that our model can handle generic, referring, and video object segmentation tasks. Our early attempts exhibit promising results on open-set segmentation and detection with visual prompting." + ], + "bbox": [ + 94, + 460, + 468, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 76, + 638, + 168, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Unified Formulation for Segmentation Tasks", + "text_level": 1, + "bbox": [ + 76, + 664, + 452, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we concentrate on visual prompting tasks involving images, encompassing both generic segmentation and referring segmentation tasks. Given $N$ reference images $\\mathcal{I} = \\{\\mathbf{I}_1,\\dots,\\mathbf{I}_N\\} \\in \\mathcal{R}^{N\\times H\\times W\\times 3}$ with the corresponding visual prompts $\\mathcal{P} = \\{p_1,\\dots,p_N\\}$ , DINOv aims to segment objects of interest on a new target image $\\mathbf{I}_{\\mathbf{t}}$ . The visual prompts include masks, boxes, scribbles, points, etc. The interested objects can be a particular object for referring segmentation or all objects of the same semantic concept for generic segmentation. Note that the reference image can be identical to the target image, in which the task reduces to single-image visual prompting segmentation.", + "bbox": [ + 75, + 688, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these tasks, DINOv utilizes a comprehensive query-based encoder-decoder architecture. This archi-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tecture comprises a vision encoder, denoted as Enc, responsible for extracting image features, a prompt encoder referred to as PromptEncoder, designed to extract visual prompt features by combining image features and user-provided visual prompts, and a general decoder represented as Decoder, which generates masks and visual concepts based on the segmentation query and visual prompt features. Upon receiving the input image and user-provided visual prompts, our initial step involves extracting image features denoted as Z using the vision encoder. Subsequently, we feed both the image features and visual prompts into the prompt encoder to extract the reference visual prompt $\\mathcal{F}$ and subsequently sample the query visual prompt features $\\mathbf{Q}_{\\mathfrak{p}}$ . Formally, we have:", + "bbox": [ + 496, + 441, + 893, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {Z} = \\operatorname {E n c} (\\mathcal {I}), \\mathbf {Z} = \\operatorname {E n c} \\left(\\mathbf {I} _ {\\mathbf {t}}\\right) \\\\ \\mathcal {F} = \\operatorname {P r o m p t E n c o d e r} (\\mathcal {P}, \\mathcal {Z}) \\tag {1} \\\\ \\mathbf {Q} _ {\\mathbf {p}} = \\text {P r o m p t S a m p l e} (\\mathcal {F}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 667, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the visual prompt features $\\mathbf{Q}_{\\mathbf{p}}$ , $DINOV$ incorporates segmentation queries $\\mathbf{Q}_{\\mathbf{s}}$ for proposal extraction. A shared decoder is employed to decode outputs for both $\\mathbf{Q}_{\\mathbf{s}}$ and $\\mathbf{Q}_{\\mathbf{p}}$ while performing cross-attention with respect to the target image feature $\\mathbf{Z}$ .", + "bbox": [ + 496, + 737, + 893, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {O} _ {\\mathbf {s}} = \\operatorname {D e c o d e r} \\left(\\mathbf {Q} _ {\\mathbf {s}}; \\mathbf {Z}\\right) \\\\ \\begin{array}{l} \\mathbf {O} _ {\\mathbf {p}} = \\operatorname {D e c o d e r} \\left(\\mathbf {Q} _ {\\mathbf {p}}; \\mathbf {Z}\\right) \\\\ \\text {P} = \\text {M a t h e r m a l} (\\mathbf {Q}) \\end{array} \\tag {2} \\\\ \\langle \\mathbf {M}, \\mathbf {B} \\rangle = \\operatorname {M a s k H e a d} \\left(\\mathbf {O} _ {\\mathbf {s}}\\right) \\\\ \\mathbf {C} _ {\\mathbf {g}}, \\mathbf {C} _ {\\mathbf {r}} = \\text {P r o m p t C l a s s i f i e r} (\\mathbf {O} _ {\\mathbf {s}}, \\mathbf {O} _ {\\mathbf {p}}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 829, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12863", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bafa3455c8142eadc2f331ab05134f3a49e81540315b4ec725dfcada581d2fee.jpg", + "image_caption": [ + "(a) Generic queries for generic segmentation" + ], + "image_footnote": [], + "bbox": [ + 98, + 98, + 307, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00119a6e77dded9b2e17a4b356c013118e2ab27baf3573ca270fde19c5d649cf.jpg", + "image_caption": [ + "(b) Interactive point queries for referring segmentation" + ], + "image_footnote": [], + "bbox": [ + 334, + 95, + 880, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7098ce3e9372dd7489996749c4924f44a92da542349a1d9657f399c38332c7f8.jpg", + "image_caption": [ + "Figure 4. DINOv query formulation of generic and referring segmentation tasks.", + "Figure 5. Prompt encoder to encode visual prompt from reference images. We use three masked cross-attention from the vision encoder small feature map to large feature map." + ], + "image_footnote": [], + "bbox": [ + 109, + 272, + 405, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{O_s}$ represents the decoded segmentation query features, $\\mathbf{O_p}$ corresponds to the decoded target visual prompt features, while $\\mathbf{M}$ and $\\mathbf{B}$ denote the predicted masks and boxes, respectively. Furthermore, we have $\\mathbf{C_g}$ and $\\mathbf{C_r}$ as the predicted matching scores for generic segmentation and referring segmentation tasks. These scores are derived through the use of a PromptClassifier, which computes the similarity between $\\mathbf{O_s}$ and $\\mathbf{O_p}$ .", + "bbox": [ + 75, + 486, + 468, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PromptClassifier. We clarify the definition of the prompt classifier, denoted as PromptClassifier $(\\cdot ,\\cdot)$ , for both generic segmentation and referring segmentation tasks here. In the case of generic segmentation tasks like instance and panoptic segmentation, the typical objective is to classify object features $\\mathbf{O_s}$ into respective categories. When employing visual prompting for generic segmentation tasks, the distinction lies in the utilization of visual prompt features $\\mathbf{O_p}$ as class embeddings. This is illustrated in the following equation:", + "bbox": [ + 75, + 607, + 470, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {g}} = \\mathbf {g} \\left(\\mathbf {O} _ {\\mathbf {s}}\\right) \\mathbf {g} \\left(\\mathbf {O} _ {\\mathbf {p}} ^ {T}\\right), \\mathbf {C} _ {\\mathbf {g}} \\in N _ {p} \\times N _ {s} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 767, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{p}$ and $N_{s}$ are the number of visual prompts and object features. $g$ is the linear projection for generic segmentation task. Each of $N_{s}$ objects is classified into one of $N_{p}$ classes. For visual referring segmentation, the objective differs. Here, each visual prompt is employed to identify the most closely matched instance within the target image. This task can be framed as a classification problem, where each", + "bbox": [ + 76, + 794, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual prompt is assigned to a specific instance within the target image. It's important to note that during our training phase, the target image and the reference image are identical. The matching score matrix for referring segmentation is structured as follows:", + "bbox": [ + 496, + 270, + 890, + 345 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C _ {r}} = \\mathbf {h} (\\mathbf {O _ {p}}) \\mathbf {h} (\\mathbf {O _ {s}} ^ {T}), \\mathbf {C _ {r}} \\in N _ {s} \\times N _ {q} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 357, + 890, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$h$ is the linear projection for referring segmentation task. Fig. 3(b) and (c) provide an illustrative representation of the two tasks. In our implementation, the generic segmentation task involves finding the most suitable visual prompt for each mask proposal, effectively pivoting the loss from a query to all prompts. Conversely, the referring segmentation task focuses on matching a given visual prompt to a specific mask proposal, with the loss pivot transitioning from a prompt to all proposals. As indicated in Equations 3 and 4, the PromptClassifier for both generic and referring segmentation tasks share a similar formulation. Consequently, they can share the entire framework, except for the two distinct linear layers denoted as $\\mathbf{g}$ and $\\mathbf{h}$ .", + "bbox": [ + 496, + 387, + 892, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2. Visual Prompt Formulation", + "text_level": 1, + "bbox": [ + 500, + 594, + 748, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The heart part of our $DINO\\nu$ is the proposed visual prompting mechanism. As shown in Eq. 1 and Eq. 2, we employ two modules to get the final visual prompt:", + "bbox": [ + 496, + 617, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A PromptEncoder to encode the reference visual prompt $\\mathcal{F}$ from the reference image features (followed by a sampling process to get query visual prompt $\\mathbf{Q}_{\\mathbf{p}}$ ).", + "- A Decoder (shared with the segmentation decoder) to decode outputs for the target visual prompt $\\mathbf{O}_{\\mathfrak{p}}$ by interacting with the target image features." + ], + "bbox": [ + 517, + 667, + 890, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This design allows our model to first encode the reference visual prompt and then adapt the prompt to the target image in a flexible way. As we attempt to express visual concepts through visual prompts, a straightforward way is to employ a pre-trained vision encoder (e.g., CLIP [29]) to process the reference images guided by user prompts [26]. However, it may encounter several challenges: $(i)$ the vision encoder takes cropped images as inputs, which causes substantial", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12864", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domain shift, especially for small objects [47]; $(ii)$ The visual features extracted from CLIP tend to be more semantic and may not meet the demands in VOS tasks. As we will show in our ablation study, employing a CLIP vision encoder to extract visual prompts has a clear inferior generalization ability.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these issues, we reuse the vision encoder in our model and develop a simple yet effective prompt encoder. It extracts visual features corresponding to the locations indicated by various forms of visual prompts. To capture visual details of different granularities, we have incorporated multiple layers (default to 3) of the Mask Cross Attention Layer, as shown in Fig. 5. Each layer takes the image features extracted at different levels (output multi-scale features from the vision encoder, ranging from lower to higher resolutions) as inputs, utilizes the regions defined by the visual inputs as masks, and employs learnable queries to process the features at the corresponding positions to get the visual prompt features.", + "bbox": [ + 75, + 183, + 473, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3. Prompt Sampling", + "text_level": 1, + "bbox": [ + 76, + 393, + 250, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce two prompt sampling strategies tailored for referring segmentation and generic segmentation.", + "bbox": [ + 75, + 417, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Referring segmentation. In the case of referring segmentation, we employ a \"self-referring\" approach during training, wherein the reference image is identical to the target image. Here, we sample a prompt from an instance and train the model to refer to the same instance. This approach allows us to leverage extensive segmentation data, such as SA-1B, for training our model effectively. Despite being trained on the same instances, our model demonstrates the capability to perform cross-image referring during inference. As illustrated in Fig. 3(c), we can change the target images to various different images, enabling the model to effectively engage in cross-image referring tasks.", + "bbox": [ + 75, + 448, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generic segmentation. The sampling strategies are slightly different during training and inference:", + "bbox": [ + 75, + 628, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Training. In the training process, it is crucial to create both positive and negative visual prompt samples. To achieve this, we generate visual prompts by utilizing a large image training batch. As depicted in Algorithm 1, our approach begins by grouping together reference visual prompt $\\mathcal{F}$ of the same semantic category across all images within a training batch. For each semantic category, we then randomly select a variable number of in-context examples, ranging from 1 to $N$ , and perform an aggregation process to generate reference visual prompt tokens $\\mathbf{Q}_{\\mathfrak{p}}$ , where each reference visual prompt token corresponds to a specific semantic category. $\\mathbf{Q}_{\\mathfrak{p}}$ is subsequently fed into the decoder, where it interacts with the target image to produce the final target visual prompt $\\mathbf{O}_{\\mathfrak{p}}$ . Consequently, the number of semantic categories corresponds to the number of target visual prompts. It's important to note that a given batch of", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "images may not encompass all semantic categories present in the dataset, resulting in a variable number of semantic categories during the training process.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Inference. During the inference stage, using the COCO Dataset as an example, we pre-extract the respective visual prompt features based on mask prompts for all semantic categories established during the training phase. For evaluation purposes, we adopt a random selection approach, where we choose N features for each semantic category. By default, N is set to 16. These selected features act as representative visual prompt features for each category. This practice ensures that our inference stage maintains the same number of categories as in traditional open-set evaluation, effectively preventing any potential information leakage.", + "bbox": [ + 496, + 136, + 893, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also provide the pseudo-code of prompting sampling in Appendix E.", + "bbox": [ + 496, + 301, + 890, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.4. Decoder Query Formulation", + "text_level": 1, + "bbox": [ + 500, + 340, + 756, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In $DINOv$ , we designed two types of segmentation queries to address two different tasks as depicted in Fig. 4. For generic segmentation, the query is a number of learnable ones similar to MaskDINO [16]. For the visual referring task, we adopt the interactive point query following Semantic-SAM [17], so that we can exploit the rich granularities in SA-1B [13]. Similar to Semantic-SAM, the visual prompts (points or boxes) are both converted into anchor box format, and then the position of each visual prompt will be encoded into position queries. Each position query is duplicated and subsequently combined with content queries of different granularities as the final segmentation queries. For the training on SA-1B, in order to avoid excessive computational overhead on the model, we selectively sample a subset of points contained within this visual concept as positive point queries. Concurrently, we randomly sample a subset of points from the remaining areas to serve as negative points. During the inference stage, we sample the initial point position queries on $20 \\times 20$ uniformly distributed grid as the initial point position for a single frame.", + "bbox": [ + 496, + 364, + 893, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 500, + 678, + 633, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We jointly train on SA-1B [13] and COCO [21]. We evaluate on open-set segmentation and video object segmentation. We provide the detailed experiment setting, implementation details, and metrics in Appendix. A, B, and C.", + "bbox": [ + 496, + 702, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Generic Segmentation and Detection", + "text_level": 1, + "bbox": [ + 500, + 771, + 816, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our visual prompt based generic segmentation performance in Table 1.", + "bbox": [ + 496, + 795, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In-domain Segmentation on COCO. Compared to other models trained for visual prompts, we achieve significantly better results. For example, we surpass SegGPT [34] and Painter [33] by 14.3 PQ and 25.5 PQ. In addition, With just a few visual in-context prompts for each category, our", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12865", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4abcddf3c3c52b134d99c75137ecb2e5e862b6b387b08066c2d64900c19472f4.jpg", + "table_caption": [ + "Table 1. One suit of weights for generic visual in-context segmentation on multiple datasets. Our model is trained on COCO and SA-1B data. Note: “-” denotes the model does not have number reported or does not have the ability for the specific task. $\\star$ means it is the test set results. † FC-CLIP adopts a frozen CLIP for open-set (text), we prompt the FC-CLIP with CLIP visual features to simulate visual promoting. # FC-CLIP and ODISE rely on frozen CLIP and Stable Diffusion knowledge. Mask DINO [18] is our baseline for comparison." + ], + "table_footnote": [], + "table_body": "
MethodSemantic DataTypeCOCO (in-domain)ADE (out-domain)SegInW (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoUAP-AverageAP-Median
Mask2Former-T [2]COCOClosed-set53.243.346.163.2-----
Mask2Former-B [2]COCO56.446.349.567.1-----
Mask2Former-L [2]COCO57.848.652.167.4-----
OneFormer-L [9]COCO57.948.9-67.4-----
MaskDINO-L [16]COCO58.350.656.267.5-----
Pano/SegFormer-B [36]COCO55.4--------
GLIPv2-H [45]COCO+O365+GOLDG+...Text Open-set-48.9*-------
MaskCLIP (L) [5]YFCC100M-----15.16.0-23.7
#ODISE-H [37]COCO (Stable diffusion))45.638.4-52.423.413.9-28.7-
#FC-CLIP-L [42]COCO (CLIP)54.444.6-63.726.816.8-34.1-
OpenSeed-T [44]COCO+O36555.447.652.064.019.814.117.022.933.921.5
X-Decoder-T [50]COCO+CC3M+...51.440.543.662.818.89.8-25.022.715.2
X-Decoder-L [50]COCO+CC3M+...56.946.7-67.521.813.1-29.636.138.7
OpenSeed-L [44]COCO+O36559.553.258.268.619.715.017.723.436.138.7
FC-CLIP†-L [42]COCOVisual Prompt----2.34.1-7.8-
SegGPT-L [34]COCO+ADE+VOC+...43.4--------
Painter-L [33]COCO+ADE+NYUv234.4--------
DINOv-T (Ours)COCO49.041.545.257.019.411.412.821.939.541.6
DINOv-L (Ours)COCO57.750.454.266.723.215.114.325.340.644.6
", + "bbox": [ + 80, + 157, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ef24e3e3088aea2d211f2e7b72af13b6951aa4f14a60f4ce68bf461012bdbeef.jpg", + "table_caption": [ + "Table 2. One suit of weights on ODinW benchmark. Average and median AP across 35 datasets are reported for simplicity." + ], + "table_footnote": [], + "table_body": "
ModelPretrain DataAverageMedian
MDETR [12]GOLDG, REFC10.73.0
GLIP-T [19]Object36511.41.6
OpenSeed (T) (Ours)Object365, COCO14.23.1
OpenSeed (L) (Ours)Object365, COCO15.25.0
DINOv (T) (Ours)COCO, SAM14.95.4
DINOv (L) (Ours)COCO, SAM15.74.78
", + "bbox": [ + 88, + 425, + 459, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model achieves comparable results with previous close-set or open-set models on COCO. For example, the panoptic segmentation performance gap between DINOv and our baseline Mask DINO is only 0.6 PQ (57.7 PQ vs 58.3 PQ). Out-domain open-set segmentation on ADE20K. After training with visual prompt on COCO and SAM, we do zero-shot evaluation on ADE20K to validate its open-set segmentation capability when seeing novel visual concepts. To our best knowledge, it is the first time to use visual prompt for open-set segmentation. Compared with previous text-prompted open-set models, we achieve comparable or better performance with only COCO semantic data and no semantic knowledge from large pre-trained models. Especially, compared with our baseline OpenSeed, we achieve better performance with much fewer data. Note that FC-CLIP [42] employs a frozen CLIP to do text-based open-set segmentation. As the text and visual features are aligned in CLIP, we also attempt to prompt a pre-trained FC-CLIP with visual features from CLIP to test its open-set ability with visual prompts. However, its visual prompting performance largely lags behind its text-prompted results. Therefore, it is non-trivial to transfer a multi-modal text-based open-set model to do visual-prompted recognition well. The results indicate that visual prompts can generalize well to new concepts.", + "bbox": [ + 73, + 523, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation and detection in the wild. We also validate the generalization capability of visual prompting on some diversified and domain-specific datasets including SegInW and ODinW, which in total encompass more than 60 datasets. These datasets contain many real-scenario or rare categories. As these datasets all focus on instance-level segmentation, we report the average and median AP (AP-Average and AP-Median) over all datasets. We first evaluate the Segmentation in the Wild (SegInW) benchmark, which consists of 25 datasets. With visual prompting, $DI-NOv$ achieves a significant performance improvement over our baseline OpenSeed. For example, Our best AP-Average exceeds OpenSeed by 4.5 AP. We further evaluate Object Detection in the Wild (ODinW), which is composed of 35 datasets with bounding box annotations. As shown in Table 2, though we only employ much fewer semantically labeled data, we achieve better performance compared with previous models under similar settings.", + "bbox": [ + 496, + 388, + 893, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Video Object Segmentation", + "text_level": 1, + "bbox": [ + 500, + 678, + 746, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video object segmentation (VOS) aims to segment an interested object in a video by giving text or visual clues. Our model focuses on the semi-supervised setting, which segments a particular object throughout a video by giving visual clues in the first frame. In $DINOv$ , the visual prompt originates from one single image (generic/referring segmentation) or other images in one batch (generic segmentation). Therefore, our model has learned to prompt with visual features from other images. Therefore, $DINOv$ is able to do video object segmentation (VOS) by replacing current frame visual prompt features with previous frames. For more accurate tracking, we also store the visual features of the predicted mask in previous frames. These fea", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12866", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ba5458e789504ff4b0a4e8651108c263b1d742f41442ab6cd6099ba925255974.jpg", + "table_caption": [ + "Table 3. Zero-shot video object segmentation. Without training with video or pairwise image data, our approach is able to do video object segmentation in a zero-shot manner. (#Concurrent work.)" + ], + "table_footnote": [], + "table_body": "
MethodSegmentation DataTypeRefer-TypeZero-ShotDAVIS17DAVIS16-InteractiveYouTube-VOS 2018
JFJFJFJFGJsFsJuFu
With Video Data
AGSS [20]VOS+DAVISVideoMask67.464.969.9---71.371.365.575.273.1
AGAME [11](Synth)VOS+DAVISMask70.067.272.7---66.066.9*61.2*
SWEM [22]Image+VOS+DAVISMask84.381.287.4---82.882.486.977.185.0
XMem [3]Image+VOS+DAVISMask------86.185.189.880.389.2
SiamMask [32]COCO+VOSBox*54.358.569.871.767.8*60.258.245.147.7
MiVOS [4]BL30K+VOS+DAVISMask84.581.787.491.089.692.482.681.185.677.786.2
ReferFormer-B [35]RefCOCO(+/g)+VOS+DAVISText61.158.164.1---*****
UNINEXT-T [41]Image+VideoGeneralistMask74.571.377.6---77.076.881.070.879.4
UNINEXT-L [41]Image+VideoMask77.273.281.2---78.179.183.571.078.9
UNINEXT-L [41]Image+VideoText66.762.371.1---*****
Without Video Data
Painter-L [33]COCO+ADE+NYUv2GeneralistMask✓34.628.540.8---24.127.635.814.318.7
SegGPT-L [34]COCO+ADE+VOC+...Mask✓75.672.578.6---74.775.180.267.475.9
PerSAM-L [46]SAM+DAVISMask✘60.356.663.9---*****
SEEM-T [51]COCO+LVISMask✓60.457.663.362.758.966.451.455.644.159.246.9
SEEM-L [51]Mask✓58.955.062.862.258.366.050.057.238.261.343.3
DINOv-T (Ours)COCO+SAMMask✓73.371.075.777.072.981.260.965.370.052.357.9
DINOv-L (Ours)✓72.369.874.875.471.379.459.661.765.752.352.358.8
", + "bbox": [ + 80, + 128, + 890, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tures, denoted as memory visual prompts, will be averaged together with the first frame's given prompt to construct the visual prompt of the current frame. Details of the memory visual prompt and ablations are in Appendix. D. By default, the memory length is set to 8. In Table 3, we conduct (interactive) video object segmentation evaluation on DAVIS17, DAVIS2016-Interactive, and Youtube-VOS 2018. The results of DAVIS2017 and Youtube-VOS 2018 indicate our model achieves better performance than SEEM and PersAM. In addition, $DINO\\nu$ can also do interactive VOS, and our performance on DAVIS16-Interactive achieves significantly better performance compared with models not using video data for training.", + "bbox": [ + 75, + 391, + 472, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3. Ablation", + "text_level": 1, + "bbox": [ + 76, + 603, + 181, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Query Formulation. In Table 4, we ablate the effectiveness of using different query formulations for different tasks. The results indicate our double query formulation outperforms using only one type of query.", + "bbox": [ + 75, + 628, + 468, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Visual Prompt Formulation. In Table 5, we attempt to use a pre-trained CLIP vision encoder to encode the features of the visual prompt by cropping the prompted region into images for CLIP to process. As CLIP features contain rich semantics with few appearance features, which could not apply to referring segmentation tasks. Therefore, we ablate on generic segmentation tasks and find that the final model could not generalize well on open-set datasets like ADE. This result verifies our hypothesis that CLIP vision features could not generalize well on in-context visual prompting.", + "bbox": [ + 75, + 688, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Unifying Tasks and Data. We unify visual generic segmentation and visual referring segmentation to use both semantically labeled data (COCO) and data with", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a192438141d31727e8d75bb8710486a0dec903100d2a7e1a23c0a578413f3cc5.jpg", + "table_caption": [ + "Table 4. Ablation of using difference queries to do both in-context reference and generic segmentation. By default, we use both generic query and interactive query. We remove one type of query at a time to ablate their effectiveness." + ], + "table_footnote": [], + "table_body": "
MethodCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinT49.642.747.058.073.371.075.7
only point query45.231.0(11.7)34.7(-12.3)52.771.468.874.0
only generic query46.238.3(-4.4)41.5(-6.0)53.368.966.571.3
", + "bbox": [ + 496, + 455, + 895, + 530 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8a8c2c413abb3701c77f7ea840c77522357e8a6114e506229771c212df793fe.jpg", + "table_caption": [ + "Table 5. Ablation of using different ways to encode the visual prompt on our Swin-T model. Under the same setting, we change our prompt encoding method and use a pre-trained CLIP to crop and encode the prompted objects in the image." + ], + "table_footnote": [], + "table_body": "
Prompt EncodingCOCO (in-domain)ADE (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoU
Ours49.642.747.058.019.411.412.821.9
CLIP48.540.743.554.912.61.41.313.3
", + "bbox": [ + 496, + 609, + 895, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only segmentation annotations (SA-1B). In Table 6, the results indicate that employing both datasets improves each individual task.", + "bbox": [ + 496, + 689, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training batch size for generic segmentation. In Table 7, the results show that increasing training batch size consistently improves the generic segmentation performance. The reason for this phenomenon is that a larger batch size helps to sample more positive and negative visual in-context examples across different images, which better matches the inference setting with random visual examples.", + "bbox": [ + 496, + 734, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference In-Context Examples. In Fig. 6, we ablate the impact of using different in-context lengths. Increasing the in-context example exhibits diminishing returns, especially when the number of examples is more than eight.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12867", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6bd13b327c728715453dcb32a1972be58bb1fb34ca8eda9e567eed18d9801d7.jpg", + "table_caption": [ + "Table 6. Ablation of the effectiveness of unifying tasks and data." + ], + "table_footnote": [], + "table_body": "
MethodDataCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinTCOCO, SAM49.642.747.058.073.371.075.7
DINOv-SwinTCOCO48.941.745.957.163.360.865.7
DINOv-SwinTSAMN/A---68.466.070.8
", + "bbox": [ + 73, + 114, + 473, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/712919a553e33b68ea2163c6a12179cdb43759d9f0bbfd1871cb95b3425c9fc0.jpg", + "table_caption": [ + "Table 7. Ablation of image batchsize sampling in training." + ], + "table_footnote": [], + "table_body": "
Method#Batchsize for Prompt SamplingCOCO
PQmask APbox APmIoU
DINOv-SwinT128.923.225.333.7
DINOv-SwinT445.137.040.450.6
DINOv-SwinT847.339.243.153.1
DINOv-SwinT3247.840.344.156.2
DINOv-SwinT6449.045.241.557.0
", + "bbox": [ + 73, + 210, + 473, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed78fa1f581541d252e0789c7a245441aba2661b4490d65021a301642e81dd1d.jpg", + "image_caption": [ + "Figure 6. DINOv query formulation of different tasks." + ], + "image_footnote": [], + "bbox": [ + 96, + 345, + 428, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Related Works", + "text_level": 1, + "bbox": [ + 76, + 578, + 225, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1. Visual Perception Through Text Prompt", + "text_level": 1, + "bbox": [ + 76, + 603, + 419, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Innovations in open-vocabulary object detection [7, 12, 19, 24, 26, 43, 44, 47] and open-vocabulary segmentation [6, 8, 14, 31, 37, 44], have shown great potential in generic visual perception, by leveraging large pre-trained vision-language models like CLIP [30] and ALIGN [10]. These approaches demonstrate significant strides in zero-shot and few-shot performance, adapting to a variety of visual contexts through text prompts. However, the reliance on text alone introduces limitations due to linguistic ambiguity and the potential mismatch between textual descriptions and complex visual scenes [40]. This highlights the ongoing need to refine the integration of visual inputs for more accurate and comprehensive image perception.", + "bbox": [ + 75, + 626, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Visual Perception Through Image Example", + "text_level": 1, + "bbox": [ + 76, + 832, + 444, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon the foundations set by text-based visual perception methodologies, the field has seen a notable shift towards incorporating image examples to enhance accu", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "racy and context sensitivity. OV-DETR [43] extends its open-vocabulary object detection capability beyond text, by utilizing both the image encoder and text encoder from CLIP [30], allowing for object detection guided by visual examples. Similarly, OWL-ViT [26] leverages large-scale image text examples in its contrastive pre-training phase, and propose to adopt image example for one-shot image-conditioned object detection. MQ-Det [40] utilizes image examples to enhance text descriptions for better open-vocabulary object detection performance. These methods typically adopt the image encoder in CLIP to extract visual features from given image examples for a more accurate perception of objects and scenes, and demonstrate that visual examples can bridge the gap between textual ambiguity and the complex nature of visual perception.", + "bbox": [ + 496, + 90, + 890, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Visual Perception Through Visual Prompt", + "text_level": 1, + "bbox": [ + 500, + 325, + 857, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different from image example-based methods that take an image as input, which are then processed by multi-modal encoder like CLIP [30], visual prompt-based methods typically use visual instructions (e.g. point, mask, scribble, and refereed regions of another image) to guide a model for a specific visual task. SAM [13], for instance, introduces a promotable model for interactive segmentation, fostering research in computer vision foundation models. SEEM [51] stands out as an interactive and versatile model for segmenting objects, accommodating various types of prompts, and is semantic-aware compared to SAM. Semantic-SAM [17] excels in semantic awareness and recognizing granularity, and can handle panoptic and part segmentation. SegGPT [34] unifies various segmentation tasks by formulating segmentation as an in-context coloring problem.", + "bbox": [ + 496, + 349, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 588, + 617, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present DINOv, a unified framework for in-context visual prompting to accommodate both referring segmentation and generic segmentation tasks. To effectively formulate in-context visual prompts, we designed a simple prompt encoder to encoder reference visual prompts from the reference image and adopted a shared decoder to decode the final target visual prompts from the target image. We also formulate generic latent queries and point queries to align different tasks and data. The experimental results indicate that DINOv demonstrates impressive referring and generic segmentation capabilities to refer and detect with in-context visual prompting. Notably, DINOv delivers competitive performance compared to close-set segmentation on in-domain datasets and show promising results on many open-set segmentation benchmarks. We hope our early exploration of visual in-context prompting could inspire the community. Limitations. We employ limited semantically labeled data (COCO), which can be scaled up for better performance and extended to text prompts for multi-modal understanding.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12868", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1", + "[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 6", + "[3] Ho Kei Cheng and Alexander G. Schwing. XMem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7", + "[4] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 7", + "[5] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022.6", + "[6] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 8", + "[7] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8", + "[8] Dat Huynh, Jason Kuen, Zhe Lin, Jiuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 8", + "[9] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022. 6", + "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8", + "[11] Joakim Johnander, Martin Danelljan, Emil Brissman, Fahad Shahbaz Khan, and Michael Felsberg. A generative appearance model for end-to-end video object segmentation, 2018. 7", + "[12] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 6, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 2, 5, 8, 12", + "[14] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 8", + "[15] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2", + "[16] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 5, 6, 12", + "[17] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 2, 5, 8, 12", + "[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2, 6", + "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 6, 8, 12", + "[20] Huaijia Lin, Xiaojuan Qi, and Jiaya Jia. Agss-vos: Attention guided single-shot video object segmentation. In ICCV, 2019. 7", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 5, 12", + "[22] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1362-1372, 2022. 7", + "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2", + "[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 8", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12869", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 12", + "[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 4, 8", + "[27] OpenAI. Gpt-4 technical report, 2023. 1", + "[28] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 12", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 8", + "[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 8", + "[32] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2019. 7", + "[33] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning, 2023. 5, 6, 7", + "[34] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2, 5, 6, 7, 8", + "[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. arXiv preprint arXiv:2201.00487, 2022. 7", + "[36] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 6", + "[37] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models, 2023. 2", + "[39] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 12", + "[40] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multimodal queried object detection in the wild. arXiv preprint arXiv:2305.18980, 2023. 8", + "[41] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. arXiv preprint arXiv:2303.06674, 2023. 7", + "[42] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. arXiv preprint arXiv:2308.02487, 2023. 2, 6", + "[43] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8", + "[44] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 6, 8", + "[45] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6", + "[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot, 2023. 7", + "[47] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 5, 8", + "[48] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 12", + "[49] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2", + "[50] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022. 2, 6, 12" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12870", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[51] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 7, 8", + "bbox": [ + 78, + 90, + 470, + 146 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_model.json b/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9c6f61f07ec2c3700caa2bd653cde87d715d0094 --- /dev/null +++ b/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.337, + 0.132, + 0.635, + 0.154 + ], + "angle": 0, + "content": "Visual In-Context Prompting" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.18, + 0.893, + 0.219 + ], + "angle": 0, + "content": "Feng Li\\*, Qing Jiang#, Hao Zhang#, Tianhe Ren†, Shilong Liu#, Xueyan Zou#, Huaizhe Xu#, Hongyang Li#, Jianwei Yang‡1, Chunyuan Li#, Lei Zhang†2, Jianfeng Gao‡2" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.226, + 0.736, + 0.242 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) HKUST \\(\\ddagger\\) Microsoft Research, Redmond \\(\\dagger\\) IDEA \\(\\sharp\\) SCUT \\(\\sharp\\) Tsinghua \\(\\S\\) UW-Madison" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.246, + 0.597, + 0.259 + ], + "angle": 0, + "content": "1. Project Lead 2. Equal Advisory Contribution" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.277, + 0.85, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.323, + 0.376, + 0.645, + 0.392 + ], + "angle": 0, + "content": "Visual Prompting Generic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.399, + 0.848, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.315, + 0.489, + 0.652, + 0.507 + ], + "angle": 0, + "content": "Visual Prompting Referring Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.517, + 0.842, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.624, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Figure 1. Our model \\(DINOV\\) supports generic and referring segmentation to associate multiple or single objects with the user input visual prompts. A user can input one or more in-context visual prompts (scribbles, masks, boxes, etc.) to improve the segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.671, + 0.314, + 0.687 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.705, + 0.474, + 0.903 + ], + "angle": 0, + "content": "In-context prompting in large language models (LLMs) has become a prevalent approach to improve zero-shot capabilities, but this idea is less explored in the vision domain. Existing visual prompting methods focus on referring segmentation to segment the most relevant object, falling short of addressing many generic vision tasks like open-set segmentation and detection. In this paper, we introduce a universal visual in-context prompting framework for both tasks, as shown in Fig. 1. In particular, we build on top of an encoder-decoder architecture, and develop a versatile prompt encoder to support a variety of prompts like strokes, boxes, and points. We further enhance it to take an arbitrary number of reference image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.809 + ], + "angle": 0, + "content": "segments as the context. Our extensive explorations show that the proposed visual in-context prompting elicits extraordinary referring and generic segmentation capabilities to refer and detect, yielding competitive performance to close-set in-domain datasets and showing promising results on many open-set segmentation datasets. By joint training on COCO and SA-1B, DINOv achieves 57.7 PQ on COCO and 23.2 PQ on ADE20K. Code will be available at https://github.com/UX-Decoder/DINOv" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.843, + 0.633, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The recent progress in large language models (LLMs) like GPT [1, 27] has shown promising results towards arti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12861" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "ficial general intelligence (AGI) by training unified models on large amounts of text data. These giant LLMs manifest themselves with intriguing emerging capabilities such as in-context learning. Nevertheless, similar paradigms have not yet succeeded in solving all vision tasks due to the diversity of scenarios in computer vision [15]. Some works [23, 49] have combined LLMs and vision models to tackle complex image understanding tasks with text outputs such as visual question answering, but challenges remain in fine-grained tasks that require pixel-level outputs, like instance masks, rather than just text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.429 + ], + "angle": 0, + "content": "To this end, the community has observed a growing interest in the development of language-enhanced vision foundation models. These models demonstrate profound competencies in open-world visual understanding tasks using text prompts, encompassing areas like open-set detection [19, 24, 44] and segmentation [38, 42, 44, 50]. Visual prompt, a different prompting mechanism has been explored in some recent segmentation models [13, 17, 51]. In these works, different visual prompting formats (e.g., points, boxes and strokes, etc) have been explored to facilitate the segmentation of visual contents specified by users." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.793 + ], + "angle": 0, + "content": "Another critical LLM technique, in-context learning, has been less explored. In-context learning specifies the new task instruction using examples, and allows models to adapt to new tasks or domains without explicit retraining by providing relevant examples. One pioneering work in this area is SegGPT [34], which demonstrates the ability to output an image mask based on visual examples. However, these works focus on associating a user visual prompt with one most relevant object and have the limited ability to identify multiple objects of the same semantic concept. Therefore, these approaches fall short of addressing many generic vision tasks like open-set object detection and segmentation, which often require the segmentation of multiple objects of a given concept. On the other hand, while text-prompted vision models [24, 50] do not align with in-context learning methodologies, they exhibit notable flexibility in managing both referring and generic tasks in detection or segmentation, but with language prompts. In this paper, we strive to develop a model that supports visual in-context prompting for all types of image segmentation tasks. A comparison between our work and previous work is shown in Fig. 2. Besides supporting both single-image and cross-image visual prompting, our model distinguishes itself by effectively handling both referring and generic segmentation problems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To achieve this goal, we build a model called DINOv to support versatile visual prompting capabilities, based on the unified detection and segmentation model MaskDINO [18]. DINOv follows the general encoder-decoder design with an extra prompt encoder to formulate and sample visual prompts. The decoder takes in segmentation queries and reference prompt queries to generate segmentation masks" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.889, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.361, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Figure 2. Comparison with related works. Generic: segment all objects of the same semantic concept that match the user prompt. Refer: segment a particular object with the user input visual prompts. Image prompt: crop the image regions as prompts. (single) Visual prompt: one image-prompt example to segment. In-context prompt: one or multiple image-prompt examples. We can do single-image and cross-image visual prompting tasks and support referring and generic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.821 + ], + "angle": 0, + "content": "and target visual prompts, and we associate the output segmentation masks with the target prompt queries for the final output. We can define the visual in-context samples with a set of reference image \\((Q)\\) - visual prompt \\((A)\\) pairs. The visual prompt can be in various types, including mask, scribble, box, etc. With the in-context examples, our model takes in a target image and outputs the masks. The creation of target visual prompts involves a prompt encoder that extracts reference visual prompts from a Q-A pair. This is followed by a decoder to get the target visual prompt by attending reference visual prompts to the target image. During training, to construct positive and negative samples for generic segmentation, we sample reference visual prompts in a batch across different images. To address task and data discrepancies, we formulate generic latent queries and point queries for generic and referring segmentation, respectively. By joint training on COCO [21] and SA-1B [13] for generic and referring segmentation, our model attains competitive performance on in-domain segmentation tasks compared with text-prompted models and shows promising generalization capability on a wide range of open-set segmentation benchmarks using purely visual prompts." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.822, + 0.828, + 0.836 + ], + "angle": 0, + "content": "To summarize, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "- We are the first to extend visual in-context prompting to support generic vision tasks like open-set generic segmentation and detection, and achieve comparable performance with text prompt-based open-set models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12862" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.096, + 0.458, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.326, + 0.335, + 0.338 + ], + "angle": 0, + "content": "(a) General Framework" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.093, + 0.892, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.325, + 0.671, + 0.338 + ], + "angle": 0, + "content": "(b) Visual Generic Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.325, + 0.866, + 0.338 + ], + "angle": 0, + "content": "(c) Visual Referring Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.894, + 0.422 + ], + "angle": 0, + "content": "Figure 3. DINov is a universal segmentation framework that can do generic segmentation and referring image segmentation. The vision encoder is used to extract image features.(b) An illustration of losses for visual generic segmentation. In the example, there are 6 visual prompts sampled from 6 masks from 3 categories. The visual prompts from the instances of the same class are averaged as the class embedding. Each column of the matching matrix is a 3-dimension one-hot vector which is a one-hot class label of the instance; (c) An illustration of losses for visual referring segmentation. Each visual prompt is classified to one of the 6 instances." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.462, + 0.47, + 0.553 + ], + "angle": 0, + "content": "- We build \\(DINOV\\), a unified framework for referring segmentation and generic segmentation based on visual in-context prompting. This unification simplifies model design and allows our model to consume both semantically-labelled and unlabelled data for better performance." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.557, + 0.47, + 0.634 + ], + "angle": 0, + "content": "- We conduct extensive experiments and visualizations to show that our model can handle generic, referring, and video object segmentation tasks. Our early attempts exhibit promising results on open-set segmentation and detection with visual prompting." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.462, + 0.47, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.64, + 0.169, + 0.655 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.665, + 0.454, + 0.681 + ], + "angle": 0, + "content": "2.1. Unified Formulation for Segmentation Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.869 + ], + "angle": 0, + "content": "In this paper, we concentrate on visual prompting tasks involving images, encompassing both generic segmentation and referring segmentation tasks. Given \\(N\\) reference images \\(\\mathcal{I} = \\{\\mathbf{I}_1,\\dots,\\mathbf{I}_N\\} \\in \\mathcal{R}^{N\\times H\\times W\\times 3}\\) with the corresponding visual prompts \\(\\mathcal{P} = \\{p_1,\\dots,p_N\\}\\), DINOv aims to segment objects of interest on a new target image \\(\\mathbf{I}_{\\mathbf{t}}\\). The visual prompts include masks, boxes, scribbles, points, etc. The interested objects can be a particular object for referring segmentation or all objects of the same semantic concept for generic segmentation. Note that the reference image can be identical to the target image, in which the task reduces to single-image visual prompting segmentation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To address these tasks, DINOv utilizes a comprehensive query-based encoder-decoder architecture. This archi-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.895, + 0.655 + ], + "angle": 0, + "content": "tecture comprises a vision encoder, denoted as Enc, responsible for extracting image features, a prompt encoder referred to as PromptEncoder, designed to extract visual prompt features by combining image features and user-provided visual prompts, and a general decoder represented as Decoder, which generates masks and visual concepts based on the segmentation query and visual prompt features. Upon receiving the input image and user-provided visual prompts, our initial step involves extracting image features denoted as Z using the vision encoder. Subsequently, we feed both the image features and visual prompts into the prompt encoder to extract the reference visual prompt \\(\\mathcal{F}\\) and subsequently sample the query visual prompt features \\(\\mathbf{Q}_{\\mathfrak{p}}\\). Formally, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.669, + 0.892, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {Z} = \\operatorname {E n c} (\\mathcal {I}), \\mathbf {Z} = \\operatorname {E n c} \\left(\\mathbf {I} _ {\\mathbf {t}}\\right) \\\\ \\mathcal {F} = \\operatorname {P r o m p t E n c o d e r} (\\mathcal {P}, \\mathcal {Z}) \\tag {1} \\\\ \\mathbf {Q} _ {\\mathbf {p}} = \\text {P r o m p t S a m p l e} (\\mathcal {F}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.894, + 0.814 + ], + "angle": 0, + "content": "In addition to the visual prompt features \\(\\mathbf{Q}_{\\mathbf{p}}\\), \\(DINOV\\) incorporates segmentation queries \\(\\mathbf{Q}_{\\mathbf{s}}\\) for proposal extraction. A shared decoder is employed to decode outputs for both \\(\\mathbf{Q}_{\\mathbf{s}}\\) and \\(\\mathbf{Q}_{\\mathbf{p}}\\) while performing cross-attention with respect to the target image feature \\(\\mathbf{Z}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.83, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {O} _ {\\mathbf {s}} = \\operatorname {D e c o d e r} \\left(\\mathbf {Q} _ {\\mathbf {s}}; \\mathbf {Z}\\right) \\\\ \\begin{array}{l} \\mathbf {O} _ {\\mathbf {p}} = \\operatorname {D e c o d e r} \\left(\\mathbf {Q} _ {\\mathbf {p}}; \\mathbf {Z}\\right) \\\\ \\text {P} = \\text {M a t h e r m a l} (\\mathbf {Q}) \\end{array} \\tag {2} \\\\ \\langle \\mathbf {M}, \\mathbf {B} \\rangle = \\operatorname {M a s k H e a d} \\left(\\mathbf {O} _ {\\mathbf {s}}\\right) \\\\ \\mathbf {C} _ {\\mathbf {g}}, \\mathbf {C} _ {\\mathbf {r}} = \\text {P r o m p t C l a s s i f i e r} (\\mathbf {O} _ {\\mathbf {s}}, \\mathbf {O} _ {\\mathbf {p}}) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.099, + 0.308, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.194, + 0.277, + 0.224 + ], + "angle": 0, + "content": "(a) Generic queries for generic segmentation" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.096, + 0.882, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.202, + 0.776, + 0.218 + ], + "angle": 0, + "content": "(b) Interactive point queries for referring segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.236, + 0.722, + 0.251 + ], + "angle": 0, + "content": "Figure 4. DINOv query formulation of generic and referring segmentation tasks." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.273, + 0.406, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Figure 5. Prompt encoder to encode visual prompt from reference images. We use three masked cross-attention from the vision encoder small feature map to large feature map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{O_s}\\) represents the decoded segmentation query features, \\(\\mathbf{O_p}\\) corresponds to the decoded target visual prompt features, while \\(\\mathbf{M}\\) and \\(\\mathbf{B}\\) denote the predicted masks and boxes, respectively. Furthermore, we have \\(\\mathbf{C_g}\\) and \\(\\mathbf{C_r}\\) as the predicted matching scores for generic segmentation and referring segmentation tasks. These scores are derived through the use of a PromptClassifier, which computes the similarity between \\(\\mathbf{O_s}\\) and \\(\\mathbf{O_p}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.471, + 0.759 + ], + "angle": 0, + "content": "PromptClassifier. We clarify the definition of the prompt classifier, denoted as PromptClassifier \\((\\cdot ,\\cdot)\\), for both generic segmentation and referring segmentation tasks here. In the case of generic segmentation tasks like instance and panoptic segmentation, the typical objective is to classify object features \\(\\mathbf{O_s}\\) into respective categories. When employing visual prompting for generic segmentation tasks, the distinction lies in the utilization of visual prompt features \\(\\mathbf{O_p}\\) as class embeddings. This is illustrated in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.768, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {g}} = \\mathbf {g} \\left(\\mathbf {O} _ {\\mathbf {s}}\\right) \\mathbf {g} \\left(\\mathbf {O} _ {\\mathbf {p}} ^ {T}\\right), \\mathbf {C} _ {\\mathbf {g}} \\in N _ {p} \\times N _ {s} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "where \\( N_{p} \\) and \\( N_{s} \\) are the number of visual prompts and object features. \\( g \\) is the linear projection for generic segmentation task. Each of \\( N_{s} \\) objects is classified into one of \\( N_{p} \\) classes. For visual referring segmentation, the objective differs. Here, each visual prompt is employed to identify the most closely matched instance within the target image. This task can be framed as a classification problem, where each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.346 + ], + "angle": 0, + "content": "visual prompt is assigned to a specific instance within the target image. It's important to note that during our training phase, the target image and the reference image are identical. The matching score matrix for referring segmentation is structured as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.358, + 0.892, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C _ {r}} = \\mathbf {h} (\\mathbf {O _ {p}}) \\mathbf {h} (\\mathbf {O _ {s}} ^ {T}), \\mathbf {C _ {r}} \\in N _ {s} \\times N _ {q} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.893, + 0.585 + ], + "angle": 0, + "content": "\\(h\\) is the linear projection for referring segmentation task. Fig. 3(b) and (c) provide an illustrative representation of the two tasks. In our implementation, the generic segmentation task involves finding the most suitable visual prompt for each mask proposal, effectively pivoting the loss from a query to all prompts. Conversely, the referring segmentation task focuses on matching a given visual prompt to a specific mask proposal, with the loss pivot transitioning from a prompt to all proposals. As indicated in Equations 3 and 4, the PromptClassifier for both generic and referring segmentation tasks share a similar formulation. Consequently, they can share the entire framework, except for the two distinct linear layers denoted as \\(\\mathbf{g}\\) and \\(\\mathbf{h}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.749, + 0.611 + ], + "angle": 0, + "content": "2.2. Visual Prompt Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.664 + ], + "angle": 0, + "content": "The heart part of our \\(DINO\\nu\\) is the proposed visual prompting mechanism. As shown in Eq. 1 and Eq. 2, we employ two modules to get the final visual prompt:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.668, + 0.892, + 0.716 + ], + "angle": 0, + "content": "- A PromptEncoder to encode the reference visual prompt \\(\\mathcal{F}\\) from the reference image features (followed by a sampling process to get query visual prompt \\(\\mathbf{Q}_{\\mathbf{p}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.73, + 0.892, + 0.775 + ], + "angle": 0, + "content": "- A Decoder (shared with the segmentation decoder) to decode outputs for the target visual prompt \\(\\mathbf{O}_{\\mathfrak{p}}\\) by interacting with the target image features." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.668, + 0.892, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This design allows our model to first encode the reference visual prompt and then adapt the prompt to the target image in a flexible way. As we attempt to express visual concepts through visual prompts, a straightforward way is to employ a pre-trained vision encoder (e.g., CLIP [29]) to process the reference images guided by user prompts [26]. However, it may encounter several challenges: \\((i)\\) the vision encoder takes cropped images as inputs, which causes substantial" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12864" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "domain shift, especially for small objects [47]; \\((ii)\\) The visual features extracted from CLIP tend to be more semantic and may not meet the demands in VOS tasks. As we will show in our ablation study, employing a CLIP vision encoder to extract visual prompts has a clear inferior generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.474, + 0.381 + ], + "angle": 0, + "content": "To address these issues, we reuse the vision encoder in our model and develop a simple yet effective prompt encoder. It extracts visual features corresponding to the locations indicated by various forms of visual prompts. To capture visual details of different granularities, we have incorporated multiple layers (default to 3) of the Mask Cross Attention Layer, as shown in Fig. 5. Each layer takes the image features extracted at different levels (output multi-scale features from the vision encoder, ranging from lower to higher resolutions) as inputs, utilizes the regions defined by the visual inputs as masks, and employs learnable queries to process the features at the corresponding positions to get the visual prompt features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.251, + 0.41 + ], + "angle": 0, + "content": "2.3. Prompt Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.448 + ], + "angle": 0, + "content": "We introduce two prompt sampling strategies tailored for referring segmentation and generic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Referring segmentation. In the case of referring segmentation, we employ a \"self-referring\" approach during training, wherein the reference image is identical to the target image. Here, we sample a prompt from an instance and train the model to refer to the same instance. This approach allows us to leverage extensive segmentation data, such as SA-1B, for training our model effectively. Despite being trained on the same instances, our model demonstrates the capability to perform cross-image referring during inference. As illustrated in Fig. 3(c), we can change the target images to various different images, enabling the model to effectively engage in cross-image referring tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Generic segmentation. The sampling strategies are slightly different during training and inference:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- Training. In the training process, it is crucial to create both positive and negative visual prompt samples. To achieve this, we generate visual prompts by utilizing a large image training batch. As depicted in Algorithm 1, our approach begins by grouping together reference visual prompt \\(\\mathcal{F}\\) of the same semantic category across all images within a training batch. For each semantic category, we then randomly select a variable number of in-context examples, ranging from 1 to \\(N\\), and perform an aggregation process to generate reference visual prompt tokens \\(\\mathbf{Q}_{\\mathfrak{p}}\\), where each reference visual prompt token corresponds to a specific semantic category. \\(\\mathbf{Q}_{\\mathfrak{p}}\\) is subsequently fed into the decoder, where it interacts with the target image to produce the final target visual prompt \\(\\mathbf{O}_{\\mathfrak{p}}\\). Consequently, the number of semantic categories corresponds to the number of target visual prompts. It's important to note that a given batch of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "images may not encompass all semantic categories present in the dataset, resulting in a variable number of semantic categories during the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.894, + 0.303 + ], + "angle": 0, + "content": "- Inference. During the inference stage, using the COCO Dataset as an example, we pre-extract the respective visual prompt features based on mask prompts for all semantic categories established during the training phase. For evaluation purposes, we adopt a random selection approach, where we choose N features for each semantic category. By default, N is set to 16. These selected features act as representative visual prompt features for each category. This practice ensures that our inference stage maintains the same number of categories as in traditional open-set evaluation, effectively preventing any potential information leakage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.334 + ], + "angle": 0, + "content": "We also provide the pseudo-code of prompting sampling in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.341, + 0.757, + 0.358 + ], + "angle": 0, + "content": "2.4. Decoder Query Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.895, + 0.668 + ], + "angle": 0, + "content": "In \\(DINOv\\), we designed two types of segmentation queries to address two different tasks as depicted in Fig. 4. For generic segmentation, the query is a number of learnable ones similar to MaskDINO [16]. For the visual referring task, we adopt the interactive point query following Semantic-SAM [17], so that we can exploit the rich granularities in SA-1B [13]. Similar to Semantic-SAM, the visual prompts (points or boxes) are both converted into anchor box format, and then the position of each visual prompt will be encoded into position queries. Each position query is duplicated and subsequently combined with content queries of different granularities as the final segmentation queries. For the training on SA-1B, in order to avoid excessive computational overhead on the model, we selectively sample a subset of points contained within this visual concept as positive point queries. Concurrently, we randomly sample a subset of points from the remaining areas to serve as negative points. During the inference stage, we sample the initial point position queries on \\(20 \\times 20\\) uniformly distributed grid as the initial point position for a single frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.634, + 0.696 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.765 + ], + "angle": 0, + "content": "We jointly train on SA-1B [13] and COCO [21]. We evaluate on open-set segmentation and video object segmentation. We provide the detailed experiment setting, implementation details, and metrics in Appendix. A, B, and C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.818, + 0.789 + ], + "angle": 0, + "content": "3.1. Generic Segmentation and Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.891, + 0.825 + ], + "angle": 0, + "content": "We evaluate our visual prompt based generic segmentation performance in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In-domain Segmentation on COCO. Compared to other models trained for visual prompts, we achieve significantly better results. For example, we surpass SegGPT [34] and Painter [33] by 14.3 PQ and 25.5 PQ. In addition, With just a few visual in-context prompts for each category, our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12865" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.149 + ], + "angle": 0, + "content": "Table 1. One suit of weights for generic visual in-context segmentation on multiple datasets. Our model is trained on COCO and SA-1B data. Note: “-” denotes the model does not have number reported or does not have the ability for the specific task. \\(\\star\\) means it is the test set results. † FC-CLIP adopts a frozen CLIP for open-set (text), we prompt the FC-CLIP with CLIP visual features to simulate visual promoting. # FC-CLIP and ODISE rely on frozen CLIP and Stable Diffusion knowledge. Mask DINO [18] is our baseline for comparison." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.158, + 0.892, + 0.365 + ], + "angle": 0, + "content": "
MethodSemantic DataTypeCOCO (in-domain)ADE (out-domain)SegInW (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoUAP-AverageAP-Median
Mask2Former-T [2]COCOClosed-set53.243.346.163.2-----
Mask2Former-B [2]COCO56.446.349.567.1-----
Mask2Former-L [2]COCO57.848.652.167.4-----
OneFormer-L [9]COCO57.948.9-67.4-----
MaskDINO-L [16]COCO58.350.656.267.5-----
Pano/SegFormer-B [36]COCO55.4--------
GLIPv2-H [45]COCO+O365+GOLDG+...Text Open-set-48.9*-------
MaskCLIP (L) [5]YFCC100M-----15.16.0-23.7
#ODISE-H [37]COCO (Stable diffusion))45.638.4-52.423.413.9-28.7-
#FC-CLIP-L [42]COCO (CLIP)54.444.6-63.726.816.8-34.1-
OpenSeed-T [44]COCO+O36555.447.652.064.019.814.117.022.933.921.5
X-Decoder-T [50]COCO+CC3M+...51.440.543.662.818.89.8-25.022.715.2
X-Decoder-L [50]COCO+CC3M+...56.946.7-67.521.813.1-29.636.138.7
OpenSeed-L [44]COCO+O36559.553.258.268.619.715.017.723.436.138.7
FC-CLIP†-L [42]COCOVisual Prompt----2.34.1-7.8-
SegGPT-L [34]COCO+ADE+VOC+...43.4--------
Painter-L [33]COCO+ADE+NYUv234.4--------
DINOv-T (Ours)COCO49.041.545.257.019.411.412.821.939.541.6
DINOv-L (Ours)COCO57.750.454.266.723.215.114.325.340.644.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.387, + 0.472, + 0.418 + ], + "angle": 0, + "content": "Table 2. One suit of weights on ODinW benchmark. Average and median AP across 35 datasets are reported for simplicity." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.426, + 0.46, + 0.518 + ], + "angle": 0, + "content": "
ModelPretrain DataAverageMedian
MDETR [12]GOLDG, REFC10.73.0
GLIP-T [19]Object36511.41.6
OpenSeed (T) (Ours)Object365, COCO14.23.1
OpenSeed (L) (Ours)Object365, COCO15.25.0
DINOv (T) (Ours)COCO, SAM14.95.4
DINOv (L) (Ours)COCO, SAM15.74.78
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.524, + 0.472, + 0.903 + ], + "angle": 0, + "content": "model achieves comparable results with previous close-set or open-set models on COCO. For example, the panoptic segmentation performance gap between DINOv and our baseline Mask DINO is only 0.6 PQ (57.7 PQ vs 58.3 PQ). Out-domain open-set segmentation on ADE20K. After training with visual prompt on COCO and SAM, we do zero-shot evaluation on ADE20K to validate its open-set segmentation capability when seeing novel visual concepts. To our best knowledge, it is the first time to use visual prompt for open-set segmentation. Compared with previous text-prompted open-set models, we achieve comparable or better performance with only COCO semantic data and no semantic knowledge from large pre-trained models. Especially, compared with our baseline OpenSeed, we achieve better performance with much fewer data. Note that FC-CLIP [42] employs a frozen CLIP to do text-based open-set segmentation. As the text and visual features are aligned in CLIP, we also attempt to prompt a pre-trained FC-CLIP with visual features from CLIP to test its open-set ability with visual prompts. However, its visual prompting performance largely lags behind its text-prompted results. Therefore, it is non-trivial to transfer a multi-modal text-based open-set model to do visual-prompted recognition well. The results indicate that visual prompts can generalize well to new concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.662 + ], + "angle": 0, + "content": "Segmentation and detection in the wild. We also validate the generalization capability of visual prompting on some diversified and domain-specific datasets including SegInW and ODinW, which in total encompass more than 60 datasets. These datasets contain many real-scenario or rare categories. As these datasets all focus on instance-level segmentation, we report the average and median AP (AP-Average and AP-Median) over all datasets. We first evaluate the Segmentation in the Wild (SegInW) benchmark, which consists of 25 datasets. With visual prompting, \\( DI-NOv \\) achieves a significant performance improvement over our baseline OpenSeed. For example, Our best AP-Average exceeds OpenSeed by 4.5 AP. We further evaluate Object Detection in the Wild (ODinW), which is composed of 35 datasets with bounding box annotations. As shown in Table 2, though we only employ much fewer semantically labeled data, we achieve better performance compared with previous models under similar settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.747, + 0.696 + ], + "angle": 0, + "content": "3.2. Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Video object segmentation (VOS) aims to segment an interested object in a video by giving text or visual clues. Our model focuses on the semi-supervised setting, which segments a particular object throughout a video by giving visual clues in the first frame. In \\(DINOv\\), the visual prompt originates from one single image (generic/referring segmentation) or other images in one batch (generic segmentation). Therefore, our model has learned to prompt with visual features from other images. Therefore, \\(DINOv\\) is able to do video object segmentation (VOS) by replacing current frame visual prompt features with previous frames. For more accurate tracking, we also store the visual features of the predicted mask in previous frames. These fea" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12866" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 3. Zero-shot video object segmentation. Without training with video or pairwise image data, our approach is able to do video object segmentation in a zero-shot manner. (#Concurrent work.)" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.13, + 0.892, + 0.368 + ], + "angle": 0, + "content": "
MethodSegmentation DataTypeRefer-TypeZero-ShotDAVIS17DAVIS16-InteractiveYouTube-VOS 2018
JFJFJFJFGJsFsJuFu
With Video Data
AGSS [20]VOS+DAVISVideoMask67.464.969.9---71.371.365.575.273.1
AGAME [11](Synth)VOS+DAVISMask70.067.272.7---66.066.9*61.2*
SWEM [22]Image+VOS+DAVISMask84.381.287.4---82.882.486.977.185.0
XMem [3]Image+VOS+DAVISMask------86.185.189.880.389.2
SiamMask [32]COCO+VOSBox*54.358.569.871.767.8*60.258.245.147.7
MiVOS [4]BL30K+VOS+DAVISMask84.581.787.491.089.692.482.681.185.677.786.2
ReferFormer-B [35]RefCOCO(+/g)+VOS+DAVISText61.158.164.1---*****
UNINEXT-T [41]Image+VideoGeneralistMask74.571.377.6---77.076.881.070.879.4
UNINEXT-L [41]Image+VideoMask77.273.281.2---78.179.183.571.078.9
UNINEXT-L [41]Image+VideoText66.762.371.1---*****
Without Video Data
Painter-L [33]COCO+ADE+NYUv2GeneralistMask✓34.628.540.8---24.127.635.814.318.7
SegGPT-L [34]COCO+ADE+VOC+...Mask✓75.672.578.6---74.775.180.267.475.9
PerSAM-L [46]SAM+DAVISMask✘60.356.663.9---*****
SEEM-T [51]COCO+LVISMask✓60.457.663.362.758.966.451.455.644.159.246.9
SEEM-L [51]Mask✓58.955.062.862.258.366.050.057.238.261.343.3
DINOv-T (Ours)COCO+SAMMask✓73.371.075.777.072.981.260.965.370.052.357.9
DINOv-L (Ours)✓72.369.874.875.471.379.459.661.765.752.352.358.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.473, + 0.589 + ], + "angle": 0, + "content": "tures, denoted as memory visual prompts, will be averaged together with the first frame's given prompt to construct the visual prompt of the current frame. Details of the memory visual prompt and ablations are in Appendix. D. By default, the memory length is set to 8. In Table 3, we conduct (interactive) video object segmentation evaluation on DAVIS17, DAVIS2016-Interactive, and Youtube-VOS 2018. The results of DAVIS2017 and Youtube-VOS 2018 indicate our model achieves better performance than SEEM and PersAM. In addition, \\(DINO\\nu\\) can also do interactive VOS, and our performance on DAVIS16-Interactive achieves significantly better performance compared with models not using video data for training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.182, + 0.619 + ], + "angle": 0, + "content": "3.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Effectiveness of Query Formulation. In Table 4, we ablate the effectiveness of using different query formulations for different tasks. The results indicate our double query formulation outperforms using only one type of query." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Effectiveness of Visual Prompt Formulation. In Table 5, we attempt to use a pre-trained CLIP vision encoder to encode the features of the visual prompt by cropping the prompted region into images for CLIP to process. As CLIP features contain rich semantics with few appearance features, which could not apply to referring segmentation tasks. Therefore, we ablate on generic segmentation tasks and find that the final model could not generalize well on open-set datasets like ADE. This result verifies our hypothesis that CLIP vision features could not generalize well on in-context visual prompting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Unifying Tasks and Data. We unify visual generic segmentation and visual referring segmentation to use both semantically labeled data (COCO) and data with" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.446 + ], + "angle": 0, + "content": "Table 4. Ablation of using difference queries to do both in-context reference and generic segmentation. By default, we use both generic query and interactive query. We remove one type of query at a time to ablate their effectiveness." + }, + { + "type": "table", + "bbox": [ + 0.498, + 0.457, + 0.897, + 0.531 + ], + "angle": 0, + "content": "
MethodCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinT49.642.747.058.073.371.075.7
only point query45.231.0(11.7)34.7(-12.3)52.771.468.874.0
only generic query46.238.3(-4.4)41.5(-6.0)53.368.966.571.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Table 5. Ablation of using different ways to encode the visual prompt on our Swin-T model. Under the same setting, we change our prompt encoding method and use a pre-trained CLIP to crop and encode the prompted objects in the image." + }, + { + "type": "table", + "bbox": [ + 0.498, + 0.611, + 0.897, + 0.67 + ], + "angle": 0, + "content": "
Prompt EncodingCOCO (in-domain)ADE (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoU
Ours49.642.747.058.019.411.412.821.9
CLIP48.540.743.554.912.61.41.313.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.734 + ], + "angle": 0, + "content": "only segmentation annotations (SA-1B). In Table 6, the results indicate that employing both datasets improves each individual task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Training batch size for generic segmentation. In Table 7, the results show that increasing training batch size consistently improves the generic segmentation performance. The reason for this phenomenon is that a larger batch size helps to sample more positive and negative visual in-context examples across different images, which better matches the inference setting with random visual examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inference In-Context Examples. In Fig. 6, we ablate the impact of using different in-context lengths. Increasing the in-context example exhibits diminishing returns, especially when the number of examples is more than eight." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12867" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.09, + 0.466, + 0.104 + ], + "angle": 0, + "content": "Table 6. Ablation of the effectiveness of unifying tasks and data." + }, + { + "type": "table", + "bbox": [ + 0.074, + 0.115, + 0.474, + 0.183 + ], + "angle": 0, + "content": "
MethodDataCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinTCOCO, SAM49.642.747.058.073.371.075.7
DINOv-SwinTCOCO48.941.745.957.163.360.865.7
DINOv-SwinTSAMN/A---68.466.070.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.185, + 0.447, + 0.2 + ], + "angle": 0, + "content": "Table 7. Ablation of image batchsize sampling in training." + }, + { + "type": "table", + "bbox": [ + 0.075, + 0.211, + 0.475, + 0.321 + ], + "angle": 0, + "content": "
Method#Batchsize for Prompt SamplingCOCO
PQmask APbox APmIoU
DINOv-SwinT128.923.225.333.7
DINOv-SwinT445.137.040.450.6
DINOv-SwinT847.339.243.153.1
DINOv-SwinT3247.840.344.156.2
DINOv-SwinT6449.045.241.557.0
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.347, + 0.429, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.555, + 0.432, + 0.57 + ], + "angle": 0, + "content": "Figure 6. DINOv query formulation of different tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.579, + 0.227, + 0.594 + ], + "angle": 0, + "content": "4. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.421, + 0.62 + ], + "angle": 0, + "content": "4.1. Visual Perception Through Text Prompt" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Innovations in open-vocabulary object detection [7, 12, 19, 24, 26, 43, 44, 47] and open-vocabulary segmentation [6, 8, 14, 31, 37, 44], have shown great potential in generic visual perception, by leveraging large pre-trained vision-language models like CLIP [30] and ALIGN [10]. These approaches demonstrate significant strides in zero-shot and few-shot performance, adapting to a variety of visual contexts through text prompts. However, the reliance on text alone introduces limitations due to linguistic ambiguity and the potential mismatch between textual descriptions and complex visual scenes [40]. This highlights the ongoing need to refine the integration of visual inputs for more accurate and comprehensive image perception." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.446, + 0.849 + ], + "angle": 0, + "content": "4.2. Visual Perception Through Image Example" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Building upon the foundations set by text-based visual perception methodologies, the field has seen a notable shift towards incorporating image examples to enhance accu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.319 + ], + "angle": 0, + "content": "racy and context sensitivity. OV-DETR [43] extends its open-vocabulary object detection capability beyond text, by utilizing both the image encoder and text encoder from CLIP [30], allowing for object detection guided by visual examples. Similarly, OWL-ViT [26] leverages large-scale image text examples in its contrastive pre-training phase, and propose to adopt image example for one-shot image-conditioned object detection. MQ-Det [40] utilizes image examples to enhance text descriptions for better open-vocabulary object detection performance. These methods typically adopt the image encoder in CLIP to extract visual features from given image examples for a more accurate perception of objects and scenes, and demonstrate that visual examples can bridge the gap between textual ambiguity and the complex nature of visual perception." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.859, + 0.344 + ], + "angle": 0, + "content": "4.3. Visual Perception Through Visual Prompt" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Different from image example-based methods that take an image as input, which are then processed by multi-modal encoder like CLIP [30], visual prompt-based methods typically use visual instructions (e.g. point, mask, scribble, and refereed regions of another image) to guide a model for a specific visual task. SAM [13], for instance, introduces a promotable model for interactive segmentation, fostering research in computer vision foundation models. SEEM [51] stands out as an interactive and versatile model for segmenting objects, accommodating various types of prompts, and is semantic-aware compared to SAM. Semantic-SAM [17] excels in semantic awareness and recognizing granularity, and can handle panoptic and part segmentation. SegGPT [34] unifies various segmentation tasks by formulating segmentation as an in-context coloring problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.618, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present DINOv, a unified framework for in-context visual prompting to accommodate both referring segmentation and generic segmentation tasks. To effectively formulate in-context visual prompts, we designed a simple prompt encoder to encoder reference visual prompts from the reference image and adopted a shared decoder to decode the final target visual prompts from the target image. We also formulate generic latent queries and point queries to align different tasks and data. The experimental results indicate that DINOv demonstrates impressive referring and generic segmentation capabilities to refer and detect with in-context visual prompting. Notably, DINOv delivers competitive performance compared to close-set segmentation on in-domain datasets and show promising results on many open-set segmentation benchmarks. We hope our early exploration of visual in-context prompting could inspire the community. Limitations. We employ limited semantically labeled data (COCO), which can be scaled up for better performance and extended to text prompts for multi-modal understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12868" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.253 + ], + "angle": 0, + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.472, + 0.324 + ], + "angle": 0, + "content": "[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.325, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[3] Ho Kei Cheng and Alexander G. Schwing. XMem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[4] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[5] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[6] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[7] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[8] Dat Huynh, Jason Kuen, Zhe Lin, Jiuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[9] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[11] Joakim Johnander, Martin Danelljan, Emil Brissman, Fahad Shahbaz Khan, and Michael Felsberg. A generative appearance model for end-to-end video object segmentation, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 2, 5, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[15] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 2, 5, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 6, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Huaijia Lin, Xiaojuan Qi, and Jiaya Jia. Agss-vos: Attention guided single-shot video object segmentation. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[22] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1362-1372, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12869" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.353, + 0.238 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-4 technical report, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.24, + 0.47, + 0.295 + ], + "angle": 0, + "content": "[28] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[32] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[33] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[34] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. arXiv preprint arXiv:2201.00487, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[36] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[37] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[39] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[40] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multimodal queried object detection in the wild. arXiv preprint arXiv:2305.18980, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[41] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. arXiv preprint arXiv:2303.06674, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[42] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. arXiv preprint arXiv:2308.02487, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[43] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[44] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[45] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[47] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[48] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[49] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022. 2, 6, 12" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12870" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[51] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 7, 8" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "12871" + } + ] +] \ No newline at end of file diff --git a/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_origin.pdf b/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..218b9056a7b8a56a28f2080559a8011cbffcf1bd --- /dev/null +++ b/2024/Visual In-Context Prompting/ecd3f7ab-166e-4fd1-8364-6864de752bd7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aef3a7a2554f540505b8500a4609f6e0921dba50ce2b78adcb23b3e5d31e0a2b +size 5267485 diff --git a/2024/Visual In-Context Prompting/full.md b/2024/Visual In-Context Prompting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8389d8cb30c42f924328fe7d440ef163cd13177 --- /dev/null +++ b/2024/Visual In-Context Prompting/full.md @@ -0,0 +1,281 @@ +# Visual In-Context Prompting + +Feng Li\*, Qing Jiang#, Hao Zhang#, Tianhe Ren†, Shilong Liu#, Xueyan Zou#, Huaizhe Xu#, Hongyang Li#, Jianwei Yang‡1, Chunyuan Li#, Lei Zhang†2, Jianfeng Gao‡2 + +$\spadesuit$ HKUST $\ddagger$ Microsoft Research, Redmond $\dagger$ IDEA $\sharp$ SCUT $\sharp$ Tsinghua $\S$ UW-Madison + +1. Project Lead 2. Equal Advisory Contribution + +![](images/59decd3dcd4ec115971bd04324fc4f2e92255dd3f0472c963e28a3be2da2ef44.jpg) + +# Visual Prompting Generic Segmentation + +![](images/3b0a5accbb7a6116badfa51cef6ddaeca37bb024a2f463b5f6bd14c33cc10722.jpg) + +# Visual Prompting Referring Segmentation + +![](images/6729c19fb7a6a9022f4850f2626c775613f7a51ad499e2b1327bbef7baccb160.jpg) +Figure 1. Our model $DINOV$ supports generic and referring segmentation to associate multiple or single objects with the user input visual prompts. A user can input one or more in-context visual prompts (scribbles, masks, boxes, etc.) to improve the segmentation performance. + +# Abstract + +In-context prompting in large language models (LLMs) has become a prevalent approach to improve zero-shot capabilities, but this idea is less explored in the vision domain. Existing visual prompting methods focus on referring segmentation to segment the most relevant object, falling short of addressing many generic vision tasks like open-set segmentation and detection. In this paper, we introduce a universal visual in-context prompting framework for both tasks, as shown in Fig. 1. In particular, we build on top of an encoder-decoder architecture, and develop a versatile prompt encoder to support a variety of prompts like strokes, boxes, and points. We further enhance it to take an arbitrary number of reference image + +segments as the context. Our extensive explorations show that the proposed visual in-context prompting elicits extraordinary referring and generic segmentation capabilities to refer and detect, yielding competitive performance to close-set in-domain datasets and showing promising results on many open-set segmentation datasets. By joint training on COCO and SA-1B, DINOv achieves 57.7 PQ on COCO and 23.2 PQ on ADE20K. Code will be available at https://github.com/UX-Decoder/DINOv + +# 1. Introduction + +The recent progress in large language models (LLMs) like GPT [1, 27] has shown promising results towards arti + +ficial general intelligence (AGI) by training unified models on large amounts of text data. These giant LLMs manifest themselves with intriguing emerging capabilities such as in-context learning. Nevertheless, similar paradigms have not yet succeeded in solving all vision tasks due to the diversity of scenarios in computer vision [15]. Some works [23, 49] have combined LLMs and vision models to tackle complex image understanding tasks with text outputs such as visual question answering, but challenges remain in fine-grained tasks that require pixel-level outputs, like instance masks, rather than just text. + +To this end, the community has observed a growing interest in the development of language-enhanced vision foundation models. These models demonstrate profound competencies in open-world visual understanding tasks using text prompts, encompassing areas like open-set detection [19, 24, 44] and segmentation [38, 42, 44, 50]. Visual prompt, a different prompting mechanism has been explored in some recent segmentation models [13, 17, 51]. In these works, different visual prompting formats (e.g., points, boxes and strokes, etc) have been explored to facilitate the segmentation of visual contents specified by users. + +Another critical LLM technique, in-context learning, has been less explored. In-context learning specifies the new task instruction using examples, and allows models to adapt to new tasks or domains without explicit retraining by providing relevant examples. One pioneering work in this area is SegGPT [34], which demonstrates the ability to output an image mask based on visual examples. However, these works focus on associating a user visual prompt with one most relevant object and have the limited ability to identify multiple objects of the same semantic concept. Therefore, these approaches fall short of addressing many generic vision tasks like open-set object detection and segmentation, which often require the segmentation of multiple objects of a given concept. On the other hand, while text-prompted vision models [24, 50] do not align with in-context learning methodologies, they exhibit notable flexibility in managing both referring and generic tasks in detection or segmentation, but with language prompts. In this paper, we strive to develop a model that supports visual in-context prompting for all types of image segmentation tasks. A comparison between our work and previous work is shown in Fig. 2. Besides supporting both single-image and cross-image visual prompting, our model distinguishes itself by effectively handling both referring and generic segmentation problems. + +To achieve this goal, we build a model called DINOv to support versatile visual prompting capabilities, based on the unified detection and segmentation model MaskDINO [18]. DINOv follows the general encoder-decoder design with an extra prompt encoder to formulate and sample visual prompts. The decoder takes in segmentation queries and reference prompt queries to generate segmentation masks + +![](images/cc9e0efdce78726f2416c72899fea74c1531f42c5e4f431a93b6fea73ba21bd0.jpg) +Figure 2. Comparison with related works. Generic: segment all objects of the same semantic concept that match the user prompt. Refer: segment a particular object with the user input visual prompts. Image prompt: crop the image regions as prompts. (single) Visual prompt: one image-prompt example to segment. In-context prompt: one or multiple image-prompt examples. We can do single-image and cross-image visual prompting tasks and support referring and generic segmentation. + +and target visual prompts, and we associate the output segmentation masks with the target prompt queries for the final output. We can define the visual in-context samples with a set of reference image $(Q)$ - visual prompt $(A)$ pairs. The visual prompt can be in various types, including mask, scribble, box, etc. With the in-context examples, our model takes in a target image and outputs the masks. The creation of target visual prompts involves a prompt encoder that extracts reference visual prompts from a Q-A pair. This is followed by a decoder to get the target visual prompt by attending reference visual prompts to the target image. During training, to construct positive and negative samples for generic segmentation, we sample reference visual prompts in a batch across different images. To address task and data discrepancies, we formulate generic latent queries and point queries for generic and referring segmentation, respectively. By joint training on COCO [21] and SA-1B [13] for generic and referring segmentation, our model attains competitive performance on in-domain segmentation tasks compared with text-prompted models and shows promising generalization capability on a wide range of open-set segmentation benchmarks using purely visual prompts. + +To summarize, our contributions are threefold: + +- We are the first to extend visual in-context prompting to support generic vision tasks like open-set generic segmentation and detection, and achieve comparable performance with text prompt-based open-set models. + +![](images/6c209b8377017c0adc0031b895dcb1ddf31ece7417925a88f798fc4639b7d47c.jpg) +(a) General Framework + +![](images/913b8314b27070f521cadf0a1098ce61fa718bd2cadf94a8607f1077c1bb0e02.jpg) +(b) Visual Generic Segmentation +(c) Visual Referring Segmentation +Figure 3. DINov is a universal segmentation framework that can do generic segmentation and referring image segmentation. The vision encoder is used to extract image features.(b) An illustration of losses for visual generic segmentation. In the example, there are 6 visual prompts sampled from 6 masks from 3 categories. The visual prompts from the instances of the same class are averaged as the class embedding. Each column of the matching matrix is a 3-dimension one-hot vector which is a one-hot class label of the instance; (c) An illustration of losses for visual referring segmentation. Each visual prompt is classified to one of the 6 instances. + +- We build $DINOV$ , a unified framework for referring segmentation and generic segmentation based on visual in-context prompting. This unification simplifies model design and allows our model to consume both semantically-labelled and unlabelled data for better performance. +- We conduct extensive experiments and visualizations to show that our model can handle generic, referring, and video object segmentation tasks. Our early attempts exhibit promising results on open-set segmentation and detection with visual prompting. + +# 2. Method + +# 2.1. Unified Formulation for Segmentation Tasks + +In this paper, we concentrate on visual prompting tasks involving images, encompassing both generic segmentation and referring segmentation tasks. Given $N$ reference images $\mathcal{I} = \{\mathbf{I}_1,\dots,\mathbf{I}_N\} \in \mathcal{R}^{N\times H\times W\times 3}$ with the corresponding visual prompts $\mathcal{P} = \{p_1,\dots,p_N\}$ , DINOv aims to segment objects of interest on a new target image $\mathbf{I}_{\mathbf{t}}$ . The visual prompts include masks, boxes, scribbles, points, etc. The interested objects can be a particular object for referring segmentation or all objects of the same semantic concept for generic segmentation. Note that the reference image can be identical to the target image, in which the task reduces to single-image visual prompting segmentation. + +To address these tasks, DINOv utilizes a comprehensive query-based encoder-decoder architecture. This archi- + +tecture comprises a vision encoder, denoted as Enc, responsible for extracting image features, a prompt encoder referred to as PromptEncoder, designed to extract visual prompt features by combining image features and user-provided visual prompts, and a general decoder represented as Decoder, which generates masks and visual concepts based on the segmentation query and visual prompt features. Upon receiving the input image and user-provided visual prompts, our initial step involves extracting image features denoted as Z using the vision encoder. Subsequently, we feed both the image features and visual prompts into the prompt encoder to extract the reference visual prompt $\mathcal{F}$ and subsequently sample the query visual prompt features $\mathbf{Q}_{\mathfrak{p}}$ . Formally, we have: + +$$ +\begin{array}{l} \mathcal {Z} = \operatorname {E n c} (\mathcal {I}), \mathbf {Z} = \operatorname {E n c} \left(\mathbf {I} _ {\mathbf {t}}\right) \\ \mathcal {F} = \operatorname {P r o m p t E n c o d e r} (\mathcal {P}, \mathcal {Z}) \tag {1} \\ \mathbf {Q} _ {\mathbf {p}} = \text {P r o m p t S a m p l e} (\mathcal {F}) \\ \end{array} +$$ + +In addition to the visual prompt features $\mathbf{Q}_{\mathbf{p}}$ , $DINOV$ incorporates segmentation queries $\mathbf{Q}_{\mathbf{s}}$ for proposal extraction. A shared decoder is employed to decode outputs for both $\mathbf{Q}_{\mathbf{s}}$ and $\mathbf{Q}_{\mathbf{p}}$ while performing cross-attention with respect to the target image feature $\mathbf{Z}$ . + +$$ +\begin{array}{l} \mathbf {O} _ {\mathbf {s}} = \operatorname {D e c o d e r} \left(\mathbf {Q} _ {\mathbf {s}}; \mathbf {Z}\right) \\ \begin{array}{l} \mathbf {O} _ {\mathbf {p}} = \operatorname {D e c o d e r} \left(\mathbf {Q} _ {\mathbf {p}}; \mathbf {Z}\right) \\ \text {P} = \text {M a t h e r m a l} (\mathbf {Q}) \end{array} \tag {2} \\ \langle \mathbf {M}, \mathbf {B} \rangle = \operatorname {M a s k H e a d} \left(\mathbf {O} _ {\mathbf {s}}\right) \\ \mathbf {C} _ {\mathbf {g}}, \mathbf {C} _ {\mathbf {r}} = \text {P r o m p t C l a s s i f i e r} (\mathbf {O} _ {\mathbf {s}}, \mathbf {O} _ {\mathbf {p}}) \\ \end{array} +$$ + +![](images/bafa3455c8142eadc2f331ab05134f3a49e81540315b4ec725dfcada581d2fee.jpg) +(a) Generic queries for generic segmentation + +![](images/00119a6e77dded9b2e17a4b356c013118e2ab27baf3573ca270fde19c5d649cf.jpg) +(b) Interactive point queries for referring segmentation + +![](images/7098ce3e9372dd7489996749c4924f44a92da542349a1d9657f399c38332c7f8.jpg) +Figure 4. DINOv query formulation of generic and referring segmentation tasks. +Figure 5. Prompt encoder to encode visual prompt from reference images. We use three masked cross-attention from the vision encoder small feature map to large feature map. + +Here, $\mathbf{O_s}$ represents the decoded segmentation query features, $\mathbf{O_p}$ corresponds to the decoded target visual prompt features, while $\mathbf{M}$ and $\mathbf{B}$ denote the predicted masks and boxes, respectively. Furthermore, we have $\mathbf{C_g}$ and $\mathbf{C_r}$ as the predicted matching scores for generic segmentation and referring segmentation tasks. These scores are derived through the use of a PromptClassifier, which computes the similarity between $\mathbf{O_s}$ and $\mathbf{O_p}$ . + +PromptClassifier. We clarify the definition of the prompt classifier, denoted as PromptClassifier $(\cdot ,\cdot)$ , for both generic segmentation and referring segmentation tasks here. In the case of generic segmentation tasks like instance and panoptic segmentation, the typical objective is to classify object features $\mathbf{O_s}$ into respective categories. When employing visual prompting for generic segmentation tasks, the distinction lies in the utilization of visual prompt features $\mathbf{O_p}$ as class embeddings. This is illustrated in the following equation: + +$$ +\mathbf {C} _ {\mathbf {g}} = \mathbf {g} \left(\mathbf {O} _ {\mathbf {s}}\right) \mathbf {g} \left(\mathbf {O} _ {\mathbf {p}} ^ {T}\right), \mathbf {C} _ {\mathbf {g}} \in N _ {p} \times N _ {s} \tag {3} +$$ + +where $N_{p}$ and $N_{s}$ are the number of visual prompts and object features. $g$ is the linear projection for generic segmentation task. Each of $N_{s}$ objects is classified into one of $N_{p}$ classes. For visual referring segmentation, the objective differs. Here, each visual prompt is employed to identify the most closely matched instance within the target image. This task can be framed as a classification problem, where each + +visual prompt is assigned to a specific instance within the target image. It's important to note that during our training phase, the target image and the reference image are identical. The matching score matrix for referring segmentation is structured as follows: + +$$ +\mathbf {C _ {r}} = \mathbf {h} (\mathbf {O _ {p}}) \mathbf {h} (\mathbf {O _ {s}} ^ {T}), \mathbf {C _ {r}} \in N _ {s} \times N _ {q} \tag {4} +$$ + +$h$ is the linear projection for referring segmentation task. Fig. 3(b) and (c) provide an illustrative representation of the two tasks. In our implementation, the generic segmentation task involves finding the most suitable visual prompt for each mask proposal, effectively pivoting the loss from a query to all prompts. Conversely, the referring segmentation task focuses on matching a given visual prompt to a specific mask proposal, with the loss pivot transitioning from a prompt to all proposals. As indicated in Equations 3 and 4, the PromptClassifier for both generic and referring segmentation tasks share a similar formulation. Consequently, they can share the entire framework, except for the two distinct linear layers denoted as $\mathbf{g}$ and $\mathbf{h}$ . + +# 2.2. Visual Prompt Formulation + +The heart part of our $DINO\nu$ is the proposed visual prompting mechanism. As shown in Eq. 1 and Eq. 2, we employ two modules to get the final visual prompt: + +- A PromptEncoder to encode the reference visual prompt $\mathcal{F}$ from the reference image features (followed by a sampling process to get query visual prompt $\mathbf{Q}_{\mathbf{p}}$ ). +- A Decoder (shared with the segmentation decoder) to decode outputs for the target visual prompt $\mathbf{O}_{\mathfrak{p}}$ by interacting with the target image features. + +This design allows our model to first encode the reference visual prompt and then adapt the prompt to the target image in a flexible way. As we attempt to express visual concepts through visual prompts, a straightforward way is to employ a pre-trained vision encoder (e.g., CLIP [29]) to process the reference images guided by user prompts [26]. However, it may encounter several challenges: $(i)$ the vision encoder takes cropped images as inputs, which causes substantial + +domain shift, especially for small objects [47]; $(ii)$ The visual features extracted from CLIP tend to be more semantic and may not meet the demands in VOS tasks. As we will show in our ablation study, employing a CLIP vision encoder to extract visual prompts has a clear inferior generalization ability. + +To address these issues, we reuse the vision encoder in our model and develop a simple yet effective prompt encoder. It extracts visual features corresponding to the locations indicated by various forms of visual prompts. To capture visual details of different granularities, we have incorporated multiple layers (default to 3) of the Mask Cross Attention Layer, as shown in Fig. 5. Each layer takes the image features extracted at different levels (output multi-scale features from the vision encoder, ranging from lower to higher resolutions) as inputs, utilizes the regions defined by the visual inputs as masks, and employs learnable queries to process the features at the corresponding positions to get the visual prompt features. + +# 2.3. Prompt Sampling + +We introduce two prompt sampling strategies tailored for referring segmentation and generic segmentation. + +Referring segmentation. In the case of referring segmentation, we employ a "self-referring" approach during training, wherein the reference image is identical to the target image. Here, we sample a prompt from an instance and train the model to refer to the same instance. This approach allows us to leverage extensive segmentation data, such as SA-1B, for training our model effectively. Despite being trained on the same instances, our model demonstrates the capability to perform cross-image referring during inference. As illustrated in Fig. 3(c), we can change the target images to various different images, enabling the model to effectively engage in cross-image referring tasks. + +Generic segmentation. The sampling strategies are slightly different during training and inference: + +- Training. In the training process, it is crucial to create both positive and negative visual prompt samples. To achieve this, we generate visual prompts by utilizing a large image training batch. As depicted in Algorithm 1, our approach begins by grouping together reference visual prompt $\mathcal{F}$ of the same semantic category across all images within a training batch. For each semantic category, we then randomly select a variable number of in-context examples, ranging from 1 to $N$ , and perform an aggregation process to generate reference visual prompt tokens $\mathbf{Q}_{\mathfrak{p}}$ , where each reference visual prompt token corresponds to a specific semantic category. $\mathbf{Q}_{\mathfrak{p}}$ is subsequently fed into the decoder, where it interacts with the target image to produce the final target visual prompt $\mathbf{O}_{\mathfrak{p}}$ . Consequently, the number of semantic categories corresponds to the number of target visual prompts. It's important to note that a given batch of + +images may not encompass all semantic categories present in the dataset, resulting in a variable number of semantic categories during the training process. + +- Inference. During the inference stage, using the COCO Dataset as an example, we pre-extract the respective visual prompt features based on mask prompts for all semantic categories established during the training phase. For evaluation purposes, we adopt a random selection approach, where we choose N features for each semantic category. By default, N is set to 16. These selected features act as representative visual prompt features for each category. This practice ensures that our inference stage maintains the same number of categories as in traditional open-set evaluation, effectively preventing any potential information leakage. + +We also provide the pseudo-code of prompting sampling in Appendix E. + +# 2.4. Decoder Query Formulation + +In $DINOv$ , we designed two types of segmentation queries to address two different tasks as depicted in Fig. 4. For generic segmentation, the query is a number of learnable ones similar to MaskDINO [16]. For the visual referring task, we adopt the interactive point query following Semantic-SAM [17], so that we can exploit the rich granularities in SA-1B [13]. Similar to Semantic-SAM, the visual prompts (points or boxes) are both converted into anchor box format, and then the position of each visual prompt will be encoded into position queries. Each position query is duplicated and subsequently combined with content queries of different granularities as the final segmentation queries. For the training on SA-1B, in order to avoid excessive computational overhead on the model, we selectively sample a subset of points contained within this visual concept as positive point queries. Concurrently, we randomly sample a subset of points from the remaining areas to serve as negative points. During the inference stage, we sample the initial point position queries on $20 \times 20$ uniformly distributed grid as the initial point position for a single frame. + +# 3. Experiments + +We jointly train on SA-1B [13] and COCO [21]. We evaluate on open-set segmentation and video object segmentation. We provide the detailed experiment setting, implementation details, and metrics in Appendix. A, B, and C. + +# 3.1. Generic Segmentation and Detection + +We evaluate our visual prompt based generic segmentation performance in Table 1. + +In-domain Segmentation on COCO. Compared to other models trained for visual prompts, we achieve significantly better results. For example, we surpass SegGPT [34] and Painter [33] by 14.3 PQ and 25.5 PQ. In addition, With just a few visual in-context prompts for each category, our + +Table 1. One suit of weights for generic visual in-context segmentation on multiple datasets. Our model is trained on COCO and SA-1B data. Note: “-” denotes the model does not have number reported or does not have the ability for the specific task. $\star$ means it is the test set results. † FC-CLIP adopts a frozen CLIP for open-set (text), we prompt the FC-CLIP with CLIP visual features to simulate visual promoting. # FC-CLIP and ODISE rely on frozen CLIP and Stable Diffusion knowledge. Mask DINO [18] is our baseline for comparison. + +
MethodSemantic DataTypeCOCO (in-domain)ADE (out-domain)SegInW (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoUAP-AverageAP-Median
Mask2Former-T [2]COCOClosed-set53.243.346.163.2-----
Mask2Former-B [2]COCO56.446.349.567.1-----
Mask2Former-L [2]COCO57.848.652.167.4-----
OneFormer-L [9]COCO57.948.9-67.4-----
MaskDINO-L [16]COCO58.350.656.267.5-----
Pano/SegFormer-B [36]COCO55.4--------
GLIPv2-H [45]COCO+O365+GOLDG+...Text Open-set-48.9*-------
MaskCLIP (L) [5]YFCC100M-----15.16.0-23.7
#ODISE-H [37]COCO (Stable diffusion))45.638.4-52.423.413.9-28.7-
#FC-CLIP-L [42]COCO (CLIP)54.444.6-63.726.816.8-34.1-
OpenSeed-T [44]COCO+O36555.447.652.064.019.814.117.022.933.921.5
X-Decoder-T [50]COCO+CC3M+...51.440.543.662.818.89.8-25.022.715.2
X-Decoder-L [50]COCO+CC3M+...56.946.7-67.521.813.1-29.636.138.7
OpenSeed-L [44]COCO+O36559.553.258.268.619.715.017.723.436.138.7
FC-CLIP†-L [42]COCOVisual Prompt----2.34.1-7.8-
SegGPT-L [34]COCO+ADE+VOC+...43.4--------
Painter-L [33]COCO+ADE+NYUv234.4--------
DINOv-T (Ours)COCO49.041.545.257.019.411.412.821.939.541.6
DINOv-L (Ours)COCO57.750.454.266.723.215.114.325.340.644.6
+ +Table 2. One suit of weights on ODinW benchmark. Average and median AP across 35 datasets are reported for simplicity. + +
ModelPretrain DataAverageMedian
MDETR [12]GOLDG, REFC10.73.0
GLIP-T [19]Object36511.41.6
OpenSeed (T) (Ours)Object365, COCO14.23.1
OpenSeed (L) (Ours)Object365, COCO15.25.0
DINOv (T) (Ours)COCO, SAM14.95.4
DINOv (L) (Ours)COCO, SAM15.74.78
+ +model achieves comparable results with previous close-set or open-set models on COCO. For example, the panoptic segmentation performance gap between DINOv and our baseline Mask DINO is only 0.6 PQ (57.7 PQ vs 58.3 PQ). Out-domain open-set segmentation on ADE20K. After training with visual prompt on COCO and SAM, we do zero-shot evaluation on ADE20K to validate its open-set segmentation capability when seeing novel visual concepts. To our best knowledge, it is the first time to use visual prompt for open-set segmentation. Compared with previous text-prompted open-set models, we achieve comparable or better performance with only COCO semantic data and no semantic knowledge from large pre-trained models. Especially, compared with our baseline OpenSeed, we achieve better performance with much fewer data. Note that FC-CLIP [42] employs a frozen CLIP to do text-based open-set segmentation. As the text and visual features are aligned in CLIP, we also attempt to prompt a pre-trained FC-CLIP with visual features from CLIP to test its open-set ability with visual prompts. However, its visual prompting performance largely lags behind its text-prompted results. Therefore, it is non-trivial to transfer a multi-modal text-based open-set model to do visual-prompted recognition well. The results indicate that visual prompts can generalize well to new concepts. + +Segmentation and detection in the wild. We also validate the generalization capability of visual prompting on some diversified and domain-specific datasets including SegInW and ODinW, which in total encompass more than 60 datasets. These datasets contain many real-scenario or rare categories. As these datasets all focus on instance-level segmentation, we report the average and median AP (AP-Average and AP-Median) over all datasets. We first evaluate the Segmentation in the Wild (SegInW) benchmark, which consists of 25 datasets. With visual prompting, $DI-NOv$ achieves a significant performance improvement over our baseline OpenSeed. For example, Our best AP-Average exceeds OpenSeed by 4.5 AP. We further evaluate Object Detection in the Wild (ODinW), which is composed of 35 datasets with bounding box annotations. As shown in Table 2, though we only employ much fewer semantically labeled data, we achieve better performance compared with previous models under similar settings. + +# 3.2. Video Object Segmentation + +Video object segmentation (VOS) aims to segment an interested object in a video by giving text or visual clues. Our model focuses on the semi-supervised setting, which segments a particular object throughout a video by giving visual clues in the first frame. In $DINOv$ , the visual prompt originates from one single image (generic/referring segmentation) or other images in one batch (generic segmentation). Therefore, our model has learned to prompt with visual features from other images. Therefore, $DINOv$ is able to do video object segmentation (VOS) by replacing current frame visual prompt features with previous frames. For more accurate tracking, we also store the visual features of the predicted mask in previous frames. These fea + +Table 3. Zero-shot video object segmentation. Without training with video or pairwise image data, our approach is able to do video object segmentation in a zero-shot manner. (#Concurrent work.) + +
MethodSegmentation DataTypeRefer-TypeZero-ShotDAVIS17DAVIS16-InteractiveYouTube-VOS 2018
JFJFJFJFGJsFsJuFu
With Video Data
AGSS [20]VOS+DAVISVideoMask67.464.969.9---71.371.365.575.273.1
AGAME [11](Synth)VOS+DAVISMask70.067.272.7---66.066.9*61.2*
SWEM [22]Image+VOS+DAVISMask84.381.287.4---82.882.486.977.185.0
XMem [3]Image+VOS+DAVISMask------86.185.189.880.389.2
SiamMask [32]COCO+VOSBox*54.358.569.871.767.8*60.258.245.147.7
MiVOS [4]BL30K+VOS+DAVISMask84.581.787.491.089.692.482.681.185.677.786.2
ReferFormer-B [35]RefCOCO(+/g)+VOS+DAVISText61.158.164.1---*****
UNINEXT-T [41]Image+VideoGeneralistMask74.571.377.6---77.076.881.070.879.4
UNINEXT-L [41]Image+VideoMask77.273.281.2---78.179.183.571.078.9
UNINEXT-L [41]Image+VideoText66.762.371.1---*****
Without Video Data
Painter-L [33]COCO+ADE+NYUv2GeneralistMask✓34.628.540.8---24.127.635.814.318.7
SegGPT-L [34]COCO+ADE+VOC+...Mask✓75.672.578.6---74.775.180.267.475.9
PerSAM-L [46]SAM+DAVISMask✘60.356.663.9---*****
SEEM-T [51]COCO+LVISMask✓60.457.663.362.758.966.451.455.644.159.246.9
SEEM-L [51]Mask✓58.955.062.862.258.366.050.057.238.261.343.3
DINOv-T (Ours)COCO+SAMMask✓73.371.075.777.072.981.260.965.370.052.357.9
DINOv-L (Ours)✓72.369.874.875.471.379.459.661.765.752.352.358.8
+ +tures, denoted as memory visual prompts, will be averaged together with the first frame's given prompt to construct the visual prompt of the current frame. Details of the memory visual prompt and ablations are in Appendix. D. By default, the memory length is set to 8. In Table 3, we conduct (interactive) video object segmentation evaluation on DAVIS17, DAVIS2016-Interactive, and Youtube-VOS 2018. The results of DAVIS2017 and Youtube-VOS 2018 indicate our model achieves better performance than SEEM and PersAM. In addition, $DINO\nu$ can also do interactive VOS, and our performance on DAVIS16-Interactive achieves significantly better performance compared with models not using video data for training. + +# 3.3. Ablation + +Effectiveness of Query Formulation. In Table 4, we ablate the effectiveness of using different query formulations for different tasks. The results indicate our double query formulation outperforms using only one type of query. + +Effectiveness of Visual Prompt Formulation. In Table 5, we attempt to use a pre-trained CLIP vision encoder to encode the features of the visual prompt by cropping the prompted region into images for CLIP to process. As CLIP features contain rich semantics with few appearance features, which could not apply to referring segmentation tasks. Therefore, we ablate on generic segmentation tasks and find that the final model could not generalize well on open-set datasets like ADE. This result verifies our hypothesis that CLIP vision features could not generalize well on in-context visual prompting. + +Effectiveness of Unifying Tasks and Data. We unify visual generic segmentation and visual referring segmentation to use both semantically labeled data (COCO) and data with + +Table 4. Ablation of using difference queries to do both in-context reference and generic segmentation. By default, we use both generic query and interactive query. We remove one type of query at a time to ablate their effectiveness. + +
MethodCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinT49.642.747.058.073.371.075.7
only point query45.231.0(11.7)34.7(-12.3)52.771.468.874.0
only generic query46.238.3(-4.4)41.5(-6.0)53.368.966.571.3
+ +Table 5. Ablation of using different ways to encode the visual prompt on our Swin-T model. Under the same setting, we change our prompt encoding method and use a pre-trained CLIP to crop and encode the prompted objects in the image. + +
Prompt EncodingCOCO (in-domain)ADE (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoU
Ours49.642.747.058.019.411.412.821.9
CLIP48.540.743.554.912.61.41.313.3
+ +only segmentation annotations (SA-1B). In Table 6, the results indicate that employing both datasets improves each individual task. + +Training batch size for generic segmentation. In Table 7, the results show that increasing training batch size consistently improves the generic segmentation performance. The reason for this phenomenon is that a larger batch size helps to sample more positive and negative visual in-context examples across different images, which better matches the inference setting with random visual examples. + +Inference In-Context Examples. In Fig. 6, we ablate the impact of using different in-context lengths. Increasing the in-context example exhibits diminishing returns, especially when the number of examples is more than eight. + +Table 6. Ablation of the effectiveness of unifying tasks and data. + +
MethodDataCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinTCOCO, SAM49.642.747.058.073.371.075.7
DINOv-SwinTCOCO48.941.745.957.163.360.865.7
DINOv-SwinTSAMN/A---68.466.070.8
+ +Table 7. Ablation of image batchsize sampling in training. + +
Method#Batchsize for Prompt SamplingCOCO
PQmask APbox APmIoU
DINOv-SwinT128.923.225.333.7
DINOv-SwinT445.137.040.450.6
DINOv-SwinT847.339.243.153.1
DINOv-SwinT3247.840.344.156.2
DINOv-SwinT6449.045.241.557.0
+ +![](images/ed78fa1f581541d252e0789c7a245441aba2661b4490d65021a301642e81dd1d.jpg) +Figure 6. DINOv query formulation of different tasks. + +# 4. Related Works + +# 4.1. Visual Perception Through Text Prompt + +Innovations in open-vocabulary object detection [7, 12, 19, 24, 26, 43, 44, 47] and open-vocabulary segmentation [6, 8, 14, 31, 37, 44], have shown great potential in generic visual perception, by leveraging large pre-trained vision-language models like CLIP [30] and ALIGN [10]. These approaches demonstrate significant strides in zero-shot and few-shot performance, adapting to a variety of visual contexts through text prompts. However, the reliance on text alone introduces limitations due to linguistic ambiguity and the potential mismatch between textual descriptions and complex visual scenes [40]. This highlights the ongoing need to refine the integration of visual inputs for more accurate and comprehensive image perception. + +# 4.2. Visual Perception Through Image Example + +Building upon the foundations set by text-based visual perception methodologies, the field has seen a notable shift towards incorporating image examples to enhance accu + +racy and context sensitivity. OV-DETR [43] extends its open-vocabulary object detection capability beyond text, by utilizing both the image encoder and text encoder from CLIP [30], allowing for object detection guided by visual examples. Similarly, OWL-ViT [26] leverages large-scale image text examples in its contrastive pre-training phase, and propose to adopt image example for one-shot image-conditioned object detection. MQ-Det [40] utilizes image examples to enhance text descriptions for better open-vocabulary object detection performance. These methods typically adopt the image encoder in CLIP to extract visual features from given image examples for a more accurate perception of objects and scenes, and demonstrate that visual examples can bridge the gap between textual ambiguity and the complex nature of visual perception. + +# 4.3. Visual Perception Through Visual Prompt + +Different from image example-based methods that take an image as input, which are then processed by multi-modal encoder like CLIP [30], visual prompt-based methods typically use visual instructions (e.g. point, mask, scribble, and refereed regions of another image) to guide a model for a specific visual task. SAM [13], for instance, introduces a promotable model for interactive segmentation, fostering research in computer vision foundation models. SEEM [51] stands out as an interactive and versatile model for segmenting objects, accommodating various types of prompts, and is semantic-aware compared to SAM. Semantic-SAM [17] excels in semantic awareness and recognizing granularity, and can handle panoptic and part segmentation. SegGPT [34] unifies various segmentation tasks by formulating segmentation as an in-context coloring problem. + +# 5. Conclusion + +We present DINOv, a unified framework for in-context visual prompting to accommodate both referring segmentation and generic segmentation tasks. To effectively formulate in-context visual prompts, we designed a simple prompt encoder to encoder reference visual prompts from the reference image and adopted a shared decoder to decode the final target visual prompts from the target image. We also formulate generic latent queries and point queries to align different tasks and data. The experimental results indicate that DINOv demonstrates impressive referring and generic segmentation capabilities to refer and detect with in-context visual prompting. Notably, DINOv delivers competitive performance compared to close-set segmentation on in-domain datasets and show promising results on many open-set segmentation benchmarks. We hope our early exploration of visual in-context prompting could inspire the community. Limitations. We employ limited semantically labeled data (COCO), which can be scaled up for better performance and extended to text prompts for multi-modal understanding. + +# References + +[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1 +[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 6 +[3] Ho Kei Cheng and Alexander G. Schwing. XMem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7 +[4] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 7 +[5] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022.6 +[6] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 8 +[7] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8 +[8] Dat Huynh, Jason Kuen, Zhe Lin, Jiuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 8 +[9] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022. 6 +[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8 +[11] Joakim Johnander, Martin Danelljan, Emil Brissman, Fahad Shahbaz Khan, and Michael Felsberg. A generative appearance model for end-to-end video object segmentation, 2018. 7 +[12] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 6, 8 + +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 2, 5, 8, 12 +[14] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 8 +[15] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2 +[16] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 5, 6, 12 +[17] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 2, 5, 8, 12 +[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2, 6 +[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 6, 8, 12 +[20] Huaijia Lin, Xiaojuan Qi, and Jiaya Jia. Agss-vos: Attention guided single-shot video object segmentation. In ICCV, 2019. 7 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 5, 12 +[22] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1362-1372, 2022. 7 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2 +[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 8 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 12 +[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 4, 8 +[27] OpenAI. Gpt-4 technical report, 2023. 1 +[28] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 12 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 8 +[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 8 +[32] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2019. 7 +[33] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning, 2023. 5, 6, 7 +[34] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2, 5, 6, 7, 8 +[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. arXiv preprint arXiv:2201.00487, 2022. 7 +[36] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 6 +[37] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 8 + +[38] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models, 2023. 2 +[39] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 12 +[40] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multimodal queried object detection in the wild. arXiv preprint arXiv:2305.18980, 2023. 8 +[41] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. arXiv preprint arXiv:2303.06674, 2023. 7 +[42] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. arXiv preprint arXiv:2308.02487, 2023. 2, 6 +[43] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8 +[44] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 6, 8 +[45] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6 +[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot, 2023. 7 +[47] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 5, 8 +[48] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 12 +[49] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2 +[50] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022. 2, 6, 12 + +[51] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 7, 8 \ No newline at end of file diff --git a/2024/Visual In-Context Prompting/images.zip b/2024/Visual In-Context Prompting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a68ec36205de3ab635477b35de0c9ba972d0dbe1 --- /dev/null +++ b/2024/Visual In-Context Prompting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1252d3bcd975825c19c00f49c3941727e3685415373a936cf63fcdf14ce23a3a +size 743738 diff --git a/2024/Visual In-Context Prompting/layout.json b/2024/Visual In-Context Prompting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c266412522139eb4b0fb4c5dcd52f31792f9b97 --- /dev/null +++ b/2024/Visual In-Context Prompting/layout.json @@ -0,0 +1,7526 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 206, + 104, + 388, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 104, + 388, + 121 + ], + "spans": [ + { + "bbox": [ + 206, + 104, + 388, + 121 + ], + "type": "text", + "content": "Visual In-Context Prompting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 142, + 546, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 546, + 173 + ], + "type": "text", + "content": "Feng Li\\*, Qing Jiang#, Hao Zhang#, Tianhe Ren†, Shilong Liu#, Xueyan Zou#, Huaizhe Xu#, Hongyang Li#, Jianwei Yang‡1, Chunyuan Li#, Lei Zhang†2, Jianfeng Gao‡2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "spans": [ + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "content": " HKUST " + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "content": " Microsoft Research, Redmond " + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "content": " IDEA " + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "content": " SCUT " + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "content": " Tsinghua " + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 140, + 178, + 450, + 191 + ], + "type": "text", + "content": " UW-Madison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 194, + 365, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 194, + 365, + 205 + ], + "spans": [ + { + "bbox": [ + 223, + 194, + 365, + 205 + ], + "type": "text", + "content": "1. Project Lead 2. Equal Advisory Contribution" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 71, + 219, + 520, + 294 + ], + "blocks": [ + { + "bbox": [ + 71, + 219, + 520, + 294 + ], + "lines": [ + { + "bbox": [ + 71, + 219, + 520, + 294 + ], + "spans": [ + { + "bbox": [ + 71, + 219, + 520, + 294 + ], + "type": "image", + "image_path": "59decd3dcd4ec115971bd04324fc4f2e92255dd3f0472c963e28a3be2da2ef44.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 297, + 394, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 297, + 394, + 310 + ], + "spans": [ + { + "bbox": [ + 197, + 297, + 394, + 310 + ], + "type": "text", + "content": "Visual Prompting Generic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 76, + 316, + 518, + 384 + ], + "blocks": [ + { + "bbox": [ + 76, + 316, + 518, + 384 + ], + "lines": [ + { + "bbox": [ + 76, + 316, + 518, + 384 + ], + "spans": [ + { + "bbox": [ + 76, + 316, + 518, + 384 + ], + "type": "image", + "image_path": "3b0a5accbb7a6116badfa51cef6ddaeca37bb024a2f463b5f6bd14c33cc10722.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 192, + 387, + 399, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 387, + 399, + 401 + ], + "spans": [ + { + "bbox": [ + 192, + 387, + 399, + 401 + ], + "type": "text", + "content": "Visual Prompting Referring Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 75, + 409, + 515, + 491 + ], + "blocks": [ + { + "bbox": [ + 75, + 409, + 515, + 491 + ], + "lines": [ + { + "bbox": [ + 75, + 409, + 515, + 491 + ], + "spans": [ + { + "bbox": [ + 75, + 409, + 515, + 491 + ], + "type": "image", + "image_path": "6729c19fb7a6a9022f4850f2626c775613f7a51ad499e2b1327bbef7baccb160.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 494, + 546, + 517 + ], + "lines": [ + { + "bbox": [ + 46, + 494, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 546, + 517 + ], + "type": "text", + "content": "Figure 1. Our model " + }, + { + "bbox": [ + 46, + 494, + 546, + 517 + ], + "type": "inline_equation", + "content": "DINOV" + }, + { + "bbox": [ + 46, + 494, + 546, + 517 + ], + "type": "text", + "content": " supports generic and referring segmentation to associate multiple or single objects with the user input visual prompts. A user can input one or more in-context visual prompts (scribbles, masks, boxes, etc.) to improve the segmentation performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 531, + 192, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 531, + 192, + 544 + ], + "spans": [ + { + "bbox": [ + 143, + 531, + 192, + 544 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 558, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 558, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 558, + 290, + 715 + ], + "type": "text", + "content": "In-context prompting in large language models (LLMs) has become a prevalent approach to improve zero-shot capabilities, but this idea is less explored in the vision domain. Existing visual prompting methods focus on referring segmentation to segment the most relevant object, falling short of addressing many generic vision tasks like open-set segmentation and detection. In this paper, we introduce a universal visual in-context prompting framework for both tasks, as shown in Fig. 1. In particular, we build on top of an encoder-decoder architecture, and develop a versatile prompt encoder to support a variety of prompts like strokes, boxes, and points. We further enhance it to take an arbitrary number of reference image" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 532, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 640 + ], + "type": "text", + "content": "segments as the context. Our extensive explorations show that the proposed visual in-context prompting elicits extraordinary referring and generic segmentation capabilities to refer and detect, yielding competitive performance to close-set in-domain datasets and showing promising results on many open-set segmentation datasets. By joint training on COCO and SA-1B, DINOv achieves 57.7 PQ on COCO and 23.2 PQ on ADE20K. Code will be available at https://github.com/UX-Decoder/DINOv" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 667, + 387, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 667, + 387, + 680 + ], + "spans": [ + { + "bbox": [ + 306, + 667, + 387, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "The recent progress in large language models (LLMs) like GPT [1, 27] has shown promising results towards arti" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12861" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "ficial general intelligence (AGI) by training unified models on large amounts of text data. These giant LLMs manifest themselves with intriguing emerging capabilities such as in-context learning. Nevertheless, similar paradigms have not yet succeeded in solving all vision tasks due to the diversity of scenarios in computer vision [15]. Some works [23, 49] have combined LLMs and vision models to tackle complex image understanding tasks with text outputs such as visual question answering, but challenges remain in fine-grained tasks that require pixel-level outputs, like instance masks, rather than just text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 339 + ], + "type": "text", + "content": "To this end, the community has observed a growing interest in the development of language-enhanced vision foundation models. These models demonstrate profound competencies in open-world visual understanding tasks using text prompts, encompassing areas like open-set detection [19, 24, 44] and segmentation [38, 42, 44, 50]. Visual prompt, a different prompting mechanism has been explored in some recent segmentation models [13, 17, 51]. In these works, different visual prompting formats (e.g., points, boxes and strokes, etc) have been explored to facilitate the segmentation of visual contents specified by users." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 628 + ], + "type": "text", + "content": "Another critical LLM technique, in-context learning, has been less explored. In-context learning specifies the new task instruction using examples, and allows models to adapt to new tasks or domains without explicit retraining by providing relevant examples. One pioneering work in this area is SegGPT [34], which demonstrates the ability to output an image mask based on visual examples. However, these works focus on associating a user visual prompt with one most relevant object and have the limited ability to identify multiple objects of the same semantic concept. Therefore, these approaches fall short of addressing many generic vision tasks like open-set object detection and segmentation, which often require the segmentation of multiple objects of a given concept. On the other hand, while text-prompted vision models [24, 50] do not align with in-context learning methodologies, they exhibit notable flexibility in managing both referring and generic tasks in detection or segmentation, but with language prompts. In this paper, we strive to develop a model that supports visual in-context prompting for all types of image segmentation tasks. A comparison between our work and previous work is shown in Fig. 2. Besides supporting both single-image and cross-image visual prompting, our model distinguishes itself by effectively handling both referring and generic segmentation problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": "To achieve this goal, we build a model called DINOv to support versatile visual prompting capabilities, based on the unified detection and segmentation model MaskDINO [18]. DINOv follows the general encoder-decoder design with an extra prompt encoder to formulate and sample visual prompts. The decoder takes in segmentation queries and reference prompt queries to generate segmentation masks" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 544, + 279 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 544, + 279 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 544, + 279 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 544, + 279 + ], + "type": "image", + "image_path": "cc9e0efdce78726f2416c72899fea74c1531f42c5e4f431a93b6fea73ba21bd0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 285, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 374 + ], + "type": "text", + "content": "Figure 2. Comparison with related works. Generic: segment all objects of the same semantic concept that match the user prompt. Refer: segment a particular object with the user input visual prompts. Image prompt: crop the image regions as prompts. (single) Visual prompt: one image-prompt example to segment. In-context prompt: one or multiple image-prompt examples. We can do single-image and cross-image visual prompting tasks and support referring and generic segmentation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "type": "text", + "content": "and target visual prompts, and we associate the output segmentation masks with the target prompt queries for the final output. We can define the visual in-context samples with a set of reference image " + }, + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "type": "text", + "content": " - visual prompt " + }, + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 304, + 387, + 545, + 650 + ], + "type": "text", + "content": " pairs. The visual prompt can be in various types, including mask, scribble, box, etc. With the in-context examples, our model takes in a target image and outputs the masks. The creation of target visual prompts involves a prompt encoder that extracts reference visual prompts from a Q-A pair. This is followed by a decoder to get the target visual prompt by attending reference visual prompts to the target image. During training, to construct positive and negative samples for generic segmentation, we sample reference visual prompts in a batch across different images. To address task and data discrepancies, we formulate generic latent queries and point queries for generic and referring segmentation, respectively. By joint training on COCO [21] and SA-1B [13] for generic and referring segmentation, our model attains competitive performance on in-domain segmentation tasks compared with text-prompted models and shows promising generalization capability on a wide range of open-set segmentation benchmarks using purely visual prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 651, + 506, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 651, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 317, + 651, + 506, + 662 + ], + "type": "text", + "content": "To summarize, our contributions are threefold:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 665, + 547, + 713 + ], + "type": "text", + "content": "- We are the first to extend visual in-context prompting to support generic vision tasks like open-set generic segmentation and detection, and achieve comparable performance with text prompt-based open-set models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12862" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 76, + 280, + 252 + ], + "blocks": [ + { + "bbox": [ + 51, + 76, + 280, + 252 + ], + "lines": [ + { + "bbox": [ + 51, + 76, + 280, + 252 + ], + "spans": [ + { + "bbox": [ + 51, + 76, + 280, + 252 + ], + "type": "image", + "image_path": "6c209b8377017c0adc0031b895dcb1ddf31ece7417925a88f798fc4639b7d47c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 258, + 205, + 267 + ], + "lines": [ + { + "bbox": [ + 124, + 258, + 205, + 267 + ], + "spans": [ + { + "bbox": [ + 124, + 258, + 205, + 267 + ], + "type": "text", + "content": "(a) General Framework" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 289, + 73, + 545, + 245 + ], + "blocks": [ + { + "bbox": [ + 289, + 73, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 289, + 73, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 289, + 73, + 545, + 245 + ], + "type": "image", + "image_path": "913b8314b27070f521cadf0a1098ce61fa718bd2cadf94a8607f1077c1bb0e02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 257, + 410, + 267 + ], + "lines": [ + { + "bbox": [ + 298, + 257, + 410, + 267 + ], + "spans": [ + { + "bbox": [ + 298, + 257, + 410, + 267 + ], + "type": "text", + "content": "(b) Visual Generic Segmentation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 257, + 529, + 267 + ], + "lines": [ + { + "bbox": [ + 414, + 257, + 529, + 267 + ], + "spans": [ + { + "bbox": [ + 414, + 257, + 529, + 267 + ], + "type": "text", + "content": "(c) Visual Referring Segmentation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 278, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 334 + ], + "type": "text", + "content": "Figure 3. DINov is a universal segmentation framework that can do generic segmentation and referring image segmentation. The vision encoder is used to extract image features.(b) An illustration of losses for visual generic segmentation. In the example, there are 6 visual prompts sampled from 6 masks from 3 categories. The visual prompts from the instances of the same class are averaged as the class embedding. Each column of the matching matrix is a 3-dimension one-hot vector which is a one-hot class label of the instance; (c) An illustration of losses for visual referring segmentation. Each visual prompt is classified to one of the 6 instances." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 365, + 287, + 502 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 365, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 365, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 58, + 365, + 287, + 437 + ], + "type": "text", + "content": "- We build " + }, + { + "bbox": [ + 58, + 365, + 287, + 437 + ], + "type": "inline_equation", + "content": "DINOV" + }, + { + "bbox": [ + 58, + 365, + 287, + 437 + ], + "type": "text", + "content": ", a unified framework for referring segmentation and generic segmentation based on visual in-context prompting. This unification simplifies model design and allows our model to consume both semantically-labelled and unlabelled data for better performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 441, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 441, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 58, + 441, + 287, + 502 + ], + "type": "text", + "content": "- We conduct extensive experiments and visualizations to show that our model can handle generic, referring, and video object segmentation tasks. Our early attempts exhibit promising results on open-set segmentation and detection with visual prompting." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 506, + 103, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 103, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 103, + 518 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 526, + 277, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 277, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 277, + 539 + ], + "type": "text", + "content": "2.1. Unified Formulation for Segmentation Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": "In this paper, we concentrate on visual prompting tasks involving images, encompassing both generic segmentation and referring segmentation tasks. Given " + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": " reference images " + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathbf{I}_1,\\dots,\\mathbf{I}_N\\} \\in \\mathcal{R}^{N\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": " with the corresponding visual prompts " + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_1,\\dots,p_N\\}" + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": ", DINOv aims to segment objects of interest on a new target image " + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{t}}" + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": ". The visual prompts include masks, boxes, scribbles, points, etc. The interested objects can be a particular object for referring segmentation or all objects of the same semantic concept for generic segmentation. Note that the reference image can be identical to the target image, in which the task reduces to single-image visual prompting segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "To address these tasks, DINOv utilizes a comprehensive query-based encoder-decoder architecture. This archi-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "type": "text", + "content": "tecture comprises a vision encoder, denoted as Enc, responsible for extracting image features, a prompt encoder referred to as PromptEncoder, designed to extract visual prompt features by combining image features and user-provided visual prompts, and a general decoder represented as Decoder, which generates masks and visual concepts based on the segmentation query and visual prompt features. Upon receiving the input image and user-provided visual prompts, our initial step involves extracting image features denoted as Z using the vision encoder. Subsequently, we feed both the image features and visual prompts into the prompt encoder to extract the reference visual prompt " + }, + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "type": "text", + "content": " and subsequently sample the query visual prompt features " + }, + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 304, + 350, + 547, + 518 + ], + "type": "text", + "content": ". Formally, we have:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 529, + 545, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 529, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 357, + 529, + 545, + 573 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {Z} = \\operatorname {E n c} (\\mathcal {I}), \\mathbf {Z} = \\operatorname {E n c} \\left(\\mathbf {I} _ {\\mathbf {t}}\\right) \\\\ \\mathcal {F} = \\operatorname {P r o m p t E n c o d e r} (\\mathcal {P}, \\mathcal {Z}) \\tag {1} \\\\ \\mathbf {Q} _ {\\mathbf {p}} = \\text {P r o m p t S a m p l e} (\\mathcal {F}) \\\\ \\end{array}", + "image_path": "ad5722f6417e5d720d3d7aca80abce52da54067e0deb61bb3f697dce156e4f3b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": "In addition to the visual prompt features " + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "inline_equation", + "content": "DINOV" + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": " incorporates segmentation queries " + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{s}}" + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": " for proposal extraction. A shared decoder is employed to decode outputs for both " + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{s}}" + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": " while performing cross-attention with respect to the target image feature " + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 584, + 547, + 644 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 657, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 657, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 342, + 657, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {O} _ {\\mathbf {s}} = \\operatorname {D e c o d e r} \\left(\\mathbf {Q} _ {\\mathbf {s}}; \\mathbf {Z}\\right) \\\\ \\begin{array}{l} \\mathbf {O} _ {\\mathbf {p}} = \\operatorname {D e c o d e r} \\left(\\mathbf {Q} _ {\\mathbf {p}}; \\mathbf {Z}\\right) \\\\ \\text {P} = \\text {M a t h e r m a l} (\\mathbf {Q}) \\end{array} \\tag {2} \\\\ \\langle \\mathbf {M}, \\mathbf {B} \\rangle = \\operatorname {M a s k H e a d} \\left(\\mathbf {O} _ {\\mathbf {s}}\\right) \\\\ \\mathbf {C} _ {\\mathbf {g}}, \\mathbf {C} _ {\\mathbf {r}} = \\text {P r o m p t C l a s s i f i e r} (\\mathbf {O} _ {\\mathbf {s}}, \\mathbf {O} _ {\\mathbf {p}}) \\\\ \\end{array}", + "image_path": "35e9f54e539e8f9dae3e30e2dc3e76394c8585cde13846cf5e1f932e0796a187.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12863" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 78, + 188, + 148 + ], + "blocks": [ + { + "bbox": [ + 60, + 78, + 188, + 148 + ], + "lines": [ + { + "bbox": [ + 60, + 78, + 188, + 148 + ], + "spans": [ + { + "bbox": [ + 60, + 78, + 188, + 148 + ], + "type": "image", + "image_path": "bafa3455c8142eadc2f331ab05134f3a49e81540315b4ec725dfcada581d2fee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 153, + 169, + 177 + ], + "lines": [ + { + "bbox": [ + 78, + 153, + 169, + 177 + ], + "spans": [ + { + "bbox": [ + 78, + 153, + 169, + 177 + ], + "type": "text", + "content": "(a) Generic queries for generic segmentation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 76, + 539, + 153 + ], + "blocks": [ + { + "bbox": [ + 205, + 76, + 539, + 153 + ], + "lines": [ + { + "bbox": [ + 205, + 76, + 539, + 153 + ], + "spans": [ + { + "bbox": [ + 205, + 76, + 539, + 153 + ], + "type": "image", + "image_path": "00119a6e77dded9b2e17a4b356c013118e2ab27baf3573ca270fde19c5d649cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 159, + 474, + 172 + ], + "lines": [ + { + "bbox": [ + 258, + 159, + 474, + 172 + ], + "spans": [ + { + "bbox": [ + 258, + 159, + 474, + 172 + ], + "type": "text", + "content": "(b) Interactive point queries for referring segmentation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 216, + 248, + 334 + ], + "blocks": [ + { + "bbox": [ + 150, + 186, + 441, + 198 + ], + "lines": [ + { + "bbox": [ + 150, + 186, + 441, + 198 + ], + "spans": [ + { + "bbox": [ + 150, + 186, + 441, + 198 + ], + "type": "text", + "content": "Figure 4. DINOv query formulation of generic and referring segmentation tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 216, + 248, + 334 + ], + "lines": [ + { + "bbox": [ + 67, + 216, + 248, + 334 + ], + "spans": [ + { + "bbox": [ + 67, + 216, + 248, + 334 + ], + "type": "image", + "image_path": "7098ce3e9372dd7489996749c4924f44a92da542349a1d9657f399c38332c7f8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 378 + ], + "type": "text", + "content": "Figure 5. Prompt encoder to encode visual prompt from reference images. We use three masked cross-attention from the vision encoder small feature map to large feature map." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{O_s}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " represents the decoded segmentation query features, " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{O_p}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " corresponds to the decoded target visual prompt features, while " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " denote the predicted masks and boxes, respectively. Furthermore, we have " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{C_g}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{C_r}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " as the predicted matching scores for generic segmentation and referring segmentation tasks. These scores are derived through the use of a PromptClassifier, which computes the similarity between " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{O_s}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{O_p}" + }, + { + "bbox": [ + 46, + 385, + 287, + 481 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "content": "PromptClassifier. We clarify the definition of the prompt classifier, denoted as PromptClassifier " + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "inline_equation", + "content": "(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "content": ", for both generic segmentation and referring segmentation tasks here. In the case of generic segmentation tasks like instance and panoptic segmentation, the typical objective is to classify object features " + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{O_s}" + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "content": " into respective categories. When employing visual prompting for generic segmentation tasks, the distinction lies in the utilization of visual prompt features " + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{O_p}" + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "content": " as class embeddings. This is illustrated in the following equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 608, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 608, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 88, + 608, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {g}} = \\mathbf {g} \\left(\\mathbf {O} _ {\\mathbf {s}}\\right) \\mathbf {g} \\left(\\mathbf {O} _ {\\mathbf {p}} ^ {T}\\right), \\mathbf {C} _ {\\mathbf {g}} \\in N _ {p} \\times N _ {s} \\tag {3}", + "image_path": "6506ff7bf3aca9005b11625ee667b0759ec4cb5c34b93b9df527bc19c220cc61.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": " are the number of visual prompts and object features. " + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": " is the linear projection for generic segmentation task. Each of " + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": " objects is classified into one of " + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": " classes. For visual referring segmentation, the objective differs. Here, each visual prompt is employed to identify the most closely matched instance within the target image. This task can be framed as a classification problem, where each" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 214, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 274 + ], + "type": "text", + "content": "visual prompt is assigned to a specific instance within the target image. It's important to note that during our training phase, the target image and the reference image are identical. The matching score matrix for referring segmentation is structured as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 283, + 545, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 283, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 347, + 283, + 545, + 299 + ], + "type": "interline_equation", + "content": "\\mathbf {C _ {r}} = \\mathbf {h} (\\mathbf {O _ {p}}) \\mathbf {h} (\\mathbf {O _ {s}} ^ {T}), \\mathbf {C _ {r}} \\in N _ {s} \\times N _ {q} \\tag {4}", + "image_path": "d81e0cbde1bd8499e0c3df936df1b95e3eec23d32ffb0cf4a6e491ecc8bc8b01.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "text", + "content": " is the linear projection for referring segmentation task. Fig. 3(b) and (c) provide an illustrative representation of the two tasks. In our implementation, the generic segmentation task involves finding the most suitable visual prompt for each mask proposal, effectively pivoting the loss from a query to all prompts. Conversely, the referring segmentation task focuses on matching a given visual prompt to a specific mask proposal, with the loss pivot transitioning from a prompt to all proposals. As indicated in Equations 3 and 4, the PromptClassifier for both generic and referring segmentation tasks share a similar formulation. Consequently, they can share the entire framework, except for the two distinct linear layers denoted as " + }, + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 304, + 307, + 546, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "type": "text", + "content": "2.2. Visual Prompt Formulation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "type": "text", + "content": "The heart part of our " + }, + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "type": "inline_equation", + "content": "DINO\\nu" + }, + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "type": "text", + "content": " is the proposed visual prompting mechanism. As shown in Eq. 1 and Eq. 2, we employ two modules to get the final visual prompt:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 529, + 545, + 613 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "type": "text", + "content": "- A PromptEncoder to encode the reference visual prompt " + }, + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "type": "text", + "content": " from the reference image features (followed by a sampling process to get query visual prompt " + }, + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{p}}" + }, + { + "bbox": [ + 317, + 529, + 545, + 567 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 578, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 545, + 613 + ], + "type": "text", + "content": "- A Decoder (shared with the segmentation decoder) to decode outputs for the target visual prompt " + }, + { + "bbox": [ + 317, + 578, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 317, + 578, + 545, + 613 + ], + "type": "text", + "content": " by interacting with the target image features." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "This design allows our model to first encode the reference visual prompt and then adapt the prompt to the target image in a flexible way. As we attempt to express visual concepts through visual prompts, a straightforward way is to employ a pre-trained vision encoder (e.g., CLIP [29]) to process the reference images guided by user prompts [26]. However, it may encounter several challenges: " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " the vision encoder takes cropped images as inputs, which causes substantial" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12864" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "domain shift, especially for small objects [47]; " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " The visual features extracted from CLIP tend to be more semantic and may not meet the demands in VOS tasks. As we will show in our ablation study, employing a CLIP vision encoder to extract visual prompts has a clear inferior generalization ability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 290, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 290, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 290, + 301 + ], + "type": "text", + "content": "To address these issues, we reuse the vision encoder in our model and develop a simple yet effective prompt encoder. It extracts visual features corresponding to the locations indicated by various forms of visual prompts. To capture visual details of different granularities, we have incorporated multiple layers (default to 3) of the Mask Cross Attention Layer, as shown in Fig. 5. Each layer takes the image features extracted at different levels (output multi-scale features from the vision encoder, ranging from lower to higher resolutions) as inputs, utilizes the regions defined by the visual inputs as masks, and employs learnable queries to process the features at the corresponding positions to get the visual prompt features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 312, + 153, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 153, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 153, + 324 + ], + "type": "text", + "content": "2.3. Prompt Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "content": "We introduce two prompt sampling strategies tailored for referring segmentation and generic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "text", + "content": "Referring segmentation. In the case of referring segmentation, we employ a \"self-referring\" approach during training, wherein the reference image is identical to the target image. Here, we sample a prompt from an instance and train the model to refer to the same instance. This approach allows us to leverage extensive segmentation data, such as SA-1B, for training our model effectively. Despite being trained on the same instances, our model demonstrates the capability to perform cross-image referring during inference. As illustrated in Fig. 3(c), we can change the target images to various different images, enabling the model to effectively engage in cross-image referring tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "text", + "content": "Generic segmentation. The sampling strategies are slightly different during training and inference:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "- Training. In the training process, it is crucial to create both positive and negative visual prompt samples. To achieve this, we generate visual prompts by utilizing a large image training batch. As depicted in Algorithm 1, our approach begins by grouping together reference visual prompt " + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": " of the same semantic category across all images within a training batch. For each semantic category, we then randomly select a variable number of in-context examples, ranging from 1 to " + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": ", and perform an aggregation process to generate reference visual prompt tokens " + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": ", where each reference visual prompt token corresponds to a specific semantic category. " + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": " is subsequently fed into the decoder, where it interacts with the target image to produce the final target visual prompt " + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": ". Consequently, the number of semantic categories corresponds to the number of target visual prompts. It's important to note that a given batch of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "images may not encompass all semantic categories present in the dataset, resulting in a variable number of semantic categories during the training process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 239 + ], + "type": "text", + "content": "- Inference. During the inference stage, using the COCO Dataset as an example, we pre-extract the respective visual prompt features based on mask prompts for all semantic categories established during the training phase. For evaluation purposes, we adopt a random selection approach, where we choose N features for each semantic category. By default, N is set to 16. These selected features act as representative visual prompt features for each category. This practice ensures that our inference stage maintains the same number of categories as in traditional open-set evaluation, effectively preventing any potential information leakage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 239, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 264 + ], + "type": "text", + "content": "We also provide the pseudo-code of prompting sampling in Appendix E." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 270, + 463, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 463, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 463, + 283 + ], + "type": "text", + "content": "2.4. Decoder Query Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "type": "inline_equation", + "content": "DINOv" + }, + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "type": "text", + "content": ", we designed two types of segmentation queries to address two different tasks as depicted in Fig. 4. For generic segmentation, the query is a number of learnable ones similar to MaskDINO [16]. For the visual referring task, we adopt the interactive point query following Semantic-SAM [17], so that we can exploit the rich granularities in SA-1B [13]. Similar to Semantic-SAM, the visual prompts (points or boxes) are both converted into anchor box format, and then the position of each visual prompt will be encoded into position queries. Each position query is duplicated and subsequently combined with content queries of different granularities as the final segmentation queries. For the training on SA-1B, in order to avoid excessive computational overhead on the model, we selectively sample a subset of points contained within this visual concept as positive point queries. Concurrently, we randomly sample a subset of points from the remaining areas to serve as negative points. During the inference stage, we sample the initial point position queries on " + }, + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "type": "inline_equation", + "content": "20 \\times 20" + }, + { + "bbox": [ + 304, + 289, + 547, + 529 + ], + "type": "text", + "content": " uniformly distributed grid as the initial point position for a single frame." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 537, + 388, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 388, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 388, + 551 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 605 + ], + "type": "text", + "content": "We jointly train on SA-1B [13] and COCO [21]. We evaluate on open-set segmentation and video object segmentation. We provide the detailed experiment setting, implementation details, and metrics in Appendix. A, B, and C." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 611, + 500, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 500, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 500, + 624 + ], + "type": "text", + "content": "3.1. Generic Segmentation and Detection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 653 + ], + "type": "text", + "content": "We evaluate our visual prompt based generic segmentation performance in Table 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "In-domain Segmentation on COCO. Compared to other models trained for visual prompts, we achieve significantly better results. For example, we surpass SegGPT [34] and Painter [33] by 14.3 PQ and 25.5 PQ. In addition, With just a few visual in-context prompts for each category, our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12865" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 125, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "text", + "content": "Table 1. One suit of weights for generic visual in-context segmentation on multiple datasets. Our model is trained on COCO and SA-1B data. Note: “-” denotes the model does not have number reported or does not have the ability for the specific task. " + }, + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "text", + "content": " means it is the test set results. † FC-CLIP adopts a frozen CLIP for open-set (text), we prompt the FC-CLIP with CLIP visual features to simulate visual promoting. # FC-CLIP and ODISE rely on frozen CLIP and Stable Diffusion knowledge. Mask DINO [18] is our baseline for comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 125, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 49, + 125, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 49, + 125, + 545, + 289 + ], + "type": "table", + "html": "
MethodSemantic DataTypeCOCO (in-domain)ADE (out-domain)SegInW (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoUAP-AverageAP-Median
Mask2Former-T [2]COCOClosed-set53.243.346.163.2-----
Mask2Former-B [2]COCO56.446.349.567.1-----
Mask2Former-L [2]COCO57.848.652.167.4-----
OneFormer-L [9]COCO57.948.9-67.4-----
MaskDINO-L [16]COCO58.350.656.267.5-----
Pano/SegFormer-B [36]COCO55.4--------
GLIPv2-H [45]COCO+O365+GOLDG+...Text Open-set-48.9*-------
MaskCLIP (L) [5]YFCC100M-----15.16.0-23.7
#ODISE-H [37]COCO (Stable diffusion))45.638.4-52.423.413.9-28.7-
#FC-CLIP-L [42]COCO (CLIP)54.444.6-63.726.816.8-34.1-
OpenSeed-T [44]COCO+O36555.447.652.064.019.814.117.022.933.921.5
X-Decoder-T [50]COCO+CC3M+...51.440.543.662.818.89.8-25.022.715.2
X-Decoder-L [50]COCO+CC3M+...56.946.7-67.521.813.1-29.636.138.7
OpenSeed-L [44]COCO+O36559.553.258.268.619.715.017.723.436.138.7
FC-CLIP†-L [42]COCOVisual Prompt----2.34.1-7.8-
SegGPT-L [34]COCO+ADE+VOC+...43.4--------
Painter-L [33]COCO+ADE+NYUv234.4--------
DINOv-T (Ours)COCO49.041.545.257.019.411.412.821.939.541.6
DINOv-L (Ours)COCO57.750.454.266.723.215.114.325.340.644.6
", + "image_path": "4abcddf3c3c52b134d99c75137ecb2e5e862b6b387b08066c2d64900c19472f4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 54, + 337, + 281, + 410 + ], + "blocks": [ + { + "bbox": [ + 47, + 306, + 288, + 331 + ], + "lines": [ + { + "bbox": [ + 47, + 306, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 288, + 331 + ], + "type": "text", + "content": "Table 2. One suit of weights on ODinW benchmark. Average and median AP across 35 datasets are reported for simplicity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 337, + 281, + 410 + ], + "lines": [ + { + "bbox": [ + 54, + 337, + 281, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 281, + 410 + ], + "type": "table", + "html": "
ModelPretrain DataAverageMedian
MDETR [12]GOLDG, REFC10.73.0
GLIP-T [19]Object36511.41.6
OpenSeed (T) (Ours)Object365, COCO14.23.1
OpenSeed (L) (Ours)Object365, COCO15.25.0
DINOv (T) (Ours)COCO, SAM14.95.4
DINOv (L) (Ours)COCO, SAM15.74.78
", + "image_path": "ef24e3e3088aea2d211f2e7b72af13b6951aa4f14a60f4ce68bf461012bdbeef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 415, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 415, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 415, + 288, + 715 + ], + "type": "text", + "content": "model achieves comparable results with previous close-set or open-set models on COCO. For example, the panoptic segmentation performance gap between DINOv and our baseline Mask DINO is only 0.6 PQ (57.7 PQ vs 58.3 PQ). Out-domain open-set segmentation on ADE20K. After training with visual prompt on COCO and SAM, we do zero-shot evaluation on ADE20K to validate its open-set segmentation capability when seeing novel visual concepts. To our best knowledge, it is the first time to use visual prompt for open-set segmentation. Compared with previous text-prompted open-set models, we achieve comparable or better performance with only COCO semantic data and no semantic knowledge from large pre-trained models. Especially, compared with our baseline OpenSeed, we achieve better performance with much fewer data. Note that FC-CLIP [42] employs a frozen CLIP to do text-based open-set segmentation. As the text and visual features are aligned in CLIP, we also attempt to prompt a pre-trained FC-CLIP with visual features from CLIP to test its open-set ability with visual prompts. However, its visual prompting performance largely lags behind its text-prompted results. Therefore, it is non-trivial to transfer a multi-modal text-based open-set model to do visual-prompted recognition well. The results indicate that visual prompts can generalize well to new concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 308, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 524 + ], + "type": "text", + "content": "Segmentation and detection in the wild. We also validate the generalization capability of visual prompting on some diversified and domain-specific datasets including SegInW and ODinW, which in total encompass more than 60 datasets. These datasets contain many real-scenario or rare categories. As these datasets all focus on instance-level segmentation, we report the average and median AP (AP-Average and AP-Median) over all datasets. We first evaluate the Segmentation in the Wild (SegInW) benchmark, which consists of 25 datasets. With visual prompting, " + }, + { + "bbox": [ + 304, + 308, + 547, + 524 + ], + "type": "inline_equation", + "content": "DI-NOv" + }, + { + "bbox": [ + 304, + 308, + 547, + 524 + ], + "type": "text", + "content": " achieves a significant performance improvement over our baseline OpenSeed. For example, Our best AP-Average exceeds OpenSeed by 4.5 AP. We further evaluate Object Detection in the Wild (ODinW), which is composed of 35 datasets with bounding box annotations. As shown in Table 2, though we only employ much fewer semantically labeled data, we achieve better performance compared with previous models under similar settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 537, + 457, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 457, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 457, + 551 + ], + "type": "text", + "content": "3.2. Video Object Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Video object segmentation (VOS) aims to segment an interested object in a video by giving text or visual clues. Our model focuses on the semi-supervised setting, which segments a particular object throughout a video by giving visual clues in the first frame. In " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "DINOv" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": ", the visual prompt originates from one single image (generic/referring segmentation) or other images in one batch (generic segmentation). Therefore, our model has learned to prompt with visual features from other images. Therefore, " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "DINOv" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " is able to do video object segmentation (VOS) by replacing current frame visual prompt features with previous frames. For more accurate tracking, we also store the visual features of the predicted mask in previous frames. These fea" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12866" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 545, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 3. Zero-shot video object segmentation. Without training with video or pairwise image data, our approach is able to do video object segmentation in a zero-shot manner. (#Concurrent work.)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 545, + 291 + ], + "type": "table", + "html": "
MethodSegmentation DataTypeRefer-TypeZero-ShotDAVIS17DAVIS16-InteractiveYouTube-VOS 2018
JFJFJFJFGJsFsJuFu
With Video Data
AGSS [20]VOS+DAVISVideoMask67.464.969.9---71.371.365.575.273.1
AGAME [11](Synth)VOS+DAVISMask70.067.272.7---66.066.9*61.2*
SWEM [22]Image+VOS+DAVISMask84.381.287.4---82.882.486.977.185.0
XMem [3]Image+VOS+DAVISMask------86.185.189.880.389.2
SiamMask [32]COCO+VOSBox*54.358.569.871.767.8*60.258.245.147.7
MiVOS [4]BL30K+VOS+DAVISMask84.581.787.491.089.692.482.681.185.677.786.2
ReferFormer-B [35]RefCOCO(+/g)+VOS+DAVISText61.158.164.1---*****
UNINEXT-T [41]Image+VideoGeneralistMask74.571.377.6---77.076.881.070.879.4
UNINEXT-L [41]Image+VideoMask77.273.281.2---78.179.183.571.078.9
UNINEXT-L [41]Image+VideoText66.762.371.1---*****
Without Video Data
Painter-L [33]COCO+ADE+NYUv2GeneralistMask✓34.628.540.8---24.127.635.814.318.7
SegGPT-L [34]COCO+ADE+VOC+...Mask✓75.672.578.6---74.775.180.267.475.9
PerSAM-L [46]SAM+DAVISMask✘60.356.663.9---*****
SEEM-T [51]COCO+LVISMask✓60.457.663.362.758.966.451.455.644.159.246.9
SEEM-L [51]Mask✓58.955.062.862.258.366.050.057.238.261.343.3
DINOv-T (Ours)COCO+SAMMask✓73.371.075.777.072.981.260.965.370.052.357.9
DINOv-L (Ours)✓72.369.874.875.471.379.459.661.765.752.352.358.8
", + "image_path": "ba5458e789504ff4b0a4e8651108c263b1d742f41442ab6cd6099ba925255974.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 310, + 289, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 289, + 466 + ], + "type": "text", + "content": "tures, denoted as memory visual prompts, will be averaged together with the first frame's given prompt to construct the visual prompt of the current frame. Details of the memory visual prompt and ablations are in Appendix. D. By default, the memory length is set to 8. In Table 3, we conduct (interactive) video object segmentation evaluation on DAVIS17, DAVIS2016-Interactive, and Youtube-VOS 2018. The results of DAVIS2017 and Youtube-VOS 2018 indicate our model achieves better performance than SEEM and PersAM. In addition, " + }, + { + "bbox": [ + 46, + 310, + 289, + 466 + ], + "type": "inline_equation", + "content": "DINO\\nu" + }, + { + "bbox": [ + 46, + 310, + 289, + 466 + ], + "type": "text", + "content": " can also do interactive VOS, and our performance on DAVIS16-Interactive achieves significantly better performance compared with models not using video data for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 478, + 111, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 111, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 111, + 490 + ], + "type": "text", + "content": "3.3. Ablation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "content": "Effectiveness of Query Formulation. In Table 4, we ablate the effectiveness of using different query formulations for different tasks. The results indicate our double query formulation outperforms using only one type of query." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": "Effectiveness of Visual Prompt Formulation. In Table 5, we attempt to use a pre-trained CLIP vision encoder to encode the features of the visual prompt by cropping the prompted region into images for CLIP to process. As CLIP features contain rich semantics with few appearance features, which could not apply to referring segmentation tasks. Therefore, we ablate on generic segmentation tasks and find that the final model could not generalize well on open-set datasets like ADE. This result verifies our hypothesis that CLIP vision features could not generalize well on in-context visual prompting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Effectiveness of Unifying Tasks and Data. We unify visual generic segmentation and visual referring segmentation to use both semantically labeled data (COCO) and data with" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 304, + 361, + 548, + 420 + ], + "blocks": [ + { + "bbox": [ + 304, + 308, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 353 + ], + "type": "text", + "content": "Table 4. Ablation of using difference queries to do both in-context reference and generic segmentation. By default, we use both generic query and interactive query. We remove one type of query at a time to ablate their effectiveness." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 361, + 548, + 420 + ], + "lines": [ + { + "bbox": [ + 304, + 361, + 548, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 548, + 420 + ], + "type": "table", + "html": "
MethodCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinT49.642.747.058.073.371.075.7
only point query45.231.0(11.7)34.7(-12.3)52.771.468.874.0
only generic query46.238.3(-4.4)41.5(-6.0)53.368.966.571.3
", + "image_path": "a192438141d31727e8d75bb8710486a0dec903100d2a7e1a23c0a578413f3cc5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 304, + 483, + 548, + 530 + ], + "blocks": [ + { + "bbox": [ + 304, + 433, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 478 + ], + "type": "text", + "content": "Table 5. Ablation of using different ways to encode the visual prompt on our Swin-T model. Under the same setting, we change our prompt encoding method and use a pre-trained CLIP to crop and encode the prompted objects in the image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 483, + 548, + 530 + ], + "lines": [ + { + "bbox": [ + 304, + 483, + 548, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 548, + 530 + ], + "type": "table", + "html": "
Prompt EncodingCOCO (in-domain)ADE (out-domain)
PQmask APbox APmIoUPQmask APbox APmIoU
Ours49.642.747.058.019.411.412.821.9
CLIP48.540.743.554.912.61.41.313.3
", + "image_path": "d8a8c2c413abb3701c77f7ea840c77522357e8a6114e506229771c212df793fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 546, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 581 + ], + "type": "text", + "content": "only segmentation annotations (SA-1B). In Table 6, the results indicate that employing both datasets improves each individual task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": "Training batch size for generic segmentation. In Table 7, the results show that increasing training batch size consistently improves the generic segmentation performance. The reason for this phenomenon is that a larger batch size helps to sample more positive and negative visual in-context examples across different images, which better matches the inference setting with random visual examples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Inference In-Context Examples. In Fig. 6, we ablate the impact of using different in-context lengths. Increasing the in-context example exhibits diminishing returns, especially when the number of examples is more than eight." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12867" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 45, + 91, + 290, + 144 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 285, + 82 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 285, + 82 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 285, + 82 + ], + "type": "text", + "content": "Table 6. Ablation of the effectiveness of unifying tasks and data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 45, + 91, + 290, + 144 + ], + "lines": [ + { + "bbox": [ + 45, + 91, + 290, + 144 + ], + "spans": [ + { + "bbox": [ + 45, + 91, + 290, + 144 + ], + "type": "table", + "html": "
MethodDataCOCODAVIS17
PQmask APbox APmIoUJFJF
DINOv-SwinTCOCO, SAM49.642.747.058.073.371.075.7
DINOv-SwinTCOCO48.941.745.957.163.360.865.7
DINOv-SwinTSAMN/A---68.466.070.8
", + "image_path": "e6bd13b327c728715453dcb32a1972be58bb1fb34ca8eda9e567eed18d9801d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 45, + 167, + 290, + 254 + ], + "blocks": [ + { + "bbox": [ + 59, + 146, + 273, + 158 + ], + "lines": [ + { + "bbox": [ + 59, + 146, + 273, + 158 + ], + "spans": [ + { + "bbox": [ + 59, + 146, + 273, + 158 + ], + "type": "text", + "content": "Table 7. Ablation of image batchsize sampling in training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 45, + 167, + 290, + 254 + ], + "lines": [ + { + "bbox": [ + 45, + 167, + 290, + 254 + ], + "spans": [ + { + "bbox": [ + 45, + 167, + 290, + 254 + ], + "type": "table", + "html": "
Method#Batchsize for Prompt SamplingCOCO
PQmask APbox APmIoU
DINOv-SwinT128.923.225.333.7
DINOv-SwinT445.137.040.450.6
DINOv-SwinT847.339.243.153.1
DINOv-SwinT3247.840.344.156.2
DINOv-SwinT6449.045.241.557.0
", + "image_path": "712919a553e33b68ea2163c6a12179cdb43759d9f0bbfd1871cb95b3425c9fc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 274, + 262, + 430 + ], + "blocks": [ + { + "bbox": [ + 59, + 274, + 262, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 274, + 262, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 274, + 262, + 430 + ], + "type": "image", + "image_path": "ed78fa1f581541d252e0789c7a245441aba2661b4490d65021a301642e81dd1d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 439, + 264, + 451 + ], + "lines": [ + { + "bbox": [ + 69, + 439, + 264, + 451 + ], + "spans": [ + { + "bbox": [ + 69, + 439, + 264, + 451 + ], + "type": "text", + "content": "Figure 6. DINOv query formulation of different tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 458, + 138, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 138, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 138, + 470 + ], + "type": "text", + "content": "4. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 478, + 257, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 257, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 257, + 491 + ], + "type": "text", + "content": "4.1. Visual Perception Through Text Prompt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 496, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 653 + ], + "type": "text", + "content": "Innovations in open-vocabulary object detection [7, 12, 19, 24, 26, 43, 44, 47] and open-vocabulary segmentation [6, 8, 14, 31, 37, 44], have shown great potential in generic visual perception, by leveraging large pre-trained vision-language models like CLIP [30] and ALIGN [10]. These approaches demonstrate significant strides in zero-shot and few-shot performance, adapting to a variety of visual contexts through text prompts. However, the reliance on text alone introduces limitations due to linguistic ambiguity and the potential mismatch between textual descriptions and complex visual scenes [40]. This highlights the ongoing need to refine the integration of visual inputs for more accurate and comprehensive image perception." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 659, + 272, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 272, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 272, + 672 + ], + "type": "text", + "content": "4.2. Visual Perception Through Image Example" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Building upon the foundations set by text-based visual perception methodologies, the field has seen a notable shift towards incorporating image examples to enhance accu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": "racy and context sensitivity. OV-DETR [43] extends its open-vocabulary object detection capability beyond text, by utilizing both the image encoder and text encoder from CLIP [30], allowing for object detection guided by visual examples. Similarly, OWL-ViT [26] leverages large-scale image text examples in its contrastive pre-training phase, and propose to adopt image example for one-shot image-conditioned object detection. MQ-Det [40] utilizes image examples to enhance text descriptions for better open-vocabulary object detection performance. These methods typically adopt the image encoder in CLIP to extract visual features from given image examples for a more accurate perception of objects and scenes, and demonstrate that visual examples can bridge the gap between textual ambiguity and the complex nature of visual perception." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 258, + 525, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 525, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 525, + 272 + ], + "type": "text", + "content": "4.3. Visual Perception Through Visual Prompt" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "content": "Different from image example-based methods that take an image as input, which are then processed by multi-modal encoder like CLIP [30], visual prompt-based methods typically use visual instructions (e.g. point, mask, scribble, and refereed regions of another image) to guide a model for a specific visual task. SAM [13], for instance, introduces a promotable model for interactive segmentation, fostering research in computer vision foundation models. SEEM [51] stands out as an interactive and versatile model for segmenting objects, accommodating various types of prompts, and is semantic-aware compared to SAM. Semantic-SAM [17] excels in semantic awareness and recognizing granularity, and can handle panoptic and part segmentation. SegGPT [34] unifies various segmentation tasks by formulating segmentation as an in-context coloring problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "We present DINOv, a unified framework for in-context visual prompting to accommodate both referring segmentation and generic segmentation tasks. To effectively formulate in-context visual prompts, we designed a simple prompt encoder to encoder reference visual prompts from the reference image and adopted a shared decoder to decode the final target visual prompts from the target image. We also formulate generic latent queries and point queries to align different tasks and data. The experimental results indicate that DINOv demonstrates impressive referring and generic segmentation capabilities to refer and detect with in-context visual prompting. Notably, DINOv delivers competitive performance compared to close-set segmentation on in-domain datasets and show promising results on many open-set segmentation benchmarks. We hope our early exploration of visual in-context prompting could inspire the community. Limitations. We employ limited semantically labeled data (COCO), which can be scaled up for better performance and extended to text prompts for multi-modal understanding." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12868" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "text", + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 201, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 256 + ], + "type": "text", + "content": "[2] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 257, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 257, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 257, + 287, + 289 + ], + "type": "text", + "content": "[3] Ho Kei Cheng and Alexander G. Schwing. XMem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "text", + "content": "[4] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 357 + ], + "type": "text", + "content": "[5] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "text", + "content": "[6] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "text", + "content": "[7] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 502 + ], + "type": "text", + "content": "[8] Dat Huynh, Jason Kuen, Zhe Lin, Jiuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "type": "text", + "content": "[9] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. arXiv preprint arXiv:2211.06220, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 612 + ], + "type": "text", + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[11] Joakim Johnander, Martin Danelljan, Emil Brissman, Fahad Shahbaz Khan, and Michael Felsberg. A generative appearance model for end-to-end video object segmentation, 2018. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 2, 5, 8, 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "text", + "content": "[14] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[15] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[16] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 5, 6, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "text", + "content": "[17] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 2, 5, 8, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 381 + ], + "type": "text", + "content": "[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 448 + ], + "type": "text", + "content": "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 6, 8, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[20] Huaijia Lin, Xiaojuan Qi, and Jiaya Jia. Agss-vos: Attention guided single-shot video object segmentation. In ICCV, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 5, 12" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[22] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1362-1372, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 635 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12869" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 174 + ], + "type": "text", + "content": "[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 216, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 216, + 188 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 216, + 188 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-4 technical report, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 190, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 190, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 190, + 287, + 233 + ], + "type": "text", + "content": "[28] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 235, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 300 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 303, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 369 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 437 + ], + "type": "text", + "content": "[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 439, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 287, + 483 + ], + "type": "text", + "content": "[32] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "text", + "content": "[33] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning, 2023. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[34] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. arXiv preprint arXiv:2201.00487, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[36] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[37] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[38] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[39] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 165, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 208 + ], + "type": "text", + "content": "[40] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multimodal queried object detection in the wild. arXiv preprint arXiv:2305.18980, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "text", + "content": "[41] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. arXiv preprint arXiv:2303.06674, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "type": "text", + "content": "[42] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. arXiv preprint arXiv:2308.02487, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 303, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 347 + ], + "type": "text", + "content": "[43] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "text", + "content": "[44] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 406, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 461 + ], + "type": "text", + "content": "[45] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "text", + "content": "[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 564 + ], + "type": "text", + "content": "[47] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "text", + "content": "[48] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 12" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "text", + "content": "[49] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[50] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. arXiv preprint arXiv:2212.11270, 2022. 2, 6, 12" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12870" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[51] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 7, 8" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12871" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_content_list.json b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1efb742c8d418919d21e94c394e2ff8a8cf3dfb --- /dev/null +++ b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "Visual LayoutComposer:Image-Vector Dual Diffusion Model for DesignLayout Generation", + "text_level": 1, + "bbox": [ + 171, + 130, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohammad Amin Shabani $^{1*}$ Zhaowen Wang $^{2}$ Difan Liu $^{2}$ Nanxuan Zhao $^{2}$ Jimei Yang $^{2}$ Yasutaka Furukawa $^{1}$ Simon Fraser University $^{2}$ Adobe Research", + "bbox": [ + 166, + 202, + 803, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/098199d519b41b62d9ea2ba15c3a8aef1a6940e4d27606ef8e18900731fad987.jpg", + "image_caption": [ + "Figure 1. Overview of our method. Left: our image-vector dual diffusion model architecture where both image and vector diffusion are conditioned on input elements and exchanging information with attention maps. Right: a comparison showing the advantage of dual-domain (image and vector) over traditional single-domain approaches." + ], + "image_footnote": [], + "bbox": [ + 86, + 271, + 584, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5f57ae3ca19cbe08b08aee90ce5581e26ce1c5ab73eeb1adb4a7d2cb4931df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 275, + 893, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 534, + 313, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes an image-vector dual diffusion model for generative layout design. Distinct from prior efforts that mostly ignores visual information of elements and the whole canvas, our approach integrates the power of a pre-trained large image diffusion model to guide layout composition in a vector diffusion model by providing enhanced salient region understanding and high-level inter-element relationship reasoning. Our proposed model simultaneously operates in two domains: it generates the overall design appearance in the image domain while optimizing the size and position of each design element in the vector domain. The proposed method achieves the state-of-the-art results on several datasets and enables new layout design applications. Project webpage: https://aminshabani.github.io/visual.layout_composer.", + "bbox": [ + 75, + 566, + 473, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 805, + 209, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Layout design is essential in creative tasks, determining the spatial arrangement of texts, images, and other visual elements, which are composed into a visual design, such as", + "bbox": [ + 75, + 830, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "book and magazine covers, posters showcasing new products, or invitation cards for holiday parties. Graphic designers rely on their keen artistic sensibility in meticulously optimizing the arrangements of the visual elements while ensuring that the composited image conveys the intended message with aesthetic appeals.", + "bbox": [ + 500, + 535, + 890, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning techniques have been successful in directly regressing the layout vector parameters, i.e., the bounding box coordinates of elements. Both transformer-based models [1] and more recent diffusion-based models [18] have achieved promising results that capture basic geometrical properties such as alignment and symmetry. However, most of these models ignore the visual information from the input elements, including image content, text typeface, and color/shape of vector arts. This leads to unsatisfactory designs as illustrated by the Vector Only results in Figure 1. The lack of layout datasets with associated images also contributes to the negligence of visual information. The most commonly used RICO [8] and Pub-LayNet [45] datasets only contain box representation of design elements. The Crello [40] dataset provides element-level image rendering for over 23K design templates, whose visual features are used in the FlexDM [18] model. However, the interplay of element appearance when rendered on", + "bbox": [ + 496, + 628, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during the first author's internship at Adobe.", + "bbox": [ + 94, + 886, + 444, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9222", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the same canvas has never been studied at any large scale.", + "bbox": [ + 76, + 90, + 460, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent success of image generative models [6, 9, 30, 32, 41] has enabled the generation of layout design in the image space. Trained on millions of high quality design images from web, these large generative models produce visually diverse and compelling layouts as raster images, demonstrating sophisticated design principles such as the placement of major subjects, style coherence, and perceptual contrast. See the Image Only results as examples in Figure 1. These visual properties, vital to layout design, are learned in a holistic way, not straightforward by previous vector domain approaches. Despite high visual quality, the layout designs in image format forbid further editing and customization. Moreover, current generative models are controlled only by text prompts, and cannot use given visual assets as required in many real-world applications.", + "bbox": [ + 75, + 108, + 468, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the complementary strengths of the vector and image domains, we propose a combined approach that generates layout designs in both image and vector domains simultaneously. An image branch is a pre-trained text-to-image stable diffusion model [30], which refines (i.e., noises) a design as an image, given a variable number of design elements with different media types. A vector branch refines per-element bounding box parameters conditioned on attention maps from the UNet of the image branch. Ground truth element masks are used to apply loss directly to these attention maps, whose manipulation at generation time will allow users to control the positioning of elements. This dual diffusion model refines the visual design in the vector and image domains simultaneously. The final vector output assembles the input elements and renders the visual design while allowing further editing by users. The overall proposed method is illustrated on the left of Figure 1.", + "bbox": [ + 75, + 335, + 470, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For training large diffusion models, we have collected a large-scale design Poster dataset from public design web pages by extracting both images and metadata. This effort yields 140K image layout paired samples as our own dataset, more than six times of the current largest visual layout dataset Crello with 23K samples. We conduct experiments on both datasets to validate the effectiveness of our dual diffusion model. Our data collection tools and processing pipeline can facilitate future efforts to curate even larger datasets for exploring visual design.", + "bbox": [ + 75, + 595, + 468, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are summarized below:", + "bbox": [ + 94, + 748, + 398, + 762 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel dual diffusion model that generates layout design in both image and vector domains conditioned on input design elements. The two diffusion processes interact by exchanging element attention maps, where the final results attain both vector editability and visual quality.", + "- Creative design applications, enabling element-wise user controls, such as canvas resizing and design variation.", + "- State-of-the-art layout generation results on the existing datasets and our Poster dataset." + ], + "bbox": [ + 76, + 763, + 467, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vector-only Layout Generation", + "text_level": 1, + "bbox": [ + 500, + 114, + 774, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models for layout generation have seen significant developments. Yamaguchi [40] introduced a VAE-based architecture for unconditional generation of vector graphic documents. Following this, Li et al. [26] proposed LayoutGAN, which uses a relational generator and a wireframe renderer for training with a pixel-based discriminator. Kikuchi et al. [20] further refined this approach with LayoutGAN++, incorporating user-specified constraints into layout generation. Chai et al. [4] proposed using DDPM for layout generation. Inoue et al. proposed LayoutDM [17], a discrete diffusion-based model for layout designs. Zhang et al. [42] and Hui et al. [16] both contributed to this domain with LayoutDiffusion and LDGM, respectively, each based on discrete diffusion models. Cheng et al. [7] proposed a unique approach using guidelines as input conditions for the latent diffusion model.", + "bbox": [ + 496, + 137, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Visual-guided Layout Generation", + "text_level": 1, + "bbox": [ + 500, + 390, + 792, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zhou et al. [46] tackled image-composition-aware layout generation with a multi-stage GAN-based method, an approach further improved by Xu et al. [39] through the use of a pixel-level discriminator. Zheng et al. [44] proposed a GAN based approach using image-based representation for layouts, which requires subsequent post-processing. Lin et al. [28] introduced a multi-step approach dedicated to the cleansing and retargeting of advertising posters. Hsu et al. [15] introduced DS-GAN, a CNN-LSTM-based conditional GAN method for poster layout generation. Inoue et al. [18] proposed Flex-DM, utilizing a transformer to predict masked attributes in design. Shimoda et al. [34] and Tang et al. [36] expanded the field by proposing new models for typography generation and layout generation, respectively. Our work differs by focusing on enhancing visual information and consistency through a large image diffusion model.", + "bbox": [ + 496, + 412, + 890, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Image-based Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 680, + 769, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image diffusion model [35] was proposed initially for unconditional image generation. With the rapid development of the large-language model (LLM) [3], it has shown groundbreaking ability for text-to-image generation tasks [14]. To enhance the computational efficiency, rather than performing diffusion steps at the pixel level, latent diffusion model (LDM) [30] conducted operations on latent space, further boosting the generation ability of diffusion models. It has been widely used in various applications, including image-to-image translation [31], style transfer [23, 37], and condition-guided editing (e.g., sketch, layout, depth map, etc.) [27, 43]. To enable multi-object composition closer to our work, Xiao et al. [38] introduced", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9223", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Localization Loss for free text-to-image generation with multiple subjects, while Sarukkai et al. [33] developed Collage Diffusion for creating realistic photos, focusing on harmonization and fidelity. Balaji et al. [2] training an ensemble of diffusion models specialized for different parts of the diffusion model to improve the overall text alignment of the model. Goel et al. [10] proposed Pair Diffusion to edit object properties in images via structure and appearance decomposition. Our model stands out by not needing predefined element positions and scales, handling multiple subjects, maintaining visual consistency, and providing enhanced editing flexibility.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 284, + 169, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Visual LayoutComposer (VLC) model consists of two diffusion models that operate in parallel: one in the vector domain and the other in the image domain. The two models exchange features at intermediate stages, enabling them to produce outputs that are not only consistent but also leverage the information from both domains. In the following subsections, we first describe our problem formulation, followed by the specifics of the diffusion models for each domain, and how they collaborate to enhance output consistency. Figure 2 shows the model overview.", + "bbox": [ + 75, + 309, + 472, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 468, + 207, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our input is a set of $N$ design elements $E = \\{e_1, e_2, \\ldots, e_N\\}$ . Each element $e_i$ is associated with a class category $c_i$ indicating its asset type (such as text, image, shape), an index $o_i$ indicating its layer order relative to other elements, and an RGBA image $\\mathbf{I}_i$ of its rendering. On a designated 2D canvas with aspect ratio $r_c$ , the task is to predict an upright bounding box for each element $e_i$ , defined by the 2D coordinates of the top left and bottom right corners as a 4D vector $x_i$ . The goal is to generate a layout design with high visual quality by composing each element $e_i$ on the canvas with the position and size specified by $x_i$ .", + "bbox": [ + 75, + 491, + 472, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Condition Embedding", + "text_level": 1, + "bbox": [ + 76, + 665, + 284, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After resizing each element image $\\mathbf{I}_i$ of aspect ratio $r_i$ to $224 \\times 224$ , we employ CLIP [29] followed by a linear layer to derive its feature representation in the model's hidden dimension $d$ . In parallel, we encode the vector attributes of each element by a vector attributes encoder. The vector condition is constructed by concatenating the element image ratio $r_i$ , canvas image ratio $r_c$ , one-hot encoded version of the element's type $c_i$ and the order of the element $o_i$ . A linear layer maps the concatenated feature to the model's dimension $d$ . The image feature and the vector feature are concatenated and processed by a linear layer to form the final condition feature of size $N \\times d$ . For each image or vector diffusion model, this condition feature is fed into a transformer-based condition processing module (Figure 2),", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "yielding outputs of sizes $N \\times 768$ for image condition feature and $N \\times d$ for vector condition feature.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Network Architecture", + "text_level": 1, + "bbox": [ + 498, + 133, + 705, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model in the image domain utilizes a pretrained Stable Diffusion V1.5 model [30], retaining all layers frozen except for the attention modules. During a given time step $t$ , a noisy latent image $z^t$ , representing the target layout rendering, will be passed to the U-Net. Contrary to the original Stable Diffusion, which uses text prompts, we pass the processed image condition feature as the input to the cross-attention layers.", + "bbox": [ + 496, + 157, + 890, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model in the vector domain produces the bounding box positions as a 4D vector $x_{i}$ . We embed the noisy input coordinates $x^{t}$ to a $N \\times d$ feature vector, concatenate with the vector condition feature, and map to the model's dimension $d$ by a linear layer. These features are fed into a transformer-based diffusion denoiser to estimate the corresponding noise for each element $x_{i}$ .", + "bbox": [ + 496, + 279, + 890, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Information exchange between the image and vector domains is challenging, where we found that direct feature space sharing as in ControlNet [43] is not effective. We use the processed attention scores from U-Net as a shared medium. The advantages of this method are: (i) it aligns the number of attention scores with the number of input elements, enabling a straightforward concatenation of each image mask to its respective vector model feature token; (ii) it provides a general indication of each element's location in the image domain, simplifying the interpretation for the vector domain; and (iii) it allows the vector domain to adjust these scores to enhance the U-Net's focusing capability, with a mechanism similar to Attention Modulation [21].", + "bbox": [ + 496, + 385, + 890, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, we extract a set of attention scores from the cross-attention module of the U-Net's middle block. Given the scores with 8 heads and size $8 \\times 8$ for each design element, we flatten the tensors to form a set of vectors with the total size of $N \\times 512$ . This feature is embedded to $N \\times 1024$ and concatenated with each input element in the vector domain prior to the transformer-based diffusion denoiser. The transformer model produces an output of $N \\times 2048$ , which is passed through linear projection layers to yield denoised estimation of $x^{t - 1}$ and the attention modulation residual of size $N \\times 512$ . We add the residual values to the original attentions to form the refined attention scores. These scores are sent back to the U-Net, promoting interactive and iterative information exchange between the two domains.", + "bbox": [ + 496, + 583, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the diffusion process, repetitive elements can lead to omissions or misalignments in the image. To mitigate this, we adopt the cross-attention localization loss [38]. This loss refines the attention scores of each region, ensuring a more accurate correspondence to its respective element. We apply this loss within the U-Net's attention modules to further enhance the quality of the attention scores.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9224", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fce675491ca02073f9bf1d1306b4b0f0f1d8b8883a2fc949cd90e621367618f5.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed model. Given a set of element images and additional attributes, the model simultaneously employs dual diffusion processes to generate a cohesive image layout and the corresponding bounding boxes for each element." + ], + "image_footnote": [], + "bbox": [ + 84, + 95, + 883, + 320 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1f6d9291135a11c36624b2838f4defc26fa548a45c43a917ed17015f80f682c.jpg", + "image_caption": [ + "Figure 3. Generated image/vector layout without or with canvas masking. By masking intermediate outputs of the diffusion model in image domain, we optimize the image output for the canvas size and increase the consistency between the vector-image domains." + ], + "image_footnote": [], + "bbox": [ + 83, + 383, + 289, + 494 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ddfc208f079e068970ec4e8e738ea1df4c2079494fafb9197b38ae9991b30b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 383, + 460, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training losses for the image and vector diffusion models are the same as the original DDPM [13].", + "bbox": [ + 75, + 597, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Controllable Element Composition", + "text_level": 1, + "bbox": [ + 76, + 647, + 382, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experiments showed that fine-tuning the image domain led the model to disregard the specified canvas aspect ratio, favoring the generation of square-shaped layout designs. This tendency not only reduces the quality and diversity of the generated designs but also increases the discrepancy between the image and vector domains, particularly given that the vector domain generally follows the canvas's aspect ratio. To address this problem, we introduce a straightforward yet efficient strategy: masking the noisy latent feature map of $z^t$ at each time step to have zero value outside the canvas region, which ensures image contents related to input condition only appear inside the canvas. The homogeneous region outside the canvas is cropped from the final generated image, maintaining the original image aspect ratio. Figure 3 shows the effect of this process on both domains.", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 376, + 633, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details: In our implementation, we use the pretrained weights of StableDiffusion v1-5 [30] for image domain, and OpenAI's clip-vit-large-patch14 vision model for embedding the image of each element to latent space. For training, we run each model for 250,000 iterations, employing a starting learning rate of 5e-5 which decreases by a factor of 0.5 every 100k steps. The vector model has a hidden dimension of 2048, accommodating a wide range of features including attention scores, vector inputs, and image information. Additionally, we incorporate classifier-free guidance into our process, randomly dropping the image inputs of the elements in $20\\%$ of the times. Training is conducted on 8 NVIDIA A100 GPUs, with a total batch size of 128. We pad each layout to square shape and normalize the coordinates to $[-1,1]$ . For both domains, we employ the DDPM scheduler with 1000 steps for training and 50 steps for inference to ensure high-quality layout generation.", + "bbox": [ + 496, + 407, + 893, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets: Our experiments use two groups of datasets.", + "bbox": [ + 500, + 686, + 861, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Poster and Crello [40] datasets: demonstrating our method's ability to capture intricate visual information and the interplay between design elements. These datasets, characterized by their rich graphic content, are instrumental in showcasing the method's effectiveness in a visually complex environment. We gathered Poster dataset, which consists of 137,781 layout designs. From this, we reserved 1,000 layouts for validation and 3,000 for testing. Each layout is represented as an RGBA image and is classified into one of five categories: Frame, Canvas, Shape, Mix, and Text. We have limited the number of elements in any given layout to a maximum of 20 to maintain structural simplicity. Given the intricate nature of", + "bbox": [ + 500, + 704, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9225", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these layouts, we have also provided the sequential order of the elements within each design. The Crello datasets include 18,768 training, 2,316 validation, and 2,331 test layouts including categories such as vector shape, image, or text placeholder.", + "bbox": [ + 89, + 90, + 467, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- RICO [8] and PubLayNet [45] datasets: having limited visual information, which challenges our method. To compensate for this, we assign distinct colors to each category and depict layouts with these colors, simplifying the visuals to fixed-color blocks representing each element. This approach tests our method's ability in generating layouts based primarily on structure rather than rich visual details. The Rico dataset consists of user interface (UI) designs sourced from various Android applications, spanning up to 25 categories of UI elements including text bars, icons, and buttons. PubLayNet includes 360,000 document layout instances from academic papers, categorized into five primary layout elements: text, titles, lists, tables, and figures. We use the processed datasets by [17], obtained by keeping only those layouts containing a maximum of 25 elements each. This results training/validation/test sets of sizes 35,851/2,109/4,218 respectively for Rico and 315,757/16,619/11,142 for PubLayNet.", + "bbox": [ + 76, + 167, + 468, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 464, + 215, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the baselines, we compare our method with recent state-of-the-art methods in Layout Design Generation [5, 11, 12, 17, 19, 20, 22, 24, 25]. We compare our method to the LayoutDM [17] as the state-of-the-art for layout generation. This model only uses element categories, adhering to their original design in discrete diffusion space. Our vector domain model operates in a continuous diffusion space and incorporates each element's image embedding. Finally, our dual-domain model combines element image embeddings with the capabilities of the image diffusion model to get the highest quality.", + "bbox": [ + 75, + 488, + 467, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 675, + 297, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: We follow the previous work [17] and use metrics including Maximum IoU [20] and FID over extracted features of the layouts [20, 24] for RICO and PubLayNet datasets. The previous methods and measurements are effective when there are no visual data or input images for layout elements. However, assessing layout designs in Poster and Crello datasets is more complex. Even layouts with well-aligned bounding boxes might look poor due to the textures and visual details of each element. Elements can also overlap for shadow or visual effects. Inoue et al. [18] approached this as a reconstruction task, using metrics like cosine similarity between the actual and predicted values. This approach works for simple designs with few elements", + "bbox": [ + 75, + 703, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/46f0378c645f1cc0822259bb960d87fd1350a3f8ea64be7cd74d9c392e675cd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelsBounding BoxComposed ImageGenerated Image
CrelloLayoutDM [17]10.2034.43—
VLC (Vector only)0.215.74—
VLC (Dual-domain)0.215.8311.42
PosterLayoutDM [17]1.8120.00—
VLC (Vector only)0.195.82—
VLC (Dual-domain)0.093.756.00
", + "bbox": [ + 501, + 89, + 883, + 222 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/03af75f581e7708cf63da6dcd0cdc8a49fbaa50989329a3d07ca645ec32c34b6.jpg", + "table_caption": [ + "Table 1. FID scores of bounding boxes, composed RGB layouts obtained from the bounding boxes, and the directly generated layouts from the image domain. Our vector-only method outperforms LayoutDM by effectively using image information, with our multi-domain approach yielding the best results." + ], + "table_footnote": [], + "table_body": "
ModelsRICOPubLayNet
FID (↓)mIoU (↑)FID (↓)mIoU (↑)
LayoutVAE [19]33.30.24926.00.316
NDN-none [24]28.40.15861.10.162
LayoutGAN++ [20]6.840.26724.00.263
LayoutTrans [12]5.570.22314.10.272
MaskGIT [5]26.10.26217.20.319
BLT [22]17.40.20272.10.215
BART [25]3.970.2539.360.320
VQDiffusion [11]4.340.25210.30.319
LayoutDM [17]3.550.2777.950.310
VLC (Vector only)2.380.4185.490.348
VLC (Dual-domain)5.600.3765.740.331
Validation Set1.850.6916.250.438
", + "bbox": [ + 501, + 316, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison of our method with the baselines using RICO and PubLayNet datasets. Our method achieves superior results compared with the baselines even with no visual information in input conditions.", + "bbox": [ + 496, + 556, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "but limits diversity and learning on more complex datasets. Instead, we propose using FID values over three outputs: bounding boxes, feature embeddings from RGB images created by composing elements within generated bounding boxes, and the directly produced RGB images. Although this is our best option, we realize that FID can be biased to image quality and content similarity which should be removed from the consideration for layout evaluation. Finding better visual metric for layout can be a future direction.", + "bbox": [ + 496, + 638, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results: We evaluate our VLC method using visual information as shown in Table 1. Our dual-domain method stands out for its ability to create visually appealing images by using a large image model. This method is a significant improvement over the LayoutDM [17]. LayoutDM uses only element categories as input, which, while effective in creating quality vectors, often fails to align with the actual visual content of each element. This mismatch results in a", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9226", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3246c79d8f14b35c7ba50b906e5c896f023b81af16e56c842e20673f4dca0761.jpg", + "image_caption": [ + "VLC (vector only)", + "VLC (dual domain)", + "Original Design", + "Figure 4. Qualitative comparison of our method with baselines on Crello dataset." + ], + "image_footnote": [], + "bbox": [ + 107, + 107, + 439, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "drop in the overall quality of the final image. Our vector domain approach is the most similar one to the transformer-based reconstruction used in FlexDM [18]. It includes both image information and other input conditions, and produces higher-quality images compared to LayoutDM, but has some drawbacks such as occasionally covering faces or not managing color contrast well when elements overlap. Our vector only model and dual-domain model perform similarly on the Crello dataset. The reason can be attributed to the larger number of elements in the dataset, which leads to the content mismatch between image and vector domain such that vector model cannot follow the guidance from image correctly.", + "bbox": [ + 75, + 383, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we compare our method with existing baselines on the RICO and PubLayNet datasets. The baseline results are reported by Inoue et al. [17]. Remarkably, even when the input elements are limited to only categorical data without any visual information, our method demonstrated its capability to produce high-quality layouts in both vector-only and dual-domain settings. In the vector-only domain, we observed a trend where training appeared less stable. However, despite this instability, the results often surpassed those generated by discrete diffusion models [11, 17]. In addition, the vector domain model tends to align with the performance of our dual-domain model. This alignment is particularly notable as our dual-domain method achieves better results than most of the baselines, underscoring its effectiveness even in the absence of visual input. It is important to note that while our method is slower than the baselines duo to the amount of computation, this trade-off is compensated by its high-quality output and robustness in processing visual information with its diffusion formulation. Overall, these findings suggest that our method not only achieves state-of-the-art performance but also offers promising po", + "bbox": [ + 75, + 583, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cb0153445454ef9fc9c631adca109a4394ea12e227d1ed84136c92f074d8616.jpg", + "image_caption": [ + "VLC (vector only)", + "VLC (dual domain)", + "Original Design", + "Figure 5. Qualitative comparison of our method with baselines on Poster dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 111, + 875, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tential for generating high-quality layouts under varied and challenging conditions.", + "bbox": [ + 500, + 652, + 890, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 700, + 710, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effectiveness of our dual-domain method is further illustrated in Figure 4&5 with visual results from the Crello and Poster datasets. These examples highlight our approach not only improves the layout quality but also ensures that the final compositions are more visually coherent and appealing, especially when compared to other methods.", + "bbox": [ + 498, + 723, + 890, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Model Analysis", + "text_level": 1, + "bbox": [ + 500, + 823, + 653, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 The Effectiveness on Dual-domain Modeling", + "text_level": 1, + "bbox": [ + 500, + 845, + 862, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addressing the complexities of dual-domain modeling, a crucial challenge lies in maintaining consistency between", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9227", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e6e15a024ca7cc46b3a8b18a49995aa2bbd6a8202fbd73c501770a5bc66746e3.jpg", + "image_caption": [ + "Original Layout" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 271, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fff2365dec227bdd301c50d558aca5762bcb55c19ce589f6a86ecdf721182771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 92, + 462, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b2249abc2e85bd50f2b4c80f7b40c8d3f8e0eb70839bedb66538846f8d286b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 148, + 462, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc1f7543131f62c3f1916abb1e0a1a534d46d571c659e4bb409ec05208358be0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 92, + 653, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41bebc02502770036f032cfacfa22c2be764fdd905cbcd75f73253b8881f533c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 148, + 653, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3d04414aa09e316a9c8f6824a1e30224e7d8a9b9f5c053d6919abf89f682466.jpg", + "image_caption": [ + "eaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeae", + "psodmooO O" + ], + "image_footnote": [], + "bbox": [ + 653, + 92, + 844, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53def6ed455f8a123750e09afd6b73b013e1c198c58c9977a21ac8413d6c04cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 148, + 844, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38f3b88969899bd45c95ca4dce11d68e96c82d48fa3cbf3c5b17449cc9cfe9cb.jpg", + "image_caption": [ + "Original Layout" + ], + "image_footnote": [], + "bbox": [ + 125, + 210, + 218, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e23fea23232d14330c9301fcf80aab7bec14800340f8ec45b1fea236684cd7b.jpg", + "image_caption": [ + "Generated", + "Image", + "Composed", + "Image" + ], + "image_footnote": [], + "bbox": [ + 274, + 209, + 462, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/583b3b8bb4c7515bd37d94927db05df5ac5016f12085dfcb38d3bf190189ef3d.jpg", + "image_caption": [ + "Generated", + "Image", + "Composed", + "Image" + ], + "image_footnote": [], + "bbox": [ + 464, + 208, + 651, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ba8ef9e6556107adac47c68c0933f45e2ddf93ca2cea8a9af33aae57a1001ed.jpg", + "image_caption": [ + "Generated", + "Image", + "Composed", + "Image" + ], + "image_footnote": [], + "bbox": [ + 656, + 209, + 841, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9c91c781261b9fab6437b95716f8e561f4ad0ba2a187fe9c1522d9ef47b71b5.jpg", + "image_caption": [ + "Figure 6. Our approach is capable of producing a wide range of consistent designs in both image and vector domains.", + "Figure 7. Attention scores are assigned to various elements in the image domain." + ], + "image_footnote": [], + "bbox": [ + 88, + 388, + 462, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the image and vector domains. This is vividly illustrated in Figure 3, which demonstrates that when these two domains adhere to a consistent denoising path, the result quality is significantly enhanced. Conversely, deviation from this path leads to a noticeable drop in quality. Complementing this, Figure 6 highlights the efficacy of our method, showcasing that both the image and vector domains can generate results that are not only consistent with each other but also retain the diversity of the generated layouts.", + "bbox": [ + 75, + 660, + 468, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 The Effect of Attention Score Maps", + "text_level": 1, + "bbox": [ + 76, + 816, + 377, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The attention score maps are the key for the visual composing capability of our model. We visualize the clear spatial correlation between attention scores and corresponding elements in the image domain in Figure 7. Our vector", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5f3c47a8147c002e6af0ada6747c65bbbc57243d963e870c8b9a3a371bbe4c8.jpg", + "image_caption": [ + "Figure 8. Through attention masking, we can decompose all the elements in the latent image space, and include them one by one to create a layered image representation. The bottom right image shows the original design." + ], + "image_footnote": [], + "bbox": [ + 506, + 388, + 883, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "domain model leverages these spatial information to predict vector layouts by adaptively tracking each element and resolving the ambiguity between similar objects.", + "bbox": [ + 496, + 656, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By manipulating the attention scores, we can also control the element generation in the image domain. For example, we can remove one element from canvas while keeping the rest layout unchanged by masking out its attention map. If we do this sequentially to each element, we can decompose image output into a layered representation organized by elements, as shown in Figure 8. This results in a virtual vector layout format, which may be used for similar purpose as our vector domain model.", + "bbox": [ + 496, + 702, + 892, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Applications", + "text_level": 1, + "bbox": [ + 500, + 847, + 633, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With its two domain outputs and high controllability through attention maps, our model can be used for a range", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9228", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d089ac2e8d8aeba559a197ded7a56ba34bd8802fb254e90c88ccab772dc96195.jpg", + "image_caption": [ + "Input Layout Boxes" + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 179, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4217ef944e4d5df95d8a1cc006ac107e06822916a387024f0748b853248345e2.jpg", + "image_caption": [ + "Generated Layout Images" + ], + "image_footnote": [], + "bbox": [ + 183, + 90, + 277, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a47382988755cfbb3c3544228bc176e68ce9bac547f6330c63a13b31a3e47b06.jpg", + "image_caption": [ + "Figure 9. Layout controlled visual style variation. Given input layout with bounding boxes, our image model can generate visual results with the same layout and different styles." + ], + "image_footnote": [], + "bbox": [ + 279, + 90, + 370, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/021c57a21423304af37281f54cdd9f0d4a866674c8107b1d9bdf3c8bd2ca861a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 90, + 462, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2fc73a3c26712e1a2cf4420abaa3f8222242bf96bb70c1886eab9c5542e98f57.jpg", + "image_caption": [ + "Figure 10. Canvas resize with different aspect ratios. Our model can adapt layout under various aspect ratios." + ], + "image_footnote": [], + "bbox": [ + 84, + 299, + 460, + 487 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef981f5c6a705963db74de49292c32dcf335c11394113a178c7cb6c8024b5a34.jpg", + "image_caption": [ + "Figure 11. Controlled layout generation. Our model allows users to add conditions on design elements to customize the result. In this example, the user specifies the location of the balloons." + ], + "image_footnote": [], + "bbox": [ + 84, + 537, + 460, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of design applications.", + "text_level": 1, + "bbox": [ + 76, + 750, + 227, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 9, we produce various visual designs with the same spatial layout conforming to given element boxes by setting the attention scores for each element accordingly. This unlocks the power of the pretrained large image model for visual style variation.", + "bbox": [ + 75, + 763, + 467, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 10, we generate layouts of different canvas sizes by changing the input condition and masking the intermediate latent states during the denoising process.", + "bbox": [ + 76, + 839, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 11, when a user specifies the position of a cer", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/adb0818d2ac48da261c4fd8d8f4954f51627ad0d8445b29ead68423f8bef0764.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 92, + 679, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eeb502ff6e0799c93e6d7a8c03074aa54b59ea0518a827af461f7ebf12b7c2b9.jpg", + "image_caption": [ + "Original Design" + ], + "image_footnote": [], + "bbox": [ + 514, + 223, + 679, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58c3547cb4d1512c4dea0f08d5683794a042432cd84b965977df31dc772e2448.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 92, + 779, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/006f2bfa06bb2aa4128048839ceccb8063f3ba411025be83642a36c38fc14564.jpg", + "image_caption": [ + "Input Style", + "Figure 12. Layout style transfer. Our model can transfer the style from the reference design to the input design." + ], + "image_footnote": [], + "bbox": [ + 681, + 223, + 779, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/371cdc007553cd04044be88736bfce118eadf8aaa8dd2f7f2fcba3aa1afd9fd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 92, + 880, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c0d5d46a62b5fd4422cdc455f2d432806d5af9880fac4aaf532f87ef870826d.jpg", + "image_caption": [ + "New Design" + ], + "image_footnote": [], + "bbox": [ + 784, + 223, + 880, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tain element (the balloons of digits “20”) to a desired position, we fix the attention map for this element during inference, and only optimize the layout for the remaining elements.", + "bbox": [ + 498, + 428, + 890, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 12, we transfer the layout from one design image to a new set of elements by propagating the attention maps from an existing design to a new one. The image model takes a given reference image as the initial noisy latent state, accompanied by a new set of elements. During inference, the image model aligns the visual content to input elements while preserving the initial reference layout. The vector model utilizes the aligned attentions to create a new design incorporating both the input elements and reference layout.", + "bbox": [ + 498, + 489, + 892, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 652, + 617, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our paper introduces a new approach to layout design, combining vector and image domains through a dual diffusion model. Compared to previous methods solely based on the vector domain, this model has shown notable improvement on layout visual quality while maintaining the flexibility of vector editing on several datasets. Our model also shows flexible controllability via element attention manipulation, enabling a set of design applications across image and vector domains.", + "bbox": [ + 496, + 678, + 890, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we have made strides in integrating visual elements into layout design, our model is an initial step in addressing this complex task. Our findings offer a new perspective to the layout generation problem, and point out a potential direction for future exploration.", + "bbox": [ + 496, + 814, + 890, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9229", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 1", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[4] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2", + "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 5", + "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2", + "[7] Chin-Yi Cheng, Forrest Huang, Gang Li, and Yang Li. Play: Parametrically conditioned layout generation using latent diffusion. arXiv preprint arXiv:2301.11529, 2023. 2", + "[8] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual ACM symposium on user interface software and technology, pages 845-854, 2017. 1, 5", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[10] Vedit Goel, Elia Peruzzo, Yifan Jiang, Dejia Xu, Nicu Sebe, Trevor Darrell, Zhangyang Wang, and Humphrey Shi. Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546, 2023. 3", + "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 5, 6", + "[12] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004-1014, 2021. 5", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4", + "[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2", + "[15] Hsiao Yuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 2", + "[16] Mude Hui, Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie, Yuwang Wang, and Yan Lu. Unifying layout generation with a decoupled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1951, 2023. 2", + "[17] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. LayoutDM: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 2, 5, 6", + "[18] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Towards flexible multi-modal document models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14287-14296, 2023. 1, 2, 5, 6", + "[19] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layoutvae: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 5", + "[20] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via latent optimization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 88–96, 2021. 2, 5", + "[21] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In ICCV, 2023. 3", + "[22] Xiang Kong, Lu Jiang, Huiwen Chang, Han Zhang, Yuan Hao, Haifeng Gong, and Irfan Essa. Blt: bidirectional layout transformer for controllable layout generation. In European Conference on Computer Vision, pages 474-490. Springer, 2022. 5", + "[23] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 2", + "[24] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 491-506. Springer, 2020. 5", + "[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9230", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 5", + "[26] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2", + "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2", + "[28] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4", + "[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[33] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. arXiv preprint arXiv:2303.00262, 2023. 3", + "[34] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. Towards diverse and consistent typography generation. arXiv preprint arXiv:2309.02099, 2023. 2", + "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[36] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Layoutnuwa: Revealing the hidden layout expertise of large language models. arXiv preprint arXiv:2309.09506, 2023. 2", + "[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2", + "[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3", + "[39] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2", + "[40] Kota Yamaguchi. Canvasvae: Learning to generate vector graphic documents. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5481-5489, 2021. 1, 2, 4", + "[41] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2", + "[42] Junyi Zhang, Jiaqi Guo, Shizhao Sun, Jian-Guang Lou, and Dongmei Zhang. Layoutdiffusion: Improving graphic layout generation by discrete diffusion probabilistic models. arXiv preprint arXiv:2303.11589, 2023. 2", + "[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3", + "[44] Xinru Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG), 38(4):1-15, 2019. 2", + "[45] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publ aaynet: largest dataset ever for document layout analysis. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1015-1022. IEEE, 2019. 1, 5", + "[46] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout gan for visual-textual presentation designs. arXiv preprint arXiv:2205.00303, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 656 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9231", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_model.json b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ea99d85c8bf31490d3f40ed44c144800102b83d8 --- /dev/null +++ b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.8, + 0.177 + ], + "angle": 0, + "content": "Visual LayoutComposer:Image-Vector Dual Diffusion Model for DesignLayout Generation" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.203, + 0.804, + 0.257 + ], + "angle": 0, + "content": "Mohammad Amin Shabani\\(^{1*}\\) Zhaowen Wang\\(^{2}\\) Difan Liu\\(^{2}\\) Nanxuan Zhao\\(^{2}\\) Jimei Yang\\(^{2}\\) Yasutaka Furukawa\\(^{1}\\) Simon Fraser University \\(^{2}\\)Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.272, + 0.585, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.276, + 0.895, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.48, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Figure 1. Overview of our method. Left: our image-vector dual diffusion model architecture where both image and vector diffusion are conditioned on input elements and exchanging information with attention maps. Right: a comparison showing the advantage of dual-domain (image and vector) over traditional single-domain approaches." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.535, + 0.314, + 0.55 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.474, + 0.794 + ], + "angle": 0, + "content": "This paper proposes an image-vector dual diffusion model for generative layout design. Distinct from prior efforts that mostly ignores visual information of elements and the whole canvas, our approach integrates the power of a pre-trained large image diffusion model to guide layout composition in a vector diffusion model by providing enhanced salient region understanding and high-level inter-element relationship reasoning. Our proposed model simultaneously operates in two domains: it generates the overall design appearance in the image domain while optimizing the size and position of each design element in the vector domain. The proposed method achieves the state-of-the-art results on several datasets and enables new layout design applications. Project webpage: https://aminshabani.github.io/visual.layout_composer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.21, + 0.821 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.831, + 0.47, + 0.878 + ], + "angle": 0, + "content": "Layout design is essential in creative tasks, determining the spatial arrangement of texts, images, and other visual elements, which are composed into a visual design, such as" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.536, + 0.892, + 0.628 + ], + "angle": 0, + "content": "book and magazine covers, posters showcasing new products, or invitation cards for holiday parties. Graphic designers rely on their keen artistic sensibility in meticulously optimizing the arrangements of the visual elements while ensuring that the composited image conveys the intended message with aesthetic appeals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Machine learning techniques have been successful in directly regressing the layout vector parameters, i.e., the bounding box coordinates of elements. Both transformer-based models [1] and more recent diffusion-based models [18] have achieved promising results that capture basic geometrical properties such as alignment and symmetry. However, most of these models ignore the visual information from the input elements, including image content, text typeface, and color/shape of vector arts. This leads to unsatisfactory designs as illustrated by the Vector Only results in Figure 1. The lack of layout datasets with associated images also contributes to the negligence of visual information. The most commonly used RICO [8] and Pub-LayNet [45] datasets only contain box representation of design elements. The Crello [40] dataset provides element-level image rendering for over 23K design templates, whose visual features are used in the FlexDM [18] model. However, the interplay of element appearance when rendered on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.446, + 0.901 + ], + "angle": 0, + "content": "*This work was done during the first author's internship at Adobe." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9222" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.462, + 0.108 + ], + "angle": 0, + "content": "the same canvas has never been studied at any large scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.336 + ], + "angle": 0, + "content": "The recent success of image generative models [6, 9, 30, 32, 41] has enabled the generation of layout design in the image space. Trained on millions of high quality design images from web, these large generative models produce visually diverse and compelling layouts as raster images, demonstrating sophisticated design principles such as the placement of major subjects, style coherence, and perceptual contrast. See the Image Only results as examples in Figure 1. These visual properties, vital to layout design, are learned in a holistic way, not straightforward by previous vector domain approaches. Despite high visual quality, the layout designs in image format forbid further editing and customization. Moreover, current generative models are controlled only by text prompts, and cannot use given visual assets as required in many real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.471, + 0.594 + ], + "angle": 0, + "content": "Motivated by the complementary strengths of the vector and image domains, we propose a combined approach that generates layout designs in both image and vector domains simultaneously. An image branch is a pre-trained text-to-image stable diffusion model [30], which refines (i.e., noises) a design as an image, given a variable number of design elements with different media types. A vector branch refines per-element bounding box parameters conditioned on attention maps from the UNet of the image branch. Ground truth element masks are used to apply loss directly to these attention maps, whose manipulation at generation time will allow users to control the positioning of elements. This dual diffusion model refines the visual design in the vector and image domains simultaneously. The final vector output assembles the input elements and renders the visual design while allowing further editing by users. The overall proposed method is illustrated on the left of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.746 + ], + "angle": 0, + "content": "For training large diffusion models, we have collected a large-scale design Poster dataset from public design web pages by extracting both images and metadata. This effort yields 140K image layout paired samples as our own dataset, more than six times of the current largest visual layout dataset Crello with 23K samples. We conduct experiments on both datasets to validate the effectiveness of our dual diffusion model. Our data collection tools and processing pipeline can facilitate future efforts to curate even larger datasets for exploring visual design." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.749, + 0.4, + 0.763 + ], + "angle": 0, + "content": "Our key contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.765, + 0.468, + 0.839 + ], + "angle": 0, + "content": "- A novel dual diffusion model that generates layout design in both image and vector domains conditioned on input design elements. The two diffusion processes interact by exchanging element attention maps, where the final results attain both vector editability and visual quality." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.468, + 0.869 + ], + "angle": 0, + "content": "- Creative design applications, enabling element-wise user controls, such as canvas resizing and design variation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.898 + ], + "angle": 0, + "content": "- State-of-the-art layout generation results on the existing datasets and our Poster dataset." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.765, + 0.468, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.776, + 0.132 + ], + "angle": 0, + "content": "2.1. Vector-only Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Generative models for layout generation have seen significant developments. Yamaguchi [40] introduced a VAE-based architecture for unconditional generation of vector graphic documents. Following this, Li et al. [26] proposed LayoutGAN, which uses a relational generator and a wireframe renderer for training with a pixel-based discriminator. Kikuchi et al. [20] further refined this approach with LayoutGAN++, incorporating user-specified constraints into layout generation. Chai et al. [4] proposed using DDPM for layout generation. Inoue et al. proposed LayoutDM [17], a discrete diffusion-based model for layout designs. Zhang et al. [42] and Hui et al. [16] both contributed to this domain with LayoutDiffusion and LDGM, respectively, each based on discrete diffusion models. Cheng et al. [7] proposed a unique approach using guidelines as input conditions for the latent diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.794, + 0.407 + ], + "angle": 0, + "content": "2.2. Visual-guided Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Zhou et al. [46] tackled image-composition-aware layout generation with a multi-stage GAN-based method, an approach further improved by Xu et al. [39] through the use of a pixel-level discriminator. Zheng et al. [44] proposed a GAN based approach using image-based representation for layouts, which requires subsequent post-processing. Lin et al. [28] introduced a multi-step approach dedicated to the cleansing and retargeting of advertising posters. Hsu et al. [15] introduced DS-GAN, a CNN-LSTM-based conditional GAN method for poster layout generation. Inoue et al. [18] proposed Flex-DM, utilizing a transformer to predict masked attributes in design. Shimoda et al. [34] and Tang et al. [36] expanded the field by proposing new models for typography generation and layout generation, respectively. Our work differs by focusing on enhancing visual information and consistency through a large image diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.771, + 0.697 + ], + "angle": 0, + "content": "2.3. Image-based Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image diffusion model [35] was proposed initially for unconditional image generation. With the rapid development of the large-language model (LLM) [3], it has shown groundbreaking ability for text-to-image generation tasks [14]. To enhance the computational efficiency, rather than performing diffusion steps at the pixel level, latent diffusion model (LDM) [30] conducted operations on latent space, further boosting the generation ability of diffusion models. It has been widely used in various applications, including image-to-image translation [31], style transfer [23, 37], and condition-guided editing (e.g., sketch, layout, depth map, etc.) [27, 43]. To enable multi-object composition closer to our work, Xiao et al. [38] introduced" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9223" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "Localization Loss for free text-to-image generation with multiple subjects, while Sarukkai et al. [33] developed Collage Diffusion for creating realistic photos, focusing on harmonization and fidelity. Balaji et al. [2] training an ensemble of diffusion models specialized for different parts of the diffusion model to improve the overall text alignment of the model. Goel et al. [10] proposed Pair Diffusion to edit object properties in images via structure and appearance decomposition. Our model stands out by not needing predefined element positions and scales, handling multiple subjects, maintaining visual consistency, and providing enhanced editing flexibility." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.285, + 0.17, + 0.301 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.473, + 0.462 + ], + "angle": 0, + "content": "Our Visual LayoutComposer (VLC) model consists of two diffusion models that operate in parallel: one in the vector domain and the other in the image domain. The two models exchange features at intermediate stages, enabling them to produce outputs that are not only consistent but also leverage the information from both domains. In the following subsections, we first describe our problem formulation, followed by the specifics of the diffusion models for each domain, and how they collaborate to enhance output consistency. Figure 2 shows the model overview." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.208, + 0.485 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.473, + 0.659 + ], + "angle": 0, + "content": "Our input is a set of \\( N \\) design elements \\( E = \\{e_1, e_2, \\ldots, e_N\\} \\). Each element \\( e_i \\) is associated with a class category \\( c_i \\) indicating its asset type (such as text, image, shape), an index \\( o_i \\) indicating its layer order relative to other elements, and an RGBA image \\( \\mathbf{I}_i \\) of its rendering. On a designated 2D canvas with aspect ratio \\( r_c \\), the task is to predict an upright bounding box for each element \\( e_i \\), defined by the 2D coordinates of the top left and bottom right corners as a 4D vector \\( x_i \\). The goal is to generate a layout design with high visual quality by composing each element \\( e_i \\) on the canvas with the position and size specified by \\( x_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.285, + 0.683 + ], + "angle": 0, + "content": "3.2. Condition Embedding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "After resizing each element image \\(\\mathbf{I}_i\\) of aspect ratio \\(r_i\\) to \\(224 \\times 224\\), we employ CLIP [29] followed by a linear layer to derive its feature representation in the model's hidden dimension \\(d\\). In parallel, we encode the vector attributes of each element by a vector attributes encoder. The vector condition is constructed by concatenating the element image ratio \\(r_i\\), canvas image ratio \\(r_c\\), one-hot encoded version of the element's type \\(c_i\\) and the order of the element \\(o_i\\). A linear layer maps the concatenated feature to the model's dimension \\(d\\). The image feature and the vector feature are concatenated and processed by a linear layer to form the final condition feature of size \\(N \\times d\\). For each image or vector diffusion model, this condition feature is fed into a transformer-based condition processing module (Figure 2)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "yielding outputs of sizes \\( N \\times 768 \\) for image condition feature and \\( N \\times d \\) for vector condition feature." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.134, + 0.707, + 0.149 + ], + "angle": 0, + "content": "3.3. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.158, + 0.892, + 0.279 + ], + "angle": 0, + "content": "The diffusion model in the image domain utilizes a pretrained Stable Diffusion V1.5 model [30], retaining all layers frozen except for the attention modules. During a given time step \\( t \\), a noisy latent image \\( z^t \\), representing the target layout rendering, will be passed to the U-Net. Contrary to the original Stable Diffusion, which uses text prompts, we pass the processed image condition feature as the input to the cross-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.385 + ], + "angle": 0, + "content": "The diffusion model in the vector domain produces the bounding box positions as a 4D vector \\( x_{i} \\). We embed the noisy input coordinates \\( x^{t} \\) to a \\( N \\times d \\) feature vector, concatenate with the vector condition feature, and map to the model's dimension \\( d \\) by a linear layer. These features are fed into a transformer-based diffusion denoiser to estimate the corresponding noise for each element \\( x_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Information exchange between the image and vector domains is challenging, where we found that direct feature space sharing as in ControlNet [43] is not effective. We use the processed attention scores from U-Net as a shared medium. The advantages of this method are: (i) it aligns the number of attention scores with the number of input elements, enabling a straightforward concatenation of each image mask to its respective vector model feature token; (ii) it provides a general indication of each element's location in the image domain, simplifying the interpretation for the vector domain; and (iii) it allows the vector domain to adjust these scores to enhance the U-Net's focusing capability, with a mechanism similar to Attention Modulation [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Concretely, we extract a set of attention scores from the cross-attention module of the U-Net's middle block. Given the scores with 8 heads and size \\(8 \\times 8\\) for each design element, we flatten the tensors to form a set of vectors with the total size of \\(N \\times 512\\). This feature is embedded to \\(N \\times 1024\\) and concatenated with each input element in the vector domain prior to the transformer-based diffusion denoiser. The transformer model produces an output of \\(N \\times 2048\\), which is passed through linear projection layers to yield denoised estimation of \\(x^{t - 1}\\) and the attention modulation residual of size \\(N \\times 512\\). We add the residual values to the original attentions to form the refined attention scores. These scores are sent back to the U-Net, promoting interactive and iterative information exchange between the two domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During the diffusion process, repetitive elements can lead to omissions or misalignments in the image. To mitigate this, we adopt the cross-attention localization loss [38]. This loss refines the attention scores of each region, ensuring a more accurate correspondence to its respective element. We apply this loss within the U-Net's attention modules to further enhance the quality of the attention scores." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9224" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.096, + 0.885, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed model. Given a set of element images and additional attributes, the model simultaneously employs dual diffusion processes to generate a cohesive image layout and the corresponding bounding boxes for each element." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.384, + 0.29, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.384, + 0.462, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.566 + ], + "angle": 0, + "content": "Figure 3. Generated image/vector layout without or with canvas masking. By masking intermediate outputs of the diffusion model in image domain, we optimize the image output for the canvas size and increase the consistency between the vector-image domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.63 + ], + "angle": 0, + "content": "The training losses for the image and vector diffusion models are the same as the original DDPM [13]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.383, + 0.665 + ], + "angle": 0, + "content": "3.4. Controllable Element Composition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our experiments showed that fine-tuning the image domain led the model to disregard the specified canvas aspect ratio, favoring the generation of square-shaped layout designs. This tendency not only reduces the quality and diversity of the generated designs but also increases the discrepancy between the image and vector domains, particularly given that the vector domain generally follows the canvas's aspect ratio. To address this problem, we introduce a straightforward yet efficient strategy: masking the noisy latent feature map of \\( z^t \\) at each time step to have zero value outside the canvas region, which ensures image contents related to input condition only appear inside the canvas. The homogeneous region outside the canvas is cropped from the final generated image, maintaining the original image aspect ratio. Figure 3 shows the effect of this process on both domains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.634, + 0.395 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.895, + 0.681 + ], + "angle": 0, + "content": "Implementation details: In our implementation, we use the pretrained weights of StableDiffusion v1-5 [30] for image domain, and OpenAI's clip-vit-large-patch14 vision model for embedding the image of each element to latent space. For training, we run each model for 250,000 iterations, employing a starting learning rate of 5e-5 which decreases by a factor of 0.5 every 100k steps. The vector model has a hidden dimension of 2048, accommodating a wide range of features including attention scores, vector inputs, and image information. Additionally, we incorporate classifier-free guidance into our process, randomly dropping the image inputs of the elements in \\(20\\%\\) of the times. Training is conducted on 8 NVIDIA A100 GPUs, with a total batch size of 128. We pad each layout to square shape and normalize the coordinates to \\([-1,1]\\). For both domains, we employ the DDPM scheduler with 1000 steps for training and 50 steps for inference to ensure high-quality layout generation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.687, + 0.862, + 0.702 + ], + "angle": 0, + "content": "Datasets: Our experiments use two groups of datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "- Poster and Crello [40] datasets: demonstrating our method's ability to capture intricate visual information and the interplay between design elements. These datasets, characterized by their rich graphic content, are instrumental in showcasing the method's effectiveness in a visually complex environment. We gathered Poster dataset, which consists of 137,781 layout designs. From this, we reserved 1,000 layouts for validation and 3,000 for testing. Each layout is represented as an RGBA image and is classified into one of five categories: Frame, Canvas, Shape, Mix, and Text. We have limited the number of elements in any given layout to a maximum of 20 to maintain structural simplicity. Given the intricate nature of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9225" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "these layouts, we have also provided the sequential order of the elements within each design. The Crello datasets include 18,768 training, 2,316 validation, and 2,331 test layouts including categories such as vector shape, image, or text placeholder." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.168, + 0.47, + 0.455 + ], + "angle": 0, + "content": "- RICO [8] and PubLayNet [45] datasets: having limited visual information, which challenges our method. To compensate for this, we assign distinct colors to each category and depict layouts with these colors, simplifying the visuals to fixed-color blocks representing each element. This approach tests our method's ability in generating layouts based primarily on structure rather than rich visual details. The Rico dataset consists of user interface (UI) designs sourced from various Android applications, spanning up to 25 categories of UI elements including text bars, icons, and buttons. PubLayNet includes 360,000 document layout instances from academic papers, categorized into five primary layout elements: text, titles, lists, tables, and figures. We use the processed datasets by [17], obtained by keeping only those layouts containing a maximum of 25 elements each. This results training/validation/test sets of sizes 35,851/2,109/4,218 respectively for Rico and 315,757/16,619/11,142 for PubLayNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.465, + 0.216, + 0.481 + ], + "angle": 0, + "content": "4.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.468, + 0.655 + ], + "angle": 0, + "content": "For the baselines, we compare our method with recent state-of-the-art methods in Layout Design Generation [5, 11, 12, 17, 19, 20, 22, 24, 25]. We compare our method to the LayoutDM [17] as the state-of-the-art for layout generation. This model only uses element categories, adhering to their original design in discrete diffusion space. Our vector domain model operates in a continuous diffusion space and incorporates each element's image embedding. Finally, our dual-domain model combines element image embeddings with the capabilities of the image diffusion model to get the highest quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.676, + 0.298, + 0.691 + ], + "angle": 0, + "content": "4.1.1 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Metrics: We follow the previous work [17] and use metrics including Maximum IoU [20] and FID over extracted features of the layouts [20, 24] for RICO and PubLayNet datasets. The previous methods and measurements are effective when there are no visual data or input images for layout elements. However, assessing layout designs in Poster and Crello datasets is more complex. Even layouts with well-aligned bounding boxes might look poor due to the textures and visual details of each element. Elements can also overlap for shadow or visual effects. Inoue et al. [18] approached this as a reconstruction task, using metrics like cosine similarity between the actual and predicted values. This approach works for simple designs with few elements" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.885, + 0.223 + ], + "angle": 0, + "content": "
DatasetModelsBounding BoxComposed ImageGenerated Image
CrelloLayoutDM [17]10.2034.43—
VLC (Vector only)0.215.74—
VLC (Dual-domain)0.215.8311.42
PosterLayoutDM [17]1.8120.00—
VLC (Vector only)0.195.82—
VLC (Dual-domain)0.093.756.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Table 1. FID scores of bounding boxes, composed RGB layouts obtained from the bounding boxes, and the directly generated layouts from the image domain. Our vector-only method outperforms LayoutDM by effectively using image information, with our multi-domain approach yielding the best results." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.317, + 0.892, + 0.547 + ], + "angle": 0, + "content": "
ModelsRICOPubLayNet
FID (↓)mIoU (↑)FID (↓)mIoU (↑)
LayoutVAE [19]33.30.24926.00.316
NDN-none [24]28.40.15861.10.162
LayoutGAN++ [20]6.840.26724.00.263
LayoutTrans [12]5.570.22314.10.272
MaskGIT [5]26.10.26217.20.319
BLT [22]17.40.20272.10.215
BART [25]3.970.2539.360.320
VQDiffusion [11]4.340.25210.30.319
LayoutDM [17]3.550.2777.950.310
VLC (Vector only)2.380.4185.490.348
VLC (Dual-domain)5.600.3765.740.331
Validation Set1.850.6916.250.438
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of our method with the baselines using RICO and PubLayNet datasets. Our method achieves superior results compared with the baselines even with no visual information in input conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.777 + ], + "angle": 0, + "content": "but limits diversity and learning on more complex datasets. Instead, we propose using FID values over three outputs: bounding boxes, feature embeddings from RGB images created by composing elements within generated bounding boxes, and the directly produced RGB images. Although this is our best option, we realize that FID can be biased to image quality and content similarity which should be removed from the consideration for layout evaluation. Finding better visual metric for layout can be a future direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results: We evaluate our VLC method using visual information as shown in Table 1. Our dual-domain method stands out for its ability to create visually appealing images by using a large image model. This method is a significant improvement over the LayoutDM [17]. LayoutDM uses only element categories as input, which, while effective in creating quality vectors, often fails to align with the actual visual content of each element. This mismatch results in a" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9226" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.094, + 0.213, + 0.106 + ], + "angle": 0, + "content": "VLC (vector only)" + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.094, + 0.324, + 0.106 + ], + "angle": 0, + "content": "VLC (dual domain)" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.094, + 0.423, + 0.106 + ], + "angle": 0, + "content": "Original Design" + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.108, + 0.44, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.343, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of our method with baselines on Crello dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.581 + ], + "angle": 0, + "content": "drop in the overall quality of the final image. Our vector domain approach is the most similar one to the transformer-based reconstruction used in FlexDM [18]. It includes both image information and other input conditions, and produces higher-quality images compared to LayoutDM, but has some drawbacks such as occasionally covering faces or not managing color contrast well when elements overlap. Our vector only model and dual-domain model perform similarly on the Crello dataset. The reason can be attributed to the larger number of elements in the dataset, which leads to the content mismatch between image and vector domain such that vector model cannot follow the guidance from image correctly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Table 2, we compare our method with existing baselines on the RICO and PubLayNet datasets. The baseline results are reported by Inoue et al. [17]. Remarkably, even when the input elements are limited to only categorical data without any visual information, our method demonstrated its capability to produce high-quality layouts in both vector-only and dual-domain settings. In the vector-only domain, we observed a trend where training appeared less stable. However, despite this instability, the results often surpassed those generated by discrete diffusion models [11, 17]. In addition, the vector domain model tends to align with the performance of our dual-domain model. This alignment is particularly notable as our dual-domain method achieves better results than most of the baselines, underscoring its effectiveness even in the absence of visual input. It is important to note that while our method is slower than the baselines duo to the amount of computation, this trade-off is compensated by its high-quality output and robustness in processing visual information with its diffusion formulation. Overall, these findings suggest that our method not only achieves state-of-the-art performance but also offers promising po" + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.097, + 0.632, + 0.11 + ], + "angle": 0, + "content": "VLC (vector only)" + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.097, + 0.752, + 0.11 + ], + "angle": 0, + "content": "VLC (dual domain)" + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.097, + 0.861, + 0.111 + ], + "angle": 0, + "content": "Original Design" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.112, + 0.877, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.609, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of our method with baselines on Poster dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.653, + 0.892, + 0.684 + ], + "angle": 0, + "content": "tential for generating high-quality layouts under varied and challenging conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.701, + 0.711, + 0.716 + ], + "angle": 0, + "content": "4.1.2 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.724, + 0.892, + 0.815 + ], + "angle": 0, + "content": "The effectiveness of our dual-domain method is further illustrated in Figure 4&5 with visual results from the Crello and Poster datasets. These examples highlight our approach not only improves the layout quality but also ensures that the final compositions are more visually coherent and appealing, especially when compared to other methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.824, + 0.655, + 0.839 + ], + "angle": 0, + "content": "4.2. Model Analysis" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.863, + 0.862 + ], + "angle": 0, + "content": "4.2.1 The Effectiveness on Dual-domain Modeling" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addressing the complexities of dual-domain modeling, a crucial challenge lies in maintaining consistency between" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9227" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.272, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.174, + 0.225, + 0.188 + ], + "angle": 0, + "content": "Original Layout" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.093, + 0.464, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.15, + 0.464, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.093, + 0.654, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.15, + 0.655, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.093, + 0.845, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.15, + 0.845, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.105, + 0.863, + 0.195 + ], + "angle": 270, + "content": "eaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeae" + }, + { + "type": "image_caption", + "bbox": [ + 0.865, + 0.097, + 0.882, + 0.204 + ], + "angle": 270, + "content": "psodmooO O" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.211, + 0.22, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.312, + 0.221, + 0.325 + ], + "angle": 0, + "content": "Original Layout" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.21, + 0.464, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.312, + 0.348, + 0.323 + ], + "angle": 0, + "content": "Generated" + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.326, + 0.337, + 0.338 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.312, + 0.442, + 0.324 + ], + "angle": 0, + "content": "Composed" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.326, + 0.429, + 0.338 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.209, + 0.653, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.312, + 0.544, + 0.323 + ], + "angle": 0, + "content": "Generated" + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.326, + 0.532, + 0.339 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.312, + 0.638, + 0.324 + ], + "angle": 0, + "content": "Composed" + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.326, + 0.624, + 0.339 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.21, + 0.843, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.312, + 0.732, + 0.323 + ], + "angle": 0, + "content": "Generated" + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.326, + 0.72, + 0.339 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.312, + 0.825, + 0.324 + ], + "angle": 0, + "content": "Composed" + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.326, + 0.812, + 0.338 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.357, + 0.831, + 0.372 + ], + "angle": 0, + "content": "Figure 6. Our approach is capable of producing a wide range of consistent designs in both image and vector domains." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.389, + 0.464, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.634 + ], + "angle": 0, + "content": "Figure 7. Attention scores are assigned to various elements in the image domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.798 + ], + "angle": 0, + "content": "the image and vector domains. This is vividly illustrated in Figure 3, which demonstrates that when these two domains adhere to a consistent denoising path, the result quality is significantly enhanced. Conversely, deviation from this path leads to a noticeable drop in quality. Complementing this, Figure 6 highlights the efficacy of our method, showcasing that both the image and vector domains can generate results that are not only consistent with each other but also retain the diversity of the generated layouts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.379, + 0.833 + ], + "angle": 0, + "content": "4.2.2 The Effect of Attention Score Maps" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The attention score maps are the key for the visual composing capability of our model. We visualize the clear spatial correlation between attention scores and corresponding elements in the image domain in Figure 7. Our vector" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.389, + 0.885, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Figure 8. Through attention masking, we can decompose all the elements in the latent image space, and include them one by one to create a layered image representation. The bottom right image shows the original design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.703 + ], + "angle": 0, + "content": "domain model leverages these spatial information to predict vector layouts by adaptively tracking each element and resolving the ambiguity between similar objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.838 + ], + "angle": 0, + "content": "By manipulating the attention scores, we can also control the element generation in the image domain. For example, we can remove one element from canvas while keeping the rest layout unchanged by masking out its attention map. If we do this sequentially to each element, we can decompose image output into a layered representation organized by elements, as shown in Figure 8. This results in a virtual vector layout format, which may be used for similar purpose as our vector domain model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.634, + 0.864 + ], + "angle": 0, + "content": "4.3. Applications" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "With its two domain outputs and high controllability through attention maps, our model can be used for a range" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.092, + 0.18, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.225, + 0.18, + 0.235 + ], + "angle": 0, + "content": "Input Layout Boxes" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.092, + 0.278, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.092, + 0.371, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.092, + 0.463, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.224, + 0.384, + 0.235 + ], + "angle": 0, + "content": "Generated Layout Images" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.249, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Figure 9. Layout controlled visual style variation. Given input layout with bounding boxes, our image model can generate visual results with the same layout and different styles." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.3, + 0.462, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.501, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Figure 10. Canvas resize with different aspect ratios. Our model can adapt layout under various aspect ratios." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.539, + 0.462, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.689, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Figure 11. Controlled layout generation. Our model allows users to add conditions on design elements to customize the result. In this example, the user specifies the location of the balloons." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.75, + 0.228, + 0.764 + ], + "angle": 0, + "content": "of design applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.84 + ], + "angle": 0, + "content": "In Figure 9, we produce various visual designs with the same spatial layout conforming to given element boxes by setting the attention scores for each element accordingly. This unlocks the power of the pretrained large image model for visual style variation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.885 + ], + "angle": 0, + "content": "In Figure 10, we generate layouts of different canvas sizes by changing the input condition and masking the intermediate latent states during the denoising process." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In Figure 11, when a user specifies the position of a cer" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.093, + 0.68, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.224, + 0.68, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.342, + 0.646, + 0.356 + ], + "angle": 0, + "content": "Original Design" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.093, + 0.78, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.224, + 0.78, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.342, + 0.766, + 0.356 + ], + "angle": 0, + "content": "Input Style" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.093, + 0.881, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.224, + 0.881, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.342, + 0.87, + 0.356 + ], + "angle": 0, + "content": "New Design" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.374, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Figure 12. Layout style transfer. Our model can transfer the style from the reference design to the input design." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.488 + ], + "angle": 0, + "content": "tain element (the balloons of digits “20”) to a desired position, we fix the attention map for this element during inference, and only optimize the layout for the remaining elements." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.49, + 0.893, + 0.641 + ], + "angle": 0, + "content": "In Figure 12, we transfer the layout from one design image to a new set of elements by propagating the attention maps from an existing design to a new one. The image model takes a given reference image as the initial noisy latent state, accompanied by a new set of elements. During inference, the image model aligns the visual content to input elements while preserving the initial reference layout. The vector model utilizes the aligned attentions to create a new design incorporating both the input elements and reference layout." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.619, + 0.67 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Our paper introduces a new approach to layout design, combining vector and image domains through a dual diffusion model. Compared to previous methods solely based on the vector domain, this model has shown notable improvement on layout visual quality while maintaining the flexibility of vector editing on several datasets. Our model also shows flexible controllability via element attention manipulation, enabling a set of design applications across image and vector domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.815, + 0.892, + 0.891 + ], + "angle": 0, + "content": "While we have made strides in integrating visual elements into layout design, our model is an initial step in addressing this complex task. Our findings offer a new perspective to the layout generation problem, and point out a potential direction for future exploration." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9229" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[4] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.454 + ], + "angle": 0, + "content": "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.469, + 0.527 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[7] Chin-Yi Cheng, Forrest Huang, Gang Li, and Yang Li. Play: Parametrically conditioned layout generation using latent diffusion. arXiv preprint arXiv:2301.11529, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[8] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual ACM symposium on user interface software and technology, pages 845-854, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[10] Vedit Goel, Elia Peruzzo, Yifan Jiang, Dejia Xu, Nicu Sebe, Trevor Darrell, Zhangyang Wang, and Humphrey Shi. Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004-1014, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[15] Hsiao Yuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[16] Mude Hui, Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie, Yuwang Wang, and Yan Lu. Unifying layout generation with a decoupled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1951, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[17] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. LayoutDM: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[18] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Towards flexible multi-modal document models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14287-14296, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[19] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layoutvae: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[20] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via latent optimization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 88–96, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[21] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[22] Xiang Kong, Lu Jiang, Huiwen Chang, Han Zhang, Yuan Hao, Haifeng Gong, and Irfan Essa. Blt: bidirectional layout transformer for controllable layout generation. In European Conference on Computer Vision, pages 474-490. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[23] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 491-506. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9230" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[26] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[28] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.471, + 0.671 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.471, + 0.714 + ], + "angle": 0, + "content": "[33] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. arXiv preprint arXiv:2303.00262, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[34] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. Towards diverse and consistent typography generation. arXiv preprint arXiv:2309.02099, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[36] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Layoutnuwa: Revealing the hidden layout expertise of large language models. arXiv preprint arXiv:2309.09506, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[39] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[40] Kota Yamaguchi. Canvasvae: Learning to generate vector graphic documents. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5481-5489, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[41] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[42] Junyi Zhang, Jiaqi Guo, Shizhao Sun, Jian-Guang Lou, and Dongmei Zhang. Layoutdiffusion: Improving graphic layout generation by discrete diffusion probabilistic models. arXiv preprint arXiv:2303.11589, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.894, + 0.543 + ], + "angle": 0, + "content": "[44] Xinru Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG), 38(4):1-15, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[45] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publ aaynet: largest dataset ever for document layout analysis. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1015-1022. IEEE, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[46] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout gan for visual-textual presentation designs. arXiv preprint arXiv:2205.00303, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9231" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_origin.pdf b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd8ee384ed2d59ed9207aeff6b3911b89897b0c2 --- /dev/null +++ b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/75d905c0-508c-4fbc-a17b-7ac7fef387c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edb19a7f88d5fcf28a0c492e7b90bbf2ab6ebfd8c3a0aab8681fc7b37f6df4d3 +size 4351408 diff --git a/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/full.md b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dce5f8a4d0575075c4f1447e1ca7ddfdd5d17c62 --- /dev/null +++ b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/full.md @@ -0,0 +1,315 @@ +# Visual LayoutComposer:Image-Vector Dual Diffusion Model for DesignLayout Generation + +Mohammad Amin Shabani $^{1*}$ Zhaowen Wang $^{2}$ Difan Liu $^{2}$ Nanxuan Zhao $^{2}$ Jimei Yang $^{2}$ Yasutaka Furukawa $^{1}$ Simon Fraser University $^{2}$ Adobe Research + +![](images/098199d519b41b62d9ea2ba15c3a8aef1a6940e4d27606ef8e18900731fad987.jpg) +Figure 1. Overview of our method. Left: our image-vector dual diffusion model architecture where both image and vector diffusion are conditioned on input elements and exchanging information with attention maps. Right: a comparison showing the advantage of dual-domain (image and vector) over traditional single-domain approaches. + +![](images/a5f57ae3ca19cbe08b08aee90ce5581e26ce1c5ab73eeb1adb4a7d2cb4931df4.jpg) + +# Abstract + +This paper proposes an image-vector dual diffusion model for generative layout design. Distinct from prior efforts that mostly ignores visual information of elements and the whole canvas, our approach integrates the power of a pre-trained large image diffusion model to guide layout composition in a vector diffusion model by providing enhanced salient region understanding and high-level inter-element relationship reasoning. Our proposed model simultaneously operates in two domains: it generates the overall design appearance in the image domain while optimizing the size and position of each design element in the vector domain. The proposed method achieves the state-of-the-art results on several datasets and enables new layout design applications. Project webpage: https://aminshabani.github.io/visual.layout_composer. + +# 1. Introduction + +Layout design is essential in creative tasks, determining the spatial arrangement of texts, images, and other visual elements, which are composed into a visual design, such as + +book and magazine covers, posters showcasing new products, or invitation cards for holiday parties. Graphic designers rely on their keen artistic sensibility in meticulously optimizing the arrangements of the visual elements while ensuring that the composited image conveys the intended message with aesthetic appeals. + +Machine learning techniques have been successful in directly regressing the layout vector parameters, i.e., the bounding box coordinates of elements. Both transformer-based models [1] and more recent diffusion-based models [18] have achieved promising results that capture basic geometrical properties such as alignment and symmetry. However, most of these models ignore the visual information from the input elements, including image content, text typeface, and color/shape of vector arts. This leads to unsatisfactory designs as illustrated by the Vector Only results in Figure 1. The lack of layout datasets with associated images also contributes to the negligence of visual information. The most commonly used RICO [8] and Pub-LayNet [45] datasets only contain box representation of design elements. The Crello [40] dataset provides element-level image rendering for over 23K design templates, whose visual features are used in the FlexDM [18] model. However, the interplay of element appearance when rendered on + +the same canvas has never been studied at any large scale. + +The recent success of image generative models [6, 9, 30, 32, 41] has enabled the generation of layout design in the image space. Trained on millions of high quality design images from web, these large generative models produce visually diverse and compelling layouts as raster images, demonstrating sophisticated design principles such as the placement of major subjects, style coherence, and perceptual contrast. See the Image Only results as examples in Figure 1. These visual properties, vital to layout design, are learned in a holistic way, not straightforward by previous vector domain approaches. Despite high visual quality, the layout designs in image format forbid further editing and customization. Moreover, current generative models are controlled only by text prompts, and cannot use given visual assets as required in many real-world applications. + +Motivated by the complementary strengths of the vector and image domains, we propose a combined approach that generates layout designs in both image and vector domains simultaneously. An image branch is a pre-trained text-to-image stable diffusion model [30], which refines (i.e., noises) a design as an image, given a variable number of design elements with different media types. A vector branch refines per-element bounding box parameters conditioned on attention maps from the UNet of the image branch. Ground truth element masks are used to apply loss directly to these attention maps, whose manipulation at generation time will allow users to control the positioning of elements. This dual diffusion model refines the visual design in the vector and image domains simultaneously. The final vector output assembles the input elements and renders the visual design while allowing further editing by users. The overall proposed method is illustrated on the left of Figure 1. + +For training large diffusion models, we have collected a large-scale design Poster dataset from public design web pages by extracting both images and metadata. This effort yields 140K image layout paired samples as our own dataset, more than six times of the current largest visual layout dataset Crello with 23K samples. We conduct experiments on both datasets to validate the effectiveness of our dual diffusion model. Our data collection tools and processing pipeline can facilitate future efforts to curate even larger datasets for exploring visual design. + +Our key contributions are summarized below: + +- A novel dual diffusion model that generates layout design in both image and vector domains conditioned on input design elements. The two diffusion processes interact by exchanging element attention maps, where the final results attain both vector editability and visual quality. +- Creative design applications, enabling element-wise user controls, such as canvas resizing and design variation. +- State-of-the-art layout generation results on the existing datasets and our Poster dataset. + +# 2. Related Works + +# 2.1. Vector-only Layout Generation + +Generative models for layout generation have seen significant developments. Yamaguchi [40] introduced a VAE-based architecture for unconditional generation of vector graphic documents. Following this, Li et al. [26] proposed LayoutGAN, which uses a relational generator and a wireframe renderer for training with a pixel-based discriminator. Kikuchi et al. [20] further refined this approach with LayoutGAN++, incorporating user-specified constraints into layout generation. Chai et al. [4] proposed using DDPM for layout generation. Inoue et al. proposed LayoutDM [17], a discrete diffusion-based model for layout designs. Zhang et al. [42] and Hui et al. [16] both contributed to this domain with LayoutDiffusion and LDGM, respectively, each based on discrete diffusion models. Cheng et al. [7] proposed a unique approach using guidelines as input conditions for the latent diffusion model. + +# 2.2. Visual-guided Layout Generation + +Zhou et al. [46] tackled image-composition-aware layout generation with a multi-stage GAN-based method, an approach further improved by Xu et al. [39] through the use of a pixel-level discriminator. Zheng et al. [44] proposed a GAN based approach using image-based representation for layouts, which requires subsequent post-processing. Lin et al. [28] introduced a multi-step approach dedicated to the cleansing and retargeting of advertising posters. Hsu et al. [15] introduced DS-GAN, a CNN-LSTM-based conditional GAN method for poster layout generation. Inoue et al. [18] proposed Flex-DM, utilizing a transformer to predict masked attributes in design. Shimoda et al. [34] and Tang et al. [36] expanded the field by proposing new models for typography generation and layout generation, respectively. Our work differs by focusing on enhancing visual information and consistency through a large image diffusion model. + +# 2.3. Image-based Diffusion Models + +Image diffusion model [35] was proposed initially for unconditional image generation. With the rapid development of the large-language model (LLM) [3], it has shown groundbreaking ability for text-to-image generation tasks [14]. To enhance the computational efficiency, rather than performing diffusion steps at the pixel level, latent diffusion model (LDM) [30] conducted operations on latent space, further boosting the generation ability of diffusion models. It has been widely used in various applications, including image-to-image translation [31], style transfer [23, 37], and condition-guided editing (e.g., sketch, layout, depth map, etc.) [27, 43]. To enable multi-object composition closer to our work, Xiao et al. [38] introduced + +Localization Loss for free text-to-image generation with multiple subjects, while Sarukkai et al. [33] developed Collage Diffusion for creating realistic photos, focusing on harmonization and fidelity. Balaji et al. [2] training an ensemble of diffusion models specialized for different parts of the diffusion model to improve the overall text alignment of the model. Goel et al. [10] proposed Pair Diffusion to edit object properties in images via structure and appearance decomposition. Our model stands out by not needing predefined element positions and scales, handling multiple subjects, maintaining visual consistency, and providing enhanced editing flexibility. + +# 3. Method + +Our Visual LayoutComposer (VLC) model consists of two diffusion models that operate in parallel: one in the vector domain and the other in the image domain. The two models exchange features at intermediate stages, enabling them to produce outputs that are not only consistent but also leverage the information from both domains. In the following subsections, we first describe our problem formulation, followed by the specifics of the diffusion models for each domain, and how they collaborate to enhance output consistency. Figure 2 shows the model overview. + +# 3.1. Preliminary + +Our input is a set of $N$ design elements $E = \{e_1, e_2, \ldots, e_N\}$ . Each element $e_i$ is associated with a class category $c_i$ indicating its asset type (such as text, image, shape), an index $o_i$ indicating its layer order relative to other elements, and an RGBA image $\mathbf{I}_i$ of its rendering. On a designated 2D canvas with aspect ratio $r_c$ , the task is to predict an upright bounding box for each element $e_i$ , defined by the 2D coordinates of the top left and bottom right corners as a 4D vector $x_i$ . The goal is to generate a layout design with high visual quality by composing each element $e_i$ on the canvas with the position and size specified by $x_i$ . + +# 3.2. Condition Embedding + +After resizing each element image $\mathbf{I}_i$ of aspect ratio $r_i$ to $224 \times 224$ , we employ CLIP [29] followed by a linear layer to derive its feature representation in the model's hidden dimension $d$ . In parallel, we encode the vector attributes of each element by a vector attributes encoder. The vector condition is constructed by concatenating the element image ratio $r_i$ , canvas image ratio $r_c$ , one-hot encoded version of the element's type $c_i$ and the order of the element $o_i$ . A linear layer maps the concatenated feature to the model's dimension $d$ . The image feature and the vector feature are concatenated and processed by a linear layer to form the final condition feature of size $N \times d$ . For each image or vector diffusion model, this condition feature is fed into a transformer-based condition processing module (Figure 2), + +yielding outputs of sizes $N \times 768$ for image condition feature and $N \times d$ for vector condition feature. + +# 3.3. Network Architecture + +The diffusion model in the image domain utilizes a pretrained Stable Diffusion V1.5 model [30], retaining all layers frozen except for the attention modules. During a given time step $t$ , a noisy latent image $z^t$ , representing the target layout rendering, will be passed to the U-Net. Contrary to the original Stable Diffusion, which uses text prompts, we pass the processed image condition feature as the input to the cross-attention layers. + +The diffusion model in the vector domain produces the bounding box positions as a 4D vector $x_{i}$ . We embed the noisy input coordinates $x^{t}$ to a $N \times d$ feature vector, concatenate with the vector condition feature, and map to the model's dimension $d$ by a linear layer. These features are fed into a transformer-based diffusion denoiser to estimate the corresponding noise for each element $x_{i}$ . + +Information exchange between the image and vector domains is challenging, where we found that direct feature space sharing as in ControlNet [43] is not effective. We use the processed attention scores from U-Net as a shared medium. The advantages of this method are: (i) it aligns the number of attention scores with the number of input elements, enabling a straightforward concatenation of each image mask to its respective vector model feature token; (ii) it provides a general indication of each element's location in the image domain, simplifying the interpretation for the vector domain; and (iii) it allows the vector domain to adjust these scores to enhance the U-Net's focusing capability, with a mechanism similar to Attention Modulation [21]. + +Concretely, we extract a set of attention scores from the cross-attention module of the U-Net's middle block. Given the scores with 8 heads and size $8 \times 8$ for each design element, we flatten the tensors to form a set of vectors with the total size of $N \times 512$ . This feature is embedded to $N \times 1024$ and concatenated with each input element in the vector domain prior to the transformer-based diffusion denoiser. The transformer model produces an output of $N \times 2048$ , which is passed through linear projection layers to yield denoised estimation of $x^{t - 1}$ and the attention modulation residual of size $N \times 512$ . We add the residual values to the original attentions to form the refined attention scores. These scores are sent back to the U-Net, promoting interactive and iterative information exchange between the two domains. + +During the diffusion process, repetitive elements can lead to omissions or misalignments in the image. To mitigate this, we adopt the cross-attention localization loss [38]. This loss refines the attention scores of each region, ensuring a more accurate correspondence to its respective element. We apply this loss within the U-Net's attention modules to further enhance the quality of the attention scores. + +![](images/fce675491ca02073f9bf1d1306b4b0f0f1d8b8883a2fc949cd90e621367618f5.jpg) +Figure 2. Illustration of the proposed model. Given a set of element images and additional attributes, the model simultaneously employs dual diffusion processes to generate a cohesive image layout and the corresponding bounding boxes for each element. + +![](images/d1f6d9291135a11c36624b2838f4defc26fa548a45c43a917ed17015f80f682c.jpg) +Figure 3. Generated image/vector layout without or with canvas masking. By masking intermediate outputs of the diffusion model in image domain, we optimize the image output for the canvas size and increase the consistency between the vector-image domains. + +![](images/2ddfc208f079e068970ec4e8e738ea1df4c2079494fafb9197b38ae9991b30b1.jpg) + +The training losses for the image and vector diffusion models are the same as the original DDPM [13]. + +# 3.4. Controllable Element Composition + +Our experiments showed that fine-tuning the image domain led the model to disregard the specified canvas aspect ratio, favoring the generation of square-shaped layout designs. This tendency not only reduces the quality and diversity of the generated designs but also increases the discrepancy between the image and vector domains, particularly given that the vector domain generally follows the canvas's aspect ratio. To address this problem, we introduce a straightforward yet efficient strategy: masking the noisy latent feature map of $z^t$ at each time step to have zero value outside the canvas region, which ensures image contents related to input condition only appear inside the canvas. The homogeneous region outside the canvas is cropped from the final generated image, maintaining the original image aspect ratio. Figure 3 shows the effect of this process on both domains. + +# 4. Experiments + +Implementation details: In our implementation, we use the pretrained weights of StableDiffusion v1-5 [30] for image domain, and OpenAI's clip-vit-large-patch14 vision model for embedding the image of each element to latent space. For training, we run each model for 250,000 iterations, employing a starting learning rate of 5e-5 which decreases by a factor of 0.5 every 100k steps. The vector model has a hidden dimension of 2048, accommodating a wide range of features including attention scores, vector inputs, and image information. Additionally, we incorporate classifier-free guidance into our process, randomly dropping the image inputs of the elements in $20\%$ of the times. Training is conducted on 8 NVIDIA A100 GPUs, with a total batch size of 128. We pad each layout to square shape and normalize the coordinates to $[-1,1]$ . For both domains, we employ the DDPM scheduler with 1000 steps for training and 50 steps for inference to ensure high-quality layout generation. + +Datasets: Our experiments use two groups of datasets. + +- Poster and Crello [40] datasets: demonstrating our method's ability to capture intricate visual information and the interplay between design elements. These datasets, characterized by their rich graphic content, are instrumental in showcasing the method's effectiveness in a visually complex environment. We gathered Poster dataset, which consists of 137,781 layout designs. From this, we reserved 1,000 layouts for validation and 3,000 for testing. Each layout is represented as an RGBA image and is classified into one of five categories: Frame, Canvas, Shape, Mix, and Text. We have limited the number of elements in any given layout to a maximum of 20 to maintain structural simplicity. Given the intricate nature of + +these layouts, we have also provided the sequential order of the elements within each design. The Crello datasets include 18,768 training, 2,316 validation, and 2,331 test layouts including categories such as vector shape, image, or text placeholder. + +- RICO [8] and PubLayNet [45] datasets: having limited visual information, which challenges our method. To compensate for this, we assign distinct colors to each category and depict layouts with these colors, simplifying the visuals to fixed-color blocks representing each element. This approach tests our method's ability in generating layouts based primarily on structure rather than rich visual details. The Rico dataset consists of user interface (UI) designs sourced from various Android applications, spanning up to 25 categories of UI elements including text bars, icons, and buttons. PubLayNet includes 360,000 document layout instances from academic papers, categorized into five primary layout elements: text, titles, lists, tables, and figures. We use the processed datasets by [17], obtained by keeping only those layouts containing a maximum of 25 elements each. This results training/validation/test sets of sizes 35,851/2,109/4,218 respectively for Rico and 315,757/16,619/11,142 for PubLayNet. + +# 4.1. Comparisons + +For the baselines, we compare our method with recent state-of-the-art methods in Layout Design Generation [5, 11, 12, 17, 19, 20, 22, 24, 25]. We compare our method to the LayoutDM [17] as the state-of-the-art for layout generation. This model only uses element categories, adhering to their original design in discrete diffusion space. Our vector domain model operates in a continuous diffusion space and incorporates each element's image embedding. Finally, our dual-domain model combines element image embeddings with the capabilities of the image diffusion model to get the highest quality. + +# 4.1.1 Quantitative Evaluation + +Metrics: We follow the previous work [17] and use metrics including Maximum IoU [20] and FID over extracted features of the layouts [20, 24] for RICO and PubLayNet datasets. The previous methods and measurements are effective when there are no visual data or input images for layout elements. However, assessing layout designs in Poster and Crello datasets is more complex. Even layouts with well-aligned bounding boxes might look poor due to the textures and visual details of each element. Elements can also overlap for shadow or visual effects. Inoue et al. [18] approached this as a reconstruction task, using metrics like cosine similarity between the actual and predicted values. This approach works for simple designs with few elements + +
DatasetModelsBounding BoxComposed ImageGenerated Image
CrelloLayoutDM [17]10.2034.43—
VLC (Vector only)0.215.74—
VLC (Dual-domain)0.215.8311.42
PosterLayoutDM [17]1.8120.00—
VLC (Vector only)0.195.82—
VLC (Dual-domain)0.093.756.00
+ +Table 1. FID scores of bounding boxes, composed RGB layouts obtained from the bounding boxes, and the directly generated layouts from the image domain. Our vector-only method outperforms LayoutDM by effectively using image information, with our multi-domain approach yielding the best results. + +
ModelsRICOPubLayNet
FID (↓)mIoU (↑)FID (↓)mIoU (↑)
LayoutVAE [19]33.30.24926.00.316
NDN-none [24]28.40.15861.10.162
LayoutGAN++ [20]6.840.26724.00.263
LayoutTrans [12]5.570.22314.10.272
MaskGIT [5]26.10.26217.20.319
BLT [22]17.40.20272.10.215
BART [25]3.970.2539.360.320
VQDiffusion [11]4.340.25210.30.319
LayoutDM [17]3.550.2777.950.310
VLC (Vector only)2.380.4185.490.348
VLC (Dual-domain)5.600.3765.740.331
Validation Set1.850.6916.250.438
+ +Table 2. Quantitative comparison of our method with the baselines using RICO and PubLayNet datasets. Our method achieves superior results compared with the baselines even with no visual information in input conditions. + +but limits diversity and learning on more complex datasets. Instead, we propose using FID values over three outputs: bounding boxes, feature embeddings from RGB images created by composing elements within generated bounding boxes, and the directly produced RGB images. Although this is our best option, we realize that FID can be biased to image quality and content similarity which should be removed from the consideration for layout evaluation. Finding better visual metric for layout can be a future direction. + +Results: We evaluate our VLC method using visual information as shown in Table 1. Our dual-domain method stands out for its ability to create visually appealing images by using a large image model. This method is a significant improvement over the LayoutDM [17]. LayoutDM uses only element categories as input, which, while effective in creating quality vectors, often fails to align with the actual visual content of each element. This mismatch results in a + +![](images/3246c79d8f14b35c7ba50b906e5c896f023b81af16e56c842e20673f4dca0761.jpg) +VLC (vector only) +VLC (dual domain) +Original Design +Figure 4. Qualitative comparison of our method with baselines on Crello dataset. + +drop in the overall quality of the final image. Our vector domain approach is the most similar one to the transformer-based reconstruction used in FlexDM [18]. It includes both image information and other input conditions, and produces higher-quality images compared to LayoutDM, but has some drawbacks such as occasionally covering faces or not managing color contrast well when elements overlap. Our vector only model and dual-domain model perform similarly on the Crello dataset. The reason can be attributed to the larger number of elements in the dataset, which leads to the content mismatch between image and vector domain such that vector model cannot follow the guidance from image correctly. + +In Table 2, we compare our method with existing baselines on the RICO and PubLayNet datasets. The baseline results are reported by Inoue et al. [17]. Remarkably, even when the input elements are limited to only categorical data without any visual information, our method demonstrated its capability to produce high-quality layouts in both vector-only and dual-domain settings. In the vector-only domain, we observed a trend where training appeared less stable. However, despite this instability, the results often surpassed those generated by discrete diffusion models [11, 17]. In addition, the vector domain model tends to align with the performance of our dual-domain model. This alignment is particularly notable as our dual-domain method achieves better results than most of the baselines, underscoring its effectiveness even in the absence of visual input. It is important to note that while our method is slower than the baselines duo to the amount of computation, this trade-off is compensated by its high-quality output and robustness in processing visual information with its diffusion formulation. Overall, these findings suggest that our method not only achieves state-of-the-art performance but also offers promising po + +![](images/9cb0153445454ef9fc9c631adca109a4394ea12e227d1ed84136c92f074d8616.jpg) +VLC (vector only) +VLC (dual domain) +Original Design +Figure 5. Qualitative comparison of our method with baselines on Poster dataset. + +tential for generating high-quality layouts under varied and challenging conditions. + +# 4.1.2 Qualitative Evaluation + +The effectiveness of our dual-domain method is further illustrated in Figure 4&5 with visual results from the Crello and Poster datasets. These examples highlight our approach not only improves the layout quality but also ensures that the final compositions are more visually coherent and appealing, especially when compared to other methods. + +# 4.2. Model Analysis + +# 4.2.1 The Effectiveness on Dual-domain Modeling + +In addressing the complexities of dual-domain modeling, a crucial challenge lies in maintaining consistency between + +![](images/e6e15a024ca7cc46b3a8b18a49995aa2bbd6a8202fbd73c501770a5bc66746e3.jpg) +Original Layout + +![](images/fff2365dec227bdd301c50d558aca5762bcb55c19ce589f6a86ecdf721182771.jpg) + +![](images/5b2249abc2e85bd50f2b4c80f7b40c8d3f8e0eb70839bedb66538846f8d286b9.jpg) + +![](images/bc1f7543131f62c3f1916abb1e0a1a534d46d571c659e4bb409ec05208358be0.jpg) + +![](images/41bebc02502770036f032cfacfa22c2be764fdd905cbcd75f73253b8881f533c.jpg) + +![](images/a3d04414aa09e316a9c8f6824a1e30224e7d8a9b9f5c053d6919abf89f682466.jpg) +eaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeae +psodmooO O + +![](images/53def6ed455f8a123750e09afd6b73b013e1c198c58c9977a21ac8413d6c04cd.jpg) + +![](images/38f3b88969899bd45c95ca4dce11d68e96c82d48fa3cbf3c5b17449cc9cfe9cb.jpg) +Original Layout + +![](images/5e23fea23232d14330c9301fcf80aab7bec14800340f8ec45b1fea236684cd7b.jpg) +Generated +Image +Composed +Image + +![](images/583b3b8bb4c7515bd37d94927db05df5ac5016f12085dfcb38d3bf190189ef3d.jpg) +Generated +Image +Composed +Image + +![](images/4ba8ef9e6556107adac47c68c0933f45e2ddf93ca2cea8a9af33aae57a1001ed.jpg) +Generated +Image +Composed +Image + +![](images/f9c91c781261b9fab6437b95716f8e561f4ad0ba2a187fe9c1522d9ef47b71b5.jpg) +Figure 6. Our approach is capable of producing a wide range of consistent designs in both image and vector domains. +Figure 7. Attention scores are assigned to various elements in the image domain. + +the image and vector domains. This is vividly illustrated in Figure 3, which demonstrates that when these two domains adhere to a consistent denoising path, the result quality is significantly enhanced. Conversely, deviation from this path leads to a noticeable drop in quality. Complementing this, Figure 6 highlights the efficacy of our method, showcasing that both the image and vector domains can generate results that are not only consistent with each other but also retain the diversity of the generated layouts. + +# 4.2.2 The Effect of Attention Score Maps + +The attention score maps are the key for the visual composing capability of our model. We visualize the clear spatial correlation between attention scores and corresponding elements in the image domain in Figure 7. Our vector + +![](images/f5f3c47a8147c002e6af0ada6747c65bbbc57243d963e870c8b9a3a371bbe4c8.jpg) +Figure 8. Through attention masking, we can decompose all the elements in the latent image space, and include them one by one to create a layered image representation. The bottom right image shows the original design. + +domain model leverages these spatial information to predict vector layouts by adaptively tracking each element and resolving the ambiguity between similar objects. + +By manipulating the attention scores, we can also control the element generation in the image domain. For example, we can remove one element from canvas while keeping the rest layout unchanged by masking out its attention map. If we do this sequentially to each element, we can decompose image output into a layered representation organized by elements, as shown in Figure 8. This results in a virtual vector layout format, which may be used for similar purpose as our vector domain model. + +# 4.3. Applications + +With its two domain outputs and high controllability through attention maps, our model can be used for a range + +![](images/d089ac2e8d8aeba559a197ded7a56ba34bd8802fb254e90c88ccab772dc96195.jpg) +Input Layout Boxes + +![](images/4217ef944e4d5df95d8a1cc006ac107e06822916a387024f0748b853248345e2.jpg) +Generated Layout Images + +![](images/a47382988755cfbb3c3544228bc176e68ce9bac547f6330c63a13b31a3e47b06.jpg) +Figure 9. Layout controlled visual style variation. Given input layout with bounding boxes, our image model can generate visual results with the same layout and different styles. + +![](images/021c57a21423304af37281f54cdd9f0d4a866674c8107b1d9bdf3c8bd2ca861a.jpg) + +![](images/2fc73a3c26712e1a2cf4420abaa3f8222242bf96bb70c1886eab9c5542e98f57.jpg) +Figure 10. Canvas resize with different aspect ratios. Our model can adapt layout under various aspect ratios. + +![](images/ef981f5c6a705963db74de49292c32dcf335c11394113a178c7cb6c8024b5a34.jpg) +Figure 11. Controlled layout generation. Our model allows users to add conditions on design elements to customize the result. In this example, the user specifies the location of the balloons. + +# of design applications. + +In Figure 9, we produce various visual designs with the same spatial layout conforming to given element boxes by setting the attention scores for each element accordingly. This unlocks the power of the pretrained large image model for visual style variation. + +In Figure 10, we generate layouts of different canvas sizes by changing the input condition and masking the intermediate latent states during the denoising process. + +In Figure 11, when a user specifies the position of a cer + +![](images/adb0818d2ac48da261c4fd8d8f4954f51627ad0d8445b29ead68423f8bef0764.jpg) + +![](images/eeb502ff6e0799c93e6d7a8c03074aa54b59ea0518a827af461f7ebf12b7c2b9.jpg) +Original Design + +![](images/58c3547cb4d1512c4dea0f08d5683794a042432cd84b965977df31dc772e2448.jpg) + +![](images/006f2bfa06bb2aa4128048839ceccb8063f3ba411025be83642a36c38fc14564.jpg) +Input Style +Figure 12. Layout style transfer. Our model can transfer the style from the reference design to the input design. + +![](images/371cdc007553cd04044be88736bfce118eadf8aaa8dd2f7f2fcba3aa1afd9fd8.jpg) + +![](images/4c0d5d46a62b5fd4422cdc455f2d432806d5af9880fac4aaf532f87ef870826d.jpg) +New Design + +tain element (the balloons of digits “20”) to a desired position, we fix the attention map for this element during inference, and only optimize the layout for the remaining elements. + +In Figure 12, we transfer the layout from one design image to a new set of elements by propagating the attention maps from an existing design to a new one. The image model takes a given reference image as the initial noisy latent state, accompanied by a new set of elements. During inference, the image model aligns the visual content to input elements while preserving the initial reference layout. The vector model utilizes the aligned attentions to create a new design incorporating both the input elements and reference layout. + +# 5. Conclusion + +Our paper introduces a new approach to layout design, combining vector and image domains through a dual diffusion model. Compared to previous methods solely based on the vector domain, this model has shown notable improvement on layout visual quality while maintaining the flexibility of vector editing on several datasets. Our model also shows flexible controllability via element attention manipulation, enabling a set of design applications across image and vector domains. + +While we have made strides in integrating visual elements into layout design, our model is an initial step in addressing this complex task. Our findings offer a new perspective to the layout generation problem, and point out a potential direction for future exploration. + +# References + +[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 1 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[4] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2 +[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 5 +[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2 +[7] Chin-Yi Cheng, Forrest Huang, Gang Li, and Yang Li. Play: Parametrically conditioned layout generation using latent diffusion. arXiv preprint arXiv:2301.11529, 2023. 2 +[8] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual ACM symposium on user interface software and technology, pages 845-854, 2017. 1, 5 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[10] Vedit Goel, Elia Peruzzo, Yifan Jiang, Dejia Xu, Nicu Sebe, Trevor Darrell, Zhangyang Wang, and Humphrey Shi. Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546, 2023. 3 +[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 5, 6 +[12] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self + +attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004-1014, 2021. 5 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 +[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2 +[15] Hsiao Yuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 2 +[16] Mude Hui, Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie, Yuwang Wang, and Yan Lu. Unifying layout generation with a decoupled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1951, 2023. 2 +[17] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. LayoutDM: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 2, 5, 6 +[18] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Towards flexible multi-modal document models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14287-14296, 2023. 1, 2, 5, 6 +[19] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layoutvae: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 5 +[20] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via latent optimization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 88–96, 2021. 2, 5 +[21] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In ICCV, 2023. 3 +[22] Xiang Kong, Lu Jiang, Huiwen Chang, Han Zhang, Yuan Hao, Haifeng Gong, and Irfan Essa. Blt: bidirectional layout transformer for controllable layout generation. In European Conference on Computer Vision, pages 474-490. Springer, 2022. 5 +[23] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 2 +[24] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 491-506. Springer, 2020. 5 +[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and + +Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 5 +[26] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2 +[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[28] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4 +[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[33] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. arXiv preprint arXiv:2303.00262, 2023. 3 +[34] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. Towards diverse and consistent typography generation. arXiv preprint arXiv:2309.02099, 2023. 2 +[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[36] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Layoutnuwa: Revealing the hidden layout expertise of large language models. arXiv preprint arXiv:2309.09506, 2023. 2 +[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. + +In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2 +[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3 +[39] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2 +[40] Kota Yamaguchi. Canvasvae: Learning to generate vector graphic documents. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5481-5489, 2021. 1, 2, 4 +[41] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2 +[42] Junyi Zhang, Jiaqi Guo, Shizhao Sun, Jian-Guang Lou, and Dongmei Zhang. Layoutdiffusion: Improving graphic layout generation by discrete diffusion probabilistic models. arXiv preprint arXiv:2303.11589, 2023. 2 +[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3 +[44] Xinru Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG), 38(4):1-15, 2019. 2 +[45] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publ aaynet: largest dataset ever for document layout analysis. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1015-1022. IEEE, 2019. 1, 5 +[46] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout gan for visual-textual presentation designs. arXiv preprint arXiv:2205.00303, 2022. 2 \ No newline at end of file diff --git a/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/images.zip b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4995a4f3ff5a9af1b5c5a9453cad1fec9fd24bb1 --- /dev/null +++ b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aa9329f46e6ebf402e095f78a5c0629653e944a09324b13dc7cffef030abff3 +size 722679 diff --git a/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/layout.json b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..32cdbdad97931b3d55ec16abe9da546fab23dd88 --- /dev/null +++ b/2024/Visual Layout Composer_ Image-Vector Dual Diffusion Model for Design Layout Generation/layout.json @@ -0,0 +1,8578 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 489, + 140 + ], + "type": "text", + "content": "Visual LayoutComposer:Image-Vector Dual Diffusion Model for DesignLayout Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": "Mohammad Amin Shabani" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": " Zhaowen Wang" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": " Difan Liu" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": " Nanxuan Zhao" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": " Jimei Yang" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": " Yasutaka Furukawa" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": " Simon Fraser University " + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 160, + 492, + 203 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 215, + 358, + 370 + ], + "blocks": [ + { + "bbox": [ + 53, + 215, + 358, + 370 + ], + "lines": [ + { + "bbox": [ + 53, + 215, + 358, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 358, + 370 + ], + "type": "image", + "image_path": "098199d519b41b62d9ea2ba15c3a8aef1a6940e4d27606ef8e18900731fad987.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 380, + 546, + 413 + ], + "lines": [ + { + "bbox": [ + 45, + 380, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 45, + 380, + 546, + 413 + ], + "type": "text", + "content": "Figure 1. Overview of our method. Left: our image-vector dual diffusion model architecture where both image and vector diffusion are conditioned on input elements and exchanging information with attention maps. Right: a comparison showing the advantage of dual-domain (image and vector) over traditional single-domain approaches." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 384, + 218, + 547, + 369 + ], + "blocks": [ + { + "bbox": [ + 384, + 218, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 384, + 218, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 384, + 218, + 547, + 369 + ], + "type": "image", + "image_path": "a5f57ae3ca19cbe08b08aee90ce5581e26ce1c5ab73eeb1adb4a7d2cb4931df4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 423, + 192, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 423, + 192, + 435 + ], + "spans": [ + { + "bbox": [ + 143, + 423, + 192, + 435 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 290, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 290, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 290, + 628 + ], + "type": "text", + "content": "This paper proposes an image-vector dual diffusion model for generative layout design. Distinct from prior efforts that mostly ignores visual information of elements and the whole canvas, our approach integrates the power of a pre-trained large image diffusion model to guide layout composition in a vector diffusion model by providing enhanced salient region understanding and high-level inter-element relationship reasoning. Our proposed model simultaneously operates in two domains: it generates the overall design appearance in the image domain while optimizing the size and position of each design element in the vector domain. The proposed method achieves the state-of-the-art results on several datasets and enables new layout design applications. Project webpage: https://aminshabani.github.io/visual.layout_composer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 638, + 128, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 128, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 128, + 650 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "type": "text", + "content": "Layout design is essential in creative tasks, determining the spatial arrangement of texts, images, and other visual elements, which are composed into a visual design, such as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 424, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 545, + 497 + ], + "type": "text", + "content": "book and magazine covers, posters showcasing new products, or invitation cards for holiday parties. Graphic designers rely on their keen artistic sensibility in meticulously optimizing the arrangements of the visual elements while ensuring that the composited image conveys the intended message with aesthetic appeals." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": "Machine learning techniques have been successful in directly regressing the layout vector parameters, i.e., the bounding box coordinates of elements. Both transformer-based models [1] and more recent diffusion-based models [18] have achieved promising results that capture basic geometrical properties such as alignment and symmetry. However, most of these models ignore the visual information from the input elements, including image content, text typeface, and color/shape of vector arts. This leads to unsatisfactory designs as illustrated by the Vector Only results in Figure 1. The lack of layout datasets with associated images also contributes to the negligence of visual information. The most commonly used RICO [8] and Pub-LayNet [45] datasets only contain box representation of design elements. The Crello [40] dataset provides element-level image rendering for over 23K design templates, whose visual features are used in the FlexDM [18] model. However, the interplay of element appearance when rendered on" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 272, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 272, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 272, + 713 + ], + "type": "text", + "content": "*This work was done during the first author's internship at Adobe." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9222" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "content": "the same canvas has never been studied at any large scale." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 266 + ], + "type": "text", + "content": "The recent success of image generative models [6, 9, 30, 32, 41] has enabled the generation of layout design in the image space. Trained on millions of high quality design images from web, these large generative models produce visually diverse and compelling layouts as raster images, demonstrating sophisticated design principles such as the placement of major subjects, style coherence, and perceptual contrast. See the Image Only results as examples in Figure 1. These visual properties, vital to layout design, are learned in a holistic way, not straightforward by previous vector domain approaches. Despite high visual quality, the layout designs in image format forbid further editing and customization. Moreover, current generative models are controlled only by text prompts, and cannot use given visual assets as required in many real-world applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 288, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 470 + ], + "type": "text", + "content": "Motivated by the complementary strengths of the vector and image domains, we propose a combined approach that generates layout designs in both image and vector domains simultaneously. An image branch is a pre-trained text-to-image stable diffusion model [30], which refines (i.e., noises) a design as an image, given a variable number of design elements with different media types. A vector branch refines per-element bounding box parameters conditioned on attention maps from the UNet of the image branch. Ground truth element masks are used to apply loss directly to these attention maps, whose manipulation at generation time will allow users to control the positioning of elements. This dual diffusion model refines the visual design in the vector and image domains simultaneously. The final vector output assembles the input elements and renders the visual design while allowing further editing by users. The overall proposed method is illustrated on the left of Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "content": "For training large diffusion models, we have collected a large-scale design Poster dataset from public design web pages by extracting both images and metadata. This effort yields 140K image layout paired samples as our own dataset, more than six times of the current largest visual layout dataset Crello with 23K samples. We conduct experiments on both datasets to validate the effectiveness of our dual diffusion model. Our data collection tools and processing pipeline can facilitate future efforts to curate even larger datasets for exploring visual design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 593, + 244, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 593, + 244, + 604 + ], + "spans": [ + { + "bbox": [ + 58, + 593, + 244, + 604 + ], + "type": "text", + "content": "Our key contributions are summarized below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 286, + 711 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 605, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 286, + 664 + ], + "type": "text", + "content": "- A novel dual diffusion model that generates layout design in both image and vector domains conditioned on input design elements. The two diffusion processes interact by exchanging element attention maps, where the final results attain both vector editability and visual quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 666, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 286, + 688 + ], + "type": "text", + "content": "- Creative design applications, enabling element-wise user controls, such as canvas resizing and design variation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 286, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 711 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 711 + ], + "type": "text", + "content": "- State-of-the-art layout generation results on the existing datasets and our Poster dataset." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 91, + 474, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 474, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 474, + 104 + ], + "type": "text", + "content": "2.1. Vector-only Layout Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "content": "Generative models for layout generation have seen significant developments. Yamaguchi [40] introduced a VAE-based architecture for unconditional generation of vector graphic documents. Following this, Li et al. [26] proposed LayoutGAN, which uses a relational generator and a wireframe renderer for training with a pixel-based discriminator. Kikuchi et al. [20] further refined this approach with LayoutGAN++, incorporating user-specified constraints into layout generation. Chai et al. [4] proposed using DDPM for layout generation. Inoue et al. proposed LayoutDM [17], a discrete diffusion-based model for layout designs. Zhang et al. [42] and Hui et al. [16] both contributed to this domain with LayoutDiffusion and LDGM, respectively, each based on discrete diffusion models. Cheng et al. [7] proposed a unique approach using guidelines as input conditions for the latent diffusion model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 309, + 485, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 485, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 485, + 322 + ], + "type": "text", + "content": "2.2. Visual-guided Layout Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 327, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 530 + ], + "type": "text", + "content": "Zhou et al. [46] tackled image-composition-aware layout generation with a multi-stage GAN-based method, an approach further improved by Xu et al. [39] through the use of a pixel-level discriminator. Zheng et al. [44] proposed a GAN based approach using image-based representation for layouts, which requires subsequent post-processing. Lin et al. [28] introduced a multi-step approach dedicated to the cleansing and retargeting of advertising posters. Hsu et al. [15] introduced DS-GAN, a CNN-LSTM-based conditional GAN method for poster layout generation. Inoue et al. [18] proposed Flex-DM, utilizing a transformer to predict masked attributes in design. Shimoda et al. [34] and Tang et al. [36] expanded the field by proposing new models for typography generation and layout generation, respectively. Our work differs by focusing on enhancing visual information and consistency through a large image diffusion model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 539, + 471, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 471, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 471, + 552 + ], + "type": "text", + "content": "2.3. Image-based Diffusion Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Image diffusion model [35] was proposed initially for unconditional image generation. With the rapid development of the large-language model (LLM) [3], it has shown groundbreaking ability for text-to-image generation tasks [14]. To enhance the computational efficiency, rather than performing diffusion steps at the pixel level, latent diffusion model (LDM) [30] conducted operations on latent space, further boosting the generation ability of diffusion models. It has been widely used in various applications, including image-to-image translation [31], style transfer [23, 37], and condition-guided editing (e.g., sketch, layout, depth map, etc.) [27, 43]. To enable multi-object composition closer to our work, Xiao et al. [38] introduced" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9223" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "Localization Loss for free text-to-image generation with multiple subjects, while Sarukkai et al. [33] developed Collage Diffusion for creating realistic photos, focusing on harmonization and fidelity. Balaji et al. [2] training an ensemble of diffusion models specialized for different parts of the diffusion model to improve the overall text alignment of the model. Goel et al. [10] proposed Pair Diffusion to edit object properties in images via structure and appearance decomposition. Our model stands out by not needing predefined element positions and scales, handling multiple subjects, maintaining visual consistency, and providing enhanced editing flexibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 225, + 104, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 104, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 104, + 238 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 245, + 289, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 289, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 289, + 365 + ], + "type": "text", + "content": "Our Visual LayoutComposer (VLC) model consists of two diffusion models that operate in parallel: one in the vector domain and the other in the image domain. The two models exchange features at intermediate stages, enabling them to produce outputs that are not only consistent but also leverage the information from both domains. In the following subsections, we first describe our problem formulation, followed by the specifics of the diffusion models for each domain, and how they collaborate to enhance output consistency. Figure 2 shows the model overview." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 371, + 127, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 127, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 127, + 384 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": "Our input is a set of " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": " design elements " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "E = \\{e_1, e_2, \\ldots, e_N\\}" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": ". Each element " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": " is associated with a class category " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": " indicating its asset type (such as text, image, shape), an index " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": " indicating its layer order relative to other elements, and an RGBA image " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": " of its rendering. On a designated 2D canvas with aspect ratio " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "r_c" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": ", the task is to predict an upright bounding box for each element " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": ", defined by the 2D coordinates of the top left and bottom right corners as a 4D vector " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": ". The goal is to generate a layout design with high visual quality by composing each element " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": " on the canvas with the position and size specified by " + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 389, + 289, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 527, + 174, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 174, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 174, + 540 + ], + "type": "text", + "content": "3.2. Condition Embedding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "After resizing each element image " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " of aspect ratio " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", we employ CLIP [29] followed by a linear layer to derive its feature representation in the model's hidden dimension " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". In parallel, we encode the vector attributes of each element by a vector attributes encoder. The vector condition is constructed by concatenating the element image ratio " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", canvas image ratio " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "r_c" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", one-hot encoded version of the element's type " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " and the order of the element " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". A linear layer maps the concatenated feature to the model's dimension " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". The image feature and the vector feature are concatenated and processed by a linear layer to form the final condition feature of size " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". For each image or vector diffusion model, this condition feature is fed into a transformer-based condition processing module (Figure 2)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "yielding outputs of sizes " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "N \\times 768" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " for image condition feature and " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " for vector condition feature." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 106, + 432, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 106, + 432, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 106, + 432, + 118 + ], + "type": "text", + "content": "3.3. Network Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "text", + "content": "The diffusion model in the image domain utilizes a pretrained Stable Diffusion V1.5 model [30], retaining all layers frozen except for the attention modules. During a given time step " + }, + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "text", + "content": ", a noisy latent image " + }, + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "inline_equation", + "content": "z^t" + }, + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "text", + "content": ", representing the target layout rendering, will be passed to the U-Net. Contrary to the original Stable Diffusion, which uses text prompts, we pass the processed image condition feature as the input to the cross-attention layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "content": "The diffusion model in the vector domain produces the bounding box positions as a 4D vector " + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "content": ". We embed the noisy input coordinates " + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "content": " feature vector, concatenate with the vector condition feature, and map to the model's dimension " + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "content": " by a linear layer. These features are fed into a transformer-based diffusion denoiser to estimate the corresponding noise for each element " + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 221, + 545, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 305, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 461 + ], + "type": "text", + "content": "Information exchange between the image and vector domains is challenging, where we found that direct feature space sharing as in ControlNet [43] is not effective. We use the processed attention scores from U-Net as a shared medium. The advantages of this method are: (i) it aligns the number of attention scores with the number of input elements, enabling a straightforward concatenation of each image mask to its respective vector model feature token; (ii) it provides a general indication of each element's location in the image domain, simplifying the interpretation for the vector domain; and (iii) it allows the vector domain to adjust these scores to enhance the U-Net's focusing capability, with a mechanism similar to Attention Modulation [21]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": "Concretely, we extract a set of attention scores from the cross-attention module of the U-Net's middle block. Given the scores with 8 heads and size " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": " for each design element, we flatten the tensors to form a set of vectors with the total size of " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "N \\times 512" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": ". This feature is embedded to " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "N \\times 1024" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": " and concatenated with each input element in the vector domain prior to the transformer-based diffusion denoiser. The transformer model produces an output of " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "N \\times 2048" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": ", which is passed through linear projection layers to yield denoised estimation of " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "x^{t - 1}" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": " and the attention modulation residual of size " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "N \\times 512" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": ". We add the residual values to the original attentions to form the refined attention scores. These scores are sent back to the U-Net, promoting interactive and iterative information exchange between the two domains." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "During the diffusion process, repetitive elements can lead to omissions or misalignments in the image. To mitigate this, we adopt the cross-attention localization loss [38]. This loss refines the attention scores of each region, ensuring a more accurate correspondence to its respective element. We apply this loss within the U-Net's attention modules to further enhance the quality of the attention scores." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9224" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 76, + 541, + 254 + ], + "blocks": [ + { + "bbox": [ + 52, + 76, + 541, + 254 + ], + "lines": [ + { + "bbox": [ + 52, + 76, + 541, + 254 + ], + "spans": [ + { + "bbox": [ + 52, + 76, + 541, + 254 + ], + "type": "image", + "image_path": "fce675491ca02073f9bf1d1306b4b0f0f1d8b8883a2fc949cd90e621367618f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed model. Given a set of element images and additional attributes, the model simultaneously employs dual diffusion processes to generate a cohesive image layout and the corresponding bounding boxes for each element." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 304, + 177, + 392 + ], + "blocks": [ + { + "bbox": [ + 51, + 304, + 177, + 392 + ], + "lines": [ + { + "bbox": [ + 51, + 304, + 177, + 392 + ], + "spans": [ + { + "bbox": [ + 51, + 304, + 177, + 392 + ], + "type": "image", + "image_path": "d1f6d9291135a11c36624b2838f4defc26fa548a45c43a917ed17015f80f682c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 288, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 448 + ], + "type": "text", + "content": "Figure 3. Generated image/vector layout without or with canvas masking. By masking intermediate outputs of the diffusion model in image domain, we optimize the image output for the canvas size and increase the consistency between the vector-image domains." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 304, + 282, + 392 + ], + "blocks": [ + { + "bbox": [ + 179, + 304, + 282, + 392 + ], + "lines": [ + { + "bbox": [ + 179, + 304, + 282, + 392 + ], + "spans": [ + { + "bbox": [ + 179, + 304, + 282, + 392 + ], + "type": "image", + "image_path": "2ddfc208f079e068970ec4e8e738ea1df4c2079494fafb9197b38ae9991b30b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 498 + ], + "type": "text", + "content": "The training losses for the image and vector diffusion models are the same as the original DDPM [13]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 513, + 234, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 234, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 234, + 526 + ], + "type": "text", + "content": "3.4. Controllable Element Composition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Our experiments showed that fine-tuning the image domain led the model to disregard the specified canvas aspect ratio, favoring the generation of square-shaped layout designs. This tendency not only reduces the quality and diversity of the generated designs but also increases the discrepancy between the image and vector domains, particularly given that the vector domain generally follows the canvas's aspect ratio. To address this problem, we introduce a straightforward yet efficient strategy: masking the noisy latent feature map of " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "z^t" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " at each time step to have zero value outside the canvas region, which ensures image contents related to input condition only appear inside the canvas. The homogeneous region outside the canvas is cropped from the final generated image, maintaining the original image aspect ratio. Figure 3 shows the effect of this process on both domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 298, + 388, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 388, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 388, + 312 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "type": "text", + "content": "Implementation details: In our implementation, we use the pretrained weights of StableDiffusion v1-5 [30] for image domain, and OpenAI's clip-vit-large-patch14 vision model for embedding the image of each element to latent space. For training, we run each model for 250,000 iterations, employing a starting learning rate of 5e-5 which decreases by a factor of 0.5 every 100k steps. The vector model has a hidden dimension of 2048, accommodating a wide range of features including attention scores, vector inputs, and image information. Additionally, we incorporate classifier-free guidance into our process, randomly dropping the image inputs of the elements in " + }, + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "type": "text", + "content": " of the times. Training is conducted on 8 NVIDIA A100 GPUs, with a total batch size of 128. We pad each layout to square shape and normalize the coordinates to " + }, + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 304, + 323, + 547, + 539 + ], + "type": "text", + "content": ". For both domains, we employ the DDPM scheduler with 1000 steps for training and 50 steps for inference to ensure high-quality layout generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 544, + 527, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 527, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 527, + 555 + ], + "type": "text", + "content": "Datasets: Our experiments use two groups of datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 547, + 713 + ], + "type": "text", + "content": "- Poster and Crello [40] datasets: demonstrating our method's ability to capture intricate visual information and the interplay between design elements. These datasets, characterized by their rich graphic content, are instrumental in showcasing the method's effectiveness in a visually complex environment. We gathered Poster dataset, which consists of 137,781 layout designs. From this, we reserved 1,000 layouts for validation and 3,000 for testing. Each layout is represented as an RGBA image and is classified into one of five categories: Frame, Canvas, Shape, Mix, and Text. We have limited the number of elements in any given layout to a maximum of 20 to maintain structural simplicity. Given the intricate nature of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9225" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 131 + ], + "type": "text", + "content": "these layouts, we have also provided the sequential order of the elements within each design. The Crello datasets include 18,768 training, 2,316 validation, and 2,331 test layouts including categories such as vector shape, image, or text placeholder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 360 + ], + "type": "text", + "content": "- RICO [8] and PubLayNet [45] datasets: having limited visual information, which challenges our method. To compensate for this, we assign distinct colors to each category and depict layouts with these colors, simplifying the visuals to fixed-color blocks representing each element. This approach tests our method's ability in generating layouts based primarily on structure rather than rich visual details. The Rico dataset consists of user interface (UI) designs sourced from various Android applications, spanning up to 25 categories of UI elements including text bars, icons, and buttons. PubLayNet includes 360,000 document layout instances from academic papers, categorized into five primary layout elements: text, titles, lists, tables, and figures. We use the processed datasets by [17], obtained by keeping only those layouts containing a maximum of 25 elements each. This results training/validation/test sets of sizes 35,851/2,109/4,218 respectively for Rico and 315,757/16,619/11,142 for PubLayNet." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 368, + 132, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 132, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 132, + 380 + ], + "type": "text", + "content": "4.1. Comparisons" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 286, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 286, + 518 + ], + "type": "text", + "content": "For the baselines, we compare our method with recent state-of-the-art methods in Layout Design Generation [5, 11, 12, 17, 19, 20, 22, 24, 25]. We compare our method to the LayoutDM [17] as the state-of-the-art for layout generation. This model only uses element categories, adhering to their original design in discrete diffusion space. Our vector domain model operates in a continuous diffusion space and incorporates each element's image embedding. Finally, our dual-domain model combines element image embeddings with the capabilities of the image diffusion model to get the highest quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 535, + 182, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 182, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 182, + 547 + ], + "type": "text", + "content": "4.1.1 Quantitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "type": "text", + "content": "Metrics: We follow the previous work [17] and use metrics including Maximum IoU [20] and FID over extracted features of the layouts [20, 24] for RICO and PubLayNet datasets. The previous methods and measurements are effective when there are no visual data or input images for layout elements. However, assessing layout designs in Poster and Crello datasets is more complex. Even layouts with well-aligned bounding boxes might look poor due to the textures and visual details of each element. Elements can also overlap for shadow or visual effects. Inoue et al. [18] approached this as a reconstruction task, using metrics like cosine similarity between the actual and predicted values. This approach works for simple designs with few elements" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 541, + 176 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 541, + 176 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 541, + 176 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 541, + 176 + ], + "type": "table", + "html": "
DatasetModelsBounding BoxComposed ImageGenerated Image
CrelloLayoutDM [17]10.2034.43—
VLC (Vector only)0.215.74—
VLC (Dual-domain)0.215.8311.42
PosterLayoutDM [17]1.8120.00—
VLC (Vector only)0.195.82—
VLC (Dual-domain)0.093.756.00
", + "image_path": "46f0378c645f1cc0822259bb960d87fd1350a3f8ea64be7cd74d9c392e675cd8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 251, + 545, + 433 + ], + "blocks": [ + { + "bbox": [ + 304, + 184, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 240 + ], + "type": "text", + "content": "Table 1. FID scores of bounding boxes, composed RGB layouts obtained from the bounding boxes, and the directly generated layouts from the image domain. Our vector-only method outperforms LayoutDM by effectively using image information, with our multi-domain approach yielding the best results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 251, + 545, + 433 + ], + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 433 + ], + "type": "table", + "html": "
ModelsRICOPubLayNet
FID (↓)mIoU (↑)FID (↓)mIoU (↑)
LayoutVAE [19]33.30.24926.00.316
NDN-none [24]28.40.15861.10.162
LayoutGAN++ [20]6.840.26724.00.263
LayoutTrans [12]5.570.22314.10.272
MaskGIT [5]26.10.26217.20.319
BLT [22]17.40.20272.10.215
BART [25]3.970.2539.360.320
VQDiffusion [11]4.340.25210.30.319
LayoutDM [17]3.550.2777.950.310
VLC (Vector only)2.380.4185.490.348
VLC (Dual-domain)5.600.3765.740.331
Validation Set1.850.6916.250.438
", + "image_path": "03af75f581e7708cf63da6dcd0cdc8a49fbaa50989329a3d07ca645ec32c34b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 441, + 545, + 485 + ], + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 485 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of our method with the baselines using RICO and PubLayNet datasets. Our method achieves superior results compared with the baselines even with no visual information in input conditions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 506, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 615 + ], + "type": "text", + "content": "but limits diversity and learning on more complex datasets. Instead, we propose using FID values over three outputs: bounding boxes, feature embeddings from RGB images created by composing elements within generated bounding boxes, and the directly produced RGB images. Although this is our best option, we realize that FID can be biased to image quality and content similarity which should be removed from the consideration for layout evaluation. Finding better visual metric for layout can be a future direction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Results: We evaluate our VLC method using visual information as shown in Table 1. Our dual-domain method stands out for its ability to create visually appealing images by using a large image model. This method is a significant improvement over the LayoutDM [17]. LayoutDM uses only element categories as input, which, while effective in creating quality vectors, often fails to align with the actual visual content of each element. This mismatch results in a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9226" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 85, + 269, + 259 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 130, + 83 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 130, + 83 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 130, + 83 + ], + "type": "text", + "content": "VLC (vector only)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 74, + 198, + 83 + ], + "lines": [ + { + "bbox": [ + 137, + 74, + 198, + 83 + ], + "spans": [ + { + "bbox": [ + 137, + 74, + 198, + 83 + ], + "type": "text", + "content": "VLC (dual domain)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 74, + 258, + 83 + ], + "lines": [ + { + "bbox": [ + 209, + 74, + 258, + 83 + ], + "spans": [ + { + "bbox": [ + 209, + 74, + 258, + 83 + ], + "type": "text", + "content": "Original Design" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 85, + 269, + 259 + ], + "lines": [ + { + "bbox": [ + 66, + 85, + 269, + 259 + ], + "spans": [ + { + "bbox": [ + 66, + 85, + 269, + 259 + ], + "type": "image", + "image_path": "3246c79d8f14b35c7ba50b906e5c896f023b81af16e56c842e20673f4dca0761.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 271, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 271, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 287, + 293 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of our method with baselines on Crello dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 460 + ], + "type": "text", + "content": "drop in the overall quality of the final image. Our vector domain approach is the most similar one to the transformer-based reconstruction used in FlexDM [18]. It includes both image information and other input conditions, and produces higher-quality images compared to LayoutDM, but has some drawbacks such as occasionally covering faces or not managing color contrast well when elements overlap. Our vector only model and dual-domain model perform similarly on the Crello dataset. The reason can be attributed to the larger number of elements in the dataset, which leads to the content mismatch between image and vector domain such that vector model cannot follow the guidance from image correctly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "In Table 2, we compare our method with existing baselines on the RICO and PubLayNet datasets. The baseline results are reported by Inoue et al. [17]. Remarkably, even when the input elements are limited to only categorical data without any visual information, our method demonstrated its capability to produce high-quality layouts in both vector-only and dual-domain settings. In the vector-only domain, we observed a trend where training appeared less stable. However, despite this instability, the results often surpassed those generated by discrete diffusion models [11, 17]. In addition, the vector domain model tends to align with the performance of our dual-domain model. This alignment is particularly notable as our dual-domain method achieves better results than most of the baselines, underscoring its effectiveness even in the absence of visual input. It is important to note that while our method is slower than the baselines duo to the amount of computation, this trade-off is compensated by its high-quality output and robustness in processing visual information with its diffusion formulation. Overall, these findings suggest that our method not only achieves state-of-the-art performance but also offers promising po" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 88, + 536, + 466 + ], + "blocks": [ + { + "bbox": [ + 322, + 76, + 386, + 87 + ], + "lines": [ + { + "bbox": [ + 322, + 76, + 386, + 87 + ], + "spans": [ + { + "bbox": [ + 322, + 76, + 386, + 87 + ], + "type": "text", + "content": "VLC (vector only)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 76, + 460, + 87 + ], + "lines": [ + { + "bbox": [ + 391, + 76, + 460, + 87 + ], + "spans": [ + { + "bbox": [ + 391, + 76, + 460, + 87 + ], + "type": "text", + "content": "VLC (dual domain)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 76, + 526, + 87 + ], + "lines": [ + { + "bbox": [ + 470, + 76, + 526, + 87 + ], + "spans": [ + { + "bbox": [ + 470, + 76, + 526, + 87 + ], + "type": "text", + "content": "Original Design" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 88, + 536, + 466 + ], + "lines": [ + { + "bbox": [ + 317, + 88, + 536, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 88, + 536, + 466 + ], + "type": "image", + "image_path": "9cb0153445454ef9fc9c631adca109a4394ea12e227d1ed84136c92f074d8616.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 482, + 545, + 503 + ], + "lines": [ + { + "bbox": [ + 306, + 482, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 545, + 503 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of our method with baselines on Poster dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 517, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 545, + 541 + ], + "type": "text", + "content": "tential for generating high-quality layouts under varied and challenging conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 555, + 435, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 435, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 435, + 567 + ], + "type": "text", + "content": "4.1.2 Qualitative Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 573, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 573, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 305, + 573, + 545, + 645 + ], + "type": "text", + "content": "The effectiveness of our dual-domain method is further illustrated in Figure 4&5 with visual results from the Crello and Poster datasets. These examples highlight our approach not only improves the layout quality but also ensures that the final compositions are more visually coherent and appealing, especially when compared to other methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 652, + 400, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 652, + 400, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 652, + 400, + 664 + ], + "type": "text", + "content": "4.2. Model Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 528, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 528, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 528, + 682 + ], + "type": "text", + "content": "4.2.1 The Effectiveness on Dual-domain Modeling" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In addressing the complexities of dual-domain modeling, a crucial challenge lies in maintaining consistency between" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9227" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 166, + 135 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 166, + 135 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 166, + 135 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 166, + 135 + ], + "type": "image", + "image_path": "e6e15a024ca7cc46b3a8b18a49995aa2bbd6a8202fbd73c501770a5bc66746e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 137, + 137, + 148 + ], + "lines": [ + { + "bbox": [ + 78, + 137, + 137, + 148 + ], + "spans": [ + { + "bbox": [ + 78, + 137, + 137, + 148 + ], + "type": "text", + "content": "Original Layout" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 73, + 283, + 118 + ], + "blocks": [ + { + "bbox": [ + 167, + 73, + 283, + 118 + ], + "lines": [ + { + "bbox": [ + 167, + 73, + 283, + 118 + ], + "spans": [ + { + "bbox": [ + 167, + 73, + 283, + 118 + ], + "type": "image", + "image_path": "fff2365dec227bdd301c50d558aca5762bcb55c19ce589f6a86ecdf721182771.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 118, + 283, + 163 + ], + "blocks": [ + { + "bbox": [ + 167, + 118, + 283, + 163 + ], + "lines": [ + { + "bbox": [ + 167, + 118, + 283, + 163 + ], + "spans": [ + { + "bbox": [ + 167, + 118, + 283, + 163 + ], + "type": "image", + "image_path": "5b2249abc2e85bd50f2b4c80f7b40c8d3f8e0eb70839bedb66538846f8d286b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 284, + 73, + 400, + 118 + ], + "blocks": [ + { + "bbox": [ + 284, + 73, + 400, + 118 + ], + "lines": [ + { + "bbox": [ + 284, + 73, + 400, + 118 + ], + "spans": [ + { + "bbox": [ + 284, + 73, + 400, + 118 + ], + "type": "image", + "image_path": "bc1f7543131f62c3f1916abb1e0a1a534d46d571c659e4bb409ec05208358be0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 284, + 118, + 400, + 163 + ], + "blocks": [ + { + "bbox": [ + 284, + 118, + 400, + 163 + ], + "lines": [ + { + "bbox": [ + 284, + 118, + 400, + 163 + ], + "spans": [ + { + "bbox": [ + 284, + 118, + 400, + 163 + ], + "type": "image", + "image_path": "41bebc02502770036f032cfacfa22c2be764fdd905cbcd75f73253b8881f533c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 400, + 73, + 517, + 118 + ], + "blocks": [ + { + "bbox": [ + 400, + 73, + 517, + 118 + ], + "lines": [ + { + "bbox": [ + 400, + 73, + 517, + 118 + ], + "spans": [ + { + "bbox": [ + 400, + 73, + 517, + 118 + ], + "type": "image", + "image_path": "a3d04414aa09e316a9c8f6824a1e30224e7d8a9b9f5c053d6919abf89f682466.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 83, + 528, + 154 + ], + "lines": [ + { + "bbox": [ + 518, + 83, + 528, + 154 + ], + "spans": [ + { + "bbox": [ + 518, + 83, + 528, + 154 + ], + "type": "text", + "content": "eaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeaeae" + } + ] + } + ], + "index": 8, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 529, + 76, + 539, + 161 + ], + "lines": [ + { + "bbox": [ + 529, + 76, + 539, + 161 + ], + "spans": [ + { + "bbox": [ + 529, + 76, + 539, + 161 + ], + "type": "text", + "content": "psodmooO O" + } + ] + } + ], + "index": 9, + "angle": 270, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 400, + 118, + 517, + 163 + ], + "blocks": [ + { + "bbox": [ + 400, + 118, + 517, + 163 + ], + "lines": [ + { + "bbox": [ + 400, + 118, + 517, + 163 + ], + "spans": [ + { + "bbox": [ + 400, + 118, + 517, + 163 + ], + "type": "image", + "image_path": "53def6ed455f8a123750e09afd6b73b013e1c198c58c9977a21ac8413d6c04cd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 77, + 167, + 134, + 243 + ], + "blocks": [ + { + "bbox": [ + 77, + 167, + 134, + 243 + ], + "lines": [ + { + "bbox": [ + 77, + 167, + 134, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 167, + 134, + 243 + ], + "type": "image", + "image_path": "38f3b88969899bd45c95ca4dce11d68e96c82d48fa3cbf3c5b17449cc9cfe9cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 247, + 135, + 257 + ], + "lines": [ + { + "bbox": [ + 76, + 247, + 135, + 257 + ], + "spans": [ + { + "bbox": [ + 76, + 247, + 135, + 257 + ], + "type": "text", + "content": "Original Layout" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 168, + 166, + 283, + 245 + ], + "blocks": [ + { + "bbox": [ + 168, + 166, + 283, + 245 + ], + "lines": [ + { + "bbox": [ + 168, + 166, + 283, + 245 + ], + "spans": [ + { + "bbox": [ + 168, + 166, + 283, + 245 + ], + "type": "image", + "image_path": "5e23fea23232d14330c9301fcf80aab7bec14800340f8ec45b1fea236684cd7b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 247, + 212, + 255 + ], + "lines": [ + { + "bbox": [ + 175, + 247, + 212, + 255 + ], + "spans": [ + { + "bbox": [ + 175, + 247, + 212, + 255 + ], + "type": "text", + "content": "Generated" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 258, + 206, + 267 + ], + "lines": [ + { + "bbox": [ + 182, + 258, + 206, + 267 + ], + "spans": [ + { + "bbox": [ + 182, + 258, + 206, + 267 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 229, + 247, + 270, + 256 + ], + "lines": [ + { + "bbox": [ + 229, + 247, + 270, + 256 + ], + "spans": [ + { + "bbox": [ + 229, + 247, + 270, + 256 + ], + "type": "text", + "content": "Composed" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 258, + 262, + 267 + ], + "lines": [ + { + "bbox": [ + 238, + 258, + 262, + 267 + ], + "spans": [ + { + "bbox": [ + 238, + 258, + 262, + 267 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 284, + 165, + 399, + 245 + ], + "blocks": [ + { + "bbox": [ + 284, + 165, + 399, + 245 + ], + "lines": [ + { + "bbox": [ + 284, + 165, + 399, + 245 + ], + "spans": [ + { + "bbox": [ + 284, + 165, + 399, + 245 + ], + "type": "image", + "image_path": "583b3b8bb4c7515bd37d94927db05df5ac5016f12085dfcb38d3bf190189ef3d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 247, + 332, + 255 + ], + "lines": [ + { + "bbox": [ + 294, + 247, + 332, + 255 + ], + "spans": [ + { + "bbox": [ + 294, + 247, + 332, + 255 + ], + "type": "text", + "content": "Generated" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 258, + 325, + 268 + ], + "lines": [ + { + "bbox": [ + 301, + 258, + 325, + 268 + ], + "spans": [ + { + "bbox": [ + 301, + 258, + 325, + 268 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 247, + 390, + 256 + ], + "lines": [ + { + "bbox": [ + 350, + 247, + 390, + 256 + ], + "spans": [ + { + "bbox": [ + 350, + 247, + 390, + 256 + ], + "type": "text", + "content": "Composed" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 258, + 381, + 268 + ], + "lines": [ + { + "bbox": [ + 358, + 258, + 381, + 268 + ], + "spans": [ + { + "bbox": [ + 358, + 258, + 381, + 268 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 402, + 166, + 515, + 245 + ], + "blocks": [ + { + "bbox": [ + 402, + 166, + 515, + 245 + ], + "lines": [ + { + "bbox": [ + 402, + 166, + 515, + 245 + ], + "spans": [ + { + "bbox": [ + 402, + 166, + 515, + 245 + ], + "type": "image", + "image_path": "4ba8ef9e6556107adac47c68c0933f45e2ddf93ca2cea8a9af33aae57a1001ed.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 247, + 447, + 255 + ], + "lines": [ + { + "bbox": [ + 408, + 247, + 447, + 255 + ], + "spans": [ + { + "bbox": [ + 408, + 247, + 447, + 255 + ], + "type": "text", + "content": "Generated" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 258, + 440, + 268 + ], + "lines": [ + { + "bbox": [ + 416, + 258, + 440, + 268 + ], + "spans": [ + { + "bbox": [ + 416, + 258, + 440, + 268 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 247, + 504, + 256 + ], + "lines": [ + { + "bbox": [ + 465, + 247, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 465, + 247, + 504, + 256 + ], + "type": "text", + "content": "Composed" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 258, + 496, + 267 + ], + "lines": [ + { + "bbox": [ + 473, + 258, + 496, + 267 + ], + "spans": [ + { + "bbox": [ + 473, + 258, + 496, + 267 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 54, + 308, + 283, + 467 + ], + "blocks": [ + { + "bbox": [ + 83, + 282, + 508, + 294 + ], + "lines": [ + { + "bbox": [ + 83, + 282, + 508, + 294 + ], + "spans": [ + { + "bbox": [ + 83, + 282, + 508, + 294 + ], + "type": "text", + "content": "Figure 6. Our approach is capable of producing a wide range of consistent designs in both image and vector domains." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 308, + 283, + 467 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 283, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 283, + 467 + ], + "type": "image", + "image_path": "f9c91c781261b9fab6437b95716f8e561f4ad0ba2a187fe9c1522d9ef47b71b5.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 479, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 502 + ], + "type": "text", + "content": "Figure 7. Attention scores are assigned to various elements in the image domain." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 523, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 632 + ], + "type": "text", + "content": "the image and vector domains. This is vividly illustrated in Figure 3, which demonstrates that when these two domains adhere to a consistent denoising path, the result quality is significantly enhanced. Conversely, deviation from this path leads to a noticeable drop in quality. Complementing this, Figure 6 highlights the efficacy of our method, showcasing that both the image and vector domains can generate results that are not only consistent with each other but also retain the diversity of the generated layouts." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 47, + 647, + 231, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 231, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 231, + 659 + ], + "type": "text", + "content": "4.2.2 The Effect of Attention Score Maps" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "The attention score maps are the key for the visual composing capability of our model. We visualize the clear spatial correlation between attention scores and corresponding elements in the image domain in Figure 7. Our vector" + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 310, + 308, + 541, + 453 + ], + "blocks": [ + { + "bbox": [ + 310, + 308, + 541, + 453 + ], + "lines": [ + { + "bbox": [ + 310, + 308, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 310, + 308, + 541, + 453 + ], + "type": "image", + "image_path": "f5f3c47a8147c002e6af0ada6747c65bbbc57243d963e870c8b9a3a371bbe4c8.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 464, + 545, + 509 + ], + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 509 + ], + "type": "text", + "content": "Figure 8. Through attention masking, we can decompose all the elements in the latent image space, and include them one by one to create a layered image representation. The bottom right image shows the original design." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 520, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 556 + ], + "type": "text", + "content": "domain model leverages these spatial information to predict vector layouts by adaptively tracking each element and resolving the ambiguity between similar objects." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 556, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 663 + ], + "type": "text", + "content": "By manipulating the attention scores, we can also control the element generation in the image domain. For example, we can remove one element from canvas while keeping the rest layout unchanged by masking out its attention map. If we do this sequentially to each element, we can decompose image output into a layered representation organized by elements, as shown in Figure 8. This results in a virtual vector layout format, which may be used for similar purpose as our vector domain model." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 671, + 388, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 388, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 388, + 684 + ], + "type": "text", + "content": "4.3. Applications" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "With its two domain outputs and high controllability through attention maps, our model can be used for a range" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9228" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 110, + 176 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 110, + 176 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 110, + 176 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 110, + 176 + ], + "type": "image", + "image_path": "d089ac2e8d8aeba559a197ded7a56ba34bd8802fb254e90c88ccab772dc96195.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 178, + 110, + 186 + ], + "lines": [ + { + "bbox": [ + 54, + 178, + 110, + 186 + ], + "spans": [ + { + "bbox": [ + 54, + 178, + 110, + 186 + ], + "type": "text", + "content": "Input Layout Boxes" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 72, + 170, + 176 + ], + "blocks": [ + { + "bbox": [ + 112, + 72, + 170, + 176 + ], + "lines": [ + { + "bbox": [ + 112, + 72, + 170, + 176 + ], + "spans": [ + { + "bbox": [ + 112, + 72, + 170, + 176 + ], + "type": "image", + "image_path": "4217ef944e4d5df95d8a1cc006ac107e06822916a387024f0748b853248345e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 177, + 235, + 186 + ], + "lines": [ + { + "bbox": [ + 163, + 177, + 235, + 186 + ], + "spans": [ + { + "bbox": [ + 163, + 177, + 235, + 186 + ], + "type": "text", + "content": "Generated Layout Images" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 227, + 176 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 227, + 176 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 227, + 176 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 227, + 176 + ], + "type": "image", + "image_path": "a47382988755cfbb3c3544228bc176e68ce9bac547f6330c63a13b31a3e47b06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 197, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 287, + 229 + ], + "type": "text", + "content": "Figure 9. Layout controlled visual style variation. Given input layout with bounding boxes, our image model can generate visual results with the same layout and different styles." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 283, + 176 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 283, + 176 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 283, + 176 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 283, + 176 + ], + "type": "image", + "image_path": "021c57a21423304af37281f54cdd9f0d4a866674c8107b1d9bdf3c8bd2ca861a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 237, + 282, + 386 + ], + "blocks": [ + { + "bbox": [ + 52, + 237, + 282, + 386 + ], + "lines": [ + { + "bbox": [ + 52, + 237, + 282, + 386 + ], + "spans": [ + { + "bbox": [ + 52, + 237, + 282, + 386 + ], + "type": "image", + "image_path": "2fc73a3c26712e1a2cf4420abaa3f8222242bf96bb70c1886eab9c5542e98f57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "content": "Figure 10. Canvas resize with different aspect ratios. Our model can adapt layout under various aspect ratios." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 52, + 426, + 282, + 534 + ], + "blocks": [ + { + "bbox": [ + 52, + 426, + 282, + 534 + ], + "lines": [ + { + "bbox": [ + 52, + 426, + 282, + 534 + ], + "spans": [ + { + "bbox": [ + 52, + 426, + 282, + 534 + ], + "type": "image", + "image_path": "ef981f5c6a705963db74de49292c32dcf335c11394113a178c7cb6c8024b5a34.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 545, + 287, + 579 + ], + "lines": [ + { + "bbox": [ + 47, + 545, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 287, + 579 + ], + "type": "text", + "content": "Figure 11. Controlled layout generation. Our model allows users to add conditions on design elements to customize the result. In this example, the user specifies the location of the balloons." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 594, + 139, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 139, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 139, + 605 + ], + "type": "text", + "content": "of design applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 665 + ], + "type": "text", + "content": "In Figure 9, we produce various visual designs with the same spatial layout conforming to given element boxes by setting the attention scores for each element accordingly. This unlocks the power of the pretrained large image model for visual style variation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 665, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 700 + ], + "type": "text", + "content": "In Figure 10, we generate layouts of different canvas sizes by changing the input condition and masking the intermediate latent states during the denoising process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "In Figure 11, when a user specifies the position of a cer" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 73, + 416, + 175 + ], + "blocks": [ + { + "bbox": [ + 315, + 73, + 416, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 73, + 416, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 416, + 175 + ], + "type": "image", + "image_path": "adb0818d2ac48da261c4fd8d8f4954f51627ad0d8445b29ead68423f8bef0764.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 177, + 416, + 268 + ], + "blocks": [ + { + "bbox": [ + 315, + 177, + 416, + 268 + ], + "lines": [ + { + "bbox": [ + 315, + 177, + 416, + 268 + ], + "spans": [ + { + "bbox": [ + 315, + 177, + 416, + 268 + ], + "type": "image", + "image_path": "eeb502ff6e0799c93e6d7a8c03074aa54b59ea0518a827af461f7ebf12b7c2b9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 270, + 395, + 281 + ], + "lines": [ + { + "bbox": [ + 335, + 270, + 395, + 281 + ], + "spans": [ + { + "bbox": [ + 335, + 270, + 395, + 281 + ], + "type": "text", + "content": "Original Design" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 417, + 73, + 477, + 175 + ], + "blocks": [ + { + "bbox": [ + 417, + 73, + 477, + 175 + ], + "lines": [ + { + "bbox": [ + 417, + 73, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 417, + 73, + 477, + 175 + ], + "type": "image", + "image_path": "58c3547cb4d1512c4dea0f08d5683794a042432cd84b965977df31dc772e2448.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 417, + 177, + 477, + 268 + ], + "blocks": [ + { + "bbox": [ + 417, + 177, + 477, + 268 + ], + "lines": [ + { + "bbox": [ + 417, + 177, + 477, + 268 + ], + "spans": [ + { + "bbox": [ + 417, + 177, + 477, + 268 + ], + "type": "image", + "image_path": "006f2bfa06bb2aa4128048839ceccb8063f3ba411025be83642a36c38fc14564.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 270, + 468, + 281 + ], + "lines": [ + { + "bbox": [ + 427, + 270, + 468, + 281 + ], + "spans": [ + { + "bbox": [ + 427, + 270, + 468, + 281 + ], + "type": "text", + "content": "Input Style" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 296, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 306, + 296, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 545, + 319 + ], + "type": "text", + "content": "Figure 12. Layout style transfer. Our model can transfer the style from the reference design to the input design." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 480, + 73, + 539, + 175 + ], + "blocks": [ + { + "bbox": [ + 480, + 73, + 539, + 175 + ], + "lines": [ + { + "bbox": [ + 480, + 73, + 539, + 175 + ], + "spans": [ + { + "bbox": [ + 480, + 73, + 539, + 175 + ], + "type": "image", + "image_path": "371cdc007553cd04044be88736bfce118eadf8aaa8dd2f7f2fcba3aa1afd9fd8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 480, + 177, + 539, + 268 + ], + "blocks": [ + { + "bbox": [ + 480, + 177, + 539, + 268 + ], + "lines": [ + { + "bbox": [ + 480, + 177, + 539, + 268 + ], + "spans": [ + { + "bbox": [ + 480, + 177, + 539, + 268 + ], + "type": "image", + "image_path": "4c0d5d46a62b5fd4422cdc455f2d432806d5af9880fac4aaf532f87ef870826d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 270, + 532, + 281 + ], + "lines": [ + { + "bbox": [ + 486, + 270, + 532, + 281 + ], + "spans": [ + { + "bbox": [ + 486, + 270, + 532, + 281 + ], + "type": "text", + "content": "New Design" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 339, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 386 + ], + "type": "text", + "content": "tain element (the balloons of digits “20”) to a desired position, we fix the attention map for this element during inference, and only optimize the layout for the remaining elements." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 388, + 546, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 546, + 507 + ], + "type": "text", + "content": "In Figure 12, we transfer the layout from one design image to a new set of elements by propagating the attention maps from an existing design to a new one. The image model takes a given reference image as the initial noisy latent state, accompanied by a new set of elements. During inference, the image model aligns the visual content to input elements while preserving the initial reference layout. The vector model utilizes the aligned attentions to create a new design incorporating both the input elements and reference layout." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 517, + 378, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 378, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 378, + 530 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "type": "text", + "content": "Our paper introduces a new approach to layout design, combining vector and image domains through a dual diffusion model. Compared to previous methods solely based on the vector domain, this model has shown notable improvement on layout visual quality while maintaining the flexibility of vector editing on several datasets. Our model also shows flexible controllability via element attention manipulation, enabling a set of design applications across image and vector domains." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 645, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 645, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 545, + 705 + ], + "type": "text", + "content": "While we have made strides in integrating visual elements into layout design, our model is an initial step in addressing this complex task. Our findings offer a new perspective to the layout generation problem, and point out a potential direction for future exploration." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9229" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 303 + ], + "type": "text", + "content": "[4] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 359 + ], + "type": "text", + "content": "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 287, + 417 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 452 + ], + "type": "text", + "content": "[7] Chin-Yi Cheng, Forrest Huang, Gang Li, and Yang Li. Play: Parametrically conditioned layout generation using latent diffusion. arXiv preprint arXiv:2301.11529, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 520 + ], + "type": "text", + "content": "[8] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual ACM symposium on user interface software and technology, pages 845-854, 2017. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 287, + 554 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "text", + "content": "[10] Vedit Goel, Elia Peruzzo, Yifan Jiang, Dejia Xu, Nicu Sebe, Trevor Darrell, Zhangyang Wang, and Humphrey Shi. Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "text", + "content": "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004-1014, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 130, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 547, + 174 + ], + "type": "text", + "content": "[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "type": "text", + "content": "[15] Hsiao Yuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "text", + "content": "[16] Mude Hui, Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie, Yuwang Wang, and Yan Lu. Unifying layout generation with a decoupled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1951, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 287, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 342 + ], + "type": "text", + "content": "[17] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. LayoutDM: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 343, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 397 + ], + "type": "text", + "content": "[18] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Towards flexible multi-modal document models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14287-14296, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[19] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layoutvae: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[20] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via latent optimization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 88–96, 2021. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "text", + "content": "[21] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "text", + "content": "[22] Xiang Kong, Lu Jiang, Huiwen Chang, Han Zhang, Yuan Hao, Haifeng Gong, and Irfan Essa. Blt: bidirectional layout transformer for controllable layout generation. In European Conference on Computer Vision, pages 474-490. Springer, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "text", + "content": "[23] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "type": "text", + "content": "[24] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 491-506. Springer, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9230" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[26] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "text", + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "type": "text", + "content": "[28] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 351 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 463 + ], + "type": "text", + "content": "[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 531 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 533, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 565 + ], + "type": "text", + "content": "[33] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. arXiv preprint arXiv:2303.00262, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 567, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 600 + ], + "type": "text", + "content": "[34] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. Towards diverse and consistent typography generation. arXiv preprint arXiv:2309.02099, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[36] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Layoutnuwa: Revealing the hidden layout expertise of large language models. arXiv preprint arXiv:2309.09506, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[37] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 520 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "text", + "content": "[39] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "text", + "content": "[40] Kota Yamaguchi. Canvasvae: Learning to generate vector graphic documents. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5481-5489, 2021. 1, 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "type": "text", + "content": "[41] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "type": "text", + "content": "[42] Junyi Zhang, Jiaqi Guo, Shizhao Sun, Jian-Guang Lou, and Dongmei Zhang. Layoutdiffusion: Improving graphic layout generation by discrete diffusion probabilistic models. arXiv preprint arXiv:2303.11589, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 386 + ], + "type": "text", + "content": "[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 547, + 430 + ], + "type": "text", + "content": "[44] Xinru Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG), 38(4):1-15, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 432, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 547, + 475 + ], + "type": "text", + "content": "[45] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publ aaynet: largest dataset ever for document layout analysis. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1015-1022. IEEE, 2019. 1, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 476, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 547, + 520 + ], + "type": "text", + "content": "[46] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout gan for visual-textual presentation designs. arXiv preprint arXiv:2205.00303, 2022. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9231" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_content_list.json b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53df64f2bcea975f16ceb60848092f54b36da251 --- /dev/null +++ b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Visual Objectification in Films: Towards a New AI Task for Video Interpretation", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julie Tores1,2", + "bbox": [ + 181, + 179, + 290, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lucile Sassatelli $^{1,3}$", + "bbox": [ + 320, + 180, + 470, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hui-Yin $\\mathbf{W}\\mathbf{u}^{4}$", + "bbox": [ + 496, + 180, + 602, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clement Bergman", + "bbox": [ + 632, + 180, + 785, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Léa Andolfi", + "bbox": [ + 179, + 198, + 284, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Victor Ecrement6", + "bbox": [ + 315, + 198, + 452, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Frédéric Precioso²", + "bbox": [ + 482, + 198, + 630, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thierry Devars6", + "bbox": [ + 660, + 198, + 789, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Magali Guaresi", + "bbox": [ + 251, + 215, + 383, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virginie Julliard", + "bbox": [ + 413, + 215, + 552, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sarah Lecossais", + "bbox": [ + 581, + 215, + 718, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Université Côte d'Azur, CNRS, I3S, France", + "bbox": [ + 94, + 233, + 447, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Université Côte d'Azur, CNRS, Inria, I3S, France", + "bbox": [ + 475, + 233, + 875, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Institut Universitaire de France", + "bbox": [ + 191, + 251, + 446, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Université Côte d'Azur, Inria, France", + "bbox": [ + 475, + 251, + 777, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5Université Côte d'Azur, CNRS, BCL, France Sorbonne Université, GRIPIC", + "bbox": [ + 163, + 268, + 805, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "7Université Sorbonne Paris Nord, LabSIC", + "bbox": [ + 320, + 286, + 650, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "julie.tores@univ-cotedazur.fr", + "bbox": [ + 352, + 306, + 611, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 312, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In film gender studies, the concept of \"male gaze\" refers to the way the characters are portrayed on-screen as objects of desire rather than subjects. In this article, we introduce a novel video-interpretation task, to detect character objectification in films. The purpose is to reveal and quantify the usage of complex temporal patterns operated in cinema to produce the cognitive perception of objectification. We introduce the ObyGaze12 dataset, made of 1914 movie clips densely annotated by experts for objectification concepts identified in film studies and psychology. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. Our new dataset and code are made available to the community.", + "bbox": [ + 75, + 388, + 472, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 632, + 209, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In film gender studies, the concept of \"male gaze\" [41] refers to the way the characters – especially women – are portrayed on-screen as objects of desire rather than subjects. Consider in Figure 1 how objectification is manifested in various ways such as camera placement and movement, the gaze interactions between characters, the choice of clothing, and arrangement of scene elements. Such disparities in how people are presented, depicted or addressed to in digital contents based on their gender has large-scale social implications such as the perpetuation of harmful stereotypes and hostile social situations.", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These disparities have been the subject of an increasing number of studies at the intersection of social and computational sciences. In online social networks, computational approaches to sexism detection have been increasingly investigated for textual data as a part of hate speech detection.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A Unreturned gaze", + "bbox": [ + 503, + 361, + 624, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6061d4d9ee8b90046c4a3e7fe37b2a5951c45445f4b706ae11dcf946a39c5278.jpg", + "image_caption": [ + "Direction of gaze" + ], + "image_footnote": [], + "bbox": [ + 503, + 372, + 617, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b509430971965b5fbd75ef6728f306b89afc9bad029a49256216fcd022b52cfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 372, + 681, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4e87c2e2cbfd7bde9098f848f7a7045e4613dd8ce6d6e9e522acd7aa4b1be18.jpg", + "image_caption": [ + "Direction of gaze" + ], + "image_footnote": [], + "bbox": [ + 687, + 372, + 795, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c10f2e48be524ec11bd908c158d0280647d7dd7a4ba3816c8d438dbd370e78a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 372, + 877, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\bullet$ Submissive posture, lower-body nudity", + "bbox": [ + 503, + 428, + 642, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0942612c1d711246e889d89dd76933f801d0a1cc400b3a0a7dc99ef90b8223fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 454, + 612, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shot - woman in parallel with animal", + "bbox": [ + 658, + 426, + 890, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e4882fc50ebcfeff8a96f46599ec3b724f4846777251e801f49e1ee686697b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 439, + 792, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd901d5686ff13217d2090174bbccabaa78c96d75fed3a9582852dac6b59134c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 439, + 866, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9d0dd0688b11567d69de0d49956748f691553a503357fe6aae05576820f6fd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 469, + 790, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d940a7225c230f84930ebc9bde10dc5e0c9421e4cf46634175efd18eff7cb89d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 469, + 866, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clothing - transparent, look-up and down the", + "bbox": [ + 506, + 502, + 661, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ent, close-up on hips, domestic activities, the body", + "bbox": [ + 637, + 500, + 882, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fdcfcac27c4d446ca009ee59f00463c6c04a6c4c7ff584c3777053eddd9e5952.jpg", + "image_caption": [ + "full body $\\rightarrow$ approach $\\rightarrow$ close-up on hips" + ], + "image_footnote": [], + "bbox": [ + 504, + 525, + 633, + 555 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6adfd1eae3ccf7b42380146c7ee1c9a75cf5e6a4d47315723b9dd82a1c3ddb0.jpg", + "image_caption": [ + "Figure 1. In modern film media, the unequal characterization of gender on screen frequently evokes concepts of objectification, such as (A) unequal gaze (Pulp Fiction, 1994), (B) Nudity and submissive postures (Pulp Fiction, 1994), (C) animalisation or infantilisation (Marley and Me, 2008), and (D) transparent clothing, camera framing, domestic gender roles, and voyeurism (Gone Girl, 2014)." + ], + "image_footnote": [], + "bbox": [ + 637, + 525, + 725, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f83c08da26f3592d3421b389c8574122ff7f8c32dc86a9db5a189a0bdd859f59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 525, + 815, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9ed2065e7ac9bb539744f8dec66bca1eb858b42ca9df25292527bf0802ba2c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 525, + 892, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As explained by Samory et al. [47], sexism is a complex sociological construct, whose high-level interpretive nature and subtle dimensions beyond offensive speech make for an unsolved challenge. In visual media such as films and TV series, the characters they depict shape our collective imagination and perception of sociological constructs, such as gender, race, and class. Currently, most large-scale approaches to understanding gender representation in these media have focused on quantifying the presence of women in the image and audio content (e.g., [35, 39, 51]). However, works in social sciences show that quantifying the presence of gender on screen is insufficient for grasping the issue of", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10864", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gender inequalities in visual media. For films, the classic Bechdel test, although useful and simple, considers neither the visual modality, which is key to analyzing gender depiction [41], nor the textual constructs of speech and dialogue [49]. While a few works have investigated sexist memes [16, 19], sexist advertisement [18], and characterized the on-screen positioning and co-occurrence of certain groups with respect to scene types and objects (e.g., Wang et al. for still images [54], Jang et al. [26] for films), computational approaches to interpretive sexism in visual media remain very scarce.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this article, we introduce a new challenging task for computer vision: detecting character objectification in films. Owing to the importance of this question, we consider it is critical to support the design of explainable methods and fine-grained model error analysis, which we address by densely annotating video data for theory-driven concepts. This presents a major step to tackle the question of subtle sexism in videos, operationalizing the popularly known concept of male gaze with the construct of objectification, and specifically considering the temporal dimension where such video patterns unfold. The end goal is to enable large-scale quantification and characterization of complex patterns producing on-screen objectification, and unveil possible correlations along the lines of the gender or race constructs.", + "bbox": [ + 75, + 261, + 472, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are:", + "text_level": 1, + "bbox": [ + 96, + 508, + 259, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a novel video-interpretation task to detect character objectification in films. This is an interpretive task, hence extending beyond the more classical yet still challenging video-understanding tasks, and involving a subjective judgement. In a team involving media studies experts, and building on results in cinematography and psychology, we design a thesaurus of visual objectification, defining coarse-grained concepts with exemplified instances. This thesaurus is then used to formulate precise annotation guidelines. We introduce the ObyGaze12 dataset to the community, with 1914 clips from 12 films densely annotated by experts for concepts of objectification, including hard negative examples. It corresponds to $25\\%$ of the MovieGraphs dataset. We verify the consistency of the obtained data, and provide first analyses showing the compositional nature of objectification. The dataset is meant to explore the complex temporal patterns producing character objectification in films. To the best of our knowledge, it is the first work proposing a computational approach to this interpretive task in videos.", + "bbox": [ + 75, + 523, + 472, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We verify that the new task of objectification detection in videos is accessible by testing recent vision and vision-language models, and that hard negative examples improve classification. We also investigate the model weaknesses in representing each objectification concept. To do so, a", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thorough analysis is carried out with Concept Bottleneck Models (CBMs) which allows us to identify the individual concepts that pose the greatest challenge, namely: Type of Shot, Look, Posture, and Appearance.", + "bbox": [ + 496, + 90, + 893, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, this is one of the few video datasets with dense concept-based annotations for a high-level construct, and the first for objectification. The dataset and code used in this article is entirely provided in https://github.com/husky-helen/ObyGaze12.", + "bbox": [ + 496, + 166, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The article is organized as follows. In Sec. 2, we first provide a review of the relevant works on visual biases in films, dataset creation, and models for video understanding. We then introduce dataset creation for ObyGaze12 and present first analyses in Sec. 3. Sec. 4 presents the evaluation of models on the new task, and the analysis of the difficulty of concept representation with CBMs. Finally, we provide discussions on ethical aspects and challenges in Sec. 5, and conclusions in Sec. 6.", + "bbox": [ + 496, + 242, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 390, + 645, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we position our contributions with respect to the relevant existing work. First, we introduce biases in visual datasets and computational approaches to the analysis of visual gender representation in films. We then discuss interpretive-level tasks, increasingly common in natural language processing, and the approaches to dataset creation to instantiate them for ML approaches. Here we highlight the scarcity of visual datasets made for high-level interpretive tasks, in particular for video data. Finally, we introduce the video understanding approaches that we consider to benchmark on our new interpretive task, specifically focusing on explainable concept-based approaches to locate the challenges ahead in video interpretation for this new task.", + "bbox": [ + 496, + 415, + 893, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual biases in film datasets", + "text_level": 1, + "bbox": [ + 500, + 619, + 756, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The task we introduce is connected to the general problem of bias detection. As exposed by Fabbrizzi et al. [15], biases in visual datasets can be classified into selection bias (how subjects are included in a dataset), framing bias (how the visual content has been artificially composed) and label bias (errors or disparities in the labelling data). Our contribution is closely related to the framing and labelling biases. In film datasets, the first studies of biases in gender representation were from a presence quantification perspective. Guha et al. [21] and Somandepalli et al. [51] automatically estimate the screen time (from video) and speaking time (from audio) of male and female characters in Hollywood movies. They show that women are seen (36% screen time) and heard (41% screen time) significantly less than male characters. Mazieres et al. [39] consider a movie dataset spanning three decades, and show a temporal trend towards general fairer representation between both binary genders.", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10865", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "They however show simultaneously that the applied framings remain unfair, with only $40\\%$ of one-face frames featuring a female (60% for males). Jang et al. specifically analyze the qualitative framing differences in gender portrayal in a dataset of 20 Hollywood movies and 20 Korean movies [26]. They find that female characters are portrayed with lower emotional diversity, spatial occupancy, temporal occupancy, intellectual image or mean age.", + "bbox": [ + 75, + 90, + 467, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast in this article, we take a first step towards detecting bias in gender representation from a high-level construct, objectification, qualitatively described in various disciplines such as cinematography [9, 41], social psychology [1, 32], and neuroscience [5-7]. Objectification is produced by complex temporal patterns never analyzed computationally in videos until now.", + "bbox": [ + 75, + 213, + 467, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Interpretive-level tasks and dataset creation", + "text_level": 1, + "bbox": [ + 76, + 330, + 447, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the same time that biases in visual datasets are uncovered and analyzed, other approaches aim to detect bias in human data. Detecting highly interpretive constructs, such as hate speech, propaganda, sexism, and racism, has been a long-standing endeavor in NLP. Da San Martino et al. [12] recruited to 4 experts to annotate news articles with text spans associated to 18 possible propaganda techniques. Samory et al. uncover the challenge of construct complexity for sexism detection [47]. They observe that multiple articles for automating this task consider widely different definitions of sexism, often referring to sub-dimensions of the broader construct. They consider existing works in social psychology where sexism dimensions have long been operationalized with sub-scale questions tested for consistency. To approach dataset creation for on-screen objectification, we inspire from these two last works, and on dense annotation approaches of image datasets recently proposed for the medical domain [13]. To the best of our knowledge, approaches for sexism detection from visual content are scarce, and almost none existing for video data. Two main types of visual content have been considered so far: hateful or sexist memes [16, 29], and sexist advertisements [18] which can also relate to symbolic advertisement understanding [25, 28].", + "bbox": [ + 75, + 356, + 467, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In films, analysis of biases in gender representation has been also automated with NLP approaches applied to film scripts [2, 37]. Specifically, Martinez et al. [37] propose a RNN-based model to automatically extract agent-verbal-patient triplets. From 912 movie scripts, they show that male characters are associated with a higher agency while female characters are more frequently the object of gaze. Su et al. introduce the more abstract task of trope understanding in movies [52]. Our work is close to this last one as we also introduce a dataset of films annotated for a higher-level construct beyond event and story understanding. However, we provide dense annotations of sequences with constituency", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tive concepts for the high-level construct of character objectification. Also, while Su et al. exploit an existing online base contributed to by the community, we design and carry out a strictly defined annotation process by experts.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this article, we set out from the concept of male gaze defined in various ways in film gender studies [9, 36, 41]. Mulvey [41] characterizes the concept of gaze by the three relations between the camera and the characters, between the characters, but also between the spectator and the characters. While some formalizations of gaze set out from the gender of the director [36], Brey defines female and male gaze only from the film content, with a corpus analysis focusing on aesthetics [9] and revolving around the construct of objectification. We consider this gaze analysis of Brey, which does not significantly rely the socio-historical context of production and reception. We build the operationalization objectification in both cinematography and psychology, to create the conditions to make a new challenge accessible to the CV community: we produce a strict annotation process to obtain densely annotated video data and analyze where the new challenges lie ahead. We generally position our approach producing a non-large scale but high-quality dataset that we hope will be useful within the lines of the call of Paullada et al. for such data-centric AI approaches [44].", + "bbox": [ + 496, + 152, + 890, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Approaches to video and movie understanding", + "text_level": 1, + "bbox": [ + 498, + 482, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal foundation models, such as CLIP [45] and ALIGN [27], learn aligned image and text representations through contrastive pre-training on large-scale closed datasets. Generalizations of the CLIP model to video data have included VideoCLIP [57] and two X-CLIP models [34, 42]. In particular, X-CLIP [42], which we employ in this article, expands CLIP with video temporal modeling and video-adaptive textual prompts.", + "bbox": [ + 496, + 506, + 890, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Legacy and large-scale pre-trained vision-language models have been leveraged for movie-related tasks. Bose et al. consider the difficulty of visual scene recognition in movies due to domain mismatch and create the MovieCLIP dataset obtained from weakly labeling movie shots from scene categories using the CLIP model [8]. An important vision-language movie-related task is audio-description for the visually-impaired, for which a major dataset introduced recently is MAD [50], gathering sparse natural language sentences grounded in over 1200 hours of movie videos. AutoAD [23] and AutoADII [22] are two recent approaches to generate audio-description from the video, both leveraging CLIP to learn to prompt GPT. To design approaches to learn human-level constructs, such as emotions, interactions or relationships, datasets generated with human supervision are also instrumental. A prominent representative is the MovieGraphs dataset, providing detailed annotations of clips of 51 movies with emotional states, character interac", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10866", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions and relationships, and other scene reasoning elements [53]. It has prompted works tackling such recognition tasks [17, 33]. In this article, we build on the MovieGraphs dataset to annotate a selection with the construct of objectification, to later analyze it in connection with the other annotated social elements.", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this article, we aim to assess the capacity of CLIP-based methods to provide relevant embeddings for the concept of objectification. We do so with a direct evaluation of classification results when an adapter (MLP) fed by X-CLIP embeddings is learnt. We analyze the results with a concept-based approach by building on Concept Bottle-neck Models (CBM) [31]. Concept-based models are an active area of research in XAI, with works tackling the accuracy-explainability tradeoff [60] and the need for user-defined concepts [58]. We specifically employ Post-hoc CBM (PCBM) [59], which consists in learning a concept subspace (made of Concept Activation Vectors [30]) in the embedding space of the pre-trained model. Data samples are then projected in this concept subspace, from where the classification task can be performed with an interpretable classifier.", + "bbox": [ + 75, + 183, + 470, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Data and methods", + "text_level": 1, + "bbox": [ + 76, + 441, + 254, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section presents our approach to create the first dataset for visual objectification in videos, specifically in films. We name this dataset ObyGaze12, short for ObjectifyingGaze12, which has the following highlights:", + "bbox": [ + 75, + 469, + 468, + 529 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- It considers the multiple dimensions of the construct of visual objectification, made of filmic (framing and editing over successive shots, camera motion, etc.) and iconographic properties (visible objects, body parts, attire, character interactions, etc.).", + "- It is based on a thesaurus articulating five sub-constructs identified from multidisciplinary literature (film studies and psychology) from which we define typical instances, then grouped into coarse-grained visual concepts.", + "- The data is annotated densely with concepts, and shows the multi-factorial property of objectification, corroborating with some recent developments in cognitive psychology [6].", + "- A hard negative category is included with the goal to perform fine-grained error analysis and improve model generalization." + ], + "bbox": [ + 76, + 530, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. A thesaurus of objectification", + "text_level": 1, + "bbox": [ + 76, + 770, + 336, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first formalize the construct of visual objectification and derive key concepts to annotate in film scenes. Together with media studies experts, we identify five sub-constructs of objectification from literature on film cognition and film gender studies, and social and cognitive psychology: male gaze (point of view of a man on a woman) [5, 9, 41], sexualization [6, 7], surveillance of the feminine body [11, 14, 40],", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "female inaction / male possession [20, 48], and infantilism / animalization [41].", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These sub-constructs come with typical instances and examples from filmmaking techniques ([9, 41]) or validated questionnaires ([11, 40]), as shown in the middle and rightmost columns in Table 1. These typical instances are then grouped into eight coarse-grained visual concepts, corresponding to the possible means of production of visual objectification. They are shown in the left-most column in Table 1: type of shot (framing and gaze of camera), look (gaze of characters on the other), body (partial or full nudity, and sexually suggestive body parts), posture (connoting, e.g., childhood, submission or inaction), clothing (in relation to context and activities), appearance (age and makeup), expression of emotion (restrained or exaggerated according to gender role), and activities (linked to gender roles). Visual examples are provided in Fig. 1, where we show video samples of objectification concepts Look, Posture, Type of shot, Clothes and Activity.", + "bbox": [ + 496, + 122, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data selection", + "text_level": 1, + "bbox": [ + 500, + 388, + 643, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Over the various existing movie datasets (see Sec. 2 and [4, 24, 39, 43, 46, 50, 53], [51, Table 1]), many have overlapping titles and only a few have rich human supervision. Amongst these, the MovieGraphs dataset [53] includes rich, high-level human annotations of 7637 clips of 51 movies, with emotional states, interactions and relationships, and other social reasoning elements. These elements are important and valuable in exploring social concepts of objectification in visual media. These movies are also frequently adopted for media research in existing work in [4, 24, 39]. The movie clips have also a short duration – mostly within 1-5 minutes – facilitating dense and granular annotations (over 100 clips for an average 2 hour film) while preserving the possibility to observe longer-term interactions and story development across a number of shots. From the 7637 clips of the MovieGraphs dataset, we select 1914 clips to annotate for objectification, amounting to $25\\%$ of the dataset and 12 complete movies, which were selected to approximately reproduce the fraction of genres in the original dataset. The list of selected films, year of release, and genre can be found in Table 3 in Appendix 7.", + "bbox": [ + 496, + 412, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data annotation", + "text_level": 1, + "bbox": [ + 500, + 739, + 660, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Every selected movie is annotated by at least two experts for objectification level and concepts over the movie scenes. Specifically, the annotators were asked to repeat a three-step process for every scene they deemed interesting from an objectification perspective: (1) watch the movie entirely and when they identify a scene worth annotating, (2) delimit the clip by using the cutting function in our annotation tool, and (3) assigning an objectification level and annotate the concept(s) involved in the objectification rating. We define", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10867", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e652fa9707f0ad109b21e28ddb1d14abf0c7a0875c09dda16c72eba1a31604a5.jpg", + "table_caption": [ + "Table 1. Thesaurus of the typical instances and examples of visual objectification in films, grouped into eight main visual concepts used for annotation. Examples are possible means to produce one of the five sub-constructs of objectification (male gaze, sexualisation, surveillance of the feminine body, female inaction/male possession, infantilism/animalisation), to be assessed by annotators." + ], + "table_footnote": [], + "table_body": "
ConceptConcept instancesExamples
Type of shotShot suggesting man perspective in presence of womanclose-up on a man's face; body parts of woman
Shot suggesting man gaze on womancamera takes the perspective of a male character with first close-up on the face followed by camera motion looking a woman from bottom up
Shot showing a woman in parallel with an animalwoman at same level and position with a dog
LookVoyerismcharacter watching another one without their knowledge
Non-reciprocal gazewoman looking at man who does not look back
BodySuggested nudityclothing on floor; silhouette behind shower curtain; nude shadow on wall
Partial nuditynude upper or lower body; partially open clothing or draping; in underwear
Full nuditynude person fully or partially shown
Body parts suggestive of sexclose-up shots on breast, buttocks, hips, or lips
PostureGesture or posture connoting seductionlip-biting; hip roll; twisting or tucking hair
Gesture or posture connoting sexualityeating phallic symbols; arching back
Gesture or posture connoting inactionbeing undressed by someone
Gesture or posture connoting submissionleaning on a man
Gesture or posture connoting dependencefollowing a man
Skippingskipping gait
ClothingWet or transparent clothingthin shirt soaked in rain
Clothing impractical to situationwear pumps for running, a skirt when gardening
Color code associated to characterwoman with pink clothing and accessories
Older woman wearing infantile clothingwoman wearing an Alice band or high socks
AppearanceDiscrepancy between appearance of woman and context or biographical elementsperfect makeup when waking up; mother of heroine too young; young girl played by older actress
Exp. of emotionAsymmetric expression of emotionboys don't cry; woman being hysterical
ActivityDoing domestic activitiesdoing laundry, cooking, cleaning, being constantly in the kitchen
", + "bbox": [ + 76, + 142, + 888, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "four levels of objectification:", + "bbox": [ + 76, + 486, + 272, + 501 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Easy Negative: there are no elements suggestive of objectification. No concept can be annotated. Default value for watched but non-selected scenes.", + "- Hard Negative: the scene contains elements of objectification from the thesaurus, but their presence does not result in an objectification effect.", + "- Not Sure: the scene indicates objectification, but does not completely fit the definition in the thesaurus.", + "- Sure: the scene contains elements of objectification from the thesaurus, and their presence results in an objectification effect." + ], + "bbox": [ + 76, + 503, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Four expert annotators were recruited to annotate the dataset. A first presentation session was held to introduce our annotation tool and the annotation procedure. All four annotators were then given the same two films – Juno and Silver Linings Playbook – to annotate using the proposed methodology. A second meeting was then set up after the annotation of the two films to analyze the reasons for divergence and remedy them by identify which elements of the annotation guidelines to clarify and how. We then randomly assigned two annotators to each of the remaining 10 films to annotate separately.", + "bbox": [ + 75, + 671, + 470, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data processing and fusion The data processing is described in detail in Appendix 7. Following the annotation step, the annotations are then projected from the delimi-", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tations provided by each annotator onto the delimitations of the MovieGraphs clips. Since multiple annotations may overlap the same MovieGraphs clip, the annotation that is projected on the MovieGraphs clip corresponds to the annotation (including objectification level and concepts) with (1) the highest level of objectification that has at least $20\\%$ overlap with the MovieGraphs clip, and (2) when multiple annotations exist at the same level of objectification, the annotated concepts for these annotations are aggregated. The same process is used to aggregate the annotations of the annotators of a same clip: the maximum objectification level is kept, with possible aggregation of concepts in case both annotators chose the same level but annotated different concepts. The merged data is shared and used in the remaining of this article.", + "bbox": [ + 496, + 486, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Analysis of the ObyGaze12 dataset", + "text_level": 1, + "bbox": [ + 498, + 724, + 790, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we comment on some interesting statistics of the resulting annotations and concepts of the 1914 clips originally delimited in the MovieGraphs dataset.", + "bbox": [ + 496, + 748, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we verify data consistency by computing the inter-annotator agreement (IAA). Given the task of annotating timespans, we choose the $\\gamma$ agreement measure introduced in [38] (and used for, e.g., annotating text spans [12]) owing to its consideration of temporal alignment, multiple annotators, and label classification at the same time. It attributes a score between 1 (complete agreement) and $-\\infty$ . A value of", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10868", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\gamma \\leq 0$ indicates no agreement. The computation details of the $\\gamma$ metric is provided in Appendix 7. Considering all four categories EN, HN, NS, S, we obtain and average $\\gamma = 0.42$ . Not considering the clips annotated Not Sure (NS), which is the uncertain and \"noisy\" class in human annotations, the IAA increases to $\\gamma = 0.69$ . This shows the consistency of the obtained annotations despite the interpretive nature of the task. Let us also mention that recent works improve learning approaches by considering explicitly the IAA in case of low number of annotators with moderate agreement [10, 55, 56].", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we analyze the obtained annotations in Fig 2. The Sure category is the least represented with $16\\%$ , the Easy Negative being, as expected, the most represented class with $52\\%$ of clips. It is interesting to note that every concept is approximately annotated with the same rate throughout the Hard Negative, Not Sure and Sure levels of objectification. Finally, it is very interesting to observe that the average number of concepts annotated per clip increases with the level of objectification: 1.26 concepts on average per Hard Negative clip, 1.71 for Not Sure, up to 2.6 for Sure. We verify that this trend is observable for every single annotator. It gives an important insight into our video interpretation data: that objectification is a compositional process. This corroborates with recent findings in neuroscience experiments that found that a single element, such as clothing on its own, is not sufficient for people to perceive a character as an object [6].", + "bbox": [ + 75, + 257, + 472, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 526, + 209, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experiments have two objectives: to verify that the new classification task is feasible, and to identify the challenges of designing efficient models. To tackle these objectives, we", + "bbox": [ + 75, + 551, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c4f81a4703555bd1db8838a1fca58ac771c1c00e703a00353fcbf028e4b4419.jpg", + "image_caption": [ + "Figure 2. Distribution of visual factors annotated for each level of objectification (HN = Hard negative, NS = Not sure, S = Sure). The percentage of the dataset for each level of objectification as well as the average number of concepts per clip are also shown. (Best viewed in colors)" + ], + "image_footnote": [], + "bbox": [ + 81, + 619, + 465, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "consider pre-trained vision models and specifically address the following research questions:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Task accuracy - What are the baseline performances by pre-trained vision models on the objectification detection task? How does the performance vary with hard negative examples?", + "- Concept representation - Can we implement interpretable models of objectification using concepts? What is the quality of representation of every concept, and what are the objectification concepts poorly captured by current models?" + ], + "bbox": [ + 500, + 122, + 893, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Task accuracy", + "text_level": 1, + "bbox": [ + 500, + 270, + 645, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup We discard the Not Sure (NS) class from the ObyGaze12 dataset, as it gathers by definition samples highly uncertain for humans, and consider the Easy Negative (EN, $62\\%$ of the clip samples), Hard Negative (HN, $19\\%$ ) and Sure (S, $19\\%$ ) classes. We approach binary classification in a progressive way, the positive class being made of the S samples. We consider two levels of classification difficulty by composing the negative class either with EN samples, or with HN samples. The implementation details of cross-validation and data balancing are provided in Appendix 8. The average performance over the test set of the best models on validation folds are shown in Table 2 with standard deviations.", + "bbox": [ + 496, + 294, + 890, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We consider video embeddings obtained from pre-trained models owing to their zero-shot classification capabilities on video tasks. We select ViViT-B/16 [3], and the available X-CLIP model, trained on Kinetics [42]. We also re-train a X-CLIP model [34] on the LSMDC [46] film dataset, and refer to Appendix 10 for corresponding results, where all implementation details are described. We keep the pre-trained models frozen and perform an adaptative max pooling of the resulting frame tokens, and feed the output to an MLP made of 2 dense layers, the hidden layer with 128 neurons and ReLU activations, the last with 2 softmax neurons.", + "bbox": [ + 496, + 507, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results To assess the quality of the models on possibly imbalanced data with a minority of positive samples, we report the F1-scores in Table 2. First, by comparing with trivial classifiers (random and an all-positive, see App. 8), we observe that the task is indeed feasible, warranted by the data consistency described in Sec. 3 despite the interpretive nature of the task. Second, we observe that the inclusion of Hard Negative examples improves the classification results, showing the importance of a fine-grained annotation for highly-interpretive tasks. Results of X-CLIP on other configurations, specifically when the movies of clips in test are different from those in train, are shown in App. 8. The best results based on existing models are of moderate qual", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10869", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b5bfe66197df7d04f8c03a5d204622793d5dee11d73c033e06e11c389840e30d.jpg", + "table_caption": [ + "Table 2. F1-score on the binary task of objectification detection for models trained with easy or with hard negatives and tested on easy or all negative samples, with standard deviations." + ], + "table_footnote": [], + "table_body": "
Test TrainEN vs. S(EN U HN) vs. S
EN vs. SHN vs. SEN vs. SHN vs. S
ViViT-B/160.53 (0.18)0.62 (0.13)0.54 (0.24)0.73 (0.1)
X-CLIP0.79 (0.05)0.71 (0.05)0.66 (0.05)0.82 (0.03)
Random0.540.55
All positive0.080.06
PCBM-DT0.680.440.580.38
PCBM-LR0.640.430.500.37
", + "bbox": [ + 76, + 141, + 470, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ity, which calls for more investigation into where the difficulties lie.", + "bbox": [ + 75, + 285, + 468, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Concept accuracy", + "text_level": 1, + "bbox": [ + 76, + 327, + 250, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To infer on-screen objectification, it is key for the model to detect the means of its production, which correspond to the eight concepts listed in Table 1. We reiterate that in the ObyGaze12 dataset, every clip annotated with a level of objectification S, NS or HN is also annotated with the presence of instances of the eight concepts. For example, if the Body concept is annotated, it means that some level of nudity and/or suggestive body parts are shown on screen, that could contribute to the production of objectification. The means of producing objectification through the eight concepts can be subtle to detect, making it difficult to provide the final interpretation. To investigate this difficulty, we implement Post-hoc Concept Bottleneck Models (PCBMs) [59], which allow us to approach a classification task with pre-trained models in an interpretable way when concept-annotated data is available. In our case, from the X-CLIP embedding space where our video clips are represented, we identify a Concept Activation Vector (CAV) [30] for every concept. We then project the X-CLIP embedding of every clip onto the subspace defined by the eight CAVs. The representation of the clip that is the output of this bottleneck is a low-dimensional vector with number-of-concepts components. This vector can then be fed to an interpretable classifier for the objectification detection task.", + "bbox": [ + 73, + 349, + 472, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CAV computation For every concept $i$ , we collect two sets of samples: positive samples where concept $i$ is present, hence made of S and HN samples with the concept annotated as present, and negative with EN, S and HN without concept $i$ . We then train a linear SVM for each concept $i$ , the CAV of concept $i$ being the normal vector of the SVM hyperplan. To train the SVM, we split the data in 10 folds, and reserve the last fold for test. We then perform an 8-fold cross-validation to select the SVM (choice of margin tolerance $c$ ), every fold training set being balanced with different draws of negative sub-sampling.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpretable classifier We then train a decision tree (DT) and a logistic regression (LR) classifier on the same 8-fold cross-validation to classify the level of objectification, the classifiers being fed with the projection of every clip onto the CAVs.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task accuracy with PCBM We first verify the quality of objectification detection with F1-scores of PCBM-DT and PCBM-LR shown in Table 2. The results lower than X-CLIP are expected owing to the known accuracy-interpretability tradeoff of CBMs [59, 60]. They are above random and all-positive predictions when training on EN vs. S, which is indicative of the relevance of information held in the concepts. However, the low results obtained when training the DT and LR to distinguish between S and HN reveals the low quality of some CAVs, where the X-CLIP embeddings cannot be linearly well-separated for these concepts. We investigate this point next. The resulting DT is discussed in App. 9.", + "bbox": [ + 496, + 184, + 892, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concept accuracy We now analyze the quality of each obtained CAV by plotting its capability to classify whether the concept is present in a test sample. We consider a positive similarity between the X-CLIP embedding and CAV of concept $i$ indicative of the presence of concept $i$ . F1-score on the test set are shown in Fig. 3. Plots correspond to CAVs obtained from classifying the presence of concept $i$ against EN only (solid bars) and against EN with S and HN without concept $i$ (hatched bars). The former is used for PCBM-DT in Table 2. We first observe that concept detection is harder when negative samples also include S and HN samples (without the concept). This is expected considering that scenes tagged EN have by definition no element possibly conducive to objectification, and are hence likely to differ visually more from scenes where the concept is present, than do S and HN scenes without the concept. However within S and HN clips with and without the concept, such shortcuts cannot be exploited anymore. We observe in this case that the X-CLIP embedding related to concepts Type of shot, Posture, Look and Appearance are harder to separate linearly. This can be correlated with the analysis of factors of error detailed next. These subtler means of on-screen objectification therefore warrant future work to be properly captured and detected.", + "bbox": [ + 496, + 398, + 892, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Error Analysis We analyze the factors impacting objectification classification errors corresponding to the results shown in Table 2. We select an average X-CLIP-based adaptation model trained on HN vs. S and tested on (ENUHN) vs. S, and consider its predictions on the clips in the test set. We label each test clip with 0 if the model fails to predict the correct label of the clip, and with 1 otherwise. We describe the clip with a one-hot encoding vector", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10870", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/738fe22f7fad2dbd95c7f88aa8448030dc9708d00f061b53e7a5cec87d1bc493.jpg", + "image_caption": [ + "Figure 3. For every concept, F1-score of the best linear SVM selected to define the CAV of this concept. Positive samples (S and HN with the concept) must be separated from: [non-hatched bars] negative samples made of EN only, or [hatched bars] negative samples made of EN and S and HN without the concept." + ], + "image_footnote": [], + "bbox": [ + 86, + 92, + 459, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "corresponding to all 11 factors shown the y-axis of Fig. 4: every of the 8 concepts, and the Sure (S), Hard negative (HN) and Easy negative (EN) labels. We then train a logistic regression model on these clip descriptors to predict the failure/success labels. Fig. 4 shows the error regression weights associated with each factor, for two film examples. A negative weight indicates a contribution of the factor to a classification failure. We first observe that the HN characteristic contributes to a classification failure of the model, while S and EN contribute to classification success. Second, we observe that there is variability over the movies on the presence of which concept strongly influences the classification success. However, the presence of the Clothing concept seems to be a strong confounder. This can be due to the frequency of appearance of this concept in HN samples, and to the subtlety of the description of this concept (provided in Table 1), which should make it difficult for a pre-trained model to discriminate between an objectifying and non-objectifying overall label on the basis of Clothing. Third, it is worth noting that the concepts shown to be poorly linearly separable when described with the X-CLIP embeddings (see CAV analysis in Sec. 4.2), are also those with a non-stable contribution to the model errors over the film examples: Type of shot, Look, Posture and Appearance.", + "bbox": [ + 75, + 449, + 472, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 828, + 191, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical aspect This work has an explicit societal motivation in its purpose to tackle, with the help of AI, the analysis of complex temporal patterns operated in cinema that pro", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/605d1d403a836609b02b61a29651571e79fc53cce48fa2531709359423eb0a59.jpg", + "image_caption": [ + "Figure 4. Analysis of the factors of error for the objectification detection task: weights of a logistic regressor predicting whether the test set examples are well classified or not. Positive (resp. negative) weights indicate a positive (resp. negative) contribution to classification success. Left: As good as it gets, Right: Up in the Air." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 893, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "duce the perception of certain characters as objects. This is a challenging but valuable task that aims to uncover and quantify differences in how various identities may be portrayed on screen.", + "bbox": [ + 496, + 362, + 893, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and challenges A distinctive element of our work is the subjective judgement involved in annotating granular video elements for objectification. Video annotation is tedious, and approaching data annotation for such an interpretive task in a rigorous way is even more so, and difficult to scale. We therefore believe that pursuing high-quality, dense annotations with well-defined concepts goes a long way to tackle this new video interpretation task, which represents a valuable new challenge for the computer vision community.", + "bbox": [ + 496, + 449, + 893, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 619, + 619, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this article, we have introduced a new video interpretation task to detect character objectification in films. We have introduced the ObyGaze12 dataset, densely annotated by experts for objectification concepts defined from five sub-constructs identified in film studies and psychology. ObyGaze12 is made available to the community. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. We show that the representation learning of the concepts of Type of shot, Look, Posture and Appearance need to be improved.", + "bbox": [ + 496, + 645, + 893, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work has been partly supported by the French National Research Agency through the ANR TRACTIVE project ANR-21-CE38-0012-01. This work was partly supported by EU Horizon 2020 project AI4Media, under contract no. 951911 (https://ai4media.eu/).", + "bbox": [ + 496, + 830, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Operationalizing self-objectification: Assessment and related methodological issues. In Self-objectification in women: Causes, consequences, and counteractions., pages 23-49. American Psychological Association, Washington, 2011. 3", + "[2] Apoorv Agarwal, Jiehan Zheng, Shruti Kamath, Sriramkumar Balasubramanian, and Shirin Ann Dey. Key Female Characters in Film Have More to Talk About Besides Men: Automating the Bechdel Test. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 830-840, Denver, Colorado, 2015. Association for Computational Linguistics. 3", + "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, Montreal, QC, Canada, 2021. IEEE. 6", + "[4] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed Movies: Story Based Retrieval with Contextual Embeddings. Technical Report arXiv:2005.04208, arXiv, 2020. arXiv:2005.04208 [cs] type: article. 4", + "[5] Philippe Bernard, Sarah J. Gervais, and Olivier Klein. Objectifying objectification: When and why people are cognitively reduced to their parts akin to objects. European Review of Social Psychology, 29(1):82-121, 2018. Publisher: Routledge _eprint: https://doi.org/10.1080/10463283.2018.1471949. 3, 4", + "[6] Philippe Bernard, Florence Hanoteau, Sarah Gervais, Lara Servais, Irene Bertolone, Paul Deltenre, and Cecile Colin. Revealing Clothing Does Not Make the Object: ERP Evidences That Cognitive Objectification is Driven by Posture Suggestiveness, Not by Revealing Clothing. Personality and Social Psychology Bulletin, 45(1):16-36, 2019. 4, 6", + "[7] Philippe Bernard, Carlotta Cogoni, and Andrea Carnaghi. The Sexualization-Objectification Link: Sexualization Affects the Way People See and Feel Toward Others. Current Directions in Psychological Science, 29(2):134-139, 2020. 3, 4", + "[8] Digbalay Bose, Rajat Hebbar, Krishna Somandepalli, Haoyang Zhang, Yin Cui, Kree Cole-McLaughlin, Huisheng Wang, and Shrikanth Narayanan. Movieclip: Visual scene recognition in movies. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2082-2091, 2023. 3", + "[9] Iris Brey. Le regard féminin-Une révolution à l'écran. Média Diffusion, 2020. 3, 4", + "[10] M. Bucarelli, L. Cassano, F. Siciliano, A. Mantrach, and F. Silvestri. Leveraging inter-rater agreement for classification in the presence of noisy labels. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3439-3448, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6, 2", + "[11] Rachel M. Calogero. A Test of Objectification Theory: The Effect of the Male Gaze on Appearance Concerns in Col-" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lege Women. Psychology of Women Quarterly, 28(1):16-21, 2004. 4", + "[12] Giovanni Da San Martino, Seunghak Yu, Alberto Barrón-Cedeno, Rostislav Petrov, and Preslav Nakov. Fine-Grained Analysis of Propaganda in News Article. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5635–5645, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 5", + "[13] Roxana Daneshjou, Mert Yuksekgonul, Zhuo Ran Cai, Roberto A. Novoa, and James Zou. Skincon: A skin disease dataset densely annotated by domain experts for fine-grained debugging and analysis. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3", + "[14] Angela Denchik. Development and Psychometric Evaluation of the Interpersonal Sexual Objectification Scale. PhD thesis, The Ohio State University, 2005. 4", + "[15] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223: 103552, 2022. 2", + "[16] Elisabetta Fersini, Francesca Gasparini, and Silvia Corchs. Detecting Sexist MEME On The Web: A Study on Textual and Visual Cues. In 2019 8th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIw), pages 226-231, 2019. 2, 3", + "[17] Bei Gan, Xiujun Shu, Ruizhi Qiao, Haoqian Wu, Keyun Chen, Hanjun Li, and Bohan Ren. Collaborative noisy label cleaner: Learning scene-aware trailers for multi-modal highlight detection in movies. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18898-18907, 2023. 4", + "[18] Francesca Gasparini, Ilaria Erba, Elisabetta Fersini, and Silvia Corchs. Multimodal Classification of Sexist Advertisements.: In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications, pages 399-406, Porto, Portugal, 2018. SCITEPRESS - Science and Technology Publications. 2, 3", + "[19] Francesca Gasparini, Giulia Rizzi, Aurora Saibene, and Elisabetta Fersini. Benchmark dataset of memes with text transcriptions for automatic detection of multi-modal misogynistic content. arXiv:2106.08409 [cs], 2021. arXiv:2106.08409.2", + "[20] Sarah J. Gervais, Gemma Sáez, Abigail R. Riemer, and Olivier Klein. The Social Interaction Model of Objectification: A process model of goal-based objectifying exchanges between men and women. British Journal of Social Psychology, 59(1):248-283, 2020. 4", + "[21] Tanaya Guha, Che-Wei Huang, Naveen Kumar, Yan Zhu, and Shrikanth S. Narayanan. Gender Representation in Cinematic Content: A Multimodal Approach. In Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, pages 31–34, Seattle Washington USA, 2015. ACM. 2", + "[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel - who," + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10872", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13645-13655, 2023. 3", + "[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie Description in Context. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18930-18940, Vancouver, BC, Canada, 2023. IEEE. 3", + "[24] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. MovieNet: A Holistic Dataset for Movie Understanding. In Computer Vision – ECCV 2020, pages 709–727. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 4", + "[25] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic Understanding of Image and Video Advertisements. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1100-1110, Honolulu, HI, 2017. IEEE. 3", + "[26] Ji Yoon Jang, Sangyoon Lee, and Byungjoo Lee. Quantification of Gender Representation Bias in Commercial Films based on Image Analysis. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW):1-29, 2019. 2, 3", + "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 3", + "[28] Nasrin Kalanat and Adriana Kovashka. Symbolic image detection using scene and knowledge graphs. Technical Report arXiv:2206.04863, arXiv, 2022. arXiv:2206.04863 [cs] type: article. 3", + "[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3", + "[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie J. Cai, James Wexler, Fernanda B. Viégas, and Rory Sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International Conference on Machine Learning, 2017. 4, 7", + "[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In Proceedings of the 37th International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 4", + "[32] Holly B. Kozee, Tracy L. Tylka, and Casey L. Augustus-Horvath. Interpersonal Sexual Objectification Scale, 2011. 3", + "[33] Anna Kukleva, Makarand Tapaswi, and Ivan Laptev. Learning Interactions and Relationships Between Movie Characters. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9846-9855, Seattle, WA, USA, 2020. IEEE. 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, page 638–647, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6", + "[35] Sarah Macharia. Global Media Monitoring Project (GMMP), pages 1-6. John Wiley Sons, Ltd, 2020. 1", + "[36] Alicia Malone. The Female Gaze: Essential Movies Made by Women. Mango Publisher, 2018. 3", + "[37] Victor R. Martinez, Krishna Somandepalli, and Shrikanth Narayanan. Boys don't cry (or kiss or dance): A computational linguistic lens into gendered actions in film. PLOS ONE, 17(12):e0278604, 2022. 3", + "[38] Yann Mathet, Antoine Widlocher, and Jean-Philippe Métivier. The unified and holistic method gamma $(\\gamma)$ for inter-annotator agreement measure and alignment. Computational Linguistics, 41(3):437-479, 2015. 5, 1", + "[39] Antoine Mazieres, Telmo Menezes, and Camille Roth. Computational appraisal of gender representativeness in popular movies. Humanities and Social Sciences Communications, 8 (1):137, 2021. 1, 2, 4", + "[40] Nita Mary McKinley and Janet Shibley Hyde. The objectified body consciousness scale: Development and validation. Psychology of women quarterly, 20(2):181-215, 1996. 4", + "[41] Laura Mulvey. Visual Pleasure and Narrative Cinema. Screen, 16(3):6-18, 1975. 1, 2, 3, 4", + "[42] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 3, 6", + "[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. MovieCuts: A New Dataset and Benchmark for Cut Type Recognition. In Computer Vision – ECCV 2022, pages 668–685. Springer Nature Switzerland, Cham, 2022. Series Title: Lecture Notes in Computer Science. 4", + "[44] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[46] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie Description. International Journal of Computer Vision, 123(1):94-120, 2017. 4, 6", + "[47] Mattia Samory, Indira Sen, Julian Kohne, Fabian Flock, and Claudia Wagner. \"Call me sexist, but...\" : Revisiting Sexism Detection Using Psychological Scales and Adversarial Samples. page 12. 1, 3" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Maarten Sap, Marcella Cindy Prasettio, Ari Holtzman, Hannah Rashkin, and Yejin Choi. Connotation frames of power and agency in modern films. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 2329-2334, 2017. 4", + "[49] Alexandra Schofield and Leo Mehr. Gender-Distinguishing Features in Film Dialogue. In Proceedings of the Fifth Workshop on Computational Linguistics for Literature, pages 32–39, San Diego, California, USA, 2016. Association for Computational Linguistics. 2", + "[50] M. Soldan, A. Pardo, J. Alcazar, F. Heilbron, C. Zhao, S. Giancola, and B. Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5016-5025, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3, 4", + "[51] Krishna Somandepalli, Tanaya Guha, Victor R. Martinez, Naveen Kumar, Hartwig Adam, and Shrikanth Narayanan. Computational Media Intelligence: Human-Centered Machine Analysis of Media. Proceedings of the IEEE, 109(5): 891-910, 2021. 1, 2, 4", + "[52] Hung-Ting Su, Po-Wei Shen, Bing-Chen Tsai, Wen-Feng Cheng, Ke-Jyun Wang, and Winston H. Hsu. TrUMAn: Trope Understanding in Movies and Animations. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 4594–4603, New York, NY, USA, 2021. Association for Computing Machinery. event-place: Virtual Event, Queensland, Australia. 3", + "[53] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. MovieGraphs: Towards Understanding Human-Centric Situations from Videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 1", + "[54] Angelina Wang, Arvind Narayanan, and Olga Russakovsky. REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets. In Computer Vision – ECCV 2020, pages 733–751. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 2", + "[55] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. ArXiv, abs/2110.12088, 2021. 6, 2", + "[56] Jiaheng Wei, Zhaowei Zhu, Tianyi Luo, Ehsan Amid, Abhishek Kumar, and Yang Liu. To Aggregate or Not? Learning with Separate Noisy Labels. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2523-2535, New York, NY, USA, 2023. Association for Computing Machinery. event-place: Long Beach, CA, USA. 6, 2", + "[57] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19187-19197, Vancouver, BC, Canada, 2023. IEEE. 4", + "[59] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR^2Struct: Privacy, Accountability, Interpretability, Robustness, Reasoning on Structured Data, 2022. 4, 7", + "[60] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models. In Advances in Neural Information Processing Systems, 2022. 4, 7" + ], + "bbox": [ + 501, + 90, + 892, + 315 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10874", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_model.json b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9437fd04ca4f9b134416da8a9ccec73b8b715738 --- /dev/null +++ b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.131, + 0.892, + 0.154 + ], + "angle": 0, + "content": "Visual Objectification in Films: Towards a New AI Task for Video Interpretation" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.18, + 0.292, + 0.198 + ], + "angle": 0, + "content": "Julie Tores1,2" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.181, + 0.471, + 0.198 + ], + "angle": 0, + "content": "Lucile Sassatelli\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.181, + 0.603, + 0.198 + ], + "angle": 0, + "content": "Hui-Yin \\(\\mathbf{W}\\mathbf{u}^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.181, + 0.786, + 0.199 + ], + "angle": 0, + "content": "Clement Bergman" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.199, + 0.285, + 0.215 + ], + "angle": 0, + "content": "Léa Andolfi" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.199, + 0.454, + 0.215 + ], + "angle": 0, + "content": "Victor Ecrement6" + }, + { + "type": "text", + "bbox": [ + 0.483, + 0.199, + 0.631, + 0.216 + ], + "angle": 0, + "content": "Frédéric Precioso²" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.199, + 0.79, + 0.216 + ], + "angle": 0, + "content": "Thierry Devars6" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.216, + 0.384, + 0.234 + ], + "angle": 0, + "content": "Magali Guaresi" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.216, + 0.553, + 0.234 + ], + "angle": 0, + "content": "Virginie Julliard" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.217, + 0.719, + 0.234 + ], + "angle": 0, + "content": "Sarah Lecossais" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.234, + 0.448, + 0.251 + ], + "angle": 0, + "content": "1Université Côte d'Azur, CNRS, I3S, France" + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.234, + 0.877, + 0.251 + ], + "angle": 0, + "content": "2Université Côte d'Azur, CNRS, Inria, I3S, France" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.252, + 0.447, + 0.268 + ], + "angle": 0, + "content": "3Institut Universitaire de France" + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.252, + 0.778, + 0.269 + ], + "angle": 0, + "content": "4Université Côte d'Azur, Inria, France" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.269, + 0.806, + 0.287 + ], + "angle": 0, + "content": "5Université Côte d'Azur, CNRS, BCL, France Sorbonne Université, GRIPIC" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.287, + 0.651, + 0.304 + ], + "angle": 0, + "content": "7Université Sorbonne Paris Nord, LabSIC" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.308, + 0.612, + 0.321 + ], + "angle": 0, + "content": "julie.tores@univ-cotedazur.fr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.313, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.601 + ], + "angle": 0, + "content": "In film gender studies, the concept of \"male gaze\" refers to the way the characters are portrayed on-screen as objects of desire rather than subjects. In this article, we introduce a novel video-interpretation task, to detect character objectification in films. The purpose is to reveal and quantify the usage of complex temporal patterns operated in cinema to produce the cognitive perception of objectification. We introduce the ObyGaze12 dataset, made of 1914 movie clips densely annotated by experts for objectification concepts identified in film studies and psychology. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. Our new dataset and code are made available to the community." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.633, + 0.21, + 0.648 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.824 + ], + "angle": 0, + "content": "In film gender studies, the concept of \"male gaze\" [41] refers to the way the characters – especially women – are portrayed on-screen as objects of desire rather than subjects. Consider in Figure 1 how objectification is manifested in various ways such as camera placement and movement, the gaze interactions between characters, the choice of clothing, and arrangement of scene elements. Such disparities in how people are presented, depicted or addressed to in digital contents based on their gender has large-scale social implications such as the perpetuation of harmful stereotypes and hostile social situations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "These disparities have been the subject of an increasing number of studies at the intersection of social and computational sciences. In online social networks, computational approaches to sexism detection have been increasingly investigated for textual data as a part of hate speech detection." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.362, + 0.625, + 0.373 + ], + "angle": 0, + "content": "A Unreturned gaze" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.373, + 0.618, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.411, + 0.665, + 0.421 + ], + "angle": 0, + "content": "Direction of gaze" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.373, + 0.683, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.373, + 0.797, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.373, + 0.878, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.411, + 0.848, + 0.421 + ], + "angle": 0, + "content": "Direction of gaze" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.429, + 0.643, + 0.449 + ], + "angle": 0, + "content": "\\(\\bullet\\) Submissive posture, lower-body nudity" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.455, + 0.613, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.659, + 0.428, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Shot - woman in parallel with animal" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.44, + 0.793, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.44, + 0.867, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.47, + 0.792, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.47, + 0.867, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.503, + 0.663, + 0.525 + ], + "angle": 0, + "content": "Clothing - transparent, look-up and down the" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.5, + 0.883, + 0.525 + ], + "angle": 0, + "content": "ent, close-up on hips, domestic activities, the body" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.526, + 0.635, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.559, + 0.622, + 0.576 + ], + "angle": 0, + "content": "full body \\(\\rightarrow\\) approach \\(\\rightarrow\\) close-up on hips" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.526, + 0.726, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.526, + 0.816, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.526, + 0.893, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.586, + 0.894, + 0.683 + ], + "angle": 0, + "content": "Figure 1. In modern film media, the unequal characterization of gender on screen frequently evokes concepts of objectification, such as (A) unequal gaze (Pulp Fiction, 1994), (B) Nudity and submissive postures (Pulp Fiction, 1994), (C) animalisation or infantilisation (Marley and Me, 2008), and (D) transparent clothing, camera framing, domestic gender roles, and voyeurism (Gone Girl, 2014)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "As explained by Samory et al. [47], sexism is a complex sociological construct, whose high-level interpretive nature and subtle dimensions beyond offensive speech make for an unsolved challenge. In visual media such as films and TV series, the characters they depict shape our collective imagination and perception of sociological constructs, such as gender, race, and class. Currently, most large-scale approaches to understanding gender representation in these media have focused on quantifying the presence of women in the image and audio content (e.g., [35, 39, 51]). However, works in social sciences show that quantifying the presence of gender on screen is insufficient for grasping the issue of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10864" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "gender inequalities in visual media. For films, the classic Bechdel test, although useful and simple, considers neither the visual modality, which is key to analyzing gender depiction [41], nor the textual constructs of speech and dialogue [49]. While a few works have investigated sexist memes [16, 19], sexist advertisement [18], and characterized the on-screen positioning and co-occurrence of certain groups with respect to scene types and objects (e.g., Wang et al. for still images [54], Jang et al. [26] for films), computational approaches to interpretive sexism in visual media remain very scarce." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.489 + ], + "angle": 0, + "content": "In this article, we introduce a new challenging task for computer vision: detecting character objectification in films. Owing to the importance of this question, we consider it is critical to support the design of explainable methods and fine-grained model error analysis, which we address by densely annotating video data for theory-driven concepts. This presents a major step to tackle the question of subtle sexism in videos, operationalizing the popularly known concept of male gaze with the construct of objectification, and specifically considering the temporal dimension where such video patterns unfold. The end goal is to enable large-scale quantification and characterization of complex patterns producing on-screen objectification, and unveil possible correlations along the lines of the gender or race constructs." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.509, + 0.26, + 0.522 + ], + "angle": 0, + "content": "Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.825 + ], + "angle": 0, + "content": "- We introduce a novel video-interpretation task to detect character objectification in films. This is an interpretive task, hence extending beyond the more classical yet still challenging video-understanding tasks, and involving a subjective judgement. In a team involving media studies experts, and building on results in cinematography and psychology, we design a thesaurus of visual objectification, defining coarse-grained concepts with exemplified instances. This thesaurus is then used to formulate precise annotation guidelines. We introduce the ObyGaze12 dataset to the community, with 1914 clips from 12 films densely annotated by experts for concepts of objectification, including hard negative examples. It corresponds to \\(25\\%\\) of the MovieGraphs dataset. We verify the consistency of the obtained data, and provide first analyses showing the compositional nature of objectification. The dataset is meant to explore the complex temporal patterns producing character objectification in films. To the best of our knowledge, it is the first work proposing a computational approach to this interpretive task in videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "- We verify that the new task of objectification detection in videos is accessible by testing recent vision and vision-language models, and that hard negative examples improve classification. We also investigate the model weaknesses in representing each objectification concept. To do so, a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.153 + ], + "angle": 0, + "content": "thorough analysis is carried out with Concept Bottleneck Models (CBMs) which allows us to identify the individual concepts that pose the greatest challenge, namely: Type of Shot, Look, Posture, and Appearance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.895, + 0.243 + ], + "angle": 0, + "content": "To the best of our knowledge, this is one of the few video datasets with dense concept-based annotations for a high-level construct, and the first for objectification. The dataset and code used in this article is entirely provided in https://github.com/husky-helen/ObyGaze12." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.895, + 0.379 + ], + "angle": 0, + "content": "The article is organized as follows. In Sec. 2, we first provide a review of the relevant works on visual biases in films, dataset creation, and models for video understanding. We then introduce dataset creation for ObyGaze12 and present first analyses in Sec. 3. Sec. 4 presents the evaluation of models on the new task, and the analysis of the difficulty of concept representation with CBMs. Finally, we provide discussions on ethical aspects and challenges in Sec. 5, and conclusions in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.646, + 0.407 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.614 + ], + "angle": 0, + "content": "In this section, we position our contributions with respect to the relevant existing work. First, we introduce biases in visual datasets and computational approaches to the analysis of visual gender representation in films. We then discuss interpretive-level tasks, increasingly common in natural language processing, and the approaches to dataset creation to instantiate them for ML approaches. Here we highlight the scarcity of visual datasets made for high-level interpretive tasks, in particular for video data. Finally, we introduce the video understanding approaches that we consider to benchmark on our new interpretive task, specifically focusing on explainable concept-based approaches to locate the challenges ahead in video interpretation for this new task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.757, + 0.637 + ], + "angle": 0, + "content": "2.1. Visual biases in film datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The task we introduce is connected to the general problem of bias detection. As exposed by Fabbrizzi et al. [15], biases in visual datasets can be classified into selection bias (how subjects are included in a dataset), framing bias (how the visual content has been artificially composed) and label bias (errors or disparities in the labelling data). Our contribution is closely related to the framing and labelling biases. In film datasets, the first studies of biases in gender representation were from a presence quantification perspective. Guha et al. [21] and Somandepalli et al. [51] automatically estimate the screen time (from video) and speaking time (from audio) of male and female characters in Hollywood movies. They show that women are seen (36% screen time) and heard (41% screen time) significantly less than male characters. Mazieres et al. [39] consider a movie dataset spanning three decades, and show a temporal trend towards general fairer representation between both binary genders." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10865" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.212 + ], + "angle": 0, + "content": "They however show simultaneously that the applied framings remain unfair, with only \\(40\\%\\) of one-face frames featuring a female (60% for males). Jang et al. specifically analyze the qualitative framing differences in gender portrayal in a dataset of 20 Hollywood movies and 20 Korean movies [26]. They find that female characters are portrayed with lower emotional diversity, spatial occupancy, temporal occupancy, intellectual image or mean age." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.468, + 0.318 + ], + "angle": 0, + "content": "In contrast in this article, we take a first step towards detecting bias in gender representation from a high-level construct, objectification, qualitatively described in various disciplines such as cinematography [9, 41], social psychology [1, 32], and neuroscience [5-7]. Objectification is produced by complex temporal patterns never analyzed computationally in videos until now." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.449, + 0.348 + ], + "angle": 0, + "content": "2.2. Interpretive-level tasks and dataset creation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.468, + 0.718 + ], + "angle": 0, + "content": "At the same time that biases in visual datasets are uncovered and analyzed, other approaches aim to detect bias in human data. Detecting highly interpretive constructs, such as hate speech, propaganda, sexism, and racism, has been a long-standing endeavor in NLP. Da San Martino et al. [12] recruited to 4 experts to annotate news articles with text spans associated to 18 possible propaganda techniques. Samory et al. uncover the challenge of construct complexity for sexism detection [47]. They observe that multiple articles for automating this task consider widely different definitions of sexism, often referring to sub-dimensions of the broader construct. They consider existing works in social psychology where sexism dimensions have long been operationalized with sub-scale questions tested for consistency. To approach dataset creation for on-screen objectification, we inspire from these two last works, and on dense annotation approaches of image datasets recently proposed for the medical domain [13]. To the best of our knowledge, approaches for sexism detection from visual content are scarce, and almost none existing for video data. Two main types of visual content have been considered so far: hateful or sexist memes [16, 29], and sexist advertisements [18] which can also relate to symbolic advertisement understanding [25, 28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In films, analysis of biases in gender representation has been also automated with NLP approaches applied to film scripts [2, 37]. Specifically, Martinez et al. [37] propose a RNN-based model to automatically extract agent-verbal-patient triplets. From 912 movie scripts, they show that male characters are associated with a higher agency while female characters are more frequently the object of gaze. Su et al. introduce the more abstract task of trope understanding in movies [52]. Our work is close to this last one as we also introduce a dataset of films annotated for a higher-level construct beyond event and story understanding. However, we provide dense annotations of sequences with constituency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "tive concepts for the high-level construct of character objectification. Also, while Su et al. exploit an existing online base contributed to by the community, we design and carry out a strictly defined annotation process by experts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.47 + ], + "angle": 0, + "content": "In this article, we set out from the concept of male gaze defined in various ways in film gender studies [9, 36, 41]. Mulvey [41] characterizes the concept of gaze by the three relations between the camera and the characters, between the characters, but also between the spectator and the characters. While some formalizations of gaze set out from the gender of the director [36], Brey defines female and male gaze only from the film content, with a corpus analysis focusing on aesthetics [9] and revolving around the construct of objectification. We consider this gaze analysis of Brey, which does not significantly rely the socio-historical context of production and reception. We build the operationalization objectification in both cinematography and psychology, to create the conditions to make a new challenge accessible to the CV community: we produce a strict annotation process to obtain densely annotated video data and analyze where the new challenges lie ahead. We generally position our approach producing a non-large scale but high-quality dataset that we hope will be useful within the lines of the call of Paullada et al. for such data-centric AI approaches [44]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.5 + ], + "angle": 0, + "content": "2.3. Approaches to video and movie understanding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Cross-modal foundation models, such as CLIP [45] and ALIGN [27], learn aligned image and text representations through contrastive pre-training on large-scale closed datasets. Generalizations of the CLIP model to video data have included VideoCLIP [57] and two X-CLIP models [34, 42]. In particular, X-CLIP [42], which we employ in this article, expands CLIP with video temporal modeling and video-adaptive textual prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Legacy and large-scale pre-trained vision-language models have been leveraged for movie-related tasks. Bose et al. consider the difficulty of visual scene recognition in movies due to domain mismatch and create the MovieCLIP dataset obtained from weakly labeling movie shots from scene categories using the CLIP model [8]. An important vision-language movie-related task is audio-description for the visually-impaired, for which a major dataset introduced recently is MAD [50], gathering sparse natural language sentences grounded in over 1200 hours of movie videos. AutoAD [23] and AutoADII [22] are two recent approaches to generate audio-description from the video, both leveraging CLIP to learn to prompt GPT. To design approaches to learn human-level constructs, such as emotions, interactions or relationships, datasets generated with human supervision are also instrumental. A prominent representative is the MovieGraphs dataset, providing detailed annotations of clips of 51 movies with emotional states, character interac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10866" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "tions and relationships, and other scene reasoning elements [53]. It has prompted works tackling such recognition tasks [17, 33]. In this article, we build on the MovieGraphs dataset to annotate a selection with the construct of objectification, to later analyze it in connection with the other annotated social elements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.424 + ], + "angle": 0, + "content": "In this article, we aim to assess the capacity of CLIP-based methods to provide relevant embeddings for the concept of objectification. We do so with a direct evaluation of classification results when an adapter (MLP) fed by X-CLIP embeddings is learnt. We analyze the results with a concept-based approach by building on Concept Bottle-neck Models (CBM) [31]. Concept-based models are an active area of research in XAI, with works tackling the accuracy-explainability tradeoff [60] and the need for user-defined concepts [58]. We specifically employ Post-hoc CBM (PCBM) [59], which consists in learning a concept subspace (made of Concept Activation Vectors [30]) in the embedding space of the pre-trained model. Data samples are then projected in this concept subspace, from where the classification task can be performed with an interpretable classifier." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.443, + 0.255, + 0.459 + ], + "angle": 0, + "content": "3. Data and methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.53 + ], + "angle": 0, + "content": "This section presents our approach to create the first dataset for visual objectification in videos, specifically in films. We name this dataset ObyGaze12, short for ObjectifyingGaze12, which has the following highlights:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.469, + 0.605 + ], + "angle": 0, + "content": "- It considers the multiple dimensions of the construct of visual objectification, made of filmic (framing and editing over successive shots, camera motion, etc.) and iconographic properties (visible objects, body parts, attire, character interactions, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.606, + 0.468, + 0.666 + ], + "angle": 0, + "content": "- It is based on a thesaurus articulating five sub-constructs identified from multidisciplinary literature (film studies and psychology) from which we define typical instances, then grouped into coarse-grained visual concepts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.667, + 0.468, + 0.711 + ], + "angle": 0, + "content": "- The data is annotated densely with concepts, and shows the multi-factorial property of objectification, corroborating with some recent developments in cognitive psychology [6]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.468, + 0.755 + ], + "angle": 0, + "content": "- A hard negative category is included with the goal to perform fine-grained error analysis and improve model generalization." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.531, + 0.469, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.338, + 0.787 + ], + "angle": 0, + "content": "3.1. A thesaurus of objectification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first formalize the construct of visual objectification and derive key concepts to annotate in film scenes. Together with media studies experts, we identify five sub-constructs of objectification from literature on film cognition and film gender studies, and social and cognitive psychology: male gaze (point of view of a man on a woman) [5, 9, 41], sexualization [6, 7], surveillance of the feminine body [11, 14, 40]," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "female inaction / male possession [20, 48], and infantilism / animalization [41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.892, + 0.379 + ], + "angle": 0, + "content": "These sub-constructs come with typical instances and examples from filmmaking techniques ([9, 41]) or validated questionnaires ([11, 40]), as shown in the middle and rightmost columns in Table 1. These typical instances are then grouped into eight coarse-grained visual concepts, corresponding to the possible means of production of visual objectification. They are shown in the left-most column in Table 1: type of shot (framing and gaze of camera), look (gaze of characters on the other), body (partial or full nudity, and sexually suggestive body parts), posture (connoting, e.g., childhood, submission or inaction), clothing (in relation to context and activities), appearance (age and makeup), expression of emotion (restrained or exaggerated according to gender role), and activities (linked to gender roles). Visual examples are provided in Fig. 1, where we show video samples of objectification concepts Look, Posture, Type of shot, Clothes and Activity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.644, + 0.404 + ], + "angle": 0, + "content": "3.2. Data selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Over the various existing movie datasets (see Sec. 2 and [4, 24, 39, 43, 46, 50, 53], [51, Table 1]), many have overlapping titles and only a few have rich human supervision. Amongst these, the MovieGraphs dataset [53] includes rich, high-level human annotations of 7637 clips of 51 movies, with emotional states, interactions and relationships, and other social reasoning elements. These elements are important and valuable in exploring social concepts of objectification in visual media. These movies are also frequently adopted for media research in existing work in [4, 24, 39]. The movie clips have also a short duration – mostly within 1-5 minutes – facilitating dense and granular annotations (over 100 clips for an average 2 hour film) while preserving the possibility to observe longer-term interactions and story development across a number of shots. From the 7637 clips of the MovieGraphs dataset, we select 1914 clips to annotate for objectification, amounting to \\(25\\%\\) of the dataset and 12 complete movies, which were selected to approximately reproduce the fraction of genres in the original dataset. The list of selected films, year of release, and genre can be found in Table 3 in Appendix 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.661, + 0.755 + ], + "angle": 0, + "content": "3.3. Data annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Every selected movie is annotated by at least two experts for objectification level and concepts over the movie scenes. Specifically, the annotators were asked to repeat a three-step process for every scene they deemed interesting from an objectification perspective: (1) watch the movie entirely and when they identify a scene worth annotating, (2) delimit the clip by using the cutting function in our annotation tool, and (3) assigning an objectification level and annotate the concept(s) involved in the objectification rating. We define" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10867" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Thesaurus of the typical instances and examples of visual objectification in films, grouped into eight main visual concepts used for annotation. Examples are possible means to produce one of the five sub-constructs of objectification (male gaze, sexualisation, surveillance of the feminine body, female inaction/male possession, infantilism/animalisation), to be assessed by annotators." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.143, + 0.89, + 0.465 + ], + "angle": 0, + "content": "
ConceptConcept instancesExamples
Type of shotShot suggesting man perspective in presence of womanclose-up on a man's face; body parts of woman
Shot suggesting man gaze on womancamera takes the perspective of a male character with first close-up on the face followed by camera motion looking a woman from bottom up
Shot showing a woman in parallel with an animalwoman at same level and position with a dog
LookVoyerismcharacter watching another one without their knowledge
Non-reciprocal gazewoman looking at man who does not look back
BodySuggested nudityclothing on floor; silhouette behind shower curtain; nude shadow on wall
Partial nuditynude upper or lower body; partially open clothing or draping; in underwear
Full nuditynude person fully or partially shown
Body parts suggestive of sexclose-up shots on breast, buttocks, hips, or lips
PostureGesture or posture connoting seductionlip-biting; hip roll; twisting or tucking hair
Gesture or posture connoting sexualityeating phallic symbols; arching back
Gesture or posture connoting inactionbeing undressed by someone
Gesture or posture connoting submissionleaning on a man
Gesture or posture connoting dependencefollowing a man
Skippingskipping gait
ClothingWet or transparent clothingthin shirt soaked in rain
Clothing impractical to situationwear pumps for running, a skirt when gardening
Color code associated to characterwoman with pink clothing and accessories
Older woman wearing infantile clothingwoman wearing an Alice band or high socks
AppearanceDiscrepancy between appearance of woman and context or biographical elementsperfect makeup when waking up; mother of heroine too young; young girl played by older actress
Exp. of emotionAsymmetric expression of emotionboys don't cry; woman being hysterical
ActivityDoing domestic activitiesdoing laundry, cooking, cleaning, being constantly in the kitchen
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.273, + 0.502 + ], + "angle": 0, + "content": "four levels of objectification:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.504, + 0.47, + 0.548 + ], + "angle": 0, + "content": "- Easy Negative: there are no elements suggestive of objectification. No concept can be annotated. Default value for watched but non-selected scenes." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.549, + 0.47, + 0.594 + ], + "angle": 0, + "content": "- Hard Negative: the scene contains elements of objectification from the thesaurus, but their presence does not result in an objectification effect." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.595, + 0.47, + 0.625 + ], + "angle": 0, + "content": "- Not Sure: the scene indicates objectification, but does not completely fit the definition in the thesaurus." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.625, + 0.47, + 0.669 + ], + "angle": 0, + "content": "- Sure: the scene contains elements of objectification from the thesaurus, and their presence results in an objectification effect." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.504, + 0.47, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Four expert annotators were recruited to annotate the dataset. A first presentation session was held to introduce our annotation tool and the annotation procedure. All four annotators were then given the same two films – Juno and Silver Linings Playbook – to annotate using the proposed methodology. A second meeting was then set up after the annotation of the two films to analyze the reasons for divergence and remedy them by identify which elements of the annotation guidelines to clarify and how. We then randomly assigned two annotators to each of the remaining 10 films to annotate separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Data processing and fusion The data processing is described in detail in Appendix 7. Following the annotation step, the annotations are then projected from the delimi-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.713 + ], + "angle": 0, + "content": "tations provided by each annotator onto the delimitations of the MovieGraphs clips. Since multiple annotations may overlap the same MovieGraphs clip, the annotation that is projected on the MovieGraphs clip corresponds to the annotation (including objectification level and concepts) with (1) the highest level of objectification that has at least \\(20\\%\\) overlap with the MovieGraphs clip, and (2) when multiple annotations exist at the same level of objectification, the annotated concepts for these annotations are aggregated. The same process is used to aggregate the annotations of the annotators of a same clip: the maximum objectification level is kept, with possible aggregation of concepts in case both annotators chose the same level but annotated different concepts. The merged data is shared and used in the remaining of this article." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.726, + 0.792, + 0.741 + ], + "angle": 0, + "content": "3.4. Analysis of the ObyGaze12 dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Here we comment on some interesting statistics of the resulting annotations and concepts of the 1914 clips originally delimited in the MovieGraphs dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "First, we verify data consistency by computing the inter-annotator agreement (IAA). Given the task of annotating timespans, we choose the \\(\\gamma\\) agreement measure introduced in [38] (and used for, e.g., annotating text spans [12]) owing to its consideration of temporal alignment, multiple annotators, and label classification at the same time. It attributes a score between 1 (complete agreement) and \\(-\\infty\\). A value of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10868" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "\\(\\gamma \\leq 0\\) indicates no agreement. The computation details of the \\(\\gamma\\) metric is provided in Appendix 7. Considering all four categories EN, HN, NS, S, we obtain and average \\(\\gamma = 0.42\\). Not considering the clips annotated Not Sure (NS), which is the uncertain and \"noisy\" class in human annotations, the IAA increases to \\(\\gamma = 0.69\\). This shows the consistency of the obtained annotations despite the interpretive nature of the task. Let us also mention that recent works improve learning approaches by considering explicitly the IAA in case of low number of annotators with moderate agreement [10, 55, 56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.515 + ], + "angle": 0, + "content": "Second, we analyze the obtained annotations in Fig 2. The Sure category is the least represented with \\(16\\%\\), the Easy Negative being, as expected, the most represented class with \\(52\\%\\) of clips. It is interesting to note that every concept is approximately annotated with the same rate throughout the Hard Negative, Not Sure and Sure levels of objectification. Finally, it is very interesting to observe that the average number of concepts annotated per clip increases with the level of objectification: 1.26 concepts on average per Hard Negative clip, 1.71 for Not Sure, up to 2.6 for Sure. We verify that this trend is observable for every single annotator. It gives an important insight into our video interpretation data: that objectification is a compositional process. This corroborates with recent findings in neuroscience experiments that found that a single element, such as clothing on its own, is not sufficient for people to perceive a character as an object [6]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.527, + 0.21, + 0.544 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.599 + ], + "angle": 0, + "content": "The experiments have two objectives: to verify that the new classification task is feasible, and to identify the challenges of designing efficient models. To tackle these objectives, we" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.62, + 0.466, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.827, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Figure 2. Distribution of visual factors annotated for each level of objectification (HN = Hard negative, NS = Not sure, S = Sure). The percentage of the dataset for each level of objectification as well as the average number of concepts per clip are also shown. (Best viewed in colors)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "consider pre-trained vision models and specifically address the following research questions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.183 + ], + "angle": 0, + "content": "- Task accuracy - What are the baseline performances by pre-trained vision models on the objectification detection task? How does the performance vary with hard negative examples?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.183, + 0.894, + 0.258 + ], + "angle": 0, + "content": "- Concept representation - Can we implement interpretable models of objectification using concepts? What is the quality of representation of every concept, and what are the objectification concepts poorly captured by current models?" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.123, + 0.894, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.271, + 0.646, + 0.287 + ], + "angle": 0, + "content": "4.1. Task accuracy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.491 + ], + "angle": 0, + "content": "Setup We discard the Not Sure (NS) class from the ObyGaze12 dataset, as it gathers by definition samples highly uncertain for humans, and consider the Easy Negative (EN, \\(62\\%\\) of the clip samples), Hard Negative (HN, \\(19\\%\\)) and Sure (S, \\(19\\%\\)) classes. We approach binary classification in a progressive way, the positive class being made of the S samples. We consider two levels of classification difficulty by composing the negative class either with EN samples, or with HN samples. The implementation details of cross-validation and data balancing are provided in Appendix 8. The average performance over the test set of the best models on validation folds are shown in Table 2 with standard deviations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Baselines We consider video embeddings obtained from pre-trained models owing to their zero-shot classification capabilities on video tasks. We select ViViT-B/16 [3], and the available X-CLIP model, trained on Kinetics [42]. We also re-train a X-CLIP model [34] on the LSMDC [46] film dataset, and refer to Appendix 10 for corresponding results, where all implementation details are described. We keep the pre-trained models frozen and perform an adaptative max pooling of the resulting frame tokens, and feed the output to an MLP made of 2 dense layers, the hidden layer with 128 neurons and ReLU activations, the last with 2 softmax neurons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results To assess the quality of the models on possibly imbalanced data with a minority of positive samples, we report the F1-scores in Table 2. First, by comparing with trivial classifiers (random and an all-positive, see App. 8), we observe that the task is indeed feasible, warranted by the data consistency described in Sec. 3 despite the interpretive nature of the task. Second, we observe that the inclusion of Hard Negative examples improves the classification results, showing the importance of a fine-grained annotation for highly-interpretive tasks. Results of X-CLIP on other configurations, specifically when the movies of clips in test are different from those in train, are shown in App. 8. The best results based on existing models are of moderate qual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10869" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 2. F1-score on the binary task of objectification detection for models trained with easy or with hard negatives and tested on easy or all negative samples, with standard deviations." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.142, + 0.472, + 0.26 + ], + "angle": 0, + "content": "
Test TrainEN vs. S(EN U HN) vs. S
EN vs. SHN vs. SEN vs. SHN vs. S
ViViT-B/160.53 (0.18)0.62 (0.13)0.54 (0.24)0.73 (0.1)
X-CLIP0.79 (0.05)0.71 (0.05)0.66 (0.05)0.82 (0.03)
Random0.540.55
All positive0.080.06
PCBM-DT0.680.440.580.38
PCBM-LR0.640.430.500.37
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.47, + 0.318 + ], + "angle": 0, + "content": "ity, which calls for more investigation into where the difficulties lie." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.328, + 0.251, + 0.344 + ], + "angle": 0, + "content": "4.2. Concept accuracy" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.351, + 0.473, + 0.715 + ], + "angle": 0, + "content": "To infer on-screen objectification, it is key for the model to detect the means of its production, which correspond to the eight concepts listed in Table 1. We reiterate that in the ObyGaze12 dataset, every clip annotated with a level of objectification S, NS or HN is also annotated with the presence of instances of the eight concepts. For example, if the Body concept is annotated, it means that some level of nudity and/or suggestive body parts are shown on screen, that could contribute to the production of objectification. The means of producing objectification through the eight concepts can be subtle to detect, making it difficult to provide the final interpretation. To investigate this difficulty, we implement Post-hoc Concept Bottleneck Models (PCBMs) [59], which allow us to approach a classification task with pre-trained models in an interpretable way when concept-annotated data is available. In our case, from the X-CLIP embedding space where our video clips are represented, we identify a Concept Activation Vector (CAV) [30] for every concept. We then project the X-CLIP embedding of every clip onto the subspace defined by the eight CAVs. The representation of the clip that is the output of this bottleneck is a low-dimensional vector with number-of-concepts components. This vector can then be fed to an interpretable classifier for the objectification detection task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "CAV computation For every concept \\( i \\), we collect two sets of samples: positive samples where concept \\( i \\) is present, hence made of S and HN samples with the concept annotated as present, and negative with EN, S and HN without concept \\( i \\). We then train a linear SVM for each concept \\( i \\), the CAV of concept \\( i \\) being the normal vector of the SVM hyperplan. To train the SVM, we split the data in 10 folds, and reserve the last fold for test. We then perform an 8-fold cross-validation to select the SVM (choice of margin tolerance \\( c \\)), every fold training set being balanced with different draws of negative sub-sampling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Interpretable classifier We then train a decision tree (DT) and a logistic regression (LR) classifier on the same 8-fold cross-validation to classify the level of objectification, the classifiers being fed with the projection of every clip onto the CAVs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.893, + 0.383 + ], + "angle": 0, + "content": "Task accuracy with PCBM We first verify the quality of objectification detection with F1-scores of PCBM-DT and PCBM-LR shown in Table 2. The results lower than X-CLIP are expected owing to the known accuracy-interpretability tradeoff of CBMs [59, 60]. They are above random and all-positive predictions when training on EN vs. S, which is indicative of the relevance of information held in the concepts. However, the low results obtained when training the DT and LR to distinguish between S and HN reveals the low quality of some CAVs, where the X-CLIP embeddings cannot be linearly well-separated for these concepts. We investigate this point next. The resulting DT is discussed in App. 9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.763 + ], + "angle": 0, + "content": "Concept accuracy We now analyze the quality of each obtained CAV by plotting its capability to classify whether the concept is present in a test sample. We consider a positive similarity between the X-CLIP embedding and CAV of concept \\( i \\) indicative of the presence of concept \\( i \\). F1-score on the test set are shown in Fig. 3. Plots correspond to CAVs obtained from classifying the presence of concept \\( i \\) against EN only (solid bars) and against EN with S and HN without concept \\( i \\) (hatched bars). The former is used for PCBM-DT in Table 2. We first observe that concept detection is harder when negative samples also include S and HN samples (without the concept). This is expected considering that scenes tagged EN have by definition no element possibly conducive to objectification, and are hence likely to differ visually more from scenes where the concept is present, than do S and HN scenes without the concept. However within S and HN clips with and without the concept, such shortcuts cannot be exploited anymore. We observe in this case that the X-CLIP embedding related to concepts Type of shot, Posture, Look and Appearance are harder to separate linearly. This can be correlated with the analysis of factors of error detailed next. These subtler means of on-screen objectification therefore warrant future work to be properly captured and detected." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Error Analysis We analyze the factors impacting objectification classification errors corresponding to the results shown in Table 2. We select an average X-CLIP-based adaptation model trained on HN vs. S and tested on (ENUHN) vs. S, and consider its predictions on the clips in the test set. We label each test clip with 0 if the model fails to predict the correct label of the clip, and with 1 otherwise. We describe the clip with a one-hot encoding vector" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10870" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.093, + 0.46, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.422 + ], + "angle": 0, + "content": "Figure 3. For every concept, F1-score of the best linear SVM selected to define the CAV of this concept. Positive samples (S and HN with the concept) must be separated from: [non-hatched bars] negative samples made of EN only, or [hatched bars] negative samples made of EN and S and HN without the concept." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.473, + 0.814 + ], + "angle": 0, + "content": "corresponding to all 11 factors shown the y-axis of Fig. 4: every of the 8 concepts, and the Sure (S), Hard negative (HN) and Easy negative (EN) labels. We then train a logistic regression model on these clip descriptors to predict the failure/success labels. Fig. 4 shows the error regression weights associated with each factor, for two film examples. A negative weight indicates a contribution of the factor to a classification failure. We first observe that the HN characteristic contributes to a classification failure of the model, while S and EN contribute to classification success. Second, we observe that there is variability over the movies on the presence of which concept strongly influences the classification success. However, the presence of the Clothing concept seems to be a strong confounder. This can be due to the frequency of appearance of this concept in HN samples, and to the subtlety of the description of this concept (provided in Table 1), which should make it difficult for a pre-trained model to discriminate between an objectifying and non-objectifying overall label on the basis of Clothing. Third, it is worth noting that the concepts shown to be poorly linearly separable when described with the X-CLIP embeddings (see CAV analysis in Sec. 4.2), are also those with a non-stable contribution to the model errors over the film examples: Type of shot, Look, Posture and Appearance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.192, + 0.845 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Ethical aspect This work has an explicit societal motivation in its purpose to tackle, with the help of AI, the analysis of complex temporal patterns operated in cinema that pro" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.895, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.333 + ], + "angle": 0, + "content": "Figure 4. Analysis of the factors of error for the objectification detection task: weights of a logistic regressor predicting whether the test set examples are well classified or not. Positive (resp. negative) weights indicate a positive (resp. negative) contribution to classification success. Left: As good as it gets, Right: Up in the Air." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.424 + ], + "angle": 0, + "content": "duce the perception of certain characters as objects. This is a challenging but valuable task that aims to uncover and quantify differences in how various identities may be portrayed on screen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.603 + ], + "angle": 0, + "content": "Limitations and challenges A distinctive element of our work is the subjective judgement involved in annotating granular video elements for objectification. Video annotation is tedious, and approaching data annotation for such an interpretive task in a rigorous way is even more so, and difficult to scale. We therefore believe that pursuing high-quality, dense annotations with well-defined concepts goes a long way to tackle this new video interpretation task, which represents a valuable new challenge for the computer vision community." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.62, + 0.636 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.894, + 0.813 + ], + "angle": 0, + "content": "In this article, we have introduced a new video interpretation task to detect character objectification in films. We have introduced the ObyGaze12 dataset, densely annotated by experts for objectification concepts defined from five sub-constructs identified in film studies and psychology. ObyGaze12 is made available to the community. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. We show that the representation learning of the concepts of Type of shot, Look, Posture and Appearance need to be improved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.832, + 0.895, + 0.888 + ], + "angle": 0, + "content": "Acknowledgements. This work has been partly supported by the French National Research Agency through the ANR TRACTIVE project ANR-21-CE38-0012-01. This work was partly supported by EU Horizon 2020 project AI4Media, under contract no. 951911 (https://ai4media.eu/)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10871" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Operationalizing self-objectification: Assessment and related methodological issues. In Self-objectification in women: Causes, consequences, and counteractions., pages 23-49. American Psychological Association, Washington, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.297 + ], + "angle": 0, + "content": "[2] Apoorv Agarwal, Jiehan Zheng, Shruti Kamath, Sriramkumar Balasubramanian, and Shirin Ann Dey. Key Female Characters in Film Have More to Talk About Besides Men: Automating the Bechdel Test. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 830-840, Denver, Colorado, 2015. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, Montreal, QC, Canada, 2021. IEEE. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed Movies: Story Based Retrieval with Contextual Embeddings. Technical Report arXiv:2005.04208, arXiv, 2020. arXiv:2005.04208 [cs] type: article. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.52 + ], + "angle": 0, + "content": "[5] Philippe Bernard, Sarah J. Gervais, and Olivier Klein. Objectifying objectification: When and why people are cognitively reduced to their parts akin to objects. European Review of Social Psychology, 29(1):82-121, 2018. Publisher: Routledge _eprint: https://doi.org/10.1080/10463283.2018.1471949. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[6] Philippe Bernard, Florence Hanoteau, Sarah Gervais, Lara Servais, Irene Bertolone, Paul Deltenre, and Cecile Colin. Revealing Clothing Does Not Make the Object: ERP Evidences That Cognitive Objectification is Driven by Posture Suggestiveness, Not by Revealing Clothing. Personality and Social Psychology Bulletin, 45(1):16-36, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[7] Philippe Bernard, Carlotta Cogoni, and Andrea Carnaghi. The Sexualization-Objectification Link: Sexualization Affects the Way People See and Feel Toward Others. Current Directions in Psychological Science, 29(2):134-139, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[8] Digbalay Bose, Rajat Hebbar, Krishna Somandepalli, Haoyang Zhang, Yin Cui, Kree Cole-McLaughlin, Huisheng Wang, and Shrikanth Narayanan. Movieclip: Visual scene recognition in movies. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2082-2091, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[9] Iris Brey. Le regard féminin-Une révolution à l'écran. Média Diffusion, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[10] M. Bucarelli, L. Cassano, F. Siciliano, A. Mantrach, and F. Silvestri. Leveraging inter-rater agreement for classification in the presence of noisy labels. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3439-3448, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Rachel M. Calogero. A Test of Objectification Theory: The Effect of the Male Gaze on Appearance Concerns in Col-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "lege Women. Psychology of Women Quarterly, 28(1):16-21, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[12] Giovanni Da San Martino, Seunghak Yu, Alberto Barrón-Cedeno, Rostislav Petrov, and Preslav Nakov. Fine-Grained Analysis of Propaganda in News Article. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5635–5645, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[13] Roxana Daneshjou, Mert Yuksekgonul, Zhuo Ran Cai, Roberto A. Novoa, and James Zou. Skincon: A skin disease dataset densely annotated by domain experts for fine-grained debugging and analysis. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[14] Angela Denchik. Development and Psychometric Evaluation of the Interpersonal Sexual Objectification Scale. PhD thesis, The Ohio State University, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[15] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223: 103552, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[16] Elisabetta Fersini, Francesca Gasparini, and Silvia Corchs. Detecting Sexist MEME On The Web: A Study on Textual and Visual Cues. In 2019 8th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIw), pages 226-231, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[17] Bei Gan, Xiujun Shu, Ruizhi Qiao, Haoqian Wu, Keyun Chen, Hanjun Li, and Bohan Ren. Collaborative noisy label cleaner: Learning scene-aware trailers for multi-modal highlight detection in movies. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18898-18907, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[18] Francesca Gasparini, Ilaria Erba, Elisabetta Fersini, and Silvia Corchs. Multimodal Classification of Sexist Advertisements.: In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications, pages 399-406, Porto, Portugal, 2018. SCITEPRESS - Science and Technology Publications. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[19] Francesca Gasparini, Giulia Rizzi, Aurora Saibene, and Elisabetta Fersini. Benchmark dataset of memes with text transcriptions for automatic detection of multi-modal misogynistic content. arXiv:2106.08409 [cs], 2021. arXiv:2106.08409.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[20] Sarah J. Gervais, Gemma Sáez, Abigail R. Riemer, and Olivier Klein. The Social Interaction Model of Objectification: A process model of goal-based objectifying exchanges between men and women. British Journal of Social Psychology, 59(1):248-283, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[21] Tanaya Guha, Che-Wei Huang, Naveen Kumar, Yan Zhu, and Shrikanth S. Narayanan. Gender Representation in Cinematic Content: A Multimodal Approach. In Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, pages 31–34, Seattle Washington USA, 2015. ACM. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel - who," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10872" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13645-13655, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie Description in Context. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18930-18940, Vancouver, BC, Canada, 2023. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[24] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. MovieNet: A Holistic Dataset for Movie Understanding. In Computer Vision – ECCV 2020, pages 709–727. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[25] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic Understanding of Image and Video Advertisements. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1100-1110, Honolulu, HI, 2017. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[26] Ji Yoon Jang, Sangyoon Lee, and Byungjoo Lee. Quantification of Gender Representation Bias in Commercial Films based on Image Analysis. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW):1-29, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[28] Nasrin Kalanat and Adriana Kovashka. Symbolic image detection using scene and knowledge graphs. Technical Report arXiv:2206.04863, arXiv, 2022. arXiv:2206.04863 [cs] type: article. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie J. Cai, James Wexler, Fernanda B. Viégas, and Rory Sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International Conference on Machine Learning, 2017. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In Proceedings of the 37th International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[32] Holly B. Kozee, Tracy L. Tylka, and Casey L. Augustus-Horvath. Interpersonal Sexual Objectification Scale, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[33] Anna Kukleva, Makarand Tapaswi, and Ivan Laptev. Learning Interactions and Relationships Between Movie Characters. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9846-9855, Seattle, WA, USA, 2020. IEEE. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, page 638–647, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[35] Sarah Macharia. Global Media Monitoring Project (GMMP), pages 1-6. John Wiley Sons, Ltd, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[36] Alicia Malone. The Female Gaze: Essential Movies Made by Women. Mango Publisher, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[37] Victor R. Martinez, Krishna Somandepalli, and Shrikanth Narayanan. Boys don't cry (or kiss or dance): A computational linguistic lens into gendered actions in film. PLOS ONE, 17(12):e0278604, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[38] Yann Mathet, Antoine Widlocher, and Jean-Philippe Métivier. The unified and holistic method gamma \\((\\gamma)\\) for inter-annotator agreement measure and alignment. Computational Linguistics, 41(3):437-479, 2015. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[39] Antoine Mazieres, Telmo Menezes, and Camille Roth. Computational appraisal of gender representativeness in popular movies. Humanities and Social Sciences Communications, 8 (1):137, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[40] Nita Mary McKinley and Janet Shibley Hyde. The objectified body consciousness scale: Development and validation. Psychology of women quarterly, 20(2):181-215, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[41] Laura Mulvey. Visual Pleasure and Narrative Cinema. Screen, 16(3):6-18, 1975. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[42] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.549, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. MovieCuts: A New Dataset and Benchmark for Cut Type Recognition. In Computer Vision – ECCV 2022, pages 668–685. Springer Nature Switzerland, Cham, 2022. Series Title: Lecture Notes in Computer Science. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[44] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[46] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie Description. International Journal of Computer Vision, 123(1):94-120, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[47] Mattia Samory, Indira Sen, Julian Kohne, Fabian Flock, and Claudia Wagner. \"Call me sexist, but...\" : Revisiting Sexism Detection Using Psychological Scales and Adversarial Samples. page 12. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10873" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[48] Maarten Sap, Marcella Cindy Prasettio, Ari Holtzman, Hannah Rashkin, and Yejin Choi. Connotation frames of power and agency in modern films. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 2329-2334, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[49] Alexandra Schofield and Leo Mehr. Gender-Distinguishing Features in Film Dialogue. In Proceedings of the Fifth Workshop on Computational Linguistics for Literature, pages 32–39, San Diego, California, USA, 2016. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[50] M. Soldan, A. Pardo, J. Alcazar, F. Heilbron, C. Zhao, S. Giancola, and B. Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5016-5025, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[51] Krishna Somandepalli, Tanaya Guha, Victor R. Martinez, Naveen Kumar, Hartwig Adam, and Shrikanth Narayanan. Computational Media Intelligence: Human-Centered Machine Analysis of Media. Proceedings of the IEEE, 109(5): 891-910, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.494 + ], + "angle": 0, + "content": "[52] Hung-Ting Su, Po-Wei Shen, Bing-Chen Tsai, Wen-Feng Cheng, Ke-Jyun Wang, and Winston H. Hsu. TrUMAn: Trope Understanding in Movies and Animations. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 4594–4603, New York, NY, USA, 2021. Association for Computing Machinery. event-place: Virtual Event, Queensland, Australia. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[53] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. MovieGraphs: Towards Understanding Human-Centric Situations from Videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[54] Angelina Wang, Arvind Narayanan, and Olga Russakovsky. REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets. In Computer Vision – ECCV 2020, pages 733–751. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[55] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. ArXiv, abs/2110.12088, 2021. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[56] Jiaheng Wei, Zhaowei Zhu, Tianyi Luo, Ehsan Amid, Abhishek Kumar, and Yang Liu. To Aggregate or Not? Learning with Separate Noisy Labels. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2523-2535, New York, NY, USA, 2023. Association for Computing Machinery. event-place: Long Beach, CA, USA. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[57] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19187-19197, Vancouver, BC, Canada, 2023. IEEE. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[59] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR^2Struct: Privacy, Accountability, Interpretability, Robustness, Reasoning on Structured Data, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.893, + 0.316 + ], + "angle": 0, + "content": "[60] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models. In Advances in Neural Information Processing Systems, 2022. 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10874" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_origin.pdf b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3b067c48799be8d54f0ab9c51d64bdf2b888dcdf --- /dev/null +++ b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/410cda22-2c59-4ec1-949f-7a68b660e50b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7beb31daca1697fc483cf53cd5d5dd76f3690250bf352b0744d737eee9b4d02a +size 3450881 diff --git a/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/full.md b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c48d1f6a42c832e08dbc6ddfaac5f27192e7fc3 --- /dev/null +++ b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/full.md @@ -0,0 +1,321 @@ +# Visual Objectification in Films: Towards a New AI Task for Video Interpretation + +Julie Tores1,2 + +Lucile Sassatelli $^{1,3}$ + +Hui-Yin $\mathbf{W}\mathbf{u}^{4}$ + +Clement Bergman + +Léa Andolfi + +Victor Ecrement6 + +Frédéric Precioso² + +Thierry Devars6 + +Magali Guaresi + +Virginie Julliard + +Sarah Lecossais + +1Université Côte d'Azur, CNRS, I3S, France + +2Université Côte d'Azur, CNRS, Inria, I3S, France + +3Institut Universitaire de France + +4Université Côte d'Azur, Inria, France + +5Université Côte d'Azur, CNRS, BCL, France Sorbonne Université, GRIPIC + +7Université Sorbonne Paris Nord, LabSIC + +julie.tores@univ-cotedazur.fr + +# Abstract + +In film gender studies, the concept of "male gaze" refers to the way the characters are portrayed on-screen as objects of desire rather than subjects. In this article, we introduce a novel video-interpretation task, to detect character objectification in films. The purpose is to reveal and quantify the usage of complex temporal patterns operated in cinema to produce the cognitive perception of objectification. We introduce the ObyGaze12 dataset, made of 1914 movie clips densely annotated by experts for objectification concepts identified in film studies and psychology. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. Our new dataset and code are made available to the community. + +# 1. Introduction + +In film gender studies, the concept of "male gaze" [41] refers to the way the characters – especially women – are portrayed on-screen as objects of desire rather than subjects. Consider in Figure 1 how objectification is manifested in various ways such as camera placement and movement, the gaze interactions between characters, the choice of clothing, and arrangement of scene elements. Such disparities in how people are presented, depicted or addressed to in digital contents based on their gender has large-scale social implications such as the perpetuation of harmful stereotypes and hostile social situations. + +These disparities have been the subject of an increasing number of studies at the intersection of social and computational sciences. In online social networks, computational approaches to sexism detection have been increasingly investigated for textual data as a part of hate speech detection. + +A Unreturned gaze + +![](images/6061d4d9ee8b90046c4a3e7fe37b2a5951c45445f4b706ae11dcf946a39c5278.jpg) +Direction of gaze + +![](images/b509430971965b5fbd75ef6728f306b89afc9bad029a49256216fcd022b52cfa.jpg) + +![](images/b4e87c2e2cbfd7bde9098f848f7a7045e4613dd8ce6d6e9e522acd7aa4b1be18.jpg) +Direction of gaze + +![](images/1c10f2e48be524ec11bd908c158d0280647d7dd7a4ba3816c8d438dbd370e78a.jpg) + +$\bullet$ Submissive posture, lower-body nudity + +![](images/0942612c1d711246e889d89dd76933f801d0a1cc400b3a0a7dc99ef90b8223fc.jpg) + +Shot - woman in parallel with animal + +![](images/4e4882fc50ebcfeff8a96f46599ec3b724f4846777251e801f49e1ee686697b3.jpg) + +![](images/fd901d5686ff13217d2090174bbccabaa78c96d75fed3a9582852dac6b59134c.jpg) + +![](images/e9d0dd0688b11567d69de0d49956748f691553a503357fe6aae05576820f6fd3.jpg) + +![](images/d940a7225c230f84930ebc9bde10dc5e0c9421e4cf46634175efd18eff7cb89d.jpg) + +Clothing - transparent, look-up and down the + +ent, close-up on hips, domestic activities, the body + +![](images/fdcfcac27c4d446ca009ee59f00463c6c04a6c4c7ff584c3777053eddd9e5952.jpg) +full body $\rightarrow$ approach $\rightarrow$ close-up on hips + +![](images/c6adfd1eae3ccf7b42380146c7ee1c9a75cf5e6a4d47315723b9dd82a1c3ddb0.jpg) +Figure 1. In modern film media, the unequal characterization of gender on screen frequently evokes concepts of objectification, such as (A) unequal gaze (Pulp Fiction, 1994), (B) Nudity and submissive postures (Pulp Fiction, 1994), (C) animalisation or infantilisation (Marley and Me, 2008), and (D) transparent clothing, camera framing, domestic gender roles, and voyeurism (Gone Girl, 2014). + +![](images/f83c08da26f3592d3421b389c8574122ff7f8c32dc86a9db5a189a0bdd859f59.jpg) + +![](images/f9ed2065e7ac9bb539744f8dec66bca1eb858b42ca9df25292527bf0802ba2c0.jpg) + +As explained by Samory et al. [47], sexism is a complex sociological construct, whose high-level interpretive nature and subtle dimensions beyond offensive speech make for an unsolved challenge. In visual media such as films and TV series, the characters they depict shape our collective imagination and perception of sociological constructs, such as gender, race, and class. Currently, most large-scale approaches to understanding gender representation in these media have focused on quantifying the presence of women in the image and audio content (e.g., [35, 39, 51]). However, works in social sciences show that quantifying the presence of gender on screen is insufficient for grasping the issue of + +gender inequalities in visual media. For films, the classic Bechdel test, although useful and simple, considers neither the visual modality, which is key to analyzing gender depiction [41], nor the textual constructs of speech and dialogue [49]. While a few works have investigated sexist memes [16, 19], sexist advertisement [18], and characterized the on-screen positioning and co-occurrence of certain groups with respect to scene types and objects (e.g., Wang et al. for still images [54], Jang et al. [26] for films), computational approaches to interpretive sexism in visual media remain very scarce. + +In this article, we introduce a new challenging task for computer vision: detecting character objectification in films. Owing to the importance of this question, we consider it is critical to support the design of explainable methods and fine-grained model error analysis, which we address by densely annotating video data for theory-driven concepts. This presents a major step to tackle the question of subtle sexism in videos, operationalizing the popularly known concept of male gaze with the construct of objectification, and specifically considering the temporal dimension where such video patterns unfold. The end goal is to enable large-scale quantification and characterization of complex patterns producing on-screen objectification, and unveil possible correlations along the lines of the gender or race constructs. + +# Our contributions are: + +- We introduce a novel video-interpretation task to detect character objectification in films. This is an interpretive task, hence extending beyond the more classical yet still challenging video-understanding tasks, and involving a subjective judgement. In a team involving media studies experts, and building on results in cinematography and psychology, we design a thesaurus of visual objectification, defining coarse-grained concepts with exemplified instances. This thesaurus is then used to formulate precise annotation guidelines. We introduce the ObyGaze12 dataset to the community, with 1914 clips from 12 films densely annotated by experts for concepts of objectification, including hard negative examples. It corresponds to $25\%$ of the MovieGraphs dataset. We verify the consistency of the obtained data, and provide first analyses showing the compositional nature of objectification. The dataset is meant to explore the complex temporal patterns producing character objectification in films. To the best of our knowledge, it is the first work proposing a computational approach to this interpretive task in videos. + +- We verify that the new task of objectification detection in videos is accessible by testing recent vision and vision-language models, and that hard negative examples improve classification. We also investigate the model weaknesses in representing each objectification concept. To do so, a + +thorough analysis is carried out with Concept Bottleneck Models (CBMs) which allows us to identify the individual concepts that pose the greatest challenge, namely: Type of Shot, Look, Posture, and Appearance. + +To the best of our knowledge, this is one of the few video datasets with dense concept-based annotations for a high-level construct, and the first for objectification. The dataset and code used in this article is entirely provided in https://github.com/husky-helen/ObyGaze12. + +The article is organized as follows. In Sec. 2, we first provide a review of the relevant works on visual biases in films, dataset creation, and models for video understanding. We then introduce dataset creation for ObyGaze12 and present first analyses in Sec. 3. Sec. 4 presents the evaluation of models on the new task, and the analysis of the difficulty of concept representation with CBMs. Finally, we provide discussions on ethical aspects and challenges in Sec. 5, and conclusions in Sec. 6. + +# 2. Related works + +In this section, we position our contributions with respect to the relevant existing work. First, we introduce biases in visual datasets and computational approaches to the analysis of visual gender representation in films. We then discuss interpretive-level tasks, increasingly common in natural language processing, and the approaches to dataset creation to instantiate them for ML approaches. Here we highlight the scarcity of visual datasets made for high-level interpretive tasks, in particular for video data. Finally, we introduce the video understanding approaches that we consider to benchmark on our new interpretive task, specifically focusing on explainable concept-based approaches to locate the challenges ahead in video interpretation for this new task. + +# 2.1. Visual biases in film datasets + +The task we introduce is connected to the general problem of bias detection. As exposed by Fabbrizzi et al. [15], biases in visual datasets can be classified into selection bias (how subjects are included in a dataset), framing bias (how the visual content has been artificially composed) and label bias (errors or disparities in the labelling data). Our contribution is closely related to the framing and labelling biases. In film datasets, the first studies of biases in gender representation were from a presence quantification perspective. Guha et al. [21] and Somandepalli et al. [51] automatically estimate the screen time (from video) and speaking time (from audio) of male and female characters in Hollywood movies. They show that women are seen (36% screen time) and heard (41% screen time) significantly less than male characters. Mazieres et al. [39] consider a movie dataset spanning three decades, and show a temporal trend towards general fairer representation between both binary genders. + +They however show simultaneously that the applied framings remain unfair, with only $40\%$ of one-face frames featuring a female (60% for males). Jang et al. specifically analyze the qualitative framing differences in gender portrayal in a dataset of 20 Hollywood movies and 20 Korean movies [26]. They find that female characters are portrayed with lower emotional diversity, spatial occupancy, temporal occupancy, intellectual image or mean age. + +In contrast in this article, we take a first step towards detecting bias in gender representation from a high-level construct, objectification, qualitatively described in various disciplines such as cinematography [9, 41], social psychology [1, 32], and neuroscience [5-7]. Objectification is produced by complex temporal patterns never analyzed computationally in videos until now. + +# 2.2. Interpretive-level tasks and dataset creation + +At the same time that biases in visual datasets are uncovered and analyzed, other approaches aim to detect bias in human data. Detecting highly interpretive constructs, such as hate speech, propaganda, sexism, and racism, has been a long-standing endeavor in NLP. Da San Martino et al. [12] recruited to 4 experts to annotate news articles with text spans associated to 18 possible propaganda techniques. Samory et al. uncover the challenge of construct complexity for sexism detection [47]. They observe that multiple articles for automating this task consider widely different definitions of sexism, often referring to sub-dimensions of the broader construct. They consider existing works in social psychology where sexism dimensions have long been operationalized with sub-scale questions tested for consistency. To approach dataset creation for on-screen objectification, we inspire from these two last works, and on dense annotation approaches of image datasets recently proposed for the medical domain [13]. To the best of our knowledge, approaches for sexism detection from visual content are scarce, and almost none existing for video data. Two main types of visual content have been considered so far: hateful or sexist memes [16, 29], and sexist advertisements [18] which can also relate to symbolic advertisement understanding [25, 28]. + +In films, analysis of biases in gender representation has been also automated with NLP approaches applied to film scripts [2, 37]. Specifically, Martinez et al. [37] propose a RNN-based model to automatically extract agent-verbal-patient triplets. From 912 movie scripts, they show that male characters are associated with a higher agency while female characters are more frequently the object of gaze. Su et al. introduce the more abstract task of trope understanding in movies [52]. Our work is close to this last one as we also introduce a dataset of films annotated for a higher-level construct beyond event and story understanding. However, we provide dense annotations of sequences with constituency + +tive concepts for the high-level construct of character objectification. Also, while Su et al. exploit an existing online base contributed to by the community, we design and carry out a strictly defined annotation process by experts. + +In this article, we set out from the concept of male gaze defined in various ways in film gender studies [9, 36, 41]. Mulvey [41] characterizes the concept of gaze by the three relations between the camera and the characters, between the characters, but also between the spectator and the characters. While some formalizations of gaze set out from the gender of the director [36], Brey defines female and male gaze only from the film content, with a corpus analysis focusing on aesthetics [9] and revolving around the construct of objectification. We consider this gaze analysis of Brey, which does not significantly rely the socio-historical context of production and reception. We build the operationalization objectification in both cinematography and psychology, to create the conditions to make a new challenge accessible to the CV community: we produce a strict annotation process to obtain densely annotated video data and analyze where the new challenges lie ahead. We generally position our approach producing a non-large scale but high-quality dataset that we hope will be useful within the lines of the call of Paullada et al. for such data-centric AI approaches [44]. + +# 2.3. Approaches to video and movie understanding + +Cross-modal foundation models, such as CLIP [45] and ALIGN [27], learn aligned image and text representations through contrastive pre-training on large-scale closed datasets. Generalizations of the CLIP model to video data have included VideoCLIP [57] and two X-CLIP models [34, 42]. In particular, X-CLIP [42], which we employ in this article, expands CLIP with video temporal modeling and video-adaptive textual prompts. + +Legacy and large-scale pre-trained vision-language models have been leveraged for movie-related tasks. Bose et al. consider the difficulty of visual scene recognition in movies due to domain mismatch and create the MovieCLIP dataset obtained from weakly labeling movie shots from scene categories using the CLIP model [8]. An important vision-language movie-related task is audio-description for the visually-impaired, for which a major dataset introduced recently is MAD [50], gathering sparse natural language sentences grounded in over 1200 hours of movie videos. AutoAD [23] and AutoADII [22] are two recent approaches to generate audio-description from the video, both leveraging CLIP to learn to prompt GPT. To design approaches to learn human-level constructs, such as emotions, interactions or relationships, datasets generated with human supervision are also instrumental. A prominent representative is the MovieGraphs dataset, providing detailed annotations of clips of 51 movies with emotional states, character interac + +tions and relationships, and other scene reasoning elements [53]. It has prompted works tackling such recognition tasks [17, 33]. In this article, we build on the MovieGraphs dataset to annotate a selection with the construct of objectification, to later analyze it in connection with the other annotated social elements. + +In this article, we aim to assess the capacity of CLIP-based methods to provide relevant embeddings for the concept of objectification. We do so with a direct evaluation of classification results when an adapter (MLP) fed by X-CLIP embeddings is learnt. We analyze the results with a concept-based approach by building on Concept Bottle-neck Models (CBM) [31]. Concept-based models are an active area of research in XAI, with works tackling the accuracy-explainability tradeoff [60] and the need for user-defined concepts [58]. We specifically employ Post-hoc CBM (PCBM) [59], which consists in learning a concept subspace (made of Concept Activation Vectors [30]) in the embedding space of the pre-trained model. Data samples are then projected in this concept subspace, from where the classification task can be performed with an interpretable classifier. + +# 3. Data and methods + +This section presents our approach to create the first dataset for visual objectification in videos, specifically in films. We name this dataset ObyGaze12, short for ObjectifyingGaze12, which has the following highlights: + +- It considers the multiple dimensions of the construct of visual objectification, made of filmic (framing and editing over successive shots, camera motion, etc.) and iconographic properties (visible objects, body parts, attire, character interactions, etc.). +- It is based on a thesaurus articulating five sub-constructs identified from multidisciplinary literature (film studies and psychology) from which we define typical instances, then grouped into coarse-grained visual concepts. +- The data is annotated densely with concepts, and shows the multi-factorial property of objectification, corroborating with some recent developments in cognitive psychology [6]. +- A hard negative category is included with the goal to perform fine-grained error analysis and improve model generalization. + +# 3.1. A thesaurus of objectification + +We first formalize the construct of visual objectification and derive key concepts to annotate in film scenes. Together with media studies experts, we identify five sub-constructs of objectification from literature on film cognition and film gender studies, and social and cognitive psychology: male gaze (point of view of a man on a woman) [5, 9, 41], sexualization [6, 7], surveillance of the feminine body [11, 14, 40], + +female inaction / male possession [20, 48], and infantilism / animalization [41]. + +These sub-constructs come with typical instances and examples from filmmaking techniques ([9, 41]) or validated questionnaires ([11, 40]), as shown in the middle and rightmost columns in Table 1. These typical instances are then grouped into eight coarse-grained visual concepts, corresponding to the possible means of production of visual objectification. They are shown in the left-most column in Table 1: type of shot (framing and gaze of camera), look (gaze of characters on the other), body (partial or full nudity, and sexually suggestive body parts), posture (connoting, e.g., childhood, submission or inaction), clothing (in relation to context and activities), appearance (age and makeup), expression of emotion (restrained or exaggerated according to gender role), and activities (linked to gender roles). Visual examples are provided in Fig. 1, where we show video samples of objectification concepts Look, Posture, Type of shot, Clothes and Activity. + +# 3.2. Data selection + +Over the various existing movie datasets (see Sec. 2 and [4, 24, 39, 43, 46, 50, 53], [51, Table 1]), many have overlapping titles and only a few have rich human supervision. Amongst these, the MovieGraphs dataset [53] includes rich, high-level human annotations of 7637 clips of 51 movies, with emotional states, interactions and relationships, and other social reasoning elements. These elements are important and valuable in exploring social concepts of objectification in visual media. These movies are also frequently adopted for media research in existing work in [4, 24, 39]. The movie clips have also a short duration – mostly within 1-5 minutes – facilitating dense and granular annotations (over 100 clips for an average 2 hour film) while preserving the possibility to observe longer-term interactions and story development across a number of shots. From the 7637 clips of the MovieGraphs dataset, we select 1914 clips to annotate for objectification, amounting to $25\%$ of the dataset and 12 complete movies, which were selected to approximately reproduce the fraction of genres in the original dataset. The list of selected films, year of release, and genre can be found in Table 3 in Appendix 7. + +# 3.3. Data annotation + +Every selected movie is annotated by at least two experts for objectification level and concepts over the movie scenes. Specifically, the annotators were asked to repeat a three-step process for every scene they deemed interesting from an objectification perspective: (1) watch the movie entirely and when they identify a scene worth annotating, (2) delimit the clip by using the cutting function in our annotation tool, and (3) assigning an objectification level and annotate the concept(s) involved in the objectification rating. We define + +Table 1. Thesaurus of the typical instances and examples of visual objectification in films, grouped into eight main visual concepts used for annotation. Examples are possible means to produce one of the five sub-constructs of objectification (male gaze, sexualisation, surveillance of the feminine body, female inaction/male possession, infantilism/animalisation), to be assessed by annotators. + +
ConceptConcept instancesExamples
Type of shotShot suggesting man perspective in presence of womanclose-up on a man's face; body parts of woman
Shot suggesting man gaze on womancamera takes the perspective of a male character with first close-up on the face followed by camera motion looking a woman from bottom up
Shot showing a woman in parallel with an animalwoman at same level and position with a dog
LookVoyerismcharacter watching another one without their knowledge
Non-reciprocal gazewoman looking at man who does not look back
BodySuggested nudityclothing on floor; silhouette behind shower curtain; nude shadow on wall
Partial nuditynude upper or lower body; partially open clothing or draping; in underwear
Full nuditynude person fully or partially shown
Body parts suggestive of sexclose-up shots on breast, buttocks, hips, or lips
PostureGesture or posture connoting seductionlip-biting; hip roll; twisting or tucking hair
Gesture or posture connoting sexualityeating phallic symbols; arching back
Gesture or posture connoting inactionbeing undressed by someone
Gesture or posture connoting submissionleaning on a man
Gesture or posture connoting dependencefollowing a man
Skippingskipping gait
ClothingWet or transparent clothingthin shirt soaked in rain
Clothing impractical to situationwear pumps for running, a skirt when gardening
Color code associated to characterwoman with pink clothing and accessories
Older woman wearing infantile clothingwoman wearing an Alice band or high socks
AppearanceDiscrepancy between appearance of woman and context or biographical elementsperfect makeup when waking up; mother of heroine too young; young girl played by older actress
Exp. of emotionAsymmetric expression of emotionboys don't cry; woman being hysterical
ActivityDoing domestic activitiesdoing laundry, cooking, cleaning, being constantly in the kitchen
+ +four levels of objectification: + +- Easy Negative: there are no elements suggestive of objectification. No concept can be annotated. Default value for watched but non-selected scenes. +- Hard Negative: the scene contains elements of objectification from the thesaurus, but their presence does not result in an objectification effect. +- Not Sure: the scene indicates objectification, but does not completely fit the definition in the thesaurus. +- Sure: the scene contains elements of objectification from the thesaurus, and their presence results in an objectification effect. + +Four expert annotators were recruited to annotate the dataset. A first presentation session was held to introduce our annotation tool and the annotation procedure. All four annotators were then given the same two films – Juno and Silver Linings Playbook – to annotate using the proposed methodology. A second meeting was then set up after the annotation of the two films to analyze the reasons for divergence and remedy them by identify which elements of the annotation guidelines to clarify and how. We then randomly assigned two annotators to each of the remaining 10 films to annotate separately. + +Data processing and fusion The data processing is described in detail in Appendix 7. Following the annotation step, the annotations are then projected from the delimi- + +tations provided by each annotator onto the delimitations of the MovieGraphs clips. Since multiple annotations may overlap the same MovieGraphs clip, the annotation that is projected on the MovieGraphs clip corresponds to the annotation (including objectification level and concepts) with (1) the highest level of objectification that has at least $20\%$ overlap with the MovieGraphs clip, and (2) when multiple annotations exist at the same level of objectification, the annotated concepts for these annotations are aggregated. The same process is used to aggregate the annotations of the annotators of a same clip: the maximum objectification level is kept, with possible aggregation of concepts in case both annotators chose the same level but annotated different concepts. The merged data is shared and used in the remaining of this article. + +# 3.4. Analysis of the ObyGaze12 dataset + +Here we comment on some interesting statistics of the resulting annotations and concepts of the 1914 clips originally delimited in the MovieGraphs dataset. + +First, we verify data consistency by computing the inter-annotator agreement (IAA). Given the task of annotating timespans, we choose the $\gamma$ agreement measure introduced in [38] (and used for, e.g., annotating text spans [12]) owing to its consideration of temporal alignment, multiple annotators, and label classification at the same time. It attributes a score between 1 (complete agreement) and $-\infty$ . A value of + +$\gamma \leq 0$ indicates no agreement. The computation details of the $\gamma$ metric is provided in Appendix 7. Considering all four categories EN, HN, NS, S, we obtain and average $\gamma = 0.42$ . Not considering the clips annotated Not Sure (NS), which is the uncertain and "noisy" class in human annotations, the IAA increases to $\gamma = 0.69$ . This shows the consistency of the obtained annotations despite the interpretive nature of the task. Let us also mention that recent works improve learning approaches by considering explicitly the IAA in case of low number of annotators with moderate agreement [10, 55, 56]. + +Second, we analyze the obtained annotations in Fig 2. The Sure category is the least represented with $16\%$ , the Easy Negative being, as expected, the most represented class with $52\%$ of clips. It is interesting to note that every concept is approximately annotated with the same rate throughout the Hard Negative, Not Sure and Sure levels of objectification. Finally, it is very interesting to observe that the average number of concepts annotated per clip increases with the level of objectification: 1.26 concepts on average per Hard Negative clip, 1.71 for Not Sure, up to 2.6 for Sure. We verify that this trend is observable for every single annotator. It gives an important insight into our video interpretation data: that objectification is a compositional process. This corroborates with recent findings in neuroscience experiments that found that a single element, such as clothing on its own, is not sufficient for people to perceive a character as an object [6]. + +# 4. Experiments + +The experiments have two objectives: to verify that the new classification task is feasible, and to identify the challenges of designing efficient models. To tackle these objectives, we + +![](images/9c4f81a4703555bd1db8838a1fca58ac771c1c00e703a00353fcbf028e4b4419.jpg) +Figure 2. Distribution of visual factors annotated for each level of objectification (HN = Hard negative, NS = Not sure, S = Sure). The percentage of the dataset for each level of objectification as well as the average number of concepts per clip are also shown. (Best viewed in colors) + +consider pre-trained vision models and specifically address the following research questions: + +- Task accuracy - What are the baseline performances by pre-trained vision models on the objectification detection task? How does the performance vary with hard negative examples? +- Concept representation - Can we implement interpretable models of objectification using concepts? What is the quality of representation of every concept, and what are the objectification concepts poorly captured by current models? + +# 4.1. Task accuracy + +Setup We discard the Not Sure (NS) class from the ObyGaze12 dataset, as it gathers by definition samples highly uncertain for humans, and consider the Easy Negative (EN, $62\%$ of the clip samples), Hard Negative (HN, $19\%$ ) and Sure (S, $19\%$ ) classes. We approach binary classification in a progressive way, the positive class being made of the S samples. We consider two levels of classification difficulty by composing the negative class either with EN samples, or with HN samples. The implementation details of cross-validation and data balancing are provided in Appendix 8. The average performance over the test set of the best models on validation folds are shown in Table 2 with standard deviations. + +Baselines We consider video embeddings obtained from pre-trained models owing to their zero-shot classification capabilities on video tasks. We select ViViT-B/16 [3], and the available X-CLIP model, trained on Kinetics [42]. We also re-train a X-CLIP model [34] on the LSMDC [46] film dataset, and refer to Appendix 10 for corresponding results, where all implementation details are described. We keep the pre-trained models frozen and perform an adaptative max pooling of the resulting frame tokens, and feed the output to an MLP made of 2 dense layers, the hidden layer with 128 neurons and ReLU activations, the last with 2 softmax neurons. + +Results To assess the quality of the models on possibly imbalanced data with a minority of positive samples, we report the F1-scores in Table 2. First, by comparing with trivial classifiers (random and an all-positive, see App. 8), we observe that the task is indeed feasible, warranted by the data consistency described in Sec. 3 despite the interpretive nature of the task. Second, we observe that the inclusion of Hard Negative examples improves the classification results, showing the importance of a fine-grained annotation for highly-interpretive tasks. Results of X-CLIP on other configurations, specifically when the movies of clips in test are different from those in train, are shown in App. 8. The best results based on existing models are of moderate qual + +Table 2. F1-score on the binary task of objectification detection for models trained with easy or with hard negatives and tested on easy or all negative samples, with standard deviations. + +
Test TrainEN vs. S(EN U HN) vs. S
EN vs. SHN vs. SEN vs. SHN vs. S
ViViT-B/160.53 (0.18)0.62 (0.13)0.54 (0.24)0.73 (0.1)
X-CLIP0.79 (0.05)0.71 (0.05)0.66 (0.05)0.82 (0.03)
Random0.540.55
All positive0.080.06
PCBM-DT0.680.440.580.38
PCBM-LR0.640.430.500.37
+ +ity, which calls for more investigation into where the difficulties lie. + +# 4.2. Concept accuracy + +To infer on-screen objectification, it is key for the model to detect the means of its production, which correspond to the eight concepts listed in Table 1. We reiterate that in the ObyGaze12 dataset, every clip annotated with a level of objectification S, NS or HN is also annotated with the presence of instances of the eight concepts. For example, if the Body concept is annotated, it means that some level of nudity and/or suggestive body parts are shown on screen, that could contribute to the production of objectification. The means of producing objectification through the eight concepts can be subtle to detect, making it difficult to provide the final interpretation. To investigate this difficulty, we implement Post-hoc Concept Bottleneck Models (PCBMs) [59], which allow us to approach a classification task with pre-trained models in an interpretable way when concept-annotated data is available. In our case, from the X-CLIP embedding space where our video clips are represented, we identify a Concept Activation Vector (CAV) [30] for every concept. We then project the X-CLIP embedding of every clip onto the subspace defined by the eight CAVs. The representation of the clip that is the output of this bottleneck is a low-dimensional vector with number-of-concepts components. This vector can then be fed to an interpretable classifier for the objectification detection task. + +CAV computation For every concept $i$ , we collect two sets of samples: positive samples where concept $i$ is present, hence made of S and HN samples with the concept annotated as present, and negative with EN, S and HN without concept $i$ . We then train a linear SVM for each concept $i$ , the CAV of concept $i$ being the normal vector of the SVM hyperplan. To train the SVM, we split the data in 10 folds, and reserve the last fold for test. We then perform an 8-fold cross-validation to select the SVM (choice of margin tolerance $c$ ), every fold training set being balanced with different draws of negative sub-sampling. + +Interpretable classifier We then train a decision tree (DT) and a logistic regression (LR) classifier on the same 8-fold cross-validation to classify the level of objectification, the classifiers being fed with the projection of every clip onto the CAVs. + +Task accuracy with PCBM We first verify the quality of objectification detection with F1-scores of PCBM-DT and PCBM-LR shown in Table 2. The results lower than X-CLIP are expected owing to the known accuracy-interpretability tradeoff of CBMs [59, 60]. They are above random and all-positive predictions when training on EN vs. S, which is indicative of the relevance of information held in the concepts. However, the low results obtained when training the DT and LR to distinguish between S and HN reveals the low quality of some CAVs, where the X-CLIP embeddings cannot be linearly well-separated for these concepts. We investigate this point next. The resulting DT is discussed in App. 9. + +Concept accuracy We now analyze the quality of each obtained CAV by plotting its capability to classify whether the concept is present in a test sample. We consider a positive similarity between the X-CLIP embedding and CAV of concept $i$ indicative of the presence of concept $i$ . F1-score on the test set are shown in Fig. 3. Plots correspond to CAVs obtained from classifying the presence of concept $i$ against EN only (solid bars) and against EN with S and HN without concept $i$ (hatched bars). The former is used for PCBM-DT in Table 2. We first observe that concept detection is harder when negative samples also include S and HN samples (without the concept). This is expected considering that scenes tagged EN have by definition no element possibly conducive to objectification, and are hence likely to differ visually more from scenes where the concept is present, than do S and HN scenes without the concept. However within S and HN clips with and without the concept, such shortcuts cannot be exploited anymore. We observe in this case that the X-CLIP embedding related to concepts Type of shot, Posture, Look and Appearance are harder to separate linearly. This can be correlated with the analysis of factors of error detailed next. These subtler means of on-screen objectification therefore warrant future work to be properly captured and detected. + +Error Analysis We analyze the factors impacting objectification classification errors corresponding to the results shown in Table 2. We select an average X-CLIP-based adaptation model trained on HN vs. S and tested on (ENUHN) vs. S, and consider its predictions on the clips in the test set. We label each test clip with 0 if the model fails to predict the correct label of the clip, and with 1 otherwise. We describe the clip with a one-hot encoding vector + +![](images/738fe22f7fad2dbd95c7f88aa8448030dc9708d00f061b53e7a5cec87d1bc493.jpg) +Figure 3. For every concept, F1-score of the best linear SVM selected to define the CAV of this concept. Positive samples (S and HN with the concept) must be separated from: [non-hatched bars] negative samples made of EN only, or [hatched bars] negative samples made of EN and S and HN without the concept. + +corresponding to all 11 factors shown the y-axis of Fig. 4: every of the 8 concepts, and the Sure (S), Hard negative (HN) and Easy negative (EN) labels. We then train a logistic regression model on these clip descriptors to predict the failure/success labels. Fig. 4 shows the error regression weights associated with each factor, for two film examples. A negative weight indicates a contribution of the factor to a classification failure. We first observe that the HN characteristic contributes to a classification failure of the model, while S and EN contribute to classification success. Second, we observe that there is variability over the movies on the presence of which concept strongly influences the classification success. However, the presence of the Clothing concept seems to be a strong confounder. This can be due to the frequency of appearance of this concept in HN samples, and to the subtlety of the description of this concept (provided in Table 1), which should make it difficult for a pre-trained model to discriminate between an objectifying and non-objectifying overall label on the basis of Clothing. Third, it is worth noting that the concepts shown to be poorly linearly separable when described with the X-CLIP embeddings (see CAV analysis in Sec. 4.2), are also those with a non-stable contribution to the model errors over the film examples: Type of shot, Look, Posture and Appearance. + +# 5. Discussion + +Ethical aspect This work has an explicit societal motivation in its purpose to tackle, with the help of AI, the analysis of complex temporal patterns operated in cinema that pro + +![](images/605d1d403a836609b02b61a29651571e79fc53cce48fa2531709359423eb0a59.jpg) +Figure 4. Analysis of the factors of error for the objectification detection task: weights of a logistic regressor predicting whether the test set examples are well classified or not. Positive (resp. negative) weights indicate a positive (resp. negative) contribution to classification success. Left: As good as it gets, Right: Up in the Air. + +duce the perception of certain characters as objects. This is a challenging but valuable task that aims to uncover and quantify differences in how various identities may be portrayed on screen. + +Limitations and challenges A distinctive element of our work is the subjective judgement involved in annotating granular video elements for objectification. Video annotation is tedious, and approaching data annotation for such an interpretive task in a rigorous way is even more so, and difficult to scale. We therefore believe that pursuing high-quality, dense annotations with well-defined concepts goes a long way to tackle this new video interpretation task, which represents a valuable new challenge for the computer vision community. + +# 6. Conclusion + +In this article, we have introduced a new video interpretation task to detect character objectification in films. We have introduced the ObyGaze12 dataset, densely annotated by experts for objectification concepts defined from five sub-constructs identified in film studies and psychology. ObyGaze12 is made available to the community. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. We show that the representation learning of the concepts of Type of shot, Look, Posture and Appearance need to be improved. + +Acknowledgements. This work has been partly supported by the French National Research Agency through the ANR TRACTIVE project ANR-21-CE38-0012-01. This work was partly supported by EU Horizon 2020 project AI4Media, under contract no. 951911 (https://ai4media.eu/). + +# References + +[1] Operationalizing self-objectification: Assessment and related methodological issues. In Self-objectification in women: Causes, consequences, and counteractions., pages 23-49. American Psychological Association, Washington, 2011. 3 +[2] Apoorv Agarwal, Jiehan Zheng, Shruti Kamath, Sriramkumar Balasubramanian, and Shirin Ann Dey. Key Female Characters in Film Have More to Talk About Besides Men: Automating the Bechdel Test. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 830-840, Denver, Colorado, 2015. Association for Computational Linguistics. 3 +[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, Montreal, QC, Canada, 2021. IEEE. 6 +[4] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed Movies: Story Based Retrieval with Contextual Embeddings. Technical Report arXiv:2005.04208, arXiv, 2020. arXiv:2005.04208 [cs] type: article. 4 +[5] Philippe Bernard, Sarah J. Gervais, and Olivier Klein. Objectifying objectification: When and why people are cognitively reduced to their parts akin to objects. European Review of Social Psychology, 29(1):82-121, 2018. Publisher: Routledge _eprint: https://doi.org/10.1080/10463283.2018.1471949. 3, 4 +[6] Philippe Bernard, Florence Hanoteau, Sarah Gervais, Lara Servais, Irene Bertolone, Paul Deltenre, and Cecile Colin. Revealing Clothing Does Not Make the Object: ERP Evidences That Cognitive Objectification is Driven by Posture Suggestiveness, Not by Revealing Clothing. Personality and Social Psychology Bulletin, 45(1):16-36, 2019. 4, 6 +[7] Philippe Bernard, Carlotta Cogoni, and Andrea Carnaghi. The Sexualization-Objectification Link: Sexualization Affects the Way People See and Feel Toward Others. Current Directions in Psychological Science, 29(2):134-139, 2020. 3, 4 +[8] Digbalay Bose, Rajat Hebbar, Krishna Somandepalli, Haoyang Zhang, Yin Cui, Kree Cole-McLaughlin, Huisheng Wang, and Shrikanth Narayanan. Movieclip: Visual scene recognition in movies. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2082-2091, 2023. 3 +[9] Iris Brey. Le regard féminin-Une révolution à l'écran. Média Diffusion, 2020. 3, 4 +[10] M. Bucarelli, L. Cassano, F. Siciliano, A. Mantrach, and F. Silvestri. Leveraging inter-rater agreement for classification in the presence of noisy labels. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3439-3448, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6, 2 +[11] Rachel M. Calogero. A Test of Objectification Theory: The Effect of the Male Gaze on Appearance Concerns in Col- + +lege Women. Psychology of Women Quarterly, 28(1):16-21, 2004. 4 +[12] Giovanni Da San Martino, Seunghak Yu, Alberto Barrón-Cedeno, Rostislav Petrov, and Preslav Nakov. Fine-Grained Analysis of Propaganda in News Article. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5635–5645, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 5 +[13] Roxana Daneshjou, Mert Yuksekgonul, Zhuo Ran Cai, Roberto A. Novoa, and James Zou. Skincon: A skin disease dataset densely annotated by domain experts for fine-grained debugging and analysis. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3 +[14] Angela Denchik. Development and Psychometric Evaluation of the Interpersonal Sexual Objectification Scale. PhD thesis, The Ohio State University, 2005. 4 +[15] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223: 103552, 2022. 2 +[16] Elisabetta Fersini, Francesca Gasparini, and Silvia Corchs. Detecting Sexist MEME On The Web: A Study on Textual and Visual Cues. In 2019 8th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIw), pages 226-231, 2019. 2, 3 +[17] Bei Gan, Xiujun Shu, Ruizhi Qiao, Haoqian Wu, Keyun Chen, Hanjun Li, and Bohan Ren. Collaborative noisy label cleaner: Learning scene-aware trailers for multi-modal highlight detection in movies. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18898-18907, 2023. 4 +[18] Francesca Gasparini, Ilaria Erba, Elisabetta Fersini, and Silvia Corchs. Multimodal Classification of Sexist Advertisements.: In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications, pages 399-406, Porto, Portugal, 2018. SCITEPRESS - Science and Technology Publications. 2, 3 +[19] Francesca Gasparini, Giulia Rizzi, Aurora Saibene, and Elisabetta Fersini. Benchmark dataset of memes with text transcriptions for automatic detection of multi-modal misogynistic content. arXiv:2106.08409 [cs], 2021. arXiv:2106.08409.2 +[20] Sarah J. Gervais, Gemma Sáez, Abigail R. Riemer, and Olivier Klein. The Social Interaction Model of Objectification: A process model of goal-based objectifying exchanges between men and women. British Journal of Social Psychology, 59(1):248-283, 2020. 4 +[21] Tanaya Guha, Che-Wei Huang, Naveen Kumar, Yan Zhu, and Shrikanth S. Narayanan. Gender Representation in Cinematic Content: A Multimodal Approach. In Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, pages 31–34, Seattle Washington USA, 2015. ACM. 2 +[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel - who, + +when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13645-13655, 2023. 3 +[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie Description in Context. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18930-18940, Vancouver, BC, Canada, 2023. IEEE. 3 +[24] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. MovieNet: A Holistic Dataset for Movie Understanding. In Computer Vision – ECCV 2020, pages 709–727. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 4 +[25] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic Understanding of Image and Video Advertisements. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1100-1110, Honolulu, HI, 2017. IEEE. 3 +[26] Ji Yoon Jang, Sangyoon Lee, and Byungjoo Lee. Quantification of Gender Representation Bias in Commercial Films based on Image Analysis. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW):1-29, 2019. 2, 3 +[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 3 +[28] Nasrin Kalanat and Adriana Kovashka. Symbolic image detection using scene and knowledge graphs. Technical Report arXiv:2206.04863, arXiv, 2022. arXiv:2206.04863 [cs] type: article. 3 +[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3 +[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie J. Cai, James Wexler, Fernanda B. Viégas, and Rory Sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International Conference on Machine Learning, 2017. 4, 7 +[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In Proceedings of the 37th International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 4 +[32] Holly B. Kozee, Tracy L. Tylka, and Casey L. Augustus-Horvath. Interpersonal Sexual Objectification Scale, 2011. 3 +[33] Anna Kukleva, Makarand Tapaswi, and Ivan Laptev. Learning Interactions and Relationships Between Movie Characters. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9846-9855, Seattle, WA, USA, 2020. IEEE. 4 + +[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, page 638–647, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6 +[35] Sarah Macharia. Global Media Monitoring Project (GMMP), pages 1-6. John Wiley Sons, Ltd, 2020. 1 +[36] Alicia Malone. The Female Gaze: Essential Movies Made by Women. Mango Publisher, 2018. 3 +[37] Victor R. Martinez, Krishna Somandepalli, and Shrikanth Narayanan. Boys don't cry (or kiss or dance): A computational linguistic lens into gendered actions in film. PLOS ONE, 17(12):e0278604, 2022. 3 +[38] Yann Mathet, Antoine Widlocher, and Jean-Philippe Métivier. The unified and holistic method gamma $(\gamma)$ for inter-annotator agreement measure and alignment. Computational Linguistics, 41(3):437-479, 2015. 5, 1 +[39] Antoine Mazieres, Telmo Menezes, and Camille Roth. Computational appraisal of gender representativeness in popular movies. Humanities and Social Sciences Communications, 8 (1):137, 2021. 1, 2, 4 +[40] Nita Mary McKinley and Janet Shibley Hyde. The objectified body consciousness scale: Development and validation. Psychology of women quarterly, 20(2):181-215, 1996. 4 +[41] Laura Mulvey. Visual Pleasure and Narrative Cinema. Screen, 16(3):6-18, 1975. 1, 2, 3, 4 +[42] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 3, 6 +[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. MovieCuts: A New Dataset and Benchmark for Cut Type Recognition. In Computer Vision – ECCV 2022, pages 668–685. Springer Nature Switzerland, Cham, 2022. Series Title: Lecture Notes in Computer Science. 4 +[44] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[46] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie Description. International Journal of Computer Vision, 123(1):94-120, 2017. 4, 6 +[47] Mattia Samory, Indira Sen, Julian Kohne, Fabian Flock, and Claudia Wagner. "Call me sexist, but..." : Revisiting Sexism Detection Using Psychological Scales and Adversarial Samples. page 12. 1, 3 + +[48] Maarten Sap, Marcella Cindy Prasettio, Ari Holtzman, Hannah Rashkin, and Yejin Choi. Connotation frames of power and agency in modern films. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 2329-2334, 2017. 4 +[49] Alexandra Schofield and Leo Mehr. Gender-Distinguishing Features in Film Dialogue. In Proceedings of the Fifth Workshop on Computational Linguistics for Literature, pages 32–39, San Diego, California, USA, 2016. Association for Computational Linguistics. 2 +[50] M. Soldan, A. Pardo, J. Alcazar, F. Heilbron, C. Zhao, S. Giancola, and B. Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5016-5025, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3, 4 +[51] Krishna Somandepalli, Tanaya Guha, Victor R. Martinez, Naveen Kumar, Hartwig Adam, and Shrikanth Narayanan. Computational Media Intelligence: Human-Centered Machine Analysis of Media. Proceedings of the IEEE, 109(5): 891-910, 2021. 1, 2, 4 +[52] Hung-Ting Su, Po-Wei Shen, Bing-Chen Tsai, Wen-Feng Cheng, Ke-Jyun Wang, and Winston H. Hsu. TrUMAn: Trope Understanding in Movies and Animations. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 4594–4603, New York, NY, USA, 2021. Association for Computing Machinery. event-place: Virtual Event, Queensland, Australia. 3 +[53] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. MovieGraphs: Towards Understanding Human-Centric Situations from Videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 1 +[54] Angelina Wang, Arvind Narayanan, and Olga Russakovsky. REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets. In Computer Vision – ECCV 2020, pages 733–751. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 2 +[55] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. ArXiv, abs/2110.12088, 2021. 6, 2 +[56] Jiaheng Wei, Zhaowei Zhu, Tianyi Luo, Ehsan Amid, Abhishek Kumar, and Yang Liu. To Aggregate or Not? Learning with Separate Noisy Labels. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2523-2535, New York, NY, USA, 2023. Association for Computing Machinery. event-place: Long Beach, CA, USA. 6, 2 +[57] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 3 + +[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19187-19197, Vancouver, BC, Canada, 2023. IEEE. 4 +[59] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR^2Struct: Privacy, Accountability, Interpretability, Robustness, Reasoning on Structured Data, 2022. 4, 7 +[60] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models. In Advances in Neural Information Processing Systems, 2022. 4, 7 \ No newline at end of file diff --git a/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/images.zip b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd94cfa2d93e9e18e3b30eff583c6260b7df6e6d --- /dev/null +++ b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fadeb69630e082906586bcfa9f9cd11aab5f0724c89887456a608934b71378e +size 323051 diff --git a/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/layout.json b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf5b805fe8640f5fba8f1fb664f56a0ac12d61fc --- /dev/null +++ b/2024/Visual Objectification in Films_ Towards a New AI Task for Video Interpretation/layout.json @@ -0,0 +1,7948 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "type": "text", + "content": "Visual Objectification in Films: Towards a New AI Task for Video Interpretation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 142, + 178, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 142, + 178, + 156 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 178, + 156 + ], + "type": "text", + "content": "Julie Tores1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 143, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 143, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 196, + 143, + 288, + 156 + ], + "type": "text", + "content": "Lucile Sassatelli" + }, + { + "bbox": [ + 196, + 143, + 288, + 156 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 143, + 369, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 143, + 369, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 369, + 156 + ], + "type": "text", + "content": "Hui-Yin " + }, + { + "bbox": [ + 304, + 143, + 369, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{4}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 387, + 143, + 481, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 143, + 481, + 157 + ], + "spans": [ + { + "bbox": [ + 387, + 143, + 481, + 157 + ], + "type": "text", + "content": "Clement Bergman" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 157, + 174, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 157, + 174, + 170 + ], + "spans": [ + { + "bbox": [ + 110, + 157, + 174, + 170 + ], + "type": "text", + "content": "Léa Andolfi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 193, + 157, + 277, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 157, + 277, + 170 + ], + "spans": [ + { + "bbox": [ + 193, + 157, + 277, + 170 + ], + "type": "text", + "content": "Victor Ecrement6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 157, + 386, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 157, + 386, + 171 + ], + "spans": [ + { + "bbox": [ + 295, + 157, + 386, + 171 + ], + "type": "text", + "content": "Frédéric Precioso²" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 404, + 157, + 483, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 157, + 483, + 171 + ], + "spans": [ + { + "bbox": [ + 404, + 157, + 483, + 171 + ], + "type": "text", + "content": "Thierry Devars6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 154, + 171, + 235, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 171, + 235, + 185 + ], + "spans": [ + { + "bbox": [ + 154, + 171, + 235, + 185 + ], + "type": "text", + "content": "Magali Guaresi" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 253, + 171, + 338, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 171, + 338, + 185 + ], + "spans": [ + { + "bbox": [ + 253, + 171, + 338, + 185 + ], + "type": "text", + "content": "Virginie Julliard" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 171, + 440, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 171, + 440, + 185 + ], + "spans": [ + { + "bbox": [ + 356, + 171, + 440, + 185 + ], + "type": "text", + "content": "Sarah Lecossais" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 185, + 274, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 185, + 274, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 185, + 274, + 198 + ], + "type": "text", + "content": "1Université Côte d'Azur, CNRS, I3S, France" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 291, + 185, + 536, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 185, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 291, + 185, + 536, + 198 + ], + "type": "text", + "content": "2Université Côte d'Azur, CNRS, Inria, I3S, France" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 199, + 273, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 199, + 273, + 212 + ], + "spans": [ + { + "bbox": [ + 117, + 199, + 273, + 212 + ], + "type": "text", + "content": "3Institut Universitaire de France" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 291, + 199, + 476, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 199, + 476, + 213 + ], + "spans": [ + { + "bbox": [ + 291, + 199, + 476, + 213 + ], + "type": "text", + "content": "4Université Côte d'Azur, Inria, France" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 100, + 213, + 493, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 213, + 493, + 227 + ], + "spans": [ + { + "bbox": [ + 100, + 213, + 493, + 227 + ], + "type": "text", + "content": "5Université Côte d'Azur, CNRS, BCL, France Sorbonne Université, GRIPIC" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 196, + 227, + 398, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 227, + 398, + 240 + ], + "spans": [ + { + "bbox": [ + 196, + 227, + 398, + 240 + ], + "type": "text", + "content": "7Université Sorbonne Paris Nord, LabSIC" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 216, + 243, + 374, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 243, + 374, + 254 + ], + "spans": [ + { + "bbox": [ + 216, + 243, + 374, + 254 + ], + "type": "text", + "content": "julie.tores@univ-cotedazur.fr" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 282, + 191, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 191, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 191, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 308, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 475 + ], + "type": "text", + "content": "In film gender studies, the concept of \"male gaze\" refers to the way the characters are portrayed on-screen as objects of desire rather than subjects. In this article, we introduce a novel video-interpretation task, to detect character objectification in films. The purpose is to reveal and quantify the usage of complex temporal patterns operated in cinema to produce the cognitive perception of objectification. We introduce the ObyGaze12 dataset, made of 1914 movie clips densely annotated by experts for objectification concepts identified in film studies and psychology. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. Our new dataset and code are made available to the community." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "In film gender studies, the concept of \"male gaze\" [41] refers to the way the characters – especially women – are portrayed on-screen as objects of desire rather than subjects. Consider in Figure 1 how objectification is manifested in various ways such as camera placement and movement, the gaze interactions between characters, the choice of clothing, and arrangement of scene elements. Such disparities in how people are presented, depicted or addressed to in digital contents based on their gender has large-scale social implications such as the perpetuation of harmful stereotypes and hostile social situations." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "These disparities have been the subject of an increasing number of studies at the intersection of social and computational sciences. In online social networks, computational approaches to sexism detection have been increasingly investigated for textual data as a part of hate speech detection." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 286, + 382, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 382, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 382, + 295 + ], + "type": "text", + "content": "A Unreturned gaze" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 308, + 295, + 378, + 322 + ], + "blocks": [ + { + "bbox": [ + 308, + 295, + 378, + 322 + ], + "lines": [ + { + "bbox": [ + 308, + 295, + 378, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 378, + 322 + ], + "type": "image", + "image_path": "6061d4d9ee8b90046c4a3e7fe37b2a5951c45445f4b706ae11dcf946a39c5278.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 325, + 406, + 333 + ], + "lines": [ + { + "bbox": [ + 362, + 325, + 406, + 333 + ], + "spans": [ + { + "bbox": [ + 362, + 325, + 406, + 333 + ], + "type": "text", + "content": "Direction of gaze" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 384, + 295, + 417, + 321 + ], + "blocks": [ + { + "bbox": [ + 384, + 295, + 417, + 321 + ], + "lines": [ + { + "bbox": [ + 384, + 295, + 417, + 321 + ], + "spans": [ + { + "bbox": [ + 384, + 295, + 417, + 321 + ], + "type": "image", + "image_path": "b509430971965b5fbd75ef6728f306b89afc9bad029a49256216fcd022b52cfa.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 421, + 295, + 487, + 321 + ], + "blocks": [ + { + "bbox": [ + 421, + 295, + 487, + 321 + ], + "lines": [ + { + "bbox": [ + 421, + 295, + 487, + 321 + ], + "spans": [ + { + "bbox": [ + 421, + 295, + 487, + 321 + ], + "type": "image", + "image_path": "b4e87c2e2cbfd7bde9098f848f7a7045e4613dd8ce6d6e9e522acd7aa4b1be18.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 325, + 518, + 333 + ], + "lines": [ + { + "bbox": [ + 474, + 325, + 518, + 333 + ], + "spans": [ + { + "bbox": [ + 474, + 325, + 518, + 333 + ], + "type": "text", + "content": "Direction of gaze" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 493, + 295, + 537, + 321 + ], + "blocks": [ + { + "bbox": [ + 493, + 295, + 537, + 321 + ], + "lines": [ + { + "bbox": [ + 493, + 295, + 537, + 321 + ], + "spans": [ + { + "bbox": [ + 493, + 295, + 537, + 321 + ], + "type": "image", + "image_path": "1c10f2e48be524ec11bd908c158d0280647d7dd7a4ba3816c8d438dbd370e78a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 339, + 393, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 393, + 355 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 393, + 355 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 308, + 339, + 393, + 355 + ], + "type": "text", + "content": " Submissive posture, lower-body nudity" + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 318, + 360, + 375, + 385 + ], + "blocks": [ + { + "bbox": [ + 318, + 360, + 375, + 385 + ], + "lines": [ + { + "bbox": [ + 318, + 360, + 375, + 385 + ], + "spans": [ + { + "bbox": [ + 318, + 360, + 375, + 385 + ], + "type": "image", + "image_path": "0942612c1d711246e889d89dd76933f801d0a1cc400b3a0a7dc99ef90b8223fc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 403, + 338, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 338, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 403, + 338, + 545, + 347 + ], + "type": "text", + "content": "Shot - woman in parallel with animal" + } + ] + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 443, + 348, + 485, + 371 + ], + "blocks": [ + { + "bbox": [ + 443, + 348, + 485, + 371 + ], + "lines": [ + { + "bbox": [ + 443, + 348, + 485, + 371 + ], + "spans": [ + { + "bbox": [ + 443, + 348, + 485, + 371 + ], + "type": "image", + "image_path": "4e4882fc50ebcfeff8a96f46599ec3b724f4846777251e801f49e1ee686697b3.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 487, + 348, + 530, + 371 + ], + "blocks": [ + { + "bbox": [ + 487, + 348, + 530, + 371 + ], + "lines": [ + { + "bbox": [ + 487, + 348, + 530, + 371 + ], + "spans": [ + { + "bbox": [ + 487, + 348, + 530, + 371 + ], + "type": "image", + "image_path": "fd901d5686ff13217d2090174bbccabaa78c96d75fed3a9582852dac6b59134c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 443, + 372, + 484, + 395 + ], + "blocks": [ + { + "bbox": [ + 443, + 372, + 484, + 395 + ], + "lines": [ + { + "bbox": [ + 443, + 372, + 484, + 395 + ], + "spans": [ + { + "bbox": [ + 443, + 372, + 484, + 395 + ], + "type": "image", + "image_path": "e9d0dd0688b11567d69de0d49956748f691553a503357fe6aae05576820f6fd3.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 487, + 372, + 530, + 395 + ], + "blocks": [ + { + "bbox": [ + 487, + 372, + 530, + 395 + ], + "lines": [ + { + "bbox": [ + 487, + 372, + 530, + 395 + ], + "spans": [ + { + "bbox": [ + 487, + 372, + 530, + 395 + ], + "type": "image", + "image_path": "d940a7225c230f84930ebc9bde10dc5e0c9421e4cf46634175efd18eff7cb89d.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 310, + 398, + 405, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 398, + 405, + 415 + ], + "spans": [ + { + "bbox": [ + 310, + 398, + 405, + 415 + ], + "type": "text", + "content": "Clothing - transparent, look-up and down the" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 390, + 396, + 540, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 396, + 540, + 415 + ], + "spans": [ + { + "bbox": [ + 390, + 396, + 540, + 415 + ], + "type": "text", + "content": "ent, close-up on hips, domestic activities, the body" + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 309, + 416, + 388, + 440 + ], + "blocks": [ + { + "bbox": [ + 309, + 416, + 388, + 440 + ], + "lines": [ + { + "bbox": [ + 309, + 416, + 388, + 440 + ], + "spans": [ + { + "bbox": [ + 309, + 416, + 388, + 440 + ], + "type": "image", + "image_path": "fdcfcac27c4d446ca009ee59f00463c6c04a6c4c7ff584c3777053eddd9e5952.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "lines": [ + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "type": "text", + "content": "full body " + }, + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "type": "text", + "content": " approach " + }, + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 308, + 442, + 380, + 456 + ], + "type": "text", + "content": " close-up on hips" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 390, + 416, + 444, + 439 + ], + "blocks": [ + { + "bbox": [ + 390, + 416, + 444, + 439 + ], + "lines": [ + { + "bbox": [ + 390, + 416, + 444, + 439 + ], + "spans": [ + { + "bbox": [ + 390, + 416, + 444, + 439 + ], + "type": "image", + "image_path": "c6adfd1eae3ccf7b42380146c7ee1c9a75cf5e6a4d47315723b9dd82a1c3ddb0.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 464, + 547, + 540 + ], + "lines": [ + { + "bbox": [ + 306, + 464, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 547, + 540 + ], + "type": "text", + "content": "Figure 1. In modern film media, the unequal characterization of gender on screen frequently evokes concepts of objectification, such as (A) unequal gaze (Pulp Fiction, 1994), (B) Nudity and submissive postures (Pulp Fiction, 1994), (C) animalisation or infantilisation (Marley and Me, 2008), and (D) transparent clothing, camera framing, domestic gender roles, and voyeurism (Gone Girl, 2014)." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 444, + 416, + 499, + 439 + ], + "blocks": [ + { + "bbox": [ + 444, + 416, + 499, + 439 + ], + "lines": [ + { + "bbox": [ + 444, + 416, + 499, + 439 + ], + "spans": [ + { + "bbox": [ + 444, + 416, + 499, + 439 + ], + "type": "image", + "image_path": "f83c08da26f3592d3421b389c8574122ff7f8c32dc86a9db5a189a0bdd859f59.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 500, + 416, + 546, + 439 + ], + "blocks": [ + { + "bbox": [ + 500, + 416, + 546, + 439 + ], + "lines": [ + { + "bbox": [ + 500, + 416, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 500, + 416, + 546, + 439 + ], + "type": "image", + "image_path": "f9ed2065e7ac9bb539744f8dec66bca1eb858b42ca9df25292527bf0802ba2c0.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "As explained by Samory et al. [47], sexism is a complex sociological construct, whose high-level interpretive nature and subtle dimensions beyond offensive speech make for an unsolved challenge. In visual media such as films and TV series, the characters they depict shape our collective imagination and perception of sociological constructs, such as gender, race, and class. Currently, most large-scale approaches to understanding gender representation in these media have focused on quantifying the presence of women in the image and audio content (e.g., [35, 39, 51]). However, works in social sciences show that quantifying the presence of gender on screen is insufficient for grasping the issue of" + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10864" + } + ] + } + ], + "index": 49 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "gender inequalities in visual media. For films, the classic Bechdel test, although useful and simple, considers neither the visual modality, which is key to analyzing gender depiction [41], nor the textual constructs of speech and dialogue [49]. While a few works have investigated sexist memes [16, 19], sexist advertisement [18], and characterized the on-screen positioning and co-occurrence of certain groups with respect to scene types and objects (e.g., Wang et al. for still images [54], Jang et al. [26] for films), computational approaches to interpretive sexism in visual media remain very scarce." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 387 + ], + "type": "text", + "content": "In this article, we introduce a new challenging task for computer vision: detecting character objectification in films. Owing to the importance of this question, we consider it is critical to support the design of explainable methods and fine-grained model error analysis, which we address by densely annotating video data for theory-driven concepts. This presents a major step to tackle the question of subtle sexism in videos, operationalizing the popularly known concept of male gaze with the construct of objectification, and specifically considering the temporal dimension where such video patterns unfold. The end goal is to enable large-scale quantification and characterization of complex patterns producing on-screen objectification, and unveil possible correlations along the lines of the gender or race constructs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 403, + 159, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 403, + 159, + 413 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 159, + 413 + ], + "type": "text", + "content": "Our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 653 + ], + "type": "text", + "content": "- We introduce a novel video-interpretation task to detect character objectification in films. This is an interpretive task, hence extending beyond the more classical yet still challenging video-understanding tasks, and involving a subjective judgement. In a team involving media studies experts, and building on results in cinematography and psychology, we design a thesaurus of visual objectification, defining coarse-grained concepts with exemplified instances. This thesaurus is then used to formulate precise annotation guidelines. We introduce the ObyGaze12 dataset to the community, with 1914 clips from 12 films densely annotated by experts for concepts of objectification, including hard negative examples. It corresponds to " + }, + { + "bbox": [ + 46, + 415, + 289, + 653 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 415, + 289, + 653 + ], + "type": "text", + "content": " of the MovieGraphs dataset. We verify the consistency of the obtained data, and provide first analyses showing the compositional nature of objectification. The dataset is meant to explore the complex temporal patterns producing character objectification in films. To the best of our knowledge, it is the first work proposing a computational approach to this interpretive task in videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "- We verify that the new task of objectification detection in videos is accessible by testing recent vision and vision-language models, and that hard negative examples improve classification. We also investigate the model weaknesses in representing each objectification concept. To do so, a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": "thorough analysis is carried out with Concept Bottleneck Models (CBMs) which allows us to identify the individual concepts that pose the greatest challenge, namely: Type of Shot, Look, Posture, and Appearance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": "To the best of our knowledge, this is one of the few video datasets with dense concept-based annotations for a high-level construct, and the first for objectification. The dataset and code used in this article is entirely provided in https://github.com/husky-helen/ObyGaze12." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 192, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 547, + 300 + ], + "type": "text", + "content": "The article is organized as follows. In Sec. 2, we first provide a review of the relevant works on visual biases in films, dataset creation, and models for video understanding. We then introduce dataset creation for ObyGaze12 and present first analyses in Sec. 3. Sec. 4 presents the evaluation of models on the new task, and the analysis of the difficulty of concept representation with CBMs. Finally, we provide discussions on ethical aspects and challenges in Sec. 5, and conclusions in Sec. 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 309, + 395, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 395, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 395, + 322 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 329, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 486 + ], + "type": "text", + "content": "In this section, we position our contributions with respect to the relevant existing work. First, we introduce biases in visual datasets and computational approaches to the analysis of visual gender representation in films. We then discuss interpretive-level tasks, increasingly common in natural language processing, and the approaches to dataset creation to instantiate them for ML approaches. Here we highlight the scarcity of visual datasets made for high-level interpretive tasks, in particular for video data. Finally, we introduce the video understanding approaches that we consider to benchmark on our new interpretive task, specifically focusing on explainable concept-based approaches to locate the challenges ahead in video interpretation for this new task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 491, + 463, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 463, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 463, + 504 + ], + "type": "text", + "content": "2.1. Visual biases in film datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "The task we introduce is connected to the general problem of bias detection. As exposed by Fabbrizzi et al. [15], biases in visual datasets can be classified into selection bias (how subjects are included in a dataset), framing bias (how the visual content has been artificially composed) and label bias (errors or disparities in the labelling data). Our contribution is closely related to the framing and labelling biases. In film datasets, the first studies of biases in gender representation were from a presence quantification perspective. Guha et al. [21] and Somandepalli et al. [51] automatically estimate the screen time (from video) and speaking time (from audio) of male and female characters in Hollywood movies. They show that women are seen (36% screen time) and heard (41% screen time) significantly less than male characters. Mazieres et al. [39] consider a movie dataset spanning three decades, and show a temporal trend towards general fairer representation between both binary genders." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10865" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "content": "They however show simultaneously that the applied framings remain unfair, with only " + }, + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "content": " of one-face frames featuring a female (60% for males). Jang et al. specifically analyze the qualitative framing differences in gender portrayal in a dataset of 20 Hollywood movies and 20 Korean movies [26]. They find that female characters are portrayed with lower emotional diversity, spatial occupancy, temporal occupancy, intellectual image or mean age." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 286, + 251 + ], + "type": "text", + "content": "In contrast in this article, we take a first step towards detecting bias in gender representation from a high-level construct, objectification, qualitatively described in various disciplines such as cinematography [9, 41], social psychology [1, 32], and neuroscience [5-7]. Objectification is produced by complex temporal patterns never analyzed computationally in videos until now." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 262, + 274, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 274, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 274, + 275 + ], + "type": "text", + "content": "2.2. Interpretive-level tasks and dataset creation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 286, + 568 + ], + "type": "text", + "content": "At the same time that biases in visual datasets are uncovered and analyzed, other approaches aim to detect bias in human data. Detecting highly interpretive constructs, such as hate speech, propaganda, sexism, and racism, has been a long-standing endeavor in NLP. Da San Martino et al. [12] recruited to 4 experts to annotate news articles with text spans associated to 18 possible propaganda techniques. Samory et al. uncover the challenge of construct complexity for sexism detection [47]. They observe that multiple articles for automating this task consider widely different definitions of sexism, often referring to sub-dimensions of the broader construct. They consider existing works in social psychology where sexism dimensions have long been operationalized with sub-scale questions tested for consistency. To approach dataset creation for on-screen objectification, we inspire from these two last works, and on dense annotation approaches of image datasets recently proposed for the medical domain [13]. To the best of our knowledge, approaches for sexism detection from visual content are scarce, and almost none existing for video data. Two main types of visual content have been considered so far: hateful or sexist memes [16, 29], and sexist advertisements [18] which can also relate to symbolic advertisement understanding [25, 28]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "In films, analysis of biases in gender representation has been also automated with NLP approaches applied to film scripts [2, 37]. Specifically, Martinez et al. [37] propose a RNN-based model to automatically extract agent-verbal-patient triplets. From 912 movie scripts, they show that male characters are associated with a higher agency while female characters are more frequently the object of gaze. Su et al. introduce the more abstract task of trope understanding in movies [52]. Our work is close to this last one as we also introduce a dataset of films annotated for a higher-level construct beyond event and story understanding. However, we provide dense annotations of sequences with constituency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "tive concepts for the high-level construct of character objectification. Also, while Su et al. exploit an existing online base contributed to by the community, we design and carry out a strictly defined annotation process by experts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 372 + ], + "type": "text", + "content": "In this article, we set out from the concept of male gaze defined in various ways in film gender studies [9, 36, 41]. Mulvey [41] characterizes the concept of gaze by the three relations between the camera and the characters, between the characters, but also between the spectator and the characters. While some formalizations of gaze set out from the gender of the director [36], Brey defines female and male gaze only from the film content, with a corpus analysis focusing on aesthetics [9] and revolving around the construct of objectification. We consider this gaze analysis of Brey, which does not significantly rely the socio-historical context of production and reception. We build the operationalization objectification in both cinematography and psychology, to create the conditions to make a new challenge accessible to the CV community: we produce a strict annotation process to obtain densely annotated video data and analyze where the new challenges lie ahead. We generally position our approach producing a non-large scale but high-quality dataset that we hope will be useful within the lines of the call of Paullada et al. for such data-centric AI approaches [44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 382, + 545, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 396 + ], + "type": "text", + "content": "2.3. Approaches to video and movie understanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 401, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 497 + ], + "type": "text", + "content": "Cross-modal foundation models, such as CLIP [45] and ALIGN [27], learn aligned image and text representations through contrastive pre-training on large-scale closed datasets. Generalizations of the CLIP model to video data have included VideoCLIP [57] and two X-CLIP models [34, 42]. In particular, X-CLIP [42], which we employ in this article, expands CLIP with video temporal modeling and video-adaptive textual prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Legacy and large-scale pre-trained vision-language models have been leveraged for movie-related tasks. Bose et al. consider the difficulty of visual scene recognition in movies due to domain mismatch and create the MovieCLIP dataset obtained from weakly labeling movie shots from scene categories using the CLIP model [8]. An important vision-language movie-related task is audio-description for the visually-impaired, for which a major dataset introduced recently is MAD [50], gathering sparse natural language sentences grounded in over 1200 hours of movie videos. AutoAD [23] and AutoADII [22] are two recent approaches to generate audio-description from the video, both leveraging CLIP to learn to prompt GPT. To design approaches to learn human-level constructs, such as emotions, interactions or relationships, datasets generated with human supervision are also instrumental. A prominent representative is the MovieGraphs dataset, providing detailed annotations of clips of 51 movies with emotional states, character interac" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10866" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "tions and relationships, and other scene reasoning elements [53]. It has prompted works tackling such recognition tasks [17, 33]. In this article, we build on the MovieGraphs dataset to annotate a selection with the construct of objectification, to later analyze it in connection with the other annotated social elements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 335 + ], + "type": "text", + "content": "In this article, we aim to assess the capacity of CLIP-based methods to provide relevant embeddings for the concept of objectification. We do so with a direct evaluation of classification results when an adapter (MLP) fed by X-CLIP embeddings is learnt. We analyze the results with a concept-based approach by building on Concept Bottle-neck Models (CBM) [31]. Concept-based models are an active area of research in XAI, with works tackling the accuracy-explainability tradeoff [60] and the need for user-defined concepts [58]. We specifically employ Post-hoc CBM (PCBM) [59], which consists in learning a concept subspace (made of Concept Activation Vectors [30]) in the embedding space of the pre-trained model. Data samples are then projected in this concept subspace, from where the classification task can be performed with an interpretable classifier." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 350, + 156, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 156, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 156, + 363 + ], + "type": "text", + "content": "3. Data and methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 419 + ], + "type": "text", + "content": "This section presents our approach to create the first dataset for visual objectification in videos, specifically in films. We name this dataset ObyGaze12, short for ObjectifyingGaze12, which has the following highlights:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 420, + 287, + 597 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "text", + "content": "- It considers the multiple dimensions of the construct of visual objectification, made of filmic (framing and editing over successive shots, camera motion, etc.) and iconographic properties (visible objects, body parts, attire, character interactions, etc.)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 286, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 286, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 286, + 527 + ], + "type": "text", + "content": "- It is based on a thesaurus articulating five sub-constructs identified from multidisciplinary literature (film studies and psychology) from which we define typical instances, then grouped into coarse-grained visual concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 528, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 286, + 563 + ], + "type": "text", + "content": "- The data is annotated densely with concepts, and shows the multi-factorial property of objectification, corroborating with some recent developments in cognitive psychology [6]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 286, + 597 + ], + "type": "text", + "content": "- A hard negative category is included with the goal to perform fine-grained error analysis and improve model generalization." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 610, + 206, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 206, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 206, + 623 + ], + "type": "text", + "content": "3.1. A thesaurus of objectification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We first formalize the construct of visual objectification and derive key concepts to annotate in film scenes. Together with media studies experts, we identify five sub-constructs of objectification from literature on film cognition and film gender studies, and social and cognitive psychology: male gaze (point of view of a man on a woman) [5, 9, 41], sexualization [6, 7], surveillance of the feminine body [11, 14, 40]," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "female inaction / male possession [20, 48], and infantilism / animalization [41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 97, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 545, + 300 + ], + "type": "text", + "content": "These sub-constructs come with typical instances and examples from filmmaking techniques ([9, 41]) or validated questionnaires ([11, 40]), as shown in the middle and rightmost columns in Table 1. These typical instances are then grouped into eight coarse-grained visual concepts, corresponding to the possible means of production of visual objectification. They are shown in the left-most column in Table 1: type of shot (framing and gaze of camera), look (gaze of characters on the other), body (partial or full nudity, and sexually suggestive body parts), posture (connoting, e.g., childhood, submission or inaction), clothing (in relation to context and activities), appearance (age and makeup), expression of emotion (restrained or exaggerated according to gender role), and activities (linked to gender roles). Visual examples are provided in Fig. 1, where we show video samples of objectification concepts Look, Posture, Type of shot, Clothes and Activity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 308, + 394, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 394, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 394, + 319 + ], + "type": "text", + "content": "3.2. Data selection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 327, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 578 + ], + "type": "text", + "content": "Over the various existing movie datasets (see Sec. 2 and [4, 24, 39, 43, 46, 50, 53], [51, Table 1]), many have overlapping titles and only a few have rich human supervision. Amongst these, the MovieGraphs dataset [53] includes rich, high-level human annotations of 7637 clips of 51 movies, with emotional states, interactions and relationships, and other social reasoning elements. These elements are important and valuable in exploring social concepts of objectification in visual media. These movies are also frequently adopted for media research in existing work in [4, 24, 39]. The movie clips have also a short duration – mostly within 1-5 minutes – facilitating dense and granular annotations (over 100 clips for an average 2 hour film) while preserving the possibility to observe longer-term interactions and story development across a number of shots. From the 7637 clips of the MovieGraphs dataset, we select 1914 clips to annotate for objectification, amounting to " + }, + { + "bbox": [ + 304, + 327, + 545, + 578 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 578 + ], + "type": "text", + "content": " of the dataset and 12 complete movies, which were selected to approximately reproduce the fraction of genres in the original dataset. The list of selected films, year of release, and genre can be found in Table 3 in Appendix 7." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 586, + 404, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 404, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 404, + 597 + ], + "type": "text", + "content": "3.3. Data annotation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Every selected movie is annotated by at least two experts for objectification level and concepts over the movie scenes. Specifically, the annotators were asked to repeat a three-step process for every scene they deemed interesting from an objectification perspective: (1) watch the movie entirely and when they identify a scene worth annotating, (2) delimit the clip by using the cutting function in our annotation tool, and (3) assigning an objectification level and annotate the concept(s) involved in the objectification rating. We define" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10867" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 113, + 544, + 368 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Thesaurus of the typical instances and examples of visual objectification in films, grouped into eight main visual concepts used for annotation. Examples are possible means to produce one of the five sub-constructs of objectification (male gaze, sexualisation, surveillance of the feminine body, female inaction/male possession, infantilism/animalisation), to be assessed by annotators." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 113, + 544, + 368 + ], + "lines": [ + { + "bbox": [ + 47, + 113, + 544, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 113, + 544, + 368 + ], + "type": "table", + "html": "
ConceptConcept instancesExamples
Type of shotShot suggesting man perspective in presence of womanclose-up on a man's face; body parts of woman
Shot suggesting man gaze on womancamera takes the perspective of a male character with first close-up on the face followed by camera motion looking a woman from bottom up
Shot showing a woman in parallel with an animalwoman at same level and position with a dog
LookVoyerismcharacter watching another one without their knowledge
Non-reciprocal gazewoman looking at man who does not look back
BodySuggested nudityclothing on floor; silhouette behind shower curtain; nude shadow on wall
Partial nuditynude upper or lower body; partially open clothing or draping; in underwear
Full nuditynude person fully or partially shown
Body parts suggestive of sexclose-up shots on breast, buttocks, hips, or lips
PostureGesture or posture connoting seductionlip-biting; hip roll; twisting or tucking hair
Gesture or posture connoting sexualityeating phallic symbols; arching back
Gesture or posture connoting inactionbeing undressed by someone
Gesture or posture connoting submissionleaning on a man
Gesture or posture connoting dependencefollowing a man
Skippingskipping gait
ClothingWet or transparent clothingthin shirt soaked in rain
Clothing impractical to situationwear pumps for running, a skirt when gardening
Color code associated to characterwoman with pink clothing and accessories
Older woman wearing infantile clothingwoman wearing an Alice band or high socks
AppearanceDiscrepancy between appearance of woman and context or biographical elementsperfect makeup when waking up; mother of heroine too young; young girl played by older actress
Exp. of emotionAsymmetric expression of emotionboys don't cry; woman being hysterical
ActivityDoing domestic activitiesdoing laundry, cooking, cleaning, being constantly in the kitchen
", + "image_path": "e652fa9707f0ad109b21e28ddb1d14abf0c7a0875c09dda16c72eba1a31604a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 385, + 167, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 167, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 167, + 397 + ], + "type": "text", + "content": "four levels of objectification:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 287, + 529 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 399, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 434 + ], + "type": "text", + "content": "- Easy Negative: there are no elements suggestive of objectification. No concept can be annotated. Default value for watched but non-selected scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 470 + ], + "type": "text", + "content": "- Hard Negative: the scene contains elements of objectification from the thesaurus, but their presence does not result in an objectification effect." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "content": "- Not Sure: the scene indicates objectification, but does not completely fit the definition in the thesaurus." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 495, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 529 + ], + "type": "text", + "content": "- Sure: the scene contains elements of objectification from the thesaurus, and their presence results in an objectification effect." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "type": "text", + "content": "Four expert annotators were recruited to annotate the dataset. A first presentation session was held to introduce our annotation tool and the annotation procedure. All four annotators were then given the same two films – Juno and Silver Linings Playbook – to annotate using the proposed methodology. A second meeting was then set up after the annotation of the two films to analyze the reasons for divergence and remedy them by identify which elements of the annotation guidelines to clarify and how. We then randomly assigned two annotators to each of the remaining 10 films to annotate separately." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Data processing and fusion The data processing is described in detail in Appendix 7. Following the annotation step, the annotations are then projected from the delimi-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 385, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 564 + ], + "type": "text", + "content": "tations provided by each annotator onto the delimitations of the MovieGraphs clips. Since multiple annotations may overlap the same MovieGraphs clip, the annotation that is projected on the MovieGraphs clip corresponds to the annotation (including objectification level and concepts) with (1) the highest level of objectification that has at least " + }, + { + "bbox": [ + 304, + 385, + 545, + 564 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 564 + ], + "type": "text", + "content": " overlap with the MovieGraphs clip, and (2) when multiple annotations exist at the same level of objectification, the annotated concepts for these annotations are aggregated. The same process is used to aggregate the annotations of the annotators of a same clip: the maximum objectification level is kept, with possible aggregation of concepts in case both annotators chose the same level but annotated different concepts. The merged data is shared and used in the remaining of this article." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 574, + 484, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 484, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 484, + 586 + ], + "type": "text", + "content": "3.4. Analysis of the ObyGaze12 dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 593, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 629 + ], + "type": "text", + "content": "Here we comment on some interesting statistics of the resulting annotations and concepts of the 1914 clips originally delimited in the MovieGraphs dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "First, we verify data consistency by computing the inter-annotator agreement (IAA). Given the task of annotating timespans, we choose the " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " agreement measure introduced in [38] (and used for, e.g., annotating text spans [12]) owing to its consideration of temporal alignment, multiple annotators, and label classification at the same time. It attributes a score between 1 (complete agreement) and " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ". A value of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10868" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\gamma \\leq 0" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " indicates no agreement. The computation details of the " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " metric is provided in Appendix 7. Considering all four categories EN, HN, NS, S, we obtain and average " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\gamma = 0.42" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ". Not considering the clips annotated Not Sure (NS), which is the uncertain and \"noisy\" class in human annotations, the IAA increases to " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\gamma = 0.69" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ". This shows the consistency of the obtained annotations despite the interpretive nature of the task. Let us also mention that recent works improve learning approaches by considering explicitly the IAA in case of low number of annotators with moderate agreement [10, 55, 56]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "content": "Second, we analyze the obtained annotations in Fig 2. The Sure category is the least represented with " + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "content": ", the Easy Negative being, as expected, the most represented class with " + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "inline_equation", + "content": "52\\%" + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "content": " of clips. It is interesting to note that every concept is approximately annotated with the same rate throughout the Hard Negative, Not Sure and Sure levels of objectification. Finally, it is very interesting to observe that the average number of concepts annotated per clip increases with the level of objectification: 1.26 concepts on average per Hard Negative clip, 1.71 for Not Sure, up to 2.6 for Sure. We verify that this trend is observable for every single annotator. It gives an important insight into our video interpretation data: that objectification is a compositional process. This corroborates with recent findings in neuroscience experiments that found that a single element, such as clothing on its own, is not sufficient for people to perceive a character as an object [6]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 417, + 128, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 128, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 128, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 474 + ], + "type": "text", + "content": "The experiments have two objectives: to verify that the new classification task is feasible, and to identify the challenges of designing efficient models. To tackle these objectives, we" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 491, + 285, + 647 + ], + "blocks": [ + { + "bbox": [ + 50, + 491, + 285, + 647 + ], + "lines": [ + { + "bbox": [ + 50, + 491, + 285, + 647 + ], + "spans": [ + { + "bbox": [ + 50, + 491, + 285, + 647 + ], + "type": "image", + "image_path": "9c4f81a4703555bd1db8838a1fca58ac771c1c00e703a00353fcbf028e4b4419.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "type": "text", + "content": "Figure 2. Distribution of visual factors annotated for each level of objectification (HN = Hard negative, NS = Not sure, S = Sure). The percentage of the dataset for each level of objectification as well as the average number of concepts per clip are also shown. (Best viewed in colors)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "consider pre-trained vision models and specifically address the following research questions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 97, + 547, + 204 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 97, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 97, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 97, + 545, + 144 + ], + "type": "text", + "content": "- Task accuracy - What are the baseline performances by pre-trained vision models on the objectification detection task? How does the performance vary with hard negative examples?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 144, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 547, + 204 + ], + "type": "text", + "content": "- Concept representation - Can we implement interpretable models of objectification using concepts? What is the quality of representation of every concept, and what are the objectification concepts poorly captured by current models?" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 214, + 395, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 395, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 395, + 227 + ], + "type": "text", + "content": "4.1. Task accuracy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "text", + "content": "Setup We discard the Not Sure (NS) class from the ObyGaze12 dataset, as it gathers by definition samples highly uncertain for humans, and consider the Easy Negative (EN, " + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "text", + "content": " of the clip samples), Hard Negative (HN, " + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "text", + "content": ") and Sure (S, " + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 304, + 233, + 545, + 388 + ], + "type": "text", + "content": ") classes. We approach binary classification in a progressive way, the positive class being made of the S samples. We consider two levels of classification difficulty by composing the negative class either with EN samples, or with HN samples. The implementation details of cross-validation and data balancing are provided in Appendix 8. The average performance over the test set of the best models on validation folds are shown in Table 2 with standard deviations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 402, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 544 + ], + "type": "text", + "content": "Baselines We consider video embeddings obtained from pre-trained models owing to their zero-shot classification capabilities on video tasks. We select ViViT-B/16 [3], and the available X-CLIP model, trained on Kinetics [42]. We also re-train a X-CLIP model [34] on the LSMDC [46] film dataset, and refer to Appendix 10 for corresponding results, where all implementation details are described. We keep the pre-trained models frozen and perform an adaptative max pooling of the resulting frame tokens, and feed the output to an MLP made of 2 dense layers, the hidden layer with 128 neurons and ReLU activations, the last with 2 softmax neurons." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Results To assess the quality of the models on possibly imbalanced data with a minority of positive samples, we report the F1-scores in Table 2. First, by comparing with trivial classifiers (random and an all-positive, see App. 8), we observe that the task is indeed feasible, warranted by the data consistency described in Sec. 3 despite the interpretive nature of the task. Second, we observe that the inclusion of Hard Negative examples improves the classification results, showing the importance of a fine-grained annotation for highly-interpretive tasks. Results of X-CLIP on other configurations, specifically when the movies of clips in test are different from those in train, are shown in App. 8. The best results based on existing models are of moderate qual" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10869" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 112, + 288, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 2. F1-score on the binary task of objectification detection for models trained with easy or with hard negatives and tested on easy or all negative samples, with standard deviations." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 112, + 288, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 112, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 112, + 288, + 205 + ], + "type": "table", + "html": "
Test TrainEN vs. S(EN U HN) vs. S
EN vs. SHN vs. SEN vs. SHN vs. S
ViViT-B/160.53 (0.18)0.62 (0.13)0.54 (0.24)0.73 (0.1)
X-CLIP0.79 (0.05)0.71 (0.05)0.66 (0.05)0.82 (0.03)
Random0.540.55
All positive0.080.06
PCBM-DT0.680.440.580.38
PCBM-LR0.640.430.500.37
", + "image_path": "b5bfe66197df7d04f8c03a5d204622793d5dee11d73c033e06e11c389840e30d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 287, + 251 + ], + "type": "text", + "content": "ity, which calls for more investigation into where the difficulties lie." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 259, + 153, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 153, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 153, + 272 + ], + "type": "text", + "content": "4.2. Concept accuracy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 277, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 277, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 45, + 277, + 289, + 566 + ], + "type": "text", + "content": "To infer on-screen objectification, it is key for the model to detect the means of its production, which correspond to the eight concepts listed in Table 1. We reiterate that in the ObyGaze12 dataset, every clip annotated with a level of objectification S, NS or HN is also annotated with the presence of instances of the eight concepts. For example, if the Body concept is annotated, it means that some level of nudity and/or suggestive body parts are shown on screen, that could contribute to the production of objectification. The means of producing objectification through the eight concepts can be subtle to detect, making it difficult to provide the final interpretation. To investigate this difficulty, we implement Post-hoc Concept Bottleneck Models (PCBMs) [59], which allow us to approach a classification task with pre-trained models in an interpretable way when concept-annotated data is available. In our case, from the X-CLIP embedding space where our video clips are represented, we identify a Concept Activation Vector (CAV) [30] for every concept. We then project the X-CLIP embedding of every clip onto the subspace defined by the eight CAVs. The representation of the clip that is the output of this bottleneck is a low-dimensional vector with number-of-concepts components. This vector can then be fed to an interpretable classifier for the objectification detection task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "CAV computation For every concept " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", we collect two sets of samples: positive samples where concept " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " is present, hence made of S and HN samples with the concept annotated as present, and negative with EN, S and HN without concept " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ". We then train a linear SVM for each concept " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", the CAV of concept " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " being the normal vector of the SVM hyperplan. To train the SVM, we split the data in 10 folds, and reserve the last fold for test. We then perform an 8-fold cross-validation to select the SVM (choice of margin tolerance " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "), every fold training set being balanced with different draws of negative sub-sampling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Interpretable classifier We then train a decision tree (DT) and a logistic regression (LR) classifier on the same 8-fold cross-validation to classify the level of objectification, the classifiers being fed with the projection of every clip onto the CAVs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 146, + 546, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 546, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 546, + 303 + ], + "type": "text", + "content": "Task accuracy with PCBM We first verify the quality of objectification detection with F1-scores of PCBM-DT and PCBM-LR shown in Table 2. The results lower than X-CLIP are expected owing to the known accuracy-interpretability tradeoff of CBMs [59, 60]. They are above random and all-positive predictions when training on EN vs. S, which is indicative of the relevance of information held in the concepts. However, the low results obtained when training the DT and LR to distinguish between S and HN reveals the low quality of some CAVs, where the X-CLIP embeddings cannot be linearly well-separated for these concepts. We investigate this point next. The resulting DT is discussed in App. 9." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "text", + "content": "Concept accuracy We now analyze the quality of each obtained CAV by plotting its capability to classify whether the concept is present in a test sample. We consider a positive similarity between the X-CLIP embedding and CAV of concept " + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "text", + "content": " indicative of the presence of concept " + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "text", + "content": ". F1-score on the test set are shown in Fig. 3. Plots correspond to CAVs obtained from classifying the presence of concept " + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "text", + "content": " against EN only (solid bars) and against EN with S and HN without concept " + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 316, + 546, + 604 + ], + "type": "text", + "content": " (hatched bars). The former is used for PCBM-DT in Table 2. We first observe that concept detection is harder when negative samples also include S and HN samples (without the concept). This is expected considering that scenes tagged EN have by definition no element possibly conducive to objectification, and are hence likely to differ visually more from scenes where the concept is present, than do S and HN scenes without the concept. However within S and HN clips with and without the concept, such shortcuts cannot be exploited anymore. We observe in this case that the X-CLIP embedding related to concepts Type of shot, Posture, Look and Appearance are harder to separate linearly. This can be correlated with the analysis of factors of error detailed next. These subtler means of on-screen objectification therefore warrant future work to be properly captured and detected." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Error Analysis We analyze the factors impacting objectification classification errors corresponding to the results shown in Table 2. We select an average X-CLIP-based adaptation model trained on HN vs. S and tested on (ENUHN) vs. S, and consider its predictions on the clips in the test set. We label each test clip with 0 if the model fails to predict the correct label of the clip, and with 1 otherwise. We describe the clip with a one-hot encoding vector" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10870" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 281, + 269 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 281, + 269 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 281, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 281, + 269 + ], + "type": "image", + "image_path": "738fe22f7fad2dbd95c7f88aa8448030dc9708d00f061b53e7a5cec87d1bc493.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "text", + "content": "Figure 3. For every concept, F1-score of the best linear SVM selected to define the CAV of this concept. Positive samples (S and HN with the concept) must be separated from: [non-hatched bars] negative samples made of EN only, or [hatched bars] negative samples made of EN and S and HN without the concept." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 356, + 289, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 289, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 289, + 644 + ], + "type": "text", + "content": "corresponding to all 11 factors shown the y-axis of Fig. 4: every of the 8 concepts, and the Sure (S), Hard negative (HN) and Easy negative (EN) labels. We then train a logistic regression model on these clip descriptors to predict the failure/success labels. Fig. 4 shows the error regression weights associated with each factor, for two film examples. A negative weight indicates a contribution of the factor to a classification failure. We first observe that the HN characteristic contributes to a classification failure of the model, while S and EN contribute to classification success. Second, we observe that there is variability over the movies on the presence of which concept strongly influences the classification success. However, the presence of the Clothing concept seems to be a strong confounder. This can be due to the frequency of appearance of this concept in HN samples, and to the subtlety of the description of this concept (provided in Table 1), which should make it difficult for a pre-trained model to discriminate between an objectifying and non-objectifying overall label on the basis of Clothing. Third, it is worth noting that the concepts shown to be poorly linearly separable when described with the X-CLIP embeddings (see CAV analysis in Sec. 4.2), are also those with a non-stable contribution to the model errors over the film examples: Type of shot, Look, Posture and Appearance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 656, + 117, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 117, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 117, + 669 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Ethical aspect This work has an explicit societal motivation in its purpose to tackle, with the help of AI, the analysis of complex temporal patterns operated in cinema that pro" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 190 + ], + "type": "image", + "image_path": "605d1d403a836609b02b61a29651571e79fc53cce48fa2531709359423eb0a59.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 198, + 547, + 263 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 263 + ], + "type": "text", + "content": "Figure 4. Analysis of the factors of error for the objectification detection task: weights of a logistic regressor predicting whether the test set examples are well classified or not. Positive (resp. negative) weights indicate a positive (resp. negative) contribution to classification success. Left: As good as it gets, Right: Up in the Air." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 287, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 335 + ], + "type": "text", + "content": "duce the perception of certain characters as objects. This is a challenging but valuable task that aims to uncover and quantify differences in how various identities may be portrayed on screen." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 356, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 477 + ], + "type": "text", + "content": "Limitations and challenges A distinctive element of our work is the subjective judgement involved in annotating granular video elements for objectification. Video annotation is tedious, and approaching data annotation for such an interpretive task in a rigorous way is even more so, and difficult to scale. We therefore believe that pursuing high-quality, dense annotations with well-defined concepts goes a long way to tackle this new video interpretation task, which represents a valuable new challenge for the computer vision community." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 491, + 379, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 379, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 379, + 503 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 511, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 547, + 643 + ], + "type": "text", + "content": "In this article, we have introduced a new video interpretation task to detect character objectification in films. We have introduced the ObyGaze12 dataset, densely annotated by experts for objectification concepts defined from five sub-constructs identified in film studies and psychology. ObyGaze12 is made available to the community. We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. We show that the representation learning of the concepts of Type of shot, Look, Posture and Appearance need to be improved." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 658, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 547, + 703 + ], + "type": "text", + "content": "Acknowledgements. This work has been partly supported by the French National Research Agency through the ANR TRACTIVE project ANR-21-CE38-0012-01. This work was partly supported by EU Horizon 2020 project AI4Media, under contract no. 951911 (https://ai4media.eu/)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10871" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Operationalizing self-objectification: Assessment and related methodological issues. In Self-objectification in women: Causes, consequences, and counteractions., pages 23-49. American Psychological Association, Washington, 2011. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 235 + ], + "type": "text", + "content": "[2] Apoorv Agarwal, Jiehan Zheng, Shruti Kamath, Sriramkumar Balasubramanian, and Shirin Ann Dey. Key Female Characters in Film Have More to Talk About Besides Men: Automating the Bechdel Test. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 830-840, Denver, Colorado, 2015. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "text", + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, Montreal, QC, Canada, 2021. IEEE. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed Movies: Story Based Retrieval with Contextual Embeddings. Technical Report arXiv:2005.04208, arXiv, 2020. arXiv:2005.04208 [cs] type: article. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 411 + ], + "type": "text", + "content": "[5] Philippe Bernard, Sarah J. Gervais, and Olivier Klein. Objectifying objectification: When and why people are cognitively reduced to their parts akin to objects. European Review of Social Psychology, 29(1):82-121, 2018. Publisher: Routledge _eprint: https://doi.org/10.1080/10463283.2018.1471949. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 413, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 479 + ], + "type": "text", + "content": "[6] Philippe Bernard, Florence Hanoteau, Sarah Gervais, Lara Servais, Irene Bertolone, Paul Deltenre, and Cecile Colin. Revealing Clothing Does Not Make the Object: ERP Evidences That Cognitive Objectification is Driven by Posture Suggestiveness, Not by Revealing Clothing. Personality and Social Psychology Bulletin, 45(1):16-36, 2019. 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 480, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 533 + ], + "type": "text", + "content": "[7] Philippe Bernard, Carlotta Cogoni, and Andrea Carnaghi. The Sexualization-Objectification Link: Sexualization Affects the Way People See and Feel Toward Others. Current Directions in Psychological Science, 29(2):134-139, 2020. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 535, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 287, + 601 + ], + "type": "text", + "content": "[8] Digbalay Bose, Rajat Hebbar, Krishna Somandepalli, Haoyang Zhang, Yin Cui, Kree Cole-McLaughlin, Huisheng Wang, and Shrikanth Narayanan. Movieclip: Visual scene recognition in movies. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2082-2091, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 602, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 624 + ], + "type": "text", + "content": "[9] Iris Brey. Le regard féminin-Une révolution à l'écran. Média Diffusion, 2020. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 690 + ], + "type": "text", + "content": "[10] M. Bucarelli, L. Cassano, F. Siciliano, A. Mantrach, and F. Silvestri. Leveraging inter-rater agreement for classification in the presence of noisy labels. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3439-3448, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[11] Rachel M. Calogero. A Test of Objectification Theory: The Effect of the Male Gaze on Appearance Concerns in Col-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "lege Women. Psychology of Women Quarterly, 28(1):16-21, 2004. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 95, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 183 + ], + "type": "text", + "content": "[12] Giovanni Da San Martino, Seunghak Yu, Alberto Barrón-Cedeno, Rostislav Petrov, and Preslav Nakov. Fine-Grained Analysis of Propaganda in News Article. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5635–5645, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 184, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 249 + ], + "type": "text", + "content": "[13] Roxana Daneshjou, Mert Yuksekgonul, Zhuo Ran Cai, Roberto A. Novoa, and James Zou. Skincon: A skin disease dataset densely annotated by domain experts for fine-grained debugging and analysis. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[14] Angela Denchik. Development and Psychometric Evaluation of the Interpersonal Sexual Objectification Scale. PhD thesis, The Ohio State University, 2005. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "text", + "content": "[15] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223: 103552, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "text", + "content": "[16] Elisabetta Fersini, Francesca Gasparini, and Silvia Corchs. Detecting Sexist MEME On The Web: A Study on Textual and Visual Cues. In 2019 8th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIw), pages 226-231, 2019. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 383, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 448 + ], + "type": "text", + "content": "[17] Bei Gan, Xiujun Shu, Ruizhi Qiao, Haoqian Wu, Keyun Chen, Hanjun Li, and Bohan Ren. Collaborative noisy label cleaner: Learning scene-aware trailers for multi-modal highlight detection in movies. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18898-18907, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "type": "text", + "content": "[18] Francesca Gasparini, Ilaria Erba, Elisabetta Fersini, and Silvia Corchs. Multimodal Classification of Sexist Advertisements.: In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications, pages 399-406, Porto, Portugal, 2018. SCITEPRESS - Science and Technology Publications. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "text", + "content": "[19] Francesca Gasparini, Giulia Rizzi, Aurora Saibene, and Elisabetta Fersini. Benchmark dataset of memes with text transcriptions for automatic detection of multi-modal misogynistic content. arXiv:2106.08409 [cs], 2021. arXiv:2106.08409.2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[20] Sarah J. Gervais, Gemma Sáez, Abigail R. Riemer, and Olivier Klein. The Social Interaction Model of Objectification: A process model of goal-based objectifying exchanges between men and women. British Journal of Social Psychology, 59(1):248-283, 2020. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 625, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 689 + ], + "type": "text", + "content": "[21] Tanaya Guha, Che-Wei Huang, Naveen Kumar, Yan Zhu, and Shrikanth S. Narayanan. Gender Representation in Cinematic Content: A Multimodal Approach. In Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, pages 31–34, Seattle Washington USA, 2015. ACM. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel - who," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10872" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13645-13655, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie Description in Context. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18930-18940, Vancouver, BC, Canada, 2023. IEEE. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 218 + ], + "type": "text", + "content": "[24] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. MovieNet: A Holistic Dataset for Movie Understanding. In Computer Vision – ECCV 2020, pages 709–727. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "text", + "content": "[25] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic Understanding of Image and Video Advertisements. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1100-1110, Honolulu, HI, 2017. IEEE. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[26] Ji Yoon Jang, Sangyoon Lee, and Byungjoo Lee. Quantification of Gender Representation Bias in Commercial Films based on Image Analysis. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW):1-29, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "text", + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "text", + "content": "[28] Nasrin Kalanat and Adriana Kovashka. Symbolic image detection using scene and knowledge graphs. Technical Report arXiv:2206.04863, arXiv, 2022. arXiv:2206.04863 [cs] type: article. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 509 + ], + "type": "text", + "content": "[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 567 + ], + "type": "text", + "content": "[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie J. Cai, James Wexler, Fernanda B. Viégas, and Rory Sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International Conference on Machine Learning, 2017. 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "text", + "content": "[31] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In Proceedings of the 37th International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "text", + "content": "[32] Holly B. Kozee, Tracy L. Tylka, and Casey L. Augustus-Horvath. Interpersonal Sexual Objectification Scale, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[33] Anna Kukleva, Makarand Tapaswi, and Ivan Laptev. Learning Interactions and Relationships Between Movie Characters. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9846-9855, Seattle, WA, USA, 2020. IEEE. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "text", + "content": "[34] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, page 638–647, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 162 + ], + "type": "text", + "content": "[35] Sarah Macharia. Global Media Monitoring Project (GMMP), pages 1-6. John Wiley Sons, Ltd, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "text", + "content": "[36] Alicia Malone. The Female Gaze: Essential Movies Made by Women. Mango Publisher, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "text", + "content": "[37] Victor R. Martinez, Krishna Somandepalli, and Shrikanth Narayanan. Boys don't cry (or kiss or dance): A computational linguistic lens into gendered actions in film. PLOS ONE, 17(12):e0278604, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[38] Yann Mathet, Antoine Widlocher, and Jean-Philippe Métivier. The unified and holistic method gamma " + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": " for inter-annotator agreement measure and alignment. Computational Linguistics, 41(3):437-479, 2015. 5, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[39] Antoine Mazieres, Telmo Menezes, and Camille Roth. Computational appraisal of gender representativeness in popular movies. Humanities and Social Sciences Communications, 8 (1):137, 2021. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 322, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 354 + ], + "type": "text", + "content": "[40] Nita Mary McKinley and Janet Shibley Hyde. The objectified body consciousness scale: Development and validation. Psychology of women quarterly, 20(2):181-215, 1996. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 355, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 376 + ], + "type": "text", + "content": "[41] Laura Mulvey. Visual Pleasure and Narrative Cinema. Screen, 16(3):6-18, 1975. 1, 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 378, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 433 + ], + "type": "text", + "content": "[42] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 499 + ], + "type": "text", + "content": "[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. MovieCuts: A New Dataset and Benchmark for Cut Type Recognition. In Computer Vision – ECCV 2022, pages 668–685. Springer Nature Switzerland, Cham, 2022. Series Title: Lecture Notes in Computer Science. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "type": "text", + "content": "[44] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 546, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 623 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[46] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie Description. International Journal of Computer Vision, 123(1):94-120, 2017. 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[47] Mattia Samory, Indira Sen, Julian Kohne, Fabian Flock, and Claudia Wagner. \"Call me sexist, but...\" : Revisiting Sexism Detection Using Psychological Scales and Adversarial Samples. page 12. 1, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10873" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[48] Maarten Sap, Marcella Cindy Prasettio, Ari Holtzman, Hannah Rashkin, and Yejin Choi. Connotation frames of power and agency in modern films. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 2329-2334, 2017. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[49] Alexandra Schofield and Leo Mehr. Gender-Distinguishing Features in Film Dialogue. In Proceedings of the Fifth Workshop on Computational Linguistics for Literature, pages 32–39, San Diego, California, USA, 2016. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "type": "text", + "content": "[50] M. Soldan, A. Pardo, J. Alcazar, F. Heilbron, C. Zhao, S. Giancola, and B. Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5016-5025, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "text", + "content": "[51] Krishna Somandepalli, Tanaya Guha, Victor R. Martinez, Naveen Kumar, Hartwig Adam, and Shrikanth Narayanan. Computational Media Intelligence: Human-Centered Machine Analysis of Media. Proceedings of the IEEE, 109(5): 891-910, 2021. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 391 + ], + "type": "text", + "content": "[52] Hung-Ting Su, Po-Wei Shen, Bing-Chen Tsai, Wen-Feng Cheng, Ke-Jyun Wang, and Winston H. Hsu. TrUMAn: Trope Understanding in Movies and Animations. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 4594–4603, New York, NY, USA, 2021. Association for Computing Machinery. event-place: Virtual Event, Queensland, Australia. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 392, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 287, + 446 + ], + "type": "text", + "content": "[53] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. MovieGraphs: Towards Understanding Human-Centric Situations from Videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[54] Angelina Wang, Arvind Narayanan, and Olga Russakovsky. REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets. In Computer Vision – ECCV 2020, pages 733–751. Springer International Publishing, Cham, 2020. Series Title: Lecture Notes in Computer Science. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[55] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. ArXiv, abs/2110.12088, 2021. 6, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 548, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 624 + ], + "type": "text", + "content": "[56] Jiaheng Wei, Zhaowei Zhu, Tianyi Luo, Ehsan Amid, Abhishek Kumar, and Yang Liu. To Aggregate or Not? Learning with Separate Noisy Labels. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2523-2535, New York, NY, USA, 2023. Association for Computing Machinery. event-place: Long Beach, CA, USA. 6, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 625, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 713 + ], + "type": "text", + "content": "[57] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 250 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "text", + "content": "[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19187-19197, Vancouver, BC, Canada, 2023. IEEE. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "text", + "content": "[59] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR^2Struct: Privacy, Accountability, Interpretability, Robustness, Reasoning on Structured Data, 2022. 4, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 184, + 546, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 546, + 250 + ], + "type": "text", + "content": "[60] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models. In Advances in Neural Information Processing Systems, 2022. 4, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10874" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_content_list.json b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a686f4e66f367faef89d37646d9a766754fcda29 --- /dev/null +++ b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_content_list.json @@ -0,0 +1,1792 @@ +[ + { + "type": "text", + "text": "Visual Point Cloud Forecasting enables Scalable Autonomous Driving", + "text_level": 1, + "bbox": [ + 135, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zetong Yang Li Chen Yanan Sun Hongyang Li", + "bbox": [ + 267, + 180, + 699, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OpenDriveLab and Shanghai AI Lab", + "bbox": [ + 336, + 205, + 630, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/OpenDriveLab/ViDAR", + "bbox": [ + 303, + 229, + 671, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72cb6304fddd876554924d65bab7ee67995c4611bf875095a360c61669bcec58.jpg", + "image_caption": [ + "Figure 1. ViDAR is a visual autonomous driving pre-training framework, which leverages the estimation of future point clouds from historical visual inputs as the pre-text task. We term this new pre-text task as visual point cloud forecasting. With the aid of ViDAR, we achieve substantial improvement spanning a diverse spectrum of downstream applications for perception, prediction, and planning." + ], + "image_footnote": [], + "bbox": [ + 78, + 257, + 895, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 553, + 313, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to extensive studies on general vision, pretraining for scalable visual autonomous driving remains seldom explored. Visual autonomous driving applications require features encompassing semantics, 3D geometry, and temporal information simultaneously for joint perception, prediction, and planning, posing dramatic challenges for pre-training. To resolve this, we bring up a new pre-training task termed as visual point cloud forecasting - predicting future point clouds from historical visual input. The key merit of this task captures the synergic learning of semantics, 3D structures, and temporal dynamics. Hence it shows superiority in various downstream tasks. To cope with this new problem, we present ViDAR, a general model to pretrain downstream visual encoders. It first extracts historical embeddings by the encoder. These representations are then transformed to 3D geometric space via a novel Latent Rendering operator for future point cloud prediction. Experiments show significant gain in downstream tasks, e.g., $3.1\\%$ NDS on 3D detection, $\\sim 10\\%$ error reduction on motion forecasting, and $\\sim 15\\%$ less collision rate on planning.", + "bbox": [ + 73, + 585, + 473, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 553, + 632, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the community has witnessed rapid development in visual, or camera-only autonomous driving, with input being monocular or multi-view images [50, 63, 69, 70, 76, 89]. Leveraging visual inputs, existing approaches demonstrate superior capability of extracting Bird's-Eye-View (BEV) features [6, 28, 45, 52, 61, 88], and performing well in perception [29, 51, 54, 67], prediction [17, 27, 65], and planning [30, 31]. Despite significant improvements in applications, these models rely on precise 3D annotations to a great extent, which are often difficult to collect, e.g., occupancy [3, 10], 3D boxes [5, 16, 64], trajectories [15], and thus are challenging to scale up for production.", + "bbox": [ + 496, + 580, + 893, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Considering the expensive labeling workflow, pretraining [2, 18, 62] has emerged as a crucial approach to scale up downstream applications. The key idea is to define pretext tasks that leverage large amounts of readily available data to learn meaningful representations. This enhances downstream performance though labeled data is limited.", + "bbox": [ + 496, + 763, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Though extensive research of pre-training in computer vision has been conducted [8, 20-22, 33, 66, 78, 79], its application in visual autonomous driving remains seldom", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "explored. Visual autonomous driving poses great challenges in pre-training as it requires the features to maintain semantics, 3D geometry, and temporal dynamics at the same time for joint perception, prediction, and planning [7, 77]. As a result, most models still rely on supervised pre-training, like 3D detection [60, 69] or occupancy [58, 67, 81], using labeled data that is often unavailable at scale [39]. Some approaches propose estimating depth [60] or rendering masked scenes [82] as pre-training. They use Image-LiDAR pairs as a means of achieving scalable annotation-free pre-training. However, they struggle in either multi-view 3D geometry or temporal modeling (Figure 2). Depth estimation retrieves depth from one image, limited in multi-view geometry; rendering techniques reconstruct scenes from multi-view images but lacking temporal modeling. However, temporal modeling is crucial in end-to-end autonomous driving systems, e.g., UniAD [27], especially for prediction and planning which are the ultimate goals and require accurate scene flow and object motion for decision-making. Due to the absence of temporal modeling, existing approaches are insufficient for pretraining the end-to-end system.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore pre-training tailored for end-to-end visual autonomous driving applications, including not only perception but also prediction and planning [7]. We formulate a new pre-text task, visual point cloud forecasting (Figure 2), to fully exploit information of semantics, 3D geometry, and temporal dynamics behind the raw Image-LiDAR sequences, with being scalable into consideration. It predicts future point clouds from historical visual images.", + "bbox": [ + 76, + 426, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main rationale of visual point cloud forecasting lies in the simultaneous supervision of semantics, 3D structure, and temporal modeling. By compelling the model to predict the future from history, it supervises the extraction of scene flow and object motion which are crucial for temporal modeling and future estimation. Meanwhile, it involves the reconstruction of point clouds from images, which supervises the multi-view geometry and semantic modeling. Therefore, features from visual point cloud forecasting embed information of both geometric and temporal hints, beneficial for perception, tracking, and planning simultaneously.", + "bbox": [ + 76, + 550, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present ViDAR, a general visual point cloud forecasting approach for pre-training (Figure 2). ViDAR includes three parts, History Encoder, Latent Rendering operator, and Future Decoder. The History Encoder is the target structure for pre-training. It could be any visual BEV encoder [45] to embed visual sequences into BEV space. These BEV features are sent to the Latent Rendering operator. Latent Rendering plays a crucial role in enabling ViDAR benefit downstream performance. It solves the ray-shaped BEV features issue [46, 90], models 3D geometric latent space, and bridges encoder and decoder. The Future Decoder is an auto-regressive transformer that takes histor", + "bbox": [ + 76, + 719, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e2e1fa4cb319e04f3374b2a35ab5b3220e7182d12c0c67ba65e33fd367d70464.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 694, + 148 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/babfffed185465b47f0547f3ecd338f53020387c1f00539a1c3e95ce6e60e2e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 888, + 147 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79a99e36f19e574617e2a210beb8cdbe7787512e7f28458008c205b34d4f5815.jpg", + "image_caption": [ + "Figure 2. Comparisons among visual autonomous driving pretraining paradigms and our ViDAR architecture. Compared to existing methods, visual point cloud forecasting jointly models multi-view geometry and temporal dynamics. We then propose ViDAR, using Image-LiDAR sequences to pre-train visual encoders." + ], + "image_footnote": [], + "bbox": [ + 503, + 148, + 888, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ical BEV features to iteratively predict future point clouds for arbitrary timestamps.", + "bbox": [ + 498, + 361, + 890, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ViDAR provides a comprehensive solution for visual autonomous driving pre-training. We test ViDAR on nuScenes dataset [5] in terms of point cloud forecasting and downstream verifications. Though with visual inputs, ViDAR still outperforms previous forecasting methods using point clouds, $\\sim 33\\%$ Chamfer Distance reduction on point cloud estimation of 1s future. ViDAR also improves downstream performance. Using Image-LiDAR sequences only, ViDAR surprisingly outperforms 3D detection pre-training [69], e.g., by $1.1\\%$ mAP and $2.5\\%$ mIoU for detection and semantic occupancy prediction, under the same data scale. If also based on the 3D detection pre-training, ViDAR boosts previous methods by $4.3\\%$ mAP and $4.6\\%$ mIoU. Further, due to the effective pre-training on the joint capture of geometric and temporal information, ViDAR improves UniAD [27] on all tasks for end-to-end autonomous driving including perception, prediction, and planning by a large margin (Figure 1). Experimental results validate that visual point cloud forecasting enables scalable autonomous driving.", + "bbox": [ + 498, + 391, + 892, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 691, + 640, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-training for Visual Autonomous Driving. Pretraining for scalable applications has been extensively studied in general vision. These approaches can be roughly divided into contrastive approaches [8, 20, 34, 66], which learn discriminative features from positive and negative pairs; and masked signal modeling [13, 21, 73, 80], which recover discarded signals from remained signals to capture a comprehensive understanding of global semantics.", + "bbox": [ + 498, + 719, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, pre-training for visual autonomous driving is still under-explored. Visual autonomous driving poses great challenges as it requires semantic understanding, 3D structural awareness, and temporal modeling at the same time,", + "bbox": [ + 498, + 840, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14674", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for joint perception, prediction, and planning. Existing vision methods mainly consider semantics; methods based on Image-LiDAR pairs [60, 82] struggle with temporal modeling; other supervised strategies [67, 69] are not scalable. In this work, we propose visual point cloud forecasting, which simultaneously models semantics, temporal dynamics, and 3D geometry by a uniform process, and is easily scaled up.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Forecasting. Point cloud forecasting, one of the most fundamental self-supervised tasks for autonomous driving, predicts future point clouds from past point cloud inputs. Previous works use range image [4, 57], a representation obtained by projecting point clouds to dense 2D images using sensor intrinsic and extrinsic parameters. Based on historical range images, they apply 3D convolutions [56], or LSTMs [74, 75] to predict future point clouds. Yet, they additionally model the motion of sensor intrinsic and extrinsic parameters. Later methods factor out the estimation of sensors by introducing 4D occupancy prediction [36] and differentiable ray-casting [35], which ensures a better modeling of the world. Compared to prior literature, we aim at visual point cloud forecasting, using past images to predict future point clouds. Meanwhile, we raise this task as a pre-training paradigm for visual autonomous driving and demonstrate its superiority in a wide range of downstream applications.", + "bbox": [ + 75, + 198, + 470, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 481, + 210, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on our ViDAR, a visual point cloud forecasting approach for general autonomous driving pre-training. We begin with an overview of ViDAR in Section 3.1, and subsequently delve into Latent Rendering and Future Decoder in Section 3.2 and Section 3.3, respectively.", + "bbox": [ + 75, + 506, + 468, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 589, + 187, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Figure 2, ViDAR comprises three components: (a) a History Encoder, also the target structure of pre-training, which extracts BEV embeddings $\\mathcal{F}_{\\mathrm{bev}}$ from visual sequence inputs $\\mathcal{I}$ ; It can be any visual BEV encoder [28, 45, 52]; (b) a Latent Rendering operator, which simulates the volume rendering operation in latent space so as to obtain geometric embedding $\\hat{\\mathcal{F}}_{\\mathrm{bev}}$ from $\\mathcal{F}_{\\mathrm{bev}}$ ; and (c) a Future Decoder, which predicts future BEV features $\\hat{\\mathcal{F}}_t$ at timestamps $t \\in \\{1, 2, \\ldots\\}$ in an auto-regressive manner. Finally, a prediction head is followed to project $\\hat{\\mathcal{F}}_t$ into 3D occupancy volume $\\mathcal{P}_t$ . This process is formulated as:", + "bbox": [ + 75, + 612, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {b e v}} = \\operatorname {E n c o d e r} (\\mathcal {I}),\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 785, + 274, + 800 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {\\text {b e v}} = \\text {L a t e n t R e n d e r} (\\mathcal {F} _ {\\text {b e v}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 806, + 468, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {t} = \\operatorname {D e c o d e r} \\left(\\hat {\\mathcal {F}} _ {t - 1}\\right), \\text {w h e r e} \\hat {\\mathcal {F}} _ {0} = \\hat {\\mathcal {F}} _ {\\text {b e v}},\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 825, + 424, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {t} = \\text {P r o j e c t i o n} (\\hat {\\mathcal {F}} _ {t}).\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 848, + 310, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point cloud predictions are obtained from the predicted occupancy volume $\\mathcal{P}_t$ . This process is similar to the previous", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "point cloud forecasting method [36]. Specifically, we first cast rays from the origin to various designated directions, then figure out the distance of waypoints along each ray with the maximum occupancy response, and finally compute the point position according to the distance and corresponding ray direction.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Latent Rendering", + "text_level": 1, + "bbox": [ + 500, + 191, + 671, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A straightforward solution of visual point cloud forecasting for pre-training is to incorporate the History Encoder and Future Decoder directly with differentiable ray-casting [36], which is the crucial component in state-of-the-art point cloud forecasting methods to render point clouds from predicted occupancy volume and compute loss for backpropagation. However, our experimental results show that this approach does not yield improvements and even has a detrimental effect on the downstream tasks due to the defective geometric feature modeling ability.", + "bbox": [ + 496, + 215, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. Differentiable ray-casting is a volume rendering process operating on an occupancy volume, denoted as $\\mathcal{P} \\in \\mathbb{R}^{L \\times H \\times W}$ . It renders depths of various rays and subsequently converts the depths with corresponding ray directions to point clouds.", + "bbox": [ + 496, + 386, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, starting from the origin sensor position, $\\mathbf{o} \\in \\mathbb{R}^3$ , differentiable ray-casting casts $n$ rays with varying directions, $\\mathbf{d} \\in \\mathbb{R}^{n \\times 3}$ . Along each ray $i$ , it uniformly samples $m$ waypoints at different distances $\\lambda^{(j)} \\in \\mathbb{R}, j \\in \\{1,2,\\dots,m\\}$ until reaching the boundary of the 3D space. The coordinates of these waypoints are calculated as:", + "bbox": [ + 498, + 462, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {(i, j)} = \\mathbf {o} + \\lambda^ {(j)} \\mathbf {d} ^ {(i)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 564, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Those waypoints coordinates, $\\mathbf{x} \\in \\mathbb{R}^{n \\times m \\times 3}$ , are used to compute occupancy values. This process is quantized [36], wherein the waypoints are discretized to occupancy volume grids. Then, the values of waypoints are derived as the associated values of volume grids, $\\mathbf{p}^{(i,j)} = \\mathcal{P}^{([\\mathbf{x}^{(i,j)}])}$ . Here $[\\cdot]$ denotes a rounding operation for discretizing waypoints.", + "bbox": [ + 496, + 595, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable ray-casting renders the corresponding depth of the $i$ -th ray, $\\hat{\\lambda}^{(i)}$ , by an integral process:", + "bbox": [ + 496, + 688, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {p}} ^ {(i, j)} = \\left[ \\prod_ {k = 1} ^ {j - 1} (1 - \\mathbf {p} ^ {(i, k)}) \\right] \\mathbf {p} ^ {(i, j)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 731, + 890, + 772 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\lambda} ^ {(i)} = \\sum_ {j = 1} ^ {m} \\hat {\\mathbf {p}} ^ {(i, j)} \\lambda^ {(j)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 777, + 890, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For simplicity, we name Eq. 3 and Eq. 4 as conditional probability function and distance expectation function. The conditional probability function determines the occupancy of a grid by considering the conditional probability that prior waypoints are unoccupied and the ray terminates at", + "bbox": [ + 496, + 825, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e757736429c0ca6d831c31ef263a2fd76473ac87aedea29fa2a46a54a5fb3119.jpg", + "image_caption": [ + "(a.) Ray-shaped Features" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 205, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c2361424183af82592f556eb04723e204ee444e99696278d01300a7ca464b9e7.jpg", + "image_caption": [ + "(b.) Geometric Features", + "Figure 3. Ray-shaped Features vs. Geometric Features. Ray-shaped features show similar feature responses on BEV grids along the same ray; while geometric features from the Latent Rendering maintain discriminative 3D geometry and can describe the 3D world in latent space." + ], + "image_footnote": [], + "bbox": [ + 218, + 88, + 326, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d5103f45b2db68050a79dbcadb34ca08671ee1f034d235325136db0a20ba0d5.jpg", + "image_caption": [ + "(c.) GT Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 343, + 90, + 446, + 169 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9dd37b522cbb4b38e23a6f3c92514bccb1b7f9feef93cd24fa62fcbfcbd68e42.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Forecasting StructureN/ADifferentiable Ray-castingLatent Rendering
NDS (%)44.1140.20 (-3.91)47.58 (+3.47)
", + "bbox": [ + 83, + 268, + 460, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Downstream detection performance under different forecasting structures. \"N/A\" represents the baseline without forecasting pre-training. We observe a performance drop when directly using History Encoder and Future Decoder with differentiable ray-casting for pre-training; while, with the Latent Rendering operator, the performance is significantly improved.", + "bbox": [ + 75, + 321, + 468, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this particular grid; the distance expectation function retrieves depths from the occupancy of grids in 3D volume. L1 loss is then applied to supervise the rendered depth for training point cloud forecasting.", + "bbox": [ + 75, + 431, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the great success of differentiable ray-casting in the task of point cloud forecasting, its application in visual point cloud forecasting pre-training does not bring any benefit for downstream performance (Table 1). After such pretraining, ray-shaped features [46, 90], where grids along the same ray tend to possess similar features (Figure 3 - (a.)), are observed. The underlying reason is that waypoints along the same ray in 3D space usually correspond to the same pixel in the visual image, resulting in a tendency to learn similar feature responses. As a consequence, these ray-shaped features are not discriminative and representative enough when transferred to downstream applications, leading to reduced performance.", + "bbox": [ + 75, + 492, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Rendering. In order to extract more discriminative and representative features, we introduce the Latent Rendering operator. It first computes the ray-wise feature through a feature expectation function, then customizes features of each grid by weighting the ray-wise feature with its associated conditional probability. The overall structure is depicted in Figure 4.", + "bbox": [ + 75, + 690, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To be specific, inspired by Eq. (4), the feature expectation function is formulated in a similar form:", + "bbox": [ + 76, + 796, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} ^ {(i)} = \\sum_ {k = 1} ^ {m} \\hat {\\mathbf {p}} ^ {(i, k)} \\mathcal {F} _ {\\mathrm {b e v}} ^ {(k)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 835, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ represents the ray extending from the origin to the", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe47be17e5e3c2f68533ab0d6f8a62b1781e37336e1b920caf564102b16c74ec.jpg", + "image_caption": [ + "Figure 4. Multi-group Latent Rendering comprises several Latent Rendering running in parallel for different channels. Latent Rendering captures geometric features by the conditional probability function and the feature extraction function. “ $\\bigoplus$ ” means concatenating multi-group features among channel dimensions." + ], + "image_footnote": [], + "bbox": [ + 503, + 94, + 890, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$i$ -th BEV grid. Here, $\\hat{\\mathbf{p}}$ is the conditional probability, computed through the conditional probability function Eq. (3), which takes as input the learnable independent probability projected from $\\mathcal{F}_{bev}$ . The ray-wise features are shared by all grids lying in the same ray.", + "bbox": [ + 496, + 281, + 890, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we compute the grid feature as:", + "bbox": [ + 517, + 358, + 769, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {\\mathrm {b e v}} = \\hat {\\mathbf {p}} \\cdot \\hat {\\mathcal {F}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 383, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which highlights the response of BEV grids with higher conditional probability so as to make $\\hat{\\mathcal{F}}_{bev}$ discriminative. This enables the BEV encoder to learn the geometric features during pre-training (Figure 3 - (b)).", + "bbox": [ + 496, + 412, + 890, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the diversity of geometric features, we further design the multi-group Latent Rendering. By parallelizing multiple Latent Rendering on different feature channels, we allow ray-wise features to maintain diverse information, leading to better downstream performance.", + "bbox": [ + 496, + 474, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in Eq. (3), the conditional probability of each BEV grid is determined not only by its own independent response, but also by the response of all its prior grids. Consequently, in the pre-training phase, once the model raises the response of a particular BEV grid, the corresponding responses of all its prior and subsequent grids are suppressed, which mitigates the issue of ray-shaped features during pre-training. After the pre-training with Latent Rendering, it is generally observed that there are only a few peaks with higher responses on a specific ray, indicating the presence of objects or structures in the scene. This effectively promotes a more accurate and consistent understanding of the 3D environment.", + "bbox": [ + 496, + 549, + 892, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Future Decoder", + "text_level": 1, + "bbox": [ + 500, + 755, + 656, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Future Decoder predicts the next BEV features $\\hat{\\mathcal{F}}_t$ of frame $t$ based on the inputs of previous BEV latent space $\\hat{\\mathcal{F}}_{t - 1}$ and the expected next ego-motion, $\\mathbf{e}_t$ . The predicted features are then used to generate point clouds as Eq. (1).", + "bbox": [ + 496, + 777, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. As depicted in Figure 5, Future Decoder is a transformer that can be iteratively used to predict future BEV features from the last roll-out embeddings in an auto-regressive manner. In the $t$ -th iteration, it first encodes", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14676", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c4188f8cdd74035c8cb8c0d03b7f996d3e73b5526d9d6d9b3dd0a1e6c9b7765.jpg", + "image_caption": [ + "Figure 5. Future Decoder iteratively predicts the next BEV features, $\\hat{\\mathcal{F}}_t$ , from the conditions of ego-motion $\\mathbf{e}_t$ and the last BEV features, to enable specific future predictions with any ego-control." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 472, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the ego-motion condition $\\mathbf{e}_t$ , which describes the expected coordinates and heading of ego-vehicle in the next frame, into high-dimensional embeddings by multi-layer perceptron (MLP), which are then added to future BEV queries as inputs of the transformer. Then, 6 transformer layers, composed of a Deformable Self-Attention [93], a Temporal Cross-Attention and a FeedForward Network [14], are used to predict the future $\\hat{\\mathcal{F}}_t$ based on the condition and the last BEV features $\\hat{\\mathcal{F}}_{t - 1}$ .", + "bbox": [ + 75, + 313, + 468, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Temporal Cross-Attention layer follows the design of Deformable Cross-Attention [93]. The difference lies in the reference coordinates of query points. In the context of Deformable Cross-Attention [93], \"reference coordinates\" refer to the corresponding positions of query points on the feature maps of keys and values. Typically, they are consistent. However, as for Future Decoder, due to the moving of the ego-vehicle, the ego-coordinate systems between the last and target frame are not necessarily aligned. Therefore, we additionally compute the reference coordinates of future BEV queries in previous BEV feature maps, according to the ego-motion condition, to align coordinate systems.", + "bbox": [ + 75, + 450, + 470, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the next BEV features $\\hat{\\mathcal{F}}_t$ , we use a projection layer to generate the occupancy volume $\\mathcal{P}_t$ .", + "bbox": [ + 75, + 632, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss. Instead of using L1 loss to supervise the depths of various rays, we directly apply ray-wise cross-entropy loss to maximize the response of points along its corresponding ray, as we have already obtained geometric features after the Latent Rendering operator. To be specific, for each ground-truth point of the $t$ -th future point clouds, we cast a ray from the origin position $\\mathbf{o}$ (the sensor position) towards the point, uniformly sample some waypoints along the ray until out of the volume, and compute cross-entropy loss for the ray to maximize the response of the point position and minimize the response of other waypoint positions. This process is formulated as:", + "bbox": [ + 75, + 664, + 470, + 845 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\frac {1}{T n} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\log \\left(\\frac {e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {g} ^ {(i)})}}}{\\sum_ {j} e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {x} ^ {(i , j)})}} + e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {g} ^ {(i)})}}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 854, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ , $n$ indicate the number of future supervisions and the number of points in the $t$ -th ground-truth point clouds. $\\mathbf{g}^{(i)}$ and $\\mathbf{x}^{(i,j)}$ are the coordinates of $i$ -th ground-truth point and $j$ -th waypoints along the same ray. $\\mathcal{P}_t^{(\\cdot)}$ is trilinear interpolation to obtain corresponding values from volume $\\mathcal{P}_t$ .", + "bbox": [ + 498, + 90, + 890, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 184, + 633, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section investigates the following questions:", + "bbox": [ + 500, + 210, + 823, + 224 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Can future point clouds be estimated from visual history, and how about ViDAR compared to point cloud methods?", + "- Can ViDAR help perception, prediction, and planning at the same time so as for scalable autonomous driving?", + "- Can ViDAR reduce the reliance of downstream applications on precise human annotations?", + "- How do different modules affect final performance?" + ], + "bbox": [ + 500, + 228, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 359, + 581, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct experiments on the challenging nuScenes [5] dataset, which is a large-scale dataset with 1,000 autonomous driving sequences. This dataset is widely utilized in perception tasks including 3D object detection [28, 43-45, 83-87], multi-object tracking [24, 59, 91], and semantic occupancy prediction [10, 67]. It has also become a popular benchmark for subsequent research on end-to-end autonomous driving, including map segmentation [40, 68], trajectory predictions [17, 47], future occupancy prediction [23, 92], and open-loop planning [26, 27, 32].", + "bbox": [ + 498, + 383, + 890, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We base our implementation on mmDet3D codebase [9] and conduct downstream verifications on BEVFormer for 3D detection, OccNet [67] for semantic occupancy prediction, and UniAD [27] for unified perception, prediction, and planning. We choose those downstream baselines due to their effectiveness on a wide range of tasks and sharing the same BEV encoder structure, BEVFormer encoder [45]. Without any specifications, the default historical encoder of ViDAR is the BEVFormer-base encoder, consisting of a ResNet101-DCN [11, 19] backbone with an FPN neck [49] and additional 6 encoder layers to extract BEV features from multi-view image sequences, which is consistent with downstream models.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To render geometric features, we use a 16-group Latent Rendering. Each group is responsible for rendering latent spaces of 16 channels given the features of 256 channels after the BEVFormer-base encoder. The Future Decoder is a 6-layer structure with a channel of 256 for each. The future BEV queries are $200 \\times 200$ learnable tokens indicating a valid perception range of $[-51.2\\mathrm{m}, 51.2\\mathrm{m}]$ for the $X$ and $Y$ axis. We then use a projection layer with output channels as 16 to transform the predicted future BEV features to occupancy volume prediction $\\mathcal{P} \\in \\mathbb{R}^{200 \\times 200 \\times 16}$ , where 16 indicates the height dimension with a range of $[-5\\mathrm{m}, 3\\mathrm{m}]$ .", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/789322680444908f439ee74e78a3a8b1069d3c7a4ec3e1aea8f5f87d76dfcf85.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
History HorizonMethodModalityChamfer Distance (m2) ↓
0.5s1.0s1.5s2.0s2.5s3.0s
1s4D-Occ [36]L1.261.88----
ViDARC1.111.251.401.571.761.97
3s4D-Occ [36]L0.911.131.301.531.722.11
ViDARC1.011.121.251.381.541.73
", + "bbox": [ + 83, + 89, + 464, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During pre-training, we use 5 frames of historical multiview images and iterate the Future Decoder 6 times to predict point clouds for the future 3 seconds (each frame has 0.5 second interval). In each training step, we randomly select 1 future prediction for computing loss and detach the gradients of the others to save GPU memory. We pre-train the system for 50 epochs by AdamW optimizer [37, 55] with an initial learning of 2e-4 adjusted by cosine annealing strategy. For fine-tuning, we follow the same training strategy as the officially released downstream models.", + "bbox": [ + 75, + 239, + 467, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 401, + 215, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now demonstrate the effectiveness of ViDAR across different tasks. First, we test the ability of ViDAR as a point cloud forecasting framework and compare it with the state-of-the-art approach that uses LiDAR inputs. Then, we show its advancement as a visual autonomous driving pre-training solution. We report the downstream comparison results in the order of perception-prediction-planning with previous state-of-the-art models on the nuScenes validation dataset.", + "bbox": [ + 75, + 424, + 467, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Downstream Settings. For downstream verifications, we test ViDAR to pre-train BEV encoders under different initialization settings, listed as the following:", + "bbox": [ + 75, + 547, + 467, + 593 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ViDAR-cls: The BEV encoders are initialized with the backbone pre-trained for ImageNet classification [12], followed by ViDAR pre-training on nuScenes dataset.", + "- ViDAR-2D-det: The BEV encoder backbones are first pre-trained for 2D object detection on COCO dataset [48] before ViDAR pre-training on nuScenes dataset.", + "- ViDAR-3D-det: The BEV encoder backbones are pretrained first by 3D detection on the nuScenes dataset, using FCOS3D [69], before ViDAR pre-training." + ], + "bbox": [ + 76, + 595, + 467, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For UniAD experiments, after ViDAR pre-training, we first fine-tune BEVFormer for 3D detection, which is then used as the initialization for subsequent two-stage fine-tuning, consistent with the UniAD official implementation.", + "bbox": [ + 75, + 743, + 467, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point Cloud Forecasting. In Table 2, we present comparisons between our ViDAR and the previous state-of-the-art point cloud forecasting method, 4D-Occ [36]. The evaluation metric is Chamfer Distance. We evaluate both methods using input time horizons of 1s and 3s, corresponding", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/498e3b4936e3b5979df92380e1557bad67c17666f2865182fd8680f14574fef5.jpg", + "table_caption": [ + "Table 2. Point cloud forecasting. ViDAR surpasses prior state-of-the-art method on future prediction, using visual input only." + ], + "table_footnote": [], + "table_body": "
MethodsEncoderPre-trainmAP (%)NDS (%)
BEV-Former [45]RN50 [19]ImageNet-cls [12]25.235.4
ViDAR-cls29.038.8
RN101 [19]ImageNet-cls [12]37.747.7
ViDAR-cls42.651.8
nus-3D-det [69]41.551.7
ViDAR-3D-det45.854.8
Intern-S [71]COCO-2D-det [48]41.551.2
ViDAR-2D-det47.656.4
Intern-B [71]COCO-2D-det [48]42.952.0
ViDAR-2D-det50.157.6
", + "bbox": [ + 506, + 89, + 887, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Detection performance of BEVFormer with and without ViDAR pre-training under various backbones and initializations.", + "bbox": [ + 500, + 258, + 890, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to input sequences of 2 frames and 6 frames, respectively, following the same setup as in 4D-Occ. To provide detailed comparisons of performance, we report the quantitative forecasting results for each future timestamp. Only points within the range of $[-51.2\\mathrm{m}, 51.2\\mathrm{m}]$ on the X- and Y-axis are considered during evaluation.", + "bbox": [ + 496, + 314, + 890, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As presented in Table 2, ViDAR consistently outperforms 4D-Occ on both 1s and 3s settings, despite utilizing visual inputs exclusively. Specifically, with 1s history input, ViDAR achieves remarkable improvement over 4D-Occ, reducing forecasting errors by $\\sim 33\\%$ for future 1s predictions. When using 3s inputs, we observe a $\\sim 18\\%$ error reduction for 3s forecasting. Moreover, due to the autoregressive design of our Future Decoder, ViDAR effectively predicts arbitrary future, though with the constraint of a limited 1s input horizon. These experiments demonstrate the effectiveness of ViDAR for point cloud forecasting.", + "bbox": [ + 496, + 405, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perception. We verify ViDAR on four downstream perception tasks, 3D object detection, semantic occupancy prediction, map segmentation, and multi-object tracking. In Table 3 and Table 4, we compare the performance of BEVFormer [45] and OccNet [67] with and without ViDAR pre-training under different backbones and initialization settings for 3D detection and semantic occupancy prediction, respectively. Notably, ViDAR, using solely Image-LiDAR sequences, outperforms 3D detection supervised pre-training (The 4th & the 5th row in Table 3, $42.6\\%$ mAP to $41.5\\%$ mAP; The 2nd & the 3rd row in Table 4, $29.57\\%$ mIoU to $26.98\\%$ mIoU). Furthermore, we also observe huge improvements in map segmentation (Table 5) and multi-object tracking (Table 6). These experiments demonstrate the effectiveness of ViDAR as a scalable pretraining method for enhancing 3D geometry modeling.", + "bbox": [ + 496, + 592, + 890, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prediction. The motion forecasting comparisons are presented in Table 7. As shown, ViDAR significantly improves the performance of UniAD [27]. For instance, we observe a", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/84e59e3ae74b903d4635f15ff9ddb9ea57ec0bdffb73ccd0f3576f385e2e7813.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsEncoderPre-trainmIoU (%) ↑
OccNet [67]RN101 [19]ImageNet-cls [12]24.35
ViDAR-cls29.57
nus-3D-det [69]26.98
ViDAR-3D-det31.67
Intern-S [71]COCO-2D-det [48]24.92
ViDAR-2D-det30.51
Intern-B [71]COCO-2D-det [48]25.24
ViDAR-2D-det31.69
", + "bbox": [ + 83, + 89, + 460, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/015b56f8320e7e1b24140a135239db5863aedc91e86bd0eb836bfff060ca2bb0.jpg", + "table_caption": [ + "Table 4. Semantic Occupancy Prediction. ViDAR consistently improves OccNet on different backbones and initializations." + ], + "table_footnote": [], + "table_body": "
MethodsEncoderPre-trainLanes (%) ↑
BEVFormer [45]RN101nus-3D-det [69]23.9
UniAD [27]RN101nus-3D-det [45]31.3
ViDAR-3D-det33.2
", + "bbox": [ + 81, + 262, + 470, + 329 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db1916240b212beff0524175e72b2acc9199d397da909f3d64d4d3dbb5ddca94.jpg", + "table_caption": [ + "Table 5. Map segmentation. ViDAR improves UniAD on online mapping. The metric is segmentation IoU." + ], + "table_footnote": [], + "table_body": "
MethodsEncoderPre-trainAMOTA (%) ↑
ViP3D [17]RN50nus-3D-det [72]21.7
QD3DT [24]RN101-24.2
MUTR3D [91]RN101ImageNet-cls29.4
DQTrack-DETR3D [44]RN101nus-3D-det [72]36.7
DQTrack-UVTR [44]RN101nus-3D-det [41]39.6
DQTrack-Stereo [44]RN101nus-3D-det [42]40.7
DQTrack-PETr v2 [44]V2-99nus-3D-det [53]44.6
UniAD-Stage1 [27]RN101nus-3D-det [45]39.0
ViDAR-3D-det45.1
", + "bbox": [ + 81, + 367, + 470, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\sim 10\\%$ error reduction in minADE and a $3.5\\%$ improvement in EPA. In Table 8, we provide the comparison between UniAD with and without ViDAR pre-training in future occupancy prediction. The results demonstrate that ViDAR enhances the performance of UniAD for all areas. We observe improvements of $2.4\\%$ IoU and $2.7\\%$ VPQ for nearby areas, as well as improvements of $2.0\\%$ IoU and $2.5\\%$ VPQ for distant areas. With ViDAR pre-training, UniAD overcomes its limitation in occupancy forecasting for distant objects and now outperforms BEVerse [92] in all areas. These experiments highlight the effectiveness of ViDAR in enhancing downstream models in utilizing temporal information and improving their prediction performance.", + "bbox": [ + 75, + 564, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Planning. Due to the effective temporal modeling and advanced future prediction capabilities, ViDAR significantly improves UniAD by reducing its average collision rate within 3 seconds by $\\sim 15\\%$ . Moreover, it achieves a substantial decrease in the average planning displacement error by $0.21\\mathrm{m}$ and enables UniAD to outperform the state-of-the-art method, VAD [32], on nuScenes open-loop planning evaluation. These improvements demonstrate the effective", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/953a9ff399d25d15f70ffefefa16e873bdd9dffed9bdef20cda49caba8627d81.jpg", + "table_caption": [ + "Table 6. Multi-object tracking. With ViDAR, UniAD outperforms previous end-to-end trackers using solely visual images." + ], + "table_footnote": [], + "table_body": "
MethodsminADE (m) ↓minFDE (m) ↓MR ↓EPA ↑
PnPNet [47]1.151.950.2260.222
ViP3D [17]2.052.840.2460.226
UniAD [27]0.751.080.1580.463
ViDAR0.670.990.1490.498
", + "bbox": [ + 504, + 89, + 895, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7f8408c33c2a5786a580c967a309b38f9369f938f1ad4c99ae60f661cf51e45.jpg", + "table_caption": [ + "Table 7. Motion forecasting. ViDAR effectively enhances temporal modeling, which in turn boosts UniAD in future motion forecasting, showing consistent improvements on different metrics." + ], + "table_footnote": [], + "table_body": "
MethodsVPQ-n. ↑VPQ-f. ↑IoU-n. ↑IoU-f. ↑
Fiery [23]50.229.959.436.7
StretchBEV [1]46.029.055.537.1
ST-P3 [26]-32.1-38.9
BEVerse [92]54.336.161.440.9
UniAD [27]54.633.962.840.1
ViDAR57.336.465.442.1
", + "bbox": [ + 504, + 222, + 883, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47ae7d0d72a05191bd6ed247c62b2988586dacd798dbc0b81a0242b48e270019.jpg", + "table_caption": [ + "Table 8. Future occupancy prediction. ViDAR improves UniAD on future occupancy prediction for objects in both near (noted as \"n\", $30\\mathrm{x}30\\mathrm{m}$ ) and far (noted as \"f\", $50\\mathrm{x}50\\mathrm{m}$ ) evaluation areas." + ], + "table_footnote": [], + "table_body": "
MethodsModalityAvg.Col. (3s) (%) ↓Avg.L2 (3s) (m) ↓
FF [25]L0.431.43
EO [35]L0.331.60
ST-P3 [26]C0.712.11
VAD [32]C0.411.05
UniAD [27]C0.271.12
ViDARC0.230.91
", + "bbox": [ + 504, + 378, + 901, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9. Planning. ViDAR improves UniAD in terms of both collision avoidance and planning accuracy. Note that, the reported numbers are obtained by averaging the results of each timestamp in the future 3 seconds, which is consistent with the reported results of VAD [32] and ST-P3 [26] at the 3s timestamp instead of the averaged one. Please refer to GitHub:Issue for more details.", + "bbox": [ + 498, + 486, + 890, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ness of ViDAR as a valuable pre-training approach for end-to-end autonomous driving. The enhanced performance in collision avoidance and planning accuracy highlights the potential of ViDAR in enhancing the safety and efficiency of downstream autonomous driving applications.", + "bbox": [ + 496, + 599, + 890, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint Perception-Prediction-Planning. Finally, we summarize the improvements of ViDAR on the state-of-the-art end-to-end visual autonomous driving system, UniAD [27], for joint perception, prediction, and planning. As depicted in Table 10, ViDAR brings substantial improvements on all sub-modules of UniAD for perception (Detection, Tracking, Mapping), prediction (Motion Forecasting and Future Occupancy Prediction), and planning at the same time. These consistent improvements illustrate that visual point cloud forecasting effectively exploits the information of semantics, 3D geometry, and temporal dynamics behind the easily obtainable Image-LiDAR sequences. This, consequently, enables scalable visual autonomous driving.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7cccab671264fc88160b387af3803080fd945ae546ab56216abbab3dc1f75c17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDetectionTrackingMappingMotion ForecastingFuture Occupancy PredictionPlanning
NDS ↑mAP ↑AMOTA↑AMOTP↓IDS↓IoU-lane↑IoU-road↑minADE↓minFDE↓MR↓IoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑avg.L2↓avg.Col.↓
UniAD49.3637.9638.31.32105431.369.10.751.080.15862.840.154.633.91.120.27
ViDAR52.5742.3342.01.2599133.271.40.670.990.14965.442.157.336.40.910.23
", + "bbox": [ + 78, + 89, + 890, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77b0bc98d2b4f373136843570bc9cd8b65772513a460b3e37f860e3127d8deaf.jpg", + "image_caption": [ + "Figure 6. Validation of ViDAR on Fine-tuning with limited supervised data. We verify ViDAR on supervision efficiency by reducing available annotations for 3D object detection during downstream fine-tuning (from the full training set to a 1/8th subset) and observe a continuous improvement on each subset." + ], + "image_footnote": [], + "bbox": [ + 86, + 180, + 460, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablative Study", + "text_level": 1, + "bbox": [ + 76, + 429, + 227, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct further analysis of ViDAR on improving the performance of downstream models with limited supervised data, and the effect of the Latent Rendering operation on learning 3D geometric latent space. More ablation studies can be found in the supplementary materials.", + "bbox": [ + 75, + 452, + 468, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency of Supervised Pre-training. The primary objective of pre-training is to minimize the dependence on precise 3D annotations. In Figure 6, we demonstrate the effectiveness of ViDAR in reducing the reliance of modern 3D detectors on accurate 3D box annotations. We fine-tune BEVFormer-base using partial 3D annotations on nuScenes, ranging from the full dataset to a $1/8$ subset.", + "bbox": [ + 75, + 547, + 468, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As depicted in Figure 6, ViDAR exhibits a remarkable reduction in the dependence on 3D annotations. Notably, BEVFormer, pre-trained by ViDAR, surpasses its counterpart under full supervision by $1.7\\%$ mAP, while using only half of the supervised samples, i.e., $39.4\\%$ mAP vs. $37.7\\%$ mAP. Consequently, through ViDAR, we can reduce half of 3D annotations without sacrificing precision. Additionally, we observe a consistent trend of increasing improvements as the available supervision decreases. For instance, the mAP improvements are $4.9\\%$ , $6.5\\%$ , $6.7\\%$ , and $7.3\\%$ when fine-tuned on the full, half, a quarter, and 1/8th subsets. These results highlight the potential of ViDAR in harnessing large amounts of Image-LiDAR sequences.", + "bbox": [ + 75, + 654, + 468, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Latent Rendering operator. Latent Rendering is the key component of ViDAR, which enables visual point", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61d3b0a1e3fe879e06fe35c871ff337a4f9c3d9a184b6e1643de95f70fc114dc.jpg", + "table_caption": [ + "Table 10. Performance gain of ViDAR for joint perception, prediction, and planning. ViDAR consistently improves UniAD [27] on all tasks towards end-to-end autonomous driving, validating its effectiveness for scalable visual autonomous driving." + ], + "table_footnote": [], + "table_body": "
Groups of Latent R.N/A124816
NDS (%)40.2039.1843.3645.5347.0147.58
", + "bbox": [ + 503, + 181, + 893, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11. Ablation of Latent Rendering for downstream finetuning. We compare the performance of 3D detection pre-trained by ViDAR without the Latent Rendering operation (denoted as \"N/A\") and by ViDAR with different groups of Latent Rendering.", + "bbox": [ + 500, + 236, + 892, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cloud forecasting to effectively contribute to downstream applications. It addresses the ray-shaped features issue encountered during pre-training. In Table 11, we verify its effectiveness by comparing the performance of downstream models pre-trained by ViDAR with the Latent Rendering or not. The downstream model is BEVFormer-small [45] for 3D object detection. For context, the baseline performance with ImageNet-cls pre-training is $44.11\\%$ NDS.", + "bbox": [ + 496, + 314, + 890, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As depicted in Table 11, when the Latent Rendering is missing (denoted as \"N/A\"), also referred to as the baseline in Section 3.2, a significant decline is observed in downstream performance after fine-tuning, from $44.11\\%$ NDS to $40.20\\%$ NDS. In contrast, with the 16-group Latent Rendering, the performance improves to $47.58\\%$ NDS, a notable $3.47\\%$ NDS improvement over the baseline.", + "bbox": [ + 496, + 436, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a comparison of Latent Rendering with different parallel groups in Table 11 as well. The results demonstrate a consistent improvement by dividing channels into more groups and integrating information separately.", + "bbox": [ + 496, + 545, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 627, + 617, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced visual point cloud forecasting, which predicts future point clouds from historical visual images, as a new pre-training task for end-to-end autonomous driving. We developed ViDAR, a general model to pre-train visual BEV encoders, and designed a Latent Rendering operator to solve the ray-shaped feature issue. To conclude, our work demonstrates that visual point cloud forecasting enables scalable autonomous driving.", + "bbox": [ + 496, + 655, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Though with the potential of scalability, in this paper, we mainly conduct pre-training on Image-LiDAR sequences from nuScenes dataset, of which the data scale is still limited. As for the future, we plan to scale up the pre-training data of ViDAR, study visual point cloud forecasting across diverse datasets, and use publicly available Image-LiDAR sequences as much as possible to train a foundation visual autonomous driving model [38].", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14680", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adil Kaan Akan and Fatma Güney. StretchBEV: Stretching Future Instance Prediction Spatially and Temporally. In ECCV, 2022. 7", + "[2] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A Cookbook of Self-supervised Learning. arXiv preprint arXiv:2304.12210, 2023. 1", + "[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In ICCV, 2019. 1", + "[4] Alex Bewley, Pei Sun, Thomas Mensink, Dragomir Anguelov, and Cristian Sminchisescu. Range Conditioned Dilated Convolutions for Scale Invariant 3D Object Detection. arXiv preprint arXiv:2005.09927, 2021. 3", + "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi-ancarlo Baldan, and Oscar Beijbom. nuScenes: A Multi-modal Dataset for Autonomous Driving. In CVPR, 2020. 1, 2, 5", + "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In ECCV, 2022. 1", + "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end Autonomous Driving: Challenges and Frontiers. arXiv preprint arXiv:2306.16927, 2023. 2", + "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint arXiv:2003.04297, 2020. 1, 2", + "[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5", + "[10] OpenScene Contributors. OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving, 2023. 1, 5", + "[11] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable Convolutional Networks. In ICCV, 2017. 5", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6, 7", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5", + "[15] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles Qi, Yin Zhou, Zoey Yang, Aurélien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset. In ICCV, 2021. 1", + "[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. I. J. Robotics Res., 2013. 1", + "[17] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. ViP3D: End-to-end visual trajectory prediction via 3d agent queries. In CVPR, 2023. 1, 5, 7", + "[18] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A Survey on Self-supervised Learning: Algorithms, Applications, and Future Trends. arXiv preprint arXiv:2301.05712, 2023. 1", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5, 6, 7", + "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In CVPR, 2020. 1, 2", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2022. 2", + "[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts. In CVPR, 2021. 1", + "[23] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: Future Instance Segmentation in Bird's-Eye view from Surround Monocular Cameras. In ICCV, 2021. 5, 7", + "[24] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular Quasi-Dense 3D Object Tracking. TPAMI, 2022. 5, 7", + "[25] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe Local Motion Planning with Self-Supervised Freespace Forecasting. In CVPR, 2021. 7", + "[26] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning. In ECCV, 2022. 5, 7", + "[27] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1, 2, 5, 6, 7, 8", + "[28] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14681", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction. In CVPR, 2023. 1", + "[30] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. DriveAdapter: Breaking the Coupling Barrier of Perception and Planning in End-to-End Autonomous Driving. In ICCV, 2023. 1", + "[31] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think Twice before Driving: Towards Scalable Decoders for End-to-End Autonomous Driving. In CVPR, 2023. 1", + "[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 5, 7", + "[33] Li Jiang, Zetong Yang, Shaoshuai Shi, Vladislav Golyanik, Dengxin Dai, and Bernt Schiele. Self-supervised Pretraining with Masked Shape Prediction for 3D Scene Understanding. In CVPR, 2023. 1", + "[34] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. arXiv preprint arXiv:2004.11362, 2020. 2", + "[35] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable Raycasting for Self-Supervised Occupancy Forecasting. In ECCV, 2022. 3, 7", + "[36] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point Cloud Forecasting as a Proxy for 4D Occupancy Forecasting. In CVPR, 2023. 3, 6", + "[37] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR, 2015. 6", + "[38] Hongyang Li, Yang Li, Huijie Wang, Jia Zeng, Pinlong Cai, Huilin Xu, Dahua Lin, Junchi Yan, Feng Xu, Lu Xiong, Jingdong Wang, Futang Zhu, Kai Yan, Chunjing Xu, Tiancai Wang, Beipeng Mu, Shaoqing Ren, Zhihui Peng, and Yu Qiao. Open-sourced Data Ecosystem in Autonomous Driving: the Present and Future. arXiv preprint arXiv:2312.03408, 2023. 8", + "[39] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, Hao Tian, Enze Xie, Jiangwei Xie, Li Chen, Tianyu Li, Yang Li, Yulu Gao, Xiaosong Jia, Si Liu, Jianping Shi, Dahua Lin, and Yu Qiao. Delving Into the Devils of Bird's-Eye-View Perception: A Review, Evaluation and Recipe. TPAMI, 2023. 2", + "[40] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based Topology Reasoning for Driving Scenes. arXiv preprint arXiv:2304.05277, 2023. 5", + "[41] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying Voxel-based Representation with Transformer for 3D Object Detection. In NeurIPS, 2022. 7", + "[42] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. BEVStereo: Enhancing Depth Estimation in" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Multi-View 3D Object Detection with Temporal Stereo. In AAAI, 2023. 7", + "[43] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object Detection. In AAAI, 2023. 5", + "[44] Yanwei Li, Zhiding Yu, Jonah Philion, Anima Anandkumar, Sanja Fidler, Jiaya Jia, and Jose Alvarez. End-to-end 3D Tracking with Decoupled Queries. In ICCV, 2023. 7", + "[45] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEV-Former: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8", + "[46] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and José Manuel Álvarez. FB-BEV: BEV Representation from Forward-Backward View Transformations. ICCV, 2023. 2, 4", + "[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. PnPNet: End-to-End Perception and Prediction With Tracking in the Loop. In CVPR, 2020. 5, 7", + "[48] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7", + "[49] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 5", + "[50] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos. arXiv preprint arXiv:2308.09244, 2023. 1", + "[51] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully Sparse 3D Panoptic Occupancy Prediction. arXiv preprint arXiv:2312.17118, 2023. 1", + "[52] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position Embedding Transformation for Multi-View 3D Object Detection. In ECCV, 2022. 1, 3", + "[53] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. In ICCV, 2023. 7", + "[54] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In ICRA, 2023. 1", + "[55] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. arXiv preprint arXiv:1711.05101, 2019. 6", + "[56] B. Mersch, X. Chen, J. Behley, and C. Stachniss. Self-supervised Point Cloud Prediction Using 3D Spatiotemporal Convolutional Networks. In CoRL, 2021. 3", + "[57] Gregory P. Meyer, Ankit Laddha, Eric Kee, Carlos Vallespi-Gonzalez, and Carl K. Wellington. LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving. In CVPR, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14682", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Chen Min, Liang Xiao, Dawei Zhao, Yiming Nie, and Bin Dai. Occupancy-MAE: Self-Supervised Pre-Training LargeScale LiDAR Point Clouds With Masked Occupancy Autoencoders. TIV, 2023. 2", + "[59] Ziqi Pang, Zhichao Li, and Naiyan Wang. SimpleTrack: Understanding and Rethinking 3D Multi-object Tracking. arXiv preprint arXiv:2111.09621, 2021. 5", + "[60] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is Pseudo-Lidar Needed for Monocular 3D Object Detection? In ICCV, 2021. 2, 3", + "[61] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical Depth DistributionNetwork for Monocular 3D Object Detection. In CVPR, 2021. 1", + "[62] Ravid Shwartz-Ziv and Yann LeCun. To Compress or Not to Compress-Self-Supervised Learning and Information Theory: A Review. arXiv preprint arXiv:2304.09355, 2023. 1", + "[63] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv preprint arXiv:2312.14150, 2023. 1", + "[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 1", + "[65] Izzeddin Teeti, Salman Khan, Ajmal Shahbaz, Andrew Bradley, and Fabio Cuzzolin. Vision-based Intention and Trajectory Prediction in Autonomous Vehicles: A Survey. In *IJCAI*, 2022. 1", + "[66] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive Multiview Coding. In ECCV, 2020. 1, 2", + "[67] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as Occupancy. In ICCV, 2023. 1, 2, 3, 5, 6, 7", + "[68] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping. In NeurIPS Datasets and Benchmarks, 2023. 5", + "[69] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In ICCV Workshops, 2021. 1, 2, 3, 6, 7", + "[70] Tai Wang, Qing Lian, Chenming Zhu, Xinge Zhu, and Wenwei Zhang. MV-FCOS3D++: Multi-View Camera-Only 4D Object Detection with Pretrained Monocular Backbones. arXiv preprint arXiv:2207.12716, 2022. 1", + "[71] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions. In CVPR, 2023. 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In CoRL, 2021. 7", + "[73] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In CVPR, 2022. 2", + "[74] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the Pose Forecasting Pipeline with SPF2: Sequential Pointcloud Forecasting for Sequential Pose Forecasting. arXiv preprint arXiv:2003.08376, 2020. 3", + "[75] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2Net: Stochastic Sequential Pointcloud Forecasting. In ECCV, 2022. 3", + "[76] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline. In NeurIPS, 2022. 1", + "[77] Penghao Wu, Li Chen, Hongyang Li, Xiaosong Jia, Junchi Yan, and Yu Qiao. Policy Pre-training for Autonomous Driving via Self-supervised Geometric Modeling. In ICLR, 2023. 2", + "[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In CVPR, 2018. 1", + "[79] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In ECCV, 2020. 1", + "[80] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2", + "[81] Xiangchao Yan, Runjian Chen, Bo Zhang, Jiakang Yuan, Xinyu Cai, Botian Shi, Wenqi Shao, Junchi Yan, Ping Luo, and Yu Qiao. SPOT: Scalable 3D Pre-training via Occupancy Prediction for Autonomous Driving. arXiv preprint arXiv:2309.10527, 2023. 2", + "[82] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, Xiaofei He, and Wanli Ouyang. UniPAD: A Universal Pre-training Paradigm for Autonomous Driving. arXiv preprint arXiv:2310.08370, 2023. 2, 3", + "[83] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. IPOD: Intensive Point-based Object Detector for Point Cloud. CoRR, 2018. 5", + "[84] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, 2019.", + "[85] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3DSSD: Point-based 3D Single Stage Object Detector, 2020.", + "[86] Zetong Yang, Yanan Sun, Shu Liu, Xiaojuan Qi, and Jiaya Jia. CN: Channel Normalization for Point Cloud Recognition. In ECCV, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14683", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3D-MAN: 3D Multi-Frame Attention Network for Object Detection. In CVPR, 2021. 5", + "[88] Zetong Yang, Li Jiang, Yanan Sun, Bernt Schiele, and Jiaya Jia. A Unified Query-based Paradigm for Point Cloud Understanding. In CVPR, 2022. 1", + "[89] Zetong Yang, Zhiding Yu, Chris Choy, Renhao Wang, Anima Anandkumar, and Jose M. Alvarez. Improving Distant 3D Object Detection Using 2D Box Supervision. arXiv preprint arXiv:2403.09230, 2024. 1", + "[90] Jia Zeng, Li Chen, Hanming Deng, Lewei Lu, Junchi Yan, Yu Qiao, and Hongyang Li. Distilling Focal Knowledge From Imperfect Expert for 3D Object Detection. In CVPR, 2023. 2, 4", + "[91] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. MUTR3D: A Multi-camera Tracking Framework via 3D-to-2D Queries. In CVPR, 2022. 5, 7", + "[92] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. BEVerse: Unified Perception and Prediction in Birds-Eye-View for Vision-Centric Autonomous Driving. arXiv preprint arXiv:2205.09743, 2022. 5, 7", + "[93] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable Transformers for End-to-End Object Detection. In ICLR, 2020. 5" + ], + "bbox": [ + 78, + 90, + 468, + 445 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14684", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_model.json b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_model.json new file mode 100644 index 0000000000000000000000000000000000000000..746abeef17795c2e4f7d5e66e7bb3ab7c026f3d4 --- /dev/null +++ b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_model.json @@ -0,0 +1,2886 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "Visual Point Cloud Forecasting enables Scalable Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.181, + 0.7, + 0.201 + ], + "angle": 0, + "content": "Zetong Yang Li Chen Yanan Sun Hongyang Li" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.206, + 0.632, + 0.224 + ], + "angle": 0, + "content": "OpenDriveLab and Shanghai AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.23, + 0.672, + 0.245 + ], + "angle": 0, + "content": "https://github.com/OpenDriveLab/ViDAR" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.258, + 0.896, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.494, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Figure 1. ViDAR is a visual autonomous driving pre-training framework, which leverages the estimation of future point clouds from historical visual inputs as the pre-text task. We term this new pre-text task as visual point cloud forecasting. With the aid of ViDAR, we achieve substantial improvement spanning a diverse spectrum of downstream applications for perception, prediction, and planning." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.554, + 0.314, + 0.57 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.587, + 0.474, + 0.892 + ], + "angle": 0, + "content": "In contrast to extensive studies on general vision, pretraining for scalable visual autonomous driving remains seldom explored. Visual autonomous driving applications require features encompassing semantics, 3D geometry, and temporal information simultaneously for joint perception, prediction, and planning, posing dramatic challenges for pre-training. To resolve this, we bring up a new pre-training task termed as visual point cloud forecasting - predicting future point clouds from historical visual input. The key merit of this task captures the synergic learning of semantics, 3D structures, and temporal dynamics. Hence it shows superiority in various downstream tasks. To cope with this new problem, we present ViDAR, a general model to pretrain downstream visual encoders. It first extracts historical embeddings by the encoder. These representations are then transformed to 3D geometric space via a novel Latent Rendering operator for future point cloud prediction. Experiments show significant gain in downstream tasks, e.g., \\(3.1\\%\\) NDS on 3D detection, \\(\\sim 10\\%\\) error reduction on motion forecasting, and \\(\\sim 15\\%\\) less collision rate on planning." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.554, + 0.633, + 0.57 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.763 + ], + "angle": 0, + "content": "Recently, the community has witnessed rapid development in visual, or camera-only autonomous driving, with input being monocular or multi-view images [50, 63, 69, 70, 76, 89]. Leveraging visual inputs, existing approaches demonstrate superior capability of extracting Bird's-Eye-View (BEV) features [6, 28, 45, 52, 61, 88], and performing well in perception [29, 51, 54, 67], prediction [17, 27, 65], and planning [30, 31]. Despite significant improvements in applications, these models rely on precise 3D annotations to a great extent, which are often difficult to collect, e.g., occupancy [3, 10], 3D boxes [5, 16, 64], trajectories [15], and thus are challenging to scale up for production." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Considering the expensive labeling workflow, pretraining [2, 18, 62] has emerged as a crucial approach to scale up downstream applications. The key idea is to define pretext tasks that leverage large amounts of readily available data to learn meaningful representations. This enhances downstream performance though labeled data is limited." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Though extensive research of pre-training in computer vision has been conducted [8, 20-22, 33, 66, 78, 79], its application in visual autonomous driving remains seldom" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14673" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "explored. Visual autonomous driving poses great challenges in pre-training as it requires the features to maintain semantics, 3D geometry, and temporal dynamics at the same time for joint perception, prediction, and planning [7, 77]. As a result, most models still rely on supervised pre-training, like 3D detection [60, 69] or occupancy [58, 67, 81], using labeled data that is often unavailable at scale [39]. Some approaches propose estimating depth [60] or rendering masked scenes [82] as pre-training. They use Image-LiDAR pairs as a means of achieving scalable annotation-free pre-training. However, they struggle in either multi-view 3D geometry or temporal modeling (Figure 2). Depth estimation retrieves depth from one image, limited in multi-view geometry; rendering techniques reconstruct scenes from multi-view images but lacking temporal modeling. However, temporal modeling is crucial in end-to-end autonomous driving systems, e.g., UniAD [27], especially for prediction and planning which are the ultimate goals and require accurate scene flow and object motion for decision-making. Due to the absence of temporal modeling, existing approaches are insufficient for pretraining the end-to-end system." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.427, + 0.47, + 0.548 + ], + "angle": 0, + "content": "In this work, we explore pre-training tailored for end-to-end visual autonomous driving applications, including not only perception but also prediction and planning [7]. We formulate a new pre-text task, visual point cloud forecasting (Figure 2), to fully exploit information of semantics, 3D geometry, and temporal dynamics behind the raw Image-LiDAR sequences, with being scalable into consideration. It predicts future point clouds from historical visual images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.551, + 0.47, + 0.718 + ], + "angle": 0, + "content": "The main rationale of visual point cloud forecasting lies in the simultaneous supervision of semantics, 3D structure, and temporal modeling. By compelling the model to predict the future from history, it supervises the extraction of scene flow and object motion which are crucial for temporal modeling and future estimation. Meanwhile, it involves the reconstruction of point clouds from images, which supervises the multi-view geometry and semantic modeling. Therefore, features from visual point cloud forecasting embed information of both geometric and temporal hints, beneficial for perception, tracking, and planning simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To this end, we present ViDAR, a general visual point cloud forecasting approach for pre-training (Figure 2). ViDAR includes three parts, History Encoder, Latent Rendering operator, and Future Decoder. The History Encoder is the target structure for pre-training. It could be any visual BEV encoder [45] to embed visual sequences into BEV space. These BEV features are sent to the Latent Rendering operator. Latent Rendering plays a crucial role in enabling ViDAR benefit downstream performance. It solves the ray-shaped BEV features issue [46, 90], models 3D geometric latent space, and bridges encoder and decoder. The Future Decoder is an auto-regressive transformer that takes histor" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.696, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.89, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.149, + 0.89, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.266, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Figure 2. Comparisons among visual autonomous driving pretraining paradigms and our ViDAR architecture. Compared to existing methods, visual point cloud forecasting jointly models multi-view geometry and temporal dynamics. We then propose ViDAR, using Image-LiDAR sequences to pre-train visual encoders." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.392 + ], + "angle": 0, + "content": "ical BEV features to iteratively predict future point clouds for arbitrary timestamps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.392, + 0.893, + 0.68 + ], + "angle": 0, + "content": "ViDAR provides a comprehensive solution for visual autonomous driving pre-training. We test ViDAR on nuScenes dataset [5] in terms of point cloud forecasting and downstream verifications. Though with visual inputs, ViDAR still outperforms previous forecasting methods using point clouds, \\(\\sim 33\\%\\) Chamfer Distance reduction on point cloud estimation of 1s future. ViDAR also improves downstream performance. Using Image-LiDAR sequences only, ViDAR surprisingly outperforms 3D detection pre-training [69], e.g., by \\(1.1\\%\\) mAP and \\(2.5\\%\\) mIoU for detection and semantic occupancy prediction, under the same data scale. If also based on the 3D detection pre-training, ViDAR boosts previous methods by \\(4.3\\%\\) mAP and \\(4.6\\%\\) mIoU. Further, due to the effective pre-training on the joint capture of geometric and temporal information, ViDAR improves UniAD [27] on all tasks for end-to-end autonomous driving including perception, prediction, and planning by a large margin (Figure 1). Experimental results validate that visual point cloud forecasting enables scalable autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.642, + 0.708 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Pre-training for Visual Autonomous Driving. Pretraining for scalable applications has been extensively studied in general vision. These approaches can be roughly divided into contrastive approaches [8, 20, 34, 66], which learn discriminative features from positive and negative pairs; and masked signal modeling [13, 21, 73, 80], which recover discarded signals from remained signals to capture a comprehensive understanding of global semantics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In contrast, pre-training for visual autonomous driving is still under-explored. Visual autonomous driving poses great challenges as it requires semantic understanding, 3D structural awareness, and temporal modeling at the same time," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "for joint perception, prediction, and planning. Existing vision methods mainly consider semantics; methods based on Image-LiDAR pairs [60, 82] struggle with temporal modeling; other supervised strategies [67, 69] are not scalable. In this work, we propose visual point cloud forecasting, which simultaneously models semantics, temporal dynamics, and 3D geometry by a uniform process, and is easily scaled up." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Point Cloud Forecasting. Point cloud forecasting, one of the most fundamental self-supervised tasks for autonomous driving, predicts future point clouds from past point cloud inputs. Previous works use range image [4, 57], a representation obtained by projecting point clouds to dense 2D images using sensor intrinsic and extrinsic parameters. Based on historical range images, they apply 3D convolutions [56], or LSTMs [74, 75] to predict future point clouds. Yet, they additionally model the motion of sensor intrinsic and extrinsic parameters. Later methods factor out the estimation of sensors by introducing 4D occupancy prediction [36] and differentiable ray-casting [35], which ensures a better modeling of the world. Compared to prior literature, we aim at visual point cloud forecasting, using past images to predict future point clouds. Meanwhile, we raise this task as a pre-training paradigm for visual autonomous driving and demonstrate its superiority in a wide range of downstream applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.212, + 0.499 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.584 + ], + "angle": 0, + "content": "In this section, we elaborate on our ViDAR, a visual point cloud forecasting approach for general autonomous driving pre-training. We begin with an overview of ViDAR in Section 3.1, and subsequently delve into Latent Rendering and Future Decoder in Section 3.2 and Section 3.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.188, + 0.605 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.78 + ], + "angle": 0, + "content": "As depicted in Figure 2, ViDAR comprises three components: (a) a History Encoder, also the target structure of pre-training, which extracts BEV embeddings \\(\\mathcal{F}_{\\mathrm{bev}}\\) from visual sequence inputs \\(\\mathcal{I}\\); It can be any visual BEV encoder [28, 45, 52]; (b) a Latent Rendering operator, which simulates the volume rendering operation in latent space so as to obtain geometric embedding \\(\\hat{\\mathcal{F}}_{\\mathrm{bev}}\\) from \\(\\mathcal{F}_{\\mathrm{bev}}\\); and (c) a Future Decoder, which predicts future BEV features \\(\\hat{\\mathcal{F}}_t\\) at timestamps \\(t \\in \\{1, 2, \\ldots\\}\\) in an auto-regressive manner. Finally, a prediction head is followed to project \\(\\hat{\\mathcal{F}}_t\\) into 3D occupancy volume \\(\\mathcal{P}_t\\). This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.786, + 0.275, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {b e v}} = \\operatorname {E n c o d e r} (\\mathcal {I}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.807, + 0.469, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {\\text {b e v}} = \\text {L a t e n t R e n d e r} (\\mathcal {F} _ {\\text {b e v}}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.826, + 0.425, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {t} = \\operatorname {D e c o d e r} \\left(\\hat {\\mathcal {F}} _ {t - 1}\\right), \\text {w h e r e} \\hat {\\mathcal {F}} _ {0} = \\hat {\\mathcal {F}} _ {\\text {b e v}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.849, + 0.312, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {t} = \\text {P r o j e c t i o n} (\\hat {\\mathcal {F}} _ {t}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Point cloud predictions are obtained from the predicted occupancy volume \\(\\mathcal{P}_t\\). This process is similar to the previous" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "point cloud forecasting method [36]. Specifically, we first cast rays from the origin to various designated directions, then figure out the distance of waypoints along each ray with the maximum occupancy response, and finally compute the point position according to the distance and corresponding ray direction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.673, + 0.209 + ], + "angle": 0, + "content": "3.2. Latent Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.216, + 0.892, + 0.368 + ], + "angle": 0, + "content": "A straightforward solution of visual point cloud forecasting for pre-training is to incorporate the History Encoder and Future Decoder directly with differentiable ray-casting [36], which is the crucial component in state-of-the-art point cloud forecasting methods to render point clouds from predicted occupancy volume and compute loss for backpropagation. However, our experimental results show that this approach does not yield improvements and even has a detrimental effect on the downstream tasks due to the defective geometric feature modeling ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Preliminary. Differentiable ray-casting is a volume rendering process operating on an occupancy volume, denoted as \\(\\mathcal{P} \\in \\mathbb{R}^{L \\times H \\times W}\\). It renders depths of various rays and subsequently converts the depths with corresponding ray directions to point clouds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Formally, starting from the origin sensor position, \\(\\mathbf{o} \\in \\mathbb{R}^3\\), differentiable ray-casting casts \\(n\\) rays with varying directions, \\(\\mathbf{d} \\in \\mathbb{R}^{n \\times 3}\\). Along each ray \\(i\\), it uniformly samples \\(m\\) waypoints at different distances \\(\\lambda^{(j)} \\in \\mathbb{R}, j \\in \\{1,2,\\dots,m\\}\\) until reaching the boundary of the 3D space. The coordinates of these waypoints are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.565, + 0.891, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {(i, j)} = \\mathbf {o} + \\lambda^ {(j)} \\mathbf {d} ^ {(i)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Those waypoints coordinates, \\(\\mathbf{x} \\in \\mathbb{R}^{n \\times m \\times 3}\\), are used to compute occupancy values. This process is quantized [36], wherein the waypoints are discretized to occupancy volume grids. Then, the values of waypoints are derived as the associated values of volume grids, \\(\\mathbf{p}^{(i,j)} = \\mathcal{P}^{([\\mathbf{x}^{(i,j)}])}\\). Here \\([\\cdot]\\) denotes a rounding operation for discretizing waypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.72 + ], + "angle": 0, + "content": "Differentiable ray-casting renders the corresponding depth of the \\(i\\)-th ray, \\(\\hat{\\lambda}^{(i)}\\), by an integral process:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.732, + 0.891, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {p}} ^ {(i, j)} = \\left[ \\prod_ {k = 1} ^ {j - 1} (1 - \\mathbf {p} ^ {(i, k)}) \\right] \\mathbf {p} ^ {(i, j)}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.779, + 0.891, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\hat {\\lambda} ^ {(i)} = \\sum_ {j = 1} ^ {m} \\hat {\\mathbf {p}} ^ {(i, j)} \\lambda^ {(j)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.902 + ], + "angle": 0, + "content": "For simplicity, we name Eq. 3 and Eq. 4 as conditional probability function and distance expectation function. The conditional probability function determines the occupancy of a grid by considering the conditional probability that prior waypoints are unoccupied and the ray terminates at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14675" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.207, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.172, + 0.212, + 0.182 + ], + "angle": 0, + "content": "(a.) Ray-shaped Features" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.089, + 0.328, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.172, + 0.333, + 0.182 + ], + "angle": 0, + "content": "(b.) Geometric Features" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.091, + 0.447, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.171, + 0.442, + 0.182 + ], + "angle": 0, + "content": "(c.) GT Point Cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.186, + 0.47, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Ray-shaped Features vs. Geometric Features. Ray-shaped features show similar feature responses on BEV grids along the same ray; while geometric features from the Latent Rendering maintain discriminative 3D geometry and can describe the 3D world in latent space." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.269, + 0.462, + 0.32 + ], + "angle": 0, + "content": "
Forecasting StructureN/ADifferentiable Ray-castingLatent Rendering
NDS (%)44.1140.20 (-3.91)47.58 (+3.47)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Table 1. Downstream detection performance under different forecasting structures. \"N/A\" represents the baseline without forecasting pre-training. We observe a performance drop when directly using History Encoder and Future Decoder with differentiable ray-casting for pre-training; while, with the Latent Rendering operator, the performance is significantly improved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.493 + ], + "angle": 0, + "content": "this particular grid; the distance expectation function retrieves depths from the occupancy of grids in 3D volume. L1 loss is then applied to supervise the rendered depth for training point cloud forecasting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Despite the great success of differentiable ray-casting in the task of point cloud forecasting, its application in visual point cloud forecasting pre-training does not bring any benefit for downstream performance (Table 1). After such pretraining, ray-shaped features [46, 90], where grids along the same ray tend to possess similar features (Figure 3 - (a.)), are observed. The underlying reason is that waypoints along the same ray in 3D space usually correspond to the same pixel in the visual image, resulting in a tendency to learn similar feature responses. As a consequence, these ray-shaped features are not discriminative and representative enough when transferred to downstream applications, leading to reduced performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.469, + 0.797 + ], + "angle": 0, + "content": "Latent Rendering. In order to extract more discriminative and representative features, we introduce the Latent Rendering operator. It first computes the ray-wise feature through a feature expectation function, then customizes features of each grid by weighting the ray-wise feature with its associated conditional probability. The overall structure is depicted in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.797, + 0.469, + 0.827 + ], + "angle": 0, + "content": "To be specific, inspired by Eq. (4), the feature expectation function is formulated in a similar form:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.836, + 0.469, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} ^ {(i)} = \\sum_ {k = 1} ^ {m} \\hat {\\mathbf {p}} ^ {(i, k)} \\mathcal {F} _ {\\mathrm {b e v}} ^ {(k)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(i\\) represents the ray extending from the origin to the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.095, + 0.891, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.195, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 4. Multi-group Latent Rendering comprises several Latent Rendering running in parallel for different channels. Latent Rendering captures geometric features by the conditional probability function and the feature extraction function. “\\(\\bigoplus\\)” means concatenating multi-group features among channel dimensions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.358 + ], + "angle": 0, + "content": "\\(i\\)-th BEV grid. Here, \\(\\hat{\\mathbf{p}}\\) is the conditional probability, computed through the conditional probability function Eq. (3), which takes as input the learnable independent probability projected from \\(\\mathcal{F}_{bev}\\). The ray-wise features are shared by all grids lying in the same ray." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.359, + 0.771, + 0.374 + ], + "angle": 0, + "content": "Then, we compute the grid feature as:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.385, + 0.892, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {\\mathrm {b e v}} = \\hat {\\mathbf {p}} \\cdot \\hat {\\mathcal {F}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.474 + ], + "angle": 0, + "content": "which highlights the response of BEV grids with higher conditional probability so as to make \\(\\hat{\\mathcal{F}}_{bev}\\) discriminative. This enables the BEV encoder to learn the geometric features during pre-training (Figure 3 - (b))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.549 + ], + "angle": 0, + "content": "To enhance the diversity of geometric features, we further design the multi-group Latent Rendering. By parallelizing multiple Latent Rendering on different feature channels, we allow ray-wise features to maintain diverse information, leading to better downstream performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.746 + ], + "angle": 0, + "content": "As described in Eq. (3), the conditional probability of each BEV grid is determined not only by its own independent response, but also by the response of all its prior grids. Consequently, in the pre-training phase, once the model raises the response of a particular BEV grid, the corresponding responses of all its prior and subsequent grids are suppressed, which mitigates the issue of ray-shaped features during pre-training. After the pre-training with Latent Rendering, it is generally observed that there are only a few peaks with higher responses on a specific ray, indicating the presence of objects or structures in the scene. This effectively promotes a more accurate and consistent understanding of the 3D environment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.658, + 0.77 + ], + "angle": 0, + "content": "3.3. Future Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The Future Decoder predicts the next BEV features \\(\\hat{\\mathcal{F}}_t\\) of frame \\(t\\) based on the inputs of previous BEV latent space \\(\\hat{\\mathcal{F}}_{t - 1}\\) and the expected next ego-motion, \\(\\mathbf{e}_t\\). The predicted features are then used to generate point clouds as Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Architecture. As depicted in Figure 5, Future Decoder is a transformer that can be iteratively used to predict future BEV features from the last roll-out embeddings in an auto-regressive manner. In the \\( t \\)-th iteration, it first encodes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.295 + ], + "angle": 0, + "content": "Figure 5. Future Decoder iteratively predicts the next BEV features, \\(\\hat{\\mathcal{F}}_t\\), from the conditions of ego-motion \\(\\mathbf{e}_t\\) and the last BEV features, to enable specific future predictions with any ego-control." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.449 + ], + "angle": 0, + "content": "the ego-motion condition \\(\\mathbf{e}_t\\), which describes the expected coordinates and heading of ego-vehicle in the next frame, into high-dimensional embeddings by multi-layer perceptron (MLP), which are then added to future BEV queries as inputs of the transformer. Then, 6 transformer layers, composed of a Deformable Self-Attention [93], a Temporal Cross-Attention and a FeedForward Network [14], are used to predict the future \\(\\hat{\\mathcal{F}}_t\\) based on the condition and the last BEV features \\(\\hat{\\mathcal{F}}_{t - 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.632 + ], + "angle": 0, + "content": "The Temporal Cross-Attention layer follows the design of Deformable Cross-Attention [93]. The difference lies in the reference coordinates of query points. In the context of Deformable Cross-Attention [93], \"reference coordinates\" refer to the corresponding positions of query points on the feature maps of keys and values. Typically, they are consistent. However, as for Future Decoder, due to the moving of the ego-vehicle, the ego-coordinate systems between the last and target frame are not necessarily aligned. Therefore, we additionally compute the reference coordinates of future BEV queries in previous BEV feature maps, according to the ego-motion condition, to align coordinate systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.663 + ], + "angle": 0, + "content": "After obtaining the next BEV features \\(\\hat{\\mathcal{F}}_t\\), we use a projection layer to generate the occupancy volume \\(\\mathcal{P}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.471, + 0.846 + ], + "angle": 0, + "content": "Loss. Instead of using L1 loss to supervise the depths of various rays, we directly apply ray-wise cross-entropy loss to maximize the response of points along its corresponding ray, as we have already obtained geometric features after the Latent Rendering operator. To be specific, for each ground-truth point of the \\( t \\)-th future point clouds, we cast a ray from the origin position \\( \\mathbf{o} \\) (the sensor position) towards the point, uniformly sample some waypoints along the ray until out of the volume, and compute cross-entropy loss for the ray to maximize the response of the point position and minimize the response of other waypoint positions. This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.856, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\frac {1}{T n} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\log \\left(\\frac {e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {g} ^ {(i)})}}}{\\sum_ {j} e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {x} ^ {(i , j)})}} + e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {g} ^ {(i)})}}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.17 + ], + "angle": 0, + "content": "where \\(T\\), \\(n\\) indicate the number of future supervisions and the number of points in the \\(t\\)-th ground-truth point clouds. \\(\\mathbf{g}^{(i)}\\) and \\(\\mathbf{x}^{(i,j)}\\) are the coordinates of \\(i\\)-th ground-truth point and \\(j\\)-th waypoints along the same ray. \\(\\mathcal{P}_t^{(\\cdot)}\\) is trilinear interpolation to obtain corresponding values from volume \\(\\mathcal{P}_t\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.185, + 0.634, + 0.203 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.211, + 0.825, + 0.226 + ], + "angle": 0, + "content": "This section investigates the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.229, + 0.892, + 0.26 + ], + "angle": 0, + "content": "- Can future point clouds be estimated from visual history, and how about ViDAR compared to point cloud methods?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.264, + 0.892, + 0.294 + ], + "angle": 0, + "content": "- Can ViDAR help perception, prediction, and planning at the same time so as for scalable autonomous driving?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.299, + 0.891, + 0.328 + ], + "angle": 0, + "content": "- Can ViDAR reduce the reliance of downstream applications on precise human annotations?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.333, + 0.857, + 0.348 + ], + "angle": 0, + "content": "- How do different modules affect final performance?" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.229, + 0.892, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.582, + 0.377 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Dataset. We conduct experiments on the challenging nuScenes [5] dataset, which is a large-scale dataset with 1,000 autonomous driving sequences. This dataset is widely utilized in perception tasks including 3D object detection [28, 43-45, 83-87], multi-object tracking [24, 59, 91], and semantic occupancy prediction [10, 67]. It has also become a popular benchmark for subsequent research on end-to-end autonomous driving, including map segmentation [40, 68], trajectory predictions [17, 47], future occupancy prediction [23, 92], and open-loop planning [26, 27, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Implementation Details. We base our implementation on mmDet3D codebase [9] and conduct downstream verifications on BEVFormer for 3D detection, OccNet [67] for semantic occupancy prediction, and UniAD [27] for unified perception, prediction, and planning. We choose those downstream baselines due to their effectiveness on a wide range of tasks and sharing the same BEV encoder structure, BEVFormer encoder [45]. Without any specifications, the default historical encoder of ViDAR is the BEVFormer-base encoder, consisting of a ResNet101-DCN [11, 19] backbone with an FPN neck [49] and additional 6 encoder layers to extract BEV features from multi-view image sequences, which is consistent with downstream models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To render geometric features, we use a 16-group Latent Rendering. Each group is responsible for rendering latent spaces of 16 channels given the features of 256 channels after the BEVFormer-base encoder. The Future Decoder is a 6-layer structure with a channel of 256 for each. The future BEV queries are \\(200 \\times 200\\) learnable tokens indicating a valid perception range of \\([-51.2\\mathrm{m}, 51.2\\mathrm{m}]\\) for the \\(X\\) and \\(Y\\) axis. We then use a projection layer with output channels as 16 to transform the predicted future BEV features to occupancy volume prediction \\(\\mathcal{P} \\in \\mathbb{R}^{200 \\times 200 \\times 16}\\), where 16 indicates the height dimension with a range of \\([-5\\mathrm{m}, 3\\mathrm{m}]\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14677" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.465, + 0.181 + ], + "angle": 0, + "content": "
History HorizonMethodModalityChamfer Distance (m2) ↓
0.5s1.0s1.5s2.0s2.5s3.0s
1s4D-Occ [36]L1.261.88----
ViDARC1.111.251.401.571.761.97
3s4D-Occ [36]L0.911.131.301.531.722.11
ViDARC1.011.121.251.381.541.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.184, + 0.468, + 0.213 + ], + "angle": 0, + "content": "Table 2. Point cloud forecasting. ViDAR surpasses prior state-of-the-art method on future prediction, using visual input only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.468, + 0.39 + ], + "angle": 0, + "content": "During pre-training, we use 5 frames of historical multiview images and iterate the Future Decoder 6 times to predict point clouds for the future 3 seconds (each frame has 0.5 second interval). In each training step, we randomly select 1 future prediction for computing loss and detach the gradients of the others to save GPU memory. We pre-train the system for 50 epochs by AdamW optimizer [37, 55] with an initial learning of 2e-4 adjusted by cosine annealing strategy. For fine-tuning, we follow the same training strategy as the officially released downstream models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.402, + 0.216, + 0.417 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.468, + 0.546 + ], + "angle": 0, + "content": "We now demonstrate the effectiveness of ViDAR across different tasks. First, we test the ability of ViDAR as a point cloud forecasting framework and compare it with the state-of-the-art approach that uses LiDAR inputs. Then, we show its advancement as a visual autonomous driving pre-training solution. We report the downstream comparison results in the order of perception-prediction-planning with previous state-of-the-art models on the nuScenes validation dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.468, + 0.594 + ], + "angle": 0, + "content": "Downstream Settings. For downstream verifications, we test ViDAR to pre-train BEV encoders under different initialization settings, listed as the following:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.596, + 0.468, + 0.641 + ], + "angle": 0, + "content": "- ViDAR-cls: The BEV encoders are initialized with the backbone pre-trained for ImageNet classification [12], followed by ViDAR pre-training on nuScenes dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.645, + 0.468, + 0.69 + ], + "angle": 0, + "content": "- ViDAR-2D-det: The BEV encoder backbones are first pre-trained for 2D object detection on COCO dataset [48] before ViDAR pre-training on nuScenes dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.696, + 0.468, + 0.741 + ], + "angle": 0, + "content": "- ViDAR-3D-det: The BEV encoder backbones are pretrained first by 3D detection on the nuScenes dataset, using FCOS3D [69], before ViDAR pre-training." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.596, + 0.468, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.468, + 0.805 + ], + "angle": 0, + "content": "For UniAD experiments, after ViDAR pre-training, we first fine-tune BEVFormer for 3D detection, which is then used as the initialization for subsequent two-stage fine-tuning, consistent with the UniAD official implementation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Point Cloud Forecasting. In Table 2, we present comparisons between our ViDAR and the previous state-of-the-art point cloud forecasting method, 4D-Occ [36]. The evaluation metric is Chamfer Distance. We evaluate both methods using input time horizons of 1s and 3s, corresponding" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.09, + 0.888, + 0.255 + ], + "angle": 0, + "content": "
MethodsEncoderPre-trainmAP (%)NDS (%)
BEV-Former [45]RN50 [19]ImageNet-cls [12]25.235.4
ViDAR-cls29.038.8
RN101 [19]ImageNet-cls [12]37.747.7
ViDAR-cls42.651.8
nus-3D-det [69]41.551.7
ViDAR-3D-det45.854.8
Intern-S [71]COCO-2D-det [48]41.551.2
ViDAR-2D-det47.656.4
Intern-B [71]COCO-2D-det [48]42.952.0
ViDAR-2D-det50.157.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.259, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Table 3. Detection performance of BEVFormer with and without ViDAR pre-training under various backbones and initializations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.405 + ], + "angle": 0, + "content": "to input sequences of 2 frames and 6 frames, respectively, following the same setup as in 4D-Occ. To provide detailed comparisons of performance, we report the quantitative forecasting results for each future timestamp. Only points within the range of \\([-51.2\\mathrm{m}, 51.2\\mathrm{m}]\\) on the X- and Y-axis are considered during evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.572 + ], + "angle": 0, + "content": "As presented in Table 2, ViDAR consistently outperforms 4D-Occ on both 1s and 3s settings, despite utilizing visual inputs exclusively. Specifically, with 1s history input, ViDAR achieves remarkable improvement over 4D-Occ, reducing forecasting errors by \\(\\sim 33\\%\\) for future 1s predictions. When using 3s inputs, we observe a \\(\\sim 18\\%\\) error reduction for 3s forecasting. Moreover, due to the autoregressive design of our Future Decoder, ViDAR effectively predicts arbitrary future, though with the constraint of a limited 1s input horizon. These experiments demonstrate the effectiveness of ViDAR for point cloud forecasting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Perception. We verify ViDAR on four downstream perception tasks, 3D object detection, semantic occupancy prediction, map segmentation, and multi-object tracking. In Table 3 and Table 4, we compare the performance of BEVFormer [45] and OccNet [67] with and without ViDAR pre-training under different backbones and initialization settings for 3D detection and semantic occupancy prediction, respectively. Notably, ViDAR, using solely Image-LiDAR sequences, outperforms 3D detection supervised pre-training (The 4th & the 5th row in Table 3, \\(42.6\\%\\) mAP to \\(41.5\\%\\) mAP; The 2nd & the 3rd row in Table 4, \\(29.57\\%\\) mIoU to \\(26.98\\%\\) mIoU). Furthermore, we also observe huge improvements in map segmentation (Table 5) and multi-object tracking (Table 6). These experiments demonstrate the effectiveness of ViDAR as a scalable pretraining method for enhancing 3D geometry modeling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Prediction. The motion forecasting comparisons are presented in Table 7. As shown, ViDAR significantly improves the performance of UniAD [27]. For instance, we observe a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14678" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.462, + 0.226 + ], + "angle": 0, + "content": "
MethodsEncoderPre-trainmIoU (%) ↑
OccNet [67]RN101 [19]ImageNet-cls [12]24.35
ViDAR-cls29.57
nus-3D-det [69]26.98
ViDAR-3D-det31.67
Intern-S [71]COCO-2D-det [48]24.92
ViDAR-2D-det30.51
Intern-B [71]COCO-2D-det [48]25.24
ViDAR-2D-det31.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.229, + 0.468, + 0.256 + ], + "angle": 0, + "content": "Table 4. Semantic Occupancy Prediction. ViDAR consistently improves OccNet on different backbones and initializations." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.263, + 0.471, + 0.33 + ], + "angle": 0, + "content": "
MethodsEncoderPre-trainLanes (%) ↑
BEVFormer [45]RN101nus-3D-det [69]23.9
UniAD [27]RN101nus-3D-det [45]31.3
ViDAR-3D-det33.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.333, + 0.468, + 0.361 + ], + "angle": 0, + "content": "Table 5. Map segmentation. ViDAR improves UniAD on online mapping. The metric is segmentation IoU." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.368, + 0.471, + 0.508 + ], + "angle": 0, + "content": "
MethodsEncoderPre-trainAMOTA (%) ↑
ViP3D [17]RN50nus-3D-det [72]21.7
QD3DT [24]RN101-24.2
MUTR3D [91]RN101ImageNet-cls29.4
DQTrack-DETR3D [44]RN101nus-3D-det [72]36.7
DQTrack-UVTR [44]RN101nus-3D-det [41]39.6
DQTrack-Stereo [44]RN101nus-3D-det [42]40.7
DQTrack-PETr v2 [44]V2-99nus-3D-det [53]44.6
UniAD-Stage1 [27]RN101nus-3D-det [45]39.0
ViDAR-3D-det45.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.51, + 0.468, + 0.539 + ], + "angle": 0, + "content": "Table 6. Multi-object tracking. With ViDAR, UniAD outperforms previous end-to-end trackers using solely visual images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.763 + ], + "angle": 0, + "content": "\\(\\sim 10\\%\\) error reduction in minADE and a \\(3.5\\%\\) improvement in EPA. In Table 8, we provide the comparison between UniAD with and without ViDAR pre-training in future occupancy prediction. The results demonstrate that ViDAR enhances the performance of UniAD for all areas. We observe improvements of \\(2.4\\%\\) IoU and \\(2.7\\%\\) VPQ for nearby areas, as well as improvements of \\(2.0\\%\\) IoU and \\(2.5\\%\\) VPQ for distant areas. With ViDAR pre-training, UniAD overcomes its limitation in occupancy forecasting for distant objects and now outperforms BEVerse [92] in all areas. These experiments highlight the effectiveness of ViDAR in enhancing downstream models in utilizing temporal information and improving their prediction performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Planning. Due to the effective temporal modeling and advanced future prediction capabilities, ViDAR significantly improves UniAD by reducing its average collision rate within 3 seconds by \\(\\sim 15\\%\\). Moreover, it achieves a substantial decrease in the average planning displacement error by \\(0.21\\mathrm{m}\\) and enables UniAD to outperform the state-of-the-art method, VAD [32], on nuScenes open-loop planning evaluation. These improvements demonstrate the effective" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.897, + 0.169 + ], + "angle": 0, + "content": "
MethodsminADE (m) ↓minFDE (m) ↓MR ↓EPA ↑
PnPNet [47]1.151.950.2260.222
ViP3D [17]2.052.840.2460.226
UniAD [27]0.751.080.1580.463
ViDAR0.670.990.1490.498
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.172, + 0.892, + 0.214 + ], + "angle": 0, + "content": "Table 7. Motion forecasting. ViDAR effectively enhances temporal modeling, which in turn boosts UniAD in future motion forecasting, showing consistent improvements on different metrics." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.223, + 0.885, + 0.326 + ], + "angle": 0, + "content": "
MethodsVPQ-n. ↑VPQ-f. ↑IoU-n. ↑IoU-f. ↑
Fiery [23]50.229.959.436.7
StretchBEV [1]46.029.055.537.1
ST-P3 [26]-32.1-38.9
BEVerse [92]54.336.161.440.9
UniAD [27]54.633.962.840.1
ViDAR57.336.465.442.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Table 8. Future occupancy prediction. ViDAR improves UniAD on future occupancy prediction for objects in both near (noted as \"n\", \\(30\\mathrm{x}30\\mathrm{m}\\)) and far (noted as \"f\", \\(50\\mathrm{x}50\\mathrm{m}\\)) evaluation areas." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.38, + 0.902, + 0.484 + ], + "angle": 0, + "content": "
MethodsModalityAvg.Col. (3s) (%) ↓Avg.L2 (3s) (m) ↓
FF [25]L0.431.43
EO [35]L0.331.60
ST-P3 [26]C0.712.11
VAD [32]C0.411.05
UniAD [27]C0.271.12
ViDARC0.230.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Table 9. Planning. ViDAR improves UniAD in terms of both collision avoidance and planning accuracy. Note that, the reported numbers are obtained by averaging the results of each timestamp in the future 3 seconds, which is consistent with the reported results of VAD [32] and ST-P3 [26] at the 3s timestamp instead of the averaged one. Please refer to GitHub:Issue for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.678 + ], + "angle": 0, + "content": "ness of ViDAR as a valuable pre-training approach for end-to-end autonomous driving. The enhanced performance in collision avoidance and planning accuracy highlights the potential of ViDAR in enhancing the safety and efficiency of downstream autonomous driving applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Joint Perception-Prediction-Planning. Finally, we summarize the improvements of ViDAR on the state-of-the-art end-to-end visual autonomous driving system, UniAD [27], for joint perception, prediction, and planning. As depicted in Table 10, ViDAR brings substantial improvements on all sub-modules of UniAD for perception (Detection, Tracking, Mapping), prediction (Motion Forecasting and Future Occupancy Prediction), and planning at the same time. These consistent improvements illustrate that visual point cloud forecasting effectively exploits the information of semantics, 3D geometry, and temporal dynamics behind the easily obtainable Image-LiDAR sequences. This, consequently, enables scalable visual autonomous driving." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14679" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.891, + 0.14 + ], + "angle": 0, + "content": "
MethodDetectionTrackingMappingMotion ForecastingFuture Occupancy PredictionPlanning
NDS ↑mAP ↑AMOTA↑AMOTP↓IDS↓IoU-lane↑IoU-road↑minADE↓minFDE↓MR↓IoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑avg.L2↓avg.Col.↓
UniAD49.3637.9638.31.32105431.369.10.751.080.15862.840.154.633.91.120.27
ViDAR52.5742.3342.01.2599133.271.40.670.990.14965.442.157.336.40.910.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.144, + 0.893, + 0.173 + ], + "angle": 0, + "content": "Table 10. Performance gain of ViDAR for joint perception, prediction, and planning. ViDAR consistently improves UniAD [27] on all tasks towards end-to-end autonomous driving, validating its effectiveness for scalable visual autonomous driving." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.181, + 0.462, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.471, + 0.412 + ], + "angle": 0, + "content": "Figure 6. Validation of ViDAR on Fine-tuning with limited supervised data. We verify ViDAR on supervision efficiency by reducing available annotations for 3D object detection during downstream fine-tuning (from the full training set to a 1/8th subset) and observe a continuous improvement on each subset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.43, + 0.228, + 0.446 + ], + "angle": 0, + "content": "4.3. Ablative Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.529 + ], + "angle": 0, + "content": "We conduct further analysis of ViDAR on improving the performance of downstream models with limited supervised data, and the effect of the Latent Rendering operation on learning 3D geometric latent space. More ablation studies can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.653 + ], + "angle": 0, + "content": "Efficiency of Supervised Pre-training. The primary objective of pre-training is to minimize the dependence on precise 3D annotations. In Figure 6, we demonstrate the effectiveness of ViDAR in reducing the reliance of modern 3D detectors on accurate 3D box annotations. We fine-tune BEVFormer-base using partial 3D annotations on nuScenes, ranging from the full dataset to a \\(1/8\\) subset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.851 + ], + "angle": 0, + "content": "As depicted in Figure 6, ViDAR exhibits a remarkable reduction in the dependence on 3D annotations. Notably, BEVFormer, pre-trained by ViDAR, surpasses its counterpart under full supervision by \\(1.7\\%\\) mAP, while using only half of the supervised samples, i.e., \\(39.4\\%\\) mAP vs. \\(37.7\\%\\) mAP. Consequently, through ViDAR, we can reduce half of 3D annotations without sacrificing precision. Additionally, we observe a consistent trend of increasing improvements as the available supervision decreases. For instance, the mAP improvements are \\(4.9\\%\\), \\(6.5\\%\\), \\(6.7\\%\\), and \\(7.3\\%\\) when fine-tuned on the full, half, a quarter, and 1/8th subsets. These results highlight the potential of ViDAR in harnessing large amounts of Image-LiDAR sequences." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Effect of Latent Rendering operator. Latent Rendering is the key component of ViDAR, which enables visual point" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.183, + 0.895, + 0.234 + ], + "angle": 0, + "content": "
Groups of Latent R.N/A124816
NDS (%)40.2039.1843.3645.5347.0147.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.237, + 0.893, + 0.293 + ], + "angle": 0, + "content": "Table 11. Ablation of Latent Rendering for downstream finetuning. We compare the performance of 3D detection pre-trained by ViDAR without the Latent Rendering operation (denoted as \"N/A\") and by ViDAR with different groups of Latent Rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.436 + ], + "angle": 0, + "content": "cloud forecasting to effectively contribute to downstream applications. It addresses the ray-shaped features issue encountered during pre-training. In Table 11, we verify its effectiveness by comparing the performance of downstream models pre-trained by ViDAR with the Latent Rendering or not. The downstream model is BEVFormer-small [45] for 3D object detection. For context, the baseline performance with ImageNet-cls pre-training is \\(44.11\\%\\) NDS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.544 + ], + "angle": 0, + "content": "As depicted in Table 11, when the Latent Rendering is missing (denoted as \"N/A\"), also referred to as the baseline in Section 3.2, a significant decline is observed in downstream performance after fine-tuning, from \\(44.11\\%\\) NDS to \\(40.20\\%\\) NDS. In contrast, with the 16-group Latent Rendering, the performance improves to \\(47.58\\%\\) NDS, a notable \\(3.47\\%\\) NDS improvement over the baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.608 + ], + "angle": 0, + "content": "We conduct a comparison of Latent Rendering with different parallel groups in Table 11 as well. The results demonstrate a consistent improvement by dividing channels into more groups and integrating information separately." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.619, + 0.644 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this paper, we introduced visual point cloud forecasting, which predicts future point clouds from historical visual images, as a new pre-training task for end-to-end autonomous driving. We developed ViDAR, a general model to pre-train visual BEV encoders, and designed a Latent Rendering operator to solve the ray-shaped feature issue. To conclude, our work demonstrates that visual point cloud forecasting enables scalable autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work. Though with the potential of scalability, in this paper, we mainly conduct pre-training on Image-LiDAR sequences from nuScenes dataset, of which the data scale is still limited. As for the future, we plan to scale up the pre-training data of ViDAR, study visual point cloud forecasting across diverse datasets, and use publicly available Image-LiDAR sequences as much as possible to train a foundation visual autonomous driving model [38]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14680" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Adil Kaan Akan and Fatma Güney. StretchBEV: Stretching Future Instance Prediction Spatially and Temporally. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A Cookbook of Self-supervised Learning. arXiv preprint arXiv:2304.12210, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[4] Alex Bewley, Pei Sun, Thomas Mensink, Dragomir Anguelov, and Cristian Sminchisescu. Range Conditioned Dilated Convolutions for Scale Invariant 3D Object Detection. arXiv preprint arXiv:2005.09927, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi-ancarlo Baldan, and Oscar Beijbom. nuScenes: A Multi-modal Dataset for Autonomous Driving. In CVPR, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.528, + 0.469, + 0.582 + ], + "angle": 0, + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end Autonomous Driving: Challenges and Frontiers. arXiv preprint arXiv:2306.16927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.585, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint arXiv:2003.04297, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.629, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[10] OpenScene Contributors. OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[11] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable Convolutional Networks. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[15] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles Qi, Yin Zhou, Zoey Yang, Aurélien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. I. J. Robotics Res., 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. ViP3D: End-to-end visual trajectory prediction via 3d agent queries. In CVPR, 2023. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[18] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A Survey on Self-supervised Learning: Algorithms, Applications, and Future Trends. arXiv preprint arXiv:2301.05712, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[23] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: Future Instance Segmentation in Bird's-Eye view from Surround Monocular Cameras. In ICCV, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[24] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular Quasi-Dense 3D Object Tracking. TPAMI, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[25] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe Local Motion Planning with Self-Supervised Freespace Forecasting. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[26] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning. In ECCV, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14681" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[30] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. DriveAdapter: Breaking the Coupling Barrier of Perception and Planning in End-to-End Autonomous Driving. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[31] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think Twice before Driving: Towards Scalable Decoders for End-to-End Autonomous Driving. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[33] Li Jiang, Zetong Yang, Shaoshuai Shi, Vladislav Golyanik, Dengxin Dai, and Bernt Schiele. Self-supervised Pretraining with Masked Shape Prediction for 3D Scene Understanding. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.47, + 0.419 + ], + "angle": 0, + "content": "[34] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. arXiv preprint arXiv:2004.11362, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[35] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable Raycasting for Self-Supervised Occupancy Forecasting. In ECCV, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.518 + ], + "angle": 0, + "content": "[36] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point Cloud Forecasting as a Proxy for 4D Occupancy Forecasting. In CVPR, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[37] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[38] Hongyang Li, Yang Li, Huijie Wang, Jia Zeng, Pinlong Cai, Huilin Xu, Dahua Lin, Junchi Yan, Feng Xu, Lu Xiong, Jingdong Wang, Futang Zhu, Kai Yan, Chunjing Xu, Tiancai Wang, Beipeng Mu, Shaoqing Ren, Zhihui Peng, and Yu Qiao. Open-sourced Data Ecosystem in Autonomous Driving: the Present and Future. arXiv preprint arXiv:2312.03408, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[39] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, Hao Tian, Enze Xie, Jiangwei Xie, Li Chen, Tianyu Li, Yang Li, Yulu Gao, Xiaosong Jia, Si Liu, Jianping Shi, Dahua Lin, and Yu Qiao. Delving Into the Devils of Bird's-Eye-View Perception: A Review, Evaluation and Recipe. TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[40] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based Topology Reasoning for Driving Scenes. arXiv preprint arXiv:2304.05277, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[41] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying Voxel-based Representation with Transformer for 3D Object Detection. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[42] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. BEVStereo: Enhancing Depth Estimation in" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Multi-View 3D Object Detection with Temporal Stereo. In AAAI, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[43] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object Detection. In AAAI, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[44] Yanwei Li, Zhiding Yu, Jonah Philion, Anima Anandkumar, Sanja Fidler, Jiaya Jia, and Jose Alvarez. End-to-end 3D Tracking with Decoupled Queries. In ICCV, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[45] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEV-Former: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[46] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and José Manuel Álvarez. FB-BEV: BEV Representation from Forward-Backward View Transformations. ICCV, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. PnPNet: End-to-End Perception and Prediction With Tracking in the Loop. In CVPR, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[48] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[49] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[50] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos. arXiv preprint arXiv:2308.09244, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[51] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully Sparse 3D Panoptic Occupancy Prediction. arXiv preprint arXiv:2312.17118, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[52] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position Embedding Transformation for Multi-View 3D Object Detection. In ECCV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[53] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. In ICCV, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[54] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In ICRA, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[55] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. arXiv preprint arXiv:1711.05101, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[56] B. Mersch, X. Chen, J. Behley, and C. Stachniss. Self-supervised Point Cloud Prediction Using 3D Spatiotemporal Convolutional Networks. In CoRL, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[57] Gregory P. Meyer, Ankit Laddha, Eric Kee, Carlos Vallespi-Gonzalez, and Carl K. Wellington. LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving. In CVPR, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14682" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[58] Chen Min, Liang Xiao, Dawei Zhao, Yiming Nie, and Bin Dai. Occupancy-MAE: Self-Supervised Pre-Training LargeScale LiDAR Point Clouds With Masked Occupancy Autoencoders. TIV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[59] Ziqi Pang, Zhichao Li, and Naiyan Wang. SimpleTrack: Understanding and Rethinking 3D Multi-object Tracking. arXiv preprint arXiv:2111.09621, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[60] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is Pseudo-Lidar Needed for Monocular 3D Object Detection? In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[61] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical Depth DistributionNetwork for Monocular 3D Object Detection. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[62] Ravid Shwartz-Ziv and Yann LeCun. To Compress or Not to Compress-Self-Supervised Learning and Information Theory: A Review. arXiv preprint arXiv:2304.09355, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.317, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[63] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv preprint arXiv:2312.14150, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.386, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[65] Izzeddin Teeti, Salman Khan, Ajmal Shahbaz, Andrew Bradley, and Fabio Cuzzolin. Vision-based Intention and Trajectory Prediction in Autonomous Vehicles: A Survey. In *IJCAI*, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.553, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[66] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive Multiview Coding. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[67] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as Occupancy. In ICCV, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[68] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping. In NeurIPS Datasets and Benchmarks, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[69] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In ICCV Workshops, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[70] Tai Wang, Qing Lian, Chenming Zhu, Xinge Zhu, and Wenwei Zhang. MV-FCOS3D++: Multi-View Camera-Only 4D Object Detection with Pretrained Monocular Backbones. arXiv preprint arXiv:2207.12716, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[71] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions. In CVPR, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[72] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In CoRL, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[73] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[74] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the Pose Forecasting Pipeline with SPF2: Sequential Pointcloud Forecasting for Sequential Pose Forecasting. arXiv preprint arXiv:2003.08376, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[75] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2Net: Stochastic Sequential Pointcloud Forecasting. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.337, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[76] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.892, + 0.449 + ], + "angle": 0, + "content": "[77] Penghao Wu, Li Chen, Hongyang Li, Xiaosong Jia, Junchi Yan, and Yu Qiao. Policy Pre-training for Autonomous Driving via Self-supervised Geometric Modeling. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[79] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.596 + ], + "angle": 0, + "content": "[80] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.666 + ], + "angle": 0, + "content": "[81] Xiangchao Yan, Runjian Chen, Bo Zhang, Jiakang Yuan, Xinyu Cai, Botian Shi, Wenqi Shao, Junchi Yan, Ping Luo, and Yu Qiao. SPOT: Scalable 3D Pre-training via Occupancy Prediction for Autonomous Driving. arXiv preprint arXiv:2309.10527, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.67, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[82] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, Xiaofei He, and Wanli Ouyang. UniPAD: A Universal Pre-training Paradigm for Autonomous Driving. arXiv preprint arXiv:2310.08370, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[83] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. IPOD: Intensive Point-based Object Detector for Point Cloud. CoRR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.825 + ], + "angle": 0, + "content": "[84] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[85] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3DSSD: Point-based 3D Single Stage Object Detector, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[86] Zetong Yang, Yanan Sun, Shu Liu, Xiaojuan Qi, and Jiaya Jia. CN: Channel Normalization for Point Cloud Recognition. In ECCV, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[87] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3D-MAN: 3D Multi-Frame Attention Network for Object Detection. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[88] Zetong Yang, Li Jiang, Yanan Sun, Bernt Schiele, and Jiaya Jia. A Unified Query-based Paradigm for Point Cloud Understanding. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[89] Zetong Yang, Zhiding Yu, Chris Choy, Renhao Wang, Anima Anandkumar, and Jose M. Alvarez. Improving Distant 3D Object Detection Using 2D Box Supervision. arXiv preprint arXiv:2403.09230, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[90] Jia Zeng, Li Chen, Hanming Deng, Lewei Lu, Junchi Yan, Yu Qiao, and Hongyang Li. Distilling Focal Knowledge From Imperfect Expert for 3D Object Detection. In CVPR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[91] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. MUTR3D: A Multi-camera Tracking Framework via 3D-to-2D Queries. In CVPR, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[92] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. BEVerse: Unified Perception and Prediction in Birds-Eye-View for Vision-Centric Autonomous Driving. arXiv preprint arXiv:2205.09743, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[93] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable Transformers for End-to-End Object Detection. In ICLR, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14684" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_origin.pdf b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..060187ed086dd33a311635ccc07f0ad891e79189 --- /dev/null +++ b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/91fee4e7-5b31-4963-9750-c742a8cbf742_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95aaf0330591bafa0cfb4355526eb1e34801dfe43bd0237983df0e4cfdfc5e59 +size 4754286 diff --git a/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/full.md b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b03facca108cf39631cc56a9ed96503d6fc60a6c --- /dev/null +++ b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/full.md @@ -0,0 +1,391 @@ +# Visual Point Cloud Forecasting enables Scalable Autonomous Driving + +Zetong Yang Li Chen Yanan Sun Hongyang Li + +OpenDriveLab and Shanghai AI Lab + +https://github.com/OpenDriveLab/ViDAR + +![](images/72cb6304fddd876554924d65bab7ee67995c4611bf875095a360c61669bcec58.jpg) +Figure 1. ViDAR is a visual autonomous driving pre-training framework, which leverages the estimation of future point clouds from historical visual inputs as the pre-text task. We term this new pre-text task as visual point cloud forecasting. With the aid of ViDAR, we achieve substantial improvement spanning a diverse spectrum of downstream applications for perception, prediction, and planning. + +# Abstract + +In contrast to extensive studies on general vision, pretraining for scalable visual autonomous driving remains seldom explored. Visual autonomous driving applications require features encompassing semantics, 3D geometry, and temporal information simultaneously for joint perception, prediction, and planning, posing dramatic challenges for pre-training. To resolve this, we bring up a new pre-training task termed as visual point cloud forecasting - predicting future point clouds from historical visual input. The key merit of this task captures the synergic learning of semantics, 3D structures, and temporal dynamics. Hence it shows superiority in various downstream tasks. To cope with this new problem, we present ViDAR, a general model to pretrain downstream visual encoders. It first extracts historical embeddings by the encoder. These representations are then transformed to 3D geometric space via a novel Latent Rendering operator for future point cloud prediction. Experiments show significant gain in downstream tasks, e.g., $3.1\%$ NDS on 3D detection, $\sim 10\%$ error reduction on motion forecasting, and $\sim 15\%$ less collision rate on planning. + +# 1. Introduction + +Recently, the community has witnessed rapid development in visual, or camera-only autonomous driving, with input being monocular or multi-view images [50, 63, 69, 70, 76, 89]. Leveraging visual inputs, existing approaches demonstrate superior capability of extracting Bird's-Eye-View (BEV) features [6, 28, 45, 52, 61, 88], and performing well in perception [29, 51, 54, 67], prediction [17, 27, 65], and planning [30, 31]. Despite significant improvements in applications, these models rely on precise 3D annotations to a great extent, which are often difficult to collect, e.g., occupancy [3, 10], 3D boxes [5, 16, 64], trajectories [15], and thus are challenging to scale up for production. + +Considering the expensive labeling workflow, pretraining [2, 18, 62] has emerged as a crucial approach to scale up downstream applications. The key idea is to define pretext tasks that leverage large amounts of readily available data to learn meaningful representations. This enhances downstream performance though labeled data is limited. + +Though extensive research of pre-training in computer vision has been conducted [8, 20-22, 33, 66, 78, 79], its application in visual autonomous driving remains seldom + +explored. Visual autonomous driving poses great challenges in pre-training as it requires the features to maintain semantics, 3D geometry, and temporal dynamics at the same time for joint perception, prediction, and planning [7, 77]. As a result, most models still rely on supervised pre-training, like 3D detection [60, 69] or occupancy [58, 67, 81], using labeled data that is often unavailable at scale [39]. Some approaches propose estimating depth [60] or rendering masked scenes [82] as pre-training. They use Image-LiDAR pairs as a means of achieving scalable annotation-free pre-training. However, they struggle in either multi-view 3D geometry or temporal modeling (Figure 2). Depth estimation retrieves depth from one image, limited in multi-view geometry; rendering techniques reconstruct scenes from multi-view images but lacking temporal modeling. However, temporal modeling is crucial in end-to-end autonomous driving systems, e.g., UniAD [27], especially for prediction and planning which are the ultimate goals and require accurate scene flow and object motion for decision-making. Due to the absence of temporal modeling, existing approaches are insufficient for pretraining the end-to-end system. + +In this work, we explore pre-training tailored for end-to-end visual autonomous driving applications, including not only perception but also prediction and planning [7]. We formulate a new pre-text task, visual point cloud forecasting (Figure 2), to fully exploit information of semantics, 3D geometry, and temporal dynamics behind the raw Image-LiDAR sequences, with being scalable into consideration. It predicts future point clouds from historical visual images. + +The main rationale of visual point cloud forecasting lies in the simultaneous supervision of semantics, 3D structure, and temporal modeling. By compelling the model to predict the future from history, it supervises the extraction of scene flow and object motion which are crucial for temporal modeling and future estimation. Meanwhile, it involves the reconstruction of point clouds from images, which supervises the multi-view geometry and semantic modeling. Therefore, features from visual point cloud forecasting embed information of both geometric and temporal hints, beneficial for perception, tracking, and planning simultaneously. + +To this end, we present ViDAR, a general visual point cloud forecasting approach for pre-training (Figure 2). ViDAR includes three parts, History Encoder, Latent Rendering operator, and Future Decoder. The History Encoder is the target structure for pre-training. It could be any visual BEV encoder [45] to embed visual sequences into BEV space. These BEV features are sent to the Latent Rendering operator. Latent Rendering plays a crucial role in enabling ViDAR benefit downstream performance. It solves the ray-shaped BEV features issue [46, 90], models 3D geometric latent space, and bridges encoder and decoder. The Future Decoder is an auto-regressive transformer that takes histor + +![](images/e2e1fa4cb319e04f3374b2a35ab5b3220e7182d12c0c67ba65e33fd367d70464.jpg) + +![](images/babfffed185465b47f0547f3ecd338f53020387c1f00539a1c3e95ce6e60e2e5.jpg) + +![](images/79a99e36f19e574617e2a210beb8cdbe7787512e7f28458008c205b34d4f5815.jpg) +Figure 2. Comparisons among visual autonomous driving pretraining paradigms and our ViDAR architecture. Compared to existing methods, visual point cloud forecasting jointly models multi-view geometry and temporal dynamics. We then propose ViDAR, using Image-LiDAR sequences to pre-train visual encoders. + +ical BEV features to iteratively predict future point clouds for arbitrary timestamps. + +ViDAR provides a comprehensive solution for visual autonomous driving pre-training. We test ViDAR on nuScenes dataset [5] in terms of point cloud forecasting and downstream verifications. Though with visual inputs, ViDAR still outperforms previous forecasting methods using point clouds, $\sim 33\%$ Chamfer Distance reduction on point cloud estimation of 1s future. ViDAR also improves downstream performance. Using Image-LiDAR sequences only, ViDAR surprisingly outperforms 3D detection pre-training [69], e.g., by $1.1\%$ mAP and $2.5\%$ mIoU for detection and semantic occupancy prediction, under the same data scale. If also based on the 3D detection pre-training, ViDAR boosts previous methods by $4.3\%$ mAP and $4.6\%$ mIoU. Further, due to the effective pre-training on the joint capture of geometric and temporal information, ViDAR improves UniAD [27] on all tasks for end-to-end autonomous driving including perception, prediction, and planning by a large margin (Figure 1). Experimental results validate that visual point cloud forecasting enables scalable autonomous driving. + +# 2. Related Work + +Pre-training for Visual Autonomous Driving. Pretraining for scalable applications has been extensively studied in general vision. These approaches can be roughly divided into contrastive approaches [8, 20, 34, 66], which learn discriminative features from positive and negative pairs; and masked signal modeling [13, 21, 73, 80], which recover discarded signals from remained signals to capture a comprehensive understanding of global semantics. + +In contrast, pre-training for visual autonomous driving is still under-explored. Visual autonomous driving poses great challenges as it requires semantic understanding, 3D structural awareness, and temporal modeling at the same time, + +for joint perception, prediction, and planning. Existing vision methods mainly consider semantics; methods based on Image-LiDAR pairs [60, 82] struggle with temporal modeling; other supervised strategies [67, 69] are not scalable. In this work, we propose visual point cloud forecasting, which simultaneously models semantics, temporal dynamics, and 3D geometry by a uniform process, and is easily scaled up. + +Point Cloud Forecasting. Point cloud forecasting, one of the most fundamental self-supervised tasks for autonomous driving, predicts future point clouds from past point cloud inputs. Previous works use range image [4, 57], a representation obtained by projecting point clouds to dense 2D images using sensor intrinsic and extrinsic parameters. Based on historical range images, they apply 3D convolutions [56], or LSTMs [74, 75] to predict future point clouds. Yet, they additionally model the motion of sensor intrinsic and extrinsic parameters. Later methods factor out the estimation of sensors by introducing 4D occupancy prediction [36] and differentiable ray-casting [35], which ensures a better modeling of the world. Compared to prior literature, we aim at visual point cloud forecasting, using past images to predict future point clouds. Meanwhile, we raise this task as a pre-training paradigm for visual autonomous driving and demonstrate its superiority in a wide range of downstream applications. + +# 3. Methodology + +In this section, we elaborate on our ViDAR, a visual point cloud forecasting approach for general autonomous driving pre-training. We begin with an overview of ViDAR in Section 3.1, and subsequently delve into Latent Rendering and Future Decoder in Section 3.2 and Section 3.3, respectively. + +# 3.1. Overview + +As depicted in Figure 2, ViDAR comprises three components: (a) a History Encoder, also the target structure of pre-training, which extracts BEV embeddings $\mathcal{F}_{\mathrm{bev}}$ from visual sequence inputs $\mathcal{I}$ ; It can be any visual BEV encoder [28, 45, 52]; (b) a Latent Rendering operator, which simulates the volume rendering operation in latent space so as to obtain geometric embedding $\hat{\mathcal{F}}_{\mathrm{bev}}$ from $\mathcal{F}_{\mathrm{bev}}$ ; and (c) a Future Decoder, which predicts future BEV features $\hat{\mathcal{F}}_t$ at timestamps $t \in \{1, 2, \ldots\}$ in an auto-regressive manner. Finally, a prediction head is followed to project $\hat{\mathcal{F}}_t$ into 3D occupancy volume $\mathcal{P}_t$ . This process is formulated as: + +$$ +\mathcal {F} _ {\mathrm {b e v}} = \operatorname {E n c o d e r} (\mathcal {I}), +$$ + +$$ +\hat {\mathcal {F}} _ {\text {b e v}} = \text {L a t e n t R e n d e r} (\mathcal {F} _ {\text {b e v}}), \tag {1} +$$ + +$$ +\hat {\mathcal {F}} _ {t} = \operatorname {D e c o d e r} \left(\hat {\mathcal {F}} _ {t - 1}\right), \text {w h e r e} \hat {\mathcal {F}} _ {0} = \hat {\mathcal {F}} _ {\text {b e v}}, +$$ + +$$ +\mathcal {P} _ {t} = \text {P r o j e c t i o n} (\hat {\mathcal {F}} _ {t}). +$$ + +Point cloud predictions are obtained from the predicted occupancy volume $\mathcal{P}_t$ . This process is similar to the previous + +point cloud forecasting method [36]. Specifically, we first cast rays from the origin to various designated directions, then figure out the distance of waypoints along each ray with the maximum occupancy response, and finally compute the point position according to the distance and corresponding ray direction. + +# 3.2. Latent Rendering + +A straightforward solution of visual point cloud forecasting for pre-training is to incorporate the History Encoder and Future Decoder directly with differentiable ray-casting [36], which is the crucial component in state-of-the-art point cloud forecasting methods to render point clouds from predicted occupancy volume and compute loss for backpropagation. However, our experimental results show that this approach does not yield improvements and even has a detrimental effect on the downstream tasks due to the defective geometric feature modeling ability. + +Preliminary. Differentiable ray-casting is a volume rendering process operating on an occupancy volume, denoted as $\mathcal{P} \in \mathbb{R}^{L \times H \times W}$ . It renders depths of various rays and subsequently converts the depths with corresponding ray directions to point clouds. + +Formally, starting from the origin sensor position, $\mathbf{o} \in \mathbb{R}^3$ , differentiable ray-casting casts $n$ rays with varying directions, $\mathbf{d} \in \mathbb{R}^{n \times 3}$ . Along each ray $i$ , it uniformly samples $m$ waypoints at different distances $\lambda^{(j)} \in \mathbb{R}, j \in \{1,2,\dots,m\}$ until reaching the boundary of the 3D space. The coordinates of these waypoints are calculated as: + +$$ +\mathbf {x} ^ {(i, j)} = \mathbf {o} + \lambda^ {(j)} \mathbf {d} ^ {(i)}. \tag {2} +$$ + +Those waypoints coordinates, $\mathbf{x} \in \mathbb{R}^{n \times m \times 3}$ , are used to compute occupancy values. This process is quantized [36], wherein the waypoints are discretized to occupancy volume grids. Then, the values of waypoints are derived as the associated values of volume grids, $\mathbf{p}^{(i,j)} = \mathcal{P}^{([\mathbf{x}^{(i,j)}])}$ . Here $[\cdot]$ denotes a rounding operation for discretizing waypoints. + +Differentiable ray-casting renders the corresponding depth of the $i$ -th ray, $\hat{\lambda}^{(i)}$ , by an integral process: + +$$ +\hat {\mathbf {p}} ^ {(i, j)} = \left[ \prod_ {k = 1} ^ {j - 1} (1 - \mathbf {p} ^ {(i, k)}) \right] \mathbf {p} ^ {(i, j)}, \tag {3} +$$ + +$$ +\hat {\lambda} ^ {(i)} = \sum_ {j = 1} ^ {m} \hat {\mathbf {p}} ^ {(i, j)} \lambda^ {(j)}. \tag {4} +$$ + +For simplicity, we name Eq. 3 and Eq. 4 as conditional probability function and distance expectation function. The conditional probability function determines the occupancy of a grid by considering the conditional probability that prior waypoints are unoccupied and the ray terminates at + +![](images/e757736429c0ca6d831c31ef263a2fd76473ac87aedea29fa2a46a54a5fb3119.jpg) +(a.) Ray-shaped Features + +![](images/c2361424183af82592f556eb04723e204ee444e99696278d01300a7ca464b9e7.jpg) +(b.) Geometric Features +Figure 3. Ray-shaped Features vs. Geometric Features. Ray-shaped features show similar feature responses on BEV grids along the same ray; while geometric features from the Latent Rendering maintain discriminative 3D geometry and can describe the 3D world in latent space. + +![](images/4d5103f45b2db68050a79dbcadb34ca08671ee1f034d235325136db0a20ba0d5.jpg) +(c.) GT Point Cloud + +
Forecasting StructureN/ADifferentiable Ray-castingLatent Rendering
NDS (%)44.1140.20 (-3.91)47.58 (+3.47)
+ +Table 1. Downstream detection performance under different forecasting structures. "N/A" represents the baseline without forecasting pre-training. We observe a performance drop when directly using History Encoder and Future Decoder with differentiable ray-casting for pre-training; while, with the Latent Rendering operator, the performance is significantly improved. + +this particular grid; the distance expectation function retrieves depths from the occupancy of grids in 3D volume. L1 loss is then applied to supervise the rendered depth for training point cloud forecasting. + +Despite the great success of differentiable ray-casting in the task of point cloud forecasting, its application in visual point cloud forecasting pre-training does not bring any benefit for downstream performance (Table 1). After such pretraining, ray-shaped features [46, 90], where grids along the same ray tend to possess similar features (Figure 3 - (a.)), are observed. The underlying reason is that waypoints along the same ray in 3D space usually correspond to the same pixel in the visual image, resulting in a tendency to learn similar feature responses. As a consequence, these ray-shaped features are not discriminative and representative enough when transferred to downstream applications, leading to reduced performance. + +Latent Rendering. In order to extract more discriminative and representative features, we introduce the Latent Rendering operator. It first computes the ray-wise feature through a feature expectation function, then customizes features of each grid by weighting the ray-wise feature with its associated conditional probability. The overall structure is depicted in Figure 4. + +To be specific, inspired by Eq. (4), the feature expectation function is formulated in a similar form: + +$$ +\hat {\mathcal {F}} ^ {(i)} = \sum_ {k = 1} ^ {m} \hat {\mathbf {p}} ^ {(i, k)} \mathcal {F} _ {\mathrm {b e v}} ^ {(k)}, \tag {5} +$$ + +where $i$ represents the ray extending from the origin to the + +![](images/fe47be17e5e3c2f68533ab0d6f8a62b1781e37336e1b920caf564102b16c74ec.jpg) +Figure 4. Multi-group Latent Rendering comprises several Latent Rendering running in parallel for different channels. Latent Rendering captures geometric features by the conditional probability function and the feature extraction function. “ $\bigoplus$ ” means concatenating multi-group features among channel dimensions. + +$i$ -th BEV grid. Here, $\hat{\mathbf{p}}$ is the conditional probability, computed through the conditional probability function Eq. (3), which takes as input the learnable independent probability projected from $\mathcal{F}_{bev}$ . The ray-wise features are shared by all grids lying in the same ray. + +Then, we compute the grid feature as: + +$$ +\hat {\mathcal {F}} _ {\mathrm {b e v}} = \hat {\mathbf {p}} \cdot \hat {\mathcal {F}}, \tag {6} +$$ + +which highlights the response of BEV grids with higher conditional probability so as to make $\hat{\mathcal{F}}_{bev}$ discriminative. This enables the BEV encoder to learn the geometric features during pre-training (Figure 3 - (b)). + +To enhance the diversity of geometric features, we further design the multi-group Latent Rendering. By parallelizing multiple Latent Rendering on different feature channels, we allow ray-wise features to maintain diverse information, leading to better downstream performance. + +As described in Eq. (3), the conditional probability of each BEV grid is determined not only by its own independent response, but also by the response of all its prior grids. Consequently, in the pre-training phase, once the model raises the response of a particular BEV grid, the corresponding responses of all its prior and subsequent grids are suppressed, which mitigates the issue of ray-shaped features during pre-training. After the pre-training with Latent Rendering, it is generally observed that there are only a few peaks with higher responses on a specific ray, indicating the presence of objects or structures in the scene. This effectively promotes a more accurate and consistent understanding of the 3D environment. + +# 3.3. Future Decoder + +The Future Decoder predicts the next BEV features $\hat{\mathcal{F}}_t$ of frame $t$ based on the inputs of previous BEV latent space $\hat{\mathcal{F}}_{t - 1}$ and the expected next ego-motion, $\mathbf{e}_t$ . The predicted features are then used to generate point clouds as Eq. (1). + +Architecture. As depicted in Figure 5, Future Decoder is a transformer that can be iteratively used to predict future BEV features from the last roll-out embeddings in an auto-regressive manner. In the $t$ -th iteration, it first encodes + +![](images/9c4188f8cdd74035c8cb8c0d03b7f996d3e73b5526d9d6d9b3dd0a1e6c9b7765.jpg) +Figure 5. Future Decoder iteratively predicts the next BEV features, $\hat{\mathcal{F}}_t$ , from the conditions of ego-motion $\mathbf{e}_t$ and the last BEV features, to enable specific future predictions with any ego-control. + +the ego-motion condition $\mathbf{e}_t$ , which describes the expected coordinates and heading of ego-vehicle in the next frame, into high-dimensional embeddings by multi-layer perceptron (MLP), which are then added to future BEV queries as inputs of the transformer. Then, 6 transformer layers, composed of a Deformable Self-Attention [93], a Temporal Cross-Attention and a FeedForward Network [14], are used to predict the future $\hat{\mathcal{F}}_t$ based on the condition and the last BEV features $\hat{\mathcal{F}}_{t - 1}$ . + +The Temporal Cross-Attention layer follows the design of Deformable Cross-Attention [93]. The difference lies in the reference coordinates of query points. In the context of Deformable Cross-Attention [93], "reference coordinates" refer to the corresponding positions of query points on the feature maps of keys and values. Typically, they are consistent. However, as for Future Decoder, due to the moving of the ego-vehicle, the ego-coordinate systems between the last and target frame are not necessarily aligned. Therefore, we additionally compute the reference coordinates of future BEV queries in previous BEV feature maps, according to the ego-motion condition, to align coordinate systems. + +After obtaining the next BEV features $\hat{\mathcal{F}}_t$ , we use a projection layer to generate the occupancy volume $\mathcal{P}_t$ . + +Loss. Instead of using L1 loss to supervise the depths of various rays, we directly apply ray-wise cross-entropy loss to maximize the response of points along its corresponding ray, as we have already obtained geometric features after the Latent Rendering operator. To be specific, for each ground-truth point of the $t$ -th future point clouds, we cast a ray from the origin position $\mathbf{o}$ (the sensor position) towards the point, uniformly sample some waypoints along the ray until out of the volume, and compute cross-entropy loss for the ray to maximize the response of the point position and minimize the response of other waypoint positions. This process is formulated as: + +$$ +\mathcal {L} = - \frac {1}{T n} \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {n} \log \left(\frac {e ^ {\mathcal {P} _ {t} ^ {(\mathbf {g} ^ {(i)})}}}{\sum_ {j} e ^ {\mathcal {P} _ {t} ^ {(\mathbf {x} ^ {(i , j)})}} + e ^ {\mathcal {P} _ {t} ^ {(\mathbf {g} ^ {(i)})}}}\right), \tag {7} +$$ + +where $T$ , $n$ indicate the number of future supervisions and the number of points in the $t$ -th ground-truth point clouds. $\mathbf{g}^{(i)}$ and $\mathbf{x}^{(i,j)}$ are the coordinates of $i$ -th ground-truth point and $j$ -th waypoints along the same ray. $\mathcal{P}_t^{(\cdot)}$ is trilinear interpolation to obtain corresponding values from volume $\mathcal{P}_t$ . + +# 4. Experiments + +This section investigates the following questions: + +- Can future point clouds be estimated from visual history, and how about ViDAR compared to point cloud methods? +- Can ViDAR help perception, prediction, and planning at the same time so as for scalable autonomous driving? +- Can ViDAR reduce the reliance of downstream applications on precise human annotations? +- How do different modules affect final performance? + +# 4.1. Setup + +Dataset. We conduct experiments on the challenging nuScenes [5] dataset, which is a large-scale dataset with 1,000 autonomous driving sequences. This dataset is widely utilized in perception tasks including 3D object detection [28, 43-45, 83-87], multi-object tracking [24, 59, 91], and semantic occupancy prediction [10, 67]. It has also become a popular benchmark for subsequent research on end-to-end autonomous driving, including map segmentation [40, 68], trajectory predictions [17, 47], future occupancy prediction [23, 92], and open-loop planning [26, 27, 32]. + +Implementation Details. We base our implementation on mmDet3D codebase [9] and conduct downstream verifications on BEVFormer for 3D detection, OccNet [67] for semantic occupancy prediction, and UniAD [27] for unified perception, prediction, and planning. We choose those downstream baselines due to their effectiveness on a wide range of tasks and sharing the same BEV encoder structure, BEVFormer encoder [45]. Without any specifications, the default historical encoder of ViDAR is the BEVFormer-base encoder, consisting of a ResNet101-DCN [11, 19] backbone with an FPN neck [49] and additional 6 encoder layers to extract BEV features from multi-view image sequences, which is consistent with downstream models. + +To render geometric features, we use a 16-group Latent Rendering. Each group is responsible for rendering latent spaces of 16 channels given the features of 256 channels after the BEVFormer-base encoder. The Future Decoder is a 6-layer structure with a channel of 256 for each. The future BEV queries are $200 \times 200$ learnable tokens indicating a valid perception range of $[-51.2\mathrm{m}, 51.2\mathrm{m}]$ for the $X$ and $Y$ axis. We then use a projection layer with output channels as 16 to transform the predicted future BEV features to occupancy volume prediction $\mathcal{P} \in \mathbb{R}^{200 \times 200 \times 16}$ , where 16 indicates the height dimension with a range of $[-5\mathrm{m}, 3\mathrm{m}]$ . + +
History HorizonMethodModalityChamfer Distance (m2) ↓
0.5s1.0s1.5s2.0s2.5s3.0s
1s4D-Occ [36]L1.261.88----
ViDARC1.111.251.401.571.761.97
3s4D-Occ [36]L0.911.131.301.531.722.11
ViDARC1.011.121.251.381.541.73
+ +During pre-training, we use 5 frames of historical multiview images and iterate the Future Decoder 6 times to predict point clouds for the future 3 seconds (each frame has 0.5 second interval). In each training step, we randomly select 1 future prediction for computing loss and detach the gradients of the others to save GPU memory. We pre-train the system for 50 epochs by AdamW optimizer [37, 55] with an initial learning of 2e-4 adjusted by cosine annealing strategy. For fine-tuning, we follow the same training strategy as the officially released downstream models. + +# 4.2. Main Results + +We now demonstrate the effectiveness of ViDAR across different tasks. First, we test the ability of ViDAR as a point cloud forecasting framework and compare it with the state-of-the-art approach that uses LiDAR inputs. Then, we show its advancement as a visual autonomous driving pre-training solution. We report the downstream comparison results in the order of perception-prediction-planning with previous state-of-the-art models on the nuScenes validation dataset. + +Downstream Settings. For downstream verifications, we test ViDAR to pre-train BEV encoders under different initialization settings, listed as the following: + +- ViDAR-cls: The BEV encoders are initialized with the backbone pre-trained for ImageNet classification [12], followed by ViDAR pre-training on nuScenes dataset. +- ViDAR-2D-det: The BEV encoder backbones are first pre-trained for 2D object detection on COCO dataset [48] before ViDAR pre-training on nuScenes dataset. +- ViDAR-3D-det: The BEV encoder backbones are pretrained first by 3D detection on the nuScenes dataset, using FCOS3D [69], before ViDAR pre-training. + +For UniAD experiments, after ViDAR pre-training, we first fine-tune BEVFormer for 3D detection, which is then used as the initialization for subsequent two-stage fine-tuning, consistent with the UniAD official implementation. + +Point Cloud Forecasting. In Table 2, we present comparisons between our ViDAR and the previous state-of-the-art point cloud forecasting method, 4D-Occ [36]. The evaluation metric is Chamfer Distance. We evaluate both methods using input time horizons of 1s and 3s, corresponding + +Table 2. Point cloud forecasting. ViDAR surpasses prior state-of-the-art method on future prediction, using visual input only. + +
MethodsEncoderPre-trainmAP (%)NDS (%)
BEV-Former [45]RN50 [19]ImageNet-cls [12]25.235.4
ViDAR-cls29.038.8
RN101 [19]ImageNet-cls [12]37.747.7
ViDAR-cls42.651.8
nus-3D-det [69]41.551.7
ViDAR-3D-det45.854.8
Intern-S [71]COCO-2D-det [48]41.551.2
ViDAR-2D-det47.656.4
Intern-B [71]COCO-2D-det [48]42.952.0
ViDAR-2D-det50.157.6
+ +Table 3. Detection performance of BEVFormer with and without ViDAR pre-training under various backbones and initializations. + +to input sequences of 2 frames and 6 frames, respectively, following the same setup as in 4D-Occ. To provide detailed comparisons of performance, we report the quantitative forecasting results for each future timestamp. Only points within the range of $[-51.2\mathrm{m}, 51.2\mathrm{m}]$ on the X- and Y-axis are considered during evaluation. + +As presented in Table 2, ViDAR consistently outperforms 4D-Occ on both 1s and 3s settings, despite utilizing visual inputs exclusively. Specifically, with 1s history input, ViDAR achieves remarkable improvement over 4D-Occ, reducing forecasting errors by $\sim 33\%$ for future 1s predictions. When using 3s inputs, we observe a $\sim 18\%$ error reduction for 3s forecasting. Moreover, due to the autoregressive design of our Future Decoder, ViDAR effectively predicts arbitrary future, though with the constraint of a limited 1s input horizon. These experiments demonstrate the effectiveness of ViDAR for point cloud forecasting. + +Perception. We verify ViDAR on four downstream perception tasks, 3D object detection, semantic occupancy prediction, map segmentation, and multi-object tracking. In Table 3 and Table 4, we compare the performance of BEVFormer [45] and OccNet [67] with and without ViDAR pre-training under different backbones and initialization settings for 3D detection and semantic occupancy prediction, respectively. Notably, ViDAR, using solely Image-LiDAR sequences, outperforms 3D detection supervised pre-training (The 4th & the 5th row in Table 3, $42.6\%$ mAP to $41.5\%$ mAP; The 2nd & the 3rd row in Table 4, $29.57\%$ mIoU to $26.98\%$ mIoU). Furthermore, we also observe huge improvements in map segmentation (Table 5) and multi-object tracking (Table 6). These experiments demonstrate the effectiveness of ViDAR as a scalable pretraining method for enhancing 3D geometry modeling. + +Prediction. The motion forecasting comparisons are presented in Table 7. As shown, ViDAR significantly improves the performance of UniAD [27]. For instance, we observe a + +
MethodsEncoderPre-trainmIoU (%) ↑
OccNet [67]RN101 [19]ImageNet-cls [12]24.35
ViDAR-cls29.57
nus-3D-det [69]26.98
ViDAR-3D-det31.67
Intern-S [71]COCO-2D-det [48]24.92
ViDAR-2D-det30.51
Intern-B [71]COCO-2D-det [48]25.24
ViDAR-2D-det31.69
+ +Table 4. Semantic Occupancy Prediction. ViDAR consistently improves OccNet on different backbones and initializations. + +
MethodsEncoderPre-trainLanes (%) ↑
BEVFormer [45]RN101nus-3D-det [69]23.9
UniAD [27]RN101nus-3D-det [45]31.3
ViDAR-3D-det33.2
+ +Table 5. Map segmentation. ViDAR improves UniAD on online mapping. The metric is segmentation IoU. + +
MethodsEncoderPre-trainAMOTA (%) ↑
ViP3D [17]RN50nus-3D-det [72]21.7
QD3DT [24]RN101-24.2
MUTR3D [91]RN101ImageNet-cls29.4
DQTrack-DETR3D [44]RN101nus-3D-det [72]36.7
DQTrack-UVTR [44]RN101nus-3D-det [41]39.6
DQTrack-Stereo [44]RN101nus-3D-det [42]40.7
DQTrack-PETr v2 [44]V2-99nus-3D-det [53]44.6
UniAD-Stage1 [27]RN101nus-3D-det [45]39.0
ViDAR-3D-det45.1
+ +$\sim 10\%$ error reduction in minADE and a $3.5\%$ improvement in EPA. In Table 8, we provide the comparison between UniAD with and without ViDAR pre-training in future occupancy prediction. The results demonstrate that ViDAR enhances the performance of UniAD for all areas. We observe improvements of $2.4\%$ IoU and $2.7\%$ VPQ for nearby areas, as well as improvements of $2.0\%$ IoU and $2.5\%$ VPQ for distant areas. With ViDAR pre-training, UniAD overcomes its limitation in occupancy forecasting for distant objects and now outperforms BEVerse [92] in all areas. These experiments highlight the effectiveness of ViDAR in enhancing downstream models in utilizing temporal information and improving their prediction performance. + +Planning. Due to the effective temporal modeling and advanced future prediction capabilities, ViDAR significantly improves UniAD by reducing its average collision rate within 3 seconds by $\sim 15\%$ . Moreover, it achieves a substantial decrease in the average planning displacement error by $0.21\mathrm{m}$ and enables UniAD to outperform the state-of-the-art method, VAD [32], on nuScenes open-loop planning evaluation. These improvements demonstrate the effective + +Table 6. Multi-object tracking. With ViDAR, UniAD outperforms previous end-to-end trackers using solely visual images. + +
MethodsminADE (m) ↓minFDE (m) ↓MR ↓EPA ↑
PnPNet [47]1.151.950.2260.222
ViP3D [17]2.052.840.2460.226
UniAD [27]0.751.080.1580.463
ViDAR0.670.990.1490.498
+ +Table 7. Motion forecasting. ViDAR effectively enhances temporal modeling, which in turn boosts UniAD in future motion forecasting, showing consistent improvements on different metrics. + +
MethodsVPQ-n. ↑VPQ-f. ↑IoU-n. ↑IoU-f. ↑
Fiery [23]50.229.959.436.7
StretchBEV [1]46.029.055.537.1
ST-P3 [26]-32.1-38.9
BEVerse [92]54.336.161.440.9
UniAD [27]54.633.962.840.1
ViDAR57.336.465.442.1
+ +Table 8. Future occupancy prediction. ViDAR improves UniAD on future occupancy prediction for objects in both near (noted as "n", $30\mathrm{x}30\mathrm{m}$ ) and far (noted as "f", $50\mathrm{x}50\mathrm{m}$ ) evaluation areas. + +
MethodsModalityAvg.Col. (3s) (%) ↓Avg.L2 (3s) (m) ↓
FF [25]L0.431.43
EO [35]L0.331.60
ST-P3 [26]C0.712.11
VAD [32]C0.411.05
UniAD [27]C0.271.12
ViDARC0.230.91
+ +Table 9. Planning. ViDAR improves UniAD in terms of both collision avoidance and planning accuracy. Note that, the reported numbers are obtained by averaging the results of each timestamp in the future 3 seconds, which is consistent with the reported results of VAD [32] and ST-P3 [26] at the 3s timestamp instead of the averaged one. Please refer to GitHub:Issue for more details. + +ness of ViDAR as a valuable pre-training approach for end-to-end autonomous driving. The enhanced performance in collision avoidance and planning accuracy highlights the potential of ViDAR in enhancing the safety and efficiency of downstream autonomous driving applications. + +Joint Perception-Prediction-Planning. Finally, we summarize the improvements of ViDAR on the state-of-the-art end-to-end visual autonomous driving system, UniAD [27], for joint perception, prediction, and planning. As depicted in Table 10, ViDAR brings substantial improvements on all sub-modules of UniAD for perception (Detection, Tracking, Mapping), prediction (Motion Forecasting and Future Occupancy Prediction), and planning at the same time. These consistent improvements illustrate that visual point cloud forecasting effectively exploits the information of semantics, 3D geometry, and temporal dynamics behind the easily obtainable Image-LiDAR sequences. This, consequently, enables scalable visual autonomous driving. + +
MethodDetectionTrackingMappingMotion ForecastingFuture Occupancy PredictionPlanning
NDS ↑mAP ↑AMOTA↑AMOTP↓IDS↓IoU-lane↑IoU-road↑minADE↓minFDE↓MR↓IoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑avg.L2↓avg.Col.↓
UniAD49.3637.9638.31.32105431.369.10.751.080.15862.840.154.633.91.120.27
ViDAR52.5742.3342.01.2599133.271.40.670.990.14965.442.157.336.40.910.23
+ +![](images/77b0bc98d2b4f373136843570bc9cd8b65772513a460b3e37f860e3127d8deaf.jpg) +Figure 6. Validation of ViDAR on Fine-tuning with limited supervised data. We verify ViDAR on supervision efficiency by reducing available annotations for 3D object detection during downstream fine-tuning (from the full training set to a 1/8th subset) and observe a continuous improvement on each subset. + +# 4.3. Ablative Study + +We conduct further analysis of ViDAR on improving the performance of downstream models with limited supervised data, and the effect of the Latent Rendering operation on learning 3D geometric latent space. More ablation studies can be found in the supplementary materials. + +Efficiency of Supervised Pre-training. The primary objective of pre-training is to minimize the dependence on precise 3D annotations. In Figure 6, we demonstrate the effectiveness of ViDAR in reducing the reliance of modern 3D detectors on accurate 3D box annotations. We fine-tune BEVFormer-base using partial 3D annotations on nuScenes, ranging from the full dataset to a $1/8$ subset. + +As depicted in Figure 6, ViDAR exhibits a remarkable reduction in the dependence on 3D annotations. Notably, BEVFormer, pre-trained by ViDAR, surpasses its counterpart under full supervision by $1.7\%$ mAP, while using only half of the supervised samples, i.e., $39.4\%$ mAP vs. $37.7\%$ mAP. Consequently, through ViDAR, we can reduce half of 3D annotations without sacrificing precision. Additionally, we observe a consistent trend of increasing improvements as the available supervision decreases. For instance, the mAP improvements are $4.9\%$ , $6.5\%$ , $6.7\%$ , and $7.3\%$ when fine-tuned on the full, half, a quarter, and 1/8th subsets. These results highlight the potential of ViDAR in harnessing large amounts of Image-LiDAR sequences. + +Effect of Latent Rendering operator. Latent Rendering is the key component of ViDAR, which enables visual point + +Table 10. Performance gain of ViDAR for joint perception, prediction, and planning. ViDAR consistently improves UniAD [27] on all tasks towards end-to-end autonomous driving, validating its effectiveness for scalable visual autonomous driving. + +
Groups of Latent R.N/A124816
NDS (%)40.2039.1843.3645.5347.0147.58
+ +Table 11. Ablation of Latent Rendering for downstream finetuning. We compare the performance of 3D detection pre-trained by ViDAR without the Latent Rendering operation (denoted as "N/A") and by ViDAR with different groups of Latent Rendering. + +cloud forecasting to effectively contribute to downstream applications. It addresses the ray-shaped features issue encountered during pre-training. In Table 11, we verify its effectiveness by comparing the performance of downstream models pre-trained by ViDAR with the Latent Rendering or not. The downstream model is BEVFormer-small [45] for 3D object detection. For context, the baseline performance with ImageNet-cls pre-training is $44.11\%$ NDS. + +As depicted in Table 11, when the Latent Rendering is missing (denoted as "N/A"), also referred to as the baseline in Section 3.2, a significant decline is observed in downstream performance after fine-tuning, from $44.11\%$ NDS to $40.20\%$ NDS. In contrast, with the 16-group Latent Rendering, the performance improves to $47.58\%$ NDS, a notable $3.47\%$ NDS improvement over the baseline. + +We conduct a comparison of Latent Rendering with different parallel groups in Table 11 as well. The results demonstrate a consistent improvement by dividing channels into more groups and integrating information separately. + +# 5. Conclusion + +In this paper, we introduced visual point cloud forecasting, which predicts future point clouds from historical visual images, as a new pre-training task for end-to-end autonomous driving. We developed ViDAR, a general model to pre-train visual BEV encoders, and designed a Latent Rendering operator to solve the ray-shaped feature issue. To conclude, our work demonstrates that visual point cloud forecasting enables scalable autonomous driving. + +Limitations and Future Work. Though with the potential of scalability, in this paper, we mainly conduct pre-training on Image-LiDAR sequences from nuScenes dataset, of which the data scale is still limited. As for the future, we plan to scale up the pre-training data of ViDAR, study visual point cloud forecasting across diverse datasets, and use publicly available Image-LiDAR sequences as much as possible to train a foundation visual autonomous driving model [38]. + +# References + +[1] Adil Kaan Akan and Fatma Güney. StretchBEV: Stretching Future Instance Prediction Spatially and Temporally. In ECCV, 2022. 7 +[2] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A Cookbook of Self-supervised Learning. arXiv preprint arXiv:2304.12210, 2023. 1 +[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In ICCV, 2019. 1 +[4] Alex Bewley, Pei Sun, Thomas Mensink, Dragomir Anguelov, and Cristian Sminchisescu. Range Conditioned Dilated Convolutions for Scale Invariant 3D Object Detection. arXiv preprint arXiv:2005.09927, 2021. 3 +[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi-ancarlo Baldan, and Oscar Beijbom. nuScenes: A Multi-modal Dataset for Autonomous Driving. In CVPR, 2020. 1, 2, 5 +[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In ECCV, 2022. 1 +[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end Autonomous Driving: Challenges and Frontiers. arXiv preprint arXiv:2306.16927, 2023. 2 +[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint arXiv:2003.04297, 2020. 1, 2 +[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5 +[10] OpenScene Contributors. OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving, 2023. 1, 5 +[11] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable Convolutional Networks. In ICCV, 2017. 5 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6, 7 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5 +[15] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles Qi, Yin Zhou, Zoey Yang, Aurélien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset. In ICCV, 2021. 1 +[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. I. J. Robotics Res., 2013. 1 +[17] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. ViP3D: End-to-end visual trajectory prediction via 3d agent queries. In CVPR, 2023. 1, 5, 7 +[18] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A Survey on Self-supervised Learning: Algorithms, Applications, and Future Trends. arXiv preprint arXiv:2301.05712, 2023. 1 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5, 6, 7 +[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In CVPR, 2020. 1, 2 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2022. 2 +[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts. In CVPR, 2021. 1 +[23] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: Future Instance Segmentation in Bird's-Eye view from Surround Monocular Cameras. In ICCV, 2021. 5, 7 +[24] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular Quasi-Dense 3D Object Tracking. TPAMI, 2022. 5, 7 +[25] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe Local Motion Planning with Self-Supervised Freespace Forecasting. In CVPR, 2021. 7 +[26] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning. In ECCV, 2022. 5, 7 +[27] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1, 2, 5, 6, 7, 8 +[28] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 5 + +[29] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction. In CVPR, 2023. 1 +[30] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. DriveAdapter: Breaking the Coupling Barrier of Perception and Planning in End-to-End Autonomous Driving. In ICCV, 2023. 1 +[31] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think Twice before Driving: Towards Scalable Decoders for End-to-End Autonomous Driving. In CVPR, 2023. 1 +[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 5, 7 +[33] Li Jiang, Zetong Yang, Shaoshuai Shi, Vladislav Golyanik, Dengxin Dai, and Bernt Schiele. Self-supervised Pretraining with Masked Shape Prediction for 3D Scene Understanding. In CVPR, 2023. 1 +[34] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. arXiv preprint arXiv:2004.11362, 2020. 2 +[35] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable Raycasting for Self-Supervised Occupancy Forecasting. In ECCV, 2022. 3, 7 +[36] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point Cloud Forecasting as a Proxy for 4D Occupancy Forecasting. In CVPR, 2023. 3, 6 +[37] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR, 2015. 6 +[38] Hongyang Li, Yang Li, Huijie Wang, Jia Zeng, Pinlong Cai, Huilin Xu, Dahua Lin, Junchi Yan, Feng Xu, Lu Xiong, Jingdong Wang, Futang Zhu, Kai Yan, Chunjing Xu, Tiancai Wang, Beipeng Mu, Shaoqing Ren, Zhihui Peng, and Yu Qiao. Open-sourced Data Ecosystem in Autonomous Driving: the Present and Future. arXiv preprint arXiv:2312.03408, 2023. 8 +[39] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, Hao Tian, Enze Xie, Jiangwei Xie, Li Chen, Tianyu Li, Yang Li, Yulu Gao, Xiaosong Jia, Si Liu, Jianping Shi, Dahua Lin, and Yu Qiao. Delving Into the Devils of Bird's-Eye-View Perception: A Review, Evaluation and Recipe. TPAMI, 2023. 2 +[40] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based Topology Reasoning for Driving Scenes. arXiv preprint arXiv:2304.05277, 2023. 5 +[41] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying Voxel-based Representation with Transformer for 3D Object Detection. In NeurIPS, 2022. 7 +[42] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. BEVStereo: Enhancing Depth Estimation in + +Multi-View 3D Object Detection with Temporal Stereo. In AAAI, 2023. 7 +[43] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object Detection. In AAAI, 2023. 5 +[44] Yanwei Li, Zhiding Yu, Jonah Philion, Anima Anandkumar, Sanja Fidler, Jiaya Jia, and Jose Alvarez. End-to-end 3D Tracking with Decoupled Queries. In ICCV, 2023. 7 +[45] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEV-Former: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8 +[46] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and José Manuel Álvarez. FB-BEV: BEV Representation from Forward-Backward View Transformations. ICCV, 2023. 2, 4 +[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. PnPNet: End-to-End Perception and Prediction With Tracking in the Loop. In CVPR, 2020. 5, 7 +[48] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7 +[49] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 5 +[50] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos. arXiv preprint arXiv:2308.09244, 2023. 1 +[51] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully Sparse 3D Panoptic Occupancy Prediction. arXiv preprint arXiv:2312.17118, 2023. 1 +[52] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position Embedding Transformation for Multi-View 3D Object Detection. In ECCV, 2022. 1, 3 +[53] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. In ICCV, 2023. 7 +[54] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In ICRA, 2023. 1 +[55] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. arXiv preprint arXiv:1711.05101, 2019. 6 +[56] B. Mersch, X. Chen, J. Behley, and C. Stachniss. Self-supervised Point Cloud Prediction Using 3D Spatiotemporal Convolutional Networks. In CoRL, 2021. 3 +[57] Gregory P. Meyer, Ankit Laddha, Eric Kee, Carlos Vallespi-Gonzalez, and Carl K. Wellington. LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving. In CVPR, 2019. 3 + +[58] Chen Min, Liang Xiao, Dawei Zhao, Yiming Nie, and Bin Dai. Occupancy-MAE: Self-Supervised Pre-Training LargeScale LiDAR Point Clouds With Masked Occupancy Autoencoders. TIV, 2023. 2 +[59] Ziqi Pang, Zhichao Li, and Naiyan Wang. SimpleTrack: Understanding and Rethinking 3D Multi-object Tracking. arXiv preprint arXiv:2111.09621, 2021. 5 +[60] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is Pseudo-Lidar Needed for Monocular 3D Object Detection? In ICCV, 2021. 2, 3 +[61] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical Depth DistributionNetwork for Monocular 3D Object Detection. In CVPR, 2021. 1 +[62] Ravid Shwartz-Ziv and Yann LeCun. To Compress or Not to Compress-Self-Supervised Learning and Information Theory: A Review. arXiv preprint arXiv:2304.09355, 2023. 1 +[63] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv preprint arXiv:2312.14150, 2023. 1 +[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 1 +[65] Izzeddin Teeti, Salman Khan, Ajmal Shahbaz, Andrew Bradley, and Fabio Cuzzolin. Vision-based Intention and Trajectory Prediction in Autonomous Vehicles: A Survey. In *IJCAI*, 2022. 1 +[66] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive Multiview Coding. In ECCV, 2020. 1, 2 +[67] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as Occupancy. In ICCV, 2023. 1, 2, 3, 5, 6, 7 +[68] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping. In NeurIPS Datasets and Benchmarks, 2023. 5 +[69] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In ICCV Workshops, 2021. 1, 2, 3, 6, 7 +[70] Tai Wang, Qing Lian, Chenming Zhu, Xinge Zhu, and Wenwei Zhang. MV-FCOS3D++: Multi-View Camera-Only 4D Object Detection with Pretrained Monocular Backbones. arXiv preprint arXiv:2207.12716, 2022. 1 +[71] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions. In CVPR, 2023. 6, 7 + +[72] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In CoRL, 2021. 7 +[73] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In CVPR, 2022. 2 +[74] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the Pose Forecasting Pipeline with SPF2: Sequential Pointcloud Forecasting for Sequential Pose Forecasting. arXiv preprint arXiv:2003.08376, 2020. 3 +[75] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2Net: Stochastic Sequential Pointcloud Forecasting. In ECCV, 2022. 3 +[76] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline. In NeurIPS, 2022. 1 +[77] Penghao Wu, Li Chen, Hongyang Li, Xiaosong Jia, Junchi Yan, and Yu Qiao. Policy Pre-training for Autonomous Driving via Self-supervised Geometric Modeling. In ICLR, 2023. 2 +[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In CVPR, 2018. 1 +[79] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In ECCV, 2020. 1 +[80] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2 +[81] Xiangchao Yan, Runjian Chen, Bo Zhang, Jiakang Yuan, Xinyu Cai, Botian Shi, Wenqi Shao, Junchi Yan, Ping Luo, and Yu Qiao. SPOT: Scalable 3D Pre-training via Occupancy Prediction for Autonomous Driving. arXiv preprint arXiv:2309.10527, 2023. 2 +[82] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, Xiaofei He, and Wanli Ouyang. UniPAD: A Universal Pre-training Paradigm for Autonomous Driving. arXiv preprint arXiv:2310.08370, 2023. 2, 3 +[83] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. IPOD: Intensive Point-based Object Detector for Point Cloud. CoRR, 2018. 5 +[84] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, 2019. +[85] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3DSSD: Point-based 3D Single Stage Object Detector, 2020. +[86] Zetong Yang, Yanan Sun, Shu Liu, Xiaojuan Qi, and Jiaya Jia. CN: Channel Normalization for Point Cloud Recognition. In ECCV, 2020. + +[87] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3D-MAN: 3D Multi-Frame Attention Network for Object Detection. In CVPR, 2021. 5 +[88] Zetong Yang, Li Jiang, Yanan Sun, Bernt Schiele, and Jiaya Jia. A Unified Query-based Paradigm for Point Cloud Understanding. In CVPR, 2022. 1 +[89] Zetong Yang, Zhiding Yu, Chris Choy, Renhao Wang, Anima Anandkumar, and Jose M. Alvarez. Improving Distant 3D Object Detection Using 2D Box Supervision. arXiv preprint arXiv:2403.09230, 2024. 1 +[90] Jia Zeng, Li Chen, Hanming Deng, Lewei Lu, Junchi Yan, Yu Qiao, and Hongyang Li. Distilling Focal Knowledge From Imperfect Expert for 3D Object Detection. In CVPR, 2023. 2, 4 +[91] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. MUTR3D: A Multi-camera Tracking Framework via 3D-to-2D Queries. In CVPR, 2022. 5, 7 +[92] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. BEVerse: Unified Perception and Prediction in Birds-Eye-View for Vision-Centric Autonomous Driving. arXiv preprint arXiv:2205.09743, 2022. 5, 7 +[93] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable Transformers for End-to-End Object Detection. In ICLR, 2020. 5 \ No newline at end of file diff --git a/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/images.zip b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b2e2b95bd364a12f79943f360f43b1ad9385345 --- /dev/null +++ b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52823277bb629c1c5167313ddd638b2848a903501f4fa7d3ffac4669b05d5b0e +size 499736 diff --git a/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/layout.json b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fa61b134346990deca773ba64cec27ba148094a4 --- /dev/null +++ b/2024/Visual Point Cloud Forecasting enables Scalable Autonomous Driving/layout.json @@ -0,0 +1,10680 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "text", + "content": "Visual Point Cloud Forecasting enables Scalable Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 143, + 428, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 143, + 428, + 159 + ], + "spans": [ + { + "bbox": [ + 164, + 143, + 428, + 159 + ], + "type": "text", + "content": "Zetong Yang Li Chen Yanan Sun Hongyang Li" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 163, + 386, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 163, + 386, + 177 + ], + "spans": [ + { + "bbox": [ + 206, + 163, + 386, + 177 + ], + "type": "text", + "content": "OpenDriveLab and Shanghai AI Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 182, + 411, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 182, + 411, + 194 + ], + "spans": [ + { + "bbox": [ + 186, + 182, + 411, + 194 + ], + "type": "text", + "content": "https://github.com/OpenDriveLab/ViDAR" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 204, + 548, + 385 + ], + "blocks": [ + { + "bbox": [ + 48, + 204, + 548, + 385 + ], + "lines": [ + { + "bbox": [ + 48, + 204, + 548, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 204, + 548, + 385 + ], + "type": "image", + "image_path": "72cb6304fddd876554924d65bab7ee67995c4611bf875095a360c61669bcec58.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 391, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 45, + 391, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 45, + 391, + 547, + 425 + ], + "type": "text", + "content": "Figure 1. ViDAR is a visual autonomous driving pre-training framework, which leverages the estimation of future point clouds from historical visual inputs as the pre-text task. We term this new pre-text task as visual point cloud forecasting. With the aid of ViDAR, we achieve substantial improvement spanning a diverse spectrum of downstream applications for perception, prediction, and planning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 438, + 192, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 438, + 192, + 451 + ], + "spans": [ + { + "bbox": [ + 143, + 438, + 192, + 451 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "text", + "content": "In contrast to extensive studies on general vision, pretraining for scalable visual autonomous driving remains seldom explored. Visual autonomous driving applications require features encompassing semantics, 3D geometry, and temporal information simultaneously for joint perception, prediction, and planning, posing dramatic challenges for pre-training. To resolve this, we bring up a new pre-training task termed as visual point cloud forecasting - predicting future point clouds from historical visual input. The key merit of this task captures the synergic learning of semantics, 3D structures, and temporal dynamics. Hence it shows superiority in various downstream tasks. To cope with this new problem, we present ViDAR, a general model to pretrain downstream visual encoders. It first extracts historical embeddings by the encoder. These representations are then transformed to 3D geometric space via a novel Latent Rendering operator for future point cloud prediction. Experiments show significant gain in downstream tasks, e.g., " + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "text", + "content": " NDS on 3D detection, " + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "text", + "content": " error reduction on motion forecasting, and " + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "inline_equation", + "content": "\\sim 15\\%" + }, + { + "bbox": [ + 45, + 464, + 290, + 706 + ], + "type": "text", + "content": " less collision rate on planning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 438, + 387, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 387, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 387, + 451 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "content": "Recently, the community has witnessed rapid development in visual, or camera-only autonomous driving, with input being monocular or multi-view images [50, 63, 69, 70, 76, 89]. Leveraging visual inputs, existing approaches demonstrate superior capability of extracting Bird's-Eye-View (BEV) features [6, 28, 45, 52, 61, 88], and performing well in perception [29, 51, 54, 67], prediction [17, 27, 65], and planning [30, 31]. Despite significant improvements in applications, these models rely on precise 3D annotations to a great extent, which are often difficult to collect, e.g., occupancy [3, 10], 3D boxes [5, 16, 64], trajectories [15], and thus are challenging to scale up for production." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": "Considering the expensive labeling workflow, pretraining [2, 18, 62] has emerged as a crucial approach to scale up downstream applications. The key idea is to define pretext tasks that leverage large amounts of readily available data to learn meaningful representations. This enhances downstream performance though labeled data is limited." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Though extensive research of pre-training in computer vision has been conducted [8, 20-22, 33, 66, 78, 79], its application in visual autonomous driving remains seldom" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14673" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "explored. Visual autonomous driving poses great challenges in pre-training as it requires the features to maintain semantics, 3D geometry, and temporal dynamics at the same time for joint perception, prediction, and planning [7, 77]. As a result, most models still rely on supervised pre-training, like 3D detection [60, 69] or occupancy [58, 67, 81], using labeled data that is often unavailable at scale [39]. Some approaches propose estimating depth [60] or rendering masked scenes [82] as pre-training. They use Image-LiDAR pairs as a means of achieving scalable annotation-free pre-training. However, they struggle in either multi-view 3D geometry or temporal modeling (Figure 2). Depth estimation retrieves depth from one image, limited in multi-view geometry; rendering techniques reconstruct scenes from multi-view images but lacking temporal modeling. However, temporal modeling is crucial in end-to-end autonomous driving systems, e.g., UniAD [27], especially for prediction and planning which are the ultimate goals and require accurate scene flow and object motion for decision-making. Due to the absence of temporal modeling, existing approaches are insufficient for pretraining the end-to-end system." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 338, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 434 + ], + "type": "text", + "content": "In this work, we explore pre-training tailored for end-to-end visual autonomous driving applications, including not only perception but also prediction and planning [7]. We formulate a new pre-text task, visual point cloud forecasting (Figure 2), to fully exploit information of semantics, 3D geometry, and temporal dynamics behind the raw Image-LiDAR sequences, with being scalable into consideration. It predicts future point clouds from historical visual images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 436, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 287, + 568 + ], + "type": "text", + "content": "The main rationale of visual point cloud forecasting lies in the simultaneous supervision of semantics, 3D structure, and temporal modeling. By compelling the model to predict the future from history, it supervises the extraction of scene flow and object motion which are crucial for temporal modeling and future estimation. Meanwhile, it involves the reconstruction of point clouds from images, which supervises the multi-view geometry and semantic modeling. Therefore, features from visual point cloud forecasting embed information of both geometric and temporal hints, beneficial for perception, tracking, and planning simultaneously." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 288, + 714 + ], + "type": "text", + "content": "To this end, we present ViDAR, a general visual point cloud forecasting approach for pre-training (Figure 2). ViDAR includes three parts, History Encoder, Latent Rendering operator, and Future Decoder. The History Encoder is the target structure for pre-training. It could be any visual BEV encoder [45] to embed visual sequences into BEV space. These BEV features are sent to the Latent Rendering operator. Latent Rendering plays a crucial role in enabling ViDAR benefit downstream performance. It solves the ray-shaped BEV features issue [46, 90], models 3D geometric latent space, and bridges encoder and decoder. The Future Decoder is an auto-regressive transformer that takes histor" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 425, + 118 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 425, + 118 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 425, + 118 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 425, + 118 + ], + "type": "image", + "image_path": "e2e1fa4cb319e04f3374b2a35ab5b3220e7182d12c0c67ba65e33fd367d70464.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 544, + 117 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 544, + 117 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 544, + 117 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 544, + 117 + ], + "type": "image", + "image_path": "babfffed185465b47f0547f3ecd338f53020387c1f00539a1c3e95ce6e60e2e5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 118, + 544, + 208 + ], + "blocks": [ + { + "bbox": [ + 308, + 118, + 544, + 208 + ], + "lines": [ + { + "bbox": [ + 308, + 118, + 544, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 544, + 208 + ], + "type": "image", + "image_path": "79a99e36f19e574617e2a210beb8cdbe7787512e7f28458008c205b34d4f5815.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 210, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 545, + 266 + ], + "type": "text", + "content": "Figure 2. Comparisons among visual autonomous driving pretraining paradigms and our ViDAR architecture. Compared to existing methods, visual point cloud forecasting jointly models multi-view geometry and temporal dynamics. We then propose ViDAR, using Image-LiDAR sequences to pre-train visual encoders." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 286, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 310 + ], + "type": "text", + "content": "ical BEV features to iteratively predict future point clouds for arbitrary timestamps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "content": "ViDAR provides a comprehensive solution for visual autonomous driving pre-training. We test ViDAR on nuScenes dataset [5] in terms of point cloud forecasting and downstream verifications. Though with visual inputs, ViDAR still outperforms previous forecasting methods using point clouds, " + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "inline_equation", + "content": "\\sim 33\\%" + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "content": " Chamfer Distance reduction on point cloud estimation of 1s future. ViDAR also improves downstream performance. Using Image-LiDAR sequences only, ViDAR surprisingly outperforms 3D detection pre-training [69], e.g., by " + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "content": " mIoU for detection and semantic occupancy prediction, under the same data scale. If also based on the 3D detection pre-training, ViDAR boosts previous methods by " + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 305, + 310, + 546, + 538 + ], + "type": "text", + "content": " mIoU. Further, due to the effective pre-training on the joint capture of geometric and temporal information, ViDAR improves UniAD [27] on all tasks for end-to-end autonomous driving including perception, prediction, and planning by a large margin (Figure 1). Experimental results validate that visual point cloud forecasting enables scalable autonomous driving." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 570, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 665 + ], + "type": "text", + "content": "Pre-training for Visual Autonomous Driving. Pretraining for scalable applications has been extensively studied in general vision. These approaches can be roughly divided into contrastive approaches [8, 20, 34, 66], which learn discriminative features from positive and negative pairs; and masked signal modeling [13, 21, 73, 80], which recover discarded signals from remained signals to capture a comprehensive understanding of global semantics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": "In contrast, pre-training for visual autonomous driving is still under-explored. Visual autonomous driving poses great challenges as it requires semantic understanding, 3D structural awareness, and temporal modeling at the same time," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14674" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "for joint perception, prediction, and planning. Existing vision methods mainly consider semantics; methods based on Image-LiDAR pairs [60, 82] struggle with temporal modeling; other supervised strategies [67, 69] are not scalable. In this work, we propose visual point cloud forecasting, which simultaneously models semantics, temporal dynamics, and 3D geometry by a uniform process, and is easily scaled up." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 373 + ], + "type": "text", + "content": "Point Cloud Forecasting. Point cloud forecasting, one of the most fundamental self-supervised tasks for autonomous driving, predicts future point clouds from past point cloud inputs. Previous works use range image [4, 57], a representation obtained by projecting point clouds to dense 2D images using sensor intrinsic and extrinsic parameters. Based on historical range images, they apply 3D convolutions [56], or LSTMs [74, 75] to predict future point clouds. Yet, they additionally model the motion of sensor intrinsic and extrinsic parameters. Later methods factor out the estimation of sensors by introducing 4D occupancy prediction [36] and differentiable ray-casting [35], which ensures a better modeling of the world. Compared to prior literature, we aim at visual point cloud forecasting, using past images to predict future point clouds. Meanwhile, we raise this task as a pre-training paradigm for visual autonomous driving and demonstrate its superiority in a wide range of downstream applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 381, + 129, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 129, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 129, + 395 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 401, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 462 + ], + "type": "text", + "content": "In this section, we elaborate on our ViDAR, a visual point cloud forecasting approach for general autonomous driving pre-training. We begin with an overview of ViDAR in Section 3.1, and subsequently delve into Latent Rendering and Future Decoder in Section 3.2 and Section 3.3, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 467, + 115, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 115, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 115, + 479 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": "As depicted in Figure 2, ViDAR comprises three components: (a) a History Encoder, also the target structure of pre-training, which extracts BEV embeddings " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": " from visual sequence inputs " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": "; It can be any visual BEV encoder [28, 45, 52]; (b) a Latent Rendering operator, which simulates the volume rendering operation in latent space so as to obtain geometric embedding " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": "; and (c) a Future Decoder, which predicts future BEV features " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": " at timestamps " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "t \\in \\{1, 2, \\ldots\\}" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": " in an auto-regressive manner. Finally, a prediction head is followed to project " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": " into 3D occupancy volume " + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 617 + ], + "type": "text", + "content": ". This process is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 622, + 168, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 622, + 168, + 634 + ], + "spans": [ + { + "bbox": [ + 75, + 622, + 168, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {b e v}} = \\operatorname {E n c o d e r} (\\mathcal {I}),", + "image_path": "b7a09ea69686301726d9d57346c55881855a0c68e544a56b93c40716beeade81.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 639, + 287, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 639, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 75, + 639, + 287, + 657 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {\\text {b e v}} = \\text {L a t e n t R e n d e r} (\\mathcal {F} _ {\\text {b e v}}), \\tag {1}", + "image_path": "264696af88184c99574752ad48062e148355aa087e8e7ede0aa9a7eca75f2de8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 654, + 260, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 654, + 260, + 667 + ], + "spans": [ + { + "bbox": [ + 75, + 654, + 260, + 667 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {t} = \\operatorname {D e c o d e r} \\left(\\hat {\\mathcal {F}} _ {t - 1}\\right), \\text {w h e r e} \\hat {\\mathcal {F}} _ {0} = \\hat {\\mathcal {F}} _ {\\text {b e v}},", + "image_path": "36dc518c4fb56a43f9fdcc7647cf34b16d5d43af0308b9004e43fc8c7cf2414f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 672, + 190, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 672, + 190, + 685 + ], + "spans": [ + { + "bbox": [ + 75, + 672, + 190, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {t} = \\text {P r o j e c t i o n} (\\hat {\\mathcal {F}} _ {t}).", + "image_path": "d6e980b98226495f65adcd7968e3472ed99ac9507aae2fc01add921b94e9ba3e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Point cloud predictions are obtained from the predicted occupancy volume " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ". This process is similar to the previous" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "point cloud forecasting method [36]. Specifically, we first cast rays from the origin to various designated directions, then figure out the distance of waypoints along each ray with the maximum occupancy response, and finally compute the point position according to the distance and corresponding ray direction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 152, + 411, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 411, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 411, + 165 + ], + "type": "text", + "content": "3.2. Latent Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 171, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 291 + ], + "type": "text", + "content": "A straightforward solution of visual point cloud forecasting for pre-training is to incorporate the History Encoder and Future Decoder directly with differentiable ray-casting [36], which is the crucial component in state-of-the-art point cloud forecasting methods to render point clouds from predicted occupancy volume and compute loss for backpropagation. However, our experimental results show that this approach does not yield improvements and even has a detrimental effect on the downstream tasks due to the defective geometric feature modeling ability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "type": "text", + "content": "Preliminary. Differentiable ray-casting is a volume rendering process operating on an occupancy volume, denoted as " + }, + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{L \\times H \\times W}" + }, + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "type": "text", + "content": ". It renders depths of various rays and subsequently converts the depths with corresponding ray directions to point clouds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": "Formally, starting from the origin sensor position, " + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": ", differentiable ray-casting casts " + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": " rays with varying directions, " + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": ". Along each ray " + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": ", it uniformly samples " + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": " waypoints at different distances " + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\lambda^{(j)} \\in \\mathbb{R}, j \\in \\{1,2,\\dots,m\\}" + }, + { + "bbox": [ + 305, + 366, + 545, + 438 + ], + "type": "text", + "content": " until reaching the boundary of the 3D space. The coordinates of these waypoints are calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 379, + 447, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 447, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 379, + 447, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {(i, j)} = \\mathbf {o} + \\lambda^ {(j)} \\mathbf {d} ^ {(i)}. \\tag {2}", + "image_path": "1f0bab4f181802e04bdd2a25e0f332accb500694775842a89fed3d9bf2429022.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": "Those waypoints coordinates, " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{n \\times m \\times 3}" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": ", are used to compute occupancy values. This process is quantized [36], wherein the waypoints are discretized to occupancy volume grids. Then, the values of waypoints are derived as the associated values of volume grids, " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{(i,j)} = \\mathcal{P}^{([\\mathbf{x}^{(i,j)}])}" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": " denotes a rounding operation for discretizing waypoints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "type": "text", + "content": "Differentiable ray-casting renders the corresponding depth of the " + }, + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "type": "text", + "content": "-th ray, " + }, + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}^{(i)}" + }, + { + "bbox": [ + 304, + 545, + 545, + 570 + ], + "type": "text", + "content": ", by an integral process:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 356, + 579, + 545, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 356, + 579, + 545, + 612 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {p}} ^ {(i, j)} = \\left[ \\prod_ {k = 1} ^ {j - 1} (1 - \\mathbf {p} ^ {(i, k)}) \\right] \\mathbf {p} ^ {(i, j)}, \\tag {3}", + "image_path": "318221625b4da73c09dfe21f7137aea2624196473438fd296e019ccff7e43801.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 381, + 616, + 545, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 616, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 381, + 616, + 545, + 648 + ], + "type": "interline_equation", + "content": "\\hat {\\lambda} ^ {(i)} = \\sum_ {j = 1} ^ {m} \\hat {\\mathbf {p}} ^ {(i, j)} \\lambda^ {(j)}. \\tag {4}", + "image_path": "0a9967f5b3c968f9ab5b5f3170c202760c427050faf1c36fa46d9a0c46f708ef.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": "For simplicity, we name Eq. 3 and Eq. 4 as conditional probability function and distance expectation function. The conditional probability function determines the occupancy of a grid by considering the conditional probability that prior waypoints are unoccupied and the ray terminates at" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14675" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 126, + 135 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 126, + 135 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 126, + 135 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 126, + 135 + ], + "type": "image", + "image_path": "e757736429c0ca6d831c31ef263a2fd76473ac87aedea29fa2a46a54a5fb3119.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 136, + 129, + 144 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 129, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 129, + 144 + ], + "type": "text", + "content": "(a.) Ray-shaped Features" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 134, + 70, + 200, + 135 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 200, + 135 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 200, + 135 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 200, + 135 + ], + "type": "image", + "image_path": "c2361424183af82592f556eb04723e204ee444e99696278d01300a7ca464b9e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 136, + 203, + 144 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 203, + 144 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 203, + 144 + ], + "type": "text", + "content": "(b.) Geometric Features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 147, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 147, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 287, + 201 + ], + "type": "text", + "content": "Figure 3. Ray-shaped Features vs. Geometric Features. Ray-shaped features show similar feature responses on BEV grids along the same ray; while geometric features from the Latent Rendering maintain discriminative 3D geometry and can describe the 3D world in latent space." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 210, + 72, + 273, + 134 + ], + "blocks": [ + { + "bbox": [ + 210, + 72, + 273, + 134 + ], + "lines": [ + { + "bbox": [ + 210, + 72, + 273, + 134 + ], + "spans": [ + { + "bbox": [ + 210, + 72, + 273, + 134 + ], + "type": "image", + "image_path": "4d5103f45b2db68050a79dbcadb34ca08671ee1f034d235325136db0a20ba0d5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 135, + 270, + 144 + ], + "lines": [ + { + "bbox": [ + 213, + 135, + 270, + 144 + ], + "spans": [ + { + "bbox": [ + 213, + 135, + 270, + 144 + ], + "type": "text", + "content": "(c.) GT Point Cloud" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 51, + 213, + 282, + 253 + ], + "blocks": [ + { + "bbox": [ + 51, + 213, + 282, + 253 + ], + "lines": [ + { + "bbox": [ + 51, + 213, + 282, + 253 + ], + "spans": [ + { + "bbox": [ + 51, + 213, + 282, + 253 + ], + "type": "table", + "html": "
Forecasting StructureN/ADifferentiable Ray-castingLatent Rendering
NDS (%)44.1140.20 (-3.91)47.58 (+3.47)
", + "image_path": "9dd37b522cbb4b38e23a6f3c92514bccb1b7f9feef93cd24fa62fcbfcbd68e42.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 255, + 287, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 321 + ], + "type": "text", + "content": "Table 1. Downstream detection performance under different forecasting structures. \"N/A\" represents the baseline without forecasting pre-training. We observe a performance drop when directly using History Encoder and Future Decoder with differentiable ray-casting for pre-training; while, with the Latent Rendering operator, the performance is significantly improved." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "type": "text", + "content": "this particular grid; the distance expectation function retrieves depths from the occupancy of grids in 3D volume. L1 loss is then applied to supervise the rendered depth for training point cloud forecasting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 390, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 545 + ], + "type": "text", + "content": "Despite the great success of differentiable ray-casting in the task of point cloud forecasting, its application in visual point cloud forecasting pre-training does not bring any benefit for downstream performance (Table 1). After such pretraining, ray-shaped features [46, 90], where grids along the same ray tend to possess similar features (Figure 3 - (a.)), are observed. The underlying reason is that waypoints along the same ray in 3D space usually correspond to the same pixel in the visual image, resulting in a tendency to learn similar feature responses. As a consequence, these ray-shaped features are not discriminative and representative enough when transferred to downstream applications, leading to reduced performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 547, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 631 + ], + "type": "text", + "content": "Latent Rendering. In order to extract more discriminative and representative features, we introduce the Latent Rendering operator. It first computes the ray-wise feature through a feature expectation function, then customizes features of each grid by weighting the ray-wise feature with its associated conditional probability. The overall structure is depicted in Figure 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 631, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 287, + 654 + ], + "type": "text", + "content": "To be specific, inspired by Eq. (4), the feature expectation function is formulated in a similar form:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 662, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 662, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 287, + 693 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} ^ {(i)} = \\sum_ {k = 1} ^ {m} \\hat {\\mathbf {p}} ^ {(i, k)} \\mathcal {F} _ {\\mathrm {b e v}} ^ {(k)}, \\tag {5}", + "image_path": "f5ba0dabed0e99e2ee48c4290ddc10954faf51e39febe844783b8795ac4fa5e3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " represents the ray extending from the origin to the" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 75, + 545, + 151 + ], + "blocks": [ + { + "bbox": [ + 308, + 75, + 545, + 151 + ], + "lines": [ + { + "bbox": [ + 308, + 75, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 75, + 545, + 151 + ], + "type": "image", + "image_path": "fe47be17e5e3c2f68533ab0d6f8a62b1781e37336e1b920caf564102b16c74ec.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 154, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 304, + 154, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 545, + 210 + ], + "type": "text", + "content": "Figure 4. Multi-group Latent Rendering comprises several Latent Rendering running in parallel for different channels. Latent Rendering captures geometric features by the conditional probability function and the feature extraction function. “" + }, + { + "bbox": [ + 304, + 154, + 545, + 210 + ], + "type": "inline_equation", + "content": "\\bigoplus" + }, + { + "bbox": [ + 304, + 154, + 545, + 210 + ], + "type": "text", + "content": "” means concatenating multi-group features among channel dimensions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "text", + "content": "-th BEV grid. Here, " + }, + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "text", + "content": " is the conditional probability, computed through the conditional probability function Eq. (3), which takes as input the learnable independent probability projected from " + }, + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{bev}" + }, + { + "bbox": [ + 304, + 223, + 545, + 283 + ], + "type": "text", + "content": ". The ray-wise features are shared by all grids lying in the same ray." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 284, + 471, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 471, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 471, + 296 + ], + "type": "text", + "content": "Then, we compute the grid feature as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 396, + 304, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 304, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 396, + 304, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {\\mathrm {b e v}} = \\hat {\\mathbf {p}} \\cdot \\hat {\\mathcal {F}}, \\tag {6}", + "image_path": "69a2a6e901834f86237ea1a557b9796d6a4266b78cc1b02c5545dfa458e68a71.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 327, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 375 + ], + "type": "text", + "content": "which highlights the response of BEV grids with higher conditional probability so as to make " + }, + { + "bbox": [ + 304, + 327, + 545, + 375 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{bev}" + }, + { + "bbox": [ + 304, + 327, + 545, + 375 + ], + "type": "text", + "content": " discriminative. This enables the BEV encoder to learn the geometric features during pre-training (Figure 3 - (b))." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 376, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 434 + ], + "type": "text", + "content": "To enhance the diversity of geometric features, we further design the multi-group Latent Rendering. By parallelizing multiple Latent Rendering on different feature channels, we allow ray-wise features to maintain diverse information, leading to better downstream performance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 435, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 590 + ], + "type": "text", + "content": "As described in Eq. (3), the conditional probability of each BEV grid is determined not only by its own independent response, but also by the response of all its prior grids. Consequently, in the pre-training phase, once the model raises the response of a particular BEV grid, the corresponding responses of all its prior and subsequent grids are suppressed, which mitigates the issue of ray-shaped features during pre-training. After the pre-training with Latent Rendering, it is generally observed that there are only a few peaks with higher responses on a specific ray, indicating the presence of objects or structures in the scene. This effectively promotes a more accurate and consistent understanding of the 3D environment." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 598, + 402, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 402, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 402, + 609 + ], + "type": "text", + "content": "3.3. Future Decoder" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": "The Future Decoder predicts the next BEV features " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": " of frame " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": " based on the inputs of previous BEV latent space " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t - 1}" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": " and the expected next ego-motion, " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": ". The predicted features are then used to generate point clouds as Eq. (1)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Architecture. As depicted in Figure 5, Future Decoder is a transformer that can be iteratively used to predict future BEV features from the last roll-out embeddings in an auto-regressive manner. In the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-th iteration, it first encodes" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14676" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 198 + ], + "type": "image", + "image_path": "9c4188f8cdd74035c8cb8c0d03b7f996d3e73b5526d9d6d9b3dd0a1e6c9b7765.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "type": "text", + "content": "Figure 5. Future Decoder iteratively predicts the next BEV features, " + }, + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "type": "text", + "content": ", from the conditions of ego-motion " + }, + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t" + }, + { + "bbox": [ + 46, + 199, + 287, + 233 + ], + "type": "text", + "content": " and the last BEV features, to enable specific future predictions with any ego-control." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "text", + "content": "the ego-motion condition " + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t" + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "text", + "content": ", which describes the expected coordinates and heading of ego-vehicle in the next frame, into high-dimensional embeddings by multi-layer perceptron (MLP), which are then added to future BEV queries as inputs of the transformer. Then, 6 transformer layers, composed of a Deformable Self-Attention [93], a Temporal Cross-Attention and a FeedForward Network [14], are used to predict the future " + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "text", + "content": " based on the condition and the last BEV features " + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t - 1}" + }, + { + "bbox": [ + 46, + 248, + 287, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 288, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 500 + ], + "type": "text", + "content": "The Temporal Cross-Attention layer follows the design of Deformable Cross-Attention [93]. The difference lies in the reference coordinates of query points. In the context of Deformable Cross-Attention [93], \"reference coordinates\" refer to the corresponding positions of query points on the feature maps of keys and values. Typically, they are consistent. However, as for Future Decoder, due to the moving of the ego-vehicle, the ego-coordinate systems between the last and target frame are not necessarily aligned. Therefore, we additionally compute the reference coordinates of future BEV queries in previous BEV feature maps, according to the ego-motion condition, to align coordinate systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "content": "After obtaining the next BEV features " + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t" + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "content": ", we use a projection layer to generate the occupancy volume " + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "type": "text", + "content": "Loss. Instead of using L1 loss to supervise the depths of various rays, we directly apply ray-wise cross-entropy loss to maximize the response of points along its corresponding ray, as we have already obtained geometric features after the Latent Rendering operator. To be specific, for each ground-truth point of the " + }, + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "type": "text", + "content": "-th future point clouds, we cast a ray from the origin position " + }, + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 526, + 288, + 670 + ], + "type": "text", + "content": " (the sensor position) towards the point, uniformly sample some waypoints along the ray until out of the volume, and compute cross-entropy loss for the ray to maximize the response of the point position and minimize the response of other waypoint positions. This process is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\frac {1}{T n} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\log \\left(\\frac {e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {g} ^ {(i)})}}}{\\sum_ {j} e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {x} ^ {(i , j)})}} + e ^ {\\mathcal {P} _ {t} ^ {(\\mathbf {g} ^ {(i)})}}}\\right), \\tag {7}", + "image_path": "b9ef355cca42a1a602e899fa2968f246b6e4aeae86ad774d7133c89a89d28e6e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": " indicate the number of future supervisions and the number of points in the " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": "-th ground-truth point clouds. " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{(i)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(i,j)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": " are the coordinates of " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": "-th ground-truth point and " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": "-th waypoints along the same ray. " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t^{(\\cdot)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": " is trilinear interpolation to obtain corresponding values from volume " + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 305, + 72, + 545, + 134 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 146, + 388, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 146, + 388, + 160 + ], + "spans": [ + { + "bbox": [ + 306, + 146, + 388, + 160 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 167, + 504, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 504, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 504, + 178 + ], + "type": "text", + "content": "This section investigates the following questions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 181, + 545, + 275 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 306, + 181, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 181, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 181, + 545, + 205 + ], + "type": "text", + "content": "- Can future point clouds be estimated from visual history, and how about ViDAR compared to point cloud methods?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 209, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 545, + 232 + ], + "type": "text", + "content": "- Can ViDAR help perception, prediction, and planning at the same time so as for scalable autonomous driving?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 236, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 545, + 259 + ], + "type": "text", + "content": "- Can ViDAR reduce the reliance of downstream applications on precise human annotations?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 263, + 524, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 524, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 524, + 275 + ], + "type": "text", + "content": "- How do different modules affect final performance?" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 285, + 356, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 356, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 356, + 298 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 304, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 424 + ], + "type": "text", + "content": "Dataset. We conduct experiments on the challenging nuScenes [5] dataset, which is a large-scale dataset with 1,000 autonomous driving sequences. This dataset is widely utilized in perception tasks including 3D object detection [28, 43-45, 83-87], multi-object tracking [24, 59, 91], and semantic occupancy prediction [10, 67]. It has also become a popular benchmark for subsequent research on end-to-end autonomous driving, including map segmentation [40, 68], trajectory predictions [17, 47], future occupancy prediction [23, 92], and open-loop planning [26, 27, 32]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "Implementation Details. We base our implementation on mmDet3D codebase [9] and conduct downstream verifications on BEVFormer for 3D detection, OccNet [67] for semantic occupancy prediction, and UniAD [27] for unified perception, prediction, and planning. We choose those downstream baselines due to their effectiveness on a wide range of tasks and sharing the same BEV encoder structure, BEVFormer encoder [45]. Without any specifications, the default historical encoder of ViDAR is the BEVFormer-base encoder, consisting of a ResNet101-DCN [11, 19] backbone with an FPN neck [49] and additional 6 encoder layers to extract BEV features from multi-view image sequences, which is consistent with downstream models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "To render geometric features, we use a 16-group Latent Rendering. Each group is responsible for rendering latent spaces of 16 channels given the features of 256 channels after the BEVFormer-base encoder. The Future Decoder is a 6-layer structure with a channel of 256 for each. The future BEV queries are " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " learnable tokens indicating a valid perception range of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "[-51.2\\mathrm{m}, 51.2\\mathrm{m}]" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " axis. We then use a projection layer with output channels as 16 to transform the predicted future BEV features to occupancy volume prediction " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{200 \\times 200 \\times 16}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", where 16 indicates the height dimension with a range of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "[-5\\mathrm{m}, 3\\mathrm{m}]" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14677" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 284, + 143 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 284, + 143 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 284, + 143 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 284, + 143 + ], + "type": "table", + "html": "
History HorizonMethodModalityChamfer Distance (m2) ↓
0.5s1.0s1.5s2.0s2.5s3.0s
1s4D-Occ [36]L1.261.88----
ViDARC1.111.251.401.571.761.97
3s4D-Occ [36]L0.911.131.301.531.722.11
ViDARC1.011.121.251.381.541.73
", + "image_path": "789322680444908f439ee74e78a3a8b1069d3c7a4ec3e1aea8f5f87d76dfcf85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 308 + ], + "type": "text", + "content": "During pre-training, we use 5 frames of historical multiview images and iterate the Future Decoder 6 times to predict point clouds for the future 3 seconds (each frame has 0.5 second interval). In each training step, we randomly select 1 future prediction for computing loss and detach the gradients of the others to save GPU memory. We pre-train the system for 50 epochs by AdamW optimizer [37, 55] with an initial learning of 2e-4 adjusted by cosine annealing strategy. For fine-tuning, we follow the same training strategy as the officially released downstream models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 132, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 132, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 132, + 330 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 286, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 286, + 432 + ], + "type": "text", + "content": "We now demonstrate the effectiveness of ViDAR across different tasks. First, we test the ability of ViDAR as a point cloud forecasting framework and compare it with the state-of-the-art approach that uses LiDAR inputs. Then, we show its advancement as a visual autonomous driving pre-training solution. We report the downstream comparison results in the order of perception-prediction-planning with previous state-of-the-art models on the nuScenes validation dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 286, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 286, + 470 + ], + "type": "text", + "content": "Downstream Settings. For downstream verifications, we test ViDAR to pre-train BEV encoders under different initialization settings, listed as the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 472, + 286, + 586 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 472, + 286, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 286, + 507 + ], + "type": "text", + "content": "- ViDAR-cls: The BEV encoders are initialized with the backbone pre-trained for ImageNet classification [12], followed by ViDAR pre-training on nuScenes dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 510, + 286, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 286, + 546 + ], + "type": "text", + "content": "- ViDAR-2D-det: The BEV encoder backbones are first pre-trained for 2D object detection on COCO dataset [48] before ViDAR pre-training on nuScenes dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "text", + "content": "- ViDAR-3D-det: The BEV encoder backbones are pretrained first by 3D detection on the nuScenes dataset, using FCOS3D [69], before ViDAR pre-training." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 589, + 286, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 286, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 286, + 637 + ], + "type": "text", + "content": "For UniAD experiments, after ViDAR pre-training, we first fine-tune BEVFormer for 3D detection, which is then used as the initialization for subsequent two-stage fine-tuning, consistent with the UniAD official implementation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "Point Cloud Forecasting. In Table 2, we present comparisons between our ViDAR and the previous state-of-the-art point cloud forecasting method, 4D-Occ [36]. The evaluation metric is Chamfer Distance. We evaluate both methods using input time horizons of 1s and 3s, corresponding" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 543, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 145, + 286, + 168 + ], + "lines": [ + { + "bbox": [ + 47, + 145, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 286, + 168 + ], + "type": "text", + "content": "Table 2. Point cloud forecasting. ViDAR surpasses prior state-of-the-art method on future prediction, using visual input only." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 71, + 543, + 201 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 543, + 201 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 543, + 201 + ], + "type": "table", + "html": "
MethodsEncoderPre-trainmAP (%)NDS (%)
BEV-Former [45]RN50 [19]ImageNet-cls [12]25.235.4
ViDAR-cls29.038.8
RN101 [19]ImageNet-cls [12]37.747.7
ViDAR-cls42.651.8
nus-3D-det [69]41.551.7
ViDAR-3D-det45.854.8
Intern-S [71]COCO-2D-det [48]41.551.2
ViDAR-2D-det47.656.4
Intern-B [71]COCO-2D-det [48]42.952.0
ViDAR-2D-det50.157.6
", + "image_path": "498e3b4936e3b5979df92380e1557bad67c17666f2865182fd8680f14574fef5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 205, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 306, + 205, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 545, + 227 + ], + "type": "text", + "content": "Table 3. Detection performance of BEVFormer with and without ViDAR pre-training under various backbones and initializations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 249, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 320 + ], + "type": "text", + "content": "to input sequences of 2 frames and 6 frames, respectively, following the same setup as in 4D-Occ. To provide detailed comparisons of performance, we report the quantitative forecasting results for each future timestamp. Only points within the range of " + }, + { + "bbox": [ + 304, + 249, + 545, + 320 + ], + "type": "inline_equation", + "content": "[-51.2\\mathrm{m}, 51.2\\mathrm{m}]" + }, + { + "bbox": [ + 304, + 249, + 545, + 320 + ], + "type": "text", + "content": " on the X- and Y-axis are considered during evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "content": "As presented in Table 2, ViDAR consistently outperforms 4D-Occ on both 1s and 3s settings, despite utilizing visual inputs exclusively. Specifically, with 1s history input, ViDAR achieves remarkable improvement over 4D-Occ, reducing forecasting errors by " + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\sim 33\\%" + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "content": " for future 1s predictions. When using 3s inputs, we observe a " + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\sim 18\\%" + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "content": " error reduction for 3s forecasting. Moreover, due to the autoregressive design of our Future Decoder, ViDAR effectively predicts arbitrary future, though with the constraint of a limited 1s input horizon. These experiments demonstrate the effectiveness of ViDAR for point cloud forecasting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "text", + "content": "Perception. We verify ViDAR on four downstream perception tasks, 3D object detection, semantic occupancy prediction, map segmentation, and multi-object tracking. In Table 3 and Table 4, we compare the performance of BEVFormer [45] and OccNet [67] with and without ViDAR pre-training under different backbones and initialization settings for 3D detection and semantic occupancy prediction, respectively. Notably, ViDAR, using solely Image-LiDAR sequences, outperforms 3D detection supervised pre-training (The 4th & the 5th row in Table 3, " + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "inline_equation", + "content": "42.6\\%" + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "text", + "content": " mAP to " + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "inline_equation", + "content": "41.5\\%" + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "text", + "content": " mAP; The 2nd & the 3rd row in Table 4, " + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "inline_equation", + "content": "29.57\\%" + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "text", + "content": " mIoU to " + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "inline_equation", + "content": "26.98\\%" + }, + { + "bbox": [ + 304, + 469, + 545, + 661 + ], + "type": "text", + "content": " mIoU). Furthermore, we also observe huge improvements in map segmentation (Table 5) and multi-object tracking (Table 6). These experiments demonstrate the effectiveness of ViDAR as a scalable pretraining method for enhancing 3D geometry modeling." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Prediction. The motion forecasting comparisons are presented in Table 7. As shown, ViDAR significantly improves the performance of UniAD [27]. For instance, we observe a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14678" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 282, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 282, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 282, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 282, + 178 + ], + "type": "table", + "html": "
MethodsEncoderPre-trainmIoU (%) ↑
OccNet [67]RN101 [19]ImageNet-cls [12]24.35
ViDAR-cls29.57
nus-3D-det [69]26.98
ViDAR-3D-det31.67
Intern-S [71]COCO-2D-det [48]24.92
ViDAR-2D-det30.51
Intern-B [71]COCO-2D-det [48]25.24
ViDAR-2D-det31.69
", + "image_path": "84e59e3ae74b903d4635f15ff9ddb9ea57ec0bdffb73ccd0f3576f385e2e7813.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 208, + 288, + 261 + ], + "blocks": [ + { + "bbox": [ + 47, + 181, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 181, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 286, + 202 + ], + "type": "text", + "content": "Table 4. Semantic Occupancy Prediction. ViDAR consistently improves OccNet on different backbones and initializations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 208, + 288, + 261 + ], + "lines": [ + { + "bbox": [ + 50, + 208, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 50, + 208, + 288, + 261 + ], + "type": "table", + "html": "
MethodsEncoderPre-trainLanes (%) ↑
BEVFormer [45]RN101nus-3D-det [69]23.9
UniAD [27]RN101nus-3D-det [45]31.3
ViDAR-3D-det33.2
", + "image_path": "015b56f8320e7e1b24140a135239db5863aedc91e86bd0eb836bfff060ca2bb0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 291, + 288, + 402 + ], + "blocks": [ + { + "bbox": [ + 47, + 263, + 286, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 263, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 286, + 285 + ], + "type": "text", + "content": "Table 5. Map segmentation. ViDAR improves UniAD on online mapping. The metric is segmentation IoU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 291, + 288, + 402 + ], + "lines": [ + { + "bbox": [ + 50, + 291, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 50, + 291, + 288, + 402 + ], + "type": "table", + "html": "
MethodsEncoderPre-trainAMOTA (%) ↑
ViP3D [17]RN50nus-3D-det [72]21.7
QD3DT [24]RN101-24.2
MUTR3D [91]RN101ImageNet-cls29.4
DQTrack-DETR3D [44]RN101nus-3D-det [72]36.7
DQTrack-UVTR [44]RN101nus-3D-det [41]39.6
DQTrack-Stereo [44]RN101nus-3D-det [42]40.7
DQTrack-PETr v2 [44]V2-99nus-3D-det [53]44.6
UniAD-Stage1 [27]RN101nus-3D-det [45]39.0
ViDAR-3D-det45.1
", + "image_path": "db1916240b212beff0524175e72b2acc9199d397da909f3d64d4d3dbb5ddca94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "content": " error reduction in minADE and a " + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "content": " improvement in EPA. In Table 8, we provide the comparison between UniAD with and without ViDAR pre-training in future occupancy prediction. The results demonstrate that ViDAR enhances the performance of UniAD for all areas. We observe improvements of " + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "content": " IoU and " + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "content": " VPQ for nearby areas, as well as improvements of " + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "content": " IoU and " + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 604 + ], + "type": "text", + "content": " VPQ for distant areas. With ViDAR pre-training, UniAD overcomes its limitation in occupancy forecasting for distant objects and now outperforms BEVerse [92] in all areas. These experiments highlight the effectiveness of ViDAR in enhancing downstream models in utilizing temporal information and improving their prediction performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Planning. Due to the effective temporal modeling and advanced future prediction capabilities, ViDAR significantly improves UniAD by reducing its average collision rate within 3 seconds by " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\sim 15\\%" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": ". Moreover, it achieves a substantial decrease in the average planning displacement error by " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.21\\mathrm{m}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and enables UniAD to outperform the state-of-the-art method, VAD [32], on nuScenes open-loop planning evaluation. These improvements demonstrate the effective" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 548, + 133 + ], + "blocks": [ + { + "bbox": [ + 47, + 403, + 286, + 426 + ], + "lines": [ + { + "bbox": [ + 47, + 403, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 286, + 426 + ], + "type": "text", + "content": "Table 6. Multi-object tracking. With ViDAR, UniAD outperforms previous end-to-end trackers using solely visual images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 548, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 548, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 548, + 133 + ], + "type": "table", + "html": "
MethodsminADE (m) ↓minFDE (m) ↓MR ↓EPA ↑
PnPNet [47]1.151.950.2260.222
ViP3D [17]2.052.840.2460.226
UniAD [27]0.751.080.1580.463
ViDAR0.670.990.1490.498
", + "image_path": "953a9ff399d25d15f70ffefefa16e873bdd9dffed9bdef20cda49caba8627d81.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 176, + 541, + 258 + ], + "blocks": [ + { + "bbox": [ + 305, + 136, + 545, + 169 + ], + "lines": [ + { + "bbox": [ + 305, + 136, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 136, + 545, + 169 + ], + "type": "text", + "content": "Table 7. Motion forecasting. ViDAR effectively enhances temporal modeling, which in turn boosts UniAD in future motion forecasting, showing consistent improvements on different metrics." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 176, + 541, + 258 + ], + "lines": [ + { + "bbox": [ + 309, + 176, + 541, + 258 + ], + "spans": [ + { + "bbox": [ + 309, + 176, + 541, + 258 + ], + "type": "table", + "html": "
MethodsVPQ-n. ↑VPQ-f. ↑IoU-n. ↑IoU-f. ↑
Fiery [23]50.229.959.436.7
StretchBEV [1]46.029.055.537.1
ST-P3 [26]-32.1-38.9
BEVerse [92]54.336.161.440.9
UniAD [27]54.633.962.840.1
ViDAR57.336.465.442.1
", + "image_path": "a7f8408c33c2a5786a580c967a309b38f9369f938f1ad4c99ae60f661cf51e45.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 300, + 552, + 383 + ], + "blocks": [ + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": "Table 8. Future occupancy prediction. ViDAR improves UniAD on future occupancy prediction for objects in both near (noted as \"n\", " + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "inline_equation", + "content": "30\\mathrm{x}30\\mathrm{m}" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": ") and far (noted as \"f\", " + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "inline_equation", + "content": "50\\mathrm{x}50\\mathrm{m}" + }, + { + "bbox": [ + 305, + 260, + 545, + 293 + ], + "type": "text", + "content": ") evaluation areas." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 300, + 552, + 383 + ], + "lines": [ + { + "bbox": [ + 309, + 300, + 552, + 383 + ], + "spans": [ + { + "bbox": [ + 309, + 300, + 552, + 383 + ], + "type": "table", + "html": "
MethodsModalityAvg.Col. (3s) (%) ↓Avg.L2 (3s) (m) ↓
FF [25]L0.431.43
EO [35]L0.331.60
ST-P3 [26]C0.712.11
VAD [32]C0.411.05
UniAD [27]C0.271.12
ViDARC0.230.91
", + "image_path": "47ae7d0d72a05191bd6ed247c62b2988586dacd798dbc0b81a0242b48e270019.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 385, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 451 + ], + "type": "text", + "content": "Table 9. Planning. ViDAR improves UniAD in terms of both collision avoidance and planning accuracy. Note that, the reported numbers are obtained by averaging the results of each timestamp in the future 3 seconds, which is consistent with the reported results of VAD [32] and ST-P3 [26] at the 3s timestamp instead of the averaged one. Please refer to GitHub:Issue for more details." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 475, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 536 + ], + "type": "text", + "content": "ness of ViDAR as a valuable pre-training approach for end-to-end autonomous driving. The enhanced performance in collision avoidance and planning accuracy highlights the potential of ViDAR in enhancing the safety and efficiency of downstream autonomous driving applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Joint Perception-Prediction-Planning. Finally, we summarize the improvements of ViDAR on the state-of-the-art end-to-end visual autonomous driving system, UniAD [27], for joint perception, prediction, and planning. As depicted in Table 10, ViDAR brings substantial improvements on all sub-modules of UniAD for perception (Detection, Tracking, Mapping), prediction (Motion Forecasting and Future Occupancy Prediction), and planning at the same time. These consistent improvements illustrate that visual point cloud forecasting effectively exploits the information of semantics, 3D geometry, and temporal dynamics behind the easily obtainable Image-LiDAR sequences. This, consequently, enables scalable visual autonomous driving." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14679" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 545, + 110 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 545, + 110 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 545, + 110 + ], + "type": "table", + "html": "
MethodDetectionTrackingMappingMotion ForecastingFuture Occupancy PredictionPlanning
NDS ↑mAP ↑AMOTA↑AMOTP↓IDS↓IoU-lane↑IoU-road↑minADE↓minFDE↓MR↓IoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑avg.L2↓avg.Col.↓
UniAD49.3637.9638.31.32105431.369.10.751.080.15862.840.154.633.91.120.27
ViDAR52.5742.3342.01.2599133.271.40.670.990.14965.442.157.336.40.910.23
", + "image_path": "7cccab671264fc88160b387af3803080fd945ae546ab56216abbab3dc1f75c17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 143, + 282, + 269 + ], + "blocks": [ + { + "bbox": [ + 53, + 143, + 282, + 269 + ], + "lines": [ + { + "bbox": [ + 53, + 143, + 282, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 282, + 269 + ], + "type": "image", + "image_path": "77b0bc98d2b4f373136843570bc9cd8b65772513a460b3e37f860e3127d8deaf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 288, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 326 + ], + "type": "text", + "content": "Figure 6. Validation of ViDAR on Fine-tuning with limited supervised data. We verify ViDAR on supervision efficiency by reducing available annotations for 3D object detection during downstream fine-tuning (from the full training set to a 1/8th subset) and observe a continuous improvement on each subset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 340, + 139, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 139, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 139, + 353 + ], + "type": "text", + "content": "4.3. Ablative Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 358, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 418 + ], + "type": "text", + "content": "We conduct further analysis of ViDAR on improving the performance of downstream models with limited supervised data, and the effect of the Latent Rendering operation on learning 3D geometric latent space. More ablation studies can be found in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 517 + ], + "type": "text", + "content": "Efficiency of Supervised Pre-training. The primary objective of pre-training is to minimize the dependence on precise 3D annotations. In Figure 6, we demonstrate the effectiveness of ViDAR in reducing the reliance of modern 3D detectors on accurate 3D box annotations. We fine-tune BEVFormer-base using partial 3D annotations on nuScenes, ranging from the full dataset to a " + }, + { + "bbox": [ + 46, + 434, + 287, + 517 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 434, + 287, + 517 + ], + "type": "text", + "content": " subset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": "As depicted in Figure 6, ViDAR exhibits a remarkable reduction in the dependence on 3D annotations. Notably, BEVFormer, pre-trained by ViDAR, surpasses its counterpart under full supervision by " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": " mAP, while using only half of the supervised samples, i.e., " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "39.4\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": " mAP vs. " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "37.7\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": " mAP. Consequently, through ViDAR, we can reduce half of 3D annotations without sacrificing precision. Additionally, we observe a consistent trend of increasing improvements as the available supervision decreases. For instance, the mAP improvements are " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 673 + ], + "type": "text", + "content": " when fine-tuned on the full, half, a quarter, and 1/8th subsets. These results highlight the potential of ViDAR in harnessing large amounts of Image-LiDAR sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Effect of Latent Rendering operator. Latent Rendering is the key component of ViDAR, which enables visual point" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 144, + 547, + 185 + ], + "blocks": [ + { + "bbox": [ + 47, + 114, + 546, + 137 + ], + "lines": [ + { + "bbox": [ + 47, + 114, + 546, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 546, + 137 + ], + "type": "text", + "content": "Table 10. Performance gain of ViDAR for joint perception, prediction, and planning. ViDAR consistently improves UniAD [27] on all tasks towards end-to-end autonomous driving, validating its effectiveness for scalable visual autonomous driving." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 144, + 547, + 185 + ], + "lines": [ + { + "bbox": [ + 308, + 144, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 144, + 547, + 185 + ], + "type": "table", + "html": "
Groups of Latent R.N/A124816
NDS (%)40.2039.1843.3645.5347.0147.58
", + "image_path": "61d3b0a1e3fe879e06fe35c871ff337a4f9c3d9a184b6e1643de95f70fc114dc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 187, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 187, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 546, + 232 + ], + "type": "text", + "content": "Table 11. Ablation of Latent Rendering for downstream finetuning. We compare the performance of 3D detection pre-trained by ViDAR without the Latent Rendering operation (denoted as \"N/A\") and by ViDAR with different groups of Latent Rendering." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 249, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 345 + ], + "type": "text", + "content": "cloud forecasting to effectively contribute to downstream applications. It addresses the ray-shaped features issue encountered during pre-training. In Table 11, we verify its effectiveness by comparing the performance of downstream models pre-trained by ViDAR with the Latent Rendering or not. The downstream model is BEVFormer-small [45] for 3D object detection. For context, the baseline performance with ImageNet-cls pre-training is " + }, + { + "bbox": [ + 304, + 249, + 545, + 345 + ], + "type": "inline_equation", + "content": "44.11\\%" + }, + { + "bbox": [ + 304, + 249, + 545, + 345 + ], + "type": "text", + "content": " NDS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": "As depicted in Table 11, when the Latent Rendering is missing (denoted as \"N/A\"), also referred to as the baseline in Section 3.2, a significant decline is observed in downstream performance after fine-tuning, from " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "44.11\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " NDS to " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "40.20\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " NDS. In contrast, with the 16-group Latent Rendering, the performance improves to " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "47.58\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " NDS, a notable " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "3.47\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " NDS improvement over the baseline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 432, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 481 + ], + "type": "text", + "content": "We conduct a comparison of Latent Rendering with different parallel groups in Table 11 as well. The results demonstrate a consistent improvement by dividing channels into more groups and integrating information separately." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 497, + 378, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 378, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 378, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 519, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 615 + ], + "type": "text", + "content": "In this paper, we introduced visual point cloud forecasting, which predicts future point clouds from historical visual images, as a new pre-training task for end-to-end autonomous driving. We developed ViDAR, a general model to pre-train visual BEV encoders, and designed a Latent Rendering operator to solve the ray-shaped feature issue. To conclude, our work demonstrates that visual point cloud forecasting enables scalable autonomous driving." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Limitations and Future Work. Though with the potential of scalability, in this paper, we mainly conduct pre-training on Image-LiDAR sequences from nuScenes dataset, of which the data scale is still limited. As for the future, we plan to scale up the pre-training data of ViDAR, study visual point cloud forecasting across diverse datasets, and use publicly available Image-LiDAR sequences as much as possible to train a foundation visual autonomous driving model [38]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14680" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Adil Kaan Akan and Fatma Güney. StretchBEV: Stretching Future Instance Prediction Spatially and Temporally. In ECCV, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 212 + ], + "type": "text", + "content": "[2] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A Cookbook of Self-supervised Learning. arXiv preprint arXiv:2304.12210, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "text", + "content": "[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In ICCV, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "type": "text", + "content": "[4] Alex Bewley, Pei Sun, Thomas Mensink, Dragomir Anguelov, and Cristian Sminchisescu. Range Conditioned Dilated Convolutions for Scale Invariant 3D Object Detection. arXiv preprint arXiv:2005.09927, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 358 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi-ancarlo Baldan, and Oscar Beijbom. nuScenes: A Multi-modal Dataset for Autonomous Driving. In CVPR, 2020. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 361, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 287, + 415 + ], + "type": "text", + "content": "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In ECCV, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 418, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 287, + 460 + ], + "type": "text", + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end Autonomous Driving: Challenges and Frontiers. arXiv preprint arXiv:2306.16927, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 463, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 463, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 463, + 287, + 496 + ], + "type": "text", + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint arXiv:2003.04297, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 540 + ], + "type": "text", + "content": "[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "text", + "content": "[10] OpenScene Contributors. OpenScene: The Largest Up-to-Date 3D Occupancy Prediction Benchmark in Autonomous Driving, 2023. 1, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "text", + "content": "[11] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable Convolutional Networks. In ICCV, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 193 + ], + "type": "text", + "content": "[15] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles Qi, Yin Zhou, Zoey Yang, Aurélien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset. In ICCV, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 194, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 226 + ], + "type": "text", + "content": "[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. I. J. Robotics Res., 2013. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[17] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. ViP3D: End-to-end visual trajectory prediction via 3d agent queries. In CVPR, 2023. 1, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "text", + "content": "[18] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A Survey on Self-supervised Learning: Algorithms, Applications, and Future Trends. arXiv preprint arXiv:2301.05712, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "text", + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 447 + ], + "type": "text", + "content": "[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts. In CVPR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[23] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: Future Instance Segmentation in Bird's-Eye view from Surround Monocular Cameras. In ICCV, 2021. 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[24] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular Quasi-Dense 3D Object Tracking. TPAMI, 2022. 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 570 + ], + "type": "text", + "content": "[25] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe Local Motion Planning with Self-Supervised Freespace Forecasting. In CVPR, 2021. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[26] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning. In ECCV, 2022. 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "text", + "content": "[27] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Junjie Huang, Guan Huang, Zheng Zhu, Ye Yun, and Dalong Du. BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14681" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction. In CVPR, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[30] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. DriveAdapter: Breaking the Coupling Barrier of Perception and Planning in End-to-End Autonomous Driving. In ICCV, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[31] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think Twice before Driving: Towards Scalable Decoders for End-to-End Autonomous Driving. In CVPR, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "text", + "content": "[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "type": "text", + "content": "[33] Li Jiang, Zetong Yang, Shaoshuai Shi, Vladislav Golyanik, Dengxin Dai, and Bernt Schiele. Self-supervised Pretraining with Masked Shape Prediction for 3D Scene Understanding. In CVPR, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "text", + "content": "[34] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. arXiv preprint arXiv:2004.11362, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[35] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable Raycasting for Self-Supervised Occupancy Forecasting. In ECCV, 2022. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 410 + ], + "type": "text", + "content": "[36] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point Cloud Forecasting as a Proxy for 4D Occupancy Forecasting. In CVPR, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 434 + ], + "type": "text", + "content": "[37] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 510 + ], + "type": "text", + "content": "[38] Hongyang Li, Yang Li, Huijie Wang, Jia Zeng, Pinlong Cai, Huilin Xu, Dahua Lin, Junchi Yan, Feng Xu, Lu Xiong, Jingdong Wang, Futang Zhu, Kai Yan, Chunjing Xu, Tiancai Wang, Beipeng Mu, Shaoqing Ren, Zhihui Peng, and Yu Qiao. Open-sourced Data Ecosystem in Autonomous Driving: the Present and Future. arXiv preprint arXiv:2312.03408, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 589 + ], + "type": "text", + "content": "[39] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, Hao Tian, Enze Xie, Jiangwei Xie, Li Chen, Tianyu Li, Yang Li, Yulu Gao, Xiaosong Jia, Si Liu, Jianping Shi, Dahua Lin, and Yu Qiao. Delving Into the Devils of Bird's-Eye-View Perception: A Review, Evaluation and Recipe. TPAMI, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[40] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based Topology Reasoning for Driving Scenes. arXiv preprint arXiv:2304.05277, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[41] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying Voxel-based Representation with Transformer for 3D Object Detection. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[42] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. BEVStereo: Enhancing Depth Estimation in" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Multi-View 3D Object Detection with Temporal Stereo. In AAAI, 2023. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[43] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object Detection. In AAAI, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[44] Yanwei Li, Zhiding Yu, Jonah Philion, Anima Anandkumar, Sanja Fidler, Jiaya Jia, and Jose Alvarez. End-to-end 3D Tracking with Decoupled Queries. In ICCV, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[45] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEV-Former: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[46] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and José Manuel Álvarez. FB-BEV: BEV Representation from Forward-Backward View Transformations. ICCV, 2023. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "text", + "content": "[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. PnPNet: End-to-End Perception and Prediction With Tracking in the Loop. In CVPR, 2020. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[48] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[49] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[50] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos. arXiv preprint arXiv:2308.09244, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[51] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully Sparse 3D Panoptic Occupancy Prediction. arXiv preprint arXiv:2312.17118, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "text", + "content": "[52] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position Embedding Transformation for Multi-View 3D Object Detection. In ECCV, 2022. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[53] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. In ICCV, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 567, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 611 + ], + "type": "text", + "content": "[54] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In ICRA, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "type": "text", + "content": "[55] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. arXiv preprint arXiv:1711.05101, 2019. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "type": "text", + "content": "[56] B. Mersch, X. Chen, J. Behley, and C. Stachniss. Self-supervised Point Cloud Prediction Using 3D Spatiotemporal Convolutional Networks. In CoRL, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[57] Gregory P. Meyer, Ankit Laddha, Eric Kee, Carlos Vallespi-Gonzalez, and Carl K. Wellington. LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving. In CVPR, 2019. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14682" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[58] Chen Min, Liang Xiao, Dawei Zhao, Yiming Nie, and Bin Dai. Occupancy-MAE: Self-Supervised Pre-Training LargeScale LiDAR Point Clouds With Masked Occupancy Autoencoders. TIV, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[59] Ziqi Pang, Zhichao Li, and Naiyan Wang. SimpleTrack: Understanding and Rethinking 3D Multi-object Tracking. arXiv preprint arXiv:2111.09621, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "type": "text", + "content": "[60] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is Pseudo-Lidar Needed for Monocular 3D Object Detection? In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "type": "text", + "content": "[61] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical Depth DistributionNetwork for Monocular 3D Object Detection. In CVPR, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "type": "text", + "content": "[62] Ravid Shwartz-Ziv and Yann LeCun. To Compress or Not to Compress-Self-Supervised Learning and Information Theory: A Review. arXiv preprint arXiv:2304.09355, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 251, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 251, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 287, + 303 + ], + "type": "text", + "content": "[63] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv preprint arXiv:2312.14150, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 305, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 305, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 287, + 392 + ], + "type": "text", + "content": "[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 394, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 436 + ], + "type": "text", + "content": "[65] Izzeddin Teeti, Salman Khan, Ajmal Shahbaz, Andrew Bradley, and Fabio Cuzzolin. Vision-based Intention and Trajectory Prediction in Autonomous Vehicles: A Survey. In *IJCAI*, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 437, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 437, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 437, + 287, + 459 + ], + "type": "text", + "content": "[66] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive Multiview Coding. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "type": "text", + "content": "[67] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as Occupancy. In ICCV, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 504, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 287, + 568 + ], + "type": "text", + "content": "[68] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping. In NeurIPS Datasets and Benchmarks, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 287, + 613 + ], + "type": "text", + "content": "[69] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In ICCV Workshops, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 658 + ], + "type": "text", + "content": "[70] Tai Wang, Qing Lian, Chenming Zhu, Xinge Zhu, and Wenwei Zhang. MV-FCOS3D++: Multi-View Camera-Only 4D Object Detection with Pretrained Monocular Backbones. arXiv preprint arXiv:2207.12716, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "[71] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[72] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In CoRL, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "type": "text", + "content": "[73] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "type": "text", + "content": "[74] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the Pose Forecasting Pipeline with SPF2: Sequential Pointcloud Forecasting for Sequential Pose Forecasting. arXiv preprint arXiv:2003.08376, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "text", + "content": "[75] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2Net: Stochastic Sequential Pointcloud Forecasting. In ECCV, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 266, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 310 + ], + "type": "text", + "content": "[76] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 312, + 545, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 355 + ], + "type": "text", + "content": "[77] Penghao Wu, Li Chen, Hongyang Li, Xiaosong Jia, Junchi Yan, and Yu Qiao. Policy Pre-training for Autonomous Driving via Self-supervised Geometric Modeling. In ICLR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 358, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 545, + 390 + ], + "type": "text", + "content": "[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In CVPR, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "type": "text", + "content": "[79] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In ECCV, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 438, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 472 + ], + "type": "text", + "content": "[80] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 474, + 545, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 527 + ], + "type": "text", + "content": "[81] Xiangchao Yan, Runjian Chen, Bo Zhang, Jiakang Yuan, Xinyu Cai, Botian Shi, Wenqi Shao, Junchi Yan, Ping Luo, and Yu Qiao. SPOT: Scalable 3D Pre-training via Occupancy Prediction for Autonomous Driving. arXiv preprint arXiv:2309.10527, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 530, + 545, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 530, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 307, + 530, + 545, + 584 + ], + "type": "text", + "content": "[82] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, Xiaofei He, and Wanli Ouyang. UniPAD: A Universal Pre-training Paradigm for Autonomous Driving. arXiv preprint arXiv:2310.08370, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 587, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 619 + ], + "type": "text", + "content": "[83] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. IPOD: Intensive Point-based Object Detector for Point Cloud. CoRR, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 622, + 545, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 653 + ], + "type": "text", + "content": "[84] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 656, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 677 + ], + "type": "text", + "content": "[85] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3DSSD: Point-based 3D Single Stage Object Detector, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[86] Zetong Yang, Yanan Sun, Shu Liu, Xiaojuan Qi, and Jiaya Jia. CN: Channel Normalization for Point Cloud Recognition. In ECCV, 2020." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14683" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 353 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[87] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3D-MAN: 3D Multi-Frame Attention Network for Object Detection. In CVPR, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[88] Zetong Yang, Li Jiang, Yanan Sun, Bernt Schiele, and Jiaya Jia. A Unified Query-based Paradigm for Point Cloud Understanding. In CVPR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "text", + "content": "[89] Zetong Yang, Zhiding Yu, Chris Choy, Renhao Wang, Anima Anandkumar, and Jose M. Alvarez. Improving Distant 3D Object Detection Using 2D Box Supervision. arXiv preprint arXiv:2403.09230, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "text", + "content": "[90] Jia Zeng, Li Chen, Hanming Deng, Lewei Lu, Junchi Yan, Yu Qiao, and Hongyang Li. Distilling Focal Knowledge From Imperfect Expert for 3D Object Detection. In CVPR, 2023. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "type": "text", + "content": "[91] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. MUTR3D: A Multi-camera Tracking Framework via 3D-to-2D Queries. In CVPR, 2022. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 287, + 318 + ], + "type": "text", + "content": "[92] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. BEVerse: Unified Perception and Prediction in Birds-Eye-View for Vision-Centric Autonomous Driving. arXiv preprint arXiv:2205.09743, 2022. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 353 + ], + "type": "text", + "content": "[93] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable Transformers for End-to-End Object Detection. In ICLR, 2020. 5" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14684" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_content_list.json b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a4483886ea65a2a6483eca920f3e0d8aec13b7f0 --- /dev/null +++ b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_content_list.json @@ -0,0 +1,1495 @@ +[ + { + "type": "text", + "text": "Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models", + "text_level": 1, + "bbox": [ + 104, + 130, + 866, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yushi Hu1,2* Otilia Strencu1 Chun-Ta Lu1 Krishnamurthy Viswanathan1 Kenji Hata1 Enming Luo1 Ranjay Krishna2 Ariel Fuxman1 1Google Research 2University of Washington https://visual-program-distillation.github.io/", + "bbox": [ + 158, + 200, + 807, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 306, + 313, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Solving complex visual tasks such as \"Who invented the musical instrument on the right?\" involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs. Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data.", + "bbox": [ + 75, + 338, + 473, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 794, + 209, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have become the pretrained backbone for many computer vision tasks [2, 4, 7, 9-11, 34, 39, 41, 60, 63, 74, 79]. Yet, all these models still fall short of solving numerous visual reasoning tasks expected", + "bbox": [ + 75, + 819, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/948f46a27311e660bc2833cdebe63714935e90d4853f13c32d2d80baf27e299c.jpg", + "image_caption": [ + "Figure 1. We introduce Visual Program Distillation (VPD), a training framework which leverages LLM-generated programs that make calls to specialist models and tools to distill cross-modal reasoning abilities and specialist skills into multimodal models." + ], + "image_footnote": [], + "bbox": [ + 503, + 292, + 890, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of competent vision models. Even state-of-the-art (SOTA) proprietary vision-language models such as GPT-4V [49] do not perform well on tasks that involve counting and spatial reasoning [71]. They find it difficult to count (TallyQA [1]), to compositionally reason (GQA [26]), and to reason with external knowledge (OK-VQA [44], A-OKVQA [53]). Many of these tasks require VLMs to conduct compositional reasoning, which still remains an unsolved challenge. For instance, answering the question \"Who invented the musical instrument on the right?\" involves a composition of skills: identifying objects, applying spatial reasoning to locate the one on the right, recognizing the musical instrument, and accessing prior knowledge to retrieve the inventor.", + "bbox": [ + 496, + 598, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, large language models (LLMs) have demonstrated remarkable performance at generating code that solves complex and compositional tasks [3, 8, 43, 49, 69]. Several recent papers [16, 20, 25, 57] capitalize on this by prompting LLMs to generate programs where each step corresponds to a reasoning step. The programs invoke specialized \"tools\" (or specialized vision models) to explicitly", + "bbox": [ + 496, + 796, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done while interning at Google Research.", + "bbox": [ + 94, + 886, + 354, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9590", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07033e3703646704ac7b1eba64c1ea5916cfc8cfbb8a6ec25a25a728543fef5a.jpg", + "image_caption": [ + "Figure 2. Overview of Visual Program Distillation (VPD). VPD uses an LLM and specialized vision tools to generate faithful chain-of-thought (CoT) training data for vision-language models (VLMs). Given a multimodal input, our 4-step data synthesis pipeline generates a CoT that answers the query. In the example above, our synthesized CoT contains a series of reasoning steps: find all buses, check if each bus is yellow, and aggregate the count into a final answer. The CoT also contains the grounding information given by object detection." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 890, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "execute each reasoning step. For the question above, the program would call an \"object detector\" tool to identify and isolate all the objects, a \"fine-grained object classification\" tool to recognize the musical instrument, and a \"knowledge-based question answering\" tool to retrieve its inventor.", + "bbox": [ + 75, + 405, + 470, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although innovative, generating explicit programs is computationally expensive in practice, prone to errors, and still underperforms end-to-end models. Programs require loading and executing multiple tools, leading to high latency and computational cost. Moreover, generated programs may omit necessary steps or include spurious ones. Even when the program is correct, vision model invocations can produce incorrect outputs, from which the overall program cannot recover. Unfortunately, empirical results show that visual programs still fall short of end-to-end fine-tuned models [16].", + "bbox": [ + 75, + 481, + 470, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of work is visual instruction tuning [39], which tries to distill the instruction-following ability of LLMs into VLMs. They prompt LLMs with image captions and bounding box annotations in order to generate queries and answers that are used to fine-tune VLMs [7, 23, 39, 65]. However, this approach has important limitations: image captions can miss fine-grained visual information, and LLM are prone to producing inconsistent outputs for custom vision representations like bounding boxes [18]. As a result, existing instruction-tuned VLMs still struggle with tasks that require complex visual reasoning [7, 14, 38, 40].", + "bbox": [ + 75, + 630, + 470, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present Visual Program Distillation (VPD), a novel distillation method that induces complex cross-modal reasoning capabilities into vision-language models (Fig. 2). As the name suggests, VPD combines two key insights to deliver a training paradigm that surpasses the sum of its parts: It relies on (1) advancements in visual programs that use tools [20] and (2) the recent breakthroughs in dis", + "bbox": [ + 75, + 796, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tillation through chain-of-thought reasoning [21]. Given a labeled training dataset of complex visual tasks, VPD generates a correct program, and then distills its reasoning steps into vision-language models. To avoid using programs that give the wrong answer, VPD prompts an LLM to generate multiple candidate programs, and executes every one of them. When labeled data is available, it then filters for programs that produce the correct answer upon execution. Therefore, our programs comprise multiple vision tools, are executable, and yield the correct answer when executed. Next, VPD rewrites the correct programs as natural language chain-of-thought instructions and uses step-by-step distillation [21] to inject the reasoning abilities into VLMs.", + "bbox": [ + 496, + 405, + 893, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our best instruction-tuned model, PaLI-X-VPD (55B), sets a new SOTA result on 8 classical VQA tasks and 2 zero-shot multimodal benchmarks. Our models even outperform the recent SOTA established by PaLI-X [10]. Importantly, we conduct a quality evaluation with human raters which shows that PaLI-X-VPD generates more consistent and faithful rationales compared to its counterpart trained using instruction-tuning data. In addition, we experiment with PaLI-3 (5B), and show that VPD also improves the performance of smaller-scale models. Finally, experiments on Hateful Memes [29] show that VPD is also helpful for adapting to new tasks, even when no labels are available.", + "bbox": [ + 496, + 599, + 893, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 794, + 635, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VPD is a general method for improving any vision-language model, and includes automatic program generation and training with chain-of-thought data as steps of the proposed framework. We discuss each of these research areas.", + "bbox": [ + 496, + 819, + 893, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language models (VLMs). Most recent generative", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9591", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLMs share a common structure: a pre-trained visual encoder, a pre-trained LLM, and a connector between the two modalities [2, 4, 9, 34, 39, 41, 60, 63, 74, 79]. The models are trained on large-scale image-text pairs from various tasks to adapt to both modalities. They are also tuned on LLM-generated visual instructions to enable models to follow versatile instructions from users [39]. Some models also include bounding boxes in the pre-training data to improve the VLM on visually grounded tasks [6, 7, 10, 41, 50, 61, 78]. The bounding boxes are usually retrieved from COCO [37] and Visual Genome [32]. Different from prior work, our method does not rely on provided dense annotations. We use LLM-generated code and specialized vision tools to generate our own visual instruction-tuning data.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual programming and agents. With the advancement of large language models (LLMs) [3, 5, 13, 49, 59], recent work has started using LLMs as an interface to solving complex reasoning tasks with tools [12, 42, 51, 54, 76], and also as an agent for vision tasks [23, 25, 70, 72]. From this line of work, the most relevant to us are VisProg [20] and ViperGPT [16], which leverage LLMs to generate executable programs with a sequence of invocations to specialized vision tools. They achieve SOTA zero-shot performance on various vision tasks, while being versatile and interpretable.", + "bbox": [ + 75, + 305, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training and inference with chain-of-thought. Chain-of-Thought (CoT) [66] has become a popular approach to improving LLM performance. Recent work such as Program-of-Thoughts (PoT) [8] and Faithful CoT [43] further improve this framework by splitting inference into two steps: first generate a program with LLM, then execute the program. This approach achieves better accuracy and reduces hallucinations in the reasoning steps. Moreover, CoT is also used to train language models. Distill step-by-step [21], PaD [80], and SCOTT [62] train smaller language models with CoT generated by LLMs, showing that this can improve model performance and reasoning consistency. Mammoth [68] trains an LLM with a hybrid of CoT and PoT rationales and achieves a SOTA model for math problems.", + "bbox": [ + 75, + 460, + 472, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Visual Program Distillation (VPD)", + "text_level": 1, + "bbox": [ + 76, + 680, + 388, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce VPD, a general model-agnostic framework that distills the reasoning power of LLM-generated programs together with the low-level image understanding abilities of vision tools into a single vision-language model (Fig. 2). At its core, VPD consists of two major steps:", + "bbox": [ + 75, + 702, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Program generation and verification: Given a textual query $q$ and a visual input $i$ , VPD first generates a program that solves the query by making use of vision modules and tools, then converts the program execution trace into a chain-of-thought $c$ (\\$3.1).", + "2. Distilling step-by-step: The visual input $i$ , textual query $q$ , and chain-of-thought $c$ produced by the previous step are distilled into a vision-language model (VLM) (§3.2)." + ], + "bbox": [ + 76, + 777, + 472, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Program generation and verification", + "text_level": 1, + "bbox": [ + 500, + 90, + 815, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our training data synthesis pipeline is illustrated in the blue boxes of Fig. 2, and consists of four stages. Given a sample $(i,q,y)$ consisting of a visual input $i$ and a textual query $q$ about its content and, when available, its ground truth answer $y$ , we perform the following sequence of steps:", + "bbox": [ + 496, + 114, + 893, + 191 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Program generation with LLM: Given $q$ , we generate a list of $k$ candidate programs $\\pi(q) = \\{z_1, z_2, \\dots, z_k\\}$ , where $\\pi$ represents a program generation function.", + "2. Program execution with vision modules: We execute each program $z_{i}$ with an execution engine $\\phi$ to obtain its final result $\\phi (i,z_i) = \\hat{y}_i$ . However, during program execution, we maintain the execution trace $t_i$ recording all the intermediate function calls and outputs. At the end of this step, we produce a list of programs, results, and execution traces $\\{(z_1,\\hat{y}_1,t_1),\\dots,(z_k,\\hat{y}_k,t_k)\\}$ .", + "3. Program filtering: Among the $k$ candidate programs from the previous step, we keep a single tuple $(z,\\hat{y},t)$ with correct answer.", + "4. Converting program execution traces into chain-of-thought rationales: We rewrite $t$ into a CoT $c$ using an LLM." + ], + "bbox": [ + 500, + 193, + 893, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now discuss in detail each of the steps above.", + "bbox": [ + 500, + 435, + 823, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Program generation. We adopt a similar approach with recent work [16, 20] in our program generation step, and use PaLM-2 [3] as LLM to generate candidate programs for a given query $q$ . We prompt PaLM-2 with the same kind of text prompt as used by ViperGPT, which contains a detailed description of the available vision modules, followed by the query $q$ (prompt in Appendix §G). The LLM is expected to directly output a Python function definition that will be executed in the following steps. However, in our experiments, we find that the success rate of top- $k$ programs is much higher than the top-1 program. Therefore, in contrast to prior work, which only samples one program $z$ , when the ground truth answer $y$ is available, we set a temperature $T$ for LLM decoding and sample a list of top- $k$ candidate programs $\\{z_1, z_2, \\dots, z_k\\}$ from the LLM. Then, we filter out one correct program $z$ in later steps. As shown in our ablation in §4.2, this becomes crucial to our performance gain. We use $T = 0.5$ and $k = 5$ for all our experiments. For unlabeled data, $k = 1$ and the filtering step can be skipped.", + "bbox": [ + 496, + 458, + 893, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Program execution with vision modules. We use the same execution engine $\\phi$ as ViperGPT [16]. An LLM-generated program $z_{i}$ is a Python function that takes the visual input $i$ as input. While the LLM outputs the program as a sequence of text, the execution engine $\\phi$ is able to interpret this as a Python program and execute it, to obtain its return result $\\hat{y}_{i}$ . Additionally, $\\phi$ also records the execution trace $t_{i}$ of the program $z_{i}$ , which keeps a record of all the vision module calls, their inputs, and outputs. We use the following tools for the program: PaLI-X [10] for simple visual queries, PaLI-X", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9592", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection (distilled from OWLv2 [47]) for object detection; Google Cloud Depth API [17] for depth estimation, and PaLM-2 [3] for external knowledge.", + "bbox": [ + 75, + 90, + 472, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Program filtering. As we discussed in §1, visual programs are error-prone for a variety of reasons. The program might be wrong, and the execution process can introduce additional errors. To overcome these issues, we employ a program filtering step. We start by sampling top- $k$ programs and attempt to execute each one. During this process, any program that fails execution is instantly discarded. We further filter the remaining programs, the strategy depending on the availability of labeled data. For tasks where human labels are available, we select a single program per input sample whose answer is correct—that is, when the program output $\\hat{y}_i$ matches the human label $y$ . In this case our visual program pipeline acts like a CoT rationale annotator. One potential problem is that some task answers are ambiguous. For example, for a question like \"Where are the horses?\", the answers \"mountain\" and \"mountains\" are both correct. We adopt the method in [28] and use an LLM to determine if the program output is correct (details in §G). If there is more than one correct program per question, we select the top scoring one, according to the scores provided by the program-generating LLM $\\phi$ . If no programs pass the test, the rated sample is not wasted—we simply use the correct answer $y$ as supervision in our fine-tuning stage, without an associated CoT. When no human-rated answer is available, we directly use the top scoring executable program1.", + "bbox": [ + 75, + 141, + 472, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Converting program execution traces into chain-of-thought rationales. After the filtering step, for each visual input $i$ and query $q$ , we will have selected at most one program $z$ together with its execution result $\\hat{y}$ and trace $t$ . Since most existing VLMs have been pre-trained with text in natural language and not code, we use an LLM to rewrite the execution trace $t$ into a natural language CoT $c$ for our VLM distillation. Some examples are shown in §B. Concretely, similar to prior work [23, 24], we hand-craft 20 examples of how an input $(q, z, t)$ can be converted into a CoT, and use these as few-shots for prompting PaLM-2 [3], which performs in-context learning and generates CoTs for new examples. We include a concrete example in §G.", + "bbox": [ + 75, + 521, + 472, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Distilling step-by-step", + "text_level": 1, + "bbox": [ + 76, + 723, + 282, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this step, we fine-tune a backbone VLM with the training data generated in §3.1, distilling the knowledge and reasoning steps of our generated programs into a single end-to-end model. We do so in a multitask fashion, where the VLM is simultaneously fine-tuned on data synthesized for multiple types tasks (e.g., free-form VQA, multiple choice).", + "bbox": [ + 75, + 746, + 468, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $f$ represent our VLM model. While the same VLM architecture could solve all these tasks, it needs to be prompted", + "bbox": [ + 75, + 837, + 470, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "differently to adapt to the task. Following prior work [39], we manually design instructions for each task. For example, for free-form VQA queries, the instruction is \"Answer with a single word or phrase\", while for multiple-choice queries we use \"Answer with the option letter from the given choices directly\". During fine-tuning, we combine the training samples generated for all tasks into a single dataset, so to account for the different types of tasks, we augment each sample with its corresponding task-specific prompt $p$ .", + "bbox": [ + 504, + 90, + 895, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using these instructions along with the data generated in §3.1, we put together the training dataset $\\mathcal{D} = \\{(i_j,q_j,\\hat{y}_j,c_j,p_j)\\}_{j = 1}^N$ , where $N$ is the total number of samples, $i_j$ is the visual input, $q_{j}$ is the textual query, $\\hat{y}_i$ is the visual program output2, and $c_{j}$ is the CoT rationale.", + "bbox": [ + 498, + 227, + 893, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train $f$ to minimize a loss for predicting both the label and the rationale. As shown in the red box of Fig. 2, similar to [21], we treat predicting the output label $\\hat{y}_j$ and the rationale $c_j$ as two separate optimization goals. However, since VLMs are open-ended text generation models, they need additional prompting to indicate whether we want a short answer or a long answer that includes a rationale. Therefore, we append the suffix $s_c =$ \"Explain the rationale to answer the question\" at the end of the prompt for generating a CoT, and use the task instruction $p_j$ for short answers. Our optimization objective is:", + "bbox": [ + 498, + 301, + 893, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\mathcal {L} _ {\\text {r a t i o n a l e}} (1) \\\\ = \\sum_ {j = 1} ^ {N} \\ell \\left(f \\left(i _ {j}, q _ {j}, p _ {j}\\right), \\hat {y} _ {j}\\right) + \\ell \\left(f \\left(i _ {j}, q _ {j}, s _ {c}\\right), c _ {j}\\right) (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 479, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\ell$ is the cross entropy-loss normalized by sequence length to ensure that the labels (typically short) and rationales (typically long) have similar weights. $\\mathcal{L}_{\\text{rationale}}$ both teaches the VLM to generate faithful reasoning steps similar to program execution traces, and carries more information beyond the label $\\hat{y}$ that also helps the VLM in better predicting the label. During test time, the rationale generation is not required. We can adjust the task instruction to directly get the short output label using $p$ , and using $s_c$ to get the human-interpretable reasoning steps if needed.", + "bbox": [ + 498, + 554, + 893, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 719, + 633, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we demonstrate VPD's effectiveness by using it to train a generalist VLM. We attempt this for two VLMs with different scales, PaLI-3 (5B) [11] and PaLI-X (55B) [10]. Detailed experimental setups are given in §4.1. Qualitatively, models fine-tuned with VPD exhibit the ability to reason step-by-step like a program, as illustrated in Fig. 3, and supported by human evaluation results (§4.3). Quantitative results show that our PaLI-3-VPD and PaLI-X-VPD", + "bbox": [ + 498, + 744, + 893, + 864 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2When labeled data is available, this is equivalent to the ground truth label $y$ due to our filtering strategy.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Studying ways to approximate program correctness strategies for unlabeled data is an interesting direction for future work.", + "bbox": [ + 75, + 875, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9593", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/279813154032f56df66d0dcf9989edf3655182e43c8d8b3529479ce489ffafdc.jpg", + "image_caption": [ + "Figure 3. Outputs of PaLI-X trained with VPD. Training with execution traces of filtered programs improves the model's ability to count, understand spatial relations, and reason compositionally. The images are from Visual Genome [32]." + ], + "image_footnote": [], + "bbox": [ + 89, + 78, + 447, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "achieve new SOTA on a broad range of tasks, on both generalist and per-task fine-tuning settings (§4.2). We also conduct a detailed analysis of the source of performance gain (§4.2). Finally, we conduct human evaluation on the quality of the rationales generated by our models, and compare it with rationales generated by an instruction-tuned model trained without VPD (§4.3).", + "bbox": [ + 75, + 328, + 472, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 445, + 264, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone models. We use two state-of-the-art VLMs, PaLI-3 (5B) [11] and PaLI-X (55B) [10] as our base models. Both take images and text as input, and generate text as output. For simplicity, we further refer to them as \"PaLI model\" when we discuss steps performed with each of them individually.", + "bbox": [ + 75, + 473, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data for Generalist Models. We fine-tune the pre-trained PaLI model on two types of datasets to make it a generalist VLM—that is, a model that performs relatively well on any task without further training on that specific task:", + "bbox": [ + 75, + 569, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Multimodal Instruction-Tuning (MMIT) tasks, created in the spirit of Self-Instruct [64]. An LLM is prompted with image captions, and generates task inputs and the desired outputs about the corresponding image. Details of MMIT tasks are covered in [65]. Note that these tasks cover a wide variety of common use cases, but do not include the specific in-domain tasks used in this work.", + "(2) Academic task-oriented VQA tasks. Since image captions contain only a coarse description of visual information in the image, and may miss details that are important for solving the task. Additionally, LLM are likely to hallucinate during this data curation process. To further boost the accuracy of our VLMs, we also fine-tune the PaLI models with academic task-oriented VQA tasks. The data mixture covers subsets of a wide variety of VQA tasks, including general VQA (VQAv2 [19]), optical character recognition (OCRVQA [48]), compositional questions and reasoning (GQA [26]), counting (TallyQA [1]), and VQA that involves" + ], + "bbox": [ + 75, + 630, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "external knowledge (OK-VQA [45] and A-OKVQA [52]). The tasks contain a textual query and a short expected label. We use the pipeline in §3.1 to synthesize CoT reasoning steps for these labels, and tune the PaLI model with these the loss in Equation 1. Notice that sometimes the pipeline fails to find a program that generates the correct answer. In that case, we set $\\mathcal{L}_{\\text{rationale}}$ to 0 and only keep $\\mathcal{L}_{\\text{label}}$ . §4.2 shows how many programs are kept after the filtering stage.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data for specialist models. While fine-tuning the generalist model with VPD, we only use a subset of each task's training data. To evaluate our model's ability on each individual task, we continue fine-tuning PaLI-3-VPD and PaLI-X-VPD on each individual task on the training splits.", + "bbox": [ + 496, + 218, + 893, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training setup. Both PaLI-3 and PaLI-X follow an encoder-decoder architecture, where images are encoded into visual tokens via a visual encoder and then passed into a UL2 model. Due to resource constraint, we use LoRA [22] to fine-tune both PaLI models. Specifically, we add LoRA weights on each linear layer in the attention blocks and the MLP blocks for both the encoder and decoder in the UL2 transformer. More training details are in §C.", + "bbox": [ + 496, + 301, + 893, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation setup. We evaluate our models on a wide range of tasks, including various VQA tasks and recent zero-shot VLM benchmarks. Noted that A-OKVQA contains two kinds of questions, multiple-choice (MC) and direct answer (DA). We report results on both. TallyQA [1] contains complex counting questions that involve object relationships, attribute identification, and reasoning to get the correct answer. It contains two evaluation partitions, simple and complex. TextVQA [55] focuses reading texts. We include it to evaluate our models on zero-shot tasks. In addition to VQA tasks, we also test our models on two popular VLM benchmarks. POPE [36] focuses on VLM hallucination, containing binary questions of whether or not an object exists in the image. MMBench [40] is a robust and comprehensive VLM benchmark testing a range of fine-grained abilities (e.g., object localization, attribute recognition, spatial relationship). Details of the evaluation sets and metrics are in §C.", + "bbox": [ + 496, + 429, + 895, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We refer to the two PaLI models fine-tuned with VPD as PaLI-3-VPD and PaLI-X-VPD, and compare them with various baselines. To evaluate the effectiveness of VPD, we experiment with removing the synthesized CoT from our data mixture, and train the PaLI models with the exact same hyper-parameters and steps. We call these models as PaLI-3 Instruct and PaLI-X Instruct. For a fair comparison, these models are also trained with the same supervised loss for predicting the ground truth answers, on the same images and textual queries as our VPD variant. Moreover, we also compare with the most recent SOTA vision-language models. These VLMs are initialized with pre-trained visual encoders and LLMs, and then trained with image-text pairs, LLM-generated data, and academic tasks.", + "bbox": [ + 496, + 691, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9594", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a7f954b3b4fbdb54a373f64988542be82ee7d54d51f968fdd94f0edebc74cdf0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Generalist ModelsVQAv2GQAOK-VQAA-OKVQATallyQATextVQAPOPEMMB
MCDASimp.Comp.
Prior generalist VLMs
Flamingo (80B) [2]82.0-57.8*----57.1--
MiniGPT-4 (Vicuna-13B) [79]-43.5-67.2-----42.3
InstructBLIP (Vicuna-13B) [14]-49.5*---75.2*57.5*50.7*78.944.0
Shikra (Vicuna-13B) [7]77.4-47.2-----84.758.8
Qwen-VL (9.7B) [4]78.859.358.6--82.6*65.8*63.8-38.2
Qwen-VL-Chat (9.7B) [4]78.257.556.6--81.1*64.0*61.5-60.6
mPLUG-Owl2 (8.2B) [74]79.456.157.7----54.3*86.264.5
LLaVA-1.5 (Vicuna-13B) [38]80.063.3---76.9*65.4*61.3*85.967.7
Our instruction-tuned baselines
PaLI-3-Instruct (5B)79.959.756.778.357.681.970.463.3*87.768.6
PaLI-X-Instruct (55B)83.663.364.384.161.585.575.465.0*88.975.0
Our visual program distillation models
PaLI-3-VPD (5B)80.461.357.578.556.583.170.963.7*88.669.0
PaLI-X-VPD (55B)83.964.964.684.562.786.276.665.4*88.876.2
", + "bbox": [ + 76, + 77, + 898, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with SOTA generalist VLMs. PaLI-X-VPD outperforms all prior models. VPD improves performance on 8/9 tasks for both PaLI-3 and PaLI-X, and is particularly effective on counting questions (TallyQA), compositional questions (GQA), and the comprehensive benchmark (MMBench). Underline indicates when PaLI-3-VPD outperforms the Instruct version. * marks tasks unseen during training. POPE and MMBench are zero-shot benchmarks for all models.", + "bbox": [ + 75, + 292, + 893, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative results", + "text_level": 1, + "bbox": [ + 76, + 363, + 266, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalist model. Table 1 compares VPD with the baselines discussed in §4.1. We infer all answers by open-ended generation with the prompt \"Answer with a single word or phrase,\" using greedy decoding without any constraint on the model's output space. For multiple-choice questions, we run inference with the prompt \"Answer with the option letter from the given choices directly.\" and generate the option letter. As Table 1 shows, PaLI-X-VPD sets the new state-of-the-art on all benchmarks. Compared with prior generalist VLMs, it achieves significant improvement on MMBench (+8.5), TallyQA complex (+9.8), and A-OKVQA(+9.5 compared with specialist SOTA[14]). PaLI-3-VPD, despite its small architecture size (5B), outperforms all prior models on VQAv2, TallyQA, POPE, and MMBench, including much larger models that use Vicuna-13B as backbones [7, 14, 38]. However, its performance on knowledge-based VQA tasks does not outperform the prior SOTA. Our hypothesis for this is that its language model (3B UL2) is too small to contain all the necessary knowledge needed for correctly solving these tasks.", + "bbox": [ + 73, + 386, + 473, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When compared to their Instruct variants, both PaLI-3-VPD and PaLI-X-VPD outperform on 11/12 tasks. Specifically, for PaLI-X, VPD obtains a $+1.6$ improvement on GQA (which is heavily focused on compositional questions, spatial relationship, and localization), $+1.2$ on TallyQA for complex questions, and $+1.2$ on MM-Bench. These results suggest that VPD is a more effective method for creating instruction-tuning data which enables VLMs to improve their visual reasoning ability.", + "bbox": [ + 73, + 685, + 470, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Per-Task Fine-Tuning (Specialist). The results for the specialist models are shown in Table 2. PaLI-X-VPD sets a new SOTA on all benchmarks. Note how the specialist models tend to have higher scores than the generalist one. We propose several hypotheses for this performance gain:", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(1) As shown in Table 2, the score gap tends to be larger on free-form VQA tasks. This may be due the fact that human annotations on these datasets have ambiguities. For example, for the question \"Who is looking up?\", GQA [26] labels are \"man\" or \"woman\" while OK-VQA [45] have more detailed labels, for example, \"worker\" or \"cook\". Per-task fine-tuning alleviates this annotation ambiguity and lets the model focus on the annotation style of that task. For multiple-choice and counting tasks, the answer has less ambiguity, and the score gaps are much smaller. (2) The performance gap between PaLI-3 (5B) specialist and generalist is larger than that of PaLI-X. We hypothesize that this shows models with larger scale has more multi-task capacity; (3) Per-task fine-tuning adds more in-domain training data, which generally improves performance.", + "bbox": [ + 496, + 363, + 893, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis: Sampling multiple programs is key to good data generation. Fig. 5 shows the success rate of finding at least one program that passes the filtering stage, when the LLM generates the top-1 or top-5 programs, respectively. There is a dramatic increase in success rate from 1 program to 5: $+45\\%$ on GQA and A-OKVQA, $+33\\%$ on OK-VQA, and $+10\\%$ on TallyQA. This design choice greatly improves our data synthesis efficiency, and, as a consequence, adding more CoT data that requires complex reasoning in our training set. We also conduct an analysis in $\\S E$ to compare the performance of VPD models with that of the visual programs they are distilled from.", + "bbox": [ + 496, + 594, + 895, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Human evaluation on rationales", + "text_level": 1, + "bbox": [ + 498, + 782, + 782, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we focus on the quality of model outputs. We performed this analysis using human annotators, who are asked to evaluate both the correctness of the final answer and the quality of the rationale behind it.", + "bbox": [ + 496, + 804, + 893, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We compare PaLI-X-VPD with PaLI-X Instruct. Among possible baselines, we chose PaLI-X", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9595", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d0ec7774655a7ee275244cc4fc1113ed0fbad2ef9aac8d0aea911e9a1d8f267.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Specialist ModelsGQA test-devOK-VQA valA-OKVQATallyQA
Multi-choice valtestDirect Answer valtestSimple testComplex test
InstructBLIP (Vicuna-7B) [14]-62.175.773.464.062.1--
PaLI-3 (5B) [11]-60.1----83.370.5
PaLI (17B) [9]-64.5----81.770.9
CogVLM (17B) [63]65.264.7------
PaLI-X (55B) [10]-66.1----86.075.6
PaLI-3-VPD (5B) generalist61.357.578.5-56.5-83.170.9
PaLI-3-VPD (5B) specialist64.760.379.776.565.563.683.370.8
PaLI-X-VPD (55B) generalist64.964.684.5-62.7-86.276.6
PaLI-X-VPD (55B) specialist67.366.885.280.471.168.286.276.4
", + "bbox": [ + 127, + 88, + 836, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of per-task fine-tuning results of specialist models. PaLI-X-VPD sets a new SOTA for all the tasks.", + "bbox": [ + 122, + 266, + 843, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a595d7b0c04990ac23bb4d038af1cadfa30941ed3dfcf7d9aabd3157eb1650e.jpg", + "image_caption": [ + "Figure 4. Human evaluation results assessing answer and rational quality for PaLI-X Instruct and PaLI-X VPD: (a) percentage of answers that are correct and have explanations; (b) rationale factuality and consistency for the answers that contain an explanation; (c) preference between the two models, when aggregating across all samples (\"All\") and across those with correct answers (\"When correct\")." + ], + "image_footnote": [], + "bbox": [ + 158, + 287, + 331, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cf7da159c3d1a5b097eea3a36149d00c37d3020592f15ed4d6c98a346b0dd6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 287, + 513, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e85a5d609dee9a1abdef080c7e96d0ccbc3273eaf73395e8d6d4be058b0979e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 287, + 803, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4708f067e55637075c087e96be6e10935beed8744c236b7f28a61bd46232ac3.jpg", + "image_caption": [ + "Figure 5. Success rate of top-1 program and top-5 programs on the training set during our data synthesis process." + ], + "image_footnote": [], + "bbox": [ + 98, + 460, + 444, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruct because it is trained to generate long-form answers [65], which allows us to assess if it is the quality of PaLI-X-VPD's answers that the annotators prefer, rather than its length. Since PaLI-X Instruct is also instruction-tuned, it can be prompted to provide long answers to alleviate this confounder.", + "bbox": [ + 75, + 609, + 470, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation protocol. We run inference with each of the two models on a combination of 600 samples from GQA (test-dev), A-OKVQA (val), and TallyQA (Simple and Complex) and record their answers and rationales. We then ask 3 human annotators to evaluate each model answer. We use prior work from natural language processing (NLP) as inspiration, and build upon it for selecting the evaluation criteria [75]. Given an image and a query, for each model-generated rationale, we ask human annotators to score the model answers along the following criteria: (1) correctness—is the final answer correct? (2) explainability—does the model explain its rationale for reaching the final answer? (3) factuality—is every step in the rationale factually cor", + "bbox": [ + 75, + 705, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rect (with respect to the image and external knowledge)? (4) consistency—does step and final answer logically follow from the previous ones? Note that a rationale may have the wrong answer while being consistent. We also conduct a side-by-side comparison, and ask annotators which of the two answers—the one provided by PaLI-X-VPD or by PaLI-X Instruct—they prefer in terms of quality of the answer and explanation. More details about the annotation protocol are in §D.", + "bbox": [ + 496, + 467, + 893, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human evaluation results. The results of the evaluation are first averaged among the human raters per sample, then aggregated across samples. Our PaLI-X-VPD model far outperforms PaLI-X Instruct along all criteria. Fig. 4 (a) shows the correctness of the final answer provided by each model (i.e. accuracy), and the proportion of the samples where the model provides a CoT rationale to explain its answer. PaLI-X-VPD's gain in accuracy of $+16.7\\%$ is even more impressive than in evaluations in §4.2 based on benchmark labels—this is because human annotators are better able to assess correctness for ambiguous questions with different possible interpretations (e.g., the model's answer \"In the living room.\" to the question \"Where is the couch located?\" was considered correct by annotators, even when the benchmark answer was \"On the right\"). Moreover, the explainability results confirm that our model is able to explain its own answer on $+24\\%$ more samples than the instruction tuned model. Additionally, Fig. 4 (b) shows impressive rationale quality: among the samples where", + "bbox": [ + 496, + 616, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9596", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an explanation is provided, PaLI-X-VPD's rationales are factual $87.2\\%$ of times, and consistent $97.8\\%$ of times, a gain of $+14.6\\%$ and $+10\\%$ , respectively, when compared to PaLI-X Instruct.", + "bbox": [ + 75, + 90, + 468, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also asked the annotators which of the two answers given by the two models they prefer. We aggregated these results in two ways: (1) across all samples, (2) across the samples where both models answered the question correctly. The results in Fig. 4 (c) show that PaLI-X-VPD is preferred on $25\\%$ more samples than PaLI-X Instruct in the \"All\" case, and on $12\\%$ more samples when both are correct. This suggests that even when PaLI-X-VPD makes mistakes, it still provides a better answer. Such examples are shown in $\\S D$ . These results confirm that a model fine-tuned with VPD leads to more faithful and consistent answers.", + "bbox": [ + 75, + 151, + 470, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Experiments on content moderation", + "text_level": 1, + "bbox": [ + 76, + 329, + 401, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prior experiments focus on training generalist VLMs. Here, we explore the effectiveness of VPD at quickly adapting models to real-world applications from a different domain than the training data [e.g., 56]. We experiment on Hateful Memes [29], a content moderation dataset where the task is to classify if a meme contains hateful content. The target labels are \"yes\" or \"no\", and models are evaluated in terms of classification accuracy and AUC-ROC. We experiment with two settings: supervised, in which the models are trained on the provided training set with 8,500 labels, and unsupervised, in which no human labels are provided.", + "bbox": [ + 75, + 353, + 470, + 518 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a22b52d3d7d759a1ba81602fe62b5a5873561289611668bc83917592e0010f8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAccAUC-ROC
Unsupervised / Zero-Shot Methods
Generated Programs (ours)69.770.1
Supervised Methods
VisualBERT [35]69.575.4
Flamingo (80B) [2]-86.6
Previous SOTA [46]78.886.7
PaLI-X-VPD (label-only FT)77.688.0
PaLI-X-VPD (specialist w/ CoT)80.889.2
Human [29]84.782.7
", + "bbox": [ + 98, + 525, + 441, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results on Hateful Memes [29] seen test set. We improve SOTA for both zero-shot and supervised settings. Unsupervised PaLI-X-VPD outperforms strong supervised baselines.", + "bbox": [ + 75, + 662, + 468, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Unsupervised setting", + "text_level": 1, + "bbox": [ + 76, + 712, + 274, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We experiment with three methods:", + "text_level": 1, + "bbox": [ + 76, + 734, + 313, + 750 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Generated Program: The program generated by PaLM-2 [3] for solving this task (shown in §F) consists of following main steps: (1) get image description with PaLI-X [10]; (2) use an OCR tool to extract embedded texts; (3) given the image description and OCR texts, ask PaLM-2 to explain if this meme is hateful.", + "2. PaLI-X-VPD (generalist): In a zero-shot setting, we directly prompt our PaLI-X-VPD (generalist) with: \"The text is $. Is this a hateful meme?\" We compute the probability of PaLI-X-VPD generating \"yes\"" + ], + "bbox": [ + 76, + 750, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "or \"no\" and measure accuracy and AUC-ROC.", + "bbox": [ + 517, + 90, + 828, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3. PaLI-X-VPD (specialist with zero-shot CoT): We follow our VPD pipeline and convert the execution traces of the program into CoTs, and then fine-tune PaLI-X-VPD to output these CoTs. Since no groundtruth labels are available, no filtering is done during the process.", + "bbox": [ + 500, + 106, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VPD significantly improves performance even when no labels are available. As shown in Table 3, PaLI-X-VPD (generalist) outperforms all other VLMs (61.4%). Interestingly, the generated programs themselves get much higher accuracy (69.7%) than on the previous datasets, perhaps because PaLM-2 is better suited at analyzing a meme than typical VQA datasets. Moreover, our PaLI-X-VPD (specialist) sets a new zero-shot SOTA on Hateful Memes, and even outperforms supervised VisualBERT [35]. Manual inspection of model outputs show a great deal of similarity with the generated code. See examples in §F.", + "bbox": [ + 496, + 186, + 893, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Supervised setting", + "text_level": 1, + "bbox": [ + 498, + 364, + 676, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We fine-tune PaLI-X-VPD in two ways:", + "bbox": [ + 498, + 388, + 774, + 402 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Label-only fine-tuning: To establish a strong baseline, we first experiment with the traditional supervised setting, where we fine-tune the model to output \"yes\" or \"no\".", + "2. PaLI-X-VPD (specialist): We use the complete VPD pipeline to train this model. We select an execution trace that leads to the correct label, and use it to tune our VLM." + ], + "bbox": [ + 500, + 402, + 893, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Our PaLI-X-VPD (specialist) sets a new SOTA for this task, with an accuracy of $80.8\\%$ and AUC-ROC of $89.2\\%$ . It outperforms the label-only fine-tuning baseline and significantly improves the SOTA metrics, achieving nearly human-level accuracy, and superhuman AUC-ROC.", + "bbox": [ + 496, + 500, + 893, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion, limitations, and future work", + "text_level": 1, + "bbox": [ + 498, + 604, + 862, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced VPD, a framework for distilling the reasoning abilities of LLMs along with the capabilities of vision tools into VLMs. VPD synthesizes training data for VLMs by generating programs that can leverage external tools. We used this technique to fine-tune some of the best existing VLMs (PaLI-3 and PaLI-X) and established new SOTA results on 8 classical VQA and 2 zero-shot multimodal benchmarks. According to human evaluations, VPD-tuned models provide more accurate answers and better rationales. Experiments on the Hateful Memes dataset show how VPD can also adapt models to new real-world domains, even when no labeled data is available, also establishing a new SOTA on this dataset. From our experiments we learnt that both a strength and a weakness of VPD is the quality of the generated programs. Better programs lead to bigger gains with VPD. We discuss multiple ideas on how to improve the program generation in Appendix H, including using additional tools, agents, and fact-checking for multimodal CoTs.", + "bbox": [ + 496, + 631, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9597", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 1, 5, 14, 15", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022. 1, 3, 6, 8", + "[3] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby Ambrose Slone,Daniel Smilkov David R. So Daniel Sohn Simon Tokumine,Dasha Valter,Vijay Vasudevan,Kiran Vodrahalli,Xuezhi Wang,Pidong Wang,Zirui Wang,Tao Wang John Wieting,Yuhuai Wu,Kelvin Xu,Yunhan Xu,Linting Xue,Pengcheng Yin,Jiahui Yu,Qiao ZhangSteven Zheng Ce ZhengWeikang ZhouDenny ZhouSlav Petrov and Yonghui Wu.Palm 2 technical report,2023.134813", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 1, 3, 6", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigptv2: large language model as a unified interface for vision-language multi-task learning, 2023. 3", + "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 1, 2, 3, 6", + "[8] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks, 2022. 1, 3", + "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 1, 3, 7", + "[10] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, AJ Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali-x: On scaling up a multilingual vision and language model, 2023. 2, 3, 4, 5, 7, 8, 15, 16", + "[11] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, Daniel Salz, Xi Xiong, Daniel Vlasic, Filip Pavetic, Keran Rong, Tianli Yu, Daniel Keysers, Xiaohua Zhai, and Radu Soricut. Pali-3 vision language models: Smaller, faster, stronger, 2023. 1, 4, 5, 7, 15, 16", + "[12] Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, et al. Binding language models in symbolic languages. arXiv preprint arXiv:2210.02875, 2022.3", + "[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robin-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9598", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "son, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways, 2022. 3", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 6, 7", + "[15] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaeldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 15", + "[16] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 1, 2, 3, 20, 24", + "[17] Ruofei Du, Eric Turner, Maksym Dzitsiuk, Luca Prasso, Ivo Duarte, Jason Dourgarian, Joao Afonso, Jose Pascoal, Josh Gladstone, Nuno Cruces, Shahram Izadi, Adarsh Kowdle, Konstantine Tsotsos, and David Kim. DepthLab: Real-Time 3D Interaction With Depth Maps for Mobile Augmented Reality. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. ACM, 2020. 4", + "[18] Weixi Feng, Wanrong Zhu, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. *ArXiv*, abs/2305.15393, 2023. 2", + "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5, 15", + "[20] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3", + "[21] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, 2023. 2, 3, 4", + "[22] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen" + ], + "bbox": [ + 78, + 90, + 472, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 5, 16", + "[23] Yushi* Hu, Hang* Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2, 3, 4", + "[24] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A. Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2627-2643, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 4", + "[25] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1, 3", + "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 5, 6, 14, 15", + "[27] Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir Ghaemmaghami, Rajendra Gottipati, William Gulland, Robert Hagemann, C. Richard Ho, Doug Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander Kaplan, Harshit Khaitan, Daniel Killebrew, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon, James Law, Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana Maggiore, Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy Nix, Thomas Norrie, Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Matt Ross, Amir Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham, Jed Souter, Dan Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma, Erick Tuttle, Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon. Indatacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture. ACM, 2017. 16", + "[28] Ehsan Kamalloo, Nouha Dziri, Charles L. A. Clarke, and Davood Rafiei. Evaluating open-domain question answering in the era of large language models, 2023. 4, 23", + "[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes, 2020. 2, 8, 18, 19", + "[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 16", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 24" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9599", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual genome: Connecting language and vision using crowdsourced dense image annotations, 2016. 3, 5, 15", + "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2023. 24", + "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 3", + "[35] Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A Simple and Performant Baseline for Vision and Language. ArXiv, abs/1908.03557, 2019. 8", + "[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 5, 15", + "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3", + "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 6", + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 3, 4", + "[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2023. 2, 5, 15", + "[41] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. ArXiv, abs/2206.08916, 2022. 1, 3", + "[42] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023. 3", + "[43] Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. Faithful chain-of-thought reasoning, 2023. 1, 3", + "[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 1", + "[45] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6, 15", + "[46] Jingbiao Mei, Jinghong Chen, Weizhe Lin, Bill Byrne, and Marcus Tomalin. Improving hateful memes detection via learning hatefulness-aware embedding space through retrieval-guided contrastive learning, 2023. 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection, 2023. 4", + "[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5, 15", + "[49] OpenAI. Gpt-4 technical report, 2023. 1, 3", + "[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world, 2023. 3", + "[51] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 3", + "[52] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 5, 14, 15", + "[53] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718, 2022. 1", + "[54] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face, 2023. 3", + "[55] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 5, 15", + "[56] Otilia Stretcu, Edward Vendrow, Kenji Hata, Krishnamurthy Viswanathan, Vittorio Ferrari, Susan Tavakkol, Wenlei Zhou, Aditya Avinash, Emming Luo, Neil Gordon Alldrin, et al. Agile modeling: From concept to classifier in minutes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22323-22334, 2023. 8", + "[57] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation, 2023. 1", + "[58] Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. UL2: Unifying Language Learning Paradigms, 2022. 15", + "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian," + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9600", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models, 2023. 3", + "[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1, 3", + "[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 3", + "[62] Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. Scott: Self-consistent chain-of-thought distillation, 2023. 3", + "[63] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 1, 3, 7", + "[64] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5", + "[65] Yaqing Wang, Jialin Wu, Tanmaya Dabral, Jiageng Zhang, Geoff Brown, Chun-Ta Lu, Frederick Liu, Yi Liang, Bo Pang, Michael Bendersky, and Radu Soricut. Non-intrusive adaptation: Input-centric parameter-efficient fine-tuning for versatile multimodal modeling, 2023. 2, 5, 7", + "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2022. 3", + "[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation. 2023. 24", + "[68] Ge Zhang Yao Fu Wenhao Huang Huan Sun Yu Su Wenhu Chen Xiang Yue, Xingwei Qu. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 3", + "[69] Yiheng Xu, Hongjin Su, Chen Xing, Boyu Mi, Qian Liu, Weijia Shi, Binyuan Hui, Fan Zhou, Yitao Liu, Tianbao Xie, Zhoujun Cheng, Siheng Zhao, Lingpeng Kong, Bailin Wang, Caiming Xiong, and Tao Yu. Lemur: Harmonizing natural language and code for language agents, 2023. 1", + "[70] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081–3089, 2022. 3", + "[71] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn" + ], + "bbox": [ + 78, + 90, + 472, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9, 2023. 1", + "[72] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 3", + "[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 24", + "[74] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration, 2023. 1, 3, 6", + "[75] Xi Ye and Greg Durrett. The unreliability of explanations in few-shot prompting for textual reasoning. In Advances in Neural Information Processing Systems, pages 30378-30392. Curran Associates, Inc., 2022. 7", + "[76] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3", + "[77] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 15", + "[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding, 2022. 3", + "[79] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 3, 6", + "[80] Xuekai Zhu, Biqing Qi, Kaiyan Zhang, Xingwei Long, and Bowen Zhou. Pad: Program-aided distillation specializes large models in reasoning, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 609 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_model.json b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..093afcdfef5fce14ab8debcb00ba55d9798c8149 --- /dev/null +++ b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_model.json @@ -0,0 +1,2457 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.131, + 0.867, + 0.177 + ], + "angle": 0, + "content": "Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.202, + 0.808, + 0.274 + ], + "angle": 0, + "content": "Yushi Hu1,2* Otilia Strencu1 Chun-Ta Lu1 Krishnamurthy Viswanathan1 Kenji Hata1 Enming Luo1 Ranjay Krishna2 Ariel Fuxman1 1Google Research 2University of Washington https://visual-program-distillation.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.308, + 0.314, + 0.324 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.474, + 0.775 + ], + "angle": 0, + "content": "Solving complex visual tasks such as \"Who invented the musical instrument on the right?\" involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs. Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.795, + 0.21, + 0.81 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.472, + 0.88 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have become the pretrained backbone for many computer vision tasks [2, 4, 7, 9-11, 34, 39, 41, 60, 63, 74, 79]. Yet, all these models still fall short of solving numerous visual reasoning tasks expected" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.294, + 0.891, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.519, + 0.895, + 0.575 + ], + "angle": 0, + "content": "Figure 1. We introduce Visual Program Distillation (VPD), a training framework which leverages LLM-generated programs that make calls to specialist models and tools to distill cross-modal reasoning abilities and specialist skills into multimodal models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.793 + ], + "angle": 0, + "content": "of competent vision models. Even state-of-the-art (SOTA) proprietary vision-language models such as GPT-4V [49] do not perform well on tasks that involve counting and spatial reasoning [71]. They find it difficult to count (TallyQA [1]), to compositionally reason (GQA [26]), and to reason with external knowledge (OK-VQA [44], A-OKVQA [53]). Many of these tasks require VLMs to conduct compositional reasoning, which still remains an unsolved challenge. For instance, answering the question \"Who invented the musical instrument on the right?\" involves a composition of skills: identifying objects, applying spatial reasoning to locate the one on the right, recognizing the musical instrument, and accessing prior knowledge to retrieve the inventor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.897, + 0.903 + ], + "angle": 0, + "content": "In contrast, large language models (LLMs) have demonstrated remarkable performance at generating code that solves complex and compositional tasks [3, 8, 43, 49, 69]. Several recent papers [16, 20, 25, 57] capitalize on this by prompting LLMs to generate programs where each step corresponds to a reasoning step. The programs invoke specialized \"tools\" (or specialized vision models) to explicitly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.355, + 0.9 + ], + "angle": 0, + "content": "*Work done while interning at Google Research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.891, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.338, + 0.897, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Overview of Visual Program Distillation (VPD). VPD uses an LLM and specialized vision tools to generate faithful chain-of-thought (CoT) training data for vision-language models (VLMs). Given a multimodal input, our 4-step data synthesis pipeline generates a CoT that answers the query. In the example above, our synthesized CoT contains a series of reasoning steps: find all buses, check if each bus is yellow, and aggregate the count into a final answer. The CoT also contains the grounding information given by object detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.481 + ], + "angle": 0, + "content": "execute each reasoning step. For the question above, the program would call an \"object detector\" tool to identify and isolate all the objects, a \"fine-grained object classification\" tool to recognize the musical instrument, and a \"knowledge-based question answering\" tool to retrieve its inventor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.471, + 0.631 + ], + "angle": 0, + "content": "Although innovative, generating explicit programs is computationally expensive in practice, prone to errors, and still underperforms end-to-end models. Programs require loading and executing multiple tools, leading to high latency and computational cost. Moreover, generated programs may omit necessary steps or include spurious ones. Even when the program is correct, vision model invocations can produce incorrect outputs, from which the overall program cannot recover. Unfortunately, empirical results show that visual programs still fall short of end-to-end fine-tuned models [16]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.471, + 0.797 + ], + "angle": 0, + "content": "Another line of work is visual instruction tuning [39], which tries to distill the instruction-following ability of LLMs into VLMs. They prompt LLMs with image captions and bounding box annotations in order to generate queries and answers that are used to fine-tune VLMs [7, 23, 39, 65]. However, this approach has important limitations: image captions can miss fine-grained visual information, and LLM are prone to producing inconsistent outputs for custom vision representations like bounding boxes [18]. As a result, existing instruction-tuned VLMs still struggle with tasks that require complex visual reasoning [7, 14, 38, 40]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this work, we present Visual Program Distillation (VPD), a novel distillation method that induces complex cross-modal reasoning capabilities into vision-language models (Fig. 2). As the name suggests, VPD combines two key insights to deliver a training paradigm that surpasses the sum of its parts: It relies on (1) advancements in visual programs that use tools [20] and (2) the recent breakthroughs in dis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.6 + ], + "angle": 0, + "content": "tillation through chain-of-thought reasoning [21]. Given a labeled training dataset of complex visual tasks, VPD generates a correct program, and then distills its reasoning steps into vision-language models. To avoid using programs that give the wrong answer, VPD prompts an LLM to generate multiple candidate programs, and executes every one of them. When labeled data is available, it then filters for programs that produce the correct answer upon execution. Therefore, our programs comprise multiple vision tools, are executable, and yield the correct answer when executed. Next, VPD rewrites the correct programs as natural language chain-of-thought instructions and uses step-by-step distillation [21] to inject the reasoning abilities into VLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.895, + 0.781 + ], + "angle": 0, + "content": "Our best instruction-tuned model, PaLI-X-VPD (55B), sets a new SOTA result on 8 classical VQA tasks and 2 zero-shot multimodal benchmarks. Our models even outperform the recent SOTA established by PaLI-X [10]. Importantly, we conduct a quality evaluation with human raters which shows that PaLI-X-VPD generates more consistent and faithful rationales compared to its counterpart trained using instruction-tuning data. In addition, we experiment with PaLI-3 (5B), and show that VPD also improves the performance of smaller-scale models. Finally, experiments on Hateful Memes [29] show that VPD is also helpful for adapting to new tasks, even when no labels are available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.795, + 0.637, + 0.81 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.894, + 0.88 + ], + "angle": 0, + "content": "VPD is a general method for improving any vision-language model, and includes automatic program generation and training with chain-of-thought data as steps of the proposed framework. We discuss each of these research areas." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Vision-language models (VLMs). Most recent generative" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "9591" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "VLMs share a common structure: a pre-trained visual encoder, a pre-trained LLM, and a connector between the two modalities [2, 4, 9, 34, 39, 41, 60, 63, 74, 79]. The models are trained on large-scale image-text pairs from various tasks to adapt to both modalities. They are also tuned on LLM-generated visual instructions to enable models to follow versatile instructions from users [39]. Some models also include bounding boxes in the pre-training data to improve the VLM on visually grounded tasks [6, 7, 10, 41, 50, 61, 78]. The bounding boxes are usually retrieved from COCO [37] and Visual Genome [32]. Different from prior work, our method does not rely on provided dense annotations. We use LLM-generated code and specialized vision tools to generate our own visual instruction-tuning data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Visual programming and agents. With the advancement of large language models (LLMs) [3, 5, 13, 49, 59], recent work has started using LLMs as an interface to solving complex reasoning tasks with tools [12, 42, 51, 54, 76], and also as an agent for vision tasks [23, 25, 70, 72]. From this line of work, the most relevant to us are VisProg [20] and ViperGPT [16], which leverage LLMs to generate executable programs with a sequence of invocations to specialized vision tools. They achieve SOTA zero-shot performance on various vision tasks, while being versatile and interpretable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.473, + 0.673 + ], + "angle": 0, + "content": "Training and inference with chain-of-thought. Chain-of-Thought (CoT) [66] has become a popular approach to improving LLM performance. Recent work such as Program-of-Thoughts (PoT) [8] and Faithful CoT [43] further improve this framework by splitting inference into two steps: first generate a program with LLM, then execute the program. This approach achieves better accuracy and reduces hallucinations in the reasoning steps. Moreover, CoT is also used to train language models. Distill step-by-step [21], PaD [80], and SCOTT [62] train smaller language models with CoT generated by LLMs, showing that this can improve model performance and reasoning consistency. Mammoth [68] trains an LLM with a hybrid of CoT and PoT rationales and achieves a SOTA model for math problems." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.39, + 0.698 + ], + "angle": 0, + "content": "3. Visual Program Distillation (VPD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We introduce VPD, a general model-agnostic framework that distills the reasoning power of LLM-generated programs together with the low-level image understanding abilities of vision tools into a single vision-language model (Fig. 2). At its core, VPD consists of two major steps:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.779, + 0.472, + 0.853 + ], + "angle": 0, + "content": "1. Program generation and verification: Given a textual query \\( q \\) and a visual input \\( i \\), VPD first generates a program that solves the query by making use of vision modules and tools, then converts the program execution trace into a chain-of-thought \\( c \\) (\\$3.1)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.853, + 0.473, + 0.899 + ], + "angle": 0, + "content": "2. Distilling step-by-step: The visual input \\( i \\), textual query \\( q \\), and chain-of-thought \\( c \\) produced by the previous step are distilled into a vision-language model (VLM) (§3.2)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.779, + 0.473, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.816, + 0.108 + ], + "angle": 0, + "content": "3.1. Program generation and verification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.894, + 0.192 + ], + "angle": 0, + "content": "Our training data synthesis pipeline is illustrated in the blue boxes of Fig. 2, and consists of four stages. Given a sample \\((i,q,y)\\) consisting of a visual input \\(i\\) and a textual query \\(q\\) about its content and, when available, its ground truth answer \\(y\\), we perform the following sequence of steps:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.194, + 0.894, + 0.239 + ], + "angle": 0, + "content": "1. Program generation with LLM: Given \\( q \\), we generate a list of \\( k \\) candidate programs \\( \\pi(q) = \\{z_1, z_2, \\dots, z_k\\} \\), where \\( \\pi \\) represents a program generation function." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.24, + 0.895, + 0.344 + ], + "angle": 0, + "content": "2. Program execution with vision modules: We execute each program \\( z_{i} \\) with an execution engine \\( \\phi \\) to obtain its final result \\( \\phi (i,z_i) = \\hat{y}_i \\). However, during program execution, we maintain the execution trace \\( t_i \\) recording all the intermediate function calls and outputs. At the end of this step, we produce a list of programs, results, and execution traces \\( \\{(z_1,\\hat{y}_1,t_1),\\dots,(z_k,\\hat{y}_k,t_k)\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.344, + 0.892, + 0.387 + ], + "angle": 0, + "content": "3. Program filtering: Among the \\(k\\) candidate programs from the previous step, we keep a single tuple \\((z,\\hat{y},t)\\) with correct answer." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.389, + 0.894, + 0.432 + ], + "angle": 0, + "content": "4. Converting program execution traces into chain-of-thought rationales: We rewrite \\( t \\) into a CoT \\( c \\) using an LLM." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.194, + 0.895, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.436, + 0.824, + 0.452 + ], + "angle": 0, + "content": "We now discuss in detail each of the steps above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Program generation. We adopt a similar approach with recent work [16, 20] in our program generation step, and use PaLM-2 [3] as LLM to generate candidate programs for a given query \\( q \\). We prompt PaLM-2 with the same kind of text prompt as used by ViperGPT, which contains a detailed description of the available vision modules, followed by the query \\( q \\) (prompt in Appendix §G). The LLM is expected to directly output a Python function definition that will be executed in the following steps. However, in our experiments, we find that the success rate of top-\\( k \\) programs is much higher than the top-1 program. Therefore, in contrast to prior work, which only samples one program \\( z \\), when the ground truth answer \\( y \\) is available, we set a temperature \\( T \\) for LLM decoding and sample a list of top-\\( k \\) candidate programs \\( \\{z_1, z_2, \\dots, z_k\\} \\) from the LLM. Then, we filter out one correct program \\( z \\) in later steps. As shown in our ablation in §4.2, this becomes crucial to our performance gain. We use \\( T = 0.5 \\) and \\( k = 5 \\) for all our experiments. For unlabeled data, \\( k = 1 \\) and the filtering step can be skipped." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Program execution with vision modules. We use the same execution engine \\(\\phi\\) as ViperGPT [16]. An LLM-generated program \\(z_{i}\\) is a Python function that takes the visual input \\(i\\) as input. While the LLM outputs the program as a sequence of text, the execution engine \\(\\phi\\) is able to interpret this as a Python program and execute it, to obtain its return result \\(\\hat{y}_{i}\\). Additionally, \\(\\phi\\) also records the execution trace \\(t_{i}\\) of the program \\(z_{i}\\), which keeps a record of all the vision module calls, their inputs, and outputs. We use the following tools for the program: PaLI-X [10] for simple visual queries, PaLI-X" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.138 + ], + "angle": 0, + "content": "detection (distilled from OWLv2 [47]) for object detection; Google Cloud Depth API [17] for depth estimation, and PaLM-2 [3] for external knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.142, + 0.473, + 0.517 + ], + "angle": 0, + "content": "Program filtering. As we discussed in §1, visual programs are error-prone for a variety of reasons. The program might be wrong, and the execution process can introduce additional errors. To overcome these issues, we employ a program filtering step. We start by sampling top-\\(k\\) programs and attempt to execute each one. During this process, any program that fails execution is instantly discarded. We further filter the remaining programs, the strategy depending on the availability of labeled data. For tasks where human labels are available, we select a single program per input sample whose answer is correct—that is, when the program output \\(\\hat{y}_i\\) matches the human label \\(y\\). In this case our visual program pipeline acts like a CoT rationale annotator. One potential problem is that some task answers are ambiguous. For example, for a question like \"Where are the horses?\", the answers \"mountain\" and \"mountains\" are both correct. We adopt the method in [28] and use an LLM to determine if the program output is correct (details in §G). If there is more than one correct program per question, we select the top scoring one, according to the scores provided by the program-generating LLM \\(\\phi\\). If no programs pass the test, the rated sample is not wasted—we simply use the correct answer \\(y\\) as supervision in our fine-tuning stage, without an associated CoT. When no human-rated answer is available, we directly use the top scoring executable program1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Converting program execution traces into chain-of-thought rationales. After the filtering step, for each visual input \\(i\\) and query \\(q\\), we will have selected at most one program \\(z\\) together with its execution result \\(\\hat{y}\\) and trace \\(t\\). Since most existing VLMs have been pre-trained with text in natural language and not code, we use an LLM to rewrite the execution trace \\(t\\) into a natural language CoT \\(c\\) for our VLM distillation. Some examples are shown in §B. Concretely, similar to prior work [23, 24], we hand-craft 20 examples of how an input \\((q, z, t)\\) can be converted into a CoT, and use these as few-shots for prompting PaLM-2 [3], which performs in-context learning and generates CoTs for new examples. We include a concrete example in §G." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.283, + 0.741 + ], + "angle": 0, + "content": "3.2. Distilling step-by-step" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.837 + ], + "angle": 0, + "content": "In this step, we fine-tune a backbone VLM with the training data generated in §3.1, distilling the knowledge and reasoning steps of our generated programs into a single end-to-end model. We do so in a multitask fashion, where the VLM is simultaneously fine-tuned on data synthesized for multiple types tasks (e.g., free-form VQA, multiple choice)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Let \\( f \\) represent our VLM model. While the same VLM architecture could solve all these tasks, it needs to be prompted" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.091, + 0.896, + 0.227 + ], + "angle": 0, + "content": "differently to adapt to the task. Following prior work [39], we manually design instructions for each task. For example, for free-form VQA queries, the instruction is \"Answer with a single word or phrase\", while for multiple-choice queries we use \"Answer with the option letter from the given choices directly\". During fine-tuning, we combine the training samples generated for all tasks into a single dataset, so to account for the different types of tasks, we augment each sample with its corresponding task-specific prompt \\( p \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Using these instructions along with the data generated in §3.1, we put together the training dataset \\(\\mathcal{D} = \\{(i_j,q_j,\\hat{y}_j,c_j,p_j)\\}_{j = 1}^N\\), where \\(N\\) is the total number of samples, \\(i_j\\) is the visual input, \\(q_{j}\\) is the textual query, \\(\\hat{y}_i\\) is the visual program output2, and \\(c_{j}\\) is the CoT rationale." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.895, + 0.468 + ], + "angle": 0, + "content": "We train \\( f \\) to minimize a loss for predicting both the label and the rationale. As shown in the red box of Fig. 2, similar to [21], we treat predicting the output label \\( \\hat{y}_j \\) and the rationale \\( c_j \\) as two separate optimization goals. However, since VLMs are open-ended text generation models, they need additional prompting to indicate whether we want a short answer or a long answer that includes a rationale. Therefore, we append the suffix \\( s_c = \\) \"Explain the rationale to answer the question\" at the end of the prompt for generating a CoT, and use the task instruction \\( p_j \\) for short answers. Our optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.481, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\mathcal {L} _ {\\text {r a t i o n a l e}} (1) \\\\ = \\sum_ {j = 1} ^ {N} \\ell \\left(f \\left(i _ {j}, q _ {j}, p _ {j}\\right), \\hat {y} _ {j}\\right) + \\ell \\left(f \\left(i _ {j}, q _ {j}, s _ {c}\\right), c _ {j}\\right) (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.895, + 0.706 + ], + "angle": 0, + "content": "Here \\(\\ell\\) is the cross entropy-loss normalized by sequence length to ensure that the labels (typically short) and rationales (typically long) have similar weights. \\(\\mathcal{L}_{\\text{rationale}}\\) both teaches the VLM to generate faithful reasoning steps similar to program execution traces, and carries more information beyond the label \\(\\hat{y}\\) that also helps the VLM in better predicting the label. During test time, the rationale generation is not required. We can adjust the task instruction to directly get the short output label using \\(p\\), and using \\(s_c\\) to get the human-interpretable reasoning steps if needed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.634, + 0.737 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.746, + 0.894, + 0.866 + ], + "angle": 0, + "content": "In this section, we demonstrate VPD's effectiveness by using it to train a generalist VLM. We attempt this for two VLMs with different scales, PaLI-3 (5B) [11] and PaLI-X (55B) [10]. Detailed experimental setups are given in §4.1. Qualitatively, models fine-tuned with VPD exhibit the ability to reason step-by-step like a program, as illustrated in Fig. 3, and supported by human evaluation results (§4.3). Quantitative results show that our PaLI-3-VPD and PaLI-X-VPD" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2When labeled data is available, this is equivalent to the ground truth label \\(y\\) due to our filtering strategy." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Studying ways to approximate program correctness strategies for unlabeled data is an interesting direction for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9593" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.079, + 0.449, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Outputs of PaLI-X trained with VPD. Training with execution traces of filtered programs improves the model's ability to count, understand spatial relations, and reason compositionally. The images are from Visual Genome [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.473, + 0.435 + ], + "angle": 0, + "content": "achieve new SOTA on a broad range of tasks, on both generalist and per-task fine-tuning settings (§4.2). We also conduct a detailed analysis of the source of performance gain (§4.2). Finally, we conduct human evaluation on the quality of the rationales generated by our models, and compare it with rationales generated by an instruction-tuned model trained without VPD (§4.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.446, + 0.266, + 0.463 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Backbone models. We use two state-of-the-art VLMs, PaLI-3 (5B) [11] and PaLI-X (55B) [10] as our base models. Both take images and text as input, and generate text as output. For simplicity, we further refer to them as \"PaLI model\" when we discuss steps performed with each of them individually." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Data for Generalist Models. We fine-tune the pre-trained PaLI model on two types of datasets to make it a generalist VLM—that is, a model that performs relatively well on any task without further training on that specific task:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.736 + ], + "angle": 0, + "content": "(1) Multimodal Instruction-Tuning (MMIT) tasks, created in the spirit of Self-Instruct [64]. An LLM is prompted with image captions, and generates task inputs and the desired outputs about the corresponding image. Details of MMIT tasks are covered in [65]. Note that these tasks cover a wide variety of common use cases, but do not include the specific in-domain tasks used in this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.471, + 0.902 + ], + "angle": 0, + "content": "(2) Academic task-oriented VQA tasks. Since image captions contain only a coarse description of visual information in the image, and may miss details that are important for solving the task. Additionally, LLM are likely to hallucinate during this data curation process. To further boost the accuracy of our VLMs, we also fine-tune the PaLI models with academic task-oriented VQA tasks. The data mixture covers subsets of a wide variety of VQA tasks, including general VQA (VQAv2 [19]), optical character recognition (OCRVQA [48]), compositional questions and reasoning (GQA [26]), counting (TallyQA [1]), and VQA that involves" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.631, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "external knowledge (OK-VQA [45] and A-OKVQA [52]). The tasks contain a textual query and a short expected label. We use the pipeline in §3.1 to synthesize CoT reasoning steps for these labels, and tune the PaLI model with these the loss in Equation 1. Notice that sometimes the pipeline fails to find a program that generates the correct answer. In that case, we set \\(\\mathcal{L}_{\\text{rationale}}\\) to 0 and only keep \\(\\mathcal{L}_{\\text{label}}\\). §4.2 shows how many programs are kept after the filtering stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Data for specialist models. While fine-tuning the generalist model with VPD, we only use a subset of each task's training data. To evaluate our model's ability on each individual task, we continue fine-tuning PaLI-3-VPD and PaLI-X-VPD on each individual task on the training splits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.895, + 0.422 + ], + "angle": 0, + "content": "Training setup. Both PaLI-3 and PaLI-X follow an encoder-decoder architecture, where images are encoded into visual tokens via a visual encoder and then passed into a UL2 model. Due to resource constraint, we use LoRA [22] to fine-tune both PaLI models. Specifically, we add LoRA weights on each linear layer in the attention blocks and the MLP blocks for both the encoder and decoder in the UL2 transformer. More training details are in §C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.897, + 0.684 + ], + "angle": 0, + "content": "Evaluation setup. We evaluate our models on a wide range of tasks, including various VQA tasks and recent zero-shot VLM benchmarks. Noted that A-OKVQA contains two kinds of questions, multiple-choice (MC) and direct answer (DA). We report results on both. TallyQA [1] contains complex counting questions that involve object relationships, attribute identification, and reasoning to get the correct answer. It contains two evaluation partitions, simple and complex. TextVQA [55] focuses reading texts. We include it to evaluate our models on zero-shot tasks. In addition to VQA tasks, we also test our models on two popular VLM benchmarks. POPE [36] focuses on VLM hallucination, containing binary questions of whether or not an object exists in the image. MMBench [40] is a robust and comprehensive VLM benchmark testing a range of fine-grained abilities (e.g., object localization, attribute recognition, spatial relationship). Details of the evaluation sets and metrics are in §C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Baselines. We refer to the two PaLI models fine-tuned with VPD as PaLI-3-VPD and PaLI-X-VPD, and compare them with various baselines. To evaluate the effectiveness of VPD, we experiment with removing the synthesized CoT from our data mixture, and train the PaLI models with the exact same hyper-parameters and steps. We call these models as PaLI-3 Instruct and PaLI-X Instruct. For a fair comparison, these models are also trained with the same supervised loss for predicting the ground truth answers, on the same images and textual queries as our VPD variant. Moreover, we also compare with the most recent SOTA vision-language models. These VLMs are initialized with pre-trained visual encoders and LLMs, and then trained with image-text pairs, LLM-generated data, and academic tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9594" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.078, + 0.899, + 0.292 + ], + "angle": 0, + "content": "
Generalist ModelsVQAv2GQAOK-VQAA-OKVQATallyQATextVQAPOPEMMB
MCDASimp.Comp.
Prior generalist VLMs
Flamingo (80B) [2]82.0-57.8*----57.1--
MiniGPT-4 (Vicuna-13B) [79]-43.5-67.2-----42.3
InstructBLIP (Vicuna-13B) [14]-49.5*---75.2*57.5*50.7*78.944.0
Shikra (Vicuna-13B) [7]77.4-47.2-----84.758.8
Qwen-VL (9.7B) [4]78.859.358.6--82.6*65.8*63.8-38.2
Qwen-VL-Chat (9.7B) [4]78.257.556.6--81.1*64.0*61.5-60.6
mPLUG-Owl2 (8.2B) [74]79.456.157.7----54.3*86.264.5
LLaVA-1.5 (Vicuna-13B) [38]80.063.3---76.9*65.4*61.3*85.967.7
Our instruction-tuned baselines
PaLI-3-Instruct (5B)79.959.756.778.357.681.970.463.3*87.768.6
PaLI-X-Instruct (55B)83.663.364.384.161.585.575.465.0*88.975.0
Our visual program distillation models
PaLI-3-VPD (5B)80.461.357.578.556.583.170.963.7*88.669.0
PaLI-X-VPD (55B)83.964.964.684.562.786.276.665.4*88.876.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.293, + 0.894, + 0.349 + ], + "angle": 0, + "content": "Table 1. Comparison with SOTA generalist VLMs. PaLI-X-VPD outperforms all prior models. VPD improves performance on 8/9 tasks for both PaLI-3 and PaLI-X, and is particularly effective on counting questions (TallyQA), compositional questions (GQA), and the comprehensive benchmark (MMBench). Underline indicates when PaLI-3-VPD outperforms the Instruct version. * marks tasks unseen during training. POPE and MMBench are zero-shot benchmarks for all models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.364, + 0.267, + 0.38 + ], + "angle": 0, + "content": "4.2. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.387, + 0.474, + 0.685 + ], + "angle": 0, + "content": "Generalist model. Table 1 compares VPD with the baselines discussed in §4.1. We infer all answers by open-ended generation with the prompt \"Answer with a single word or phrase,\" using greedy decoding without any constraint on the model's output space. For multiple-choice questions, we run inference with the prompt \"Answer with the option letter from the given choices directly.\" and generate the option letter. As Table 1 shows, PaLI-X-VPD sets the new state-of-the-art on all benchmarks. Compared with prior generalist VLMs, it achieves significant improvement on MMBench (+8.5), TallyQA complex (+9.8), and A-OKVQA(+9.5 compared with specialist SOTA[14]). PaLI-3-VPD, despite its small architecture size (5B), outperforms all prior models on VQAv2, TallyQA, POPE, and MMBench, including much larger models that use Vicuna-13B as backbones [7, 14, 38]. However, its performance on knowledge-based VQA tasks does not outperform the prior SOTA. Our hypothesis for this is that its language model (3B UL2) is too small to contain all the necessary knowledge needed for correctly solving these tasks." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.686, + 0.472, + 0.821 + ], + "angle": 0, + "content": "When compared to their Instruct variants, both PaLI-3-VPD and PaLI-X-VPD outperform on 11/12 tasks. Specifically, for PaLI-X, VPD obtains a \\(+1.6\\) improvement on GQA (which is heavily focused on compositional questions, spatial relationship, and localization), \\(+1.2\\) on TallyQA for complex questions, and \\(+1.2\\) on MM-Bench. These results suggest that VPD is a more effective method for creating instruction-tuning data which enables VLMs to improve their visual reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Per-Task Fine-Tuning (Specialist). The results for the specialist models are shown in Table 2. PaLI-X-VPD sets a new SOTA on all benchmarks. Note how the specialist models tend to have higher scores than the generalist one. We propose several hypotheses for this performance gain:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.895, + 0.59 + ], + "angle": 0, + "content": "(1) As shown in Table 2, the score gap tends to be larger on free-form VQA tasks. This may be due the fact that human annotations on these datasets have ambiguities. For example, for the question \"Who is looking up?\", GQA [26] labels are \"man\" or \"woman\" while OK-VQA [45] have more detailed labels, for example, \"worker\" or \"cook\". Per-task fine-tuning alleviates this annotation ambiguity and lets the model focus on the annotation style of that task. For multiple-choice and counting tasks, the answer has less ambiguity, and the score gaps are much smaller. (2) The performance gap between PaLI-3 (5B) specialist and generalist is larger than that of PaLI-X. We hypothesize that this shows models with larger scale has more multi-task capacity; (3) Per-task fine-tuning adds more in-domain training data, which generally improves performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.897, + 0.775 + ], + "angle": 0, + "content": "Analysis: Sampling multiple programs is key to good data generation. Fig. 5 shows the success rate of finding at least one program that passes the filtering stage, when the LLM generates the top-1 or top-5 programs, respectively. There is a dramatic increase in success rate from 1 program to 5: \\(+45\\%\\) on GQA and A-OKVQA, \\(+33\\%\\) on OK-VQA, and \\(+10\\%\\) on TallyQA. This design choice greatly improves our data synthesis efficiency, and, as a consequence, adding more CoT data that requires complex reasoning in our training set. We also conduct an analysis in \\(\\S E\\) to compare the performance of VPD models with that of the visual programs they are distilled from." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.783, + 0.783, + 0.798 + ], + "angle": 0, + "content": "4.3. Human evaluation on rationales" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.894, + 0.867 + ], + "angle": 0, + "content": "In this section, we focus on the quality of model outputs. We performed this analysis using human annotators, who are asked to evaluate both the correctness of the final answer and the quality of the rationale behind it." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Models. We compare PaLI-X-VPD with PaLI-X Instruct. Among possible baselines, we chose PaLI-X" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9595" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.089, + 0.837, + 0.263 + ], + "angle": 0, + "content": "
Specialist ModelsGQA test-devOK-VQA valA-OKVQATallyQA
Multi-choice valtestDirect Answer valtestSimple testComplex test
InstructBLIP (Vicuna-7B) [14]-62.175.773.464.062.1--
PaLI-3 (5B) [11]-60.1----83.370.5
PaLI (17B) [9]-64.5----81.770.9
CogVLM (17B) [63]65.264.7------
PaLI-X (55B) [10]-66.1----86.075.6
PaLI-3-VPD (5B) generalist61.357.578.5-56.5-83.170.9
PaLI-3-VPD (5B) specialist64.760.379.776.565.563.683.370.8
PaLI-X-VPD (55B) generalist64.964.684.5-62.7-86.276.6
PaLI-X-VPD (55B) specialist67.366.885.280.471.168.286.276.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.267, + 0.844, + 0.281 + ], + "angle": 0, + "content": "Table 2. Comparison of per-task fine-tuning results of specialist models. PaLI-X-VPD sets a new SOTA for all the tasks." + }, + { + "type": "image", + "bbox": [ + 0.159, + 0.288, + 0.332, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.288, + 0.514, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.288, + 0.804, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Human evaluation results assessing answer and rational quality for PaLI-X Instruct and PaLI-X VPD: (a) percentage of answers that are correct and have explanations; (b) rationale factuality and consistency for the answers that contain an explanation; (c) preference between the two models, when aggregating across all samples (\"All\") and across those with correct answers (\"When correct\")." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.462, + 0.445, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.566, + 0.471, + 0.595 + ], + "angle": 0, + "content": "Figure 5. Success rate of top-1 program and top-5 programs on the training set during our data synthesis process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.472, + 0.701 + ], + "angle": 0, + "content": "Instruct because it is trained to generate long-form answers [65], which allows us to assess if it is the quality of PaLI-X-VPD's answers that the annotators prefer, rather than its length. Since PaLI-X Instruct is also instruction-tuned, it can be prompted to provide long answers to alleviate this confounder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Annotation protocol. We run inference with each of the two models on a combination of 600 samples from GQA (test-dev), A-OKVQA (val), and TallyQA (Simple and Complex) and record their answers and rationales. We then ask 3 human annotators to evaluate each model answer. We use prior work from natural language processing (NLP) as inspiration, and build upon it for selecting the evaluation criteria [75]. Given an image and a query, for each model-generated rationale, we ask human annotators to score the model answers along the following criteria: (1) correctness—is the final answer correct? (2) explainability—does the model explain its rationale for reaching the final answer? (3) factuality—is every step in the rationale factually cor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.895, + 0.604 + ], + "angle": 0, + "content": "rect (with respect to the image and external knowledge)? (4) consistency—does step and final answer logically follow from the previous ones? Note that a rationale may have the wrong answer while being consistent. We also conduct a side-by-side comparison, and ask annotators which of the two answers—the one provided by PaLI-X-VPD or by PaLI-X Instruct—they prefer in terms of quality of the answer and explanation. More details about the annotation protocol are in §D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Human evaluation results. The results of the evaluation are first averaged among the human raters per sample, then aggregated across samples. Our PaLI-X-VPD model far outperforms PaLI-X Instruct along all criteria. Fig. 4 (a) shows the correctness of the final answer provided by each model (i.e. accuracy), and the proportion of the samples where the model provides a CoT rationale to explain its answer. PaLI-X-VPD's gain in accuracy of \\(+16.7\\%\\) is even more impressive than in evaluations in §4.2 based on benchmark labels—this is because human annotators are better able to assess correctness for ambiguous questions with different possible interpretations (e.g., the model's answer \"In the living room.\" to the question \"Where is the couch located?\" was considered correct by annotators, even when the benchmark answer was \"On the right\"). Moreover, the explainability results confirm that our model is able to explain its own answer on \\(+24\\%\\) more samples than the instruction tuned model. Additionally, Fig. 4 (b) shows impressive rationale quality: among the samples where" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9596" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.15 + ], + "angle": 0, + "content": "an explanation is provided, PaLI-X-VPD's rationales are factual \\(87.2\\%\\) of times, and consistent \\(97.8\\%\\) of times, a gain of \\(+14.6\\%\\) and \\(+10\\%\\), respectively, when compared to PaLI-X Instruct." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.471, + 0.314 + ], + "angle": 0, + "content": "We also asked the annotators which of the two answers given by the two models they prefer. We aggregated these results in two ways: (1) across all samples, (2) across the samples where both models answered the question correctly. The results in Fig. 4 (c) show that PaLI-X-VPD is preferred on \\(25\\%\\) more samples than PaLI-X Instruct in the \"All\" case, and on \\(12\\%\\) more samples when both are correct. This suggests that even when PaLI-X-VPD makes mistakes, it still provides a better answer. Such examples are shown in \\(\\S D\\). These results confirm that a model fine-tuned with VPD leads to more faithful and consistent answers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.33, + 0.403, + 0.347 + ], + "angle": 0, + "content": "5. Experiments on content moderation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.471, + 0.519 + ], + "angle": 0, + "content": "Prior experiments focus on training generalist VLMs. Here, we explore the effectiveness of VPD at quickly adapting models to real-world applications from a different domain than the training data [e.g., 56]. We experiment on Hateful Memes [29], a content moderation dataset where the task is to classify if a meme contains hateful content. The target labels are \"yes\" or \"no\", and models are evaluated in terms of classification accuracy and AUC-ROC. We experiment with two settings: supervised, in which the models are trained on the provided training set with 8,500 labels, and unsupervised, in which no human labels are provided." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.526, + 0.442, + 0.657 + ], + "angle": 0, + "content": "
ModelAccAUC-ROC
Unsupervised / Zero-Shot Methods
Generated Programs (ours)69.770.1
Supervised Methods
VisualBERT [35]69.575.4
Flamingo (80B) [2]-86.6
Previous SOTA [46]78.886.7
PaLI-X-VPD (label-only FT)77.688.0
PaLI-X-VPD (specialist w/ CoT)80.889.2
Human [29]84.782.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Table 3. Results on Hateful Memes [29] seen test set. We improve SOTA for both zero-shot and supervised settings. Unsupervised PaLI-X-VPD outperforms strong supervised baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.275, + 0.73 + ], + "angle": 0, + "content": "5.1. Unsupervised setting" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.736, + 0.314, + 0.751 + ], + "angle": 0, + "content": "We experiment with three methods:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.751, + 0.47, + 0.84 + ], + "angle": 0, + "content": "1. Generated Program: The program generated by PaLM-2 [3] for solving this task (shown in §F) consists of following main steps: (1) get image description with PaLI-X [10]; (2) use an OCR tool to extract embedded texts; (3) given the image description and OCR texts, ask PaLM-2 to explain if this meme is hateful." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "2. PaLI-X-VPD (generalist): In a zero-shot setting, we directly prompt our PaLI-X-VPD (generalist) with: \"The text is \\(. Is this a hateful meme?\" We compute the probability of PaLI-X-VPD generating \"yes\"" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.751, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.829, + 0.106 + ], + "angle": 0, + "content": "or \"no\" and measure accuracy and AUC-ROC." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.182 + ], + "angle": 0, + "content": "3. PaLI-X-VPD (specialist with zero-shot CoT): We follow our VPD pipeline and convert the execution traces of the program into CoTs, and then fine-tune PaLI-X-VPD to output these CoTs. Since no groundtruth labels are available, no filtering is done during the process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.895, + 0.353 + ], + "angle": 0, + "content": "VPD significantly improves performance even when no labels are available. As shown in Table 3, PaLI-X-VPD (generalist) outperforms all other VLMs (61.4%). Interestingly, the generated programs themselves get much higher accuracy (69.7%) than on the previous datasets, perhaps because PaLM-2 is better suited at analyzing a meme than typical VQA datasets. Moreover, our PaLI-X-VPD (specialist) sets a new zero-shot SOTA on Hateful Memes, and even outperforms supervised VisualBERT [35]. Manual inspection of model outputs show a great deal of similarity with the generated code. See examples in §F." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.365, + 0.677, + 0.381 + ], + "angle": 0, + "content": "5.2. Supervised setting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.389, + 0.776, + 0.403 + ], + "angle": 0, + "content": "We fine-tune PaLI-X-VPD in two ways:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.404, + 0.894, + 0.449 + ], + "angle": 0, + "content": "1. Label-only fine-tuning: To establish a strong baseline, we first experiment with the traditional supervised setting, where we fine-tune the model to output \"yes\" or \"no\"." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.449, + 0.894, + 0.493 + ], + "angle": 0, + "content": "2. PaLI-X-VPD (specialist): We use the complete VPD pipeline to train this model. We select an execution trace that leads to the correct label, and use it to tune our VLM." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.404, + 0.894, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.894, + 0.589 + ], + "angle": 0, + "content": "Results. Our PaLI-X-VPD (specialist) sets a new SOTA for this task, with an accuracy of \\(80.8\\%\\) and AUC-ROC of \\(89.2\\%\\). It outperforms the label-only fine-tuning baseline and significantly improves the SOTA metrics, achieving nearly human-level accuracy, and superhuman AUC-ROC." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.606, + 0.863, + 0.622 + ], + "angle": 0, + "content": "6. Conclusion, limitations, and future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced VPD, a framework for distilling the reasoning abilities of LLMs along with the capabilities of vision tools into VLMs. VPD synthesizes training data for VLMs by generating programs that can leverage external tools. We used this technique to fine-tune some of the best existing VLMs (PaLI-3 and PaLI-X) and established new SOTA results on 8 classical VQA and 2 zero-shot multimodal benchmarks. According to human evaluations, VPD-tuned models provide more accurate answers and better rationales. Experiments on the Hateful Memes dataset show how VPD can also adapt models to new real-world domains, even when no labeled data is available, also establishing a new SOTA on this dataset. From our experiments we learnt that both a strength and a weakness of VPD is the quality of the generated programs. Better programs lead to bigger gains with VPD. We discuss multiple ideas on how to improve the program generation in Appendix H, including using additional tools, agents, and fact-checking for multimodal CoTs." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9597" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.169 + ], + "angle": 0, + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 1, 5, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.295 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.296, + 0.472, + 0.762 + ], + "angle": 0, + "content": "[3] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby Ambrose Slone,Daniel Smilkov David R. So Daniel Sohn Simon Tokumine,Dasha Valter,Vijay Vasudevan,Kiran Vodrahalli,Xuezhi Wang,Pidong Wang,Zirui Wang,Tao Wang John Wieting,Yuhuai Wu,Kelvin Xu,Yunhan Xu,Linting Xue,Pengcheng Yin,Jiahui Yu,Qiao ZhangSteven Zheng Ce ZhengWeikang ZhouDenny ZhouSlav Petrov and Yonghui Wu.Palm 2 technical report,2023.134813" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.763, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigptv2: large language model as a unified interface for vision-language multi-task learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.163, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.206, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[8] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.263, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[10] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, AJ Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali-x: On scaling up a multilingual vision and language model, 2023. 2, 3, 4, 5, 7, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[11] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, Daniel Salz, Xi Xiong, Daniel Vlasic, Filip Pavetic, Keran Rong, Tianli Yu, Daniel Keysers, Xiaohua Zhai, and Radu Soricut. Pali-3 vision language models: Smaller, faster, stronger, 2023. 1, 4, 5, 7, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[12] Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, et al. Binding language models in symbolic languages. arXiv preprint arXiv:2210.02875, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robin-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.473, + 0.23 + ], + "angle": 0, + "content": "son, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.472, + 0.286 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.473, + 0.479 + ], + "angle": 0, + "content": "[15] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaeldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.481, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[16] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 1, 2, 3, 20, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[17] Ruofei Du, Eric Turner, Maksym Dzitsiuk, Luca Prasso, Ivo Duarte, Jason Dourgarian, Joao Afonso, Jose Pascoal, Josh Gladstone, Nuno Cruces, Shahram Izadi, Adarsh Kowdle, Konstantine Tsotsos, and David Kim. DepthLab: Real-Time 3D Interaction With Depth Maps for Mobile Augmented Reality. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. ACM, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.472, + 0.689 + ], + "angle": 0, + "content": "[18] Weixi Feng, Wanrong Zhu, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. *ArXiv*, abs/2305.15393, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[20] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.885 + ], + "angle": 0, + "content": "[21] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[22] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[23] Yushi* Hu, Hang* Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[24] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A. Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2627-2643, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[25] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.894, + 0.4 + ], + "angle": 0, + "content": "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 5, 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[27] Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir Ghaemmaghami, Rajendra Gottipati, William Gulland, Robert Hagemann, C. Richard Ho, Doug Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander Kaplan, Harshit Khaitan, Daniel Killebrew, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon, James Law, Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana Maggiore, Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy Nix, Thomas Norrie, Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Matt Ross, Amir Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham, Jed Souter, Dan Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma, Erick Tuttle, Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon. Indatacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture. ACM, 2017. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[28] Ehsan Kamalloo, Nouha Dziri, Charles L. A. Clarke, and Davood Rafiei. Evaluating open-domain question answering in the era of large language models, 2023. 4, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes, 2020. 2, 8, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 24" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9599" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual genome: Connecting language and vision using crowdsourced dense image annotations, 2016. 3, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2023. 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.472, + 0.258 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[35] Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A Simple and Performant Baseline for Vision and Language. ArXiv, abs/1908.03557, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.472, + 0.495 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2023. 2, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[41] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. ArXiv, abs/2206.08916, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[42] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.472, + 0.706 + ], + "angle": 0, + "content": "[43] Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. Faithful chain-of-thought reasoning, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[45] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[46] Jingbiao Mei, Jinghong Chen, Weizhe Lin, Bill Byrne, and Marcus Tomalin. Improving hateful memes detection via learning hatefulness-aware embedding space through retrieval-guided contrastive learning, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "[47] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.791, + 0.177 + ], + "angle": 0, + "content": "[49] OpenAI. Gpt-4 technical report, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[51] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.276, + 0.893, + 0.33 + ], + "angle": 0, + "content": "[52] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 5, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.893, + 0.386 + ], + "angle": 0, + "content": "[53] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.388, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[54] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.893, + 0.497 + ], + "angle": 0, + "content": "[55] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.5, + 0.893, + 0.581 + ], + "angle": 0, + "content": "[56] Otilia Stretcu, Edward Vendrow, Kenji Hata, Krishnamurthy Viswanathan, Vittorio Ferrari, Susan Tavakkol, Wenlei Zhou, Aditya Avinash, Emming Luo, Neil Gordon Alldrin, et al. Agile modeling: From concept to classifier in minutes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22323-22334, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.583, + 0.893, + 0.638 + ], + "angle": 0, + "content": "[57] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.639, + 0.893, + 0.707 + ], + "angle": 0, + "content": "[58] Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. UL2: Unifying Language Learning Paradigms, 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.709, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9600" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.473, + 0.174 + ], + "angle": 0, + "content": "Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[62] Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. Scott: Self-consistent chain-of-thought distillation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[63] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[64] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[65] Yaqing Wang, Jialin Wu, Tanmaya Dabral, Jiageng Zhang, Geoff Brown, Chun-Ta Lu, Frederick Liu, Yi Liang, Bo Pang, Michael Bendersky, and Radu Soricut. Non-intrusive adaptation: Input-centric parameter-efficient fine-tuning for versatile multimodal modeling, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation. 2023. 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[68] Ge Zhang Yao Fu Wenhao Huang Huan Sun Yu Su Wenhu Chen Xiang Yue, Xingwei Qu. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[69] Yiheng Xu, Hongjin Su, Chen Xing, Boyu Mi, Qian Liu, Weijia Shi, Binyuan Hui, Fan Zhou, Yitao Liu, Tianbao Xie, Zhoujun Cheng, Siheng Zhao, Lingpeng Kong, Bailin Wang, Caiming Xiong, and Tao Yu. Lemur: Harmonizing natural language and code for language agents, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[70] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081–3089, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[71] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.473, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[72] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[74] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[75] Xi Ye and Greg Durrett. The unreliability of explanations in few-shot prompting for textual reasoning. In Advances in Neural Information Processing Systems, pages 30378-30392. Curran Associates, Inc., 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[76] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[77] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.458, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.515, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[79] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[80] Xuekai Zhu, Biqing Qi, Kaiyan Zhang, Xingwei Long, and Bowen Zhou. Pad: Program-aided distillation specializes large models in reasoning, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9601" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_origin.pdf b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23080a0ad891b427c017d8c190e7a58a36e47d49 --- /dev/null +++ b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/d3d264df-7a0f-49c2-be47-325ff84855b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9ce5074e13c8827ce0eba50dab42420a22d389e05aa90877f8018d447b5295e +size 2336346 diff --git a/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/full.md b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02c9b859b534383048ea94a4908639df22573ede --- /dev/null +++ b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/full.md @@ -0,0 +1,295 @@ +# Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models + +Yushi Hu1,2* Otilia Strencu1 Chun-Ta Lu1 Krishnamurthy Viswanathan1 Kenji Hata1 Enming Luo1 Ranjay Krishna2 Ariel Fuxman1 1Google Research 2University of Washington https://visual-program-distillation.github.io/ + +# Abstract + +Solving complex visual tasks such as "Who invented the musical instrument on the right?" involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs. Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data. + +# 1. Introduction + +Vision-language models (VLMs) have become the pretrained backbone for many computer vision tasks [2, 4, 7, 9-11, 34, 39, 41, 60, 63, 74, 79]. Yet, all these models still fall short of solving numerous visual reasoning tasks expected + +![](images/948f46a27311e660bc2833cdebe63714935e90d4853f13c32d2d80baf27e299c.jpg) +Figure 1. We introduce Visual Program Distillation (VPD), a training framework which leverages LLM-generated programs that make calls to specialist models and tools to distill cross-modal reasoning abilities and specialist skills into multimodal models. + +of competent vision models. Even state-of-the-art (SOTA) proprietary vision-language models such as GPT-4V [49] do not perform well on tasks that involve counting and spatial reasoning [71]. They find it difficult to count (TallyQA [1]), to compositionally reason (GQA [26]), and to reason with external knowledge (OK-VQA [44], A-OKVQA [53]). Many of these tasks require VLMs to conduct compositional reasoning, which still remains an unsolved challenge. For instance, answering the question "Who invented the musical instrument on the right?" involves a composition of skills: identifying objects, applying spatial reasoning to locate the one on the right, recognizing the musical instrument, and accessing prior knowledge to retrieve the inventor. + +In contrast, large language models (LLMs) have demonstrated remarkable performance at generating code that solves complex and compositional tasks [3, 8, 43, 49, 69]. Several recent papers [16, 20, 25, 57] capitalize on this by prompting LLMs to generate programs where each step corresponds to a reasoning step. The programs invoke specialized "tools" (or specialized vision models) to explicitly + +![](images/07033e3703646704ac7b1eba64c1ea5916cfc8cfbb8a6ec25a25a728543fef5a.jpg) +Figure 2. Overview of Visual Program Distillation (VPD). VPD uses an LLM and specialized vision tools to generate faithful chain-of-thought (CoT) training data for vision-language models (VLMs). Given a multimodal input, our 4-step data synthesis pipeline generates a CoT that answers the query. In the example above, our synthesized CoT contains a series of reasoning steps: find all buses, check if each bus is yellow, and aggregate the count into a final answer. The CoT also contains the grounding information given by object detection. + +execute each reasoning step. For the question above, the program would call an "object detector" tool to identify and isolate all the objects, a "fine-grained object classification" tool to recognize the musical instrument, and a "knowledge-based question answering" tool to retrieve its inventor. + +Although innovative, generating explicit programs is computationally expensive in practice, prone to errors, and still underperforms end-to-end models. Programs require loading and executing multiple tools, leading to high latency and computational cost. Moreover, generated programs may omit necessary steps or include spurious ones. Even when the program is correct, vision model invocations can produce incorrect outputs, from which the overall program cannot recover. Unfortunately, empirical results show that visual programs still fall short of end-to-end fine-tuned models [16]. + +Another line of work is visual instruction tuning [39], which tries to distill the instruction-following ability of LLMs into VLMs. They prompt LLMs with image captions and bounding box annotations in order to generate queries and answers that are used to fine-tune VLMs [7, 23, 39, 65]. However, this approach has important limitations: image captions can miss fine-grained visual information, and LLM are prone to producing inconsistent outputs for custom vision representations like bounding boxes [18]. As a result, existing instruction-tuned VLMs still struggle with tasks that require complex visual reasoning [7, 14, 38, 40]. + +In this work, we present Visual Program Distillation (VPD), a novel distillation method that induces complex cross-modal reasoning capabilities into vision-language models (Fig. 2). As the name suggests, VPD combines two key insights to deliver a training paradigm that surpasses the sum of its parts: It relies on (1) advancements in visual programs that use tools [20] and (2) the recent breakthroughs in dis + +tillation through chain-of-thought reasoning [21]. Given a labeled training dataset of complex visual tasks, VPD generates a correct program, and then distills its reasoning steps into vision-language models. To avoid using programs that give the wrong answer, VPD prompts an LLM to generate multiple candidate programs, and executes every one of them. When labeled data is available, it then filters for programs that produce the correct answer upon execution. Therefore, our programs comprise multiple vision tools, are executable, and yield the correct answer when executed. Next, VPD rewrites the correct programs as natural language chain-of-thought instructions and uses step-by-step distillation [21] to inject the reasoning abilities into VLMs. + +Our best instruction-tuned model, PaLI-X-VPD (55B), sets a new SOTA result on 8 classical VQA tasks and 2 zero-shot multimodal benchmarks. Our models even outperform the recent SOTA established by PaLI-X [10]. Importantly, we conduct a quality evaluation with human raters which shows that PaLI-X-VPD generates more consistent and faithful rationales compared to its counterpart trained using instruction-tuning data. In addition, we experiment with PaLI-3 (5B), and show that VPD also improves the performance of smaller-scale models. Finally, experiments on Hateful Memes [29] show that VPD is also helpful for adapting to new tasks, even when no labels are available. + +# 2. Related work + +VPD is a general method for improving any vision-language model, and includes automatic program generation and training with chain-of-thought data as steps of the proposed framework. We discuss each of these research areas. + +Vision-language models (VLMs). Most recent generative + +VLMs share a common structure: a pre-trained visual encoder, a pre-trained LLM, and a connector between the two modalities [2, 4, 9, 34, 39, 41, 60, 63, 74, 79]. The models are trained on large-scale image-text pairs from various tasks to adapt to both modalities. They are also tuned on LLM-generated visual instructions to enable models to follow versatile instructions from users [39]. Some models also include bounding boxes in the pre-training data to improve the VLM on visually grounded tasks [6, 7, 10, 41, 50, 61, 78]. The bounding boxes are usually retrieved from COCO [37] and Visual Genome [32]. Different from prior work, our method does not rely on provided dense annotations. We use LLM-generated code and specialized vision tools to generate our own visual instruction-tuning data. + +Visual programming and agents. With the advancement of large language models (LLMs) [3, 5, 13, 49, 59], recent work has started using LLMs as an interface to solving complex reasoning tasks with tools [12, 42, 51, 54, 76], and also as an agent for vision tasks [23, 25, 70, 72]. From this line of work, the most relevant to us are VisProg [20] and ViperGPT [16], which leverage LLMs to generate executable programs with a sequence of invocations to specialized vision tools. They achieve SOTA zero-shot performance on various vision tasks, while being versatile and interpretable. + +Training and inference with chain-of-thought. Chain-of-Thought (CoT) [66] has become a popular approach to improving LLM performance. Recent work such as Program-of-Thoughts (PoT) [8] and Faithful CoT [43] further improve this framework by splitting inference into two steps: first generate a program with LLM, then execute the program. This approach achieves better accuracy and reduces hallucinations in the reasoning steps. Moreover, CoT is also used to train language models. Distill step-by-step [21], PaD [80], and SCOTT [62] train smaller language models with CoT generated by LLMs, showing that this can improve model performance and reasoning consistency. Mammoth [68] trains an LLM with a hybrid of CoT and PoT rationales and achieves a SOTA model for math problems. + +# 3. Visual Program Distillation (VPD) + +We introduce VPD, a general model-agnostic framework that distills the reasoning power of LLM-generated programs together with the low-level image understanding abilities of vision tools into a single vision-language model (Fig. 2). At its core, VPD consists of two major steps: + +1. Program generation and verification: Given a textual query $q$ and a visual input $i$ , VPD first generates a program that solves the query by making use of vision modules and tools, then converts the program execution trace into a chain-of-thought $c$ (\$3.1). +2. Distilling step-by-step: The visual input $i$ , textual query $q$ , and chain-of-thought $c$ produced by the previous step are distilled into a vision-language model (VLM) (§3.2). + +# 3.1. Program generation and verification + +Our training data synthesis pipeline is illustrated in the blue boxes of Fig. 2, and consists of four stages. Given a sample $(i,q,y)$ consisting of a visual input $i$ and a textual query $q$ about its content and, when available, its ground truth answer $y$ , we perform the following sequence of steps: + +1. Program generation with LLM: Given $q$ , we generate a list of $k$ candidate programs $\pi(q) = \{z_1, z_2, \dots, z_k\}$ , where $\pi$ represents a program generation function. +2. Program execution with vision modules: We execute each program $z_{i}$ with an execution engine $\phi$ to obtain its final result $\phi (i,z_i) = \hat{y}_i$ . However, during program execution, we maintain the execution trace $t_i$ recording all the intermediate function calls and outputs. At the end of this step, we produce a list of programs, results, and execution traces $\{(z_1,\hat{y}_1,t_1),\dots,(z_k,\hat{y}_k,t_k)\}$ . +3. Program filtering: Among the $k$ candidate programs from the previous step, we keep a single tuple $(z,\hat{y},t)$ with correct answer. +4. Converting program execution traces into chain-of-thought rationales: We rewrite $t$ into a CoT $c$ using an LLM. + +We now discuss in detail each of the steps above. + +Program generation. We adopt a similar approach with recent work [16, 20] in our program generation step, and use PaLM-2 [3] as LLM to generate candidate programs for a given query $q$ . We prompt PaLM-2 with the same kind of text prompt as used by ViperGPT, which contains a detailed description of the available vision modules, followed by the query $q$ (prompt in Appendix §G). The LLM is expected to directly output a Python function definition that will be executed in the following steps. However, in our experiments, we find that the success rate of top- $k$ programs is much higher than the top-1 program. Therefore, in contrast to prior work, which only samples one program $z$ , when the ground truth answer $y$ is available, we set a temperature $T$ for LLM decoding and sample a list of top- $k$ candidate programs $\{z_1, z_2, \dots, z_k\}$ from the LLM. Then, we filter out one correct program $z$ in later steps. As shown in our ablation in §4.2, this becomes crucial to our performance gain. We use $T = 0.5$ and $k = 5$ for all our experiments. For unlabeled data, $k = 1$ and the filtering step can be skipped. + +Program execution with vision modules. We use the same execution engine $\phi$ as ViperGPT [16]. An LLM-generated program $z_{i}$ is a Python function that takes the visual input $i$ as input. While the LLM outputs the program as a sequence of text, the execution engine $\phi$ is able to interpret this as a Python program and execute it, to obtain its return result $\hat{y}_{i}$ . Additionally, $\phi$ also records the execution trace $t_{i}$ of the program $z_{i}$ , which keeps a record of all the vision module calls, their inputs, and outputs. We use the following tools for the program: PaLI-X [10] for simple visual queries, PaLI-X + +detection (distilled from OWLv2 [47]) for object detection; Google Cloud Depth API [17] for depth estimation, and PaLM-2 [3] for external knowledge. + +Program filtering. As we discussed in §1, visual programs are error-prone for a variety of reasons. The program might be wrong, and the execution process can introduce additional errors. To overcome these issues, we employ a program filtering step. We start by sampling top- $k$ programs and attempt to execute each one. During this process, any program that fails execution is instantly discarded. We further filter the remaining programs, the strategy depending on the availability of labeled data. For tasks where human labels are available, we select a single program per input sample whose answer is correct—that is, when the program output $\hat{y}_i$ matches the human label $y$ . In this case our visual program pipeline acts like a CoT rationale annotator. One potential problem is that some task answers are ambiguous. For example, for a question like "Where are the horses?", the answers "mountain" and "mountains" are both correct. We adopt the method in [28] and use an LLM to determine if the program output is correct (details in §G). If there is more than one correct program per question, we select the top scoring one, according to the scores provided by the program-generating LLM $\phi$ . If no programs pass the test, the rated sample is not wasted—we simply use the correct answer $y$ as supervision in our fine-tuning stage, without an associated CoT. When no human-rated answer is available, we directly use the top scoring executable program1. + +Converting program execution traces into chain-of-thought rationales. After the filtering step, for each visual input $i$ and query $q$ , we will have selected at most one program $z$ together with its execution result $\hat{y}$ and trace $t$ . Since most existing VLMs have been pre-trained with text in natural language and not code, we use an LLM to rewrite the execution trace $t$ into a natural language CoT $c$ for our VLM distillation. Some examples are shown in §B. Concretely, similar to prior work [23, 24], we hand-craft 20 examples of how an input $(q, z, t)$ can be converted into a CoT, and use these as few-shots for prompting PaLM-2 [3], which performs in-context learning and generates CoTs for new examples. We include a concrete example in §G. + +# 3.2. Distilling step-by-step + +In this step, we fine-tune a backbone VLM with the training data generated in §3.1, distilling the knowledge and reasoning steps of our generated programs into a single end-to-end model. We do so in a multitask fashion, where the VLM is simultaneously fine-tuned on data synthesized for multiple types tasks (e.g., free-form VQA, multiple choice). + +Let $f$ represent our VLM model. While the same VLM architecture could solve all these tasks, it needs to be prompted + +differently to adapt to the task. Following prior work [39], we manually design instructions for each task. For example, for free-form VQA queries, the instruction is "Answer with a single word or phrase", while for multiple-choice queries we use "Answer with the option letter from the given choices directly". During fine-tuning, we combine the training samples generated for all tasks into a single dataset, so to account for the different types of tasks, we augment each sample with its corresponding task-specific prompt $p$ . + +Using these instructions along with the data generated in §3.1, we put together the training dataset $\mathcal{D} = \{(i_j,q_j,\hat{y}_j,c_j,p_j)\}_{j = 1}^N$ , where $N$ is the total number of samples, $i_j$ is the visual input, $q_{j}$ is the textual query, $\hat{y}_i$ is the visual program output2, and $c_{j}$ is the CoT rationale. + +We train $f$ to minimize a loss for predicting both the label and the rationale. As shown in the red box of Fig. 2, similar to [21], we treat predicting the output label $\hat{y}_j$ and the rationale $c_j$ as two separate optimization goals. However, since VLMs are open-ended text generation models, they need additional prompting to indicate whether we want a short answer or a long answer that includes a rationale. Therefore, we append the suffix $s_c =$ "Explain the rationale to answer the question" at the end of the prompt for generating a CoT, and use the task instruction $p_j$ for short answers. Our optimization objective is: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\text {l a b e l}} + \mathcal {L} _ {\text {r a t i o n a l e}} (1) \\ = \sum_ {j = 1} ^ {N} \ell \left(f \left(i _ {j}, q _ {j}, p _ {j}\right), \hat {y} _ {j}\right) + \ell \left(f \left(i _ {j}, q _ {j}, s _ {c}\right), c _ {j}\right) (2) \\ \end{array} +$$ + +Here $\ell$ is the cross entropy-loss normalized by sequence length to ensure that the labels (typically short) and rationales (typically long) have similar weights. $\mathcal{L}_{\text{rationale}}$ both teaches the VLM to generate faithful reasoning steps similar to program execution traces, and carries more information beyond the label $\hat{y}$ that also helps the VLM in better predicting the label. During test time, the rationale generation is not required. We can adjust the task instruction to directly get the short output label using $p$ , and using $s_c$ to get the human-interpretable reasoning steps if needed. + +# 4. Experiments + +In this section, we demonstrate VPD's effectiveness by using it to train a generalist VLM. We attempt this for two VLMs with different scales, PaLI-3 (5B) [11] and PaLI-X (55B) [10]. Detailed experimental setups are given in §4.1. Qualitatively, models fine-tuned with VPD exhibit the ability to reason step-by-step like a program, as illustrated in Fig. 3, and supported by human evaluation results (§4.3). Quantitative results show that our PaLI-3-VPD and PaLI-X-VPD + +![](images/279813154032f56df66d0dcf9989edf3655182e43c8d8b3529479ce489ffafdc.jpg) +Figure 3. Outputs of PaLI-X trained with VPD. Training with execution traces of filtered programs improves the model's ability to count, understand spatial relations, and reason compositionally. The images are from Visual Genome [32]. + +achieve new SOTA on a broad range of tasks, on both generalist and per-task fine-tuning settings (§4.2). We also conduct a detailed analysis of the source of performance gain (§4.2). Finally, we conduct human evaluation on the quality of the rationales generated by our models, and compare it with rationales generated by an instruction-tuned model trained without VPD (§4.3). + +# 4.1. Experimental setup + +Backbone models. We use two state-of-the-art VLMs, PaLI-3 (5B) [11] and PaLI-X (55B) [10] as our base models. Both take images and text as input, and generate text as output. For simplicity, we further refer to them as "PaLI model" when we discuss steps performed with each of them individually. + +Data for Generalist Models. We fine-tune the pre-trained PaLI model on two types of datasets to make it a generalist VLM—that is, a model that performs relatively well on any task without further training on that specific task: + +(1) Multimodal Instruction-Tuning (MMIT) tasks, created in the spirit of Self-Instruct [64]. An LLM is prompted with image captions, and generates task inputs and the desired outputs about the corresponding image. Details of MMIT tasks are covered in [65]. Note that these tasks cover a wide variety of common use cases, but do not include the specific in-domain tasks used in this work. +(2) Academic task-oriented VQA tasks. Since image captions contain only a coarse description of visual information in the image, and may miss details that are important for solving the task. Additionally, LLM are likely to hallucinate during this data curation process. To further boost the accuracy of our VLMs, we also fine-tune the PaLI models with academic task-oriented VQA tasks. The data mixture covers subsets of a wide variety of VQA tasks, including general VQA (VQAv2 [19]), optical character recognition (OCRVQA [48]), compositional questions and reasoning (GQA [26]), counting (TallyQA [1]), and VQA that involves + +external knowledge (OK-VQA [45] and A-OKVQA [52]). The tasks contain a textual query and a short expected label. We use the pipeline in §3.1 to synthesize CoT reasoning steps for these labels, and tune the PaLI model with these the loss in Equation 1. Notice that sometimes the pipeline fails to find a program that generates the correct answer. In that case, we set $\mathcal{L}_{\text{rationale}}$ to 0 and only keep $\mathcal{L}_{\text{label}}$ . §4.2 shows how many programs are kept after the filtering stage. + +Data for specialist models. While fine-tuning the generalist model with VPD, we only use a subset of each task's training data. To evaluate our model's ability on each individual task, we continue fine-tuning PaLI-3-VPD and PaLI-X-VPD on each individual task on the training splits. + +Training setup. Both PaLI-3 and PaLI-X follow an encoder-decoder architecture, where images are encoded into visual tokens via a visual encoder and then passed into a UL2 model. Due to resource constraint, we use LoRA [22] to fine-tune both PaLI models. Specifically, we add LoRA weights on each linear layer in the attention blocks and the MLP blocks for both the encoder and decoder in the UL2 transformer. More training details are in §C. + +Evaluation setup. We evaluate our models on a wide range of tasks, including various VQA tasks and recent zero-shot VLM benchmarks. Noted that A-OKVQA contains two kinds of questions, multiple-choice (MC) and direct answer (DA). We report results on both. TallyQA [1] contains complex counting questions that involve object relationships, attribute identification, and reasoning to get the correct answer. It contains two evaluation partitions, simple and complex. TextVQA [55] focuses reading texts. We include it to evaluate our models on zero-shot tasks. In addition to VQA tasks, we also test our models on two popular VLM benchmarks. POPE [36] focuses on VLM hallucination, containing binary questions of whether or not an object exists in the image. MMBench [40] is a robust and comprehensive VLM benchmark testing a range of fine-grained abilities (e.g., object localization, attribute recognition, spatial relationship). Details of the evaluation sets and metrics are in §C. + +Baselines. We refer to the two PaLI models fine-tuned with VPD as PaLI-3-VPD and PaLI-X-VPD, and compare them with various baselines. To evaluate the effectiveness of VPD, we experiment with removing the synthesized CoT from our data mixture, and train the PaLI models with the exact same hyper-parameters and steps. We call these models as PaLI-3 Instruct and PaLI-X Instruct. For a fair comparison, these models are also trained with the same supervised loss for predicting the ground truth answers, on the same images and textual queries as our VPD variant. Moreover, we also compare with the most recent SOTA vision-language models. These VLMs are initialized with pre-trained visual encoders and LLMs, and then trained with image-text pairs, LLM-generated data, and academic tasks. + +
Generalist ModelsVQAv2GQAOK-VQAA-OKVQATallyQATextVQAPOPEMMB
MCDASimp.Comp.
Prior generalist VLMs
Flamingo (80B) [2]82.0-57.8*----57.1--
MiniGPT-4 (Vicuna-13B) [79]-43.5-67.2-----42.3
InstructBLIP (Vicuna-13B) [14]-49.5*---75.2*57.5*50.7*78.944.0
Shikra (Vicuna-13B) [7]77.4-47.2-----84.758.8
Qwen-VL (9.7B) [4]78.859.358.6--82.6*65.8*63.8-38.2
Qwen-VL-Chat (9.7B) [4]78.257.556.6--81.1*64.0*61.5-60.6
mPLUG-Owl2 (8.2B) [74]79.456.157.7----54.3*86.264.5
LLaVA-1.5 (Vicuna-13B) [38]80.063.3---76.9*65.4*61.3*85.967.7
Our instruction-tuned baselines
PaLI-3-Instruct (5B)79.959.756.778.357.681.970.463.3*87.768.6
PaLI-X-Instruct (55B)83.663.364.384.161.585.575.465.0*88.975.0
Our visual program distillation models
PaLI-3-VPD (5B)80.461.357.578.556.583.170.963.7*88.669.0
PaLI-X-VPD (55B)83.964.964.684.562.786.276.665.4*88.876.2
+ +Table 1. Comparison with SOTA generalist VLMs. PaLI-X-VPD outperforms all prior models. VPD improves performance on 8/9 tasks for both PaLI-3 and PaLI-X, and is particularly effective on counting questions (TallyQA), compositional questions (GQA), and the comprehensive benchmark (MMBench). Underline indicates when PaLI-3-VPD outperforms the Instruct version. * marks tasks unseen during training. POPE and MMBench are zero-shot benchmarks for all models. + +# 4.2. Quantitative results + +Generalist model. Table 1 compares VPD with the baselines discussed in §4.1. We infer all answers by open-ended generation with the prompt "Answer with a single word or phrase," using greedy decoding without any constraint on the model's output space. For multiple-choice questions, we run inference with the prompt "Answer with the option letter from the given choices directly." and generate the option letter. As Table 1 shows, PaLI-X-VPD sets the new state-of-the-art on all benchmarks. Compared with prior generalist VLMs, it achieves significant improvement on MMBench (+8.5), TallyQA complex (+9.8), and A-OKVQA(+9.5 compared with specialist SOTA[14]). PaLI-3-VPD, despite its small architecture size (5B), outperforms all prior models on VQAv2, TallyQA, POPE, and MMBench, including much larger models that use Vicuna-13B as backbones [7, 14, 38]. However, its performance on knowledge-based VQA tasks does not outperform the prior SOTA. Our hypothesis for this is that its language model (3B UL2) is too small to contain all the necessary knowledge needed for correctly solving these tasks. + +When compared to their Instruct variants, both PaLI-3-VPD and PaLI-X-VPD outperform on 11/12 tasks. Specifically, for PaLI-X, VPD obtains a $+1.6$ improvement on GQA (which is heavily focused on compositional questions, spatial relationship, and localization), $+1.2$ on TallyQA for complex questions, and $+1.2$ on MM-Bench. These results suggest that VPD is a more effective method for creating instruction-tuning data which enables VLMs to improve their visual reasoning ability. + +Per-Task Fine-Tuning (Specialist). The results for the specialist models are shown in Table 2. PaLI-X-VPD sets a new SOTA on all benchmarks. Note how the specialist models tend to have higher scores than the generalist one. We propose several hypotheses for this performance gain: + +(1) As shown in Table 2, the score gap tends to be larger on free-form VQA tasks. This may be due the fact that human annotations on these datasets have ambiguities. For example, for the question "Who is looking up?", GQA [26] labels are "man" or "woman" while OK-VQA [45] have more detailed labels, for example, "worker" or "cook". Per-task fine-tuning alleviates this annotation ambiguity and lets the model focus on the annotation style of that task. For multiple-choice and counting tasks, the answer has less ambiguity, and the score gaps are much smaller. (2) The performance gap between PaLI-3 (5B) specialist and generalist is larger than that of PaLI-X. We hypothesize that this shows models with larger scale has more multi-task capacity; (3) Per-task fine-tuning adds more in-domain training data, which generally improves performance. + +Analysis: Sampling multiple programs is key to good data generation. Fig. 5 shows the success rate of finding at least one program that passes the filtering stage, when the LLM generates the top-1 or top-5 programs, respectively. There is a dramatic increase in success rate from 1 program to 5: $+45\%$ on GQA and A-OKVQA, $+33\%$ on OK-VQA, and $+10\%$ on TallyQA. This design choice greatly improves our data synthesis efficiency, and, as a consequence, adding more CoT data that requires complex reasoning in our training set. We also conduct an analysis in $\S E$ to compare the performance of VPD models with that of the visual programs they are distilled from. + +# 4.3. Human evaluation on rationales + +In this section, we focus on the quality of model outputs. We performed this analysis using human annotators, who are asked to evaluate both the correctness of the final answer and the quality of the rationale behind it. + +Models. We compare PaLI-X-VPD with PaLI-X Instruct. Among possible baselines, we chose PaLI-X + +
Specialist ModelsGQA test-devOK-VQA valA-OKVQATallyQA
Multi-choice valtestDirect Answer valtestSimple testComplex test
InstructBLIP (Vicuna-7B) [14]-62.175.773.464.062.1--
PaLI-3 (5B) [11]-60.1----83.370.5
PaLI (17B) [9]-64.5----81.770.9
CogVLM (17B) [63]65.264.7------
PaLI-X (55B) [10]-66.1----86.075.6
PaLI-3-VPD (5B) generalist61.357.578.5-56.5-83.170.9
PaLI-3-VPD (5B) specialist64.760.379.776.565.563.683.370.8
PaLI-X-VPD (55B) generalist64.964.684.5-62.7-86.276.6
PaLI-X-VPD (55B) specialist67.366.885.280.471.168.286.276.4
+ +Table 2. Comparison of per-task fine-tuning results of specialist models. PaLI-X-VPD sets a new SOTA for all the tasks. + +![](images/1a595d7b0c04990ac23bb4d038af1cadfa30941ed3dfcf7d9aabd3157eb1650e.jpg) +Figure 4. Human evaluation results assessing answer and rational quality for PaLI-X Instruct and PaLI-X VPD: (a) percentage of answers that are correct and have explanations; (b) rationale factuality and consistency for the answers that contain an explanation; (c) preference between the two models, when aggregating across all samples ("All") and across those with correct answers ("When correct"). + +![](images/4cf7da159c3d1a5b097eea3a36149d00c37d3020592f15ed4d6c98a346b0dd6a.jpg) + +![](images/e85a5d609dee9a1abdef080c7e96d0ccbc3273eaf73395e8d6d4be058b0979e7.jpg) + +![](images/c4708f067e55637075c087e96be6e10935beed8744c236b7f28a61bd46232ac3.jpg) +Figure 5. Success rate of top-1 program and top-5 programs on the training set during our data synthesis process. + +Instruct because it is trained to generate long-form answers [65], which allows us to assess if it is the quality of PaLI-X-VPD's answers that the annotators prefer, rather than its length. Since PaLI-X Instruct is also instruction-tuned, it can be prompted to provide long answers to alleviate this confounder. + +Annotation protocol. We run inference with each of the two models on a combination of 600 samples from GQA (test-dev), A-OKVQA (val), and TallyQA (Simple and Complex) and record their answers and rationales. We then ask 3 human annotators to evaluate each model answer. We use prior work from natural language processing (NLP) as inspiration, and build upon it for selecting the evaluation criteria [75]. Given an image and a query, for each model-generated rationale, we ask human annotators to score the model answers along the following criteria: (1) correctness—is the final answer correct? (2) explainability—does the model explain its rationale for reaching the final answer? (3) factuality—is every step in the rationale factually cor + +rect (with respect to the image and external knowledge)? (4) consistency—does step and final answer logically follow from the previous ones? Note that a rationale may have the wrong answer while being consistent. We also conduct a side-by-side comparison, and ask annotators which of the two answers—the one provided by PaLI-X-VPD or by PaLI-X Instruct—they prefer in terms of quality of the answer and explanation. More details about the annotation protocol are in §D. + +Human evaluation results. The results of the evaluation are first averaged among the human raters per sample, then aggregated across samples. Our PaLI-X-VPD model far outperforms PaLI-X Instruct along all criteria. Fig. 4 (a) shows the correctness of the final answer provided by each model (i.e. accuracy), and the proportion of the samples where the model provides a CoT rationale to explain its answer. PaLI-X-VPD's gain in accuracy of $+16.7\%$ is even more impressive than in evaluations in §4.2 based on benchmark labels—this is because human annotators are better able to assess correctness for ambiguous questions with different possible interpretations (e.g., the model's answer "In the living room." to the question "Where is the couch located?" was considered correct by annotators, even when the benchmark answer was "On the right"). Moreover, the explainability results confirm that our model is able to explain its own answer on $+24\%$ more samples than the instruction tuned model. Additionally, Fig. 4 (b) shows impressive rationale quality: among the samples where + +an explanation is provided, PaLI-X-VPD's rationales are factual $87.2\%$ of times, and consistent $97.8\%$ of times, a gain of $+14.6\%$ and $+10\%$ , respectively, when compared to PaLI-X Instruct. + +We also asked the annotators which of the two answers given by the two models they prefer. We aggregated these results in two ways: (1) across all samples, (2) across the samples where both models answered the question correctly. The results in Fig. 4 (c) show that PaLI-X-VPD is preferred on $25\%$ more samples than PaLI-X Instruct in the "All" case, and on $12\%$ more samples when both are correct. This suggests that even when PaLI-X-VPD makes mistakes, it still provides a better answer. Such examples are shown in $\S D$ . These results confirm that a model fine-tuned with VPD leads to more faithful and consistent answers. + +# 5. Experiments on content moderation + +Prior experiments focus on training generalist VLMs. Here, we explore the effectiveness of VPD at quickly adapting models to real-world applications from a different domain than the training data [e.g., 56]. We experiment on Hateful Memes [29], a content moderation dataset where the task is to classify if a meme contains hateful content. The target labels are "yes" or "no", and models are evaluated in terms of classification accuracy and AUC-ROC. We experiment with two settings: supervised, in which the models are trained on the provided training set with 8,500 labels, and unsupervised, in which no human labels are provided. + +
ModelAccAUC-ROC
Unsupervised / Zero-Shot Methods
Generated Programs (ours)69.770.1
Supervised Methods
VisualBERT [35]69.575.4
Flamingo (80B) [2]-86.6
Previous SOTA [46]78.886.7
PaLI-X-VPD (label-only FT)77.688.0
PaLI-X-VPD (specialist w/ CoT)80.889.2
Human [29]84.782.7
+ +Table 3. Results on Hateful Memes [29] seen test set. We improve SOTA for both zero-shot and supervised settings. Unsupervised PaLI-X-VPD outperforms strong supervised baselines. + +# 5.1. Unsupervised setting + +# We experiment with three methods: + +1. Generated Program: The program generated by PaLM-2 [3] for solving this task (shown in §F) consists of following main steps: (1) get image description with PaLI-X [10]; (2) use an OCR tool to extract embedded texts; (3) given the image description and OCR texts, ask PaLM-2 to explain if this meme is hateful. +2. PaLI-X-VPD (generalist): In a zero-shot setting, we directly prompt our PaLI-X-VPD (generalist) with: "The text is $. Is this a hateful meme?" We compute the probability of PaLI-X-VPD generating "yes" + +or "no" and measure accuracy and AUC-ROC. + +3. PaLI-X-VPD (specialist with zero-shot CoT): We follow our VPD pipeline and convert the execution traces of the program into CoTs, and then fine-tune PaLI-X-VPD to output these CoTs. Since no groundtruth labels are available, no filtering is done during the process. + +VPD significantly improves performance even when no labels are available. As shown in Table 3, PaLI-X-VPD (generalist) outperforms all other VLMs (61.4%). Interestingly, the generated programs themselves get much higher accuracy (69.7%) than on the previous datasets, perhaps because PaLM-2 is better suited at analyzing a meme than typical VQA datasets. Moreover, our PaLI-X-VPD (specialist) sets a new zero-shot SOTA on Hateful Memes, and even outperforms supervised VisualBERT [35]. Manual inspection of model outputs show a great deal of similarity with the generated code. See examples in §F. + +# 5.2. Supervised setting + +We fine-tune PaLI-X-VPD in two ways: + +1. Label-only fine-tuning: To establish a strong baseline, we first experiment with the traditional supervised setting, where we fine-tune the model to output "yes" or "no". +2. PaLI-X-VPD (specialist): We use the complete VPD pipeline to train this model. We select an execution trace that leads to the correct label, and use it to tune our VLM. + +Results. Our PaLI-X-VPD (specialist) sets a new SOTA for this task, with an accuracy of $80.8\%$ and AUC-ROC of $89.2\%$ . It outperforms the label-only fine-tuning baseline and significantly improves the SOTA metrics, achieving nearly human-level accuracy, and superhuman AUC-ROC. + +# 6. Conclusion, limitations, and future work + +In this paper, we introduced VPD, a framework for distilling the reasoning abilities of LLMs along with the capabilities of vision tools into VLMs. VPD synthesizes training data for VLMs by generating programs that can leverage external tools. We used this technique to fine-tune some of the best existing VLMs (PaLI-3 and PaLI-X) and established new SOTA results on 8 classical VQA and 2 zero-shot multimodal benchmarks. According to human evaluations, VPD-tuned models provide more accurate answers and better rationales. Experiments on the Hateful Memes dataset show how VPD can also adapt models to new real-world domains, even when no labeled data is available, also establishing a new SOTA on this dataset. From our experiments we learnt that both a strength and a weakness of VPD is the quality of the generated programs. Better programs lead to bigger gains with VPD. We discuss multiple ideas on how to improve the program generation in Appendix H, including using additional tools, agents, and fact-checking for multimodal CoTs. + +# References + +[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 1, 5, 14, 15 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022. 1, 3, 6, 8 +[3] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby Ambrose Slone,Daniel Smilkov David R. So Daniel Sohn Simon Tokumine,Dasha Valter,Vijay Vasudevan,Kiran Vodrahalli,Xuezhi Wang,Pidong Wang,Zirui Wang,Tao Wang John Wieting,Yuhuai Wu,Kelvin Xu,Yunhan Xu,Linting Xue,Pengcheng Yin,Jiahui Yu,Qiao ZhangSteven Zheng Ce ZhengWeikang ZhouDenny ZhouSlav Petrov and Yonghui Wu.Palm 2 technical report,2023.134813 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 1, 3, 6 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 + +[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigptv2: large language model as a unified interface for vision-language multi-task learning, 2023. 3 +[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 1, 2, 3, 6 +[8] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks, 2022. 1, 3 +[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 1, 3, 7 +[10] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, AJ Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali-x: On scaling up a multilingual vision and language model, 2023. 2, 3, 4, 5, 7, 8, 15, 16 +[11] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, Daniel Salz, Xi Xiong, Daniel Vlasic, Filip Pavetic, Keran Rong, Tianli Yu, Daniel Keysers, Xiaohua Zhai, and Radu Soricut. Pali-3 vision language models: Smaller, faster, stronger, 2023. 1, 4, 5, 7, 15, 16 +[12] Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, et al. Binding language models in symbolic languages. arXiv preprint arXiv:2210.02875, 2022.3 +[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robin- + +son, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways, 2022. 3 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 6, 7 +[15] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaeldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 15 +[16] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 1, 2, 3, 20, 24 +[17] Ruofei Du, Eric Turner, Maksym Dzitsiuk, Luca Prasso, Ivo Duarte, Jason Dourgarian, Joao Afonso, Jose Pascoal, Josh Gladstone, Nuno Cruces, Shahram Izadi, Adarsh Kowdle, Konstantine Tsotsos, and David Kim. DepthLab: Real-Time 3D Interaction With Depth Maps for Mobile Augmented Reality. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. ACM, 2020. 4 +[18] Weixi Feng, Wanrong Zhu, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. *ArXiv*, abs/2305.15393, 2023. 2 +[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5, 15 +[20] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3 +[21] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, 2023. 2, 3, 4 +[22] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen + +Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 5, 16 +[23] Yushi* Hu, Hang* Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2, 3, 4 +[24] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A. Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2627-2643, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 4 +[25] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1, 3 +[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 5, 6, 14, 15 +[27] Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir Ghaemmaghami, Rajendra Gottipati, William Gulland, Robert Hagemann, C. Richard Ho, Doug Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander Kaplan, Harshit Khaitan, Daniel Killebrew, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon, James Law, Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana Maggiore, Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy Nix, Thomas Norrie, Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Matt Ross, Amir Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham, Jed Souter, Dan Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma, Erick Tuttle, Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon. Indatacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture. ACM, 2017. 16 +[28] Ehsan Kamalloo, Nouha Dziri, Charles L. A. Clarke, and Davood Rafiei. Evaluating open-domain question answering in the era of large language models, 2023. 4, 23 +[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes, 2020. 2, 8, 18, 19 +[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 16 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 24 + +[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual genome: Connecting language and vision using crowdsourced dense image annotations, 2016. 3, 5, 15 +[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2023. 24 +[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 3 +[35] Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A Simple and Performant Baseline for Vision and Language. ArXiv, abs/1908.03557, 2019. 8 +[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 5, 15 +[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3 +[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 6 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 3, 4 +[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2023. 2, 5, 15 +[41] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. ArXiv, abs/2206.08916, 2022. 1, 3 +[42] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023. 3 +[43] Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. Faithful chain-of-thought reasoning, 2023. 1, 3 +[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 1 +[45] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6, 15 +[46] Jingbiao Mei, Jinghong Chen, Weizhe Lin, Bill Byrne, and Marcus Tomalin. Improving hateful memes detection via learning hatefulness-aware embedding space through retrieval-guided contrastive learning, 2023. 8 + +[47] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection, 2023. 4 +[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5, 15 +[49] OpenAI. Gpt-4 technical report, 2023. 1, 3 +[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world, 2023. 3 +[51] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 3 +[52] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 5, 14, 15 +[53] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718, 2022. 1 +[54] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face, 2023. 3 +[55] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 5, 15 +[56] Otilia Stretcu, Edward Vendrow, Kenji Hata, Krishnamurthy Viswanathan, Vittorio Ferrari, Susan Tavakkol, Wenlei Zhou, Aditya Avinash, Emming Luo, Neil Gordon Alldrin, et al. Agile modeling: From concept to classifier in minutes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22323-22334, 2023. 8 +[57] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation, 2023. 1 +[58] Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. UL2: Unifying Language Learning Paradigms, 2022. 15 +[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, + +Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models, 2023. 3 +[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1, 3 +[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 3 +[62] Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. Scott: Self-consistent chain-of-thought distillation, 2023. 3 +[63] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 1, 3, 7 +[64] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5 +[65] Yaqing Wang, Jialin Wu, Tanmaya Dabral, Jiageng Zhang, Geoff Brown, Chun-Ta Lu, Frederick Liu, Yi Liang, Bo Pang, Michael Bendersky, and Radu Soricut. Non-intrusive adaptation: Input-centric parameter-efficient fine-tuning for versatile multimodal modeling, 2023. 2, 5, 7 +[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2022. 3 +[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation. 2023. 24 +[68] Ge Zhang Yao Fu Wenhao Huang Huan Sun Yu Su Wenhu Chen Xiang Yue, Xingwei Qu. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 3 +[69] Yiheng Xu, Hongjin Su, Chen Xing, Boyu Mi, Qian Liu, Weijia Shi, Binyuan Hui, Fan Zhou, Yitao Liu, Tianbao Xie, Zhoujun Cheng, Siheng Zhao, Lingpeng Kong, Bailin Wang, Caiming Xiong, and Tao Yu. Lemur: Harmonizing natural language and code for language agents, 2023. 1 +[70] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081–3089, 2022. 3 +[71] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn + +of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9, 2023. 1 +[72] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 3 +[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 24 +[74] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration, 2023. 1, 3, 6 +[75] Xi Ye and Greg Durrett. The unreliability of explanations in few-shot prompting for textual reasoning. In Advances in Neural Information Processing Systems, pages 30378-30392. Curran Associates, Inc., 2022. 7 +[76] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3 +[77] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 15 +[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding, 2022. 3 +[79] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 3, 6 +[80] Xuekai Zhu, Biqing Qi, Kaiyan Zhang, Xingwei Long, and Bowen Zhou. Pad: Program-aided distillation specializes large models in reasoning, 2023. 3 \ No newline at end of file diff --git a/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/images.zip b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f9f1dcec780b98525edc8d237664b37cdb03cbb --- /dev/null +++ b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78d31275311743d45721c8e19fdc1ac4415372557c7e2bfa1f823578bc6d0d02 +size 411620 diff --git a/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/layout.json b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d0794aa3e6ed95aece9f9c944642c5ea507b077 --- /dev/null +++ b/2024/Visual Program Distillation_ Distilling Tools and Programmatic Reasoning into Vision-Language Models/layout.json @@ -0,0 +1,9238 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 530, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 530, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 530, + 140 + ], + "type": "text", + "content": "Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 159, + 494, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 159, + 494, + 217 + ], + "spans": [ + { + "bbox": [ + 97, + 159, + 494, + 217 + ], + "type": "text", + "content": "Yushi Hu1,2* Otilia Strencu1 Chun-Ta Lu1 Krishnamurthy Viswanathan1 Kenji Hata1 Enming Luo1 Ranjay Krishna2 Ariel Fuxman1 1Google Research 2University of Washington https://visual-program-distillation.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 243, + 192, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 243, + 192, + 256 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 192, + 256 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 268, + 290, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 290, + 613 + ], + "type": "text", + "content": "Solving complex visual tasks such as \"Who invented the musical instrument on the right?\" involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs. Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 629, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 128, + 641 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 649, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 288, + 696 + ], + "type": "text", + "content": "Vision-language models (VLMs) have become the pretrained backbone for many computer vision tasks [2, 4, 7, 9-11, 34, 39, 41, 60, 63, 74, 79]. Yet, all these models still fall short of solving numerous visual reasoning tasks expected" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 232, + 545, + 404 + ], + "blocks": [ + { + "bbox": [ + 308, + 232, + 545, + 404 + ], + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 404 + ], + "type": "image", + "image_path": "948f46a27311e660bc2833cdebe63714935e90d4853f13c32d2d80baf27e299c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 411, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 455 + ], + "type": "text", + "content": "Figure 1. We introduce Visual Program Distillation (VPD), a training framework which leverages LLM-generated programs that make calls to specialist models and tools to distill cross-modal reasoning abilities and specialist skills into multimodal models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "type": "text", + "content": "of competent vision models. Even state-of-the-art (SOTA) proprietary vision-language models such as GPT-4V [49] do not perform well on tasks that involve counting and spatial reasoning [71]. They find it difficult to count (TallyQA [1]), to compositionally reason (GQA [26]), and to reason with external knowledge (OK-VQA [44], A-OKVQA [53]). Many of these tasks require VLMs to conduct compositional reasoning, which still remains an unsolved challenge. For instance, answering the question \"Who invented the musical instrument on the right?\" involves a composition of skills: identifying objects, applying spatial reasoning to locate the one on the right, recognizing the musical instrument, and accessing prior knowledge to retrieve the inventor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 631, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 548, + 715 + ], + "type": "text", + "content": "In contrast, large language models (LLMs) have demonstrated remarkable performance at generating code that solves complex and compositional tasks [3, 8, 43, 49, 69]. Several recent papers [16, 20, 25, 57] capitalize on this by prompting LLMs to generate programs where each step corresponds to a reasoning step. The programs invoke specialized \"tools\" (or specialized vision models) to explicitly" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 217, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 217, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 217, + 712 + ], + "type": "text", + "content": "*Work done while interning at Google Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9590" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 545, + 259 + ], + "type": "image", + "image_path": "07033e3703646704ac7b1eba64c1ea5916cfc8cfbb8a6ec25a25a728543fef5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 267, + 548, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 548, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 548, + 313 + ], + "type": "text", + "content": "Figure 2. Overview of Visual Program Distillation (VPD). VPD uses an LLM and specialized vision tools to generate faithful chain-of-thought (CoT) training data for vision-language models (VLMs). Given a multimodal input, our 4-step data synthesis pipeline generates a CoT that answers the query. In the example above, our synthesized CoT contains a series of reasoning steps: find all buses, check if each bus is yellow, and aggregate the count into a final answer. The CoT also contains the grounding information given by object detection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 321, + 288, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 380 + ], + "type": "text", + "content": "execute each reasoning step. For the question above, the program would call an \"object detector\" tool to identify and isolate all the objects, a \"fine-grained object classification\" tool to recognize the musical instrument, and a \"knowledge-based question answering\" tool to retrieve its inventor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 381, + 288, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 288, + 499 + ], + "type": "text", + "content": "Although innovative, generating explicit programs is computationally expensive in practice, prone to errors, and still underperforms end-to-end models. Programs require loading and executing multiple tools, leading to high latency and computational cost. Moreover, generated programs may omit necessary steps or include spurious ones. Even when the program is correct, vision model invocations can produce incorrect outputs, from which the overall program cannot recover. Unfortunately, empirical results show that visual programs still fall short of end-to-end fine-tuned models [16]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 499, + 288, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 288, + 631 + ], + "type": "text", + "content": "Another line of work is visual instruction tuning [39], which tries to distill the instruction-following ability of LLMs into VLMs. They prompt LLMs with image captions and bounding box annotations in order to generate queries and answers that are used to fine-tune VLMs [7, 23, 39, 65]. However, this approach has important limitations: image captions can miss fine-grained visual information, and LLM are prone to producing inconsistent outputs for custom vision representations like bounding boxes [18]. As a result, existing instruction-tuned VLMs still struggle with tasks that require complex visual reasoning [7, 14, 38, 40]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 631, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 288, + 714 + ], + "type": "text", + "content": "In this work, we present Visual Program Distillation (VPD), a novel distillation method that induces complex cross-modal reasoning capabilities into vision-language models (Fig. 2). As the name suggests, VPD combines two key insights to deliver a training paradigm that surpasses the sum of its parts: It relies on (1) advancements in visual programs that use tools [20] and (2) the recent breakthroughs in dis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 321, + 547, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 475 + ], + "type": "text", + "content": "tillation through chain-of-thought reasoning [21]. Given a labeled training dataset of complex visual tasks, VPD generates a correct program, and then distills its reasoning steps into vision-language models. To avoid using programs that give the wrong answer, VPD prompts an LLM to generate multiple candidate programs, and executes every one of them. When labeled data is available, it then filters for programs that produce the correct answer upon execution. Therefore, our programs comprise multiple vision tools, are executable, and yield the correct answer when executed. Next, VPD rewrites the correct programs as natural language chain-of-thought instructions and uses step-by-step distillation [21] to inject the reasoning abilities into VLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 475, + 547, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 618 + ], + "type": "text", + "content": "Our best instruction-tuned model, PaLI-X-VPD (55B), sets a new SOTA result on 8 classical VQA tasks and 2 zero-shot multimodal benchmarks. Our models even outperform the recent SOTA established by PaLI-X [10]. Importantly, we conduct a quality evaluation with human raters which shows that PaLI-X-VPD generates more consistent and faithful rationales compared to its counterpart trained using instruction-tuning data. In addition, we experiment with PaLI-3 (5B), and show that VPD also improves the performance of smaller-scale models. Finally, experiments on Hateful Memes [29] show that VPD is also helpful for adapting to new tasks, even when no labels are available." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 629, + 389, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 389, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 389, + 641 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 649, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 547, + 696 + ], + "type": "text", + "content": "VPD is a general method for improving any vision-language model, and includes automatic program generation and training with chain-of-thought data as steps of the proposed framework. We discuss each of these research areas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Vision-language models (VLMs). Most recent generative" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "9591" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "VLMs share a common structure: a pre-trained visual encoder, a pre-trained LLM, and a connector between the two modalities [2, 4, 9, 34, 39, 41, 60, 63, 74, 79]. The models are trained on large-scale image-text pairs from various tasks to adapt to both modalities. They are also tuned on LLM-generated visual instructions to enable models to follow versatile instructions from users [39]. Some models also include bounding boxes in the pre-training data to improve the VLM on visually grounded tasks [6, 7, 10, 41, 50, 61, 78]. The bounding boxes are usually retrieved from COCO [37] and Visual Genome [32]. Different from prior work, our method does not rely on provided dense annotations. We use LLM-generated code and specialized vision tools to generate our own visual instruction-tuning data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 361 + ], + "type": "text", + "content": "Visual programming and agents. With the advancement of large language models (LLMs) [3, 5, 13, 49, 59], recent work has started using LLMs as an interface to solving complex reasoning tasks with tools [12, 42, 51, 54, 76], and also as an agent for vision tasks [23, 25, 70, 72]. From this line of work, the most relevant to us are VisProg [20] and ViperGPT [16], which leverage LLMs to generate executable programs with a sequence of invocations to specialized vision tools. They achieve SOTA zero-shot performance on various vision tasks, while being versatile and interpretable." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 365, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 289, + 533 + ], + "type": "text", + "content": "Training and inference with chain-of-thought. Chain-of-Thought (CoT) [66] has become a popular approach to improving LLM performance. Recent work such as Program-of-Thoughts (PoT) [8] and Faithful CoT [43] further improve this framework by splitting inference into two steps: first generate a program with LLM, then execute the program. This approach achieves better accuracy and reduces hallucinations in the reasoning steps. Moreover, CoT is also used to train language models. Distill step-by-step [21], PaD [80], and SCOTT [62] train smaller language models with CoT generated by LLMs, showing that this can improve model performance and reasoning consistency. Mammoth [68] trains an LLM with a hybrid of CoT and PoT rationales and achieves a SOTA model for math problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 539, + 238, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 238, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 238, + 552 + ], + "type": "text", + "content": "3. Visual Program Distillation (VPD)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": "We introduce VPD, a general model-agnostic framework that distills the reasoning power of LLM-generated programs together with the low-level image understanding abilities of vision tools into a single vision-language model (Fig. 2). At its core, VPD consists of two major steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 616, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "text", + "content": "1. Program generation and verification: Given a textual query " + }, + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "text", + "content": " and a visual input " + }, + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "text", + "content": ", VPD first generates a program that solves the query by making use of vision modules and tools, then converts the program execution trace into a chain-of-thought " + }, + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 616, + 288, + 675 + ], + "type": "text", + "content": " (\\$3.1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "text", + "content": "2. Distilling step-by-step: The visual input " + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "text", + "content": ", textual query " + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "text", + "content": ", and chain-of-thought " + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 675, + 289, + 712 + ], + "type": "text", + "content": " produced by the previous step are distilled into a vision-language model (VLM) (§3.2)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 499, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 499, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 499, + 85 + ], + "type": "text", + "content": "3.1. Program generation and verification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": "Our training data synthesis pipeline is illustrated in the blue boxes of Fig. 2, and consists of four stages. Given a sample " + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "inline_equation", + "content": "(i,q,y)" + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": " consisting of a visual input " + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": " and a textual query " + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": " about its content and, when available, its ground truth answer " + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": ", we perform the following sequence of steps:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 153, + 547, + 342 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "text", + "content": "1. Program generation with LLM: Given " + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "text", + "content": ", we generate a list of " + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "text", + "content": " candidate programs " + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "inline_equation", + "content": "\\pi(q) = \\{z_1, z_2, \\dots, z_k\\}" + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 306, + 153, + 547, + 189 + ], + "type": "text", + "content": " represents a program generation function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "content": "2. Program execution with vision modules: We execute each program " + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "content": " with an execution engine " + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "content": " to obtain its final result " + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "inline_equation", + "content": "\\phi (i,z_i) = \\hat{y}_i" + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "content": ". However, during program execution, we maintain the execution trace " + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "content": " recording all the intermediate function calls and outputs. At the end of this step, we produce a list of programs, results, and execution traces " + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "inline_equation", + "content": "\\{(z_1,\\hat{y}_1,t_1),\\dots,(z_k,\\hat{y}_k,t_k)\\}" + }, + { + "bbox": [ + 306, + 190, + 547, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "type": "text", + "content": "3. Program filtering: Among the " + }, + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "type": "text", + "content": " candidate programs from the previous step, we keep a single tuple " + }, + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "type": "inline_equation", + "content": "(z,\\hat{y},t)" + }, + { + "bbox": [ + 306, + 272, + 545, + 306 + ], + "type": "text", + "content": " with correct answer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "type": "text", + "content": "4. Converting program execution traces into chain-of-thought rationales: We rewrite " + }, + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "type": "text", + "content": " into a CoT " + }, + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 308, + 547, + 342 + ], + "type": "text", + "content": " using an LLM." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "type": "text", + "content": "We now discuss in detail each of the steps above." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": "Program generation. We adopt a similar approach with recent work [16, 20] in our program generation step, and use PaLM-2 [3] as LLM to generate candidate programs for a given query " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": ". We prompt PaLM-2 with the same kind of text prompt as used by ViperGPT, which contains a detailed description of the available vision modules, followed by the query " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " (prompt in Appendix §G). The LLM is expected to directly output a Python function definition that will be executed in the following steps. However, in our experiments, we find that the success rate of top-" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " programs is much higher than the top-1 program. Therefore, in contrast to prior work, which only samples one program " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": ", when the ground truth answer " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " is available, we set a temperature " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " for LLM decoding and sample a list of top-" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " candidate programs " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\{z_1, z_2, \\dots, z_k\\}" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " from the LLM. Then, we filter out one correct program " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " in later steps. As shown in our ablation in §4.2, this becomes crucial to our performance gain. We use " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "T = 0.5" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " for all our experiments. For unlabeled data, " + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 304, + 363, + 547, + 589 + ], + "type": "text", + "content": " and the filtering step can be skipped." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Program execution with vision modules. We use the same execution engine " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " as ViperGPT [16]. An LLM-generated program " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is a Python function that takes the visual input " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " as input. While the LLM outputs the program as a sequence of text, the execution engine " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is able to interpret this as a Python program and execute it, to obtain its return result " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " also records the execution trace " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " of the program " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", which keeps a record of all the vision module calls, their inputs, and outputs. We use the following tools for the program: PaLI-X [10] for simple visual queries, PaLI-X" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9592" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": "detection (distilled from OWLv2 [47]) for object detection; Google Cloud Depth API [17] for depth estimation, and PaLM-2 [3] for external knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "content": "Program filtering. As we discussed in §1, visual programs are error-prone for a variety of reasons. The program might be wrong, and the execution process can introduce additional errors. To overcome these issues, we employ a program filtering step. We start by sampling top-" + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "content": " programs and attempt to execute each one. During this process, any program that fails execution is instantly discarded. We further filter the remaining programs, the strategy depending on the availability of labeled data. For tasks where human labels are available, we select a single program per input sample whose answer is correct—that is, when the program output " + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "content": " matches the human label " + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "content": ". In this case our visual program pipeline acts like a CoT rationale annotator. One potential problem is that some task answers are ambiguous. For example, for a question like \"Where are the horses?\", the answers \"mountain\" and \"mountains\" are both correct. We adopt the method in [28] and use an LLM to determine if the program output is correct (details in §G). If there is more than one correct program per question, we select the top scoring one, according to the scores provided by the program-generating LLM " + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "content": ". If no programs pass the test, the rated sample is not wasted—we simply use the correct answer " + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 112, + 289, + 409 + ], + "type": "text", + "content": " as supervision in our fine-tuning stage, without an associated CoT. When no human-rated answer is available, we directly use the top scoring executable program1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": "Converting program execution traces into chain-of-thought rationales. After the filtering step, for each visual input " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": ", we will have selected at most one program " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": " together with its execution result " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": " and trace " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": ". Since most existing VLMs have been pre-trained with text in natural language and not code, we use an LLM to rewrite the execution trace " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": " into a natural language CoT " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": " for our VLM distillation. Some examples are shown in §B. Concretely, similar to prior work [23, 24], we hand-craft 20 examples of how an input " + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "inline_equation", + "content": "(q, z, t)" + }, + { + "bbox": [ + 46, + 413, + 289, + 567 + ], + "type": "text", + "content": " can be converted into a CoT, and use these as few-shots for prompting PaLM-2 [3], which performs in-context learning and generates CoTs for new examples. We include a concrete example in §G." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 573, + 173, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 173, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 173, + 586 + ], + "type": "text", + "content": "3.2. Distilling step-by-step" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 591, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 662 + ], + "type": "text", + "content": "In this step, we fine-tune a backbone VLM with the training data generated in §3.1, distilling the knowledge and reasoning steps of our generated programs into a single end-to-end model. We do so in a multitask fashion, where the VLM is simultaneously fine-tuned on data synthesized for multiple types tasks (e.g., free-form VQA, multiple choice)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 663, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 288, + 687 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 663, + 288, + 687 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 663, + 288, + 687 + ], + "type": "text", + "content": " represent our VLM model. While the same VLM architecture could solve all these tasks, it needs to be prompted" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 72, + 548, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 72, + 548, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 548, + 179 + ], + "type": "text", + "content": "differently to adapt to the task. Following prior work [39], we manually design instructions for each task. For example, for free-form VQA queries, the instruction is \"Answer with a single word or phrase\", while for multiple-choice queries we use \"Answer with the option letter from the given choices directly\". During fine-tuning, we combine the training samples generated for all tasks into a single dataset, so to account for the different types of tasks, we augment each sample with its corresponding task-specific prompt " + }, + { + "bbox": [ + 309, + 72, + 548, + 179 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 309, + 72, + 548, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": "Using these instructions along with the data generated in §3.1, we put together the training dataset " + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(i_j,q_j,\\hat{y}_j,c_j,p_j)\\}_{j = 1}^N" + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": " is the total number of samples, " + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "inline_equation", + "content": "i_j" + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": " is the visual input, " + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": " is the textual query, " + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": " is the visual program output2, and " + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 305, + 180, + 547, + 239 + ], + "type": "text", + "content": " is the CoT rationale." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "content": "We train " + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "content": " to minimize a loss for predicting both the label and the rationale. As shown in the red box of Fig. 2, similar to [21], we treat predicting the output label " + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "inline_equation", + "content": "\\hat{y}_j" + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "content": " and the rationale " + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "content": " as two separate optimization goals. However, since VLMs are open-ended text generation models, they need additional prompting to indicate whether we want a short answer or a long answer that includes a rationale. Therefore, we append the suffix " + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "inline_equation", + "content": "s_c =" + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "content": " \"Explain the rationale to answer the question\" at the end of the prompt for generating a CoT, and use the task instruction " + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 305, + 239, + 547, + 370 + ], + "type": "text", + "content": " for short answers. Our optimization objective is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 380, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 380, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 321, + 380, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\mathcal {L} _ {\\text {r a t i o n a l e}} (1) \\\\ = \\sum_ {j = 1} ^ {N} \\ell \\left(f \\left(i _ {j}, q _ {j}, p _ {j}\\right), \\hat {y} _ {j}\\right) + \\ell \\left(f \\left(i _ {j}, q _ {j}, s _ {c}\\right), c _ {j}\\right) (2) \\\\ \\end{array}", + "image_path": "1b74a08270d115234dce5b0cb20edd556ffa0f98feeb6297e1847dadb3664ba4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "content": " is the cross entropy-loss normalized by sequence length to ensure that the labels (typically short) and rationales (typically long) have similar weights. " + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{rationale}}" + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "content": " both teaches the VLM to generate faithful reasoning steps similar to program execution traces, and carries more information beyond the label " + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "content": " that also helps the VLM in better predicting the label. During test time, the rationale generation is not required. We can adjust the task instruction to directly get the short output label using " + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "content": ", and using " + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 305, + 439, + 547, + 559 + ], + "type": "text", + "content": " to get the human-interpretable reasoning steps if needed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 570, + 388, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 388, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 388, + 583 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 590, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 547, + 685 + ], + "type": "text", + "content": "In this section, we demonstrate VPD's effectiveness by using it to train a generalist VLM. We attempt this for two VLMs with different scales, PaLI-3 (5B) [11] and PaLI-X (55B) [10]. Detailed experimental setups are given in §4.1. Qualitatively, models fine-tuned with VPD exhibit the ability to reason step-by-step like a program, as illustrated in Fig. 3, and supported by human evaluation results (§4.3). Quantitative results show that our PaLI-3-VPD and PaLI-X-VPD" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2When labeled data is available, this is equivalent to the ground truth label " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " due to our filtering strategy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "text", + "content": "Studying ways to approximate program correctness strategies for unlabeled data is an interesting direction for future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9593" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 62, + 274, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 62, + 274, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 62, + 274, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 62, + 274, + 201 + ], + "type": "image", + "image_path": "279813154032f56df66d0dcf9989edf3655182e43c8d8b3529479ce489ffafdc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 289, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 251 + ], + "type": "text", + "content": "Figure 3. Outputs of PaLI-X trained with VPD. Training with execution traces of filtered programs improves the model's ability to count, understand spatial relations, and reason compositionally. The images are from Visual Genome [32]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 260, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 289, + 344 + ], + "type": "text", + "content": "achieve new SOTA on a broad range of tasks, on both generalist and per-task fine-tuning settings (§4.2). We also conduct a detailed analysis of the source of performance gain (§4.2). Finally, we conduct human evaluation on the quality of the rationales generated by our models, and compare it with rationales generated by an instruction-tuned model trained without VPD (§4.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 353, + 162, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 162, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 162, + 366 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 375, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 448 + ], + "type": "text", + "content": "Backbone models. We use two state-of-the-art VLMs, PaLI-3 (5B) [11] and PaLI-X (55B) [10] as our base models. Both take images and text as input, and generate text as output. For simplicity, we further refer to them as \"PaLI model\" when we discuss steps performed with each of them individually." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "type": "text", + "content": "Data for Generalist Models. We fine-tune the pre-trained PaLI model on two types of datasets to make it a generalist VLM—that is, a model that performs relatively well on any task without further training on that specific task:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 499, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 46, + 499, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 582 + ], + "type": "text", + "content": "(1) Multimodal Instruction-Tuning (MMIT) tasks, created in the spirit of Self-Instruct [64]. An LLM is prompted with image captions, and generates task inputs and the desired outputs about the corresponding image. Details of MMIT tasks are covered in [65]. Note that these tasks cover a wide variety of common use cases, but do not include the specific in-domain tasks used in this work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 583, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 288, + 714 + ], + "type": "text", + "content": "(2) Academic task-oriented VQA tasks. Since image captions contain only a coarse description of visual information in the image, and may miss details that are important for solving the task. Additionally, LLM are likely to hallucinate during this data curation process. To further boost the accuracy of our VLMs, we also fine-tune the PaLI models with academic task-oriented VQA tasks. The data mixture covers subsets of a wide variety of VQA tasks, including general VQA (VQAv2 [19]), optical character recognition (OCRVQA [48]), compositional questions and reasoning (GQA [26]), counting (TallyQA [1]), and VQA that involves" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "external knowledge (OK-VQA [45] and A-OKVQA [52]). The tasks contain a textual query and a short expected label. We use the pipeline in §3.1 to synthesize CoT reasoning steps for these labels, and tune the PaLI model with these the loss in Equation 1. Notice that sometimes the pipeline fails to find a program that generates the correct answer. In that case, we set " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{rationale}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " to 0 and only keep " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{label}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": ". §4.2 shows how many programs are kept after the filtering stage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 173, + 547, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 234 + ], + "type": "text", + "content": "Data for specialist models. While fine-tuning the generalist model with VPD, we only use a subset of each task's training data. To evaluate our model's ability on each individual task, we continue fine-tuning PaLI-3-VPD and PaLI-X-VPD on each individual task on the training splits." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 239, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 334 + ], + "type": "text", + "content": "Training setup. Both PaLI-3 and PaLI-X follow an encoder-decoder architecture, where images are encoded into visual tokens via a visual encoder and then passed into a UL2 model. Due to resource constraint, we use LoRA [22] to fine-tune both PaLI models. Specifically, we add LoRA weights on each linear layer in the attention blocks and the MLP blocks for both the encoder and decoder in the UL2 transformer. More training details are in §C." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 340, + 548, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 548, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 548, + 541 + ], + "type": "text", + "content": "Evaluation setup. We evaluate our models on a wide range of tasks, including various VQA tasks and recent zero-shot VLM benchmarks. Noted that A-OKVQA contains two kinds of questions, multiple-choice (MC) and direct answer (DA). We report results on both. TallyQA [1] contains complex counting questions that involve object relationships, attribute identification, and reasoning to get the correct answer. It contains two evaluation partitions, simple and complex. TextVQA [55] focuses reading texts. We include it to evaluate our models on zero-shot tasks. In addition to VQA tasks, we also test our models on two popular VLM benchmarks. POPE [36] focuses on VLM hallucination, containing binary questions of whether or not an object exists in the image. MMBench [40] is a robust and comprehensive VLM benchmark testing a range of fine-grained abilities (e.g., object localization, attribute recognition, spatial relationship). Details of the evaluation sets and metrics are in §C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 548, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 713 + ], + "type": "text", + "content": "Baselines. We refer to the two PaLI models fine-tuned with VPD as PaLI-3-VPD and PaLI-X-VPD, and compare them with various baselines. To evaluate the effectiveness of VPD, we experiment with removing the synthesized CoT from our data mixture, and train the PaLI models with the exact same hyper-parameters and steps. We call these models as PaLI-3 Instruct and PaLI-X Instruct. For a fair comparison, these models are also trained with the same supervised loss for predicting the ground truth answers, on the same images and textual queries as our VPD variant. Moreover, we also compare with the most recent SOTA vision-language models. These VLMs are initialized with pre-trained visual encoders and LLMs, and then trained with image-text pairs, LLM-generated data, and academic tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "9594" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 61, + 550, + 231 + ], + "blocks": [ + { + "bbox": [ + 47, + 61, + 550, + 231 + ], + "lines": [ + { + "bbox": [ + 47, + 61, + 550, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 61, + 550, + 231 + ], + "type": "table", + "html": "
Generalist ModelsVQAv2GQAOK-VQAA-OKVQATallyQATextVQAPOPEMMB
MCDASimp.Comp.
Prior generalist VLMs
Flamingo (80B) [2]82.0-57.8*----57.1--
MiniGPT-4 (Vicuna-13B) [79]-43.5-67.2-----42.3
InstructBLIP (Vicuna-13B) [14]-49.5*---75.2*57.5*50.7*78.944.0
Shikra (Vicuna-13B) [7]77.4-47.2-----84.758.8
Qwen-VL (9.7B) [4]78.859.358.6--82.6*65.8*63.8-38.2
Qwen-VL-Chat (9.7B) [4]78.257.556.6--81.1*64.0*61.5-60.6
mPLUG-Owl2 (8.2B) [74]79.456.157.7----54.3*86.264.5
LLaVA-1.5 (Vicuna-13B) [38]80.063.3---76.9*65.4*61.3*85.967.7
Our instruction-tuned baselines
PaLI-3-Instruct (5B)79.959.756.778.357.681.970.463.3*87.768.6
PaLI-X-Instruct (55B)83.663.364.384.161.585.575.465.0*88.975.0
Our visual program distillation models
PaLI-3-VPD (5B)80.461.357.578.556.583.170.963.7*88.669.0
PaLI-X-VPD (55B)83.964.964.684.562.786.276.665.4*88.876.2
", + "image_path": "a7f954b3b4fbdb54a373f64988542be82ee7d54d51f968fdd94f0edebc74cdf0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 276 + ], + "type": "text", + "content": "Table 1. Comparison with SOTA generalist VLMs. PaLI-X-VPD outperforms all prior models. VPD improves performance on 8/9 tasks for both PaLI-3 and PaLI-X, and is particularly effective on counting questions (TallyQA), compositional questions (GQA), and the comprehensive benchmark (MMBench). Underline indicates when PaLI-3-VPD outperforms the Instruct version. * marks tasks unseen during training. POPE and MMBench are zero-shot benchmarks for all models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 288, + 163, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 163, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 163, + 300 + ], + "type": "text", + "content": "4.2. Quantitative results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 306, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 306, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 306, + 290, + 542 + ], + "type": "text", + "content": "Generalist model. Table 1 compares VPD with the baselines discussed in §4.1. We infer all answers by open-ended generation with the prompt \"Answer with a single word or phrase,\" using greedy decoding without any constraint on the model's output space. For multiple-choice questions, we run inference with the prompt \"Answer with the option letter from the given choices directly.\" and generate the option letter. As Table 1 shows, PaLI-X-VPD sets the new state-of-the-art on all benchmarks. Compared with prior generalist VLMs, it achieves significant improvement on MMBench (+8.5), TallyQA complex (+9.8), and A-OKVQA(+9.5 compared with specialist SOTA[14]). PaLI-3-VPD, despite its small architecture size (5B), outperforms all prior models on VQAv2, TallyQA, POPE, and MMBench, including much larger models that use Vicuna-13B as backbones [7, 14, 38]. However, its performance on knowledge-based VQA tasks does not outperform the prior SOTA. Our hypothesis for this is that its language model (3B UL2) is too small to contain all the necessary knowledge needed for correctly solving these tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "text", + "content": "When compared to their Instruct variants, both PaLI-3-VPD and PaLI-X-VPD outperform on 11/12 tasks. Specifically, for PaLI-X, VPD obtains a " + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "inline_equation", + "content": "+1.6" + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "text", + "content": " improvement on GQA (which is heavily focused on compositional questions, spatial relationship, and localization), " + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "text", + "content": " on TallyQA for complex questions, and " + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 45, + 543, + 288, + 650 + ], + "type": "text", + "content": " on MM-Bench. These results suggest that VPD is a more effective method for creating instruction-tuning data which enables VLMs to improve their visual reasoning ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "Per-Task Fine-Tuning (Specialist). The results for the specialist models are shown in Table 2. PaLI-X-VPD sets a new SOTA on all benchmarks. Note how the specialist models tend to have higher scores than the generalist one. We propose several hypotheses for this performance gain:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 288, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 547, + 467 + ], + "type": "text", + "content": "(1) As shown in Table 2, the score gap tends to be larger on free-form VQA tasks. This may be due the fact that human annotations on these datasets have ambiguities. For example, for the question \"Who is looking up?\", GQA [26] labels are \"man\" or \"woman\" while OK-VQA [45] have more detailed labels, for example, \"worker\" or \"cook\". Per-task fine-tuning alleviates this annotation ambiguity and lets the model focus on the annotation style of that task. For multiple-choice and counting tasks, the answer has less ambiguity, and the score gaps are much smaller. (2) The performance gap between PaLI-3 (5B) specialist and generalist is larger than that of PaLI-X. We hypothesize that this shows models with larger scale has more multi-task capacity; (3) Per-task fine-tuning adds more in-domain training data, which generally improves performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "text", + "content": "Analysis: Sampling multiple programs is key to good data generation. Fig. 5 shows the success rate of finding at least one program that passes the filtering stage, when the LLM generates the top-1 or top-5 programs, respectively. There is a dramatic increase in success rate from 1 program to 5: " + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "inline_equation", + "content": "+45\\%" + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "text", + "content": " on GQA and A-OKVQA, " + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "inline_equation", + "content": "+33\\%" + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "text", + "content": " on OK-VQA, and " + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "text", + "content": " on TallyQA. This design choice greatly improves our data synthesis efficiency, and, as a consequence, adding more CoT data that requires complex reasoning in our training set. We also conduct an analysis in " + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "inline_equation", + "content": "\\S E" + }, + { + "bbox": [ + 304, + 471, + 548, + 613 + ], + "type": "text", + "content": " to compare the performance of VPD models with that of the visual programs they are distilled from." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 620, + 479, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 620, + 479, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 620, + 479, + 632 + ], + "type": "text", + "content": "4.3. Human evaluation on rationales" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 637, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 637, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 637, + 547, + 686 + ], + "type": "text", + "content": "In this section, we focus on the quality of model outputs. We performed this analysis using human annotators, who are asked to evaluate both the correctness of the final answer and the quality of the rationale behind it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Models. We compare PaLI-X-VPD with PaLI-X Instruct. Among possible baselines, we chose PaLI-X" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "9595" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 70, + 512, + 208 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 512, + 208 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 512, + 208 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 512, + 208 + ], + "type": "table", + "html": "
Specialist ModelsGQA test-devOK-VQA valA-OKVQATallyQA
Multi-choice valtestDirect Answer valtestSimple testComplex test
InstructBLIP (Vicuna-7B) [14]-62.175.773.464.062.1--
PaLI-3 (5B) [11]-60.1----83.370.5
PaLI (17B) [9]-64.5----81.770.9
CogVLM (17B) [63]65.264.7------
PaLI-X (55B) [10]-66.1----86.075.6
PaLI-3-VPD (5B) generalist61.357.578.5-56.5-83.170.9
PaLI-3-VPD (5B) specialist64.760.379.776.565.563.683.370.8
PaLI-X-VPD (55B) generalist64.964.684.5-62.7-86.276.6
PaLI-X-VPD (55B) specialist67.366.885.280.471.168.286.276.4
", + "image_path": "8d0ec7774655a7ee275244cc4fc1113ed0fbad2ef9aac8d0aea911e9a1d8f267.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 211, + 516, + 222 + ], + "lines": [ + { + "bbox": [ + 75, + 211, + 516, + 222 + ], + "spans": [ + { + "bbox": [ + 75, + 211, + 516, + 222 + ], + "type": "text", + "content": "Table 2. Comparison of per-task fine-tuning results of specialist models. PaLI-X-VPD sets a new SOTA for all the tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 97, + 228, + 203, + 320 + ], + "blocks": [ + { + "bbox": [ + 97, + 228, + 203, + 320 + ], + "lines": [ + { + "bbox": [ + 97, + 228, + 203, + 320 + ], + "spans": [ + { + "bbox": [ + 97, + 228, + 203, + 320 + ], + "type": "image", + "image_path": "1a595d7b0c04990ac23bb4d038af1cadfa30941ed3dfcf7d9aabd3157eb1650e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "text", + "content": "Figure 4. Human evaluation results assessing answer and rational quality for PaLI-X Instruct and PaLI-X VPD: (a) percentage of answers that are correct and have explanations; (b) rationale factuality and consistency for the answers that contain an explanation; (c) preference between the two models, when aggregating across all samples (\"All\") and across those with correct answers (\"When correct\")." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 207, + 228, + 314, + 320 + ], + "blocks": [ + { + "bbox": [ + 207, + 228, + 314, + 320 + ], + "lines": [ + { + "bbox": [ + 207, + 228, + 314, + 320 + ], + "spans": [ + { + "bbox": [ + 207, + 228, + 314, + 320 + ], + "type": "image", + "image_path": "4cf7da159c3d1a5b097eea3a36149d00c37d3020592f15ed4d6c98a346b0dd6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 492, + 320 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 492, + 320 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 492, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 492, + 320 + ], + "type": "image", + "image_path": "e85a5d609dee9a1abdef080c7e96d0ccbc3273eaf73395e8d6d4be058b0979e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 365, + 272, + 445 + ], + "blocks": [ + { + "bbox": [ + 60, + 365, + 272, + 445 + ], + "lines": [ + { + "bbox": [ + 60, + 365, + 272, + 445 + ], + "spans": [ + { + "bbox": [ + 60, + 365, + 272, + 445 + ], + "type": "image", + "image_path": "c4708f067e55637075c087e96be6e10935beed8744c236b7f28a61bd46232ac3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 448, + 288, + 471 + ], + "lines": [ + { + "bbox": [ + 46, + 448, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 288, + 471 + ], + "type": "text", + "content": "Figure 5. Success rate of top-1 program and top-5 programs on the training set during our data synthesis process." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 555 + ], + "type": "text", + "content": "Instruct because it is trained to generate long-form answers [65], which allows us to assess if it is the quality of PaLI-X-VPD's answers that the annotators prefer, rather than its length. Since PaLI-X Instruct is also instruction-tuned, it can be prompted to provide long answers to alleviate this confounder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 559, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 289, + 715 + ], + "type": "text", + "content": "Annotation protocol. We run inference with each of the two models on a combination of 600 samples from GQA (test-dev), A-OKVQA (val), and TallyQA (Simple and Complex) and record their answers and rationales. We then ask 3 human annotators to evaluate each model answer. We use prior work from natural language processing (NLP) as inspiration, and build upon it for selecting the evaluation criteria [75]. Given an image and a query, for each model-generated rationale, we ask human annotators to score the model answers along the following criteria: (1) correctness—is the final answer correct? (2) explainability—does the model explain its rationale for reaching the final answer? (3) factuality—is every step in the rationale factually cor" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 370, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 547, + 478 + ], + "type": "text", + "content": "rect (with respect to the image and external knowledge)? (4) consistency—does step and final answer logically follow from the previous ones? Note that a rationale may have the wrong answer while being consistent. We also conduct a side-by-side comparison, and ask annotators which of the two answers—the one provided by PaLI-X-VPD or by PaLI-X Instruct—they prefer in terms of quality of the answer and explanation. More details about the annotation protocol are in §D." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "type": "text", + "content": "Human evaluation results. The results of the evaluation are first averaged among the human raters per sample, then aggregated across samples. Our PaLI-X-VPD model far outperforms PaLI-X Instruct along all criteria. Fig. 4 (a) shows the correctness of the final answer provided by each model (i.e. accuracy), and the proportion of the samples where the model provides a CoT rationale to explain its answer. PaLI-X-VPD's gain in accuracy of " + }, + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "type": "inline_equation", + "content": "+16.7\\%" + }, + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "type": "text", + "content": " is even more impressive than in evaluations in §4.2 based on benchmark labels—this is because human annotators are better able to assess correctness for ambiguous questions with different possible interpretations (e.g., the model's answer \"In the living room.\" to the question \"Where is the couch located?\" was considered correct by annotators, even when the benchmark answer was \"On the right\"). Moreover, the explainability results confirm that our model is able to explain its own answer on " + }, + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "type": "inline_equation", + "content": "+24\\%" + }, + { + "bbox": [ + 304, + 488, + 548, + 715 + ], + "type": "text", + "content": " more samples than the instruction tuned model. Additionally, Fig. 4 (b) shows impressive rationale quality: among the samples where" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9596" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": "an explanation is provided, PaLI-X-VPD's rationales are factual " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "87.2\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " of times, and consistent " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "97.8\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " of times, a gain of " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "+14.6\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": ", respectively, when compared to PaLI-X Instruct." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "text", + "content": "We also asked the annotators which of the two answers given by the two models they prefer. We aggregated these results in two ways: (1) across all samples, (2) across the samples where both models answered the question correctly. The results in Fig. 4 (c) show that PaLI-X-VPD is preferred on " + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "text", + "content": " more samples than PaLI-X Instruct in the \"All\" case, and on " + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "text", + "content": " more samples when both are correct. This suggests that even when PaLI-X-VPD makes mistakes, it still provides a better answer. Such examples are shown in " + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "inline_equation", + "content": "\\S D" + }, + { + "bbox": [ + 46, + 120, + 288, + 248 + ], + "type": "text", + "content": ". These results confirm that a model fine-tuned with VPD leads to more faithful and consistent answers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 261, + 246, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 246, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 246, + 274 + ], + "type": "text", + "content": "5. Experiments on content moderation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 411 + ], + "type": "text", + "content": "Prior experiments focus on training generalist VLMs. Here, we explore the effectiveness of VPD at quickly adapting models to real-world applications from a different domain than the training data [e.g., 56]. We experiment on Hateful Memes [29], a content moderation dataset where the task is to classify if a meme contains hateful content. The target labels are \"yes\" or \"no\", and models are evaluated in terms of classification accuracy and AUC-ROC. We experiment with two settings: supervised, in which the models are trained on the provided training set with 8,500 labels, and unsupervised, in which no human labels are provided." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 60, + 416, + 270, + 520 + ], + "blocks": [ + { + "bbox": [ + 60, + 416, + 270, + 520 + ], + "lines": [ + { + "bbox": [ + 60, + 416, + 270, + 520 + ], + "spans": [ + { + "bbox": [ + 60, + 416, + 270, + 520 + ], + "type": "table", + "html": "
ModelAccAUC-ROC
Unsupervised / Zero-Shot Methods
Generated Programs (ours)69.770.1
Supervised Methods
VisualBERT [35]69.575.4
Flamingo (80B) [2]-86.6
Previous SOTA [46]78.886.7
PaLI-X-VPD (label-only FT)77.688.0
PaLI-X-VPD (specialist w/ CoT)80.889.2
Human [29]84.782.7
", + "image_path": "a22b52d3d7d759a1ba81602fe62b5a5873561289611668bc83917592e0010f8a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 525, + 287, + 557 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 557 + ], + "type": "text", + "content": "Table 3. Results on Hateful Memes [29] seen test set. We improve SOTA for both zero-shot and supervised settings. Unsupervised PaLI-X-VPD outperforms strong supervised baselines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 564, + 168, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 168, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 168, + 578 + ], + "type": "text", + "content": "5.1. Unsupervised setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 582, + 192, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 192, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 192, + 594 + ], + "type": "text", + "content": "We experiment with three methods:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 594, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "content": "1. Generated Program: The program generated by PaLM-2 [3] for solving this task (shown in §F) consists of following main steps: (1) get image description with PaLI-X [10]; (2) use an OCR tool to extract embedded texts; (3) given the image description and OCR texts, ask PaLM-2 to explain if this meme is hateful." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 289, + 714 + ], + "type": "text", + "content": "2. PaLI-X-VPD (generalist): In a zero-shot setting, we directly prompt our PaLI-X-VPD (generalist) with: \"The text is " + }, + { + "bbox": [ + 47, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": ". Is this a hateful meme?\" We compute the probability of PaLI-X-VPD generating \"yes\"" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 507, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 507, + 83 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 507, + 83 + ], + "type": "text", + "content": "or \"no\" and measure accuracy and AUC-ROC." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 84, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 144 + ], + "type": "text", + "content": "3. PaLI-X-VPD (specialist with zero-shot CoT): We follow our VPD pipeline and convert the execution traces of the program into CoTs, and then fine-tune PaLI-X-VPD to output these CoTs. Since no groundtruth labels are available, no filtering is done during the process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 148, + 547, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 547, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 547, + 279 + ], + "type": "text", + "content": "VPD significantly improves performance even when no labels are available. As shown in Table 3, PaLI-X-VPD (generalist) outperforms all other VLMs (61.4%). Interestingly, the generated programs themselves get much higher accuracy (69.7%) than on the previous datasets, perhaps because PaLM-2 is better suited at analyzing a meme than typical VQA datasets. Moreover, our PaLI-X-VPD (specialist) sets a new zero-shot SOTA on Hateful Memes, and even outperforms supervised VisualBERT [35]. Manual inspection of model outputs show a great deal of similarity with the generated code. See examples in §F." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 289, + 414, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 414, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 414, + 301 + ], + "type": "text", + "content": "5.2. Supervised setting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 308, + 474, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 474, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 474, + 319 + ], + "type": "text", + "content": "We fine-tune PaLI-X-VPD in two ways:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 319, + 547, + 390 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 319, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 547, + 355 + ], + "type": "text", + "content": "1. Label-only fine-tuning: To establish a strong baseline, we first experiment with the traditional supervised setting, where we fine-tune the model to output \"yes\" or \"no\"." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 355, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 355, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 355, + 547, + 390 + ], + "type": "text", + "content": "2. PaLI-X-VPD (specialist): We use the complete VPD pipeline to train this model. We select an execution trace that leads to the correct label, and use it to tune our VLM." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "type": "text", + "content": "Results. Our PaLI-X-VPD (specialist) sets a new SOTA for this task, with an accuracy of " + }, + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "type": "inline_equation", + "content": "80.8\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "type": "text", + "content": " and AUC-ROC of " + }, + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "type": "inline_equation", + "content": "89.2\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 466 + ], + "type": "text", + "content": ". It outperforms the label-only fine-tuning baseline and significantly improves the SOTA metrics, achieving nearly human-level accuracy, and superhuman AUC-ROC." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 479, + 528, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 528, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 528, + 492 + ], + "type": "text", + "content": "6. Conclusion, limitations, and future work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 500, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we introduced VPD, a framework for distilling the reasoning abilities of LLMs along with the capabilities of vision tools into VLMs. VPD synthesizes training data for VLMs by generating programs that can leverage external tools. We used this technique to fine-tune some of the best existing VLMs (PaLI-3 and PaLI-X) and established new SOTA results on 8 classical VQA and 2 zero-shot multimodal benchmarks. According to human evaluations, VPD-tuned models provide more accurate answers and better rationales. Experiments on the Hateful Memes dataset show how VPD can also adapt models to new real-world domains, even when no labeled data is available, also establishing a new SOTA on this dataset. From our experiments we learnt that both a strength and a weakness of VPD is the quality of the generated programs. Better programs lead to bigger gains with VPD. We discuss multiple ideas on how to improve the program generation in Appendix H, including using additional tools, agents, and fact-checking for multimodal CoTs." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9597" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "type": "text", + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 1, 5, 14, 15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 233 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022. 1, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 234, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 234, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 53, + 234, + 288, + 603 + ], + "type": "text", + "content": "[3] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby Ambrose Slone,Daniel Smilkov David R. So Daniel Sohn Simon Tokumine,Dasha Valter,Vijay Vasudevan,Kiran Vodrahalli,Xuezhi Wang,Pidong Wang,Zirui Wang,Tao Wang John Wieting,Yuhuai Wu,Kelvin Xu,Yunhan Xu,Linting Xue,Pengcheng Yin,Jiahui Yu,Qiao ZhangSteven Zheng Ce ZhengWeikang ZhouDenny ZhouSlav Petrov and Yonghui Wu.Palm 2 technical report,2023.134813" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 604, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 604, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 604, + 288, + 657 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 128 + ], + "type": "text", + "content": "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigptv2: large language model as a unified interface for vision-language multi-task learning, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 129, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 547, + 161 + ], + "type": "text", + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 547, + 206 + ], + "type": "text", + "content": "[8] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 208, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 547, + 316 + ], + "type": "text", + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 1, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 318, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 547, + 469 + ], + "type": "text", + "content": "[10] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, AJ Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali-x: On scaling up a multilingual vision and language model, 2023. 2, 3, 4, 5, 7, 8, 15, 16" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 472, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 547, + 547 + ], + "type": "text", + "content": "[11] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, Daniel Salz, Xi Xiong, Daniel Vlasic, Filip Pavetic, Keran Rong, Tianli Yu, Daniel Keysers, Xiaohua Zhai, and Radu Soricut. Pali-3 vision language models: Smaller, faster, stronger, 2023. 1, 4, 5, 7, 15, 16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 548, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 602 + ], + "type": "text", + "content": "[12] Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, et al. Binding language models in symbolic languages. arXiv preprint arXiv:2210.02875, 2022.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 604, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 547, + 713 + ], + "type": "text", + "content": "[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robin-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9598" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 72, + 289, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 289, + 182 + ], + "type": "text", + "content": "son, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 183, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 288, + 226 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 228, + 289, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 289, + 379 + ], + "type": "text", + "content": "[15] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaeldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 288, + 413 + ], + "type": "text", + "content": "[16] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 1, 2, 3, 20, 24" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 415, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 491 + ], + "type": "text", + "content": "[17] Ruofei Du, Eric Turner, Maksym Dzitsiuk, Luca Prasso, Ivo Duarte, Jason Dourgarian, Joao Afonso, Jose Pascoal, Josh Gladstone, Nuno Cruces, Shahram Izadi, Adarsh Kowdle, Konstantine Tsotsos, and David Kim. DepthLab: Real-Time 3D Interaction With Depth Maps for Mobile Augmented Reality. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. ACM, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 491, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 545 + ], + "type": "text", + "content": "[18] Weixi Feng, Wanrong Zhu, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. *ArXiv*, abs/2305.15393, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "text", + "content": "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "text", + "content": "[20] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 647, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 700 + ], + "type": "text", + "content": "[21] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, 2023. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "type": "text", + "content": "[22] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "text", + "content": "Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 5, 16" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 107, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 149 + ], + "type": "text", + "content": "[23] Yushi* Hu, Hang* Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "text", + "content": "[24] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A. Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2627-2643, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 218, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 547, + 261 + ], + "type": "text", + "content": "[25] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 262, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 547, + 316 + ], + "type": "text", + "content": "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 5, 6, 14, 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 318, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 547, + 567 + ], + "type": "text", + "content": "[27] Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir Ghaemmaghami, Rajendra Gottipati, William Gulland, Robert Hagemann, C. Richard Ho, Doug Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander Kaplan, Harshit Khaitan, Daniel Killebrew, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon, James Law, Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana Maggiore, Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy Nix, Thomas Norrie, Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Matt Ross, Amir Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham, Jed Souter, Dan Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma, Erick Tuttle, Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon. Indatacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture. ACM, 2017. 16" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 569, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 547, + 602 + ], + "type": "text", + "content": "[28] Ehsan Kamalloo, Nouha Dziri, Charles L. A. Clarke, and Davood Rafiei. Evaluating open-domain question answering in the era of large language models, 2023. 4, 23" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 602, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 547, + 645 + ], + "type": "text", + "content": "[29] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes, 2020. 2, 8, 18, 19" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 647, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 667 + ], + "type": "text", + "content": "[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 16" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 24" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9599" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual genome: Connecting language and vision using crowdsourced dense image annotations, 2016. 3, 5, 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "type": "text", + "content": "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2023. 24" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 248 + ], + "type": "text", + "content": "[35] Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A Simple and Performant Baseline for Vision and Language. ArXiv, abs/1908.03557, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 249, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 288, + 282 + ], + "type": "text", + "content": "[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 5, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 337 + ], + "type": "text", + "content": "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 360 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 394, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 437 + ], + "type": "text", + "content": "[40] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2023. 2, 5, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 438, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 288, + 480 + ], + "type": "text", + "content": "[41] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. ArXiv, abs/2206.08916, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "type": "text", + "content": "[42] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "type": "text", + "content": "[43] Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. Faithful chain-of-thought reasoning, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 615, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 288, + 668 + ], + "type": "text", + "content": "[45] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6, 15" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[46] Jingbiao Mei, Jinghong Chen, Weizhe Lin, Bill Byrne, and Marcus Tomalin. Improving hateful memes detection via learning hatefulness-aware embedding space through retrieval-guided contrastive learning, 2023. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 95 + ], + "type": "text", + "content": "[47] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 546, + 128 + ], + "type": "text", + "content": "[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5, 15" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 484, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 484, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 484, + 140 + ], + "type": "text", + "content": "[49] OpenAI. Gpt-4 technical report, 2023. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 546, + 173 + ], + "type": "text", + "content": "[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 175, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 546, + 217 + ], + "type": "text", + "content": "[51] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 218, + 546, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 218, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 218, + 546, + 261 + ], + "type": "text", + "content": "[52] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 5, 14, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 262, + 546, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 546, + 305 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 546, + 305 + ], + "type": "text", + "content": "[53] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 307, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 307, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 546, + 340 + ], + "type": "text", + "content": "[54] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 341, + 546, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 546, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 546, + 393 + ], + "type": "text", + "content": "[55] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 5, 15" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 396, + 546, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 546, + 460 + ], + "type": "text", + "content": "[56] Otilia Stretcu, Edward Vendrow, Kenji Hata, Krishnamurthy Viswanathan, Vittorio Ferrari, Susan Tavakkol, Wenlei Zhou, Aditya Avinash, Emming Luo, Neil Gordon Alldrin, et al. Agile modeling: From concept to classifier in minutes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22323-22334, 2023. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 461, + 546, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 461, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 308, + 461, + 546, + 505 + ], + "type": "text", + "content": "[57] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 506, + 546, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 506, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 506, + 546, + 559 + ], + "type": "text", + "content": "[58] Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. UL2: Unifying Language Learning Paradigms, 2022. 15" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 561, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 561, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 561, + 546, + 712 + ], + "type": "text", + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9600" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 289, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 289, + 137 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 289, + 137 + ], + "type": "text", + "content": "Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "type": "text", + "content": "[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 274 + ], + "type": "text", + "content": "[62] Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. Scott: Self-consistent chain-of-thought distillation, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "text", + "content": "[63] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 1, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 374 + ], + "type": "text", + "content": "[64] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "type": "text", + "content": "[65] Yaqing Wang, Jialin Wu, Tanmaya Dabral, Jiageng Zhang, Geoff Brown, Chun-Ta Lu, Frederick Liu, Yi Liang, Bo Pang, Michael Bendersky, and Radu Soricut. Non-intrusive adaptation: Input-centric parameter-efficient fine-tuning for versatile multimodal modeling, 2023. 2, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 476 + ], + "type": "text", + "content": "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "text", + "content": "[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation. 2023. 24" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[68] Ge Zhang Yao Fu Wenhao Huang Huan Sun Yu Su Wenhu Chen Xiang Yue, Xingwei Qu. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 634 + ], + "type": "text", + "content": "[69] Yiheng Xu, Hongjin Su, Chen Xing, Boyu Mi, Qian Liu, Weijia Shi, Binyuan Hui, Fan Zhou, Yitao Liu, Tianbao Xie, Zhoujun Cheng, Siheng Zhao, Lingpeng Kong, Bailin Wang, Caiming Xiong, and Tao Yu. Lemur: Harmonizing natural language and code for language agents, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[70] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081–3089, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[71] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 483 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 139 + ], + "type": "text", + "content": "[72] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "text", + "content": "[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 24" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[74] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration, 2023. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 271 + ], + "type": "text", + "content": "[75] Xi Ye and Greg Durrett. The unreliability of explanations in few-shot prompting for textual reasoning. In Advances in Neural Information Processing Systems, pages 30378-30392. Curran Associates, Inc., 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "text", + "content": "[76] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 329, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 360 + ], + "type": "text", + "content": "[77] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 15" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 362, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 362, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 362, + 545, + 406 + ], + "type": "text", + "content": "[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 407, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 449 + ], + "type": "text", + "content": "[79] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "type": "text", + "content": "[80] Xuekai Zhu, Biqing Qi, Kaiyan Zhang, Xingwei Long, and Bowen Zhou. Pad: Program-aided distillation specializes large models in reasoning, 2023. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9601" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_content_list.json b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e888ef438d60592251ec071da9d5b894c02d5336 --- /dev/null +++ b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_content_list.json @@ -0,0 +1,1792 @@ +[ + { + "type": "text", + "text": "Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 138, + 161, + 833, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihao Yuan $^{1,2}$ , Jinke Ren $^{1,2}$ , Chun-Mei Feng $^{3}$ , Hengshuang Zhao $^{4}$ , Shuguang Cui $^{2,1}$ , Zhen Li $^{2,1*}$", + "bbox": [ + 135, + 205, + 839, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FNii, CUHKSZ $^{2}$ SSE, CUHKSZ $^{3}$ IHPC, A*STAR, Singapore $^{4}$ HKU", + "bbox": [ + 200, + 239, + 767, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 287, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Visual Grounding (3DVG) aims at localizing 3D object based on textual descriptions. Conventional supervised methods for 3DVG often necessitate extensive annotations and a predefined vocabulary, which can be restrictive. To address this issue, we propose a novel visual programming approach for zero-shot open-vocabulary 3DVG, leveraging the capabilities of large language models (LLMs). Our approach begins with a unique dialog-based method, engaging with LLMs to establish a foundational understanding of zero-shot 3DVG. Building on this, we design a visual program that consists of three types of modules, i.e., view-independent, view-dependent, and functional modules. These modules, specifically tailored for 3D scenarios, work collaboratively to perform complex reasoning and inference. Furthermore, we develop an innovative language-object correlation module to extend the scope of existing 3D object detectors into open-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot approach can outperform some supervised baselines, marking a significant stride towards effective 3DVG. Code is available at https://curryyuan.github.io/ZSVG3D.", + "bbox": [ + 109, + 316, + 473, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 637, + 233, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Visual Grounding (3DVG) aims to localize specific objects within 3D scenes by using a series of textual descriptions. This has become a crucial component in a variety of burgeoning applications, such as autonomous robotics [12, 53, 58], virtual reality [40, 55], and metaverse [10, 33]. For illustration, given a 3D scan in Figure 1(a) along with its description — It is the keyboard closest to the door, the goal of 3DVG is to accurately pinpoint the keyboard in the green box, while eliminating potential distractions such as tables and desks. Despite the apparent simplicity of this task for humans, it poses a significant challenge for machines due to their inherently limited perceptual capabilities.", + "bbox": [ + 109, + 661, + 470, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba5f103fe443dd646aa6fa4627d25c841b3e7ce9197e3635f39d6d0fdf63692a.jpg", + "image_caption": [ + "(a) Supervised 3D Visual Grounding" + ], + "image_footnote": [], + "bbox": [ + 504, + 287, + 857, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7178641d035cdc56f39dcfb55cf758c5f7657f527053888d08fbd2b208116e68.jpg", + "image_caption": [ + "(b) Zero-shot 3D Visual Grounding", + "Figure 1. Comparative overview of two 3DVG approaches, where (a) Supervised 3DVG involves input from 3D scans combined with text queries, guided by object-text pair annotations, (b) Zero-shot 3DVG identifies the location of target objects using programmatic representation generated by LLMs, i.e., target category, anchor category, and relation grounding, thereby highlighting its superiority in decoding spatial relations and object identifiers within a given space, e.g., the location of the keyboard (outlined in green) can be retrieved based on the distance between the keyboard and the door (outlined in blue)." + ], + "image_footnote": [], + "bbox": [ + 503, + 393, + 854, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional supervised 3DVG approaches [19, 60, 62] achieve this objective by leveraging the rich annotations in public datasets, such as ScanRefer [4] and Referit3D [1]. These approaches typically define 3DVG as a matching problem, generating possible objects via 3D detectors [22, 38], and identifying the best match by fusing the visual and textual features. While these approaches can yield precise results, the acquisition of sufficient annotations is prohibitively resource-intensive for real-world applications. Furthermore, these approaches are often constrained by the pre-defined vocabulary during training, making them suboptimal in open-vocabulary scenarios.", + "bbox": [ + 496, + 657, + 857, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel visual programming approach for 3DVG that integrates zero-shot learning and large language models (LLMs). Zero-shot", + "bbox": [ + 498, + 824, + 857, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 259, + 42, + 781, + 56 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 338, + 56, + 704, + 70 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 308, + 71, + 732, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 130, + 853, + 245, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20623", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning [65, 67, 70] can generalize across new categories by leveraging the pre-trained capabilities of CLIP [41] in the 3D domain. LLMs [34, 47, 49] can facilitate 3DVG due to their strong planning and reasoning capabilities. Regarding this, we first propose a vanilla version dialog with LLM. It describes the location and size of all objects in the scene and instructs the LLM to distinguish the object of interest through an interactive dialog. Despite the simplicity of the base approach, the inherent stochasticity and control limitations of LLMs make it hard to capture the view-dependent queries and decipher spatial relations in 3D space, which are the main challenges of 3DVG. To overcome this limitation, we further develop a new visual programming approach, as shown in Figure 1(b). It mainly consists of three steps: (1) generating a 3D visual program using LLMs, (2) interpreting the program into Python code, and (3) identifying the target bounding box by executing the code. To enhance the localization accuracy, we further introduce a novel language-object correlation (LOC) module capable of merging the geometric discernment of 3D point clouds with the fine-grained appearance acumen of 2D images.", + "bbox": [ + 109, + 125, + 472, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, contributions are summarized as follows:", + "bbox": [ + 129, + 416, + 459, + 429 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an innovative 3D visual programming approach. It eliminates the need for extensive object-text pair annotations required in supervised approaches.", + "- We transform the visual program to Python code by designing two types of modules, i.e., relation modules and LOC modules. The former explicitly defines the view-dependent and view-independent relations in 3D space, while the latter captures both the geometric and appearance information for open-vocabulary localization.", + "- We conduct extensive experiments on two popular datasets, i.e., ScanRefer [4] and Nr3d [1]. We for the first time evaluate the whole validation set rather than a few samples. The results demonstrate the superior performance of our approach, even comparable with existing supervised approaches." + ], + "bbox": [ + 112, + 430, + 470, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 112, + 648, + 240, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised 3DVG. 3DVG has received much attention in many emerging applications ranging from automatic robotics [12, 53, 58] to metaverse [10, 33]. On the one hand, densely-annotated datasets like ScanRefer [4] and Referit3D [1] can provide well-aligned object-text pairs for ScanNet [9]. On the other hand, most existing methods [6, 14, 42, 60, 62] treat 3DVG as a matching problem, where object identifiers [22, 38] are utilized to generate candidate objects and find the best-matching one by fusing visual and textual features. Building on this, [57, 63] attempt to explore the object attributes and relations between different proposals. Moreover, some works [7, 23] have also investigated 3D language pretraining using advanced techniques, such as mask modeling and contrastive learning on paired", + "bbox": [ + 109, + 672, + 472, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "object-caption data, followed by finetuning on downstream tasks. Additionally, NS3D [17] has employed CodeX [5] to generate hierarchical programs. However, it still needs annotations to train the neuro-symbolic networks, thus lacking open-vocabulary and zero-shot capabilities.", + "bbox": [ + 496, + 126, + 857, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Indoor 3D Scene Understanding. 3D scene understanding of indoor environments has been widely studied. In specific, the emergence of RBG-D scans datasets [9, 44, 52] greatly push the boundary of several tasks, including 3D object classification [36, 37], 3D object detection [31, 38], 3D semantic segmentation [8, 66], 3D instance segmentation [22, 43, 51], and so on [16, 59, 64]. However, these methods are often constrained to a closed set of semantic class labels, limiting their applicability in real-world scenarios. Recent progress in open-vocabulary image segmentation [13, 29] has inspired research into 3D scene understanding under the open-vocabulary setting. For instance, LERF [24] learns a language field inside NeRF [32] by volume rendering CLIP [41] features along training rays, enabling it to generate 3D relevancy maps for arbitrary language queries. OpenScene [35] extracts image features using 2D open-vocabulary segmentation models [13, 27], then trained a 3D network to produce point features aligned with multi-view fused pixel features. OpenMask3D [46] utilizes the closed-vocabulary network to generate instance masks while discarded the classification head. Despite these advancements, these methods still lack spatial and commonsense reasoning abilities.", + "bbox": [ + 496, + 195, + 857, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs for Vision-Language Tasks. Recent progress on LLMs has provided impressive zero-shot planning and reasoning abilities [34, 47, 49]. Advanced prompting technologies such as Least-to-Most [69], Think-Step-by-Step [26], and Chain-of-Thought [54] are proposed to elicit the capabilities of LLMs. These methods can understand human instructions, break complex goals into sub-goals, and control robot agents to execute tasks without additional training [2, 20, 30]. Moreover, when combined with specialized vision models, LLMs can significantly enhance the performance of vision-language tasks. For instance, Visual ChatGPT [56] uses ChatGPT as a central orchestrator, interfacing with a variety of visual foundation models to solve more challenging problems. VISPROG [15] leverages the incontext learning ability to generate high-level modular programs for solving complex and compositional natural language reasoning and image editing tasks. ViperGPT [45] directly feeds the API of available modules to LLM and then generated executable Python code for image grounding. However, leveraging these capabilities for zero-shot 3D language grounding remains an unexplored area.", + "bbox": [ + 496, + 512, + 857, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 815, + 622, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Section 3.1, we introduce the vanilla approach, i.e., dialog with LLM to overcome the annotation issue in 3DVG.", + "bbox": [ + 496, + 838, + 857, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20624", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/967e80552cedfab2ef3bf8cca03cba085e8ba269a9d5789ab6585bde0f309aa6.jpg", + "image_caption": [ + "Figure 2. Overview of two zero-shot approaches for 3DVG. (a) shows the working mechanism of the vanilla dialog with LLM approach. First, we describe the 3DVG task and provide the text descriptions of the room. Then, LLMs identify the objects relevant to the query sentence and perform human-like reasoning. (b) presents the 3D visual programming approach. We first input in-context examples into LLMs. Then, LLMs generate 3D visual programs through the grounding descriptions and perform human-like reasoning. Next, these programs are transformed into executable Python codes via the LOC module for predicting the location of the object. For example, the upper example uses the view-independent module, i.e., CLOSEST to determine the proximity in 3D space, while the lower example applies the view-dependent module, i.e., RIGHT to establish the relative positioning." + ], + "image_footnote": [], + "bbox": [ + 116, + 127, + 854, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From Section 3.2 to Section 3.4, we present the visual programming approach, address the issue of view-dependent relations, and design the LOC module, respectively.", + "bbox": [ + 109, + 506, + 473, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dialog with LLM", + "text_level": 1, + "bbox": [ + 112, + 556, + 269, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To accomplish the goal of 3DVG, we propose to initiate a dialogue with LLMs. The input for the dialogue consists of a real-world RGB-D scan and a free-form text description $\\mathcal{T}$ . The text description provides specific information about the target object within a point cloud representation $\\mathbf{P} \\in \\mathbb{R}^{N \\times 6}$ , where $\\mathbf{P}$ is a collection of color-enriched 3D points and $N$ is the total number of such points. The LLM acts as an agent located in the scanned room, which aims to identify the specified object based on the given text description. To bridge the gap between the model's proficiency in understanding text and the spatial nature of the 3DVG task, we first transform the scene into a textual narrative. This narrative can provide a comprehensive account of the objects $\\mathcal{O}$ presented in the scene, including their positions and dimensions, which can be expressed as:", + "bbox": [ + 109, + 577, + 473, + 785 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Object is a located at (x, y, z) with sizes (width, length, height).", + "guess_lang": "txt", + "bbox": [ + 142, + 791, + 424, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given such textual layout, we dialog with the LLM by providing the scene's description and query. Our objec", + "bbox": [ + 111, + 838, + 473, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tive is to guide the LLM to identify the object mentioned in the query, while also understand and explain its reasoning process in the identification duration. Particularly, LLM is capable of mimicking the reasoning steps undertaken by humans. As illustrated in Figure 2(a), if the LLM get the object information, it can extract the objects relevant to the query sentence, i.e., targets keyboard and anchors door, and successfully identify the correct target keyboard by calculating its distance with door.", + "bbox": [ + 496, + 507, + 859, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While LLMs show powerful human-like reasoning capabilities, they still have some limitations. First, it cannot handle the view-dependent issue such as the right window. This is because the 3D scene can freely rotate to different views while it keeps static in 2D images. LLMs usually make decisions by comparing their x-y values of 3D coordinates despite hinting it in the conversation. Second, mathematical calculation is a common weakness of LLMs but is necessary for 3DVG [11]. For example, in Figure 2(a), distance computing is crucial to solve the closest relations, whereas the LLMs cannot always provide accurate results. These two issues stem from LLM's training limitations, which affect the reliability of the dialog with LLM approach.", + "bbox": [ + 496, + 631, + 859, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Visual Programming", + "text_level": 1, + "bbox": [ + 500, + 816, + 705, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above two issues, we now introduce a new approach that generates visual programs through LLMs. As", + "bbox": [ + 498, + 838, + 857, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20625", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/37f762afbbb0139afc769f0cadc3cc69b07de6e3edff0b15eb53b4f13325f885.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
View-independentnear, close, next to, far, above, below, under, top, on, opposite, middle
View-dependentfront, behind, back, right, left, facing, leftmost, rightmost, looking, across, between
Functionalmin, max, size, length, width
", + "bbox": [ + 114, + 125, + 472, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Common relations in 3DVG.", + "bbox": [ + 187, + 215, + 395, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shown in Figure 2(b), we first construct a set of sample programs to encapsulate human-like problem-solving tactics in 3DVG. Each program includes a sequence of operations, where each operation contains a module name, several input parameters, and an assigned output variable. The output of each step can be reused in the subsequent step, thus creating an interlinked sequence that reflects logical reasoning within a visual context.", + "bbox": [ + 111, + 244, + 470, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We transform the reasoning process of 3DVG into a scripted visual program. Specifically, we collect a set of in-context examples and the corresponding grounding descriptions, and then use LLMs to extrapolate new visual programs tailored to the task. For example, in Figure 2(b), we consider the task prompted by the following description:", + "bbox": [ + 111, + 354, + 472, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The round cocktail table in the corner of the room with the blue and yellow poster.", + "bbox": [ + 142, + 446, + 424, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this case, the objective is to identify the round cocktail table, which can be transformed into a operation: $\\mathrm{BOX0} = \\mathrm{LOC}($ 'round cocktail table'), where the LOC operator processes the textual query and outputs the bounding boxes for the target objects. We will elaborate the design of LOC module in Section 3.4. Nevertheless, since there may exist multiple similar objects in 3D scenarios, the identified results may not be unique. To overcome this issue, we further pinpoint the blue and yellow poster as an auxiliary reference point by a operation: $\\mathrm{BOX1} = \\mathrm{LOC}($ 'blue and yellow poster'). Then, the CLOSEST module computes the proximity between BOX0 (potential tables) and BOX1 (poster), and selects the table closest to the poster as the result.", + "bbox": [ + 111, + 492, + 470, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 summarizes the common relations in 3DVG. Based on this, we present the detailed visual program by developing three types of modules tailored for 3D contexts:", + "bbox": [ + 111, + 672, + 470, + 714 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- View-independent modules: They operate on the 3D spatial relations between objects. For example, the CLOS-EST module can discern proximity independent of the viewer's position.", + "- View-dependent modules: They depend on the observer's vantage point. For instance, the RIGHT module determines the right window (TARGET) when looking at cabinets (BOX1) from all windows (BOX0).", + "- Functional modules: They include multiple operations such as MIN and MAX, which select objects based on the extremal criteria." + ], + "bbox": [ + 112, + 714, + 470, + 864 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/989c65269447548de2e43e5f6b17cba535202781d22b6c24b26ab650058ef08d.jpg", + "image_caption": [ + "Figure 3. Addressing view-dependent relations: A shift to 2D ego-centric view." + ], + "image_footnote": [], + "bbox": [ + 506, + 130, + 851, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These three types of modules allow the output of one operation to be fed into another operation, thus providing flexible composability. They not only facilitate structured and accurate inference sequences, but also integrate 3D and 2D data to yield a robust and interpretable result for 3DVG.", + "bbox": [ + 498, + 324, + 857, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Addressing View-Dependent Relations", + "text_level": 1, + "bbox": [ + 500, + 402, + 802, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we discuss the intricacies of the view-dependent relations, which are essential for interpreting spatial relations within 3D space. Particularly, the main challenge is the dynamic nature of these relations that will change with the observer's viewpoint. Although traditional supervised approaches can learn these relations implicitly, they cannot provide a definitive resolution.", + "bbox": [ + 496, + 422, + 857, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On 2D planes, the relations, especially the left and right are well defined. More specifically, right often corresponds to the positive direction of the x-axis while the left implies the negative direction. Motivated by this, we adopt a 2D egocentric view approach to ensure a consistent frame of reference for the spatial relations in Table 1.", + "bbox": [ + 496, + 520, + 857, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our view-dependent modules accept a target argument and an optional anchors parameter. They output the target objects that fulfill the spatial relation to the anchors. When grounding queries do not specify targets, we treat targets as anchors as well. This approach aligns with our intuition, such as identifying the left window by treating all windows themselves as the anchors.", + "bbox": [ + 496, + 603, + 857, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, we assume there is a virtual camera in the center of the room, i.e., $P_{\\mathrm{center}}$ , which can rotate to align with the location of the anchor objects, i.e., $P_{\\mathrm{o}_a}$ . The 3D objects are projected onto a 2D plane from this vantage point. Assume that the orthogonal camera has an intrinsic matrix $I$ , then the 2D projections can be obtained by", + "bbox": [ + 496, + 699, + 857, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR, T = \\operatorname {L o o k a t} \\left(P _ {\\text {c e n t e r}}, P _ {\\mathrm {o} _ {a}}, u p\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 792, + 857, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(u, v, w) ^ {\\mathrm {T}} = I \\cdot (R | t) \\cdot P, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 811, + 857, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Lookat}(\\cdot)$ is a view transformation function that computes the rotation matrix $R$ and translation matrix $T$", + "bbox": [ + 500, + 838, + 857, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20626", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[50], $P = (x, y, z, 1)^{\\mathrm{T}}$ denotes the 3D coordinate vector, $u$ and $v$ respectively signify the x-axis and y-axis on the 2D plane, and $w$ is the depth value. According to the value of $u$ of an object's center, we can determine its left or right position — a lower $u$ value indicates left. Similarly, $w$ allows us to distinguish the front from behind. By synthesizing these concepts, we can define the between relation.", + "bbox": [ + 111, + 125, + 473, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The transition from 3D to 2D egocentric perspective provides a clear and consistent solution to interpret view-dependent relations in 3D space, thus enhancing our model's spatial reasoning ability.", + "bbox": [ + 111, + 223, + 473, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Language-Object Correlation Module", + "text_level": 1, + "bbox": [ + 112, + 289, + 413, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although our zero-shot 3DVG approach does not need extensive grounding annotations, it still requires a basic vision model for object localization. To overcome this issue, previous works [4, 62] usually use pre-trained 3D detectors [22, 38] to generate object proposals and the corresponding labels within a fixed vocabulary. However, this approach is restricted to a predefined object class set, thus limiting the scope of class prediction. To enable open-vocabulary segmentation, we develop an LOC module, combining the advantages of 3D and 2D networks to extend the labeling capability beyond the closed set. For example, in Figure 4, considering the operation: $\\mathrm{BOX0} = \\mathrm{LOC}($ 'round cocktail table'), we first filter a subset of objects whose predicted label is table using a 3D instance segmentation network [43]. Then we only need to identify a round cocktail table from this subset using the corresponding 2D imagery. By mapping each 3D proposal to its 2D image, we can extract the color and texture details pertinent to our query. To further pinpoint the round cocktail table shown in the Figure 4, we consider three types of 2D multi-modal models:", + "bbox": [ + 111, + 311, + 473, + 588 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image classification models: We construct a dynamic vocabulary, including both the query term \"round cocktail table\" and the class \"table\" using popular tools such as CLIP [41]. Then we evaluate the cosine similarity between these terms and the imagery to find the best correlation to our query.", + "- Visual question answering models: We raise the question: Is there a [query]? to the model such as ViLT [25]. Then the model sifts through its dictionary to suggest the most likely answer, i.e., yes or no.", + "- General large models: We submit the same inquiry and anticipate a response based on the generated text. This process is crucial for verifying the alignment between the detected table and the query." + ], + "bbox": [ + 112, + 589, + 473, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We shall note that our approach is not limited to specific 3D or 2D models, allowing versatile incorporation of various models. In the experiments, we will demonstrate that the benefit of the LOC modules by comparing with the 3D-only and 2D-only counterparts. Our design indicates a leap forward in 3D open-vocabulary instance segmentation and", + "bbox": [ + 111, + 782, + 473, + 867 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e7d8fe31afa4235c96087cf3d70416266f850cfd2aa4ae9951261021ac40fdc4.jpg", + "image_caption": [ + "BOX0=LOC(object='round cocktail table')", + "Figure 4. Illustration of the language-object correlation module." + ], + "image_footnote": [], + "bbox": [ + 501, + 140, + 857, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can improve the object recognition accuracy in 3DVG.", + "bbox": [ + 500, + 327, + 830, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 352, + 620, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 375, + 689, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use two popular datasets, i.e., ScanRefer [4] and Nr3D [1] for experiments. ScanRefer is tailored for 3DVG that contains 51,500 sentence descriptions for 800 ScanNet scenes [9]. Nr3D is a human-written and free-form dataset for 3DVG, collected by 2-player reference game in 3D scenes. The sentences are divided into \"easy\" and \"hard\" subsets, where the target object only contains one same-class distractor in the \"easy\" subset but contains multiple ones in the \"hard\" subset. Depending on whether the sentence requires a specific viewpoint to ground the referred object, the dataset can also be partitioned into \"view dependent\" and \"view independent\" subsets. For both datasets, we evaluate the zero-shot approaches on the validation split.", + "bbox": [ + 496, + 396, + 859, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We consider two settings for performance evaluation. The first one mandates the generation of object proposals, aligning closely with real-world applications. The evaluation metrics are Acc@0.25 and Acc@0.5, representing the percentage of correctly predicted bounding boxes whose IoU exceeds 0.25 or 0.5 with the ground-truth, respectively. This is the default setting for ScanRefer dataset. The second one furnishes ground-truth object masks, necessitating only classification, with an objective to eradicate localization error and achieve high grounding accuracy. This is the default setting for Nr3D dataset.", + "bbox": [ + 496, + 575, + 859, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We use six supervised and two open-vocabulary 3D scene understanding approaches for performance comparison. For supervised approaches, ScanRefer [4] and ReferIt3DNet [1] encode the 3D point clouds and language separately, and then fuse them to rank the objects by predicted scores. TGNN [18] and InstanceRefer [62] make one further step by learning instance-wise features. 3DVG-Transformer [68] and BUTD-DETR [21] respectively utilize the Transformer [48] and DETR [3] architectures, representing the SoTA approaches. For open-vocabulary ap", + "bbox": [ + 496, + 728, + 859, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20627", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0dbea280a87e6f3dcb60e728d6239ab45d45cda77e4e23c0c4ed491aa2a8e531.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSupervisionUniqueMultipleOverall
Acc@0.25Acc@0.5Acc@0.25Acc@0.5Acc@0.25Acc@0.5
ScanRefer [4]fully65.043.330.619.837.324.3
TGNN [18]fully64.553.027.021.934.329.7
InstanceRefer [62]fully77.566.831.324.840.232.9
3DVG-Transformer [68]fully81.960.639.328.447.634.7
BUTD-DETR [21]fully84.266.346.635.152.239.8
LERF [24]-----4.80.9
OpenScene [35]-20.113.111.14.413.26.5
Ours (2D only)-32.527.816.114.620.017.6
Ours (3D only)-57.149.425.923.333.129.3
Ours-63.858.427.724.636.432.7
", + "bbox": [ + 125, + 123, + 841, + 296 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0e0518423041f85b502b990cb67f8bf06f80509c45ad34d47003f90365cd3ed0.jpg", + "table_caption": [ + "Table 2. 3DVG results on ScanRefer validation set. The accuracy on the \"unique\" subset, \"multiple\" subset, and whole validation set are all provided. Following [4], we label the scene as \"unique\" if it only contains a single object of its class. Otherwise, we label it as \"multiple\"." + ], + "table_footnote": [], + "table_body": "
MethodEasyHardDep.Indep.Overall
ReferIt3DNet [1]43.627.932.537.135.6
InstanceRefer [62]46.031.834.541.938.8
3DVG-Transformer [68]48.534.834.843.740.8
BUTD-DETR [21]60.748.446.058.054.6
Ours (2D only)29.418.423.023.923.6
Ours (3D only)45.927.934.938.436.7
Ours46.531.736.840.039.0
", + "bbox": [ + 117, + 340, + 467, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Performance analysis of language grounding on Nr3D. We evaluate the top-1 accuracy using ground-truth boxes.", + "bbox": [ + 112, + 446, + 470, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proaches, OpenScene [35] and LERF [24] aims to learn a 3D representation aligned with the 2D CLIP feature, thus enabling free-form language grounding. The query $\\mathcal{T}$ is processed by the CLIP text encoder, and its similarity is computed against the extracted point features. Finally, they cluster the points with the highest score to determine the target object.", + "bbox": [ + 111, + 484, + 473, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 112, + 594, + 290, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanRefer. Table 2 provides a quantitative assessment of the proposed approach on the ScanRefer dataset. We can see that our zero-shot approach outperforms all baseline approaches. Specifically, our approach can achieve a 32.7 Acc@0.5 score, which surpasses part of supervised approaches such as ScanRefer and TGNN. On the other hand, the open-vocabulary approaches LERF and OpenScene can respectively achieve the overall accuracy of 4.8 and 13.2, even with the 0.25 IoU threshold. This is due to their limitations in reasoning and localization precision. Moreover, our zero-shot approach outperforms the approaches that only utilize the 3D or 2D information in the LOC module. This result demonstrates the effectiveness of incorporating visual programming and perception modules, highlighting our zero-shot approach in navigating 3DVG.", + "bbox": [ + 111, + 616, + 470, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nr3D. Table 3 shows the performance of different approaches on the Nr3D dataset, in which the ground-truth instance mask is also provided. We can see that our zero", + "bbox": [ + 111, + 824, + 470, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shot approach further excels the supervised approach InstanceRefer. Specifically, our zero-shot approach on the \"view-dependent\" split can achieve a $2\\%$ accuracy gain than the 3DVG-Transformer approach. This performance gain comes from the relation modules, strengthening the potential of our zero-shot approach for 3DVG tasks.", + "bbox": [ + 496, + 344, + 857, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 498, + 436, + 666, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 shows the visualizations of the selected samples from the ScanRefer validation set. The four columns present the ground-truth result, the supervised approach BUTD-DETR, the dialog with LLM, and the visual programming approaches, respectively. From Figure 5(a) and Figure 5(b), we can observe that the dialog with LLM and the visual programming approaches can achieve accurate prediction results for view-independent relations, i.e., (above, under) without much training. On the contrary, both the BUTD-DETR and the dialog with LLM approaches cannot address the view-dependent relations, i.e., (left, front), as shown in Figure 5(c) and Figure 5(d). The inherent uncertainty of these relations reflects the limitations of existing methods. However, our visual programming approach can leverage the 2D egocentric views, thus achieving accurate predictions in 3D scenarios.", + "bbox": [ + 496, + 459, + 857, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5(e) presents a failure case, where the dialog with LLM approach cannot recognize chair has wheels since it lacks open-vocabulary detection ability. Besides, the visual programming approach makes wrong predictions because the LLM cannot correctly recognize the relation pushed. Fortunately, when we correct the program using the CLOsE ST module, the visual programming approach can make correct predictions.", + "bbox": [ + 496, + 681, + 857, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 801, + 648, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dialog with LLM vs. visual programming. We compare the performance of the two proposed zero-shot 3DVG approaches on the ScanRefer validation set with 700 exam", + "bbox": [ + 496, + 824, + 857, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20628", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbabafc01d0e613c1a286cf7d153edc1ad7a2af009162624d08d3c7a96d26618.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 204, + 125, + 312, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/663372fcc575527cdadb0a2616789a4b8ca5cbdae45ad1074d62a5cf02604efc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 125, + 434, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8e529c60bea8d85d165407542a0047f975da91c29aba43ace5a447ce1b28fc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 125, + 558, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/059b07e9a2dc57e0542ca06a77b4c86d8c8f8f8b9de7b6d0ef6d14432d90fcea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 125, + 681, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d21a3ffea9c8efd25c257ede92c1f2b99caf3d58f9407ac293967409beb255b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 125, + 805, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0e273c6ef725411e30020c0a5ce9ad515b75d19d5da72e53068830a66f9249e.jpg", + "image_caption": [ + "BUTD-DETR" + ], + "image_footnote": [], + "bbox": [ + 205, + 191, + 310, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71e666e749c290cacd6753be38393b19dd2cdde973281576979007d677149f72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 191, + 434, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67be939add2aa8f86ceba45bb88f0d502bf9c0ae31f9840363187c7a696751d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 191, + 558, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d2b28f6e8797a5a8c6a3f868f92fd6c9907fe89e93fd46df04b7ea59598d555.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 191, + 681, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b279d8359f8d86eb7a1eac1662b9c0aa53dd25941d07bf7a38e130c1a10807f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 191, + 803, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/234491251d07a6675289c89e06ebe712f07a0213505e76b41491407aa0936fa9.jpg", + "image_caption": [ + "Ours (Dialog)" + ], + "image_footnote": [], + "bbox": [ + 204, + 261, + 310, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f81117909218875e75e2de5e8b92bafe7139e7f8511ceba47ecf1e24d2520a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 261, + 434, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f21d468f1678d9a1380ef613f8a65035d2704c51ad6594cdcf8b7536e86c9de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 261, + 558, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0a94a61edc601d5004913c5e5be966429f345b0bb7643d4010c8b3795970d1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 261, + 681, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c7e229a58b25ae8df5f6bbb5709351a411c856b14d0a89c5e35fc7e0ebfd837.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 261, + 803, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc34a793a88f176b9b7725922d0d908e7cb94c9b2ff723b50d444e03ac728929.jpg", + "image_caption": [ + "Ours (Program)", + "It is a window. It is located above a recycle bin that has a blue top.", + "(a)", + "Figure 5. Visualization results of 3D visual grounding. Rendered images of 3D scans are presented, including the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green)." + ], + "image_footnote": [], + "bbox": [ + 204, + 330, + 310, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a38076226530917f1a5fd4642de42e54682f8823abf7bdb5673c03a7d581e82.jpg", + "image_caption": [ + "The rolling office chair. The chair is under the desk.", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 325, + 330, + 434, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d255a7bd054511bc317938843d326d757644262ef78a799c2139607988fc498.jpg", + "image_caption": [ + "There is a square beige armchair. It is left of a square table.", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 450, + 330, + 558, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a13609d7839721aeeb3f30421bc00cc4eb7cbff21e2165025c56ddf99fa60dd3.jpg", + "image_caption": [ + "This is a brown piano bench. It is in front of the piano.", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 571, + 330, + 681, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0bc77b8506f99caf5444fc935798c34e269e2efb8cb2a398a7faa214b3a3368.jpg", + "image_caption": [ + "A desk chair is pushed into a small computer desk. The chair has wheels.", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 694, + 330, + 803, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plies. For both approaches, we use two GPT versions, i.e., GPT-3.5-turbo-0613 and GPT-4-0613. The cost of each GPT version depends on the number of input and output tokens. The experimental results are shown in Table 4. We can observe that for both zero-shot approaches, GPT4-based approach can achieve higher accuracy than GPT3.5-based approach, even it induces a larger economic cost. On the other hand, the visual programming approach always outperforms the dialog with LLM approach in terms of accuracy and cost, which demonstrates the effectiveness of the proposed visual programming approach. For other experiments, we use GPT3.5 to save cost.", + "bbox": [ + 109, + 482, + 472, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relation modules. We now ablate different relation modules in Section 3.2 to analyze their impact on the system performance. The most important view-dependent and view-independent modules are presented in Table 5 and 6, respectively. We can see that LEFT and RIGHT are the most important view-dependent relations, while CLOSEST is the most important view-independent relation. This result is coherent with our motivation and design.", + "bbox": [ + 111, + 652, + 472, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LOC module. We juxtapose our approach by separately omitting the 3D component and 2D component. Both models utilize the instance mask prediction of Mask3D [43]. Particularly, the 2D-only model solely employs the paired 2D images for classification, while the 3D-only model just uses the 3D result. As can be seen from Tables 2 and 3, the 2D-only model performs worst when the images of in", + "bbox": [ + 111, + 768, + 472, + 867 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f674e0fac8a930d3bc1e1e6d4ea3fc620fcc0d682a27d920456214095f76430c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLLMAcc@0.5TokensCost
DialogGPT3.525.41959k$3.05
DialogGPT427.51916k$62.6
ProgramGPT3.532.1121k$0.19
ProgramGPT435.4115k$4.24
", + "bbox": [ + 527, + 479, + 831, + 554 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35e4a3352510d4fb8e4af113a04158c1de96d2429c3fe12ed25a222eee83b2cf.jpg", + "table_caption": [ + "Table 4. Performance comparison of the dialog with LLM and the visual programming approaches." + ], + "table_footnote": [], + "table_body": "
LEFTRIGHTFRONTBEHINDBETWEENAccuracy
26.5
✓32.4
✓✓35.9
✓✓✓36.8
✓✓✓✓38.4
✓✓✓✓✓39.0
", + "bbox": [ + 513, + 599, + 841, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study of different view-dependent modules.", + "bbox": [ + 506, + 700, + 846, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "door scenes are complicated and have domain gaps with the training samples. The 3D-only model performs better since it can utilize the geometric information and is trained on closed-set labels. Our full model can always achieve the best performance because it integrates the geometric distinctiveness of point clouds and the open-vocabulary ability of the image models.", + "bbox": [ + 496, + 724, + 857, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization. As discussed in Section 3.4, our framework has strong adaptability for a spectrum of 3D and 2D perception models. To validate this claim, we conduct ex", + "bbox": [ + 496, + 824, + 857, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20629", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0d15a31d4f1aa3e0eb134efdd93698857caac1625d00fd5137b99784cf767c6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLOSESTFARTHESTLOWERHIGHERAccuracy
18.8
✓30.7
✓✓34.0
✓✓✓36.8
✓✓✓✓39.0
", + "bbox": [ + 122, + 122, + 459, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "periments using several representative models. For 3D perception, we utilize three backbones, i.e., PointNet++ [37], PointNeXt [39], and PointBERT [61]. For 2D perception, we use an image classification model proposed in [41], a visual question answering model in [25], and a general large model BLIP-2 [28] for testing. The results are shown in Tables 7 and 8. We can observe that our framework is compatible with other models. Also, it can leverage the advancements within both 2D and 3D foundational models to improve the performance. This cross-model effectiveness demonstrates the robustness and future-proof nature of our approach in the ever-evolving landscape of visual perception models.", + "bbox": [ + 111, + 234, + 472, + 412 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ae3a2a8b4633821fb35e43a33a0c49ed40dc0eed662f0a9d25835bf87d10e46b.jpg", + "table_caption": [ + "Table 6. Ablation study of different view-independent modules." + ], + "table_footnote": [], + "table_body": "
2D AssistanceUniqueMultipleAcc@0.25
CLIP62.527.135.7
ViLT60.327.135.1
BLIP-263.827.736.4
", + "bbox": [ + 142, + 426, + 439, + 483 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/31a73e8d6de3fd3a84cc9a5d49faca58d7d0b73f23b6b42a678f9ea74817f2bf.jpg", + "table_caption": [ + "Table 7. Ablation study on different 2D models." + ], + "table_footnote": [], + "table_body": "
3D BackboneView-dep.View-indep.Overall
PointNet++35.839.438.2
PointBert36.039.838.6
PointNeXt36.840.039.0
", + "bbox": [ + 135, + 515, + 444, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study on different 3D backbones.", + "bbox": [ + 151, + 575, + 431, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of prompt size. We use different numbers of incontext examples in the prompt for program generation. The result is shown in Figure 6. It can be seen that the performance on ScanRefer and Nr3d improve with the number of examples. This is because more examples can guide LLMs to handle more cases in the visual program generation process. Meanwhile, it still follows the law of diminishing marginal utility. Moreover, we test the voting technique [15] to aggregate the results from multiple runs, which brings some performance gains.", + "bbox": [ + 111, + 601, + 470, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Error analysis. We select a representative subset for each dataset with around 100 samples and manually check the rationales offered by the visual program. As illustrated in Figure 7, the generation of accurate visual programs is the primary error source. Therefore, the performance can be improved by using more in-context examples and more powerful LLMs. The second error source is the object localization and classification, indicating that object detection and classification in 3D space remains a critical component.", + "bbox": [ + 111, + 741, + 472, + 866 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81a1d1b34240c8702a3a2779df88df2b625304a1a4a601ae1fcfb421705fa0db.jpg", + "image_caption": [ + "Figure 6. Ablation study on the number of in-context examples. The performance on Nr3D and ScanRefer improves with the number of in-context examples." + ], + "image_footnote": [], + "bbox": [ + 517, + 125, + 681, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72f79338576ca87efd0ee73cbc9f8c0035c7951e1923b36375360755cfa68e3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 125, + 836, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78d7351c3a5f5cfd39dd44ab859742cfc02ef8e61dbf123accc98330310c64f4.jpg", + "image_caption": [ + "Figure 7. Breakdown of error sources." + ], + "image_footnote": [], + "bbox": [ + 521, + 273, + 834, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 415, + 607, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a novel zero-shot approach for 3DVG to eliminate the need for extensive annotations and predefined vocabularies. A vanilla dialog with LLM approach is first proposed by taking interactive dialog with LLMs. A visual programming approach is further developed, which leverages three types of modules to navigate the intricate 3D relations. To adapt to open-vocabulary scenarios, we also develop a LOC module to seamlessly integrate both 3D and 2D features. Experimental results demonstrate the superiority of the proposed approach and highlight its potential to advance the field of 3DVG.", + "bbox": [ + 496, + 438, + 859, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 602, + 663, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the National Key R&D Program of China with grant No. 2018YFB1800800, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and the Key Area R&D Program of Guangdong Province with grant No. 2018B03033800, by Tencent & Huawei Open Fund.", + "bbox": [ + 496, + 625, + 859, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20630", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 125, + 200, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 1, 2, 5, 6", + "[2] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In Conference on Robot Learning, pages 287-318. PMLR, 2023. 2", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5", + "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 2, 5, 6", + "[5] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 2", + "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 2", + "[7] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2", + "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2, 5", + "[10] John David N Dionisio, William G Burns Iii, and Richard Gilbert. 3d virtual worlds and the metaverse: Current status and future possibilities. ACM Computing Surveys (CSUR), 45(3):1-38, 2013. 1, 2", + "[11] Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D Hwang, et al. Faith and fate: Limits of transformers on compositionality. arXiv preprint arXiv:2305.18654, 2023. 3", + "[12] Qi Feng, Vitaly Ablavsky, and Stan Sclaroff. Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint arXiv:2101.04741, 2021. 1, 2", + "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level" + ], + "bbox": [ + 114, + 147, + 472, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2", + "[14] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 2", + "[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 8", + "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 2", + "[17] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2614-2623, 2023. 2", + "[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. 35th AAAI Conference on Artificial Intelligence, 2021. 5, 6", + "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1", + "[20] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2", + "[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 417-433. Springer, 2022. 5, 6", + "[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4867-4876, 2020. 1, 2, 5", + "[23] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2", + "[24] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 6", + "[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region su" + ], + "bbox": [ + 501, + 127, + 856, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20631", + "bbox": [ + 480, + 891, + 514, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5, 8", + "[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2", + "[27] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8", + "[29] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2", + "[30] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2", + "[31] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[33] Stylianos Mystakidis. Metaverse. Encyclopedia, 2(1):486-497, 2022. 1, 2", + "[34] OpenAI OpenAI. Gpt-4 technical report. 2023. 2", + "[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2, 6", + "[36] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[37] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 8", + "[38] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5", + "[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard" + ], + "bbox": [ + 116, + 127, + 470, + 864 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 8", + "[40] Xing-yue Qiu, Chuang-Kai Chiu, Lu-Lu Zhao, Cai-Feng Sun, and Shu-jie Chen. Trends in vr/ar technology-supporting language learning from 2008 to 2019: A research perspective. Interactive Learning Environments, 31 (4):2090-2113, 2023. 1", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5, 8", + "[42] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In 5th Annual Conference on Robot Learning, 2021. 2", + "[43] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 5, 7", + "[44] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 2", + "[45] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2", + "[46] Ayça Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2", + "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5", + "[49] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2:20, 2023. 2", + "[50] John Vince and John A Vince. Mathematics for computer graphics. Springer, 2006. 5", + "[51] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 2", + "[52] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance relocalization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7658-7667, 2019. 2" + ], + "bbox": [ + 501, + 127, + 856, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20632", + "bbox": [ + 480, + 891, + 517, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6629-6638, 2019. 1, 2", + "[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2", + "[55] Wei Wei. Research progress on virtual reality (vr) and augmented reality (ar) in tourism and hospitality: A critical review of publications from 2000 to 2018. Journal of Hospitality and Tourism Technology, 10(4):539-570, 2019. 1", + "[56] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2", + "[57] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2", + "[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1, 2", + "[59] Xu Yan, Zhihao Yuan, Yuhao Du, Yinghong Liao, Yao Guo, Shuguang Cui, and Zhen Li. Comprehensive visual question answering on point clouds through compositional scene manipulation. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[60] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2", + "[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8", + "[62] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 1, 2, 5, 6", + "[63] Zhihao Yuan, Xu Yan, Zhuo Li, Xuhao Li, Yao Guo, Shuguang Cui, and Zhen Li. Toward explainable and fine-grained 3d grounding through referring textual phrases. arXiv preprint arXiv:2207.01821, 2022. 2", + "[64] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Zhen Li, and Shuguang Cui. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense caption" + ], + "bbox": [ + 114, + 126, + 472, + 866 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingquiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2023. 2", + "[66] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4534-4543, 2020. 2", + "[67] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2", + "[68] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 5, 6", + "[69] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 2", + "[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 2" + ], + "bbox": [ + 501, + 126, + 856, + 551 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20633", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_model.json b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..131b39384cf224afc9f17881f45bb93c0b569e8a --- /dev/null +++ b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_model.json @@ -0,0 +1,2785 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.26, + 0.044, + 0.782, + 0.058 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.339, + 0.058, + 0.705, + 0.071 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.31, + 0.072, + 0.733, + 0.085 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.162, + 0.834, + 0.183 + ], + "angle": 0, + "content": "Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.207, + 0.841, + 0.226 + ], + "angle": 0, + "content": "Zhihao Yuan\\(^{1,2}\\), Jinke Ren\\(^{1,2}\\), Chun-Mei Feng\\(^{3}\\), Hengshuang Zhao\\(^{4}\\), Shuguang Cui\\(^{2,1}\\), Zhen Li\\(^{2,1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.24, + 0.768, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\) FNii, CUHKSZ \\(^{2}\\) SSE, CUHKSZ \\(^{3}\\) IHPC, A*STAR, Singapore \\(^{4}\\) HKU" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.289, + 0.33, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.318, + 0.474, + 0.61 + ], + "angle": 0, + "content": "3D Visual Grounding (3DVG) aims at localizing 3D object based on textual descriptions. Conventional supervised methods for 3DVG often necessitate extensive annotations and a predefined vocabulary, which can be restrictive. To address this issue, we propose a novel visual programming approach for zero-shot open-vocabulary 3DVG, leveraging the capabilities of large language models (LLMs). Our approach begins with a unique dialog-based method, engaging with LLMs to establish a foundational understanding of zero-shot 3DVG. Building on this, we design a visual program that consists of three types of modules, i.e., view-independent, view-dependent, and functional modules. These modules, specifically tailored for 3D scenarios, work collaboratively to perform complex reasoning and inference. Furthermore, we develop an innovative language-object correlation module to extend the scope of existing 3D object detectors into open-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot approach can outperform some supervised baselines, marking a significant stride towards effective 3DVG. Code is available at https://curryyuan.github.io/ZSVG3D." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.638, + 0.235, + 0.652 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.662, + 0.472, + 0.842 + ], + "angle": 0, + "content": "3D Visual Grounding (3DVG) aims to localize specific objects within 3D scenes by using a series of textual descriptions. This has become a crucial component in a variety of burgeoning applications, such as autonomous robotics [12, 53, 58], virtual reality [40, 55], and metaverse [10, 33]. For illustration, given a 3D scan in Figure 1(a) along with its description — It is the keyboard closest to the door, the goal of 3DVG is to accurately pinpoint the keyboard in the green box, while eliminating potential distractions such as tables and desks. Despite the apparent simplicity of this task for humans, it poses a significant challenge for machines due to their inherently limited perceptual capabilities." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.288, + 0.859, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.379, + 0.758, + 0.39 + ], + "angle": 0, + "content": "(a) Supervised 3D Visual Grounding" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.394, + 0.856, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.502, + 0.761, + 0.513 + ], + "angle": 0, + "content": "(b) Zero-shot 3D Visual Grounding" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.52, + 0.86, + 0.647 + ], + "angle": 0, + "content": "Figure 1. Comparative overview of two 3DVG approaches, where (a) Supervised 3DVG involves input from 3D scans combined with text queries, guided by object-text pair annotations, (b) Zero-shot 3DVG identifies the location of target objects using programmatic representation generated by LLMs, i.e., target category, anchor category, and relation grounding, thereby highlighting its superiority in decoding spatial relations and object identifiers within a given space, e.g., the location of the keyboard (outlined in green) can be retrieved based on the distance between the keyboard and the door (outlined in blue)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.859, + 0.824 + ], + "angle": 0, + "content": "Traditional supervised 3DVG approaches [19, 60, 62] achieve this objective by leveraging the rich annotations in public datasets, such as ScanRefer [4] and Referit3D [1]. These approaches typically define 3DVG as a matching problem, generating possible objects via 3D detectors [22, 38], and identifying the best match by fusing the visual and textual features. While these approaches can yield precise results, the acquisition of sufficient annotations is prohibitively resource-intensive for real-world applications. Furthermore, these approaches are often constrained by the pre-defined vocabulary during training, making them suboptimal in open-vocabulary scenarios." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.859, + 0.868 + ], + "angle": 0, + "content": "To address these issues, we propose a novel visual programming approach for 3DVG that integrates zero-shot learning and large language models (LLMs). Zero-shot" + }, + { + "type": "page_footnote", + "bbox": [ + 0.131, + 0.854, + 0.246, + 0.867 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20623" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.126, + 0.473, + 0.416 + ], + "angle": 0, + "content": "learning [65, 67, 70] can generalize across new categories by leveraging the pre-trained capabilities of CLIP [41] in the 3D domain. LLMs [34, 47, 49] can facilitate 3DVG due to their strong planning and reasoning capabilities. Regarding this, we first propose a vanilla version dialog with LLM. It describes the location and size of all objects in the scene and instructs the LLM to distinguish the object of interest through an interactive dialog. Despite the simplicity of the base approach, the inherent stochasticity and control limitations of LLMs make it hard to capture the view-dependent queries and decipher spatial relations in 3D space, which are the main challenges of 3DVG. To overcome this limitation, we further develop a new visual programming approach, as shown in Figure 1(b). It mainly consists of three steps: (1) generating a 3D visual program using LLMs, (2) interpreting the program into Python code, and (3) identifying the target bounding box by executing the code. To enhance the localization accuracy, we further introduce a novel language-object correlation (LOC) module capable of merging the geometric discernment of 3D point clouds with the fine-grained appearance acumen of 2D images." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.417, + 0.46, + 0.43 + ], + "angle": 0, + "content": "In summary, contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.431, + 0.472, + 0.471 + ], + "angle": 0, + "content": "- We propose an innovative 3D visual programming approach. It eliminates the need for extensive object-text pair annotations required in supervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.473, + 0.472, + 0.554 + ], + "angle": 0, + "content": "- We transform the visual program to Python code by designing two types of modules, i.e., relation modules and LOC modules. The former explicitly defines the view-dependent and view-independent relations in 3D space, while the latter captures both the geometric and appearance information for open-vocabulary localization." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.555, + 0.472, + 0.638 + ], + "angle": 0, + "content": "- We conduct extensive experiments on two popular datasets, i.e., ScanRefer [4] and Nr3d [1]. We for the first time evaluate the whole validation set rather than a few samples. The results demonstrate the superior performance of our approach, even comparable with existing supervised approaches." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.431, + 0.472, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.65, + 0.241, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.673, + 0.473, + 0.867 + ], + "angle": 0, + "content": "Supervised 3DVG. 3DVG has received much attention in many emerging applications ranging from automatic robotics [12, 53, 58] to metaverse [10, 33]. On the one hand, densely-annotated datasets like ScanRefer [4] and Referit3D [1] can provide well-aligned object-text pairs for ScanNet [9]. On the other hand, most existing methods [6, 14, 42, 60, 62] treat 3DVG as a matching problem, where object identifiers [22, 38] are utilized to generate candidate objects and find the best-matching one by fusing visual and textual features. Building on this, [57, 63] attempt to explore the object attributes and relations between different proposals. Moreover, some works [7, 23] have also investigated 3D language pretraining using advanced techniques, such as mask modeling and contrastive learning on paired" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.858, + 0.195 + ], + "angle": 0, + "content": "object-caption data, followed by finetuning on downstream tasks. Additionally, NS3D [17] has employed CodeX [5] to generate hierarchical programs. However, it still needs annotations to train the neuro-symbolic networks, thus lacking open-vocabulary and zero-shot capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.196, + 0.859, + 0.513 + ], + "angle": 0, + "content": "Indoor 3D Scene Understanding. 3D scene understanding of indoor environments has been widely studied. In specific, the emergence of RBG-D scans datasets [9, 44, 52] greatly push the boundary of several tasks, including 3D object classification [36, 37], 3D object detection [31, 38], 3D semantic segmentation [8, 66], 3D instance segmentation [22, 43, 51], and so on [16, 59, 64]. However, these methods are often constrained to a closed set of semantic class labels, limiting their applicability in real-world scenarios. Recent progress in open-vocabulary image segmentation [13, 29] has inspired research into 3D scene understanding under the open-vocabulary setting. For instance, LERF [24] learns a language field inside NeRF [32] by volume rendering CLIP [41] features along training rays, enabling it to generate 3D relevancy maps for arbitrary language queries. OpenScene [35] extracts image features using 2D open-vocabulary segmentation models [13, 27], then trained a 3D network to produce point features aligned with multi-view fused pixel features. OpenMask3D [46] utilizes the closed-vocabulary network to generate instance masks while discarded the classification head. Despite these advancements, these methods still lack spatial and commonsense reasoning abilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.859, + 0.804 + ], + "angle": 0, + "content": "LLMs for Vision-Language Tasks. Recent progress on LLMs has provided impressive zero-shot planning and reasoning abilities [34, 47, 49]. Advanced prompting technologies such as Least-to-Most [69], Think-Step-by-Step [26], and Chain-of-Thought [54] are proposed to elicit the capabilities of LLMs. These methods can understand human instructions, break complex goals into sub-goals, and control robot agents to execute tasks without additional training [2, 20, 30]. Moreover, when combined with specialized vision models, LLMs can significantly enhance the performance of vision-language tasks. For instance, Visual ChatGPT [56] uses ChatGPT as a central orchestrator, interfacing with a variety of visual foundation models to solve more challenging problems. VISPROG [15] leverages the incontext learning ability to generate high-level modular programs for solving complex and compositional natural language reasoning and image editing tasks. ViperGPT [45] directly feeds the API of available modules to LLM and then generated executable Python code for image grounding. However, leveraging these capabilities for zero-shot 3D language grounding remains an unexplored area." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.623, + 0.832 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.858, + 0.867 + ], + "angle": 0, + "content": "In Section 3.1, we introduce the vanilla approach, i.e., dialog with LLM to overcome the annotation issue in 3DVG." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.128, + 0.856, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.394, + 0.862, + 0.486 + ], + "angle": 0, + "content": "Figure 2. Overview of two zero-shot approaches for 3DVG. (a) shows the working mechanism of the vanilla dialog with LLM approach. First, we describe the 3DVG task and provide the text descriptions of the room. Then, LLMs identify the objects relevant to the query sentence and perform human-like reasoning. (b) presents the 3D visual programming approach. We first input in-context examples into LLMs. Then, LLMs generate 3D visual programs through the grounding descriptions and perform human-like reasoning. Next, these programs are transformed into executable Python codes via the LOC module for predicting the location of the object. For example, the upper example uses the view-independent module, i.e., CLOSEST to determine the proximity in 3D space, while the lower example applies the view-dependent module, i.e., RIGHT to establish the relative positioning." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.507, + 0.475, + 0.551 + ], + "angle": 0, + "content": "From Section 3.2 to Section 3.4, we present the visual programming approach, address the issue of view-dependent relations, and design the LOC module, respectively." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.557, + 0.271, + 0.571 + ], + "angle": 0, + "content": "3.1. Dialog with LLM" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.578, + 0.475, + 0.786 + ], + "angle": 0, + "content": "To accomplish the goal of 3DVG, we propose to initiate a dialogue with LLMs. The input for the dialogue consists of a real-world RGB-D scan and a free-form text description \\(\\mathcal{T}\\). The text description provides specific information about the target object within a point cloud representation \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 6}\\), where \\(\\mathbf{P}\\) is a collection of color-enriched 3D points and \\(N\\) is the total number of such points. The LLM acts as an agent located in the scanned room, which aims to identify the specified object based on the given text description. To bridge the gap between the model's proficiency in understanding text and the spatial nature of the 3DVG task, we first transform the scene into a textual narrative. This narrative can provide a comprehensive account of the objects \\(\\mathcal{O}\\) presented in the scene, including their positions and dimensions, which can be expressed as:" + }, + { + "type": "code", + "bbox": [ + 0.143, + 0.792, + 0.426, + 0.834 + ], + "angle": 0, + "content": "Object is a located at (x, y, z) with sizes (width, length, height)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.839, + 0.474, + 0.869 + ], + "angle": 0, + "content": "Given such textual layout, we dialog with the LLM by providing the scene's description and query. Our objec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.86, + 0.631 + ], + "angle": 0, + "content": "tive is to guide the LLM to identify the object mentioned in the query, while also understand and explain its reasoning process in the identification duration. Particularly, LLM is capable of mimicking the reasoning steps undertaken by humans. As illustrated in Figure 2(a), if the LLM get the object information, it can extract the objects relevant to the query sentence, i.e., targets keyboard and anchors door, and successfully identify the correct target keyboard by calculating its distance with door." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.861, + 0.812 + ], + "angle": 0, + "content": "While LLMs show powerful human-like reasoning capabilities, they still have some limitations. First, it cannot handle the view-dependent issue such as the right window. This is because the 3D scene can freely rotate to different views while it keeps static in 2D images. LLMs usually make decisions by comparing their x-y values of 3D coordinates despite hinting it in the conversation. Second, mathematical calculation is a common weakness of LLMs but is necessary for 3DVG [11]. For example, in Figure 2(a), distance computing is crucial to solve the closest relations, whereas the LLMs cannot always provide accurate results. These two issues stem from LLM's training limitations, which affect the reliability of the dialog with LLM approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.707, + 0.833 + ], + "angle": 0, + "content": "3.2. 3D Visual Programming" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.859, + 0.868 + ], + "angle": 0, + "content": "To address the above two issues, we now introduce a new approach that generates visual programs through LLMs. As" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20625" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.126, + 0.473, + 0.212 + ], + "angle": 0, + "content": "
View-independentnear, close, next to, far, above, below, under, top, on, opposite, middle
View-dependentfront, behind, back, right, left, facing, leftmost, rightmost, looking, across, between
Functionalmin, max, size, length, width
" + }, + { + "type": "table_caption", + "bbox": [ + 0.189, + 0.217, + 0.396, + 0.228 + ], + "angle": 0, + "content": "Table 1. Common relations in 3DVG." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.245, + 0.471, + 0.355 + ], + "angle": 0, + "content": "shown in Figure 2(b), we first construct a set of sample programs to encapsulate human-like problem-solving tactics in 3DVG. Each program includes a sequence of operations, where each operation contains a module name, several input parameters, and an assigned output variable. The output of each step can be reused in the subsequent step, thus creating an interlinked sequence that reflects logical reasoning within a visual context." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.356, + 0.473, + 0.44 + ], + "angle": 0, + "content": "We transform the reasoning process of 3DVG into a scripted visual program. Specifically, we collect a set of in-context examples and the corresponding grounding descriptions, and then use LLMs to extrapolate new visual programs tailored to the task. For example, in Figure 2(b), we consider the task prompted by the following description:" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.447, + 0.425, + 0.487 + ], + "angle": 0, + "content": "The round cocktail table in the corner of the room with the blue and yellow poster." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.493, + 0.472, + 0.672 + ], + "angle": 0, + "content": "In this case, the objective is to identify the round cocktail table, which can be transformed into a operation: \\(\\mathrm{BOX0} = \\mathrm{LOC}(\\) 'round cocktail table'), where the LOC operator processes the textual query and outputs the bounding boxes for the target objects. We will elaborate the design of LOC module in Section 3.4. Nevertheless, since there may exist multiple similar objects in 3D scenarios, the identified results may not be unique. To overcome this issue, we further pinpoint the blue and yellow poster as an auxiliary reference point by a operation: \\(\\mathrm{BOX1} = \\mathrm{LOC}(\\) 'blue and yellow poster'). Then, the CLOSEST module computes the proximity between BOX0 (potential tables) and BOX1 (poster), and selects the table closest to the poster as the result." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.673, + 0.472, + 0.715 + ], + "angle": 0, + "content": "Table 1 summarizes the common relations in 3DVG. Based on this, we present the detailed visual program by developing three types of modules tailored for 3D contexts:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.715, + 0.472, + 0.769 + ], + "angle": 0, + "content": "- View-independent modules: They operate on the 3D spatial relations between objects. For example, the CLOS-EST module can discern proximity independent of the viewer's position." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.77, + 0.472, + 0.825 + ], + "angle": 0, + "content": "- View-dependent modules: They depend on the observer's vantage point. For instance, the RIGHT module determines the right window (TARGET) when looking at cabinets (BOX1) from all windows (BOX0)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.826, + 0.472, + 0.866 + ], + "angle": 0, + "content": "- Functional modules: They include multiple operations such as MIN and MAX, which select objects based on the extremal criteria." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.715, + 0.472, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.131, + 0.852, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.275, + 0.858, + 0.3 + ], + "angle": 0, + "content": "Figure 3. Addressing view-dependent relations: A shift to 2D ego-centric view." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.325, + 0.859, + 0.395 + ], + "angle": 0, + "content": "These three types of modules allow the output of one operation to be fed into another operation, thus providing flexible composability. They not only facilitate structured and accurate inference sequences, but also integrate 3D and 2D data to yield a robust and interpretable result for 3DVG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.803, + 0.417 + ], + "angle": 0, + "content": "3.3. Addressing View-Dependent Relations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.859, + 0.52 + ], + "angle": 0, + "content": "In this section, we discuss the intricacies of the view-dependent relations, which are essential for interpreting spatial relations within 3D space. Particularly, the main challenge is the dynamic nature of these relations that will change with the observer's viewpoint. Although traditional supervised approaches can learn these relations implicitly, they cannot provide a definitive resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.859, + 0.603 + ], + "angle": 0, + "content": "On 2D planes, the relations, especially the left and right are well defined. More specifically, right often corresponds to the positive direction of the x-axis while the left implies the negative direction. Motivated by this, we adopt a 2D egocentric view approach to ensure a consistent frame of reference for the spatial relations in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.859, + 0.699 + ], + "angle": 0, + "content": "Our view-dependent modules accept a target argument and an optional anchors parameter. They output the target objects that fulfill the spatial relation to the anchors. When grounding queries do not specify targets, we treat targets as anchors as well. This approach aligns with our intuition, such as identifying the left window by treating all windows themselves as the anchors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.859, + 0.784 + ], + "angle": 0, + "content": "As shown in Figure 3, we assume there is a virtual camera in the center of the room, i.e., \\( P_{\\mathrm{center}} \\), which can rotate to align with the location of the anchor objects, i.e., \\( P_{\\mathrm{o}_a} \\). The 3D objects are projected onto a 2D plane from this vantage point. Assume that the orthogonal camera has an intrinsic matrix \\( I \\), then the 2D projections can be obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.794, + 0.858, + 0.809 + ], + "angle": 0, + "content": "\\[\nR, T = \\operatorname {L o o k a t} \\left(P _ {\\text {c e n t e r}}, P _ {\\mathrm {o} _ {a}}, u p\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.812, + 0.858, + 0.828 + ], + "angle": 0, + "content": "\\[\n(u, v, w) ^ {\\mathrm {T}} = I \\cdot (R | t) \\cdot P, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.859, + 0.867 + ], + "angle": 0, + "content": "where \\(\\mathrm{Lookat}(\\cdot)\\) is a view transformation function that computes the rotation matrix \\(R\\) and translation matrix \\(T\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.126, + 0.474, + 0.223 + ], + "angle": 0, + "content": "[50], \\( P = (x, y, z, 1)^{\\mathrm{T}} \\) denotes the 3D coordinate vector, \\( u \\) and \\( v \\) respectively signify the x-axis and y-axis on the 2D plane, and \\( w \\) is the depth value. According to the value of \\( u \\) of an object's center, we can determine its left or right position — a lower \\( u \\) value indicates left. Similarly, \\( w \\) allows us to distinguish the front from behind. By synthesizing these concepts, we can define the between relation." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.224, + 0.474, + 0.28 + ], + "angle": 0, + "content": "The transition from 3D to 2D egocentric perspective provides a clear and consistent solution to interpret view-dependent relations in 3D space, thus enhancing our model's spatial reasoning ability." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.29, + 0.414, + 0.306 + ], + "angle": 0, + "content": "3.4. Language-Object Correlation Module" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.313, + 0.474, + 0.589 + ], + "angle": 0, + "content": "Although our zero-shot 3DVG approach does not need extensive grounding annotations, it still requires a basic vision model for object localization. To overcome this issue, previous works [4, 62] usually use pre-trained 3D detectors [22, 38] to generate object proposals and the corresponding labels within a fixed vocabulary. However, this approach is restricted to a predefined object class set, thus limiting the scope of class prediction. To enable open-vocabulary segmentation, we develop an LOC module, combining the advantages of 3D and 2D networks to extend the labeling capability beyond the closed set. For example, in Figure 4, considering the operation: \\(\\mathrm{BOX0} = \\mathrm{LOC}(\\) 'round cocktail table'), we first filter a subset of objects whose predicted label is table using a 3D instance segmentation network [43]. Then we only need to identify a round cocktail table from this subset using the corresponding 2D imagery. By mapping each 3D proposal to its 2D image, we can extract the color and texture details pertinent to our query. To further pinpoint the round cocktail table shown in the Figure 4, we consider three types of 2D multi-modal models:" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.59, + 0.474, + 0.673 + ], + "angle": 0, + "content": "- Image classification models: We construct a dynamic vocabulary, including both the query term \"round cocktail table\" and the class \"table\" using popular tools such as CLIP [41]. Then we evaluate the cosine similarity between these terms and the imagery to find the best correlation to our query." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.672, + 0.474, + 0.728 + ], + "angle": 0, + "content": "- Visual question answering models: We raise the question: Is there a [query]? to the model such as ViLT [25]. Then the model sifts through its dictionary to suggest the most likely answer, i.e., yes or no." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.728, + 0.473, + 0.784 + ], + "angle": 0, + "content": "- General large models: We submit the same inquiry and anticipate a response based on the generated text. This process is crucial for verifying the alignment between the detected table and the query." + }, + { + "type": "list", + "bbox": [ + 0.113, + 0.59, + 0.474, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.784, + 0.474, + 0.868 + ], + "angle": 0, + "content": "We shall note that our approach is not limited to specific 3D or 2D models, allowing versatile incorporation of various models. In the experiments, we will demonstrate that the benefit of the LOC modules by comparing with the 3D-only and 2D-only counterparts. Our design indicates a leap forward in 3D open-vocabulary instance segmentation and" + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.127, + 0.778, + 0.138 + ], + "angle": 0, + "content": "BOX0=LOC(object='round cocktail table')" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.141, + 0.858, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.306, + 0.854, + 0.319 + ], + "angle": 0, + "content": "Figure 4. Illustration of the language-object correlation module." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.831, + 0.342 + ], + "angle": 0, + "content": "can improve the object recognition accuracy in 3DVG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.353, + 0.622, + 0.369 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.691, + 0.391 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.86, + 0.577 + ], + "angle": 0, + "content": "Datasets. We use two popular datasets, i.e., ScanRefer [4] and Nr3D [1] for experiments. ScanRefer is tailored for 3DVG that contains 51,500 sentence descriptions for 800 ScanNet scenes [9]. Nr3D is a human-written and free-form dataset for 3DVG, collected by 2-player reference game in 3D scenes. The sentences are divided into \"easy\" and \"hard\" subsets, where the target object only contains one same-class distractor in the \"easy\" subset but contains multiple ones in the \"hard\" subset. Depending on whether the sentence requires a specific viewpoint to ground the referred object, the dataset can also be partitioned into \"view dependent\" and \"view independent\" subsets. For both datasets, we evaluate the zero-shot approaches on the validation split." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.86, + 0.729 + ], + "angle": 0, + "content": "Evaluation metrics. We consider two settings for performance evaluation. The first one mandates the generation of object proposals, aligning closely with real-world applications. The evaluation metrics are Acc@0.25 and Acc@0.5, representing the percentage of correctly predicted bounding boxes whose IoU exceeds 0.25 or 0.5 with the ground-truth, respectively. This is the default setting for ScanRefer dataset. The second one furnishes ground-truth object masks, necessitating only classification, with an objective to eradicate localization error and achieve high grounding accuracy. This is the default setting for Nr3D dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.861, + 0.868 + ], + "angle": 0, + "content": "Baselines. We use six supervised and two open-vocabulary 3D scene understanding approaches for performance comparison. For supervised approaches, ScanRefer [4] and ReferIt3DNet [1] encode the 3D point clouds and language separately, and then fuse them to rank the objects by predicted scores. TGNN [18] and InstanceRefer [62] make one further step by learning instance-wise features. 3DVG-Transformer [68] and BUTD-DETR [21] respectively utilize the Transformer [48] and DETR [3] architectures, representing the SoTA approaches. For open-vocabulary ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20627" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.124, + 0.842, + 0.297 + ], + "angle": 0, + "content": "
MethodsSupervisionUniqueMultipleOverall
Acc@0.25Acc@0.5Acc@0.25Acc@0.5Acc@0.25Acc@0.5
ScanRefer [4]fully65.043.330.619.837.324.3
TGNN [18]fully64.553.027.021.934.329.7
InstanceRefer [62]fully77.566.831.324.840.232.9
3DVG-Transformer [68]fully81.960.639.328.447.634.7
BUTD-DETR [21]fully84.266.346.635.152.239.8
LERF [24]-----4.80.9
OpenScene [35]-20.113.111.14.413.26.5
Ours (2D only)-32.527.816.114.620.017.6
Ours (3D only)-57.149.425.923.333.129.3
Ours-63.858.427.724.636.432.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.303, + 0.859, + 0.331 + ], + "angle": 0, + "content": "Table 2. 3DVG results on ScanRefer validation set. The accuracy on the \"unique\" subset, \"multiple\" subset, and whole validation set are all provided. Following [4], we label the scene as \"unique\" if it only contains a single object of its class. Otherwise, we label it as \"multiple\"." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.342, + 0.468, + 0.441 + ], + "angle": 0, + "content": "
MethodEasyHardDep.Indep.Overall
ReferIt3DNet [1]43.627.932.537.135.6
InstanceRefer [62]46.031.834.541.938.8
3DVG-Transformer [68]48.534.834.843.740.8
BUTD-DETR [21]60.748.446.058.054.6
Ours (2D only)29.418.423.023.923.6
Ours (3D only)45.927.934.938.436.7
Ours46.531.736.840.039.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.448, + 0.472, + 0.474 + ], + "angle": 0, + "content": "Table 3. Performance analysis of language grounding on Nr3D. We evaluate the top-1 accuracy using ground-truth boxes." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.486, + 0.474, + 0.584 + ], + "angle": 0, + "content": "proaches, OpenScene [35] and LERF [24] aims to learn a 3D representation aligned with the 2D CLIP feature, thus enabling free-form language grounding. The query \\(\\mathcal{T}\\) is processed by the CLIP text encoder, and its similarity is computed against the extracted point features. Finally, they cluster the points with the highest score to determine the target object." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.595, + 0.291, + 0.609 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.617, + 0.472, + 0.824 + ], + "angle": 0, + "content": "ScanRefer. Table 2 provides a quantitative assessment of the proposed approach on the ScanRefer dataset. We can see that our zero-shot approach outperforms all baseline approaches. Specifically, our approach can achieve a 32.7 Acc@0.5 score, which surpasses part of supervised approaches such as ScanRefer and TGNN. On the other hand, the open-vocabulary approaches LERF and OpenScene can respectively achieve the overall accuracy of 4.8 and 13.2, even with the 0.25 IoU threshold. This is due to their limitations in reasoning and localization precision. Moreover, our zero-shot approach outperforms the approaches that only utilize the 3D or 2D information in the LOC module. This result demonstrates the effectiveness of incorporating visual programming and perception modules, highlighting our zero-shot approach in navigating 3DVG." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.825, + 0.472, + 0.867 + ], + "angle": 0, + "content": "Nr3D. Table 3 shows the performance of different approaches on the Nr3D dataset, in which the ground-truth instance mask is also provided. We can see that our zero" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.859, + 0.429 + ], + "angle": 0, + "content": "shot approach further excels the supervised approach InstanceRefer. Specifically, our zero-shot approach on the \"view-dependent\" split can achieve a \\(2\\%\\) accuracy gain than the 3DVG-Transformer approach. This performance gain comes from the relation modules, strengthening the potential of our zero-shot approach for 3DVG tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.438, + 0.667, + 0.453 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.859, + 0.68 + ], + "angle": 0, + "content": "Figure 5 shows the visualizations of the selected samples from the ScanRefer validation set. The four columns present the ground-truth result, the supervised approach BUTD-DETR, the dialog with LLM, and the visual programming approaches, respectively. From Figure 5(a) and Figure 5(b), we can observe that the dialog with LLM and the visual programming approaches can achieve accurate prediction results for view-independent relations, i.e., (above, under) without much training. On the contrary, both the BUTD-DETR and the dialog with LLM approaches cannot address the view-dependent relations, i.e., (left, front), as shown in Figure 5(c) and Figure 5(d). The inherent uncertainty of these relations reflects the limitations of existing methods. However, our visual programming approach can leverage the 2D egocentric views, thus achieving accurate predictions in 3D scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.859, + 0.793 + ], + "angle": 0, + "content": "Figure 5(e) presents a failure case, where the dialog with LLM approach cannot recognize chair has wheels since it lacks open-vocabulary detection ability. Besides, the visual programming approach makes wrong predictions because the LLM cannot correctly recognize the relation pushed. Fortunately, when we correct the program using the CLOsE ST module, the visual programming approach can make correct predictions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.65, + 0.817 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.859, + 0.867 + ], + "angle": 0, + "content": "Dialog with LLM vs. visual programming. We compare the performance of the two proposed zero-shot 3DVG approaches on the ScanRefer validation set with 700 exam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20628" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.147, + 0.193, + 0.166 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.126, + 0.313, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.126, + 0.436, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.126, + 0.559, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.126, + 0.683, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.126, + 0.806, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.216, + 0.191, + 0.234 + ], + "angle": 0, + "content": "BUTD-DETR" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.193, + 0.312, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.193, + 0.436, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.193, + 0.559, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.193, + 0.683, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.193, + 0.805, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.284, + 0.191, + 0.304 + ], + "angle": 0, + "content": "Ours (Dialog)" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.262, + 0.312, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.262, + 0.436, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.262, + 0.559, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.262, + 0.683, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.262, + 0.805, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.356, + 0.197, + 0.377 + ], + "angle": 0, + "content": "Ours (Program)" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.331, + 0.312, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.331, + 0.436, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.331, + 0.559, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.331, + 0.683, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.331, + 0.805, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.402, + 0.315, + 0.432 + ], + "angle": 0, + "content": "It is a window. It is located above a recycle bin that has a blue top." + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.402, + 0.431, + 0.431 + ], + "angle": 0, + "content": "The rolling office chair. The chair is under the desk." + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.401, + 0.555, + 0.431 + ], + "angle": 0, + "content": "There is a square beige armchair. It is left of a square table." + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.402, + 0.668, + 0.432 + ], + "angle": 0, + "content": "This is a brown piano bench. It is in front of the piano." + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.402, + 0.815, + 0.431 + ], + "angle": 0, + "content": "A desk chair is pushed into a small computer desk. The chair has wheels." + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.436, + 0.273, + 0.448 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.436, + 0.39, + 0.448 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.436, + 0.508, + 0.448 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.436, + 0.637, + 0.448 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.436, + 0.765, + 0.448 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.45, + 0.858, + 0.477 + ], + "angle": 0, + "content": "Figure 5. Visualization results of 3D visual grounding. Rendered images of 3D scans are presented, including the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.483, + 0.473, + 0.649 + ], + "angle": 0, + "content": "plies. For both approaches, we use two GPT versions, i.e., GPT-3.5-turbo-0613 and GPT-4-0613. The cost of each GPT version depends on the number of input and output tokens. The experimental results are shown in Table 4. We can observe that for both zero-shot approaches, GPT4-based approach can achieve higher accuracy than GPT3.5-based approach, even it induces a larger economic cost. On the other hand, the visual programming approach always outperforms the dialog with LLM approach in terms of accuracy and cost, which demonstrates the effectiveness of the proposed visual programming approach. For other experiments, we use GPT3.5 to save cost." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.654, + 0.473, + 0.766 + ], + "angle": 0, + "content": "Relation modules. We now ablate different relation modules in Section 3.2 to analyze their impact on the system performance. The most important view-dependent and view-independent modules are presented in Table 5 and 6, respectively. We can see that LEFT and RIGHT are the most important view-dependent relations, while CLOSEST is the most important view-independent relation. This result is coherent with our motivation and design." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.77, + 0.473, + 0.868 + ], + "angle": 0, + "content": "LOC module. We juxtapose our approach by separately omitting the 3D component and 2D component. Both models utilize the instance mask prediction of Mask3D [43]. Particularly, the 2D-only model solely employs the paired 2D images for classification, while the 3D-only model just uses the 3D result. As can be seen from Tables 2 and 3, the 2D-only model performs worst when the images of in" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.481, + 0.832, + 0.555 + ], + "angle": 0, + "content": "
MethodLLMAcc@0.5TokensCost
DialogGPT3.525.41959k$3.05
DialogGPT427.51916k$62.6
ProgramGPT3.532.1121k$0.19
ProgramGPT435.4115k$4.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.564, + 0.859, + 0.591 + ], + "angle": 0, + "content": "Table 4. Performance comparison of the dialog with LLM and the visual programming approaches." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.6, + 0.842, + 0.698 + ], + "angle": 0, + "content": "
LEFTRIGHTFRONTBEHINDBETWEENAccuracy
26.5
✓32.4
✓✓35.9
✓✓✓36.8
✓✓✓✓38.4
✓✓✓✓✓39.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.701, + 0.848, + 0.714 + ], + "angle": 0, + "content": "Table 5. Ablation study of different view-dependent modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.859, + 0.822 + ], + "angle": 0, + "content": "door scenes are complicated and have domain gaps with the training samples. The 3D-only model performs better since it can utilize the geometric information and is trained on closed-set labels. Our full model can always achieve the best performance because it integrates the geometric distinctiveness of point clouds and the open-vocabulary ability of the image models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.859, + 0.868 + ], + "angle": 0, + "content": "Generalization. As discussed in Section 3.4, our framework has strong adaptability for a spectrum of 3D and 2D perception models. To validate this claim, we conduct ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "20629" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.123, + 0.46, + 0.208 + ], + "angle": 0, + "content": "
CLOSESTFARTHESTLOWERHIGHERAccuracy
18.8
✓30.7
✓✓34.0
✓✓✓36.8
✓✓✓✓39.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.212, + 0.468, + 0.225 + ], + "angle": 0, + "content": "Table 6. Ablation study of different view-independent modules." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.235, + 0.473, + 0.414 + ], + "angle": 0, + "content": "periments using several representative models. For 3D perception, we utilize three backbones, i.e., PointNet++ [37], PointNeXt [39], and PointBERT [61]. For 2D perception, we use an image classification model proposed in [41], a visual question answering model in [25], and a general large model BLIP-2 [28] for testing. The results are shown in Tables 7 and 8. We can observe that our framework is compatible with other models. Also, it can leverage the advancements within both 2D and 3D foundational models to improve the performance. This cross-model effectiveness demonstrates the robustness and future-proof nature of our approach in the ever-evolving landscape of visual perception models." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.427, + 0.44, + 0.484 + ], + "angle": 0, + "content": "
2D AssistanceUniqueMultipleAcc@0.25
CLIP62.527.135.7
ViLT60.327.135.1
BLIP-263.827.736.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.161, + 0.49, + 0.424, + 0.503 + ], + "angle": 0, + "content": "Table 7. Ablation study on different 2D models." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.516, + 0.446, + 0.574 + ], + "angle": 0, + "content": "
3D BackboneView-dep.View-indep.Overall
PointNet++35.839.438.2
PointBert36.039.838.6
PointNeXt36.840.039.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.577, + 0.432, + 0.59 + ], + "angle": 0, + "content": "Table 8. Ablation study on different 3D backbones." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.602, + 0.472, + 0.741 + ], + "angle": 0, + "content": "Effect of prompt size. We use different numbers of incontext examples in the prompt for program generation. The result is shown in Figure 6. It can be seen that the performance on ScanRefer and Nr3d improve with the number of examples. This is because more examples can guide LLMs to handle more cases in the visual program generation process. Meanwhile, it still follows the law of diminishing marginal utility. Moreover, we test the voting technique [15] to aggregate the results from multiple runs, which brings some performance gains." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.742, + 0.473, + 0.867 + ], + "angle": 0, + "content": "Error analysis. We select a representative subset for each dataset with around 100 samples and manually check the rationales offered by the visual program. As illustrated in Figure 7, the generation of accurate visual programs is the primary error source. Therefore, the performance can be improved by using more in-context examples and more powerful LLMs. The second error source is the object localization and classification, indicating that object detection and classification in 3D space remains a critical component." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.125, + 0.683, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.125, + 0.837, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.229, + 0.858, + 0.267 + ], + "angle": 0, + "content": "Figure 6. Ablation study on the number of in-context examples. The performance on Nr3D and ScanRefer improves with the number of in-context examples." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.274, + 0.835, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.391, + 0.783, + 0.404 + ], + "angle": 0, + "content": "Figure 7. Breakdown of error sources." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.609, + 0.43 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.86, + 0.591 + ], + "angle": 0, + "content": "In this paper, we present a novel zero-shot approach for 3DVG to eliminate the need for extensive annotations and predefined vocabularies. A vanilla dialog with LLM approach is first proposed by taking interactive dialog with LLMs. A visual programming approach is further developed, which leverages three types of modules to navigate the intricate 3D relations. To adapt to open-vocabulary scenarios, we also develop a LOC module to seamlessly integrate both 3D and 2D features. Experimental results demonstrate the superiority of the proposed approach and highlight its potential to advance the field of 3DVG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.664, + 0.619 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.86, + 0.867 + ], + "angle": 0, + "content": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the National Key R&D Program of China with grant No. 2018YFB1800800, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and the Key Area R&D Program of Guangdong Province with grant No. 2018B03033800, by Tencent & Huawei Open Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "20630" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.201, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.148, + 0.473, + 0.211 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.213, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[2] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In Conference on Robot Learning, pages 287-318. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.279, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.332, + 0.472, + 0.382 + ], + "angle": 0, + "content": "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.383, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[5] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.448, + 0.472, + 0.511 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.513, + 0.472, + 0.577 + ], + "angle": 0, + "content": "[7] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.579, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.619, + 0.472, + 0.669 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.671, + 0.472, + 0.721 + ], + "angle": 0, + "content": "[10] John David N Dionisio, William G Burns Iii, and Richard Gilbert. 3d virtual worlds and the metaverse: Current status and future possibilities. ACM Computing Surveys (CSUR), 45(3):1-38, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.723, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[11] Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D Hwang, et al. Faith and fate: Limits of transformers on compositionality. arXiv preprint arXiv:2305.18654, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.789, + 0.472, + 0.839 + ], + "angle": 0, + "content": "[12] Qi Feng, Vitaly Ablavsky, and Stan Sclaroff. Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint arXiv:2101.04741, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.841, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.148, + 0.473, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.531, + 0.128, + 0.857, + 0.153 + ], + "angle": 0, + "content": "labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.857, + 0.218 + ], + "angle": 0, + "content": "[14] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.857, + 0.281 + ], + "angle": 0, + "content": "[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.857, + 0.346 + ], + "angle": 0, + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.857, + 0.4 + ], + "angle": 0, + "content": "[17] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2614-2623, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.857, + 0.451 + ], + "angle": 0, + "content": "[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. 35th AAAI Conference on Artificial Intelligence, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.857, + 0.503 + ], + "angle": 0, + "content": "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.857, + 0.568 + ], + "angle": 0, + "content": "[20] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.57, + 0.857, + 0.646 + ], + "angle": 0, + "content": "[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 417-433. Springer, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.857, + 0.711 + ], + "angle": 0, + "content": "[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4867-4876, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.857, + 0.774 + ], + "angle": 0, + "content": "[23] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.857, + 0.839 + ], + "angle": 0, + "content": "[24] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.857, + 0.867 + ], + "angle": 0, + "content": "[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region su" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.892, + 0.516, + 0.902 + ], + "angle": 0, + "content": "20631" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.144, + 0.128, + 0.471, + 0.153 + ], + "angle": 0, + "content": "pervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.155, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.205, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[27] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.256, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.307, + 0.472, + 0.38 + ], + "angle": 0, + "content": "[29] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.383, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[30] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.447, + 0.472, + 0.496 + ], + "angle": 0, + "content": "[31] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.498, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.561, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[33] Stylianos Mystakidis. Metaverse. Encyclopedia, 2(1):486-497, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.587, + 0.414, + 0.599 + ], + "angle": 0, + "content": "[34] OpenAI OpenAI. Gpt-4 technical report. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.601, + 0.472, + 0.661 + ], + "angle": 0, + "content": "[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.664, + 0.472, + 0.724 + ], + "angle": 0, + "content": "[36] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.727, + 0.472, + 0.777 + ], + "angle": 0, + "content": "[37] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.779, + 0.472, + 0.839 + ], + "angle": 0, + "content": "[38] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.842, + 0.472, + 0.866 + ], + "angle": 0, + "content": "[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.128, + 0.472, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.128, + 0.857, + 0.165 + ], + "angle": 0, + "content": "Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.857, + 0.228 + ], + "angle": 0, + "content": "[40] Xing-yue Qiu, Chuang-Kai Chiu, Lu-Lu Zhao, Cai-Feng Sun, and Shu-jie Chen. Trends in vr/ar technology-supporting language learning from 2008 to 2019: A research perspective. Interactive Learning Environments, 31 (4):2090-2113, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.231, + 0.857, + 0.305 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.857, + 0.354 + ], + "angle": 0, + "content": "[42] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In 5th Annual Conference on Robot Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.857, + 0.42 + ], + "angle": 0, + "content": "[43] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.422, + 0.857, + 0.47 + ], + "angle": 0, + "content": "[44] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.857, + 0.509 + ], + "angle": 0, + "content": "[45] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.857, + 0.56 + ], + "angle": 0, + "content": "[46] Ayça Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.857, + 0.624 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.626, + 0.857, + 0.674 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.677, + 0.857, + 0.713 + ], + "angle": 0, + "content": "[49] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2:20, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.857, + 0.739 + ], + "angle": 0, + "content": "[50] John Vince and John A Vince. Mathematics for computer graphics. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.857, + 0.801 + ], + "angle": 0, + "content": "[51] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.857, + 0.866 + ], + "angle": 0, + "content": "[52] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance relocalization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7658-7667, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.892, + 0.518, + 0.902 + ], + "angle": 0, + "content": "20632" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.204 + ], + "angle": 0, + "content": "[53] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6629-6638, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.205, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.269, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[55] Wei Wei. Research progress on virtual reality (vr) and augmented reality (ar) in tourism and hospitality: A critical review of publications from 2000 to 2018. Journal of Hospitality and Tourism Technology, 10(4):539-570, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.322, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[56] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.374, + 0.472, + 0.438 + ], + "angle": 0, + "content": "[57] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.439, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.504, + 0.472, + 0.568 + ], + "angle": 0, + "content": "[59] Xu Yan, Zhihao Yuan, Yuhao Du, Yinghong Liao, Yao Guo, Shuguang Cui, and Zhen Li. Comprehensive visual question answering on point clouds through compositional scene manipulation. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.569, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[60] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.633, + 0.472, + 0.697 + ], + "angle": 0, + "content": "[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.699, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[62] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.776, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[63] Zhihao Yuan, Xu Yan, Zhuo Li, Xuhao Li, Yao Guo, Shuguang Cui, and Zhen Li. Toward explainable and fine-grained 3d grounding through referring textual phrases. arXiv preprint arXiv:2207.01821, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.828, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[64] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Zhen Li, and Shuguang Cui. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense caption" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.531, + 0.127, + 0.857, + 0.153 + ], + "angle": 0, + "content": "ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.857, + 0.229 + ], + "angle": 0, + "content": "[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingquiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.231, + 0.857, + 0.293 + ], + "angle": 0, + "content": "[66] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4534-4543, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.857, + 0.358 + ], + "angle": 0, + "content": "[67] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.36, + 0.857, + 0.422 + ], + "angle": 0, + "content": "[68] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.857, + 0.487 + ], + "angle": 0, + "content": "[69] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.857, + 0.552 + ], + "angle": 0, + "content": "[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.857, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "20633" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_origin.pdf b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e191c2f7f24a5b241e5bd8282262be0bdf059e83 --- /dev/null +++ b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/7da4d511-ff1b-4061-9760-a8f55231549a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae5fa2c6b4d9094d2082f15ce4e5322c0ab4f884a5e4efaa1196dc65be536b85 +size 1581426 diff --git a/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/full.md b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..13908ca4f41f3b5cfadca1d7879abe69b4131e44 --- /dev/null +++ b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/full.md @@ -0,0 +1,359 @@ +# Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding + +Zhihao Yuan $^{1,2}$ , Jinke Ren $^{1,2}$ , Chun-Mei Feng $^{3}$ , Hengshuang Zhao $^{4}$ , Shuguang Cui $^{2,1}$ , Zhen Li $^{2,1*}$ + +$^{1}$ FNii, CUHKSZ $^{2}$ SSE, CUHKSZ $^{3}$ IHPC, A*STAR, Singapore $^{4}$ HKU + +# Abstract + +3D Visual Grounding (3DVG) aims at localizing 3D object based on textual descriptions. Conventional supervised methods for 3DVG often necessitate extensive annotations and a predefined vocabulary, which can be restrictive. To address this issue, we propose a novel visual programming approach for zero-shot open-vocabulary 3DVG, leveraging the capabilities of large language models (LLMs). Our approach begins with a unique dialog-based method, engaging with LLMs to establish a foundational understanding of zero-shot 3DVG. Building on this, we design a visual program that consists of three types of modules, i.e., view-independent, view-dependent, and functional modules. These modules, specifically tailored for 3D scenarios, work collaboratively to perform complex reasoning and inference. Furthermore, we develop an innovative language-object correlation module to extend the scope of existing 3D object detectors into open-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot approach can outperform some supervised baselines, marking a significant stride towards effective 3DVG. Code is available at https://curryyuan.github.io/ZSVG3D. + +# 1. Introduction + +3D Visual Grounding (3DVG) aims to localize specific objects within 3D scenes by using a series of textual descriptions. This has become a crucial component in a variety of burgeoning applications, such as autonomous robotics [12, 53, 58], virtual reality [40, 55], and metaverse [10, 33]. For illustration, given a 3D scan in Figure 1(a) along with its description — It is the keyboard closest to the door, the goal of 3DVG is to accurately pinpoint the keyboard in the green box, while eliminating potential distractions such as tables and desks. Despite the apparent simplicity of this task for humans, it poses a significant challenge for machines due to their inherently limited perceptual capabilities. + +![](images/ba5f103fe443dd646aa6fa4627d25c841b3e7ce9197e3635f39d6d0fdf63692a.jpg) +(a) Supervised 3D Visual Grounding + +![](images/7178641d035cdc56f39dcfb55cf758c5f7657f527053888d08fbd2b208116e68.jpg) +(b) Zero-shot 3D Visual Grounding +Figure 1. Comparative overview of two 3DVG approaches, where (a) Supervised 3DVG involves input from 3D scans combined with text queries, guided by object-text pair annotations, (b) Zero-shot 3DVG identifies the location of target objects using programmatic representation generated by LLMs, i.e., target category, anchor category, and relation grounding, thereby highlighting its superiority in decoding spatial relations and object identifiers within a given space, e.g., the location of the keyboard (outlined in green) can be retrieved based on the distance between the keyboard and the door (outlined in blue). + +Traditional supervised 3DVG approaches [19, 60, 62] achieve this objective by leveraging the rich annotations in public datasets, such as ScanRefer [4] and Referit3D [1]. These approaches typically define 3DVG as a matching problem, generating possible objects via 3D detectors [22, 38], and identifying the best match by fusing the visual and textual features. While these approaches can yield precise results, the acquisition of sufficient annotations is prohibitively resource-intensive for real-world applications. Furthermore, these approaches are often constrained by the pre-defined vocabulary during training, making them suboptimal in open-vocabulary scenarios. + +To address these issues, we propose a novel visual programming approach for 3DVG that integrates zero-shot learning and large language models (LLMs). Zero-shot + +learning [65, 67, 70] can generalize across new categories by leveraging the pre-trained capabilities of CLIP [41] in the 3D domain. LLMs [34, 47, 49] can facilitate 3DVG due to their strong planning and reasoning capabilities. Regarding this, we first propose a vanilla version dialog with LLM. It describes the location and size of all objects in the scene and instructs the LLM to distinguish the object of interest through an interactive dialog. Despite the simplicity of the base approach, the inherent stochasticity and control limitations of LLMs make it hard to capture the view-dependent queries and decipher spatial relations in 3D space, which are the main challenges of 3DVG. To overcome this limitation, we further develop a new visual programming approach, as shown in Figure 1(b). It mainly consists of three steps: (1) generating a 3D visual program using LLMs, (2) interpreting the program into Python code, and (3) identifying the target bounding box by executing the code. To enhance the localization accuracy, we further introduce a novel language-object correlation (LOC) module capable of merging the geometric discernment of 3D point clouds with the fine-grained appearance acumen of 2D images. + +In summary, contributions are summarized as follows: + +- We propose an innovative 3D visual programming approach. It eliminates the need for extensive object-text pair annotations required in supervised approaches. +- We transform the visual program to Python code by designing two types of modules, i.e., relation modules and LOC modules. The former explicitly defines the view-dependent and view-independent relations in 3D space, while the latter captures both the geometric and appearance information for open-vocabulary localization. +- We conduct extensive experiments on two popular datasets, i.e., ScanRefer [4] and Nr3d [1]. We for the first time evaluate the whole validation set rather than a few samples. The results demonstrate the superior performance of our approach, even comparable with existing supervised approaches. + +# 2. Related Work + +Supervised 3DVG. 3DVG has received much attention in many emerging applications ranging from automatic robotics [12, 53, 58] to metaverse [10, 33]. On the one hand, densely-annotated datasets like ScanRefer [4] and Referit3D [1] can provide well-aligned object-text pairs for ScanNet [9]. On the other hand, most existing methods [6, 14, 42, 60, 62] treat 3DVG as a matching problem, where object identifiers [22, 38] are utilized to generate candidate objects and find the best-matching one by fusing visual and textual features. Building on this, [57, 63] attempt to explore the object attributes and relations between different proposals. Moreover, some works [7, 23] have also investigated 3D language pretraining using advanced techniques, such as mask modeling and contrastive learning on paired + +object-caption data, followed by finetuning on downstream tasks. Additionally, NS3D [17] has employed CodeX [5] to generate hierarchical programs. However, it still needs annotations to train the neuro-symbolic networks, thus lacking open-vocabulary and zero-shot capabilities. + +Indoor 3D Scene Understanding. 3D scene understanding of indoor environments has been widely studied. In specific, the emergence of RBG-D scans datasets [9, 44, 52] greatly push the boundary of several tasks, including 3D object classification [36, 37], 3D object detection [31, 38], 3D semantic segmentation [8, 66], 3D instance segmentation [22, 43, 51], and so on [16, 59, 64]. However, these methods are often constrained to a closed set of semantic class labels, limiting their applicability in real-world scenarios. Recent progress in open-vocabulary image segmentation [13, 29] has inspired research into 3D scene understanding under the open-vocabulary setting. For instance, LERF [24] learns a language field inside NeRF [32] by volume rendering CLIP [41] features along training rays, enabling it to generate 3D relevancy maps for arbitrary language queries. OpenScene [35] extracts image features using 2D open-vocabulary segmentation models [13, 27], then trained a 3D network to produce point features aligned with multi-view fused pixel features. OpenMask3D [46] utilizes the closed-vocabulary network to generate instance masks while discarded the classification head. Despite these advancements, these methods still lack spatial and commonsense reasoning abilities. + +LLMs for Vision-Language Tasks. Recent progress on LLMs has provided impressive zero-shot planning and reasoning abilities [34, 47, 49]. Advanced prompting technologies such as Least-to-Most [69], Think-Step-by-Step [26], and Chain-of-Thought [54] are proposed to elicit the capabilities of LLMs. These methods can understand human instructions, break complex goals into sub-goals, and control robot agents to execute tasks without additional training [2, 20, 30]. Moreover, when combined with specialized vision models, LLMs can significantly enhance the performance of vision-language tasks. For instance, Visual ChatGPT [56] uses ChatGPT as a central orchestrator, interfacing with a variety of visual foundation models to solve more challenging problems. VISPROG [15] leverages the incontext learning ability to generate high-level modular programs for solving complex and compositional natural language reasoning and image editing tasks. ViperGPT [45] directly feeds the API of available modules to LLM and then generated executable Python code for image grounding. However, leveraging these capabilities for zero-shot 3D language grounding remains an unexplored area. + +# 3. Methodology + +In Section 3.1, we introduce the vanilla approach, i.e., dialog with LLM to overcome the annotation issue in 3DVG. + +![](images/967e80552cedfab2ef3bf8cca03cba085e8ba269a9d5789ab6585bde0f309aa6.jpg) +Figure 2. Overview of two zero-shot approaches for 3DVG. (a) shows the working mechanism of the vanilla dialog with LLM approach. First, we describe the 3DVG task and provide the text descriptions of the room. Then, LLMs identify the objects relevant to the query sentence and perform human-like reasoning. (b) presents the 3D visual programming approach. We first input in-context examples into LLMs. Then, LLMs generate 3D visual programs through the grounding descriptions and perform human-like reasoning. Next, these programs are transformed into executable Python codes via the LOC module for predicting the location of the object. For example, the upper example uses the view-independent module, i.e., CLOSEST to determine the proximity in 3D space, while the lower example applies the view-dependent module, i.e., RIGHT to establish the relative positioning. + +From Section 3.2 to Section 3.4, we present the visual programming approach, address the issue of view-dependent relations, and design the LOC module, respectively. + +# 3.1. Dialog with LLM + +To accomplish the goal of 3DVG, we propose to initiate a dialogue with LLMs. The input for the dialogue consists of a real-world RGB-D scan and a free-form text description $\mathcal{T}$ . The text description provides specific information about the target object within a point cloud representation $\mathbf{P} \in \mathbb{R}^{N \times 6}$ , where $\mathbf{P}$ is a collection of color-enriched 3D points and $N$ is the total number of such points. The LLM acts as an agent located in the scanned room, which aims to identify the specified object based on the given text description. To bridge the gap between the model's proficiency in understanding text and the spatial nature of the 3DVG task, we first transform the scene into a textual narrative. This narrative can provide a comprehensive account of the objects $\mathcal{O}$ presented in the scene, including their positions and dimensions, which can be expressed as: + +```txt +Object is a located at (x, y, z) with sizes (width, length, height). +``` + +Given such textual layout, we dialog with the LLM by providing the scene's description and query. Our objec + +tive is to guide the LLM to identify the object mentioned in the query, while also understand and explain its reasoning process in the identification duration. Particularly, LLM is capable of mimicking the reasoning steps undertaken by humans. As illustrated in Figure 2(a), if the LLM get the object information, it can extract the objects relevant to the query sentence, i.e., targets keyboard and anchors door, and successfully identify the correct target keyboard by calculating its distance with door. + +While LLMs show powerful human-like reasoning capabilities, they still have some limitations. First, it cannot handle the view-dependent issue such as the right window. This is because the 3D scene can freely rotate to different views while it keeps static in 2D images. LLMs usually make decisions by comparing their x-y values of 3D coordinates despite hinting it in the conversation. Second, mathematical calculation is a common weakness of LLMs but is necessary for 3DVG [11]. For example, in Figure 2(a), distance computing is crucial to solve the closest relations, whereas the LLMs cannot always provide accurate results. These two issues stem from LLM's training limitations, which affect the reliability of the dialog with LLM approach. + +# 3.2. 3D Visual Programming + +To address the above two issues, we now introduce a new approach that generates visual programs through LLMs. As + +
View-independentnear, close, next to, far, above, below, under, top, on, opposite, middle
View-dependentfront, behind, back, right, left, facing, leftmost, rightmost, looking, across, between
Functionalmin, max, size, length, width
+ +Table 1. Common relations in 3DVG. + +shown in Figure 2(b), we first construct a set of sample programs to encapsulate human-like problem-solving tactics in 3DVG. Each program includes a sequence of operations, where each operation contains a module name, several input parameters, and an assigned output variable. The output of each step can be reused in the subsequent step, thus creating an interlinked sequence that reflects logical reasoning within a visual context. + +We transform the reasoning process of 3DVG into a scripted visual program. Specifically, we collect a set of in-context examples and the corresponding grounding descriptions, and then use LLMs to extrapolate new visual programs tailored to the task. For example, in Figure 2(b), we consider the task prompted by the following description: + +The round cocktail table in the corner of the room with the blue and yellow poster. + +In this case, the objective is to identify the round cocktail table, which can be transformed into a operation: $\mathrm{BOX0} = \mathrm{LOC}($ 'round cocktail table'), where the LOC operator processes the textual query and outputs the bounding boxes for the target objects. We will elaborate the design of LOC module in Section 3.4. Nevertheless, since there may exist multiple similar objects in 3D scenarios, the identified results may not be unique. To overcome this issue, we further pinpoint the blue and yellow poster as an auxiliary reference point by a operation: $\mathrm{BOX1} = \mathrm{LOC}($ 'blue and yellow poster'). Then, the CLOSEST module computes the proximity between BOX0 (potential tables) and BOX1 (poster), and selects the table closest to the poster as the result. + +Table 1 summarizes the common relations in 3DVG. Based on this, we present the detailed visual program by developing three types of modules tailored for 3D contexts: + +- View-independent modules: They operate on the 3D spatial relations between objects. For example, the CLOS-EST module can discern proximity independent of the viewer's position. +- View-dependent modules: They depend on the observer's vantage point. For instance, the RIGHT module determines the right window (TARGET) when looking at cabinets (BOX1) from all windows (BOX0). +- Functional modules: They include multiple operations such as MIN and MAX, which select objects based on the extremal criteria. + +![](images/989c65269447548de2e43e5f6b17cba535202781d22b6c24b26ab650058ef08d.jpg) +Figure 3. Addressing view-dependent relations: A shift to 2D ego-centric view. + +These three types of modules allow the output of one operation to be fed into another operation, thus providing flexible composability. They not only facilitate structured and accurate inference sequences, but also integrate 3D and 2D data to yield a robust and interpretable result for 3DVG. + +# 3.3. Addressing View-Dependent Relations + +In this section, we discuss the intricacies of the view-dependent relations, which are essential for interpreting spatial relations within 3D space. Particularly, the main challenge is the dynamic nature of these relations that will change with the observer's viewpoint. Although traditional supervised approaches can learn these relations implicitly, they cannot provide a definitive resolution. + +On 2D planes, the relations, especially the left and right are well defined. More specifically, right often corresponds to the positive direction of the x-axis while the left implies the negative direction. Motivated by this, we adopt a 2D egocentric view approach to ensure a consistent frame of reference for the spatial relations in Table 1. + +Our view-dependent modules accept a target argument and an optional anchors parameter. They output the target objects that fulfill the spatial relation to the anchors. When grounding queries do not specify targets, we treat targets as anchors as well. This approach aligns with our intuition, such as identifying the left window by treating all windows themselves as the anchors. + +As shown in Figure 3, we assume there is a virtual camera in the center of the room, i.e., $P_{\mathrm{center}}$ , which can rotate to align with the location of the anchor objects, i.e., $P_{\mathrm{o}_a}$ . The 3D objects are projected onto a 2D plane from this vantage point. Assume that the orthogonal camera has an intrinsic matrix $I$ , then the 2D projections can be obtained by + +$$ +R, T = \operatorname {L o o k a t} \left(P _ {\text {c e n t e r}}, P _ {\mathrm {o} _ {a}}, u p\right), \tag {1} +$$ + +$$ +(u, v, w) ^ {\mathrm {T}} = I \cdot (R | t) \cdot P, \tag {2} +$$ + +where $\mathrm{Lookat}(\cdot)$ is a view transformation function that computes the rotation matrix $R$ and translation matrix $T$ + +[50], $P = (x, y, z, 1)^{\mathrm{T}}$ denotes the 3D coordinate vector, $u$ and $v$ respectively signify the x-axis and y-axis on the 2D plane, and $w$ is the depth value. According to the value of $u$ of an object's center, we can determine its left or right position — a lower $u$ value indicates left. Similarly, $w$ allows us to distinguish the front from behind. By synthesizing these concepts, we can define the between relation. + +The transition from 3D to 2D egocentric perspective provides a clear and consistent solution to interpret view-dependent relations in 3D space, thus enhancing our model's spatial reasoning ability. + +# 3.4. Language-Object Correlation Module + +Although our zero-shot 3DVG approach does not need extensive grounding annotations, it still requires a basic vision model for object localization. To overcome this issue, previous works [4, 62] usually use pre-trained 3D detectors [22, 38] to generate object proposals and the corresponding labels within a fixed vocabulary. However, this approach is restricted to a predefined object class set, thus limiting the scope of class prediction. To enable open-vocabulary segmentation, we develop an LOC module, combining the advantages of 3D and 2D networks to extend the labeling capability beyond the closed set. For example, in Figure 4, considering the operation: $\mathrm{BOX0} = \mathrm{LOC}($ 'round cocktail table'), we first filter a subset of objects whose predicted label is table using a 3D instance segmentation network [43]. Then we only need to identify a round cocktail table from this subset using the corresponding 2D imagery. By mapping each 3D proposal to its 2D image, we can extract the color and texture details pertinent to our query. To further pinpoint the round cocktail table shown in the Figure 4, we consider three types of 2D multi-modal models: + +- Image classification models: We construct a dynamic vocabulary, including both the query term "round cocktail table" and the class "table" using popular tools such as CLIP [41]. Then we evaluate the cosine similarity between these terms and the imagery to find the best correlation to our query. +- Visual question answering models: We raise the question: Is there a [query]? to the model such as ViLT [25]. Then the model sifts through its dictionary to suggest the most likely answer, i.e., yes or no. +- General large models: We submit the same inquiry and anticipate a response based on the generated text. This process is crucial for verifying the alignment between the detected table and the query. + +We shall note that our approach is not limited to specific 3D or 2D models, allowing versatile incorporation of various models. In the experiments, we will demonstrate that the benefit of the LOC modules by comparing with the 3D-only and 2D-only counterparts. Our design indicates a leap forward in 3D open-vocabulary instance segmentation and + +![](images/e7d8fe31afa4235c96087cf3d70416266f850cfd2aa4ae9951261021ac40fdc4.jpg) +BOX0=LOC(object='round cocktail table') +Figure 4. Illustration of the language-object correlation module. + +can improve the object recognition accuracy in 3DVG. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We use two popular datasets, i.e., ScanRefer [4] and Nr3D [1] for experiments. ScanRefer is tailored for 3DVG that contains 51,500 sentence descriptions for 800 ScanNet scenes [9]. Nr3D is a human-written and free-form dataset for 3DVG, collected by 2-player reference game in 3D scenes. The sentences are divided into "easy" and "hard" subsets, where the target object only contains one same-class distractor in the "easy" subset but contains multiple ones in the "hard" subset. Depending on whether the sentence requires a specific viewpoint to ground the referred object, the dataset can also be partitioned into "view dependent" and "view independent" subsets. For both datasets, we evaluate the zero-shot approaches on the validation split. + +Evaluation metrics. We consider two settings for performance evaluation. The first one mandates the generation of object proposals, aligning closely with real-world applications. The evaluation metrics are Acc@0.25 and Acc@0.5, representing the percentage of correctly predicted bounding boxes whose IoU exceeds 0.25 or 0.5 with the ground-truth, respectively. This is the default setting for ScanRefer dataset. The second one furnishes ground-truth object masks, necessitating only classification, with an objective to eradicate localization error and achieve high grounding accuracy. This is the default setting for Nr3D dataset. + +Baselines. We use six supervised and two open-vocabulary 3D scene understanding approaches for performance comparison. For supervised approaches, ScanRefer [4] and ReferIt3DNet [1] encode the 3D point clouds and language separately, and then fuse them to rank the objects by predicted scores. TGNN [18] and InstanceRefer [62] make one further step by learning instance-wise features. 3DVG-Transformer [68] and BUTD-DETR [21] respectively utilize the Transformer [48] and DETR [3] architectures, representing the SoTA approaches. For open-vocabulary ap + +
MethodsSupervisionUniqueMultipleOverall
Acc@0.25Acc@0.5Acc@0.25Acc@0.5Acc@0.25Acc@0.5
ScanRefer [4]fully65.043.330.619.837.324.3
TGNN [18]fully64.553.027.021.934.329.7
InstanceRefer [62]fully77.566.831.324.840.232.9
3DVG-Transformer [68]fully81.960.639.328.447.634.7
BUTD-DETR [21]fully84.266.346.635.152.239.8
LERF [24]-----4.80.9
OpenScene [35]-20.113.111.14.413.26.5
Ours (2D only)-32.527.816.114.620.017.6
Ours (3D only)-57.149.425.923.333.129.3
Ours-63.858.427.724.636.432.7
+ +Table 2. 3DVG results on ScanRefer validation set. The accuracy on the "unique" subset, "multiple" subset, and whole validation set are all provided. Following [4], we label the scene as "unique" if it only contains a single object of its class. Otherwise, we label it as "multiple". + +
MethodEasyHardDep.Indep.Overall
ReferIt3DNet [1]43.627.932.537.135.6
InstanceRefer [62]46.031.834.541.938.8
3DVG-Transformer [68]48.534.834.843.740.8
BUTD-DETR [21]60.748.446.058.054.6
Ours (2D only)29.418.423.023.923.6
Ours (3D only)45.927.934.938.436.7
Ours46.531.736.840.039.0
+ +Table 3. Performance analysis of language grounding on Nr3D. We evaluate the top-1 accuracy using ground-truth boxes. + +proaches, OpenScene [35] and LERF [24] aims to learn a 3D representation aligned with the 2D CLIP feature, thus enabling free-form language grounding. The query $\mathcal{T}$ is processed by the CLIP text encoder, and its similarity is computed against the extracted point features. Finally, they cluster the points with the highest score to determine the target object. + +# 4.2. Quantitative Results + +ScanRefer. Table 2 provides a quantitative assessment of the proposed approach on the ScanRefer dataset. We can see that our zero-shot approach outperforms all baseline approaches. Specifically, our approach can achieve a 32.7 Acc@0.5 score, which surpasses part of supervised approaches such as ScanRefer and TGNN. On the other hand, the open-vocabulary approaches LERF and OpenScene can respectively achieve the overall accuracy of 4.8 and 13.2, even with the 0.25 IoU threshold. This is due to their limitations in reasoning and localization precision. Moreover, our zero-shot approach outperforms the approaches that only utilize the 3D or 2D information in the LOC module. This result demonstrates the effectiveness of incorporating visual programming and perception modules, highlighting our zero-shot approach in navigating 3DVG. + +Nr3D. Table 3 shows the performance of different approaches on the Nr3D dataset, in which the ground-truth instance mask is also provided. We can see that our zero + +shot approach further excels the supervised approach InstanceRefer. Specifically, our zero-shot approach on the "view-dependent" split can achieve a $2\%$ accuracy gain than the 3DVG-Transformer approach. This performance gain comes from the relation modules, strengthening the potential of our zero-shot approach for 3DVG tasks. + +# 4.3. Qualitative Results + +Figure 5 shows the visualizations of the selected samples from the ScanRefer validation set. The four columns present the ground-truth result, the supervised approach BUTD-DETR, the dialog with LLM, and the visual programming approaches, respectively. From Figure 5(a) and Figure 5(b), we can observe that the dialog with LLM and the visual programming approaches can achieve accurate prediction results for view-independent relations, i.e., (above, under) without much training. On the contrary, both the BUTD-DETR and the dialog with LLM approaches cannot address the view-dependent relations, i.e., (left, front), as shown in Figure 5(c) and Figure 5(d). The inherent uncertainty of these relations reflects the limitations of existing methods. However, our visual programming approach can leverage the 2D egocentric views, thus achieving accurate predictions in 3D scenarios. + +Figure 5(e) presents a failure case, where the dialog with LLM approach cannot recognize chair has wheels since it lacks open-vocabulary detection ability. Besides, the visual programming approach makes wrong predictions because the LLM cannot correctly recognize the relation pushed. Fortunately, when we correct the program using the CLOsE ST module, the visual programming approach can make correct predictions. + +# 4.4. Ablation Studies + +Dialog with LLM vs. visual programming. We compare the performance of the two proposed zero-shot 3DVG approaches on the ScanRefer validation set with 700 exam + +![](images/cbabafc01d0e613c1a286cf7d153edc1ad7a2af009162624d08d3c7a96d26618.jpg) +Ground Truth + +![](images/663372fcc575527cdadb0a2616789a4b8ca5cbdae45ad1074d62a5cf02604efc.jpg) + +![](images/f8e529c60bea8d85d165407542a0047f975da91c29aba43ace5a447ce1b28fc0.jpg) + +![](images/059b07e9a2dc57e0542ca06a77b4c86d8c8f8f8b9de7b6d0ef6d14432d90fcea.jpg) + +![](images/5d21a3ffea9c8efd25c257ede92c1f2b99caf3d58f9407ac293967409beb255b.jpg) + +![](images/d0e273c6ef725411e30020c0a5ce9ad515b75d19d5da72e53068830a66f9249e.jpg) +BUTD-DETR + +![](images/71e666e749c290cacd6753be38393b19dd2cdde973281576979007d677149f72.jpg) + +![](images/67be939add2aa8f86ceba45bb88f0d502bf9c0ae31f9840363187c7a696751d5.jpg) + +![](images/3d2b28f6e8797a5a8c6a3f868f92fd6c9907fe89e93fd46df04b7ea59598d555.jpg) + +![](images/b279d8359f8d86eb7a1eac1662b9c0aa53dd25941d07bf7a38e130c1a10807f6.jpg) + +![](images/234491251d07a6675289c89e06ebe712f07a0213505e76b41491407aa0936fa9.jpg) +Ours (Dialog) + +![](images/f81117909218875e75e2de5e8b92bafe7139e7f8511ceba47ecf1e24d2520a26.jpg) + +![](images/5f21d468f1678d9a1380ef613f8a65035d2704c51ad6594cdcf8b7536e86c9de.jpg) + +![](images/c0a94a61edc601d5004913c5e5be966429f345b0bb7643d4010c8b3795970d1e.jpg) + +![](images/8c7e229a58b25ae8df5f6bbb5709351a411c856b14d0a89c5e35fc7e0ebfd837.jpg) + +![](images/bc34a793a88f176b9b7725922d0d908e7cb94c9b2ff723b50d444e03ac728929.jpg) +Ours (Program) +It is a window. It is located above a recycle bin that has a blue top. +(a) +Figure 5. Visualization results of 3D visual grounding. Rendered images of 3D scans are presented, including the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green). + +![](images/3a38076226530917f1a5fd4642de42e54682f8823abf7bdb5673c03a7d581e82.jpg) +The rolling office chair. The chair is under the desk. +(b) + +![](images/9d255a7bd054511bc317938843d326d757644262ef78a799c2139607988fc498.jpg) +There is a square beige armchair. It is left of a square table. +(c) + +![](images/a13609d7839721aeeb3f30421bc00cc4eb7cbff21e2165025c56ddf99fa60dd3.jpg) +This is a brown piano bench. It is in front of the piano. +(d) + +![](images/a0bc77b8506f99caf5444fc935798c34e269e2efb8cb2a398a7faa214b3a3368.jpg) +A desk chair is pushed into a small computer desk. The chair has wheels. +(e) + +plies. For both approaches, we use two GPT versions, i.e., GPT-3.5-turbo-0613 and GPT-4-0613. The cost of each GPT version depends on the number of input and output tokens. The experimental results are shown in Table 4. We can observe that for both zero-shot approaches, GPT4-based approach can achieve higher accuracy than GPT3.5-based approach, even it induces a larger economic cost. On the other hand, the visual programming approach always outperforms the dialog with LLM approach in terms of accuracy and cost, which demonstrates the effectiveness of the proposed visual programming approach. For other experiments, we use GPT3.5 to save cost. + +Relation modules. We now ablate different relation modules in Section 3.2 to analyze their impact on the system performance. The most important view-dependent and view-independent modules are presented in Table 5 and 6, respectively. We can see that LEFT and RIGHT are the most important view-dependent relations, while CLOSEST is the most important view-independent relation. This result is coherent with our motivation and design. + +LOC module. We juxtapose our approach by separately omitting the 3D component and 2D component. Both models utilize the instance mask prediction of Mask3D [43]. Particularly, the 2D-only model solely employs the paired 2D images for classification, while the 3D-only model just uses the 3D result. As can be seen from Tables 2 and 3, the 2D-only model performs worst when the images of in + +
MethodLLMAcc@0.5TokensCost
DialogGPT3.525.41959k$3.05
DialogGPT427.51916k$62.6
ProgramGPT3.532.1121k$0.19
ProgramGPT435.4115k$4.24
+ +Table 4. Performance comparison of the dialog with LLM and the visual programming approaches. + +
LEFTRIGHTFRONTBEHINDBETWEENAccuracy
26.5
✓32.4
✓✓35.9
✓✓✓36.8
✓✓✓✓38.4
✓✓✓✓✓39.0
+ +Table 5. Ablation study of different view-dependent modules. + +door scenes are complicated and have domain gaps with the training samples. The 3D-only model performs better since it can utilize the geometric information and is trained on closed-set labels. Our full model can always achieve the best performance because it integrates the geometric distinctiveness of point clouds and the open-vocabulary ability of the image models. + +Generalization. As discussed in Section 3.4, our framework has strong adaptability for a spectrum of 3D and 2D perception models. To validate this claim, we conduct ex + +
CLOSESTFARTHESTLOWERHIGHERAccuracy
18.8
✓30.7
✓✓34.0
✓✓✓36.8
✓✓✓✓39.0
+ +periments using several representative models. For 3D perception, we utilize three backbones, i.e., PointNet++ [37], PointNeXt [39], and PointBERT [61]. For 2D perception, we use an image classification model proposed in [41], a visual question answering model in [25], and a general large model BLIP-2 [28] for testing. The results are shown in Tables 7 and 8. We can observe that our framework is compatible with other models. Also, it can leverage the advancements within both 2D and 3D foundational models to improve the performance. This cross-model effectiveness demonstrates the robustness and future-proof nature of our approach in the ever-evolving landscape of visual perception models. + +Table 6. Ablation study of different view-independent modules. + +
2D AssistanceUniqueMultipleAcc@0.25
CLIP62.527.135.7
ViLT60.327.135.1
BLIP-263.827.736.4
+ +Table 7. Ablation study on different 2D models. + +
3D BackboneView-dep.View-indep.Overall
PointNet++35.839.438.2
PointBert36.039.838.6
PointNeXt36.840.039.0
+ +Table 8. Ablation study on different 3D backbones. + +Effect of prompt size. We use different numbers of incontext examples in the prompt for program generation. The result is shown in Figure 6. It can be seen that the performance on ScanRefer and Nr3d improve with the number of examples. This is because more examples can guide LLMs to handle more cases in the visual program generation process. Meanwhile, it still follows the law of diminishing marginal utility. Moreover, we test the voting technique [15] to aggregate the results from multiple runs, which brings some performance gains. + +Error analysis. We select a representative subset for each dataset with around 100 samples and manually check the rationales offered by the visual program. As illustrated in Figure 7, the generation of accurate visual programs is the primary error source. Therefore, the performance can be improved by using more in-context examples and more powerful LLMs. The second error source is the object localization and classification, indicating that object detection and classification in 3D space remains a critical component. + +![](images/81a1d1b34240c8702a3a2779df88df2b625304a1a4a601ae1fcfb421705fa0db.jpg) +Figure 6. Ablation study on the number of in-context examples. The performance on Nr3D and ScanRefer improves with the number of in-context examples. + +![](images/72f79338576ca87efd0ee73cbc9f8c0035c7951e1923b36375360755cfa68e3d.jpg) + +![](images/78d7351c3a5f5cfd39dd44ab859742cfc02ef8e61dbf123accc98330310c64f4.jpg) +Figure 7. Breakdown of error sources. + +# 5. Conclusion + +In this paper, we present a novel zero-shot approach for 3DVG to eliminate the need for extensive annotations and predefined vocabularies. A vanilla dialog with LLM approach is first proposed by taking interactive dialog with LLMs. A visual programming approach is further developed, which leverages three types of modules to navigate the intricate 3D relations. To adapt to open-vocabulary scenarios, we also develop a LOC module to seamlessly integrate both 3D and 2D features. Experimental results demonstrate the superiority of the proposed approach and highlight its potential to advance the field of 3DVG. + +# 6. Acknowledgement + +This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the National Key R&D Program of China with grant No. 2018YFB1800800, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and the Key Area R&D Program of Guangdong Province with grant No. 2018B03033800, by Tencent & Huawei Open Fund. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 1, 2, 5, 6 +[2] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In Conference on Robot Learning, pages 287-318. PMLR, 2023. 2 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5 +[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 2, 5, 6 +[5] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 2 +[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 2 +[7] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2 +[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2, 5 +[10] John David N Dionisio, William G Burns Iii, and Richard Gilbert. 3d virtual worlds and the metaverse: Current status and future possibilities. ACM Computing Surveys (CSUR), 45(3):1-38, 2013. 1, 2 +[11] Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D Hwang, et al. Faith and fate: Limits of transformers on compositionality. arXiv preprint arXiv:2305.18654, 2023. 3 +[12] Qi Feng, Vitaly Ablavsky, and Stan Sclaroff. Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint arXiv:2101.04741, 2021. 1, 2 +[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level + +labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2 +[14] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 2 +[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 8 +[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 2 +[17] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2614-2623, 2023. 2 +[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. 35th AAAI Conference on Artificial Intelligence, 2021. 5, 6 +[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1 +[20] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2 +[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 417-433. Springer, 2022. 5, 6 +[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4867-4876, 2020. 1, 2, 5 +[23] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2 +[24] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 6 +[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region su + +pervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5, 8 +[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2 +[27] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8 +[29] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2 +[30] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2 +[31] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[33] Stylianos Mystakidis. Metaverse. Encyclopedia, 2(1):486-497, 2022. 1, 2 +[34] OpenAI OpenAI. Gpt-4 technical report. 2023. 2 +[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2, 6 +[36] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[37] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 8 +[38] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5 +[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard + +Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 8 +[40] Xing-yue Qiu, Chuang-Kai Chiu, Lu-Lu Zhao, Cai-Feng Sun, and Shu-jie Chen. Trends in vr/ar technology-supporting language learning from 2008 to 2019: A research perspective. Interactive Learning Environments, 31 (4):2090-2113, 2023. 1 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5, 8 +[42] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In 5th Annual Conference on Robot Learning, 2021. 2 +[43] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 5, 7 +[44] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 2 +[45] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2 +[46] Ayça Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2 +[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5 +[49] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2:20, 2023. 2 +[50] John Vince and John A Vince. Mathematics for computer graphics. Springer, 2006. 5 +[51] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 2 +[52] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance relocalization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7658-7667, 2019. 2 + +[53] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6629-6638, 2019. 1, 2 +[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2 +[55] Wei Wei. Research progress on virtual reality (vr) and augmented reality (ar) in tourism and hospitality: A critical review of publications from 2000 to 2018. Journal of Hospitality and Tourism Technology, 10(4):539-570, 2019. 1 +[56] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2 +[57] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2 +[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1, 2 +[59] Xu Yan, Zhihao Yuan, Yuhao Du, Yinghong Liao, Yao Guo, Shuguang Cui, and Zhen Li. Comprehensive visual question answering on point clouds through compositional scene manipulation. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[60] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2 +[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8 +[62] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 1, 2, 5, 6 +[63] Zhihao Yuan, Xu Yan, Zhuo Li, Xuhao Li, Yao Guo, Shuguang Cui, and Zhen Li. Toward explainable and fine-grained 3d grounding through referring textual phrases. arXiv preprint arXiv:2207.01821, 2022. 2 +[64] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Zhen Li, and Shuguang Cui. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense caption + +ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingquiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2023. 2 +[66] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4534-4543, 2020. 2 +[67] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2 +[68] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 5, 6 +[69] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 2 +[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 2 \ No newline at end of file diff --git a/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/images.zip b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5177b546c98816afdfae2e36e8a059737b371440 --- /dev/null +++ b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46df21f32b03943896f86b6c6fc4b320347212b8a7a389ba9a8e264b36f5ffb7 +size 566240 diff --git a/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/layout.json b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38d5a6ace2b3416ec47c855b4689d9bed6dd1ec1 --- /dev/null +++ b/2024/Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding/layout.json @@ -0,0 +1,9325 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 128, + 510, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 128, + 510, + 144 + ], + "spans": [ + { + "bbox": [ + 85, + 128, + 510, + 144 + ], + "type": "text", + "content": "Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "content": "Zhihao Yuan" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "content": ", Jinke Ren" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "content": ", Chun-Mei Feng" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "content": ", Hengshuang Zhao" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "content": ", Shuguang Cui" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "text", + "content": ", Zhen Li" + }, + { + "bbox": [ + 83, + 163, + 514, + 178 + ], + "type": "inline_equation", + "content": "^{2,1*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "text", + "content": " FNii, CUHKSZ " + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "text", + "content": " SSE, CUHKSZ " + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "text", + "content": " IHPC, A*STAR, Singapore " + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 123, + 190, + 470, + 204 + ], + "type": "text", + "content": " HKU" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 251, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 251, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 290, + 483 + ], + "type": "text", + "content": "3D Visual Grounding (3DVG) aims at localizing 3D object based on textual descriptions. Conventional supervised methods for 3DVG often necessitate extensive annotations and a predefined vocabulary, which can be restrictive. To address this issue, we propose a novel visual programming approach for zero-shot open-vocabulary 3DVG, leveraging the capabilities of large language models (LLMs). Our approach begins with a unique dialog-based method, engaging with LLMs to establish a foundational understanding of zero-shot 3DVG. Building on this, we design a visual program that consists of three types of modules, i.e., view-independent, view-dependent, and functional modules. These modules, specifically tailored for 3D scenarios, work collaboratively to perform complex reasoning and inference. Furthermore, we develop an innovative language-object correlation module to extend the scope of existing 3D object detectors into open-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot approach can outperform some supervised baselines, marking a significant stride towards effective 3DVG. Code is available at https://curryyuan.github.io/ZSVG3D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 505, + 143, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 505, + 143, + 516 + ], + "spans": [ + { + "bbox": [ + 69, + 505, + 143, + 516 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 524, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 524, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 67, + 524, + 288, + 666 + ], + "type": "text", + "content": "3D Visual Grounding (3DVG) aims to localize specific objects within 3D scenes by using a series of textual descriptions. This has become a crucial component in a variety of burgeoning applications, such as autonomous robotics [12, 53, 58], virtual reality [40, 55], and metaverse [10, 33]. For illustration, given a 3D scan in Figure 1(a) along with its description — It is the keyboard closest to the door, the goal of 3DVG is to accurately pinpoint the keyboard in the green box, while eliminating potential distractions such as tables and desks. Despite the apparent simplicity of this task for humans, it poses a significant challenge for machines due to their inherently limited perceptual capabilities." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 228, + 525, + 297 + ], + "blocks": [ + { + "bbox": [ + 309, + 228, + 525, + 297 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 525, + 297 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 525, + 297 + ], + "type": "image", + "image_path": "ba5f103fe443dd646aa6fa4627d25c841b3e7ce9197e3635f39d6d0fdf63692a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 300, + 463, + 308 + ], + "lines": [ + { + "bbox": [ + 372, + 300, + 463, + 308 + ], + "spans": [ + { + "bbox": [ + 372, + 300, + 463, + 308 + ], + "type": "text", + "content": "(a) Supervised 3D Visual Grounding" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 312, + 523, + 394 + ], + "blocks": [ + { + "bbox": [ + 308, + 312, + 523, + 394 + ], + "lines": [ + { + "bbox": [ + 308, + 312, + 523, + 394 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 523, + 394 + ], + "type": "image", + "image_path": "7178641d035cdc56f39dcfb55cf758c5f7657f527053888d08fbd2b208116e68.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 397, + 465, + 406 + ], + "lines": [ + { + "bbox": [ + 376, + 397, + 465, + 406 + ], + "spans": [ + { + "bbox": [ + 376, + 397, + 465, + 406 + ], + "type": "text", + "content": "(b) Zero-shot 3D Visual Grounding" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 411, + 526, + 512 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 526, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 526, + 512 + ], + "type": "text", + "content": "Figure 1. Comparative overview of two 3DVG approaches, where (a) Supervised 3DVG involves input from 3D scans combined with text queries, guided by object-text pair annotations, (b) Zero-shot 3DVG identifies the location of target objects using programmatic representation generated by LLMs, i.e., target category, anchor category, and relation grounding, thereby highlighting its superiority in decoding spatial relations and object identifiers within a given space, e.g., the location of the keyboard (outlined in green) can be retrieved based on the distance between the keyboard and the door (outlined in blue)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 525, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 525, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 525, + 652 + ], + "type": "text", + "content": "Traditional supervised 3DVG approaches [19, 60, 62] achieve this objective by leveraging the rich annotations in public datasets, such as ScanRefer [4] and Referit3D [1]. These approaches typically define 3DVG as a matching problem, generating possible objects via 3D detectors [22, 38], and identifying the best match by fusing the visual and textual features. While these approaches can yield precise results, the acquisition of sufficient annotations is prohibitively resource-intensive for real-world applications. Furthermore, these approaches are often constrained by the pre-defined vocabulary during training, making them suboptimal in open-vocabulary scenarios." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 653, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 525, + 687 + ], + "type": "text", + "content": "To address these issues, we propose a novel visual programming approach for 3DVG that integrates zero-shot learning and large language models (LLMs). Zero-shot" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "spans": [ + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "spans": [ + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 676, + 150, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 676, + 150, + 686 + ], + "spans": [ + { + "bbox": [ + 80, + 676, + 150, + 686 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20623" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 99, + 289, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 289, + 329 + ], + "type": "text", + "content": "learning [65, 67, 70] can generalize across new categories by leveraging the pre-trained capabilities of CLIP [41] in the 3D domain. LLMs [34, 47, 49] can facilitate 3DVG due to their strong planning and reasoning capabilities. Regarding this, we first propose a vanilla version dialog with LLM. It describes the location and size of all objects in the scene and instructs the LLM to distinguish the object of interest through an interactive dialog. Despite the simplicity of the base approach, the inherent stochasticity and control limitations of LLMs make it hard to capture the view-dependent queries and decipher spatial relations in 3D space, which are the main challenges of 3DVG. To overcome this limitation, we further develop a new visual programming approach, as shown in Figure 1(b). It mainly consists of three steps: (1) generating a 3D visual program using LLMs, (2) interpreting the program into Python code, and (3) identifying the target bounding box by executing the code. To enhance the localization accuracy, we further introduce a novel language-object correlation (LOC) module capable of merging the geometric discernment of 3D point clouds with the fine-grained appearance acumen of 2D images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 330, + 281, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 330, + 281, + 340 + ], + "spans": [ + { + "bbox": [ + 79, + 330, + 281, + 340 + ], + "type": "text", + "content": "In summary, contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 341, + 288, + 505 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 69, + 341, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 341, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 69, + 341, + 288, + 373 + ], + "type": "text", + "content": "- We propose an innovative 3D visual programming approach. It eliminates the need for extensive object-text pair annotations required in supervised approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 374, + 288, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 374, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 69, + 374, + 288, + 438 + ], + "type": "text", + "content": "- We transform the visual program to Python code by designing two types of modules, i.e., relation modules and LOC modules. The former explicitly defines the view-dependent and view-independent relations in 3D space, while the latter captures both the geometric and appearance information for open-vocabulary localization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 439, + 288, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 439, + 288, + 505 + ], + "spans": [ + { + "bbox": [ + 69, + 439, + 288, + 505 + ], + "type": "text", + "content": "- We conduct extensive experiments on two popular datasets, i.e., ScanRefer [4] and Nr3d [1]. We for the first time evaluate the whole validation set rather than a few samples. The results demonstrate the superior performance of our approach, even comparable with existing supervised approaches." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 514, + 147, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 514, + 147, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 514, + 147, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 533, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 533, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 533, + 289, + 686 + ], + "type": "text", + "content": "Supervised 3DVG. 3DVG has received much attention in many emerging applications ranging from automatic robotics [12, 53, 58] to metaverse [10, 33]. On the one hand, densely-annotated datasets like ScanRefer [4] and Referit3D [1] can provide well-aligned object-text pairs for ScanNet [9]. On the other hand, most existing methods [6, 14, 42, 60, 62] treat 3DVG as a matching problem, where object identifiers [22, 38] are utilized to generate candidate objects and find the best-matching one by fusing visual and textual features. Building on this, [57, 63] attempt to explore the object attributes and relations between different proposals. Moreover, some works [7, 23] have also investigated 3D language pretraining using advanced techniques, such as mask modeling and contrastive learning on paired" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 100, + 525, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 525, + 154 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 525, + 154 + ], + "type": "text", + "content": "object-caption data, followed by finetuning on downstream tasks. Additionally, NS3D [17] has employed CodeX [5] to generate hierarchical programs. However, it still needs annotations to train the neuro-symbolic networks, thus lacking open-vocabulary and zero-shot capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 155, + 525, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 525, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 525, + 406 + ], + "type": "text", + "content": "Indoor 3D Scene Understanding. 3D scene understanding of indoor environments has been widely studied. In specific, the emergence of RBG-D scans datasets [9, 44, 52] greatly push the boundary of several tasks, including 3D object classification [36, 37], 3D object detection [31, 38], 3D semantic segmentation [8, 66], 3D instance segmentation [22, 43, 51], and so on [16, 59, 64]. However, these methods are often constrained to a closed set of semantic class labels, limiting their applicability in real-world scenarios. Recent progress in open-vocabulary image segmentation [13, 29] has inspired research into 3D scene understanding under the open-vocabulary setting. For instance, LERF [24] learns a language field inside NeRF [32] by volume rendering CLIP [41] features along training rays, enabling it to generate 3D relevancy maps for arbitrary language queries. OpenScene [35] extracts image features using 2D open-vocabulary segmentation models [13, 27], then trained a 3D network to produce point features aligned with multi-view fused pixel features. OpenMask3D [46] utilizes the closed-vocabulary network to generate instance masks while discarded the classification head. Despite these advancements, these methods still lack spatial and commonsense reasoning abilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 406, + 525, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 525, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 525, + 636 + ], + "type": "text", + "content": "LLMs for Vision-Language Tasks. Recent progress on LLMs has provided impressive zero-shot planning and reasoning abilities [34, 47, 49]. Advanced prompting technologies such as Least-to-Most [69], Think-Step-by-Step [26], and Chain-of-Thought [54] are proposed to elicit the capabilities of LLMs. These methods can understand human instructions, break complex goals into sub-goals, and control robot agents to execute tasks without additional training [2, 20, 30]. Moreover, when combined with specialized vision models, LLMs can significantly enhance the performance of vision-language tasks. For instance, Visual ChatGPT [56] uses ChatGPT as a central orchestrator, interfacing with a variety of visual foundation models to solve more challenging problems. VISPROG [15] leverages the incontext learning ability to generate high-level modular programs for solving complex and compositional natural language reasoning and image editing tasks. ViperGPT [45] directly feeds the API of available modules to LLM and then generated executable Python code for image grounding. However, leveraging these capabilities for zero-shot 3D language grounding remains an unexplored area." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 646, + 381, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 381, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 381, + 658 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 664, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 525, + 686 + ], + "type": "text", + "content": "In Section 3.1, we introduce the vanilla approach, i.e., dialog with LLM to overcome the annotation issue in 3DVG." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20624" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 101, + 523, + 304 + ], + "blocks": [ + { + "bbox": [ + 71, + 101, + 523, + 304 + ], + "lines": [ + { + "bbox": [ + 71, + 101, + 523, + 304 + ], + "spans": [ + { + "bbox": [ + 71, + 101, + 523, + 304 + ], + "type": "image", + "image_path": "967e80552cedfab2ef3bf8cca03cba085e8ba269a9d5789ab6585bde0f309aa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 312, + 527, + 384 + ], + "lines": [ + { + "bbox": [ + 67, + 312, + 527, + 384 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 527, + 384 + ], + "type": "text", + "content": "Figure 2. Overview of two zero-shot approaches for 3DVG. (a) shows the working mechanism of the vanilla dialog with LLM approach. First, we describe the 3DVG task and provide the text descriptions of the room. Then, LLMs identify the objects relevant to the query sentence and perform human-like reasoning. (b) presents the 3D visual programming approach. We first input in-context examples into LLMs. Then, LLMs generate 3D visual programs through the grounding descriptions and perform human-like reasoning. Next, these programs are transformed into executable Python codes via the LOC module for predicting the location of the object. For example, the upper example uses the view-independent module, i.e., CLOSEST to determine the proximity in 3D space, while the lower example applies the view-dependent module, i.e., RIGHT to establish the relative positioning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 401, + 290, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 401, + 290, + 436 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 290, + 436 + ], + "type": "text", + "content": "From Section 3.2 to Section 3.4, we present the visual programming approach, address the issue of view-dependent relations, and design the LOC module, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 441, + 165, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 441, + 165, + 452 + ], + "spans": [ + { + "bbox": [ + 69, + 441, + 165, + 452 + ], + "type": "text", + "content": "3.1. Dialog with LLM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "content": "To accomplish the goal of 3DVG, we propose to initiate a dialogue with LLMs. The input for the dialogue consists of a real-world RGB-D scan and a free-form text description " + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "content": ". The text description provides specific information about the target object within a point cloud representation " + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 6}" + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "content": " is a collection of color-enriched 3D points and " + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "content": " is the total number of such points. The LLM acts as an agent located in the scanned room, which aims to identify the specified object based on the given text description. To bridge the gap between the model's proficiency in understanding text and the spatial nature of the 3DVG task, we first transform the scene into a textual narrative. This narrative can provide a comprehensive account of the objects " + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 67, + 457, + 290, + 622 + ], + "type": "text", + "content": " presented in the scene, including their positions and dimensions, which can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 87, + 627, + 260, + 660 + ], + "blocks": [ + { + "bbox": [ + 87, + 627, + 260, + 660 + ], + "lines": [ + { + "bbox": [ + 87, + 627, + 260, + 660 + ], + "spans": [ + { + "bbox": [ + 87, + 627, + 260, + 660 + ], + "type": "text", + "content": "Object is a located at (x, y, z) with sizes (width, length, height)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 68, + 664, + 290, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 664, + 290, + 688 + ], + "spans": [ + { + "bbox": [ + 68, + 664, + 290, + 688 + ], + "type": "text", + "content": "Given such textual layout, we dialog with the LLM by providing the scene's description and query. Our objec" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 402, + 526, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 526, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 526, + 499 + ], + "type": "text", + "content": "tive is to guide the LLM to identify the object mentioned in the query, while also understand and explain its reasoning process in the identification duration. Particularly, LLM is capable of mimicking the reasoning steps undertaken by humans. As illustrated in Figure 2(a), if the LLM get the object information, it can extract the objects relevant to the query sentence, i.e., targets keyboard and anchors door, and successfully identify the correct target keyboard by calculating its distance with door." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 500, + 526, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 526, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 526, + 643 + ], + "type": "text", + "content": "While LLMs show powerful human-like reasoning capabilities, they still have some limitations. First, it cannot handle the view-dependent issue such as the right window. This is because the 3D scene can freely rotate to different views while it keeps static in 2D images. LLMs usually make decisions by comparing their x-y values of 3D coordinates despite hinting it in the conversation. Second, mathematical calculation is a common weakness of LLMs but is necessary for 3DVG [11]. For example, in Figure 2(a), distance computing is crucial to solve the closest relations, whereas the LLMs cannot always provide accurate results. These two issues stem from LLM's training limitations, which affect the reliability of the dialog with LLM approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 647, + 432, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 432, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 432, + 659 + ], + "type": "text", + "content": "3.2. 3D Visual Programming" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "content": "To address the above two issues, we now introduce a new approach that generates visual programs through LLMs. As" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20625" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 99, + 289, + 167 + ], + "blocks": [ + { + "bbox": [ + 70, + 99, + 289, + 167 + ], + "lines": [ + { + "bbox": [ + 70, + 99, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 289, + 167 + ], + "type": "table", + "html": "
View-independentnear, close, next to, far, above, below, under, top, on, opposite, middle
View-dependentfront, behind, back, right, left, facing, leftmost, rightmost, looking, across, between
Functionalmin, max, size, length, width
", + "image_path": "37f762afbbb0139afc769f0cadc3cc69b07de6e3edff0b15eb53b4f13325f885.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 171, + 242, + 180 + ], + "lines": [ + { + "bbox": [ + 115, + 171, + 242, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 242, + 180 + ], + "type": "text", + "content": "Table 1. Common relations in 3DVG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 194, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 194, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 68, + 194, + 288, + 281 + ], + "type": "text", + "content": "shown in Figure 2(b), we first construct a set of sample programs to encapsulate human-like problem-solving tactics in 3DVG. Each program includes a sequence of operations, where each operation contains a module name, several input parameters, and an assigned output variable. The output of each step can be reused in the subsequent step, thus creating an interlinked sequence that reflects logical reasoning within a visual context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 281, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 281, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 68, + 281, + 289, + 348 + ], + "type": "text", + "content": "We transform the reasoning process of 3DVG into a scripted visual program. Specifically, we collect a set of in-context examples and the corresponding grounding descriptions, and then use LLMs to extrapolate new visual programs tailored to the task. For example, in Figure 2(b), we consider the task prompted by the following description:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 354, + 260, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 354, + 260, + 385 + ], + "spans": [ + { + "bbox": [ + 87, + 354, + 260, + 385 + ], + "type": "text", + "content": "The round cocktail table in the corner of the room with the blue and yellow poster." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "type": "text", + "content": "In this case, the objective is to identify the round cocktail table, which can be transformed into a operation: " + }, + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{BOX0} = \\mathrm{LOC}(" + }, + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "type": "text", + "content": " 'round cocktail table'), where the LOC operator processes the textual query and outputs the bounding boxes for the target objects. We will elaborate the design of LOC module in Section 3.4. Nevertheless, since there may exist multiple similar objects in 3D scenarios, the identified results may not be unique. To overcome this issue, we further pinpoint the blue and yellow poster as an auxiliary reference point by a operation: " + }, + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{BOX1} = \\mathrm{LOC}(" + }, + { + "bbox": [ + 68, + 390, + 288, + 532 + ], + "type": "text", + "content": " 'blue and yellow poster'). Then, the CLOSEST module computes the proximity between BOX0 (potential tables) and BOX1 (poster), and selects the table closest to the poster as the result." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 533, + 288, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 533, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 68, + 533, + 288, + 566 + ], + "type": "text", + "content": "Table 1 summarizes the common relations in 3DVG. Based on this, we present the detailed visual program by developing three types of modules tailored for 3D contexts:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 566, + 288, + 685 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 566, + 288, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 566, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 69, + 566, + 288, + 609 + ], + "type": "text", + "content": "- View-independent modules: They operate on the 3D spatial relations between objects. For example, the CLOS-EST module can discern proximity independent of the viewer's position." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 609, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 609, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 609, + 288, + 653 + ], + "type": "text", + "content": "- View-dependent modules: They depend on the observer's vantage point. For instance, the RIGHT module determines the right window (TARGET) when looking at cabinets (BOX1) from all windows (BOX0)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 654, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 654, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 654, + 288, + 685 + ], + "type": "text", + "content": "- Functional modules: They include multiple operations such as MIN and MAX, which select objects based on the extremal criteria." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 103, + 521, + 205 + ], + "blocks": [ + { + "bbox": [ + 310, + 103, + 521, + 205 + ], + "lines": [ + { + "bbox": [ + 310, + 103, + 521, + 205 + ], + "spans": [ + { + "bbox": [ + 310, + 103, + 521, + 205 + ], + "type": "image", + "image_path": "989c65269447548de2e43e5f6b17cba535202781d22b6c24b26ab650058ef08d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 217, + 525, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 525, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 525, + 237 + ], + "type": "text", + "content": "Figure 3. Addressing view-dependent relations: A shift to 2D ego-centric view." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 257, + 525, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 525, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 525, + 312 + ], + "type": "text", + "content": "These three types of modules allow the output of one operation to be fed into another operation, thus providing flexible composability. They not only facilitate structured and accurate inference sequences, but also integrate 3D and 2D data to yield a robust and interpretable result for 3DVG." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 319, + 491, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 491, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 491, + 330 + ], + "type": "text", + "content": "3.3. Addressing View-Dependent Relations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 335, + 525, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 525, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 525, + 411 + ], + "type": "text", + "content": "In this section, we discuss the intricacies of the view-dependent relations, which are essential for interpreting spatial relations within 3D space. Particularly, the main challenge is the dynamic nature of these relations that will change with the observer's viewpoint. Although traditional supervised approaches can learn these relations implicitly, they cannot provide a definitive resolution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 412, + 525, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 525, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 525, + 477 + ], + "type": "text", + "content": "On 2D planes, the relations, especially the left and right are well defined. More specifically, right often corresponds to the positive direction of the x-axis while the left implies the negative direction. Motivated by this, we adopt a 2D egocentric view approach to ensure a consistent frame of reference for the spatial relations in Table 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 478, + 525, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 525, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 525, + 553 + ], + "type": "text", + "content": "Our view-dependent modules accept a target argument and an optional anchors parameter. They output the target objects that fulfill the spatial relation to the anchors. When grounding queries do not specify targets, we treat targets as anchors as well. This approach aligns with our intuition, such as identifying the left window by treating all windows themselves as the anchors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "text", + "content": "As shown in Figure 3, we assume there is a virtual camera in the center of the room, i.e., " + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{center}}" + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "text", + "content": ", which can rotate to align with the location of the anchor objects, i.e., " + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{o}_a}" + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "text", + "content": ". The 3D objects are projected onto a 2D plane from this vantage point. Assume that the orthogonal camera has an intrinsic matrix " + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 554, + 525, + 620 + ], + "type": "text", + "content": ", then the 2D projections can be obtained by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 628, + 525, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 628, + 525, + 640 + ], + "spans": [ + { + "bbox": [ + 353, + 628, + 525, + 640 + ], + "type": "interline_equation", + "content": "R, T = \\operatorname {L o o k a t} \\left(P _ {\\text {c e n t e r}}, P _ {\\mathrm {o} _ {a}}, u p\\right), \\tag {1}", + "image_path": "748907e5d6411da4667e53541c9d145198e1ff2e603d8ac4aaea4783fe61c451.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 353, + 643, + 525, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 643, + 525, + 655 + ], + "spans": [ + { + "bbox": [ + 353, + 643, + 525, + 655 + ], + "type": "interline_equation", + "content": "(u, v, w) ^ {\\mathrm {T}} = I \\cdot (R | t) \\cdot P, \\tag {2}", + "image_path": "8fbc9a195bd4072f44538f3cf9ce99b7886660514d63f7a3b017f61c28163fdb.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{Lookat}(\\cdot)" + }, + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "text", + "content": " is a view transformation function that computes the rotation matrix " + }, + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "text", + "content": " and translation matrix " + }, + { + "bbox": [ + 306, + 664, + 525, + 686 + ], + "type": "inline_equation", + "content": "T" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20626" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "spans": [ + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": "[50], " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "P = (x, y, z, 1)^{\\mathrm{T}}" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " denotes the 3D coordinate vector, " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " respectively signify the x-axis and y-axis on the 2D plane, and " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " is the depth value. According to the value of " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " of an object's center, we can determine its left or right position — a lower " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " value indicates left. Similarly, " + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 68, + 99, + 290, + 176 + ], + "type": "text", + "content": " allows us to distinguish the front from behind. By synthesizing these concepts, we can define the between relation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 177, + 290, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 177, + 290, + 221 + ], + "spans": [ + { + "bbox": [ + 68, + 177, + 290, + 221 + ], + "type": "text", + "content": "The transition from 3D to 2D egocentric perspective provides a clear and consistent solution to interpret view-dependent relations in 3D space, thus enhancing our model's spatial reasoning ability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 229, + 253, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 229, + 253, + 242 + ], + "spans": [ + { + "bbox": [ + 69, + 229, + 253, + 242 + ], + "type": "text", + "content": "3.4. Language-Object Correlation Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 247, + 290, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 247, + 290, + 466 + ], + "spans": [ + { + "bbox": [ + 68, + 247, + 290, + 466 + ], + "type": "text", + "content": "Although our zero-shot 3DVG approach does not need extensive grounding annotations, it still requires a basic vision model for object localization. To overcome this issue, previous works [4, 62] usually use pre-trained 3D detectors [22, 38] to generate object proposals and the corresponding labels within a fixed vocabulary. However, this approach is restricted to a predefined object class set, thus limiting the scope of class prediction. To enable open-vocabulary segmentation, we develop an LOC module, combining the advantages of 3D and 2D networks to extend the labeling capability beyond the closed set. For example, in Figure 4, considering the operation: " + }, + { + "bbox": [ + 68, + 247, + 290, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{BOX0} = \\mathrm{LOC}(" + }, + { + "bbox": [ + 68, + 247, + 290, + 466 + ], + "type": "text", + "content": " 'round cocktail table'), we first filter a subset of objects whose predicted label is table using a 3D instance segmentation network [43]. Then we only need to identify a round cocktail table from this subset using the corresponding 2D imagery. By mapping each 3D proposal to its 2D image, we can extract the color and texture details pertinent to our query. To further pinpoint the round cocktail table shown in the Figure 4, we consider three types of 2D multi-modal models:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 467, + 290, + 620 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 69, + 467, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 467, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 69, + 467, + 290, + 533 + ], + "type": "text", + "content": "- Image classification models: We construct a dynamic vocabulary, including both the query term \"round cocktail table\" and the class \"table\" using popular tools such as CLIP [41]. Then we evaluate the cosine similarity between these terms and the imagery to find the best correlation to our query." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 532, + 290, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 532, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 69, + 532, + 290, + 576 + ], + "type": "text", + "content": "- Visual question answering models: We raise the question: Is there a [query]? to the model such as ViLT [25]. Then the model sifts through its dictionary to suggest the most likely answer, i.e., yes or no." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 576, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 576, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 69, + 576, + 289, + 620 + ], + "type": "text", + "content": "- General large models: We submit the same inquiry and anticipate a response based on the generated text. This process is crucial for verifying the alignment between the detected table and the query." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 620, + 290, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 620, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 620, + 290, + 687 + ], + "type": "text", + "content": "We shall note that our approach is not limited to specific 3D or 2D models, allowing versatile incorporation of various models. In the experiments, we will demonstrate that the benefit of the LOC modules by comparing with the 3D-only and 2D-only counterparts. Our design indicates a leap forward in 3D open-vocabulary instance segmentation and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 111, + 525, + 235 + ], + "blocks": [ + { + "bbox": [ + 360, + 100, + 476, + 109 + ], + "lines": [ + { + "bbox": [ + 360, + 100, + 476, + 109 + ], + "spans": [ + { + "bbox": [ + 360, + 100, + 476, + 109 + ], + "type": "text", + "content": "BOX0=LOC(object='round cocktail table')" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 111, + 525, + 235 + ], + "lines": [ + { + "bbox": [ + 307, + 111, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 307, + 111, + 525, + 235 + ], + "type": "image", + "image_path": "e7d8fe31afa4235c96087cf3d70416266f850cfd2aa4ae9951261021ac40fdc4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 242, + 522, + 252 + ], + "lines": [ + { + "bbox": [ + 307, + 242, + 522, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 522, + 252 + ], + "type": "text", + "content": "Figure 4. Illustration of the language-object correlation module." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 259, + 508, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 508, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 508, + 270 + ], + "type": "text", + "content": "can improve the object recognition accuracy in 3DVG." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 279, + 380, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 279, + 380, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 279, + 380, + 292 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 297, + 422, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 422, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 422, + 309 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 314, + 526, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 526, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 526, + 456 + ], + "type": "text", + "content": "Datasets. We use two popular datasets, i.e., ScanRefer [4] and Nr3D [1] for experiments. ScanRefer is tailored for 3DVG that contains 51,500 sentence descriptions for 800 ScanNet scenes [9]. Nr3D is a human-written and free-form dataset for 3DVG, collected by 2-player reference game in 3D scenes. The sentences are divided into \"easy\" and \"hard\" subsets, where the target object only contains one same-class distractor in the \"easy\" subset but contains multiple ones in the \"hard\" subset. Depending on whether the sentence requires a specific viewpoint to ground the referred object, the dataset can also be partitioned into \"view dependent\" and \"view independent\" subsets. For both datasets, we evaluate the zero-shot approaches on the validation split." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 456, + 526, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 526, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 526, + 577 + ], + "type": "text", + "content": "Evaluation metrics. We consider two settings for performance evaluation. The first one mandates the generation of object proposals, aligning closely with real-world applications. The evaluation metrics are Acc@0.25 and Acc@0.5, representing the percentage of correctly predicted bounding boxes whose IoU exceeds 0.25 or 0.5 with the ground-truth, respectively. This is the default setting for ScanRefer dataset. The second one furnishes ground-truth object masks, necessitating only classification, with an objective to eradicate localization error and achieve high grounding accuracy. This is the default setting for Nr3D dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 577, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 526, + 687 + ], + "type": "text", + "content": "Baselines. We use six supervised and two open-vocabulary 3D scene understanding approaches for performance comparison. For supervised approaches, ScanRefer [4] and ReferIt3DNet [1] encode the 3D point clouds and language separately, and then fuse them to rank the objects by predicted scores. TGNN [18] and InstanceRefer [62] make one further step by learning instance-wise features. 3DVG-Transformer [68] and BUTD-DETR [21] respectively utilize the Transformer [48] and DETR [3] architectures, representing the SoTA approaches. For open-vocabulary ap" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20627" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 98, + 515, + 235 + ], + "blocks": [ + { + "bbox": [ + 77, + 98, + 515, + 235 + ], + "lines": [ + { + "bbox": [ + 77, + 98, + 515, + 235 + ], + "spans": [ + { + "bbox": [ + 77, + 98, + 515, + 235 + ], + "type": "table", + "html": "
MethodsSupervisionUniqueMultipleOverall
Acc@0.25Acc@0.5Acc@0.25Acc@0.5Acc@0.25Acc@0.5
ScanRefer [4]fully65.043.330.619.837.324.3
TGNN [18]fully64.553.027.021.934.329.7
InstanceRefer [62]fully77.566.831.324.840.232.9
3DVG-Transformer [68]fully81.960.639.328.447.634.7
BUTD-DETR [21]fully84.266.346.635.152.239.8
LERF [24]-----4.80.9
OpenScene [35]-20.113.111.14.413.26.5
Ours (2D only)-32.527.816.114.620.017.6
Ours (3D only)-57.149.425.923.333.129.3
Ours-63.858.427.724.636.432.7
", + "image_path": "0dbea280a87e6f3dcb60e728d6239ab45d45cda77e4e23c0c4ed491aa2a8e531.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 270, + 286, + 349 + ], + "blocks": [ + { + "bbox": [ + 69, + 239, + 525, + 262 + ], + "lines": [ + { + "bbox": [ + 69, + 239, + 525, + 262 + ], + "spans": [ + { + "bbox": [ + 69, + 239, + 525, + 262 + ], + "type": "text", + "content": "Table 2. 3DVG results on ScanRefer validation set. The accuracy on the \"unique\" subset, \"multiple\" subset, and whole validation set are all provided. Following [4], we label the scene as \"unique\" if it only contains a single object of its class. Otherwise, we label it as \"multiple\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 270, + 286, + 349 + ], + "lines": [ + { + "bbox": [ + 72, + 270, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 72, + 270, + 286, + 349 + ], + "type": "table", + "html": "
MethodEasyHardDep.Indep.Overall
ReferIt3DNet [1]43.627.932.537.135.6
InstanceRefer [62]46.031.834.541.938.8
3DVG-Transformer [68]48.534.834.843.740.8
BUTD-DETR [21]60.748.446.058.054.6
Ours (2D only)29.418.423.023.923.6
Ours (3D only)45.927.934.938.436.7
Ours46.531.736.840.039.0
", + "image_path": "0e0518423041f85b502b990cb67f8bf06f80509c45ad34d47003f90365cd3ed0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 354, + 288, + 375 + ], + "lines": [ + { + "bbox": [ + 69, + 354, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 69, + 354, + 288, + 375 + ], + "type": "text", + "content": "Table 3. Performance analysis of language grounding on Nr3D. We evaluate the top-1 accuracy using ground-truth boxes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 384, + 290, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 384, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 68, + 384, + 290, + 462 + ], + "type": "text", + "content": "proaches, OpenScene [35] and LERF [24] aims to learn a 3D representation aligned with the 2D CLIP feature, thus enabling free-form language grounding. The query " + }, + { + "bbox": [ + 68, + 384, + 290, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 68, + 384, + 290, + 462 + ], + "type": "text", + "content": " is processed by the CLIP text encoder, and its similarity is computed against the extracted point features. Finally, they cluster the points with the highest score to determine the target object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 471, + 178, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 471, + 178, + 482 + ], + "spans": [ + { + "bbox": [ + 69, + 471, + 178, + 482 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 488, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 488, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 68, + 488, + 288, + 652 + ], + "type": "text", + "content": "ScanRefer. Table 2 provides a quantitative assessment of the proposed approach on the ScanRefer dataset. We can see that our zero-shot approach outperforms all baseline approaches. Specifically, our approach can achieve a 32.7 Acc@0.5 score, which surpasses part of supervised approaches such as ScanRefer and TGNN. On the other hand, the open-vocabulary approaches LERF and OpenScene can respectively achieve the overall accuracy of 4.8 and 13.2, even with the 0.25 IoU threshold. This is due to their limitations in reasoning and localization precision. Moreover, our zero-shot approach outperforms the approaches that only utilize the 3D or 2D information in the LOC module. This result demonstrates the effectiveness of incorporating visual programming and perception modules, highlighting our zero-shot approach in navigating 3DVG." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 653, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 653, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 653, + 288, + 686 + ], + "type": "text", + "content": "Nr3D. Table 3 shows the performance of different approaches on the Nr3D dataset, in which the ground-truth instance mask is also provided. We can see that our zero" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 273, + 525, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 525, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 525, + 339 + ], + "type": "text", + "content": "shot approach further excels the supervised approach InstanceRefer. Specifically, our zero-shot approach on the \"view-dependent\" split can achieve a " + }, + { + "bbox": [ + 304, + 273, + 525, + 339 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 273, + 525, + 339 + ], + "type": "text", + "content": " accuracy gain than the 3DVG-Transformer approach. This performance gain comes from the relation modules, strengthening the potential of our zero-shot approach for 3DVG tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 346, + 408, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 408, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 408, + 358 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 364, + 525, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 525, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 525, + 538 + ], + "type": "text", + "content": "Figure 5 shows the visualizations of the selected samples from the ScanRefer validation set. The four columns present the ground-truth result, the supervised approach BUTD-DETR, the dialog with LLM, and the visual programming approaches, respectively. From Figure 5(a) and Figure 5(b), we can observe that the dialog with LLM and the visual programming approaches can achieve accurate prediction results for view-independent relations, i.e., (above, under) without much training. On the contrary, both the BUTD-DETR and the dialog with LLM approaches cannot address the view-dependent relations, i.e., (left, front), as shown in Figure 5(c) and Figure 5(d). The inherent uncertainty of these relations reflects the limitations of existing methods. However, our visual programming approach can leverage the 2D egocentric views, thus achieving accurate predictions in 3D scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 540, + 525, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 525, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 525, + 628 + ], + "type": "text", + "content": "Figure 5(e) presents a failure case, where the dialog with LLM approach cannot recognize chair has wheels since it lacks open-vocabulary detection ability. Besides, the visual programming approach makes wrong predictions because the LLM cannot correctly recognize the relation pushed. Fortunately, when we correct the program using the CLOsE ST module, the visual programming approach can make correct predictions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 525, + 686 + ], + "type": "text", + "content": "Dialog with LLM vs. visual programming. We compare the performance of the two proposed zero-shot 3DVG approaches on the ScanRefer validation set with 700 exam" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20628" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 99, + 191, + 150 + ], + "blocks": [ + { + "bbox": [ + 93, + 116, + 118, + 131 + ], + "lines": [ + { + "bbox": [ + 93, + 116, + 118, + 131 + ], + "spans": [ + { + "bbox": [ + 93, + 116, + 118, + 131 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 125, + 99, + 191, + 150 + ], + "lines": [ + { + "bbox": [ + 125, + 99, + 191, + 150 + ], + "spans": [ + { + "bbox": [ + 125, + 99, + 191, + 150 + ], + "type": "image", + "image_path": "cbabafc01d0e613c1a286cf7d153edc1ad7a2af009162624d08d3c7a96d26618.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 199, + 99, + 266, + 149 + ], + "blocks": [ + { + "bbox": [ + 199, + 99, + 266, + 149 + ], + "lines": [ + { + "bbox": [ + 199, + 99, + 266, + 149 + ], + "spans": [ + { + "bbox": [ + 199, + 99, + 266, + 149 + ], + "type": "image", + "image_path": "663372fcc575527cdadb0a2616789a4b8ca5cbdae45ad1074d62a5cf02604efc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 276, + 99, + 342, + 149 + ], + "blocks": [ + { + "bbox": [ + 276, + 99, + 342, + 149 + ], + "lines": [ + { + "bbox": [ + 276, + 99, + 342, + 149 + ], + "spans": [ + { + "bbox": [ + 276, + 99, + 342, + 149 + ], + "type": "image", + "image_path": "f8e529c60bea8d85d165407542a0047f975da91c29aba43ace5a447ce1b28fc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 99, + 417, + 149 + ], + "blocks": [ + { + "bbox": [ + 350, + 99, + 417, + 149 + ], + "lines": [ + { + "bbox": [ + 350, + 99, + 417, + 149 + ], + "spans": [ + { + "bbox": [ + 350, + 99, + 417, + 149 + ], + "type": "image", + "image_path": "059b07e9a2dc57e0542ca06a77b4c86d8c8f8f8b9de7b6d0ef6d14432d90fcea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 99, + 493, + 149 + ], + "blocks": [ + { + "bbox": [ + 425, + 99, + 493, + 149 + ], + "lines": [ + { + "bbox": [ + 425, + 99, + 493, + 149 + ], + "spans": [ + { + "bbox": [ + 425, + 99, + 493, + 149 + ], + "type": "image", + "image_path": "5d21a3ffea9c8efd25c257ede92c1f2b99caf3d58f9407ac293967409beb255b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 126, + 152, + 190, + 202 + ], + "blocks": [ + { + "bbox": [ + 96, + 171, + 116, + 185 + ], + "lines": [ + { + "bbox": [ + 96, + 171, + 116, + 185 + ], + "spans": [ + { + "bbox": [ + 96, + 171, + 116, + 185 + ], + "type": "text", + "content": "BUTD-DETR" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 152, + 190, + 202 + ], + "lines": [ + { + "bbox": [ + 126, + 152, + 190, + 202 + ], + "spans": [ + { + "bbox": [ + 126, + 152, + 190, + 202 + ], + "type": "image", + "image_path": "d0e273c6ef725411e30020c0a5ce9ad515b75d19d5da72e53068830a66f9249e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 199, + 152, + 266, + 202 + ], + "blocks": [ + { + "bbox": [ + 199, + 152, + 266, + 202 + ], + "lines": [ + { + "bbox": [ + 199, + 152, + 266, + 202 + ], + "spans": [ + { + "bbox": [ + 199, + 152, + 266, + 202 + ], + "type": "image", + "image_path": "71e666e749c290cacd6753be38393b19dd2cdde973281576979007d677149f72.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 276, + 152, + 342, + 202 + ], + "blocks": [ + { + "bbox": [ + 276, + 152, + 342, + 202 + ], + "lines": [ + { + "bbox": [ + 276, + 152, + 342, + 202 + ], + "spans": [ + { + "bbox": [ + 276, + 152, + 342, + 202 + ], + "type": "image", + "image_path": "67be939add2aa8f86ceba45bb88f0d502bf9c0ae31f9840363187c7a696751d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 350, + 152, + 417, + 202 + ], + "blocks": [ + { + "bbox": [ + 350, + 152, + 417, + 202 + ], + "lines": [ + { + "bbox": [ + 350, + 152, + 417, + 202 + ], + "spans": [ + { + "bbox": [ + 350, + 152, + 417, + 202 + ], + "type": "image", + "image_path": "3d2b28f6e8797a5a8c6a3f868f92fd6c9907fe89e93fd46df04b7ea59598d555.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 152, + 492, + 202 + ], + "blocks": [ + { + "bbox": [ + 425, + 152, + 492, + 202 + ], + "lines": [ + { + "bbox": [ + 425, + 152, + 492, + 202 + ], + "spans": [ + { + "bbox": [ + 425, + 152, + 492, + 202 + ], + "type": "image", + "image_path": "b279d8359f8d86eb7a1eac1662b9c0aa53dd25941d07bf7a38e130c1a10807f6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 125, + 207, + 190, + 257 + ], + "blocks": [ + { + "bbox": [ + 94, + 224, + 116, + 240 + ], + "lines": [ + { + "bbox": [ + 94, + 224, + 116, + 240 + ], + "spans": [ + { + "bbox": [ + 94, + 224, + 116, + 240 + ], + "type": "text", + "content": "Ours (Dialog)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 125, + 207, + 190, + 257 + ], + "lines": [ + { + "bbox": [ + 125, + 207, + 190, + 257 + ], + "spans": [ + { + "bbox": [ + 125, + 207, + 190, + 257 + ], + "type": "image", + "image_path": "234491251d07a6675289c89e06ebe712f07a0213505e76b41491407aa0936fa9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 199, + 207, + 266, + 257 + ], + "blocks": [ + { + "bbox": [ + 199, + 207, + 266, + 257 + ], + "lines": [ + { + "bbox": [ + 199, + 207, + 266, + 257 + ], + "spans": [ + { + "bbox": [ + 199, + 207, + 266, + 257 + ], + "type": "image", + "image_path": "f81117909218875e75e2de5e8b92bafe7139e7f8511ceba47ecf1e24d2520a26.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 276, + 207, + 342, + 256 + ], + "blocks": [ + { + "bbox": [ + 276, + 207, + 342, + 256 + ], + "lines": [ + { + "bbox": [ + 276, + 207, + 342, + 256 + ], + "spans": [ + { + "bbox": [ + 276, + 207, + 342, + 256 + ], + "type": "image", + "image_path": "5f21d468f1678d9a1380ef613f8a65035d2704c51ad6594cdcf8b7536e86c9de.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 350, + 207, + 417, + 257 + ], + "blocks": [ + { + "bbox": [ + 350, + 207, + 417, + 257 + ], + "lines": [ + { + "bbox": [ + 350, + 207, + 417, + 257 + ], + "spans": [ + { + "bbox": [ + 350, + 207, + 417, + 257 + ], + "type": "image", + "image_path": "c0a94a61edc601d5004913c5e5be966429f345b0bb7643d4010c8b3795970d1e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 425, + 207, + 492, + 257 + ], + "blocks": [ + { + "bbox": [ + 425, + 207, + 492, + 257 + ], + "lines": [ + { + "bbox": [ + 425, + 207, + 492, + 257 + ], + "spans": [ + { + "bbox": [ + 425, + 207, + 492, + 257 + ], + "type": "image", + "image_path": "8c7e229a58b25ae8df5f6bbb5709351a411c856b14d0a89c5e35fc7e0ebfd837.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 125, + 262, + 190, + 311 + ], + "blocks": [ + { + "bbox": [ + 93, + 281, + 120, + 298 + ], + "lines": [ + { + "bbox": [ + 93, + 281, + 120, + 298 + ], + "spans": [ + { + "bbox": [ + 93, + 281, + 120, + 298 + ], + "type": "text", + "content": "Ours (Program)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 125, + 262, + 190, + 311 + ], + "lines": [ + { + "bbox": [ + 125, + 262, + 190, + 311 + ], + "spans": [ + { + "bbox": [ + 125, + 262, + 190, + 311 + ], + "type": "image", + "image_path": "bc34a793a88f176b9b7725922d0d908e7cb94c9b2ff723b50d444e03ac728929.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 318, + 192, + 342 + ], + "lines": [ + { + "bbox": [ + 128, + 318, + 192, + 342 + ], + "spans": [ + { + "bbox": [ + 128, + 318, + 192, + 342 + ], + "type": "text", + "content": "It is a window. It is located above a recycle bin that has a blue top." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 345, + 167, + 354 + ], + "lines": [ + { + "bbox": [ + 154, + 345, + 167, + 354 + ], + "spans": [ + { + "bbox": [ + 154, + 345, + 167, + 354 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 68, + 356, + 525, + 377 + ], + "lines": [ + { + "bbox": [ + 68, + 356, + 525, + 377 + ], + "spans": [ + { + "bbox": [ + 68, + 356, + 525, + 377 + ], + "type": "text", + "content": "Figure 5. Visualization results of 3D visual grounding. Rendered images of 3D scans are presented, including the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green)." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 199, + 262, + 266, + 311 + ], + "blocks": [ + { + "bbox": [ + 199, + 262, + 266, + 311 + ], + "lines": [ + { + "bbox": [ + 199, + 262, + 266, + 311 + ], + "spans": [ + { + "bbox": [ + 199, + 262, + 266, + 311 + ], + "type": "image", + "image_path": "3a38076226530917f1a5fd4642de42e54682f8823abf7bdb5673c03a7d581e82.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 318, + 263, + 341 + ], + "lines": [ + { + "bbox": [ + 201, + 318, + 263, + 341 + ], + "spans": [ + { + "bbox": [ + 201, + 318, + 263, + 341 + ], + "type": "text", + "content": "The rolling office chair. The chair is under the desk." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 345, + 238, + 354 + ], + "lines": [ + { + "bbox": [ + 227, + 345, + 238, + 354 + ], + "spans": [ + { + "bbox": [ + 227, + 345, + 238, + 354 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 276, + 262, + 342, + 311 + ], + "blocks": [ + { + "bbox": [ + 276, + 262, + 342, + 311 + ], + "lines": [ + { + "bbox": [ + 276, + 262, + 342, + 311 + ], + "spans": [ + { + "bbox": [ + 276, + 262, + 342, + 311 + ], + "type": "image", + "image_path": "9d255a7bd054511bc317938843d326d757644262ef78a799c2139607988fc498.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 317, + 339, + 341 + ], + "lines": [ + { + "bbox": [ + 277, + 317, + 339, + 341 + ], + "spans": [ + { + "bbox": [ + 277, + 317, + 339, + 341 + ], + "type": "text", + "content": "There is a square beige armchair. It is left of a square table." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 345, + 310, + 354 + ], + "lines": [ + { + "bbox": [ + 299, + 345, + 310, + 354 + ], + "spans": [ + { + "bbox": [ + 299, + 345, + 310, + 354 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 350, + 262, + 417, + 311 + ], + "blocks": [ + { + "bbox": [ + 350, + 262, + 417, + 311 + ], + "lines": [ + { + "bbox": [ + 350, + 262, + 417, + 311 + ], + "spans": [ + { + "bbox": [ + 350, + 262, + 417, + 311 + ], + "type": "image", + "image_path": "a13609d7839721aeeb3f30421bc00cc4eb7cbff21e2165025c56ddf99fa60dd3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 318, + 408, + 342 + ], + "lines": [ + { + "bbox": [ + 356, + 318, + 408, + 342 + ], + "spans": [ + { + "bbox": [ + 356, + 318, + 408, + 342 + ], + "type": "text", + "content": "This is a brown piano bench. It is in front of the piano." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 345, + 389, + 354 + ], + "lines": [ + { + "bbox": [ + 378, + 345, + 389, + 354 + ], + "spans": [ + { + "bbox": [ + 378, + 345, + 389, + 354 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 425, + 262, + 492, + 311 + ], + "blocks": [ + { + "bbox": [ + 425, + 262, + 492, + 311 + ], + "lines": [ + { + "bbox": [ + 425, + 262, + 492, + 311 + ], + "spans": [ + { + "bbox": [ + 425, + 262, + 492, + 311 + ], + "type": "image", + "image_path": "a0bc77b8506f99caf5444fc935798c34e269e2efb8cb2a398a7faa214b3a3368.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 318, + 498, + 341 + ], + "lines": [ + { + "bbox": [ + 425, + 318, + 498, + 341 + ], + "spans": [ + { + "bbox": [ + 425, + 318, + 498, + 341 + ], + "type": "text", + "content": "A desk chair is pushed into a small computer desk. The chair has wheels." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 345, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 457, + 345, + 468, + 354 + ], + "spans": [ + { + "bbox": [ + 457, + 345, + 468, + 354 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 382, + 289, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 382, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 67, + 382, + 289, + 514 + ], + "type": "text", + "content": "plies. For both approaches, we use two GPT versions, i.e., GPT-3.5-turbo-0613 and GPT-4-0613. The cost of each GPT version depends on the number of input and output tokens. The experimental results are shown in Table 4. We can observe that for both zero-shot approaches, GPT4-based approach can achieve higher accuracy than GPT3.5-based approach, even it induces a larger economic cost. On the other hand, the visual programming approach always outperforms the dialog with LLM approach in terms of accuracy and cost, which demonstrates the effectiveness of the proposed visual programming approach. For other experiments, we use GPT3.5 to save cost." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 517, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 517, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 68, + 517, + 289, + 606 + ], + "type": "text", + "content": "Relation modules. We now ablate different relation modules in Section 3.2 to analyze their impact on the system performance. The most important view-dependent and view-independent modules are presented in Table 5 and 6, respectively. We can see that LEFT and RIGHT are the most important view-dependent relations, while CLOSEST is the most important view-independent relation. This result is coherent with our motivation and design." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 609, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 609, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 609, + 289, + 687 + ], + "type": "text", + "content": "LOC module. We juxtapose our approach by separately omitting the 3D component and 2D component. Both models utilize the instance mask prediction of Mask3D [43]. Particularly, the 2D-only model solely employs the paired 2D images for classification, while the 3D-only model just uses the 3D result. As can be seen from Tables 2 and 3, the 2D-only model performs worst when the images of in" + } + ] + } + ], + "index": 37 + }, + { + "type": "table", + "bbox": [ + 323, + 380, + 509, + 439 + ], + "blocks": [ + { + "bbox": [ + 323, + 380, + 509, + 439 + ], + "lines": [ + { + "bbox": [ + 323, + 380, + 509, + 439 + ], + "spans": [ + { + "bbox": [ + 323, + 380, + 509, + 439 + ], + "type": "table", + "html": "
MethodLLMAcc@0.5TokensCost
DialogGPT3.525.41959k$3.05
DialogGPT427.51916k$62.6
ProgramGPT3.532.1121k$0.19
ProgramGPT435.4115k$4.24
", + "image_path": "f674e0fac8a930d3bc1e1e6d4ea3fc620fcc0d682a27d920456214095f76430c.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 314, + 475, + 515, + 552 + ], + "blocks": [ + { + "bbox": [ + 305, + 446, + 525, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 446, + 525, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 525, + 468 + ], + "type": "text", + "content": "Table 4. Performance comparison of the dialog with LLM and the visual programming approaches." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 475, + 515, + 552 + ], + "lines": [ + { + "bbox": [ + 314, + 475, + 515, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 515, + 552 + ], + "type": "table", + "html": "
LEFTRIGHTFRONTBEHINDBETWEENAccuracy
26.5
✓32.4
✓✓35.9
✓✓✓36.8
✓✓✓✓38.4
✓✓✓✓✓39.0
", + "image_path": "35e4a3352510d4fb8e4af113a04158c1de96d2429c3fe12ed25a222eee83b2cf.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 310, + 555, + 518, + 565 + ], + "lines": [ + { + "bbox": [ + 310, + 555, + 518, + 565 + ], + "spans": [ + { + "bbox": [ + 310, + 555, + 518, + 565 + ], + "type": "text", + "content": "Table 5. Ablation study of different view-dependent modules." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 574, + 525, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 525, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 525, + 651 + ], + "type": "text", + "content": "door scenes are complicated and have domain gaps with the training samples. The 3D-only model performs better since it can utilize the geometric information and is trained on closed-set labels. Our full model can always achieve the best performance because it integrates the geometric distinctiveness of point clouds and the open-vocabulary ability of the image models." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "text", + "content": "Generalization. As discussed in Section 3.4, our framework has strong adaptability for a spectrum of 3D and 2D perception models. To validate this claim, we conduct ex" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "20629" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 97, + 281, + 164 + ], + "blocks": [ + { + "bbox": [ + 75, + 97, + 281, + 164 + ], + "lines": [ + { + "bbox": [ + 75, + 97, + 281, + 164 + ], + "spans": [ + { + "bbox": [ + 75, + 97, + 281, + 164 + ], + "type": "table", + "html": "
CLOSESTFARTHESTLOWERHIGHERAccuracy
18.8
✓30.7
✓✓34.0
✓✓✓36.8
✓✓✓✓39.0
", + "image_path": "0d15a31d4f1aa3e0eb134efdd93698857caac1625d00fd5137b99784cf767c6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 186, + 289, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 186, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 68, + 186, + 289, + 327 + ], + "type": "text", + "content": "periments using several representative models. For 3D perception, we utilize three backbones, i.e., PointNet++ [37], PointNeXt [39], and PointBERT [61]. For 2D perception, we use an image classification model proposed in [41], a visual question answering model in [25], and a general large model BLIP-2 [28] for testing. The results are shown in Tables 7 and 8. We can observe that our framework is compatible with other models. Also, it can leverage the advancements within both 2D and 3D foundational models to improve the performance. This cross-model effectiveness demonstrates the robustness and future-proof nature of our approach in the ever-evolving landscape of visual perception models." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 87, + 338, + 269, + 383 + ], + "blocks": [ + { + "bbox": [ + 70, + 167, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 70, + 167, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 70, + 167, + 286, + 178 + ], + "type": "text", + "content": "Table 6. Ablation study of different view-independent modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 338, + 269, + 383 + ], + "lines": [ + { + "bbox": [ + 87, + 338, + 269, + 383 + ], + "spans": [ + { + "bbox": [ + 87, + 338, + 269, + 383 + ], + "type": "table", + "html": "
2D AssistanceUniqueMultipleAcc@0.25
CLIP62.527.135.7
ViLT60.327.135.1
BLIP-263.827.736.4
", + "image_path": "ae3a2a8b4633821fb35e43a33a0c49ed40dc0eed662f0a9d25835bf87d10e46b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 83, + 408, + 272, + 454 + ], + "blocks": [ + { + "bbox": [ + 98, + 388, + 259, + 398 + ], + "lines": [ + { + "bbox": [ + 98, + 388, + 259, + 398 + ], + "spans": [ + { + "bbox": [ + 98, + 388, + 259, + 398 + ], + "type": "text", + "content": "Table 7. Ablation study on different 2D models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 408, + 272, + 454 + ], + "lines": [ + { + "bbox": [ + 83, + 408, + 272, + 454 + ], + "spans": [ + { + "bbox": [ + 83, + 408, + 272, + 454 + ], + "type": "table", + "html": "
3D BackboneView-dep.View-indep.Overall
PointNet++35.839.438.2
PointBert36.039.838.6
PointNeXt36.840.039.0
", + "image_path": "31a73e8d6de3fd3a84cc9a5d49faca58d7d0b73f23b6b42a678f9ea74817f2bf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 456, + 264, + 467 + ], + "lines": [ + { + "bbox": [ + 93, + 456, + 264, + 467 + ], + "spans": [ + { + "bbox": [ + 93, + 456, + 264, + 467 + ], + "type": "text", + "content": "Table 8. Ablation study on different 3D backbones." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 476, + 288, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 476, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 68, + 476, + 288, + 586 + ], + "type": "text", + "content": "Effect of prompt size. We use different numbers of incontext examples in the prompt for program generation. The result is shown in Figure 6. It can be seen that the performance on ScanRefer and Nr3d improve with the number of examples. This is because more examples can guide LLMs to handle more cases in the visual program generation process. Meanwhile, it still follows the law of diminishing marginal utility. Moreover, we test the voting technique [15] to aggregate the results from multiple runs, which brings some performance gains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 587, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 587, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 587, + 289, + 686 + ], + "type": "text", + "content": "Error analysis. We select a representative subset for each dataset with around 100 samples and manually check the rationales offered by the visual program. As illustrated in Figure 7, the generation of accurate visual programs is the primary error source. Therefore, the performance can be improved by using more in-context examples and more powerful LLMs. The second error source is the object localization and classification, indicating that object detection and classification in 3D space remains a critical component." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 99, + 417, + 178 + ], + "blocks": [ + { + "bbox": [ + 317, + 99, + 417, + 178 + ], + "lines": [ + { + "bbox": [ + 317, + 99, + 417, + 178 + ], + "spans": [ + { + "bbox": [ + 317, + 99, + 417, + 178 + ], + "type": "image", + "image_path": "81a1d1b34240c8702a3a2779df88df2b625304a1a4a601ae1fcfb421705fa0db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 181, + 525, + 211 + ], + "lines": [ + { + "bbox": [ + 304, + 181, + 525, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 525, + 211 + ], + "type": "text", + "content": "Figure 6. Ablation study on the number of in-context examples. The performance on Nr3D and ScanRefer improves with the number of in-context examples." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 418, + 99, + 512, + 178 + ], + "blocks": [ + { + "bbox": [ + 418, + 99, + 512, + 178 + ], + "lines": [ + { + "bbox": [ + 418, + 99, + 512, + 178 + ], + "spans": [ + { + "bbox": [ + 418, + 99, + 512, + 178 + ], + "type": "image", + "image_path": "72f79338576ca87efd0ee73cbc9f8c0035c7951e1923b36375360755cfa68e3d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 217, + 511, + 299 + ], + "blocks": [ + { + "bbox": [ + 319, + 217, + 511, + 299 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 511, + 299 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 511, + 299 + ], + "type": "image", + "image_path": "78d7351c3a5f5cfd39dd44ab859742cfc02ef8e61dbf123accc98330310c64f4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 309, + 479, + 319 + ], + "lines": [ + { + "bbox": [ + 351, + 309, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 351, + 309, + 479, + 319 + ], + "type": "text", + "content": "Figure 7. Breakdown of error sources." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 329, + 372, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 372, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 372, + 340 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 347, + 526, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 526, + 468 + ], + "type": "text", + "content": "In this paper, we present a novel zero-shot approach for 3DVG to eliminate the need for extensive annotations and predefined vocabularies. A vanilla dialog with LLM approach is first proposed by taking interactive dialog with LLMs. A visual programming approach is further developed, which leverages three types of modules to navigate the intricate 3D relations. To adapt to open-vocabulary scenarios, we also develop a LOC module to seamlessly integrate both 3D and 2D features. Experimental results demonstrate the superiority of the proposed approach and highlight its potential to advance the field of 3DVG." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 477, + 406, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 406, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 406, + 490 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 526, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 526, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 526, + 686 + ], + "type": "text", + "content": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the National Key R&D Program of China with grant No. 2018YFB1800800, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and the Key Area R&D Program of Guangdong Province with grant No. 2018B03033800, by Tencent & Huawei Open Fund." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "20630" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 117, + 289, + 686 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 74, + 117, + 289, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 117, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 74, + 117, + 289, + 167 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 168, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 168, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 74, + 168, + 288, + 219 + ], + "type": "text", + "content": "[2] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In Conference on Robot Learning, pages 287-318. PMLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 220, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 220, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 74, + 220, + 288, + 261 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 262, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 262, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 74, + 262, + 288, + 302 + ], + "type": "text", + "content": "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 303, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 303, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 74, + 303, + 288, + 353 + ], + "type": "text", + "content": "[5] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 354, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 354, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 74, + 354, + 288, + 404 + ], + "type": "text", + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 406, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 406, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 74, + 406, + 288, + 456 + ], + "type": "text", + "content": "[7] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 458, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 458, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 74, + 458, + 288, + 488 + ], + "type": "text", + "content": "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 490, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 490, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 74, + 490, + 288, + 529 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 531, + 288, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 531, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 70, + 531, + 288, + 571 + ], + "type": "text", + "content": "[10] John David N Dionisio, William G Burns Iii, and Richard Gilbert. 3d virtual worlds and the metaverse: Current status and future possibilities. ACM Computing Surveys (CSUR), 45(3):1-38, 2013. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 572, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 572, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 70, + 572, + 288, + 622 + ], + "type": "text", + "content": "[11] Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D Hwang, et al. Faith and fate: Limits of transformers on compositionality. arXiv preprint arXiv:2305.18654, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 624, + 288, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 624, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 70, + 624, + 288, + 664 + ], + "type": "text", + "content": "[12] Qi Feng, Vitaly Ablavsky, and Stan Sclaroff. Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint arXiv:2101.04741, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 666, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 666, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 666, + 288, + 686 + ], + "type": "text", + "content": "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 524, + 686 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 324, + 101, + 524, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 101, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 324, + 101, + 524, + 121 + ], + "type": "text", + "content": "labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 122, + 524, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 122, + 524, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 122, + 524, + 172 + ], + "type": "text", + "content": "[14] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15372-15383, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 524, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 524, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 524, + 222 + ], + "type": "text", + "content": "[15] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 224, + 524, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 524, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 524, + 274 + ], + "type": "text", + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 275, + 524, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 524, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 524, + 316 + ], + "type": "text", + "content": "[17] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2614-2623, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 317, + 524, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 524, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 524, + 357 + ], + "type": "text", + "content": "[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. 35th AAAI Conference on Artificial Intelligence, 2021. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 358, + 524, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 524, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 524, + 398 + ], + "type": "text", + "content": "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 524, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 524, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 524, + 449 + ], + "type": "text", + "content": "[20] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 451, + 524, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 451, + 524, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 451, + 524, + 511 + ], + "type": "text", + "content": "[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 417-433. Springer, 2022. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 524, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 524, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 524, + 563 + ], + "type": "text", + "content": "[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4867-4876, 2020. 1, 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 563, + 524, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 524, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 524, + 613 + ], + "type": "text", + "content": "[23] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 524, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 524, + 664 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 524, + 664 + ], + "type": "text", + "content": "[24] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "type": "text", + "content": "[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region su" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 706, + 315, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 706, + 315, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 706, + 315, + 714 + ], + "type": "text", + "content": "20631" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 71, + 101, + 288, + 685 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 88, + 101, + 288, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 101, + 288, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 101, + 288, + 121 + ], + "type": "text", + "content": "pervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 122, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 122, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 71, + 122, + 288, + 160 + ], + "type": "text", + "content": "[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 162, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 162, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 71, + 162, + 288, + 201 + ], + "type": "text", + "content": "[27] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 202, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 202, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 71, + 202, + 288, + 241 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 243, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 243, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 71, + 243, + 288, + 300 + ], + "type": "text", + "content": "[29] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 303, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 303, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 71, + 303, + 288, + 352 + ], + "type": "text", + "content": "[30] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 354, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 354, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 71, + 354, + 288, + 392 + ], + "type": "text", + "content": "[31] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 394, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 394, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 71, + 394, + 288, + 441 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 444, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 444, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 71, + 444, + 288, + 463 + ], + "type": "text", + "content": "[33] Stylianos Mystakidis. Metaverse. Encyclopedia, 2(1):486-497, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 464, + 253, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 464, + 253, + 474 + ], + "spans": [ + { + "bbox": [ + 71, + 464, + 253, + 474 + ], + "type": "text", + "content": "[34] OpenAI OpenAI. Gpt-4 technical report. 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 475, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 475, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 71, + 475, + 288, + 523 + ], + "type": "text", + "content": "[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 71, + 525, + 288, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 525, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 71, + 525, + 288, + 573 + ], + "type": "text", + "content": "[36] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 575, + 288, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 575, + 288, + 615 + ], + "spans": [ + { + "bbox": [ + 71, + 575, + 288, + 615 + ], + "type": "text", + "content": "[37] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 616, + 288, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 616, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 71, + 616, + 288, + 664 + ], + "type": "text", + "content": "[38] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 666, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 666, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 71, + 666, + 288, + 685 + ], + "type": "text", + "content": "[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 524, + 685 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 325, + 101, + 524, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 101, + 524, + 130 + ], + "spans": [ + { + "bbox": [ + 325, + 101, + 524, + 130 + ], + "type": "text", + "content": "Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 132, + 524, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 524, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 524, + 180 + ], + "type": "text", + "content": "[40] Xing-yue Qiu, Chuang-Kai Chiu, Lu-Lu Zhao, Cai-Feng Sun, and Shu-jie Chen. Trends in vr/ar technology-supporting language learning from 2008 to 2019: A research perspective. Interactive Learning Environments, 31 (4):2090-2113, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 182, + 524, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 182, + 524, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 524, + 241 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 524, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 524, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 524, + 280 + ], + "type": "text", + "content": "[42] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In 5th Annual Conference on Robot Learning, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 282, + 524, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 524, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 524, + 332 + ], + "type": "text", + "content": "[43] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 334, + 524, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 334, + 524, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 334, + 524, + 372 + ], + "type": "text", + "content": "[44] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 373, + 524, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 524, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 524, + 403 + ], + "type": "text", + "content": "[45] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 524, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 524, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 524, + 443 + ], + "type": "text", + "content": "[46] Ayça Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 524, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 524, + 494 + ], + "type": "text", + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 495, + 524, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 495, + 524, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 495, + 524, + 533 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 536, + 524, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 524, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 524, + 564 + ], + "type": "text", + "content": "[49] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2:20, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 566, + 524, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 524, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 524, + 585 + ], + "type": "text", + "content": "[50] John Vince and John A Vince. Mathematics for computer graphics. Springer, 2006. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 586, + 524, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 524, + 634 + ], + "type": "text", + "content": "[51] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 636, + 524, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 524, + 685 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 524, + 685 + ], + "type": "text", + "content": "[52] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance relocalization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7658-7667, 2019. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 706, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 706, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 706, + 317, + 714 + ], + "type": "text", + "content": "20632" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 686 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 100, + 289, + 161 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 289, + 161 + ], + "type": "text", + "content": "[53] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6629-6638, 2019. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 162, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 162, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 70, + 162, + 288, + 212 + ], + "type": "text", + "content": "[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 213, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 213, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 70, + 213, + 288, + 253 + ], + "type": "text", + "content": "[55] Wei Wei. Research progress on virtual reality (vr) and augmented reality (ar) in tourism and hospitality: A critical review of publications from 2000 to 2018. Journal of Hospitality and Tourism Technology, 10(4):539-570, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 255, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 255, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 288, + 295 + ], + "type": "text", + "content": "[56] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 296, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 296, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 70, + 296, + 288, + 346 + ], + "type": "text", + "content": "[57] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 347, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 347, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 70, + 347, + 288, + 397 + ], + "type": "text", + "content": "[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 399, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 399, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 70, + 399, + 288, + 449 + ], + "type": "text", + "content": "[59] Xu Yan, Zhihao Yuan, Yuhao Du, Yinghong Liao, Yao Guo, Shuguang Cui, and Zhen Li. Comprehensive visual question answering on point clouds through compositional scene manipulation. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 450, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 450, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 70, + 450, + 288, + 499 + ], + "type": "text", + "content": "[60] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1856-1866, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 501, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 501, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 70, + 501, + 288, + 552 + ], + "type": "text", + "content": "[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 553, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 553, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 70, + 553, + 288, + 613 + ], + "type": "text", + "content": "[62] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1791-1800, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 614, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 614, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 70, + 614, + 288, + 654 + ], + "type": "text", + "content": "[63] Zhihao Yuan, Xu Yan, Zhuo Li, Xuhao Li, Yao Guo, Shuguang Cui, and Zhen Li. Toward explainable and fine-grained 3d grounding through referring textual phrases. arXiv preprint arXiv:2207.01821, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 655, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 655, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 655, + 288, + 686 + ], + "type": "text", + "content": "[64] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Zhen Li, and Shuguang Cui. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense caption" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 100, + 524, + 437 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 324, + 100, + 524, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 100, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 324, + 100, + 524, + 121 + ], + "type": "text", + "content": "ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 122, + 524, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 122, + 524, + 181 + ], + "spans": [ + { + "bbox": [ + 307, + 122, + 524, + 181 + ], + "type": "text", + "content": "[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingquiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 182, + 524, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 182, + 524, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 524, + 232 + ], + "type": "text", + "content": "[66] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4534-4543, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 234, + 524, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 524, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 524, + 283 + ], + "type": "text", + "content": "[67] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 285, + 524, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 524, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 524, + 334 + ], + "type": "text", + "content": "[68] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 335, + 524, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 524, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 524, + 385 + ], + "type": "text", + "content": "[69] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 387, + 524, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 524, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 524, + 437 + ], + "type": "text", + "content": "[70] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "20633" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_content_list.json b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..de74d5bac95ce775b765106dc7487d92432fec98 --- /dev/null +++ b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach", + "text_level": 1, + "bbox": [ + 189, + 130, + 779, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mir Rayat Imtiaz Hossain $^{1,2}$", + "bbox": [ + 99, + 202, + 352, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of British Columbia", + "bbox": [ + 122, + 220, + 377, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mennatullah Siam $^{3,1}$", + "bbox": [ + 364, + 203, + 529, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Vector Institute for AI", + "bbox": [ + 418, + 220, + 602, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Canada CIFAR AI Chair", + "bbox": [ + 403, + 238, + 607, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leonid Sigal $^{1,2,4}$", + "bbox": [ + 571, + 203, + 704, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "James J. Little", + "bbox": [ + 745, + 203, + 867, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Ontario Tech University", + "bbox": [ + 640, + 220, + 846, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base categories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mechanism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-20 $^i$ and Pascal5 $^i$ , without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabeled test data can be used to improve the prompts, which we refer to as transductive prompt tuning.", + "bbox": [ + 75, + 323, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of foundational models, trained on broad datasets, in natural language processing (e.g., GPT-3 [5]) and in vision-language (e.g., CLIP [31]), exhibits powerful generalization and strong performance on multiple downstream tasks. These models have been adapted through different prompting techniques [6, 43] for them to be used in the few-shot scenario.", + "bbox": [ + 75, + 763, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, despite successes, including in localization [6, 43, 50], prompting, which allows for few-shot demon", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdbde3ca844e54cf00bd6dbaf756c69e556440150096a7dce4fe7910aa402885.jpg", + "image_caption": [ + "Figure 1. Overview of the Proposed Approach. Prompting multi-scale transformer decoders for generalized few-shot segmentation. Our approach is a simple approach that allows test-time transductive prompt tuning (see red arrows)." + ], + "image_footnote": [], + "bbox": [ + 526, + 290, + 864, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strations in dense prediction, and specifically in semantic segmentation tasks, is relatively under-explored.", + "bbox": [ + 498, + 657, + 890, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot semantic segmentation [24, 40, 45, 46, 48, 49] aims to segment novel (unseen) classes with few labelled training examples. Most state-of-the-art methods rely on meta-learning, to leverage abundant training data as a form of \"data augmentation\" to construct many tasks that resemble test-time few-shot inference [26, 45]. In general, few-shot learning approaches can be classified as either inductive or transductive. Inductive approaches mainly rely on the training data, while transductive ones [4], in addition, exploit unlabelled test data in an unsupervised manner to improve the performance $e.g.$ , by leveraging entropy priors over predictive class distributions. However, both cases largely focus on the performance of the novel categories; which isn't particularly realistic. A recently proposed gen", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23470", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eralized few-shot segmentation (GFSS) setting defines a more realistic scenario where the goal is to perform well on all classes, novel and base [36]. This is considerably more challenging.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompting has been proven to be effective for learning from few demonstrations, as seen in LLMs [5]. It was also explored in visual prompt tuning [16] to efficiently fine-tune vision transformers for new tasks. We posit that prompting within a transformer-based [38] architecture would similarly offer an effective and flexible mechanism for GFSS. This involves learning prompts that can be utilized to perform cross-attention with input image to make predictions. While learning prompts for base classes, with abundant data, is relatively straightforward, it becomes considerably more challenging for novel classes with few examples. Specifically, one must ensure that novel prompts learned from few samples are sufficiently distinct from the base prompts to avoid novel-base class misclassification.", + "bbox": [ + 75, + 155, + 470, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we develop a simple yet highly effective visual prompting of transformer decoders for dense prediction at multiple scale that relies on novel-to-base causal attention without meta-training. We look upon queries in DETR-style architectures as a form of visual prompts and devise a mechanism to initialize and learn novel prompts. Then the novel-to-base causal attention allows base prompts to affect novel prompt representations, but not the other way around. This, intuitively, allows novel prompts to be repulsed and/or attracted towards their base counterparts. This attention is shared across scales (layers of the transformer) which, as we show, leads to more robust learning and improved performance. The multi-scale refinement of prompts helps interacting and reasoning between image features at multiple scales which in turn helps in better dense prediction. Finally, we extended this architecture in a transductive setting, where both novel and base prompts can be fine-tuned at test-time with respect to unsupervised objective to further improve performance. The proposed architecture and methodology is shown in Fig. 1.", + "bbox": [ + 75, + 369, + 470, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 96, + 676, + 326, + 691 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Designing a multi-scale visual prompting transformer decoder architecture for GFSS, featuring learnable prompts that allow the creation of new prompts for novel classes, initialized through masked average pooling of support images (with their masks).", + "- Within this architecture, proposing and learning a multiscale (shared) novel-to-base cross-attention mechanism between the novel and base prompts.", + "- Proposing a novel transductive prompt tuning, which allows the visual prompts to be tuned on test (unlabeled) images, hence the term transductive." + ], + "bbox": [ + 76, + 695, + 468, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 89, + 637, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot Learning. Few-shot learning, a longstanding concept in computer vision, serves as a paradigm to evaluate the data efficiency and adaptability of learning algorithms. It focuses on making predictions for novel classes with a few labeled examples, often by leveraging information from disjoint base classes for which there is abundant data. Few-shot learning has been extensively studied for image classification [12, 32-34, 39]. Most approaches can be categorized as either transfer or meta-learning-based. Transfer learning methods [11] typically pre-train the model on base classes with abundant data and then fine-tune this model for novel classes with few samples. Meta-learning [12, 32] approaches, on the other hand, simulate the inference stage during meta-training by sampling tasks of support and target sets. Meta-learning approaches can be sub-categorized into gradient-based and metric-learning based. Gradient-based approaches [12, 32] learn the commonalities among different tasks. Metric-based (or prototype-based) methods [34, 39] learn image embeddings and compare the distances between them. Despite its popularity and strong performance, meta-learning is found to be sensitive to the meta-training setup [7] and worse performance with domain shift [9] is observed.", + "bbox": [ + 496, + 119, + 890, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalized few-shot learning, introduced in [14] as a more realistic variant, allows query images from either base or novel categories. Recently, the transductive setting has been proven to be effective for few-shot learning [3, 4, 13, 29, 52]. Transductive approaches exploit information of unlabelled test images during inference to improve classification [11, 22, 27, 30]. These approaches build on classic semi-supervised learning approaches, such as graph-based label propagation [23] or entropy minimization [11]. Recent variants of few-shot learning problems have also been established in other more granular domains, e.g., detection [17, 41, 42, 44] and segmentation.", + "bbox": [ + 496, + 470, + 890, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot segmentation. Less common than classification, few-shot segmentation is emerging as the dense counterpart to the aforementioned approaches. Few-shot segmentation approaches can be categorized to prototype-based and dense comparison-based methods. Prototype-based methods mostly relied on extracted prototypes from the support set to guide the segmentation of the target image, either as a single prototype [40, 49], part-aware [24] or a prototype mixture [45]. Dense comparison methods rely on dense comparisons between the few template images and the target image [46, 48] with extension to multiscale [26]. A recent approach proposed generalized few-shot segmentation (GFSS) task where the unlabelled target images are used to improve the segmentation of novel classes while maintaining the performance on base classes through knowledge distillation [13]. A concurrent work [21] on GFSS learns a set", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1The term 'causal' [28] indicates its unidirectional nature.", + "bbox": [ + 94, + 886, + 401, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of orthogonal prototypes for base and novel categories, yet their GFSS setup is easier as they sample in their inference balanced base class images along with the novel ones, unlike the setup we follow [13]. Recent work [50] proposed a prompting strategy within a meta-learning framework using few examples. This approach is designed for 1-way segmentation only and performs prompting at a single scale unlike our proposed approach, a non-meta learning scheme, which performs multi-scale prompting and is designed for $k$ -way GFSS.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompting. An emerging direction in learning with limited labelled data is to utilize foundation models [2] that have strong generalization capabilities. These models, pretrained on large-scale data, demonstrate strong generalization in few-shot settings [1]. Prompting is often used to adapt these foundation models to few-shot tasks, where it formulates the inference problem similarly to the training. Examples include auto-regressive modelling [1] where inference involves generating the next tokens after receiving the correct prompt. For few-shot learning, prompts include few-shot demonstrations followed by the query. Prompting can employ discrete prompts [1] or learning continuous prompts [20] through back-propagation. In computer vision, both vision-language prompting [1] and visual prompting [16] of transformer-based models have shown success in various tasks, particularly in few-shot scenarios. However, effectively prompting models designed for localization and segmentation remains an open question.", + "bbox": [ + 75, + 244, + 472, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Work. A recent method [6] proposed prompting of transformer decoders for few-shot object detection. Although conceptually related, their approach is designed for a different task (detection) compared to ours (semantic segmentation), trained within a meta-learning framework unlike ours. Further, [6] lacks novel-to-base attention and multi-scale prompt refinement which are at the core of our architectural design. A recent foundation model for semantic segmentation, SAM [18], exhibits strong performance when prompted by point, bounding box, or text. However, it is not specifically designed for few-shot inference, potentially requiring fine-tuning with the support set and other non-trivial changes. More importantly, the use of extensive data to train the model makes it unsuitable for few-shot tasks since it has likely encountered many target classes during training. In addition to aforementioned differences, including multi-scale prompting and novel-to-base attention, our method is the first to explore transductive prompt-tuning where the visual prompts can be further optimized with the unlabelled target images.", + "bbox": [ + 75, + 518, + 472, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 830, + 284, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classic Few-shot Segmentation. In the classic few-shot segmentation scenario, the categories are divided into two sets of disjoint sets: base and novel. The problem is", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "framed as an $n$ -way $k$ -shot segmentation task, where $n$ is the number of novel classes and $k$ is the number of examples available for each class. A typical solution involves meta-learning with two-stage training. In the meta-training stage, a support set is created with $n$ randomly selected base classes, and $k$ examples from each class are used to train the model. In the meta-testing stage, the model is evaluated on tasks with $k$ examples from $n$ novel classes. An alternative approach is non-meta learning, where a model is initially trained on abundant data for base classes, then frozen and fine-tuned for novel classes using few examples. The focus is on segmenting novel classes in test images without considering base class segmentation.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalized Few-shot Segmentation (GFSS). Generalized few-shot segmentation, proposed by [36], extends the classic few-shot setting. In this new setup the model must preserve the segmentation performance on base classes in addition to generalizing to novel classes. While during finetuning the support set (typically) only contains $k$ -shot examples for each novel category, the model is still expected to make predictions of both novel and base classes on the test images. We follow the recent more practical setup of generalized few-shot segmentation [13] within a non-meta learning framework, details provided in Section 5.1.", + "bbox": [ + 496, + 287, + 893, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 459, + 591, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform prompting using a set of learnable visual prompts, where, intuitively $^2$ , each prompt corresponds to a class embedding. These learned visual prompts cross-attend with image features at multiple scales. Initially, we pre-train a base model with abundant examples of base classes, learning visual prompts for them. We then fix these base prompts and fine-tune new visual prompts for novel classes with a few examples. However, learning prompts from limited data is challenging because it leads to confusion between base and novel categories. To improve novel class generalization, we introduce a uni-directional causal attention mechanism between novel and base prompts. Intuitively, this allows better contextualization and repulsion / attraction between base and novel prompts. Section 5 demonstrates the empirical effectiveness of this mechanism.", + "bbox": [ + 496, + 484, + 893, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows our proposed architecture. It centers around learnable embeddings called visual prompts, which interact with image features at multiple scales. Our design is inspired by Mask2former [10] with a couple of key differences: (i) the number of visual prompts aligns with the number of classes, each representing a class as a learned embedding; (ii) we adopt per-pixel classification, instead of mask classification, thereby eliminating the query classification component of [10] and defining custom segmentation heads capable of making per-pixel predictions.", + "bbox": [ + 496, + 712, + 893, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2This may not be precisely true in practice, but is a reasonable abstraction for intuitive understanding.", + "bbox": [ + 500, + 875, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23472", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c837ac7f32204fe304d6787618d26b719d8b773a7770dce2790e294b53ff944f.jpg", + "image_caption": [ + "Figure 2. Detailed architecture of our proposed visual prompting of multiscale transformer decoder. Our design initializes the novel visual prompts using the support set. This is followed by consecutive novel-to-base causal attention, $\\mathcal{CA}$ , and prompt-to-target features cross attention, $\\mathcal{C}$ , across scales. Note that causal attention uses shared weights across the scales and decoder layers. Our design allows for transductive fine-tuning of the visual prompts leveraging the unlabelled test image." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 823, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Visual Prompting Multiscale Transformers", + "text_level": 1, + "bbox": [ + 76, + 430, + 442, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Base Visual Prompts. Given $B$ base classes, we define $B$ visual prompts each representing a class and initialize them randomly. These visual prompts first self-attend to learn the relationship between themselves and then are used to prompt the image features at multiple scales. The refinement of base prompts at each decoder layer is given by:", + "bbox": [ + 76, + 453, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {B} ^ {(l)} = \\mathcal {A} \\left(V _ {B} ^ {(l - 1)}\\right) + \\mathcal {C} \\left(V _ {B} ^ {(l - 1)}, F ^ {(l - 1)}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 551, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ and $\\mathcal{C}$ refer to multi-headed self and cross attention respectively; $V_{B}^{(l)}\\in \\mathbb{R}^{B\\times C}$ are the base visual prompts at layer $l$ ; $F^{(l)}\\in \\mathbb{R}^{H_lW_l\\times C_l}$ is the flattened image feature at layer $l$ ; $H_{l}$ and $W_{l}$ denote image dimensions at that layer; $C_l$ denotes the embedding dimension.", + "bbox": [ + 76, + 579, + 468, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After refining the prompts through multiple levels of transformer attention, a segmentation head, denoted as $\\mathcal{H}(\\cdot)$ , is defined. This segmentation head takes the final refined prompts, $V_{B}^{(L)}$ , and the highest resolution image features, $F^{(L)}$ , as input, producing per-pixel per-class predictions $O_{base} \\in \\mathbb{R}^{B \\times H \\times W}$ (see Section 4.3).", + "bbox": [ + 76, + 655, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Novel Visual Prompts. After training on the base classes, we freeze all the layers of the model except the visual prompts representing the base classes. Given there are $N$ novel classes, we add $N$ new visual prompts. We further add a causal unidirectional attention module that allows the novel prompts to be enhanced by the base prompts (see Section 4.2). The output is passed to a segmentation head to make predictions for the novel classes.", + "bbox": [ + 75, + 750, + 468, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The novel visual prompts are first refined by a unidirectional novel-to-base causal attention module. The enhanced", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "novel prompts are then concatenated with the base prompts. Overall the concatenated prompts are refined by:", + "bbox": [ + 498, + 431, + 888, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {N} ^ {(l)} = \\mathcal {C A} \\left(V _ {B} ^ {(l)}, V _ {N} ^ {(l)}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 479, + 725, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {A} ^ {(l)} = \\left[ V _ {B} ^ {(l)}, V _ {N} ^ {(l)} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 502, + 704, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {A} ^ {(l)} = \\mathcal {A} (V _ {A} ^ {(l - 1)}) + \\mathcal {C} (V _ {A} ^ {(l - 1)}, F ^ {(l - 1)}),\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 529, + 841, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{N}^{(l)}\\in \\mathbb{R}^{N\\times C}$ are novel prompts at layer $l$ ; $N$ is the total number of novel class and $C$ is the embedding size; $V_{A}^{(l)}\\in \\mathbb{R}^{(B + N)\\times C}$ are all visual prompts including base and novel; $[.,.]$ is the concatenation operator. The operation $\\mathcal{CA}$ denotes the causal novel-to-base attention and is explained in Section 4.2.", + "bbox": [ + 498, + 570, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each of the $N$ novel visual prompts (representing a novel class) is initialized using masked average global pooling of the k-shot support images corresponding to the class masked by the ground truth binary masks.", + "bbox": [ + 496, + 667, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n} ^ {(0)} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\sum_ {x , y} M _ {n} ^ {k} (x , y) F _ {n} ^ {k} (x , y)}{\\sum_ {x , y} M _ {n} ^ {k} (x , y)}, \\forall n \\in N \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 747, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{n}^{(0)} \\in \\mathbb{R}^{1 \\times C}$ is the initial novel visual prompt for novel class $n$ ; $F_{n}^{k}$ denotes the image features of $k$ -th shot for novel class; $M_{n}^{k}$ denotes the corresponding ground-truth binary mask; $x, y$ are the spatial locations; $N$ is the total number of novel classes and $K$ is the total number of shots.", + "bbox": [ + 496, + 821, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Novel-to-base Causal Attention", + "text_level": 1, + "bbox": [ + 76, + 90, + 354, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key component of our proposed approach is the unidirectional novel-to-base prompt causal3 attention. The base visual prompts are learned with abundant examples; this is not the case for novel prompts. Learning novel prompts directly or having novel prompts impact the base ones could therefore lead to degradation of performance. However, having novel prompts aware of their base counterparts could be quite helpful. We hypothesize that causal attention between the novel visual prompts and base prompts could help contextualize novel class embeddings, reducing confusion with the base counterparts. Motivated by this intuition, we propose a novel-to-base causal attention layer that is repeated at each scale (or layer of the transformer decoder); see Fig. 2.", + "bbox": [ + 75, + 113, + 468, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed causal attention module is a single layer cross-attention where the Key $K^{(l)}$ , Value $V^{(l)}$ and $Q^{(l)}$ for layer $l$ of the decoder are given by:", + "bbox": [ + 75, + 325, + 468, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {(l)} = V _ {N} ^ {(l)} W ^ {Q}; \\quad K ^ {(l)} = V _ {B} ^ {(l)} W ^ {K}; \\quad V ^ {(l)} = V _ {B} ^ {(l)} W ^ {V}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 378, + 460, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W^{Q}, W^{K}, W^{V}$ are weight matrices for $Q, K$ and $V$ respectively. The parameters $W^{Q}, W^{K}, W^{V}$ are shared at all the layers of this cross-attention module to reduce the number of trainable parameters during fine-tuning on novel classes to prevent overfitting and poor generalization.", + "bbox": [ + 75, + 409, + 468, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This uni-directional novel-to-base causal attention refines the novel visual prompts $V_{N}^{(l)}$ at each decoder layer $l$ by:", + "bbox": [ + 75, + 486, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {N} ^ {(l)} = \\operatorname {s o f t m a x} (Q ^ {(l)} (K ^ {(l)}) ^ {\\mathrm {T}}) V ^ {(l)}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 530, + 388, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This results in a set of contextualized novel visual prompts.", + "bbox": [ + 76, + 555, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Segmentation Head", + "text_level": 1, + "bbox": [ + 76, + 589, + 264, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base Segmentation Head. The base segmentation head, $\\mathcal{H}_B$ , during the base training in the proposed model is relatively simple and takes the form of a three layer MLP that projects a refined prompt into, effectively, a class prototype; computes the (dot product) similarity of each pixel feature with that prototype, and then applies softmax along the channel / class dimension to obtain per-pixel and per-class probability:", + "bbox": [ + 75, + 613, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} O _ {b a s e} = \\mathcal {H} _ {B} (V _ {B} ^ {(L)}, F ^ {(L)}) \\\\ = \\operatorname {M L P} \\left(V _ {B} ^ {(L)}\\right) \\cdot \\left[ F ^ {(L)} \\right] ^ {\\mathrm {T}}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 742, + 468, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O_{base} \\in \\mathbb{R}^{B \\times H \\times W}$ and the class assignment can then be made by an argmax along the channel dimension.", + "bbox": [ + 75, + 803, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Novel Segmentation Head. While we could, in principle, use the same form of segmentation head during few-shot", + "bbox": [ + 75, + 835, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "inference as that defined for the base classes (see Eq. 3), simply fine-tuning it for novel classes, this proves ineffective in practice. Specifically, fine-tuning the MLP leads to significant overfitting given the few examples and lack of base class data during this step. So, instead, we resort to a simpler form of the novel segmentation head, $\\mathcal{H}_N$ , where we only learn a residual to the MLP learned during base training (which we keep fixed), mainly:", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} O _ {n o v e l} = \\mathcal {H} _ {N} (V _ {N} ^ {(L)}, F ^ {(L)}) \\\\ = \\left[ \\operatorname {M L P} _ {\\bullet} \\left(V _ {N} ^ {(L)}\\right) + W _ {N} \\right] \\cdot \\left[ F ^ {(L)} \\right] ^ {\\mathrm {T}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 217, + 857, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O_{\\text{novel}}$ is output for novel classes and $\\bullet$ designates that parameters of MLP are frozen and only $W_N \\in \\mathbb{R}^{N \\times C}$ weight matrix is learned. This weight matrix is initialized in the same manner as novel prompts, in Eq. 2.", + "bbox": [ + 496, + 273, + 890, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Transductive Prompt Tuning", + "text_level": 1, + "bbox": [ + 498, + 335, + 759, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent approaches [4, 13] show the effectiveness of a good transductive setting for few-shot dense prediction tasks. Transductive approaches perform test-time optimization for each unlabeled test image and utilize specific characteristics and information of the test instances to improve prediction.", + "bbox": [ + 496, + 359, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the flexibility of our approach, in this section we outline the process of adapting our model for transductive inference, improving GFSS performance through transductive fine-tuning of the visual prompts. We build on the transductive losses proposed in DlaM [13] and we describe them here for completeness.", + "bbox": [ + 496, + 435, + 890, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The losses are designed to maximize the mutual information between the learned features and the corresponding predictions on the test image, which is achieved through maximizing: $\\mathrm{H}(O) - \\mathrm{H}(O|I)$ , where $I, O$ are random variables associated with the pixels and predictions distributions respectively; $\\mathrm{H}(O)$ is the marginal and $\\mathrm{H}(O|I)$ is the conditional entropy. The conditional entropy $\\mathrm{H}(O|I)$ involved is given by sum of cross-entropy loss over the supervised support set and entropy of the predicted probability given the test image $I$ . The marginal entropy is given by KL divergence loss relying on an estimated region proportion prior following [4]. Additional details are in [13].", + "bbox": [ + 496, + 526, + 890, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, similar to [13], we use knowledge-distillation loss to preserve base class performance, given by:", + "bbox": [ + 496, + 707, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D}} = \\mathrm {K L} \\left(O _ {\\text {b a s e}} ^ {\\text {n e w}} \\mid \\mid O _ {\\text {b a s e}} ^ {\\text {o l d}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 750, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O_{base}^{old}$ is the prediction probability of the base classes of the frozen model after base training; $O_{base}^{new}$ is the prediction probability of the base classes on the model with our additional components to handle novel categories that is being fine-tuned. These include the multi-scale causal novel-to-base cross attention and the novel prompts.", + "bbox": [ + 496, + 771, + 890, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall transductive objective function is given by:", + "bbox": [ + 517, + 863, + 880, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a n s .}} = \\alpha \\mathrm {H} (O | I) - \\mathrm {H} (O) + \\gamma \\mathcal {L} _ {\\mathrm {K D}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 883, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3We use the term causal borrowing from its use in the look-ahead-attention-masking for decoder Transformers [38] (e.g., in language).", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23474", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\gamma$ are balancing hyper-parameters.", + "bbox": [ + 76, + 90, + 392, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We noticed that applying transduction from the first iteration underperforms, likely due to a poor initial estimate of the marginal distribution. To address this, we only apply per-pixel cross-entropy loss for a set number of iterations and then incorporate the remaining transductive losses from Eq. 5 when we achieve a more accurate estimate of the marginal distribution.", + "bbox": [ + 75, + 106, + 468, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 215, + 209, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 239, + 388, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on two few-shot segmentation benchmarks [36]: COCO- $20^{i}$ and PASCAL- $5^{i}$ . Following [13] we report average performance of our model on 10K test images for COCO- $20^{i}$ and on all test images on PASCAL- $5^{i}$ . The results are reported as an average across five different runs on four different splits.", + "bbox": [ + 75, + 263, + 468, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Pre-processing. We adopt the same setup as DIAM [13]. During base training, images containing both base and novel categories are kept, but the pixels corresponding to novel classes are relabeled as background. This approach, as noted by [13], presents challenges due to potential ambiguity when the base model predicts novel categories as background. In fine-tuning, only novel classes are labeled in the support set, with base classes marked as background. To avoid unfair shot increases, images with multiple novel categories are excluded from the support set.", + "bbox": [ + 75, + 354, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. We evaluate performance using the standard mean intersection-over-union (mIoU) metric. Following [13], the mean score for generalized few-shot segmentation is computed as the average between base and novel mIoUs. This approach addresses bias in datasets like PASCAL- $5^{i}$ and COCO- $20^{i}$ , where base classes outnumber novel classes threefold.", + "bbox": [ + 75, + 507, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use PSPNet [51] with Resnet-50 [15] as feature extractor. We utilize MSDeformAttn-based pixel decoder [53] from Mask2Former [10]. The dimension $C$ for visual prompts is chosen to be 256. Contrary to [10], the base model is trained with per-pixel cross-entropy loss. We use AdamW optimizer [25] with base learning rate of $1 \\times 10^{-4}$ , a weight decay of 0.05 and a batch size of 16. Following [13, 19], we train base model for 20 epochs for COCO- $20^{i}$ and 100 epochs for PASCAL- $5^{i}$ .", + "bbox": [ + 75, + 613, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During novel fine-tuning, we freeze the parameters of the feature extractor, pixel decoder, and base segmentation head, and only optimize the parameters of the visual prompts, novel segmentation head, and novel-to-base causal attention module. For the inductive case, the model is finetuned with per-pixel cross-entropy loss for 100 iterations. For the transductive setting, the model is fine-tuned for 40 iterations with only cross-entropy loss and then fine-tuned for 60 more iterations with full transductive losses mentioned in Eq. 5. The hyper-parameters $\\alpha$ and $\\gamma$ are set to", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "be 100 and 25 respectively. We use AdamW optimizer with a learning rate of $5 \\times 10^{-3}$ and a weight decay of 0.05 for fine-tuning.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 147, + 591, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with the State-of-The-Art. Table 1 compares our approach in both inductive and transductive setting against the state-of-the-art. As observed, our model outperforms existing approaches. Interestingly, our approach, in an inductive setting (where optimization is performed only using a supervised few-shot support set, and evaluation is conducted on all test images), outperforms transductive approaches like DIaM [13] for both 1-shot and 5-shot cases on the COCO-20 $^i$ dataset and for the 1-shot case on Pascal-5 $^i$ . This is notable despite transductive approaches requiring test-time optimization for each unlabeled test image separately, leading to significantly longer inference times.", + "bbox": [ + 496, + 166, + 890, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison in the inductive setting, we finetuned DIaM without the transduction losses. We also finetuned POP [21] in our inductive setup. POP [21] randomly samples balanced base class image, making it easier for them to retain base class information and uses a form of transduction where their model relabels novel class pixels initially labelled as background. We observe that the performance of POP [21] degrades substantially, especially for the base classes in our inductive setting.", + "bbox": [ + 496, + 348, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our inductive approach has $5.78\\%$ higher mIoU for 1-shot and $11.16\\%$ higher mIoU for 5-shot on COCO-20 $^i$ than inductive DIaM. For PASCAL-5 $^i$ , the improvement from inductive DIaM is $7.71\\%$ for 1-shot and $13.29\\%$ for 5-shot. In fact, compared to the best inductive GFSS approach BAM [19], our inductive setting achieves an improvement of approximately $2.78\\%$ and $7.59\\%$ for 1-shot and 5-shot respectively on COCO-20 $^i$ and $5.24\\%$ and $12.32\\%$ for 1-shot and 5-shot respectively on PASCAL-5 $^i$ . Our method, in the inductive setting, even outperforms transductive DIaM [13] by $2.03\\%$ and $2.28\\%$ for 1-shot and 5-shot respectively on COCO-20 $^i$ and by $1.79\\%$ for 1-shot on PASCAL-5 $^i$", + "bbox": [ + 496, + 484, + 890, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On applying transductive fine-tuning to our method, overall mIoU increases considerably, outperforming transductive DIaM [13] and POP [21] on both 1-shot and 5-shot inference for both datasets. The improvement over DIaM for $\\mathrm{COCO - }20^{i}$ is $3.3\\%$ for 1-shot and $3.93\\%$ for 5-shot. For PASCAL- $5^{i}$ , our transductive setting outperforms DIaM by $5.11\\%$ and $3.15\\%$ for 1 and 5-shots.", + "bbox": [ + 496, + 667, + 890, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation", + "text_level": 1, + "bbox": [ + 500, + 784, + 602, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We performed ablation studies (refer to Table 2) to explain our design choices and contribution of each components.", + "bbox": [ + 496, + 806, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Prompts Initialization. As seen in (1) and (2) of the Table 2, initializing novel prompts with masked mean pooling from the support set images improves both base and novel class accuracy compared to random initialization.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1b3ee162ac5f82561f2c4a98507bfaccdf461f5bb815cfb82aa291c7d35a6253.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLearningCOCO-20i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CAPL [36] (CVPR22)Inductive43.217.2125.2143.7111.0027.36
BAM [19] (CVPR22)Inductive49.8414.1632.0049.8516.6333.24
DIaM (w/o trans.)* [13] (CVPR23)Inductive42.6915.3229.0038.4720.8729.67
POP† [21] (CVPR23)Inductive30.389.6320.0024.5316.1920.36
Ours (w/o trans.)Inductive51.5518.0034.7851.5930.0640.83
RePRI [4] (CVPR21)Transductive5.624.745.188.858.848.85
DIaM [13] (CVPR23)Transductive48.2817.2232.7548.3728.7338.55
POP† [21] (CVPR23)Transductive54.7115.3135.0154.9029.9742.44
Ours (w/ trans.)Transductive53.8018.3036.0553.8131.1442.48
", + "bbox": [ + 197, + 89, + 772, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea07f875f1c9e1d9f8c8a123e0e51b2941c5a76531ad1190c510f84238666a38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLearningPASCAL-5i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CANet [49] (CVPR19)Inductive8.732.425.589.051.525.29
PFENET [35] (TPAMI20)Inductive8.322.675.508.831.895.36
PANET [40] (ICCV19)Inductive31.8811.2521.5732.9515.2524.1
SCL [47] (CVPR21)Inductive8.882.445.669.111.835.47
MiB [8] (CVPR20)Inductive63.808.8636.3368.6028.9348.77
CAPL [36] (CVPR22)Inductive64.8017.4641.1365.4324.4344.93
BAM [19] (CVPR22)Inductive71.6027.4949.5571.6028.9650.28
DIaM (w/o trans.)* [13] (CVPR23)Inductive66.7927.3647.0864.0534.5649.31
POP† [21] (CVPR23)Inductive46.6819.9633.3241.5036.2638.80
Ours (w/o trans.)Inductive74.5834.9954.7974.8650.3462.60
RePRI [4] (CVPR21)Transductive20.7610.5015.6334.0620.9827.52
DIaM [13] (CVPR23)Transductive70.8935.1153.0070.8555.3163.08
POP† [21] (CVPR23)Transductive73.9235.5154.7274.7855.8765.33
Ours (w/ trans.)Transductive76.3939.8358.1176.4256.1266.27
", + "bbox": [ + 197, + 267, + 772, + 488 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c20aaf93c884ebf1a1bb863e732d9b0545485eacb951e2ab849abd5418eb97f.jpg", + "table_caption": [ + "Table 1. Comparison to SoTA methods on COCO-20 $^i$ and PASCAL-5 $^i$ . The mIoU results are reported as the average across 5 different runs. All methods use ResNet-50 backbone. $\\dagger$ POP [21] uses an easier and less practical GFSS setup, e.g., balanced base class training in addition to support set during novel fine-tuning and uses a form of transduction; inductive setting is implemented by us following our setup. Results of the rest of the methods are obtained from [13] except the inductive setting of DIA ${M}^{*}$ , which we implemented ourselves. The w/o trans., w/ trans. denote without and with transduction respectively." + ], + "table_footnote": [], + "table_body": "
Causal AttentionPrompt InitializationTransductionmIoU
BaseNovelMean
(1)NoneRandomNo50.1811.2230.70
(2)NoneMasked PoolingNo50.4311.3130.87
(3)First Layer OnlyMasked PoolingNo51.5312.2631.90
(4)Separate WeightsMasked PoolingNo51.0618.0534.56
(5)Shared WeightsRandomNo50.7517.4134.08
(6)Shared WeightsMasked PoolingNo51.5518.0034.78
(7)Shared WeightsMasked PoolingYes53.8018.3036.05
", + "bbox": [ + 78, + 583, + 468, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation Study on COCO- ${20}^{i}$ . The 1-shot performane, averaged over all folds.", + "bbox": [ + 75, + 681, + 468, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This is more pronounced with novel-to-base causal attention (5)-(6), amplifying the performance gain from masked pooling initialization.", + "bbox": [ + 75, + 718, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel-to-base Causal Attention. As seen in (4)-(7) of Table 2, novel-to-base causal attention yields the most significant improvement in novel class accuracy compared to (1)-(2). This aligns with our hypothesis that uni-directional causal attention between rich base prompts (learned with many examples) and novel prompts (learned with few examples) increases separation among novel classes and pushes the feature space of base prompts away from the novel prompts. It is further confirmed by Figure 3 and Fig-", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0913caea56e1a20fe6b9e35147adf3fe3496a7b9ab746fe2ea30d39e2c1fddd2.jpg", + "image_caption": [ + "Figure 3. Confusion Matrix for: (left) model with Novel-to-Base Causal Attention; (right) model without Novel-to-Base Causal Attention. Note that Novel-to-Base Causal Attention reduces confusion between novel and base categories (bottom-left block)." + ], + "image_footnote": [], + "bbox": [ + 500, + 583, + 908, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ure 4. As observed in Figure 3, the novel-to-base causal attention leads to lower confusion between base and novel classes. Figure 4 shows a t-SNE [37] visualization of the features of base and novel visual prompts for 10 different", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23476", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23f666fdbb8253d19932f086feb2fb76eb6041f3a7d88db61e4da710a9e9d5f3.jpg", + "image_caption": [ + "Figure 4. TSNE visualization of the learned base and novel prompt features: (left) with Novel-to-Base Causal Attention; (right) without Novel-to-Base Casual Attention." + ], + "image_footnote": [], + "bbox": [ + 78, + 92, + 486, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "test images. As observed, the incorporation of novel-to-base causal attention leads to much clearer separation of visual prompt features, particularly for novel prompts.", + "bbox": [ + 75, + 354, + 468, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiscale Causal Attention. We also ablated the placement of the causal attention module. We observe that applying it only at the first layer (3) slightly improves novel mIoU compared to the model without causal attention. Yet, it is notably inferior to applying at every decoder layer (4)-(7). We could have opted for separate parameters for the causal attention module at each layer. However, as seen in (4), having separate weights offers little improvement to novel mIoU despite significant increase in total number of parameters to optimize. In fact, it decreases the base accuracy, leading to overall worse performance.", + "bbox": [ + 75, + 401, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 5 compares the qualitative results of our model for 1-shot GFSS on Pascal- $5^{i}$ in both inductive and transductive settings with the baseline lacking causal attention and DIaM [13]. Notably, for first row, both our inductive and transductive approaches predict more accurate segmentation of the cow compared to DIaM. In second row, the causal attention improves person segmentation followed by transduction. In the third row, DIaM completely misses the novel category chair. The causal attention allows our model to better differentiate the base class dining table from the novel class chairs. Transduction improves accuracy of the segmentation of dining table. In fourth row, we can observe that both DIaM and our model without causal attention misdetects novel class cat for cow at multiple locations. Causal attention allows our model to make a uniform and more accurate prediction of cat. In the fifth row, we observe that DIaM [13] completely misses detecting the base category potted plant and poorly segments person. The baseline without causal attention misdetects the potted plant with novel class bus and misdetects dining table with chair. Causal attention improves the accuracy which is further improved by transduction. Last row", + "bbox": [ + 75, + 569, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f675ac8b2fdd0e8dedcc775892190cc201f7b23bbfb609070141f9c1ff19f758.jpg", + "image_caption": [ + "Figure 5. Qualitative Results for 1-shot on Pascal- $5^{i}$ . The left-most two columns show image and ground truth mask; (Third) Baseline without causal attention; (Fourth); Ours in inductive setting; (Fifth) Ours in transductive setting; (Last) DIaM [13]. Last row illustrates a failure." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 910, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrates a case where our model performs worse than DlaM [13]. Again, we observe that adding causal attention reduces misclassification of person with car, although overall the segmentation of car is less accurate than DlaM [13].", + "bbox": [ + 496, + 452, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the qualitative results, the effectiveness of novel-to-base causal attention at distinguishing base from novel classes becomes apparent. Our model in the inductive setting outperforms transductive DIaM in most cases, despite the inference time being faster.", + "bbox": [ + 496, + 512, + 890, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 594, + 617, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a technique to learn visual prompts from few-shot examples of novel unseen cateogries to prompt a pre-trained segmentation model that not only generalizes well on the unseen novel class but strongly retains performance over the base classes on which the model was pre-trained. We achieve this by prompting the image features at multiple scales and proposing a uni-directional novel-to-base causal attention mechanism that enriches the novel visual prompts through interaction with the base visual prompts. Our proposed method can be applied in both inductive and transductive settings and achieves state-of-the-art performance on two benchmark few-shot segmentation datasets.", + "bbox": [ + 496, + 613, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Acknowledgments and Disclosure of Funding. This work was partially funded by the Vector Institute for AI, Canada CIFAR AI Chair, NSERC CRC and NSERC DG. Hardware resources were partially provided by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute. Additional support was provided by JELF CFI grant and Digital Research Alliance of Canada under the RAC award.", + "bbox": [ + 496, + 814, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23477", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, E. Rutherford, S. Cabi, T. Han, Z. Gong, S. Samangooei, M. Monteiro, J. L. Menick, S. Borgeaud, A. Brock, A. Nematzadeh, S. Sharifzadeh, M. Binkowski, R. Barreira, O. Vinyals, A. Zisserman, and K. Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3", + "[2] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S.v.ArxM.S.BernsteinA.Bosselut J.BohgE.Brunskill, E. Brynjolfsson, S. Buch, D.Card, R.Castellon, N.Chatterji, A. Chen, K.Creel, J.Q.Davis,D.Demszky,C.Donahue,M.Doumbouya,S.Ermon E.Durmus,J.Etchemendy, K.Ethayarajh, L.Fei-Fei, C.FinnT.GaleL.GillespieK. Goel, N. Goodman, S. Grossman, N. Guha, T. Hashimoto, P. Henderson, J. Hewitt, D.E.HoJ.HongK.HsuJ.Huang, T.IcardS.JainD.JurafskyP.KalluriS.KaramchetiG. KeelingF.KhaniO.KhattabP.W.KohM.KrassR. Krishna,R.Kuditipudi,A.Kumar,F.Ladhak,M.Lee,T. LeeJ.LeskovecI.LeventX.L.LiX.LiT.MaA.MlikC.D.ManningS.Mirchandani,E.MitchellZ.MunyikwaS.NairD.Narayanan A.NarayanB.NewmanA. NieJ.C.Niebles,H.NilforoshanJ.Nyarko,G.OgutI. PapadimitriouJ.S.ParkC.PiechE.PortelanceC.Potts, A.RaghunathanR.ReichH.RenF.RongY.RoohaniC. RuizJ.RyanC.ReD.SadighS.SagawaK.Santhanam, A.ShihK.SrinivasanA.TamkinR.TaoriA.W.Thomas F.TramereR.E.WangW.WangB.WuJ.WuY.WuS.M XieM.YasunagaJ.YouM.ZahariaM.ZhangT.Zhang X.ZhangY.ZhangL.ZhengK.Zhou and P.Liang.On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021.3", + "[3] Malik Boudiaf, Imtiaz Ziko, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Information maximization for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 33:2445-2457, 2020. 2", + "[4] M. Boudiaf, H. Kervadec, Z. Masud, P. Piantanida, I. Ben Ayed, and J. Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13979-13988, 2021. 1, 2, 5, 7", + "[5] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, G. Sastry P. Shyam, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33: 1877-1901, 2020. 1, 2", + "[6] A. Bulat, R. Guerrero, B. Martinez, and G. Tzimiropoulos. Fs-detr: Few-shot detection transformer with prompting and without re-training. In International Conference on Computer Vision (ICCV), pages 11793-11802, 2023. 1, 3", + "[7] T. Cao, M. Law, and S. Fidler. A theoretical analysis of the" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "number of shots in few-shot learning. International Conference on Learning Representations (ICLR), 2020. 2", + "[8] F. Cermelli, M. Mancini, S.R. Bulo, E. Ricci, and B. Caputo. Modeling the background for incremental learning in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233–9242, 2020. 7", + "[9] W. Chen, Y. Liu, Z. Kira, Y. Frank Wang, and J. Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), 2019. 2", + "[10] B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 3, 6", + "[11] G. S. Dhillon, P. Chaudhari, A. Ravichandran, and S. Soatto. A baseline for few-shot image classification. In International Conference on Learning Representations (ICLR), 2020. 2", + "[12] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning (ICML), 2017. 2", + "[13] S. Hajimiri, M. Boudiaf, I. B. Ayed, and J. Dolz. A strong baseline for generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11269-11278, 2023. 2, 3, 5, 6, 7, 8", + "[14] B. Hariharan and R. B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In International Conference on Computer Vision (ICCV), 2017. 2", + "[15] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[16] M. Jia, L. Tang, B. Chen, C. Cardie, S. Belongie, B. Hariharan, and S. Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 2, 3", + "[17] B. Kang, Z. Liu, X. Wang, F. Yu, J. Feng, and T. Darrell. Few-shot object detection via feature reweighting. In International Conference on Computer Vision (ICCV), 2019. 2", + "[18] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, A. C. Berg S. Whitehead, W. Lo, P. Dollár, and R. Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3", + "[19] C. Lang, G. Cheng, B. Tu, and J. Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8067, 2022. 6, 7", + "[20] X. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021.3", + "[21] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11319-11328, 2023. 2, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23478", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2", + "[23] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. J. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2", + "[24] Y. Liu, X. Zhang, S. Zhang, and X. He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), 2020. 1, 2", + "[25] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6", + "[26] J. Min, D. Kang, and M. Cho. Hypercorrelation squeeze for few-shot segmentation. In International Conference on Computer Vision (ICCV), pages 6941-6952, 2021. 1, 2", + "[27] A. Nichol, J. Achiam, and J. Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2", + "[28] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 2", + "[29] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8412-8422, 2021. 2", + "[30] L. Qiao, Y. Shi, J. Li, Y. Wang, T. Huang, and Y. Tian. Transductive episodic-wise adaptive metric for few-shot learning. In International Conference on Computer Vision (ICCV), pages 3603-3612, 2019. 2", + "[31] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1", + "[32] S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. 2", + "[33] M. Ren, E. Triantafillou, S. Ravi, J.Snell, K. Swersky, J. B. Tenenbaum, H. Larochelle, and R. S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), 2018.", + "[34] J. Snell, K. Swersky, and R. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2", + "[35] Z. Tian, H. Zhao, M. Shu, Z. Yang, R. Li, and J. Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(2):1050-1065, 2020. 7", + "[36] Z. Tian, X. Lai, L. Jiang, S. Liu, M. Shu, H. Zhao, and J. Jia. Generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11563-11572, 2022. 2, 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 7", + "[38] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 5", + "[39] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 3630-3638, 2016. 2", + "[40] K. Wang, J. Liew, Y. Zou, D. Zhou, and J. Feng. PANet: Few-shot image semantic segmentation with prototype alignment. In International Conference on Computer Vision (ICCV), pages 9197-9206, 2019. 1, 2, 7", + "[41] X. Wang, T. E Huang, T. Darrell, J. E. Gonzalez, and F. Yu. Frustratingly simple few-shot object detection. In International Conference on Machine Learning (ICML), 2020. 2", + "[42] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In International Conference on Computer Vision (ICCV), 2019. 2", + "[43] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1", + "[44] X. Yan, Z. Chen, A. Xu, X. Wang, X. Liang, and L. Lin. Meta R-CNN: Towards general solver for instance-level low-shot learning. In International Conference on Computer Vision (ICCV), 2019. 2", + "[45] B. Yang, C. Liu, B. Li, J. Jiao, and Q. Ye. Prototype mixture models for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), pages 763-778, 2020. 1, 2", + "[46] X. Yang, B. Wang, K. Chen, X. Zhou, S. Yi, W. Ouyang, and L. Zhou. BriNet: Towards bridging the intra-class and inter-class gaps in one-shot segmentation. In British Machine Vision Conference (BMVC), 2020. 1, 2", + "[47] B. Zhang, J. Xiao, and T. Qin. Self-guided and cross-guided learning for few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8321, 2021. 7", + "[48] C. Zhang, G. Lin, F. Liu, J. Guo, Q. Wu, and R. Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In International Conference on Computer Vision (ICCV), pages 9587-9595, 2019, 1, 2", + "[49] C. Zhang, G. Lin, F. Liu, R. Yao, and C. Shen. CANet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217–5226, 2019. 1, 2, 7", + "[50] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 1, 3", + "[51] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In IEEE/CVF Conference on Computer" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23479", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6", + "[52] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2", + "[53] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representations (ICLR), 2021. 6" + ], + "bbox": [ + 78, + 90, + 468, + 232 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23480", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_model.json b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6216c5ec1799b0fed5e5d6fd5d1154a0a663c9fa --- /dev/null +++ b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.191, + 0.131, + 0.78, + 0.176 + ], + "angle": 0, + "content": "Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.203, + 0.353, + 0.221 + ], + "angle": 0, + "content": "Mir Rayat Imtiaz Hossain\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.221, + 0.378, + 0.239 + ], + "angle": 0, + "content": "1University of British Columbia" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.204, + 0.531, + 0.221 + ], + "angle": 0, + "content": "Mennatullah Siam\\(^{3,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.221, + 0.603, + 0.238 + ], + "angle": 0, + "content": "\\(^{2}\\)Vector Institute for AI" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.239, + 0.609, + 0.255 + ], + "angle": 0, + "content": "4Canada CIFAR AI Chair" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.204, + 0.705, + 0.222 + ], + "angle": 0, + "content": "Leonid Sigal\\(^{1,2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.746, + 0.204, + 0.868, + 0.221 + ], + "angle": 0, + "content": "James J. Little" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.222, + 0.848, + 0.24 + ], + "angle": 0, + "content": "3 Ontario Tech University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.732 + ], + "angle": 0, + "content": "The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base categories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mechanism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-20\\(^i\\) and Pascal5\\(^i\\), without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabeled test data can be used to improve the prompts, which we refer to as transductive prompt tuning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.869 + ], + "angle": 0, + "content": "The emergence of foundational models, trained on broad datasets, in natural language processing (e.g., GPT-3 [5]) and in vision-language (e.g., CLIP [31]), exhibits powerful generalization and strong performance on multiple downstream tasks. These models have been adapted through different prompting techniques [6, 43] for them to be used in the few-shot scenario." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "However, despite successes, including in localization [6, 43, 50], prompting, which allows for few-shot demon" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.291, + 0.865, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 1. Overview of the Proposed Approach. Prompting multi-scale transformer decoders for generalized few-shot segmentation. Our approach is a simple approach that allows test-time transductive prompt tuning (see red arrows)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.688 + ], + "angle": 0, + "content": "strations in dense prediction, and specifically in semantic segmentation tasks, is relatively under-explored." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Few-shot semantic segmentation [24, 40, 45, 46, 48, 49] aims to segment novel (unseen) classes with few labelled training examples. Most state-of-the-art methods rely on meta-learning, to leverage abundant training data as a form of \"data augmentation\" to construct many tasks that resemble test-time few-shot inference [26, 45]. In general, few-shot learning approaches can be classified as either inductive or transductive. Inductive approaches mainly rely on the training data, while transductive ones [4], in addition, exploit unlabelled test data in an unsupervised manner to improve the performance \\( e.g. \\), by leveraging entropy priors over predictive class distributions. However, both cases largely focus on the performance of the novel categories; which isn't particularly realistic. A recently proposed gen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "eralized few-shot segmentation (GFSS) setting defines a more realistic scenario where the goal is to perform well on all classes, novel and base [36]. This is considerably more challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.471, + 0.368 + ], + "angle": 0, + "content": "Prompting has been proven to be effective for learning from few demonstrations, as seen in LLMs [5]. It was also explored in visual prompt tuning [16] to efficiently fine-tune vision transformers for new tasks. We posit that prompting within a transformer-based [38] architecture would similarly offer an effective and flexible mechanism for GFSS. This involves learning prompts that can be utilized to perform cross-attention with input image to make predictions. While learning prompts for base classes, with abundant data, is relatively straightforward, it becomes considerably more challenging for novel classes with few examples. Specifically, one must ensure that novel prompts learned from few samples are sufficiently distinct from the base prompts to avoid novel-base class misclassification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.471, + 0.675 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we develop a simple yet highly effective visual prompting of transformer decoders for dense prediction at multiple scale that relies on novel-to-base causal attention without meta-training. We look upon queries in DETR-style architectures as a form of visual prompts and devise a mechanism to initialize and learn novel prompts. Then the novel-to-base causal attention allows base prompts to affect novel prompt representations, but not the other way around. This, intuitively, allows novel prompts to be repulsed and/or attracted towards their base counterparts. This attention is shared across scales (layers of the transformer) which, as we show, leads to more robust learning and improved performance. The multi-scale refinement of prompts helps interacting and reasoning between image features at multiple scales which in turn helps in better dense prediction. Finally, we extended this architecture in a transductive setting, where both novel and base prompts can be fine-tuned at test-time with respect to unsupervised objective to further improve performance. The proposed architecture and methodology is shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.678, + 0.327, + 0.692 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.696, + 0.469, + 0.771 + ], + "angle": 0, + "content": "- Designing a multi-scale visual prompting transformer decoder architecture for GFSS, featuring learnable prompts that allow the creation of new prompts for novel classes, initialized through masked average pooling of support images (with their masks)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.772, + 0.469, + 0.817 + ], + "angle": 0, + "content": "- Within this architecture, proposing and learning a multiscale (shared) novel-to-base cross-attention mechanism between the novel and base prompts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.818, + 0.469, + 0.862 + ], + "angle": 0, + "content": "- Proposing a novel transductive prompt tuning, which allows the visual prompts to be tuned on test (unlabeled) images, hence the term transductive." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.696, + 0.469, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.638, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.12, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Few-shot Learning. Few-shot learning, a longstanding concept in computer vision, serves as a paradigm to evaluate the data efficiency and adaptability of learning algorithms. It focuses on making predictions for novel classes with a few labeled examples, often by leveraging information from disjoint base classes for which there is abundant data. Few-shot learning has been extensively studied for image classification [12, 32-34, 39]. Most approaches can be categorized as either transfer or meta-learning-based. Transfer learning methods [11] typically pre-train the model on base classes with abundant data and then fine-tune this model for novel classes with few samples. Meta-learning [12, 32] approaches, on the other hand, simulate the inference stage during meta-training by sampling tasks of support and target sets. Meta-learning approaches can be sub-categorized into gradient-based and metric-learning based. Gradient-based approaches [12, 32] learn the commonalities among different tasks. Metric-based (or prototype-based) methods [34, 39] learn image embeddings and compare the distances between them. Despite its popularity and strong performance, meta-learning is found to be sensitive to the meta-training setup [7] and worse performance with domain shift [9] is observed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Generalized few-shot learning, introduced in [14] as a more realistic variant, allows query images from either base or novel categories. Recently, the transductive setting has been proven to be effective for few-shot learning [3, 4, 13, 29, 52]. Transductive approaches exploit information of unlabelled test images during inference to improve classification [11, 22, 27, 30]. These approaches build on classic semi-supervised learning approaches, such as graph-based label propagation [23] or entropy minimization [11]. Recent variants of few-shot learning problems have also been established in other more granular domains, e.g., detection [17, 41, 42, 44] and segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Few-shot segmentation. Less common than classification, few-shot segmentation is emerging as the dense counterpart to the aforementioned approaches. Few-shot segmentation approaches can be categorized to prototype-based and dense comparison-based methods. Prototype-based methods mostly relied on extracted prototypes from the support set to guide the segmentation of the target image, either as a single prototype [40, 49], part-aware [24] or a prototype mixture [45]. Dense comparison methods rely on dense comparisons between the few template images and the target image [46, 48] with extension to multiscale [26]. A recent approach proposed generalized few-shot segmentation (GFSS) task where the unlabelled target images are used to improve the segmentation of novel classes while maintaining the performance on base classes through knowledge distillation [13]. A concurrent work [21] on GFSS learns a set" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.402, + 0.9 + ], + "angle": 0, + "content": "1The term 'causal' [28] indicates its unidirectional nature." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "of orthogonal prototypes for base and novel categories, yet their GFSS setup is easier as they sample in their inference balanced base class images along with the novel ones, unlike the setup we follow [13]. Recent work [50] proposed a prompting strategy within a meta-learning framework using few examples. This approach is designed for 1-way segmentation only and performs prompting at a single scale unlike our proposed approach, a non-meta learning scheme, which performs multi-scale prompting and is designed for \\( k \\)-way GFSS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.518 + ], + "angle": 0, + "content": "Prompting. An emerging direction in learning with limited labelled data is to utilize foundation models [2] that have strong generalization capabilities. These models, pretrained on large-scale data, demonstrate strong generalization in few-shot settings [1]. Prompting is often used to adapt these foundation models to few-shot tasks, where it formulates the inference problem similarly to the training. Examples include auto-regressive modelling [1] where inference involves generating the next tokens after receiving the correct prompt. For few-shot learning, prompts include few-shot demonstrations followed by the query. Prompting can employ discrete prompts [1] or learning continuous prompts [20] through back-propagation. In computer vision, both vision-language prompting [1] and visual prompting [16] of transformer-based models have shown success in various tasks, particularly in few-shot scenarios. However, effectively prompting models designed for localization and segmentation remains an open question." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.473, + 0.822 + ], + "angle": 0, + "content": "Concurrent Work. A recent method [6] proposed prompting of transformer decoders for few-shot object detection. Although conceptually related, their approach is designed for a different task (detection) compared to ours (semantic segmentation), trained within a meta-learning framework unlike ours. Further, [6] lacks novel-to-base attention and multi-scale prompt refinement which are at the core of our architectural design. A recent foundation model for semantic segmentation, SAM [18], exhibits strong performance when prompted by point, bounding box, or text. However, it is not specifically designed for few-shot inference, potentially requiring fine-tuning with the support set and other non-trivial changes. More importantly, the use of extensive data to train the model makes it unsuitable for few-shot tasks since it has likely encountered many target classes during training. In addition to aforementioned differences, including multi-scale prompting and novel-to-base attention, our method is the first to explore transductive prompt-tuning where the visual prompts can be further optimized with the unlabelled target images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.285, + 0.849 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Classic Few-shot Segmentation. In the classic few-shot segmentation scenario, the categories are divided into two sets of disjoint sets: base and novel. The problem is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "framed as an \\( n \\)-way \\( k \\)-shot segmentation task, where \\( n \\) is the number of novel classes and \\( k \\) is the number of examples available for each class. A typical solution involves meta-learning with two-stage training. In the meta-training stage, a support set is created with \\( n \\) randomly selected base classes, and \\( k \\) examples from each class are used to train the model. In the meta-testing stage, the model is evaluated on tasks with \\( k \\) examples from \\( n \\) novel classes. An alternative approach is non-meta learning, where a model is initially trained on abundant data for base classes, then frozen and fine-tuned for novel classes using few examples. The focus is on segmenting novel classes in test images without considering base class segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Generalized Few-shot Segmentation (GFSS). Generalized few-shot segmentation, proposed by [36], extends the classic few-shot setting. In this new setup the model must preserve the segmentation performance on base classes in addition to generalizing to novel classes. While during finetuning the support set (typically) only contains \\( k \\)-shot examples for each novel category, the model is still expected to make predictions of both novel and base classes on the test images. We follow the recent more practical setup of generalized few-shot segmentation [13] within a non-meta learning framework, details provided in Section 5.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.593, + 0.476 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.895, + 0.713 + ], + "angle": 0, + "content": "We perform prompting using a set of learnable visual prompts, where, intuitively\\(^2\\), each prompt corresponds to a class embedding. These learned visual prompts cross-attend with image features at multiple scales. Initially, we pre-train a base model with abundant examples of base classes, learning visual prompts for them. We then fix these base prompts and fine-tune new visual prompts for novel classes with a few examples. However, learning prompts from limited data is challenging because it leads to confusion between base and novel categories. To improve novel class generalization, we introduce a uni-directional causal attention mechanism between novel and base prompts. Intuitively, this allows better contextualization and repulsion / attraction between base and novel prompts. Section 5 demonstrates the empirical effectiveness of this mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.895, + 0.865 + ], + "angle": 0, + "content": "Figure 2 shows our proposed architecture. It centers around learnable embeddings called visual prompts, which interact with image features at multiple scales. Our design is inspired by Mask2former [10] with a couple of key differences: (i) the number of visual prompts aligns with the number of classes, each representing a class as a learned embedding; (ii) we adopt per-pixel classification, instead of mask classification, thereby eliminating the query classification component of [10] and defining custom segmentation heads capable of making per-pixel predictions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.902 + ], + "angle": 0, + "content": "2This may not be precisely true in practice, but is a reasonable abstraction for intuitive understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.088, + 0.825, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Detailed architecture of our proposed visual prompting of multiscale transformer decoder. Our design initializes the novel visual prompts using the support set. This is followed by consecutive novel-to-base causal attention, \\(\\mathcal{CA}\\), and prompt-to-target features cross attention, \\(\\mathcal{C}\\), across scales. Note that causal attention uses shared weights across the scales and decoder layers. Our design allows for transductive fine-tuning of the visual prompts leveraging the unlabelled test image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.431, + 0.443, + 0.447 + ], + "angle": 0, + "content": "4.1. Visual Prompting Multiscale Transformers" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.454, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Base Visual Prompts. Given \\( B \\) base classes, we define \\( B \\) visual prompts each representing a class and initialize them randomly. These visual prompts first self-attend to learn the relationship between themselves and then are used to prompt the image features at multiple scales. The refinement of base prompts at each decoder layer is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.552, + 0.469, + 0.573 + ], + "angle": 0, + "content": "\\[\nV _ {B} ^ {(l)} = \\mathcal {A} \\left(V _ {B} ^ {(l - 1)}\\right) + \\mathcal {C} \\left(V _ {B} ^ {(l - 1)}, F ^ {(l - 1)}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.58, + 0.469, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\) and \\(\\mathcal{C}\\) refer to multi-headed self and cross attention respectively; \\(V_{B}^{(l)}\\in \\mathbb{R}^{B\\times C}\\) are the base visual prompts at layer \\(l\\); \\(F^{(l)}\\in \\mathbb{R}^{H_lW_l\\times C_l}\\) is the flattened image feature at layer \\(l\\); \\(H_{l}\\) and \\(W_{l}\\) denote image dimensions at that layer; \\(C_l\\) denotes the embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.656, + 0.469, + 0.748 + ], + "angle": 0, + "content": "After refining the prompts through multiple levels of transformer attention, a segmentation head, denoted as \\(\\mathcal{H}(\\cdot)\\), is defined. This segmentation head takes the final refined prompts, \\(V_{B}^{(L)}\\), and the highest resolution image features, \\(F^{(L)}\\), as input, producing per-pixel per-class predictions \\(O_{base} \\in \\mathbb{R}^{B \\times H \\times W}\\) (see Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Novel Visual Prompts. After training on the base classes, we freeze all the layers of the model except the visual prompts representing the base classes. Given there are \\( N \\) novel classes, we add \\( N \\) new visual prompts. We further add a causal unidirectional attention module that allows the novel prompts to be enhanced by the base prompts (see Section 4.2). The output is passed to a segmentation head to make predictions for the novel classes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.899 + ], + "angle": 0, + "content": "The novel visual prompts are first refined by a unidirectional novel-to-base causal attention module. The enhanced" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.432, + 0.89, + 0.463 + ], + "angle": 0, + "content": "novel prompts are then concatenated with the base prompts. Overall the concatenated prompts are refined by:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.48, + 0.726, + 0.501 + ], + "angle": 0, + "content": "\\[\nV _ {N} ^ {(l)} = \\mathcal {C A} \\left(V _ {B} ^ {(l)}, V _ {N} ^ {(l)}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.503, + 0.705, + 0.528 + ], + "angle": 0, + "content": "\\[\nV _ {A} ^ {(l)} = \\left[ V _ {B} ^ {(l)}, V _ {N} ^ {(l)} \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.53, + 0.842, + 0.551 + ], + "angle": 0, + "content": "\\[\nV _ {A} ^ {(l)} = \\mathcal {A} (V _ {A} ^ {(l - 1)}) + \\mathcal {C} (V _ {A} ^ {(l - 1)}, F ^ {(l - 1)}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.571, + 0.891, + 0.664 + ], + "angle": 0, + "content": "where \\(V_{N}^{(l)}\\in \\mathbb{R}^{N\\times C}\\) are novel prompts at layer \\(l\\); \\(N\\) is the total number of novel class and \\(C\\) is the embedding size; \\(V_{A}^{(l)}\\in \\mathbb{R}^{(B + N)\\times C}\\) are all visual prompts including base and novel; \\([.,.]\\) is the concatenation operator. The operation \\(\\mathcal{CA}\\) denotes the causal novel-to-base attention and is explained in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Each of the \\(N\\) novel visual prompts (representing a novel class) is initialized using masked average global pooling of the k-shot support images corresponding to the class masked by the ground truth binary masks." + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "\\[\nV _ {n} ^ {(0)} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\sum_ {x , y} M _ {n} ^ {k} (x , y) F _ {n} ^ {k} (x , y)}{\\sum_ {x , y} M _ {n} ^ {k} (x , y)}, \\forall n \\in N \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.823, + 0.892, + 0.899 + ], + "angle": 0, + "content": "where \\( V_{n}^{(0)} \\in \\mathbb{R}^{1 \\times C} \\) is the initial novel visual prompt for novel class \\( n \\); \\( F_{n}^{k} \\) denotes the image features of \\( k \\)-th shot for novel class; \\( M_{n}^{k} \\) denotes the corresponding ground-truth binary mask; \\( x, y \\) are the spatial locations; \\( N \\) is the total number of novel classes and \\( K \\) is the total number of shots." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23473" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.355, + 0.106 + ], + "angle": 0, + "content": "4.2. Novel-to-base Causal Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.325 + ], + "angle": 0, + "content": "A key component of our proposed approach is the unidirectional novel-to-base prompt causal3 attention. The base visual prompts are learned with abundant examples; this is not the case for novel prompts. Learning novel prompts directly or having novel prompts impact the base ones could therefore lead to degradation of performance. However, having novel prompts aware of their base counterparts could be quite helpful. We hypothesize that causal attention between the novel visual prompts and base prompts could help contextualize novel class embeddings, reducing confusion with the base counterparts. Motivated by this intuition, we propose a novel-to-base causal attention layer that is repeated at each scale (or layer of the transformer decoder); see Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.372 + ], + "angle": 0, + "content": "The proposed causal attention module is a single layer cross-attention where the Key \\( K^{(l)} \\), Value \\( V^{(l)} \\) and \\( Q^{(l)} \\) for layer \\( l \\) of the decoder are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.38, + 0.462, + 0.402 + ], + "angle": 0, + "content": "\\[\nQ ^ {(l)} = V _ {N} ^ {(l)} W ^ {Q}; \\quad K ^ {(l)} = V _ {B} ^ {(l)} W ^ {K}; \\quad V ^ {(l)} = V _ {B} ^ {(l)} W ^ {V}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.469, + 0.486 + ], + "angle": 0, + "content": "where \\( W^{Q}, W^{K}, W^{V} \\) are weight matrices for \\( Q, K \\) and \\( V \\) respectively. The parameters \\( W^{Q}, W^{K}, W^{V} \\) are shared at all the layers of this cross-attention module to reduce the number of trainable parameters during fine-tuning on novel classes to prevent overfitting and poor generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.531 + ], + "angle": 0, + "content": "This uni-directional novel-to-base causal attention refines the novel visual prompts \\( V_{N}^{(l)} \\) at each decoder layer \\( l \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.531, + 0.39, + 0.55 + ], + "angle": 0, + "content": "\\[\nV _ {N} ^ {(l)} = \\operatorname {s o f t m a x} (Q ^ {(l)} (K ^ {(l)}) ^ {\\mathrm {T}}) V ^ {(l)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.556, + 0.469, + 0.572 + ], + "angle": 0, + "content": "This results in a set of contextualized novel visual prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.266, + 0.607 + ], + "angle": 0, + "content": "4.3. Segmentation Head" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Base Segmentation Head. The base segmentation head, \\(\\mathcal{H}_B\\), during the base training in the proposed model is relatively simple and takes the form of a three layer MLP that projects a refined prompt into, effectively, a class prototype; computes the (dot product) similarity of each pixel feature with that prototype, and then applies softmax along the channel / class dimension to obtain per-pixel and per-class probability:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.743, + 0.469, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} O _ {b a s e} = \\mathcal {H} _ {B} (V _ {B} ^ {(L)}, F ^ {(L)}) \\\\ = \\operatorname {M L P} \\left(V _ {B} ^ {(L)}\\right) \\cdot \\left[ F ^ {(L)} \\right] ^ {\\mathrm {T}}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.469, + 0.836 + ], + "angle": 0, + "content": "where \\(O_{base} \\in \\mathbb{R}^{B \\times H \\times W}\\) and the class assignment can then be made by an argmax along the channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.837, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Novel Segmentation Head. While we could, in principle, use the same form of segmentation head during few-shot" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "3We use the term causal borrowing from its use in the look-ahead-attention-masking for decoder Transformers [38] (e.g., in language)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "inference as that defined for the base classes (see Eq. 3), simply fine-tuning it for novel classes, this proves ineffective in practice. Specifically, fine-tuning the MLP leads to significant overfitting given the few examples and lack of base class data during this step. So, instead, we resort to a simpler form of the novel segmentation head, \\(\\mathcal{H}_N\\), where we only learn a residual to the MLP learned during base training (which we keep fixed), mainly:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.218, + 0.858, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} O _ {n o v e l} = \\mathcal {H} _ {N} (V _ {N} ^ {(L)}, F ^ {(L)}) \\\\ = \\left[ \\operatorname {M L P} _ {\\bullet} \\left(V _ {N} ^ {(L)}\\right) + W _ {N} \\right] \\cdot \\left[ F ^ {(L)} \\right] ^ {\\mathrm {T}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.335 + ], + "angle": 0, + "content": "where \\(O_{\\text{novel}}\\) is output for novel classes and \\(\\bullet\\) designates that parameters of MLP are frozen and only \\(W_N \\in \\mathbb{R}^{N \\times C}\\) weight matrix is learned. This weight matrix is initialized in the same manner as novel prompts, in Eq. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.76, + 0.354 + ], + "angle": 0, + "content": "4.4. Transductive Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Recent approaches [4, 13] show the effectiveness of a good transductive setting for few-shot dense prediction tasks. Transductive approaches perform test-time optimization for each unlabeled test image and utilize specific characteristics and information of the test instances to improve prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.526 + ], + "angle": 0, + "content": "To demonstrate the flexibility of our approach, in this section we outline the process of adapting our model for transductive inference, improving GFSS performance through transductive fine-tuning of the visual prompts. We build on the transductive losses proposed in DlaM [13] and we describe them here for completeness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.708 + ], + "angle": 0, + "content": "The losses are designed to maximize the mutual information between the learned features and the corresponding predictions on the test image, which is achieved through maximizing: \\(\\mathrm{H}(O) - \\mathrm{H}(O|I)\\), where \\(I, O\\) are random variables associated with the pixels and predictions distributions respectively; \\(\\mathrm{H}(O)\\) is the marginal and \\(\\mathrm{H}(O|I)\\) is the conditional entropy. The conditional entropy \\(\\mathrm{H}(O|I)\\) involved is given by sum of cross-entropy loss over the supervised support set and entropy of the predicted probability given the test image \\(I\\). The marginal entropy is given by KL divergence loss relying on an estimated region proportion prior following [4]. Additional details are in [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Additionally, similar to [13], we use knowledge-distillation loss to preserve base class performance, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.751, + 0.892, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D}} = \\mathrm {K L} \\left(O _ {\\text {b a s e}} ^ {\\text {n e w}} \\mid \\mid O _ {\\text {b a s e}} ^ {\\text {o l d}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.863 + ], + "angle": 0, + "content": "where \\( O_{base}^{old} \\) is the prediction probability of the base classes of the frozen model after base training; \\( O_{base}^{new} \\) is the prediction probability of the base classes on the model with our additional components to handle novel categories that is being fine-tuned. These include the multi-scale causal novel-to-base cross attention and the novel prompts." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.864, + 0.882, + 0.879 + ], + "angle": 0, + "content": "The overall transductive objective function is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a n s .}} = \\alpha \\mathrm {H} (O | I) - \\mathrm {H} (O) + \\gamma \\mathcal {L} _ {\\mathrm {K D}}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23474" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.393, + 0.106 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\gamma\\) are balancing hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.213 + ], + "angle": 0, + "content": "We noticed that applying transduction from the first iteration underperforms, likely due to a poor initial estimate of the marginal distribution. To address this, we only apply per-pixel cross-entropy loss for a set number of iterations and then incorporate the remaining transductive losses from Eq. 5 when we achieve a more accurate estimate of the marginal distribution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.217, + 0.21, + 0.233 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.241, + 0.39, + 0.258 + ], + "angle": 0, + "content": "5.1. Dataset and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.469, + 0.354 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on two few-shot segmentation benchmarks [36]: COCO- \\(20^{i}\\) and PASCAL- \\(5^{i}\\). Following [13] we report average performance of our model on 10K test images for COCO- \\(20^{i}\\) and on all test images on PASCAL- \\(5^{i}\\). The results are reported as an average across five different runs on four different splits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.469, + 0.506 + ], + "angle": 0, + "content": "Data Pre-processing. We adopt the same setup as DIAM [13]. During base training, images containing both base and novel categories are kept, but the pixels corresponding to novel classes are relabeled as background. This approach, as noted by [13], presents challenges due to potential ambiguity when the base model predicts novel categories as background. In fine-tuning, only novel classes are labeled in the support set, with base classes marked as background. To avoid unfair shot increases, images with multiple novel categories are excluded from the support set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.611 + ], + "angle": 0, + "content": "Evaluation Protocol. We evaluate performance using the standard mean intersection-over-union (mIoU) metric. Following [13], the mean score for generalized few-shot segmentation is computed as the average between base and novel mIoUs. This approach addresses bias in datasets like PASCAL- \\(5^{i}\\) and COCO- \\(20^{i}\\), where base classes outnumber novel classes threefold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Implementation Details. We use PSPNet [51] with Resnet-50 [15] as feature extractor. We utilize MSDeformAttn-based pixel decoder [53] from Mask2Former [10]. The dimension \\(C\\) for visual prompts is chosen to be 256. Contrary to [10], the base model is trained with per-pixel cross-entropy loss. We use AdamW optimizer [25] with base learning rate of \\(1 \\times 10^{-4}\\), a weight decay of 0.05 and a batch size of 16. Following [13, 19], we train base model for 20 epochs for COCO- \\(20^{i}\\) and 100 epochs for PASCAL- \\(5^{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "During novel fine-tuning, we freeze the parameters of the feature extractor, pixel decoder, and base segmentation head, and only optimize the parameters of the visual prompts, novel segmentation head, and novel-to-base causal attention module. For the inductive case, the model is finetuned with per-pixel cross-entropy loss for 100 iterations. For the transductive setting, the model is fine-tuned for 40 iterations with only cross-entropy loss and then fine-tuned for 60 more iterations with full transductive losses mentioned in Eq. 5. The hyper-parameters \\(\\alpha\\) and \\(\\gamma\\) are set to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "be 100 and 25 respectively. We use AdamW optimizer with a learning rate of \\(5 \\times 10^{-3}\\) and a weight decay of 0.05 for fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.148, + 0.593, + 0.162 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Comparison with the State-of-The-Art. Table 1 compares our approach in both inductive and transductive setting against the state-of-the-art. As observed, our model outperforms existing approaches. Interestingly, our approach, in an inductive setting (where optimization is performed only using a supervised few-shot support set, and evaluation is conducted on all test images), outperforms transductive approaches like DIaM [13] for both 1-shot and 5-shot cases on the COCO-20\\(^i\\) dataset and for the 1-shot case on Pascal-5\\(^i\\). This is notable despite transductive approaches requiring test-time optimization for each unlabeled test image separately, leading to significantly longer inference times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.485 + ], + "angle": 0, + "content": "For a fair comparison in the inductive setting, we finetuned DIaM without the transduction losses. We also finetuned POP [21] in our inductive setup. POP [21] randomly samples balanced base class image, making it easier for them to retain base class information and uses a form of transduction where their model relabels novel class pixels initially labelled as background. We observe that the performance of POP [21] degrades substantially, especially for the base classes in our inductive setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.666 + ], + "angle": 0, + "content": "Our inductive approach has \\(5.78\\%\\) higher mIoU for 1-shot and \\(11.16\\%\\) higher mIoU for 5-shot on COCO-20\\(^i\\) than inductive DIaM. For PASCAL-5\\(^i\\), the improvement from inductive DIaM is \\(7.71\\%\\) for 1-shot and \\(13.29\\%\\) for 5-shot. In fact, compared to the best inductive GFSS approach BAM [19], our inductive setting achieves an improvement of approximately \\(2.78\\%\\) and \\(7.59\\%\\) for 1-shot and 5-shot respectively on COCO-20\\(^i\\) and \\(5.24\\%\\) and \\(12.32\\%\\) for 1-shot and 5-shot respectively on PASCAL-5\\(^i\\). Our method, in the inductive setting, even outperforms transductive DIaM [13] by \\(2.03\\%\\) and \\(2.28\\%\\) for 1-shot and 5-shot respectively on COCO-20\\(^i\\) and by \\(1.79\\%\\) for 1-shot on PASCAL-5\\(^i\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.772 + ], + "angle": 0, + "content": "On applying transductive fine-tuning to our method, overall mIoU increases considerably, outperforming transductive DIaM [13] and POP [21] on both 1-shot and 5-shot inference for both datasets. The improvement over DIaM for \\(\\mathrm{COCO - }20^{i}\\) is \\(3.3\\%\\) for 1-shot and \\(3.93\\%\\) for 5-shot. For PASCAL- \\(5^{i}\\), our transductive setting outperforms DIaM by \\(5.11\\%\\) and \\(3.15\\%\\) for 1 and 5-shots." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.603, + 0.799 + ], + "angle": 0, + "content": "5.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.891, + 0.839 + ], + "angle": 0, + "content": "We performed ablation studies (refer to Table 2) to explain our design choices and contribution of each components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual Prompts Initialization. As seen in (1) and (2) of the Table 2, initializing novel prompts with masked mean pooling from the support set images improves both base and novel class accuracy compared to random initialization." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23475" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.198, + 0.09, + 0.773, + 0.263 + ], + "angle": 0, + "content": "
MethodLearningCOCO-20i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CAPL [36] (CVPR22)Inductive43.217.2125.2143.7111.0027.36
BAM [19] (CVPR22)Inductive49.8414.1632.0049.8516.6333.24
DIaM (w/o trans.)* [13] (CVPR23)Inductive42.6915.3229.0038.4720.8729.67
POP† [21] (CVPR23)Inductive30.389.6320.0024.5316.1920.36
Ours (w/o trans.)Inductive51.5518.0034.7851.5930.0640.83
RePRI [4] (CVPR21)Transductive5.624.745.188.858.848.85
DIaM [13] (CVPR23)Transductive48.2817.2232.7548.3728.7338.55
POP† [21] (CVPR23)Transductive54.7115.3135.0154.9029.9742.44
Ours (w/ trans.)Transductive53.8018.3036.0553.8131.1442.48
" + }, + { + "type": "table", + "bbox": [ + 0.199, + 0.268, + 0.773, + 0.489 + ], + "angle": 0, + "content": "
MethodLearningPASCAL-5i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CANet [49] (CVPR19)Inductive8.732.425.589.051.525.29
PFENET [35] (TPAMI20)Inductive8.322.675.508.831.895.36
PANET [40] (ICCV19)Inductive31.8811.2521.5732.9515.2524.1
SCL [47] (CVPR21)Inductive8.882.445.669.111.835.47
MiB [8] (CVPR20)Inductive63.808.8636.3368.6028.9348.77
CAPL [36] (CVPR22)Inductive64.8017.4641.1365.4324.4344.93
BAM [19] (CVPR22)Inductive71.6027.4949.5571.6028.9650.28
DIaM (w/o trans.)* [13] (CVPR23)Inductive66.7927.3647.0864.0534.5649.31
POP† [21] (CVPR23)Inductive46.6819.9633.3241.5036.2638.80
Ours (w/o trans.)Inductive74.5834.9954.7974.8650.3462.60
RePRI [4] (CVPR21)Transductive20.7610.5015.6334.0620.9827.52
DIaM [13] (CVPR23)Transductive70.8935.1153.0070.8555.3163.08
POP† [21] (CVPR23)Transductive73.9235.5154.7274.7855.8765.33
Ours (w/ trans.)Transductive76.3939.8358.1176.4256.1266.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.5, + 0.896, + 0.571 + ], + "angle": 0, + "content": "Table 1. Comparison to SoTA methods on COCO-20\\(^i\\) and PASCAL-5\\(^i\\). The mIoU results are reported as the average across 5 different runs. All methods use ResNet-50 backbone. \\(\\dagger\\) POP [21] uses an easier and less practical GFSS setup, e.g., balanced base class training in addition to support set during novel fine-tuning and uses a form of transduction; inductive setting is implemented by us following our setup. Results of the rest of the methods are obtained from [13] except the inductive setting of DIA\\({M}^{*}\\), which we implemented ourselves. The w/o trans., w/ trans. denote without and with transduction respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.584, + 0.47, + 0.676 + ], + "angle": 0, + "content": "
Causal AttentionPrompt InitializationTransductionmIoU
BaseNovelMean
(1)NoneRandomNo50.1811.2230.70
(2)NoneMasked PoolingNo50.4311.3130.87
(3)First Layer OnlyMasked PoolingNo51.5312.2631.90
(4)Separate WeightsMasked PoolingNo51.0618.0534.56
(5)Shared WeightsRandomNo50.7517.4134.08
(6)Shared WeightsMasked PoolingNo51.5518.0034.78
(7)Shared WeightsMasked PoolingYes53.8018.3036.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Table 2. Ablation Study on COCO- \\( {20}^{i} \\) . The 1-shot performane, averaged over all folds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.763 + ], + "angle": 0, + "content": "This is more pronounced with novel-to-base causal attention (5)-(6), amplifying the performance gain from masked pooling initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Novel-to-base Causal Attention. As seen in (4)-(7) of Table 2, novel-to-base causal attention yields the most significant improvement in novel class accuracy compared to (1)-(2). This aligns with our hypothesis that uni-directional causal attention between rich base prompts (learned with many examples) and novel prompts (learned with few examples) increases separation among novel classes and pushes the feature space of base prompts away from the novel prompts. It is further confirmed by Figure 3 and Fig-" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.584, + 0.91, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.775, + 0.895, + 0.832 + ], + "angle": 0, + "content": "Figure 3. Confusion Matrix for: (left) model with Novel-to-Base Causal Attention; (right) model without Novel-to-Base Causal Attention. Note that Novel-to-Base Causal Attention reduces confusion between novel and base categories (bottom-left block)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ure 4. As observed in Figure 3, the novel-to-base causal attention leads to lower confusion between base and novel classes. Figure 4 shows a t-SNE [37] visualization of the features of base and novel visual prompts for 10 different" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23476" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.093, + 0.487, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.471, + 0.337 + ], + "angle": 0, + "content": "Figure 4. TSNE visualization of the learned base and novel prompt features: (left) with Novel-to-Base Causal Attention; (right) without Novel-to-Base Casual Attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.401 + ], + "angle": 0, + "content": "test images. As observed, the incorporation of novel-to-base causal attention leads to much clearer separation of visual prompt features, particularly for novel prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Multiscale Causal Attention. We also ablated the placement of the causal attention module. We observe that applying it only at the first layer (3) slightly improves novel mIoU compared to the model without causal attention. Yet, it is notably inferior to applying at every decoder layer (4)-(7). We could have opted for separate parameters for the causal attention module at each layer. However, as seen in (4), having separate weights offers little improvement to novel mIoU despite significant increase in total number of parameters to optimize. In fact, it decreases the base accuracy, leading to overall worse performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. Figure 5 compares the qualitative results of our model for 1-shot GFSS on Pascal- \\(5^{i}\\) in both inductive and transductive settings with the baseline lacking causal attention and DIaM [13]. Notably, for first row, both our inductive and transductive approaches predict more accurate segmentation of the cow compared to DIaM. In second row, the causal attention improves person segmentation followed by transduction. In the third row, DIaM completely misses the novel category chair. The causal attention allows our model to better differentiate the base class dining table from the novel class chairs. Transduction improves accuracy of the segmentation of dining table. In fourth row, we can observe that both DIaM and our model without causal attention misdetects novel class cat for cow at multiple locations. Causal attention allows our model to make a uniform and more accurate prediction of cat. In the fifth row, we observe that DIaM [13] completely misses detecting the base category potted plant and poorly segments person. The baseline without causal attention misdetects the potted plant with novel class bus and misdetects dining table with chair. Causal attention improves the accuracy which is further improved by transduction. Last row" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.911, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.366, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results for 1-shot on Pascal- \\(5^{i}\\). The left-most two columns show image and ground truth mask; (Third) Baseline without causal attention; (Fourth); Ours in inductive setting; (Fifth) Ours in transductive setting; (Last) DIaM [13]. Last row illustrates a failure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.514 + ], + "angle": 0, + "content": "demonstrates a case where our model performs worse than DlaM [13]. Again, we observe that adding causal attention reduces misclassification of person with car, although overall the segmentation of car is less accurate than DlaM [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.589 + ], + "angle": 0, + "content": "From the qualitative results, the effectiveness of novel-to-base causal attention at distinguishing base from novel classes becomes apparent. Our model in the inductive setting outperforms transductive DIaM in most cases, despite the inference time being faster." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.619, + 0.609 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this work, we proposed a technique to learn visual prompts from few-shot examples of novel unseen cateogries to prompt a pre-trained segmentation model that not only generalizes well on the unseen novel class but strongly retains performance over the base classes on which the model was pre-trained. We achieve this by prompting the image features at multiple scales and proposing a uni-directional novel-to-base causal attention mechanism that enriches the novel visual prompts through interaction with the base visual prompts. Our proposed method can be applied in both inductive and transductive settings and achieves state-of-the-art performance on two benchmark few-shot segmentation datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.815, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments and Disclosure of Funding. This work was partially funded by the Vector Institute for AI, Canada CIFAR AI Chair, NSERC CRC and NSERC DG. Hardware resources were partially provided by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute. Additional support was provided by JELF CFI grant and Digital Research Alliance of Canada under the RAC award." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23477" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.239 + ], + "angle": 0, + "content": "[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, E. Rutherford, S. Cabi, T. Han, Z. Gong, S. Samangooei, M. Monteiro, J. L. Menick, S. Borgeaud, A. Brock, A. Nematzadeh, S. Sharifzadeh, M. Binkowski, R. Barreira, O. Vinyals, A. Zisserman, and K. Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[2] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S.v.ArxM.S.BernsteinA.Bosselut J.BohgE.Brunskill, E. Brynjolfsson, S. Buch, D.Card, R.Castellon, N.Chatterji, A. Chen, K.Creel, J.Q.Davis,D.Demszky,C.Donahue,M.Doumbouya,S.Ermon E.Durmus,J.Etchemendy, K.Ethayarajh, L.Fei-Fei, C.FinnT.GaleL.GillespieK. Goel, N. Goodman, S. Grossman, N. Guha, T. Hashimoto, P. Henderson, J. Hewitt, D.E.HoJ.HongK.HsuJ.Huang, T.IcardS.JainD.JurafskyP.KalluriS.KaramchetiG. KeelingF.KhaniO.KhattabP.W.KohM.KrassR. Krishna,R.Kuditipudi,A.Kumar,F.Ladhak,M.Lee,T. LeeJ.LeskovecI.LeventX.L.LiX.LiT.MaA.MlikC.D.ManningS.Mirchandani,E.MitchellZ.MunyikwaS.NairD.Narayanan A.NarayanB.NewmanA. NieJ.C.Niebles,H.NilforoshanJ.Nyarko,G.OgutI. PapadimitriouJ.S.ParkC.PiechE.PortelanceC.Potts, A.RaghunathanR.ReichH.RenF.RongY.RoohaniC. RuizJ.RyanC.ReD.SadighS.SagawaK.Santhanam, A.ShihK.SrinivasanA.TamkinR.TaoriA.W.Thomas F.TramereR.E.WangW.WangB.WuJ.WuY.WuS.M XieM.YasunagaJ.YouM.ZahariaM.ZhangT.Zhang X.ZhangY.ZhangL.ZhengK.Zhou and P.Liang.On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.576, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[3] Malik Boudiaf, Imtiaz Ziko, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Information maximization for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 33:2445-2457, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.633, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[4] M. Boudiaf, H. Kervadec, Z. Masud, P. Piantanida, I. Ben Ayed, and J. Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13979-13988, 2021. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[5] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, G. Sastry P. Shyam, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33: 1877-1901, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.83, + 0.47, + 0.886 + ], + "angle": 0, + "content": "[6] A. Bulat, R. Guerrero, B. Martinez, and G. Tzimiropoulos. Fs-detr: Few-shot detection transformer with prompting and without re-training. In International Conference on Computer Vision (ICCV), pages 11793-11802, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.887, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[7] T. Cao, M. Law, and S. Fidler. A theoretical analysis of the" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "number of shots in few-shot learning. International Conference on Learning Representations (ICLR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[8] F. Cermelli, M. Mancini, S.R. Bulo, E. Ricci, and B. Caputo. Modeling the background for incremental learning in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233–9242, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[9] W. Chen, Y. Liu, Z. Kira, Y. Frank Wang, and J. Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[10] B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[11] G. S. Dhillon, P. Chaudhari, A. Ravichandran, and S. Soatto. A baseline for few-shot image classification. In International Conference on Learning Representations (ICLR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[12] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning (ICML), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[13] S. Hajimiri, M. Boudiaf, I. B. Ayed, and J. Dolz. A strong baseline for generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11269-11278, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[14] B. Hariharan and R. B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[15] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[16] M. Jia, L. Tang, B. Chen, C. Cardie, S. Belongie, B. Hariharan, and S. Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[17] B. Kang, Z. Liu, X. Wang, F. Yu, J. Feng, and T. Darrell. Few-shot object detection via feature reweighting. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[18] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, A. C. Berg S. Whitehead, W. Lo, P. Dollár, and R. Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[19] C. Lang, G. Cheng, B. Tu, and J. Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8067, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[20] X. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[21] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11319-11328, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[22] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[23] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. J. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[24] Y. Liu, X. Zhang, S. Zhang, and X. He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[25] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[26] J. Min, D. Kang, and M. Cho. Hypercorrelation squeeze for few-shot segmentation. In International Conference on Computer Vision (ICCV), pages 6941-6952, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[27] A. Nichol, J. Achiam, and J. Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[28] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[29] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8412-8422, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.506, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[30] L. Qiao, Y. Shi, J. Li, Y. Wang, T. Huang, and Y. Tian. Transductive episodic-wise adaptive metric for few-shot learning. In International Conference on Computer Vision (ICCV), pages 3603-3612, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.563, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[31] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[32] S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[33] M. Ren, E. Triantafillou, S. Ravi, J.Snell, K. Swersky, J. B. Tenenbaum, H. Larochelle, and R. S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[34] J. Snell, K. Swersky, and R. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[35] Z. Tian, H. Zhao, M. Shu, Z. Yang, R. Li, and J. Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(2):1050-1065, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Z. Tian, X. Lai, L. Jiang, S. Liu, M. Shu, H. Zhao, and J. Jia. Generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11563-11572, 2022. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[38] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[39] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 3630-3638, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[40] K. Wang, J. Liew, Y. Zou, D. Zhou, and J. Feng. PANet: Few-shot image semantic segmentation with prototype alignment. In International Conference on Computer Vision (ICCV), pages 9197-9206, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[41] X. Wang, T. E Huang, T. Darrell, J. E. Gonzalez, and F. Yu. Frustratingly simple few-shot object detection. In International Conference on Machine Learning (ICML), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[42] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.387, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[43] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[44] X. Yan, Z. Chen, A. Xu, X. Wang, X. Liang, and L. Lin. Meta R-CNN: Towards general solver for instance-level low-shot learning. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[45] B. Yang, C. Liu, B. Li, J. Jiao, and Q. Ye. Prototype mixture models for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), pages 763-778, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[46] X. Yang, B. Wang, K. Chen, X. Zhou, S. Yi, W. Ouyang, and L. Zhou. BriNet: Towards bridging the intra-class and inter-class gaps in one-shot segmentation. In British Machine Vision Conference (BMVC), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[47] B. Zhang, J. Xiao, and T. Qin. Self-guided and cross-guided learning for few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8321, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[48] C. Zhang, G. Lin, F. Liu, J. Guo, Q. Wu, and R. Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In International Conference on Computer Vision (ICCV), pages 9587-9595, 2019, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[49] C. Zhang, G. Lin, F. Liu, R. Yao, and C. Shen. CANet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217–5226, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[50] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[51] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23479" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[52] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[53] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representations (ICLR), 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23480" + } + ] +] \ No newline at end of file diff --git a/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_origin.pdf b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7bb361497770044509fc5328b18e85b1486fb5d --- /dev/null +++ b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/d4543bc4-aa84-4173-9885-2737fde4ca07_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbf4564bf095e82c0952fece2c7d043e2f1281631a12c3e121d963221b552566 +size 5776351 diff --git a/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/full.md b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c51551c008807d498143918df6d500aaba3ac11 --- /dev/null +++ b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/full.md @@ -0,0 +1,316 @@ +# Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach + +Mir Rayat Imtiaz Hossain $^{1,2}$ + +1University of British Columbia + +Mennatullah Siam $^{3,1}$ + +$^{2}$ Vector Institute for AI + +4Canada CIFAR AI Chair + +Leonid Sigal $^{1,2,4}$ + +James J. Little + +3 Ontario Tech University + +# Abstract + +The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base categories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mechanism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-20 $^i$ and Pascal5 $^i$ , without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabeled test data can be used to improve the prompts, which we refer to as transductive prompt tuning. + +# 1. Introduction + +The emergence of foundational models, trained on broad datasets, in natural language processing (e.g., GPT-3 [5]) and in vision-language (e.g., CLIP [31]), exhibits powerful generalization and strong performance on multiple downstream tasks. These models have been adapted through different prompting techniques [6, 43] for them to be used in the few-shot scenario. + +However, despite successes, including in localization [6, 43, 50], prompting, which allows for few-shot demon + +![](images/cdbde3ca844e54cf00bd6dbaf756c69e556440150096a7dce4fe7910aa402885.jpg) +Figure 1. Overview of the Proposed Approach. Prompting multi-scale transformer decoders for generalized few-shot segmentation. Our approach is a simple approach that allows test-time transductive prompt tuning (see red arrows). + +strations in dense prediction, and specifically in semantic segmentation tasks, is relatively under-explored. + +Few-shot semantic segmentation [24, 40, 45, 46, 48, 49] aims to segment novel (unseen) classes with few labelled training examples. Most state-of-the-art methods rely on meta-learning, to leverage abundant training data as a form of "data augmentation" to construct many tasks that resemble test-time few-shot inference [26, 45]. In general, few-shot learning approaches can be classified as either inductive or transductive. Inductive approaches mainly rely on the training data, while transductive ones [4], in addition, exploit unlabelled test data in an unsupervised manner to improve the performance $e.g.$ , by leveraging entropy priors over predictive class distributions. However, both cases largely focus on the performance of the novel categories; which isn't particularly realistic. A recently proposed gen + +eralized few-shot segmentation (GFSS) setting defines a more realistic scenario where the goal is to perform well on all classes, novel and base [36]. This is considerably more challenging. + +Prompting has been proven to be effective for learning from few demonstrations, as seen in LLMs [5]. It was also explored in visual prompt tuning [16] to efficiently fine-tune vision transformers for new tasks. We posit that prompting within a transformer-based [38] architecture would similarly offer an effective and flexible mechanism for GFSS. This involves learning prompts that can be utilized to perform cross-attention with input image to make predictions. While learning prompts for base classes, with abundant data, is relatively straightforward, it becomes considerably more challenging for novel classes with few examples. Specifically, one must ensure that novel prompts learned from few samples are sufficiently distinct from the base prompts to avoid novel-base class misclassification. + +To address the aforementioned challenges, we develop a simple yet highly effective visual prompting of transformer decoders for dense prediction at multiple scale that relies on novel-to-base causal attention without meta-training. We look upon queries in DETR-style architectures as a form of visual prompts and devise a mechanism to initialize and learn novel prompts. Then the novel-to-base causal attention allows base prompts to affect novel prompt representations, but not the other way around. This, intuitively, allows novel prompts to be repulsed and/or attracted towards their base counterparts. This attention is shared across scales (layers of the transformer) which, as we show, leads to more robust learning and improved performance. The multi-scale refinement of prompts helps interacting and reasoning between image features at multiple scales which in turn helps in better dense prediction. Finally, we extended this architecture in a transductive setting, where both novel and base prompts can be fine-tuned at test-time with respect to unsupervised objective to further improve performance. The proposed architecture and methodology is shown in Fig. 1. + +In summary, our contributions are: + +- Designing a multi-scale visual prompting transformer decoder architecture for GFSS, featuring learnable prompts that allow the creation of new prompts for novel classes, initialized through masked average pooling of support images (with their masks). +- Within this architecture, proposing and learning a multiscale (shared) novel-to-base cross-attention mechanism between the novel and base prompts. +- Proposing a novel transductive prompt tuning, which allows the visual prompts to be tuned on test (unlabeled) images, hence the term transductive. + +# 2. Related work + +Few-shot Learning. Few-shot learning, a longstanding concept in computer vision, serves as a paradigm to evaluate the data efficiency and adaptability of learning algorithms. It focuses on making predictions for novel classes with a few labeled examples, often by leveraging information from disjoint base classes for which there is abundant data. Few-shot learning has been extensively studied for image classification [12, 32-34, 39]. Most approaches can be categorized as either transfer or meta-learning-based. Transfer learning methods [11] typically pre-train the model on base classes with abundant data and then fine-tune this model for novel classes with few samples. Meta-learning [12, 32] approaches, on the other hand, simulate the inference stage during meta-training by sampling tasks of support and target sets. Meta-learning approaches can be sub-categorized into gradient-based and metric-learning based. Gradient-based approaches [12, 32] learn the commonalities among different tasks. Metric-based (or prototype-based) methods [34, 39] learn image embeddings and compare the distances between them. Despite its popularity and strong performance, meta-learning is found to be sensitive to the meta-training setup [7] and worse performance with domain shift [9] is observed. + +Generalized few-shot learning, introduced in [14] as a more realistic variant, allows query images from either base or novel categories. Recently, the transductive setting has been proven to be effective for few-shot learning [3, 4, 13, 29, 52]. Transductive approaches exploit information of unlabelled test images during inference to improve classification [11, 22, 27, 30]. These approaches build on classic semi-supervised learning approaches, such as graph-based label propagation [23] or entropy minimization [11]. Recent variants of few-shot learning problems have also been established in other more granular domains, e.g., detection [17, 41, 42, 44] and segmentation. + +Few-shot segmentation. Less common than classification, few-shot segmentation is emerging as the dense counterpart to the aforementioned approaches. Few-shot segmentation approaches can be categorized to prototype-based and dense comparison-based methods. Prototype-based methods mostly relied on extracted prototypes from the support set to guide the segmentation of the target image, either as a single prototype [40, 49], part-aware [24] or a prototype mixture [45]. Dense comparison methods rely on dense comparisons between the few template images and the target image [46, 48] with extension to multiscale [26]. A recent approach proposed generalized few-shot segmentation (GFSS) task where the unlabelled target images are used to improve the segmentation of novel classes while maintaining the performance on base classes through knowledge distillation [13]. A concurrent work [21] on GFSS learns a set + +of orthogonal prototypes for base and novel categories, yet their GFSS setup is easier as they sample in their inference balanced base class images along with the novel ones, unlike the setup we follow [13]. Recent work [50] proposed a prompting strategy within a meta-learning framework using few examples. This approach is designed for 1-way segmentation only and performs prompting at a single scale unlike our proposed approach, a non-meta learning scheme, which performs multi-scale prompting and is designed for $k$ -way GFSS. + +Prompting. An emerging direction in learning with limited labelled data is to utilize foundation models [2] that have strong generalization capabilities. These models, pretrained on large-scale data, demonstrate strong generalization in few-shot settings [1]. Prompting is often used to adapt these foundation models to few-shot tasks, where it formulates the inference problem similarly to the training. Examples include auto-regressive modelling [1] where inference involves generating the next tokens after receiving the correct prompt. For few-shot learning, prompts include few-shot demonstrations followed by the query. Prompting can employ discrete prompts [1] or learning continuous prompts [20] through back-propagation. In computer vision, both vision-language prompting [1] and visual prompting [16] of transformer-based models have shown success in various tasks, particularly in few-shot scenarios. However, effectively prompting models designed for localization and segmentation remains an open question. + +Concurrent Work. A recent method [6] proposed prompting of transformer decoders for few-shot object detection. Although conceptually related, their approach is designed for a different task (detection) compared to ours (semantic segmentation), trained within a meta-learning framework unlike ours. Further, [6] lacks novel-to-base attention and multi-scale prompt refinement which are at the core of our architectural design. A recent foundation model for semantic segmentation, SAM [18], exhibits strong performance when prompted by point, bounding box, or text. However, it is not specifically designed for few-shot inference, potentially requiring fine-tuning with the support set and other non-trivial changes. More importantly, the use of extensive data to train the model makes it unsuitable for few-shot tasks since it has likely encountered many target classes during training. In addition to aforementioned differences, including multi-scale prompting and novel-to-base attention, our method is the first to explore transductive prompt-tuning where the visual prompts can be further optimized with the unlabelled target images. + +# 3. Problem Formulation + +Classic Few-shot Segmentation. In the classic few-shot segmentation scenario, the categories are divided into two sets of disjoint sets: base and novel. The problem is + +framed as an $n$ -way $k$ -shot segmentation task, where $n$ is the number of novel classes and $k$ is the number of examples available for each class. A typical solution involves meta-learning with two-stage training. In the meta-training stage, a support set is created with $n$ randomly selected base classes, and $k$ examples from each class are used to train the model. In the meta-testing stage, the model is evaluated on tasks with $k$ examples from $n$ novel classes. An alternative approach is non-meta learning, where a model is initially trained on abundant data for base classes, then frozen and fine-tuned for novel classes using few examples. The focus is on segmenting novel classes in test images without considering base class segmentation. + +Generalized Few-shot Segmentation (GFSS). Generalized few-shot segmentation, proposed by [36], extends the classic few-shot setting. In this new setup the model must preserve the segmentation performance on base classes in addition to generalizing to novel classes. While during finetuning the support set (typically) only contains $k$ -shot examples for each novel category, the model is still expected to make predictions of both novel and base classes on the test images. We follow the recent more practical setup of generalized few-shot segmentation [13] within a non-meta learning framework, details provided in Section 5.1. + +# 4. Method + +We perform prompting using a set of learnable visual prompts, where, intuitively $^2$ , each prompt corresponds to a class embedding. These learned visual prompts cross-attend with image features at multiple scales. Initially, we pre-train a base model with abundant examples of base classes, learning visual prompts for them. We then fix these base prompts and fine-tune new visual prompts for novel classes with a few examples. However, learning prompts from limited data is challenging because it leads to confusion between base and novel categories. To improve novel class generalization, we introduce a uni-directional causal attention mechanism between novel and base prompts. Intuitively, this allows better contextualization and repulsion / attraction between base and novel prompts. Section 5 demonstrates the empirical effectiveness of this mechanism. + +Figure 2 shows our proposed architecture. It centers around learnable embeddings called visual prompts, which interact with image features at multiple scales. Our design is inspired by Mask2former [10] with a couple of key differences: (i) the number of visual prompts aligns with the number of classes, each representing a class as a learned embedding; (ii) we adopt per-pixel classification, instead of mask classification, thereby eliminating the query classification component of [10] and defining custom segmentation heads capable of making per-pixel predictions. + +![](images/c837ac7f32204fe304d6787618d26b719d8b773a7770dce2790e294b53ff944f.jpg) +Figure 2. Detailed architecture of our proposed visual prompting of multiscale transformer decoder. Our design initializes the novel visual prompts using the support set. This is followed by consecutive novel-to-base causal attention, $\mathcal{CA}$ , and prompt-to-target features cross attention, $\mathcal{C}$ , across scales. Note that causal attention uses shared weights across the scales and decoder layers. Our design allows for transductive fine-tuning of the visual prompts leveraging the unlabelled test image. + +# 4.1. Visual Prompting Multiscale Transformers + +Base Visual Prompts. Given $B$ base classes, we define $B$ visual prompts each representing a class and initialize them randomly. These visual prompts first self-attend to learn the relationship between themselves and then are used to prompt the image features at multiple scales. The refinement of base prompts at each decoder layer is given by: + +$$ +V _ {B} ^ {(l)} = \mathcal {A} \left(V _ {B} ^ {(l - 1)}\right) + \mathcal {C} \left(V _ {B} ^ {(l - 1)}, F ^ {(l - 1)}\right), \tag {1} +$$ + +where $\mathcal{A}$ and $\mathcal{C}$ refer to multi-headed self and cross attention respectively; $V_{B}^{(l)}\in \mathbb{R}^{B\times C}$ are the base visual prompts at layer $l$ ; $F^{(l)}\in \mathbb{R}^{H_lW_l\times C_l}$ is the flattened image feature at layer $l$ ; $H_{l}$ and $W_{l}$ denote image dimensions at that layer; $C_l$ denotes the embedding dimension. + +After refining the prompts through multiple levels of transformer attention, a segmentation head, denoted as $\mathcal{H}(\cdot)$ , is defined. This segmentation head takes the final refined prompts, $V_{B}^{(L)}$ , and the highest resolution image features, $F^{(L)}$ , as input, producing per-pixel per-class predictions $O_{base} \in \mathbb{R}^{B \times H \times W}$ (see Section 4.3). + +Novel Visual Prompts. After training on the base classes, we freeze all the layers of the model except the visual prompts representing the base classes. Given there are $N$ novel classes, we add $N$ new visual prompts. We further add a causal unidirectional attention module that allows the novel prompts to be enhanced by the base prompts (see Section 4.2). The output is passed to a segmentation head to make predictions for the novel classes. + +The novel visual prompts are first refined by a unidirectional novel-to-base causal attention module. The enhanced + +novel prompts are then concatenated with the base prompts. Overall the concatenated prompts are refined by: + +$$ +V _ {N} ^ {(l)} = \mathcal {C A} \left(V _ {B} ^ {(l)}, V _ {N} ^ {(l)}\right) +$$ + +$$ +V _ {A} ^ {(l)} = \left[ V _ {B} ^ {(l)}, V _ {N} ^ {(l)} \right] +$$ + +$$ +V _ {A} ^ {(l)} = \mathcal {A} (V _ {A} ^ {(l - 1)}) + \mathcal {C} (V _ {A} ^ {(l - 1)}, F ^ {(l - 1)}), +$$ + +where $V_{N}^{(l)}\in \mathbb{R}^{N\times C}$ are novel prompts at layer $l$ ; $N$ is the total number of novel class and $C$ is the embedding size; $V_{A}^{(l)}\in \mathbb{R}^{(B + N)\times C}$ are all visual prompts including base and novel; $[.,.]$ is the concatenation operator. The operation $\mathcal{CA}$ denotes the causal novel-to-base attention and is explained in Section 4.2. + +Each of the $N$ novel visual prompts (representing a novel class) is initialized using masked average global pooling of the k-shot support images corresponding to the class masked by the ground truth binary masks. + +$$ +V _ {n} ^ {(0)} = \frac {1}{K} \sum_ {k = 1} ^ {K} \frac {\sum_ {x , y} M _ {n} ^ {k} (x , y) F _ {n} ^ {k} (x , y)}{\sum_ {x , y} M _ {n} ^ {k} (x , y)}, \forall n \in N \tag {2} +$$ + +where $V_{n}^{(0)} \in \mathbb{R}^{1 \times C}$ is the initial novel visual prompt for novel class $n$ ; $F_{n}^{k}$ denotes the image features of $k$ -th shot for novel class; $M_{n}^{k}$ denotes the corresponding ground-truth binary mask; $x, y$ are the spatial locations; $N$ is the total number of novel classes and $K$ is the total number of shots. + +# 4.2. Novel-to-base Causal Attention + +A key component of our proposed approach is the unidirectional novel-to-base prompt causal3 attention. The base visual prompts are learned with abundant examples; this is not the case for novel prompts. Learning novel prompts directly or having novel prompts impact the base ones could therefore lead to degradation of performance. However, having novel prompts aware of their base counterparts could be quite helpful. We hypothesize that causal attention between the novel visual prompts and base prompts could help contextualize novel class embeddings, reducing confusion with the base counterparts. Motivated by this intuition, we propose a novel-to-base causal attention layer that is repeated at each scale (or layer of the transformer decoder); see Fig. 2. + +The proposed causal attention module is a single layer cross-attention where the Key $K^{(l)}$ , Value $V^{(l)}$ and $Q^{(l)}$ for layer $l$ of the decoder are given by: + +$$ +Q ^ {(l)} = V _ {N} ^ {(l)} W ^ {Q}; \quad K ^ {(l)} = V _ {B} ^ {(l)} W ^ {K}; \quad V ^ {(l)} = V _ {B} ^ {(l)} W ^ {V} +$$ + +where $W^{Q}, W^{K}, W^{V}$ are weight matrices for $Q, K$ and $V$ respectively. The parameters $W^{Q}, W^{K}, W^{V}$ are shared at all the layers of this cross-attention module to reduce the number of trainable parameters during fine-tuning on novel classes to prevent overfitting and poor generalization. + +This uni-directional novel-to-base causal attention refines the novel visual prompts $V_{N}^{(l)}$ at each decoder layer $l$ by: + +$$ +V _ {N} ^ {(l)} = \operatorname {s o f t m a x} (Q ^ {(l)} (K ^ {(l)}) ^ {\mathrm {T}}) V ^ {(l)} +$$ + +This results in a set of contextualized novel visual prompts. + +# 4.3. Segmentation Head + +Base Segmentation Head. The base segmentation head, $\mathcal{H}_B$ , during the base training in the proposed model is relatively simple and takes the form of a three layer MLP that projects a refined prompt into, effectively, a class prototype; computes the (dot product) similarity of each pixel feature with that prototype, and then applies softmax along the channel / class dimension to obtain per-pixel and per-class probability: + +$$ +\begin{array}{l} O _ {b a s e} = \mathcal {H} _ {B} (V _ {B} ^ {(L)}, F ^ {(L)}) \\ = \operatorname {M L P} \left(V _ {B} ^ {(L)}\right) \cdot \left[ F ^ {(L)} \right] ^ {\mathrm {T}}, \tag {3} \\ \end{array} +$$ + +where $O_{base} \in \mathbb{R}^{B \times H \times W}$ and the class assignment can then be made by an argmax along the channel dimension. + +Novel Segmentation Head. While we could, in principle, use the same form of segmentation head during few-shot + +inference as that defined for the base classes (see Eq. 3), simply fine-tuning it for novel classes, this proves ineffective in practice. Specifically, fine-tuning the MLP leads to significant overfitting given the few examples and lack of base class data during this step. So, instead, we resort to a simpler form of the novel segmentation head, $\mathcal{H}_N$ , where we only learn a residual to the MLP learned during base training (which we keep fixed), mainly: + +$$ +\begin{array}{l} O _ {n o v e l} = \mathcal {H} _ {N} (V _ {N} ^ {(L)}, F ^ {(L)}) \\ = \left[ \operatorname {M L P} _ {\bullet} \left(V _ {N} ^ {(L)}\right) + W _ {N} \right] \cdot \left[ F ^ {(L)} \right] ^ {\mathrm {T}}, \\ \end{array} +$$ + +where $O_{\text{novel}}$ is output for novel classes and $\bullet$ designates that parameters of MLP are frozen and only $W_N \in \mathbb{R}^{N \times C}$ weight matrix is learned. This weight matrix is initialized in the same manner as novel prompts, in Eq. 2. + +# 4.4. Transductive Prompt Tuning + +Recent approaches [4, 13] show the effectiveness of a good transductive setting for few-shot dense prediction tasks. Transductive approaches perform test-time optimization for each unlabeled test image and utilize specific characteristics and information of the test instances to improve prediction. + +To demonstrate the flexibility of our approach, in this section we outline the process of adapting our model for transductive inference, improving GFSS performance through transductive fine-tuning of the visual prompts. We build on the transductive losses proposed in DlaM [13] and we describe them here for completeness. + +The losses are designed to maximize the mutual information between the learned features and the corresponding predictions on the test image, which is achieved through maximizing: $\mathrm{H}(O) - \mathrm{H}(O|I)$ , where $I, O$ are random variables associated with the pixels and predictions distributions respectively; $\mathrm{H}(O)$ is the marginal and $\mathrm{H}(O|I)$ is the conditional entropy. The conditional entropy $\mathrm{H}(O|I)$ involved is given by sum of cross-entropy loss over the supervised support set and entropy of the predicted probability given the test image $I$ . The marginal entropy is given by KL divergence loss relying on an estimated region proportion prior following [4]. Additional details are in [13]. + +Additionally, similar to [13], we use knowledge-distillation loss to preserve base class performance, given by: + +$$ +\mathcal {L} _ {\mathrm {K D}} = \mathrm {K L} \left(O _ {\text {b a s e}} ^ {\text {n e w}} \mid \mid O _ {\text {b a s e}} ^ {\text {o l d}}\right), \tag {4} +$$ + +where $O_{base}^{old}$ is the prediction probability of the base classes of the frozen model after base training; $O_{base}^{new}$ is the prediction probability of the base classes on the model with our additional components to handle novel categories that is being fine-tuned. These include the multi-scale causal novel-to-base cross attention and the novel prompts. + +The overall transductive objective function is given by: + +$$ +\mathcal {L} _ {\text {t r a n s .}} = \alpha \mathrm {H} (O | I) - \mathrm {H} (O) + \gamma \mathcal {L} _ {\mathrm {K D}}, \tag {5} +$$ + +where $\alpha$ and $\gamma$ are balancing hyper-parameters. + +We noticed that applying transduction from the first iteration underperforms, likely due to a poor initial estimate of the marginal distribution. To address this, we only apply per-pixel cross-entropy loss for a set number of iterations and then incorporate the remaining transductive losses from Eq. 5 when we achieve a more accurate estimate of the marginal distribution. + +# 5. Experiments + +# 5.1. Dataset and Implementation Details + +Datasets. We evaluate our method on two few-shot segmentation benchmarks [36]: COCO- $20^{i}$ and PASCAL- $5^{i}$ . Following [13] we report average performance of our model on 10K test images for COCO- $20^{i}$ and on all test images on PASCAL- $5^{i}$ . The results are reported as an average across five different runs on four different splits. + +Data Pre-processing. We adopt the same setup as DIAM [13]. During base training, images containing both base and novel categories are kept, but the pixels corresponding to novel classes are relabeled as background. This approach, as noted by [13], presents challenges due to potential ambiguity when the base model predicts novel categories as background. In fine-tuning, only novel classes are labeled in the support set, with base classes marked as background. To avoid unfair shot increases, images with multiple novel categories are excluded from the support set. + +Evaluation Protocol. We evaluate performance using the standard mean intersection-over-union (mIoU) metric. Following [13], the mean score for generalized few-shot segmentation is computed as the average between base and novel mIoUs. This approach addresses bias in datasets like PASCAL- $5^{i}$ and COCO- $20^{i}$ , where base classes outnumber novel classes threefold. + +Implementation Details. We use PSPNet [51] with Resnet-50 [15] as feature extractor. We utilize MSDeformAttn-based pixel decoder [53] from Mask2Former [10]. The dimension $C$ for visual prompts is chosen to be 256. Contrary to [10], the base model is trained with per-pixel cross-entropy loss. We use AdamW optimizer [25] with base learning rate of $1 \times 10^{-4}$ , a weight decay of 0.05 and a batch size of 16. Following [13, 19], we train base model for 20 epochs for COCO- $20^{i}$ and 100 epochs for PASCAL- $5^{i}$ . + +During novel fine-tuning, we freeze the parameters of the feature extractor, pixel decoder, and base segmentation head, and only optimize the parameters of the visual prompts, novel segmentation head, and novel-to-base causal attention module. For the inductive case, the model is finetuned with per-pixel cross-entropy loss for 100 iterations. For the transductive setting, the model is fine-tuned for 40 iterations with only cross-entropy loss and then fine-tuned for 60 more iterations with full transductive losses mentioned in Eq. 5. The hyper-parameters $\alpha$ and $\gamma$ are set to + +be 100 and 25 respectively. We use AdamW optimizer with a learning rate of $5 \times 10^{-3}$ and a weight decay of 0.05 for fine-tuning. + +# 5.2. Results + +Comparison with the State-of-The-Art. Table 1 compares our approach in both inductive and transductive setting against the state-of-the-art. As observed, our model outperforms existing approaches. Interestingly, our approach, in an inductive setting (where optimization is performed only using a supervised few-shot support set, and evaluation is conducted on all test images), outperforms transductive approaches like DIaM [13] for both 1-shot and 5-shot cases on the COCO-20 $^i$ dataset and for the 1-shot case on Pascal-5 $^i$ . This is notable despite transductive approaches requiring test-time optimization for each unlabeled test image separately, leading to significantly longer inference times. + +For a fair comparison in the inductive setting, we finetuned DIaM without the transduction losses. We also finetuned POP [21] in our inductive setup. POP [21] randomly samples balanced base class image, making it easier for them to retain base class information and uses a form of transduction where their model relabels novel class pixels initially labelled as background. We observe that the performance of POP [21] degrades substantially, especially for the base classes in our inductive setting. + +Our inductive approach has $5.78\%$ higher mIoU for 1-shot and $11.16\%$ higher mIoU for 5-shot on COCO-20 $^i$ than inductive DIaM. For PASCAL-5 $^i$ , the improvement from inductive DIaM is $7.71\%$ for 1-shot and $13.29\%$ for 5-shot. In fact, compared to the best inductive GFSS approach BAM [19], our inductive setting achieves an improvement of approximately $2.78\%$ and $7.59\%$ for 1-shot and 5-shot respectively on COCO-20 $^i$ and $5.24\%$ and $12.32\%$ for 1-shot and 5-shot respectively on PASCAL-5 $^i$ . Our method, in the inductive setting, even outperforms transductive DIaM [13] by $2.03\%$ and $2.28\%$ for 1-shot and 5-shot respectively on COCO-20 $^i$ and by $1.79\%$ for 1-shot on PASCAL-5 $^i$ + +On applying transductive fine-tuning to our method, overall mIoU increases considerably, outperforming transductive DIaM [13] and POP [21] on both 1-shot and 5-shot inference for both datasets. The improvement over DIaM for $\mathrm{COCO - }20^{i}$ is $3.3\%$ for 1-shot and $3.93\%$ for 5-shot. For PASCAL- $5^{i}$ , our transductive setting outperforms DIaM by $5.11\%$ and $3.15\%$ for 1 and 5-shots. + +# 5.3. Ablation + +We performed ablation studies (refer to Table 2) to explain our design choices and contribution of each components. + +Visual Prompts Initialization. As seen in (1) and (2) of the Table 2, initializing novel prompts with masked mean pooling from the support set images improves both base and novel class accuracy compared to random initialization. + +
MethodLearningCOCO-20i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CAPL [36] (CVPR22)Inductive43.217.2125.2143.7111.0027.36
BAM [19] (CVPR22)Inductive49.8414.1632.0049.8516.6333.24
DIaM (w/o trans.)* [13] (CVPR23)Inductive42.6915.3229.0038.4720.8729.67
POP† [21] (CVPR23)Inductive30.389.6320.0024.5316.1920.36
Ours (w/o trans.)Inductive51.5518.0034.7851.5930.0640.83
RePRI [4] (CVPR21)Transductive5.624.745.188.858.848.85
DIaM [13] (CVPR23)Transductive48.2817.2232.7548.3728.7338.55
POP† [21] (CVPR23)Transductive54.7115.3135.0154.9029.9742.44
Ours (w/ trans.)Transductive53.8018.3036.0553.8131.1442.48
+ +
MethodLearningPASCAL-5i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CANet [49] (CVPR19)Inductive8.732.425.589.051.525.29
PFENET [35] (TPAMI20)Inductive8.322.675.508.831.895.36
PANET [40] (ICCV19)Inductive31.8811.2521.5732.9515.2524.1
SCL [47] (CVPR21)Inductive8.882.445.669.111.835.47
MiB [8] (CVPR20)Inductive63.808.8636.3368.6028.9348.77
CAPL [36] (CVPR22)Inductive64.8017.4641.1365.4324.4344.93
BAM [19] (CVPR22)Inductive71.6027.4949.5571.6028.9650.28
DIaM (w/o trans.)* [13] (CVPR23)Inductive66.7927.3647.0864.0534.5649.31
POP† [21] (CVPR23)Inductive46.6819.9633.3241.5036.2638.80
Ours (w/o trans.)Inductive74.5834.9954.7974.8650.3462.60
RePRI [4] (CVPR21)Transductive20.7610.5015.6334.0620.9827.52
DIaM [13] (CVPR23)Transductive70.8935.1153.0070.8555.3163.08
POP† [21] (CVPR23)Transductive73.9235.5154.7274.7855.8765.33
Ours (w/ trans.)Transductive76.3939.8358.1176.4256.1266.27
+ +Table 1. Comparison to SoTA methods on COCO-20 $^i$ and PASCAL-5 $^i$ . The mIoU results are reported as the average across 5 different runs. All methods use ResNet-50 backbone. $\dagger$ POP [21] uses an easier and less practical GFSS setup, e.g., balanced base class training in addition to support set during novel fine-tuning and uses a form of transduction; inductive setting is implemented by us following our setup. Results of the rest of the methods are obtained from [13] except the inductive setting of DIA ${M}^{*}$ , which we implemented ourselves. The w/o trans., w/ trans. denote without and with transduction respectively. + +
Causal AttentionPrompt InitializationTransductionmIoU
BaseNovelMean
(1)NoneRandomNo50.1811.2230.70
(2)NoneMasked PoolingNo50.4311.3130.87
(3)First Layer OnlyMasked PoolingNo51.5312.2631.90
(4)Separate WeightsMasked PoolingNo51.0618.0534.56
(5)Shared WeightsRandomNo50.7517.4134.08
(6)Shared WeightsMasked PoolingNo51.5518.0034.78
(7)Shared WeightsMasked PoolingYes53.8018.3036.05
+ +Table 2. Ablation Study on COCO- ${20}^{i}$ . The 1-shot performane, averaged over all folds. + +This is more pronounced with novel-to-base causal attention (5)-(6), amplifying the performance gain from masked pooling initialization. + +Novel-to-base Causal Attention. As seen in (4)-(7) of Table 2, novel-to-base causal attention yields the most significant improvement in novel class accuracy compared to (1)-(2). This aligns with our hypothesis that uni-directional causal attention between rich base prompts (learned with many examples) and novel prompts (learned with few examples) increases separation among novel classes and pushes the feature space of base prompts away from the novel prompts. It is further confirmed by Figure 3 and Fig- + +![](images/0913caea56e1a20fe6b9e35147adf3fe3496a7b9ab746fe2ea30d39e2c1fddd2.jpg) +Figure 3. Confusion Matrix for: (left) model with Novel-to-Base Causal Attention; (right) model without Novel-to-Base Causal Attention. Note that Novel-to-Base Causal Attention reduces confusion between novel and base categories (bottom-left block). + +ure 4. As observed in Figure 3, the novel-to-base causal attention leads to lower confusion between base and novel classes. Figure 4 shows a t-SNE [37] visualization of the features of base and novel visual prompts for 10 different + +![](images/23f666fdbb8253d19932f086feb2fb76eb6041f3a7d88db61e4da710a9e9d5f3.jpg) +Figure 4. TSNE visualization of the learned base and novel prompt features: (left) with Novel-to-Base Causal Attention; (right) without Novel-to-Base Casual Attention. + +test images. As observed, the incorporation of novel-to-base causal attention leads to much clearer separation of visual prompt features, particularly for novel prompts. + +Multiscale Causal Attention. We also ablated the placement of the causal attention module. We observe that applying it only at the first layer (3) slightly improves novel mIoU compared to the model without causal attention. Yet, it is notably inferior to applying at every decoder layer (4)-(7). We could have opted for separate parameters for the causal attention module at each layer. However, as seen in (4), having separate weights offers little improvement to novel mIoU despite significant increase in total number of parameters to optimize. In fact, it decreases the base accuracy, leading to overall worse performance. + +Qualitative Results. Figure 5 compares the qualitative results of our model for 1-shot GFSS on Pascal- $5^{i}$ in both inductive and transductive settings with the baseline lacking causal attention and DIaM [13]. Notably, for first row, both our inductive and transductive approaches predict more accurate segmentation of the cow compared to DIaM. In second row, the causal attention improves person segmentation followed by transduction. In the third row, DIaM completely misses the novel category chair. The causal attention allows our model to better differentiate the base class dining table from the novel class chairs. Transduction improves accuracy of the segmentation of dining table. In fourth row, we can observe that both DIaM and our model without causal attention misdetects novel class cat for cow at multiple locations. Causal attention allows our model to make a uniform and more accurate prediction of cat. In the fifth row, we observe that DIaM [13] completely misses detecting the base category potted plant and poorly segments person. The baseline without causal attention misdetects the potted plant with novel class bus and misdetects dining table with chair. Causal attention improves the accuracy which is further improved by transduction. Last row + +![](images/f675ac8b2fdd0e8dedcc775892190cc201f7b23bbfb609070141f9c1ff19f758.jpg) +Figure 5. Qualitative Results for 1-shot on Pascal- $5^{i}$ . The left-most two columns show image and ground truth mask; (Third) Baseline without causal attention; (Fourth); Ours in inductive setting; (Fifth) Ours in transductive setting; (Last) DIaM [13]. Last row illustrates a failure. + +demonstrates a case where our model performs worse than DlaM [13]. Again, we observe that adding causal attention reduces misclassification of person with car, although overall the segmentation of car is less accurate than DlaM [13]. + +From the qualitative results, the effectiveness of novel-to-base causal attention at distinguishing base from novel classes becomes apparent. Our model in the inductive setting outperforms transductive DIaM in most cases, despite the inference time being faster. + +# 6. Conclusion + +In this work, we proposed a technique to learn visual prompts from few-shot examples of novel unseen cateogries to prompt a pre-trained segmentation model that not only generalizes well on the unseen novel class but strongly retains performance over the base classes on which the model was pre-trained. We achieve this by prompting the image features at multiple scales and proposing a uni-directional novel-to-base causal attention mechanism that enriches the novel visual prompts through interaction with the base visual prompts. Our proposed method can be applied in both inductive and transductive settings and achieves state-of-the-art performance on two benchmark few-shot segmentation datasets. + +# References + +[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, E. Rutherford, S. Cabi, T. Han, Z. Gong, S. Samangooei, M. Monteiro, J. L. Menick, S. Borgeaud, A. Brock, A. Nematzadeh, S. Sharifzadeh, M. Binkowski, R. Barreira, O. Vinyals, A. Zisserman, and K. Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3 +[2] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S.v.ArxM.S.BernsteinA.Bosselut J.BohgE.Brunskill, E. Brynjolfsson, S. Buch, D.Card, R.Castellon, N.Chatterji, A. Chen, K.Creel, J.Q.Davis,D.Demszky,C.Donahue,M.Doumbouya,S.Ermon E.Durmus,J.Etchemendy, K.Ethayarajh, L.Fei-Fei, C.FinnT.GaleL.GillespieK. Goel, N. Goodman, S. Grossman, N. Guha, T. Hashimoto, P. Henderson, J. Hewitt, D.E.HoJ.HongK.HsuJ.Huang, T.IcardS.JainD.JurafskyP.KalluriS.KaramchetiG. KeelingF.KhaniO.KhattabP.W.KohM.KrassR. Krishna,R.Kuditipudi,A.Kumar,F.Ladhak,M.Lee,T. LeeJ.LeskovecI.LeventX.L.LiX.LiT.MaA.MlikC.D.ManningS.Mirchandani,E.MitchellZ.MunyikwaS.NairD.Narayanan A.NarayanB.NewmanA. NieJ.C.Niebles,H.NilforoshanJ.Nyarko,G.OgutI. PapadimitriouJ.S.ParkC.PiechE.PortelanceC.Potts, A.RaghunathanR.ReichH.RenF.RongY.RoohaniC. RuizJ.RyanC.ReD.SadighS.SagawaK.Santhanam, A.ShihK.SrinivasanA.TamkinR.TaoriA.W.Thomas F.TramereR.E.WangW.WangB.WuJ.WuY.WuS.M XieM.YasunagaJ.YouM.ZahariaM.ZhangT.Zhang X.ZhangY.ZhangL.ZhengK.Zhou and P.Liang.On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021.3 +[3] Malik Boudiaf, Imtiaz Ziko, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Information maximization for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 33:2445-2457, 2020. 2 +[4] M. Boudiaf, H. Kervadec, Z. Masud, P. Piantanida, I. Ben Ayed, and J. Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13979-13988, 2021. 1, 2, 5, 7 +[5] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, G. Sastry P. Shyam, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33: 1877-1901, 2020. 1, 2 +[6] A. Bulat, R. Guerrero, B. Martinez, and G. Tzimiropoulos. Fs-detr: Few-shot detection transformer with prompting and without re-training. In International Conference on Computer Vision (ICCV), pages 11793-11802, 2023. 1, 3 +[7] T. Cao, M. Law, and S. Fidler. A theoretical analysis of the + +number of shots in few-shot learning. International Conference on Learning Representations (ICLR), 2020. 2 +[8] F. Cermelli, M. Mancini, S.R. Bulo, E. Ricci, and B. Caputo. Modeling the background for incremental learning in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233–9242, 2020. 7 +[9] W. Chen, Y. Liu, Z. Kira, Y. Frank Wang, and J. Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), 2019. 2 +[10] B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 3, 6 +[11] G. S. Dhillon, P. Chaudhari, A. Ravichandran, and S. Soatto. A baseline for few-shot image classification. In International Conference on Learning Representations (ICLR), 2020. 2 +[12] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning (ICML), 2017. 2 +[13] S. Hajimiri, M. Boudiaf, I. B. Ayed, and J. Dolz. A strong baseline for generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11269-11278, 2023. 2, 3, 5, 6, 7, 8 +[14] B. Hariharan and R. B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In International Conference on Computer Vision (ICCV), 2017. 2 +[15] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[16] M. Jia, L. Tang, B. Chen, C. Cardie, S. Belongie, B. Hariharan, and S. Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 2, 3 +[17] B. Kang, Z. Liu, X. Wang, F. Yu, J. Feng, and T. Darrell. Few-shot object detection via feature reweighting. In International Conference on Computer Vision (ICCV), 2019. 2 +[18] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, A. C. Berg S. Whitehead, W. Lo, P. Dollár, and R. Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3 +[19] C. Lang, G. Cheng, B. Tu, and J. Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8067, 2022. 6, 7 +[20] X. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021.3 +[21] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11319-11328, 2023. 2, 6, 7 + +[22] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2 +[23] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. J. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2 +[24] Y. Liu, X. Zhang, S. Zhang, and X. He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), 2020. 1, 2 +[25] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6 +[26] J. Min, D. Kang, and M. Cho. Hypercorrelation squeeze for few-shot segmentation. In International Conference on Computer Vision (ICCV), pages 6941-6952, 2021. 1, 2 +[27] A. Nichol, J. Achiam, and J. Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2 +[28] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 2 +[29] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8412-8422, 2021. 2 +[30] L. Qiao, Y. Shi, J. Li, Y. Wang, T. Huang, and Y. Tian. Transductive episodic-wise adaptive metric for few-shot learning. In International Conference on Computer Vision (ICCV), pages 3603-3612, 2019. 2 +[31] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1 +[32] S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. 2 +[33] M. Ren, E. Triantafillou, S. Ravi, J.Snell, K. Swersky, J. B. Tenenbaum, H. Larochelle, and R. S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), 2018. +[34] J. Snell, K. Swersky, and R. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2 +[35] Z. Tian, H. Zhao, M. Shu, Z. Yang, R. Li, and J. Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(2):1050-1065, 2020. 7 +[36] Z. Tian, X. Lai, L. Jiang, S. Liu, M. Shu, H. Zhao, and J. Jia. Generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11563-11572, 2022. 2, 3, 6, 7 + +[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 7 +[38] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 5 +[39] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 3630-3638, 2016. 2 +[40] K. Wang, J. Liew, Y. Zou, D. Zhou, and J. Feng. PANet: Few-shot image semantic segmentation with prototype alignment. In International Conference on Computer Vision (ICCV), pages 9197-9206, 2019. 1, 2, 7 +[41] X. Wang, T. E Huang, T. Darrell, J. E. Gonzalez, and F. Yu. Frustratingly simple few-shot object detection. In International Conference on Machine Learning (ICML), 2020. 2 +[42] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In International Conference on Computer Vision (ICCV), 2019. 2 +[43] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1 +[44] X. Yan, Z. Chen, A. Xu, X. Wang, X. Liang, and L. Lin. Meta R-CNN: Towards general solver for instance-level low-shot learning. In International Conference on Computer Vision (ICCV), 2019. 2 +[45] B. Yang, C. Liu, B. Li, J. Jiao, and Q. Ye. Prototype mixture models for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), pages 763-778, 2020. 1, 2 +[46] X. Yang, B. Wang, K. Chen, X. Zhou, S. Yi, W. Ouyang, and L. Zhou. BriNet: Towards bridging the intra-class and inter-class gaps in one-shot segmentation. In British Machine Vision Conference (BMVC), 2020. 1, 2 +[47] B. Zhang, J. Xiao, and T. Qin. Self-guided and cross-guided learning for few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8321, 2021. 7 +[48] C. Zhang, G. Lin, F. Liu, J. Guo, Q. Wu, and R. Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In International Conference on Computer Vision (ICCV), pages 9587-9595, 2019, 1, 2 +[49] C. Zhang, G. Lin, F. Liu, R. Yao, and C. Shen. CANet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217–5226, 2019. 1, 2, 7 +[50] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 1, 3 +[51] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In IEEE/CVF Conference on Computer + +Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6 +[52] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2 +[53] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representations (ICLR), 2021. 6 \ No newline at end of file diff --git a/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/images.zip b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9297817e7d3ab3eaf45b0c3ddbc7b90b0005447 --- /dev/null +++ b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bfc50de9db3dbf82cd9fbd5d420320e8dd3ac4b4b7a5c2a2d9f5f56397c8ce8 +size 507447 diff --git a/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/layout.json b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f9a32ab20fe8f14a51f81f8980a2b19cb5999db --- /dev/null +++ b/2024/Visual Prompting for Generalized Few-shot Segmentation_ A Multi-scale Approach/layout.json @@ -0,0 +1,9120 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 477, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 477, + 139 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 477, + 139 + ], + "type": "text", + "content": "Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 216, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 216, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 216, + 175 + ], + "type": "text", + "content": "Mir Rayat Imtiaz Hossain" + }, + { + "bbox": [ + 61, + 160, + 216, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 175, + 231, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 175, + 231, + 189 + ], + "spans": [ + { + "bbox": [ + 75, + 175, + 231, + 189 + ], + "type": "text", + "content": "1University of British Columbia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 161, + 324, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 161, + 324, + 175 + ], + "spans": [ + { + "bbox": [ + 223, + 161, + 324, + 175 + ], + "type": "text", + "content": "Mennatullah Siam" + }, + { + "bbox": [ + 223, + 161, + 324, + 175 + ], + "type": "inline_equation", + "content": "^{3,1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 175, + 369, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 175, + 369, + 188 + ], + "spans": [ + { + "bbox": [ + 256, + 175, + 369, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 256, + 175, + 369, + 188 + ], + "type": "text", + "content": "Vector Institute for AI" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 189, + 372, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 189, + 372, + 201 + ], + "spans": [ + { + "bbox": [ + 247, + 189, + 372, + 201 + ], + "type": "text", + "content": "4Canada CIFAR AI Chair" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 161, + 431, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 161, + 431, + 175 + ], + "spans": [ + { + "bbox": [ + 350, + 161, + 431, + 175 + ], + "type": "text", + "content": "Leonid Sigal" + }, + { + "bbox": [ + 350, + 161, + 431, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 456, + 161, + 531, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 456, + 161, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 456, + 161, + 531, + 175 + ], + "type": "text", + "content": "James J. Little" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 392, + 175, + 518, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 175, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 392, + 175, + 518, + 190 + ], + "type": "text", + "content": "3 Ontario Tech University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "content": "The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base categories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mechanism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-20" + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "content": " and Pascal5" + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "content": ", without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabeled test data can be used to improve the prompts, which we refer to as transductive prompt tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": "The emergence of foundational models, trained on broad datasets, in natural language processing (e.g., GPT-3 [5]) and in vision-language (e.g., CLIP [31]), exhibits powerful generalization and strong performance on multiple downstream tasks. These models have been adapted through different prompting techniques [6, 43] for them to be used in the few-shot scenario." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "However, despite successes, including in localization [6, 43, 50], prompting, which allows for few-shot demon" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 322, + 230, + 529, + 468 + ], + "blocks": [ + { + "bbox": [ + 322, + 230, + 529, + 468 + ], + "lines": [ + { + "bbox": [ + 322, + 230, + 529, + 468 + ], + "spans": [ + { + "bbox": [ + 322, + 230, + 529, + 468 + ], + "type": "image", + "image_path": "cdbde3ca844e54cf00bd6dbaf756c69e556440150096a7dce4fe7910aa402885.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 471, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 515 + ], + "type": "text", + "content": "Figure 1. Overview of the Proposed Approach. Prompting multi-scale transformer decoders for generalized few-shot segmentation. Our approach is a simple approach that allows test-time transductive prompt tuning (see red arrows)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "type": "text", + "content": "strations in dense prediction, and specifically in semantic segmentation tasks, is relatively under-explored." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Few-shot semantic segmentation [24, 40, 45, 46, 48, 49] aims to segment novel (unseen) classes with few labelled training examples. Most state-of-the-art methods rely on meta-learning, to leverage abundant training data as a form of \"data augmentation\" to construct many tasks that resemble test-time few-shot inference [26, 45]. In general, few-shot learning approaches can be classified as either inductive or transductive. Inductive approaches mainly rely on the training data, while transductive ones [4], in addition, exploit unlabelled test data in an unsupervised manner to improve the performance " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "e.g." + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": ", by leveraging entropy priors over predictive class distributions. However, both cases largely focus on the performance of the novel categories; which isn't particularly realistic. A recently proposed gen" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23470" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "eralized few-shot segmentation (GFSS) setting defines a more realistic scenario where the goal is to perform well on all classes, novel and base [36]. This is considerably more challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 123, + 288, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 123, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 288, + 291 + ], + "type": "text", + "content": "Prompting has been proven to be effective for learning from few demonstrations, as seen in LLMs [5]. It was also explored in visual prompt tuning [16] to efficiently fine-tune vision transformers for new tasks. We posit that prompting within a transformer-based [38] architecture would similarly offer an effective and flexible mechanism for GFSS. This involves learning prompts that can be utilized to perform cross-attention with input image to make predictions. While learning prompts for base classes, with abundant data, is relatively straightforward, it becomes considerably more challenging for novel classes with few examples. Specifically, one must ensure that novel prompts learned from few samples are sufficiently distinct from the base prompts to avoid novel-base class misclassification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 288, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 534 + ], + "type": "text", + "content": "To address the aforementioned challenges, we develop a simple yet highly effective visual prompting of transformer decoders for dense prediction at multiple scale that relies on novel-to-base causal attention without meta-training. We look upon queries in DETR-style architectures as a form of visual prompts and devise a mechanism to initialize and learn novel prompts. Then the novel-to-base causal attention allows base prompts to affect novel prompt representations, but not the other way around. This, intuitively, allows novel prompts to be repulsed and/or attracted towards their base counterparts. This attention is shared across scales (layers of the transformer) which, as we show, leads to more robust learning and improved performance. The multi-scale refinement of prompts helps interacting and reasoning between image features at multiple scales which in turn helps in better dense prediction. Finally, we extended this architecture in a transductive setting, where both novel and base prompts can be fine-tuned at test-time with respect to unsupervised objective to further improve performance. The proposed architecture and methodology is shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 536, + 200, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 536, + 200, + 548 + ], + "spans": [ + { + "bbox": [ + 59, + 536, + 200, + 548 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 287, + 682 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 551, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 287, + 610 + ], + "type": "text", + "content": "- Designing a multi-scale visual prompting transformer decoder architecture for GFSS, featuring learnable prompts that allow the creation of new prompts for novel classes, initialized through masked average pooling of support images (with their masks)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 611, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 287, + 647 + ], + "type": "text", + "content": "- Within this architecture, proposing and learning a multiscale (shared) novel-to-base cross-attention mechanism between the novel and base prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "type": "text", + "content": "- Proposing a novel transductive prompt tuning, which allows the visual prompts to be tuned on test (unlabeled) images, hence the term transductive." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 95, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 370 + ], + "type": "text", + "content": "Few-shot Learning. Few-shot learning, a longstanding concept in computer vision, serves as a paradigm to evaluate the data efficiency and adaptability of learning algorithms. It focuses on making predictions for novel classes with a few labeled examples, often by leveraging information from disjoint base classes for which there is abundant data. Few-shot learning has been extensively studied for image classification [12, 32-34, 39]. Most approaches can be categorized as either transfer or meta-learning-based. Transfer learning methods [11] typically pre-train the model on base classes with abundant data and then fine-tune this model for novel classes with few samples. Meta-learning [12, 32] approaches, on the other hand, simulate the inference stage during meta-training by sampling tasks of support and target sets. Meta-learning approaches can be sub-categorized into gradient-based and metric-learning based. Gradient-based approaches [12, 32] learn the commonalities among different tasks. Metric-based (or prototype-based) methods [34, 39] learn image embeddings and compare the distances between them. Despite its popularity and strong performance, meta-learning is found to be sensitive to the meta-training setup [7] and worse performance with domain shift [9] is observed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 517 + ], + "type": "text", + "content": "Generalized few-shot learning, introduced in [14] as a more realistic variant, allows query images from either base or novel categories. Recently, the transductive setting has been proven to be effective for few-shot learning [3, 4, 13, 29, 52]. Transductive approaches exploit information of unlabelled test images during inference to improve classification [11, 22, 27, 30]. These approaches build on classic semi-supervised learning approaches, such as graph-based label propagation [23] or entropy minimization [11]. Recent variants of few-shot learning problems have also been established in other more granular domains, e.g., detection [17, 41, 42, 44] and segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Few-shot segmentation. Less common than classification, few-shot segmentation is emerging as the dense counterpart to the aforementioned approaches. Few-shot segmentation approaches can be categorized to prototype-based and dense comparison-based methods. Prototype-based methods mostly relied on extracted prototypes from the support set to guide the segmentation of the target image, either as a single prototype [40, 49], part-aware [24] or a prototype mixture [45]. Dense comparison methods rely on dense comparisons between the few template images and the target image [46, 48] with extension to multiscale [26]. A recent approach proposed generalized few-shot segmentation (GFSS) task where the unlabelled target images are used to improve the segmentation of novel classes while maintaining the performance on base classes through knowledge distillation [13]. A concurrent work [21] on GFSS learns a set" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 246, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 246, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 246, + 712 + ], + "type": "text", + "content": "1The term 'causal' [28] indicates its unidirectional nature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23471" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "of orthogonal prototypes for base and novel categories, yet their GFSS setup is easier as they sample in their inference balanced base class images along with the novel ones, unlike the setup we follow [13]. Recent work [50] proposed a prompting strategy within a meta-learning framework using few examples. This approach is designed for 1-way segmentation only and performs prompting at a single scale unlike our proposed approach, a non-meta learning scheme, which performs multi-scale prompting and is designed for " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "-way GFSS." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 410 + ], + "type": "text", + "content": "Prompting. An emerging direction in learning with limited labelled data is to utilize foundation models [2] that have strong generalization capabilities. These models, pretrained on large-scale data, demonstrate strong generalization in few-shot settings [1]. Prompting is often used to adapt these foundation models to few-shot tasks, where it formulates the inference problem similarly to the training. Examples include auto-regressive modelling [1] where inference involves generating the next tokens after receiving the correct prompt. For few-shot learning, prompts include few-shot demonstrations followed by the query. Prompting can employ discrete prompts [1] or learning continuous prompts [20] through back-propagation. In computer vision, both vision-language prompting [1] and visual prompting [16] of transformer-based models have shown success in various tasks, particularly in few-shot scenarios. However, effectively prompting models designed for localization and segmentation remains an open question." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 651 + ], + "type": "text", + "content": "Concurrent Work. A recent method [6] proposed prompting of transformer decoders for few-shot object detection. Although conceptually related, their approach is designed for a different task (detection) compared to ours (semantic segmentation), trained within a meta-learning framework unlike ours. Further, [6] lacks novel-to-base attention and multi-scale prompt refinement which are at the core of our architectural design. A recent foundation model for semantic segmentation, SAM [18], exhibits strong performance when prompted by point, bounding box, or text. However, it is not specifically designed for few-shot inference, potentially requiring fine-tuning with the support set and other non-trivial changes. More importantly, the use of extensive data to train the model makes it unsuitable for few-shot tasks since it has likely encountered many target classes during training. In addition to aforementioned differences, including multi-scale prompting and novel-to-base attention, our method is the first to explore transductive prompt-tuning where the visual prompts can be further optimized with the unlabelled target images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Classic Few-shot Segmentation. In the classic few-shot segmentation scenario, the categories are divided into two sets of disjoint sets: base and novel. The problem is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "framed as an " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "-shot segmentation task, where " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " is the number of novel classes and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " is the number of examples available for each class. A typical solution involves meta-learning with two-stage training. In the meta-training stage, a support set is created with " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " randomly selected base classes, and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " examples from each class are used to train the model. In the meta-testing stage, the model is evaluated on tasks with " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " examples from " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " novel classes. An alternative approach is non-meta learning, where a model is initially trained on abundant data for base classes, then frozen and fine-tuned for novel classes using few examples. The focus is on segmenting novel classes in test images without considering base class segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "type": "text", + "content": "Generalized Few-shot Segmentation (GFSS). Generalized few-shot segmentation, proposed by [36], extends the classic few-shot setting. In this new setup the model must preserve the segmentation performance on base classes in addition to generalizing to novel classes. While during finetuning the support set (typically) only contains " + }, + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "type": "text", + "content": "-shot examples for each novel category, the model is still expected to make predictions of both novel and base classes on the test images. We follow the recent more practical setup of generalized few-shot segmentation [13] within a non-meta learning framework, details provided in Section 5.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 364, + 362, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 362, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 362, + 376 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "content": "We perform prompting using a set of learnable visual prompts, where, intuitively" + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "content": ", each prompt corresponds to a class embedding. These learned visual prompts cross-attend with image features at multiple scales. Initially, we pre-train a base model with abundant examples of base classes, learning visual prompts for them. We then fix these base prompts and fine-tune new visual prompts for novel classes with a few examples. However, learning prompts from limited data is challenging because it leads to confusion between base and novel categories. To improve novel class generalization, we introduce a uni-directional causal attention mechanism between novel and base prompts. Intuitively, this allows better contextualization and repulsion / attraction between base and novel prompts. Section 5 demonstrates the empirical effectiveness of this mechanism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 564, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 547, + 685 + ], + "type": "text", + "content": "Figure 2 shows our proposed architecture. It centers around learnable embeddings called visual prompts, which interact with image features at multiple scales. Our design is inspired by Mask2former [10] with a couple of key differences: (i) the number of visual prompts aligns with the number of classes, each representing a class as a learned embedding; (ii) we adopt per-pixel classification, instead of mask classification, thereby eliminating the query classification component of [10] and defining custom segmentation heads capable of making per-pixel predictions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "text", + "content": "2This may not be precisely true in practice, but is a reasonable abstraction for intuitive understanding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23472" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 504, + 284 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 504, + 284 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 504, + 284 + ], + "type": "image", + "image_path": "c837ac7f32204fe304d6787618d26b719d8b773a7770dce2790e294b53ff944f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "type": "text", + "content": "Figure 2. Detailed architecture of our proposed visual prompting of multiscale transformer decoder. Our design initializes the novel visual prompts using the support set. This is followed by consecutive novel-to-base causal attention, " + }, + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{CA}" + }, + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "type": "text", + "content": ", and prompt-to-target features cross attention, " + }, + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 285, + 547, + 329 + ], + "type": "text", + "content": ", across scales. Note that causal attention uses shared weights across the scales and decoder layers. Our design allows for transductive fine-tuning of the visual prompts leveraging the unlabelled test image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 341, + 271, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 271, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 271, + 354 + ], + "type": "text", + "content": "4.1. Visual Prompting Multiscale Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "type": "text", + "content": "Base Visual Prompts. Given " + }, + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "type": "text", + "content": " base classes, we define " + }, + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 359, + 287, + 432 + ], + "type": "text", + "content": " visual prompts each representing a class and initialize them randomly. These visual prompts first self-attend to learn the relationship between themselves and then are used to prompt the image features at multiple scales. The refinement of base prompts at each decoder layer is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 437, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 437, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 84, + 437, + 287, + 453 + ], + "type": "interline_equation", + "content": "V _ {B} ^ {(l)} = \\mathcal {A} \\left(V _ {B} ^ {(l - 1)}\\right) + \\mathcal {C} \\left(V _ {B} ^ {(l - 1)}, F ^ {(l - 1)}\\right), \\tag {1}", + "image_path": "195fe1a5c8762a7054811e9a29e09ee6c6de504025432fed2630cdbc68d50846.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " refer to multi-headed self and cross attention respectively; " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "V_{B}^{(l)}\\in \\mathbb{R}^{B\\times C}" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " are the base visual prompts at layer " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "F^{(l)}\\in \\mathbb{R}^{H_lW_l\\times C_l}" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " is the flattened image feature at layer " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " denote image dimensions at that layer; " + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 47, + 459, + 287, + 518 + ], + "type": "text", + "content": " denotes the embedding dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "text", + "content": "After refining the prompts through multiple levels of transformer attention, a segmentation head, denoted as " + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "text", + "content": ", is defined. This segmentation head takes the final refined prompts, " + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "inline_equation", + "content": "V_{B}^{(L)}" + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "text", + "content": ", and the highest resolution image features, " + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "inline_equation", + "content": "F^{(L)}" + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "text", + "content": ", as input, producing per-pixel per-class predictions " + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "inline_equation", + "content": "O_{base} \\in \\mathbb{R}^{B \\times H \\times W}" + }, + { + "bbox": [ + 47, + 519, + 287, + 592 + ], + "type": "text", + "content": " (see Section 4.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "type": "text", + "content": "Novel Visual Prompts. After training on the base classes, we freeze all the layers of the model except the visual prompts representing the base classes. Given there are " + }, + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "type": "text", + "content": " novel classes, we add " + }, + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 594, + 287, + 688 + ], + "type": "text", + "content": " new visual prompts. We further add a causal unidirectional attention module that allows the novel prompts to be enhanced by the base prompts (see Section 4.2). The output is passed to a segmentation head to make predictions for the novel classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "The novel visual prompts are first refined by a unidirectional novel-to-base causal attention module. The enhanced" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 342, + 544, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 544, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 544, + 366 + ], + "type": "text", + "content": "novel prompts are then concatenated with the base prompts. Overall the concatenated prompts are refined by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 380, + 444, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 380, + 444, + 396 + ], + "spans": [ + { + "bbox": [ + 335, + 380, + 444, + 396 + ], + "type": "interline_equation", + "content": "V _ {N} ^ {(l)} = \\mathcal {C A} \\left(V _ {B} ^ {(l)}, V _ {N} ^ {(l)}\\right)", + "image_path": "39b65da0ec5a1a8625ebf2de1fb27c5129f27d184a37f6cfbda9136fedb7d0c3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 398, + 431, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 398, + 431, + 418 + ], + "spans": [ + { + "bbox": [ + 337, + 398, + 431, + 418 + ], + "type": "interline_equation", + "content": "V _ {A} ^ {(l)} = \\left[ V _ {B} ^ {(l)}, V _ {N} ^ {(l)} \\right]", + "image_path": "c490cedc67ccb8662dc2cfae649cc755d73406a06ca9134f95bcfdcdd499d222.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 337, + 419, + 515, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 419, + 515, + 436 + ], + "spans": [ + { + "bbox": [ + 337, + 419, + 515, + 436 + ], + "type": "interline_equation", + "content": "V _ {A} ^ {(l)} = \\mathcal {A} (V _ {A} ^ {(l - 1)}) + \\mathcal {C} (V _ {A} ^ {(l - 1)}, F ^ {(l - 1)}),", + "image_path": "6b5ddbc981257cb5d6cb0d87330e3ec7c464854656d2de190cb0df013c87cf3f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "V_{N}^{(l)}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": " are novel prompts at layer " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": " is the total number of novel class and " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": " is the embedding size; " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "V_{A}^{(l)}\\in \\mathbb{R}^{(B + N)\\times C}" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": " are all visual prompts including base and novel; " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "[.,.]" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": " is the concatenation operator. The operation " + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{CA}" + }, + { + "bbox": [ + 305, + 452, + 545, + 525 + ], + "type": "text", + "content": " denotes the causal novel-to-base attention and is explained in Section 4.2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 529, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 578 + ], + "type": "text", + "content": "Each of the " + }, + { + "bbox": [ + 304, + 529, + 545, + 578 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 529, + 545, + 578 + ], + "type": "text", + "content": " novel visual prompts (representing a novel class) is initialized using masked average global pooling of the k-shot support images corresponding to the class masked by the ground truth binary masks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 329, + 592, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 329, + 592, + 545, + 635 + ], + "type": "interline_equation", + "content": "V _ {n} ^ {(0)} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\sum_ {x , y} M _ {n} ^ {k} (x , y) F _ {n} ^ {k} (x , y)}{\\sum_ {x , y} M _ {n} ^ {k} (x , y)}, \\forall n \\in N \\tag {2}", + "image_path": "a2bae2d88ed21fb73df1dd089c0b9bd7625fccfa9edfd0a608bc89201ad50d7f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "V_{n}^{(0)} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": " is the initial novel visual prompt for novel class " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "F_{n}^{k}" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": " denotes the image features of " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": "-th shot for novel class; " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "M_{n}^{k}" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": " denotes the corresponding ground-truth binary mask; " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": " are the spatial locations; " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": " is the total number of novel classes and " + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 651, + 545, + 712 + ], + "type": "text", + "content": " is the total number of shots." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23473" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "content": "4.2. Novel-to-base Causal Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": "A key component of our proposed approach is the unidirectional novel-to-base prompt causal3 attention. The base visual prompts are learned with abundant examples; this is not the case for novel prompts. Learning novel prompts directly or having novel prompts impact the base ones could therefore lead to degradation of performance. However, having novel prompts aware of their base counterparts could be quite helpful. We hypothesize that causal attention between the novel visual prompts and base prompts could help contextualize novel class embeddings, reducing confusion with the base counterparts. Motivated by this intuition, we propose a novel-to-base causal attention layer that is repeated at each scale (or layer of the transformer decoder); see Fig. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "text", + "content": "The proposed causal attention module is a single layer cross-attention where the Key " + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "inline_equation", + "content": "K^{(l)}" + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "text", + "content": ", Value " + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "inline_equation", + "content": "V^{(l)}" + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "inline_equation", + "content": "Q^{(l)}" + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "text", + "content": " for layer " + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 258, + 287, + 294 + ], + "type": "text", + "content": " of the decoder are given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 300, + 282, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 300, + 282, + 318 + ], + "spans": [ + { + "bbox": [ + 51, + 300, + 282, + 318 + ], + "type": "interline_equation", + "content": "Q ^ {(l)} = V _ {N} ^ {(l)} W ^ {Q}; \\quad K ^ {(l)} = V _ {B} ^ {(l)} W ^ {K}; \\quad V ^ {(l)} = V _ {B} ^ {(l)} W ^ {V}", + "image_path": "31a983a295b52e203351cb1f6066bb337c6d4c4d7eb6943375da6be4a19954fa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "inline_equation", + "content": "W^{Q}, W^{K}, W^{V}" + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "content": " are weight matrices for " + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "content": " respectively. The parameters " + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "inline_equation", + "content": "W^{Q}, W^{K}, W^{V}" + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "content": " are shared at all the layers of this cross-attention module to reduce the number of trainable parameters during fine-tuning on novel classes to prevent overfitting and poor generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "text", + "content": "This uni-directional novel-to-base causal attention refines the novel visual prompts " + }, + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "inline_equation", + "content": "V_{N}^{(l)}" + }, + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "text", + "content": " at each decoder layer " + }, + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 420, + 238, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 420, + 238, + 435 + ], + "spans": [ + { + "bbox": [ + 95, + 420, + 238, + 435 + ], + "type": "interline_equation", + "content": "V _ {N} ^ {(l)} = \\operatorname {s o f t m a x} (Q ^ {(l)} (K ^ {(l)}) ^ {\\mathrm {T}}) V ^ {(l)}", + "image_path": "2ce0125fb0ec65329033ee3bb435bf466645a88e3126ecaa04e4bdaf035a2fa2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 440, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 287, + 453 + ], + "type": "text", + "content": "This results in a set of contextualized novel visual prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 467, + 162, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 162, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 162, + 480 + ], + "type": "text", + "content": "4.3. Segmentation Head" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": "Base Segmentation Head. The base segmentation head, " + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_B" + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": ", during the base training in the proposed model is relatively simple and takes the form of a three layer MLP that projects a refined prompt into, effectively, a class prototype; computes the (dot product) similarity of each pixel feature with that prototype, and then applies softmax along the channel / class dimension to obtain per-pixel and per-class probability:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 588, + 287, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 588, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 86, + 588, + 287, + 630 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} O _ {b a s e} = \\mathcal {H} _ {B} (V _ {B} ^ {(L)}, F ^ {(L)}) \\\\ = \\operatorname {M L P} \\left(V _ {B} ^ {(L)}\\right) \\cdot \\left[ F ^ {(L)} \\right] ^ {\\mathrm {T}}, \\tag {3} \\\\ \\end{array}", + "image_path": "4adb26de59afbe17b3178ad16a60af5d4183794f15b98e312eb9788d192c539e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 636, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 636, + 287, + 662 + ], + "type": "inline_equation", + "content": "O_{base} \\in \\mathbb{R}^{B \\times H \\times W}" + }, + { + "bbox": [ + 46, + 636, + 287, + 662 + ], + "type": "text", + "content": " and the class assignment can then be made by an argmax along the channel dimension." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 662, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 662, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 287, + 686 + ], + "type": "text", + "content": "Novel Segmentation Head. While we could, in principle, use the same form of segmentation head during few-shot" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "inference as that defined for the base classes (see Eq. 3), simply fine-tuning it for novel classes, this proves ineffective in practice. Specifically, fine-tuning the MLP leads to significant overfitting given the few examples and lack of base class data during this step. So, instead, we resort to a simpler form of the novel segmentation head, " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_N" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", where we only learn a residual to the MLP learned during base training (which we keep fixed), mainly:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 172, + 525, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 172, + 525, + 213 + ], + "spans": [ + { + "bbox": [ + 326, + 172, + 525, + 213 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} O _ {n o v e l} = \\mathcal {H} _ {N} (V _ {N} ^ {(L)}, F ^ {(L)}) \\\\ = \\left[ \\operatorname {M L P} _ {\\bullet} \\left(V _ {N} ^ {(L)}\\right) + W _ {N} \\right] \\cdot \\left[ F ^ {(L)} \\right] ^ {\\mathrm {T}}, \\\\ \\end{array}", + "image_path": "57a26d251619caac0b6455b5c8b32934f4737fb4c82f3c74fd8219dbfc108700.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "inline_equation", + "content": "O_{\\text{novel}}" + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": " is output for novel classes and " + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": " designates that parameters of MLP are frozen and only " + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "inline_equation", + "content": "W_N \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": " weight matrix is learned. This weight matrix is initialized in the same manner as novel prompts, in Eq. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 266, + 465, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 465, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 465, + 280 + ], + "type": "text", + "content": "4.4. Transductive Prompt Tuning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 285, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 344 + ], + "type": "text", + "content": "Recent approaches [4, 13] show the effectiveness of a good transductive setting for few-shot dense prediction tasks. Transductive approaches perform test-time optimization for each unlabeled test image and utilize specific characteristics and information of the test instances to improve prediction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 345, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 416 + ], + "type": "text", + "content": "To demonstrate the flexibility of our approach, in this section we outline the process of adapting our model for transductive inference, improving GFSS performance through transductive fine-tuning of the visual prompts. We build on the transductive losses proposed in DlaM [13] and we describe them here for completeness." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": "The losses are designed to maximize the mutual information between the learned features and the corresponding predictions on the test image, which is achieved through maximizing: " + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(O) - \\mathrm{H}(O|I)" + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "inline_equation", + "content": "I, O" + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": " are random variables associated with the pixels and predictions distributions respectively; " + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(O)" + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": " is the marginal and " + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(O|I)" + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": " is the conditional entropy. The conditional entropy " + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(O|I)" + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": " involved is given by sum of cross-entropy loss over the supervised support set and entropy of the predicted probability given the test image " + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 417, + 545, + 560 + ], + "type": "text", + "content": ". The marginal entropy is given by KL divergence loss relying on an estimated region proportion prior following [4]. Additional details are in [13]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 560, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 594 + ], + "type": "text", + "content": "Additionally, similar to [13], we use knowledge-distillation loss to preserve base class performance, given by:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 364, + 594, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 594, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 364, + 594, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D}} = \\mathrm {K L} \\left(O _ {\\text {b a s e}} ^ {\\text {n e w}} \\mid \\mid O _ {\\text {b a s e}} ^ {\\text {o l d}}\\right), \\tag {4}", + "image_path": "7020d644aa8b5271510778bb24eb4247ea9dfbba90f9181e7614c255aca27a35.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "O_{base}^{old}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " is the prediction probability of the base classes of the frozen model after base training; " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "O_{base}^{new}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " is the prediction probability of the base classes on the model with our additional components to handle novel categories that is being fine-tuned. These include the multi-scale causal novel-to-base cross attention and the novel prompts." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 684, + 539, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 684, + 539, + 696 + ], + "spans": [ + { + "bbox": [ + 317, + 684, + 539, + 696 + ], + "type": "text", + "content": "The overall transductive objective function is given by:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 347, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 347, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a n s .}} = \\alpha \\mathrm {H} (O | I) - \\mathrm {H} (O) + \\gamma \\mathcal {L} _ {\\mathrm {K D}}, \\tag {5}", + "image_path": "f8d760d58ec9f4057fc1ce768a18f571f0315341e3130b8a25e216d886cd9c82.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "3We use the term causal borrowing from its use in the look-ahead-attention-masking for decoder Transformers [38] (e.g., in language)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23474" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "text", + "content": " are balancing hyper-parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 168 + ], + "type": "text", + "content": "We noticed that applying transduction from the first iteration underperforms, likely due to a poor initial estimate of the marginal distribution. To address this, we only apply per-pixel cross-entropy loss for a set number of iterations and then incorporate the remaining transductive losses from Eq. 5 when we achieve a more accurate estimate of the marginal distribution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 171, + 128, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 128, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 128, + 184 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 190, + 238, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 238, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 238, + 204 + ], + "type": "text", + "content": "5.1. Dataset and Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "text", + "content": "Datasets. We evaluate our method on two few-shot segmentation benchmarks [36]: COCO- " + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "text", + "content": ". Following [13] we report average performance of our model on 10K test images for COCO- " + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "text", + "content": " and on all test images on PASCAL- " + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 209, + 287, + 280 + ], + "type": "text", + "content": ". The results are reported as an average across five different runs on four different splits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "content": "Data Pre-processing. We adopt the same setup as DIAM [13]. During base training, images containing both base and novel categories are kept, but the pixels corresponding to novel classes are relabeled as background. This approach, as noted by [13], presents challenges due to potential ambiguity when the base model predicts novel categories as background. In fine-tuning, only novel classes are labeled in the support set, with base classes marked as background. To avoid unfair shot increases, images with multiple novel categories are excluded from the support set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "type": "text", + "content": "Evaluation Protocol. We evaluate performance using the standard mean intersection-over-union (mIoU) metric. Following [13], the mean score for generalized few-shot segmentation is computed as the average between base and novel mIoUs. This approach addresses bias in datasets like PASCAL- " + }, + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 46, + 402, + 287, + 483 + ], + "type": "text", + "content": ", where base classes outnumber novel classes threefold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": "Implementation Details. We use PSPNet [51] with Resnet-50 [15] as feature extractor. We utilize MSDeformAttn-based pixel decoder [53] from Mask2Former [10]. The dimension " + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": " for visual prompts is chosen to be 256. Contrary to [10], the base model is trained with per-pixel cross-entropy loss. We use AdamW optimizer [25] with base learning rate of " + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": ", a weight decay of 0.05 and a batch size of 16. Following [13, 19], we train base model for 20 epochs for COCO- " + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": " and 100 epochs for PASCAL- " + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "During novel fine-tuning, we freeze the parameters of the feature extractor, pixel decoder, and base segmentation head, and only optimize the parameters of the visual prompts, novel segmentation head, and novel-to-base causal attention module. For the inductive case, the model is finetuned with per-pixel cross-entropy loss for 100 iterations. For the transductive setting, the model is fine-tuned for 40 iterations with only cross-entropy loss and then fine-tuned for 60 more iterations with full transductive losses mentioned in Eq. 5. The hyper-parameters " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " are set to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "be 100 and 25 respectively. We use AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " and a weight decay of 0.05 for fine-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 117, + 362, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 117, + 362, + 128 + ], + "spans": [ + { + "bbox": [ + 306, + 117, + 362, + 128 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "type": "text", + "content": "Comparison with the State-of-The-Art. Table 1 compares our approach in both inductive and transductive setting against the state-of-the-art. As observed, our model outperforms existing approaches. Interestingly, our approach, in an inductive setting (where optimization is performed only using a supervised few-shot support set, and evaluation is conducted on all test images), outperforms transductive approaches like DIaM [13] for both 1-shot and 5-shot cases on the COCO-20" + }, + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "type": "text", + "content": " dataset and for the 1-shot case on Pascal-5" + }, + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 132, + 545, + 275 + ], + "type": "text", + "content": ". This is notable despite transductive approaches requiring test-time optimization for each unlabeled test image separately, leading to significantly longer inference times." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 276, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 384 + ], + "type": "text", + "content": "For a fair comparison in the inductive setting, we finetuned DIaM without the transduction losses. We also finetuned POP [21] in our inductive setup. POP [21] randomly samples balanced base class image, making it easier for them to retain base class information and uses a form of transduction where their model relabels novel class pixels initially labelled as background. We observe that the performance of POP [21] degrades substantially, especially for the base classes in our inductive setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": "Our inductive approach has " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "5.78\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " higher mIoU for 1-shot and " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "11.16\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " higher mIoU for 5-shot on COCO-20" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " than inductive DIaM. For PASCAL-5" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": ", the improvement from inductive DIaM is " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "7.71\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " for 1-shot and " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "13.29\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " for 5-shot. In fact, compared to the best inductive GFSS approach BAM [19], our inductive setting achieves an improvement of approximately " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "2.78\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "7.59\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " for 1-shot and 5-shot respectively on COCO-20" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "5.24\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "12.32\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " for 1-shot and 5-shot respectively on PASCAL-5" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": ". Our method, in the inductive setting, even outperforms transductive DIaM [13] by " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "2.03\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "2.28\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " for 1-shot and 5-shot respectively on COCO-20" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " and by " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " for 1-shot on PASCAL-5" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "^i" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": "On applying transductive fine-tuning to our method, overall mIoU increases considerably, outperforming transductive DIaM [13] and POP [21] on both 1-shot and 5-shot inference for both datasets. The improvement over DIaM for " + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO - }20^{i}" + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": " for 1-shot and " + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "inline_equation", + "content": "3.93\\%" + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": " for 5-shot. For PASCAL- " + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": ", our transductive setting outperforms DIaM by " + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "inline_equation", + "content": "5.11\\%" + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "inline_equation", + "content": "3.15\\%" + }, + { + "bbox": [ + 304, + 529, + 545, + 611 + ], + "type": "text", + "content": " for 1 and 5-shots." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 621, + 369, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 369, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 369, + 632 + ], + "type": "text", + "content": "5.3. Ablation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 639, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 664 + ], + "type": "text", + "content": "We performed ablation studies (refer to Table 2) to explain our design choices and contribution of each components." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Visual Prompts Initialization. As seen in (1) and (2) of the Table 2, initializing novel prompts with masked mean pooling from the support set images improves both base and novel class accuracy compared to random initialization." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23475" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 121, + 71, + 473, + 208 + ], + "blocks": [ + { + "bbox": [ + 121, + 71, + 473, + 208 + ], + "lines": [ + { + "bbox": [ + 121, + 71, + 473, + 208 + ], + "spans": [ + { + "bbox": [ + 121, + 71, + 473, + 208 + ], + "type": "table", + "html": "
MethodLearningCOCO-20i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CAPL [36] (CVPR22)Inductive43.217.2125.2143.7111.0027.36
BAM [19] (CVPR22)Inductive49.8414.1632.0049.8516.6333.24
DIaM (w/o trans.)* [13] (CVPR23)Inductive42.6915.3229.0038.4720.8729.67
POP† [21] (CVPR23)Inductive30.389.6320.0024.5316.1920.36
Ours (w/o trans.)Inductive51.5518.0034.7851.5930.0640.83
RePRI [4] (CVPR21)Transductive5.624.745.188.858.848.85
DIaM [13] (CVPR23)Transductive48.2817.2232.7548.3728.7338.55
POP† [21] (CVPR23)Transductive54.7115.3135.0154.9029.9742.44
Ours (w/ trans.)Transductive53.8018.3036.0553.8131.1442.48
", + "image_path": "1b3ee162ac5f82561f2c4a98507bfaccdf461f5bb815cfb82aa291c7d35a6253.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 121, + 212, + 473, + 387 + ], + "blocks": [ + { + "bbox": [ + 121, + 212, + 473, + 387 + ], + "lines": [ + { + "bbox": [ + 121, + 212, + 473, + 387 + ], + "spans": [ + { + "bbox": [ + 121, + 212, + 473, + 387 + ], + "type": "table", + "html": "
MethodLearningPASCAL-5i
1-shot5-shot
BaseNovelMeanBaseNovelMean
CANet [49] (CVPR19)Inductive8.732.425.589.051.525.29
PFENET [35] (TPAMI20)Inductive8.322.675.508.831.895.36
PANET [40] (ICCV19)Inductive31.8811.2521.5732.9515.2524.1
SCL [47] (CVPR21)Inductive8.882.445.669.111.835.47
MiB [8] (CVPR20)Inductive63.808.8636.3368.6028.9348.77
CAPL [36] (CVPR22)Inductive64.8017.4641.1365.4324.4344.93
BAM [19] (CVPR22)Inductive71.6027.4949.5571.6028.9650.28
DIaM (w/o trans.)* [13] (CVPR23)Inductive66.7927.3647.0864.0534.5649.31
POP† [21] (CVPR23)Inductive46.6819.9633.3241.5036.2638.80
Ours (w/o trans.)Inductive74.5834.9954.7974.8650.3462.60
RePRI [4] (CVPR21)Transductive20.7610.5015.6334.0620.9827.52
DIaM [13] (CVPR23)Transductive70.8935.1153.0070.8555.3163.08
POP† [21] (CVPR23)Transductive73.9235.5154.7274.7855.8765.33
Ours (w/ trans.)Transductive76.3939.8358.1176.4256.1266.27
", + "image_path": "ea07f875f1c9e1d9f8c8a123e0e51b2941c5a76531ad1190c510f84238666a38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 462, + 287, + 535 + ], + "blocks": [ + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "text", + "content": "Table 1. Comparison to SoTA methods on COCO-20" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "text", + "content": " and PASCAL-5" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "text", + "content": ". The mIoU results are reported as the average across 5 different runs. All methods use ResNet-50 backbone. " + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "text", + "content": " POP [21] uses an easier and less practical GFSS setup, e.g., balanced base class training in addition to support set during novel fine-tuning and uses a form of transduction; inductive setting is implemented by us following our setup. Results of the rest of the methods are obtained from [13] except the inductive setting of DIA" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "inline_equation", + "content": "{M}^{*}" + }, + { + "bbox": [ + 46, + 396, + 548, + 452 + ], + "type": "text", + "content": ", which we implemented ourselves. The w/o trans., w/ trans. denote without and with transduction respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 462, + 287, + 535 + ], + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 535 + ], + "type": "table", + "html": "
Causal AttentionPrompt InitializationTransductionmIoU
BaseNovelMean
(1)NoneRandomNo50.1811.2230.70
(2)NoneMasked PoolingNo50.4311.3130.87
(3)First Layer OnlyMasked PoolingNo51.5312.2631.90
(4)Separate WeightsMasked PoolingNo51.0618.0534.56
(5)Shared WeightsRandomNo50.7517.4134.08
(6)Shared WeightsMasked PoolingNo51.5518.0034.78
(7)Shared WeightsMasked PoolingYes53.8018.3036.05
", + "image_path": "8c20aaf93c884ebf1a1bb863e732d9b0545485eacb951e2ab849abd5418eb97f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 540, + 287, + 562 + ], + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 562 + ], + "type": "text", + "content": "Table 2. Ablation Study on COCO- " + }, + { + "bbox": [ + 46, + 540, + 287, + 562 + ], + "type": "inline_equation", + "content": "{20}^{i}" + }, + { + "bbox": [ + 46, + 540, + 287, + 562 + ], + "type": "text", + "content": " . The 1-shot performane, averaged over all folds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 569, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 604 + ], + "type": "text", + "content": "This is more pronounced with novel-to-base causal attention (5)-(6), amplifying the performance gain from masked pooling initialization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Novel-to-base Causal Attention. As seen in (4)-(7) of Table 2, novel-to-base causal attention yields the most significant improvement in novel class accuracy compared to (1)-(2). This aligns with our hypothesis that uni-directional causal attention between rich base prompts (learned with many examples) and novel prompts (learned with few examples) increases separation among novel classes and pushes the feature space of base prompts away from the novel prompts. It is further confirmed by Figure 3 and Fig-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 462, + 556, + 602 + ], + "blocks": [ + { + "bbox": [ + 306, + 462, + 556, + 602 + ], + "lines": [ + { + "bbox": [ + 306, + 462, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 556, + 602 + ], + "type": "image", + "image_path": "0913caea56e1a20fe6b9e35147adf3fe3496a7b9ab746fe2ea30d39e2c1fddd2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 613, + 547, + 658 + ], + "lines": [ + { + "bbox": [ + 304, + 613, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 547, + 658 + ], + "type": "text", + "content": "Figure 3. Confusion Matrix for: (left) model with Novel-to-Base Causal Attention; (right) model without Novel-to-Base Causal Attention. Note that Novel-to-Base Causal Attention reduces confusion between novel and base categories (bottom-left block)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "ure 4. As observed in Figure 3, the novel-to-base causal attention leads to lower confusion between base and novel classes. Figure 4 shows a t-SNE [37] visualization of the features of base and novel visual prompts for 10 different" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23476" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 73, + 298, + 227 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 298, + 227 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 298, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 298, + 227 + ], + "type": "image", + "image_path": "23f666fdbb8253d19932f086feb2fb76eb6041f3a7d88db61e4da710a9e9d5f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 266 + ], + "type": "text", + "content": "Figure 4. TSNE visualization of the learned base and novel prompt features: (left) with Novel-to-Base Causal Attention; (right) without Novel-to-Base Casual Attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "content": "test images. As observed, the incorporation of novel-to-base causal attention leads to much clearer separation of visual prompt features, particularly for novel prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": "Multiscale Causal Attention. We also ablated the placement of the causal attention module. We observe that applying it only at the first layer (3) slightly improves novel mIoU compared to the model without causal attention. Yet, it is notably inferior to applying at every decoder layer (4)-(7). We could have opted for separate parameters for the causal attention module at each layer. However, as seen in (4), having separate weights offers little improvement to novel mIoU despite significant increase in total number of parameters to optimize. In fact, it decreases the base accuracy, leading to overall worse performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "type": "text", + "content": "Qualitative Results. Figure 5 compares the qualitative results of our model for 1-shot GFSS on Pascal- " + }, + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "type": "text", + "content": " in both inductive and transductive settings with the baseline lacking causal attention and DIaM [13]. Notably, for first row, both our inductive and transductive approaches predict more accurate segmentation of the cow compared to DIaM. In second row, the causal attention improves person segmentation followed by transduction. In the third row, DIaM completely misses the novel category chair. The causal attention allows our model to better differentiate the base class dining table from the novel class chairs. Transduction improves accuracy of the segmentation of dining table. In fourth row, we can observe that both DIaM and our model without causal attention misdetects novel class cat for cow at multiple locations. Causal attention allows our model to make a uniform and more accurate prediction of cat. In the fifth row, we observe that DIaM [13] completely misses detecting the base category potted plant and poorly segments person. The baseline without causal attention misdetects the potted plant with novel class bus and misdetects dining table with chair. Causal attention improves the accuracy which is further improved by transduction. Last row" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 557, + 281 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 557, + 281 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 557, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 557, + 281 + ], + "type": "image", + "image_path": "f675ac8b2fdd0e8dedcc775892190cc201f7b23bbfb609070141f9c1ff19f758.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 289, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 344 + ], + "type": "text", + "content": "Figure 5. Qualitative Results for 1-shot on Pascal- " + }, + { + "bbox": [ + 304, + 289, + 546, + 344 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 289, + 546, + 344 + ], + "type": "text", + "content": ". The left-most two columns show image and ground truth mask; (Third) Baseline without causal attention; (Fourth); Ours in inductive setting; (Fifth) Ours in transductive setting; (Last) DIaM [13]. Last row illustrates a failure." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": "demonstrates a case where our model performs worse than DlaM [13]. Again, we observe that adding causal attention reduces misclassification of person with car, although overall the segmentation of car is less accurate than DlaM [13]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "type": "text", + "content": "From the qualitative results, the effectiveness of novel-to-base causal attention at distinguishing base from novel classes becomes apparent. Our model in the inductive setting outperforms transductive DIaM in most cases, despite the inference time being faster." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 471, + 378, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 378, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 378, + 482 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 629 + ], + "type": "text", + "content": "In this work, we proposed a technique to learn visual prompts from few-shot examples of novel unseen cateogries to prompt a pre-trained segmentation model that not only generalizes well on the unseen novel class but strongly retains performance over the base classes on which the model was pre-trained. We achieve this by prompting the image features at multiple scales and proposing a uni-directional novel-to-base causal attention mechanism that enriches the novel visual prompts through interaction with the base visual prompts. Our proposed method can be applied in both inductive and transductive settings and achieves state-of-the-art performance on two benchmark few-shot segmentation datasets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 645, + 546, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 645, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgments and Disclosure of Funding. This work was partially funded by the Vector Institute for AI, Canada CIFAR AI Chair, NSERC CRC and NSERC DG. Hardware resources were partially provided by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute. Additional support was provided by JELF CFI grant and Digital Research Alliance of Canada under the RAC award." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23477" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 189 + ], + "type": "text", + "content": "[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, E. Rutherford, S. Cabi, T. Han, Z. Gong, S. Samangooei, M. Monteiro, J. L. Menick, S. Borgeaud, A. Brock, A. Nematzadeh, S. Sharifzadeh, M. Binkowski, R. Barreira, O. Vinyals, A. Zisserman, and K. Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 191, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 453 + ], + "type": "text", + "content": "[2] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S.v.ArxM.S.BernsteinA.Bosselut J.BohgE.Brunskill, E. Brynjolfsson, S. Buch, D.Card, R.Castellon, N.Chatterji, A. Chen, K.Creel, J.Q.Davis,D.Demszky,C.Donahue,M.Doumbouya,S.Ermon E.Durmus,J.Etchemendy, K.Ethayarajh, L.Fei-Fei, C.FinnT.GaleL.GillespieK. Goel, N. Goodman, S. Grossman, N. Guha, T. Hashimoto, P. Henderson, J. Hewitt, D.E.HoJ.HongK.HsuJ.Huang, T.IcardS.JainD.JurafskyP.KalluriS.KaramchetiG. KeelingF.KhaniO.KhattabP.W.KohM.KrassR. Krishna,R.Kuditipudi,A.Kumar,F.Ladhak,M.Lee,T. LeeJ.LeskovecI.LeventX.L.LiX.LiT.MaA.MlikC.D.ManningS.Mirchandani,E.MitchellZ.MunyikwaS.NairD.Narayanan A.NarayanB.NewmanA. NieJ.C.Niebles,H.NilforoshanJ.Nyarko,G.OgutI. PapadimitriouJ.S.ParkC.PiechE.PortelanceC.Potts, A.RaghunathanR.ReichH.RenF.RongY.RoohaniC. RuizJ.RyanC.ReD.SadighS.SagawaK.Santhanam, A.ShihK.SrinivasanA.TamkinR.TaoriA.W.Thomas F.TramereR.E.WangW.WangB.WuJ.WuY.WuS.M XieM.YasunagaJ.YouM.ZahariaM.ZhangT.Zhang X.ZhangY.ZhangL.ZhengK.Zhou and P.Liang.On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 287, + 499 + ], + "type": "text", + "content": "[3] Malik Boudiaf, Imtiaz Ziko, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Information maximization for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 33:2445-2457, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 287, + 555 + ], + "type": "text", + "content": "[4] M. Boudiaf, H. Kervadec, Z. Masud, P. Piantanida, I. Ben Ayed, and J. Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13979-13988, 2021. 1, 2, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 557, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 655 + ], + "type": "text", + "content": "[5] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, G. Sastry P. Shyam, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33: 1877-1901, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 657, + 287, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 657, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 53, + 657, + 287, + 701 + ], + "type": "text", + "content": "[6] A. Bulat, R. Guerrero, B. Martinez, and G. Tzimiropoulos. Fs-detr: Few-shot detection transformer with prompting and without re-training. In International Conference on Computer Vision (ICCV), pages 11793-11802, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 702, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 702, + 287, + 713 + ], + "type": "text", + "content": "[7] T. Cao, M. Law, and S. Fidler. A theoretical analysis of the" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "number of shots in few-shot learning. International Conference on Learning Representations (ICLR), 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 149 + ], + "type": "text", + "content": "[8] F. Cermelli, M. Mancini, S.R. Bulo, E. Ricci, and B. Caputo. Modeling the background for incremental learning in semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233–9242, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "type": "text", + "content": "[9] W. Chen, Y. Liu, Z. Kira, Y. Frank Wang, and J. Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[10] B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 241, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 275 + ], + "type": "text", + "content": "[11] G. S. Dhillon, P. Chaudhari, A. Ravichandran, and S. Soatto. A baseline for few-shot image classification. In International Conference on Learning Representations (ICLR), 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[12] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning (ICML), 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "text", + "content": "[13] S. Hajimiri, M. Boudiaf, I. B. Ayed, and J. Dolz. A strong baseline for generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11269-11278, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 365, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 399 + ], + "type": "text", + "content": "[14] B. Hariharan and R. B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[15] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "text", + "content": "[16] M. Jia, L. Tang, B. Chen, C. Cardie, S. Belongie, B. Hariharan, and S. Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 489, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 522 + ], + "type": "text", + "content": "[17] B. Kang, Z. Liu, X. Wang, F. Yu, J. Feng, and T. Darrell. Few-shot object detection via feature reweighting. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[18] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, A. C. Berg S. Whitehead, W. Lo, P. Dollár, and R. Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 568, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 613 + ], + "type": "text", + "content": "[19] C. Lang, G. Cheng, B. Tu, and J. Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8067, 2022. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 645 + ], + "type": "text", + "content": "[20] X. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[21] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11319-11328, 2023. 2, 6, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23478" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[22] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[23] Y. Liu, J. Lee, M. Park, S. Kim, E. Yang, S. J. Hwang, and Y. Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 196 + ], + "type": "text", + "content": "[24] Y. Liu, X. Zhang, S. Zhang, and X. He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "type": "text", + "content": "[25] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 263 + ], + "type": "text", + "content": "[26] J. Min, D. Kang, and M. Cho. Hypercorrelation squeeze for few-shot segmentation. In International Conference on Computer Vision (ICCV), pages 6941-6952, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 287, + 297 + ], + "type": "text", + "content": "[27] A. Nichol, J. Achiam, and J. Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 299, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 353 + ], + "type": "text", + "content": "[28] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 287, + 399 + ], + "type": "text", + "content": "[29] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8412-8422, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 400, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 400, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 287, + 443 + ], + "type": "text", + "content": "[30] L. Qiao, Y. Shi, J. Li, Y. Wang, T. Huang, and Y. Tian. Transductive episodic-wise adaptive metric for few-shot learning. In International Conference on Computer Vision (ICCV), pages 3603-3612, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 445, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 509 + ], + "type": "text", + "content": "[31] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 287, + 544 + ], + "type": "text", + "content": "[32] S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 287, + 589 + ], + "type": "text", + "content": "[33] M. Ren, E. Triantafillou, S. Ravi, J.Snell, K. Swersky, J. B. Tenenbaum, H. Larochelle, and R. S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 623 + ], + "type": "text", + "content": "[34] J. Snell, K. Swersky, and R. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "type": "text", + "content": "[35] Z. Tian, H. Zhao, M. Shu, Z. Yang, R. Li, and J. Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(2):1050-1065, 2020. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[36] Z. Tian, X. Lai, L. Jiang, S. Liu, M. Shu, H. Zhao, and J. Jia. Generalized few-shot semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11563-11572, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "text", + "content": "[38] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[39] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 3630-3638, 2016. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "text", + "content": "[40] K. Wang, J. Liew, Y. Zou, D. Zhou, and J. Feng. PANet: Few-shot image semantic segmentation with prototype alignment. In International Conference on Computer Vision (ICCV), pages 9197-9206, 2019. 1, 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[41] X. Wang, T. E Huang, T. Darrell, J. E. Gonzalez, and F. Yu. Frustratingly simple few-shot object detection. In International Conference on Machine Learning (ICML), 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "type": "text", + "content": "[42] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 306, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 306, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 306, + 545, + 358 + ], + "type": "text", + "content": "[43] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "text", + "content": "[44] X. Yan, Z. Chen, A. Xu, X. Wang, X. Liang, and L. Lin. Meta R-CNN: Towards general solver for instance-level low-shot learning. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "type": "text", + "content": "[45] B. Yang, C. Liu, B. Li, J. Jiao, and Q. Ye. Prototype mixture models for few-shot semantic segmentation. In European Conference on Computer Vision (ECCV), pages 763-778, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[46] X. Yang, B. Wang, K. Chen, X. Zhou, S. Yi, W. Ouyang, and L. Zhou. BriNet: Towards bridging the intra-class and inter-class gaps in one-shot segmentation. In British Machine Vision Conference (BMVC), 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[47] B. Zhang, J. Xiao, and T. Qin. Self-guided and cross-guided learning for few-shot segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8321, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "text", + "content": "[48] C. Zhang, G. Lin, F. Liu, J. Guo, Q. Wu, and R. Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In International Conference on Computer Vision (ICCV), pages 9587-9595, 2019, 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 592, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 646 + ], + "type": "text", + "content": "[49] C. Zhang, G. Lin, F. Liu, R. Yao, and C. Shen. CANet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217–5226, 2019. 1, 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[50] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[51] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23479" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 184 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[52] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[53] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representations (ICLR), 2021. 6" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23480" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_content_list.json b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97bab2a4cb3ec6710fc391980ab322331fe9fc34 --- /dev/null +++ b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_content_list.json @@ -0,0 +1,1508 @@ +[ + { + "type": "text", + "text": "Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjin Hou $^{1,4}$ , Shiming Chen $^{2*}$ , Shuhuang Chen $^{1}$ , Ziming Hong $^{3}$ , Yan Wang $^{4}$ , Xuetao Feng $^{4}$ , Salman Khan $^{2,5}$ , Fahad Shahbaz Khan $^{2,6}$ , Xinge You $^{1*}$", + "bbox": [ + 114, + 202, + 854, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Huazhong University of Science and Technology (HUST), China", + "bbox": [ + 223, + 238, + 748, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Mohamed bin Zayed University of AI $^{3}$ The University of Sydney", + "bbox": [ + 212, + 256, + 754, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Alibaba Group 5Australian National University 6Linkoping University", + "bbox": [ + 194, + 273, + 794, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{houwj17, gchenshiming}@gmail.com youxg@mail.hust.edu.cn", + "bbox": [ + 227, + 294, + 736, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Zero-shot learning (ZSL) learns a generator to synthesize visual samples for unseen classes, which is an effective way to advance ZSL. However, existing generative methods rely on the conditions of Gaussian noise and the predefined semantic prototype, which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). To address this issue, we propose a novel Visual-Augmented Dynamic Semantic prototype method (termed VADS) to boost the generator to learn accurate semantic-visual mapping by fully exploiting the visual-augmented knowledge into semantic conditions. In detail, VADS consists of two modules: (1) Visual-aware Domain Knowledge Learning module (VDKL) learns the local bias and global prior of the visual features (referred to as domain visual knowledge), which replace pure Gaussian noise to provide richer prior noise information; (2) Vision-Oriented Semantic Updation module (VOSU) updates the semantic prototype according to the visual representations of the samples. Ultimately, we concatenate their output as a dynamic semantic prototype, which serves as the condition of the generator. Extensive experiments demonstrate that our VADS achieves superior CZSL and GZSL performances on three prominent datasets and outperforms other state-of-the-art methods with averaging increases by $6.4\\%$ , $5.9\\%$ and $4.2\\%$ on SUN, CUB and AWA2, respectively.", + "bbox": [ + 76, + 376, + 472, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 801, + 209, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-shot learning [33], which transfers knowledge from seen classes to unseen classes, has garnered much attention recently. By establishing interactions between visual fea", + "bbox": [ + 75, + 828, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tures and semantic prototypes (also referred to as attribute vectors, side information, or semantic embeddings [47]), generative ZSL methods exhibit impressive performance, demonstrating the potential of feature synthesis. One of the most successful frameworks is the conditional generative adversarial network (GAN) [17]. The main idea of generative ZSL methods is to align semantic prototypes and visual features to synthesize feature of unseen classes. Recent emerging studies have either designed more effective frameworks [7, 19, 20, 31, 46] or addressed more specific issues related to visual-semantic alignment [4, 8, 12, 43, 53]. These methods have achieved significant improvements.", + "bbox": [ + 500, + 345, + 893, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, these methods rely on the conditions of Gaussian noise and the predefined semantic prototype (referred to as the static semantic prototype), which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). Fig. 1 illustrates these issues: (1) The noise is sampled from a Gaussian distribution $\\mathcal{N}(0,1)$ , which lacks the dataset-specific visual prior knowledge (e.g., global visual information \"flying\" and \"still\" and background information \"sky\" and \"grass\"). As a result, the domain knowledge shared between seen and unseen classes cannot be utilized for feature synthesis of unseen classes, limiting the knowledge transfer. (2) The predefined semantic prototype fails to characterize each instance well. For example, the attributes \"wing black\", \"breast white\" and \"bill orange\" of the Laysan Albatross are not fixed on different images. Due to these limitations, the visual features synthesized by existing works [4, 6, 21, 24, 31, 43, 46] struggle to represent the distribution of real features, leading to poor generalization to unseen classes, as shown in Fig. 1(b). More intuitively, as shown in Fig. 1(d), the features of unseen classes synthesized by these methods are confusing, resulting in the decision boundary overfitting to the seen classes.", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 94, + 886, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23627", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38bcce7d3c04da7ae89f2b372b3204866c504d38b6464ae121b7f8bfd8236ab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 92, + 290, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67cb4509fa7045f419b184adb3d350086b1557ff3f285d5661caba16a6303bf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 93, + 465, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a6fb2437c9d2299a0ace19e0727ffeae5f5c078a27dd5f88ce67f95a73e7bdc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 93, + 638, + 226 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a52a90e9f396589cc762815c104667531bcc7786000477168e194edbb475667d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 93, + 867, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4af521c52b95043c9d12143da9279ff25d0005307d0bfcbd985bfd82b4118f60.jpg", + "image_caption": [ + "(a) Input image" + ], + "image_footnote": [], + "bbox": [ + 101, + 228, + 289, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/231add22ada34f480de0f3711fa79bdcd8426404e547fac4b13411f49c65c7c8.jpg", + "image_caption": [ + "(b) Existing works" + ], + "image_footnote": [], + "bbox": [ + 290, + 228, + 465, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b53f1f0ecec2b8820955ccf422c0e2db2770a6e4ea9db8c81760100a7b26a62f.jpg", + "image_caption": [ + "(c) Our VADS", + "Figure 1. An illustration of the core idea of our method. (a) The semantic prototype (i.e., attribute) of different images of the same category is not fixed, so the predefined semantic prototype is inaccurate in characterizing each instance. (b) Most existing works utilize Gaussian noise and the predefined semantic prototype as conditions to train a semantic→visual generator on seen classes, which fails to generalize to unseen classes. (c)(d) Our method incorporates rich visual prior with an updated semantic prototype to construct a visual-augmented dynamic semantic prototype of each instance, empowering the generator to synthesize features that faithfully represent the real distribution of unseen classes. Thus, our method achieves better generalization on seen and unseen classes than existing works (e.g., CLSWGAN [46])." + ], + "image_footnote": [], + "bbox": [ + 465, + 228, + 638, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c2fde26257060e3e09fd9027fd21b60e4f6c3677b84083f55b79436cb3a1525.jpg", + "image_caption": [ + "(d) Feature visualization" + ], + "image_footnote": [], + "bbox": [ + 643, + 237, + 867, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Drawing inspiration from image captioning [36], which highlights the generalization of instance-conditional learning, we aim to fully leverage the visual-augmented knowledge into semantic conditions to tackle the aforementioned challenges. On the one hand, we can exploit rich domain visual prior knowledge, serving as a prior noise, to enhance the adaptation and generalization of models [18, 51, 54]. On the other hand, we can update the predefined semantic prototype to align visual representations based on visual features. As such, incorporating richer and more accurate visual information acts as the semantic condition to train an instance-conditional generative model, which is optimized to characterize each instance (more robust to class shift) rather than to serve only for specific classes. Accordingly, the generative model can synthesize features of unseen classes closer to the real ones, facilitating the classifier in learning an appropriate decision boundary (see the right of Fig. 1(d)).", + "bbox": [ + 75, + 489, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an approach called Visual-Augmented Dynamic Semantic prototype (VADS) to improve generative ZSL methods. Specifically, VADS consists of two learnable modules: a Visual-aware Domain Knowledge Learning module (VDKL) and a Vision-Oriented Semantic Updation module (VOSU). The VDKL explores domain visual prior knowledge derived from visual information, which provides richer information for representing instances. The VOSU predicts instance-level semantics", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through visual $\\rightarrow$ semantic mapping, guiding the updation of the predefined semantic prototype and promoting accurate semantic prototype learning. Finally, the extracted visual prior and the updated semantic prototype are concatenated as a visual-augmented dynamic semantic prototype, which serves as the condition of the generator during training and feature synthesis, as illustrated in Fig. 1(c). Extensive experiments demonstrate the effectiveness of our VADS.", + "bbox": [ + 496, + 489, + 890, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 517, + 616, + 843, + 630 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a Visual-Augmented Dynamic Semantic prototype (VADS) to enhance the generalization of generative ZSL methods, facilitating substantial knowledge transfer.", + "- We devise the VDKL to leverage domain visual prior knowledge from visual features and design the VOSU to dynamically update the predefined semantic prototype. Their outputs together serve as the generator's conditions, providing richer and more accurate visual information.", + "- We conduct extensive experiments on AWA2 [47], SUN [34] and CUB [44] datasets. The comprehensive results demonstrate that visual prior knowledge significantly improves the generalization of generative ZSL methods, i.e., average improvements of the harmonic mean over existing generative methods (e.g., f-CLSWGAN [46], TFVAEGAN [31] and FREE [7]) $6.4\\%$ , $5.9\\%$ and $4.2\\%$ on SUN, CUB and AWA2, respectively." + ], + "bbox": [ + 500, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23628", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embedding-based Zero-Shot Learning. Embedding-based ZSL methods are one of the mainstream branches that project visual information into semantic space to align with semantic prototypes. Earlier works [25, 39, 48] directly mapped global visual features to semantic space, failing to capture local discriminative representation, resulting in suboptimal ZSL performance. Also, embeddings are learned only in seen classes, leading to inevitable seen class bias. In this regard, some studies [9, 10, 22] have attempted to use calibration loss to balance the prediction results between seen and unseen classes. Recently, attention mechanisms [41] have emerged with surprising localization abilities, so semantic-guided methods [9, 11, 29, 30, 32, 42, 48, 49] learn to discover attribute-related local regions, providing more accurate inter-class discrimination. Among these methods, APN [49] proposed an attribute prototype network to learn local features, and DPPN [42] updated attribute and category prototypes. Inspired by their work, we introduce a dynamic semantic prototype for generative ZSL methods.", + "bbox": [ + 76, + 119, + 472, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Zero-Shot Learning. Generative ZSL methods learn semantic $\\rightarrow$ visual mapping to synthesize unseen class features, effectively alleviating the lack of unseen class data. Consequently, the quality of synthesized features, which preserves visual-semantic correspondence, plays a crucial role in classification. Thus, TF-VAEGAN [31] forced semantic alignment at all stages, and FREE [7] fine-tuned visual features to address cross-dataset biases. CE-GZSL [20] and ICCE [24] projected visual features into the latent space for classification. However, these methods constructed projection spaces on seen classes, resulting in inferior generalization ability on unseen classes. Moreover, they uniformly utilize the predefined semantic prototype as a condition, making it difficult to achieve accurate visual-semantic alignment. The method most related to ours is DSP [12], which updates the prototype by simply adding the evolved and predefined semantic prototype.", + "bbox": [ + 76, + 411, + 472, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large-Scale Vision-Language Models Generalization. Vision-language models like CLIP [35], pre-trained on large-scale image-text pairs, have demonstrated significant potential for downstream tasks. When performing zero-shot recognition, the class prompts are input into the text encoder to obtain the classification weights, and the cosine similarity between the test image and the weights determines the resulting classification score. It is different from the classical ZSL methods [9, 25, 39, 42, 48, 49]. Recent research has focused on improving the generalization to unseen classes, with several previous works proposing prompt learning [1, 51, 54]. Motivated by optimizing visual conditional prompts, we introduce visual-aware domain knowledge learning into generative ZSL methods, facilitating knowledge transfer to unseen classes.", + "bbox": [ + 76, + 674, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Visual-Augmented Dynamic Semantic Prototype Method", + "text_level": 1, + "bbox": [ + 498, + 89, + 890, + 125 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the framework of our VADS. Next, we first present the problem formulation and briefly review the generative ZSL model. Then, we introduce the detailed design of our method.", + "bbox": [ + 496, + 132, + 893, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Conventional zero-shot learning (CZSL) recognizes unseen classes in the inference stage. Generalized zero-shot learning (GZSL) recognizes both seen and unseen classes. Both settings generalize from seen data $\\mathcal{D}^s$ to unseen domains $\\mathcal{D}^u$ . $\\mathcal{D}^s = \\{(x_i^s, y_i^s) | x_i^s \\in \\mathcal{X}^s, y_i^s \\in \\mathcal{Y}^s\\}_{i=1}^{N_s}$ , where $N_s$ is the sample number of seen classes, $x_i^s$ is a feature vector in $\\mathcal{X}^s$ and $y_i^s$ is the corresponding lable from $\\mathcal{Y}^s$ . The $\\mathcal{D}^s$ is split into a training set $\\mathcal{D}_{tr}^s$ and a testing set $\\mathcal{D}_{te}^s$ following Xian et al. [47]. Similarly, $\\mathcal{D}^u = \\{(x_i^u, y_i^u) | x_i^u \\in \\mathcal{X}^u, y_i^u \\in \\mathcal{Y}^u\\}_{i=1}^{N_u}$ , where $x_i^u$ is a feature vector in $\\mathcal{X}^u$ and $y_i^u$ is the lable from $\\mathcal{Y}^u$ . $\\mathcal{Y}^s$ and $\\mathcal{Y}^u$ are disjoint. Define attribute semantic prototypes $\\mathcal{A} = \\mathcal{A}^s \\cup \\mathcal{A}^u$ , corresponding to each category, as a bridge to transfer knowledge from seen classes to unseen classes. In this paper, we dynamically update $\\mathcal{A}$ to learn accurate visual-semantic alignment.", + "bbox": [ + 496, + 193, + 893, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generative ZSL Model", + "text_level": 1, + "bbox": [ + 498, + 443, + 714, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the generative ZSL methods is to learn a semantic $\\rightarrow$ visual generative model $(G)$ on seen classes and then use it to synthesize samples of unseen classes to train a classifier. Existing methods use Gaussian noise and the predefined semantic prototype as input conditions to supervise $G$ synthetic features (i.e., $\\mathcal{A} \\times \\mathcal{Z} \\rightarrow \\hat{X}$ ). In our method, $G$ represents an off-the-shelf CLSWGAN [46], which contains a generator and a discriminator. We develop the dynamic semantic prototype as a condition, allowing $G$ to characterize more accurate visual-semantic relationships.", + "bbox": [ + 496, + 465, + 893, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual-aware Domain Knowledge Learning (VDKL)", + "text_level": 1, + "bbox": [ + 498, + 626, + 890, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing inspiration from previous prompt learning [51, 54], we exploit the rich information in visual features to assist in synthesizing features. VDKL is a data-efficient module allowing the visual features to be used to improve generalization. As shown in Fig. 2, we design a Visual Encoder (VE) and a Domain Knowledge Learning network (DKL). First, the VE encodes visual features into a latent feature $l$ and a latent code $z$ . The latent feature enables inter-class alignment of visual features, and latent code is subsequently confined to a prior distribution $Z$ . The optimization of the VE is achieved via contrastive loss [13] and evidence-lower bound given by the equation as follows:", + "bbox": [ + 496, + 664, + 893, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = \\mathbb {E} [ \\log \\frac {\\exp \\left(l _ {i} ^ {T} l ^ {+} / \\tau\\right)}{\\exp \\left(l _ {i} ^ {T} l ^ {+} / \\tau\\right) + \\sum_ {k = 1} ^ {K} \\exp \\left(l _ {i} ^ {T} l _ {k} ^ {-} / \\tau\\right)} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 867, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23629", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d06d19ef86e59dd2205450a03a94b64dc3fbbd5037dfd1ed8b67bb677ad80d91.jpg", + "image_caption": [ + "Figure 2. The architecture of our proposed VADS. It consists of two learnable modules: a Visual-Oriented Semantic updation module (VOSU) and a Visual-aware Domain Knowledge Learning module (VDKL). First, we obtain the prior distribution $\\mathbf{Z}$ by the Visual Encoder $(VE)$ . Following this, the Domain Knowledge Learning network $(DKL)$ transforms $\\mathbf{Z}$ into a local bias $\\mathbf{b}$ , which is subsequently added to global learnable prior vectors $(\\mathbf{p})$ to construct the domain visual prior noise (i.e., $\\mathbf{Z}'$ ). At the bottom, VOSU notably updates the semantic prototype in two stages (depicted by the blue and green arrows). Finally, the visual prior noise and the updated semantic prototype together form a dynamic semantic prototype, used for the reconstruction of features by the generator." + ], + "image_footnote": [], + "bbox": [ + 109, + 92, + 857, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k l} = K L (V E (x) \\| p (z)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 588, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l^{+}$ and $l_{k}^{-}$ represent positive and negative latent features, $\\tau$ is a temperature parameter set as 0.15, $K$ is the class numbers, $KL$ denotes the Kullback-Leibler divergence and $p(z)$ is a prior distribution that is assumed to be $\\mathcal{N}(0,1)$ .", + "bbox": [ + 75, + 613, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further utilize visual prior knowledge during the training and synthesis stages, we propose a Domain Knowledge Learning network (DKL) to obtain a local bias $\\pmb{b}$ of visual features (i.e., $\\pmb{b} = DKL(\\pmb{Z})$ ). Additionally, we employ a learnable prior vector $\\pmb{p}$ to capture global visual information ( $\\pmb{p}$ is randomly initialized). Subsequently, we obtain domain-specific visual prior noise as follows:", + "bbox": [ + 75, + 674, + 470, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} ^ {\\prime} = \\alpha \\cdot (\\boldsymbol {b} + \\boldsymbol {p}) + (1 - \\alpha) \\cdot \\text {n o i s e} (\\mathbf {0}, \\mathbf {1}), \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 792, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the noise $(0,1)$ represents Gaussian noise aimed at enhancing diversity in synthesis, $\\alpha$ is the combination coefficient set as 0.9. Through this operation, we argue that $Z^{\\prime}$ includes rich domain visual knowledge and feeds it into the generator to provide instance conditions, promoting the", + "bbox": [ + 76, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generator learning and utilizing it for feature synthesis of unseen classes. Note that unseen class samples are unavailable in the feature synthesis stage, so we randomly sample Gaussian noise input to $DKL$ , transferring the domain knowledge acquired from seen classes to unseen classes.", + "bbox": [ + 496, + 589, + 893, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Visual-Oriented Semantic updation (VOSU)", + "text_level": 1, + "bbox": [ + 498, + 676, + 877, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the predefined semantic prototype struggles to represent each visual sample accurately, so we propose a Visual-Oriented Semantic Updation module (VOSU), optimizing the semantic prototype dynamically. Our semantic prototype updation involves a two-stage process. In the first stage, we feed visual features $x_{s}$ into the Visual-Semantic Prediction network (VSP) to generate a predictive semantic $\\hat{a}$ that explicitly captures specific visual patterns of the target image. Then, the predefined semantic prototype is input into a Semantic Updation Mapping network (SUM) to learn an updated semantic $\\dot{a}$ . This mapping can be expressed as:", + "bbox": [ + 496, + 700, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {a} = S U M (a). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 880, + 890, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23630", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To maintain the attribute information of the prototype and integrate the visual information, we jointly optimize them by the cross-entropy loss $\\mathcal{L}_{ce}$ . $\\mathcal{L}_{ce}$ is defined as:", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(V S P \\left(x _ {i}\\right) ^ {T} \\dot {a} ^ {y}\\right)}{\\sum_ {\\bar {c} \\in C ^ {s} \\cup u} \\exp \\left(V S P \\left(x _ {i}\\right) ^ {T} \\dot {a} ^ {\\bar {c}}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 148, + 468, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the batch numbers. Accordingly, the updated semantic prototype incorporates rich visual information. In the second stage, we employ SUM to update $\\dot{a}$ during the conditional generative model training and use it as a condition to learn together with the generator $G$ and the discriminator $D$ . This implementation facilitates dynamic updation and accurate visual-semantic matching. To this end, we propose the semantic consistency loss $\\mathcal{L}_{sc}$ as follows:", + "bbox": [ + 76, + 199, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s c} = \\mathbb {E} \\left[ \\| S U M (\\dot {a}) - \\dot {a} \\| _ {1} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 333, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the first stage leverages visual features to assist semantic updation, and the second stage dynamically updates the prototype of each sample. Then, we concatenate the updated semantic prototype with the visual prior noise $Z^{\\prime}$ , called the dynamic semantic prototype, which serves as the condition for the generator, as depicted in Fig. 2.", + "bbox": [ + 76, + 362, + 468, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Objective and Inference", + "text_level": 1, + "bbox": [ + 76, + 462, + 359, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VADS Objective Loss Function. Overall, the objective loss function of VADS is:", + "bbox": [ + 76, + 486, + 468, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {G} + \\lambda_ {\\text {c o n}} \\mathcal {L} _ {\\text {c o n}} + \\lambda_ {k l} \\mathcal {L} _ {k l} + \\lambda_ {s c} \\mathcal {L} _ {s c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 529, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_G$ is the loss of conditional generative model G, $\\lambda_{con}$ , $\\lambda_{kl}$ and $\\lambda_{sc}$ are the hyper-parameter to balance each loss term. To fully validate our method, by using this loss, we train on various mainstream generative models (e.g., CLSWAGN [46], TFVAEGAN [31], and FREE [7]). Next, we illustrate feature synthesis and classifier training.", + "bbox": [ + 76, + 556, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual-Augmented Feature Synthesis for Unseen Classes. To fully utilize the visual knowledge and accurate semantic prototypes, we sample Gaussian noise input $DKL$ to obtain the prior noise $\\mathbf{Z}^{\\prime}$ (i.e., Eq. (3)) and use SUM to update the semantic prototypes of unseen classes (i.e., $\\dot{a}_u = SUM(\\hat{a}_u)$ ). They serve as conditions to synthesize visual samples closer to the real features for training the classifier. The form can be written as:", + "bbox": [ + 76, + 662, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {u} = \\mathrm {G} \\left(\\boldsymbol {Z} ^ {\\prime}, \\dot {a} _ {u}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 794, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\dot{a}_u$ is updated semantic prototype of the unseen classes and $\\hat{x}_u$ is the synthesized features of unseen classes.", + "bbox": [ + 76, + 825, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ZSL Classifier Training and Inference. After synthesizing features, we input the seen class training features and", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "synthesized unseen class features into $VE$ to extract latent features and concatenate them to enhance the original features, alleviating cross-dataset bias [7]. Then, we train a CZSL classifier using enhanced-synthetic features (i.e., $f_{CZSL}:\\mathcal{X}\\to \\mathcal{Y}^u$ ) and train a GZSL classifier using enhanced seen class training features and enhanced-synthetic features (i.e., $f_{GZSL}:\\mathcal{X}\\to \\mathcal{Y}^s\\cup \\mathcal{Y}^u$ ). Finally, we perform inference using the test sets $\\mathcal{D}_{te}^{s}$ and $\\mathcal{D}^u$ .", + "bbox": [ + 498, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 227, + 630, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 252, + 689, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Datasets. We conduct extensive experiments on three prominent ZSL benchmark datasets: Animals with Attributes 2 (AWA2 [47]), SUN Attribute (SUN [34]) and Caltech-USCD Birds-200-2011 (CUB [44]). We follow the Proposed Split (PS) setting [47] to split each dataset into seen and unseen classes, as detailed in Tab. 2.", + "bbox": [ + 498, + 276, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. During inference (i.e., performing CZSL and GZSL classification), we follow the evaluation protocols in [47]. In the CZSL setting, we calculate the average per-class Top-1 accuracy of unseen classes, denoted as $\\mathbf{A}$ . For the GZSL scenario, we measure the Top-1 accuracy of the seen and unseen classes, defined as $\\mathbf{S}$ and $\\mathbf{U}$ , respectively. We also compute the harmonic mean, defined as $\\mathbf{H} = (2 \\times \\mathbf{S} \\times \\mathbf{U}) / (\\mathbf{S} + \\mathbf{U})$ .", + "bbox": [ + 498, + 367, + 890, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We follow PSVMA [26] using the ViT-Base Backbone [15] without fine-tuning as the feature extractor, obtaining 768-dimensional visual features for all samples. The global prior $\\pmb{p}$ has the same dimension as the semantic prototype. We set the mini-batch to 64, 128 and 128 for AWA2, SUN and CUB, respectively. We use the Adam optimizer [23] with $\\beta_{1} = 0.5$ , $\\beta_{2} = 0.999$ , and set the initial learning rate to 0.0001. We synthesize 5600, 100, and 400 samples for each class on AWA2, SUN and CUB. Our experiments are based on the PyTorch and implemented on a NVIDIA GeForce RTX 3090 GPU.", + "bbox": [ + 498, + 488, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 500, + 665, + 867, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report the performance of our proposed VADS using CLSWGAN as a generative model compared to state-of-the-art methods. Tab. 1 shows the results, including the embedding-based and generative ZSL methods. In the CZSL scenario, our method notably outperforms the suboptimal results by $8.4\\%$ , $10.3\\%$ , and $8.4\\%$ , achieving the best results on AWA2, SUN, and CUB. The results confirm that our method incorporating dynamic semantic prototypes is more generalizable to unseen classes than static semantic prototypes. In the GZSL scenario, our method obtains the best harmonic mean $\\pmb{H}$ on all datasets (i.e., AWA2 $(H = 79.3)$ , SUN $(H = 55.7)$ and CUB $(H = 74.3)$ ). Our method significantly outperforms CLSWGAN+DSP [12], which proposed an evolved semantic prototype, indicat", + "bbox": [ + 498, + 689, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23631", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7b848c2839264b4b5adc4fa1d3d19e291302a3b4a12d7f06e3251cf7700cc547.jpg", + "table_caption": [ + "Table 1. Compared our VADS with the state-of-the-art on AWA2, SUN and CUB benchmark datasets in the CZSL and GZSL settings. The best and second-best results are marked in Red and Blue, respectively. Symbol “-” denotes no results are reported." + ], + "table_footnote": [], + "table_body": "
TypeMethodsVenueBackboneAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
EmbeddingPREN [52]CVPR'19ResNet-10174.132.488.647.462.935.427.230.866.435.255.843.1
DAZLE [22]CVPR'20ResNet-10167.960.375.767.159.452.324.333.266.056.759.658.1
DVBE [28]CVPR'20ResNet-101-63.670.867.0-45.037.240.7-53.260.256.5
CN [38]ICLR'21ResNet-101-60.277.167.6-44.741.643.1-49.950.750.3
GEM-ZSL [27]CVPR'21ResNet-10167.364.877.570.662.838.135.736.977.864.877.170.4
ViT-ZSL [3]IMVIP'21ViT-Large-51.990.065.8-44.555.349.3-67.375.271.0
IEAM-ZSL [2]DGAM'21ViT-Large-53.789.967.2-48.254.751.3-68.673.871.1
DUET [14]AAAI'23ViT-Base69.963.784.772.764.445.745.845.872.362.972.867.5
PSVMA [26]CVPR'23ViT-Base-73.677.375.4-61.745.352.3-70.177.873.8
Generativef-VAEGAN-D2 [47]CVPR'19ResNet-10171.157.670.663.564.745.138.041.361.048.460.153.6
TF-VAEGAN [31]ECCV'20ResNet-10172.259.875.166.666.045.640.743.064.952.864.758.1
FREE [7]ICCV'21ResNet-101-60.475.467.1-47.437.241.7-55.759.957.7
HSVA [8]NeurIPS'21ResNet-101-56.779.866.3-48.639.043.3-52.758.355.3
CE-GZSL [20]CVPR'21ResNet-10170.463.178.670.063.348.838.643.177.563.966.865.3
FREE+ESZSL [5]ICLR'22ResNet-101-51.378.061.8-48.236.541.5-51.660.455.7
ICCE [24]CVPR'22ResNet-10172.765.382.372.8----78.467.365.566.4
TDCSS [16]CVPR'22ResNet-101-59.274.966.1-----44.262.851.9
CDL + OSCO [4]TPAMI'23ResNet-101-48.071.057.1-32.065.042.9-29.069.040.6
CLSWGAN+DSP [12]ICML'23ResNet-101-60.086.070.7-48.343.045.5-51.463.856.9
TFVAEGAN+SHIP [43]ICCV'23ViT-Base-61.295.974.7-----22.582.235.3
VADS (Ours)-ViT-Base82.575.483.679.376.364.649.055.786.874.174.674.3
", + "bbox": [ + 80, + 127, + 893, + 415 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a5d15ab5fdb53e9e4870636d99d5b6469d722165a4a43e1f202f1aaa4ae0446.jpg", + "table_caption": [ + "Table 2. A detailed illustration of the ZSL benchmark datasets. $s$ and $u$ are the number of seen and unseen classes. ${N}_{A}$ refers to semantic prototype dimensions. We follow CE-GZSL [20] using 1024-dimensional semantic prototypes generated from textual descriptions [37] on CUB." + ], + "table_footnote": [], + "table_body": "
Datasets# images# classes (s | u)# NA
AWA2 [47]3732250 (40 | 10)85
SUN [34]14340717 (645 | 72)102
CUB [44]11788200 (150 | 50)1024
", + "bbox": [ + 91, + 517, + 446, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing that our design is reasonable. Meanwhile, our method is also competitive compared to the recent PSVMA [26], DUET [14], and TFVAEGAN+SHIP [43] methods using the ViT Backbone. Furthermore, our method achieves optimal results in unseen class accuracy $U$ , demonstrating that the features synthesized by the generator are closer to the real features of unseen classes, effectively alleviating the over-fitting problem. Noted that TFVAEGAN+SHIP [43] using CLIP Encoder and ViT-ZSL [2] using ViT-Large achieve the best and second-best accuracy for seen classes, but they fail to generalize well to unseen classes. These results consistently demonstrate that our method synthesizes reliable features of unseen classes to facilitate classifier learning, resulting in superior ZSL performance.", + "bbox": [ + 75, + 594, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 816, + 230, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Component Analysis. In this section, we perform a series of experiments to analyze the effectiveness of significant components. Tab. 3 summarizes the results of ablation studies on CUB and AWA2. We first use ViT-Base", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/87ef7d699d47d60f53e1d67f86d2d4c9e861f12cfa5a73b01816b0284a4e07b0.jpg", + "table_caption": [ + "Table 3. Ablation study of VADS on modules, feature enhancement and loss terms on CUB and AWA2. We use CLSWGAN [45] as a generative model. The best result is marked in boldface." + ], + "table_footnote": [], + "table_body": "
ConfigurationsCUBAWA2
AccHAccH
(1) VADS w/o VDKL & VOSU80.165.271.369.3
(2) VADS w/o VDKL84.972.878.178.5
(3) VADS w/o VOSU83.870.475.477.0
(4) VADS w/o enhancement85.173.481.879.1
(5) VADS w/o Lcon (i.e., Eq. (1))85.373.179.478.6
(6) VADS w/o Lsc (i.e., Eq. (6))86.073.579.578.7
(7) VADS (full)86.874.382.579.3
", + "bbox": [ + 514, + 489, + 879, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to extract visual features to train CLSWAGN as the baseline. Compared to the baseline (i.e., configuration (1)), our VADS (configuration (7)) performance improves significantly (i.e., the $Acc / H$ increases by $6.7\\% / 9.1\\%$ on CUB and $11.5\\% / 10\\%$ on AWA2). Configurations (2) and (3) correspond to the effectiveness of the VDKL and VOSU modules. When there is no VDKL, that is, no visual prior is introduced to generate samples, the performance drops by $1.9\\% / 1.5\\%$ and $4.4\\% / 0.8\\%$ for CUB and AWA2, indicating that visual prior is beneficial for transferring knowledge to unseen classes. When without VOSU, the performance drops most severely, verifying visual-semantic alignment is crucial for learning an accurate generator $G$ . Next, without $VE$ enhancing classification features, the results of configuration (4) drop slightly in both CZSL and GZSL settings, which shows that feature enhancement mitigates cross-dataset bias. Lastly, we conduct experiments to study the impact of the loss terms on performance. Configuration (5) without contrastive loss $\\mathcal{L}_{con}$ , $Acc / H$ decreases", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23632", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f35e236a9c3ad42d0aa4c3a6a624d5388b489e1f75690ee605aea09ef31219e3.jpg", + "table_caption": [ + "Table 4. Evaluation of VADS with multiple generative ZSL models on three prominent datasets using ViT-Base Backbone. Each row pair shows the effect of adding VADS to a particular generative ZSL model. We use the same hyperparameter optimization policy in all cases to make results comparable." + ], + "table_footnote": [], + "table_body": "
Generative ZSL MethodsAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
CLSWGAN [46]71.366.272.669.366.046.942.644.680.160.071.365.2
CLSWGAN + VADS82.5+11.575.483.679.3+10.076.3+10.364.649.055.7+11.186.8+6.774.174.674.3+9.1
TFVAEGAN [31]78.266.787.175.673.160.648.654.081.664.874.669.3
TFVAEGAN + VADS80.2+2.075.783.379.3+3.776.3+3.261.951.055.9+1.983.6+2.070.170.970.5+1.2
FREE [7]70.662.985.972.671.745.450.447.884.368.773.570.9
FREE + VADS79.4+8.870.184.676.6+4.075.0+3.357.650.753.9+6.185.5+1.270.975.473.1+3.2
", + "bbox": [ + 76, + 141, + 893, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5149f26aff000651bc3f3ece0de284bb2acc273238e457bf7a2aecf9baf841f4.jpg", + "image_caption": [ + "Figure 3. t-SNE visualizations on CUB. The 10 different colors refer to the 5 seen classes and 5 unseen classes that are randomly selected. Please zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 83, + 280, + 464, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by $1.5\\% / 1.2\\%$ and $3.1\\% / 0.7\\%$ on CUB and AWA2, respectively. The main reason is that contrastive loss achieves feature alignment, without which the network may learn category-agnostic redundant information. For the lack of semantic consistency loss $\\mathcal{L}_{sc}$ (i.e., configuration (6)), there is no guarantee that the semantic prototype maintains the original definition when training on seen classes, thus hurting the performance of seen classes. Note that the cross-entropy loss $\\mathcal{L}_{ce}$ and KL loss $\\mathcal{L}_{kl}$ are necessary for single network training, so we did not perform ablation.", + "bbox": [ + 75, + 477, + 470, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. To further demonstrate that our method synthesizes reliable unseen class features for classifier training, we project various visual features into two principal components via t-SNE [40]. Fig. 3 shows the t-SNE visualization on CUB. Each color of $\\star$ represents a class, and we randomly select 5 seen classes and 5 unseen classes. From left to right, they represent the real visual features extracted by ViT-Base [15], the features synthesized by CLSWGAN, and the features synthesized by our VADS. We observe that the real visual features are interclass dispersion and intra-class aggregation. In Fig. 3(b), Gaussian noise and the predefined semantic prototype serve as conditions to synthesize samples by CLSWGAN. There are two apparent phenomena: first, the feature distributions are scattered, which cannot truly reflect each class; second, the synthesized seen and unseen class features are confusing (e.g., “ $\\star$ ” and “ $\\star$ ”), which denote seen class", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"Yellow breasted Chat\" and unseen class \"Yellowthroat\", respectively). Therefore, the decision boundary of the CZSL/GZSL classifier trained with these features is unclear, consistent with the motivation of Fig. 2(d). In contrast, the features generated with our VADS are closer to the real features and are inter-class separated, as shown in Fig. 3(c). On the one hand, we analyze that visual-augmented dynamic semantic prototype motivates the generator to learn accurate semantic→visual mapping. On the other hand, synthesized unseen class features are more reliable, leading to learning appropriate classification boundaries of unseen classes.", + "bbox": [ + 496, + 280, + 893, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generative ZSL Models with VADS", + "text_level": 1, + "bbox": [ + 498, + 457, + 808, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate VADS as a generic technology to improve generative ZSL, we integrate it into three prevalent generative ZSL frameworks: CLSWGAN [45], TF-VAEGAN [31], and FREE [7]. We use the official repository to reproduce the results and then insert our module to verify our method's effectiveness. Note that TF-VAEGAN and FREE contain a visual encoder, so we maintain their design. When inserting our modules, we keep the hyperparameters unchanged to make the results comparable. The results on the three datasets are presented in Tab. 4. In terms of $Acc$ and $H$ , we observe varying degrees of performance improvements (e.g., a maximum of 11.5 points and a minimum of 1.2 points). Average growth is $7.4\\% / 5.9\\%$ , $5.6\\% / 6.4\\%$ and $3.3\\% / 4.2\\%$ for $Acc / H$ on AWA2, SUN and CUB. Overall, the consistent improvement over competitive benchmarks validates the effectiveness of our proposed method.", + "bbox": [ + 496, + 481, + 893, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Generalization Analysis of Visual Prior", + "text_level": 1, + "bbox": [ + 498, + 748, + 836, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our method, we learn a local bias and a fixed global prior vector representing domain visual prior knowledge to generalize to unseen classes. Therefore, we investigate the impact of different forms of prior knowledge on performance. The results are detailed in Tab. 5. \"Random\" refers to the prior knowledge sampled from Gaussian distribution, \"VGSE\" [50] indicates that we take the semantic knowledge extracted from the visual representation as the prior, and \"Other domain\" means using the global vector learned", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23633", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5897e4193b5a634d87eef120515a2cf3d2b779360dd96f575a6baf1c13cc4720.jpg", + "image_caption": [ + "Figure 4. Effect of (a) synthesized samples $N_{syn}$ , (b) loss weights $\\lambda_{con}$ , (c) loss weights $\\lambda_{kl}$ , and (d) loss weights $\\lambda_{sc}$ on CUB." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 302, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7284c7164519396e30295f55718079cdd1979921a6a78a3d82b259fcee33eb1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 90, + 475, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3060fdfd8a091e672a7aed5831a98f35b66c2a12a37364dcb8dc038110741e99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 90, + 650, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea761a70f0db4f0b9cdab00b58cd52edc14c8bb05fe9112ad403d575f63814df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 90, + 826, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/07a5438ad193208f09c5c2eff46d599018c425d10d42396f26a51a212d421c12.jpg", + "table_caption": [ + "Table 5. We evaluate different forms of prior knowledge on CUB. The best result is marked in boldface." + ], + "table_footnote": [], + "table_body": "
ConditionLearnableCUB
AccUSH
w/o prior84.969.175.572.2
Random85.672.774.773.4
VGSE [50]✓85.569.272.370.7
Other domain✓84.066.772.069.3
VADS (Ours)✓86.874.174.674.3
", + "bbox": [ + 88, + 268, + 455, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from SUN to transfer to CUB. The results indicate that our method utilizes the global prior and the local bias, which yields the best performance. Knowledge transfer from SUN to CUB suffers from a negative impact, underscoring the importance of dataset-specific global information. VGSE captures semantic side information from seen classes, limiting the performance of unseen classes.", + "bbox": [ + 75, + 395, + 468, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Hyper-parameters Analysis", + "text_level": 1, + "bbox": [ + 76, + 511, + 323, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study the impact of different hyper-parameters of our VADS on the CUB dataset. Fig. 4(a) shows the results of synthesizing different numbers per unseen class. The unseen class accuracy varies with the number of synthesized samples, and when $N_{syn} = 400$ , the performance is optimal. This result demonstrates that the features synthesized by our method alleviate the lack of unseen class data. Next, we evaluate the influence of individual loss weights (i.e., $\\lambda_{con}$ of the $\\mathcal{L}_{con}$ , $\\lambda_{kl}$ of the $\\mathcal{L}_{kl}$ and $\\lambda_{sc}$ of the $\\mathcal{L}_{sc}$ ). The results are presented in Fig. 4(b)(c)(d). The performance of $S$ and $U$ changes slightly as $\\mathcal{L}_{con}$ increases. When $\\mathcal{L}_{con}$ is set to 1.5, $H$ obtains the maximum value. For $\\mathcal{L}_{kl}$ loss, we find that larger $\\lambda_{kl}$ achieves better performance because the prior distribution assumption of the data is crucial. Finally, $\\mathcal{L}_{sc}$ loss forces the semantic update process to maintain semantic consistency. Its weight $\\lambda_{sc}$ is insensitive to performance.", + "bbox": [ + 75, + 534, + 470, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Predefined Semantic Prototype vs Updated Semantic Prototype", + "text_level": 1, + "bbox": [ + 76, + 801, + 468, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To give a clearer insight into the predefined semantic prototypes and the semantic prototypes updated by our method, we randomly select 10 classes on CUB and calculate their cosine similarity. Then, we visualize them in Fig. 5. We", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9420cbfcfe02778e9b392ec3d7cb37637dc42acbb0a3cc5c1538ba485a4a58c8.jpg", + "image_caption": [ + "(a) Predefined semantic prototype" + ], + "image_footnote": [], + "bbox": [ + 506, + 232, + 714, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/716f6489b306001fa2125ecbd4acc0c38180d7dba3dd38a8a367601bad64889e.jpg", + "image_caption": [ + "(b) Updated semantic prototype", + "Figure 5. Visualization of the heatmap of semantic prototype similarity. We randomly select 10 classes on CUB." + ], + "image_footnote": [], + "bbox": [ + 733, + 232, + 885, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "observe that the similarity between the predefined semantic prototypes is very high. Our method dynamically refines the visual-semantic relationships of each instance based on visual information, making the updated semantic prototype easier to distinguish between categories and achieving more accurate semantic-visual alignment.", + "bbox": [ + 496, + 395, + 890, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 506, + 617, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel Visual-Augmented Dynamic Semantic prototype method (VADS) to boost the generator to synthesize reliable features of unseen classes. Considering that rich visual knowledge can effectively generalize to unseen classes, our proposed VADS fully leverages visual information. Specifically, we design a Visual-aware Domain Knowledge Learning module (VDKL) to acquire visual prior and a Vision-Oriented Semantic Updation module (VOSU) to dynamically update the predefined semantic prototype. Ultimately, we concatenate their output to form a dynamic semantic prototype, serving as the condition of the generator to learn accurate semantic-visual mapping and synthesize features of unseen classes. Extensive experiments demonstrate remarkable results in both CZSL and GZSL scenarios. In summary, our study provides a timely insight into reliable feature synthesis, improving the generalization to unseen classes. Additionally, tasks related to knowledge transfer can draw inspiration from this concept.", + "bbox": [ + 496, + 534, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 842, + 666, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by National Key R&D Program of China (2022YFC3301000).", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23634", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[2] Faisal Alamri and Anjan Dutta. Implicit and explicit attention for zero-shot learning. In DAGM German Conference on Pattern Recognition, pages 467-483. Springer, 2021. 6", + "[3] Faisal Alamri and Anjan Dutta. Multi-head self-attention via vision transformer for zero-shot learning. arXiv preprint arXiv:2108.00045, 2021. 6", + "[4] Jacopo Cavazza, Vittorio Murino, and Alessio Del Bue. No adversaries to zero-shot learning: Distilling an ensemble of gaussian feature generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 6", + "[5] Samet Cetin. Closed-form sample probing for training generative models in zero-shot learning. Master's thesis, Middle East Technical University, 2022. 6", + "[6] Dubing Chen, Yuming Shen, Haofeng Zhang, and Philip HS Torr. Zero-shot logit adjustment. arXiv preprint arXiv:2204.11822, 2022. 1", + "[7] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 122-131, 2021. 1, 2, 3, 5, 6, 7", + "[8] Shiming Chen, GuoSen Xie, Yang Liu, Qinmu Peng, Baigui Sun, Hao Li, Xinge You, and Ling Shao. Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems, 34:16622–16634, 2021. 1, 6", + "[9] Shiming Chen, Ziming Hong, Wenjin Hou, Guo-Sen Xie, Yibing Song, Jian Zhao, Xinge You, Shuicheng Yan, and Ling Shao. Transzero++: Cross attribute-guided transformer for zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3", + "[10] Shiming Chen, Ziming Hong, Yang Liu, Guo-Sen Xie, Baigui Sun, Hao Li, Qinmu Peng, Ke Lu, and Xinge You. Transzero: Attribute-guided transformer for zero-shot learning. In AAAI, page 3, 2022. 3", + "[11] Shiming Chen, Ziming Hong, Guo-Sen Xie, Wenhan Yang, Qinmu Peng, Kai Wang, Jian Zhao, and Xinge You. Msdn: Mutually semantic distillation network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7612-7621, 2022. 3", + "[12] Shiming Chen, Wenjin Hou, Ziming Hong, Xiaohan Ding, Yibing Song, Xinge You, Tongliang Liu, and Kun Zhang. Evolving semantic prototype improves generative zero-shot learning. arXiv preprint arXiv:2306.06931, 2023. 1, 3, 5, 6", + "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhuo Chen, Yufeng Huang, Jiaoyan Chen, Yuxia Geng, Wen Zhang, Yin Fang, Jeff Z Pan, and Huajun Chen. Duet: Cross-modal semantic grounding for contrastive zero-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 405-413, 2023. 6", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5, 7", + "[16] Yaogong Feng, Xiaowen Huang, Pengbo Yang, Jian Yu, and Jitao Sang. Non-generative generalized zero-shot learning via task-correlated disentanglement and controllable samples synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2022. 6", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[18] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, pages 11494–11503, 2023. 2", + "[19] Akshita Gupta, Sanath Narayan, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Joost van de Weijer. Generative multi-label zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[20] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2371-2381, 2021. 1, 3, 6", + "[21] Ziming Hong, Shiming Chen, G Xie, Wenhan Yang, Jian Zhao, Yuanjie Shao, Qinmu Peng, and Xinge You. Semantic compression embedding for generative zero-shot learning. *IJCAI*, Vienna, Austria, 7:956–963, 2022. 1", + "[22] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4483-4493, 2020. 3, 6", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[24] Xia Kong, Zuodong Gao, Xiaofan Li, Ming Hong, Jun Liu, Chengjie Wang, Yuan Xie, and Yanyun Qu. En-compactness: Self-distillation embedding & contrastive generation for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9306–9315, 2022. 1, 3, 6", + "[25] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 3", + "[26] Man Liu, Feng Li, Chunjie Zhang, Yunchao Wei, Huihui Bai, and Yao Zhao. Progressive semantic-visual mutual adaptation for generalized zero-shot learning. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23635", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15337-15346, 2023. 5, 6", + "[27] Yang Liu, Lei Zhou, Xiao Bai, Yifei Huang, Lin Gu, Jun Zhou, and Tatsuya Harada. Goal-oriented gaze estimation for zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3794-3803, 2021. 6", + "[28] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12664-12673, 2020. 6", + "[29] Muhammad Ferjad Naeem, Yongqin Xian, Luc V Gool, and Federico Tombari. I2dformer: Learning image to document attention for zero-shot image classification. Advances in Neural Information Processing Systems, 35:12283-12294, 2022. 3", + "[30] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. I2mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3", + "[31] Sanath Narayan, Akshita Gupta, Fahad Shahbaz Khan, Cees GM Snoek, and Ling Shao. Latent embedding feedback and discriminative features for zero-shot classification. In European Conference on Computer Vision, pages 479-495. Springer, 2020. 1, 2, 3, 5, 6, 7", + "[32] Sanath Narayan, Akshita Gupta, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Mubarak Shah. Discriminative region-based multi-label zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8731-8740, 2021. 3", + "[33] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. Advances in neural information processing systems, 22, 2009. 1", + "[34] Genevieve Patterson and James Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2751-2758. IEEE, 2012. 2, 5, 6", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[36] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: lightweight image captioning prompted with retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2840-2849, 2023. 2", + "[37] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 49-58, 2016. 6" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ivan Skorokhodov and Mohamed Elhoseiny. Class normalization for (continual)? generalized zero-shot learning. arXiv preprint arXiv:2006.11328, 2020. 6", + "[39] Jie Song, Chengchao Shen, Yezhou Yang, Yang Liu, and Mingli Song. Transductive unbiased embedding for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1024-1033, 2018. 3", + "[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[42] Chaoqun Wang, Shaobo Min, Xuejin Chen, Xiaoyan Sun, and Houqiang Li. Dual progressive prototype network for generalized zero-shot learning. Advances in Neural Information Processing Systems, 34:2936-2948, 2021. 3", + "[43] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3032-3042, 2023. 1, 6", + "[44] P. Welinder, S. Branson, T. Mita, C. Wah, Florian Schroff, Serge J. Belongie, and P. Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-2010-001, Caltech., 2010. 2, 5, 6", + "[45] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6, 7", + "[46] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5542-5551, 2018. 1, 2, 3, 5, 7", + "[47] Yongqin Xian, Saurabh Sharma, Bernt Schiele, and Zeynep Akata. f-vaegan-d2: A feature generating framework for any-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10275-10284, 2019. 1, 2, 3, 5, 6", + "[48] Guo-Sen Xie, Li Liu, Xiaobo Jin, Fan Zhu, Zheng Zhang, Jie Qin, Yazhou Yao, and Ling Shao. Attentive region embedding network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9384-9393, 2019. 3", + "[49] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 3", + "[50] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Vgse: Visually-grounded semantic embeddings for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9316-9325, 2022. 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23636", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3", + "[52] Meng Ye and Yuhong Guo. Progressive ensemble networks for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11728-11736, 2019. 6", + "[53] Sheng Zhang, Muzammal Naseer, Guangyi Chen, Zhiqiang Shen, Salman Khan, Kun Zhang, and Fahad Shahbaz Khan. S3a: Towards realistic zero-shot classification via self structural semantic alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7278-7286, 2024. 1", + "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 357 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23637", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_model.json b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5bc28c869c6fb88642bd4467b45bece8bd21698b --- /dev/null +++ b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.131, + 0.73, + 0.177 + ], + "angle": 0, + "content": "Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.203, + 0.856, + 0.24 + ], + "angle": 0, + "content": "Wenjin Hou\\(^{1,4}\\), Shiming Chen\\(^{2*}\\), Shuhuang Chen\\(^{1}\\), Ziming Hong\\(^{3}\\), Yan Wang\\(^{4}\\), Xuetao Feng\\(^{4}\\), Salman Khan\\(^{2,5}\\), Fahad Shahbaz Khan\\(^{2,6}\\), Xinge You\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.239, + 0.749, + 0.257 + ], + "angle": 0, + "content": "1Huazhong University of Science and Technology (HUST), China" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.257, + 0.756, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Mohamed bin Zayed University of AI \\(^{3}\\)The University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.275, + 0.795, + 0.293 + ], + "angle": 0, + "content": "4Alibaba Group 5Australian National University 6Linkoping University" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.295, + 0.737, + 0.31 + ], + "angle": 0, + "content": "{houwj17, gchenshiming}@gmail.com youxg@mail.hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.377, + 0.473, + 0.771 + ], + "angle": 0, + "content": "Generative Zero-shot learning (ZSL) learns a generator to synthesize visual samples for unseen classes, which is an effective way to advance ZSL. However, existing generative methods rely on the conditions of Gaussian noise and the predefined semantic prototype, which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). To address this issue, we propose a novel Visual-Augmented Dynamic Semantic prototype method (termed VADS) to boost the generator to learn accurate semantic-visual mapping by fully exploiting the visual-augmented knowledge into semantic conditions. In detail, VADS consists of two modules: (1) Visual-aware Domain Knowledge Learning module (VDKL) learns the local bias and global prior of the visual features (referred to as domain visual knowledge), which replace pure Gaussian noise to provide richer prior noise information; (2) Vision-Oriented Semantic Updation module (VOSU) updates the semantic prototype according to the visual representations of the samples. Ultimately, we concatenate their output as a dynamic semantic prototype, which serves as the condition of the generator. Extensive experiments demonstrate that our VADS achieves superior CZSL and GZSL performances on three prominent datasets and outperforms other state-of-the-art methods with averaging increases by \\(6.4\\%\\), \\(5.9\\%\\) and \\(4.2\\%\\) on SUN, CUB and AWA2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.21, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.829, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Zero-shot learning [33], which transfers knowledge from seen classes to unseen classes, has garnered much attention recently. By establishing interactions between visual fea" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.346, + 0.895, + 0.528 + ], + "angle": 0, + "content": "tures and semantic prototypes (also referred to as attribute vectors, side information, or semantic embeddings [47]), generative ZSL methods exhibit impressive performance, demonstrating the potential of feature synthesis. One of the most successful frameworks is the conditional generative adversarial network (GAN) [17]. The main idea of generative ZSL methods is to align semantic prototypes and visual features to synthesize feature of unseen classes. Recent emerging studies have either designed more effective frameworks [7, 19, 20, 31, 46] or addressed more specific issues related to visual-semantic alignment [4, 8, 12, 43, 53]. These methods have achieved significant improvements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "However, these methods rely on the conditions of Gaussian noise and the predefined semantic prototype (referred to as the static semantic prototype), which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). Fig. 1 illustrates these issues: (1) The noise is sampled from a Gaussian distribution \\(\\mathcal{N}(0,1)\\), which lacks the dataset-specific visual prior knowledge (e.g., global visual information \"flying\" and \"still\" and background information \"sky\" and \"grass\"). As a result, the domain knowledge shared between seen and unseen classes cannot be utilized for feature synthesis of unseen classes, limiting the knowledge transfer. (2) The predefined semantic prototype fails to characterize each instance well. For example, the attributes \"wing black\", \"breast white\" and \"bill orange\" of the Laysan Albatross are not fixed on different images. Due to these limitations, the visual features synthesized by existing works [4, 6, 21, 24, 31, 43, 46] struggle to represent the distribution of real features, leading to poor generalization to unseen classes, as shown in Fig. 1(b). More intuitively, as shown in Fig. 1(d), the features of unseen classes synthesized by these methods are confusing, resulting in the decision boundary overfitting to the seen classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.093, + 0.291, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.094, + 0.466, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.094, + 0.639, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.094, + 0.868, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.229, + 0.29, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.364, + 0.235, + 0.375 + ], + "angle": 0, + "content": "(a) Input image" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.229, + 0.466, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.363, + 0.423, + 0.375 + ], + "angle": 0, + "content": "(b) Existing works" + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.229, + 0.639, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.363, + 0.588, + 0.374 + ], + "angle": 0, + "content": "(c) Our VADS" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.238, + 0.868, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.364, + 0.812, + 0.375 + ], + "angle": 0, + "content": "(d) Feature visualization" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.391, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Figure 1. An illustration of the core idea of our method. (a) The semantic prototype (i.e., attribute) of different images of the same category is not fixed, so the predefined semantic prototype is inaccurate in characterizing each instance. (b) Most existing works utilize Gaussian noise and the predefined semantic prototype as conditions to train a semantic→visual generator on seen classes, which fails to generalize to unseen classes. (c)(d) Our method incorporates rich visual prior with an updated semantic prototype to construct a visual-augmented dynamic semantic prototype of each instance, empowering the generator to synthesize features that faithfully represent the real distribution of unseen classes. Thus, our method achieves better generalization on seen and unseen classes than existing works (e.g., CLSWGAN [46])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Drawing inspiration from image captioning [36], which highlights the generalization of instance-conditional learning, we aim to fully leverage the visual-augmented knowledge into semantic conditions to tackle the aforementioned challenges. On the one hand, we can exploit rich domain visual prior knowledge, serving as a prior noise, to enhance the adaptation and generalization of models [18, 51, 54]. On the other hand, we can update the predefined semantic prototype to align visual representations based on visual features. As such, incorporating richer and more accurate visual information acts as the semantic condition to train an instance-conditional generative model, which is optimized to characterize each instance (more robust to class shift) rather than to serve only for specific classes. Accordingly, the generative model can synthesize features of unseen classes closer to the real ones, facilitating the classifier in learning an appropriate decision boundary (see the right of Fig. 1(d))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose an approach called Visual-Augmented Dynamic Semantic prototype (VADS) to improve generative ZSL methods. Specifically, VADS consists of two learnable modules: a Visual-aware Domain Knowledge Learning module (VDKL) and a Vision-Oriented Semantic Updation module (VOSU). The VDKL explores domain visual prior knowledge derived from visual information, which provides richer information for representing instances. The VOSU predicts instance-level semantics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.612 + ], + "angle": 0, + "content": "through visual \\(\\rightarrow\\) semantic mapping, guiding the updation of the predefined semantic prototype and promoting accurate semantic prototype learning. Finally, the extracted visual prior and the updated semantic prototype are concatenated as a visual-augmented dynamic semantic prototype, which serves as the condition of the generator during training and feature synthesis, as illustrated in Fig. 1(c). Extensive experiments demonstrate the effectiveness of our VADS." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.617, + 0.844, + 0.631 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.645, + 0.891, + 0.703 + ], + "angle": 0, + "content": "- We introduce a Visual-Augmented Dynamic Semantic prototype (VADS) to enhance the generalization of generative ZSL methods, facilitating substantial knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.705, + 0.891, + 0.78 + ], + "angle": 0, + "content": "- We devise the VDKL to leverage domain visual prior knowledge from visual features and design the VOSU to dynamically update the predefined semantic prototype. Their outputs together serve as the generator's conditions, providing richer and more accurate visual information." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We conduct extensive experiments on AWA2 [47], SUN [34] and CUB [44] datasets. The comprehensive results demonstrate that visual prior knowledge significantly improves the generalization of generative ZSL methods, i.e., average improvements of the harmonic mean over existing generative methods (e.g., f-CLSWGAN [46], TFVAEGAN [31] and FREE [7]) \\(6.4\\%\\), \\(5.9\\%\\) and \\(4.2\\%\\) on SUN, CUB and AWA2, respectively." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23628" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.12, + 0.473, + 0.408 + ], + "angle": 0, + "content": "Embedding-based Zero-Shot Learning. Embedding-based ZSL methods are one of the mainstream branches that project visual information into semantic space to align with semantic prototypes. Earlier works [25, 39, 48] directly mapped global visual features to semantic space, failing to capture local discriminative representation, resulting in suboptimal ZSL performance. Also, embeddings are learned only in seen classes, leading to inevitable seen class bias. In this regard, some studies [9, 10, 22] have attempted to use calibration loss to balance the prediction results between seen and unseen classes. Recently, attention mechanisms [41] have emerged with surprising localization abilities, so semantic-guided methods [9, 11, 29, 30, 32, 42, 48, 49] learn to discover attribute-related local regions, providing more accurate inter-class discrimination. Among these methods, APN [49] proposed an attribute prototype network to learn local features, and DPPN [42] updated attribute and category prototypes. Inspired by their work, we introduce a dynamic semantic prototype for generative ZSL methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.473, + 0.671 + ], + "angle": 0, + "content": "Generative Zero-Shot Learning. Generative ZSL methods learn semantic \\(\\rightarrow\\) visual mapping to synthesize unseen class features, effectively alleviating the lack of unseen class data. Consequently, the quality of synthesized features, which preserves visual-semantic correspondence, plays a crucial role in classification. Thus, TF-VAEGAN [31] forced semantic alignment at all stages, and FREE [7] fine-tuned visual features to address cross-dataset biases. CE-GZSL [20] and ICCE [24] projected visual features into the latent space for classification. However, these methods constructed projection spaces on seen classes, resulting in inferior generalization ability on unseen classes. Moreover, they uniformly utilize the predefined semantic prototype as a condition, making it difficult to achieve accurate visual-semantic alignment. The method most related to ours is DSP [12], which updates the prototype by simply adding the evolved and predefined semantic prototype." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Large-Scale Vision-Language Models Generalization. Vision-language models like CLIP [35], pre-trained on large-scale image-text pairs, have demonstrated significant potential for downstream tasks. When performing zero-shot recognition, the class prompts are input into the text encoder to obtain the classification weights, and the cosine similarity between the test image and the weights determines the resulting classification score. It is different from the classical ZSL methods [9, 25, 39, 42, 48, 49]. Recent research has focused on improving the generalization to unseen classes, with several previous works proposing prompt learning [1, 51, 54]. Motivated by optimizing visual conditional prompts, we introduce visual-aware domain knowledge learning into generative ZSL methods, facilitating knowledge transfer to unseen classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.126 + ], + "angle": 0, + "content": "3. Visual-Augmented Dynamic Semantic Prototype Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.133, + 0.894, + 0.193 + ], + "angle": 0, + "content": "Fig. 2 shows the framework of our VADS. Next, we first present the problem formulation and briefly review the generative ZSL model. Then, we introduce the detailed design of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.895, + 0.436 + ], + "angle": 0, + "content": "Problem Formulation. Conventional zero-shot learning (CZSL) recognizes unseen classes in the inference stage. Generalized zero-shot learning (GZSL) recognizes both seen and unseen classes. Both settings generalize from seen data \\(\\mathcal{D}^s\\) to unseen domains \\(\\mathcal{D}^u\\). \\(\\mathcal{D}^s = \\{(x_i^s, y_i^s) | x_i^s \\in \\mathcal{X}^s, y_i^s \\in \\mathcal{Y}^s\\}_{i=1}^{N_s}\\), where \\(N_s\\) is the sample number of seen classes, \\(x_i^s\\) is a feature vector in \\(\\mathcal{X}^s\\) and \\(y_i^s\\) is the corresponding lable from \\(\\mathcal{Y}^s\\). The \\(\\mathcal{D}^s\\) is split into a training set \\(\\mathcal{D}_{tr}^s\\) and a testing set \\(\\mathcal{D}_{te}^s\\) following Xian et al. [47]. Similarly, \\(\\mathcal{D}^u = \\{(x_i^u, y_i^u) | x_i^u \\in \\mathcal{X}^u, y_i^u \\in \\mathcal{Y}^u\\}_{i=1}^{N_u}\\), where \\(x_i^u\\) is a feature vector in \\(\\mathcal{X}^u\\) and \\(y_i^u\\) is the lable from \\(\\mathcal{Y}^u\\). \\(\\mathcal{Y}^s\\) and \\(\\mathcal{Y}^u\\) are disjoint. Define attribute semantic prototypes \\(\\mathcal{A} = \\mathcal{A}^s \\cup \\mathcal{A}^u\\), corresponding to each category, as a bridge to transfer knowledge from seen classes to unseen classes. In this paper, we dynamically update \\(\\mathcal{A}\\) to learn accurate visual-semantic alignment." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.444, + 0.715, + 0.459 + ], + "angle": 0, + "content": "3.1. Generative ZSL Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.895, + 0.619 + ], + "angle": 0, + "content": "The goal of the generative ZSL methods is to learn a semantic \\(\\rightarrow\\) visual generative model \\((G)\\) on seen classes and then use it to synthesize samples of unseen classes to train a classifier. Existing methods use Gaussian noise and the predefined semantic prototype as input conditions to supervise \\(G\\) synthetic features (i.e., \\(\\mathcal{A} \\times \\mathcal{Z} \\rightarrow \\hat{X}\\)). In our method, \\(G\\) represents an off-the-shelf CLSWGAN [46], which contains a generator and a discriminator. We develop the dynamic semantic prototype as a condition, allowing \\(G\\) to characterize more accurate visual-semantic relationships." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.627, + 0.892, + 0.657 + ], + "angle": 0, + "content": "3.2. Visual-aware Domain Knowledge Learning (VDKL)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.895, + 0.848 + ], + "angle": 0, + "content": "Drawing inspiration from previous prompt learning [51, 54], we exploit the rich information in visual features to assist in synthesizing features. VDKL is a data-efficient module allowing the visual features to be used to improve generalization. As shown in Fig. 2, we design a Visual Encoder (VE) and a Domain Knowledge Learning network (DKL). First, the VE encodes visual features into a latent feature \\( l \\) and a latent code \\( z \\). The latent feature enables inter-class alignment of visual features, and latent code is subsequently confined to a prior distribution \\( Z \\). The optimization of the VE is achieved via contrastive loss [13] and evidence-lower bound given by the equation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.868, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = \\mathbb {E} [ \\log \\frac {\\exp \\left(l _ {i} ^ {T} l ^ {+} / \\tau\\right)}{\\exp \\left(l _ {i} ^ {T} l ^ {+} / \\tau\\right) + \\sum_ {k = 1} ^ {K} \\exp \\left(l _ {i} ^ {T} l _ {k} ^ {-} / \\tau\\right)} ], \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.093, + 0.859, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.896, + 0.568 + ], + "angle": 0, + "content": "Figure 2. The architecture of our proposed VADS. It consists of two learnable modules: a Visual-Oriented Semantic updation module (VOSU) and a Visual-aware Domain Knowledge Learning module (VDKL). First, we obtain the prior distribution \\( \\mathbf{Z} \\) by the Visual Encoder \\( (VE) \\). Following this, the Domain Knowledge Learning network \\( (DKL) \\) transforms \\( \\mathbf{Z} \\) into a local bias \\( \\mathbf{b} \\), which is subsequently added to global learnable prior vectors \\( (\\mathbf{p}) \\) to construct the domain visual prior noise (i.e., \\( \\mathbf{Z}' \\)). At the bottom, VOSU notably updates the semantic prototype in two stages (depicted by the blue and green arrows). Finally, the visual prior noise and the updated semantic prototype together form a dynamic semantic prototype, used for the reconstruction of features by the generator." + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.589, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k l} = K L (V E (x) \\| p (z)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.675 + ], + "angle": 0, + "content": "where \\(l^{+}\\) and \\(l_{k}^{-}\\) represent positive and negative latent features, \\(\\tau\\) is a temperature parameter set as 0.15, \\(K\\) is the class numbers, \\(KL\\) denotes the Kullback-Leibler divergence and \\(p(z)\\) is a prior distribution that is assumed to be \\(\\mathcal{N}(0,1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.782 + ], + "angle": 0, + "content": "To further utilize visual prior knowledge during the training and synthesis stages, we propose a Domain Knowledge Learning network (DKL) to obtain a local bias \\( \\pmb{b} \\) of visual features (i.e., \\( \\pmb{b} = DKL(\\pmb{Z}) \\)). Additionally, we employ a learnable prior vector \\( \\pmb{p} \\) to capture global visual information (\\( \\pmb{p} \\) is randomly initialized). Subsequently, we obtain domain-specific visual prior noise as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.793, + 0.47, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} ^ {\\prime} = \\alpha \\cdot (\\boldsymbol {b} + \\boldsymbol {p}) + (1 - \\alpha) \\cdot \\text {n o i s e} (\\mathbf {0}, \\mathbf {1}), \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where the noise \\((0,1)\\) represents Gaussian noise aimed at enhancing diversity in synthesis, \\(\\alpha\\) is the combination coefficient set as 0.9. Through this operation, we argue that \\(Z^{\\prime}\\) includes rich domain visual knowledge and feeds it into the generator to provide instance conditions, promoting the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.895, + 0.667 + ], + "angle": 0, + "content": "generator learning and utilizing it for feature synthesis of unseen classes. Note that unseen class samples are unavailable in the feature synthesis stage, so we randomly sample Gaussian noise input to \\(DKL\\), transferring the domain knowledge acquired from seen classes to unseen classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.677, + 0.878, + 0.695 + ], + "angle": 0, + "content": "3.3. Visual-Oriented Semantic updation (VOSU)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.868 + ], + "angle": 0, + "content": "We observe that the predefined semantic prototype struggles to represent each visual sample accurately, so we propose a Visual-Oriented Semantic Updation module (VOSU), optimizing the semantic prototype dynamically. Our semantic prototype updation involves a two-stage process. In the first stage, we feed visual features \\( x_{s} \\) into the Visual-Semantic Prediction network (VSP) to generate a predictive semantic \\( \\hat{a} \\) that explicitly captures specific visual patterns of the target image. Then, the predefined semantic prototype is input into a Semantic Updation Mapping network (SUM) to learn an updated semantic \\( \\dot{a} \\). This mapping can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.881, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\dot {a} = S U M (a). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23630" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "To maintain the attribute information of the prototype and integrate the visual information, we jointly optimize them by the cross-entropy loss \\(\\mathcal{L}_{ce}\\). \\(\\mathcal{L}_{ce}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(V S P \\left(x _ {i}\\right) ^ {T} \\dot {a} ^ {y}\\right)}{\\sum_ {\\bar {c} \\in C ^ {s} \\cup u} \\exp \\left(V S P \\left(x _ {i}\\right) ^ {T} \\dot {a} ^ {\\bar {c}}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.2, + 0.469, + 0.322 + ], + "angle": 0, + "content": "where \\(N\\) is the batch numbers. Accordingly, the updated semantic prototype incorporates rich visual information. In the second stage, we employ SUM to update \\(\\dot{a}\\) during the conditional generative model training and use it as a condition to learn together with the generator \\(G\\) and the discriminator \\(D\\). This implementation facilitates dynamic updation and accurate visual-semantic matching. To this end, we propose the semantic consistency loss \\(\\mathcal{L}_{sc}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.334, + 0.469, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s c} = \\mathbb {E} \\left[ \\| S U M (\\dot {a}) - \\dot {a} \\| _ {1} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.363, + 0.469, + 0.453 + ], + "angle": 0, + "content": "In summary, the first stage leverages visual features to assist semantic updation, and the second stage dynamically updates the prototype of each sample. Then, we concatenate the updated semantic prototype with the visual prior noise \\( Z^{\\prime} \\), called the dynamic semantic prototype, which serves as the condition for the generator, as depicted in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.463, + 0.36, + 0.479 + ], + "angle": 0, + "content": "3.4. Overall Objective and Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.469, + 0.516 + ], + "angle": 0, + "content": "VADS Objective Loss Function. Overall, the objective loss function of VADS is:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.53, + 0.469, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {G} + \\lambda_ {\\text {c o n}} \\mathcal {L} _ {\\text {c o n}} + \\lambda_ {k l} \\mathcal {L} _ {k l} + \\lambda_ {s c} \\mathcal {L} _ {s c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.469, + 0.649 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_G\\) is the loss of conditional generative model G, \\(\\lambda_{con}\\), \\(\\lambda_{kl}\\) and \\(\\lambda_{sc}\\) are the hyper-parameter to balance each loss term. To fully validate our method, by using this loss, we train on various mainstream generative models (e.g., CLSWAGN [46], TFVAEGAN [31], and FREE [7]). Next, we illustrate feature synthesis and classifier training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.664, + 0.469, + 0.783 + ], + "angle": 0, + "content": "Visual-Augmented Feature Synthesis for Unseen Classes. To fully utilize the visual knowledge and accurate semantic prototypes, we sample Gaussian noise input \\(DKL\\) to obtain the prior noise \\(\\mathbf{Z}^{\\prime}\\) (i.e., Eq. (3)) and use SUM to update the semantic prototypes of unseen classes (i.e., \\(\\dot{a}_u = SUM(\\hat{a}_u)\\)). They serve as conditions to synthesize visual samples closer to the real features for training the classifier. The form can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.795, + 0.469, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {u} = \\mathrm {G} \\left(\\boldsymbol {Z} ^ {\\prime}, \\dot {a} _ {u}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(\\dot{a}_u\\) is updated semantic prototype of the unseen classes and \\(\\hat{x}_u\\) is the synthesized features of unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "ZSL Classifier Training and Inference. After synthesizing features, we input the seen class training features and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "synthesized unseen class features into \\(VE\\) to extract latent features and concatenate them to enhance the original features, alleviating cross-dataset bias [7]. Then, we train a CZSL classifier using enhanced-synthetic features (i.e., \\(f_{CZSL}:\\mathcal{X}\\to \\mathcal{Y}^u\\)) and train a GZSL classifier using enhanced seen class training features and enhanced-synthetic features (i.e., \\(f_{GZSL}:\\mathcal{X}\\to \\mathcal{Y}^s\\cup \\mathcal{Y}^u\\)). Finally, we perform inference using the test sets \\(\\mathcal{D}_{te}^{s}\\) and \\(\\mathcal{D}^u\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.632, + 0.244 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.253, + 0.691, + 0.269 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Benchmark Datasets. We conduct extensive experiments on three prominent ZSL benchmark datasets: Animals with Attributes 2 (AWA2 [47]), SUN Attribute (SUN [34]) and Caltech-USCD Birds-200-2011 (CUB [44]). We follow the Proposed Split (PS) setting [47] to split each dataset into seen and unseen classes, as detailed in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Evaluation Protocols. During inference (i.e., performing CZSL and GZSL classification), we follow the evaluation protocols in [47]. In the CZSL setting, we calculate the average per-class Top-1 accuracy of unseen classes, denoted as \\( \\mathbf{A} \\). For the GZSL scenario, we measure the Top-1 accuracy of the seen and unseen classes, defined as \\( \\mathbf{S} \\) and \\( \\mathbf{U} \\), respectively. We also compute the harmonic mean, defined as \\( \\mathbf{H} = (2 \\times \\mathbf{S} \\times \\mathbf{U}) / (\\mathbf{S} + \\mathbf{U}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Implementation Details. We follow PSVMA [26] using the ViT-Base Backbone [15] without fine-tuning as the feature extractor, obtaining 768-dimensional visual features for all samples. The global prior \\(\\pmb{p}\\) has the same dimension as the semantic prototype. We set the mini-batch to 64, 128 and 128 for AWA2, SUN and CUB, respectively. We use the Adam optimizer [23] with \\(\\beta_{1} = 0.5\\), \\(\\beta_{2} = 0.999\\), and set the initial learning rate to 0.0001. We synthesize 5600, 100, and 400 samples for each class on AWA2, SUN and CUB. Our experiments are based on the PyTorch and implemented on a NVIDIA GeForce RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.868, + 0.682 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We report the performance of our proposed VADS using CLSWGAN as a generative model compared to state-of-the-art methods. Tab. 1 shows the results, including the embedding-based and generative ZSL methods. In the CZSL scenario, our method notably outperforms the suboptimal results by \\(8.4\\%\\), \\(10.3\\%\\), and \\(8.4\\%\\), achieving the best results on AWA2, SUN, and CUB. The results confirm that our method incorporating dynamic semantic prototypes is more generalizable to unseen classes than static semantic prototypes. In the GZSL scenario, our method obtains the best harmonic mean \\(\\pmb{H}\\) on all datasets (i.e., AWA2 \\((H = 79.3)\\), SUN \\((H = 55.7)\\) and CUB \\((H = 74.3)\\)). Our method significantly outperforms CLSWGAN+DSP [12], which proposed an evolved semantic prototype, indicat" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23631" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Compared our VADS with the state-of-the-art on AWA2, SUN and CUB benchmark datasets in the CZSL and GZSL settings. The best and second-best results are marked in Red and Blue, respectively. Symbol “-” denotes no results are reported." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.895, + 0.416 + ], + "angle": 0, + "content": "
TypeMethodsVenueBackboneAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
EmbeddingPREN [52]CVPR'19ResNet-10174.132.488.647.462.935.427.230.866.435.255.843.1
DAZLE [22]CVPR'20ResNet-10167.960.375.767.159.452.324.333.266.056.759.658.1
DVBE [28]CVPR'20ResNet-101-63.670.867.0-45.037.240.7-53.260.256.5
CN [38]ICLR'21ResNet-101-60.277.167.6-44.741.643.1-49.950.750.3
GEM-ZSL [27]CVPR'21ResNet-10167.364.877.570.662.838.135.736.977.864.877.170.4
ViT-ZSL [3]IMVIP'21ViT-Large-51.990.065.8-44.555.349.3-67.375.271.0
IEAM-ZSL [2]DGAM'21ViT-Large-53.789.967.2-48.254.751.3-68.673.871.1
DUET [14]AAAI'23ViT-Base69.963.784.772.764.445.745.845.872.362.972.867.5
PSVMA [26]CVPR'23ViT-Base-73.677.375.4-61.745.352.3-70.177.873.8
Generativef-VAEGAN-D2 [47]CVPR'19ResNet-10171.157.670.663.564.745.138.041.361.048.460.153.6
TF-VAEGAN [31]ECCV'20ResNet-10172.259.875.166.666.045.640.743.064.952.864.758.1
FREE [7]ICCV'21ResNet-101-60.475.467.1-47.437.241.7-55.759.957.7
HSVA [8]NeurIPS'21ResNet-101-56.779.866.3-48.639.043.3-52.758.355.3
CE-GZSL [20]CVPR'21ResNet-10170.463.178.670.063.348.838.643.177.563.966.865.3
FREE+ESZSL [5]ICLR'22ResNet-101-51.378.061.8-48.236.541.5-51.660.455.7
ICCE [24]CVPR'22ResNet-10172.765.382.372.8----78.467.365.566.4
TDCSS [16]CVPR'22ResNet-101-59.274.966.1-----44.262.851.9
CDL + OSCO [4]TPAMI'23ResNet-101-48.071.057.1-32.065.042.9-29.069.040.6
CLSWGAN+DSP [12]ICML'23ResNet-101-60.086.070.7-48.343.045.5-51.463.856.9
TFVAEGAN+SHIP [43]ICCV'23ViT-Base-61.295.974.7-----22.582.235.3
VADS (Ours)-ViT-Base82.575.483.679.376.364.649.055.786.874.174.674.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.437, + 0.472, + 0.507 + ], + "angle": 0, + "content": "Table 2. A detailed illustration of the ZSL benchmark datasets. \\( s \\) and \\( u \\) are the number of seen and unseen classes. \\( {N}_{A} \\) refers to semantic prototype dimensions. We follow CE-GZSL [20] using 1024-dimensional semantic prototypes generated from textual descriptions [37] on CUB." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.518, + 0.447, + 0.584 + ], + "angle": 0, + "content": "
Datasets# images# classes (s | u)# NA
AWA2 [47]3732250 (40 | 10)85
SUN [34]14340717 (645 | 72)102
CUB [44]11788200 (150 | 50)1024
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.807 + ], + "angle": 0, + "content": "ing that our design is reasonable. Meanwhile, our method is also competitive compared to the recent PSVMA [26], DUET [14], and TFVAEGAN+SHIP [43] methods using the ViT Backbone. Furthermore, our method achieves optimal results in unseen class accuracy \\( U \\), demonstrating that the features synthesized by the generator are closer to the real features of unseen classes, effectively alleviating the over-fitting problem. Noted that TFVAEGAN+SHIP [43] using CLIP Encoder and ViT-ZSL [2] using ViT-Large achieve the best and second-best accuracy for seen classes, but they fail to generalize well to unseen classes. These results consistently demonstrate that our method synthesizes reliable features of unseen classes to facilitate classifier learning, resulting in superior ZSL performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.231, + 0.833 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Component Analysis. In this section, we perform a series of experiments to analyze the effectiveness of significant components. Tab. 3 summarizes the results of ablation studies on CUB and AWA2. We first use ViT-Base" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Table 3. Ablation study of VADS on modules, feature enhancement and loss terms on CUB and AWA2. We use CLSWGAN [45] as a generative model. The best result is marked in boldface." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.49, + 0.88, + 0.607 + ], + "angle": 0, + "content": "
ConfigurationsCUBAWA2
AccHAccH
(1) VADS w/o VDKL & VOSU80.165.271.369.3
(2) VADS w/o VDKL84.972.878.178.5
(3) VADS w/o VOSU83.870.475.477.0
(4) VADS w/o enhancement85.173.481.879.1
(5) VADS w/o Lcon (i.e., Eq. (1))85.373.179.478.6
(6) VADS w/o Lsc (i.e., Eq. (6))86.073.579.578.7
(7) VADS (full)86.874.382.579.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "to extract visual features to train CLSWAGN as the baseline. Compared to the baseline (i.e., configuration (1)), our VADS (configuration (7)) performance improves significantly (i.e., the \\( Acc / H \\) increases by \\( 6.7\\% / 9.1\\% \\) on CUB and \\( 11.5\\% / 10\\% \\) on AWA2). Configurations (2) and (3) correspond to the effectiveness of the VDKL and VOSU modules. When there is no VDKL, that is, no visual prior is introduced to generate samples, the performance drops by \\( 1.9\\% / 1.5\\% \\) and \\( 4.4\\% / 0.8\\% \\) for CUB and AWA2, indicating that visual prior is beneficial for transferring knowledge to unseen classes. When without VOSU, the performance drops most severely, verifying visual-semantic alignment is crucial for learning an accurate generator \\( G \\). Next, without \\( VE \\) enhancing classification features, the results of configuration (4) drop slightly in both CZSL and GZSL settings, which shows that feature enhancement mitigates cross-dataset bias. Lastly, we conduct experiments to study the impact of the loss terms on performance. Configuration (5) without contrastive loss \\( \\mathcal{L}_{con} \\), \\( Acc / H \\) decreases" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23632" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 4. Evaluation of VADS with multiple generative ZSL models on three prominent datasets using ViT-Base Backbone. Each row pair shows the effect of adding VADS to a particular generative ZSL model. We use the same hyperparameter optimization policy in all cases to make results comparable." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.142, + 0.895, + 0.269 + ], + "angle": 0, + "content": "
Generative ZSL MethodsAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
CLSWGAN [46]71.366.272.669.366.046.942.644.680.160.071.365.2
CLSWGAN + VADS82.5+11.575.483.679.3+10.076.3+10.364.649.055.7+11.186.8+6.774.174.674.3+9.1
TFVAEGAN [31]78.266.787.175.673.160.648.654.081.664.874.669.3
TFVAEGAN + VADS80.2+2.075.783.379.3+3.776.3+3.261.951.055.9+1.983.6+2.070.170.970.5+1.2
FREE [7]70.662.985.972.671.745.450.447.884.368.773.570.9
FREE + VADS79.4+8.870.184.676.6+4.075.0+3.357.650.753.9+6.185.5+1.270.975.473.1+3.2
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.281, + 0.465, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.472, + 0.464 + ], + "angle": 0, + "content": "Figure 3. t-SNE visualizations on CUB. The 10 different colors refer to the 5 seen classes and 5 unseen classes that are randomly selected. Please zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.472, + 0.63 + ], + "angle": 0, + "content": "by \\(1.5\\% / 1.2\\%\\) and \\(3.1\\% / 0.7\\%\\) on CUB and AWA2, respectively. The main reason is that contrastive loss achieves feature alignment, without which the network may learn category-agnostic redundant information. For the lack of semantic consistency loss \\(\\mathcal{L}_{sc}\\) (i.e., configuration (6)), there is no guarantee that the semantic prototype maintains the original definition when training on seen classes, thus hurting the performance of seen classes. Note that the cross-entropy loss \\(\\mathcal{L}_{ce}\\) and KL loss \\(\\mathcal{L}_{kl}\\) are necessary for single network training, so we did not perform ablation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Qualitative Evaluation. To further demonstrate that our method synthesizes reliable unseen class features for classifier training, we project various visual features into two principal components via t-SNE [40]. Fig. 3 shows the t-SNE visualization on CUB. Each color of \\(\\star\\) represents a class, and we randomly select 5 seen classes and 5 unseen classes. From left to right, they represent the real visual features extracted by ViT-Base [15], the features synthesized by CLSWGAN, and the features synthesized by our VADS. We observe that the real visual features are interclass dispersion and intra-class aggregation. In Fig. 3(b), Gaussian noise and the predefined semantic prototype serve as conditions to synthesize samples by CLSWGAN. There are two apparent phenomena: first, the feature distributions are scattered, which cannot truly reflect each class; second, the synthesized seen and unseen class features are confusing (e.g., “\\(\\star\\)” and “\\(\\star\\)”), which denote seen class" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.449 + ], + "angle": 0, + "content": "\"Yellow breasted Chat\" and unseen class \"Yellowthroat\", respectively). Therefore, the decision boundary of the CZSL/GZSL classifier trained with these features is unclear, consistent with the motivation of Fig. 2(d). In contrast, the features generated with our VADS are closer to the real features and are inter-class separated, as shown in Fig. 3(c). On the one hand, we analyze that visual-augmented dynamic semantic prototype motivates the generator to learn accurate semantic→visual mapping. On the other hand, synthesized unseen class features are more reliable, leading to learning appropriate classification boundaries of unseen classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.458, + 0.81, + 0.473 + ], + "angle": 0, + "content": "4.4. Generative ZSL Models with VADS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.895, + 0.739 + ], + "angle": 0, + "content": "To further evaluate VADS as a generic technology to improve generative ZSL, we integrate it into three prevalent generative ZSL frameworks: CLSWGAN [45], TF-VAEGAN [31], and FREE [7]. We use the official repository to reproduce the results and then insert our module to verify our method's effectiveness. Note that TF-VAEGAN and FREE contain a visual encoder, so we maintain their design. When inserting our modules, we keep the hyperparameters unchanged to make the results comparable. The results on the three datasets are presented in Tab. 4. In terms of \\( Acc \\) and \\( H \\), we observe varying degrees of performance improvements (e.g., a maximum of 11.5 points and a minimum of 1.2 points). Average growth is \\( 7.4\\% / 5.9\\% \\), \\( 5.6\\% / 6.4\\% \\) and \\( 3.3\\% / 4.2\\% \\) for \\( Acc / H \\) on AWA2, SUN and CUB. Overall, the consistent improvement over competitive benchmarks validates the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.749, + 0.838, + 0.764 + ], + "angle": 0, + "content": "4.5. Generalization Analysis of Visual Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In our method, we learn a local bias and a fixed global prior vector representing domain visual prior knowledge to generalize to unseen classes. Therefore, we investigate the impact of different forms of prior knowledge on performance. The results are detailed in Tab. 5. \"Random\" refers to the prior knowledge sampled from Gaussian distribution, \"VGSE\" [50] indicates that we take the semantic knowledge extracted from the visual representation as the prior, and \"Other domain\" means using the global vector learned" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23633" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.092, + 0.303, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.092, + 0.477, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.092, + 0.651, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.092, + 0.828, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.196, + 0.86, + 0.211 + ], + "angle": 0, + "content": "Figure 4. Effect of (a) synthesized samples \\(N_{syn}\\), (b) loss weights \\(\\lambda_{con}\\), (c) loss weights \\(\\lambda_{kl}\\), and (d) loss weights \\(\\lambda_{sc}\\) on CUB." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.23, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Table 5. We evaluate different forms of prior knowledge on CUB. The best result is marked in boldface." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.269, + 0.456, + 0.371 + ], + "angle": 0, + "content": "
ConditionLearnableCUB
AccUSH
w/o prior84.969.175.572.2
Random85.672.774.773.4
VGSE [50]✓85.569.272.370.7
Other domain✓84.066.772.069.3
VADS (Ours)✓86.874.174.674.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.502 + ], + "angle": 0, + "content": "from SUN to transfer to CUB. The results indicate that our method utilizes the global prior and the local bias, which yields the best performance. Knowledge transfer from SUN to CUB suffers from a negative impact, underscoring the importance of dataset-specific global information. VGSE captures semantic side information from seen classes, limiting the performance of unseen classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.325, + 0.529 + ], + "angle": 0, + "content": "4.6. Hyper-parameters Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.793 + ], + "angle": 0, + "content": "We study the impact of different hyper-parameters of our VADS on the CUB dataset. Fig. 4(a) shows the results of synthesizing different numbers per unseen class. The unseen class accuracy varies with the number of synthesized samples, and when \\( N_{syn} = 400 \\), the performance is optimal. This result demonstrates that the features synthesized by our method alleviate the lack of unseen class data. Next, we evaluate the influence of individual loss weights (i.e., \\( \\lambda_{con} \\) of the \\( \\mathcal{L}_{con} \\), \\( \\lambda_{kl} \\) of the \\( \\mathcal{L}_{kl} \\) and \\( \\lambda_{sc} \\) of the \\( \\mathcal{L}_{sc} \\)). The results are presented in Fig. 4(b)(c)(d). The performance of \\( S \\) and \\( U \\) changes slightly as \\( \\mathcal{L}_{con} \\) increases. When \\( \\mathcal{L}_{con} \\) is set to 1.5, \\( H \\) obtains the maximum value. For \\( \\mathcal{L}_{kl} \\) loss, we find that larger \\( \\lambda_{kl} \\) achieves better performance because the prior distribution assumption of the data is crucial. Finally, \\( \\mathcal{L}_{sc} \\) loss forces the semantic update process to maintain semantic consistency. Its weight \\( \\lambda_{sc} \\) is insensitive to performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.469, + 0.833 + ], + "angle": 0, + "content": "4.7. Predefined Semantic Prototype vs Updated Semantic Prototype" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To give a clearer insight into the predefined semantic prototypes and the semantic prototypes updated by our method, we randomly select 10 classes on CUB and calculate their cosine similarity. Then, we visualize them in Fig. 5. We" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.233, + 0.715, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.331, + 0.676, + 0.341 + ], + "angle": 0, + "content": "(a) Predefined semantic prototype" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.233, + 0.886, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.331, + 0.864, + 0.341 + ], + "angle": 0, + "content": "(b) Updated semantic prototype" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Figure 5. Visualization of the heatmap of semantic prototype similarity. We randomly select 10 classes on CUB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.892, + 0.487 + ], + "angle": 0, + "content": "observe that the similarity between the predefined semantic prototypes is very high. Our method dynamically refines the visual-semantic relationships of each instance based on visual information, making the updated semantic prototype easier to distinguish between categories and achieving more accurate semantic-visual alignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.619, + 0.523 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.823 + ], + "angle": 0, + "content": "In this work, we propose a novel Visual-Augmented Dynamic Semantic prototype method (VADS) to boost the generator to synthesize reliable features of unseen classes. Considering that rich visual knowledge can effectively generalize to unseen classes, our proposed VADS fully leverages visual information. Specifically, we design a Visual-aware Domain Knowledge Learning module (VDKL) to acquire visual prior and a Vision-Oriented Semantic Updation module (VOSU) to dynamically update the predefined semantic prototype. Ultimately, we concatenate their output to form a dynamic semantic prototype, serving as the condition of the generator to learn accurate semantic-visual mapping and synthesize features of unseen classes. Extensive experiments demonstrate remarkable results in both CZSL and GZSL scenarios. In summary, our study provides a timely insight into reliable feature synthesis, improving the generalization to unseen classes. Additionally, tasks related to knowledge transfer can draw inspiration from this concept." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.843, + 0.667, + 0.86 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "This work is partially supported by National Key R&D Program of China (2022YFC3301000)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23634" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Faisal Alamri and Anjan Dutta. Implicit and explicit attention for zero-shot learning. In DAGM German Conference on Pattern Recognition, pages 467-483. Springer, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Faisal Alamri and Anjan Dutta. Multi-head self-attention via vision transformer for zero-shot learning. arXiv preprint arXiv:2108.00045, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[4] Jacopo Cavazza, Vittorio Murino, and Alessio Del Bue. No adversaries to zero-shot learning: Distilling an ensemble of gaussian feature generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[5] Samet Cetin. Closed-form sample probing for training generative models in zero-shot learning. Master's thesis, Middle East Technical University, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.471, + 0.43 + ], + "angle": 0, + "content": "[6] Dubing Chen, Yuming Shen, Haofeng Zhang, and Philip HS Torr. Zero-shot logit adjustment. arXiv preprint arXiv:2204.11822, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[7] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 122-131, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[8] Shiming Chen, GuoSen Xie, Yang Liu, Qinmu Peng, Baigui Sun, Hao Li, Xinge You, and Ling Shao. Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems, 34:16622–16634, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[9] Shiming Chen, Ziming Hong, Wenjin Hou, Guo-Sen Xie, Yibing Song, Jian Zhao, Xinge You, Shuicheng Yan, and Ling Shao. Transzero++: Cross attribute-guided transformer for zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.646, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[10] Shiming Chen, Ziming Hong, Yang Liu, Guo-Sen Xie, Baigui Sun, Hao Li, Qinmu Peng, Ke Lu, and Xinge You. Transzero: Attribute-guided transformer for zero-shot learning. In AAAI, page 3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.703, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[11] Shiming Chen, Ziming Hong, Guo-Sen Xie, Wenhan Yang, Qinmu Peng, Kai Wang, Jian Zhao, and Xinge You. Msdn: Mutually semantic distillation network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7612-7621, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[12] Shiming Chen, Wenjin Hou, Ziming Hong, Xiaohan Ding, Yibing Song, Xinge You, Tongliang Liu, and Kun Zhang. Evolving semantic prototype improves generative zero-shot learning. arXiv preprint arXiv:2306.06931, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Zhuo Chen, Yufeng Huang, Jiaoyan Chen, Yuxia Geng, Wen Zhang, Yin Fang, Jeff Z Pan, and Huajun Chen. Duet: Cross-modal semantic grounding for contrastive zero-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 405-413, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.895, + 0.245 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.895, + 0.327 + ], + "angle": 0, + "content": "[16] Yaogong Feng, Xiaowen Huang, Pengbo Yang, Jian Yu, and Jitao Sang. Non-generative generalized zero-shot learning via task-correlated disentanglement and controllable samples synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.895, + 0.385 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.894, + 0.441 + ], + "angle": 0, + "content": "[18] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, pages 11494–11503, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[19] Akshita Gupta, Sanath Narayan, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Joost van de Weijer. Generative multi-label zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.894, + 0.564 + ], + "angle": 0, + "content": "[20] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2371-2381, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.894, + 0.621 + ], + "angle": 0, + "content": "[21] Ziming Hong, Shiming Chen, G Xie, Wenhan Yang, Jian Zhao, Yuanjie Shao, Qinmu Peng, and Xinge You. Semantic compression embedding for generative zero-shot learning. *IJCAI*, Vienna, Austria, 7:956–963, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[22] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4483-4493, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[24] Xia Kong, Zuodong Gao, Xiaofan Li, Ming Hong, Jun Liu, Chengjie Wang, Yuan Xie, and Yanyun Qu. En-compactness: Self-distillation embedding & contrastive generation for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9306–9315, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[25] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[26] Man Liu, Feng Li, Chunjie Zhang, Yunchao Wei, Huihui Bai, and Yao Zhao. Progressive semantic-visual mutual adaptation for generalized zero-shot learning. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23635" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15337-15346, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[27] Yang Liu, Lei Zhou, Xiao Bai, Yifei Huang, Lin Gu, Jun Zhou, and Tatsuya Harada. Goal-oriented gaze estimation for zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3794-3803, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[28] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12664-12673, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[29] Muhammad Ferjad Naeem, Yongqin Xian, Luc V Gool, and Federico Tombari. I2dformer: Learning image to document attention for zero-shot image classification. Advances in Neural Information Processing Systems, 35:12283-12294, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[30] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. I2mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[31] Sanath Narayan, Akshita Gupta, Fahad Shahbaz Khan, Cees GM Snoek, and Ling Shao. Latent embedding feedback and discriminative features for zero-shot classification. In European Conference on Computer Vision, pages 479-495. Springer, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[32] Sanath Narayan, Akshita Gupta, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Mubarak Shah. Discriminative region-based multi-label zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8731-8740, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[33] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. Advances in neural information processing systems, 22, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[34] Genevieve Patterson and James Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2751-2758. IEEE, 2012. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[36] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: lightweight image captioning prompted with retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2840-2849, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 49-58, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[38] Ivan Skorokhodov and Mohamed Elhoseiny. Class normalization for (continual)? generalized zero-shot learning. arXiv preprint arXiv:2006.11328, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[39] Jie Song, Chengchao Shen, Yezhou Yang, Yang Liu, and Mingli Song. Transductive unbiased embedding for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1024-1033, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.254, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.368 + ], + "angle": 0, + "content": "[42] Chaoqun Wang, Shaobo Min, Xuejin Chen, Xiaoyan Sun, and Houqiang Li. Dual progressive prototype network for generalized zero-shot learning. Advances in Neural Information Processing Systems, 34:2936-2948, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.37, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[43] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3032-3042, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.484 + ], + "angle": 0, + "content": "[44] P. Welinder, S. Branson, T. Mita, C. Wah, Florian Schroff, Serge J. Belongie, and P. Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-2010-001, Caltech., 2010. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.486, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[45] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[46] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5542-5551, 2018. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[47] Yongqin Xian, Saurabh Sharma, Bernt Schiele, and Zeynep Akata. f-vaegan-d2: A feature generating framework for any-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10275-10284, 2019. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[48] Guo-Sen Xie, Li Liu, Xiaobo Jin, Fan Zhu, Zheng Zhang, Jie Qin, Yazhou Yao, and Ling Shao. Attentive region embedding network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9384-9393, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[49] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Vgse: Visually-grounded semantic embeddings for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9316-9325, 2022. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23636" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[52] Meng Ye and Yuhong Guo. Progressive ensemble networks for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11728-11736, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[53] Sheng Zhang, Muzammal Naseer, Guangyi Chen, Zhiqiang Shen, Salman Khan, Kun Zhang, and Fahad Shahbaz Khan. S3a: Towards realistic zero-shot classification via self structural semantic alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7278-7286, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23637" + } + ] +] \ No newline at end of file diff --git a/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_origin.pdf b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fbbe0f89e2e62d9b7389041636b4305936ba9a5f --- /dev/null +++ b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/7c08a69a-09ad-40db-a963-6aaf195fbfa6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1511c811472e8deb629d5dd13734b03f047f08add7a929a5a5e494283328af16 +size 2133491 diff --git a/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/full.md b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75ed8b170eb1f93cfd0557ef600b76adef44d11d --- /dev/null +++ b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/full.md @@ -0,0 +1,302 @@ +# Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning + +Wenjin Hou $^{1,4}$ , Shiming Chen $^{2*}$ , Shuhuang Chen $^{1}$ , Ziming Hong $^{3}$ , Yan Wang $^{4}$ , Xuetao Feng $^{4}$ , Salman Khan $^{2,5}$ , Fahad Shahbaz Khan $^{2,6}$ , Xinge You $^{1*}$ + +1Huazhong University of Science and Technology (HUST), China + +$^{2}$ Mohamed bin Zayed University of AI $^{3}$ The University of Sydney + +4Alibaba Group 5Australian National University 6Linkoping University + +{houwj17, gchenshiming}@gmail.com youxg@mail.hust.edu.cn + +# Abstract + +Generative Zero-shot learning (ZSL) learns a generator to synthesize visual samples for unseen classes, which is an effective way to advance ZSL. However, existing generative methods rely on the conditions of Gaussian noise and the predefined semantic prototype, which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). To address this issue, we propose a novel Visual-Augmented Dynamic Semantic prototype method (termed VADS) to boost the generator to learn accurate semantic-visual mapping by fully exploiting the visual-augmented knowledge into semantic conditions. In detail, VADS consists of two modules: (1) Visual-aware Domain Knowledge Learning module (VDKL) learns the local bias and global prior of the visual features (referred to as domain visual knowledge), which replace pure Gaussian noise to provide richer prior noise information; (2) Vision-Oriented Semantic Updation module (VOSU) updates the semantic prototype according to the visual representations of the samples. Ultimately, we concatenate their output as a dynamic semantic prototype, which serves as the condition of the generator. Extensive experiments demonstrate that our VADS achieves superior CZSL and GZSL performances on three prominent datasets and outperforms other state-of-the-art methods with averaging increases by $6.4\%$ , $5.9\%$ and $4.2\%$ on SUN, CUB and AWA2, respectively. + +# 1. Introduction + +Zero-shot learning [33], which transfers knowledge from seen classes to unseen classes, has garnered much attention recently. By establishing interactions between visual fea + +tures and semantic prototypes (also referred to as attribute vectors, side information, or semantic embeddings [47]), generative ZSL methods exhibit impressive performance, demonstrating the potential of feature synthesis. One of the most successful frameworks is the conditional generative adversarial network (GAN) [17]. The main idea of generative ZSL methods is to align semantic prototypes and visual features to synthesize feature of unseen classes. Recent emerging studies have either designed more effective frameworks [7, 19, 20, 31, 46] or addressed more specific issues related to visual-semantic alignment [4, 8, 12, 43, 53]. These methods have achieved significant improvements. + +However, these methods rely on the conditions of Gaussian noise and the predefined semantic prototype (referred to as the static semantic prototype), which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). Fig. 1 illustrates these issues: (1) The noise is sampled from a Gaussian distribution $\mathcal{N}(0,1)$ , which lacks the dataset-specific visual prior knowledge (e.g., global visual information "flying" and "still" and background information "sky" and "grass"). As a result, the domain knowledge shared between seen and unseen classes cannot be utilized for feature synthesis of unseen classes, limiting the knowledge transfer. (2) The predefined semantic prototype fails to characterize each instance well. For example, the attributes "wing black", "breast white" and "bill orange" of the Laysan Albatross are not fixed on different images. Due to these limitations, the visual features synthesized by existing works [4, 6, 21, 24, 31, 43, 46] struggle to represent the distribution of real features, leading to poor generalization to unseen classes, as shown in Fig. 1(b). More intuitively, as shown in Fig. 1(d), the features of unseen classes synthesized by these methods are confusing, resulting in the decision boundary overfitting to the seen classes. + +![](images/38bcce7d3c04da7ae89f2b372b3204866c504d38b6464ae121b7f8bfd8236ab6.jpg) + +![](images/67cb4509fa7045f419b184adb3d350086b1557ff3f285d5661caba16a6303bf6.jpg) + +![](images/a6fb2437c9d2299a0ace19e0727ffeae5f5c078a27dd5f88ce67f95a73e7bdc0.jpg) + +![](images/a52a90e9f396589cc762815c104667531bcc7786000477168e194edbb475667d.jpg) + +![](images/4af521c52b95043c9d12143da9279ff25d0005307d0bfcbd985bfd82b4118f60.jpg) +(a) Input image + +![](images/231add22ada34f480de0f3711fa79bdcd8426404e547fac4b13411f49c65c7c8.jpg) +(b) Existing works + +![](images/b53f1f0ecec2b8820955ccf422c0e2db2770a6e4ea9db8c81760100a7b26a62f.jpg) +(c) Our VADS +Figure 1. An illustration of the core idea of our method. (a) The semantic prototype (i.e., attribute) of different images of the same category is not fixed, so the predefined semantic prototype is inaccurate in characterizing each instance. (b) Most existing works utilize Gaussian noise and the predefined semantic prototype as conditions to train a semantic→visual generator on seen classes, which fails to generalize to unseen classes. (c)(d) Our method incorporates rich visual prior with an updated semantic prototype to construct a visual-augmented dynamic semantic prototype of each instance, empowering the generator to synthesize features that faithfully represent the real distribution of unseen classes. Thus, our method achieves better generalization on seen and unseen classes than existing works (e.g., CLSWGAN [46]). + +![](images/5c2fde26257060e3e09fd9027fd21b60e4f6c3677b84083f55b79436cb3a1525.jpg) +(d) Feature visualization + +Drawing inspiration from image captioning [36], which highlights the generalization of instance-conditional learning, we aim to fully leverage the visual-augmented knowledge into semantic conditions to tackle the aforementioned challenges. On the one hand, we can exploit rich domain visual prior knowledge, serving as a prior noise, to enhance the adaptation and generalization of models [18, 51, 54]. On the other hand, we can update the predefined semantic prototype to align visual representations based on visual features. As such, incorporating richer and more accurate visual information acts as the semantic condition to train an instance-conditional generative model, which is optimized to characterize each instance (more robust to class shift) rather than to serve only for specific classes. Accordingly, the generative model can synthesize features of unseen classes closer to the real ones, facilitating the classifier in learning an appropriate decision boundary (see the right of Fig. 1(d)). + +In this paper, we propose an approach called Visual-Augmented Dynamic Semantic prototype (VADS) to improve generative ZSL methods. Specifically, VADS consists of two learnable modules: a Visual-aware Domain Knowledge Learning module (VDKL) and a Vision-Oriented Semantic Updation module (VOSU). The VDKL explores domain visual prior knowledge derived from visual information, which provides richer information for representing instances. The VOSU predicts instance-level semantics + +through visual $\rightarrow$ semantic mapping, guiding the updation of the predefined semantic prototype and promoting accurate semantic prototype learning. Finally, the extracted visual prior and the updated semantic prototype are concatenated as a visual-augmented dynamic semantic prototype, which serves as the condition of the generator during training and feature synthesis, as illustrated in Fig. 1(c). Extensive experiments demonstrate the effectiveness of our VADS. + +Our contributions can be summarized as follows: + +- We introduce a Visual-Augmented Dynamic Semantic prototype (VADS) to enhance the generalization of generative ZSL methods, facilitating substantial knowledge transfer. +- We devise the VDKL to leverage domain visual prior knowledge from visual features and design the VOSU to dynamically update the predefined semantic prototype. Their outputs together serve as the generator's conditions, providing richer and more accurate visual information. +- We conduct extensive experiments on AWA2 [47], SUN [34] and CUB [44] datasets. The comprehensive results demonstrate that visual prior knowledge significantly improves the generalization of generative ZSL methods, i.e., average improvements of the harmonic mean over existing generative methods (e.g., f-CLSWGAN [46], TFVAEGAN [31] and FREE [7]) $6.4\%$ , $5.9\%$ and $4.2\%$ on SUN, CUB and AWA2, respectively. + +# 2. Related Work + +Embedding-based Zero-Shot Learning. Embedding-based ZSL methods are one of the mainstream branches that project visual information into semantic space to align with semantic prototypes. Earlier works [25, 39, 48] directly mapped global visual features to semantic space, failing to capture local discriminative representation, resulting in suboptimal ZSL performance. Also, embeddings are learned only in seen classes, leading to inevitable seen class bias. In this regard, some studies [9, 10, 22] have attempted to use calibration loss to balance the prediction results between seen and unseen classes. Recently, attention mechanisms [41] have emerged with surprising localization abilities, so semantic-guided methods [9, 11, 29, 30, 32, 42, 48, 49] learn to discover attribute-related local regions, providing more accurate inter-class discrimination. Among these methods, APN [49] proposed an attribute prototype network to learn local features, and DPPN [42] updated attribute and category prototypes. Inspired by their work, we introduce a dynamic semantic prototype for generative ZSL methods. + +Generative Zero-Shot Learning. Generative ZSL methods learn semantic $\rightarrow$ visual mapping to synthesize unseen class features, effectively alleviating the lack of unseen class data. Consequently, the quality of synthesized features, which preserves visual-semantic correspondence, plays a crucial role in classification. Thus, TF-VAEGAN [31] forced semantic alignment at all stages, and FREE [7] fine-tuned visual features to address cross-dataset biases. CE-GZSL [20] and ICCE [24] projected visual features into the latent space for classification. However, these methods constructed projection spaces on seen classes, resulting in inferior generalization ability on unseen classes. Moreover, they uniformly utilize the predefined semantic prototype as a condition, making it difficult to achieve accurate visual-semantic alignment. The method most related to ours is DSP [12], which updates the prototype by simply adding the evolved and predefined semantic prototype. + +Large-Scale Vision-Language Models Generalization. Vision-language models like CLIP [35], pre-trained on large-scale image-text pairs, have demonstrated significant potential for downstream tasks. When performing zero-shot recognition, the class prompts are input into the text encoder to obtain the classification weights, and the cosine similarity between the test image and the weights determines the resulting classification score. It is different from the classical ZSL methods [9, 25, 39, 42, 48, 49]. Recent research has focused on improving the generalization to unseen classes, with several previous works proposing prompt learning [1, 51, 54]. Motivated by optimizing visual conditional prompts, we introduce visual-aware domain knowledge learning into generative ZSL methods, facilitating knowledge transfer to unseen classes. + +# 3. Visual-Augmented Dynamic Semantic Prototype Method + +Fig. 2 shows the framework of our VADS. Next, we first present the problem formulation and briefly review the generative ZSL model. Then, we introduce the detailed design of our method. + +Problem Formulation. Conventional zero-shot learning (CZSL) recognizes unseen classes in the inference stage. Generalized zero-shot learning (GZSL) recognizes both seen and unseen classes. Both settings generalize from seen data $\mathcal{D}^s$ to unseen domains $\mathcal{D}^u$ . $\mathcal{D}^s = \{(x_i^s, y_i^s) | x_i^s \in \mathcal{X}^s, y_i^s \in \mathcal{Y}^s\}_{i=1}^{N_s}$ , where $N_s$ is the sample number of seen classes, $x_i^s$ is a feature vector in $\mathcal{X}^s$ and $y_i^s$ is the corresponding lable from $\mathcal{Y}^s$ . The $\mathcal{D}^s$ is split into a training set $\mathcal{D}_{tr}^s$ and a testing set $\mathcal{D}_{te}^s$ following Xian et al. [47]. Similarly, $\mathcal{D}^u = \{(x_i^u, y_i^u) | x_i^u \in \mathcal{X}^u, y_i^u \in \mathcal{Y}^u\}_{i=1}^{N_u}$ , where $x_i^u$ is a feature vector in $\mathcal{X}^u$ and $y_i^u$ is the lable from $\mathcal{Y}^u$ . $\mathcal{Y}^s$ and $\mathcal{Y}^u$ are disjoint. Define attribute semantic prototypes $\mathcal{A} = \mathcal{A}^s \cup \mathcal{A}^u$ , corresponding to each category, as a bridge to transfer knowledge from seen classes to unseen classes. In this paper, we dynamically update $\mathcal{A}$ to learn accurate visual-semantic alignment. + +# 3.1. Generative ZSL Model + +The goal of the generative ZSL methods is to learn a semantic $\rightarrow$ visual generative model $(G)$ on seen classes and then use it to synthesize samples of unseen classes to train a classifier. Existing methods use Gaussian noise and the predefined semantic prototype as input conditions to supervise $G$ synthetic features (i.e., $\mathcal{A} \times \mathcal{Z} \rightarrow \hat{X}$ ). In our method, $G$ represents an off-the-shelf CLSWGAN [46], which contains a generator and a discriminator. We develop the dynamic semantic prototype as a condition, allowing $G$ to characterize more accurate visual-semantic relationships. + +# 3.2. Visual-aware Domain Knowledge Learning (VDKL) + +Drawing inspiration from previous prompt learning [51, 54], we exploit the rich information in visual features to assist in synthesizing features. VDKL is a data-efficient module allowing the visual features to be used to improve generalization. As shown in Fig. 2, we design a Visual Encoder (VE) and a Domain Knowledge Learning network (DKL). First, the VE encodes visual features into a latent feature $l$ and a latent code $z$ . The latent feature enables inter-class alignment of visual features, and latent code is subsequently confined to a prior distribution $Z$ . The optimization of the VE is achieved via contrastive loss [13] and evidence-lower bound given by the equation as follows: + +$$ +\mathcal {L} _ {c o n} = \mathbb {E} [ \log \frac {\exp \left(l _ {i} ^ {T} l ^ {+} / \tau\right)}{\exp \left(l _ {i} ^ {T} l ^ {+} / \tau\right) + \sum_ {k = 1} ^ {K} \exp \left(l _ {i} ^ {T} l _ {k} ^ {-} / \tau\right)} ], \tag {1} +$$ + +![](images/d06d19ef86e59dd2205450a03a94b64dc3fbbd5037dfd1ed8b67bb677ad80d91.jpg) +Figure 2. The architecture of our proposed VADS. It consists of two learnable modules: a Visual-Oriented Semantic updation module (VOSU) and a Visual-aware Domain Knowledge Learning module (VDKL). First, we obtain the prior distribution $\mathbf{Z}$ by the Visual Encoder $(VE)$ . Following this, the Domain Knowledge Learning network $(DKL)$ transforms $\mathbf{Z}$ into a local bias $\mathbf{b}$ , which is subsequently added to global learnable prior vectors $(\mathbf{p})$ to construct the domain visual prior noise (i.e., $\mathbf{Z}'$ ). At the bottom, VOSU notably updates the semantic prototype in two stages (depicted by the blue and green arrows). Finally, the visual prior noise and the updated semantic prototype together form a dynamic semantic prototype, used for the reconstruction of features by the generator. + +$$ +\mathcal {L} _ {k l} = K L (V E (x) \| p (z)), \tag {2} +$$ + +where $l^{+}$ and $l_{k}^{-}$ represent positive and negative latent features, $\tau$ is a temperature parameter set as 0.15, $K$ is the class numbers, $KL$ denotes the Kullback-Leibler divergence and $p(z)$ is a prior distribution that is assumed to be $\mathcal{N}(0,1)$ . + +To further utilize visual prior knowledge during the training and synthesis stages, we propose a Domain Knowledge Learning network (DKL) to obtain a local bias $\pmb{b}$ of visual features (i.e., $\pmb{b} = DKL(\pmb{Z})$ ). Additionally, we employ a learnable prior vector $\pmb{p}$ to capture global visual information ( $\pmb{p}$ is randomly initialized). Subsequently, we obtain domain-specific visual prior noise as follows: + +$$ +\boldsymbol {Z} ^ {\prime} = \alpha \cdot (\boldsymbol {b} + \boldsymbol {p}) + (1 - \alpha) \cdot \text {n o i s e} (\mathbf {0}, \mathbf {1}), \quad (3) +$$ + +where the noise $(0,1)$ represents Gaussian noise aimed at enhancing diversity in synthesis, $\alpha$ is the combination coefficient set as 0.9. Through this operation, we argue that $Z^{\prime}$ includes rich domain visual knowledge and feeds it into the generator to provide instance conditions, promoting the + +generator learning and utilizing it for feature synthesis of unseen classes. Note that unseen class samples are unavailable in the feature synthesis stage, so we randomly sample Gaussian noise input to $DKL$ , transferring the domain knowledge acquired from seen classes to unseen classes. + +# 3.3. Visual-Oriented Semantic updation (VOSU) + +We observe that the predefined semantic prototype struggles to represent each visual sample accurately, so we propose a Visual-Oriented Semantic Updation module (VOSU), optimizing the semantic prototype dynamically. Our semantic prototype updation involves a two-stage process. In the first stage, we feed visual features $x_{s}$ into the Visual-Semantic Prediction network (VSP) to generate a predictive semantic $\hat{a}$ that explicitly captures specific visual patterns of the target image. Then, the predefined semantic prototype is input into a Semantic Updation Mapping network (SUM) to learn an updated semantic $\dot{a}$ . This mapping can be expressed as: + +$$ +\dot {a} = S U M (a). \tag {4} +$$ + +To maintain the attribute information of the prototype and integrate the visual information, we jointly optimize them by the cross-entropy loss $\mathcal{L}_{ce}$ . $\mathcal{L}_{ce}$ is defined as: + +$$ +\mathcal {L} _ {c e} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(V S P \left(x _ {i}\right) ^ {T} \dot {a} ^ {y}\right)}{\sum_ {\bar {c} \in C ^ {s} \cup u} \exp \left(V S P \left(x _ {i}\right) ^ {T} \dot {a} ^ {\bar {c}}\right)}, \tag {5} +$$ + +where $N$ is the batch numbers. Accordingly, the updated semantic prototype incorporates rich visual information. In the second stage, we employ SUM to update $\dot{a}$ during the conditional generative model training and use it as a condition to learn together with the generator $G$ and the discriminator $D$ . This implementation facilitates dynamic updation and accurate visual-semantic matching. To this end, we propose the semantic consistency loss $\mathcal{L}_{sc}$ as follows: + +$$ +\mathcal {L} _ {s c} = \mathbb {E} \left[ \| S U M (\dot {a}) - \dot {a} \| _ {1} \right]. \tag {6} +$$ + +In summary, the first stage leverages visual features to assist semantic updation, and the second stage dynamically updates the prototype of each sample. Then, we concatenate the updated semantic prototype with the visual prior noise $Z^{\prime}$ , called the dynamic semantic prototype, which serves as the condition for the generator, as depicted in Fig. 2. + +# 3.4. Overall Objective and Inference + +VADS Objective Loss Function. Overall, the objective loss function of VADS is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {G} + \lambda_ {\text {c o n}} \mathcal {L} _ {\text {c o n}} + \lambda_ {k l} \mathcal {L} _ {k l} + \lambda_ {s c} \mathcal {L} _ {s c}, \tag {7} +$$ + +where $\mathcal{L}_G$ is the loss of conditional generative model G, $\lambda_{con}$ , $\lambda_{kl}$ and $\lambda_{sc}$ are the hyper-parameter to balance each loss term. To fully validate our method, by using this loss, we train on various mainstream generative models (e.g., CLSWAGN [46], TFVAEGAN [31], and FREE [7]). Next, we illustrate feature synthesis and classifier training. + +Visual-Augmented Feature Synthesis for Unseen Classes. To fully utilize the visual knowledge and accurate semantic prototypes, we sample Gaussian noise input $DKL$ to obtain the prior noise $\mathbf{Z}^{\prime}$ (i.e., Eq. (3)) and use SUM to update the semantic prototypes of unseen classes (i.e., $\dot{a}_u = SUM(\hat{a}_u)$ ). They serve as conditions to synthesize visual samples closer to the real features for training the classifier. The form can be written as: + +$$ +\hat {x} _ {u} = \mathrm {G} \left(\boldsymbol {Z} ^ {\prime}, \dot {a} _ {u}\right), \tag {8} +$$ + +where $\dot{a}_u$ is updated semantic prototype of the unseen classes and $\hat{x}_u$ is the synthesized features of unseen classes. + +ZSL Classifier Training and Inference. After synthesizing features, we input the seen class training features and + +synthesized unseen class features into $VE$ to extract latent features and concatenate them to enhance the original features, alleviating cross-dataset bias [7]. Then, we train a CZSL classifier using enhanced-synthetic features (i.e., $f_{CZSL}:\mathcal{X}\to \mathcal{Y}^u$ ) and train a GZSL classifier using enhanced seen class training features and enhanced-synthetic features (i.e., $f_{GZSL}:\mathcal{X}\to \mathcal{Y}^s\cup \mathcal{Y}^u$ ). Finally, we perform inference using the test sets $\mathcal{D}_{te}^{s}$ and $\mathcal{D}^u$ . + +# 4. Experiments + +# 4.1. Experimental Setup + +Benchmark Datasets. We conduct extensive experiments on three prominent ZSL benchmark datasets: Animals with Attributes 2 (AWA2 [47]), SUN Attribute (SUN [34]) and Caltech-USCD Birds-200-2011 (CUB [44]). We follow the Proposed Split (PS) setting [47] to split each dataset into seen and unseen classes, as detailed in Tab. 2. + +Evaluation Protocols. During inference (i.e., performing CZSL and GZSL classification), we follow the evaluation protocols in [47]. In the CZSL setting, we calculate the average per-class Top-1 accuracy of unseen classes, denoted as $\mathbf{A}$ . For the GZSL scenario, we measure the Top-1 accuracy of the seen and unseen classes, defined as $\mathbf{S}$ and $\mathbf{U}$ , respectively. We also compute the harmonic mean, defined as $\mathbf{H} = (2 \times \mathbf{S} \times \mathbf{U}) / (\mathbf{S} + \mathbf{U})$ . + +Implementation Details. We follow PSVMA [26] using the ViT-Base Backbone [15] without fine-tuning as the feature extractor, obtaining 768-dimensional visual features for all samples. The global prior $\pmb{p}$ has the same dimension as the semantic prototype. We set the mini-batch to 64, 128 and 128 for AWA2, SUN and CUB, respectively. We use the Adam optimizer [23] with $\beta_{1} = 0.5$ , $\beta_{2} = 0.999$ , and set the initial learning rate to 0.0001. We synthesize 5600, 100, and 400 samples for each class on AWA2, SUN and CUB. Our experiments are based on the PyTorch and implemented on a NVIDIA GeForce RTX 3090 GPU. + +# 4.2. Comparison with State-of-the-Art Methods + +We report the performance of our proposed VADS using CLSWGAN as a generative model compared to state-of-the-art methods. Tab. 1 shows the results, including the embedding-based and generative ZSL methods. In the CZSL scenario, our method notably outperforms the suboptimal results by $8.4\%$ , $10.3\%$ , and $8.4\%$ , achieving the best results on AWA2, SUN, and CUB. The results confirm that our method incorporating dynamic semantic prototypes is more generalizable to unseen classes than static semantic prototypes. In the GZSL scenario, our method obtains the best harmonic mean $\pmb{H}$ on all datasets (i.e., AWA2 $(H = 79.3)$ , SUN $(H = 55.7)$ and CUB $(H = 74.3)$ ). Our method significantly outperforms CLSWGAN+DSP [12], which proposed an evolved semantic prototype, indicat + +Table 1. Compared our VADS with the state-of-the-art on AWA2, SUN and CUB benchmark datasets in the CZSL and GZSL settings. The best and second-best results are marked in Red and Blue, respectively. Symbol “-” denotes no results are reported. + +
TypeMethodsVenueBackboneAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
EmbeddingPREN [52]CVPR'19ResNet-10174.132.488.647.462.935.427.230.866.435.255.843.1
DAZLE [22]CVPR'20ResNet-10167.960.375.767.159.452.324.333.266.056.759.658.1
DVBE [28]CVPR'20ResNet-101-63.670.867.0-45.037.240.7-53.260.256.5
CN [38]ICLR'21ResNet-101-60.277.167.6-44.741.643.1-49.950.750.3
GEM-ZSL [27]CVPR'21ResNet-10167.364.877.570.662.838.135.736.977.864.877.170.4
ViT-ZSL [3]IMVIP'21ViT-Large-51.990.065.8-44.555.349.3-67.375.271.0
IEAM-ZSL [2]DGAM'21ViT-Large-53.789.967.2-48.254.751.3-68.673.871.1
DUET [14]AAAI'23ViT-Base69.963.784.772.764.445.745.845.872.362.972.867.5
PSVMA [26]CVPR'23ViT-Base-73.677.375.4-61.745.352.3-70.177.873.8
Generativef-VAEGAN-D2 [47]CVPR'19ResNet-10171.157.670.663.564.745.138.041.361.048.460.153.6
TF-VAEGAN [31]ECCV'20ResNet-10172.259.875.166.666.045.640.743.064.952.864.758.1
FREE [7]ICCV'21ResNet-101-60.475.467.1-47.437.241.7-55.759.957.7
HSVA [8]NeurIPS'21ResNet-101-56.779.866.3-48.639.043.3-52.758.355.3
CE-GZSL [20]CVPR'21ResNet-10170.463.178.670.063.348.838.643.177.563.966.865.3
FREE+ESZSL [5]ICLR'22ResNet-101-51.378.061.8-48.236.541.5-51.660.455.7
ICCE [24]CVPR'22ResNet-10172.765.382.372.8----78.467.365.566.4
TDCSS [16]CVPR'22ResNet-101-59.274.966.1-----44.262.851.9
CDL + OSCO [4]TPAMI'23ResNet-101-48.071.057.1-32.065.042.9-29.069.040.6
CLSWGAN+DSP [12]ICML'23ResNet-101-60.086.070.7-48.343.045.5-51.463.856.9
TFVAEGAN+SHIP [43]ICCV'23ViT-Base-61.295.974.7-----22.582.235.3
VADS (Ours)-ViT-Base82.575.483.679.376.364.649.055.786.874.174.674.3
+ +Table 2. A detailed illustration of the ZSL benchmark datasets. $s$ and $u$ are the number of seen and unseen classes. ${N}_{A}$ refers to semantic prototype dimensions. We follow CE-GZSL [20] using 1024-dimensional semantic prototypes generated from textual descriptions [37] on CUB. + +
Datasets# images# classes (s | u)# NA
AWA2 [47]3732250 (40 | 10)85
SUN [34]14340717 (645 | 72)102
CUB [44]11788200 (150 | 50)1024
+ +ing that our design is reasonable. Meanwhile, our method is also competitive compared to the recent PSVMA [26], DUET [14], and TFVAEGAN+SHIP [43] methods using the ViT Backbone. Furthermore, our method achieves optimal results in unseen class accuracy $U$ , demonstrating that the features synthesized by the generator are closer to the real features of unseen classes, effectively alleviating the over-fitting problem. Noted that TFVAEGAN+SHIP [43] using CLIP Encoder and ViT-ZSL [2] using ViT-Large achieve the best and second-best accuracy for seen classes, but they fail to generalize well to unseen classes. These results consistently demonstrate that our method synthesizes reliable features of unseen classes to facilitate classifier learning, resulting in superior ZSL performance. + +# 4.3. Ablation Study + +Component Analysis. In this section, we perform a series of experiments to analyze the effectiveness of significant components. Tab. 3 summarizes the results of ablation studies on CUB and AWA2. We first use ViT-Base + +Table 3. Ablation study of VADS on modules, feature enhancement and loss terms on CUB and AWA2. We use CLSWGAN [45] as a generative model. The best result is marked in boldface. + +
ConfigurationsCUBAWA2
AccHAccH
(1) VADS w/o VDKL & VOSU80.165.271.369.3
(2) VADS w/o VDKL84.972.878.178.5
(3) VADS w/o VOSU83.870.475.477.0
(4) VADS w/o enhancement85.173.481.879.1
(5) VADS w/o Lcon (i.e., Eq. (1))85.373.179.478.6
(6) VADS w/o Lsc (i.e., Eq. (6))86.073.579.578.7
(7) VADS (full)86.874.382.579.3
+ +to extract visual features to train CLSWAGN as the baseline. Compared to the baseline (i.e., configuration (1)), our VADS (configuration (7)) performance improves significantly (i.e., the $Acc / H$ increases by $6.7\% / 9.1\%$ on CUB and $11.5\% / 10\%$ on AWA2). Configurations (2) and (3) correspond to the effectiveness of the VDKL and VOSU modules. When there is no VDKL, that is, no visual prior is introduced to generate samples, the performance drops by $1.9\% / 1.5\%$ and $4.4\% / 0.8\%$ for CUB and AWA2, indicating that visual prior is beneficial for transferring knowledge to unseen classes. When without VOSU, the performance drops most severely, verifying visual-semantic alignment is crucial for learning an accurate generator $G$ . Next, without $VE$ enhancing classification features, the results of configuration (4) drop slightly in both CZSL and GZSL settings, which shows that feature enhancement mitigates cross-dataset bias. Lastly, we conduct experiments to study the impact of the loss terms on performance. Configuration (5) without contrastive loss $\mathcal{L}_{con}$ , $Acc / H$ decreases + +Table 4. Evaluation of VADS with multiple generative ZSL models on three prominent datasets using ViT-Base Backbone. Each row pair shows the effect of adding VADS to a particular generative ZSL model. We use the same hyperparameter optimization policy in all cases to make results comparable. + +
Generative ZSL MethodsAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
CLSWGAN [46]71.366.272.669.366.046.942.644.680.160.071.365.2
CLSWGAN + VADS82.5+11.575.483.679.3+10.076.3+10.364.649.055.7+11.186.8+6.774.174.674.3+9.1
TFVAEGAN [31]78.266.787.175.673.160.648.654.081.664.874.669.3
TFVAEGAN + VADS80.2+2.075.783.379.3+3.776.3+3.261.951.055.9+1.983.6+2.070.170.970.5+1.2
FREE [7]70.662.985.972.671.745.450.447.884.368.773.570.9
FREE + VADS79.4+8.870.184.676.6+4.075.0+3.357.650.753.9+6.185.5+1.270.975.473.1+3.2
+ +![](images/5149f26aff000651bc3f3ece0de284bb2acc273238e457bf7a2aecf9baf841f4.jpg) +Figure 3. t-SNE visualizations on CUB. The 10 different colors refer to the 5 seen classes and 5 unseen classes that are randomly selected. Please zoom in for a better view. + +by $1.5\% / 1.2\%$ and $3.1\% / 0.7\%$ on CUB and AWA2, respectively. The main reason is that contrastive loss achieves feature alignment, without which the network may learn category-agnostic redundant information. For the lack of semantic consistency loss $\mathcal{L}_{sc}$ (i.e., configuration (6)), there is no guarantee that the semantic prototype maintains the original definition when training on seen classes, thus hurting the performance of seen classes. Note that the cross-entropy loss $\mathcal{L}_{ce}$ and KL loss $\mathcal{L}_{kl}$ are necessary for single network training, so we did not perform ablation. + +Qualitative Evaluation. To further demonstrate that our method synthesizes reliable unseen class features for classifier training, we project various visual features into two principal components via t-SNE [40]. Fig. 3 shows the t-SNE visualization on CUB. Each color of $\star$ represents a class, and we randomly select 5 seen classes and 5 unseen classes. From left to right, they represent the real visual features extracted by ViT-Base [15], the features synthesized by CLSWGAN, and the features synthesized by our VADS. We observe that the real visual features are interclass dispersion and intra-class aggregation. In Fig. 3(b), Gaussian noise and the predefined semantic prototype serve as conditions to synthesize samples by CLSWGAN. There are two apparent phenomena: first, the feature distributions are scattered, which cannot truly reflect each class; second, the synthesized seen and unseen class features are confusing (e.g., “ $\star$ ” and “ $\star$ ”), which denote seen class + +"Yellow breasted Chat" and unseen class "Yellowthroat", respectively). Therefore, the decision boundary of the CZSL/GZSL classifier trained with these features is unclear, consistent with the motivation of Fig. 2(d). In contrast, the features generated with our VADS are closer to the real features and are inter-class separated, as shown in Fig. 3(c). On the one hand, we analyze that visual-augmented dynamic semantic prototype motivates the generator to learn accurate semantic→visual mapping. On the other hand, synthesized unseen class features are more reliable, leading to learning appropriate classification boundaries of unseen classes. + +# 4.4. Generative ZSL Models with VADS + +To further evaluate VADS as a generic technology to improve generative ZSL, we integrate it into three prevalent generative ZSL frameworks: CLSWGAN [45], TF-VAEGAN [31], and FREE [7]. We use the official repository to reproduce the results and then insert our module to verify our method's effectiveness. Note that TF-VAEGAN and FREE contain a visual encoder, so we maintain their design. When inserting our modules, we keep the hyperparameters unchanged to make the results comparable. The results on the three datasets are presented in Tab. 4. In terms of $Acc$ and $H$ , we observe varying degrees of performance improvements (e.g., a maximum of 11.5 points and a minimum of 1.2 points). Average growth is $7.4\% / 5.9\%$ , $5.6\% / 6.4\%$ and $3.3\% / 4.2\%$ for $Acc / H$ on AWA2, SUN and CUB. Overall, the consistent improvement over competitive benchmarks validates the effectiveness of our proposed method. + +# 4.5. Generalization Analysis of Visual Prior + +In our method, we learn a local bias and a fixed global prior vector representing domain visual prior knowledge to generalize to unseen classes. Therefore, we investigate the impact of different forms of prior knowledge on performance. The results are detailed in Tab. 5. "Random" refers to the prior knowledge sampled from Gaussian distribution, "VGSE" [50] indicates that we take the semantic knowledge extracted from the visual representation as the prior, and "Other domain" means using the global vector learned + +![](images/5897e4193b5a634d87eef120515a2cf3d2b779360dd96f575a6baf1c13cc4720.jpg) +Figure 4. Effect of (a) synthesized samples $N_{syn}$ , (b) loss weights $\lambda_{con}$ , (c) loss weights $\lambda_{kl}$ , and (d) loss weights $\lambda_{sc}$ on CUB. + +![](images/7284c7164519396e30295f55718079cdd1979921a6a78a3d82b259fcee33eb1d.jpg) + +![](images/3060fdfd8a091e672a7aed5831a98f35b66c2a12a37364dcb8dc038110741e99.jpg) + +![](images/ea761a70f0db4f0b9cdab00b58cd52edc14c8bb05fe9112ad403d575f63814df.jpg) + +Table 5. We evaluate different forms of prior knowledge on CUB. The best result is marked in boldface. + +
ConditionLearnableCUB
AccUSH
w/o prior84.969.175.572.2
Random85.672.774.773.4
VGSE [50]✓85.569.272.370.7
Other domain✓84.066.772.069.3
VADS (Ours)✓86.874.174.674.3
+ +from SUN to transfer to CUB. The results indicate that our method utilizes the global prior and the local bias, which yields the best performance. Knowledge transfer from SUN to CUB suffers from a negative impact, underscoring the importance of dataset-specific global information. VGSE captures semantic side information from seen classes, limiting the performance of unseen classes. + +# 4.6. Hyper-parameters Analysis + +We study the impact of different hyper-parameters of our VADS on the CUB dataset. Fig. 4(a) shows the results of synthesizing different numbers per unseen class. The unseen class accuracy varies with the number of synthesized samples, and when $N_{syn} = 400$ , the performance is optimal. This result demonstrates that the features synthesized by our method alleviate the lack of unseen class data. Next, we evaluate the influence of individual loss weights (i.e., $\lambda_{con}$ of the $\mathcal{L}_{con}$ , $\lambda_{kl}$ of the $\mathcal{L}_{kl}$ and $\lambda_{sc}$ of the $\mathcal{L}_{sc}$ ). The results are presented in Fig. 4(b)(c)(d). The performance of $S$ and $U$ changes slightly as $\mathcal{L}_{con}$ increases. When $\mathcal{L}_{con}$ is set to 1.5, $H$ obtains the maximum value. For $\mathcal{L}_{kl}$ loss, we find that larger $\lambda_{kl}$ achieves better performance because the prior distribution assumption of the data is crucial. Finally, $\mathcal{L}_{sc}$ loss forces the semantic update process to maintain semantic consistency. Its weight $\lambda_{sc}$ is insensitive to performance. + +# 4.7. Predefined Semantic Prototype vs Updated Semantic Prototype + +To give a clearer insight into the predefined semantic prototypes and the semantic prototypes updated by our method, we randomly select 10 classes on CUB and calculate their cosine similarity. Then, we visualize them in Fig. 5. We + +![](images/9420cbfcfe02778e9b392ec3d7cb37637dc42acbb0a3cc5c1538ba485a4a58c8.jpg) +(a) Predefined semantic prototype + +![](images/716f6489b306001fa2125ecbd4acc0c38180d7dba3dd38a8a367601bad64889e.jpg) +(b) Updated semantic prototype +Figure 5. Visualization of the heatmap of semantic prototype similarity. We randomly select 10 classes on CUB. + +observe that the similarity between the predefined semantic prototypes is very high. Our method dynamically refines the visual-semantic relationships of each instance based on visual information, making the updated semantic prototype easier to distinguish between categories and achieving more accurate semantic-visual alignment. + +# 5. Conclusion + +In this work, we propose a novel Visual-Augmented Dynamic Semantic prototype method (VADS) to boost the generator to synthesize reliable features of unseen classes. Considering that rich visual knowledge can effectively generalize to unseen classes, our proposed VADS fully leverages visual information. Specifically, we design a Visual-aware Domain Knowledge Learning module (VDKL) to acquire visual prior and a Vision-Oriented Semantic Updation module (VOSU) to dynamically update the predefined semantic prototype. Ultimately, we concatenate their output to form a dynamic semantic prototype, serving as the condition of the generator to learn accurate semantic-visual mapping and synthesize features of unseen classes. Extensive experiments demonstrate remarkable results in both CZSL and GZSL scenarios. In summary, our study provides a timely insight into reliable feature synthesis, improving the generalization to unseen classes. Additionally, tasks related to knowledge transfer can draw inspiration from this concept. + +# Acknowledgements + +This work is partially supported by National Key R&D Program of China (2022YFC3301000). + +# References + +[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. Advances in Neural Information Processing Systems, 36, 2024. 3 +[2] Faisal Alamri and Anjan Dutta. Implicit and explicit attention for zero-shot learning. In DAGM German Conference on Pattern Recognition, pages 467-483. Springer, 2021. 6 +[3] Faisal Alamri and Anjan Dutta. Multi-head self-attention via vision transformer for zero-shot learning. arXiv preprint arXiv:2108.00045, 2021. 6 +[4] Jacopo Cavazza, Vittorio Murino, and Alessio Del Bue. No adversaries to zero-shot learning: Distilling an ensemble of gaussian feature generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 6 +[5] Samet Cetin. Closed-form sample probing for training generative models in zero-shot learning. Master's thesis, Middle East Technical University, 2022. 6 +[6] Dubing Chen, Yuming Shen, Haofeng Zhang, and Philip HS Torr. Zero-shot logit adjustment. arXiv preprint arXiv:2204.11822, 2022. 1 +[7] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 122-131, 2021. 1, 2, 3, 5, 6, 7 +[8] Shiming Chen, GuoSen Xie, Yang Liu, Qinmu Peng, Baigui Sun, Hao Li, Xinge You, and Ling Shao. Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems, 34:16622–16634, 2021. 1, 6 +[9] Shiming Chen, Ziming Hong, Wenjin Hou, Guo-Sen Xie, Yibing Song, Jian Zhao, Xinge You, Shuicheng Yan, and Ling Shao. Transzero++: Cross attribute-guided transformer for zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3 +[10] Shiming Chen, Ziming Hong, Yang Liu, Guo-Sen Xie, Baigui Sun, Hao Li, Qinmu Peng, Ke Lu, and Xinge You. Transzero: Attribute-guided transformer for zero-shot learning. In AAAI, page 3, 2022. 3 +[11] Shiming Chen, Ziming Hong, Guo-Sen Xie, Wenhan Yang, Qinmu Peng, Kai Wang, Jian Zhao, and Xinge You. Msdn: Mutually semantic distillation network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7612-7621, 2022. 3 +[12] Shiming Chen, Wenjin Hou, Ziming Hong, Xiaohan Ding, Yibing Song, Xinge You, Tongliang Liu, and Kun Zhang. Evolving semantic prototype improves generative zero-shot learning. arXiv preprint arXiv:2306.06931, 2023. 1, 3, 5, 6 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3 + +[14] Zhuo Chen, Yufeng Huang, Jiaoyan Chen, Yuxia Geng, Wen Zhang, Yin Fang, Jeff Z Pan, and Huajun Chen. Duet: Cross-modal semantic grounding for contrastive zero-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 405-413, 2023. 6 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5, 7 +[16] Yaogong Feng, Xiaowen Huang, Pengbo Yang, Jian Yu, and Jitao Sang. Non-generative generalized zero-shot learning via task-correlated disentanglement and controllable samples synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2022. 6 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[18] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, pages 11494–11503, 2023. 2 +[19] Akshita Gupta, Sanath Narayan, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Joost van de Weijer. Generative multi-label zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[20] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2371-2381, 2021. 1, 3, 6 +[21] Ziming Hong, Shiming Chen, G Xie, Wenhan Yang, Jian Zhao, Yuanjie Shao, Qinmu Peng, and Xinge You. Semantic compression embedding for generative zero-shot learning. *IJCAI*, Vienna, Austria, 7:956–963, 2022. 1 +[22] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4483-4493, 2020. 3, 6 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[24] Xia Kong, Zuodong Gao, Xiaofan Li, Ming Hong, Jun Liu, Chengjie Wang, Yuan Xie, and Yanyun Qu. En-compactness: Self-distillation embedding & contrastive generation for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9306–9315, 2022. 1, 3, 6 +[25] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 3 +[26] Man Liu, Feng Li, Chunjie Zhang, Yunchao Wei, Huihui Bai, and Yao Zhao. Progressive semantic-visual mutual adaptation for generalized zero-shot learning. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15337-15346, 2023. 5, 6 +[27] Yang Liu, Lei Zhou, Xiao Bai, Yifei Huang, Lin Gu, Jun Zhou, and Tatsuya Harada. Goal-oriented gaze estimation for zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3794-3803, 2021. 6 +[28] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12664-12673, 2020. 6 +[29] Muhammad Ferjad Naeem, Yongqin Xian, Luc V Gool, and Federico Tombari. I2dformer: Learning image to document attention for zero-shot image classification. Advances in Neural Information Processing Systems, 35:12283-12294, 2022. 3 +[30] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. I2mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3 +[31] Sanath Narayan, Akshita Gupta, Fahad Shahbaz Khan, Cees GM Snoek, and Ling Shao. Latent embedding feedback and discriminative features for zero-shot classification. In European Conference on Computer Vision, pages 479-495. Springer, 2020. 1, 2, 3, 5, 6, 7 +[32] Sanath Narayan, Akshita Gupta, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Mubarak Shah. Discriminative region-based multi-label zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8731-8740, 2021. 3 +[33] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. Advances in neural information processing systems, 22, 2009. 1 +[34] Genevieve Patterson and James Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2751-2758. IEEE, 2012. 2, 5, 6 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[36] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: lightweight image captioning prompted with retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2840-2849, 2023. 2 +[37] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 49-58, 2016. 6 + +[38] Ivan Skorokhodov and Mohamed Elhoseiny. Class normalization for (continual)? generalized zero-shot learning. arXiv preprint arXiv:2006.11328, 2020. 6 +[39] Jie Song, Chengchao Shen, Yezhou Yang, Yang Liu, and Mingli Song. Transductive unbiased embedding for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1024-1033, 2018. 3 +[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[42] Chaoqun Wang, Shaobo Min, Xuejin Chen, Xiaoyan Sun, and Houqiang Li. Dual progressive prototype network for generalized zero-shot learning. Advances in Neural Information Processing Systems, 34:2936-2948, 2021. 3 +[43] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3032-3042, 2023. 1, 6 +[44] P. Welinder, S. Branson, T. Mita, C. Wah, Florian Schroff, Serge J. Belongie, and P. Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-2010-001, Caltech., 2010. 2, 5, 6 +[45] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6, 7 +[46] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5542-5551, 2018. 1, 2, 3, 5, 7 +[47] Yongqin Xian, Saurabh Sharma, Bernt Schiele, and Zeynep Akata. f-vaegan-d2: A feature generating framework for any-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10275-10284, 2019. 1, 2, 3, 5, 6 +[48] Guo-Sen Xie, Li Liu, Xiaobo Jin, Fan Zhu, Zheng Zhang, Jie Qin, Yazhou Yao, and Ling Shao. Attentive region embedding network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9384-9393, 2019. 3 +[49] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 3 +[50] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Vgse: Visually-grounded semantic embeddings for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9316-9325, 2022. 7, 8 + +[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3 +[52] Meng Ye and Yuhong Guo. Progressive ensemble networks for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11728-11736, 2019. 6 +[53] Sheng Zhang, Muzammal Naseer, Guangyi Chen, Zhiqiang Shen, Salman Khan, Kun Zhang, and Fahad Shahbaz Khan. S3a: Towards realistic zero-shot classification via self structural semantic alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7278-7286, 2024. 1 +[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3 \ No newline at end of file diff --git a/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/images.zip b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe4300d1f901569e0918e385c1fc57bef00c4e9f --- /dev/null +++ b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12949bfe39352cf26a5b87b272b0905caf68c88ab36a403ce305b432127ab0a4 +size 675982 diff --git a/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/layout.json b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bef12ef2b0010d2e15cba8db4d14c8829517b249 --- /dev/null +++ b/2024/Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning/layout.json @@ -0,0 +1,9833 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "type": "text", + "content": "Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": "Wenjin Hou" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Shiming Chen" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Shuhuang Chen" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Ziming Hong" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Yan Wang" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Xuetao Feng" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Salman Khan" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Fahad Shahbaz Khan" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{2,6}" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "text", + "content": ", Xinge You" + }, + { + "bbox": [ + 70, + 160, + 523, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 189, + 458, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 189, + 458, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 189, + 458, + 203 + ], + "type": "text", + "content": "1Huazhong University of Science and Technology (HUST), China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 462, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 462, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 462, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 203, + 462, + 217 + ], + "type": "text", + "content": "Mohamed bin Zayed University of AI " + }, + { + "bbox": [ + 130, + 203, + 462, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 203, + 462, + 217 + ], + "type": "text", + "content": "The University of Sydney" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 217, + 486, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 217, + 486, + 232 + ], + "spans": [ + { + "bbox": [ + 119, + 217, + 486, + 232 + ], + "type": "text", + "content": "4Alibaba Group 5Australian National University 6Linkoping University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 233, + 451, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 233, + 451, + 245 + ], + "spans": [ + { + "bbox": [ + 139, + 233, + 451, + 245 + ], + "type": "text", + "content": "{houwj17, gchenshiming}@gmail.com youxg@mail.hust.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "text", + "content": "Generative Zero-shot learning (ZSL) learns a generator to synthesize visual samples for unseen classes, which is an effective way to advance ZSL. However, existing generative methods rely on the conditions of Gaussian noise and the predefined semantic prototype, which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). To address this issue, we propose a novel Visual-Augmented Dynamic Semantic prototype method (termed VADS) to boost the generator to learn accurate semantic-visual mapping by fully exploiting the visual-augmented knowledge into semantic conditions. In detail, VADS consists of two modules: (1) Visual-aware Domain Knowledge Learning module (VDKL) learns the local bias and global prior of the visual features (referred to as domain visual knowledge), which replace pure Gaussian noise to provide richer prior noise information; (2) Vision-Oriented Semantic Updation module (VOSU) updates the semantic prototype according to the visual representations of the samples. Ultimately, we concatenate their output as a dynamic semantic prototype, which serves as the condition of the generator. Extensive experiments demonstrate that our VADS achieves superior CZSL and GZSL performances on three prominent datasets and outperforms other state-of-the-art methods with averaging increases by " + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 47, + 298, + 289, + 610 + ], + "type": "text", + "content": " on SUN, CUB and AWA2, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 656, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 656, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 656, + 288, + 693 + ], + "type": "text", + "content": "Zero-shot learning [33], which transfers knowledge from seen classes to unseen classes, has garnered much attention recently. By establishing interactions between visual fea" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 274, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 547, + 418 + ], + "type": "text", + "content": "tures and semantic prototypes (also referred to as attribute vectors, side information, or semantic embeddings [47]), generative ZSL methods exhibit impressive performance, demonstrating the potential of feature synthesis. One of the most successful frameworks is the conditional generative adversarial network (GAN) [17]. The main idea of generative ZSL methods is to align semantic prototypes and visual features to synthesize feature of unseen classes. Recent emerging studies have either designed more effective frameworks [7, 19, 20, 31, 46] or addressed more specific issues related to visual-semantic alignment [4, 8, 12, 43, 53]. These methods have achieved significant improvements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "However, these methods rely on the conditions of Gaussian noise and the predefined semantic prototype (referred to as the static semantic prototype), which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). Fig. 1 illustrates these issues: (1) The noise is sampled from a Gaussian distribution " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", which lacks the dataset-specific visual prior knowledge (e.g., global visual information \"flying\" and \"still\" and background information \"sky\" and \"grass\"). As a result, the domain knowledge shared between seen and unseen classes cannot be utilized for feature synthesis of unseen classes, limiting the knowledge transfer. (2) The predefined semantic prototype fails to characterize each instance well. For example, the attributes \"wing black\", \"breast white\" and \"bill orange\" of the Laysan Albatross are not fixed on different images. Due to these limitations, the visual features synthesized by existing works [4, 6, 21, 24, 31, 43, 46] struggle to represent the distribution of real features, leading to poor generalization to unseen classes, as shown in Fig. 1(b). More intuitively, as shown in Fig. 1(d), the features of unseen classes synthesized by these methods are confusing, resulting in the decision boundary overfitting to the seen classes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23627" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 178, + 180 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 178, + 180 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 178, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 178, + 180 + ], + "type": "image", + "image_path": "38bcce7d3c04da7ae89f2b372b3204866c504d38b6464ae121b7f8bfd8236ab6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 74, + 285, + 180 + ], + "blocks": [ + { + "bbox": [ + 178, + 74, + 285, + 180 + ], + "lines": [ + { + "bbox": [ + 178, + 74, + 285, + 180 + ], + "spans": [ + { + "bbox": [ + 178, + 74, + 285, + 180 + ], + "type": "image", + "image_path": "67cb4509fa7045f419b184adb3d350086b1557ff3f285d5661caba16a6303bf6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 285, + 74, + 391, + 179 + ], + "blocks": [ + { + "bbox": [ + 285, + 74, + 391, + 179 + ], + "lines": [ + { + "bbox": [ + 285, + 74, + 391, + 179 + ], + "spans": [ + { + "bbox": [ + 285, + 74, + 391, + 179 + ], + "type": "image", + "image_path": "a6fb2437c9d2299a0ace19e0727ffeae5f5c078a27dd5f88ce67f95a73e7bdc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 74, + 531, + 180 + ], + "blocks": [ + { + "bbox": [ + 391, + 74, + 531, + 180 + ], + "lines": [ + { + "bbox": [ + 391, + 74, + 531, + 180 + ], + "spans": [ + { + "bbox": [ + 391, + 74, + 531, + 180 + ], + "type": "image", + "image_path": "a52a90e9f396589cc762815c104667531bcc7786000477168e194edbb475667d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 181, + 177, + 286 + ], + "blocks": [ + { + "bbox": [ + 62, + 181, + 177, + 286 + ], + "lines": [ + { + "bbox": [ + 62, + 181, + 177, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 181, + 177, + 286 + ], + "type": "image", + "image_path": "4af521c52b95043c9d12143da9279ff25d0005307d0bfcbd985bfd82b4118f60.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 288, + 143, + 297 + ], + "lines": [ + { + "bbox": [ + 96, + 288, + 143, + 297 + ], + "spans": [ + { + "bbox": [ + 96, + 288, + 143, + 297 + ], + "type": "text", + "content": "(a) Input image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 181, + 285, + 286 + ], + "blocks": [ + { + "bbox": [ + 178, + 181, + 285, + 286 + ], + "lines": [ + { + "bbox": [ + 178, + 181, + 285, + 286 + ], + "spans": [ + { + "bbox": [ + 178, + 181, + 285, + 286 + ], + "type": "image", + "image_path": "231add22ada34f480de0f3711fa79bdcd8426404e547fac4b13411f49c65c7c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 287, + 258, + 297 + ], + "lines": [ + { + "bbox": [ + 201, + 287, + 258, + 297 + ], + "spans": [ + { + "bbox": [ + 201, + 287, + 258, + 297 + ], + "type": "text", + "content": "(b) Existing works" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 285, + 181, + 391, + 286 + ], + "blocks": [ + { + "bbox": [ + 285, + 181, + 391, + 286 + ], + "lines": [ + { + "bbox": [ + 285, + 181, + 391, + 286 + ], + "spans": [ + { + "bbox": [ + 285, + 181, + 391, + 286 + ], + "type": "image", + "image_path": "b53f1f0ecec2b8820955ccf422c0e2db2770a6e4ea9db8c81760100a7b26a62f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 287, + 359, + 296 + ], + "lines": [ + { + "bbox": [ + 315, + 287, + 359, + 296 + ], + "spans": [ + { + "bbox": [ + 315, + 287, + 359, + 296 + ], + "type": "text", + "content": "(c) Our VADS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 309, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 545, + 376 + ], + "type": "text", + "content": "Figure 1. An illustration of the core idea of our method. (a) The semantic prototype (i.e., attribute) of different images of the same category is not fixed, so the predefined semantic prototype is inaccurate in characterizing each instance. (b) Most existing works utilize Gaussian noise and the predefined semantic prototype as conditions to train a semantic→visual generator on seen classes, which fails to generalize to unseen classes. (c)(d) Our method incorporates rich visual prior with an updated semantic prototype to construct a visual-augmented dynamic semantic prototype of each instance, empowering the generator to synthesize features that faithfully represent the real distribution of unseen classes. Thus, our method achieves better generalization on seen and unseen classes than existing works (e.g., CLSWGAN [46])." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 188, + 531, + 286 + ], + "blocks": [ + { + "bbox": [ + 394, + 188, + 531, + 286 + ], + "lines": [ + { + "bbox": [ + 394, + 188, + 531, + 286 + ], + "spans": [ + { + "bbox": [ + 394, + 188, + 531, + 286 + ], + "type": "image", + "image_path": "5c2fde26257060e3e09fd9027fd21b60e4f6c3677b84083f55b79436cb3a1525.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 288, + 496, + 297 + ], + "lines": [ + { + "bbox": [ + 423, + 288, + 496, + 297 + ], + "spans": [ + { + "bbox": [ + 423, + 288, + 496, + 297 + ], + "type": "text", + "content": "(d) Feature visualization" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 388, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 604 + ], + "type": "text", + "content": "Drawing inspiration from image captioning [36], which highlights the generalization of instance-conditional learning, we aim to fully leverage the visual-augmented knowledge into semantic conditions to tackle the aforementioned challenges. On the one hand, we can exploit rich domain visual prior knowledge, serving as a prior noise, to enhance the adaptation and generalization of models [18, 51, 54]. On the other hand, we can update the predefined semantic prototype to align visual representations based on visual features. As such, incorporating richer and more accurate visual information acts as the semantic condition to train an instance-conditional generative model, which is optimized to characterize each instance (more robust to class shift) rather than to serve only for specific classes. Accordingly, the generative model can synthesize features of unseen classes closer to the real ones, facilitating the classifier in learning an appropriate decision boundary (see the right of Fig. 1(d))." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we propose an approach called Visual-Augmented Dynamic Semantic prototype (VADS) to improve generative ZSL methods. Specifically, VADS consists of two learnable modules: a Visual-aware Domain Knowledge Learning module (VDKL) and a Vision-Oriented Semantic Updation module (VOSU). The VDKL explores domain visual prior knowledge derived from visual information, which provides richer information for representing instances. The VOSU predicts instance-level semantics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": "through visual " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " semantic mapping, guiding the updation of the predefined semantic prototype and promoting accurate semantic prototype learning. Finally, the extracted visual prior and the updated semantic prototype are concatenated as a visual-augmented dynamic semantic prototype, which serves as the condition of the generator during training and feature synthesis, as illustrated in Fig. 1(c). Extensive experiments demonstrate the effectiveness of our VADS." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 488, + 516, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 516, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 516, + 499 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 510, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 306, + 510, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 545, + 556 + ], + "type": "text", + "content": "- We introduce a Visual-Augmented Dynamic Semantic prototype (VADS) to enhance the generalization of generative ZSL methods, facilitating substantial knowledge transfer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 558, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 545, + 617 + ], + "type": "text", + "content": "- We devise the VDKL to leverage domain visual prior knowledge from visual features and design the VOSU to dynamically update the predefined semantic prototype. Their outputs together serve as the generator's conditions, providing richer and more accurate visual information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "content": "- We conduct extensive experiments on AWA2 [47], SUN [34] and CUB [44] datasets. The comprehensive results demonstrate that visual prior knowledge significantly improves the generalization of generative ZSL methods, i.e., average improvements of the harmonic mean over existing generative methods (e.g., f-CLSWGAN [46], TFVAEGAN [31] and FREE [7]) " + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "content": " on SUN, CUB and AWA2, respectively." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23628" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 95, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 95, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 289, + 323 + ], + "type": "text", + "content": "Embedding-based Zero-Shot Learning. Embedding-based ZSL methods are one of the mainstream branches that project visual information into semantic space to align with semantic prototypes. Earlier works [25, 39, 48] directly mapped global visual features to semantic space, failing to capture local discriminative representation, resulting in suboptimal ZSL performance. Also, embeddings are learned only in seen classes, leading to inevitable seen class bias. In this regard, some studies [9, 10, 22] have attempted to use calibration loss to balance the prediction results between seen and unseen classes. Recently, attention mechanisms [41] have emerged with surprising localization abilities, so semantic-guided methods [9, 11, 29, 30, 32, 42, 48, 49] learn to discover attribute-related local regions, providing more accurate inter-class discrimination. Among these methods, APN [49] proposed an attribute prototype network to learn local features, and DPPN [42] updated attribute and category prototypes. Inspired by their work, we introduce a dynamic semantic prototype for generative ZSL methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 326, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 289, + 531 + ], + "type": "text", + "content": "Generative Zero-Shot Learning. Generative ZSL methods learn semantic " + }, + { + "bbox": [ + 47, + 326, + 289, + 531 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 326, + 289, + 531 + ], + "type": "text", + "content": " visual mapping to synthesize unseen class features, effectively alleviating the lack of unseen class data. Consequently, the quality of synthesized features, which preserves visual-semantic correspondence, plays a crucial role in classification. Thus, TF-VAEGAN [31] forced semantic alignment at all stages, and FREE [7] fine-tuned visual features to address cross-dataset biases. CE-GZSL [20] and ICCE [24] projected visual features into the latent space for classification. However, these methods constructed projection spaces on seen classes, resulting in inferior generalization ability on unseen classes. Moreover, they uniformly utilize the predefined semantic prototype as a condition, making it difficult to achieve accurate visual-semantic alignment. The method most related to ours is DSP [12], which updates the prototype by simply adding the evolved and predefined semantic prototype." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 289, + 714 + ], + "type": "text", + "content": "Large-Scale Vision-Language Models Generalization. Vision-language models like CLIP [35], pre-trained on large-scale image-text pairs, have demonstrated significant potential for downstream tasks. When performing zero-shot recognition, the class prompts are input into the text encoder to obtain the classification weights, and the cosine similarity between the test image and the weights determines the resulting classification score. It is different from the classical ZSL methods [9, 25, 39, 42, 48, 49]. Recent research has focused on improving the generalization to unseen classes, with several previous works proposing prompt learning [1, 51, 54]. Motivated by optimizing visual conditional prompts, we introduce visual-aware domain knowledge learning into generative ZSL methods, facilitating knowledge transfer to unseen classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 545, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 99 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 99 + ], + "type": "text", + "content": "3. Visual-Augmented Dynamic Semantic Prototype Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 105, + 547, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 105, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 105, + 547, + 152 + ], + "type": "text", + "content": "Fig. 2 shows the framework of our VADS. Next, we first present the problem formulation and briefly review the generative ZSL model. Then, we introduce the detailed design of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": "Problem Formulation. Conventional zero-shot learning (CZSL) recognizes unseen classes in the inference stage. Generalized zero-shot learning (GZSL) recognizes both seen and unseen classes. Both settings generalize from seen data " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " to unseen domains " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s = \\{(x_i^s, y_i^s) | x_i^s \\in \\mathcal{X}^s, y_i^s \\in \\mathcal{Y}^s\\}_{i=1}^{N_s}" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " is the sample number of seen classes, " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "x_i^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " is a feature vector in " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "y_i^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " is the corresponding lable from " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " is split into a training set " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{tr}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " and a testing set " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{te}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " following Xian et al. [47]. Similarly, " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{(x_i^u, y_i^u) | x_i^u \\in \\mathcal{X}^u, y_i^u \\in \\mathcal{Y}^u\\}_{i=1}^{N_u}" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "x_i^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " is a feature vector in " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "y_i^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " is the lable from " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^s" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " are disjoint. Define attribute semantic prototypes " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\mathcal{A}^s \\cup \\mathcal{A}^u" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": ", corresponding to each category, as a bridge to transfer knowledge from seen classes to unseen classes. In this paper, we dynamically update " + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 153, + 547, + 345 + ], + "type": "text", + "content": " to learn accurate visual-semantic alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 351, + 437, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 437, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 437, + 363 + ], + "type": "text", + "content": "3.1. Generative ZSL Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": "The goal of the generative ZSL methods is to learn a semantic " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": " visual generative model " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "(G)" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": " on seen classes and then use it to synthesize samples of unseen classes to train a classifier. Existing methods use Gaussian noise and the predefined semantic prototype as input conditions to supervise " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": " synthetic features (i.e., " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\times \\mathcal{Z} \\rightarrow \\hat{X}" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": "). In our method, " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": " represents an off-the-shelf CLSWGAN [46], which contains a generator and a discriminator. We develop the dynamic semantic prototype as a condition, allowing " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": " to characterize more accurate visual-semantic relationships." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 496, + 545, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 496, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 496, + 545, + 520 + ], + "type": "text", + "content": "3.2. Visual-aware Domain Knowledge Learning (VDKL)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "text", + "content": "Drawing inspiration from previous prompt learning [51, 54], we exploit the rich information in visual features to assist in synthesizing features. VDKL is a data-efficient module allowing the visual features to be used to improve generalization. As shown in Fig. 2, we design a Visual Encoder (VE) and a Domain Knowledge Learning network (DKL). First, the VE encodes visual features into a latent feature " + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "text", + "content": " and a latent code " + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "text", + "content": ". The latent feature enables inter-class alignment of visual features, and latent code is subsequently confined to a prior distribution " + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 526, + 547, + 671 + ], + "type": "text", + "content": ". The optimization of the VE is achieved via contrastive loss [13] and evidence-lower bound given by the equation as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 687, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 687, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 687, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = \\mathbb {E} [ \\log \\frac {\\exp \\left(l _ {i} ^ {T} l ^ {+} / \\tau\\right)}{\\exp \\left(l _ {i} ^ {T} l ^ {+} / \\tau\\right) + \\sum_ {k = 1} ^ {K} \\exp \\left(l _ {i} ^ {T} l _ {k} ^ {-} / \\tau\\right)} ], \\tag {1}", + "image_path": "46fca2a0459b5425cd7bd99f5639693fb71ad1821cf8cdf83a5df5bf8e8fd049.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23629" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 73, + 525, + 366 + ], + "blocks": [ + { + "bbox": [ + 67, + 73, + 525, + 366 + ], + "lines": [ + { + "bbox": [ + 67, + 73, + 525, + 366 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 525, + 366 + ], + "type": "image", + "image_path": "d06d19ef86e59dd2205450a03a94b64dc3fbbd5037dfd1ed8b67bb677ad80d91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": "Figure 2. The architecture of our proposed VADS. It consists of two learnable modules: a Visual-Oriented Semantic updation module (VOSU) and a Visual-aware Domain Knowledge Learning module (VDKL). First, we obtain the prior distribution " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": " by the Visual Encoder " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "(VE)" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": ". Following this, the Domain Knowledge Learning network " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "(DKL)" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": " transforms " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": " into a local bias " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": ", which is subsequently added to global learnable prior vectors " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": " to construct the domain visual prior noise (i.e., " + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}'" + }, + { + "bbox": [ + 46, + 378, + 548, + 449 + ], + "type": "text", + "content": "). At the bottom, VOSU notably updates the semantic prototype in two stages (depicted by the blue and green arrows). Finally, the visual prior noise and the updated semantic prototype together form a dynamic semantic prototype, used for the reconstruction of features by the generator." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 466, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 466, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 111, + 466, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k l} = K L (V E (x) \\| p (z)), \\tag {2}", + "image_path": "4eebc72c727c9a05171633234dfbb28334ef0dcb496253334956cd2ec281fae5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "l^{+}" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "l_{k}^{-}" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": " represent positive and negative latent features, " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": " is a temperature parameter set as 0.15, " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": " is the class numbers, " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": " denotes the Kullback-Leibler divergence and " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": " is a prior distribution that is assumed to be " + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 486, + 287, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "text", + "content": "To further utilize visual prior knowledge during the training and synthesis stages, we propose a Domain Knowledge Learning network (DKL) to obtain a local bias " + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "text", + "content": " of visual features (i.e., " + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{b} = DKL(\\pmb{Z})" + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "text", + "content": "). Additionally, we employ a learnable prior vector " + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "text", + "content": " to capture global visual information (" + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 534, + 288, + 619 + ], + "type": "text", + "content": " is randomly initialized). Subsequently, we obtain domain-specific visual prior noise as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 628, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 628, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 77, + 628, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} ^ {\\prime} = \\alpha \\cdot (\\boldsymbol {b} + \\boldsymbol {p}) + (1 - \\alpha) \\cdot \\text {n o i s e} (\\mathbf {0}, \\mathbf {1}), \\quad (3)", + "image_path": "331fced4954647c82b54a16ffd13fe53e0de2876560ac16100a5ff108e33a44f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": "where the noise " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " represents Gaussian noise aimed at enhancing diversity in synthesis, " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " is the combination coefficient set as 0.9. Through this operation, we argue that " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "Z^{\\prime}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " includes rich domain visual knowledge and feeds it into the generator to provide instance conditions, promoting the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "type": "text", + "content": "generator learning and utilizing it for feature synthesis of unseen classes. Note that unseen class samples are unavailable in the feature synthesis stage, so we randomly sample Gaussian noise input to " + }, + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "type": "inline_equation", + "content": "DKL" + }, + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "type": "text", + "content": ", transferring the domain knowledge acquired from seen classes to unseen classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 536, + 537, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 536, + 537, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 536, + 537, + 550 + ], + "type": "text", + "content": "3.3. Visual-Oriented Semantic updation (VOSU)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": "We observe that the predefined semantic prototype struggles to represent each visual sample accurately, so we propose a Visual-Oriented Semantic Updation module (VOSU), optimizing the semantic prototype dynamically. Our semantic prototype updation involves a two-stage process. In the first stage, we feed visual features " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " into the Visual-Semantic Prediction network (VSP) to generate a predictive semantic " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\hat{a}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " that explicitly captures specific visual patterns of the target image. Then, the predefined semantic prototype is input into a Semantic Updation Mapping network (SUM) to learn an updated semantic " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\dot{a}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": ". This mapping can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 394, + 697, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 697, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 394, + 697, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\dot {a} = S U M (a). \\tag {4}", + "image_path": "fb15c5a4d935ea365ce44c8fd5dbd92b4b59838393bbf74d2ce9115ee7d0772d.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23630" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "To maintain the attribute information of the prototype and integrate the visual information, we jointly optimize them by the cross-entropy loss " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 118, + 287, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 287, + 150 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(V S P \\left(x _ {i}\\right) ^ {T} \\dot {a} ^ {y}\\right)}{\\sum_ {\\bar {c} \\in C ^ {s} \\cup u} \\exp \\left(V S P \\left(x _ {i}\\right) ^ {T} \\dot {a} ^ {\\bar {c}}\\right)}, \\tag {5}", + "image_path": "6dfa55f37122b876678632f0508dc6129785830f66ad3fadb67e5a2fd40d0d76.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "content": " is the batch numbers. Accordingly, the updated semantic prototype incorporates rich visual information. In the second stage, we employ SUM to update " + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\dot{a}" + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "content": " during the conditional generative model training and use it as a condition to learn together with the generator " + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "content": " and the discriminator " + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "content": ". This implementation facilitates dynamic updation and accurate visual-semantic matching. To this end, we propose the semantic consistency loss " + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sc}" + }, + { + "bbox": [ + 47, + 158, + 287, + 255 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 264, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 264, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 108, + 264, + 287, + 277 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s c} = \\mathbb {E} \\left[ \\| S U M (\\dot {a}) - \\dot {a} \\| _ {1} \\right]. \\tag {6}", + "image_path": "d63a3dc615445c64f920185faa461ccd60a7d1645fccfc24c2b39408fcd5060c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 287, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 287, + 358 + ], + "type": "text", + "content": "In summary, the first stage leverages visual features to assist semantic updation, and the second stage dynamically updates the prototype of each sample. Then, we concatenate the updated semantic prototype with the visual prior noise " + }, + { + "bbox": [ + 47, + 287, + 287, + 358 + ], + "type": "inline_equation", + "content": "Z^{\\prime}" + }, + { + "bbox": [ + 47, + 287, + 287, + 358 + ], + "type": "text", + "content": ", called the dynamic semantic prototype, which serves as the condition for the generator, as depicted in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 366, + 220, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 220, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 220, + 379 + ], + "type": "text", + "content": "3.4. Overall Objective and Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 385, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 287, + 408 + ], + "type": "text", + "content": "VADS Objective Loss Function. Overall, the objective loss function of VADS is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 419, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 419, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 73, + 419, + 287, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {G} + \\lambda_ {\\text {c o n}} \\mathcal {L} _ {\\text {c o n}} + \\lambda_ {k l} \\mathcal {L} _ {k l} + \\lambda_ {s c} \\mathcal {L} _ {s c}, \\tag {7}", + "image_path": "c77a4f0c1324b71b66b036ce194c05c170076375437199294b0cd5e4d6051736.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G" + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "text", + "content": " is the loss of conditional generative model G, " + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{con}" + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{kl}" + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{sc}" + }, + { + "bbox": [ + 47, + 441, + 287, + 514 + ], + "type": "text", + "content": " are the hyper-parameter to balance each loss term. To fully validate our method, by using this loss, we train on various mainstream generative models (e.g., CLSWAGN [46], TFVAEGAN [31], and FREE [7]). Next, we illustrate feature synthesis and classifier training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "text", + "content": "Visual-Augmented Feature Synthesis for Unseen Classes. To fully utilize the visual knowledge and accurate semantic prototypes, we sample Gaussian noise input " + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "inline_equation", + "content": "DKL" + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "text", + "content": " to obtain the prior noise " + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{\\prime}" + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "text", + "content": " (i.e., Eq. (3)) and use SUM to update the semantic prototypes of unseen classes (i.e., " + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\dot{a}_u = SUM(\\hat{a}_u)" + }, + { + "bbox": [ + 47, + 525, + 287, + 620 + ], + "type": "text", + "content": "). They serve as conditions to synthesize visual samples closer to the real features for training the classifier. The form can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {u} = \\mathrm {G} \\left(\\boldsymbol {Z} ^ {\\prime}, \\dot {a} _ {u}\\right), \\tag {8}", + "image_path": "d8efd059acc6927cf51542a70c331f2bdce2b754ad794584a714a09cbf0fbc2b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\dot{a}_u" + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": " is updated semantic prototype of the unseen classes and " + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\hat{x}_u" + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": " is the synthesized features of unseen classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "ZSL Classifier Training and Inference. After synthesizing features, we input the seen class training features and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": "synthesized unseen class features into " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": " to extract latent features and concatenate them to enhance the original features, alleviating cross-dataset bias [7]. Then, we train a CZSL classifier using enhanced-synthetic features (i.e., " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "f_{CZSL}:\\mathcal{X}\\to \\mathcal{Y}^u" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": ") and train a GZSL classifier using enhanced seen class training features and enhanced-synthetic features (i.e., " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "f_{GZSL}:\\mathcal{X}\\to \\mathcal{Y}^s\\cup \\mathcal{Y}^u" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": "). Finally, we perform inference using the test sets " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{te}^{s}" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 180, + 386, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 386, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 386, + 193 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 200, + 422, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 422, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 422, + 213 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 219, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 289 + ], + "type": "text", + "content": "Benchmark Datasets. We conduct extensive experiments on three prominent ZSL benchmark datasets: Animals with Attributes 2 (AWA2 [47]), SUN Attribute (SUN [34]) and Caltech-USCD Birds-200-2011 (CUB [44]). We follow the Proposed Split (PS) setting [47] to split each dataset into seen and unseen classes, as detailed in Tab. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "text", + "content": "Evaluation Protocols. During inference (i.e., performing CZSL and GZSL classification), we follow the evaluation protocols in [47]. In the CZSL setting, we calculate the average per-class Top-1 accuracy of unseen classes, denoted as " + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "text", + "content": ". For the GZSL scenario, we measure the Top-1 accuracy of the seen and unseen classes, defined as " + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "text", + "content": ", respectively. We also compute the harmonic mean, defined as " + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = (2 \\times \\mathbf{S} \\times \\mathbf{U}) / (\\mathbf{S} + \\mathbf{U})" + }, + { + "bbox": [ + 305, + 291, + 545, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "text", + "content": "Implementation Details. We follow PSVMA [26] using the ViT-Base Backbone [15] without fine-tuning as the feature extractor, obtaining 768-dimensional visual features for all samples. The global prior " + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "text", + "content": " has the same dimension as the semantic prototype. We set the mini-batch to 64, 128 and 128 for AWA2, SUN and CUB, respectively. We use the Adam optimizer [23] with " + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.5" + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 305, + 387, + 545, + 517 + ], + "type": "text", + "content": ", and set the initial learning rate to 0.0001. We synthesize 5600, 100, and 400 samples for each class on AWA2, SUN and CUB. Our experiments are based on the PyTorch and implemented on a NVIDIA GeForce RTX 3090 GPU." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 527, + 531, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 531, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 531, + 540 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": "We report the performance of our proposed VADS using CLSWGAN as a generative model compared to state-of-the-art methods. Tab. 1 shows the results, including the embedding-based and generative ZSL methods. In the CZSL scenario, our method notably outperforms the suboptimal results by " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", achieving the best results on AWA2, SUN, and CUB. The results confirm that our method incorporating dynamic semantic prototypes is more generalizable to unseen classes than static semantic prototypes. In the GZSL scenario, our method obtains the best harmonic mean " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": " on all datasets (i.e., AWA2 " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "(H = 79.3)" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", SUN " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "(H = 55.7)" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": " and CUB " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "(H = 74.3)" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": "). Our method significantly outperforms CLSWGAN+DSP [12], which proposed an evolved semantic prototype, indicat" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23631" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 547, + 329 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Compared our VADS with the state-of-the-art on AWA2, SUN and CUB benchmark datasets in the CZSL and GZSL settings. The best and second-best results are marked in Red and Blue, respectively. Symbol “-” denotes no results are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 547, + 329 + ], + "type": "table", + "html": "
TypeMethodsVenueBackboneAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
EmbeddingPREN [52]CVPR'19ResNet-10174.132.488.647.462.935.427.230.866.435.255.843.1
DAZLE [22]CVPR'20ResNet-10167.960.375.767.159.452.324.333.266.056.759.658.1
DVBE [28]CVPR'20ResNet-101-63.670.867.0-45.037.240.7-53.260.256.5
CN [38]ICLR'21ResNet-101-60.277.167.6-44.741.643.1-49.950.750.3
GEM-ZSL [27]CVPR'21ResNet-10167.364.877.570.662.838.135.736.977.864.877.170.4
ViT-ZSL [3]IMVIP'21ViT-Large-51.990.065.8-44.555.349.3-67.375.271.0
IEAM-ZSL [2]DGAM'21ViT-Large-53.789.967.2-48.254.751.3-68.673.871.1
DUET [14]AAAI'23ViT-Base69.963.784.772.764.445.745.845.872.362.972.867.5
PSVMA [26]CVPR'23ViT-Base-73.677.375.4-61.745.352.3-70.177.873.8
Generativef-VAEGAN-D2 [47]CVPR'19ResNet-10171.157.670.663.564.745.138.041.361.048.460.153.6
TF-VAEGAN [31]ECCV'20ResNet-10172.259.875.166.666.045.640.743.064.952.864.758.1
FREE [7]ICCV'21ResNet-101-60.475.467.1-47.437.241.7-55.759.957.7
HSVA [8]NeurIPS'21ResNet-101-56.779.866.3-48.639.043.3-52.758.355.3
CE-GZSL [20]CVPR'21ResNet-10170.463.178.670.063.348.838.643.177.563.966.865.3
FREE+ESZSL [5]ICLR'22ResNet-101-51.378.061.8-48.236.541.5-51.660.455.7
ICCE [24]CVPR'22ResNet-10172.765.382.372.8----78.467.365.566.4
TDCSS [16]CVPR'22ResNet-101-59.274.966.1-----44.262.851.9
CDL + OSCO [4]TPAMI'23ResNet-101-48.071.057.1-32.065.042.9-29.069.040.6
CLSWGAN+DSP [12]ICML'23ResNet-101-60.086.070.7-48.343.045.5-51.463.856.9
TFVAEGAN+SHIP [43]ICCV'23ViT-Base-61.295.974.7-----22.582.235.3
VADS (Ours)-ViT-Base82.575.483.679.376.364.649.055.786.874.174.674.3
", + "image_path": "7b848c2839264b4b5adc4fa1d3d19e291302a3b4a12d7f06e3251cf7700cc547.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 410, + 273, + 462 + ], + "blocks": [ + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "lines": [ + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "text", + "content": "Table 2. A detailed illustration of the ZSL benchmark datasets. " + }, + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "text", + "content": " are the number of seen and unseen classes. " + }, + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "inline_equation", + "content": "{N}_{A}" + }, + { + "bbox": [ + 47, + 346, + 288, + 401 + ], + "type": "text", + "content": " refers to semantic prototype dimensions. We follow CE-GZSL [20] using 1024-dimensional semantic prototypes generated from textual descriptions [37] on CUB." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 410, + 273, + 462 + ], + "lines": [ + { + "bbox": [ + 56, + 410, + 273, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 273, + 462 + ], + "type": "table", + "html": "
Datasets# images# classes (s | u)# NA
AWA2 [47]3732250 (40 | 10)85
SUN [34]14340717 (645 | 72)102
CUB [44]11788200 (150 | 50)1024
", + "image_path": "7a5d15ab5fdb53e9e4870636d99d5b6469d722165a4a43e1f202f1aaa4ae0446.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 471, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 639 + ], + "type": "text", + "content": "ing that our design is reasonable. Meanwhile, our method is also competitive compared to the recent PSVMA [26], DUET [14], and TFVAEGAN+SHIP [43] methods using the ViT Backbone. Furthermore, our method achieves optimal results in unseen class accuracy " + }, + { + "bbox": [ + 46, + 471, + 287, + 639 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 471, + 287, + 639 + ], + "type": "text", + "content": ", demonstrating that the features synthesized by the generator are closer to the real features of unseen classes, effectively alleviating the over-fitting problem. Noted that TFVAEGAN+SHIP [43] using CLIP Encoder and ViT-ZSL [2] using ViT-Large achieve the best and second-best accuracy for seen classes, but they fail to generalize well to unseen classes. These results consistently demonstrate that our method synthesizes reliable features of unseen classes to facilitate classifier learning, resulting in superior ZSL performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 141, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 141, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 141, + 659 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Component Analysis. In this section, we perform a series of experiments to analyze the effectiveness of significant components. Tab. 3 summarizes the results of ablation studies on CUB and AWA2. We first use ViT-Base" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 388, + 538, + 480 + ], + "blocks": [ + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "type": "text", + "content": "Table 3. Ablation study of VADS on modules, feature enhancement and loss terms on CUB and AWA2. We use CLSWGAN [45] as a generative model. The best result is marked in boldface." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 388, + 538, + 480 + ], + "lines": [ + { + "bbox": [ + 315, + 388, + 538, + 480 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 538, + 480 + ], + "type": "table", + "html": "
ConfigurationsCUBAWA2
AccHAccH
(1) VADS w/o VDKL & VOSU80.165.271.369.3
(2) VADS w/o VDKL84.972.878.178.5
(3) VADS w/o VOSU83.870.475.477.0
(4) VADS w/o enhancement85.173.481.879.1
(5) VADS w/o Lcon (i.e., Eq. (1))85.373.179.478.6
(6) VADS w/o Lsc (i.e., Eq. (6))86.073.579.578.7
(7) VADS (full)86.874.382.579.3
", + "image_path": "87ef7d699d47d60f53e1d67f86d2d4c9e861f12cfa5a73b01816b0284a4e07b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "to extract visual features to train CLSWAGN as the baseline. Compared to the baseline (i.e., configuration (1)), our VADS (configuration (7)) performance improves significantly (i.e., the " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "Acc / H" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " increases by " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "6.7\\% / 9.1\\%" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " on CUB and " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "11.5\\% / 10\\%" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " on AWA2). Configurations (2) and (3) correspond to the effectiveness of the VDKL and VOSU modules. When there is no VDKL, that is, no visual prior is introduced to generate samples, the performance drops by " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.9\\% / 1.5\\%" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.4\\% / 0.8\\%" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " for CUB and AWA2, indicating that visual prior is beneficial for transferring knowledge to unseen classes. When without VOSU, the performance drops most severely, verifying visual-semantic alignment is crucial for learning an accurate generator " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": ". Next, without " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " enhancing classification features, the results of configuration (4) drop slightly in both CZSL and GZSL settings, which shows that feature enhancement mitigates cross-dataset bias. Lastly, we conduct experiments to study the impact of the loss terms on performance. Configuration (5) without contrastive loss " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "Acc / H" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " decreases" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23632" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 112, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 4. Evaluation of VADS with multiple generative ZSL models on three prominent datasets using ViT-Base Backbone. Each row pair shows the effect of adding VADS to a particular generative ZSL model. We use the same hyperparameter optimization policy in all cases to make results comparable." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 112, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 112, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 112, + 547, + 213 + ], + "type": "table", + "html": "
Generative ZSL MethodsAWA2SUNCUB
CZSLGZSLCZSLGZSLCZSLGZSL
AccUSHAccUSHAccUSH
CLSWGAN [46]71.366.272.669.366.046.942.644.680.160.071.365.2
CLSWGAN + VADS82.5+11.575.483.679.3+10.076.3+10.364.649.055.7+11.186.8+6.774.174.674.3+9.1
TFVAEGAN [31]78.266.787.175.673.160.648.654.081.664.874.669.3
TFVAEGAN + VADS80.2+2.075.783.379.3+3.776.3+3.261.951.055.9+1.983.6+2.070.170.970.5+1.2
FREE [7]70.662.985.972.671.745.450.447.884.368.773.570.9
FREE + VADS79.4+8.870.184.676.6+4.075.0+3.357.650.753.9+6.185.5+1.270.975.473.1+3.2
", + "image_path": "f35e236a9c3ad42d0aa4c3a6a624d5388b489e1f75690ee605aea09ef31219e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 222, + 284, + 325 + ], + "blocks": [ + { + "bbox": [ + 51, + 222, + 284, + 325 + ], + "lines": [ + { + "bbox": [ + 51, + 222, + 284, + 325 + ], + "spans": [ + { + "bbox": [ + 51, + 222, + 284, + 325 + ], + "type": "image", + "image_path": "5149f26aff000651bc3f3ece0de284bb2acc273238e457bf7a2aecf9baf841f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "type": "text", + "content": "Figure 3. t-SNE visualizations on CUB. The 10 different colors refer to the 5 seen classes and 5 unseen classes that are randomly selected. Please zoom in for a better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "inline_equation", + "content": "1.5\\% / 1.2\\%" + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "inline_equation", + "content": "3.1\\% / 0.7\\%" + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "content": " on CUB and AWA2, respectively. The main reason is that contrastive loss achieves feature alignment, without which the network may learn category-agnostic redundant information. For the lack of semantic consistency loss " + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sc}" + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "content": " (i.e., configuration (6)), there is no guarantee that the semantic prototype maintains the original definition when training on seen classes, thus hurting the performance of seen classes. Note that the cross-entropy loss " + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "content": " and KL loss " + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 46, + 378, + 288, + 498 + ], + "type": "text", + "content": " are necessary for single network training, so we did not perform ablation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Qualitative Evaluation. To further demonstrate that our method synthesizes reliable unseen class features for classifier training, we project various visual features into two principal components via t-SNE [40]. Fig. 3 shows the t-SNE visualization on CUB. Each color of " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " represents a class, and we randomly select 5 seen classes and 5 unseen classes. From left to right, they represent the real visual features extracted by ViT-Base [15], the features synthesized by CLSWGAN, and the features synthesized by our VADS. We observe that the real visual features are interclass dispersion and intra-class aggregation. In Fig. 3(b), Gaussian noise and the predefined semantic prototype serve as conditions to synthesize samples by CLSWGAN. There are two apparent phenomena: first, the feature distributions are scattered, which cannot truly reflect each class; second, the synthesized seen and unseen class features are confusing (e.g., “" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "” and “" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "”), which denote seen class" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 222, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 355 + ], + "type": "text", + "content": "\"Yellow breasted Chat\" and unseen class \"Yellowthroat\", respectively). Therefore, the decision boundary of the CZSL/GZSL classifier trained with these features is unclear, consistent with the motivation of Fig. 2(d). In contrast, the features generated with our VADS are closer to the real features and are inter-class separated, as shown in Fig. 3(c). On the one hand, we analyze that visual-augmented dynamic semantic prototype motivates the generator to learn accurate semantic→visual mapping. On the other hand, synthesized unseen class features are more reliable, leading to learning appropriate classification boundaries of unseen classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 362, + 495, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 495, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 495, + 374 + ], + "type": "text", + "content": "4.4. Generative ZSL Models with VADS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": "To further evaluate VADS as a generic technology to improve generative ZSL, we integrate it into three prevalent generative ZSL frameworks: CLSWGAN [45], TF-VAEGAN [31], and FREE [7]. We use the official repository to reproduce the results and then insert our module to verify our method's effectiveness. Note that TF-VAEGAN and FREE contain a visual encoder, so we maintain their design. When inserting our modules, we keep the hyperparameters unchanged to make the results comparable. The results on the three datasets are presented in Tab. 4. In terms of " + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "inline_equation", + "content": "Acc" + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": ", we observe varying degrees of performance improvements (e.g., a maximum of 11.5 points and a minimum of 1.2 points). Average growth is " + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "inline_equation", + "content": "7.4\\% / 5.9\\%" + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "inline_equation", + "content": "5.6\\% / 6.4\\%" + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "inline_equation", + "content": "3.3\\% / 4.2\\%" + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "inline_equation", + "content": "Acc / H" + }, + { + "bbox": [ + 304, + 381, + 547, + 585 + ], + "type": "text", + "content": " on AWA2, SUN and CUB. Overall, the consistent improvement over competitive benchmarks validates the effectiveness of our proposed method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 593, + 512, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 512, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 512, + 605 + ], + "type": "text", + "content": "4.5. Generalization Analysis of Visual Prior" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "In our method, we learn a local bias and a fixed global prior vector representing domain visual prior knowledge to generalize to unseen classes. Therefore, we investigate the impact of different forms of prior knowledge on performance. The results are detailed in Tab. 5. \"Random\" refers to the prior knowledge sampled from Gaussian distribution, \"VGSE\" [50] indicates that we take the semantic knowledge extracted from the visual representation as the prior, and \"Other domain\" means using the global vector learned" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23633" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 185, + 153 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 185, + 153 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 185, + 153 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 185, + 153 + ], + "type": "image", + "image_path": "5897e4193b5a634d87eef120515a2cf3d2b779360dd96f575a6baf1c13cc4720.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "lines": [ + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "spans": [ + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "text", + "content": "Figure 4. Effect of (a) synthesized samples " + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "inline_equation", + "content": "N_{syn}" + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "text", + "content": ", (b) loss weights " + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "inline_equation", + "content": "\\lambda_{con}" + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "text", + "content": ", (c) loss weights " + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "inline_equation", + "content": "\\lambda_{kl}" + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "text", + "content": ", and (d) loss weights " + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "inline_equation", + "content": "\\lambda_{sc}" + }, + { + "bbox": [ + 64, + 155, + 526, + 167 + ], + "type": "text", + "content": " on CUB." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 188, + 72, + 291, + 153 + ], + "blocks": [ + { + "bbox": [ + 188, + 72, + 291, + 153 + ], + "lines": [ + { + "bbox": [ + 188, + 72, + 291, + 153 + ], + "spans": [ + { + "bbox": [ + 188, + 72, + 291, + 153 + ], + "type": "image", + "image_path": "7284c7164519396e30295f55718079cdd1979921a6a78a3d82b259fcee33eb1d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 398, + 152 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 398, + 152 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 398, + 152 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 398, + 152 + ], + "type": "image", + "image_path": "3060fdfd8a091e672a7aed5831a98f35b66c2a12a37364dcb8dc038110741e99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 72, + 506, + 153 + ], + "blocks": [ + { + "bbox": [ + 402, + 72, + 506, + 153 + ], + "lines": [ + { + "bbox": [ + 402, + 72, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 402, + 72, + 506, + 153 + ], + "type": "image", + "image_path": "ea761a70f0db4f0b9cdab00b58cd52edc14c8bb05fe9112ad403d575f63814df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 54, + 213, + 279, + 293 + ], + "blocks": [ + { + "bbox": [ + 47, + 182, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 204 + ], + "type": "text", + "content": "Table 5. We evaluate different forms of prior knowledge on CUB. The best result is marked in boldface." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 213, + 279, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 279, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 279, + 293 + ], + "type": "table", + "html": "
ConditionLearnableCUB
AccUSH
w/o prior84.969.175.572.2
Random85.672.774.773.4
VGSE [50]✓85.569.272.370.7
Other domain✓84.066.772.069.3
VADS (Ours)✓86.874.174.674.3
", + "image_path": "07a5438ad193208f09c5c2eff46d599018c425d10d42396f26a51a212d421c12.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 313, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 397 + ], + "type": "text", + "content": "from SUN to transfer to CUB. The results indicate that our method utilizes the global prior and the local bias, which yields the best performance. Knowledge transfer from SUN to CUB suffers from a negative impact, underscoring the importance of dataset-specific global information. VGSE captures semantic side information from seen classes, limiting the performance of unseen classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 405, + 198, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 198, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 198, + 418 + ], + "type": "text", + "content": "4.6. Hyper-parameters Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": "We study the impact of different hyper-parameters of our VADS on the CUB dataset. Fig. 4(a) shows the results of synthesizing different numbers per unseen class. The unseen class accuracy varies with the number of synthesized samples, and when " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "N_{syn} = 400" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": ", the performance is optimal. This result demonstrates that the features synthesized by our method alleviate the lack of unseen class data. Next, we evaluate the influence of individual loss weights (i.e., " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{con}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{kl}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{sc}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sc}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": "). The results are presented in Fig. 4(b)(c)(d). The performance of " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " changes slightly as " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " increases. When " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " is set to 1.5, " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " obtains the maximum value. For " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " loss, we find that larger " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{kl}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " achieves better performance because the prior distribution assumption of the data is crucial. Finally, " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sc}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " loss forces the semantic update process to maintain semantic consistency. Its weight " + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{sc}" + }, + { + "bbox": [ + 46, + 423, + 288, + 628 + ], + "type": "text", + "content": " is insensitive to performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "type": "text", + "content": "4.7. Predefined Semantic Prototype vs Updated Semantic Prototype" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "To give a clearer insight into the predefined semantic prototypes and the semantic prototypes updated by our method, we randomly select 10 classes on CUB and calculate their cosine similarity. Then, we visualize them in Fig. 5. We" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 184, + 437, + 260 + ], + "blocks": [ + { + "bbox": [ + 310, + 184, + 437, + 260 + ], + "lines": [ + { + "bbox": [ + 310, + 184, + 437, + 260 + ], + "spans": [ + { + "bbox": [ + 310, + 184, + 437, + 260 + ], + "type": "image", + "image_path": "9420cbfcfe02778e9b392ec3d7cb37637dc42acbb0a3cc5c1538ba485a4a58c8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 262, + 413, + 270 + ], + "lines": [ + { + "bbox": [ + 332, + 262, + 413, + 270 + ], + "spans": [ + { + "bbox": [ + 332, + 262, + 413, + 270 + ], + "type": "text", + "content": "(a) Predefined semantic prototype" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 449, + 184, + 542, + 261 + ], + "blocks": [ + { + "bbox": [ + 449, + 184, + 542, + 261 + ], + "lines": [ + { + "bbox": [ + 449, + 184, + 542, + 261 + ], + "spans": [ + { + "bbox": [ + 449, + 184, + 542, + 261 + ], + "type": "image", + "image_path": "716f6489b306001fa2125ecbd4acc0c38180d7dba3dd38a8a367601bad64889e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 262, + 528, + 270 + ], + "lines": [ + { + "bbox": [ + 452, + 262, + 528, + 270 + ], + "spans": [ + { + "bbox": [ + 452, + 262, + 528, + 270 + ], + "type": "text", + "content": "(b) Updated semantic prototype" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 281, + 545, + 303 + ], + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 303 + ], + "type": "text", + "content": "Figure 5. Visualization of the heatmap of semantic prototype similarity. We randomly select 10 classes on CUB." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 313, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 545, + 385 + ], + "type": "text", + "content": "observe that the similarity between the predefined semantic prototypes is very high. Our method dynamically refines the visual-semantic relationships of each instance based on visual information, making the updated semantic prototype easier to distinguish between categories and achieving more accurate semantic-visual alignment." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 423, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 651 + ], + "type": "text", + "content": "In this work, we propose a novel Visual-Augmented Dynamic Semantic prototype method (VADS) to boost the generator to synthesize reliable features of unseen classes. Considering that rich visual knowledge can effectively generalize to unseen classes, our proposed VADS fully leverages visual information. Specifically, we design a Visual-aware Domain Knowledge Learning module (VDKL) to acquire visual prior and a Vision-Oriented Semantic Updation module (VOSU) to dynamically update the predefined semantic prototype. Ultimately, we concatenate their output to form a dynamic semantic prototype, serving as the condition of the generator to learn accurate semantic-visual mapping and synthesize features of unseen classes. Extensive experiments demonstrate remarkable results in both CZSL and GZSL scenarios. In summary, our study provides a timely insight into reliable feature synthesis, improving the generalization to unseen classes. Additionally, tasks related to knowledge transfer can draw inspiration from this concept." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 667, + 408, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 667, + 408, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 667, + 408, + 681 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work is partially supported by National Key R&D Program of China (2022YFC3301000)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23634" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "text", + "content": "[2] Faisal Alamri and Anjan Dutta. Implicit and explicit attention for zero-shot learning. In DAGM German Conference on Pattern Recognition, pages 467-483. Springer, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 225 + ], + "type": "text", + "content": "[3] Faisal Alamri and Anjan Dutta. Multi-head self-attention via vision transformer for zero-shot learning. arXiv preprint arXiv:2108.00045, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Jacopo Cavazza, Vittorio Murino, and Alessio Del Bue. No adversaries to zero-shot learning: Distilling an ensemble of gaussian feature generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 305 + ], + "type": "text", + "content": "[5] Samet Cetin. Closed-form sample probing for training generative models in zero-shot learning. Master's thesis, Middle East Technical University, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 340 + ], + "type": "text", + "content": "[6] Dubing Chen, Yuming Shen, Haofeng Zhang, and Philip HS Torr. Zero-shot logit adjustment. arXiv preprint arXiv:2204.11822, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 342, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 396 + ], + "type": "text", + "content": "[7] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 122-131, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 398, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 288, + 453 + ], + "type": "text", + "content": "[8] Shiming Chen, GuoSen Xie, Yang Liu, Qinmu Peng, Baigui Sun, Hao Li, Xinge You, and Ling Shao. Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems, 34:16622–16634, 2021. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 509 + ], + "type": "text", + "content": "[9] Shiming Chen, Ziming Hong, Wenjin Hou, Guo-Sen Xie, Yibing Song, Jian Zhao, Xinge You, Shuicheng Yan, and Ling Shao. Transzero++: Cross attribute-guided transformer for zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "type": "text", + "content": "[10] Shiming Chen, Ziming Hong, Yang Liu, Guo-Sen Xie, Baigui Sun, Hao Li, Qinmu Peng, Ke Lu, and Xinge You. Transzero: Attribute-guided transformer for zero-shot learning. In AAAI, page 3, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 621 + ], + "type": "text", + "content": "[11] Shiming Chen, Ziming Hong, Guo-Sen Xie, Wenhan Yang, Qinmu Peng, Kai Wang, Jian Zhao, and Xinge You. Msdn: Mutually semantic distillation network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7612-7621, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "text", + "content": "[12] Shiming Chen, Wenjin Hou, Ziming Hong, Xiaohan Ding, Yibing Song, Xinge You, Tongliang Liu, and Kun Zhang. Evolving semantic prototype improves generative zero-shot learning. arXiv preprint arXiv:2306.06931, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[14] Zhuo Chen, Yufeng Huang, Jiaoyan Chen, Yuxia Geng, Wen Zhang, Yin Fang, Jeff Z Pan, and Huajun Chen. Duet: Cross-modal semantic grounding for contrastive zero-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 405-413, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 547, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 258 + ], + "type": "text", + "content": "[16] Yaogong Feng, Xiaowen Huang, Pengbo Yang, Jian Yu, and Jitao Sang. Non-generative generalized zero-shot learning via task-correlated disentanglement and controllable samples synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 260, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 547, + 304 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 304, + 547, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 547, + 349 + ], + "type": "text", + "content": "[18] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In CVPR, pages 11494–11503, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "text", + "content": "[19] Akshita Gupta, Sanath Narayan, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Joost van de Weijer. Generative multi-label zero-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 393, + 547, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 547, + 446 + ], + "type": "text", + "content": "[20] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2371-2381, 2021. 1, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "type": "text", + "content": "[21] Ziming Hong, Shiming Chen, G Xie, Wenhan Yang, Jian Zhao, Yuanjie Shao, Qinmu Peng, and Xinge You. Semantic compression embedding for generative zero-shot learning. *IJCAI*, Vienna, Austria, 7:956–963, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 492, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 547, + 536 + ], + "type": "text", + "content": "[22] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4483-4493, 2020. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 537, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 568 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 570, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 635 + ], + "type": "text", + "content": "[24] Xia Kong, Zuodong Gao, Xiaofan Li, Ming Hong, Jun Liu, Chengjie Wang, Yuan Xie, and Yanyun Qu. En-compactness: Self-distillation embedding & contrastive generation for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9306–9315, 2022. 1, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 636, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 547, + 680 + ], + "type": "text", + "content": "[25] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "text", + "content": "[26] Man Liu, Feng Li, Chunjie Zhang, Yunchao Wei, Huihui Bai, and Yao Zhao. Progressive semantic-visual mutual adaptation for generalized zero-shot learning. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23635" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15337-15346, 2023. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[27] Yang Liu, Lei Zhou, Xiao Bai, Yifei Huang, Lin Gu, Jun Zhou, and Tatsuya Harada. Goal-oriented gaze estimation for zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3794-3803, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[28] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12664-12673, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[29] Muhammad Ferjad Naeem, Yongqin Xian, Luc V Gool, and Federico Tombari. I2dformer: Learning image to document attention for zero-shot image classification. Advances in Neural Information Processing Systems, 35:12283-12294, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 342 + ], + "type": "text", + "content": "[30] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. I2mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "type": "text", + "content": "[31] Sanath Narayan, Akshita Gupta, Fahad Shahbaz Khan, Cees GM Snoek, and Ling Shao. Latent embedding feedback and discriminative features for zero-shot classification. In European Conference on Computer Vision, pages 479-495. Springer, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[32] Sanath Narayan, Akshita Gupta, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Mubarak Shah. Discriminative region-based multi-label zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8731-8740, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "text", + "content": "[33] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. Advances in neural information processing systems, 22, 2009. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[34] Genevieve Patterson and James Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2751-2758. IEEE, 2012. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[36] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: lightweight image captioning prompted with retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2840-2849, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[37] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 49-58, 2016. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[38] Ivan Skorokhodov and Mohamed Elhoseiny. Class normalization for (continual)? generalized zero-shot learning. arXiv preprint arXiv:2006.11328, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[39] Jie Song, Chengchao Shen, Yezhou Yang, Yang Liu, and Mingli Song. Transductive unbiased embedding for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1024-1033, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 164, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 198 + ], + "type": "text", + "content": "[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 201, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 244 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 246, + 545, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 291 + ], + "type": "text", + "content": "[42] Chaoqun Wang, Shaobo Min, Xuejin Chen, Xiaoyan Sun, and Houqiang Li. Dual progressive prototype network for generalized zero-shot learning. Advances in Neural Information Processing Systems, 34:2936-2948, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 293, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 346 + ], + "type": "text", + "content": "[43] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3032-3042, 2023. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 349, + 545, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 383 + ], + "type": "text", + "content": "[44] P. Welinder, S. Branson, T. Mita, C. Wah, Florian Schroff, Serge J. Belongie, and P. Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-2010-001, Caltech., 2010. 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 384, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 439 + ], + "type": "text", + "content": "[45] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 441, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 495 + ], + "type": "text", + "content": "[46] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5542-5551, 2018. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 553 + ], + "type": "text", + "content": "[47] Yongqin Xian, Saurabh Sharma, Bernt Schiele, and Zeynep Akata. f-vaegan-d2: A feature generating framework for any-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10275-10284, 2019. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 555, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 610 + ], + "type": "text", + "content": "[48] Guo-Sen Xie, Li Liu, Xiaobo Jin, Fan Zhu, Zheng Zhang, Jie Qin, Yazhou Yao, and Ling Shao. Attentive region embedding network for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9384-9393, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[49] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[50] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Vgse: Visually-grounded semantic embeddings for zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9316-9325, 2022. 7, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23636" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 283 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[52] Meng Ye and Yuhong Guo. Progressive ensemble networks for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11728-11736, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[53] Sheng Zhang, Muzammal Naseer, Guangyi Chen, Zhiqiang Shen, Salman Khan, Kun Zhang, and Fahad Shahbaz Khan. S3a: Towards realistic zero-shot classification via self structural semantic alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7278-7286, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "text", + "content": "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23637" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_content_list.json b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f069ae0173f544ff8165607f7fa31f3e5e429d4d --- /dev/null +++ b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_content_list.json @@ -0,0 +1,1521 @@ +[ + { + "type": "text", + "text": "$V_{k}D:$ Improving Knowledge Distillation using Orthogonal Projections", + "text_level": 1, + "bbox": [ + 130, + 128, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Roy Miles Ismail Elezi Jiankang Deng Huawei Noah's Ark Lab", + "bbox": [ + 313, + 180, + 651, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 253, + 313, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to a $4.4\\%$ relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. Code and models are publicly available1.", + "bbox": [ + 75, + 287, + 473, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 560, + 209, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has achieved remarkable success across a wide variety of tasks in computer vision [38], audio [58], and language [17] domains. However, its adoption is often coupled with increasing computational costs which has limited its application on resource constrained devices such as mobile phones. Fortunately, there have been many techniques proposed to train fast and efficient networks, such as weight pruning [24, 39, 41], quantisation [6, 88], and knowledge distillation [9, 34, 47, 69].", + "bbox": [ + 75, + 585, + 468, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Knowledge distillation (KD) in particular has shown great success [4, 20, 45, 50, 69]. Its main idea is to utilise the pre-trained knowledge of a much larger (teacher) model to supervise the training of a much smaller (student) model. Traditional KD [3, 34] methods have focused on image classification by using the softmax predictions of the teacher as ground-truth labels for the student. However, doing so has made them limited and only applicable to specific modalities, or tasks. Although feature distillation [8, 69] can relax this constraint on the downstream task or modality, its", + "bbox": [ + 75, + 723, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d6689c5adc3d3616bd25be69be41774b049c4c2a7275a2cc911e976320b1ad9.jpg", + "image_caption": [ + "Figure 1. Comparison to both DeiT and DeiT [71] on ImageNet-1K, where DeiT is a distilled DeiT model using a distillation token. Our proposed distillation method achieves significant improvements over DeiT-Ti, while effectively bridging the gap between the teacher and student performance for DeiT-S." + ], + "image_footnote": [], + "bbox": [ + 514, + 268, + 854, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adoption can incur significant computational costs due to the construction of expensive relational objects [49, 50, 53] and memory banks [69]. Most feature distillation pipelines can be described as using some projection [64], alignment [12], or fusion module [9]. These components, though widely used, tend to rely on heuristic design choices. Such heuristic-driven approaches often fall short in delivering new insights into the underlying mechanics of distillation and struggle to adapt to diverse tasks without the introduction of several auxiliary losses. [9, 16, 44, 85]. Furthermore, these additional losses introduce additional hyperparameters which will require tuning for specific tasks or settings.", + "bbox": [ + 496, + 566, + 892, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a novel projection layer that is derived from a principal concept. Our approach focuses on one key idea: Preserving the intra-batch feature similarity. We highlight that if the similarity between features is preserved, then the projection layer will not change or alter the underlying student representation. This constraint is important since it will maximise the amount of knowledge being distilled to the student backbone. For example, if the projector is too expressive it may shortcut the distillation objective by learning some complex non-linear mapping between the", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/roymiles/vkd", + "bbox": [ + 94, + 886, + 349, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15720", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two spaces. This result would significantly diminish the efficacy of distillation and is especially detrimental since the projector is thrown away after training.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By enforcing the preservation of the feature similarity, we derive a reparameterisation of the projection layer itself using the set of orthogonal matrices. We propose to efficiently implement this reparameterisation by projecting the weights onto $SO(n)$ and then truncating the excess rows. In doing so, we enforce the property of row-wise orthogonality, while avoiding the need to compute any expensive matrix inversions or factorisations. We show that this constraint not only improves the student performance but also improves the training convergence and the efficacy of distilling inductive biases.", + "bbox": [ + 75, + 136, + 468, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, a common component in the application of knowledge distillation for generative tasks is the use of additional auxiliary losses. These losses can encourage the generation of diverse features, which will subsequently lead to the generation of more diverse images. However, these auxiliary losses often conflict with the distillation objective and will subsequently degrade the student performance. To address this limitation, we propose a unified framework for incorporating these auxiliary objectives into the distillation loss itself using a task-specific normalisation step. Thus, we show that simply whitening the teacher features can implicitly encourage feature diversity, while removing the need for fine-tuning the hyperparameters of many additional losses. We further demonstrate the importance of this whitening step for data-limited image generation, whereby we achieve a consistent and substantial performance improvements over state-of-the-art. In summary, our contributions are outlined as follows:", + "bbox": [ + 75, + 301, + 470, + 571 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel orthogonal projection layer to maximise the knowledge being distilled through to the student backbone.", + "- We complement our projection with a task-wise normalisation that enables knowledge distillation in generative tasks.", + "- We apply our method to a wide range of tasks and modalities, improving over the state-of-the-art by up to $4.4\\%$ on ImageNet-1K (see Fig. 1)." + ], + "bbox": [ + 76, + 575, + 468, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 724, + 215, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation utilises the knowledge of a pretrained model as supervision for a much smaller model, which can then enable the application and deployment within resource constrained environments. The field can be broadly divided into two main areas: logits distillation [3, 14, 34, 37, 52] and feature distillation [9, 31, 49, 64, 69, 79, 80]. Logit distillation focuses on classification based tasks and introduces an additional objective to minimise the distance between the student and teacher predictions. This was originally proposed using the KL divergence [34], how-", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever it has since been extended using spherical normalisation [23], label decoupling [86], and probability reweighting [52]. In our work, we focus on feature distillation due to its generality to other tasks [9, 50] and modalities [65, 81]. Unfortunately, there is no underlying metric for the intermediate representation spaces, which has led to many heuristically derived solutions. For example, the hand-crafted FSP matrices were proposed [83] to capture the relation between features before and after a set of residual layers. Similarly, many other works have proposed to transfer knowledge using the construction of various Gram [28, 43, 74] or correlation-based matrices [40, 49, 57]. The activation boundaries have also been shown to be an effective supervisory signal for distillation [32], along with the gradients to capture the loss landscape [68, 89]. Another line of work can be loosely grouped together by their inspirations from the self-supervision [69, 77] or information theory literature [49, 50]. In contrast to these methods, we take a step back from the conventional curation of hand-crafted relational objects or objectives. We instead derive and demonstrate that an orthogonal projection is much more effective and can be used in conjunction with just a simple $L2$ loss.", + "bbox": [ + 496, + 90, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning describes the family of pretext tasks used to learn good representations of data in the absence of any ground truth labels. This is an important topic with the overwhelming abundance of unlabelled data available and the increasing costs for human annotations. Self-supervised learning shares some significant similarities with the knowledge distillation literature. For example, contrastive learning [10, 29, 35] has already inspired many distillation methods [8, 69, 78] and asymmetric architectures [11, 21, 22, 62] are often described as a form of self-distillation [2]. However, the most salient overlap with our work instead lies with the use of a predictor. The predictor is a learnable model that maps from the online network to the momentum network. Its usage is very similar to what is described as a projector [13] in the knowledge distillation literature. DirectPred [70] explored the training dynamics of this predictor, which allowed the derivation of a closed form solution for its weights. This work was later simplified by either removing the expensive eigen-decomposition [76] or using fast matrix iterations [63]. In contrast to these works, we explore the role of the projector for knowledge distillation. By building upon a simple set of principles for KD, we are able to derive a cheap reparameterisation of the projector weights that can maximise the knowledge transfer.", + "bbox": [ + 496, + 428, + 890, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Layer reparameterisation has been widely adopted as a technique for constraining weights to introduce favourable properties. For instance, unitary matrices have been shown to address the gradient issues in RNNs [1], positive definite matrices enhance the robustness of batch normalization layers [5], and orthogonal matrices offer spectral regularization for improved generalization [1, 30]. More recent works", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15721", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ee7a3b03f147c2f59fb5ec4a78e990bb1157b1a44a421770d2bdda347cad8754.jpg", + "image_caption": [ + "Figure 2. Illustration of our proposed feature distillation using an orthonormal projection and task-specific feature normalisation. The orthonormal projection (a) maximises the knowledge being distilled to the student backbone, while the task-specific normalisation (b) can introduce domain-specific priors to improve model performance. $\\diamond$ denotes trainable weights, while $\\clubsuit$ denotes weights which are frozen." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 500, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a70a9d133dde5a25ba3f5ed76a4b865eaf18c44cd171aeaf82517ec99dc0500.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 89, + 849, + 203 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/25d3359b53a84e5130ddeb8f9251bdc0581084e96943a35ba30eab2650efeab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 205, + 848, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have shown that low-rank matrices are effective in reducing the cost of fine-tuning large language models [36], while orthogonal matrices enable the cheap controllable fine-tuning of text-image diffusion models [59]. In our work we take these ideas into the context of knowledge distillation with an orthogonal projection. We show that this orthogonal constraint improves both the efficacy of distillation and improves the overall model convergence.", + "bbox": [ + 75, + 357, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Orthogonal Projections", + "text_level": 1, + "bbox": [ + 76, + 489, + 299, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the generality of feature distillation, its use is often coupled with many design decisions and heuristics. These decisions arise from the construction of multiple losses between intermediate feature maps which incur significant and unnecessary training overheads. To address these constraints, we adopt a simple feature distillation pipeline (see Fig. 2) using only the features directly before the classifier, or in the case of generative tasks, the latent representation.", + "bbox": [ + 75, + 515, + 468, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In section 3.1 we motivate the necessary conditions to maximise the efficacy of distillation through the projection. This leads to a reparameterisation of the projection as an orthogonal matrix, which is then efficiently implemented in section 3.2. In section 3.3 we provide some interesting additional insights into the properties of these orthogonal projections, while in section 3.4 we extend our distillation pipeline to improve the performance on both generative and discriminative tasks by using an additional task-specific normalisation step.", + "bbox": [ + 75, + 637, + 470, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Why use orthogonal projections?", + "text_level": 1, + "bbox": [ + 76, + 800, + 366, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main objective is to mitigate the possibility of the projection layer learning any new representation of the data that is not shared by the feature extractor. This is important because the projection layer is thrown away after training and we want to match the feature extractor with the", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "teacher, rather than solely matching the projected features. To achieve this, we propose to preserve the structural information through the projection. We describe this structural information using a kernel matrix $\\mathbf{K} \\in \\mathbb{R}^{b \\times b}$ , where $b$ is the batch-size. This kernel matrix captures the pairwise similarity between all features within a batch:", + "bbox": [ + 496, + 357, + 890, + 448 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {i j} = k \\left(\\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s}\\right) = \\left\\langle \\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s} \\right\\rangle_ {\\mathcal {H}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 465, + 890, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ is some Hilbert space implicitly defined by the positive-definite real-valued kernel $k$ and $\\mathbf{Z}^s\\in \\mathbb{R}^{b\\times d_s}$ are the student features of dimension $d_{s}$ . We aim to preserve $\\mathbf{K}$ under the application of a linear transformation of its arguments. This is equivalent to preserving the structural information of the features. We can express many practical kernels, such as the radial basis function kernel or the polynomial kernel, using a Taylor series expansion [15]:", + "bbox": [ + 496, + 492, + 892, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk \\left(\\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s}\\right) = \\sum_ {n = 0} ^ {\\infty} a _ {n} \\left\\langle \\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s} \\right\\rangle^ {n}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 628, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $a_{n}$ are the coefficients. This expression shows that we simply need a transformation $\\mathbf{P}$ that preserves inner products. Using the canonical inner product in $\\mathbb{R}^{d_s}$ , we derive this constraint on $\\mathbf{P}$ as follows:", + "bbox": [ + 496, + 672, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Z} _ {i} ^ {s} \\left(\\mathbf {Z} _ {j} ^ {s}\\right) ^ {T} = \\mathbf {Z} _ {i} ^ {s} \\mathbf {P} \\left(\\mathbf {Z} _ {j} ^ {s} \\mathbf {P}\\right) ^ {T} (3) \\\\ = \\mathbf {Z} _ {i} ^ {s} \\mathbf {P P} ^ {T} \\left(\\mathbf {Z} _ {j} ^ {s}\\right) ^ {T}, (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 755, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which holds if $\\boxed{\\mathbf{P}^T = \\mathbf{P}^{-1}}$ . This constraint conveniently defines the special orthogonal group $SO(d_s)$ with $d_{s} = d_{t}$ . This group parameterises the set of all rotations in $\\mathbb{R}^{d_s}$ , thus very naturally and intuitively preserves the idea of structural information. However, in the general case where $d_{s}\\neq d_{t}$ , our projection matrix $\\mathbf{P}$ is no longer square. This means", + "bbox": [ + 496, + 804, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15722", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that there exists no canonical inverse for our derived constraint to hold. Choosing the right-inverse defines the set of matrices with orthonormal rows, whereas choosing the left-inverse defines the set of matrices with orthonormal columns. Motivated by the need for an efficient reparameterisation, we focus our attention on the right-inverse, which defines the set of matrices with orthonormal rows, its transpose of which is conveniently represented as a Stiefel matrix manifold [18], denoted $V_{d_t}(\\mathbb{R}^{d_s})$ . To simplify further notation, we omit this distinction between the two.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the Stiefel manifold is smooth, it facilitates the use of standard gradient descent techniques using reparameterisations2. In the next section we provide an efficient implementation of this reparameterisation.", + "bbox": [ + 75, + 243, + 470, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Orthogonal reparameterisation", + "text_level": 1, + "bbox": [ + 76, + 319, + 354, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are a few convenient ways to ensure orthogonality of the projection matrix $\\mathbf{P}$ . One of these ways is to use a Cayley transformation [25] that constructs an orthogonal matrix $\\mathbf{P}$ from a skew-symmetric matrix: $\\mathbf{P} = (\\mathbf{I} - \\mathbf{W})(\\mathbf{I} + \\mathbf{W})^{-1}$ where $\\mathbf{W} = -\\mathbf{W}^T$ . Unfortunately, this closed-form parameterisation, despite its simplicity, requires the expensive computation of a large matrix inverse. Using a QR decomposition can also be considered, but will require the use of expensive iterative algorithms, such as the Gram Schmidt process. Since an orthogonal reparameterisations map is needed for each iteration during training, it is critical for its evaluation to be computationally cheap. To address this computational constraint we propose an efficient algorithm that avoids the need for any expensive matrix inversions or factorisations. We propose to instead perform a cheap parameterisation map onto $SO(d_t)$ using the matrix exponential $\\exp(\\mathbf{A})$ , which can be efficiently implemented using the padé approximation. Knowing that $\\mathbf{W}$ is skew-symmetric, we show that $\\exp(\\mathbf{W})$ is an orthogonal matrix using a few properties of the exponential: $\\exp(\\mathbf{W}) \\cdot \\exp(\\mathbf{W})^T = \\exp(\\mathbf{W} + \\mathbf{W}^T) = \\exp(-\\mathbf{W}^T + \\mathbf{W}^T) = \\exp(\\mathbf{0}) = \\mathbf{I}$ . We then project back to $V_{d_t}(\\mathbb{R}^{d_s})$ by dropping the last $d_t - d_s$ rows. These two sequential steps are given more compactly as follows::", + "bbox": [ + 75, + 344, + 472, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi : \\mathbf {W} \\xrightarrow {\\exp (\\mathbf {W})} \\mathbf {A} \\in S O (d _ {t}) \\xrightarrow {\\mathbf {A} _ {: d _ {s}}} \\mathbf {P} \\in V _ {d _ {t}} \\left(\\mathbb {R} ^ {d _ {s}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 728, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the subscript notation follows from the slicing convention in Pytorch [56] and Numpy [27]. This detour only requires the computation of one exponential, which can be cheaply evaluated using the Padé approximation [33]. We show an illustration of this re-parameterisation in Fig. 2a with the map $\\phi$ described in equation 5.", + "bbox": [ + 75, + 767, + 468, + 858 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/04f390bcea065e25e79cba5ea02bdd68dd520328134f8b191d0936f28f5aacf8.jpg", + "image_caption": [ + "(a) original" + ], + "image_footnote": [], + "bbox": [ + 516, + 93, + 629, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4ab843a6e1f09303384d5ea4010e8b4960df0e1aa9d0efa1c7290a1f3dae78f7.jpg", + "image_caption": [ + "(b) orthogonal" + ], + "image_footnote": [], + "bbox": [ + 640, + 95, + 754, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d781ed62e16d47241f67a1008b3cd63e579bdc29eda96730c3ad6b349271129c.jpg", + "image_caption": [ + "(c) linear", + "Figure 3. t-SNE visualisation [75] of features undergoing either a linear or orthogonal transformation. The orthogonal transformation preserves all of the structural feature information, whereas the linear projection can distort a lot of structure, which can diminish the efficacy of distillation." + ], + "image_footnote": [], + "bbox": [ + 761, + 94, + 872, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Orthogonal projections minimise redundancy", + "text_level": 1, + "bbox": [ + 498, + 301, + 883, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The orthogonal transformations ensure that the projected features $\\mathbf{Z}^s\\mathbf{P}$ will not only be a linear combination of the original features, but also be a transformation that preserves the notion of distance between features. This result can be explained more concretely by observing that the singular values of $\\mathbf{P}$ are all 1. Geometrically, this means that the projection is not squashing or distorting along any of the dimensions, which would bias the loss toward reconstructing the teachers features using only a subset of features. We illustrate this phenomenon in Fig. 3 where we find that even a simple linear projection distorts the features making them overlap with each other. This is caused because it attempts to align the space with the teacher. This level of distortion can degrade the linear separability of the features for the classifier, which impacts the model performance. In contrast, the orthogonal projection preserves the underlying feature manifold.", + "bbox": [ + 496, + 324, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Introducing domain-specific priors", + "text_level": 1, + "bbox": [ + 498, + 590, + 803, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For many tasks it is important to invoke domain specific priors or auxiliary losses to improve model performance [16, 44]. Unfortunately, many of these auxiliary losses conflict with the distillation objective and hinder its efficacy. Instead, we propose a general framework for normalisation that naturally and implicitly incorporates these priors into the distillation objective itself. We show that standardisation is very effective for discriminative tasks by improving the model convergence. This convergence property can be attributed to the improved robustness of the distillation loss to random input perturbations. Similarly, we also show that whitening is a critical step for generative tasks by providing an implicit and soft encouragement of diverse features, which has been proven effective for generating diverse images [19]. Whitening can be much more effective and significantly cheaper than introducing additional auxiliary losses, which has been previously proposed in the literature [16]. We now provide a more detailed illustration of these ideas.", + "bbox": [ + 496, + 613, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2Reparameterisations can be seen as surjective functions that map from Euclidean space back onto the manifold.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15723", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea776fef6ad95ad6697ee688a50e7ec8d31733d1e56c7aa8033dee3b52c6d054.jpg", + "image_caption": [ + "77.7% top-1 Acc", + "Figure 4. Visualisation of the $V_{k}$ D-Ti $\\mathcal{L}_{distill}$ loss landscape with perturbations of the input image across two random dimensions. Normalisation significantly reduces the sensitivity of the loss to random perturbations, which leads to improved robustness and convergence for training." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 267, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/164209a46e6936a42ec8de4efee4794e934f85cb2c2d161806e40bc55ff2cac3.jpg", + "image_caption": [ + "(a) without normalisation", + "78.3% top-1 Acc", + "(b) with normalisation" + ], + "image_footnote": [], + "bbox": [ + 281, + 101, + 442, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Standardisation improves model convergence. In our application of knowledge distillation to discriminative tasks, we observe that a straightforward normalisation of the teacher's representation yields a notable improvement in the robustness of the distillation loss to spurious deformations of the input image (see Fig. 4). These spurious deformations arise from the increasingly expressive families of data augmentation strategies being commonly employed for knowledge distillation [71]. We find that minimizing this loss variance can significantly improve the overall model convergence and performance.", + "bbox": [ + 75, + 301, + 470, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whitening improves feature diversity. By whitening the teacher features, we derive a lower bound that resembles a feature diversity loss [16]. We start with an $L2$ loss between $\\mathbf{Z}_s\\mathbf{P} \\in \\mathbb{R}^{b \\times d}$ and $\\mathbf{Z}_t \\in \\mathbb{R}^{b \\times d}$ . Since $\\mathbf{P}$ is an orthogonal projection that preserves inner products (see section 3.1), we omit its usage to simplify analysis:", + "bbox": [ + 76, + 474, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d i s t i l l} = \\left\\| \\mathbf {Z} ^ {s} - \\mathbf {Z} ^ {t} \\right\\| ^ {2} \\\\ = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} - \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\\\ = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + \\left\\| \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\tag {6} \\\\ - 2 \\left\\langle \\mathbf {Z} _ {:; j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t}, \\mathbf {Z} _ {:; j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\rangle \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 585, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\mathbf{Z}^t$ is whitened such that $(\\mathbf{Z}^t)^T (\\mathbf{Z}^t) = \\mathbf{I}$ , i.e., perfect decorrelation of features, we can significantly simplify the expression above:", + "bbox": [ + 75, + 694, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s t i l l}} = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + 2 - 2 \\langle \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t}, \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\rangle \\\\ \\geq \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + 2 - 2 \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\left\\| \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\\\ = \\boxed {\\operatorname {c o n s t} - \\lambda \\sum_ {i \\neq j} \\mathbf {C} _ {j, i} ^ {2}}, \\quad \\text {w h e r e} \\quad \\mathbf {C} _ {i, j} = \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 758, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where const and $\\lambda \\geq 3$ are both constants that do not depend on the model parameters. Here $\\mathbf{C}$ is the euclidean", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cross-correlation matrix that captures the distance between all the pairs of student and teacher features. This derivation simply shows that minimising an $L2$ loss subject to an explicit whitening constraint on the teacher features provides a cross-feature objective. This cross-feature objective maximises the off-diagonal entries in the cross-correlation matrix, thus encouraging all the features to be decorrelated with respect to the teacher. We describe this process of encouraging decorrelation of features as increasing the feature diversity. We validate this connection of whitening with feature diversity in section 4.3, where we employ knowledge distillation in the context of data-efficient image generation.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 284, + 630, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the generality of our simple knowledge distillation pipeline across three distinct vision tasks: Image classification, object detection, and image generation. In each of these tasks, we consider the harder distillation settings, such as distilling cross-architecture, or in the data-efficient regimes. Throughout we use $V_{k}D$ to denote our KD method using an orthogonal projection, i.e., a matrix projection from the Stiefel manifold $\\mathbf{V}_{\\mathbf{k}}(\\mathbb{R}^{\\mathbf{d}})$ .", + "bbox": [ + 496, + 310, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4181ad8a6a05b3e680ea30f079ebda8143f1acab4efb33be218e76c6844a8540.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Networkacc@1Teacher#params
RegNetY-160 [60] CVPR2082.6none84M
CaiT-S24 [72] ICCV2183.4none47M
DeiT-B [73] ECCV2283.8none87M
DeiT-Ti [71] ICML2172.2none5M
CivT-Ti [61] CVPR2274.9regnety-600m + rednet-266M
MaskedKD [66] arXiv2375.4cait-s246M
Manifold [26] NeurIPS2276.5cait-s246M
DeiT-Ti [71] ICML2174.5regnety-1606M
↓, 1000 epochs76.6regnety-1606M
DearKD [12] CVPR2274.8regnety-1606M
↓, 1000 epochs77.0regnety-1606M
USKD [82] ICCV2375.0regnety-1606M
SRD [48] AAAI2477.2regnety-1606M
VkD-Ti78.3regnety-1606M
DeiT-S [71] ICML2179.8none22M
CivT-S [61] CVPR2282.0regnety-4gf + rednet-5022M
MaskedKD [66] arXiv2381.4deit3-b22M
DeiT-S [71] ICML2181.2regnety-16022M
↓, 1000 epochs82.6regnety-16022M
DearKD [12] CVPR2281.5regnety-16022M
↓, 1000 epochs82.8regnety-16022M
USKD [82] ICCV2380.8regnety-16022M
SRD [48] AAAI2482.1regnety-16022M
VkD-S82.3regnety-16022M
", + "bbox": [ + 501, + 459, + 885, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Data-efficient training of transformers using knowledge distillation on the ImageNet-1K dataset. Unless specified, each model is only trained for 300 epochs.", + "bbox": [ + 498, + 763, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We train all models in Pytorch [55] using 2 NVIDIA V100 GPUs. For the ImageNet experiments, we follow DeiT [71] using the same training schedule and optimization parameters. We also adopt Mixup [84] augmentation, but replace rand-augment", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15724", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/38c134a2555ac1cf20f952e5118ad8d43cb8ccf9a40d9cb43d8c39bdb250fc81.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Param.FPS
ViDTSwin-nano5040.459.643.323.242.555.816M20.0 (45.8)
\\( \\downarrow w/ V_kD \\)5043.062.346.224.845.360.1
Swin-tiny5044.864.548.725.947.662.138M17.2 (26.5)
\\( \\downarrow w/ V_kD \\)5046.966.650.927.849.864.6
Swin-small5047.567.751.429.250.764.861M12.1 (16.5)
\\( \\downarrow w/ V_kD \\)5048.568.452.430.852.266.0
", + "bbox": [ + 176, + 88, + 790, + 188 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2575f4949b90a5654a78e35593a9dc546481cddaa5f97dc0345395835106bf8c.jpg", + "table_caption": [ + "Table 2. Comparison with other detectors on COCO2017 val set. FPS is measured with batch size 1 of $800 \\times 1333$ resolution on a single Tesla V100 GPU, where the value inside the parentheses is measured with batch size 4 of the same resolution to maximize GPU utilisation. All of the student models are distilled from a pre-trained ViDT-base." + ], + "table_footnote": [], + "table_body": "
StudentViDT (Swin-nano)ViDT (Swin-tiny)
TeacherViDT (small)ViDT (base)ViDT (small)ViDT (base)
No Distillation [67] ICLR2240.444.8
Token Matching [67] ICLR2241.541.945.846.5
VkD42.243.045.946.9
", + "bbox": [ + 78, + 250, + 460, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of ViDT on COCO2017 val set. We report AP for the student models distilled from different teacher models, with random gray scaling, gaussian blurring, and solarization. We use AdamW [46] optimizer with learning rate set to 0.001 and weight decay to 0.05. For the object detection experiments, we follow the same training methodology as ViDT [67] except that we replace the original token matching loss with our $V_{k}D$ . Finally, for the image generation task, we use the same training methodology as KD-DLGAN [16] except that we remove the auxiliary diversity losses and instead replace it with either teacher standardisation or whitening. We also remove any distillation from the text encoders, thus further reducing the cost of our method in comparison.", + "bbox": [ + 75, + 352, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Data efficient training of transformers", + "text_level": 1, + "bbox": [ + 76, + 575, + 406, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment with vision transformers, due to their proven success across a variety of fields. However, despite this success, they demand excessive training data and long training schedules. This limitation has motivated the use of knowledge distillation for improving the data efficiency of transformer models [71]. We compare our method with several others using the common knowledge distillation setting proposed alongside DeiT [71] that uses a CNN teacher pre-trained on ImageNet-21K. We train each student model for 300 epochs on ImageNet-1K for the image classification task. Unlike other methods that propose to leverage the efficacy of distilling through distillation tokens alone, we propose to distill directly through the patch tokens. We present the results in Tab. 1 where we massively outperform the previous state-of-the-art. In the tiny architecture, we outperform the baseline by 6.1 percentage points $(pp)$ , and the previous best method that uses the same teacher, the USKD by $3.3pp$ , or a relative improvement of $4.4\\%$ . Interestingly, we perform better than DearKD trained for 1000 epochs by $1.3pp$ , showing that for distillation, it is not nec", + "bbox": [ + 75, + 598, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "essary to train that long. We reach similar results in the small architecture too, where we outperform all the other methods that use similar training resources, and reach competitive results with the methods that use more than 3 times as long training time. In fact, unlike other methods, our approach bridges the gap between the teacher model that reaches $82.6\\%$ accuracy without needing to introduce any excessively long training schedules.", + "bbox": [ + 496, + 253, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Object detection", + "text_level": 1, + "bbox": [ + 500, + 387, + 661, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider the object detection task using the common MS-COCO benchmark [42]. We use the ViDT transformer architecture [67] due to its task performance and its efficiency on consumer hardware. We present the results in Tab. 2, and observe a significant and consistent improvement across a wide range of different ViDT variants. We improve using Swin-nano backbone by $2.6pp$ , in Swin-tiny backbone by $2.1pp$ and in Swin-small backbone by $1pp$ . Furthermore, we also compare against an alternative distillation method, described as token matching [67]. We present these results in Tab. 3. Our method outperforms token matching by up to $1.1pp$ , reaching the best results when we use a larger teacher, demonstrating that our method is not limited in the cases of larger capacity gaps.", + "bbox": [ + 496, + 411, + 890, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Data limited image generation", + "text_level": 1, + "bbox": [ + 498, + 633, + 767, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the generality of our feature distillation framework, we consider an image generation task and compare to the recent KD-DLGAN [16]. KD-DLGAN proposes to use both the text and feature embeddings for the feature guidance followed by an additional diversity loss. Using our novel framework, we show that neither of these explicit additional losses is necessary. Instead, we use a simple whitening of features to encourage the generation of diverse images. We show these results in Tab. 4. A noticeable observation in these results is that whitening can obtain the most significant improvements in the more extreme data-limited regimes. For example, when training on $10\\%$ data with CIFAR-100, whitening the teacher features can improve the FID by up to 9.09. This result highlights that feature diversity is much more critical when there is insufficient training data. To show that our method is", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b90d8b1fa0dec873f4a6987351a80181fabe8be97de2308a808e9988c713d12f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100
10% Data20% Data100% Data10% Data20% Data100% Data
DA + KD (CLIP)22.03 ± 0.0713.70 ± 0.088.70 ± 0.0233.93 ± 0.0921.76 ± 0.0611.74 ± 0.02
DA (Baseline)23.34 ± 0.0914.53 ± 0.108.75 ± 0.0335.39 ± 0.0822.55 ± 0.0611.99 ± 0.02
KD-DLGAN14.20 ± 0.0611.01 ± 0.078.42 ± 0.0118.03 ± 0.1115.60 ± 0.0810.28 ± 0.03
DA + VkD16.42 ± 0.0710.94 ± 0.078.28 ± 0.0124.92 ± 0.1517.97 ± 0.1710.61 ± 0.08
↓ w/ whitening13.16 ± 0.0610.24 ± 0.068.20 ± 0.0316.87 ± 0.0914.00 ± 0.0710.41 ± 0.01
", + "bbox": [ + 129, + 88, + 834, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cafdcc83afb164ad5558d31907f90710bd7e1ba34e613b49cbb18ccfe6ee58b6.jpg", + "table_caption": [ + "Table 4. Comparison with the state-of-the-art over CIFAR-10 and CIFAR 100. Competitive performance is achieved using the orthogonal projection alone, however, introducing a simple whitening step is sufficient in outperforming state-of-the-art by a significant margin. All the compared methods employ BigGAN as the backbone. FID is averaged over three runs." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10\n10% DataCIFAR-100\n10% Data
DA (Baseline) [87] NeurIPS2023.34 ± 0.0935.39 ± 0.08
FitNets [64] ICLR1422.03 ± 0.0733.93 ± 0.09
Label Distillation [34] arXiv1520.46 ± 0.1034.14 ± 0.11
PKD [54] ECCV1921.34 ± 0.0832.15 ± 0.13
SPKD [74] ICCV1919.11 ± 0.0731.97 ± 0.10
KD-DLGAN [16] CVPR2314.20 ± 0.0618.03 ± 0.11
VkD16.47 ± 0.0724.92 ± 0.15
↓ w/ whitening13.16 ± 0.0616.87 ± 0.09
", + "bbox": [ + 80, + 244, + 464, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "much more general than other KD methods in the literature, we also include a comparison for the hardest data-efficient regime in Tab. 5. The results show a consistent improvement in performance, which highlight the importance of engineering both the projector architecture and the normalisation scheme, as opposed to focusing on the distance metrics alone [16, 74].", + "bbox": [ + 75, + 452, + 468, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 566, + 227, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We do a series of ablation studies to highlight the importance of our proposed building blocks. We also provide qualitative results that provide additional insights into explaining the efficacy of our distillation framework.", + "bbox": [ + 75, + 590, + 468, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of orthogonal projections. To demonstrate the effectiveness of constraining the projection weights to be orthogonal, we consider the use of various other projection variants. We analyze the use of a projector ensemble [13], a multi-layer perceptron [51], and a standard linear layer. We present the results in Fig. 5. We observe that both the MLP and projector ensembles show improved performance over a linear layer when under short training schedules. However, when we extend the training schedule, the linear layer becomes much more effective. This is a consequence of the expressive projections beginning to learn new representations that are no longer shared by the student feature extractor. In contrast, our orthogonal projection not only improves the final accuracy but also improves the convergence properties for training, reaching state-of-the-art results in only $\\sim 200$ epochs.", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f13c7773bec2662c1bbb54f9ad3bfa616193a0608e806e42c60d50a3653ae1f.jpg", + "image_caption": [ + "Figure 5. Comparing the performance and convergence of various projector reparameterisations. Although the MLP layer initially trains fast, it begins to saturate as it starts to learn a new representation of the data." + ], + "image_footnote": [], + "bbox": [ + 517, + 271, + 852, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of each block. We now disentangle the contribution of each block of our framework. In table 6 we report the final accuracy with and without normalisation or an orthogonal projection. We observe that, in the classification task, most of the performance improvement is obtained by our orthogonal projection. For example, the orthogonal projection alone boosts the performance from $76.3\\%$ to $77.9\\%$ . This observation is in contrast to the generative tasks, whereby we observe a necessity to use normalisation for strong performance.", + "bbox": [ + 496, + 541, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23b71ff014a7a1c4d5e4da06f26fa7be0abb16bec431d2997c1b6beaa32409cf.jpg", + "table_caption": [ + "Table 5. Comparison to other knowledge distillation methods for image generation. Results were originally reported in [16] and highlight the importance of incorporating domain-specific priors - in this case, encouraging diverse features." + ], + "table_footnote": [], + "table_body": "
Orth.Norm.Acc@1Acc@5
XX76.393.3
X✓76.993.5
✓X77.993.9
✓✓78.394.1
", + "bbox": [ + 583, + 704, + 805, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Highlighting the primary importance of an orthonormal projection. Image classification on ImageNet-1K using a DeiT-student and a RegNety-160 teacher.", + "bbox": [ + 496, + 792, + 890, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Whitening for generative tasks. To empirically confirm the importance of feature whitening for generative tasks, we perform an evaluation with and with it being used. We", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15726", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b83b3b97016be1ab7a44cbbf2e10efe31cfa256a964c500ab1bc1d57872aca43.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 106, + 207, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40a7a843000654db953b2351ccf45635b95159f85088f49477bc10e76d6f67bf.jpg", + "image_caption": [ + "Deit-Ti?" + ], + "image_footnote": [], + "bbox": [ + 210, + 106, + 336, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72679dca277f42e7031d98cb84edb7c9ccdb9f589d34159beabe73a434fd43f3.jpg", + "image_caption": [ + "$V_{k}$ D-Ti" + ], + "image_footnote": [], + "bbox": [ + 338, + 106, + 467, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afed8edae0044c0bebc9a63933f921fa5916f3af5e028f3211f0e82fa8f33d2c.jpg", + "image_caption": [ + "Figure 6. Evaluating the translational equivariance of attention maps. We select the best channel for the first translation and observe its attention maps after translating the input image again." + ], + "image_footnote": [], + "bbox": [ + 80, + 205, + 207, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c866549e74e584b3ff6470023aa22ec3a3670de423880d5db9ff88ed66d9ee37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 205, + 336, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3eb74a15ab7bc75593fd2053202cf2d56c8ea9ff7ee18ba9ac7a86a862593d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 205, + 465, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "present the results in Tab. 4 and 5. We observe that not only is whitening necessary to outperform previous state-of-the-art image generation, but it also leads to a larger increase in performance in the data-limited regime. This result highlights the more prominent importance of diverse features when limited training data is available.", + "bbox": [ + 76, + 364, + 467, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distilling inductive biases. Knowledge distillation has proven effective for improving the data efficiency in training transformer models, especially when the teacher is a CNN. Unfortunately, there has been little qualitative analysis on explaining why this cross-architecture setting helps. We now quantify that this result is a consequence of providing a soft distillation of inductive biases (in this case translational equivariance). In Fig. 6, we explore the impact of applying a translation on the attention maps of a given layer. We observe that any translation of an object is reflected with a translation of the attention maps. Interestingly, this is unlike other methods, such as Deit-Ti, where the attention maps become messy after the translation of the original image. This observation suggests that their improvements may instead be attributed to some other factor, such as an implicit regularisation of the model.", + "bbox": [ + 76, + 470, + 467, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improved localisation of attention maps. We provide further insights into our orthogonal feature distillation method by analyzing the attention maps of various images. These attention maps show how well the model is attending to the salient objects in an image [7]. We compare with various other distillation methods and show the results in Fig. 7. We observe that the attention maps of our method are clustered around the boundaries of the objects, unlike the other two methods where the attention maps are spread over the entire image. In fact, we observe that our distilled model can attend much more to the salient object than the much larger CiT-S model.", + "bbox": [ + 76, + 724, + 467, + 906 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0cd4488dd766c4f571541cee8195077218b2edce01d9e6aaaad2da6f539d2835.jpg", + "image_caption": [ + "Input", + "Deit-Ti", + "CiT-S", + "$V_{k}$ D-Ti", + "Figure 7. Qualitative comparison to other transformer distillation methods. The best channel is selected qualitatively for all examples shown. $V_{k}$ D-Ti is compared against both the same size Deitt-Ti and a much larger $(3.7\\times)$ CiT-S. Best viewed in colour." + ], + "image_footnote": [], + "bbox": [ + 501, + 106, + 890, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture agnostic. Our method is agnostic in the choice of features and can be applied to various classifiers, object detectors, or generative models. For example, in section 4.1, we distill from a CNN to a transformer, in section 4.2, we distill between two transformers, and in section 4.3 we perform distillation in the other direction, from a transformer to a CNN.", + "bbox": [ + 498, + 488, + 890, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 614, + 617, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present a novel projection layer with a principled foundation centered on preserving the intra-batch feature similarity. The core idea of maintaining feature similarity ensures that the projection layer does not distort the underlying student representation, thus maximizing the knowledge transfer to the student backbone. We show that enforcing this constraint is equivalent to parameterising the projection weights to have orthonormal rows or columns. Our simple drop-in replacement for the projection layer leads to improved performance across a wide range of distillation tasks, from image classification to object detection. To further improve the generality of this framework, we show that whitening the teachers' features is sufficient and more effective in extending to generative tasks than other methods. We show in the experiments that our method improves state-of-the-art by up to $4.4\\%$ for image classification and $2.6\\%$ for object detection.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15727", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. ICML, 2016. 2", + "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning. ICLR, 2022. 2", + "[3] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. CVPR, 2022. 1, 2", + "[4] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. ICML Joint Workshop on On-Device Machine Learning and Compact Deep Neural Network Representations (ODML-CDNNR), 2019. 1", + "[5] Daniel Brooks, Olivier Schwander, Frederic Barbaresco, Jean-Yves Schneider, and Matthieu Cord. Riemannian batch normalization for spd neural networks. NeurIPS, 2019. 2", + "[6] Adrian Bulat and Georgios Tzimiropoulos. XNOR-Net++: Improved Binary Neural Networks. BMVC, 2019. 1", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 8", + "[8] Liquin Chen, Dong Wang, Zhe Gan, Jingjing Liu, Ricardo Henao, and Lawrence Carin. Wasserstein Contrastive Representation Distillation. CVPR, 2020. 1, 2", + "[9] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. CVPR, 2021. 1, 2", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020. 2", + "[11] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. CVPR, 2021. 2", + "[12] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. CVPR, 2022. 1, 5", + "[13] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved Feature Distillation via Projector Ensemble. NeurIPS, 2022. 2, 7", + "[14] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. ICCV, 2019. 2", + "[15] Andrew Cotter, Joseph Keshet, and Nathan Srebro. Explicit approximations of the gaussian kernel. arXiv preprint, 2011. 3", + "[16] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu1, and Eric Xing. Kd-dlgan: Data limited image generation via knowledge distillation. CVPR, 2023. 1, 4, 5, 6, 7", + "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2019. 1", + "[18] Alan Edelman, T. A. Arias, and Steven T. Smith. The geometry of algorithms with orthogonality constraints. SIAM Journal of Matrix Analysis and Applications, 1998. 4" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Mohamed Elfeki, Camille Couprie, Morgane Riviere, and Mohamed Elhoseiny. Gdpp: Learning diverse generations using determinantal point process. ICML, 2019. 4", + "[20] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive Model Inversion for Data-Free Knowledge Distillation. *IJCAI*, 2021. 1", + "[21] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. CVPR, 2021. 2", + "[22] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. NeurIPS, 2020. 2", + "[23] Jia Guo, Minghao Chen, Yao Hu, Chen Zhu, Xiaofei He, and Deng Cai. Reducing the Teacher-Student Gap via Spherical Knowledge Distillation. arXiv preprint, 2020. 2", + "[24] Shaopeng Guo, Yujie Wang, Quanquan Li, and Junjie Yan. DMCP: Differentiable Markov Channel Pruning for Neural Networks. CVPR, 2020. 1", + "[25] George Bruce Halsted. The collected mathematical papers of arthur Cayley. The American Mathematical Monthly, 1899. 4", + "[26] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. NeurIPS, 2022. 5", + "[27] Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Río, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. Array programming with NumPy. Nature, 2020. 4", + "[28] Bobby He and Mete Ozay. Feature Kernel Distillation. ICLR, 2022. 2", + "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. CVPR, 2020. 2", + "[30] Kyle Helfrich, Devin Willmott, and Qiang Ye. Orthogonal recurrent neural networks with scaled cayley transform. PMLR, 2018. 2", + "[31] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. ICCV, 2019. 2", + "[32] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. AAAI, 2019. 2", + "[33] Nicholas J. Higham. The scaling and squaring method for the matrix exponential revisited. SIAM Journal on Matrix Analysis and Applications, 2005. 4", + "[34] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2015. 1, 2, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15728", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. *ICLR*, 2019. 2", + "[36] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint, 2021. 3", + "[37] Youmin Kim, Jinbae Park, Younho Jang, Muhammad Ali, and Tae-hyun Oh Sung-ho Bae. Distilling Global and Local Logits with Densely Connected Relations. ICCV, 2021. 2", + "[38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. NeurIPS, 2012. 1", + "[39] Yann Lecun. Optimal Brain Damage. NeurIPS, 1990. 1", + "[40] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In ECCV, 2020. 2", + "[41] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter Pruning using High-Rank Feature Map. CVPR, 2020. 1", + "[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. arXiv preprint, 2015. 6", + "[43] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. CVPR, 2019. 2", + "[44] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured Knowledge Distillation for Semantic Segmentation. CVPR, 2019. 1, 4", + "[45] David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and privileged information. *ICLR*, 2016. 1", + "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6", + "[47] Roy Miles and Krystian Mokolajczyk. Cascaded channel pruning using hierarchical self-distillation. BMVC, 2020. 1", + "[48] Roy Miles and Krystian Mokolajczyk. Understanding the role of the projector in knowledge distillation. AAAI, 2024. 5", + "[49] Roy Miles, Adrian Lopez Rodriguez, and Krystian Miko-lajczyk. Information Theoretic Representation Distillation. BMVC, 2022. 1, 2", + "[50] Roy Miles, Mehmet Kerim Yucel, Bruno Manganelli, and Albert Saa-Garriga. MobileVOS: Real-Time Video Object Segmentation Contrastive Learning meets Knowledge Distillation. CVPR, 2023. 1, 2", + "[51] K L Navaneet, Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. SimReg: Regression as a Simple Yet Effective Tool for Self-supervised Knowledge Distillation. BMVC, 2021. 7", + "[52] Yulei Niu, Long Chen, Chang Zhou, and Hanwang Zhang. Respecting transfer gap in knowledge distillation. NeurIPS, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Wonpyo Park, Kakao Corp, Dongju Kim, and Yan Lu. Relational Knowledge Distillation. CVPR, 2019. 1", + "[54] Nikolaos Passalis and Anastasios Tefas. Learning Deep Representations with Probabilistic Knowledge Transfer. ECCV, 2018. 7", + "[55] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. NeurIPS 2017 Workshop Autodiff homepage, 2017. 5", + "[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4", + "[57] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. CVPR, 2019. 2", + "[58] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schluter, Shuo-Yiin Chang, and Tara Sainath. Deep learning for audio signal processing. IEEE Journal of Selected Topics in Signal Processing, 2019. 1", + "[59] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. arXiv preprint, 2023. 3", + "[60] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing Network Design Spaces. CVPR, 2020. 5", + "[61] Sucheng Ren, Zhengqi Gao, Tianyu Hua, Zihui Xue, Yonglong Tian, Shengfeng He, and Hang Zhao. Co-advise: Cross Inductive Bias Distillation. CVPR, 2022. 5", + "[62] Pierre H. Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, and Michal Valko. Byol works even without batch statistics. arXiv preprint, 2020. 2", + "[63] Pierre H. Richemond, Allison Tam, Yunhao Tang, Florian Strub, Bilal Piot, and Felix Hill. The edge of orthogonality: A simple view of what makes byol tick. arXiv preprint, 2023. 2", + "[64] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints For Thin Deep Nets. ICLR, 2015. 1, 2, 7", + "[65] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. NeurIPS Workshop on Energy Efficient Machine Learning and Cognitive Computing, 2019. 2", + "[66] Seungwoo Son, Namhoon Lee, and Jaeho Lee. Maskedkd: Efficient distillation of vision transformers with masked images. arXiv preprint, 2023. 5", + "[67] Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo, Wonjae Kim, and" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15729", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ming-Hsuan Yang. Vidt: An efficient and effective fully transformer-based object detector. ICLR, 2021. 6", + "[68] Suraj Srinivas and François Fleuret. Knowledge transfer with jacobian matching. ICML, 2018. 2", + "[69] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. *ICLR*, 2019. 1, 2", + "[70] Yuandong Tian, Xinlei Chen, and Surya Ganguli. Understanding self-supervised Learning Dynamics without Contrastive Pairs. ICML, 2021. 2", + "[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. PMLR, 2021. 1, 5, 6", + "[72] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 5", + "[73] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 5", + "[74] Fred Tung and Greg Mori. Similarity-preserving knowledge distillation. ICCV, 2019. 2, 7", + "[75] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 4", + "[76] Xiang Wang, Xinlei Chen, Simon S. Du, and Yuandong Tian. Towards demystifying representation learning with non-contrastive self-supervision. arXiv preprint, 2022. 2", + "[77] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-supervision. ECCV, 2020. 2", + "[78] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. ECCV, 2020. 2", + "[79] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 2", + "[80] Jing Yang, Xiatian Zhu, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Knowledge distillation meets open-set semi-supervised learning. arXiv:2205.06701, 2022. 2", + "[81] Ziqing Yang, Yiming Cui, Zhipeng Chen, Wanxiang Che, Ting Liu, Shijin Wang, and Guoping Hu. TextBrewer: An Open-Source Knowledge Distillation Toolkit for Natural Language Processing. ACL, 2020. 2", + "[82] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. ICCV, 2023. 5", + "[83] Junho Yim. A Gift from Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning. CVPR, 2017. 2", + "[84] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. *ICLR*, 2017. 5", + "[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaisheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. CVPR, 2022. 1", + "[86] Borui Zhao, Renjie Song, and Yiyu Qiu. Decoupled Knowledge Distillation. CVPR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. NeurIPS, 2020. 7", + "[88] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients. arXiv preprint, 2016. 1", + "[89] Jinguo Zhu, Shixiang Tang, Dapeng Chen, and Shijie Yu. Complementary Relation Contrastive Distillation. CVPR, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15730", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_model.json b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_model.json new file mode 100644 index 0000000000000000000000000000000000000000..16191e1d1fe1849da4bdabcef03ee6c59f5c4691 --- /dev/null +++ b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.13, + 0.842, + 0.154 + ], + "angle": 0, + "content": "\\(V_{k}D:\\) Improving Knowledge Distillation using Orthogonal Projections" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.181, + 0.652, + 0.22 + ], + "angle": 0, + "content": "Roy Miles Ismail Elezi Jiankang Deng Huawei Noah's Ark Lab" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.255, + 0.314, + 0.272 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.474, + 0.531 + ], + "angle": 0, + "content": "Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to a \\(4.4\\%\\) relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. Code and models are publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.561, + 0.21, + 0.577 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.723 + ], + "angle": 0, + "content": "Deep learning has achieved remarkable success across a wide variety of tasks in computer vision [38], audio [58], and language [17] domains. However, its adoption is often coupled with increasing computational costs which has limited its application on resource constrained devices such as mobile phones. Fortunately, there have been many techniques proposed to train fast and efficient networks, such as weight pruning [24, 39, 41], quantisation [6, 88], and knowledge distillation [9, 34, 47, 69]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Knowledge distillation (KD) in particular has shown great success [4, 20, 45, 50, 69]. Its main idea is to utilise the pre-trained knowledge of a much larger (teacher) model to supervise the training of a much smaller (student) model. Traditional KD [3, 34] methods have focused on image classification by using the softmax predictions of the teacher as ground-truth labels for the student. However, doing so has made them limited and only applicable to specific modalities, or tasks. Although feature distillation [8, 69] can relax this constraint on the downstream task or modality, its" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.269, + 0.856, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.49, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Figure 1. Comparison to both DeiT and DeiT [71] on ImageNet-1K, where DeiT is a distilled DeiT model using a distillation token. Our proposed distillation method achieves significant improvements over DeiT-Ti, while effectively bridging the gap between the teacher and student performance for DeiT-S." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.749 + ], + "angle": 0, + "content": "adoption can incur significant computational costs due to the construction of expensive relational objects [49, 50, 53] and memory banks [69]. Most feature distillation pipelines can be described as using some projection [64], alignment [12], or fusion module [9]. These components, though widely used, tend to rely on heuristic design choices. Such heuristic-driven approaches often fall short in delivering new insights into the underlying mechanics of distillation and struggle to adapt to diverse tasks without the introduction of several auxiliary losses. [9, 16, 44, 85]. Furthermore, these additional losses introduce additional hyperparameters which will require tuning for specific tasks or settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a novel projection layer that is derived from a principal concept. Our approach focuses on one key idea: Preserving the intra-batch feature similarity. We highlight that if the similarity between features is preserved, then the projection layer will not change or alter the underlying student representation. This constraint is important since it will maximise the amount of knowledge being distilled to the student backbone. For example, if the projector is too expressive it may shortcut the distillation objective by learning some complex non-linear mapping between the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.351, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/roymiles/vkd" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15720" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "two spaces. This result would significantly diminish the efficacy of distillation and is especially detrimental since the projector is thrown away after training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.302 + ], + "angle": 0, + "content": "By enforcing the preservation of the feature similarity, we derive a reparameterisation of the projection layer itself using the set of orthogonal matrices. We propose to efficiently implement this reparameterisation by projecting the weights onto \\( SO(n) \\) and then truncating the excess rows. In doing so, we enforce the property of row-wise orthogonality, while avoiding the need to compute any expensive matrix inversions or factorisations. We show that this constraint not only improves the student performance but also improves the training convergence and the efficacy of distilling inductive biases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.471, + 0.573 + ], + "angle": 0, + "content": "Finally, a common component in the application of knowledge distillation for generative tasks is the use of additional auxiliary losses. These losses can encourage the generation of diverse features, which will subsequently lead to the generation of more diverse images. However, these auxiliary losses often conflict with the distillation objective and will subsequently degrade the student performance. To address this limitation, we propose a unified framework for incorporating these auxiliary objectives into the distillation loss itself using a task-specific normalisation step. Thus, we show that simply whitening the teacher features can implicitly encourage feature diversity, while removing the need for fine-tuning the hyperparameters of many additional losses. We further demonstrate the importance of this whitening step for data-limited image generation, whereby we achieve a consistent and substantial performance improvements over state-of-the-art. In summary, our contributions are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.577, + 0.469, + 0.62 + ], + "angle": 0, + "content": "- We propose a novel orthogonal projection layer to maximise the knowledge being distilled through to the student backbone." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.622, + 0.469, + 0.665 + ], + "angle": 0, + "content": "- We complement our projection with a task-wise normalisation that enables knowledge distillation in generative tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.667, + 0.469, + 0.713 + ], + "angle": 0, + "content": "- We apply our method to a wide range of tasks and modalities, improving over the state-of-the-art by up to \\(4.4\\%\\) on ImageNet-1K (see Fig. 1)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.577, + 0.469, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.216, + 0.74 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Knowledge Distillation utilises the knowledge of a pretrained model as supervision for a much smaller model, which can then enable the application and deployment within resource constrained environments. The field can be broadly divided into two main areas: logits distillation [3, 14, 34, 37, 52] and feature distillation [9, 31, 49, 64, 69, 79, 80]. Logit distillation focuses on classification based tasks and introduces an additional objective to minimise the distance between the student and teacher predictions. This was originally proposed using the KL divergence [34], how-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.424 + ], + "angle": 0, + "content": "ever it has since been extended using spherical normalisation [23], label decoupling [86], and probability reweighting [52]. In our work, we focus on feature distillation due to its generality to other tasks [9, 50] and modalities [65, 81]. Unfortunately, there is no underlying metric for the intermediate representation spaces, which has led to many heuristically derived solutions. For example, the hand-crafted FSP matrices were proposed [83] to capture the relation between features before and after a set of residual layers. Similarly, many other works have proposed to transfer knowledge using the construction of various Gram [28, 43, 74] or correlation-based matrices [40, 49, 57]. The activation boundaries have also been shown to be an effective supervisory signal for distillation [32], along with the gradients to capture the loss landscape [68, 89]. Another line of work can be loosely grouped together by their inspirations from the self-supervision [69, 77] or information theory literature [49, 50]. In contrast to these methods, we take a step back from the conventional curation of hand-crafted relational objects or objectives. We instead derive and demonstrate that an orthogonal projection is much more effective and can be used in conjunction with just a simple \\(L2\\) loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Self-supervised learning describes the family of pretext tasks used to learn good representations of data in the absence of any ground truth labels. This is an important topic with the overwhelming abundance of unlabelled data available and the increasing costs for human annotations. Self-supervised learning shares some significant similarities with the knowledge distillation literature. For example, contrastive learning [10, 29, 35] has already inspired many distillation methods [8, 69, 78] and asymmetric architectures [11, 21, 22, 62] are often described as a form of self-distillation [2]. However, the most salient overlap with our work instead lies with the use of a predictor. The predictor is a learnable model that maps from the online network to the momentum network. Its usage is very similar to what is described as a projector [13] in the knowledge distillation literature. DirectPred [70] explored the training dynamics of this predictor, which allowed the derivation of a closed form solution for its weights. This work was later simplified by either removing the expensive eigen-decomposition [76] or using fast matrix iterations [63]. In contrast to these works, we explore the role of the projector for knowledge distillation. By building upon a simple set of principles for KD, we are able to derive a cheap reparameterisation of the projector weights that can maximise the knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Layer reparameterisation has been widely adopted as a technique for constraining weights to introduce favourable properties. For instance, unitary matrices have been shown to address the gradient issues in RNNs [1], positive definite matrices enhance the robustness of batch normalization layers [5], and orthogonal matrices offer spectral regularization for improved generalization [1, 30]. More recent works" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15721" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.501, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.85, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.207, + 0.849, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Illustration of our proposed feature distillation using an orthonormal projection and task-specific feature normalisation. The orthonormal projection (a) maximises the knowledge being distilled to the student backbone, while the task-specific normalisation (b) can introduce domain-specific priors to improve model performance. \\(\\diamond\\) denotes trainable weights, while \\(\\clubsuit\\) denotes weights which are frozen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.48 + ], + "angle": 0, + "content": "have shown that low-rank matrices are effective in reducing the cost of fine-tuning large language models [36], while orthogonal matrices enable the cheap controllable fine-tuning of text-image diffusion models [59]. In our work we take these ideas into the context of knowledge distillation with an orthogonal projection. We show that this orthogonal constraint improves both the efficacy of distillation and improves the overall model convergence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.491, + 0.3, + 0.509 + ], + "angle": 0, + "content": "3. Orthogonal Projections" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Despite the generality of feature distillation, its use is often coupled with many design decisions and heuristics. These decisions arise from the construction of multiple losses between intermediate feature maps which incur significant and unnecessary training overheads. To address these constraints, we adopt a simple feature distillation pipeline (see Fig. 2) using only the features directly before the classifier, or in the case of generative tasks, the latent representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.471, + 0.789 + ], + "angle": 0, + "content": "In section 3.1 we motivate the necessary conditions to maximise the efficacy of distillation through the projection. This leads to a reparameterisation of the projection as an orthogonal matrix, which is then efficiently implemented in section 3.2. In section 3.3 we provide some interesting additional insights into the properties of these orthogonal projections, while in section 3.4 we extend our distillation pipeline to improve the performance on both generative and discriminative tasks by using an additional task-specific normalisation step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.367, + 0.818 + ], + "angle": 0, + "content": "3.1. Why use orthogonal projections?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our main objective is to mitigate the possibility of the projection layer learning any new representation of the data that is not shared by the feature extractor. This is important because the projection layer is thrown away after training and we want to match the feature extractor with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.449 + ], + "angle": 0, + "content": "teacher, rather than solely matching the projected features. To achieve this, we propose to preserve the structural information through the projection. We describe this structural information using a kernel matrix \\(\\mathbf{K} \\in \\mathbb{R}^{b \\times b}\\), where \\(b\\) is the batch-size. This kernel matrix captures the pairwise similarity between all features within a batch:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.466, + 0.892, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {i j} = k \\left(\\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s}\\right) = \\left\\langle \\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s} \\right\\rangle_ {\\mathcal {H}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.616 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) is some Hilbert space implicitly defined by the positive-definite real-valued kernel \\(k\\) and \\(\\mathbf{Z}^s\\in \\mathbb{R}^{b\\times d_s}\\) are the student features of dimension \\(d_{s}\\). We aim to preserve \\(\\mathbf{K}\\) under the application of a linear transformation of its arguments. This is equivalent to preserving the structural information of the features. We can express many practical kernels, such as the radial basis function kernel or the polynomial kernel, using a Taylor series expansion [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.629, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\nk \\left(\\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s}\\right) = \\sum_ {n = 0} ^ {\\infty} a _ {n} \\left\\langle \\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s} \\right\\rangle^ {n}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.735 + ], + "angle": 0, + "content": "where \\(a_{n}\\) are the coefficients. This expression shows that we simply need a transformation \\(\\mathbf{P}\\) that preserves inner products. Using the canonical inner product in \\(\\mathbb{R}^{d_s}\\), we derive this constraint on \\(\\mathbf{P}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.756, + 0.891, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Z} _ {i} ^ {s} \\left(\\mathbf {Z} _ {j} ^ {s}\\right) ^ {T} = \\mathbf {Z} _ {i} ^ {s} \\mathbf {P} \\left(\\mathbf {Z} _ {j} ^ {s} \\mathbf {P}\\right) ^ {T} (3) \\\\ = \\mathbf {Z} _ {i} ^ {s} \\mathbf {P P} ^ {T} \\left(\\mathbf {Z} _ {j} ^ {s}\\right) ^ {T}, (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.893, + 0.902 + ], + "angle": 0, + "content": "which holds if \\(\\boxed{\\mathbf{P}^T = \\mathbf{P}^{-1}}\\). This constraint conveniently defines the special orthogonal group \\(SO(d_s)\\) with \\(d_{s} = d_{t}\\). This group parameterises the set of all rotations in \\(\\mathbb{R}^{d_s}\\), thus very naturally and intuitively preserves the idea of structural information. However, in the general case where \\(d_{s}\\neq d_{t}\\), our projection matrix \\(\\mathbf{P}\\) is no longer square. This means" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "that there exists no canonical inverse for our derived constraint to hold. Choosing the right-inverse defines the set of matrices with orthonormal rows, whereas choosing the left-inverse defines the set of matrices with orthonormal columns. Motivated by the need for an efficient reparameterisation, we focus our attention on the right-inverse, which defines the set of matrices with orthonormal rows, its transpose of which is conveniently represented as a Stiefel matrix manifold [18], denoted \\( V_{d_t}(\\mathbb{R}^{d_s}) \\). To simplify further notation, we omit this distinction between the two." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.472, + 0.306 + ], + "angle": 0, + "content": "Since the Stiefel manifold is smooth, it facilitates the use of standard gradient descent techniques using reparameterisations2. In the next section we provide an efficient implementation of this reparameterisation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.32, + 0.356, + 0.337 + ], + "angle": 0, + "content": "3.2. Orthogonal reparameterisation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.473, + 0.707 + ], + "angle": 0, + "content": "There are a few convenient ways to ensure orthogonality of the projection matrix \\(\\mathbf{P}\\). One of these ways is to use a Cayley transformation [25] that constructs an orthogonal matrix \\(\\mathbf{P}\\) from a skew-symmetric matrix: \\(\\mathbf{P} = (\\mathbf{I} - \\mathbf{W})(\\mathbf{I} + \\mathbf{W})^{-1}\\) where \\(\\mathbf{W} = -\\mathbf{W}^T\\). Unfortunately, this closed-form parameterisation, despite its simplicity, requires the expensive computation of a large matrix inverse. Using a QR decomposition can also be considered, but will require the use of expensive iterative algorithms, such as the Gram Schmidt process. Since an orthogonal reparameterisations map is needed for each iteration during training, it is critical for its evaluation to be computationally cheap. To address this computational constraint we propose an efficient algorithm that avoids the need for any expensive matrix inversions or factorisations. We propose to instead perform a cheap parameterisation map onto \\(SO(d_t)\\) using the matrix exponential \\(\\exp(\\mathbf{A})\\), which can be efficiently implemented using the padé approximation. Knowing that \\(\\mathbf{W}\\) is skew-symmetric, we show that \\(\\exp(\\mathbf{W})\\) is an orthogonal matrix using a few properties of the exponential: \\(\\exp(\\mathbf{W}) \\cdot \\exp(\\mathbf{W})^T = \\exp(\\mathbf{W} + \\mathbf{W}^T) = \\exp(-\\mathbf{W}^T + \\mathbf{W}^T) = \\exp(\\mathbf{0}) = \\mathbf{I}\\). We then project back to \\(V_{d_t}(\\mathbb{R}^{d_s})\\) by dropping the last \\(d_t - d_s\\) rows. These two sequential steps are given more compactly as follows::" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.729, + 0.47, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\phi : \\mathbf {W} \\xrightarrow {\\exp (\\mathbf {W})} \\mathbf {A} \\in S O (d _ {t}) \\xrightarrow {\\mathbf {A} _ {: d _ {s}}} \\mathbf {P} \\in V _ {d _ {t}} \\left(\\mathbb {R} ^ {d _ {s}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.768, + 0.47, + 0.859 + ], + "angle": 0, + "content": "where the subscript notation follows from the slicing convention in Pytorch [56] and Numpy [27]. This detour only requires the computation of one exponential, which can be cheaply evaluated using the Padé approximation [33]. We show an illustration of this re-parameterisation in Fig. 2a with the map \\(\\phi\\) described in equation 5." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.094, + 0.63, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.204, + 0.601, + 0.217 + ], + "angle": 0, + "content": "(a) original" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.096, + 0.755, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.204, + 0.736, + 0.216 + ], + "angle": 0, + "content": "(b) orthogonal" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.095, + 0.874, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.204, + 0.85, + 0.216 + ], + "angle": 0, + "content": "(c) linear" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Figure 3. t-SNE visualisation [75] of features undergoing either a linear or orthogonal transformation. The orthogonal transformation preserves all of the structural feature information, whereas the linear projection can distort a lot of structure, which can diminish the efficacy of distillation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.302, + 0.885, + 0.319 + ], + "angle": 0, + "content": "3.3. Orthogonal projections minimise redundancy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.581 + ], + "angle": 0, + "content": "The orthogonal transformations ensure that the projected features \\(\\mathbf{Z}^s\\mathbf{P}\\) will not only be a linear combination of the original features, but also be a transformation that preserves the notion of distance between features. This result can be explained more concretely by observing that the singular values of \\(\\mathbf{P}\\) are all 1. Geometrically, this means that the projection is not squashing or distorting along any of the dimensions, which would bias the loss toward reconstructing the teachers features using only a subset of features. We illustrate this phenomenon in Fig. 3 where we find that even a simple linear projection distorts the features making them overlap with each other. This is caused because it attempts to align the space with the teacher. This level of distortion can degrade the linear separability of the features for the classifier, which impacts the model performance. In contrast, the orthogonal projection preserves the underlying feature manifold." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.591, + 0.804, + 0.607 + ], + "angle": 0, + "content": "3.4. Introducing domain-specific priors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.9 + ], + "angle": 0, + "content": "For many tasks it is important to invoke domain specific priors or auxiliary losses to improve model performance [16, 44]. Unfortunately, many of these auxiliary losses conflict with the distillation objective and hinder its efficacy. Instead, we propose a general framework for normalisation that naturally and implicitly incorporates these priors into the distillation objective itself. We show that standardisation is very effective for discriminative tasks by improving the model convergence. This convergence property can be attributed to the improved robustness of the distillation loss to random input perturbations. Similarly, we also show that whitening is a critical step for generative tasks by providing an implicit and soft encouragement of diverse features, which has been proven effective for generating diverse images [19]. Whitening can be much more effective and significantly cheaper than introducing additional auxiliary losses, which has been previously proposed in the literature [16]. We now provide a more detailed illustration of these ideas." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2Reparameterisations can be seen as surjective functions that map from Euclidean space back onto the manifold." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.268, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.182, + 0.237, + 0.195 + ], + "angle": 0, + "content": "77.7% top-1 Acc" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.196, + 0.258, + 0.209 + ], + "angle": 0, + "content": "(a) without normalisation" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.102, + 0.443, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.182, + 0.415, + 0.195 + ], + "angle": 0, + "content": "78.3% top-1 Acc" + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.196, + 0.431, + 0.21 + ], + "angle": 0, + "content": "(b) with normalisation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.471, + 0.29 + ], + "angle": 0, + "content": "Figure 4. Visualisation of the \\( V_{k} \\) D-Ti \\( \\mathcal{L}_{distill} \\) loss landscape with perturbations of the input image across two random dimensions. Normalisation significantly reduces the sensitivity of the loss to random perturbations, which leads to improved robustness and convergence for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.471, + 0.469 + ], + "angle": 0, + "content": "Standardisation improves model convergence. In our application of knowledge distillation to discriminative tasks, we observe that a straightforward normalisation of the teacher's representation yields a notable improvement in the robustness of the distillation loss to spurious deformations of the input image (see Fig. 4). These spurious deformations arise from the increasingly expressive families of data augmentation strategies being commonly employed for knowledge distillation [71]. We find that minimizing this loss variance can significantly improve the overall model convergence and performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Whitening improves feature diversity. By whitening the teacher features, we derive a lower bound that resembles a feature diversity loss [16]. We start with an \\(L2\\) loss between \\(\\mathbf{Z}_s\\mathbf{P} \\in \\mathbb{R}^{b \\times d}\\) and \\(\\mathbf{Z}_t \\in \\mathbb{R}^{b \\times d}\\). Since \\(\\mathbf{P}\\) is an orthogonal projection that preserves inner products (see section 3.1), we omit its usage to simplify analysis:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.586, + 0.469, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d i s t i l l} = \\left\\| \\mathbf {Z} ^ {s} - \\mathbf {Z} ^ {t} \\right\\| ^ {2} \\\\ = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} - \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\\\ = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + \\left\\| \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\tag {6} \\\\ - 2 \\left\\langle \\mathbf {Z} _ {:; j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t}, \\mathbf {Z} _ {:; j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\rangle \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Since \\(\\mathbf{Z}^t\\) is whitened such that \\((\\mathbf{Z}^t)^T (\\mathbf{Z}^t) = \\mathbf{I}\\), i.e., perfect decorrelation of features, we can significantly simplify the expression above:" + }, + { + "type": "equation", + "bbox": [ + 0.083, + 0.76, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s t i l l}} = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + 2 - 2 \\langle \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t}, \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\rangle \\\\ \\geq \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + 2 - 2 \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\left\\| \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\\\ = \\boxed {\\operatorname {c o n s t} - \\lambda \\sum_ {i \\neq j} \\mathbf {C} _ {j, i} ^ {2}}, \\quad \\text {w h e r e} \\quad \\mathbf {C} _ {i, j} = \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where const and \\(\\lambda \\geq 3\\) are both constants that do not depend on the model parameters. Here \\(\\mathbf{C}\\) is the euclidean" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "cross-correlation matrix that captures the distance between all the pairs of student and teacher features. This derivation simply shows that minimising an \\( L2 \\) loss subject to an explicit whitening constraint on the teacher features provides a cross-feature objective. This cross-feature objective maximises the off-diagonal entries in the cross-correlation matrix, thus encouraging all the features to be decorrelated with respect to the teacher. We describe this process of encouraging decorrelation of features as increasing the feature diversity. We validate this connection of whitening with feature diversity in section 4.3, where we employ knowledge distillation in the context of data-efficient image generation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.285, + 0.632, + 0.303 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.433 + ], + "angle": 0, + "content": "In this section, we evaluate the generality of our simple knowledge distillation pipeline across three distinct vision tasks: Image classification, object detection, and image generation. In each of these tasks, we consider the harder distillation settings, such as distilling cross-architecture, or in the data-efficient regimes. Throughout we use \\( V_{k}D \\) to denote our KD method using an orthogonal projection, i.e., a matrix projection from the Stiefel manifold \\( \\mathbf{V}_{\\mathbf{k}}(\\mathbb{R}^{\\mathbf{d}}) \\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.46, + 0.887, + 0.759 + ], + "angle": 0, + "content": "
Networkacc@1Teacher#params
RegNetY-160 [60] CVPR2082.6none84M
CaiT-S24 [72] ICCV2183.4none47M
DeiT-B [73] ECCV2283.8none87M
DeiT-Ti [71] ICML2172.2none5M
CivT-Ti [61] CVPR2274.9regnety-600m + rednet-266M
MaskedKD [66] arXiv2375.4cait-s246M
Manifold [26] NeurIPS2276.5cait-s246M
DeiT-Ti [71] ICML2174.5regnety-1606M
↓, 1000 epochs76.6regnety-1606M
DearKD [12] CVPR2274.8regnety-1606M
↓, 1000 epochs77.0regnety-1606M
USKD [82] ICCV2375.0regnety-1606M
SRD [48] AAAI2477.2regnety-1606M
VkD-Ti78.3regnety-1606M
DeiT-S [71] ICML2179.8none22M
CivT-S [61] CVPR2282.0regnety-4gf + rednet-5022M
MaskedKD [66] arXiv2381.4deit3-b22M
DeiT-S [71] ICML2181.2regnety-16022M
↓, 1000 epochs82.6regnety-16022M
DearKD [12] CVPR2281.5regnety-16022M
↓, 1000 epochs82.8regnety-16022M
USKD [82] ICCV2380.8regnety-16022M
SRD [48] AAAI2482.1regnety-16022M
VkD-S82.3regnety-16022M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Table 1. Data-efficient training of transformers using knowledge distillation on the ImageNet-1K dataset. Unless specified, each model is only trained for 300 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We train all models in Pytorch [55] using 2 NVIDIA V100 GPUs. For the ImageNet experiments, we follow DeiT [71] using the same training schedule and optimization parameters. We also adopt Mixup [84] augmentation, but replace rand-augment" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15724" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.177, + 0.089, + 0.791, + 0.189 + ], + "angle": 0, + "content": "
MethodBackboneEpochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Param.FPS
ViDTSwin-nano5040.459.643.323.242.555.816M20.0 (45.8)
\\( \\downarrow w/ V_kD \\)5043.062.346.224.845.360.1
Swin-tiny5044.864.548.725.947.662.138M17.2 (26.5)
\\( \\downarrow w/ V_kD \\)5046.966.650.927.849.864.6
Swin-small5047.567.751.429.250.764.861M12.1 (16.5)
\\( \\downarrow w/ V_kD \\)5048.568.452.430.852.266.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.2, + 0.894, + 0.243 + ], + "angle": 0, + "content": "Table 2. Comparison with other detectors on COCO2017 val set. FPS is measured with batch size 1 of \\(800 \\times 1333\\) resolution on a single Tesla V100 GPU, where the value inside the parentheses is measured with batch size 4 of the same resolution to maximize GPU utilisation. All of the student models are distilled from a pre-trained ViDT-base." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.251, + 0.462, + 0.343 + ], + "angle": 0, + "content": "
StudentViDT (Swin-nano)ViDT (Swin-tiny)
TeacherViDT (small)ViDT (base)ViDT (small)ViDT (base)
No Distillation [67] ICLR2240.444.8
Token Matching [67] ICLR2241.541.945.846.5
VkD42.243.045.946.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Table 3. Comparison of ViDT on COCO2017 val set. We report AP for the student models distilled from different teacher models, with random gray scaling, gaussian blurring, and solarization. We use AdamW [46] optimizer with learning rate set to 0.001 and weight decay to 0.05. For the object detection experiments, we follow the same training methodology as ViDT [67] except that we replace the original token matching loss with our \\( V_{k}D \\). Finally, for the image generation task, we use the same training methodology as KD-DLGAN [16] except that we remove the auxiliary diversity losses and instead replace it with either teacher standardisation or whitening. We also remove any distillation from the text encoders, thus further reducing the cost of our method in comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.407, + 0.592 + ], + "angle": 0, + "content": "4.1. Data efficient training of transformers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We experiment with vision transformers, due to their proven success across a variety of fields. However, despite this success, they demand excessive training data and long training schedules. This limitation has motivated the use of knowledge distillation for improving the data efficiency of transformer models [71]. We compare our method with several others using the common knowledge distillation setting proposed alongside DeiT [71] that uses a CNN teacher pre-trained on ImageNet-21K. We train each student model for 300 epochs on ImageNet-1K for the image classification task. Unlike other methods that propose to leverage the efficacy of distilling through distillation tokens alone, we propose to distill directly through the patch tokens. We present the results in Tab. 1 where we massively outperform the previous state-of-the-art. In the tiny architecture, we outperform the baseline by 6.1 percentage points \\((pp)\\), and the previous best method that uses the same teacher, the USKD by \\(3.3pp\\), or a relative improvement of \\(4.4\\%\\). Interestingly, we perform better than DearKD trained for 1000 epochs by \\(1.3pp\\), showing that for distillation, it is not nec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.378 + ], + "angle": 0, + "content": "essary to train that long. We reach similar results in the small architecture too, where we outperform all the other methods that use similar training resources, and reach competitive results with the methods that use more than 3 times as long training time. In fact, unlike other methods, our approach bridges the gap between the teacher model that reaches \\(82.6\\%\\) accuracy without needing to introduce any excessively long training schedules." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.663, + 0.404 + ], + "angle": 0, + "content": "4.2. Object detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.624 + ], + "angle": 0, + "content": "We consider the object detection task using the common MS-COCO benchmark [42]. We use the ViDT transformer architecture [67] due to its task performance and its efficiency on consumer hardware. We present the results in Tab. 2, and observe a significant and consistent improvement across a wide range of different ViDT variants. We improve using Swin-nano backbone by \\(2.6pp\\), in Swin-tiny backbone by \\(2.1pp\\) and in Swin-small backbone by \\(1pp\\). Furthermore, we also compare against an alternative distillation method, described as token matching [67]. We present these results in Tab. 3. Our method outperforms token matching by up to \\(1.1pp\\), reaching the best results when we use a larger teacher, demonstrating that our method is not limited in the cases of larger capacity gaps." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.768, + 0.651 + ], + "angle": 0, + "content": "4.3. Data limited image generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the generality of our feature distillation framework, we consider an image generation task and compare to the recent KD-DLGAN [16]. KD-DLGAN proposes to use both the text and feature embeddings for the feature guidance followed by an additional diversity loss. Using our novel framework, we show that neither of these explicit additional losses is necessary. Instead, we use a simple whitening of features to encourage the generation of diverse images. We show these results in Tab. 4. A noticeable observation in these results is that whitening can obtain the most significant improvements in the more extreme data-limited regimes. For example, when training on \\(10\\%\\) data with CIFAR-100, whitening the teacher features can improve the FID by up to 9.09. This result highlights that feature diversity is much more critical when there is insufficient training data. To show that our method is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15725" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.089, + 0.835, + 0.189 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100
10% Data20% Data100% Data10% Data20% Data100% Data
DA + KD (CLIP)22.03 ± 0.0713.70 ± 0.088.70 ± 0.0233.93 ± 0.0921.76 ± 0.0611.74 ± 0.02
DA (Baseline)23.34 ± 0.0914.53 ± 0.108.75 ± 0.0335.39 ± 0.0822.55 ± 0.0611.99 ± 0.02
KD-DLGAN14.20 ± 0.0611.01 ± 0.078.42 ± 0.0118.03 ± 0.1115.60 ± 0.0810.28 ± 0.03
DA + VkD16.42 ± 0.0710.94 ± 0.078.28 ± 0.0124.92 ± 0.1517.97 ± 0.1710.61 ± 0.08
↓ w/ whitening13.16 ± 0.0610.24 ± 0.068.20 ± 0.0316.87 ± 0.0914.00 ± 0.0710.41 ± 0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.895, + 0.237 + ], + "angle": 0, + "content": "Table 4. Comparison with the state-of-the-art over CIFAR-10 and CIFAR 100. Competitive performance is achieved using the orthogonal projection alone, however, introducing a simple whitening step is sufficient in outperforming state-of-the-art by a significant margin. All the compared methods employ BigGAN as the backbone. FID is averaged over three runs." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.246, + 0.465, + 0.383 + ], + "angle": 0, + "content": "
MethodCIFAR-10\n10% DataCIFAR-100\n10% Data
DA (Baseline) [87] NeurIPS2023.34 ± 0.0935.39 ± 0.08
FitNets [64] ICLR1422.03 ± 0.0733.93 ± 0.09
Label Distillation [34] arXiv1520.46 ± 0.1034.14 ± 0.11
PKD [54] ECCV1921.34 ± 0.0832.15 ± 0.13
SPKD [74] ICCV1919.11 ± 0.0731.97 ± 0.10
KD-DLGAN [16] CVPR2314.20 ± 0.0618.03 ± 0.11
VkD16.47 ± 0.0724.92 ± 0.15
↓ w/ whitening13.16 ± 0.0616.87 ± 0.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.45 + ], + "angle": 0, + "content": "Table 5. Comparison to other knowledge distillation methods for image generation. Results were originally reported in [16] and highlight the importance of incorporating domain-specific priors - in this case, encouraging diverse features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.56 + ], + "angle": 0, + "content": "much more general than other KD methods in the literature, we also include a comparison for the hardest data-efficient regime in Tab. 5. The results show a consistent improvement in performance, which highlight the importance of engineering both the projector architecture and the normalisation scheme, as opposed to focusing on the distance metrics alone [16, 74]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.228, + 0.584 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.652 + ], + "angle": 0, + "content": "We do a series of ablation studies to highlight the importance of our proposed building blocks. We also provide qualitative results that provide additional insights into explaining the efficacy of our distillation framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of orthogonal projections. To demonstrate the effectiveness of constraining the projection weights to be orthogonal, we consider the use of various other projection variants. We analyze the use of a projector ensemble [13], a multi-layer perceptron [51], and a standard linear layer. We present the results in Fig. 5. We observe that both the MLP and projector ensembles show improved performance over a linear layer when under short training schedules. However, when we extend the training schedule, the linear layer becomes much more effective. This is a consequence of the expressive projections beginning to learn new representations that are no longer shared by the student feature extractor. In contrast, our orthogonal projection not only improves the final accuracy but also improves the convergence properties for training, reaching state-of-the-art results in only \\(\\sim 200\\) epochs." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.272, + 0.853, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Figure 5. Comparing the performance and convergence of various projector reparameterisations. Although the MLP layer initially trains fast, it begins to saturate as it starts to learn a new representation of the data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.693 + ], + "angle": 0, + "content": "The effect of each block. We now disentangle the contribution of each block of our framework. In table 6 we report the final accuracy with and without normalisation or an orthogonal projection. We observe that, in the classification task, most of the performance improvement is obtained by our orthogonal projection. For example, the orthogonal projection alone boosts the performance from \\(76.3\\%\\) to \\(77.9\\%\\). This observation is in contrast to the generative tasks, whereby we observe a necessity to use normalisation for strong performance." + }, + { + "type": "table", + "bbox": [ + 0.584, + 0.705, + 0.807, + 0.782 + ], + "angle": 0, + "content": "
Orth.Norm.Acc@1Acc@5
XX76.393.3
X✓76.993.5
✓X77.993.9
✓✓78.394.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.793, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Table 6. Highlighting the primary importance of an orthonormal projection. Image classification on ImageNet-1K using a DeiT-student and a RegNety-160 teacher." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Whitening for generative tasks. To empirically confirm the importance of feature whitening for generative tasks, we perform an evaluation with and with it being used. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15726" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.094, + 0.16, + 0.105 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.094, + 0.298, + 0.104 + ], + "angle": 0, + "content": "Deit-Ti?" + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.094, + 0.425, + 0.105 + ], + "angle": 0, + "content": "\\(V_{k}\\) D-Ti" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.107, + 0.209, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.107, + 0.338, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.107, + 0.468, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.207, + 0.208, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.207, + 0.338, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.207, + 0.467, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.311, + 0.468, + 0.354 + ], + "angle": 0, + "content": "Figure 6. Evaluating the translational equivariance of attention maps. We select the best channel for the first translation and observe its attention maps after translating the input image again." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.468, + 0.456 + ], + "angle": 0, + "content": "present the results in Tab. 4 and 5. We observe that not only is whitening necessary to outperform previous state-of-the-art image generation, but it also leads to a larger increase in performance in the data-limited regime. This result highlights the more prominent importance of diverse features when limited training data is available." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.468, + 0.713 + ], + "angle": 0, + "content": "Distilling inductive biases. Knowledge distillation has proven effective for improving the data efficiency in training transformer models, especially when the teacher is a CNN. Unfortunately, there has been little qualitative analysis on explaining why this cross-architecture setting helps. We now quantify that this result is a consequence of providing a soft distillation of inductive biases (in this case translational equivariance). In Fig. 6, we explore the impact of applying a translation on the attention maps of a given layer. We observe that any translation of an object is reflected with a translation of the attention maps. Interestingly, this is unlike other methods, such as Deit-Ti, where the attention maps become messy after the translation of the original image. This observation suggests that their improvements may instead be attributed to some other factor, such as an implicit regularisation of the model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.468, + 0.907 + ], + "angle": 0, + "content": "Improved localisation of attention maps. We provide further insights into our orthogonal feature distillation method by analyzing the attention maps of various images. These attention maps show how well the model is attending to the salient objects in an image [7]. We compare with various other distillation methods and show the results in Fig. 7. We observe that the attention maps of our method are clustered around the boundaries of the objects, unlike the other two methods where the attention maps are spread over the entire image. In fact, we observe that our distilled model can attend much more to the salient object than the much larger CiT-S model." + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.094, + 0.567, + 0.105 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.094, + 0.671, + 0.104 + ], + "angle": 0, + "content": "Deit-Ti" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.094, + 0.762, + 0.105 + ], + "angle": 0, + "content": "CiT-S" + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.094, + 0.864, + 0.105 + ], + "angle": 0, + "content": "\\(V_{k}\\) D-Ti" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.107, + 0.892, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.415, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison to other transformer distillation methods. The best channel is selected qualitatively for all examples shown. \\( V_{k} \\) D-Ti is compared against both the same size Deitt-Ti and a much larger \\( (3.7\\times) \\) CiT-S. Best viewed in colour." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Architecture agnostic. Our method is agnostic in the choice of features and can be applied to various classifiers, object detectors, or generative models. For example, in section 4.1, we distill from a CNN to a transformer, in section 4.2, we distill between two transformers, and in section 4.3 we perform distillation in the other direction, from a transformer to a CNN." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.616, + 0.618, + 0.632 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we present a novel projection layer with a principled foundation centered on preserving the intra-batch feature similarity. The core idea of maintaining feature similarity ensures that the projection layer does not distort the underlying student representation, thus maximizing the knowledge transfer to the student backbone. We show that enforcing this constraint is equivalent to parameterising the projection weights to have orthonormal rows or columns. Our simple drop-in replacement for the projection layer leads to improved performance across a wide range of distillation tasks, from image classification to object detection. To further improve the generality of this framework, we show that whitening the teachers' features is sufficient and more effective in extending to generative tasks than other methods. We show in the experiments that our method improves state-of-the-art by up to \\(4.4\\%\\) for image classification and \\(2.6\\%\\) for object detection." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15727" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning. ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[3] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[4] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. ICML Joint Workshop on On-Device Machine Learning and Compact Deep Neural Network Representations (ODML-CDNNR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.468, + 0.353 + ], + "angle": 0, + "content": "[5] Daniel Brooks, Olivier Schwander, Frederic Barbaresco, Jean-Yves Schneider, and Matthieu Cord. Riemannian batch normalization for spd neural networks. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.468, + 0.381 + ], + "angle": 0, + "content": "[6] Adrian Bulat and Georgios Tzimiropoulos. XNOR-Net++: Improved Binary Neural Networks. BMVC, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.478 + ], + "angle": 0, + "content": "[8] Liquin Chen, Dong Wang, Zhe Gan, Jingjing Liu, Ricardo Henao, and Lawrence Carin. Wasserstein Contrastive Representation Distillation. CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.468, + 0.52 + ], + "angle": 0, + "content": "[9] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[12] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. CVPR, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[13] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved Feature Distillation via Projector Ensemble. NeurIPS, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[14] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[15] Andrew Cotter, Joseph Keshet, and Nathan Srebro. Explicit approximations of the gaussian kernel. arXiv preprint, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[16] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu1, and Eric Xing. Kd-dlgan: Data limited image generation via knowledge distillation. CVPR, 2023. 1, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[18] Alan Edelman, T. A. Arias, and Steven T. Smith. The geometry of algorithms with orthogonality constraints. SIAM Journal of Matrix Analysis and Applications, 1998. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[19] Mohamed Elfeki, Camille Couprie, Morgane Riviere, and Mohamed Elhoseiny. Gdpp: Learning diverse generations using determinantal point process. ICML, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[20] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive Model Inversion for Data-Free Knowledge Distillation. *IJCAI*, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[21] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[22] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[23] Jia Guo, Minghao Chen, Yao Hu, Chen Zhu, Xiaofei He, and Deng Cai. Reducing the Teacher-Student Gap via Spherical Knowledge Distillation. arXiv preprint, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[24] Shaopeng Guo, Yujie Wang, Quanquan Li, and Junjie Yan. DMCP: Differentiable Markov Channel Pruning for Neural Networks. CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[25] George Bruce Halsted. The collected mathematical papers of arthur Cayley. The American Mathematical Monthly, 1899. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[26] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[27] Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Río, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. Array programming with NumPy. Nature, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[28] Bobby He and Mete Ozay. Feature Kernel Distillation. ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[30] Kyle Helfrich, Devin Willmott, and Qiang Ye. Orthogonal recurrent neural networks with scaled cayley transform. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[31] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[32] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. AAAI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[33] Nicholas J. Higham. The scaling and squaring method for the matrix exponential revisited. SIAM Journal on Matrix Analysis and Applications, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2015. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15728" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[35] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. *ICLR*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[36] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[37] Youmin Kim, Jinbae Park, Younho Jang, Muhammad Ali, and Tae-hyun Oh Sung-ho Bae. Distilling Global and Local Logits with Densely Connected Relations. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. NeurIPS, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.294, + 0.445, + 0.308 + ], + "angle": 0, + "content": "[39] Yann Lecun. Optimal Brain Damage. NeurIPS, 1990. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.351 + ], + "angle": 0, + "content": "[40] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[41] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter Pruning using High-Rank Feature Map. CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. arXiv preprint, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[43] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[44] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured Knowledge Distillation for Semantic Segmentation. CVPR, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[45] David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and privileged information. *ICLR*, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.656 + ], + "angle": 0, + "content": "[47] Roy Miles and Krystian Mokolajczyk. Cascaded channel pruning using hierarchical self-distillation. BMVC, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[48] Roy Miles and Krystian Mokolajczyk. Understanding the role of the projector in knowledge distillation. AAAI, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[49] Roy Miles, Adrian Lopez Rodriguez, and Krystian Miko-lajczyk. Information Theoretic Representation Distillation. BMVC, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[50] Roy Miles, Mehmet Kerim Yucel, Bruno Manganelli, and Albert Saa-Garriga. MobileVOS: Real-Time Video Object Segmentation Contrastive Learning meets Knowledge Distillation. CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[51] K L Navaneet, Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. SimReg: Regression as a Simple Yet Effective Tool for Self-supervised Knowledge Distillation. BMVC, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[52] Yulei Niu, Long Chen, Chang Zhou, and Hanwang Zhang. Respecting transfer gap in knowledge distillation. NeurIPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[53] Wonpyo Park, Kakao Corp, Dongju Kim, and Yan Lu. Relational Knowledge Distillation. CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[54] Nikolaos Passalis and Anastasios Tefas. Learning Deep Representations with Probabilistic Knowledge Transfer. ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[55] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. NeurIPS 2017 Workshop Autodiff homepage, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[57] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[58] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schluter, Shuo-Yiin Chang, and Tara Sainath. Deep learning for audio signal processing. IEEE Journal of Selected Topics in Signal Processing, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[59] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. arXiv preprint, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[60] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing Network Design Spaces. CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[61] Sucheng Ren, Zhengqi Gao, Tianyu Hua, Zihui Xue, Yonglong Tian, Shengfeng He, and Hang Zhao. Co-advise: Cross Inductive Bias Distillation. CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[62] Pierre H. Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, and Michal Valko. Byol works even without batch statistics. arXiv preprint, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[63] Pierre H. Richemond, Allison Tam, Yunhao Tang, Florian Strub, Bilal Piot, and Felix Hill. The edge of orthogonality: A simple view of what makes byol tick. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[64] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints For Thin Deep Nets. ICLR, 2015. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[65] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. NeurIPS Workshop on Energy Efficient Machine Learning and Cognitive Computing, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[66] Seungwoo Son, Namhoon Lee, and Jaeho Lee. Maskedkd: Efficient distillation of vision transformers with masked images. arXiv preprint, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[67] Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo, Wonjae Kim, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15729" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Ming-Hsuan Yang. Vidt: An efficient and effective fully transformer-based object detector. ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[68] Suraj Srinivas and François Fleuret. Knowledge transfer with jacobian matching. ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.469, + 0.178 + ], + "angle": 0, + "content": "[69] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. *ICLR*, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[70] Yuandong Tian, Xinlei Chen, and Surya Ganguli. Understanding self-supervised Learning Dynamics without Contrastive Pairs. ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. PMLR, 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[72] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[73] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[74] Fred Tung and Greg Mori. Similarity-preserving knowledge distillation. ICCV, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.468, + 0.405 + ], + "angle": 0, + "content": "[75] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.468, + 0.448 + ], + "angle": 0, + "content": "[76] Xiang Wang, Xinlei Chen, Simon S. Du, and Yuandong Tian. Towards demystifying representation learning with non-contrastive self-supervision. arXiv preprint, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[77] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-supervision. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.468, + 0.532 + ], + "angle": 0, + "content": "[78] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.468, + 0.575 + ], + "angle": 0, + "content": "[79] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.468, + 0.63 + ], + "angle": 0, + "content": "[80] Jing Yang, Xiatian Zhu, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Knowledge distillation meets open-set semi-supervised learning. arXiv:2205.06701, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.688 + ], + "angle": 0, + "content": "[81] Ziqing Yang, Yiming Cui, Zhipeng Chen, Wanxiang Che, Ting Liu, Shijin Wang, and Guoping Hu. TextBrewer: An Open-Source Knowledge Distillation Toolkit for Natural Language Processing. ACL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.468, + 0.744 + ], + "angle": 0, + "content": "[82] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[83] Junho Yim. A Gift from Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning. CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[84] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. *ICLR*, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.872 + ], + "angle": 0, + "content": "[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaisheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[86] Borui Zhao, Renjie Song, and Yiyu Qiu. Decoupled Knowledge Distillation. CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[87] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[88] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients. arXiv preprint, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[89] Jinguo Zhu, Shixiang Tang, Dapeng Chen, and Shijie Yu. Complementary Relation Contrastive Distillation. CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15730" + } + ] +] \ No newline at end of file diff --git a/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_origin.pdf b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82d22bc33de3bc71afde8d6603369540a001badf --- /dev/null +++ b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/260ea0d3-c32f-4988-98e9-ada076a48353_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16fe47fd971580224b789b604ea627bf174282b06309f821221027cd1e4157b1 +size 3297587 diff --git a/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/full.md b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e19c5799f75517cedc2f2c985e28adb2b476217f --- /dev/null +++ b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/full.md @@ -0,0 +1,332 @@ +# $V_{k}D:$ Improving Knowledge Distillation using Orthogonal Projections + +Roy Miles Ismail Elezi Jiankang Deng Huawei Noah's Ark Lab + +# Abstract + +Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to a $4.4\%$ relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. Code and models are publicly available1. + +# 1. Introduction + +Deep learning has achieved remarkable success across a wide variety of tasks in computer vision [38], audio [58], and language [17] domains. However, its adoption is often coupled with increasing computational costs which has limited its application on resource constrained devices such as mobile phones. Fortunately, there have been many techniques proposed to train fast and efficient networks, such as weight pruning [24, 39, 41], quantisation [6, 88], and knowledge distillation [9, 34, 47, 69]. + +Knowledge distillation (KD) in particular has shown great success [4, 20, 45, 50, 69]. Its main idea is to utilise the pre-trained knowledge of a much larger (teacher) model to supervise the training of a much smaller (student) model. Traditional KD [3, 34] methods have focused on image classification by using the softmax predictions of the teacher as ground-truth labels for the student. However, doing so has made them limited and only applicable to specific modalities, or tasks. Although feature distillation [8, 69] can relax this constraint on the downstream task or modality, its + +![](images/8d6689c5adc3d3616bd25be69be41774b049c4c2a7275a2cc911e976320b1ad9.jpg) +Figure 1. Comparison to both DeiT and DeiT [71] on ImageNet-1K, where DeiT is a distilled DeiT model using a distillation token. Our proposed distillation method achieves significant improvements over DeiT-Ti, while effectively bridging the gap between the teacher and student performance for DeiT-S. + +adoption can incur significant computational costs due to the construction of expensive relational objects [49, 50, 53] and memory banks [69]. Most feature distillation pipelines can be described as using some projection [64], alignment [12], or fusion module [9]. These components, though widely used, tend to rely on heuristic design choices. Such heuristic-driven approaches often fall short in delivering new insights into the underlying mechanics of distillation and struggle to adapt to diverse tasks without the introduction of several auxiliary losses. [9, 16, 44, 85]. Furthermore, these additional losses introduce additional hyperparameters which will require tuning for specific tasks or settings. + +In this work, we propose a novel projection layer that is derived from a principal concept. Our approach focuses on one key idea: Preserving the intra-batch feature similarity. We highlight that if the similarity between features is preserved, then the projection layer will not change or alter the underlying student representation. This constraint is important since it will maximise the amount of knowledge being distilled to the student backbone. For example, if the projector is too expressive it may shortcut the distillation objective by learning some complex non-linear mapping between the + +two spaces. This result would significantly diminish the efficacy of distillation and is especially detrimental since the projector is thrown away after training. + +By enforcing the preservation of the feature similarity, we derive a reparameterisation of the projection layer itself using the set of orthogonal matrices. We propose to efficiently implement this reparameterisation by projecting the weights onto $SO(n)$ and then truncating the excess rows. In doing so, we enforce the property of row-wise orthogonality, while avoiding the need to compute any expensive matrix inversions or factorisations. We show that this constraint not only improves the student performance but also improves the training convergence and the efficacy of distilling inductive biases. + +Finally, a common component in the application of knowledge distillation for generative tasks is the use of additional auxiliary losses. These losses can encourage the generation of diverse features, which will subsequently lead to the generation of more diverse images. However, these auxiliary losses often conflict with the distillation objective and will subsequently degrade the student performance. To address this limitation, we propose a unified framework for incorporating these auxiliary objectives into the distillation loss itself using a task-specific normalisation step. Thus, we show that simply whitening the teacher features can implicitly encourage feature diversity, while removing the need for fine-tuning the hyperparameters of many additional losses. We further demonstrate the importance of this whitening step for data-limited image generation, whereby we achieve a consistent and substantial performance improvements over state-of-the-art. In summary, our contributions are outlined as follows: + +- We propose a novel orthogonal projection layer to maximise the knowledge being distilled through to the student backbone. +- We complement our projection with a task-wise normalisation that enables knowledge distillation in generative tasks. +- We apply our method to a wide range of tasks and modalities, improving over the state-of-the-art by up to $4.4\%$ on ImageNet-1K (see Fig. 1). + +# 2. Related work + +Knowledge Distillation utilises the knowledge of a pretrained model as supervision for a much smaller model, which can then enable the application and deployment within resource constrained environments. The field can be broadly divided into two main areas: logits distillation [3, 14, 34, 37, 52] and feature distillation [9, 31, 49, 64, 69, 79, 80]. Logit distillation focuses on classification based tasks and introduces an additional objective to minimise the distance between the student and teacher predictions. This was originally proposed using the KL divergence [34], how- + +ever it has since been extended using spherical normalisation [23], label decoupling [86], and probability reweighting [52]. In our work, we focus on feature distillation due to its generality to other tasks [9, 50] and modalities [65, 81]. Unfortunately, there is no underlying metric for the intermediate representation spaces, which has led to many heuristically derived solutions. For example, the hand-crafted FSP matrices were proposed [83] to capture the relation between features before and after a set of residual layers. Similarly, many other works have proposed to transfer knowledge using the construction of various Gram [28, 43, 74] or correlation-based matrices [40, 49, 57]. The activation boundaries have also been shown to be an effective supervisory signal for distillation [32], along with the gradients to capture the loss landscape [68, 89]. Another line of work can be loosely grouped together by their inspirations from the self-supervision [69, 77] or information theory literature [49, 50]. In contrast to these methods, we take a step back from the conventional curation of hand-crafted relational objects or objectives. We instead derive and demonstrate that an orthogonal projection is much more effective and can be used in conjunction with just a simple $L2$ loss. + +Self-supervised learning describes the family of pretext tasks used to learn good representations of data in the absence of any ground truth labels. This is an important topic with the overwhelming abundance of unlabelled data available and the increasing costs for human annotations. Self-supervised learning shares some significant similarities with the knowledge distillation literature. For example, contrastive learning [10, 29, 35] has already inspired many distillation methods [8, 69, 78] and asymmetric architectures [11, 21, 22, 62] are often described as a form of self-distillation [2]. However, the most salient overlap with our work instead lies with the use of a predictor. The predictor is a learnable model that maps from the online network to the momentum network. Its usage is very similar to what is described as a projector [13] in the knowledge distillation literature. DirectPred [70] explored the training dynamics of this predictor, which allowed the derivation of a closed form solution for its weights. This work was later simplified by either removing the expensive eigen-decomposition [76] or using fast matrix iterations [63]. In contrast to these works, we explore the role of the projector for knowledge distillation. By building upon a simple set of principles for KD, we are able to derive a cheap reparameterisation of the projector weights that can maximise the knowledge transfer. + +Layer reparameterisation has been widely adopted as a technique for constraining weights to introduce favourable properties. For instance, unitary matrices have been shown to address the gradient issues in RNNs [1], positive definite matrices enhance the robustness of batch normalization layers [5], and orthogonal matrices offer spectral regularization for improved generalization [1, 30]. More recent works + +![](images/ee7a3b03f147c2f59fb5ec4a78e990bb1157b1a44a421770d2bdda347cad8754.jpg) +Figure 2. Illustration of our proposed feature distillation using an orthonormal projection and task-specific feature normalisation. The orthonormal projection (a) maximises the knowledge being distilled to the student backbone, while the task-specific normalisation (b) can introduce domain-specific priors to improve model performance. $\diamond$ denotes trainable weights, while $\clubsuit$ denotes weights which are frozen. + +![](images/5a70a9d133dde5a25ba3f5ed76a4b865eaf18c44cd171aeaf82517ec99dc0500.jpg) + +![](images/25d3359b53a84e5130ddeb8f9251bdc0581084e96943a35ba30eab2650efeab6.jpg) + +have shown that low-rank matrices are effective in reducing the cost of fine-tuning large language models [36], while orthogonal matrices enable the cheap controllable fine-tuning of text-image diffusion models [59]. In our work we take these ideas into the context of knowledge distillation with an orthogonal projection. We show that this orthogonal constraint improves both the efficacy of distillation and improves the overall model convergence. + +# 3. Orthogonal Projections + +Despite the generality of feature distillation, its use is often coupled with many design decisions and heuristics. These decisions arise from the construction of multiple losses between intermediate feature maps which incur significant and unnecessary training overheads. To address these constraints, we adopt a simple feature distillation pipeline (see Fig. 2) using only the features directly before the classifier, or in the case of generative tasks, the latent representation. + +In section 3.1 we motivate the necessary conditions to maximise the efficacy of distillation through the projection. This leads to a reparameterisation of the projection as an orthogonal matrix, which is then efficiently implemented in section 3.2. In section 3.3 we provide some interesting additional insights into the properties of these orthogonal projections, while in section 3.4 we extend our distillation pipeline to improve the performance on both generative and discriminative tasks by using an additional task-specific normalisation step. + +# 3.1. Why use orthogonal projections? + +Our main objective is to mitigate the possibility of the projection layer learning any new representation of the data that is not shared by the feature extractor. This is important because the projection layer is thrown away after training and we want to match the feature extractor with the + +teacher, rather than solely matching the projected features. To achieve this, we propose to preserve the structural information through the projection. We describe this structural information using a kernel matrix $\mathbf{K} \in \mathbb{R}^{b \times b}$ , where $b$ is the batch-size. This kernel matrix captures the pairwise similarity between all features within a batch: + +$$ +\mathbf {K} _ {i j} = k \left(\mathbf {Z} _ {i} ^ {s}, \mathbf {Z} _ {j} ^ {s}\right) = \left\langle \mathbf {Z} _ {i} ^ {s}, \mathbf {Z} _ {j} ^ {s} \right\rangle_ {\mathcal {H}}, \tag {1} +$$ + +where $\mathcal{H}$ is some Hilbert space implicitly defined by the positive-definite real-valued kernel $k$ and $\mathbf{Z}^s\in \mathbb{R}^{b\times d_s}$ are the student features of dimension $d_{s}$ . We aim to preserve $\mathbf{K}$ under the application of a linear transformation of its arguments. This is equivalent to preserving the structural information of the features. We can express many practical kernels, such as the radial basis function kernel or the polynomial kernel, using a Taylor series expansion [15]: + +$$ +k \left(\mathbf {Z} _ {i} ^ {s}, \mathbf {Z} _ {j} ^ {s}\right) = \sum_ {n = 0} ^ {\infty} a _ {n} \left\langle \mathbf {Z} _ {i} ^ {s}, \mathbf {Z} _ {j} ^ {s} \right\rangle^ {n}, \tag {2} +$$ + +where $a_{n}$ are the coefficients. This expression shows that we simply need a transformation $\mathbf{P}$ that preserves inner products. Using the canonical inner product in $\mathbb{R}^{d_s}$ , we derive this constraint on $\mathbf{P}$ as follows: + +$$ +\begin{array}{l} \mathbf {Z} _ {i} ^ {s} \left(\mathbf {Z} _ {j} ^ {s}\right) ^ {T} = \mathbf {Z} _ {i} ^ {s} \mathbf {P} \left(\mathbf {Z} _ {j} ^ {s} \mathbf {P}\right) ^ {T} (3) \\ = \mathbf {Z} _ {i} ^ {s} \mathbf {P P} ^ {T} \left(\mathbf {Z} _ {j} ^ {s}\right) ^ {T}, (4) \\ \end{array} +$$ + +which holds if $\boxed{\mathbf{P}^T = \mathbf{P}^{-1}}$ . This constraint conveniently defines the special orthogonal group $SO(d_s)$ with $d_{s} = d_{t}$ . This group parameterises the set of all rotations in $\mathbb{R}^{d_s}$ , thus very naturally and intuitively preserves the idea of structural information. However, in the general case where $d_{s}\neq d_{t}$ , our projection matrix $\mathbf{P}$ is no longer square. This means + +that there exists no canonical inverse for our derived constraint to hold. Choosing the right-inverse defines the set of matrices with orthonormal rows, whereas choosing the left-inverse defines the set of matrices with orthonormal columns. Motivated by the need for an efficient reparameterisation, we focus our attention on the right-inverse, which defines the set of matrices with orthonormal rows, its transpose of which is conveniently represented as a Stiefel matrix manifold [18], denoted $V_{d_t}(\mathbb{R}^{d_s})$ . To simplify further notation, we omit this distinction between the two. + +Since the Stiefel manifold is smooth, it facilitates the use of standard gradient descent techniques using reparameterisations2. In the next section we provide an efficient implementation of this reparameterisation. + +# 3.2. Orthogonal reparameterisation + +There are a few convenient ways to ensure orthogonality of the projection matrix $\mathbf{P}$ . One of these ways is to use a Cayley transformation [25] that constructs an orthogonal matrix $\mathbf{P}$ from a skew-symmetric matrix: $\mathbf{P} = (\mathbf{I} - \mathbf{W})(\mathbf{I} + \mathbf{W})^{-1}$ where $\mathbf{W} = -\mathbf{W}^T$ . Unfortunately, this closed-form parameterisation, despite its simplicity, requires the expensive computation of a large matrix inverse. Using a QR decomposition can also be considered, but will require the use of expensive iterative algorithms, such as the Gram Schmidt process. Since an orthogonal reparameterisations map is needed for each iteration during training, it is critical for its evaluation to be computationally cheap. To address this computational constraint we propose an efficient algorithm that avoids the need for any expensive matrix inversions or factorisations. We propose to instead perform a cheap parameterisation map onto $SO(d_t)$ using the matrix exponential $\exp(\mathbf{A})$ , which can be efficiently implemented using the padé approximation. Knowing that $\mathbf{W}$ is skew-symmetric, we show that $\exp(\mathbf{W})$ is an orthogonal matrix using a few properties of the exponential: $\exp(\mathbf{W}) \cdot \exp(\mathbf{W})^T = \exp(\mathbf{W} + \mathbf{W}^T) = \exp(-\mathbf{W}^T + \mathbf{W}^T) = \exp(\mathbf{0}) = \mathbf{I}$ . We then project back to $V_{d_t}(\mathbb{R}^{d_s})$ by dropping the last $d_t - d_s$ rows. These two sequential steps are given more compactly as follows:: + +$$ +\phi : \mathbf {W} \xrightarrow {\exp (\mathbf {W})} \mathbf {A} \in S O (d _ {t}) \xrightarrow {\mathbf {A} _ {: d _ {s}}} \mathbf {P} \in V _ {d _ {t}} \left(\mathbb {R} ^ {d _ {s}}\right), \tag {5} +$$ + +where the subscript notation follows from the slicing convention in Pytorch [56] and Numpy [27]. This detour only requires the computation of one exponential, which can be cheaply evaluated using the Padé approximation [33]. We show an illustration of this re-parameterisation in Fig. 2a with the map $\phi$ described in equation 5. + +![](images/04f390bcea065e25e79cba5ea02bdd68dd520328134f8b191d0936f28f5aacf8.jpg) +(a) original + +![](images/4ab843a6e1f09303384d5ea4010e8b4960df0e1aa9d0efa1c7290a1f3dae78f7.jpg) +(b) orthogonal + +![](images/d781ed62e16d47241f67a1008b3cd63e579bdc29eda96730c3ad6b349271129c.jpg) +(c) linear +Figure 3. t-SNE visualisation [75] of features undergoing either a linear or orthogonal transformation. The orthogonal transformation preserves all of the structural feature information, whereas the linear projection can distort a lot of structure, which can diminish the efficacy of distillation. + +# 3.3. Orthogonal projections minimise redundancy + +The orthogonal transformations ensure that the projected features $\mathbf{Z}^s\mathbf{P}$ will not only be a linear combination of the original features, but also be a transformation that preserves the notion of distance between features. This result can be explained more concretely by observing that the singular values of $\mathbf{P}$ are all 1. Geometrically, this means that the projection is not squashing or distorting along any of the dimensions, which would bias the loss toward reconstructing the teachers features using only a subset of features. We illustrate this phenomenon in Fig. 3 where we find that even a simple linear projection distorts the features making them overlap with each other. This is caused because it attempts to align the space with the teacher. This level of distortion can degrade the linear separability of the features for the classifier, which impacts the model performance. In contrast, the orthogonal projection preserves the underlying feature manifold. + +# 3.4. Introducing domain-specific priors + +For many tasks it is important to invoke domain specific priors or auxiliary losses to improve model performance [16, 44]. Unfortunately, many of these auxiliary losses conflict with the distillation objective and hinder its efficacy. Instead, we propose a general framework for normalisation that naturally and implicitly incorporates these priors into the distillation objective itself. We show that standardisation is very effective for discriminative tasks by improving the model convergence. This convergence property can be attributed to the improved robustness of the distillation loss to random input perturbations. Similarly, we also show that whitening is a critical step for generative tasks by providing an implicit and soft encouragement of diverse features, which has been proven effective for generating diverse images [19]. Whitening can be much more effective and significantly cheaper than introducing additional auxiliary losses, which has been previously proposed in the literature [16]. We now provide a more detailed illustration of these ideas. + +![](images/ea776fef6ad95ad6697ee688a50e7ec8d31733d1e56c7aa8033dee3b52c6d054.jpg) +77.7% top-1 Acc +Figure 4. Visualisation of the $V_{k}$ D-Ti $\mathcal{L}_{distill}$ loss landscape with perturbations of the input image across two random dimensions. Normalisation significantly reduces the sensitivity of the loss to random perturbations, which leads to improved robustness and convergence for training. + +![](images/164209a46e6936a42ec8de4efee4794e934f85cb2c2d161806e40bc55ff2cac3.jpg) +(a) without normalisation +78.3% top-1 Acc +(b) with normalisation + +Standardisation improves model convergence. In our application of knowledge distillation to discriminative tasks, we observe that a straightforward normalisation of the teacher's representation yields a notable improvement in the robustness of the distillation loss to spurious deformations of the input image (see Fig. 4). These spurious deformations arise from the increasingly expressive families of data augmentation strategies being commonly employed for knowledge distillation [71]. We find that minimizing this loss variance can significantly improve the overall model convergence and performance. + +Whitening improves feature diversity. By whitening the teacher features, we derive a lower bound that resembles a feature diversity loss [16]. We start with an $L2$ loss between $\mathbf{Z}_s\mathbf{P} \in \mathbb{R}^{b \times d}$ and $\mathbf{Z}_t \in \mathbb{R}^{b \times d}$ . Since $\mathbf{P}$ is an orthogonal projection that preserves inner products (see section 3.1), we omit its usage to simplify analysis: + +$$ +\begin{array}{l} \mathcal {L} _ {d i s t i l l} = \left\| \mathbf {Z} ^ {s} - \mathbf {Z} ^ {t} \right\| ^ {2} \\ = \sum_ {i \neq j} \left\| \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t} - \mathbf {Z} _ {:, j} ^ {t} + \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} \\ = \sum_ {i \neq j} \left\| \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} + \left\| \mathbf {Z} _ {:, j} ^ {t} + \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} \tag {6} \\ - 2 \left\langle \mathbf {Z} _ {:; j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t}, \mathbf {Z} _ {:; j} ^ {t} + \mathbf {Z} _ {:, i} ^ {t} \right\rangle \\ \end{array} +$$ + +Since $\mathbf{Z}^t$ is whitened such that $(\mathbf{Z}^t)^T (\mathbf{Z}^t) = \mathbf{I}$ , i.e., perfect decorrelation of features, we can significantly simplify the expression above: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i s t i l l}} = \sum_ {i \neq j} \left\| \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} + 2 - 2 \langle \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t}, \mathbf {Z} _ {:, j} ^ {t} + \mathbf {Z} _ {:, i} ^ {t} \rangle \\ \geq \sum_ {i \neq j} \left\| \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} + 2 - 2 \left\| \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} \left\| \mathbf {Z} _ {:, j} ^ {t} + \mathbf {Z} _ {:, i} ^ {t} \right\| ^ {2} \\ = \boxed {\operatorname {c o n s t} - \lambda \sum_ {i \neq j} \mathbf {C} _ {j, i} ^ {2}}, \quad \text {w h e r e} \quad \mathbf {C} _ {i, j} = \left\| \mathbf {Z} _ {:, j} ^ {s} - \mathbf {Z} _ {:, i} ^ {t} \right\| \tag {7} \\ \end{array} +$$ + +where const and $\lambda \geq 3$ are both constants that do not depend on the model parameters. Here $\mathbf{C}$ is the euclidean + +cross-correlation matrix that captures the distance between all the pairs of student and teacher features. This derivation simply shows that minimising an $L2$ loss subject to an explicit whitening constraint on the teacher features provides a cross-feature objective. This cross-feature objective maximises the off-diagonal entries in the cross-correlation matrix, thus encouraging all the features to be decorrelated with respect to the teacher. We describe this process of encouraging decorrelation of features as increasing the feature diversity. We validate this connection of whitening with feature diversity in section 4.3, where we employ knowledge distillation in the context of data-efficient image generation. + +# 4. Experiments + +In this section, we evaluate the generality of our simple knowledge distillation pipeline across three distinct vision tasks: Image classification, object detection, and image generation. In each of these tasks, we consider the harder distillation settings, such as distilling cross-architecture, or in the data-efficient regimes. Throughout we use $V_{k}D$ to denote our KD method using an orthogonal projection, i.e., a matrix projection from the Stiefel manifold $\mathbf{V}_{\mathbf{k}}(\mathbb{R}^{\mathbf{d}})$ . + +
Networkacc@1Teacher#params
RegNetY-160 [60] CVPR2082.6none84M
CaiT-S24 [72] ICCV2183.4none47M
DeiT-B [73] ECCV2283.8none87M
DeiT-Ti [71] ICML2172.2none5M
CivT-Ti [61] CVPR2274.9regnety-600m + rednet-266M
MaskedKD [66] arXiv2375.4cait-s246M
Manifold [26] NeurIPS2276.5cait-s246M
DeiT-Ti [71] ICML2174.5regnety-1606M
↓, 1000 epochs76.6regnety-1606M
DearKD [12] CVPR2274.8regnety-1606M
↓, 1000 epochs77.0regnety-1606M
USKD [82] ICCV2375.0regnety-1606M
SRD [48] AAAI2477.2regnety-1606M
VkD-Ti78.3regnety-1606M
DeiT-S [71] ICML2179.8none22M
CivT-S [61] CVPR2282.0regnety-4gf + rednet-5022M
MaskedKD [66] arXiv2381.4deit3-b22M
DeiT-S [71] ICML2181.2regnety-16022M
↓, 1000 epochs82.6regnety-16022M
DearKD [12] CVPR2281.5regnety-16022M
↓, 1000 epochs82.8regnety-16022M
USKD [82] ICCV2380.8regnety-16022M
SRD [48] AAAI2482.1regnety-16022M
VkD-S82.3regnety-16022M
+ +Table 1. Data-efficient training of transformers using knowledge distillation on the ImageNet-1K dataset. Unless specified, each model is only trained for 300 epochs. + +Implementation details. We train all models in Pytorch [55] using 2 NVIDIA V100 GPUs. For the ImageNet experiments, we follow DeiT [71] using the same training schedule and optimization parameters. We also adopt Mixup [84] augmentation, but replace rand-augment + +
MethodBackboneEpochsAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)Param.FPS
ViDTSwin-nano5040.459.643.323.242.555.816M20.0 (45.8)
\( \downarrow w/ V_kD \)5043.062.346.224.845.360.1
Swin-tiny5044.864.548.725.947.662.138M17.2 (26.5)
\( \downarrow w/ V_kD \)5046.966.650.927.849.864.6
Swin-small5047.567.751.429.250.764.861M12.1 (16.5)
\( \downarrow w/ V_kD \)5048.568.452.430.852.266.0
+ +Table 2. Comparison with other detectors on COCO2017 val set. FPS is measured with batch size 1 of $800 \times 1333$ resolution on a single Tesla V100 GPU, where the value inside the parentheses is measured with batch size 4 of the same resolution to maximize GPU utilisation. All of the student models are distilled from a pre-trained ViDT-base. + +
StudentViDT (Swin-nano)ViDT (Swin-tiny)
TeacherViDT (small)ViDT (base)ViDT (small)ViDT (base)
No Distillation [67] ICLR2240.444.8
Token Matching [67] ICLR2241.541.945.846.5
VkD42.243.045.946.9
+ +Table 3. Comparison of ViDT on COCO2017 val set. We report AP for the student models distilled from different teacher models, with random gray scaling, gaussian blurring, and solarization. We use AdamW [46] optimizer with learning rate set to 0.001 and weight decay to 0.05. For the object detection experiments, we follow the same training methodology as ViDT [67] except that we replace the original token matching loss with our $V_{k}D$ . Finally, for the image generation task, we use the same training methodology as KD-DLGAN [16] except that we remove the auxiliary diversity losses and instead replace it with either teacher standardisation or whitening. We also remove any distillation from the text encoders, thus further reducing the cost of our method in comparison. + +# 4.1. Data efficient training of transformers + +We experiment with vision transformers, due to their proven success across a variety of fields. However, despite this success, they demand excessive training data and long training schedules. This limitation has motivated the use of knowledge distillation for improving the data efficiency of transformer models [71]. We compare our method with several others using the common knowledge distillation setting proposed alongside DeiT [71] that uses a CNN teacher pre-trained on ImageNet-21K. We train each student model for 300 epochs on ImageNet-1K for the image classification task. Unlike other methods that propose to leverage the efficacy of distilling through distillation tokens alone, we propose to distill directly through the patch tokens. We present the results in Tab. 1 where we massively outperform the previous state-of-the-art. In the tiny architecture, we outperform the baseline by 6.1 percentage points $(pp)$ , and the previous best method that uses the same teacher, the USKD by $3.3pp$ , or a relative improvement of $4.4\%$ . Interestingly, we perform better than DearKD trained for 1000 epochs by $1.3pp$ , showing that for distillation, it is not nec + +essary to train that long. We reach similar results in the small architecture too, where we outperform all the other methods that use similar training resources, and reach competitive results with the methods that use more than 3 times as long training time. In fact, unlike other methods, our approach bridges the gap between the teacher model that reaches $82.6\%$ accuracy without needing to introduce any excessively long training schedules. + +# 4.2. Object detection + +We consider the object detection task using the common MS-COCO benchmark [42]. We use the ViDT transformer architecture [67] due to its task performance and its efficiency on consumer hardware. We present the results in Tab. 2, and observe a significant and consistent improvement across a wide range of different ViDT variants. We improve using Swin-nano backbone by $2.6pp$ , in Swin-tiny backbone by $2.1pp$ and in Swin-small backbone by $1pp$ . Furthermore, we also compare against an alternative distillation method, described as token matching [67]. We present these results in Tab. 3. Our method outperforms token matching by up to $1.1pp$ , reaching the best results when we use a larger teacher, demonstrating that our method is not limited in the cases of larger capacity gaps. + +# 4.3. Data limited image generation + +To demonstrate the generality of our feature distillation framework, we consider an image generation task and compare to the recent KD-DLGAN [16]. KD-DLGAN proposes to use both the text and feature embeddings for the feature guidance followed by an additional diversity loss. Using our novel framework, we show that neither of these explicit additional losses is necessary. Instead, we use a simple whitening of features to encourage the generation of diverse images. We show these results in Tab. 4. A noticeable observation in these results is that whitening can obtain the most significant improvements in the more extreme data-limited regimes. For example, when training on $10\%$ data with CIFAR-100, whitening the teacher features can improve the FID by up to 9.09. This result highlights that feature diversity is much more critical when there is insufficient training data. To show that our method is + +
MethodCIFAR-10CIFAR-100
10% Data20% Data100% Data10% Data20% Data100% Data
DA + KD (CLIP)22.03 ± 0.0713.70 ± 0.088.70 ± 0.0233.93 ± 0.0921.76 ± 0.0611.74 ± 0.02
DA (Baseline)23.34 ± 0.0914.53 ± 0.108.75 ± 0.0335.39 ± 0.0822.55 ± 0.0611.99 ± 0.02
KD-DLGAN14.20 ± 0.0611.01 ± 0.078.42 ± 0.0118.03 ± 0.1115.60 ± 0.0810.28 ± 0.03
DA + VkD16.42 ± 0.0710.94 ± 0.078.28 ± 0.0124.92 ± 0.1517.97 ± 0.1710.61 ± 0.08
↓ w/ whitening13.16 ± 0.0610.24 ± 0.068.20 ± 0.0316.87 ± 0.0914.00 ± 0.0710.41 ± 0.01
+ +Table 4. Comparison with the state-of-the-art over CIFAR-10 and CIFAR 100. Competitive performance is achieved using the orthogonal projection alone, however, introducing a simple whitening step is sufficient in outperforming state-of-the-art by a significant margin. All the compared methods employ BigGAN as the backbone. FID is averaged over three runs. + +
MethodCIFAR-10 +10% DataCIFAR-100 +10% Data
DA (Baseline) [87] NeurIPS2023.34 ± 0.0935.39 ± 0.08
FitNets [64] ICLR1422.03 ± 0.0733.93 ± 0.09
Label Distillation [34] arXiv1520.46 ± 0.1034.14 ± 0.11
PKD [54] ECCV1921.34 ± 0.0832.15 ± 0.13
SPKD [74] ICCV1919.11 ± 0.0731.97 ± 0.10
KD-DLGAN [16] CVPR2314.20 ± 0.0618.03 ± 0.11
VkD16.47 ± 0.0724.92 ± 0.15
↓ w/ whitening13.16 ± 0.0616.87 ± 0.09
+ +much more general than other KD methods in the literature, we also include a comparison for the hardest data-efficient regime in Tab. 5. The results show a consistent improvement in performance, which highlight the importance of engineering both the projector architecture and the normalisation scheme, as opposed to focusing on the distance metrics alone [16, 74]. + +# 4.4. Ablation study + +We do a series of ablation studies to highlight the importance of our proposed building blocks. We also provide qualitative results that provide additional insights into explaining the efficacy of our distillation framework. + +Effectiveness of orthogonal projections. To demonstrate the effectiveness of constraining the projection weights to be orthogonal, we consider the use of various other projection variants. We analyze the use of a projector ensemble [13], a multi-layer perceptron [51], and a standard linear layer. We present the results in Fig. 5. We observe that both the MLP and projector ensembles show improved performance over a linear layer when under short training schedules. However, when we extend the training schedule, the linear layer becomes much more effective. This is a consequence of the expressive projections beginning to learn new representations that are no longer shared by the student feature extractor. In contrast, our orthogonal projection not only improves the final accuracy but also improves the convergence properties for training, reaching state-of-the-art results in only $\sim 200$ epochs. + +![](images/6f13c7773bec2662c1bbb54f9ad3bfa616193a0608e806e42c60d50a3653ae1f.jpg) +Figure 5. Comparing the performance and convergence of various projector reparameterisations. Although the MLP layer initially trains fast, it begins to saturate as it starts to learn a new representation of the data. + +The effect of each block. We now disentangle the contribution of each block of our framework. In table 6 we report the final accuracy with and without normalisation or an orthogonal projection. We observe that, in the classification task, most of the performance improvement is obtained by our orthogonal projection. For example, the orthogonal projection alone boosts the performance from $76.3\%$ to $77.9\%$ . This observation is in contrast to the generative tasks, whereby we observe a necessity to use normalisation for strong performance. + +Table 5. Comparison to other knowledge distillation methods for image generation. Results were originally reported in [16] and highlight the importance of incorporating domain-specific priors - in this case, encouraging diverse features. + +
Orth.Norm.Acc@1Acc@5
XX76.393.3
X✓76.993.5
✓X77.993.9
✓✓78.394.1
+ +Table 6. Highlighting the primary importance of an orthonormal projection. Image classification on ImageNet-1K using a DeiT-student and a RegNety-160 teacher. + +Whitening for generative tasks. To empirically confirm the importance of feature whitening for generative tasks, we perform an evaluation with and with it being used. We + +![](images/b83b3b97016be1ab7a44cbbf2e10efe31cfa256a964c500ab1bc1d57872aca43.jpg) +Input + +![](images/40a7a843000654db953b2351ccf45635b95159f85088f49477bc10e76d6f67bf.jpg) +Deit-Ti? + +![](images/72679dca277f42e7031d98cb84edb7c9ccdb9f589d34159beabe73a434fd43f3.jpg) +$V_{k}$ D-Ti + +![](images/afed8edae0044c0bebc9a63933f921fa5916f3af5e028f3211f0e82fa8f33d2c.jpg) +Figure 6. Evaluating the translational equivariance of attention maps. We select the best channel for the first translation and observe its attention maps after translating the input image again. + +![](images/c866549e74e584b3ff6470023aa22ec3a3670de423880d5db9ff88ed66d9ee37.jpg) + +![](images/e3eb74a15ab7bc75593fd2053202cf2d56c8ea9ff7ee18ba9ac7a86a862593d9.jpg) + +present the results in Tab. 4 and 5. We observe that not only is whitening necessary to outperform previous state-of-the-art image generation, but it also leads to a larger increase in performance in the data-limited regime. This result highlights the more prominent importance of diverse features when limited training data is available. + +Distilling inductive biases. Knowledge distillation has proven effective for improving the data efficiency in training transformer models, especially when the teacher is a CNN. Unfortunately, there has been little qualitative analysis on explaining why this cross-architecture setting helps. We now quantify that this result is a consequence of providing a soft distillation of inductive biases (in this case translational equivariance). In Fig. 6, we explore the impact of applying a translation on the attention maps of a given layer. We observe that any translation of an object is reflected with a translation of the attention maps. Interestingly, this is unlike other methods, such as Deit-Ti, where the attention maps become messy after the translation of the original image. This observation suggests that their improvements may instead be attributed to some other factor, such as an implicit regularisation of the model. + +Improved localisation of attention maps. We provide further insights into our orthogonal feature distillation method by analyzing the attention maps of various images. These attention maps show how well the model is attending to the salient objects in an image [7]. We compare with various other distillation methods and show the results in Fig. 7. We observe that the attention maps of our method are clustered around the boundaries of the objects, unlike the other two methods where the attention maps are spread over the entire image. In fact, we observe that our distilled model can attend much more to the salient object than the much larger CiT-S model. + +![](images/0cd4488dd766c4f571541cee8195077218b2edce01d9e6aaaad2da6f539d2835.jpg) +Input +Deit-Ti +CiT-S +$V_{k}$ D-Ti +Figure 7. Qualitative comparison to other transformer distillation methods. The best channel is selected qualitatively for all examples shown. $V_{k}$ D-Ti is compared against both the same size Deitt-Ti and a much larger $(3.7\times)$ CiT-S. Best viewed in colour. + +Architecture agnostic. Our method is agnostic in the choice of features and can be applied to various classifiers, object detectors, or generative models. For example, in section 4.1, we distill from a CNN to a transformer, in section 4.2, we distill between two transformers, and in section 4.3 we perform distillation in the other direction, from a transformer to a CNN. + +# 5. Conclusion + +In this work, we present a novel projection layer with a principled foundation centered on preserving the intra-batch feature similarity. The core idea of maintaining feature similarity ensures that the projection layer does not distort the underlying student representation, thus maximizing the knowledge transfer to the student backbone. We show that enforcing this constraint is equivalent to parameterising the projection weights to have orthonormal rows or columns. Our simple drop-in replacement for the projection layer leads to improved performance across a wide range of distillation tasks, from image classification to object detection. To further improve the generality of this framework, we show that whitening the teachers' features is sufficient and more effective in extending to generative tasks than other methods. We show in the experiments that our method improves state-of-the-art by up to $4.4\%$ for image classification and $2.6\%$ for object detection. + +# References + +[1] Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. ICML, 2016. 2 +[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning. ICLR, 2022. 2 +[3] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. CVPR, 2022. 1, 2 +[4] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. ICML Joint Workshop on On-Device Machine Learning and Compact Deep Neural Network Representations (ODML-CDNNR), 2019. 1 +[5] Daniel Brooks, Olivier Schwander, Frederic Barbaresco, Jean-Yves Schneider, and Matthieu Cord. Riemannian batch normalization for spd neural networks. NeurIPS, 2019. 2 +[6] Adrian Bulat and Georgios Tzimiropoulos. XNOR-Net++: Improved Binary Neural Networks. BMVC, 2019. 1 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 8 +[8] Liquin Chen, Dong Wang, Zhe Gan, Jingjing Liu, Ricardo Henao, and Lawrence Carin. Wasserstein Contrastive Representation Distillation. CVPR, 2020. 1, 2 +[9] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. CVPR, 2021. 1, 2 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020. 2 +[11] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. CVPR, 2021. 2 +[12] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. CVPR, 2022. 1, 5 +[13] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved Feature Distillation via Projector Ensemble. NeurIPS, 2022. 2, 7 +[14] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. ICCV, 2019. 2 +[15] Andrew Cotter, Joseph Keshet, and Nathan Srebro. Explicit approximations of the gaussian kernel. arXiv preprint, 2011. 3 +[16] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu1, and Eric Xing. Kd-dlgan: Data limited image generation via knowledge distillation. CVPR, 2023. 1, 4, 5, 6, 7 +[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2019. 1 +[18] Alan Edelman, T. A. Arias, and Steven T. Smith. The geometry of algorithms with orthogonality constraints. SIAM Journal of Matrix Analysis and Applications, 1998. 4 + +[19] Mohamed Elfeki, Camille Couprie, Morgane Riviere, and Mohamed Elhoseiny. Gdpp: Learning diverse generations using determinantal point process. ICML, 2019. 4 +[20] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive Model Inversion for Data-Free Knowledge Distillation. *IJCAI*, 2021. 1 +[21] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. CVPR, 2021. 2 +[22] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. NeurIPS, 2020. 2 +[23] Jia Guo, Minghao Chen, Yao Hu, Chen Zhu, Xiaofei He, and Deng Cai. Reducing the Teacher-Student Gap via Spherical Knowledge Distillation. arXiv preprint, 2020. 2 +[24] Shaopeng Guo, Yujie Wang, Quanquan Li, and Junjie Yan. DMCP: Differentiable Markov Channel Pruning for Neural Networks. CVPR, 2020. 1 +[25] George Bruce Halsted. The collected mathematical papers of arthur Cayley. The American Mathematical Monthly, 1899. 4 +[26] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. NeurIPS, 2022. 5 +[27] Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Río, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. Array programming with NumPy. Nature, 2020. 4 +[28] Bobby He and Mete Ozay. Feature Kernel Distillation. ICLR, 2022. 2 +[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. CVPR, 2020. 2 +[30] Kyle Helfrich, Devin Willmott, and Qiang Ye. Orthogonal recurrent neural networks with scaled cayley transform. PMLR, 2018. 2 +[31] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. ICCV, 2019. 2 +[32] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. AAAI, 2019. 2 +[33] Nicholas J. Higham. The scaling and squaring method for the matrix exponential revisited. SIAM Journal on Matrix Analysis and Applications, 2005. 4 +[34] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2015. 1, 2, 7 + +[35] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. *ICLR*, 2019. 2 +[36] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint, 2021. 3 +[37] Youmin Kim, Jinbae Park, Younho Jang, Muhammad Ali, and Tae-hyun Oh Sung-ho Bae. Distilling Global and Local Logits with Densely Connected Relations. ICCV, 2021. 2 +[38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. NeurIPS, 2012. 1 +[39] Yann Lecun. Optimal Brain Damage. NeurIPS, 1990. 1 +[40] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In ECCV, 2020. 2 +[41] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter Pruning using High-Rank Feature Map. CVPR, 2020. 1 +[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. arXiv preprint, 2015. 6 +[43] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. CVPR, 2019. 2 +[44] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured Knowledge Distillation for Semantic Segmentation. CVPR, 2019. 1, 4 +[45] David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and privileged information. *ICLR*, 2016. 1 +[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6 +[47] Roy Miles and Krystian Mokolajczyk. Cascaded channel pruning using hierarchical self-distillation. BMVC, 2020. 1 +[48] Roy Miles and Krystian Mokolajczyk. Understanding the role of the projector in knowledge distillation. AAAI, 2024. 5 +[49] Roy Miles, Adrian Lopez Rodriguez, and Krystian Miko-lajczyk. Information Theoretic Representation Distillation. BMVC, 2022. 1, 2 +[50] Roy Miles, Mehmet Kerim Yucel, Bruno Manganelli, and Albert Saa-Garriga. MobileVOS: Real-Time Video Object Segmentation Contrastive Learning meets Knowledge Distillation. CVPR, 2023. 1, 2 +[51] K L Navaneet, Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. SimReg: Regression as a Simple Yet Effective Tool for Self-supervised Knowledge Distillation. BMVC, 2021. 7 +[52] Yulei Niu, Long Chen, Chang Zhou, and Hanwang Zhang. Respecting transfer gap in knowledge distillation. NeurIPS, 2022. 2 + +[53] Wonpyo Park, Kakao Corp, Dongju Kim, and Yan Lu. Relational Knowledge Distillation. CVPR, 2019. 1 +[54] Nikolaos Passalis and Anastasios Tefas. Learning Deep Representations with Probabilistic Knowledge Transfer. ECCV, 2018. 7 +[55] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. NeurIPS 2017 Workshop Autodiff homepage, 2017. 5 +[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4 +[57] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. CVPR, 2019. 2 +[58] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schluter, Shuo-Yiin Chang, and Tara Sainath. Deep learning for audio signal processing. IEEE Journal of Selected Topics in Signal Processing, 2019. 1 +[59] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. arXiv preprint, 2023. 3 +[60] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing Network Design Spaces. CVPR, 2020. 5 +[61] Sucheng Ren, Zhengqi Gao, Tianyu Hua, Zihui Xue, Yonglong Tian, Shengfeng He, and Hang Zhao. Co-advise: Cross Inductive Bias Distillation. CVPR, 2022. 5 +[62] Pierre H. Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, and Michal Valko. Byol works even without batch statistics. arXiv preprint, 2020. 2 +[63] Pierre H. Richemond, Allison Tam, Yunhao Tang, Florian Strub, Bilal Piot, and Felix Hill. The edge of orthogonality: A simple view of what makes byol tick. arXiv preprint, 2023. 2 +[64] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints For Thin Deep Nets. ICLR, 2015. 1, 2, 7 +[65] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. NeurIPS Workshop on Energy Efficient Machine Learning and Cognitive Computing, 2019. 2 +[66] Seungwoo Son, Namhoon Lee, and Jaeho Lee. Maskedkd: Efficient distillation of vision transformers with masked images. arXiv preprint, 2023. 5 +[67] Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo, Wonjae Kim, and + +Ming-Hsuan Yang. Vidt: An efficient and effective fully transformer-based object detector. ICLR, 2021. 6 +[68] Suraj Srinivas and François Fleuret. Knowledge transfer with jacobian matching. ICML, 2018. 2 +[69] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. *ICLR*, 2019. 1, 2 +[70] Yuandong Tian, Xinlei Chen, and Surya Ganguli. Understanding self-supervised Learning Dynamics without Contrastive Pairs. ICML, 2021. 2 +[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. PMLR, 2021. 1, 5, 6 +[72] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 5 +[73] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 5 +[74] Fred Tung and Greg Mori. Similarity-preserving knowledge distillation. ICCV, 2019. 2, 7 +[75] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 4 +[76] Xiang Wang, Xinlei Chen, Simon S. Du, and Yuandong Tian. Towards demystifying representation learning with non-contrastive self-supervision. arXiv preprint, 2022. 2 +[77] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-supervision. ECCV, 2020. 2 +[78] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. ECCV, 2020. 2 +[79] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 2 +[80] Jing Yang, Xiatian Zhu, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Knowledge distillation meets open-set semi-supervised learning. arXiv:2205.06701, 2022. 2 +[81] Ziqing Yang, Yiming Cui, Zhipeng Chen, Wanxiang Che, Ting Liu, Shijin Wang, and Guoping Hu. TextBrewer: An Open-Source Knowledge Distillation Toolkit for Natural Language Processing. ACL, 2020. 2 +[82] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. ICCV, 2023. 5 +[83] Junho Yim. A Gift from Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning. CVPR, 2017. 2 +[84] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. *ICLR*, 2017. 5 +[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaisheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. CVPR, 2022. 1 +[86] Borui Zhao, Renjie Song, and Yiyu Qiu. Decoupled Knowledge Distillation. CVPR, 2022. 2 + +[87] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. NeurIPS, 2020. 7 +[88] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients. arXiv preprint, 2016. 1 +[89] Jinguo Zhu, Shixiang Tang, Dapeng Chen, and Shijie Yu. Complementary Relation Contrastive Distillation. CVPR, 2021. 2 \ No newline at end of file diff --git a/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/images.zip b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..946cfef47c53aaf44868f924c734c34f8df023c7 --- /dev/null +++ b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2370ba17c54b4d6f34b261cf5c0c561a17bd3df6d8814d05b9bef7c264c03408 +size 556494 diff --git a/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/layout.json b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3b0d57f3b1f4b722e6420d029d6dee40dabe646 --- /dev/null +++ b/2024/VkD_ Improving Knowledge Distillation using Orthogonal Projections/layout.json @@ -0,0 +1,9467 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 102, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 102, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 102, + 515, + 121 + ], + "type": "inline_equation", + "content": "V_{k}D:" + }, + { + "bbox": [ + 80, + 102, + 515, + 121 + ], + "type": "text", + "content": " Improving Knowledge Distillation using Orthogonal Projections" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 143, + 399, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 143, + 399, + 174 + ], + "spans": [ + { + "bbox": [ + 192, + 143, + 399, + 174 + ], + "type": "text", + "content": "Roy Miles Ismail Elezi Jiankang Deng Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 290, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 420 + ], + "type": "text", + "content": "Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to a " + }, + { + "bbox": [ + 46, + 228, + 290, + 420 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 46, + 228, + 290, + 420 + ], + "type": "text", + "content": " relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. Code and models are publicly available1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 128, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 128, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 128, + 456 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 464, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 572 + ], + "type": "text", + "content": "Deep learning has achieved remarkable success across a wide variety of tasks in computer vision [38], audio [58], and language [17] domains. However, its adoption is often coupled with increasing computational costs which has limited its application on resource constrained devices such as mobile phones. Fortunately, there have been many techniques proposed to train fast and efficient networks, such as weight pruning [24, 39, 41], quantisation [6, 88], and knowledge distillation [9, 34, 47, 69]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 573, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 288, + 693 + ], + "type": "text", + "content": "Knowledge distillation (KD) in particular has shown great success [4, 20, 45, 50, 69]. Its main idea is to utilise the pre-trained knowledge of a much larger (teacher) model to supervise the training of a much smaller (student) model. Traditional KD [3, 34] methods have focused on image classification by using the softmax predictions of the teacher as ground-truth labels for the student. However, doing so has made them limited and only applicable to specific modalities, or tasks. Although feature distillation [8, 69] can relax this constraint on the downstream task or modality, its" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 213, + 523, + 380 + ], + "blocks": [ + { + "bbox": [ + 315, + 213, + 523, + 380 + ], + "lines": [ + { + "bbox": [ + 315, + 213, + 523, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 523, + 380 + ], + "type": "image", + "image_path": "8d6689c5adc3d3616bd25be69be41774b049c4c2a7275a2cc911e976320b1ad9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 388, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 443 + ], + "type": "text", + "content": "Figure 1. Comparison to both DeiT and DeiT [71] on ImageNet-1K, where DeiT is a distilled DeiT model using a distillation token. Our proposed distillation method achieves significant improvements over DeiT-Ti, while effectively bridging the gap between the teacher and student performance for DeiT-S." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 593 + ], + "type": "text", + "content": "adoption can incur significant computational costs due to the construction of expensive relational objects [49, 50, 53] and memory banks [69]. Most feature distillation pipelines can be described as using some projection [64], alignment [12], or fusion module [9]. These components, though widely used, tend to rely on heuristic design choices. Such heuristic-driven approaches often fall short in delivering new insights into the underlying mechanics of distillation and struggle to adapt to diverse tasks without the introduction of several auxiliary losses. [9, 16, 44, 85]. Furthermore, these additional losses introduce additional hyperparameters which will require tuning for specific tasks or settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "In this work, we propose a novel projection layer that is derived from a principal concept. Our approach focuses on one key idea: Preserving the intra-batch feature similarity. We highlight that if the similarity between features is preserved, then the projection layer will not change or alter the underlying student representation. This constraint is important since it will maximise the amount of knowledge being distilled to the student backbone. For example, if the projector is too expressive it may shortcut the distillation objective by learning some complex non-linear mapping between the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 214, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 214, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 214, + 712 + ], + "type": "text", + "content": "1https://github.com/roymiles/vkd" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15720" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "two spaces. This result would significantly diminish the efficacy of distillation and is especially detrimental since the projector is thrown away after training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "content": "By enforcing the preservation of the feature similarity, we derive a reparameterisation of the projection layer itself using the set of orthogonal matrices. We propose to efficiently implement this reparameterisation by projecting the weights onto " + }, + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "inline_equation", + "content": "SO(n)" + }, + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "content": " and then truncating the excess rows. In doing so, we enforce the property of row-wise orthogonality, while avoiding the need to compute any expensive matrix inversions or factorisations. We show that this constraint not only improves the student performance but also improves the training convergence and the efficacy of distilling inductive biases." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 288, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 453 + ], + "type": "text", + "content": "Finally, a common component in the application of knowledge distillation for generative tasks is the use of additional auxiliary losses. These losses can encourage the generation of diverse features, which will subsequently lead to the generation of more diverse images. However, these auxiliary losses often conflict with the distillation objective and will subsequently degrade the student performance. To address this limitation, we propose a unified framework for incorporating these auxiliary objectives into the distillation loss itself using a task-specific normalisation step. Thus, we show that simply whitening the teacher features can implicitly encourage feature diversity, while removing the need for fine-tuning the hyperparameters of many additional losses. We further demonstrate the importance of this whitening step for data-limited image generation, whereby we achieve a consistent and substantial performance improvements over state-of-the-art. In summary, our contributions are outlined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 456, + 287, + 564 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 456, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 491 + ], + "type": "text", + "content": "- We propose a novel orthogonal projection layer to maximise the knowledge being distilled through to the student backbone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 492, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 287, + 526 + ], + "type": "text", + "content": "- We complement our projection with a task-wise normalisation that enables knowledge distillation in generative tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "type": "text", + "content": "- We apply our method to a wide range of tasks and modalities, improving over the state-of-the-art by up to " + }, + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "type": "text", + "content": " on ImageNet-1K (see Fig. 1)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 574, + 132, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 132, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 132, + 586 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Knowledge Distillation utilises the knowledge of a pretrained model as supervision for a much smaller model, which can then enable the application and deployment within resource constrained environments. The field can be broadly divided into two main areas: logits distillation [3, 14, 34, 37, 52] and feature distillation [9, 31, 49, 64, 69, 79, 80]. Logit distillation focuses on classification based tasks and introduces an additional objective to minimise the distance between the student and teacher predictions. This was originally proposed using the KL divergence [34], how-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": "ever it has since been extended using spherical normalisation [23], label decoupling [86], and probability reweighting [52]. In our work, we focus on feature distillation due to its generality to other tasks [9, 50] and modalities [65, 81]. Unfortunately, there is no underlying metric for the intermediate representation spaces, which has led to many heuristically derived solutions. For example, the hand-crafted FSP matrices were proposed [83] to capture the relation between features before and after a set of residual layers. Similarly, many other works have proposed to transfer knowledge using the construction of various Gram [28, 43, 74] or correlation-based matrices [40, 49, 57]. The activation boundaries have also been shown to be an effective supervisory signal for distillation [32], along with the gradients to capture the loss landscape [68, 89]. Another line of work can be loosely grouped together by their inspirations from the self-supervision [69, 77] or information theory literature [49, 50]. In contrast to these methods, we take a step back from the conventional curation of hand-crafted relational objects or objectives. We instead derive and demonstrate that an orthogonal projection is much more effective and can be used in conjunction with just a simple " + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 625 + ], + "type": "text", + "content": "Self-supervised learning describes the family of pretext tasks used to learn good representations of data in the absence of any ground truth labels. This is an important topic with the overwhelming abundance of unlabelled data available and the increasing costs for human annotations. Self-supervised learning shares some significant similarities with the knowledge distillation literature. For example, contrastive learning [10, 29, 35] has already inspired many distillation methods [8, 69, 78] and asymmetric architectures [11, 21, 22, 62] are often described as a form of self-distillation [2]. However, the most salient overlap with our work instead lies with the use of a predictor. The predictor is a learnable model that maps from the online network to the momentum network. Its usage is very similar to what is described as a projector [13] in the knowledge distillation literature. DirectPred [70] explored the training dynamics of this predictor, which allowed the derivation of a closed form solution for its weights. This work was later simplified by either removing the expensive eigen-decomposition [76] or using fast matrix iterations [63]. In contrast to these works, we explore the role of the projector for knowledge distillation. By building upon a simple set of principles for KD, we are able to derive a cheap reparameterisation of the projector weights that can maximise the knowledge transfer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Layer reparameterisation has been widely adopted as a technique for constraining weights to introduce favourable properties. For instance, unitary matrices have been shown to address the gradient issues in RNNs [1], positive definite matrices enhance the robustness of batch normalization layers [5], and orthogonal matrices offer spectral regularization for improved generalization [1, 30]. More recent works" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15721" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 306, + 235 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 306, + 235 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 306, + 235 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 306, + 235 + ], + "type": "image", + "image_path": "ee7a3b03f147c2f59fb5ec4a78e990bb1157b1a44a421770d2bdda347cad8754.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "type": "text", + "content": "Figure 2. Illustration of our proposed feature distillation using an orthonormal projection and task-specific feature normalisation. The orthonormal projection (a) maximises the knowledge being distilled to the student backbone, while the task-specific normalisation (b) can introduce domain-specific priors to improve model performance. " + }, + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "type": "text", + "content": " denotes trainable weights, while " + }, + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 46, + 242, + 545, + 277 + ], + "type": "text", + "content": " denotes weights which are frozen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 520, + 161 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 520, + 161 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 520, + 161 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 520, + 161 + ], + "type": "image", + "image_path": "5a70a9d133dde5a25ba3f5ed76a4b865eaf18c44cd171aeaf82517ec99dc0500.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 163, + 519, + 232 + ], + "blocks": [ + { + "bbox": [ + 306, + 163, + 519, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 163, + 519, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 519, + 232 + ], + "type": "image", + "image_path": "25d3359b53a84e5130ddeb8f9251bdc0581084e96943a35ba30eab2650efeab6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 380 + ], + "type": "text", + "content": "have shown that low-rank matrices are effective in reducing the cost of fine-tuning large language models [36], while orthogonal matrices enable the cheap controllable fine-tuning of text-image diffusion models [59]. In our work we take these ideas into the context of knowledge distillation with an orthogonal projection. We show that this orthogonal constraint improves both the efficacy of distillation and improves the overall model convergence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 388, + 183, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 183, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 183, + 403 + ], + "type": "text", + "content": "3. Orthogonal Projections" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "type": "text", + "content": "Despite the generality of feature distillation, its use is often coupled with many design decisions and heuristics. These decisions arise from the construction of multiple losses between intermediate feature maps which incur significant and unnecessary training overheads. To address these constraints, we adopt a simple feature distillation pipeline (see Fig. 2) using only the features directly before the classifier, or in the case of generative tasks, the latent representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 505, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 288, + 624 + ], + "type": "text", + "content": "In section 3.1 we motivate the necessary conditions to maximise the efficacy of distillation through the projection. This leads to a reparameterisation of the projection as an orthogonal matrix, which is then efficiently implemented in section 3.2. In section 3.3 we provide some interesting additional insights into the properties of these orthogonal projections, while in section 3.4 we extend our distillation pipeline to improve the performance on both generative and discriminative tasks by using an additional task-specific normalisation step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 224, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 224, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 224, + 647 + ], + "type": "text", + "content": "3.1. Why use orthogonal projections?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Our main objective is to mitigate the possibility of the projection layer learning any new representation of the data that is not shared by the feature extractor. This is important because the projection layer is thrown away after training and we want to match the feature extractor with the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "text", + "content": "teacher, rather than solely matching the projected features. To achieve this, we propose to preserve the structural information through the projection. We describe this structural information using a kernel matrix " + }, + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{b \\times b}" + }, + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "text", + "content": " is the batch-size. This kernel matrix captures the pairwise similarity between all features within a batch:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 369, + 545, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 369, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 364, + 369, + 545, + 383 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {i j} = k \\left(\\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s}\\right) = \\left\\langle \\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s} \\right\\rangle_ {\\mathcal {H}}, \\tag {1}", + "image_path": "807e928e4d6cd96484dc0d3fc1cf13fd480355d6c774d6bea1a64d5e34884b7d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "content": " is some Hilbert space implicitly defined by the positive-definite real-valued kernel " + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^s\\in \\mathbb{R}^{b\\times d_s}" + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "content": " are the student features of dimension " + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "content": ". We aim to preserve " + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 390, + 546, + 487 + ], + "type": "text", + "content": " under the application of a linear transformation of its arguments. This is equivalent to preserving the structural information of the features. We can express many practical kernels, such as the radial basis function kernel or the polynomial kernel, using a Taylor series expansion [15]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 498, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 498, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 364, + 498, + 545, + 526 + ], + "type": "interline_equation", + "content": "k \\left(\\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s}\\right) = \\sum_ {n = 0} ^ {\\infty} a _ {n} \\left\\langle \\mathbf {Z} _ {i} ^ {s}, \\mathbf {Z} _ {j} ^ {s} \\right\\rangle^ {n}, \\tag {2}", + "image_path": "9c2523117b59cdaf1b20455f50374b73983351e899b1b96e10db9cb1880fd168.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "inline_equation", + "content": "a_{n}" + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": " are the coefficients. This expression shows that we simply need a transformation " + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": " that preserves inner products. Using the canonical inner product in " + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_s}" + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": ", we derive this constraint on " + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 372, + 598, + 545, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 598, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 372, + 598, + 545, + 629 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Z} _ {i} ^ {s} \\left(\\mathbf {Z} _ {j} ^ {s}\\right) ^ {T} = \\mathbf {Z} _ {i} ^ {s} \\mathbf {P} \\left(\\mathbf {Z} _ {j} ^ {s} \\mathbf {P}\\right) ^ {T} (3) \\\\ = \\mathbf {Z} _ {i} ^ {s} \\mathbf {P P} ^ {T} \\left(\\mathbf {Z} _ {j} ^ {s}\\right) ^ {T}, (4) \\\\ \\end{array}", + "image_path": "9c5be500ac63501bbb7a77ade7021dafd9c27f32b24f1a0a8ea371ddaaef55e3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": "which holds if " + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\boxed{\\mathbf{P}^T = \\mathbf{P}^{-1}}" + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": ". This constraint conveniently defines the special orthogonal group " + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "inline_equation", + "content": "SO(d_s)" + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "inline_equation", + "content": "d_{s} = d_{t}" + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": ". This group parameterises the set of all rotations in " + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_s}" + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": ", thus very naturally and intuitively preserves the idea of structural information. However, in the general case where " + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "inline_equation", + "content": "d_{s}\\neq d_{t}" + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": ", our projection matrix " + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 637, + 546, + 714 + ], + "type": "text", + "content": " is no longer square. This means" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "15722" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "that there exists no canonical inverse for our derived constraint to hold. Choosing the right-inverse defines the set of matrices with orthonormal rows, whereas choosing the left-inverse defines the set of matrices with orthonormal columns. Motivated by the need for an efficient reparameterisation, we focus our attention on the right-inverse, which defines the set of matrices with orthonormal rows, its transpose of which is conveniently represented as a Stiefel matrix manifold [18], denoted " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "V_{d_t}(\\mathbb{R}^{d_s})" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". To simplify further notation, we omit this distinction between the two." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 288, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 288, + 242 + ], + "type": "text", + "content": "Since the Stiefel manifold is smooth, it facilitates the use of standard gradient descent techniques using reparameterisations2. In the next section we provide an efficient implementation of this reparameterisation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 253, + 217, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 217, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 217, + 266 + ], + "type": "text", + "content": "3.2. Orthogonal reparameterisation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": "There are a few convenient ways to ensure orthogonality of the projection matrix " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": ". One of these ways is to use a Cayley transformation [25] that constructs an orthogonal matrix " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " from a skew-symmetric matrix: " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = (\\mathbf{I} - \\mathbf{W})(\\mathbf{I} + \\mathbf{W})^{-1}" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = -\\mathbf{W}^T" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": ". Unfortunately, this closed-form parameterisation, despite its simplicity, requires the expensive computation of a large matrix inverse. Using a QR decomposition can also be considered, but will require the use of expensive iterative algorithms, such as the Gram Schmidt process. Since an orthogonal reparameterisations map is needed for each iteration during training, it is critical for its evaluation to be computationally cheap. To address this computational constraint we propose an efficient algorithm that avoids the need for any expensive matrix inversions or factorisations. We propose to instead perform a cheap parameterisation map onto " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "SO(d_t)" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " using the matrix exponential " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\exp(\\mathbf{A})" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": ", which can be efficiently implemented using the padé approximation. Knowing that " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " is skew-symmetric, we show that " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\exp(\\mathbf{W})" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " is an orthogonal matrix using a few properties of the exponential: " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\exp(\\mathbf{W}) \\cdot \\exp(\\mathbf{W})^T = \\exp(\\mathbf{W} + \\mathbf{W}^T) = \\exp(-\\mathbf{W}^T + \\mathbf{W}^T) = \\exp(\\mathbf{0}) = \\mathbf{I}" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": ". We then project back to " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "V_{d_t}(\\mathbb{R}^{d_s})" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " by dropping the last " + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "inline_equation", + "content": "d_t - d_s" + }, + { + "bbox": [ + 46, + 273, + 289, + 559 + ], + "type": "text", + "content": " rows. These two sequential steps are given more compactly as follows::" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 577, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\phi : \\mathbf {W} \\xrightarrow {\\exp (\\mathbf {W})} \\mathbf {A} \\in S O (d _ {t}) \\xrightarrow {\\mathbf {A} _ {: d _ {s}}} \\mathbf {P} \\in V _ {d _ {t}} \\left(\\mathbb {R} ^ {d _ {s}}\\right), \\tag {5}", + "image_path": "be3d4f12e6f0df30e1ead08fb6ce87684d521b01bc5d7e1c22d90b5ba9cf8aef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 608, + 287, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 608, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 287, + 680 + ], + "type": "text", + "content": "where the subscript notation follows from the slicing convention in Pytorch [56] and Numpy [27]. This detour only requires the computation of one exponential, which can be cheaply evaluated using the Padé approximation [33]. We show an illustration of this re-parameterisation in Fig. 2a with the map " + }, + { + "bbox": [ + 46, + 608, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 608, + 287, + 680 + ], + "type": "text", + "content": " described in equation 5." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 74, + 385, + 144 + ], + "blocks": [ + { + "bbox": [ + 316, + 74, + 385, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 74, + 385, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 74, + 385, + 144 + ], + "type": "image", + "image_path": "04f390bcea065e25e79cba5ea02bdd68dd520328134f8b191d0936f28f5aacf8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 161, + 367, + 171 + ], + "lines": [ + { + "bbox": [ + 329, + 161, + 367, + 171 + ], + "spans": [ + { + "bbox": [ + 329, + 161, + 367, + 171 + ], + "type": "text", + "content": "(a) original" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 76, + 462, + 144 + ], + "blocks": [ + { + "bbox": [ + 392, + 76, + 462, + 144 + ], + "lines": [ + { + "bbox": [ + 392, + 76, + 462, + 144 + ], + "spans": [ + { + "bbox": [ + 392, + 76, + 462, + 144 + ], + "type": "image", + "image_path": "4ab843a6e1f09303384d5ea4010e8b4960df0e1aa9d0efa1c7290a1f3dae78f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 161, + 450, + 171 + ], + "lines": [ + { + "bbox": [ + 402, + 161, + 450, + 171 + ], + "spans": [ + { + "bbox": [ + 402, + 161, + 450, + 171 + ], + "type": "text", + "content": "(b) orthogonal" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 466, + 75, + 534, + 144 + ], + "blocks": [ + { + "bbox": [ + 466, + 75, + 534, + 144 + ], + "lines": [ + { + "bbox": [ + 466, + 75, + 534, + 144 + ], + "spans": [ + { + "bbox": [ + 466, + 75, + 534, + 144 + ], + "type": "image", + "image_path": "d781ed62e16d47241f67a1008b3cd63e579bdc29eda96730c3ad6b349271129c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 161, + 520, + 171 + ], + "lines": [ + { + "bbox": [ + 488, + 161, + 520, + 171 + ], + "spans": [ + { + "bbox": [ + 488, + 161, + 520, + 171 + ], + "type": "text", + "content": "(c) linear" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 180, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 236 + ], + "type": "text", + "content": "Figure 3. t-SNE visualisation [75] of features undergoing either a linear or orthogonal transformation. The orthogonal transformation preserves all of the structural feature information, whereas the linear projection can distort a lot of structure, which can diminish the efficacy of distillation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 239, + 541, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 541, + 252 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 541, + 252 + ], + "type": "text", + "content": "3.3. Orthogonal projections minimise redundancy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "type": "text", + "content": "The orthogonal transformations ensure that the projected features " + }, + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^s\\mathbf{P}" + }, + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "type": "text", + "content": " will not only be a linear combination of the original features, but also be a transformation that preserves the notion of distance between features. This result can be explained more concretely by observing that the singular values of " + }, + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 257, + 545, + 460 + ], + "type": "text", + "content": " are all 1. Geometrically, this means that the projection is not squashing or distorting along any of the dimensions, which would bias the loss toward reconstructing the teachers features using only a subset of features. We illustrate this phenomenon in Fig. 3 where we find that even a simple linear projection distorts the features making them overlap with each other. This is caused because it attempts to align the space with the teacher. This level of distortion can degrade the linear separability of the features for the classifier, which impacts the model performance. In contrast, the orthogonal projection preserves the underlying feature manifold." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 468, + 492, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 492, + 480 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 492, + 480 + ], + "type": "text", + "content": "3.4. Introducing domain-specific priors" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 486, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 712 + ], + "type": "text", + "content": "For many tasks it is important to invoke domain specific priors or auxiliary losses to improve model performance [16, 44]. Unfortunately, many of these auxiliary losses conflict with the distillation objective and hinder its efficacy. Instead, we propose a general framework for normalisation that naturally and implicitly incorporates these priors into the distillation objective itself. We show that standardisation is very effective for discriminative tasks by improving the model convergence. This convergence property can be attributed to the improved robustness of the distillation loss to random input perturbations. Similarly, we also show that whitening is a critical step for generative tasks by providing an implicit and soft encouragement of diverse features, which has been proven effective for generating diverse images [19]. Whitening can be much more effective and significantly cheaper than introducing additional auxiliary losses, which has been previously proposed in the literature [16]. We now provide a more detailed illustration of these ideas." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "2Reparameterisations can be seen as surjective functions that map from Euclidean space back onto the manifold." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15723" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 164, + 140 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 164, + 140 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 164, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 164, + 140 + ], + "type": "image", + "image_path": "ea776fef6ad95ad6697ee688a50e7ec8d31733d1e56c7aa8033dee3b52c6d054.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 144, + 145, + 154 + ], + "lines": [ + { + "bbox": [ + 80, + 144, + 145, + 154 + ], + "spans": [ + { + "bbox": [ + 80, + 144, + 145, + 154 + ], + "type": "text", + "content": "77.7% top-1 Acc" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "type": "text", + "content": "Figure 4. Visualisation of the " + }, + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "type": "text", + "content": " D-Ti " + }, + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{distill}" + }, + { + "bbox": [ + 46, + 174, + 288, + 229 + ], + "type": "text", + "content": " loss landscape with perturbations of the input image across two random dimensions. Normalisation significantly reduces the sensitivity of the loss to random perturbations, which leads to improved robustness and convergence for training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 80, + 271, + 135 + ], + "blocks": [ + { + "bbox": [ + 63, + 155, + 157, + 165 + ], + "lines": [ + { + "bbox": [ + 63, + 155, + 157, + 165 + ], + "spans": [ + { + "bbox": [ + 63, + 155, + 157, + 165 + ], + "type": "text", + "content": "(a) without normalisation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 80, + 271, + 135 + ], + "lines": [ + { + "bbox": [ + 172, + 80, + 271, + 135 + ], + "spans": [ + { + "bbox": [ + 172, + 80, + 271, + 135 + ], + "type": "image", + "image_path": "164209a46e6936a42ec8de4efee4794e934f85cb2c2d161806e40bc55ff2cac3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 144, + 253, + 154 + ], + "lines": [ + { + "bbox": [ + 189, + 144, + 253, + 154 + ], + "spans": [ + { + "bbox": [ + 189, + 144, + 253, + 154 + ], + "type": "text", + "content": "78.3% top-1 Acc" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 155, + 263, + 166 + ], + "lines": [ + { + "bbox": [ + 179, + 155, + 263, + 166 + ], + "spans": [ + { + "bbox": [ + 179, + 155, + 263, + 166 + ], + "type": "text", + "content": "(b) with normalisation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 239, + 288, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 371 + ], + "type": "text", + "content": "Standardisation improves model convergence. In our application of knowledge distillation to discriminative tasks, we observe that a straightforward normalisation of the teacher's representation yields a notable improvement in the robustness of the distillation loss to spurious deformations of the input image (see Fig. 4). These spurious deformations arise from the increasingly expressive families of data augmentation strategies being commonly employed for knowledge distillation [71]. We find that minimizing this loss variance can significantly improve the overall model convergence and performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": "Whitening improves feature diversity. By whitening the teacher features, we derive a lower bound that resembles a feature diversity loss [16]. We start with an " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": " loss between " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_s\\mathbf{P} \\in \\mathbb{R}^{b \\times d}" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t \\in \\mathbb{R}^{b \\times d}" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": " is an orthogonal projection that preserves inner products (see section 3.1), we omit its usage to simplify analysis:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 464, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 464, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 59, + 464, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d i s t i l l} = \\left\\| \\mathbf {Z} ^ {s} - \\mathbf {Z} ^ {t} \\right\\| ^ {2} \\\\ = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} - \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\\\ = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + \\left\\| \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\tag {6} \\\\ - 2 \\left\\langle \\mathbf {Z} _ {:; j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t}, \\mathbf {Z} _ {:; j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\rangle \\\\ \\end{array}", + "image_path": "5ddabbaba4d2bc18bc87faca8277df2052208adf3c48bd9692cdbb508e23a7a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^t" + }, + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "type": "text", + "content": " is whitened such that " + }, + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}^t)^T (\\mathbf{Z}^t) = \\mathbf{I}" + }, + { + "bbox": [ + 46, + 550, + 287, + 586 + ], + "type": "text", + "content": ", i.e., perfect decorrelation of features, we can significantly simplify the expression above:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 601, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 601, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 50, + 601, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i s t i l l}} = \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + 2 - 2 \\langle \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t}, \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\rangle \\\\ \\geq \\sum_ {i \\neq j} \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} + 2 - 2 \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\left\\| \\mathbf {Z} _ {:, j} ^ {t} + \\mathbf {Z} _ {:, i} ^ {t} \\right\\| ^ {2} \\\\ = \\boxed {\\operatorname {c o n s t} - \\lambda \\sum_ {i \\neq j} \\mathbf {C} _ {j, i} ^ {2}}, \\quad \\text {w h e r e} \\quad \\mathbf {C} _ {i, j} = \\left\\| \\mathbf {Z} _ {:, j} ^ {s} - \\mathbf {Z} _ {:, i} ^ {t} \\right\\| \\tag {7} \\\\ \\end{array}", + "image_path": "79cc0e217a73aba998641431d88fcaea1335fd6955ce1b663d4d2d57c6b0e47f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where const and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda \\geq 3" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " are both constants that do not depend on the model parameters. Here " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the euclidean" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "cross-correlation matrix that captures the distance between all the pairs of student and teacher features. This derivation simply shows that minimising an " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " loss subject to an explicit whitening constraint on the teacher features provides a cross-feature objective. This cross-feature objective maximises the off-diagonal entries in the cross-correlation matrix, thus encouraging all the features to be decorrelated with respect to the teacher. We describe this process of encouraging decorrelation of features as increasing the feature diversity. We validate this connection of whitening with feature diversity in section 4.3, where we employ knowledge distillation in the context of data-efficient image generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 225, + 386, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 225, + 386, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 386, + 239 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "text", + "content": "In this section, we evaluate the generality of our simple knowledge distillation pipeline across three distinct vision tasks: Image classification, object detection, and image generation. In each of these tasks, we consider the harder distillation settings, such as distilling cross-architecture, or in the data-efficient regimes. Throughout we use " + }, + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "inline_equation", + "content": "V_{k}D" + }, + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "text", + "content": " to denote our KD method using an orthogonal projection, i.e., a matrix projection from the Stiefel manifold " + }, + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{k}}(\\mathbb{R}^{\\mathbf{d}})" + }, + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 307, + 364, + 542, + 601 + ], + "blocks": [ + { + "bbox": [ + 307, + 364, + 542, + 601 + ], + "lines": [ + { + "bbox": [ + 307, + 364, + 542, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 542, + 601 + ], + "type": "table", + "html": "
Networkacc@1Teacher#params
RegNetY-160 [60] CVPR2082.6none84M
CaiT-S24 [72] ICCV2183.4none47M
DeiT-B [73] ECCV2283.8none87M
DeiT-Ti [71] ICML2172.2none5M
CivT-Ti [61] CVPR2274.9regnety-600m + rednet-266M
MaskedKD [66] arXiv2375.4cait-s246M
Manifold [26] NeurIPS2276.5cait-s246M
DeiT-Ti [71] ICML2174.5regnety-1606M
↓, 1000 epochs76.6regnety-1606M
DearKD [12] CVPR2274.8regnety-1606M
↓, 1000 epochs77.0regnety-1606M
USKD [82] ICCV2375.0regnety-1606M
SRD [48] AAAI2477.2regnety-1606M
VkD-Ti78.3regnety-1606M
DeiT-S [71] ICML2179.8none22M
CivT-S [61] CVPR2282.0regnety-4gf + rednet-5022M
MaskedKD [66] arXiv2381.4deit3-b22M
DeiT-S [71] ICML2181.2regnety-16022M
↓, 1000 epochs82.6regnety-16022M
DearKD [12] CVPR2281.5regnety-16022M
↓, 1000 epochs82.8regnety-16022M
USKD [82] ICCV2380.8regnety-16022M
SRD [48] AAAI2482.1regnety-16022M
VkD-S82.3regnety-16022M
", + "image_path": "4181ad8a6a05b3e680ea30f079ebda8143f1acab4efb33be218e76c6844a8540.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 605, + 545, + 639 + ], + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 639 + ], + "type": "text", + "content": "Table 1. Data-efficient training of transformers using knowledge distillation on the ImageNet-1K dataset. Unless specified, each model is only trained for 300 epochs." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. We train all models in Pytorch [55] using 2 NVIDIA V100 GPUs. For the ImageNet experiments, we follow DeiT [71] using the same training schedule and optimization parameters. We also adopt Mixup [84] augmentation, but replace rand-augment" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15724" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 484, + 149 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 484, + 149 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 484, + 149 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 484, + 149 + ], + "type": "table", + "html": "
MethodBackboneEpochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Param.FPS
ViDTSwin-nano5040.459.643.323.242.555.816M20.0 (45.8)
\\( \\downarrow w/ V_kD \\)5043.062.346.224.845.360.1
Swin-tiny5044.864.548.725.947.662.138M17.2 (26.5)
\\( \\downarrow w/ V_kD \\)5046.966.650.927.849.864.6
Swin-small5047.567.751.429.250.764.861M12.1 (16.5)
\\( \\downarrow w/ V_kD \\)5048.568.452.430.852.266.0
", + "image_path": "38c134a2555ac1cf20f952e5118ad8d43cb8ccf9a40d9cb43d8c39bdb250fc81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 198, + 282, + 271 + ], + "blocks": [ + { + "bbox": [ + 46, + 158, + 547, + 192 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 547, + 192 + ], + "type": "text", + "content": "Table 2. Comparison with other detectors on COCO2017 val set. FPS is measured with batch size 1 of " + }, + { + "bbox": [ + 46, + 158, + 547, + 192 + ], + "type": "inline_equation", + "content": "800 \\times 1333" + }, + { + "bbox": [ + 46, + 158, + 547, + 192 + ], + "type": "text", + "content": " resolution on a single Tesla V100 GPU, where the value inside the parentheses is measured with batch size 4 of the same resolution to maximize GPU utilisation. All of the student models are distilled from a pre-trained ViDT-base." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 198, + 282, + 271 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 282, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 282, + 271 + ], + "type": "table", + "html": "
StudentViDT (Swin-nano)ViDT (Swin-tiny)
TeacherViDT (small)ViDT (base)ViDT (small)ViDT (base)
No Distillation [67] ICLR2240.444.8
Token Matching [67] ICLR2241.541.945.846.5
VkD42.243.045.946.9
", + "image_path": "2575f4949b90a5654a78e35593a9dc546481cddaa5f97dc0345395835106bf8c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 449 + ], + "type": "text", + "content": "Table 3. Comparison of ViDT on COCO2017 val set. We report AP for the student models distilled from different teacher models, with random gray scaling, gaussian blurring, and solarization. We use AdamW [46] optimizer with learning rate set to 0.001 and weight decay to 0.05. For the object detection experiments, we follow the same training methodology as ViDT [67] except that we replace the original token matching loss with our " + }, + { + "bbox": [ + 46, + 279, + 287, + 449 + ], + "type": "inline_equation", + "content": "V_{k}D" + }, + { + "bbox": [ + 46, + 279, + 287, + 449 + ], + "type": "text", + "content": ". Finally, for the image generation task, we use the same training methodology as KD-DLGAN [16] except that we remove the auxiliary diversity losses and instead replace it with either teacher standardisation or whitening. We also remove any distillation from the text encoders, thus further reducing the cost of our method in comparison." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 456, + 249, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 249, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 249, + 468 + ], + "type": "text", + "content": "4.1. Data efficient training of transformers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": "We experiment with vision transformers, due to their proven success across a variety of fields. However, despite this success, they demand excessive training data and long training schedules. This limitation has motivated the use of knowledge distillation for improving the data efficiency of transformer models [71]. We compare our method with several others using the common knowledge distillation setting proposed alongside DeiT [71] that uses a CNN teacher pre-trained on ImageNet-21K. We train each student model for 300 epochs on ImageNet-1K for the image classification task. Unlike other methods that propose to leverage the efficacy of distilling through distillation tokens alone, we propose to distill directly through the patch tokens. We present the results in Tab. 1 where we massively outperform the previous state-of-the-art. In the tiny architecture, we outperform the baseline by 6.1 percentage points " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "(pp)" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": ", and the previous best method that uses the same teacher, the USKD by " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "3.3pp" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": ", or a relative improvement of " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": ". Interestingly, we perform better than DearKD trained for 1000 epochs by " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.3pp" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": ", showing that for distillation, it is not nec" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 201, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 299 + ], + "type": "text", + "content": "essary to train that long. We reach similar results in the small architecture too, where we outperform all the other methods that use similar training resources, and reach competitive results with the methods that use more than 3 times as long training time. In fact, unlike other methods, our approach bridges the gap between the teacher model that reaches " + }, + { + "bbox": [ + 304, + 201, + 545, + 299 + ], + "type": "inline_equation", + "content": "82.6\\%" + }, + { + "bbox": [ + 304, + 201, + 545, + 299 + ], + "type": "text", + "content": " accuracy without needing to introduce any excessively long training schedules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 307, + 405, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 405, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 405, + 319 + ], + "type": "text", + "content": "4.2. Object detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "text", + "content": "We consider the object detection task using the common MS-COCO benchmark [42]. We use the ViDT transformer architecture [67] due to its task performance and its efficiency on consumer hardware. We present the results in Tab. 2, and observe a significant and consistent improvement across a wide range of different ViDT variants. We improve using Swin-nano backbone by " + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "inline_equation", + "content": "2.6pp" + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "text", + "content": ", in Swin-tiny backbone by " + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "inline_equation", + "content": "2.1pp" + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "text", + "content": " and in Swin-small backbone by " + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "inline_equation", + "content": "1pp" + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "text", + "content": ". Furthermore, we also compare against an alternative distillation method, described as token matching [67]. We present these results in Tab. 3. Our method outperforms token matching by up to " + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "inline_equation", + "content": "1.1pp" + }, + { + "bbox": [ + 304, + 326, + 545, + 494 + ], + "type": "text", + "content": ", reaching the best results when we use a larger teacher, demonstrating that our method is not limited in the cases of larger capacity gaps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 502, + 470, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 470, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 470, + 515 + ], + "type": "text", + "content": "4.3. Data limited image generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "To demonstrate the generality of our feature distillation framework, we consider an image generation task and compare to the recent KD-DLGAN [16]. KD-DLGAN proposes to use both the text and feature embeddings for the feature guidance followed by an additional diversity loss. Using our novel framework, we show that neither of these explicit additional losses is necessary. Instead, we use a simple whitening of features to encourage the generation of diverse images. We show these results in Tab. 4. A noticeable observation in these results is that whitening can obtain the most significant improvements in the more extreme data-limited regimes. For example, when training on " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " data with CIFAR-100, whitening the teacher features can improve the FID by up to 9.09. This result highlights that feature diversity is much more critical when there is insufficient training data. To show that our method is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15725" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 70, + 511, + 149 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 511, + 149 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 511, + 149 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 511, + 149 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100
10% Data20% Data100% Data10% Data20% Data100% Data
DA + KD (CLIP)22.03 ± 0.0713.70 ± 0.088.70 ± 0.0233.93 ± 0.0921.76 ± 0.0611.74 ± 0.02
DA (Baseline)23.34 ± 0.0914.53 ± 0.108.75 ± 0.0335.39 ± 0.0822.55 ± 0.0611.99 ± 0.02
KD-DLGAN14.20 ± 0.0611.01 ± 0.078.42 ± 0.0118.03 ± 0.1115.60 ± 0.0810.28 ± 0.03
DA + VkD16.42 ± 0.0710.94 ± 0.078.28 ± 0.0124.92 ± 0.1517.97 ± 0.1710.61 ± 0.08
↓ w/ whitening13.16 ± 0.0610.24 ± 0.068.20 ± 0.0316.87 ± 0.0914.00 ± 0.0710.41 ± 0.01
", + "image_path": "b90d8b1fa0dec873f4a6987351a80181fabe8be97de2308a808e9988c713d12f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 194, + 284, + 303 + ], + "blocks": [ + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "type": "text", + "content": "Table 4. Comparison with the state-of-the-art over CIFAR-10 and CIFAR 100. Competitive performance is achieved using the orthogonal projection alone, however, introducing a simple whitening step is sufficient in outperforming state-of-the-art by a significant margin. All the compared methods employ BigGAN as the backbone. FID is averaged over three runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 194, + 284, + 303 + ], + "lines": [ + { + "bbox": [ + 49, + 194, + 284, + 303 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 284, + 303 + ], + "type": "table", + "html": "
MethodCIFAR-10\n10% DataCIFAR-100\n10% Data
DA (Baseline) [87] NeurIPS2023.34 ± 0.0935.39 ± 0.08
FitNets [64] ICLR1422.03 ± 0.0733.93 ± 0.09
Label Distillation [34] arXiv1520.46 ± 0.1034.14 ± 0.11
PKD [54] ECCV1921.34 ± 0.0832.15 ± 0.13
SPKD [74] ICCV1919.11 ± 0.0731.97 ± 0.10
KD-DLGAN [16] CVPR2314.20 ± 0.0618.03 ± 0.11
VkD16.47 ± 0.0724.92 ± 0.15
↓ w/ whitening13.16 ± 0.0616.87 ± 0.09
", + "image_path": "cafdcc83afb164ad5558d31907f90710bd7e1ba34e613b49cbb18ccfe6ee58b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 358, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 443 + ], + "type": "text", + "content": "much more general than other KD methods in the literature, we also include a comparison for the hardest data-efficient regime in Tab. 5. The results show a consistent improvement in performance, which highlight the importance of engineering both the projector architecture and the normalisation scheme, as opposed to focusing on the distance metrics alone [16, 74]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 449, + 139, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 139, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 139, + 462 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 516 + ], + "type": "text", + "content": "We do a series of ablation studies to highlight the importance of our proposed building blocks. We also provide qualitative results that provide additional insights into explaining the efficacy of our distillation framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "Effectiveness of orthogonal projections. To demonstrate the effectiveness of constraining the projection weights to be orthogonal, we consider the use of various other projection variants. We analyze the use of a projector ensemble [13], a multi-layer perceptron [51], and a standard linear layer. We present the results in Fig. 5. We observe that both the MLP and projector ensembles show improved performance over a linear layer when under short training schedules. However, when we extend the training schedule, the linear layer becomes much more effective. This is a consequence of the expressive projections beginning to learn new representations that are no longer shared by the student feature extractor. In contrast, our orthogonal projection not only improves the final accuracy but also improves the convergence properties for training, reaching state-of-the-art results in only " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sim 200" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": " epochs." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 215, + 522, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 215, + 522, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 215, + 522, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 215, + 522, + 373 + ], + "type": "image", + "image_path": "6f13c7773bec2662c1bbb54f9ad3bfa616193a0608e806e42c60d50a3653ae1f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 382, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 426 + ], + "type": "text", + "content": "Figure 5. Comparing the performance and convergence of various projector reparameterisations. Although the MLP layer initially trains fast, it begins to saturate as it starts to learn a new representation of the data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": "The effect of each block. We now disentangle the contribution of each block of our framework. In table 6 we report the final accuracy with and without normalisation or an orthogonal projection. We observe that, in the classification task, most of the performance improvement is obtained by our orthogonal projection. For example, the orthogonal projection alone boosts the performance from " + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "inline_equation", + "content": "76.3\\%" + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "inline_equation", + "content": "77.9\\%" + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": ". This observation is in contrast to the generative tasks, whereby we observe a necessity to use normalisation for strong performance." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 357, + 558, + 493, + 619 + ], + "blocks": [ + { + "bbox": [ + 46, + 312, + 287, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 356 + ], + "type": "text", + "content": "Table 5. Comparison to other knowledge distillation methods for image generation. Results were originally reported in [16] and highlight the importance of incorporating domain-specific priors - in this case, encouraging diverse features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 357, + 558, + 493, + 619 + ], + "lines": [ + { + "bbox": [ + 357, + 558, + 493, + 619 + ], + "spans": [ + { + "bbox": [ + 357, + 558, + 493, + 619 + ], + "type": "table", + "html": "
Orth.Norm.Acc@1Acc@5
XX76.393.3
X✓76.993.5
✓X77.993.9
✓✓78.394.1
", + "image_path": "23b71ff014a7a1c4d5e4da06f26fa7be0abb16bec431d2997c1b6beaa32409cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 628, + 545, + 661 + ], + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 661 + ], + "type": "text", + "content": "Table 6. Highlighting the primary importance of an orthonormal projection. Image classification on ImageNet-1K using a DeiT-student and a RegNety-160 teacher." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Whitening for generative tasks. To empirically confirm the importance of feature whitening for generative tasks, we perform an evaluation with and with it being used. We" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15726" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 84, + 127, + 163 + ], + "blocks": [ + { + "bbox": [ + 78, + 74, + 97, + 83 + ], + "lines": [ + { + "bbox": [ + 78, + 74, + 97, + 83 + ], + "spans": [ + { + "bbox": [ + 78, + 74, + 97, + 83 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 84, + 127, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 84, + 127, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 84, + 127, + 163 + ], + "type": "image", + "image_path": "b83b3b97016be1ab7a44cbbf2e10efe31cfa256a964c500ab1bc1d57872aca43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 84, + 206, + 163 + ], + "blocks": [ + { + "bbox": [ + 151, + 74, + 182, + 82 + ], + "lines": [ + { + "bbox": [ + 151, + 74, + 182, + 82 + ], + "spans": [ + { + "bbox": [ + 151, + 74, + 182, + 82 + ], + "type": "text", + "content": "Deit-Ti?" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 84, + 206, + 163 + ], + "lines": [ + { + "bbox": [ + 129, + 84, + 206, + 163 + ], + "spans": [ + { + "bbox": [ + 129, + 84, + 206, + 163 + ], + "type": "image", + "image_path": "40a7a843000654db953b2351ccf45635b95159f85088f49477bc10e76d6f67bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 207, + 84, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 233, + 74, + 260, + 83 + ], + "lines": [ + { + "bbox": [ + 233, + 74, + 260, + 83 + ], + "spans": [ + { + "bbox": [ + 233, + 74, + 260, + 83 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 233, + 74, + 260, + 83 + ], + "type": "text", + "content": " D-Ti" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 207, + 84, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 207, + 84, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 207, + 84, + 286, + 163 + ], + "type": "image", + "image_path": "72679dca277f42e7031d98cb84edb7c9ccdb9f589d34159beabe73a434fd43f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 163, + 127, + 242 + ], + "blocks": [ + { + "bbox": [ + 49, + 163, + 127, + 242 + ], + "lines": [ + { + "bbox": [ + 49, + 163, + 127, + 242 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 127, + 242 + ], + "type": "image", + "image_path": "afed8edae0044c0bebc9a63933f921fa5916f3af5e028f3211f0e82fa8f33d2c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 246, + 286, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 246, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 286, + 280 + ], + "type": "text", + "content": "Figure 6. Evaluating the translational equivariance of attention maps. We select the best channel for the first translation and observe its attention maps after translating the input image again." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 128, + 163, + 206, + 242 + ], + "blocks": [ + { + "bbox": [ + 128, + 163, + 206, + 242 + ], + "lines": [ + { + "bbox": [ + 128, + 163, + 206, + 242 + ], + "spans": [ + { + "bbox": [ + 128, + 163, + 206, + 242 + ], + "type": "image", + "image_path": "c866549e74e584b3ff6470023aa22ec3a3670de423880d5db9ff88ed66d9ee37.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 207, + 163, + 285, + 242 + ], + "blocks": [ + { + "bbox": [ + 207, + 163, + 285, + 242 + ], + "lines": [ + { + "bbox": [ + 207, + 163, + 285, + 242 + ], + "spans": [ + { + "bbox": [ + 207, + 163, + 285, + 242 + ], + "type": "image", + "image_path": "e3eb74a15ab7bc75593fd2053202cf2d56c8ea9ff7ee18ba9ac7a86a862593d9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 289, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 286, + 361 + ], + "type": "text", + "content": "present the results in Tab. 4 and 5. We observe that not only is whitening necessary to outperform previous state-of-the-art image generation, but it also leads to a larger increase in performance in the data-limited regime. This result highlights the more prominent importance of diverse features when limited training data is available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 373, + 286, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 286, + 564 + ], + "type": "text", + "content": "Distilling inductive biases. Knowledge distillation has proven effective for improving the data efficiency in training transformer models, especially when the teacher is a CNN. Unfortunately, there has been little qualitative analysis on explaining why this cross-architecture setting helps. We now quantify that this result is a consequence of providing a soft distillation of inductive biases (in this case translational equivariance). In Fig. 6, we explore the impact of applying a translation on the attention maps of a given layer. We observe that any translation of an object is reflected with a translation of the attention maps. Interestingly, this is unlike other methods, such as Deit-Ti, where the attention maps become messy after the translation of the original image. This observation suggests that their improvements may instead be attributed to some other factor, such as an implicit regularisation of the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 574, + 286, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 286, + 718 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 286, + 718 + ], + "type": "text", + "content": "Improved localisation of attention maps. We provide further insights into our orthogonal feature distillation method by analyzing the attention maps of various images. These attention maps show how well the model is attending to the salient objects in an image [7]. We compare with various other distillation methods and show the results in Fig. 7. We observe that the attention maps of our method are clustered around the boundaries of the objects, unlike the other two methods where the attention maps are spread over the entire image. In fact, we observe that our distilled model can attend much more to the salient object than the much larger CiT-S model." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 84, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 328, + 74, + 347, + 83 + ], + "lines": [ + { + "bbox": [ + 328, + 74, + 347, + 83 + ], + "spans": [ + { + "bbox": [ + 328, + 74, + 347, + 83 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 74, + 410, + 82 + ], + "lines": [ + { + "bbox": [ + 381, + 74, + 410, + 82 + ], + "spans": [ + { + "bbox": [ + 381, + 74, + 410, + 82 + ], + "type": "text", + "content": "Deit-Ti" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 74, + 466, + 83 + ], + "lines": [ + { + "bbox": [ + 446, + 74, + 466, + 83 + ], + "spans": [ + { + "bbox": [ + 446, + 74, + 466, + 83 + ], + "type": "text", + "content": "CiT-S" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 502, + 74, + 528, + 83 + ], + "lines": [ + { + "bbox": [ + 502, + 74, + 528, + 83 + ], + "spans": [ + { + "bbox": [ + 502, + 74, + 528, + 83 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 502, + 74, + 528, + 83 + ], + "type": "text", + "content": " D-Ti" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 84, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 307, + 84, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 84, + 545, + 324 + ], + "type": "image", + "image_path": "0cd4488dd766c4f571541cee8195077218b2edce01d9e6aaaad2da6f539d2835.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison to other transformer distillation methods. The best channel is selected qualitatively for all examples shown. " + }, + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "type": "text", + "content": " D-Ti is compared against both the same size Deitt-Ti and a much larger " + }, + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "type": "inline_equation", + "content": "(3.7\\times)" + }, + { + "bbox": [ + 306, + 328, + 545, + 372 + ], + "type": "text", + "content": " CiT-S. Best viewed in colour." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 387, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 469 + ], + "type": "text", + "content": "Architecture agnostic. Our method is agnostic in the choice of features and can be applied to various classifiers, object detectors, or generative models. For example, in section 4.1, we distill from a CNN to a transformer, in section 4.2, we distill between two transformers, and in section 4.3 we perform distillation in the other direction, from a transformer to a CNN." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In this work, we present a novel projection layer with a principled foundation centered on preserving the intra-batch feature similarity. The core idea of maintaining feature similarity ensures that the projection layer does not distort the underlying student representation, thus maximizing the knowledge transfer to the student backbone. We show that enforcing this constraint is equivalent to parameterising the projection weights to have orthonormal rows or columns. Our simple drop-in replacement for the projection layer leads to improved performance across a wide range of distillation tasks, from image classification to object detection. To further improve the generality of this framework, we show that whitening the teachers' features is sufficient and more effective in extending to generative tasks than other methods. We show in the experiments that our method improves state-of-the-art by up to " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " for image classification and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " for object detection." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15727" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. ICML, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 146 + ], + "type": "text", + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning. ICLR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 147, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 287, + 190 + ], + "type": "text", + "content": "[3] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "type": "text", + "content": "[4] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. ICML Joint Workshop on On-Device Machine Learning and Compact Deep Neural Network Representations (ODML-CDNNR), 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 279 + ], + "type": "text", + "content": "[5] Daniel Brooks, Olivier Schwander, Frederic Barbaresco, Jean-Yves Schneider, and Matthieu Cord. Riemannian batch normalization for spd neural networks. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 286, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 286, + 301 + ], + "type": "text", + "content": "[6] Adrian Bulat and Georgios Tzimiropoulos. XNOR-Net++: Improved Binary Neural Networks. BMVC, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 286, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 378 + ], + "type": "text", + "content": "[8] Liquin Chen, Dong Wang, Zhe Gan, Jingjing Liu, Ricardo Henao, and Lawrence Carin. Wasserstein Contrastive Representation Distillation. CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 411 + ], + "type": "text", + "content": "[9] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 413, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 287, + 445 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 447, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 468 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 470, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 512 + ], + "type": "text", + "content": "[12] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. CVPR, 2022. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "text", + "content": "[13] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved Feature Distillation via Projector Ensemble. NeurIPS, 2022. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "type": "text", + "content": "[14] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. ICCV, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "text", + "content": "[15] Andrew Cotter, Joseph Keshet, and Nathan Srebro. Explicit approximations of the gaussian kernel. arXiv preprint, 2011. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "text", + "content": "[16] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu1, and Eric Xing. Kd-dlgan: Data limited image generation via knowledge distillation. CVPR, 2023. 1, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[18] Alan Edelman, T. A. Arias, and Steven T. Smith. The geometry of algorithms with orthogonality constraints. SIAM Journal of Matrix Analysis and Applications, 1998. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Mohamed Elfeki, Camille Couprie, Morgane Riviere, and Mohamed Elhoseiny. Gdpp: Learning diverse generations using determinantal point process. ICML, 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[20] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive Model Inversion for Data-Free Knowledge Distillation. *IJCAI*, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[21] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 186, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 251 + ], + "type": "text", + "content": "[22] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[23] Jia Guo, Minghao Chen, Yao Hu, Chen Zhu, Xiaofei He, and Deng Cai. Reducing the Teacher-Student Gap via Spherical Knowledge Distillation. arXiv preprint, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[24] Shaopeng Guo, Yujie Wang, Quanquan Li, and Junjie Yan. DMCP: Differentiable Markov Channel Pruning for Neural Networks. CVPR, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 320, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 351 + ], + "type": "text", + "content": "[25] George Bruce Halsted. The collected mathematical papers of arthur Cayley. The American Mathematical Monthly, 1899. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[26] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. NeurIPS, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 399, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 498 + ], + "type": "text", + "content": "[27] Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Río, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. Array programming with NumPy. Nature, 2020. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 498, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 518 + ], + "type": "text", + "content": "[28] Bobby He and Mete Ozay. Feature Kernel Distillation. ICLR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 521, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 554 + ], + "type": "text", + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. CVPR, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "type": "text", + "content": "[30] Kyle Helfrich, Devin Willmott, and Qiang Ye. Orthogonal recurrent neural networks with scaled cayley transform. PMLR, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "type": "text", + "content": "[31] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. ICCV, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "text", + "content": "[32] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. AAAI, 2019. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[33] Nicholas J. Higham. The scaling and squaring method for the matrix exponential revisited. SIAM Journal on Matrix Analysis and Applications, 2005. 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[34] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2015. 1, 2, 7" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15728" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[35] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. *ICLR*, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "text", + "content": "[36] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "text", + "content": "[37] Youmin Kim, Jinbae Park, Younho Jang, Muhammad Ali, and Tae-hyun Oh Sung-ho Bae. Distilling Global and Local Logits with Densely Connected Relations. ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 231 + ], + "type": "text", + "content": "[38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. NeurIPS, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 272, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 272, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 272, + 243 + ], + "type": "text", + "content": "[39] Yann Lecun. Optimal Brain Damage. NeurIPS, 1990. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "text", + "content": "[40] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "type": "text", + "content": "[41] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter Pruning using High-Rank Feature Map. CVPR, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 325, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 369 + ], + "type": "text", + "content": "[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. arXiv preprint, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[43] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. CVPR, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 405, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 437 + ], + "type": "text", + "content": "[44] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured Knowledge Distillation for Semantic Segmentation. CVPR, 2019. 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 440, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 472 + ], + "type": "text", + "content": "[45] David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and privileged information. *ICLR*, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 474, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 495 + ], + "type": "text", + "content": "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 498, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 519 + ], + "type": "text", + "content": "[47] Roy Miles and Krystian Mokolajczyk. Cascaded channel pruning using hierarchical self-distillation. BMVC, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 521, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 552 + ], + "type": "text", + "content": "[48] Roy Miles and Krystian Mokolajczyk. Understanding the role of the projector in knowledge distillation. AAAI, 2024. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[49] Roy Miles, Adrian Lopez Rodriguez, and Krystian Miko-lajczyk. Information Theoretic Representation Distillation. BMVC, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[50] Roy Miles, Mehmet Kerim Yucel, Bruno Manganelli, and Albert Saa-Garriga. MobileVOS: Real-Time Video Object Segmentation Contrastive Learning meets Knowledge Distillation. CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[51] K L Navaneet, Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. SimReg: Regression as a Simple Yet Effective Tool for Self-supervised Knowledge Distillation. BMVC, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[52] Yulei Niu, Long Chen, Chang Zhou, and Hanwang Zhang. Respecting transfer gap in knowledge distillation. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[53] Wonpyo Park, Kakao Corp, Dongju Kim, and Yan Lu. Relational Knowledge Distillation. CVPR, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 96, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 126 + ], + "type": "text", + "content": "[54] Nikolaos Passalis and Anastasios Tefas. Learning Deep Representations with Probabilistic Knowledge Transfer. ECCV, 2018. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[55] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. NeurIPS 2017 Workshop Autodiff homepage, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 185, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 270 + ], + "type": "text", + "content": "[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 272, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 313 + ], + "type": "text", + "content": "[57] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "text", + "content": "[58] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schluter, Shuo-Yiin Chang, and Tara Sainath. Deep learning for audio signal processing. IEEE Journal of Selected Topics in Signal Processing, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "text", + "content": "[59] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. arXiv preprint, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "text", + "content": "[60] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing Network Design Spaces. CVPR, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "type": "text", + "content": "[61] Sucheng Ren, Zhengqi Gao, Tianyu Hua, Zihui Xue, Yonglong Tian, Shengfeng He, and Hang Zhao. Co-advise: Cross Inductive Bias Distillation. CVPR, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "text", + "content": "[62] Pierre H. Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, and Michal Valko. Byol works even without batch statistics. arXiv preprint, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 526, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 567 + ], + "type": "text", + "content": "[63] Pierre H. Richemond, Allison Tam, Yunhao Tang, Florian Strub, Bilal Piot, and Felix Hill. The edge of orthogonality: A simple view of what makes byol tick. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "text", + "content": "[64] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints For Thin Deep Nets. ICLR, 2015. 1, 2, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "type": "text", + "content": "[65] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. NeurIPS Workshop on Energy Efficient Machine Learning and Cognitive Computing, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[66] Seungwoo Son, Namhoon Lee, and Jaeho Lee. Maskedkd: Efficient distillation of vision transformers with masked images. arXiv preprint, 2023. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "text", + "content": "[67] Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo, Wonjae Kim, and" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15729" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Ming-Hsuan Yang. Vidt: An efficient and effective fully transformer-based object detector. ICLR, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "text", + "content": "[68] Suraj Srinivas and François Fleuret. Knowledge transfer with jacobian matching. ICML, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 119, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 140 + ], + "type": "text", + "content": "[69] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. *ICLR*, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 142, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 287, + 174 + ], + "type": "text", + "content": "[70] Yuandong Tian, Xinlei Chen, and Surya Ganguli. Understanding self-supervised Learning Dynamics without Contrastive Pairs. ICML, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. PMLR, 2021. 1, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[72] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 254, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 275 + ], + "type": "text", + "content": "[73] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 277, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 286, + 297 + ], + "type": "text", + "content": "[74] Fred Tung and Greg Mori. Similarity-preserving knowledge distillation. ICCV, 2019. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 299, + 286, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 286, + 320 + ], + "type": "text", + "content": "[75] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 322, + 286, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 286, + 354 + ], + "type": "text", + "content": "[76] Xiang Wang, Xinlei Chen, Simon S. Du, and Yuandong Tian. Towards demystifying representation learning with non-contrastive self-supervision. arXiv preprint, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 355, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 286, + 387 + ], + "type": "text", + "content": "[77] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-supervision. ECCV, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 388, + 286, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 286, + 421 + ], + "type": "text", + "content": "[78] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. ECCV, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 422, + 286, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 286, + 455 + ], + "type": "text", + "content": "[79] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 456, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 498 + ], + "type": "text", + "content": "[80] Jing Yang, Xiatian Zhu, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Knowledge distillation meets open-set semi-supervised learning. arXiv:2205.06701, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 501, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 286, + 544 + ], + "type": "text", + "content": "[81] Ziqing Yang, Yiming Cui, Zhipeng Chen, Wanxiang Che, Ting Liu, Shijin Wang, and Guoping Hu. TextBrewer: An Open-Source Knowledge Distillation Toolkit for Natural Language Processing. ACL, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 545, + 286, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 286, + 589 + ], + "type": "text", + "content": "[82] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. ICCV, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "type": "text", + "content": "[83] Junho Yim. A Gift from Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning. CVPR, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "text", + "content": "[84] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. *ICLR*, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "type": "text", + "content": "[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaisheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. CVPR, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "text", + "content": "[86] Borui Zhao, Renjie Song, and Yiyu Qiu. Decoupled Knowledge Distillation. CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 183 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[87] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. NeurIPS, 2020. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[88] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients. arXiv preprint, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "text", + "content": "[89] Jinguo Zhu, Shixiang Tang, Dapeng Chen, and Shijie Yu. Complementary Relation Contrastive Distillation. CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15730" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_content_list.json b/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69ed70b0c0c1f942999e2fd2c551ca3bd4d6d6fb --- /dev/null +++ b/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_content_list.json @@ -0,0 +1,1744 @@ +[ + { + "type": "image", + "img_path": "images/afb986863174c39bf308b50146f3c2bf17f125c1acd1db62eb0d9a84d3631dc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 122, + 333, + 164 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vlogger: Make Your Dream A Vlog", + "text_level": 1, + "bbox": [ + 336, + 138, + 697, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaobin Zhuang $^{1,2}$ Kunchang Li $^{3,4,2}$ Xinyuan Chen $^{2}$ \nYaohui Wang $^{2}$ Ziwei Liu $^{5}$ Yu Qiao $^{2}$ Yali Wang $^{3,2}$", + "bbox": [ + 246, + 198, + 722, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University $^{2}$ Shanghai AI Laboratory", + "bbox": [ + 316, + 238, + 637, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences", + "bbox": [ + 272, + 255, + 679, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4University of Chinese Academy of Sciences 5S-Lab, Nanyang Technological University", + "bbox": [ + 233, + 273, + 715, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 323, + 312, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present Vlogger, a generic AI system for generating a minute-level video blog (i.e., vlog) of user descriptions. Different from short videos with a few seconds, vlog often contains a complex storyline with diversified scenes, which is challenging for most existing video generation approaches. To break through this bottleneck, our Vlogger smartly leverages Large Language Model (LLM) as Director and decomposes a long video generation task of vlog into four key stages, where we invoke various foundation models to play the critical roles of vlog professionals, including (1) Script, (2) Actor, (3) ShowMaker, and (4) Voicer. With such a design of mimicking human beings, our Vlogger can generate vlogs through explainable cooperation of top-down planning and bottom-up shooting. Moreover, we introduce a novel video diffusion model, ShowMaker, which serves as a videographer in our Vlogger for generating the video snippet of each shooting scene. By incorporating Script and Actor attentively as textual and visual prompts, it can effectively enhance spatial-temporal coherence in the snippet. Besides, we design a concise mixed training paradigm for ShowMaker, boosting its capacity for both T2V generation and prediction. Finally, the extensive experiments show that our method achieves state-of-the-art performance on zero-shot T2V generation and prediction tasks. More importantly, Vlogger can generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor.", + "bbox": [ + 76, + 354, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 791, + 207, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vlogs represent a unique form of blogging, distinguished by their utilization of video as the primary medium rather than text. Due to more lively expression in the dynamic scenes, blog has become one of the most popular online-sharing", + "bbox": [ + 76, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49b8f596ce7c194b305802120405d1253c5f8fbef479cbb63b4017879c956a07.jpg", + "image_caption": [ + "Figure 1. Overview of Vlogger. Based on the user story, our Vlogger leverages Large Language Model (LLM) as Director, and decomposes a minute-long vlog generation task into four key stages with Script, Actor, ShowMaker, and Voicer. Moreover, ShowMaker is a novel video diffusion model to generate video snippet of each shooting scene, with script and actor coherence." + ], + "image_footnote": [], + "bbox": [ + 501, + 323, + 942, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ways in the digital world. In the past two years, the remarkable success of diffusion models [31, 32, 59] has shown a great impact on video creation [8, 25, 33, 57, 70, 76, 83, 87]. Hence, there is a natural question, can we build a generic AI system to generate wonderful vlogs automatically?", + "bbox": [ + 496, + 674, + 890, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Regrettably, the majority of current video diffusion approaches mainly generate short videos with a few seconds, by temporal adaptation of image diffusion models [8, 25, 57, 70, 76, 87]. In contrast, a vlog typically constitutes a minute-level long video in the open world. Recently, there have been some attempts in long video generation [67, 83]. However, these early works either require extensive training on large well-captioned long video datasets [83] or suffer from noticeable incoherence of shot changes [67]. Hence, it is still challenging to generate a minute-level", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\spadesuit$ Interns at Shanghai AI Laboratory. $\\heartsuit$ Corresponding authors.", + "bbox": [ + 99, + 886, + 441, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8806", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vlog with complex narratives and multiple scenes.", + "bbox": [ + 76, + 90, + 408, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, we notice that a successful blog production is a systematical work in our realistic world, where the key staffs are involved in script creation, actor design, video shooting, and editing [22]. Drawing inspiration from this, we believe that, generating a long-form video blog requires elaborate systematical planning and shooting processes, rather than simply designing a generative model.", + "bbox": [ + 76, + 108, + 467, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, we propose a generic AI system for vlog generation in this paper, namely Vlogger, which can smartly address this difficult task by mimicking vlog professionals with various foundation models in the core steps. As shown in Fig. 1, we first hire a Large Language Model (LLM) as Director (e.g., GPT-4 [46]), due to its great power of linguistic knowledge understanding. Given a user story, this director schedules a four-step plan for vlog generation. (1) Script. First, we introduce a progressive script creation paradigm for the LLM Director. By the coarse-to-fine instructions in this paradigm, the LLM Director can effectively convert a user story into a script, which sufficiently describes the story by a number of shooting scenes and their corresponding shooting duration. (2) Actor. After creating the script, the LLM Director reads it again to summarize the actors, and then invokes a character designer (e.g., SD-XL [48]) to generate the reference images of these actors in the vlog. (3) ShowMaker. With the guidance of script texts and actor images, we develop a novel ShowMaker as a videographer, which can effectively generate a controllable-duration snippet for each shooting scene, with spatial-temporal coherence. (4) Voicer. Finally, the LLM Director invokes a voicer (e.g., Bark [1]) to dub the vlog with script subtitles.", + "bbox": [ + 76, + 217, + 467, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It should be noted that our Vlogger overcomes the challenges previously encountered in long video generation tasks. On one hand, it elegantly decomposes the user story into a number of shooting scenes and designs actor images that can participate in different scenes in the vlog. In this case, it can reduce spatial-temporal incoherence of abrupt shot changes, with explicit guidance of scene texts and actor images. On the other hand, Vlogger crafts individual video snippets for every scene and seamlessly integrates them into a single cohesive vlog. Consequently, this bypasses the tedious training process with large-scale long video datasets. Via such collaboration between top-down planning and bottom-up shooting, Vlogger can effectively transform an open-world story into a minute-long vlog.", + "bbox": [ + 76, + 566, + 467, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we would like to emphasize that, ShowMaker is a distinctive video diffusion model designed for generating video snippets of each shooting scene. From the structural perspective, we introduce a novel Spatial-Temporal Enhanced Block (STEB) in this model. This block can adaptively leverage scene descriptions and actor images as textual and visual prompts, which attentively guide ShowMaker to enhance spatial-temporal coherence of", + "bbox": [ + 76, + 779, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "script and actors. From the training perspective, we develop a probabilistic mode selection mechanism, which can boost the capacity of ShowMaker by mixed training of Text-to-Video (T2V) generation and prediction. More notably, by sequential combination of generation and prediction mode in the inference stage, ShowMaker can produce a video snippet with a controllable duration. This allows our Vlogger to generate a blog with a preferable duration, according to the planning of each scene in the script by LLM director.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, our method achieves the state-of-the-art performance on both zero-shot T2V generation and prediction, by expensive experiments within the popular video benchmarks. More importantly, our Vlogger outperforms the well-known long video generation method, i.e., Phenaki [67], despite utilizing only $66.7\\%$ of the training videos. Remarkably, our Vlogger is capable of generating over 5-minute vlogs, without loss of script and actor coherence in the video. The models and codes will be released afterward.", + "bbox": [ + 496, + 227, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 376, + 648, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Video Generation. Distinct from conventional unconditional and class-conditional video generation [7, 9, 13, 17, 23, 58, 63, 64, 72-74, 81, 85], T2V generation focuses on automatically converting textual descriptions into videos. This is a challenging task, as it involves understanding text semantics and translating it into video content. This often requires powerful cross-modal algorithms [50], large computing resources [15, 29], and extensive video data [3, 55, 56, 75, 76, 80]. Based on the success of diffusion models in the Text-to-Image (T2I) generation [4, 6, 18, 40, 45, 51, 52, 54], a series of such works have been recently transferred to T2V generation [8, 25, 33, 57, 70, 71, 76, 87]. However, whether trained from scratch [33, 57] or finetuned from T2I model [8, 25, 70, 71, 76, 87], most of these approaches mainly work on generating short videos with few seconds from simple descriptions. In contrast, our Vlogger can generate a minute-long vlog with complex stories.", + "bbox": [ + 496, + 402, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long Video Generation. The generation of long videos predominantly relies on parallel [83] or autoregressive [67] structures. However, these early works still face challenges for vlog generation. On the one hand, the parallel manner can relax spatial-temporal content incoherence problems by coarse-to-fine generation. However, this approach necessitates extensive and laborious training on large, well-annotated long video datasets [83]. On the other hand, the autoregressive manner can relax heavy data requirements for training long videos, by applying short video generation models iteratively with sliding windows. However, this solution often suffers from noticeable shot change and long-term incoherence [14, 28, 49, 67, 69] which becomes problematic when generating vlogs encompassing complex narratives and multiple scenes. It is worth mentioning that the community has gradually realized that delegating higher", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/02251ca8a9605faca7c26330df738c3971969d0686f1459410bb9afcc90382b2.jpg", + "image_caption": [ + "Figure 2. Top-Down Planning. Through four rounds of dialogue with the LLM, we gradually convert the user story into the final script. Based on this script, we further extract actor reference images, and then determine which actor would star in each script scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "order reasoning tasks to LLM is of great help to visual tasks [6, 26, 27, 37, 41, 44, 79, 88, 90]. Our Vlogger introduces LLM to the field of long video generation, and effectively addresses the problems of training burden and content incoherence in the previous methods, by distinct cooperation of top-down planning and bottom-up shooting.", + "bbox": [ + 75, + 378, + 472, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 486, + 169, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our Vlogger framework in detail. First, we describe how to make a vlog by planning and shooting with our Vlogger. Then, we further explain the novel design of ShowMaker. As the videographer in our Vlogger, it is critical for video generation in the vlog.", + "bbox": [ + 75, + 512, + 468, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework of Vlogger", + "text_level": 1, + "bbox": [ + 76, + 599, + 352, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate a minute-level vlog, our Vlogger leverages LLM as the director, which can effectively decompose this generation task by four key roles within the planning and shooting stages. As shown in Fig. 1, the LLM Director first creates Script and designs Actor in the planning stage. Based on Script and Actor, ShowMaker generates a video snippet for each scene in the shooting stage, and Voicer dubs subtitles of this snippet. Finally, one can combine the snippets of all the scenes as a vlog.", + "bbox": [ + 75, + 625, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Top-Down Planning", + "text_level": 1, + "bbox": [ + 76, + 784, + 272, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A vlog often comes from a user story that contains diversified content within many shot changes. Clearly, it is challenging to produce a minute-long vlog, by directly feeding such a complex story into video generation models. To this end, we propose to leverage LLM as director and decompose the user story by top-down planning in Fig. 2.", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Script Creating. First, we parse the use story into a script, which describes this story explicitly by a number of shooting scenes. In this case, we can generate a video snippet for each shooting scene, instead of learning a long video tediously from the entire story. Since LLM has shown an impressive capacity in language understanding [2, 10, 16, 46, 61, 62, 65, 86], we feed the user story into such a director for script generation. As shown in Fig. 2, we introduce a progressive creation paradigm, which can effectively parse the story by coarse-to-fine steps,", + "bbox": [ + 496, + 378, + 892, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {(i)} = \\operatorname {L L M} \\left(\\mathcal {S} ^ {(i - 1)}, \\mathcal {I} _ {S} ^ {(i)}, \\mathcal {U}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 537, + 890, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the user story $\\mathcal{U}$ and creation instruction $\\mathcal{I}_S^{(i)}$ , the LLM Director generates the current script $S^{(i)}$ from the previous one $S^{(i-1)}$ . More specifically, there are four steps including (1) Rough. First, LLM generates a basic draft of the script from the story. (2) Detailed. Then, LLM refines the rough script with story details. (3) Completed. Next, LLM checks if the detailed script misses the important parts of the story. (4) Scheduled. Finally, LLM allocates a shooting duration for each scene in the completed script, according to the scene content. For convenience, we denote the final script as $\\mathcal{S}$ in the following. It contains the descriptions of $N$ shooting scenes $\\{S_1, \\dots, S_N\\}$ and their allocated duration $\\{\\mathcal{T}_1, \\dots, \\mathcal{T}_N\\}$ . Additionally, due to the limited pages, please read the full descriptions of instructions and script in the supplementary doc.", + "bbox": [ + 496, + 566, + 892, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Actor Designing. After generating the script, it is time to design actors in the vlog. As shown in Fig. 2, we ask the LLM Director to summarize the actor list $\\mathcal{A}$ from script $S$ ,", + "bbox": [ + 496, + 794, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = \\operatorname {L L M} \\left(\\mathcal {S}, \\mathcal {I} _ {A}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 845, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_A$ is the instruction for actor summarization. Then, according to actor descriptions, the LLM Director invokes", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8808", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6bc924f6c8f59cefba01209a3d54d76064c7466e074cc3a4117ee7160ae8c73f.jpg", + "image_caption": [ + "Figure 3. Bottom-Up Shooting. For each scene, ShowMaker can generate the video snippet with script and actor coherence, by using script description and actor image as textual and visual prompts. Moreover, ShowMaker can effectively control the snippet duration, by performing generation and prediction sequentially in the inference. Finally, we apply a Text-To-Speech (TTS) model as Voicer for dubbing." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 872, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a designer to generate reference images of these actors $\\mathcal{R}$", + "bbox": [ + 76, + 410, + 460, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} = \\text {S t a b l e - D i f f u s i o n} (\\mathcal {A}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 439, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We choose Stable Diffusion XL [48] as the character designer, due to its high-quality generation. Finally, based on script $S$ and actor $\\mathcal{A}$ , the LLM Director decides the leading actor (i.e., protagonist) in each shooting scene of the script,", + "bbox": [ + 75, + 465, + 468, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} = \\operatorname {L L M} (\\mathcal {S}, \\mathcal {A}, \\mathcal {I} _ {P}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 542, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_P$ is the instruction for protagonist selection. The resulting doc $\\mathcal{P}$ is aligned with the script $S$ , where $\\mathcal{P}_n$ describes which actor appears in the scene $S_n$ . For example, $\\{\\mathcal{S}_6$ is starred by $\\mathcal{A}_1\\}$ in Fig. 2.", + "bbox": [ + 76, + 568, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Bottom-Up Shooting", + "text_level": 1, + "bbox": [ + 76, + 648, + 276, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Via the top-down planning above, the LLM Director flexibly decomposes a complex user story into several script scenes and designs actor reference image for each scene. Such a manner largely reduces the difficulty of vlog generation, since we can generate vlogs by bottom-up shooting, i.e., we just need to generate the video snippet for each shooting scene and combine all of them as a vlog.", + "bbox": [ + 75, + 672, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ShowMaker Shooting. To generate the video snippet of a shooting scene, we introduce a novel ShowMaker as the videographer, which is a video diffusion model with two distinct designs. First, it is important to maintain spatial-temporal coherence of both script and actor in the generated snippet. Hence, our ShowMaker not only takes the scene description $S_{n}$ as a textual prompt but also takes actor image of this scene $\\mathcal{R}_n$ as a visual prompt. Second, each", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scene is allocated with a shooting duration in the script. To control the duration of each scene, our ShowMaker contains two learning modes including generation and prediction in Fig.3. Specifically, it starts with the generation mode. For the shooting scene $n$ , we feed its script description $S_{n}$ and actor reference image $\\mathcal{R}_n$ into ShowMaker,", + "bbox": [ + 496, + 410, + 893, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {n} ^ {(1)} = \\operatorname {S h o w M a k e r} \\left(\\mathcal {N} _ {n} ^ {(1)} \\mid S _ {n}, \\mathcal {R} _ {n}, \\text {G e n e r a t e}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 511, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which generates the first video clip of this scene $\\mathcal{C}_n^{(1)}$ from the noisy clip $\\mathcal{N}_n^{(1)}$ . If the duration of this clip is smaller than the allocated duration $\\mathcal{T}_n$ in the script, we continue to perform the prediction mode, i.e., the last $k$ frames of the current clip $\\mathcal{C}_n^{(j)}(k)$ are used as context, when generating the next clip from the noisy input $\\mathcal{N}_n^{(j + 1)}$ ,", + "bbox": [ + 496, + 544, + 893, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {n} ^ {(j + 1)} = \\operatorname {S h o w M a k e r} \\left(\\mathcal {N} _ {n} ^ {(j + 1)} \\mid \\mathcal {S} _ {n}, \\mathcal {C} _ {n} ^ {(j)} (k), \\text {P r e d i c t}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 652, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, actor reference images are not necessary in the prediction mode, since such an actor's appearance has been shown in the current clip $\\mathcal{C}_n^{(j)}(k)$ that is used as input for prediction. This prediction procedure stops until the total duration achieves the allocated $\\mathcal{T}_n$ of this scene. Subsequently, we combine all the clips as the video snippet of this scene, i.e., $\\mathcal{C}_n = \\{\\mathcal{C}_n^{(1)},\\dots,\\mathcal{C}_n^{(J)}\\}$ .", + "bbox": [ + 496, + 684, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Voiceer Speaking. To enhance the completeness of the vlog, we apply a Text-To-Speech model (e.g., Bark [1]) as a Voiceer, which converts the scene description $S_{n}$ into the corresponding audio $\\mathcal{O}_n = \\mathrm{Bark}(S_n)$ . Finally, we add this audio to the corresponding video snippet $\\mathcal{V}_n = \\mathcal{O}_n \\oplus \\mathcal{C}_n$ , and combine all the sounded video snippets as a complete vlog, i.e., $\\mathcal{V} = \\{\\mathcal{V}_1, \\dots, \\mathcal{V}_N\\}$ .", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8809", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59984be092b7681a5995ae5e04a4a0e17682fa4eea8060b8e6037928736c90da.jpg", + "image_caption": [ + "(a) Model Architecture" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 459, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc5840383611892868a806d6cca32574491db4088501d78b1a54efd28426e880.jpg", + "image_caption": [ + "(b) Spatial-Temporal Enhanced Block (STEB)" + ], + "image_footnote": [], + "bbox": [ + 477, + 88, + 875, + 198 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be821abd5ca9dafa9157bdf171838e846e1071c5aec2aff65e117e69db17a442.jpg", + "image_caption": [ + "(c) Mode Selection", + "Figure 4. ShowMaker. (a) The Overall Architecture. (b) Spatial-Temporal Enhanced Block (STEB). Via spatial-actor and temporal-text cross attention, our STEB can further enhance actor and script coherence in the snippet. (c) Mode Selection. We introduce a mixed training paradigm of T2V generation and prediction, via probabilistic selection of masked frames." + ], + "image_footnote": [], + "bbox": [ + 475, + 229, + 879, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. ShowMaker", + "text_level": 1, + "bbox": [ + 75, + 378, + 207, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Section 3.1.2, ShowMaker plays a critical role in generating video snippet of a shooting scene. In this work, we introduce a text-to-video diffusion model for it. It follows the style of latent diffusion model [52]. In the diffusing stage, we add Gaussian noise progressively to the latent code of a training snippet. In the denoising stage, we reconstruct the latent code from the noisy latent code at any iteration step. For simplicity, we just show the denoising stage in Fig.4 (a). First, we forward a noisy training snippet into the encoder of the autoencoder to extract its latent code. Then, we feed this into a denoising U-Net [53] to learn the clean latent code. Finally, we leverage the decoder to reconstruct the original snippet with the clean latent code.", + "bbox": [ + 73, + 402, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "But, compared with the existing video diffusion models [8, 25, 33, 34, 76], our novel ShowMaker contains two distinct designs, in terms of model structure (i.e., Spatial-Temporal Enhanced Block) in Fig.4 (b), and training paradigm (i.e., Mode Selection) in Fig.4 (c).", + "bbox": [ + 75, + 599, + 468, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Spatial-Temporal Enhanced Block (STEB)", + "text_level": 1, + "bbox": [ + 76, + 691, + 426, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reconstruct the clean latent code of a training video snippet, each block in the denoising U-Net consists of both spatial and temporal operations in previous works [76]. First, spatial operations encode the feature of each frame separately in the snippet. Typically, three operations are inherited from text-to-image generation approaches [4, 45, 51, 52, 54], including spatial ConVolution (CV), spatial Self Attention (SA), and spatial Cross Attention (CA),", + "bbox": [ + 75, + 715, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {c v} = \\operatorname {C V - S p a t i a l} \\left(\\mathcal {X} _ {i n}, \\mathcal {F} _ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 842, + 468, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {s a} = \\operatorname {S A} - \\operatorname {S p a t i a l} \\left(\\mathcal {X} _ {c v}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 862, + 468, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {c a} = \\operatorname {C A} - \\text {S p a t i a l - T e x t} \\left(\\mathcal {X} _ {s a}, \\mathcal {S} _ {n}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 881, + 468, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{in}$ is the noisy feature of the training snippet, and $\\mathcal{F}_t$ is the positional embedding of the iteration step $t$ . To guide denoising with the given text, we use the scene description $S_{n}$ as Key and Value of cross attention in Eq. (9).", + "bbox": [ + 500, + 380, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, we notice that such spatial encoding does not consider actors. Hence, it inevitably suffers from actor incoherence when generating a snippet. To tackle this problem, we introduce a spatial image cross attention,", + "bbox": [ + 496, + 441, + 890, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} _ {c a} = \\operatorname {C A} - \\text {S p a t i a l - A c t o r} \\left(\\mathcal {X} _ {s a}, \\mathcal {R} _ {n}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 518, + 890, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a shooting scene $S_{n}$ , we leverage the protagonist doc (Eq. 4) to find the corresponding actor in this scene. Then, we leverage actor reference image $\\mathcal{R}_{n}$ as the visual context of spatial cross attention. Subsequently, we enhance spatial embedding with the complementary guidance of both script and actor, i.e., $\\mathcal{Z}_{se} = \\mathcal{X}_{ca} + \\beta \\mathcal{Y}_{ca}$ with a weight $\\beta$ .", + "bbox": [ + 496, + 549, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After spatial encoding, it is time to learn the correlation across frames in the snippet. Hence, the typical operation is to perform self attention along the temporal dimension,", + "bbox": [ + 496, + 641, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} _ {s a} = \\operatorname {S A - T e m p o r a l} \\left(\\mathcal {Z} _ {s e}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 702, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, such a temporal operation does not take the constraints of scene text into account. It often leads to text incoherence when generating video snippets. Hence, we introduce a temporal text cross attention,", + "bbox": [ + 496, + 732, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} _ {c a} = \\operatorname {C A} - \\text {T e m p o r a l - T e x t} \\left(\\mathcal {Z} _ {s a}, S _ {n}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 809, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we leverage the scene description $S_{n}$ as the textual context of temporal encoding. Via spatial-actor (Eq. 10) and temporal-text (Eq. 12) cross attention, our STEB can further enhance actor and script coherence in the snippet.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8810", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/87c03f2c3a895b1ebcbec990458c4721af331c5dab047ca71939fe809a118ce3.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Hand-crafted prompt." + ], + "table_body": "
MethodFVD (↓)
VideoFactory [71]410.00
Make-A-Video [57]367.23
PYoCo [25]355.19
Ours292.43
", + "bbox": [ + 83, + 88, + 267, + 166 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0f0207750094ee3dc44da23589e0052ddc719151ca847e2657c073290cd95bb7.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Class label.", + "Table 1. Zero-shot comparison with the state-of-the-art methods on UCF-101. Hand-crafted prompt from [25]." + ], + "table_body": "
MethodFVD (↓)
CogVideo (Chinese) [35]751.34
CogVideo (English) [35]701.59
MagicVideo [89]699.00
Video LDM [8]550.61
Ours525.01
", + "bbox": [ + 274, + 88, + 470, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2 Mixed Training Paradigm with Mode Selection", + "text_level": 1, + "bbox": [ + 76, + 223, + 459, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed in Section 3.1.2, ShowMaker aims to generate a video snippet with the allocated duration in the script. To this goal, it leverages the combination of both generation and prediction modes in the inference stage. Next, we explain how to train our ShowMaker to learn these modes.", + "bbox": [ + 75, + 247, + 468, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig.4 (c), we design a mode selection mechanism, which selects $k$ frames of the clean snippet as the context of the noisy snippet. The $k = 0$ setting refers to the generation mode since there is no context from the clean snippet. Alternatively, the $k > 0$ setting refers to the prediction mode, since $k$ frames of the clean snippet are already available. The goal is to generate the rest frames of the clean snippet from the noisy one. To integrate both modes into training, we design a probabilistic manner to select $k$ ,", + "bbox": [ + 75, + 324, + 468, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} (k) = \\left\\{ \\begin{array}{c c} \\alpha^ {k} - \\alpha^ {k + 1}, & k \\in [ 0, m) \\\\ \\alpha^ {k}, & k = m \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 483, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{P}(k)$ is the selection probability distribution of $k$ . Moreover, $0 < \\alpha < 1$ and $0 \\leq m$ are the manual parameters, which respectively control the mode selection tendencies of $\\mathrm{P}(k)$ and the maximum number of preserved frames.", + "bbox": [ + 75, + 535, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After determining $k$ , we introduce a frame mask $\\mathcal{M}_k$ on the latent code of the clean snippet $\\mathcal{X}_{clean}$", + "bbox": [ + 75, + 595, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {k} = \\mathcal {X} _ {\\text {c l e a n}} \\odot \\mathcal {M} _ {k}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 638, + 468, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we only preserve $k$ frames and mask the rest of the frames. Then, we concatenate the mask $\\mathcal{M}_k$ with the latent code of snippet $\\mathcal{X}_{noise}$ and $\\mathcal{X}_k$ ,", + "bbox": [ + 75, + 667, + 468, + 714 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {i n} = \\operatorname {C o n c a t} \\left(\\mathcal {X} _ {\\text {n o i s e}}, \\mathcal {X} _ {k}, \\mathcal {M} _ {k}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 726, + 468, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This produces the input feature $\\mathcal{X}_{in}$ for training the denoising U-Net in Section 3.2.1. Via such a concise probabilistic manner, we can effectively integrate both generation $(k = 0)$ and prediction $(k > 0)$ modes in the training procedure.", + "bbox": [ + 75, + 753, + 468, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 829, + 209, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To make state-of-the-art comparison, We conduct zero-shot evaluation on the popular video benchmarks, i.e., UCF-101 [60], Kinetics-400 [39], and MSR-VTT [12].", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/033a29263103d726d2c5ec699f52e16918809d199c747feab0ef9641edbb84e3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodZero-ShotPre-training VideosFID (↓)
T2V [42]XX82.13
SC [5]XX33.51
TFGAN [5]XX31.76
NUWA [78]X0.97M28.46
Phenaki [67]✓15M37.74
Ours✓10M37.23
", + "bbox": [ + 500, + 88, + 888, + 195 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2068ed2dfa66722521dcab8d9c6b48106e0ab613ed2d249a1b8300633d62cfa2.jpg", + "table_caption": [ + "Table 2. Comparison with the state-of-the-art methods on Kinetics-400. Both under a zero-shot setting, our FID is better than that of Phenaki, with only $66.7\\%$ pre-training videos." + ], + "table_footnote": [], + "table_body": "
MethodZero-ShotCLIPSIM (↑)CLIP-FID (↓)
GODIVA [77]✘0.2402-
NüWA [78]✘0.243947.68
CogVideo (Chinese) [35]✓0.261424.78
CogVideo (English) [35]✓0.263123.59
Video LDM [8]✓0.2929-
Make-A-Video [57]✓0.304913.17
PYoCo [25]✓-10.21
Ours✓0.290812.67
", + "bbox": [ + 501, + 242, + 888, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with the state-of-the-art methods on MSR-VTT. PYoCo [25] sample 59,794 videos for CLIP-FID (noted in gray) while others only sample less than 3,000 videos.", + "bbox": [ + 498, + 377, + 890, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCF-101 contains videos of 101 action categories. Following [66], we use FVD to evaluate the distance between the generated video and the real video. Kinetics-400 is a dataset comprising videos of 400 action categories. Following [5, 42, 67, 78], we use FID [30] to assess the performance of video generation. MSR-VTT is a video dataset with open-vocabulary captions, wherein CLIPSIM [77] and CLIP-FID [47] are commonly employed for evaluating the T2V generation. Moreover, since these existing benchmarks either have a small number of testing videos or contain only action labels without complex descriptions, we propose to collect an evaluation benchmark for ablation studies. It is called as Vimeo11k, where we collect 11,293 open-world videos along with their captions from 10 mainstream categories in Vimeo. To our best knowledge, it is the largest testing benchmark for zero-shot video generation. More details and experiments can be found in the supplementary doc.", + "bbox": [ + 496, + 431, + 892, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. (1) Director & Script & Actor & Voicer. We choose GPT-4 [46] as our LLM director to generate script. The specific instructions can be found in the supplementary doc. Then we employ Stable Diffusion XL [48] and Bark [1] as our designer and voicer to generate reference actor images and convert scripts into speech. (2) ShowMaker. We choose SD-1.4 [52] as our base model, and follow [57] to add temporal self attention. Then, we add our temporal text cross attention on top of each temporal self attention. We expand the input channel of the conv-in layer of U-Net from 4 to 9, so that model can take the concatenated feature in Eq. (15) as input. We use zero initialization for newly added channels. We use CLIP ViT-L/14 [50] as text encoder $\\varepsilon_{T}$ , VQVAE [20] as autoencoder", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8811", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/138b28faf8d1c260ede75d2a4290bff96e4e216c06206a80870633c8c5e729dd.jpg", + "image_caption": [ + "Figure 5. Long video generation. Comparison on UCF-101 (1000 frames). The lower FVD, the better generation performance." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 446, + 243 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aeb7ecc62ee8f684e38e81aefed1075bdbe96901c8641243a4bfab96b83cb189.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIP-I (↑)CLIP-T (↑)Vlog Duration (↑)
w/o autoregressive [91]0.56830.27521min03s
only autoregressive [67]0.56420.25353min58s
Ours0.62940.27563min58s
", + "bbox": [ + 78, + 275, + 467, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "consisting of $\\varepsilon$ and $\\mathcal{D}$ . Besides, we use OpenCLIP ViT-H/14 [38] as image encoder $\\varepsilon_{I}$ , and add spatial image cross attention in the STEB block. The diffusion step $T$ is set to 1000 as [52]. For training, we set the parameter of mode selection as $\\alpha = 0.6$ and $m = 6$ in Eq. (13), and $\\beta$ is set to 1. Following [19, 33, 34, 67, 76], we employ publicly available image dataset (i.e., Laion400M [55]) and video dataset (i.e., WebVid10M [3]) for joint training. More training details can be found in the supplementary doc.", + "bbox": [ + 75, + 383, + 467, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 527, + 366, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 shows that, no matter whether the input text is the class label or hand-crafted prompt, our method achieves the best FVD performance of zero-shot video generation on UCF-101. Tab. 2 shows that, compared to Phenaki [67], our method achieves a better FID performance of zero-shot setting on Kinetics-400, but only using $66.7\\%$ training videos. Furthermore, our generated videos have a resolution of $320 \\times 512$ , which is higher than Phenaki's $256 \\times 256$ . Tab. 3 shows that, our method achieves remarkably competitive FID and CLIPSIM performance on MSR-VTT.", + "bbox": [ + 75, + 551, + 467, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, Fig. 5 illustrates that, we significantly surpass TATS [24] (i.e., the only open-source long video generation model within our knowledge), for generating 1000-frame videos on UCF-101. Moreover, our method does not encounter the issue of TATS, where the video quality continuously declines as the number of frames increases. It is noteworthy that, our method achieves this performance by zero-shot generation, without any finetuning on UCF-101.", + "bbox": [ + 75, + 702, + 467, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 832, + 228, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vlog Generation Process. Tab. 4 presents a comparison between different generation processes. For a fair comparison, we use the same scripts and employ our ShowMaker", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efc26d8a5e467255e0cf3a20598a6737b9421e2d25a3c27942748fea7e853f67.jpg", + "image_caption": [ + "Figure 6. Ablation for spatial image cross attention. The values around a point on the curve correspond to the $\\beta$ . While $\\beta = 0$ means the original network without spatial image cross attention." + ], + "image_footnote": [], + "bbox": [ + 516, + 85, + 870, + 242 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3189e3a083c1a4fb2c73c8a3a1cc29cea76cb30815d51297714f7b39831cfff5.jpg", + "table_caption": [ + "Table 4. Ablation for generation process. [91] cannot generate a long video in a single scene and both [67, 91] can't refer to images." + ], + "table_footnote": [], + "table_body": "
TTCAMTP0/16 (↓)1/15 (↓)3/13 (↓)5/11 (↓)
✗✗348.36217.39297.37246.56
✓✗333.63178.27230.56193.52
✓✓257.70123.51118.02109.31
", + "bbox": [ + 527, + 292, + 859, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation for temporal text cross attention and mixed training paradigm. \"TTCA\" and \"MTP\" refer to temporal text cross attention and mixed training paradigm respectively. \"3/13\" denotes predicting the following 13 frames after being given 3 frames of a ground truth video, and we evaluate the FVD between the newly generated 13 frames and the corresponding frames in the ground truth. \"0/16\", \"1/15\", \"5/11\" follow this pattern.", + "bbox": [ + 498, + 362, + 890, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as videographer. We compare three approaches including no autoregression at all as MovieFactory [91], fully autoregressive processes in one go as Phenaki [67], and the generation process of our Vlogger. Specifically, We use GPT-4 to generate five stories and go through the Vlogger's planning process to get five vlog scripts. We set 20 different random seeds, so that each of the above generation processes generates 20 different vlogs for each script. We use CLIP-I, CLIP-T [82] and video duration to evaluate the quality of the generated vlogs. The results demonstrate that, even using the same script and videographer, our Vlogger outshines other existing frameworks for preferable vlog generation. More details can be found in the supplementary doc.", + "bbox": [ + 496, + 472, + 890, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial-Temporal Enhanced Block. Fig. 6 and Tab. 5 evaluate two important operations in our STEB block, i.e., spatial image cross attention and temporal text cross attention. With spatial image cross attention in Fig. 6, the CLIP-I is significantly improved as $\\beta$ increased, while the CLIP-T reached its maximum at $\\beta = 0.5$ on COCO2017 [43]. With temporal text cross attention (TTCA) in Tab. 5, both T2V generation (0/16) and prediction (1/15, 3/13 and 5/11) show a significant improvement on Vineo11k.", + "bbox": [ + 496, + 671, + 890, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixed Training Paradigm. Tab. 5 also presents a model comparison by incorporating our mixed training paradigm or the random mask training method in [11, 14, 21, 36, 68, 84]. The results show that the mixed training paradigm is effective in enhancing the model's T2V generation and prediction capabilities. More details in the supplementary doc.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lots of traffic in futuristic city. ... The camera moves away from the astronaut. The astronaut leaves the keyboard and walks to the left. ... We are in an office room with empty desks. A lion runs on top of the office desks. ... Timelapse of sunset in the modern city.", + "bbox": [ + 96, + 87, + 869, + 112 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a87289cd4d58944740a66d11ac1a26019ba3e80ace55bc9d7c582d7b82fa5b56.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with the state-of-the-art methods on long video generation. The story and long video of Phenaki [67] are available at phenaki.github.io. The scene diversity and picture quality generated by Vlogger are significantly better than Phenaki." + ], + "image_footnote": [], + "bbox": [ + 96, + 113, + 869, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91ddcd3217c96cbde8e784ce97c5f06cd4f4437bf8482916a1b819e1efda8449.jpg", + "image_caption": [ + "Figure 8. Qualitative ablation for STEB and training paradigm. \"TTCA\" and \"MTP\" refer to temporal the text cross attention and mixed training paradigm respectively. The sample without a visual prompt corresponds to the model without spatial image cross attention." + ], + "image_footnote": [], + "bbox": [ + 93, + 311, + 869, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Visualization", + "text_level": 1, + "bbox": [ + 76, + 647, + 212, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we visualize long video generation, by comparison with the well-known Phenaki [67]. Note that, since Phenaki does not have the open-sourced codes, we alternatively use the demo shown in its official website. Specifically, we feed the same story description into our Vlogger. As shown in Fig. 7, our Vlogger shows superior and more diverse video content, compared to Phenaki. Moreover, our Vlogger sufficiently exhibits the story with a much longer duration (i.e., 7min59s), according to our LLM-created script.", + "bbox": [ + 75, + 669, + 468, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further visualize the ablation of T2V video generation and prediction by ShowMaker. As depicted in Fig. 8, ShowMaker had significant improvements in generation and prediction performance, by incorporating our ShowMaker designs. Additionally, it can leverage visual prompts in the spatial actor cross attention, for distinguishing the", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Teddy\" concept within text prompts.", + "bbox": [ + 500, + 647, + 750, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 669, + 617, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a generic system Vlogger, to generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor. Moreover, we present a novel video diffusion model ShowMaker for boosting state-of-the-art T2V generation and prediction. Finally, we will release all the models, data and codes afterward, allowing to develop further designs toward long video generation in the open world.", + "bbox": [ + 496, + 688, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Key R&D Program of China (NO.2022ZD0160102), the National Natural Science Foundation of China under Grant No. 62102150, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Suno AI. bark. https://github.com/suno-ai/bark#-usage-in-cython, 2023.2,4,6", + "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenhang Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, K. Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Yu Bowen, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xing Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. ArXiv, abs/2309.16609, 2023. 3", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 7", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 2, 5", + "[5] Yogesh Balaji1, Martin Renqiang Min, Bing Bai, Rama Chellappa1, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 6", + "[6] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 2, 3", + "[7] Sarthak Bhagat, Shagun Uppal, Vivian Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In ECCV, 2020. 2", + "[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5, 6", + "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A. Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2", + "[10] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 3", + "[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 7", + "[12] Haoran Chen, Jianmin Li, Simone Frintrop, and Xiaolin Hu. The msr-video to text dataset with clean annotations. Comput. Vis. Image Underst., 225:103581, 2021. 6", + "[13] Xinyuan Chen, Chang Xu, Xiaokang Yang, and Dacheng Tao. Long-term video prediction via criticization and retrospection. IEEE TIP, 29:7090-7103, 2020. 2", + "[14] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. ArXiv, abs/2310.20700, 2023. 2, 7", + "[15] François Chollet. On the measure of intelligence. ArXiv, abs/1911.01547, 2019. 2", + "[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. JMLR, 24:240:1-240:113, 2022. 3", + "[17] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv: Computer Vision and Pattern Recognition, 2019. 2", + "[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 2", + "[19] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay P. Namboodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. WACV, pages 3017-3031, 2019. 7", + "[20] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2020. 6", + "[21] Tsu-Jui Fu, Licheng Yu, N. Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 7", + "[22] Wen Gao, Yonghong Tian, Tiejun Huang, and Qiang Yang. Vlogging: A survey of videoblogging technology on the web. ACM Computing Surveys, 2010. 2", + "[23] Songwei Ge, Thomas Hayes, Harry Yang, Xiaoyue Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2", + "[24] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 7", + "[25] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. ArXiv, abs/2305.10474, 2023. 1, 2, 5, 6", + "[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3", + "[27] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. Animate-a-story: Storytelling with retrieval-augmented video generation. ArXiv, abs/2307.06940, 2023. 3", + "[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. ArXiv, abs/2211.13221, 2023. 2", + "[29] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Frederick Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. ArXiv, abs/1712.00409, 2017. 2", + "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6", + "[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 1", + "[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1", + "[33] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1, 2, 5, 7", + "[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. ArXiv, abs/2204.03458, 2022. 5, 7", + "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In ICLR, 2023. 6", + "[36] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. TMLR, 2022, 2022. 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. ArXiv, abs/2309.14494, 2023. 3", + "[38] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.7", + "[39] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Apostol Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. ArXiv, abs/1705.06950, 2017. 6", + "[40] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In NeurIPS, 2022. 2", + "[41] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. ArXiv, abs/2305.06355, 2023. 3", + "[42] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2017. 6", + "[43] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7", + "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2022. 3", + "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 2, 5", + "[46] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 3, 6", + "[47] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 6", + "[48] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 4, 6", + "[49] Haonan Qiu, Menghan Xia, Yong Zhang, Yin-Yin He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv, abs/2310.15169, 2023. 2", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8815", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2, 5", + "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 6, 7", + "[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. ArXiv, abs/1505.04597, 2015. 5", + "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 5", + "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2, 7", + "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2", + "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1, 2, 6", + "[58] Ivan Skorokhodov, S. Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2021. 2", + "[59] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1", + "[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. ArXiv, abs/1212.0402, 2012. 6", + "[61] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, Weixin Liu, Zhihua Wu, Weibao Gong, Jianzhong Liang, Zhizhou Shang, Peng Sun, Wei Liu, Ouyang Xuan, Dianhai Yu, Hao Tian, Hua Wu, and Haifeng Wang. Ernie 3.0: Large-scale knowledge enhanced pretraining for language understanding and generation. ArXiv, abs/2107.02137, 2021. 3", + "[62] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023.3", + "[63] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. Motion-based generator model: Unsupervised disentanglement of appearance," + ], + "bbox": [ + 78, + 92, + 467, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trackable and intractable motions in dynamic patterns. In ICLR, 2019. 2", + "[64] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 2", + "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. *ArXiv*, abs/2302.13971, 2023. 3", + "[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. In ICLR, 2019. 6", + "[67] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 1, 2, 6, 7, 8", + "[68] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 7", + "[69] Fu Lee Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. ArXiv, abs/2305.18264, 2023. 2", + "[70] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. ArXiv, abs/2308.06571, 2023. 1, 2", + "[71] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. ArXiv, abs/2305.10874, 2023. 2, 6", + "[72] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. G3an: Disentangling appearance and motion for video generation. In CVPR, 2020. 2", + "[73] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. Imaginar: Conditional spatiotemporal gan for video generation. WACV, pages 1149-1158, 2020.", + "[74] Yaohui Wang, François Bremond, and Antitza Dantcheva. Inmodegan: Interpretable motion decomposition generative adversarial network for video generation. ArXiv, abs/2101.03049, 2021. 2", + "[75] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. *Internvideo: General video foundation models via generative and discriminative learning.* ArXiv, abs/2212.03191, 2022. 2", + "[76] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Pe der Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Y. Qiao, and Ziwei Liu. Lavie: High-quality video" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8816", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation with cascaded latent diffusion models. ArXiv, abs/2309.15103, 2023. 1, 2, 5, 7", + "[77] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. ArXiv, abs/2104.14806, 2021. 6", + "[78] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6", + "[79] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 3", + "[80] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2021. 2", + "[81] Wilson Yan, Yunzhi Zhang, P. Abbeel, and A. Srinivas. Videogpt: Video generation using vq-vae and transformers. ArXiv, abs/2104.10157, 2021. 2", + "[82] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 7", + "[83] Sheng-Siang Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation. In Annual Meeting of the Association for Computational Linguistics, 2023. 1, 2", + "[84] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit: Masked generative video transformer. In CVPR, 2023. 7", + "[85] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2", + "[86] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, P. Zhang, Yuxiao Dong, and Jie Tang. Glm-130b: An open bilingual pre-trained model. ArXiv, abs/2210.02414, 2022. 3", + "[87] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lin Hao Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. ArXiv, abs/2309.15818, 2023. 1, 2", + "[88] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. ArXiv, abs/2305.18583, 2023. 3", + "[89] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. ArXiv, abs/2211.11018, 2022. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 3", + "[91] Junchen Zhu, Huan Yang, Huiguo He, Wenjing Wang, Zixi Tuo, Wen-Huang Cheng, Lianli Gao, Jingkuan Song, and Jianlong Fu. Moviefactory: Automatic movie creation from text using large generative models for language and images. In ACM MM, 2023. 7" + ], + "bbox": [ + 501, + 92, + 890, + 215 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8817", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_model.json b/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ae9f7a8cd2faa7031b6cab9535b08225377d9c9 --- /dev/null +++ b/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.123, + 0.334, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.338, + 0.14, + 0.699, + 0.162 + ], + "angle": 0, + "content": "Vlogger: Make Your Dream A Vlog" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.199, + 0.723, + 0.236 + ], + "angle": 0, + "content": "Shaobin Zhuang\\(^{1,2}\\) Kunchang Li\\(^{3,4,2}\\) Xinyuan Chen\\(^{2}\\) \nYaohui Wang\\(^{2}\\) Ziwei Liu\\(^{5}\\) Yu Qiao\\(^{2}\\) Yali Wang\\(^{3,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.239, + 0.638, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Jiao Tong University \\(^{2}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.256, + 0.68, + 0.27 + ], + "angle": 0, + "content": "3Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.274, + 0.717, + 0.288 + ], + "angle": 0, + "content": "4University of Chinese Academy of Sciences 5S-Lab, Nanyang Technological University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.324, + 0.313, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.474, + 0.765 + ], + "angle": 0, + "content": "In this work, we present Vlogger, a generic AI system for generating a minute-level video blog (i.e., vlog) of user descriptions. Different from short videos with a few seconds, vlog often contains a complex storyline with diversified scenes, which is challenging for most existing video generation approaches. To break through this bottleneck, our Vlogger smartly leverages Large Language Model (LLM) as Director and decomposes a long video generation task of vlog into four key stages, where we invoke various foundation models to play the critical roles of vlog professionals, including (1) Script, (2) Actor, (3) ShowMaker, and (4) Voicer. With such a design of mimicking human beings, our Vlogger can generate vlogs through explainable cooperation of top-down planning and bottom-up shooting. Moreover, we introduce a novel video diffusion model, ShowMaker, which serves as a videographer in our Vlogger for generating the video snippet of each shooting scene. By incorporating Script and Actor attentively as textual and visual prompts, it can effectively enhance spatial-temporal coherence in the snippet. Besides, we design a concise mixed training paradigm for ShowMaker, boosting its capacity for both T2V generation and prediction. Finally, the extensive experiments show that our method achieves state-of-the-art performance on zero-shot T2V generation and prediction tasks. More importantly, Vlogger can generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.208, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Vlogs represent a unique form of blogging, distinguished by their utilization of video as the primary medium rather than text. Due to more lively expression in the dynamic scenes, blog has become one of the most popular online-sharing" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.324, + 0.943, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Figure 1. Overview of Vlogger. Based on the user story, our Vlogger leverages Large Language Model (LLM) as Director, and decomposes a minute-long vlog generation task into four key stages with Script, Actor, ShowMaker, and Voicer. Moreover, ShowMaker is a novel video diffusion model to generate video snippet of each shooting scene, with script and actor coherence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.75 + ], + "angle": 0, + "content": "ways in the digital world. In the past two years, the remarkable success of diffusion models [31, 32, 59] has shown a great impact on video creation [8, 25, 33, 57, 70, 76, 83, 87]. Hence, there is a natural question, can we build a generic AI system to generate wonderful vlogs automatically?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Regrettably, the majority of current video diffusion approaches mainly generate short videos with a few seconds, by temporal adaptation of image diffusion models [8, 25, 57, 70, 76, 87]. In contrast, a vlog typically constitutes a minute-level long video in the open world. Recently, there have been some attempts in long video generation [67, 83]. However, these early works either require extensive training on large well-captioned long video datasets [83] or suffer from noticeable incoherence of shot changes [67]. Hence, it is still challenging to generate a minute-level" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.442, + 0.901 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Interns at Shanghai AI Laboratory. \\(\\heartsuit\\) Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.409, + 0.107 + ], + "angle": 0, + "content": "vlog with complex narratives and multiple scenes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.109, + 0.468, + 0.215 + ], + "angle": 0, + "content": "Alternatively, we notice that a successful blog production is a systematical work in our realistic world, where the key staffs are involved in script creation, actor design, video shooting, and editing [22]. Drawing inspiration from this, we believe that, generating a long-form video blog requires elaborate systematical planning and shooting processes, rather than simply designing a generative model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.218, + 0.468, + 0.563 + ], + "angle": 0, + "content": "Hence, we propose a generic AI system for vlog generation in this paper, namely Vlogger, which can smartly address this difficult task by mimicking vlog professionals with various foundation models in the core steps. As shown in Fig. 1, we first hire a Large Language Model (LLM) as Director (e.g., GPT-4 [46]), due to its great power of linguistic knowledge understanding. Given a user story, this director schedules a four-step plan for vlog generation. (1) Script. First, we introduce a progressive script creation paradigm for the LLM Director. By the coarse-to-fine instructions in this paradigm, the LLM Director can effectively convert a user story into a script, which sufficiently describes the story by a number of shooting scenes and their corresponding shooting duration. (2) Actor. After creating the script, the LLM Director reads it again to summarize the actors, and then invokes a character designer (e.g., SD-XL [48]) to generate the reference images of these actors in the vlog. (3) ShowMaker. With the guidance of script texts and actor images, we develop a novel ShowMaker as a videographer, which can effectively generate a controllable-duration snippet for each shooting scene, with spatial-temporal coherence. (4) Voicer. Finally, the LLM Director invokes a voicer (e.g., Bark [1]) to dub the vlog with script subtitles." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.567, + 0.468, + 0.778 + ], + "angle": 0, + "content": "It should be noted that our Vlogger overcomes the challenges previously encountered in long video generation tasks. On one hand, it elegantly decomposes the user story into a number of shooting scenes and designs actor images that can participate in different scenes in the vlog. In this case, it can reduce spatial-temporal incoherence of abrupt shot changes, with explicit guidance of scene texts and actor images. On the other hand, Vlogger crafts individual video snippets for every scene and seamlessly integrates them into a single cohesive vlog. Consequently, this bypasses the tedious training process with large-scale long video datasets. Via such collaboration between top-down planning and bottom-up shooting, Vlogger can effectively transform an open-world story into a minute-long vlog." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we would like to emphasize that, ShowMaker is a distinctive video diffusion model designed for generating video snippets of each shooting scene. From the structural perspective, we introduce a novel Spatial-Temporal Enhanced Block (STEB) in this model. This block can adaptively leverage scene descriptions and actor images as textual and visual prompts, which attentively guide ShowMaker to enhance spatial-temporal coherence of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "script and actors. From the training perspective, we develop a probabilistic mode selection mechanism, which can boost the capacity of ShowMaker by mixed training of Text-to-Video (T2V) generation and prediction. More notably, by sequential combination of generation and prediction mode in the inference stage, ShowMaker can produce a video snippet with a controllable duration. This allows our Vlogger to generate a blog with a preferable duration, according to the planning of each scene in the script by LLM director." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Finally, our method achieves the state-of-the-art performance on both zero-shot T2V generation and prediction, by expensive experiments within the popular video benchmarks. More importantly, our Vlogger outperforms the well-known long video generation method, i.e., Phenaki [67], despite utilizing only \\(66.7\\%\\) of the training videos. Remarkably, our Vlogger is capable of generating over 5-minute vlogs, without loss of script and actor coherence in the video. The models and codes will be released afterward." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.65, + 0.393 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Text-to-Video Generation. Distinct from conventional unconditional and class-conditional video generation [7, 9, 13, 17, 23, 58, 63, 64, 72-74, 81, 85], T2V generation focuses on automatically converting textual descriptions into videos. This is a challenging task, as it involves understanding text semantics and translating it into video content. This often requires powerful cross-modal algorithms [50], large computing resources [15, 29], and extensive video data [3, 55, 56, 75, 76, 80]. Based on the success of diffusion models in the Text-to-Image (T2I) generation [4, 6, 18, 40, 45, 51, 52, 54], a series of such works have been recently transferred to T2V generation [8, 25, 33, 57, 70, 71, 76, 87]. However, whether trained from scratch [33, 57] or finetuned from T2I model [8, 25, 70, 71, 76, 87], most of these approaches mainly work on generating short videos with few seconds from simple descriptions. In contrast, our Vlogger can generate a minute-long vlog with complex stories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Long Video Generation. The generation of long videos predominantly relies on parallel [83] or autoregressive [67] structures. However, these early works still face challenges for vlog generation. On the one hand, the parallel manner can relax spatial-temporal content incoherence problems by coarse-to-fine generation. However, this approach necessitates extensive and laborious training on large, well-annotated long video datasets [83]. On the other hand, the autoregressive manner can relax heavy data requirements for training long videos, by applying short video generation models iteratively with sliding windows. However, this solution often suffers from noticeable shot change and long-term incoherence [14, 28, 49, 67, 69] which becomes problematic when generating vlogs encompassing complex narratives and multiple scenes. It is worth mentioning that the community has gradually realized that delegating higher" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.893, + 0.372 + ], + "angle": 0, + "content": "Figure 2. Top-Down Planning. Through four rounds of dialogue with the LLM, we gradually convert the user story into the final script. Based on this script, we further extract actor reference images, and then determine which actor would star in each script scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.473 + ], + "angle": 0, + "content": "order reasoning tasks to LLM is of great help to visual tasks [6, 26, 27, 37, 41, 44, 79, 88, 90]. Our Vlogger introduces LLM to the field of long video generation, and effectively addresses the problems of training burden and content incoherence in the previous methods, by distinct cooperation of top-down planning and bottom-up shooting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.17, + 0.503 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.59 + ], + "angle": 0, + "content": "In this section, we introduce our Vlogger framework in detail. First, we describe how to make a vlog by planning and shooting with our Vlogger. Then, we further explain the novel design of ShowMaker. As the videographer in our Vlogger, it is critical for video generation in the vlog." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.353, + 0.618 + ], + "angle": 0, + "content": "3.1. Overall Framework of Vlogger" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.763 + ], + "angle": 0, + "content": "To generate a minute-level vlog, our Vlogger leverages LLM as the director, which can effectively decompose this generation task by four key roles within the planning and shooting stages. As shown in Fig. 1, the LLM Director first creates Script and designs Actor in the planning stage. Based on Script and Actor, ShowMaker generates a video snippet for each scene in the shooting stage, and Voicer dubs subtitles of this snippet. Finally, one can combine the snippets of all the scenes as a vlog." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.273, + 0.801 + ], + "angle": 0, + "content": "3.1.1 Top-Down Planning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "A vlog often comes from a user story that contains diversified content within many shot changes. Clearly, it is challenging to produce a minute-long vlog, by directly feeding such a complex story into video generation models. To this end, we propose to leverage LLM as director and decompose the user story by top-down planning in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.893, + 0.532 + ], + "angle": 0, + "content": "Script Creating. First, we parse the use story into a script, which describes this story explicitly by a number of shooting scenes. In this case, we can generate a video snippet for each shooting scene, instead of learning a long video tediously from the entire story. Since LLM has shown an impressive capacity in language understanding [2, 10, 16, 46, 61, 62, 65, 86], we feed the user story into such a director for script generation. As shown in Fig. 2, we introduce a progressive creation paradigm, which can effectively parse the story by coarse-to-fine steps," + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.538, + 0.892, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {(i)} = \\operatorname {L L M} \\left(\\mathcal {S} ^ {(i - 1)}, \\mathcal {I} _ {S} ^ {(i)}, \\mathcal {U}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Given the user story \\(\\mathcal{U}\\) and creation instruction \\(\\mathcal{I}_S^{(i)}\\), the LLM Director generates the current script \\(S^{(i)}\\) from the previous one \\(S^{(i-1)}\\). More specifically, there are four steps including (1) Rough. First, LLM generates a basic draft of the script from the story. (2) Detailed. Then, LLM refines the rough script with story details. (3) Completed. Next, LLM checks if the detailed script misses the important parts of the story. (4) Scheduled. Finally, LLM allocates a shooting duration for each scene in the completed script, according to the scene content. For convenience, we denote the final script as \\(\\mathcal{S}\\) in the following. It contains the descriptions of \\(N\\) shooting scenes \\(\\{S_1, \\dots, S_N\\}\\) and their allocated duration \\(\\{\\mathcal{T}_1, \\dots, \\mathcal{T}_N\\}\\). Additionally, due to the limited pages, please read the full descriptions of instructions and script in the supplementary doc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Actor Designing. After generating the script, it is time to design actors in the vlog. As shown in Fig. 2, we ask the LLM Director to summarize the actor list \\(\\mathcal{A}\\) from script \\(S\\)," + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.847, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = \\operatorname {L L M} \\left(\\mathcal {S}, \\mathcal {I} _ {A}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_A\\) is the instruction for actor summarization. Then, according to actor descriptions, the LLM Director invokes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.873, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 3. Bottom-Up Shooting. For each scene, ShowMaker can generate the video snippet with script and actor coherence, by using script description and actor image as textual and visual prompts. Moreover, ShowMaker can effectively control the snippet duration, by performing generation and prediction sequentially in the inference. Finally, we apply a Text-To-Speech (TTS) model as Voicer for dubbing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.411, + 0.462, + 0.427 + ], + "angle": 0, + "content": "a designer to generate reference images of these actors \\(\\mathcal{R}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.44, + 0.47, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} = \\text {S t a b l e - D i f f u s i o n} (\\mathcal {A}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.53 + ], + "angle": 0, + "content": "We choose Stable Diffusion XL [48] as the character designer, due to its high-quality generation. Finally, based on script \\( S \\) and actor \\( \\mathcal{A} \\), the LLM Director decides the leading actor (i.e., protagonist) in each shooting scene of the script," + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.544, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} = \\operatorname {L L M} (\\mathcal {S}, \\mathcal {A}, \\mathcal {I} _ {P}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.47, + 0.63 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_P\\) is the instruction for protagonist selection. The resulting doc \\(\\mathcal{P}\\) is aligned with the script \\(S\\), where \\(\\mathcal{P}_n\\) describes which actor appears in the scene \\(S_n\\). For example, \\(\\{\\mathcal{S}_6\\) is starred by \\(\\mathcal{A}_1\\}\\) in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.277, + 0.667 + ], + "angle": 0, + "content": "3.1.2 Bottom-Up Shooting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Via the top-down planning above, the LLM Director flexibly decomposes a complex user story into several script scenes and designs actor reference image for each scene. Such a manner largely reduces the difficulty of vlog generation, since we can generate vlogs by bottom-up shooting, i.e., we just need to generate the video snippet for each shooting scene and combine all of them as a vlog." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "ShowMaker Shooting. To generate the video snippet of a shooting scene, we introduce a novel ShowMaker as the videographer, which is a video diffusion model with two distinct designs. First, it is important to maintain spatial-temporal coherence of both script and actor in the generated snippet. Hence, our ShowMaker not only takes the scene description \\( S_{n} \\) as a textual prompt but also takes actor image of this scene \\( \\mathcal{R}_n \\) as a visual prompt. Second, each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.895, + 0.502 + ], + "angle": 0, + "content": "scene is allocated with a shooting duration in the script. To control the duration of each scene, our ShowMaker contains two learning modes including generation and prediction in Fig.3. Specifically, it starts with the generation mode. For the shooting scene \\( n \\), we feed its script description \\( S_{n} \\) and actor reference image \\( \\mathcal{R}_n \\) into ShowMaker," + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.512, + 0.892, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {n} ^ {(1)} = \\operatorname {S h o w M a k e r} \\left(\\mathcal {N} _ {n} ^ {(1)} \\mid S _ {n}, \\mathcal {R} _ {n}, \\text {G e n e r a t e}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.894, + 0.645 + ], + "angle": 0, + "content": "which generates the first video clip of this scene \\(\\mathcal{C}_n^{(1)}\\) from the noisy clip \\(\\mathcal{N}_n^{(1)}\\). If the duration of this clip is smaller than the allocated duration \\(\\mathcal{T}_n\\) in the script, we continue to perform the prediction mode, i.e., the last \\(k\\) frames of the current clip \\(\\mathcal{C}_n^{(j)}(k)\\) are used as context, when generating the next clip from the noisy input \\(\\mathcal{N}_n^{(j + 1)}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.654, + 0.892, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {n} ^ {(j + 1)} = \\operatorname {S h o w M a k e r} \\left(\\mathcal {N} _ {n} ^ {(j + 1)} \\mid \\mathcal {S} _ {n}, \\mathcal {C} _ {n} ^ {(j)} (k), \\text {P r e d i c t}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Note that, actor reference images are not necessary in the prediction mode, since such an actor's appearance has been shown in the current clip \\(\\mathcal{C}_n^{(j)}(k)\\) that is used as input for prediction. This prediction procedure stops until the total duration achieves the allocated \\(\\mathcal{T}_n\\) of this scene. Subsequently, we combine all the clips as the video snippet of this scene, i.e., \\(\\mathcal{C}_n = \\{\\mathcal{C}_n^{(1)},\\dots,\\mathcal{C}_n^{(J)}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Voiceer Speaking. To enhance the completeness of the vlog, we apply a Text-To-Speech model (e.g., Bark [1]) as a Voiceer, which converts the scene description \\( S_{n} \\) into the corresponding audio \\( \\mathcal{O}_n = \\mathrm{Bark}(S_n) \\). Finally, we add this audio to the corresponding video snippet \\( \\mathcal{V}_n = \\mathcal{O}_n \\oplus \\mathcal{C}_n \\), and combine all the sounded video snippets as a complete vlog, i.e., \\( \\mathcal{V} = \\{\\mathcal{V}_1, \\dots, \\mathcal{V}_N\\} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.46, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.306, + 0.321, + 0.319 + ], + "angle": 0, + "content": "(a) Model Architecture" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.089, + 0.877, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.206, + 0.828, + 0.22 + ], + "angle": 0, + "content": "(b) Spatial-Temporal Enhanced Block (STEB)" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.231, + 0.88, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.306, + 0.76, + 0.318 + ], + "angle": 0, + "content": "(c) Mode Selection" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.33, + 0.895, + 0.372 + ], + "angle": 0, + "content": "Figure 4. ShowMaker. (a) The Overall Architecture. (b) Spatial-Temporal Enhanced Block (STEB). Via spatial-actor and temporal-text cross attention, our STEB can further enhance actor and script coherence in the snippet. (c) Mode Selection. We introduce a mixed training paradigm of T2V generation and prediction, via probabilistic selection of masked frames." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.38, + 0.208, + 0.395 + ], + "angle": 0, + "content": "3.2. ShowMaker" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.403, + 0.47, + 0.599 + ], + "angle": 0, + "content": "As discussed in Section 3.1.2, ShowMaker plays a critical role in generating video snippet of a shooting scene. In this work, we introduce a text-to-video diffusion model for it. It follows the style of latent diffusion model [52]. In the diffusing stage, we add Gaussian noise progressively to the latent code of a training snippet. In the denoising stage, we reconstruct the latent code from the noisy latent code at any iteration step. For simplicity, we just show the denoising stage in Fig.4 (a). First, we forward a noisy training snippet into the encoder of the autoencoder to extract its latent code. Then, we feed this into a denoising U-Net [53] to learn the clean latent code. Finally, we leverage the decoder to reconstruct the original snippet with the clean latent code." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.676 + ], + "angle": 0, + "content": "But, compared with the existing video diffusion models [8, 25, 33, 34, 76], our novel ShowMaker contains two distinct designs, in terms of model structure (i.e., Spatial-Temporal Enhanced Block) in Fig.4 (b), and training paradigm (i.e., Mode Selection) in Fig.4 (c)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.428, + 0.708 + ], + "angle": 0, + "content": "3.2.1 Spatial-Temporal Enhanced Block (STEB)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.837 + ], + "angle": 0, + "content": "To reconstruct the clean latent code of a training video snippet, each block in the denoising U-Net consists of both spatial and temporal operations in previous works [76]. First, spatial operations encode the feature of each frame separately in the snippet. Typically, three operations are inherited from text-to-image generation approaches [4, 45, 51, 52, 54], including spatial ConVolution (CV), spatial Self Attention (SA), and spatial Cross Attention (CA)," + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.843, + 0.469, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {c v} = \\operatorname {C V - S p a t i a l} \\left(\\mathcal {X} _ {i n}, \\mathcal {F} _ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.863, + 0.469, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {s a} = \\operatorname {S A} - \\operatorname {S p a t i a l} \\left(\\mathcal {X} _ {c v}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.882, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {c a} = \\operatorname {C A} - \\text {S p a t i a l - T e x t} \\left(\\mathcal {X} _ {s a}, \\mathcal {S} _ {n}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.381, + 0.892, + 0.442 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{in}\\) is the noisy feature of the training snippet, and \\(\\mathcal{F}_t\\) is the positional embedding of the iteration step \\(t\\). To guide denoising with the given text, we use the scene description \\(S_{n}\\) as Key and Value of cross attention in Eq. (9)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.503 + ], + "angle": 0, + "content": "However, we notice that such spatial encoding does not consider actors. Hence, it inevitably suffers from actor incoherence when generating a snippet. To tackle this problem, we introduce a spatial image cross attention," + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.519, + 0.892, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} _ {c a} = \\operatorname {C A} - \\text {S p a t i a l - A c t o r} \\left(\\mathcal {X} _ {s a}, \\mathcal {R} _ {n}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.641 + ], + "angle": 0, + "content": "For a shooting scene \\( S_{n} \\), we leverage the protagonist doc (Eq. 4) to find the corresponding actor in this scene. Then, we leverage actor reference image \\( \\mathcal{R}_{n} \\) as the visual context of spatial cross attention. Subsequently, we enhance spatial embedding with the complementary guidance of both script and actor, i.e., \\( \\mathcal{Z}_{se} = \\mathcal{X}_{ca} + \\beta \\mathcal{Y}_{ca} \\) with a weight \\( \\beta \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.687 + ], + "angle": 0, + "content": "After spatial encoding, it is time to learn the correlation across frames in the snippet. Hence, the typical operation is to perform self attention along the temporal dimension," + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.703, + 0.891, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} _ {s a} = \\operatorname {S A - T e m p o r a l} \\left(\\mathcal {Z} _ {s e}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.794 + ], + "angle": 0, + "content": "However, such a temporal operation does not take the constraints of scene text into account. It often leads to text incoherence when generating video snippets. Hence, we introduce a temporal text cross attention," + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.81, + 0.891, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} _ {c a} = \\operatorname {C A} - \\text {T e m p o r a l - T e x t} \\left(\\mathcal {Z} _ {s a}, S _ {n}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where we leverage the scene description \\( S_{n} \\) as the textual context of temporal encoding. Via spatial-actor (Eq. 10) and temporal-text (Eq. 12) cross attention, our STEB can further enhance actor and script coherence in the snippet." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.268, + 0.167 + ], + "angle": 0, + "content": "
MethodFVD (↓)
VideoFactory [71]410.00
Make-A-Video [57]367.23
PYoCo [25]355.19
Ours292.43
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.108, + 0.176, + 0.243, + 0.189 + ], + "angle": 0, + "content": "(a) Hand-crafted prompt." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.089, + 0.472, + 0.176 + ], + "angle": 0, + "content": "
MethodFVD (↓)
CogVideo (Chinese) [35]751.34
CogVideo (English) [35]701.59
MagicVideo [89]699.00
Video LDM [8]550.61
Ours525.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.328, + 0.178, + 0.41, + 0.189 + ], + "angle": 0, + "content": "(b) Class label." + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.191, + 0.47, + 0.219 + ], + "angle": 0, + "content": "Table 1. Zero-shot comparison with the state-of-the-art methods on UCF-101. Hand-crafted prompt from [25]." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.46, + 0.24 + ], + "angle": 0, + "content": "3.2.2 Mixed Training Paradigm with Mode Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.469, + 0.324 + ], + "angle": 0, + "content": "As discussed in Section 3.1.2, ShowMaker aims to generate a video snippet with the allocated duration in the script. To this goal, it leverages the combination of both generation and prediction modes in the inference stage. Next, we explain how to train our ShowMaker to learn these modes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.469, + 0.461 + ], + "angle": 0, + "content": "As shown in Fig.4 (c), we design a mode selection mechanism, which selects \\( k \\) frames of the clean snippet as the context of the noisy snippet. The \\( k = 0 \\) setting refers to the generation mode since there is no context from the clean snippet. Alternatively, the \\( k > 0 \\) setting refers to the prediction mode, since \\( k \\) frames of the clean snippet are already available. The goal is to generate the rest frames of the clean snippet from the noisy one. To integrate both modes into training, we design a probabilistic manner to select \\( k \\)," + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.484, + 0.469, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} (k) = \\left\\{ \\begin{array}{c c} \\alpha^ {k} - \\alpha^ {k + 1}, & k \\in [ 0, m) \\\\ \\alpha^ {k}, & k = m \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.597 + ], + "angle": 0, + "content": "where \\(\\mathrm{P}(k)\\) is the selection probability distribution of \\(k\\). Moreover, \\(0 < \\alpha < 1\\) and \\(0 \\leq m\\) are the manual parameters, which respectively control the mode selection tendencies of \\(\\mathrm{P}(k)\\) and the maximum number of preserved frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.629 + ], + "angle": 0, + "content": "After determining \\(k\\), we introduce a frame mask \\(\\mathcal{M}_k\\) on the latent code of the clean snippet \\(\\mathcal{X}_{clean}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.64, + 0.469, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {k} = \\mathcal {X} _ {\\text {c l e a n}} \\odot \\mathcal {M} _ {k}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.715 + ], + "angle": 0, + "content": "where we only preserve \\(k\\) frames and mask the rest of the frames. Then, we concatenate the mask \\(\\mathcal{M}_k\\) with the latent code of snippet \\(\\mathcal{X}_{noise}\\) and \\(\\mathcal{X}_k\\)," + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.727, + 0.469, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {i n} = \\operatorname {C o n c a t} \\left(\\mathcal {X} _ {\\text {n o i s e}}, \\mathcal {X} _ {k}, \\mathcal {M} _ {k}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.469, + 0.816 + ], + "angle": 0, + "content": "This produces the input feature \\(\\mathcal{X}_{in}\\) for training the denoising U-Net in Section 3.2.1. Via such a concise probabilistic manner, we can effectively integrate both generation \\((k = 0)\\) and prediction \\((k > 0)\\) modes in the training procedure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.21, + 0.847 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Datasets. To make state-of-the-art comparison, We conduct zero-shot evaluation on the popular video benchmarks, i.e., UCF-101 [60], Kinetics-400 [39], and MSR-VTT [12]." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.889, + 0.196 + ], + "angle": 0, + "content": "
MethodZero-ShotPre-training VideosFID (↓)
T2V [42]XX82.13
SC [5]XX33.51
TFGAN [5]XX31.76
NUWA [78]X0.97M28.46
Phenaki [67]✓15M37.74
Ours✓10M37.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.199, + 0.892, + 0.241 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art methods on Kinetics-400. Both under a zero-shot setting, our FID is better than that of Phenaki, with only \\(66.7\\%\\) pre-training videos." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.243, + 0.89, + 0.376 + ], + "angle": 0, + "content": "
MethodZero-ShotCLIPSIM (↑)CLIP-FID (↓)
GODIVA [77]✘0.2402-
NüWA [78]✘0.243947.68
CogVideo (Chinese) [35]✓0.261424.78
CogVideo (English) [35]✓0.263123.59
Video LDM [8]✓0.2929-
Make-A-Video [57]✓0.304913.17
PYoCo [25]✓-10.21
Ours✓0.290812.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Table 3. Comparison with the state-of-the-art methods on MSR-VTT. PYoCo [25] sample 59,794 videos for CLIP-FID (noted in gray) while others only sample less than 3,000 videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.689 + ], + "angle": 0, + "content": "UCF-101 contains videos of 101 action categories. Following [66], we use FVD to evaluate the distance between the generated video and the real video. Kinetics-400 is a dataset comprising videos of 400 action categories. Following [5, 42, 67, 78], we use FID [30] to assess the performance of video generation. MSR-VTT is a video dataset with open-vocabulary captions, wherein CLIPSIM [77] and CLIP-FID [47] are commonly employed for evaluating the T2V generation. Moreover, since these existing benchmarks either have a small number of testing videos or contain only action labels without complex descriptions, we propose to collect an evaluation benchmark for ablation studies. It is called as Vimeo11k, where we collect 11,293 open-world videos along with their captions from 10 mainstream categories in Vimeo. To our best knowledge, it is the largest testing benchmark for zero-shot video generation. More details and experiments can be found in the supplementary doc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. (1) Director & Script & Actor & Voicer. We choose GPT-4 [46] as our LLM director to generate script. The specific instructions can be found in the supplementary doc. Then we employ Stable Diffusion XL [48] and Bark [1] as our designer and voicer to generate reference actor images and convert scripts into speech. (2) ShowMaker. We choose SD-1.4 [52] as our base model, and follow [57] to add temporal self attention. Then, we add our temporal text cross attention on top of each temporal self attention. We expand the input channel of the conv-in layer of U-Net from 4 to 9, so that model can take the concatenated feature in Eq. (15) as input. We use zero initialization for newly added channels. We use CLIP ViT-L/14 [50] as text encoder \\(\\varepsilon_{T}\\), VQVAE [20] as autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.087, + 0.447, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.246, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Figure 5. Long video generation. Comparison on UCF-101 (1000 frames). The lower FVD, the better generation performance." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.276, + 0.468, + 0.338 + ], + "angle": 0, + "content": "
MethodCLIP-I (↑)CLIP-T (↑)Vlog Duration (↑)
w/o autoregressive [91]0.56830.27521min03s
only autoregressive [67]0.56420.25353min58s
Ours0.62940.27563min58s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.343, + 0.468, + 0.371 + ], + "angle": 0, + "content": "Table 4. Ablation for generation process. [91] cannot generate a long video in a single scene and both [67, 91] can't refer to images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.468, + 0.52 + ], + "angle": 0, + "content": "consisting of \\(\\varepsilon\\) and \\(\\mathcal{D}\\). Besides, we use OpenCLIP ViT-H/14 [38] as image encoder \\(\\varepsilon_{I}\\), and add spatial image cross attention in the STEB block. The diffusion step \\(T\\) is set to 1000 as [52]. For training, we set the parameter of mode selection as \\(\\alpha = 0.6\\) and \\(m = 6\\) in Eq. (13), and \\(\\beta\\) is set to 1. Following [19, 33, 34, 67, 76], we employ publicly available image dataset (i.e., Laion400M [55]) and video dataset (i.e., WebVid10M [3]) for joint training. More training details can be found in the supplementary doc." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.529, + 0.367, + 0.545 + ], + "angle": 0, + "content": "4.1. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.468, + 0.702 + ], + "angle": 0, + "content": "Tab. 1 shows that, no matter whether the input text is the class label or hand-crafted prompt, our method achieves the best FVD performance of zero-shot video generation on UCF-101. Tab. 2 shows that, compared to Phenaki [67], our method achieves a better FID performance of zero-shot setting on Kinetics-400, but only using \\(66.7\\%\\) training videos. Furthermore, our generated videos have a resolution of \\(320 \\times 512\\), which is higher than Phenaki's \\(256 \\times 256\\). Tab. 3 shows that, our method achieves remarkably competitive FID and CLIPSIM performance on MSR-VTT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Furthermore, Fig. 5 illustrates that, we significantly surpass TATS [24] (i.e., the only open-source long video generation model within our knowledge), for generating 1000-frame videos on UCF-101. Moreover, our method does not encounter the issue of TATS, where the video quality continuously declines as the number of frames increases. It is noteworthy that, our method achieves this performance by zero-shot generation, without any finetuning on UCF-101." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.229, + 0.849 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Vlog Generation Process. Tab. 4 presents a comparison between different generation processes. For a fair comparison, we use the same scripts and employ our ShowMaker" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.087, + 0.871, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.245, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Figure 6. Ablation for spatial image cross attention. The values around a point on the curve correspond to the \\(\\beta\\). While \\(\\beta = 0\\) means the original network without spatial image cross attention." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.294, + 0.861, + 0.357 + ], + "angle": 0, + "content": "
TTCAMTP0/16 (↓)1/15 (↓)3/13 (↓)5/11 (↓)
✗✗348.36217.39297.37246.56
✓✗333.63178.27230.56193.52
✓✓257.70123.51118.02109.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.363, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Table 5. Ablation for temporal text cross attention and mixed training paradigm. \"TTCA\" and \"MTP\" refer to temporal text cross attention and mixed training paradigm respectively. \"3/13\" denotes predicting the following 13 frames after being given 3 frames of a ground truth video, and we evaluate the FVD between the newly generated 13 frames and the corresponding frames in the ground truth. \"0/16\", \"1/15\", \"5/11\" follow this pattern." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.67 + ], + "angle": 0, + "content": "as videographer. We compare three approaches including no autoregression at all as MovieFactory [91], fully autoregressive processes in one go as Phenaki [67], and the generation process of our Vlogger. Specifically, We use GPT-4 to generate five stories and go through the Vlogger's planning process to get five vlog scripts. We set 20 different random seeds, so that each of the above generation processes generates 20 different vlogs for each script. We use CLIP-I, CLIP-T [82] and video duration to evaluate the quality of the generated vlogs. The results demonstrate that, even using the same script and videographer, our Vlogger outshines other existing frameworks for preferable vlog generation. More details can be found in the supplementary doc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.808 + ], + "angle": 0, + "content": "Spatial-Temporal Enhanced Block. Fig. 6 and Tab. 5 evaluate two important operations in our STEB block, i.e., spatial image cross attention and temporal text cross attention. With spatial image cross attention in Fig. 6, the CLIP-I is significantly improved as \\(\\beta\\) increased, while the CLIP-T reached its maximum at \\(\\beta = 0.5\\) on COCO2017 [43]. With temporal text cross attention (TTCA) in Tab. 5, both T2V generation (0/16) and prediction (1/15, 3/13 and 5/11) show a significant improvement on Vineo11k." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Mixed Training Paradigm. Tab. 5 also presents a model comparison by incorporating our mixed training paradigm or the random mask training method in [11, 14, 21, 36, 68, 84]. The results show that the mixed training paradigm is effective in enhancing the model's T2V generation and prediction capabilities. More details in the supplementary doc." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8812" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.098, + 0.088, + 0.87, + 0.113 + ], + "angle": 0, + "content": "Lots of traffic in futuristic city. ... The camera moves away from the astronaut. The astronaut leaves the keyboard and walks to the left. ... We are in an office room with empty desks. A lion runs on top of the office desks. ... Timelapse of sunset in the modern city." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.114, + 0.87, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.28, + 0.892, + 0.309 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with the state-of-the-art methods on long video generation. The story and long video of Phenaki [67] are available at phenaki.github.io. The scene diversity and picture quality generated by Vlogger are significantly better than Phenaki." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.313, + 0.87, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.608, + 0.893, + 0.637 + ], + "angle": 0, + "content": "Figure 8. Qualitative ablation for STEB and training paradigm. \"TTCA\" and \"MTP\" refer to temporal the text cross attention and mixed training paradigm respectively. The sample without a visual prompt corresponds to the model without spatial image cross attention." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.214, + 0.663 + ], + "angle": 0, + "content": "4.3. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.806 + ], + "angle": 0, + "content": "First, we visualize long video generation, by comparison with the well-known Phenaki [67]. Note that, since Phenaki does not have the open-sourced codes, we alternatively use the demo shown in its official website. Specifically, we feed the same story description into our Vlogger. As shown in Fig. 7, our Vlogger shows superior and more diverse video content, compared to Phenaki. Moreover, our Vlogger sufficiently exhibits the story with a much longer duration (i.e., 7min59s), according to our LLM-created script." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We further visualize the ablation of T2V video generation and prediction by ShowMaker. As depicted in Fig. 8, ShowMaker had significant improvements in generation and prediction performance, by incorporating our ShowMaker designs. Additionally, it can leverage visual prompts in the spatial actor cross attention, for distinguishing the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.648, + 0.75, + 0.665 + ], + "angle": 0, + "content": "\"Teddy\" concept within text prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.67, + 0.619, + 0.685 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.81 + ], + "angle": 0, + "content": "In this paper, we introduce a generic system Vlogger, to generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor. Moreover, we present a novel video diffusion model ShowMaker for boosting state-of-the-art T2V generation and prediction. Finally, we will release all the models, data and codes afterward, allowing to develop further designs toward long video generation in the open world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Key R&D Program of China (NO.2022ZD0160102), the National Natural Science Foundation of China under Grant No. 62102150, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Suno AI. bark. https://github.com/suno-ai/bark#-usage-in-cython, 2023.2,4,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenhang Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, K. Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Yu Bowen, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xing Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. ArXiv, abs/2309.16609, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.468, + 0.354 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.468, + 0.44 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.443, + 0.468, + 0.497 + ], + "angle": 0, + "content": "[5] Yogesh Balaji1, Martin Renqiang Min, Bing Bai, Rama Chellappa1, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[6] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.468, + 0.627 + ], + "angle": 0, + "content": "[7] Sarthak Bhagat, Shagun Uppal, Vivian Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.468, + 0.685 + ], + "angle": 0, + "content": "[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.688, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A. Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.883 + ], + "angle": 0, + "content": "[10] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[12] Haoran Chen, Jianmin Li, Simone Frintrop, and Xiaolin Hu. The msr-video to text dataset with clean annotations. Comput. Vis. Image Underst., 225:103581, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[13] Xinyuan Chen, Chang Xu, Xiaokang Yang, and Dacheng Tao. Long-term video prediction via criticization and retrospection. IEEE TIP, 29:7090-7103, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[14] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. ArXiv, abs/2310.20700, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[15] François Chollet. On the measure of intelligence. ArXiv, abs/1911.01547, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. JMLR, 24:240:1-240:113, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[17] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv: Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[19] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay P. Namboodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. WACV, pages 3017-3031, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[20] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Tsu-Jui Fu, Licheng Yu, N. Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[22] Wen Gao, Yonghong Tian, Tiejun Huang, and Qiang Yang. Vlogging: A survey of videoblogging technology on the web. ACM Computing Surveys, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.166, + 0.47, + 0.221 + ], + "angle": 0, + "content": "[23] Songwei Ge, Thomas Hayes, Harry Yang, Xiaoyue Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[24] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[25] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. ArXiv, abs/2305.10474, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[27] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. Animate-a-story: Storytelling with retrieval-augmented video generation. ArXiv, abs/2307.06940, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. ArXiv, abs/2211.13221, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[29] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Frederick Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. ArXiv, abs/1712.00409, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.583, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[33] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. ArXiv, abs/2204.03458, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In ICLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. TMLR, 2022, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[37] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. ArXiv, abs/2309.14494, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[38] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[39] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Apostol Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. ArXiv, abs/1705.06950, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[40] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[41] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. ArXiv, abs/2305.06355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[42] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[46] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[47] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[48] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[49] Haonan Qiu, Menghan Xia, Yong Zhang, Yin-Yin He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv, abs/2310.15169, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.468, + 0.188 + ], + "angle": 0, + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. ArXiv, abs/1505.04597, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.468, + 0.328 + ], + "angle": 0, + "content": "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.468, + 0.398 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.4, + 0.468, + 0.496 + ], + "angle": 0, + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.468, + 0.565 + ], + "angle": 0, + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.567, + 0.468, + 0.608 + ], + "angle": 0, + "content": "[58] Ivan Skorokhodov, S. Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.468, + 0.663 + ], + "angle": 0, + "content": "[59] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.468, + 0.705 + ], + "angle": 0, + "content": "[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. ArXiv, abs/1212.0402, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.707, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[61] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, Weixin Liu, Zhihua Wu, Weibao Gong, Jianzhong Liang, Zhizhou Shang, Peng Sun, Wei Liu, Ouyang Xuan, Dianhai Yu, Hao Tian, Hua Wu, and Haifeng Wang. Ernie 3.0: Large-scale knowledge enhanced pretraining for language understanding and generation. ArXiv, abs/2107.02137, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.468, + 0.859 + ], + "angle": 0, + "content": "[62] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[63] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. Motion-based generator model: Unsupervised disentanglement of appearance," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "trackable and intractable motions in dynamic patterns. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[64] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. *ArXiv*, abs/2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[67] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[68] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[69] Fu Lee Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. ArXiv, abs/2305.18264, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[70] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. ArXiv, abs/2308.06571, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[71] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. ArXiv, abs/2305.10874, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[72] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. G3an: Disentangling appearance and motion for video generation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[73] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. Imaginar: Conditional spatiotemporal gan for video generation. WACV, pages 1149-1158, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[74] Yaohui Wang, François Bremond, and Antitza Dantcheva. Inmodegan: Interpretable motion decomposition generative adversarial network for video generation. ArXiv, abs/2101.03049, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[75] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. *Internvideo: General video foundation models via generative and discriminative learning.* ArXiv, abs/2212.03191, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[76] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Pe der Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Y. Qiao, and Ziwei Liu. Lavie: High-quality video" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8816" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "generation with cascaded latent diffusion models. ArXiv, abs/2309.15103, 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[77] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. ArXiv, abs/2104.14806, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[78] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[79] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[80] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.468, + 0.39 + ], + "angle": 0, + "content": "[81] Wilson Yan, Yunzhi Zhang, P. Abbeel, and A. Srinivas. Videogpt: Video generation using vq-vae and transformers. ArXiv, abs/2104.10157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[82] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.436, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[83] Sheng-Siang Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation. In Annual Meeting of the Association for Computational Linguistics, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[84] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit: Masked generative video transformer. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[85] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[86] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, P. Zhang, Yuxiao Dong, and Jie Tang. Glm-130b: An open bilingual pre-trained model. ArXiv, abs/2210.02414, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[87] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lin Hao Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. ArXiv, abs/2309.15818, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[88] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. ArXiv, abs/2305.18583, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[89] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. ArXiv, abs/2211.11018, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[90] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[91] Junchen Zhu, Huan Yang, Huiguo He, Wenjing Wang, Zixi Tuo, Wen-Huang Cheng, Lianli Gao, Jingkuan Song, and Jianlong Fu. Moviefactory: Automatic movie creation from text using large generative models for language and images. In ACM MM, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8817" + } + ] +] \ No newline at end of file diff --git a/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_origin.pdf b/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d64bb41eac30db3328d151cffcb95910f44f125b --- /dev/null +++ b/2024/Vlogger_ Make Your Dream A Vlog/3067237f-4763-4f29-bd3f-1f0c7a112ee8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66c0bb5d227b0b90ab1c6981e24ed23875667eb6a9f39b8ad8f2c82a9ee2f816 +size 3657374 diff --git a/2024/Vlogger_ Make Your Dream A Vlog/full.md b/2024/Vlogger_ Make Your Dream A Vlog/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5348c564a263536ac0257d3ff98de398f7a40f7c --- /dev/null +++ b/2024/Vlogger_ Make Your Dream A Vlog/full.md @@ -0,0 +1,385 @@ +![](images/afb986863174c39bf308b50146f3c2bf17f125c1acd1db62eb0d9a84d3631dc9.jpg) + +# Vlogger: Make Your Dream A Vlog + +Shaobin Zhuang $^{1,2}$ Kunchang Li $^{3,4,2}$ Xinyuan Chen $^{2}$ +Yaohui Wang $^{2}$ Ziwei Liu $^{5}$ Yu Qiao $^{2}$ Yali Wang $^{3,2}$ + +$^{1}$ Shanghai Jiao Tong University $^{2}$ Shanghai AI Laboratory + +3Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences + +4University of Chinese Academy of Sciences 5S-Lab, Nanyang Technological University + +# Abstract + +In this work, we present Vlogger, a generic AI system for generating a minute-level video blog (i.e., vlog) of user descriptions. Different from short videos with a few seconds, vlog often contains a complex storyline with diversified scenes, which is challenging for most existing video generation approaches. To break through this bottleneck, our Vlogger smartly leverages Large Language Model (LLM) as Director and decomposes a long video generation task of vlog into four key stages, where we invoke various foundation models to play the critical roles of vlog professionals, including (1) Script, (2) Actor, (3) ShowMaker, and (4) Voicer. With such a design of mimicking human beings, our Vlogger can generate vlogs through explainable cooperation of top-down planning and bottom-up shooting. Moreover, we introduce a novel video diffusion model, ShowMaker, which serves as a videographer in our Vlogger for generating the video snippet of each shooting scene. By incorporating Script and Actor attentively as textual and visual prompts, it can effectively enhance spatial-temporal coherence in the snippet. Besides, we design a concise mixed training paradigm for ShowMaker, boosting its capacity for both T2V generation and prediction. Finally, the extensive experiments show that our method achieves state-of-the-art performance on zero-shot T2V generation and prediction tasks. More importantly, Vlogger can generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor. + +# 1. Introduction + +Vlogs represent a unique form of blogging, distinguished by their utilization of video as the primary medium rather than text. Due to more lively expression in the dynamic scenes, blog has become one of the most popular online-sharing + +![](images/49b8f596ce7c194b305802120405d1253c5f8fbef479cbb63b4017879c956a07.jpg) +Figure 1. Overview of Vlogger. Based on the user story, our Vlogger leverages Large Language Model (LLM) as Director, and decomposes a minute-long vlog generation task into four key stages with Script, Actor, ShowMaker, and Voicer. Moreover, ShowMaker is a novel video diffusion model to generate video snippet of each shooting scene, with script and actor coherence. + +ways in the digital world. In the past two years, the remarkable success of diffusion models [31, 32, 59] has shown a great impact on video creation [8, 25, 33, 57, 70, 76, 83, 87]. Hence, there is a natural question, can we build a generic AI system to generate wonderful vlogs automatically? + +Regrettably, the majority of current video diffusion approaches mainly generate short videos with a few seconds, by temporal adaptation of image diffusion models [8, 25, 57, 70, 76, 87]. In contrast, a vlog typically constitutes a minute-level long video in the open world. Recently, there have been some attempts in long video generation [67, 83]. However, these early works either require extensive training on large well-captioned long video datasets [83] or suffer from noticeable incoherence of shot changes [67]. Hence, it is still challenging to generate a minute-level + +vlog with complex narratives and multiple scenes. + +Alternatively, we notice that a successful blog production is a systematical work in our realistic world, where the key staffs are involved in script creation, actor design, video shooting, and editing [22]. Drawing inspiration from this, we believe that, generating a long-form video blog requires elaborate systematical planning and shooting processes, rather than simply designing a generative model. + +Hence, we propose a generic AI system for vlog generation in this paper, namely Vlogger, which can smartly address this difficult task by mimicking vlog professionals with various foundation models in the core steps. As shown in Fig. 1, we first hire a Large Language Model (LLM) as Director (e.g., GPT-4 [46]), due to its great power of linguistic knowledge understanding. Given a user story, this director schedules a four-step plan for vlog generation. (1) Script. First, we introduce a progressive script creation paradigm for the LLM Director. By the coarse-to-fine instructions in this paradigm, the LLM Director can effectively convert a user story into a script, which sufficiently describes the story by a number of shooting scenes and their corresponding shooting duration. (2) Actor. After creating the script, the LLM Director reads it again to summarize the actors, and then invokes a character designer (e.g., SD-XL [48]) to generate the reference images of these actors in the vlog. (3) ShowMaker. With the guidance of script texts and actor images, we develop a novel ShowMaker as a videographer, which can effectively generate a controllable-duration snippet for each shooting scene, with spatial-temporal coherence. (4) Voicer. Finally, the LLM Director invokes a voicer (e.g., Bark [1]) to dub the vlog with script subtitles. + +It should be noted that our Vlogger overcomes the challenges previously encountered in long video generation tasks. On one hand, it elegantly decomposes the user story into a number of shooting scenes and designs actor images that can participate in different scenes in the vlog. In this case, it can reduce spatial-temporal incoherence of abrupt shot changes, with explicit guidance of scene texts and actor images. On the other hand, Vlogger crafts individual video snippets for every scene and seamlessly integrates them into a single cohesive vlog. Consequently, this bypasses the tedious training process with large-scale long video datasets. Via such collaboration between top-down planning and bottom-up shooting, Vlogger can effectively transform an open-world story into a minute-long vlog. + +Furthermore, we would like to emphasize that, ShowMaker is a distinctive video diffusion model designed for generating video snippets of each shooting scene. From the structural perspective, we introduce a novel Spatial-Temporal Enhanced Block (STEB) in this model. This block can adaptively leverage scene descriptions and actor images as textual and visual prompts, which attentively guide ShowMaker to enhance spatial-temporal coherence of + +script and actors. From the training perspective, we develop a probabilistic mode selection mechanism, which can boost the capacity of ShowMaker by mixed training of Text-to-Video (T2V) generation and prediction. More notably, by sequential combination of generation and prediction mode in the inference stage, ShowMaker can produce a video snippet with a controllable duration. This allows our Vlogger to generate a blog with a preferable duration, according to the planning of each scene in the script by LLM director. + +Finally, our method achieves the state-of-the-art performance on both zero-shot T2V generation and prediction, by expensive experiments within the popular video benchmarks. More importantly, our Vlogger outperforms the well-known long video generation method, i.e., Phenaki [67], despite utilizing only $66.7\%$ of the training videos. Remarkably, our Vlogger is capable of generating over 5-minute vlogs, without loss of script and actor coherence in the video. The models and codes will be released afterward. + +# 2. Related Works + +Text-to-Video Generation. Distinct from conventional unconditional and class-conditional video generation [7, 9, 13, 17, 23, 58, 63, 64, 72-74, 81, 85], T2V generation focuses on automatically converting textual descriptions into videos. This is a challenging task, as it involves understanding text semantics and translating it into video content. This often requires powerful cross-modal algorithms [50], large computing resources [15, 29], and extensive video data [3, 55, 56, 75, 76, 80]. Based on the success of diffusion models in the Text-to-Image (T2I) generation [4, 6, 18, 40, 45, 51, 52, 54], a series of such works have been recently transferred to T2V generation [8, 25, 33, 57, 70, 71, 76, 87]. However, whether trained from scratch [33, 57] or finetuned from T2I model [8, 25, 70, 71, 76, 87], most of these approaches mainly work on generating short videos with few seconds from simple descriptions. In contrast, our Vlogger can generate a minute-long vlog with complex stories. + +Long Video Generation. The generation of long videos predominantly relies on parallel [83] or autoregressive [67] structures. However, these early works still face challenges for vlog generation. On the one hand, the parallel manner can relax spatial-temporal content incoherence problems by coarse-to-fine generation. However, this approach necessitates extensive and laborious training on large, well-annotated long video datasets [83]. On the other hand, the autoregressive manner can relax heavy data requirements for training long videos, by applying short video generation models iteratively with sliding windows. However, this solution often suffers from noticeable shot change and long-term incoherence [14, 28, 49, 67, 69] which becomes problematic when generating vlogs encompassing complex narratives and multiple scenes. It is worth mentioning that the community has gradually realized that delegating higher + +![](images/02251ca8a9605faca7c26330df738c3971969d0686f1459410bb9afcc90382b2.jpg) +Figure 2. Top-Down Planning. Through four rounds of dialogue with the LLM, we gradually convert the user story into the final script. Based on this script, we further extract actor reference images, and then determine which actor would star in each script scene. + +order reasoning tasks to LLM is of great help to visual tasks [6, 26, 27, 37, 41, 44, 79, 88, 90]. Our Vlogger introduces LLM to the field of long video generation, and effectively addresses the problems of training burden and content incoherence in the previous methods, by distinct cooperation of top-down planning and bottom-up shooting. + +# 3. Method + +In this section, we introduce our Vlogger framework in detail. First, we describe how to make a vlog by planning and shooting with our Vlogger. Then, we further explain the novel design of ShowMaker. As the videographer in our Vlogger, it is critical for video generation in the vlog. + +# 3.1. Overall Framework of Vlogger + +To generate a minute-level vlog, our Vlogger leverages LLM as the director, which can effectively decompose this generation task by four key roles within the planning and shooting stages. As shown in Fig. 1, the LLM Director first creates Script and designs Actor in the planning stage. Based on Script and Actor, ShowMaker generates a video snippet for each scene in the shooting stage, and Voicer dubs subtitles of this snippet. Finally, one can combine the snippets of all the scenes as a vlog. + +# 3.1.1 Top-Down Planning + +A vlog often comes from a user story that contains diversified content within many shot changes. Clearly, it is challenging to produce a minute-long vlog, by directly feeding such a complex story into video generation models. To this end, we propose to leverage LLM as director and decompose the user story by top-down planning in Fig. 2. + +Script Creating. First, we parse the use story into a script, which describes this story explicitly by a number of shooting scenes. In this case, we can generate a video snippet for each shooting scene, instead of learning a long video tediously from the entire story. Since LLM has shown an impressive capacity in language understanding [2, 10, 16, 46, 61, 62, 65, 86], we feed the user story into such a director for script generation. As shown in Fig. 2, we introduce a progressive creation paradigm, which can effectively parse the story by coarse-to-fine steps, + +$$ +\mathcal {S} ^ {(i)} = \operatorname {L L M} \left(\mathcal {S} ^ {(i - 1)}, \mathcal {I} _ {S} ^ {(i)}, \mathcal {U}\right). \tag {1} +$$ + +Given the user story $\mathcal{U}$ and creation instruction $\mathcal{I}_S^{(i)}$ , the LLM Director generates the current script $S^{(i)}$ from the previous one $S^{(i-1)}$ . More specifically, there are four steps including (1) Rough. First, LLM generates a basic draft of the script from the story. (2) Detailed. Then, LLM refines the rough script with story details. (3) Completed. Next, LLM checks if the detailed script misses the important parts of the story. (4) Scheduled. Finally, LLM allocates a shooting duration for each scene in the completed script, according to the scene content. For convenience, we denote the final script as $\mathcal{S}$ in the following. It contains the descriptions of $N$ shooting scenes $\{S_1, \dots, S_N\}$ and their allocated duration $\{\mathcal{T}_1, \dots, \mathcal{T}_N\}$ . Additionally, due to the limited pages, please read the full descriptions of instructions and script in the supplementary doc. + +Actor Designing. After generating the script, it is time to design actors in the vlog. As shown in Fig. 2, we ask the LLM Director to summarize the actor list $\mathcal{A}$ from script $S$ , + +$$ +\mathcal {A} = \operatorname {L L M} \left(\mathcal {S}, \mathcal {I} _ {A}\right), \tag {2} +$$ + +where $\mathcal{I}_A$ is the instruction for actor summarization. Then, according to actor descriptions, the LLM Director invokes + +![](images/6bc924f6c8f59cefba01209a3d54d76064c7466e074cc3a4117ee7160ae8c73f.jpg) +Figure 3. Bottom-Up Shooting. For each scene, ShowMaker can generate the video snippet with script and actor coherence, by using script description and actor image as textual and visual prompts. Moreover, ShowMaker can effectively control the snippet duration, by performing generation and prediction sequentially in the inference. Finally, we apply a Text-To-Speech (TTS) model as Voicer for dubbing. + +a designer to generate reference images of these actors $\mathcal{R}$ + +$$ +\mathcal {R} = \text {S t a b l e - D i f f u s i o n} (\mathcal {A}). \tag {3} +$$ + +We choose Stable Diffusion XL [48] as the character designer, due to its high-quality generation. Finally, based on script $S$ and actor $\mathcal{A}$ , the LLM Director decides the leading actor (i.e., protagonist) in each shooting scene of the script, + +$$ +\mathcal {P} = \operatorname {L L M} (\mathcal {S}, \mathcal {A}, \mathcal {I} _ {P}). \tag {4} +$$ + +where $\mathcal{I}_P$ is the instruction for protagonist selection. The resulting doc $\mathcal{P}$ is aligned with the script $S$ , where $\mathcal{P}_n$ describes which actor appears in the scene $S_n$ . For example, $\{\mathcal{S}_6$ is starred by $\mathcal{A}_1\}$ in Fig. 2. + +# 3.1.2 Bottom-Up Shooting + +Via the top-down planning above, the LLM Director flexibly decomposes a complex user story into several script scenes and designs actor reference image for each scene. Such a manner largely reduces the difficulty of vlog generation, since we can generate vlogs by bottom-up shooting, i.e., we just need to generate the video snippet for each shooting scene and combine all of them as a vlog. + +ShowMaker Shooting. To generate the video snippet of a shooting scene, we introduce a novel ShowMaker as the videographer, which is a video diffusion model with two distinct designs. First, it is important to maintain spatial-temporal coherence of both script and actor in the generated snippet. Hence, our ShowMaker not only takes the scene description $S_{n}$ as a textual prompt but also takes actor image of this scene $\mathcal{R}_n$ as a visual prompt. Second, each + +scene is allocated with a shooting duration in the script. To control the duration of each scene, our ShowMaker contains two learning modes including generation and prediction in Fig.3. Specifically, it starts with the generation mode. For the shooting scene $n$ , we feed its script description $S_{n}$ and actor reference image $\mathcal{R}_n$ into ShowMaker, + +$$ +\mathcal {C} _ {n} ^ {(1)} = \operatorname {S h o w M a k e r} \left(\mathcal {N} _ {n} ^ {(1)} \mid S _ {n}, \mathcal {R} _ {n}, \text {G e n e r a t e}\right), \tag {5} +$$ + +which generates the first video clip of this scene $\mathcal{C}_n^{(1)}$ from the noisy clip $\mathcal{N}_n^{(1)}$ . If the duration of this clip is smaller than the allocated duration $\mathcal{T}_n$ in the script, we continue to perform the prediction mode, i.e., the last $k$ frames of the current clip $\mathcal{C}_n^{(j)}(k)$ are used as context, when generating the next clip from the noisy input $\mathcal{N}_n^{(j + 1)}$ , + +$$ +\mathcal {C} _ {n} ^ {(j + 1)} = \operatorname {S h o w M a k e r} \left(\mathcal {N} _ {n} ^ {(j + 1)} \mid \mathcal {S} _ {n}, \mathcal {C} _ {n} ^ {(j)} (k), \text {P r e d i c t}\right). \tag {6} +$$ + +Note that, actor reference images are not necessary in the prediction mode, since such an actor's appearance has been shown in the current clip $\mathcal{C}_n^{(j)}(k)$ that is used as input for prediction. This prediction procedure stops until the total duration achieves the allocated $\mathcal{T}_n$ of this scene. Subsequently, we combine all the clips as the video snippet of this scene, i.e., $\mathcal{C}_n = \{\mathcal{C}_n^{(1)},\dots,\mathcal{C}_n^{(J)}\}$ . + +Voiceer Speaking. To enhance the completeness of the vlog, we apply a Text-To-Speech model (e.g., Bark [1]) as a Voiceer, which converts the scene description $S_{n}$ into the corresponding audio $\mathcal{O}_n = \mathrm{Bark}(S_n)$ . Finally, we add this audio to the corresponding video snippet $\mathcal{V}_n = \mathcal{O}_n \oplus \mathcal{C}_n$ , and combine all the sounded video snippets as a complete vlog, i.e., $\mathcal{V} = \{\mathcal{V}_1, \dots, \mathcal{V}_N\}$ . + +![](images/59984be092b7681a5995ae5e04a4a0e17682fa4eea8060b8e6037928736c90da.jpg) +(a) Model Architecture + +![](images/cc5840383611892868a806d6cca32574491db4088501d78b1a54efd28426e880.jpg) +(b) Spatial-Temporal Enhanced Block (STEB) + +![](images/be821abd5ca9dafa9157bdf171838e846e1071c5aec2aff65e117e69db17a442.jpg) +(c) Mode Selection +Figure 4. ShowMaker. (a) The Overall Architecture. (b) Spatial-Temporal Enhanced Block (STEB). Via spatial-actor and temporal-text cross attention, our STEB can further enhance actor and script coherence in the snippet. (c) Mode Selection. We introduce a mixed training paradigm of T2V generation and prediction, via probabilistic selection of masked frames. + +# 3.2. ShowMaker + +As discussed in Section 3.1.2, ShowMaker plays a critical role in generating video snippet of a shooting scene. In this work, we introduce a text-to-video diffusion model for it. It follows the style of latent diffusion model [52]. In the diffusing stage, we add Gaussian noise progressively to the latent code of a training snippet. In the denoising stage, we reconstruct the latent code from the noisy latent code at any iteration step. For simplicity, we just show the denoising stage in Fig.4 (a). First, we forward a noisy training snippet into the encoder of the autoencoder to extract its latent code. Then, we feed this into a denoising U-Net [53] to learn the clean latent code. Finally, we leverage the decoder to reconstruct the original snippet with the clean latent code. + +But, compared with the existing video diffusion models [8, 25, 33, 34, 76], our novel ShowMaker contains two distinct designs, in terms of model structure (i.e., Spatial-Temporal Enhanced Block) in Fig.4 (b), and training paradigm (i.e., Mode Selection) in Fig.4 (c). + +# 3.2.1 Spatial-Temporal Enhanced Block (STEB) + +To reconstruct the clean latent code of a training video snippet, each block in the denoising U-Net consists of both spatial and temporal operations in previous works [76]. First, spatial operations encode the feature of each frame separately in the snippet. Typically, three operations are inherited from text-to-image generation approaches [4, 45, 51, 52, 54], including spatial ConVolution (CV), spatial Self Attention (SA), and spatial Cross Attention (CA), + +$$ +\mathcal {X} _ {c v} = \operatorname {C V - S p a t i a l} \left(\mathcal {X} _ {i n}, \mathcal {F} _ {t}\right), \tag {7} +$$ + +$$ +\mathcal {X} _ {s a} = \operatorname {S A} - \operatorname {S p a t i a l} \left(\mathcal {X} _ {c v}\right), \tag {8} +$$ + +$$ +\mathcal {X} _ {c a} = \operatorname {C A} - \text {S p a t i a l - T e x t} \left(\mathcal {X} _ {s a}, \mathcal {S} _ {n}\right), \tag {9} +$$ + +where $\mathcal{X}_{in}$ is the noisy feature of the training snippet, and $\mathcal{F}_t$ is the positional embedding of the iteration step $t$ . To guide denoising with the given text, we use the scene description $S_{n}$ as Key and Value of cross attention in Eq. (9). + +However, we notice that such spatial encoding does not consider actors. Hence, it inevitably suffers from actor incoherence when generating a snippet. To tackle this problem, we introduce a spatial image cross attention, + +$$ +\mathcal {Y} _ {c a} = \operatorname {C A} - \text {S p a t i a l - A c t o r} \left(\mathcal {X} _ {s a}, \mathcal {R} _ {n}\right). \tag {10} +$$ + +For a shooting scene $S_{n}$ , we leverage the protagonist doc (Eq. 4) to find the corresponding actor in this scene. Then, we leverage actor reference image $\mathcal{R}_{n}$ as the visual context of spatial cross attention. Subsequently, we enhance spatial embedding with the complementary guidance of both script and actor, i.e., $\mathcal{Z}_{se} = \mathcal{X}_{ca} + \beta \mathcal{Y}_{ca}$ with a weight $\beta$ . + +After spatial encoding, it is time to learn the correlation across frames in the snippet. Hence, the typical operation is to perform self attention along the temporal dimension, + +$$ +\mathcal {Z} _ {s a} = \operatorname {S A - T e m p o r a l} \left(\mathcal {Z} _ {s e}\right). \tag {11} +$$ + +However, such a temporal operation does not take the constraints of scene text into account. It often leads to text incoherence when generating video snippets. Hence, we introduce a temporal text cross attention, + +$$ +\mathcal {Z} _ {c a} = \operatorname {C A} - \text {T e m p o r a l - T e x t} \left(\mathcal {Z} _ {s a}, S _ {n}\right), \tag {12} +$$ + +where we leverage the scene description $S_{n}$ as the textual context of temporal encoding. Via spatial-actor (Eq. 10) and temporal-text (Eq. 12) cross attention, our STEB can further enhance actor and script coherence in the snippet. + +
MethodFVD (↓)
VideoFactory [71]410.00
Make-A-Video [57]367.23
PYoCo [25]355.19
Ours292.43
+ +(a) Hand-crafted prompt. + +
MethodFVD (↓)
CogVideo (Chinese) [35]751.34
CogVideo (English) [35]701.59
MagicVideo [89]699.00
Video LDM [8]550.61
Ours525.01
+ +(b) Class label. +Table 1. Zero-shot comparison with the state-of-the-art methods on UCF-101. Hand-crafted prompt from [25]. + +# 3.2.2 Mixed Training Paradigm with Mode Selection + +As discussed in Section 3.1.2, ShowMaker aims to generate a video snippet with the allocated duration in the script. To this goal, it leverages the combination of both generation and prediction modes in the inference stage. Next, we explain how to train our ShowMaker to learn these modes. + +As shown in Fig.4 (c), we design a mode selection mechanism, which selects $k$ frames of the clean snippet as the context of the noisy snippet. The $k = 0$ setting refers to the generation mode since there is no context from the clean snippet. Alternatively, the $k > 0$ setting refers to the prediction mode, since $k$ frames of the clean snippet are already available. The goal is to generate the rest frames of the clean snippet from the noisy one. To integrate both modes into training, we design a probabilistic manner to select $k$ , + +$$ +\mathrm {P} (k) = \left\{ \begin{array}{c c} \alpha^ {k} - \alpha^ {k + 1}, & k \in [ 0, m) \\ \alpha^ {k}, & k = m \end{array} \right. \tag {13} +$$ + +where $\mathrm{P}(k)$ is the selection probability distribution of $k$ . Moreover, $0 < \alpha < 1$ and $0 \leq m$ are the manual parameters, which respectively control the mode selection tendencies of $\mathrm{P}(k)$ and the maximum number of preserved frames. + +After determining $k$ , we introduce a frame mask $\mathcal{M}_k$ on the latent code of the clean snippet $\mathcal{X}_{clean}$ + +$$ +\mathcal {X} _ {k} = \mathcal {X} _ {\text {c l e a n}} \odot \mathcal {M} _ {k}, \tag {14} +$$ + +where we only preserve $k$ frames and mask the rest of the frames. Then, we concatenate the mask $\mathcal{M}_k$ with the latent code of snippet $\mathcal{X}_{noise}$ and $\mathcal{X}_k$ , + +$$ +\mathcal {X} _ {i n} = \operatorname {C o n c a t} \left(\mathcal {X} _ {\text {n o i s e}}, \mathcal {X} _ {k}, \mathcal {M} _ {k}\right). \tag {15} +$$ + +This produces the input feature $\mathcal{X}_{in}$ for training the denoising U-Net in Section 3.2.1. Via such a concise probabilistic manner, we can effectively integrate both generation $(k = 0)$ and prediction $(k > 0)$ modes in the training procedure. + +# 4. Experiments + +Datasets. To make state-of-the-art comparison, We conduct zero-shot evaluation on the popular video benchmarks, i.e., UCF-101 [60], Kinetics-400 [39], and MSR-VTT [12]. + +
MethodZero-ShotPre-training VideosFID (↓)
T2V [42]XX82.13
SC [5]XX33.51
TFGAN [5]XX31.76
NUWA [78]X0.97M28.46
Phenaki [67]✓15M37.74
Ours✓10M37.23
+ +Table 2. Comparison with the state-of-the-art methods on Kinetics-400. Both under a zero-shot setting, our FID is better than that of Phenaki, with only $66.7\%$ pre-training videos. + +
MethodZero-ShotCLIPSIM (↑)CLIP-FID (↓)
GODIVA [77]✘0.2402-
NüWA [78]✘0.243947.68
CogVideo (Chinese) [35]✓0.261424.78
CogVideo (English) [35]✓0.263123.59
Video LDM [8]✓0.2929-
Make-A-Video [57]✓0.304913.17
PYoCo [25]✓-10.21
Ours✓0.290812.67
+ +Table 3. Comparison with the state-of-the-art methods on MSR-VTT. PYoCo [25] sample 59,794 videos for CLIP-FID (noted in gray) while others only sample less than 3,000 videos. + +UCF-101 contains videos of 101 action categories. Following [66], we use FVD to evaluate the distance between the generated video and the real video. Kinetics-400 is a dataset comprising videos of 400 action categories. Following [5, 42, 67, 78], we use FID [30] to assess the performance of video generation. MSR-VTT is a video dataset with open-vocabulary captions, wherein CLIPSIM [77] and CLIP-FID [47] are commonly employed for evaluating the T2V generation. Moreover, since these existing benchmarks either have a small number of testing videos or contain only action labels without complex descriptions, we propose to collect an evaluation benchmark for ablation studies. It is called as Vimeo11k, where we collect 11,293 open-world videos along with their captions from 10 mainstream categories in Vimeo. To our best knowledge, it is the largest testing benchmark for zero-shot video generation. More details and experiments can be found in the supplementary doc. + +Implementation Details. (1) Director & Script & Actor & Voicer. We choose GPT-4 [46] as our LLM director to generate script. The specific instructions can be found in the supplementary doc. Then we employ Stable Diffusion XL [48] and Bark [1] as our designer and voicer to generate reference actor images and convert scripts into speech. (2) ShowMaker. We choose SD-1.4 [52] as our base model, and follow [57] to add temporal self attention. Then, we add our temporal text cross attention on top of each temporal self attention. We expand the input channel of the conv-in layer of U-Net from 4 to 9, so that model can take the concatenated feature in Eq. (15) as input. We use zero initialization for newly added channels. We use CLIP ViT-L/14 [50] as text encoder $\varepsilon_{T}$ , VQVAE [20] as autoencoder + +![](images/138b28faf8d1c260ede75d2a4290bff96e4e216c06206a80870633c8c5e729dd.jpg) +Figure 5. Long video generation. Comparison on UCF-101 (1000 frames). The lower FVD, the better generation performance. + +
MethodCLIP-I (↑)CLIP-T (↑)Vlog Duration (↑)
w/o autoregressive [91]0.56830.27521min03s
only autoregressive [67]0.56420.25353min58s
Ours0.62940.27563min58s
+ +consisting of $\varepsilon$ and $\mathcal{D}$ . Besides, we use OpenCLIP ViT-H/14 [38] as image encoder $\varepsilon_{I}$ , and add spatial image cross attention in the STEB block. The diffusion step $T$ is set to 1000 as [52]. For training, we set the parameter of mode selection as $\alpha = 0.6$ and $m = 6$ in Eq. (13), and $\beta$ is set to 1. Following [19, 33, 34, 67, 76], we employ publicly available image dataset (i.e., Laion400M [55]) and video dataset (i.e., WebVid10M [3]) for joint training. More training details can be found in the supplementary doc. + +# 4.1. Comparison with state-of-the-art + +Tab. 1 shows that, no matter whether the input text is the class label or hand-crafted prompt, our method achieves the best FVD performance of zero-shot video generation on UCF-101. Tab. 2 shows that, compared to Phenaki [67], our method achieves a better FID performance of zero-shot setting on Kinetics-400, but only using $66.7\%$ training videos. Furthermore, our generated videos have a resolution of $320 \times 512$ , which is higher than Phenaki's $256 \times 256$ . Tab. 3 shows that, our method achieves remarkably competitive FID and CLIPSIM performance on MSR-VTT. + +Furthermore, Fig. 5 illustrates that, we significantly surpass TATS [24] (i.e., the only open-source long video generation model within our knowledge), for generating 1000-frame videos on UCF-101. Moreover, our method does not encounter the issue of TATS, where the video quality continuously declines as the number of frames increases. It is noteworthy that, our method achieves this performance by zero-shot generation, without any finetuning on UCF-101. + +# 4.2. Ablation Study + +Vlog Generation Process. Tab. 4 presents a comparison between different generation processes. For a fair comparison, we use the same scripts and employ our ShowMaker + +![](images/efc26d8a5e467255e0cf3a20598a6737b9421e2d25a3c27942748fea7e853f67.jpg) +Figure 6. Ablation for spatial image cross attention. The values around a point on the curve correspond to the $\beta$ . While $\beta = 0$ means the original network without spatial image cross attention. + +Table 4. Ablation for generation process. [91] cannot generate a long video in a single scene and both [67, 91] can't refer to images. + +
TTCAMTP0/16 (↓)1/15 (↓)3/13 (↓)5/11 (↓)
✗✗348.36217.39297.37246.56
✓✗333.63178.27230.56193.52
✓✓257.70123.51118.02109.31
+ +Table 5. Ablation for temporal text cross attention and mixed training paradigm. "TTCA" and "MTP" refer to temporal text cross attention and mixed training paradigm respectively. "3/13" denotes predicting the following 13 frames after being given 3 frames of a ground truth video, and we evaluate the FVD between the newly generated 13 frames and the corresponding frames in the ground truth. "0/16", "1/15", "5/11" follow this pattern. + +as videographer. We compare three approaches including no autoregression at all as MovieFactory [91], fully autoregressive processes in one go as Phenaki [67], and the generation process of our Vlogger. Specifically, We use GPT-4 to generate five stories and go through the Vlogger's planning process to get five vlog scripts. We set 20 different random seeds, so that each of the above generation processes generates 20 different vlogs for each script. We use CLIP-I, CLIP-T [82] and video duration to evaluate the quality of the generated vlogs. The results demonstrate that, even using the same script and videographer, our Vlogger outshines other existing frameworks for preferable vlog generation. More details can be found in the supplementary doc. + +Spatial-Temporal Enhanced Block. Fig. 6 and Tab. 5 evaluate two important operations in our STEB block, i.e., spatial image cross attention and temporal text cross attention. With spatial image cross attention in Fig. 6, the CLIP-I is significantly improved as $\beta$ increased, while the CLIP-T reached its maximum at $\beta = 0.5$ on COCO2017 [43]. With temporal text cross attention (TTCA) in Tab. 5, both T2V generation (0/16) and prediction (1/15, 3/13 and 5/11) show a significant improvement on Vineo11k. + +Mixed Training Paradigm. Tab. 5 also presents a model comparison by incorporating our mixed training paradigm or the random mask training method in [11, 14, 21, 36, 68, 84]. The results show that the mixed training paradigm is effective in enhancing the model's T2V generation and prediction capabilities. More details in the supplementary doc. + +Lots of traffic in futuristic city. ... The camera moves away from the astronaut. The astronaut leaves the keyboard and walks to the left. ... We are in an office room with empty desks. A lion runs on top of the office desks. ... Timelapse of sunset in the modern city. + +![](images/a87289cd4d58944740a66d11ac1a26019ba3e80ace55bc9d7c582d7b82fa5b56.jpg) +Figure 7. Qualitative comparison with the state-of-the-art methods on long video generation. The story and long video of Phenaki [67] are available at phenaki.github.io. The scene diversity and picture quality generated by Vlogger are significantly better than Phenaki. + +![](images/91ddcd3217c96cbde8e784ce97c5f06cd4f4437bf8482916a1b819e1efda8449.jpg) +Figure 8. Qualitative ablation for STEB and training paradigm. "TTCA" and "MTP" refer to temporal the text cross attention and mixed training paradigm respectively. The sample without a visual prompt corresponds to the model without spatial image cross attention. + +# 4.3. Visualization + +First, we visualize long video generation, by comparison with the well-known Phenaki [67]. Note that, since Phenaki does not have the open-sourced codes, we alternatively use the demo shown in its official website. Specifically, we feed the same story description into our Vlogger. As shown in Fig. 7, our Vlogger shows superior and more diverse video content, compared to Phenaki. Moreover, our Vlogger sufficiently exhibits the story with a much longer duration (i.e., 7min59s), according to our LLM-created script. + +We further visualize the ablation of T2V video generation and prediction by ShowMaker. As depicted in Fig. 8, ShowMaker had significant improvements in generation and prediction performance, by incorporating our ShowMaker designs. Additionally, it can leverage visual prompts in the spatial actor cross attention, for distinguishing the + +"Teddy" concept within text prompts. + +# 5. Conclusion + +In this paper, we introduce a generic system Vlogger, to generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor. Moreover, we present a novel video diffusion model ShowMaker for boosting state-of-the-art T2V generation and prediction. Finally, we will release all the models, data and codes afterward, allowing to develop further designs toward long video generation in the open world. + +Acknowledgement. This work is supported by the National Key R&D Program of China (NO.2022ZD0160102), the National Natural Science Foundation of China under Grant No. 62102150, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800. + +# References + +[1] Suno AI. bark. https://github.com/suno-ai/bark#-usage-in-cython, 2023.2,4,6 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenhang Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, K. Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Yu Bowen, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xing Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. ArXiv, abs/2309.16609, 2023. 3 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 7 +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 2, 5 +[5] Yogesh Balaji1, Martin Renqiang Min, Bing Bai, Rama Chellappa1, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 6 +[6] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 2, 3 +[7] Sarthak Bhagat, Shagun Uppal, Vivian Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In ECCV, 2020. 2 +[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5, 6 +[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A. Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2 +[10] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 3 +[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and + +William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 7 +[12] Haoran Chen, Jianmin Li, Simone Frintrop, and Xiaolin Hu. The msr-video to text dataset with clean annotations. Comput. Vis. Image Underst., 225:103581, 2021. 6 +[13] Xinyuan Chen, Chang Xu, Xiaokang Yang, and Dacheng Tao. Long-term video prediction via criticization and retrospection. IEEE TIP, 29:7090-7103, 2020. 2 +[14] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. ArXiv, abs/2310.20700, 2023. 2, 7 +[15] François Chollet. On the measure of intelligence. ArXiv, abs/1911.01547, 2019. 2 +[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. JMLR, 24:240:1-240:113, 2022. 3 +[17] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv: Computer Vision and Pattern Recognition, 2019. 2 +[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 2 +[19] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay P. Namboodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. WACV, pages 3017-3031, 2019. 7 +[20] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2020. 6 +[21] Tsu-Jui Fu, Licheng Yu, N. Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what + +happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 7 +[22] Wen Gao, Yonghong Tian, Tiejun Huang, and Qiang Yang. Vlogging: A survey of videoblogging technology on the web. ACM Computing Surveys, 2010. 2 +[23] Songwei Ge, Thomas Hayes, Harry Yang, Xiaoyue Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2 +[24] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 7 +[25] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. ArXiv, abs/2305.10474, 2023. 1, 2, 5, 6 +[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3 +[27] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. Animate-a-story: Storytelling with retrieval-augmented video generation. ArXiv, abs/2307.06940, 2023. 3 +[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. ArXiv, abs/2211.13221, 2023. 2 +[29] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Frederick Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. ArXiv, abs/1712.00409, 2017. 2 +[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6 +[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 1 +[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1 +[33] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1, 2, 5, 7 +[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. ArXiv, abs/2204.03458, 2022. 5, 7 +[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In ICLR, 2023. 6 +[36] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. TMLR, 2022, 2022. 7 + +[37] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. ArXiv, abs/2309.14494, 2023. 3 +[38] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.7 +[39] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Apostol Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. ArXiv, abs/1705.06950, 2017. 6 +[40] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In NeurIPS, 2022. 2 +[41] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. ArXiv, abs/2305.06355, 2023. 3 +[42] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2017. 6 +[43] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7 +[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2022. 3 +[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 2, 5 +[46] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 3, 6 +[47] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 6 +[48] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 4, 6 +[49] Haonan Qiu, Menghan Xia, Yong Zhang, Yin-Yin He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv, abs/2310.15169, 2023. 2 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 6 + +[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2, 5 +[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 6, 7 +[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. ArXiv, abs/1505.04597, 2015. 5 +[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 5 +[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2, 7 +[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2 +[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1, 2, 6 +[58] Ivan Skorokhodov, S. Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2021. 2 +[59] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1 +[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. ArXiv, abs/1212.0402, 2012. 6 +[61] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, Weixin Liu, Zhihua Wu, Weibao Gong, Jianzhong Liang, Zhizhou Shang, Peng Sun, Wei Liu, Ouyang Xuan, Dianhai Yu, Hao Tian, Hua Wu, and Haifeng Wang. Ernie 3.0: Large-scale knowledge enhanced pretraining for language understanding and generation. ArXiv, abs/2107.02137, 2021. 3 +[62] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023.3 +[63] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. Motion-based generator model: Unsupervised disentanglement of appearance, + +trackable and intractable motions in dynamic patterns. In ICLR, 2019. 2 +[64] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 2 +[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. *ArXiv*, abs/2302.13971, 2023. 3 +[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. In ICLR, 2019. 6 +[67] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 1, 2, 6, 7, 8 +[68] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 7 +[69] Fu Lee Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. ArXiv, abs/2305.18264, 2023. 2 +[70] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. ArXiv, abs/2308.06571, 2023. 1, 2 +[71] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. ArXiv, abs/2305.10874, 2023. 2, 6 +[72] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. G3an: Disentangling appearance and motion for video generation. In CVPR, 2020. 2 +[73] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. Imaginar: Conditional spatiotemporal gan for video generation. WACV, pages 1149-1158, 2020. +[74] Yaohui Wang, François Bremond, and Antitza Dantcheva. Inmodegan: Interpretable motion decomposition generative adversarial network for video generation. ArXiv, abs/2101.03049, 2021. 2 +[75] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. *Internvideo: General video foundation models via generative and discriminative learning.* ArXiv, abs/2212.03191, 2022. 2 +[76] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Pe der Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Y. Qiao, and Ziwei Liu. Lavie: High-quality video + +generation with cascaded latent diffusion models. ArXiv, abs/2309.15103, 2023. 1, 2, 5, 7 +[77] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. ArXiv, abs/2104.14806, 2021. 6 +[78] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6 +[79] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 3 +[80] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2021. 2 +[81] Wilson Yan, Yunzhi Zhang, P. Abbeel, and A. Srinivas. Videogpt: Video generation using vq-vae and transformers. ArXiv, abs/2104.10157, 2021. 2 +[82] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 7 +[83] Sheng-Siang Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation. In Annual Meeting of the Association for Computational Linguistics, 2023. 1, 2 +[84] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit: Masked generative video transformer. In CVPR, 2023. 7 +[85] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2 +[86] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, P. Zhang, Yuxiao Dong, and Jie Tang. Glm-130b: An open bilingual pre-trained model. ArXiv, abs/2210.02414, 2022. 3 +[87] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lin Hao Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. ArXiv, abs/2309.15818, 2023. 1, 2 +[88] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. ArXiv, abs/2305.18583, 2023. 3 +[89] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. ArXiv, abs/2211.11018, 2022. 6 + +[90] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 3 +[91] Junchen Zhu, Huan Yang, Huiguo He, Wenjing Wang, Zixi Tuo, Wen-Huang Cheng, Lianli Gao, Jingkuan Song, and Jianlong Fu. Moviefactory: Automatic movie creation from text using large generative models for language and images. In ACM MM, 2023. 7 \ No newline at end of file diff --git a/2024/Vlogger_ Make Your Dream A Vlog/images.zip b/2024/Vlogger_ Make Your Dream A Vlog/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25d2a718bc3c28deb7ba418331032ec6c374f725 --- /dev/null +++ b/2024/Vlogger_ Make Your Dream A Vlog/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2a2ad975106804620caad78899248d57c8f5ddc0bb9b6f1074f82f68cb1e269 +size 769888 diff --git a/2024/Vlogger_ Make Your Dream A Vlog/layout.json b/2024/Vlogger_ Make Your Dream A Vlog/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8035e7bdc05d1fcd83ae388256daa911644ccac --- /dev/null +++ b/2024/Vlogger_ Make Your Dream A Vlog/layout.json @@ -0,0 +1,10331 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 97, + 204, + 130 + ], + "blocks": [ + { + "bbox": [ + 166, + 97, + 204, + 130 + ], + "lines": [ + { + "bbox": [ + 166, + 97, + 204, + 130 + ], + "spans": [ + { + "bbox": [ + 166, + 97, + 204, + 130 + ], + "type": "image", + "image_path": "afb986863174c39bf308b50146f3c2bf17f125c1acd1db62eb0d9a84d3631dc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 110, + 427, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 110, + 427, + 128 + ], + "spans": [ + { + "bbox": [ + 206, + 110, + 427, + 128 + ], + "type": "text", + "content": "Vlogger: Make Your Dream A Vlog" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "spans": [ + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": "Shaobin Zhuang" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": " Kunchang Li" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{3,4,2}" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": " Xinyuan Chen" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": " \nYaohui Wang" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "text", + "content": " Yali Wang" + }, + { + "bbox": [ + 151, + 157, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{3,2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 189, + 390, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 189, + 390, + 200 + ], + "spans": [ + { + "bbox": [ + 194, + 189, + 390, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 189, + 390, + 200 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 194, + 189, + 390, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 189, + 390, + 200 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 202, + 416, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 202, + 416, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 202, + 416, + 213 + ], + "type": "text", + "content": "3Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 217, + 438, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 438, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 438, + 228 + ], + "type": "text", + "content": "4University of Chinese Academy of Sciences 5S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 256, + 191, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 256, + 191, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 256, + 191, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 281, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 290, + 605 + ], + "type": "text", + "content": "In this work, we present Vlogger, a generic AI system for generating a minute-level video blog (i.e., vlog) of user descriptions. Different from short videos with a few seconds, vlog often contains a complex storyline with diversified scenes, which is challenging for most existing video generation approaches. To break through this bottleneck, our Vlogger smartly leverages Large Language Model (LLM) as Director and decomposes a long video generation task of vlog into four key stages, where we invoke various foundation models to play the critical roles of vlog professionals, including (1) Script, (2) Actor, (3) ShowMaker, and (4) Voicer. With such a design of mimicking human beings, our Vlogger can generate vlogs through explainable cooperation of top-down planning and bottom-up shooting. Moreover, we introduce a novel video diffusion model, ShowMaker, which serves as a videographer in our Vlogger for generating the video snippet of each shooting scene. By incorporating Script and Actor attentively as textual and visual prompts, it can effectively enhance spatial-temporal coherence in the snippet. Besides, we design a concise mixed training paradigm for ShowMaker, boosting its capacity for both T2V generation and prediction. Finally, the extensive experiments show that our method achieves state-of-the-art performance on zero-shot T2V generation and prediction tasks. More importantly, Vlogger can generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "content": "Vlogs represent a unique form of blogging, distinguished by their utilization of video as the primary medium rather than text. Due to more lively expression in the dynamic scenes, blog has become one of the most popular online-sharing" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 256, + 577, + 457 + ], + "blocks": [ + { + "bbox": [ + 307, + 256, + 577, + 457 + ], + "lines": [ + { + "bbox": [ + 307, + 256, + 577, + 457 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 577, + 457 + ], + "type": "image", + "image_path": "49b8f596ce7c194b305802120405d1253c5f8fbef479cbb63b4017879c956a07.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 463, + 545, + 529 + ], + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 529 + ], + "type": "text", + "content": "Figure 1. Overview of Vlogger. Based on the user story, our Vlogger leverages Large Language Model (LLM) as Director, and decomposes a minute-long vlog generation task into four key stages with Script, Actor, ShowMaker, and Voicer. Moreover, ShowMaker is a novel video diffusion model to generate video snippet of each shooting scene, with script and actor coherence." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 594 + ], + "type": "text", + "content": "ways in the digital world. In the past two years, the remarkable success of diffusion models [31, 32, 59] has shown a great impact on video creation [8, 25, 33, 57, 70, 76, 83, 87]. Hence, there is a natural question, can we build a generic AI system to generate wonderful vlogs automatically?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Regrettably, the majority of current video diffusion approaches mainly generate short videos with a few seconds, by temporal adaptation of image diffusion models [8, 25, 57, 70, 76, 87]. In contrast, a vlog typically constitutes a minute-level long video in the open world. Recently, there have been some attempts in long video generation [67, 83]. However, these early works either require extensive training on large well-captioned long video datasets [83] or suffer from noticeable incoherence of shot changes [67]. Hence, it is still challenging to generate a minute-level" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 270, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 270, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 270, + 713 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 61, + 702, + 270, + 713 + ], + "type": "text", + "content": " Interns at Shanghai AI Laboratory. " + }, + { + "bbox": [ + 61, + 702, + 270, + 713 + ], + "type": "inline_equation", + "content": "\\heartsuit" + }, + { + "bbox": [ + 61, + 702, + 270, + 713 + ], + "type": "text", + "content": " Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8806" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 250, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 250, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 250, + 84 + ], + "type": "text", + "content": "vlog with complex narratives and multiple scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 86, + 286, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 86, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 286, + 170 + ], + "type": "text", + "content": "Alternatively, we notice that a successful blog production is a systematical work in our realistic world, where the key staffs are involved in script creation, actor design, video shooting, and editing [22]. Drawing inspiration from this, we believe that, generating a long-form video blog requires elaborate systematical planning and shooting processes, rather than simply designing a generative model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 172, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 172, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 286, + 445 + ], + "type": "text", + "content": "Hence, we propose a generic AI system for vlog generation in this paper, namely Vlogger, which can smartly address this difficult task by mimicking vlog professionals with various foundation models in the core steps. As shown in Fig. 1, we first hire a Large Language Model (LLM) as Director (e.g., GPT-4 [46]), due to its great power of linguistic knowledge understanding. Given a user story, this director schedules a four-step plan for vlog generation. (1) Script. First, we introduce a progressive script creation paradigm for the LLM Director. By the coarse-to-fine instructions in this paradigm, the LLM Director can effectively convert a user story into a script, which sufficiently describes the story by a number of shooting scenes and their corresponding shooting duration. (2) Actor. After creating the script, the LLM Director reads it again to summarize the actors, and then invokes a character designer (e.g., SD-XL [48]) to generate the reference images of these actors in the vlog. (3) ShowMaker. With the guidance of script texts and actor images, we develop a novel ShowMaker as a videographer, which can effectively generate a controllable-duration snippet for each shooting scene, with spatial-temporal coherence. (4) Voicer. Finally, the LLM Director invokes a voicer (e.g., Bark [1]) to dub the vlog with script subtitles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 286, + 616 + ], + "type": "text", + "content": "It should be noted that our Vlogger overcomes the challenges previously encountered in long video generation tasks. On one hand, it elegantly decomposes the user story into a number of shooting scenes and designs actor images that can participate in different scenes in the vlog. In this case, it can reduce spatial-temporal incoherence of abrupt shot changes, with explicit guidance of scene texts and actor images. On the other hand, Vlogger crafts individual video snippets for every scene and seamlessly integrates them into a single cohesive vlog. Consequently, this bypasses the tedious training process with large-scale long video datasets. Via such collaboration between top-down planning and bottom-up shooting, Vlogger can effectively transform an open-world story into a minute-long vlog." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 286, + 713 + ], + "type": "text", + "content": "Furthermore, we would like to emphasize that, ShowMaker is a distinctive video diffusion model designed for generating video snippets of each shooting scene. From the structural perspective, we introduce a novel Spatial-Temporal Enhanced Block (STEB) in this model. This block can adaptively leverage scene descriptions and actor images as textual and visual prompts, which attentively guide ShowMaker to enhance spatial-temporal coherence of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "script and actors. From the training perspective, we develop a probabilistic mode selection mechanism, which can boost the capacity of ShowMaker by mixed training of Text-to-Video (T2V) generation and prediction. More notably, by sequential combination of generation and prediction mode in the inference stage, ShowMaker can produce a video snippet with a controllable duration. This allows our Vlogger to generate a blog with a preferable duration, according to the planning of each scene in the script by LLM director." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 180, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 287 + ], + "type": "text", + "content": "Finally, our method achieves the state-of-the-art performance on both zero-shot T2V generation and prediction, by expensive experiments within the popular video benchmarks. More importantly, our Vlogger outperforms the well-known long video generation method, i.e., Phenaki [67], despite utilizing only " + }, + { + "bbox": [ + 304, + 180, + 545, + 287 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 304, + 180, + 545, + 287 + ], + "type": "text", + "content": " of the training videos. Remarkably, our Vlogger is capable of generating over 5-minute vlogs, without loss of script and actor coherence in the video. The models and codes will be released afterward." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 298, + 397, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 397, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 397, + 311 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 521 + ], + "type": "text", + "content": "Text-to-Video Generation. Distinct from conventional unconditional and class-conditional video generation [7, 9, 13, 17, 23, 58, 63, 64, 72-74, 81, 85], T2V generation focuses on automatically converting textual descriptions into videos. This is a challenging task, as it involves understanding text semantics and translating it into video content. This often requires powerful cross-modal algorithms [50], large computing resources [15, 29], and extensive video data [3, 55, 56, 75, 76, 80]. Based on the success of diffusion models in the Text-to-Image (T2I) generation [4, 6, 18, 40, 45, 51, 52, 54], a series of such works have been recently transferred to T2V generation [8, 25, 33, 57, 70, 71, 76, 87]. However, whether trained from scratch [33, 57] or finetuned from T2I model [8, 25, 70, 71, 76, 87], most of these approaches mainly work on generating short videos with few seconds from simple descriptions. In contrast, our Vlogger can generate a minute-long vlog with complex stories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Long Video Generation. The generation of long videos predominantly relies on parallel [83] or autoregressive [67] structures. However, these early works still face challenges for vlog generation. On the one hand, the parallel manner can relax spatial-temporal content incoherence problems by coarse-to-fine generation. However, this approach necessitates extensive and laborious training on large, well-annotated long video datasets [83]. On the other hand, the autoregressive manner can relax heavy data requirements for training long videos, by applying short video generation models iteratively with sliding windows. However, this solution often suffers from noticeable shot change and long-term incoherence [14, 28, 49, 67, 69] which becomes problematic when generating vlogs encompassing complex narratives and multiple scenes. It is worth mentioning that the community has gradually realized that delegating higher" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8807" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 269 + ], + "type": "image", + "image_path": "02251ca8a9605faca7c26330df738c3971969d0686f1459410bb9afcc90382b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 546, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 546, + 294 + ], + "type": "text", + "content": "Figure 2. Top-Down Planning. Through four rounds of dialogue with the LLM, we gradually convert the user story into the final script. Based on this script, we further extract actor reference images, and then determine which actor would star in each script scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 374 + ], + "type": "text", + "content": "order reasoning tasks to LLM is of great help to visual tasks [6, 26, 27, 37, 41, 44, 79, 88, 90]. Our Vlogger introduces LLM to the field of long video generation, and effectively addresses the problems of training burden and content incoherence in the previous methods, by distinct cooperation of top-down planning and bottom-up shooting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 385, + 104, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 104, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 104, + 398 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 467 + ], + "type": "text", + "content": "In this section, we introduce our Vlogger framework in detail. First, we describe how to make a vlog by planning and shooting with our Vlogger. Then, we further explain the novel design of ShowMaker. As the videographer in our Vlogger, it is critical for video generation in the vlog." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 475, + 216, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 216, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 216, + 489 + ], + "type": "text", + "content": "3.1. Overall Framework of Vlogger" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "type": "text", + "content": "To generate a minute-level vlog, our Vlogger leverages LLM as the director, which can effectively decompose this generation task by four key roles within the planning and shooting stages. As shown in Fig. 1, the LLM Director first creates Script and designs Actor in the planning stage. Based on Script and Actor, ShowMaker generates a video snippet for each scene in the shooting stage, and Voicer dubs subtitles of this snippet. Finally, one can combine the snippets of all the scenes as a vlog." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 621, + 167, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 167, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 167, + 634 + ], + "type": "text", + "content": "3.1.1 Top-Down Planning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "A vlog often comes from a user story that contains diversified content within many shot changes. Clearly, it is challenging to produce a minute-long vlog, by directly feeding such a complex story into video generation models. To this end, we propose to leverage LLM as director and decompose the user story by top-down planning in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 300, + 546, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 421 + ], + "type": "text", + "content": "Script Creating. First, we parse the use story into a script, which describes this story explicitly by a number of shooting scenes. In this case, we can generate a video snippet for each shooting scene, instead of learning a long video tediously from the entire story. Since LLM has shown an impressive capacity in language understanding [2, 10, 16, 46, 61, 62, 65, 86], we feed the user story into such a director for script generation. As shown in Fig. 2, we introduce a progressive creation paradigm, which can effectively parse the story by coarse-to-fine steps," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 426, + 545, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 426, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 364, + 426, + 545, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {(i)} = \\operatorname {L L M} \\left(\\mathcal {S} ^ {(i - 1)}, \\mathcal {I} _ {S} ^ {(i)}, \\mathcal {U}\\right). \\tag {1}", + "image_path": "3b6d21567cb063facd3091222cc3247ebe9c7038c6b144827e32d0771a89cc1c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": "Given the user story " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " and creation instruction " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_S^{(i)}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": ", the LLM Director generates the current script " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "S^{(i)}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " from the previous one " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "S^{(i-1)}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": ". More specifically, there are four steps including (1) Rough. First, LLM generates a basic draft of the script from the story. (2) Detailed. Then, LLM refines the rough script with story details. (3) Completed. Next, LLM checks if the detailed script misses the important parts of the story. (4) Scheduled. Finally, LLM allocates a shooting duration for each scene in the completed script, according to the scene content. For convenience, we denote the final script as " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " in the following. It contains the descriptions of " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " shooting scenes " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\{S_1, \\dots, S_N\\}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " and their allocated duration " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_1, \\dots, \\mathcal{T}_N\\}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": ". Additionally, due to the limited pages, please read the full descriptions of instructions and script in the supplementary doc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": "Actor Designing. After generating the script, it is time to design actors in the vlog. As shown in Fig. 2, we ask the LLM Director to summarize the actor list " + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": " from script " + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 386, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = \\operatorname {L L M} \\left(\\mathcal {S}, \\mathcal {I} _ {A}\\right), \\tag {2}", + "image_path": "1866de2b322b4474fc9a845705323a5f4303d831e246f0530921bdbcebe10ded.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_A" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the instruction for actor summarization. Then, according to actor descriptions, the LLM Director invokes" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8808" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 534, + 281 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 534, + 281 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 534, + 281 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 534, + 281 + ], + "type": "image", + "image_path": "6bc924f6c8f59cefba01209a3d54d76064c7466e074cc3a4117ee7160ae8c73f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "text", + "content": "Figure 3. Bottom-Up Shooting. For each scene, ShowMaker can generate the video snippet with script and actor coherence, by using script description and actor image as textual and visual prompts. Moreover, ShowMaker can effectively control the snippet duration, by performing generation and prediction sequentially in the inference. Finally, we apply a Text-To-Speech (TTS) model as Voicer for dubbing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 325, + 282, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 282, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 282, + 338 + ], + "type": "text", + "content": "a designer to generate reference images of these actors " + }, + { + "bbox": [ + 47, + 325, + 282, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 348, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 348, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 112, + 348, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {R} = \\text {S t a b l e - D i f f u s i o n} (\\mathcal {A}). \\tag {3}", + "image_path": "6f136956d32e8c0cc41a96d78d95e205cecc4c82d168b474735b99ab61be5f83.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "type": "text", + "content": "We choose Stable Diffusion XL [48] as the character designer, due to its high-quality generation. Finally, based on script " + }, + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "type": "text", + "content": " and actor " + }, + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 369, + 287, + 419 + ], + "type": "text", + "content": ", the LLM Director decides the leading actor (i.e., protagonist) in each shooting scene of the script," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 430, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 430, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 121, + 430, + 287, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {P} = \\operatorname {L L M} (\\mathcal {S}, \\mathcal {A}, \\mathcal {I} _ {P}). \\tag {4}", + "image_path": "4642b529de6094944bb9cc75fddda96fba96b2fecc00ed5f80c4e4485ba7db96.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_P" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": " is the instruction for protagonist selection. The resulting doc " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": " is aligned with the script " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": " describes which actor appears in the scene " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "S_n" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": ". For example, " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{S}_6" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": " is starred by " + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_1\\}" + }, + { + "bbox": [ + 47, + 450, + 287, + 498 + ], + "type": "text", + "content": " in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 514, + 169, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 169, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 169, + 528 + ], + "type": "text", + "content": "3.1.2 Bottom-Up Shooting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "content": "Via the top-down planning above, the LLM Director flexibly decomposes a complex user story into several script scenes and designs actor reference image for each scene. Such a manner largely reduces the difficulty of vlog generation, since we can generate vlogs by bottom-up shooting, i.e., we just need to generate the video snippet for each shooting scene and combine all of them as a vlog." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "ShowMaker Shooting. To generate the video snippet of a shooting scene, we introduce a novel ShowMaker as the videographer, which is a video diffusion model with two distinct designs. First, it is important to maintain spatial-temporal coherence of both script and actor in the generated snippet. Hence, our ShowMaker not only takes the scene description " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " as a textual prompt but also takes actor image of this scene " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_n" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " as a visual prompt. Second, each" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "text", + "content": "scene is allocated with a shooting duration in the script. To control the duration of each scene, our ShowMaker contains two learning modes including generation and prediction in Fig.3. Specifically, it starts with the generation mode. For the shooting scene " + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "text", + "content": ", we feed its script description " + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "text", + "content": " and actor reference image " + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_n" + }, + { + "bbox": [ + 304, + 325, + 547, + 397 + ], + "type": "text", + "content": " into ShowMaker," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 405, + 545, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 405, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 329, + 405, + 545, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {n} ^ {(1)} = \\operatorname {S h o w M a k e r} \\left(\\mathcal {N} _ {n} ^ {(1)} \\mid S _ {n}, \\mathcal {R} _ {n}, \\text {G e n e r a t e}\\right), \\tag {5}", + "image_path": "44b70f09ce67a986cb5e707e8482de6106eb32000c53727accb12300e49354d6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": "which generates the first video clip of this scene " + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n^{(1)}" + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": " from the noisy clip " + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_n^{(1)}" + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": ". If the duration of this clip is smaller than the allocated duration " + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_n" + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": " in the script, we continue to perform the prediction mode, i.e., the last " + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": " frames of the current clip " + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n^{(j)}(k)" + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": " are used as context, when generating the next clip from the noisy input " + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_n^{(j + 1)}" + }, + { + "bbox": [ + 304, + 431, + 547, + 510 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 517, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 517, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 310, + 517, + 545, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {n} ^ {(j + 1)} = \\operatorname {S h o w M a k e r} \\left(\\mathcal {N} _ {n} ^ {(j + 1)} \\mid \\mathcal {S} _ {n}, \\mathcal {C} _ {n} ^ {(j)} (k), \\text {P r e d i c t}\\right). \\tag {6}", + "image_path": "1d8976222ae6826f7bf94c2f25f31c904d77e3d70343fa91341c612a238d6832.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "text", + "content": "Note that, actor reference images are not necessary in the prediction mode, since such an actor's appearance has been shown in the current clip " + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n^{(j)}(k)" + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "text", + "content": " that is used as input for prediction. This prediction procedure stops until the total duration achieves the allocated " + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_n" + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "text", + "content": " of this scene. Subsequently, we combine all the clips as the video snippet of this scene, i.e., " + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n = \\{\\mathcal{C}_n^{(1)},\\dots,\\mathcal{C}_n^{(J)}\\}" + }, + { + "bbox": [ + 304, + 542, + 545, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Voiceer Speaking. To enhance the completeness of the vlog, we apply a Text-To-Speech model (e.g., Bark [1]) as a Voiceer, which converts the scene description " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": " into the corresponding audio " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_n = \\mathrm{Bark}(S_n)" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": ". Finally, we add this audio to the corresponding video snippet " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_n = \\mathcal{O}_n \\oplus \\mathcal{C}_n" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": ", and combine all the sounded video snippets as a complete vlog, i.e., " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathcal{V}_1, \\dots, \\mathcal{V}_N\\}" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8809" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 281, + 240 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 281, + 240 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 281, + 240 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 281, + 240 + ], + "type": "image", + "image_path": "59984be092b7681a5995ae5e04a4a0e17682fa4eea8060b8e6037928736c90da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 242, + 196, + 252 + ], + "lines": [ + { + "bbox": [ + 111, + 242, + 196, + 252 + ], + "spans": [ + { + "bbox": [ + 111, + 242, + 196, + 252 + ], + "type": "text", + "content": "(a) Model Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 292, + 70, + 536, + 157 + ], + "blocks": [ + { + "bbox": [ + 292, + 70, + 536, + 157 + ], + "lines": [ + { + "bbox": [ + 292, + 70, + 536, + 157 + ], + "spans": [ + { + "bbox": [ + 292, + 70, + 536, + 157 + ], + "type": "image", + "image_path": "cc5840383611892868a806d6cca32574491db4088501d78b1a54efd28426e880.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 163, + 506, + 174 + ], + "lines": [ + { + "bbox": [ + 339, + 163, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 339, + 163, + 506, + 174 + ], + "type": "text", + "content": "(b) Spatial-Temporal Enhanced Block (STEB)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 291, + 182, + 538, + 237 + ], + "blocks": [ + { + "bbox": [ + 291, + 182, + 538, + 237 + ], + "lines": [ + { + "bbox": [ + 291, + 182, + 538, + 237 + ], + "spans": [ + { + "bbox": [ + 291, + 182, + 538, + 237 + ], + "type": "image", + "image_path": "be821abd5ca9dafa9157bdf171838e846e1071c5aec2aff65e117e69db17a442.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 242, + 465, + 251 + ], + "lines": [ + { + "bbox": [ + 396, + 242, + 465, + 251 + ], + "spans": [ + { + "bbox": [ + 396, + 242, + 465, + 251 + ], + "type": "text", + "content": "(c) Mode Selection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 261, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 45, + 261, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 45, + 261, + 547, + 294 + ], + "type": "text", + "content": "Figure 4. ShowMaker. (a) The Overall Architecture. (b) Spatial-Temporal Enhanced Block (STEB). Via spatial-actor and temporal-text cross attention, our STEB can further enhance actor and script coherence in the snippet. (c) Mode Selection. We introduce a mixed training paradigm of T2V generation and prediction, via probabilistic selection of masked frames." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 300, + 127, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 127, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 127, + 312 + ], + "type": "text", + "content": "3.2. ShowMaker" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 319, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 319, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 45, + 319, + 287, + 474 + ], + "type": "text", + "content": "As discussed in Section 3.1.2, ShowMaker plays a critical role in generating video snippet of a shooting scene. In this work, we introduce a text-to-video diffusion model for it. It follows the style of latent diffusion model [52]. In the diffusing stage, we add Gaussian noise progressively to the latent code of a training snippet. In the denoising stage, we reconstruct the latent code from the noisy latent code at any iteration step. For simplicity, we just show the denoising stage in Fig.4 (a). First, we forward a noisy training snippet into the encoder of the autoencoder to extract its latent code. Then, we feed this into a denoising U-Net [53] to learn the clean latent code. Finally, we leverage the decoder to reconstruct the original snippet with the clean latent code." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 475, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 535 + ], + "type": "text", + "content": "But, compared with the existing video diffusion models [8, 25, 33, 34, 76], our novel ShowMaker contains two distinct designs, in terms of model structure (i.e., Spatial-Temporal Enhanced Block) in Fig.4 (b), and training paradigm (i.e., Mode Selection) in Fig.4 (c)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 548, + 261, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 261, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 261, + 560 + ], + "type": "text", + "content": "3.2.1 Spatial-Temporal Enhanced Block (STEB)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 567, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 662 + ], + "type": "text", + "content": "To reconstruct the clean latent code of a training video snippet, each block in the denoising U-Net consists of both spatial and temporal operations in previous works [76]. First, spatial operations encode the feature of each frame separately in the snippet. Typically, three operations are inherited from text-to-image generation approaches [4, 45, 51, 52, 54], including spatial ConVolution (CV), spatial Self Attention (SA), and spatial Cross Attention (CA)," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 96, + 667, + 287, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 667, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 96, + 667, + 287, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {c v} = \\operatorname {C V - S p a t i a l} \\left(\\mathcal {X} _ {i n}, \\mathcal {F} _ {t}\\right), \\tag {7}", + "image_path": "427f006efb0b840419f2bc49a0cc1b47cb1c13c0d7792522ff3d73d559570c2c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 97, + 683, + 287, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 683, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 97, + 683, + 287, + 695 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {s a} = \\operatorname {S A} - \\operatorname {S p a t i a l} \\left(\\mathcal {X} _ {c v}\\right), \\tag {8}", + "image_path": "09fdbce2765ed11ae9866c30fd40848c4f15690de5347a0183e2116f22188af2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 97, + 698, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 698, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 97, + 698, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {c a} = \\operatorname {C A} - \\text {S p a t i a l - T e x t} \\left(\\mathcal {X} _ {s a}, \\mathcal {S} _ {n}\\right), \\tag {9}", + "image_path": "116c6c50503259033a598cdaed75bb6ebef05c005efef415aa5c5631a6bc2835.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{in}" + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "text", + "content": " is the noisy feature of the training snippet, and " + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "text", + "content": " is the positional embedding of the iteration step " + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "text", + "content": ". To guide denoising with the given text, we use the scene description " + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 306, + 301, + 545, + 350 + ], + "type": "text", + "content": " as Key and Value of cross attention in Eq. (9)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "content": "However, we notice that such spatial encoding does not consider actors. Hence, it inevitably suffers from actor incoherence when generating a snippet. To tackle this problem, we introduce a spatial image cross attention," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 351, + 411, + 545, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 411, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 351, + 411, + 545, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} _ {c a} = \\operatorname {C A} - \\text {S p a t i a l - A c t o r} \\left(\\mathcal {X} _ {s a}, \\mathcal {R} _ {n}\\right). \\tag {10}", + "image_path": "9a07ca22f0e4e4c66186fa62ce14bb840dbbe880a705d4a77ee8582cf1cb400f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "text", + "content": "For a shooting scene " + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "text", + "content": ", we leverage the protagonist doc (Eq. 4) to find the corresponding actor in this scene. Then, we leverage actor reference image " + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{n}" + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "text", + "content": " as the visual context of spatial cross attention. Subsequently, we enhance spatial embedding with the complementary guidance of both script and actor, i.e., " + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{se} = \\mathcal{X}_{ca} + \\beta \\mathcal{Y}_{ca}" + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "text", + "content": " with a weight " + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 435, + 545, + 507 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 508, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 544 + ], + "type": "text", + "content": "After spatial encoding, it is time to learn the correlation across frames in the snippet. Hence, the typical operation is to perform self attention along the temporal dimension," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 370, + 556, + 545, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 556, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 370, + 556, + 545, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} _ {s a} = \\operatorname {S A - T e m p o r a l} \\left(\\mathcal {Z} _ {s e}\\right). \\tag {11}", + "image_path": "fac4e8482621afa7e76bf3a29a95d0822ba3cc9c1828ffa82b13c644f4ea830f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "type": "text", + "content": "However, such a temporal operation does not take the constraints of scene text into account. It often leads to text incoherence when generating video snippets. Hence, we introduce a temporal text cross attention," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 351, + 641, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 641, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 351, + 641, + 545, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} _ {c a} = \\operatorname {C A} - \\text {T e m p o r a l - T e x t} \\left(\\mathcal {Z} _ {s a}, S _ {n}\\right), \\tag {12}", + "image_path": "03e754364ecabf6b3737a56c9a66d44200e069689eed8cb059b344558249d48f.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "where we leverage the scene description " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " as the textual context of temporal encoding. Via spatial-actor (Eq. 10) and temporal-text (Eq. 12) cross attention, our STEB can further enhance actor and script coherence in the snippet." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8810" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 164, + 132 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 164, + 132 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 164, + 132 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 164, + 132 + ], + "type": "table", + "html": "
MethodFVD (↓)
VideoFactory [71]410.00
Make-A-Video [57]367.23
PYoCo [25]355.19
Ours292.43
", + "image_path": "87c03f2c3a895b1ebcbec990458c4721af331c5dab047ca71939fe809a118ce3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 66, + 139, + 148, + 149 + ], + "lines": [ + { + "bbox": [ + 66, + 139, + 148, + 149 + ], + "spans": [ + { + "bbox": [ + 66, + 139, + 148, + 149 + ], + "type": "text", + "content": "(a) Hand-crafted prompt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 168, + 70, + 288, + 139 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 288, + 139 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 288, + 139 + ], + "type": "table", + "html": "
MethodFVD (↓)
CogVideo (Chinese) [35]751.34
CogVideo (English) [35]701.59
MagicVideo [89]699.00
Video LDM [8]550.61
Ours525.01
", + "image_path": "0f0207750094ee3dc44da23589e0052ddc719151ca847e2657c073290cd95bb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 200, + 140, + 250, + 149 + ], + "lines": [ + { + "bbox": [ + 200, + 140, + 250, + 149 + ], + "spans": [ + { + "bbox": [ + 200, + 140, + 250, + 149 + ], + "type": "text", + "content": "(b) Class label." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 46, + 151, + 287, + 173 + ], + "lines": [ + { + "bbox": [ + 46, + 151, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 46, + 151, + 287, + 173 + ], + "type": "text", + "content": "Table 1. Zero-shot comparison with the state-of-the-art methods on UCF-101. Hand-crafted prompt from [25]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 177, + 281, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 281, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 281, + 190 + ], + "type": "text", + "content": "3.2.2 Mixed Training Paradigm with Mode Selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": "As discussed in Section 3.1.2, ShowMaker aims to generate a video snippet with the allocated duration in the script. To this goal, it leverages the combination of both generation and prediction modes in the inference stage. Next, we explain how to train our ShowMaker to learn these modes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": "As shown in Fig.4 (c), we design a mode selection mechanism, which selects " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " frames of the clean snippet as the context of the noisy snippet. The " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " setting refers to the generation mode since there is no context from the clean snippet. Alternatively, the " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "k > 0" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " setting refers to the prediction mode, since " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " frames of the clean snippet are already available. The goal is to generate the rest frames of the clean snippet from the noisy one. To integrate both modes into training, we design a probabilistic manner to select " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 383, + 287, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 91, + 383, + 287, + 415 + ], + "type": "interline_equation", + "content": "\\mathrm {P} (k) = \\left\\{ \\begin{array}{c c} \\alpha^ {k} - \\alpha^ {k + 1}, & k \\in [ 0, m) \\\\ \\alpha^ {k}, & k = m \\end{array} \\right. \\tag {13}", + "image_path": "fd4532668cb300ede74a8bcc326928a5a781873e97799138e4e86fb4903002e9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(k)" + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "content": " is the selection probability distribution of " + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "content": ". Moreover, " + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "inline_equation", + "content": "0 < \\alpha < 1" + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "inline_equation", + "content": "0 \\leq m" + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "content": " are the manual parameters, which respectively control the mode selection tendencies of " + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(k)" + }, + { + "bbox": [ + 46, + 424, + 287, + 472 + ], + "type": "text", + "content": " and the maximum number of preserved frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "text", + "content": "After determining " + }, + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "text", + "content": ", we introduce a frame mask " + }, + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "text", + "content": " on the latent code of the clean snippet " + }, + { + "bbox": [ + 46, + 472, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{clean}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 123, + 506, + 287, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 506, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 123, + 506, + 287, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {k} = \\mathcal {X} _ {\\text {c l e a n}} \\odot \\mathcal {M} _ {k}, \\tag {14}", + "image_path": "64d7deb2d930cf3696267790dcfb377be5d5b1e3006d6aa61679ba121c09a8df.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "text", + "content": "where we only preserve " + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "text", + "content": " frames and mask the rest of the frames. Then, we concatenate the mask " + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "text", + "content": " with the latent code of snippet " + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{noise}" + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_k" + }, + { + "bbox": [ + 46, + 529, + 287, + 566 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 96, + 575, + 287, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 575, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 96, + 575, + 287, + 588 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {i n} = \\operatorname {C o n c a t} \\left(\\mathcal {X} _ {\\text {n o i s e}}, \\mathcal {X} _ {k}, \\mathcal {M} _ {k}\\right). \\tag {15}", + "image_path": "506844ba90962e061e165832a3ec159a58094a941eddc71aa0d21dc981c911bc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": "This produces the input feature " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{in}" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " for training the denoising U-Net in Section 3.2.1. Via such a concise probabilistic manner, we can effectively integrate both generation " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "(k = 0)" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " and prediction " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "(k > 0)" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " modes in the training procedure." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 657, + 128, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 128, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 128, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Datasets. To make state-of-the-art comparison, We conduct zero-shot evaluation on the popular video benchmarks, i.e., UCF-101 [60], Kinetics-400 [39], and MSR-VTT [12]." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 544, + 155 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 544, + 155 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 544, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 544, + 155 + ], + "type": "table", + "html": "
MethodZero-ShotPre-training VideosFID (↓)
T2V [42]XX82.13
SC [5]XX33.51
TFGAN [5]XX31.76
NUWA [78]X0.97M28.46
Phenaki [67]✓15M37.74
Ours✓10M37.23
", + "image_path": "033a29263103d726d2c5ec699f52e16918809d199c747feab0ef9641edbb84e3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 307, + 192, + 544, + 297 + ], + "blocks": [ + { + "bbox": [ + 305, + 157, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 157, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 545, + 190 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art methods on Kinetics-400. Both under a zero-shot setting, our FID is better than that of Phenaki, with only " + }, + { + "bbox": [ + 305, + 157, + 545, + 190 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 305, + 157, + 545, + 190 + ], + "type": "text", + "content": " pre-training videos." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 192, + 544, + 297 + ], + "lines": [ + { + "bbox": [ + 307, + 192, + 544, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 192, + 544, + 297 + ], + "type": "table", + "html": "
MethodZero-ShotCLIPSIM (↑)CLIP-FID (↓)
GODIVA [77]✘0.2402-
NüWA [78]✘0.243947.68
CogVideo (Chinese) [35]✓0.261424.78
CogVideo (English) [35]✓0.263123.59
Video LDM [8]✓0.2929-
Make-A-Video [57]✓0.304913.17
PYoCo [25]✓-10.21
Ours✓0.290812.67
", + "image_path": "2068ed2dfa66722521dcab8d9c6b48106e0ab613ed2d249a1b8300633d62cfa2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 299, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 332 + ], + "type": "text", + "content": "Table 3. Comparison with the state-of-the-art methods on MSR-VTT. PYoCo [25] sample 59,794 videos for CLIP-FID (noted in gray) while others only sample less than 3,000 videos." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 342, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 545 + ], + "type": "text", + "content": "UCF-101 contains videos of 101 action categories. Following [66], we use FVD to evaluate the distance between the generated video and the real video. Kinetics-400 is a dataset comprising videos of 400 action categories. Following [5, 42, 67, 78], we use FID [30] to assess the performance of video generation. MSR-VTT is a video dataset with open-vocabulary captions, wherein CLIPSIM [77] and CLIP-FID [47] are commonly employed for evaluating the T2V generation. Moreover, since these existing benchmarks either have a small number of testing videos or contain only action labels without complex descriptions, we propose to collect an evaluation benchmark for ablation studies. It is called as Vimeo11k, where we collect 11,293 open-world videos along with their captions from 10 mainstream categories in Vimeo. To our best knowledge, it is the largest testing benchmark for zero-shot video generation. More details and experiments can be found in the supplementary doc." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "Implementation Details. (1) Director & Script & Actor & Voicer. We choose GPT-4 [46] as our LLM director to generate script. The specific instructions can be found in the supplementary doc. Then we employ Stable Diffusion XL [48] and Bark [1] as our designer and voicer to generate reference actor images and convert scripts into speech. (2) ShowMaker. We choose SD-1.4 [52] as our base model, and follow [57] to add temporal self attention. Then, we add our temporal text cross attention on top of each temporal self attention. We expand the input channel of the conv-in layer of U-Net from 4 to 9, so that model can take the concatenated feature in Eq. (15) as input. We use zero initialization for newly added channels. We use CLIP ViT-L/14 [50] as text encoder " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon_{T}" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": ", VQVAE [20] as autoencoder" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8811" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 273, + 193 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 273, + 193 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 273, + 193 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 273, + 193 + ], + "type": "image", + "image_path": "138b28faf8d1c260ede75d2a4290bff96e4e216c06206a80870633c8c5e729dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 194, + 286, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 286, + 216 + ], + "type": "text", + "content": "Figure 5. Long video generation. Comparison on UCF-101 (1000 frames). The lower FVD, the better generation performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 218, + 286, + 267 + ], + "blocks": [ + { + "bbox": [ + 48, + 218, + 286, + 267 + ], + "lines": [ + { + "bbox": [ + 48, + 218, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 286, + 267 + ], + "type": "table", + "html": "
MethodCLIP-I (↑)CLIP-T (↑)Vlog Duration (↑)
w/o autoregressive [91]0.56830.27521min03s
only autoregressive [67]0.56420.25353min58s
Ours0.62940.27563min58s
", + "image_path": "aeb7ecc62ee8f684e38e81aefed1075bdbe96901c8641243a4bfab96b83cb189.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": "consisting of " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": ". Besides, we use OpenCLIP ViT-H/14 [38] as image encoder " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\varepsilon_{I}" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": ", and add spatial image cross attention in the STEB block. The diffusion step " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": " is set to 1000 as [52]. For training, we set the parameter of mode selection as " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "m = 6" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": " in Eq. (13), and " + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 304, + 286, + 411 + ], + "type": "text", + "content": " is set to 1. Following [19, 33, 34, 67, 76], we employ publicly available image dataset (i.e., Laion400M [55]) and video dataset (i.e., WebVid10M [3]) for joint training. More training details can be found in the supplementary doc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 418, + 224, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 224, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 224, + 431 + ], + "type": "text", + "content": "4.1. Comparison with state-of-the-art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "text", + "content": "Tab. 1 shows that, no matter whether the input text is the class label or hand-crafted prompt, our method achieves the best FVD performance of zero-shot video generation on UCF-101. Tab. 2 shows that, compared to Phenaki [67], our method achieves a better FID performance of zero-shot setting on Kinetics-400, but only using " + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "text", + "content": " training videos. Furthermore, our generated videos have a resolution of " + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "inline_equation", + "content": "320 \\times 512" + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "text", + "content": ", which is higher than Phenaki's " + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 437, + 286, + 555 + ], + "type": "text", + "content": ". Tab. 3 shows that, our method achieves remarkably competitive FID and CLIPSIM performance on MSR-VTT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 286, + 652 + ], + "type": "text", + "content": "Furthermore, Fig. 5 illustrates that, we significantly surpass TATS [24] (i.e., the only open-source long video generation model within our knowledge), for generating 1000-frame videos on UCF-101. Moreover, our method does not encounter the issue of TATS, where the video quality continuously declines as the number of frames increases. It is noteworthy that, our method achieves this performance by zero-shot generation, without any finetuning on UCF-101." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 140, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 140, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 140, + 672 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Vlog Generation Process. Tab. 4 presents a comparison between different generation processes. For a fair comparison, we use the same scripts and employ our ShowMaker" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 68, + 533, + 192 + ], + "blocks": [ + { + "bbox": [ + 316, + 68, + 533, + 192 + ], + "lines": [ + { + "bbox": [ + 316, + 68, + 533, + 192 + ], + "spans": [ + { + "bbox": [ + 316, + 68, + 533, + 192 + ], + "type": "image", + "image_path": "efc26d8a5e467255e0cf3a20598a6737b9421e2d25a3c27942748fea7e853f67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "type": "text", + "content": "Figure 6. Ablation for spatial image cross attention. The values around a point on the curve correspond to the " + }, + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "type": "text", + "content": ". While " + }, + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 306, + 194, + 545, + 227 + ], + "type": "text", + "content": " means the original network without spatial image cross attention." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 232, + 526, + 282 + ], + "blocks": [ + { + "bbox": [ + 47, + 271, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 271, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 286, + 293 + ], + "type": "text", + "content": "Table 4. Ablation for generation process. [91] cannot generate a long video in a single scene and both [67, 91] can't refer to images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 232, + 526, + 282 + ], + "lines": [ + { + "bbox": [ + 323, + 232, + 526, + 282 + ], + "spans": [ + { + "bbox": [ + 323, + 232, + 526, + 282 + ], + "type": "table", + "html": "
TTCAMTP0/16 (↓)1/15 (↓)3/13 (↓)5/11 (↓)
✗✗348.36217.39297.37246.56
✓✗333.63178.27230.56193.52
✓✓257.70123.51118.02109.31
", + "image_path": "3189e3a083c1a4fb2c73c8a3a1cc29cea76cb30815d51297714f7b39831cfff5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 287, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 305, + 287, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 545, + 364 + ], + "type": "text", + "content": "Table 5. Ablation for temporal text cross attention and mixed training paradigm. \"TTCA\" and \"MTP\" refer to temporal text cross attention and mixed training paradigm respectively. \"3/13\" denotes predicting the following 13 frames after being given 3 frames of a ground truth video, and we evaluate the FVD between the newly generated 13 frames and the corresponding frames in the ground truth. \"0/16\", \"1/15\", \"5/11\" follow this pattern." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 374, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 530 + ], + "type": "text", + "content": "as videographer. We compare three approaches including no autoregression at all as MovieFactory [91], fully autoregressive processes in one go as Phenaki [67], and the generation process of our Vlogger. Specifically, We use GPT-4 to generate five stories and go through the Vlogger's planning process to get five vlog scripts. We set 20 different random seeds, so that each of the above generation processes generates 20 different vlogs for each script. We use CLIP-I, CLIP-T [82] and video duration to evaluate the quality of the generated vlogs. The results demonstrate that, even using the same script and videographer, our Vlogger outshines other existing frameworks for preferable vlog generation. More details can be found in the supplementary doc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "type": "text", + "content": "Spatial-Temporal Enhanced Block. Fig. 6 and Tab. 5 evaluate two important operations in our STEB block, i.e., spatial image cross attention and temporal text cross attention. With spatial image cross attention in Fig. 6, the CLIP-I is significantly improved as " + }, + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "type": "text", + "content": " increased, while the CLIP-T reached its maximum at " + }, + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 304, + 532, + 545, + 639 + ], + "type": "text", + "content": " on COCO2017 [43]. With temporal text cross attention (TTCA) in Tab. 5, both T2V generation (0/16) and prediction (1/15, 3/13 and 5/11) show a significant improvement on Vineo11k." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Mixed Training Paradigm. Tab. 5 also presents a model comparison by incorporating our mixed training paradigm or the random mask training method in [11, 14, 21, 36, 68, 84]. The results show that the mixed training paradigm is effective in enhancing the model's T2V generation and prediction capabilities. More details in the supplementary doc." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8812" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 69, + 532, + 89 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 69, + 532, + 89 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 532, + 89 + ], + "type": "text", + "content": "Lots of traffic in futuristic city. ... The camera moves away from the astronaut. The astronaut leaves the keyboard and walks to the left. ... We are in an office room with empty desks. A lion runs on top of the office desks. ... Timelapse of sunset in the modern city." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 90, + 532, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 90, + 532, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 90, + 532, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 90, + 532, + 220 + ], + "type": "image", + "image_path": "a87289cd4d58944740a66d11ac1a26019ba3e80ace55bc9d7c582d7b82fa5b56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 221, + 545, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 221, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 545, + 244 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with the state-of-the-art methods on long video generation. The story and long video of Phenaki [67] are available at phenaki.github.io. The scene diversity and picture quality generated by Vlogger are significantly better than Phenaki." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 247, + 532, + 479 + ], + "blocks": [ + { + "bbox": [ + 57, + 247, + 532, + 479 + ], + "lines": [ + { + "bbox": [ + 57, + 247, + 532, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 532, + 479 + ], + "type": "image", + "image_path": "91ddcd3217c96cbde8e784ce97c5f06cd4f4437bf8482916a1b819e1efda8449.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 481, + 546, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 481, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 546, + 504 + ], + "type": "text", + "content": "Figure 8. Qualitative ablation for STEB and training paradigm. \"TTCA\" and \"MTP\" refer to temporal the text cross attention and mixed training paradigm respectively. The sample without a visual prompt corresponds to the model without spatial image cross attention." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 513, + 130, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 130, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 130, + 525 + ], + "type": "text", + "content": "4.3. Visualization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 530, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 638 + ], + "type": "text", + "content": "First, we visualize long video generation, by comparison with the well-known Phenaki [67]. Note that, since Phenaki does not have the open-sourced codes, we alternatively use the demo shown in its official website. Specifically, we feed the same story description into our Vlogger. As shown in Fig. 7, our Vlogger shows superior and more diverse video content, compared to Phenaki. Moreover, our Vlogger sufficiently exhibits the story with a much longer duration (i.e., 7min59s), according to our LLM-created script." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "We further visualize the ablation of T2V video generation and prediction by ShowMaker. As depicted in Fig. 8, ShowMaker had significant improvements in generation and prediction performance, by incorporating our ShowMaker designs. Additionally, it can leverage visual prompts in the spatial actor cross attention, for distinguishing the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 513, + 459, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 459, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 459, + 526 + ], + "type": "text", + "content": "\"Teddy\" concept within text prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "content": "In this paper, we introduce a generic system Vlogger, to generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor. Moreover, we present a novel video diffusion model ShowMaker for boosting state-of-the-art T2V generation and prediction. Finally, we will release all the models, data and codes afterward, allowing to develop further designs toward long video generation in the open world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Key R&D Program of China (NO.2022ZD0160102), the National Natural Science Foundation of China under Grant No. 62102150, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8813" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Suno AI. bark. https://github.com/suno-ai/bark#-usage-in-cython, 2023.2,4,6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 286, + 246 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenhang Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, K. Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Yu Bowen, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xing Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. ArXiv, abs/2309.16609, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 248, + 286, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 286, + 280 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 282, + 286, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 286, + 348 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 286, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 286, + 393 + ], + "type": "text", + "content": "[5] Yogesh Balaji1, Martin Renqiang Min, Bing Bai, Rama Chellappa1, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 286, + 452 + ], + "type": "text", + "content": "[6] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 286, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 286, + 496 + ], + "type": "text", + "content": "[7] Sarthak Bhagat, Shagun Uppal, Vivian Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In ECCV, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 499, + 286, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 286, + 542 + ], + "type": "text", + "content": "[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 544, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 286, + 588 + ], + "type": "text", + "content": "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A. Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 286, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 699 + ], + "type": "text", + "content": "[10] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "text", + "content": "[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[12] Haoran Chen, Jianmin Li, Simone Frintrop, and Xiaolin Hu. The msr-video to text dataset with clean annotations. Comput. Vis. Image Underst., 225:103581, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "text", + "content": "[13] Xinyuan Chen, Chang Xu, Xiaokang Yang, and Dacheng Tao. Long-term video prediction via criticization and retrospection. IEEE TIP, 29:7090-7103, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 216 + ], + "type": "text", + "content": "[14] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. ArXiv, abs/2310.20700, 2023. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 239 + ], + "type": "text", + "content": "[15] François Chollet. On the measure of intelligence. ArXiv, abs/1911.01547, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 240, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 479 + ], + "type": "text", + "content": "[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. JMLR, 24:240:1-240:113, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 481, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 514 + ], + "type": "text", + "content": "[17] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv: Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 515, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 612 + ], + "type": "text", + "content": "[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "text", + "content": "[19] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay P. Namboodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. WACV, pages 3017-3031, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[20] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[21] Tsu-Jui Fu, Licheng Yu, N. Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8814" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[22] Wen Gao, Yonghong Tian, Tiejun Huang, and Qiang Yang. Vlogging: A survey of videoblogging technology on the web. ACM Computing Surveys, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 131, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 131, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 131, + 287, + 175 + ], + "type": "text", + "content": "[23] Songwei Ge, Thomas Hayes, Harry Yang, Xiaoyue Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "type": "text", + "content": "[24] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 277 + ], + "type": "text", + "content": "[25] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. ArXiv, abs/2305.10474, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "text", + "content": "[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 314, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 367 + ], + "type": "text", + "content": "[27] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. Animate-a-story: Storytelling with retrieval-augmented video generation. ArXiv, abs/2307.06940, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "type": "text", + "content": "[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. ArXiv, abs/2211.13221, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "text", + "content": "[29] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Frederick Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. ArXiv, abs/1712.00409, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "type": "text", + "content": "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 507, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 507, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 507, + 287, + 529 + ], + "type": "text", + "content": "[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 531, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 552 + ], + "type": "text", + "content": "[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[33] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. ArXiv, abs/2204.03458, 2022. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In ICLR, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[36] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. TMLR, 2022, 2022. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[37] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. ArXiv, abs/2309.14494, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "text", + "content": "[38] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, 2021.7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[39] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Apostol Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. ArXiv, abs/1705.06950, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "text", + "content": "[40] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "type": "text", + "content": "[41] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. ArXiv, abs/2305.06355, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "text", + "content": "[42] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 422, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 443 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "type": "text", + "content": "[46] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[47] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 558, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 601 + ], + "type": "text", + "content": "[48] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "text", + "content": "[49] Haonan Qiu, Menghan Xia, Yong Zhang, Yin-Yin He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv, abs/2310.15169, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8815" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "text", + "content": "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 286, + 148 + ], + "type": "text", + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 286, + 183 + ], + "type": "text", + "content": "[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. ArXiv, abs/1505.04597, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 286, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 286, + 259 + ], + "type": "text", + "content": "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 286, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 286, + 315 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 316, + 286, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 316, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 316, + 286, + 392 + ], + "type": "text", + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 393, + 286, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 393, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 286, + 447 + ], + "type": "text", + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 449, + 286, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 286, + 481 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 286, + 481 + ], + "type": "text", + "content": "[58] Ivan Skorokhodov, S. Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 482, + 286, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 286, + 525 + ], + "type": "text", + "content": "[59] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 526, + 286, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 526, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 49, + 526, + 286, + 558 + ], + "type": "text", + "content": "[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. ArXiv, abs/1212.0402, 2012. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 559, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 286, + 646 + ], + "type": "text", + "content": "[61] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, Weixin Liu, Zhihua Wu, Weibao Gong, Jianzhong Liang, Zhizhou Shang, Peng Sun, Wei Liu, Ouyang Xuan, Dianhai Yu, Hao Tian, Hua Wu, and Haifeng Wang. Ernie 3.0: Large-scale knowledge enhanced pretraining for language understanding and generation. ArXiv, abs/2107.02137, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 286, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 680 + ], + "type": "text", + "content": "[62] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 713 + ], + "type": "text", + "content": "[63] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. Motion-based generator model: Unsupervised disentanglement of appearance," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "trackable and intractable motions in dynamic patterns. In ICLR, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[64] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and S. Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "text", + "content": "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. *ArXiv*, abs/2302.13971, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. In ICLR, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[67] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "text", + "content": "[68] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "text", + "content": "[69] Fu Lee Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. ArXiv, abs/2305.18264, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "text", + "content": "[70] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. ArXiv, abs/2308.06571, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[71] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. ArXiv, abs/2305.10874, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "type": "text", + "content": "[72] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. G3an: Disentangling appearance and motion for video generation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 502, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 544 + ], + "type": "text", + "content": "[73] Yaohui Wang, Piotr Tadeusz Bilinski, François Brémond, and Antitza Dantcheva. Imaginar: Conditional spatiotemporal gan for video generation. WACV, pages 1149-1158, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "type": "text", + "content": "[74] Yaohui Wang, François Bremond, and Antitza Dantcheva. Inmodegan: Interpretable motion decomposition generative adversarial network for video generation. ArXiv, abs/2101.03049, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "type": "text", + "content": "[75] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. *Internvideo: General video foundation models via generative and discriminative learning.* ArXiv, abs/2212.03191, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[76] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Pe der Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Y. Qiao, and Ziwei Liu. Lavie: High-quality video" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8816" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "generation with cascaded latent diffusion models. ArXiv, abs/2309.15103, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[77] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. ArXiv, abs/2104.14806, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[78] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "text", + "content": "[79] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "text", + "content": "[80] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 286, + 308 + ], + "type": "text", + "content": "[81] Wilson Yan, Yunzhi Zhang, P. Abbeel, and A. Srinivas. Videogpt: Video generation using vq-vae and transformers. ArXiv, abs/2104.10157, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 311, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 287, + 342 + ], + "type": "text", + "content": "[82] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 345, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 345, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 345, + 287, + 420 + ], + "type": "text", + "content": "[83] Sheng-Siang Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation. In Annual Meeting of the Association for Computational Linguistics, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 465 + ], + "type": "text", + "content": "[84] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit: Masked generative video transformer. In CVPR, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "text", + "content": "[85] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 577 + ], + "type": "text", + "content": "[86] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, P. Zhang, Yuxiao Dong, and Jie Tang. Glm-130b: An open bilingual pre-trained model. ArXiv, abs/2210.02414, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 632 + ], + "type": "text", + "content": "[87] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lin Hao Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. ArXiv, abs/2309.15818, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[88] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. ArXiv, abs/2305.18583, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[89] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. ArXiv, abs/2211.11018, 2022. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 171 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[90] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[91] Junchen Zhu, Huan Yang, Huiguo He, Wenjing Wang, Zixi Tuo, Wen-Huang Cheng, Lianli Gao, Jingkuan Song, and Jianlong Fu. Moviefactory: Automatic movie creation from text using large generative models for language and images. In ACM MM, 2023. 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8817" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_content_list.json b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44e1d6c9bf99ce080aa4220f342c0d1f916136eb --- /dev/null +++ b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_content_list.json @@ -0,0 +1,1563 @@ +[ + { + "type": "text", + "text": "VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis", + "text_level": 1, + "bbox": [ + 86, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linshan Wu", + "bbox": [ + 145, + 204, + 246, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxin Zhuang", + "bbox": [ + 431, + 204, + 549, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Chen *", + "bbox": [ + 735, + 204, + 828, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hong Kong University of Science and Technology", + "bbox": [ + 290, + 222, + 689, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-Supervised Learning (SSL) has demonstrated promising results in 3D medical image analysis. However, the lack of high-level semantics in pre-training still heavily hinders the performance of downstream tasks. We observe that 3D medical images contain relatively consistent contextual position information, i.e., consistent geometric relations between different organs, which leads to a potential way for us to learn consistent semantic representations in pre-training. In this paper, we propose a simple yet-effective Volume Contrast (VoCo) framework to leverage the contextual position priors for pre-training. Specifically, we first generate a group of base crops from different regions while enforcing feature discrepancy among them, where we employ them as class assignments of different regions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which region) by contrasting their similarity to different base crops, which can be seen as predicting contextual positions of different sub-volumes. Through this pretext task, VoCo implicitly encodes the contextual position priors into model representations without the guidance of annotations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experimental results on six downstream tasks demonstrate the superior effectiveness of VoCo. Code will be available at https://github.com/Luffy03/VoCo.", + "bbox": [ + 75, + 323, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 744, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has demonstrated outstanding achievements in 3D medical image analysis [53, 21, 40, 33, 39], yet is heavily hampered by the expensive cost of the required expert annotations [50, 23]. To address this problem, Self-Supervised Learning (SSL) has received significant attention due to its promising ability to learn representations without annotations [10, 11, 6, 28, 20], which has become", + "bbox": [ + 75, + 771, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/864fec36362eafd2c1ad151515fa3a58137346649cea1e8cb91415b5af2a8e16.jpg", + "image_caption": [ + "(a) Relatively consistent contextual positions in 3D medical images" + ], + "image_footnote": [], + "bbox": [ + 508, + 289, + 885, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abde1250c2e2349e645a2e82439a7d4c9b3b032678330d1dd9fdab3f990dc6d3.jpg", + "image_caption": [ + "(b) Volume Contrast (VoCo) for 3D Medical Image Analysis", + "Figure 1. (a) In 3D medical images, the contextual positions, i.e., the geometric relations between different organs are relatively consistent. (b) Aiming to leverage contextual position priors for pretraining, we proposed a Volume Contrast (VoCo) framework for 3D Medical Image Analysis." + ], + "image_footnote": [], + "bbox": [ + 516, + 375, + 887, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an important label-efficient solution in 3D medical image analysis [71, 51, 32, 2, 34, 36].", + "bbox": [ + 498, + 595, + 890, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods [50, 75, 71, 13] are mostly based on information reconstructions to learn augment-invariant representations of 3D medical images, which first employ strong data augmentation to the images and then reconstruct the raw information. Specifically, rotate-and-reconstruct [50, 51, 75, 52] proposed to randomly rotate the 3D volumetric images and learn to recover them, which encourages models to learn rotational invariant features. Recent methods [70, 71, 32, 25, 62] further proposed to restore information among different views of the image. PCRL [70, 71] cropped global and local patches then conducted multiscale restorations. GVSL [32] further explored the geometric similarity between multi-scans by affine augmentation and matching. Mask-reconstruct methods [13, 73, 55] are also widely used, which are introduced from MAE [28] and aim to learn representations by masking images and reconstructing the missing pixels. Although promising results have been demonstrated, previous works [52, 32] have", + "bbox": [ + 496, + 628, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: jhc@cse.ust.hk", + "bbox": [ + 94, + 886, + 307, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22873", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1fbbea86f9466374034fabf6a28da9cf19a0135a7f820aa2b9593741e46bafc5.jpg", + "image_caption": [ + "(a) Instance-level" + ], + "image_footnote": [], + "bbox": [ + 158, + 90, + 361, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79a7a20cd8fab64dad775dc75401d3df656deea6b5e6faf3ce22de6ab4626638.jpg", + "image_caption": [ + "(b) Prototype-level", + "Figure 2. Typical contrastive learning frameworks. (a) Instance-level contrastive learning [10, 11, 29, 22, 7] employs strong data augmentation or model perturbation on input data to acquire different views of instance, then regularizes their consistency. (b) Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] conducts (1) online clustering or (2) randomly initialize then online update process to obtain prototypes as class assignments, then leverage the prototypes to contrast each input image. (c) Our VoCo follows the idea of prototype-level contrastive learning. Specifically, instead of using time-consuming online clustering and updating procedures, we leverage the valuable contextual position priors of 3D medical images and leverage the base crops to generate prototypes (bases)." + ], + "image_footnote": [], + "bbox": [ + 367, + 90, + 602, + 214 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/80fbd5499995cb2ac570b093809e1502ef0fb9b200e7b699afbe748d23a3280e.jpg", + "image_caption": [ + "(c) Our VoCo" + ], + "image_footnote": [], + "bbox": [ + 604, + 90, + 813, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proved that the lack of high-level semantics in pre-training will heavily hinder the performance of downstream tasks. To address this challenge, we argue that stronger high-level semantics should be further involved into 3D medical image pre-training.", + "bbox": [ + 75, + 338, + 468, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we argue that the contextual position priors of 3D medical images should be further exploited. As shown in Fig. 1(a), we observe that in 3D medical images, different organs (semantic regions) contain relatively consistent contextual positions with relatively consistent anatomic characteristics (shapes). Thus, the consistency of geometric relations between different organs leads to a potential way for us to learn consistent semantic representations for 3D medical images pre-training. In this paper, we propose a pretext task for contextual position predictions, which aims to encode contextual position priors into model representations and enables us to effectively improve the performance of downstream tasks that require high-level semantics.", + "bbox": [ + 75, + 415, + 470, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple-yet-effective Volume Contrast (VoCo) framework for 3D medical image analysis, as shown in Fig. 1(b). Specifically, we first crop a group of non-overlap volumes from different positions while enforcing feature discrepancy among them. We represent these volumes as a group of bases in the learned high-dimension space, where we employ them as class assignments of different positions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which position) by contrasting their similarity to different bases, which can be seen as predicting contextual positions of different sub-volumes. In this way, we formulate a contextual position prediction pretext task for 3D medical image SSL. Through learning to predict contextual positions, we implicitly involve the high-level semantic priors into the model representations, which enables us to significantly improve the performance of downstream tasks. Extensive experimental results on six downstream tasks demonstrate that our", + "bbox": [ + 75, + 628, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proposed VoCo clearly outperforms existing state-of-the-art 3D medical image SSL methods.", + "bbox": [ + 498, + 338, + 890, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 386, + 648, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the previous mainstream contrastive learning paradigms. Then, we survey the existing SSL methods for medical image analysis, especially for 3D medical images. Finally, we review the position-related SSL methods for comparisons with our method and highlight the differences.", + "bbox": [ + 496, + 412, + 890, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning. Contrastive learning is one of the mainstream paradigms in SSL, which aims to learn consistent representations by contrasting positive and negative pairs of samples without extra annotations [10, 6, 29, 11]. According to [6], instance- and prototype-level contrastive learning are two typical types of contrastive learning, as shown in Fig. 2.", + "bbox": [ + 496, + 505, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instance-level contrastive learning [10, 11, 29, 22, 7] transforms input images with different augmentations or model perturbations, aiming to compare the features from each other. Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] proposes to generate prototypes (also called clusters or bases) for contrasting each input image. Specifically, there are two typical ways to generate prototypes. First, Caron et al. proposed DeepCluster [5] to conduct online clustering on the whole dataset to generate prototypes. However, it is very time-consuming to calculate clusters on a large dataset. Thus, some recent works [6, 56, 15, 16] propose to randomly initialize a group of prototypes and then update them through back-propagation during training, which has demonstrated promising results. However, there is still no explicit guarantee that these randomly initialized prototypes can be updated well during training.", + "bbox": [ + 496, + 612, + 892, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our VoCo follows the primary idea of prototype-level constrasive learning. As shown in Fig. 2(c), to address the existing problems mentioned above, instead of randomly", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22874", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "initializing and updating prototypes, VoCo leverages the valuable contextual position priors of 3D medical images to generate base crops as prototypes, which also requires no time-consuming clustering on a large dataset.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SSL for medical image analysis. Due to the high potential in label-efficient learning [29, 59, 61, 58, 57, 60, 37], SSL has also received significant attention in the field of medical image analysis [70, 32, 31, 51, 19]. Existing methods are mainly based on comparative SSL [71]. Specifically, Zhou et al. [69] combined Mixup [66] into MoCo [29] to learn the diversity of positive and negative samples in InfoNCE [44]. Azizi et al. used multi-instance learning to compare multiple views of images from each patient. There are also a number of approaches [25, 70, 71] that supervising the models via restoring low-level information from raw images.", + "bbox": [ + 75, + 161, + 468, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In 3D medical image analysis, reconstructing raw information is a popular pretext task for learning representations [50, 51, 71]. Existing methods are mainly based on reconstructing information from augmented images. These previous methods first conducted strong data augmentation, e.g., rotate [51, 75, 52], multi-view crops [70, 71, 32], and mask [13, 73, 55], then supervised the model by reconstructing raw 3D information. Although promising results have been demonstrated, most of these methods still largely ignore the importance of integrating high-level semantics into model representations, which heavily hinders the performance of downstream tasks.", + "bbox": [ + 75, + 352, + 468, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Position-related SSL. Position-related SSL methods are also explored in a number of previous works [8, 9, 41, 47, 43, 17, 64, 68] in the field of natural images. Noroozi et al. [43] proposed to predict the order of a set of shuffled patches. Zhai et al. [64] and Caron et al. [8] proposed to train a ViT [18] to predict the locations of each input patch. However, since the geometric relations of different objects are not very consistent in natural images, it is still difficult to effectively learn consistent position representations given visual appearance only (as stated in [68]). In addition, previous works [64, 8, 68] mainly trained a linear layer to output the positions directly, which works in a black-box manner.", + "bbox": [ + 75, + 542, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce the pretext task of contextual position prediction into the field of 3D medical images, where the geometric relations between different organs are relatively consistent, which guides us to learn consistent semantic representations in pre-training. Different from the previous methods, in this paper, we introduce a totally different position prediction paradigm. Specifically, instead of using a linear layer to output positions directly, we predict the contextual positions based on volume contrast, which is more intuitive and effective.", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 89, + 635, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the overall framework of our proposed VoCo in Section 3.1. After that, we present the process of contextual position prediction in Section 3.2. Then, the regularization process via volume contrast in our proposed VoCo framework is described in Section 3.3.", + "bbox": [ + 496, + 114, + 893, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 498, + 202, + 689, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of our proposed VoCo is presented in Fig. 3, which contains a contextual position prediction branch and a regularization branch. The prediction branch is used to predict the contextual positions between different cropped volumes. Specifically, given an input volume, we first crop it into non-overlap base volumes, which cover the whole input volume. Then, we randomly crop a volume and transform it into the high-dimension feature space using a typical backbone (CNN [30] or Transformer [18]). The goal is to predict the contextual positions between the randomly cropped volumes and base volumes. In this paper, instead of training a linear classifier to predict positions as in previous works [64, 9, 8, 68], we propose to establish this goal by volume contrast. We develop a loss function $L_{pred}$ to supervise the final predictions. In addition, we further use a loss function $L_{reg}$ to regularize the feature discrepancy from different bases by enlarging their distance, aiming to learn more discriminative class assignments. The details are presented in Section 3.2 and 3.3.", + "bbox": [ + 496, + 226, + 892, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Contextual Position Prediction", + "text_level": 1, + "bbox": [ + 498, + 523, + 769, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Base and random crops. Given an input volume, we first crop it into $n$ non-overlap base volumes, which cover the whole input volume. We then employ the extracted features $z$ as class assignments (we call them bases), which present the prototype-level features from different positions. Then, following previous SSL works [10, 11, 29], a projector with linear layers is used to project $z$ into latent features $q$ . Then, we randomly crop a volume and transform it into high-dimension feature space as $p$ . The backbone and projector are also used to project the features from the randomly cropped volumes.", + "bbox": [ + 496, + 547, + 890, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Volume contrast for contextual position prediction. With features extracted from the backbone and projector, following previous SSL works [10, 11, 29], we first conduct 3D adaptive average pooling to resize them to one dimension, i.e., $p \\in \\mathbb{R}^{1 \\times C}$ and $q \\in \\mathbb{R}^{1 \\times C}$ , where $C$ is the number of channels. Specifically, we empirically set $C$ to 2048 as in [10, 11, 29].", + "bbox": [ + 496, + 714, + 890, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we calculate the similarity logits $l$ between $p$ and $q_{i}$ . Specifically, we use cosine similarity to compute $l$ as follows:", + "bbox": [ + 496, + 820, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} = \\operatorname {C o s S i m} (p, q _ {i}) = \\frac {p \\cdot q _ {i}}{\\| p \\| \\| q _ {i} \\|}, i \\in n \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 875, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7f4b4b75e18847a633dde5f16f7d9d40aab1d9327b86900306c6e446583248b9.jpg", + "image_caption": [ + "Figure 3. The overall framework of VoCo. VoCo contains a prediction branch and a regularization branch. The prediction branch is responsible for predicting contextual positions between different sub-volumes. The regularization branch is employed to enforce the feature discrepancy between different bases, which aims to learn more discriminative class assignments." + ], + "image_footnote": [], + "bbox": [ + 158, + 85, + 802, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{i}$ is the projected feature of each base crop. $l_{i}$ denotes the similarity between $p$ in $q_{i}$ , which ranges from 0 to 1. It is worth noting that, we stop the gradients of $q$ when computing Eq. 1, which aims to avoid feature collapse [10, 11, 6].", + "bbox": [ + 75, + 424, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, higher $l_{i}$ represents that $p$ has higher probabilities to share overlap regions with $q_{i}$ . In this way, we can explicitly associate the similarity value with the position information, i.e., $p$ with higher $l_{i}$ is more likely to be located in the region of the $i_{th}$ base. Thus, instead of training a black-box linear layer, we predict the contextual positions by volume contrast, which is more intuitive and effective.", + "bbox": [ + 75, + 503, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position labels generation. The process of generating position labels is shown in Fig. 4. As shown in Fig. 4, when we generate $n = 4 \\times 4$ base crops, there will be $n$ class assignments. Then we calculate the overlap area between a randomly cropped volume and $n$ base crops. The proportions of the overlap area are then assigned as position labels $y$ , which also range from 0 to 1. Thus, we can easily supervise the model by calculating the distance between the prediction logits $l$ and position labels $y$ . The setting of the number $n$ of base crops will be discussed in Section 4.4.", + "bbox": [ + 75, + 613, + 470, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss function for contextual position prediction. The formulation of prediction loss function $L_{pred}$ is based on entropy. Specifically, we first calculate the distance $d$ between prediction logits $l$ and position labels $y$ :", + "bbox": [ + 75, + 768, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i} = \\left| y _ {i} - l _ {i} \\right|, i \\in n, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 849, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|.|$ denotes the absolute value. Then, $L_{pred}$ is formu", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/52712bc928b1b014ead126e6371ce21df02539f86b5fabfa38b6e3e7eaacc0b1.jpg", + "image_caption": [ + "Figure 4. The process of generating position labels. We calculate the proportions of overlap area as position labels $y$ , e.g., the randomly cropped volume in the figure is assigned to class 5, 6, 9, and 10 with probabilities of 0.25, 0.1, 0.5, and 0.15, respectively." + ], + "image_footnote": [], + "bbox": [ + 540, + 421, + 846, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lated as follows:", + "bbox": [ + 500, + 660, + 609, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p r e d} = - \\frac {1}{n} \\sum_ {i \\in n} ^ {n} \\log \\left(1 - d _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 683, + 890, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that VoCo predicts contextual positions of a volume (high similarities with all its contextual overlapped volumes), thus don't need one-to-one correspondence: e.g., in Fig. 4, high-value $l_{i}$ pertain to $y_{i} > 0, i = 5,6,9,10$ simultaneously. Then we calculate the distance between $l_{i}$ and $y_{i}$ (Eq. 2).", + "bbox": [ + 498, + 732, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Volume Contrast for Regularization", + "text_level": 1, + "bbox": [ + 500, + 832, + 810, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to learn more discriminative class assignments (bases) for volume contrast. Since intuitively, different subvolumes tend to contain different organs (semantic discrepancy).", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22876", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ancy). Thus, we aim to enlarge the high-dimension feature discrepancy between different bases. To this end, we design a regularization loss $L_{reg}$ to enlarge the feature discrepancy between different bases $z$ .", + "bbox": [ + 75, + 90, + 467, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, given projected bases $q$ , we also calculate the cosine similarity $s_{ij}$ between different $q_i$ and $q_j$ as follows:", + "bbox": [ + 76, + 152, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} = \\operatorname {C o s S i m} \\left(q _ {i}, q _ {j}\\right) = \\frac {q _ {i} \\cdot q _ {j}}{\\| q _ {i} \\| \\| q _ {j} \\|}, i, j \\in n, i \\neq j, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 191, + 467, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we aim to regularize $s_{ij}$ to 0, enforcing feature discrepancy between different bases. Thus, the loss function $L_{reg}$ is formulated as:", + "bbox": [ + 76, + 232, + 467, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e g} = \\frac {2}{n (n - 1)} \\sum_ {i, j \\in n, i \\neq j} ^ {n} | s _ {i j} |, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 289, + 467, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|.|$ denotes the absolute value. With loss $L_{reg}$ , we aim to optimize $q$ as linearly independent bases:", + "bbox": [ + 76, + 340, + 467, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} \\bot q _ {j}, i \\neq j, i, j \\in n. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 383, + 467, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the regularization loss function $L_{reg}$ , we aim to learn a group of linearly independent bases to represent all directions of high-dimension features [6]. In this way, we can learn a group of more discriminative class assignments to supervise the final position predictions.", + "bbox": [ + 76, + 411, + 467, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall loss function. Thus, the total loss function $L$ is the combination of $L_{reg}$ and $L_{pred}$ :", + "bbox": [ + 76, + 487, + 467, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {p r e d} + \\lambda L _ {r e g}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 530, + 467, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is used to balance the relative contributions of these two loss terms and set to 1.0 in experiments empirically since we consider their importance equally. The ablation studies of $\\lambda$ are provided in the supplementary materials.", + "bbox": [ + 76, + 556, + 467, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 631, + 207, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the datasets used in the pre-training and downstream tasks. Then, we briefly introduce the implementation details of VoCo. Finally, we report detailed experiment results of our proposed VoCo compared with other state-of-the-art SSL methods in 3D medical images. More details are in the supplementary materials.", + "bbox": [ + 76, + 656, + 467, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 756, + 178, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training datasets. Aiming to conduct fair comparisons with the previous works [51, 55, 70, 71, 13, 73], we also carry out pre-training experiments on the same three public datasets, i.e., BTCV [35], TCIA Covid19 [14], and LUNA [48] datasets, including about totally $1.6\\mathrm{k}$ CT scans for pre-training. It is worth noting that, aiming to conduct fair comparisons with previous works [73, 13], we only use BTCV [35] and TCIA Covid-19 [14] for pre-training in", + "bbox": [ + 76, + 780, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the downstream experiments of BTCV [35]. For the other downstream tasks, we use all three datasets for pre-training. Details are provided in the supplementary materials.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream datasets. To evaluate the effectiveness of our VoCo, we conduct downstream experiments on six public datasets, i.e., BTCV [35], LiTs [4], MSD Spleen [1], MM-WHS [74], BraTS 21 [49], and CC-CCII [67], including segmentation and classification tasks. The first five datasets are developed for segmentation, while CC-CCII [67] is for COVID-19 classification. Note that only BTCV [35] is used in pre-training, the other datasets are unseen in pre-training. In addition, to evaluate the cross-modality generalization ability, we transfer the model pre-trained on the CT dataset to the MRI dataset BraTS 21 [49]. We adopt consistent settings as previous works [13, 73, 26, 55, 32]. We also evaluate the performance on 2D medical dataset [54]. Details are provided in the supplementary materials.", + "bbox": [ + 498, + 137, + 890, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 359, + 712, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous works [51, 55], we use SwinUNETR [26] for both pre-training and downstream tasks. We use AdamW [38] optimizer and cosine learning rate scheduler for all experiments. We set 100K training steps in the pre-training process and applied a slicing window inference for fair comparisons with previous works [51, 55, 13, 73]. Aiming to evaluate the pure effectiveness, we do not use foundation models or post-processing [34, 36]. Details are provided in the supplementary materials.", + "bbox": [ + 498, + 382, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison methods. We compare our VoCo with both General and Medical SSL methods. First, we compare with the typical SSL methods MAE [28, 13] and MoCo v3 [29, 12], since they represent the two mainstream SSL paradigms, i.e., mask-autoencoder and constrastive learning. Since it is not practical to set a large batch size for 3D medical images due to computation cost, for fair comparisons, we adopt consistent settings with other methods in MAE [28, 13] and MoCo v3 [29, 12]. We also report the results of SimCLR [10] and SimMIM [63] according to [13]. We further evaluate the performance of Jiasaw [9] and PositionLabel [68], since they are related to our position-aware methods. Most existing state-of-the-art medical SSL methods are compared in our experiments.", + "bbox": [ + 498, + 518, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Experiments on downstream tasks", + "text_level": 1, + "bbox": [ + 500, + 741, + 797, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Outperform existing methods on the BTCV dataset. We first conduct experiments on BTCV [35], as shown in Table 1. Specifically, among the comparison methods, MAE3D [28, 13], SimCLR [10], SimMIM [63], MoCo v3 [29, 12], and GL-MAE [73] use UNETR [27]. Other methods including our VoCo adopt Swin-UNETR [26] as the default settings of the previous work [51].", + "bbox": [ + 498, + 763, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark. It can be seen in Table 1 that the general SSL methods perform worse than most medical SSL methods.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22877", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5347f340f77a652b22584b00373f13bf4cb9663e0ca8572f8d0917e7e2123d27.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDice Score(%)AVG
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAG
From Scratch
UNETR [27]93.0294.1394.1266.9970.8796.1177.2789.2282.1070.1676.6565.3259.2179.82
Swin-UNETR† [26]94.0693.5493.8065.5174.6097.0975.9491.8082.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [28, 13]93.9894.3794.1869.8674.6596.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [10]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [63]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3† [29, 12]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw† [9]94.6293.4193.5575.6373.2195.7180.8089.4184.7871.0279.5765.6860.2281.35
PositionLabel† [68]94.3593.1593.2175.3973.2495.7680.6988.8084.0471.1879.0265.1160.0781.09
With Medical SSL
MG [72]91.9993.5291.8165.1176.1495.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [50]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [3]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++† [52]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1† [70]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2† [71]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [26, 51]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [55]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [73]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL† [32]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
", + "bbox": [ + 76, + 88, + 906, + 385 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3bd736bc744be5c5e196c599bb21a3fd6296895e5a45fb5b283a699af478e2d2.jpg", + "table_caption": [ + "Table 1. Experimental results on BTCV [35]. The best results are bolded. 'From Scratch' denotes the supervised baseline without self-supervised pre-training. † denotes we re-implement the approach. Most results are drawn from [13, 65, 73] or their own papers." + ], + "table_footnote": [], + "table_body": "
MethodNetworkDice Score(%)
From Scratch
3D UNet [46]-90.70
UNETR† [27]-93.25
Swin-UNETR† [26]-93.42
With General SSL
Jigsaw [9]3D UNet94.36
MAE3D [28, 13]UNETR94.02
MoCo v3† [29, 12]UNETR93.86
Jigsaw† [9]Swin-UNETR95.24
PositionLabel† [68]Swin-UNETR94.13
With Medical SSL
MG [72]3D UNet91.30
TransVW [24]3D UNet91.42
ROT [50]3D UNet94.49
PCRLv1 [70]3D UNet93.87
PCRLv2 [71]3D UNet94.50
Rubik [75]3D UNet94.93
Rubik++ [52]3D UNet95.46
Rubik++† [52]Swin-UNETR95.72
Swin-UNETR [51, 26]Swin-UNETR95.33
SwinMM [55]Swin-UNETR95.52
VoCo3D UNet96.03
VoCoSwin-UNETR96.52
", + "bbox": [ + 106, + 436, + 442, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, MoCo v3 [29, 12] can only achieve $79.54\\%$ Dice Score. Since MoCo v3 [29, 12] heavily relies on a large batch size to acquire adequate negative samples, which is not practical in 3D medical images due to the huge computation burden. In addition, the negative relation between different images used in MoCo [29, 12] is not ap", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2853e1a7ade31cc86b7df1644bcd50855c616ccc8b5dd1dc3c75009ad7ab4a4.jpg", + "table_caption": [ + "Table 2. Experimental results on LiTs [4]. We report the Dice Scores of liver segmentation. $\\dagger$ denotes we re-implement the approach. Most results are drawn from [65, 71]." + ], + "table_footnote": [], + "table_body": "
MethodMSD SpleenMM-WHS
From Scratch
3D UNet [46]93.7183.09
UNETR† [27]94.2085.85
Swin-UNETR† [26]94.6386.11
With General SSL
MAE3D [28, 13]95.2086.03
MoCo v3† [29, 12]94.3284.16
Jiasaw† [9]94.2985.88
PositionLabel† [68]94.1685.52
With Medical SSL
MG [72]94.4086.36
VicRegl [3]94.1284.72
UniMiss [62]95.0984.68
PCRLv1† [70]94.3286.58
PCRLv2† [71]94.9486.82
Rubik++† [52]95.1187.02
Swin-UNETR [51, 26]95.0287.06
SwinMM [55]95.3486.98
JSSL [42]94.9284.89
GVSL [32]95.4788.27
VoCo96.3490.54
", + "bbox": [ + 540, + 436, + 852, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Experimental results on MSD Spleen [1] and MM-WHS [74]. We report the Dice Scores of segmentation prediction. $\\dagger$ denotes we re-implement the approach.", + "bbox": [ + 498, + 710, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "propriate in medical images. MAE [28, 13], SimCLR [10], and SimMIM [63] (results from [13]) also gain limited performance. Our VoCo also outperforms the position-based methods Jigsaw [9] and PositionLabel [68] by a clear margin. Thus, we conclude that general SSL methods are not very suitable for 3D medical images. It is crucial to consider the characteristics of medical images in medical SSL.", + "bbox": [ + 496, + 777, + 890, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The scratch Swin-UNETR [26] only achieves $80.53\\%$", + "bbox": [ + 519, + 883, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22878", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/68063f4546c105b98685ac041e69423c4674723c7825de98fb7832c240d35d2f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNet.Dice Score(%)
TCWTETAVG
From Scratch
UNETR [27]-81.6287.8157.3475.58
Swin-UNETR [26]-81.2888.6757.7375.89
With General SSL
MAE3D [28, 13]UNETR82.3490.3559.1877.29
SimMIM [63]UNETR84.0690.4359.0777.85
SimCLR [10]UNETR83.1389.4458.4276.99
MoCo v3† [29, 12]UNETR82.6088.8957.6976.39
Jigsaw† [9]Sw-UNE.81.6289.4559.1076.72
PositionLabel† [68]Sw-UNE.81.3589.6258.7376.64
With Medical SSL
PCRLv1†[70]Sw-UNE.81.9688.8357.5876.12
PCRLv2†[71]Sw-UNE.82.1390.0657.7076.63
Rubik++†[52]Sw-UNE.84.3290.2358.0177.51
Swin-UNETR [51, 26]Sw-UNE.82.5189.0858.1576.58
SwinMM [55]Sw-UNE.83.4890.4758.7277.56
VoCoSw-UNE.85.2790.4559.8778.53
", + "bbox": [ + 78, + 88, + 486, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b4087d451304d31dd049f0c307a5747d18b5362e2c75a5214c2b42c5428a7c0.jpg", + "table_caption": [ + "Table 4. Experimental results on BraTS 21 [49]. WT, TC, and ET denote the whole tumor, tumor core, and enhancing tumor, respectively. $\\dagger$ denotes we re-implement the approach." + ], + "table_footnote": [], + "table_body": "
MethodNetworkAccuracy(%)
From Scratch
UNETR [27]-88.92
Swin-UNETR [26]-88.04
With General SSL
MAE3D [28, 13]UNETR89.47
MoCo v3 [29, 12]UNETR84.95
Jiasaw [9]Swin-UNETR86.88
PositionLabel [68]Swin-UNETR87.54
With Medical SSL
PCRLv1 [70]Swin-UNETR88.72
PCRLv2 [71]Swin-UNETR89.15
Rubik++ [52]Swin-UNETR89.23
Swin-UNETR [51, 26]Swin-UNETR89.45
SwinMM [55]Swin-UNETR89.61
VoCoSwin-UNETR90.83
", + "bbox": [ + 107, + 383, + 437, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dice Score. With VoCo pre-training, we gain $3.32\\%$ improvements with $83.85\\%$ Dice Score, which also outperforms existing methods by a clear margin. Among the compared methods, GL-MAE [73] achieves the highest Dice Score $(82.01\\%)$ . Our VoCo surpasses it by $1.84\\%$ Dice Score, which is a clear improvement in this dataset.", + "bbox": [ + 75, + 627, + 467, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Promising performance on Unseen datasets. We further conduct experiments on unseen datasets in pre-training, i.e., LiTs [4], MSD Spleen [1], and MM-WHS [74]. The results on LiTs [4] are shown in Table 2. We report the results of compared methods according to [70, 71, 65]. Since the scratch Swin-UNETR [26] can obtain a higher Dice Score (93.42%), we further pre-train a 3D UNet [46] based on VoCo, aiming to conduct fair comparisons. It can be seen that with VoCo pre-training, Swin-UNETR [26] gains $3.10\\%$ improvements and achieves $96.52\\%$ Dice Score. With 3D UNet [46] as the backbone, VoCo also achieves $96.03\\%$ Dice Score, proving the effectiveness of VoCo with", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a0f0ab2186be6e026c2b52d348737a6b476da5fb8f1e306b15f06f5ba9cdb5f.jpg", + "image_caption": [ + "Figure 5. Overall comparisons with state-of-the-art methods on six different datasets." + ], + "image_footnote": [], + "bbox": [ + 500, + 89, + 887, + 256 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67f2ffb27484c8407e0b37605da899ffd57e9b6c525d584124ebdb6a6d25fb40.jpg", + "table_caption": [ + "Table 5. Experimental results of CC-CCII [67] classification." + ], + "table_footnote": [], + "table_body": "
Loss FunctionsBTCVMM-WHS
LpredLreg
XX80.5386.11
✓X82.9688.82
✓✓83.8590.54
", + "bbox": [ + 570, + 299, + 821, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f17d4bf7ed6e2b383241780984495d4b357b654980653c3c0de10e9668b84c3.jpg", + "table_caption": [ + "Table 6. Evaluation of loss functions $L_{pred}$ and $L_{reg}$ . We report the average Dice Score on BTCV [35] and MM-WHS [74]." + ], + "table_footnote": [], + "table_body": "
Number of bases nBTCVMM-WHS
(x,y,z)
2×2×181.5686.73
3×3×182.7789.31
4×4×183.8590.54
5×5×183.6090.49
3×3×282.8689.14
4×4×283.4790.52
", + "bbox": [ + 563, + 417, + 828, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Evaluation of the value of bases $n$ . We report the average Dice Score on BTCV [35] and MM-WHS [74].", + "bbox": [ + 500, + 525, + 890, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different network architectures.", + "bbox": [ + 500, + 580, + 707, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on MSD Spleen [1] and MM-WHS [74] datasets are shown in Table 3. In previous methods, GVSL [32] achieves the best performance with $95.47\\%$ and $88.27\\%$ Dice Score, while our VoCo surpasses all previous methods with $96.34\\%$ and $90.54\\%$ Dice Score on MSD Spleen [1] and MM-WHS [74] datasets, respectively.", + "bbox": [ + 496, + 595, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization capacity on MRI dataset. To verify the generalization capacity on the MRI dataset, we further evaluate the performance of VoCo on BraTS 21 [49]. As shown in Table 4, VoCo achieves $78.53\\%$ Dice Score and outperforms existing state-of-the-art methods, demonstrating the cross-model generalization capacity of VoCo.", + "bbox": [ + 496, + 686, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of COVID-19 classification. We further evaluate the performance of the classification task on the CC-CCII [67] dataset in Table 5. Since existing SSL methods did not conduct experiments on this dataset, we reproduce the related methods for comparisons. It can be seen that VoCo can also achieve superior results with $90.83\\%$ accuracy, proving its effectiveness in the classification task.", + "bbox": [ + 496, + 777, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall comparisons on six downstream datasets. The", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22879", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ab02bfc61edd815b2b5f76aeffa029525242aa7f5b890fa09ffa6cf03bb3607.jpg", + "image_caption": [ + "Figure 6. Qualitative visualization of segmentation results for the BTCV [35] dataset. We compare VoCo with MoCo v3 [12], PositionLabel [68], SwinUNETR [51, 26], and GVSL [32]." + ], + "image_footnote": [], + "bbox": [ + 220, + 88, + 750, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "overall comparisons are shown in Fig. 5. Specifically, we compare with the existing state-of-the-art methods on six different downstream datasets. It can be seen that our VoCo outperforms them by a clear margin.", + "bbox": [ + 75, + 324, + 470, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 393, + 230, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further conduct ablation studies to evaluate the loss functions and the settings in VoCo, which are verified on the BTCV [35] and MM-WHS [74] datasets.", + "bbox": [ + 75, + 417, + 468, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss functions. We first study the importance of the two loss functions, i.e., $L_{pred}$ and $L_{reg}$ , as shown in Table 6. It can be seen that with our proposed $L_{pred}$ loss function, the performance is significantly improved, i.e., $80.53\\%$ to $82.96\\%$ on BTCV, $86.11\\%$ to $88.82\\%$ on MM-WHS. These results demonstrate the effectiveness of our proposed position prediction pretext task. In addition, with the proposed regularization loss $L_{reg}$ , the performance can be further improved. Thus, we can see that it is crucial to learn discriminative bases in VoCo.", + "bbox": [ + 75, + 463, + 468, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of bases. We further evaluate different settings of the number of bases $n$ in VoCo. We compare with different settings of $n$ in the ablation studies, as shown in Table 7. It is worth noting that due to the ROI size inconsistency in the $Z$ direction, it is not practical to crop multiple bases in the $Z$ direction, since we have to resize the volume after crops, which will result in inconsistent volume scales. In addition, due to the computation limitation, it is costly to increase the values of $n$ . As shown in Table 7, with $n = 2 \\times 2 \\times 1$ , the VoCo only achieves $81.56\\%$ and $86.73\\%$ Dice Score on BTCV and MM-WHS, respectively. When we increase the values of $n$ to $3 \\times 3 \\times 1$ and $4 \\times 4 \\times 1$ , the performances are improved obviously. Specifically, with $n$ as $4 \\times 4 \\times 1$ , we achieve $83.85\\%$ and $90.54\\%$ on BTCV and MM-WHS, respectively. However, we observe that higher $n$ ( $5 \\times 5 \\times 1$ ) cannot further bring higher performance. We further verify the performance of generating base crops in the $Z$ direction. It can be seen that $3 \\times 3 \\times 2$ and $4 \\times 4 \\times 2$ cannot yield improvements. Thus, aiming to balance the performance", + "bbox": [ + 75, + 613, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and efficiency, we set $n$ as $4 \\times 4 \\times 1$ in VoCo. It can be seen that the setting of $n$ is crucial to VoCo.", + "bbox": [ + 498, + 324, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization results on BTCV [35] are shown in Fig. 6. It can be seen that VoCo can yield improved segmentation accuracy and completeness. More visualization results are in the supplementary materials.", + "bbox": [ + 498, + 354, + 890, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Directions", + "text_level": 1, + "bbox": [ + 500, + 428, + 805, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we develop a simple-yet-effective SSL framework VoCo for 3D medical image analysis. Motivated by the observation that 3D medical images contain relatively consistent contextual positions between different organs, we propose to leverage the contextual position priors to learn consistent semantic representations in pre-training. Specifically, we crop volumes from different positions in an input volume and represent them as a group of bases to represent features in different directions. Then, we predict the contextual position of a randomly cropped volume by contrasting its similarity to different bases. In this way, VoCo effectively encodes the contextual position priors into model representations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experiments demonstrate that VoCo achieves superior performance.", + "bbox": [ + 496, + 454, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We will further consider several ways of extension: (1) Scale up the pre-training dataset to evaluate the upper performance of VoCo. (2) Experiments on more downstream datasets. (3) Evaluate the label-efficient performance of VoCo (e.g., semi-supervised learning). (4) Explore the capacity of VoCo in mining inter-volume relationships.", + "bbox": [ + 496, + 696, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 800, + 658, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Hong Kong Innovation and Technology Fund (Project No. ITS/028/21FP and No. MHP/002/22), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. T45-401/22-N).", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22880", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michela Antonelli et al. The medical segmentation decathlon. Nature Commun., 13(1):4128, 2022. 5, 6, 7", + "[2] Shekoofeh Azizi et al. Big self-supervised models advance medical image classification. In ICCV, pages 3478-3488, 2021. 1", + "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. NIPS, 35:8799-8810, 2022. 6", + "[4] Patrick Bilic et al. The liver tumor segmentation benchmark (lits). Medical Image Analy., 84:102680, 2023. 5, 6, 7", + "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2", + "[6] Mathilde Caron et al. Unsupervised learning of visual features by contrasting cluster assignments. NIPS, 33:9912-9924, 2020. 1, 2, 4, 5", + "[7] Mathilde Caron et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2", + "[8] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 3", + "[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In CVPR, pages 11526-11535, 2021. 3, 5, 6, 7", + "[10] Ting Chen et al. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 2, 3, 4, 5, 6, 7", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 1, 2, 3, 4", + "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 5, 6, 7, 8", + "[13] Zekai Chen et al. Masked image modeling advances 3d medical image analysis. In WACV, pages 1970-1980, 2023. 1, 3, 5, 6, 7", + "[14] Kenneth Clark and Bruce and others Vendl. The cancer imaging archive (tcia): maintaining and operating a public information repository. Jour. of Dig. Imag., 26:1045-1057, 2013. 5", + "[15] Jiequan Cui et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 2", + "[16] Jiequan Cui et al. Generalized parametric contrastive learning. IEEE Trans. Pattern Analy. Mach. Intell., 2023. 2", + "[17] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 3", + "[18] Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 3", + "[19] Hao Du, Qihua Dong, Yan Xu, and Jing Liao. Weakly-supervised 3d medical image segmentation using geometric prior and contrastive similarity. IEEE Trans. Medi. Imag., 2023. 3", + "[20] Yuting Gao, Jia-Xin Zhuang, et al. Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In ECCV, pages 237-253, 2022. 1" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Florin-Cristian Ghesu et al. Multi-scale deep reinforcement learning for real-time 3d-landmark detection in ct scans. IEEE Trans. Pattern Anal. Mach. Intell., 41(1):176-189, 2017. 1", + "[22] Jean-Bastien Grill et al. Bootstrap your own latent-a new approach to self-supervised learning. NIPS, 33:21271-21284, 2020. 2", + "[23] Katharina Grünberg et al. Annotating medical image data. Medical Image Analy., pages 45-67, 2017. 1", + "[24] Fatemeh Haghighi et al. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE Trans. Medical Imag., 40(10):2857-2868, 2021. 6", + "[25] Fatemeh Haghighi et al. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In CVPR, pages 20824-20834, 2022. 1, 3", + "[26] Ali Hatamizadeh et al. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In MICCAIW, pages 272-284, 2021. 5, 6, 7, 8", + "[27] Ali Hatamizadeh et al. Unetr: Transformers for 3d medical image segmentation. In WACV, pages 574-584, 2022. 5, 6, 7", + "[28] Kaiming He et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 5, 6, 7", + "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5, 6, 7", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3", + "[31] Xingxin He, Leyuan Fang, Mingkui Tan, and Xiangdong Chen. Intra-and inter-slice contrastive learning for point supervised oct fluid segmentation. IEEE Trans. Image Process., 31:1870-1881, 2022. 3", + "[32] Yuting He et al. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In CVPR, pages 9538-9547, 2023. 1, 3, 5, 6, 7, 8", + "[33] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021. 1", + "[34] Yankai Jiang et al. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In ICCV, pages 15859-15869, 2023. 1, 5", + "[35] Bennett Landman et al. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In MICCAIW, volume 5, page 12, 2015. 5, 6, 7, 8", + "[36] Jie Liu et al. Clip-driven universal model for organ segmentation and tumor detection. In ICCV, pages 21152-21164, 2023. 1, 5", + "[37] Qiang Liu et al. A multi-level label-aware semi-supervised framework for remote sensing scene classification. IEEE Trans. Geosci. Remote Sens., 2023. 3", + "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22881", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Xiangde Luo et al. WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from ct image. Medical Image Analysis, 82:102642, 2022. 1", + "[40] Jun Ma et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2021. 1", + "[41] T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In CVPR, pages 9339-9348, 2018. 3", + "[42] Nguyen et al. Joint self-supervised image-volume representation learning with intra-inter contrastive clustering. In AAAI, pages 14426-14435, 2023. 6", + "[43] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84. Springer, 2016. 3", + "[44] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3", + "[45] Maxime Oquab et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2", + "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 6, 7", + "[47] Rodrigo Santa Cruz et al. Deeppermnet: Visual permutation learning. In CVPR, pages 3949-3957, 2017. 3", + "[48] Arnaud Arindra Adiyoso Setio et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analy., 42:1-13, 2017. 5", + "[49] Amber L Simpson et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 5, 7", + "[50] Aiham Taleb et al. 3d self-supervised methods for medical imaging. NIPS, 33:18158-18172, 2020. 1, 3, 6", + "[51] Yucheng Tang et al. Self-supervised pre-training of swin transformers for 3d medical image analysis. In CVPR, pages 20730-20740, 2022. 1, 3, 5, 6, 7, 8", + "[52] Xing Tao et al. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In MICCAI, pages 238-248, 2020. 1, 3, 6, 7", + "[53] Guotai Wang et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 41(7):1559-1572, 2018. 1", + "[54] Xiaosong Wang et al. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, pages 2097-2106, 2017. 5", + "[55] Yiqing Wang et al. Swinmm: masked multi-view with swim transformers for 3d medical image segmentation. In MIC-CAI, 2023. 1, 3, 5, 6, 7", + "[56] Xin Wen et al. Self-supervised visual representation learning with semantic grouping. NIPS, 35:16423-16438, 2022. 2", + "[57] Linshan Wu et al. Deep bilateral filtering network for point-supervised semantic segmentation in remote sensing images. IEEE Trans. Image Process., 31:7419-7434, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Linshan Wu et al. Modeling the label distributions for weakly-supervised semantic segmentation, 2024. 3", + "[59] Linshan Wu, Leyuan Fang, Xingxin He, Min He, Jiayi Ma, and Zhun Zhong. Querying labeled for unlabeled: Cross-image semantic consistency guided semi-supervised semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8827-8844, Jul. 2023. 3", + "[60] Linshan Wu, Ming Lu, and Leyuan Fang. Deep covariance alignment for domain adaptive remote sensing image segmentation. IEEE Trans. Geosci. Remote Sens., 60:1-11, 2022. 3", + "[61] Linshan Wu, Zhun Zhong, Leyuan Fang, Xingxin He, Qiang Liu, Jiayi Ma, and Hao Chen. Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In CVPR, pages 15454-15464, 2023. 3", + "[62] Yutong Xie, Jianpeng Zhang, Yong Xia, and Qi Wu. Unimiss: Universal medical self-supervised learning via breaking dimensionality barrier. In ECCV, pages 558-575, 2022. 1, 6", + "[63] Zhenda Xie et al. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5, 6, 7", + "[64] Shuangfei Zhai et al. Position prediction as an effective pretraining strategy. arXiv preprint arXiv:2207.07611, 2022. 3", + "[65] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Anal., 89:102879, 2023. 6, 7", + "[66] Hongyi Zhang et al. Mixup: Beyond empirical risk minimization. In ICLR, 2018. 3", + "[67] Kang Zhang et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181(6):1423-1433, 2020. 5, 7", + "[68] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In AAAI, pages 3516-3524, 2023. 3, 5, 6, 7, 8", + "[69] Hong-Yu Zhou et al. Comparing to learn: Surpassing image-net pretraining on radiographs by comparing image representations. In MICCAI, pages 398-407, 2020. 3", + "[70] Hong-Yu Zhou et al. Preservational learning improves self-supervised medical image models by reconstructing diverse contexts. In ICCV, pages 3499-3509, 2021. 1, 3, 5, 6, 7", + "[71] Hong-Yu Zhou et al. A unified visual information preservation framework for self-supervised pre-training in medical image analysis. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 3, 5, 6, 7", + "[72] Zongwei Zhou et al. Models genesis. Medical Image Analy., 67:101840, 2021. 6", + "[73] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 1, 3, 5, 6, 7", + "[74] Xiahai Zhuang. Multivariate mixture model for myocardial segmentation combining multi-source images. IEEE Trans. Pattern Analy. Mach. Intell., 41(12):2933-2946, 2018. 5, 6, 7, 8", + "[75] Xinrui Zhuang et al. Self-supervised feature learning for 3d medical images by playing a rubik's cube. In MICCAI, pages 420-428, 2019. 1, 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22882", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_model.json b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ea6e4816bc77de619639b1f5d2fbc2e2e98ca2ab --- /dev/null +++ b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.205, + 0.248, + 0.221 + ], + "angle": 0, + "content": "Linshan Wu" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.205, + 0.55, + 0.222 + ], + "angle": 0, + "content": "Jiaxin Zhuang" + }, + { + "type": "text", + "bbox": [ + 0.736, + 0.205, + 0.83, + 0.22 + ], + "angle": 0, + "content": "Hao Chen *" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.223, + 0.69, + 0.24 + ], + "angle": 0, + "content": "Hong Kong University of Science and Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Self-Supervised Learning (SSL) has demonstrated promising results in 3D medical image analysis. However, the lack of high-level semantics in pre-training still heavily hinders the performance of downstream tasks. We observe that 3D medical images contain relatively consistent contextual position information, i.e., consistent geometric relations between different organs, which leads to a potential way for us to learn consistent semantic representations in pre-training. In this paper, we propose a simple yet-effective Volume Contrast (VoCo) framework to leverage the contextual position priors for pre-training. Specifically, we first generate a group of base crops from different regions while enforcing feature discrepancy among them, where we employ them as class assignments of different regions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which region) by contrasting their similarity to different base crops, which can be seen as predicting contextual positions of different sub-volumes. Through this pretext task, VoCo implicitly encodes the contextual position priors into model representations without the guidance of annotations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experimental results on six downstream tasks demonstrate the superior effectiveness of VoCo. Code will be available at https://github.com/Luffy03/VoCo." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Deep learning has demonstrated outstanding achievements in 3D medical image analysis [53, 21, 40, 33, 39], yet is heavily hampered by the expensive cost of the required expert annotations [50, 23]. To address this problem, Self-Supervised Learning (SSL) has received significant attention due to its promising ability to learn representations without annotations [10, 11, 6, 28, 20], which has become" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.29, + 0.887, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.363, + 0.873, + 0.373 + ], + "angle": 0, + "content": "(a) Relatively consistent contextual positions in 3D medical images" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.376, + 0.888, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.481, + 0.86, + 0.492 + ], + "angle": 0, + "content": "(b) Volume Contrast (VoCo) for 3D Medical Image Analysis" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.495, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1. (a) In 3D medical images, the contextual positions, i.e., the geometric relations between different organs are relatively consistent. (b) Aiming to leverage contextual position priors for pretraining, we proposed a Volume Contrast (VoCo) framework for 3D Medical Image Analysis." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.596, + 0.892, + 0.627 + ], + "angle": 0, + "content": "an important label-efficient solution in 3D medical image analysis [71, 51, 32, 2, 34, 36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Existing methods [50, 75, 71, 13] are mostly based on information reconstructions to learn augment-invariant representations of 3D medical images, which first employ strong data augmentation to the images and then reconstruct the raw information. Specifically, rotate-and-reconstruct [50, 51, 75, 52] proposed to randomly rotate the 3D volumetric images and learn to recover them, which encourages models to learn rotational invariant features. Recent methods [70, 71, 32, 25, 62] further proposed to restore information among different views of the image. PCRL [70, 71] cropped global and local patches then conducted multiscale restorations. GVSL [32] further explored the geometric similarity between multi-scans by affine augmentation and matching. Mask-reconstruct methods [13, 73, 55] are also widely used, which are introduced from MAE [28] and aim to learn representations by masking images and reconstructing the missing pixels. Although promising results have been demonstrated, previous works [52, 32] have" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.308, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author: jhc@cse.ust.hk" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22873" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.092, + 0.362, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.215, + 0.317, + 0.228 + ], + "angle": 0, + "content": "(a) Instance-level" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.092, + 0.603, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.216, + 0.543, + 0.228 + ], + "angle": 0, + "content": "(b) Prototype-level" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.092, + 0.815, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.216, + 0.749, + 0.228 + ], + "angle": 0, + "content": "(c) Our VoCo" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 2. Typical contrastive learning frameworks. (a) Instance-level contrastive learning [10, 11, 29, 22, 7] employs strong data augmentation or model perturbation on input data to acquire different views of instance, then regularizes their consistency. (b) Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] conducts (1) online clustering or (2) randomly initialize then online update process to obtain prototypes as class assignments, then leverage the prototypes to contrast each input image. (c) Our VoCo follows the idea of prototype-level contrastive learning. Specifically, instead of using time-consuming online clustering and updating procedures, we leverage the valuable contextual position priors of 3D medical images and leverage the base crops to generate prototypes (bases)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.414 + ], + "angle": 0, + "content": "proved that the lack of high-level semantics in pre-training will heavily hinder the performance of downstream tasks. To address this challenge, we argue that stronger high-level semantics should be further involved into 3D medical image pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.627 + ], + "angle": 0, + "content": "To this end, we argue that the contextual position priors of 3D medical images should be further exploited. As shown in Fig. 1(a), we observe that in 3D medical images, different organs (semantic regions) contain relatively consistent contextual positions with relatively consistent anatomic characteristics (shapes). Thus, the consistency of geometric relations between different organs leads to a potential way for us to learn consistent semantic representations for 3D medical images pre-training. In this paper, we propose a pretext task for contextual position predictions, which aims to encode contextual position priors into model representations and enables us to effectively improve the performance of downstream tasks that require high-level semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a simple-yet-effective Volume Contrast (VoCo) framework for 3D medical image analysis, as shown in Fig. 1(b). Specifically, we first crop a group of non-overlap volumes from different positions while enforcing feature discrepancy among them. We represent these volumes as a group of bases in the learned high-dimension space, where we employ them as class assignments of different positions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which position) by contrasting their similarity to different bases, which can be seen as predicting contextual positions of different sub-volumes. In this way, we formulate a contextual position prediction pretext task for 3D medical image SSL. Through learning to predict contextual positions, we implicitly involve the high-level semantic priors into the model representations, which enables us to significantly improve the performance of downstream tasks. Extensive experimental results on six downstream tasks demonstrate that our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.339, + 0.892, + 0.369 + ], + "angle": 0, + "content": "proposed VoCo clearly outperforms existing state-of-the-art 3D medical image SSL methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.65, + 0.403 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.504 + ], + "angle": 0, + "content": "In this section, we first introduce the previous mainstream contrastive learning paradigms. Then, we survey the existing SSL methods for medical image analysis, especially for 3D medical images. Finally, we review the position-related SSL methods for comparisons with our method and highlight the differences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Contrastive learning. Contrastive learning is one of the mainstream paradigms in SSL, which aims to learn consistent representations by contrasting positive and negative pairs of samples without extra annotations [10, 6, 29, 11]. According to [6], instance- and prototype-level contrastive learning are two typical types of contrastive learning, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Instance-level contrastive learning [10, 11, 29, 22, 7] transforms input images with different augmentations or model perturbations, aiming to compare the features from each other. Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] proposes to generate prototypes (also called clusters or bases) for contrasting each input image. Specifically, there are two typical ways to generate prototypes. First, Caron et al. proposed DeepCluster [5] to conduct online clustering on the whole dataset to generate prototypes. However, it is very time-consuming to calculate clusters on a large dataset. Thus, some recent works [6, 56, 15, 16] propose to randomly initialize a group of prototypes and then update them through back-propagation during training, which has demonstrated promising results. However, there is still no explicit guarantee that these randomly initialized prototypes can be updated well during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our VoCo follows the primary idea of prototype-level constrasive learning. As shown in Fig. 2(c), to address the existing problems mentioned above, instead of randomly" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22874" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "initializing and updating prototypes, VoCo leverages the valuable contextual position priors of 3D medical images to generate base crops as prototypes, which also requires no time-consuming clustering on a large dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.162, + 0.47, + 0.344 + ], + "angle": 0, + "content": "SSL for medical image analysis. Due to the high potential in label-efficient learning [29, 59, 61, 58, 57, 60, 37], SSL has also received significant attention in the field of medical image analysis [70, 32, 31, 51, 19]. Existing methods are mainly based on comparative SSL [71]. Specifically, Zhou et al. [69] combined Mixup [66] into MoCo [29] to learn the diversity of positive and negative samples in InfoNCE [44]. Azizi et al. used multi-instance learning to compare multiple views of images from each patient. There are also a number of approaches [25, 70, 71] that supervising the models via restoring low-level information from raw images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.534 + ], + "angle": 0, + "content": "In 3D medical image analysis, reconstructing raw information is a popular pretext task for learning representations [50, 51, 71]. Existing methods are mainly based on reconstructing information from augmented images. These previous methods first conducted strong data augmentation, e.g., rotate [51, 75, 52], multi-view crops [70, 71, 32], and mask [13, 73, 55], then supervised the model by reconstructing raw 3D information. Although promising results have been demonstrated, most of these methods still largely ignore the importance of integrating high-level semantics into model representations, which heavily hinders the performance of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.74 + ], + "angle": 0, + "content": "Position-related SSL. Position-related SSL methods are also explored in a number of previous works [8, 9, 41, 47, 43, 17, 64, 68] in the field of natural images. Noroozi et al. [43] proposed to predict the order of a set of shuffled patches. Zhai et al. [64] and Caron et al. [8] proposed to train a ViT [18] to predict the locations of each input patch. However, since the geometric relations of different objects are not very consistent in natural images, it is still difficult to effectively learn consistent position representations given visual appearance only (as stated in [68]). In addition, previous works [64, 8, 68] mainly trained a linear layer to output the positions directly, which works in a black-box manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce the pretext task of contextual position prediction into the field of 3D medical images, where the geometric relations between different organs are relatively consistent, which guides us to learn consistent semantic representations in pre-training. Different from the previous methods, in this paper, we introduce a totally different position prediction paradigm. Specifically, instead of using a linear layer to output positions directly, we predict the contextual positions based on volume contrast, which is more intuitive and effective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.636, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.894, + 0.193 + ], + "angle": 0, + "content": "In this section, we first introduce the overall framework of our proposed VoCo in Section 3.1. After that, we present the process of contextual position prediction in Section 3.2. Then, the regularization process via volume contrast in our proposed VoCo framework is described in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.203, + 0.691, + 0.218 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.227, + 0.893, + 0.513 + ], + "angle": 0, + "content": "The overall framework of our proposed VoCo is presented in Fig. 3, which contains a contextual position prediction branch and a regularization branch. The prediction branch is used to predict the contextual positions between different cropped volumes. Specifically, given an input volume, we first crop it into non-overlap base volumes, which cover the whole input volume. Then, we randomly crop a volume and transform it into the high-dimension feature space using a typical backbone (CNN [30] or Transformer [18]). The goal is to predict the contextual positions between the randomly cropped volumes and base volumes. In this paper, instead of training a linear classifier to predict positions as in previous works [64, 9, 8, 68], we propose to establish this goal by volume contrast. We develop a loss function \\( L_{pred} \\) to supervise the final predictions. In addition, we further use a loss function \\( L_{reg} \\) to regularize the feature discrepancy from different bases by enlarging their distance, aiming to learn more discriminative class assignments. The details are presented in Section 3.2 and 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.525, + 0.771, + 0.54 + ], + "angle": 0, + "content": "3.2. Contextual Position Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Base and random crops. Given an input volume, we first crop it into \\( n \\) non-overlap base volumes, which cover the whole input volume. We then employ the extracted features \\( z \\) as class assignments (we call them bases), which present the prototype-level features from different positions. Then, following previous SSL works [10, 11, 29], a projector with linear layers is used to project \\( z \\) into latent features \\( q \\). Then, we randomly crop a volume and transform it into high-dimension feature space as \\( p \\). The backbone and projector are also used to project the features from the randomly cropped volumes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.82 + ], + "angle": 0, + "content": "Volume contrast for contextual position prediction. With features extracted from the backbone and projector, following previous SSL works [10, 11, 29], we first conduct 3D adaptive average pooling to resize them to one dimension, i.e., \\( p \\in \\mathbb{R}^{1 \\times C} \\) and \\( q \\in \\mathbb{R}^{1 \\times C} \\), where \\( C \\) is the number of channels. Specifically, we empirically set \\( C \\) to 2048 as in [10, 11, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.821, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Then, we calculate the similarity logits \\( l \\) between \\( p \\) and \\( q_{i} \\). Specifically, we use cosine similarity to compute \\( l \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.875, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nl _ {i} = \\operatorname {C o s S i m} (p, q _ {i}) = \\frac {p \\cdot q _ {i}}{\\| p \\| \\| q _ {i} \\|}, i \\in n \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22875" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.087, + 0.803, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Figure 3. The overall framework of VoCo. VoCo contains a prediction branch and a regularization branch. The prediction branch is responsible for predicting contextual positions between different sub-volumes. The regularization branch is employed to enforce the feature discrepancy between different bases, which aims to learn more discriminative class assignments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.501 + ], + "angle": 0, + "content": "where \\( q_{i} \\) is the projected feature of each base crop. \\( l_{i} \\) denotes the similarity between \\( p \\) in \\( q_{i} \\), which ranges from 0 to 1. It is worth noting that, we stop the gradients of \\( q \\) when computing Eq. 1, which aims to avoid feature collapse [10, 11, 6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Intuitively, higher \\( l_{i} \\) represents that \\( p \\) has higher probabilities to share overlap regions with \\( q_{i} \\). In this way, we can explicitly associate the similarity value with the position information, i.e., \\( p \\) with higher \\( l_{i} \\) is more likely to be located in the region of the \\( i_{th} \\) base. Thus, instead of training a black-box linear layer, we predict the contextual positions by volume contrast, which is more intuitive and effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.765 + ], + "angle": 0, + "content": "Position labels generation. The process of generating position labels is shown in Fig. 4. As shown in Fig. 4, when we generate \\( n = 4 \\times 4 \\) base crops, there will be \\( n \\) class assignments. Then we calculate the overlap area between a randomly cropped volume and \\( n \\) base crops. The proportions of the overlap area are then assigned as position labels \\( y \\), which also range from 0 to 1. Thus, we can easily supervise the model by calculating the distance between the prediction logits \\( l \\) and position labels \\( y \\). The setting of the number \\( n \\) of base crops will be discussed in Section 4.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Loss function for contextual position prediction. The formulation of prediction loss function \\( L_{pred} \\) is based on entropy. Specifically, we first calculate the distance \\( d \\) between prediction logits \\( l \\) and position labels \\( y \\):" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.85, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\nd _ {i} = \\left| y _ {i} - l _ {i} \\right|, i \\in n, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( |.| \\) denotes the absolute value. Then, \\( L_{pred} \\) is formu" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.422, + 0.847, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.581, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Figure 4. The process of generating position labels. We calculate the proportions of overlap area as position labels \\( y \\), e.g., the randomly cropped volume in the figure is assigned to class 5, 6, 9, and 10 with probabilities of 0.25, 0.1, 0.5, and 0.15, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.661, + 0.611, + 0.675 + ], + "angle": 0, + "content": "lated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.684, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\nL _ {p r e d} = - \\frac {1}{n} \\sum_ {i \\in n} ^ {n} \\log \\left(1 - d _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.733, + 0.892, + 0.824 + ], + "angle": 0, + "content": "It is worth noting that VoCo predicts contextual positions of a volume (high similarities with all its contextual overlapped volumes), thus don't need one-to-one correspondence: e.g., in Fig. 4, high-value \\( l_{i} \\) pertain to \\( y_{i} > 0, i = 5,6,9,10 \\) simultaneously. Then we calculate the distance between \\( l_{i} \\) and \\( y_{i} \\) (Eq. 2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.812, + 0.849 + ], + "angle": 0, + "content": "3.3. Volume Contrast for Regularization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We aim to learn more discriminative class assignments (bases) for volume contrast. Since intuitively, different subvolumes tend to contain different organs (semantic discrepancy)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22876" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "ancy). Thus, we aim to enlarge the high-dimension feature discrepancy between different bases. To this end, we design a regularization loss \\( L_{reg} \\) to enlarge the feature discrepancy between different bases \\( z \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.153, + 0.468, + 0.183 + ], + "angle": 0, + "content": "First, given projected bases \\( q \\), we also calculate the cosine similarity \\( s_{ij} \\) between different \\( q_i \\) and \\( q_j \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.193, + 0.468, + 0.224 + ], + "angle": 0, + "content": "\\[\ns _ {i j} = \\operatorname {C o s S i m} \\left(q _ {i}, q _ {j}\\right) = \\frac {q _ {i} \\cdot q _ {j}}{\\| q _ {i} \\| \\| q _ {j} \\|}, i, j \\in n, i \\neq j, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.233, + 0.468, + 0.279 + ], + "angle": 0, + "content": "where we aim to regularize \\( s_{ij} \\) to 0, enforcing feature discrepancy between different bases. Thus, the loss function \\( L_{reg} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.29, + 0.468, + 0.332 + ], + "angle": 0, + "content": "\\[\nL _ {r e g} = \\frac {2}{n (n - 1)} \\sum_ {i, j \\in n, i \\neq j} ^ {n} | s _ {i j} |, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.468, + 0.372 + ], + "angle": 0, + "content": "where \\( |.| \\) denotes the absolute value. With loss \\( L_{reg} \\), we aim to optimize \\( q \\) as linearly independent bases:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.385, + 0.468, + 0.402 + ], + "angle": 0, + "content": "\\[\nq _ {i} \\bot q _ {j}, i \\neq j, i, j \\in n. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.468, + 0.488 + ], + "angle": 0, + "content": "With the regularization loss function \\( L_{reg} \\), we aim to learn a group of linearly independent bases to represent all directions of high-dimension features [6]. In this way, we can learn a group of more discriminative class assignments to supervise the final position predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.468, + 0.519 + ], + "angle": 0, + "content": "Overall loss function. Thus, the total loss function \\( L \\) is the combination of \\( L_{reg} \\) and \\( L_{pred} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.531, + 0.468, + 0.547 + ], + "angle": 0, + "content": "\\[\nL = L _ {p r e d} + \\lambda L _ {r e g}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.468, + 0.618 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is used to balance the relative contributions of these two loss terms and set to 1.0 in experiments empirically since we consider their importance equally. The ablation studies of \\(\\lambda\\) are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.208, + 0.648 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.468, + 0.748 + ], + "angle": 0, + "content": "In this section, we first describe the datasets used in the pre-training and downstream tasks. Then, we briefly introduce the implementation details of VoCo. Finally, we report detailed experiment results of our proposed VoCo compared with other state-of-the-art SSL methods in 3D medical images. More details are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.179, + 0.771 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Pre-training datasets. Aiming to conduct fair comparisons with the previous works [51, 55, 70, 71, 13, 73], we also carry out pre-training experiments on the same three public datasets, i.e., BTCV [35], TCIA Covid19 [14], and LUNA [48] datasets, including about totally \\(1.6\\mathrm{k}\\) CT scans for pre-training. It is worth noting that, aiming to conduct fair comparisons with previous works [73, 13], we only use BTCV [35] and TCIA Covid-19 [14] for pre-training in" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "the downstream experiments of BTCV [35]. For the other downstream tasks, we use all three datasets for pre-training. Details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Downstream datasets. To evaluate the effectiveness of our VoCo, we conduct downstream experiments on six public datasets, i.e., BTCV [35], LiTs [4], MSD Spleen [1], MM-WHS [74], BraTS 21 [49], and CC-CCII [67], including segmentation and classification tasks. The first five datasets are developed for segmentation, while CC-CCII [67] is for COVID-19 classification. Note that only BTCV [35] is used in pre-training, the other datasets are unseen in pre-training. In addition, to evaluate the cross-modality generalization ability, we transfer the model pre-trained on the CT dataset to the MRI dataset BraTS 21 [49]. We adopt consistent settings as previous works [13, 73, 26, 55, 32]. We also evaluate the performance on 2D medical dataset [54]. Details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.714, + 0.374 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.891, + 0.518 + ], + "angle": 0, + "content": "Following previous works [51, 55], we use SwinUNETR [26] for both pre-training and downstream tasks. We use AdamW [38] optimizer and cosine learning rate scheduler for all experiments. We set 100K training steps in the pre-training process and applied a slicing window inference for fair comparisons with previous works [51, 55, 13, 73]. Aiming to evaluate the pure effectiveness, we do not use foundation models or post-processing [34, 36]. Details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.519, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Comparison methods. We compare our VoCo with both General and Medical SSL methods. First, we compare with the typical SSL methods MAE [28, 13] and MoCo v3 [29, 12], since they represent the two mainstream SSL paradigms, i.e., mask-autoencoder and constrastive learning. Since it is not practical to set a large batch size for 3D medical images due to computation cost, for fair comparisons, we adopt consistent settings with other methods in MAE [28, 13] and MoCo v3 [29, 12]. We also report the results of SimCLR [10] and SimMIM [63] according to [13]. We further evaluate the performance of Jiasaw [9] and PositionLabel [68], since they are related to our position-aware methods. Most existing state-of-the-art medical SSL methods are compared in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.799, + 0.756 + ], + "angle": 0, + "content": "4.3. Experiments on downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.891, + 0.87 + ], + "angle": 0, + "content": "Outperform existing methods on the BTCV dataset. We first conduct experiments on BTCV [35], as shown in Table 1. Specifically, among the comparison methods, MAE3D [28, 13], SimCLR [10], SimMIM [63], MoCo v3 [29, 12], and GL-MAE [73] use UNETR [27]. Other methods including our VoCo adopt Swin-UNETR [26] as the default settings of the previous work [51]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Remark. It can be seen in Table 1 that the general SSL methods perform worse than most medical SSL methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22877" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.907, + 0.386 + ], + "angle": 0, + "content": "
MethodDice Score(%)AVG
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAG
From Scratch
UNETR [27]93.0294.1394.1266.9970.8796.1177.2789.2282.1070.1676.6565.3259.2179.82
Swin-UNETR† [26]94.0693.5493.8065.5174.6097.0975.9491.8082.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [28, 13]93.9894.3794.1869.8674.6596.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [10]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [63]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3† [29, 12]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw† [9]94.6293.4193.5575.6373.2195.7180.8089.4184.7871.0279.5765.6860.2281.35
PositionLabel† [68]94.3593.1593.2175.3973.2495.7680.6988.8084.0471.1879.0265.1160.0781.09
With Medical SSL
MG [72]91.9993.5291.8165.1176.1495.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [50]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [3]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++† [52]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1† [70]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2† [71]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [26, 51]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [55]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [73]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL† [32]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.387, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Table 1. Experimental results on BTCV [35]. The best results are bolded. 'From Scratch' denotes the supervised baseline without self-supervised pre-training. † denotes we re-implement the approach. Most results are drawn from [13, 65, 73] or their own papers." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.437, + 0.443, + 0.734 + ], + "angle": 0, + "content": "
MethodNetworkDice Score(%)
From Scratch
3D UNet [46]-90.70
UNETR† [27]-93.25
Swin-UNETR† [26]-93.42
With General SSL
Jigsaw [9]3D UNet94.36
MAE3D [28, 13]UNETR94.02
MoCo v3† [29, 12]UNETR93.86
Jigsaw† [9]Swin-UNETR95.24
PositionLabel† [68]Swin-UNETR94.13
With Medical SSL
MG [72]3D UNet91.30
TransVW [24]3D UNet91.42
ROT [50]3D UNet94.49
PCRLv1 [70]3D UNet93.87
PCRLv2 [71]3D UNet94.50
Rubik [75]3D UNet94.93
Rubik++ [52]3D UNet95.46
Rubik++† [52]Swin-UNETR95.72
Swin-UNETR [51, 26]Swin-UNETR95.33
SwinMM [55]Swin-UNETR95.52
VoCo3D UNet96.03
VoCoSwin-UNETR96.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Table 2. Experimental results on LiTs [4]. We report the Dice Scores of liver segmentation. \\(\\dagger\\) denotes we re-implement the approach. Most results are drawn from [65, 71]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Specifically, MoCo v3 [29, 12] can only achieve \\(79.54\\%\\) Dice Score. Since MoCo v3 [29, 12] heavily relies on a large batch size to acquire adequate negative samples, which is not practical in 3D medical images due to the huge computation burden. In addition, the negative relation between different images used in MoCo [29, 12] is not ap" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.437, + 0.853, + 0.711 + ], + "angle": 0, + "content": "
MethodMSD SpleenMM-WHS
From Scratch
3D UNet [46]93.7183.09
UNETR† [27]94.2085.85
Swin-UNETR† [26]94.6386.11
With General SSL
MAE3D [28, 13]95.2086.03
MoCo v3† [29, 12]94.3284.16
Jiasaw† [9]94.2985.88
PositionLabel† [68]94.1685.52
With Medical SSL
MG [72]94.4086.36
VicRegl [3]94.1284.72
UniMiss [62]95.0984.68
PCRLv1† [70]94.3286.58
PCRLv2† [71]94.9486.82
Rubik++† [52]95.1187.02
Swin-UNETR [51, 26]95.0287.06
SwinMM [55]95.3486.98
JSSL [42]94.9284.89
GVSL [32]95.4788.27
VoCo96.3490.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.712, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Table 3. Experimental results on MSD Spleen [1] and MM-WHS [74]. We report the Dice Scores of segmentation prediction. \\(\\dagger\\) denotes we re-implement the approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.884 + ], + "angle": 0, + "content": "propriate in medical images. MAE [28, 13], SimCLR [10], and SimMIM [63] (results from [13]) also gain limited performance. Our VoCo also outperforms the position-based methods Jigsaw [9] and PositionLabel [68] by a clear margin. Thus, we conclude that general SSL methods are not very suitable for 3D medical images. It is crucial to consider the characteristics of medical images in medical SSL." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The scratch Swin-UNETR [26] only achieves \\(80.53\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22878" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.487, + 0.326 + ], + "angle": 0, + "content": "
MethodNet.Dice Score(%)
TCWTETAVG
From Scratch
UNETR [27]-81.6287.8157.3475.58
Swin-UNETR [26]-81.2888.6757.7375.89
With General SSL
MAE3D [28, 13]UNETR82.3490.3559.1877.29
SimMIM [63]UNETR84.0690.4359.0777.85
SimCLR [10]UNETR83.1389.4458.4276.99
MoCo v3† [29, 12]UNETR82.6088.8957.6976.39
Jigsaw† [9]Sw-UNE.81.6289.4559.1076.72
PositionLabel† [68]Sw-UNE.81.3589.6258.7376.64
With Medical SSL
PCRLv1†[70]Sw-UNE.81.9688.8357.5876.12
PCRLv2†[71]Sw-UNE.82.1390.0657.7076.63
Rubik++†[52]Sw-UNE.84.3290.2358.0177.51
Swin-UNETR [51, 26]Sw-UNE.82.5189.0858.1576.58
SwinMM [55]Sw-UNE.83.4890.4758.7277.56
VoCoSw-UNE.85.2790.4559.8778.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.327, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Table 4. Experimental results on BraTS 21 [49]. WT, TC, and ET denote the whole tumor, tumor core, and enhancing tumor, respectively. \\(\\dagger\\) denotes we re-implement the approach." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.384, + 0.439, + 0.586 + ], + "angle": 0, + "content": "
MethodNetworkAccuracy(%)
From Scratch
UNETR [27]-88.92
Swin-UNETR [26]-88.04
With General SSL
MAE3D [28, 13]UNETR89.47
MoCo v3 [29, 12]UNETR84.95
Jiasaw [9]Swin-UNETR86.88
PositionLabel [68]Swin-UNETR87.54
With Medical SSL
PCRLv1 [70]Swin-UNETR88.72
PCRLv2 [71]Swin-UNETR89.15
Rubik++ [52]Swin-UNETR89.23
Swin-UNETR [51, 26]Swin-UNETR89.45
SwinMM [55]Swin-UNETR89.61
VoCoSwin-UNETR90.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.587, + 0.438, + 0.601 + ], + "angle": 0, + "content": "Table 5. Experimental results of CC-CCII [67] classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Dice Score. With VoCo pre-training, we gain \\(3.32\\%\\) improvements with \\(83.85\\%\\) Dice Score, which also outperforms existing methods by a clear margin. Among the compared methods, GL-MAE [73] achieves the highest Dice Score \\((82.01\\%)\\). Our VoCo surpasses it by \\(1.84\\%\\) Dice Score, which is a clear improvement in this dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Promising performance on Unseen datasets. We further conduct experiments on unseen datasets in pre-training, i.e., LiTs [4], MSD Spleen [1], and MM-WHS [74]. The results on LiTs [4] are shown in Table 2. We report the results of compared methods according to [70, 71, 65]. Since the scratch Swin-UNETR [26] can obtain a higher Dice Score (93.42%), we further pre-train a 3D UNet [46] based on VoCo, aiming to conduct fair comparisons. It can be seen that with VoCo pre-training, Swin-UNETR [26] gains \\(3.10\\%\\) improvements and achieves \\(96.52\\%\\) Dice Score. With 3D UNet [46] as the backbone, VoCo also achieves \\(96.03\\%\\) Dice Score, proving the effectiveness of VoCo with" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.09, + 0.888, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.261, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Overall comparisons with state-of-the-art methods on six different datasets." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.3, + 0.822, + 0.375 + ], + "angle": 0, + "content": "
Loss FunctionsBTCVMM-WHS
LpredLreg
XX80.5386.11
✓X82.9688.82
✓✓83.8590.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Table 6. Evaluation of loss functions \\( L_{pred} \\) and \\( L_{reg} \\). We report the average Dice Score on BTCV [35] and MM-WHS [74]." + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.418, + 0.829, + 0.524 + ], + "angle": 0, + "content": "
Number of bases nBTCVMM-WHS
(x,y,z)
2×2×181.5686.73
3×3×182.7789.31
4×4×183.8590.54
5×5×183.6090.49
3×3×282.8689.14
4×4×283.4790.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.526, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Table 7. Evaluation of the value of bases \\( n \\) . We report the average Dice Score on BTCV [35] and MM-WHS [74]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.581, + 0.709, + 0.594 + ], + "angle": 0, + "content": "different network architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.687 + ], + "angle": 0, + "content": "The results on MSD Spleen [1] and MM-WHS [74] datasets are shown in Table 3. In previous methods, GVSL [32] achieves the best performance with \\(95.47\\%\\) and \\(88.27\\%\\) Dice Score, while our VoCo surpasses all previous methods with \\(96.34\\%\\) and \\(90.54\\%\\) Dice Score on MSD Spleen [1] and MM-WHS [74] datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Generalization capacity on MRI dataset. To verify the generalization capacity on the MRI dataset, we further evaluate the performance of VoCo on BraTS 21 [49]. As shown in Table 4, VoCo achieves \\(78.53\\%\\) Dice Score and outperforms existing state-of-the-art methods, demonstrating the cross-model generalization capacity of VoCo." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Evaluation of COVID-19 classification. We further evaluate the performance of the classification task on the CC-CCII [67] dataset in Table 5. Since existing SSL methods did not conduct experiments on this dataset, we reproduce the related methods for comparisons. It can be seen that VoCo can also achieve superior results with \\(90.83\\%\\) accuracy, proving its effectiveness in the classification task." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overall comparisons on six downstream datasets. The" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22879" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.222, + 0.089, + 0.75, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.302 + ], + "angle": 0, + "content": "Figure 6. Qualitative visualization of segmentation results for the BTCV [35] dataset. We compare VoCo with MoCo v3 [12], PositionLabel [68], SwinUNETR [51, 26], and GVSL [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.471, + 0.388 + ], + "angle": 0, + "content": "overall comparisons are shown in Fig. 5. Specifically, we compare with the existing state-of-the-art methods on six different downstream datasets. It can be seen that our VoCo outperforms them by a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.231, + 0.411 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.463 + ], + "angle": 0, + "content": "We further conduct ablation studies to evaluate the loss functions and the settings in VoCo, which are verified on the BTCV [35] and MM-WHS [74] datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Loss functions. We first study the importance of the two loss functions, i.e., \\( L_{pred} \\) and \\( L_{reg} \\), as shown in Table 6. It can be seen that with our proposed \\( L_{pred} \\) loss function, the performance is significantly improved, i.e., \\( 80.53\\% \\) to \\( 82.96\\% \\) on BTCV, \\( 86.11\\% \\) to \\( 88.82\\% \\) on MM-WHS. These results demonstrate the effectiveness of our proposed position prediction pretext task. In addition, with the proposed regularization loss \\( L_{reg} \\), the performance can be further improved. Thus, we can see that it is crucial to learn discriminative bases in VoCo." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Number of bases. We further evaluate different settings of the number of bases \\( n \\) in VoCo. We compare with different settings of \\( n \\) in the ablation studies, as shown in Table 7. It is worth noting that due to the ROI size inconsistency in the \\( Z \\) direction, it is not practical to crop multiple bases in the \\( Z \\) direction, since we have to resize the volume after crops, which will result in inconsistent volume scales. In addition, due to the computation limitation, it is costly to increase the values of \\( n \\). As shown in Table 7, with \\( n = 2 \\times 2 \\times 1 \\), the VoCo only achieves \\( 81.56\\% \\) and \\( 86.73\\% \\) Dice Score on BTCV and MM-WHS, respectively. When we increase the values of \\( n \\) to \\( 3 \\times 3 \\times 1 \\) and \\( 4 \\times 4 \\times 1 \\), the performances are improved obviously. Specifically, with \\( n \\) as \\( 4 \\times 4 \\times 1 \\), we achieve \\( 83.85\\% \\) and \\( 90.54\\% \\) on BTCV and MM-WHS, respectively. However, we observe that higher \\( n \\) (\\( 5 \\times 5 \\times 1 \\)) cannot further bring higher performance. We further verify the performance of generating base crops in the \\( Z \\) direction. It can be seen that \\( 3 \\times 3 \\times 2 \\) and \\( 4 \\times 4 \\times 2 \\) cannot yield improvements. Thus, aiming to balance the performance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.325, + 0.892, + 0.354 + ], + "angle": 0, + "content": "and efficiency, we set \\( n \\) as \\( 4 \\times 4 \\times 1 \\) in VoCo. It can be seen that the setting of \\( n \\) is crucial to VoCo." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Visualization results on BTCV [35] are shown in Fig. 6. It can be seen that VoCo can yield improved segmentation accuracy and completeness. More visualization results are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.807, + 0.445 + ], + "angle": 0, + "content": "5. Conclusion and Future Directions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.697 + ], + "angle": 0, + "content": "In this paper, we develop a simple-yet-effective SSL framework VoCo for 3D medical image analysis. Motivated by the observation that 3D medical images contain relatively consistent contextual positions between different organs, we propose to leverage the contextual position priors to learn consistent semantic representations in pre-training. Specifically, we crop volumes from different positions in an input volume and represent them as a group of bases to represent features in different directions. Then, we predict the contextual position of a randomly cropped volume by contrasting its similarity to different bases. In this way, VoCo effectively encodes the contextual position priors into model representations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experiments demonstrate that VoCo achieves superior performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.788 + ], + "angle": 0, + "content": "We will further consider several ways of extension: (1) Scale up the pre-training dataset to evaluate the upper performance of VoCo. (2) Experiments on more downstream datasets. (3) Evaluate the label-efficient performance of VoCo (e.g., semi-supervised learning). (4) Explore the capacity of VoCo in mining inter-volume relationships." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.659, + 0.817 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by Hong Kong Innovation and Technology Fund (Project No. ITS/028/21FP and No. MHP/002/22), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. T45-401/22-N)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22880" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Michela Antonelli et al. The medical segmentation decathlon. Nature Commun., 13(1):4128, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.183 + ], + "angle": 0, + "content": "[2] Shekoofeh Azizi et al. Big self-supervised models advance medical image classification. In ICCV, pages 3478-3488, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.468, + 0.225 + ], + "angle": 0, + "content": "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. NIPS, 35:8799-8810, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[4] Patrick Bilic et al. The liver tumor segmentation benchmark (lits). Medical Image Analy., 84:102680, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.297 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[6] Mathilde Caron et al. Unsupervised learning of visual features by contrasting cluster assignments. NIPS, 33:9912-9924, 2020. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.368 + ], + "angle": 0, + "content": "[7] Mathilde Caron et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.468, + 0.451 + ], + "angle": 0, + "content": "[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In CVPR, pages 11526-11535, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[10] Ting Chen et al. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.468, + 0.578 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[13] Zekai Chen et al. Masked image modeling advances 3d medical image analysis. In WACV, pages 1970-1980, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[14] Kenneth Clark and Bruce and others Vendl. The cancer imaging archive (tcia): maintaining and operating a public information repository. Jour. of Dig. Imag., 26:1045-1057, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[15] Jiequan Cui et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.733 + ], + "angle": 0, + "content": "[16] Jiequan Cui et al. Generalized parametric contrastive learning. IEEE Trans. Pattern Analy. Mach. Intell., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.774 + ], + "angle": 0, + "content": "[17] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[19] Hao Du, Qihua Dong, Yan Xu, and Jing Liao. Weakly-supervised 3d medical image segmentation using geometric prior and contrastive similarity. IEEE Trans. Medi. Imag., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[20] Yuting Gao, Jia-Xin Zhuang, et al. Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In ECCV, pages 237-253, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[21] Florin-Cristian Ghesu et al. Multi-scale deep reinforcement learning for real-time 3d-landmark detection in ct scans. IEEE Trans. Pattern Anal. Mach. Intell., 41(1):176-189, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[22] Jean-Bastien Grill et al. Bootstrap your own latent-a new approach to self-supervised learning. NIPS, 33:21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[23] Katharina Grünberg et al. Annotating medical image data. Medical Image Analy., pages 45-67, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[24] Fatemeh Haghighi et al. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE Trans. Medical Imag., 40(10):2857-2868, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[25] Fatemeh Haghighi et al. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In CVPR, pages 20824-20834, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[26] Ali Hatamizadeh et al. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In MICCAIW, pages 272-284, 2021. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[27] Ali Hatamizadeh et al. Unetr: Transformers for 3d medical image segmentation. In WACV, pages 574-584, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[28] Kaiming He et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.596 + ], + "angle": 0, + "content": "[31] Xingxin He, Leyuan Fang, Mingkui Tan, and Xiangdong Chen. Intra-and inter-slice contrastive learning for point supervised oct fluid segmentation. IEEE Trans. Image Process., 31:1870-1881, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[32] Yuting He et al. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In CVPR, pages 9538-9547, 2023. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[33] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[34] Yankai Jiang et al. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In ICCV, pages 15859-15869, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[35] Bennett Landman et al. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In MICCAIW, volume 5, page 12, 2015. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[36] Jie Liu et al. Clip-driven universal model for organ segmentation and tumor detection. In ICCV, pages 21152-21164, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[37] Qiang Liu et al. A multi-level label-aware semi-supervised framework for remote sensing scene classification. IEEE Trans. Geosci. Remote Sens., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22881" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[39] Xiangde Luo et al. WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from ct image. Medical Image Analysis, 82:102642, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[40] Jun Ma et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[41] T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In CVPR, pages 9339-9348, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[42] Nguyen et al. Joint self-supervised image-volume representation learning with intra-inter contrastive clustering. In AAAI, pages 14426-14435, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[43] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[44] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[45] Maxime Oquab et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[47] Rodrigo Santa Cruz et al. Deeppermnet: Visual permutation learning. In CVPR, pages 3949-3957, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[48] Arnaud Arindra Adiyoso Setio et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analy., 42:1-13, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[49] Amber L Simpson et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.595 + ], + "angle": 0, + "content": "[50] Aiham Taleb et al. 3d self-supervised methods for medical imaging. NIPS, 33:18158-18172, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[51] Yucheng Tang et al. Self-supervised pre-training of swin transformers for 3d medical image analysis. In CVPR, pages 20730-20740, 2022. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[52] Xing Tao et al. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In MICCAI, pages 238-248, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[53] Guotai Wang et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 41(7):1559-1572, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[54] Xiaosong Wang et al. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, pages 2097-2106, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[55] Yiqing Wang et al. Swinmm: masked multi-view with swim transformers for 3d medical image segmentation. In MIC-CAI, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[56] Xin Wen et al. Self-supervised visual representation learning with semantic grouping. NIPS, 35:16423-16438, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[57] Linshan Wu et al. Deep bilateral filtering network for point-supervised semantic segmentation in remote sensing images. IEEE Trans. Image Process., 31:7419-7434, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[58] Linshan Wu et al. Modeling the label distributions for weakly-supervised semantic segmentation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[59] Linshan Wu, Leyuan Fang, Xingxin He, Min He, Jiayi Ma, and Zhun Zhong. Querying labeled for unlabeled: Cross-image semantic consistency guided semi-supervised semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8827-8844, Jul. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[60] Linshan Wu, Ming Lu, and Leyuan Fang. Deep covariance alignment for domain adaptive remote sensing image segmentation. IEEE Trans. Geosci. Remote Sens., 60:1-11, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.245, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[61] Linshan Wu, Zhun Zhong, Leyuan Fang, Xingxin He, Qiang Liu, Jiayi Ma, and Hao Chen. Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In CVPR, pages 15454-15464, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[62] Yutong Xie, Jianpeng Zhang, Yong Xia, and Qi Wu. Unimiss: Universal medical self-supervised learning via breaking dimensionality barrier. In ECCV, pages 558-575, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[63] Zhenda Xie et al. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[64] Shuangfei Zhai et al. Position prediction as an effective pretraining strategy. arXiv preprint arXiv:2207.07611, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[65] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Anal., 89:102879, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[66] Hongyi Zhang et al. Mixup: Beyond empirical risk minimization. In ICLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.481, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[67] Kang Zhang et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181(6):1423-1433, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.537, + 0.892, + 0.577 + ], + "angle": 0, + "content": "[68] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In AAAI, pages 3516-3524, 2023. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.579, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[69] Hong-Yu Zhou et al. Comparing to learn: Surpassing image-net pretraining on radiographs by comparing image representations. In MICCAI, pages 398-407, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.621, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[70] Hong-Yu Zhou et al. Preservational learning improves self-supervised medical image models by reconstructing diverse contexts. In ICCV, pages 3499-3509, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[71] Hong-Yu Zhou et al. A unified visual information preservation framework for self-supervised pre-training in medical image analysis. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[72] Zongwei Zhou et al. Models genesis. Medical Image Analy., 67:101840, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[73] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[74] Xiahai Zhuang. Multivariate mixture model for myocardial segmentation combining multi-source images. IEEE Trans. Pattern Analy. Mach. Intell., 41(12):2933-2946, 2018. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.856, + 0.892, + 0.896 + ], + "angle": 0, + "content": "[75] Xinrui Zhuang et al. Self-supervised feature learning for 3d medical images by playing a rubik's cube. In MICCAI, pages 420-428, 2019. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22882" + } + ] +] \ No newline at end of file diff --git a/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_origin.pdf b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac987a59a2e540e83fefd51b0037200e8c76e8b5 --- /dev/null +++ b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/8902ed6b-c8cb-4bc7-a572-045797bd1786_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a97fd5159ba1ead906f750108a67eba02af160ae6af1e65afcc7da27d5c9a32 +size 3149302 diff --git a/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/full.md b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4cb38237d80ad9f94c22b4b533686ba85da5f0c --- /dev/null +++ b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/full.md @@ -0,0 +1,331 @@ +# VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis + +Linshan Wu + +Jiaxin Zhuang + +Hao Chen * + +Hong Kong University of Science and Technology + +# Abstract + +Self-Supervised Learning (SSL) has demonstrated promising results in 3D medical image analysis. However, the lack of high-level semantics in pre-training still heavily hinders the performance of downstream tasks. We observe that 3D medical images contain relatively consistent contextual position information, i.e., consistent geometric relations between different organs, which leads to a potential way for us to learn consistent semantic representations in pre-training. In this paper, we propose a simple yet-effective Volume Contrast (VoCo) framework to leverage the contextual position priors for pre-training. Specifically, we first generate a group of base crops from different regions while enforcing feature discrepancy among them, where we employ them as class assignments of different regions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which region) by contrasting their similarity to different base crops, which can be seen as predicting contextual positions of different sub-volumes. Through this pretext task, VoCo implicitly encodes the contextual position priors into model representations without the guidance of annotations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experimental results on six downstream tasks demonstrate the superior effectiveness of VoCo. Code will be available at https://github.com/Luffy03/VoCo. + +# 1. Introduction + +Deep learning has demonstrated outstanding achievements in 3D medical image analysis [53, 21, 40, 33, 39], yet is heavily hampered by the expensive cost of the required expert annotations [50, 23]. To address this problem, Self-Supervised Learning (SSL) has received significant attention due to its promising ability to learn representations without annotations [10, 11, 6, 28, 20], which has become + +![](images/864fec36362eafd2c1ad151515fa3a58137346649cea1e8cb91415b5af2a8e16.jpg) +(a) Relatively consistent contextual positions in 3D medical images + +![](images/abde1250c2e2349e645a2e82439a7d4c9b3b032678330d1dd9fdab3f990dc6d3.jpg) +(b) Volume Contrast (VoCo) for 3D Medical Image Analysis +Figure 1. (a) In 3D medical images, the contextual positions, i.e., the geometric relations between different organs are relatively consistent. (b) Aiming to leverage contextual position priors for pretraining, we proposed a Volume Contrast (VoCo) framework for 3D Medical Image Analysis. + +an important label-efficient solution in 3D medical image analysis [71, 51, 32, 2, 34, 36]. + +Existing methods [50, 75, 71, 13] are mostly based on information reconstructions to learn augment-invariant representations of 3D medical images, which first employ strong data augmentation to the images and then reconstruct the raw information. Specifically, rotate-and-reconstruct [50, 51, 75, 52] proposed to randomly rotate the 3D volumetric images and learn to recover them, which encourages models to learn rotational invariant features. Recent methods [70, 71, 32, 25, 62] further proposed to restore information among different views of the image. PCRL [70, 71] cropped global and local patches then conducted multiscale restorations. GVSL [32] further explored the geometric similarity between multi-scans by affine augmentation and matching. Mask-reconstruct methods [13, 73, 55] are also widely used, which are introduced from MAE [28] and aim to learn representations by masking images and reconstructing the missing pixels. Although promising results have been demonstrated, previous works [52, 32] have + +![](images/1fbbea86f9466374034fabf6a28da9cf19a0135a7f820aa2b9593741e46bafc5.jpg) +(a) Instance-level + +![](images/79a7a20cd8fab64dad775dc75401d3df656deea6b5e6faf3ce22de6ab4626638.jpg) +(b) Prototype-level +Figure 2. Typical contrastive learning frameworks. (a) Instance-level contrastive learning [10, 11, 29, 22, 7] employs strong data augmentation or model perturbation on input data to acquire different views of instance, then regularizes their consistency. (b) Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] conducts (1) online clustering or (2) randomly initialize then online update process to obtain prototypes as class assignments, then leverage the prototypes to contrast each input image. (c) Our VoCo follows the idea of prototype-level contrastive learning. Specifically, instead of using time-consuming online clustering and updating procedures, we leverage the valuable contextual position priors of 3D medical images and leverage the base crops to generate prototypes (bases). + +![](images/80fbd5499995cb2ac570b093809e1502ef0fb9b200e7b699afbe748d23a3280e.jpg) +(c) Our VoCo + +proved that the lack of high-level semantics in pre-training will heavily hinder the performance of downstream tasks. To address this challenge, we argue that stronger high-level semantics should be further involved into 3D medical image pre-training. + +To this end, we argue that the contextual position priors of 3D medical images should be further exploited. As shown in Fig. 1(a), we observe that in 3D medical images, different organs (semantic regions) contain relatively consistent contextual positions with relatively consistent anatomic characteristics (shapes). Thus, the consistency of geometric relations between different organs leads to a potential way for us to learn consistent semantic representations for 3D medical images pre-training. In this paper, we propose a pretext task for contextual position predictions, which aims to encode contextual position priors into model representations and enables us to effectively improve the performance of downstream tasks that require high-level semantics. + +In this paper, we propose a simple-yet-effective Volume Contrast (VoCo) framework for 3D medical image analysis, as shown in Fig. 1(b). Specifically, we first crop a group of non-overlap volumes from different positions while enforcing feature discrepancy among them. We represent these volumes as a group of bases in the learned high-dimension space, where we employ them as class assignments of different positions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which position) by contrasting their similarity to different bases, which can be seen as predicting contextual positions of different sub-volumes. In this way, we formulate a contextual position prediction pretext task for 3D medical image SSL. Through learning to predict contextual positions, we implicitly involve the high-level semantic priors into the model representations, which enables us to significantly improve the performance of downstream tasks. Extensive experimental results on six downstream tasks demonstrate that our + +proposed VoCo clearly outperforms existing state-of-the-art 3D medical image SSL methods. + +# 2. Related Works + +In this section, we first introduce the previous mainstream contrastive learning paradigms. Then, we survey the existing SSL methods for medical image analysis, especially for 3D medical images. Finally, we review the position-related SSL methods for comparisons with our method and highlight the differences. + +Contrastive learning. Contrastive learning is one of the mainstream paradigms in SSL, which aims to learn consistent representations by contrasting positive and negative pairs of samples without extra annotations [10, 6, 29, 11]. According to [6], instance- and prototype-level contrastive learning are two typical types of contrastive learning, as shown in Fig. 2. + +Instance-level contrastive learning [10, 11, 29, 22, 7] transforms input images with different augmentations or model perturbations, aiming to compare the features from each other. Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] proposes to generate prototypes (also called clusters or bases) for contrasting each input image. Specifically, there are two typical ways to generate prototypes. First, Caron et al. proposed DeepCluster [5] to conduct online clustering on the whole dataset to generate prototypes. However, it is very time-consuming to calculate clusters on a large dataset. Thus, some recent works [6, 56, 15, 16] propose to randomly initialize a group of prototypes and then update them through back-propagation during training, which has demonstrated promising results. However, there is still no explicit guarantee that these randomly initialized prototypes can be updated well during training. + +Our VoCo follows the primary idea of prototype-level constrasive learning. As shown in Fig. 2(c), to address the existing problems mentioned above, instead of randomly + +initializing and updating prototypes, VoCo leverages the valuable contextual position priors of 3D medical images to generate base crops as prototypes, which also requires no time-consuming clustering on a large dataset. + +SSL for medical image analysis. Due to the high potential in label-efficient learning [29, 59, 61, 58, 57, 60, 37], SSL has also received significant attention in the field of medical image analysis [70, 32, 31, 51, 19]. Existing methods are mainly based on comparative SSL [71]. Specifically, Zhou et al. [69] combined Mixup [66] into MoCo [29] to learn the diversity of positive and negative samples in InfoNCE [44]. Azizi et al. used multi-instance learning to compare multiple views of images from each patient. There are also a number of approaches [25, 70, 71] that supervising the models via restoring low-level information from raw images. + +In 3D medical image analysis, reconstructing raw information is a popular pretext task for learning representations [50, 51, 71]. Existing methods are mainly based on reconstructing information from augmented images. These previous methods first conducted strong data augmentation, e.g., rotate [51, 75, 52], multi-view crops [70, 71, 32], and mask [13, 73, 55], then supervised the model by reconstructing raw 3D information. Although promising results have been demonstrated, most of these methods still largely ignore the importance of integrating high-level semantics into model representations, which heavily hinders the performance of downstream tasks. + +Position-related SSL. Position-related SSL methods are also explored in a number of previous works [8, 9, 41, 47, 43, 17, 64, 68] in the field of natural images. Noroozi et al. [43] proposed to predict the order of a set of shuffled patches. Zhai et al. [64] and Caron et al. [8] proposed to train a ViT [18] to predict the locations of each input patch. However, since the geometric relations of different objects are not very consistent in natural images, it is still difficult to effectively learn consistent position representations given visual appearance only (as stated in [68]). In addition, previous works [64, 8, 68] mainly trained a linear layer to output the positions directly, which works in a black-box manner. + +In this paper, we introduce the pretext task of contextual position prediction into the field of 3D medical images, where the geometric relations between different organs are relatively consistent, which guides us to learn consistent semantic representations in pre-training. Different from the previous methods, in this paper, we introduce a totally different position prediction paradigm. Specifically, instead of using a linear layer to output positions directly, we predict the contextual positions based on volume contrast, which is more intuitive and effective. + +# 3. Methodology + +In this section, we first introduce the overall framework of our proposed VoCo in Section 3.1. After that, we present the process of contextual position prediction in Section 3.2. Then, the regularization process via volume contrast in our proposed VoCo framework is described in Section 3.3. + +# 3.1. Overall Framework + +The overall framework of our proposed VoCo is presented in Fig. 3, which contains a contextual position prediction branch and a regularization branch. The prediction branch is used to predict the contextual positions between different cropped volumes. Specifically, given an input volume, we first crop it into non-overlap base volumes, which cover the whole input volume. Then, we randomly crop a volume and transform it into the high-dimension feature space using a typical backbone (CNN [30] or Transformer [18]). The goal is to predict the contextual positions between the randomly cropped volumes and base volumes. In this paper, instead of training a linear classifier to predict positions as in previous works [64, 9, 8, 68], we propose to establish this goal by volume contrast. We develop a loss function $L_{pred}$ to supervise the final predictions. In addition, we further use a loss function $L_{reg}$ to regularize the feature discrepancy from different bases by enlarging their distance, aiming to learn more discriminative class assignments. The details are presented in Section 3.2 and 3.3. + +# 3.2. Contextual Position Prediction + +Base and random crops. Given an input volume, we first crop it into $n$ non-overlap base volumes, which cover the whole input volume. We then employ the extracted features $z$ as class assignments (we call them bases), which present the prototype-level features from different positions. Then, following previous SSL works [10, 11, 29], a projector with linear layers is used to project $z$ into latent features $q$ . Then, we randomly crop a volume and transform it into high-dimension feature space as $p$ . The backbone and projector are also used to project the features from the randomly cropped volumes. + +Volume contrast for contextual position prediction. With features extracted from the backbone and projector, following previous SSL works [10, 11, 29], we first conduct 3D adaptive average pooling to resize them to one dimension, i.e., $p \in \mathbb{R}^{1 \times C}$ and $q \in \mathbb{R}^{1 \times C}$ , where $C$ is the number of channels. Specifically, we empirically set $C$ to 2048 as in [10, 11, 29]. + +Then, we calculate the similarity logits $l$ between $p$ and $q_{i}$ . Specifically, we use cosine similarity to compute $l$ as follows: + +$$ +l _ {i} = \operatorname {C o s S i m} (p, q _ {i}) = \frac {p \cdot q _ {i}}{\| p \| \| q _ {i} \|}, i \in n \tag {1} +$$ + +![](images/7f4b4b75e18847a633dde5f16f7d9d40aab1d9327b86900306c6e446583248b9.jpg) +Figure 3. The overall framework of VoCo. VoCo contains a prediction branch and a regularization branch. The prediction branch is responsible for predicting contextual positions between different sub-volumes. The regularization branch is employed to enforce the feature discrepancy between different bases, which aims to learn more discriminative class assignments. + +where $q_{i}$ is the projected feature of each base crop. $l_{i}$ denotes the similarity between $p$ in $q_{i}$ , which ranges from 0 to 1. It is worth noting that, we stop the gradients of $q$ when computing Eq. 1, which aims to avoid feature collapse [10, 11, 6]. + +Intuitively, higher $l_{i}$ represents that $p$ has higher probabilities to share overlap regions with $q_{i}$ . In this way, we can explicitly associate the similarity value with the position information, i.e., $p$ with higher $l_{i}$ is more likely to be located in the region of the $i_{th}$ base. Thus, instead of training a black-box linear layer, we predict the contextual positions by volume contrast, which is more intuitive and effective. + +Position labels generation. The process of generating position labels is shown in Fig. 4. As shown in Fig. 4, when we generate $n = 4 \times 4$ base crops, there will be $n$ class assignments. Then we calculate the overlap area between a randomly cropped volume and $n$ base crops. The proportions of the overlap area are then assigned as position labels $y$ , which also range from 0 to 1. Thus, we can easily supervise the model by calculating the distance between the prediction logits $l$ and position labels $y$ . The setting of the number $n$ of base crops will be discussed in Section 4.4. + +Loss function for contextual position prediction. The formulation of prediction loss function $L_{pred}$ is based on entropy. Specifically, we first calculate the distance $d$ between prediction logits $l$ and position labels $y$ : + +$$ +d _ {i} = \left| y _ {i} - l _ {i} \right|, i \in n, \tag {2} +$$ + +where $|.|$ denotes the absolute value. Then, $L_{pred}$ is formu + +![](images/52712bc928b1b014ead126e6371ce21df02539f86b5fabfa38b6e3e7eaacc0b1.jpg) +Figure 4. The process of generating position labels. We calculate the proportions of overlap area as position labels $y$ , e.g., the randomly cropped volume in the figure is assigned to class 5, 6, 9, and 10 with probabilities of 0.25, 0.1, 0.5, and 0.15, respectively. + +lated as follows: + +$$ +L _ {p r e d} = - \frac {1}{n} \sum_ {i \in n} ^ {n} \log \left(1 - d _ {i}\right). \tag {3} +$$ + +It is worth noting that VoCo predicts contextual positions of a volume (high similarities with all its contextual overlapped volumes), thus don't need one-to-one correspondence: e.g., in Fig. 4, high-value $l_{i}$ pertain to $y_{i} > 0, i = 5,6,9,10$ simultaneously. Then we calculate the distance between $l_{i}$ and $y_{i}$ (Eq. 2). + +# 3.3. Volume Contrast for Regularization + +We aim to learn more discriminative class assignments (bases) for volume contrast. Since intuitively, different subvolumes tend to contain different organs (semantic discrepancy). + +ancy). Thus, we aim to enlarge the high-dimension feature discrepancy between different bases. To this end, we design a regularization loss $L_{reg}$ to enlarge the feature discrepancy between different bases $z$ . + +First, given projected bases $q$ , we also calculate the cosine similarity $s_{ij}$ between different $q_i$ and $q_j$ as follows: + +$$ +s _ {i j} = \operatorname {C o s S i m} \left(q _ {i}, q _ {j}\right) = \frac {q _ {i} \cdot q _ {j}}{\| q _ {i} \| \| q _ {j} \|}, i, j \in n, i \neq j, \tag {4} +$$ + +where we aim to regularize $s_{ij}$ to 0, enforcing feature discrepancy between different bases. Thus, the loss function $L_{reg}$ is formulated as: + +$$ +L _ {r e g} = \frac {2}{n (n - 1)} \sum_ {i, j \in n, i \neq j} ^ {n} | s _ {i j} |, \tag {5} +$$ + +where $|.|$ denotes the absolute value. With loss $L_{reg}$ , we aim to optimize $q$ as linearly independent bases: + +$$ +q _ {i} \bot q _ {j}, i \neq j, i, j \in n. \tag {6} +$$ + +With the regularization loss function $L_{reg}$ , we aim to learn a group of linearly independent bases to represent all directions of high-dimension features [6]. In this way, we can learn a group of more discriminative class assignments to supervise the final position predictions. + +Overall loss function. Thus, the total loss function $L$ is the combination of $L_{reg}$ and $L_{pred}$ : + +$$ +L = L _ {p r e d} + \lambda L _ {r e g}, \tag {7} +$$ + +where $\lambda$ is used to balance the relative contributions of these two loss terms and set to 1.0 in experiments empirically since we consider their importance equally. The ablation studies of $\lambda$ are provided in the supplementary materials. + +# 4. Experiments + +In this section, we first describe the datasets used in the pre-training and downstream tasks. Then, we briefly introduce the implementation details of VoCo. Finally, we report detailed experiment results of our proposed VoCo compared with other state-of-the-art SSL methods in 3D medical images. More details are in the supplementary materials. + +# 4.1. Datasets + +Pre-training datasets. Aiming to conduct fair comparisons with the previous works [51, 55, 70, 71, 13, 73], we also carry out pre-training experiments on the same three public datasets, i.e., BTCV [35], TCIA Covid19 [14], and LUNA [48] datasets, including about totally $1.6\mathrm{k}$ CT scans for pre-training. It is worth noting that, aiming to conduct fair comparisons with previous works [73, 13], we only use BTCV [35] and TCIA Covid-19 [14] for pre-training in + +the downstream experiments of BTCV [35]. For the other downstream tasks, we use all three datasets for pre-training. Details are provided in the supplementary materials. + +Downstream datasets. To evaluate the effectiveness of our VoCo, we conduct downstream experiments on six public datasets, i.e., BTCV [35], LiTs [4], MSD Spleen [1], MM-WHS [74], BraTS 21 [49], and CC-CCII [67], including segmentation and classification tasks. The first five datasets are developed for segmentation, while CC-CCII [67] is for COVID-19 classification. Note that only BTCV [35] is used in pre-training, the other datasets are unseen in pre-training. In addition, to evaluate the cross-modality generalization ability, we transfer the model pre-trained on the CT dataset to the MRI dataset BraTS 21 [49]. We adopt consistent settings as previous works [13, 73, 26, 55, 32]. We also evaluate the performance on 2D medical dataset [54]. Details are provided in the supplementary materials. + +# 4.2. Implementation details + +Following previous works [51, 55], we use SwinUNETR [26] for both pre-training and downstream tasks. We use AdamW [38] optimizer and cosine learning rate scheduler for all experiments. We set 100K training steps in the pre-training process and applied a slicing window inference for fair comparisons with previous works [51, 55, 13, 73]. Aiming to evaluate the pure effectiveness, we do not use foundation models or post-processing [34, 36]. Details are provided in the supplementary materials. + +Comparison methods. We compare our VoCo with both General and Medical SSL methods. First, we compare with the typical SSL methods MAE [28, 13] and MoCo v3 [29, 12], since they represent the two mainstream SSL paradigms, i.e., mask-autoencoder and constrastive learning. Since it is not practical to set a large batch size for 3D medical images due to computation cost, for fair comparisons, we adopt consistent settings with other methods in MAE [28, 13] and MoCo v3 [29, 12]. We also report the results of SimCLR [10] and SimMIM [63] according to [13]. We further evaluate the performance of Jiasaw [9] and PositionLabel [68], since they are related to our position-aware methods. Most existing state-of-the-art medical SSL methods are compared in our experiments. + +# 4.3. Experiments on downstream tasks + +Outperform existing methods on the BTCV dataset. We first conduct experiments on BTCV [35], as shown in Table 1. Specifically, among the comparison methods, MAE3D [28, 13], SimCLR [10], SimMIM [63], MoCo v3 [29, 12], and GL-MAE [73] use UNETR [27]. Other methods including our VoCo adopt Swin-UNETR [26] as the default settings of the previous work [51]. + +Remark. It can be seen in Table 1 that the general SSL methods perform worse than most medical SSL methods. + +
MethodDice Score(%)AVG
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAG
From Scratch
UNETR [27]93.0294.1394.1266.9970.8796.1177.2789.2282.1070.1676.6565.3259.2179.82
Swin-UNETR† [26]94.0693.5493.8065.5174.6097.0975.9491.8082.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [28, 13]93.9894.3794.1869.8674.6596.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [10]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [63]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3† [29, 12]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw† [9]94.6293.4193.5575.6373.2195.7180.8089.4184.7871.0279.5765.6860.2281.35
PositionLabel† [68]94.3593.1593.2175.3973.2495.7680.6988.8084.0471.1879.0265.1160.0781.09
With Medical SSL
MG [72]91.9993.5291.8165.1176.1495.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [50]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [3]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++† [52]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1† [70]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2† [71]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [26, 51]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [55]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [73]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL† [32]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
+ +Table 1. Experimental results on BTCV [35]. The best results are bolded. 'From Scratch' denotes the supervised baseline without self-supervised pre-training. † denotes we re-implement the approach. Most results are drawn from [13, 65, 73] or their own papers. + +
MethodNetworkDice Score(%)
From Scratch
3D UNet [46]-90.70
UNETR† [27]-93.25
Swin-UNETR† [26]-93.42
With General SSL
Jigsaw [9]3D UNet94.36
MAE3D [28, 13]UNETR94.02
MoCo v3† [29, 12]UNETR93.86
Jigsaw† [9]Swin-UNETR95.24
PositionLabel† [68]Swin-UNETR94.13
With Medical SSL
MG [72]3D UNet91.30
TransVW [24]3D UNet91.42
ROT [50]3D UNet94.49
PCRLv1 [70]3D UNet93.87
PCRLv2 [71]3D UNet94.50
Rubik [75]3D UNet94.93
Rubik++ [52]3D UNet95.46
Rubik++† [52]Swin-UNETR95.72
Swin-UNETR [51, 26]Swin-UNETR95.33
SwinMM [55]Swin-UNETR95.52
VoCo3D UNet96.03
VoCoSwin-UNETR96.52
+ +Specifically, MoCo v3 [29, 12] can only achieve $79.54\%$ Dice Score. Since MoCo v3 [29, 12] heavily relies on a large batch size to acquire adequate negative samples, which is not practical in 3D medical images due to the huge computation burden. In addition, the negative relation between different images used in MoCo [29, 12] is not ap + +Table 2. Experimental results on LiTs [4]. We report the Dice Scores of liver segmentation. $\dagger$ denotes we re-implement the approach. Most results are drawn from [65, 71]. + +
MethodMSD SpleenMM-WHS
From Scratch
3D UNet [46]93.7183.09
UNETR† [27]94.2085.85
Swin-UNETR† [26]94.6386.11
With General SSL
MAE3D [28, 13]95.2086.03
MoCo v3† [29, 12]94.3284.16
Jiasaw† [9]94.2985.88
PositionLabel† [68]94.1685.52
With Medical SSL
MG [72]94.4086.36
VicRegl [3]94.1284.72
UniMiss [62]95.0984.68
PCRLv1† [70]94.3286.58
PCRLv2† [71]94.9486.82
Rubik++† [52]95.1187.02
Swin-UNETR [51, 26]95.0287.06
SwinMM [55]95.3486.98
JSSL [42]94.9284.89
GVSL [32]95.4788.27
VoCo96.3490.54
+ +Table 3. Experimental results on MSD Spleen [1] and MM-WHS [74]. We report the Dice Scores of segmentation prediction. $\dagger$ denotes we re-implement the approach. + +propriate in medical images. MAE [28, 13], SimCLR [10], and SimMIM [63] (results from [13]) also gain limited performance. Our VoCo also outperforms the position-based methods Jigsaw [9] and PositionLabel [68] by a clear margin. Thus, we conclude that general SSL methods are not very suitable for 3D medical images. It is crucial to consider the characteristics of medical images in medical SSL. + +The scratch Swin-UNETR [26] only achieves $80.53\%$ + +
MethodNet.Dice Score(%)
TCWTETAVG
From Scratch
UNETR [27]-81.6287.8157.3475.58
Swin-UNETR [26]-81.2888.6757.7375.89
With General SSL
MAE3D [28, 13]UNETR82.3490.3559.1877.29
SimMIM [63]UNETR84.0690.4359.0777.85
SimCLR [10]UNETR83.1389.4458.4276.99
MoCo v3† [29, 12]UNETR82.6088.8957.6976.39
Jigsaw† [9]Sw-UNE.81.6289.4559.1076.72
PositionLabel† [68]Sw-UNE.81.3589.6258.7376.64
With Medical SSL
PCRLv1†[70]Sw-UNE.81.9688.8357.5876.12
PCRLv2†[71]Sw-UNE.82.1390.0657.7076.63
Rubik++†[52]Sw-UNE.84.3290.2358.0177.51
Swin-UNETR [51, 26]Sw-UNE.82.5189.0858.1576.58
SwinMM [55]Sw-UNE.83.4890.4758.7277.56
VoCoSw-UNE.85.2790.4559.8778.53
+ +Table 4. Experimental results on BraTS 21 [49]. WT, TC, and ET denote the whole tumor, tumor core, and enhancing tumor, respectively. $\dagger$ denotes we re-implement the approach. + +
MethodNetworkAccuracy(%)
From Scratch
UNETR [27]-88.92
Swin-UNETR [26]-88.04
With General SSL
MAE3D [28, 13]UNETR89.47
MoCo v3 [29, 12]UNETR84.95
Jiasaw [9]Swin-UNETR86.88
PositionLabel [68]Swin-UNETR87.54
With Medical SSL
PCRLv1 [70]Swin-UNETR88.72
PCRLv2 [71]Swin-UNETR89.15
Rubik++ [52]Swin-UNETR89.23
Swin-UNETR [51, 26]Swin-UNETR89.45
SwinMM [55]Swin-UNETR89.61
VoCoSwin-UNETR90.83
+ +Dice Score. With VoCo pre-training, we gain $3.32\%$ improvements with $83.85\%$ Dice Score, which also outperforms existing methods by a clear margin. Among the compared methods, GL-MAE [73] achieves the highest Dice Score $(82.01\%)$ . Our VoCo surpasses it by $1.84\%$ Dice Score, which is a clear improvement in this dataset. + +Promising performance on Unseen datasets. We further conduct experiments on unseen datasets in pre-training, i.e., LiTs [4], MSD Spleen [1], and MM-WHS [74]. The results on LiTs [4] are shown in Table 2. We report the results of compared methods according to [70, 71, 65]. Since the scratch Swin-UNETR [26] can obtain a higher Dice Score (93.42%), we further pre-train a 3D UNet [46] based on VoCo, aiming to conduct fair comparisons. It can be seen that with VoCo pre-training, Swin-UNETR [26] gains $3.10\%$ improvements and achieves $96.52\%$ Dice Score. With 3D UNet [46] as the backbone, VoCo also achieves $96.03\%$ Dice Score, proving the effectiveness of VoCo with + +![](images/8a0f0ab2186be6e026c2b52d348737a6b476da5fb8f1e306b15f06f5ba9cdb5f.jpg) +Figure 5. Overall comparisons with state-of-the-art methods on six different datasets. + +Table 5. Experimental results of CC-CCII [67] classification. + +
Loss FunctionsBTCVMM-WHS
LpredLreg
XX80.5386.11
✓X82.9688.82
✓✓83.8590.54
+ +Table 6. Evaluation of loss functions $L_{pred}$ and $L_{reg}$ . We report the average Dice Score on BTCV [35] and MM-WHS [74]. + +
Number of bases nBTCVMM-WHS
(x,y,z)
2×2×181.5686.73
3×3×182.7789.31
4×4×183.8590.54
5×5×183.6090.49
3×3×282.8689.14
4×4×283.4790.52
+ +Table 7. Evaluation of the value of bases $n$ . We report the average Dice Score on BTCV [35] and MM-WHS [74]. + +different network architectures. + +The results on MSD Spleen [1] and MM-WHS [74] datasets are shown in Table 3. In previous methods, GVSL [32] achieves the best performance with $95.47\%$ and $88.27\%$ Dice Score, while our VoCo surpasses all previous methods with $96.34\%$ and $90.54\%$ Dice Score on MSD Spleen [1] and MM-WHS [74] datasets, respectively. + +Generalization capacity on MRI dataset. To verify the generalization capacity on the MRI dataset, we further evaluate the performance of VoCo on BraTS 21 [49]. As shown in Table 4, VoCo achieves $78.53\%$ Dice Score and outperforms existing state-of-the-art methods, demonstrating the cross-model generalization capacity of VoCo. + +Evaluation of COVID-19 classification. We further evaluate the performance of the classification task on the CC-CCII [67] dataset in Table 5. Since existing SSL methods did not conduct experiments on this dataset, we reproduce the related methods for comparisons. It can be seen that VoCo can also achieve superior results with $90.83\%$ accuracy, proving its effectiveness in the classification task. + +Overall comparisons on six downstream datasets. The + +![](images/4ab02bfc61edd815b2b5f76aeffa029525242aa7f5b890fa09ffa6cf03bb3607.jpg) +Figure 6. Qualitative visualization of segmentation results for the BTCV [35] dataset. We compare VoCo with MoCo v3 [12], PositionLabel [68], SwinUNETR [51, 26], and GVSL [32]. + +overall comparisons are shown in Fig. 5. Specifically, we compare with the existing state-of-the-art methods on six different downstream datasets. It can be seen that our VoCo outperforms them by a clear margin. + +# 4.4. Ablation Study + +We further conduct ablation studies to evaluate the loss functions and the settings in VoCo, which are verified on the BTCV [35] and MM-WHS [74] datasets. + +Loss functions. We first study the importance of the two loss functions, i.e., $L_{pred}$ and $L_{reg}$ , as shown in Table 6. It can be seen that with our proposed $L_{pred}$ loss function, the performance is significantly improved, i.e., $80.53\%$ to $82.96\%$ on BTCV, $86.11\%$ to $88.82\%$ on MM-WHS. These results demonstrate the effectiveness of our proposed position prediction pretext task. In addition, with the proposed regularization loss $L_{reg}$ , the performance can be further improved. Thus, we can see that it is crucial to learn discriminative bases in VoCo. + +Number of bases. We further evaluate different settings of the number of bases $n$ in VoCo. We compare with different settings of $n$ in the ablation studies, as shown in Table 7. It is worth noting that due to the ROI size inconsistency in the $Z$ direction, it is not practical to crop multiple bases in the $Z$ direction, since we have to resize the volume after crops, which will result in inconsistent volume scales. In addition, due to the computation limitation, it is costly to increase the values of $n$ . As shown in Table 7, with $n = 2 \times 2 \times 1$ , the VoCo only achieves $81.56\%$ and $86.73\%$ Dice Score on BTCV and MM-WHS, respectively. When we increase the values of $n$ to $3 \times 3 \times 1$ and $4 \times 4 \times 1$ , the performances are improved obviously. Specifically, with $n$ as $4 \times 4 \times 1$ , we achieve $83.85\%$ and $90.54\%$ on BTCV and MM-WHS, respectively. However, we observe that higher $n$ ( $5 \times 5 \times 1$ ) cannot further bring higher performance. We further verify the performance of generating base crops in the $Z$ direction. It can be seen that $3 \times 3 \times 2$ and $4 \times 4 \times 2$ cannot yield improvements. Thus, aiming to balance the performance + +and efficiency, we set $n$ as $4 \times 4 \times 1$ in VoCo. It can be seen that the setting of $n$ is crucial to VoCo. + +Visualization results on BTCV [35] are shown in Fig. 6. It can be seen that VoCo can yield improved segmentation accuracy and completeness. More visualization results are in the supplementary materials. + +# 5. Conclusion and Future Directions + +In this paper, we develop a simple-yet-effective SSL framework VoCo for 3D medical image analysis. Motivated by the observation that 3D medical images contain relatively consistent contextual positions between different organs, we propose to leverage the contextual position priors to learn consistent semantic representations in pre-training. Specifically, we crop volumes from different positions in an input volume and represent them as a group of bases to represent features in different directions. Then, we predict the contextual position of a randomly cropped volume by contrasting its similarity to different bases. In this way, VoCo effectively encodes the contextual position priors into model representations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experiments demonstrate that VoCo achieves superior performance. + +We will further consider several ways of extension: (1) Scale up the pre-training dataset to evaluate the upper performance of VoCo. (2) Experiments on more downstream datasets. (3) Evaluate the label-efficient performance of VoCo (e.g., semi-supervised learning). (4) Explore the capacity of VoCo in mining inter-volume relationships. + +# Acknowledgments + +This work was supported by Hong Kong Innovation and Technology Fund (Project No. ITS/028/21FP and No. MHP/002/22), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. T45-401/22-N). + +# References + +[1] Michela Antonelli et al. The medical segmentation decathlon. Nature Commun., 13(1):4128, 2022. 5, 6, 7 +[2] Shekoofeh Azizi et al. Big self-supervised models advance medical image classification. In ICCV, pages 3478-3488, 2021. 1 +[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. NIPS, 35:8799-8810, 2022. 6 +[4] Patrick Bilic et al. The liver tumor segmentation benchmark (lits). Medical Image Analy., 84:102680, 2023. 5, 6, 7 +[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2 +[6] Mathilde Caron et al. Unsupervised learning of visual features by contrasting cluster assignments. NIPS, 33:9912-9924, 2020. 1, 2, 4, 5 +[7] Mathilde Caron et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2 +[8] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 3 +[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In CVPR, pages 11526-11535, 2021. 3, 5, 6, 7 +[10] Ting Chen et al. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 2, 3, 4, 5, 6, 7 +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 1, 2, 3, 4 +[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 5, 6, 7, 8 +[13] Zekai Chen et al. Masked image modeling advances 3d medical image analysis. In WACV, pages 1970-1980, 2023. 1, 3, 5, 6, 7 +[14] Kenneth Clark and Bruce and others Vendl. The cancer imaging archive (tcia): maintaining and operating a public information repository. Jour. of Dig. Imag., 26:1045-1057, 2013. 5 +[15] Jiequan Cui et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 2 +[16] Jiequan Cui et al. Generalized parametric contrastive learning. IEEE Trans. Pattern Analy. Mach. Intell., 2023. 2 +[17] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 3 +[18] Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 3 +[19] Hao Du, Qihua Dong, Yan Xu, and Jing Liao. Weakly-supervised 3d medical image segmentation using geometric prior and contrastive similarity. IEEE Trans. Medi. Imag., 2023. 3 +[20] Yuting Gao, Jia-Xin Zhuang, et al. Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In ECCV, pages 237-253, 2022. 1 + +[21] Florin-Cristian Ghesu et al. Multi-scale deep reinforcement learning for real-time 3d-landmark detection in ct scans. IEEE Trans. Pattern Anal. Mach. Intell., 41(1):176-189, 2017. 1 +[22] Jean-Bastien Grill et al. Bootstrap your own latent-a new approach to self-supervised learning. NIPS, 33:21271-21284, 2020. 2 +[23] Katharina Grünberg et al. Annotating medical image data. Medical Image Analy., pages 45-67, 2017. 1 +[24] Fatemeh Haghighi et al. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE Trans. Medical Imag., 40(10):2857-2868, 2021. 6 +[25] Fatemeh Haghighi et al. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In CVPR, pages 20824-20834, 2022. 1, 3 +[26] Ali Hatamizadeh et al. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In MICCAIW, pages 272-284, 2021. 5, 6, 7, 8 +[27] Ali Hatamizadeh et al. Unetr: Transformers for 3d medical image segmentation. In WACV, pages 574-584, 2022. 5, 6, 7 +[28] Kaiming He et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 5, 6, 7 +[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5, 6, 7 +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3 +[31] Xingxin He, Leyuan Fang, Mingkui Tan, and Xiangdong Chen. Intra-and inter-slice contrastive learning for point supervised oct fluid segmentation. IEEE Trans. Image Process., 31:1870-1881, 2022. 3 +[32] Yuting He et al. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In CVPR, pages 9538-9547, 2023. 1, 3, 5, 6, 7, 8 +[33] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021. 1 +[34] Yankai Jiang et al. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In ICCV, pages 15859-15869, 2023. 1, 5 +[35] Bennett Landman et al. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In MICCAIW, volume 5, page 12, 2015. 5, 6, 7, 8 +[36] Jie Liu et al. Clip-driven universal model for organ segmentation and tumor detection. In ICCV, pages 21152-21164, 2023. 1, 5 +[37] Qiang Liu et al. A multi-level label-aware semi-supervised framework for remote sensing scene classification. IEEE Trans. Geosci. Remote Sens., 2023. 3 +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 + +[39] Xiangde Luo et al. WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from ct image. Medical Image Analysis, 82:102642, 2022. 1 +[40] Jun Ma et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2021. 1 +[41] T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In CVPR, pages 9339-9348, 2018. 3 +[42] Nguyen et al. Joint self-supervised image-volume representation learning with intra-inter contrastive clustering. In AAAI, pages 14426-14435, 2023. 6 +[43] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84. Springer, 2016. 3 +[44] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3 +[45] Maxime Oquab et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2 +[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 6, 7 +[47] Rodrigo Santa Cruz et al. Deeppermnet: Visual permutation learning. In CVPR, pages 3949-3957, 2017. 3 +[48] Arnaud Arindra Adiyoso Setio et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analy., 42:1-13, 2017. 5 +[49] Amber L Simpson et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 5, 7 +[50] Aiham Taleb et al. 3d self-supervised methods for medical imaging. NIPS, 33:18158-18172, 2020. 1, 3, 6 +[51] Yucheng Tang et al. Self-supervised pre-training of swin transformers for 3d medical image analysis. In CVPR, pages 20730-20740, 2022. 1, 3, 5, 6, 7, 8 +[52] Xing Tao et al. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In MICCAI, pages 238-248, 2020. 1, 3, 6, 7 +[53] Guotai Wang et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 41(7):1559-1572, 2018. 1 +[54] Xiaosong Wang et al. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, pages 2097-2106, 2017. 5 +[55] Yiqing Wang et al. Swinmm: masked multi-view with swim transformers for 3d medical image segmentation. In MIC-CAI, 2023. 1, 3, 5, 6, 7 +[56] Xin Wen et al. Self-supervised visual representation learning with semantic grouping. NIPS, 35:16423-16438, 2022. 2 +[57] Linshan Wu et al. Deep bilateral filtering network for point-supervised semantic segmentation in remote sensing images. IEEE Trans. Image Process., 31:7419-7434, 2022. 3 + +[58] Linshan Wu et al. Modeling the label distributions for weakly-supervised semantic segmentation, 2024. 3 +[59] Linshan Wu, Leyuan Fang, Xingxin He, Min He, Jiayi Ma, and Zhun Zhong. Querying labeled for unlabeled: Cross-image semantic consistency guided semi-supervised semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8827-8844, Jul. 2023. 3 +[60] Linshan Wu, Ming Lu, and Leyuan Fang. Deep covariance alignment for domain adaptive remote sensing image segmentation. IEEE Trans. Geosci. Remote Sens., 60:1-11, 2022. 3 +[61] Linshan Wu, Zhun Zhong, Leyuan Fang, Xingxin He, Qiang Liu, Jiayi Ma, and Hao Chen. Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In CVPR, pages 15454-15464, 2023. 3 +[62] Yutong Xie, Jianpeng Zhang, Yong Xia, and Qi Wu. Unimiss: Universal medical self-supervised learning via breaking dimensionality barrier. In ECCV, pages 558-575, 2022. 1, 6 +[63] Zhenda Xie et al. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5, 6, 7 +[64] Shuangfei Zhai et al. Position prediction as an effective pretraining strategy. arXiv preprint arXiv:2207.07611, 2022. 3 +[65] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Anal., 89:102879, 2023. 6, 7 +[66] Hongyi Zhang et al. Mixup: Beyond empirical risk minimization. In ICLR, 2018. 3 +[67] Kang Zhang et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181(6):1423-1433, 2020. 5, 7 +[68] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In AAAI, pages 3516-3524, 2023. 3, 5, 6, 7, 8 +[69] Hong-Yu Zhou et al. Comparing to learn: Surpassing image-net pretraining on radiographs by comparing image representations. In MICCAI, pages 398-407, 2020. 3 +[70] Hong-Yu Zhou et al. Preservational learning improves self-supervised medical image models by reconstructing diverse contexts. In ICCV, pages 3499-3509, 2021. 1, 3, 5, 6, 7 +[71] Hong-Yu Zhou et al. A unified visual information preservation framework for self-supervised pre-training in medical image analysis. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 3, 5, 6, 7 +[72] Zongwei Zhou et al. Models genesis. Medical Image Analy., 67:101840, 2021. 6 +[73] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 1, 3, 5, 6, 7 +[74] Xiahai Zhuang. Multivariate mixture model for myocardial segmentation combining multi-source images. IEEE Trans. Pattern Analy. Mach. Intell., 41(12):2933-2946, 2018. 5, 6, 7, 8 +[75] Xinrui Zhuang et al. Self-supervised feature learning for 3d medical images by playing a rubik's cube. In MICCAI, pages 420-428, 2019. 1, 3, 6 \ No newline at end of file diff --git a/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/images.zip b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1745572f67ba66550e9453b750c46c720a4480ad --- /dev/null +++ b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11565de4c156160260ceb761f4a34273eaaa5360e14df81d9c8b475fefa1da8e +size 723202 diff --git a/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/layout.json b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..205fa465259a88a9ac20d264cf94d8b18f64dff7 --- /dev/null +++ b/2024/VoCo_ A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis/layout.json @@ -0,0 +1,9677 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "type": "text", + "content": "VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 151, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 151, + 175 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 151, + 175 + ], + "type": "text", + "content": "Linshan Wu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 264, + 162, + 336, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 162, + 336, + 175 + ], + "spans": [ + { + "bbox": [ + 264, + 162, + 336, + 175 + ], + "type": "text", + "content": "Jiaxin Zhuang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 450, + 162, + 507, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 450, + 162, + 507, + 174 + ], + "spans": [ + { + "bbox": [ + 450, + 162, + 507, + 174 + ], + "type": "text", + "content": "Hao Chen *" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 178, + 176, + 422, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 176, + 422, + 190 + ], + "spans": [ + { + "bbox": [ + 178, + 176, + 422, + 190 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "content": "Self-Supervised Learning (SSL) has demonstrated promising results in 3D medical image analysis. However, the lack of high-level semantics in pre-training still heavily hinders the performance of downstream tasks. We observe that 3D medical images contain relatively consistent contextual position information, i.e., consistent geometric relations between different organs, which leads to a potential way for us to learn consistent semantic representations in pre-training. In this paper, we propose a simple yet-effective Volume Contrast (VoCo) framework to leverage the contextual position priors for pre-training. Specifically, we first generate a group of base crops from different regions while enforcing feature discrepancy among them, where we employ them as class assignments of different regions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which region) by contrasting their similarity to different base crops, which can be seen as predicting contextual positions of different sub-volumes. Through this pretext task, VoCo implicitly encodes the contextual position priors into model representations without the guidance of annotations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experimental results on six downstream tasks demonstrate the superior effectiveness of VoCo. Code will be available at https://github.com/Luffy03/VoCo." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 590, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "content": "Deep learning has demonstrated outstanding achievements in 3D medical image analysis [53, 21, 40, 33, 39], yet is heavily hampered by the expensive cost of the required expert annotations [50, 23]. To address this problem, Self-Supervised Learning (SSL) has received significant attention due to its promising ability to learn representations without annotations [10, 11, 6, 28, 20], which has become" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 229, + 542, + 287 + ], + "blocks": [ + { + "bbox": [ + 311, + 229, + 542, + 287 + ], + "lines": [ + { + "bbox": [ + 311, + 229, + 542, + 287 + ], + "spans": [ + { + "bbox": [ + 311, + 229, + 542, + 287 + ], + "type": "image", + "image_path": "864fec36362eafd2c1ad151515fa3a58137346649cea1e8cb91415b5af2a8e16.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 287, + 534, + 295 + ], + "lines": [ + { + "bbox": [ + 328, + 287, + 534, + 295 + ], + "spans": [ + { + "bbox": [ + 328, + 287, + 534, + 295 + ], + "type": "text", + "content": "(a) Relatively consistent contextual positions in 3D medical images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 297, + 543, + 378 + ], + "blocks": [ + { + "bbox": [ + 316, + 297, + 543, + 378 + ], + "lines": [ + { + "bbox": [ + 316, + 297, + 543, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 543, + 378 + ], + "type": "image", + "image_path": "abde1250c2e2349e645a2e82439a7d4c9b3b032678330d1dd9fdab3f990dc6d3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 380, + 526, + 389 + ], + "lines": [ + { + "bbox": [ + 347, + 380, + 526, + 389 + ], + "spans": [ + { + "bbox": [ + 347, + 380, + 526, + 389 + ], + "type": "text", + "content": "(b) Volume Contrast (VoCo) for 3D Medical Image Analysis" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 392, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 547, + 448 + ], + "type": "text", + "content": "Figure 1. (a) In 3D medical images, the contextual positions, i.e., the geometric relations between different organs are relatively consistent. (b) Aiming to leverage contextual position priors for pretraining, we proposed a Volume Contrast (VoCo) framework for 3D Medical Image Analysis." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 472, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 496 + ], + "type": "text", + "content": "an important label-efficient solution in 3D medical image analysis [71, 51, 32, 2, 34, 36]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": "Existing methods [50, 75, 71, 13] are mostly based on information reconstructions to learn augment-invariant representations of 3D medical images, which first employ strong data augmentation to the images and then reconstruct the raw information. Specifically, rotate-and-reconstruct [50, 51, 75, 52] proposed to randomly rotate the 3D volumetric images and learn to recover them, which encourages models to learn rotational invariant features. Recent methods [70, 71, 32, 25, 62] further proposed to restore information among different views of the image. PCRL [70, 71] cropped global and local patches then conducted multiscale restorations. GVSL [32] further explored the geometric similarity between multi-scans by affine augmentation and matching. Mask-reconstruct methods [13, 73, 55] are also widely used, which are introduced from MAE [28] and aim to learn representations by masking images and reconstructing the missing pixels. Although promising results have been demonstrated, previous works [52, 32] have" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 188, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 188, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 188, + 712 + ], + "type": "text", + "content": "*Corresponding author: jhc@cse.ust.hk" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22873" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 72, + 221, + 160 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 221, + 160 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 221, + 160 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 221, + 160 + ], + "type": "image", + "image_path": "1fbbea86f9466374034fabf6a28da9cf19a0135a7f820aa2b9593741e46bafc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 170, + 194, + 180 + ], + "lines": [ + { + "bbox": [ + 127, + 170, + 194, + 180 + ], + "spans": [ + { + "bbox": [ + 127, + 170, + 194, + 180 + ], + "type": "text", + "content": "(a) Instance-level" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 369, + 170 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 369, + 170 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 369, + 170 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 369, + 170 + ], + "type": "image", + "image_path": "79a7a20cd8fab64dad775dc75401d3df656deea6b5e6faf3ce22de6ab4626638.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 171, + 332, + 180 + ], + "lines": [ + { + "bbox": [ + 260, + 171, + 332, + 180 + ], + "spans": [ + { + "bbox": [ + 260, + 171, + 332, + 180 + ], + "type": "text", + "content": "(b) Prototype-level" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 248 + ], + "type": "text", + "content": "Figure 2. Typical contrastive learning frameworks. (a) Instance-level contrastive learning [10, 11, 29, 22, 7] employs strong data augmentation or model perturbation on input data to acquire different views of instance, then regularizes their consistency. (b) Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] conducts (1) online clustering or (2) randomly initialize then online update process to obtain prototypes as class assignments, then leverage the prototypes to contrast each input image. (c) Our VoCo follows the idea of prototype-level contrastive learning. Specifically, instead of using time-consuming online clustering and updating procedures, we leverage the valuable contextual position priors of 3D medical images and leverage the base crops to generate prototypes (bases)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 498, + 170 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 498, + 170 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 498, + 170 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 498, + 170 + ], + "type": "image", + "image_path": "80fbd5499995cb2ac570b093809e1502ef0fb9b200e7b699afbe748d23a3280e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 171, + 458, + 180 + ], + "lines": [ + { + "bbox": [ + 408, + 171, + 458, + 180 + ], + "spans": [ + { + "bbox": [ + 408, + 171, + 458, + 180 + ], + "type": "text", + "content": "(c) Our VoCo" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 268, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 327 + ], + "type": "text", + "content": "proved that the lack of high-level semantics in pre-training will heavily hinder the performance of downstream tasks. To address this challenge, we argue that stronger high-level semantics should be further involved into 3D medical image pre-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 329, + 288, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 496 + ], + "type": "text", + "content": "To this end, we argue that the contextual position priors of 3D medical images should be further exploited. As shown in Fig. 1(a), we observe that in 3D medical images, different organs (semantic regions) contain relatively consistent contextual positions with relatively consistent anatomic characteristics (shapes). Thus, the consistency of geometric relations between different organs leads to a potential way for us to learn consistent semantic representations for 3D medical images pre-training. In this paper, we propose a pretext task for contextual position predictions, which aims to encode contextual position priors into model representations and enables us to effectively improve the performance of downstream tasks that require high-level semantics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "In this paper, we propose a simple-yet-effective Volume Contrast (VoCo) framework for 3D medical image analysis, as shown in Fig. 1(b). Specifically, we first crop a group of non-overlap volumes from different positions while enforcing feature discrepancy among them. We represent these volumes as a group of bases in the learned high-dimension space, where we employ them as class assignments of different positions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which position) by contrasting their similarity to different bases, which can be seen as predicting contextual positions of different sub-volumes. In this way, we formulate a contextual position prediction pretext task for 3D medical image SSL. Through learning to predict contextual positions, we implicitly involve the high-level semantic priors into the model representations, which enables us to significantly improve the performance of downstream tasks. Extensive experimental results on six downstream tasks demonstrate that our" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 268, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 545, + 292 + ], + "type": "text", + "content": "proposed VoCo clearly outperforms existing state-of-the-art 3D medical image SSL methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 306, + 397, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 397, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 397, + 319 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 327, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 399 + ], + "type": "text", + "content": "In this section, we first introduce the previous mainstream contrastive learning paradigms. Then, we survey the existing SSL methods for medical image analysis, especially for 3D medical images. Finally, we review the position-related SSL methods for comparisons with our method and highlight the differences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "content": "Contrastive learning. Contrastive learning is one of the mainstream paradigms in SSL, which aims to learn consistent representations by contrasting positive and negative pairs of samples without extra annotations [10, 6, 29, 11]. According to [6], instance- and prototype-level contrastive learning are two typical types of contrastive learning, as shown in Fig. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 485, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 676 + ], + "type": "text", + "content": "Instance-level contrastive learning [10, 11, 29, 22, 7] transforms input images with different augmentations or model perturbations, aiming to compare the features from each other. Prototype-level contrastive learning [5, 6, 56, 45, 15, 16] proposes to generate prototypes (also called clusters or bases) for contrasting each input image. Specifically, there are two typical ways to generate prototypes. First, Caron et al. proposed DeepCluster [5] to conduct online clustering on the whole dataset to generate prototypes. However, it is very time-consuming to calculate clusters on a large dataset. Thus, some recent works [6, 56, 15, 16] propose to randomly initialize a group of prototypes and then update them through back-propagation during training, which has demonstrated promising results. However, there is still no explicit guarantee that these randomly initialized prototypes can be updated well during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Our VoCo follows the primary idea of prototype-level constrasive learning. As shown in Fig. 2(c), to address the existing problems mentioned above, instead of randomly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22874" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "initializing and updating prototypes, VoCo leverages the valuable contextual position priors of 3D medical images to generate base crops as prototypes, which also requires no time-consuming clustering on a large dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 128, + 287, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 128, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 128, + 287, + 272 + ], + "type": "text", + "content": "SSL for medical image analysis. Due to the high potential in label-efficient learning [29, 59, 61, 58, 57, 60, 37], SSL has also received significant attention in the field of medical image analysis [70, 32, 31, 51, 19]. Existing methods are mainly based on comparative SSL [71]. Specifically, Zhou et al. [69] combined Mixup [66] into MoCo [29] to learn the diversity of positive and negative samples in InfoNCE [44]. Azizi et al. used multi-instance learning to compare multiple views of images from each patient. There are also a number of approaches [25, 70, 71] that supervising the models via restoring low-level information from raw images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 422 + ], + "type": "text", + "content": "In 3D medical image analysis, reconstructing raw information is a popular pretext task for learning representations [50, 51, 71]. Existing methods are mainly based on reconstructing information from augmented images. These previous methods first conducted strong data augmentation, e.g., rotate [51, 75, 52], multi-view crops [70, 71, 32], and mask [13, 73, 55], then supervised the model by reconstructing raw 3D information. Although promising results have been demonstrated, most of these methods still largely ignore the importance of integrating high-level semantics into model representations, which heavily hinders the performance of downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": "Position-related SSL. Position-related SSL methods are also explored in a number of previous works [8, 9, 41, 47, 43, 17, 64, 68] in the field of natural images. Noroozi et al. [43] proposed to predict the order of a set of shuffled patches. Zhai et al. [64] and Caron et al. [8] proposed to train a ViT [18] to predict the locations of each input patch. However, since the geometric relations of different objects are not very consistent in natural images, it is still difficult to effectively learn consistent position representations given visual appearance only (as stated in [68]). In addition, previous works [64, 8, 68] mainly trained a linear layer to output the positions directly, which works in a black-box manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "In this paper, we introduce the pretext task of contextual position prediction into the field of 3D medical images, where the geometric relations between different organs are relatively consistent, which guides us to learn consistent semantic representations in pre-training. Different from the previous methods, in this paper, we introduce a totally different position prediction paradigm. Specifically, instead of using a linear layer to output positions directly, we predict the contextual positions based on volume contrast, which is more intuitive and effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": "In this section, we first introduce the overall framework of our proposed VoCo in Section 3.1. After that, we present the process of contextual position prediction in Section 3.2. Then, the regularization process via volume contrast in our proposed VoCo framework is described in Section 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 160, + 422, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 160, + 422, + 172 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 422, + 172 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "type": "text", + "content": "The overall framework of our proposed VoCo is presented in Fig. 3, which contains a contextual position prediction branch and a regularization branch. The prediction branch is used to predict the contextual positions between different cropped volumes. Specifically, given an input volume, we first crop it into non-overlap base volumes, which cover the whole input volume. Then, we randomly crop a volume and transform it into the high-dimension feature space using a typical backbone (CNN [30] or Transformer [18]). The goal is to predict the contextual positions between the randomly cropped volumes and base volumes. In this paper, instead of training a linear classifier to predict positions as in previous works [64, 9, 8, 68], we propose to establish this goal by volume contrast. We develop a loss function " + }, + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "type": "text", + "content": " to supervise the final predictions. In addition, we further use a loss function " + }, + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 304, + 179, + 546, + 406 + ], + "type": "text", + "content": " to regularize the feature discrepancy from different bases by enlarging their distance, aiming to learn more discriminative class assignments. The details are presented in Section 3.2 and 3.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 415, + 471, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 471, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 471, + 427 + ], + "type": "text", + "content": "3.2. Contextual Position Prediction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": "Base and random crops. Given an input volume, we first crop it into " + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": " non-overlap base volumes, which cover the whole input volume. We then employ the extracted features " + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": " as class assignments (we call them bases), which present the prototype-level features from different positions. Then, following previous SSL works [10, 11, 29], a projector with linear layers is used to project " + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": " into latent features " + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": ". Then, we randomly crop a volume and transform it into high-dimension feature space as " + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": ". The backbone and projector are also used to project the features from the randomly cropped volumes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "text", + "content": "Volume contrast for contextual position prediction. With features extracted from the backbone and projector, following previous SSL works [10, 11, 29], we first conduct 3D adaptive average pooling to resize them to one dimension, i.e., " + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "text", + "content": " is the number of channels. Specifically, we empirically set " + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 566, + 545, + 649 + ], + "type": "text", + "content": " to 2048 as in [10, 11, 29]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "text", + "content": "Then, we calculate the similarity logits " + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "text", + "content": ". Specifically, we use cosine similarity to compute " + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 650, + 545, + 685 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 693, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 693, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 343, + 693, + 545, + 717 + ], + "type": "interline_equation", + "content": "l _ {i} = \\operatorname {C o s S i m} (p, q _ {i}) = \\frac {p \\cdot q _ {i}}{\\| p \\| \\| q _ {i} \\|}, i \\in n \\tag {1}", + "image_path": "47daade1be827036ce1b0ae14b78cd1ede9f79be44413b1d539728da580a6324.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22875" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 68, + 491, + 281 + ], + "blocks": [ + { + "bbox": [ + 97, + 68, + 491, + 281 + ], + "lines": [ + { + "bbox": [ + 97, + 68, + 491, + 281 + ], + "spans": [ + { + "bbox": [ + 97, + 68, + 491, + 281 + ], + "type": "image", + "image_path": "7f4b4b75e18847a633dde5f16f7d9d40aab1d9327b86900306c6e446583248b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 283, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 547, + 319 + ], + "type": "text", + "content": "Figure 3. The overall framework of VoCo. VoCo contains a prediction branch and a regularization branch. The prediction branch is responsible for predicting contextual positions between different sub-volumes. The regularization branch is employed to enforce the feature discrepancy between different bases, which aims to learn more discriminative class assignments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "content": " is the projected feature of each base crop. " + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "content": " denotes the similarity between " + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "content": ", which ranges from 0 to 1. It is worth noting that, we stop the gradients of " + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 336, + 287, + 396 + ], + "type": "text", + "content": " when computing Eq. 1, which aims to avoid feature collapse [10, 11, 6]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "Intuitively, higher " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": " represents that " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": " has higher probabilities to share overlap regions with " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": ". In this way, we can explicitly associate the similarity value with the position information, i.e., " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": " with higher " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": " is more likely to be located in the region of the " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": " base. Thus, instead of training a black-box linear layer, we predict the contextual positions by volume contrast, which is more intuitive and effective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": "Position labels generation. The process of generating position labels is shown in Fig. 4. As shown in Fig. 4, when we generate " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "n = 4 \\times 4" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": " base crops, there will be " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": " class assignments. Then we calculate the overlap area between a randomly cropped volume and " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": " base crops. The proportions of the overlap area are then assigned as position labels " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": ", which also range from 0 to 1. Thus, we can easily supervise the model by calculating the distance between the prediction logits " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": " and position labels " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": ". The setting of the number " + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 605 + ], + "type": "text", + "content": " of base crops will be discussed in Section 4.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": "Loss function for contextual position prediction. The formulation of prediction loss function " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": " is based on entropy. Specifically, we first calculate the distance " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": " between prediction logits " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": " and position labels " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 673, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 673, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 287, + 685 + ], + "type": "interline_equation", + "content": "d _ {i} = \\left| y _ {i} - l _ {i} \\right|, i \\in n, \\tag {2}", + "image_path": "56b9d430d128ed12bbc991eb5aee81f6a54af49e20222d543b95ca8f45939f54.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "|.|" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " denotes the absolute value. Then, " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is formu" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 334, + 518, + 458 + ], + "blocks": [ + { + "bbox": [ + 331, + 334, + 518, + 458 + ], + "lines": [ + { + "bbox": [ + 331, + 334, + 518, + 458 + ], + "spans": [ + { + "bbox": [ + 331, + 334, + 518, + 458 + ], + "type": "image", + "image_path": "52712bc928b1b014ead126e6371ce21df02539f86b5fabfa38b6e3e7eaacc0b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 460, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 506 + ], + "type": "text", + "content": "Figure 4. The process of generating position labels. We calculate the proportions of overlap area as position labels " + }, + { + "bbox": [ + 305, + 460, + 545, + 506 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 460, + 545, + 506 + ], + "type": "text", + "content": ", e.g., the randomly cropped volume in the figure is assigned to class 5, 6, 9, and 10 with probabilities of 0.25, 0.1, 0.5, and 0.15, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 523, + 373, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 373, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 373, + 534 + ], + "type": "text", + "content": "lated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 541, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 541, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 364, + 541, + 545, + 572 + ], + "type": "interline_equation", + "content": "L _ {p r e d} = - \\frac {1}{n} \\sum_ {i \\in n} ^ {n} \\log \\left(1 - d _ {i}\\right). \\tag {3}", + "image_path": "4ec0ae6727e300a284b0acfb9cec823116a7040b1525e9df83e204085d0425e1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "text", + "content": "It is worth noting that VoCo predicts contextual positions of a volume (high similarities with all its contextual overlapped volumes), thus don't need one-to-one correspondence: e.g., in Fig. 4, high-value " + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "text", + "content": " pertain to " + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "inline_equation", + "content": "y_{i} > 0, i = 5,6,9,10" + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "text", + "content": " simultaneously. Then we calculate the distance between " + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 580, + 545, + 652 + ], + "type": "text", + "content": " (Eq. 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 659, + 496, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 496, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 496, + 672 + ], + "type": "text", + "content": "3.3. Volume Contrast for Regularization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We aim to learn more discriminative class assignments (bases) for volume contrast. Since intuitively, different subvolumes tend to contain different organs (semantic discrepancy)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22876" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "content": "ancy). Thus, we aim to enlarge the high-dimension feature discrepancy between different bases. To this end, we design a regularization loss " + }, + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "content": " to enlarge the feature discrepancy between different bases " + }, + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "text", + "content": "First, given projected bases " + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "text", + "content": ", we also calculate the cosine similarity " + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "text", + "content": " between different " + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "inline_equation", + "content": "q_j" + }, + { + "bbox": [ + 47, + 121, + 286, + 144 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 286, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 286, + 177 + ], + "type": "interline_equation", + "content": "s _ {i j} = \\operatorname {C o s S i m} \\left(q _ {i}, q _ {j}\\right) = \\frac {q _ {i} \\cdot q _ {j}}{\\| q _ {i} \\| \\| q _ {j} \\|}, i, j \\in n, i \\neq j, \\tag {4}", + "image_path": "6ff61a6aea5211497059ea4d0f7f78d98ca7a8550395d40f42a99116bb76e151.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "type": "text", + "content": "where we aim to regularize " + }, + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "type": "text", + "content": " to 0, enforcing feature discrepancy between different bases. Thus, the loss function " + }, + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 47, + 184, + 286, + 220 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 229, + 286, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 229, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 100, + 229, + 286, + 262 + ], + "type": "interline_equation", + "content": "L _ {r e g} = \\frac {2}{n (n - 1)} \\sum_ {i, j \\in n, i \\neq j} ^ {n} | s _ {i j} |, \\tag {5}", + "image_path": "c9572fb928206e7f2120ee9de210b55d2aecb389f03d66ce110e8806d18c7c1c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "inline_equation", + "content": "|.|" + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": " denotes the absolute value. With loss " + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": ", we aim to optimize " + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": " as linearly independent bases:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 304, + 286, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 304, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 127, + 304, + 286, + 318 + ], + "type": "interline_equation", + "content": "q _ {i} \\bot q _ {j}, i \\neq j, i, j \\in n. \\tag {6}", + "image_path": "82f844c6c9e453e196685546e0ac3a3b32ba430f7e0684361914bdaecd61151f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 326, + 286, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 286, + 386 + ], + "type": "text", + "content": "With the regularization loss function " + }, + { + "bbox": [ + 47, + 326, + 286, + 386 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 47, + 326, + 286, + 386 + ], + "type": "text", + "content": ", we aim to learn a group of linearly independent bases to represent all directions of high-dimension features [6]. In this way, we can learn a group of more discriminative class assignments to supervise the final position predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "text", + "content": "Overall loss function. Thus, the total loss function " + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "text", + "content": " is the combination of " + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 47, + 386, + 286, + 411 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 420, + 286, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 420, + 286, + 433 + ], + "spans": [ + { + "bbox": [ + 124, + 420, + 286, + 433 + ], + "type": "interline_equation", + "content": "L = L _ {p r e d} + \\lambda L _ {r e g}, \\tag {7}", + "image_path": "ccbd438f129872560ef7ca66f09f4dafad6e71ab218b8833e1a4bbd7ce7fe8f3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "type": "text", + "content": " is used to balance the relative contributions of these two loss terms and set to 1.0 in experiments empirically since we consider their importance equally. The ablation studies of " + }, + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 441, + 286, + 489 + ], + "type": "text", + "content": " are provided in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 500, + 127, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 127, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 127, + 513 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 520, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 286, + 592 + ], + "type": "text", + "content": "In this section, we first describe the datasets used in the pre-training and downstream tasks. Then, we briefly introduce the implementation details of VoCo. Finally, we report detailed experiment results of our proposed VoCo compared with other state-of-the-art SSL methods in 3D medical images. More details are in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 599, + 109, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 109, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 109, + 610 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "content": "Pre-training datasets. Aiming to conduct fair comparisons with the previous works [51, 55, 70, 71, 13, 73], we also carry out pre-training experiments on the same three public datasets, i.e., BTCV [35], TCIA Covid19 [14], and LUNA [48] datasets, including about totally " + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{k}" + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "content": " CT scans for pre-training. It is worth noting that, aiming to conduct fair comparisons with previous works [73, 13], we only use BTCV [35] and TCIA Covid-19 [14] for pre-training in" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "the downstream experiments of BTCV [35]. For the other downstream tasks, we use all three datasets for pre-training. Details are provided in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 109, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 545, + 276 + ], + "type": "text", + "content": "Downstream datasets. To evaluate the effectiveness of our VoCo, we conduct downstream experiments on six public datasets, i.e., BTCV [35], LiTs [4], MSD Spleen [1], MM-WHS [74], BraTS 21 [49], and CC-CCII [67], including segmentation and classification tasks. The first five datasets are developed for segmentation, while CC-CCII [67] is for COVID-19 classification. Note that only BTCV [35] is used in pre-training, the other datasets are unseen in pre-training. In addition, to evaluate the cross-modality generalization ability, we transfer the model pre-trained on the CT dataset to the MRI dataset BraTS 21 [49]. We adopt consistent settings as previous works [13, 73, 26, 55, 32]. We also evaluate the performance on 2D medical dataset [54]. Details are provided in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 285, + 436, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 436, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 436, + 296 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 303, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 410 + ], + "type": "text", + "content": "Following previous works [51, 55], we use SwinUNETR [26] for both pre-training and downstream tasks. We use AdamW [38] optimizer and cosine learning rate scheduler for all experiments. We set 100K training steps in the pre-training process and applied a slicing window inference for fair comparisons with previous works [51, 55, 13, 73]. Aiming to evaluate the pure effectiveness, we do not use foundation models or post-processing [34, 36]. Details are provided in the supplementary materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 411, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 578 + ], + "type": "text", + "content": "Comparison methods. We compare our VoCo with both General and Medical SSL methods. First, we compare with the typical SSL methods MAE [28, 13] and MoCo v3 [29, 12], since they represent the two mainstream SSL paradigms, i.e., mask-autoencoder and constrastive learning. Since it is not practical to set a large batch size for 3D medical images due to computation cost, for fair comparisons, we adopt consistent settings with other methods in MAE [28, 13] and MoCo v3 [29, 12]. We also report the results of SimCLR [10] and SimMIM [63] according to [13]. We further evaluate the performance of Jiasaw [9] and PositionLabel [68], since they are related to our position-aware methods. Most existing state-of-the-art medical SSL methods are compared in our experiments." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 587, + 488, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 488, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 488, + 598 + ], + "type": "text", + "content": "4.3. Experiments on downstream tasks" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 605, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 689 + ], + "type": "text", + "content": "Outperform existing methods on the BTCV dataset. We first conduct experiments on BTCV [35], as shown in Table 1. Specifically, among the comparison methods, MAE3D [28, 13], SimCLR [10], SimMIM [63], MoCo v3 [29, 12], and GL-MAE [73] use UNETR [27]. Other methods including our VoCo adopt Swin-UNETR [26] as the default settings of the previous work [51]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Remark. It can be seen in Table 1 that the general SSL methods perform worse than most medical SSL methods." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22877" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 555, + 305 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 555, + 305 + ], + "type": "table", + "html": "
MethodDice Score(%)AVG
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAG
From Scratch
UNETR [27]93.0294.1394.1266.9970.8796.1177.2789.2282.1070.1676.6565.3259.2179.82
Swin-UNETR† [26]94.0693.5493.8065.5174.6097.0975.9491.8082.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [28, 13]93.9894.3794.1869.8674.6596.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [10]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [63]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3† [29, 12]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw† [9]94.6293.4193.5575.6373.2195.7180.8089.4184.7871.0279.5765.6860.2281.35
PositionLabel† [68]94.3593.1593.2175.3973.2495.7680.6988.8084.0471.1879.0265.1160.0781.09
With Medical SSL
MG [72]91.9993.5291.8165.1176.1495.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [50]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [3]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++† [52]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1† [70]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2† [71]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [26, 51]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [55]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [73]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL† [32]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
", + "image_path": "5347f340f77a652b22584b00373f13bf4cb9663e0ca8572f8d0917e7e2123d27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 346, + 271, + 581 + ], + "blocks": [ + { + "bbox": [ + 47, + 306, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 47, + 306, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 545, + 331 + ], + "type": "text", + "content": "Table 1. Experimental results on BTCV [35]. The best results are bolded. 'From Scratch' denotes the supervised baseline without self-supervised pre-training. † denotes we re-implement the approach. Most results are drawn from [13, 65, 73] or their own papers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 346, + 271, + 581 + ], + "lines": [ + { + "bbox": [ + 65, + 346, + 271, + 581 + ], + "spans": [ + { + "bbox": [ + 65, + 346, + 271, + 581 + ], + "type": "table", + "html": "
MethodNetworkDice Score(%)
From Scratch
3D UNet [46]-90.70
UNETR† [27]-93.25
Swin-UNETR† [26]-93.42
With General SSL
Jigsaw [9]3D UNet94.36
MAE3D [28, 13]UNETR94.02
MoCo v3† [29, 12]UNETR93.86
Jigsaw† [9]Swin-UNETR95.24
PositionLabel† [68]Swin-UNETR94.13
With Medical SSL
MG [72]3D UNet91.30
TransVW [24]3D UNet91.42
ROT [50]3D UNet94.49
PCRLv1 [70]3D UNet93.87
PCRLv2 [71]3D UNet94.50
Rubik [75]3D UNet94.93
Rubik++ [52]3D UNet95.46
Rubik++† [52]Swin-UNETR95.72
Swin-UNETR [51, 26]Swin-UNETR95.33
SwinMM [55]Swin-UNETR95.52
VoCo3D UNet96.03
VoCoSwin-UNETR96.52
", + "image_path": "3bd736bc744be5c5e196c599bb21a3fd6296895e5a45fb5b283a699af478e2d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Specifically, MoCo v3 [29, 12] can only achieve " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "79.54\\%" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " Dice Score. Since MoCo v3 [29, 12] heavily relies on a large batch size to acquire adequate negative samples, which is not practical in 3D medical images due to the huge computation burden. In addition, the negative relation between different images used in MoCo [29, 12] is not ap" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 331, + 346, + 522, + 563 + ], + "blocks": [ + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "type": "text", + "content": "Table 2. Experimental results on LiTs [4]. We report the Dice Scores of liver segmentation. " + }, + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "type": "text", + "content": " denotes we re-implement the approach. Most results are drawn from [65, 71]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 346, + 522, + 563 + ], + "lines": [ + { + "bbox": [ + 331, + 346, + 522, + 563 + ], + "spans": [ + { + "bbox": [ + 331, + 346, + 522, + 563 + ], + "type": "table", + "html": "
MethodMSD SpleenMM-WHS
From Scratch
3D UNet [46]93.7183.09
UNETR† [27]94.2085.85
Swin-UNETR† [26]94.6386.11
With General SSL
MAE3D [28, 13]95.2086.03
MoCo v3† [29, 12]94.3284.16
Jiasaw† [9]94.2985.88
PositionLabel† [68]94.1685.52
With Medical SSL
MG [72]94.4086.36
VicRegl [3]94.1284.72
UniMiss [62]95.0984.68
PCRLv1† [70]94.3286.58
PCRLv2† [71]94.9486.82
Rubik++† [52]95.1187.02
Swin-UNETR [51, 26]95.0287.06
SwinMM [55]95.3486.98
JSSL [42]94.9284.89
GVSL [32]95.4788.27
VoCo96.3490.54
", + "image_path": "b2853e1a7ade31cc86b7df1644bcd50855c616ccc8b5dd1dc3c75009ad7ab4a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "content": "Table 3. Experimental results on MSD Spleen [1] and MM-WHS [74]. We report the Dice Scores of segmentation prediction. " + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "content": " denotes we re-implement the approach." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": "propriate in medical images. MAE [28, 13], SimCLR [10], and SimMIM [63] (results from [13]) also gain limited performance. Our VoCo also outperforms the position-based methods Jigsaw [9] and PositionLabel [68] by a clear margin. Thus, we conclude that general SSL methods are not very suitable for 3D medical images. It is crucial to consider the characteristics of medical images in medical SSL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 700, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 700, + 545, + 714 + ], + "type": "text", + "content": "The scratch Swin-UNETR [26] only achieves " + }, + { + "bbox": [ + 318, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "80.53\\%" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22878" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 298, + 258 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 298, + 258 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 298, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 298, + 258 + ], + "type": "table", + "html": "
MethodNet.Dice Score(%)
TCWTETAVG
From Scratch
UNETR [27]-81.6287.8157.3475.58
Swin-UNETR [26]-81.2888.6757.7375.89
With General SSL
MAE3D [28, 13]UNETR82.3490.3559.1877.29
SimMIM [63]UNETR84.0690.4359.0777.85
SimCLR [10]UNETR83.1389.4458.4276.99
MoCo v3† [29, 12]UNETR82.6088.8957.6976.39
Jigsaw† [9]Sw-UNE.81.6289.4559.1076.72
PositionLabel† [68]Sw-UNE.81.3589.6258.7376.64
With Medical SSL
PCRLv1†[70]Sw-UNE.81.9688.8357.5876.12
PCRLv2†[71]Sw-UNE.82.1390.0657.7076.63
Rubik++†[52]Sw-UNE.84.3290.2358.0177.51
Swin-UNETR [51, 26]Sw-UNE.82.5189.0858.1576.58
SwinMM [55]Sw-UNE.83.4890.4758.7277.56
VoCoSw-UNE.85.2790.4559.8778.53
", + "image_path": "68063f4546c105b98685ac041e69423c4674723c7825de98fb7832c240d35d2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 304, + 268, + 464 + ], + "blocks": [ + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "type": "text", + "content": "Table 4. Experimental results on BraTS 21 [49]. WT, TC, and ET denote the whole tumor, tumor core, and enhancing tumor, respectively. " + }, + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "type": "text", + "content": " denotes we re-implement the approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 304, + 268, + 464 + ], + "lines": [ + { + "bbox": [ + 66, + 304, + 268, + 464 + ], + "spans": [ + { + "bbox": [ + 66, + 304, + 268, + 464 + ], + "type": "table", + "html": "
MethodNetworkAccuracy(%)
From Scratch
UNETR [27]-88.92
Swin-UNETR [26]-88.04
With General SSL
MAE3D [28, 13]UNETR89.47
MoCo v3 [29, 12]UNETR84.95
Jiasaw [9]Swin-UNETR86.88
PositionLabel [68]Swin-UNETR87.54
With Medical SSL
PCRLv1 [70]Swin-UNETR88.72
PCRLv2 [71]Swin-UNETR89.15
Rubik++ [52]Swin-UNETR89.23
Swin-UNETR [51, 26]Swin-UNETR89.45
SwinMM [55]Swin-UNETR89.61
VoCoSwin-UNETR90.83
", + "image_path": "9b4087d451304d31dd049f0c307a5747d18b5362e2c75a5214c2b42c5428a7c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": "Dice Score. With VoCo pre-training, we gain " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "3.32\\%" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " improvements with " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "83.85\\%" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " Dice Score, which also outperforms existing methods by a clear margin. Among the compared methods, GL-MAE [73] achieves the highest Dice Score " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "(82.01\\%)" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": ". Our VoCo surpasses it by " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "1.84\\%" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " Dice Score, which is a clear improvement in this dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Promising performance on Unseen datasets. We further conduct experiments on unseen datasets in pre-training, i.e., LiTs [4], MSD Spleen [1], and MM-WHS [74]. The results on LiTs [4] are shown in Table 2. We report the results of compared methods according to [70, 71, 65]. Since the scratch Swin-UNETR [26] can obtain a higher Dice Score (93.42%), we further pre-train a 3D UNet [46] based on VoCo, aiming to conduct fair comparisons. It can be seen that with VoCo pre-training, Swin-UNETR [26] gains " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "3.10\\%" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " improvements and achieves " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "96.52\\%" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " Dice Score. With 3D UNet [46] as the backbone, VoCo also achieves " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "96.03\\%" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " Dice Score, proving the effectiveness of VoCo with" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 543, + 203 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 543, + 203 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 543, + 203 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 543, + 203 + ], + "type": "image", + "image_path": "8a0f0ab2186be6e026c2b52d348737a6b476da5fb8f1e306b15f06f5ba9cdb5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 206, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 306, + 206, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 545, + 227 + ], + "type": "text", + "content": "Figure 5. Overall comparisons with state-of-the-art methods on six different datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 349, + 237, + 503, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 464, + 268, + 475 + ], + "lines": [ + { + "bbox": [ + 47, + 464, + 268, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 268, + 475 + ], + "type": "text", + "content": "Table 5. Experimental results of CC-CCII [67] classification." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 237, + 503, + 297 + ], + "lines": [ + { + "bbox": [ + 349, + 237, + 503, + 297 + ], + "spans": [ + { + "bbox": [ + 349, + 237, + 503, + 297 + ], + "type": "table", + "html": "
Loss FunctionsBTCVMM-WHS
LpredLreg
XX80.5386.11
✓X82.9688.82
✓✓83.8590.54
", + "image_path": "67f2ffb27484c8407e0b37605da899ffd57e9b6c525d584124ebdb6a6d25fb40.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 345, + 331, + 507, + 415 + ], + "blocks": [ + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "text", + "content": "Table 6. Evaluation of loss functions " + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "text", + "content": ". We report the average Dice Score on BTCV [35] and MM-WHS [74]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 331, + 507, + 415 + ], + "lines": [ + { + "bbox": [ + 345, + 331, + 507, + 415 + ], + "spans": [ + { + "bbox": [ + 345, + 331, + 507, + 415 + ], + "type": "table", + "html": "
Number of bases nBTCVMM-WHS
(x,y,z)
2×2×181.5686.73
3×3×182.7789.31
4×4×183.8590.54
5×5×183.6090.49
3×3×282.8689.14
4×4×283.4790.52
", + "image_path": "2f17d4bf7ed6e2b383241780984495d4b357b654980653c3c0de10e9668b84c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 416, + 545, + 438 + ], + "lines": [ + { + "bbox": [ + 306, + 416, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 545, + 438 + ], + "type": "text", + "content": "Table 7. Evaluation of the value of bases " + }, + { + "bbox": [ + 306, + 416, + 545, + 438 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 416, + 545, + 438 + ], + "type": "text", + "content": " . We report the average Dice Score on BTCV [35] and MM-WHS [74]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 460, + 433, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 433, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 433, + 470 + ], + "type": "text", + "content": "different network architectures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": "The results on MSD Spleen [1] and MM-WHS [74] datasets are shown in Table 3. In previous methods, GVSL [32] achieves the best performance with " + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "95.47\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "88.27\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": " Dice Score, while our VoCo surpasses all previous methods with " + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "96.34\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "90.54\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": " Dice Score on MSD Spleen [1] and MM-WHS [74] datasets, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": "Generalization capacity on MRI dataset. To verify the generalization capacity on the MRI dataset, we further evaluate the performance of VoCo on BraTS 21 [49]. As shown in Table 4, VoCo achieves " + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "inline_equation", + "content": "78.53\\%" + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": " Dice Score and outperforms existing state-of-the-art methods, demonstrating the cross-model generalization capacity of VoCo." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": "Evaluation of COVID-19 classification. We further evaluate the performance of the classification task on the CC-CCII [67] dataset in Table 5. Since existing SSL methods did not conduct experiments on this dataset, we reproduce the related methods for comparisons. It can be seen that VoCo can also achieve superior results with " + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "inline_equation", + "content": "90.83\\%" + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": " accuracy, proving its effectiveness in the classification task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Overall comparisons on six downstream datasets. The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22879" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 70, + 459, + 213 + ], + "blocks": [ + { + "bbox": [ + 135, + 70, + 459, + 213 + ], + "lines": [ + { + "bbox": [ + 135, + 70, + 459, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 70, + 459, + 213 + ], + "type": "image", + "image_path": "4ab02bfc61edd815b2b5f76aeffa029525242aa7f5b890fa09ffa6cf03bb3607.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": "Figure 6. Qualitative visualization of segmentation results for the BTCV [35] dataset. We compare VoCo with MoCo v3 [12], PositionLabel [68], SwinUNETR [51, 26], and GVSL [32]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 307 + ], + "type": "text", + "content": "overall comparisons are shown in Fig. 5. Specifically, we compare with the existing state-of-the-art methods on six different downstream datasets. It can be seen that our VoCo outperforms them by a clear margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 141, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 141, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 141, + 325 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 366 + ], + "type": "text", + "content": "We further conduct ablation studies to evaluate the loss functions and the settings in VoCo, which are verified on the BTCV [35] and MM-WHS [74] datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": "Loss functions. We first study the importance of the two loss functions, i.e., " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": ", as shown in Table 6. It can be seen that with our proposed " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": " loss function, the performance is significantly improved, i.e., " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "80.53\\%" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "82.96\\%" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": " on BTCV, " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "86.11\\%" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "88.82\\%" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": " on MM-WHS. These results demonstrate the effectiveness of our proposed position prediction pretext task. In addition, with the proposed regularization loss " + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 46, + 367, + 287, + 485 + ], + "type": "text", + "content": ", the performance can be further improved. Thus, we can see that it is crucial to learn discriminative bases in VoCo." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "Number of bases. We further evaluate different settings of the number of bases " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " in VoCo. We compare with different settings of " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " in the ablation studies, as shown in Table 7. It is worth noting that due to the ROI size inconsistency in the " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " direction, it is not practical to crop multiple bases in the " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " direction, since we have to resize the volume after crops, which will result in inconsistent volume scales. In addition, due to the computation limitation, it is costly to increase the values of " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": ". As shown in Table 7, with " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n = 2 \\times 2 \\times 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": ", the VoCo only achieves " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "81.56\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "86.73\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " Dice Score on BTCV and MM-WHS, respectively. When we increase the values of " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 3 \\times 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": ", the performances are improved obviously. Specifically, with " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": ", we achieve " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "83.85\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "90.54\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " on BTCV and MM-WHS, respectively. However, we observe that higher " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "5 \\times 5 \\times 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": ") cannot further bring higher performance. We further verify the performance of generating base crops in the " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " direction. It can be seen that " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 3 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " cannot yield improvements. Thus, aiming to balance the performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "content": "and efficiency, we set " + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 1" + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "content": " in VoCo. It can be seen that the setting of " + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "content": " is crucial to VoCo." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 281, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 329 + ], + "type": "text", + "content": "Visualization results on BTCV [35] are shown in Fig. 6. It can be seen that VoCo can yield improved segmentation accuracy and completeness. More visualization results are in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 339, + 493, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 493, + 352 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 493, + 352 + ], + "type": "text", + "content": "5. Conclusion and Future Directions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 552 + ], + "type": "text", + "content": "In this paper, we develop a simple-yet-effective SSL framework VoCo for 3D medical image analysis. Motivated by the observation that 3D medical images contain relatively consistent contextual positions between different organs, we propose to leverage the contextual position priors to learn consistent semantic representations in pre-training. Specifically, we crop volumes from different positions in an input volume and represent them as a group of bases to represent features in different directions. Then, we predict the contextual position of a randomly cropped volume by contrasting its similarity to different bases. In this way, VoCo effectively encodes the contextual position priors into model representations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experiments demonstrate that VoCo achieves superior performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": "We will further consider several ways of extension: (1) Scale up the pre-training dataset to evaluate the upper performance of VoCo. (2) Experiments on more downstream datasets. (3) Evaluate the label-efficient performance of VoCo (e.g., semi-supervised learning). (4) Explore the capacity of VoCo in mining inter-volume relationships." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 634, + 403, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 403, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 403, + 647 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by Hong Kong Innovation and Technology Fund (Project No. ITS/028/21FP and No. MHP/002/22), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. T45-401/22-N)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22880" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Michela Antonelli et al. The medical segmentation decathlon. Nature Commun., 13(1):4128, 2022. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 144 + ], + "type": "text", + "content": "[2] Shekoofeh Azizi et al. Big self-supervised models advance medical image classification. In ICCV, pages 3478-3488, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 146, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 286, + 178 + ], + "type": "text", + "content": "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. NIPS, 35:8799-8810, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 180, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 286, + 201 + ], + "type": "text", + "content": "[4] Patrick Bilic et al. The liver tumor segmentation benchmark (lits). Medical Image Analy., 84:102680, 2023. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 203, + 286, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 235 + ], + "type": "text", + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "type": "text", + "content": "[6] Mathilde Caron et al. Unsupervised learning of visual features by contrasting cluster assignments. NIPS, 33:9912-9924, 2020. 1, 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 269, + 286, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 291 + ], + "type": "text", + "content": "[7] Mathilde Caron et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 292, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 323 + ], + "type": "text", + "content": "[8] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 325, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 357 + ], + "type": "text", + "content": "[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In CVPR, pages 11526-11535, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 358, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 286, + 390 + ], + "type": "text", + "content": "[10] Ting Chen et al. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 392, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 286, + 423 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 425, + 286, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 286, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 286, + 457 + ], + "type": "text", + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 459, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 286, + 490 + ], + "type": "text", + "content": "[13] Zekai Chen et al. Masked image modeling advances 3d medical image analysis. In WACV, pages 1970-1980, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "text", + "content": "[14] Kenneth Clark and Bruce and others Vendl. The cancer imaging archive (tcia): maintaining and operating a public information repository. Jour. of Dig. Imag., 26:1045-1057, 2013. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "type": "text", + "content": "[15] Jiequan Cui et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 558, + 286, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 580 + ], + "type": "text", + "content": "[16] Jiequan Cui et al. Generalized parametric contrastive learning. IEEE Trans. Pattern Analy. Mach. Intell., 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 581, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 613 + ], + "type": "text", + "content": "[17] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 614, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 635 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "type": "text", + "content": "[19] Hao Du, Qihua Dong, Yan Xu, and Jing Liao. Weakly-supervised 3d medical image segmentation using geometric prior and contrastive similarity. IEEE Trans. Medi. Imag., 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[20] Yuting Gao, Jia-Xin Zhuang, et al. Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In ECCV, pages 237-253, 2022. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[21] Florin-Cristian Ghesu et al. Multi-scale deep reinforcement learning for real-time 3d-landmark detection in ct scans. IEEE Trans. Pattern Anal. Mach. Intell., 41(1):176-189, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[22] Jean-Bastien Grill et al. Bootstrap your own latent-a new approach to self-supervised learning. NIPS, 33:21271-21284, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "text", + "content": "[23] Katharina Grünberg et al. Annotating medical image data. Medical Image Analy., pages 45-67, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 177, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 219 + ], + "type": "text", + "content": "[24] Fatemeh Haghighi et al. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE Trans. Medical Imag., 40(10):2857-2868, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[25] Fatemeh Haghighi et al. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In CVPR, pages 20824-20834, 2022. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "type": "text", + "content": "[26] Ali Hatamizadeh et al. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In MICCAIW, pages 272-284, 2021. 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 291, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 322 + ], + "type": "text", + "content": "[27] Ali Hatamizadeh et al. Unetr: Transformers for 3d medical image segmentation. In WACV, pages 574-584, 2022. 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 325, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 346 + ], + "type": "text", + "content": "[28] Kaiming He et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "type": "text", + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 428, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 472 + ], + "type": "text", + "content": "[31] Xingxin He, Leyuan Fang, Mingkui Tan, and Xiangdong Chen. Intra-and inter-slice contrastive learning for point supervised oct fluid segmentation. IEEE Trans. Image Process., 31:1870-1881, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "text", + "content": "[32] Yuting He et al. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In CVPR, pages 9538-9547, 2023. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 508, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 552 + ], + "type": "text", + "content": "[33] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "text", + "content": "[34] Yankai Jiang et al. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In ICCV, pages 15859-15869, 2023. 1, 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 588, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 620 + ], + "type": "text", + "content": "[35] Bennett Landman et al. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In MICCAIW, volume 5, page 12, 2015. 5, 6, 7, 8" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 623, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 654 + ], + "type": "text", + "content": "[36] Jie Liu et al. Clip-driven universal model for organ segmentation and tumor detection. In ICCV, pages 21152-21164, 2023. 1, 5" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "text", + "content": "[37] Qiang Liu et al. A multi-level label-aware semi-supervised framework for remote sensing scene classification. IEEE Trans. Geosci. Remote Sens., 2023. 3" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22881" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[39] Xiangde Luo et al. WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from ct image. Medical Image Analysis, 82:102642, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 149 + ], + "type": "text", + "content": "[40] Jun Ma et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[41] T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In CVPR, pages 9339-9348, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "text", + "content": "[42] Nguyen et al. Joint self-supervised image-volume representation learning with intra-inter contrastive clustering. In AAAI, pages 14426-14435, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "text", + "content": "[43] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84. Springer, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "text", + "content": "[44] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "type": "text", + "content": "[45] Maxime Oquab et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 371 + ], + "type": "text", + "content": "[47] Rodrigo Santa Cruz et al. Deeppermnet: Visual permutation learning. In CVPR, pages 3949-3957, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "text", + "content": "[48] Arnaud Arindra Adiyoso Setio et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analy., 42:1-13, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "type": "text", + "content": "[49] Amber L Simpson et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 449, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 471 + ], + "type": "text", + "content": "[50] Aiham Taleb et al. 3d self-supervised methods for medical imaging. NIPS, 33:18158-18172, 2020. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 472, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 503 + ], + "type": "text", + "content": "[51] Yucheng Tang et al. Self-supervised pre-training of swin transformers for 3d medical image analysis. In CVPR, pages 20730-20740, 2022. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "text", + "content": "[52] Xing Tao et al. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In MICCAI, pages 238-248, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[53] Guotai Wang et al. Deepigeos: a deep interactive geodesic framework for medical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 41(7):1559-1572, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[54] Xiaosong Wang et al. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, pages 2097-2106, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "text", + "content": "[55] Yiqing Wang et al. Swinmm: masked multi-view with swim transformers for 3d medical image segmentation. In MIC-CAI, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "type": "text", + "content": "[56] Xin Wen et al. Self-supervised visual representation learning with semantic grouping. NIPS, 35:16423-16438, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[57] Linshan Wu et al. Deep bilateral filtering network for point-supervised semantic segmentation in remote sensing images. IEEE Trans. Image Process., 31:7419-7434, 2022. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 709 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[58] Linshan Wu et al. Modeling the label distributions for weakly-supervised semantic segmentation, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "text", + "content": "[59] Linshan Wu, Leyuan Fang, Xingxin He, Min He, Jiayi Ma, and Zhun Zhong. Querying labeled for unlabeled: Cross-image semantic consistency guided semi-supervised semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8827-8844, Jul. 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 150, + 545, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 192 + ], + "type": "text", + "content": "[60] Linshan Wu, Ming Lu, and Leyuan Fang. Deep covariance alignment for domain adaptive remote sensing image segmentation. IEEE Trans. Geosci. Remote Sens., 60:1-11, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "text", + "content": "[61] Linshan Wu, Zhun Zhong, Leyuan Fang, Xingxin He, Qiang Liu, Jiayi Ma, and Hao Chen. Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In CVPR, pages 15454-15464, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 238, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 545, + 281 + ], + "type": "text", + "content": "[62] Yutong Xie, Jianpeng Zhang, Yong Xia, and Qi Wu. Unimiss: Universal medical self-supervised learning via breaking dimensionality barrier. In ECCV, pages 558-575, 2022. 1, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "type": "text", + "content": "[63] Zhenda Xie et al. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 304, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 326 + ], + "type": "text", + "content": "[64] Shuangfei Zhai et al. Position prediction as an effective pretraining strategy. arXiv preprint arXiv:2207.07611, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "type": "text", + "content": "[65] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Anal., 89:102879, 2023. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 360, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 379 + ], + "type": "text", + "content": "[66] Hongyi Zhang et al. Mixup: Beyond empirical risk minimization. In ICLR, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 380, + 545, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 545, + 423 + ], + "type": "text", + "content": "[67] Kang Zhang et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181(6):1423-1433, 2020. 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 425, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 425, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 425, + 545, + 456 + ], + "type": "text", + "content": "[68] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In AAAI, pages 3516-3524, 2023. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 458, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 458, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 458, + 545, + 491 + ], + "type": "text", + "content": "[69] Hong-Yu Zhou et al. Comparing to learn: Surpassing image-net pretraining on radiographs by comparing image representations. In MICCAI, pages 398-407, 2020. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 491, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 523 + ], + "type": "text", + "content": "[70] Hong-Yu Zhou et al. Preservational learning improves self-supervised medical image models by reconstructing diverse contexts. In ICCV, pages 3499-3509, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "text", + "content": "[71] Hong-Yu Zhou et al. A unified visual information preservation framework for self-supervised pre-training in medical image analysis. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 567, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 589 + ], + "type": "text", + "content": "[72] Zongwei Zhou et al. Models genesis. Medical Image Analy., 67:101840, 2021. 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "text", + "content": "[73] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "text", + "content": "[74] Xiahai Zhuang. Multivariate mixture model for myocardial segmentation combining multi-source images. IEEE Trans. Pattern Analy. Mach. Intell., 41(12):2933-2946, 2018. 5, 6, 7, 8" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 677, + 545, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 545, + 709 + ], + "type": "text", + "content": "[75] Xinrui Zhuang et al. Self-supervised feature learning for 3d medical images by playing a rubik's cube. In MICCAI, pages 420-428, 2019. 1, 3, 6" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22882" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_content_list.json b/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..16a661dd0e069541caefa5d8a4edc184aaf7735a --- /dev/null +++ b/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_content_list.json @@ -0,0 +1,1874 @@ +[ + { + "type": "text", + "text": "Volumetric Environment Representation for Vision-Language Navigation", + "text_level": 1, + "bbox": [ + 148, + 161, + 821, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Liu Wenguan Wang Yi Yang*", + "bbox": [ + 346, + 208, + 629, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ReLER, CCAI, Zhejiang University", + "bbox": [ + 387, + 226, + 584, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/DefaultRui/VLN-VER", + "bbox": [ + 333, + 242, + 635, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 287, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language navigation (VLN) requires an agent to navigate through an 3D environment based on visual observations and natural language instructions. It is clear that the pivotal factor for successful navigation lies in the comprehensive scene understanding. Previous VLN agents employ monocular frameworks to extract 2D features of perspective views directly. Though straightforward, they struggle for capturing 3D geometry and semantics, leading to a partial and incomplete environment representation. To achieve a comprehensive 3D representation with fine-grained details, we introduce a Volumetric Environment Representation (VER), which voxelizes the physical world into structured 3D cells. For each cell, VER aggregates multi-view 2D features into such a unified 3D space via 2D-3D sampling. Through coarse-to-fine feature extraction and multi-task learning for VER, our agent predicts 3D occupancy, 3D room layout, and 3D bounding boxes jointly. Based on online collected VERs, our agent performs volume state estimation and builds episodic memory for predicting the next step. Experimental results show our environment representations from multi-task learning lead to evident performance gains on VLN. Our model achieves state-of-the-art performance across VLN benchmarks (R2R, REVERIE, and R4R).", + "bbox": [ + 109, + 316, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 678, + 233, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language navigation (VLN) requires an agent to navigate in a 3D environment following natural language instructions [3, 64]. As a holistic understanding of the environment plays a pivotal role in decision-making within VLN, environment representation learning serves as a foundation for formulating accurate navigation policies.", + "bbox": [ + 109, + 702, + 473, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early VLN approaches [3, 23] typically learn the navigation policy through the sequence-to-sequence (Seq2Seq) framework [72], which directly maps instructions and multi-view perspective observations to actions. They sim", + "bbox": [ + 109, + 786, + 470, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f19d63d8b8ddb11e9961a01e4077fc09ef996fc26f68a809ff701d3ad8f3b53.jpg", + "image_caption": [ + "Previous Methods", + "Perspective", + "Observations", + "(1)", + "#", + "#", + "", + "#", + "2D Features of", + "Candidate View", + "#", + "Topological Graph", + "$\\rightarrow$", + "", + "$\\frac{1}{2}x - 1 > 0$", + "Semantic Map", + "1", + "2D-3D Sampling", + "23. 15 sampling", + "#", + "", + "multi-view", + "D Features", + "$\\therefore m - 1 \\neq 0$ ;", + "3D Occupancy", + "(1)", + "3D Detection -", + "$\\therefore m : x = 1$ 或 ${3x} + {4y} + 1 = 0$", + "Room Layout", + "Estimation", + "#", + "Volume State Space", + "Volume State Space", + "Figure 1. The agent observes its surroundings with corresponding perspective features of different candidate views (O). Previous methods construct the topological graph or semantic map based on these 2D features. Our VER aggregates the multi-view features into structured 3D cells via 2D-3D sampling. VER is a powerful representation for both 3D perception tasks and VLN, providing a volume state space for decision-making." + ], + "image_footnote": [], + "bbox": [ + 504, + 286, + 645, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our Method", + "bbox": [ + 506, + 388, + 571, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our Method", + "bbox": [ + 506, + 388, + 571, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ply embed their immediate observation of local environment into the hidden states of recurrent units. As a result, they lack of explicit environment representations and struggle to access their past states during long-time exploration [61, 82]. To address this issue, later VLN agents are equipped with an external memory module [61], which stores the environment representations distinctly from navigation states. In this way, they can explicitly model and maintain the environment layouts and contents in a form of topological graph [2, 11, 16, 19, 32, 62] or semantic map [1, 4, 10, 12, 26, 34, 38, 55, 75, 94] (Fig. 1). Despite their promising performance with advanced frameworks (e.g., graph neural network [44] and Transformer [79]), their environment representations are still built upon 2D perspective features extracted by monocular frameworks. While straightforward, they compress depth information onto the perspective plane, sacrificing the integral scene structure in the 3D space. Thus, they encounter challenges in capturing 3D geometry and semantics in complex scenes. Such an incomplete environment representation easily leads to suboptimal navigation decisions.", + "bbox": [ + 496, + 575, + 859, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 259, + 42, + 781, + 56 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 338, + 56, + 704, + 70 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 308, + 71, + 732, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Yi Yang.", + "bbox": [ + 129, + 853, + 292, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16317", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this article, we propose a Volumetric Environment Representation (VER) that quantizes the physical world into structured 3D cells (Fig. 1). These cells, arranged within a predefined volumetric space, maintain both height and depth dimensions. Each cell corresponds to local context of the 3D space. VER aggregates multi-perspective 2D features within these cells through an environment encoder (§3.1). Compared to previous partial representations derived from hidden states and external memory, our VER captures the full geometry and semantics of the physical world. These 3D cells store the properties of the corresponding space in the scene by predicting 3D occupancy [70, 77], room layout [106], and 3D object boxes [51]. However, directly reconstructing the high-quality VER from 2D perspective views is challenging to capture the fine-grained details. As a response, we propose a coarse-to-fine VER extraction architecture, which uses learnable up-sampling operations to construct the representations progressively. It is supervised by multi-resolution semantic labels at different scales, utilizing the coarse-to-fine representations as hierarchical inputs. The annotations of the 3D tasks are collected for multi-task learning (§3.4).", + "bbox": [ + 109, + 126, + 472, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At each navigation step, our agent initially encodes the multi-view observations into VER (§3.2). With VER, instructions can be more effectively grounded in the 3D context. This is achieved by establishing cross-modal correlations between linguistic words and 3D cells of VER. Based on the correlations, a volume state estimation module is proposed to calculate transition probabilities over the surrounding cells. With the help of this module, our agent performs comprehensive decision-making in volumetric space, and then maps the volume state into local action space. In addition, an episodic memory module is established to online collect the information of observed viewpoints and build a topological graph providing global action space (§3.3). The node embeddings in the graph are from neighbor pillar representations in VER corresponding to the respective viewpoints. To balance the long-range action reasoning and language grounding, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory.", + "bbox": [ + 109, + 430, + 472, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our agent is evaluated on three VLN benchmarks, i.e., R2R [3], REVERIE [64], and R4R [39] ( $\\S 4.1$ ). It yields solid performance gains (about $3\\%$ SR and $4\\%$ SPL on R2R test, $4\\%$ SR and $4\\%$ SPL on REVERIE val unseen). The ablation study confirms the efficacy of core model designs ( $\\S 4.2$ ). Additional results show our model achieves promising performance in 3D occupancy prediction, 3D detection, and room layout estimation ( $\\S 4.3$ ).", + "bbox": [ + 109, + 691, + 470, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 112, + 815, + 243, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Navigation (VLN). Early VLN agents [3, 23] are built upon Seq2Seq [66, 72] framework to re", + "bbox": [ + 109, + 838, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "serve the observation history in hidden state. Thus they struggle to capture long-range context as the path length increases. Later efforts are devoted to multimodal representation learning, navigation strategy learning, and data generation [27]. As a primary step, multimodal representation learning helps agents understand the environments and establish relations between the instructions and visual observations. Inspired by the success of vision-language pretraining [58, 65, 73], recent approaches [14, 33] use transformer-based architectures [42, 79] for joint visual and textual representations. Some attempts further exploit the visual information by modeling semantic relation [37] and spatial information [11, 19, 32, 83, 86, 100]. For navigation strategy learning, many VLN models [24, 88] use imitation and reinforcement learning-based training strategies. Previous solutions [45, 85] introduce world models [29] to perform mental simulations and make mental planning. Furthermore, the scarcity of human instructions and limited diversity of the scene hinder the agent to learn navigation policy and generalize to unseen environments well [54]. Therefore, several VLN data generation strategies have been proposed to create new trajectories from existing datasets [15, 28, 93], generate more instructions [41, 74, 84, 90], or create synthetic environments [48, 49]. In addition, driven by large models [5], existing agents [13, 57, 71, 101] demonstrate promising zero-shot performance.", + "bbox": [ + 496, + 126, + 859, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their outstanding contributions, most of them rely on 2D visual cues in perspective observations. These representations are constrained by occlusion and limited geometric information, especially in complex scenes. In this paper, we propose VER, a unified environment representation learned by 3D perception tasks. During navigation, our agent performs volume state estimation on VER, facilitating comprehensive decision-making within the 3D space.", + "bbox": [ + 496, + 489, + 859, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Environment Representation. Existing VLN models introduce various representations for environment modeling, including topological graphs [2, 11, 16, 19, 32, 82], and semantic spatial representations [1, 4, 12, 26, 34, 38, 55, 75, 94]. With a broader view, diverse representations have been proposed for robotics and autonomous driving [25, 47, 68, 76, 96]. In the early stage, 2D occupancy grid maps [22] model occupied and free space in the surroundings based on Bayesian estimation for robot navigation. Classic SLAM systems [21] construct a map directly by integrating information from various sensors, including LiDAR and cameras. However, the representations in SLAM still rely on primitives such as 3D point clouds and image patches. Some efforts [6, 8, 31] focus on developing learnable semantic map representations. To enhance spatial reasoning, scene graph representations [9, 80] define the topological relations between spatial elements of the environment. In addition, neural scene representations [46, 50] embed image observations into latent codes", + "bbox": [ + 496, + 603, + 859, + 867 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16318", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f94df0618eedd39947ddf9e7e46ca8f5ffecdd45e20241f110237182613679f.jpg", + "image_caption": [ + "Figure 2. Overview of our model. Given the perspective features of candidate views, a group of 3D queries are used to sample and aggregate them into VER (§3.1). To encode VER, we adopt coarse-to-fine extraction and perform multi-task learning on 3D perception. Based on VER, a volume state estimation module is proposed to predict state transition (§3.2). The episodic memory is used to store past observations using neighboring pillar representations for each viewpoint (§3.3). For decision-making, our agent combines both the local action probabilities from the volume state and the global action probabilities obtained from the episodic memory. See §3 for more details." + ], + "image_footnote": [], + "bbox": [ + 122, + 111, + 851, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for object categories, showcasing scalability to large scenes [40, 59]. Vision-centric BEV perception, which transforms perspective-view inputs to BEV grid representations, has recently received increasing attention [47, 51, 63]. As the BEV representations simplify the vertical geometry, 3D occupancy prediction [36, 78] is further proposed to infer the 3D geometry from perspective images [68, 70, 87, 98].", + "bbox": [ + 111, + 338, + 472, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, existing VLN agents mainly employ abstract relations or compressed spatial maps, lacking the ability to access complete scene information. Providing more world context can be beneficial for the subsequent decision-making and policy learning. Motivated by this insight, we explore a holistic environment representation VER that voxelizes the world into structured 3D cells. VER captures both semantic information and geometric details of the whole scene. Building upon VER, our agent is able to predict the 3D occupancy, room layout, and 3D boxes accurately.", + "bbox": [ + 111, + 436, + 473, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 112, + 604, + 212, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. For brevity, we present the technical description in the context of R2R [3]. The navigable area of the environment is organized as an undirected graph, containing a set of nodes (viewpoints) and connectivity edges. In R2R, an embodied agent needs to navigate to a target location in the 3D environment following human instructions with $L$ words (embedded as $E \\in \\mathbb{R}^{D_w \\times L}$ , where $D_w$ is the channel dimension). At time step $t$ , the agent looks around and obtains multi-view observations of its surrounding scene from the current location. Each view is represented by a 2D visual feature $F_t^{2d} \\in \\mathbb{R}^{D_i \\times H \\times W}$ , where $H$ and $W$ are the spatial shape of image plane, $D_i$ denotes the channel dimension. The local action space $A_t \\in \\mathbb{R}^{N_t + 1}$ is defined by $N_t$ candidate views, which correspond to neighboring navigable nodes $\\{v_{t,n}^*\\}_{n=1}^{N_t}$ , as well as a [STOP] action. Previous agents predict the action probabilities $p_t^{2d} \\in \\mathbb{R}^{N_t + 1}$ directly based on $F_t^{2d}$ of each candidate view. However,", + "bbox": [ + 111, + 628, + 473, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these 2D features with limited geometric information are partial representations of the 3D environment, easily leading to suboptimal decision making.", + "bbox": [ + 496, + 338, + 857, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. To achieve comprehensive scene understanding, we introduce VER, which voxelizes the 3D world into structured 3D cells (Fig. 2). At step $t$ , an environment encoder is proposed to sample multi-view features $(F_{t}^{2\\mathrm{d}}$ of each view) into the volumetric space of VER, forming a unified representation $F_{t}^{3\\mathrm{d}} \\in \\mathbb{R}^{D_{e} \\times X \\times Y \\times Z}$ (\\$3.1). $X$ and $Y$ are the shapes of horizontal plane, $Z$ reserves the height information of 3D space, and $D_{e}$ represents the channel dimension. The volumetric space aligns with gravity in the world coordinate system based on the Manhattan assumption [106]. To encode VER, we devise coarse-to-fine extraction with multiple 3D perception tasks supervised by multi-resolution annotations (\\$3.4). Based on VER, a volume state estimation module is proposed to predict state transition probabilities $p_{t}^{3\\mathrm{d}} \\in \\mathbb{R}^{X \\times Y \\times Z}$ over surrounding 3D cells (\\$3.2). With this module, our agent performs comprehensive decision-making in the 3D space and maps $p_{t}^{3\\mathrm{d}}$ to $p_{t}^{2\\mathrm{d}}$ . To predict the next step, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory (\\$3.3).", + "bbox": [ + 496, + 378, + 859, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.Environment Encoder", + "text_level": 1, + "bbox": [ + 500, + 665, + 689, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D-3D Sampling. At step $t$ , the agent observes its surroundings and acquires the multi-view images. We introduce cross-view attention (CVA) to aggregate their features $(F^{2d}$ for each view) into a unified volumetric representation $F^{3d}$ with a group of learnable volume queries $Q \\in \\mathbb{R}^{D_e \\times X \\times Y \\times Z}$ ( $t$ is omitted for simplicity). Specifically, for the 3D cell positioned at $(x,y,z)$ within the egocentric world, a single query $Q(x,y,z) \\in \\mathbb{R}^{D_e}$ is used to sample each image feature $F^{2d}$ as:", + "bbox": [ + 496, + 686, + 859, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {3 d} (x, y, z) = \\operatorname {C V A} \\left(\\boldsymbol {Q} (x, y, z), \\boldsymbol {F} ^ {2 d} \\left(h ^ {\\prime}, w ^ {\\prime}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 815, + 857, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $(h', w')$ denotes the location of corresponding sampling point on the image plane. Note that we only show", + "bbox": [ + 496, + 838, + 857, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16319", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2d216cda22b946887bfdb1b7ccd0e3a5ae64a419ee1d12997b24a489f0ffbc98.jpg", + "image_caption": [ + "Figure 3. Our coarse-to-fine VER representation extraction (§3.1) adopts cascade up-sampling operations with 3D deconvolutions (Eq. 2) and 3D queries (Eq. 1). The training process is supervised at different scales by multi-resolution semantic labels." + ], + "image_footnote": [], + "bbox": [ + 117, + 125, + 467, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the formulation of a single sampling point for conciseness. Since the sampling strategy of vanilla cross-attention is computationally expensive, the deformable attention [51, 104] is introduced and extended in CVA. In this way, $Q(x,y,z)$ selectively attends to a set of key sampling points around a reference instead of the entire $F^{2\\mathrm{d}}$ .", + "bbox": [ + 111, + 340, + 470, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-Fine VER Extraction Architecture. Directly recovering the fine-grained VER from perspective features easily leads to performance and efficiency degradation [17, 95]. The coarse-to-fine extraction is proposed to reconstruct VER progressively. Our approach involves cascade up-sampling operations (Fig. 3), dividing this extraction into $M$ levels. At each level, 3D deconvolutions are utilized for lifting spatial resolution, and then CVA is used to query the multi-view 2D features (Eq. 1) for refining the detailed geometry. This enables the direct learning of details and avoids the inaccuracy of interpolation [56, 77, 92] (see Table 7). Between the input coarse feature $F^{3\\mathrm{d}(0)} \\in \\mathbb{R}^{D_e \\times \\frac{X}{2^M} \\times \\frac{Y}{2^M} \\times \\frac{Z}{2^M}}$ from Eq. (1) and the target fine feature $F^{3\\mathrm{d}(M)} \\in \\mathbb{R}^{D_e \\times X \\times Y \\times Z}$ , the intermediate features with varying shapes are calculated as follows:", + "bbox": [ + 111, + 424, + 473, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {\\mathrm {3 d} (1)} = \\uparrow \\boldsymbol {F} ^ {\\mathrm {3 d} (0)}, \\dots , \\boldsymbol {F} ^ {\\mathrm {3 d} (M)} = \\uparrow \\boldsymbol {F} ^ {\\mathrm {3 d} (M - 1)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 636, + 470, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F^{3\\mathrm{d}(1)}, \\dots, F^{3\\mathrm{d}(M-1)}$ denote the intermediate features from different levels, and $\\uparrow$ denotes the up-sampling. Multi-task Learning. Our VER offers a unified scene representation for various 3D perception tasks. Existing studies [68, 70] highlight that semantics and geometry are tightly intertwined. We train our environment encoder to extract VER under the supervision of multiple 3D perception tasks (§3.4). This process utilizes $\\{F^{3\\mathrm{d}(0)}, \\dots, F^{3\\mathrm{d}(M)}\\}$ as inputs and is supervised at different scales (Fig. 3). For 3D occupancy prediction, the decoder is implemented as MLPs with the focal loss [52]. For 3D layout estimation, we adopt a query-based head to yield the Manhattan room layouts. A combination of the L1 loss and the IoU loss [67] is used as the training objective. For 3D detection, we employ a detection head to predict the 3D boxes [51]. The bipartite match", + "bbox": [ + 111, + 657, + 472, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing and the bounding box loss [51, 104] are employed for detection. A weight vector $[2.0, 0.25, 0.25]$ is used to balance the optimization of the three tasks, respectively. During navigation, the agent traverses between different viewpoints and encodes VERs through the frozen environment encoder.", + "bbox": [ + 496, + 125, + 859, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Volume State Estimation", + "text_level": 1, + "bbox": [ + 500, + 217, + 705, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLN task is typically viewed as a state estimation and transition problem [76]. With our VER, the agent state is represented as \"volume state\". As such, the state transition within the locally observed 3D environment, computed by Eq. (3)&(4), is referred as \"volume state estimation\". Different from previous plane-level state models [1, 4], VER introduces additional height dimension for 3D state estimation. This enables a more accurate action prediction.", + "bbox": [ + 496, + 238, + 857, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Volume State. At the beginning of a navigation episode, the agent is located at a start viewpoint $(x_0,y_0,z_0)$ . Based on its perception range, a volume state space $\\mathcal{X}\\in \\mathbb{R}^{X\\times Y\\times Z}$ is defined corresponding to the 3D physical world with an initial state $s_0 = (x_0,y_0,z_0)$ . At step $t$ , the next intermediate state $s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})$ is determined by the instruction embeddings $\\pmb{E}$ and VER $\\pmb{F}_t^{3\\mathrm{d}}$ for reaching the goal state $s_T$ $(0 < t < T)$ . As the entire environment is partially observable, the current state transition $(s_t\\rightarrow s_{t + 1}$ in $\\mathcal{X}$ ) is regarded as a local consideration for action prediction.", + "bbox": [ + 496, + 349, + 859, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State Estimation. A volume state estimation module is devised to predict the probability distribution $\\pmb{p}_t^{3\\mathrm{d}} \\in \\mathbb{R}^{X \\times Y \\times Z}$ of the intermediate state $s_{t+1}$ conditioned on $\\pmb{E}$ and $\\pmb{F}_t^{3\\mathrm{d}}$ . The environment representation is first reshaped as $\\pmb{F}_t^{3\\mathrm{d}''} \\in \\mathbb{R}^{D_e \\times XYZ}$ , and then adopt multi-layer transformers (MLT) to model the relations between $\\pmb{E}$ and $\\pmb{F}_t^{3\\mathrm{d}''}$ as follows:", + "bbox": [ + 496, + 487, + 857, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {F}} _ {t} ^ {3 d} = \\mathrm {M L T} \\left(\\left[ \\boldsymbol {E}; \\boldsymbol {F} _ {t} ^ {3 d ^ {\\prime}} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times X Y Z}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 574, + 857, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widetilde{F}_t^{3\\mathrm{d}}$ is the updated representations, and $[\\cdot ]$ denotes the concatenation operation. MLT consists of stacked selfattention blocks. Then we use MLPs for state estimation:", + "bbox": [ + 496, + 598, + 857, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} ^ {\\mathrm {3 d}} = \\operatorname {S o f t m a x} \\left(\\mathrm {M L P} \\left(\\widetilde {\\boldsymbol {F}} _ {t} ^ {\\mathrm {3 d}}\\right)\\right) \\in [ 0, 1 ] ^ {X \\times Y \\times Z}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 642, + 857, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Height-aware Group. The computational and memory efficiency of Eq. (3) is compromised due to the resolution of $F_{t}^{3\\mathrm{d}}$ ( $XYZ \\gg L$ ). Considering the similarity and sparsity of information along the height direction [97, 99], we partition $F_{t}^{3\\mathrm{d}}$ into several uniform groups $\\{F_{t,z}^{g} \\in \\mathbb{R}^{D_e \\times XY}\\}_{z=1}^Z$ along this axis. Then we apply MLT to each group, and Eq. (3) is reformulated as:", + "bbox": [ + 496, + 662, + 857, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {F}} _ {t, i} ^ {g} = \\operatorname {M L T} \\left(\\left[ \\boldsymbol {E}; \\mathbf {F} _ {t, i} ^ {g} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times X Y}, \\\\ \\widetilde {\\boldsymbol {F}} _ {t} ^ {3 \\mathrm {d}} = \\left\\{\\widetilde {\\boldsymbol {F}} _ {t, z} ^ {g} \\right\\} _ {z = 1} ^ {Z} \\in \\mathbb {R} ^ {D _ {e} \\times X Y Z}, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 767, + 856, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the weights of MLT are shared among different groups. The updated features from different groups $\\{\\widetilde{F}_{t,z}^{g}\\}_{z = 1}^{Z}$ are aggregated along the height to leverage complementary information. For ease of notation, the symbol", + "bbox": [ + 496, + 810, + 857, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16320", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\widetilde{F}_t^{3\\mathrm{d}}$ is slightly reused for the gathered 3D representations. Then, $\\pmb{p}_t^{3\\mathrm{d}}$ is calculated by Eq. (4).", + "bbox": [ + 112, + 125, + 470, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Action Prediction", + "text_level": 1, + "bbox": [ + 112, + 162, + 269, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For action prediction across the entire explored scene, a topological graph $\\mathcal{G}_t = \\{\\mathcal{V}_t,\\mathcal{E}_t\\}$ is constructed and updated online to represent episodic memory during navigation. Specifically, $\\mathcal{V}_t$ denotes the observed viewpoints, i.e., all visited viewpoints and their candidate viewpoints. These viewpoints are encoded by compressing previous VERs. The edge $\\mathcal{E}_t$ denotes the navigable connections between these viewpoints. To predict the next step, our agent combines both the volume state estimation and episodic navigation memory for decision-making.", + "bbox": [ + 111, + 184, + 470, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mapping Volume State to Action. As our agent navigates on the horizontal plane to reach the adjacent candidate viewpoints $\\{v_{n}^{*}\\}_{n = 1}^{N_{t}}$ , we map the volume state space into 2D space to align with this movement pattern. Specifically, we average $p_t^{3d}$ along the height (z-axis) axis as $p_t^h\\in \\mathbb{R}^{X\\times Y}$ . Then we sum probability values in the neighborhood of $\\{v_{t,n}^{*}\\}_{n = 0}^{N_{t}}$ ( $v_{0}^{*}$ for the current viewpoint, i.e., [STOP]), and normalize them as the local action probabilities:", + "bbox": [ + 112, + 323, + 473, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} ^ {\\mathrm {2 d}} = \\left\\{\\sum \\boldsymbol {p} _ {t} ^ {h} \\left(x _ {n}, y _ {n}\\right) \\right\\} _ {n = 0} ^ {N _ {t}} \\in [ 0, 1 ] ^ {N _ {t} + 1}, \\left(x _ {n}, y _ {n}\\right) \\in \\Omega_ {n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 439, + 470, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}_t^h (x_n,y_n)$ is the value at the coordinate $(x_{n},y_{n})$ and $\\Omega_{n}$ is the neighborhood of $v_{n}^{*}$ in the horizontal plane. In the training stage, a heatmap [102] with a Gaussian kernel is used to supervise this action prediction (§3.5).", + "bbox": [ + 111, + 465, + 470, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global Action Prediction. The episodic memory module is used to store past environment representations and allows easy access to them. To memory the environment representations efficiently, we use the neighboring pillar [91] representations $\\pmb{F}_{t,n}^{p} \\in \\mathbb{R}^{D_{e} \\times |\\Omega_{n}| \\times Z}$ , corresponding to the current observed viewpoints $\\{v_{n}^{*}\\}_{n=0}^{N_{t}}$ at step $t$ :", + "bbox": [ + 111, + 522, + 470, + 608 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {t, n} ^ {P} = \\left\\{\\tilde {\\boldsymbol {F}} _ {t} ^ {3 d} \\left(x _ {n}, y _ {n}, z _ {n}\\right) \\right\\} _ {\\left(x _ {n}, y _ {n}\\right) \\in \\Omega_ {n}, z _ {n} \\in [ 1, Z ]} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 609, + 470, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{F}_t^{3\\mathrm{d}}(x_n,y_n,z_n)$ is the representation at position $(x_{n},y_{n},z_{n})$ of $\\widetilde{F}_t^{3\\mathrm{d}}$ (Eq. 5). $\\{\\bar{F}_{t,n}^p\\in \\mathbb{R}^{D_e}\\}_{n = 0}^{N_t}$ is obtained by average pooling as the corresponding node embeddings, which are then incorporated into $\\mathcal{G}_t$ . For previously observed nodes, we compute the average of their features.", + "bbox": [ + 112, + 637, + 470, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The episodic memory $\\mathcal{G}_t$ , which includes the observed viewpoints, offers a global action space $\\mathcal{A}_t^* \\in \\mathbb{R}^{|\\mathcal{V}_t|}$ . This enables our agent to change its current navigation state easily by 'jumping' directly to another viewpoint, which may be even observed several steps ago. The global action probabilities on $\\mathcal{G}_t$ are calculated as:", + "bbox": [ + 111, + 709, + 470, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {G}} _ {t} = \\operatorname {M L T} \\left(\\left[ \\boldsymbol {E}; \\boldsymbol {G} _ {t} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times | \\mathcal {V} _ {t} |}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 796, + 470, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} ^ {g} = \\operatorname {S o f t m a x} \\left(\\operatorname {M L P} \\left(\\hat {\\boldsymbol {G}} _ {t}\\right)\\right) \\in [ 0, 1 ] ^ {| \\nu_ {t} |},\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 816, + 400, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G_{t}$ denotes the node embeddings of $\\mathcal{G}_t$ . The ultimate action probabilities are given as:", + "bbox": [ + 112, + 838, + 470, + 866 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} ^ {2 \\mathrm {d}} = \\left[ \\boldsymbol {p} _ {t} ^ {2 \\mathrm {d} \\rightarrow g}; \\boldsymbol {p} _ {t} ^ {2 \\mathrm {d}} \\right] \\in [ 0, 1 ] ^ {| \\mathcal {V} _ {t} |}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 122, + 857, + 145 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} _ {t} ^ {g} = W _ {g} \\boldsymbol {p} _ {t} ^ {g} + (1 - W _ {g}) \\hat {\\boldsymbol {p}} _ {t} ^ {2 d} \\in [ 0, 1 ] ^ {| \\mathcal {V} _ {t} |},\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 142, + 792, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}_t^{2\\mathrm{d}\\rightarrow g} \\in \\mathbb{R}^{|\\mathcal{V}_t| - (N_t + 1)}$ denotes the probabilities of global backtracking and we use the same value as the local [STOP] probability in Eq. (6); $W_{g}$ is a learnable weight.", + "bbox": [ + 496, + 165, + 857, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State Transition and Memory Update. After executing the action in $\\mathcal{A}_t^*$ , our agent reaches the next viewpoint $v_{t + 1,0}^{*}$ , and will iteratively: (1) encode its current observation as $F_{t + 1}^{3\\mathrm{d}}$ through Eq. (1); (2) update its volume state as $s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})$ ; (3) add the node embeddings of $\\{v_{t + 1,n}^{*}\\}_{n = 0}^{N_{t + 1}}$ into the episodic memory $\\mathcal{G}_{t + 1}$ ; and (4) predict the next step with the updated episodic memory (i.e., $\\mathcal{G}_{t + 1}$ ) and volume state (i.e., $s_{t + 1}$ ). Our agent repeats the above process until it chooses the [STOP] action or reaches the maximum step limit.", + "bbox": [ + 496, + 209, + 857, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Annotation Generation", + "text_level": 1, + "bbox": [ + 500, + 359, + 694, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A multi-task learning framework is proposed to extract and encode the VERs (§3.1). To achieve this, we generate annotations on Matterport3D dataset [7] for 3D occupancy prediction, object detection, and room layout estimation. We design a room-object-voxel pipeline to automatically generate these annotations. This pipeline leverages existing LiDAR point labels without additional human annotations (more details in Appendix). We utilize the egocentric observations with multi-view images as input.", + "bbox": [ + 496, + 381, + 857, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Room Layout. The room layout in our context specifies the positions, orientations, and heights of the walls, relative to the camera center. It aims to reconstruct cuboid room shapes within the Manhattan world [18]. Given that the horizontal plane is aligned on the $x - z$ axis, we parameterize the layout with center coordinate, width, length, height, and rotation. In contrast to directly operating on a single panoramic image [105, 106], we use the embodied observations with multi-view images as input.", + "bbox": [ + 496, + 507, + 857, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Detection. Based on the room layout, we collect the surrounding objects if the agent locates in a room. In a non-closed environment, we collect information about nearby objects based on their distance from the agent. For each object, the eigenvectors of its vertices are used to define an oriented bounding box that tightly encloses the object [35]. Considering some objects may disappear from view due to occlusion but still exist in the environment (referred to as permanence [98]), we also include them in the analysis.", + "bbox": [ + 496, + 631, + 857, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Accumulation for Occupancy. To generate voxel labels for occupancy [77, 95], we accumulate the sparse LiDAR points and utilize 3D boxes. Given dense background and object points, we first voxelize the 3D space and label each voxel based on the majority vote of labelled points in that voxel. Due to the limited number of LiDAR points, we leverage the Nearest Neighbors algorithm to generate dense labels for remaining voxels by searching the nearest", + "bbox": [ + 496, + 756, + 857, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16321", + "bbox": [ + 480, + 890, + 514, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "semantic label. Moreover, the agent infers the complete 3D occupancy of each object (amodal perception), including regions that are not directly observed. This attribute enables the agent to predict the entire object instead of only visible surfaces [98].", + "bbox": [ + 111, + 125, + 470, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Statistics. For high-resolution labels, we define $120 \\times 120 \\times$ 35 voxel grids in world coordinates, where the scene voxel size equals to $0.1\\mathrm{m}$ (see more details about multi-resolution labels in Appendix). We annotate over 50 billion voxels and 16 classes, including 11 foreground objects and 5 background stuffs. It comprises about $100k$ annotated bounding boxes and 1,500 room layouts within the scenes. These annotations follow the same train/val/test splits as R2R [3]. There are 61 scenes for train/val seen, 11 scenes for val unseen, and 18 scenes for test.", + "bbox": [ + 111, + 196, + 473, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 112, + 344, + 312, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initially, the environment encoder (§3.1) is introduced for VER through coarse-to-fine extraction. Then multi-task learning is performed across multiple 3D perception tasks, including 3D occupancy prediction, 3D layout estimation, and 3D detection. During navigation, our agent is equipped with the frozen environment encoder and predicts the next step. Following recent VLN practice [14, 16, 33], both offline pretraining and finetuning are adopted. In this section, we will mainly introduce the details of architecture and training (see more details in Appendix).", + "bbox": [ + 111, + 366, + 470, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Environment Representation Learning. For the multiview images, we adopt ViT-B/16 [20] pretrained on ImageNet to extract features. The number of 3D volume queries is $15 \\times 15 \\times 4$ . For each query, it is projected to sample 2D views according to intrinsic and extrinsic parameters of the camera. We set the perception range as $[-6m, 6m]$ for $x - y$ axis and $[-1.5m, 2m]$ for the height ( $z$ axis). We adopt six layers of CVA (Eq. 1) for 2D-3D sampling, and then use $M = 3$ cascade deconvolutions for up-sampling (Eq. 2). The feature dimension is 768 (i.e., $D_{i} = D_{w} = D_{e} = 768$ ).", + "bbox": [ + 111, + 506, + 470, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Navigation Network. MLT with 4 layers is initialized from [73] for state estimation (Eq. 4) and global action prediction (Eq. 8), respectively. The range of neighborhood for each candidate is set as $|\\Omega_n| = 9$ . The standard deviation of the Gaussian kernel for the heat map is set as 3.0. Based on this heat map, the focal loss [52] is used to supervise the local action prediction (Eq. 6). We also use a cross-entropy loss for the global action prediction (Eq. 9).", + "bbox": [ + 111, + 645, + 470, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining. For R2R [3] and R4R [39], Masked Language Modeling [14, 42] and Single-step Action Prediction [14, 33] are adopted as auxiliary tasks on offline-sampled instruction-route pairs [30]. For REVERIE [64], an additional Object Grounding (OG) [16, 53] is used for object reasoning. During pretraining, we train the model with a batch size of 64 for $100k$ iterations, using Adam [43] optimizer with 1e-4 learning rate. Only one task is adopted", + "bbox": [ + 111, + 755, + 470, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c33933339691c0e610aa0a6e7bf9dd1b1aa1828212e276cf849463c33141eed7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsR2R
val unseentest unseen
TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑
Seq2Seq [3]8.397.8122-8.137.852018
SF [23]-6.6235-14.826.623528
EnvDrop [74]10.705.22524811.665.235147
AuxRN [103]-5.285550-5.155551
Active [81]20.604.36584021.604.336041
RecBERT [33]12.013.93635712.354.096357
HAMT [14]11.462.29666112.273.936560
SOAT [60]12.154.28595312.264.495853
SSM [82]20.74.32624520.44.576146
CCC [84]-5.205046-5.305148
HOP [65]12.273.80645712.683.836459
DUET [16]13.943.31726014.733.656959
LANA [89]12.0-686212.6-6560
TD-STP [100]-3.227063-3.736761
BSG [55]14.902.89746214.863.197362
BEVbert [1]14.552.817564-3.137362
Ours14.832.80766515.232.747666
", + "bbox": [ + 509, + 123, + 864, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on R2R [3] (more details in §4.1).", + "bbox": [ + 506, + 369, + 849, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "at each mini-batch with the same sampling ratio.", + "bbox": [ + 500, + 402, + 792, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finetuning. Following the standard protocol [1, 16, 94], we finetune the navigation network using Dagger [69] techniques. In addition, the OG loss [1, 16, 53] is employed on REVERIE. In this stage, we set the learning rate to 1e-5 and batch size to 8 with $20k$ iterations.", + "bbox": [ + 498, + 417, + 857, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. During the testing phase, the agent receives the multi-view images and encodes them as VERs through the frozen environment encoder (§3.1). Based on VERs, the agent performs volume state estimation (§3.2) and models episodic memory (§3.3). By combining both of them, the agent predicts the next step accurately until stops.", + "bbox": [ + 498, + 486, + 857, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reproducibility. Our model is implemented in PyTorch and trained on eight RTX 4090 GPUs with a 24GB memory per-card. Testing is conducted on the same machine.", + "bbox": [ + 498, + 570, + 857, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 628, + 612, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Performance on VLN", + "text_level": 1, + "bbox": [ + 500, + 651, + 684, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The experiments are conducted on three datasets. R2R [3] contains 7,189 trajectories sampled from 90 real-world indoor scenes. It consists of $22k$ human-annotated navigational instructions. The dataset is split into train, val seen, val unseen, and test unseen sets, which mainly focus on the generalization capability in unseen environments. REVERIE [64] contains high-level instructions which describe target locations and objects, with a focus on grounding remote target objects. R4R [39] is an extended version of R2R with longer trajectories.", + "bbox": [ + 498, + 672, + 857, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For R2R, Success Rate (SR), Trajectory Length (TL), Oracle Success Rate (OSR), Success rate weighted by Path Length (SPL), and Navigation Error (NE) are used. For REVERIE, Remote Grounding Success rate", + "bbox": [ + 498, + 811, + 857, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16322", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f11c19036038233cf3e877dad30380869f383a61e265d7cbc4930195e4ad2217.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelsval seenREVERIEtest unseen
TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑
RCM [88]10.7029.4423.3321.8216.2315.3611.9814.239.296.974.893.8910.6011.687.846.673.673.14
FAST-M [64]16.3555.1750.5345.5031.9729.6645.2828.2014.407.197.844.6739.0530.6319.8811.6111.286.08
SIA [53]13.6165.8561.9157.0845.9642.6541.5344.6731.5316.2822.4111.5648.6144.5630.8014.8519.029.20
RecBERT [33]13.4453.9051.7947.9638.2335.6116.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [28]15.1648.9847.0142.3432.7530.0118.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [14]12.7947.6543.2940.1927.2025.1814.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
HOP [65]13.8054.8853.7647.1938.6533.8516.4636.2431.7826.1118.8515.7316.3833.0630.1724.3417.6914.34
DUET [16]13.8673.8671.7563.9457.4151.1422.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
TD-STP [100]-------39.4834.8827.3221.1616.56-40.2635.8927.5119.8815.40
BEVBert [1]-76.1873.7265.3257.7051.73-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
GridMM [94]------23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
LANA [89]15.9174.2871.9462.7759.0250.3423.1852.9748.3133.8632.8622.7718.8357.2051.7236.4532.9522.85
BSG [55]15.2678.3676.1866.6961.5654.0224.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
Ours16.1380.4975.8366.1961.7156.2023.0361.0955.9839.6633.7123.7024.7462.2256.8238.7633.8823.19
", + "bbox": [ + 122, + 111, + 856, + 313 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/085610d1312f93098f13f8f040e3e6a93467224efc066f41a842d950bb5fab36.jpg", + "table_caption": [ + "Table 2. Quantitative comparison results on REVERIE [64]. ' -': unavailable statistics (see §4.1 for more details)." + ], + "table_footnote": [], + "table_body": "
ModelsR4R val unseen
NE↓SR↑CLS↑nDTW↑SDTW↑
SF [3]8.472430--
RCM [88]-29353013
EGP [19]8.0030443718
SSM [82]8.2732533919
RelGraph [32]7.4336414734
RecBERT [33]6.6744514530
HAMT [14]6.0945585032
BSG [55]6.1247595334
LANA [89]-43605232
Ours6.1047615433
", + "bbox": [ + 124, + 345, + 478, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(RGS), and Remote Grounding Success weighted by Path Length (RGSPL) are also employed for object grounding. For R4R, Coverage weighted by Length Score (CLS), normalized Dynamic Time Warping (nDTW), and Success rate weighted nDTW (SDTW) are used.", + "bbox": [ + 111, + 534, + 470, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on R2R. Table 1 compares our model with the state-of-the-art models on R2R. As we find that our model yields SR of $76\\%$ and SPL of $66\\%$ on text unseen, which leads to promising gains of $3\\%$ and $4\\%$ over BEVBert [1], respectively. It verifies that using VER to represent the environment leads to better decision-making.", + "bbox": [ + 111, + 603, + 470, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on REVERIE. Table 2 presents the comparison results on REVERIE. Compared with the recent state-of-the-art VLN agent [55], our agent improves SR and SPL by $3.86\\%$ and $4.07\\%$ on the val unseen split. This highlights the effectiveness of our architecture design.", + "bbox": [ + 111, + 686, + 470, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on R4R. Table 3 shows results on R4R. Our approach outperforms others in most metrics with a promising gain on nDTW (i.e., $1\\%$ ). This suggests the episodic memory module is able to retrieve the long-time context.", + "bbox": [ + 111, + 755, + 470, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Results. Fig. 4 depicts one exemplar navigation episode from val unseen set of R2R. In this complex environment, there are many rooms with different objects and 3D layout. From the visualization of 3D occupancy pre", + "bbox": [ + 111, + 811, + 470, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/63cdb91647543ea4a4838fa9410ca0145aaec1bdcd27109dc61431466ccddac2.jpg", + "table_caption": [ + "Table 3. Quantitative results on R4R [39] (more details in §4.1)." + ], + "table_footnote": [], + "table_body": "
ModelsREVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
w/o Volume State52.3134.9132.7572.7161.13
w/o Episodic Memory49.3333.7131.3668.2161.70
Full model55.9839.6633.7175.8065.37
", + "bbox": [ + 509, + 347, + 864, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of overall design on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details).", + "bbox": [ + 498, + 424, + 859, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dition at the key steps, we find the geometric details and semantics can be captured well by VER. The room layout estimation can help the agent to understand \"enter the bedroom\". Finally, our agent finds the \"bed\" and accomplishes the instruction successfully.", + "bbox": [ + 496, + 462, + 857, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Diagnostic Experiment", + "text_level": 1, + "bbox": [ + 500, + 540, + 694, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To thoroughly test the efficacy of crucial components of our model, we conduct a series of diagnostic studies on val unseen split of REVERIE and R2R.", + "bbox": [ + 496, + 561, + 857, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall Design. We first investigate the effectiveness of our core design. The results in Table 4 indicate that adding Volume State leads to a substantial performance gain (i.e., $3.67\\%$ on SR). After using Episodic Memory, a higher score (i.e., $31.36\\% \\rightarrow 33.71\\%$ on RGS) is achieved.", + "bbox": [ + 496, + 602, + 857, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Neighborhood Range of Viewpoints. We use the neighborhood of each viewpoint in the state space for local action prediction (Eq. 6). Moreover, corresponding pillar representations of the neighborhood are also used for node embeddings of the episodic memory (Eq. 7). From Table 5, the limited range of neighborhood is insufficient to represent the candidate viewpoint for navigation (e.g., $75.80\\% \\rightarrow 73.75\\%$ of SR on R2R). However, too large neighborhood range will contain irrelevant information, leading to inferior performance (e.g., $55.98\\% \\rightarrow 53.49\\%$ of SR on REVERIE).", + "bbox": [ + 496, + 671, + 857, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis on 3D Representation Learning", + "text_level": 1, + "bbox": [ + 498, + 816, + 816, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metric. Following standard protocols, we employ Intersection over Union (IoU) to evaluate the occu", + "bbox": [ + 498, + 838, + 857, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16323", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4a6f24632285fe7044140dcc4020ce6d26b47870a7018f11edebb1103f923dc.jpg", + "image_caption": [], + "image_footnote": [ + "Turn around and go passed the couch. Cross the hall to the sitting room. Turn left and enter the bedroom on the left. Wait near the bed." + ], + "bbox": [ + 124, + 112, + 401, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/195d147b8ba1ef9d7fad116043938d71e4705a12a4934b6e382c6921c4aede83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 112, + 550, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa1038e304bdc59c608be196534f17d6bd8b7498e3becb779165840da106f647.jpg", + "image_caption": [ + "Figure 4. A representative visual result on val unseen of R2R [3]. We first visualize the 3D occupancy prediction at the key steps. In addition, we provide the prediction of 3D boxes and 3D room layout at step $⑦$ . We find that VER can capture the geometric details of 'couch' and the structure of 'bedroom'. With VER, our agent easily finds the 'bed' and succeeds. See §4.1 for more details." + ], + "image_footnote": [], + "bbox": [ + 406, + 194, + 550, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5908d47a2d36734e0b9fd62d5d03fcbeb4b559cc792b31156a350c86376d249.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 112, + 694, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/899ef8050b2b013484e8c9e41202b1035aeae1c27a6cc31ba4154c0f4eee0efa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 194, + 696, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcba29e178db031f005031f1b115b7228239839994a0ce5b1a12bf0ff4084555.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 112, + 851, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b0e0ad220f3711e9200bc4cddcba689c5484c9818dd6876b84a1216d2c548aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 194, + 851, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3b9c8bc5efdcb853e14c3e1407042d21422ed9cc31487973493b9e62f359b4e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
|Ωn|REVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
453.3035.9034.1673.7562.77
955.9839.6633.7175.8065.37
1652.1836.3733.1473.8263.49
2553.4935.8733.4674.6363.24
", + "bbox": [ + 122, + 316, + 470, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/23fe3b696faea661afd5fa067f2c9190a3274320101deadcfb4f0f7765114069.jpg", + "table_caption": [ + "Table 5. Ablation study of neighborhood range on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details)." + ], + "table_footnote": [], + "table_body": "
ModelsOccupancyDetectionLayout
IoU↑mIoU↑mAP↑mAR↑3D IoU↑
BEVFormer [51]20.388.9727.3043.8862.71
OccNet [78]22.1210.6629.9147.1564.04
Ours24.3112.9333.5751.6066.45
", + "bbox": [ + 122, + 443, + 475, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pancy prediction quality, regardless of the semantic labels. The mean IoU (mIoU) of 15 classes is also used to assess the performance of semantic occupancy. For 3D object detection, we utilize mean Average Precision (mAP) and mean Average Recall (mAR) with IoU thresholds of 0.50. For room layout, we adopt 3D IoU for cuboid layout evaluation.", + "bbox": [ + 111, + 559, + 470, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on 3D Tasks. In Table 6, our network (§3.1) outperforms other methods [51, 78] by a significant margin (2.19% on IoU of occupancy, 3.66% on mAP of 3D detection, and 2.41% on 3D IoU of room layout estimation). The mIoU of occupancy is also exhibits improvement (2.27%), underscoring the network's proficiency in capturing both scene geometry and fine-grained semantics.", + "bbox": [ + 111, + 643, + 470, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Coarse-to-Fine Extraction. Table 7 lists the scores with different up-sampling operations (Eq. 2). Our approach improves the performance by solid margins (e.g., $11.03\\% \\rightarrow 12.93\\%$ for 3D occupancy, $75.14\\% \\rightarrow 75.80\\%$ on SR of R2R). This verifies the efficacy of our design of the coarse-to-fine extraction and learnable up-sampling operations.", + "bbox": [ + 111, + 739, + 470, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-task Learning. Table 8 reports performance comparison with different perception tasks (§3.1). We find that multi-task learning yields a substantial performance gain.", + "bbox": [ + 111, + 824, + 472, + 867 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f886aab8ce16523699c8a8f6697e0c4281d563142c1a38784a74ee63a631ce54.jpg", + "table_caption": [ + "Table 6. Quantitative results on 3D occupancy, 3D detection, and room layout prediction (see §4.3 for more details)." + ], + "table_footnote": [], + "table_body": "
Up-Sampling3D PerceptionR2R
mIoU↑mAP↑3D IoU↑SR↑SPL↑
w/o Coarse-to-Fine12.3932.9566.57--
Trilinear Interpolation11.0329.4263.4575.1464.30
3D Deconvolution12.9333.5766.4575.8065.37
", + "bbox": [ + 509, + 316, + 867, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/799d26cf106cf3aeffcc3c4c19e35236296554a5c2fce8e74b41f2345086d628.jpg", + "table_caption": [ + "Table 7. Ablation study of Coarse-to-Fine Extraction on occupancy prediction (mIoU), 3D detection (mAP), room layout (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details)." + ], + "table_footnote": [], + "table_body": "
Multi-task Learning3D PerceptionR2R
Occ.Obj.Room.mIoU↑mAP↑3D IoU↑SR↑SPL↑
✓12.09--74.9063.82
✓✓12.1432.64-75.2164.79
✓✓-33.1164.5874.0363.51
✓✓11.37-63.2974.9764.66
✓✓✓12.9333.5766.4575.8065.37
", + "bbox": [ + 509, + 431, + 862, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study of Multi-task Learning on occupancy prediction (mIoU), 3D detection (mAP), room layout estimation (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details).", + "bbox": [ + 498, + 531, + 857, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This suggests these 3D perception tasks are complementary to each other in capturing geometric and semantic properties of scenes, further facilitating the decision-making.", + "bbox": [ + 498, + 590, + 857, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 607, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a Volumetric Environment Representation (VER), which aggregates the perspective features into structured 3D cells. Through coarse-to-fine feature extraction, we can efficiently perform several 3D perception tasks. Based on this comprehensive representation, we develop the volume state for local action prediction and the episodic memory for retrieving the global context. We demonstrate that our agent achieves promising performance on VLN benchmarks (R2R, REVERIE, and R4R).", + "bbox": [ + 496, + 671, + 857, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (No. 62372405), the Fundamental Research Funds for the Central Universities (No. 226-2022-00051), and CCF-Tencent Open Fund.", + "bbox": [ + 496, + 797, + 857, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16324", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 125, + 200, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. In ICCV, 2023. 1, 2, 4, 6, 7", + "[2] Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv:2304.03047, 2023. 1, 2", + "[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, 2018. 1, 2, 3, 6, 7, 8", + "[4] Peter Anderson, Ayush Shrivastava, Devi Parikh, Dhruv Batra, and Stefan Lee. Chasing ghosts: Instruction following as bayesian state tracking. In NeurIPS, 2019. 1, 2, 4", + "[5] Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuzhhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt4.arXiv:2303.12712, 2023.2", + "[6] Vincent Cartillier, Zhile Ren, Neha Jain, Stefan Lee, Irfan Essa, and Dhruv Batra. Semantic mapnet: Building allocentric semantic maps and representations from egocentric views. In AAAI, 2021. 2", + "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 5", + "[8] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In ICLR, 2019. 2", + "[9] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In CVPR, 2020. 2", + "[10] Jinyu Chen, Wenguan Wang, Si Liu, Hongsheng Li, and Yi Yang. Omnidirectional information gathering for knowledge transfer-based audio-visual navigation. In ICCV, 2023. 1", + "[11] Kevin Chen, Junshen K Chen, Jo Chuang, Marynel Vázquez, and Silvio Savarese. Topological planning with transformers for vision-and-language navigation. In CVPR, 2021. 1, 2", + "[12] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurIPS, 2022. 1, 2", + "[13] Peihao Chen, Xinyu Sun, Hongyan Zhi, Runhao Zeng, Thomas H Li, Gaowen Liu, Mingkui Tan, and Chuang Gan. $a^2$ nav: Action-aware zero-shot robot navigation by exploiting vision-and-language ability of foundation models. In NeurIPS Workshop, 2023. 2", + "[14] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 6, 7" + ], + "bbox": [ + 120, + 147, + 472, + 864 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 2", + "[16] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 1, 2, 6, 7", + "[17] Ian Cherabier, Johannes L Schonberger, Martin R Oswald, Marc Pollefeys, and Andreas Geiger. Learning priors for semantic 3d reconstruction. In ECCV, 2018. 4", + "[18] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In ICCV, 1999. 5", + "[19] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In NeurIPS, 2020. 1, 2, 7", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6", + "[21] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2", + "[22] Alberto Elfes. Occupancy grids: A stochastic spatial representation for active robot perception. arXiv:1304.1098, 2013. 2", + "[23] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, 2018. 1, 2, 6", + "[24] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, 2023. 2", + "[25] Sourav Garg, Niko Sunderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 2", + "[26] Georgios Georgakis, Karl Schmeckpeper, Karan Wanchoo, Soham Dan, Eleni Miltsakaki, Dan Roth, and Kostas Dani-ilidis. Cross-modal map learning for vision and language navigation. In CVPR, 2022. 1, 2", + "[27] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In ACL, 2022. 2", + "[28] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, 2021. 2, 7", + "[29] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In NeurIPS, 2018. 2" + ], + "bbox": [ + 508, + 127, + 856, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16325", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, 2020. 6", + "[31] Joao F Henriques and Andrea Vedaldi. Mapnet: An allocentric spatial memory for mapping environments. In CVPR, 2018. 2", + "[32] Yicong Hong, Cristian Rodriguez, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 1, 2, 7", + "[33] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In CVPR, 2021. 2, 6, 7", + "[34] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In CVPR, 2023. 1, 2", + "[35] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 5", + "[36] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, 2023. 3", + "[37] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In CVPR, 2023. 2", + "[38] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, 2022. 1, 2", + "[39] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, 2019. 2, 6, 7", + "[40] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 3", + "[41] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In CVPR, 2023. 2", + "[42] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 6", + "[43] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[44] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2016. 1", + "[45] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, 2021. 2" + ], + "bbox": [ + 120, + 126, + 472, + 864 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Obin Kwon, Jeongho Park, and Songhwai Oh. Rendering neural radiance map for visual navigation. In CVPR, 2023. 2", + "[47] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE TPAMI, 2023. 2, 3", + "[48] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurIPS, 2023. 2", + "[49] Jialu Li, Hao Tan, and Mohit Bansal. Envide: Environment editing for vision-and-language navigation. In CVPR, 2022. 2", + "[50] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In CoRL, 2022. 2", + "[51] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 2, 3, 4, 8", + "[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 4, 6", + "[53] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In CVPR, 2021. 6, 7", + "[54] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, 2021. 2", + "[55] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, 2023. 1, 2, 6, 7", + "[56] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS, 2018. 4", + "[57] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zero-shot video grounding with pseudo query lookup and verification. IEEE TIP, 33:1643-1654, 2024. 2", + "[58] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, 2020. 2", + "[59] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[60] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. In NeurIPS, 2021. 6", + "[61] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In ICLR, 2018. 1" + ], + "bbox": [ + 509, + 126, + 856, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16326", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, 2021. 1", + "[63] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3", + "[64] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR, 2020. 1, 2, 6, 7, 8", + "[65] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop: history-and-order aware pretraining for vision-and-language navigation. In CVPR, 2022. 2, 6, 7", + "[66] Ruijie Quan, Linchao Zhu, Yu Wu, and Yi Yang. Holistic lstm for pedestrian trajectory prediction. IEEE TIP, 30: 3229-3239, 2021. 2", + "[67] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4", + "[68] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. IJCV, 130(8):1978-2005, 2022. 2, 3, 4", + "[69] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 6", + "[70] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2, 3, 4", + "[71] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 2", + "[72] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In NeurIPS, 2014. 1, 2", + "[73] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2, 6", + "[74] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In NAACL, 2019. 2, 6", + "[75] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Self-supervised 3d semantic representation learning for vision-and-language navigation. arXiv:2201.10788, 2022. 1, 2", + "[76] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 2, 4", + "[77] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv:2304.14365, 2023. 2, 4, 5", + "[78] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 3, 8" + ], + "bbox": [ + 120, + 126, + 470, + 866 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2", + "[80] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In CVPR, 2020. 2", + "[81] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, 2020. 6", + "[82] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, 2021. 1, 2, 6, 7", + "[83] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. In NeurIPS, 2022. 2", + "[84] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, 2022. 2, 6", + "[85] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, 2023. 2", + "[86] Hanqing Wang, Wenguan Wang, Wei Liang, Steven CH Hoi, Jianbing Shen, and Luc Van Gool. Active perception for visual-language navigation. IJCV, 131(3):607-625, 2023. 2", + "[87] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 3", + "[88] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 2, 7", + "[89] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023. 6, 7", + "[90] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Learning to follow and generate instructions for language-capable navigation. IEEE TPAMI, 2023. 2", + "[91] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, 2020. 5", + "[92] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In CVPR, 2024. 4", + "[93] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV, 2023. 2", + "[94] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In ICCV, 2023. 1, 2, 6, 7" + ], + "bbox": [ + 509, + 126, + 857, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16327", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, 2023. 4, 5", + "[96] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, 2021. 2", + "[97] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv:2307.13510, 2023. 4", + "[98] Zhenjia Xu, Zhanpeng He, Jiajun Wu, and Shuran Song. Learning 3d dynamic scene representations for robot manipulation. In CoRL, 2020. 3, 5, 6", + "[99] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, 2018. 4", + "[100] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In ACM MM, 2022. 2, 6, 7", + "[101] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 2", + "[102] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv: 1904.07850, 2019. 5", + "[103] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, 2020. 6", + "[104] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 4", + "[105] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In CVPR, 2018. 5", + "[106] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 2, 3, 5" + ], + "bbox": [ + 114, + 126, + 470, + 657 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16328", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_model.json b/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b20ffd2ff72c9ddf9b3b370a12089b09d8fd36f2 --- /dev/null +++ b/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_model.json @@ -0,0 +1,3293 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.26, + 0.044, + 0.782, + 0.058 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.339, + 0.058, + 0.705, + 0.071 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.31, + 0.072, + 0.733, + 0.085 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.162, + 0.823, + 0.183 + ], + "angle": 0, + "content": "Volumetric Environment Representation for Vision-Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.209, + 0.63, + 0.225 + ], + "angle": 0, + "content": "Rui Liu Wenguan Wang Yi Yang*" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.227, + 0.585, + 0.241 + ], + "angle": 0, + "content": "ReLER, CCAI, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.243, + 0.636, + 0.256 + ], + "angle": 0, + "content": "https://github.com/DefaultRui/VLN-VER" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.289, + 0.33, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.318, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Vision-language navigation (VLN) requires an agent to navigate through an 3D environment based on visual observations and natural language instructions. It is clear that the pivotal factor for successful navigation lies in the comprehensive scene understanding. Previous VLN agents employ monocular frameworks to extract 2D features of perspective views directly. Though straightforward, they struggle for capturing 3D geometry and semantics, leading to a partial and incomplete environment representation. To achieve a comprehensive 3D representation with fine-grained details, we introduce a Volumetric Environment Representation (VER), which voxelizes the physical world into structured 3D cells. For each cell, VER aggregates multi-view 2D features into such a unified 3D space via 2D-3D sampling. Through coarse-to-fine feature extraction and multi-task learning for VER, our agent predicts 3D occupancy, 3D room layout, and 3D bounding boxes jointly. Based on online collected VERs, our agent performs volume state estimation and builds episodic memory for predicting the next step. Experimental results show our environment representations from multi-task learning lead to evident performance gains on VLN. Our model achieves state-of-the-art performance across VLN benchmarks (R2R, REVERIE, and R4R)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.679, + 0.234, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.703, + 0.475, + 0.786 + ], + "angle": 0, + "content": "Vision-language navigation (VLN) requires an agent to navigate in a 3D environment following natural language instructions [3, 64]. As a holistic understanding of the environment plays a pivotal role in decision-making within VLN, environment representation learning serves as a foundation for formulating accurate navigation policies." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.787, + 0.472, + 0.843 + ], + "angle": 0, + "content": "Early VLN approaches [3, 23] typically learn the navigation policy through the sequence-to-sequence (Seq2Seq) framework [72], which directly maps instructions and multi-view perspective observations to actions. They sim" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.287, + 0.646, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.39, + 0.573, + 0.4 + ], + "angle": 0, + "content": "Our Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.289, + 0.742, + 0.3 + ], + "angle": 0, + "content": "Previous Methods" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.301, + 0.737, + 0.31 + ], + "angle": 0, + "content": "Perspective" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.31, + 0.737, + 0.32 + ], + "angle": 0, + "content": "Observations" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.32, + 0.757, + 0.33 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.33, + 0.757, + 0.34 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.34, + 0.757, + 0.35 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.35, + 0.757, + 0.36 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.36, + 0.737, + 0.37 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.37, + 0.737, + 0.38 + ], + "angle": 0, + "content": "2D Features of" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.38, + 0.737, + 0.389 + ], + "angle": 0, + "content": "Candidate View" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.301, + 0.852, + 0.33 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.33, + 0.852, + 0.34 + ], + "angle": 0, + "content": "Topological Graph" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.34, + 0.852, + 0.35 + ], + "angle": 0, + "content": "\\( \\rightarrow \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.35, + 0.852, + 0.36 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.36, + 0.852, + 0.37 + ], + "angle": 0, + "content": "\\( \\frac{1}{2}x - 1 > 0 \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.37, + 0.852, + 0.38 + ], + "angle": 0, + "content": "Semantic Map" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.38, + 0.852, + 0.389 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.39, + 0.573, + 0.4 + ], + "angle": 0, + "content": "Our Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.401, + 0.581, + 0.41 + ], + "angle": 0, + "content": "2D-3D Sampling" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.41, + 0.581, + 0.42 + ], + "angle": 0, + "content": "23. 15 sampling" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.42, + 0.581, + 0.43 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.43, + 0.581, + 0.44 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.44, + 0.581, + 0.45 + ], + "angle": 0, + "content": "multi-view" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.45, + 0.581, + 0.456 + ], + "angle": 0, + "content": "D Features" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.39, + 0.75, + 0.4 + ], + "angle": 0, + "content": "\\( \\therefore m - 1 \\neq 0 \\) ;" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.4, + 0.75, + 0.41 + ], + "angle": 0, + "content": "3D Occupancy" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.41, + 0.75, + 0.42 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.42, + 0.75, + 0.43 + ], + "angle": 0, + "content": "3D Detection -" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.43, + 0.747, + 0.44 + ], + "angle": 0, + "content": "\\( \\therefore m : x = 1 \\) 或 \\( {3x} + {4y} + 1 = 0 \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.44, + 0.747, + 0.45 + ], + "angle": 0, + "content": "Room Layout" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.45, + 0.747, + 0.456 + ], + "angle": 0, + "content": "Estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.39, + 0.852, + 0.44 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.44, + 0.852, + 0.45 + ], + "angle": 0, + "content": "Volume State Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.45, + 0.852, + 0.456 + ], + "angle": 0, + "content": "Volume State Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.466, + 0.86, + 0.556 + ], + "angle": 0, + "content": "Figure 1. The agent observes its surroundings with corresponding perspective features of different candidate views (O). Previous methods construct the topological graph or semantic map based on these 2D features. Our VER aggregates the multi-view features into structured 3D cells via 2D-3D sampling. VER is a powerful representation for both 3D perception tasks and VLN, providing a volume state space for decision-making." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.86, + 0.868 + ], + "angle": 0, + "content": "ply embed their immediate observation of local environment into the hidden states of recurrent units. As a result, they lack of explicit environment representations and struggle to access their past states during long-time exploration [61, 82]. To address this issue, later VLN agents are equipped with an external memory module [61], which stores the environment representations distinctly from navigation states. In this way, they can explicitly model and maintain the environment layouts and contents in a form of topological graph [2, 11, 16, 19, 32, 62] or semantic map [1, 4, 10, 12, 26, 34, 38, 55, 75, 94] (Fig. 1). Despite their promising performance with advanced frameworks (e.g., graph neural network [44] and Transformer [79]), their environment representations are still built upon 2D perspective features extracted by monocular frameworks. While straightforward, they compress depth information onto the perspective plane, sacrificing the integral scene structure in the 3D space. Thus, they encounter challenges in capturing 3D geometry and semantics in complex scenes. Such an incomplete environment representation easily leads to suboptimal navigation decisions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.13, + 0.854, + 0.293, + 0.867 + ], + "angle": 0, + "content": "*Corresponding author: Yi Yang." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.127, + 0.473, + 0.43 + ], + "angle": 0, + "content": "In this article, we propose a Volumetric Environment Representation (VER) that quantizes the physical world into structured 3D cells (Fig. 1). These cells, arranged within a predefined volumetric space, maintain both height and depth dimensions. Each cell corresponds to local context of the 3D space. VER aggregates multi-perspective 2D features within these cells through an environment encoder (§3.1). Compared to previous partial representations derived from hidden states and external memory, our VER captures the full geometry and semantics of the physical world. These 3D cells store the properties of the corresponding space in the scene by predicting 3D occupancy [70, 77], room layout [106], and 3D object boxes [51]. However, directly reconstructing the high-quality VER from 2D perspective views is challenging to capture the fine-grained details. As a response, we propose a coarse-to-fine VER extraction architecture, which uses learnable up-sampling operations to construct the representations progressively. It is supervised by multi-resolution semantic labels at different scales, utilizing the coarse-to-fine representations as hierarchical inputs. The annotations of the 3D tasks are collected for multi-task learning (§3.4)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.431, + 0.473, + 0.692 + ], + "angle": 0, + "content": "At each navigation step, our agent initially encodes the multi-view observations into VER (§3.2). With VER, instructions can be more effectively grounded in the 3D context. This is achieved by establishing cross-modal correlations between linguistic words and 3D cells of VER. Based on the correlations, a volume state estimation module is proposed to calculate transition probabilities over the surrounding cells. With the help of this module, our agent performs comprehensive decision-making in volumetric space, and then maps the volume state into local action space. In addition, an episodic memory module is established to online collect the information of observed viewpoints and build a topological graph providing global action space (§3.3). The node embeddings in the graph are from neighbor pillar representations in VER corresponding to the respective viewpoints. To balance the long-range action reasoning and language grounding, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.693, + 0.472, + 0.803 + ], + "angle": 0, + "content": "Our agent is evaluated on three VLN benchmarks, i.e., R2R [3], REVERIE [64], and R4R [39] (\\(\\S 4.1\\)). It yields solid performance gains (about \\(3\\%\\) SR and \\(4\\%\\) SPL on R2R test, \\(4\\%\\) SR and \\(4\\%\\) SPL on REVERIE val unseen). The ablation study confirms the efficacy of core model designs (\\(\\S 4.2\\)). Additional results show our model achieves promising performance in 3D occupancy prediction, 3D detection, and room layout estimation (\\(\\S 4.3\\))." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.816, + 0.245, + 0.83 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.839, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Vision-Language Navigation (VLN). Early VLN agents [3, 23] are built upon Seq2Seq [66, 72] framework to re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.86, + 0.486 + ], + "angle": 0, + "content": "serve the observation history in hidden state. Thus they struggle to capture long-range context as the path length increases. Later efforts are devoted to multimodal representation learning, navigation strategy learning, and data generation [27]. As a primary step, multimodal representation learning helps agents understand the environments and establish relations between the instructions and visual observations. Inspired by the success of vision-language pretraining [58, 65, 73], recent approaches [14, 33] use transformer-based architectures [42, 79] for joint visual and textual representations. Some attempts further exploit the visual information by modeling semantic relation [37] and spatial information [11, 19, 32, 83, 86, 100]. For navigation strategy learning, many VLN models [24, 88] use imitation and reinforcement learning-based training strategies. Previous solutions [45, 85] introduce world models [29] to perform mental simulations and make mental planning. Furthermore, the scarcity of human instructions and limited diversity of the scene hinder the agent to learn navigation policy and generalize to unseen environments well [54]. Therefore, several VLN data generation strategies have been proposed to create new trajectories from existing datasets [15, 28, 93], generate more instructions [41, 74, 84, 90], or create synthetic environments [48, 49]. In addition, driven by large models [5], existing agents [13, 57, 71, 101] demonstrate promising zero-shot performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.86, + 0.601 + ], + "angle": 0, + "content": "Despite their outstanding contributions, most of them rely on 2D visual cues in perspective observations. These representations are constrained by occlusion and limited geometric information, especially in complex scenes. In this paper, we propose VER, a unified environment representation learned by 3D perception tasks. During navigation, our agent performs volume state estimation on VER, facilitating comprehensive decision-making within the 3D space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.861, + 0.868 + ], + "angle": 0, + "content": "Environment Representation. Existing VLN models introduce various representations for environment modeling, including topological graphs [2, 11, 16, 19, 32, 82], and semantic spatial representations [1, 4, 12, 26, 34, 38, 55, 75, 94]. With a broader view, diverse representations have been proposed for robotics and autonomous driving [25, 47, 68, 76, 96]. In the early stage, 2D occupancy grid maps [22] model occupied and free space in the surroundings based on Bayesian estimation for robot navigation. Classic SLAM systems [21] construct a map directly by integrating information from various sensors, including LiDAR and cameras. However, the representations in SLAM still rely on primitives such as 3D point clouds and image patches. Some efforts [6, 8, 31] focus on developing learnable semantic map representations. To enhance spatial reasoning, scene graph representations [9, 80] define the topological relations between spatial elements of the environment. In addition, neural scene representations [46, 50] embed image observations into latent codes" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16318" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.112, + 0.852, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.252, + 0.862, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overview of our model. Given the perspective features of candidate views, a group of 3D queries are used to sample and aggregate them into VER (§3.1). To encode VER, we adopt coarse-to-fine extraction and perform multi-task learning on 3D perception. Based on VER, a volume state estimation module is proposed to predict state transition (§3.2). The episodic memory is used to store past observations using neighboring pillar representations for each viewpoint (§3.3). For decision-making, our agent combines both the local action probabilities from the volume state and the global action probabilities obtained from the episodic memory. See §3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.339, + 0.473, + 0.436 + ], + "angle": 0, + "content": "for object categories, showcasing scalability to large scenes [40, 59]. Vision-centric BEV perception, which transforms perspective-view inputs to BEV grid representations, has recently received increasing attention [47, 51, 63]. As the BEV representations simplify the vertical geometry, 3D occupancy prediction [36, 78] is further proposed to infer the 3D geometry from perspective images [68, 70, 87, 98]." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.438, + 0.474, + 0.576 + ], + "angle": 0, + "content": "However, existing VLN agents mainly employ abstract relations or compressed spatial maps, lacking the ability to access complete scene information. Providing more world context can be beneficial for the subsequent decision-making and policy learning. Motivated by this insight, we explore a holistic environment representation VER that voxelizes the world into structured 3D cells. VER captures both semantic information and geometric details of the whole scene. Building upon VER, our agent is able to predict the 3D occupancy, room layout, and 3D boxes accurately." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.606, + 0.214, + 0.621 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.63, + 0.474, + 0.867 + ], + "angle": 0, + "content": "Problem Formulation. For brevity, we present the technical description in the context of R2R [3]. The navigable area of the environment is organized as an undirected graph, containing a set of nodes (viewpoints) and connectivity edges. In R2R, an embodied agent needs to navigate to a target location in the 3D environment following human instructions with \\( L \\) words (embedded as \\( E \\in \\mathbb{R}^{D_w \\times L} \\), where \\( D_w \\) is the channel dimension). At time step \\( t \\), the agent looks around and obtains multi-view observations of its surrounding scene from the current location. Each view is represented by a 2D visual feature \\( F_t^{2d} \\in \\mathbb{R}^{D_i \\times H \\times W} \\), where \\( H \\) and \\( W \\) are the spatial shape of image plane, \\( D_i \\) denotes the channel dimension. The local action space \\( A_t \\in \\mathbb{R}^{N_t + 1} \\) is defined by \\( N_t \\) candidate views, which correspond to neighboring navigable nodes \\( \\{v_{t,n}^*\\}_{n=1}^{N_t} \\), as well as a [STOP] action. Previous agents predict the action probabilities \\( p_t^{2d} \\in \\mathbb{R}^{N_t + 1} \\) directly based on \\( F_t^{2d} \\) of each candidate view. However," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.859, + 0.38 + ], + "angle": 0, + "content": "these 2D features with limited geometric information are partial representations of the 3D environment, easily leading to suboptimal decision making." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.86, + 0.657 + ], + "angle": 0, + "content": "Overview. To achieve comprehensive scene understanding, we introduce VER, which voxelizes the 3D world into structured 3D cells (Fig. 2). At step \\( t \\), an environment encoder is proposed to sample multi-view features \\( (F_{t}^{2\\mathrm{d}} \\) of each view) into the volumetric space of VER, forming a unified representation \\( F_{t}^{3\\mathrm{d}} \\in \\mathbb{R}^{D_{e} \\times X \\times Y \\times Z} \\) (\\$3.1). \\( X \\) and \\( Y \\) are the shapes of horizontal plane, \\( Z \\) reserves the height information of 3D space, and \\( D_{e} \\) represents the channel dimension. The volumetric space aligns with gravity in the world coordinate system based on the Manhattan assumption [106]. To encode VER, we devise coarse-to-fine extraction with multiple 3D perception tasks supervised by multi-resolution annotations (\\$3.4). Based on VER, a volume state estimation module is proposed to predict state transition probabilities \\( p_{t}^{3\\mathrm{d}} \\in \\mathbb{R}^{X \\times Y \\times Z} \\) over surrounding 3D cells (\\$3.2). With this module, our agent performs comprehensive decision-making in the 3D space and maps \\( p_{t}^{3\\mathrm{d}} \\) to \\( p_{t}^{2\\mathrm{d}} \\). To predict the next step, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory (\\$3.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.691, + 0.68 + ], + "angle": 0, + "content": "3.1.Environment Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.86, + 0.812 + ], + "angle": 0, + "content": "2D-3D Sampling. At step \\(t\\), the agent observes its surroundings and acquires the multi-view images. We introduce cross-view attention (CVA) to aggregate their features \\((F^{2d}\\) for each view) into a unified volumetric representation \\(F^{3d}\\) with a group of learnable volume queries \\(Q \\in \\mathbb{R}^{D_e \\times X \\times Y \\times Z}\\) (\\(t\\) is omitted for simplicity). Specifically, for the 3D cell positioned at \\((x,y,z)\\) within the egocentric world, a single query \\(Q(x,y,z) \\in \\mathbb{R}^{D_e}\\) is used to sample each image feature \\(F^{2d}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.816, + 0.858, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {3 d} (x, y, z) = \\operatorname {C V A} \\left(\\boldsymbol {Q} (x, y, z), \\boldsymbol {F} ^ {2 d} \\left(h ^ {\\prime}, w ^ {\\prime}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.859, + 0.868 + ], + "angle": 0, + "content": "where \\((h', w')\\) denotes the location of corresponding sampling point on the image plane. Note that we only show" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16319" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.125, + 0.468, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.268, + 0.473, + 0.32 + ], + "angle": 0, + "content": "Figure 3. Our coarse-to-fine VER representation extraction (§3.1) adopts cascade up-sampling operations with 3D deconvolutions (Eq. 2) and 3D queries (Eq. 1). The training process is supervised at different scales by multi-resolution semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.342, + 0.472, + 0.424 + ], + "angle": 0, + "content": "the formulation of a single sampling point for conciseness. Since the sampling strategy of vanilla cross-attention is computationally expensive, the deformable attention [51, 104] is introduced and extended in CVA. In this way, \\( Q(x,y,z) \\) selectively attends to a set of key sampling points around a reference instead of the entire \\( F^{2\\mathrm{d}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.425, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Coarse-to-Fine VER Extraction Architecture. Directly recovering the fine-grained VER from perspective features easily leads to performance and efficiency degradation [17, 95]. The coarse-to-fine extraction is proposed to reconstruct VER progressively. Our approach involves cascade up-sampling operations (Fig. 3), dividing this extraction into \\(M\\) levels. At each level, 3D deconvolutions are utilized for lifting spatial resolution, and then CVA is used to query the multi-view 2D features (Eq. 1) for refining the detailed geometry. This enables the direct learning of details and avoids the inaccuracy of interpolation [56, 77, 92] (see Table 7). Between the input coarse feature \\(F^{3\\mathrm{d}(0)} \\in \\mathbb{R}^{D_e \\times \\frac{X}{2^M} \\times \\frac{Y}{2^M} \\times \\frac{Z}{2^M}}\\) from Eq. (1) and the target fine feature \\(F^{3\\mathrm{d}(M)} \\in \\mathbb{R}^{D_e \\times X \\times Y \\times Z}\\), the intermediate features with varying shapes are calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.637, + 0.472, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {\\mathrm {3 d} (1)} = \\uparrow \\boldsymbol {F} ^ {\\mathrm {3 d} (0)}, \\dots , \\boldsymbol {F} ^ {\\mathrm {3 d} (M)} = \\uparrow \\boldsymbol {F} ^ {\\mathrm {3 d} (M - 1)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.658, + 0.473, + 0.867 + ], + "angle": 0, + "content": "where \\( F^{3\\mathrm{d}(1)}, \\dots, F^{3\\mathrm{d}(M-1)} \\) denote the intermediate features from different levels, and \\( \\uparrow \\) denotes the up-sampling. Multi-task Learning. Our VER offers a unified scene representation for various 3D perception tasks. Existing studies [68, 70] highlight that semantics and geometry are tightly intertwined. We train our environment encoder to extract VER under the supervision of multiple 3D perception tasks (§3.4). This process utilizes \\( \\{F^{3\\mathrm{d}(0)}, \\dots, F^{3\\mathrm{d}(M)}\\} \\) as inputs and is supervised at different scales (Fig. 3). For 3D occupancy prediction, the decoder is implemented as MLPs with the focal loss [52]. For 3D layout estimation, we adopt a query-based head to yield the Manhattan room layouts. A combination of the L1 loss and the IoU loss [67] is used as the training objective. For 3D detection, we employ a detection head to predict the 3D boxes [51]. The bipartite match" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.86, + 0.209 + ], + "angle": 0, + "content": "ing and the bounding box loss [51, 104] are employed for detection. A weight vector \\([2.0, 0.25, 0.25]\\) is used to balance the optimization of the three tasks, respectively. During navigation, the agent traverses between different viewpoints and encodes VERs through the frozen environment encoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.218, + 0.707, + 0.231 + ], + "angle": 0, + "content": "3.2. Volume State Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.859, + 0.349 + ], + "angle": 0, + "content": "VLN task is typically viewed as a state estimation and transition problem [76]. With our VER, the agent state is represented as \"volume state\". As such, the state transition within the locally observed 3D environment, computed by Eq. (3)&(4), is referred as \"volume state estimation\". Different from previous plane-level state models [1, 4], VER introduces additional height dimension for 3D state estimation. This enables a more accurate action prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.86, + 0.488 + ], + "angle": 0, + "content": "Volume State. At the beginning of a navigation episode, the agent is located at a start viewpoint \\((x_0,y_0,z_0)\\). Based on its perception range, a volume state space \\(\\mathcal{X}\\in \\mathbb{R}^{X\\times Y\\times Z}\\) is defined corresponding to the 3D physical world with an initial state \\(s_0 = (x_0,y_0,z_0)\\). At step \\(t\\), the next intermediate state \\(s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})\\) is determined by the instruction embeddings \\(\\pmb{E}\\) and VER \\(\\pmb{F}_t^{3\\mathrm{d}}\\) for reaching the goal state \\(s_T\\) \\((0 < t < T)\\). As the entire environment is partially observable, the current state transition \\((s_t\\rightarrow s_{t + 1}\\) in \\(\\mathcal{X}\\)) is regarded as a local consideration for action prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.859, + 0.572 + ], + "angle": 0, + "content": "State Estimation. A volume state estimation module is devised to predict the probability distribution \\( \\pmb{p}_t^{3\\mathrm{d}} \\in \\mathbb{R}^{X \\times Y \\times Z} \\) of the intermediate state \\( s_{t+1} \\) conditioned on \\( \\pmb{E} \\) and \\( \\pmb{F}_t^{3\\mathrm{d}} \\). The environment representation is first reshaped as \\( \\pmb{F}_t^{3\\mathrm{d}''} \\in \\mathbb{R}^{D_e \\times XYZ} \\), and then adopt multi-layer transformers (MLT) to model the relations between \\( \\pmb{E} \\) and \\( \\pmb{F}_t^{3\\mathrm{d}''} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.575, + 0.858, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {F}} _ {t} ^ {3 d} = \\mathrm {M L T} \\left(\\left[ \\boldsymbol {E}; \\boldsymbol {F} _ {t} ^ {3 d ^ {\\prime}} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times X Y Z}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.859, + 0.64 + ], + "angle": 0, + "content": "where \\(\\widetilde{F}_t^{3\\mathrm{d}}\\) is the updated representations, and \\([\\cdot ]\\) denotes the concatenation operation. MLT consists of stacked selfattention blocks. Then we use MLPs for state estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.643, + 0.858, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} ^ {\\mathrm {3 d}} = \\operatorname {S o f t m a x} \\left(\\mathrm {M L P} \\left(\\widetilde {\\boldsymbol {F}} _ {t} ^ {\\mathrm {3 d}}\\right)\\right) \\in [ 0, 1 ] ^ {X \\times Y \\times Z}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.859, + 0.76 + ], + "angle": 0, + "content": "Efficient Height-aware Group. The computational and memory efficiency of Eq. (3) is compromised due to the resolution of \\( F_{t}^{3\\mathrm{d}} \\) (\\( XYZ \\gg L \\)). Considering the similarity and sparsity of information along the height direction [97, 99], we partition \\( F_{t}^{3\\mathrm{d}} \\) into several uniform groups \\( \\{F_{t,z}^{g} \\in \\mathbb{R}^{D_e \\times XY}\\}_{z=1}^Z \\) along this axis. Then we apply MLT to each group, and Eq. (3) is reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.768, + 0.857, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {F}} _ {t, i} ^ {g} = \\operatorname {M L T} \\left(\\left[ \\boldsymbol {E}; \\mathbf {F} _ {t, i} ^ {g} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times X Y}, \\\\ \\widetilde {\\boldsymbol {F}} _ {t} ^ {3 \\mathrm {d}} = \\left\\{\\widetilde {\\boldsymbol {F}} _ {t, z} ^ {g} \\right\\} _ {z = 1} ^ {Z} \\in \\mathbb {R} ^ {D _ {e} \\times X Y Z}, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.859, + 0.868 + ], + "angle": 0, + "content": "where the weights of MLT are shared among different groups. The updated features from different groups \\(\\{\\widetilde{F}_{t,z}^{g}\\}_{z = 1}^{Z}\\) are aggregated along the height to leverage complementary information. For ease of notation, the symbol" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16320" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.125, + 0.471, + 0.154 + ], + "angle": 0, + "content": "\\(\\widetilde{F}_t^{3\\mathrm{d}}\\) is slightly reused for the gathered 3D representations. Then, \\(\\pmb{p}_t^{3\\mathrm{d}}\\) is calculated by Eq. (4)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.164, + 0.27, + 0.177 + ], + "angle": 0, + "content": "3.3. Action Prediction" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.185, + 0.472, + 0.323 + ], + "angle": 0, + "content": "For action prediction across the entire explored scene, a topological graph \\(\\mathcal{G}_t = \\{\\mathcal{V}_t,\\mathcal{E}_t\\}\\) is constructed and updated online to represent episodic memory during navigation. Specifically, \\(\\mathcal{V}_t\\) denotes the observed viewpoints, i.e., all visited viewpoints and their candidate viewpoints. These viewpoints are encoded by compressing previous VERs. The edge \\(\\mathcal{E}_t\\) denotes the navigable connections between these viewpoints. To predict the next step, our agent combines both the volume state estimation and episodic navigation memory for decision-making." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.324, + 0.474, + 0.435 + ], + "angle": 0, + "content": "Mapping Volume State to Action. As our agent navigates on the horizontal plane to reach the adjacent candidate viewpoints \\(\\{v_{n}^{*}\\}_{n = 1}^{N_{t}}\\), we map the volume state space into 2D space to align with this movement pattern. Specifically, we average \\(p_t^{3d}\\) along the height (z-axis) axis as \\(p_t^h\\in \\mathbb{R}^{X\\times Y}\\). Then we sum probability values in the neighborhood of \\(\\{v_{t,n}^{*}\\}_{n = 0}^{N_{t}}\\) (\\(v_{0}^{*}\\) for the current viewpoint, i.e., [STOP]), and normalize them as the local action probabilities:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.44, + 0.472, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} ^ {\\mathrm {2 d}} = \\left\\{\\sum \\boldsymbol {p} _ {t} ^ {h} \\left(x _ {n}, y _ {n}\\right) \\right\\} _ {n = 0} ^ {N _ {t}} \\in [ 0, 1 ] ^ {N _ {t} + 1}, \\left(x _ {n}, y _ {n}\\right) \\in \\Omega_ {n}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.467, + 0.472, + 0.522 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_t^h (x_n,y_n)\\) is the value at the coordinate \\((x_{n},y_{n})\\) and \\(\\Omega_{n}\\) is the neighborhood of \\(v_{n}^{*}\\) in the horizontal plane. In the training stage, a heatmap [102] with a Gaussian kernel is used to supervise this action prediction (§3.5)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.523, + 0.472, + 0.609 + ], + "angle": 0, + "content": "Global Action Prediction. The episodic memory module is used to store past environment representations and allows easy access to them. To memory the environment representations efficiently, we use the neighboring pillar [91] representations \\( \\pmb{F}_{t,n}^{p} \\in \\mathbb{R}^{D_{e} \\times |\\Omega_{n}| \\times Z} \\), corresponding to the current observed viewpoints \\( \\{v_{n}^{*}\\}_{n=0}^{N_{t}} \\) at step \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.611, + 0.472, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {t, n} ^ {P} = \\left\\{\\tilde {\\boldsymbol {F}} _ {t} ^ {3 d} \\left(x _ {n}, y _ {n}, z _ {n}\\right) \\right\\} _ {\\left(x _ {n}, y _ {n}\\right) \\in \\Omega_ {n}, z _ {n} \\in [ 1, Z ]} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.638, + 0.472, + 0.709 + ], + "angle": 0, + "content": "where \\(\\widetilde{F}_t^{3\\mathrm{d}}(x_n,y_n,z_n)\\) is the representation at position \\((x_{n},y_{n},z_{n})\\) of \\(\\widetilde{F}_t^{3\\mathrm{d}}\\) (Eq. 5). \\(\\{\\bar{F}_{t,n}^p\\in \\mathbb{R}^{D_e}\\}_{n = 0}^{N_t}\\) is obtained by average pooling as the corresponding node embeddings, which are then incorporated into \\(\\mathcal{G}_t\\). For previously observed nodes, we compute the average of their features." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.71, + 0.472, + 0.792 + ], + "angle": 0, + "content": "The episodic memory \\(\\mathcal{G}_t\\), which includes the observed viewpoints, offers a global action space \\(\\mathcal{A}_t^* \\in \\mathbb{R}^{|\\mathcal{V}_t|}\\). This enables our agent to change its current navigation state easily by 'jumping' directly to another viewpoint, which may be even observed several steps ago. The global action probabilities on \\(\\mathcal{G}_t\\) are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.797, + 0.472, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {G}} _ {t} = \\operatorname {M L T} \\left(\\left[ \\boldsymbol {E}; \\boldsymbol {G} _ {t} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times | \\mathcal {V} _ {t} |}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.818, + 0.401, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} ^ {g} = \\operatorname {S o f t m a x} \\left(\\operatorname {M L P} \\left(\\hat {\\boldsymbol {G}} _ {t}\\right)\\right) \\in [ 0, 1 ] ^ {| \\nu_ {t} |},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.839, + 0.472, + 0.867 + ], + "angle": 0, + "content": "where \\(G_{t}\\) denotes the node embeddings of \\(\\mathcal{G}_t\\). The ultimate action probabilities are given as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.123, + 0.858, + 0.146 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} ^ {2 \\mathrm {d}} = \\left[ \\boldsymbol {p} _ {t} ^ {2 \\mathrm {d} \\rightarrow g}; \\boldsymbol {p} _ {t} ^ {2 \\mathrm {d}} \\right] \\in [ 0, 1 ] ^ {| \\mathcal {V} _ {t} |}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.143, + 0.794, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} _ {t} ^ {g} = W _ {g} \\boldsymbol {p} _ {t} ^ {g} + (1 - W _ {g}) \\hat {\\boldsymbol {p}} _ {t} ^ {2 d} \\in [ 0, 1 ] ^ {| \\mathcal {V} _ {t} |},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.166, + 0.859, + 0.21 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_t^{2\\mathrm{d}\\rightarrow g} \\in \\mathbb{R}^{|\\mathcal{V}_t| - (N_t + 1)}\\) denotes the probabilities of global backtracking and we use the same value as the local [STOP] probability in Eq. (6); \\(W_{g}\\) is a learnable weight." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.859, + 0.352 + ], + "angle": 0, + "content": "State Transition and Memory Update. After executing the action in \\(\\mathcal{A}_t^*\\), our agent reaches the next viewpoint \\(v_{t + 1,0}^{*}\\), and will iteratively: (1) encode its current observation as \\(F_{t + 1}^{3\\mathrm{d}}\\) through Eq. (1); (2) update its volume state as \\(s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})\\); (3) add the node embeddings of \\(\\{v_{t + 1,n}^{*}\\}_{n = 0}^{N_{t + 1}}\\) into the episodic memory \\(\\mathcal{G}_{t + 1}\\); and (4) predict the next step with the updated episodic memory (i.e., \\(\\mathcal{G}_{t + 1}\\)) and volume state (i.e., \\(s_{t + 1}\\)). Our agent repeats the above process until it chooses the [STOP] action or reaches the maximum step limit." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.361, + 0.696, + 0.374 + ], + "angle": 0, + "content": "3.4. Annotation Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.859, + 0.507 + ], + "angle": 0, + "content": "A multi-task learning framework is proposed to extract and encode the VERs (§3.1). To achieve this, we generate annotations on Matterport3D dataset [7] for 3D occupancy prediction, object detection, and room layout estimation. We design a room-object-voxel pipeline to automatically generate these annotations. This pipeline leverages existing LiDAR point labels without additional human annotations (more details in Appendix). We utilize the egocentric observations with multi-view images as input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.859, + 0.632 + ], + "angle": 0, + "content": "Room Layout. The room layout in our context specifies the positions, orientations, and heights of the walls, relative to the camera center. It aims to reconstruct cuboid room shapes within the Manhattan world [18]. Given that the horizontal plane is aligned on the \\(x - z\\) axis, we parameterize the layout with center coordinate, width, length, height, and rotation. In contrast to directly operating on a single panoramic image [105, 106], we use the embodied observations with multi-view images as input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.859, + 0.756 + ], + "angle": 0, + "content": "Object Detection. Based on the room layout, we collect the surrounding objects if the agent locates in a room. In a non-closed environment, we collect information about nearby objects based on their distance from the agent. For each object, the eigenvectors of its vertices are used to define an oriented bounding box that tightly encloses the object [35]. Considering some objects may disappear from view due to occlusion but still exist in the environment (referred to as permanence [98]), we also include them in the analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.859, + 0.867 + ], + "angle": 0, + "content": "Point Accumulation for Occupancy. To generate voxel labels for occupancy [77, 95], we accumulate the sparse LiDAR points and utilize 3D boxes. Given dense background and object points, we first voxelize the 3D space and label each voxel based on the majority vote of labelled points in that voxel. Due to the limited number of LiDAR points, we leverage the Nearest Neighbors algorithm to generate dense labels for remaining voxels by searching the nearest" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.516, + 0.903 + ], + "angle": 0, + "content": "16321" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.126, + 0.471, + 0.195 + ], + "angle": 0, + "content": "semantic label. Moreover, the agent infers the complete 3D occupancy of each object (amodal perception), including regions that are not directly observed. This attribute enables the agent to predict the entire object instead of only visible surfaces [98]." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.197, + 0.474, + 0.335 + ], + "angle": 0, + "content": "Statistics. For high-resolution labels, we define \\(120 \\times 120 \\times\\) 35 voxel grids in world coordinates, where the scene voxel size equals to \\(0.1\\mathrm{m}\\) (see more details about multi-resolution labels in Appendix). We annotate over 50 billion voxels and 16 classes, including 11 foreground objects and 5 background stuffs. It comprises about \\(100k\\) annotated bounding boxes and 1,500 room layouts within the scenes. These annotations follow the same train/val/test splits as R2R [3]. There are 61 scenes for train/val seen, 11 scenes for val unseen, and 18 scenes for test." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.345, + 0.313, + 0.36 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.367, + 0.472, + 0.505 + ], + "angle": 0, + "content": "Initially, the environment encoder (§3.1) is introduced for VER through coarse-to-fine extraction. Then multi-task learning is performed across multiple 3D perception tasks, including 3D occupancy prediction, 3D layout estimation, and 3D detection. During navigation, our agent is equipped with the frozen environment encoder and predicts the next step. Following recent VLN practice [14, 16, 33], both offline pretraining and finetuning are adopted. In this section, we will mainly introduce the details of architecture and training (see more details in Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.507, + 0.472, + 0.644 + ], + "angle": 0, + "content": "Environment Representation Learning. For the multiview images, we adopt ViT-B/16 [20] pretrained on ImageNet to extract features. The number of 3D volume queries is \\(15 \\times 15 \\times 4\\). For each query, it is projected to sample 2D views according to intrinsic and extrinsic parameters of the camera. We set the perception range as \\([-6m, 6m]\\) for \\(x - y\\) axis and \\([-1.5m, 2m]\\) for the height (\\(z\\) axis). We adopt six layers of CVA (Eq. 1) for 2D-3D sampling, and then use \\(M = 3\\) cascade deconvolutions for up-sampling (Eq. 2). The feature dimension is 768 (i.e., \\(D_{i} = D_{w} = D_{e} = 768\\))." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.646, + 0.472, + 0.756 + ], + "angle": 0, + "content": "Navigation Network. MLT with 4 layers is initialized from [73] for state estimation (Eq. 4) and global action prediction (Eq. 8), respectively. The range of neighborhood for each candidate is set as \\( |\\Omega_n| = 9 \\). The standard deviation of the Gaussian kernel for the heat map is set as 3.0. Based on this heat map, the focal loss [52] is used to supervise the local action prediction (Eq. 6). We also use a cross-entropy loss for the global action prediction (Eq. 9)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.756, + 0.472, + 0.867 + ], + "angle": 0, + "content": "Pretraining. For R2R [3] and R4R [39], Masked Language Modeling [14, 42] and Single-step Action Prediction [14, 33] are adopted as auxiliary tasks on offline-sampled instruction-route pairs [30]. For REVERIE [64], an additional Object Grounding (OG) [16, 53] is used for object reasoning. During pretraining, we train the model with a batch size of 64 for \\(100k\\) iterations, using Adam [43] optimizer with 1e-4 learning rate. Only one task is adopted" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.124, + 0.865, + 0.365 + ], + "angle": 0, + "content": "
ModelsR2R
val unseentest unseen
TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑
Seq2Seq [3]8.397.8122-8.137.852018
SF [23]-6.6235-14.826.623528
EnvDrop [74]10.705.22524811.665.235147
AuxRN [103]-5.285550-5.155551
Active [81]20.604.36584021.604.336041
RecBERT [33]12.013.93635712.354.096357
HAMT [14]11.462.29666112.273.936560
SOAT [60]12.154.28595312.264.495853
SSM [82]20.74.32624520.44.576146
CCC [84]-5.205046-5.305148
HOP [65]12.273.80645712.683.836459
DUET [16]13.943.31726014.733.656959
LANA [89]12.0-686212.6-6560
TD-STP [100]-3.227063-3.736761
BSG [55]14.902.89746214.863.197362
BEVbert [1]14.552.817564-3.137362
Ours14.832.80766515.232.747666
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.37, + 0.85, + 0.383 + ], + "angle": 0, + "content": "Table 1. Quantitative results on R2R [3] (more details in §4.1)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.404, + 0.794, + 0.417 + ], + "angle": 0, + "content": "at each mini-batch with the same sampling ratio." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.858, + 0.486 + ], + "angle": 0, + "content": "Finetuning. Following the standard protocol [1, 16, 94], we finetune the navigation network using Dagger [69] techniques. In addition, the OG loss [1, 16, 53] is employed on REVERIE. In this stage, we set the learning rate to 1e-5 and batch size to 8 with \\(20k\\) iterations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.487, + 0.858, + 0.57 + ], + "angle": 0, + "content": "Inference. During the testing phase, the agent receives the multi-view images and encodes them as VERs through the frozen environment encoder (§3.1). Based on VERs, the agent performs volume state estimation (§3.2) and models episodic memory (§3.3). By combining both of them, the agent predicts the next step accurately until stops." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.571, + 0.858, + 0.612 + ], + "angle": 0, + "content": "Reproducibility. Our model is implemented in PyTorch and trained on eight RTX 4090 GPUs with a 24GB memory per-card. Testing is conducted on the same machine." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.614, + 0.645 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.685, + 0.665 + ], + "angle": 0, + "content": "4.1. Performance on VLN" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.673, + 0.859, + 0.811 + ], + "angle": 0, + "content": "Datasets. The experiments are conducted on three datasets. R2R [3] contains 7,189 trajectories sampled from 90 real-world indoor scenes. It consists of \\(22k\\) human-annotated navigational instructions. The dataset is split into train, val seen, val unseen, and test unseen sets, which mainly focus on the generalization capability in unseen environments. REVERIE [64] contains high-level instructions which describe target locations and objects, with a focus on grounding remote target objects. R4R [39] is an extended version of R2R with longer trajectories." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.812, + 0.859, + 0.867 + ], + "angle": 0, + "content": "Evaluation Metrics. For R2R, Success Rate (SR), Trajectory Length (TL), Oracle Success Rate (OSR), Success rate weighted by Path Length (SPL), and Navigation Error (NE) are used. For REVERIE, Remote Grounding Success rate" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16322" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.112, + 0.857, + 0.314 + ], + "angle": 0, + "content": "
Modelsval seenREVERIEtest unseen
TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑
RCM [88]10.7029.4423.3321.8216.2315.3611.9814.239.296.974.893.8910.6011.687.846.673.673.14
FAST-M [64]16.3555.1750.5345.5031.9729.6645.2828.2014.407.197.844.6739.0530.6319.8811.6111.286.08
SIA [53]13.6165.8561.9157.0845.9642.6541.5344.6731.5316.2822.4111.5648.6144.5630.8014.8519.029.20
RecBERT [33]13.4453.9051.7947.9638.2335.6116.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [28]15.1648.9847.0142.3432.7530.0118.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [14]12.7947.6543.2940.1927.2025.1814.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
HOP [65]13.8054.8853.7647.1938.6533.8516.4636.2431.7826.1118.8515.7316.3833.0630.1724.3417.6914.34
DUET [16]13.8673.8671.7563.9457.4151.1422.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
TD-STP [100]-------39.4834.8827.3221.1616.56-40.2635.8927.5119.8815.40
BEVBert [1]-76.1873.7265.3257.7051.73-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
GridMM [94]------23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
LANA [89]15.9174.2871.9462.7759.0250.3423.1852.9748.3133.8632.8622.7718.8357.2051.7236.4532.9522.85
BSG [55]15.2678.3676.1866.6961.5654.0224.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
Ours16.1380.4975.8366.1961.7156.2023.0361.0955.9839.6633.7123.7024.7462.2256.8238.7633.8823.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.318, + 0.796, + 0.332 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison results on REVERIE [64]. ' -': unavailable statistics (see §4.1 for more details)." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.347, + 0.479, + 0.498 + ], + "angle": 0, + "content": "
ModelsR4R val unseen
NE↓SR↑CLS↑nDTW↑SDTW↑
SF [3]8.472430--
RCM [88]-29353013
EGP [19]8.0030443718
SSM [82]8.2732533919
RelGraph [32]7.4336414734
RecBERT [33]6.6744514530
HAMT [14]6.0945585032
BSG [55]6.1247595334
LANA [89]-43605232
Ours6.1047615433
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.503, + 0.466, + 0.516 + ], + "angle": 0, + "content": "Table 3. Quantitative results on R4R [39] (more details in §4.1)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.535, + 0.471, + 0.603 + ], + "angle": 0, + "content": "(RGS), and Remote Grounding Success weighted by Path Length (RGSPL) are also employed for object grounding. For R4R, Coverage weighted by Length Score (CLS), normalized Dynamic Time Warping (nDTW), and Success rate weighted nDTW (SDTW) are used." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.604, + 0.472, + 0.687 + ], + "angle": 0, + "content": "Performance on R2R. Table 1 compares our model with the state-of-the-art models on R2R. As we find that our model yields SR of \\(76\\%\\) and SPL of \\(66\\%\\) on text unseen, which leads to promising gains of \\(3\\%\\) and \\(4\\%\\) over BEVBert [1], respectively. It verifies that using VER to represent the environment leads to better decision-making." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.687, + 0.472, + 0.756 + ], + "angle": 0, + "content": "Performance on REVERIE. Table 2 presents the comparison results on REVERIE. Compared with the recent state-of-the-art VLN agent [55], our agent improves SR and SPL by \\(3.86\\%\\) and \\(4.07\\%\\) on the val unseen split. This highlights the effectiveness of our architecture design." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.756, + 0.472, + 0.811 + ], + "angle": 0, + "content": "Performance on R4R. Table 3 shows results on R4R. Our approach outperforms others in most metrics with a promising gain on nDTW (i.e., \\(1\\%\\)). This suggests the episodic memory module is able to retrieve the long-time context." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.812, + 0.472, + 0.867 + ], + "angle": 0, + "content": "Visual Results. Fig. 4 depicts one exemplar navigation episode from val unseen set of R2R. In this complex environment, there are many rooms with different objects and 3D layout. From the visualization of 3D occupancy pre" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.348, + 0.865, + 0.416 + ], + "angle": 0, + "content": "
ModelsREVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
w/o Volume State52.3134.9132.7572.7161.13
w/o Episodic Memory49.3333.7131.3668.2161.70
Full model55.9839.6633.7175.8065.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.425, + 0.86, + 0.451 + ], + "angle": 0, + "content": "Table 4. Ablation study of overall design on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.859, + 0.533 + ], + "angle": 0, + "content": "dition at the key steps, we find the geometric details and semantics can be captured well by VER. The room layout estimation can help the agent to understand \"enter the bedroom\". Finally, our agent finds the \"bed\" and accomplishes the instruction successfully." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.695, + 0.555 + ], + "angle": 0, + "content": "4.2. Diagnostic Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.859, + 0.602 + ], + "angle": 0, + "content": "To thoroughly test the efficacy of crucial components of our model, we conduct a series of diagnostic studies on val unseen split of REVERIE and R2R." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.859, + 0.671 + ], + "angle": 0, + "content": "Overall Design. We first investigate the effectiveness of our core design. The results in Table 4 indicate that adding Volume State leads to a substantial performance gain (i.e., \\(3.67\\%\\) on SR). After using Episodic Memory, a higher score (i.e., \\(31.36\\% \\rightarrow 33.71\\%\\) on RGS) is achieved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.859, + 0.811 + ], + "angle": 0, + "content": "Neighborhood Range of Viewpoints. We use the neighborhood of each viewpoint in the state space for local action prediction (Eq. 6). Moreover, corresponding pillar representations of the neighborhood are also used for node embeddings of the episodic memory (Eq. 7). From Table 5, the limited range of neighborhood is insufficient to represent the candidate viewpoint for navigation (e.g., \\(75.80\\% \\rightarrow 73.75\\%\\) of SR on R2R). However, too large neighborhood range will contain irrelevant information, leading to inferior performance (e.g., \\(55.98\\% \\rightarrow 53.49\\%\\) of SR on REVERIE)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.818, + 0.818, + 0.833 + ], + "angle": 0, + "content": "4.3. Analysis on 3D Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.859, + 0.867 + ], + "angle": 0, + "content": "Evaluation Metric. Following standard protocols, we employ Intersection over Union (IoU) to evaluate the occu" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "16323" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.113, + 0.403, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.119, + 0.21, + 0.277, + 0.271 + ], + "angle": 0, + "content": "Turn around and go passed the couch. Cross the hall to the sitting room. Turn left and enter the bedroom on the left. Wait near the bed." + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.113, + 0.551, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.195, + 0.551, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.113, + 0.695, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.195, + 0.697, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.113, + 0.852, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.195, + 0.852, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.274, + 0.859, + 0.311 + ], + "angle": 0, + "content": "Figure 4. A representative visual result on val unseen of R2R [3]. We first visualize the 3D occupancy prediction at the key steps. In addition, we provide the prediction of 3D boxes and 3D room layout at step \\(⑦\\). We find that VER can capture the geometric details of 'couch' and the structure of 'bedroom'. With VER, our agent easily finds the 'bed' and succeeds. See §4.1 for more details." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.317, + 0.471, + 0.4 + ], + "angle": 0, + "content": "
|Ωn|REVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
453.3035.9034.1673.7562.77
955.9839.6633.7175.8065.37
1652.1836.3733.1473.8263.49
2553.4935.8733.4674.6363.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.404, + 0.473, + 0.43 + ], + "angle": 0, + "content": "Table 5. Ablation study of neighborhood range on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details)." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.444, + 0.477, + 0.51 + ], + "angle": 0, + "content": "
ModelsOccupancyDetectionLayout
IoU↑mIoU↑mAP↑mAR↑3D IoU↑
BEVFormer [51]20.388.9727.3043.8862.71
OccNet [78]22.1210.6629.9147.1564.04
Ours24.3112.9333.5751.6066.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.514, + 0.472, + 0.54 + ], + "angle": 0, + "content": "Table 6. Quantitative results on 3D occupancy, 3D detection, and room layout prediction (see §4.3 for more details)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.56, + 0.472, + 0.643 + ], + "angle": 0, + "content": "pancy prediction quality, regardless of the semantic labels. The mean IoU (mIoU) of 15 classes is also used to assess the performance of semantic occupancy. For 3D object detection, we utilize mean Average Precision (mAP) and mean Average Recall (mAR) with IoU thresholds of 0.50. For room layout, we adopt 3D IoU for cuboid layout evaluation." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.644, + 0.472, + 0.74 + ], + "angle": 0, + "content": "Performance on 3D Tasks. In Table 6, our network (§3.1) outperforms other methods [51, 78] by a significant margin (2.19% on IoU of occupancy, 3.66% on mAP of 3D detection, and 2.41% on 3D IoU of room layout estimation). The mIoU of occupancy is also exhibits improvement (2.27%), underscoring the network's proficiency in capturing both scene geometry and fine-grained semantics." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.741, + 0.472, + 0.825 + ], + "angle": 0, + "content": "Coarse-to-Fine Extraction. Table 7 lists the scores with different up-sampling operations (Eq. 2). Our approach improves the performance by solid margins (e.g., \\(11.03\\% \\rightarrow 12.93\\%\\) for 3D occupancy, \\(75.14\\% \\rightarrow 75.80\\%\\) on SR of R2R). This verifies the efficacy of our design of the coarse-to-fine extraction and learnable up-sampling operations." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.825, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Multi-task Learning. Table 8 reports performance comparison with different perception tasks (§3.1). We find that multi-task learning yields a substantial performance gain." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.318, + 0.869, + 0.38 + ], + "angle": 0, + "content": "
Up-Sampling3D PerceptionR2R
mIoU↑mAP↑3D IoU↑SR↑SPL↑
w/o Coarse-to-Fine12.3932.9566.57--
Trilinear Interpolation11.0329.4263.4575.1464.30
3D Deconvolution12.9333.5766.4575.8065.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.385, + 0.86, + 0.423 + ], + "angle": 0, + "content": "Table 7. Ablation study of Coarse-to-Fine Extraction on occupancy prediction (mIoU), 3D detection (mAP), room layout (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details)." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.433, + 0.864, + 0.526 + ], + "angle": 0, + "content": "
Multi-task Learning3D PerceptionR2R
Occ.Obj.Room.mIoU↑mAP↑3D IoU↑SR↑SPL↑
✓12.09--74.9063.82
✓✓12.1432.64-75.2164.79
✓✓-33.1164.5874.0363.51
✓✓11.37-63.2974.9764.66
✓✓✓12.9333.5766.4575.8065.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.532, + 0.859, + 0.571 + ], + "angle": 0, + "content": "Table 8. Ablation study of Multi-task Learning on occupancy prediction (mIoU), 3D detection (mAP), room layout estimation (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.859, + 0.634 + ], + "angle": 0, + "content": "This suggests these 3D perception tasks are complementary to each other in capturing geometric and semantic properties of scenes, further facilitating the decision-making." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.608, + 0.663 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.859, + 0.796 + ], + "angle": 0, + "content": "In this paper, we propose a Volumetric Environment Representation (VER), which aggregates the perspective features into structured 3D cells. Through coarse-to-fine feature extraction, we can efficiently perform several 3D perception tasks. Based on this comprehensive representation, we develop the volume state for local action prediction and the episodic memory for retrieving the global context. We demonstrate that our agent achieves promising performance on VLN benchmarks (R2R, REVERIE, and R4R)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.859, + 0.868 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (No. 62372405), the Fundamental Research Funds for the Central Universities (No. 226-2022-00051), and CCF-Tencent Open Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "16324" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.201, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.148, + 0.473, + 0.198 + ], + "angle": 0, + "content": "[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. In ICCV, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.2, + 0.472, + 0.249 + ], + "angle": 0, + "content": "[2] Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv:2304.03047, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.251, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, 2018. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.315, + 0.472, + 0.353 + ], + "angle": 0, + "content": "[4] Peter Anderson, Ayush Shrivastava, Devi Parikh, Dhruv Batra, and Stefan Lee. Chasing ghosts: Instruction following as bayesian state tracking. In NeurIPS, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.354, + 0.472, + 0.416 + ], + "angle": 0, + "content": "[5] Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuzhhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt4.arXiv:2303.12712, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.418, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[6] Vincent Cartillier, Zhile Ren, Neha Jain, Stefan Lee, Irfan Essa, and Dhruv Batra. Semantic mapnet: Building allocentric semantic maps and representations from egocentric views. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.469, + 0.472, + 0.52 + ], + "angle": 0, + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.521, + 0.472, + 0.559 + ], + "angle": 0, + "content": "[8] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.129, + 0.56, + 0.472, + 0.597 + ], + "angle": 0, + "content": "[9] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.598, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[10] Jinyu Chen, Wenguan Wang, Si Liu, Hongsheng Li, and Yi Yang. Omnidirectional information gathering for knowledge transfer-based audio-visual navigation. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.649, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[11] Kevin Chen, Junshen K Chen, Jo Chuang, Marynel Vázquez, and Silvio Savarese. Topological planning with transformers for vision-and-language navigation. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.701, + 0.472, + 0.751 + ], + "angle": 0, + "content": "[12] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.752, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[13] Peihao Chen, Xinyu Sun, Hongyan Zhi, Runhao Zeng, Thomas H Li, Gaowen Liu, Mingkui Tan, and Chuang Gan. \\(a^2\\) nav: Action-aware zero-shot robot navigation by exploiting vision-and-language ability of foundation models. In NeurIPS Workshop, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.816, + 0.472, + 0.865 + ], + "angle": 0, + "content": "[14] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.122, + 0.148, + 0.473, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.128, + 0.857, + 0.178 + ], + "angle": 0, + "content": "[15] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.18, + 0.857, + 0.231 + ], + "angle": 0, + "content": "[16] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.233, + 0.857, + 0.271 + ], + "angle": 0, + "content": "[17] Ian Cherabier, Johannes L Schonberger, Martin R Oswald, Marc Pollefeys, and Andreas Geiger. Learning priors for semantic 3d reconstruction. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.273, + 0.857, + 0.311 + ], + "angle": 0, + "content": "[18] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In ICCV, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.313, + 0.857, + 0.352 + ], + "angle": 0, + "content": "[19] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In NeurIPS, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.353, + 0.857, + 0.416 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.418, + 0.857, + 0.457 + ], + "angle": 0, + "content": "[21] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.459, + 0.857, + 0.496 + ], + "angle": 0, + "content": "[22] Alberto Elfes. Occupancy grids: A stochastic spatial representation for active robot perception. arXiv:1304.1098, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.499, + 0.857, + 0.562 + ], + "angle": 0, + "content": "[23] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.565, + 0.857, + 0.614 + ], + "angle": 0, + "content": "[24] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.617, + 0.857, + 0.692 + ], + "angle": 0, + "content": "[25] Sourav Garg, Niko Sunderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.695, + 0.857, + 0.747 + ], + "angle": 0, + "content": "[26] Georgios Georgakis, Karl Schmeckpeper, Karan Wanchoo, Soham Dan, Eleni Miltsakaki, Dan Roth, and Kostas Dani-ilidis. Cross-modal map learning for vision and language navigation. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.748, + 0.857, + 0.786 + ], + "angle": 0, + "content": "[27] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In ACL, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.789, + 0.857, + 0.838 + ], + "angle": 0, + "content": "[28] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.841, + 0.857, + 0.867 + ], + "angle": 0, + "content": "[29] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In NeurIPS, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.128, + 0.857, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16325" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.127, + 0.473, + 0.177 + ], + "angle": 0, + "content": "[30] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.179, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[31] Joao F Henriques and Andrea Vedaldi. Mapnet: An allocentric spatial memory for mapping environments. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.217, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[32] Yicong Hong, Cristian Rodriguez, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.256, + 0.472, + 0.304 + ], + "angle": 0, + "content": "[33] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In CVPR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.307, + 0.472, + 0.356 + ], + "angle": 0, + "content": "[34] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.358, + 0.472, + 0.394 + ], + "angle": 0, + "content": "[35] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.396, + 0.472, + 0.433 + ], + "angle": 0, + "content": "[36] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.435, + 0.472, + 0.484 + ], + "angle": 0, + "content": "[37] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.485, + 0.472, + 0.547 + ], + "angle": 0, + "content": "[38] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.549, + 0.472, + 0.598 + ], + "angle": 0, + "content": "[39] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.6, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[40] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.65, + 0.472, + 0.713 + ], + "angle": 0, + "content": "[41] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.715, + 0.472, + 0.763 + ], + "angle": 0, + "content": "[42] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.765, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[43] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.79, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[44] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.829, + 0.472, + 0.866 + ], + "angle": 0, + "content": "[45] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.122, + 0.127, + 0.473, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.127, + 0.857, + 0.164 + ], + "angle": 0, + "content": "[46] Obin Kwon, Jeongho Park, and Songhwai Oh. Rendering neural radiance map for visual navigation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.167, + 0.857, + 0.229 + ], + "angle": 0, + "content": "[47] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE TPAMI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.231, + 0.857, + 0.269 + ], + "angle": 0, + "content": "[48] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.271, + 0.857, + 0.307 + ], + "angle": 0, + "content": "[49] Jialu Li, Hao Tan, and Mohit Bansal. Envide: Environment editing for vision-and-language navigation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.31, + 0.857, + 0.346 + ], + "angle": 0, + "content": "[50] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In CoRL, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.349, + 0.857, + 0.41 + ], + "angle": 0, + "content": "[51] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.413, + 0.857, + 0.45 + ], + "angle": 0, + "content": "[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.452, + 0.857, + 0.489 + ], + "angle": 0, + "content": "[53] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In CVPR, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.491, + 0.857, + 0.54 + ], + "angle": 0, + "content": "[54] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.543, + 0.857, + 0.58 + ], + "angle": 0, + "content": "[55] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.582, + 0.857, + 0.62 + ], + "angle": 0, + "content": "[56] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.622, + 0.857, + 0.658 + ], + "angle": 0, + "content": "[57] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zero-shot video grounding with pseudo query lookup and verification. IEEE TIP, 33:1643-1654, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.661, + 0.857, + 0.71 + ], + "angle": 0, + "content": "[58] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.713, + 0.857, + 0.773 + ], + "angle": 0, + "content": "[59] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.776, + 0.857, + 0.827 + ], + "angle": 0, + "content": "[60] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.829, + 0.857, + 0.866 + ], + "angle": 0, + "content": "[61] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In ICLR, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.127, + 0.857, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16326" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.127, + 0.472, + 0.165 + ], + "angle": 0, + "content": "[62] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.166, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[63] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.205, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[64] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.27, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[65] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop: history-and-order aware pretraining for vision-and-language navigation. In CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.321, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[66] Ruijie Quan, Linchao Zhu, Yu Wu, and Yi Yang. Holistic lstm for pedestrian trajectory prediction. IEEE TIP, 30: 3229-3239, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.36, + 0.472, + 0.411 + ], + "angle": 0, + "content": "[67] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.412, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[68] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. IJCV, 130(8):1978-2005, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.451, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[69] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.491, + 0.472, + 0.54 + ], + "angle": 0, + "content": "[70] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.542, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[71] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.581, + 0.472, + 0.618 + ], + "angle": 0, + "content": "[72] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In NeurIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.621, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[73] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.66, + 0.472, + 0.697 + ], + "angle": 0, + "content": "[74] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In NAACL, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.698, + 0.472, + 0.748 + ], + "angle": 0, + "content": "[75] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Self-supervised 3d semantic representation learning for vision-and-language navigation. arXiv:2201.10788, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.75, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[76] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.777, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[77] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv:2304.14365, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.828, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[78] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.122, + 0.127, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.127, + 0.858, + 0.177 + ], + "angle": 0, + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.179, + 0.858, + 0.216 + ], + "angle": 0, + "content": "[80] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.217, + 0.858, + 0.254 + ], + "angle": 0, + "content": "[81] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.255, + 0.858, + 0.292 + ], + "angle": 0, + "content": "[82] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.293, + 0.858, + 0.33 + ], + "angle": 0, + "content": "[83] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.332, + 0.858, + 0.381 + ], + "angle": 0, + "content": "[84] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.383, + 0.858, + 0.42 + ], + "angle": 0, + "content": "[85] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.421, + 0.858, + 0.47 + ], + "angle": 0, + "content": "[86] Hanqing Wang, Wenguan Wang, Wei Liang, Steven CH Hoi, Jianbing Shen, and Luc Van Gool. Active perception for visual-language navigation. IJCV, 131(3):607-625, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.472, + 0.858, + 0.535 + ], + "angle": 0, + "content": "[87] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.536, + 0.858, + 0.598 + ], + "angle": 0, + "content": "[88] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.6, + 0.858, + 0.637 + ], + "angle": 0, + "content": "[89] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.638, + 0.858, + 0.674 + ], + "angle": 0, + "content": "[90] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Learning to follow and generate instructions for language-capable navigation. IEEE TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.676, + 0.858, + 0.725 + ], + "angle": 0, + "content": "[91] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.727, + 0.858, + 0.775 + ], + "angle": 0, + "content": "[92] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.777, + 0.858, + 0.827 + ], + "angle": 0, + "content": "[93] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.828, + 0.858, + 0.867 + ], + "angle": 0, + "content": "[94] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.127, + 0.858, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16327" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.127, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[95] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.179, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[96] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.218, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[97] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv:2307.13510, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.27, + 0.472, + 0.308 + ], + "angle": 0, + "content": "[98] Zhenjia Xu, Zhanpeng He, Jiajun Wu, and Shuran Song. Learning 3d dynamic scene representations for robot manipulation. In CoRL, 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.31, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[99] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.361, + 0.472, + 0.411 + ], + "angle": 0, + "content": "[100] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In ACM MM, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.413, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[101] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.452, + 0.472, + 0.478 + ], + "angle": 0, + "content": "[102] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv: 1904.07850, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.479, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[103] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.518, + 0.472, + 0.556 + ], + "angle": 0, + "content": "[104] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.557, + 0.472, + 0.595 + ], + "angle": 0, + "content": "[105] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.596, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[106] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.472, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "16328" + } + ] +] \ No newline at end of file diff --git a/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_origin.pdf b/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81a6648b6f3df5c5d6c990119f38d99a6be0b73a --- /dev/null +++ b/2024/Volumetric Environment Representation for Vision-Language Navigation/6562b0ae-fedd-4e65-b84a-d2cbb4f40727_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b42948a8bd8fac23a8adbf885c49f4ca43252cb779e9481bbe75c782614e07ee +size 1429941 diff --git a/2024/Volumetric Environment Representation for Vision-Language Navigation/full.md b/2024/Volumetric Environment Representation for Vision-Language Navigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dcc73f3eab93e8d757fd4e0af29c53a815a5a3d5 --- /dev/null +++ b/2024/Volumetric Environment Representation for Vision-Language Navigation/full.md @@ -0,0 +1,430 @@ +# Volumetric Environment Representation for Vision-Language Navigation + +Rui Liu Wenguan Wang Yi Yang* + +ReLER, CCAI, Zhejiang University + +https://github.com/DefaultRui/VLN-VER + +# Abstract + +Vision-language navigation (VLN) requires an agent to navigate through an 3D environment based on visual observations and natural language instructions. It is clear that the pivotal factor for successful navigation lies in the comprehensive scene understanding. Previous VLN agents employ monocular frameworks to extract 2D features of perspective views directly. Though straightforward, they struggle for capturing 3D geometry and semantics, leading to a partial and incomplete environment representation. To achieve a comprehensive 3D representation with fine-grained details, we introduce a Volumetric Environment Representation (VER), which voxelizes the physical world into structured 3D cells. For each cell, VER aggregates multi-view 2D features into such a unified 3D space via 2D-3D sampling. Through coarse-to-fine feature extraction and multi-task learning for VER, our agent predicts 3D occupancy, 3D room layout, and 3D bounding boxes jointly. Based on online collected VERs, our agent performs volume state estimation and builds episodic memory for predicting the next step. Experimental results show our environment representations from multi-task learning lead to evident performance gains on VLN. Our model achieves state-of-the-art performance across VLN benchmarks (R2R, REVERIE, and R4R). + +# 1. Introduction + +Vision-language navigation (VLN) requires an agent to navigate in a 3D environment following natural language instructions [3, 64]. As a holistic understanding of the environment plays a pivotal role in decision-making within VLN, environment representation learning serves as a foundation for formulating accurate navigation policies. + +Early VLN approaches [3, 23] typically learn the navigation policy through the sequence-to-sequence (Seq2Seq) framework [72], which directly maps instructions and multi-view perspective observations to actions. They sim + +![](images/5f19d63d8b8ddb11e9961a01e4077fc09ef996fc26f68a809ff701d3ad8f3b53.jpg) +Previous Methods +Perspective +Observations +(1) +# +# + +# +2D Features of +Candidate View +# +Topological Graph +$\rightarrow$ + +$\frac{1}{2}x - 1 > 0$ +Semantic Map +1 +2D-3D Sampling +23. 15 sampling +# + +multi-view +D Features +$\therefore m - 1 \neq 0$ ; +3D Occupancy +(1) +3D Detection - +$\therefore m : x = 1$ 或 ${3x} + {4y} + 1 = 0$ +Room Layout +Estimation +# +Volume State Space +Volume State Space +Figure 1. The agent observes its surroundings with corresponding perspective features of different candidate views (O). Previous methods construct the topological graph or semantic map based on these 2D features. Our VER aggregates the multi-view features into structured 3D cells via 2D-3D sampling. VER is a powerful representation for both 3D perception tasks and VLN, providing a volume state space for decision-making. + +Our Method + +Our Method + +ply embed their immediate observation of local environment into the hidden states of recurrent units. As a result, they lack of explicit environment representations and struggle to access their past states during long-time exploration [61, 82]. To address this issue, later VLN agents are equipped with an external memory module [61], which stores the environment representations distinctly from navigation states. In this way, they can explicitly model and maintain the environment layouts and contents in a form of topological graph [2, 11, 16, 19, 32, 62] or semantic map [1, 4, 10, 12, 26, 34, 38, 55, 75, 94] (Fig. 1). Despite their promising performance with advanced frameworks (e.g., graph neural network [44] and Transformer [79]), their environment representations are still built upon 2D perspective features extracted by monocular frameworks. While straightforward, they compress depth information onto the perspective plane, sacrificing the integral scene structure in the 3D space. Thus, they encounter challenges in capturing 3D geometry and semantics in complex scenes. Such an incomplete environment representation easily leads to suboptimal navigation decisions. + +In this article, we propose a Volumetric Environment Representation (VER) that quantizes the physical world into structured 3D cells (Fig. 1). These cells, arranged within a predefined volumetric space, maintain both height and depth dimensions. Each cell corresponds to local context of the 3D space. VER aggregates multi-perspective 2D features within these cells through an environment encoder (§3.1). Compared to previous partial representations derived from hidden states and external memory, our VER captures the full geometry and semantics of the physical world. These 3D cells store the properties of the corresponding space in the scene by predicting 3D occupancy [70, 77], room layout [106], and 3D object boxes [51]. However, directly reconstructing the high-quality VER from 2D perspective views is challenging to capture the fine-grained details. As a response, we propose a coarse-to-fine VER extraction architecture, which uses learnable up-sampling operations to construct the representations progressively. It is supervised by multi-resolution semantic labels at different scales, utilizing the coarse-to-fine representations as hierarchical inputs. The annotations of the 3D tasks are collected for multi-task learning (§3.4). + +At each navigation step, our agent initially encodes the multi-view observations into VER (§3.2). With VER, instructions can be more effectively grounded in the 3D context. This is achieved by establishing cross-modal correlations between linguistic words and 3D cells of VER. Based on the correlations, a volume state estimation module is proposed to calculate transition probabilities over the surrounding cells. With the help of this module, our agent performs comprehensive decision-making in volumetric space, and then maps the volume state into local action space. In addition, an episodic memory module is established to online collect the information of observed viewpoints and build a topological graph providing global action space (§3.3). The node embeddings in the graph are from neighbor pillar representations in VER corresponding to the respective viewpoints. To balance the long-range action reasoning and language grounding, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory. + +Our agent is evaluated on three VLN benchmarks, i.e., R2R [3], REVERIE [64], and R4R [39] ( $\S 4.1$ ). It yields solid performance gains (about $3\%$ SR and $4\%$ SPL on R2R test, $4\%$ SR and $4\%$ SPL on REVERIE val unseen). The ablation study confirms the efficacy of core model designs ( $\S 4.2$ ). Additional results show our model achieves promising performance in 3D occupancy prediction, 3D detection, and room layout estimation ( $\S 4.3$ ). + +# 2. Related Work + +Vision-Language Navigation (VLN). Early VLN agents [3, 23] are built upon Seq2Seq [66, 72] framework to re + +serve the observation history in hidden state. Thus they struggle to capture long-range context as the path length increases. Later efforts are devoted to multimodal representation learning, navigation strategy learning, and data generation [27]. As a primary step, multimodal representation learning helps agents understand the environments and establish relations between the instructions and visual observations. Inspired by the success of vision-language pretraining [58, 65, 73], recent approaches [14, 33] use transformer-based architectures [42, 79] for joint visual and textual representations. Some attempts further exploit the visual information by modeling semantic relation [37] and spatial information [11, 19, 32, 83, 86, 100]. For navigation strategy learning, many VLN models [24, 88] use imitation and reinforcement learning-based training strategies. Previous solutions [45, 85] introduce world models [29] to perform mental simulations and make mental planning. Furthermore, the scarcity of human instructions and limited diversity of the scene hinder the agent to learn navigation policy and generalize to unseen environments well [54]. Therefore, several VLN data generation strategies have been proposed to create new trajectories from existing datasets [15, 28, 93], generate more instructions [41, 74, 84, 90], or create synthetic environments [48, 49]. In addition, driven by large models [5], existing agents [13, 57, 71, 101] demonstrate promising zero-shot performance. + +Despite their outstanding contributions, most of them rely on 2D visual cues in perspective observations. These representations are constrained by occlusion and limited geometric information, especially in complex scenes. In this paper, we propose VER, a unified environment representation learned by 3D perception tasks. During navigation, our agent performs volume state estimation on VER, facilitating comprehensive decision-making within the 3D space. + +Environment Representation. Existing VLN models introduce various representations for environment modeling, including topological graphs [2, 11, 16, 19, 32, 82], and semantic spatial representations [1, 4, 12, 26, 34, 38, 55, 75, 94]. With a broader view, diverse representations have been proposed for robotics and autonomous driving [25, 47, 68, 76, 96]. In the early stage, 2D occupancy grid maps [22] model occupied and free space in the surroundings based on Bayesian estimation for robot navigation. Classic SLAM systems [21] construct a map directly by integrating information from various sensors, including LiDAR and cameras. However, the representations in SLAM still rely on primitives such as 3D point clouds and image patches. Some efforts [6, 8, 31] focus on developing learnable semantic map representations. To enhance spatial reasoning, scene graph representations [9, 80] define the topological relations between spatial elements of the environment. In addition, neural scene representations [46, 50] embed image observations into latent codes + +![](images/3f94df0618eedd39947ddf9e7e46ca8f5ffecdd45e20241f110237182613679f.jpg) +Figure 2. Overview of our model. Given the perspective features of candidate views, a group of 3D queries are used to sample and aggregate them into VER (§3.1). To encode VER, we adopt coarse-to-fine extraction and perform multi-task learning on 3D perception. Based on VER, a volume state estimation module is proposed to predict state transition (§3.2). The episodic memory is used to store past observations using neighboring pillar representations for each viewpoint (§3.3). For decision-making, our agent combines both the local action probabilities from the volume state and the global action probabilities obtained from the episodic memory. See §3 for more details. + +for object categories, showcasing scalability to large scenes [40, 59]. Vision-centric BEV perception, which transforms perspective-view inputs to BEV grid representations, has recently received increasing attention [47, 51, 63]. As the BEV representations simplify the vertical geometry, 3D occupancy prediction [36, 78] is further proposed to infer the 3D geometry from perspective images [68, 70, 87, 98]. + +However, existing VLN agents mainly employ abstract relations or compressed spatial maps, lacking the ability to access complete scene information. Providing more world context can be beneficial for the subsequent decision-making and policy learning. Motivated by this insight, we explore a holistic environment representation VER that voxelizes the world into structured 3D cells. VER captures both semantic information and geometric details of the whole scene. Building upon VER, our agent is able to predict the 3D occupancy, room layout, and 3D boxes accurately. + +# 3. Approach + +Problem Formulation. For brevity, we present the technical description in the context of R2R [3]. The navigable area of the environment is organized as an undirected graph, containing a set of nodes (viewpoints) and connectivity edges. In R2R, an embodied agent needs to navigate to a target location in the 3D environment following human instructions with $L$ words (embedded as $E \in \mathbb{R}^{D_w \times L}$ , where $D_w$ is the channel dimension). At time step $t$ , the agent looks around and obtains multi-view observations of its surrounding scene from the current location. Each view is represented by a 2D visual feature $F_t^{2d} \in \mathbb{R}^{D_i \times H \times W}$ , where $H$ and $W$ are the spatial shape of image plane, $D_i$ denotes the channel dimension. The local action space $A_t \in \mathbb{R}^{N_t + 1}$ is defined by $N_t$ candidate views, which correspond to neighboring navigable nodes $\{v_{t,n}^*\}_{n=1}^{N_t}$ , as well as a [STOP] action. Previous agents predict the action probabilities $p_t^{2d} \in \mathbb{R}^{N_t + 1}$ directly based on $F_t^{2d}$ of each candidate view. However, + +these 2D features with limited geometric information are partial representations of the 3D environment, easily leading to suboptimal decision making. + +Overview. To achieve comprehensive scene understanding, we introduce VER, which voxelizes the 3D world into structured 3D cells (Fig. 2). At step $t$ , an environment encoder is proposed to sample multi-view features $(F_{t}^{2\mathrm{d}}$ of each view) into the volumetric space of VER, forming a unified representation $F_{t}^{3\mathrm{d}} \in \mathbb{R}^{D_{e} \times X \times Y \times Z}$ (\$3.1). $X$ and $Y$ are the shapes of horizontal plane, $Z$ reserves the height information of 3D space, and $D_{e}$ represents the channel dimension. The volumetric space aligns with gravity in the world coordinate system based on the Manhattan assumption [106]. To encode VER, we devise coarse-to-fine extraction with multiple 3D perception tasks supervised by multi-resolution annotations (\$3.4). Based on VER, a volume state estimation module is proposed to predict state transition probabilities $p_{t}^{3\mathrm{d}} \in \mathbb{R}^{X \times Y \times Z}$ over surrounding 3D cells (\$3.2). With this module, our agent performs comprehensive decision-making in the 3D space and maps $p_{t}^{3\mathrm{d}}$ to $p_{t}^{2\mathrm{d}}$ . To predict the next step, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory (\$3.3). + +# 3.1.Environment Encoder + +2D-3D Sampling. At step $t$ , the agent observes its surroundings and acquires the multi-view images. We introduce cross-view attention (CVA) to aggregate their features $(F^{2d}$ for each view) into a unified volumetric representation $F^{3d}$ with a group of learnable volume queries $Q \in \mathbb{R}^{D_e \times X \times Y \times Z}$ ( $t$ is omitted for simplicity). Specifically, for the 3D cell positioned at $(x,y,z)$ within the egocentric world, a single query $Q(x,y,z) \in \mathbb{R}^{D_e}$ is used to sample each image feature $F^{2d}$ as: + +$$ +\boldsymbol {F} ^ {3 d} (x, y, z) = \operatorname {C V A} \left(\boldsymbol {Q} (x, y, z), \boldsymbol {F} ^ {2 d} \left(h ^ {\prime}, w ^ {\prime}\right)\right), \tag {1} +$$ + +where $(h', w')$ denotes the location of corresponding sampling point on the image plane. Note that we only show + +![](images/2d216cda22b946887bfdb1b7ccd0e3a5ae64a419ee1d12997b24a489f0ffbc98.jpg) +Figure 3. Our coarse-to-fine VER representation extraction (§3.1) adopts cascade up-sampling operations with 3D deconvolutions (Eq. 2) and 3D queries (Eq. 1). The training process is supervised at different scales by multi-resolution semantic labels. + +the formulation of a single sampling point for conciseness. Since the sampling strategy of vanilla cross-attention is computationally expensive, the deformable attention [51, 104] is introduced and extended in CVA. In this way, $Q(x,y,z)$ selectively attends to a set of key sampling points around a reference instead of the entire $F^{2\mathrm{d}}$ . + +Coarse-to-Fine VER Extraction Architecture. Directly recovering the fine-grained VER from perspective features easily leads to performance and efficiency degradation [17, 95]. The coarse-to-fine extraction is proposed to reconstruct VER progressively. Our approach involves cascade up-sampling operations (Fig. 3), dividing this extraction into $M$ levels. At each level, 3D deconvolutions are utilized for lifting spatial resolution, and then CVA is used to query the multi-view 2D features (Eq. 1) for refining the detailed geometry. This enables the direct learning of details and avoids the inaccuracy of interpolation [56, 77, 92] (see Table 7). Between the input coarse feature $F^{3\mathrm{d}(0)} \in \mathbb{R}^{D_e \times \frac{X}{2^M} \times \frac{Y}{2^M} \times \frac{Z}{2^M}}$ from Eq. (1) and the target fine feature $F^{3\mathrm{d}(M)} \in \mathbb{R}^{D_e \times X \times Y \times Z}$ , the intermediate features with varying shapes are calculated as follows: + +$$ +\boldsymbol {F} ^ {\mathrm {3 d} (1)} = \uparrow \boldsymbol {F} ^ {\mathrm {3 d} (0)}, \dots , \boldsymbol {F} ^ {\mathrm {3 d} (M)} = \uparrow \boldsymbol {F} ^ {\mathrm {3 d} (M - 1)}, \tag {2} +$$ + +where $F^{3\mathrm{d}(1)}, \dots, F^{3\mathrm{d}(M-1)}$ denote the intermediate features from different levels, and $\uparrow$ denotes the up-sampling. Multi-task Learning. Our VER offers a unified scene representation for various 3D perception tasks. Existing studies [68, 70] highlight that semantics and geometry are tightly intertwined. We train our environment encoder to extract VER under the supervision of multiple 3D perception tasks (§3.4). This process utilizes $\{F^{3\mathrm{d}(0)}, \dots, F^{3\mathrm{d}(M)}\}$ as inputs and is supervised at different scales (Fig. 3). For 3D occupancy prediction, the decoder is implemented as MLPs with the focal loss [52]. For 3D layout estimation, we adopt a query-based head to yield the Manhattan room layouts. A combination of the L1 loss and the IoU loss [67] is used as the training objective. For 3D detection, we employ a detection head to predict the 3D boxes [51]. The bipartite match + +ing and the bounding box loss [51, 104] are employed for detection. A weight vector $[2.0, 0.25, 0.25]$ is used to balance the optimization of the three tasks, respectively. During navigation, the agent traverses between different viewpoints and encodes VERs through the frozen environment encoder. + +# 3.2. Volume State Estimation + +VLN task is typically viewed as a state estimation and transition problem [76]. With our VER, the agent state is represented as "volume state". As such, the state transition within the locally observed 3D environment, computed by Eq. (3)&(4), is referred as "volume state estimation". Different from previous plane-level state models [1, 4], VER introduces additional height dimension for 3D state estimation. This enables a more accurate action prediction. + +Volume State. At the beginning of a navigation episode, the agent is located at a start viewpoint $(x_0,y_0,z_0)$ . Based on its perception range, a volume state space $\mathcal{X}\in \mathbb{R}^{X\times Y\times Z}$ is defined corresponding to the 3D physical world with an initial state $s_0 = (x_0,y_0,z_0)$ . At step $t$ , the next intermediate state $s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})$ is determined by the instruction embeddings $\pmb{E}$ and VER $\pmb{F}_t^{3\mathrm{d}}$ for reaching the goal state $s_T$ $(0 < t < T)$ . As the entire environment is partially observable, the current state transition $(s_t\rightarrow s_{t + 1}$ in $\mathcal{X}$ ) is regarded as a local consideration for action prediction. + +State Estimation. A volume state estimation module is devised to predict the probability distribution $\pmb{p}_t^{3\mathrm{d}} \in \mathbb{R}^{X \times Y \times Z}$ of the intermediate state $s_{t+1}$ conditioned on $\pmb{E}$ and $\pmb{F}_t^{3\mathrm{d}}$ . The environment representation is first reshaped as $\pmb{F}_t^{3\mathrm{d}''} \in \mathbb{R}^{D_e \times XYZ}$ , and then adopt multi-layer transformers (MLT) to model the relations between $\pmb{E}$ and $\pmb{F}_t^{3\mathrm{d}''}$ as follows: + +$$ +\tilde {\boldsymbol {F}} _ {t} ^ {3 d} = \mathrm {M L T} \left(\left[ \boldsymbol {E}; \boldsymbol {F} _ {t} ^ {3 d ^ {\prime}} \right]\right) \in \mathbb {R} ^ {D _ {e} \times X Y Z}, \tag {3} +$$ + +where $\widetilde{F}_t^{3\mathrm{d}}$ is the updated representations, and $[\cdot ]$ denotes the concatenation operation. MLT consists of stacked selfattention blocks. Then we use MLPs for state estimation: + +$$ +\boldsymbol {p} _ {t} ^ {\mathrm {3 d}} = \operatorname {S o f t m a x} \left(\mathrm {M L P} \left(\widetilde {\boldsymbol {F}} _ {t} ^ {\mathrm {3 d}}\right)\right) \in [ 0, 1 ] ^ {X \times Y \times Z}. \tag {4} +$$ + +Efficient Height-aware Group. The computational and memory efficiency of Eq. (3) is compromised due to the resolution of $F_{t}^{3\mathrm{d}}$ ( $XYZ \gg L$ ). Considering the similarity and sparsity of information along the height direction [97, 99], we partition $F_{t}^{3\mathrm{d}}$ into several uniform groups $\{F_{t,z}^{g} \in \mathbb{R}^{D_e \times XY}\}_{z=1}^Z$ along this axis. Then we apply MLT to each group, and Eq. (3) is reformulated as: + +$$ +\begin{array}{l} \tilde {\mathbf {F}} _ {t, i} ^ {g} = \operatorname {M L T} \left(\left[ \boldsymbol {E}; \mathbf {F} _ {t, i} ^ {g} \right]\right) \in \mathbb {R} ^ {D _ {e} \times X Y}, \\ \widetilde {\boldsymbol {F}} _ {t} ^ {3 \mathrm {d}} = \left\{\widetilde {\boldsymbol {F}} _ {t, z} ^ {g} \right\} _ {z = 1} ^ {Z} \in \mathbb {R} ^ {D _ {e} \times X Y Z}, \tag {5} \\ \end{array} +$$ + +where the weights of MLT are shared among different groups. The updated features from different groups $\{\widetilde{F}_{t,z}^{g}\}_{z = 1}^{Z}$ are aggregated along the height to leverage complementary information. For ease of notation, the symbol + +$\widetilde{F}_t^{3\mathrm{d}}$ is slightly reused for the gathered 3D representations. Then, $\pmb{p}_t^{3\mathrm{d}}$ is calculated by Eq. (4). + +# 3.3. Action Prediction + +For action prediction across the entire explored scene, a topological graph $\mathcal{G}_t = \{\mathcal{V}_t,\mathcal{E}_t\}$ is constructed and updated online to represent episodic memory during navigation. Specifically, $\mathcal{V}_t$ denotes the observed viewpoints, i.e., all visited viewpoints and their candidate viewpoints. These viewpoints are encoded by compressing previous VERs. The edge $\mathcal{E}_t$ denotes the navigable connections between these viewpoints. To predict the next step, our agent combines both the volume state estimation and episodic navigation memory for decision-making. + +Mapping Volume State to Action. As our agent navigates on the horizontal plane to reach the adjacent candidate viewpoints $\{v_{n}^{*}\}_{n = 1}^{N_{t}}$ , we map the volume state space into 2D space to align with this movement pattern. Specifically, we average $p_t^{3d}$ along the height (z-axis) axis as $p_t^h\in \mathbb{R}^{X\times Y}$ . Then we sum probability values in the neighborhood of $\{v_{t,n}^{*}\}_{n = 0}^{N_{t}}$ ( $v_{0}^{*}$ for the current viewpoint, i.e., [STOP]), and normalize them as the local action probabilities: + +$$ +\boldsymbol {p} _ {t} ^ {\mathrm {2 d}} = \left\{\sum \boldsymbol {p} _ {t} ^ {h} \left(x _ {n}, y _ {n}\right) \right\} _ {n = 0} ^ {N _ {t}} \in [ 0, 1 ] ^ {N _ {t} + 1}, \left(x _ {n}, y _ {n}\right) \in \Omega_ {n}, \tag {6} +$$ + +where $\pmb{p}_t^h (x_n,y_n)$ is the value at the coordinate $(x_{n},y_{n})$ and $\Omega_{n}$ is the neighborhood of $v_{n}^{*}$ in the horizontal plane. In the training stage, a heatmap [102] with a Gaussian kernel is used to supervise this action prediction (§3.5). + +Global Action Prediction. The episodic memory module is used to store past environment representations and allows easy access to them. To memory the environment representations efficiently, we use the neighboring pillar [91] representations $\pmb{F}_{t,n}^{p} \in \mathbb{R}^{D_{e} \times |\Omega_{n}| \times Z}$ , corresponding to the current observed viewpoints $\{v_{n}^{*}\}_{n=0}^{N_{t}}$ at step $t$ : + +$$ +\boldsymbol {F} _ {t, n} ^ {P} = \left\{\tilde {\boldsymbol {F}} _ {t} ^ {3 d} \left(x _ {n}, y _ {n}, z _ {n}\right) \right\} _ {\left(x _ {n}, y _ {n}\right) \in \Omega_ {n}, z _ {n} \in [ 1, Z ]} \tag {7} +$$ + +where $\widetilde{F}_t^{3\mathrm{d}}(x_n,y_n,z_n)$ is the representation at position $(x_{n},y_{n},z_{n})$ of $\widetilde{F}_t^{3\mathrm{d}}$ (Eq. 5). $\{\bar{F}_{t,n}^p\in \mathbb{R}^{D_e}\}_{n = 0}^{N_t}$ is obtained by average pooling as the corresponding node embeddings, which are then incorporated into $\mathcal{G}_t$ . For previously observed nodes, we compute the average of their features. + +The episodic memory $\mathcal{G}_t$ , which includes the observed viewpoints, offers a global action space $\mathcal{A}_t^* \in \mathbb{R}^{|\mathcal{V}_t|}$ . This enables our agent to change its current navigation state easily by 'jumping' directly to another viewpoint, which may be even observed several steps ago. The global action probabilities on $\mathcal{G}_t$ are calculated as: + +$$ +\hat {\boldsymbol {G}} _ {t} = \operatorname {M L T} \left(\left[ \boldsymbol {E}; \boldsymbol {G} _ {t} \right]\right) \in \mathbb {R} ^ {D _ {e} \times | \mathcal {V} _ {t} |}, \tag {8} +$$ + +$$ +\boldsymbol {p} _ {t} ^ {g} = \operatorname {S o f t m a x} \left(\operatorname {M L P} \left(\hat {\boldsymbol {G}} _ {t}\right)\right) \in [ 0, 1 ] ^ {| \nu_ {t} |}, +$$ + +where $G_{t}$ denotes the node embeddings of $\mathcal{G}_t$ . The ultimate action probabilities are given as: + +$$ +\boldsymbol {p} _ {t} ^ {2 \mathrm {d}} = \left[ \boldsymbol {p} _ {t} ^ {2 \mathrm {d} \rightarrow g}; \boldsymbol {p} _ {t} ^ {2 \mathrm {d}} \right] \in [ 0, 1 ] ^ {| \mathcal {V} _ {t} |}, \tag {9} +$$ + +$$ +\hat {\boldsymbol {p}} _ {t} ^ {g} = W _ {g} \boldsymbol {p} _ {t} ^ {g} + (1 - W _ {g}) \hat {\boldsymbol {p}} _ {t} ^ {2 d} \in [ 0, 1 ] ^ {| \mathcal {V} _ {t} |}, +$$ + +where $\pmb{p}_t^{2\mathrm{d}\rightarrow g} \in \mathbb{R}^{|\mathcal{V}_t| - (N_t + 1)}$ denotes the probabilities of global backtracking and we use the same value as the local [STOP] probability in Eq. (6); $W_{g}$ is a learnable weight. + +State Transition and Memory Update. After executing the action in $\mathcal{A}_t^*$ , our agent reaches the next viewpoint $v_{t + 1,0}^{*}$ , and will iteratively: (1) encode its current observation as $F_{t + 1}^{3\mathrm{d}}$ through Eq. (1); (2) update its volume state as $s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})$ ; (3) add the node embeddings of $\{v_{t + 1,n}^{*}\}_{n = 0}^{N_{t + 1}}$ into the episodic memory $\mathcal{G}_{t + 1}$ ; and (4) predict the next step with the updated episodic memory (i.e., $\mathcal{G}_{t + 1}$ ) and volume state (i.e., $s_{t + 1}$ ). Our agent repeats the above process until it chooses the [STOP] action or reaches the maximum step limit. + +# 3.4. Annotation Generation + +A multi-task learning framework is proposed to extract and encode the VERs (§3.1). To achieve this, we generate annotations on Matterport3D dataset [7] for 3D occupancy prediction, object detection, and room layout estimation. We design a room-object-voxel pipeline to automatically generate these annotations. This pipeline leverages existing LiDAR point labels without additional human annotations (more details in Appendix). We utilize the egocentric observations with multi-view images as input. + +Room Layout. The room layout in our context specifies the positions, orientations, and heights of the walls, relative to the camera center. It aims to reconstruct cuboid room shapes within the Manhattan world [18]. Given that the horizontal plane is aligned on the $x - z$ axis, we parameterize the layout with center coordinate, width, length, height, and rotation. In contrast to directly operating on a single panoramic image [105, 106], we use the embodied observations with multi-view images as input. + +Object Detection. Based on the room layout, we collect the surrounding objects if the agent locates in a room. In a non-closed environment, we collect information about nearby objects based on their distance from the agent. For each object, the eigenvectors of its vertices are used to define an oriented bounding box that tightly encloses the object [35]. Considering some objects may disappear from view due to occlusion but still exist in the environment (referred to as permanence [98]), we also include them in the analysis. + +Point Accumulation for Occupancy. To generate voxel labels for occupancy [77, 95], we accumulate the sparse LiDAR points and utilize 3D boxes. Given dense background and object points, we first voxelize the 3D space and label each voxel based on the majority vote of labelled points in that voxel. Due to the limited number of LiDAR points, we leverage the Nearest Neighbors algorithm to generate dense labels for remaining voxels by searching the nearest + +semantic label. Moreover, the agent infers the complete 3D occupancy of each object (amodal perception), including regions that are not directly observed. This attribute enables the agent to predict the entire object instead of only visible surfaces [98]. + +Statistics. For high-resolution labels, we define $120 \times 120 \times$ 35 voxel grids in world coordinates, where the scene voxel size equals to $0.1\mathrm{m}$ (see more details about multi-resolution labels in Appendix). We annotate over 50 billion voxels and 16 classes, including 11 foreground objects and 5 background stuffs. It comprises about $100k$ annotated bounding boxes and 1,500 room layouts within the scenes. These annotations follow the same train/val/test splits as R2R [3]. There are 61 scenes for train/val seen, 11 scenes for val unseen, and 18 scenes for test. + +# 3.5. Implementation Details + +Initially, the environment encoder (§3.1) is introduced for VER through coarse-to-fine extraction. Then multi-task learning is performed across multiple 3D perception tasks, including 3D occupancy prediction, 3D layout estimation, and 3D detection. During navigation, our agent is equipped with the frozen environment encoder and predicts the next step. Following recent VLN practice [14, 16, 33], both offline pretraining and finetuning are adopted. In this section, we will mainly introduce the details of architecture and training (see more details in Appendix). + +Environment Representation Learning. For the multiview images, we adopt ViT-B/16 [20] pretrained on ImageNet to extract features. The number of 3D volume queries is $15 \times 15 \times 4$ . For each query, it is projected to sample 2D views according to intrinsic and extrinsic parameters of the camera. We set the perception range as $[-6m, 6m]$ for $x - y$ axis and $[-1.5m, 2m]$ for the height ( $z$ axis). We adopt six layers of CVA (Eq. 1) for 2D-3D sampling, and then use $M = 3$ cascade deconvolutions for up-sampling (Eq. 2). The feature dimension is 768 (i.e., $D_{i} = D_{w} = D_{e} = 768$ ). + +Navigation Network. MLT with 4 layers is initialized from [73] for state estimation (Eq. 4) and global action prediction (Eq. 8), respectively. The range of neighborhood for each candidate is set as $|\Omega_n| = 9$ . The standard deviation of the Gaussian kernel for the heat map is set as 3.0. Based on this heat map, the focal loss [52] is used to supervise the local action prediction (Eq. 6). We also use a cross-entropy loss for the global action prediction (Eq. 9). + +Pretraining. For R2R [3] and R4R [39], Masked Language Modeling [14, 42] and Single-step Action Prediction [14, 33] are adopted as auxiliary tasks on offline-sampled instruction-route pairs [30]. For REVERIE [64], an additional Object Grounding (OG) [16, 53] is used for object reasoning. During pretraining, we train the model with a batch size of 64 for $100k$ iterations, using Adam [43] optimizer with 1e-4 learning rate. Only one task is adopted + +
ModelsR2R
val unseentest unseen
TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑
Seq2Seq [3]8.397.8122-8.137.852018
SF [23]-6.6235-14.826.623528
EnvDrop [74]10.705.22524811.665.235147
AuxRN [103]-5.285550-5.155551
Active [81]20.604.36584021.604.336041
RecBERT [33]12.013.93635712.354.096357
HAMT [14]11.462.29666112.273.936560
SOAT [60]12.154.28595312.264.495853
SSM [82]20.74.32624520.44.576146
CCC [84]-5.205046-5.305148
HOP [65]12.273.80645712.683.836459
DUET [16]13.943.31726014.733.656959
LANA [89]12.0-686212.6-6560
TD-STP [100]-3.227063-3.736761
BSG [55]14.902.89746214.863.197362
BEVbert [1]14.552.817564-3.137362
Ours14.832.80766515.232.747666
+ +Table 1. Quantitative results on R2R [3] (more details in §4.1). + +at each mini-batch with the same sampling ratio. + +Finetuning. Following the standard protocol [1, 16, 94], we finetune the navigation network using Dagger [69] techniques. In addition, the OG loss [1, 16, 53] is employed on REVERIE. In this stage, we set the learning rate to 1e-5 and batch size to 8 with $20k$ iterations. + +Inference. During the testing phase, the agent receives the multi-view images and encodes them as VERs through the frozen environment encoder (§3.1). Based on VERs, the agent performs volume state estimation (§3.2) and models episodic memory (§3.3). By combining both of them, the agent predicts the next step accurately until stops. + +Reproducibility. Our model is implemented in PyTorch and trained on eight RTX 4090 GPUs with a 24GB memory per-card. Testing is conducted on the same machine. + +# 4. Experiment + +# 4.1. Performance on VLN + +Datasets. The experiments are conducted on three datasets. R2R [3] contains 7,189 trajectories sampled from 90 real-world indoor scenes. It consists of $22k$ human-annotated navigational instructions. The dataset is split into train, val seen, val unseen, and test unseen sets, which mainly focus on the generalization capability in unseen environments. REVERIE [64] contains high-level instructions which describe target locations and objects, with a focus on grounding remote target objects. R4R [39] is an extended version of R2R with longer trajectories. + +Evaluation Metrics. For R2R, Success Rate (SR), Trajectory Length (TL), Oracle Success Rate (OSR), Success rate weighted by Path Length (SPL), and Navigation Error (NE) are used. For REVERIE, Remote Grounding Success rate + +
Modelsval seenREVERIEtest unseen
TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑
RCM [88]10.7029.4423.3321.8216.2315.3611.9814.239.296.974.893.8910.6011.687.846.673.673.14
FAST-M [64]16.3555.1750.5345.5031.9729.6645.2828.2014.407.197.844.6739.0530.6319.8811.6111.286.08
SIA [53]13.6165.8561.9157.0845.9642.6541.5344.6731.5316.2822.4111.5648.6144.5630.8014.8519.029.20
RecBERT [33]13.4453.9051.7947.9638.2335.6116.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [28]15.1648.9847.0142.3432.7530.0118.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [14]12.7947.6543.2940.1927.2025.1814.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
HOP [65]13.8054.8853.7647.1938.6533.8516.4636.2431.7826.1118.8515.7316.3833.0630.1724.3417.6914.34
DUET [16]13.8673.8671.7563.9457.4151.1422.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
TD-STP [100]-------39.4834.8827.3221.1616.56-40.2635.8927.5119.8815.40
BEVBert [1]-76.1873.7265.3257.7051.73-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
GridMM [94]------23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
LANA [89]15.9174.2871.9462.7759.0250.3423.1852.9748.3133.8632.8622.7718.8357.2051.7236.4532.9522.85
BSG [55]15.2678.3676.1866.6961.5654.0224.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
Ours16.1380.4975.8366.1961.7156.2023.0361.0955.9839.6633.7123.7024.7462.2256.8238.7633.8823.19
+ +Table 2. Quantitative comparison results on REVERIE [64]. ' -': unavailable statistics (see §4.1 for more details). + +
ModelsR4R val unseen
NE↓SR↑CLS↑nDTW↑SDTW↑
SF [3]8.472430--
RCM [88]-29353013
EGP [19]8.0030443718
SSM [82]8.2732533919
RelGraph [32]7.4336414734
RecBERT [33]6.6744514530
HAMT [14]6.0945585032
BSG [55]6.1247595334
LANA [89]-43605232
Ours6.1047615433
+ +(RGS), and Remote Grounding Success weighted by Path Length (RGSPL) are also employed for object grounding. For R4R, Coverage weighted by Length Score (CLS), normalized Dynamic Time Warping (nDTW), and Success rate weighted nDTW (SDTW) are used. + +Performance on R2R. Table 1 compares our model with the state-of-the-art models on R2R. As we find that our model yields SR of $76\%$ and SPL of $66\%$ on text unseen, which leads to promising gains of $3\%$ and $4\%$ over BEVBert [1], respectively. It verifies that using VER to represent the environment leads to better decision-making. + +Performance on REVERIE. Table 2 presents the comparison results on REVERIE. Compared with the recent state-of-the-art VLN agent [55], our agent improves SR and SPL by $3.86\%$ and $4.07\%$ on the val unseen split. This highlights the effectiveness of our architecture design. + +Performance on R4R. Table 3 shows results on R4R. Our approach outperforms others in most metrics with a promising gain on nDTW (i.e., $1\%$ ). This suggests the episodic memory module is able to retrieve the long-time context. + +Visual Results. Fig. 4 depicts one exemplar navigation episode from val unseen set of R2R. In this complex environment, there are many rooms with different objects and 3D layout. From the visualization of 3D occupancy pre + +Table 3. Quantitative results on R4R [39] (more details in §4.1). + +
ModelsREVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
w/o Volume State52.3134.9132.7572.7161.13
w/o Episodic Memory49.3333.7131.3668.2161.70
Full model55.9839.6633.7175.8065.37
+ +Table 4. Ablation study of overall design on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details). + +dition at the key steps, we find the geometric details and semantics can be captured well by VER. The room layout estimation can help the agent to understand "enter the bedroom". Finally, our agent finds the "bed" and accomplishes the instruction successfully. + +# 4.2. Diagnostic Experiment + +To thoroughly test the efficacy of crucial components of our model, we conduct a series of diagnostic studies on val unseen split of REVERIE and R2R. + +Overall Design. We first investigate the effectiveness of our core design. The results in Table 4 indicate that adding Volume State leads to a substantial performance gain (i.e., $3.67\%$ on SR). After using Episodic Memory, a higher score (i.e., $31.36\% \rightarrow 33.71\%$ on RGS) is achieved. + +Neighborhood Range of Viewpoints. We use the neighborhood of each viewpoint in the state space for local action prediction (Eq. 6). Moreover, corresponding pillar representations of the neighborhood are also used for node embeddings of the episodic memory (Eq. 7). From Table 5, the limited range of neighborhood is insufficient to represent the candidate viewpoint for navigation (e.g., $75.80\% \rightarrow 73.75\%$ of SR on R2R). However, too large neighborhood range will contain irrelevant information, leading to inferior performance (e.g., $55.98\% \rightarrow 53.49\%$ of SR on REVERIE). + +# 4.3. Analysis on 3D Representation Learning + +Evaluation Metric. Following standard protocols, we employ Intersection over Union (IoU) to evaluate the occu + +![](images/b4a6f24632285fe7044140dcc4020ce6d26b47870a7018f11edebb1103f923dc.jpg) +Turn around and go passed the couch. Cross the hall to the sitting room. Turn left and enter the bedroom on the left. Wait near the bed. + +![](images/195d147b8ba1ef9d7fad116043938d71e4705a12a4934b6e382c6921c4aede83.jpg) + +![](images/fa1038e304bdc59c608be196534f17d6bd8b7498e3becb779165840da106f647.jpg) +Figure 4. A representative visual result on val unseen of R2R [3]. We first visualize the 3D occupancy prediction at the key steps. In addition, we provide the prediction of 3D boxes and 3D room layout at step $⑦$ . We find that VER can capture the geometric details of 'couch' and the structure of 'bedroom'. With VER, our agent easily finds the 'bed' and succeeds. See §4.1 for more details. + +![](images/e5908d47a2d36734e0b9fd62d5d03fcbeb4b559cc792b31156a350c86376d249.jpg) + +![](images/899ef8050b2b013484e8c9e41202b1035aeae1c27a6cc31ba4154c0f4eee0efa.jpg) + +![](images/bcba29e178db031f005031f1b115b7228239839994a0ce5b1a12bf0ff4084555.jpg) + +![](images/3b0e0ad220f3711e9200bc4cddcba689c5484c9818dd6876b84a1216d2c548aa.jpg) + +
|Ωn|REVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
453.3035.9034.1673.7562.77
955.9839.6633.7175.8065.37
1652.1836.3733.1473.8263.49
2553.4935.8733.4674.6363.24
+ +Table 5. Ablation study of neighborhood range on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details). + +
ModelsOccupancyDetectionLayout
IoU↑mIoU↑mAP↑mAR↑3D IoU↑
BEVFormer [51]20.388.9727.3043.8862.71
OccNet [78]22.1210.6629.9147.1564.04
Ours24.3112.9333.5751.6066.45
+ +pancy prediction quality, regardless of the semantic labels. The mean IoU (mIoU) of 15 classes is also used to assess the performance of semantic occupancy. For 3D object detection, we utilize mean Average Precision (mAP) and mean Average Recall (mAR) with IoU thresholds of 0.50. For room layout, we adopt 3D IoU for cuboid layout evaluation. + +Performance on 3D Tasks. In Table 6, our network (§3.1) outperforms other methods [51, 78] by a significant margin (2.19% on IoU of occupancy, 3.66% on mAP of 3D detection, and 2.41% on 3D IoU of room layout estimation). The mIoU of occupancy is also exhibits improvement (2.27%), underscoring the network's proficiency in capturing both scene geometry and fine-grained semantics. + +Coarse-to-Fine Extraction. Table 7 lists the scores with different up-sampling operations (Eq. 2). Our approach improves the performance by solid margins (e.g., $11.03\% \rightarrow 12.93\%$ for 3D occupancy, $75.14\% \rightarrow 75.80\%$ on SR of R2R). This verifies the efficacy of our design of the coarse-to-fine extraction and learnable up-sampling operations. + +Multi-task Learning. Table 8 reports performance comparison with different perception tasks (§3.1). We find that multi-task learning yields a substantial performance gain. + +Table 6. Quantitative results on 3D occupancy, 3D detection, and room layout prediction (see §4.3 for more details). + +
Up-Sampling3D PerceptionR2R
mIoU↑mAP↑3D IoU↑SR↑SPL↑
w/o Coarse-to-Fine12.3932.9566.57--
Trilinear Interpolation11.0329.4263.4575.1464.30
3D Deconvolution12.9333.5766.4575.8065.37
+ +Table 7. Ablation study of Coarse-to-Fine Extraction on occupancy prediction (mIoU), 3D detection (mAP), room layout (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details). + +
Multi-task Learning3D PerceptionR2R
Occ.Obj.Room.mIoU↑mAP↑3D IoU↑SR↑SPL↑
✓12.09--74.9063.82
✓✓12.1432.64-75.2164.79
✓✓-33.1164.5874.0363.51
✓✓11.37-63.2974.9764.66
✓✓✓12.9333.5766.4575.8065.37
+ +Table 8. Ablation study of Multi-task Learning on occupancy prediction (mIoU), 3D detection (mAP), room layout estimation (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details). + +This suggests these 3D perception tasks are complementary to each other in capturing geometric and semantic properties of scenes, further facilitating the decision-making. + +# 5. Conclusion + +In this paper, we propose a Volumetric Environment Representation (VER), which aggregates the perspective features into structured 3D cells. Through coarse-to-fine feature extraction, we can efficiently perform several 3D perception tasks. Based on this comprehensive representation, we develop the volume state for local action prediction and the episodic memory for retrieving the global context. We demonstrate that our agent achieves promising performance on VLN benchmarks (R2R, REVERIE, and R4R). + +Acknowledgment. This work was supported by the National Natural Science Foundation of China (No. 62372405), the Fundamental Research Funds for the Central Universities (No. 226-2022-00051), and CCF-Tencent Open Fund. + +# References + +[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. In ICCV, 2023. 1, 2, 4, 6, 7 +[2] Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv:2304.03047, 2023. 1, 2 +[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, 2018. 1, 2, 3, 6, 7, 8 +[4] Peter Anderson, Ayush Shrivastava, Devi Parikh, Dhruv Batra, and Stefan Lee. Chasing ghosts: Instruction following as bayesian state tracking. In NeurIPS, 2019. 1, 2, 4 +[5] Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuzhhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt4.arXiv:2303.12712, 2023.2 +[6] Vincent Cartillier, Zhile Ren, Neha Jain, Stefan Lee, Irfan Essa, and Dhruv Batra. Semantic mapnet: Building allocentric semantic maps and representations from egocentric views. In AAAI, 2021. 2 +[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 5 +[8] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In ICLR, 2019. 2 +[9] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In CVPR, 2020. 2 +[10] Jinyu Chen, Wenguan Wang, Si Liu, Hongsheng Li, and Yi Yang. Omnidirectional information gathering for knowledge transfer-based audio-visual navigation. In ICCV, 2023. 1 +[11] Kevin Chen, Junshen K Chen, Jo Chuang, Marynel Vázquez, and Silvio Savarese. Topological planning with transformers for vision-and-language navigation. In CVPR, 2021. 1, 2 +[12] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurIPS, 2022. 1, 2 +[13] Peihao Chen, Xinyu Sun, Hongyan Zhi, Runhao Zeng, Thomas H Li, Gaowen Liu, Mingkui Tan, and Chuang Gan. $a^2$ nav: Action-aware zero-shot robot navigation by exploiting vision-and-language ability of foundation models. In NeurIPS Workshop, 2023. 2 +[14] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 6, 7 + +[15] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 2 +[16] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 1, 2, 6, 7 +[17] Ian Cherabier, Johannes L Schonberger, Martin R Oswald, Marc Pollefeys, and Andreas Geiger. Learning priors for semantic 3d reconstruction. In ECCV, 2018. 4 +[18] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In ICCV, 1999. 5 +[19] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In NeurIPS, 2020. 1, 2, 7 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6 +[21] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2 +[22] Alberto Elfes. Occupancy grids: A stochastic spatial representation for active robot perception. arXiv:1304.1098, 2013. 2 +[23] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, 2018. 1, 2, 6 +[24] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, 2023. 2 +[25] Sourav Garg, Niko Sunderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 2 +[26] Georgios Georgakis, Karl Schmeckpeper, Karan Wanchoo, Soham Dan, Eleni Miltsakaki, Dan Roth, and Kostas Dani-ilidis. Cross-modal map learning for vision and language navigation. In CVPR, 2022. 1, 2 +[27] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In ACL, 2022. 2 +[28] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, 2021. 2, 7 +[29] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In NeurIPS, 2018. 2 + +[30] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, 2020. 6 +[31] Joao F Henriques and Andrea Vedaldi. Mapnet: An allocentric spatial memory for mapping environments. In CVPR, 2018. 2 +[32] Yicong Hong, Cristian Rodriguez, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 1, 2, 7 +[33] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In CVPR, 2021. 2, 6, 7 +[34] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In CVPR, 2023. 1, 2 +[35] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 5 +[36] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, 2023. 3 +[37] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In CVPR, 2023. 2 +[38] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, 2022. 1, 2 +[39] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, 2019. 2, 6, 7 +[40] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 3 +[41] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In CVPR, 2023. 2 +[42] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 6 +[43] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[44] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2016. 1 +[45] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, 2021. 2 + +[46] Obin Kwon, Jeongho Park, and Songhwai Oh. Rendering neural radiance map for visual navigation. In CVPR, 2023. 2 +[47] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE TPAMI, 2023. 2, 3 +[48] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurIPS, 2023. 2 +[49] Jialu Li, Hao Tan, and Mohit Bansal. Envide: Environment editing for vision-and-language navigation. In CVPR, 2022. 2 +[50] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In CoRL, 2022. 2 +[51] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 2, 3, 4, 8 +[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 4, 6 +[53] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In CVPR, 2021. 6, 7 +[54] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, 2021. 2 +[55] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, 2023. 1, 2, 6, 7 +[56] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS, 2018. 4 +[57] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zero-shot video grounding with pseudo query lookup and verification. IEEE TIP, 33:1643-1654, 2024. 2 +[58] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, 2020. 2 +[59] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[60] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. In NeurIPS, 2021. 6 +[61] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In ICLR, 2018. 1 + +[62] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, 2021. 1 +[63] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3 +[64] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR, 2020. 1, 2, 6, 7, 8 +[65] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop: history-and-order aware pretraining for vision-and-language navigation. In CVPR, 2022. 2, 6, 7 +[66] Ruijie Quan, Linchao Zhu, Yu Wu, and Yi Yang. Holistic lstm for pedestrian trajectory prediction. IEEE TIP, 30: 3229-3239, 2021. 2 +[67] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4 +[68] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. IJCV, 130(8):1978-2005, 2022. 2, 3, 4 +[69] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 6 +[70] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2, 3, 4 +[71] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 2 +[72] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In NeurIPS, 2014. 1, 2 +[73] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2, 6 +[74] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In NAACL, 2019. 2, 6 +[75] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Self-supervised 3d semantic representation learning for vision-and-language navigation. arXiv:2201.10788, 2022. 1, 2 +[76] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 2, 4 +[77] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv:2304.14365, 2023. 2, 4, 5 +[78] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 3, 8 + +[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2 +[80] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In CVPR, 2020. 2 +[81] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, 2020. 6 +[82] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, 2021. 1, 2, 6, 7 +[83] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. In NeurIPS, 2022. 2 +[84] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, 2022. 2, 6 +[85] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, 2023. 2 +[86] Hanqing Wang, Wenguan Wang, Wei Liang, Steven CH Hoi, Jianbing Shen, and Luc Van Gool. Active perception for visual-language navigation. IJCV, 131(3):607-625, 2023. 2 +[87] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 3 +[88] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 2, 7 +[89] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023. 6, 7 +[90] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Learning to follow and generate instructions for language-capable navigation. IEEE TPAMI, 2023. 2 +[91] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, 2020. 5 +[92] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In CVPR, 2024. 4 +[93] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV, 2023. 2 +[94] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In ICCV, 2023. 1, 2, 6, 7 + +[95] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, 2023. 4, 5 +[96] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, 2021. 2 +[97] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv:2307.13510, 2023. 4 +[98] Zhenjia Xu, Zhanpeng He, Jiajun Wu, and Shuran Song. Learning 3d dynamic scene representations for robot manipulation. In CoRL, 2020. 3, 5, 6 +[99] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, 2018. 4 +[100] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In ACM MM, 2022. 2, 6, 7 +[101] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 2 +[102] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv: 1904.07850, 2019. 5 +[103] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, 2020. 6 +[104] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 4 +[105] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In CVPR, 2018. 5 +[106] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 2, 3, 5 \ No newline at end of file diff --git a/2024/Volumetric Environment Representation for Vision-Language Navigation/images.zip b/2024/Volumetric Environment Representation for Vision-Language Navigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5a0e5c3c13c8c4dc8ac12ea8cf261fd405ce4bf9 --- /dev/null +++ b/2024/Volumetric Environment Representation for Vision-Language Navigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e118d6b1f8731be5979a2b345e38a18396add343e1f1c0fb591e311326a26c0 +size 559491 diff --git a/2024/Volumetric Environment Representation for Vision-Language Navigation/layout.json b/2024/Volumetric Environment Representation for Vision-Language Navigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..36184a67a78650c849e13f6dc6ded6950d66e5e6 --- /dev/null +++ b/2024/Volumetric Environment Representation for Vision-Language Navigation/layout.json @@ -0,0 +1,12851 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 128, + 503, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 128, + 503, + 144 + ], + "spans": [ + { + "bbox": [ + 91, + 128, + 503, + 144 + ], + "type": "text", + "content": "Volumetric Environment Representation for Vision-Language Navigation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 165, + 385, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 165, + 385, + 178 + ], + "spans": [ + { + "bbox": [ + 212, + 165, + 385, + 178 + ], + "type": "text", + "content": "Rui Liu Wenguan Wang Yi Yang*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 237, + 179, + 358, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 179, + 358, + 190 + ], + "spans": [ + { + "bbox": [ + 237, + 179, + 358, + 190 + ], + "type": "text", + "content": "ReLER, CCAI, Zhejiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 204, + 192, + 389, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 192, + 389, + 202 + ], + "spans": [ + { + "bbox": [ + 204, + 192, + 389, + 202 + ], + "type": "text", + "content": "https://github.com/DefaultRui/VLN-VER" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 251, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 251, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 290, + 515 + ], + "type": "text", + "content": "Vision-language navigation (VLN) requires an agent to navigate through an 3D environment based on visual observations and natural language instructions. It is clear that the pivotal factor for successful navigation lies in the comprehensive scene understanding. Previous VLN agents employ monocular frameworks to extract 2D features of perspective views directly. Though straightforward, they struggle for capturing 3D geometry and semantics, leading to a partial and incomplete environment representation. To achieve a comprehensive 3D representation with fine-grained details, we introduce a Volumetric Environment Representation (VER), which voxelizes the physical world into structured 3D cells. For each cell, VER aggregates multi-view 2D features into such a unified 3D space via 2D-3D sampling. Through coarse-to-fine feature extraction and multi-task learning for VER, our agent predicts 3D occupancy, 3D room layout, and 3D bounding boxes jointly. Based on online collected VERs, our agent performs volume state estimation and builds episodic memory for predicting the next step. Experimental results show our environment representations from multi-task learning lead to evident performance gains on VLN. Our model achieves state-of-the-art performance across VLN benchmarks (R2R, REVERIE, and R4R)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 537, + 143, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 537, + 143, + 549 + ], + "spans": [ + { + "bbox": [ + 69, + 537, + 143, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 556, + 290, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 290, + 622 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 290, + 622 + ], + "type": "text", + "content": "Vision-language navigation (VLN) requires an agent to navigate in a 3D environment following natural language instructions [3, 64]. As a holistic understanding of the environment plays a pivotal role in decision-making within VLN, environment representation learning serves as a foundation for formulating accurate navigation policies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 623, + 288, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 623, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 67, + 623, + 288, + 667 + ], + "type": "text", + "content": "Early VLN approaches [3, 23] typically learn the navigation policy through the sequence-to-sequence (Seq2Seq) framework [72], which directly maps instructions and multi-view perspective observations to actions. They sim" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 227, + 395, + 308 + ], + "blocks": [ + { + "bbox": [ + 309, + 227, + 395, + 308 + ], + "lines": [ + { + "bbox": [ + 309, + 227, + 395, + 308 + ], + "spans": [ + { + "bbox": [ + 309, + 227, + 395, + 308 + ], + "type": "image", + "image_path": "5f19d63d8b8ddb11e9961a01e4077fc09ef996fc26f68a809ff701d3ad8f3b53.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 228, + 454, + 237 + ], + "lines": [ + { + "bbox": [ + 398, + 228, + 454, + 237 + ], + "spans": [ + { + "bbox": [ + 398, + 228, + 454, + 237 + ], + "type": "text", + "content": "Previous Methods" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 238, + 451, + 245 + ], + "lines": [ + { + "bbox": [ + 399, + 238, + 451, + 245 + ], + "spans": [ + { + "bbox": [ + 399, + 238, + 451, + 245 + ], + "type": "text", + "content": "Perspective" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 245, + 451, + 253 + ], + "lines": [ + { + "bbox": [ + 399, + 245, + 451, + 253 + ], + "spans": [ + { + "bbox": [ + 399, + 245, + 451, + 253 + ], + "type": "text", + "content": "Observations" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 253, + 463, + 261 + ], + "lines": [ + { + "bbox": [ + 399, + 253, + 463, + 261 + ], + "spans": [ + { + "bbox": [ + 399, + 253, + 463, + 261 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 261, + 463, + 269 + ], + "lines": [ + { + "bbox": [ + 399, + 261, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 399, + 261, + 463, + 269 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 269, + 463, + 277 + ], + "lines": [ + { + "bbox": [ + 399, + 269, + 463, + 277 + ], + "spans": [ + { + "bbox": [ + 399, + 269, + 463, + 277 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 277, + 463, + 285 + ], + "lines": [ + { + "bbox": [ + 399, + 277, + 463, + 285 + ], + "spans": [ + { + "bbox": [ + 399, + 277, + 463, + 285 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 285, + 451, + 293 + ], + "lines": [ + { + "bbox": [ + 399, + 285, + 451, + 293 + ], + "spans": [ + { + "bbox": [ + 399, + 285, + 451, + 293 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 293, + 451, + 300 + ], + "lines": [ + { + "bbox": [ + 399, + 293, + 451, + 300 + ], + "spans": [ + { + "bbox": [ + 399, + 293, + 451, + 300 + ], + "type": "text", + "content": "2D Features of" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 300, + 451, + 308 + ], + "lines": [ + { + "bbox": [ + 399, + 300, + 451, + 308 + ], + "spans": [ + { + "bbox": [ + 399, + 300, + 451, + 308 + ], + "type": "text", + "content": "Candidate View" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 238, + 521, + 261 + ], + "lines": [ + { + "bbox": [ + 463, + 238, + 521, + 261 + ], + "spans": [ + { + "bbox": [ + 463, + 238, + 521, + 261 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 261, + 521, + 269 + ], + "lines": [ + { + "bbox": [ + 463, + 261, + 521, + 269 + ], + "spans": [ + { + "bbox": [ + 463, + 261, + 521, + 269 + ], + "type": "text", + "content": "Topological Graph" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 269, + 521, + 277 + ], + "lines": [ + { + "bbox": [ + 463, + 269, + 521, + 277 + ], + "spans": [ + { + "bbox": [ + 463, + 269, + 521, + 277 + ], + "type": "inline_equation", + "content": "\\rightarrow" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 277, + 521, + 285 + ], + "lines": [ + { + "bbox": [ + 463, + 277, + 521, + 285 + ], + "spans": [ + { + "bbox": [ + 463, + 277, + 521, + 285 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 285, + 521, + 293 + ], + "lines": [ + { + "bbox": [ + 463, + 285, + 521, + 293 + ], + "spans": [ + { + "bbox": [ + 463, + 285, + 521, + 293 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}x - 1 > 0" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 293, + 521, + 300 + ], + "lines": [ + { + "bbox": [ + 463, + 293, + 521, + 300 + ], + "spans": [ + { + "bbox": [ + 463, + 293, + 521, + 300 + ], + "type": "text", + "content": "Semantic Map" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 300, + 521, + 308 + ], + "lines": [ + { + "bbox": [ + 463, + 300, + 521, + 308 + ], + "spans": [ + { + "bbox": [ + 463, + 300, + 521, + 308 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 317, + 355, + 324 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 355, + 324 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 355, + 324 + ], + "type": "text", + "content": "2D-3D Sampling" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 324, + 355, + 332 + ], + "lines": [ + { + "bbox": [ + 317, + 324, + 355, + 332 + ], + "spans": [ + { + "bbox": [ + 317, + 324, + 355, + 332 + ], + "type": "text", + "content": "23. 15 sampling" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 332, + 355, + 340 + ], + "lines": [ + { + "bbox": [ + 317, + 332, + 355, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 355, + 340 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 340, + 355, + 348 + ], + "lines": [ + { + "bbox": [ + 317, + 340, + 355, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 340, + 355, + 348 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 348, + 355, + 356 + ], + "lines": [ + { + "bbox": [ + 317, + 348, + 355, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 348, + 355, + 356 + ], + "type": "text", + "content": "multi-view" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 356, + 355, + 361 + ], + "lines": [ + { + "bbox": [ + 317, + 356, + 355, + 361 + ], + "spans": [ + { + "bbox": [ + 317, + 356, + 355, + 361 + ], + "type": "text", + "content": "D Features" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 308, + 459, + 316 + ], + "lines": [ + { + "bbox": [ + 399, + 308, + 459, + 316 + ], + "spans": [ + { + "bbox": [ + 399, + 308, + 459, + 316 + ], + "type": "inline_equation", + "content": "\\therefore m - 1 \\neq 0" + }, + { + "bbox": [ + 399, + 308, + 459, + 316 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 316, + 459, + 324 + ], + "lines": [ + { + "bbox": [ + 399, + 316, + 459, + 324 + ], + "spans": [ + { + "bbox": [ + 399, + 316, + 459, + 324 + ], + "type": "text", + "content": "3D Occupancy" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 324, + 459, + 332 + ], + "lines": [ + { + "bbox": [ + 399, + 324, + 459, + 332 + ], + "spans": [ + { + "bbox": [ + 399, + 324, + 459, + 332 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 332, + 459, + 340 + ], + "lines": [ + { + "bbox": [ + 399, + 332, + 459, + 340 + ], + "spans": [ + { + "bbox": [ + 399, + 332, + 459, + 340 + ], + "type": "text", + "content": "3D Detection -" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 340, + 457, + 348 + ], + "lines": [ + { + "bbox": [ + 399, + 340, + 457, + 348 + ], + "spans": [ + { + "bbox": [ + 399, + 340, + 457, + 348 + ], + "type": "inline_equation", + "content": "\\therefore m : x = 1" + }, + { + "bbox": [ + 399, + 340, + 457, + 348 + ], + "type": "text", + "content": " 或 " + }, + { + "bbox": [ + 399, + 340, + 457, + 348 + ], + "type": "inline_equation", + "content": "{3x} + {4y} + 1 = 0" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 348, + 457, + 356 + ], + "lines": [ + { + "bbox": [ + 399, + 348, + 457, + 356 + ], + "spans": [ + { + "bbox": [ + 399, + 348, + 457, + 356 + ], + "type": "text", + "content": "Room Layout" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 356, + 457, + 361 + ], + "lines": [ + { + "bbox": [ + 399, + 356, + 457, + 361 + ], + "spans": [ + { + "bbox": [ + 399, + 356, + 457, + 361 + ], + "type": "text", + "content": "Estimation" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 308, + 521, + 348 + ], + "lines": [ + { + "bbox": [ + 463, + 308, + 521, + 348 + ], + "spans": [ + { + "bbox": [ + 463, + 308, + 521, + 348 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 348, + 521, + 356 + ], + "lines": [ + { + "bbox": [ + 463, + 348, + 521, + 356 + ], + "spans": [ + { + "bbox": [ + 463, + 348, + 521, + 356 + ], + "type": "text", + "content": "Volume State Space" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 356, + 521, + 361 + ], + "lines": [ + { + "bbox": [ + 463, + 356, + 521, + 361 + ], + "spans": [ + { + "bbox": [ + 463, + 356, + 521, + 361 + ], + "type": "text", + "content": "Volume State Space" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 369, + 526, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 369, + 526, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 526, + 440 + ], + "type": "text", + "content": "Figure 1. The agent observes its surroundings with corresponding perspective features of different candidate views (O). Previous methods construct the topological graph or semantic map based on these 2D features. Our VER aggregates the multi-view features into structured 3D cells via 2D-3D sampling. VER is a powerful representation for both 3D perception tasks and VLN, providing a volume state space for decision-making." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 308, + 350, + 316 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 308, + 350, + 316 + ], + "spans": [ + { + "bbox": [ + 310, + 308, + 350, + 316 + ], + "type": "text", + "content": "Our Method" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 310, + 308, + 350, + 316 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 308, + 350, + 316 + ], + "spans": [ + { + "bbox": [ + 310, + 308, + 350, + 316 + ], + "type": "text", + "content": "Our Method" + } + ] + } + ], + "index": 32, + "type": "text" + }, + { + "bbox": [ + 304, + 456, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 526, + 687 + ], + "type": "text", + "content": "ply embed their immediate observation of local environment into the hidden states of recurrent units. As a result, they lack of explicit environment representations and struggle to access their past states during long-time exploration [61, 82]. To address this issue, later VLN agents are equipped with an external memory module [61], which stores the environment representations distinctly from navigation states. In this way, they can explicitly model and maintain the environment layouts and contents in a form of topological graph [2, 11, 16, 19, 32, 62] or semantic map [1, 4, 10, 12, 26, 34, 38, 55, 75, 94] (Fig. 1). Despite their promising performance with advanced frameworks (e.g., graph neural network [44] and Transformer [79]), their environment representations are still built upon 2D perspective features extracted by monocular frameworks. While straightforward, they compress depth information onto the perspective plane, sacrificing the integral scene structure in the 3D space. Thus, they encounter challenges in capturing 3D geometry and semantics in complex scenes. Such an incomplete environment representation easily leads to suboptimal navigation decisions." + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "spans": [ + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "spans": [ + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 676, + 179, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 676, + 179, + 686 + ], + "spans": [ + { + "bbox": [ + 79, + 676, + 179, + 686 + ], + "type": "text", + "content": "*Corresponding author: Yi Yang." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16317" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 100, + 289, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 100, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 67, + 100, + 289, + 340 + ], + "type": "text", + "content": "In this article, we propose a Volumetric Environment Representation (VER) that quantizes the physical world into structured 3D cells (Fig. 1). These cells, arranged within a predefined volumetric space, maintain both height and depth dimensions. Each cell corresponds to local context of the 3D space. VER aggregates multi-perspective 2D features within these cells through an environment encoder (§3.1). Compared to previous partial representations derived from hidden states and external memory, our VER captures the full geometry and semantics of the physical world. These 3D cells store the properties of the corresponding space in the scene by predicting 3D occupancy [70, 77], room layout [106], and 3D object boxes [51]. However, directly reconstructing the high-quality VER from 2D perspective views is challenging to capture the fine-grained details. As a response, we propose a coarse-to-fine VER extraction architecture, which uses learnable up-sampling operations to construct the representations progressively. It is supervised by multi-resolution semantic labels at different scales, utilizing the coarse-to-fine representations as hierarchical inputs. The annotations of the 3D tasks are collected for multi-task learning (§3.4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 341, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 341, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 67, + 341, + 289, + 548 + ], + "type": "text", + "content": "At each navigation step, our agent initially encodes the multi-view observations into VER (§3.2). With VER, instructions can be more effectively grounded in the 3D context. This is achieved by establishing cross-modal correlations between linguistic words and 3D cells of VER. Based on the correlations, a volume state estimation module is proposed to calculate transition probabilities over the surrounding cells. With the help of this module, our agent performs comprehensive decision-making in volumetric space, and then maps the volume state into local action space. In addition, an episodic memory module is established to online collect the information of observed viewpoints and build a topological graph providing global action space (§3.3). The node embeddings in the graph are from neighbor pillar representations in VER corresponding to the respective viewpoints. To balance the long-range action reasoning and language grounding, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": "Our agent is evaluated on three VLN benchmarks, i.e., R2R [3], REVERIE [64], and R4R [39] (" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "\\S 4.1" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": "). It yields solid performance gains (about " + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": " SR and " + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": " SPL on R2R test, " + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": " SR and " + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": " SPL on REVERIE val unseen). The ablation study confirms the efficacy of core model designs (" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": "). Additional results show our model achieves promising performance in 3D occupancy prediction, 3D detection, and room layout estimation (" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "inline_equation", + "content": "\\S 4.3" + }, + { + "bbox": [ + 67, + 548, + 288, + 635 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 646, + 149, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 646, + 149, + 657 + ], + "spans": [ + { + "bbox": [ + 69, + 646, + 149, + 657 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 664, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 289, + 687 + ], + "type": "text", + "content": "Vision-Language Navigation (VLN). Early VLN agents [3, 23] are built upon Seq2Seq [66, 72] framework to re" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 100, + 526, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 526, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 526, + 384 + ], + "type": "text", + "content": "serve the observation history in hidden state. Thus they struggle to capture long-range context as the path length increases. Later efforts are devoted to multimodal representation learning, navigation strategy learning, and data generation [27]. As a primary step, multimodal representation learning helps agents understand the environments and establish relations between the instructions and visual observations. Inspired by the success of vision-language pretraining [58, 65, 73], recent approaches [14, 33] use transformer-based architectures [42, 79] for joint visual and textual representations. Some attempts further exploit the visual information by modeling semantic relation [37] and spatial information [11, 19, 32, 83, 86, 100]. For navigation strategy learning, many VLN models [24, 88] use imitation and reinforcement learning-based training strategies. Previous solutions [45, 85] introduce world models [29] to perform mental simulations and make mental planning. Furthermore, the scarcity of human instructions and limited diversity of the scene hinder the agent to learn navigation policy and generalize to unseen environments well [54]. Therefore, several VLN data generation strategies have been proposed to create new trajectories from existing datasets [15, 28, 93], generate more instructions [41, 74, 84, 90], or create synthetic environments [48, 49]. In addition, driven by large models [5], existing agents [13, 57, 71, 101] demonstrate promising zero-shot performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 388, + 526, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 526, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 526, + 475 + ], + "type": "text", + "content": "Despite their outstanding contributions, most of them rely on 2D visual cues in perspective observations. These representations are constrained by occlusion and limited geometric information, especially in complex scenes. In this paper, we propose VER, a unified environment representation learned by 3D perception tasks. During navigation, our agent performs volume state estimation on VER, facilitating comprehensive decision-making within the 3D space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 478, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 526, + 687 + ], + "type": "text", + "content": "Environment Representation. Existing VLN models introduce various representations for environment modeling, including topological graphs [2, 11, 16, 19, 32, 82], and semantic spatial representations [1, 4, 12, 26, 34, 38, 55, 75, 94]. With a broader view, diverse representations have been proposed for robotics and autonomous driving [25, 47, 68, 76, 96]. In the early stage, 2D occupancy grid maps [22] model occupied and free space in the surroundings based on Bayesian estimation for robot navigation. Classic SLAM systems [21] construct a map directly by integrating information from various sensors, including LiDAR and cameras. However, the representations in SLAM still rely on primitives such as 3D point clouds and image patches. Some efforts [6, 8, 31] focus on developing learnable semantic map representations. To enhance spatial reasoning, scene graph representations [9, 80] define the topological relations between spatial elements of the environment. In addition, neural scene representations [46, 50] embed image observations into latent codes" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16318" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 88, + 521, + 190 + ], + "blocks": [ + { + "bbox": [ + 75, + 88, + 521, + 190 + ], + "lines": [ + { + "bbox": [ + 75, + 88, + 521, + 190 + ], + "spans": [ + { + "bbox": [ + 75, + 88, + 521, + 190 + ], + "type": "image", + "image_path": "3f94df0618eedd39947ddf9e7e46ca8f5ffecdd45e20241f110237182613679f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 199, + 527, + 251 + ], + "lines": [ + { + "bbox": [ + 67, + 199, + 527, + 251 + ], + "spans": [ + { + "bbox": [ + 67, + 199, + 527, + 251 + ], + "type": "text", + "content": "Figure 2. Overview of our model. Given the perspective features of candidate views, a group of 3D queries are used to sample and aggregate them into VER (§3.1). To encode VER, we adopt coarse-to-fine extraction and perform multi-task learning on 3D perception. Based on VER, a volume state estimation module is proposed to predict state transition (§3.2). The episodic memory is used to store past observations using neighboring pillar representations for each viewpoint (§3.3). For decision-making, our agent combines both the local action probabilities from the volume state and the global action probabilities obtained from the episodic memory. See §3 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 268, + 289, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 268, + 289, + 345 + ], + "spans": [ + { + "bbox": [ + 68, + 268, + 289, + 345 + ], + "type": "text", + "content": "for object categories, showcasing scalability to large scenes [40, 59]. Vision-centric BEV perception, which transforms perspective-view inputs to BEV grid representations, has recently received increasing attention [47, 51, 63]. As the BEV representations simplify the vertical geometry, 3D occupancy prediction [36, 78] is further proposed to infer the 3D geometry from perspective images [68, 70, 87, 98]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 346, + 290, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 346, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 68, + 346, + 290, + 456 + ], + "type": "text", + "content": "However, existing VLN agents mainly employ abstract relations or compressed spatial maps, lacking the ability to access complete scene information. Providing more world context can be beneficial for the subsequent decision-making and policy learning. Motivated by this insight, we explore a holistic environment representation VER that voxelizes the world into structured 3D cells. VER captures both semantic information and geometric details of the whole scene. Building upon VER, our agent is able to predict the 3D occupancy, room layout, and 3D boxes accurately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 479, + 130, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 479, + 130, + 491 + ], + "spans": [ + { + "bbox": [ + 69, + 479, + 130, + 491 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": "Problem Formulation. For brevity, we present the technical description in the context of R2R [3]. The navigable area of the environment is organized as an undirected graph, containing a set of nodes (viewpoints) and connectivity edges. In R2R, an embodied agent needs to navigate to a target location in the 3D environment following human instructions with " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " words (embedded as " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{D_w \\times L}" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "D_w" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " is the channel dimension). At time step " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": ", the agent looks around and obtains multi-view observations of its surrounding scene from the current location. Each view is represented by a 2D visual feature " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "F_t^{2d} \\in \\mathbb{R}^{D_i \\times H \\times W}" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " are the spatial shape of image plane, " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " denotes the channel dimension. The local action space " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "A_t \\in \\mathbb{R}^{N_t + 1}" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " candidate views, which correspond to neighboring navigable nodes " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "\\{v_{t,n}^*\\}_{n=1}^{N_t}" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": ", as well as a [STOP] action. Previous agents predict the action probabilities " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "p_t^{2d} \\in \\mathbb{R}^{N_t + 1}" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " directly based on " + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "inline_equation", + "content": "F_t^{2d}" + }, + { + "bbox": [ + 68, + 498, + 290, + 686 + ], + "type": "text", + "content": " of each candidate view. However," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 268, + 525, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 525, + 300 + ], + "type": "text", + "content": "these 2D features with limited geometric information are partial representations of the 3D environment, easily leading to suboptimal decision making." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": "Overview. To achieve comprehensive scene understanding, we introduce VER, which voxelizes the 3D world into structured 3D cells (Fig. 2). At step " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": ", an environment encoder is proposed to sample multi-view features " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "(F_{t}^{2\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " of each view) into the volumetric space of VER, forming a unified representation " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "F_{t}^{3\\mathrm{d}} \\in \\mathbb{R}^{D_{e} \\times X \\times Y \\times Z}" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " (\\$3.1). " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " are the shapes of horizontal plane, " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " reserves the height information of 3D space, and " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "D_{e}" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " represents the channel dimension. The volumetric space aligns with gravity in the world coordinate system based on the Manhattan assumption [106]. To encode VER, we devise coarse-to-fine extraction with multiple 3D perception tasks supervised by multi-resolution annotations (\\$3.4). Based on VER, a volume state estimation module is proposed to predict state transition probabilities " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "p_{t}^{3\\mathrm{d}} \\in \\mathbb{R}^{X \\times Y \\times Z}" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " over surrounding 3D cells (\\$3.2). With this module, our agent performs comprehensive decision-making in the 3D space and maps " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "p_{t}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "inline_equation", + "content": "p_{t}^{2\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 300, + 526, + 520 + ], + "type": "text", + "content": ". To predict the next step, our agent combines both the local action probabilities derived from the volume state and the global action probabilities obtained from the episodic memory (\\$3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 527, + 422, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 422, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 422, + 538 + ], + "type": "text", + "content": "3.1.Environment Encoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": "2D-3D Sampling. At step " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": ", the agent observes its surroundings and acquires the multi-view images. We introduce cross-view attention (CVA) to aggregate their features " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "(F^{2d}" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " for each view) into a unified volumetric representation " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "F^{3d}" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " with a group of learnable volume queries " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{D_e \\times X \\times Y \\times Z}" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " is omitted for simplicity). Specifically, for the 3D cell positioned at " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " within the egocentric world, a single query " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "Q(x,y,z) \\in \\mathbb{R}^{D_e}" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " is used to sample each image feature " + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "inline_equation", + "content": "F^{2d}" + }, + { + "bbox": [ + 304, + 544, + 526, + 643 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 336, + 646, + 525, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 646, + 525, + 659 + ], + "spans": [ + { + "bbox": [ + 336, + 646, + 525, + 659 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {3 d} (x, y, z) = \\operatorname {C V A} \\left(\\boldsymbol {Q} (x, y, z), \\boldsymbol {F} ^ {2 d} \\left(h ^ {\\prime}, w ^ {\\prime}\\right)\\right), \\tag {1}", + "image_path": "34922800f0ea352453f0e1eef2e055ef741e325d413c3b02dda90e96ab494306.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 664, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 525, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 664, + 525, + 687 + ], + "type": "inline_equation", + "content": "(h', w')" + }, + { + "bbox": [ + 304, + 664, + 525, + 687 + ], + "type": "text", + "content": " denotes the location of corresponding sampling point on the image plane. Note that we only show" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16319" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 99, + 286, + 202 + ], + "blocks": [ + { + "bbox": [ + 72, + 99, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 72, + 99, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 72, + 99, + 286, + 202 + ], + "type": "image", + "image_path": "2d216cda22b946887bfdb1b7ccd0e3a5ae64a419ee1d12997b24a489f0ffbc98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 212, + 289, + 253 + ], + "lines": [ + { + "bbox": [ + 68, + 212, + 289, + 253 + ], + "spans": [ + { + "bbox": [ + 68, + 212, + 289, + 253 + ], + "type": "text", + "content": "Figure 3. Our coarse-to-fine VER representation extraction (§3.1) adopts cascade up-sampling operations with 3D deconvolutions (Eq. 2) and 3D queries (Eq. 1). The training process is supervised at different scales by multi-resolution semantic labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "type": "text", + "content": "the formulation of a single sampling point for conciseness. Since the sampling strategy of vanilla cross-attention is computationally expensive, the deformable attention [51, 104] is introduced and extended in CVA. In this way, " + }, + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "type": "inline_equation", + "content": "Q(x,y,z)" + }, + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "type": "text", + "content": " selectively attends to a set of key sampling points around a reference instead of the entire " + }, + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "type": "inline_equation", + "content": "F^{2\\mathrm{d}}" + }, + { + "bbox": [ + 68, + 270, + 288, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "text", + "content": "Coarse-to-Fine VER Extraction Architecture. Directly recovering the fine-grained VER from perspective features easily leads to performance and efficiency degradation [17, 95]. The coarse-to-fine extraction is proposed to reconstruct VER progressively. Our approach involves cascade up-sampling operations (Fig. 3), dividing this extraction into " + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "text", + "content": " levels. At each level, 3D deconvolutions are utilized for lifting spatial resolution, and then CVA is used to query the multi-view 2D features (Eq. 1) for refining the detailed geometry. This enables the direct learning of details and avoids the inaccuracy of interpolation [56, 77, 92] (see Table 7). Between the input coarse feature " + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "inline_equation", + "content": "F^{3\\mathrm{d}(0)} \\in \\mathbb{R}^{D_e \\times \\frac{X}{2^M} \\times \\frac{Y}{2^M} \\times \\frac{Z}{2^M}}" + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "text", + "content": " from Eq. (1) and the target fine feature " + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "inline_equation", + "content": "F^{3\\mathrm{d}(M)} \\in \\mathbb{R}^{D_e \\times X \\times Y \\times Z}" + }, + { + "bbox": [ + 68, + 336, + 290, + 502 + ], + "type": "text", + "content": ", the intermediate features with varying shapes are calculated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 504, + 288, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 504, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 99, + 504, + 288, + 517 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {\\mathrm {3 d} (1)} = \\uparrow \\boldsymbol {F} ^ {\\mathrm {3 d} (0)}, \\dots , \\boldsymbol {F} ^ {\\mathrm {3 d} (M)} = \\uparrow \\boldsymbol {F} ^ {\\mathrm {3 d} (M - 1)}, \\tag {2}", + "image_path": "9a00d239038d1334631eeec8e198c9de8636603c556090672d16f5cb96ec8cd4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "inline_equation", + "content": "F^{3\\mathrm{d}(1)}, \\dots, F^{3\\mathrm{d}(M-1)}" + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "text", + "content": " denote the intermediate features from different levels, and " + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "text", + "content": " denotes the up-sampling. Multi-task Learning. Our VER offers a unified scene representation for various 3D perception tasks. Existing studies [68, 70] highlight that semantics and geometry are tightly intertwined. We train our environment encoder to extract VER under the supervision of multiple 3D perception tasks (§3.4). This process utilizes " + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "inline_equation", + "content": "\\{F^{3\\mathrm{d}(0)}, \\dots, F^{3\\mathrm{d}(M)}\\}" + }, + { + "bbox": [ + 68, + 521, + 289, + 686 + ], + "type": "text", + "content": " as inputs and is supervised at different scales (Fig. 3). For 3D occupancy prediction, the decoder is implemented as MLPs with the focal loss [52]. For 3D layout estimation, we adopt a query-based head to yield the Manhattan room layouts. A combination of the L1 loss and the IoU loss [67] is used as the training objective. For 3D detection, we employ a detection head to predict the 3D boxes [51]. The bipartite match" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 99, + 526, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 99, + 526, + 165 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 526, + 165 + ], + "type": "text", + "content": "ing and the bounding box loss [51, 104] are employed for detection. A weight vector " + }, + { + "bbox": [ + 304, + 99, + 526, + 165 + ], + "type": "inline_equation", + "content": "[2.0, 0.25, 0.25]" + }, + { + "bbox": [ + 304, + 99, + 526, + 165 + ], + "type": "text", + "content": " is used to balance the optimization of the three tasks, respectively. During navigation, the agent traverses between different viewpoints and encodes VERs through the frozen environment encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 172, + 432, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 172, + 432, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 172, + 432, + 182 + ], + "type": "text", + "content": "3.2. Volume State Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 189, + 525, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 525, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 525, + 276 + ], + "type": "text", + "content": "VLN task is typically viewed as a state estimation and transition problem [76]. With our VER, the agent state is represented as \"volume state\". As such, the state transition within the locally observed 3D environment, computed by Eq. (3)&(4), is referred as \"volume state estimation\". Different from previous plane-level state models [1, 4], VER introduces additional height dimension for 3D state estimation. This enables a more accurate action prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": "Volume State. At the beginning of a navigation episode, the agent is located at a start viewpoint " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "(x_0,y_0,z_0)" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": ". Based on its perception range, a volume state space " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{X}\\in \\mathbb{R}^{X\\times Y\\times Z}" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": " is defined corresponding to the 3D physical world with an initial state " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "s_0 = (x_0,y_0,z_0)" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": ". At step " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": ", the next intermediate state " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": " is determined by the instruction embeddings " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": " and VER " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": " for reaching the goal state " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "s_T" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "(0 < t < T)" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": ". As the entire environment is partially observable, the current state transition " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "(s_t\\rightarrow s_{t + 1}" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 277, + 526, + 386 + ], + "type": "text", + "content": ") is regarded as a local consideration for action prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": "State Estimation. A volume state estimation module is devised to predict the probability distribution " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^{3\\mathrm{d}} \\in \\mathbb{R}^{X \\times Y \\times Z}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": " of the intermediate state " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "s_{t+1}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": ". The environment representation is first reshaped as " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^{3\\mathrm{d}''} \\in \\mathbb{R}^{D_e \\times XYZ}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": ", and then adopt multi-layer transformers (MLT) to model the relations between " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^{3\\mathrm{d}''}" + }, + { + "bbox": [ + 304, + 386, + 525, + 453 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 455, + 525, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 455, + 525, + 469 + ], + "spans": [ + { + "bbox": [ + 350, + 455, + 525, + 469 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {F}} _ {t} ^ {3 d} = \\mathrm {M L T} \\left(\\left[ \\boldsymbol {E}; \\boldsymbol {F} _ {t} ^ {3 d ^ {\\prime}} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times X Y Z}, \\tag {3}", + "image_path": "5440f0c6bc9b1d899adf219e597dd78db18eaecdcfea7043ba6cc256bbea933d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_t^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "type": "text", + "content": " is the updated representations, and " + }, + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 304, + 474, + 525, + 506 + ], + "type": "text", + "content": " denotes the concatenation operation. MLT consists of stacked selfattention blocks. Then we use MLPs for state estimation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 509, + 525, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 509, + 525, + 522 + ], + "spans": [ + { + "bbox": [ + 339, + 509, + 525, + 522 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} ^ {\\mathrm {3 d}} = \\operatorname {S o f t m a x} \\left(\\mathrm {M L P} \\left(\\widetilde {\\boldsymbol {F}} _ {t} ^ {\\mathrm {3 d}}\\right)\\right) \\in [ 0, 1 ] ^ {X \\times Y \\times Z}. \\tag {4}", + "image_path": "43d452e7230de3c4d88a0c9c9ab0be0a6aa907e735d61bc9c53c913dd2aae0e0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "text", + "content": "Efficient Height-aware Group. The computational and memory efficiency of Eq. (3) is compromised due to the resolution of " + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "inline_equation", + "content": "F_{t}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "inline_equation", + "content": "XYZ \\gg L" + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "text", + "content": "). Considering the similarity and sparsity of information along the height direction [97, 99], we partition " + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "inline_equation", + "content": "F_{t}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "text", + "content": " into several uniform groups " + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "inline_equation", + "content": "\\{F_{t,z}^{g} \\in \\mathbb{R}^{D_e \\times XY}\\}_{z=1}^Z" + }, + { + "bbox": [ + 304, + 525, + 525, + 601 + ], + "type": "text", + "content": " along this axis. Then we apply MLT to each group, and Eq. (3) is reformulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 608, + 524, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 608, + 524, + 636 + ], + "spans": [ + { + "bbox": [ + 352, + 608, + 524, + 636 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {F}} _ {t, i} ^ {g} = \\operatorname {M L T} \\left(\\left[ \\boldsymbol {E}; \\mathbf {F} _ {t, i} ^ {g} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times X Y}, \\\\ \\widetilde {\\boldsymbol {F}} _ {t} ^ {3 \\mathrm {d}} = \\left\\{\\widetilde {\\boldsymbol {F}} _ {t, z} ^ {g} \\right\\} _ {z = 1} ^ {Z} \\in \\mathbb {R} ^ {D _ {e} \\times X Y Z}, \\tag {5} \\\\ \\end{array}", + "image_path": "a0e0a7ad354ce34e3a97fb4a720b08d5080080be6d556e0780fe3dd3fe633897.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 525, + 687 + ], + "type": "text", + "content": "where the weights of MLT are shared among different groups. The updated features from different groups " + }, + { + "bbox": [ + 304, + 642, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{F}_{t,z}^{g}\\}_{z = 1}^{Z}" + }, + { + "bbox": [ + 304, + 642, + 525, + 687 + ], + "type": "text", + "content": " are aggregated along the height to leverage complementary information. For ease of notation, the symbol" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16320" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 99, + 288, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 99, + 288, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 99, + 288, + 121 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_t^{3\\mathrm{d}}" + }, + { + "bbox": [ + 69, + 99, + 288, + 121 + ], + "type": "text", + "content": " is slightly reused for the gathered 3D representations. Then, " + }, + { + "bbox": [ + 69, + 99, + 288, + 121 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^{3\\mathrm{d}}" + }, + { + "bbox": [ + 69, + 99, + 288, + 121 + ], + "type": "text", + "content": " is calculated by Eq. (4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 129, + 165, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 129, + 165, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 129, + 165, + 140 + ], + "type": "text", + "content": "3.3. Action Prediction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "text", + "content": "For action prediction across the entire explored scene, a topological graph " + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t = \\{\\mathcal{V}_t,\\mathcal{E}_t\\}" + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "text", + "content": " is constructed and updated online to represent episodic memory during navigation. Specifically, " + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_t" + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "text", + "content": " denotes the observed viewpoints, i.e., all visited viewpoints and their candidate viewpoints. These viewpoints are encoded by compressing previous VERs. The edge " + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 68, + 146, + 288, + 255 + ], + "type": "text", + "content": " denotes the navigable connections between these viewpoints. To predict the next step, our agent combines both the volume state estimation and episodic navigation memory for decision-making." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "spans": [ + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "content": "Mapping Volume State to Action. As our agent navigates on the horizontal plane to reach the adjacent candidate viewpoints " + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "inline_equation", + "content": "\\{v_{n}^{*}\\}_{n = 1}^{N_{t}}" + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "content": ", we map the volume state space into 2D space to align with this movement pattern. Specifically, we average " + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "inline_equation", + "content": "p_t^{3d}" + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "content": " along the height (z-axis) axis as " + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "inline_equation", + "content": "p_t^h\\in \\mathbb{R}^{X\\times Y}" + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "content": ". Then we sum probability values in the neighborhood of " + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "inline_equation", + "content": "\\{v_{t,n}^{*}\\}_{n = 0}^{N_{t}}" + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "inline_equation", + "content": "v_{0}^{*}" + }, + { + "bbox": [ + 69, + 256, + 290, + 344 + ], + "type": "text", + "content": " for the current viewpoint, i.e., [STOP]), and normalize them as the local action probabilities:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 348, + 288, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 348, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 74, + 348, + 288, + 364 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} ^ {\\mathrm {2 d}} = \\left\\{\\sum \\boldsymbol {p} _ {t} ^ {h} \\left(x _ {n}, y _ {n}\\right) \\right\\} _ {n = 0} ^ {N _ {t}} \\in [ 0, 1 ] ^ {N _ {t} + 1}, \\left(x _ {n}, y _ {n}\\right) \\in \\Omega_ {n}, \\tag {6}", + "image_path": "a4e44aa6fd0b411b8578919d8130fe79014144fd93cf8b1f3978ef772c189dd4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^h (x_n,y_n)" + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "text", + "content": " is the value at the coordinate " + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "inline_equation", + "content": "(x_{n},y_{n})" + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\Omega_{n}" + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "text", + "content": " is the neighborhood of " + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "inline_equation", + "content": "v_{n}^{*}" + }, + { + "bbox": [ + 68, + 369, + 288, + 413 + ], + "type": "text", + "content": " in the horizontal plane. In the training stage, a heatmap [102] with a Gaussian kernel is used to supervise this action prediction (§3.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "text", + "content": "Global Action Prediction. The episodic memory module is used to store past environment representations and allows easy access to them. To memory the environment representations efficiently, we use the neighboring pillar [91] representations " + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{t,n}^{p} \\in \\mathbb{R}^{D_{e} \\times |\\Omega_{n}| \\times Z}" + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "text", + "content": ", corresponding to the current observed viewpoints " + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\{v_{n}^{*}\\}_{n=0}^{N_{t}}" + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 68, + 414, + 288, + 482 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 483, + 288, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 483, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 100, + 483, + 288, + 497 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {t, n} ^ {P} = \\left\\{\\tilde {\\boldsymbol {F}} _ {t} ^ {3 d} \\left(x _ {n}, y _ {n}, z _ {n}\\right) \\right\\} _ {\\left(x _ {n}, y _ {n}\\right) \\in \\Omega_ {n}, z _ {n} \\in [ 1, Z ]} \\tag {7}", + "image_path": "a239641a7e700287b1b142e9bd46f450d2272073d1db3eb54ff27a1d67395b5d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_t^{3\\mathrm{d}}(x_n,y_n,z_n)" + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "content": " is the representation at position " + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "inline_equation", + "content": "(x_{n},y_{n},z_{n})" + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_t^{3\\mathrm{d}}" + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "content": " (Eq. 5). " + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "inline_equation", + "content": "\\{\\bar{F}_{t,n}^p\\in \\mathbb{R}^{D_e}\\}_{n = 0}^{N_t}" + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "content": " is obtained by average pooling as the corresponding node embeddings, which are then incorporated into " + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 69, + 505, + 288, + 561 + ], + "type": "text", + "content": ". For previously observed nodes, we compute the average of their features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "text", + "content": "The episodic memory " + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "text", + "content": ", which includes the observed viewpoints, offers a global action space " + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t^* \\in \\mathbb{R}^{|\\mathcal{V}_t|}" + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "text", + "content": ". This enables our agent to change its current navigation state easily by 'jumping' directly to another viewpoint, which may be even observed several steps ago. The global action probabilities on " + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 68, + 562, + 288, + 627 + ], + "type": "text", + "content": " are calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 631, + 288, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 631, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 111, + 631, + 288, + 649 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {G}} _ {t} = \\operatorname {M L T} \\left(\\left[ \\boldsymbol {E}; \\boldsymbol {G} _ {t} \\right]\\right) \\in \\mathbb {R} ^ {D _ {e} \\times | \\mathcal {V} _ {t} |}, \\tag {8}", + "image_path": "421ae29ebf098664be79f16c0cd34ddb8ea456a2267d8b540b9bcdc9fe2ab178.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 647, + 245, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 647, + 245, + 659 + ], + "spans": [ + { + "bbox": [ + 114, + 647, + 245, + 659 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} ^ {g} = \\operatorname {S o f t m a x} \\left(\\operatorname {M L P} \\left(\\hat {\\boldsymbol {G}} _ {t}\\right)\\right) \\in [ 0, 1 ] ^ {| \\nu_ {t} |},", + "image_path": "60f59a4f5dff895f47fb098e5a17b41f97fc67827201af822700b9f884525582.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "type": "text", + "content": " denotes the node embeddings of " + }, + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 69, + 664, + 288, + 686 + ], + "type": "text", + "content": ". The ultimate action probabilities are given as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 97, + 525, + 115 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 97, + 525, + 115 + ], + "spans": [ + { + "bbox": [ + 344, + 97, + 525, + 115 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} ^ {2 \\mathrm {d}} = \\left[ \\boldsymbol {p} _ {t} ^ {2 \\mathrm {d} \\rightarrow g}; \\boldsymbol {p} _ {t} ^ {2 \\mathrm {d}} \\right] \\in [ 0, 1 ] ^ {| \\mathcal {V} _ {t} |}, \\tag {9}", + "image_path": "46d437c33b03129b0fc198d55942da00e74db24aa51ca47589a6b42bfef4483f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 113, + 485, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 113, + 485, + 125 + ], + "spans": [ + { + "bbox": [ + 348, + 113, + 485, + 125 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} _ {t} ^ {g} = W _ {g} \\boldsymbol {p} _ {t} ^ {g} + (1 - W _ {g}) \\hat {\\boldsymbol {p}} _ {t} ^ {2 d} \\in [ 0, 1 ] ^ {| \\mathcal {V} _ {t} |},", + "image_path": "09fabd3a8345c4fac5f22d0570bbde3645e1b5c68d30506a082fa60bf8771fdf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "spans": [ + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^{2\\mathrm{d}\\rightarrow g} \\in \\mathbb{R}^{|\\mathcal{V}_t| - (N_t + 1)}" + }, + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "type": "text", + "content": " denotes the probabilities of global backtracking and we use the same value as the local [STOP] probability in Eq. (6); " + }, + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "type": "inline_equation", + "content": "W_{g}" + }, + { + "bbox": [ + 304, + 131, + 525, + 166 + ], + "type": "text", + "content": " is a learnable weight." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": "State Transition and Memory Update. After executing the action in " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t^*" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": ", our agent reaches the next viewpoint " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "v_{t + 1,0}^{*}" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": ", and will iteratively: (1) encode its current observation as " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "F_{t + 1}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": " through Eq. (1); (2) update its volume state as " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "s_{t + 1} = (x_{t + 1},y_{t + 1},z_{t + 1})" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": "; (3) add the node embeddings of " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "\\{v_{t + 1,n}^{*}\\}_{n = 0}^{N_{t + 1}}" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": " into the episodic memory " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{t + 1}" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": "; and (4) predict the next step with the updated episodic memory (i.e., " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{t + 1}" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": ") and volume state (i.e., " + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "inline_equation", + "content": "s_{t + 1}" + }, + { + "bbox": [ + 304, + 166, + 525, + 278 + ], + "type": "text", + "content": "). Our agent repeats the above process until it chooses the [STOP] action or reaches the maximum step limit." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 285, + 425, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 425, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 425, + 296 + ], + "type": "text", + "content": "3.4. Annotation Generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 302, + 525, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 525, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 525, + 401 + ], + "type": "text", + "content": "A multi-task learning framework is proposed to extract and encode the VERs (§3.1). To achieve this, we generate annotations on Matterport3D dataset [7] for 3D occupancy prediction, object detection, and room layout estimation. We design a room-object-voxel pipeline to automatically generate these annotations. This pipeline leverages existing LiDAR point labels without additional human annotations (more details in Appendix). We utilize the egocentric observations with multi-view images as input." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 402, + 525, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 525, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 525, + 500 + ], + "type": "text", + "content": "Room Layout. The room layout in our context specifies the positions, orientations, and heights of the walls, relative to the camera center. It aims to reconstruct cuboid room shapes within the Manhattan world [18]. Given that the horizontal plane is aligned on the " + }, + { + "bbox": [ + 304, + 402, + 525, + 500 + ], + "type": "inline_equation", + "content": "x - z" + }, + { + "bbox": [ + 304, + 402, + 525, + 500 + ], + "type": "text", + "content": " axis, we parameterize the layout with center coordinate, width, length, height, and rotation. In contrast to directly operating on a single panoramic image [105, 106], we use the embodied observations with multi-view images as input." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 500, + 525, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 525, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 525, + 598 + ], + "type": "text", + "content": "Object Detection. Based on the room layout, we collect the surrounding objects if the agent locates in a room. In a non-closed environment, we collect information about nearby objects based on their distance from the agent. For each object, the eigenvectors of its vertices are used to define an oriented bounding box that tightly encloses the object [35]. Considering some objects may disappear from view due to occlusion but still exist in the environment (referred to as permanence [98]), we also include them in the analysis." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 599, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 525, + 686 + ], + "type": "text", + "content": "Point Accumulation for Occupancy. To generate voxel labels for occupancy [77, 95], we accumulate the sparse LiDAR points and utilize 3D boxes. Given dense background and object points, we first voxelize the 3D space and label each voxel based on the majority vote of labelled points in that voxel. Due to the limited number of LiDAR points, we leverage the Nearest Neighbors algorithm to generate dense labels for remaining voxels by searching the nearest" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 315, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 315, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 315, + 715 + ], + "type": "text", + "content": "16321" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 99, + 288, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 99, + 288, + 154 + ], + "spans": [ + { + "bbox": [ + 68, + 99, + 288, + 154 + ], + "type": "text", + "content": "semantic label. Moreover, the agent infers the complete 3D occupancy of each object (amodal perception), including regions that are not directly observed. This attribute enables the agent to predict the entire object instead of only visible surfaces [98]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "spans": [ + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "text", + "content": "Statistics. For high-resolution labels, we define " + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "inline_equation", + "content": "120 \\times 120 \\times" + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "text", + "content": " 35 voxel grids in world coordinates, where the scene voxel size equals to " + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "text", + "content": " (see more details about multi-resolution labels in Appendix). We annotate over 50 billion voxels and 16 classes, including 11 foreground objects and 5 background stuffs. It comprises about " + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 68, + 156, + 290, + 265 + ], + "type": "text", + "content": " annotated bounding boxes and 1,500 room layouts within the scenes. These annotations follow the same train/val/test splits as R2R [3]. There are 61 scenes for train/val seen, 11 scenes for val unseen, and 18 scenes for test." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 273, + 191, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 273, + 191, + 285 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 191, + 285 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 290, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 290, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 68, + 290, + 288, + 399 + ], + "type": "text", + "content": "Initially, the environment encoder (§3.1) is introduced for VER through coarse-to-fine extraction. Then multi-task learning is performed across multiple 3D perception tasks, including 3D occupancy prediction, 3D layout estimation, and 3D detection. During navigation, our agent is equipped with the frozen environment encoder and predicts the next step. Following recent VLN practice [14, 16, 33], both offline pretraining and finetuning are adopted. In this section, we will mainly introduce the details of architecture and training (see more details in Appendix)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": "Environment Representation Learning. For the multiview images, we adopt ViT-B/16 [20] pretrained on ImageNet to extract features. The number of 3D volume queries is " + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "15 \\times 15 \\times 4" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": ". For each query, it is projected to sample 2D views according to intrinsic and extrinsic parameters of the camera. We set the perception range as " + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "[-6m, 6m]" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": " axis and " + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "[-1.5m, 2m]" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": " for the height (" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": " axis). We adopt six layers of CVA (Eq. 1) for 2D-3D sampling, and then use " + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": " cascade deconvolutions for up-sampling (Eq. 2). The feature dimension is 768 (i.e., " + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "inline_equation", + "content": "D_{i} = D_{w} = D_{e} = 768" + }, + { + "bbox": [ + 68, + 401, + 288, + 510 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 511, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 511, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 68, + 511, + 288, + 598 + ], + "type": "text", + "content": "Navigation Network. MLT with 4 layers is initialized from [73] for state estimation (Eq. 4) and global action prediction (Eq. 8), respectively. The range of neighborhood for each candidate is set as " + }, + { + "bbox": [ + 68, + 511, + 288, + 598 + ], + "type": "inline_equation", + "content": "|\\Omega_n| = 9" + }, + { + "bbox": [ + 68, + 511, + 288, + 598 + ], + "type": "text", + "content": ". The standard deviation of the Gaussian kernel for the heat map is set as 3.0. Based on this heat map, the focal loss [52] is used to supervise the local action prediction (Eq. 6). We also use a cross-entropy loss for the global action prediction (Eq. 9)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 598, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 598, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 598, + 288, + 686 + ], + "type": "text", + "content": "Pretraining. For R2R [3] and R4R [39], Masked Language Modeling [14, 42] and Single-step Action Prediction [14, 33] are adopted as auxiliary tasks on offline-sampled instruction-route pairs [30]. For REVERIE [64], an additional Object Grounding (OG) [16, 53] is used for object reasoning. During pretraining, we train the model with a batch size of 64 for " + }, + { + "bbox": [ + 68, + 598, + 288, + 686 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 68, + 598, + 288, + 686 + ], + "type": "text", + "content": " iterations, using Adam [43] optimizer with 1e-4 learning rate. Only one task is adopted" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 98, + 529, + 289 + ], + "blocks": [ + { + "bbox": [ + 312, + 98, + 529, + 289 + ], + "lines": [ + { + "bbox": [ + 312, + 98, + 529, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 98, + 529, + 289 + ], + "type": "table", + "html": "
ModelsR2R
val unseentest unseen
TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑
Seq2Seq [3]8.397.8122-8.137.852018
SF [23]-6.6235-14.826.623528
EnvDrop [74]10.705.22524811.665.235147
AuxRN [103]-5.285550-5.155551
Active [81]20.604.36584021.604.336041
RecBERT [33]12.013.93635712.354.096357
HAMT [14]11.462.29666112.273.936560
SOAT [60]12.154.28595312.264.495853
SSM [82]20.74.32624520.44.576146
CCC [84]-5.205046-5.305148
HOP [65]12.273.80645712.683.836459
DUET [16]13.943.31726014.733.656959
LANA [89]12.0-686212.6-6560
TD-STP [100]-3.227063-3.736761
BSG [55]14.902.89746214.863.197362
BEVbert [1]14.552.817564-3.137362
Ours14.832.80766515.232.747666
", + "image_path": "c33933339691c0e610aa0a6e7bf9dd1b1aa1828212e276cf849463c33141eed7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 293, + 520, + 303 + ], + "lines": [ + { + "bbox": [ + 310, + 293, + 520, + 303 + ], + "spans": [ + { + "bbox": [ + 310, + 293, + 520, + 303 + ], + "type": "text", + "content": "Table 1. Quantitative results on R2R [3] (more details in §4.1)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 319, + 485, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 485, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 485, + 330 + ], + "type": "text", + "content": "at each mini-batch with the same sampling ratio." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 331, + 525, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 525, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 525, + 384 + ], + "type": "text", + "content": "Finetuning. Following the standard protocol [1, 16, 94], we finetune the navigation network using Dagger [69] techniques. In addition, the OG loss [1, 16, 53] is employed on REVERIE. In this stage, we set the learning rate to 1e-5 and batch size to 8 with " + }, + { + "bbox": [ + 305, + 331, + 525, + 384 + ], + "type": "inline_equation", + "content": "20k" + }, + { + "bbox": [ + 305, + 331, + 525, + 384 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 385, + 525, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 525, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 525, + 451 + ], + "type": "text", + "content": "Inference. During the testing phase, the agent receives the multi-view images and encodes them as VERs through the frozen environment encoder (§3.1). Based on VERs, the agent performs volume state estimation (§3.2) and models episodic memory (§3.3). By combining both of them, the agent predicts the next step accurately until stops." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 452, + 525, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 525, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 525, + 484 + ], + "type": "text", + "content": "Reproducibility. Our model is implemented in PyTorch and trained on eight RTX 4090 GPUs with a 24GB memory per-card. Testing is conducted on the same machine." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 498, + 375, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 375, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 375, + 510 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 516, + 419, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 419, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 419, + 526 + ], + "type": "text", + "content": "4.1. Performance on VLN" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 533, + 525, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 525, + 642 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 525, + 642 + ], + "type": "text", + "content": "Datasets. The experiments are conducted on three datasets. R2R [3] contains 7,189 trajectories sampled from 90 real-world indoor scenes. It consists of " + }, + { + "bbox": [ + 305, + 533, + 525, + 642 + ], + "type": "inline_equation", + "content": "22k" + }, + { + "bbox": [ + 305, + 533, + 525, + 642 + ], + "type": "text", + "content": " human-annotated navigational instructions. The dataset is split into train, val seen, val unseen, and test unseen sets, which mainly focus on the generalization capability in unseen environments. REVERIE [64] contains high-level instructions which describe target locations and objects, with a focus on grounding remote target objects. R4R [39] is an extended version of R2R with longer trajectories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 643, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 643, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 643, + 525, + 686 + ], + "type": "text", + "content": "Evaluation Metrics. For R2R, Success Rate (SR), Trajectory Length (TL), Oracle Success Rate (OSR), Success rate weighted by Path Length (SPL), and Navigation Error (NE) are used. For REVERIE, Remote Grounding Success rate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16322" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 88, + 524, + 248 + ], + "blocks": [ + { + "bbox": [ + 75, + 88, + 524, + 248 + ], + "lines": [ + { + "bbox": [ + 75, + 88, + 524, + 248 + ], + "spans": [ + { + "bbox": [ + 75, + 88, + 524, + 248 + ], + "type": "table", + "html": "
Modelsval seenREVERIEtest unseen
TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑TL↓OSR↑SR↑SPL↑RGS↑RGSPL↑
RCM [88]10.7029.4423.3321.8216.2315.3611.9814.239.296.974.893.8910.6011.687.846.673.673.14
FAST-M [64]16.3555.1750.5345.5031.9729.6645.2828.2014.407.197.844.6739.0530.6319.8811.6111.286.08
SIA [53]13.6165.8561.9157.0845.9642.6541.5344.6731.5316.2822.4111.5648.6144.5630.8014.8519.029.20
RecBERT [33]13.4453.9051.7947.9638.2335.6116.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [28]15.1648.9847.0142.3432.7530.0118.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [14]12.7947.6543.2940.1927.2025.1814.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
HOP [65]13.8054.8853.7647.1938.6533.8516.4636.2431.7826.1118.8515.7316.3833.0630.1724.3417.6914.34
DUET [16]13.8673.8671.7563.9457.4151.1422.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
TD-STP [100]-------39.4834.8827.3221.1616.56-40.2635.8927.5119.8815.40
BEVBert [1]-76.1873.7265.3257.7051.73-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
GridMM [94]------23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
LANA [89]15.9174.2871.9462.7759.0250.3423.1852.9748.3133.8632.8622.7718.8357.2051.7236.4532.9522.85
BSG [55]15.2678.3676.1866.6961.5654.0224.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
Ours16.1380.4975.8366.1961.7156.2023.0361.0955.9839.6633.7123.7024.7462.2256.8238.7633.8823.19
", + "image_path": "f11c19036038233cf3e877dad30380869f383a61e265d7cbc4930195e4ad2217.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 274, + 293, + 394 + ], + "blocks": [ + { + "bbox": [ + 108, + 251, + 487, + 262 + ], + "lines": [ + { + "bbox": [ + 108, + 251, + 487, + 262 + ], + "spans": [ + { + "bbox": [ + 108, + 251, + 487, + 262 + ], + "type": "text", + "content": "Table 2. Quantitative comparison results on REVERIE [64]. ' -': unavailable statistics (see §4.1 for more details)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 274, + 293, + 394 + ], + "lines": [ + { + "bbox": [ + 76, + 274, + 293, + 394 + ], + "spans": [ + { + "bbox": [ + 76, + 274, + 293, + 394 + ], + "type": "table", + "html": "
ModelsR4R val unseen
NE↓SR↑CLS↑nDTW↑SDTW↑
SF [3]8.472430--
RCM [88]-29353013
EGP [19]8.0030443718
SSM [82]8.2732533919
RelGraph [32]7.4336414734
RecBERT [33]6.6744514530
HAMT [14]6.0945585032
BSG [55]6.1247595334
LANA [89]-43605232
Ours6.1047615433
", + "image_path": "085610d1312f93098f13f8f040e3e6a93467224efc066f41a842d950bb5fab36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 423, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 423, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 68, + 423, + 288, + 477 + ], + "type": "text", + "content": "(RGS), and Remote Grounding Success weighted by Path Length (RGSPL) are also employed for object grounding. For R4R, Coverage weighted by Length Score (CLS), normalized Dynamic Time Warping (nDTW), and Success rate weighted nDTW (SDTW) are used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "text", + "content": "Performance on R2R. Table 1 compares our model with the state-of-the-art models on R2R. As we find that our model yields SR of " + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "text", + "content": " and SPL of " + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "text", + "content": " on text unseen, which leads to promising gains of " + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 68, + 478, + 288, + 544 + ], + "type": "text", + "content": " over BEVBert [1], respectively. It verifies that using VER to represent the environment leads to better decision-making." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "type": "text", + "content": "Performance on REVERIE. Table 2 presents the comparison results on REVERIE. Compared with the recent state-of-the-art VLN agent [55], our agent improves SR and SPL by " + }, + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "type": "inline_equation", + "content": "3.86\\%" + }, + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "type": "inline_equation", + "content": "4.07\\%" + }, + { + "bbox": [ + 68, + 544, + 288, + 598 + ], + "type": "text", + "content": " on the val unseen split. This highlights the effectiveness of our architecture design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 598, + 288, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 598, + 288, + 642 + ], + "spans": [ + { + "bbox": [ + 68, + 598, + 288, + 642 + ], + "type": "text", + "content": "Performance on R4R. Table 3 shows results on R4R. Our approach outperforms others in most metrics with a promising gain on nDTW (i.e., " + }, + { + "bbox": [ + 68, + 598, + 288, + 642 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 68, + 598, + 288, + 642 + ], + "type": "text", + "content": "). This suggests the episodic memory module is able to retrieve the long-time context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 643, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 288, + 686 + ], + "type": "text", + "content": "Visual Results. Fig. 4 depicts one exemplar navigation episode from val unseen set of R2R. In this complex environment, there are many rooms with different objects and 3D layout. From the visualization of 3D occupancy pre" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 275, + 529, + 329 + ], + "blocks": [ + { + "bbox": [ + 72, + 398, + 285, + 408 + ], + "lines": [ + { + "bbox": [ + 72, + 398, + 285, + 408 + ], + "spans": [ + { + "bbox": [ + 72, + 398, + 285, + 408 + ], + "type": "text", + "content": "Table 3. Quantitative results on R4R [39] (more details in §4.1)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 275, + 529, + 329 + ], + "lines": [ + { + "bbox": [ + 312, + 275, + 529, + 329 + ], + "spans": [ + { + "bbox": [ + 312, + 275, + 529, + 329 + ], + "type": "table", + "html": "
ModelsREVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
w/o Volume State52.3134.9132.7572.7161.13
w/o Episodic Memory49.3333.7131.3668.2161.70
Full model55.9839.6633.7175.8065.37
", + "image_path": "63cdb91647543ea4a4838fa9410ca0145aaec1bdcd27109dc61431466ccddac2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 336, + 526, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 336, + 526, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 526, + 357 + ], + "type": "text", + "content": "Table 4. Ablation study of overall design on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 366, + 525, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 525, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 525, + 422 + ], + "type": "text", + "content": "dition at the key steps, we find the geometric details and semantics can be captured well by VER. The room layout estimation can help the agent to understand \"enter the bedroom\". Finally, our agent finds the \"bed\" and accomplishes the instruction successfully." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 428, + 425, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 425, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 425, + 439 + ], + "type": "text", + "content": "4.2. Diagnostic Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 525, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 525, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 525, + 476 + ], + "type": "text", + "content": "To thoroughly test the efficacy of crucial components of our model, we conduct a series of diagnostic studies on val unseen split of REVERIE and R2R." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "type": "text", + "content": "Overall Design. We first investigate the effectiveness of our core design. The results in Table 4 indicate that adding Volume State leads to a substantial performance gain (i.e., " + }, + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "type": "inline_equation", + "content": "3.67\\%" + }, + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "type": "text", + "content": " on SR). After using Episodic Memory, a higher score (i.e., " + }, + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "type": "inline_equation", + "content": "31.36\\% \\rightarrow 33.71\\%" + }, + { + "bbox": [ + 304, + 477, + 525, + 531 + ], + "type": "text", + "content": " on RGS) is achieved." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "type": "text", + "content": "Neighborhood Range of Viewpoints. We use the neighborhood of each viewpoint in the state space for local action prediction (Eq. 6). Moreover, corresponding pillar representations of the neighborhood are also used for node embeddings of the episodic memory (Eq. 7). From Table 5, the limited range of neighborhood is insufficient to represent the candidate viewpoint for navigation (e.g., " + }, + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "type": "inline_equation", + "content": "75.80\\% \\rightarrow 73.75\\%" + }, + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "type": "text", + "content": " of SR on R2R). However, too large neighborhood range will contain irrelevant information, leading to inferior performance (e.g., " + }, + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "type": "inline_equation", + "content": "55.98\\% \\rightarrow 53.49\\%" + }, + { + "bbox": [ + 304, + 532, + 525, + 642 + ], + "type": "text", + "content": " of SR on REVERIE)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 647, + 500, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 500, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 500, + 659 + ], + "type": "text", + "content": "4.3. Analysis on 3D Representation Learning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 664, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 664, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 664, + 525, + 686 + ], + "type": "text", + "content": "Evaluation Metric. Following standard protocols, we employ Intersection over Union (IoU) to evaluate the occu" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "16323" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 89, + 246, + 163 + ], + "blocks": [ + { + "bbox": [ + 76, + 89, + 246, + 163 + ], + "lines": [ + { + "bbox": [ + 76, + 89, + 246, + 163 + ], + "spans": [ + { + "bbox": [ + 76, + 89, + 246, + 163 + ], + "type": "image", + "image_path": "b4a6f24632285fe7044140dcc4020ce6d26b47870a7018f11edebb1103f923dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 166, + 169, + 214 + ], + "lines": [ + { + "bbox": [ + 72, + 166, + 169, + 214 + ], + "spans": [ + { + "bbox": [ + 72, + 166, + 169, + 214 + ], + "type": "text", + "content": "Turn around and go passed the couch. Cross the hall to the sitting room. Turn left and enter the bedroom on the left. Wait near the bed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 249, + 89, + 337, + 153 + ], + "blocks": [ + { + "bbox": [ + 249, + 89, + 337, + 153 + ], + "lines": [ + { + "bbox": [ + 249, + 89, + 337, + 153 + ], + "spans": [ + { + "bbox": [ + 249, + 89, + 337, + 153 + ], + "type": "image", + "image_path": "195d147b8ba1ef9d7fad116043938d71e4705a12a4934b6e382c6921c4aede83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 154, + 337, + 216 + ], + "blocks": [ + { + "bbox": [ + 249, + 154, + 337, + 216 + ], + "lines": [ + { + "bbox": [ + 249, + 154, + 337, + 216 + ], + "spans": [ + { + "bbox": [ + 249, + 154, + 337, + 216 + ], + "type": "image", + "image_path": "fa1038e304bdc59c608be196534f17d6bd8b7498e3becb779165840da106f647.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 217, + 525, + 246 + ], + "lines": [ + { + "bbox": [ + 69, + 217, + 525, + 246 + ], + "spans": [ + { + "bbox": [ + 69, + 217, + 525, + 246 + ], + "type": "text", + "content": "Figure 4. A representative visual result on val unseen of R2R [3]. We first visualize the 3D occupancy prediction at the key steps. In addition, we provide the prediction of 3D boxes and 3D room layout at step " + }, + { + "bbox": [ + 69, + 217, + 525, + 246 + ], + "type": "inline_equation", + "content": "⑦" + }, + { + "bbox": [ + 69, + 217, + 525, + 246 + ], + "type": "text", + "content": ". We find that VER can capture the geometric details of 'couch' and the structure of 'bedroom'. With VER, our agent easily finds the 'bed' and succeeds. See §4.1 for more details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 337, + 89, + 425, + 153 + ], + "blocks": [ + { + "bbox": [ + 337, + 89, + 425, + 153 + ], + "lines": [ + { + "bbox": [ + 337, + 89, + 425, + 153 + ], + "spans": [ + { + "bbox": [ + 337, + 89, + 425, + 153 + ], + "type": "image", + "image_path": "e5908d47a2d36734e0b9fd62d5d03fcbeb4b559cc792b31156a350c86376d249.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 337, + 154, + 426, + 215 + ], + "blocks": [ + { + "bbox": [ + 337, + 154, + 426, + 215 + ], + "lines": [ + { + "bbox": [ + 337, + 154, + 426, + 215 + ], + "spans": [ + { + "bbox": [ + 337, + 154, + 426, + 215 + ], + "type": "image", + "image_path": "899ef8050b2b013484e8c9e41202b1035aeae1c27a6cc31ba4154c0f4eee0efa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 426, + 89, + 521, + 154 + ], + "blocks": [ + { + "bbox": [ + 426, + 89, + 521, + 154 + ], + "lines": [ + { + "bbox": [ + 426, + 89, + 521, + 154 + ], + "spans": [ + { + "bbox": [ + 426, + 89, + 521, + 154 + ], + "type": "image", + "image_path": "bcba29e178db031f005031f1b115b7228239839994a0ce5b1a12bf0ff4084555.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 154, + 521, + 215 + ], + "blocks": [ + { + "bbox": [ + 427, + 154, + 521, + 215 + ], + "lines": [ + { + "bbox": [ + 427, + 154, + 521, + 215 + ], + "spans": [ + { + "bbox": [ + 427, + 154, + 521, + 215 + ], + "type": "image", + "image_path": "3b0e0ad220f3711e9200bc4cddcba689c5484c9818dd6876b84a1216d2c548aa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 75, + 251, + 288, + 316 + ], + "blocks": [ + { + "bbox": [ + 75, + 251, + 288, + 316 + ], + "lines": [ + { + "bbox": [ + 75, + 251, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 75, + 251, + 288, + 316 + ], + "type": "table", + "html": "
|Ωn|REVERIER2R
SR↑SPL↑RGS↑SR↑SPL↑
453.3035.9034.1673.7562.77
955.9839.6633.7175.8065.37
1652.1836.3733.1473.8263.49
2553.4935.8733.4674.6363.24
", + "image_path": "3b9c8bc5efdcb853e14c3e1407042d21422ed9cc31487973493b9e62f359b4e4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 75, + 351, + 291, + 403 + ], + "blocks": [ + { + "bbox": [ + 69, + 319, + 289, + 340 + ], + "lines": [ + { + "bbox": [ + 69, + 319, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 69, + 319, + 289, + 340 + ], + "type": "text", + "content": "Table 5. Ablation study of neighborhood range on val unseen of REVERIE [64] and R2R [3] (see §4.2 for more details)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 351, + 291, + 403 + ], + "lines": [ + { + "bbox": [ + 75, + 351, + 291, + 403 + ], + "spans": [ + { + "bbox": [ + 75, + 351, + 291, + 403 + ], + "type": "table", + "html": "
ModelsOccupancyDetectionLayout
IoU↑mIoU↑mAP↑mAR↑3D IoU↑
BEVFormer [51]20.388.9727.3043.8862.71
OccNet [78]22.1210.6629.9147.1564.04
Ours24.3112.9333.5751.6066.45
", + "image_path": "23fe3b696faea661afd5fa067f2c9190a3274320101deadcfb4f0f7765114069.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 443, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 443, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 68, + 443, + 288, + 509 + ], + "type": "text", + "content": "pancy prediction quality, regardless of the semantic labels. The mean IoU (mIoU) of 15 classes is also used to assess the performance of semantic occupancy. For 3D object detection, we utilize mean Average Precision (mAP) and mean Average Recall (mAR) with IoU thresholds of 0.50. For room layout, we adopt 3D IoU for cuboid layout evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 510, + 288, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 510, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 68, + 510, + 288, + 586 + ], + "type": "text", + "content": "Performance on 3D Tasks. In Table 6, our network (§3.1) outperforms other methods [51, 78] by a significant margin (2.19% on IoU of occupancy, 3.66% on mAP of 3D detection, and 2.41% on 3D IoU of room layout estimation). The mIoU of occupancy is also exhibits improvement (2.27%), underscoring the network's proficiency in capturing both scene geometry and fine-grained semantics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "type": "text", + "content": "Coarse-to-Fine Extraction. Table 7 lists the scores with different up-sampling operations (Eq. 2). Our approach improves the performance by solid margins (e.g., " + }, + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "type": "inline_equation", + "content": "11.03\\% \\rightarrow 12.93\\%" + }, + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "type": "text", + "content": " for 3D occupancy, " + }, + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "type": "inline_equation", + "content": "75.14\\% \\rightarrow 75.80\\%" + }, + { + "bbox": [ + 68, + 586, + 288, + 653 + ], + "type": "text", + "content": " on SR of R2R). This verifies the efficacy of our design of the coarse-to-fine extraction and learnable up-sampling operations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 68, + 653, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 653, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 653, + 289, + 687 + ], + "type": "text", + "content": "Multi-task Learning. Table 8 reports performance comparison with different perception tasks (§3.1). We find that multi-task learning yields a substantial performance gain." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 312, + 251, + 531, + 300 + ], + "blocks": [ + { + "bbox": [ + 69, + 407, + 288, + 427 + ], + "lines": [ + { + "bbox": [ + 69, + 407, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 69, + 407, + 288, + 427 + ], + "type": "text", + "content": "Table 6. Quantitative results on 3D occupancy, 3D detection, and room layout prediction (see §4.3 for more details)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 251, + 531, + 300 + ], + "lines": [ + { + "bbox": [ + 312, + 251, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 531, + 300 + ], + "type": "table", + "html": "
Up-Sampling3D PerceptionR2R
mIoU↑mAP↑3D IoU↑SR↑SPL↑
w/o Coarse-to-Fine12.3932.9566.57--
Trilinear Interpolation11.0329.4263.4575.1464.30
3D Deconvolution12.9333.5766.4575.8065.37
", + "image_path": "f886aab8ce16523699c8a8f6697e0c4281d563142c1a38784a74ee63a631ce54.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 312, + 342, + 528, + 416 + ], + "blocks": [ + { + "bbox": [ + 305, + 304, + 526, + 335 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 526, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 526, + 335 + ], + "type": "text", + "content": "Table 7. Ablation study of Coarse-to-Fine Extraction on occupancy prediction (mIoU), 3D detection (mAP), room layout (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 342, + 528, + 416 + ], + "lines": [ + { + "bbox": [ + 312, + 342, + 528, + 416 + ], + "spans": [ + { + "bbox": [ + 312, + 342, + 528, + 416 + ], + "type": "table", + "html": "
Multi-task Learning3D PerceptionR2R
Occ.Obj.Room.mIoU↑mAP↑3D IoU↑SR↑SPL↑
✓12.09--74.9063.82
✓✓12.1432.64-75.2164.79
✓✓-33.1164.5874.0363.51
✓✓11.37-63.2974.9764.66
✓✓✓12.9333.5766.4575.8065.37
", + "image_path": "799d26cf106cf3aeffcc3c4c19e35236296554a5c2fce8e74b41f2345086d628.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 421, + 525, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 421, + 525, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 525, + 452 + ], + "type": "text", + "content": "Table 8. Ablation study of Multi-task Learning on occupancy prediction (mIoU), 3D detection (mAP), room layout estimation (3D IoU), and val unseen set of R2R [3] (see §4.3 for more details)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 468, + 525, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 525, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 525, + 502 + ], + "type": "text", + "content": "This suggests these 3D perception tasks are complementary to each other in capturing geometric and semantic properties of scenes, further facilitating the decision-making." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 514, + 372, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 372, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 372, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 532, + 525, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 525, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 525, + 630 + ], + "type": "text", + "content": "In this paper, we propose a Volumetric Environment Representation (VER), which aggregates the perspective features into structured 3D cells. Through coarse-to-fine feature extraction, we can efficiently perform several 3D perception tasks. Based on this comprehensive representation, we develop the volume state for local action prediction and the episodic memory for retrieving the global context. We demonstrate that our agent achieves promising performance on VLN benchmarks (R2R, REVERIE, and R4R)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 632, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 525, + 687 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (No. 62372405), the Fundamental Research Funds for the Central Universities (No. 226-2022-00051), and CCF-Tencent Open Fund." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "16324" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 117, + 289, + 685 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 78, + 117, + 289, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 117, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 78, + 117, + 289, + 156 + ], + "type": "text", + "content": "[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. In ICCV, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 158, + 288, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 158, + 288, + 197 + ], + "spans": [ + { + "bbox": [ + 78, + 158, + 288, + 197 + ], + "type": "text", + "content": "[2] Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv:2304.03047, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 198, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 198, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 78, + 198, + 288, + 248 + ], + "type": "text", + "content": "[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, 2018. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 249, + 288, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 249, + 288, + 279 + ], + "spans": [ + { + "bbox": [ + 78, + 249, + 288, + 279 + ], + "type": "text", + "content": "[4] Peter Anderson, Ayush Shrivastava, Devi Parikh, Dhruv Batra, and Stefan Lee. Chasing ghosts: Instruction following as bayesian state tracking. In NeurIPS, 2019. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 280, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 280, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 78, + 280, + 288, + 329 + ], + "type": "text", + "content": "[5] Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuzhhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt4.arXiv:2303.12712, 2023.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 331, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 331, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 78, + 331, + 288, + 369 + ], + "type": "text", + "content": "[6] Vincent Cartillier, Zhile Ren, Neha Jain, Stefan Lee, Irfan Essa, and Dhruv Batra. Semantic mapnet: Building allocentric semantic maps and representations from egocentric views. In AAAI, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 371, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 371, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 78, + 371, + 288, + 411 + ], + "type": "text", + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 412, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 412, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 78, + 412, + 288, + 442 + ], + "type": "text", + "content": "[8] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In ICLR, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 443, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 443, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 78, + 443, + 288, + 472 + ], + "type": "text", + "content": "[9] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 473, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 473, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 74, + 473, + 288, + 513 + ], + "type": "text", + "content": "[10] Jinyu Chen, Wenguan Wang, Si Liu, Hongsheng Li, and Yi Yang. Omnidirectional information gathering for knowledge transfer-based audio-visual navigation. In ICCV, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 514, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 514, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 74, + 514, + 288, + 553 + ], + "type": "text", + "content": "[11] Kevin Chen, Junshen K Chen, Jo Chuang, Marynel Vázquez, and Silvio Savarese. Topological planning with transformers for vision-and-language navigation. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 74, + 555, + 288, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 555, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 74, + 555, + 288, + 594 + ], + "type": "text", + "content": "[12] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 74, + 595, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 595, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 74, + 595, + 288, + 645 + ], + "type": "text", + "content": "[13] Peihao Chen, Xinyu Sun, Hongyan Zhi, Runhao Zeng, Thomas H Li, Gaowen Liu, Mingkui Tan, and Chuang Gan. " + }, + { + "bbox": [ + 74, + 595, + 288, + 645 + ], + "type": "inline_equation", + "content": "a^2" + }, + { + "bbox": [ + 74, + 595, + 288, + 645 + ], + "type": "text", + "content": " nav: Action-aware zero-shot robot navigation by exploiting vision-and-language ability of foundation models. In NeurIPS Workshop, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 74, + 646, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 646, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 74, + 646, + 288, + 685 + ], + "type": "text", + "content": "[14] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 311, + 101, + 524, + 686 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 311, + 101, + 524, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 101, + 524, + 140 + ], + "spans": [ + { + "bbox": [ + 311, + 101, + 524, + 140 + ], + "type": "text", + "content": "[15] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 142, + 524, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 142, + 524, + 182 + ], + "spans": [ + { + "bbox": [ + 311, + 142, + 524, + 182 + ], + "type": "text", + "content": "[16] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 184, + 524, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 184, + 524, + 214 + ], + "spans": [ + { + "bbox": [ + 311, + 184, + 524, + 214 + ], + "type": "text", + "content": "[17] Ian Cherabier, Johannes L Schonberger, Martin R Oswald, Marc Pollefeys, and Andreas Geiger. Learning priors for semantic 3d reconstruction. In ECCV, 2018. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 216, + 524, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 216, + 524, + 246 + ], + "spans": [ + { + "bbox": [ + 311, + 216, + 524, + 246 + ], + "type": "text", + "content": "[18] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In ICCV, 1999. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 247, + 524, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 247, + 524, + 278 + ], + "spans": [ + { + "bbox": [ + 311, + 247, + 524, + 278 + ], + "type": "text", + "content": "[19] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In NeurIPS, 2020. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 279, + 524, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 279, + 524, + 329 + ], + "spans": [ + { + "bbox": [ + 311, + 279, + 524, + 329 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 331, + 524, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 331, + 524, + 361 + ], + "spans": [ + { + "bbox": [ + 311, + 331, + 524, + 361 + ], + "type": "text", + "content": "[21] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 363, + 524, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 363, + 524, + 392 + ], + "spans": [ + { + "bbox": [ + 311, + 363, + 524, + 392 + ], + "type": "text", + "content": "[22] Alberto Elfes. Occupancy grids: A stochastic spatial representation for active robot perception. arXiv:1304.1098, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 311, + 395, + 524, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 395, + 524, + 445 + ], + "spans": [ + { + "bbox": [ + 311, + 395, + 524, + 445 + ], + "type": "text", + "content": "[23] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, 2018. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 447, + 524, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 447, + 524, + 486 + ], + "spans": [ + { + "bbox": [ + 311, + 447, + 524, + 486 + ], + "type": "text", + "content": "[24] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 311, + 488, + 524, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 488, + 524, + 548 + ], + "spans": [ + { + "bbox": [ + 311, + 488, + 524, + 548 + ], + "type": "text", + "content": "[25] Sourav Garg, Niko Sunderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 311, + 550, + 524, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 550, + 524, + 591 + ], + "spans": [ + { + "bbox": [ + 311, + 550, + 524, + 591 + ], + "type": "text", + "content": "[26] Georgios Georgakis, Karl Schmeckpeper, Karan Wanchoo, Soham Dan, Eleni Miltsakaki, Dan Roth, and Kostas Dani-ilidis. Cross-modal map learning for vision and language navigation. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 311, + 592, + 524, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 592, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 311, + 592, + 524, + 622 + ], + "type": "text", + "content": "[27] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In ACL, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 311, + 624, + 524, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 524, + 663 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 524, + 663 + ], + "type": "text", + "content": "[28] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 311, + 666, + 524, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 666, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 311, + 666, + 524, + 686 + ], + "type": "text", + "content": "[29] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16325" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 74, + 100, + 289, + 685 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 74, + 100, + 289, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 100, + 289, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 100, + 289, + 140 + ], + "type": "text", + "content": "[30] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 141, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 141, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 74, + 141, + 288, + 170 + ], + "type": "text", + "content": "[31] Joao F Henriques and Andrea Vedaldi. Mapnet: An allocentric spatial memory for mapping environments. In CVPR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 171, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 171, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 74, + 171, + 288, + 201 + ], + "type": "text", + "content": "[32] Yicong Hong, Cristian Rodriguez, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 202, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 202, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 74, + 202, + 288, + 240 + ], + "type": "text", + "content": "[33] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In CVPR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 243, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 243, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 74, + 243, + 288, + 281 + ], + "type": "text", + "content": "[34] Yicong Hong, Yang Zhou, Ruiyi Zhang, Franck Dernoncourt, Trung Bui, Stephen Gould, and Hao Tan. Learning navigational visual representations with semantic map supervision. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 283, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 283, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 74, + 283, + 288, + 312 + ], + "type": "text", + "content": "[35] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 313, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 313, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 74, + 313, + 288, + 342 + ], + "type": "text", + "content": "[36] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 344, + 288, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 344, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 74, + 344, + 288, + 383 + ], + "type": "text", + "content": "[37] Jingyang Huo, Qiang Sun, Boyan Jiang, Haitao Lin, and Yanwei Fu. Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 384, + 288, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 384, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 74, + 384, + 288, + 433 + ], + "type": "text", + "content": "[38] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 434, + 288, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 434, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 74, + 434, + 288, + 473 + ], + "type": "text", + "content": "[39] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, 2019. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 475, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 475, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 74, + 475, + 288, + 513 + ], + "type": "text", + "content": "[40] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 514, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 514, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 74, + 514, + 288, + 564 + ], + "type": "text", + "content": "[41] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In CVPR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 74, + 566, + 288, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 566, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 74, + 566, + 288, + 604 + ], + "type": "text", + "content": "[42] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 74, + 605, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 605, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 74, + 605, + 288, + 624 + ], + "type": "text", + "content": "[43] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 74, + 625, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 625, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 74, + 625, + 288, + 654 + ], + "type": "text", + "content": "[44] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 74, + 656, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 656, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 74, + 656, + 288, + 685 + ], + "type": "text", + "content": "[45] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 100, + 524, + 685 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 312, + 100, + 524, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 100, + 524, + 129 + ], + "spans": [ + { + "bbox": [ + 312, + 100, + 524, + 129 + ], + "type": "text", + "content": "[46] Obin Kwon, Jeongho Park, and Songhwai Oh. Rendering neural radiance map for visual navigation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 132, + 524, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 132, + 524, + 181 + ], + "spans": [ + { + "bbox": [ + 312, + 132, + 524, + 181 + ], + "type": "text", + "content": "[47] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE TPAMI, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 182, + 524, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 182, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 182, + 524, + 213 + ], + "type": "text", + "content": "[48] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 214, + 524, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 214, + 524, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 214, + 524, + 243 + ], + "type": "text", + "content": "[49] Jialu Li, Hao Tan, and Mohit Bansal. Envide: Environment editing for vision-and-language navigation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 245, + 524, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 245, + 524, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 245, + 524, + 274 + ], + "type": "text", + "content": "[50] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In CoRL, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 276, + 524, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 524, + 324 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 524, + 324 + ], + "type": "text", + "content": "[51] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 2, 3, 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 327, + 524, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 524, + 356 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 524, + 356 + ], + "type": "text", + "content": "[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 357, + 524, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 357, + 524, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 357, + 524, + 387 + ], + "type": "text", + "content": "[53] Xiangru Lin, Guanbin Li, and Yizhou Yu. Scene-intuitive agent for remote embodied visual grounding. In CVPR, 2021. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 388, + 524, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 524, + 427 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 524, + 427 + ], + "type": "text", + "content": "[54] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 430, + 524, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 430, + 524, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 430, + 524, + 459 + ], + "type": "text", + "content": "[55] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 460, + 524, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 524, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 524, + 491 + ], + "type": "text", + "content": "[56] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS, 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 492, + 524, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 492, + 524, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 492, + 524, + 521 + ], + "type": "text", + "content": "[57] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zero-shot video grounding with pseudo query lookup and verification. IEEE TIP, 33:1643-1654, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 523, + 524, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 524, + 562 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 524, + 562 + ], + "type": "text", + "content": "[58] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 564, + 524, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 564, + 524, + 612 + ], + "spans": [ + { + "bbox": [ + 312, + 564, + 524, + 612 + ], + "type": "text", + "content": "[59] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 614, + 524, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 524, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 524, + 654 + ], + "type": "text", + "content": "[60] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. Soat: A scene-and object-aware transformer for vision-and-language navigation. In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 656, + 524, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 656, + 524, + 685 + ], + "spans": [ + { + "bbox": [ + 312, + 656, + 524, + 685 + ], + "type": "text", + "content": "[61] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In ICLR, 2018. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16326" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 74, + 100, + 288, + 686 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 74, + 100, + 288, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 100, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 74, + 100, + 288, + 130 + ], + "type": "text", + "content": "[62] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 131, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 131, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 74, + 131, + 288, + 161 + ], + "type": "text", + "content": "[63] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 162, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 162, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 74, + 162, + 288, + 212 + ], + "type": "text", + "content": "[64] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 213, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 213, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 74, + 213, + 288, + 253 + ], + "type": "text", + "content": "[65] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop: history-and-order aware pretraining for vision-and-language navigation. In CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 254, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 254, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 74, + 254, + 288, + 283 + ], + "type": "text", + "content": "[66] Ruijie Quan, Linchao Zhu, Yu Wu, and Yi Yang. Holistic lstm for pedestrian trajectory prediction. IEEE TIP, 30: 3229-3239, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 285, + 288, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 285, + 288, + 325 + ], + "spans": [ + { + "bbox": [ + 74, + 285, + 288, + 325 + ], + "type": "text", + "content": "[67] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 326, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 326, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 74, + 326, + 288, + 356 + ], + "type": "text", + "content": "[68] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. IJCV, 130(8):1978-2005, 2022. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 357, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 357, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 74, + 357, + 288, + 388 + ], + "type": "text", + "content": "[69] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 388, + 288, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 388, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 74, + 388, + 288, + 427 + ], + "type": "text", + "content": "[70] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 429, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 429, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 74, + 429, + 288, + 459 + ], + "type": "text", + "content": "[71] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 460, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 460, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 74, + 460, + 288, + 489 + ], + "type": "text", + "content": "[72] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In NeurIPS, 2014. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 491, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 491, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 74, + 491, + 288, + 520 + ], + "type": "text", + "content": "[73] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 74, + 522, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 522, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 74, + 522, + 288, + 552 + ], + "type": "text", + "content": "[74] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In NAACL, 2019. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 74, + 552, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 552, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 74, + 552, + 288, + 592 + ], + "type": "text", + "content": "[75] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Self-supervised 3d semantic representation learning for vision-and-language navigation. arXiv:2201.10788, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 74, + 594, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 594, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 74, + 594, + 288, + 613 + ], + "type": "text", + "content": "[76] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 74, + 615, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 615, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 74, + 615, + 288, + 654 + ], + "type": "text", + "content": "[77] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv:2304.14365, 2023. 2, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 74, + 655, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 655, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 74, + 655, + 288, + 686 + ], + "type": "text", + "content": "[78] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 3, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 100, + 525, + 686 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 312, + 100, + 525, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 100, + 525, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 100, + 525, + 140 + ], + "type": "text", + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 141, + 525, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 525, + 171 + ], + "type": "text", + "content": "[80] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In CVPR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 171, + 525, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 171, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 312, + 171, + 525, + 201 + ], + "type": "text", + "content": "[81] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 201, + 525, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 201, + 525, + 231 + ], + "spans": [ + { + "bbox": [ + 312, + 201, + 525, + 231 + ], + "type": "text", + "content": "[82] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 232, + 525, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 525, + 261 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 525, + 261 + ], + "type": "text", + "content": "[83] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 262, + 525, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 525, + 301 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 525, + 301 + ], + "type": "text", + "content": "[84] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 303, + 525, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 303, + 525, + 332 + ], + "spans": [ + { + "bbox": [ + 312, + 303, + 525, + 332 + ], + "type": "text", + "content": "[85] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 333, + 525, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 333, + 525, + 372 + ], + "spans": [ + { + "bbox": [ + 312, + 333, + 525, + 372 + ], + "type": "text", + "content": "[86] Hanqing Wang, Wenguan Wang, Wei Liang, Steven CH Hoi, Jianbing Shen, and Luc Van Gool. Active perception for visual-language navigation. IJCV, 131(3):607-625, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 373, + 525, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 525, + 423 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 525, + 423 + ], + "type": "text", + "content": "[87] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 424, + 525, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 424, + 525, + 473 + ], + "spans": [ + { + "bbox": [ + 312, + 424, + 525, + 473 + ], + "type": "text", + "content": "[88] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 475, + 525, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 475, + 525, + 504 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 525, + 504 + ], + "type": "text", + "content": "[89] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 505, + 525, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 505, + 525, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 505, + 525, + 533 + ], + "type": "text", + "content": "[90] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Learning to follow and generate instructions for language-capable navigation. IEEE TPAMI, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 535, + 525, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 535, + 525, + 574 + ], + "spans": [ + { + "bbox": [ + 312, + 535, + 525, + 574 + ], + "type": "text", + "content": "[91] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, 2020. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 575, + 525, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 575, + 525, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 575, + 525, + 613 + ], + "type": "text", + "content": "[92] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In CVPR, 2024. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 615, + 525, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 615, + 525, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 615, + 525, + 654 + ], + "type": "text", + "content": "[93] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 655, + 525, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 655, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 312, + 655, + 525, + 686 + ], + "type": "text", + "content": "[94] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Gridmm: Grid memory map for vision-and-language navigation. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16327" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 521 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 73, + 100, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 100, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 100, + 288, + 140 + ], + "type": "text", + "content": "[95] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, 2023. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 141, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 141, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 73, + 141, + 288, + 171 + ], + "type": "text", + "content": "[96] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 172, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 172, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 74, + 172, + 288, + 212 + ], + "type": "text", + "content": "[97] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv:2307.13510, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 213, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 213, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 74, + 213, + 288, + 243 + ], + "type": "text", + "content": "[98] Zhenjia Xu, Zhanpeng He, Jiajun Wu, and Shuran Song. Learning 3d dynamic scene representations for robot manipulation. In CoRL, 2020. 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 245, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 245, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 74, + 245, + 288, + 283 + ], + "type": "text", + "content": "[99] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 285, + 288, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 285, + 288, + 325 + ], + "spans": [ + { + "bbox": [ + 70, + 285, + 288, + 325 + ], + "type": "text", + "content": "[100] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. In ACM MM, 2022. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 327, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 327, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 70, + 327, + 288, + 357 + ], + "type": "text", + "content": "[101] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 357, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 357, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 70, + 357, + 288, + 378 + ], + "type": "text", + "content": "[102] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv: 1904.07850, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 379, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 379, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 70, + 379, + 288, + 408 + ], + "type": "text", + "content": "[103] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 410, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 410, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 70, + 410, + 288, + 440 + ], + "type": "text", + "content": "[104] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 441, + 288, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 441, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 70, + 441, + 288, + 471 + ], + "type": "text", + "content": "[105] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In CVPR, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 472, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 472, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 70, + 472, + 288, + 521 + ], + "type": "text", + "content": "[106] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 2, 3, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "16328" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_content_list.json b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9beedfd1a1e5720862c53392159d0edaec79d26c --- /dev/null +++ b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_content_list.json @@ -0,0 +1,2196 @@ +[ + { + "type": "text", + "text": "WALT3D: Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects under Occlusion", + "text_level": 1, + "bbox": [ + 117, + 128, + 852, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Khiem Vuong $^{1,*}$", + "bbox": [ + 106, + 202, + 238, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "N Dinesh Reddy $^{2,*}$", + "bbox": [ + 277, + 202, + 429, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robert Tamburo1", + "bbox": [ + 470, + 203, + 609, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Srinivasa G. Narasimhan1", + "bbox": [ + 651, + 203, + 854, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ Amazon https://www.cs.cmu.edu/~walt3d", + "bbox": [ + 307, + 227, + 661, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/737840bcd566ca8f0201d04a7c47075eabd7e326e7882f2c454a2bd541023056.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 303, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1ec3367fc2b5d9696b0fe17b98f269d9ec30a777d4e688db180f9d705bcd0d9.jpg", + "image_caption": [ + "Predicted Segmentation", + "Figure 1. Models trained on our automatically generated data from time-lapse imagery can reliably estimate amodal 2D bounding box, segmentation as well as 3D shape and pose despite the complex occlusions presented in the input image." + ], + "image_footnote": [], + "bbox": [ + 303, + 275, + 527, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d62346d8dd0e265d0ce83df131a5765dea8a9d26dc01018b9659960f2b7d2e3e.jpg", + "image_caption": [ + "Predicted Shape" + ], + "image_footnote": [], + "bbox": [ + 527, + 275, + 751, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e298ce3a518fd885aa05f40994654b37ee973e06b1c7a111b15a459ba56f1f9c.jpg", + "image_caption": [ + "3D View" + ], + "image_footnote": [], + "bbox": [ + 753, + 277, + 892, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 433, + 312, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current methods for 2D and 3D object understanding struggle with severe occlusions in busy urban environments, partly due to the lack of large-scale labeled ground-truth annotations for learning occlusion. In this work, we introduce a novel framework for automatically generating a large, realistic dataset of dynamic objects under occlusions using freely available time-lapse imagery. By leveraging off-the-shelf 2D (bounding box, segmentation, keypoint) and 3D (pose, shape) predictions as pseudo-groundtruth, unoccluded 3D objects are identified automatically and composited into the background in a clip-art style, ensuring realistic appearances and physically accurate occlusion configurations. The resulting clip-art image with pseudogroundtruth enables efficient training of object reconstruction methods that are robust to occlusions. Our method demonstrates significant improvements in both 2D and 3D reconstruction, particularly in scenarios with heavily occluded objects like vehicles and people in urban scenes.", + "bbox": [ + 75, + 464, + 472, + 737 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 763, + 207, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, remarkable progress has been made in advancing scene understanding tasks such as object detection [49, 51, 63], tracking [14, 78, 81], segmentation [10, 31], and 3D reconstruction [35, 44, 70]. These achievements are mainly attributed to the availability of large-scale datasets [13, 15, 23, 50, 77] and architectural innova", + "bbox": [ + 75, + 789, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions [17, 30, 42]. Despite this progress, a notable challenge persists in scenarios with severe occlusion, where only a portion of the object is visible: an object may be partially occluded by other objects or truncated by the camera's field-of-view. This phenomenon is referred to as amodal perception [36], and it is hard to intuitively infer their complete shapes. Overcoming these challenges is important to advance many smart cities applications as well as robotics applications, where the number of cameras on vehicles and city infrastructure is rapidly increasing [3, 25, 55].", + "bbox": [ + 500, + 419, + 890, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efforts to learn holistic representations necessitate a substantially annotated and realistic dataset. While recent works such as KINS [58], COCO-Amodal [84], and Ithaca365 [16] have contributed by annotating some amodal ground-truth data, the available annotations remain limited. This scarcity is primarily due to the inherent difficulty in obtaining supervision for amodal representations, as labeling hidden parts of objects is a difficult task for people to accomplish consistently [58, 60, 84]. Moreover, 3D annotations under occlusions, including object shape and pose [7, 43, 74], are even meager due to the difficulty of annotating 3D data, posing challenges for 3D prediction tasks.", + "bbox": [ + 496, + 577, + 890, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Expanding on WALT [62], we utilize time-lapse videos from stationary cameras to synthesize realistic occlusion scenarios by extracting unoccluded objects and composite them back into the background image at their original positions. Unlike WALT which focuses solely on compositing and learning 2D tasks, our approach extends to generating high-quality 3D pseudo-groundtruth data for robust 3D object reconstruction under occlusion. Additionally, our 3D-based compositing method, named WALT3D, in", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes equal contribution and joint first author", + "bbox": [ + 99, + 887, + 359, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9514", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "contrast to WALT's simplistic 2D compositing, produces physically accurate occlusion configurations, leading to increased training data efficiency and scalability.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start with the observation that, although not perfect, existing off-the-shelf methods demonstrate good accuracy in both 2D (segmentation [31], keypoints [67, 68]) and 3D (pose, shape [37, 54, 73]) prediction tasks, especially on unoccluded objects. Thus, they can be used as \"pseudo-groundtruth\" to improve the robustness of existing approaches in occlusion scenarios. Next, we randomly select these unoccluded, non-intersecting 3D objects and put them back into the background image at their original positions (i.e., clip-art style). Specifically, we arrange them based on their distance from the camera, ensuring physically accurate and realistic occlusion configurations. Each such resulting clip-art image is accompanied by amodal bounding box, segmentation masks, and 3D poses and shapes – referred to as \"pseudo-groundtruth\" as these were predicted by off-the-shelf methods on the original unoccluded objects.", + "bbox": [ + 75, + 137, + 468, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments, we demonstrate the effectiveness of our data in both vehicle and human reconstruction, particularly in scenarios with heavy occlusions. It is important to note that our method does not require any human labeling and hence is easily scalable and serves as an effective method to automatically generate realistic training data for reconstructing dynamic objects under occlusion. Our notable technical contributions are summarized below:", + "bbox": [ + 75, + 378, + 468, + 498 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel method that automatically generates 2D/3D supervision data from time-lapse imagery with realistic occlusion configurations without human labeling.", + "- We demonstrate that the utilization of our generated data significantly enhances training efficiency and the accuracy of 2D/3D object reconstruction on real-world data, particularly in scenarios with high occlusion." + ], + "bbox": [ + 76, + 500, + 467, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 618, + 225, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Occlusion Reasoning: Understanding and reasoning occlusions has been extensively studied for decades [21, 22, 65]. Bad predictions due to occlusions are dealt with as noise/outliers in robust estimators. On the other hand, occlusions are explicitly treated as missing parts in model fitting methods [71, 82]. But severe occlusions, such as when a large part of an object is blocked, can result in poor model fitting [28, 85], especially when they attempt to simultaneously estimate the model fit as well as the missing parts.", + "bbox": [ + 75, + 643, + 467, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Amodal Representation: Although the effects of occlusion on visual reasoning has been widely studied, estimating the amodal representation (i.e. both the occluded and visible regions) has only been recently explored. Initial attempts [20, 29, 58, 84] use a supervised learning paradigm using small datasets [58, 84] where humans have annotated occlusions to the best of their abilities. Some methods [59, 60, 66] have explored using multiple views to", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "provide accurate supervision for occluded parts but are not scalable due to capture limitations. To expand supervision, several methods synthesize occlusions to varying degrees of realism. But pure CG renderings [2, 18, 19, 32, 38, 47, 80] suffer from a wide domain-gap [41, 65]. While driving datasets [6, 8, 23, 79] provide 2D amodal bounding boxes, annotating additional representations like segmentation and keypoints under occlusion remains challenging. Some methods, like Ghiasi et al. [24], address this by randomly copy-pasting objects onto diverse background images. However, as these methods solely rely on 2D information, they generate unrealistic occlusion configurations leading to poor performance and limited training efficiency.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Reconstruction Under Occlusion: Reconstruction under severe occlusion is still in the nascent stages of research. Most algorithms developed focus on self-occluded objects with shape completion from partial observations [12, 56, 83]. On the other hand, shape models fitting for objects only with the visible regions either from images [28, 33, 45, 61, 64] or depth sensors [1, 11, 57, 76] have been explored. Due to the inherent challenges in annotating 3D information, ground-truth data for 3D object understanding is limited, often confined to specific object sets [43, 74], indoor [5], or driving scenarios [6, 23, 67], and struggles to generalize well to novel viewpoints such as stationary traffic cameras. Recent methods addressing pose estimation for vehicles [54, 73] and people [26, 40] have demonstrated robustness in handling occlusion, either implicitly or explicitly. Thus, our dataset generation method can significantly enhance the robustness of these approaches to occlusion.", + "bbox": [ + 496, + 287, + 890, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Generating Realistic Supervision Data", + "text_level": 1, + "bbox": [ + 498, + 558, + 844, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An overview of our approach is shown in Figure 2. From a time-lapse video, we identify unoccluded objects and extract their 2D attributes such as segmentation and keypoints. We then use off-the-shelf 3D object reconstruction methods to obtain the pose and shape of these objects, constrained by the camera intrinsics and ground plane. After that, we re-insert these non-intersecting 3D objects into the background image at their original positions in a \"clip-art\" style, arranged based on their distance from the camera to ensure the occlusion configurations are physically accurate and realistic. Finally, we use the clip-art composited image together with its pseudo-groundtruth supervision data to learn robust 2D/3D object reconstruction under occlusion.", + "bbox": [ + 496, + 583, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mining Unoccluded Objects: Given a stream of time-lapse data from a camera, our goal is to mine for unoccluded objects. In this context, \"unoccluded\" denotes instances where a detected object is not obstructed by any other object or truncated out of the field-of-view. On the time-lapse feed from a camera, we run instance segmentation [51] on each frame and use a simple object tracker [75] to track the detected bounding box and segmentation that belongs to \"per", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9515", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/76caf37dc8b6f5399b6adbb1219f3e0f7ac36b8c00daf73ae90ff43a15313786.jpg", + "image_caption": [ + "Figure 2. Given a time-lapse video, we automatically generate 2D/3D training data under severe occlusions. We start by detecting each object in the video, and unoccluded (fully visible) objects are identified. Each unoccluded object is then reconstructed using the ground plane and camera parameters. With the 3D pose, unoccluded objects are composited back into the same location (i.e., clip-art style) in a geometrically consistent approach. The composited image and its pseudo-groundtruth from off-the-shelf methods (e.g., segmentation, keypoints, shapes) are utilized to train a model that can produce accurate 2D/3D object reconstruction under severe occlusions." + ], + "image_footnote": [], + "bbox": [ + 116, + 90, + 851, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "son” and “car” classes. The simplest heuristic for detecting unoccluded objects involves calculating the Intersection over Union (IOU) of detected objects. In previous methods like [62], objects were considered unoccluded if the intersection of their bounding box bottom with any other box was smaller than a threshold $\\delta$ . However, this strategy becomes less reliable in scenarios where dynamic objects are occluded by static elements, such as vehicles being occluded by buildings, trees, poles, etc.", + "bbox": [ + 75, + 306, + 470, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this limitation, we train a simple Occlusion Classifier (OC) to categorize each detected object as either unoccluded or occluded, which is trained using a supervised approach with human annotators who label objects as unoccluded/occluded. Despite the OC module outperforming the heuristic filter in classifying objects' occlusion status, we did not observe a significant improvement in downstream evaluations (less than $\\pm 0.5$ AP) due to the small amount of outliers overall in the training data. Therefore, it is not a mandatory component in our pipeline: the simple heuristics proposed in WALT [62] can be employed for the unoccluded object detection task without additional training. Nevertheless, we believe this data can be useful for occlusion reasoning tasks and we will publicly release it.", + "bbox": [ + 75, + 445, + 468, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstructing Unoccluded Objects: Once unoccluded objects are identified, we describe the 3D reconstruction process of two primary classes: vehicles and humans.", + "bbox": [ + 76, + 657, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each mined unoccluded vehicle is reconstructed following Li et al. [48]. We parameterize each vehicle's 3D keypoints $\\mathbf{X}$ by a linear combination of the mean shape $\\bar{\\mathbf{Q}}$ and $K$ principal components $\\mathbf{Q}_1, \\ldots, \\mathbf{Q}_K$ computed from an object CAD model dataset [47]: $\\mathbf{X} = \\bar{\\mathbf{Q}} + \\sum_{k=1}^{K} \\alpha_k \\mathbf{Q}_k$ , where $\\alpha_k$ is the shape coefficient that needs to be optimized. Starting from the mean shape $\\bar{\\mathbf{Q}}$ , we first detect the 2D keypoints for each unoccluded vehicle and initialize the 6-DoF poses using EPnP [46]. Subsequently, for each track of detected vehicles, we optimize for the 6-DoF object poses while regularizing the shape parameter $\\alpha_k$ to be constant for the same vehicle in different frames by minimizing reprojection errors. Additionally, we also constrain the objects to", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lie on the constant global ground plane, ensuring physically plausible reconstructions.", + "bbox": [ + 498, + 306, + 890, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of human reconstruction, we employ the Human Mesh Recovery (HMR) method HMR 2.0 from Goel et al. [26]. Using HMR 2.0, we predict the SMPL pose and shape parameters for each detected unoccluded human [52]. To determine camera translation, we find the intersection between the backprojected ray from the bottom of the 2D bounding box with the ground plane. Given the camera's intrinsic matrix $\\mathbf{K}$ , ground plane equation $\\left[\\mathbf{n} \\quad d\\right]$ , and $\\mathbf{p}_b$ as the pixel coordinate of the bottom of the box, its depth is computed as $z_b = -\\frac{d}{\\mathbf{n}^T (\\mathbf{K}^{-1} \\mathbf{p}_b)}$ . For camera intrinsic parameters and ground plane equation, we employ a recent method by Vuong et al. [72] that utilizes Google Street View [27] to automatically calibrate stationary traffic cameras and infer the ground plane. The mined unoccluded object image with its corresponding segmentation mask, keypoint locations, and 3D poses are used in a clip-art based framework to generate 2D and 3D supervision data in severe occlusions as illustrated in Fig. 2.", + "bbox": [ + 496, + 339, + 892, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo-groundtruth Data Generation: To obtain supervision for occluded objects, we use image-based compositing to generate composited images, serving as input for training a network to learn 2D/3D reconstruction in severe occlusions. During the compositing process, various 2D/3D representations for each unoccluded object, including keypoints, segmentation, pose, and shape obtained from off-the-shelf methods, serve as pseudo-groundtruth supervision signals even after compositing. By using the clip-art-based compositing approach described below, we automatically generate a large number of realistic supervision signals in severe occlusions. Leveraging the 3D poses of mined unoccluded objects, we geometrically compose the object's image in a clip-art manner. Non-intersecting 3D objects are randomly sampled and reinserted into the background image (i.e., median image) at their original positions, ranging from the farthest to the closest. This approach ensures the creation of physically accurate and realistic occlusion configurations (see Fig. 3). Crucially, our geometry-based", + "bbox": [ + 496, + 613, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9516", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5cc224c2cd719fd14109731cf2343fa539e61ba4e3c67519f5ba32065aab93cf.jpg", + "image_caption": [ + "Cannnnnne" + ], + "image_footnote": [], + "bbox": [ + 151, + 90, + 325, + 165 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39b738b48528dde64a75b36740e0078e0e32247088e1e836c12598d5aa8fb93f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 166, + 323, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/91b2602258577f7eba6f1fb4b739cfe5e2e8f744a06836928c7ff839d9bfef07.jpg", + "image_caption": [ + "Segmentation" + ], + "image_footnote": [], + "bbox": [ + 326, + 90, + 495, + 165 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/01bf58d3fd137e25b5084594f6099cefdcd4e81d344cc612e5be4793664aeb92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 167, + 493, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2bf9951cc0751685e1e8a01d895eb874dcc0cccf6a3de3b3b37a5e0b159334d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 132, + 663, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/11365d4c43fa789325e85b67ddf0f44e3269c21515f0532fea9e05dc477f75e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 169, + 665, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4bcbdcfa6c2473ca6ee58da991fc9528dc1a6a61b0724bb67c564ec341a725f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 90, + 838, + 166 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8231a37bff3ef88a0325ccf3f6a1b4b95d274dafc5eb3d757e40e4b6c231b0ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 167, + 823, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c024c3653f48e58d274cf3b0ac10a7a9379f68ea2e6a6d230dc374460684148.jpg", + "image_caption": [ + "Cannnnnne" + ], + "image_footnote": [], + "bbox": [ + 153, + 253, + 323, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1abe2ee2679dfb1498037657903327519bf9650027f24c1a9cb55206d929163a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 153, + 329, + 323, + 404 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5cd928f9cade3fd1e3171de5d5bf46cae30074121ff18584b58b1b158a52b866.jpg", + "image_caption": [ + "Segmentation" + ], + "image_footnote": [], + "bbox": [ + 325, + 253, + 495, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c0bbd1200c6dc6d3afee36980b1ba8ce95d69c65fa3c06fb5e62228104bd61d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 330, + 493, + 404 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a7f123a7fee0fda041c96f77c194f5032171c3c4bf21a2c9c38b0e2cec97324.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 280, + 651, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3bc7db2cd20d4dc8c16db483453989dc1bf817053d1535a52d251b57f2638283.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 329, + 666, + 404 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b9965965d87343fdda67b2b6a9b44d07bba9f1b060c0c49539939cea52f9df2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 253, + 838, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f01dbe634590424c56d0b017ca24c3b8d46bf4bff448ef04216495e3e0815fbe.jpg", + "image_caption": [ + "3D View" + ], + "image_footnote": [], + "bbox": [ + 700, + 342, + 816, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c4085b37ac6d4e93956bcd6d0b91de2f3692636e35ce9229a80e857bea9ba43.jpg", + "image_caption": [ + "Cannnnnnae" + ], + "image_footnote": [], + "bbox": [ + 153, + 414, + 323, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ca6651d5e30208a2eb1bdbef8da45efa4e957121837f2bdd9a522decaef839a.jpg", + "image_caption": [ + "Figure 3. Automatically generated 2D and 3D Clip-Art to supervise our network: Unoccluded objects are first mined using time-lapse imagery of WALT dataset [62]. Non-intersecting unoccluded objects are composited back into the background image in their respective original positions to preserve correct appearances. The resulting clip-art images, along with their corresponding amodal pseudo-groundtruth information, such as segmentation, keypoints, depth/normal maps, and 3D shapes, are shown. Our method generates realistic appearances from any stationary camera, incorporating diverse viewing geometries, weather conditions, lighting, and occlusion configurations." + ], + "image_footnote": [], + "bbox": [ + 153, + 489, + 323, + 565 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6970d72d18ebf3a91156e6a9d5f134e178610db9a7663e2161dfa8a8e2940b7.jpg", + "image_caption": [ + "Segmentation" + ], + "image_footnote": [], + "bbox": [ + 325, + 414, + 495, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/240e09b24c6b33e3d3ac8931ab281d38a1a419f09cea53dbda947d5ea7a8efe5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 491, + 493, + 565 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a41f1f22920137e3d9a29bebfe76bebe6f456f0e277cabbb40d89837f00b518d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 436, + 665, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e372824fa0e3cbde0b1a31516205794f9b4e709e11c089788c544f6deed3730.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 491, + 665, + 565 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42d96b3aa2cffbd6141e140573d10e01bc3b2ea61376f68a40adc6835b085af0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 414, + 838, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a749abfe66bcca2aefa611bbc548825ceb3fe862c6f0c845c0d98bd5ce505ff.jpg", + "image_caption": [ + "3D View" + ], + "image_footnote": [], + "bbox": [ + 666, + 498, + 836, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "approach distinguishes itself from the mere compositing of 2D images such as WALT [62], which can result in unrealistic composition (see the comparison in Fig. 4). Our data with physically accurate and realistic occlusion configurations contributes to more efficient and scalable training, as demonstrated in our experimental results. Each such generated clip-art image is accompanied by amodal segmentation masks, keypoints, 3D poses and shapes, provided by off-the-shelf methods that serve as pseudo-groundtruth supervision to learn object reconstruction under occlusion.", + "bbox": [ + 75, + 651, + 468, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Learning Reconstruction under Occlusion", + "text_level": 1, + "bbox": [ + 76, + 814, + 452, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have generated an extensive clip-art image dataset with corresponding 2D/3D pseudo-groundtruth representations (see Fig. 3). Using this supervision signal, we train a model capable of inferring holistic object representations in the", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "presence of severe occlusions.", + "bbox": [ + 500, + 652, + 700, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Base Network and 2D Amodal Representations: Using the Swin Transformer [51] backbone with MaskRCNN-based [31] detection heads for its simplicity, our goal is to show improvement irrespective of the base model used. We train the default network with a multi-task loss $\\mathcal{L}_{2\\mathrm{D}}$ for each representation. These losses, computed on each sampled Region of Interest (RoI), include label classification loss $\\mathcal{L}_{\\mathrm{cls}}$ , bounding-box loss $\\mathcal{L}_{\\mathrm{box}}$ , binary cross-entropy loss on the mask branch $\\mathcal{L}_{\\mathrm{mask}}$ and keypoints $\\mathcal{L}_{\\mathrm{kp}}$ : $\\mathcal{L}_{2\\mathrm{D}} = \\mathcal{L}_{\\mathrm{cls}} + \\mathcal{L}_{\\mathrm{box}} + \\mathcal{L}_{\\mathrm{mask}} + \\mathcal{L}_{\\mathrm{kp}}$ . Crucially, with our clip-art dataset including amodal 2D bounding box, segmentation, and keypoints, our method effectively learns amodal instance segmentation and keypoint detection.", + "bbox": [ + 496, + 670, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Object Reconstruction: To regress object 3D pose and shape under occlusion, we extend the base network with a", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9517", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42d251e99c5694a20e861c16969fac73391067739ebf39a996517643465cb399.jpg", + "image_caption": [ + "Composited using WALT" + ], + "image_footnote": [], + "bbox": [ + 137, + 95, + 472, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60fa54432d89f3e4683718d36e268c041fbe69429b9d8864391310c8b711d4b9.jpg", + "image_caption": [ + "Composed using our method", + "Figure 4. Comparison between images composited using the 2D-based method WALT2D [62] (left) and our 3D-based method WALT3D (right). It is evident that our 3D-based compositing method generates realistic and geometrically accurate occlusion configurations, in contrast to the 2D-based method (e.g., cars and people overlapping in an unfeasible way)." + ], + "image_footnote": [], + "bbox": [ + 491, + 95, + 826, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D regression branch. Each vehicle $N$ semantic 3D keypoints $\\mathbf{X}$ are represented by a linear combination of the mean shape $\\bar{\\mathbf{Q}}$ and $K$ principal components $\\mathbf{Q}_{1\\dots K}$ , where $\\alpha_{k}$ is the shape coefficients: $\\mathbf{X} = \\bar{\\mathbf{Q}} +\\sum_{k = 1}^{K}\\alpha_{k}\\mathbf{Q}_{k}$ . Thus, we want to regress for its 6-DoF pose $\\{(\\mathbf{R},\\mathbf{t})\\in \\mathbf{SE}(3)\\}$ and shape coefficients $\\alpha_{k}$ . Given the predicted 2D keypoints and the corresponding 3D keypoints $\\mathbf{X}$ , we can obtain its 6-DoF pose through a differentiable PnP layer [9]. Defining $\\pi (\\cdot)$ as the projection function with known intrinsic $\\mathbf{K}$ , we can optimize for the keypoints reprojection loss $\\mathcal{L}_{\\mathrm{kp2D}}$ , shape loss $\\mathcal{L}_{\\mathrm{shape}}$ , and pose loss $\\mathcal{L}_{\\mathrm{pose}}$ as:", + "bbox": [ + 76, + 383, + 472, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k p 2 D}} = \\sum_ {i = 1} ^ {N} \\left| \\left| \\pi \\left(\\mathbf {R} \\mathbf {X} _ {i} + \\mathbf {t}\\right) - \\mathbf {x} _ {i} ^ {*} \\right| \\right| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 559, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s h a p e}} = | | \\alpha_ {k} - \\alpha_ {k} ^ {*} | | _ {2} ^ {2}, \\mathcal {L} _ {\\mathrm {p o s e}} = | | [ \\mathbf {R}, \\mathbf {t} ] - [ \\mathbf {R} ^ {*}, \\mathbf {t} ^ {*} ] | |\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 602, + 442, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{x}_i^*,\\alpha_k^*,\\mathbf{R}^*,\\mathbf{t}^*$ represents the (pseudo)-groundtruth 2D keypoints, shape coefficients, rotation, and translation respectively. The final loss term is given as the weighted sum of the losses, and we learn amodal bounding box, segmentation, keypoint locations, 3D shape and pose in an end-to-end differentiable manner.", + "bbox": [ + 76, + 628, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For human reconstruction, we train our HMR predictor branch like HMR 2.0 [26]. Supervised by pseudogroundtruth SMPL pose $\\theta^{*}$ and shape $\\beta^{*}$ parameters [52], our network is guided by the loss functions on SMPL parameters $\\mathcal{L}_{\\mathrm{smp1}}$ , 3D keypoints $\\mathcal{L}_{\\mathrm{kp3D}}$ , 2D keypoints $\\mathcal{L}_{\\mathrm{kp2D}}$ and an adversarial loss $\\mathcal{L}_{\\mathrm{adv}}$ that ensures the model predicts valid 3D poses through the discriminator $D_{k}$ :", + "bbox": [ + 76, + 719, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {s m p l}} = | | [ \\theta , \\beta ] - [ \\theta^ {*}, \\beta^ {*} ] | | _ {2} ^ {2}, \\\\ \\mathcal {L} _ {\\mathrm {k p} 3 \\mathrm {D}} = | | \\mathbf {X} - \\mathbf {X} ^ {*} | | _ {1}, \\mathcal {L} _ {\\mathrm {k p} 2 \\mathrm {D}} = | | \\pi (\\mathbf {X}) - \\mathbf {x} ^ {*} | | _ {1}, \\tag {2} \\\\ \\mathcal {L} _ {\\mathrm {a d v}} = \\sum_ {k} (D _ {k} (\\theta_ {b}, \\beta) - 1) ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 832, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{x}^*$ and $\\mathbf{X}^*$ denote the (pseudo)-groundtruth 2D and 3D keypoints respectively. With a weighted combination of these losses, we optimize for human 3D shape and pose in an end-to-end fashion.", + "bbox": [ + 498, + 383, + 892, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Dataset and Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 462, + 825, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Raw Time-lapse Dataset [62]: This dataset contains images from 20 stationary cameras in urban scenes captured over multiple years. The images are either 4K or HD and are captured at 60fps in short bursts. We used 30 days of data from 10 cameras resulting in approximately 3.3 million car and people instances for our experiments.", + "bbox": [ + 498, + 488, + 892, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WALT2D [62]: From the raw time-lapse dataset, we use the 2D-based compositing method proposed by WALT [62] to generate supervision data by pasting unoccluded objects back into the scene with different backgrounds resulting in 10000 training and 500 testing images per camera. Fig. 4 (left) illustrates an example of this dataset. It is important to note that this compositing method, relying solely on 2D information, leads to intersecting objects in the real world that may not faithfully represent the physical configuration.", + "bbox": [ + 498, + 580, + 892, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WALT3D (Ours): Also from the raw time-lapse dataset, our approach involves mining unoccluded objects and reinsering them into the scene, resulting in a comparable number of object instances as 2D-based compositing in WALT2D [62]. However, our approach employs a 3D-based compositing method to generate supervision data, as outlined in Section 3, in contrast to relying solely on 2D-based composition. Fig. 4 (right) shows an example clip-art image composited using our method, with our approach yielding training data with greater realism and physical accuracy compared to WALT2D.", + "bbox": [ + 498, + 718, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vehicle 2D Keypoints Dataset (Ours): To determine 6-", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9518", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f68f3d22999f7fbdb3f9bcee3ba4191e3c4b073c162f6c0e47ff466845502f76.jpg", + "image_caption": [ + "Figure 5. Sample images from our new vehicle 2D keypoints dataset. The dataset contains a wide range of appearance variations including day and night and various traffic scenarios." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 444, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DoF vehicle pose during compositing, we require high-quality 2D vehicle keypoints. Existing datasets [47, 59, 67, 77] offer human-annotated 2D vehicle keypoints but mainly focus on driving scenes or have limited training examples, lacking necessary appearance diversity for novel viewpoints like our stationary traffic cameras. Thus, we propose a new dataset with 7,018 images and 42,547 annotated instances from diverse viewpoints (see Fig. 5), each keypoint containing 2D location and occlusion status. With our dataset, we observe significant improvement in keypoint localization accuracy (66.41% to 80.12% on PCK@0.1), and consequently, an improvement in pseudo-groundtruth data quality. Further details are in the Supplementary.", + "bbox": [ + 75, + 305, + 472, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Baselines, Metrics, and Evaluations", + "text_level": 1, + "bbox": [ + 76, + 516, + 398, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Baselines", + "text_level": 1, + "bbox": [ + 76, + 541, + 184, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detection and Instance Segmentation: All baselines share the same network architecture, employing a Swin [51] backbone as detailed in Section 4. The SWIN baseline is initially trained on the COCO dataset [50]. Subsequently, we perform further finetuning on the WALT2D and WALT3D (Ours) datasets, as detailed in Section 5.", + "bbox": [ + 75, + 564, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vehicle 3D Reconstruction: We compare our approach with Occ-Net [62] and 3DRCNN [45] by changing their feature extraction backbone to be as close as possible to ours for fair comparison. They are trained on a combination of PASCAL3D+ [77], KITTI-3D [47], Carfusion [59], and ApolloCar3D [67], where we further finetune our baselines on the WALT2D and our WALT3D dataset, respectively.", + "bbox": [ + 75, + 655, + 468, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human 3D Reconstruction: Our architecture mirrors HMR 2.0 [26] as described in Section 4. We initialize from the pretrained model, then finetune on our WALT3D data.", + "bbox": [ + 75, + 761, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 816, + 261, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2D Metrics: We report Average Precision (AP) on bounding box detection $(\\mathrm{AP}^{\\mathrm{box}})$ and instance segmentation $(\\mathrm{AP}^{\\mathrm{mask}})$ following [50]. For keypoints, we use Percentage of Correct Keypoints (PCK) metric where a keypoint", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e92bac6db158d80d0018b980a556c89a147dbbd8ced869af683550db799e3df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 87, + 694, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36e7458477813ad70a527e4d524f9e5ea4b478bf73def2efd2a3f09efdae5321.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 87, + 885, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06891abaceb83f32acbb20cc538638e61f4dbd78dfea43b1822c61c13213c2e5.jpg", + "image_caption": [ + "Figure 6. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal vehicle detection, segmentation, 2D and 3D keypoint estimation. Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage." + ], + "image_footnote": [], + "bbox": [ + 503, + 203, + 694, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43523d45f41f7fbacf664b7ab73b87d60582bc3d5a9ada1688f924422af87126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 203, + 883, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e4910d1c02e2ab20730c2066f4d9643a6edad906756cbde257fa7aa55e67d81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 388, + 696, + 500 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e631aac7db9bba79c36bd3bfed551a5f36af5327c9e6228d41994b54fa71c7a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 388, + 885, + 500 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a362ce4b63535c3da1816bb1be0c3da5bd513f1d21590a905c50376a6ae1b04b.jpg", + "image_caption": [ + "Figure 7. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal human detection, segmentation, and human mesh recovery (HMR). Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage." + ], + "image_footnote": [], + "bbox": [ + 503, + 503, + 694, + 612 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c302e39b214b053277cddb179afeb84a9be8eace253ad9f4fde636c36e9b5b75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 503, + 883, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is considered correct if it lies within the radius $\\alpha$ (with $0 < \\alpha < 1$ ) of the ground-truth keypoint.", + "bbox": [ + 496, + 694, + 890, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Metrics: We use the standard metrics of previous work [34], reporting the Mean Per Joint Position Error (MPJME) between predicted and ground-truth 3D keypoints (aligned using the root joint) as well as 3D PCK.", + "bbox": [ + 496, + 726, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Ablation Analysis", + "text_level": 1, + "bbox": [ + 500, + 800, + 671, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness to Occlusions: We evaluate the effectiveness of our algorithm with varying occlusion levels. Similar to WALT [62], we use the pseudo-groundtruth segmentation masks from the evaluation set to group objects based on occlusion percentage. In 2D tasks such as detection and", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9519", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11bc1846cbde4d155582ab20d4e781c6b9fd4bbcc9e6237881d0bc22b8d0922f.jpg", + "image_caption": [ + "Figure 8. Comparison between WALT2D [62]) and our WALT3D approach. Since our method produces higher quality training data with more realistic occlusion configuration, our approach is especially useful in low-data regime." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 444, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "segmentation, the model trained with our WALT3D data significantly outperforms the baseline SWIN model trained on COCO [50], which only includes modal masks. This improvement is particularly evident in high occlusion percentages for both vehicle (Fig. 6) and human (Fig. 7). Compared to WALT2D [62], our data demonstrates slightly better accuracy across all occlusion levels, thanks to our 3D-based compositing method WALT3D producing physically accurate occlusion configurations with the same amount of training data. For 3D tasks such as vehicle (Fig. 6) and human (Fig. 7) 3D pose estimation, using our WALT3D data consistently outperforms other baselines, particularly in high occlusion percentages. Thus, models trained with our WALT3D data, even with imperfect pseudo-groundtruth from off-the-shelf methods, are robust to occlusion.", + "bbox": [ + 75, + 352, + 468, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D-based Compositing helps Data Efficiency: Fig. 8 shows that our WALT3D data significantly improves performance across all training data fractions, including at low data regime (at $15\\%$ of training data with $+13.9\\mathrm{AP}^{\\mathrm{mask}}$ improvement compared to training with WALT2D data). Essentially, a model trained on $25\\%$ of WALT3D data achieves comparable AP to one trained on $60\\%$ of WALT2D data. This improvement stems from our WALT3D approach generating higher quality and physically accurate occlusion configurations with the same training data volume, aligning more closely with the real-world distribution as in Fig. 4. Importantly, this data efficiency enables scaling to a larger number of scenes more efficiently.", + "bbox": [ + 75, + 580, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-evaluations on other datasets: Table 1 presents quantitative results on COCO-Amodal [84] (COCO-A) (natural images) and KINS [58] (driving images), each containing human-annotated ground-truth 2D amodal instance masks. It shows that additional fine-tuning with our WALT3D data improves results in both KINS and COCO-A, highlighting the cross-evaluation performance of our data, with potential for even better generalization with the", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c84d1ce64a46986ad9bb44fc1ce2a2aa3a8e491d20513f9bc3714d0a2065f4a.jpg", + "image_caption": [ + "Previous SOTA" + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 694, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b4695871f6f385107405b370d782682899a6788e61c1d433fd3de35e56e7dbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 88, + 800, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c89720c2764fe66dfb94bdaf2a708c453d99f8424d6650c5732393ce6a6f56b.jpg", + "image_caption": [ + "Figure 9. Qualitative results showing that our data improves amodal segmentation, keypoint prediction and 3D reconstruction compared to previous SOTA." + ], + "image_footnote": [], + "bbox": [ + 714, + 217, + 799, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48c14d578cf8539c8da2c81d4552ff38a63a996db8f4ed73a7c87d6160c38bd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 88, + 883, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/233b070b6eb184501aca11923d44eb232dcd11fa446b7786071b516b4e97eca0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 800, + 217, + 883, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b2a65d1aea2c6c7a75075b859b1048e3b00120b7b4fc3a02eeb9f0a4096d39b5.jpg", + "table_caption": [ + "increasing amount of automatically generated data." + ], + "table_footnote": [], + "table_body": "
Testing Data Training DataBox APMask AP
KINSCOCO-AKINSCOCO-A
COCO27.838.020.831.2
COCO + WALT2D28.240.422.436.4
COCO + WALT3D28.843.223.538.7
COCO + COCO-A + KINS36.846.932.942.7
COCO + COCO-A + KINS + WALT2D37.447.933.343.5
COCO + COCO-A + KINS + WALT3D38.848.735.645.3
", + "bbox": [ + 500, + 372, + 890, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Finetuning with our WALT3D data improves generalization across KINS and COCO-A (vehicle & people). Additionally, combining our data with domain-specific data like KINS and COCO-A further enhances performance in their domains.", + "bbox": [ + 498, + 474, + 890, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-evaluations on pedestrian tracking: We evaluate how our WALT3D data helps pedestrian tracking in Table 2. To highlight how better detections lead to better tracking, we use a simple tracker SORT [4] with two detectors: one pretrained on COCO and one finetuned with our WALT3D data. On MOT17-train, the model finetuned with WALT3D data produces high-quality detections under strong occlusions, resulting in improvements over the baseline in all metrics, particularly those favoring better detection quality like MOTA and DetA (see metrics details in [53]).", + "bbox": [ + 496, + 532, + 890, + 683 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f455f0e2eaba1e3537e39ddcd3decf2ea88ea85e8e4c1885422f5d8dab828710.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Detector Train DataMOTA↑DetA↑IDF1↑HOTA↑
COCO42.240.346.740.1
COCO + WALT3D47.243.249.542.0
", + "bbox": [ + 517, + 686, + 870, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. MOT17-train results using SORT with two detectors: one pretrained on COCO and one fine-tuned with our WALT3D data.", + "bbox": [ + 498, + 741, + 890, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results: Results of our method can be seen in Fig. 9 and Fig. 10. Comparing to previous methods, observe that our method can produce robust 2D/3D predictions, even under challenging high-occlusion scenarios.", + "bbox": [ + 498, + 776, + 890, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 849, + 746, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our data generation framework is method-agnostic, accommodating various robust object pose estimation alterna", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9520", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ccbf88efd1bb580fa878e4ae6e0b5c316184ee4d5a613fc3d4e6346df7781084.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 89, + 250, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51ad2b5dcc6bed33f649f678b501cb62c5f053f008395fb8e3ec9eb651f72abe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 89, + 403, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2751cadcf00ee45e6c78370eef789f1576401f2b517e912b8caa89e39ffe3f0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 89, + 560, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/452d2486b5d256d025f88c7730d8f836df23d730c5be0d9807e654129184494c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 90, + 717, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ecb1b346aa61a46ea543f2a456906e80ff3108edfa464a80b61bf4d59f6a5a5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 90, + 872, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a25042638fd834a5e91dab330423a7b9efbcfa678c956b1bb1f4868e00ba3bc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 191, + 250, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/543584f73f8c002f24b62ad3c012bdadb89a4041eac4b19f608686e037032b9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 193, + 403, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f6e3100e3c1ce33eeba61d0470fe127f1ab2e82a59e0d2f95d4a4531c4cdb97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 193, + 560, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/235807ed81f953666ecaeac0082d3ea38de823bb91d5b8b60392e8714036552e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 193, + 717, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5399533297c714632f8830e8efce5c5f417d8092d5fb0f6f81463deb023d2a03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 193, + 872, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb4b8542f64e925969b965f299c478e2c14ea405820cfc8e25505b3289148d35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 295, + 248, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e1a6bcc678c48932502e78b9b1851e67cbe98f7a37cf4bc777dbe745be99734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 295, + 403, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6dce8c450b37d38d0b352f38bddd37ba81b0e5b98a40784f10aa89209521e6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 295, + 560, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6524742e917981b262cdc6c17c523d73466fe04d04a80c6149bc074b05c28126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 295, + 717, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f9f2a98ee5fedec2f926dbec7bb8a148aa45e868b37213aeca9c69127f64d05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 295, + 872, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3bc3618d5e4807490fda3ff465da208174db30136c1fec39896c05fbecd3f711.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 390, + 248, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a89a001ca50e1525be270798caeb74aa75826f4cc067985fea0a3afe01fa724.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 390, + 403, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d92175515ce611fb0cb3047af03db622c1c19ec6659ac6b755ec280b5e4406ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 390, + 560, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed8424bb2ff5ea25edd655c2f3f038386c6af56830d7f419cf98f968387582ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 390, + 717, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d7c6ce62630e89dec5784ff8e2ea92bda020383dae2c2b8e1b66e2ac6eb93a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 390, + 872, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9351148fd96970ee0295c863557a24279f0eea6ea48a6f3118af9e4a8273585.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 508, + 250, + 637 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b8f5970b9821bb857dcf272c919d374721c794de0eba53b2b1169cfa26eddcd.jpg", + "image_caption": [ + "Amodal Segmentation" + ], + "image_footnote": [], + "bbox": [ + 250, + 508, + 403, + 636 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3346e2b90ad0b9e4d655e650d22d93f2cf77cec3a62ccfc1e4639a2e2c9f979.jpg", + "image_caption": [ + "Amodal Keypoints/Shapes" + ], + "image_footnote": [], + "bbox": [ + 405, + 508, + 560, + 637 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f19aff6d7c4ef2bedd6f0f4575dc04bb206b314bcddc660ee6b5362d6c1a7db.jpg", + "image_caption": [ + "3D View-1" + ], + "image_footnote": [], + "bbox": [ + 562, + 520, + 715, + 636 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c98d465c2f01795d93b15c5811bcbf24729b242edaaf45c4a0ce086cdc29d1dc.jpg", + "image_caption": [ + "3D View-2", + "Figure 10. We show qualitative results of our method on real data. Our method produces accurate amodal segmentation, keypoints, as well as 3D poses and shapes across diverse poses and occlusion configurations. In particular, we show results on different level and types of occlusions like vehicle-vehicle (row 1, 2, 3), vehicle-people (row 3), and people-people occlusion (row 4, 5)." + ], + "image_footnote": [], + "bbox": [ + 717, + 547, + 872, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tives [54, 73] as well as methods like VIBE [39] for high-quality human reconstructions from videos. Moreover, as our pseudo-groundtruth data includes metric-scale depth information using ground plane, we can augment existing datasets like Relative Human [69] to enhance the robustness of 3D multi-human reconstruction methods to occlusion.", + "bbox": [ + 75, + 704, + 470, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Further research is required for generalization across diverse views for it to be used as a generic solution. It also assumes a mean shape or a parametric object model is available, posing challenges for rare objects. Addressing appearance inconsistencies (e.g., variations in lighting) in clip-art images is a promising research direction.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, our method introduces an automated approach to generate a realistic dataset for reconstructing dynamic objects under occlusions from time-lapse imagery. It demonstrates significant improvements in both 2D and 3D object reconstruction, particularly in busy urban scenes with diverse occlusion configurations.", + "bbox": [ + 496, + 705, + 890, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Societal Impact. We do not perform any human subjects research from these cameras. For privacy, we blur faces and license plates in all images intended for release.", + "bbox": [ + 496, + 796, + 890, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported in part by an NSF Grant CNS-2038612, a US DOT grant 69A3551747111 through the Mobility21 UTC and grants 69A3552344811 and 69A3552348316 through the Safety21 UTC.", + "bbox": [ + 496, + 840, + 890, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9521", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] William Agnew, Christopher Xie, Aaron Walsman, Octavian Murad, Yubo Wang, Pedro Domingos, and Siddhartha Srinivasa. Amodal 3d reconstruction for robotic manipulation via stability and connectivity. In CoRL, 2021. 2", + "[2] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets deep learning for car instance segmentation in urban scenes. In BMVC, 2017. 2", + "[3] Asra Aslam. Detecting objects in less response time for processing multimedia events in smart cities. In CVPR, 2022. 1", + "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 7", + "[5] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 2", + "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2", + "[7] Berk Calli, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. IJRR, 2017. 1", + "[8] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In CVPR, 2019. 2", + "[9] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In CVPR, 2020. 5", + "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1", + "[11] Falak Chhaya, Dinesh Reddy, Sarthak Upadhyay, Visesh Chari, M. Zeeshan Zia, and K. Madhava Krishna. Monocular reconstruction of vehicles: Combining slam with shape priors. In ICRA, 2016. 2", + "[12] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In ECCV, 2016. 2", + "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1", + "[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1", + "[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 1", + "[16] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In CVPR, 2022. 1", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[18] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2", + "[19] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 2", + "[20] Patrick Follmann, Rebecca König, Philipp Hartinger, Michael Klostermann, and Tobias Böttger. Learning to see the invisible: End-to-end trainable amodal instance segmentation. In WACV, 2019. 2", + "[21] Rik Fransens, Christoph Strecha, and Luc Van Gool. A mean field em-algorithm for coherent occlusion handling in mape-estimation prob. In CVPR, 2006. 2", + "[22] Tianshi Gao, Benjamin Packer, and Daphne Koller. A segmentation-aware object detection model with occlusion handling. In CVPR, 2011. 2", + "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1, 2", + "[24] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2", + "[25] Holger Glasl, David Schreiber, Nikolaus Viertl, Stephan Veigl, and Gustavo Fernandez. Video based traffic congestion prediction on an embedded system. In ITSC, 2008. 1", + "[26] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3, 5, 6", + "[27] Google. Google Street View. https://www.google.com/streetview/.3", + "[28] Can Gumeli, Angela Dai, and Matthias Nießner. Roca: robust cad model retrieval and alignment from a single image. In CVPR, 2022. 2", + "[29] Ruiqi Guo and Derek Hoiem. Beyond the line of sight: labeling the underlying surfaces. In ECCV, 2012. 2", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1", + "[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 1, 2, 4", + "[32] Y.-T. Hu, H.-S. Chen, K. Hui, J.-B. Huang, and A. G. Schwing. SAIL-VOS: Semantic Amodal Instance Level Video Object Segmentation - A Synthetic Dataset and Baselines. In CVPR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9522", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In ECCV, 2020. 2", + "[34] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 6", + "[35] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1", + "[36] Gaetano Kanizsa. Organization in vision: Essays on gestalt perception. Praeger Publishers, 1979. 1", + "[37] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, 2020. 2", + "[38] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2", + "[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020. 8", + "[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, 2021. 2", + "[41] Philipp Krahenbuhl. Free supervision from video games. In CVPR, 2018. 2", + "[42] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. 1", + "[43] Alexander Krull, Eric Brachmann, Frank Michel, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Learning analysis-by-synthesis for 6d pose estimation in rgb-d images. In ICCV, 2015. 1, 2", + "[44] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 1", + "[45] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, 2018. 2, 6", + "[46] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 3", + "[47] Chi Li, M Zeeshan Zia, Quoc-Huy Tran, Xiang Yu, Gregory D Hager, and Manmohan Chandraker. Deep supervision with shape concepts for occlusion-aware 3d object parsing. In CVPR, 2017. 2, 3, 6", + "[48] Fangyu Li, N Dinesh Reddy, Xudong Chen, and Srinivasa G Narasimhan. Traffic4d: Single view longitudinal 4d reconstruction of repetitious activity using self-supervised experts. In IV, 2021. 3", + "[49] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 1", + "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 6", + "[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. SIGGRAPH Asia, 2015. 3, 5", + "[53] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 2021. 7", + "[54] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In ECCV, 2022. 2, 8", + "[55] M. Naphade, S. Wang, D. C. Anastasiu, Z. Tang, M. Chang, Y. Yao, L. Zheng, M. Shaiqur Rahman, A. Venkatachalapathy, A. Sharma, Q. Feng, V. Ablavsky, S. Sclaroff, P. Chakraborty, A. Li, S. Li, and R. Chellappa. The 6th ai city challenge. In CVPRW, 2022. 1", + "[56] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2", + "[57] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2", + "[58] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 1, 2, 7", + "[59] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In CVPR, 2018. 2, 6", + "[60] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In CVPR, 2019. 1, 2", + "[61] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15190-15200, 2021. 2", + "[62] N. Dinesh Reddy, Robert Tamburo, and Srinivasa G. Narasimhan. Walt: Watch and learn 2d amodal representation from time-lapse imagery. In CVPR, 2022. 1, 3, 4, 5, 6, 7", + "[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1", + "[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2", + "[65] Samuel Schulter, Menghua Zhai, Nathan Jacobs, and Manmohan Chandraker. Learning to look around objects for top-view representations of outdoor scenes. In ECCV, 2018. 2", + "[66] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9523", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In CVPR, 2019. 2, 6", + "[68] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2", + "[69] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J. Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, 2022. 8", + "[70] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In CVPR, 2015. 1", + "[71] Andrea Vedaldi and Andrew Zisserman. Structured output regression for detection with partial truncation. In NeurIPS, 2009. 2", + "[72] Khiem Vuong, Robert Tamburo, and Srinivasa G. Narasimhan. Toward planet-wide traffic camera calibration. In WACV, 2024. 3", + "[73] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. In ICLR, 2021. 2, 8", + "[74] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1, 2", + "[75] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2", + "[76] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2", + "[77] Y. Xiang, R. Mottaghi, and S. Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 1, 6", + "[78] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015. 1", + "[79] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2", + "[80] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust instance segmentation through reasoning about multi-object occlusion. In CVPR, 2021. 2", + "[81] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1", + "[82] Xiaowei Zhou, Spyridon Leonardos, Xiaoyan Hu, and Kostas Daniilidis. 3d shape estimation from 2d landmarks: A convex relaxation approach. In CVPR, 2015. 2", + "[83] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2", + "[84] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 1, 2, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[85] M Zeeshan Zia, Michael Stark, and Konrad Schindler. Towards scene understanding with detailed 3d object representations. IJCV, 2015. 2", + "bbox": [ + 501, + 90, + 890, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9524", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_model.json b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6a0793a68f6f4bfb7e67d83bb0ea14965bbaeb43 --- /dev/null +++ b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_model.json @@ -0,0 +1,3203 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.853, + 0.175 + ], + "angle": 0, + "content": "WALT3D: Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects under Occlusion" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.203, + 0.239, + 0.222 + ], + "angle": 0, + "content": "Khiem Vuong\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.203, + 0.431, + 0.222 + ], + "angle": 0, + "content": "N Dinesh Reddy\\(^{2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.204, + 0.61, + 0.221 + ], + "angle": 0, + "content": "Robert Tamburo1" + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.204, + 0.856, + 0.221 + ], + "angle": 0, + "content": "Srinivasa G. Narasimhan1" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.228, + 0.662, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)Amazon https://www.cs.cmu.edu/~walt3d" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.275, + 0.305, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.364, + 0.218, + 0.377 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.276, + 0.528, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.364, + 0.485, + 0.376 + ], + "angle": 0, + "content": "Predicted Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.276, + 0.753, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.364, + 0.677, + 0.377 + ], + "angle": 0, + "content": "Predicted Shape" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.278, + 0.893, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.364, + 0.848, + 0.375 + ], + "angle": 0, + "content": "3D View" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.89, + 0.406 + ], + "angle": 0, + "content": "Figure 1. Models trained on our automatically generated data from time-lapse imagery can reliably estimate amodal 2D bounding box, segmentation as well as 3D shape and pose despite the complex occlusions presented in the input image." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.434, + 0.313, + 0.45 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.473, + 0.738 + ], + "angle": 0, + "content": "Current methods for 2D and 3D object understanding struggle with severe occlusions in busy urban environments, partly due to the lack of large-scale labeled ground-truth annotations for learning occlusion. In this work, we introduce a novel framework for automatically generating a large, realistic dataset of dynamic objects under occlusions using freely available time-lapse imagery. By leveraging off-the-shelf 2D (bounding box, segmentation, keypoint) and 3D (pose, shape) predictions as pseudo-groundtruth, unoccluded 3D objects are identified automatically and composited into the background in a clip-art style, ensuring realistic appearances and physically accurate occlusion configurations. The resulting clip-art image with pseudogroundtruth enables efficient training of object reconstruction methods that are robust to occlusions. Our method demonstrates significant improvements in both 2D and 3D reconstruction, particularly in scenarios with heavily occluded objects like vehicles and people in urban scenes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.764, + 0.208, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.469, + 0.88 + ], + "angle": 0, + "content": "In recent years, remarkable progress has been made in advancing scene understanding tasks such as object detection [49, 51, 63], tracking [14, 78, 81], segmentation [10, 31], and 3D reconstruction [35, 44, 70]. These achievements are mainly attributed to the availability of large-scale datasets [13, 15, 23, 50, 77] and architectural innova" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.42, + 0.892, + 0.571 + ], + "angle": 0, + "content": "tions [17, 30, 42]. Despite this progress, a notable challenge persists in scenarios with severe occlusion, where only a portion of the object is visible: an object may be partially occluded by other objects or truncated by the camera's field-of-view. This phenomenon is referred to as amodal perception [36], and it is hard to intuitively infer their complete shapes. Overcoming these challenges is important to advance many smart cities applications as well as robotics applications, where the number of cameras on vehicles and city infrastructure is rapidly increasing [3, 25, 55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Efforts to learn holistic representations necessitate a substantially annotated and realistic dataset. While recent works such as KINS [58], COCO-Amodal [84], and Ithaca365 [16] have contributed by annotating some amodal ground-truth data, the available annotations remain limited. This scarcity is primarily due to the inherent difficulty in obtaining supervision for amodal representations, as labeling hidden parts of objects is a difficult task for people to accomplish consistently [58, 60, 84]. Moreover, 3D annotations under occlusions, including object shape and pose [7, 43, 74], are even meager due to the difficulty of annotating 3D data, posing challenges for 3D prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Expanding on WALT [62], we utilize time-lapse videos from stationary cameras to synthesize realistic occlusion scenarios by extracting unoccluded objects and composite them back into the background image at their original positions. Unlike WALT which focuses solely on compositing and learning 2D tasks, our approach extends to generating high-quality 3D pseudo-groundtruth data for robust 3D object reconstruction under occlusion. Additionally, our 3D-based compositing method, named WALT3D, in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.361, + 0.9 + ], + "angle": 0, + "content": "* denotes equal contribution and joint first author" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "contrast to WALT's simplistic 2D compositing, produces physically accurate occlusion configurations, leading to increased training data efficiency and scalability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.379 + ], + "angle": 0, + "content": "We start with the observation that, although not perfect, existing off-the-shelf methods demonstrate good accuracy in both 2D (segmentation [31], keypoints [67, 68]) and 3D (pose, shape [37, 54, 73]) prediction tasks, especially on unoccluded objects. Thus, they can be used as \"pseudo-groundtruth\" to improve the robustness of existing approaches in occlusion scenarios. Next, we randomly select these unoccluded, non-intersecting 3D objects and put them back into the background image at their original positions (i.e., clip-art style). Specifically, we arrange them based on their distance from the camera, ensuring physically accurate and realistic occlusion configurations. Each such resulting clip-art image is accompanied by amodal bounding box, segmentation masks, and 3D poses and shapes – referred to as \"pseudo-groundtruth\" as these were predicted by off-the-shelf methods on the original unoccluded objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Through extensive experiments, we demonstrate the effectiveness of our data in both vehicle and human reconstruction, particularly in scenarios with heavy occlusions. It is important to note that our method does not require any human labeling and hence is easily scalable and serves as an effective method to automatically generate realistic training data for reconstructing dynamic objects under occlusion. Our notable technical contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.468, + 0.543 + ], + "angle": 0, + "content": "- We introduce a novel method that automatically generates 2D/3D supervision data from time-lapse imagery with realistic occlusion configurations without human labeling." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.545, + 0.468, + 0.605 + ], + "angle": 0, + "content": "- We demonstrate that the utilization of our generated data significantly enhances training efficiency and the accuracy of 2D/3D object reconstruction on real-world data, particularly in scenarios with high occlusion." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.5, + 0.468, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.619, + 0.227, + 0.635 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.468, + 0.78 + ], + "angle": 0, + "content": "Occlusion Reasoning: Understanding and reasoning occlusions has been extensively studied for decades [21, 22, 65]. Bad predictions due to occlusions are dealt with as noise/outliers in robust estimators. On the other hand, occlusions are explicitly treated as missing parts in model fitting methods [71, 82]. But severe occlusions, such as when a large part of an object is blocked, can result in poor model fitting [28, 85], especially when they attempt to simultaneously estimate the model fit as well as the missing parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2D Amodal Representation: Although the effects of occlusion on visual reasoning has been widely studied, estimating the amodal representation (i.e. both the occluded and visible regions) has only been recently explored. Initial attempts [20, 29, 58, 84] use a supervised learning paradigm using small datasets [58, 84] where humans have annotated occlusions to the best of their abilities. Some methods [59, 60, 66] have explored using multiple views to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "provide accurate supervision for occluded parts but are not scalable due to capture limitations. To expand supervision, several methods synthesize occlusions to varying degrees of realism. But pure CG renderings [2, 18, 19, 32, 38, 47, 80] suffer from a wide domain-gap [41, 65]. While driving datasets [6, 8, 23, 79] provide 2D amodal bounding boxes, annotating additional representations like segmentation and keypoints under occlusion remains challenging. Some methods, like Ghiasi et al. [24], address this by randomly copy-pasting objects onto diverse background images. However, as these methods solely rely on 2D information, they generate unrealistic occlusion configurations leading to poor performance and limited training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.545 + ], + "angle": 0, + "content": "3D Reconstruction Under Occlusion: Reconstruction under severe occlusion is still in the nascent stages of research. Most algorithms developed focus on self-occluded objects with shape completion from partial observations [12, 56, 83]. On the other hand, shape models fitting for objects only with the visible regions either from images [28, 33, 45, 61, 64] or depth sensors [1, 11, 57, 76] have been explored. Due to the inherent challenges in annotating 3D information, ground-truth data for 3D object understanding is limited, often confined to specific object sets [43, 74], indoor [5], or driving scenarios [6, 23, 67], and struggles to generalize well to novel viewpoints such as stationary traffic cameras. Recent methods addressing pose estimation for vehicles [54, 73] and people [26, 40] have demonstrated robustness in handling occlusion, either implicitly or explicitly. Thus, our dataset generation method can significantly enhance the robustness of these approaches to occlusion." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.559, + 0.845, + 0.575 + ], + "angle": 0, + "content": "3. Generating Realistic Supervision Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.78 + ], + "angle": 0, + "content": "An overview of our approach is shown in Figure 2. From a time-lapse video, we identify unoccluded objects and extract their 2D attributes such as segmentation and keypoints. We then use off-the-shelf 3D object reconstruction methods to obtain the pose and shape of these objects, constrained by the camera intrinsics and ground plane. After that, we re-insert these non-intersecting 3D objects into the background image at their original positions in a \"clip-art\" style, arranged based on their distance from the camera to ensure the occlusion configurations are physically accurate and realistic. Finally, we use the clip-art composited image together with its pseudo-groundtruth supervision data to learn robust 2D/3D object reconstruction under occlusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Mining Unoccluded Objects: Given a stream of time-lapse data from a camera, our goal is to mine for unoccluded objects. In this context, \"unoccluded\" denotes instances where a detected object is not obstructed by any other object or truncated out of the field-of-view. On the time-lapse feed from a camera, we run instance segmentation [51] on each frame and use a simple object tracker [75] to track the detected bounding box and segmentation that belongs to \"per" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9515" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.092, + 0.852, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 2. Given a time-lapse video, we automatically generate 2D/3D training data under severe occlusions. We start by detecting each object in the video, and unoccluded (fully visible) objects are identified. Each unoccluded object is then reconstructed using the ground plane and camera parameters. With the 3D pose, unoccluded objects are composited back into the same location (i.e., clip-art style) in a geometrically consistent approach. The composited image and its pseudo-groundtruth from off-the-shelf methods (e.g., segmentation, keypoints, shapes) are utilized to train a model that can produce accurate 2D/3D object reconstruction under severe occlusions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.471, + 0.444 + ], + "angle": 0, + "content": "son” and “car” classes. The simplest heuristic for detecting unoccluded objects involves calculating the Intersection over Union (IOU) of detected objects. In previous methods like [62], objects were considered unoccluded if the intersection of their bounding box bottom with any other box was smaller than a threshold \\(\\delta\\). However, this strategy becomes less reliable in scenarios where dynamic objects are occluded by static elements, such as vehicles being occluded by buildings, trees, poles, etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.656 + ], + "angle": 0, + "content": "To address this limitation, we train a simple Occlusion Classifier (OC) to categorize each detected object as either unoccluded or occluded, which is trained using a supervised approach with human annotators who label objects as unoccluded/occluded. Despite the OC module outperforming the heuristic filter in classifying objects' occlusion status, we did not observe a significant improvement in downstream evaluations (less than \\(\\pm 0.5\\) AP) due to the small amount of outliers overall in the training data. Therefore, it is not a mandatory component in our pipeline: the simple heuristics proposed in WALT [62] can be employed for the unoccluded object detection task without additional training. Nevertheless, we believe this data can be useful for occlusion reasoning tasks and we will publicly release it." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Reconstructing Unoccluded Objects: Once unoccluded objects are identified, we describe the 3D reconstruction process of two primary classes: vehicles and humans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Each mined unoccluded vehicle is reconstructed following Li et al. [48]. We parameterize each vehicle's 3D keypoints \\(\\mathbf{X}\\) by a linear combination of the mean shape \\(\\bar{\\mathbf{Q}}\\) and \\(K\\) principal components \\(\\mathbf{Q}_1, \\ldots, \\mathbf{Q}_K\\) computed from an object CAD model dataset [47]: \\(\\mathbf{X} = \\bar{\\mathbf{Q}} + \\sum_{k=1}^{K} \\alpha_k \\mathbf{Q}_k\\), where \\(\\alpha_k\\) is the shape coefficient that needs to be optimized. Starting from the mean shape \\(\\bar{\\mathbf{Q}}\\), we first detect the 2D keypoints for each unoccluded vehicle and initialize the 6-DoF poses using EPnP [46]. Subsequently, for each track of detected vehicles, we optimize for the 6-DoF object poses while regularizing the shape parameter \\(\\alpha_k\\) to be constant for the same vehicle in different frames by minimizing reprojection errors. Additionally, we also constrain the objects to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.338 + ], + "angle": 0, + "content": "lie on the constant global ground plane, ensuring physically plausible reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.612 + ], + "angle": 0, + "content": "In the context of human reconstruction, we employ the Human Mesh Recovery (HMR) method HMR 2.0 from Goel et al. [26]. Using HMR 2.0, we predict the SMPL pose and shape parameters for each detected unoccluded human [52]. To determine camera translation, we find the intersection between the backprojected ray from the bottom of the 2D bounding box with the ground plane. Given the camera's intrinsic matrix \\(\\mathbf{K}\\), ground plane equation \\(\\left[\\mathbf{n} \\quad d\\right]\\), and \\(\\mathbf{p}_b\\) as the pixel coordinate of the bottom of the box, its depth is computed as \\(z_b = -\\frac{d}{\\mathbf{n}^T (\\mathbf{K}^{-1} \\mathbf{p}_b)}\\). For camera intrinsic parameters and ground plane equation, we employ a recent method by Vuong et al. [72] that utilizes Google Street View [27] to automatically calibrate stationary traffic cameras and infer the ground plane. The mined unoccluded object image with its corresponding segmentation mask, keypoint locations, and 3D poses are used in a clip-art based framework to generate 2D and 3D supervision data in severe occlusions as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Pseudo-groundtruth Data Generation: To obtain supervision for occluded objects, we use image-based compositing to generate composited images, serving as input for training a network to learn 2D/3D reconstruction in severe occlusions. During the compositing process, various 2D/3D representations for each unoccluded object, including keypoints, segmentation, pose, and shape obtained from off-the-shelf methods, serve as pseudo-groundtruth supervision signals even after compositing. By using the clip-art-based compositing approach described below, we automatically generate a large number of realistic supervision signals in severe occlusions. Leveraging the 3D poses of mined unoccluded objects, we geometrically compose the object's image in a clip-art manner. Non-intersecting 3D objects are randomly sampled and reinserted into the background image (i.e., median image) at their original positions, ranging from the farthest to the closest. This approach ensures the creation of physically accurate and realistic occlusion configurations (see Fig. 3). Crucially, our geometry-based" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9516" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.145, + 0.141, + 0.188 + ], + "angle": 0, + "content": "Cannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.091, + 0.326, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.167, + 0.325, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.091, + 0.496, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.168, + 0.495, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.092, + 0.582, + 0.106 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.133, + 0.665, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.17, + 0.666, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.092, + 0.839, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.169, + 0.824, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.305, + 0.141, + 0.35 + ], + "angle": 0, + "content": "Cannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.254, + 0.325, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.33, + 0.324, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.255, + 0.496, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.331, + 0.495, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.257, + 0.577, + 0.27 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.281, + 0.653, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.33, + 0.668, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.254, + 0.839, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.332, + 0.73, + 0.343 + ], + "angle": 0, + "content": "3D View" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.343, + 0.818, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.467, + 0.141, + 0.512 + ], + "angle": 0, + "content": "Cannnnnnae" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.415, + 0.325, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.491, + 0.324, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.415, + 0.496, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.492, + 0.495, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.418, + 0.577, + 0.431 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.438, + 0.666, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.492, + 0.666, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.415, + 0.839, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.494, + 0.728, + 0.505 + ], + "angle": 0, + "content": "3D View" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.499, + 0.838, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.57, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Figure 3. Automatically generated 2D and 3D Clip-Art to supervise our network: Unoccluded objects are first mined using time-lapse imagery of WALT dataset [62]. Non-intersecting unoccluded objects are composited back into the background image in their respective original positions to preserve correct appearances. The resulting clip-art images, along with their corresponding amodal pseudo-groundtruth information, such as segmentation, keypoints, depth/normal maps, and 3D shapes, are shown. Our method generates realistic appearances from any stationary camera, incorporating diverse viewing geometries, weather conditions, lighting, and occlusion configurations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.804 + ], + "angle": 0, + "content": "approach distinguishes itself from the mere compositing of 2D images such as WALT [62], which can result in unrealistic composition (see the comparison in Fig. 4). Our data with physically accurate and realistic occlusion configurations contributes to more efficient and scalable training, as demonstrated in our experimental results. Each such generated clip-art image is accompanied by amodal segmentation masks, keypoints, 3D poses and shapes, provided by off-the-shelf methods that serve as pseudo-groundtruth supervision to learn object reconstruction under occlusion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.453, + 0.833 + ], + "angle": 0, + "content": "4. Learning Reconstruction under Occlusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We have generated an extensive clip-art image dataset with corresponding 2D/3D pseudo-groundtruth representations (see Fig. 3). Using this supervision signal, we train a model capable of inferring holistic object representations in the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.653, + 0.702, + 0.667 + ], + "angle": 0, + "content": "presence of severe occlusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Base Network and 2D Amodal Representations: Using the Swin Transformer [51] backbone with MaskRCNN-based [31] detection heads for its simplicity, our goal is to show improvement irrespective of the base model used. We train the default network with a multi-task loss \\(\\mathcal{L}_{2\\mathrm{D}}\\) for each representation. These losses, computed on each sampled Region of Interest (RoI), include label classification loss \\(\\mathcal{L}_{\\mathrm{cls}}\\), bounding-box loss \\(\\mathcal{L}_{\\mathrm{box}}\\), binary cross-entropy loss on the mask branch \\(\\mathcal{L}_{\\mathrm{mask}}\\) and keypoints \\(\\mathcal{L}_{\\mathrm{kp}}\\): \\(\\mathcal{L}_{2\\mathrm{D}} = \\mathcal{L}_{\\mathrm{cls}} + \\mathcal{L}_{\\mathrm{box}} + \\mathcal{L}_{\\mathrm{mask}} + \\mathcal{L}_{\\mathrm{kp}}\\). Crucially, with our clip-art dataset including amodal 2D bounding box, segmentation, and keypoints, our method effectively learns amodal instance segmentation and keypoint detection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Object Reconstruction: To regress object 3D pose and shape under occlusion, we extend the base network with a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9517" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.096, + 0.473, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.313, + 0.364, + 0.326 + ], + "angle": 0, + "content": "Composited using WALT" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.097, + 0.828, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.313, + 0.74, + 0.326 + ], + "angle": 0, + "content": "Composed using our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.894, + 0.372 + ], + "angle": 0, + "content": "Figure 4. Comparison between images composited using the 2D-based method WALT2D [62] (left) and our 3D-based method WALT3D (right). It is evident that our 3D-based compositing method generates realistic and geometrically accurate occlusion configurations, in contrast to the 2D-based method (e.g., cars and people overlapping in an unfeasible way)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.473, + 0.552 + ], + "angle": 0, + "content": "3D regression branch. Each vehicle \\(N\\) semantic 3D keypoints \\(\\mathbf{X}\\) are represented by a linear combination of the mean shape \\(\\bar{\\mathbf{Q}}\\) and \\(K\\) principal components \\(\\mathbf{Q}_{1\\dots K}\\), where \\(\\alpha_{k}\\) is the shape coefficients: \\(\\mathbf{X} = \\bar{\\mathbf{Q}} +\\sum_{k = 1}^{K}\\alpha_{k}\\mathbf{Q}_{k}\\). Thus, we want to regress for its 6-DoF pose \\(\\{(\\mathbf{R},\\mathbf{t})\\in \\mathbf{SE}(3)\\}\\) and shape coefficients \\(\\alpha_{k}\\). Given the predicted 2D keypoints and the corresponding 3D keypoints \\(\\mathbf{X}\\), we can obtain its 6-DoF pose through a differentiable PnP layer [9]. Defining \\(\\pi (\\cdot)\\) as the projection function with known intrinsic \\(\\mathbf{K}\\), we can optimize for the keypoints reprojection loss \\(\\mathcal{L}_{\\mathrm{kp2D}}\\), shape loss \\(\\mathcal{L}_{\\mathrm{shape}}\\), and pose loss \\(\\mathcal{L}_{\\mathrm{pose}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.56, + 0.469, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k p 2 D}} = \\sum_ {i = 1} ^ {N} \\left| \\left| \\pi \\left(\\mathbf {R} \\mathbf {X} _ {i} + \\mathbf {t}\\right) - \\mathbf {x} _ {i} ^ {*} \\right| \\right| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.603, + 0.443, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s h a p e}} = | | \\alpha_ {k} - \\alpha_ {k} ^ {*} | | _ {2} ^ {2}, \\mathcal {L} _ {\\mathrm {p o s e}} = | | [ \\mathbf {R}, \\mathbf {t} ] - [ \\mathbf {R} ^ {*}, \\mathbf {t} ^ {*} ] | |\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{x}_i^*,\\alpha_k^*,\\mathbf{R}^*,\\mathbf{t}^*\\) represents the (pseudo)-groundtruth 2D keypoints, shape coefficients, rotation, and translation respectively. The final loss term is given as the weighted sum of the losses, and we learn amodal bounding box, segmentation, keypoint locations, 3D shape and pose in an end-to-end differentiable manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.826 + ], + "angle": 0, + "content": "For human reconstruction, we train our HMR predictor branch like HMR 2.0 [26]. Supervised by pseudogroundtruth SMPL pose \\(\\theta^{*}\\) and shape \\(\\beta^{*}\\) parameters [52], our network is guided by the loss functions on SMPL parameters \\(\\mathcal{L}_{\\mathrm{smp1}}\\), 3D keypoints \\(\\mathcal{L}_{\\mathrm{kp3D}}\\), 2D keypoints \\(\\mathcal{L}_{\\mathrm{kp2D}}\\) and an adversarial loss \\(\\mathcal{L}_{\\mathrm{adv}}\\) that ensures the model predicts valid 3D poses through the discriminator \\(D_{k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.833, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {s m p l}} = | | [ \\theta , \\beta ] - [ \\theta^ {*}, \\beta^ {*} ] | | _ {2} ^ {2}, \\\\ \\mathcal {L} _ {\\mathrm {k p} 3 \\mathrm {D}} = | | \\mathbf {X} - \\mathbf {X} ^ {*} | | _ {1}, \\mathcal {L} _ {\\mathrm {k p} 2 \\mathrm {D}} = | | \\pi (\\mathbf {X}) - \\mathbf {x} ^ {*} | | _ {1}, \\tag {2} \\\\ \\mathcal {L} _ {\\mathrm {a d v}} = \\sum_ {k} (D _ {k} (\\theta_ {b}, \\beta) - 1) ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.893, + 0.445 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{x}^*\\) and \\(\\mathbf{X}^*\\) denote the (pseudo)-groundtruth 2D and 3D keypoints respectively. With a weighted combination of these losses, we optimize for human 3D shape and pose in an end-to-end fashion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.826, + 0.48 + ], + "angle": 0, + "content": "5. Dataset and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.893, + 0.58 + ], + "angle": 0, + "content": "Raw Time-lapse Dataset [62]: This dataset contains images from 20 stationary cameras in urban scenes captured over multiple years. The images are either 4K or HD and are captured at 60fps in short bursts. We used 30 days of data from 10 cameras resulting in approximately 3.3 million car and people instances for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.581, + 0.893, + 0.718 + ], + "angle": 0, + "content": "WALT2D [62]: From the raw time-lapse dataset, we use the 2D-based compositing method proposed by WALT [62] to generate supervision data by pasting unoccluded objects back into the scene with different backgrounds resulting in 10000 training and 500 testing images per camera. Fig. 4 (left) illustrates an example of this dataset. It is important to note that this compositing method, relying solely on 2D information, leads to intersecting objects in the real world that may not faithfully represent the physical configuration." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.719, + 0.892, + 0.884 + ], + "angle": 0, + "content": "WALT3D (Ours): Also from the raw time-lapse dataset, our approach involves mining unoccluded objects and reinsering them into the scene, resulting in a comparable number of object instances as 2D-based compositing in WALT2D [62]. However, our approach employs a 3D-based compositing method to generate supervision data, as outlined in Section 3, in contrast to relying solely on 2D-based composition. Fig. 4 (right) shows an example clip-art image composited using our method, with our approach yielding training data with greater realism and physical accuracy compared to WALT2D." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Vehicle 2D Keypoints Dataset (Ours): To determine 6-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.445, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.298 + ], + "angle": 0, + "content": "Figure 5. Sample images from our new vehicle 2D keypoints dataset. The dataset contains a wide range of appearance variations including day and night and various traffic scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.504 + ], + "angle": 0, + "content": "DoF vehicle pose during compositing, we require high-quality 2D vehicle keypoints. Existing datasets [47, 59, 67, 77] offer human-annotated 2D vehicle keypoints but mainly focus on driving scenes or have limited training examples, lacking necessary appearance diversity for novel viewpoints like our stationary traffic cameras. Thus, we propose a new dataset with 7,018 images and 42,547 annotated instances from diverse viewpoints (see Fig. 5), each keypoint containing 2D location and occlusion status. With our dataset, we observe significant improvement in keypoint localization accuracy (66.41% to 80.12% on PCK@0.1), and consequently, an improvement in pseudo-groundtruth data quality. Further details are in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.517, + 0.4, + 0.533 + ], + "angle": 0, + "content": "6. Baselines, Metrics, and Evaluations" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.186, + 0.556 + ], + "angle": 0, + "content": "6.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Detection and Instance Segmentation: All baselines share the same network architecture, employing a Swin [51] backbone as detailed in Section 4. The SWIN baseline is initially trained on the COCO dataset [50]. Subsequently, we perform further finetuning on the WALT2D and WALT3D (Ours) datasets, as detailed in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Vehicle 3D Reconstruction: We compare our approach with Occ-Net [62] and 3DRCNN [45] by changing their feature extraction backbone to be as close as possible to ours for fair comparison. They are trained on a combination of PASCAL3D+ [77], KITTI-3D [47], Carfusion [59], and ApolloCar3D [67], where we further finetune our baselines on the WALT2D and our WALT3D dataset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.469, + 0.807 + ], + "angle": 0, + "content": "Human 3D Reconstruction: Our architecture mirrors HMR 2.0 [26] as described in Section 4. We initialize from the pretrained model, then finetune on our WALT3D data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.262, + 0.832 + ], + "angle": 0, + "content": "6.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "2D Metrics: We report Average Precision (AP) on bounding box detection \\((\\mathrm{AP}^{\\mathrm{box}})\\) and instance segmentation \\((\\mathrm{AP}^{\\mathrm{mask}})\\) following [50]. For keypoints, we use Percentage of Correct Keypoints (PCK) metric where a keypoint" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.696, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.088, + 0.886, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.204, + 0.695, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.204, + 0.885, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 6. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal vehicle detection, segmentation, 2D and 3D keypoint estimation. Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.39, + 0.697, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.39, + 0.886, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.505, + 0.695, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.505, + 0.884, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Figure 7. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal human detection, segmentation, and human mesh recovery (HMR). Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.725 + ], + "angle": 0, + "content": "is considered correct if it lies within the radius \\(\\alpha\\) (with \\(0 < \\alpha < 1\\)) of the ground-truth keypoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.788 + ], + "angle": 0, + "content": "3D Metrics: We use the standard metrics of previous work [34], reporting the Mean Per Joint Position Error (MPJME) between predicted and ground-truth 3D keypoints (aligned using the root joint) as well as 3D PCK." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.673, + 0.817 + ], + "angle": 0, + "content": "6.3. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Robustness to Occlusions: We evaluate the effectiveness of our algorithm with varying occlusion levels. Similar to WALT [62], we use the pseudo-groundtruth segmentation masks from the evaluation set to group objects based on occlusion percentage. In 2D tasks such as detection and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.445, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Figure 8. Comparison between WALT2D [62]) and our WALT3D approach. Since our method produces higher quality training data with more realistic occlusion configuration, our approach is especially useful in low-data regime." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.579 + ], + "angle": 0, + "content": "segmentation, the model trained with our WALT3D data significantly outperforms the baseline SWIN model trained on COCO [50], which only includes modal masks. This improvement is particularly evident in high occlusion percentages for both vehicle (Fig. 6) and human (Fig. 7). Compared to WALT2D [62], our data demonstrates slightly better accuracy across all occlusion levels, thanks to our 3D-based compositing method WALT3D producing physically accurate occlusion configurations with the same amount of training data. For 3D tasks such as vehicle (Fig. 6) and human (Fig. 7) 3D pose estimation, using our WALT3D data consistently outperforms other baselines, particularly in high occlusion percentages. Thus, models trained with our WALT3D data, even with imperfect pseudo-groundtruth from off-the-shelf methods, are robust to occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.779 + ], + "angle": 0, + "content": "3D-based Compositing helps Data Efficiency: Fig. 8 shows that our WALT3D data significantly improves performance across all training data fractions, including at low data regime (at \\(15\\%\\) of training data with \\(+13.9\\mathrm{AP}^{\\mathrm{mask}}\\) improvement compared to training with WALT2D data). Essentially, a model trained on \\(25\\%\\) of WALT3D data achieves comparable AP to one trained on \\(60\\%\\) of WALT2D data. This improvement stems from our WALT3D approach generating higher quality and physically accurate occlusion configurations with the same training data volume, aligning more closely with the real-world distribution as in Fig. 4. Importantly, this data efficiency enables scaling to a larger number of scenes more efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Cross-evaluations on other datasets: Table 1 presents quantitative results on COCO-Amodal [84] (COCO-A) (natural images) and KINS [58] (driving images), each containing human-annotated ground-truth 2D amodal instance masks. It shows that additional fine-tuning with our WALT3D data improves results in both KINS and COCO-A, highlighting the cross-evaluation performance of our data, with potential for even better generalization with the" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.088, + 0.695, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.285, + 0.676, + 0.297 + ], + "angle": 0, + "content": "Previous SOTA" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.089, + 0.801, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.218, + 0.8, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.089, + 0.885, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.218, + 0.885, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.285, + 0.822, + 0.297 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.299, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 9. Qualitative results showing that our data improves amodal segmentation, keypoint prediction and 3D reconstruction compared to previous SOTA." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.347, + 0.838, + 0.362 + ], + "angle": 0, + "content": "increasing amount of automatically generated data." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.373, + 0.892, + 0.47 + ], + "angle": 0, + "content": "
Testing Data Training DataBox APMask AP
KINSCOCO-AKINSCOCO-A
COCO27.838.020.831.2
COCO + WALT2D28.240.422.436.4
COCO + WALT3D28.843.223.538.7
COCO + COCO-A + KINS36.846.932.942.7
COCO + COCO-A + KINS + WALT2D37.447.933.343.5
COCO + COCO-A + KINS + WALT3D38.848.735.645.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.475, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Table 1. Finetuning with our WALT3D data improves generalization across KINS and COCO-A (vehicle & people). Additionally, combining our data with domain-specific data like KINS and COCO-A further enhances performance in their domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Cross-evaluations on pedestrian tracking: We evaluate how our WALT3D data helps pedestrian tracking in Table 2. To highlight how better detections lead to better tracking, we use a simple tracker SORT [4] with two detectors: one pretrained on COCO and one finetuned with our WALT3D data. On MOT17-train, the model finetuned with WALT3D data produces high-quality detections under strong occlusions, resulting in improvements over the baseline in all metrics, particularly those favoring better detection quality like MOTA and DetA (see metrics details in [53])." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.687, + 0.871, + 0.739 + ], + "angle": 0, + "content": "
Detector Train DataMOTA↑DetA↑IDF1↑HOTA↑
COCO42.240.346.740.1
COCO + WALT3D47.243.249.542.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.742, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Table 2. MOT17-train results using SORT with two detectors: one pretrained on COCO and one fine-tuned with our WALT3D data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Qualitative Results: Results of our method can be seen in Fig. 9 and Fig. 10. Comparing to previous methods, observe that our method can produce robust 2D/3D predictions, even under challenging high-occlusion scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.85, + 0.748, + 0.865 + ], + "angle": 0, + "content": "7. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our data generation framework is method-agnostic, accommodating various robust object pose estimation alterna" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9520" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.25, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.09, + 0.405, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.09, + 0.562, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.091, + 0.718, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.091, + 0.874, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.193, + 0.25, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.194, + 0.405, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.194, + 0.562, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.194, + 0.718, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.194, + 0.874, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.296, + 0.249, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.296, + 0.405, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.296, + 0.562, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.296, + 0.718, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.296, + 0.874, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.391, + 0.249, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.391, + 0.405, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.391, + 0.562, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.391, + 0.718, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.391, + 0.874, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.509, + 0.25, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.639, + 0.214, + 0.653 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.509, + 0.405, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.639, + 0.396, + 0.652 + ], + "angle": 0, + "content": "Amodal Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.509, + 0.562, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.639, + 0.565, + 0.652 + ], + "angle": 0, + "content": "Amodal Keypoints/Shapes" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.521, + 0.717, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.64, + 0.671, + 0.652 + ], + "angle": 0, + "content": "3D View-1" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.549, + 0.874, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.639, + 0.824, + 0.651 + ], + "angle": 0, + "content": "3D View-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.655, + 0.895, + 0.699 + ], + "angle": 0, + "content": "Figure 10. We show qualitative results of our method on real data. Our method produces accurate amodal segmentation, keypoints, as well as 3D poses and shapes across diverse poses and occlusion configurations. In particular, we show results on different level and types of occlusions like vehicle-vehicle (row 1, 2, 3), vehicle-people (row 3), and people-people occlusion (row 4, 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.796 + ], + "angle": 0, + "content": "tives [54, 73] as well as methods like VIBE [39] for high-quality human reconstructions from videos. Moreover, as our pseudo-groundtruth data includes metric-scale depth information using ground plane, we can augment existing datasets like Relative Human [69] to enhance the robustness of 3D multi-human reconstruction methods to occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Limitations: Further research is required for generalization across diverse views for it to be used as a generic solution. It also assumes a mean shape or a parametric object model is available, posing challenges for rare objects. Addressing appearance inconsistencies (e.g., variations in lighting) in clip-art images is a promising research direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.796 + ], + "angle": 0, + "content": "In conclusion, our method introduces an automated approach to generate a realistic dataset for reconstructing dynamic objects under occlusions from time-lapse imagery. It demonstrates significant improvements in both 2D and 3D object reconstruction, particularly in busy urban scenes with diverse occlusion configurations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.842 + ], + "angle": 0, + "content": "Potential Societal Impact. We do not perform any human subjects research from these cameras. For privacy, we blur faces and license plates in all images intended for release." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.842, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported in part by an NSF Grant CNS-2038612, a US DOT grant 69A3551747111 through the Mobility21 UTC and grants 69A3552344811 and 69A3552348316 through the Safety21 UTC." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9521" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] William Agnew, Christopher Xie, Aaron Walsman, Octavian Murad, Yubo Wang, Pedro Domingos, and Siddhartha Srinivasa. Amodal 3d reconstruction for robotic manipulation via stability and connectivity. In CoRL, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets deep learning for car instance segmentation in urban scenes. In BMVC, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.47, + 0.271 + ], + "angle": 0, + "content": "[3] Asra Aslam. Detecting objects in less response time for processing multimedia events in smart cities. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[5] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.486 + ], + "angle": 0, + "content": "[7] Berk Calli, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. IJRR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[8] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[9] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.588, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.645, + 0.47, + 0.7 + ], + "angle": 0, + "content": "[11] Falak Chhaya, Dinesh Reddy, Sarthak Upadhyay, Visesh Chari, M. Zeeshan Zia, and K. Madhava Krishna. Monocular reconstruction of vehicles: Combining slam with shape priors. In ICRA, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.702, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[12] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.758, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.47, + 0.884 + ], + "angle": 0, + "content": "[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[16] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[19] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[20] Patrick Follmann, Rebecca König, Philipp Hartinger, Michael Klostermann, and Tobias Böttger. Learning to see the invisible: End-to-end trainable amodal instance segmentation. In WACV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[21] Rik Fransens, Christoph Strecha, and Luc Van Gool. A mean field em-algorithm for coherent occlusion handling in mape-estimation prob. In CVPR, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[22] Tianshi Gao, Benjamin Packer, and Daphne Koller. A segmentation-aware object detection model with occlusion handling. In CVPR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[24] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[25] Holger Glasl, David Schreiber, Nikolaus Viertl, Stephan Veigl, and Gustavo Fernandez. Video based traffic congestion prediction on an embedded system. In ITSC, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[26] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[27] Google. Google Street View. https://www.google.com/streetview/.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[28] Can Gumeli, Angela Dai, and Matthias Nießner. Roca: robust cad model retrieval and alignment from a single image. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[29] Ruiqi Guo and Derek Hoiem. Beyond the line of sight: labeling the underlying surfaces. In ECCV, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.777, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] Y.-T. Hu, H.-S. Chen, K. Hui, J.-B. Huang, and A. G. Schwing. SAIL-VOS: Semantic Amodal Instance Level Video Object Segmentation - A Synthetic Dataset and Baselines. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9522" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[34] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[35] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[36] Gaetano Kanizsa. Organization in vision: Essays on gestalt perception. Praeger Publishers, 1979. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.261, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[37] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[38] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[41] Philipp Krahenbuhl. Free supervision from video games. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.47, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[42] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[43] Alexander Krull, Eric Brachmann, Frank Michel, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Learning analysis-by-synthesis for 6d pose estimation in rgb-d images. In ICCV, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.568, + 0.469, + 0.608 + ], + "angle": 0, + "content": "[44] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.61, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[45] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.652, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[46] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[47] Chi Li, M Zeeshan Zia, Quoc-Huy Tran, Xiang Yu, Gregory D Hager, and Manmohan Chandraker. Deep supervision with shape concepts for occlusion-aware 3d object parsing. In CVPR, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[48] Fangyu Li, N Dinesh Reddy, Xudong Chen, and Srinivasa G Narasimhan. Traffic4d: Single view longitudinal 4d reconstruction of repetitious activity using self-supervised experts. In IV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[49] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.847, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[51] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. SIGGRAPH Asia, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[53] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[54] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In ECCV, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[55] M. Naphade, S. Wang, D. C. Anastasiu, Z. Tang, M. Chang, Y. Yao, L. Zheng, M. Shaiqur Rahman, A. Venkatachalapathy, A. Sharma, Q. Feng, V. Ablavsky, S. Sclaroff, P. Chakraborty, A. Li, S. Li, and R. Chellappa. The 6th ai city challenge. In CVPRW, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[56] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[57] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[58] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[59] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In CVPR, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[60] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[61] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15190-15200, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[62] N. Dinesh Reddy, Robert Tamburo, and Srinivasa G. Narasimhan. Walt: Watch and learn 2d amodal representation from time-lapse imagery. In CVPR, 2022. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[65] Samuel Schulter, Menghua Zhai, Nathan Jacobs, and Manmohan Chandraker. Learning to look around objects for top-view representations of outdoor scenes. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[66] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9523" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[67] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[68] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[69] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J. Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.237, + 0.469, + 0.264 + ], + "angle": 0, + "content": "[70] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.267, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[71] Andrea Vedaldi and Andrew Zisserman. Structured output regression for detection with partial truncation. In NeurIPS, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.31, + 0.469, + 0.351 + ], + "angle": 0, + "content": "[72] Khiem Vuong, Robert Tamburo, and Srinivasa G. Narasimhan. Toward planet-wide traffic camera calibration. In WACV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.354, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[73] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. In ICLR, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.398, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[74] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.454, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[75] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[76] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.469, + 0.581 + ], + "angle": 0, + "content": "[77] Y. Xiang, R. Mottaghi, and S. Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.585, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[78] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.629, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[79] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.686, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[80] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust instance segmentation through reasoning about multi-object occlusion. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.729, + 0.469, + 0.783 + ], + "angle": 0, + "content": "[81] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.786, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[82] Xiaowei Zhou, Spyridon Leonardos, Xiaoyan Hu, and Kostas Daniilidis. 3d shape estimation from 2d landmarks: A convex relaxation approach. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.83, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[83] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[84] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[85] M Zeeshan Zia, Michael Stark, and Konrad Schindler. Towards scene understanding with detailed 3d object representations. IJCV, 2015. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9524" + } + ] +] \ No newline at end of file diff --git a/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_origin.pdf b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23248e2087b0881e558fc654ac2fcf5f41db0877 --- /dev/null +++ b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/048cb13a-fa04-482f-97a0-4548a36785ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c9e8e98b1ae1926bc4125164b43f2a5862b31faa5fbeaa77dd8770d0077ec50 +size 8839489 diff --git a/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/full.md b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..38c40d94199c66edf42426710f4a3ee3fc809fb9 --- /dev/null +++ b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/full.md @@ -0,0 +1,431 @@ +# WALT3D: Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects under Occlusion + +Khiem Vuong $^{1,*}$ + +N Dinesh Reddy $^{2,*}$ + +Robert Tamburo1 + +Srinivasa G. Narasimhan1 + +$^{1}$ Carnegie Mellon University $^{2}$ Amazon https://www.cs.cmu.edu/~walt3d + +![](images/737840bcd566ca8f0201d04a7c47075eabd7e326e7882f2c454a2bd541023056.jpg) +Input Image + +![](images/f1ec3367fc2b5d9696b0fe17b98f269d9ec30a777d4e688db180f9d705bcd0d9.jpg) +Predicted Segmentation +Figure 1. Models trained on our automatically generated data from time-lapse imagery can reliably estimate amodal 2D bounding box, segmentation as well as 3D shape and pose despite the complex occlusions presented in the input image. + +![](images/d62346d8dd0e265d0ce83df131a5765dea8a9d26dc01018b9659960f2b7d2e3e.jpg) +Predicted Shape + +![](images/e298ce3a518fd885aa05f40994654b37ee973e06b1c7a111b15a459ba56f1f9c.jpg) +3D View + +# Abstract + +Current methods for 2D and 3D object understanding struggle with severe occlusions in busy urban environments, partly due to the lack of large-scale labeled ground-truth annotations for learning occlusion. In this work, we introduce a novel framework for automatically generating a large, realistic dataset of dynamic objects under occlusions using freely available time-lapse imagery. By leveraging off-the-shelf 2D (bounding box, segmentation, keypoint) and 3D (pose, shape) predictions as pseudo-groundtruth, unoccluded 3D objects are identified automatically and composited into the background in a clip-art style, ensuring realistic appearances and physically accurate occlusion configurations. The resulting clip-art image with pseudogroundtruth enables efficient training of object reconstruction methods that are robust to occlusions. Our method demonstrates significant improvements in both 2D and 3D reconstruction, particularly in scenarios with heavily occluded objects like vehicles and people in urban scenes. + +# 1. Introduction + +In recent years, remarkable progress has been made in advancing scene understanding tasks such as object detection [49, 51, 63], tracking [14, 78, 81], segmentation [10, 31], and 3D reconstruction [35, 44, 70]. These achievements are mainly attributed to the availability of large-scale datasets [13, 15, 23, 50, 77] and architectural innova + +tions [17, 30, 42]. Despite this progress, a notable challenge persists in scenarios with severe occlusion, where only a portion of the object is visible: an object may be partially occluded by other objects or truncated by the camera's field-of-view. This phenomenon is referred to as amodal perception [36], and it is hard to intuitively infer their complete shapes. Overcoming these challenges is important to advance many smart cities applications as well as robotics applications, where the number of cameras on vehicles and city infrastructure is rapidly increasing [3, 25, 55]. + +Efforts to learn holistic representations necessitate a substantially annotated and realistic dataset. While recent works such as KINS [58], COCO-Amodal [84], and Ithaca365 [16] have contributed by annotating some amodal ground-truth data, the available annotations remain limited. This scarcity is primarily due to the inherent difficulty in obtaining supervision for amodal representations, as labeling hidden parts of objects is a difficult task for people to accomplish consistently [58, 60, 84]. Moreover, 3D annotations under occlusions, including object shape and pose [7, 43, 74], are even meager due to the difficulty of annotating 3D data, posing challenges for 3D prediction tasks. + +Expanding on WALT [62], we utilize time-lapse videos from stationary cameras to synthesize realistic occlusion scenarios by extracting unoccluded objects and composite them back into the background image at their original positions. Unlike WALT which focuses solely on compositing and learning 2D tasks, our approach extends to generating high-quality 3D pseudo-groundtruth data for robust 3D object reconstruction under occlusion. Additionally, our 3D-based compositing method, named WALT3D, in + +contrast to WALT's simplistic 2D compositing, produces physically accurate occlusion configurations, leading to increased training data efficiency and scalability. + +We start with the observation that, although not perfect, existing off-the-shelf methods demonstrate good accuracy in both 2D (segmentation [31], keypoints [67, 68]) and 3D (pose, shape [37, 54, 73]) prediction tasks, especially on unoccluded objects. Thus, they can be used as "pseudo-groundtruth" to improve the robustness of existing approaches in occlusion scenarios. Next, we randomly select these unoccluded, non-intersecting 3D objects and put them back into the background image at their original positions (i.e., clip-art style). Specifically, we arrange them based on their distance from the camera, ensuring physically accurate and realistic occlusion configurations. Each such resulting clip-art image is accompanied by amodal bounding box, segmentation masks, and 3D poses and shapes – referred to as "pseudo-groundtruth" as these were predicted by off-the-shelf methods on the original unoccluded objects. + +Through extensive experiments, we demonstrate the effectiveness of our data in both vehicle and human reconstruction, particularly in scenarios with heavy occlusions. It is important to note that our method does not require any human labeling and hence is easily scalable and serves as an effective method to automatically generate realistic training data for reconstructing dynamic objects under occlusion. Our notable technical contributions are summarized below: + +- We introduce a novel method that automatically generates 2D/3D supervision data from time-lapse imagery with realistic occlusion configurations without human labeling. +- We demonstrate that the utilization of our generated data significantly enhances training efficiency and the accuracy of 2D/3D object reconstruction on real-world data, particularly in scenarios with high occlusion. + +# 2. Related Works + +Occlusion Reasoning: Understanding and reasoning occlusions has been extensively studied for decades [21, 22, 65]. Bad predictions due to occlusions are dealt with as noise/outliers in robust estimators. On the other hand, occlusions are explicitly treated as missing parts in model fitting methods [71, 82]. But severe occlusions, such as when a large part of an object is blocked, can result in poor model fitting [28, 85], especially when they attempt to simultaneously estimate the model fit as well as the missing parts. + +2D Amodal Representation: Although the effects of occlusion on visual reasoning has been widely studied, estimating the amodal representation (i.e. both the occluded and visible regions) has only been recently explored. Initial attempts [20, 29, 58, 84] use a supervised learning paradigm using small datasets [58, 84] where humans have annotated occlusions to the best of their abilities. Some methods [59, 60, 66] have explored using multiple views to + +provide accurate supervision for occluded parts but are not scalable due to capture limitations. To expand supervision, several methods synthesize occlusions to varying degrees of realism. But pure CG renderings [2, 18, 19, 32, 38, 47, 80] suffer from a wide domain-gap [41, 65]. While driving datasets [6, 8, 23, 79] provide 2D amodal bounding boxes, annotating additional representations like segmentation and keypoints under occlusion remains challenging. Some methods, like Ghiasi et al. [24], address this by randomly copy-pasting objects onto diverse background images. However, as these methods solely rely on 2D information, they generate unrealistic occlusion configurations leading to poor performance and limited training efficiency. + +3D Reconstruction Under Occlusion: Reconstruction under severe occlusion is still in the nascent stages of research. Most algorithms developed focus on self-occluded objects with shape completion from partial observations [12, 56, 83]. On the other hand, shape models fitting for objects only with the visible regions either from images [28, 33, 45, 61, 64] or depth sensors [1, 11, 57, 76] have been explored. Due to the inherent challenges in annotating 3D information, ground-truth data for 3D object understanding is limited, often confined to specific object sets [43, 74], indoor [5], or driving scenarios [6, 23, 67], and struggles to generalize well to novel viewpoints such as stationary traffic cameras. Recent methods addressing pose estimation for vehicles [54, 73] and people [26, 40] have demonstrated robustness in handling occlusion, either implicitly or explicitly. Thus, our dataset generation method can significantly enhance the robustness of these approaches to occlusion. + +# 3. Generating Realistic Supervision Data + +An overview of our approach is shown in Figure 2. From a time-lapse video, we identify unoccluded objects and extract their 2D attributes such as segmentation and keypoints. We then use off-the-shelf 3D object reconstruction methods to obtain the pose and shape of these objects, constrained by the camera intrinsics and ground plane. After that, we re-insert these non-intersecting 3D objects into the background image at their original positions in a "clip-art" style, arranged based on their distance from the camera to ensure the occlusion configurations are physically accurate and realistic. Finally, we use the clip-art composited image together with its pseudo-groundtruth supervision data to learn robust 2D/3D object reconstruction under occlusion. + +Mining Unoccluded Objects: Given a stream of time-lapse data from a camera, our goal is to mine for unoccluded objects. In this context, "unoccluded" denotes instances where a detected object is not obstructed by any other object or truncated out of the field-of-view. On the time-lapse feed from a camera, we run instance segmentation [51] on each frame and use a simple object tracker [75] to track the detected bounding box and segmentation that belongs to "per + +![](images/76caf37dc8b6f5399b6adbb1219f3e0f7ac36b8c00daf73ae90ff43a15313786.jpg) +Figure 2. Given a time-lapse video, we automatically generate 2D/3D training data under severe occlusions. We start by detecting each object in the video, and unoccluded (fully visible) objects are identified. Each unoccluded object is then reconstructed using the ground plane and camera parameters. With the 3D pose, unoccluded objects are composited back into the same location (i.e., clip-art style) in a geometrically consistent approach. The composited image and its pseudo-groundtruth from off-the-shelf methods (e.g., segmentation, keypoints, shapes) are utilized to train a model that can produce accurate 2D/3D object reconstruction under severe occlusions. + +son” and “car” classes. The simplest heuristic for detecting unoccluded objects involves calculating the Intersection over Union (IOU) of detected objects. In previous methods like [62], objects were considered unoccluded if the intersection of their bounding box bottom with any other box was smaller than a threshold $\delta$ . However, this strategy becomes less reliable in scenarios where dynamic objects are occluded by static elements, such as vehicles being occluded by buildings, trees, poles, etc. + +To address this limitation, we train a simple Occlusion Classifier (OC) to categorize each detected object as either unoccluded or occluded, which is trained using a supervised approach with human annotators who label objects as unoccluded/occluded. Despite the OC module outperforming the heuristic filter in classifying objects' occlusion status, we did not observe a significant improvement in downstream evaluations (less than $\pm 0.5$ AP) due to the small amount of outliers overall in the training data. Therefore, it is not a mandatory component in our pipeline: the simple heuristics proposed in WALT [62] can be employed for the unoccluded object detection task without additional training. Nevertheless, we believe this data can be useful for occlusion reasoning tasks and we will publicly release it. + +Reconstructing Unoccluded Objects: Once unoccluded objects are identified, we describe the 3D reconstruction process of two primary classes: vehicles and humans. + +Each mined unoccluded vehicle is reconstructed following Li et al. [48]. We parameterize each vehicle's 3D keypoints $\mathbf{X}$ by a linear combination of the mean shape $\bar{\mathbf{Q}}$ and $K$ principal components $\mathbf{Q}_1, \ldots, \mathbf{Q}_K$ computed from an object CAD model dataset [47]: $\mathbf{X} = \bar{\mathbf{Q}} + \sum_{k=1}^{K} \alpha_k \mathbf{Q}_k$ , where $\alpha_k$ is the shape coefficient that needs to be optimized. Starting from the mean shape $\bar{\mathbf{Q}}$ , we first detect the 2D keypoints for each unoccluded vehicle and initialize the 6-DoF poses using EPnP [46]. Subsequently, for each track of detected vehicles, we optimize for the 6-DoF object poses while regularizing the shape parameter $\alpha_k$ to be constant for the same vehicle in different frames by minimizing reprojection errors. Additionally, we also constrain the objects to + +lie on the constant global ground plane, ensuring physically plausible reconstructions. + +In the context of human reconstruction, we employ the Human Mesh Recovery (HMR) method HMR 2.0 from Goel et al. [26]. Using HMR 2.0, we predict the SMPL pose and shape parameters for each detected unoccluded human [52]. To determine camera translation, we find the intersection between the backprojected ray from the bottom of the 2D bounding box with the ground plane. Given the camera's intrinsic matrix $\mathbf{K}$ , ground plane equation $\left[\mathbf{n} \quad d\right]$ , and $\mathbf{p}_b$ as the pixel coordinate of the bottom of the box, its depth is computed as $z_b = -\frac{d}{\mathbf{n}^T (\mathbf{K}^{-1} \mathbf{p}_b)}$ . For camera intrinsic parameters and ground plane equation, we employ a recent method by Vuong et al. [72] that utilizes Google Street View [27] to automatically calibrate stationary traffic cameras and infer the ground plane. The mined unoccluded object image with its corresponding segmentation mask, keypoint locations, and 3D poses are used in a clip-art based framework to generate 2D and 3D supervision data in severe occlusions as illustrated in Fig. 2. + +Pseudo-groundtruth Data Generation: To obtain supervision for occluded objects, we use image-based compositing to generate composited images, serving as input for training a network to learn 2D/3D reconstruction in severe occlusions. During the compositing process, various 2D/3D representations for each unoccluded object, including keypoints, segmentation, pose, and shape obtained from off-the-shelf methods, serve as pseudo-groundtruth supervision signals even after compositing. By using the clip-art-based compositing approach described below, we automatically generate a large number of realistic supervision signals in severe occlusions. Leveraging the 3D poses of mined unoccluded objects, we geometrically compose the object's image in a clip-art manner. Non-intersecting 3D objects are randomly sampled and reinserted into the background image (i.e., median image) at their original positions, ranging from the farthest to the closest. This approach ensures the creation of physically accurate and realistic occlusion configurations (see Fig. 3). Crucially, our geometry-based + +![](images/5cc224c2cd719fd14109731cf2343fa539e61ba4e3c67519f5ba32065aab93cf.jpg) +Cannnnnne + +![](images/39b738b48528dde64a75b36740e0078e0e32247088e1e836c12598d5aa8fb93f.jpg) + +![](images/91b2602258577f7eba6f1fb4b739cfe5e2e8f744a06836928c7ff839d9bfef07.jpg) +Segmentation + +![](images/01bf58d3fd137e25b5084594f6099cefdcd4e81d344cc612e5be4793664aeb92.jpg) + +![](images/2bf9951cc0751685e1e8a01d895eb874dcc0cccf6a3de3b3b37a5e0b159334d3.jpg) + +![](images/11365d4c43fa789325e85b67ddf0f44e3269c21515f0532fea9e05dc477f75e1.jpg) + +![](images/d4bcbdcfa6c2473ca6ee58da991fc9528dc1a6a61b0724bb67c564ec341a725f.jpg) + +![](images/8231a37bff3ef88a0325ccf3f6a1b4b95d274dafc5eb3d757e40e4b6c231b0ce.jpg) + +![](images/6c024c3653f48e58d274cf3b0ac10a7a9379f68ea2e6a6d230dc374460684148.jpg) +Cannnnnne + +![](images/1abe2ee2679dfb1498037657903327519bf9650027f24c1a9cb55206d929163a.jpg) + +![](images/5cd928f9cade3fd1e3171de5d5bf46cae30074121ff18584b58b1b158a52b866.jpg) +Segmentation + +![](images/3c0bbd1200c6dc6d3afee36980b1ba8ce95d69c65fa3c06fb5e62228104bd61d.jpg) + +![](images/7a7f123a7fee0fda041c96f77c194f5032171c3c4bf21a2c9c38b0e2cec97324.jpg) + +![](images/3bc7db2cd20d4dc8c16db483453989dc1bf817053d1535a52d251b57f2638283.jpg) + +![](images/2b9965965d87343fdda67b2b6a9b44d07bba9f1b060c0c49539939cea52f9df2.jpg) + +![](images/f01dbe634590424c56d0b017ca24c3b8d46bf4bff448ef04216495e3e0815fbe.jpg) +3D View + +![](images/8c4085b37ac6d4e93956bcd6d0b91de2f3692636e35ce9229a80e857bea9ba43.jpg) +Cannnnnnae + +![](images/3ca6651d5e30208a2eb1bdbef8da45efa4e957121837f2bdd9a522decaef839a.jpg) +Figure 3. Automatically generated 2D and 3D Clip-Art to supervise our network: Unoccluded objects are first mined using time-lapse imagery of WALT dataset [62]. Non-intersecting unoccluded objects are composited back into the background image in their respective original positions to preserve correct appearances. The resulting clip-art images, along with their corresponding amodal pseudo-groundtruth information, such as segmentation, keypoints, depth/normal maps, and 3D shapes, are shown. Our method generates realistic appearances from any stationary camera, incorporating diverse viewing geometries, weather conditions, lighting, and occlusion configurations. + +![](images/f6970d72d18ebf3a91156e6a9d5f134e178610db9a7663e2161dfa8a8e2940b7.jpg) +Segmentation + +![](images/240e09b24c6b33e3d3ac8931ab281d38a1a419f09cea53dbda947d5ea7a8efe5.jpg) + +![](images/a41f1f22920137e3d9a29bebfe76bebe6f456f0e277cabbb40d89837f00b518d.jpg) + +![](images/8e372824fa0e3cbde0b1a31516205794f9b4e709e11c089788c544f6deed3730.jpg) + +![](images/42d96b3aa2cffbd6141e140573d10e01bc3b2ea61376f68a40adc6835b085af0.jpg) + +![](images/6a749abfe66bcca2aefa611bbc548825ceb3fe862c6f0c845c0d98bd5ce505ff.jpg) +3D View + +approach distinguishes itself from the mere compositing of 2D images such as WALT [62], which can result in unrealistic composition (see the comparison in Fig. 4). Our data with physically accurate and realistic occlusion configurations contributes to more efficient and scalable training, as demonstrated in our experimental results. Each such generated clip-art image is accompanied by amodal segmentation masks, keypoints, 3D poses and shapes, provided by off-the-shelf methods that serve as pseudo-groundtruth supervision to learn object reconstruction under occlusion. + +# 4. Learning Reconstruction under Occlusion + +We have generated an extensive clip-art image dataset with corresponding 2D/3D pseudo-groundtruth representations (see Fig. 3). Using this supervision signal, we train a model capable of inferring holistic object representations in the + +presence of severe occlusions. + +Base Network and 2D Amodal Representations: Using the Swin Transformer [51] backbone with MaskRCNN-based [31] detection heads for its simplicity, our goal is to show improvement irrespective of the base model used. We train the default network with a multi-task loss $\mathcal{L}_{2\mathrm{D}}$ for each representation. These losses, computed on each sampled Region of Interest (RoI), include label classification loss $\mathcal{L}_{\mathrm{cls}}$ , bounding-box loss $\mathcal{L}_{\mathrm{box}}$ , binary cross-entropy loss on the mask branch $\mathcal{L}_{\mathrm{mask}}$ and keypoints $\mathcal{L}_{\mathrm{kp}}$ : $\mathcal{L}_{2\mathrm{D}} = \mathcal{L}_{\mathrm{cls}} + \mathcal{L}_{\mathrm{box}} + \mathcal{L}_{\mathrm{mask}} + \mathcal{L}_{\mathrm{kp}}$ . Crucially, with our clip-art dataset including amodal 2D bounding box, segmentation, and keypoints, our method effectively learns amodal instance segmentation and keypoint detection. + +3D Object Reconstruction: To regress object 3D pose and shape under occlusion, we extend the base network with a + +![](images/42d251e99c5694a20e861c16969fac73391067739ebf39a996517643465cb399.jpg) +Composited using WALT + +![](images/60fa54432d89f3e4683718d36e268c041fbe69429b9d8864391310c8b711d4b9.jpg) +Composed using our method +Figure 4. Comparison between images composited using the 2D-based method WALT2D [62] (left) and our 3D-based method WALT3D (right). It is evident that our 3D-based compositing method generates realistic and geometrically accurate occlusion configurations, in contrast to the 2D-based method (e.g., cars and people overlapping in an unfeasible way). + +3D regression branch. Each vehicle $N$ semantic 3D keypoints $\mathbf{X}$ are represented by a linear combination of the mean shape $\bar{\mathbf{Q}}$ and $K$ principal components $\mathbf{Q}_{1\dots K}$ , where $\alpha_{k}$ is the shape coefficients: $\mathbf{X} = \bar{\mathbf{Q}} +\sum_{k = 1}^{K}\alpha_{k}\mathbf{Q}_{k}$ . Thus, we want to regress for its 6-DoF pose $\{(\mathbf{R},\mathbf{t})\in \mathbf{SE}(3)\}$ and shape coefficients $\alpha_{k}$ . Given the predicted 2D keypoints and the corresponding 3D keypoints $\mathbf{X}$ , we can obtain its 6-DoF pose through a differentiable PnP layer [9]. Defining $\pi (\cdot)$ as the projection function with known intrinsic $\mathbf{K}$ , we can optimize for the keypoints reprojection loss $\mathcal{L}_{\mathrm{kp2D}}$ , shape loss $\mathcal{L}_{\mathrm{shape}}$ , and pose loss $\mathcal{L}_{\mathrm{pose}}$ as: + +$$ +\mathcal {L} _ {\mathrm {k p 2 D}} = \sum_ {i = 1} ^ {N} \left| \left| \pi \left(\mathbf {R} \mathbf {X} _ {i} + \mathbf {t}\right) - \mathbf {x} _ {i} ^ {*} \right| \right| _ {2} ^ {2}, \tag {1} +$$ + +$$ +\mathcal {L} _ {\mathrm {s h a p e}} = | | \alpha_ {k} - \alpha_ {k} ^ {*} | | _ {2} ^ {2}, \mathcal {L} _ {\mathrm {p o s e}} = | | [ \mathbf {R}, \mathbf {t} ] - [ \mathbf {R} ^ {*}, \mathbf {t} ^ {*} ] | | +$$ + +Here, $\mathbf{x}_i^*,\alpha_k^*,\mathbf{R}^*,\mathbf{t}^*$ represents the (pseudo)-groundtruth 2D keypoints, shape coefficients, rotation, and translation respectively. The final loss term is given as the weighted sum of the losses, and we learn amodal bounding box, segmentation, keypoint locations, 3D shape and pose in an end-to-end differentiable manner. + +For human reconstruction, we train our HMR predictor branch like HMR 2.0 [26]. Supervised by pseudogroundtruth SMPL pose $\theta^{*}$ and shape $\beta^{*}$ parameters [52], our network is guided by the loss functions on SMPL parameters $\mathcal{L}_{\mathrm{smp1}}$ , 3D keypoints $\mathcal{L}_{\mathrm{kp3D}}$ , 2D keypoints $\mathcal{L}_{\mathrm{kp2D}}$ and an adversarial loss $\mathcal{L}_{\mathrm{adv}}$ that ensures the model predicts valid 3D poses through the discriminator $D_{k}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {s m p l}} = | | [ \theta , \beta ] - [ \theta^ {*}, \beta^ {*} ] | | _ {2} ^ {2}, \\ \mathcal {L} _ {\mathrm {k p} 3 \mathrm {D}} = | | \mathbf {X} - \mathbf {X} ^ {*} | | _ {1}, \mathcal {L} _ {\mathrm {k p} 2 \mathrm {D}} = | | \pi (\mathbf {X}) - \mathbf {x} ^ {*} | | _ {1}, \tag {2} \\ \mathcal {L} _ {\mathrm {a d v}} = \sum_ {k} (D _ {k} (\theta_ {b}, \beta) - 1) ^ {2} \\ \end{array} +$$ + +Here, $\mathbf{x}^*$ and $\mathbf{X}^*$ denote the (pseudo)-groundtruth 2D and 3D keypoints respectively. With a weighted combination of these losses, we optimize for human 3D shape and pose in an end-to-end fashion. + +# 5. Dataset and Implementation Details + +Raw Time-lapse Dataset [62]: This dataset contains images from 20 stationary cameras in urban scenes captured over multiple years. The images are either 4K or HD and are captured at 60fps in short bursts. We used 30 days of data from 10 cameras resulting in approximately 3.3 million car and people instances for our experiments. + +WALT2D [62]: From the raw time-lapse dataset, we use the 2D-based compositing method proposed by WALT [62] to generate supervision data by pasting unoccluded objects back into the scene with different backgrounds resulting in 10000 training and 500 testing images per camera. Fig. 4 (left) illustrates an example of this dataset. It is important to note that this compositing method, relying solely on 2D information, leads to intersecting objects in the real world that may not faithfully represent the physical configuration. + +WALT3D (Ours): Also from the raw time-lapse dataset, our approach involves mining unoccluded objects and reinsering them into the scene, resulting in a comparable number of object instances as 2D-based compositing in WALT2D [62]. However, our approach employs a 3D-based compositing method to generate supervision data, as outlined in Section 3, in contrast to relying solely on 2D-based composition. Fig. 4 (right) shows an example clip-art image composited using our method, with our approach yielding training data with greater realism and physical accuracy compared to WALT2D. + +Vehicle 2D Keypoints Dataset (Ours): To determine 6- + +![](images/f68f3d22999f7fbdb3f9bcee3ba4191e3c4b073c162f6c0e47ff466845502f76.jpg) +Figure 5. Sample images from our new vehicle 2D keypoints dataset. The dataset contains a wide range of appearance variations including day and night and various traffic scenarios. + +DoF vehicle pose during compositing, we require high-quality 2D vehicle keypoints. Existing datasets [47, 59, 67, 77] offer human-annotated 2D vehicle keypoints but mainly focus on driving scenes or have limited training examples, lacking necessary appearance diversity for novel viewpoints like our stationary traffic cameras. Thus, we propose a new dataset with 7,018 images and 42,547 annotated instances from diverse viewpoints (see Fig. 5), each keypoint containing 2D location and occlusion status. With our dataset, we observe significant improvement in keypoint localization accuracy (66.41% to 80.12% on PCK@0.1), and consequently, an improvement in pseudo-groundtruth data quality. Further details are in the Supplementary. + +# 6. Baselines, Metrics, and Evaluations + +# 6.1. Baselines + +Detection and Instance Segmentation: All baselines share the same network architecture, employing a Swin [51] backbone as detailed in Section 4. The SWIN baseline is initially trained on the COCO dataset [50]. Subsequently, we perform further finetuning on the WALT2D and WALT3D (Ours) datasets, as detailed in Section 5. + +Vehicle 3D Reconstruction: We compare our approach with Occ-Net [62] and 3DRCNN [45] by changing their feature extraction backbone to be as close as possible to ours for fair comparison. They are trained on a combination of PASCAL3D+ [77], KITTI-3D [47], Carfusion [59], and ApolloCar3D [67], where we further finetune our baselines on the WALT2D and our WALT3D dataset, respectively. + +Human 3D Reconstruction: Our architecture mirrors HMR 2.0 [26] as described in Section 4. We initialize from the pretrained model, then finetune on our WALT3D data. + +# 6.2. Evaluation Metrics + +2D Metrics: We report Average Precision (AP) on bounding box detection $(\mathrm{AP}^{\mathrm{box}})$ and instance segmentation $(\mathrm{AP}^{\mathrm{mask}})$ following [50]. For keypoints, we use Percentage of Correct Keypoints (PCK) metric where a keypoint + +![](images/3e92bac6db158d80d0018b980a556c89a147dbbd8ced869af683550db799e3df.jpg) + +![](images/36e7458477813ad70a527e4d524f9e5ea4b478bf73def2efd2a3f09efdae5321.jpg) + +![](images/06891abaceb83f32acbb20cc538638e61f4dbd78dfea43b1822c61c13213c2e5.jpg) +Figure 6. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal vehicle detection, segmentation, 2D and 3D keypoint estimation. Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage. + +![](images/43523d45f41f7fbacf664b7ab73b87d60582bc3d5a9ada1688f924422af87126.jpg) + +![](images/1e4910d1c02e2ab20730c2066f4d9643a6edad906756cbde257fa7aa55e67d81.jpg) + +![](images/e631aac7db9bba79c36bd3bfed551a5f36af5327c9e6228d41994b54fa71c7a6.jpg) + +![](images/a362ce4b63535c3da1816bb1be0c3da5bd513f1d21590a905c50376a6ae1b04b.jpg) +Figure 7. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal human detection, segmentation, and human mesh recovery (HMR). Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage. + +![](images/c302e39b214b053277cddb179afeb84a9be8eace253ad9f4fde636c36e9b5b75.jpg) + +is considered correct if it lies within the radius $\alpha$ (with $0 < \alpha < 1$ ) of the ground-truth keypoint. + +3D Metrics: We use the standard metrics of previous work [34], reporting the Mean Per Joint Position Error (MPJME) between predicted and ground-truth 3D keypoints (aligned using the root joint) as well as 3D PCK. + +# 6.3. Ablation Analysis + +Robustness to Occlusions: We evaluate the effectiveness of our algorithm with varying occlusion levels. Similar to WALT [62], we use the pseudo-groundtruth segmentation masks from the evaluation set to group objects based on occlusion percentage. In 2D tasks such as detection and + +![](images/11bc1846cbde4d155582ab20d4e781c6b9fd4bbcc9e6237881d0bc22b8d0922f.jpg) +Figure 8. Comparison between WALT2D [62]) and our WALT3D approach. Since our method produces higher quality training data with more realistic occlusion configuration, our approach is especially useful in low-data regime. + +segmentation, the model trained with our WALT3D data significantly outperforms the baseline SWIN model trained on COCO [50], which only includes modal masks. This improvement is particularly evident in high occlusion percentages for both vehicle (Fig. 6) and human (Fig. 7). Compared to WALT2D [62], our data demonstrates slightly better accuracy across all occlusion levels, thanks to our 3D-based compositing method WALT3D producing physically accurate occlusion configurations with the same amount of training data. For 3D tasks such as vehicle (Fig. 6) and human (Fig. 7) 3D pose estimation, using our WALT3D data consistently outperforms other baselines, particularly in high occlusion percentages. Thus, models trained with our WALT3D data, even with imperfect pseudo-groundtruth from off-the-shelf methods, are robust to occlusion. + +3D-based Compositing helps Data Efficiency: Fig. 8 shows that our WALT3D data significantly improves performance across all training data fractions, including at low data regime (at $15\%$ of training data with $+13.9\mathrm{AP}^{\mathrm{mask}}$ improvement compared to training with WALT2D data). Essentially, a model trained on $25\%$ of WALT3D data achieves comparable AP to one trained on $60\%$ of WALT2D data. This improvement stems from our WALT3D approach generating higher quality and physically accurate occlusion configurations with the same training data volume, aligning more closely with the real-world distribution as in Fig. 4. Importantly, this data efficiency enables scaling to a larger number of scenes more efficiently. + +Cross-evaluations on other datasets: Table 1 presents quantitative results on COCO-Amodal [84] (COCO-A) (natural images) and KINS [58] (driving images), each containing human-annotated ground-truth 2D amodal instance masks. It shows that additional fine-tuning with our WALT3D data improves results in both KINS and COCO-A, highlighting the cross-evaluation performance of our data, with potential for even better generalization with the + +![](images/5c84d1ce64a46986ad9bb44fc1ce2a2aa3a8e491d20513f9bc3714d0a2065f4a.jpg) +Previous SOTA + +![](images/7b4695871f6f385107405b370d782682899a6788e61c1d433fd3de35e56e7dbf.jpg) + +![](images/1c89720c2764fe66dfb94bdaf2a708c453d99f8424d6650c5732393ce6a6f56b.jpg) +Figure 9. Qualitative results showing that our data improves amodal segmentation, keypoint prediction and 3D reconstruction compared to previous SOTA. + +![](images/48c14d578cf8539c8da2c81d4552ff38a63a996db8f4ed73a7c87d6160c38bd5.jpg) + +![](images/233b070b6eb184501aca11923d44eb232dcd11fa446b7786071b516b4e97eca0.jpg) +Ours + +increasing amount of automatically generated data. + +
Testing Data Training DataBox APMask AP
KINSCOCO-AKINSCOCO-A
COCO27.838.020.831.2
COCO + WALT2D28.240.422.436.4
COCO + WALT3D28.843.223.538.7
COCO + COCO-A + KINS36.846.932.942.7
COCO + COCO-A + KINS + WALT2D37.447.933.343.5
COCO + COCO-A + KINS + WALT3D38.848.735.645.3
+ +Table 1. Finetuning with our WALT3D data improves generalization across KINS and COCO-A (vehicle & people). Additionally, combining our data with domain-specific data like KINS and COCO-A further enhances performance in their domains. + +Cross-evaluations on pedestrian tracking: We evaluate how our WALT3D data helps pedestrian tracking in Table 2. To highlight how better detections lead to better tracking, we use a simple tracker SORT [4] with two detectors: one pretrained on COCO and one finetuned with our WALT3D data. On MOT17-train, the model finetuned with WALT3D data produces high-quality detections under strong occlusions, resulting in improvements over the baseline in all metrics, particularly those favoring better detection quality like MOTA and DetA (see metrics details in [53]). + +
Detector Train DataMOTA↑DetA↑IDF1↑HOTA↑
COCO42.240.346.740.1
COCO + WALT3D47.243.249.542.0
+ +Table 2. MOT17-train results using SORT with two detectors: one pretrained on COCO and one fine-tuned with our WALT3D data. + +Qualitative Results: Results of our method can be seen in Fig. 9 and Fig. 10. Comparing to previous methods, observe that our method can produce robust 2D/3D predictions, even under challenging high-occlusion scenarios. + +# 7. Discussion and Conclusion + +Our data generation framework is method-agnostic, accommodating various robust object pose estimation alterna + +![](images/ccbf88efd1bb580fa878e4ae6e0b5c316184ee4d5a613fc3d4e6346df7781084.jpg) + +![](images/51ad2b5dcc6bed33f649f678b501cb62c5f053f008395fb8e3ec9eb651f72abe.jpg) + +![](images/2751cadcf00ee45e6c78370eef789f1576401f2b517e912b8caa89e39ffe3f0b.jpg) + +![](images/452d2486b5d256d025f88c7730d8f836df23d730c5be0d9807e654129184494c.jpg) + +![](images/ecb1b346aa61a46ea543f2a456906e80ff3108edfa464a80b61bf4d59f6a5a5a.jpg) + +![](images/a25042638fd834a5e91dab330423a7b9efbcfa678c956b1bb1f4868e00ba3bc2.jpg) + +![](images/543584f73f8c002f24b62ad3c012bdadb89a4041eac4b19f608686e037032b9d.jpg) + +![](images/2f6e3100e3c1ce33eeba61d0470fe127f1ab2e82a59e0d2f95d4a4531c4cdb97.jpg) + +![](images/235807ed81f953666ecaeac0082d3ea38de823bb91d5b8b60392e8714036552e.jpg) + +![](images/5399533297c714632f8830e8efce5c5f417d8092d5fb0f6f81463deb023d2a03.jpg) + +![](images/cb4b8542f64e925969b965f299c478e2c14ea405820cfc8e25505b3289148d35.jpg) + +![](images/0e1a6bcc678c48932502e78b9b1851e67cbe98f7a37cf4bc777dbe745be99734.jpg) + +![](images/a6dce8c450b37d38d0b352f38bddd37ba81b0e5b98a40784f10aa89209521e6e.jpg) + +![](images/6524742e917981b262cdc6c17c523d73466fe04d04a80c6149bc074b05c28126.jpg) + +![](images/8f9f2a98ee5fedec2f926dbec7bb8a148aa45e868b37213aeca9c69127f64d05.jpg) + +![](images/3bc3618d5e4807490fda3ff465da208174db30136c1fec39896c05fbecd3f711.jpg) + +![](images/5a89a001ca50e1525be270798caeb74aa75826f4cc067985fea0a3afe01fa724.jpg) + +![](images/d92175515ce611fb0cb3047af03db622c1c19ec6659ac6b755ec280b5e4406ca.jpg) + +![](images/ed8424bb2ff5ea25edd655c2f3f038386c6af56830d7f419cf98f968387582ed.jpg) + +![](images/5d7c6ce62630e89dec5784ff8e2ea92bda020383dae2c2b8e1b66e2ac6eb93a6.jpg) + +![](images/a9351148fd96970ee0295c863557a24279f0eea6ea48a6f3118af9e4a8273585.jpg) +Input Image + +![](images/2b8f5970b9821bb857dcf272c919d374721c794de0eba53b2b1169cfa26eddcd.jpg) +Amodal Segmentation + +![](images/d3346e2b90ad0b9e4d655e650d22d93f2cf77cec3a62ccfc1e4639a2e2c9f979.jpg) +Amodal Keypoints/Shapes + +![](images/7f19aff6d7c4ef2bedd6f0f4575dc04bb206b314bcddc660ee6b5362d6c1a7db.jpg) +3D View-1 + +![](images/c98d465c2f01795d93b15c5811bcbf24729b242edaaf45c4a0ce086cdc29d1dc.jpg) +3D View-2 +Figure 10. We show qualitative results of our method on real data. Our method produces accurate amodal segmentation, keypoints, as well as 3D poses and shapes across diverse poses and occlusion configurations. In particular, we show results on different level and types of occlusions like vehicle-vehicle (row 1, 2, 3), vehicle-people (row 3), and people-people occlusion (row 4, 5). + +tives [54, 73] as well as methods like VIBE [39] for high-quality human reconstructions from videos. Moreover, as our pseudo-groundtruth data includes metric-scale depth information using ground plane, we can augment existing datasets like Relative Human [69] to enhance the robustness of 3D multi-human reconstruction methods to occlusion. + +Limitations: Further research is required for generalization across diverse views for it to be used as a generic solution. It also assumes a mean shape or a parametric object model is available, posing challenges for rare objects. Addressing appearance inconsistencies (e.g., variations in lighting) in clip-art images is a promising research direction. + +In conclusion, our method introduces an automated approach to generate a realistic dataset for reconstructing dynamic objects under occlusions from time-lapse imagery. It demonstrates significant improvements in both 2D and 3D object reconstruction, particularly in busy urban scenes with diverse occlusion configurations. + +Potential Societal Impact. We do not perform any human subjects research from these cameras. For privacy, we blur faces and license plates in all images intended for release. + +Acknowledgements: This work was supported in part by an NSF Grant CNS-2038612, a US DOT grant 69A3551747111 through the Mobility21 UTC and grants 69A3552344811 and 69A3552348316 through the Safety21 UTC. + +# References + +[1] William Agnew, Christopher Xie, Aaron Walsman, Octavian Murad, Yubo Wang, Pedro Domingos, and Siddhartha Srinivasa. Amodal 3d reconstruction for robotic manipulation via stability and connectivity. In CoRL, 2021. 2 +[2] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets deep learning for car instance segmentation in urban scenes. In BMVC, 2017. 2 +[3] Asra Aslam. Detecting objects in less response time for processing multimedia events in smart cities. In CVPR, 2022. 1 +[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 7 +[5] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 2 +[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2 +[7] Berk Calli, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. IJRR, 2017. 1 +[8] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In CVPR, 2019. 2 +[9] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In CVPR, 2020. 5 +[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1 +[11] Falak Chhaya, Dinesh Reddy, Sarthak Upadhyay, Visesh Chari, M. Zeeshan Zia, and K. Madhava Krishna. Monocular reconstruction of vehicles: Combining slam with shape priors. In ICRA, 2016. 2 +[12] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In ECCV, 2016. 2 +[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1 +[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1 +[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. + +ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 1 +[16] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In CVPR, 2022. 1 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[18] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2 +[19] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 2 +[20] Patrick Follmann, Rebecca König, Philipp Hartinger, Michael Klostermann, and Tobias Böttger. Learning to see the invisible: End-to-end trainable amodal instance segmentation. In WACV, 2019. 2 +[21] Rik Fransens, Christoph Strecha, and Luc Van Gool. A mean field em-algorithm for coherent occlusion handling in mape-estimation prob. In CVPR, 2006. 2 +[22] Tianshi Gao, Benjamin Packer, and Daphne Koller. A segmentation-aware object detection model with occlusion handling. In CVPR, 2011. 2 +[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1, 2 +[24] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2 +[25] Holger Glasl, David Schreiber, Nikolaus Viertl, Stephan Veigl, and Gustavo Fernandez. Video based traffic congestion prediction on an embedded system. In ITSC, 2008. 1 +[26] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3, 5, 6 +[27] Google. Google Street View. https://www.google.com/streetview/.3 +[28] Can Gumeli, Angela Dai, and Matthias Nießner. Roca: robust cad model retrieval and alignment from a single image. In CVPR, 2022. 2 +[29] Ruiqi Guo and Derek Hoiem. Beyond the line of sight: labeling the underlying surfaces. In ECCV, 2012. 2 +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1 +[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 1, 2, 4 +[32] Y.-T. Hu, H.-S. Chen, K. Hui, J.-B. Huang, and A. G. Schwing. SAIL-VOS: Semantic Amodal Instance Level Video Object Segmentation - A Synthetic Dataset and Baselines. In CVPR, 2019. 2 + +[33] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In ECCV, 2020. 2 +[34] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 6 +[35] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1 +[36] Gaetano Kanizsa. Organization in vision: Essays on gestalt perception. Praeger Publishers, 1979. 1 +[37] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, 2020. 2 +[38] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2 +[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020. 8 +[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, 2021. 2 +[41] Philipp Krahenbuhl. Free supervision from video games. In CVPR, 2018. 2 +[42] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. 1 +[43] Alexander Krull, Eric Brachmann, Frank Michel, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Learning analysis-by-synthesis for 6d pose estimation in rgb-d images. In ICCV, 2015. 1, 2 +[44] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 1 +[45] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, 2018. 2, 6 +[46] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 3 +[47] Chi Li, M Zeeshan Zia, Quoc-Huy Tran, Xiang Yu, Gregory D Hager, and Manmohan Chandraker. Deep supervision with shape concepts for occlusion-aware 3d object parsing. In CVPR, 2017. 2, 3, 6 +[48] Fangyu Li, N Dinesh Reddy, Xudong Chen, and Srinivasa G Narasimhan. Traffic4d: Single view longitudinal 4d reconstruction of repetitious activity using self-supervised experts. In IV, 2021. 3 +[49] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 1 +[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7 + +[51] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 6 +[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. SIGGRAPH Asia, 2015. 3, 5 +[53] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 2021. 7 +[54] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In ECCV, 2022. 2, 8 +[55] M. Naphade, S. Wang, D. C. Anastasiu, Z. Tang, M. Chang, Y. Yao, L. Zheng, M. Shaiqur Rahman, A. Venkatachalapathy, A. Sharma, Q. Feng, V. Ablavsky, S. Sclaroff, P. Chakraborty, A. Li, S. Li, and R. Chellappa. The 6th ai city challenge. In CVPRW, 2022. 1 +[56] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2 +[57] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2 +[58] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 1, 2, 7 +[59] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In CVPR, 2018. 2, 6 +[60] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In CVPR, 2019. 1, 2 +[61] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15190-15200, 2021. 2 +[62] N. Dinesh Reddy, Robert Tamburo, and Srinivasa G. Narasimhan. Walt: Watch and learn 2d amodal representation from time-lapse imagery. In CVPR, 2022. 1, 3, 4, 5, 6, 7 +[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1 +[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2 +[65] Samuel Schulter, Menghua Zhai, Nathan Jacobs, and Manmohan Chandraker. Learning to look around objects for top-view representations of outdoor scenes. In ECCV, 2018. 2 +[66] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 2 + +[67] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In CVPR, 2019. 2, 6 +[68] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2 +[69] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J. Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, 2022. 8 +[70] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In CVPR, 2015. 1 +[71] Andrea Vedaldi and Andrew Zisserman. Structured output regression for detection with partial truncation. In NeurIPS, 2009. 2 +[72] Khiem Vuong, Robert Tamburo, and Srinivasa G. Narasimhan. Toward planet-wide traffic camera calibration. In WACV, 2024. 3 +[73] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. In ICLR, 2021. 2, 8 +[74] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1, 2 +[75] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2 +[76] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2 +[77] Y. Xiang, R. Mottaghi, and S. Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 1, 6 +[78] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015. 1 +[79] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2 +[80] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust instance segmentation through reasoning about multi-object occlusion. In CVPR, 2021. 2 +[81] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1 +[82] Xiaowei Zhou, Spyridon Leonardos, Xiaoyan Hu, and Kostas Daniilidis. 3d shape estimation from 2d landmarks: A convex relaxation approach. In CVPR, 2015. 2 +[83] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2 +[84] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 1, 2, 7 + +[85] M Zeeshan Zia, Michael Stark, and Konrad Schindler. Towards scene understanding with detailed 3d object representations. IJCV, 2015. 2 \ No newline at end of file diff --git a/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/images.zip b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba979951ee184d37c705b923b2fde8aff2b17dde --- /dev/null +++ b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:388a085fa9402be2736ed84fe0c37a2ca8be5891c89a4f0764d099e27a451d88 +size 1082940 diff --git a/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/layout.json b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a88155ba775342b13c1c1c82297b9ebf1ffe137d --- /dev/null +++ b/2024/WALT3D_ Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects Under Occlusion/layout.json @@ -0,0 +1,11541 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "type": "text", + "content": "WALT3D: Generating Realistic Training Data from Time-Lapse Imagery for Reconstructing Dynamic Objects under Occlusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 160, + 146, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 160, + 146, + 175 + ], + "spans": [ + { + "bbox": [ + 65, + 160, + 146, + 175 + ], + "type": "text", + "content": "Khiem Vuong" + }, + { + "bbox": [ + 65, + 160, + 146, + 175 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 160, + 263, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 160, + 263, + 175 + ], + "spans": [ + { + "bbox": [ + 170, + 160, + 263, + 175 + ], + "type": "text", + "content": "N Dinesh Reddy" + }, + { + "bbox": [ + 170, + 160, + 263, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 288, + 161, + 373, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 161, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 288, + 161, + 373, + 175 + ], + "type": "text", + "content": "Robert Tamburo1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 399, + 161, + 523, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 161, + 523, + 175 + ], + "spans": [ + { + "bbox": [ + 399, + 161, + 523, + 175 + ], + "type": "text", + "content": "Srinivasa G. Narasimhan1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 180, + 405, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 180, + 405, + 209 + ], + "spans": [ + { + "bbox": [ + 188, + 180, + 405, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 188, + 180, + 405, + 209 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 188, + 180, + 405, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 188, + 180, + 405, + 209 + ], + "type": "text", + "content": "Amazon https://www.cs.cmu.edu/~walt3d" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 186, + 285 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 186, + 285 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 186, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 186, + 285 + ], + "type": "image", + "image_path": "737840bcd566ca8f0201d04a7c47075eabd7e326e7882f2c454a2bd541023056.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 288, + 133, + 298 + ], + "lines": [ + { + "bbox": [ + 89, + 288, + 133, + 298 + ], + "spans": [ + { + "bbox": [ + 89, + 288, + 133, + 298 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 186, + 218, + 323, + 286 + ], + "blocks": [ + { + "bbox": [ + 186, + 218, + 323, + 286 + ], + "lines": [ + { + "bbox": [ + 186, + 218, + 323, + 286 + ], + "spans": [ + { + "bbox": [ + 186, + 218, + 323, + 286 + ], + "type": "image", + "image_path": "f1ec3367fc2b5d9696b0fe17b98f269d9ec30a777d4e688db180f9d705bcd0d9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 288, + 296, + 297 + ], + "lines": [ + { + "bbox": [ + 213, + 288, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 213, + 288, + 296, + 297 + ], + "type": "text", + "content": "Predicted Segmentation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 299, + 544, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 544, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 544, + 321 + ], + "type": "text", + "content": "Figure 1. Models trained on our automatically generated data from time-lapse imagery can reliably estimate amodal 2D bounding box, segmentation as well as 3D shape and pose despite the complex occlusions presented in the input image." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 218, + 460, + 286 + ], + "blocks": [ + { + "bbox": [ + 323, + 218, + 460, + 286 + ], + "lines": [ + { + "bbox": [ + 323, + 218, + 460, + 286 + ], + "spans": [ + { + "bbox": [ + 323, + 218, + 460, + 286 + ], + "type": "image", + "image_path": "d62346d8dd0e265d0ce83df131a5765dea8a9d26dc01018b9659960f2b7d2e3e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 288, + 414, + 298 + ], + "lines": [ + { + "bbox": [ + 356, + 288, + 414, + 298 + ], + "spans": [ + { + "bbox": [ + 356, + 288, + 414, + 298 + ], + "type": "text", + "content": "Predicted Shape" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 461, + 220, + 546, + 286 + ], + "blocks": [ + { + "bbox": [ + 461, + 220, + 546, + 286 + ], + "lines": [ + { + "bbox": [ + 461, + 220, + 546, + 286 + ], + "spans": [ + { + "bbox": [ + 461, + 220, + 546, + 286 + ], + "type": "image", + "image_path": "e298ce3a518fd885aa05f40994654b37ee973e06b1c7a111b15a459ba56f1f9c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 288, + 518, + 297 + ], + "lines": [ + { + "bbox": [ + 487, + 288, + 518, + 297 + ], + "spans": [ + { + "bbox": [ + 487, + 288, + 518, + 297 + ], + "type": "text", + "content": "3D View" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 343, + 191, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 343, + 191, + 356 + ], + "spans": [ + { + "bbox": [ + 143, + 343, + 191, + 356 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 368, + 289, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 289, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 289, + 584 + ], + "type": "text", + "content": "Current methods for 2D and 3D object understanding struggle with severe occlusions in busy urban environments, partly due to the lack of large-scale labeled ground-truth annotations for learning occlusion. In this work, we introduce a novel framework for automatically generating a large, realistic dataset of dynamic objects under occlusions using freely available time-lapse imagery. By leveraging off-the-shelf 2D (bounding box, segmentation, keypoint) and 3D (pose, shape) predictions as pseudo-groundtruth, unoccluded 3D objects are identified automatically and composited into the background in a clip-art style, ensuring realistic appearances and physically accurate occlusion configurations. The resulting clip-art image with pseudogroundtruth enables efficient training of object reconstruction methods that are robust to occlusions. Our method demonstrates significant improvements in both 2D and 3D reconstruction, particularly in scenarios with heavily occluded objects like vehicles and people in urban scenes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 605, + 127, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 127, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 127, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 625, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 696 + ], + "type": "text", + "content": "In recent years, remarkable progress has been made in advancing scene understanding tasks such as object detection [49, 51, 63], tracking [14, 78, 81], segmentation [10, 31], and 3D reconstruction [35, 44, 70]. These achievements are mainly attributed to the availability of large-scale datasets [13, 15, 23, 50, 77] and architectural innova" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 332, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 332, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 332, + 545, + 452 + ], + "type": "text", + "content": "tions [17, 30, 42]. Despite this progress, a notable challenge persists in scenarios with severe occlusion, where only a portion of the object is visible: an object may be partially occluded by other objects or truncated by the camera's field-of-view. This phenomenon is referred to as amodal perception [36], and it is hard to intuitively infer their complete shapes. Overcoming these challenges is important to advance many smart cities applications as well as robotics applications, where the number of cameras on vehicles and city infrastructure is rapidly increasing [3, 25, 55]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 457, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 601 + ], + "type": "text", + "content": "Efforts to learn holistic representations necessitate a substantially annotated and realistic dataset. While recent works such as KINS [58], COCO-Amodal [84], and Ithaca365 [16] have contributed by annotating some amodal ground-truth data, the available annotations remain limited. This scarcity is primarily due to the inherent difficulty in obtaining supervision for amodal representations, as labeling hidden parts of objects is a difficult task for people to accomplish consistently [58, 60, 84]. Moreover, 3D annotations under occlusions, including object shape and pose [7, 43, 74], are even meager due to the difficulty of annotating 3D data, posing challenges for 3D prediction tasks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Expanding on WALT [62], we utilize time-lapse videos from stationary cameras to synthesize realistic occlusion scenarios by extracting unoccluded objects and composite them back into the background image at their original positions. Unlike WALT which focuses solely on compositing and learning 2D tasks, our approach extends to generating high-quality 3D pseudo-groundtruth data for robust 3D object reconstruction under occlusion. Additionally, our 3D-based compositing method, named WALT3D, in" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 220, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 220, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 220, + 712 + ], + "type": "text", + "content": "* denotes equal contribution and joint first author" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9514" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "contrast to WALT's simplistic 2D compositing, produces physically accurate occlusion configurations, leading to increased training data efficiency and scalability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "content": "We start with the observation that, although not perfect, existing off-the-shelf methods demonstrate good accuracy in both 2D (segmentation [31], keypoints [67, 68]) and 3D (pose, shape [37, 54, 73]) prediction tasks, especially on unoccluded objects. Thus, they can be used as \"pseudo-groundtruth\" to improve the robustness of existing approaches in occlusion scenarios. Next, we randomly select these unoccluded, non-intersecting 3D objects and put them back into the background image at their original positions (i.e., clip-art style). Specifically, we arrange them based on their distance from the camera, ensuring physically accurate and realistic occlusion configurations. Each such resulting clip-art image is accompanied by amodal bounding box, segmentation masks, and 3D poses and shapes – referred to as \"pseudo-groundtruth\" as these were predicted by off-the-shelf methods on the original unoccluded objects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 395 + ], + "type": "text", + "content": "Through extensive experiments, we demonstrate the effectiveness of our data in both vehicle and human reconstruction, particularly in scenarios with heavy occlusions. It is important to note that our method does not require any human labeling and hence is easily scalable and serves as an effective method to automatically generate realistic training data for reconstructing dynamic objects under occlusion. Our notable technical contributions are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 396, + 286, + 479 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 396, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 430 + ], + "type": "text", + "content": "- We introduce a novel method that automatically generates 2D/3D supervision data from time-lapse imagery with realistic occlusion configurations without human labeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 286, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 286, + 479 + ], + "type": "text", + "content": "- We demonstrate that the utilization of our generated data significantly enhances training efficiency and the accuracy of 2D/3D object reconstruction on real-world data, particularly in scenarios with high occlusion." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 490, + 138, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 138, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 138, + 502 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 617 + ], + "type": "text", + "content": "Occlusion Reasoning: Understanding and reasoning occlusions has been extensively studied for decades [21, 22, 65]. Bad predictions due to occlusions are dealt with as noise/outliers in robust estimators. On the other hand, occlusions are explicitly treated as missing parts in model fitting methods [71, 82]. But severe occlusions, such as when a large part of an object is blocked, can result in poor model fitting [28, 85], especially when they attempt to simultaneously estimate the model fit as well as the missing parts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "2D Amodal Representation: Although the effects of occlusion on visual reasoning has been widely studied, estimating the amodal representation (i.e. both the occluded and visible regions) has only been recently explored. Initial attempts [20, 29, 58, 84] use a supervised learning paradigm using small datasets [58, 84] where humans have annotated occlusions to the best of their abilities. Some methods [59, 60, 66] have explored using multiple views to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "provide accurate supervision for occluded parts but are not scalable due to capture limitations. To expand supervision, several methods synthesize occlusions to varying degrees of realism. But pure CG renderings [2, 18, 19, 32, 38, 47, 80] suffer from a wide domain-gap [41, 65]. While driving datasets [6, 8, 23, 79] provide 2D amodal bounding boxes, annotating additional representations like segmentation and keypoints under occlusion remains challenging. Some methods, like Ghiasi et al. [24], address this by randomly copy-pasting objects onto diverse background images. However, as these methods solely rely on 2D information, they generate unrealistic occlusion configurations leading to poor performance and limited training efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 431 + ], + "type": "text", + "content": "3D Reconstruction Under Occlusion: Reconstruction under severe occlusion is still in the nascent stages of research. Most algorithms developed focus on self-occluded objects with shape completion from partial observations [12, 56, 83]. On the other hand, shape models fitting for objects only with the visible regions either from images [28, 33, 45, 61, 64] or depth sensors [1, 11, 57, 76] have been explored. Due to the inherent challenges in annotating 3D information, ground-truth data for 3D object understanding is limited, often confined to specific object sets [43, 74], indoor [5], or driving scenarios [6, 23, 67], and struggles to generalize well to novel viewpoints such as stationary traffic cameras. Recent methods addressing pose estimation for vehicles [54, 73] and people [26, 40] have demonstrated robustness in handling occlusion, either implicitly or explicitly. Thus, our dataset generation method can significantly enhance the robustness of these approaches to occlusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 442, + 517, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 517, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 517, + 455 + ], + "type": "text", + "content": "3. Generating Realistic Supervision Data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": "An overview of our approach is shown in Figure 2. From a time-lapse video, we identify unoccluded objects and extract their 2D attributes such as segmentation and keypoints. We then use off-the-shelf 3D object reconstruction methods to obtain the pose and shape of these objects, constrained by the camera intrinsics and ground plane. After that, we re-insert these non-intersecting 3D objects into the background image at their original positions in a \"clip-art\" style, arranged based on their distance from the camera to ensure the occlusion configurations are physically accurate and realistic. Finally, we use the clip-art composited image together with its pseudo-groundtruth supervision data to learn robust 2D/3D object reconstruction under occlusion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Mining Unoccluded Objects: Given a stream of time-lapse data from a camera, our goal is to mine for unoccluded objects. In this context, \"unoccluded\" denotes instances where a detected object is not obstructed by any other object or truncated out of the field-of-view. On the time-lapse feed from a camera, we run instance segmentation [51] on each frame and use a simple object tracker [75] to track the detected bounding box and segmentation that belongs to \"per" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9515" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 72, + 521, + 175 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 521, + 175 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 521, + 175 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 521, + 175 + ], + "type": "image", + "image_path": "76caf37dc8b6f5399b6adbb1219f3e0f7ac36b8c00daf73ae90ff43a15313786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 234 + ], + "type": "text", + "content": "Figure 2. Given a time-lapse video, we automatically generate 2D/3D training data under severe occlusions. We start by detecting each object in the video, and unoccluded (fully visible) objects are identified. Each unoccluded object is then reconstructed using the ground plane and camera parameters. With the 3D pose, unoccluded objects are composited back into the same location (i.e., clip-art style) in a geometrically consistent approach. The composited image and its pseudo-groundtruth from off-the-shelf methods (e.g., segmentation, keypoints, shapes) are utilized to train a model that can produce accurate 2D/3D object reconstruction under severe occlusions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 288, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 351 + ], + "type": "text", + "content": "son” and “car” classes. The simplest heuristic for detecting unoccluded objects involves calculating the Intersection over Union (IOU) of detected objects. In previous methods like [62], objects were considered unoccluded if the intersection of their bounding box bottom with any other box was smaller than a threshold " + }, + { + "bbox": [ + 46, + 243, + 288, + 351 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 243, + 288, + 351 + ], + "type": "text", + "content": ". However, this strategy becomes less reliable in scenarios where dynamic objects are occluded by static elements, such as vehicles being occluded by buildings, trees, poles, etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 519 + ], + "type": "text", + "content": "To address this limitation, we train a simple Occlusion Classifier (OC) to categorize each detected object as either unoccluded or occluded, which is trained using a supervised approach with human annotators who label objects as unoccluded/occluded. Despite the OC module outperforming the heuristic filter in classifying objects' occlusion status, we did not observe a significant improvement in downstream evaluations (less than " + }, + { + "bbox": [ + 46, + 353, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\pm 0.5" + }, + { + "bbox": [ + 46, + 353, + 287, + 519 + ], + "type": "text", + "content": " AP) due to the small amount of outliers overall in the training data. Therefore, it is not a mandatory component in our pipeline: the simple heuristics proposed in WALT [62] can be employed for the unoccluded object detection task without additional training. Nevertheless, we believe this data can be useful for occlusion reasoning tasks and we will publicly release it." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 521, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 556 + ], + "type": "text", + "content": "Reconstructing Unoccluded Objects: Once unoccluded objects are identified, we describe the 3D reconstruction process of two primary classes: vehicles and humans." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Each mined unoccluded vehicle is reconstructed following Li et al. [48]. We parameterize each vehicle's 3D keypoints " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " by a linear combination of the mean shape " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Q}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " principal components " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_1, \\ldots, \\mathbf{Q}_K" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " computed from an object CAD model dataset [47]: " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\bar{\\mathbf{Q}} + \\sum_{k=1}^{K} \\alpha_k \\mathbf{Q}_k" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " is the shape coefficient that needs to be optimized. Starting from the mean shape " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Q}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", we first detect the 2D keypoints for each unoccluded vehicle and initialize the 6-DoF poses using EPnP [46]. Subsequently, for each track of detected vehicles, we optimize for the 6-DoF object poses while regularizing the shape parameter " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " to be constant for the same vehicle in different frames by minimizing reprojection errors. Additionally, we also constrain the objects to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "text", + "content": "lie on the constant global ground plane, ensuring physically plausible reconstructions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "text", + "content": "In the context of human reconstruction, we employ the Human Mesh Recovery (HMR) method HMR 2.0 from Goel et al. [26]. Using HMR 2.0, we predict the SMPL pose and shape parameters for each detected unoccluded human [52]. To determine camera translation, we find the intersection between the backprojected ray from the bottom of the 2D bounding box with the ground plane. Given the camera's intrinsic matrix " + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "text", + "content": ", ground plane equation " + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{n} \\quad d\\right]" + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_b" + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "text", + "content": " as the pixel coordinate of the bottom of the box, its depth is computed as " + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "inline_equation", + "content": "z_b = -\\frac{d}{\\mathbf{n}^T (\\mathbf{K}^{-1} \\mathbf{p}_b)}" + }, + { + "bbox": [ + 304, + 269, + 546, + 484 + ], + "type": "text", + "content": ". For camera intrinsic parameters and ground plane equation, we employ a recent method by Vuong et al. [72] that utilizes Google Street View [27] to automatically calibrate stationary traffic cameras and infer the ground plane. The mined unoccluded object image with its corresponding segmentation mask, keypoint locations, and 3D poses are used in a clip-art based framework to generate 2D and 3D supervision data in severe occlusions as illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": "Pseudo-groundtruth Data Generation: To obtain supervision for occluded objects, we use image-based compositing to generate composited images, serving as input for training a network to learn 2D/3D reconstruction in severe occlusions. During the compositing process, various 2D/3D representations for each unoccluded object, including keypoints, segmentation, pose, and shape obtained from off-the-shelf methods, serve as pseudo-groundtruth supervision signals even after compositing. By using the clip-art-based compositing approach described below, we automatically generate a large number of realistic supervision signals in severe occlusions. Leveraging the 3D poses of mined unoccluded objects, we geometrically compose the object's image in a clip-art manner. Non-intersecting 3D objects are randomly sampled and reinserted into the background image (i.e., median image) at their original positions, ranging from the farthest to the closest. This approach ensures the creation of physically accurate and realistic occlusion configurations (see Fig. 3). Crucially, our geometry-based" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9516" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 72, + 199, + 131 + ], + "blocks": [ + { + "bbox": [ + 77, + 114, + 86, + 148 + ], + "lines": [ + { + "bbox": [ + 77, + 114, + 86, + 148 + ], + "spans": [ + { + "bbox": [ + 77, + 114, + 86, + 148 + ], + "type": "text", + "content": "Cannnnnne" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 72, + 199, + 131 + ], + "lines": [ + { + "bbox": [ + 93, + 72, + 199, + 131 + ], + "spans": [ + { + "bbox": [ + 93, + 72, + 199, + 131 + ], + "type": "image", + "image_path": "5cc224c2cd719fd14109731cf2343fa539e61ba4e3c67519f5ba32065aab93cf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 93, + 132, + 198, + 191 + ], + "blocks": [ + { + "bbox": [ + 93, + 132, + 198, + 191 + ], + "lines": [ + { + "bbox": [ + 93, + 132, + 198, + 191 + ], + "spans": [ + { + "bbox": [ + 93, + 132, + 198, + 191 + ], + "type": "image", + "image_path": "39b738b48528dde64a75b36740e0078e0e32247088e1e836c12598d5aa8fb93f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 72, + 303, + 131 + ], + "blocks": [ + { + "bbox": [ + 200, + 72, + 303, + 131 + ], + "lines": [ + { + "bbox": [ + 200, + 72, + 303, + 131 + ], + "spans": [ + { + "bbox": [ + 200, + 72, + 303, + 131 + ], + "type": "image", + "image_path": "91b2602258577f7eba6f1fb4b739cfe5e2e8f744a06836928c7ff839d9bfef07.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 72, + 356, + 83 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 356, + 83 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 356, + 83 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 200, + 133, + 302, + 191 + ], + "blocks": [ + { + "bbox": [ + 200, + 133, + 302, + 191 + ], + "lines": [ + { + "bbox": [ + 200, + 133, + 302, + 191 + ], + "spans": [ + { + "bbox": [ + 200, + 133, + 302, + 191 + ], + "type": "image", + "image_path": "01bf58d3fd137e25b5084594f6099cefdcd4e81d344cc612e5be4793664aeb92.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 105, + 406, + 133 + ], + "blocks": [ + { + "bbox": [ + 304, + 105, + 406, + 133 + ], + "lines": [ + { + "bbox": [ + 304, + 105, + 406, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 105, + 406, + 133 + ], + "type": "image", + "image_path": "2bf9951cc0751685e1e8a01d895eb874dcc0cccf6a3de3b3b37a5e0b159334d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 134, + 407, + 191 + ], + "blocks": [ + { + "bbox": [ + 304, + 134, + 407, + 191 + ], + "lines": [ + { + "bbox": [ + 304, + 134, + 407, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 407, + 191 + ], + "type": "image", + "image_path": "11365d4c43fa789325e85b67ddf0f44e3269c21515f0532fea9e05dc477f75e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 513, + 132 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 513, + 132 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 513, + 132 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 513, + 132 + ], + "type": "image", + "image_path": "d4bcbdcfa6c2473ca6ee58da991fc9528dc1a6a61b0724bb67c564ec341a725f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 408, + 133, + 504, + 191 + ], + "blocks": [ + { + "bbox": [ + 408, + 133, + 504, + 191 + ], + "lines": [ + { + "bbox": [ + 408, + 133, + 504, + 191 + ], + "spans": [ + { + "bbox": [ + 408, + 133, + 504, + 191 + ], + "type": "image", + "image_path": "8231a37bff3ef88a0325ccf3f6a1b4b95d274dafc5eb3d757e40e4b6c231b0ce.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 94, + 201, + 198, + 260 + ], + "blocks": [ + { + "bbox": [ + 76, + 241, + 86, + 277 + ], + "lines": [ + { + "bbox": [ + 76, + 241, + 86, + 277 + ], + "spans": [ + { + "bbox": [ + 76, + 241, + 86, + 277 + ], + "type": "text", + "content": "Cannnnnne" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 201, + 198, + 260 + ], + "lines": [ + { + "bbox": [ + 94, + 201, + 198, + 260 + ], + "spans": [ + { + "bbox": [ + 94, + 201, + 198, + 260 + ], + "type": "image", + "image_path": "6c024c3653f48e58d274cf3b0ac10a7a9379f68ea2e6a6d230dc374460684148.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 94, + 261, + 198, + 320 + ], + "blocks": [ + { + "bbox": [ + 94, + 261, + 198, + 320 + ], + "lines": [ + { + "bbox": [ + 94, + 261, + 198, + 320 + ], + "spans": [ + { + "bbox": [ + 94, + 261, + 198, + 320 + ], + "type": "image", + "image_path": "1abe2ee2679dfb1498037657903327519bf9650027f24c1a9cb55206d929163a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 199, + 201, + 303, + 260 + ], + "blocks": [ + { + "bbox": [ + 199, + 201, + 303, + 260 + ], + "lines": [ + { + "bbox": [ + 199, + 201, + 303, + 260 + ], + "spans": [ + { + "bbox": [ + 199, + 201, + 303, + 260 + ], + "type": "image", + "image_path": "5cd928f9cade3fd1e3171de5d5bf46cae30074121ff18584b58b1b158a52b866.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 203, + 353, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 203, + 353, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 353, + 213 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 200, + 262, + 302, + 320 + ], + "blocks": [ + { + "bbox": [ + 200, + 262, + 302, + 320 + ], + "lines": [ + { + "bbox": [ + 200, + 262, + 302, + 320 + ], + "spans": [ + { + "bbox": [ + 200, + 262, + 302, + 320 + ], + "type": "image", + "image_path": "3c0bbd1200c6dc6d3afee36980b1ba8ce95d69c65fa3c06fb5e62228104bd61d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 304, + 222, + 399, + 259 + ], + "blocks": [ + { + "bbox": [ + 304, + 222, + 399, + 259 + ], + "lines": [ + { + "bbox": [ + 304, + 222, + 399, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 399, + 259 + ], + "type": "image", + "image_path": "7a7f123a7fee0fda041c96f77c194f5032171c3c4bf21a2c9c38b0e2cec97324.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 304, + 261, + 408, + 320 + ], + "blocks": [ + { + "bbox": [ + 304, + 261, + 408, + 320 + ], + "lines": [ + { + "bbox": [ + 304, + 261, + 408, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 408, + 320 + ], + "type": "image", + "image_path": "3bc7db2cd20d4dc8c16db483453989dc1bf817053d1535a52d251b57f2638283.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 409, + 201, + 513, + 260 + ], + "blocks": [ + { + "bbox": [ + 409, + 201, + 513, + 260 + ], + "lines": [ + { + "bbox": [ + 409, + 201, + 513, + 260 + ], + "spans": [ + { + "bbox": [ + 409, + 201, + 513, + 260 + ], + "type": "image", + "image_path": "2b9965965d87343fdda67b2b6a9b44d07bba9f1b060c0c49539939cea52f9df2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 429, + 271, + 500, + 316 + ], + "blocks": [ + { + "bbox": [ + 411, + 262, + 446, + 271 + ], + "lines": [ + { + "bbox": [ + 411, + 262, + 446, + 271 + ], + "spans": [ + { + "bbox": [ + 411, + 262, + 446, + 271 + ], + "type": "text", + "content": "3D View" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 271, + 500, + 316 + ], + "lines": [ + { + "bbox": [ + 429, + 271, + 500, + 316 + ], + "spans": [ + { + "bbox": [ + 429, + 271, + 500, + 316 + ], + "type": "image", + "image_path": "f01dbe634590424c56d0b017ca24c3b8d46bf4bff448ef04216495e3e0815fbe.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 94, + 328, + 198, + 388 + ], + "blocks": [ + { + "bbox": [ + 77, + 369, + 86, + 405 + ], + "lines": [ + { + "bbox": [ + 77, + 369, + 86, + 405 + ], + "spans": [ + { + "bbox": [ + 77, + 369, + 86, + 405 + ], + "type": "text", + "content": "Cannnnnnae" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 328, + 198, + 388 + ], + "lines": [ + { + "bbox": [ + 94, + 328, + 198, + 388 + ], + "spans": [ + { + "bbox": [ + 94, + 328, + 198, + 388 + ], + "type": "image", + "image_path": "8c4085b37ac6d4e93956bcd6d0b91de2f3692636e35ce9229a80e857bea9ba43.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 94, + 388, + 198, + 448 + ], + "blocks": [ + { + "bbox": [ + 94, + 388, + 198, + 448 + ], + "lines": [ + { + "bbox": [ + 94, + 388, + 198, + 448 + ], + "spans": [ + { + "bbox": [ + 94, + 388, + 198, + 448 + ], + "type": "image", + "image_path": "3ca6651d5e30208a2eb1bdbef8da45efa4e957121837f2bdd9a522decaef839a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 451, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 545, + 506 + ], + "type": "text", + "content": "Figure 3. Automatically generated 2D and 3D Clip-Art to supervise our network: Unoccluded objects are first mined using time-lapse imagery of WALT dataset [62]. Non-intersecting unoccluded objects are composited back into the background image in their respective original positions to preserve correct appearances. The resulting clip-art images, along with their corresponding amodal pseudo-groundtruth information, such as segmentation, keypoints, depth/normal maps, and 3D shapes, are shown. Our method generates realistic appearances from any stationary camera, incorporating diverse viewing geometries, weather conditions, lighting, and occlusion configurations." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 199, + 328, + 303, + 388 + ], + "blocks": [ + { + "bbox": [ + 199, + 328, + 303, + 388 + ], + "lines": [ + { + "bbox": [ + 199, + 328, + 303, + 388 + ], + "spans": [ + { + "bbox": [ + 199, + 328, + 303, + 388 + ], + "type": "image", + "image_path": "f6970d72d18ebf3a91156e6a9d5f134e178610db9a7663e2161dfa8a8e2940b7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 331, + 353, + 341 + ], + "lines": [ + { + "bbox": [ + 307, + 331, + 353, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 353, + 341 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 200, + 389, + 302, + 448 + ], + "blocks": [ + { + "bbox": [ + 200, + 389, + 302, + 448 + ], + "lines": [ + { + "bbox": [ + 200, + 389, + 302, + 448 + ], + "spans": [ + { + "bbox": [ + 200, + 389, + 302, + 448 + ], + "type": "image", + "image_path": "240e09b24c6b33e3d3ac8931ab281d38a1a419f09cea53dbda947d5ea7a8efe5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 304, + 346, + 407, + 388 + ], + "blocks": [ + { + "bbox": [ + 304, + 346, + 407, + 388 + ], + "lines": [ + { + "bbox": [ + 304, + 346, + 407, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 407, + 388 + ], + "type": "image", + "image_path": "a41f1f22920137e3d9a29bebfe76bebe6f456f0e277cabbb40d89837f00b518d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 304, + 389, + 407, + 448 + ], + "blocks": [ + { + "bbox": [ + 304, + 389, + 407, + 448 + ], + "lines": [ + { + "bbox": [ + 304, + 389, + 407, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 407, + 448 + ], + "type": "image", + "image_path": "8e372824fa0e3cbde0b1a31516205794f9b4e709e11c089788c544f6deed3730.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 409, + 328, + 513, + 388 + ], + "blocks": [ + { + "bbox": [ + 409, + 328, + 513, + 388 + ], + "lines": [ + { + "bbox": [ + 409, + 328, + 513, + 388 + ], + "spans": [ + { + "bbox": [ + 409, + 328, + 513, + 388 + ], + "type": "image", + "image_path": "42d96b3aa2cffbd6141e140573d10e01bc3b2ea61376f68a40adc6835b085af0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 408, + 395, + 512, + 444 + ], + "blocks": [ + { + "bbox": [ + 410, + 391, + 445, + 399 + ], + "lines": [ + { + "bbox": [ + 410, + 391, + 445, + 399 + ], + "spans": [ + { + "bbox": [ + 410, + 391, + 445, + 399 + ], + "type": "text", + "content": "3D View" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 395, + 512, + 444 + ], + "lines": [ + { + "bbox": [ + 408, + 395, + 512, + 444 + ], + "spans": [ + { + "bbox": [ + 408, + 395, + 512, + 444 + ], + "type": "image", + "image_path": "6a749abfe66bcca2aefa611bbc548825ceb3fe862c6f0c845c0d98bd5ce505ff.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 516, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 636 + ], + "type": "text", + "content": "approach distinguishes itself from the mere compositing of 2D images such as WALT [62], which can result in unrealistic composition (see the comparison in Fig. 4). Our data with physically accurate and realistic occlusion configurations contributes to more efficient and scalable training, as demonstrated in our experimental results. Each such generated clip-art image is accompanied by amodal segmentation masks, keypoints, 3D poses and shapes, provided by off-the-shelf methods that serve as pseudo-groundtruth supervision to learn object reconstruction under occlusion." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 645, + 277, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 277, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 277, + 659 + ], + "type": "text", + "content": "4. Learning Reconstruction under Occlusion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "We have generated an extensive clip-art image dataset with corresponding 2D/3D pseudo-groundtruth representations (see Fig. 3). Using this supervision signal, we train a model capable of inferring holistic object representations in the" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 517, + 429, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 429, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 429, + 528 + ], + "type": "text", + "content": "presence of severe occlusions." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": "Base Network and 2D Amodal Representations: Using the Swin Transformer [51] backbone with MaskRCNN-based [31] detection heads for its simplicity, our goal is to show improvement irrespective of the base model used. We train the default network with a multi-task loss " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": " for each representation. These losses, computed on each sampled Region of Interest (RoI), include label classification loss " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": ", bounding-box loss " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": ", binary cross-entropy loss on the mask branch " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": " and keypoints " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2\\mathrm{D}} = \\mathcal{L}_{\\mathrm{cls}} + \\mathcal{L}_{\\mathrm{box}} + \\mathcal{L}_{\\mathrm{mask}} + \\mathcal{L}_{\\mathrm{kp}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": ". Crucially, with our clip-art dataset including amodal 2D bounding box, segmentation, and keypoints, our method effectively learns amodal instance segmentation and keypoint detection." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "3D Object Reconstruction: To regress object 3D pose and shape under occlusion, we extend the base network with a" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9517" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 76, + 289, + 243 + ], + "blocks": [ + { + "bbox": [ + 84, + 76, + 289, + 243 + ], + "lines": [ + { + "bbox": [ + 84, + 76, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 84, + 76, + 289, + 243 + ], + "type": "image", + "image_path": "42d251e99c5694a20e861c16969fac73391067739ebf39a996517643465cb399.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 247, + 222, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 247, + 222, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 247, + 222, + 258 + ], + "type": "text", + "content": "Composited using WALT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 76, + 506, + 243 + ], + "blocks": [ + { + "bbox": [ + 301, + 76, + 506, + 243 + ], + "lines": [ + { + "bbox": [ + 301, + 76, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 301, + 76, + 506, + 243 + ], + "type": "image", + "image_path": "60fa54432d89f3e4683718d36e268c041fbe69429b9d8864391310c8b711d4b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 247, + 452, + 258 + ], + "lines": [ + { + "bbox": [ + 348, + 247, + 452, + 258 + ], + "spans": [ + { + "bbox": [ + 348, + 247, + 452, + 258 + ], + "type": "text", + "content": "Composed using our method" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 261, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 294 + ], + "type": "text", + "content": "Figure 4. Comparison between images composited using the 2D-based method WALT2D [62] (left) and our 3D-based method WALT3D (right). It is evident that our 3D-based compositing method generates realistic and geometrically accurate occlusion configurations, in contrast to the 2D-based method (e.g., cars and people overlapping in an unfeasible way)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": "3D regression branch. Each vehicle " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " semantic 3D keypoints " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " are represented by a linear combination of the mean shape " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Q}}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " principal components " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{1\\dots K}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " is the shape coefficients: " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\bar{\\mathbf{Q}} +\\sum_{k = 1}^{K}\\alpha_{k}\\mathbf{Q}_{k}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ". Thus, we want to regress for its 6-DoF pose " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{R},\\mathbf{t})\\in \\mathbf{SE}(3)\\}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " and shape coefficients " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ". Given the predicted 2D keypoints and the corresponding 3D keypoints " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ", we can obtain its 6-DoF pose through a differentiable PnP layer [9]. Defining " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " as the projection function with known intrinsic " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ", we can optimize for the keypoints reprojection loss " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp2D}}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ", shape loss " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shape}}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": ", and pose loss " + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pose}}" + }, + { + "bbox": [ + 47, + 304, + 289, + 437 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 443, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 443, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 79, + 443, + 287, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k p 2 D}} = \\sum_ {i = 1} ^ {N} \\left| \\left| \\pi \\left(\\mathbf {R} \\mathbf {X} _ {i} + \\mathbf {t}\\right) - \\mathbf {x} _ {i} ^ {*} \\right| \\right| _ {2} ^ {2}, \\tag {1}", + "image_path": "76b653976db1ccbb34332b0d30afa2097ea75e7f7c4c8f99a3774e11b306d825.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 477, + 271, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 477, + 271, + 493 + ], + "spans": [ + { + "bbox": [ + 52, + 477, + 271, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s h a p e}} = | | \\alpha_ {k} - \\alpha_ {k} ^ {*} | | _ {2} ^ {2}, \\mathcal {L} _ {\\mathrm {p o s e}} = | | [ \\mathbf {R}, \\mathbf {t} ] - [ \\mathbf {R} ^ {*}, \\mathbf {t} ^ {*} ] | |", + "image_path": "650536e1e4df717de4b677f7232d47abfa8b89d7cc9dcbd287f85fd159011f9b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 498, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 568 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 498, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^*,\\alpha_k^*,\\mathbf{R}^*,\\mathbf{t}^*" + }, + { + "bbox": [ + 47, + 498, + 287, + 568 + ], + "type": "text", + "content": " represents the (pseudo)-groundtruth 2D keypoints, shape coefficients, rotation, and translation respectively. The final loss term is given as the weighted sum of the losses, and we learn amodal bounding box, segmentation, keypoint locations, 3D shape and pose in an end-to-end differentiable manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": "For human reconstruction, we train our HMR predictor branch like HMR 2.0 [26]. Supervised by pseudogroundtruth SMPL pose " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": " and shape " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\beta^{*}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": " parameters [52], our network is guided by the loss functions on SMPL parameters " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smp1}}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": ", 3D keypoints " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp3D}}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": ", 2D keypoints " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp2D}}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": " and an adversarial loss " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": " that ensures the model predicts valid 3D poses through the discriminator " + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 47, + 570, + 287, + 654 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 659, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {s m p l}} = | | [ \\theta , \\beta ] - [ \\theta^ {*}, \\beta^ {*} ] | | _ {2} ^ {2}, \\\\ \\mathcal {L} _ {\\mathrm {k p} 3 \\mathrm {D}} = | | \\mathbf {X} - \\mathbf {X} ^ {*} | | _ {1}, \\mathcal {L} _ {\\mathrm {k p} 2 \\mathrm {D}} = | | \\pi (\\mathbf {X}) - \\mathbf {x} ^ {*} | | _ {1}, \\tag {2} \\\\ \\mathcal {L} _ {\\mathrm {a d v}} = \\sum_ {k} (D _ {k} (\\theta_ {b}, \\beta) - 1) ^ {2} \\\\ \\end{array}", + "image_path": "b776174911b801b372c9195b1ca454abca5aafa796b4d912829641e01871e304.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^*" + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^*" + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "content": " denote the (pseudo)-groundtruth 2D and 3D keypoints respectively. With a weighted combination of these losses, we optimize for human 3D shape and pose in an end-to-end fashion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 366, + 505, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 505, + 380 + ], + "type": "text", + "content": "5. Dataset and Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 387, + 546, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 546, + 459 + ], + "type": "text", + "content": "Raw Time-lapse Dataset [62]: This dataset contains images from 20 stationary cameras in urban scenes captured over multiple years. The images are either 4K or HD and are captured at 60fps in short bursts. We used 30 days of data from 10 cameras resulting in approximately 3.3 million car and people instances for our experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 460, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 546, + 568 + ], + "type": "text", + "content": "WALT2D [62]: From the raw time-lapse dataset, we use the 2D-based compositing method proposed by WALT [62] to generate supervision data by pasting unoccluded objects back into the scene with different backgrounds resulting in 10000 training and 500 testing images per camera. Fig. 4 (left) illustrates an example of this dataset. It is important to note that this compositing method, relying solely on 2D information, leads to intersecting objects in the real world that may not faithfully represent the physical configuration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 569, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 569, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 569, + 545, + 700 + ], + "type": "text", + "content": "WALT3D (Ours): Also from the raw time-lapse dataset, our approach involves mining unoccluded objects and reinsering them into the scene, resulting in a comparable number of object instances as 2D-based compositing in WALT2D [62]. However, our approach employs a 3D-based compositing method to generate supervision data, as outlined in Section 3, in contrast to relying solely on 2D-based composition. Fig. 4 (right) shows an example clip-art image composited using our method, with our approach yielding training data with greater realism and physical accuracy compared to WALT2D." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Vehicle 2D Keypoints Dataset (Ours): To determine 6-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9518" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 272, + 201 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 272, + 201 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 272, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 272, + 201 + ], + "type": "image", + "image_path": "f68f3d22999f7fbdb3f9bcee3ba4191e3c4b073c162f6c0e47ff466845502f76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "type": "text", + "content": "Figure 5. Sample images from our new vehicle 2D keypoints dataset. The dataset contains a wide range of appearance variations including day and night and various traffic scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 399 + ], + "type": "text", + "content": "DoF vehicle pose during compositing, we require high-quality 2D vehicle keypoints. Existing datasets [47, 59, 67, 77] offer human-annotated 2D vehicle keypoints but mainly focus on driving scenes or have limited training examples, lacking necessary appearance diversity for novel viewpoints like our stationary traffic cameras. Thus, we propose a new dataset with 7,018 images and 42,547 annotated instances from diverse viewpoints (see Fig. 5), each keypoint containing 2D location and occlusion status. With our dataset, we observe significant improvement in keypoint localization accuracy (66.41% to 80.12% on PCK@0.1), and consequently, an improvement in pseudo-groundtruth data quality. Further details are in the Supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 409, + 244, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 244, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 244, + 422 + ], + "type": "text", + "content": "6. Baselines, Metrics, and Evaluations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 429, + 113, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 113, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 113, + 440 + ], + "type": "text", + "content": "6.1. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 518 + ], + "type": "text", + "content": "Detection and Instance Segmentation: All baselines share the same network architecture, employing a Swin [51] backbone as detailed in Section 4. The SWIN baseline is initially trained on the COCO dataset [50]. Subsequently, we perform further finetuning on the WALT2D and WALT3D (Ours) datasets, as detailed in Section 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "type": "text", + "content": "Vehicle 3D Reconstruction: We compare our approach with Occ-Net [62] and 3DRCNN [45] by changing their feature extraction backbone to be as close as possible to ours for fair comparison. They are trained on a combination of PASCAL3D+ [77], KITTI-3D [47], Carfusion [59], and ApolloCar3D [67], where we further finetune our baselines on the WALT2D and our WALT3D dataset, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 639 + ], + "type": "text", + "content": "Human 3D Reconstruction: Our architecture mirrors HMR 2.0 [26] as described in Section 4. We initialize from the pretrained model, then finetune on our WALT3D data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 160, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 160, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 160, + 658 + ], + "type": "text", + "content": "6.2. Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "2D Metrics: We report Average Precision (AP) on bounding box detection " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{box}})" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and instance segmentation " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{mask}})" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " following [50]. For keypoints, we use Percentage of Correct Keypoints (PCK) metric where a keypoint" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 425, + 156 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 425, + 156 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 425, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 425, + 156 + ], + "type": "image", + "image_path": "3e92bac6db158d80d0018b980a556c89a147dbbd8ced869af683550db799e3df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 69, + 542, + 156 + ], + "blocks": [ + { + "bbox": [ + 425, + 69, + 542, + 156 + ], + "lines": [ + { + "bbox": [ + 425, + 69, + 542, + 156 + ], + "spans": [ + { + "bbox": [ + 425, + 69, + 542, + 156 + ], + "type": "image", + "image_path": "36e7458477813ad70a527e4d524f9e5ea4b478bf73def2efd2a3f09efdae5321.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 161, + 425, + 247 + ], + "blocks": [ + { + "bbox": [ + 308, + 161, + 425, + 247 + ], + "lines": [ + { + "bbox": [ + 308, + 161, + 425, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 425, + 247 + ], + "type": "image", + "image_path": "06891abaceb83f32acbb20cc538638e61f4dbd78dfea43b1822c61c13213c2e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 249, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 304 + ], + "type": "text", + "content": "Figure 6. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal vehicle detection, segmentation, 2D and 3D keypoint estimation. Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 161, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 425, + 161, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 425, + 161, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 425, + 161, + 541, + 247 + ], + "type": "image", + "image_path": "43523d45f41f7fbacf664b7ab73b87d60582bc3d5a9ada1688f924422af87126.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 308, + 426, + 396 + ], + "blocks": [ + { + "bbox": [ + 309, + 308, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 309, + 308, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 426, + 396 + ], + "type": "image", + "image_path": "1e4910d1c02e2ab20730c2066f4d9643a6edad906756cbde257fa7aa55e67d81.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 426, + 308, + 542, + 396 + ], + "blocks": [ + { + "bbox": [ + 426, + 308, + 542, + 396 + ], + "lines": [ + { + "bbox": [ + 426, + 308, + 542, + 396 + ], + "spans": [ + { + "bbox": [ + 426, + 308, + 542, + 396 + ], + "type": "image", + "image_path": "e631aac7db9bba79c36bd3bfed551a5f36af5327c9e6228d41994b54fa71c7a6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 399, + 425, + 485 + ], + "blocks": [ + { + "bbox": [ + 308, + 399, + 425, + 485 + ], + "lines": [ + { + "bbox": [ + 308, + 399, + 425, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 425, + 485 + ], + "type": "image", + "image_path": "a362ce4b63535c3da1816bb1be0c3da5bd513f1d21590a905c50376a6ae1b04b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 487, + 545, + 542 + ], + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 542 + ], + "type": "text", + "content": "Figure 7. We show the accuracy of our method with respect to increasing percentage of occlusion on multiple tasks like amodal human detection, segmentation, and human mesh recovery (HMR). Observe that our method consistently performs better than other baselines showing robustness to increasing occlusion percentage." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 425, + 399, + 541, + 485 + ], + "blocks": [ + { + "bbox": [ + 425, + 399, + 541, + 485 + ], + "lines": [ + { + "bbox": [ + 425, + 399, + 541, + 485 + ], + "spans": [ + { + "bbox": [ + 425, + 399, + 541, + 485 + ], + "type": "image", + "image_path": "c302e39b214b053277cddb179afeb84a9be8eace253ad9f4fde636c36e9b5b75.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "type": "text", + "content": "is considered correct if it lies within the radius " + }, + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "type": "inline_equation", + "content": "0 < \\alpha < 1" + }, + { + "bbox": [ + 304, + 550, + 545, + 574 + ], + "type": "text", + "content": ") of the ground-truth keypoint." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 575, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 624 + ], + "type": "text", + "content": "3D Metrics: We use the standard metrics of previous work [34], reporting the Mean Per Joint Position Error (MPJME) between predicted and ground-truth 3D keypoints (aligned using the root joint) as well as 3D PCK." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "type": "text", + "content": "6.3. Ablation Analysis" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Robustness to Occlusions: We evaluate the effectiveness of our algorithm with varying occlusion levels. Similar to WALT [62], we use the pseudo-groundtruth segmentation masks from the evaluation set to group objects based on occlusion percentage. In 2D tasks such as detection and" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9519" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 272, + 223 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 272, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 272, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 272, + 223 + ], + "type": "image", + "image_path": "11bc1846cbde4d155582ab20d4e781c6b9fd4bbcc9e6237881d0bc22b8d0922f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 287, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 269 + ], + "type": "text", + "content": "Figure 8. Comparison between WALT2D [62]) and our WALT3D approach. Since our method produces higher quality training data with more realistic occlusion configuration, our approach is especially useful in low-data regime." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 279, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 458 + ], + "type": "text", + "content": "segmentation, the model trained with our WALT3D data significantly outperforms the baseline SWIN model trained on COCO [50], which only includes modal masks. This improvement is particularly evident in high occlusion percentages for both vehicle (Fig. 6) and human (Fig. 7). Compared to WALT2D [62], our data demonstrates slightly better accuracy across all occlusion levels, thanks to our 3D-based compositing method WALT3D producing physically accurate occlusion configurations with the same amount of training data. For 3D tasks such as vehicle (Fig. 6) and human (Fig. 7) 3D pose estimation, using our WALT3D data consistently outperforms other baselines, particularly in high occlusion percentages. Thus, models trained with our WALT3D data, even with imperfect pseudo-groundtruth from off-the-shelf methods, are robust to occlusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": "3D-based Compositing helps Data Efficiency: Fig. 8 shows that our WALT3D data significantly improves performance across all training data fractions, including at low data regime (at " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " of training data with " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "+13.9\\mathrm{AP}^{\\mathrm{mask}}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " improvement compared to training with WALT2D data). Essentially, a model trained on " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " of WALT3D data achieves comparable AP to one trained on " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " of WALT2D data. This improvement stems from our WALT3D approach generating higher quality and physically accurate occlusion configurations with the same training data volume, aligning more closely with the real-world distribution as in Fig. 4. Importantly, this data efficiency enables scaling to a larger number of scenes more efficiently." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Cross-evaluations on other datasets: Table 1 presents quantitative results on COCO-Amodal [84] (COCO-A) (natural images) and KINS [58] (driving images), each containing human-annotated ground-truth 2D amodal instance masks. It shows that additional fine-tuning with our WALT3D data improves results in both KINS and COCO-A, highlighting the cross-evaluation performance of our data, with potential for even better generalization with the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 425, + 223 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 425, + 223 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 425, + 223 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 425, + 223 + ], + "type": "image", + "image_path": "5c84d1ce64a46986ad9bb44fc1ce2a2aa3a8e491d20513f9bc3714d0a2065f4a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 225, + 413, + 235 + ], + "lines": [ + { + "bbox": [ + 350, + 225, + 413, + 235 + ], + "spans": [ + { + "bbox": [ + 350, + 225, + 413, + 235 + ], + "type": "text", + "content": "Previous SOTA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 437, + 70, + 490, + 172 + ], + "blocks": [ + { + "bbox": [ + 437, + 70, + 490, + 172 + ], + "lines": [ + { + "bbox": [ + 437, + 70, + 490, + 172 + ], + "spans": [ + { + "bbox": [ + 437, + 70, + 490, + 172 + ], + "type": "image", + "image_path": "7b4695871f6f385107405b370d782682899a6788e61c1d433fd3de35e56e7dbf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 172, + 489, + 223 + ], + "blocks": [ + { + "bbox": [ + 437, + 172, + 489, + 223 + ], + "lines": [ + { + "bbox": [ + 437, + 172, + 489, + 223 + ], + "spans": [ + { + "bbox": [ + 437, + 172, + 489, + 223 + ], + "type": "image", + "image_path": "1c89720c2764fe66dfb94bdaf2a708c453d99f8424d6650c5732393ce6a6f56b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "type": "text", + "content": "Figure 9. Qualitative results showing that our data improves amodal segmentation, keypoint prediction and 3D reconstruction compared to previous SOTA." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 490, + 70, + 541, + 172 + ], + "blocks": [ + { + "bbox": [ + 490, + 70, + 541, + 172 + ], + "lines": [ + { + "bbox": [ + 490, + 70, + 541, + 172 + ], + "spans": [ + { + "bbox": [ + 490, + 70, + 541, + 172 + ], + "type": "image", + "image_path": "48c14d578cf8539c8da2c81d4552ff38a63a996db8f4ed73a7c87d6160c38bd5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 490, + 172, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 490, + 172, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 490, + 172, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 490, + 172, + 541, + 223 + ], + "type": "image", + "image_path": "233b070b6eb184501aca11923d44eb232dcd11fa446b7786071b516b4e97eca0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 225, + 503, + 235 + ], + "lines": [ + { + "bbox": [ + 481, + 225, + 503, + 235 + ], + "spans": [ + { + "bbox": [ + 481, + 225, + 503, + 235 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 306, + 295, + 545, + 372 + ], + "blocks": [ + { + "bbox": [ + 306, + 274, + 512, + 286 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 512, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 512, + 286 + ], + "type": "text", + "content": "increasing amount of automatically generated data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 295, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 306, + 295, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 295, + 545, + 372 + ], + "type": "table", + "html": "
Testing Data Training DataBox APMask AP
KINSCOCO-AKINSCOCO-A
COCO27.838.020.831.2
COCO + WALT2D28.240.422.436.4
COCO + WALT3D28.843.223.538.7
COCO + COCO-A + KINS36.846.932.942.7
COCO + COCO-A + KINS + WALT2D37.447.933.343.5
COCO + COCO-A + KINS + WALT3D38.848.735.645.3
", + "image_path": "b2a65d1aea2c6c7a75075b859b1048e3b00120b7b4fc3a02eeb9f0a4096d39b5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 376, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 420 + ], + "type": "text", + "content": "Table 1. Finetuning with our WALT3D data improves generalization across KINS and COCO-A (vehicle & people). Additionally, combining our data with domain-specific data like KINS and COCO-A further enhances performance in their domains." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 422, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 541 + ], + "type": "text", + "content": "Cross-evaluations on pedestrian tracking: We evaluate how our WALT3D data helps pedestrian tracking in Table 2. To highlight how better detections lead to better tracking, we use a simple tracker SORT [4] with two detectors: one pretrained on COCO and one finetuned with our WALT3D data. On MOT17-train, the model finetuned with WALT3D data produces high-quality detections under strong occlusions, resulting in improvements over the baseline in all metrics, particularly those favoring better detection quality like MOTA and DetA (see metrics details in [53])." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 317, + 544, + 533, + 585 + ], + "blocks": [ + { + "bbox": [ + 317, + 544, + 533, + 585 + ], + "lines": [ + { + "bbox": [ + 317, + 544, + 533, + 585 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 533, + 585 + ], + "type": "table", + "html": "
Detector Train DataMOTA↑DetA↑IDF1↑HOTA↑
COCO42.240.346.740.1
COCO + WALT3D47.243.249.542.0
", + "image_path": "f455f0e2eaba1e3537e39ddcd3decf2ea88ea85e8e4c1885422f5d8dab828710.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 587, + 545, + 610 + ], + "lines": [ + { + "bbox": [ + 305, + 587, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 545, + 610 + ], + "type": "text", + "content": "Table 2. MOT17-train results using SORT with two detectors: one pretrained on COCO and one fine-tuned with our WALT3D data." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 615, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 663 + ], + "type": "text", + "content": "Qualitative Results: Results of our method can be seen in Fig. 9 and Fig. 10. Comparing to previous methods, observe that our method can produce robust 2D/3D predictions, even under challenging high-occlusion scenarios." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 673, + 457, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 457, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 457, + 685 + ], + "type": "text", + "content": "7. Discussion and Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our data generation framework is method-agnostic, accommodating various robust object pose estimation alterna" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9520" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 153, + 151 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 153, + 151 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 153, + 151 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 153, + 151 + ], + "type": "image", + "image_path": "ccbf88efd1bb580fa878e4ae6e0b5c316184ee4d5a613fc3d4e6346df7781084.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 153, + 71, + 247, + 150 + ], + "blocks": [ + { + "bbox": [ + 153, + 71, + 247, + 150 + ], + "lines": [ + { + "bbox": [ + 153, + 71, + 247, + 150 + ], + "spans": [ + { + "bbox": [ + 153, + 71, + 247, + 150 + ], + "type": "image", + "image_path": "51ad2b5dcc6bed33f649f678b501cb62c5f053f008395fb8e3ec9eb651f72abe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 248, + 71, + 343, + 150 + ], + "blocks": [ + { + "bbox": [ + 248, + 71, + 343, + 150 + ], + "lines": [ + { + "bbox": [ + 248, + 71, + 343, + 150 + ], + "spans": [ + { + "bbox": [ + 248, + 71, + 343, + 150 + ], + "type": "image", + "image_path": "2751cadcf00ee45e6c78370eef789f1576401f2b517e912b8caa89e39ffe3f0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 344, + 72, + 439, + 150 + ], + "blocks": [ + { + "bbox": [ + 344, + 72, + 439, + 150 + ], + "lines": [ + { + "bbox": [ + 344, + 72, + 439, + 150 + ], + "spans": [ + { + "bbox": [ + 344, + 72, + 439, + 150 + ], + "type": "image", + "image_path": "452d2486b5d256d025f88c7730d8f836df23d730c5be0d9807e654129184494c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 534, + 150 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 534, + 150 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 534, + 150 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 534, + 150 + ], + "type": "image", + "image_path": "ecb1b346aa61a46ea543f2a456906e80ff3108edfa464a80b61bf4d59f6a5a5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 152, + 153, + 231 + ], + "blocks": [ + { + "bbox": [ + 57, + 152, + 153, + 231 + ], + "lines": [ + { + "bbox": [ + 57, + 152, + 153, + 231 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 153, + 231 + ], + "type": "image", + "image_path": "a25042638fd834a5e91dab330423a7b9efbcfa678c956b1bb1f4868e00ba3bc2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 153, + 153, + 247, + 231 + ], + "blocks": [ + { + "bbox": [ + 153, + 153, + 247, + 231 + ], + "lines": [ + { + "bbox": [ + 153, + 153, + 247, + 231 + ], + "spans": [ + { + "bbox": [ + 153, + 153, + 247, + 231 + ], + "type": "image", + "image_path": "543584f73f8c002f24b62ad3c012bdadb89a4041eac4b19f608686e037032b9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 248, + 153, + 343, + 231 + ], + "blocks": [ + { + "bbox": [ + 248, + 153, + 343, + 231 + ], + "lines": [ + { + "bbox": [ + 248, + 153, + 343, + 231 + ], + "spans": [ + { + "bbox": [ + 248, + 153, + 343, + 231 + ], + "type": "image", + "image_path": "2f6e3100e3c1ce33eeba61d0470fe127f1ab2e82a59e0d2f95d4a4531c4cdb97.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 344, + 153, + 439, + 231 + ], + "blocks": [ + { + "bbox": [ + 344, + 153, + 439, + 231 + ], + "lines": [ + { + "bbox": [ + 344, + 153, + 439, + 231 + ], + "spans": [ + { + "bbox": [ + 344, + 153, + 439, + 231 + ], + "type": "image", + "image_path": "235807ed81f953666ecaeac0082d3ea38de823bb91d5b8b60392e8714036552e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 153, + 534, + 231 + ], + "blocks": [ + { + "bbox": [ + 440, + 153, + 534, + 231 + ], + "lines": [ + { + "bbox": [ + 440, + 153, + 534, + 231 + ], + "spans": [ + { + "bbox": [ + 440, + 153, + 534, + 231 + ], + "type": "image", + "image_path": "5399533297c714632f8830e8efce5c5f417d8092d5fb0f6f81463deb023d2a03.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 234, + 152, + 306 + ], + "blocks": [ + { + "bbox": [ + 58, + 234, + 152, + 306 + ], + "lines": [ + { + "bbox": [ + 58, + 234, + 152, + 306 + ], + "spans": [ + { + "bbox": [ + 58, + 234, + 152, + 306 + ], + "type": "image", + "image_path": "cb4b8542f64e925969b965f299c478e2c14ea405820cfc8e25505b3289148d35.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 153, + 234, + 247, + 306 + ], + "blocks": [ + { + "bbox": [ + 153, + 234, + 247, + 306 + ], + "lines": [ + { + "bbox": [ + 153, + 234, + 247, + 306 + ], + "spans": [ + { + "bbox": [ + 153, + 234, + 247, + 306 + ], + "type": "image", + "image_path": "0e1a6bcc678c48932502e78b9b1851e67cbe98f7a37cf4bc777dbe745be99734.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 249, + 234, + 343, + 306 + ], + "blocks": [ + { + "bbox": [ + 249, + 234, + 343, + 306 + ], + "lines": [ + { + "bbox": [ + 249, + 234, + 343, + 306 + ], + "spans": [ + { + "bbox": [ + 249, + 234, + 343, + 306 + ], + "type": "image", + "image_path": "a6dce8c450b37d38d0b352f38bddd37ba81b0e5b98a40784f10aa89209521e6e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 344, + 234, + 439, + 306 + ], + "blocks": [ + { + "bbox": [ + 344, + 234, + 439, + 306 + ], + "lines": [ + { + "bbox": [ + 344, + 234, + 439, + 306 + ], + "spans": [ + { + "bbox": [ + 344, + 234, + 439, + 306 + ], + "type": "image", + "image_path": "6524742e917981b262cdc6c17c523d73466fe04d04a80c6149bc074b05c28126.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 440, + 234, + 534, + 306 + ], + "blocks": [ + { + "bbox": [ + 440, + 234, + 534, + 306 + ], + "lines": [ + { + "bbox": [ + 440, + 234, + 534, + 306 + ], + "spans": [ + { + "bbox": [ + 440, + 234, + 534, + 306 + ], + "type": "image", + "image_path": "8f9f2a98ee5fedec2f926dbec7bb8a148aa45e868b37213aeca9c69127f64d05.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 309, + 152, + 400 + ], + "blocks": [ + { + "bbox": [ + 58, + 309, + 152, + 400 + ], + "lines": [ + { + "bbox": [ + 58, + 309, + 152, + 400 + ], + "spans": [ + { + "bbox": [ + 58, + 309, + 152, + 400 + ], + "type": "image", + "image_path": "3bc3618d5e4807490fda3ff465da208174db30136c1fec39896c05fbecd3f711.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 153, + 309, + 247, + 400 + ], + "blocks": [ + { + "bbox": [ + 153, + 309, + 247, + 400 + ], + "lines": [ + { + "bbox": [ + 153, + 309, + 247, + 400 + ], + "spans": [ + { + "bbox": [ + 153, + 309, + 247, + 400 + ], + "type": "image", + "image_path": "5a89a001ca50e1525be270798caeb74aa75826f4cc067985fea0a3afe01fa724.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 249, + 309, + 343, + 400 + ], + "blocks": [ + { + "bbox": [ + 249, + 309, + 343, + 400 + ], + "lines": [ + { + "bbox": [ + 249, + 309, + 343, + 400 + ], + "spans": [ + { + "bbox": [ + 249, + 309, + 343, + 400 + ], + "type": "image", + "image_path": "d92175515ce611fb0cb3047af03db622c1c19ec6659ac6b755ec280b5e4406ca.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 344, + 309, + 439, + 400 + ], + "blocks": [ + { + "bbox": [ + 344, + 309, + 439, + 400 + ], + "lines": [ + { + "bbox": [ + 344, + 309, + 439, + 400 + ], + "spans": [ + { + "bbox": [ + 344, + 309, + 439, + 400 + ], + "type": "image", + "image_path": "ed8424bb2ff5ea25edd655c2f3f038386c6af56830d7f419cf98f968387582ed.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 440, + 309, + 534, + 400 + ], + "blocks": [ + { + "bbox": [ + 440, + 309, + 534, + 400 + ], + "lines": [ + { + "bbox": [ + 440, + 309, + 534, + 400 + ], + "spans": [ + { + "bbox": [ + 440, + 309, + 534, + 400 + ], + "type": "image", + "image_path": "5d7c6ce62630e89dec5784ff8e2ea92bda020383dae2c2b8e1b66e2ac6eb93a6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 58, + 403, + 153, + 505 + ], + "blocks": [ + { + "bbox": [ + 58, + 403, + 153, + 505 + ], + "lines": [ + { + "bbox": [ + 58, + 403, + 153, + 505 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 153, + 505 + ], + "type": "image", + "image_path": "a9351148fd96970ee0295c863557a24279f0eea6ea48a6f3118af9e4a8273585.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 506, + 130, + 517 + ], + "lines": [ + { + "bbox": [ + 85, + 506, + 130, + 517 + ], + "spans": [ + { + "bbox": [ + 85, + 506, + 130, + 517 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 153, + 403, + 247, + 504 + ], + "blocks": [ + { + "bbox": [ + 153, + 403, + 247, + 504 + ], + "lines": [ + { + "bbox": [ + 153, + 403, + 247, + 504 + ], + "spans": [ + { + "bbox": [ + 153, + 403, + 247, + 504 + ], + "type": "image", + "image_path": "2b8f5970b9821bb857dcf272c919d374721c794de0eba53b2b1169cfa26eddcd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 506, + 242, + 516 + ], + "lines": [ + { + "bbox": [ + 161, + 506, + 242, + 516 + ], + "spans": [ + { + "bbox": [ + 161, + 506, + 242, + 516 + ], + "type": "text", + "content": "Amodal Segmentation" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 248, + 403, + 343, + 505 + ], + "blocks": [ + { + "bbox": [ + 248, + 403, + 343, + 505 + ], + "lines": [ + { + "bbox": [ + 248, + 403, + 343, + 505 + ], + "spans": [ + { + "bbox": [ + 248, + 403, + 343, + 505 + ], + "type": "image", + "image_path": "d3346e2b90ad0b9e4d655e650d22d93f2cf77cec3a62ccfc1e4639a2e2c9f979.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 506, + 345, + 516 + ], + "lines": [ + { + "bbox": [ + 249, + 506, + 345, + 516 + ], + "spans": [ + { + "bbox": [ + 249, + 506, + 345, + 516 + ], + "type": "text", + "content": "Amodal Keypoints/Shapes" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 344, + 412, + 438, + 504 + ], + "blocks": [ + { + "bbox": [ + 344, + 412, + 438, + 504 + ], + "lines": [ + { + "bbox": [ + 344, + 412, + 438, + 504 + ], + "spans": [ + { + "bbox": [ + 344, + 412, + 438, + 504 + ], + "type": "image", + "image_path": "7f19aff6d7c4ef2bedd6f0f4575dc04bb206b314bcddc660ee6b5362d6c1a7db.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 506, + 410, + 516 + ], + "lines": [ + { + "bbox": [ + 369, + 506, + 410, + 516 + ], + "spans": [ + { + "bbox": [ + 369, + 506, + 410, + 516 + ], + "type": "text", + "content": "3D View-1" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 439, + 434, + 534, + 505 + ], + "blocks": [ + { + "bbox": [ + 439, + 434, + 534, + 505 + ], + "lines": [ + { + "bbox": [ + 439, + 434, + 534, + 505 + ], + "spans": [ + { + "bbox": [ + 439, + 434, + 534, + 505 + ], + "type": "image", + "image_path": "c98d465c2f01795d93b15c5811bcbf24729b242edaaf45c4a0ce086cdc29d1dc.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 506, + 504, + 515 + ], + "lines": [ + { + "bbox": [ + 463, + 506, + 504, + 515 + ], + "spans": [ + { + "bbox": [ + 463, + 506, + 504, + 515 + ], + "type": "text", + "content": "3D View-2" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 518, + 547, + 553 + ], + "lines": [ + { + "bbox": [ + 46, + 518, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 547, + 553 + ], + "type": "text", + "content": "Figure 10. We show qualitative results of our method on real data. Our method produces accurate amodal segmentation, keypoints, as well as 3D poses and shapes across diverse poses and occlusion configurations. In particular, we show results on different level and types of occlusions like vehicle-vehicle (row 1, 2, 3), vehicle-people (row 3), and people-people occlusion (row 4, 5)." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 558, + 288, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 630 + ], + "type": "text", + "content": "tives [54, 73] as well as methods like VIBE [39] for high-quality human reconstructions from videos. Moreover, as our pseudo-groundtruth data includes metric-scale depth information using ground plane, we can augment existing datasets like Relative Human [69] to enhance the robustness of 3D multi-human reconstruction methods to occlusion." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Limitations: Further research is required for generalization across diverse views for it to be used as a generic solution. It also assumes a mean shape or a parametric object model is available, posing challenges for rare objects. Addressing appearance inconsistencies (e.g., variations in lighting) in clip-art images is a promising research direction." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 559, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 630 + ], + "type": "text", + "content": "In conclusion, our method introduces an automated approach to generate a realistic dataset for reconstructing dynamic objects under occlusions from time-lapse imagery. It demonstrates significant improvements in both 2D and 3D object reconstruction, particularly in busy urban scenes with diverse occlusion configurations." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 631, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 545, + 666 + ], + "type": "text", + "content": "Potential Societal Impact. We do not perform any human subjects research from these cameras. For privacy, we blur faces and license plates in all images intended for release." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 666, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 711 + ], + "type": "text", + "content": "Acknowledgements: This work was supported in part by an NSF Grant CNS-2038612, a US DOT grant 69A3551747111 through the Mobility21 UTC and grants 69A3552344811 and 69A3552348316 through the Safety21 UTC." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9521" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] William Agnew, Christopher Xie, Aaron Walsman, Octavian Murad, Yubo Wang, Pedro Domingos, and Siddhartha Srinivasa. Amodal 3d reconstruction for robotic manipulation via stability and connectivity. In CoRL, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "text", + "content": "[2] Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets deep learning for car instance segmentation in urban scenes. In BMVC, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 214 + ], + "type": "text", + "content": "[3] Asra Aslam. Detecting objects in less response time for processing multimedia events in smart cities. In CVPR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "text", + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "type": "text", + "content": "[5] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "text", + "content": "[7] Berk Calli, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. IJRR, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[8] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In CVPR, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "type": "text", + "content": "[9] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In CVPR, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 508 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[11] Falak Chhaya, Dinesh Reddy, Sarthak Upadhyay, Visesh Chari, M. Zeeshan Zia, and K. Madhava Krishna. Monocular reconstruction of vehicles: Combining slam with shape priors. In ICRA, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[12] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In ECCV, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 700 + ], + "type": "text", + "content": "[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "text", + "content": "[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[16] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In CVPR, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 206, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 227 + ], + "type": "text", + "content": "[18] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "text", + "content": "[19] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "type": "text", + "content": "[20] Patrick Follmann, Rebecca König, Philipp Hartinger, Michael Klostermann, and Tobias Böttger. Learning to see the invisible: End-to-end trainable amodal instance segmentation. In WACV, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[21] Rik Fransens, Christoph Strecha, and Luc Van Gool. A mean field em-algorithm for coherent occlusion handling in mape-estimation prob. In CVPR, 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "text", + "content": "[22] Tianshi Gao, Benjamin Packer, and Daphne Koller. A segmentation-aware object detection model with occlusion handling. In CVPR, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 383, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 414 + ], + "type": "text", + "content": "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "text", + "content": "[24] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[25] Holger Glasl, David Schreiber, Nikolaus Viertl, Stephan Veigl, and Gustavo Fernandez. Video based traffic congestion prediction on an embedded system. In ITSC, 2008. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "text", + "content": "[26] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 537, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 558 + ], + "type": "text", + "content": "[27] Google. Google Street View. https://www.google.com/streetview/.3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 559, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 590 + ], + "type": "text", + "content": "[28] Can Gumeli, Angela Dai, and Matthias Nießner. Roca: robust cad model retrieval and alignment from a single image. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 592, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 613 + ], + "type": "text", + "content": "[29] Ruiqi Guo and Derek Hoiem. Beyond the line of sight: labeling the underlying surfaces. In ECCV, 2012. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 615, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 615, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 615, + 545, + 646 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 668 + ], + "type": "text", + "content": "[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 1, 2, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Y.-T. Hu, H.-S. Chen, K. Hui, J.-B. Huang, and A. G. Schwing. SAIL-VOS: Semantic Amodal Instance Level Video Object Segmentation - A Synthetic Dataset and Baselines. In CVPR, 2019. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9522" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In ECCV, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[34] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "type": "text", + "content": "[35] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 205 + ], + "type": "text", + "content": "[36] Gaetano Kanizsa. Organization in vision: Essays on gestalt perception. Praeger Publishers, 1979. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 206, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 248 + ], + "type": "text", + "content": "[37] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 287, + 282 + ], + "type": "text", + "content": "[38] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 284, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 316 + ], + "type": "text", + "content": "[39] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "type": "text", + "content": "[40] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 350, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 370 + ], + "type": "text", + "content": "[41] Philipp Krahenbuhl. Free supervision from video games. In CVPR, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 372, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 372, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 372, + 287, + 403 + ], + "type": "text", + "content": "[42] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "type": "text", + "content": "[43] Alexander Krull, Eric Brachmann, Frank Michel, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Learning analysis-by-synthesis for 6d pose estimation in rgb-d images. In ICCV, 2015. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 449, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 481 + ], + "type": "text", + "content": "[44] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 483, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 483, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 483, + 287, + 514 + ], + "type": "text", + "content": "[45] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, 2018. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 516, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 516, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 49, + 516, + 287, + 547 + ], + "type": "text", + "content": "[46] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "type": "text", + "content": "[47] Chi Li, M Zeeshan Zia, Quoc-Huy Tran, Xiang Yu, Gregory D Hager, and Manmohan Chandraker. Deep supervision with shape concepts for occlusion-aware 3d object parsing. In CVPR, 2017. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "text", + "content": "[48] Fangyu Li, N Dinesh Reddy, Xudong Chen, and Srinivasa G Narasimhan. Traffic4d: Single view longitudinal 4d reconstruction of repetitious activity using self-supervised experts. In IV, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 637, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 287, + 669 + ], + "type": "text", + "content": "[49] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[51] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "text", + "content": "[52] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. SIGGRAPH Asia, 2015. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[53] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[54] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In ECCV, 2022. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[55] M. Naphade, S. Wang, D. C. Anastasiu, Z. Tang, M. Chang, Y. Yao, L. Zheng, M. Shaiqur Rahman, A. Venkatachalapathy, A. Sharma, Q. Feng, V. Ablavsky, S. Sclaroff, P. Chakraborty, A. Li, S. Li, and R. Chellappa. The 6th ai city challenge. In CVPRW, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "text", + "content": "[56] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[57] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "text", + "content": "[58] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 1, 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[59] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In CVPR, 2018. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "type": "text", + "content": "[60] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 472, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 535 + ], + "type": "text", + "content": "[61] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15190-15200, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 537, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 579 + ], + "type": "text", + "content": "[62] N. Dinesh Reddy, Robert Tamburo, and Srinivasa G. Narasimhan. Walt: Watch and learn 2d amodal representation from time-lapse imagery. In CVPR, 2022. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 581, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 613 + ], + "type": "text", + "content": "[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 615, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 646 + ], + "type": "text", + "content": "[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[65] Samuel Schulter, Menghua Zhai, Nathan Jacobs, and Manmohan Chandraker. Learning to look around objects for top-view representations of outdoor scenes. In ECCV, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[66] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9523" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[67] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 119, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 287, + 151 + ], + "type": "text", + "content": "[68] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 185 + ], + "type": "text", + "content": "[69] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J. Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 187, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 187, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 187, + 287, + 209 + ], + "type": "text", + "content": "[70] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In CVPR, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 211, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 211, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 211, + 287, + 243 + ], + "type": "text", + "content": "[71] Andrea Vedaldi and Andrew Zisserman. Structured output regression for detection with partial truncation. In NeurIPS, 2009. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 245, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 245, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 245, + 287, + 277 + ], + "type": "text", + "content": "[72] Khiem Vuong, Robert Tamburo, and Srinivasa G. Narasimhan. Toward planet-wide traffic camera calibration. In WACV, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 49, + 280, + 287, + 312 + ], + "type": "text", + "content": "[73] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. In ICLR, 2021. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 315, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 315, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 315, + 287, + 357 + ], + "type": "text", + "content": "[74] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 359, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 359, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 359, + 287, + 392 + ], + "type": "text", + "content": "[75] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 394, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 425 + ], + "type": "text", + "content": "[76] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 428, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 287, + 460 + ], + "type": "text", + "content": "[77] Y. Xiang, R. Mottaghi, and S. Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 463, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 463, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 463, + 287, + 495 + ], + "type": "text", + "content": "[78] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 287, + 540 + ], + "type": "text", + "content": "[79] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 543, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 543, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 49, + 543, + 287, + 575 + ], + "type": "text", + "content": "[80] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust instance segmentation through reasoning about multi-object occlusion. In CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 577, + 287, + 620 + ], + "type": "text", + "content": "[81] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 622, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 622, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 49, + 622, + 287, + 654 + ], + "type": "text", + "content": "[82] Xiaowei Zhou, Spyridon Leonardos, Xiaoyan Hu, and Kostas Daniilidis. 3d shape estimation from 2d landmarks: A convex relaxation approach. In CVPR, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 657, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 657, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 657, + 287, + 678 + ], + "type": "text", + "content": "[83] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[84] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 1, 2, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[85] M Zeeshan Zia, Michael Stark, and Konrad Schindler. Towards scene understanding with detailed 3d object representations. IJCV, 2015. 2" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9524" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_content_list.json b/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9dc049282867b1428d907e28db0445634e9d2a3b --- /dev/null +++ b/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_content_list.json @@ -0,0 +1,1536 @@ +[ + { + "type": "text", + "text": "WANDR: Intention-guided Human Motion Generation", + "text_level": 1, + "bbox": [ + 207, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Markos Diomataris $^{1,2}$ Nikos Athanasiou $^{1}$ Omid Taheri $^{1}$ Xi Wang $^{2}$ Otmar Hilliges $^{2}$ Michael J. Black", + "bbox": [ + 196, + 179, + 772, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany $^{2}$ ETH Zürich, Switzerland", + "bbox": [ + 194, + 217, + 776, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f6ffb3bc9c796b2db9a211351ebbea7733ac7f0758be8b079cfcc56c28c80de.jpg", + "image_caption": [ + "Figure 1. WANDR starts from an arbitrary body pose and generates precise and realistic human motions that reach a specified 3D goal (depicted as a red sphere). Employing a purely data-driven approach, WANDR is a conditional Variational Autoencoder guided by intention features (depicted arrows) that steer the human's orientation (yellow), position (cyan) and wrist (pink) towards the goal. WANDR is able to reach a wide range of goals even if they deviate significantly from the training data." + ], + "image_footnote": [], + "bbox": [ + 81, + 238, + 893, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 609, + 313, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing natural human motions that enable a 3D human avatar to walk and reach for arbitrary goals in 3D space remains an unsolved problem with many applications. Existing methods (data-driven or using reinforcement learning) are limited in terms of generalization and motion naturalness. A primary obstacle is the scarcity of training data that combines locomotion with goal reaching. To address this, we introduce WANDR, a data-driven model that takes an avatar's initial pose and a goal's 3D position and generates natural human motions that place the end effector (wrist) on the goal location. To solve this, we introduce novel intention features that drive rich goal-oriented movement. Intention guides the agent to the goal, and interactively adapts the generation to novel situations without needing to define sub-goals or the entire motion path. Crucially, intention allows training on datasets that have goal-oriented motions as well as those that do not. WANDR is", + "bbox": [ + 75, + 643, + 472, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a conditional Variational Auto-Encoder (c-VAE), which we train using the AMASS and CIRCLE datasets. We evaluate our method extensively and demonstrate its ability to generate natural and long-term motions that reach 3D goals and generalize to unseen goal locations. Our models and code are available for research purposes at wandr.is.tue.mpg.de.", + "bbox": [ + 496, + 611, + 893, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 738, + 632, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Goals drive our motions. Even the simplest goal can give rise to intricate motions. Consider reaching for a coffee cup – it can be as straightforward as an arm extension or can involve the coordinated action of our entire body. Actions like bending down, extending our arm, and walking must come together to achieve the goal. At a granular level, we continuously make subtle adjustments to maintain balance and stay on course towards our objective. The result is a fluid motion that seamlessly integrates numerous smaller movements, all", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "927", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "converging toward a common and simple goal: placing our hand on the cup. Generating this hierarchy of motions, from the overarching goal to the moment-to-moment individual actions, remains a longstanding challenge in computer vision, graphics, and robotics.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here we focus on a representative task, illustrated in Fig. 1: given a goal location in space and a starting pose, a humanoid agent must place an end effector (wrist joint) on the goal location while moving in a natural human-like way. To solve the task, the agent needs to be able to approach the goal, orient itself towards it, and reach out such that its wrist makes contact with the goal. Our primary emphasis is on ensuring autonomy for human agents. Consequently, we strive to minimize the guidance information provided, limiting it only to the human's initial pose and the goal's position. Diverging from prior data-driven approaches [4, 17], we choose to refrain from evaluating the model solely on limited labeled data. Instead, we devise an evaluation pipeline that requires agents to reach goals positioned in diverse locations around them. Considering the arbitrary selection of the goal during evaluation, and the minimal guidance information provided, tackling this task is challenging, demanding an approach with the capacity to generalize beyond the distribution of the training dataset.", + "bbox": [ + 75, + 167, + 470, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods approach this problem either using reinforcement learning (RL) [11, 15, 24, 43] or by capturing task-specific datasets [4, 8, 32]. While RL provides a principled way to explore the solution space, it comes with considerable shortcomings. The \"trial and error\" of exploratory learning, in combination with the high dimensionality of human motion result in policies requiring an enormous amount of training even to achieve simple tasks such as walking to a waypoint [5, 43]. In addition, since motion naturalness is better captured by data and not reward functions, RL approaches tend to produce motions that lack naturalness and expressiveness. Data-driven approaches on the other hand, rely on plentiful training motions that are acquired through motion capture and carefully curated for the downstream tasks [4, 10]. Such approaches do not scale and do not generalize well to out-of-distribution tasks.", + "bbox": [ + 75, + 458, + 468, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In prioritizing both motion realism and training efficiency, we adopt a data-driven approach. However, current data-driven methods lack the ability to learn both from smaller datasets that provide high-quality human reaching motions with goal labels, and from unlabeled larger scale datasets that contain necessary motion skills such as navigating to a goal position. This raises two key challenges. First, how do we model human motion in a way that generated motions can combine skills from different datasets? Second, what should the training objective be in the cases where goal labels are absent?", + "bbox": [ + 75, + 700, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose WANDR (Wrist-driven Autonomous Navigation for Data-based goal", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reaching). We observe that by modeling human motion generation as an autoregressive stochastic process that produces motions frame by frame, WANDR is able to combine pieces of different dataset distributions when generating a motion sequence. Each generation step is conditioned on goal-related information that we call intention (visualized arrows in Fig. 1). We carefully design intention in a way that strikes a balance between being informative enough to guide the avatar to reach the goal, while also being abstract enough to promote generalization to unseen goals. This allows our generated motions to reach goals that were never encountered during the training phase in a completely zero-shot evaluation scenario. By generating the motion in an autoregressive way, we disentangle the spatial and temporal dimensions of motion. This is necessary as it allows our model to generate novel long-term sequences while being realistic in terms of local dynamic details.", + "bbox": [ + 496, + 90, + 890, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In more detail, our method is based on a conditional Variational Auto-Encoder (c-VAE) that learns to model motion as a frame-by-frame generation process by auto-encoding the pose difference between two adjacent frames. The condition signal consists of the human's current pose and dynamics along with the intention information. Intention is a function of both the current pose and the goal location and therefore actively guides the avatar during the motion generation in a closed loop manner. Through training, the c-VAE learns the distribution of potential subsequent poses conditioned on the current dynamic state of the human and its intention towards a specific goal. We train WANDR using two datasets: AMASS [18], which captures a wide range of motions including locomotion, and CIRCLE [4], which captures reaching motions.", + "bbox": [ + 496, + 349, + 890, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although AMASS is large, it lacks any explicit label of goals or intentions. To address this, inspired by the Hind-sight Experience Replay paradigm in robotics [3], we define intention using a hallucinated goal derived from the ground-truth wrist position in a future frame. This approach allows us to establish a unified training objective spanning AMASS and CIRCLE. Consequently, our model learns to combine motions from both datasets, enabling it to effectively reach arbitrary goals during testing.", + "bbox": [ + 496, + 580, + 890, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we present WANDR, a data-driven method that combines an autoregressive motion prior with a novel intention guiding mechanism and is able to generate avatars that realistically move in space and reach arbitrary goals. We experimentally evaluate our approach, including the benefit of combining multiple datasets as well as the generalization capabilities of our motion generator. Our results underscore the efficacy of the intention mechanism as an elegant way of guiding the motion generation process while also enabling the incorporation of pseudo goal labels for datasets lacking explicit goal annotations. The model and code are available for research purposes.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "928", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early research in motion generation focuses on tasks like motion prediction [1, 2, 7, 9, 19, 27] and unconstrained motion generation [6, 16, 21, 23, 28, 29, 36-38, 40]. More recently, significant effort has been devoted into improving controllability, with a focus on motion generation conditioned on different types of goals [25, 39], enabling interactions with scenes [12, 13, 20] and objects [31, 33, 42, 45]. Methods that attempt goal-driven motion generation can be broadly divided into reinforcement learning or data-driven approaches.", + "bbox": [ + 75, + 114, + 470, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Reinforcement Learning for Motion Synthesis", + "text_level": 1, + "bbox": [ + 76, + 280, + 465, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many existing works employ Reinforcement Learning (RL) for the generation of task-specific long motion sequences. Representative work includes MotionVAE [15] and AMP [24]. MotionVAE [15] employs a two-step process where it initially leverages an autoregressive conditional Variational Autoencoder (VAE) to construct a latent space that encapsulates possible human movements. Subsequently, it utilizes RL to sample from this action space to reach a designated target location while avoiding obstacles by monitoring the area ahead. Similar to MotionVAE, GAMMA [43] learns a policy to extract samples from a latent space and then employs a tree-based search algorithm to find viable motions that steer clear of obstacles by considering the environmental geometric constraints. DIMOS [44] further extends the GAMMA framework by introducing two specialized policy networks: one for locomotion and one for interaction. Together these networks generate goal-conditioned motion sequences that dynamically interact with objects and the environment. AMP [24] learns an adversarial motion prior from unstructured datasets and then applies goal-conditioned RL. This approach involves the formulation of a style reward to encourage the resemblance of the generated sequences to those in the dataset, complemented by a task-specific reward aimed at achieving a particular objective. Hassan et al. [11] extend AMP to produce motions that facilitate interactions with the scene, by conditioning both the discriminator and the policy network on the scene context. However, RL requires significant computation and struggles to generate natural and expressive motion sequences.", + "bbox": [ + 75, + 303, + 470, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Data-driven Approaches for Motion Synthesis", + "text_level": 1, + "bbox": [ + 76, + 770, + 465, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most data-driven approaches use existing motion capture (MoCap) datasets [18, 35] to train their models through supervised learning. The pioneering Neural State Machine method [30] is a data-driven technique for generating motion with character-scene interactions, focusing on scenarios with a limited number of objects and interactions. HuMoR [26] proposes a robust model for 3D human shape and", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "temporal pose estimation, yet it falls short of generating motions that are conditioned on specific goals. The SAMP [10] method, designed for real-time stochastic motion synthesis, generates diverse human-scene interaction movements by breaking down the process into predicting goals, planning paths, and generating motion along a predefined route. The GOAL [32] method, trained on the GRAB dataset [31], produces motion sequences in which humans walk towards and grasp 3D objects. However, the generated motions exhibit minimal movements, especially in the feet. To address these constraints, the newly introduced CIRCLE [4] dataset provides a collection of reaching motion data. This dataset is used to train a neural network that generates diverse scene-aware reaching motions. Recently, diffusion models have seen the most success at generating motions conditioned on textural input [34, 41] and spatial data [14]. This advancement enables the synthesized motion to accurately reach specified target locations or navigate around obstacles. Nevertheless, the effectiveness of data-driven approaches is constrained by the amount of training data and they lack generalization to out-of-distribution scenarios.", + "bbox": [ + 496, + 90, + 893, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 449, + 589, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to have a virtual human that can autonomously and realistically move from an initial pose to an arbitrary goal position and accurately place its right hand on the target. This challenge requires a nuanced understanding of human motion and the intricate dynamics involved in goal-oriented motions. For example, when the human tries to reach a distant goal, the motions are mostly focused on the legs and navigating the body to approach the object, but when it gets close to the object, the focus will be on moving the arms and upper body to reach the target. Using these observations, we develop a method named WANDR, which, although trained in a supervised setting on motion capture data, exhibits generalization in reaching unseen goal locations during test time. WANDR is designed to generate human motion in an autoregressive frame-by-frame fashion, conditioned on novel intention features. During training, these features are extracted by picking a future frame as the goal for the wrist. During inference, the intention features are dynamically computed based on the goal's position in a feedback loop, guiding the virtual human to reach the goal. See Fig. 2 for the network overview.", + "bbox": [ + 496, + 483, + 893, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first consider the different distributions of the datasets we will be using (section 3.1). Following this, we detail the components of the intention features and how they are computed during both the training and inference phases (Section 3.2). Finally, we define the motion representation and motion generator network (Section 3.3).", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "929", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/40453d55b9cf3fcc28638003879d9578750b23cf414d6914c1f57b4eaa8dbca2.jpg", + "image_caption": [ + "Figure 2. WANDR architecture. During training, our model conditions on the intention vectors $I^p$ , $I^r$ and $I^w$ , learning to associate them with actions that result into reaching goals realistically. When the training data has no defined goal, we create a goal based on the wrist location in future frames; see Sec. 3.2. The state of the avatar, $p_i^{dyn}$ expresses the SMPL-X local pose parameters $p_i$ , as well as the deltas $d_{i-1}$ the body parameters have in frame $i-1$ . During inference, WANDR takes the intention features, the state, and random noise and returns the change in pose, $\\hat{d}_i$ . The next pose, $\\hat{p}_i$ is obtained by integrating the $\\hat{d}_i$ with the previous pose $\\hat{p}_{i-1}$ ." + ], + "image_footnote": [], + "bbox": [ + 116, + 85, + 854, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a0993d495849048161452f010dd815479f228bd758fd187ae25d9f04bde8d14c.jpg", + "image_caption": [ + "Figure 3. In training, if goals are not specified, they are determined by the future wrist location at a randomly selected future timestep, compensating for the lack of paired ground-truth data in AMASS and direct human motion through intention vectors. During inference, target locations are used as goals with intention vectors calculated based on these specific locations." + ], + "image_footnote": [], + "bbox": [ + 89, + 386, + 272, + 500 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/491de65aee91085c701c2ca5ddf33e417a06f71cd35ccc8e52632abb781c9ed1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 386, + 455, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Two Complementing Datasets", + "text_level": 1, + "bbox": [ + 76, + 650, + 343, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the development of WANDR, we use two key datasets: AMASS and CIRCLE. AMASS is a large-scale dataset that offers a broad range of general human motions but lacks a specific focus on goal-reaching tasks. Its diverse collection of movements provides a solid base for understanding human locomotion and body movement when the person is far away from the goal location. In contrast to AMASS, CIRCLE is tailored towards movements involving reaching specific target positions, particularly capturing the nuances of upper body and arm movements. By integrating AMASS's general motion diversity with CIRCLE's targeted goal-reaching data, we equip WANDR with the ability to generate the entire process of reaching a distant goal, from the initial navigation to the precise target-reaching actions.", + "bbox": [ + 75, + 688, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Intention Features", + "text_level": 1, + "bbox": [ + 500, + 388, + 679, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We represent 3D human motion as a sequence of SMPL-X [22] body poses $\\mathbf{p} = \\{p_1,\\dots,p_N\\}$ . Each pose $p_i\\in \\mathbb{R}^{135}$ consists of three concatenated components: the body's translation $t_i\\in \\mathbb{R}^3$ , root orientation $r_i\\in \\mathbb{R}^6$ , and the body pose $\\theta_{i}\\in \\mathbb{R}^{21\\times 6}$ , both in 6D format [46].", + "bbox": [ + 498, + 411, + 892, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the avatar to reach the goal, it is important to be informed about the spatial relation of the goal location with respect to its current pose, as well as to have a sense of time to reach the goal promptly. We achieve this by introducing the intention features, which are central to our approach.", + "bbox": [ + 498, + 487, + 890, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To define the intention features $I_{i}$ at timestep $i$ , it is crucial to first establish the selection criteria for the goal $G \\in \\mathbb{R}^3$ , within a motion sequence. Our training involves both label-available scenarios (e.g., CIRCLE dataset) and label-absent scenarios (e.g., AMASS dataset). In label-available cases, where the goal position $G$ and the frame index $t_{G}$ which it is reached are known, calculating $I_{i}$ is straightforward. Conversely, for label-absent scenarios like in AMASS, we pick a random future frame as the $t_{G}$ and define the human's right wrist joint location as the goal $G$ (Fig. 3).", + "bbox": [ + 498, + 563, + 892, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In both scenarios, the intention features are defined as:", + "bbox": [ + 517, + 729, + 879, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} = I _ {i} \\left(p _ {i}, G, t _ {G}, i\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 755, + 764, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These features are essentially a function of the current body pose, the goal position, and time, offering both spatial and temporal insights required for the motion to reach the goal. They are designed to provide sufficient information to reach goals while also enabling test-time generalization. We define them as three distinct components:", + "bbox": [ + 498, + 782, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} = \\left(I _ {i} ^ {w}, I _ {i} ^ {r}, I _ {i} ^ {p}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 883, + 756, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "930", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These components represent wrist intention, body orientation intention, and pelvis intention, respectively.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Wrist Intention: This is the main time-dependent component that guides the wrist to reach the goal. It is calculated as the necessary average velocity for the wrist to be at the goal location in time, defined as:", + "bbox": [ + 75, + 121, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {w} = \\frac {G - W _ {i}}{t _ {G} - i}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 191, + 325, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_G$ is the frame when the goal should be reached. During training, we know the goal frame and the time to reach it. At test time, we know the goal location but need an externally-supplied time to reach it. An emergent behavior of this formulation is that, at inference, the model is able to adjust its movement speed and reach the goal just in time.", + "bbox": [ + 75, + 234, + 467, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Orientation Intention: This component captures the body orientation when reaching the goal location. By conditioning on this, we ensure that the human model orients towards the goal and smoothly navigates towards it, preventing unnatural motions during inference, like walking backward. During training, this is defined as the difference between the forward direction of the current body frame, $H_{i}^{xy}$ , and the goal body, $H_{tG}^{xy}$ where $xy$ signifies removing the $z$ component. During inference, since we do not have the goal body, we use the pelvis position $P_{i}$ to calculate the pelvisto-goal direction as the desired orientation. This feature is formulated as:", + "bbox": [ + 75, + 325, + 467, + 505 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {r} = \\left\\{ \\begin{array}{l l} H _ {t _ {G}} ^ {x y} - H _ {i} ^ {x y} & \\text {d u r i n g t r a i n i n g} \\\\ (G - P _ {i}) ^ {x y} - H _ {i} ^ {x y} & \\text {d u r i n g i n f e r e n c e}. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 516, + 426, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pelvis Intention: This feature captures information about the position of the goal relative to the body. It is the difference between the goal and the pelvis joint, excluding the z (height) component. Following the approach in [32], we scale this distance by an exponential function that saturates this vector to have a maximum norm of 2. This formulation helps the method generalize to navigating towards the goal during longer motions and helps the model learn since the distance from the goal does not grow indefinitely in extreme scenarios. This intention is defined by the equation:", + "bbox": [ + 75, + 568, + 467, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {p} = 2 \\times \\left(1 - e ^ {\\left\\| G ^ {x y} - P _ {i} ^ {x y} \\right\\| _ {2}}\\right) \\times \\frac {G ^ {x y} - P _ {i} ^ {x y}}{\\left\\| G ^ {x y} - P _ {i} ^ {x y} \\right\\| _ {2}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 729, + 436, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experimental section, we delve into the significance of each of these intention features and discuss the rationale behind our design choices, illustrating their impact on the effectiveness of our model.", + "bbox": [ + 75, + 776, + 467, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Motion Network (WANDR)", + "text_level": 1, + "bbox": [ + 76, + 845, + 323, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WANDR is designed as a conditional Variational AutoEncoder (c-VAE) network, operating in an autoregressive", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "manner to generate sequential motion frames. This framework is pivotal in predicting the subsequent pose in a motion sequence, emphasizing an incremental, frame-by-frame approach.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Central to our approach is the training of the c-VAE to autoencode pose deltas, denoted as $d_{i} \\in \\mathbb{R}^{135}$ . These deltas represent the difference between two consecutive poses, $p_{i}$ and $p_{i - 1}$ . By focusing on pose deltas rather than absolute pose values, our model benefits from an important inductive bias, enhancing its learning efficiency and performance, as supported by prior research [15, 26]. We separate rotational differences into: body orientation $(d_i^r)$ and body pose $(d_i^\\theta)$ , each expressed in a 6-D rotational format. Translation deltas are denoted as $d_{i}^{t} = t_{i} - t_{i - 1}$ .", + "bbox": [ + 496, + 151, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the motion representation's invariance, we remove information related to the global $z$ -orientation. This is accomplished by subtracting the global $z$ Euler angle of $r_{i-1}$ from both the translational $(d_i^t)$ and rotational $(d_i^r)$ deltas. The resulting deltas, $d_i^{t-z}$ for translation and $d_i^{r-z}$ for orientation, provide a more robust and consistent representation of motion, irrespective of global direction. Consequently, the delta pose features for any given frame $i$ are composed as follows:", + "bbox": [ + 496, + 303, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i} = \\left(d _ {i} ^ {t - z}, d _ {i} ^ {r - z}, d _ {i} ^ {\\theta}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 450, + 769, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An advantage of this representation is its consistency across different motion global orientations. For instance, in the scenario of a person walking, the delta representation remains agnostic to the walking direction. This attribute underscores the efficacy of our method in capturing the essence of motion without being biased towards any specific orientation or direction.", + "bbox": [ + 496, + 479, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Condition Inputs: For each motion frame, the decoder is conditioned on a combination of state and intention features. Specifically, this condition signal is formulated as $c_{i} = (p_{i}^{dyn}, I_{i})$ . The state features, $p_{i}^{dyn}$ , encapsulate the avatar's current local pose, focusing on the z-component of translation and a modified orientation that excludes the global z Euler angle, along with the pose deltas $di - 1$ of the previous step. That combination ensures that the generated motion at each step is informed by both the local pose configuration of the avatar, its dynamics and its directional intention towards the set goal, vital for producing realistic, goal-oriented human motions. An overview of the network architecture is shown in Fig. 2.", + "bbox": [ + 496, + 587, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Losses", + "text_level": 1, + "bbox": [ + 500, + 792, + 656, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training objective is a composite of three distinct loss functions:", + "bbox": [ + 496, + 816, + 888, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r e c} + \\alpha \\mathcal {L} _ {K L} + \\mathcal {L} _ {J}.\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 847, + 782, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reconstruction loss, $\\mathcal{L}_{rec}$ , measures the accuracy of the motion reconstruction, quantified as the mean square er", + "bbox": [ + 496, + 869, + 888, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "931", + "bbox": [ + 485, + 945, + 508, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ror (MSE) between the input pose delta, $d_{i}$ , and its reconstructed counterpart, $\\hat{d}_{i}$ . It ensures the network's ability to faithfully replicate the input motion.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The KL Divergence Loss, $(\\mathcal{L}_{KL})$ , evaluates the deviation of the encoded distribution from a standard normal distribution. It is formulated as:", + "bbox": [ + 75, + 137, + 468, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} = \\mathcal {K L} (\\mathcal {N} (0, I) | | \\mathcal {N} (\\mu_ {i}, \\sigma_ {i})).\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 196, + 385, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mu_{i}$ and $\\sigma_{i}$ represent the mean and variance of the Gaussian distribution predicted by the encoder. We balance this term with $\\alpha = 10^{-2}$ to prevent the over-dominance of $\\mathcal{L}_{KL}$ , thereby aiding the decoder in avoiding collapse to mean predictions.", + "bbox": [ + 75, + 227, + 468, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we use a Joint Error Loss $(\\mathcal{L}_J)$ , to ensure perceptual accuracy, by integrating the predicted $\\hat{d}_i$ to get the predicted next pose $\\hat{p}_i$ , which is then fed into the SMPL-X model to obtain the predicted joint positions, $\\hat{J}$ . The loss $\\mathcal{L}_J$ is the MSE between these predicted joints $\\hat{J}$ and the ground truth joints $J$ , addressing errors that might not be apparent in parameter space but are perceptually significant, such as incorrect body orientation.", + "bbox": [ + 75, + 304, + 468, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, our approach does not incorporate any explicit loss functions directly related to reaching a goal. This omission is a deliberate choice, aligning with our method's emphasis on generalizing to diverse goal-reaching scenarios without being constrained by goal-specific training losses.", + "bbox": [ + 75, + 425, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Motion Generation", + "text_level": 1, + "bbox": [ + 76, + 513, + 261, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the inference phase of WANDR, our primary objective is to generate human motion that is driven towards a specific goal. Using the decoder of the WANDR c-VAE, we iteratively generate and integrate pose deltas. This process is initiated from the starting pose and progressively builds upon each subsequent pose.", + "bbox": [ + 75, + 537, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The intention features are recalculated at each step based on the current predicted pose and the goal location. They serve as a guiding mechanism, ensuring that the generated motion is consistently oriented towards placing the human's right wrist on the target.", + "bbox": [ + 75, + 630, + 468, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The user can control the motion's pace by specifying the time to reach the goal $t_{G}$ . This directly affects the wrist intention feature, enabling adjustments from fast to slow motions to suit various scenarios and constraints.", + "bbox": [ + 75, + 705, + 468, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 782, + 207, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we outline the datasets used for training and evaluation and benchmark how each dataset affects the goal-reaching ability and the quality of the generated motions. Furthermore, we compare our approach with several baselines and ablate the effect of the different components of our intention vector.", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets & Processing", + "text_level": 1, + "bbox": [ + 500, + 90, + 707, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is trained on two datasets: AMASS [18] and CIRCLE [4]. AMASS is a collection of 17k sequences, containing a wide range of motion types including long-term navigational skills like walking and turning. CIRCLE, on the other hand, contains 7.2k shorter sequences, each marked with a specific goal reached by a hand. For training, we refine AMASS by excluding sequences where feet are more than $20cm$ above the ground, resulting in a combined dataset of nearly 20k sequences. This dataset is split into $80\\%$ training, $10\\%$ validation, and $10\\%$ test sets. All motions are re-sampled to 30 frames per second (fps).", + "bbox": [ + 496, + 114, + 890, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Strategy", + "text_level": 1, + "bbox": [ + 500, + 292, + 687, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our evaluation procedure aims at testing the degree which WANDR can generalize to generating reaching motions that start from unseen poses and reach the whole range of 3D space around the starting pose. This is why we choose not to evaluate on held-out motion-goal pairs from the training data. Instead, we only hold out initial poses. During evaluation, starting from these unseen poses, we generate motions that attempt to reach goals that uniformly cover the volume of a cylinder centered on the human, including completely out-of-distribution goal locations (see Sup. Mat. Sec. 5).", + "bbox": [ + 496, + 316, + 890, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, the set of evaluation goals is defined in a cylindrical coordinate frame by taking all the combinations of (1) 5 angles equally separating the 360 degrees around the human, (2) 5 different goal heights ranging from 0 to 1.8 meters and (3) 5 distances from 0.5 to 5 meters. We generate motions from 6 different initial poses, with an 8-second duration specified for reaching each goal. Five motions are sampled for each pose-goal combination, resulting in $5 \\times 5 \\times 5 \\times 6 \\times 5 = 3750$ unique motion sequences from which our metrics are computed. This setup allows us to thoroughly test our model in diverse scenarios, including long-term movements, navigational skills, and reaching motions at various heights and distances.", + "bbox": [ + 496, + 468, + 890, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 676, + 683, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To accurately assess the effectiveness of our approach in generating realistic, goal-oriented human motion, we employ a set of metrics focused on both the ability to successfully reach the intended goal and the naturalness of the motion. These metrics are:", + "bbox": [ + 496, + 700, + 890, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- **Success Rate (SR):** This quantifies the percentage of motions where the right wrist reaches within $10\\mathrm{cm}$ of the goal, indicating successful goal attainment. The criterion for success aligns with that used in [4].", + "bbox": [ + 500, + 780, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Foot Skating (FS): FS evaluates the naturalness of the motion based on foot skating, where a frame is considered as having foot skating if the lowest vertex of the human mesh moves more than $0.66 \\mathrm{~cm}$ between consecutive", + "bbox": [ + 500, + 839, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "932", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2a54a6a851db091d9d8890cd4f6bffba33ab5aa51cdd843f5eee8b3634ab48b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 282, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92bb0804f869f5d58f8efcb7d32025f886e70701e90af6e669dcb19ba2f95564.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 282, + 89, + 483, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02c60a9a8cd2c0c0b8abdef35b801868fdf8dc0370a2cd5e1d219fa6fc84e006.jpg", + "image_caption": [ + "(c)", + "Figure 4. Diverse motion generated with WANDR: Displaying a range of motions generated by WANDR from various initial poses towards arbitrary goals. Examples include navigating towards goals from initial orientations not facing the goal (a, b, c, d), elevating the right hand to reach higher targets (c), and bending down to access goals near the floor (d), showcasing the model's ability to adapt to novel goal locations." + ], + "image_footnote": [], + "bbox": [ + 486, + 89, + 684, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d64e743a7666665c40dcbe4fe1ba1c10a548d588277e1d5533fad2213c9f0c51.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 686, + 89, + 890, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames (adjusted from the 1cm threshold used in [4] to accommodate our 30fps motion generation).", + "bbox": [ + 89, + 446, + 468, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Distance to Goal (DTG): DTG records the closest distance in cm that the right wrist gets to the goal during a motion. This metric offers a nuanced view of the model's capability to guide the motion towards the goal.", + "bbox": [ + 76, + 477, + 470, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results", + "text_level": 1, + "bbox": [ + 76, + 547, + 169, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 573, + 272, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combining AMASS and CIRCLE: Our evaluation in Table 1 validates our hypothesis about the benefits of training with both the AMASS and CIRCLE datasets. On the one hand, training only on CIRCLE (line 1) is not sufficient for the model to learn necessary navigational skills, such as walking, due to the dataset's narrow focus or reaching motions. This is apparent from the very high foot-skating. On the other hand, training only on AMASS (line 2) results in high-quality motion generation with low foot-skating, but a relatively low success rate in goal-reaching. The Distance to Goal (DTG) metric, suggests that the model is able to navigate close to the target, but it lacks the precise movement needed to successfully reach for the goal. Using both datasets (line 3) illustrates how our approach effectively merges the broad motion vocabulary of AMASS with the goal-oriented precision of CIRCLE, leading to both high-quality motion and improved goal-reaching capability.", + "bbox": [ + 75, + 597, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare with GOAL [32], a method that generates human motions that reach and grasp objects. GOAL is trained on GRAB [31], a dataset purely consisting of mo", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b16be26b025b25f20a868500e6c49b36c49d47a82a35be22c7708a5b93336b60.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR/Circle0%56%205.4
WANDR/AMASS16%19%48.0
WANDR32%16%24.8
GOAL [32]0%29%149.2
", + "bbox": [ + 521, + 444, + 870, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. We evaluate WANDR trained on different datasets and compare with GOAL [32]. Training solely on CIRCLE results in unrealistic motions, whereas AMASS excels in motion quality but struggles with finer goal-reaching skills. WANDR, leveraging both of what these datasets offer, demonstrates realistic motions as well as better ability to reach goals compared to baselines and existing methods.", + "bbox": [ + 498, + 535, + 890, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions of humans grasping and manipulating objects. Since the relative positioning of the human and the object as well as the motions in GRAB have very small variations, GOAL does not succeed in any of the evaluation configurations (line 4).", + "bbox": [ + 498, + 655, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Intention Features: In order to demonstrate the contribution of each component of the intention feature we conduct an ablation study (Table 2). Using only wrist intention (line 1) results in the lowest foot skating, due to the minimal constraints applied to the motion, allowing for more adaptable motion planning. But wrist intention features, are time-dependent and do not carry information about the absolute distance to the goal. This is why it can lead to the avatar over- or under-shooting and thus achieving a low success ratio (SR). The addition of pelvis intention (line 2) enables the avatar to sense the distance to the", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "933", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "goal while the orientation intention (line 3) properly aligns the body to face the goal. Since pelvis and orientation intention add more constraints to the motion, they can sometimes cause more challenging body dynamics and produce motions with higher foot-skating (FS) (e.g. turning around in place instead of walking in a U-turn). We also try removing the intention from the motion prior and optimizing the latent space of a randomly generated initial motion to minimize the distance between the wrist and the goal (VAE + opt). This approach fails since the result is heavily dependent on the initialization of the latent variables of the motion. Our results confirm that each component of the intention feature is essential to achieving the overall performance of the model. For more details on the optimization see Sup. Mat.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5127dbd43d7465ee282e8efc44852aad78303ff1600ee684ffda2075b5ce74a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR (Iw)15%13%62.9
WANDR (Iw + Ip)18%17%44.9
WANDR (Iw + Ir)19%19%36.0
WANDR (full intention)32%16%24.8
VAE + opt3%4%217.0
", + "bbox": [ + 78, + 329, + 468, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation Study. We evaluate the impact of each component of the intention vector. We also compare with an optimization baseline that does not use any condition signals. The results highlight the effectiveness of all of the components of intention as well as the fact that the complexity of the task makes \"brute-forcing\" with optimization unsuccessful.", + "bbox": [ + 75, + 438, + 468, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Success Ratio Distribution: Our decomposition of the model's success ratio, presented in Fig. 5, offers insights into how the model's performance varies with respect to different goal positions. WANDR demonstrates a consistent ability to reach goals across various distances (blue) and directions (green). It is more capable at reaching goals that are closer to the natural position of the wrist and do not require extensive bending or stretching (yellow). This trend likely results from the abundance of standing or upright motion sequences in the training data, as opposed to motions involving bending or crouching. This analysis provides valuable information for future improvements and dataset balancing.", + "bbox": [ + 75, + 539, + 472, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2 Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 739, + 264, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 4, we show a variety of motion sequences generated with our network featuring reaching goals located at varying distances and heights, highlighting the model's ability to realistically and smoothly orient, navigate, and reach for goals. These goals require actions such as bending down, turning, or stretching upwards. A critical aspect observed is the model's ability to decelerate as it approaches the goal, seamlessly coordinating body and arm movements to achieve a natural-looking reaching motion. Overall, the", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/553e7be86e57d9c1cd546370764f29503c78f0a949f4f792b35eafcf2d99bdfe.jpg", + "image_caption": [ + "Figure 5. We show the success rates of reaching goals at various heights, angles, and distances from the initial human pose. It highlights how goal position affects the model in accurately navigating and achieving the goals." + ], + "image_footnote": [], + "bbox": [ + 496, + 85, + 893, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "qualitative results show that our network generalizes well to novel goal locations while generating realistic motions. For more results please see Sup. Mat. and the video.", + "bbox": [ + 496, + 313, + 893, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 372, + 617, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, our research presents a novel data-driven approach to human motion generation, focusing on the task of reaching arbitrary goals in space. We introduce novel intention features that enable learning both general navigational skills from AMASS and goal-reaching skills from the CIRCLE dataset under the same distribution. We evaluate our model's ability to reach unseen goals that cover the whole space an avatar should be able to reach around it. The autoregressive design of WANDR demonstrates generalizability in generating realistic human motions that reach unseen goals without requiring any extra guidance information such as a pre-defined trajectory.", + "bbox": [ + 496, + 398, + 890, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work: Our approach is not without its limitations. Currently, error accumulation can sometimes bring the avatar to states where it can no longer recover. Additionally, our model shows less proficiency in reaching extremely low or high goals, reflecting a need for more diverse training data encompassing a wider range of body movements. Future work could focus on incorporating realistic grasping mechanisms and interactions with objects, as well as including scene navigation capabilities. This could involve integrating more complex datasets or developing advanced algorithms capable of understanding and interacting with varied environmental contexts, thereby pushing the boundaries of realistic human motion simulation.", + "bbox": [ + 496, + 579, + 892, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: Markos Diomataris was supported in part by the Max Planck ETH Center for Learning Systems. We thank P. Ghosh, O. Ben Dov S. Tripathi, for the fruitful discussions and A. Cseke, T. Niewiadomski, T. Alexiadis, T. McConnell, for conducting the user studies.", + "bbox": [ + 496, + 790, + 890, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conflicts of Interest:", + "text_level": 1, + "bbox": [ + 500, + 866, + 647, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt", + "bbox": [ + 519, + 881, + 864, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "934", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pages 565-574. IEEE, 2021. 3", + "[2] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5223-5232, 2020. 3", + "[3] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. In Advances in Neural Information Processing Systems, 2017. 2", + "[4] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 2, 3, 6, 7", + "[5] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In International Conference on 3D Vision (3DV), 2024. 2", + "[6] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3D human motion synthesis model via conditional variational auto-encoder. In Computer Vision and Pattern Recognition (CVPR), pages 11645–11655, 2021. 3", + "[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision (ECCV), pages 387-404. Springer, 2020. 3", + "[8] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[9] Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 2017 International Conference on 3D Vision (3DV), pages 458-466. IEEE, 2017. 3", + "[10] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. arXiv:2108.08284, 2021. 2, 3", + "[11] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 2, 3", + "[12] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. arXiv:1904.05767, 2019. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3", + "[14] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[15] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion VAEs. Transactions on Graphics (TOG), 2020. 2, 3, 5", + "[16] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. Transactions on Graphics (TOG), 2020. 3", + "[17] Matthew Loper, Naureen Mahmood, and Michael J Black. Mosh: motion and shape capture from sparse markers. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 33(6):220-1, 2014. 2", + "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 6", + "[19] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[20] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In International Conference on 3D Vision (3DV), 2024. 3", + "[21] Dirk Ormoneit, Hedvig Sidenbladh, Michael Black, and Trevor Hastie. Learning and tracking cyclic human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3", + "[22] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 4", + "[23] Vladimir Pavlovic, James M Rehg, and John MacCormick. Learning switching linear models of human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3", + "[24] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: adversarial motion priors for stylized physics-based character control. Transactions on Graphics (TOG), 2021. 2, 3", + "[25] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[26] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "935", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44 (6):3300-3315, 2021. 3", + "[28] Hedvig Sidenbladh, Michael J Black, and David J Fleet. Stochastic tracking of 3d human figures using 2d image motion. In European Conference on Computer Vision (ECCV). Springer, 2000. 3", + "[29] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision (IJCV), 87 (1-2):4-27, 2010. 3", + "[30] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. Transactions on Graphics (TOG), 2019. 3", + "[31] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3, 7", + "[32] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7", + "[33] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. Grip: Generating interaction poses using latent consistency and spatial cues, 2023. 3", + "[34] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[35] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In British Machine Vision Conference (BMVC), pages 1-13, 2017. 3", + "[36] Raquel Urtasun, David J Fleet, and Pascal Fua. Temporal motion models for monocular and multiview 3d human body tracking. Computer Vision and Image Understanding (CVIU), 104(2-3):157-177, 2006. 3", + "[37] Qi Wang, Thierry Artières, Mickael Chen, and Ludovic Denoyer. Adversarial learning for modeling human motion. The Visual Computer, 36(1):141-160, 2020.", + "[38] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI conference on artificial intelligence, pages 12281-12288, 2020. 3", + "[39] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In International Conference on Computer Vision (ICCV), pages 2228-2238, 2023. 3", + "[40] Ping Yu, Yang Zhao, Chunyuan Li, Junsong Yuan, and Changyou Chen. Structure-aware human-action generation." + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 18-34. Springer, 2020. 3", + "[41] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3", + "[42] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision, pages 518-535. Springer, 2022. 3", + "[43] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[44] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3", + "[45] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision (ECCV), pages 676-694. Springer, 2022. 3", + "[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 4" + ], + "bbox": [ + 501, + 90, + 892, + 488 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "936", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_model.json b/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d68720c64a1e7d789701fb08f05c9c872df51cc7 --- /dev/null +++ b/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.764, + 0.154 + ], + "angle": 0, + "content": "WANDR: Intention-guided Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.18, + 0.774, + 0.217 + ], + "angle": 0, + "content": "Markos Diomataris \\(^{1,2}\\) Nikos Athanasiou \\(^{1}\\) Omid Taheri \\(^{1}\\) Xi Wang \\(^{2}\\) Otmar Hilliges \\(^{2}\\) Michael J. Black" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.218, + 0.777, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Max Planck Institute for Intelligent Systems, Tübingen, Germany \\(^{2}\\)ETH Zürich, Switzerland" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.239, + 0.894, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.531, + 0.893, + 0.588 + ], + "angle": 0, + "content": "Figure 1. WANDR starts from an arbitrary body pose and generates precise and realistic human motions that reach a specified 3D goal (depicted as a red sphere). Employing a purely data-driven approach, WANDR is a conditional Variational Autoencoder guided by intention features (depicted arrows) that steer the human's orientation (yellow), position (cyan) and wrist (pink) towards the goal. WANDR is able to reach a wide range of goals even if they deviate significantly from the training data." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.611, + 0.314, + 0.627 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Synthesizing natural human motions that enable a 3D human avatar to walk and reach for arbitrary goals in 3D space remains an unsolved problem with many applications. Existing methods (data-driven or using reinforcement learning) are limited in terms of generalization and motion naturalness. A primary obstacle is the scarcity of training data that combines locomotion with goal reaching. To address this, we introduce WANDR, a data-driven model that takes an avatar's initial pose and a goal's 3D position and generates natural human motions that place the end effector (wrist) on the goal location. To solve this, we introduce novel intention features that drive rich goal-oriented movement. Intention guides the agent to the goal, and interactively adapts the generation to novel situations without needing to define sub-goals or the entire motion path. Crucially, intention allows training on datasets that have goal-oriented motions as well as those that do not. WANDR is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.704 + ], + "angle": 0, + "content": "a conditional Variational Auto-Encoder (c-VAE), which we train using the AMASS and CIRCLE datasets. We evaluate our method extensively and demonstrate its ability to generate natural and long-term motions that reach 3D goals and generalize to unseen goal locations. Our models and code are available for research purposes at wandr.is.tue.mpg.de." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.739, + 0.633, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Goals drive our motions. Even the simplest goal can give rise to intricate motions. Consider reaching for a coffee cup – it can be as straightforward as an arm extension or can involve the coordinated action of our entire body. Actions like bending down, extending our arm, and walking must come together to achieve the goal. At a granular level, we continuously make subtle adjustments to maintain balance and stay on course towards our objective. The result is a fluid motion that seamlessly integrates numerous smaller movements, all" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "converging toward a common and simple goal: placing our hand on the cup. Generating this hierarchy of motions, from the overarching goal to the moment-to-moment individual actions, remains a longstanding challenge in computer vision, graphics, and robotics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.456 + ], + "angle": 0, + "content": "Here we focus on a representative task, illustrated in Fig. 1: given a goal location in space and a starting pose, a humanoid agent must place an end effector (wrist joint) on the goal location while moving in a natural human-like way. To solve the task, the agent needs to be able to approach the goal, orient itself towards it, and reach out such that its wrist makes contact with the goal. Our primary emphasis is on ensuring autonomy for human agents. Consequently, we strive to minimize the guidance information provided, limiting it only to the human's initial pose and the goal's position. Diverging from prior data-driven approaches [4, 17], we choose to refrain from evaluating the model solely on limited labeled data. Instead, we devise an evaluation pipeline that requires agents to reach goals positioned in diverse locations around them. Considering the arbitrary selection of the goal during evaluation, and the minimal guidance information provided, tackling this task is challenging, demanding an approach with the capacity to generalize beyond the distribution of the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.7 + ], + "angle": 0, + "content": "Existing methods approach this problem either using reinforcement learning (RL) [11, 15, 24, 43] or by capturing task-specific datasets [4, 8, 32]. While RL provides a principled way to explore the solution space, it comes with considerable shortcomings. The \"trial and error\" of exploratory learning, in combination with the high dimensionality of human motion result in policies requiring an enormous amount of training even to achieve simple tasks such as walking to a waypoint [5, 43]. In addition, since motion naturalness is better captured by data and not reward functions, RL approaches tend to produce motions that lack naturalness and expressiveness. Data-driven approaches on the other hand, rely on plentiful training motions that are acquired through motion capture and carefully curated for the downstream tasks [4, 10]. Such approaches do not scale and do not generalize well to out-of-distribution tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.868 + ], + "angle": 0, + "content": "In prioritizing both motion realism and training efficiency, we adopt a data-driven approach. However, current data-driven methods lack the ability to learn both from smaller datasets that provide high-quality human reaching motions with goal labels, and from unlabeled larger scale datasets that contain necessary motion skills such as navigating to a goal position. This raises two key challenges. First, how do we model human motion in a way that generated motions can combine skills from different datasets? Second, what should the training objective be in the cases where goal labels are absent?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To address these challenges, we propose WANDR (Wrist-driven Autonomous Navigation for Data-based goal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Reaching). We observe that by modeling human motion generation as an autoregressive stochastic process that produces motions frame by frame, WANDR is able to combine pieces of different dataset distributions when generating a motion sequence. Each generation step is conditioned on goal-related information that we call intention (visualized arrows in Fig. 1). We carefully design intention in a way that strikes a balance between being informative enough to guide the avatar to reach the goal, while also being abstract enough to promote generalization to unseen goals. This allows our generated motions to reach goals that were never encountered during the training phase in a completely zero-shot evaluation scenario. By generating the motion in an autoregressive way, we disentangle the spatial and temporal dimensions of motion. This is necessary as it allows our model to generate novel long-term sequences while being realistic in terms of local dynamic details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.578 + ], + "angle": 0, + "content": "In more detail, our method is based on a conditional Variational Auto-Encoder (c-VAE) that learns to model motion as a frame-by-frame generation process by auto-encoding the pose difference between two adjacent frames. The condition signal consists of the human's current pose and dynamics along with the intention information. Intention is a function of both the current pose and the goal location and therefore actively guides the avatar during the motion generation in a closed loop manner. Through training, the c-VAE learns the distribution of potential subsequent poses conditioned on the current dynamic state of the human and its intention towards a specific goal. We train WANDR using two datasets: AMASS [18], which captures a wide range of motions including locomotion, and CIRCLE [4], which captures reaching motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Although AMASS is large, it lacks any explicit label of goals or intentions. To address this, inspired by the Hind-sight Experience Replay paradigm in robotics [3], we define intention using a hallucinated goal derived from the ground-truth wrist position in a future frame. This approach allows us to establish a unified training objective spanning AMASS and CIRCLE. Consequently, our model learns to combine motions from both datasets, enabling it to effectively reach arbitrary goals during testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In summary, we present WANDR, a data-driven method that combines an autoregressive motion prior with a novel intention guiding mechanism and is able to generate avatars that realistically move in space and reach arbitrary goals. We experimentally evaluate our approach, including the benefit of combining multiple datasets as well as the generalization capabilities of our motion generator. Our results underscore the efficacy of the intention mechanism as an elegant way of guiding the motion generation process while also enabling the incorporation of pseudo goal labels for datasets lacking explicit goal annotations. The model and code are available for research purposes." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "928" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.471, + 0.268 + ], + "angle": 0, + "content": "Early research in motion generation focuses on tasks like motion prediction [1, 2, 7, 9, 19, 27] and unconstrained motion generation [6, 16, 21, 23, 28, 29, 36-38, 40]. More recently, significant effort has been devoted into improving controllability, with a focus on motion generation conditioned on different types of goals [25, 39], enabling interactions with scenes [12, 13, 20] and objects [31, 33, 42, 45]. Methods that attempt goal-driven motion generation can be broadly divided into reinforcement learning or data-driven approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.281, + 0.466, + 0.297 + ], + "angle": 0, + "content": "2.1. Reinforcement Learning for Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.76 + ], + "angle": 0, + "content": "Many existing works employ Reinforcement Learning (RL) for the generation of task-specific long motion sequences. Representative work includes MotionVAE [15] and AMP [24]. MotionVAE [15] employs a two-step process where it initially leverages an autoregressive conditional Variational Autoencoder (VAE) to construct a latent space that encapsulates possible human movements. Subsequently, it utilizes RL to sample from this action space to reach a designated target location while avoiding obstacles by monitoring the area ahead. Similar to MotionVAE, GAMMA [43] learns a policy to extract samples from a latent space and then employs a tree-based search algorithm to find viable motions that steer clear of obstacles by considering the environmental geometric constraints. DIMOS [44] further extends the GAMMA framework by introducing two specialized policy networks: one for locomotion and one for interaction. Together these networks generate goal-conditioned motion sequences that dynamically interact with objects and the environment. AMP [24] learns an adversarial motion prior from unstructured datasets and then applies goal-conditioned RL. This approach involves the formulation of a style reward to encourage the resemblance of the generated sequences to those in the dataset, complemented by a task-specific reward aimed at achieving a particular objective. Hassan et al. [11] extend AMP to produce motions that facilitate interactions with the scene, by conditioning both the discriminator and the policy network on the scene context. However, RL requires significant computation and struggles to generate natural and expressive motion sequences." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.466, + 0.788 + ], + "angle": 0, + "content": "2.2. Data-driven Approaches for Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Most data-driven approaches use existing motion capture (MoCap) datasets [18, 35] to train their models through supervised learning. The pioneering Neural State Machine method [30] is a data-driven technique for generating motion with character-scene interactions, focusing on scenarios with a limited number of objects and interactions. HuMoR [26] proposes a robust model for 3D human shape and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.41 + ], + "angle": 0, + "content": "temporal pose estimation, yet it falls short of generating motions that are conditioned on specific goals. The SAMP [10] method, designed for real-time stochastic motion synthesis, generates diverse human-scene interaction movements by breaking down the process into predicting goals, planning paths, and generating motion along a predefined route. The GOAL [32] method, trained on the GRAB dataset [31], produces motion sequences in which humans walk towards and grasp 3D objects. However, the generated motions exhibit minimal movements, especially in the feet. To address these constraints, the newly introduced CIRCLE [4] dataset provides a collection of reaching motion data. This dataset is used to train a neural network that generates diverse scene-aware reaching motions. Recently, diffusion models have seen the most success at generating motions conditioned on textural input [34, 41] and spatial data [14]. This advancement enables the synthesized motion to accurately reach specified target locations or navigate around obstacles. Nevertheless, the effectiveness of data-driven approaches is constrained by the amount of training data and they lack generalization to out-of-distribution scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.45, + 0.591, + 0.466 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.895, + 0.802 + ], + "angle": 0, + "content": "Our goal is to have a virtual human that can autonomously and realistically move from an initial pose to an arbitrary goal position and accurately place its right hand on the target. This challenge requires a nuanced understanding of human motion and the intricate dynamics involved in goal-oriented motions. For example, when the human tries to reach a distant goal, the motions are mostly focused on the legs and navigating the body to approach the object, but when it gets close to the object, the focus will be on moving the arms and upper body to reach the target. Using these observations, we develop a method named WANDR, which, although trained in a supervised setting on motion capture data, exhibits generalization in reaching unseen goal locations during test time. WANDR is designed to generate human motion in an autoregressive frame-by-frame fashion, conditioned on novel intention features. During training, these features are extracted by picking a future frame as the goal for the wrist. During inference, the intention features are dynamically computed based on the goal's position in a feedback loop, guiding the virtual human to reach the goal. See Fig. 2 for the network overview." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this section, we first consider the different distributions of the datasets we will be using (section 3.1). Following this, we detail the components of the intention features and how they are computed during both the training and inference phases (Section 3.2). Finally, we define the motion representation and motion generator network (Section 3.3)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.086, + 0.855, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Figure 2. WANDR architecture. During training, our model conditions on the intention vectors \\( I^p \\), \\( I^r \\) and \\( I^w \\), learning to associate them with actions that result into reaching goals realistically. When the training data has no defined goal, we create a goal based on the wrist location in future frames; see Sec. 3.2. The state of the avatar, \\( p_i^{dyn} \\) expresses the SMPL-X local pose parameters \\( p_i \\), as well as the deltas \\( d_{i-1} \\) the body parameters have in frame \\( i-1 \\). During inference, WANDR takes the intention features, the state, and random noise and returns the change in pose, \\( \\hat{d}_i \\). The next pose, \\( \\hat{p}_i \\) is obtained by integrating the \\( \\hat{d}_i \\) with the previous pose \\( \\hat{p}_{i-1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.387, + 0.273, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.387, + 0.457, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.471, + 0.594 + ], + "angle": 0, + "content": "Figure 3. In training, if goals are not specified, they are determined by the future wrist location at a randomly selected future timestep, compensating for the lack of paired ground-truth data in AMASS and direct human motion through intention vectors. During inference, target locations are used as goals with intention vectors calculated based on these specific locations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.344, + 0.669 + ], + "angle": 0, + "content": "3.1. Two Complementing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.903 + ], + "angle": 0, + "content": "For the development of WANDR, we use two key datasets: AMASS and CIRCLE. AMASS is a large-scale dataset that offers a broad range of general human motions but lacks a specific focus on goal-reaching tasks. Its diverse collection of movements provides a solid base for understanding human locomotion and body movement when the person is far away from the goal location. In contrast to AMASS, CIRCLE is tailored towards movements involving reaching specific target positions, particularly capturing the nuances of upper body and arm movements. By integrating AMASS's general motion diversity with CIRCLE's targeted goal-reaching data, we equip WANDR with the ability to generate the entire process of reaching a distant goal, from the initial navigation to the precise target-reaching actions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.68, + 0.404 + ], + "angle": 0, + "content": "3.2. Intention Features" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.893, + 0.488 + ], + "angle": 0, + "content": "We represent 3D human motion as a sequence of SMPL-X [22] body poses \\(\\mathbf{p} = \\{p_1,\\dots,p_N\\}\\). Each pose \\(p_i\\in \\mathbb{R}^{135}\\) consists of three concatenated components: the body's translation \\(t_i\\in \\mathbb{R}^3\\), root orientation \\(r_i\\in \\mathbb{R}^6\\), and the body pose \\(\\theta_{i}\\in \\mathbb{R}^{21\\times 6}\\), both in 6D format [46]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.563 + ], + "angle": 0, + "content": "For the avatar to reach the goal, it is important to be informed about the spatial relation of the goal location with respect to its current pose, as well as to have a sense of time to reach the goal promptly. We achieve this by introducing the intention features, which are central to our approach." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.893, + 0.729 + ], + "angle": 0, + "content": "To define the intention features \\( I_{i} \\) at timestep \\( i \\), it is crucial to first establish the selection criteria for the goal \\( G \\in \\mathbb{R}^3 \\), within a motion sequence. Our training involves both label-available scenarios (e.g., CIRCLE dataset) and label-absent scenarios (e.g., AMASS dataset). In label-available cases, where the goal position \\( G \\) and the frame index \\( t_{G} \\) which it is reached are known, calculating \\( I_{i} \\) is straightforward. Conversely, for label-absent scenarios like in AMASS, we pick a random future frame as the \\( t_{G} \\) and define the human's right wrist joint location as the goal \\( G \\) (Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.73, + 0.88, + 0.743 + ], + "angle": 0, + "content": "In both scenarios, the intention features are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.756, + 0.766, + 0.772 + ], + "angle": 0, + "content": "\\[\nI _ {i} = I _ {i} \\left(p _ {i}, G, t _ {G}, i\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.783, + 0.892, + 0.874 + ], + "angle": 0, + "content": "These features are essentially a function of the current body pose, the goal position, and time, offering both spatial and temporal insights required for the motion to reach the goal. They are designed to provide sufficient information to reach goals while also enabling test-time generalization. We define them as three distinct components:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.885, + 0.757, + 0.903 + ], + "angle": 0, + "content": "\\[\nI _ {i} = \\left(I _ {i} ^ {w}, I _ {i} ^ {r}, I _ {i} ^ {p}\\right).\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "930" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "These components represent wrist intention, body orientation intention, and pelvis intention, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.182 + ], + "angle": 0, + "content": "Wrist Intention: This is the main time-dependent component that guides the wrist to reach the goal. It is calculated as the necessary average velocity for the wrist to be at the goal location in time, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.193, + 0.326, + 0.226 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {w} = \\frac {G - W _ {i}}{t _ {G} - i}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.468, + 0.326 + ], + "angle": 0, + "content": "where \\( t_G \\) is the frame when the goal should be reached. During training, we know the goal frame and the time to reach it. At test time, we know the goal location but need an externally-supplied time to reach it. An emergent behavior of this formulation is that, at inference, the model is able to adjust its movement speed and reach the goal just in time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.468, + 0.506 + ], + "angle": 0, + "content": "Orientation Intention: This component captures the body orientation when reaching the goal location. By conditioning on this, we ensure that the human model orients towards the goal and smoothly navigates towards it, preventing unnatural motions during inference, like walking backward. During training, this is defined as the difference between the forward direction of the current body frame, \\( H_{i}^{xy} \\), and the goal body, \\( H_{tG}^{xy} \\) where \\( xy \\) signifies removing the \\( z \\) component. During inference, since we do not have the goal body, we use the pelvis position \\( P_{i} \\) to calculate the pelvisto-goal direction as the desired orientation. This feature is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.517, + 0.427, + 0.557 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {r} = \\left\\{ \\begin{array}{l l} H _ {t _ {G}} ^ {x y} - H _ {i} ^ {x y} & \\text {d u r i n g t r a i n i n g} \\\\ (G - P _ {i}) ^ {x y} - H _ {i} ^ {x y} & \\text {d u r i n g i n f e r e n c e}. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.719 + ], + "angle": 0, + "content": "Pelvis Intention: This feature captures information about the position of the goal relative to the body. It is the difference between the goal and the pelvis joint, excluding the z (height) component. Following the approach in [32], we scale this distance by an exponential function that saturates this vector to have a maximum norm of 2. This formulation helps the method generalize to navigating towards the goal during longer motions and helps the model learn since the distance from the goal does not grow indefinitely in extreme scenarios. This intention is defined by the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.731, + 0.437, + 0.766 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {p} = 2 \\times \\left(1 - e ^ {\\left\\| G ^ {x y} - P _ {i} ^ {x y} \\right\\| _ {2}}\\right) \\times \\frac {G ^ {x y} - P _ {i} ^ {x y}}{\\left\\| G ^ {x y} - P _ {i} ^ {x y} \\right\\| _ {2}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.468, + 0.835 + ], + "angle": 0, + "content": "In the experimental section, we delve into the significance of each of these intention features and discuss the rationale behind our design choices, illustrating their impact on the effectiveness of our model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.325, + 0.862 + ], + "angle": 0, + "content": "3.3. Motion Network (WANDR)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "WANDR is designed as a conditional Variational AutoEncoder (c-VAE) network, operating in an autoregressive" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "manner to generate sequential motion frames. This framework is pivotal in predicting the subsequent pose in a motion sequence, emphasizing an incremental, frame-by-frame approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Central to our approach is the training of the c-VAE to autoencode pose deltas, denoted as \\( d_{i} \\in \\mathbb{R}^{135} \\). These deltas represent the difference between two consecutive poses, \\( p_{i} \\) and \\( p_{i - 1} \\). By focusing on pose deltas rather than absolute pose values, our model benefits from an important inductive bias, enhancing its learning efficiency and performance, as supported by prior research [15, 26]. We separate rotational differences into: body orientation \\( (d_i^r) \\) and body pose \\( (d_i^\\theta) \\), each expressed in a 6-D rotational format. Translation deltas are denoted as \\( d_{i}^{t} = t_{i} - t_{i - 1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.44 + ], + "angle": 0, + "content": "To enhance the motion representation's invariance, we remove information related to the global \\(z\\)-orientation. This is accomplished by subtracting the global \\(z\\) Euler angle of \\(r_{i-1}\\) from both the translational \\((d_i^t)\\) and rotational \\((d_i^r)\\) deltas. The resulting deltas, \\(d_i^{t-z}\\) for translation and \\(d_i^{r-z}\\) for orientation, provide a more robust and consistent representation of motion, irrespective of global direction. Consequently, the delta pose features for any given frame \\(i\\) are composed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.451, + 0.77, + 0.47 + ], + "angle": 0, + "content": "\\[\nd _ {i} = \\left(d _ {i} ^ {t - z}, d _ {i} ^ {r - z}, d _ {i} ^ {\\theta}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.586 + ], + "angle": 0, + "content": "An advantage of this representation is its consistency across different motion global orientations. For instance, in the scenario of a person walking, the delta representation remains agnostic to the walking direction. This attribute underscores the efficacy of our method in capturing the essence of motion without being biased towards any specific orientation or direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Condition Inputs: For each motion frame, the decoder is conditioned on a combination of state and intention features. Specifically, this condition signal is formulated as \\( c_{i} = (p_{i}^{dyn}, I_{i}) \\). The state features, \\( p_{i}^{dyn} \\), encapsulate the avatar's current local pose, focusing on the z-component of translation and a modified orientation that excludes the global z Euler angle, along with the pose deltas \\( di - 1 \\) of the previous step. That combination ensures that the generated motion at each step is informed by both the local pose configuration of the avatar, its dynamics and its directional intention towards the set goal, vital for producing realistic, goal-oriented human motions. An overview of the network architecture is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.794, + 0.658, + 0.809 + ], + "angle": 0, + "content": "3.4. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.89, + 0.845 + ], + "angle": 0, + "content": "Our training objective is a composite of three distinct loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.848, + 0.783, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r e c} + \\alpha \\mathcal {L} _ {K L} + \\mathcal {L} _ {J}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.89, + 0.901 + ], + "angle": 0, + "content": "The reconstruction loss, \\(\\mathcal{L}_{rec}\\), measures the accuracy of the motion reconstruction, quantified as the mean square er" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.509, + 0.957 + ], + "angle": 0, + "content": "931" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "ror (MSE) between the input pose delta, \\( d_{i} \\), and its reconstructed counterpart, \\( \\hat{d}_{i} \\). It ensures the network's ability to faithfully replicate the input motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.182 + ], + "angle": 0, + "content": "The KL Divergence Loss, \\((\\mathcal{L}_{KL})\\), evaluates the deviation of the encoded distribution from a standard normal distribution. It is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.198, + 0.387, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} = \\mathcal {K L} (\\mathcal {N} (0, I) | | \\mathcal {N} (\\mu_ {i}, \\sigma_ {i})).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.469, + 0.303 + ], + "angle": 0, + "content": "Here, \\(\\mu_{i}\\) and \\(\\sigma_{i}\\) represent the mean and variance of the Gaussian distribution predicted by the encoder. We balance this term with \\(\\alpha = 10^{-2}\\) to prevent the over-dominance of \\(\\mathcal{L}_{KL}\\), thereby aiding the decoder in avoiding collapse to mean predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.469, + 0.425 + ], + "angle": 0, + "content": "Finally, we use a Joint Error Loss \\((\\mathcal{L}_J)\\), to ensure perceptual accuracy, by integrating the predicted \\(\\hat{d}_i\\) to get the predicted next pose \\(\\hat{p}_i\\), which is then fed into the SMPL-X model to obtain the predicted joint positions, \\(\\hat{J}\\). The loss \\(\\mathcal{L}_J\\) is the MSE between these predicted joints \\(\\hat{J}\\) and the ground truth joints \\(J\\), addressing errors that might not be apparent in parameter space but are perceptually significant, such as incorrect body orientation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.469, + 0.502 + ], + "angle": 0, + "content": "Notably, our approach does not incorporate any explicit loss functions directly related to reaching a goal. This omission is a deliberate choice, aligning with our method's emphasis on generalizing to diverse goal-reaching scenarios without being constrained by goal-specific training losses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.262, + 0.529 + ], + "angle": 0, + "content": "3.5. Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.629 + ], + "angle": 0, + "content": "In the inference phase of WANDR, our primary objective is to generate human motion that is driven towards a specific goal. Using the decoder of the WANDR c-VAE, we iteratively generate and integrate pose deltas. This process is initiated from the starting pose and progressively builds upon each subsequent pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.469, + 0.705 + ], + "angle": 0, + "content": "The intention features are recalculated at each step based on the current predicted pose and the goal location. They serve as a guiding mechanism, ensuring that the generated motion is consistently oriented towards placing the human's right wrist on the target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.469, + 0.767 + ], + "angle": 0, + "content": "The user can control the motion's pace by specifying the time to reach the goal \\( t_{G} \\). This directly affects the wrist intention feature, enabling adjustments from fast to slow motions to suit various scenarios and constraints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.209, + 0.801 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this section, we outline the datasets used for training and evaluation and benchmark how each dataset affects the goal-reaching ability and the quality of the generated motions. Furthermore, we compare our approach with several baselines and ablate the effect of the different components of our intention vector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.708, + 0.107 + ], + "angle": 0, + "content": "4.1. Datasets & Processing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.282 + ], + "angle": 0, + "content": "Our model is trained on two datasets: AMASS [18] and CIRCLE [4]. AMASS is a collection of 17k sequences, containing a wide range of motion types including long-term navigational skills like walking and turning. CIRCLE, on the other hand, contains 7.2k shorter sequences, each marked with a specific goal reached by a hand. For training, we refine AMASS by excluding sequences where feet are more than \\(20cm\\) above the ground, resulting in a combined dataset of nearly 20k sequences. This dataset is split into \\(80\\%\\) training, \\(10\\%\\) validation, and \\(10\\%\\) test sets. All motions are re-sampled to 30 frames per second (fps)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.293, + 0.689, + 0.309 + ], + "angle": 0, + "content": "4.2. Evaluation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Our evaluation procedure aims at testing the degree which WANDR can generalize to generating reaching motions that start from unseen poses and reach the whole range of 3D space around the starting pose. This is why we choose not to evaluate on held-out motion-goal pairs from the training data. Instead, we only hold out initial poses. During evaluation, starting from these unseen poses, we generate motions that attempt to reach goals that uniformly cover the volume of a cylinder centered on the human, including completely out-of-distribution goal locations (see Sup. Mat. Sec. 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.665 + ], + "angle": 0, + "content": "In particular, the set of evaluation goals is defined in a cylindrical coordinate frame by taking all the combinations of (1) 5 angles equally separating the 360 degrees around the human, (2) 5 different goal heights ranging from 0 to 1.8 meters and (3) 5 distances from 0.5 to 5 meters. We generate motions from 6 different initial poses, with an 8-second duration specified for reaching each goal. Five motions are sampled for each pose-goal combination, resulting in \\(5 \\times 5 \\times 5 \\times 6 \\times 5 = 3750\\) unique motion sequences from which our metrics are computed. This setup allows us to thoroughly test our model in diverse scenarios, including long-term movements, navigational skills, and reaching motions at various heights and distances." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.684, + 0.691 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.891, + 0.775 + ], + "angle": 0, + "content": "To accurately assess the effectiveness of our approach in generating realistic, goal-oriented human motion, we employ a set of metrics focused on both the ability to successfully reach the intended goal and the naturalness of the motion. These metrics are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.891, + 0.84 + ], + "angle": 0, + "content": "- **Success Rate (SR):** This quantifies the percentage of motions where the right wrist reaches within \\(10\\mathrm{cm}\\) of the goal, indicating successful goal attainment. The criterion for success aligns with that used in [4]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Foot Skating (FS): FS evaluates the naturalness of the motion based on foot skating, where a frame is considered as having foot skating if the lowest vertex of the human mesh moves more than \\(0.66 \\mathrm{~cm}\\) between consecutive" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "932" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.283, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.341, + 0.18, + 0.353 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.485, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.341, + 0.395, + 0.353 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.09, + 0.686, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.341, + 0.586, + 0.353 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.09, + 0.892, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.341, + 0.801, + 0.353 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.894, + 0.419 + ], + "angle": 0, + "content": "Figure 4. Diverse motion generated with WANDR: Displaying a range of motions generated by WANDR from various initial poses towards arbitrary goals. Examples include navigating towards goals from initial orientations not facing the goal (a, b, c, d), elevating the right hand to reach higher targets (c), and bending down to access goals near the floor (d), showcasing the model's ability to adapt to novel goal locations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.469, + 0.477 + ], + "angle": 0, + "content": "frames (adjusted from the 1cm threshold used in [4] to accommodate our 30fps motion generation)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.478, + 0.471, + 0.539 + ], + "angle": 0, + "content": "- Distance to Goal (DTG): DTG records the closest distance in cm that the right wrist gets to the goal during a motion. This metric offers a nuanced view of the model's capability to guide the motion towards the goal." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.549, + 0.171, + 0.563 + ], + "angle": 0, + "content": "4.4. Results" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.574, + 0.274, + 0.588 + ], + "angle": 0, + "content": "4.4.1 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Combining AMASS and CIRCLE: Our evaluation in Table 1 validates our hypothesis about the benefits of training with both the AMASS and CIRCLE datasets. On the one hand, training only on CIRCLE (line 1) is not sufficient for the model to learn necessary navigational skills, such as walking, due to the dataset's narrow focus or reaching motions. This is apparent from the very high foot-skating. On the other hand, training only on AMASS (line 2) results in high-quality motion generation with low foot-skating, but a relatively low success rate in goal-reaching. The Distance to Goal (DTG) metric, suggests that the model is able to navigate close to the target, but it lacks the precise movement needed to successfully reach for the goal. Using both datasets (line 3) illustrates how our approach effectively merges the broad motion vocabulary of AMASS with the goal-oriented precision of CIRCLE, leading to both high-quality motion and improved goal-reaching capability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We also compare with GOAL [32], a method that generates human motions that reach and grasp objects. GOAL is trained on GRAB [31], a dataset purely consisting of mo" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.445, + 0.872, + 0.526 + ], + "angle": 0, + "content": "
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR/Circle0%56%205.4
WANDR/AMASS16%19%48.0
WANDR32%16%24.8
GOAL [32]0%29%149.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Table 1. We evaluate WANDR trained on different datasets and compare with GOAL [32]. Training solely on CIRCLE results in unrealistic motions, whereas AMASS excels in motion quality but struggles with finer goal-reaching skills. WANDR, leveraging both of what these datasets offer, demonstrates realistic motions as well as better ability to reach goals compared to baselines and existing methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.656, + 0.892, + 0.732 + ], + "angle": 0, + "content": "tions of humans grasping and manipulating objects. Since the relative positioning of the human and the object as well as the motions in GRAB have very small variations, GOAL does not succeed in any of the evaluation configurations (line 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation of Intention Features: In order to demonstrate the contribution of each component of the intention feature we conduct an ablation study (Table 2). Using only wrist intention (line 1) results in the lowest foot skating, due to the minimal constraints applied to the motion, allowing for more adaptable motion planning. But wrist intention features, are time-dependent and do not carry information about the absolute distance to the goal. This is why it can lead to the avatar over- or under-shooting and thus achieving a low success ratio (SR). The addition of pelvis intention (line 2) enables the avatar to sense the distance to the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "933" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "goal while the orientation intention (line 3) properly aligns the body to face the goal. Since pelvis and orientation intention add more constraints to the motion, they can sometimes cause more challenging body dynamics and produce motions with higher foot-skating (FS) (e.g. turning around in place instead of walking in a U-turn). We also try removing the intention from the motion prior and optimizing the latent space of a randomly generated initial motion to minimize the distance between the wrist and the goal (VAE + opt). This approach fails since the result is heavily dependent on the initialization of the latent variables of the motion. Our results confirm that each component of the intention feature is essential to achieving the overall performance of the model. For more details on the optimization see Sup. Mat." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.429 + ], + "angle": 0, + "content": "
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR (Iw)15%13%62.9
WANDR (Iw + Ip)18%17%44.9
WANDR (Iw + Ir)19%19%36.0
WANDR (full intention)32%16%24.8
VAE + opt3%4%217.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Table 2. Ablation Study. We evaluate the impact of each component of the intention vector. We also compare with an optimization baseline that does not use any condition signals. The results highlight the effectiveness of all of the components of intention as well as the fact that the complexity of the task makes \"brute-forcing\" with optimization unsuccessful." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.473, + 0.722 + ], + "angle": 0, + "content": "Success Ratio Distribution: Our decomposition of the model's success ratio, presented in Fig. 5, offers insights into how the model's performance varies with respect to different goal positions. WANDR demonstrates a consistent ability to reach goals across various distances (blue) and directions (green). It is more capable at reaching goals that are closer to the natural position of the wrist and do not require extensive bending or stretching (yellow). This trend likely results from the abundance of standing or upright motion sequences in the training data, as opposed to motions involving bending or crouching. This analysis provides valuable information for future improvements and dataset balancing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.265, + 0.756 + ], + "angle": 0, + "content": "4.4.2 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Fig. 4, we show a variety of motion sequences generated with our network featuring reaching goals located at varying distances and heights, highlighting the model's ability to realistically and smoothly orient, navigate, and reach for goals. These goals require actions such as bending down, turning, or stretching upwards. A critical aspect observed is the model's ability to decelerate as it approaches the goal, seamlessly coordinating body and arm movements to achieve a natural-looking reaching motion. Overall, the" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.087, + 0.895, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.232, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Figure 5. We show the success rates of reaching goals at various heights, angles, and distances from the initial human pose. It highlights how goal position affects the model in accurately navigating and achieving the goals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.895, + 0.361 + ], + "angle": 0, + "content": "qualitative results show that our network generalizes well to novel goal locations while generating realistic motions. For more results please see Sup. Mat. and the video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.373, + 0.619, + 0.388 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.579 + ], + "angle": 0, + "content": "In conclusion, our research presents a novel data-driven approach to human motion generation, focusing on the task of reaching arbitrary goals in space. We introduce novel intention features that enable learning both general navigational skills from AMASS and goal-reaching skills from the CIRCLE dataset under the same distribution. We evaluate our model's ability to reach unseen goals that cover the whole space an avatar should be able to reach around it. The autoregressive design of WANDR demonstrates generalizability in generating realistic human motions that reach unseen goals without requiring any extra guidance information such as a pre-defined trajectory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Limitations and Future Work: Our approach is not without its limitations. Currently, error accumulation can sometimes bring the avatar to states where it can no longer recover. Additionally, our model shows less proficiency in reaching extremely low or high goals, reflecting a need for more diverse training data encompassing a wider range of body movements. Future work could focus on incorporating realistic grasping mechanisms and interactions with objects, as well as including scene navigation capabilities. This could involve integrating more complex datasets or developing advanced algorithms capable of understanding and interacting with varied environmental contexts, thereby pushing the boundaries of realistic human motion simulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Acknowledgments: Markos Diomataris was supported in part by the Max Planck ETH Center for Learning Systems. We thank P. Ghosh, O. Ben Dov S. Tripathi, for the fruitful discussions and A. Cseke, T. Niewiadomski, T. Alexiadis, T. McConnell, for conducting the user studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.867, + 0.648, + 0.88 + ], + "angle": 0, + "content": "Conflicts of Interest:" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.882, + 0.866, + 0.897 + ], + "angle": 0, + "content": "https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "934" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pages 565-574. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5223-5232, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[3] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. In Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[5] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In International Conference on 3D Vision (3DV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3D human motion synthesis model via conditional variational auto-encoder. In Computer Vision and Pattern Recognition (CVPR), pages 11645–11655, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision (ECCV), pages 387-404. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[9] Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 2017 International Conference on 3D Vision (3DV), pages 458-466. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[10] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. arXiv:2108.08284, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[11] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. arXiv:1904.05767, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[13] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[14] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[15] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion VAEs. Transactions on Graphics (TOG), 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[16] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. Transactions on Graphics (TOG), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[17] Matthew Loper, Naureen Mahmood, and Michael J Black. Mosh: motion and shape capture from sparse markers. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 33(6):220-1, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In International Conference on 3D Vision (3DV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[21] Dirk Ormoneit, Hedvig Sidenbladh, Michael Black, and Trevor Hastie. Learning and tracking cyclic human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[22] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[23] Vladimir Pavlovic, James M Rehg, and John MacCormick. Learning switching linear models of human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[24] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: adversarial motion priors for stylized physics-based character control. Transactions on Graphics (TOG), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[25] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "935" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44 (6):3300-3315, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[28] Hedvig Sidenbladh, Michael J Black, and David J Fleet. Stochastic tracking of 3d human figures using 2d image motion. In European Conference on Computer Vision (ECCV). Springer, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[29] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision (IJCV), 87 (1-2):4-27, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[30] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. Transactions on Graphics (TOG), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[31] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[32] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[33] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. Grip: Generating interaction poses using latent consistency and spatial cues, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[34] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[35] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In British Machine Vision Conference (BMVC), pages 1-13, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[36] Raquel Urtasun, David J Fleet, and Pascal Fua. Temporal motion models for monocular and multiview 3d human body tracking. Computer Vision and Image Understanding (CVIU), 104(2-3):157-177, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[37] Qi Wang, Thierry Artières, Mickael Chen, and Ludovic Denoyer. Adversarial learning for modeling human motion. The Visual Computer, 36(1):141-160, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.79 + ], + "angle": 0, + "content": "[38] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI conference on artificial intelligence, pages 12281-12288, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[39] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In International Conference on Computer Vision (ICCV), pages 2228-2238, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[40] Ping Yu, Yang Zhao, Chunyuan Li, Junsong Yuan, and Changyou Chen. Structure-aware human-action generation." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.893, + 0.135 + ], + "angle": 0, + "content": "In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 18-34. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[41] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.259 + ], + "angle": 0, + "content": "[42] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision, pages 518-535. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.893, + 0.304 + ], + "angle": 0, + "content": "[43] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[44] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[45] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision (ECCV), pages 676-694. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.513, + 0.957 + ], + "angle": 0, + "content": "936" + } + ] +] \ No newline at end of file diff --git a/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_origin.pdf b/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fcebe15df21555e17098b46df67d12fe2f79fb6b --- /dev/null +++ b/2024/WANDR_ Intention-guided Human Motion Generation/f6756573-8e2b-41a5-9d35-381ea4ebb252_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44d4708b697c560c430e3336b97f44457383f2f39650782a614b7c5fee87439c +size 1119658 diff --git a/2024/WANDR_ Intention-guided Human Motion Generation/full.md b/2024/WANDR_ Intention-guided Human Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae0806bdb0bdfc909c23b3953f441267d41429b9 --- /dev/null +++ b/2024/WANDR_ Intention-guided Human Motion Generation/full.md @@ -0,0 +1,299 @@ +# WANDR: Intention-guided Human Motion Generation + +Markos Diomataris $^{1,2}$ Nikos Athanasiou $^{1}$ Omid Taheri $^{1}$ Xi Wang $^{2}$ Otmar Hilliges $^{2}$ Michael J. Black + +$^{1}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany $^{2}$ ETH Zürich, Switzerland + +![](images/6f6ffb3bc9c796b2db9a211351ebbea7733ac7f0758be8b079cfcc56c28c80de.jpg) +Figure 1. WANDR starts from an arbitrary body pose and generates precise and realistic human motions that reach a specified 3D goal (depicted as a red sphere). Employing a purely data-driven approach, WANDR is a conditional Variational Autoencoder guided by intention features (depicted arrows) that steer the human's orientation (yellow), position (cyan) and wrist (pink) towards the goal. WANDR is able to reach a wide range of goals even if they deviate significantly from the training data. + +# Abstract + +Synthesizing natural human motions that enable a 3D human avatar to walk and reach for arbitrary goals in 3D space remains an unsolved problem with many applications. Existing methods (data-driven or using reinforcement learning) are limited in terms of generalization and motion naturalness. A primary obstacle is the scarcity of training data that combines locomotion with goal reaching. To address this, we introduce WANDR, a data-driven model that takes an avatar's initial pose and a goal's 3D position and generates natural human motions that place the end effector (wrist) on the goal location. To solve this, we introduce novel intention features that drive rich goal-oriented movement. Intention guides the agent to the goal, and interactively adapts the generation to novel situations without needing to define sub-goals or the entire motion path. Crucially, intention allows training on datasets that have goal-oriented motions as well as those that do not. WANDR is + +a conditional Variational Auto-Encoder (c-VAE), which we train using the AMASS and CIRCLE datasets. We evaluate our method extensively and demonstrate its ability to generate natural and long-term motions that reach 3D goals and generalize to unseen goal locations. Our models and code are available for research purposes at wandr.is.tue.mpg.de. + +# 1. Introduction + +Goals drive our motions. Even the simplest goal can give rise to intricate motions. Consider reaching for a coffee cup – it can be as straightforward as an arm extension or can involve the coordinated action of our entire body. Actions like bending down, extending our arm, and walking must come together to achieve the goal. At a granular level, we continuously make subtle adjustments to maintain balance and stay on course towards our objective. The result is a fluid motion that seamlessly integrates numerous smaller movements, all + +converging toward a common and simple goal: placing our hand on the cup. Generating this hierarchy of motions, from the overarching goal to the moment-to-moment individual actions, remains a longstanding challenge in computer vision, graphics, and robotics. + +Here we focus on a representative task, illustrated in Fig. 1: given a goal location in space and a starting pose, a humanoid agent must place an end effector (wrist joint) on the goal location while moving in a natural human-like way. To solve the task, the agent needs to be able to approach the goal, orient itself towards it, and reach out such that its wrist makes contact with the goal. Our primary emphasis is on ensuring autonomy for human agents. Consequently, we strive to minimize the guidance information provided, limiting it only to the human's initial pose and the goal's position. Diverging from prior data-driven approaches [4, 17], we choose to refrain from evaluating the model solely on limited labeled data. Instead, we devise an evaluation pipeline that requires agents to reach goals positioned in diverse locations around them. Considering the arbitrary selection of the goal during evaluation, and the minimal guidance information provided, tackling this task is challenging, demanding an approach with the capacity to generalize beyond the distribution of the training dataset. + +Existing methods approach this problem either using reinforcement learning (RL) [11, 15, 24, 43] or by capturing task-specific datasets [4, 8, 32]. While RL provides a principled way to explore the solution space, it comes with considerable shortcomings. The "trial and error" of exploratory learning, in combination with the high dimensionality of human motion result in policies requiring an enormous amount of training even to achieve simple tasks such as walking to a waypoint [5, 43]. In addition, since motion naturalness is better captured by data and not reward functions, RL approaches tend to produce motions that lack naturalness and expressiveness. Data-driven approaches on the other hand, rely on plentiful training motions that are acquired through motion capture and carefully curated for the downstream tasks [4, 10]. Such approaches do not scale and do not generalize well to out-of-distribution tasks. + +In prioritizing both motion realism and training efficiency, we adopt a data-driven approach. However, current data-driven methods lack the ability to learn both from smaller datasets that provide high-quality human reaching motions with goal labels, and from unlabeled larger scale datasets that contain necessary motion skills such as navigating to a goal position. This raises two key challenges. First, how do we model human motion in a way that generated motions can combine skills from different datasets? Second, what should the training objective be in the cases where goal labels are absent? + +To address these challenges, we propose WANDR (Wrist-driven Autonomous Navigation for Data-based goal + +Reaching). We observe that by modeling human motion generation as an autoregressive stochastic process that produces motions frame by frame, WANDR is able to combine pieces of different dataset distributions when generating a motion sequence. Each generation step is conditioned on goal-related information that we call intention (visualized arrows in Fig. 1). We carefully design intention in a way that strikes a balance between being informative enough to guide the avatar to reach the goal, while also being abstract enough to promote generalization to unseen goals. This allows our generated motions to reach goals that were never encountered during the training phase in a completely zero-shot evaluation scenario. By generating the motion in an autoregressive way, we disentangle the spatial and temporal dimensions of motion. This is necessary as it allows our model to generate novel long-term sequences while being realistic in terms of local dynamic details. + +In more detail, our method is based on a conditional Variational Auto-Encoder (c-VAE) that learns to model motion as a frame-by-frame generation process by auto-encoding the pose difference between two adjacent frames. The condition signal consists of the human's current pose and dynamics along with the intention information. Intention is a function of both the current pose and the goal location and therefore actively guides the avatar during the motion generation in a closed loop manner. Through training, the c-VAE learns the distribution of potential subsequent poses conditioned on the current dynamic state of the human and its intention towards a specific goal. We train WANDR using two datasets: AMASS [18], which captures a wide range of motions including locomotion, and CIRCLE [4], which captures reaching motions. + +Although AMASS is large, it lacks any explicit label of goals or intentions. To address this, inspired by the Hind-sight Experience Replay paradigm in robotics [3], we define intention using a hallucinated goal derived from the ground-truth wrist position in a future frame. This approach allows us to establish a unified training objective spanning AMASS and CIRCLE. Consequently, our model learns to combine motions from both datasets, enabling it to effectively reach arbitrary goals during testing. + +In summary, we present WANDR, a data-driven method that combines an autoregressive motion prior with a novel intention guiding mechanism and is able to generate avatars that realistically move in space and reach arbitrary goals. We experimentally evaluate our approach, including the benefit of combining multiple datasets as well as the generalization capabilities of our motion generator. Our results underscore the efficacy of the intention mechanism as an elegant way of guiding the motion generation process while also enabling the incorporation of pseudo goal labels for datasets lacking explicit goal annotations. The model and code are available for research purposes. + +# 2. Related Work + +Early research in motion generation focuses on tasks like motion prediction [1, 2, 7, 9, 19, 27] and unconstrained motion generation [6, 16, 21, 23, 28, 29, 36-38, 40]. More recently, significant effort has been devoted into improving controllability, with a focus on motion generation conditioned on different types of goals [25, 39], enabling interactions with scenes [12, 13, 20] and objects [31, 33, 42, 45]. Methods that attempt goal-driven motion generation can be broadly divided into reinforcement learning or data-driven approaches. + +# 2.1. Reinforcement Learning for Motion Synthesis + +Many existing works employ Reinforcement Learning (RL) for the generation of task-specific long motion sequences. Representative work includes MotionVAE [15] and AMP [24]. MotionVAE [15] employs a two-step process where it initially leverages an autoregressive conditional Variational Autoencoder (VAE) to construct a latent space that encapsulates possible human movements. Subsequently, it utilizes RL to sample from this action space to reach a designated target location while avoiding obstacles by monitoring the area ahead. Similar to MotionVAE, GAMMA [43] learns a policy to extract samples from a latent space and then employs a tree-based search algorithm to find viable motions that steer clear of obstacles by considering the environmental geometric constraints. DIMOS [44] further extends the GAMMA framework by introducing two specialized policy networks: one for locomotion and one for interaction. Together these networks generate goal-conditioned motion sequences that dynamically interact with objects and the environment. AMP [24] learns an adversarial motion prior from unstructured datasets and then applies goal-conditioned RL. This approach involves the formulation of a style reward to encourage the resemblance of the generated sequences to those in the dataset, complemented by a task-specific reward aimed at achieving a particular objective. Hassan et al. [11] extend AMP to produce motions that facilitate interactions with the scene, by conditioning both the discriminator and the policy network on the scene context. However, RL requires significant computation and struggles to generate natural and expressive motion sequences. + +# 2.2. Data-driven Approaches for Motion Synthesis + +Most data-driven approaches use existing motion capture (MoCap) datasets [18, 35] to train their models through supervised learning. The pioneering Neural State Machine method [30] is a data-driven technique for generating motion with character-scene interactions, focusing on scenarios with a limited number of objects and interactions. HuMoR [26] proposes a robust model for 3D human shape and + +temporal pose estimation, yet it falls short of generating motions that are conditioned on specific goals. The SAMP [10] method, designed for real-time stochastic motion synthesis, generates diverse human-scene interaction movements by breaking down the process into predicting goals, planning paths, and generating motion along a predefined route. The GOAL [32] method, trained on the GRAB dataset [31], produces motion sequences in which humans walk towards and grasp 3D objects. However, the generated motions exhibit minimal movements, especially in the feet. To address these constraints, the newly introduced CIRCLE [4] dataset provides a collection of reaching motion data. This dataset is used to train a neural network that generates diverse scene-aware reaching motions. Recently, diffusion models have seen the most success at generating motions conditioned on textural input [34, 41] and spatial data [14]. This advancement enables the synthesized motion to accurately reach specified target locations or navigate around obstacles. Nevertheless, the effectiveness of data-driven approaches is constrained by the amount of training data and they lack generalization to out-of-distribution scenarios. + +# 3. Method + +Our goal is to have a virtual human that can autonomously and realistically move from an initial pose to an arbitrary goal position and accurately place its right hand on the target. This challenge requires a nuanced understanding of human motion and the intricate dynamics involved in goal-oriented motions. For example, when the human tries to reach a distant goal, the motions are mostly focused on the legs and navigating the body to approach the object, but when it gets close to the object, the focus will be on moving the arms and upper body to reach the target. Using these observations, we develop a method named WANDR, which, although trained in a supervised setting on motion capture data, exhibits generalization in reaching unseen goal locations during test time. WANDR is designed to generate human motion in an autoregressive frame-by-frame fashion, conditioned on novel intention features. During training, these features are extracted by picking a future frame as the goal for the wrist. During inference, the intention features are dynamically computed based on the goal's position in a feedback loop, guiding the virtual human to reach the goal. See Fig. 2 for the network overview. + +In this section, we first consider the different distributions of the datasets we will be using (section 3.1). Following this, we detail the components of the intention features and how they are computed during both the training and inference phases (Section 3.2). Finally, we define the motion representation and motion generator network (Section 3.3). + +![](images/40453d55b9cf3fcc28638003879d9578750b23cf414d6914c1f57b4eaa8dbca2.jpg) +Figure 2. WANDR architecture. During training, our model conditions on the intention vectors $I^p$ , $I^r$ and $I^w$ , learning to associate them with actions that result into reaching goals realistically. When the training data has no defined goal, we create a goal based on the wrist location in future frames; see Sec. 3.2. The state of the avatar, $p_i^{dyn}$ expresses the SMPL-X local pose parameters $p_i$ , as well as the deltas $d_{i-1}$ the body parameters have in frame $i-1$ . During inference, WANDR takes the intention features, the state, and random noise and returns the change in pose, $\hat{d}_i$ . The next pose, $\hat{p}_i$ is obtained by integrating the $\hat{d}_i$ with the previous pose $\hat{p}_{i-1}$ . + +![](images/a0993d495849048161452f010dd815479f228bd758fd187ae25d9f04bde8d14c.jpg) +Figure 3. In training, if goals are not specified, they are determined by the future wrist location at a randomly selected future timestep, compensating for the lack of paired ground-truth data in AMASS and direct human motion through intention vectors. During inference, target locations are used as goals with intention vectors calculated based on these specific locations. + +![](images/491de65aee91085c701c2ca5ddf33e417a06f71cd35ccc8e52632abb781c9ed1.jpg) + +# 3.1. Two Complementing Datasets + +For the development of WANDR, we use two key datasets: AMASS and CIRCLE. AMASS is a large-scale dataset that offers a broad range of general human motions but lacks a specific focus on goal-reaching tasks. Its diverse collection of movements provides a solid base for understanding human locomotion and body movement when the person is far away from the goal location. In contrast to AMASS, CIRCLE is tailored towards movements involving reaching specific target positions, particularly capturing the nuances of upper body and arm movements. By integrating AMASS's general motion diversity with CIRCLE's targeted goal-reaching data, we equip WANDR with the ability to generate the entire process of reaching a distant goal, from the initial navigation to the precise target-reaching actions. + +# 3.2. Intention Features + +We represent 3D human motion as a sequence of SMPL-X [22] body poses $\mathbf{p} = \{p_1,\dots,p_N\}$ . Each pose $p_i\in \mathbb{R}^{135}$ consists of three concatenated components: the body's translation $t_i\in \mathbb{R}^3$ , root orientation $r_i\in \mathbb{R}^6$ , and the body pose $\theta_{i}\in \mathbb{R}^{21\times 6}$ , both in 6D format [46]. + +For the avatar to reach the goal, it is important to be informed about the spatial relation of the goal location with respect to its current pose, as well as to have a sense of time to reach the goal promptly. We achieve this by introducing the intention features, which are central to our approach. + +To define the intention features $I_{i}$ at timestep $i$ , it is crucial to first establish the selection criteria for the goal $G \in \mathbb{R}^3$ , within a motion sequence. Our training involves both label-available scenarios (e.g., CIRCLE dataset) and label-absent scenarios (e.g., AMASS dataset). In label-available cases, where the goal position $G$ and the frame index $t_{G}$ which it is reached are known, calculating $I_{i}$ is straightforward. Conversely, for label-absent scenarios like in AMASS, we pick a random future frame as the $t_{G}$ and define the human's right wrist joint location as the goal $G$ (Fig. 3). + +In both scenarios, the intention features are defined as: + +$$ +I _ {i} = I _ {i} \left(p _ {i}, G, t _ {G}, i\right). +$$ + +These features are essentially a function of the current body pose, the goal position, and time, offering both spatial and temporal insights required for the motion to reach the goal. They are designed to provide sufficient information to reach goals while also enabling test-time generalization. We define them as three distinct components: + +$$ +I _ {i} = \left(I _ {i} ^ {w}, I _ {i} ^ {r}, I _ {i} ^ {p}\right). +$$ + +These components represent wrist intention, body orientation intention, and pelvis intention, respectively. + +Wrist Intention: This is the main time-dependent component that guides the wrist to reach the goal. It is calculated as the necessary average velocity for the wrist to be at the goal location in time, defined as: + +$$ +I _ {i} ^ {w} = \frac {G - W _ {i}}{t _ {G} - i} +$$ + +where $t_G$ is the frame when the goal should be reached. During training, we know the goal frame and the time to reach it. At test time, we know the goal location but need an externally-supplied time to reach it. An emergent behavior of this formulation is that, at inference, the model is able to adjust its movement speed and reach the goal just in time. + +Orientation Intention: This component captures the body orientation when reaching the goal location. By conditioning on this, we ensure that the human model orients towards the goal and smoothly navigates towards it, preventing unnatural motions during inference, like walking backward. During training, this is defined as the difference between the forward direction of the current body frame, $H_{i}^{xy}$ , and the goal body, $H_{tG}^{xy}$ where $xy$ signifies removing the $z$ component. During inference, since we do not have the goal body, we use the pelvis position $P_{i}$ to calculate the pelvisto-goal direction as the desired orientation. This feature is formulated as: + +$$ +I _ {i} ^ {r} = \left\{ \begin{array}{l l} H _ {t _ {G}} ^ {x y} - H _ {i} ^ {x y} & \text {d u r i n g t r a i n i n g} \\ (G - P _ {i}) ^ {x y} - H _ {i} ^ {x y} & \text {d u r i n g i n f e r e n c e}. \end{array} \right. +$$ + +Pelvis Intention: This feature captures information about the position of the goal relative to the body. It is the difference between the goal and the pelvis joint, excluding the z (height) component. Following the approach in [32], we scale this distance by an exponential function that saturates this vector to have a maximum norm of 2. This formulation helps the method generalize to navigating towards the goal during longer motions and helps the model learn since the distance from the goal does not grow indefinitely in extreme scenarios. This intention is defined by the equation: + +$$ +I _ {i} ^ {p} = 2 \times \left(1 - e ^ {\left\| G ^ {x y} - P _ {i} ^ {x y} \right\| _ {2}}\right) \times \frac {G ^ {x y} - P _ {i} ^ {x y}}{\left\| G ^ {x y} - P _ {i} ^ {x y} \right\| _ {2}}. +$$ + +In the experimental section, we delve into the significance of each of these intention features and discuss the rationale behind our design choices, illustrating their impact on the effectiveness of our model. + +# 3.3. Motion Network (WANDR) + +WANDR is designed as a conditional Variational AutoEncoder (c-VAE) network, operating in an autoregressive + +manner to generate sequential motion frames. This framework is pivotal in predicting the subsequent pose in a motion sequence, emphasizing an incremental, frame-by-frame approach. + +Central to our approach is the training of the c-VAE to autoencode pose deltas, denoted as $d_{i} \in \mathbb{R}^{135}$ . These deltas represent the difference between two consecutive poses, $p_{i}$ and $p_{i - 1}$ . By focusing on pose deltas rather than absolute pose values, our model benefits from an important inductive bias, enhancing its learning efficiency and performance, as supported by prior research [15, 26]. We separate rotational differences into: body orientation $(d_i^r)$ and body pose $(d_i^\theta)$ , each expressed in a 6-D rotational format. Translation deltas are denoted as $d_{i}^{t} = t_{i} - t_{i - 1}$ . + +To enhance the motion representation's invariance, we remove information related to the global $z$ -orientation. This is accomplished by subtracting the global $z$ Euler angle of $r_{i-1}$ from both the translational $(d_i^t)$ and rotational $(d_i^r)$ deltas. The resulting deltas, $d_i^{t-z}$ for translation and $d_i^{r-z}$ for orientation, provide a more robust and consistent representation of motion, irrespective of global direction. Consequently, the delta pose features for any given frame $i$ are composed as follows: + +$$ +d _ {i} = \left(d _ {i} ^ {t - z}, d _ {i} ^ {r - z}, d _ {i} ^ {\theta}\right). +$$ + +An advantage of this representation is its consistency across different motion global orientations. For instance, in the scenario of a person walking, the delta representation remains agnostic to the walking direction. This attribute underscores the efficacy of our method in capturing the essence of motion without being biased towards any specific orientation or direction. + +Condition Inputs: For each motion frame, the decoder is conditioned on a combination of state and intention features. Specifically, this condition signal is formulated as $c_{i} = (p_{i}^{dyn}, I_{i})$ . The state features, $p_{i}^{dyn}$ , encapsulate the avatar's current local pose, focusing on the z-component of translation and a modified orientation that excludes the global z Euler angle, along with the pose deltas $di - 1$ of the previous step. That combination ensures that the generated motion at each step is informed by both the local pose configuration of the avatar, its dynamics and its directional intention towards the set goal, vital for producing realistic, goal-oriented human motions. An overview of the network architecture is shown in Fig. 2. + +# 3.4. Training Losses + +Our training objective is a composite of three distinct loss functions: + +$$ +\mathcal {L} = \mathcal {L} _ {r e c} + \alpha \mathcal {L} _ {K L} + \mathcal {L} _ {J}. +$$ + +The reconstruction loss, $\mathcal{L}_{rec}$ , measures the accuracy of the motion reconstruction, quantified as the mean square er + +ror (MSE) between the input pose delta, $d_{i}$ , and its reconstructed counterpart, $\hat{d}_{i}$ . It ensures the network's ability to faithfully replicate the input motion. + +The KL Divergence Loss, $(\mathcal{L}_{KL})$ , evaluates the deviation of the encoded distribution from a standard normal distribution. It is formulated as: + +$$ +\mathcal {L} _ {K L} = \mathcal {K L} (\mathcal {N} (0, I) | | \mathcal {N} (\mu_ {i}, \sigma_ {i})). +$$ + +Here, $\mu_{i}$ and $\sigma_{i}$ represent the mean and variance of the Gaussian distribution predicted by the encoder. We balance this term with $\alpha = 10^{-2}$ to prevent the over-dominance of $\mathcal{L}_{KL}$ , thereby aiding the decoder in avoiding collapse to mean predictions. + +Finally, we use a Joint Error Loss $(\mathcal{L}_J)$ , to ensure perceptual accuracy, by integrating the predicted $\hat{d}_i$ to get the predicted next pose $\hat{p}_i$ , which is then fed into the SMPL-X model to obtain the predicted joint positions, $\hat{J}$ . The loss $\mathcal{L}_J$ is the MSE between these predicted joints $\hat{J}$ and the ground truth joints $J$ , addressing errors that might not be apparent in parameter space but are perceptually significant, such as incorrect body orientation. + +Notably, our approach does not incorporate any explicit loss functions directly related to reaching a goal. This omission is a deliberate choice, aligning with our method's emphasis on generalizing to diverse goal-reaching scenarios without being constrained by goal-specific training losses. + +# 3.5. Motion Generation + +In the inference phase of WANDR, our primary objective is to generate human motion that is driven towards a specific goal. Using the decoder of the WANDR c-VAE, we iteratively generate and integrate pose deltas. This process is initiated from the starting pose and progressively builds upon each subsequent pose. + +The intention features are recalculated at each step based on the current predicted pose and the goal location. They serve as a guiding mechanism, ensuring that the generated motion is consistently oriented towards placing the human's right wrist on the target. + +The user can control the motion's pace by specifying the time to reach the goal $t_{G}$ . This directly affects the wrist intention feature, enabling adjustments from fast to slow motions to suit various scenarios and constraints. + +# 4. Experiments + +In this section, we outline the datasets used for training and evaluation and benchmark how each dataset affects the goal-reaching ability and the quality of the generated motions. Furthermore, we compare our approach with several baselines and ablate the effect of the different components of our intention vector. + +# 4.1. Datasets & Processing + +Our model is trained on two datasets: AMASS [18] and CIRCLE [4]. AMASS is a collection of 17k sequences, containing a wide range of motion types including long-term navigational skills like walking and turning. CIRCLE, on the other hand, contains 7.2k shorter sequences, each marked with a specific goal reached by a hand. For training, we refine AMASS by excluding sequences where feet are more than $20cm$ above the ground, resulting in a combined dataset of nearly 20k sequences. This dataset is split into $80\%$ training, $10\%$ validation, and $10\%$ test sets. All motions are re-sampled to 30 frames per second (fps). + +# 4.2. Evaluation Strategy + +Our evaluation procedure aims at testing the degree which WANDR can generalize to generating reaching motions that start from unseen poses and reach the whole range of 3D space around the starting pose. This is why we choose not to evaluate on held-out motion-goal pairs from the training data. Instead, we only hold out initial poses. During evaluation, starting from these unseen poses, we generate motions that attempt to reach goals that uniformly cover the volume of a cylinder centered on the human, including completely out-of-distribution goal locations (see Sup. Mat. Sec. 5). + +In particular, the set of evaluation goals is defined in a cylindrical coordinate frame by taking all the combinations of (1) 5 angles equally separating the 360 degrees around the human, (2) 5 different goal heights ranging from 0 to 1.8 meters and (3) 5 distances from 0.5 to 5 meters. We generate motions from 6 different initial poses, with an 8-second duration specified for reaching each goal. Five motions are sampled for each pose-goal combination, resulting in $5 \times 5 \times 5 \times 6 \times 5 = 3750$ unique motion sequences from which our metrics are computed. This setup allows us to thoroughly test our model in diverse scenarios, including long-term movements, navigational skills, and reaching motions at various heights and distances. + +# 4.3. Evaluation Metrics + +To accurately assess the effectiveness of our approach in generating realistic, goal-oriented human motion, we employ a set of metrics focused on both the ability to successfully reach the intended goal and the naturalness of the motion. These metrics are: + +- **Success Rate (SR):** This quantifies the percentage of motions where the right wrist reaches within $10\mathrm{cm}$ of the goal, indicating successful goal attainment. The criterion for success aligns with that used in [4]. + +- Foot Skating (FS): FS evaluates the naturalness of the motion based on foot skating, where a frame is considered as having foot skating if the lowest vertex of the human mesh moves more than $0.66 \mathrm{~cm}$ between consecutive + +![](images/e2a54a6a851db091d9d8890cd4f6bffba33ab5aa51cdd843f5eee8b3634ab48b.jpg) +(a) + +![](images/92bb0804f869f5d58f8efcb7d32025f886e70701e90af6e669dcb19ba2f95564.jpg) +(b) + +![](images/02c60a9a8cd2c0c0b8abdef35b801868fdf8dc0370a2cd5e1d219fa6fc84e006.jpg) +(c) +Figure 4. Diverse motion generated with WANDR: Displaying a range of motions generated by WANDR from various initial poses towards arbitrary goals. Examples include navigating towards goals from initial orientations not facing the goal (a, b, c, d), elevating the right hand to reach higher targets (c), and bending down to access goals near the floor (d), showcasing the model's ability to adapt to novel goal locations. + +![](images/d64e743a7666665c40dcbe4fe1ba1c10a548d588277e1d5533fad2213c9f0c51.jpg) +(d) + +frames (adjusted from the 1cm threshold used in [4] to accommodate our 30fps motion generation). + +- Distance to Goal (DTG): DTG records the closest distance in cm that the right wrist gets to the goal during a motion. This metric offers a nuanced view of the model's capability to guide the motion towards the goal. + +# 4.4. Results + +# 4.4.1 Quantitative Results + +Combining AMASS and CIRCLE: Our evaluation in Table 1 validates our hypothesis about the benefits of training with both the AMASS and CIRCLE datasets. On the one hand, training only on CIRCLE (line 1) is not sufficient for the model to learn necessary navigational skills, such as walking, due to the dataset's narrow focus or reaching motions. This is apparent from the very high foot-skating. On the other hand, training only on AMASS (line 2) results in high-quality motion generation with low foot-skating, but a relatively low success rate in goal-reaching. The Distance to Goal (DTG) metric, suggests that the model is able to navigate close to the target, but it lacks the precise movement needed to successfully reach for the goal. Using both datasets (line 3) illustrates how our approach effectively merges the broad motion vocabulary of AMASS with the goal-oriented precision of CIRCLE, leading to both high-quality motion and improved goal-reaching capability. + +We also compare with GOAL [32], a method that generates human motions that reach and grasp objects. GOAL is trained on GRAB [31], a dataset purely consisting of mo + +
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR/Circle0%56%205.4
WANDR/AMASS16%19%48.0
WANDR32%16%24.8
GOAL [32]0%29%149.2
+ +Table 1. We evaluate WANDR trained on different datasets and compare with GOAL [32]. Training solely on CIRCLE results in unrealistic motions, whereas AMASS excels in motion quality but struggles with finer goal-reaching skills. WANDR, leveraging both of what these datasets offer, demonstrates realistic motions as well as better ability to reach goals compared to baselines and existing methods. + +tions of humans grasping and manipulating objects. Since the relative positioning of the human and the object as well as the motions in GRAB have very small variations, GOAL does not succeed in any of the evaluation configurations (line 4). + +Ablation of Intention Features: In order to demonstrate the contribution of each component of the intention feature we conduct an ablation study (Table 2). Using only wrist intention (line 1) results in the lowest foot skating, due to the minimal constraints applied to the motion, allowing for more adaptable motion planning. But wrist intention features, are time-dependent and do not carry information about the absolute distance to the goal. This is why it can lead to the avatar over- or under-shooting and thus achieving a low success ratio (SR). The addition of pelvis intention (line 2) enables the avatar to sense the distance to the + +goal while the orientation intention (line 3) properly aligns the body to face the goal. Since pelvis and orientation intention add more constraints to the motion, they can sometimes cause more challenging body dynamics and produce motions with higher foot-skating (FS) (e.g. turning around in place instead of walking in a U-turn). We also try removing the intention from the motion prior and optimizing the latent space of a randomly generated initial motion to minimize the distance between the wrist and the goal (VAE + opt). This approach fails since the result is heavily dependent on the initialization of the latent variables of the motion. Our results confirm that each component of the intention feature is essential to achieving the overall performance of the model. For more details on the optimization see Sup. Mat. + +
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR (Iw)15%13%62.9
WANDR (Iw + Ip)18%17%44.9
WANDR (Iw + Ir)19%19%36.0
WANDR (full intention)32%16%24.8
VAE + opt3%4%217.0
+ +Table 2. Ablation Study. We evaluate the impact of each component of the intention vector. We also compare with an optimization baseline that does not use any condition signals. The results highlight the effectiveness of all of the components of intention as well as the fact that the complexity of the task makes "brute-forcing" with optimization unsuccessful. + +Success Ratio Distribution: Our decomposition of the model's success ratio, presented in Fig. 5, offers insights into how the model's performance varies with respect to different goal positions. WANDR demonstrates a consistent ability to reach goals across various distances (blue) and directions (green). It is more capable at reaching goals that are closer to the natural position of the wrist and do not require extensive bending or stretching (yellow). This trend likely results from the abundance of standing or upright motion sequences in the training data, as opposed to motions involving bending or crouching. This analysis provides valuable information for future improvements and dataset balancing. + +# 4.4.2 Qualitative Results + +In Fig. 4, we show a variety of motion sequences generated with our network featuring reaching goals located at varying distances and heights, highlighting the model's ability to realistically and smoothly orient, navigate, and reach for goals. These goals require actions such as bending down, turning, or stretching upwards. A critical aspect observed is the model's ability to decelerate as it approaches the goal, seamlessly coordinating body and arm movements to achieve a natural-looking reaching motion. Overall, the + +![](images/553e7be86e57d9c1cd546370764f29503c78f0a949f4f792b35eafcf2d99bdfe.jpg) +Figure 5. We show the success rates of reaching goals at various heights, angles, and distances from the initial human pose. It highlights how goal position affects the model in accurately navigating and achieving the goals. + +qualitative results show that our network generalizes well to novel goal locations while generating realistic motions. For more results please see Sup. Mat. and the video. + +# 5. Conclusion + +In conclusion, our research presents a novel data-driven approach to human motion generation, focusing on the task of reaching arbitrary goals in space. We introduce novel intention features that enable learning both general navigational skills from AMASS and goal-reaching skills from the CIRCLE dataset under the same distribution. We evaluate our model's ability to reach unseen goals that cover the whole space an avatar should be able to reach around it. The autoregressive design of WANDR demonstrates generalizability in generating realistic human motions that reach unseen goals without requiring any extra guidance information such as a pre-defined trajectory. + +Limitations and Future Work: Our approach is not without its limitations. Currently, error accumulation can sometimes bring the avatar to states where it can no longer recover. Additionally, our model shows less proficiency in reaching extremely low or high goals, reflecting a need for more diverse training data encompassing a wider range of body movements. Future work could focus on incorporating realistic grasping mechanisms and interactions with objects, as well as including scene navigation capabilities. This could involve integrating more complex datasets or developing advanced algorithms capable of understanding and interacting with varied environmental contexts, thereby pushing the boundaries of realistic human motion simulation. + +Acknowledgments: Markos Diomataris was supported in part by the Max Planck ETH Center for Learning Systems. We thank P. Ghosh, O. Ben Dov S. Tripathi, for the fruitful discussions and A. Cseke, T. Niewiadomski, T. Alexiadis, T. McConnell, for conducting the user studies. + +# Conflicts of Interest: + +https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt + +# References + +[1] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pages 565-574. IEEE, 2021. 3 +[2] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5223-5232, 2020. 3 +[3] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. In Advances in Neural Information Processing Systems, 2017. 2 +[4] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 2, 3, 6, 7 +[5] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In International Conference on 3D Vision (3DV), 2024. 2 +[6] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3D human motion synthesis model via conditional variational auto-encoder. In Computer Vision and Pattern Recognition (CVPR), pages 11645–11655, 2021. 3 +[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision (ECCV), pages 387-404. Springer, 2020. 3 +[8] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[9] Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 2017 International Conference on 3D Vision (3DV), pages 458-466. IEEE, 2017. 3 +[10] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. arXiv:2108.08284, 2021. 2, 3 +[11] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 2, 3 +[12] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. arXiv:1904.05767, 2019. 3 + +[13] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3 +[14] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[15] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion VAEs. Transactions on Graphics (TOG), 2020. 2, 3, 5 +[16] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. Transactions on Graphics (TOG), 2020. 3 +[17] Matthew Loper, Naureen Mahmood, and Michael J Black. Mosh: motion and shape capture from sparse markers. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 33(6):220-1, 2014. 2 +[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 6 +[19] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[20] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In International Conference on 3D Vision (3DV), 2024. 3 +[21] Dirk Ormoneit, Hedvig Sidenbladh, Michael Black, and Trevor Hastie. Learning and tracking cyclic human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3 +[22] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 4 +[23] Vladimir Pavlovic, James M Rehg, and John MacCormick. Learning switching linear models of human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3 +[24] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: adversarial motion priors for stylized physics-based character control. Transactions on Graphics (TOG), 2021. 2, 3 +[25] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[26] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 5 + +[27] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44 (6):3300-3315, 2021. 3 +[28] Hedvig Sidenbladh, Michael J Black, and David J Fleet. Stochastic tracking of 3d human figures using 2d image motion. In European Conference on Computer Vision (ECCV). Springer, 2000. 3 +[29] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision (IJCV), 87 (1-2):4-27, 2010. 3 +[30] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. Transactions on Graphics (TOG), 2019. 3 +[31] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3, 7 +[32] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7 +[33] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. Grip: Generating interaction poses using latent consistency and spatial cues, 2023. 3 +[34] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3 +[35] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In British Machine Vision Conference (BMVC), pages 1-13, 2017. 3 +[36] Raquel Urtasun, David J Fleet, and Pascal Fua. Temporal motion models for monocular and multiview 3d human body tracking. Computer Vision and Image Understanding (CVIU), 104(2-3):157-177, 2006. 3 +[37] Qi Wang, Thierry Artières, Mickael Chen, and Ludovic Denoyer. Adversarial learning for modeling human motion. The Visual Computer, 36(1):141-160, 2020. +[38] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI conference on artificial intelligence, pages 12281-12288, 2020. 3 +[39] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In International Conference on Computer Vision (ICCV), pages 2228-2238, 2023. 3 +[40] Ping Yu, Yang Zhao, Chunyuan Li, Junsong Yuan, and Changyou Chen. Structure-aware human-action generation. + +In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 18-34. Springer, 2020. 3 +[41] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3 +[42] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision, pages 518-535. Springer, 2022. 3 +[43] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[44] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3 +[45] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision (ECCV), pages 676-694. Springer, 2022. 3 +[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 4 \ No newline at end of file diff --git a/2024/WANDR_ Intention-guided Human Motion Generation/images.zip b/2024/WANDR_ Intention-guided Human Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..afce1b92ab0ce870369066b12899e5cb2089a8ef --- /dev/null +++ b/2024/WANDR_ Intention-guided Human Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e9f7a15718889a9b59e3719557d32b28f7d973bad0683da8f1acca39f3ad9ca +size 267527 diff --git a/2024/WANDR_ Intention-guided Human Motion Generation/layout.json b/2024/WANDR_ Intention-guided Human Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a6f904f19a3f51c5638d8a666ca13324f1c387d3 --- /dev/null +++ b/2024/WANDR_ Intention-guided Human Motion Generation/layout.json @@ -0,0 +1,7784 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "text", + "content": "WANDR: Intention-guided Human Motion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "content": "Markos Diomataris " + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "content": " Nikos Athanasiou " + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "content": " Omid Taheri " + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "content": " Xi Wang " + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "content": " Otmar Hilliges " + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 142, + 473, + 171 + ], + "type": "text", + "content": " Michael J. Black" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 172, + 475, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 172, + 475, + 185 + ], + "spans": [ + { + "bbox": [ + 119, + 172, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 172, + 475, + 185 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems, Tübingen, Germany " + }, + { + "bbox": [ + 119, + 172, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 172, + 475, + 185 + ], + "type": "text", + "content": "ETH Zürich, Switzerland" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 189, + 547, + 415 + ], + "blocks": [ + { + "bbox": [ + 50, + 189, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 50, + 189, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 50, + 189, + 547, + 415 + ], + "type": "image", + "image_path": "6f6ffb3bc9c796b2db9a211351ebbea7733ac7f0758be8b079cfcc56c28c80de.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 420, + 546, + 465 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 546, + 465 + ], + "type": "text", + "content": "Figure 1. WANDR starts from an arbitrary body pose and generates precise and realistic human motions that reach a specified 3D goal (depicted as a red sphere). Employing a purely data-driven approach, WANDR is a conditional Variational Autoencoder guided by intention features (depicted arrows) that steer the human's orientation (yellow), position (cyan) and wrist (pink) towards the goal. WANDR is able to reach a wide range of goals even if they deviate significantly from the training data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 483, + 192, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 483, + 192, + 496 + ], + "spans": [ + { + "bbox": [ + 143, + 483, + 192, + 496 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": "Synthesizing natural human motions that enable a 3D human avatar to walk and reach for arbitrary goals in 3D space remains an unsolved problem with many applications. Existing methods (data-driven or using reinforcement learning) are limited in terms of generalization and motion naturalness. A primary obstacle is the scarcity of training data that combines locomotion with goal reaching. To address this, we introduce WANDR, a data-driven model that takes an avatar's initial pose and a goal's 3D position and generates natural human motions that place the end effector (wrist) on the goal location. To solve this, we introduce novel intention features that drive rich goal-oriented movement. Intention guides the agent to the goal, and interactively adapts the generation to novel situations without needing to define sub-goals or the entire motion path. Crucially, intention allows training on datasets that have goal-oriented motions as well as those that do not. WANDR is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 484, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 557 + ], + "type": "text", + "content": "a conditional Variational Auto-Encoder (c-VAE), which we train using the AMASS and CIRCLE datasets. We evaluate our method extensively and demonstrate its ability to generate natural and long-term motions that reach 3D goals and generalize to unseen goal locations. Our models and code are available for research purposes at wandr.is.tue.mpg.de." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 585, + 387, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 387, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 387, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Goals drive our motions. Even the simplest goal can give rise to intricate motions. Consider reaching for a coffee cup – it can be as straightforward as an arm extension or can involve the coordinated action of our entire body. Actions like bending down, extending our arm, and walking must come together to achieve the goal. At a granular level, we continuously make subtle adjustments to maintain balance and stay on course towards our objective. The result is a fluid motion that seamlessly integrates numerous smaller movements, all" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "927" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "converging toward a common and simple goal: placing our hand on the cup. Generating this hierarchy of motions, from the overarching goal to the moment-to-moment individual actions, remains a longstanding challenge in computer vision, graphics, and robotics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 361 + ], + "type": "text", + "content": "Here we focus on a representative task, illustrated in Fig. 1: given a goal location in space and a starting pose, a humanoid agent must place an end effector (wrist joint) on the goal location while moving in a natural human-like way. To solve the task, the agent needs to be able to approach the goal, orient itself towards it, and reach out such that its wrist makes contact with the goal. Our primary emphasis is on ensuring autonomy for human agents. Consequently, we strive to minimize the guidance information provided, limiting it only to the human's initial pose and the goal's position. Diverging from prior data-driven approaches [4, 17], we choose to refrain from evaluating the model solely on limited labeled data. Instead, we devise an evaluation pipeline that requires agents to reach goals positioned in diverse locations around them. Considering the arbitrary selection of the goal during evaluation, and the minimal guidance information provided, tackling this task is challenging, demanding an approach with the capacity to generalize beyond the distribution of the training dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 363, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 554 + ], + "type": "text", + "content": "Existing methods approach this problem either using reinforcement learning (RL) [11, 15, 24, 43] or by capturing task-specific datasets [4, 8, 32]. While RL provides a principled way to explore the solution space, it comes with considerable shortcomings. The \"trial and error\" of exploratory learning, in combination with the high dimensionality of human motion result in policies requiring an enormous amount of training even to achieve simple tasks such as walking to a waypoint [5, 43]. In addition, since motion naturalness is better captured by data and not reward functions, RL approaches tend to produce motions that lack naturalness and expressiveness. Data-driven approaches on the other hand, rely on plentiful training motions that are acquired through motion capture and carefully curated for the downstream tasks [4, 10]. Such approaches do not scale and do not generalize well to out-of-distribution tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": "In prioritizing both motion realism and training efficiency, we adopt a data-driven approach. However, current data-driven methods lack the ability to learn both from smaller datasets that provide high-quality human reaching motions with goal labels, and from unlabeled larger scale datasets that contain necessary motion skills such as navigating to a goal position. This raises two key challenges. First, how do we model human motion in a way that generated motions can combine skills from different datasets? Second, what should the training objective be in the cases where goal labels are absent?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "To address these challenges, we propose WANDR (Wrist-driven Autonomous Navigation for Data-based goal" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "type": "text", + "content": "Reaching). We observe that by modeling human motion generation as an autoregressive stochastic process that produces motions frame by frame, WANDR is able to combine pieces of different dataset distributions when generating a motion sequence. Each generation step is conditioned on goal-related information that we call intention (visualized arrows in Fig. 1). We carefully design intention in a way that strikes a balance between being informative enough to guide the avatar to reach the goal, while also being abstract enough to promote generalization to unseen goals. This allows our generated motions to reach goals that were never encountered during the training phase in a completely zero-shot evaluation scenario. By generating the motion in an autoregressive way, we disentangle the spatial and temporal dimensions of motion. This is necessary as it allows our model to generate novel long-term sequences while being realistic in terms of local dynamic details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 277, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 457 + ], + "type": "text", + "content": "In more detail, our method is based on a conditional Variational Auto-Encoder (c-VAE) that learns to model motion as a frame-by-frame generation process by auto-encoding the pose difference between two adjacent frames. The condition signal consists of the human's current pose and dynamics along with the intention information. Intention is a function of both the current pose and the goal location and therefore actively guides the avatar during the motion generation in a closed loop manner. Through training, the c-VAE learns the distribution of potential subsequent poses conditioned on the current dynamic state of the human and its intention towards a specific goal. We train WANDR using two datasets: AMASS [18], which captures a wide range of motions including locomotion, and CIRCLE [4], which captures reaching motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "text", + "content": "Although AMASS is large, it lacks any explicit label of goals or intentions. To address this, inspired by the Hind-sight Experience Replay paradigm in robotics [3], we define intention using a hallucinated goal derived from the ground-truth wrist position in a future frame. This approach allows us to establish a unified training objective spanning AMASS and CIRCLE. Consequently, our model learns to combine motions from both datasets, enabling it to effectively reach arbitrary goals during testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "In summary, we present WANDR, a data-driven method that combines an autoregressive motion prior with a novel intention guiding mechanism and is able to generate avatars that realistically move in space and reach arbitrary goals. We experimentally evaluate our approach, including the benefit of combining multiple datasets as well as the generalization capabilities of our motion generator. Our results underscore the efficacy of the intention mechanism as an elegant way of guiding the motion generation process while also enabling the incorporation of pseudo goal labels for datasets lacking explicit goal annotations. The model and code are available for research purposes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "928" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 212 + ], + "type": "text", + "content": "Early research in motion generation focuses on tasks like motion prediction [1, 2, 7, 9, 19, 27] and unconstrained motion generation [6, 16, 21, 23, 28, 29, 36-38, 40]. More recently, significant effort has been devoted into improving controllability, with a focus on motion generation conditioned on different types of goals [25, 39], enabling interactions with scenes [12, 13, 20] and objects [31, 33, 42, 45]. Methods that attempt goal-driven motion generation can be broadly divided into reinforcement learning or data-driven approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 222, + 285, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 222, + 285, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 285, + 235 + ], + "type": "text", + "content": "2.1. Reinforcement Learning for Motion Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 601 + ], + "type": "text", + "content": "Many existing works employ Reinforcement Learning (RL) for the generation of task-specific long motion sequences. Representative work includes MotionVAE [15] and AMP [24]. MotionVAE [15] employs a two-step process where it initially leverages an autoregressive conditional Variational Autoencoder (VAE) to construct a latent space that encapsulates possible human movements. Subsequently, it utilizes RL to sample from this action space to reach a designated target location while avoiding obstacles by monitoring the area ahead. Similar to MotionVAE, GAMMA [43] learns a policy to extract samples from a latent space and then employs a tree-based search algorithm to find viable motions that steer clear of obstacles by considering the environmental geometric constraints. DIMOS [44] further extends the GAMMA framework by introducing two specialized policy networks: one for locomotion and one for interaction. Together these networks generate goal-conditioned motion sequences that dynamically interact with objects and the environment. AMP [24] learns an adversarial motion prior from unstructured datasets and then applies goal-conditioned RL. This approach involves the formulation of a style reward to encourage the resemblance of the generated sequences to those in the dataset, complemented by a task-specific reward aimed at achieving a particular objective. Hassan et al. [11] extend AMP to produce motions that facilitate interactions with the scene, by conditioning both the discriminator and the policy network on the scene context. However, RL requires significant computation and struggles to generate natural and expressive motion sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 285, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 285, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 285, + 624 + ], + "type": "text", + "content": "2.2. Data-driven Approaches for Motion Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Most data-driven approaches use existing motion capture (MoCap) datasets [18, 35] to train their models through supervised learning. The pioneering Neural State Machine method [30] is a data-driven technique for generating motion with character-scene interactions, focusing on scenarios with a limited number of objects and interactions. HuMoR [26] proposes a robust model for 3D human shape and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": "temporal pose estimation, yet it falls short of generating motions that are conditioned on specific goals. The SAMP [10] method, designed for real-time stochastic motion synthesis, generates diverse human-scene interaction movements by breaking down the process into predicting goals, planning paths, and generating motion along a predefined route. The GOAL [32] method, trained on the GRAB dataset [31], produces motion sequences in which humans walk towards and grasp 3D objects. However, the generated motions exhibit minimal movements, especially in the feet. To address these constraints, the newly introduced CIRCLE [4] dataset provides a collection of reaching motion data. This dataset is used to train a neural network that generates diverse scene-aware reaching motions. Recently, diffusion models have seen the most success at generating motions conditioned on textural input [34, 41] and spatial data [14]. This advancement enables the synthesized motion to accurately reach specified target locations or navigate around obstacles. Nevertheless, the effectiveness of data-driven approaches is constrained by the amount of training data and they lack generalization to out-of-distribution scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 356, + 361, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 361, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 361, + 369 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 383, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 547, + 635 + ], + "type": "text", + "content": "Our goal is to have a virtual human that can autonomously and realistically move from an initial pose to an arbitrary goal position and accurately place its right hand on the target. This challenge requires a nuanced understanding of human motion and the intricate dynamics involved in goal-oriented motions. For example, when the human tries to reach a distant goal, the motions are mostly focused on the legs and navigating the body to approach the object, but when it gets close to the object, the focus will be on moving the arms and upper body to reach the target. Using these observations, we develop a method named WANDR, which, although trained in a supervised setting on motion capture data, exhibits generalization in reaching unseen goal locations during test time. WANDR is designed to generate human motion in an autoregressive frame-by-frame fashion, conditioned on novel intention features. During training, these features are extracted by picking a future frame as the goal for the wrist. During inference, the intention features are dynamically computed based on the goal's position in a feedback loop, guiding the virtual human to reach the goal. See Fig. 2 for the network overview." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "In this section, we first consider the different distributions of the datasets we will be using (section 3.1). Following this, we detail the components of the intention features and how they are computed during both the training and inference phases (Section 3.2). Finally, we define the motion representation and motion generator network (Section 3.3)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "929" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 68, + 523, + 222 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 523, + 222 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 523, + 222 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 523, + 222 + ], + "type": "image", + "image_path": "40453d55b9cf3fcc28638003879d9578750b23cf414d6914c1f57b4eaa8dbca2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": "Figure 2. WANDR architecture. During training, our model conditions on the intention vectors " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "I^p" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "I^r" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "I^w" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": ", learning to associate them with actions that result into reaching goals realistically. When the training data has no defined goal, we create a goal based on the wrist location in future frames; see Sec. 3.2. The state of the avatar, " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "p_i^{dyn}" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": " expresses the SMPL-X local pose parameters " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": ", as well as the deltas " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "d_{i-1}" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": " the body parameters have in frame " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "i-1" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": ". During inference, WANDR takes the intention features, the state, and random noise and returns the change in pose, " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": ". The next pose, " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": " is obtained by integrating the " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": " with the previous pose " + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\hat{p}_{i-1}" + }, + { + "bbox": [ + 46, + 232, + 546, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 306, + 167, + 396 + ], + "blocks": [ + { + "bbox": [ + 55, + 306, + 167, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 167, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 167, + 396 + ], + "type": "image", + "image_path": "a0993d495849048161452f010dd815479f228bd758fd187ae25d9f04bde8d14c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 404, + 288, + 470 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 470 + ], + "type": "text", + "content": "Figure 3. In training, if goals are not specified, they are determined by the future wrist location at a randomly selected future timestep, compensating for the lack of paired ground-truth data in AMASS and direct human motion through intention vectors. During inference, target locations are used as goals with intention vectors calculated based on these specific locations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 306, + 279, + 396 + ], + "blocks": [ + { + "bbox": [ + 168, + 306, + 279, + 396 + ], + "lines": [ + { + "bbox": [ + 168, + 306, + 279, + 396 + ], + "spans": [ + { + "bbox": [ + 168, + 306, + 279, + 396 + ], + "type": "image", + "image_path": "491de65aee91085c701c2ca5ddf33e417a06f71cd35ccc8e52632abb781c9ed1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 210, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 210, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 210, + 529 + ], + "type": "text", + "content": "3.1. Two Complementing Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 715 + ], + "type": "text", + "content": "For the development of WANDR, we use two key datasets: AMASS and CIRCLE. AMASS is a large-scale dataset that offers a broad range of general human motions but lacks a specific focus on goal-reaching tasks. Its diverse collection of movements provides a solid base for understanding human locomotion and body movement when the person is far away from the goal location. In contrast to AMASS, CIRCLE is tailored towards movements involving reaching specific target positions, particularly capturing the nuances of upper body and arm movements. By integrating AMASS's general motion diversity with CIRCLE's targeted goal-reaching data, we equip WANDR with the ability to generate the entire process of reaching a distant goal, from the initial navigation to the precise target-reaching actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 308, + 416, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 416, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 416, + 319 + ], + "type": "text", + "content": "3.2. Intention Features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "content": "We represent 3D human motion as a sequence of SMPL-X [22] body poses " + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\{p_1,\\dots,p_N\\}" + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "content": ". Each pose " + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "inline_equation", + "content": "p_i\\in \\mathbb{R}^{135}" + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "content": " consists of three concatenated components: the body's translation " + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "inline_equation", + "content": "t_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "content": ", root orientation " + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "inline_equation", + "content": "r_i\\in \\mathbb{R}^6" + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "content": ", and the body pose " + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "inline_equation", + "content": "\\theta_{i}\\in \\mathbb{R}^{21\\times 6}" + }, + { + "bbox": [ + 305, + 326, + 546, + 386 + ], + "type": "text", + "content": ", both in 6D format [46]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 386, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 445 + ], + "type": "text", + "content": "For the avatar to reach the goal, it is important to be informed about the spatial relation of the goal location with respect to its current pose, as well as to have a sense of time to reach the goal promptly. We achieve this by introducing the intention features, which are central to our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": "To define the intention features " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": ", it is crucial to first establish the selection criteria for the goal " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": ", within a motion sequence. Our training involves both label-available scenarios (e.g., CIRCLE dataset) and label-absent scenarios (e.g., AMASS dataset). In label-available cases, where the goal position " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": " and the frame index " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "t_{G}" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": " which it is reached are known, calculating " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": " is straightforward. Conversely, for label-absent scenarios like in AMASS, we pick a random future frame as the " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "t_{G}" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": " and define the human's right wrist joint location as the goal " + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 446, + 546, + 577 + ], + "type": "text", + "content": " (Fig. 3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 578, + 538, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 538, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 538, + 588 + ], + "type": "text", + "content": "In both scenarios, the intention features are defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 382, + 598, + 468, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 598, + 468, + 611 + ], + "spans": [ + { + "bbox": [ + 382, + 598, + 468, + 611 + ], + "type": "interline_equation", + "content": "I _ {i} = I _ {i} \\left(p _ {i}, G, t _ {G}, i\\right).", + "image_path": "4d7bcf260fd1b549b1a63b9d5dd5066f4b457aa8df247303adb3e0466b2033d5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 620, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 620, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 305, + 620, + 545, + 692 + ], + "type": "text", + "content": "These features are essentially a function of the current body pose, the goal position, and time, offering both spatial and temporal insights required for the motion to reach the goal. They are designed to provide sufficient information to reach goals while also enabling test-time generalization. We define them as three distinct components:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 700, + 463, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 700, + 463, + 715 + ], + "spans": [ + { + "bbox": [ + 388, + 700, + 463, + 715 + ], + "type": "interline_equation", + "content": "I _ {i} = \\left(I _ {i} ^ {w}, I _ {i} ^ {r}, I _ {i} ^ {p}\\right).", + "image_path": "29fca013a1046007fee3bd0fbc17ecc10c61b857151a1434b337f56b4f2f146b.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "930" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "These components represent wrist intention, body orientation intention, and pelvis intention, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": "Wrist Intention: This is the main time-dependent component that guides the wrist to reach the goal. It is calculated as the necessary average velocity for the wrist to be at the goal location in time, defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 152, + 199, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 152, + 199, + 178 + ], + "spans": [ + { + "bbox": [ + 135, + 152, + 199, + 178 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {w} = \\frac {G - W _ {i}}{t _ {G} - i}", + "image_path": "4d91892588a7e57b553a89f6dcb1b22b8848801e4780dec5bdfaf84b3ab3f5d9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "type": "inline_equation", + "content": "t_G" + }, + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "type": "text", + "content": " is the frame when the goal should be reached. During training, we know the goal frame and the time to reach it. At test time, we know the goal location but need an externally-supplied time to reach it. An emergent behavior of this formulation is that, at inference, the model is able to adjust its movement speed and reach the goal just in time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "content": "Orientation Intention: This component captures the body orientation when reaching the goal location. By conditioning on this, we ensure that the human model orients towards the goal and smoothly navigates towards it, preventing unnatural motions during inference, like walking backward. During training, this is defined as the difference between the forward direction of the current body frame, " + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "inline_equation", + "content": "H_{i}^{xy}" + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "content": ", and the goal body, " + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "inline_equation", + "content": "H_{tG}^{xy}" + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "content": " signifies removing the " + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "content": " component. During inference, since we do not have the goal body, we use the pelvis position " + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 46, + 258, + 286, + 400 + ], + "type": "text", + "content": " to calculate the pelvisto-goal direction as the desired orientation. This feature is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 409, + 261, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 409, + 261, + 441 + ], + "spans": [ + { + "bbox": [ + 71, + 409, + 261, + 441 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {r} = \\left\\{ \\begin{array}{l l} H _ {t _ {G}} ^ {x y} - H _ {i} ^ {x y} & \\text {d u r i n g t r a i n i n g} \\\\ (G - P _ {i}) ^ {x y} - H _ {i} ^ {x y} & \\text {d u r i n g i n f e r e n c e}. \\end{array} \\right.", + "image_path": "f10cb814df9d9ae0f5b43430501f0ce10339291201825df23cb7950b97f4fcd5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 569 + ], + "type": "text", + "content": "Pelvis Intention: This feature captures information about the position of the goal relative to the body. It is the difference between the goal and the pelvis joint, excluding the z (height) component. Following the approach in [32], we scale this distance by an exponential function that saturates this vector to have a maximum norm of 2. This formulation helps the method generalize to navigating towards the goal during longer motions and helps the model learn since the distance from the goal does not grow indefinitely in extreme scenarios. This intention is defined by the equation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 578, + 267, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 578, + 267, + 606 + ], + "spans": [ + { + "bbox": [ + 66, + 578, + 267, + 606 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {p} = 2 \\times \\left(1 - e ^ {\\left\\| G ^ {x y} - P _ {i} ^ {x y} \\right\\| _ {2}}\\right) \\times \\frac {G ^ {x y} - P _ {i} ^ {x y}}{\\left\\| G ^ {x y} - P _ {i} ^ {x y} \\right\\| _ {2}}.", + "image_path": "9abfd78fce0c62910ca276975e079c2f8038fe6c37aa739b3321a62193c02738.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 615, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 286, + 661 + ], + "type": "text", + "content": "In the experimental section, we delve into the significance of each of these intention features and discuss the rationale behind our design choices, illustrating their impact on the effectiveness of our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 198, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 198, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 198, + 682 + ], + "type": "text", + "content": "3.3. Motion Network (WANDR)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "WANDR is designed as a conditional Variational AutoEncoder (c-VAE) network, operating in an autoregressive" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "manner to generate sequential motion frames. This framework is pivotal in predicting the subsequent pose in a motion sequence, emphasizing an incremental, frame-by-frame approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": "Central to our approach is the training of the c-VAE to autoencode pose deltas, denoted as " + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "inline_equation", + "content": "d_{i} \\in \\mathbb{R}^{135}" + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": ". These deltas represent the difference between two consecutive poses, " + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "inline_equation", + "content": "p_{i - 1}" + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": ". By focusing on pose deltas rather than absolute pose values, our model benefits from an important inductive bias, enhancing its learning efficiency and performance, as supported by prior research [15, 26]. We separate rotational differences into: body orientation " + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "inline_equation", + "content": "(d_i^r)" + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": " and body pose " + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "inline_equation", + "content": "(d_i^\\theta)" + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": ", each expressed in a 6-D rotational format. Translation deltas are denoted as " + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "inline_equation", + "content": "d_{i}^{t} = t_{i} - t_{i - 1}" + }, + { + "bbox": [ + 304, + 120, + 545, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": "To enhance the motion representation's invariance, we remove information related to the global " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": "-orientation. This is accomplished by subtracting the global " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " Euler angle of " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "r_{i-1}" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " from both the translational " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "(d_i^t)" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " and rotational " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "(d_i^r)" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " deltas. The resulting deltas, " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "d_i^{t-z}" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " for translation and " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "d_i^{r-z}" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " for orientation, provide a more robust and consistent representation of motion, irrespective of global direction. Consequently, the delta pose features for any given frame " + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 240, + 545, + 348 + ], + "type": "text", + "content": " are composed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 357, + 471, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 357, + 471, + 372 + ], + "spans": [ + { + "bbox": [ + 380, + 357, + 471, + 372 + ], + "type": "interline_equation", + "content": "d _ {i} = \\left(d _ {i} ^ {t - z}, d _ {i} ^ {r - z}, d _ {i} ^ {\\theta}\\right).", + "image_path": "55d43223cc1229fa1403587d5241ae06a28c4b7b69b77728a60cf2f743e9cf72.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 380, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 464 + ], + "type": "text", + "content": "An advantage of this representation is its consistency across different motion global orientations. For instance, in the scenario of a person walking, the delta representation remains agnostic to the walking direction. This attribute underscores the efficacy of our method in capturing the essence of motion without being biased towards any specific orientation or direction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "text", + "content": "Condition Inputs: For each motion frame, the decoder is conditioned on a combination of state and intention features. Specifically, this condition signal is formulated as " + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "inline_equation", + "content": "c_{i} = (p_{i}^{dyn}, I_{i})" + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "text", + "content": ". The state features, " + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "inline_equation", + "content": "p_{i}^{dyn}" + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "text", + "content": ", encapsulate the avatar's current local pose, focusing on the z-component of translation and a modified orientation that excludes the global z Euler angle, along with the pose deltas " + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "inline_equation", + "content": "di - 1" + }, + { + "bbox": [ + 304, + 465, + 545, + 620 + ], + "type": "text", + "content": " of the previous step. That combination ensures that the generated motion at each step is informed by both the local pose configuration of the avatar, its dynamics and its directional intention towards the set goal, vital for producing realistic, goal-oriented human motions. An overview of the network architecture is shown in Fig. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 628, + 402, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 402, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 402, + 640 + ], + "type": "text", + "content": "3.4. Training Losses" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 647, + 544, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 544, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 544, + 669 + ], + "type": "text", + "content": "Our training objective is a composite of three distinct loss functions:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 372, + 671, + 479, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 671, + 479, + 683 + ], + "spans": [ + { + "bbox": [ + 372, + 671, + 479, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r e c} + \\alpha \\mathcal {L} _ {K L} + \\mathcal {L} _ {J}.", + "image_path": "d85d745020f54a194bd48d4a0679df98c4ef8bd61abbfadd8b3eee26df600649.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "content": "The reconstruction loss, " + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "content": ", measures the accuracy of the motion reconstruction, quantified as the mean square er" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 311, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 311, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 311, + 757 + ], + "type": "text", + "content": "931" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "ror (MSE) between the input pose delta, " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ", and its reconstructed counterpart, " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\hat{d}_{i}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ". It ensures the network's ability to faithfully replicate the input motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 144 + ], + "type": "text", + "content": "The KL Divergence Loss, " + }, + { + "bbox": [ + 46, + 109, + 287, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{KL})" + }, + { + "bbox": [ + 46, + 109, + 287, + 144 + ], + "type": "text", + "content": ", evaluates the deviation of the encoded distribution from a standard normal distribution. It is formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 156, + 236, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 156, + 236, + 170 + ], + "spans": [ + { + "bbox": [ + 97, + 156, + 236, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} = \\mathcal {K L} (\\mathcal {N} (0, I) | | \\mathcal {N} (\\mu_ {i}, \\sigma_ {i})).", + "image_path": "f6f3af914ace7b0dc22e921fcee899641a616cb84733d8bec540e8edc5f5d6f9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": " represent the mean and variance of the Gaussian distribution predicted by the encoder. We balance this term with " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\alpha = 10^{-2}" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": " to prevent the over-dominance of " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": ", thereby aiding the decoder in avoiding collapse to mean predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": "Finally, we use a Joint Error Loss " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_J)" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": ", to ensure perceptual accuracy, by integrating the predicted " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": " to get the predicted next pose " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": ", which is then fed into the SMPL-X model to obtain the predicted joint positions, " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": ". The loss " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_J" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": " is the MSE between these predicted joints " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": " and the ground truth joints " + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 241, + 287, + 336 + ], + "type": "text", + "content": ", addressing errors that might not be apparent in parameter space but are perceptually significant, such as incorrect body orientation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 337, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 397 + ], + "type": "text", + "content": "Notably, our approach does not incorporate any explicit loss functions directly related to reaching a goal. This omission is a deliberate choice, aligning with our method's emphasis on generalizing to diverse goal-reaching scenarios without being constrained by goal-specific training losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 407, + 160, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 160, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 160, + 418 + ], + "type": "text", + "content": "3.5. Motion Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "content": "In the inference phase of WANDR, our primary objective is to generate human motion that is driven towards a specific goal. Using the decoder of the WANDR c-VAE, we iteratively generate and integrate pose deltas. This process is initiated from the starting pose and progressively builds upon each subsequent pose." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 499, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 558 + ], + "type": "text", + "content": "The intention features are recalculated at each step based on the current predicted pose and the goal location. They serve as a guiding mechanism, ensuring that the generated motion is consistently oriented towards placing the human's right wrist on the target." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 559, + 287, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 607 + ], + "type": "text", + "content": "The user can control the motion's pace by specifying the time to reach the goal " + }, + { + "bbox": [ + 46, + 559, + 287, + 607 + ], + "type": "inline_equation", + "content": "t_{G}" + }, + { + "bbox": [ + 46, + 559, + 287, + 607 + ], + "type": "text", + "content": ". This directly affects the wrist intention feature, enabling adjustments from fast to slow motions to suit various scenarios and constraints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 620, + 127, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 127, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 127, + 634 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "In this section, we outline the datasets used for training and evaluation and benchmark how each dataset affects the goal-reaching ability and the quality of the generated motions. Furthermore, we compare our approach with several baselines and ablate the effect of the different components of our intention vector." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 433, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 433, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 433, + 84 + ], + "type": "text", + "content": "4.1. Datasets & Processing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "text", + "content": "Our model is trained on two datasets: AMASS [18] and CIRCLE [4]. AMASS is a collection of 17k sequences, containing a wide range of motion types including long-term navigational skills like walking and turning. CIRCLE, on the other hand, contains 7.2k shorter sequences, each marked with a specific goal reached by a hand. For training, we refine AMASS by excluding sequences where feet are more than " + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "inline_equation", + "content": "20cm" + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "text", + "content": " above the ground, resulting in a combined dataset of nearly 20k sequences. This dataset is split into " + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "text", + "content": " training, " + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "text", + "content": " validation, and " + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 91, + 545, + 223 + ], + "type": "text", + "content": " test sets. All motions are re-sampled to 30 frames per second (fps)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 232, + 421, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 421, + 244 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 421, + 244 + ], + "type": "text", + "content": "4.2. Evaluation Strategy" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 251, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 370 + ], + "type": "text", + "content": "Our evaluation procedure aims at testing the degree which WANDR can generalize to generating reaching motions that start from unseen poses and reach the whole range of 3D space around the starting pose. This is why we choose not to evaluate on held-out motion-goal pairs from the training data. Instead, we only hold out initial poses. During evaluation, starting from these unseen poses, we generate motions that attempt to reach goals that uniformly cover the volume of a cylinder centered on the human, including completely out-of-distribution goal locations (see Sup. Mat. Sec. 5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 371, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 526 + ], + "type": "text", + "content": "In particular, the set of evaluation goals is defined in a cylindrical coordinate frame by taking all the combinations of (1) 5 angles equally separating the 360 degrees around the human, (2) 5 different goal heights ranging from 0 to 1.8 meters and (3) 5 distances from 0.5 to 5 meters. We generate motions from 6 different initial poses, with an 8-second duration specified for reaching each goal. Five motions are sampled for each pose-goal combination, resulting in " + }, + { + "bbox": [ + 304, + 371, + 545, + 526 + ], + "type": "inline_equation", + "content": "5 \\times 5 \\times 5 \\times 6 \\times 5 = 3750" + }, + { + "bbox": [ + 304, + 371, + 545, + 526 + ], + "type": "text", + "content": " unique motion sequences from which our metrics are computed. This setup allows us to thoroughly test our model in diverse scenarios, including long-term movements, navigational skills, and reaching motions at various heights and distances." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 536, + 418, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 418, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 418, + 547 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 555, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 613 + ], + "type": "text", + "content": "To accurately assess the effectiveness of our approach in generating realistic, goal-oriented human motion, we employ a set of metrics focused on both the ability to successfully reach the intended goal and the naturalness of the motion. These metrics are:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 618, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 665 + ], + "type": "text", + "content": "- **Success Rate (SR):** This quantifies the percentage of motions where the right wrist reaches within " + }, + { + "bbox": [ + 306, + 618, + 545, + 665 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 306, + 618, + 545, + 665 + ], + "type": "text", + "content": " of the goal, indicating successful goal attainment. The criterion for success aligns with that used in [4]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 545, + 712 + ], + "type": "text", + "content": "- Foot Skating (FS): FS evaluates the naturalness of the motion based on foot skating, where a frame is considered as having foot skating if the lowest vertex of the human mesh moves more than " + }, + { + "bbox": [ + 306, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "0.66 \\mathrm{~cm}" + }, + { + "bbox": [ + 306, + 665, + 545, + 712 + ], + "type": "text", + "content": " between consecutive" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "932" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 173, + 269 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 173, + 269 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 173, + 269 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 173, + 269 + ], + "type": "image", + "image_path": "e2a54a6a851db091d9d8890cd4f6bffba33ab5aa51cdd843f5eee8b3634ab48b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 270, + 110, + 279 + ], + "lines": [ + { + "bbox": [ + 98, + 270, + 110, + 279 + ], + "spans": [ + { + "bbox": [ + 98, + 270, + 110, + 279 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 296, + 269 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 296, + 269 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 296, + 269 + ], + "type": "image", + "image_path": "92bb0804f869f5d58f8efcb7d32025f886e70701e90af6e669dcb19ba2f95564.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 270, + 241, + 279 + ], + "lines": [ + { + "bbox": [ + 229, + 270, + 241, + 279 + ], + "spans": [ + { + "bbox": [ + 229, + 270, + 241, + 279 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 71, + 419, + 269 + ], + "blocks": [ + { + "bbox": [ + 298, + 71, + 419, + 269 + ], + "lines": [ + { + "bbox": [ + 298, + 71, + 419, + 269 + ], + "spans": [ + { + "bbox": [ + 298, + 71, + 419, + 269 + ], + "type": "image", + "image_path": "02c60a9a8cd2c0c0b8abdef35b801868fdf8dc0370a2cd5e1d219fa6fc84e006.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 270, + 358, + 279 + ], + "lines": [ + { + "bbox": [ + 347, + 270, + 358, + 279 + ], + "spans": [ + { + "bbox": [ + 347, + 270, + 358, + 279 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 288, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 547, + 331 + ], + "type": "text", + "content": "Figure 4. Diverse motion generated with WANDR: Displaying a range of motions generated by WANDR from various initial poses towards arbitrary goals. Examples include navigating towards goals from initial orientations not facing the goal (a, b, c, d), elevating the right hand to reach higher targets (c), and bending down to access goals near the floor (d), showcasing the model's ability to adapt to novel goal locations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 420, + 71, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 420, + 71, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 420, + 71, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 420, + 71, + 545, + 269 + ], + "type": "image", + "image_path": "d64e743a7666665c40dcbe4fe1ba1c10a548d588277e1d5533fad2213c9f0c51.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 270, + 490, + 279 + ], + "lines": [ + { + "bbox": [ + 477, + 270, + 490, + 279 + ], + "spans": [ + { + "bbox": [ + 477, + 270, + 490, + 279 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 354, + 287, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 287, + 377 + ], + "type": "text", + "content": "frames (adjusted from the 1cm threshold used in [4] to accommodate our 30fps motion generation)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 378, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 288, + 426 + ], + "type": "text", + "content": "- Distance to Goal (DTG): DTG records the closest distance in cm that the right wrist gets to the goal during a motion. This metric offers a nuanced view of the model's capability to guide the motion towards the goal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 434, + 104, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 104, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 104, + 445 + ], + "type": "text", + "content": "4.4. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 454, + 167, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 167, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 167, + 465 + ], + "type": "text", + "content": "4.4.1 Quantitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "content": "Combining AMASS and CIRCLE: Our evaluation in Table 1 validates our hypothesis about the benefits of training with both the AMASS and CIRCLE datasets. On the one hand, training only on CIRCLE (line 1) is not sufficient for the model to learn necessary navigational skills, such as walking, due to the dataset's narrow focus or reaching motions. This is apparent from the very high foot-skating. On the other hand, training only on AMASS (line 2) results in high-quality motion generation with low foot-skating, but a relatively low success rate in goal-reaching. The Distance to Goal (DTG) metric, suggests that the model is able to navigate close to the target, but it lacks the precise movement needed to successfully reach for the goal. Using both datasets (line 3) illustrates how our approach effectively merges the broad motion vocabulary of AMASS with the goal-oriented precision of CIRCLE, leading to both high-quality motion and improved goal-reaching capability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "We also compare with GOAL [32], a method that generates human motions that reach and grasp objects. GOAL is trained on GRAB [31], a dataset purely consisting of mo" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 319, + 352, + 533, + 416 + ], + "blocks": [ + { + "bbox": [ + 319, + 352, + 533, + 416 + ], + "lines": [ + { + "bbox": [ + 319, + 352, + 533, + 416 + ], + "spans": [ + { + "bbox": [ + 319, + 352, + 533, + 416 + ], + "type": "table", + "html": "
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR/Circle0%56%205.4
WANDR/AMASS16%19%48.0
WANDR32%16%24.8
GOAL [32]0%29%149.2
", + "image_path": "b16be26b025b25f20a868500e6c49b36c49d47a82a35be22c7708a5b93336b60.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 424, + 545, + 501 + ], + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 501 + ], + "type": "text", + "content": "Table 1. We evaluate WANDR trained on different datasets and compare with GOAL [32]. Training solely on CIRCLE results in unrealistic motions, whereas AMASS excels in motion quality but struggles with finer goal-reaching skills. WANDR, leveraging both of what these datasets offer, demonstrates realistic motions as well as better ability to reach goals compared to baselines and existing methods." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 519, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 519, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 519, + 545, + 579 + ], + "type": "text", + "content": "tions of humans grasping and manipulating objects. Since the relative positioning of the human and the object as well as the motions in GRAB have very small variations, GOAL does not succeed in any of the evaluation configurations (line 4)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Ablation of Intention Features: In order to demonstrate the contribution of each component of the intention feature we conduct an ablation study (Table 2). Using only wrist intention (line 1) results in the lowest foot skating, due to the minimal constraints applied to the motion, allowing for more adaptable motion planning. But wrist intention features, are time-dependent and do not carry information about the absolute distance to the goal. This is why it can lead to the avatar over- or under-shooting and thus achieving a low success ratio (SR). The addition of pelvis intention (line 2) enables the avatar to sense the distance to the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "933" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "goal while the orientation intention (line 3) properly aligns the body to face the goal. Since pelvis and orientation intention add more constraints to the motion, they can sometimes cause more challenging body dynamics and produce motions with higher foot-skating (FS) (e.g. turning around in place instead of walking in a U-turn). We also try removing the intention from the motion prior and optimizing the latent space of a randomly generated initial motion to minimize the distance between the wrist and the goal (VAE + opt). This approach fails since the result is heavily dependent on the initialization of the latent variables of the motion. Our results confirm that each component of the intention feature is essential to achieving the overall performance of the model. For more details on the optimization see Sup. Mat." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 261, + 287, + 339 + ], + "blocks": [ + { + "bbox": [ + 48, + 261, + 287, + 339 + ], + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 339 + ], + "type": "table", + "html": "
Train SetSR ↑FS ↓DTG (cm) ↓
WANDR (Iw)15%13%62.9
WANDR (Iw + Ip)18%17%44.9
WANDR (Iw + Ir)19%19%36.0
WANDR (full intention)32%16%24.8
VAE + opt3%4%217.0
", + "image_path": "5127dbd43d7465ee282e8efc44852aad78303ff1600ee684ffda2075b5ce74a2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 347, + 287, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 413 + ], + "type": "text", + "content": "Table 2. Ablation Study. We evaluate the impact of each component of the intention vector. We also compare with an optimization baseline that does not use any condition signals. The results highlight the effectiveness of all of the components of intention as well as the fact that the complexity of the task makes \"brute-forcing\" with optimization unsuccessful." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 427, + 289, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 289, + 571 + ], + "type": "text", + "content": "Success Ratio Distribution: Our decomposition of the model's success ratio, presented in Fig. 5, offers insights into how the model's performance varies with respect to different goal positions. WANDR demonstrates a consistent ability to reach goals across various distances (blue) and directions (green). It is more capable at reaching goals that are closer to the natural position of the wrist and do not require extensive bending or stretching (yellow). This trend likely results from the abundance of standing or upright motion sequences in the training data, as opposed to motions involving bending or crouching. This analysis provides valuable information for future improvements and dataset balancing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 162, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 162, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 162, + 598 + ], + "type": "text", + "content": "4.4.2 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "In Fig. 4, we show a variety of motion sequences generated with our network featuring reaching goals located at varying distances and heights, highlighting the model's ability to realistically and smoothly orient, navigate, and reach for goals. These goals require actions such as bending down, turning, or stretching upwards. A critical aspect observed is the model's ability to decelerate as it approaches the goal, seamlessly coordinating body and arm movements to achieve a natural-looking reaching motion. Overall, the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 547, + 175 + ], + "type": "image", + "image_path": "553e7be86e57d9c1cd546370764f29503c78f0a949f4f792b35eafcf2d99bdfe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 183, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 183, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 547, + 228 + ], + "type": "text", + "content": "Figure 5. We show the success rates of reaching goals at various heights, angles, and distances from the initial human pose. It highlights how goal position affects the model in accurately navigating and achieving the goals." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 248, + 547, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 547, + 285 + ], + "type": "text", + "content": "qualitative results show that our network generalizes well to novel goal locations while generating realistic motions. For more results please see Sup. Mat. and the video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 295, + 378, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 295, + 378, + 307 + ], + "spans": [ + { + "bbox": [ + 306, + 295, + 378, + 307 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 316, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 458 + ], + "type": "text", + "content": "In conclusion, our research presents a novel data-driven approach to human motion generation, focusing on the task of reaching arbitrary goals in space. We introduce novel intention features that enable learning both general navigational skills from AMASS and goal-reaching skills from the CIRCLE dataset under the same distribution. We evaluate our model's ability to reach unseen goals that cover the whole space an avatar should be able to reach around it. The autoregressive design of WANDR demonstrates generalizability in generating realistic human motions that reach unseen goals without requiring any extra guidance information such as a pre-defined trajectory." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 624 + ], + "type": "text", + "content": "Limitations and Future Work: Our approach is not without its limitations. Currently, error accumulation can sometimes bring the avatar to states where it can no longer recover. Additionally, our model shows less proficiency in reaching extremely low or high goals, reflecting a need for more diverse training data encompassing a wider range of body movements. Future work could focus on incorporating realistic grasping mechanisms and interactions with objects, as well as including scene navigation capabilities. This could involve integrating more complex datasets or developing advanced algorithms capable of understanding and interacting with varied environmental contexts, thereby pushing the boundaries of realistic human motion simulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 626, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 545, + 685 + ], + "type": "text", + "content": "Acknowledgments: Markos Diomataris was supported in part by the Max Planck ETH Center for Learning Systems. We thank P. Ghosh, O. Ben Dov S. Tripathi, for the fruitful discussions and A. Cseke, T. Niewiadomski, T. Alexiadis, T. McConnell, for conducting the user studies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 686, + 396, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 686, + 396, + 696 + ], + "spans": [ + { + "bbox": [ + 306, + 686, + 396, + 696 + ], + "type": "text", + "content": "Conflicts of Interest:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 698, + 529, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 698, + 529, + 710 + ], + "spans": [ + { + "bbox": [ + 318, + 698, + 529, + 710 + ], + "type": "text", + "content": "https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "934" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pages 565-574. IEEE, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5223-5232, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 246 + ], + "type": "text", + "content": "[3] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. In Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "type": "text", + "content": "[4] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "type": "text", + "content": "[5] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In International Conference on 3D Vision (3DV), 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 423 + ], + "type": "text", + "content": "[6] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3D human motion synthesis model via conditional variational auto-encoder. In Computer Vision and Pattern Recognition (CVPR), pages 11645–11655, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 468 + ], + "type": "text", + "content": "[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision (ECCV), pages 387-404. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "text", + "content": "[8] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 524, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 287, + 568 + ], + "type": "text", + "content": "[9] Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 2017 International Conference on 3D Vision (3DV), pages 458-466. IEEE, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[10] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. arXiv:2108.08284, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[11] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[12] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. arXiv:1904.05767, 2019. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[13] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 185 + ], + "type": "text", + "content": "[14] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "text", + "content": "[15] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion VAEs. Transactions on Graphics (TOG), 2020. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 255 + ], + "type": "text", + "content": "[16] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. Transactions on Graphics (TOG), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "text", + "content": "[17] Matthew Loper, Naureen Mahmood, and Michael J Black. Mosh: motion and shape capture from sparse markers. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 33(6):220-1, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "text", + "content": "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 347, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 381 + ], + "type": "text", + "content": "[19] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "text", + "content": "[20] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In International Conference on 3D Vision (3DV), 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 428, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 473 + ], + "type": "text", + "content": "[21] Dirk Ormoneit, Hedvig Sidenbladh, Michael Black, and Trevor Hastie. Learning and tracking cyclic human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "text", + "content": "[22] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 531, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 575 + ], + "type": "text", + "content": "[23] Vladimir Pavlovic, James M Rehg, and John MacCormick. Learning switching linear models of human motion. Conference on Neural Information Processing Systems (NeurIPS), 13, 2000. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 577, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 621 + ], + "type": "text", + "content": "[24] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: adversarial motion priors for stylized physics-based character control. Transactions on Graphics (TOG), 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "text", + "content": "[25] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[26] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "935" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44 (6):3300-3315, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[28] Hedvig Sidenbladh, Michael J Black, and David J Fleet. Stochastic tracking of 3d human figures using 2d image motion. In European Conference on Computer Vision (ECCV). Springer, 2000. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[29] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision (IJCV), 87 (1-2):4-27, 2010. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "text", + "content": "[30] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. Transactions on Graphics (TOG), 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[31] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "text", + "content": "[32] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[33] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. Grip: Generating interaction poses using latent consistency and spatial cues, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[34] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[35] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In British Machine Vision Conference (BMVC), pages 1-13, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[36] Raquel Urtasun, David J Fleet, and Pascal Fua. Temporal motion models for monocular and multiview 3d human body tracking. Computer Vision and Image Understanding (CVIU), 104(2-3):157-177, 2006. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "type": "text", + "content": "[37] Qi Wang, Thierry Artières, Mickael Chen, and Ludovic Denoyer. Adversarial learning for modeling human motion. The Visual Computer, 36(1):141-160, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 625 + ], + "type": "text", + "content": "[38] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI conference on artificial intelligence, pages 12281-12288, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 691 + ], + "type": "text", + "content": "[39] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In International Conference on Computer Vision (ICCV), pages 2228-2238, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[40] Ping Yu, Yang Zhao, Chunyuan Li, Junsong Yuan, and Changyou Chen. Structure-aware human-action generation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 387 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 546, + 106 + ], + "type": "text", + "content": "In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 18-34. Springer, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "text", + "content": "[41] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 546, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 205 + ], + "type": "text", + "content": "[42] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision, pages 518-535. Springer, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "type": "text", + "content": "[43] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "type": "text", + "content": "[44] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 285, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 546, + 342 + ], + "type": "text", + "content": "[45] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In European Conference on Computer Vision (ECCV), pages 676-694. Springer, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 546, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 546, + 387 + ], + "type": "text", + "content": "[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "936" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_content_list.json b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54be3921242d12da914693a957f8332233ebf6c7 --- /dev/null +++ b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Soyong Shin", + "bbox": [ + 189, + 180, + 303, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juyong Kim", + "bbox": [ + 343, + 180, + 450, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eni Halilaj", + "bbox": [ + 491, + 180, + 586, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michael J. Black", + "bbox": [ + 627, + 180, + 772, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ Max Planck Institute for Intelligent System $\\{$ soyongs, juyongk, ehalilaj\\} $@$ andrew.cmu.edu black@tuebingen.mpg.de", + "bbox": [ + 178, + 208, + 789, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02de85de1d553150e6003c73f47b9ebc3857dde697d07567bdcb53729d637c96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 257, + 156, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55c6900b3437267b9e552ea97719da27a26c57afb96518b6d56a592b526da379.jpg", + "image_caption": [ + "Figure 1. WHAM: World-grounded Humans with Accurate Motion. State-of-the-art methods like TRACE [43] and SLAHMR [53] fail to capture global 3D human trajectories accurately when given in-the-wild videos captured using a moving camera, producing implausible world-grounded motion (e.g., foot sliding). To address this, WHAM uses two novel strategies: (1) feature integration from 2D keypoints and pixels to reconstruct precise and pixel-aligned 3D human motion and (2) contact-aware trajectory recovery to place the human in a global coordinate system without foot sliding. Gray dots show the ground-truth global trajectory. See Supplemental Video." + ], + "image_footnote": [], + "bbox": [ + 81, + 359, + 156, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ae0c45593a6d230d633af96a4346b8b41e66b5f7a71c46867f2cf51b181089b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 258, + 235, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ab30b270c67489a9234fcd8cfe611b1651f1f304a0fd6449f4e05c71146dd30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 359, + 236, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fb411d59626f103f39d2d91c24cab0859a9b7601efbdb1576815ca16ecd8bda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 258, + 313, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ecb7c67e515e8b0761b9d7b4ade6e1456b734e032aedeb3dde91e5fca698f96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 359, + 313, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ca2367ede506337ffe951dd7b79558904968a72a0f20da31a3d929ea266edc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 258, + 392, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9dab6a5c367779b47e013dc525d716607eae46c73878ca694cb371e98180d571.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 359, + 390, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/166cfae0e92bc02bb9a9dcb94f3bdb6cff144eed3a63d790b4dac2ba4ad0a522.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 258, + 418, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 392, + 359, + 416, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96dbba8d7c47704d0c29f7ac033f83d5e4c1802c42563e5f977cb486c214962a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 258, + 498, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 418, + 359, + 498, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 398, + 258, + 418, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 398, + 359, + 416, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GT trajectory", + "bbox": [ + 419, + 271, + 498, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6e3f307f386fe7c11473a5c836a1f82fff316b3d6473f3767fb9a932184ec23.jpg", + "image_caption": [ + "TRACE" + ], + "image_footnote": [], + "bbox": [ + 522, + 271, + 552, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f12dbcac6531be2968ed4fb4678344e83764c9a79f918ec134d3e023463cdafc.jpg", + "image_caption": [ + "SLAHMR" + ], + "image_footnote": [], + "bbox": [ + 635, + 271, + 661, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21bd141faa0eb986fb6901c57fbfa0b36b452804fb648cee1717e232c87589d7.jpg", + "image_caption": [ + "WHAM (Ours)" + ], + "image_footnote": [], + "bbox": [ + 753, + 271, + 779, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1bcb5be1c007b058b27d63decabfdb1aef80b7fea3189cb61302638df128f592.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 301, + 885, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 566, + 312, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The estimation of 3D human motion from video has progressed rapidly but current methods still have several key limitations. First, most methods estimate the human in camera coordinates. Second, prior work on estimating humans in global coordinates often assumes a flat ground plane and produces foot sliding. Third, the most accurate methods rely on computationally expensive optimization pipelines, limiting their use to offline applications. Finally, existing video-based methods are surprisingly less accurate than single-frame methods. We address these limitations with WHAM (World-grounded Humans with Accurate Motion), which accurately and efficiently reconstructs 3D human motion in a global coordinate system from video. WHAM learns to lift 2D keypoint sequences to 3D using motion capture data and fuses this with video features, integrating motion context and visual information. WHAM exploits camera angular velocity estimated from a SLAM method together with human motion to estimate the body's global trajectory. We combine this with a contact-aware trajectory refinement method that", + "bbox": [ + 73, + 613, + 473, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lets WHAM capture human motion in diverse conditions, such as climbing stairs. WHAM outperforms all existing 3D human motion recovery methods across multiple in-the-wild benchmarks. Code is available for research purposes at http://wham.is.tue.mpg.de/.", + "bbox": [ + 498, + 568, + 893, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 676, + 632, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our goal is to accurately estimate the 3D pose and shape of a person from monocular video. This is a longstanding problem and, while the field has made rapid progress, several key challenges remain. First, human motion should be computed in a consistent global coordinate system. Second, the method should be computationally efficient. Third, the results should be accurate, temporally smooth, detailed, natural looking, and have realistic foot-ground contact. Fourth, the capture should work with an arbitrary moving camera. These constraints need to be satisfied to make markerless human motion capture widely available for applications in gaming, AR/VR, autonomous driving, sports analysis, and human-robot interaction. We address these challenges with", + "bbox": [ + 496, + 704, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2070", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "WHAM (World-grounded Humans with Accurate Motion), which enables fast and accurate recovery of 3D human motion from a moving camera; see Fig. 1.", + "bbox": [ + 76, + 90, + 470, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It seems natural that, in estimating 3D humans from video, we should be able to exploit the temporal nature of video. Counter-intuitively, existing video-based methods for 3D human pose and shape (HPS) estimation [5, 14, 16, 29, 42, 50] are less accurate than the best single-frame methods [6, 13, 17, 20, 22, 24, 25, 57, 57]. This may be an issue of training data. There are large datasets of single images with ground-truth 3D human poses containing a diversity of body shapes, poses, backgrounds, lighting, etc. In contrast, video datasets with ground truth are much more limited.", + "bbox": [ + 75, + 138, + 468, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, WHAM leverages both the large-scale AMASS motion capture (mocap) dataset [31] and video datasets. Our key idea is to learn about 3D human motion from AMASS and then learn to fuse this information with temporal image cues from video, getting the best of both. Similar to previous work [58, 59], we use AMASS to generate synthetic 2D keypoints and ground-truth motion sequence pairs, from which we pretrain a motion encoder, which captures the motion context, and decoder that lifts sequences of 2D keypoints to sequences of 3D poses. Given the robustness of recent 2D keypoint detection models [52, 56], our pretrained model does a good job of predicting human pose from video.", + "bbox": [ + 75, + 291, + 470, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keypoints alone, however, are too sparse for accurate 3D mesh estimation. To improve accuracy, we jointly train a feature integrator network that merges information from video and 2D-keypoint sequences. We use a pretrained image encoder from previous work [1, 6, 20, 24] and train the feature integrator using video datasets [10, 14, 32, 47]. This integration process supplements the motion context extracted from the sparse 2D keypoints with dense visual context, significantly improving the recovered pose and shape accuracy.", + "bbox": [ + 75, + 489, + 468, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the above approach produces accurate motion, we want this motion in global coordinates, unlike most previous methods that compute the body in camera coordinates. Estimating the global human trajectory is challenging when the camera is moving because the motions of the body and the camera are entangled. Recent work addresses this with optimization fitting based on a learned human motion prior and camera information from SLAM methods [19, 40, 53] or dense 3D scene information from COLMAP [27]. However, these methods are computationally expensive and far from real time. Recent regression-based methods are faster but either constrain the problem with static or known camera conditions [41, 58] or have temporal jitter and limited accuracy [43]. We tackle this challenge with two additional modules. First, we predict the global orientation and root velocity of the human from the sequence of 2D keypoints by training a global trajectory decoder. Specifically, we concatenate the camera's angular velocity to the context and", + "bbox": [ + 75, + 628, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "train the global trajectory decoder to recursively predict the current orientation and root velocity, effectively factoring camera motion from human motion. WHAM takes the camera's angular velocity either from the output of a SLAM method or from a camera's gyroscope when available.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above solution relies on knowledge of human motion learned from AMASS. Therefore, it can fail to capture elevation changes when the surface is not flat, e.g. when ascending the stairs because AMASS has a limited amount of such data. To address this, we introduce foot contact as an additional explicit form of motion context. We train WHAM to predict the likelihood of foot-ground contact using estimated contact labels from both AMASS and 3D video datasets. We then train a trajectory refinement network that outputs an update to the root orientation and velocity based on the information about the foot contact/velocity. This refinement enables WHAM to accurately estimate human motion in a global coordinate system even when the terrain is not flat.", + "bbox": [ + 496, + 167, + 892, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "WHAM has very low computational overhead because it is an on-line algorithm that recursively predicts the pose, shape, and global motion parameters. The network, excluding preprocessing (bounding box detection, keypoint detection, and person identification), runs at 200 fps, significantly faster than prior methods. Also, despite not using global optimization like [53], we obtain accurate 3D camera trajectories and global body motions with minimal drift. Through extensive comparisons on multiple in-the-wild datasets as well as detailed ablation studies, we find that WHAM achieves state-of-the-art (SOTA) accuracy on 3D human pose estimation as well as global trajectory estimation (see Fig. 1).", + "bbox": [ + 496, + 363, + 893, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, in this paper we: (1) introduce the first approach to effectively fuse 3D human motion context and video context for 3D HPS regression; (2) propose a novel global trajectory estimation framework that leverages motion context and foot contact to effectively address foot sliding and enable the 3D tracking of people on non-planar surfaces; (3) efficiently perform HPS regression in global coordinates; (4) achieve state-of-the-art (SOTA) performance on multiple in-the-wild benchmark datasets (3DPW [47], RICH [9], EMDB [15]). WHAM is the first video-based method to outperform all image-based and video-based methods on per-frame accuracy while maintaining temporal smoothness.", + "bbox": [ + 496, + 544, + 893, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 738, + 640, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based 3D HPS Estimation. There are two broad classes of methods for recovering 3D HPS from images: model-free [21, 26, 33] and model-based [6, 11-13, 20, 24, 34, 35]. Here we focus on model-based methods, which estimate the low-dimensional parameters of a statistical body model [28, 36, 37, 51]. While early work explores optimization-based methods [2], here we focus on direct regression methods based on deep learning.", + "bbox": [ + 496, + 763, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many existing methods follow the architecture of HMR", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[13], which uses a pretrained backbone to predict image features followed by a multilayer perceptron (MLP) that regresses SMPL [28] pose parameters from image features. Training such networks typically leverages paired images with SMPL parameters; these parameters are often pseudo-groundtruth (p-GT), estimated from 2D keypoints [11, 20, 24, 34, 35]. Other architectures for HPS regression have also been proposed [6, 17, 18, 22, 23, 57]. None of these methods use video or estimate the body in global coordinates. While quite accurate, when these image-based models are applied independently to frames of a video sequence, the shape and pose can be temporally inconsistent. In contrast, WHAM effectively aggregates temporal information to provide frame-accurate and temporally-coherent 3D HPS estimation.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-Based 3D HPS. Video-based methods typically encode temporal information by combining static features extracted by a backbone from each frame. HMMR [14] uses a convolutional encoder, while VIBE [16] and MEVA [29] employ recurrent neural networks. TCMR [5] divides sequences into past, future, and whole frames, aggregating information to strongly constrain the output with motion consistency. MPS-Net [50] uses attention to capture non-local motion context and a hierarchical architecture to aggregate temporal features. Both MAED [48] and GLoT [42] use transformer architectures [46] to encode videos. MAED encodes videos in both temporal (across frames) and spatial (within each frame) dimensions and leverages the kinematic tree to iteratively regress each joint angle. GLoT encodes long-term temporal correlations and refines local details by focusing on nearby frames. Despite integrating information across frames, all existing video-based methods have lower accuracy than the best single-frame methods.", + "bbox": [ + 75, + 320, + 472, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given limited video training data with ground truth SMPL poses, several single-frame methods infer a mesh from 2D/3D keypoints [4, 7, 30, 33, 38] and use the keypoints as a proxy for training. Another set of approaches exploits 3D mocap data, which is plentiful [31], to train a network to lift 2D joints to 3D poses, which are used as a proxy for ground truth 3D. MotionBERT [59] synthesizes 2D keypoints through orthographic projection to learn a unified motion representation. ProxyCap [58] projects synthetic 3D keypoints into virtual cameras using a heuristic camera pose distribution. Despite benefiting from the scale of mocap datasets, these approaches do not fully utilize the visual information available in the video at test time. Here, we propose a combined network architecture and training strategy that leverages both proxy representations of human pose (lifting) and visual context extracted from video.", + "bbox": [ + 75, + 595, + 472, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global 3D Motion Estimation with Dense Sensors. Several methods augment video data with other sensors to estimate 3D HPS in world coordinates. The 3DPW dataset [47] employs pre-calibrated body-worn inertial sensors and", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a handheld camera to jointly optimize the camera and human motion in challenging environments. Similarly, the EMDB dataset [15] uses electromagnetic sensors with an RGB-D camera, enabling accurate human motion capture in the world. While body-worn sensors aid global human motion reconstruction, they are intrusive, require cooperation, and do not help with archival video. BodySLAM++ [8] uses an optimization method with a visual-inertial sensor, comprising stereo cameras and an IMU. In contrast, we use a standard monocular camera, balancing accessibility and accuracy without using specialized equipment. While WHAM can take the camera gyro as input, this is not required.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular Global 3D Human Trajectory Estimation. Estimating the global human trajectory from a monocular dynamic camera is challenging. Previous work relies on learned prior distributions of human motion to separate human motion from camera motion. GLAMR [55] computes the global trajectory based on a predicted and inflated 3D motion sequence and optimizes it across multiple individuals in the scene. However, since GLAMR does not consider camera motion cues, the output trajectory may be noisy when the camera is rotating. SLAHMR [53] and PACE [19] use off-the-shelf SLAM algorithms [44, 45] and jointly optimize the camera and human motion to minimize the negative log-likelihood of a learned motion prior [39]. While they achieve good results, their optimization approach is computationally expensive. TRACE [43] is a pure regression method that utilizes optical flow as a motion cue and estimates multiple people at once, but lacks temporal consistency. GloPro [41] regresses the uncertainty of the global human motion in real-time, but requires known camera poses. In contrast, WHAM leverages both explicit and implicit prior knowledge of human motion and efficiently reconstructs accurate and temporally coherent 3D human motion in world coordinates.", + "bbox": [ + 496, + 273, + 895, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 638, + 599, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 666, + 609, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our World-grounded Human with Accurate Motion (WHAM) framework is illustrated in Fig. 2. The input to WHAM is a raw video data $\\{I^{(t)}\\}_{t = 0}^{T}$ , captured by a camera with possibly unknown motion. Our goal is to predict the corresponding sequence of the SMPL model parameters $\\{\\Theta^{(t)}\\}_{t = 0}^{T}$ , as well as the root orientation $\\{\\Gamma^{(t)}\\}_{t = 0}^{T}$ and translation $\\{\\tau^{(t)}\\}_{t = 0}^{T}$ , expressed in the world coordinate system. We use ViTPose [52] to detect 2D keypoints $\\{x_{2D}^{(t)}\\}_{t = 0}^{T}$ from which we obtain motion features $\\{\\phi_m^{(t)}\\}_{t = 0}^T$ using the motion encoder. Additionally, we use a pretrained image encoder [6, 20, 24] to extract static image features $\\{\\phi_i^{(t)}\\}_{t = 0}^T$ and integrate them with $\\{\\phi_m^{(t)}\\}_{t = 0}^T$ to obtain fine-grained motion features $\\{\\hat{\\phi}_m^{(t)}\\}_{t = 0}^T$ from which we regress 3D human", + "bbox": [ + 496, + 691, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c1794a5e6d8edd55623e19d803efd1d0e3c8bdfa15231e1bd8783c3de45e9b88.jpg", + "image_caption": [ + "Figure 2. An Overview of WHAM. WHAM takes the sequence of 2D keypoints estimated by a pretrained detector and encodes it into a motion feature. WHAM then updates the motion feature using another sequence of image features extracted from the image encoder through the feature integrator. From the updated motion feature, the Local Motion Decoder estimates 3D motion in the camera coordinate system and foot-ground contact probability. The Trajectory Decoder takes the motion feature and camera angular velocity to initially estimate the global root orientation and egocentric velocity, which are then updated through the Trajectory Refiner using the foot-ground contact. The final output of WHAM is pixel-aligned 3D human motion with the 3D trajectory in the global coordinates." + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 851, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "motion in the world coordinate system.", + "bbox": [ + 76, + 375, + 336, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 400, + 282, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uni-directional Motion Encoder and Decoder. In contrast to existing methods [5, 29, 42, 50, 59], which use windows with a fixed time duration, we use uni-directional recurrent neural networks (RNN) for the motion encoder and decoder, making WHAM suitable for online inference. The objective of the motion encoder, $E_{M}$ , is to extract the motion context, $\\phi_{m}^{(t)}$ , from the current and previous sequence of 2D keypoints and the initial hidden state, $h_{E}^{(0)}$ :", + "bbox": [ + 75, + 424, + 470, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {m} ^ {(t)} = E _ {M} \\big (x _ {2 D} ^ {(0)}, x _ {2 D} ^ {(1)},..., x _ {2 D} ^ {(t)} | h _ {E} ^ {(0)} \\big).\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 559, + 395, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We normalize keypoints to a bounding box around the person and concatenate the box's center and scale to the keypoints, similar to CLIFF [24]. The role of the motion decoder, $D_{M}$ is to recover SMPL parameters, $(\\theta ,\\beta)$ , weak-perspective camera translation, $c$ , and foot-ground contact probability, $p$ from the motion feature history:", + "bbox": [ + 75, + 592, + 468, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\theta^ {(t)}, \\beta^ {(t)}, c ^ {(t)}, p ^ {(t)}\\right) = D _ {M} \\left(\\hat {\\phi} _ {m} ^ {(0)}, \\dots , \\hat {\\phi} _ {m} ^ {(t)} | h _ {D} ^ {(0)}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 694, + 431, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\hat{\\phi}_m^{(t)}$ is the motion feature integrated with the image feature $\\phi_i^{(t)}$ (described below). During pretraining on synthetic data, the image feature is not available and we set $\\hat{\\phi}_m^{(t)} = \\phi_m^{(t)}$ . As the encoder and decoder are tasked with reconstructing a dense 3D representation $\\Theta$ from a sparse 2D input signal $x_{2D}$ , we design an intermediate task to predict the 3D keypoints $x_{3D}$ and use them as the intermediate motion representation. This cascaded approach guides $\\phi_m$ to represent the implicit context of motion and the 3D spatial structure of the body. Similar to PIP [54], we use Neural Initialization that uses MLP to initialize the hidden state of", + "bbox": [ + 75, + 729, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the motion encoder and decoder, $\\left(h_{E}^{(0)}, h_{D}^{(0)}\\right)$ ; see section 4.3 and Sup. Mat. for details.", + "bbox": [ + 496, + 372, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion and Visual Feature Integrator. We use the AMASS dataset to synthetically generate 2D sequences by projecting 3D SMPL joints into images with varied camera motions. This provides effectively limitless training data that is far more diverse than existing video datasets that contain ground truth 3D pose and shape. Although we leverage the temporal human motion context, lifting 2D keypoints to 3D meshes is an ambiguous task. A key idea is to augment this 2D keypoint information with image cues that can help disambiguate the 3D pose. Specifically, we use an image encoder [1, 6, 20, 24], pretrained on the human mesh recovery task, to extract image features $\\phi_{i}$ , which contain dense visual contextual information related to the 3D human pose and shape. We then train a feature integrator network, $F_{I}$ , to combine $\\phi_{m}$ with $\\phi_{i}$ , integrating motion and visual context. The feature integrator uses a simple yet effective residual connection:", + "bbox": [ + 496, + 406, + 892, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\phi} _ {m} ^ {(t)} = \\phi_ {m} ^ {(t)} + F _ {I} \\left(\\operatorname {c o n c a t} \\left(\\phi_ {m} ^ {(t)}, \\phi_ {i} ^ {(t)}\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 660, + 823, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This supplements motion features pretrained on the 2D-to-3D lifting task using AMASS with visual context, resulting in enriched motion features that use image evidence to help disambiguate the task.", + "bbox": [ + 496, + 696, + 893, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Trajectory Decoder. We design an additional decoder, $D_{T}$ , to predict the rough global root orientation $\\Gamma_0^{(t)}$ and root velocity $v_0^{(t)}$ from the motion feature $\\phi_m^{(t)}$ . Since $\\phi_m$ is derived from the input signals in the camera coordinates, it is highly challenging to decouple the human and camera motion from it. To address this ambiguity, we append the angular velocity of the camera, $\\omega^{(t)}$ , to the motion feature, $\\phi_m^{(t)}$ , to create a camera-agnostic motion context. This design choice makes WHAM compatible with both off-the-shelf", + "bbox": [ + 496, + 758, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SLAM algorithms [44, 45] and gyroscope measurements that are widely available from modern digital cameras. We recursively predict global orientation, $\\Gamma_0^{(t)}$ , using the unidirectional RNN.", + "bbox": [ + 75, + 90, + 470, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\big (\\Gamma_ {0} ^ {(t)}, v _ {0} ^ {(t)} \\big) = D _ {T} \\big (\\phi_ {m} ^ {(0)}, \\omega^ {(0)}, \\dots , \\phi_ {m} ^ {(t)}, \\omega^ {(t)} \\big).\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 161, + 419, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contact Aware Trajectory Refinement. Good 3D motion in world coordinates in most scenarios implies accurate foot-ground contact without sliding. We want WHAM to generalize beyond flat ground planes, which are typically assumed in prior work. Specifically, our new trajectory refiner aims to resolve foot sliding and enables WHAM to generalize well to diverse motions, including climbing stairs. The refinement involves two stages. First, we adjust the ego-centric root velocity to $\\tilde{v}^{(t)}$ to minimize foot sliding, based on the foot-ground contact probability $p^{(t)}$ , estimated from the motion decoder $D_M$ :", + "bbox": [ + 75, + 191, + 470, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {v} ^ {(t)} = v _ {0} ^ {(t)} - \\left(\\Gamma_ {0} ^ {(t)}\\right) ^ {- 1} \\bar {v} _ {f} ^ {(t)},\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 367, + 364, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{v}_f^{(t)}$ is the averaged velocity of the toes and heels in world coordinates when their contact probability, $p^{(t)}$ , is higher than a threshold. However, this velocity adjustment often introduces noisy translation when the contact and pose estimation is inaccurate. Therefore, we propose a simple learning mechanism in which a trajectory refining network, $R_{T}$ , updates the root orientation and velocity to address this issue. Finally, the global translation is computed through a roll-out operation:", + "bbox": [ + 75, + 401, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\Gamma^ {(t)}, v ^ {(t)}\\right) = R _ {T} \\big (\\phi_ {m} ^ {(0)}, \\Gamma_ {0} ^ {(0)}, \\tilde {v} ^ {(0)},..., \\phi_ {m} ^ {(t)}, \\Gamma_ {0} ^ {(t)}, \\tilde {v} ^ {(t)} \\big), \\\\ \\tau^ {(t)} = \\sum_ {i = 0} ^ {t - 1} \\Gamma^ {(i)} v ^ {(i)}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 551, + 460, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, this full process reconstructs accurate 3D human pose and shape in both the camera and world coordinates from a single monocular video sequence (Fig. 2).", + "bbox": [ + 75, + 626, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 76, + 680, + 181, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining on AMASS. We train in two stages: (1) pretraining with synthetic data, and (2) fine-tuning with real data (Fig. 3). The objective of the pretraining stage is to teach the motion encoder to extract motion context from the input 2D keypoint sequence. The motion and trajectory decoders then map this motion context to the corresponding 3D motion and global trajectory spaces (i.e. they lift the encoding to 3D). We use the AMASS dataset [31] to generate an extensive set of synthetic pairs consisting of sequences of 2D keypoints together with the ground truth SMPL parameters.", + "bbox": [ + 75, + 703, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To synthesize 2D keypoints from AMASS, we create virtual cameras onto which we project 3D keypoints derived from the ground truth mesh. Unlike MotionBERT [59] and", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7969ce51517c8e6e1f5aeb41fee4efc9565047caa0171225060958dbbd635b69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 92, + 888, + 169 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/392b61eee0dc7ddbda862f69f72d090dd8d6d7edf4ec41ebb641348740e34203.jpg", + "image_caption": [ + "Figure 3. WHAM's Two-Stage Training Scheme. During pre-taining, we generate synthetic 2D keypoint sequences from AMASS [31] and train a motion encoder and decoder on the generated data (top). We then leverage video datasets with ground truth SMPL parameters, for which there is much less data. We use the fixed-weight pre-trained image encoder and keypoints detector $(\\text{※})$ to extract image features and 2D keypoints. In this stage, we train the feature integration network while fine-tuning the motion encoder and motion/trajectory decoders, marked $\\text{※}$ (bottom)." + ], + "image_footnote": [], + "bbox": [ + 504, + 170, + 888, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ProxyCap [58], which use static cameras for keypoint projection, we employ dynamic cameras that incorporate both rotational and translational motion. This choice has two main motivations. First, it accounts for the inherent differences between human motions captured in static and dynamic camera setups. Second, it enables the learning of a camera-agnostic motion representation, from which the trajectory decoder can reconstruct the global trajectory. We also augment the 2D data with noise and masking. For details of the synthetic generation process see Sup. Mat.", + "bbox": [ + 496, + 401, + 892, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-tuning on Video Datasets. Starting with the pretrained network, we fine-tune WHAM on four video datasets: 3DPW [47], Human3.6M [10], MPI-INF-3DHP [32], and InstaVariety [14]. For the human mesh recovery task, we supervise WHAM on ground-truth SMPL parameters from AMASS and 3DPW, 3D keypoints from Human3.6M and MPI-INF-3DHP, and 2D keypoints from InstaVariety. For the global trajectory estimation task, we use AMASS, Human3.6M, and MPI-INF-3DHP. Additionally, during training we experiment with adding BEDLAM [1], a large synthetic dataset with realistic video and ground truth SMPL parameters.", + "bbox": [ + 496, + 551, + 893, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The fine-tuning has two objectives: 1) exposing the network to real 2D keypoints, instead of training it solely on synthetic data, and 2) training the feature integrator network to aggregate motion and image features. To achieve these goals, we jointly train the entire network on the video datasets while setting a smaller learning rate on the pretrained modules (see Fig. 3). Consistent with prior work [5, 16, 29, 42, 50], we employ a pretrained and fixed-weight image encoder [20] to extract image features. However, to leverage recent network architectures and training strategies, we also experiment with different types of encoders [1, 6, 24] in the following section. Implementation Details. For the pretraining stage, we train", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WHAM on AMASS for 80 epochs with the learning rate of $5 \\times 10^{-4}$ . Then we finetune WHAM on 3DPW, MPI-INF-3DHP, Human3.6M, and InstaVariety for 30 epochs. We use a learning rate of $1 \\times 10^{-4}$ for the feature integrator and $1 \\times 10^{-5}$ for the pretrained components during finetuning. During training, we use the Adam optimizer and a batch size of 64 and split sequences into 81-frame chunks.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 208, + 209, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate WHAM on three in-the-wild benchmarks: 3DPW [47], RICH [9], and EMDB [15]. Following previous work [1, 5, 13, 16, 20, 24, 29], we perform the evaluation in camera coordinates. The estimated global trajectory is evaluated on a subset of EMDB (EMDB 2) for which they provide ground truth global motion with dynamic cameras (used for evaluation). We also test on new sequences captured using an iPhone with the gyroscope. See Sup. Mat. for more details on the datasets and iPhone results.", + "bbox": [ + 75, + 233, + 472, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. To evaluate the accuracy of 3D human pose and shape estimation, we compute Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PA-MPJPE), and Per Vertex Error (PVE) measured in millimeters $(mm)$ . We compute Acceleration error (Accel, in $m / s^2$ ) to measure the inter-frame smoothness of the reconstructed motion. We also evaluate the motion reconstruction and trajectory estimation accuracy in the world-frame. Following previous work [19, 53], we split sequences into smaller segments of 100 frames and align each output segment with the ground-truth data using the first two frames $\\mathrm{(W - MPJPE_{100})}$ or the entire segment $\\mathrm{(WA - MPJPE_{100})}$ in mm. These previous metrics give an unrealistic picture of 3D performance as they do not measure drift over long sequences. Therefore, we also evaluate the error over the entire trajectory after the rigid alignment and measure Root Translation Error (RTE in %) normalized by the actual displacement of the person. We also assess the jitter of the motion in the world coordinate system in $10m / s^3$ and foot sliding during the contact (FS in mm).", + "bbox": [ + 75, + 369, + 473, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. 3D Human Motion Recovery", + "text_level": 1, + "bbox": [ + 76, + 678, + 333, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Per-frame accuracy. Table 1 presents a comprehensive comparison of WHAM and the existing state-of-the-art per-frame and video-based methods across three benchmark datasets [9, 15, 47]. Because none of the methods are exposed to data from RICH or EMDB during training, results on these datasets are indicative of each method's ability to generalize. WHAM (Res), WHAM (HR), and WHAM (ViT) correspond to different architectures for the pretrained image encoders, derived from SPIN (ResNet-50) [20], CLIFF (HRNet-W48) [1, 24], and HMR2.0 (ViT-H/16) [6], respectively. Not surprisingly, WHAM (HR) is more accurate than WHAM (Res),", + "bbox": [ + 75, + 700, + 470, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous work follows [14] in reporting Accel in $mm / \\text{frame}^2$ . To remove the dependency on frame rate, we convert all previous results to $m / s^2$ .", + "bbox": [ + 76, + 875, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e78d245f4a53bc4d530334e5d581d76085a986972024f3f22c9b8864f7e6c9a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 89, + 581, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cddb34224c4ffaa183f9d634e571b3e2821a426e0e1ef72268b37d680accbf65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 89, + 656, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db656ab550ff811dbb5d14f2723b23f772b914999419f8b23c57a16056e2277a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 90, + 728, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1895002a458b99197ef2a95b6f752f200fbeffd1ca827bbcda74092f79670b6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 89, + 803, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30be85eceac94604ab062647e1b1eca513fca27cc6bd38d7e3eda92a730824fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 89, + 880, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20477e4785edbd57c64195ccc0e3fefd676f995afcd2bb721c9f08a40ba31b8e.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with previous state-of-the-art methods for 3D human pose and shape estimation. See text." + ], + "image_footnote": [], + "bbox": [ + 513, + 176, + 870, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while the transformer-based version, WHAM (ViT), is the most accurate. The backbone matters, with WHAM (ViT) outperforming all previous methods on all per-frame metrics (MPJPE, PA-MPJPE, and PVE) on all benchmarks. Even with the simplest ResNet backbone, WHAM (Res) outperforms every method except for ReFit on RICH.", + "bbox": [ + 496, + 364, + 893, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training on BEDLAM consistently improves accuracy in prior work (see [1, 3, 49]), and we find the same here as shown in Table 2. However, compared to ReFit, WHAM exhibits relatively smaller performance improvement by adding BEDLAM. One possible reason for this is that, unlike ReFit, we did not finetune the image encoder during training to limit training time.", + "bbox": [ + 496, + 455, + 893, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inter-frame smoothness. We also evaluate the inter-frame smoothness using the acceleration error. Compared with state-of-the-art per-frame methods [1, 6, 22, 24], WHAM has significantly lower acceleration error. This indicates that WHAM reconstructs smooth and more plausible 3D human motion across frames while not sacrificing high per-frame accuracy. On the other hand, when compared to recent temporal methods [5, 42, 50], WHAM exhibits comparable or slightly higher acceleration error. However, we observe that these video-based methods tend to over-smooth the human motion, resulting in lower accuracy on per-frame metrics.", + "bbox": [ + 496, + 563, + 893, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To provide intuition for these numbers, we qualitatively compare WHAM with TCMR [5] and GLoT [42] in Fig. 4. While producing smooth results, TCMR and GLoT fail to capture the bending of the left knee when the subject is ascending the stairs, while WHAM more accurately reconstructs the 3D human pose.", + "bbox": [ + 496, + 729, + 893, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. 3D Global Trajectory Recovery", + "text_level": 1, + "bbox": [ + 500, + 830, + 774, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate global trajectory recovery, we compare WHAM with the state-of-the-art methods and a baseline that combines a SLAM method (DPVO [45]) and a per-frame", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa320a60d3acd2892ff715e85d8ff8761203f6fbd184c8e772af8fcba63b56eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Models3DPW (14)RICH (24)EMDB (24)
PA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccel
per-frameSPIN [20]59.296.9112.831.469.7122.9144.235.287.1140.7166.141.3
PARE* [17]46.574.588.6-60.7109.2123.5-72.2113.9133.2-
CLIFF* [24]43.069.081.222.556.6102.6115.022.468.3103.5123.724.5
HybrIK* [22]41.871.682.3-56.496.8110.4-65.6103.0122.2-
HMR2.0 [6]44.469.882.218.148.196.0110.918.860.798.3120.819.9
ReFit* [49]40.565.375.118.547.980.792.917.158.688.0104.520.7
temporalTCMR* [5]52.786.5101.46.065.6119.1137.75.079.8127.7150.25.3
VIBE* [16]51.982.998.418.568.4120.5140.221.881.6126.1149.926.5
MPS-Net* [50]52.184.399.06.567.1118.2136.75.881.4123.3143.96.2
GLoT* [42]50.680.796.46.065.6114.3132.75.279.1119.9140.85.4
GLAMR [55]51.1--8.079.9--107.773.8113.8134.933.0
TRACE* [43]50.979.195.428.6----71.5110.0129.625.5
SLAHMR [53]55.9---52.5--9.469.793.7111.37.1
PACE [19]----49.3--8.8----
WHAM (Res)*40.262.775.16.351.889.5103.25.057.884.099.75.2
WHAM (HR)*39.062.674.86.449.184.696.45.257.185.7103.25.6
WHAM (ViT)*35.957.868.76.644.380.091.25.350.479.794.45.3
", + "bbox": [ + 160, + 90, + 805, + 414 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a6eb3b90b669912d09dba271b938e435aef15ca23b98623bfe7d60c7baaa878a.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of state-of-the-art models on the 3DPW [47], RICH [9], and EMDB [15] datasets. Ordering of per-frame and temporal methods is done separately by descending MPJPE on EMDB (except for PACE). For testing on EMDB, we follow the protocol of EMDB 1 [15]. Parenthesis denotes the number of body joints used to compute MPJPE and PA-MPJPE, and * denotes models trained with the 3DPW training set. Bold numbers denote the most accurate method in each column. Accel is in $m / s^2$ , all other errors are in mm." + ], + "table_footnote": [], + "table_body": "
ModelsDataset3DPW (14)
PA-MPJPEMPJPEPVEAccel
CLIFF [1, 24]R43.668.882.119.2
R+B43.066.978.531.0
ReFit [49]R40.565.375.118.5
R+B38.257.667.621.4
WHAM (ViT)R35.957.868.76.6
R+B35.756.967.46.7
", + "bbox": [ + 107, + 503, + 434, + 625 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/30c7cb09d2354554a3d8761707ec2c379c8f33c01670b8ae8cdc76aefa6c9899.jpg", + "table_caption": [ + "Table 2. Dataset ablation experiments on 3DPW [47]. R denotes the use of real datasets and B denotes BEDLAM." + ], + "table_footnote": [], + "table_body": "
ModelsEMDB 2
WA-MPJPE100W-MPJPE100RTEJitterFS
DPVO (+ HMR2.0) [6, 45]647.82231.415.8537.3107.6
GLAMR [55]280.8726.611.446.320.7
TRACE [43]529.01702.317.72987.6370.7
SLAHMR [53]326.9776.110.231.314.5
WHAM (w/DPVO [45])135.6354.86.022.54.4
WHAM (w/DROID [44])133.3343.94.621.54.4
WHAM (w/ GT gyro)131.1335.34.121.04.4
", + "bbox": [ + 81, + 676, + 467, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Global motion estimation accuracy on EMDB [15].", + "bbox": [ + 93, + 799, + 450, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method (HMR2.0 [6]); see Table 3. WHAM is agnostic to the source of the camera angular velocity and we compare results using DPVO, DROID-SLAM [44] and the ground truth angular velocity (gyro).", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 3, WHAM outperforms the existing methods on all metrics. Specifically, combining WHAM with DPVO is more accurate than the global trajectory estimation of DPVO combined with HMR2.0, illustrating that our method actively refines the global trajectory instead of performing a simple integration. DROID-SLAM gives slightly better results than DPVO. Furthermore, WHAM significantly outperforms the regression-based method, TRACE, on jitter and foot sliding metrics. We further demonstrate this in Figs. 5 and 1, where WHAM captures more consistent and plausible human motion in the global coordinate system than TRACE and SLAHMR for videos captured by dynamic cameras. As depicted in Fig. 6, WHAM outperforms GLAMR, TRACE, and SLAHMR in capturing the pattern of human motion in the global coordinate system.", + "bbox": [ + 496, + 503, + 893, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 739, + 653, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To provide further insight into our approach, we conduct ablation studies to analyze the contribution of each component to the performance. As shown in Table 4, our entire system (WHAM) outperforms the different variants of WHAM that ablate a single component. To be specific, we first observe that adding feature integration improves both motion and global trajectory estimation accuracy when compared with an ablated version without feature integration $(\\mathrm{w / o}F_I)$ . Similarly, the removal of the pretraining on", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2076", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/182970dcdeee0d1bf6d9a67512ac5f6d202d89b4dfbaafa7646a53c8ef95001e.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with TRACE [43] and SLAHMR [53] on global human motion estimation with dynamic cameras." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 460, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06647819792b926e3c1cea4ad54a3ec6ec7a526c29b5544816ef00e894718ba5.jpg", + "image_caption": [ + "Figure 6. Comparison of global trajectory estimation on EMDB [15]. Overall, WHAM shows better alignment to ground truth data compared to GLAMR [55], TRACE [43], and SLAHMR [53]." + ], + "image_footnote": [], + "bbox": [ + 78, + 441, + 467, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the 2D-to-3D lifting task using AMASS [31] (w/o lifting) shows significant performance degradation. WHAM also outperforms the ablation of the Neural Initialization (w/o NI), particularly in the MPJPE metric. In addition, we experiment with WHAM to decode trajectory solely based on the motion context without using the estimated camera angular velocity $(\\mathrm{w / o}\\omega)$ . Although this version shows similar performance in predicting 3D human pose, it suffers from the entanglement of camera and human motion, resulting in significantly high overall trajectory error (RTE). Last, we observe that WHAM without the trajectory refinement (w/o traj. ref.) gives larger global trajectory and foot sliding errors in return for less jitter, indicating that our approach contributes to the global trajectory accuracy and helps reduce foot sliding. Core details are presented here; see Sup. Mat. for more details and information on run-time cost.", + "bbox": [ + 75, + 590, + 472, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "WHAM is a new method to recover accurate 3D human motion in global coordinates from a moving camera more ef", + "bbox": [ + 76, + 869, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5cd46d0d8745a81d652dde20e53d183ccd31b92aaba5218b0870942a84a4f109.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsEMDB 2
PA-MPJPEMPJPEWA-MPJPE100W-MPJPE100RTEJitterFS
w/o F144.269.0147.6377.96.323.15.5
w/o lifting60.383.0238.0693.011.524.55.0
w/o NI40.466.3142.7368.16.822.34.6
w/o ω39.162.0156.5422.010.122.15.0
w/o traj. ref.38.259.3154.7407.56.318.86.5
WHAM (Ours)38.259.3135.6354.86.022.54.4
", + "bbox": [ + 504, + 90, + 890, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation experiments. See text.", + "bbox": [ + 575, + 190, + 815, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ficiently and accurately than the state-of-the-art approaches. Our approach leverages the AMASS dataset to train a network to recursively lift 2D sequences of keypoints to sequences of 3D SMPL parameters. But keypoints alone lack valuable information about the body and its movement. Consequently, we integrate image context information over time and learn to combine it with the motion context to better estimate human body shape and pose. Additionally, our method takes an estimate of the camera angular velocity, which can either be computed from a SLAM method or from the camera's gyro when available. Finally, we combine all this information with an estimate of foot contact to recover the 3D human motion in global coordinates from a monocular video sequence. WHAM significantly outperforms the existing state-of-the-art methods (both image-based and video-based) on challenging in-the-wild benchmarks in both 3D HPS and the world-coordinate trajectory estimation accuracy. Because of its speed and accuracy, WHAM provides a foundation for in-the-wild motion capture applications.", + "bbox": [ + 496, + 229, + 893, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future directions: WHAM learns about human motion from AMASS, limiting generalization to motions that are out of distribution. While we employ random masking as part of our data synthesis process, our generating approach mainly assumes the scenario where the full body is within the field of view. See Sup. Mat. for more details.", + "bbox": [ + 496, + 523, + 893, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "WHAM opens up many directions for future work. For example, while we use SLAM to estimate the camera's angular velocity, SLAM could also provide camera intrinsics and extrinsics as well as information about the 3D scene that could be used to enforce consistency between the scene and the human. While WHAM is an online method, designed for real-time applications, it could also initialize an optimization-based post-processing akin to bundle adjustment, which would optimize the camera, scene, and human motion together. Furthermore, a real-time and phone-based implementation of WHAM should be feasible.", + "bbox": [ + 496, + 616, + 893, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Part of this work was done when the first author was an intern at the Max Planck Institute for Intelligence System. This work was partially supported through an NSF CAREER Award (CBET 2145473) and a Chan-Zuckerberg Essential Open Source Software for Science Award. CoI Disclosure. https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt.", + "bbox": [ + 496, + 787, + 895, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8726-8737, 2023. 2, 4, 5, 6, 7", + "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016, pages 561-578. Springer International Publishing, 2016. 2", + "[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. SMPLer-X: Scaling up expressive human pose and shape estimation. In Proc. Neural Information Processing Systems (NeurIPS), 2023. 6", + "[4] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020. 3", + "[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 4, 5, 6, 7", + "[6] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091, 2023. 2, 3, 4, 5, 6, 7", + "[7] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[8] Dorian F. Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. *Bodyslam++*: Fast and tightly-coupled visual-inertial camera and human motion tracking, 2023. 3", + "[9] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 2, 6, 7", + "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 5", + "[11] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020. 2, 3", + "[12] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020.", + "[13] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 6", + "[14] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. arXiv preprint arXiv:1812.01601, 2018. 2, 3, 5, 6", + "[15] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 6, 7, 8", + "[16] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5252-5262. IEEE, 2020. 2, 3, 5, 6, 7", + "[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2, 3, 7", + "[18] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11015-11025, Piscataway, NJ, 2021. IEEE. 3", + "[19] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2, 3, 6, 7", + "[20] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2, 3, 4, 5, 6, 7", + "[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 2", + "[22] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021. 2, 3, 6, 7", + "[23] Jiefeng Li, Siyuan Bian, Qi Liu, Jiasheng Tang, Fan Wang, and Cewu Lu. NIKI: Neural inverse kinematics with invertible neural networks for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[24] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2, 3, 4, 5, 6, 7", + "[25] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 2" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kevin Lin, Chung-Ching Lin, Lin Liang, Zicheng Liu, and Lijuan Wang. Mpt: Mesh pre-training with transformers for human pose and mesh reconstruction, 2023. 2", + "[27] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, 2021. 2", + "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3", + "[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3, 4, 5, 6", + "[30] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 534-543, 2023. 3", + "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 5, 8", + "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3D Vision (3DV), 2017 Fifth International Conference on. IEEE, 2017. 2, 5", + "[33] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2, 3", + "[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2, 3", + "[35] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudo-gts of 3d human mesh estimation in the wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2, 3", + "[36] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Computer Vision - ECCV 2020, pages 598-613. Springer, 2020. 2", + "[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2", + "[38] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459-468, 2019. 3", + "[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3", + "[40] Nitin Saini, Chun-Hao P. Huang, Michael J. Black, and Aamir Ahmad. Smartmocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, pages 1-8, 2023. 2", + "[41] Simon Schaefer, Dorian F. Henning, and Stefan Leutenegger. Glopro: Globally-consistent uncertainty-aware 3d human pose estimation & tracking in the wild, 2023. 2, 3", + "[42] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 2, 3, 4, 5, 6, 7", + "[43] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 7, 8", + "[44] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 5, 7", + "[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 3, 5, 6, 7", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3", + "[47] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 5, 6, 7", + "[48] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In The IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[49] Yufu Wang and Kostas Daniilidis. Refit: Recurrent fitting network for 3d human recovery. In International Conference on Computer Vision, 2023. 6, 7", + "[50] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 7", + "[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: generative 3D human shape and articulated pose models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 2", + "[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2079", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation. In Advances in Neural Information Processing Systems, 2022. 2, 3", + "[53] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8", + "[54] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4", + "[55] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 7, 8", + "[56] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[57] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3", + "[58] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Hongwei Yi, Shengping Zhang, and Yebin Liu. Real-time monocular full-body capture in world space via sequential proxy-to-motion learning. arXiv preprint arXiv:2307.01200, 2023. 2, 3, 5", + "[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 5" + ], + "bbox": [ + 78, + 90, + 468, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2080", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_model.json b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ff964eb7ab14e0efca93aff9be0bfde1c97a811 --- /dev/null +++ b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.152 + ], + "angle": 0, + "content": "WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.181, + 0.304, + 0.2 + ], + "angle": 0, + "content": "Soyong Shin" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.181, + 0.451, + 0.2 + ], + "angle": 0, + "content": "Juyong Kim" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.181, + 0.588, + 0.199 + ], + "angle": 0, + "content": "Eni Halilaj" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.181, + 0.773, + 0.198 + ], + "angle": 0, + "content": "Michael J. Black" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.209, + 0.79, + 0.245 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)Max Planck Institute for Intelligent System \\(\\{\\) soyongs, juyongk, ehalilaj\\} \\(@\\) andrew.cmu.edu black@tuebingen.mpg.de" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.258, + 0.158, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.36, + 0.157, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.259, + 0.236, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.36, + 0.237, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.259, + 0.315, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.36, + 0.314, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.259, + 0.393, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.36, + 0.392, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.259, + 0.419, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.36, + 0.418, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.259, + 0.499, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.36, + 0.499, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.259, + 0.419, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.36, + 0.418, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.272, + 0.499, + 0.285 + ], + "angle": 0, + "content": "GT trajectory" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.272, + 0.553, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.273, + 0.601, + 0.284 + ], + "angle": 0, + "content": "TRACE" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.272, + 0.663, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.273, + 0.722, + 0.284 + ], + "angle": 0, + "content": "SLAHMR" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.272, + 0.781, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.272, + 0.867, + 0.285 + ], + "angle": 0, + "content": "WHAM (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.302, + 0.887, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.47, + 0.893, + 0.541 + ], + "angle": 0, + "content": "Figure 1. WHAM: World-grounded Humans with Accurate Motion. State-of-the-art methods like TRACE [43] and SLAHMR [53] fail to capture global 3D human trajectories accurately when given in-the-wild videos captured using a moving camera, producing implausible world-grounded motion (e.g., foot sliding). To address this, WHAM uses two novel strategies: (1) feature integration from 2D keypoints and pixels to reconstruct precise and pixel-aligned 3D human motion and (2) contact-aware trajectory recovery to place the human in a global coordinate system without foot sliding. Gray dots show the ground-truth global trajectory. See Supplemental Video." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.568, + 0.313, + 0.583 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.614, + 0.474, + 0.901 + ], + "angle": 0, + "content": "The estimation of 3D human motion from video has progressed rapidly but current methods still have several key limitations. First, most methods estimate the human in camera coordinates. Second, prior work on estimating humans in global coordinates often assumes a flat ground plane and produces foot sliding. Third, the most accurate methods rely on computationally expensive optimization pipelines, limiting their use to offline applications. Finally, existing video-based methods are surprisingly less accurate than single-frame methods. We address these limitations with WHAM (World-grounded Humans with Accurate Motion), which accurately and efficiently reconstructs 3D human motion in a global coordinate system from video. WHAM learns to lift 2D keypoint sequences to 3D using motion capture data and fuses this with video features, integrating motion context and visual information. WHAM exploits camera angular velocity estimated from a SLAM method together with human motion to estimate the body's global trajectory. We combine this with a contact-aware trajectory refinement method that" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.569, + 0.895, + 0.645 + ], + "angle": 0, + "content": "lets WHAM capture human motion in diverse conditions, such as climbing stairs. WHAM outperforms all existing 3D human motion recovery methods across multiple in-the-wild benchmarks. Code is available for research purposes at http://wham.is.tue.mpg.de/." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.678, + 0.633, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Our goal is to accurately estimate the 3D pose and shape of a person from monocular video. This is a longstanding problem and, while the field has made rapid progress, several key challenges remain. First, human motion should be computed in a consistent global coordinate system. Second, the method should be computationally efficient. Third, the results should be accurate, temporally smooth, detailed, natural looking, and have realistic foot-ground contact. Fourth, the capture should work with an arbitrary moving camera. These constraints need to be satisfied to make markerless human motion capture widely available for applications in gaming, AR/VR, autonomous driving, sports analysis, and human-robot interaction. We address these challenges with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2070" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "WHAM (World-grounded Humans with Accurate Motion), which enables fast and accurate recovery of 3D human motion from a moving camera; see Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.29 + ], + "angle": 0, + "content": "It seems natural that, in estimating 3D humans from video, we should be able to exploit the temporal nature of video. Counter-intuitively, existing video-based methods for 3D human pose and shape (HPS) estimation [5, 14, 16, 29, 42, 50] are less accurate than the best single-frame methods [6, 13, 17, 20, 22, 24, 25, 57, 57]. This may be an issue of training data. There are large datasets of single images with ground-truth 3D human poses containing a diversity of body shapes, poses, backgrounds, lighting, etc. In contrast, video datasets with ground truth are much more limited." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.488 + ], + "angle": 0, + "content": "To address this, WHAM leverages both the large-scale AMASS motion capture (mocap) dataset [31] and video datasets. Our key idea is to learn about 3D human motion from AMASS and then learn to fuse this information with temporal image cues from video, getting the best of both. Similar to previous work [58, 59], we use AMASS to generate synthetic 2D keypoints and ground-truth motion sequence pairs, from which we pretrain a motion encoder, which captures the motion context, and decoder that lifts sequences of 2D keypoints to sequences of 3D poses. Given the robustness of recent 2D keypoint detection models [52, 56], our pretrained model does a good job of predicting human pose from video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Keypoints alone, however, are too sparse for accurate 3D mesh estimation. To improve accuracy, we jointly train a feature integrator network that merges information from video and 2D-keypoint sequences. We use a pretrained image encoder from previous work [1, 6, 20, 24] and train the feature integrator using video datasets [10, 14, 32, 47]. This integration process supplements the motion context extracted from the sparse 2D keypoints with dense visual context, significantly improving the recovered pose and shape accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.901 + ], + "angle": 0, + "content": "While the above approach produces accurate motion, we want this motion in global coordinates, unlike most previous methods that compute the body in camera coordinates. Estimating the global human trajectory is challenging when the camera is moving because the motions of the body and the camera are entangled. Recent work addresses this with optimization fitting based on a learned human motion prior and camera information from SLAM methods [19, 40, 53] or dense 3D scene information from COLMAP [27]. However, these methods are computationally expensive and far from real time. Recent regression-based methods are faster but either constrain the problem with static or known camera conditions [41, 58] or have temporal jitter and limited accuracy [43]. We tackle this challenge with two additional modules. First, we predict the global orientation and root velocity of the human from the sequence of 2D keypoints by training a global trajectory decoder. Specifically, we concatenate the camera's angular velocity to the context and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "train the global trajectory decoder to recursively predict the current orientation and root velocity, effectively factoring camera motion from human motion. WHAM takes the camera's angular velocity either from the output of a SLAM method or from a camera's gyroscope when available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.362 + ], + "angle": 0, + "content": "The above solution relies on knowledge of human motion learned from AMASS. Therefore, it can fail to capture elevation changes when the surface is not flat, e.g. when ascending the stairs because AMASS has a limited amount of such data. To address this, we introduce foot contact as an additional explicit form of motion context. We train WHAM to predict the likelihood of foot-ground contact using estimated contact labels from both AMASS and 3D video datasets. We then train a trajectory refinement network that outputs an update to the root orientation and velocity based on the information about the foot contact/velocity. This refinement enables WHAM to accurately estimate human motion in a global coordinate system even when the terrain is not flat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.894, + 0.544 + ], + "angle": 0, + "content": "WHAM has very low computational overhead because it is an on-line algorithm that recursively predicts the pose, shape, and global motion parameters. The network, excluding preprocessing (bounding box detection, keypoint detection, and person identification), runs at 200 fps, significantly faster than prior methods. Also, despite not using global optimization like [53], we obtain accurate 3D camera trajectories and global body motions with minimal drift. Through extensive comparisons on multiple in-the-wild datasets as well as detailed ablation studies, we find that WHAM achieves state-of-the-art (SOTA) accuracy on 3D human pose estimation as well as global trajectory estimation (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.894, + 0.727 + ], + "angle": 0, + "content": "In summary, in this paper we: (1) introduce the first approach to effectively fuse 3D human motion context and video context for 3D HPS regression; (2) propose a novel global trajectory estimation framework that leverages motion context and foot contact to effectively address foot sliding and enable the 3D tracking of people on non-planar surfaces; (3) efficiently perform HPS regression in global coordinates; (4) achieve state-of-the-art (SOTA) performance on multiple in-the-wild benchmark datasets (3DPW [47], RICH [9], EMDB [15]). WHAM is the first video-based method to outperform all image-based and video-based methods on per-frame accuracy while maintaining temporal smoothness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.642, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Image-based 3D HPS Estimation. There are two broad classes of methods for recovering 3D HPS from images: model-free [21, 26, 33] and model-based [6, 11-13, 20, 24, 34, 35]. Here we focus on model-based methods, which estimate the low-dimensional parameters of a statistical body model [28, 36, 37, 51]. While early work explores optimization-based methods [2], here we focus on direct regression methods based on deep learning." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Many existing methods follow the architecture of HMR" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.318 + ], + "angle": 0, + "content": "[13], which uses a pretrained backbone to predict image features followed by a multilayer perceptron (MLP) that regresses SMPL [28] pose parameters from image features. Training such networks typically leverages paired images with SMPL parameters; these parameters are often pseudo-groundtruth (p-GT), estimated from 2D keypoints [11, 20, 24, 34, 35]. Other architectures for HPS regression have also been proposed [6, 17, 18, 22, 23, 57]. None of these methods use video or estimate the body in global coordinates. While quite accurate, when these image-based models are applied independently to frames of a video sequence, the shape and pose can be temporally inconsistent. In contrast, WHAM effectively aggregates temporal information to provide frame-accurate and temporally-coherent 3D HPS estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.473, + 0.593 + ], + "angle": 0, + "content": "Video-Based 3D HPS. Video-based methods typically encode temporal information by combining static features extracted by a backbone from each frame. HMMR [14] uses a convolutional encoder, while VIBE [16] and MEVA [29] employ recurrent neural networks. TCMR [5] divides sequences into past, future, and whole frames, aggregating information to strongly constrain the output with motion consistency. MPS-Net [50] uses attention to capture non-local motion context and a hierarchical architecture to aggregate temporal features. Both MAED [48] and GLoT [42] use transformer architectures [46] to encode videos. MAED encodes videos in both temporal (across frames) and spatial (within each frame) dimensions and leverages the kinematic tree to iteratively regress each joint angle. GLoT encodes long-term temporal correlations and refines local details by focusing on nearby frames. Despite integrating information across frames, all existing video-based methods have lower accuracy than the best single-frame methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.473, + 0.838 + ], + "angle": 0, + "content": "Given limited video training data with ground truth SMPL poses, several single-frame methods infer a mesh from 2D/3D keypoints [4, 7, 30, 33, 38] and use the keypoints as a proxy for training. Another set of approaches exploits 3D mocap data, which is plentiful [31], to train a network to lift 2D joints to 3D poses, which are used as a proxy for ground truth 3D. MotionBERT [59] synthesizes 2D keypoints through orthographic projection to learn a unified motion representation. ProxyCap [58] projects synthetic 3D keypoints into virtual cameras using a heuristic camera pose distribution. Despite benefiting from the scale of mocap datasets, these approaches do not fully utilize the visual information available in the video at test time. Here, we propose a combined network architecture and training strategy that leverages both proxy representations of human pose (lifting) and visual context extracted from video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Global 3D Motion Estimation with Dense Sensors. Several methods augment video data with other sensors to estimate 3D HPS in world coordinates. The 3DPW dataset [47] employs pre-calibrated body-worn inertial sensors and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "a handheld camera to jointly optimize the camera and human motion in challenging environments. Similarly, the EMDB dataset [15] uses electromagnetic sensors with an RGB-D camera, enabling accurate human motion capture in the world. While body-worn sensors aid global human motion reconstruction, they are intrusive, require cooperation, and do not help with archival video. BodySLAM++ [8] uses an optimization method with a visual-inertial sensor, comprising stereo cameras and an IMU. In contrast, we use a standard monocular camera, balancing accessibility and accuracy without using specialized equipment. While WHAM can take the camera gyro as input, this is not required." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.897, + 0.607 + ], + "angle": 0, + "content": "Monocular Global 3D Human Trajectory Estimation. Estimating the global human trajectory from a monocular dynamic camera is challenging. Previous work relies on learned prior distributions of human motion to separate human motion from camera motion. GLAMR [55] computes the global trajectory based on a predicted and inflated 3D motion sequence and optimizes it across multiple individuals in the scene. However, since GLAMR does not consider camera motion cues, the output trajectory may be noisy when the camera is rotating. SLAHMR [53] and PACE [19] use off-the-shelf SLAM algorithms [44, 45] and jointly optimize the camera and human motion to minimize the negative log-likelihood of a learned motion prior [39]. While they achieve good results, their optimization approach is computationally expensive. TRACE [43] is a pure regression method that utilizes optical flow as a motion cue and estimates multiple people at once, but lacks temporal consistency. GloPro [41] regresses the uncertainty of the global human motion in real-time, but requires known camera poses. In contrast, WHAM leverages both explicit and implicit prior knowledge of human motion and efficiently reconstructs accurate and temporally coherent 3D human motion in world coordinates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.6, + 0.656 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.611, + 0.682 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.897, + 0.903 + ], + "angle": 0, + "content": "An overview of our World-grounded Human with Accurate Motion (WHAM) framework is illustrated in Fig. 2. The input to WHAM is a raw video data \\(\\{I^{(t)}\\}_{t = 0}^{T}\\), captured by a camera with possibly unknown motion. Our goal is to predict the corresponding sequence of the SMPL model parameters \\(\\{\\Theta^{(t)}\\}_{t = 0}^{T}\\), as well as the root orientation \\(\\{\\Gamma^{(t)}\\}_{t = 0}^{T}\\) and translation \\(\\{\\tau^{(t)}\\}_{t = 0}^{T}\\), expressed in the world coordinate system. We use ViTPose [52] to detect 2D keypoints \\(\\{x_{2D}^{(t)}\\}_{t = 0}^{T}\\) from which we obtain motion features \\(\\{\\phi_m^{(t)}\\}_{t = 0}^T\\) using the motion encoder. Additionally, we use a pretrained image encoder [6, 20, 24] to extract static image features \\(\\{\\phi_i^{(t)}\\}_{t = 0}^T\\) and integrate them with \\(\\{\\phi_m^{(t)}\\}_{t = 0}^T\\) to obtain fine-grained motion features \\(\\{\\hat{\\phi}_m^{(t)}\\}_{t = 0}^T\\) from which we regress 3D human" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.852, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 2. An Overview of WHAM. WHAM takes the sequence of 2D keypoints estimated by a pretrained detector and encodes it into a motion feature. WHAM then updates the motion feature using another sequence of image features extracted from the image encoder through the feature integrator. From the updated motion feature, the Local Motion Decoder estimates 3D motion in the camera coordinate system and foot-ground contact probability. The Trajectory Decoder takes the motion feature and camera angular velocity to initially estimate the global root orientation and egocentric velocity, which are then updated through the Trajectory Refiner using the foot-ground contact. The final output of WHAM is pixel-aligned 3D human motion with the 3D trajectory in the global coordinates." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.375, + 0.338, + 0.391 + ], + "angle": 0, + "content": "motion in the world coordinate system." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.401, + 0.283, + 0.416 + ], + "angle": 0, + "content": "3.2. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.549 + ], + "angle": 0, + "content": "Uni-directional Motion Encoder and Decoder. In contrast to existing methods [5, 29, 42, 50, 59], which use windows with a fixed time duration, we use uni-directional recurrent neural networks (RNN) for the motion encoder and decoder, making WHAM suitable for online inference. The objective of the motion encoder, \\( E_{M} \\), is to extract the motion context, \\( \\phi_{m}^{(t)} \\), from the current and previous sequence of 2D keypoints and the initial hidden state, \\( h_{E}^{(0)} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.56, + 0.396, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\phi_ {m} ^ {(t)} = E _ {M} \\big (x _ {2 D} ^ {(0)}, x _ {2 D} ^ {(1)},..., x _ {2 D} ^ {(t)} | h _ {E} ^ {(0)} \\big).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.684 + ], + "angle": 0, + "content": "We normalize keypoints to a bounding box around the person and concatenate the box's center and scale to the keypoints, similar to CLIFF [24]. The role of the motion decoder, \\(D_{M}\\) is to recover SMPL parameters, \\((\\theta ,\\beta)\\), weak-perspective camera translation, \\(c\\), and foot-ground contact probability, \\(p\\) from the motion feature history:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.695, + 0.433, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\left(\\theta^ {(t)}, \\beta^ {(t)}, c ^ {(t)}, p ^ {(t)}\\right) = D _ {M} \\left(\\hat {\\phi} _ {m} ^ {(0)}, \\dots , \\hat {\\phi} _ {m} ^ {(t)} | h _ {D} ^ {(0)}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Here, \\(\\hat{\\phi}_m^{(t)}\\) is the motion feature integrated with the image feature \\(\\phi_i^{(t)}\\) (described below). During pretraining on synthetic data, the image feature is not available and we set \\(\\hat{\\phi}_m^{(t)} = \\phi_m^{(t)}\\). As the encoder and decoder are tasked with reconstructing a dense 3D representation \\(\\Theta\\) from a sparse 2D input signal \\(x_{2D}\\), we design an intermediate task to predict the 3D keypoints \\(x_{3D}\\) and use them as the intermediate motion representation. This cascaded approach guides \\(\\phi_m\\) to represent the implicit context of motion and the 3D spatial structure of the body. Similar to PIP [54], we use Neural Initialization that uses MLP to initialize the hidden state of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.405 + ], + "angle": 0, + "content": "the motion encoder and decoder, \\(\\left(h_{E}^{(0)}, h_{D}^{(0)}\\right)\\); see section 4.3 and Sup. Mat. for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.649 + ], + "angle": 0, + "content": "Motion and Visual Feature Integrator. We use the AMASS dataset to synthetically generate 2D sequences by projecting 3D SMPL joints into images with varied camera motions. This provides effectively limitless training data that is far more diverse than existing video datasets that contain ground truth 3D pose and shape. Although we leverage the temporal human motion context, lifting 2D keypoints to 3D meshes is an ambiguous task. A key idea is to augment this 2D keypoint information with image cues that can help disambiguate the 3D pose. Specifically, we use an image encoder [1, 6, 20, 24], pretrained on the human mesh recovery task, to extract image features \\(\\phi_{i}\\), which contain dense visual contextual information related to the 3D human pose and shape. We then train a feature integrator network, \\(F_{I}\\), to combine \\(\\phi_{m}\\) with \\(\\phi_{i}\\), integrating motion and visual context. The feature integrator uses a simple yet effective residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.661, + 0.824, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\hat {\\phi} _ {m} ^ {(t)} = \\phi_ {m} ^ {(t)} + F _ {I} \\left(\\operatorname {c o n c a t} \\left(\\phi_ {m} ^ {(t)}, \\phi_ {i} ^ {(t)}\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.894, + 0.757 + ], + "angle": 0, + "content": "This supplements motion features pretrained on the 2D-to-3D lifting task using AMASS with visual context, resulting in enriched motion features that use image evidence to help disambiguate the task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Global Trajectory Decoder. We design an additional decoder, \\(D_{T}\\), to predict the rough global root orientation \\(\\Gamma_0^{(t)}\\) and root velocity \\(v_0^{(t)}\\) from the motion feature \\(\\phi_m^{(t)}\\). Since \\(\\phi_m\\) is derived from the input signals in the camera coordinates, it is highly challenging to decouple the human and camera motion from it. To address this ambiguity, we append the angular velocity of the camera, \\(\\omega^{(t)}\\), to the motion feature, \\(\\phi_m^{(t)}\\), to create a camera-agnostic motion context. This design choice makes WHAM compatible with both off-the-shelf" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2073" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.155 + ], + "angle": 0, + "content": "SLAM algorithms [44, 45] and gyroscope measurements that are widely available from modern digital cameras. We recursively predict global orientation, \\(\\Gamma_0^{(t)}\\), using the unidirectional RNN." + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.162, + 0.42, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\big (\\Gamma_ {0} ^ {(t)}, v _ {0} ^ {(t)} \\big) = D _ {T} \\big (\\phi_ {m} ^ {(0)}, \\omega^ {(0)}, \\dots , \\phi_ {m} ^ {(t)}, \\omega^ {(t)} \\big).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.193, + 0.471, + 0.359 + ], + "angle": 0, + "content": "Contact Aware Trajectory Refinement. Good 3D motion in world coordinates in most scenarios implies accurate foot-ground contact without sliding. We want WHAM to generalize beyond flat ground planes, which are typically assumed in prior work. Specifically, our new trajectory refiner aims to resolve foot sliding and enables WHAM to generalize well to diverse motions, including climbing stairs. The refinement involves two stages. First, we adjust the ego-centric root velocity to \\(\\tilde{v}^{(t)}\\) to minimize foot sliding, based on the foot-ground contact probability \\(p^{(t)}\\), estimated from the motion decoder \\(D_M\\):" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.368, + 0.366, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\tilde {v} ^ {(t)} = v _ {0} ^ {(t)} - \\left(\\Gamma_ {0} ^ {(t)}\\right) ^ {- 1} \\bar {v} _ {f} ^ {(t)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.543 + ], + "angle": 0, + "content": "where \\(\\bar{v}_f^{(t)}\\) is the averaged velocity of the toes and heels in world coordinates when their contact probability, \\(p^{(t)}\\), is higher than a threshold. However, this velocity adjustment often introduces noisy translation when the contact and pose estimation is inaccurate. Therefore, we propose a simple learning mechanism in which a trajectory refining network, \\(R_{T}\\), updates the root orientation and velocity to address this issue. Finally, the global translation is computed through a roll-out operation:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.553, + 0.462, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\Gamma^ {(t)}, v ^ {(t)}\\right) = R _ {T} \\big (\\phi_ {m} ^ {(0)}, \\Gamma_ {0} ^ {(0)}, \\tilde {v} ^ {(0)},..., \\phi_ {m} ^ {(t)}, \\Gamma_ {0} ^ {(t)}, \\tilde {v} ^ {(t)} \\big), \\\\ \\tau^ {(t)} = \\sum_ {i = 0} ^ {t - 1} \\Gamma^ {(i)} v ^ {(i)}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.673 + ], + "angle": 0, + "content": "In summary, this full process reconstructs accurate 3D human pose and shape in both the camera and world coordinates from a single monocular video sequence (Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.182, + 0.697 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Pretraining on AMASS. We train in two stages: (1) pretraining with synthetic data, and (2) fine-tuning with real data (Fig. 3). The objective of the pretraining stage is to teach the motion encoder to extract motion context from the input 2D keypoint sequence. The motion and trajectory decoders then map this motion context to the corresponding 3D motion and global trajectory spaces (i.e. they lift the encoding to 3D). We use the AMASS dataset [31] to generate an extensive set of synthetic pairs consisting of sequences of 2D keypoints together with the ground truth SMPL parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To synthesize 2D keypoints from AMASS, we create virtual cameras onto which we project 3D keypoints derived from the ground truth mesh. Unlike MotionBERT [59] and" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.093, + 0.889, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.171, + 0.889, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.257, + 0.894, + 0.382 + ], + "angle": 0, + "content": "Figure 3. WHAM's Two-Stage Training Scheme. During pre-taining, we generate synthetic 2D keypoint sequences from AMASS [31] and train a motion encoder and decoder on the generated data (top). We then leverage video datasets with ground truth SMPL parameters, for which there is much less data. We use the fixed-weight pre-trained image encoder and keypoints detector \\((\\text{※})\\) to extract image features and 2D keypoints. In this stage, we train the feature integration network while fine-tuning the motion encoder and motion/trajectory decoders, marked \\(\\text{※}\\) (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.553 + ], + "angle": 0, + "content": "ProxyCap [58], which use static cameras for keypoint projection, we employ dynamic cameras that incorporate both rotational and translational motion. This choice has two main motivations. First, it accounts for the inherent differences between human motions captured in static and dynamic camera setups. Second, it enables the learning of a camera-agnostic motion representation, from which the trajectory decoder can reconstruct the global trajectory. We also augment the 2D data with noise and masking. For details of the synthetic generation process see Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.894, + 0.718 + ], + "angle": 0, + "content": "Fine-tuning on Video Datasets. Starting with the pretrained network, we fine-tune WHAM on four video datasets: 3DPW [47], Human3.6M [10], MPI-INF-3DHP [32], and InstaVariety [14]. For the human mesh recovery task, we supervise WHAM on ground-truth SMPL parameters from AMASS and 3DPW, 3D keypoints from Human3.6M and MPI-INF-3DHP, and 2D keypoints from InstaVariety. For the global trajectory estimation task, we use AMASS, Human3.6M, and MPI-INF-3DHP. Additionally, during training we experiment with adding BEDLAM [1], a large synthetic dataset with realistic video and ground truth SMPL parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The fine-tuning has two objectives: 1) exposing the network to real 2D keypoints, instead of training it solely on synthetic data, and 2) training the feature integrator network to aggregate motion and image features. To achieve these goals, we jointly train the entire network on the video datasets while setting a smaller learning rate on the pretrained modules (see Fig. 3). Consistent with prior work [5, 16, 29, 42, 50], we employ a pretrained and fixed-weight image encoder [20] to extract image features. However, to leverage recent network architectures and training strategies, we also experiment with different types of encoders [1, 6, 24] in the following section. Implementation Details. For the pretraining stage, we train" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.198 + ], + "angle": 0, + "content": "WHAM on AMASS for 80 epochs with the learning rate of \\(5 \\times 10^{-4}\\). Then we finetune WHAM on 3DPW, MPI-INF-3DHP, Human3.6M, and InstaVariety for 30 epochs. We use a learning rate of \\(1 \\times 10^{-4}\\) for the feature integrator and \\(1 \\times 10^{-5}\\) for the pretrained components during finetuning. During training, we use the Adam optimizer and a batch size of 64 and split sequences into 81-frame chunks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.209, + 0.21, + 0.226 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.473, + 0.369 + ], + "angle": 0, + "content": "Datasets. We evaluate WHAM on three in-the-wild benchmarks: 3DPW [47], RICH [9], and EMDB [15]. Following previous work [1, 5, 13, 16, 20, 24, 29], we perform the evaluation in camera coordinates. The estimated global trajectory is evaluated on a subset of EMDB (EMDB 2) for which they provide ground truth global motion with dynamic cameras (used for evaluation). We also test on new sequences captured using an iPhone with the gyroscope. See Sup. Mat. for more details on the datasets and iPhone results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.474, + 0.673 + ], + "angle": 0, + "content": "Evaluation metrics. To evaluate the accuracy of 3D human pose and shape estimation, we compute Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PA-MPJPE), and Per Vertex Error (PVE) measured in millimeters \\((mm)\\). We compute Acceleration error (Accel, in \\(m / s^2\\)) to measure the inter-frame smoothness of the reconstructed motion. We also evaluate the motion reconstruction and trajectory estimation accuracy in the world-frame. Following previous work [19, 53], we split sequences into smaller segments of 100 frames and align each output segment with the ground-truth data using the first two frames \\(\\mathrm{(W - MPJPE_{100})}\\) or the entire segment \\(\\mathrm{(WA - MPJPE_{100})}\\) in mm. These previous metrics give an unrealistic picture of 3D performance as they do not measure drift over long sequences. Therefore, we also evaluate the error over the entire trajectory after the rigid alignment and measure Root Translation Error (RTE in %) normalized by the actual displacement of the person. We also assess the jitter of the motion in the world coordinate system in \\(10m / s^3\\) and foot sliding during the contact (FS in mm)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.334, + 0.695 + ], + "angle": 0, + "content": "4.1. 3D Human Motion Recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.472, + 0.869 + ], + "angle": 0, + "content": "Per-frame accuracy. Table 1 presents a comprehensive comparison of WHAM and the existing state-of-the-art per-frame and video-based methods across three benchmark datasets [9, 15, 47]. Because none of the methods are exposed to data from RICH or EMDB during training, results on these datasets are indicative of each method's ability to generalize. WHAM (Res), WHAM (HR), and WHAM (ViT) correspond to different architectures for the pretrained image encoders, derived from SPIN (ResNet-50) [20], CLIFF (HRNet-W48) [1, 24], and HMR2.0 (ViT-H/16) [6], respectively. Not surprisingly, WHAM (HR) is more accurate than WHAM (Res)," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.875, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Previous work follows [14] in reporting Accel in \\( mm / \\text{frame}^2 \\). To remove the dependency on frame rate, we convert all previous results to \\( m / s^2 \\)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.09, + 0.582, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.09, + 0.657, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.091, + 0.73, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.09, + 0.805, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.09, + 0.881, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.177, + 0.872, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.317, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with previous state-of-the-art methods for 3D human pose and shape estimation. See text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.457 + ], + "angle": 0, + "content": "while the transformer-based version, WHAM (ViT), is the most accurate. The backbone matters, with WHAM (ViT) outperforming all previous methods on all per-frame metrics (MPJPE, PA-MPJPE, and PVE) on all benchmarks. Even with the simplest ResNet backbone, WHAM (Res) outperforms every method except for ReFit on RICH." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.895, + 0.563 + ], + "angle": 0, + "content": "Training on BEDLAM consistently improves accuracy in prior work (see [1, 3, 49]), and we find the same here as shown in Table 2. However, compared to ReFit, WHAM exhibits relatively smaller performance improvement by adding BEDLAM. One possible reason for this is that, unlike ReFit, we did not finetune the image encoder during training to limit training time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.895, + 0.73 + ], + "angle": 0, + "content": "Inter-frame smoothness. We also evaluate the inter-frame smoothness using the acceleration error. Compared with state-of-the-art per-frame methods [1, 6, 22, 24], WHAM has significantly lower acceleration error. This indicates that WHAM reconstructs smooth and more plausible 3D human motion across frames while not sacrificing high per-frame accuracy. On the other hand, when compared to recent temporal methods [5, 42, 50], WHAM exhibits comparable or slightly higher acceleration error. However, we observe that these video-based methods tend to over-smooth the human motion, resulting in lower accuracy on per-frame metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.895, + 0.821 + ], + "angle": 0, + "content": "To provide intuition for these numbers, we qualitatively compare WHAM with TCMR [5] and GLoT [42] in Fig. 4. While producing smooth results, TCMR and GLoT fail to capture the bending of the left knee when the subject is ascending the stairs, while WHAM more accurately reconstructs the 3D human pose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.776, + 0.848 + ], + "angle": 0, + "content": "4.2. 3D Global Trajectory Recovery" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To evaluate global trajectory recovery, we compare WHAM with the state-of-the-art methods and a baseline that combines a SLAM method (DPVO [45]) and a per-frame" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2075" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.161, + 0.091, + 0.807, + 0.415 + ], + "angle": 0, + "content": "
Models3DPW (14)RICH (24)EMDB (24)
PA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccel
per-frameSPIN [20]59.296.9112.831.469.7122.9144.235.287.1140.7166.141.3
PARE* [17]46.574.588.6-60.7109.2123.5-72.2113.9133.2-
CLIFF* [24]43.069.081.222.556.6102.6115.022.468.3103.5123.724.5
HybrIK* [22]41.871.682.3-56.496.8110.4-65.6103.0122.2-
HMR2.0 [6]44.469.882.218.148.196.0110.918.860.798.3120.819.9
ReFit* [49]40.565.375.118.547.980.792.917.158.688.0104.520.7
temporalTCMR* [5]52.786.5101.46.065.6119.1137.75.079.8127.7150.25.3
VIBE* [16]51.982.998.418.568.4120.5140.221.881.6126.1149.926.5
MPS-Net* [50]52.184.399.06.567.1118.2136.75.881.4123.3143.96.2
GLoT* [42]50.680.796.46.065.6114.3132.75.279.1119.9140.85.4
GLAMR [55]51.1--8.079.9--107.773.8113.8134.933.0
TRACE* [43]50.979.195.428.6----71.5110.0129.625.5
SLAHMR [53]55.9---52.5--9.469.793.7111.37.1
PACE [19]----49.3--8.8----
WHAM (Res)*40.262.775.16.351.889.5103.25.057.884.099.75.2
WHAM (HR)*39.062.674.86.449.184.696.45.257.185.7103.25.6
WHAM (ViT)*35.957.868.76.644.380.091.25.350.479.794.45.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.422, + 0.895, + 0.479 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of state-of-the-art models on the 3DPW [47], RICH [9], and EMDB [15] datasets. Ordering of per-frame and temporal methods is done separately by descending MPJPE on EMDB (except for PACE). For testing on EMDB, we follow the protocol of EMDB 1 [15]. Parenthesis denotes the number of body joints used to compute MPJPE and PA-MPJPE, and * denotes models trained with the 3DPW training set. Bold numbers denote the most accurate method in each column. Accel is in \\( m / s^2 \\), all other errors are in mm." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.504, + 0.435, + 0.626 + ], + "angle": 0, + "content": "
ModelsDataset3DPW (14)
PA-MPJPEMPJPEPVEAccel
CLIFF [1, 24]R43.668.882.119.2
R+B43.066.978.531.0
ReFit [49]R40.565.375.118.5
R+B38.257.667.621.4
WHAM (ViT)R35.957.868.76.6
R+B35.756.967.46.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.631, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Table 2. Dataset ablation experiments on 3DPW [47]. R denotes the use of real datasets and B denotes BEDLAM." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.678, + 0.468, + 0.796 + ], + "angle": 0, + "content": "
ModelsEMDB 2
WA-MPJPE100W-MPJPE100RTEJitterFS
DPVO (+ HMR2.0) [6, 45]647.82231.415.8537.3107.6
GLAMR [55]280.8726.611.446.320.7
TRACE [43]529.01702.317.72987.6370.7
SLAHMR [53]326.9776.110.231.314.5
WHAM (w/DPVO [45])135.6354.86.022.54.4
WHAM (w/DROID [44])133.3343.94.621.54.4
WHAM (w/ GT gyro)131.1335.34.121.04.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.8, + 0.452, + 0.814 + ], + "angle": 0, + "content": "Table 3. Global motion estimation accuracy on EMDB [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "method (HMR2.0 [6]); see Table 3. WHAM is agnostic to the source of the camera angular velocity and we compare results using DPVO, DROID-SLAM [44] and the ground truth angular velocity (gyro)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.732 + ], + "angle": 0, + "content": "As shown in Table 3, WHAM outperforms the existing methods on all metrics. Specifically, combining WHAM with DPVO is more accurate than the global trajectory estimation of DPVO combined with HMR2.0, illustrating that our method actively refines the global trajectory instead of performing a simple integration. DROID-SLAM gives slightly better results than DPVO. Furthermore, WHAM significantly outperforms the regression-based method, TRACE, on jitter and foot sliding metrics. We further demonstrate this in Figs. 5 and 1, where WHAM captures more consistent and plausible human motion in the global coordinate system than TRACE and SLAHMR for videos captured by dynamic cameras. As depicted in Fig. 6, WHAM outperforms GLAMR, TRACE, and SLAHMR in capturing the pattern of human motion in the global coordinate system." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.655, + 0.759 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To provide further insight into our approach, we conduct ablation studies to analyze the contribution of each component to the performance. As shown in Table 4, our entire system (WHAM) outperforms the different variants of WHAM that ablate a single component. To be specific, we first observe that adding feature integration improves both motion and global trajectory estimation accuracy when compared with an ablated version without feature integration \\((\\mathrm{w / o}F_I)\\). Similarly, the removal of the pretraining on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.462, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.401, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with TRACE [43] and SLAHMR [53] on global human motion estimation with dynamic cameras." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.443, + 0.468, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.53, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Figure 6. Comparison of global trajectory estimation on EMDB [15]. Overall, WHAM shows better alignment to ground truth data compared to GLAMR [55], TRACE [43], and SLAHMR [53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.473, + 0.834 + ], + "angle": 0, + "content": "the 2D-to-3D lifting task using AMASS [31] (w/o lifting) shows significant performance degradation. WHAM also outperforms the ablation of the Neural Initialization (w/o NI), particularly in the MPJPE metric. In addition, we experiment with WHAM to decode trajectory solely based on the motion context without using the estimated camera angular velocity \\((\\mathrm{w / o}\\omega)\\). Although this version shows similar performance in predicting 3D human pose, it suffers from the entanglement of camera and human motion, resulting in significantly high overall trajectory error (RTE). Last, we observe that WHAM without the trajectory refinement (w/o traj. ref.) gives larger global trajectory and foot sliding errors in return for less jitter, indicating that our approach contributes to the global trajectory accuracy and helps reduce foot sliding. Core details are presented here; see Sup. Mat. for more details and information on run-time cost." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.196, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "WHAM is a new method to recover accurate 3D human motion in global coordinates from a moving camera more ef" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.091, + 0.891, + 0.187 + ], + "angle": 0, + "content": "
ModelsEMDB 2
PA-MPJPEMPJPEWA-MPJPE100W-MPJPE100RTEJitterFS
w/o F144.269.0147.6377.96.323.15.5
w/o lifting60.383.0238.0693.011.524.55.0
w/o NI40.466.3142.7368.16.822.34.6
w/o ω39.162.0156.5422.010.122.15.0
w/o traj. ref.38.259.3154.7407.56.318.86.5
WHAM (Ours)38.259.3135.6354.86.022.54.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.576, + 0.191, + 0.816, + 0.204 + ], + "angle": 0, + "content": "Table 4. Ablation experiments. See text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.519 + ], + "angle": 0, + "content": "ficiently and accurately than the state-of-the-art approaches. Our approach leverages the AMASS dataset to train a network to recursively lift 2D sequences of keypoints to sequences of 3D SMPL parameters. But keypoints alone lack valuable information about the body and its movement. Consequently, we integrate image context information over time and learn to combine it with the motion context to better estimate human body shape and pose. Additionally, our method takes an estimate of the camera angular velocity, which can either be computed from a SLAM method or from the camera's gyro when available. Finally, we combine all this information with an estimate of foot contact to recover the 3D human motion in global coordinates from a monocular video sequence. WHAM significantly outperforms the existing state-of-the-art methods (both image-based and video-based) on challenging in-the-wild benchmarks in both 3D HPS and the world-coordinate trajectory estimation accuracy. Because of its speed and accuracy, WHAM provides a foundation for in-the-wild motion capture applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.894, + 0.616 + ], + "angle": 0, + "content": "Limitations and future directions: WHAM learns about human motion from AMASS, limiting generalization to motions that are out of distribution. While we employ random masking as part of our data synthesis process, our generating approach mainly assumes the scenario where the full body is within the field of view. See Sup. Mat. for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.895, + 0.783 + ], + "angle": 0, + "content": "WHAM opens up many directions for future work. For example, while we use SLAM to estimate the camera's angular velocity, SLAM could also provide camera intrinsics and extrinsics as well as information about the 3D scene that could be used to enforce consistency between the scene and the human. While WHAM is an online method, designed for real-time applications, it could also initialize an optimization-based post-processing akin to bundle adjustment, which would optimize the camera, scene, and human motion together. Furthermore, a real-time and phone-based implementation of WHAM should be feasible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.897, + 0.888 + ], + "angle": 0, + "content": "Acknowledgements. Part of this work was done when the first author was an intern at the Max Planck Institute for Intelligence System. This work was partially supported through an NSF CAREER Award (CBET 2145473) and a Chan-Zuckerberg Essential Open Source Software for Science Award. CoI Disclosure. https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2077" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8726-8737, 2023. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016, pages 561-578. Springer International Publishing, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. SMPLer-X: Scaling up expressive human pose and shape estimation. In Proc. Neural Information Processing Systems (NeurIPS), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[4] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.527 + ], + "angle": 0, + "content": "[6] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.529, + 0.471, + 0.584 + ], + "angle": 0, + "content": "[7] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[8] Dorian F. Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. *Bodyslam++*: Fast and tightly-coupled visual-inertial camera and human motion tracking, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[9] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.471, + 0.784 + ], + "angle": 0, + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[11] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.124, + 0.894, + 0.165 + ], + "angle": 0, + "content": "[14] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. arXiv preprint arXiv:1812.01601, 2018. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.168, + 0.894, + 0.236 + ], + "angle": 0, + "content": "[15] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[16] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5252-5262. IEEE, 2020. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.311, + 0.894, + 0.379 + ], + "angle": 0, + "content": "[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.383, + 0.893, + 0.453 + ], + "angle": 0, + "content": "[18] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11015-11025, Piscataway, NJ, 2021. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[19] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.893, + 0.568 + ], + "angle": 0, + "content": "[20] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.893, + 0.613 + ], + "angle": 0, + "content": "[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.894, + 0.698 + ], + "angle": 0, + "content": "[22] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.701, + 0.893, + 0.771 + ], + "angle": 0, + "content": "[23] Jiefeng Li, Siyuan Bian, Qi Liu, Jiasheng Tang, Fan Wang, and Cewu Lu. NIKI: Neural inverse kinematics with invertible neural networks for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.893, + 0.828 + ], + "angle": 0, + "content": "[24] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[25] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2078" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[26] Kevin Lin, Chung-Ching Lin, Lin Liang, Zicheng Liu, and Lijuan Wang. Mpt: Mesh pre-training with transformers for human pose and mesh reconstruction, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[27] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[30] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 534-543, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.471, + 0.419 + ], + "angle": 0, + "content": "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3D Vision (3DV), 2017 Fifth International Conference on. IEEE, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[33] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[35] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudo-gts of 3d human mesh estimation in the wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.715 + ], + "angle": 0, + "content": "[36] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Computer Vision - ECCV 2020, pages 598-613. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[38] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459-468, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[40] Nitin Saini, Chun-Hao P. Huang, Michael J. Black, and Aamir Ahmad. Smartmocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, pages 1-8, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.221 + ], + "angle": 0, + "content": "[41] Simon Schaefer, Dorian F. Henning, and Stefan Leutenegger. Glopro: Globally-consistent uncertainty-aware 3d human pose estimation & tracking in the wild, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[42] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[43] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.894, + 0.431 + ], + "angle": 0, + "content": "[44] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.461 + ], + "angle": 0, + "content": "[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[47] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[48] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In The IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[49] Yufu Wang and Kostas Daniilidis. Refit: Recurrent fitting network for 3d human recovery. In International Conference on Computer Vision, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[50] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.894, + 0.87 + ], + "angle": 0, + "content": "[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: generative 3D human shape and articulated pose models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2079" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "estimation. In Advances in Neural Information Processing Systems, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[53] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[54] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[55] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.388 + ], + "angle": 0, + "content": "[56] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.458 + ], + "angle": 0, + "content": "[57] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[58] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Hongwei Yi, Shengping Zhang, and Yebin Liu. Real-time monocular full-body capture in world space via sequential proxy-to-motion learning. arXiv preprint arXiv:2307.01200, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2080" + } + ] +] \ No newline at end of file diff --git a/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_origin.pdf b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..52a8f02488eb20300e86bc4eb381f19fae8bd95d --- /dev/null +++ b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/63a7b8f6-97aa-4651-a01d-91f5ef8cd504_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88e8d6e4397aa622889a2edb2a9761e8733f67762ee6f63d813d6e77346e77ca +size 8582280 diff --git a/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/full.md b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e5469f40e0600d16b00096daf3c1a929030e161 --- /dev/null +++ b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/full.md @@ -0,0 +1,319 @@ +# WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion + +Soyong Shin + +Juyong Kim + +Eni Halilaj + +Michael J. Black + +$^{1}$ Carnegie Mellon University $^{2}$ Max Planck Institute for Intelligent System $\{$ soyongs, juyongk, ehalilaj\} $@$ andrew.cmu.edu black@tuebingen.mpg.de + +![](images/02de85de1d553150e6003c73f47b9ebc3857dde697d07567bdcb53729d637c96.jpg) + +![](images/55c6900b3437267b9e552ea97719da27a26c57afb96518b6d56a592b526da379.jpg) +Figure 1. WHAM: World-grounded Humans with Accurate Motion. State-of-the-art methods like TRACE [43] and SLAHMR [53] fail to capture global 3D human trajectories accurately when given in-the-wild videos captured using a moving camera, producing implausible world-grounded motion (e.g., foot sliding). To address this, WHAM uses two novel strategies: (1) feature integration from 2D keypoints and pixels to reconstruct precise and pixel-aligned 3D human motion and (2) contact-aware trajectory recovery to place the human in a global coordinate system without foot sliding. Gray dots show the ground-truth global trajectory. See Supplemental Video. + +![](images/6ae0c45593a6d230d633af96a4346b8b41e66b5f7a71c46867f2cf51b181089b.jpg) + +![](images/1ab30b270c67489a9234fcd8cfe611b1651f1f304a0fd6449f4e05c71146dd30.jpg) + +![](images/5fb411d59626f103f39d2d91c24cab0859a9b7601efbdb1576815ca16ecd8bda.jpg) + +![](images/1ecb7c67e515e8b0761b9d7b4ade6e1456b734e032aedeb3dde91e5fca698f96.jpg) + +![](images/7ca2367ede506337ffe951dd7b79558904968a72a0f20da31a3d929ea266edc7.jpg) + +![](images/9dab6a5c367779b47e013dc525d716607eae46c73878ca694cb371e98180d571.jpg) + +![](images/166cfae0e92bc02bb9a9dcb94f3bdb6cff144eed3a63d790b4dac2ba4ad0a522.jpg) + +![](images/96dbba8d7c47704d0c29f7ac033f83d5e4c1802c42563e5f977cb486c214962a.jpg) + +GT trajectory + +![](images/c6e3f307f386fe7c11473a5c836a1f82fff316b3d6473f3767fb9a932184ec23.jpg) +TRACE + +![](images/f12dbcac6531be2968ed4fb4678344e83764c9a79f918ec134d3e023463cdafc.jpg) +SLAHMR + +![](images/21bd141faa0eb986fb6901c57fbfa0b36b452804fb648cee1717e232c87589d7.jpg) +WHAM (Ours) + +![](images/1bcb5be1c007b058b27d63decabfdb1aef80b7fea3189cb61302638df128f592.jpg) + +# Abstract + +The estimation of 3D human motion from video has progressed rapidly but current methods still have several key limitations. First, most methods estimate the human in camera coordinates. Second, prior work on estimating humans in global coordinates often assumes a flat ground plane and produces foot sliding. Third, the most accurate methods rely on computationally expensive optimization pipelines, limiting their use to offline applications. Finally, existing video-based methods are surprisingly less accurate than single-frame methods. We address these limitations with WHAM (World-grounded Humans with Accurate Motion), which accurately and efficiently reconstructs 3D human motion in a global coordinate system from video. WHAM learns to lift 2D keypoint sequences to 3D using motion capture data and fuses this with video features, integrating motion context and visual information. WHAM exploits camera angular velocity estimated from a SLAM method together with human motion to estimate the body's global trajectory. We combine this with a contact-aware trajectory refinement method that + +lets WHAM capture human motion in diverse conditions, such as climbing stairs. WHAM outperforms all existing 3D human motion recovery methods across multiple in-the-wild benchmarks. Code is available for research purposes at http://wham.is.tue.mpg.de/. + +# 1. Introduction + +Our goal is to accurately estimate the 3D pose and shape of a person from monocular video. This is a longstanding problem and, while the field has made rapid progress, several key challenges remain. First, human motion should be computed in a consistent global coordinate system. Second, the method should be computationally efficient. Third, the results should be accurate, temporally smooth, detailed, natural looking, and have realistic foot-ground contact. Fourth, the capture should work with an arbitrary moving camera. These constraints need to be satisfied to make markerless human motion capture widely available for applications in gaming, AR/VR, autonomous driving, sports analysis, and human-robot interaction. We address these challenges with + +WHAM (World-grounded Humans with Accurate Motion), which enables fast and accurate recovery of 3D human motion from a moving camera; see Fig. 1. + +It seems natural that, in estimating 3D humans from video, we should be able to exploit the temporal nature of video. Counter-intuitively, existing video-based methods for 3D human pose and shape (HPS) estimation [5, 14, 16, 29, 42, 50] are less accurate than the best single-frame methods [6, 13, 17, 20, 22, 24, 25, 57, 57]. This may be an issue of training data. There are large datasets of single images with ground-truth 3D human poses containing a diversity of body shapes, poses, backgrounds, lighting, etc. In contrast, video datasets with ground truth are much more limited. + +To address this, WHAM leverages both the large-scale AMASS motion capture (mocap) dataset [31] and video datasets. Our key idea is to learn about 3D human motion from AMASS and then learn to fuse this information with temporal image cues from video, getting the best of both. Similar to previous work [58, 59], we use AMASS to generate synthetic 2D keypoints and ground-truth motion sequence pairs, from which we pretrain a motion encoder, which captures the motion context, and decoder that lifts sequences of 2D keypoints to sequences of 3D poses. Given the robustness of recent 2D keypoint detection models [52, 56], our pretrained model does a good job of predicting human pose from video. + +Keypoints alone, however, are too sparse for accurate 3D mesh estimation. To improve accuracy, we jointly train a feature integrator network that merges information from video and 2D-keypoint sequences. We use a pretrained image encoder from previous work [1, 6, 20, 24] and train the feature integrator using video datasets [10, 14, 32, 47]. This integration process supplements the motion context extracted from the sparse 2D keypoints with dense visual context, significantly improving the recovered pose and shape accuracy. + +While the above approach produces accurate motion, we want this motion in global coordinates, unlike most previous methods that compute the body in camera coordinates. Estimating the global human trajectory is challenging when the camera is moving because the motions of the body and the camera are entangled. Recent work addresses this with optimization fitting based on a learned human motion prior and camera information from SLAM methods [19, 40, 53] or dense 3D scene information from COLMAP [27]. However, these methods are computationally expensive and far from real time. Recent regression-based methods are faster but either constrain the problem with static or known camera conditions [41, 58] or have temporal jitter and limited accuracy [43]. We tackle this challenge with two additional modules. First, we predict the global orientation and root velocity of the human from the sequence of 2D keypoints by training a global trajectory decoder. Specifically, we concatenate the camera's angular velocity to the context and + +train the global trajectory decoder to recursively predict the current orientation and root velocity, effectively factoring camera motion from human motion. WHAM takes the camera's angular velocity either from the output of a SLAM method or from a camera's gyroscope when available. + +The above solution relies on knowledge of human motion learned from AMASS. Therefore, it can fail to capture elevation changes when the surface is not flat, e.g. when ascending the stairs because AMASS has a limited amount of such data. To address this, we introduce foot contact as an additional explicit form of motion context. We train WHAM to predict the likelihood of foot-ground contact using estimated contact labels from both AMASS and 3D video datasets. We then train a trajectory refinement network that outputs an update to the root orientation and velocity based on the information about the foot contact/velocity. This refinement enables WHAM to accurately estimate human motion in a global coordinate system even when the terrain is not flat. + +WHAM has very low computational overhead because it is an on-line algorithm that recursively predicts the pose, shape, and global motion parameters. The network, excluding preprocessing (bounding box detection, keypoint detection, and person identification), runs at 200 fps, significantly faster than prior methods. Also, despite not using global optimization like [53], we obtain accurate 3D camera trajectories and global body motions with minimal drift. Through extensive comparisons on multiple in-the-wild datasets as well as detailed ablation studies, we find that WHAM achieves state-of-the-art (SOTA) accuracy on 3D human pose estimation as well as global trajectory estimation (see Fig. 1). + +In summary, in this paper we: (1) introduce the first approach to effectively fuse 3D human motion context and video context for 3D HPS regression; (2) propose a novel global trajectory estimation framework that leverages motion context and foot contact to effectively address foot sliding and enable the 3D tracking of people on non-planar surfaces; (3) efficiently perform HPS regression in global coordinates; (4) achieve state-of-the-art (SOTA) performance on multiple in-the-wild benchmark datasets (3DPW [47], RICH [9], EMDB [15]). WHAM is the first video-based method to outperform all image-based and video-based methods on per-frame accuracy while maintaining temporal smoothness. + +# 2. Related Work + +Image-based 3D HPS Estimation. There are two broad classes of methods for recovering 3D HPS from images: model-free [21, 26, 33] and model-based [6, 11-13, 20, 24, 34, 35]. Here we focus on model-based methods, which estimate the low-dimensional parameters of a statistical body model [28, 36, 37, 51]. While early work explores optimization-based methods [2], here we focus on direct regression methods based on deep learning. + +Many existing methods follow the architecture of HMR + +[13], which uses a pretrained backbone to predict image features followed by a multilayer perceptron (MLP) that regresses SMPL [28] pose parameters from image features. Training such networks typically leverages paired images with SMPL parameters; these parameters are often pseudo-groundtruth (p-GT), estimated from 2D keypoints [11, 20, 24, 34, 35]. Other architectures for HPS regression have also been proposed [6, 17, 18, 22, 23, 57]. None of these methods use video or estimate the body in global coordinates. While quite accurate, when these image-based models are applied independently to frames of a video sequence, the shape and pose can be temporally inconsistent. In contrast, WHAM effectively aggregates temporal information to provide frame-accurate and temporally-coherent 3D HPS estimation. + +Video-Based 3D HPS. Video-based methods typically encode temporal information by combining static features extracted by a backbone from each frame. HMMR [14] uses a convolutional encoder, while VIBE [16] and MEVA [29] employ recurrent neural networks. TCMR [5] divides sequences into past, future, and whole frames, aggregating information to strongly constrain the output with motion consistency. MPS-Net [50] uses attention to capture non-local motion context and a hierarchical architecture to aggregate temporal features. Both MAED [48] and GLoT [42] use transformer architectures [46] to encode videos. MAED encodes videos in both temporal (across frames) and spatial (within each frame) dimensions and leverages the kinematic tree to iteratively regress each joint angle. GLoT encodes long-term temporal correlations and refines local details by focusing on nearby frames. Despite integrating information across frames, all existing video-based methods have lower accuracy than the best single-frame methods. + +Given limited video training data with ground truth SMPL poses, several single-frame methods infer a mesh from 2D/3D keypoints [4, 7, 30, 33, 38] and use the keypoints as a proxy for training. Another set of approaches exploits 3D mocap data, which is plentiful [31], to train a network to lift 2D joints to 3D poses, which are used as a proxy for ground truth 3D. MotionBERT [59] synthesizes 2D keypoints through orthographic projection to learn a unified motion representation. ProxyCap [58] projects synthetic 3D keypoints into virtual cameras using a heuristic camera pose distribution. Despite benefiting from the scale of mocap datasets, these approaches do not fully utilize the visual information available in the video at test time. Here, we propose a combined network architecture and training strategy that leverages both proxy representations of human pose (lifting) and visual context extracted from video. + +Global 3D Motion Estimation with Dense Sensors. Several methods augment video data with other sensors to estimate 3D HPS in world coordinates. The 3DPW dataset [47] employs pre-calibrated body-worn inertial sensors and + +a handheld camera to jointly optimize the camera and human motion in challenging environments. Similarly, the EMDB dataset [15] uses electromagnetic sensors with an RGB-D camera, enabling accurate human motion capture in the world. While body-worn sensors aid global human motion reconstruction, they are intrusive, require cooperation, and do not help with archival video. BodySLAM++ [8] uses an optimization method with a visual-inertial sensor, comprising stereo cameras and an IMU. In contrast, we use a standard monocular camera, balancing accessibility and accuracy without using specialized equipment. While WHAM can take the camera gyro as input, this is not required. + +Monocular Global 3D Human Trajectory Estimation. Estimating the global human trajectory from a monocular dynamic camera is challenging. Previous work relies on learned prior distributions of human motion to separate human motion from camera motion. GLAMR [55] computes the global trajectory based on a predicted and inflated 3D motion sequence and optimizes it across multiple individuals in the scene. However, since GLAMR does not consider camera motion cues, the output trajectory may be noisy when the camera is rotating. SLAHMR [53] and PACE [19] use off-the-shelf SLAM algorithms [44, 45] and jointly optimize the camera and human motion to minimize the negative log-likelihood of a learned motion prior [39]. While they achieve good results, their optimization approach is computationally expensive. TRACE [43] is a pure regression method that utilizes optical flow as a motion cue and estimates multiple people at once, but lacks temporal consistency. GloPro [41] regresses the uncertainty of the global human motion in real-time, but requires known camera poses. In contrast, WHAM leverages both explicit and implicit prior knowledge of human motion and efficiently reconstructs accurate and temporally coherent 3D human motion in world coordinates. + +# 3. Methods + +# 3.1. Overview + +An overview of our World-grounded Human with Accurate Motion (WHAM) framework is illustrated in Fig. 2. The input to WHAM is a raw video data $\{I^{(t)}\}_{t = 0}^{T}$ , captured by a camera with possibly unknown motion. Our goal is to predict the corresponding sequence of the SMPL model parameters $\{\Theta^{(t)}\}_{t = 0}^{T}$ , as well as the root orientation $\{\Gamma^{(t)}\}_{t = 0}^{T}$ and translation $\{\tau^{(t)}\}_{t = 0}^{T}$ , expressed in the world coordinate system. We use ViTPose [52] to detect 2D keypoints $\{x_{2D}^{(t)}\}_{t = 0}^{T}$ from which we obtain motion features $\{\phi_m^{(t)}\}_{t = 0}^T$ using the motion encoder. Additionally, we use a pretrained image encoder [6, 20, 24] to extract static image features $\{\phi_i^{(t)}\}_{t = 0}^T$ and integrate them with $\{\phi_m^{(t)}\}_{t = 0}^T$ to obtain fine-grained motion features $\{\hat{\phi}_m^{(t)}\}_{t = 0}^T$ from which we regress 3D human + +![](images/c1794a5e6d8edd55623e19d803efd1d0e3c8bdfa15231e1bd8783c3de45e9b88.jpg) +Figure 2. An Overview of WHAM. WHAM takes the sequence of 2D keypoints estimated by a pretrained detector and encodes it into a motion feature. WHAM then updates the motion feature using another sequence of image features extracted from the image encoder through the feature integrator. From the updated motion feature, the Local Motion Decoder estimates 3D motion in the camera coordinate system and foot-ground contact probability. The Trajectory Decoder takes the motion feature and camera angular velocity to initially estimate the global root orientation and egocentric velocity, which are then updated through the Trajectory Refiner using the foot-ground contact. The final output of WHAM is pixel-aligned 3D human motion with the 3D trajectory in the global coordinates. + +motion in the world coordinate system. + +# 3.2. Network Architecture + +Uni-directional Motion Encoder and Decoder. In contrast to existing methods [5, 29, 42, 50, 59], which use windows with a fixed time duration, we use uni-directional recurrent neural networks (RNN) for the motion encoder and decoder, making WHAM suitable for online inference. The objective of the motion encoder, $E_{M}$ , is to extract the motion context, $\phi_{m}^{(t)}$ , from the current and previous sequence of 2D keypoints and the initial hidden state, $h_{E}^{(0)}$ : + +$$ +\phi_ {m} ^ {(t)} = E _ {M} \big (x _ {2 D} ^ {(0)}, x _ {2 D} ^ {(1)},..., x _ {2 D} ^ {(t)} | h _ {E} ^ {(0)} \big). +$$ + +We normalize keypoints to a bounding box around the person and concatenate the box's center and scale to the keypoints, similar to CLIFF [24]. The role of the motion decoder, $D_{M}$ is to recover SMPL parameters, $(\theta ,\beta)$ , weak-perspective camera translation, $c$ , and foot-ground contact probability, $p$ from the motion feature history: + +$$ +\left(\theta^ {(t)}, \beta^ {(t)}, c ^ {(t)}, p ^ {(t)}\right) = D _ {M} \left(\hat {\phi} _ {m} ^ {(0)}, \dots , \hat {\phi} _ {m} ^ {(t)} | h _ {D} ^ {(0)}\right). +$$ + +Here, $\hat{\phi}_m^{(t)}$ is the motion feature integrated with the image feature $\phi_i^{(t)}$ (described below). During pretraining on synthetic data, the image feature is not available and we set $\hat{\phi}_m^{(t)} = \phi_m^{(t)}$ . As the encoder and decoder are tasked with reconstructing a dense 3D representation $\Theta$ from a sparse 2D input signal $x_{2D}$ , we design an intermediate task to predict the 3D keypoints $x_{3D}$ and use them as the intermediate motion representation. This cascaded approach guides $\phi_m$ to represent the implicit context of motion and the 3D spatial structure of the body. Similar to PIP [54], we use Neural Initialization that uses MLP to initialize the hidden state of + +the motion encoder and decoder, $\left(h_{E}^{(0)}, h_{D}^{(0)}\right)$ ; see section 4.3 and Sup. Mat. for details. + +Motion and Visual Feature Integrator. We use the AMASS dataset to synthetically generate 2D sequences by projecting 3D SMPL joints into images with varied camera motions. This provides effectively limitless training data that is far more diverse than existing video datasets that contain ground truth 3D pose and shape. Although we leverage the temporal human motion context, lifting 2D keypoints to 3D meshes is an ambiguous task. A key idea is to augment this 2D keypoint information with image cues that can help disambiguate the 3D pose. Specifically, we use an image encoder [1, 6, 20, 24], pretrained on the human mesh recovery task, to extract image features $\phi_{i}$ , which contain dense visual contextual information related to the 3D human pose and shape. We then train a feature integrator network, $F_{I}$ , to combine $\phi_{m}$ with $\phi_{i}$ , integrating motion and visual context. The feature integrator uses a simple yet effective residual connection: + +$$ +\hat {\phi} _ {m} ^ {(t)} = \phi_ {m} ^ {(t)} + F _ {I} \left(\operatorname {c o n c a t} \left(\phi_ {m} ^ {(t)}, \phi_ {i} ^ {(t)}\right)\right). +$$ + +This supplements motion features pretrained on the 2D-to-3D lifting task using AMASS with visual context, resulting in enriched motion features that use image evidence to help disambiguate the task. + +Global Trajectory Decoder. We design an additional decoder, $D_{T}$ , to predict the rough global root orientation $\Gamma_0^{(t)}$ and root velocity $v_0^{(t)}$ from the motion feature $\phi_m^{(t)}$ . Since $\phi_m$ is derived from the input signals in the camera coordinates, it is highly challenging to decouple the human and camera motion from it. To address this ambiguity, we append the angular velocity of the camera, $\omega^{(t)}$ , to the motion feature, $\phi_m^{(t)}$ , to create a camera-agnostic motion context. This design choice makes WHAM compatible with both off-the-shelf + +SLAM algorithms [44, 45] and gyroscope measurements that are widely available from modern digital cameras. We recursively predict global orientation, $\Gamma_0^{(t)}$ , using the unidirectional RNN. + +$$ +\big (\Gamma_ {0} ^ {(t)}, v _ {0} ^ {(t)} \big) = D _ {T} \big (\phi_ {m} ^ {(0)}, \omega^ {(0)}, \dots , \phi_ {m} ^ {(t)}, \omega^ {(t)} \big). +$$ + +Contact Aware Trajectory Refinement. Good 3D motion in world coordinates in most scenarios implies accurate foot-ground contact without sliding. We want WHAM to generalize beyond flat ground planes, which are typically assumed in prior work. Specifically, our new trajectory refiner aims to resolve foot sliding and enables WHAM to generalize well to diverse motions, including climbing stairs. The refinement involves two stages. First, we adjust the ego-centric root velocity to $\tilde{v}^{(t)}$ to minimize foot sliding, based on the foot-ground contact probability $p^{(t)}$ , estimated from the motion decoder $D_M$ : + +$$ +\tilde {v} ^ {(t)} = v _ {0} ^ {(t)} - \left(\Gamma_ {0} ^ {(t)}\right) ^ {- 1} \bar {v} _ {f} ^ {(t)}, +$$ + +where $\bar{v}_f^{(t)}$ is the averaged velocity of the toes and heels in world coordinates when their contact probability, $p^{(t)}$ , is higher than a threshold. However, this velocity adjustment often introduces noisy translation when the contact and pose estimation is inaccurate. Therefore, we propose a simple learning mechanism in which a trajectory refining network, $R_{T}$ , updates the root orientation and velocity to address this issue. Finally, the global translation is computed through a roll-out operation: + +$$ +\begin{array}{l} \left(\Gamma^ {(t)}, v ^ {(t)}\right) = R _ {T} \big (\phi_ {m} ^ {(0)}, \Gamma_ {0} ^ {(0)}, \tilde {v} ^ {(0)},..., \phi_ {m} ^ {(t)}, \Gamma_ {0} ^ {(t)}, \tilde {v} ^ {(t)} \big), \\ \tau^ {(t)} = \sum_ {i = 0} ^ {t - 1} \Gamma^ {(i)} v ^ {(i)}. \\ \end{array} +$$ + +In summary, this full process reconstructs accurate 3D human pose and shape in both the camera and world coordinates from a single monocular video sequence (Fig. 2). + +# 3.3. Training + +Pretraining on AMASS. We train in two stages: (1) pretraining with synthetic data, and (2) fine-tuning with real data (Fig. 3). The objective of the pretraining stage is to teach the motion encoder to extract motion context from the input 2D keypoint sequence. The motion and trajectory decoders then map this motion context to the corresponding 3D motion and global trajectory spaces (i.e. they lift the encoding to 3D). We use the AMASS dataset [31] to generate an extensive set of synthetic pairs consisting of sequences of 2D keypoints together with the ground truth SMPL parameters. + +To synthesize 2D keypoints from AMASS, we create virtual cameras onto which we project 3D keypoints derived from the ground truth mesh. Unlike MotionBERT [59] and + +![](images/7969ce51517c8e6e1f5aeb41fee4efc9565047caa0171225060958dbbd635b69.jpg) + +![](images/392b61eee0dc7ddbda862f69f72d090dd8d6d7edf4ec41ebb641348740e34203.jpg) +Figure 3. WHAM's Two-Stage Training Scheme. During pre-taining, we generate synthetic 2D keypoint sequences from AMASS [31] and train a motion encoder and decoder on the generated data (top). We then leverage video datasets with ground truth SMPL parameters, for which there is much less data. We use the fixed-weight pre-trained image encoder and keypoints detector $(\text{※})$ to extract image features and 2D keypoints. In this stage, we train the feature integration network while fine-tuning the motion encoder and motion/trajectory decoders, marked $\text{※}$ (bottom). + +ProxyCap [58], which use static cameras for keypoint projection, we employ dynamic cameras that incorporate both rotational and translational motion. This choice has two main motivations. First, it accounts for the inherent differences between human motions captured in static and dynamic camera setups. Second, it enables the learning of a camera-agnostic motion representation, from which the trajectory decoder can reconstruct the global trajectory. We also augment the 2D data with noise and masking. For details of the synthetic generation process see Sup. Mat. + +Fine-tuning on Video Datasets. Starting with the pretrained network, we fine-tune WHAM on four video datasets: 3DPW [47], Human3.6M [10], MPI-INF-3DHP [32], and InstaVariety [14]. For the human mesh recovery task, we supervise WHAM on ground-truth SMPL parameters from AMASS and 3DPW, 3D keypoints from Human3.6M and MPI-INF-3DHP, and 2D keypoints from InstaVariety. For the global trajectory estimation task, we use AMASS, Human3.6M, and MPI-INF-3DHP. Additionally, during training we experiment with adding BEDLAM [1], a large synthetic dataset with realistic video and ground truth SMPL parameters. + +The fine-tuning has two objectives: 1) exposing the network to real 2D keypoints, instead of training it solely on synthetic data, and 2) training the feature integrator network to aggregate motion and image features. To achieve these goals, we jointly train the entire network on the video datasets while setting a smaller learning rate on the pretrained modules (see Fig. 3). Consistent with prior work [5, 16, 29, 42, 50], we employ a pretrained and fixed-weight image encoder [20] to extract image features. However, to leverage recent network architectures and training strategies, we also experiment with different types of encoders [1, 6, 24] in the following section. Implementation Details. For the pretraining stage, we train + +WHAM on AMASS for 80 epochs with the learning rate of $5 \times 10^{-4}$ . Then we finetune WHAM on 3DPW, MPI-INF-3DHP, Human3.6M, and InstaVariety for 30 epochs. We use a learning rate of $1 \times 10^{-4}$ for the feature integrator and $1 \times 10^{-5}$ for the pretrained components during finetuning. During training, we use the Adam optimizer and a batch size of 64 and split sequences into 81-frame chunks. + +# 4. Experiments + +Datasets. We evaluate WHAM on three in-the-wild benchmarks: 3DPW [47], RICH [9], and EMDB [15]. Following previous work [1, 5, 13, 16, 20, 24, 29], we perform the evaluation in camera coordinates. The estimated global trajectory is evaluated on a subset of EMDB (EMDB 2) for which they provide ground truth global motion with dynamic cameras (used for evaluation). We also test on new sequences captured using an iPhone with the gyroscope. See Sup. Mat. for more details on the datasets and iPhone results. + +Evaluation metrics. To evaluate the accuracy of 3D human pose and shape estimation, we compute Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PA-MPJPE), and Per Vertex Error (PVE) measured in millimeters $(mm)$ . We compute Acceleration error (Accel, in $m / s^2$ ) to measure the inter-frame smoothness of the reconstructed motion. We also evaluate the motion reconstruction and trajectory estimation accuracy in the world-frame. Following previous work [19, 53], we split sequences into smaller segments of 100 frames and align each output segment with the ground-truth data using the first two frames $\mathrm{(W - MPJPE_{100})}$ or the entire segment $\mathrm{(WA - MPJPE_{100})}$ in mm. These previous metrics give an unrealistic picture of 3D performance as they do not measure drift over long sequences. Therefore, we also evaluate the error over the entire trajectory after the rigid alignment and measure Root Translation Error (RTE in %) normalized by the actual displacement of the person. We also assess the jitter of the motion in the world coordinate system in $10m / s^3$ and foot sliding during the contact (FS in mm). + +# 4.1. 3D Human Motion Recovery + +Per-frame accuracy. Table 1 presents a comprehensive comparison of WHAM and the existing state-of-the-art per-frame and video-based methods across three benchmark datasets [9, 15, 47]. Because none of the methods are exposed to data from RICH or EMDB during training, results on these datasets are indicative of each method's ability to generalize. WHAM (Res), WHAM (HR), and WHAM (ViT) correspond to different architectures for the pretrained image encoders, derived from SPIN (ResNet-50) [20], CLIFF (HRNet-W48) [1, 24], and HMR2.0 (ViT-H/16) [6], respectively. Not surprisingly, WHAM (HR) is more accurate than WHAM (Res), + +Previous work follows [14] in reporting Accel in $mm / \text{frame}^2$ . To remove the dependency on frame rate, we convert all previous results to $m / s^2$ . + +![](images/e78d245f4a53bc4d530334e5d581d76085a986972024f3f22c9b8864f7e6c9a0.jpg) + +![](images/cddb34224c4ffaa183f9d634e571b3e2821a426e0e1ef72268b37d680accbf65.jpg) + +![](images/db656ab550ff811dbb5d14f2723b23f772b914999419f8b23c57a16056e2277a.jpg) + +![](images/1895002a458b99197ef2a95b6f752f200fbeffd1ca827bbcda74092f79670b6e.jpg) + +![](images/30be85eceac94604ab062647e1b1eca513fca27cc6bd38d7e3eda92a730824fb.jpg) + +![](images/20477e4785edbd57c64195ccc0e3fefd676f995afcd2bb721c9f08a40ba31b8e.jpg) +Figure 4. Qualitative comparison with previous state-of-the-art methods for 3D human pose and shape estimation. See text. + +while the transformer-based version, WHAM (ViT), is the most accurate. The backbone matters, with WHAM (ViT) outperforming all previous methods on all per-frame metrics (MPJPE, PA-MPJPE, and PVE) on all benchmarks. Even with the simplest ResNet backbone, WHAM (Res) outperforms every method except for ReFit on RICH. + +Training on BEDLAM consistently improves accuracy in prior work (see [1, 3, 49]), and we find the same here as shown in Table 2. However, compared to ReFit, WHAM exhibits relatively smaller performance improvement by adding BEDLAM. One possible reason for this is that, unlike ReFit, we did not finetune the image encoder during training to limit training time. + +Inter-frame smoothness. We also evaluate the inter-frame smoothness using the acceleration error. Compared with state-of-the-art per-frame methods [1, 6, 22, 24], WHAM has significantly lower acceleration error. This indicates that WHAM reconstructs smooth and more plausible 3D human motion across frames while not sacrificing high per-frame accuracy. On the other hand, when compared to recent temporal methods [5, 42, 50], WHAM exhibits comparable or slightly higher acceleration error. However, we observe that these video-based methods tend to over-smooth the human motion, resulting in lower accuracy on per-frame metrics. + +To provide intuition for these numbers, we qualitatively compare WHAM with TCMR [5] and GLoT [42] in Fig. 4. While producing smooth results, TCMR and GLoT fail to capture the bending of the left knee when the subject is ascending the stairs, while WHAM more accurately reconstructs the 3D human pose. + +# 4.2. 3D Global Trajectory Recovery + +To evaluate global trajectory recovery, we compare WHAM with the state-of-the-art methods and a baseline that combines a SLAM method (DPVO [45]) and a per-frame + +
Models3DPW (14)RICH (24)EMDB (24)
PA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccel
per-frameSPIN [20]59.296.9112.831.469.7122.9144.235.287.1140.7166.141.3
PARE* [17]46.574.588.6-60.7109.2123.5-72.2113.9133.2-
CLIFF* [24]43.069.081.222.556.6102.6115.022.468.3103.5123.724.5
HybrIK* [22]41.871.682.3-56.496.8110.4-65.6103.0122.2-
HMR2.0 [6]44.469.882.218.148.196.0110.918.860.798.3120.819.9
ReFit* [49]40.565.375.118.547.980.792.917.158.688.0104.520.7
temporalTCMR* [5]52.786.5101.46.065.6119.1137.75.079.8127.7150.25.3
VIBE* [16]51.982.998.418.568.4120.5140.221.881.6126.1149.926.5
MPS-Net* [50]52.184.399.06.567.1118.2136.75.881.4123.3143.96.2
GLoT* [42]50.680.796.46.065.6114.3132.75.279.1119.9140.85.4
GLAMR [55]51.1--8.079.9--107.773.8113.8134.933.0
TRACE* [43]50.979.195.428.6----71.5110.0129.625.5
SLAHMR [53]55.9---52.5--9.469.793.7111.37.1
PACE [19]----49.3--8.8----
WHAM (Res)*40.262.775.16.351.889.5103.25.057.884.099.75.2
WHAM (HR)*39.062.674.86.449.184.696.45.257.185.7103.25.6
WHAM (ViT)*35.957.868.76.644.380.091.25.350.479.794.45.3
+ +Table 1. Quantitative comparison of state-of-the-art models on the 3DPW [47], RICH [9], and EMDB [15] datasets. Ordering of per-frame and temporal methods is done separately by descending MPJPE on EMDB (except for PACE). For testing on EMDB, we follow the protocol of EMDB 1 [15]. Parenthesis denotes the number of body joints used to compute MPJPE and PA-MPJPE, and * denotes models trained with the 3DPW training set. Bold numbers denote the most accurate method in each column. Accel is in $m / s^2$ , all other errors are in mm. + +
ModelsDataset3DPW (14)
PA-MPJPEMPJPEPVEAccel
CLIFF [1, 24]R43.668.882.119.2
R+B43.066.978.531.0
ReFit [49]R40.565.375.118.5
R+B38.257.667.621.4
WHAM (ViT)R35.957.868.76.6
R+B35.756.967.46.7
+ +Table 2. Dataset ablation experiments on 3DPW [47]. R denotes the use of real datasets and B denotes BEDLAM. + +
ModelsEMDB 2
WA-MPJPE100W-MPJPE100RTEJitterFS
DPVO (+ HMR2.0) [6, 45]647.82231.415.8537.3107.6
GLAMR [55]280.8726.611.446.320.7
TRACE [43]529.01702.317.72987.6370.7
SLAHMR [53]326.9776.110.231.314.5
WHAM (w/DPVO [45])135.6354.86.022.54.4
WHAM (w/DROID [44])133.3343.94.621.54.4
WHAM (w/ GT gyro)131.1335.34.121.04.4
+ +Table 3. Global motion estimation accuracy on EMDB [15]. + +method (HMR2.0 [6]); see Table 3. WHAM is agnostic to the source of the camera angular velocity and we compare results using DPVO, DROID-SLAM [44] and the ground truth angular velocity (gyro). + +As shown in Table 3, WHAM outperforms the existing methods on all metrics. Specifically, combining WHAM with DPVO is more accurate than the global trajectory estimation of DPVO combined with HMR2.0, illustrating that our method actively refines the global trajectory instead of performing a simple integration. DROID-SLAM gives slightly better results than DPVO. Furthermore, WHAM significantly outperforms the regression-based method, TRACE, on jitter and foot sliding metrics. We further demonstrate this in Figs. 5 and 1, where WHAM captures more consistent and plausible human motion in the global coordinate system than TRACE and SLAHMR for videos captured by dynamic cameras. As depicted in Fig. 6, WHAM outperforms GLAMR, TRACE, and SLAHMR in capturing the pattern of human motion in the global coordinate system. + +# 4.3. Ablation Study + +To provide further insight into our approach, we conduct ablation studies to analyze the contribution of each component to the performance. As shown in Table 4, our entire system (WHAM) outperforms the different variants of WHAM that ablate a single component. To be specific, we first observe that adding feature integration improves both motion and global trajectory estimation accuracy when compared with an ablated version without feature integration $(\mathrm{w / o}F_I)$ . Similarly, the removal of the pretraining on + +![](images/182970dcdeee0d1bf6d9a67512ac5f6d202d89b4dfbaafa7646a53c8ef95001e.jpg) +Figure 5. Qualitative comparison with TRACE [43] and SLAHMR [53] on global human motion estimation with dynamic cameras. + +![](images/06647819792b926e3c1cea4ad54a3ec6ec7a526c29b5544816ef00e894718ba5.jpg) +Figure 6. Comparison of global trajectory estimation on EMDB [15]. Overall, WHAM shows better alignment to ground truth data compared to GLAMR [55], TRACE [43], and SLAHMR [53]. + +the 2D-to-3D lifting task using AMASS [31] (w/o lifting) shows significant performance degradation. WHAM also outperforms the ablation of the Neural Initialization (w/o NI), particularly in the MPJPE metric. In addition, we experiment with WHAM to decode trajectory solely based on the motion context without using the estimated camera angular velocity $(\mathrm{w / o}\omega)$ . Although this version shows similar performance in predicting 3D human pose, it suffers from the entanglement of camera and human motion, resulting in significantly high overall trajectory error (RTE). Last, we observe that WHAM without the trajectory refinement (w/o traj. ref.) gives larger global trajectory and foot sliding errors in return for less jitter, indicating that our approach contributes to the global trajectory accuracy and helps reduce foot sliding. Core details are presented here; see Sup. Mat. for more details and information on run-time cost. + +# 5. Conclusion + +WHAM is a new method to recover accurate 3D human motion in global coordinates from a moving camera more ef + +
ModelsEMDB 2
PA-MPJPEMPJPEWA-MPJPE100W-MPJPE100RTEJitterFS
w/o F144.269.0147.6377.96.323.15.5
w/o lifting60.383.0238.0693.011.524.55.0
w/o NI40.466.3142.7368.16.822.34.6
w/o ω39.162.0156.5422.010.122.15.0
w/o traj. ref.38.259.3154.7407.56.318.86.5
WHAM (Ours)38.259.3135.6354.86.022.54.4
+ +Table 4. Ablation experiments. See text. + +ficiently and accurately than the state-of-the-art approaches. Our approach leverages the AMASS dataset to train a network to recursively lift 2D sequences of keypoints to sequences of 3D SMPL parameters. But keypoints alone lack valuable information about the body and its movement. Consequently, we integrate image context information over time and learn to combine it with the motion context to better estimate human body shape and pose. Additionally, our method takes an estimate of the camera angular velocity, which can either be computed from a SLAM method or from the camera's gyro when available. Finally, we combine all this information with an estimate of foot contact to recover the 3D human motion in global coordinates from a monocular video sequence. WHAM significantly outperforms the existing state-of-the-art methods (both image-based and video-based) on challenging in-the-wild benchmarks in both 3D HPS and the world-coordinate trajectory estimation accuracy. Because of its speed and accuracy, WHAM provides a foundation for in-the-wild motion capture applications. + +Limitations and future directions: WHAM learns about human motion from AMASS, limiting generalization to motions that are out of distribution. While we employ random masking as part of our data synthesis process, our generating approach mainly assumes the scenario where the full body is within the field of view. See Sup. Mat. for more details. + +WHAM opens up many directions for future work. For example, while we use SLAM to estimate the camera's angular velocity, SLAM could also provide camera intrinsics and extrinsics as well as information about the 3D scene that could be used to enforce consistency between the scene and the human. While WHAM is an online method, designed for real-time applications, it could also initialize an optimization-based post-processing akin to bundle adjustment, which would optimize the camera, scene, and human motion together. Furthermore, a real-time and phone-based implementation of WHAM should be feasible. + +Acknowledgements. Part of this work was done when the first author was an intern at the Max Planck Institute for Intelligence System. This work was partially supported through an NSF CAREER Award (CBET 2145473) and a Chan-Zuckerberg Essential Open Source Software for Science Award. CoI Disclosure. https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt. + +# References + +[1] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8726-8737, 2023. 2, 4, 5, 6, 7 +[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016, pages 561-578. Springer International Publishing, 2016. 2 +[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. SMPLer-X: Scaling up expressive human pose and shape estimation. In Proc. Neural Information Processing Systems (NeurIPS), 2023. 6 +[4] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020. 3 +[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 4, 5, 6, 7 +[6] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091, 2023. 2, 3, 4, 5, 6, 7 +[7] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[8] Dorian F. Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. *Bodyslam++*: Fast and tightly-coupled visual-inertial camera and human motion tracking, 2023. 3 +[9] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 2, 6, 7 +[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 5 +[11] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020. 2, 3 +[12] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020. +[13] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and + +pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 6 +[14] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. arXiv preprint arXiv:1812.01601, 2018. 2, 3, 5, 6 +[15] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 6, 7, 8 +[16] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5252-5262. IEEE, 2020. 2, 3, 5, 6, 7 +[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2, 3, 7 +[18] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11015-11025, Piscataway, NJ, 2021. IEEE. 3 +[19] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2, 3, 6, 7 +[20] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2, 3, 4, 5, 6, 7 +[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 2 +[22] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021. 2, 3, 6, 7 +[23] Jiefeng Li, Siyuan Bian, Qi Liu, Jiasheng Tang, Fan Wang, and Cewu Lu. NIKI: Neural inverse kinematics with invertible neural networks for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[24] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2, 3, 4, 5, 6, 7 +[25] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 2 + +[26] Kevin Lin, Chung-Ching Lin, Lin Liang, Zicheng Liu, and Lijuan Wang. Mpt: Mesh pre-training with transformers for human pose and mesh reconstruction, 2023. 2 +[27] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, 2021. 2 +[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3 +[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3, 4, 5, 6 +[30] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 534-543, 2023. 3 +[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 5, 8 +[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3D Vision (3DV), 2017 Fifth International Conference on. IEEE, 2017. 2, 5 +[33] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2, 3 +[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2, 3 +[35] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudo-gts of 3d human mesh estimation in the wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2, 3 +[36] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Computer Vision - ECCV 2020, pages 598-613. Springer, 2020. 2 +[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2 +[38] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459-468, 2019. 3 +[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human + +motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3 +[40] Nitin Saini, Chun-Hao P. Huang, Michael J. Black, and Aamir Ahmad. Smartmocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, pages 1-8, 2023. 2 +[41] Simon Schaefer, Dorian F. Henning, and Stefan Leutenegger. Glopro: Globally-consistent uncertainty-aware 3d human pose estimation & tracking in the wild, 2023. 2, 3 +[42] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 2, 3, 4, 5, 6, 7 +[43] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 7, 8 +[44] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 5, 7 +[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 3, 5, 6, 7 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3 +[47] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 5, 6, 7 +[48] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In The IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[49] Yufu Wang and Kostas Daniilidis. Refit: Recurrent fitting network for 3d human recovery. In International Conference on Computer Vision, 2023. 6, 7 +[50] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 7 +[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: generative 3D human shape and articulated pose models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 2 +[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose + +estimation. In Advances in Neural Information Processing Systems, 2022. 2, 3 +[53] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8 +[54] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4 +[55] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 7, 8 +[56] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[57] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3 +[58] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Hongwei Yi, Shengping Zhang, and Yebin Liu. Real-time monocular full-body capture in world space via sequential proxy-to-motion learning. arXiv preprint arXiv:2307.01200, 2023. 2, 3, 5 +[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 5 \ No newline at end of file diff --git a/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/images.zip b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9506dfe49fc1457cccdde06a90d5a76c7ee4dfd8 --- /dev/null +++ b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f259be18d6c5a9215afbdd3473e03f52784e4efbbbd7cb9f46eeca57a78f2880 +size 586772 diff --git a/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/layout.json b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..132762282ee81bb58975c8892649b41b4d5f7a39 --- /dev/null +++ b/2024/WHAM_ Reconstructing World-grounded Humans with Accurate 3D Motion/layout.json @@ -0,0 +1,8497 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 120 + ], + "type": "text", + "content": "WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 143, + 186, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 143, + 186, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 143, + 186, + 158 + ], + "type": "text", + "content": "Soyong Shin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 143, + 276, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 143, + 276, + 158 + ], + "spans": [ + { + "bbox": [ + 210, + 143, + 276, + 158 + ], + "type": "text", + "content": "Juyong Kim" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 301, + 143, + 359, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 143, + 359, + 157 + ], + "spans": [ + { + "bbox": [ + 301, + 143, + 359, + 157 + ], + "type": "text", + "content": "Eni Halilaj" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 384, + 143, + 473, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 143, + 473, + 156 + ], + "spans": [ + { + "bbox": [ + 384, + 143, + 473, + 156 + ], + "type": "text", + "content": "Michael J. Black" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "spans": [ + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent System " + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "inline_equation", + "content": "\\{" + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "text", + "content": " soyongs, juyongk, ehalilaj\\} " + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 109, + 165, + 483, + 194 + ], + "type": "text", + "content": " andrew.cmu.edu black@tuebingen.mpg.de" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 204, + 96, + 284 + ], + "blocks": [ + { + "bbox": [ + 50, + 204, + 96, + 284 + ], + "lines": [ + { + "bbox": [ + 50, + 204, + 96, + 284 + ], + "spans": [ + { + "bbox": [ + 50, + 204, + 96, + 284 + ], + "type": "image", + "image_path": "02de85de1d553150e6003c73f47b9ebc3857dde697d07567bdcb53729d637c96.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 285, + 96, + 364 + ], + "blocks": [ + { + "bbox": [ + 50, + 285, + 96, + 364 + ], + "lines": [ + { + "bbox": [ + 50, + 285, + 96, + 364 + ], + "spans": [ + { + "bbox": [ + 50, + 285, + 96, + 364 + ], + "type": "image", + "image_path": "55c6900b3437267b9e552ea97719da27a26c57afb96518b6d56a592b526da379.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 372, + 546, + 428 + ], + "lines": [ + { + "bbox": [ + 46, + 372, + 546, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 546, + 428 + ], + "type": "text", + "content": "Figure 1. WHAM: World-grounded Humans with Accurate Motion. State-of-the-art methods like TRACE [43] and SLAHMR [53] fail to capture global 3D human trajectories accurately when given in-the-wild videos captured using a moving camera, producing implausible world-grounded motion (e.g., foot sliding). To address this, WHAM uses two novel strategies: (1) feature integration from 2D keypoints and pixels to reconstruct precise and pixel-aligned 3D human motion and (2) contact-aware trajectory recovery to place the human in a global coordinate system without foot sliding. Gray dots show the ground-truth global trajectory. See Supplemental Video." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 97, + 205, + 144, + 284 + ], + "blocks": [ + { + "bbox": [ + 97, + 205, + 144, + 284 + ], + "lines": [ + { + "bbox": [ + 97, + 205, + 144, + 284 + ], + "spans": [ + { + "bbox": [ + 97, + 205, + 144, + 284 + ], + "type": "image", + "image_path": "6ae0c45593a6d230d633af96a4346b8b41e66b5f7a71c46867f2cf51b181089b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 97, + 285, + 145, + 364 + ], + "blocks": [ + { + "bbox": [ + 97, + 285, + 145, + 364 + ], + "lines": [ + { + "bbox": [ + 97, + 285, + 145, + 364 + ], + "spans": [ + { + "bbox": [ + 97, + 285, + 145, + 364 + ], + "type": "image", + "image_path": "1ab30b270c67489a9234fcd8cfe611b1651f1f304a0fd6449f4e05c71146dd30.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 148, + 205, + 192, + 284 + ], + "blocks": [ + { + "bbox": [ + 148, + 205, + 192, + 284 + ], + "lines": [ + { + "bbox": [ + 148, + 205, + 192, + 284 + ], + "spans": [ + { + "bbox": [ + 148, + 205, + 192, + 284 + ], + "type": "image", + "image_path": "5fb411d59626f103f39d2d91c24cab0859a9b7601efbdb1576815ca16ecd8bda.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 148, + 285, + 192, + 364 + ], + "blocks": [ + { + "bbox": [ + 148, + 285, + 192, + 364 + ], + "lines": [ + { + "bbox": [ + 148, + 285, + 192, + 364 + ], + "spans": [ + { + "bbox": [ + 148, + 285, + 192, + 364 + ], + "type": "image", + "image_path": "1ecb7c67e515e8b0761b9d7b4ade6e1456b734e032aedeb3dde91e5fca698f96.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 194, + 205, + 240, + 284 + ], + "blocks": [ + { + "bbox": [ + 194, + 205, + 240, + 284 + ], + "lines": [ + { + "bbox": [ + 194, + 205, + 240, + 284 + ], + "spans": [ + { + "bbox": [ + 194, + 205, + 240, + 284 + ], + "type": "image", + "image_path": "7ca2367ede506337ffe951dd7b79558904968a72a0f20da31a3d929ea266edc7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 194, + 285, + 239, + 363 + ], + "blocks": [ + { + "bbox": [ + 194, + 285, + 239, + 363 + ], + "lines": [ + { + "bbox": [ + 194, + 285, + 239, + 363 + ], + "spans": [ + { + "bbox": [ + 194, + 285, + 239, + 363 + ], + "type": "image", + "image_path": "9dab6a5c367779b47e013dc525d716607eae46c73878ca694cb371e98180d571.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 240, + 205, + 256, + 284 + ], + "blocks": [ + { + "bbox": [ + 240, + 205, + 256, + 284 + ], + "lines": [ + { + "bbox": [ + 240, + 205, + 256, + 284 + ], + "spans": [ + { + "bbox": [ + 240, + 205, + 256, + 284 + ], + "type": "image", + "image_path": "166cfae0e92bc02bb9a9dcb94f3bdb6cff144eed3a63d790b4dac2ba4ad0a522.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 240, + 285, + 255, + 363 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 285, + 255, + 363 + ], + "spans": [ + { + "bbox": [ + 240, + 285, + 255, + 363 + ], + "type": "image", + "image_path": "6a29e0dec3e496a067707fbac78e952690a5fe10bd8c3155f28f07dc11d3028d.jpg" + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 256, + 205, + 305, + 284 + ], + "blocks": [ + { + "bbox": [ + 256, + 205, + 305, + 284 + ], + "lines": [ + { + "bbox": [ + 256, + 205, + 305, + 284 + ], + "spans": [ + { + "bbox": [ + 256, + 205, + 305, + 284 + ], + "type": "image", + "image_path": "96dbba8d7c47704d0c29f7ac033f83d5e4c1802c42563e5f977cb486c214962a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 256, + 285, + 305, + 363 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 285, + 305, + 363 + ], + "spans": [ + { + "bbox": [ + 256, + 285, + 305, + 363 + ], + "type": "image", + "image_path": "c8a7722e375f3169d7f403599342adc7f487319be1a60f1159699c8a2e13118f.jpg" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "bbox": [ + 244, + 205, + 256, + 284 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 205, + 256, + 284 + ], + "spans": [ + { + "bbox": [ + 244, + 205, + 256, + 284 + ], + "type": "image", + "image_path": "6d21151ffb6ab332f2b19db1e134435648696c95dac9e12ad504f8ac1c69e0e3.jpg" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "bbox": [ + 244, + 285, + 255, + 363 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 285, + 255, + 363 + ], + "spans": [ + { + "bbox": [ + 244, + 285, + 255, + 363 + ], + "type": "image", + "image_path": "761592e9d0e83445de2fc90bd2863fe10ad14cb5bf739a059af3b5a2bf4e3026.jpg" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "bbox": [ + 257, + 215, + 305, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 215, + 305, + 225 + ], + "spans": [ + { + "bbox": [ + 257, + 215, + 305, + 225 + ], + "type": "text", + "content": "GT trajectory" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 320, + 215, + 338, + 225 + ], + "blocks": [ + { + "bbox": [ + 320, + 215, + 338, + 225 + ], + "lines": [ + { + "bbox": [ + 320, + 215, + 338, + 225 + ], + "spans": [ + { + "bbox": [ + 320, + 215, + 338, + 225 + ], + "type": "image", + "image_path": "c6e3f307f386fe7c11473a5c836a1f82fff316b3d6473f3767fb9a932184ec23.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 216, + 367, + 224 + ], + "lines": [ + { + "bbox": [ + 340, + 216, + 367, + 224 + ], + "spans": [ + { + "bbox": [ + 340, + 216, + 367, + 224 + ], + "type": "text", + "content": "TRACE" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 389, + 215, + 405, + 225 + ], + "blocks": [ + { + "bbox": [ + 389, + 215, + 405, + 225 + ], + "lines": [ + { + "bbox": [ + 389, + 215, + 405, + 225 + ], + "spans": [ + { + "bbox": [ + 389, + 215, + 405, + 225 + ], + "type": "image", + "image_path": "f12dbcac6531be2968ed4fb4678344e83764c9a79f918ec134d3e023463cdafc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 216, + 441, + 224 + ], + "lines": [ + { + "bbox": [ + 408, + 216, + 441, + 224 + ], + "spans": [ + { + "bbox": [ + 408, + 216, + 441, + 224 + ], + "type": "text", + "content": "SLAHMR" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 461, + 215, + 477, + 225 + ], + "blocks": [ + { + "bbox": [ + 461, + 215, + 477, + 225 + ], + "lines": [ + { + "bbox": [ + 461, + 215, + 477, + 225 + ], + "spans": [ + { + "bbox": [ + 461, + 215, + 477, + 225 + ], + "type": "image", + "image_path": "21bd141faa0eb986fb6901c57fbfa0b36b452804fb648cee1717e232c87589d7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 215, + 530, + 225 + ], + "lines": [ + { + "bbox": [ + 481, + 215, + 530, + 225 + ], + "spans": [ + { + "bbox": [ + 481, + 215, + 530, + 225 + ], + "type": "text", + "content": "WHAM (Ours)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 239, + 239, + 542, + 364 + ], + "blocks": [ + { + "bbox": [ + 239, + 239, + 542, + 364 + ], + "lines": [ + { + "bbox": [ + 239, + 239, + 542, + 364 + ], + "spans": [ + { + "bbox": [ + 239, + 239, + 542, + 364 + ], + "type": "image", + "image_path": "1bcb5be1c007b058b27d63decabfdb1aef80b7fea3189cb61302638df128f592.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 143, + 449, + 191, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 449, + 191, + 461 + ], + "spans": [ + { + "bbox": [ + 143, + 449, + 191, + 461 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 45, + 486, + 290, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 486, + 290, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 486, + 290, + 713 + ], + "type": "text", + "content": "The estimation of 3D human motion from video has progressed rapidly but current methods still have several key limitations. First, most methods estimate the human in camera coordinates. Second, prior work on estimating humans in global coordinates often assumes a flat ground plane and produces foot sliding. Third, the most accurate methods rely on computationally expensive optimization pipelines, limiting their use to offline applications. Finally, existing video-based methods are surprisingly less accurate than single-frame methods. We address these limitations with WHAM (World-grounded Humans with Accurate Motion), which accurately and efficiently reconstructs 3D human motion in a global coordinate system from video. WHAM learns to lift 2D keypoint sequences to 3D using motion capture data and fuses this with video features, integrating motion context and visual information. WHAM exploits camera angular velocity estimated from a SLAM method together with human motion to estimate the body's global trajectory. We combine this with a contact-aware trajectory refinement method that" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 450, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 547, + 510 + ], + "type": "text", + "content": "lets WHAM capture human motion in diverse conditions, such as climbing stairs. WHAM outperforms all existing 3D human motion recovery methods across multiple in-the-wild benchmarks. Code is available for research purposes at http://wham.is.tue.mpg.de/." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 536, + 387, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 387, + 549 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 387, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 558, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 713 + ], + "type": "text", + "content": "Our goal is to accurately estimate the 3D pose and shape of a person from monocular video. This is a longstanding problem and, while the field has made rapid progress, several key challenges remain. First, human motion should be computed in a consistent global coordinate system. Second, the method should be computationally efficient. Third, the results should be accurate, temporally smooth, detailed, natural looking, and have realistic foot-ground contact. Fourth, the capture should work with an arbitrary moving camera. These constraints need to be satisfied to make markerless human motion capture widely available for applications in gaming, AR/VR, autonomous driving, sports analysis, and human-robot interaction. We address these challenges with" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2070" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "content": "WHAM (World-grounded Humans with Accurate Motion), which enables fast and accurate recovery of 3D human motion from a moving camera; see Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 229 + ], + "type": "text", + "content": "It seems natural that, in estimating 3D humans from video, we should be able to exploit the temporal nature of video. Counter-intuitively, existing video-based methods for 3D human pose and shape (HPS) estimation [5, 14, 16, 29, 42, 50] are less accurate than the best single-frame methods [6, 13, 17, 20, 22, 24, 25, 57, 57]. This may be an issue of training data. There are large datasets of single images with ground-truth 3D human poses containing a diversity of body shapes, poses, backgrounds, lighting, etc. In contrast, video datasets with ground truth are much more limited." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 386 + ], + "type": "text", + "content": "To address this, WHAM leverages both the large-scale AMASS motion capture (mocap) dataset [31] and video datasets. Our key idea is to learn about 3D human motion from AMASS and then learn to fuse this information with temporal image cues from video, getting the best of both. Similar to previous work [58, 59], we use AMASS to generate synthetic 2D keypoints and ground-truth motion sequence pairs, from which we pretrain a motion encoder, which captures the motion context, and decoder that lifts sequences of 2D keypoints to sequences of 3D poses. Given the robustness of recent 2D keypoint detection models [52, 56], our pretrained model does a good job of predicting human pose from video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 496 + ], + "type": "text", + "content": "Keypoints alone, however, are too sparse for accurate 3D mesh estimation. To improve accuracy, we jointly train a feature integrator network that merges information from video and 2D-keypoint sequences. We use a pretrained image encoder from previous work [1, 6, 20, 24] and train the feature integrator using video datasets [10, 14, 32, 47]. This integration process supplements the motion context extracted from the sparse 2D keypoints with dense visual context, significantly improving the recovered pose and shape accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "While the above approach produces accurate motion, we want this motion in global coordinates, unlike most previous methods that compute the body in camera coordinates. Estimating the global human trajectory is challenging when the camera is moving because the motions of the body and the camera are entangled. Recent work addresses this with optimization fitting based on a learned human motion prior and camera information from SLAM methods [19, 40, 53] or dense 3D scene information from COLMAP [27]. However, these methods are computationally expensive and far from real time. Recent regression-based methods are faster but either constrain the problem with static or known camera conditions [41, 58] or have temporal jitter and limited accuracy [43]. We tackle this challenge with two additional modules. First, we predict the global orientation and root velocity of the human from the sequence of 2D keypoints by training a global trajectory decoder. Specifically, we concatenate the camera's angular velocity to the context and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "train the global trajectory decoder to recursively predict the current orientation and root velocity, effectively factoring camera motion from human motion. WHAM takes the camera's angular velocity either from the output of a SLAM method or from a camera's gyroscope when available." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 546, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 286 + ], + "type": "text", + "content": "The above solution relies on knowledge of human motion learned from AMASS. Therefore, it can fail to capture elevation changes when the surface is not flat, e.g. when ascending the stairs because AMASS has a limited amount of such data. To address this, we introduce foot contact as an additional explicit form of motion context. We train WHAM to predict the likelihood of foot-ground contact using estimated contact labels from both AMASS and 3D video datasets. We then train a trajectory refinement network that outputs an update to the root orientation and velocity based on the information about the foot contact/velocity. This refinement enables WHAM to accurately estimate human motion in a global coordinate system even when the terrain is not flat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 288, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 547, + 430 + ], + "type": "text", + "content": "WHAM has very low computational overhead because it is an on-line algorithm that recursively predicts the pose, shape, and global motion parameters. The network, excluding preprocessing (bounding box detection, keypoint detection, and person identification), runs at 200 fps, significantly faster than prior methods. Also, despite not using global optimization like [53], we obtain accurate 3D camera trajectories and global body motions with minimal drift. Through extensive comparisons on multiple in-the-wild datasets as well as detailed ablation studies, we find that WHAM achieves state-of-the-art (SOTA) accuracy on 3D human pose estimation as well as global trajectory estimation (see Fig. 1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 431, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 575 + ], + "type": "text", + "content": "In summary, in this paper we: (1) introduce the first approach to effectively fuse 3D human motion context and video context for 3D HPS regression; (2) propose a novel global trajectory estimation framework that leverages motion context and foot contact to effectively address foot sliding and enable the 3D tracking of people on non-planar surfaces; (3) efficiently perform HPS regression in global coordinates; (4) achieve state-of-the-art (SOTA) performance on multiple in-the-wild benchmark datasets (3DPW [47], RICH [9], EMDB [15]). WHAM is the first video-based method to outperform all image-based and video-based methods on per-frame accuracy while maintaining temporal smoothness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 585, + 392, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 392, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 392, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": "Image-based 3D HPS Estimation. There are two broad classes of methods for recovering 3D HPS from images: model-free [21, 26, 33] and model-based [6, 11-13, 20, 24, 34, 35]. Here we focus on model-based methods, which estimate the low-dimensional parameters of a statistical body model [28, 36, 37, 51]. While early work explores optimization-based methods [2], here we focus on direct regression methods based on deep learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Many existing methods follow the architecture of HMR" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2071" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "[13], which uses a pretrained backbone to predict image features followed by a multilayer perceptron (MLP) that regresses SMPL [28] pose parameters from image features. Training such networks typically leverages paired images with SMPL parameters; these parameters are often pseudo-groundtruth (p-GT), estimated from 2D keypoints [11, 20, 24, 34, 35]. Other architectures for HPS regression have also been proposed [6, 17, 18, 22, 23, 57]. None of these methods use video or estimate the body in global coordinates. While quite accurate, when these image-based models are applied independently to frames of a video sequence, the shape and pose can be temporally inconsistent. In contrast, WHAM effectively aggregates temporal information to provide frame-accurate and temporally-coherent 3D HPS estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 254, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 289, + 469 + ], + "type": "text", + "content": "Video-Based 3D HPS. Video-based methods typically encode temporal information by combining static features extracted by a backbone from each frame. HMMR [14] uses a convolutional encoder, while VIBE [16] and MEVA [29] employ recurrent neural networks. TCMR [5] divides sequences into past, future, and whole frames, aggregating information to strongly constrain the output with motion consistency. MPS-Net [50] uses attention to capture non-local motion context and a hierarchical architecture to aggregate temporal features. Both MAED [48] and GLoT [42] use transformer architectures [46] to encode videos. MAED encodes videos in both temporal (across frames) and spatial (within each frame) dimensions and leverages the kinematic tree to iteratively regress each joint angle. GLoT encodes long-term temporal correlations and refines local details by focusing on nearby frames. Despite integrating information across frames, all existing video-based methods have lower accuracy than the best single-frame methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 472, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 289, + 663 + ], + "type": "text", + "content": "Given limited video training data with ground truth SMPL poses, several single-frame methods infer a mesh from 2D/3D keypoints [4, 7, 30, 33, 38] and use the keypoints as a proxy for training. Another set of approaches exploits 3D mocap data, which is plentiful [31], to train a network to lift 2D joints to 3D poses, which are used as a proxy for ground truth 3D. MotionBERT [59] synthesizes 2D keypoints through orthographic projection to learn a unified motion representation. ProxyCap [58] projects synthetic 3D keypoints into virtual cameras using a heuristic camera pose distribution. Despite benefiting from the scale of mocap datasets, these approaches do not fully utilize the visual information available in the video at test time. Here, we propose a combined network architecture and training strategy that leverages both proxy representations of human pose (lifting) and visual context extracted from video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Global 3D Motion Estimation with Dense Sensors. Several methods augment video data with other sensors to estimate 3D HPS in world coordinates. The 3DPW dataset [47] employs pre-calibrated body-worn inertial sensors and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "a handheld camera to jointly optimize the camera and human motion in challenging environments. Similarly, the EMDB dataset [15] uses electromagnetic sensors with an RGB-D camera, enabling accurate human motion capture in the world. While body-worn sensors aid global human motion reconstruction, they are intrusive, require cooperation, and do not help with archival video. BodySLAM++ [8] uses an optimization method with a visual-inertial sensor, comprising stereo cameras and an IMU. In contrast, we use a standard monocular camera, balancing accessibility and accuracy without using specialized equipment. While WHAM can take the camera gyro as input, this is not required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 217, + 548, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 548, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 548, + 480 + ], + "type": "text", + "content": "Monocular Global 3D Human Trajectory Estimation. Estimating the global human trajectory from a monocular dynamic camera is challenging. Previous work relies on learned prior distributions of human motion to separate human motion from camera motion. GLAMR [55] computes the global trajectory based on a predicted and inflated 3D motion sequence and optimizes it across multiple individuals in the scene. However, since GLAMR does not consider camera motion cues, the output trajectory may be noisy when the camera is rotating. SLAHMR [53] and PACE [19] use off-the-shelf SLAM algorithms [44, 45] and jointly optimize the camera and human motion to minimize the negative log-likelihood of a learned motion prior [39]. While they achieve good results, their optimization approach is computationally expensive. TRACE [43] is a pure regression method that utilizes optical flow as a motion cue and estimates multiple people at once, but lacks temporal consistency. GloPro [41] regresses the uncertainty of the global human motion in real-time, but requires known camera poses. In contrast, WHAM leverages both explicit and implicit prior knowledge of human motion and efficiently reconstructs accurate and temporally coherent 3D human motion in world coordinates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 506, + 367, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 367, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 367, + 519 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 528, + 373, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 373, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 373, + 540 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": "An overview of our World-grounded Human with Accurate Motion (WHAM) framework is illustrated in Fig. 2. The input to WHAM is a raw video data " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{I^{(t)}\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": ", captured by a camera with possibly unknown motion. Our goal is to predict the corresponding sequence of the SMPL model parameters " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\Theta^{(t)}\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": ", as well as the root orientation " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\Gamma^{(t)}\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\tau^{(t)}\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": ", expressed in the world coordinate system. We use ViTPose [52] to detect 2D keypoints " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{x_{2D}^{(t)}\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": " from which we obtain motion features " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\phi_m^{(t)}\\}_{t = 0}^T" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": " using the motion encoder. Additionally, we use a pretrained image encoder [6, 20, 24] to extract static image features " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\phi_i^{(t)}\\}_{t = 0}^T" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": " and integrate them with " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\phi_m^{(t)}\\}_{t = 0}^T" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": " to obtain fine-grained motion features " + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\phi}_m^{(t)}\\}_{t = 0}^T" + }, + { + "bbox": [ + 304, + 548, + 548, + 715 + ], + "type": "text", + "content": " from which we regress 3D human" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2072" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 521, + 210 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 521, + 210 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 521, + 210 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 521, + 210 + ], + "type": "image", + "image_path": "c1794a5e6d8edd55623e19d803efd1d0e3c8bdfa15231e1bd8783c3de45e9b88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 282 + ], + "type": "text", + "content": "Figure 2. An Overview of WHAM. WHAM takes the sequence of 2D keypoints estimated by a pretrained detector and encodes it into a motion feature. WHAM then updates the motion feature using another sequence of image features extracted from the image encoder through the feature integrator. From the updated motion feature, the Local Motion Decoder estimates 3D motion in the camera coordinate system and foot-ground contact probability. The Trajectory Decoder takes the motion feature and camera angular velocity to initially estimate the global root orientation and egocentric velocity, which are then updated through the Trajectory Refiner using the foot-ground contact. The final output of WHAM is pixel-aligned 3D human motion with the 3D trajectory in the global coordinates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 297, + 206, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 206, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 206, + 309 + ], + "type": "text", + "content": "motion in the world coordinate system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 317, + 173, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 173, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 173, + 329 + ], + "type": "text", + "content": "3.2. Network Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "text", + "content": "Uni-directional Motion Encoder and Decoder. In contrast to existing methods [5, 29, 42, 50, 59], which use windows with a fixed time duration, we use uni-directional recurrent neural networks (RNN) for the motion encoder and decoder, making WHAM suitable for online inference. The objective of the motion encoder, " + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "inline_equation", + "content": "E_{M}" + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "text", + "content": ", is to extract the motion context, " + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\phi_{m}^{(t)}" + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "text", + "content": ", from the current and previous sequence of 2D keypoints and the initial hidden state, " + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "inline_equation", + "content": "h_{E}^{(0)}" + }, + { + "bbox": [ + 46, + 336, + 288, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 443, + 242, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 443, + 242, + 460 + ], + "spans": [ + { + "bbox": [ + 91, + 443, + 242, + 460 + ], + "type": "interline_equation", + "content": "\\phi_ {m} ^ {(t)} = E _ {M} \\big (x _ {2 D} ^ {(0)}, x _ {2 D} ^ {(1)},..., x _ {2 D} ^ {(t)} | h _ {E} ^ {(0)} \\big).", + "image_path": "5aff4da25c95f218b454afbf3b5c051f27773fa25c7863c894cea06e8f8433bb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "text", + "content": "We normalize keypoints to a bounding box around the person and concatenate the box's center and scale to the keypoints, similar to CLIFF [24]. The role of the motion decoder, " + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "inline_equation", + "content": "D_{M}" + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "text", + "content": " is to recover SMPL parameters, " + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "inline_equation", + "content": "(\\theta ,\\beta)" + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "text", + "content": ", weak-perspective camera translation, " + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "text", + "content": ", and foot-ground contact probability, " + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 469, + 287, + 541 + ], + "type": "text", + "content": " from the motion feature history:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 550, + 264, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 550, + 264, + 567 + ], + "spans": [ + { + "bbox": [ + 69, + 550, + 264, + 567 + ], + "type": "interline_equation", + "content": "\\left(\\theta^ {(t)}, \\beta^ {(t)}, c ^ {(t)}, p ^ {(t)}\\right) = D _ {M} \\left(\\hat {\\phi} _ {m} ^ {(0)}, \\dots , \\hat {\\phi} _ {m} ^ {(t)} | h _ {D} ^ {(0)}\\right).", + "image_path": "4a368cef59eb8e522e38c31a1165f5313bfb5708293f495a2ae93c44ffb7fa19.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_m^{(t)}" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": " is the motion feature integrated with the image feature " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\phi_i^{(t)}" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": " (described below). During pretraining on synthetic data, the image feature is not available and we set " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_m^{(t)} = \\phi_m^{(t)}" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": ". As the encoder and decoder are tasked with reconstructing a dense 3D representation " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": " from a sparse 2D input signal " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "x_{2D}" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": ", we design an intermediate task to predict the 3D keypoints " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "x_{3D}" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": " and use them as the intermediate motion representation. This cascaded approach guides " + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\phi_m" + }, + { + "bbox": [ + 46, + 578, + 289, + 713 + ], + "type": "text", + "content": " to represent the implicit context of motion and the 3D spatial structure of the body. Similar to PIP [54], we use Neural Initialization that uses MLP to initialize the hidden state of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 295, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 320 + ], + "type": "text", + "content": "the motion encoder and decoder, " + }, + { + "bbox": [ + 304, + 295, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\left(h_{E}^{(0)}, h_{D}^{(0)}\\right)" + }, + { + "bbox": [ + 304, + 295, + 545, + 320 + ], + "type": "text", + "content": "; see section 4.3 and Sup. Mat. for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "text", + "content": "Motion and Visual Feature Integrator. We use the AMASS dataset to synthetically generate 2D sequences by projecting 3D SMPL joints into images with varied camera motions. This provides effectively limitless training data that is far more diverse than existing video datasets that contain ground truth 3D pose and shape. Although we leverage the temporal human motion context, lifting 2D keypoints to 3D meshes is an ambiguous task. A key idea is to augment this 2D keypoint information with image cues that can help disambiguate the 3D pose. Specifically, we use an image encoder [1, 6, 20, 24], pretrained on the human mesh recovery task, to extract image features " + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "text", + "content": ", which contain dense visual contextual information related to the 3D human pose and shape. We then train a feature integrator network, " + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "text", + "content": ", to combine " + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "inline_equation", + "content": "\\phi_{m}" + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 322, + 546, + 514 + ], + "type": "text", + "content": ", integrating motion and visual context. The feature integrator uses a simple yet effective residual connection:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 523, + 504, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 523, + 504, + 543 + ], + "spans": [ + { + "bbox": [ + 347, + 523, + 504, + 543 + ], + "type": "interline_equation", + "content": "\\hat {\\phi} _ {m} ^ {(t)} = \\phi_ {m} ^ {(t)} + F _ {I} \\left(\\operatorname {c o n c a t} \\left(\\phi_ {m} ^ {(t)}, \\phi_ {i} ^ {(t)}\\right)\\right).", + "image_path": "ea84ca6d24e9ba1b27ea5ea3bd14a75aaad135b1528475118677c2415ffee928.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 552, + 547, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 599 + ], + "type": "text", + "content": "This supplements motion features pretrained on the 2D-to-3D lifting task using AMASS with visual context, resulting in enriched motion features that use image evidence to help disambiguate the task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": "Global Trajectory Decoder. We design an additional decoder, " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": ", to predict the rough global root orientation " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma_0^{(t)}" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": " and root velocity " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "v_0^{(t)}" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": " from the motion feature " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\phi_m^{(t)}" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\phi_m" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": " is derived from the input signals in the camera coordinates, it is highly challenging to decouple the human and camera motion from it. To address this ambiguity, we append the angular velocity of the camera, " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\omega^{(t)}" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": ", to the motion feature, " + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\phi_m^{(t)}" + }, + { + "bbox": [ + 304, + 601, + 547, + 713 + ], + "type": "text", + "content": ", to create a camera-agnostic motion context. This design choice makes WHAM compatible with both off-the-shelf" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2073" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 122 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 122 + ], + "type": "text", + "content": "SLAM algorithms [44, 45] and gyroscope measurements that are widely available from modern digital cameras. We recursively predict global orientation, " + }, + { + "bbox": [ + 46, + 72, + 288, + 122 + ], + "type": "inline_equation", + "content": "\\Gamma_0^{(t)}" + }, + { + "bbox": [ + 46, + 72, + 288, + 122 + ], + "type": "text", + "content": ", using the unidirectional RNN." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 128, + 257, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 128, + 257, + 146 + ], + "spans": [ + { + "bbox": [ + 78, + 128, + 257, + 146 + ], + "type": "interline_equation", + "content": "\\big (\\Gamma_ {0} ^ {(t)}, v _ {0} ^ {(t)} \\big) = D _ {T} \\big (\\phi_ {m} ^ {(0)}, \\omega^ {(0)}, \\dots , \\phi_ {m} ^ {(t)}, \\omega^ {(t)} \\big).", + "image_path": "a9675930d4c1356dd2513fcb7c907c05871a372759d8d7660da44882bc32c597.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "text", + "content": "Contact Aware Trajectory Refinement. Good 3D motion in world coordinates in most scenarios implies accurate foot-ground contact without sliding. We want WHAM to generalize beyond flat ground planes, which are typically assumed in prior work. Specifically, our new trajectory refiner aims to resolve foot sliding and enables WHAM to generalize well to diverse motions, including climbing stairs. The refinement involves two stages. First, we adjust the ego-centric root velocity to " + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "inline_equation", + "content": "\\tilde{v}^{(t)}" + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "text", + "content": " to minimize foot sliding, based on the foot-ground contact probability " + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "inline_equation", + "content": "p^{(t)}" + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "text", + "content": ", estimated from the motion decoder " + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "inline_equation", + "content": "D_M" + }, + { + "bbox": [ + 46, + 152, + 288, + 284 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 291, + 223, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 291, + 223, + 309 + ], + "spans": [ + { + "bbox": [ + 110, + 291, + 223, + 309 + ], + "type": "interline_equation", + "content": "\\tilde {v} ^ {(t)} = v _ {0} ^ {(t)} - \\left(\\Gamma_ {0} ^ {(t)}\\right) ^ {- 1} \\bar {v} _ {f} ^ {(t)},", + "image_path": "94efcf35961c4f31e9befb4843813e95c2a37b0079c4ee46abe79b315a63393c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\bar{v}_f^{(t)}" + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "text", + "content": " is the averaged velocity of the toes and heels in world coordinates when their contact probability, " + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "inline_equation", + "content": "p^{(t)}" + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "text", + "content": ", is higher than a threshold. However, this velocity adjustment often introduces noisy translation when the contact and pose estimation is inaccurate. Therefore, we propose a simple learning mechanism in which a trajectory refining network, " + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "inline_equation", + "content": "R_{T}" + }, + { + "bbox": [ + 46, + 318, + 287, + 430 + ], + "type": "text", + "content": ", updates the root orientation and velocity to address this issue. Finally, the global translation is computed through a roll-out operation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 437, + 282, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 437, + 282, + 488 + ], + "spans": [ + { + "bbox": [ + 50, + 437, + 282, + 488 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\Gamma^ {(t)}, v ^ {(t)}\\right) = R _ {T} \\big (\\phi_ {m} ^ {(0)}, \\Gamma_ {0} ^ {(0)}, \\tilde {v} ^ {(0)},..., \\phi_ {m} ^ {(t)}, \\Gamma_ {0} ^ {(t)}, \\tilde {v} ^ {(t)} \\big), \\\\ \\tau^ {(t)} = \\sum_ {i = 0} ^ {t - 1} \\Gamma^ {(i)} v ^ {(i)}. \\\\ \\end{array}", + "image_path": "89dfaa5a330684e7425b1d72d9155dca47e2a32f1f42dc3af964424a38bbd7d4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 496, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 533 + ], + "type": "text", + "content": "In summary, this full process reconstructs accurate 3D human pose and shape in both the camera and world coordinates from a single monocular video sequence (Fig. 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 539, + 111, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 111, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 111, + 552 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "Pretraining on AMASS. We train in two stages: (1) pretraining with synthetic data, and (2) fine-tuning with real data (Fig. 3). The objective of the pretraining stage is to teach the motion encoder to extract motion context from the input 2D keypoint sequence. The motion and trajectory decoders then map this motion context to the corresponding 3D motion and global trajectory spaces (i.e. they lift the encoding to 3D). We use the AMASS dataset [31] to generate an extensive set of synthetic pairs consisting of sequences of 2D keypoints together with the ground truth SMPL parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "To synthesize 2D keypoints from AMASS, we create virtual cameras onto which we project 3D keypoints derived from the ground truth mesh. Unlike MotionBERT [59] and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 544, + 134 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 544, + 134 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 544, + 134 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 544, + 134 + ], + "type": "image", + "image_path": "7969ce51517c8e6e1f5aeb41fee4efc9565047caa0171225060958dbbd635b69.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 135, + 544, + 198 + ], + "blocks": [ + { + "bbox": [ + 309, + 135, + 544, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 135, + 544, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 135, + 544, + 198 + ], + "type": "image", + "image_path": "392b61eee0dc7ddbda862f69f72d090dd8d6d7edf4ec41ebb641348740e34203.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "type": "text", + "content": "Figure 3. WHAM's Two-Stage Training Scheme. During pre-taining, we generate synthetic 2D keypoint sequences from AMASS [31] and train a motion encoder and decoder on the generated data (top). We then leverage video datasets with ground truth SMPL parameters, for which there is much less data. We use the fixed-weight pre-trained image encoder and keypoints detector " + }, + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "type": "inline_equation", + "content": "(\\text{※})" + }, + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "type": "text", + "content": " to extract image features and 2D keypoints. In this stage, we train the feature integration network while fine-tuning the motion encoder and motion/trajectory decoders, marked " + }, + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\text{※}" + }, + { + "bbox": [ + 304, + 203, + 547, + 302 + ], + "type": "text", + "content": " (bottom)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": "ProxyCap [58], which use static cameras for keypoint projection, we employ dynamic cameras that incorporate both rotational and translational motion. This choice has two main motivations. First, it accounts for the inherent differences between human motions captured in static and dynamic camera setups. Second, it enables the learning of a camera-agnostic motion representation, from which the trajectory decoder can reconstruct the global trajectory. We also augment the 2D data with noise and masking. For details of the synthetic generation process see Sup. Mat." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 568 + ], + "type": "text", + "content": "Fine-tuning on Video Datasets. Starting with the pretrained network, we fine-tune WHAM on four video datasets: 3DPW [47], Human3.6M [10], MPI-INF-3DHP [32], and InstaVariety [14]. For the human mesh recovery task, we supervise WHAM on ground-truth SMPL parameters from AMASS and 3DPW, 3D keypoints from Human3.6M and MPI-INF-3DHP, and 2D keypoints from InstaVariety. For the global trajectory estimation task, we use AMASS, Human3.6M, and MPI-INF-3DHP. Additionally, during training we experiment with adding BEDLAM [1], a large synthetic dataset with realistic video and ground truth SMPL parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "The fine-tuning has two objectives: 1) exposing the network to real 2D keypoints, instead of training it solely on synthetic data, and 2) training the feature integrator network to aggregate motion and image features. To achieve these goals, we jointly train the entire network on the video datasets while setting a smaller learning rate on the pretrained modules (see Fig. 3). Consistent with prior work [5, 16, 29, 42, 50], we employ a pretrained and fixed-weight image encoder [20] to extract image features. However, to leverage recent network architectures and training strategies, we also experiment with different types of encoders [1, 6, 24] in the following section. Implementation Details. For the pretraining stage, we train" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2074" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "WHAM on AMASS for 80 epochs with the learning rate of " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". Then we finetune WHAM on 3DPW, MPI-INF-3DHP, Human3.6M, and InstaVariety for 30 epochs. We use a learning rate of " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " for the feature integrator and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " for the pretrained components during finetuning. During training, we use the Adam optimizer and a batch size of 64 and split sequences into 81-frame chunks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 165, + 128, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 128, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 128, + 178 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 289, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 289, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 289, + 292 + ], + "type": "text", + "content": "Datasets. We evaluate WHAM on three in-the-wild benchmarks: 3DPW [47], RICH [9], and EMDB [15]. Following previous work [1, 5, 13, 16, 20, 24, 29], we perform the evaluation in camera coordinates. The estimated global trajectory is evaluated on a subset of EMDB (EMDB 2) for which they provide ground truth global motion with dynamic cameras (used for evaluation). We also test on new sequences captured using an iPhone with the gyroscope. See Sup. Mat. for more details on the datasets and iPhone results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "content": "Evaluation metrics. To evaluate the accuracy of 3D human pose and shape estimation, we compute Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PA-MPJPE), and Per Vertex Error (PVE) measured in millimeters " + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "inline_equation", + "content": "(mm)" + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "content": ". We compute Acceleration error (Accel, in " + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "inline_equation", + "content": "m / s^2" + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "content": ") to measure the inter-frame smoothness of the reconstructed motion. We also evaluate the motion reconstruction and trajectory estimation accuracy in the world-frame. Following previous work [19, 53], we split sequences into smaller segments of 100 frames and align each output segment with the ground-truth data using the first two frames " + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{(W - MPJPE_{100})}" + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "content": " or the entire segment " + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{(WA - MPJPE_{100})}" + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "content": " in mm. These previous metrics give an unrealistic picture of 3D performance as they do not measure drift over long sequences. Therefore, we also evaluate the error over the entire trajectory after the rigid alignment and measure Root Translation Error (RTE in %) normalized by the actual displacement of the person. We also assess the jitter of the motion in the world coordinate system in " + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "inline_equation", + "content": "10m / s^3" + }, + { + "bbox": [ + 46, + 293, + 290, + 533 + ], + "type": "text", + "content": " and foot sliding during the contact (FS in mm)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 537, + 204, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 204, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 204, + 550 + ], + "type": "text", + "content": "4.1. 3D Human Motion Recovery" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 688 + ], + "type": "text", + "content": "Per-frame accuracy. Table 1 presents a comprehensive comparison of WHAM and the existing state-of-the-art per-frame and video-based methods across three benchmark datasets [9, 15, 47]. Because none of the methods are exposed to data from RICH or EMDB during training, results on these datasets are indicative of each method's ability to generalize. WHAM (Res), WHAM (HR), and WHAM (ViT) correspond to different architectures for the pretrained image encoders, derived from SPIN (ResNet-50) [20], CLIFF (HRNet-W48) [1, 24], and HMR2.0 (ViT-H/16) [6], respectively. Not surprisingly, WHAM (HR) is more accurate than WHAM (Res)," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "text", + "content": "Previous work follows [14] in reporting Accel in " + }, + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "inline_equation", + "content": "mm / \\text{frame}^2" + }, + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "text", + "content": ". To remove the dependency on frame rate, we convert all previous results to " + }, + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "inline_equation", + "content": "m / s^2" + }, + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 71, + 356, + 132 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 356, + 132 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 356, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 356, + 132 + ], + "type": "image", + "image_path": "e78d245f4a53bc4d530334e5d581d76085a986972024f3f22c9b8864f7e6c9a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 359, + 71, + 402, + 131 + ], + "blocks": [ + { + "bbox": [ + 359, + 71, + 402, + 131 + ], + "lines": [ + { + "bbox": [ + 359, + 71, + 402, + 131 + ], + "spans": [ + { + "bbox": [ + 359, + 71, + 402, + 131 + ], + "type": "image", + "image_path": "cddb34224c4ffaa183f9d634e571b3e2821a426e0e1ef72268b37d680accbf65.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 405, + 72, + 446, + 130 + ], + "blocks": [ + { + "bbox": [ + 405, + 72, + 446, + 130 + ], + "lines": [ + { + "bbox": [ + 405, + 72, + 446, + 130 + ], + "spans": [ + { + "bbox": [ + 405, + 72, + 446, + 130 + ], + "type": "image", + "image_path": "db656ab550ff811dbb5d14f2723b23f772b914999419f8b23c57a16056e2277a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 450, + 71, + 492, + 132 + ], + "blocks": [ + { + "bbox": [ + 450, + 71, + 492, + 132 + ], + "lines": [ + { + "bbox": [ + 450, + 71, + 492, + 132 + ], + "spans": [ + { + "bbox": [ + 450, + 71, + 492, + 132 + ], + "type": "image", + "image_path": "1895002a458b99197ef2a95b6f752f200fbeffd1ca827bbcda74092f79670b6e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 495, + 71, + 539, + 131 + ], + "blocks": [ + { + "bbox": [ + 495, + 71, + 539, + 131 + ], + "lines": [ + { + "bbox": [ + 495, + 71, + 539, + 131 + ], + "spans": [ + { + "bbox": [ + 495, + 71, + 539, + 131 + ], + "type": "image", + "image_path": "30be85eceac94604ab062647e1b1eca513fca27cc6bd38d7e3eda92a730824fb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 314, + 140, + 533, + 248 + ], + "blocks": [ + { + "bbox": [ + 314, + 140, + 533, + 248 + ], + "lines": [ + { + "bbox": [ + 314, + 140, + 533, + 248 + ], + "spans": [ + { + "bbox": [ + 314, + 140, + 533, + 248 + ], + "type": "image", + "image_path": "20477e4785edbd57c64195ccc0e3fefd676f995afcd2bb721c9f08a40ba31b8e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 251, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 273 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with previous state-of-the-art methods for 3D human pose and shape estimation. See text." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "content": "while the transformer-based version, WHAM (ViT), is the most accurate. The backbone matters, with WHAM (ViT) outperforming all previous methods on all per-frame metrics (MPJPE, PA-MPJPE, and PVE) on all benchmarks. Even with the simplest ResNet backbone, WHAM (Res) outperforms every method except for ReFit on RICH." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "type": "text", + "content": "Training on BEDLAM consistently improves accuracy in prior work (see [1, 3, 49]), and we find the same here as shown in Table 2. However, compared to ReFit, WHAM exhibits relatively smaller performance improvement by adding BEDLAM. One possible reason for this is that, unlike ReFit, we did not finetune the image encoder during training to limit training time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 446, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 578 + ], + "type": "text", + "content": "Inter-frame smoothness. We also evaluate the inter-frame smoothness using the acceleration error. Compared with state-of-the-art per-frame methods [1, 6, 22, 24], WHAM has significantly lower acceleration error. This indicates that WHAM reconstructs smooth and more plausible 3D human motion across frames while not sacrificing high per-frame accuracy. On the other hand, when compared to recent temporal methods [5, 42, 50], WHAM exhibits comparable or slightly higher acceleration error. However, we observe that these video-based methods tend to over-smooth the human motion, resulting in lower accuracy on per-frame metrics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 578, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 650 + ], + "type": "text", + "content": "To provide intuition for these numbers, we qualitatively compare WHAM with TCMR [5] and GLoT [42] in Fig. 4. While producing smooth results, TCMR and GLoT fail to capture the bending of the left knee when the subject is ascending the stairs, while WHAM more accurately reconstructs the 3D human pose." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 658, + 474, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 474, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 474, + 671 + ], + "type": "text", + "content": "4.2. 3D Global Trajectory Recovery" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "To evaluate global trajectory recovery, we compare WHAM with the state-of-the-art methods and a baseline that combines a SLAM method (DPVO [45]) and a per-frame" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2075" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 98, + 72, + 493, + 328 + ], + "blocks": [ + { + "bbox": [ + 98, + 72, + 493, + 328 + ], + "lines": [ + { + "bbox": [ + 98, + 72, + 493, + 328 + ], + "spans": [ + { + "bbox": [ + 98, + 72, + 493, + 328 + ], + "type": "table", + "html": "
Models3DPW (14)RICH (24)EMDB (24)
PA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccelPA-MPJPEMPIJPEPVEAccel
per-frameSPIN [20]59.296.9112.831.469.7122.9144.235.287.1140.7166.141.3
PARE* [17]46.574.588.6-60.7109.2123.5-72.2113.9133.2-
CLIFF* [24]43.069.081.222.556.6102.6115.022.468.3103.5123.724.5
HybrIK* [22]41.871.682.3-56.496.8110.4-65.6103.0122.2-
HMR2.0 [6]44.469.882.218.148.196.0110.918.860.798.3120.819.9
ReFit* [49]40.565.375.118.547.980.792.917.158.688.0104.520.7
temporalTCMR* [5]52.786.5101.46.065.6119.1137.75.079.8127.7150.25.3
VIBE* [16]51.982.998.418.568.4120.5140.221.881.6126.1149.926.5
MPS-Net* [50]52.184.399.06.567.1118.2136.75.881.4123.3143.96.2
GLoT* [42]50.680.796.46.065.6114.3132.75.279.1119.9140.85.4
GLAMR [55]51.1--8.079.9--107.773.8113.8134.933.0
TRACE* [43]50.979.195.428.6----71.5110.0129.625.5
SLAHMR [53]55.9---52.5--9.469.793.7111.37.1
PACE [19]----49.3--8.8----
WHAM (Res)*40.262.775.16.351.889.5103.25.057.884.099.75.2
WHAM (HR)*39.062.674.86.449.184.696.45.257.185.7103.25.6
WHAM (ViT)*35.957.868.76.644.380.091.25.350.479.794.45.3
", + "image_path": "fa320a60d3acd2892ff715e85d8ff8761203f6fbd184c8e772af8fcba63b56eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 399, + 266, + 495 + ], + "blocks": [ + { + "bbox": [ + 46, + 334, + 547, + 379 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 547, + 379 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of state-of-the-art models on the 3DPW [47], RICH [9], and EMDB [15] datasets. Ordering of per-frame and temporal methods is done separately by descending MPJPE on EMDB (except for PACE). For testing on EMDB, we follow the protocol of EMDB 1 [15]. Parenthesis denotes the number of body joints used to compute MPJPE and PA-MPJPE, and * denotes models trained with the 3DPW training set. Bold numbers denote the most accurate method in each column. Accel is in " + }, + { + "bbox": [ + 46, + 334, + 547, + 379 + ], + "type": "inline_equation", + "content": "m / s^2" + }, + { + "bbox": [ + 46, + 334, + 547, + 379 + ], + "type": "text", + "content": ", all other errors are in mm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 399, + 266, + 495 + ], + "lines": [ + { + "bbox": [ + 66, + 399, + 266, + 495 + ], + "spans": [ + { + "bbox": [ + 66, + 399, + 266, + 495 + ], + "type": "table", + "html": "
ModelsDataset3DPW (14)
PA-MPJPEMPJPEPVEAccel
CLIFF [1, 24]R43.668.882.119.2
R+B43.066.978.531.0
ReFit [49]R40.565.375.118.5
R+B38.257.667.621.4
WHAM (ViT)R35.957.868.76.6
R+B35.756.967.46.7
", + "image_path": "a6eb3b90b669912d09dba271b938e435aef15ca23b98623bfe7d60c7baaa878a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 536, + 286, + 630 + ], + "blocks": [ + { + "bbox": [ + 47, + 499, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 47, + 499, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 287, + 521 + ], + "type": "text", + "content": "Table 2. Dataset ablation experiments on 3DPW [47]. R denotes the use of real datasets and B denotes BEDLAM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 536, + 286, + 630 + ], + "lines": [ + { + "bbox": [ + 50, + 536, + 286, + 630 + ], + "spans": [ + { + "bbox": [ + 50, + 536, + 286, + 630 + ], + "type": "table", + "html": "
ModelsEMDB 2
WA-MPJPE100W-MPJPE100RTEJitterFS
DPVO (+ HMR2.0) [6, 45]647.82231.415.8537.3107.6
GLAMR [55]280.8726.611.446.320.7
TRACE [43]529.01702.317.72987.6370.7
SLAHMR [53]326.9776.110.231.314.5
WHAM (w/DPVO [45])135.6354.86.022.54.4
WHAM (w/DROID [44])133.3343.94.621.54.4
WHAM (w/ GT gyro)131.1335.34.121.04.4
", + "image_path": "30c7cb09d2354554a3d8761707ec2c379c8f33c01670b8ae8cdc76aefa6c9899.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 633, + 276, + 644 + ], + "lines": [ + { + "bbox": [ + 57, + 633, + 276, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 633, + 276, + 644 + ], + "type": "text", + "content": "Table 3. Global motion estimation accuracy on EMDB [15]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "method (HMR2.0 [6]); see Table 3. WHAM is agnostic to the source of the camera angular velocity and we compare results using DPVO, DROID-SLAM [44] and the ground truth angular velocity (gyro)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 399, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 579 + ], + "type": "text", + "content": "As shown in Table 3, WHAM outperforms the existing methods on all metrics. Specifically, combining WHAM with DPVO is more accurate than the global trajectory estimation of DPVO combined with HMR2.0, illustrating that our method actively refines the global trajectory instead of performing a simple integration. DROID-SLAM gives slightly better results than DPVO. Furthermore, WHAM significantly outperforms the regression-based method, TRACE, on jitter and foot sliding metrics. We further demonstrate this in Figs. 5 and 1, where WHAM captures more consistent and plausible human motion in the global coordinate system than TRACE and SLAHMR for videos captured by dynamic cameras. As depicted in Fig. 6, WHAM outperforms GLAMR, TRACE, and SLAHMR in capturing the pattern of human motion in the global coordinate system." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 586, + 400, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 400, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 400, + 601 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "To provide further insight into our approach, we conduct ablation studies to analyze the contribution of each component to the performance. As shown in Table 4, our entire system (WHAM) outperforms the different variants of WHAM that ablate a single component. To be specific, we first observe that adding feature integration improves both motion and global trajectory estimation accuracy when compared with an ablated version without feature integration " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{w / o}F_I)" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ". Similarly, the removal of the pretraining on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2076" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 282, + 312 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 282, + 312 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 282, + 312 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 282, + 312 + ], + "type": "image", + "image_path": "182970dcdeee0d1bf6d9a67512ac5f6d202d89b4dfbaafa7646a53c8ef95001e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 317, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 317, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 287, + 340 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with TRACE [43] and SLAHMR [53] on global human motion estimation with dynamic cameras." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 350, + 286, + 416 + ], + "blocks": [ + { + "bbox": [ + 48, + 350, + 286, + 416 + ], + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 416 + ], + "type": "image", + "image_path": "06647819792b926e3c1cea4ad54a3ec6ec7a526c29b5544816ef00e894718ba5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 419, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 453 + ], + "type": "text", + "content": "Figure 6. Comparison of global trajectory estimation on EMDB [15]. Overall, WHAM shows better alignment to ground truth data compared to GLAMR [55], TRACE [43], and SLAHMR [53]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 660 + ], + "type": "text", + "content": "the 2D-to-3D lifting task using AMASS [31] (w/o lifting) shows significant performance degradation. WHAM also outperforms the ablation of the Neural Initialization (w/o NI), particularly in the MPJPE metric. In addition, we experiment with WHAM to decode trajectory solely based on the motion context without using the estimated camera angular velocity " + }, + { + "bbox": [ + 46, + 468, + 289, + 660 + ], + "type": "inline_equation", + "content": "(\\mathrm{w / o}\\omega)" + }, + { + "bbox": [ + 46, + 468, + 289, + 660 + ], + "type": "text", + "content": ". Although this version shows similar performance in predicting 3D human pose, it suffers from the entanglement of camera and human motion, resulting in significantly high overall trajectory error (RTE). Last, we observe that WHAM without the trajectory refinement (w/o traj. ref.) gives larger global trajectory and foot sliding errors in return for less jitter, indicating that our approach contributes to the global trajectory accuracy and helps reduce foot sliding. Core details are presented here; see Sup. Mat. for more details and information on run-time cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": "WHAM is a new method to recover accurate 3D human motion in global coordinates from a moving camera more ef" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 72, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 148 + ], + "type": "table", + "html": "
ModelsEMDB 2
PA-MPJPEMPJPEWA-MPJPE100W-MPJPE100RTEJitterFS
w/o F144.269.0147.6377.96.323.15.5
w/o lifting60.383.0238.0693.011.524.55.0
w/o NI40.466.3142.7368.16.822.34.6
w/o ω39.162.0156.5422.010.122.15.0
w/o traj. ref.38.259.3154.7407.56.318.86.5
WHAM (Ours)38.259.3135.6354.86.022.54.4
", + "image_path": "5cd46d0d8745a81d652dde20e53d183ccd31b92aaba5218b0870942a84a4f109.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 352, + 151, + 499, + 161 + ], + "lines": [ + { + "bbox": [ + 352, + 151, + 499, + 161 + ], + "spans": [ + { + "bbox": [ + 352, + 151, + 499, + 161 + ], + "type": "text", + "content": "Table 4. Ablation experiments. See text." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 182, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 411 + ], + "type": "text", + "content": "ficiently and accurately than the state-of-the-art approaches. Our approach leverages the AMASS dataset to train a network to recursively lift 2D sequences of keypoints to sequences of 3D SMPL parameters. But keypoints alone lack valuable information about the body and its movement. Consequently, we integrate image context information over time and learn to combine it with the motion context to better estimate human body shape and pose. Additionally, our method takes an estimate of the camera angular velocity, which can either be computed from a SLAM method or from the camera's gyro when available. Finally, we combine all this information with an estimate of foot contact to recover the 3D human motion in global coordinates from a monocular video sequence. WHAM significantly outperforms the existing state-of-the-art methods (both image-based and video-based) on challenging in-the-wild benchmarks in both 3D HPS and the world-coordinate trajectory estimation accuracy. Because of its speed and accuracy, WHAM provides a foundation for in-the-wild motion capture applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 487 + ], + "type": "text", + "content": "Limitations and future directions: WHAM learns about human motion from AMASS, limiting generalization to motions that are out of distribution. While we employ random masking as part of our data synthesis process, our generating approach mainly assumes the scenario where the full body is within the field of view. See Sup. Mat. for more details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 488, + 547, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 547, + 620 + ], + "type": "text", + "content": "WHAM opens up many directions for future work. For example, while we use SLAM to estimate the camera's angular velocity, SLAM could also provide camera intrinsics and extrinsics as well as information about the 3D scene that could be used to enforce consistency between the scene and the human. While WHAM is an online method, designed for real-time applications, it could also initialize an optimization-based post-processing akin to bundle adjustment, which would optimize the camera, scene, and human motion together. Furthermore, a real-time and phone-based implementation of WHAM should be feasible." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 624, + 548, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 548, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 548, + 703 + ], + "type": "text", + "content": "Acknowledgements. Part of this work was done when the first author was an intern at the Max Planck Institute for Intelligence System. This work was partially supported through an NSF CAREER Award (CBET 2145473) and a Chan-Zuckerberg Essential Open Source Software for Science Award. CoI Disclosure. https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2077" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8726-8737, 2023. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "text", + "content": "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016, pages 561-578. Springer International Publishing, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 269 + ], + "type": "text", + "content": "[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. SMPLer-X: Scaling up expressive human pose and shape estimation. In Proc. Neural Information Processing Systems (NeurIPS), 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 271, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 315 + ], + "type": "text", + "content": "[4] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 370 + ], + "type": "text", + "content": "[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "text", + "content": "[6] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 418, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 288, + 462 + ], + "type": "text", + "content": "[7] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 497 + ], + "type": "text", + "content": "[8] Dorian F. Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. *Bodyslam++*: Fast and tightly-coupled visual-inertial camera and human motion tracking, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 564 + ], + "type": "text", + "content": "[9] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "text", + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 622, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 288, + 655 + ], + "type": "text", + "content": "[11] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "text", + "content": "[12] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In 3DV, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[13] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 95 + ], + "type": "text", + "content": "pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 98, + 547, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 98, + 547, + 130 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 547, + 130 + ], + "type": "text", + "content": "[14] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. arXiv preprint arXiv:1812.01601, 2018. 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 133, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 133, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 133, + 547, + 186 + ], + "type": "text", + "content": "[15] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 189, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 547, + 243 + ], + "type": "text", + "content": "[16] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5252-5262. IEEE, 2020. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 246, + 547, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 547, + 300 + ], + "type": "text", + "content": "[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 303, + 546, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 303, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 546, + 358 + ], + "type": "text", + "content": "[18] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11015-11025, Piscataway, NJ, 2021. IEEE. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 546, + 404 + ], + "type": "text", + "content": "[19] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 406, + 546, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 546, + 449 + ], + "type": "text", + "content": "[20] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 452, + 546, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 546, + 485 + ], + "type": "text", + "content": "[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 487, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 547, + 552 + ], + "type": "text", + "content": "[22] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 555, + 546, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 546, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 546, + 610 + ], + "type": "text", + "content": "[23] Jiefeng Li, Siyuan Bian, Qi Liu, Jiasheng Tang, Fan Wang, and Cewu Lu. NIKI: Neural inverse kinematics with invertible neural networks for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 612, + 546, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 546, + 655 + ], + "type": "text", + "content": "[24] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "text", + "content": "[25] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2078" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[26] Kevin Lin, Chung-Ching Lin, Lin Liang, Zicheng Liu, and Lijuan Wang. Mpt: Mesh pre-training with transformers for human pose and mesh reconstruction, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[27] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 185 + ], + "type": "text", + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "text", + "content": "[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 288, + 285 + ], + "type": "text", + "content": "[30] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 534-543, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 331 + ], + "type": "text", + "content": "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 288, + 387 + ], + "type": "text", + "content": "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3D Vision (3DV), 2017 Fifth International Conference on. IEEE, 2017. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "type": "text", + "content": "[33] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "text", + "content": "[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 521 + ], + "type": "text", + "content": "[35] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudo-gts of 3d human mesh estimation in the wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 566 + ], + "type": "text", + "content": "[36] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Computer Vision - ECCV 2020, pages 598-613. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 633 + ], + "type": "text", + "content": "[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[38] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459-468, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[40] Nitin Saini, Chun-Hao P. Huang, Michael J. Black, and Aamir Ahmad. Smartmocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, pages 1-8, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 175 + ], + "type": "text", + "content": "[41] Simon Schaefer, Dorian F. Henning, and Stefan Leutenegger. Glopro: Globally-consistent uncertainty-aware 3d human pose estimation & tracking in the wild, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 176, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 547, + 239 + ], + "type": "text", + "content": "[42] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 296 + ], + "type": "text", + "content": "[43] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 547, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 341 + ], + "type": "text", + "content": "[44] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 546, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 365 + ], + "type": "text", + "content": "[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 3, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 366, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 547, + 420 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 476 + ], + "type": "text", + "content": "[47] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 479, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 546, + 532 + ], + "type": "text", + "content": "[48] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In The IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 535, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 546, + 567 + ], + "type": "text", + "content": "[49] Yufu Wang and Kostas Daniilidis. Refit: Recurrent fitting network for 3d human recovery. In International Conference on Computer Vision, 2023. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 547, + 623 + ], + "type": "text", + "content": "[50] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 689 + ], + "type": "text", + "content": "[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: generative 3D human shape and articulated pose models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 712 + ], + "type": "text", + "content": "[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2079" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "estimation. In Advances in Neural Information Processing Systems, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[53] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "type": "text", + "content": "[54] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 261 + ], + "type": "text", + "content": "[55] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "text", + "content": "[56] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "type": "text", + "content": "[57] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 407 + ], + "type": "text", + "content": "[58] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Hongwei Yi, Shengping Zhang, and Yebin Liu. Real-time monocular full-body capture in world space via sequential proxy-to-motion learning. arXiv preprint arXiv:2307.01200, 2023. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 463 + ], + "type": "text", + "content": "[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2080" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_content_list.json b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d01617a64ec8fad49b8d054140f604f20e158288 --- /dev/null +++ b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_content_list.json @@ -0,0 +1,1563 @@ +[ + { + "type": "text", + "text": "WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 132, + 128, + 839, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changhoon Kim $^{1}$ Kyle Min $^{2}$ Maitreya Patel $^{1}$ Sheng Cheng $^{1}$ Yezhou Yang $^{1}$ $^{1}$ Arizona State University $^{2}$ Intel Labs", + "bbox": [ + 142, + 202, + 825, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kch,maitreya.patel,scheng53,yz.yang}@asu.edu kyle.min@intel.com", + "bbox": [ + 189, + 248, + 771, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of generative models, facilitating the creation of hyper-realistic images from textual descriptions, has concurrently escalated critical societal concerns such as misinformation. Although providing some mitigation, traditional fingerprinting mechanisms fall short in attributing responsibility for the malicious use of synthetic images. This paper introduces a novel approach to model fingerprinting that assigns responsibility for the generated images, thereby serving as a potential countermeasure to model misuse. Our method modifies generative models based on each user's unique digital fingerprint, imprinting a unique identifier onto the resultant content that can be traced back to the user. This approach, incorporating fine-tuning into Text-to-Image (T2I) tasks using the Stable Diffusion Model, demonstrates near-perfect attribution accuracy with a minimal impact on output quality. Through extensive evaluation, we show that our method outperforms baseline methods with an average improvement of $11\\%$ in handling image post-processes. Our method presents a promising and novel avenue for accountable model distribution and responsible use. Our code is available in https://github.com/kylemin/WOUAF.", + "bbox": [ + 75, + 330, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 683, + 209, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in generative models have propelled their proficiency, expanding their repertoire to include not just the generation of photorealistic images [4, 14] but also the synthesis of images from textual prompts [21, 24, 26, 28]. These significant strides have equipped individuals with the capacity to leverage these models to create hyperrealistic images that correspond seamlessly with given textual instructions.", + "bbox": [ + 75, + 709, + 468, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonetheless, the escalating prominence of generative models instigates pressing societal apprehensions. A case in point is Deepfake, intentionally crafted to disseminate", + "bbox": [ + 75, + 830, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dcc97cf11cc5d6d310a8a6556f801e8e70d24f1e1ca82fdc1f79fdbbe32c9294.jpg", + "image_caption": [ + "Figure 1. Illustration of user attribution based on our method. Please refer to the main text for detailed descriptions." + ], + "image_footnote": [], + "bbox": [ + 540, + 299, + 852, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "misinformation, fostering a climate of fake news and political disarray [2, 17, 23]. The gravity of these concerns necessitates calls for governmental intervention to regulate the indiscriminate application of generative models1.", + "bbox": [ + 496, + 579, + 890, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A feasible method to counteract malicious use involves assigning accountability for generated images. One approach to achieve this is by integrating independent fingerprinting modules that can embed user-specific information on top of image generation. The open-source Text-to-Image (T2I) project Stable Diffusion (SD) [26] currently employs this technique using discrete wavelet transform or RivaGAN [34]. However, in the open-source setting, bypassing the fingerprinting module is straightforward and can be achieved by commeting just a single line in the source code [5].", + "bbox": [ + 496, + 640, + 892, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Is it feasible to achieve user attribution without an independent fingerprinting module? In response, we propose a distributor-oriented methodology named WOUAF, standing for Weight mOdupulation for User Attribution and", + "bbox": [ + 496, + 806, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 886, + 351, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence. The White House", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8974", + "bbox": [ + 480, + 943, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fingerprinting. In practical terms, when a model inventor opens-sources their work to a model distributor such as Huggingface, the distributor could utilize our proposed method to create a generic version. Upon receiving a download request from an end-user, the distributor can adjust the model weights using our technique and deploy a fingerprinted version to the user, simultaneously registering the user's fingerprint into their database. In the event of a model's malicious exploitation, the distributor can decode the fingerprint from the misused image and cross-reference it with their database to identify the responsible user. Consequently, this provides the distributor with an actionable method to counteract malicious uses of the model (see Fig. 1 for a comprehensive framework of our methodology).", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our methodology, designed for T2I tasks, is integrated into the Stable Diffusion (SD) framework without necessitating any structural changes to the model. This design choice effectively prevents end-users from bypassing the fingerprinting process. Consistent with prior research [5, 16, 22, 32, 33], our primary goal is to maintain high attribution accuracy while ensuring minimal impact on output quality, as elaborated in Sec. 3. Our rigorous evaluations of this method concentrate initially on assessing both attribution accuracy and image quality. We have found that our approach attains nearly flawless attribution accuracy with only a slight influence on image quality. Moreover, we evaluate the robustness of our method in various scenarios involving post-processing manipulations that images might undergo. Our method outperforms baseline methods in these robustness tests, showing an average improvement of $11\\%$ in handling such manipulations (refer to Sec. 4 for further details).", + "bbox": [ + 75, + 303, + 472, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are four main contributions: (1) We introduce WOUAF, a distinctive distributor-centered fine-tuning methodology. This approach embeds fingerprints within the model in such a way that end-users cannot easily circumvent or remove them. (2) Our method successfully achieves high attribution accuracy, while maintaining the quality of the output images. (3) Our approach exhibits marked resilience against a diverse array of image post-processes, a vital attribute for practical applications. (4) We conduct thorough assessments to balance attribution accuracy with manipulations to intentionally remove fingerprints, including strategies like image compression via auto-encoders and obfuscation through model fine-tuning.", + "bbox": [ + 75, + 575, + 472, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 787, + 218, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss related works of model fingerprinting in generative models. More related works are available in the appendix.", + "bbox": [ + 75, + 814, + 468, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inventor-oriented Model Fingerprinting. Yu et al. [33] leveraged a pre-trained deep steganography model to em", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bed fingerprints into the training set for fingerprinted GANs. However, this approach suffers from limited scalability, as it necessitates training a GAN from scratch for each distinct fingerprint. To address this issue, Yu et al. [32] introduced a weight modulation method [12] that directly embeds a user's fingerprint into the generator's weights. Despite these advancements, current methods are predominantly tailored for GAN-based models and typically require training from scratch. This raises important questions regarding their suitability for diffusion-based models, which have a different structural makeup compared to GANs, and the feasibility of avoiding the requirement for training from scratch. The adoption of fine-tuning as a method for embedding fingerprints presents a promising solution. It facilitates the incorporation of fingerprints into pre-trained diffusion models, eliminating the necessity for comprehensive retraining from the ground up [5, 36]. This approach significantly streamlines the process, allowing model inventors to concentrate on core model development without the complexities of embedding fingerprints during training.", + "bbox": [ + 496, + 90, + 893, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distributor-oriented Model Fingerprinting. Kim et al. [16] proposed a technique for achieving user attribution by explicitly incorporating user-specific fingerprints into the generator's output. While this simplified attribution method allowed for the derivation of sufficient fingerprint conditions, it necessitates a trade-off between the quality of the generated output and attribution accuracy, which is further exacerbated when image post-processes are taken into account. To tackle this issue, an approach has been proposed that utilizes subtle semantic variations along latent dimensions as fingerprints, generated by perturbations of eigenvectors in the latent distribution [22]. This method demonstrates an improved balance between generation quality and attribution accuracy. However, its applicability is restricted to unconditional image generation, as eigenvectors are computed by sampling the learned latent representation. In the context of conditional image generation, estimating eigenvectors of latent representation becomes challenging due to the vast space of conditions, such as those found in text conditions.", + "bbox": [ + 496, + 405, + 893, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent Advances in Fingerprinting for Text-to-Image Diffusion Models. Recent studies [5, 31, 36] have scrutinized fingerprinting techniques in the Stable Diffusion model [26], uncovering vulnerabilities in existing methods [34] that facilitate easy circumvention [5] or robust post-hoc fingerprinting [31]. Fernandez et al. [5] achieved near-perfect attribution accuracy by fine-tuning user-specific models to align with steganography module [37], demonstrating a viable alternative to conventional post-hoc fingerprinting modules [34]. However, this approach scales linearly in computational demand with the number of users since it necessitates fixed-time fine-tuning", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8975", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13be24e73fdd824db63437548662696957246b1c8bd2a9a3852cabdcb8418f54.jpg", + "image_caption": [ + "(a) The overall pipeline." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 504, + 256 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1118a0670691935fe7c445f3c5a6ab745c029dedda7eec74a9a5cb962ef28fb2.jpg", + "image_caption": [ + "(b) Weight modulation.", + "Figure 2. Depiction of our method's pipeline and weight modulation: (a) The model fingerprinting procedure encompasses encoding via the mapping network and weight modulation, along with decoding through the fingerprint decoding network. (b) Weight modulation of the decoding network $\\mathcal{D}$ to incorporate the fingerprint." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 883, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for each individual. In contrast, our method requires only a one-time training followed by a negligible forward pass time to generate user-specific models. Furthermore, our approach shows superior robustness against common image post-processing techniques compared to that of Stable Signature [5] (refer to Section Sec. 4.6 for details).", + "bbox": [ + 75, + 345, + 467, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another notable contribution is by Wen et al. [31], who introduced an alternative fingerprinting method for the Stable Diffusion model [26]. Their method, similar to posthoc style fingerprinting [34], depends on user-driven embedding, which allows end-users the option to exclude the fingerprint. Moreover, it is confined to the DDIM scheduler [30]. Our method, in contrast, is adaptable to both the DDIM [30] and Euler schedulers [15], underscoring its versatility and wider applicability (refer to the Appendix).", + "bbox": [ + 75, + 436, + 467, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 585, + 174, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section outlines our approach, beginning with an overview of the Text-to-Image (T2I) diffusion model with a focus on the Stable Diffusion (SD) model [26] detailed in Sec. 3.1. We then introduce our key component, the user-specific weight modulation, in Sec. 3.2. The section concludes with a detailed explanation of our training objectives and methods, outlined in Sec. 3.3.", + "bbox": [ + 75, + 609, + 467, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 726, + 217, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach utilizes the Stable Diffusion (SD) model, which functions within the latent space framework of an autoencoder. SD comprises two main elements: Firstly, an autoencoder is pre-trained on an extensive dataset of images. Its encoder, $\\mathcal{E}(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_z}$ , converts an image $x\\sim p_{data}$ into a latent representation $z = \\mathcal{E}(x)$ . The decoder, $\\mathcal{D}(\\cdot):\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_x}$ , then reconstructs the original image from this latent representation, resulting in $\\hat{x} = \\mathcal{D}(z)$ . The secondary element is a diffusion model, based on the U-Net architecture [27], represented as $\\epsilon_{\\theta}$ . This model is", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adept at generating latent representations and can be conditioned using pre-trained text embeddings", + "bbox": [ + 498, + 345, + 890, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. User-specific Weight Modulation", + "text_level": 1, + "bbox": [ + 498, + 383, + 785, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is fundamentally based on integrating fingerprints into the parameters of the SD through weight modulation [12, 32].", + "bbox": [ + 496, + 407, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of our method is illustrated in Fig. 2(a). A user-specific fingerprint is drawn from a Bernoulli distribution with a probability of 0.5, represented as $\\phi \\in \\Phi \\coloneqq \\operatorname{Bernoulli}(0.5)^{d_{\\phi}}$ , where $d_{\\phi}$ signifies the fingerprint length in bits. We employ a mapping network $\\mathcal{M}(\\cdot): \\mathbb{R}^{d_{\\phi}} \\to \\mathbb{R}^{d_M}$ to convert the sampled fingerprint $\\phi$ into an intermediate fingerprint representation within the $d_M$ dimension. For modulating each layer in the SD component, we introduce an affine transformation layer, $\\mathcal{A}_l(\\cdot): \\mathbb{R}^{d_M} \\to \\mathbb{R}^{d_j}$ , for all layers $l$ . As depicted in Fig. 2(b), this transformation matches the dimensions between $d_M$ and the $j$ -th channel in weight $W \\in \\mathbb{R}^{i,j,k}$ , where $i,j,k$ denote input, output, and kernel dimensions, respectively. The weight modulation for the $l$ -th layer is defined as:", + "bbox": [ + 496, + 452, + 892, + 664 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i, j, k} ^ {\\phi} = u _ {j} * W _ {i, j, k}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 672, + 890, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W$ and $W^{\\phi}$ denote the pre-trained and fingerprinted weights respectively, $u_{j} = \\mathcal{A}_{l}(\\mathcal{M}(\\phi))$ is the scale of the fingerprint representation corresponding to the $j$ th output channel.", + "bbox": [ + 496, + 702, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We incorporate fingerprints into the SD by applying weight modulation exclusively to the weights in the decoder $\\mathcal{D}$ . The rationale for not applying modulation to both the diffusion model $\\epsilon_{\\theta}$ and decoder $\\mathcal{D}$ , an approach that mirrors GAN-based models [32], is explained in Sec. 4.5.", + "bbox": [ + 496, + 763, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Training Objectives", + "text_level": 1, + "bbox": [ + 500, + 847, + 687, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our training architecture comprises two primary objectives. The initial objective is to decode fingerprints from the pro", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8976", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/b3177fc4b8cf17757cd731de50a44e76577cf8a0816dd4889521d6491d18a3d5.jpg", + "table_caption": [ + "Table 1. Evaluation of Attribution Accuracy and Image Generation Quality. We conducted validation using the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which were excluded from our training phase. Symbols $\\uparrow$ and $\\downarrow$ denote preferred higher and lower values, respectively." + ], + "table_footnote": [], + "table_body": "
ModelFingerprinting Time (↓)MS-COCOLAION
Attribution Acc (↑)CLIP-score (↑)FID (↓)Attribution Acc (↑)CLIP-score (↑)FID (↓)
Original SD [26]--0.7324.48-0.5019.67
DAG [16]8.4 hr0.700.7326.540.710.4923.13
Stable Signature [5]< 1 min0.990.7324.550.980.5020.02
WOUAF-conv< 1 sec0.990.7324.430.980.5120.46
WOUAF-all< 1 sec0.990.7324.420.990.5119.91
", + "bbox": [ + 76, + 141, + 893, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vided images. We train a fingerprint decoding network $\\mathcal{F}(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_\\phi}$ , which is instantiated by ResNet-50 [6], as follows:", + "bbox": [ + 75, + 299, + 470, + 344 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\phi} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\sum_ {i = 1} ^ {d _ {\\phi}} [ \\phi_ {i} \\log \\sigma (\\mathcal {F} (\\mathcal {D} (\\phi , z)) _ {i} \\\\ \\left. + \\left(1 - \\phi_ {i}\\right) \\log \\left(1 - \\sigma \\left(\\mathcal {F} \\left(\\mathcal {D} (\\phi , z)\\right)\\right) _ {i} \\right], \\right. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 357, + 468, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ refers to the sigmoid activation function, constraining the output of $\\mathcal{F}$ to the range [0, 1]. Thus, this loss function effectively combines binary cross-entropy for all bits of the fingerprint. During training time, fingerprint $\\phi$ is sampled from Bernoulli distribution. However, after training, the model distributor initially samples a user-specific fingerprint $\\phi_{\\alpha}$ and subsequently modulates the decoder $\\mathcal{D}$ using $\\phi_{\\alpha}$ . The user will receive the fingerprinted decoder $\\mathcal{D}(\\phi_{\\alpha}, \\cdot)$ , which solely permits latent input.", + "bbox": [ + 75, + 434, + 468, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The secondary objective endeavors to regularize the quality of outputs. Ideally, this regularization inhibits the decoder $\\mathcal{D}$ from compromising image quality while minimizing $L_{\\phi}$ in Eq. (2):", + "bbox": [ + 75, + 571, + 470, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {q u a l i t y}} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\left[ \\ell (x, \\mathcal {D} (\\phi , z)) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 647, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\ell$ represents the distance metric between original images and fingerprinted images. For practical applications, we utilize perceptual distance [35] to gauge the perceptual difference between $x$ and $\\mathcal{D}(\\phi, z)$ .", + "bbox": [ + 75, + 679, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final objective function can be formulated as:", + "bbox": [ + 96, + 741, + 424, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {A}, \\mathcal {M}, \\mathcal {D}, \\mathcal {F}} \\lambda_ {1} L _ {\\phi} + \\lambda_ {2} L _ {\\text {q u a l i t y}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 772, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where both $\\lambda_{1}$ and $\\lambda_{2}$ are set to 1.0. Fundamentally, the loss function aspires to reconstruct fingerprints while maintaining the quality of the generated outputs. To assess the efficacy of our proposed method, we employ attribution accuracy and image quality metrics (Refer to Sec. 4.1 for details).", + "bbox": [ + 75, + 809, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 297, + 633, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 498, + 323, + 692, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. Our approach is fine-tuned on the MS-COCO [19] dataset, adopting the Karpathy split. For methodological evaluation, we harness the test set from MS-COCO and randomly sample from the LAION-aesthetics [29] dataset. For T2I image generation, we adopt the Euler scheduler [15] with timestep $T = 20$ , and the classifier-free guidance scale [9] is set to 7.5 unless otherwise specified. Evaluation for DDIM scheduler [30] and various image generation hyperparameters are available in the Appendix.", + "bbox": [ + 496, + 345, + 890, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental Setting. We implement the weight modulation following the design specified in the source code of StyleGAN2-ADA [13]. Our mapping network $M$ is designed with a series of fully connected layers, wherein all experiments are conducted using a two-layer configuration. To train robust models against image post-processing transformations, differentiable post-processes are necessary. To this end, we incorporate the Kornia library [25]. For Stable Signature [5], we utilize the official code provided by the authors. We note that its post-processing transformations are replaced with our version for fair comparison. Appendix includes details on mapping network dimensions, training parameters, and optimizer.", + "bbox": [ + 496, + 505, + 892, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Evaluations.** User attribution accuracy is gauged by the formula: $\\frac{1}{d_{\\phi}} \\sum_{i=1}^{d_{\\phi}} \\mathbb{1}(\\phi_i = \\hat{\\phi}_i)$ , where $\\phi$ is the true fingerprint and $\\hat{\\phi} = \\mathbb{1}\\left[\\sigma(\\mathcal{F}(x_{\\phi})) > 0.5\\right]$ is the estimated fingerprint from image $x_{\\phi}$ . Unless otherwise stated, $d_{\\phi}$ is set to 32 in our experiments (Refer to Sec. 4.2 for additional information). We further employ a statistical test [32, 33] to evaluate matching bits between $\\hat{\\phi}$ and $\\phi$ . The null hypothesis $H_0$ suggests that the number of matching bits arises by chance. The test uses a binomial distribution, with a $p$ -value derived as: $P(X \\geq k|H_0) = \\sum_{i=k}^{d_{\\phi}} \\binom{d_{\\phi}}{i} 0.5^{d_{\\phi}}$ . A $p$ -value below 0.05 leads to the rejection of $H_0$ , with $1 - p$ serving as an indicator of verification confidence. Lastly, to", + "bbox": [ + 496, + 709, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8977", + "bbox": [ + 480, + 943, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81b5af490550c7849846a48c1d9a2b3989c323c3e67c7f8c1823de3b1950f898.jpg", + "image_caption": [ + "(a) MS COCO" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 480, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/414dc7ecef5c313ac54f68e4af6d44b95bec8eec265e1b21765a60fcacfcb7bb.jpg", + "image_caption": [ + "(b) LAION Aesthetics", + "Figure 3. Qualitative comparison of the original and fingerprinted Stable Diffusion models on MS-COCO [19] and LAION aesthetics [29] (Pixel-wise differences $\\times$ 5: they are multiplied by a factor of 5 for better view). We can observe that our method maintains high image quality." + ], + "image_footnote": [], + "bbox": [ + 491, + 88, + 890, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "validate the quality of our method, we assess image quality using the Fréchet Inception Distance (FID) [8] and employ the Clip-score [7] to determine the alignment between text and generated images. Additional experimental details can be found in the Appendix.", + "bbox": [ + 75, + 416, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. For evaluating our methodology, we benchmark against two established baseline methods: DAG [16] and Stable Signature [5]. Both these methods, conceptualized from the model distributor's standpoint, incorporate finetuning for model fingerprinting. To ensure a fair comparison, we retrain the Stable Signature method within our training settings by replacing its post-processing scheme. Additionally, we evaluate our method against three distinct variants based on the specific layers chosen for weight modulation implementation. The first variant, WOUAF-conv, applies modulation to only the convolutional layers in $\\mathcal{D}$ . In contrast, WOUAF-all extends this approach across all layers of $\\mathcal{D}$ , covering both self-attention and convolution layers. The final variant implements weight modulation in both the diffusion model $\\epsilon_{\\theta}$ and the decoder $\\mathcal{D}$ , mirroring the approach used in GAN-based methods [32]. Further details on why this variant is not used in our experiments are discussed in Sec. 4.5.", + "bbox": [ + 75, + 501, + 470, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Fingerprint Capacity", + "text_level": 1, + "bbox": [ + 76, + 785, + 276, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The capacity of our method depends on the maximum number of unique user-specific fingerprints it can support without significant crosstalk. This capacity is primarily influenced by the fingerprint dimension $(d_{\\phi})$ . Selecting an optimal $d_{\\phi}$ presents a challenge: while a larger $d_{\\phi}$ can accommodate more users, it also complicates effective fingerprint", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f6ba285f34e2143b648010dc29f8e6f7883e6a707254c867f0851e2a18346c2d.jpg", + "table_caption": [ + "Table 2. Experiments of attribution accuracy across various fingerprint dimensions $(d_{\\phi})$" + ], + "table_footnote": [], + "table_body": "
Fingerprint Dims.Attribution Accuracy
163264128
WOUAF-conv0.990.990.980.94
WOUAF-all0.990.990.990.97
", + "bbox": [ + 524, + 453, + 859, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "decoding [18].", + "bbox": [ + 500, + 556, + 599, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To investigate this trade-off, we conduct an analysis with varying fingerprint dimensions, specifically $d_{\\phi}$ values of 16, 32, 64, and 128. Tab. 2 presents the user attribution accuracy for each $d_{\\phi}$ value. As shown in Tab. 2, attribution accuracy tends to decrease monotonically as $d_{\\phi}$ increases. Importantly, both our variant models achieve a near-perfect attribution accuracy of 0.99 for $d_{\\phi}$ values of 16, 32, and 64. However, for $d_{\\phi} = 128$ , WOUAF-all variant outperforms the WOUAF-conv variant. For a balanced comparison with existing methods, we choose $d_{\\phi} = 32$ , which notably can support a substantial user base exceeding 4 billion $\\approx 2^{32}$ .", + "bbox": [ + 496, + 573, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Attribution Accuracy and Image Quality", + "text_level": 1, + "bbox": [ + 500, + 753, + 849, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a comprehensive evaluation of WOUAF, focusing on attribution accuracy and image quality. The assessment involves the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which are excluded from the training phase. The results, detailed in Tab. 1, showcase the efficacy of our method.", + "bbox": [ + 496, + 777, + 890, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our variants, namely WOUAF-conv and WOUAF-all, demonstrate superior performance in attribution accuracy", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8978", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "over DAG [16], indicating their proficiency in accurately decoding embedded fingerprints from the generated images. These variants also show competitive results when compared to Stable Signature [5], reinforcing our methodology's robustness. Notably, we achieve this high level of accuracy without significantly compromising image quality. Both FID scores and Clip-scores showed minimal variation from the baseline SD model, indicating that our approach has a negligible impact on image output quality. This is further corroborated by qualitative examples in Fig. 3, which highlight WOUAF's ability to reliably incorporate fingerprinting without degrading image generation quality. For additional insights, uncurated image collections are provided in the Appendix.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the growing importance of T2I models, computation time for fingerprinting emerges as a key metric. Our method stands out in computational efficiency. It contrasts with approaches like Stable Signature that need fine-tuning for each new fingerprint. Our method requires just a single forward pass, markedly reducing computational overhead.", + "bbox": [ + 75, + 301, + 468, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Attribution Analysis for Diverse Image Sources", + "text_level": 1, + "bbox": [ + 76, + 402, + 468, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Investigating the attribution of generated images to responsible users, we explore the potential for images from non-fingerprinted or varied sources to bypass our system. Our analysis aims to determine if decoded fingerprints from such images match any entries in the model distributor's database. A mismatch indicates the image's external origin, absolving users in the database.", + "bbox": [ + 75, + 426, + 468, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the experimental setup from [32], compiling a dataset with different image types: authentic images from the MS-COCO test set [19], non-fingerprinted images from Stable Diffusion [26], and synthesized images from ProGAN [10], StyleGAN [11], and StyleGAN2 [12], with each category containing 1,000 samples. Given our extensive user database of 1 million entries, we set a threshold at $32 * 0.95 \\approx 30$ bits, aligning with our 0.99 attribution accuracy as shown in Tab. 1.", + "bbox": [ + 75, + 534, + 470, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our rigorous experiments revealed that, irrespective of the source, no images were incorrectly attributed as possessing a fingerprint from our 1 million fingerprint database. This reinforces the reliability of our attribution approach as detailed in Sec. 4.3 demonstrating the robustness of our system against diverse image sources.", + "bbox": [ + 75, + 670, + 470, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Benefits of Finetuning only Decoder", + "text_level": 1, + "bbox": [ + 76, + 770, + 387, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When developing our last variant that incorporates weight modulation into both the diffusion model $\\epsilon_{\\theta}$ and the decoder $\\mathcal{D}$ , we note that the resultant pipeline demonstrates similarities with the GAN-based method [32]. A direct comparison between ours and the GAN-based methods may not be entirely straightforward, given the fundamental differences in their training methodologies. This is because the", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c1ba3c77858c1dcf5632e7f38a2ef68307c48119e24f1286bc3936d6d630958.jpg", + "image_caption": [ + "Figure 4. Comparative analysis of weight modulation on decoder $\\mathcal{D}$ and diffusion model $\\epsilon_{\\theta}$ with decoder $\\mathcal{D}$ . Modulating the diffusion model negatively affects image quality." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 890, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GAN-based methods entail training from scratch, whereas our proposed approach leans towards fine-tuning. Nevertheless, both methodologies share a common mechanism: they aim to modulate the weights of the layers instrumental in learning the latent space. The shared characteristic underscores the fundamental objective of optimizing the balance between attribution accuracy and generation quality.", + "bbox": [ + 496, + 375, + 890, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, our empirical observations suggest that this variant does not consistently achieve commendable performance as an attribution model. Specifically, it appears that this variant can only optimize either attribution accuracy or generation quality, but not both simultaneously. In our tests, the highest attribution accuracy reached by this variant is $89\\%$ , with a Clip-score of 0.68 and FID of 63.48 (detailed in Fig. 4). The inherent trade-off observed here further reinforces the challenge of balancing these two critical parameters in the context of model fingerprinting techniques.", + "bbox": [ + 496, + 479, + 892, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Robust User Attribution against Image Post-processes", + "text_level": 1, + "bbox": [ + 498, + 638, + 890, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section evaluates the robustness of our method in scenarios where generated images undergo post-processing. These processes could potentially alter the embedded fingerprint within the images.", + "bbox": [ + 496, + 676, + 890, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistent with methodologies outlined in previous research [5, 16, 22, 32, 33], we examine our model's resilience to various image post-processing operations. We simulate the effect of post-processing at random intensities before inputting data into the fingerprint decoding network, $\\mathcal{F}$ :", + "bbox": [ + 496, + 738, + 890, + 827 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {r o b u s t}} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\sum_ {i = 1} ^ {d _ {\\phi}} [ \\phi_ {i} \\log \\sigma (\\mathcal {F} (T (\\mathcal {D} (\\phi , z))) _ {i} \\\\ \\left. + \\left(1 - \\phi_ {i}\\right) \\log \\left(1 - \\sigma \\left(\\mathcal {F} \\left(T (\\mathcal {D} (\\phi , z))\\right)\\right) _ {i} \\right], \\quad (5) \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 835, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8979", + "bbox": [ + 480, + 943, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7848319d6cc20b05cc460cd7b7c0433e3e1d16411d4a5aba0f755eb22835a257.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 90, + 299, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8095f54cd666872ea3960f117a907aa8a82d240a2bc5d013d5d4913ddb3f0f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 90, + 495, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38edb2069f0a835330a80ec59409407d8d05890dd5d89985087e71f8f44374cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 687, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e74b07f35158e93c83bb77396777bb8a17a8ae1f43824742fc0eac6fc636517e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 90, + 883, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/718c6d942aaac1bfa24d360faff29276214095031d49ea1856a331c8795ff322.jpg", + "image_caption": [ + "Figure 5. Enhanced Robustness Against Image Post-Processes. For almost all scenarios, WOUAF consistently exceeds the performance of DAG [16] and Stable Signature [5]." + ], + "image_footnote": [], + "bbox": [ + 83, + 237, + 299, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53192df47f6f52d5757238393de518074e78bf82bf537204d8a29119d839ba56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 237, + 495, + 372 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/550be5844220d3111342a5bf94c8bb462e2062122bdcb71e750a28ab4bcdb556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 237, + 687, + 372 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77d36631f3c4cd4f45fdbfeb210c65e17ca059b162c682f88885bcc6971db789.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 237, + 883, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $T(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_x}$ denotes the post-processing function. In the optimization process, we employ an objective function akin to the one detailed in Eq. (4), with $L_{\\phi}$ substituted by $L_{\\mathrm{robust}}$", + "bbox": [ + 75, + 428, + 467, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our exploration, we contemplate eight different post-processing techniques: Erasing, Rotation, Gaussian Blurring, Cropping, Brightness jittering, the addition of Gaussian Noise, JPEG compression, and a Combination of all these post-processes. The parameters for these post-processes are designed as follows: For random erasing, we use a random erase ratio within the range $[5\\%, 10\\%, 15\\%, 20\\%]$ . Rotation involves randomly sampling a degree within the range (-30, 30). For Gaussian Blurring, we randomly select a kernel size from [3, 5, 7]. For Cropping, we use a random cropping-out ratio within the range $[5\\%, 10\\%, 15\\%, 20\\%]$ . The Brightness factor is randomly sampled within the range (-0.3, 0.3). For Gaussian Noise, we add noise with a standard deviation randomly sampled from a uniform distribution $U[0, 0.2]$ . JPEG compression quality level is selected from [90, 80, 70, 60, 50]. The Combination technique randomly selects a subset of these seven post-processing methods with a probability of 0.5.", + "bbox": [ + 75, + 489, + 467, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User attribution accuracy for each post-process is evaluated under these parameters. Our tests, depicted in Fig.5, offer a comparative analysis of user attribution accuracy across robust versions of DAG [16], Stable Signature [5], and WOUAF. Remarkably, our method demonstrates robustness across a range of post-processes, achieving an attribution accuracy improvement of $11\\%$ over Stable Signature and $29\\%$ over DAG. A notable trend across all transformations is the monotonic decrease in user attribution accu", + "bbox": [ + 75, + 763, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "racy as the intensity of post-processing increases. This reinforces the challenges posed by post-processing in maintaining accurate user attribution. However, our results also underscore the benefits of robust training in overcoming these challenges, emphasizing the importance of resilient training strategies for fingerprinting methods in the face of post-processing transformations. Considering the robustness of our method against various post-processes, it becomes a viable choice for model distributors seeking reliable fingerprinting solutions. Detailed results of FID scores and visual examples are available in the Appendix.", + "bbox": [ + 496, + 428, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Deliberate Fingerprint Manipulations", + "text_level": 1, + "bbox": [ + 498, + 609, + 841, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section delves into our method's robustness against deliberate attempts to remove fingerprints, which include malicious manipulations via auto-encoders and model purification. Further details and extended attack scenarios are provided in the Appendix.", + "bbox": [ + 496, + 636, + 890, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Resilience Against Deep Classifier", + "text_level": 1, + "bbox": [ + 498, + 724, + 795, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The imperceptibility of the fingerprint in generated images is crucial to prevent its detection and subsequent tampering by malicious entities. To assess the secrecy of our method, we adopt an attack scenario akin to the one in [32], assuming an attacker aims to train a classifier to detect the presence of a fingerprint.", + "bbox": [ + 496, + 748, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assume that the attacker seeks to train a classifier capable of detecting the presence of a fingerprint. To assess this scenario, we utilize a pretrained ResNet-50 [6] based binary classifier, trained using 10K SD generated images", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8980", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(5K original SD images and 5K fingerprinted SD images). This configuration is deemed valid as detecting the presence of a fingerprint necessitates using both non-fingerprinted and fingerprinted images in the training set. The binary classifier achieve $98\\%$ accuracy in the training stage. In subsequent evaluations using a separate set of 5K images from our variant models, the binary classification accuracy is 0.66 for WOUAF-conv and just 0.56 for WOUAF-all, which is nearly equivalent to random chance.", + "bbox": [ + 75, + 90, + 467, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These findings imply that detecting our embedded fingerprint, particularly in the WOUAF-all variant, poses a challenge to detect. Upcoming subsections will delve into further evaluations, predicated on the stringent assumption that users are cognizant of the fingerprint's presence and endeavor to eliminate it by employing auto-encoder methods or fine-tuning techniques.", + "bbox": [ + 75, + 226, + 467, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Resilience Against Auto-Encoders", + "text_level": 1, + "bbox": [ + 76, + 343, + 372, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contexts where adversaries aim to alter output images, leveraging deep learning techniques such as neural auto-encoders [1, 3, 20] becomes a common strategy for the purpose of obfuscating or removing fingerprints embedded in images [5]. To assess the resilience of our approach, we utilize the robust model against JPEG described in Sec. 4.6. This comparison is appropriate as JPEG represents a conventional image compression method. However, the auto-encoders [1, 3, 20] employed in our evaluation exhibit superior compression performance compared to JPEG. Our research explores the resilience of our proposed method against these sophisticated auto-encoders, focusing particularly on the impact of their varying compression rates.", + "bbox": [ + 75, + 367, + 467, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As depicted on the left side of Fig. 6, our investigations reveal a notable trend: attribution accuracy progressively declines towards a near-random level (approximately $50\\%$ ) as the compression rate employed by the auto-encoders escalates. This trend highlights a critical trade-off: the reduction in attribution accuracy is achievable solely by compromising the quality of the image [18]. Our findings indicate that compromising the integrity of the image is a necessary consequence to effectively obscure the fingerprinting process.", + "bbox": [ + 75, + 564, + 467, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Resilience Against Model Purification", + "text_level": 1, + "bbox": [ + 76, + 724, + 401, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This subsection addresses the scenario where an adversary, upon recognizing the presence of fingerprints within the images generated by the image decoder $\\mathcal{D}$ , opts to fine-tune $\\mathcal{D}$ with the objective of obliterating the embedded fingerprint. This strategy, known as model purification, is a sophisticated approach to altering the model's output to erase traceable imprints [5].", + "bbox": [ + 75, + 750, + 467, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this adversarial setting, the primary aim is to refine the downloaded fingerprinted model by optimizing the reconstruction error between the adversary's proprietary im", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ccbb89d3c93e09bc396edd8c5054b2550c96dbae9f717ab2e0da0d5bd0d83e0.jpg", + "image_caption": [ + "Figure 6. Left: Auto-Encoder-based Fingerprint Removal. With heightened compression rates, both image quality and attribution accuracy experience a decrease. Right: Model Purification. Progressive fine-tuning leads to concurrent declines in both image quality and attribution accuracy. Note that a lower FID score is preferable, indicating better image quality." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 702, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2230b67358d1b95f9e022c1ba4474a5a5a10b3b03576790ada6ecc763fa446eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 890, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age dataset and the output from the fingerprinted model. By adhering to the experimental framework outlined in [5], we charted the interplay between FID scores and attribution accuracy, as presented on the right side of Fig. 6. Our empirical analysis reveals a significant challenge: efforts to decrease the attribution accuracy lead to a decline in the quality of the generated images. This finding underscores the inherent complexity in fine-tuning processes aimed at model purification, particularly when striving to maintain the visual quality of the output while endeavoring to obscure its traceable characteristics.", + "bbox": [ + 496, + 319, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 505, + 617, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we have delved into user attribution for Stable Diffusion-based Text-to-Image (T2I) model, employing a weight modulation-based fingerprinting approach. Our method, WOUAF, not only achieves near-perfect accuracy but also preserves the high quality of generated images. A key aspect of WOUAF is its computational efficiency coupled with enhanced robustness against various image post-processing techniques compared to existing baselines. Our results lay a solid groundwork for future exploration into the broader implications and challenges posed by generative models. In future work, we plan to expand and refine our methodology to encompass various data types including text, audio, and video, necessitating tailored adjustments in model fingerprinting techniques.", + "bbox": [ + 496, + 532, + 890, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 762, + 668, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by the National Science Foundation under Grant No. 2038666, No. 2101052, and a grant from Meta AI Learning Alliance. The HuggingFace demo of this work is funded by Intel Corporation. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers.", + "bbox": [ + 496, + 787, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8981", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 8", + "[2] Ali Breland. The bizarre and terrifying case of the \"deep-fake\" video that helped bring an african nation to the brink. motherjones, 2019. 1", + "[3] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8", + "[4] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1", + "[5] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7", + "[7] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5", + "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5", + "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[10] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6", + "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 6", + "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. arXiv preprint arXiv:1912.04958, 2019. 2, 3, 6", + "[13] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4", + "[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1", + "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 3, 4", + "[16] Changhoon Kim, Yi Ren, and Yezhou Yang. Decentralized attribution of generative models. In International Conference on Learning Representations, 2021. 2, 4, 5, 6, 7", + "[17] Arijeta Lajka. New ai voice-cloning tools 'add fuel' to misinformation fire. AP News, 2023. 1", + "[18] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. ArXiv, abs/2103.09274, 2021. 5, 8", + "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6", + "[20] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, 3-8 December 2018, Montréal, Canada, pages 10794–10803, 2018. 8", + "[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1", + "[22] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2, 6", + "[23] Matt Novak. Ai image creator midjourney halts free trials but it has nothing to do with the pope's jacket. forbes, 2023. 1", + "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[25] E. Riba, D. Mishkin, D. Ponsa, E. Rublee, and G. Bradski. Kornia: an open source differentiable computer vision library for pytorch. In Winter Conference on Applications of Computer Vision, 2020. 4", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4, 6", + "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8982", + "bbox": [ + 482, + 944, + 514, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 4, 5", + "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4", + "[31] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3", + "[32] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2, 3, 4, 5, 6, 7", + "[33] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2, 4, 6", + "[34] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. 2019. 1, 2, 3", + "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4", + "[36] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. ArXiv, abs/2303.10137, 2023. 2", + "[37] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 657–672, 2018. 2" + ], + "bbox": [ + 78, + 90, + 470, + 698 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8983", + "bbox": [ + 482, + 944, + 514, + 954 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_model.json b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91894ac8d018cd131c71ec971d78ddc5142592ca --- /dev/null +++ b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.13, + 0.84, + 0.175 + ], + "angle": 0, + "content": "WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.826, + 0.247 + ], + "angle": 0, + "content": "Changhoon Kim\\(^{1}\\) Kyle Min\\(^{2}\\) Maitreya Patel\\(^{1}\\) Sheng Cheng\\(^{1}\\) Yezhou Yang\\(^{1}\\) \n\\(^{1}\\)Arizona State University \\(^{2}\\)Intel Labs" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.249, + 0.772, + 0.264 + ], + "angle": 0, + "content": "{kch,maitreya.patel,scheng53,yz.yang}@asu.edu kyle.min@intel.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.298, + 0.314, + 0.314 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.474, + 0.664 + ], + "angle": 0, + "content": "The rapid advancement of generative models, facilitating the creation of hyper-realistic images from textual descriptions, has concurrently escalated critical societal concerns such as misinformation. Although providing some mitigation, traditional fingerprinting mechanisms fall short in attributing responsibility for the malicious use of synthetic images. This paper introduces a novel approach to model fingerprinting that assigns responsibility for the generated images, thereby serving as a potential countermeasure to model misuse. Our method modifies generative models based on each user's unique digital fingerprint, imprinting a unique identifier onto the resultant content that can be traced back to the user. This approach, incorporating fine-tuning into Text-to-Image (T2I) tasks using the Stable Diffusion Model, demonstrates near-perfect attribution accuracy with a minimal impact on output quality. Through extensive evaluation, we show that our method outperforms baseline methods with an average improvement of \\(11\\%\\) in handling image post-processes. Our method presents a promising and novel avenue for accountable model distribution and responsible use. Our code is available in https://github.com/kylemin/WOUAF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.684, + 0.21, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Recent advancements in generative models have propelled their proficiency, expanding their repertoire to include not just the generation of photorealistic images [4, 14] but also the synthesis of images from textual prompts [21, 24, 26, 28]. These significant strides have equipped individuals with the capacity to leverage these models to create hyperrealistic images that correspond seamlessly with given textual instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.831, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Nonetheless, the escalating prominence of generative models instigates pressing societal apprehensions. A case in point is Deepfake, intentionally crafted to disseminate" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.3, + 0.853, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.54, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Illustration of user attribution based on our method. Please refer to the main text for detailed descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.64 + ], + "angle": 0, + "content": "misinformation, fostering a climate of fake news and political disarray [2, 17, 23]. The gravity of these concerns necessitates calls for governmental intervention to regulate the indiscriminate application of generative models1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.893, + 0.806 + ], + "angle": 0, + "content": "A feasible method to counteract malicious use involves assigning accountability for generated images. One approach to achieve this is by integrating independent fingerprinting modules that can embed user-specific information on top of image generation. The open-source Text-to-Image (T2I) project Stable Diffusion (SD) [26] currently employs this technique using discrete wavelet transform or RivaGAN [34]. However, in the open-source setting, bypassing the fingerprinting module is straightforward and can be achieved by commeting just a single line in the source code [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Is it feasible to achieve user attribution without an independent fingerprinting module? In response, we propose a distributor-oriented methodology named WOUAF, standing for Weight mOdupulation for User Attribution and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.352, + 0.9 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence. The White House" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8974" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Fingerprinting. In practical terms, when a model inventor opens-sources their work to a model distributor such as Huggingface, the distributor could utilize our proposed method to create a generic version. Upon receiving a download request from an end-user, the distributor can adjust the model weights using our technique and deploy a fingerprinted version to the user, simultaneously registering the user's fingerprint into their database. In the event of a model's malicious exploitation, the distributor can decode the fingerprint from the misused image and cross-reference it with their database to identify the responsible user. Consequently, this provides the distributor with an actionable method to counteract malicious uses of the model (see Fig. 1 for a comprehensive framework of our methodology)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Our methodology, designed for T2I tasks, is integrated into the Stable Diffusion (SD) framework without necessitating any structural changes to the model. This design choice effectively prevents end-users from bypassing the fingerprinting process. Consistent with prior research [5, 16, 22, 32, 33], our primary goal is to maintain high attribution accuracy while ensuring minimal impact on output quality, as elaborated in Sec. 3. Our rigorous evaluations of this method concentrate initially on assessing both attribution accuracy and image quality. We have found that our approach attains nearly flawless attribution accuracy with only a slight influence on image quality. Moreover, we evaluate the robustness of our method in various scenarios involving post-processing manipulations that images might undergo. Our method outperforms baseline methods in these robustness tests, showing an average improvement of \\(11\\%\\) in handling such manipulations (refer to Sec. 4 for further details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.473, + 0.775 + ], + "angle": 0, + "content": "There are four main contributions: (1) We introduce WOUAF, a distinctive distributor-centered fine-tuning methodology. This approach embeds fingerprints within the model in such a way that end-users cannot easily circumvent or remove them. (2) Our method successfully achieves high attribution accuracy, while maintaining the quality of the output images. (3) Our approach exhibits marked resilience against a diverse array of image post-processes, a vital attribute for practical applications. (4) We conduct thorough assessments to balance attribution accuracy with manipulations to intentionally remove fingerprints, including strategies like image compression via auto-encoders and obfuscation through model fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.789, + 0.22, + 0.805 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.815, + 0.47, + 0.862 + ], + "angle": 0, + "content": "In this section, we discuss related works of model fingerprinting in generative models. More related works are available in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Inventor-oriented Model Fingerprinting. Yu et al. [33] leveraged a pre-trained deep steganography model to em" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.396 + ], + "angle": 0, + "content": "bed fingerprints into the training set for fingerprinted GANs. However, this approach suffers from limited scalability, as it necessitates training a GAN from scratch for each distinct fingerprint. To address this issue, Yu et al. [32] introduced a weight modulation method [12] that directly embeds a user's fingerprint into the generator's weights. Despite these advancements, current methods are predominantly tailored for GAN-based models and typically require training from scratch. This raises important questions regarding their suitability for diffusion-based models, which have a different structural makeup compared to GANs, and the feasibility of avoiding the requirement for training from scratch. The adoption of fine-tuning as a method for embedding fingerprints presents a promising solution. It facilitates the incorporation of fingerprints into pre-trained diffusion models, eliminating the necessity for comprehensive retraining from the ground up [5, 36]. This approach significantly streamlines the process, allowing model inventors to concentrate on core model development without the complexities of embedding fingerprints during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.709 + ], + "angle": 0, + "content": "Distributor-oriented Model Fingerprinting. Kim et al. [16] proposed a technique for achieving user attribution by explicitly incorporating user-specific fingerprints into the generator's output. While this simplified attribution method allowed for the derivation of sufficient fingerprint conditions, it necessitates a trade-off between the quality of the generated output and attribution accuracy, which is further exacerbated when image post-processes are taken into account. To tackle this issue, an approach has been proposed that utilizes subtle semantic variations along latent dimensions as fingerprints, generated by perturbations of eigenvectors in the latent distribution [22]. This method demonstrates an improved balance between generation quality and attribution accuracy. However, its applicability is restricted to unconditional image generation, as eigenvectors are computed by sampling the learned latent representation. In the context of conditional image generation, estimating eigenvectors of latent representation becomes challenging due to the vast space of conditions, such as those found in text conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recent Advances in Fingerprinting for Text-to-Image Diffusion Models. Recent studies [5, 31, 36] have scrutinized fingerprinting techniques in the Stable Diffusion model [26], uncovering vulnerabilities in existing methods [34] that facilitate easy circumvention [5] or robust post-hoc fingerprinting [31]. Fernandez et al. [5] achieved near-perfect attribution accuracy by fine-tuning user-specific models to align with steganography module [37], demonstrating a viable alternative to conventional post-hoc fingerprinting modules [34]. However, this approach scales linearly in computational demand with the number of users since it necessitates fixed-time fine-tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8975" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.506, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.265, + 0.381, + 0.281 + ], + "angle": 0, + "content": "(a) The overall pipeline." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.885, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.265, + 0.782, + 0.28 + ], + "angle": 0, + "content": "(b) Weight modulation." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.291, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Figure 2. Depiction of our method's pipeline and weight modulation: (a) The model fingerprinting procedure encompasses encoding via the mapping network and weight modulation, along with decoding through the fingerprint decoding network. (b) Weight modulation of the decoding network \\(\\mathcal{D}\\) to incorporate the fingerprint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.468, + 0.435 + ], + "angle": 0, + "content": "for each individual. In contrast, our method requires only a one-time training followed by a negligible forward pass time to generate user-specific models. Furthermore, our approach shows superior robustness against common image post-processing techniques compared to that of Stable Signature [5] (refer to Section Sec. 4.6 for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.468, + 0.573 + ], + "angle": 0, + "content": "Another notable contribution is by Wen et al. [31], who introduced an alternative fingerprinting method for the Stable Diffusion model [26]. Their method, similar to posthoc style fingerprinting [34], depends on user-driven embedding, which allows end-users the option to exclude the fingerprint. Moreover, it is confined to the DDIM scheduler [30]. Our method, in contrast, is adaptable to both the DDIM [30] and Euler schedulers [15], underscoring its versatility and wider applicability (refer to the Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.176, + 0.601 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.468, + 0.716 + ], + "angle": 0, + "content": "This section outlines our approach, beginning with an overview of the Text-to-Image (T2I) diffusion model with a focus on the Stable Diffusion (SD) model [26] detailed in Sec. 3.1. We then introduce our key component, the user-specific weight modulation, in Sec. 3.2. The section concludes with a detailed explanation of our training objectives and methods, outlined in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.218, + 0.74 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our approach utilizes the Stable Diffusion (SD) model, which functions within the latent space framework of an autoencoder. SD comprises two main elements: Firstly, an autoencoder is pre-trained on an extensive dataset of images. Its encoder, \\(\\mathcal{E}(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_z}\\), converts an image \\(x\\sim p_{data}\\) into a latent representation \\(z = \\mathcal{E}(x)\\). The decoder, \\(\\mathcal{D}(\\cdot):\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_x}\\), then reconstructs the original image from this latent representation, resulting in \\(\\hat{x} = \\mathcal{D}(z)\\). The secondary element is a diffusion model, based on the U-Net architecture [27], represented as \\(\\epsilon_{\\theta}\\). This model is" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.346, + 0.892, + 0.377 + ], + "angle": 0, + "content": "adept at generating latent representations and can be conditioned using pre-trained text embeddings" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.385, + 0.787, + 0.401 + ], + "angle": 0, + "content": "3.2. User-specific Weight Modulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Our method is fundamentally based on integrating fingerprints into the parameters of the SD through weight modulation [12, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.893, + 0.665 + ], + "angle": 0, + "content": "The overall pipeline of our method is illustrated in Fig. 2(a). A user-specific fingerprint is drawn from a Bernoulli distribution with a probability of 0.5, represented as \\(\\phi \\in \\Phi \\coloneqq \\operatorname{Bernoulli}(0.5)^{d_{\\phi}}\\), where \\(d_{\\phi}\\) signifies the fingerprint length in bits. We employ a mapping network \\(\\mathcal{M}(\\cdot): \\mathbb{R}^{d_{\\phi}} \\to \\mathbb{R}^{d_M}\\) to convert the sampled fingerprint \\(\\phi\\) into an intermediate fingerprint representation within the \\(d_M\\) dimension. For modulating each layer in the SD component, we introduce an affine transformation layer, \\(\\mathcal{A}_l(\\cdot): \\mathbb{R}^{d_M} \\to \\mathbb{R}^{d_j}\\), for all layers \\(l\\). As depicted in Fig. 2(b), this transformation matches the dimensions between \\(d_M\\) and the \\(j\\)-th channel in weight \\(W \\in \\mathbb{R}^{i,j,k}\\), where \\(i,j,k\\) denote input, output, and kernel dimensions, respectively. The weight modulation for the \\(l\\)-th layer is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.673, + 0.891, + 0.694 + ], + "angle": 0, + "content": "\\[\nW _ {i, j, k} ^ {\\phi} = u _ {j} * W _ {i, j, k}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where \\(W\\) and \\(W^{\\phi}\\) denote the pre-trained and fingerprinted weights respectively, \\(u_{j} = \\mathcal{A}_{l}(\\mathcal{M}(\\phi))\\) is the scale of the fingerprint representation corresponding to the \\(j\\)th output channel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We incorporate fingerprints into the SD by applying weight modulation exclusively to the weights in the decoder \\(\\mathcal{D}\\). The rationale for not applying modulation to both the diffusion model \\(\\epsilon_{\\theta}\\) and decoder \\(\\mathcal{D}\\), an approach that mirrors GAN-based models [32], is explained in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.689, + 0.864 + ], + "angle": 0, + "content": "3.3. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our training architecture comprises two primary objectives. The initial objective is to decode fingerprints from the pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8976" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Evaluation of Attribution Accuracy and Image Generation Quality. We conducted validation using the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which were excluded from our training phase. Symbols \\(\\uparrow\\) and \\(\\downarrow\\) denote preferred higher and lower values, respectively." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.142, + 0.895, + 0.284 + ], + "angle": 0, + "content": "
ModelFingerprinting Time (↓)MS-COCOLAION
Attribution Acc (↑)CLIP-score (↑)FID (↓)Attribution Acc (↑)CLIP-score (↑)FID (↓)
Original SD [26]--0.7324.48-0.5019.67
DAG [16]8.4 hr0.700.7326.540.710.4923.13
Stable Signature [5]< 1 min0.990.7324.550.980.5020.02
WOUAF-conv< 1 sec0.990.7324.430.980.5120.46
WOUAF-all< 1 sec0.990.7324.420.990.5119.91
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.345 + ], + "angle": 0, + "content": "vided images. We train a fingerprint decoding network \\(\\mathcal{F}(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_\\phi}\\), which is instantiated by ResNet-50 [6], as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.358, + 0.47, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\phi} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\sum_ {i = 1} ^ {d _ {\\phi}} [ \\phi_ {i} \\log \\sigma (\\mathcal {F} (\\mathcal {D} (\\phi , z)) _ {i} \\\\ \\left. + \\left(1 - \\phi_ {i}\\right) \\log \\left(1 - \\sigma \\left(\\mathcal {F} \\left(\\mathcal {D} (\\phi , z)\\right)\\right) _ {i} \\right], \\right. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.571 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) refers to the sigmoid activation function, constraining the output of \\(\\mathcal{F}\\) to the range [0, 1]. Thus, this loss function effectively combines binary cross-entropy for all bits of the fingerprint. During training time, fingerprint \\(\\phi\\) is sampled from Bernoulli distribution. However, after training, the model distributor initially samples a user-specific fingerprint \\(\\phi_{\\alpha}\\) and subsequently modulates the decoder \\(\\mathcal{D}\\) using \\(\\phi_{\\alpha}\\). The user will receive the fingerprinted decoder \\(\\mathcal{D}(\\phi_{\\alpha}, \\cdot)\\), which solely permits latent input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.471, + 0.635 + ], + "angle": 0, + "content": "The secondary objective endeavors to regularize the quality of outputs. Ideally, this regularization inhibits the decoder \\(\\mathcal{D}\\) from compromising image quality while minimizing \\(L_{\\phi}\\) in Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.648, + 0.47, + 0.668 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {q u a l i t y}} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\left[ \\ell (x, \\mathcal {D} (\\phi , z)) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.741 + ], + "angle": 0, + "content": "\\(\\ell\\) represents the distance metric between original images and fingerprinted images. For practical applications, we utilize perceptual distance [35] to gauge the perceptual difference between \\(x\\) and \\(\\mathcal{D}(\\phi, z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.742, + 0.426, + 0.758 + ], + "angle": 0, + "content": "The final objective function can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.773, + 0.47, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {A}, \\mathcal {M}, \\mathcal {D}, \\mathcal {F}} \\lambda_ {1} L _ {\\phi} + \\lambda_ {2} L _ {\\text {q u a l i t y}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.9 + ], + "angle": 0, + "content": "where both \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are set to 1.0. Fundamentally, the loss function aspires to reconstruct fingerprints while maintaining the quality of the generated outputs. To assess the efficacy of our proposed method, we employ attribution accuracy and image quality metrics (Refer to Sec. 4.1 for details)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.298, + 0.634, + 0.316 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.324, + 0.694, + 0.341 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Datasets. Our approach is fine-tuned on the MS-COCO [19] dataset, adopting the Karpathy split. For methodological evaluation, we harness the test set from MS-COCO and randomly sample from the LAION-aesthetics [29] dataset. For T2I image generation, we adopt the Euler scheduler [15] with timestep \\(T = 20\\), and the classifier-free guidance scale [9] is set to 7.5 unless otherwise specified. Evaluation for DDIM scheduler [30] and various image generation hyperparameters are available in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.893, + 0.704 + ], + "angle": 0, + "content": "Experimental Setting. We implement the weight modulation following the design specified in the source code of StyleGAN2-ADA [13]. Our mapping network \\( M \\) is designed with a series of fully connected layers, wherein all experiments are conducted using a two-layer configuration. To train robust models against image post-processing transformations, differentiable post-processes are necessary. To this end, we incorporate the Kornia library [25]. For Stable Signature [5], we utilize the official code provided by the authors. We note that its post-processing transformations are replaced with our version for fair comparison. Appendix includes details on mapping network dimensions, training parameters, and optimizer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.902 + ], + "angle": 0, + "content": "**Evaluations.** User attribution accuracy is gauged by the formula: \\(\\frac{1}{d_{\\phi}} \\sum_{i=1}^{d_{\\phi}} \\mathbb{1}(\\phi_i = \\hat{\\phi}_i)\\), where \\(\\phi\\) is the true fingerprint and \\(\\hat{\\phi} = \\mathbb{1}\\left[\\sigma(\\mathcal{F}(x_{\\phi})) > 0.5\\right]\\) is the estimated fingerprint from image \\(x_{\\phi}\\). Unless otherwise stated, \\(d_{\\phi}\\) is set to 32 in our experiments (Refer to Sec. 4.2 for additional information). We further employ a statistical test [32, 33] to evaluate matching bits between \\(\\hat{\\phi}\\) and \\(\\phi\\). The null hypothesis \\(H_0\\) suggests that the number of matching bits arises by chance. The test uses a binomial distribution, with a \\(p\\)-value derived as: \\(P(X \\geq k|H_0) = \\sum_{i=k}^{d_{\\phi}} \\binom{d_{\\phi}}{i} 0.5^{d_{\\phi}}\\). A \\(p\\)-value below 0.05 leads to the rejection of \\(H_0\\), with \\(1 - p\\) serving as an indicator of verification confidence. Lastly, to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8977" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.482, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.337, + 0.383, + 0.352 + ], + "angle": 0, + "content": "(a) MS COCO" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.089, + 0.892, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.337, + 0.808, + 0.351 + ], + "angle": 0, + "content": "(b) LAION Aesthetics" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.894, + 0.405 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of the original and fingerprinted Stable Diffusion models on MS-COCO [19] and LAION aesthetics [29] (Pixel-wise differences \\(\\times\\) 5: they are multiplied by a factor of 5 for better view). We can observe that our method maintains high image quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.493 + ], + "angle": 0, + "content": "validate the quality of our method, we assess image quality using the Fréchet Inception Distance (FID) [8] and employ the Clip-score [7] to determine the alignment between text and generated images. Additional experimental details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.471, + 0.775 + ], + "angle": 0, + "content": "Models. For evaluating our methodology, we benchmark against two established baseline methods: DAG [16] and Stable Signature [5]. Both these methods, conceptualized from the model distributor's standpoint, incorporate finetuning for model fingerprinting. To ensure a fair comparison, we retrain the Stable Signature method within our training settings by replacing its post-processing scheme. Additionally, we evaluate our method against three distinct variants based on the specific layers chosen for weight modulation implementation. The first variant, WOUAF-conv, applies modulation to only the convolutional layers in \\(\\mathcal{D}\\). In contrast, WOUAF-all extends this approach across all layers of \\(\\mathcal{D}\\), covering both self-attention and convolution layers. The final variant implements weight modulation in both the diffusion model \\(\\epsilon_{\\theta}\\) and the decoder \\(\\mathcal{D}\\), mirroring the approach used in GAN-based methods [32]. Further details on why this variant is not used in our experiments are discussed in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.277, + 0.803 + ], + "angle": 0, + "content": "4.2. Fingerprint Capacity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The capacity of our method depends on the maximum number of unique user-specific fingerprints it can support without significant crosstalk. This capacity is primarily influenced by the fingerprint dimension \\((d_{\\phi})\\). Selecting an optimal \\(d_{\\phi}\\) presents a challenge: while a larger \\(d_{\\phi}\\) can accommodate more users, it also complicates effective fingerprint" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.416, + 0.892, + 0.443 + ], + "angle": 0, + "content": "Table 2. Experiments of attribution accuracy across various fingerprint dimensions \\((d_{\\phi})\\)" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.454, + 0.86, + 0.539 + ], + "angle": 0, + "content": "
Fingerprint Dims.Attribution Accuracy
163264128
WOUAF-conv0.990.990.980.94
WOUAF-all0.990.990.990.97
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.557, + 0.6, + 0.572 + ], + "angle": 0, + "content": "decoding [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.74 + ], + "angle": 0, + "content": "To investigate this trade-off, we conduct an analysis with varying fingerprint dimensions, specifically \\( d_{\\phi} \\) values of 16, 32, 64, and 128. Tab. 2 presents the user attribution accuracy for each \\( d_{\\phi} \\) value. As shown in Tab. 2, attribution accuracy tends to decrease monotonically as \\( d_{\\phi} \\) increases. Importantly, both our variant models achieve a near-perfect attribution accuracy of 0.99 for \\( d_{\\phi} \\) values of 16, 32, and 64. However, for \\( d_{\\phi} = 128 \\), WOUAF-all variant outperforms the WOUAF-conv variant. For a balanced comparison with existing methods, we choose \\( d_{\\phi} = 32 \\), which notably can support a substantial user base exceeding 4 billion \\( \\approx 2^{32} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.754, + 0.85, + 0.77 + ], + "angle": 0, + "content": "4.3. Attribution Accuracy and Image Quality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.891, + 0.868 + ], + "angle": 0, + "content": "We conduct a comprehensive evaluation of WOUAF, focusing on attribution accuracy and image quality. The assessment involves the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which are excluded from the training phase. The results, detailed in Tab. 1, showcase the efficacy of our method." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our variants, namely WOUAF-conv and WOUAF-all, demonstrate superior performance in attribution accuracy" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8978" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "over DAG [16], indicating their proficiency in accurately decoding embedded fingerprints from the generated images. These variants also show competitive results when compared to Stable Signature [5], reinforcing our methodology's robustness. Notably, we achieve this high level of accuracy without significantly compromising image quality. Both FID scores and Clip-scores showed minimal variation from the baseline SD model, indicating that our approach has a negligible impact on image output quality. This is further corroborated by qualitative examples in Fig. 3, which highlight WOUAF's ability to reliably incorporate fingerprinting without degrading image generation quality. For additional insights, uncurated image collections are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.47, + 0.395 + ], + "angle": 0, + "content": "Given the growing importance of T2I models, computation time for fingerprinting emerges as a key metric. Our method stands out in computational efficiency. It contrasts with approaches like Stable Signature that need fine-tuning for each new fingerprint. Our method requires just a single forward pass, markedly reducing computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.47, + 0.421 + ], + "angle": 0, + "content": "4.4. Attribution Analysis for Diverse Image Sources" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Investigating the attribution of generated images to responsible users, we explore the potential for images from non-fingerprinted or varied sources to bypass our system. Our analysis aims to determine if decoded fingerprints from such images match any entries in the model distributor's database. A mismatch indicates the image's external origin, absolving users in the database." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.67 + ], + "angle": 0, + "content": "We adopt the experimental setup from [32], compiling a dataset with different image types: authentic images from the MS-COCO test set [19], non-fingerprinted images from Stable Diffusion [26], and synthesized images from ProGAN [10], StyleGAN [11], and StyleGAN2 [12], with each category containing 1,000 samples. Given our extensive user database of 1 million entries, we set a threshold at \\(32 * 0.95 \\approx 30\\) bits, aligning with our 0.99 attribution accuracy as shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.471, + 0.762 + ], + "angle": 0, + "content": "Our rigorous experiments revealed that, irrespective of the source, no images were incorrectly attributed as possessing a fingerprint from our 1 million fingerprint database. This reinforces the reliability of our attribution approach as detailed in Sec. 4.3 demonstrating the robustness of our system against diverse image sources." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.388, + 0.789 + ], + "angle": 0, + "content": "4.5. Benefits of Finetuning only Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "When developing our last variant that incorporates weight modulation into both the diffusion model \\(\\epsilon_{\\theta}\\) and the decoder \\(\\mathcal{D}\\), we note that the resultant pipeline demonstrates similarities with the GAN-based method [32]. A direct comparison between ours and the GAN-based methods may not be entirely straightforward, given the fundamental differences in their training methodologies. This is because the" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.092, + 0.892, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.317, + 0.894, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Comparative analysis of weight modulation on decoder \\(\\mathcal{D}\\) and diffusion model \\(\\epsilon_{\\theta}\\) with decoder \\(\\mathcal{D}\\). Modulating the diffusion model negatively affects image quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.481 + ], + "angle": 0, + "content": "GAN-based methods entail training from scratch, whereas our proposed approach leans towards fine-tuning. Nevertheless, both methodologies share a common mechanism: they aim to modulate the weights of the layers instrumental in learning the latent space. The shared characteristic underscores the fundamental objective of optimizing the balance between attribution accuracy and generation quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.893, + 0.633 + ], + "angle": 0, + "content": "However, our empirical observations suggest that this variant does not consistently achieve commendable performance as an attribution model. Specifically, it appears that this variant can only optimize either attribution accuracy or generation quality, but not both simultaneously. In our tests, the highest attribution accuracy reached by this variant is \\(89\\%\\), with a Clip-score of 0.68 and FID of 63.48 (detailed in Fig. 4). The inherent trade-off observed here further reinforces the challenge of balancing these two critical parameters in the context of model fingerprinting techniques." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.64, + 0.892, + 0.671 + ], + "angle": 0, + "content": "4.6. Robust User Attribution against Image Post-processes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.892, + 0.738 + ], + "angle": 0, + "content": "This section evaluates the robustness of our method in scenarios where generated images undergo post-processing. These processes could potentially alter the embedded fingerprint within the images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Consistent with methodologies outlined in previous research [5, 16, 22, 32, 33], we examine our model's resilience to various image post-processing operations. We simulate the effect of post-processing at random intensities before inputting data into the fingerprint decoding network, \\(\\mathcal{F}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.836, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {r o b u s t}} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\sum_ {i = 1} ^ {d _ {\\phi}} [ \\phi_ {i} \\log \\sigma (\\mathcal {F} (T (\\mathcal {D} (\\phi , z))) _ {i} \\\\ \\left. + \\left(1 - \\phi_ {i}\\right) \\log \\left(1 - \\sigma \\left(\\mathcal {F} \\left(T (\\mathcal {D} (\\phi , z))\\right)\\right) _ {i} \\right], \\quad (5) \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8979" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.3, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.092, + 0.496, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.092, + 0.689, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.092, + 0.885, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.238, + 0.3, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.238, + 0.496, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.238, + 0.689, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.238, + 0.885, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.893, + 0.416 + ], + "angle": 0, + "content": "Figure 5. Enhanced Robustness Against Image Post-Processes. For almost all scenarios, WOUAF consistently exceeds the performance of DAG [16] and Stable Signature [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.468, + 0.49 + ], + "angle": 0, + "content": "where, \\(T(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_x}\\) denotes the post-processing function. In the optimization process, we employ an objective function akin to the one detailed in Eq. (4), with \\(L_{\\phi}\\) substituted by \\(L_{\\mathrm{robust}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.763 + ], + "angle": 0, + "content": "In our exploration, we contemplate eight different post-processing techniques: Erasing, Rotation, Gaussian Blurring, Cropping, Brightness jittering, the addition of Gaussian Noise, JPEG compression, and a Combination of all these post-processes. The parameters for these post-processes are designed as follows: For random erasing, we use a random erase ratio within the range \\([5\\%, 10\\%, 15\\%, 20\\%]\\). Rotation involves randomly sampling a degree within the range (-30, 30). For Gaussian Blurring, we randomly select a kernel size from [3, 5, 7]. For Cropping, we use a random cropping-out ratio within the range \\([5\\%, 10\\%, 15\\%, 20\\%]\\). The Brightness factor is randomly sampled within the range (-0.3, 0.3). For Gaussian Noise, we add noise with a standard deviation randomly sampled from a uniform distribution \\(U[0, 0.2]\\). JPEG compression quality level is selected from [90, 80, 70, 60, 50]. The Combination technique randomly selects a subset of these seven post-processing methods with a probability of 0.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.9 + ], + "angle": 0, + "content": "User attribution accuracy for each post-process is evaluated under these parameters. Our tests, depicted in Fig.5, offer a comparative analysis of user attribution accuracy across robust versions of DAG [16], Stable Signature [5], and WOUAF. Remarkably, our method demonstrates robustness across a range of post-processes, achieving an attribution accuracy improvement of \\(11\\%\\) over Stable Signature and \\(29\\%\\) over DAG. A notable trend across all transformations is the monotonic decrease in user attribution accu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.596 + ], + "angle": 0, + "content": "racy as the intensity of post-processing increases. This reinforces the challenges posed by post-processing in maintaining accurate user attribution. However, our results also underscore the benefits of robust training in overcoming these challenges, emphasizing the importance of resilient training strategies for fingerprinting methods in the face of post-processing transformations. Considering the robustness of our method against various post-processes, it becomes a viable choice for model distributors seeking reliable fingerprinting solutions. Detailed results of FID scores and visual examples are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.611, + 0.842, + 0.629 + ], + "angle": 0, + "content": "5. Deliberate Fingerprint Manipulations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.714 + ], + "angle": 0, + "content": "This section delves into our method's robustness against deliberate attempts to remove fingerprints, which include malicious manipulations via auto-encoders and model purification. Further details and extended attack scenarios are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.725, + 0.796, + 0.741 + ], + "angle": 0, + "content": "5.1. Resilience Against Deep Classifier" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The imperceptibility of the fingerprint in generated images is crucial to prevent its detection and subsequent tampering by malicious entities. To assess the secrecy of our method, we adopt an attack scenario akin to the one in [32], assuming an attacker aims to train a classifier to detect the presence of a fingerprint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We assume that the attacker seeks to train a classifier capable of detecting the presence of a fingerprint. To assess this scenario, we utilize a pretrained ResNet-50 [6] based binary classifier, trained using 10K SD generated images" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "8980" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.226 + ], + "angle": 0, + "content": "(5K original SD images and 5K fingerprinted SD images). This configuration is deemed valid as detecting the presence of a fingerprint necessitates using both non-fingerprinted and fingerprinted images in the training set. The binary classifier achieve \\(98\\%\\) accuracy in the training stage. In subsequent evaluations using a separate set of 5K images from our variant models, the binary classification accuracy is 0.66 for WOUAF-conv and just 0.56 for WOUAF-all, which is nearly equivalent to random chance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.468, + 0.333 + ], + "angle": 0, + "content": "These findings imply that detecting our embedded fingerprint, particularly in the WOUAF-all variant, poses a challenge to detect. Upcoming subsections will delve into further evaluations, predicated on the stringent assumption that users are cognizant of the fingerprint's presence and endeavor to eliminate it by employing auto-encoder methods or fine-tuning techniques." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.344, + 0.373, + 0.36 + ], + "angle": 0, + "content": "5.2. Resilience Against Auto-Encoders" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.468, + 0.563 + ], + "angle": 0, + "content": "In contexts where adversaries aim to alter output images, leveraging deep learning techniques such as neural auto-encoders [1, 3, 20] becomes a common strategy for the purpose of obfuscating or removing fingerprints embedded in images [5]. To assess the resilience of our approach, we utilize the robust model against JPEG described in Sec. 4.6. This comparison is appropriate as JPEG represents a conventional image compression method. However, the auto-encoders [1, 3, 20] employed in our evaluation exhibit superior compression performance compared to JPEG. Our research explores the resilience of our proposed method against these sophisticated auto-encoders, focusing particularly on the impact of their varying compression rates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.468, + 0.713 + ], + "angle": 0, + "content": "As depicted on the left side of Fig. 6, our investigations reveal a notable trend: attribution accuracy progressively declines towards a near-random level (approximately \\(50\\%\\)) as the compression rate employed by the auto-encoders escalates. This trend highlights a critical trade-off: the reduction in attribution accuracy is achievable solely by compromising the quality of the image [18]. Our findings indicate that compromising the integrity of the image is a necessary consequence to effectively obscure the fingerprinting process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.402, + 0.741 + ], + "angle": 0, + "content": "5.3. Resilience Against Model Purification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.854 + ], + "angle": 0, + "content": "This subsection addresses the scenario where an adversary, upon recognizing the presence of fingerprints within the images generated by the image decoder \\(\\mathcal{D}\\), opts to fine-tune \\(\\mathcal{D}\\) with the objective of obliterating the embedded fingerprint. This strategy, known as model purification, is a sophisticated approach to altering the model's output to erase traceable imprints [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this adversarial setting, the primary aim is to refine the downloaded fingerprinted model by optimizing the reconstruction error between the adversary's proprietary im" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.703, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.892, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 6. Left: Auto-Encoder-based Fingerprint Removal. With heightened compression rates, both image quality and attribution accuracy experience a decrease. Right: Model Purification. Progressive fine-tuning leads to concurrent declines in both image quality and attribution accuracy. Note that a lower FID score is preferable, indicating better image quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.485 + ], + "angle": 0, + "content": "age dataset and the output from the fingerprinted model. By adhering to the experimental framework outlined in [5], we charted the interplay between FID scores and attribution accuracy, as presented on the right side of Fig. 6. Our empirical analysis reveals a significant challenge: efforts to decrease the attribution accuracy lead to a decline in the quality of the generated images. This finding underscores the inherent complexity in fine-tuning processes aimed at model purification, particularly when striving to maintain the visual quality of the output while endeavoring to obscure its traceable characteristics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.506, + 0.618, + 0.521 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.744 + ], + "angle": 0, + "content": "In this study, we have delved into user attribution for Stable Diffusion-based Text-to-Image (T2I) model, employing a weight modulation-based fingerprinting approach. Our method, WOUAF, not only achieves near-perfect accuracy but also preserves the high quality of generated images. A key aspect of WOUAF is its computational efficiency coupled with enhanced robustness against various image post-processing techniques compared to existing baselines. Our results lay a solid groundwork for future exploration into the broader implications and challenges posed by generative models. In future work, we plan to expand and refine our methodology to encompass various data types including text, audio, and video, necessitating tailored adjustments in model fingerprinting techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.669, + 0.78 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.887 + ], + "angle": 0, + "content": "This work is partially supported by the National Science Foundation under Grant No. 2038666, No. 2101052, and a grant from Meta AI Learning Alliance. The HuggingFace demo of this work is funded by Intel Corporation. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8981" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[2] Ali Breland. The bizarre and terrifying case of the \"deep-fake\" video that helped bring an african nation to the brink. motherjones, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.47, + 0.312 + ], + "angle": 0, + "content": "[3] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[5] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.469, + 0.541 + ], + "angle": 0, + "content": "[7] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.543, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.645, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[10] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. arXiv preprint arXiv:1912.04958, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[13] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[16] Changhoon Kim, Yi Ren, and Yezhou Yang. Decentralized attribution of generative models. In International Conference on Learning Representations, 2021. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[17] Arijeta Lajka. New ai voice-cloning tools 'add fuel' to misinformation fire. AP News, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[18] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. ArXiv, abs/2103.09274, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.28, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.365, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[20] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, 3-8 December 2018, Montréal, Canada, pages 10794–10803, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[22] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[23] Matt Novak. Ai image creator midjourney halts free trials but it has nothing to do with the pope's jacket. forbes, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[25] E. Riba, D. Mishkin, D. Ponsa, E. Rublee, and G. Bradski. Kornia: an open source differentiable computer vision library for pytorch. In Winter Conference on Applications of Computer Vision, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.955 + ], + "angle": 0, + "content": "8982" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[31] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[32] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.472, + 0.486 + ], + "angle": 0, + "content": "[33] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[34] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[36] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. ArXiv, abs/2303.10137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.471, + 0.699 + ], + "angle": 0, + "content": "[37] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 657–672, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.955 + ], + "angle": 0, + "content": "8983" + } + ] +] \ No newline at end of file diff --git a/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_origin.pdf b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..65623312616ddefdaea0bf988a639a784d821d0f --- /dev/null +++ b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/92d9125f-3522-4960-9c33-6d71a67f7675_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20a2b3928331d44737341b773cb7c3936e4b8daa8e0847ad880ea29a696eb274 +size 1615154 diff --git a/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/full.md b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b5eb5ac52ee299c7a3de3083dd999cda20506150 --- /dev/null +++ b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/full.md @@ -0,0 +1,290 @@ +# WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models + +Changhoon Kim $^{1}$ Kyle Min $^{2}$ Maitreya Patel $^{1}$ Sheng Cheng $^{1}$ Yezhou Yang $^{1}$ $^{1}$ Arizona State University $^{2}$ Intel Labs + +{kch,maitreya.patel,scheng53,yz.yang}@asu.edu kyle.min@intel.com + +# Abstract + +The rapid advancement of generative models, facilitating the creation of hyper-realistic images from textual descriptions, has concurrently escalated critical societal concerns such as misinformation. Although providing some mitigation, traditional fingerprinting mechanisms fall short in attributing responsibility for the malicious use of synthetic images. This paper introduces a novel approach to model fingerprinting that assigns responsibility for the generated images, thereby serving as a potential countermeasure to model misuse. Our method modifies generative models based on each user's unique digital fingerprint, imprinting a unique identifier onto the resultant content that can be traced back to the user. This approach, incorporating fine-tuning into Text-to-Image (T2I) tasks using the Stable Diffusion Model, demonstrates near-perfect attribution accuracy with a minimal impact on output quality. Through extensive evaluation, we show that our method outperforms baseline methods with an average improvement of $11\%$ in handling image post-processes. Our method presents a promising and novel avenue for accountable model distribution and responsible use. Our code is available in https://github.com/kylemin/WOUAF. + +# 1. Introduction + +Recent advancements in generative models have propelled their proficiency, expanding their repertoire to include not just the generation of photorealistic images [4, 14] but also the synthesis of images from textual prompts [21, 24, 26, 28]. These significant strides have equipped individuals with the capacity to leverage these models to create hyperrealistic images that correspond seamlessly with given textual instructions. + +Nonetheless, the escalating prominence of generative models instigates pressing societal apprehensions. A case in point is Deepfake, intentionally crafted to disseminate + +![](images/dcc97cf11cc5d6d310a8a6556f801e8e70d24f1e1ca82fdc1f79fdbbe32c9294.jpg) +Figure 1. Illustration of user attribution based on our method. Please refer to the main text for detailed descriptions. + +misinformation, fostering a climate of fake news and political disarray [2, 17, 23]. The gravity of these concerns necessitates calls for governmental intervention to regulate the indiscriminate application of generative models1. + +A feasible method to counteract malicious use involves assigning accountability for generated images. One approach to achieve this is by integrating independent fingerprinting modules that can embed user-specific information on top of image generation. The open-source Text-to-Image (T2I) project Stable Diffusion (SD) [26] currently employs this technique using discrete wavelet transform or RivaGAN [34]. However, in the open-source setting, bypassing the fingerprinting module is straightforward and can be achieved by commeting just a single line in the source code [5]. + +Is it feasible to achieve user attribution without an independent fingerprinting module? In response, we propose a distributor-oriented methodology named WOUAF, standing for Weight mOdupulation for User Attribution and + +Fingerprinting. In practical terms, when a model inventor opens-sources their work to a model distributor such as Huggingface, the distributor could utilize our proposed method to create a generic version. Upon receiving a download request from an end-user, the distributor can adjust the model weights using our technique and deploy a fingerprinted version to the user, simultaneously registering the user's fingerprint into their database. In the event of a model's malicious exploitation, the distributor can decode the fingerprint from the misused image and cross-reference it with their database to identify the responsible user. Consequently, this provides the distributor with an actionable method to counteract malicious uses of the model (see Fig. 1 for a comprehensive framework of our methodology). + +Our methodology, designed for T2I tasks, is integrated into the Stable Diffusion (SD) framework without necessitating any structural changes to the model. This design choice effectively prevents end-users from bypassing the fingerprinting process. Consistent with prior research [5, 16, 22, 32, 33], our primary goal is to maintain high attribution accuracy while ensuring minimal impact on output quality, as elaborated in Sec. 3. Our rigorous evaluations of this method concentrate initially on assessing both attribution accuracy and image quality. We have found that our approach attains nearly flawless attribution accuracy with only a slight influence on image quality. Moreover, we evaluate the robustness of our method in various scenarios involving post-processing manipulations that images might undergo. Our method outperforms baseline methods in these robustness tests, showing an average improvement of $11\%$ in handling such manipulations (refer to Sec. 4 for further details). + +There are four main contributions: (1) We introduce WOUAF, a distinctive distributor-centered fine-tuning methodology. This approach embeds fingerprints within the model in such a way that end-users cannot easily circumvent or remove them. (2) Our method successfully achieves high attribution accuracy, while maintaining the quality of the output images. (3) Our approach exhibits marked resilience against a diverse array of image post-processes, a vital attribute for practical applications. (4) We conduct thorough assessments to balance attribution accuracy with manipulations to intentionally remove fingerprints, including strategies like image compression via auto-encoders and obfuscation through model fine-tuning. + +# 2. Related Work + +In this section, we discuss related works of model fingerprinting in generative models. More related works are available in the appendix. + +Inventor-oriented Model Fingerprinting. Yu et al. [33] leveraged a pre-trained deep steganography model to em + +bed fingerprints into the training set for fingerprinted GANs. However, this approach suffers from limited scalability, as it necessitates training a GAN from scratch for each distinct fingerprint. To address this issue, Yu et al. [32] introduced a weight modulation method [12] that directly embeds a user's fingerprint into the generator's weights. Despite these advancements, current methods are predominantly tailored for GAN-based models and typically require training from scratch. This raises important questions regarding their suitability for diffusion-based models, which have a different structural makeup compared to GANs, and the feasibility of avoiding the requirement for training from scratch. The adoption of fine-tuning as a method for embedding fingerprints presents a promising solution. It facilitates the incorporation of fingerprints into pre-trained diffusion models, eliminating the necessity for comprehensive retraining from the ground up [5, 36]. This approach significantly streamlines the process, allowing model inventors to concentrate on core model development without the complexities of embedding fingerprints during training. + +Distributor-oriented Model Fingerprinting. Kim et al. [16] proposed a technique for achieving user attribution by explicitly incorporating user-specific fingerprints into the generator's output. While this simplified attribution method allowed for the derivation of sufficient fingerprint conditions, it necessitates a trade-off between the quality of the generated output and attribution accuracy, which is further exacerbated when image post-processes are taken into account. To tackle this issue, an approach has been proposed that utilizes subtle semantic variations along latent dimensions as fingerprints, generated by perturbations of eigenvectors in the latent distribution [22]. This method demonstrates an improved balance between generation quality and attribution accuracy. However, its applicability is restricted to unconditional image generation, as eigenvectors are computed by sampling the learned latent representation. In the context of conditional image generation, estimating eigenvectors of latent representation becomes challenging due to the vast space of conditions, such as those found in text conditions. + +Recent Advances in Fingerprinting for Text-to-Image Diffusion Models. Recent studies [5, 31, 36] have scrutinized fingerprinting techniques in the Stable Diffusion model [26], uncovering vulnerabilities in existing methods [34] that facilitate easy circumvention [5] or robust post-hoc fingerprinting [31]. Fernandez et al. [5] achieved near-perfect attribution accuracy by fine-tuning user-specific models to align with steganography module [37], demonstrating a viable alternative to conventional post-hoc fingerprinting modules [34]. However, this approach scales linearly in computational demand with the number of users since it necessitates fixed-time fine-tuning + +![](images/13be24e73fdd824db63437548662696957246b1c8bd2a9a3852cabdcb8418f54.jpg) +(a) The overall pipeline. + +![](images/1118a0670691935fe7c445f3c5a6ab745c029dedda7eec74a9a5cb962ef28fb2.jpg) +(b) Weight modulation. +Figure 2. Depiction of our method's pipeline and weight modulation: (a) The model fingerprinting procedure encompasses encoding via the mapping network and weight modulation, along with decoding through the fingerprint decoding network. (b) Weight modulation of the decoding network $\mathcal{D}$ to incorporate the fingerprint. + +for each individual. In contrast, our method requires only a one-time training followed by a negligible forward pass time to generate user-specific models. Furthermore, our approach shows superior robustness against common image post-processing techniques compared to that of Stable Signature [5] (refer to Section Sec. 4.6 for details). + +Another notable contribution is by Wen et al. [31], who introduced an alternative fingerprinting method for the Stable Diffusion model [26]. Their method, similar to posthoc style fingerprinting [34], depends on user-driven embedding, which allows end-users the option to exclude the fingerprint. Moreover, it is confined to the DDIM scheduler [30]. Our method, in contrast, is adaptable to both the DDIM [30] and Euler schedulers [15], underscoring its versatility and wider applicability (refer to the Appendix). + +# 3. Methods + +This section outlines our approach, beginning with an overview of the Text-to-Image (T2I) diffusion model with a focus on the Stable Diffusion (SD) model [26] detailed in Sec. 3.1. We then introduce our key component, the user-specific weight modulation, in Sec. 3.2. The section concludes with a detailed explanation of our training objectives and methods, outlined in Sec. 3.3. + +# 3.1. Preliminaries + +Our approach utilizes the Stable Diffusion (SD) model, which functions within the latent space framework of an autoencoder. SD comprises two main elements: Firstly, an autoencoder is pre-trained on an extensive dataset of images. Its encoder, $\mathcal{E}(\cdot):\mathbb{R}^{d_x}\to \mathbb{R}^{d_z}$ , converts an image $x\sim p_{data}$ into a latent representation $z = \mathcal{E}(x)$ . The decoder, $\mathcal{D}(\cdot):\mathbb{R}^{d_z}\to \mathbb{R}^{d_x}$ , then reconstructs the original image from this latent representation, resulting in $\hat{x} = \mathcal{D}(z)$ . The secondary element is a diffusion model, based on the U-Net architecture [27], represented as $\epsilon_{\theta}$ . This model is + +adept at generating latent representations and can be conditioned using pre-trained text embeddings + +# 3.2. User-specific Weight Modulation + +Our method is fundamentally based on integrating fingerprints into the parameters of the SD through weight modulation [12, 32]. + +The overall pipeline of our method is illustrated in Fig. 2(a). A user-specific fingerprint is drawn from a Bernoulli distribution with a probability of 0.5, represented as $\phi \in \Phi \coloneqq \operatorname{Bernoulli}(0.5)^{d_{\phi}}$ , where $d_{\phi}$ signifies the fingerprint length in bits. We employ a mapping network $\mathcal{M}(\cdot): \mathbb{R}^{d_{\phi}} \to \mathbb{R}^{d_M}$ to convert the sampled fingerprint $\phi$ into an intermediate fingerprint representation within the $d_M$ dimension. For modulating each layer in the SD component, we introduce an affine transformation layer, $\mathcal{A}_l(\cdot): \mathbb{R}^{d_M} \to \mathbb{R}^{d_j}$ , for all layers $l$ . As depicted in Fig. 2(b), this transformation matches the dimensions between $d_M$ and the $j$ -th channel in weight $W \in \mathbb{R}^{i,j,k}$ , where $i,j,k$ denote input, output, and kernel dimensions, respectively. The weight modulation for the $l$ -th layer is defined as: + +$$ +W _ {i, j, k} ^ {\phi} = u _ {j} * W _ {i, j, k}, \tag {1} +$$ + +where $W$ and $W^{\phi}$ denote the pre-trained and fingerprinted weights respectively, $u_{j} = \mathcal{A}_{l}(\mathcal{M}(\phi))$ is the scale of the fingerprint representation corresponding to the $j$ th output channel. + +We incorporate fingerprints into the SD by applying weight modulation exclusively to the weights in the decoder $\mathcal{D}$ . The rationale for not applying modulation to both the diffusion model $\epsilon_{\theta}$ and decoder $\mathcal{D}$ , an approach that mirrors GAN-based models [32], is explained in Sec. 4.5. + +# 3.3. Training Objectives + +Our training architecture comprises two primary objectives. The initial objective is to decode fingerprints from the pro + +Table 1. Evaluation of Attribution Accuracy and Image Generation Quality. We conducted validation using the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which were excluded from our training phase. Symbols $\uparrow$ and $\downarrow$ denote preferred higher and lower values, respectively. + +
ModelFingerprinting Time (↓)MS-COCOLAION
Attribution Acc (↑)CLIP-score (↑)FID (↓)Attribution Acc (↑)CLIP-score (↑)FID (↓)
Original SD [26]--0.7324.48-0.5019.67
DAG [16]8.4 hr0.700.7326.540.710.4923.13
Stable Signature [5]< 1 min0.990.7324.550.980.5020.02
WOUAF-conv< 1 sec0.990.7324.430.980.5120.46
WOUAF-all< 1 sec0.990.7324.420.990.5119.91
+ +vided images. We train a fingerprint decoding network $\mathcal{F}(\cdot):\mathbb{R}^{d_x}\to \mathbb{R}^{d_\phi}$ , which is instantiated by ResNet-50 [6], as follows: + +$$ +\begin{array}{l} L _ {\phi} = \mathbb {E} _ {z = \mathcal {E} (x), \phi \sim \Phi} \sum_ {i = 1} ^ {d _ {\phi}} [ \phi_ {i} \log \sigma (\mathcal {F} (\mathcal {D} (\phi , z)) _ {i} \\ \left. + \left(1 - \phi_ {i}\right) \log \left(1 - \sigma \left(\mathcal {F} \left(\mathcal {D} (\phi , z)\right)\right) _ {i} \right], \right. \tag {2} \\ \end{array} +$$ + +where $\sigma(\cdot)$ refers to the sigmoid activation function, constraining the output of $\mathcal{F}$ to the range [0, 1]. Thus, this loss function effectively combines binary cross-entropy for all bits of the fingerprint. During training time, fingerprint $\phi$ is sampled from Bernoulli distribution. However, after training, the model distributor initially samples a user-specific fingerprint $\phi_{\alpha}$ and subsequently modulates the decoder $\mathcal{D}$ using $\phi_{\alpha}$ . The user will receive the fingerprinted decoder $\mathcal{D}(\phi_{\alpha}, \cdot)$ , which solely permits latent input. + +The secondary objective endeavors to regularize the quality of outputs. Ideally, this regularization inhibits the decoder $\mathcal{D}$ from compromising image quality while minimizing $L_{\phi}$ in Eq. (2): + +$$ +L _ {\text {q u a l i t y}} = \mathbb {E} _ {z = \mathcal {E} (x), \phi \sim \Phi} \left[ \ell (x, \mathcal {D} (\phi , z)) \right], \tag {3} +$$ + +$\ell$ represents the distance metric between original images and fingerprinted images. For practical applications, we utilize perceptual distance [35] to gauge the perceptual difference between $x$ and $\mathcal{D}(\phi, z)$ . + +The final objective function can be formulated as: + +$$ +\min _ {\mathcal {A}, \mathcal {M}, \mathcal {D}, \mathcal {F}} \lambda_ {1} L _ {\phi} + \lambda_ {2} L _ {\text {q u a l i t y}}, \tag {4} +$$ + +where both $\lambda_{1}$ and $\lambda_{2}$ are set to 1.0. Fundamentally, the loss function aspires to reconstruct fingerprints while maintaining the quality of the generated outputs. To assess the efficacy of our proposed method, we employ attribution accuracy and image quality metrics (Refer to Sec. 4.1 for details). + +# 4. Experiments + +# 4.1. Experiment Settings + +Datasets. Our approach is fine-tuned on the MS-COCO [19] dataset, adopting the Karpathy split. For methodological evaluation, we harness the test set from MS-COCO and randomly sample from the LAION-aesthetics [29] dataset. For T2I image generation, we adopt the Euler scheduler [15] with timestep $T = 20$ , and the classifier-free guidance scale [9] is set to 7.5 unless otherwise specified. Evaluation for DDIM scheduler [30] and various image generation hyperparameters are available in the Appendix. + +Experimental Setting. We implement the weight modulation following the design specified in the source code of StyleGAN2-ADA [13]. Our mapping network $M$ is designed with a series of fully connected layers, wherein all experiments are conducted using a two-layer configuration. To train robust models against image post-processing transformations, differentiable post-processes are necessary. To this end, we incorporate the Kornia library [25]. For Stable Signature [5], we utilize the official code provided by the authors. We note that its post-processing transformations are replaced with our version for fair comparison. Appendix includes details on mapping network dimensions, training parameters, and optimizer. + +**Evaluations.** User attribution accuracy is gauged by the formula: $\frac{1}{d_{\phi}} \sum_{i=1}^{d_{\phi}} \mathbb{1}(\phi_i = \hat{\phi}_i)$ , where $\phi$ is the true fingerprint and $\hat{\phi} = \mathbb{1}\left[\sigma(\mathcal{F}(x_{\phi})) > 0.5\right]$ is the estimated fingerprint from image $x_{\phi}$ . Unless otherwise stated, $d_{\phi}$ is set to 32 in our experiments (Refer to Sec. 4.2 for additional information). We further employ a statistical test [32, 33] to evaluate matching bits between $\hat{\phi}$ and $\phi$ . The null hypothesis $H_0$ suggests that the number of matching bits arises by chance. The test uses a binomial distribution, with a $p$ -value derived as: $P(X \geq k|H_0) = \sum_{i=k}^{d_{\phi}} \binom{d_{\phi}}{i} 0.5^{d_{\phi}}$ . A $p$ -value below 0.05 leads to the rejection of $H_0$ , with $1 - p$ serving as an indicator of verification confidence. Lastly, to + +![](images/81b5af490550c7849846a48c1d9a2b3989c323c3e67c7f8c1823de3b1950f898.jpg) +(a) MS COCO + +![](images/414dc7ecef5c313ac54f68e4af6d44b95bec8eec265e1b21765a60fcacfcb7bb.jpg) +(b) LAION Aesthetics +Figure 3. Qualitative comparison of the original and fingerprinted Stable Diffusion models on MS-COCO [19] and LAION aesthetics [29] (Pixel-wise differences $\times$ 5: they are multiplied by a factor of 5 for better view). We can observe that our method maintains high image quality. + +validate the quality of our method, we assess image quality using the Fréchet Inception Distance (FID) [8] and employ the Clip-score [7] to determine the alignment between text and generated images. Additional experimental details can be found in the Appendix. + +Models. For evaluating our methodology, we benchmark against two established baseline methods: DAG [16] and Stable Signature [5]. Both these methods, conceptualized from the model distributor's standpoint, incorporate finetuning for model fingerprinting. To ensure a fair comparison, we retrain the Stable Signature method within our training settings by replacing its post-processing scheme. Additionally, we evaluate our method against three distinct variants based on the specific layers chosen for weight modulation implementation. The first variant, WOUAF-conv, applies modulation to only the convolutional layers in $\mathcal{D}$ . In contrast, WOUAF-all extends this approach across all layers of $\mathcal{D}$ , covering both self-attention and convolution layers. The final variant implements weight modulation in both the diffusion model $\epsilon_{\theta}$ and the decoder $\mathcal{D}$ , mirroring the approach used in GAN-based methods [32]. Further details on why this variant is not used in our experiments are discussed in Sec. 4.5. + +# 4.2. Fingerprint Capacity + +The capacity of our method depends on the maximum number of unique user-specific fingerprints it can support without significant crosstalk. This capacity is primarily influenced by the fingerprint dimension $(d_{\phi})$ . Selecting an optimal $d_{\phi}$ presents a challenge: while a larger $d_{\phi}$ can accommodate more users, it also complicates effective fingerprint + +Table 2. Experiments of attribution accuracy across various fingerprint dimensions $(d_{\phi})$ + +
Fingerprint Dims.Attribution Accuracy
163264128
WOUAF-conv0.990.990.980.94
WOUAF-all0.990.990.990.97
+ +decoding [18]. + +To investigate this trade-off, we conduct an analysis with varying fingerprint dimensions, specifically $d_{\phi}$ values of 16, 32, 64, and 128. Tab. 2 presents the user attribution accuracy for each $d_{\phi}$ value. As shown in Tab. 2, attribution accuracy tends to decrease monotonically as $d_{\phi}$ increases. Importantly, both our variant models achieve a near-perfect attribution accuracy of 0.99 for $d_{\phi}$ values of 16, 32, and 64. However, for $d_{\phi} = 128$ , WOUAF-all variant outperforms the WOUAF-conv variant. For a balanced comparison with existing methods, we choose $d_{\phi} = 32$ , which notably can support a substantial user base exceeding 4 billion $\approx 2^{32}$ . + +# 4.3. Attribution Accuracy and Image Quality + +We conduct a comprehensive evaluation of WOUAF, focusing on attribution accuracy and image quality. The assessment involves the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which are excluded from the training phase. The results, detailed in Tab. 1, showcase the efficacy of our method. + +Our variants, namely WOUAF-conv and WOUAF-all, demonstrate superior performance in attribution accuracy + +over DAG [16], indicating their proficiency in accurately decoding embedded fingerprints from the generated images. These variants also show competitive results when compared to Stable Signature [5], reinforcing our methodology's robustness. Notably, we achieve this high level of accuracy without significantly compromising image quality. Both FID scores and Clip-scores showed minimal variation from the baseline SD model, indicating that our approach has a negligible impact on image output quality. This is further corroborated by qualitative examples in Fig. 3, which highlight WOUAF's ability to reliably incorporate fingerprinting without degrading image generation quality. For additional insights, uncurated image collections are provided in the Appendix. + +Given the growing importance of T2I models, computation time for fingerprinting emerges as a key metric. Our method stands out in computational efficiency. It contrasts with approaches like Stable Signature that need fine-tuning for each new fingerprint. Our method requires just a single forward pass, markedly reducing computational overhead. + +# 4.4. Attribution Analysis for Diverse Image Sources + +Investigating the attribution of generated images to responsible users, we explore the potential for images from non-fingerprinted or varied sources to bypass our system. Our analysis aims to determine if decoded fingerprints from such images match any entries in the model distributor's database. A mismatch indicates the image's external origin, absolving users in the database. + +We adopt the experimental setup from [32], compiling a dataset with different image types: authentic images from the MS-COCO test set [19], non-fingerprinted images from Stable Diffusion [26], and synthesized images from ProGAN [10], StyleGAN [11], and StyleGAN2 [12], with each category containing 1,000 samples. Given our extensive user database of 1 million entries, we set a threshold at $32 * 0.95 \approx 30$ bits, aligning with our 0.99 attribution accuracy as shown in Tab. 1. + +Our rigorous experiments revealed that, irrespective of the source, no images were incorrectly attributed as possessing a fingerprint from our 1 million fingerprint database. This reinforces the reliability of our attribution approach as detailed in Sec. 4.3 demonstrating the robustness of our system against diverse image sources. + +# 4.5. Benefits of Finetuning only Decoder + +When developing our last variant that incorporates weight modulation into both the diffusion model $\epsilon_{\theta}$ and the decoder $\mathcal{D}$ , we note that the resultant pipeline demonstrates similarities with the GAN-based method [32]. A direct comparison between ours and the GAN-based methods may not be entirely straightforward, given the fundamental differences in their training methodologies. This is because the + +![](images/2c1ba3c77858c1dcf5632e7f38a2ef68307c48119e24f1286bc3936d6d630958.jpg) +Figure 4. Comparative analysis of weight modulation on decoder $\mathcal{D}$ and diffusion model $\epsilon_{\theta}$ with decoder $\mathcal{D}$ . Modulating the diffusion model negatively affects image quality. + +GAN-based methods entail training from scratch, whereas our proposed approach leans towards fine-tuning. Nevertheless, both methodologies share a common mechanism: they aim to modulate the weights of the layers instrumental in learning the latent space. The shared characteristic underscores the fundamental objective of optimizing the balance between attribution accuracy and generation quality. + +However, our empirical observations suggest that this variant does not consistently achieve commendable performance as an attribution model. Specifically, it appears that this variant can only optimize either attribution accuracy or generation quality, but not both simultaneously. In our tests, the highest attribution accuracy reached by this variant is $89\%$ , with a Clip-score of 0.68 and FID of 63.48 (detailed in Fig. 4). The inherent trade-off observed here further reinforces the challenge of balancing these two critical parameters in the context of model fingerprinting techniques. + +# 4.6. Robust User Attribution against Image Post-processes + +This section evaluates the robustness of our method in scenarios where generated images undergo post-processing. These processes could potentially alter the embedded fingerprint within the images. + +Consistent with methodologies outlined in previous research [5, 16, 22, 32, 33], we examine our model's resilience to various image post-processing operations. We simulate the effect of post-processing at random intensities before inputting data into the fingerprint decoding network, $\mathcal{F}$ : + +$$ +\begin{array}{l} L _ {\text {r o b u s t}} = \mathbb {E} _ {z = \mathcal {E} (x), \phi \sim \Phi} \sum_ {i = 1} ^ {d _ {\phi}} [ \phi_ {i} \log \sigma (\mathcal {F} (T (\mathcal {D} (\phi , z))) _ {i} \\ \left. + \left(1 - \phi_ {i}\right) \log \left(1 - \sigma \left(\mathcal {F} \left(T (\mathcal {D} (\phi , z))\right)\right) _ {i} \right], \quad (5) \right. \\ \end{array} +$$ + +![](images/7848319d6cc20b05cc460cd7b7c0433e3e1d16411d4a5aba0f755eb22835a257.jpg) + +![](images/d8095f54cd666872ea3960f117a907aa8a82d240a2bc5d013d5d4913ddb3f0f6.jpg) + +![](images/38edb2069f0a835330a80ec59409407d8d05890dd5d89985087e71f8f44374cc.jpg) + +![](images/e74b07f35158e93c83bb77396777bb8a17a8ae1f43824742fc0eac6fc636517e.jpg) + +![](images/718c6d942aaac1bfa24d360faff29276214095031d49ea1856a331c8795ff322.jpg) +Figure 5. Enhanced Robustness Against Image Post-Processes. For almost all scenarios, WOUAF consistently exceeds the performance of DAG [16] and Stable Signature [5]. + +![](images/53192df47f6f52d5757238393de518074e78bf82bf537204d8a29119d839ba56.jpg) + +![](images/550be5844220d3111342a5bf94c8bb462e2062122bdcb71e750a28ab4bcdb556.jpg) + +![](images/77d36631f3c4cd4f45fdbfeb210c65e17ca059b162c682f88885bcc6971db789.jpg) + +where, $T(\cdot):\mathbb{R}^{d_x}\to \mathbb{R}^{d_x}$ denotes the post-processing function. In the optimization process, we employ an objective function akin to the one detailed in Eq. (4), with $L_{\phi}$ substituted by $L_{\mathrm{robust}}$ + +In our exploration, we contemplate eight different post-processing techniques: Erasing, Rotation, Gaussian Blurring, Cropping, Brightness jittering, the addition of Gaussian Noise, JPEG compression, and a Combination of all these post-processes. The parameters for these post-processes are designed as follows: For random erasing, we use a random erase ratio within the range $[5\%, 10\%, 15\%, 20\%]$ . Rotation involves randomly sampling a degree within the range (-30, 30). For Gaussian Blurring, we randomly select a kernel size from [3, 5, 7]. For Cropping, we use a random cropping-out ratio within the range $[5\%, 10\%, 15\%, 20\%]$ . The Brightness factor is randomly sampled within the range (-0.3, 0.3). For Gaussian Noise, we add noise with a standard deviation randomly sampled from a uniform distribution $U[0, 0.2]$ . JPEG compression quality level is selected from [90, 80, 70, 60, 50]. The Combination technique randomly selects a subset of these seven post-processing methods with a probability of 0.5. + +User attribution accuracy for each post-process is evaluated under these parameters. Our tests, depicted in Fig.5, offer a comparative analysis of user attribution accuracy across robust versions of DAG [16], Stable Signature [5], and WOUAF. Remarkably, our method demonstrates robustness across a range of post-processes, achieving an attribution accuracy improvement of $11\%$ over Stable Signature and $29\%$ over DAG. A notable trend across all transformations is the monotonic decrease in user attribution accu + +racy as the intensity of post-processing increases. This reinforces the challenges posed by post-processing in maintaining accurate user attribution. However, our results also underscore the benefits of robust training in overcoming these challenges, emphasizing the importance of resilient training strategies for fingerprinting methods in the face of post-processing transformations. Considering the robustness of our method against various post-processes, it becomes a viable choice for model distributors seeking reliable fingerprinting solutions. Detailed results of FID scores and visual examples are available in the Appendix. + +# 5. Deliberate Fingerprint Manipulations + +This section delves into our method's robustness against deliberate attempts to remove fingerprints, which include malicious manipulations via auto-encoders and model purification. Further details and extended attack scenarios are provided in the Appendix. + +# 5.1. Resilience Against Deep Classifier + +The imperceptibility of the fingerprint in generated images is crucial to prevent its detection and subsequent tampering by malicious entities. To assess the secrecy of our method, we adopt an attack scenario akin to the one in [32], assuming an attacker aims to train a classifier to detect the presence of a fingerprint. + +We assume that the attacker seeks to train a classifier capable of detecting the presence of a fingerprint. To assess this scenario, we utilize a pretrained ResNet-50 [6] based binary classifier, trained using 10K SD generated images + +(5K original SD images and 5K fingerprinted SD images). This configuration is deemed valid as detecting the presence of a fingerprint necessitates using both non-fingerprinted and fingerprinted images in the training set. The binary classifier achieve $98\%$ accuracy in the training stage. In subsequent evaluations using a separate set of 5K images from our variant models, the binary classification accuracy is 0.66 for WOUAF-conv and just 0.56 for WOUAF-all, which is nearly equivalent to random chance. + +These findings imply that detecting our embedded fingerprint, particularly in the WOUAF-all variant, poses a challenge to detect. Upcoming subsections will delve into further evaluations, predicated on the stringent assumption that users are cognizant of the fingerprint's presence and endeavor to eliminate it by employing auto-encoder methods or fine-tuning techniques. + +# 5.2. Resilience Against Auto-Encoders + +In contexts where adversaries aim to alter output images, leveraging deep learning techniques such as neural auto-encoders [1, 3, 20] becomes a common strategy for the purpose of obfuscating or removing fingerprints embedded in images [5]. To assess the resilience of our approach, we utilize the robust model against JPEG described in Sec. 4.6. This comparison is appropriate as JPEG represents a conventional image compression method. However, the auto-encoders [1, 3, 20] employed in our evaluation exhibit superior compression performance compared to JPEG. Our research explores the resilience of our proposed method against these sophisticated auto-encoders, focusing particularly on the impact of their varying compression rates. + +As depicted on the left side of Fig. 6, our investigations reveal a notable trend: attribution accuracy progressively declines towards a near-random level (approximately $50\%$ ) as the compression rate employed by the auto-encoders escalates. This trend highlights a critical trade-off: the reduction in attribution accuracy is achievable solely by compromising the quality of the image [18]. Our findings indicate that compromising the integrity of the image is a necessary consequence to effectively obscure the fingerprinting process. + +# 5.3. Resilience Against Model Purification + +This subsection addresses the scenario where an adversary, upon recognizing the presence of fingerprints within the images generated by the image decoder $\mathcal{D}$ , opts to fine-tune $\mathcal{D}$ with the objective of obliterating the embedded fingerprint. This strategy, known as model purification, is a sophisticated approach to altering the model's output to erase traceable imprints [5]. + +In this adversarial setting, the primary aim is to refine the downloaded fingerprinted model by optimizing the reconstruction error between the adversary's proprietary im + +![](images/0ccbb89d3c93e09bc396edd8c5054b2550c96dbae9f717ab2e0da0d5bd0d83e0.jpg) +Figure 6. Left: Auto-Encoder-based Fingerprint Removal. With heightened compression rates, both image quality and attribution accuracy experience a decrease. Right: Model Purification. Progressive fine-tuning leads to concurrent declines in both image quality and attribution accuracy. Note that a lower FID score is preferable, indicating better image quality. + +![](images/2230b67358d1b95f9e022c1ba4474a5a5a10b3b03576790ada6ecc763fa446eb.jpg) + +age dataset and the output from the fingerprinted model. By adhering to the experimental framework outlined in [5], we charted the interplay between FID scores and attribution accuracy, as presented on the right side of Fig. 6. Our empirical analysis reveals a significant challenge: efforts to decrease the attribution accuracy lead to a decline in the quality of the generated images. This finding underscores the inherent complexity in fine-tuning processes aimed at model purification, particularly when striving to maintain the visual quality of the output while endeavoring to obscure its traceable characteristics. + +# 6. Conclusion + +In this study, we have delved into user attribution for Stable Diffusion-based Text-to-Image (T2I) model, employing a weight modulation-based fingerprinting approach. Our method, WOUAF, not only achieves near-perfect accuracy but also preserves the high quality of generated images. A key aspect of WOUAF is its computational efficiency coupled with enhanced robustness against various image post-processing techniques compared to existing baselines. Our results lay a solid groundwork for future exploration into the broader implications and challenges posed by generative models. In future work, we plan to expand and refine our methodology to encompass various data types including text, audio, and video, necessitating tailored adjustments in model fingerprinting techniques. + +# 7. Acknowledgment + +This work is partially supported by the National Science Foundation under Grant No. 2038666, No. 2101052, and a grant from Meta AI Learning Alliance. The HuggingFace demo of this work is funded by Intel Corporation. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers. + +# References + +[1] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 8 +[2] Ali Breland. The bizarre and terrifying case of the "deep-fake" video that helped bring an african nation to the brink. motherjones, 2019. 1 +[3] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8 +[4] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1 +[5] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7 +[7] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5 +[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5 +[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[10] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6 +[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 6 +[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. arXiv preprint arXiv:1912.04958, 2019. 2, 3, 6 +[13] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4 +[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1 +[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 3, 4 +[16] Changhoon Kim, Yi Ren, and Yezhou Yang. Decentralized attribution of generative models. In International Conference on Learning Representations, 2021. 2, 4, 5, 6, 7 +[17] Arijeta Lajka. New ai voice-cloning tools 'add fuel' to misinformation fire. AP News, 2023. 1 +[18] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. ArXiv, abs/2103.09274, 2021. 5, 8 +[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6 +[20] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, 3-8 December 2018, Montréal, Canada, pages 10794–10803, 2018. 8 +[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1 +[22] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2, 6 +[23] Matt Novak. Ai image creator midjourney halts free trials but it has nothing to do with the pope's jacket. forbes, 2023. 1 +[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[25] E. Riba, D. Mishkin, D. Ponsa, E. Rublee, and G. Bradski. Kornia: an open source differentiable computer vision library for pytorch. In Winter Conference on Applications of Computer Vision, 2020. 4 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4, 6 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 + +[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 4, 5 +[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4 +[31] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3 +[32] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2, 3, 4, 5, 6, 7 +[33] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2, 4, 6 +[34] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. 2019. 1, 2, 3 +[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4 +[36] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. ArXiv, abs/2303.10137, 2023. 2 +[37] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 657–672, 2018. 2 \ No newline at end of file diff --git a/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/images.zip b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e8d23c5c4b1276bae3d1e03b2dc256e099c3274 --- /dev/null +++ b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9867783402fcc3e6ee9ae1db43e953601028be8ecac95a7bac84f1306589c7ab +size 506836 diff --git a/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/layout.json b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad18f04f8bcacb78ef9276f22c9dd45c3eb902b4 --- /dev/null +++ b/2024/WOUAF_ Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,8242 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 102, + 514, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 102, + 514, + 138 + ], + "spans": [ + { + "bbox": [ + 81, + 102, + 514, + 138 + ], + "type": "text", + "content": "WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": "Changhoon Kim" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": " Kyle Min" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": " Maitreya Patel" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": " Sheng Cheng" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": " Yezhou Yang" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": "Arizona State University " + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 160, + 505, + 195 + ], + "type": "text", + "content": "Intel Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 197, + 472, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 197, + 472, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 197, + 472, + 209 + ], + "type": "text", + "content": "{kch,maitreya.patel,scheng53,yz.yang}@asu.edu kyle.min@intel.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 236, + 192, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 248 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 248 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "text", + "content": "The rapid advancement of generative models, facilitating the creation of hyper-realistic images from textual descriptions, has concurrently escalated critical societal concerns such as misinformation. Although providing some mitigation, traditional fingerprinting mechanisms fall short in attributing responsibility for the malicious use of synthetic images. This paper introduces a novel approach to model fingerprinting that assigns responsibility for the generated images, thereby serving as a potential countermeasure to model misuse. Our method modifies generative models based on each user's unique digital fingerprint, imprinting a unique identifier onto the resultant content that can be traced back to the user. This approach, incorporating fine-tuning into Text-to-Image (T2I) tasks using the Stable Diffusion Model, demonstrates near-perfect attribution accuracy with a minimal impact on output quality. Through extensive evaluation, we show that our method outperforms baseline methods with an average improvement of " + }, + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "text", + "content": " in handling image post-processes. Our method presents a promising and novel avenue for accountable model distribution and responsible use. Our code is available in https://github.com/kylemin/WOUAF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 541, + 128, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 128, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 128, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 562, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 656 + ], + "type": "text", + "content": "Recent advancements in generative models have propelled their proficiency, expanding their repertoire to include not just the generation of photorealistic images [4, 14] but also the synthesis of images from textual prompts [21, 24, 26, 28]. These significant strides have equipped individuals with the capacity to leverage these models to create hyperrealistic images that correspond seamlessly with given textual instructions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 658, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 694 + ], + "type": "text", + "content": "Nonetheless, the escalating prominence of generative models instigates pressing societal apprehensions. A case in point is Deepfake, intentionally crafted to disseminate" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 237, + 522, + 418 + ], + "blocks": [ + { + "bbox": [ + 331, + 237, + 522, + 418 + ], + "lines": [ + { + "bbox": [ + 331, + 237, + 522, + 418 + ], + "spans": [ + { + "bbox": [ + 331, + 237, + 522, + 418 + ], + "type": "image", + "image_path": "dcc97cf11cc5d6d310a8a6556f801e8e70d24f1e1ca82fdc1f79fdbbe32c9294.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "type": "text", + "content": "Figure 1. Illustration of user attribution based on our method. Please refer to the main text for detailed descriptions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "content": "misinformation, fostering a climate of fake news and political disarray [2, 17, 23]. The gravity of these concerns necessitates calls for governmental intervention to regulate the indiscriminate application of generative models1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 507, + 546, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 546, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 546, + 638 + ], + "type": "text", + "content": "A feasible method to counteract malicious use involves assigning accountability for generated images. One approach to achieve this is by integrating independent fingerprinting modules that can embed user-specific information on top of image generation. The open-source Text-to-Image (T2I) project Stable Diffusion (SD) [26] currently employs this technique using discrete wavelet transform or RivaGAN [34]. However, in the open-source setting, bypassing the fingerprinting module is straightforward and can be achieved by commeting just a single line in the source code [5]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 639, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 686 + ], + "type": "text", + "content": "Is it feasible to achieve user attribution without an independent fingerprinting module? In response, we propose a distributor-oriented methodology named WOUAF, standing for Weight mOdupulation for User Attribution and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 215, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 215, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 215, + 712 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "1President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence. The White House" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "type": "text", + "content": "8974" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "Fingerprinting. In practical terms, when a model inventor opens-sources their work to a model distributor such as Huggingface, the distributor could utilize our proposed method to create a generic version. Upon receiving a download request from an end-user, the distributor can adjust the model weights using our technique and deploy a fingerprinted version to the user, simultaneously registering the user's fingerprint into their database. In the event of a model's malicious exploitation, the distributor can decode the fingerprint from the misused image and cross-reference it with their database to identify the responsible user. Consequently, this provides the distributor with an actionable method to counteract malicious uses of the model (see Fig. 1 for a comprehensive framework of our methodology)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 456 + ], + "type": "text", + "content": "Our methodology, designed for T2I tasks, is integrated into the Stable Diffusion (SD) framework without necessitating any structural changes to the model. This design choice effectively prevents end-users from bypassing the fingerprinting process. Consistent with prior research [5, 16, 22, 32, 33], our primary goal is to maintain high attribution accuracy while ensuring minimal impact on output quality, as elaborated in Sec. 3. Our rigorous evaluations of this method concentrate initially on assessing both attribution accuracy and image quality. We have found that our approach attains nearly flawless attribution accuracy with only a slight influence on image quality. Moreover, we evaluate the robustness of our method in various scenarios involving post-processing manipulations that images might undergo. Our method outperforms baseline methods in these robustness tests, showing an average improvement of " + }, + { + "bbox": [ + 46, + 240, + 289, + 456 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 240, + 289, + 456 + ], + "type": "text", + "content": " in handling such manipulations (refer to Sec. 4 for further details)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 456, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 613 + ], + "type": "text", + "content": "There are four main contributions: (1) We introduce WOUAF, a distinctive distributor-centered fine-tuning methodology. This approach embeds fingerprints within the model in such a way that end-users cannot easily circumvent or remove them. (2) Our method successfully achieves high attribution accuracy, while maintaining the quality of the output images. (3) Our approach exhibits marked resilience against a diverse array of image post-processes, a vital attribute for practical applications. (4) We conduct thorough assessments to balance attribution accuracy with manipulations to intentionally remove fingerprints, including strategies like image compression via auto-encoders and obfuscation through model fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 624, + 134, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 134, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 134, + 637 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 645, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 287, + 682 + ], + "type": "text", + "content": "In this section, we discuss related works of model fingerprinting in generative models. More related works are available in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Inventor-oriented Model Fingerprinting. Yu et al. [33] leveraged a pre-trained deep steganography model to em" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 313 + ], + "type": "text", + "content": "bed fingerprints into the training set for fingerprinted GANs. However, this approach suffers from limited scalability, as it necessitates training a GAN from scratch for each distinct fingerprint. To address this issue, Yu et al. [32] introduced a weight modulation method [12] that directly embeds a user's fingerprint into the generator's weights. Despite these advancements, current methods are predominantly tailored for GAN-based models and typically require training from scratch. This raises important questions regarding their suitability for diffusion-based models, which have a different structural makeup compared to GANs, and the feasibility of avoiding the requirement for training from scratch. The adoption of fine-tuning as a method for embedding fingerprints presents a promising solution. It facilitates the incorporation of fingerprints into pre-trained diffusion models, eliminating the necessity for comprehensive retraining from the ground up [5, 36]. This approach significantly streamlines the process, allowing model inventors to concentrate on core model development without the complexities of embedding fingerprints during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 321, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 561 + ], + "type": "text", + "content": "Distributor-oriented Model Fingerprinting. Kim et al. [16] proposed a technique for achieving user attribution by explicitly incorporating user-specific fingerprints into the generator's output. While this simplified attribution method allowed for the derivation of sufficient fingerprint conditions, it necessitates a trade-off between the quality of the generated output and attribution accuracy, which is further exacerbated when image post-processes are taken into account. To tackle this issue, an approach has been proposed that utilizes subtle semantic variations along latent dimensions as fingerprints, generated by perturbations of eigenvectors in the latent distribution [22]. This method demonstrates an improved balance between generation quality and attribution accuracy. However, its applicability is restricted to unconditional image generation, as eigenvectors are computed by sampling the learned latent representation. In the context of conditional image generation, estimating eigenvectors of latent representation becomes challenging due to the vast space of conditions, such as those found in text conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Recent Advances in Fingerprinting for Text-to-Image Diffusion Models. Recent studies [5, 31, 36] have scrutinized fingerprinting techniques in the Stable Diffusion model [26], uncovering vulnerabilities in existing methods [34] that facilitate easy circumvention [5] or robust post-hoc fingerprinting [31]. Fernandez et al. [5] achieved near-perfect attribution accuracy by fine-tuning user-specific models to align with steganography module [37], demonstrating a viable alternative to conventional post-hoc fingerprinting modules [34]. However, this approach scales linearly in computational demand with the number of users since it necessitates fixed-time fine-tuning" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8975" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 309, + 203 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 309, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 309, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 309, + 203 + ], + "type": "image", + "image_path": "13be24e73fdd824db63437548662696957246b1c8bd2a9a3852cabdcb8418f54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 209, + 233, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 209, + 233, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 209, + 233, + 222 + ], + "type": "text", + "content": "(a) The overall pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 541, + 200 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 541, + 200 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 541, + 200 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 541, + 200 + ], + "type": "image", + "image_path": "1118a0670691935fe7c445f3c5a6ab745c029dedda7eec74a9a5cb962ef28fb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 209, + 478, + 221 + ], + "lines": [ + { + "bbox": [ + 383, + 209, + 478, + 221 + ], + "spans": [ + { + "bbox": [ + 383, + 209, + 478, + 221 + ], + "type": "text", + "content": "(b) Weight modulation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 230, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 230, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 547, + 264 + ], + "type": "text", + "content": "Figure 2. Depiction of our method's pipeline and weight modulation: (a) The model fingerprinting procedure encompasses encoding via the mapping network and weight modulation, along with decoding through the fingerprint decoding network. (b) Weight modulation of the decoding network " + }, + { + "bbox": [ + 46, + 230, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 230, + 547, + 264 + ], + "type": "text", + "content": " to incorporate the fingerprint." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 286, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 286, + 344 + ], + "type": "text", + "content": "for each individual. In contrast, our method requires only a one-time training followed by a negligible forward pass time to generate user-specific models. Furthermore, our approach shows superior robustness against common image post-processing techniques compared to that of Stable Signature [5] (refer to Section Sec. 4.6 for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 346, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 286, + 453 + ], + "type": "text", + "content": "Another notable contribution is by Wen et al. [31], who introduced an alternative fingerprinting method for the Stable Diffusion model [26]. Their method, similar to posthoc style fingerprinting [34], depends on user-driven embedding, which allows end-users the option to exclude the fingerprint. Moreover, it is confined to the DDIM scheduler [30]. Our method, in contrast, is adaptable to both the DDIM [30] and Euler schedulers [15], underscoring its versatility and wider applicability (refer to the Appendix)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 464, + 107, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 107, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 107, + 475 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 483, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 286, + 567 + ], + "type": "text", + "content": "This section outlines our approach, beginning with an overview of the Text-to-Image (T2I) diffusion model with a focus on the Stable Diffusion (SD) model [26] detailed in Sec. 3.1. We then introduce our key component, the user-specific weight modulation, in Sec. 3.2. The section concludes with a detailed explanation of our training objectives and methods, outlined in Sec. 3.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 575, + 133, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 133, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 133, + 586 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Our approach utilizes the Stable Diffusion (SD) model, which functions within the latent space framework of an autoencoder. SD comprises two main elements: Firstly, an autoencoder is pre-trained on an extensive dataset of images. Its encoder, " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", converts an image " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "x\\sim p_{data}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ". The decoder, " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot):\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_x}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", then reconstructs the original image from this latent representation, resulting in " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\mathcal{D}(z)" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ". The secondary element is a diffusion model, based on the U-Net architecture [27], represented as " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ". This model is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 274, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 298 + ], + "type": "text", + "content": "adept at generating latent representations and can be conditioned using pre-trained text embeddings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 304, + 481, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 481, + 317 + ], + "type": "text", + "content": "3.2. User-specific Weight Modulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "content": "Our method is fundamentally based on integrating fingerprints into the parameters of the SD through weight modulation [12, 32]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": "The overall pipeline of our method is illustrated in Fig. 2(a). A user-specific fingerprint is drawn from a Bernoulli distribution with a probability of 0.5, represented as " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\phi \\in \\Phi \\coloneqq \\operatorname{Bernoulli}(0.5)^{d_{\\phi}}" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": " signifies the fingerprint length in bits. We employ a mapping network " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot): \\mathbb{R}^{d_{\\phi}} \\to \\mathbb{R}^{d_M}" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": " to convert the sampled fingerprint " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": " into an intermediate fingerprint representation within the " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "d_M" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": " dimension. For modulating each layer in the SD component, we introduce an affine transformation layer, " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_l(\\cdot): \\mathbb{R}^{d_M} \\to \\mathbb{R}^{d_j}" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": ", for all layers " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": ". As depicted in Fig. 2(b), this transformation matches the dimensions between " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "d_M" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": "-th channel in weight " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{i,j,k}" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "i,j,k" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": " denote input, output, and kernel dimensions, respectively. The weight modulation for the " + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 358, + 546, + 526 + ], + "type": "text", + "content": "-th layer is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 533, + 545, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 533, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 381, + 533, + 545, + 549 + ], + "type": "interline_equation", + "content": "W _ {i, j, k} ^ {\\phi} = u _ {j} * W _ {i, j, k}, \\tag {1}", + "image_path": "bc03e049296dfe9458fce7ccbc466e9da0da2217b2a897c049866b93d28322ea.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "inline_equation", + "content": "W^{\\phi}" + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "content": " denote the pre-trained and fingerprinted weights respectively, " + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "inline_equation", + "content": "u_{j} = \\mathcal{A}_{l}(\\mathcal{M}(\\phi))" + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "content": " is the scale of the fingerprint representation corresponding to the " + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "content": "th output channel." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": "We incorporate fingerprints into the SD by applying weight modulation exclusively to the weights in the decoder " + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": ". The rationale for not applying modulation to both the diffusion model " + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": ", an approach that mirrors GAN-based models [32], is explained in Sec. 4.5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 421, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 421, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 421, + 684 + ], + "type": "text", + "content": "3.3. Training Objectives" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Our training architecture comprises two primary objectives. The initial objective is to decode fingerprints from the pro" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "8976" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 112, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Evaluation of Attribution Accuracy and Image Generation Quality. We conducted validation using the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which were excluded from our training phase. Symbols " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " denote preferred higher and lower values, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 112, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 47, + 112, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 112, + 547, + 224 + ], + "type": "table", + "html": "
ModelFingerprinting Time (↓)MS-COCOLAION
Attribution Acc (↑)CLIP-score (↑)FID (↓)Attribution Acc (↑)CLIP-score (↑)FID (↓)
Original SD [26]--0.7324.48-0.5019.67
DAG [16]8.4 hr0.700.7326.540.710.4923.13
Stable Signature [5]< 1 min0.990.7324.550.980.5020.02
WOUAF-conv< 1 sec0.990.7324.430.980.5120.46
WOUAF-all< 1 sec0.990.7324.420.990.5119.91
", + "image_path": "b3177fc4b8cf17757cd731de50a44e76577cf8a0816dd4889521d6491d18a3d5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 237, + 288, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 273 + ], + "type": "text", + "content": "vided images. We train a fingerprint decoding network " + }, + { + "bbox": [ + 46, + 237, + 288, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_\\phi}" + }, + { + "bbox": [ + 46, + 237, + 288, + 273 + ], + "type": "text", + "content": ", which is instantiated by ResNet-50 [6], as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 283, + 287, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 287, + 333 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\phi} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\sum_ {i = 1} ^ {d _ {\\phi}} [ \\phi_ {i} \\log \\sigma (\\mathcal {F} (\\mathcal {D} (\\phi , z)) _ {i} \\\\ \\left. + \\left(1 - \\phi_ {i}\\right) \\log \\left(1 - \\sigma \\left(\\mathcal {F} \\left(\\mathcal {D} (\\phi , z)\\right)\\right) _ {i} \\right], \\right. \\tag {2} \\\\ \\end{array}", + "image_path": "2c69913644057df8deb9767fcd1d00ead4b983a0aef3e0f05a8ab29530ed8278.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": " refers to the sigmoid activation function, constraining the output of " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": " to the range [0, 1]. Thus, this loss function effectively combines binary cross-entropy for all bits of the fingerprint. During training time, fingerprint " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": " is sampled from Bernoulli distribution. However, after training, the model distributor initially samples a user-specific fingerprint " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\phi_{\\alpha}" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": " and subsequently modulates the decoder " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\phi_{\\alpha}" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": ". The user will receive the fingerprinted decoder " + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\phi_{\\alpha}, \\cdot)" + }, + { + "bbox": [ + 46, + 344, + 287, + 452 + ], + "type": "text", + "content": ", which solely permits latent input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "type": "text", + "content": "The secondary objective endeavors to regularize the quality of outputs. Ideally, this regularization inhibits the decoder " + }, + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "type": "text", + "content": " from compromising image quality while minimizing " + }, + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "type": "inline_equation", + "content": "L_{\\phi}" + }, + { + "bbox": [ + 46, + 453, + 288, + 502 + ], + "type": "text", + "content": " in Eq. (2):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 513, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 513, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 86, + 513, + 287, + 529 + ], + "type": "interline_equation", + "content": "L _ {\\text {q u a l i t y}} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\left[ \\ell (x, \\mathcal {D} (\\phi , z)) \\right], \\tag {3}", + "image_path": "9498659e8bbcb5fe7be1c5881466e767752ebef4357591d8058b674e198b772d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "text", + "content": " represents the distance metric between original images and fingerprinted images. For practical applications, we utilize perceptual distance [35] to gauge the perceptual difference between " + }, + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\phi, z)" + }, + { + "bbox": [ + 46, + 538, + 287, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 587, + 260, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 587, + 260, + 600 + ], + "spans": [ + { + "bbox": [ + 59, + 587, + 260, + 600 + ], + "type": "text", + "content": "The final objective function can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 612, + 287, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 612, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 110, + 612, + 287, + 631 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {A}, \\mathcal {M}, \\mathcal {D}, \\mathcal {F}} \\lambda_ {1} L _ {\\phi} + \\lambda_ {2} L _ {\\text {q u a l i t y}}, \\tag {4}", + "image_path": "0b93b24599172defbd4e307b2cd01f331b005317350205291f32b5bfd796cc1a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": "where both " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " are set to 1.0. Fundamentally, the loss function aspires to reconstruct fingerprints while maintaining the quality of the generated outputs. To assess the efficacy of our proposed method, we employ attribution accuracy and image quality metrics (Refer to Sec. 4.1 for details)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 236, + 388, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 388, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 388, + 250 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 256, + 424, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 424, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 424, + 270 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "text", + "content": "Datasets. Our approach is fine-tuned on the MS-COCO [19] dataset, adopting the Karpathy split. For methodological evaluation, we harness the test set from MS-COCO and randomly sample from the LAION-aesthetics [29] dataset. For T2I image generation, we adopt the Euler scheduler [15] with timestep " + }, + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "text", + "content": ", and the classifier-free guidance scale [9] is set to 7.5 unless otherwise specified. Evaluation for DDIM scheduler [30] and various image generation hyperparameters are available in the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 400, + 546, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 546, + 557 + ], + "type": "text", + "content": "Experimental Setting. We implement the weight modulation following the design specified in the source code of StyleGAN2-ADA [13]. Our mapping network " + }, + { + "bbox": [ + 304, + 400, + 546, + 557 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 400, + 546, + 557 + ], + "type": "text", + "content": " is designed with a series of fully connected layers, wherein all experiments are conducted using a two-layer configuration. To train robust models against image post-processing transformations, differentiable post-processes are necessary. To this end, we incorporate the Kornia library [25]. For Stable Signature [5], we utilize the official code provided by the authors. We note that its post-processing transformations are replaced with our version for fair comparison. Appendix includes details on mapping network dimensions, training parameters, and optimizer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": "**Evaluations.** User attribution accuracy is gauged by the formula: " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{d_{\\phi}} \\sum_{i=1}^{d_{\\phi}} \\mathbb{1}(\\phi_i = \\hat{\\phi}_i)" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": " is the true fingerprint and " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\phi} = \\mathbb{1}\\left[\\sigma(\\mathcal{F}(x_{\\phi})) > 0.5\\right]" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": " is the estimated fingerprint from image " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{\\phi}" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": ". Unless otherwise stated, " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": " is set to 32 in our experiments (Refer to Sec. 4.2 for additional information). We further employ a statistical test [32, 33] to evaluate matching bits between " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": ". The null hypothesis " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": " suggests that the number of matching bits arises by chance. The test uses a binomial distribution, with a " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": "-value derived as: " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "P(X \\geq k|H_0) = \\sum_{i=k}^{d_{\\phi}} \\binom{d_{\\phi}}{i} 0.5^{d_{\\phi}}" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": ". A " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": "-value below 0.05 leads to the rejection of " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "inline_equation", + "content": "1 - p" + }, + { + "bbox": [ + 304, + 562, + 547, + 714 + ], + "type": "text", + "content": " serving as an indicator of verification confidence. Lastly, to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "type": "text", + "content": "8977" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 294, + 264 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 294, + 264 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 294, + 264 + ], + "type": "image", + "image_path": "81b5af490550c7849846a48c1d9a2b3989c323c3e67c7f8c1823de3b1950f898.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 266, + 234, + 278 + ], + "lines": [ + { + "bbox": [ + 174, + 266, + 234, + 278 + ], + "spans": [ + { + "bbox": [ + 174, + 266, + 234, + 278 + ], + "type": "text", + "content": "(a) MS COCO" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 545, + 264 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 545, + 264 + ], + "type": "image", + "image_path": "414dc7ecef5c313ac54f68e4af6d44b95bec8eec265e1b21765a60fcacfcb7bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 266, + 494, + 277 + ], + "lines": [ + { + "bbox": [ + 404, + 266, + 494, + 277 + ], + "spans": [ + { + "bbox": [ + 404, + 266, + 494, + 277 + ], + "type": "text", + "content": "(b) LAION Aesthetics" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 286, + 547, + 320 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 320 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of the original and fingerprinted Stable Diffusion models on MS-COCO [19] and LAION aesthetics [29] (Pixel-wise differences " + }, + { + "bbox": [ + 46, + 286, + 547, + 320 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 286, + 547, + 320 + ], + "type": "text", + "content": " 5: they are multiplied by a factor of 5 for better view). We can observe that our method maintains high image quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 330, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 390 + ], + "type": "text", + "content": "validate the quality of our method, we assess image quality using the Fréchet Inception Distance (FID) [8] and employ the Clip-score [7] to determine the alignment between text and generated images. Additional experimental details can be found in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "text", + "content": "Models. For evaluating our methodology, we benchmark against two established baseline methods: DAG [16] and Stable Signature [5]. Both these methods, conceptualized from the model distributor's standpoint, incorporate finetuning for model fingerprinting. To ensure a fair comparison, we retrain the Stable Signature method within our training settings by replacing its post-processing scheme. Additionally, we evaluate our method against three distinct variants based on the specific layers chosen for weight modulation implementation. The first variant, WOUAF-conv, applies modulation to only the convolutional layers in " + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "text", + "content": ". In contrast, WOUAF-all extends this approach across all layers of " + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "text", + "content": ", covering both self-attention and convolution layers. The final variant implements weight modulation in both the diffusion model " + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 397, + 288, + 613 + ], + "type": "text", + "content": ", mirroring the approach used in GAN-based methods [32]. Further details on why this variant is not used in our experiments are discussed in Sec. 4.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 622, + 169, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 169, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 169, + 635 + ], + "type": "text", + "content": "4.2. Fingerprint Capacity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "The capacity of our method depends on the maximum number of unique user-specific fingerprints it can support without significant crosstalk. This capacity is primarily influenced by the fingerprint dimension " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "(d_{\\phi})" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ". Selecting an optimal " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " presents a challenge: while a larger " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " can accommodate more users, it also complicates effective fingerprint" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 359, + 526, + 426 + ], + "blocks": [ + { + "bbox": [ + 305, + 329, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 329, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 545, + 350 + ], + "type": "text", + "content": "Table 2. Experiments of attribution accuracy across various fingerprint dimensions " + }, + { + "bbox": [ + 305, + 329, + 545, + 350 + ], + "type": "inline_equation", + "content": "(d_{\\phi})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 359, + 526, + 426 + ], + "lines": [ + { + "bbox": [ + 321, + 359, + 526, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 359, + 526, + 426 + ], + "type": "table", + "html": "
Fingerprint Dims.Attribution Accuracy
163264128
WOUAF-conv0.990.990.980.94
WOUAF-all0.990.990.990.97
", + "image_path": "f6ba285f34e2143b648010dc29f8e6f7883e6a707254c867f0851e2a18346c2d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 441, + 367, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 367, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 367, + 453 + ], + "type": "text", + "content": "decoding [18]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": "To investigate this trade-off, we conduct an analysis with varying fingerprint dimensions, specifically " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": " values of 16, 32, 64, and 128. Tab. 2 presents the user attribution accuracy for each " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": " value. As shown in Tab. 2, attribution accuracy tends to decrease monotonically as " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": " increases. Importantly, both our variant models achieve a near-perfect attribution accuracy of 0.99 for " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "d_{\\phi}" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": " values of 16, 32, and 64. However, for " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "d_{\\phi} = 128" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": ", WOUAF-all variant outperforms the WOUAF-conv variant. For a balanced comparison with existing methods, we choose " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "d_{\\phi} = 32" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": ", which notably can support a substantial user base exceeding 4 billion " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\approx 2^{32}" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 597, + 520, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 520, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 520, + 609 + ], + "type": "text", + "content": "4.3. Attribution Accuracy and Image Quality" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 616, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 687 + ], + "type": "text", + "content": "We conduct a comprehensive evaluation of WOUAF, focusing on attribution accuracy and image quality. The assessment involves the MS-COCO [19] test set and the LAION-Aesthetics [29] dataset, which are excluded from the training phase. The results, detailed in Tab. 1, showcase the efficacy of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our variants, namely WOUAF-conv and WOUAF-all, demonstrate superior performance in attribution accuracy" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "8978" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "over DAG [16], indicating their proficiency in accurately decoding embedded fingerprints from the generated images. These variants also show competitive results when compared to Stable Signature [5], reinforcing our methodology's robustness. Notably, we achieve this high level of accuracy without significantly compromising image quality. Both FID scores and Clip-scores showed minimal variation from the baseline SD model, indicating that our approach has a negligible impact on image output quality. This is further corroborated by qualitative examples in Fig. 3, which highlight WOUAF's ability to reliably incorporate fingerprinting without degrading image generation quality. For additional insights, uncurated image collections are provided in the Appendix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 312 + ], + "type": "text", + "content": "Given the growing importance of T2I models, computation time for fingerprinting emerges as a key metric. Our method stands out in computational efficiency. It contrasts with approaches like Stable Signature that need fine-tuning for each new fingerprint. Our method requires just a single forward pass, markedly reducing computational overhead." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 319, + 287, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 333 + ], + "type": "text", + "content": "4.4. Attribution Analysis for Diverse Image Sources" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": "Investigating the attribution of generated images to responsible users, we explore the potential for images from non-fingerprinted or varied sources to bypass our system. Our analysis aims to determine if decoded fingerprints from such images match any entries in the model distributor's database. A mismatch indicates the image's external origin, absolving users in the database." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 288, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 530 + ], + "type": "text", + "content": "We adopt the experimental setup from [32], compiling a dataset with different image types: authentic images from the MS-COCO test set [19], non-fingerprinted images from Stable Diffusion [26], and synthesized images from ProGAN [10], StyleGAN [11], and StyleGAN2 [12], with each category containing 1,000 samples. Given our extensive user database of 1 million entries, we set a threshold at " + }, + { + "bbox": [ + 46, + 423, + 288, + 530 + ], + "type": "inline_equation", + "content": "32 * 0.95 \\approx 30" + }, + { + "bbox": [ + 46, + 423, + 288, + 530 + ], + "type": "text", + "content": " bits, aligning with our 0.99 attribution accuracy as shown in Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 531, + 288, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 603 + ], + "type": "text", + "content": "Our rigorous experiments revealed that, irrespective of the source, no images were incorrectly attributed as possessing a fingerprint from our 1 million fingerprint database. This reinforces the reliability of our attribution approach as detailed in Sec. 4.3 demonstrating the robustness of our system against diverse image sources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 610, + 237, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 237, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 237, + 624 + ], + "type": "text", + "content": "4.5. Benefits of Finetuning only Decoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "When developing our last variant that incorporates weight modulation into both the diffusion model " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ", we note that the resultant pipeline demonstrates similarities with the GAN-based method [32]. A direct comparison between ours and the GAN-based methods may not be entirely straightforward, given the fundamental differences in their training methodologies. This is because the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 545, + 242 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 545, + 242 + ], + "type": "image", + "image_path": "2c1ba3c77858c1dcf5632e7f38a2ef68307c48119e24f1286bc3936d6d630958.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "text", + "content": "Figure 4. Comparative analysis of weight modulation on decoder " + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "text", + "content": " and diffusion model " + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "text", + "content": " with decoder " + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 251, + 547, + 285 + ], + "type": "text", + "content": ". Modulating the diffusion model negatively affects image quality." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "content": "GAN-based methods entail training from scratch, whereas our proposed approach leans towards fine-tuning. Nevertheless, both methodologies share a common mechanism: they aim to modulate the weights of the layers instrumental in learning the latent space. The shared characteristic underscores the fundamental objective of optimizing the balance between attribution accuracy and generation quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 546, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 501 + ], + "type": "text", + "content": "However, our empirical observations suggest that this variant does not consistently achieve commendable performance as an attribution model. Specifically, it appears that this variant can only optimize either attribution accuracy or generation quality, but not both simultaneously. In our tests, the highest attribution accuracy reached by this variant is " + }, + { + "bbox": [ + 304, + 380, + 546, + 501 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 304, + 380, + 546, + 501 + ], + "type": "text", + "content": ", with a Clip-score of 0.68 and FID of 63.48 (detailed in Fig. 4). The inherent trade-off observed here further reinforces the challenge of balancing these two critical parameters in the context of model fingerprinting techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "type": "text", + "content": "4.6. Robust User Attribution against Image Post-processes" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 536, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 584 + ], + "type": "text", + "content": "This section evaluates the robustness of our method in scenarios where generated images undergo post-processing. These processes could potentially alter the embedded fingerprint within the images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 585, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 655 + ], + "type": "text", + "content": "Consistent with methodologies outlined in previous research [5, 16, 22, 32, 33], we examine our model's resilience to various image post-processing operations. We simulate the effect of post-processing at random intensities before inputting data into the fingerprint decoding network, " + }, + { + "bbox": [ + 304, + 585, + 545, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 585, + 545, + 655 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 662, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 662, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 662, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {r o b u s t}} = \\mathbb {E} _ {z = \\mathcal {E} (x), \\phi \\sim \\Phi} \\sum_ {i = 1} ^ {d _ {\\phi}} [ \\phi_ {i} \\log \\sigma (\\mathcal {F} (T (\\mathcal {D} (\\phi , z))) _ {i} \\\\ \\left. + \\left(1 - \\phi_ {i}\\right) \\log \\left(1 - \\sigma \\left(\\mathcal {F} \\left(T (\\mathcal {D} (\\phi , z))\\right)\\right) _ {i} \\right], \\quad (5) \\right. \\\\ \\end{array}", + "image_path": "37e127e53e29217273ff04b3af2e57c9732930737069f052e03d071c9cc9846c.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 315, + 757 + ], + "type": "text", + "content": "8979" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 183, + 178 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 183, + 178 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 183, + 178 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 183, + 178 + ], + "type": "image", + "image_path": "7848319d6cc20b05cc460cd7b7c0433e3e1d16411d4a5aba0f755eb22835a257.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 185, + 72, + 303, + 178 + ], + "blocks": [ + { + "bbox": [ + 185, + 72, + 303, + 178 + ], + "lines": [ + { + "bbox": [ + 185, + 72, + 303, + 178 + ], + "spans": [ + { + "bbox": [ + 185, + 72, + 303, + 178 + ], + "type": "image", + "image_path": "d8095f54cd666872ea3960f117a907aa8a82d240a2bc5d013d5d4913ddb3f0f6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 421, + 178 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 421, + 178 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 421, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 421, + 178 + ], + "type": "image", + "image_path": "38edb2069f0a835330a80ec59409407d8d05890dd5d89985087e71f8f44374cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 541, + 180 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 541, + 180 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 541, + 180 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 541, + 180 + ], + "type": "image", + "image_path": "e74b07f35158e93c83bb77396777bb8a17a8ae1f43824742fc0eac6fc636517e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 188, + 183, + 296 + ], + "blocks": [ + { + "bbox": [ + 51, + 188, + 183, + 296 + ], + "lines": [ + { + "bbox": [ + 51, + 188, + 183, + 296 + ], + "spans": [ + { + "bbox": [ + 51, + 188, + 183, + 296 + ], + "type": "image", + "image_path": "718c6d942aaac1bfa24d360faff29276214095031d49ea1856a331c8795ff322.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 546, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 546, + 329 + ], + "type": "text", + "content": "Figure 5. Enhanced Robustness Against Image Post-Processes. For almost all scenarios, WOUAF consistently exceeds the performance of DAG [16] and Stable Signature [5]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 188, + 303, + 295 + ], + "blocks": [ + { + "bbox": [ + 186, + 188, + 303, + 295 + ], + "lines": [ + { + "bbox": [ + 186, + 188, + 303, + 295 + ], + "spans": [ + { + "bbox": [ + 186, + 188, + 303, + 295 + ], + "type": "image", + "image_path": "53192df47f6f52d5757238393de518074e78bf82bf537204d8a29119d839ba56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 188, + 421, + 295 + ], + "blocks": [ + { + "bbox": [ + 304, + 188, + 421, + 295 + ], + "lines": [ + { + "bbox": [ + 304, + 188, + 421, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 421, + 295 + ], + "type": "image", + "image_path": "550be5844220d3111342a5bf94c8bb462e2062122bdcb71e750a28ab4bcdb556.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 188, + 541, + 295 + ], + "blocks": [ + { + "bbox": [ + 425, + 188, + 541, + 295 + ], + "lines": [ + { + "bbox": [ + 425, + 188, + 541, + 295 + ], + "spans": [ + { + "bbox": [ + 425, + 188, + 541, + 295 + ], + "type": "image", + "image_path": "77d36631f3c4cd4f45fdbfeb210c65e17ca059b162c682f88885bcc6971db789.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "inline_equation", + "content": "T(\\cdot):\\mathbb{R}^{d_x}\\to \\mathbb{R}^{d_x}" + }, + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "text", + "content": " denotes the post-processing function. In the optimization process, we employ an objective function akin to the one detailed in Eq. (4), with " + }, + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "inline_equation", + "content": "L_{\\phi}" + }, + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "text", + "content": " substituted by " + }, + { + "bbox": [ + 46, + 339, + 286, + 388 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{robust}}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "text", + "content": "In our exploration, we contemplate eight different post-processing techniques: Erasing, Rotation, Gaussian Blurring, Cropping, Brightness jittering, the addition of Gaussian Noise, JPEG compression, and a Combination of all these post-processes. The parameters for these post-processes are designed as follows: For random erasing, we use a random erase ratio within the range " + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "inline_equation", + "content": "[5\\%, 10\\%, 15\\%, 20\\%]" + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "text", + "content": ". Rotation involves randomly sampling a degree within the range (-30, 30). For Gaussian Blurring, we randomly select a kernel size from [3, 5, 7]. For Cropping, we use a random cropping-out ratio within the range " + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "inline_equation", + "content": "[5\\%, 10\\%, 15\\%, 20\\%]" + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "text", + "content": ". The Brightness factor is randomly sampled within the range (-0.3, 0.3). For Gaussian Noise, we add noise with a standard deviation randomly sampled from a uniform distribution " + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "inline_equation", + "content": "U[0, 0.2]" + }, + { + "bbox": [ + 46, + 388, + 286, + 604 + ], + "type": "text", + "content": ". JPEG compression quality level is selected from [90, 80, 70, 60, 50]. The Combination technique randomly selects a subset of these seven post-processing methods with a probability of 0.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "content": "User attribution accuracy for each post-process is evaluated under these parameters. Our tests, depicted in Fig.5, offer a comparative analysis of user attribution accuracy across robust versions of DAG [16], Stable Signature [5], and WOUAF. Remarkably, our method demonstrates robustness across a range of post-processes, achieving an attribution accuracy improvement of " + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "content": " over Stable Signature and " + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "content": " over DAG. A notable trend across all transformations is the monotonic decrease in user attribution accu" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 339, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 472 + ], + "type": "text", + "content": "racy as the intensity of post-processing increases. This reinforces the challenges posed by post-processing in maintaining accurate user attribution. However, our results also underscore the benefits of robust training in overcoming these challenges, emphasizing the importance of resilient training strategies for fingerprinting methods in the face of post-processing transformations. Considering the robustness of our method against various post-processes, it becomes a viable choice for model distributors seeking reliable fingerprinting solutions. Detailed results of FID scores and visual examples are available in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 483, + 515, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 515, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 515, + 498 + ], + "type": "text", + "content": "5. Deliberate Fingerprint Manipulations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 504, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 565 + ], + "type": "text", + "content": "This section delves into our method's robustness against deliberate attempts to remove fingerprints, which include malicious manipulations via auto-encoders and model purification. Further details and extended attack scenarios are provided in the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 574, + 487, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 487, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 487, + 586 + ], + "type": "text", + "content": "5.1. Resilience Against Deep Classifier" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 593, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 665 + ], + "type": "text", + "content": "The imperceptibility of the fingerprint in generated images is crucial to prevent its detection and subsequent tampering by malicious entities. To assess the secrecy of our method, we adopt an attack scenario akin to the one in [32], assuming an attacker aims to train a classifier to detect the presence of a fingerprint." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We assume that the attacker seeks to train a classifier capable of detecting the presence of a fingerprint. To assess this scenario, we utilize a pretrained ResNet-50 [6] based binary classifier, trained using 10K SD generated images" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "8980" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": "(5K original SD images and 5K fingerprinted SD images). This configuration is deemed valid as detecting the presence of a fingerprint necessitates using both non-fingerprinted and fingerprinted images in the training set. The binary classifier achieve " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": " accuracy in the training stage. In subsequent evaluations using a separate set of 5K images from our variant models, the binary classification accuracy is 0.66 for WOUAF-conv and just 0.56 for WOUAF-all, which is nearly equivalent to random chance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 286, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 286, + 263 + ], + "type": "text", + "content": "These findings imply that detecting our embedded fingerprint, particularly in the WOUAF-all variant, poses a challenge to detect. Upcoming subsections will delve into further evaluations, predicated on the stringent assumption that users are cognizant of the fingerprint's presence and endeavor to eliminate it by employing auto-encoder methods or fine-tuning techniques." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 272, + 228, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 228, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 228, + 285 + ], + "type": "text", + "content": "5.2. Resilience Against Auto-Encoders" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 286, + 445 + ], + "type": "text", + "content": "In contexts where adversaries aim to alter output images, leveraging deep learning techniques such as neural auto-encoders [1, 3, 20] becomes a common strategy for the purpose of obfuscating or removing fingerprints embedded in images [5]. To assess the resilience of our approach, we utilize the robust model against JPEG described in Sec. 4.6. This comparison is appropriate as JPEG represents a conventional image compression method. However, the auto-encoders [1, 3, 20] employed in our evaluation exhibit superior compression performance compared to JPEG. Our research explores the resilience of our proposed method against these sophisticated auto-encoders, focusing particularly on the impact of their varying compression rates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 447, + 286, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 286, + 564 + ], + "type": "text", + "content": "As depicted on the left side of Fig. 6, our investigations reveal a notable trend: attribution accuracy progressively declines towards a near-random level (approximately " + }, + { + "bbox": [ + 46, + 447, + 286, + 564 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 447, + 286, + 564 + ], + "type": "text", + "content": ") as the compression rate employed by the auto-encoders escalates. This trend highlights a critical trade-off: the reduction in attribution accuracy is achievable solely by compromising the quality of the image [18]. Our findings indicate that compromising the integrity of the image is a necessary consequence to effectively obscure the fingerprinting process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 246, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 246, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 246, + 586 + ], + "type": "text", + "content": "5.3. Resilience Against Model Purification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "type": "text", + "content": "This subsection addresses the scenario where an adversary, upon recognizing the presence of fingerprints within the images generated by the image decoder " + }, + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "type": "text", + "content": ", opts to fine-tune " + }, + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 594, + 286, + 676 + ], + "type": "text", + "content": " with the objective of obliterating the embedded fingerprint. This strategy, known as model purification, is a sophisticated approach to altering the model's output to erase traceable imprints [5]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "In this adversarial setting, the primary aim is to refine the downloaded fingerprinted model by optimizing the reconstruction error between the adversary's proprietary im" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 430, + 166 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 430, + 166 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 430, + 166 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 430, + 166 + ], + "type": "image", + "image_path": "0ccbb89d3c93e09bc396edd8c5054b2550c96dbae9f717ab2e0da0d5bd0d83e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 174, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 240 + ], + "type": "text", + "content": "Figure 6. Left: Auto-Encoder-based Fingerprint Removal. With heightened compression rates, both image quality and attribution accuracy experience a decrease. Right: Model Purification. Progressive fine-tuning leads to concurrent declines in both image quality and attribution accuracy. Note that a lower FID score is preferable, indicating better image quality." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 545, + 166 + ], + "type": "image", + "image_path": "2230b67358d1b95f9e022c1ba4474a5a5a10b3b03576790ada6ecc763fa446eb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 253, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 384 + ], + "type": "text", + "content": "age dataset and the output from the fingerprinted model. By adhering to the experimental framework outlined in [5], we charted the interplay between FID scores and attribution accuracy, as presented on the right side of Fig. 6. Our empirical analysis reveals a significant challenge: efforts to decrease the attribution accuracy lead to a decline in the quality of the generated images. This finding underscores the inherent complexity in fine-tuning processes aimed at model purification, particularly when striving to maintain the visual quality of the output while endeavoring to obscure its traceable characteristics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 400, + 378, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 378, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 378, + 412 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "content": "In this study, we have delved into user attribution for Stable Diffusion-based Text-to-Image (T2I) model, employing a weight modulation-based fingerprinting approach. Our method, WOUAF, not only achieves near-perfect accuracy but also preserves the high quality of generated images. A key aspect of WOUAF is its computational efficiency coupled with enhanced robustness against various image post-processing techniques compared to existing baselines. Our results lay a solid groundwork for future exploration into the broader implications and challenges posed by generative models. In future work, we plan to expand and refine our methodology to encompass various data types including text, audio, and video, necessitating tailored adjustments in model fingerprinting techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 604, + 409, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 409, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 409, + 617 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 624, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 702 + ], + "type": "text", + "content": "This work is partially supported by the National Science Foundation under Grant No. 2038666, No. 2101052, and a grant from Meta AI Learning Alliance. The HuggingFace demo of this work is funded by Intel Corporation. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8981" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "text", + "content": "[2] Ali Breland. The bizarre and terrifying case of the \"deep-fake\" video that helped bring an african nation to the brink. motherjones, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 247 + ], + "type": "text", + "content": "[3] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 248, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 287, + 293 + ], + "type": "text", + "content": "[4] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 338 + ], + "type": "text", + "content": "[5] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 384 + ], + "type": "text", + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "type": "text", + "content": "[7] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "text", + "content": "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 509 + ], + "type": "text", + "content": "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "text", + "content": "[10] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. arXiv preprint arXiv:1912.04958, 2019. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[13] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "type": "text", + "content": "[16] Changhoon Kim, Yi Ren, and Yezhou Yang. Decentralized attribution of generative models. In International Conference on Learning Representations, 2021. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 164, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 186 + ], + "type": "text", + "content": "[17] Arijeta Lajka. New ai voice-cloning tools 'add fuel' to misinformation fire. AP News, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 186, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 220 + ], + "type": "text", + "content": "[18] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. ArXiv, abs/2103.09274, 2021. 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 221, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 221, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 221, + 545, + 287 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 289, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 545, + 354 + ], + "type": "text", + "content": "[20] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, 3-8 December 2018, Montréal, Canada, pages 10794–10803, 2018. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 355, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 421 + ], + "type": "text", + "content": "[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "type": "text", + "content": "[22] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "type": "text", + "content": "[23] Matt Novak. Ai image creator midjourney halts free trials but it has nothing to do with the pope's jacket. forbes, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "text", + "content": "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "text", + "content": "[25] E. Riba, D. Mishkin, D. Ponsa, E. Rublee, and G. Bradski. Kornia: an open source differentiable computer vision library for pytorch. In Winter Conference on Applications of Computer Vision, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 647, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 714 + ], + "type": "text", + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "type": "text", + "content": "8982" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "type": "text", + "content": "[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 207, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 288, + 239 + ], + "type": "text", + "content": "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 283 + ], + "type": "text", + "content": "[31] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 288, + 328 + ], + "type": "text", + "content": "[32] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 384 + ], + "type": "text", + "content": "[33] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 288, + 418 + ], + "type": "text", + "content": "[34] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. 2019. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "text", + "content": "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "type": "text", + "content": "[36] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. ArXiv, abs/2303.10137, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 553 + ], + "type": "text", + "content": "[37] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 657–672, 2018. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "type": "text", + "content": "8983" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_content_list.json b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1c2b3ffebd1fc8173946ff600e28dce4b4b7f30 --- /dev/null +++ b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_content_list.json @@ -0,0 +1,1356 @@ +[ + { + "type": "text", + "text": "WWW: A Unified Framework for Explaining What, Where and Why of Neural Networks by Interpretation of Neuron Concepts", + "text_level": 1, + "bbox": [ + 86, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong Hyun Ahn, $^{1}$ , Hyeon Bae Kim, $^{1}$ , Seong Tae Kim $^{2*}$", + "bbox": [ + 261, + 202, + 712, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Artificial Intelligence, Kyung Hee University, Republic of Korea $^{2}$ Department of Computer Science and Engineering, Kyung Hee University, Republic of Korea", + "bbox": [ + 111, + 220, + 857, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yhahn, hyeonbae.kim, st.kim}@khu.ac.kr", + "bbox": [ + 310, + 258, + 656, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in neural networks have showcased their remarkable capabilities across various domains. Despite these successes, the \"black box\" problem still remains. To address this, we propose a novel framework, WWW, that offers the 'what', 'where', and 'why' of the neural network decisions in human-understandable terms. Specifically, WWW utilizes adaptive selection for concept discovery, employing adaptive cosine similarity and thresholding techniques to effectively explain 'what'. To address the 'where' and 'why', we proposed a novel combination of neuron activation maps (NAMs) with Shapley values, generating localized concept maps and heatmaps for individual inputs. Furthermore, WWW introduces a method for predicting uncertainty, leveraging heatmap similarities to estimate the prediction's reliability. Experimental evaluations of WWW demonstrate superior performance in both quantitative and qualitative metrics, outperforming existing methods in interpretability. WWW provides a unified solution for explaining 'what', 'where', and 'why', introducing a method for localized explanations from global interpretations and offering a plug-and-play solution adaptable to various architectures. Code is available at: https://github.com/ailab-kyunghee/WWW", + "bbox": [ + 73, + 340, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural networks have demonstrated impressive performance in various fields in recent years. Despite these successes, their widespread adoption in more diverse areas is slowed by several challenges. A fundamental issue is the \"black box\" problem, referring to the often hidden and unclear decision-making processes of neural network models. This lack of clarity raises concerns about the stability and reliability of these models, leading to a growing consensus that artificial intelligence should be reliable, robust,", + "bbox": [ + 75, + 741, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/5070b4a805a306447b80efe7e8d43bd62e11e36df8eca54b98cfeee5e4ce4924.jpg", + "table_caption": [ + "Table 1. Illustration of What, Where, and Why of recent concept-based neural network interpretation methods. Green, yellow, and red marks illustrate that the method is able to interpret well, is partially interpretable, and has limitations for interpretation, respectively. Recent methods are able to interpret one or two 'w's but have limitations regarding interpreting three 'w's at once." + ], + "table_footnote": [], + "table_body": "
MethodWhatWhereWhy
CLIP-Dissect [20] (ICLR'23)✓XX
FALCON [14] (ICML'23)✓X▲
CRAFT [7] (CVPR'23)X✓▲
WWW (Ours)✓✓✓
", + "bbox": [ + 504, + 402, + 870, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and safe [4, 13]. As a response to this need for trustworthy AI, there has been an emergence of laws and regulations [15, 18] that require neural networks to base their decisions on principles that are understandable to humans.", + "bbox": [ + 496, + 517, + 892, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "According to Doshi-Velez et al. [4], interpretability is defined as the capability to provide explanations in terms that are understandable to humans. Zhang et al. [29] expand on this definition, emphasizing that interpretability involves providing explanations in understandable terms to humans. This requirement for explanations in human-understandable terms is a consistent theme in the literature on interpretable methods. Furthermore, this concept aligns with a fundamental principle in journalism and problem-solving: the importance of clear and understandable communication. This principle is often encapsulated in the five Ws (Who, What, When, Where, Why) [25]. Building on this, we propose that practical explanations or interpretations should include three critical elements: 'what' (the nature of the decision or outcome), 'where' (the context or specific region of the input that affects the decision-making process), and 'why' (the reasoning or factors behind a decision). This approach aims to make the interpretations more accessible and relevant to human understanding.", + "bbox": [ + 496, + 580, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From this point of view, former works can be re-grouped into several groups which those works aim to solve. For ex-", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10968", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ample, a series of works in feature attribution [10, 23, 28] can be classified as works mainly focusing on 'where'. These methods explain model decisions by identifying highly contributed input regions. However, identifying highly related input regions primarily focuses on 'where' only, not enough explanation for 'what' and 'why'.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, concept-based explanations mainly focus on 'why'. Concept-based explanations [6, 7, 10, 17] explain model decisions by decomposing model responses into smaller units called concepts. Kim et al. proposed TCAV [17], understanding and interpreting model decisions into a form of concept activation vectors. Recently, Fel et al. proposed CRAFT [7], which utilizes non-negative matrix factorization for identifying concept vectors. The works mentioned above mainly decompose activations into the form of vectors, which gives a great advantage for finding out 'why' but often fails to match or annotate the human-understandable term (i.e., name) that is needed for 'what'.", + "bbox": [ + 75, + 184, + 467, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other streams of work, called Neuron-concept association, aim to name what each neuron (e.g. convolution filters, layer outputs) represents, which answers 'what' the model cares about. Bau et al. proposed Network Dissection [3], which identifies each neuron's representing concept from Broaden dataset [3]. Recently, CLIP-Dissect [20] and FALCON [14] have shown that representing the concept of neurons can be matched automatically by using a pre-trained CLIP model [21]. The aforementioned methods are suitable for explaining 'what' but partially explain or have limitations for explaining 'where' and 'why'. Furthermore, these methods generate global explanations focusing on each neuron's role in the network, not the local explanations for each sample input.", + "bbox": [ + 75, + 367, + 467, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel framework that can explain 'What', 'Where', and 'Why' (WWW) at once. WWW introduces adaptive selection for discovering each neuron's concept and interpreting each neuron to explain 'what'. By leveraging adaptive cosine similarity (ACS) and adaptive selection techniques, we achieve advanced performance compared to competitive methods. Moreover, we combined neuron activation map (NAM) and Shapley value [2, 16, 24] to generate class and sample concept maps and heatmaps to explain 'where' and 'why'. We also conducted various objective evaluations to assess the performance of the proposed method. In the experiments, WWW achieves better results in both quantitative and qualitative evaluations. Our key contributions are summarized as follows:", + "bbox": [ + 75, + 580, + 467, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a novel and effective way to generate high-quality explanations that explain 'what', 'where', and 'why' at once. Due to the powerful performance of the adaptive selection for concept discovery, WWW is able to achieve higher quantitative and better qualitative results in various metrics.", + "bbox": [ + 76, + 809, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a way to generate localized explanations from global neural network interpretation. By the novel combination of Shapley value and neuron activation maps, WWW is able to generate localized explanations with concept annotations for sample input.", + "- WWW can be attached to various target models with different architectures, from conventional convolution neural networks to the recent attention-based Vision transformers, in a plug-and-play manner." + ], + "bbox": [ + 500, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 247, + 648, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neuron-Concept Association", + "text_level": 1, + "bbox": [ + 500, + 273, + 754, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bau et al. introduced Network Dissection [3], using the Broden dataset to identify which concepts individual neurons in a network represent. They use overlap between segmentation masks and feature maps to annotate concepts for neurons. Fong and Vedaldi expanded on this with Net2Vec [9], which looks at individual neurons and their combinations. Mu and Andreas further extended these ideas with Compositional Explanation [19], aiming to generate more complex and detailed explanations. These methods primarily focused on understanding 'what' a neuron represents. However, the aforementioned methods fell short in explaining the 'where' and 'why' of neuron representations. Additionally, they relied heavily on image-concept-matched datasets like Broden, which are often costly and hard to collect due to the need for pixel-wise labels. To address these limitations, recent approaches like HINT [27] and MILAN [12] have been developed. These methods train concept classifiers or models that reduce the dependency on image-concept-matched datasets. Moreover, approaches like CLIP-Dissect [20] and FALCON [14] leverage the pretrained CLIP model to use separate sets of image datasets and concept datasets. Despite these advancements in understanding 'what' neurons represent, there remains a gap in fully explaining the 'where' and 'why' of the neuron representations.", + "bbox": [ + 496, + 299, + 890, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vector-based Explanation", + "text_level": 1, + "bbox": [ + 500, + 693, + 733, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The field of interpretability has significantly advanced, particularly in understanding and interpreting model decisions using Concept Activation Vectors (CAVs). Kim et al. introduced TCAV [17], understanding and interpreting model decisions into the form of CAVs. Recently, CRAFT [7] was introduced, leveraging non-negative matrix factorization to identify CAVs and localize the most relevant input regions for each CAV. Also, Achtibat et al. [1] and Fel et al. [6] proposed methods to interpret each CAV's role in decision-making. Despite these developments, a key challenge remains: translating CAVs into human-understandable terms.", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10969", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf758565ed0917fa187b3806594c33d01b812147cbee75e8c2ca58e3bf17fb5f.jpg", + "image_caption": [ + "Figure 1. Overall flow of Concept Discovery module identifying concepts for a single neuron. We first calculate the cosine similarity of CLIP features between images and concepts with the template from the selected high-activating images. Then, we subtract the cosine similarity of CLIP features between images and the base template by only considering the similarity between the concept and image. From calculated adaptive cosine similarity (ACS), we generate concept score $S$ by the average similarity of images. Note that concept score $S = \\{s_1, s_2, \\dots, s_m\\}$ are a group of scores, not a single scalar. From the calculated concept scores $S$ , we select major concepts by adaptive selection. We also discover minor concepts using the same process but with crop images." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 861, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 547, + 169, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Method Overview", + "text_level": 1, + "bbox": [ + 76, + 571, + 251, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WWW consists of three modules: Concept discovery, Localization, and Reasoning. The concept discovery module identifies each neuron's concept, which explains part 'what'. It selects neuron concepts from the concept set $(D_{\\text{concept}})$ by leveraging adaptive cosine similarity and adaptive selection. The localization module is to identify highly contributed input regions of the test sample, which concept is present at 'where'. Also, the combination of the neuron activation map leveraging Shapley value helps identify 'where' the concept is and tells the 'why' of the individual predictions. The reasoning module identifies important neurons of the test sample and predicted class. By comparing the differences between the sample and class explanations, users can understand the 'why' of the model prediction and, even more, whether the prediction is reliable or not.", + "bbox": [ + 75, + 595, + 473, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Concept Discovery Module", + "text_level": 1, + "bbox": [ + 76, + 847, + 321, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The concept discovery module aims to identify proper major and minor concepts from $D_{\\text{concept}}$ that match an", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "example-based representation of each neuron. Let target model $f$ d, and let $(l,i)$ as $i$ -th neuron in layer $l$ of the target model. From the images in probing dataset $D_{probe}$ , we select high activating images for neuron $(l,i)$ as $D_{rep}^{(l,i)}$ . With selected images in $D_{rep}^{(l,i)}$ and concepts in $D_{concept}$ , we calculate CLIP visual feature of $D_{rep}^{(l,i)}$ as $V^{(l,i)} = \\left[v_1^{(l,i)}, v_2^{(l,i)}, \\dots, v_n^{(l,i)}\\right]$ and CLIP text feature of $D_{concept}$ as $T = [t_1, t_2, \\dots, t_m]$ . $n$ denotes number of example images for neuron $(l,i)$ and $m$ stands for the number of concept in concept set $D_{concept}$ .", + "bbox": [ + 496, + 549, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We calculate concept score $s^{(l,i)}$ by calculating adaptive cosine similarity (ACS), which allows us to reduce the effect of the base template and only consider the similarity between the image and the concept itself. Concept score $s_{j}^{(l,i)}$ for $j$ -th concept $t_j$ of neuron $(l,i)$ is calculated as follows:", + "bbox": [ + 496, + 715, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {j} ^ {(l, i)} = \\frac {1}{n} \\sum_ {o = 1} ^ {n} \\left\\{\\cos \\left(v _ {o} ^ {(l, i)}, t _ {j}\\right) - \\cos \\left(v _ {o} ^ {(l, i)}, t _ {t e m}\\right) \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 811, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $1 \\leq j \\leq m$ , $t_{tem}$ denotes CLIP text feature of base template(e.g. 'a photo of') and $\\cos(x, y)$ denotes cosine similarity between $x$ and $y$ .", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10970", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d04ee3b5409b6957433008bf52fbadf8ad0cb6c743557fbf5ee0b869389d07bc.jpg", + "image_caption": [ + "Figure 2. Illustration of overall test time flow of WWW. In the test time (i.e., inference), class explanation selects important neurons with pre-computed Shapley value of the predicted class. On the other hand, the sample explanation selects important neurons with a Shapley value for the input sample. With selected important neurons, pre-computed concepts are annotated. After the concept annotation, a class heatmap is generated with the pre-computed Shapley value of the predicted class. On the other hand, a sample heatmap is generated with the Shapley value of the sample input." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the calculated concept scores $s^{(l,i)} = \\left[s_1^{(l,i)},\\dots ,s_m^{(l,i)}\\right]$ , we select concepts where $s_j^{(l,i)} > \\delta^{(l,i)}$ . Adaptive selection threshold $\\delta^{(l,i)}$ for neuron $(l,i)$ is calculated as follows:", + "bbox": [ + 76, + 554, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {(l, i)} = \\alpha \\times \\max (s ^ {(l, i)}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 636, + 468, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes concept sensitivity for major and minor concepts. For discovering minor concepts, we select $D_{rep}^{(l,i)}$ with cropped images of probing dataset $D_{probe}$ .", + "bbox": [ + 75, + 667, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Localization Module", + "text_level": 1, + "bbox": [ + 76, + 724, + 272, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The localization module aims to identify highly contributed input regions of each concept and generate a concept map and concept heatmap. We select important neurons with Taylor approximation of Shapley value introduced in [16] for generating a concept region map. For input image sample $x$ , Neuron contribution $w^{(l,i)}$ of neuron $(l,i)$ is calculated as follows:", + "bbox": [ + 75, + 748, + 470, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {(l, i)} (x) = \\left| f (x) - f (x; a ^ {(l, i)} \\leftarrow 0) \\right| = \\left| a ^ {(l, i)} \\nabla_ {a ^ {(l, i)}} f (x) \\right|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 862, + 475, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $a^{(l,i)}$ denotes activation of neuron $(l,i)$ . After calculating each layer's Neuron contribution, we rank neurons by calculated contribution (i.e., Shapley value) and select the top- $k$ important neurons for the sample. The concept heatmap $M$ is calculated with the weighted sum of important neurons' neuron activation maps (NAMs). Concept heatmap $M^l(x)$ of important neurons is calculated as follows:", + "bbox": [ + 496, + 555, + 890, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {l} (x) = \\sum^ {u} w ^ {(l, u)} (x) A ^ {(l, u)} (x) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 676, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u$ denotes the index of important neurons, $w$ denotes the neuron contribution, and $A$ denotes the neuron activation map of the neuron. Note that $M^l$ shows only related regions of a single or combination of important neurons, not the whole network.", + "bbox": [ + 496, + 713, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Reasoning Module", + "text_level": 1, + "bbox": [ + 500, + 800, + 679, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reasoning module is designed to help users understand the 'why' of the model output. This not only explains the result but also includes 'why' this prediction is reliable or not. By leveraging the class-wise Shapley value introduced in [2], we can understand each neuron's class-wise contri", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10971", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bution and identify important neurons for each class. From Ahn et al. [2], the class-wise contribution of a neuron can be calculated by the average contribution of a neuron in class samples. From the class-wise contribution, we rank top- $k$ important neurons for each class. We can also generate a class-wise concept region map and concept heatmap with the index of important neurons. Class-wise maps can be used as a guideline for understanding the general case of the prediction. By comparing Class-wise maps and sample maps, users can identify which concept and region of the sample differs from general cases. This naturally helps users understand the 'why' of the model output and also 'why' the results are reliable or not. Figure 2 for further understanding the WWW flow for generating class and sample explanations. Figure 5 shows the generated explanation example.", + "bbox": [ + 76, + 90, + 472, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Overall flow and Output of WWW", + "text_level": 1, + "bbox": [ + 76, + 340, + 379, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described above, WWW has three main modules explaining three 'w' of 'what', 'where', and 'why', respectively. In Section 3.5, we are going to follow the overall flow of our method by time sequence. Before starting inference, WWW needs pre-computing for discovering major and minor concepts with the concept discovery module (Sec 3.2) and also a class-wise contribution for class-wise analysis in the Reasoning Module (Sec 3.4). After pre-computing concepts and contributions, WWW is ready to generate test sample explanations. Figure 2 shows the overall flow of explaining the generation of WWW in the inference time. In the test time (i.e., inference), WWW generates a class-wise explanation of the predicted class and sample explanation into two tracks. WWW leverages the predicted class and class-wise contribution matrix to generate a class-wise explanation to select the most critical neurons for the predicted class. After identifying important neurons, WWW leverages the localization module (Sec 3.3) to generate a class concept region map and class concept combination heatmap. On the other hand, for the sample explanation, WWW calculates the Taylor-approximated Shapley value of the sample for the predicted class and identifies critical neurons for the prediction of the sample. After identifying important neurons for the sample, WWW leverages the sample Shapley value to generate the sample concept region map and sample concept combination heat map.", + "bbox": [ + 76, + 364, + 472, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 768, + 200, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comprehensive experiments have been conducted to evaluate our method. In section 4.1, we evaluate the performance of the concept discovery module both qualitatively (Sec. 4.1.1) and quantitatively (Sec. 4.1.2). Section 4.2 is an ablation study for the concept discovery module. In Section 4.3, we analyze generated explanations of both correct and wrong predictions with examples.", + "bbox": [ + 76, + 794, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/753df40db424fd15b1f841e8e88684930ffa708dcf10ae125ef6c8c63c9fcabd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 109, + 542, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b1833a0838b902a2981c64443502d4cc113aa2959862ad74f7930757081e718a.jpg", + "image_caption": [ + "ResNet - 50 Layer 4", + "Neuron 1152", + "WWW (Ours): goldfish - flame fish", + "CLIP-Dissect: swordtail", + "MILAN: ORANGE objects FALCON: TRICK, CALE", + "FALCON:Insh,sale,series,fare", + "Neuron 816", + "CLIP-Dissect: genus Chaetodon" + ], + "image_footnote": [], + "bbox": [ + 544, + 109, + 575, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b5ce624862242000be52437a649e115e43d3d4fe5ca4ddfde5852d1f553ef214.jpg", + "image_caption": [ + "WWW (Ours): saltwater fish - swordtail", + "MILAN: Aquatic life" + ], + "image_footnote": [], + "bbox": [ + 503, + 152, + 544, + 178 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a722818b2a6fcdedf84eccc43634f97f7b2039eb7cabe1612166beb09f42a729.jpg", + "image_caption": [ + "FALCON: cover, black, white, home, blue", + "ResNet - 50 Final layer", + "Neuron 1", + "Ground Truth: goldfish", + "www", + "CLIP-Dissect: swordtail", + "MILAN: Nature", + "FALCON: trish, size, storage assorted, gap" + ], + "image_footnote": [], + "bbox": [ + 722, + 130, + 794, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f58cb48ab48c3973fe85d6f598d284e057f1b1d629b63eefc161804b9c3fe750.jpg", + "image_caption": [ + "Neuron 2018", + "WWW (Ours): Dutch oven - stewing pan", + "CLIP-Dissect: electric frying pan", + "Neuron 1911", + "CLIP-Dissect: cookstove" + ], + "image_footnote": [], + "bbox": [ + 503, + 195, + 542, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cac259dfc7409f347023906ad41d2053160759e097e650a6dab4ba1ef3d200c9.jpg", + "image_caption": [ + "MILAN: Rounded edges in pictures", + "FALCON: black, card, woman, home, white", + "Neuron 544", + "Ground Truth: Dutch o", + "WWW (Ours): Dutch oven", + "CLIP-Dissect: stewing pan", + "MILAN: Circular objects", + "FACLCON:men top game", + "award, guard", + "WWW (Ours): cookfire - cookfire", + "MILAN: Animals and circular objects", + "FALCON: shirt, school, mini, woman, house", + "Figure 3. Qualitative comparison of WWW with other baselines. We compared WWW with three competitive baselines (CLIP-Dissect [20], MILAN [12], FALCON [14]) in two final layer neurons and four penultimate layer (i.e., layer 4) neurons with each neuron's highly activating images. layer-4 neurons are top-2 important neurons of the final layer class. We have colored the descriptions green if they match the images, yellow if they match but are too generic or similar, and red if they do not match." + ], + "image_footnote": [], + "bbox": [ + 723, + 220, + 797, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Performance Evaluation for Concept Module", + "text_level": 1, + "bbox": [ + 498, + 417, + 880, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this experiment, we evaluate the concept matching performance of the concept discovery module with four other baselines, Network Dissection [3], MILAN [12], CLIP-Dissect [20] and FALCON [14]. We evaluate the performance of methods on the various models (e.g., ResNet-18 [11], ResNet-50 [11], and ViT-B/16 [5]), various probing datasets (e.g.,Imagenet, Places365) and various concept sets (e.g., Wordnet nouns, and labels of Places365, Broaden, and ImageNet)", + "bbox": [ + 496, + 441, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 602, + 686, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. We compared WWW with the three most comparable methods (CLIP-Dissect [20], MILAN [12], FALCON [14]) in the penultimate layer and final layer of the model. We do not compare with Network Dissection [3] due to the limitation that probe image data $D_{probe}$ and concept set $D_{concept}$ is fixed to Broaden. We used a ResNet-50 [11] model pre-trained in the ImageNet-1k [22] dataset. For probe image data $D_{probe}$ , we used the ImageNet-1k validation set, and we extracted all nouns in Wordnet[8] (about 80k) dataset for a concept set $D_{concept}$ .", + "bbox": [ + 496, + 627, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. Figure 3 shows examples of descriptions for hidden neurons in the penultimate and final layers. Neurons in the penultimate layer are top-2 important neurons of the final layer neuron's ground truth label class. We observed that WWW not only interpreted each neuron well but also showed robust interpretation that the most important neuron of the class in the penultimate layer represents the same major concept as the final layer neuron.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10972", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c23793173ec8aef1fcf56021743112b26f950bf910c4fe157e31c58e7657083b.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on final layer concept matching performance of ResNet-50 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for $D_{probe}$ . Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported." + ], + "table_footnote": [], + "table_body": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
Network Dissection [3]BrodenBroden(1.2k)0.7229 ± 0.0030.2989 ± 0.0050.0010 ± 0.0010.001
MILAN(b) [12]ImageNet val-0.7300 ± 0.0030.2485 ± 0.0050.0005 ± 0.0000.001
FALCON [14]ImageNet valLAION-400m0.7065 ± 0.0030.1790 ± 0.0010.0002 ± 0.0000.001
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9340 ± 0.0030.8376 ± 0.0060.7286 ± 0.0090.933
ImageNet valBroden (1.2k)0.7369 ± 0.0030.3432 ± 0.0040.0328 ± 0.0030.108
ImageNet valWordnet (80k)0.8689 ± 0.0040.6846 ± 0.0080.3647 ± 0.0140.456
WWW (Ours)ImageNet valImageNet (1k)0.9325 ± 0.0030.8327 ± 0.0060.7719 ± 0.0090.955
ImageNet valBroden (1.2k)0.7758 ± 0.0040.4414 ± 0.0070.0645 ± 0.0070.091
ImageNet valWordnet (80k)0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.0120.645
", + "bbox": [ + 98, + 142, + 870, + 306 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1457275cca632cf28803e1f7df6c56aa7bb8dd7b60009ad5102edfc38dab5a65.jpg", + "table_caption": [ + "Table 3. Quantitative comparison on final layer concept matching performance of ViT-B/16 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for $D_{probe}$ . Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported." + ], + "table_footnote": [], + "table_body": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9337 ± 0.0030.8375 ± 0.0060.7289 ± 0.0090.933
ImageNet valBroden (1.2k)0.7365 ± 0.0030.3416 ± 0.0040.0319 ± 0.0030.106
ImageNet valWordnet (80k)0.8700 ± 0.0040.6886 ± 0.0080.3679 ± 0.0140.460
WWW (Ours)ImageNet valImageNet (1k)0.9331 ± 0.0030.8347 ± 0.0060.7718 ± 0.0090.955
ImageNet valBroden (1.2k)0.7754 ± 0.0040.4407 ± 0.0070.0651 ± 0.0070.091
ImageNet valWordnet (80k)0.8857 ± 0.0030.6970 ± 0.0080.4166 ± 0.0120.634
", + "bbox": [ + 112, + 371, + 854, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 512, + 274, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare the performance of our methods with baselines. As introduced in [20], we evaluate final layer neuron concepts with the class labels with various metrics. By comparing generated explanations with the class labels, we can objectively evaluate the quality of the generated neuron labels with what each neuron is trained to represent.", + "bbox": [ + 75, + 537, + 470, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. CLIP cos and mpnet cos are measured as cosine similarities between the encoded feature of the class label and selected concepts with CLIP [21] model and mpnet [26] model, respectively. F1-score is measured to evaluate the discovered concept's balance of exactness and flexibility. Also, the hit rate is calculated as a rate of selected concepts that exactly match the class label.", + "bbox": [ + 75, + 643, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison on ResNet-50. Table 2 compares WWW with Network Dissection [3], MILAN(b) [12], FALCON [14], and CLIP-Dissect [20]. We evaluate the final layer neuron concepts of ResNet-50 with the class labels of ImageNet-1k. In table 2, WWW showed better performance as the concept set $D_{\\text{concept}}$ gets larger and outperformed all other baselines when concept set $D_{\\text{concept}}$ is set to Broden and Wordnet nouns. In the results of Wordnet nouns ( $D_{\\text{concept}}$ ), comparison between WWW and CLIP-Dissect [20] are statistically significant across all metrics", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$(p < 0.05)$", + "bbox": [ + 500, + 513, + 578, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison on ViT-B/16. We compared predicted concepts of the final layer to ground truth labels of ViT-B/16 pre-trained on ImageNet. In Table 3, WWW showed better performance as the concept set $D_{\\text{concept}}$ gets larger and outperformed other baselines when concept set $D_{\\text{concept}}$ is set to Broden and Wordnet nouns.", + "bbox": [ + 496, + 529, + 893, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison on ResNet-18 pre-trained in Places365. We compared predicted labels to ground truth labels in final layer neurons of ResNet-18 pre-trained on Places365. In table 4, WWW showed better performance as the concept set $D_{\\text{concept}}$ gets larger and outperformed other baselines when concept set $D_{\\text{concept}}$ is set to Broden and Wordnet nouns.", + "bbox": [ + 496, + 621, + 893, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 739, + 650, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on the effect of leveraging base template and ACS. Table 5 shows an ablation study over various parts used in WWW. Without using a base template, WWW showed the lowest performance overall. With the use of a base template, WWW shows slightly increased performance in CLIP cos and F1-score but not much advance overall. However, performance was significantly increased, and the highest scores in all three metrics were achieved when leveraging ACS. The findings suggest that while us-", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10973", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7e25587ff550ae5010ee71add0b48c9e2060a67e3000f5fc14602f2e29da9809.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on final layer concept matching performance of ResNet-18 trained on Places365. We compared predicted neuron concepts with ground truth labels of Places365. We used the Places365 test set for $D_{probe}$ . Bold numbers represent the best scores between the same settings. The average score and standard errors of the 365 final layer neurons are reported." + ], + "table_footnote": [], + "table_body": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]Places365 testPlaces365 (0.4k)0.9562 ± 0.0040.8687 ± 0.0120.7233 ± 0.0230.723
Places365 testBroden (1.2k)0.8304 ± 0.0030.4678 ± 0.0070.1096 ± 0.0080.329
Places365 testWordnet (80k)0.8378 ± 0.0060.5107 ± 0.0140.1041 ± 0.0160.104
WWW (Ours)Places365 testPlaces365 (0.4k)0.9402 ± 0.0040.8204 ± 0.0130.6361 ± 0.0240.660
Places365 testBroden (1.2k)0.8925 ± 0.0050.6415 ± 0.0130.2242 ± 0.0210.255
Places365 testWordnet (80k)0.8492 ± 0.0040.5000 ± 0.0130.1106 ± 0.0150.142
", + "bbox": [ + 104, + 142, + 864, + 260 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7424b59fa03525a8ab5bd6efc26db2d2277dc271b5a9b461f751d903a2028945.jpg", + "table_caption": [ + "Table 5. Ablation study on the use of the template and ACS. '√' in the Template represents WWW using a template (e.g., 'a photo of word') '√' in the ACS represents WWW using adaptive cosine similarity." + ], + "table_footnote": [], + "table_body": "
TemplateACSCLIP cosmpnet cosF1-score
0.8499 ± 0.0030.6123 ± 0.0070.3265 ± 0.009
✓0.8547 ± 0.0030.6075 ± 0.0070.3361 ± 0.009
✓✓0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.012
", + "bbox": [ + 81, + 348, + 467, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing a base template in the CLIP model offers some benefits, it also has its limitations, particularly in terms of embedding concepts within a similar CLIP feature space. In contrast, ACS appears to reduce the uniformity between concepts, allowing the concept discovery module to discover more distinct and accurate concepts for each neuron representation. This highlights the effectiveness of ACS in enhancing the overall performance of the concept discovery module. Ablation study on Concept Sensitivity $(\\alpha)$ . The ablation study depicted on the left side of Figure 4 examined the impact on the F1-score by varying levels of major concept sensitivity (denoted as $\\alpha$ ). We found that as concept sensitivity decreases, the F1-score initially increases $(\\alpha > 0.95)$ and then decreases. This pattern is due to a trade-off. Higher concept sensitivity leads to more precise concept identification but at the cost of identifying fewer concepts. Conversely, lower sensitivity results in more concepts being identified that may be less similar to the target concept. The point at which the F1-score is maximized can be seen as the optimal balance in this trade-off, providing a guideline for setting the most effective level of concept sensitivity for the major concept discovery.", + "bbox": [ + 75, + 433, + 472, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 76, + 776, + 194, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Overall Explanation Generated by WWW", + "text_level": 1, + "bbox": [ + 76, + 800, + 424, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 illustrates an explanation example generated by WWW, where the left side of the figure provides a class explanation for a test sample by highlighting important neurons for the predicted class, and the right side depicts a sample explanation showing the important neurons identi", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41abbd05d7ff74a9a13265c79cf6a52df4e98f8828677d3ec3efc0840011f7c1.jpg", + "image_caption": [ + "Figure 4. Ablation of concept sensitivity and heatmap similarity feasibility result. Left figure illustrates the F1 score with respect to Concept Sensitivity $(\\alpha)$ changes. Concept sensitivity $(\\alpha)$ that maximizes the F1-score is illustrated as the red line. The right figure illustrates the rejection test result of heatmap similarity and maximum softmax probability (MSP). # of hit denotes the number of correctly detected samples as a misprediction." + ], + "image_footnote": [], + "bbox": [ + 501, + 281, + 694, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/557ea332b44f6e5abb6c279d82309619a8e74eed22ebfb0dd5cc8e67bd31be50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 281, + 890, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcd3a6304ac012c6e51a47bb15f0a9de8ece2ad445a2e3ac85310b692351500a.jpg", + "image_caption": [ + "Figure 5. Example of generated explanation by WWW. From top to bottom, important neurons are displayed in the order of importance. Images in each neuron are examples of a major and minor concept, respectively. Colors in the top localization image show highly related regions for each concept. The bottom localization image is a weighted sum of important neuron activation maps displayed as a heatmap." + ], + "image_footnote": [], + "bbox": [ + 500, + 518, + 890, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fied for the test sample image prediction. This explanation target neurons from the penultimate layer of a ResNet-50 model pre-trained on ImageNet, using the ImageNet validation set $(D_{probe})$ and Wordnet nouns $(D_{concept})$ for the", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10974", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "explanation. Notably, neuron 30 is consistently identified as the most crucial in both explanations. It represents the 'tench' concept, which matches the prediction and the test sample's ground truth. Interestingly, while neuron 1246 is the third most important for the concept, it is the second most significant for the test sample's prediction, and neuron 1480, not highlighted in the class explanation, emerges as the third important neuron in the sample explanation. Despite these neuron selection and ranking variations, the model accurately predicts the correct class. The overall heatmap of the class and sample explanations can explain this. The heatmap explanation reveals a high cosine similarity between the two heatmaps, indicating that the calculated weighted sum of NAMs is similar. Despite the difference in individual neuron importance, the highly related input region is remarkably similar in both cases.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Anlaysis of Explanation of Failure Case", + "text_level": 1, + "bbox": [ + 76, + 353, + 405, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 provides an example of a mispredicted sample explanation generated by the WWW. In this particular case of failure, not only do the selected important neurons differ between the class and sample explanations, but the cosine similarity between their respective heatmaps is relatively low, measuring at 0.19. Interestingly, even though the ground-truth class explanations and the sample explanations highlight different important neurons, they both localize to similar regions in the heatmap, showing a relatively high similarity score of 0.47. From this observation, we explore the potential utility of heatmap similarity to predict uncertainty. On the right side of Figure 4, we present the results of a rejection test that evaluates heatmap similarity and the maximum softmax probability (MSP). When the rejection method detects the mispredicted sample correctly, we consider that as a hit. The number of hit is measured with respect to the rejected sample rate. As the rejection rate increases based on their uncertainty level, it becomes evident that heatmap similarity outperforms MSP regarding the detection of misprediction samples. This suggests that heatmap similarity can serve as a more effective measure of uncertainty compared to MSP. These findings indicate that heatmap similarity can be used as a tool for predicting uncertainty.", + "bbox": [ + 75, + 377, + 472, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 753, + 194, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed WWW, a unified framework that provides comprehensive explanations for the 'what', 'where', and 'why' of neural network decisions. WWW demonstrates superior performance in both quantitative and qualitative measures, offering a deeper and more detailed understanding of neural network behavior. This is achieved through a novel integration of adaptive selection for concept discovery, neuron activation maps, and Shapley values. WWW's", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/831385207359ed7e0d6f249e7eb3bc7cd0824bb87729be4ffe7505aa3f878c67.jpg", + "image_caption": [ + "Figure 6. Example of failure case explanation by WWW. The explanations of the predicted label are presented on the left side. On the right side, the explanation of the ground-truth label is shown. In the upper half, we displayed concept attribution maps, which show important concepts and their respective regions. In the bottom half, we showed the overall heatmap which shows important regions for the model decision." + ], + "image_footnote": [], + "bbox": [ + 511, + 87, + 883, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "adaptability is also shown across various neural network architectures, including convolutional networks and attention-based Vision Transformers. Additionally, our approach to predicting uncertainty through heatmap similarity analysis introduces a new way to obtain the reliability of their predictions. By offering localized explanations with concept annotations for individual inputs, WWW enhances the transparency of the model's decision-making process, contributing to the broader goal of making AI more reliable and trustworthy.", + "bbox": [ + 496, + 482, + 890, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 648, + 668, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Research Foundation of Korea (NRF) Grant funded by the Korea Government (MSIT) under Grant 2021R1G1A1094990, by the Institute of Information and Communications Technology Planning and Evaluation (IITP) Grant funded by the Korea Government (MSIT) under Grant 2022-0-00078 (Explainable Logical Reasoning for Medical Knowledge Generation), Grant IITP-2024-RS-2023-00258649 (ITRC(Information Technology Research Center) support program), Grant 2021-0-02068 (Artificial Intelligence Innovation Hub), Grant RS-2022-00155911 (Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)), and by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD).", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10975", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 2", + "[2] Yong Hyun Ahn, Gyeong-Moon Park, and Seong Tae Kim. Line: Out-of-distribution detection by leveraging important neurons. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19852-19862, 2023. 2, 4, 5", + "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 5, 6", + "[4] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[6] Thomas Fel, Victor Boutin, Mazda Moayeri, Rémi Cadène, Louis Bethune, Léo andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. arXiv preprint arXiv:2306.07304, 2023. 2", + "[7] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2", + "[8] Christiane Fellbaum. Wordnet and wordnets. encyclopedia of language and linguistics, 2005. 5", + "[9] Ruth Fong and Andrea Vedaldi. Net2vec: Quantifying and explaining how concepts are encoded by filters in deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8730-8738, 2018. 2", + "[10] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proceedings of the European conference on computer vision, pages 630-645. Springer, 2016. 5", + "[12] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2, 5, 6, 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alon Jacovi, Ana Marasović, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1", + "[14] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. International Conference on Machine Learning, 2023. 1, 2, 5, 6", + "[15] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1", + "[16] Ashkan Khakzar, Soroosh Baselizadeh, Saurabh Khanduja, Christian Rupprecht, Seong Tae Kim, and Nassir Navab. Neural response interpretation through the lens of critical pathways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13528-13538, 2021. 2, 4", + "[17] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2", + "[18] Mauritz Kop. Eu artificial intelligence act: The european approach to ai. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust ..., 2021. 1", + "[19] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2", + "[20] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-dissect: Automatic description of neuron representations in deep vision networks. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 7", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6", + "[22] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5", + "[23] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2", + "[24] Lloyd S Shapley. A value for n-person games. Classics in game theory, 69, 1997. 2", + "[25] Michael C Sloan. Aristotle's nicomachean ethics as the original locus for the septem circumstantiae. Classical Philology, 105(3):236-251, 2010. 1", + "[26] K. Song et al. Mpnet: Masked and permuted pre-training for language understanding. NeurIPS, 33:16857-16867, 2020. 6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10976", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Andong Wang, Wei-Ning Lee, and Xiaojuan Qi. Hint: Hierarchical neuron concept explainer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10254–10264, 2022. 2", + "[28] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 2", + "[29] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10977", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_model.json b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_model.json new file mode 100644 index 0000000000000000000000000000000000000000..faa2270c0f9e058ebc52da8bc2cb08acb2c35341 --- /dev/null +++ b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "WWW: A Unified Framework for Explaining What, Where and Why of Neural Networks by Interpretation of Neuron Concepts" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.203, + 0.714, + 0.221 + ], + "angle": 0, + "content": "Yong Hyun Ahn,\\(^{1}\\), Hyeon Bae Kim,\\(^{1}\\), Seong Tae Kim\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.222, + 0.859, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Artificial Intelligence, Kyung Hee University, Republic of Korea \\(^{2}\\)Department of Computer Science and Engineering, Kyung Hee University, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.259, + 0.658, + 0.275 + ], + "angle": 0, + "content": "{yhahn, hyeonbae.kim, st.kim}@khu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.342, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Recent advancements in neural networks have showcased their remarkable capabilities across various domains. Despite these successes, the \"black box\" problem still remains. To address this, we propose a novel framework, WWW, that offers the 'what', 'where', and 'why' of the neural network decisions in human-understandable terms. Specifically, WWW utilizes adaptive selection for concept discovery, employing adaptive cosine similarity and thresholding techniques to effectively explain 'what'. To address the 'where' and 'why', we proposed a novel combination of neuron activation maps (NAMs) with Shapley values, generating localized concept maps and heatmaps for individual inputs. Furthermore, WWW introduces a method for predicting uncertainty, leveraging heatmap similarities to estimate the prediction's reliability. Experimental evaluations of WWW demonstrate superior performance in both quantitative and qualitative metrics, outperforming existing methods in interpretability. WWW provides a unified solution for explaining 'what', 'where', and 'why', introducing a method for localized explanations from global interpretations and offering a plug-and-play solution adaptable to various architectures. Code is available at: https://github.com/ailab-kyunghee/WWW" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Neural networks have demonstrated impressive performance in various fields in recent years. Despite these successes, their widespread adoption in more diverse areas is slowed by several challenges. A fundamental issue is the \"black box\" problem, referring to the often hidden and unclear decision-making processes of neural network models. This lack of clarity raises concerns about the stability and reliability of these models, leading to a growing consensus that artificial intelligence should be reliable, robust," + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.309, + 0.895, + 0.394 + ], + "angle": 0, + "content": "Table 1. Illustration of What, Where, and Why of recent concept-based neural network interpretation methods. Green, yellow, and red marks illustrate that the method is able to interpret well, is partially interpretable, and has limitations for interpretation, respectively. Recent methods are able to interpret one or two 'w's but have limitations regarding interpreting three 'w's at once." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.403, + 0.872, + 0.49 + ], + "angle": 0, + "content": "
MethodWhatWhereWhy
CLIP-Dissect [20] (ICLR'23)✓XX
FALCON [14] (ICML'23)✓X▲
CRAFT [7] (CVPR'23)X✓▲
WWW (Ours)✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.579 + ], + "angle": 0, + "content": "and safe [4, 13]. As a response to this need for trustworthy AI, there has been an emergence of laws and regulations [15, 18] that require neural networks to base their decisions on principles that are understandable to humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.869 + ], + "angle": 0, + "content": "According to Doshi-Velez et al. [4], interpretability is defined as the capability to provide explanations in terms that are understandable to humans. Zhang et al. [29] expand on this definition, emphasizing that interpretability involves providing explanations in understandable terms to humans. This requirement for explanations in human-understandable terms is a consistent theme in the literature on interpretable methods. Furthermore, this concept aligns with a fundamental principle in journalism and problem-solving: the importance of clear and understandable communication. This principle is often encapsulated in the five Ws (Who, What, When, Where, Why) [25]. Building on this, we propose that practical explanations or interpretations should include three critical elements: 'what' (the nature of the decision or outcome), 'where' (the context or specific region of the input that affects the decision-making process), and 'why' (the reasoning or factors behind a decision). This approach aims to make the interpretations more accessible and relevant to human understanding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "From this point of view, former works can be re-grouped into several groups which those works aim to solve. For ex-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10968" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "ample, a series of works in feature attribution [10, 23, 28] can be classified as works mainly focusing on 'where'. These methods explain model decisions by identifying highly contributed input regions. However, identifying highly related input regions primarily focuses on 'where' only, not enough explanation for 'what' and 'why'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.468, + 0.365 + ], + "angle": 0, + "content": "On the other hand, concept-based explanations mainly focus on 'why'. Concept-based explanations [6, 7, 10, 17] explain model decisions by decomposing model responses into smaller units called concepts. Kim et al. proposed TCAV [17], understanding and interpreting model decisions into a form of concept activation vectors. Recently, Fel et al. proposed CRAFT [7], which utilizes non-negative matrix factorization for identifying concept vectors. The works mentioned above mainly decompose activations into the form of vectors, which gives a great advantage for finding out 'why' but often fails to match or annotate the human-understandable term (i.e., name) that is needed for 'what'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.468, + 0.579 + ], + "angle": 0, + "content": "Other streams of work, called Neuron-concept association, aim to name what each neuron (e.g. convolution filters, layer outputs) represents, which answers 'what' the model cares about. Bau et al. proposed Network Dissection [3], which identifies each neuron's representing concept from Broaden dataset [3]. Recently, CLIP-Dissect [20] and FALCON [14] have shown that representing the concept of neurons can be matched automatically by using a pre-trained CLIP model [21]. The aforementioned methods are suitable for explaining 'what' but partially explain or have limitations for explaining 'where' and 'why'. Furthermore, these methods generate global explanations focusing on each neuron's role in the network, not the local explanations for each sample input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.807 + ], + "angle": 0, + "content": "To address these issues, we propose a novel framework that can explain 'What', 'Where', and 'Why' (WWW) at once. WWW introduces adaptive selection for discovering each neuron's concept and interpreting each neuron to explain 'what'. By leveraging adaptive cosine similarity (ACS) and adaptive selection techniques, we achieve advanced performance compared to competitive methods. Moreover, we combined neuron activation map (NAM) and Shapley value [2, 16, 24] to generate class and sample concept maps and heatmaps to explain 'where' and 'why'. We also conducted various objective evaluations to assess the performance of the proposed method. In the experiments, WWW achieves better results in both quantitative and qualitative evaluations. Our key contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- We introduce a novel and effective way to generate high-quality explanations that explain 'what', 'where', and 'why' at once. Due to the powerful performance of the adaptive selection for concept discovery, WWW is able to achieve higher quantitative and better qualitative results in various metrics." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "- We introduce a way to generate localized explanations from global neural network interpretation. By the novel combination of Shapley value and neuron activation maps, WWW is able to generate localized explanations with concept annotations for sample input." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.167, + 0.892, + 0.227 + ], + "angle": 0, + "content": "- WWW can be attached to various target models with different architectures, from conventional convolution neural networks to the recent attention-based Vision transformers, in a plug-and-play manner." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.248, + 0.65, + 0.264 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.275, + 0.755, + 0.291 + ], + "angle": 0, + "content": "2.1. Neuron-Concept Association" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Bau et al. introduced Network Dissection [3], using the Broden dataset to identify which concepts individual neurons in a network represent. They use overlap between segmentation masks and feature maps to annotate concepts for neurons. Fong and Vedaldi expanded on this with Net2Vec [9], which looks at individual neurons and their combinations. Mu and Andreas further extended these ideas with Compositional Explanation [19], aiming to generate more complex and detailed explanations. These methods primarily focused on understanding 'what' a neuron represents. However, the aforementioned methods fell short in explaining the 'where' and 'why' of neuron representations. Additionally, they relied heavily on image-concept-matched datasets like Broden, which are often costly and hard to collect due to the need for pixel-wise labels. To address these limitations, recent approaches like HINT [27] and MILAN [12] have been developed. These methods train concept classifiers or models that reduce the dependency on image-concept-matched datasets. Moreover, approaches like CLIP-Dissect [20] and FALCON [14] leverage the pretrained CLIP model to use separate sets of image datasets and concept datasets. Despite these advancements in understanding 'what' neurons represent, there remains a gap in fully explaining the 'where' and 'why' of the neuron representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.734, + 0.71 + ], + "angle": 0, + "content": "2.2. Vector-based Explanation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The field of interpretability has significantly advanced, particularly in understanding and interpreting model decisions using Concept Activation Vectors (CAVs). Kim et al. introduced TCAV [17], understanding and interpreting model decisions into the form of CAVs. Recently, CRAFT [7] was introduced, leveraging non-negative matrix factorization to identify CAVs and localize the most relevant input regions for each CAV. Also, Achtibat et al. [1] and Fel et al. [6] proposed methods to interpret each CAV's role in decision-making. Despite these developments, a key challenge remains: translating CAVs into human-understandable terms." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10969" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.862, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.44, + 0.895, + 0.526 + ], + "angle": 0, + "content": "Figure 1. Overall flow of Concept Discovery module identifying concepts for a single neuron. We first calculate the cosine similarity of CLIP features between images and concepts with the template from the selected high-activating images. Then, we subtract the cosine similarity of CLIP features between images and the base template by only considering the similarity between the concept and image. From calculated adaptive cosine similarity (ACS), we generate concept score \\( S \\) by the average similarity of images. Note that concept score \\( S = \\{s_1, s_2, \\dots, s_m\\} \\) are a group of scores, not a single scalar. From the calculated concept scores \\( S \\), we select major concepts by adaptive selection. We also discover minor concepts using the same process but with crop images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.548, + 0.17, + 0.563 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.573, + 0.252, + 0.588 + ], + "angle": 0, + "content": "3.1. Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.474, + 0.838 + ], + "angle": 0, + "content": "WWW consists of three modules: Concept discovery, Localization, and Reasoning. The concept discovery module identifies each neuron's concept, which explains part 'what'. It selects neuron concepts from the concept set \\((D_{\\text{concept}})\\) by leveraging adaptive cosine similarity and adaptive selection. The localization module is to identify highly contributed input regions of the test sample, which concept is present at 'where'. Also, the combination of the neuron activation map leveraging Shapley value helps identify 'where' the concept is and tells the 'why' of the individual predictions. The reasoning module identifies important neurons of the test sample and predicted class. By comparing the differences between the sample and class explanations, users can understand the 'why' of the model prediction and, even more, whether the prediction is reliable or not." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.322, + 0.864 + ], + "angle": 0, + "content": "3.2. Concept Discovery Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The concept discovery module aims to identify proper major and minor concepts from \\( D_{\\text{concept}} \\) that match an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.717 + ], + "angle": 0, + "content": "example-based representation of each neuron. Let target model \\(f\\) d, and let \\((l,i)\\) as \\(i\\)-th neuron in layer \\(l\\) of the target model. From the images in probing dataset \\(D_{probe}\\), we select high activating images for neuron \\((l,i)\\) as \\(D_{rep}^{(l,i)}\\). With selected images in \\(D_{rep}^{(l,i)}\\) and concepts in \\(D_{concept}\\), we calculate CLIP visual feature of \\(D_{rep}^{(l,i)}\\) as \\(V^{(l,i)} = \\left[v_1^{(l,i)}, v_2^{(l,i)}, \\dots, v_n^{(l,i)}\\right]\\) and CLIP text feature of \\(D_{concept}\\) as \\(T = [t_1, t_2, \\dots, t_m]\\). \\(n\\) denotes number of example images for neuron \\((l,i)\\) and \\(m\\) stands for the number of concept in concept set \\(D_{concept}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We calculate concept score \\( s^{(l,i)} \\) by calculating adaptive cosine similarity (ACS), which allows us to reduce the effect of the base template and only consider the similarity between the image and the concept itself. Concept score \\( s_{j}^{(l,i)} \\) for \\( j \\)-th concept \\( t_j \\) of neuron \\( (l,i) \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.812, + 0.892, + 0.851 + ], + "angle": 0, + "content": "\\[\ns _ {j} ^ {(l, i)} = \\frac {1}{n} \\sum_ {o = 1} ^ {n} \\left\\{\\cos \\left(v _ {o} ^ {(l, i)}, t _ {j}\\right) - \\cos \\left(v _ {o} ^ {(l, i)}, t _ {t e m}\\right) \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(1 \\leq j \\leq m\\), \\(t_{tem}\\) denotes CLIP text feature of base template(e.g. 'a photo of') and \\(\\cos(x, y)\\) denotes cosine similarity between \\(x\\) and \\(y\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10970" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.46, + 0.895, + 0.531 + ], + "angle": 0, + "content": "Figure 2. Illustration of overall test time flow of WWW. In the test time (i.e., inference), class explanation selects important neurons with pre-computed Shapley value of the predicted class. On the other hand, the sample explanation selects important neurons with a Shapley value for the input sample. With selected important neurons, pre-computed concepts are annotated. After the concept annotation, a class heatmap is generated with the pre-computed Shapley value of the predicted class. On the other hand, a sample heatmap is generated with the Shapley value of the sample input." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.555, + 0.47, + 0.626 + ], + "angle": 0, + "content": "From the calculated concept scores \\( s^{(l,i)} = \\left[s_1^{(l,i)},\\dots ,s_m^{(l,i)}\\right] \\), we select concepts where \\( s_j^{(l,i)} > \\delta^{(l,i)} \\). Adaptive selection threshold \\( \\delta^{(l,i)} \\) for neuron \\( (l,i) \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.637, + 0.47, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\delta^ {(l, i)} = \\alpha \\times \\max (s ^ {(l, i)}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.718 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes concept sensitivity for major and minor concepts. For discovering minor concepts, we select \\(D_{rep}^{(l,i)}\\) with cropped images of probing dataset \\(D_{probe}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.273, + 0.74 + ], + "angle": 0, + "content": "3.3. Localization Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.471, + 0.854 + ], + "angle": 0, + "content": "The localization module aims to identify highly contributed input regions of each concept and generate a concept map and concept heatmap. We select important neurons with Taylor approximation of Shapley value introduced in [16] for generating a concept region map. For input image sample \\( x \\), Neuron contribution \\( w^{(l,i)} \\) of neuron \\( (l,i) \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.863, + 0.476, + 0.902 + ], + "angle": 0, + "content": "\\[\nw ^ {(l, i)} (x) = \\left| f (x) - f (x; a ^ {(l, i)} \\leftarrow 0) \\right| = \\left| a ^ {(l, i)} \\nabla_ {a ^ {(l, i)}} f (x) \\right|. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Where \\( a^{(l,i)} \\) denotes activation of neuron \\( (l,i) \\). After calculating each layer's Neuron contribution, we rank neurons by calculated contribution (i.e., Shapley value) and select the top- \\( k \\) important neurons for the sample. The concept heatmap \\( M \\) is calculated with the weighted sum of important neurons' neuron activation maps (NAMs). Concept heatmap \\( M^l(x) \\) of important neurons is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.677, + 0.892, + 0.707 + ], + "angle": 0, + "content": "\\[\nM ^ {l} (x) = \\sum^ {u} w ^ {(l, u)} (x) A ^ {(l, u)} (x) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.789 + ], + "angle": 0, + "content": "where \\( u \\) denotes the index of important neurons, \\( w \\) denotes the neuron contribution, and \\( A \\) denotes the neuron activation map of the neuron. Note that \\( M^l \\) shows only related regions of a single or combination of important neurons, not the whole network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.681, + 0.817 + ], + "angle": 0, + "content": "3.4. Reasoning Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The reasoning module is designed to help users understand the 'why' of the model output. This not only explains the result but also includes 'why' this prediction is reliable or not. By leveraging the class-wise Shapley value introduced in [2], we can understand each neuron's class-wise contri" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "bution and identify important neurons for each class. From Ahn et al. [2], the class-wise contribution of a neuron can be calculated by the average contribution of a neuron in class samples. From the class-wise contribution, we rank top-\\(k\\) important neurons for each class. We can also generate a class-wise concept region map and concept heatmap with the index of important neurons. Class-wise maps can be used as a guideline for understanding the general case of the prediction. By comparing Class-wise maps and sample maps, users can identify which concept and region of the sample differs from general cases. This naturally helps users understand the 'why' of the model output and also 'why' the results are reliable or not. Figure 2 for further understanding the WWW flow for generating class and sample explanations. Figure 5 shows the generated explanation example." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.341, + 0.38, + 0.358 + ], + "angle": 0, + "content": "3.5. Overall flow and Output of WWW" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.365, + 0.473, + 0.76 + ], + "angle": 0, + "content": "As described above, WWW has three main modules explaining three 'w' of 'what', 'where', and 'why', respectively. In Section 3.5, we are going to follow the overall flow of our method by time sequence. Before starting inference, WWW needs pre-computing for discovering major and minor concepts with the concept discovery module (Sec 3.2) and also a class-wise contribution for class-wise analysis in the Reasoning Module (Sec 3.4). After pre-computing concepts and contributions, WWW is ready to generate test sample explanations. Figure 2 shows the overall flow of explaining the generation of WWW in the inference time. In the test time (i.e., inference), WWW generates a class-wise explanation of the predicted class and sample explanation into two tracks. WWW leverages the predicted class and class-wise contribution matrix to generate a class-wise explanation to select the most critical neurons for the predicted class. After identifying important neurons, WWW leverages the localization module (Sec 3.3) to generate a class concept region map and class concept combination heatmap. On the other hand, for the sample explanation, WWW calculates the Taylor-approximated Shapley value of the sample for the predicted class and identifies critical neurons for the prediction of the sample. After identifying important neurons for the sample, WWW leverages the sample Shapley value to generate the sample concept region map and sample concept combination heat map." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.202, + 0.788 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Comprehensive experiments have been conducted to evaluate our method. In section 4.1, we evaluate the performance of the concept discovery module both qualitatively (Sec. 4.1.1) and quantitatively (Sec. 4.1.2). Section 4.2 is an ablation study for the concept discovery module. In Section 4.3, we analyze generated explanations of both correct and wrong predictions with examples." + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.089, + 0.623, + 0.098 + ], + "angle": 0, + "content": "ResNet - 50 Layer 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.102, + 0.6, + 0.11 + ], + "angle": 0, + "content": "Neuron 1152" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.111, + 0.544, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.111, + 0.576, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.112, + 0.692, + 0.12 + ], + "angle": 0, + "content": "WWW (Ours): goldfish - flame fish" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.12, + 0.659, + 0.127 + ], + "angle": 0, + "content": "CLIP-Dissect: swordtail" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.128, + 0.657, + 0.135 + ], + "angle": 0, + "content": "MILAN: ORANGE objects FALCON: TRICK, CALE" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.135, + 0.682, + 0.139 + ], + "angle": 0, + "content": "FALCON:Insh,sale,series,fare" + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.146, + 0.599, + 0.153 + ], + "angle": 0, + "content": "Neuron 816" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.153, + 0.545, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.154, + 0.707, + 0.161 + ], + "angle": 0, + "content": "WWW (Ours): saltwater fish - swordtail" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.162, + 0.687, + 0.168 + ], + "angle": 0, + "content": "CLIP-Dissect: genus Chaetodon" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.169, + 0.646, + 0.175 + ], + "angle": 0, + "content": "MILAN: Aquatic life" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.176, + 0.715, + 0.181 + ], + "angle": 0, + "content": "FALCON: cover, black, white, home, blue" + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.089, + 0.819, + 0.099 + ], + "angle": 0, + "content": "ResNet - 50 Final layer" + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.117, + 0.776, + 0.124 + ], + "angle": 0, + "content": "Neuron 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.125, + 0.795, + 0.131 + ], + "angle": 0, + "content": "Ground Truth: goldfish" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.131, + 0.795, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.131, + 0.82, + 0.138 + ], + "angle": 0, + "content": "www" + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.138, + 0.874, + 0.144 + ], + "angle": 0, + "content": "CLIP-Dissect: swordtail" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.145, + 0.846, + 0.151 + ], + "angle": 0, + "content": "MILAN: Nature" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.152, + 0.891, + 0.165 + ], + "angle": 0, + "content": "FALCON: trish, size, storage assorted, gap" + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.186, + 0.6, + 0.194 + ], + "angle": 0, + "content": "Neuron 2018" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.196, + 0.544, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.197, + 0.713, + 0.205 + ], + "angle": 0, + "content": "WWW (Ours): Dutch oven - stewing pan" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.206, + 0.685, + 0.211 + ], + "angle": 0, + "content": "CLIP-Dissect: electric frying pan" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.212, + 0.695, + 0.218 + ], + "angle": 0, + "content": "MILAN: Rounded edges in pictures" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.218, + 0.721, + 0.224 + ], + "angle": 0, + "content": "FALCON: black, card, woman, home, white" + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.207, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Neuron 544" + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.214, + 0.799, + 0.221 + ], + "angle": 0, + "content": "Ground Truth: Dutch o" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.221, + 0.799, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.222, + 0.888, + 0.229 + ], + "angle": 0, + "content": "WWW (Ours): Dutch oven" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.229, + 0.888, + 0.235 + ], + "angle": 0, + "content": "CLIP-Dissect: stewing pan" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.236, + 0.878, + 0.241 + ], + "angle": 0, + "content": "MILAN: Circular objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.241, + 0.888, + 0.247 + ], + "angle": 0, + "content": "FACLCON:men top game" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.248, + 0.846, + 0.255 + ], + "angle": 0, + "content": "award, guard" + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.231, + 0.6, + 0.238 + ], + "angle": 0, + "content": "Neuron 1911" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.24, + 0.687, + 0.247 + ], + "angle": 0, + "content": "WWW (Ours): cookfire - cookfire" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.248, + 0.663, + 0.254 + ], + "angle": 0, + "content": "CLIP-Dissect: cookstove" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.255, + 0.7, + 0.261 + ], + "angle": 0, + "content": "MILAN: Animals and circular objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.261, + 0.723, + 0.267 + ], + "angle": 0, + "content": "FALCON: shirt, school, mini, woman, house" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of WWW with other baselines. We compared WWW with three competitive baselines (CLIP-Dissect [20], MILAN [12], FALCON [14]) in two final layer neurons and four penultimate layer (i.e., layer 4) neurons with each neuron's highly activating images. layer-4 neurons are top-2 important neurons of the final layer class. We have colored the descriptions green if they match the images, yellow if they match but are too generic or similar, and red if they do not match." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.418, + 0.882, + 0.435 + ], + "angle": 0, + "content": "4.1. Performance Evaluation for Concept Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.579 + ], + "angle": 0, + "content": "In this experiment, we evaluate the concept matching performance of the concept discovery module with four other baselines, Network Dissection [3], MILAN [12], CLIP-Dissect [20] and FALCON [14]. We evaluate the performance of methods on the various models (e.g., ResNet-18 [11], ResNet-50 [11], and ViT-B/16 [5]), various probing datasets (e.g.,Imagenet, Places365) and various concept sets (e.g., Wordnet nouns, and labels of Places365, Broaden, and ImageNet)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.687, + 0.618 + ], + "angle": 0, + "content": "4.1.1 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Settings. We compared WWW with the three most comparable methods (CLIP-Dissect [20], MILAN [12], FALCON [14]) in the penultimate layer and final layer of the model. We do not compare with Network Dissection [3] due to the limitation that probe image data \\( D_{probe} \\) and concept set \\( D_{concept} \\) is fixed to Broaden. We used a ResNet-50 [11] model pre-trained in the ImageNet-1k [22] dataset. For probe image data \\( D_{probe} \\), we used the ImageNet-1k validation set, and we extracted all nouns in Wordnet[8] (about 80k) dataset for a concept set \\( D_{concept} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results. Figure 3 shows examples of descriptions for hidden neurons in the penultimate and final layers. Neurons in the penultimate layer are top-2 important neurons of the final layer neuron's ground truth label class. We observed that WWW not only interpreted each neuron well but also showed robust interpretation that the most important neuron of the class in the penultimate layer represents the same major concept as the final layer neuron." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10972" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.134 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on final layer concept matching performance of ResNet-50 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for \\(D_{probe}\\). Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.143, + 0.872, + 0.307 + ], + "angle": 0, + "content": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
Network Dissection [3]BrodenBroden(1.2k)0.7229 ± 0.0030.2989 ± 0.0050.0010 ± 0.0010.001
MILAN(b) [12]ImageNet val-0.7300 ± 0.0030.2485 ± 0.0050.0005 ± 0.0000.001
FALCON [14]ImageNet valLAION-400m0.7065 ± 0.0030.1790 ± 0.0010.0002 ± 0.0000.001
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9340 ± 0.0030.8376 ± 0.0060.7286 ± 0.0090.933
ImageNet valBroden (1.2k)0.7369 ± 0.0030.3432 ± 0.0040.0328 ± 0.0030.108
ImageNet valWordnet (80k)0.8689 ± 0.0040.6846 ± 0.0080.3647 ± 0.0140.456
WWW (Ours)ImageNet valImageNet (1k)0.9325 ± 0.0030.8327 ± 0.0060.7719 ± 0.0090.955
ImageNet valBroden (1.2k)0.7758 ± 0.0040.4414 ± 0.0070.0645 ± 0.0070.091
ImageNet valWordnet (80k)0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.0120.645
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on final layer concept matching performance of ViT-B/16 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for \\(D_{probe}\\). Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.372, + 0.855, + 0.49 + ], + "angle": 0, + "content": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9337 ± 0.0030.8375 ± 0.0060.7289 ± 0.0090.933
ImageNet valBroden (1.2k)0.7365 ± 0.0030.3416 ± 0.0040.0319 ± 0.0030.106
ImageNet valWordnet (80k)0.8700 ± 0.0040.6886 ± 0.0080.3679 ± 0.0140.460
WWW (Ours)ImageNet valImageNet (1k)0.9331 ± 0.0030.8347 ± 0.0060.7718 ± 0.0090.955
ImageNet valBroden (1.2k)0.7754 ± 0.0040.4407 ± 0.0070.0651 ± 0.0070.091
ImageNet valWordnet (80k)0.8857 ± 0.0030.6970 ± 0.0080.4166 ± 0.0120.634
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.275, + 0.529 + ], + "angle": 0, + "content": "4.1.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.644 + ], + "angle": 0, + "content": "In this section, we compare the performance of our methods with baselines. As introduced in [20], we evaluate final layer neuron concepts with the class labels with various metrics. By comparing generated explanations with the class labels, we can objectively evaluate the quality of the generated neuron labels with what each neuron is trained to represent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Metrics. CLIP cos and mpnet cos are measured as cosine similarities between the encoded feature of the class label and selected concepts with CLIP [21] model and mpnet [26] model, respectively. F1-score is measured to evaluate the discovered concept's balance of exactness and flexibility. Also, the hit rate is calculated as a rate of selected concepts that exactly match the class label." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Quantitative comparison on ResNet-50. Table 2 compares WWW with Network Dissection [3], MILAN(b) [12], FALCON [14], and CLIP-Dissect [20]. We evaluate the final layer neuron concepts of ResNet-50 with the class labels of ImageNet-1k. In table 2, WWW showed better performance as the concept set \\( D_{\\text{concept}} \\) gets larger and outperformed all other baselines when concept set \\( D_{\\text{concept}} \\) is set to Broden and Wordnet nouns. In the results of Wordnet nouns (\\( D_{\\text{concept}} \\)), comparison between WWW and CLIP-Dissect [20] are statistically significant across all metrics" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.514, + 0.58, + 0.528 + ], + "angle": 0, + "content": "\\((p < 0.05)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.621 + ], + "angle": 0, + "content": "Quantitative comparison on ViT-B/16. We compared predicted concepts of the final layer to ground truth labels of ViT-B/16 pre-trained on ImageNet. In Table 3, WWW showed better performance as the concept set \\( D_{\\text{concept}} \\) gets larger and outperformed other baselines when concept set \\( D_{\\text{concept}} \\) is set to Broden and Wordnet nouns." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.895, + 0.728 + ], + "angle": 0, + "content": "Quantitative comparison on ResNet-18 pre-trained in Places365. We compared predicted labels to ground truth labels in final layer neurons of ResNet-18 pre-trained on Places365. In table 4, WWW showed better performance as the concept set \\( D_{\\text{concept}} \\) gets larger and outperformed other baselines when concept set \\( D_{\\text{concept}} \\) is set to Broden and Wordnet nouns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.651, + 0.758 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Evaluation on the effect of leveraging base template and ACS. Table 5 shows an ablation study over various parts used in WWW. Without using a base template, WWW showed the lowest performance overall. With the use of a base template, WWW shows slightly increased performance in CLIP cos and F1-score but not much advance overall. However, performance was significantly increased, and the highest scores in all three metrics were achieved when leveraging ACS. The findings suggest that while us-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10973" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on final layer concept matching performance of ResNet-18 trained on Places365. We compared predicted neuron concepts with ground truth labels of Places365. We used the Places365 test set for \\(D_{probe}\\). Bold numbers represent the best scores between the same settings. The average score and standard errors of the 365 final layer neurons are reported." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.143, + 0.865, + 0.261 + ], + "angle": 0, + "content": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]Places365 testPlaces365 (0.4k)0.9562 ± 0.0040.8687 ± 0.0120.7233 ± 0.0230.723
Places365 testBroden (1.2k)0.8304 ± 0.0030.4678 ± 0.0070.1096 ± 0.0080.329
Places365 testWordnet (80k)0.8378 ± 0.0060.5107 ± 0.0140.1041 ± 0.0160.104
WWW (Ours)Places365 testPlaces365 (0.4k)0.9402 ± 0.0040.8204 ± 0.0130.6361 ± 0.0240.660
Places365 testBroden (1.2k)0.8925 ± 0.0050.6415 ± 0.0130.2242 ± 0.0210.255
Places365 testWordnet (80k)0.8492 ± 0.0040.5000 ± 0.0130.1106 ± 0.0150.142
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.282, + 0.47, + 0.338 + ], + "angle": 0, + "content": "Table 5. Ablation study on the use of the template and ACS. '√' in the Template represents WWW using a template (e.g., 'a photo of word') '√' in the ACS represents WWW using adaptive cosine similarity." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.349, + 0.468, + 0.407 + ], + "angle": 0, + "content": "
TemplateACSCLIP cosmpnet cosF1-score
0.8499 ± 0.0030.6123 ± 0.0070.3265 ± 0.009
✓0.8547 ± 0.0030.6075 ± 0.0070.3361 ± 0.009
✓✓0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.012
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.473, + 0.766 + ], + "angle": 0, + "content": "ing a base template in the CLIP model offers some benefits, it also has its limitations, particularly in terms of embedding concepts within a similar CLIP feature space. In contrast, ACS appears to reduce the uniformity between concepts, allowing the concept discovery module to discover more distinct and accurate concepts for each neuron representation. This highlights the effectiveness of ACS in enhancing the overall performance of the concept discovery module. Ablation study on Concept Sensitivity \\((\\alpha)\\). The ablation study depicted on the left side of Figure 4 examined the impact on the F1-score by varying levels of major concept sensitivity (denoted as \\(\\alpha\\)). We found that as concept sensitivity decreases, the F1-score initially increases \\((\\alpha > 0.95)\\) and then decreases. This pattern is due to a trade-off. Higher concept sensitivity leads to more precise concept identification but at the cost of identifying fewer concepts. Conversely, lower sensitivity results in more concepts being identified that may be less similar to the target concept. The point at which the F1-score is maximized can be seen as the optimal balance in this trade-off, providing a guideline for setting the most effective level of concept sensitivity for the major concept discovery." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.195, + 0.79 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.425, + 0.816 + ], + "angle": 0, + "content": "4.3.1 Overall Explanation Generated by WWW" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Figure 5 illustrates an explanation example generated by WWW, where the left side of the figure provides a class explanation for a test sample by highlighting important neurons for the predicted class, and the right side depicts a sample explanation showing the important neurons identi" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.282, + 0.695, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.282, + 0.892, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.407, + 0.895, + 0.505 + ], + "angle": 0, + "content": "Figure 4. Ablation of concept sensitivity and heatmap similarity feasibility result. Left figure illustrates the F1 score with respect to Concept Sensitivity \\((\\alpha)\\) changes. Concept sensitivity \\((\\alpha)\\) that maximizes the F1-score is illustrated as the red line. The right figure illustrates the rejection test result of heatmap similarity and maximum softmax probability (MSP). # of hit denotes the number of correctly detected samples as a misprediction." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.52, + 0.892, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.714, + 0.895, + 0.812 + ], + "angle": 0, + "content": "Figure 5. Example of generated explanation by WWW. From top to bottom, important neurons are displayed in the order of importance. Images in each neuron are examples of a major and minor concept, respectively. Colors in the top localization image show highly related regions for each concept. The bottom localization image is a weighted sum of important neuron activation maps displayed as a heatmap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "fied for the test sample image prediction. This explanation target neurons from the penultimate layer of a ResNet-50 model pre-trained on ImageNet, using the ImageNet validation set \\((D_{probe})\\) and Wordnet nouns \\((D_{concept})\\) for the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10974" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "explanation. Notably, neuron 30 is consistently identified as the most crucial in both explanations. It represents the 'tench' concept, which matches the prediction and the test sample's ground truth. Interestingly, while neuron 1246 is the third most important for the concept, it is the second most significant for the test sample's prediction, and neuron 1480, not highlighted in the class explanation, emerges as the third important neuron in the sample explanation. Despite these neuron selection and ranking variations, the model accurately predicts the correct class. The overall heatmap of the class and sample explanations can explain this. The heatmap explanation reveals a high cosine similarity between the two heatmaps, indicating that the calculated weighted sum of NAMs is similar. Despite the difference in individual neuron importance, the highly related input region is remarkably similar in both cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.354, + 0.406, + 0.369 + ], + "angle": 0, + "content": "4.3.2 Anlaysis of Explanation of Failure Case" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.473, + 0.741 + ], + "angle": 0, + "content": "Figure 6 provides an example of a mispredicted sample explanation generated by the WWW. In this particular case of failure, not only do the selected important neurons differ between the class and sample explanations, but the cosine similarity between their respective heatmaps is relatively low, measuring at 0.19. Interestingly, even though the ground-truth class explanations and the sample explanations highlight different important neurons, they both localize to similar regions in the heatmap, showing a relatively high similarity score of 0.47. From this observation, we explore the potential utility of heatmap similarity to predict uncertainty. On the right side of Figure 4, we present the results of a rejection test that evaluates heatmap similarity and the maximum softmax probability (MSP). When the rejection method detects the mispredicted sample correctly, we consider that as a hit. The number of hit is measured with respect to the rejected sample rate. As the rejection rate increases based on their uncertainty level, it becomes evident that heatmap similarity outperforms MSP regarding the detection of misprediction samples. This suggests that heatmap similarity can serve as a more effective measure of uncertainty compared to MSP. These findings indicate that heatmap similarity can be used as a tool for predicting uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.196, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We proposed WWW, a unified framework that provides comprehensive explanations for the 'what', 'where', and 'why' of neural network decisions. WWW demonstrates superior performance in both quantitative and qualitative measures, offering a deeper and more detailed understanding of neural network behavior. This is achieved through a novel integration of adaptive selection for concept discovery, neuron activation maps, and Shapley values. WWW's" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.088, + 0.885, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.358, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 6. Example of failure case explanation by WWW. The explanations of the predicted label are presented on the left side. On the right side, the explanation of the ground-truth label is shown. In the upper half, we displayed concept attribution maps, which show important concepts and their respective regions. In the bottom half, we showed the overall heatmap which shows important regions for the model decision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.633 + ], + "angle": 0, + "content": "adaptability is also shown across various neural network architectures, including convolutional networks and attention-based Vision Transformers. Additionally, our approach to predicting uncertainty through heatmap similarity analysis introduces a new way to obtain the reliability of their predictions. By offering localized explanations with concept annotations for individual inputs, WWW enhances the transparency of the model's decision-making process, contributing to the broader goal of making AI more reliable and trustworthy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.669, + 0.666 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Research Foundation of Korea (NRF) Grant funded by the Korea Government (MSIT) under Grant 2021R1G1A1094990, by the Institute of Information and Communications Technology Planning and Evaluation (IITP) Grant funded by the Korea Government (MSIT) under Grant 2022-0-00078 (Explainable Logical Reasoning for Medical Knowledge Generation), Grant IITP-2024-RS-2023-00258649 (ITRC(Information Technology Research Center) support program), Grant 2021-0-02068 (Artificial Intelligence Innovation Hub), Grant RS-2022-00155911 (Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)), and by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10975" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Yong Hyun Ahn, Gyeong-Moon Park, and Seong Tae Kim. Line: Out-of-distribution detection by leveraging important neurons. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19852-19862, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[4] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.464 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Thomas Fel, Victor Boutin, Mazda Moayeri, Rémi Cadène, Louis Bethune, Léo andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. arXiv preprint arXiv:2306.07304, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[7] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[8] Christiane Fellbaum. Wordnet and wordnets. encyclopedia of language and linguistics, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[9] Ruth Fong and Andrea Vedaldi. Net2vec: Quantifying and explaining how concepts are encoded by filters in deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8730-8738, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[10] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.776, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proceedings of the European conference on computer vision, pages 630-645. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[12] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2, 5, 6, 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[13] Alon Jacovi, Ana Marasović, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[14] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. International Conference on Machine Learning, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[15] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[16] Ashkan Khakzar, Soroosh Baselizadeh, Saurabh Khanduja, Christian Rupprecht, Seong Tae Kim, and Nassir Navab. Neural response interpretation through the lens of critical pathways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13528-13538, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[17] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.446 + ], + "angle": 0, + "content": "[18] Mauritz Kop. Eu artificial intelligence act: The european approach to ai. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust ..., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[19] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.893, + 0.546 + ], + "angle": 0, + "content": "[20] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-dissect: Automatic description of neuron representations in deep vision networks. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.894, + 0.63 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[22] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.893, + 0.785 + ], + "angle": 0, + "content": "[23] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[24] Lloyd S Shapley. A value for n-person games. Classics in game theory, 69, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[25] Michael C Sloan. Aristotle's nicomachean ethics as the original locus for the septem circumstantiae. Classical Philology, 105(3):236-251, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[26] K. Song et al. Mpnet: Masked and permuted pre-training for language understanding. NeurIPS, 33:16857-16867, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10976" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Andong Wang, Wei-Ning Lee, and Xiaojuan Qi. Hint: Hierarchical neuron concept explainer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10254–10264, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[28] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[29] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10977" + } + ] +] \ No newline at end of file diff --git a/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_origin.pdf b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e3528e7a228701ddd884fe37732bea699ee3998 --- /dev/null +++ b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/81e2b3ad-69ba-401e-815e-45595c76f065_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c68e6ddbff48a55cd7856691fbf7d8b272d8bee89cbaa2ca7c142c4a933ae095 +size 9051309 diff --git a/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/full.md b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..455f985ac069bce8208d5fa75d5cd9afa0b7d395 --- /dev/null +++ b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/full.md @@ -0,0 +1,275 @@ +# WWW: A Unified Framework for Explaining What, Where and Why of Neural Networks by Interpretation of Neuron Concepts + +Yong Hyun Ahn, $^{1}$ , Hyeon Bae Kim, $^{1}$ , Seong Tae Kim $^{2*}$ + +$^{1}$ Department of Artificial Intelligence, Kyung Hee University, Republic of Korea $^{2}$ Department of Computer Science and Engineering, Kyung Hee University, Republic of Korea + +{yhahn, hyeonbae.kim, st.kim}@khu.ac.kr + +# Abstract + +Recent advancements in neural networks have showcased their remarkable capabilities across various domains. Despite these successes, the "black box" problem still remains. To address this, we propose a novel framework, WWW, that offers the 'what', 'where', and 'why' of the neural network decisions in human-understandable terms. Specifically, WWW utilizes adaptive selection for concept discovery, employing adaptive cosine similarity and thresholding techniques to effectively explain 'what'. To address the 'where' and 'why', we proposed a novel combination of neuron activation maps (NAMs) with Shapley values, generating localized concept maps and heatmaps for individual inputs. Furthermore, WWW introduces a method for predicting uncertainty, leveraging heatmap similarities to estimate the prediction's reliability. Experimental evaluations of WWW demonstrate superior performance in both quantitative and qualitative metrics, outperforming existing methods in interpretability. WWW provides a unified solution for explaining 'what', 'where', and 'why', introducing a method for localized explanations from global interpretations and offering a plug-and-play solution adaptable to various architectures. Code is available at: https://github.com/ailab-kyunghee/WWW + +# 1. Introduction + +Neural networks have demonstrated impressive performance in various fields in recent years. Despite these successes, their widespread adoption in more diverse areas is slowed by several challenges. A fundamental issue is the "black box" problem, referring to the often hidden and unclear decision-making processes of neural network models. This lack of clarity raises concerns about the stability and reliability of these models, leading to a growing consensus that artificial intelligence should be reliable, robust, + +Table 1. Illustration of What, Where, and Why of recent concept-based neural network interpretation methods. Green, yellow, and red marks illustrate that the method is able to interpret well, is partially interpretable, and has limitations for interpretation, respectively. Recent methods are able to interpret one or two 'w's but have limitations regarding interpreting three 'w's at once. + +
MethodWhatWhereWhy
CLIP-Dissect [20] (ICLR'23)✓XX
FALCON [14] (ICML'23)✓X▲
CRAFT [7] (CVPR'23)X✓▲
WWW (Ours)✓✓✓
+ +and safe [4, 13]. As a response to this need for trustworthy AI, there has been an emergence of laws and regulations [15, 18] that require neural networks to base their decisions on principles that are understandable to humans. + +According to Doshi-Velez et al. [4], interpretability is defined as the capability to provide explanations in terms that are understandable to humans. Zhang et al. [29] expand on this definition, emphasizing that interpretability involves providing explanations in understandable terms to humans. This requirement for explanations in human-understandable terms is a consistent theme in the literature on interpretable methods. Furthermore, this concept aligns with a fundamental principle in journalism and problem-solving: the importance of clear and understandable communication. This principle is often encapsulated in the five Ws (Who, What, When, Where, Why) [25]. Building on this, we propose that practical explanations or interpretations should include three critical elements: 'what' (the nature of the decision or outcome), 'where' (the context or specific region of the input that affects the decision-making process), and 'why' (the reasoning or factors behind a decision). This approach aims to make the interpretations more accessible and relevant to human understanding. + +From this point of view, former works can be re-grouped into several groups which those works aim to solve. For ex- + +ample, a series of works in feature attribution [10, 23, 28] can be classified as works mainly focusing on 'where'. These methods explain model decisions by identifying highly contributed input regions. However, identifying highly related input regions primarily focuses on 'where' only, not enough explanation for 'what' and 'why'. + +On the other hand, concept-based explanations mainly focus on 'why'. Concept-based explanations [6, 7, 10, 17] explain model decisions by decomposing model responses into smaller units called concepts. Kim et al. proposed TCAV [17], understanding and interpreting model decisions into a form of concept activation vectors. Recently, Fel et al. proposed CRAFT [7], which utilizes non-negative matrix factorization for identifying concept vectors. The works mentioned above mainly decompose activations into the form of vectors, which gives a great advantage for finding out 'why' but often fails to match or annotate the human-understandable term (i.e., name) that is needed for 'what'. + +Other streams of work, called Neuron-concept association, aim to name what each neuron (e.g. convolution filters, layer outputs) represents, which answers 'what' the model cares about. Bau et al. proposed Network Dissection [3], which identifies each neuron's representing concept from Broaden dataset [3]. Recently, CLIP-Dissect [20] and FALCON [14] have shown that representing the concept of neurons can be matched automatically by using a pre-trained CLIP model [21]. The aforementioned methods are suitable for explaining 'what' but partially explain or have limitations for explaining 'where' and 'why'. Furthermore, these methods generate global explanations focusing on each neuron's role in the network, not the local explanations for each sample input. + +To address these issues, we propose a novel framework that can explain 'What', 'Where', and 'Why' (WWW) at once. WWW introduces adaptive selection for discovering each neuron's concept and interpreting each neuron to explain 'what'. By leveraging adaptive cosine similarity (ACS) and adaptive selection techniques, we achieve advanced performance compared to competitive methods. Moreover, we combined neuron activation map (NAM) and Shapley value [2, 16, 24] to generate class and sample concept maps and heatmaps to explain 'where' and 'why'. We also conducted various objective evaluations to assess the performance of the proposed method. In the experiments, WWW achieves better results in both quantitative and qualitative evaluations. Our key contributions are summarized as follows: + +- We introduce a novel and effective way to generate high-quality explanations that explain 'what', 'where', and 'why' at once. Due to the powerful performance of the adaptive selection for concept discovery, WWW is able to achieve higher quantitative and better qualitative results in various metrics. + +- We introduce a way to generate localized explanations from global neural network interpretation. By the novel combination of Shapley value and neuron activation maps, WWW is able to generate localized explanations with concept annotations for sample input. +- WWW can be attached to various target models with different architectures, from conventional convolution neural networks to the recent attention-based Vision transformers, in a plug-and-play manner. + +# 2. Related Works + +# 2.1. Neuron-Concept Association + +Bau et al. introduced Network Dissection [3], using the Broden dataset to identify which concepts individual neurons in a network represent. They use overlap between segmentation masks and feature maps to annotate concepts for neurons. Fong and Vedaldi expanded on this with Net2Vec [9], which looks at individual neurons and their combinations. Mu and Andreas further extended these ideas with Compositional Explanation [19], aiming to generate more complex and detailed explanations. These methods primarily focused on understanding 'what' a neuron represents. However, the aforementioned methods fell short in explaining the 'where' and 'why' of neuron representations. Additionally, they relied heavily on image-concept-matched datasets like Broden, which are often costly and hard to collect due to the need for pixel-wise labels. To address these limitations, recent approaches like HINT [27] and MILAN [12] have been developed. These methods train concept classifiers or models that reduce the dependency on image-concept-matched datasets. Moreover, approaches like CLIP-Dissect [20] and FALCON [14] leverage the pretrained CLIP model to use separate sets of image datasets and concept datasets. Despite these advancements in understanding 'what' neurons represent, there remains a gap in fully explaining the 'where' and 'why' of the neuron representations. + +# 2.2. Vector-based Explanation + +The field of interpretability has significantly advanced, particularly in understanding and interpreting model decisions using Concept Activation Vectors (CAVs). Kim et al. introduced TCAV [17], understanding and interpreting model decisions into the form of CAVs. Recently, CRAFT [7] was introduced, leveraging non-negative matrix factorization to identify CAVs and localize the most relevant input regions for each CAV. Also, Achtibat et al. [1] and Fel et al. [6] proposed methods to interpret each CAV's role in decision-making. Despite these developments, a key challenge remains: translating CAVs into human-understandable terms. + +![](images/bf758565ed0917fa187b3806594c33d01b812147cbee75e8c2ca58e3bf17fb5f.jpg) +Figure 1. Overall flow of Concept Discovery module identifying concepts for a single neuron. We first calculate the cosine similarity of CLIP features between images and concepts with the template from the selected high-activating images. Then, we subtract the cosine similarity of CLIP features between images and the base template by only considering the similarity between the concept and image. From calculated adaptive cosine similarity (ACS), we generate concept score $S$ by the average similarity of images. Note that concept score $S = \{s_1, s_2, \dots, s_m\}$ are a group of scores, not a single scalar. From the calculated concept scores $S$ , we select major concepts by adaptive selection. We also discover minor concepts using the same process but with crop images. + +# 3. Method + +# 3.1. Method Overview + +WWW consists of three modules: Concept discovery, Localization, and Reasoning. The concept discovery module identifies each neuron's concept, which explains part 'what'. It selects neuron concepts from the concept set $(D_{\text{concept}})$ by leveraging adaptive cosine similarity and adaptive selection. The localization module is to identify highly contributed input regions of the test sample, which concept is present at 'where'. Also, the combination of the neuron activation map leveraging Shapley value helps identify 'where' the concept is and tells the 'why' of the individual predictions. The reasoning module identifies important neurons of the test sample and predicted class. By comparing the differences between the sample and class explanations, users can understand the 'why' of the model prediction and, even more, whether the prediction is reliable or not. + +# 3.2. Concept Discovery Module + +The concept discovery module aims to identify proper major and minor concepts from $D_{\text{concept}}$ that match an + +example-based representation of each neuron. Let target model $f$ d, and let $(l,i)$ as $i$ -th neuron in layer $l$ of the target model. From the images in probing dataset $D_{probe}$ , we select high activating images for neuron $(l,i)$ as $D_{rep}^{(l,i)}$ . With selected images in $D_{rep}^{(l,i)}$ and concepts in $D_{concept}$ , we calculate CLIP visual feature of $D_{rep}^{(l,i)}$ as $V^{(l,i)} = \left[v_1^{(l,i)}, v_2^{(l,i)}, \dots, v_n^{(l,i)}\right]$ and CLIP text feature of $D_{concept}$ as $T = [t_1, t_2, \dots, t_m]$ . $n$ denotes number of example images for neuron $(l,i)$ and $m$ stands for the number of concept in concept set $D_{concept}$ . + +We calculate concept score $s^{(l,i)}$ by calculating adaptive cosine similarity (ACS), which allows us to reduce the effect of the base template and only consider the similarity between the image and the concept itself. Concept score $s_{j}^{(l,i)}$ for $j$ -th concept $t_j$ of neuron $(l,i)$ is calculated as follows: + +$$ +s _ {j} ^ {(l, i)} = \frac {1}{n} \sum_ {o = 1} ^ {n} \left\{\cos \left(v _ {o} ^ {(l, i)}, t _ {j}\right) - \cos \left(v _ {o} ^ {(l, i)}, t _ {t e m}\right) \right\}, \tag {1} +$$ + +where $1 \leq j \leq m$ , $t_{tem}$ denotes CLIP text feature of base template(e.g. 'a photo of') and $\cos(x, y)$ denotes cosine similarity between $x$ and $y$ . + +![](images/d04ee3b5409b6957433008bf52fbadf8ad0cb6c743557fbf5ee0b869389d07bc.jpg) +Figure 2. Illustration of overall test time flow of WWW. In the test time (i.e., inference), class explanation selects important neurons with pre-computed Shapley value of the predicted class. On the other hand, the sample explanation selects important neurons with a Shapley value for the input sample. With selected important neurons, pre-computed concepts are annotated. After the concept annotation, a class heatmap is generated with the pre-computed Shapley value of the predicted class. On the other hand, a sample heatmap is generated with the Shapley value of the sample input. + +From the calculated concept scores $s^{(l,i)} = \left[s_1^{(l,i)},\dots ,s_m^{(l,i)}\right]$ , we select concepts where $s_j^{(l,i)} > \delta^{(l,i)}$ . Adaptive selection threshold $\delta^{(l,i)}$ for neuron $(l,i)$ is calculated as follows: + +$$ +\delta^ {(l, i)} = \alpha \times \max (s ^ {(l, i)}), \tag {2} +$$ + +where $\alpha$ denotes concept sensitivity for major and minor concepts. For discovering minor concepts, we select $D_{rep}^{(l,i)}$ with cropped images of probing dataset $D_{probe}$ . + +# 3.3. Localization Module + +The localization module aims to identify highly contributed input regions of each concept and generate a concept map and concept heatmap. We select important neurons with Taylor approximation of Shapley value introduced in [16] for generating a concept region map. For input image sample $x$ , Neuron contribution $w^{(l,i)}$ of neuron $(l,i)$ is calculated as follows: + +$$ +w ^ {(l, i)} (x) = \left| f (x) - f (x; a ^ {(l, i)} \leftarrow 0) \right| = \left| a ^ {(l, i)} \nabla_ {a ^ {(l, i)}} f (x) \right|. \tag {3} +$$ + +Where $a^{(l,i)}$ denotes activation of neuron $(l,i)$ . After calculating each layer's Neuron contribution, we rank neurons by calculated contribution (i.e., Shapley value) and select the top- $k$ important neurons for the sample. The concept heatmap $M$ is calculated with the weighted sum of important neurons' neuron activation maps (NAMs). Concept heatmap $M^l(x)$ of important neurons is calculated as follows: + +$$ +M ^ {l} (x) = \sum^ {u} w ^ {(l, u)} (x) A ^ {(l, u)} (x) \tag {4} +$$ + +where $u$ denotes the index of important neurons, $w$ denotes the neuron contribution, and $A$ denotes the neuron activation map of the neuron. Note that $M^l$ shows only related regions of a single or combination of important neurons, not the whole network. + +# 3.4. Reasoning Module + +The reasoning module is designed to help users understand the 'why' of the model output. This not only explains the result but also includes 'why' this prediction is reliable or not. By leveraging the class-wise Shapley value introduced in [2], we can understand each neuron's class-wise contri + +bution and identify important neurons for each class. From Ahn et al. [2], the class-wise contribution of a neuron can be calculated by the average contribution of a neuron in class samples. From the class-wise contribution, we rank top- $k$ important neurons for each class. We can also generate a class-wise concept region map and concept heatmap with the index of important neurons. Class-wise maps can be used as a guideline for understanding the general case of the prediction. By comparing Class-wise maps and sample maps, users can identify which concept and region of the sample differs from general cases. This naturally helps users understand the 'why' of the model output and also 'why' the results are reliable or not. Figure 2 for further understanding the WWW flow for generating class and sample explanations. Figure 5 shows the generated explanation example. + +# 3.5. Overall flow and Output of WWW + +As described above, WWW has three main modules explaining three 'w' of 'what', 'where', and 'why', respectively. In Section 3.5, we are going to follow the overall flow of our method by time sequence. Before starting inference, WWW needs pre-computing for discovering major and minor concepts with the concept discovery module (Sec 3.2) and also a class-wise contribution for class-wise analysis in the Reasoning Module (Sec 3.4). After pre-computing concepts and contributions, WWW is ready to generate test sample explanations. Figure 2 shows the overall flow of explaining the generation of WWW in the inference time. In the test time (i.e., inference), WWW generates a class-wise explanation of the predicted class and sample explanation into two tracks. WWW leverages the predicted class and class-wise contribution matrix to generate a class-wise explanation to select the most critical neurons for the predicted class. After identifying important neurons, WWW leverages the localization module (Sec 3.3) to generate a class concept region map and class concept combination heatmap. On the other hand, for the sample explanation, WWW calculates the Taylor-approximated Shapley value of the sample for the predicted class and identifies critical neurons for the prediction of the sample. After identifying important neurons for the sample, WWW leverages the sample Shapley value to generate the sample concept region map and sample concept combination heat map. + +# 4. Experiment + +Comprehensive experiments have been conducted to evaluate our method. In section 4.1, we evaluate the performance of the concept discovery module both qualitatively (Sec. 4.1.1) and quantitatively (Sec. 4.1.2). Section 4.2 is an ablation study for the concept discovery module. In Section 4.3, we analyze generated explanations of both correct and wrong predictions with examples. + +![](images/753df40db424fd15b1f841e8e88684930ffa708dcf10ae125ef6c8c63c9fcabd.jpg) + +![](images/b1833a0838b902a2981c64443502d4cc113aa2959862ad74f7930757081e718a.jpg) +ResNet - 50 Layer 4 +Neuron 1152 +WWW (Ours): goldfish - flame fish +CLIP-Dissect: swordtail +MILAN: ORANGE objects FALCON: TRICK, CALE +FALCON:Insh,sale,series,fare +Neuron 816 +CLIP-Dissect: genus Chaetodon + +![](images/b5ce624862242000be52437a649e115e43d3d4fe5ca4ddfde5852d1f553ef214.jpg) +WWW (Ours): saltwater fish - swordtail +MILAN: Aquatic life + +![](images/a722818b2a6fcdedf84eccc43634f97f7b2039eb7cabe1612166beb09f42a729.jpg) +FALCON: cover, black, white, home, blue +ResNet - 50 Final layer +Neuron 1 +Ground Truth: goldfish +www +CLIP-Dissect: swordtail +MILAN: Nature +FALCON: trish, size, storage assorted, gap + +![](images/f58cb48ab48c3973fe85d6f598d284e057f1b1d629b63eefc161804b9c3fe750.jpg) +Neuron 2018 +WWW (Ours): Dutch oven - stewing pan +CLIP-Dissect: electric frying pan +Neuron 1911 +CLIP-Dissect: cookstove + +![](images/cac259dfc7409f347023906ad41d2053160759e097e650a6dab4ba1ef3d200c9.jpg) +MILAN: Rounded edges in pictures +FALCON: black, card, woman, home, white +Neuron 544 +Ground Truth: Dutch o +WWW (Ours): Dutch oven +CLIP-Dissect: stewing pan +MILAN: Circular objects +FACLCON:men top game +award, guard +WWW (Ours): cookfire - cookfire +MILAN: Animals and circular objects +FALCON: shirt, school, mini, woman, house +Figure 3. Qualitative comparison of WWW with other baselines. We compared WWW with three competitive baselines (CLIP-Dissect [20], MILAN [12], FALCON [14]) in two final layer neurons and four penultimate layer (i.e., layer 4) neurons with each neuron's highly activating images. layer-4 neurons are top-2 important neurons of the final layer class. We have colored the descriptions green if they match the images, yellow if they match but are too generic or similar, and red if they do not match. + +# 4.1. Performance Evaluation for Concept Module + +In this experiment, we evaluate the concept matching performance of the concept discovery module with four other baselines, Network Dissection [3], MILAN [12], CLIP-Dissect [20] and FALCON [14]. We evaluate the performance of methods on the various models (e.g., ResNet-18 [11], ResNet-50 [11], and ViT-B/16 [5]), various probing datasets (e.g.,Imagenet, Places365) and various concept sets (e.g., Wordnet nouns, and labels of Places365, Broaden, and ImageNet) + +# 4.1.1 Qualitative Results + +Settings. We compared WWW with the three most comparable methods (CLIP-Dissect [20], MILAN [12], FALCON [14]) in the penultimate layer and final layer of the model. We do not compare with Network Dissection [3] due to the limitation that probe image data $D_{probe}$ and concept set $D_{concept}$ is fixed to Broaden. We used a ResNet-50 [11] model pre-trained in the ImageNet-1k [22] dataset. For probe image data $D_{probe}$ , we used the ImageNet-1k validation set, and we extracted all nouns in Wordnet[8] (about 80k) dataset for a concept set $D_{concept}$ . + +Results. Figure 3 shows examples of descriptions for hidden neurons in the penultimate and final layers. Neurons in the penultimate layer are top-2 important neurons of the final layer neuron's ground truth label class. We observed that WWW not only interpreted each neuron well but also showed robust interpretation that the most important neuron of the class in the penultimate layer represents the same major concept as the final layer neuron. + +Table 2. Quantitative comparison on final layer concept matching performance of ResNet-50 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for $D_{probe}$ . Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported. + +
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
Network Dissection [3]BrodenBroden(1.2k)0.7229 ± 0.0030.2989 ± 0.0050.0010 ± 0.0010.001
MILAN(b) [12]ImageNet val-0.7300 ± 0.0030.2485 ± 0.0050.0005 ± 0.0000.001
FALCON [14]ImageNet valLAION-400m0.7065 ± 0.0030.1790 ± 0.0010.0002 ± 0.0000.001
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9340 ± 0.0030.8376 ± 0.0060.7286 ± 0.0090.933
ImageNet valBroden (1.2k)0.7369 ± 0.0030.3432 ± 0.0040.0328 ± 0.0030.108
ImageNet valWordnet (80k)0.8689 ± 0.0040.6846 ± 0.0080.3647 ± 0.0140.456
WWW (Ours)ImageNet valImageNet (1k)0.9325 ± 0.0030.8327 ± 0.0060.7719 ± 0.0090.955
ImageNet valBroden (1.2k)0.7758 ± 0.0040.4414 ± 0.0070.0645 ± 0.0070.091
ImageNet valWordnet (80k)0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.0120.645
+ +Table 3. Quantitative comparison on final layer concept matching performance of ViT-B/16 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for $D_{probe}$ . Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported. + +
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9337 ± 0.0030.8375 ± 0.0060.7289 ± 0.0090.933
ImageNet valBroden (1.2k)0.7365 ± 0.0030.3416 ± 0.0040.0319 ± 0.0030.106
ImageNet valWordnet (80k)0.8700 ± 0.0040.6886 ± 0.0080.3679 ± 0.0140.460
WWW (Ours)ImageNet valImageNet (1k)0.9331 ± 0.0030.8347 ± 0.0060.7718 ± 0.0090.955
ImageNet valBroden (1.2k)0.7754 ± 0.0040.4407 ± 0.0070.0651 ± 0.0070.091
ImageNet valWordnet (80k)0.8857 ± 0.0030.6970 ± 0.0080.4166 ± 0.0120.634
+ +# 4.1.2 Quantitative Results + +In this section, we compare the performance of our methods with baselines. As introduced in [20], we evaluate final layer neuron concepts with the class labels with various metrics. By comparing generated explanations with the class labels, we can objectively evaluate the quality of the generated neuron labels with what each neuron is trained to represent. + +Metrics. CLIP cos and mpnet cos are measured as cosine similarities between the encoded feature of the class label and selected concepts with CLIP [21] model and mpnet [26] model, respectively. F1-score is measured to evaluate the discovered concept's balance of exactness and flexibility. Also, the hit rate is calculated as a rate of selected concepts that exactly match the class label. + +Quantitative comparison on ResNet-50. Table 2 compares WWW with Network Dissection [3], MILAN(b) [12], FALCON [14], and CLIP-Dissect [20]. We evaluate the final layer neuron concepts of ResNet-50 with the class labels of ImageNet-1k. In table 2, WWW showed better performance as the concept set $D_{\text{concept}}$ gets larger and outperformed all other baselines when concept set $D_{\text{concept}}$ is set to Broden and Wordnet nouns. In the results of Wordnet nouns ( $D_{\text{concept}}$ ), comparison between WWW and CLIP-Dissect [20] are statistically significant across all metrics + +$(p < 0.05)$ + +Quantitative comparison on ViT-B/16. We compared predicted concepts of the final layer to ground truth labels of ViT-B/16 pre-trained on ImageNet. In Table 3, WWW showed better performance as the concept set $D_{\text{concept}}$ gets larger and outperformed other baselines when concept set $D_{\text{concept}}$ is set to Broden and Wordnet nouns. + +Quantitative comparison on ResNet-18 pre-trained in Places365. We compared predicted labels to ground truth labels in final layer neurons of ResNet-18 pre-trained on Places365. In table 4, WWW showed better performance as the concept set $D_{\text{concept}}$ gets larger and outperformed other baselines when concept set $D_{\text{concept}}$ is set to Broden and Wordnet nouns. + +# 4.2. Ablation study + +Evaluation on the effect of leveraging base template and ACS. Table 5 shows an ablation study over various parts used in WWW. Without using a base template, WWW showed the lowest performance overall. With the use of a base template, WWW shows slightly increased performance in CLIP cos and F1-score but not much advance overall. However, performance was significantly increased, and the highest scores in all three metrics were achieved when leveraging ACS. The findings suggest that while us- + +Table 4. Quantitative comparison on final layer concept matching performance of ResNet-18 trained on Places365. We compared predicted neuron concepts with ground truth labels of Places365. We used the Places365 test set for $D_{probe}$ . Bold numbers represent the best scores between the same settings. The average score and standard errors of the 365 final layer neurons are reported. + +
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]Places365 testPlaces365 (0.4k)0.9562 ± 0.0040.8687 ± 0.0120.7233 ± 0.0230.723
Places365 testBroden (1.2k)0.8304 ± 0.0030.4678 ± 0.0070.1096 ± 0.0080.329
Places365 testWordnet (80k)0.8378 ± 0.0060.5107 ± 0.0140.1041 ± 0.0160.104
WWW (Ours)Places365 testPlaces365 (0.4k)0.9402 ± 0.0040.8204 ± 0.0130.6361 ± 0.0240.660
Places365 testBroden (1.2k)0.8925 ± 0.0050.6415 ± 0.0130.2242 ± 0.0210.255
Places365 testWordnet (80k)0.8492 ± 0.0040.5000 ± 0.0130.1106 ± 0.0150.142
+ +Table 5. Ablation study on the use of the template and ACS. '√' in the Template represents WWW using a template (e.g., 'a photo of word') '√' in the ACS represents WWW using adaptive cosine similarity. + +
TemplateACSCLIP cosmpnet cosF1-score
0.8499 ± 0.0030.6123 ± 0.0070.3265 ± 0.009
✓0.8547 ± 0.0030.6075 ± 0.0070.3361 ± 0.009
✓✓0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.012
+ +ing a base template in the CLIP model offers some benefits, it also has its limitations, particularly in terms of embedding concepts within a similar CLIP feature space. In contrast, ACS appears to reduce the uniformity between concepts, allowing the concept discovery module to discover more distinct and accurate concepts for each neuron representation. This highlights the effectiveness of ACS in enhancing the overall performance of the concept discovery module. Ablation study on Concept Sensitivity $(\alpha)$ . The ablation study depicted on the left side of Figure 4 examined the impact on the F1-score by varying levels of major concept sensitivity (denoted as $\alpha$ ). We found that as concept sensitivity decreases, the F1-score initially increases $(\alpha > 0.95)$ and then decreases. This pattern is due to a trade-off. Higher concept sensitivity leads to more precise concept identification but at the cost of identifying fewer concepts. Conversely, lower sensitivity results in more concepts being identified that may be less similar to the target concept. The point at which the F1-score is maximized can be seen as the optimal balance in this trade-off, providing a guideline for setting the most effective level of concept sensitivity for the major concept discovery. + +# 4.3. Discussion + +# 4.3.1 Overall Explanation Generated by WWW + +Figure 5 illustrates an explanation example generated by WWW, where the left side of the figure provides a class explanation for a test sample by highlighting important neurons for the predicted class, and the right side depicts a sample explanation showing the important neurons identi + +![](images/41abbd05d7ff74a9a13265c79cf6a52df4e98f8828677d3ec3efc0840011f7c1.jpg) +Figure 4. Ablation of concept sensitivity and heatmap similarity feasibility result. Left figure illustrates the F1 score with respect to Concept Sensitivity $(\alpha)$ changes. Concept sensitivity $(\alpha)$ that maximizes the F1-score is illustrated as the red line. The right figure illustrates the rejection test result of heatmap similarity and maximum softmax probability (MSP). # of hit denotes the number of correctly detected samples as a misprediction. + +![](images/557ea332b44f6e5abb6c279d82309619a8e74eed22ebfb0dd5cc8e67bd31be50.jpg) + +![](images/fcd3a6304ac012c6e51a47bb15f0a9de8ece2ad445a2e3ac85310b692351500a.jpg) +Figure 5. Example of generated explanation by WWW. From top to bottom, important neurons are displayed in the order of importance. Images in each neuron are examples of a major and minor concept, respectively. Colors in the top localization image show highly related regions for each concept. The bottom localization image is a weighted sum of important neuron activation maps displayed as a heatmap. + +fied for the test sample image prediction. This explanation target neurons from the penultimate layer of a ResNet-50 model pre-trained on ImageNet, using the ImageNet validation set $(D_{probe})$ and Wordnet nouns $(D_{concept})$ for the + +explanation. Notably, neuron 30 is consistently identified as the most crucial in both explanations. It represents the 'tench' concept, which matches the prediction and the test sample's ground truth. Interestingly, while neuron 1246 is the third most important for the concept, it is the second most significant for the test sample's prediction, and neuron 1480, not highlighted in the class explanation, emerges as the third important neuron in the sample explanation. Despite these neuron selection and ranking variations, the model accurately predicts the correct class. The overall heatmap of the class and sample explanations can explain this. The heatmap explanation reveals a high cosine similarity between the two heatmaps, indicating that the calculated weighted sum of NAMs is similar. Despite the difference in individual neuron importance, the highly related input region is remarkably similar in both cases. + +# 4.3.2 Anlaysis of Explanation of Failure Case + +Figure 6 provides an example of a mispredicted sample explanation generated by the WWW. In this particular case of failure, not only do the selected important neurons differ between the class and sample explanations, but the cosine similarity between their respective heatmaps is relatively low, measuring at 0.19. Interestingly, even though the ground-truth class explanations and the sample explanations highlight different important neurons, they both localize to similar regions in the heatmap, showing a relatively high similarity score of 0.47. From this observation, we explore the potential utility of heatmap similarity to predict uncertainty. On the right side of Figure 4, we present the results of a rejection test that evaluates heatmap similarity and the maximum softmax probability (MSP). When the rejection method detects the mispredicted sample correctly, we consider that as a hit. The number of hit is measured with respect to the rejected sample rate. As the rejection rate increases based on their uncertainty level, it becomes evident that heatmap similarity outperforms MSP regarding the detection of misprediction samples. This suggests that heatmap similarity can serve as a more effective measure of uncertainty compared to MSP. These findings indicate that heatmap similarity can be used as a tool for predicting uncertainty. + +# 5. Conclusion + +We proposed WWW, a unified framework that provides comprehensive explanations for the 'what', 'where', and 'why' of neural network decisions. WWW demonstrates superior performance in both quantitative and qualitative measures, offering a deeper and more detailed understanding of neural network behavior. This is achieved through a novel integration of adaptive selection for concept discovery, neuron activation maps, and Shapley values. WWW's + +![](images/831385207359ed7e0d6f249e7eb3bc7cd0824bb87729be4ffe7505aa3f878c67.jpg) +Figure 6. Example of failure case explanation by WWW. The explanations of the predicted label are presented on the left side. On the right side, the explanation of the ground-truth label is shown. In the upper half, we displayed concept attribution maps, which show important concepts and their respective regions. In the bottom half, we showed the overall heatmap which shows important regions for the model decision. + +adaptability is also shown across various neural network architectures, including convolutional networks and attention-based Vision Transformers. Additionally, our approach to predicting uncertainty through heatmap similarity analysis introduces a new way to obtain the reliability of their predictions. By offering localized explanations with concept annotations for individual inputs, WWW enhances the transparency of the model's decision-making process, contributing to the broader goal of making AI more reliable and trustworthy. + +# Acknowledgements + +This work was supported in part by the National Research Foundation of Korea (NRF) Grant funded by the Korea Government (MSIT) under Grant 2021R1G1A1094990, by the Institute of Information and Communications Technology Planning and Evaluation (IITP) Grant funded by the Korea Government (MSIT) under Grant 2022-0-00078 (Explainable Logical Reasoning for Medical Knowledge Generation), Grant IITP-2024-RS-2023-00258649 (ITRC(Information Technology Research Center) support program), Grant 2021-0-02068 (Artificial Intelligence Innovation Hub), Grant RS-2022-00155911 (Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)), and by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD). + +# References + +[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 2 +[2] Yong Hyun Ahn, Gyeong-Moon Park, and Seong Tae Kim. Line: Out-of-distribution detection by leveraging important neurons. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19852-19862, 2023. 2, 4, 5 +[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 5, 6 +[4] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[6] Thomas Fel, Victor Boutin, Mazda Moayeri, Rémi Cadène, Louis Bethune, Léo andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. arXiv preprint arXiv:2306.07304, 2023. 2 +[7] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2 +[8] Christiane Fellbaum. Wordnet and wordnets. encyclopedia of language and linguistics, 2005. 5 +[9] Ruth Fong and Andrea Vedaldi. Net2vec: Quantifying and explaining how concepts are encoded by filters in deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8730-8738, 2018. 2 +[10] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proceedings of the European conference on computer vision, pages 630-645. Springer, 2016. 5 +[12] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2, 5, 6, 1 + +[13] Alon Jacovi, Ana Marasović, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1 +[14] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. International Conference on Machine Learning, 2023. 1, 2, 5, 6 +[15] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1 +[16] Ashkan Khakzar, Soroosh Baselizadeh, Saurabh Khanduja, Christian Rupprecht, Seong Tae Kim, and Nassir Navab. Neural response interpretation through the lens of critical pathways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13528-13538, 2021. 2, 4 +[17] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2 +[18] Mauritz Kop. Eu artificial intelligence act: The european approach to ai. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust ..., 2021. 1 +[19] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2 +[20] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-dissect: Automatic description of neuron representations in deep vision networks. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 7 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6 +[22] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5 +[23] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2 +[24] Lloyd S Shapley. A value for n-person games. Classics in game theory, 69, 1997. 2 +[25] Michael C Sloan. Aristotle's nicomachean ethics as the original locus for the septem circumstantiae. Classical Philology, 105(3):236-251, 2010. 1 +[26] K. Song et al. Mpnet: Masked and permuted pre-training for language understanding. NeurIPS, 33:16857-16867, 2020. 6 + +[27] Andong Wang, Wei-Ning Lee, and Xiaojuan Qi. Hint: Hierarchical neuron concept explainer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10254–10264, 2022. 2 +[28] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 2 +[29] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 1 \ No newline at end of file diff --git a/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/images.zip b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3838349673aa43f1214f15810ed6369970a91071 --- /dev/null +++ b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29647c4305f74d29e898b81e117e19b9762a4d97546290bbe72b75ca75d630ae +size 611163 diff --git a/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/layout.json b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..54aef814df29c3c70539f6a7cce24dd2f800bb41 --- /dev/null +++ b/2024/WWW_ A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts/layout.json @@ -0,0 +1,7898 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "type": "text", + "content": "WWW: A Unified Framework for Explaining What, Where and Why of Neural Networks by Interpretation of Neuron Concepts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "spans": [ + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "text", + "content": "Yong Hyun Ahn," + }, + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Hyeon Bae Kim," + }, + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Seong Tae Kim" + }, + { + "bbox": [ + 160, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 175, + 525, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 175, + 525, + 203 + ], + "spans": [ + { + "bbox": [ + 68, + 175, + 525, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 175, + 525, + 203 + ], + "type": "text", + "content": "Department of Artificial Intelligence, Kyung Hee University, Republic of Korea " + }, + { + "bbox": [ + 68, + 175, + 525, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 68, + 175, + 525, + 203 + ], + "type": "text", + "content": "Department of Computer Science and Engineering, Kyung Hee University, Republic of Korea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 205, + 402, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 205, + 402, + 217 + ], + "spans": [ + { + "bbox": [ + 190, + 205, + 402, + 217 + ], + "type": "text", + "content": "{yhahn, hyeonbae.kim, st.kim}@khu.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 270, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 270, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 45, + 270, + 290, + 547 + ], + "type": "text", + "content": "Recent advancements in neural networks have showcased their remarkable capabilities across various domains. Despite these successes, the \"black box\" problem still remains. To address this, we propose a novel framework, WWW, that offers the 'what', 'where', and 'why' of the neural network decisions in human-understandable terms. Specifically, WWW utilizes adaptive selection for concept discovery, employing adaptive cosine similarity and thresholding techniques to effectively explain 'what'. To address the 'where' and 'why', we proposed a novel combination of neuron activation maps (NAMs) with Shapley values, generating localized concept maps and heatmaps for individual inputs. Furthermore, WWW introduces a method for predicting uncertainty, leveraging heatmap similarities to estimate the prediction's reliability. Experimental evaluations of WWW demonstrate superior performance in both quantitative and qualitative metrics, outperforming existing methods in interpretability. WWW provides a unified solution for explaining 'what', 'where', and 'why', introducing a method for localized explanations from global interpretations and offering a plug-and-play solution adaptable to various architectures. Code is available at: https://github.com/ailab-kyunghee/WWW" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "content": "Neural networks have demonstrated impressive performance in various fields in recent years. Despite these successes, their widespread adoption in more diverse areas is slowed by several challenges. A fundamental issue is the \"black box\" problem, referring to the often hidden and unclear decision-making processes of neural network models. This lack of clarity raises concerns about the stability and reliability of these models, leading to a growing consensus that artificial intelligence should be reliable, robust," + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 319, + 533, + 388 + ], + "blocks": [ + { + "bbox": [ + 306, + 244, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 306, + 244, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 244, + 547, + 312 + ], + "type": "text", + "content": "Table 1. Illustration of What, Where, and Why of recent concept-based neural network interpretation methods. Green, yellow, and red marks illustrate that the method is able to interpret well, is partially interpretable, and has limitations for interpretation, respectively. Recent methods are able to interpret one or two 'w's but have limitations regarding interpreting three 'w's at once." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 319, + 533, + 388 + ], + "lines": [ + { + "bbox": [ + 309, + 319, + 533, + 388 + ], + "spans": [ + { + "bbox": [ + 309, + 319, + 533, + 388 + ], + "type": "table", + "html": "
MethodWhatWhereWhy
CLIP-Dissect [20] (ICLR'23)✓XX
FALCON [14] (ICML'23)✓X▲
CRAFT [7] (CVPR'23)X✓▲
WWW (Ours)✓✓✓
", + "image_path": "5070b4a805a306447b80efe7e8d43bd62e11e36df8eca54b98cfeee5e4ce4924.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 410, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 546, + 458 + ], + "type": "text", + "content": "and safe [4, 13]. As a response to this need for trustworthy AI, there has been an emergence of laws and regulations [15, 18] that require neural networks to base their decisions on principles that are understandable to humans." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 460, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 688 + ], + "type": "text", + "content": "According to Doshi-Velez et al. [4], interpretability is defined as the capability to provide explanations in terms that are understandable to humans. Zhang et al. [29] expand on this definition, emphasizing that interpretability involves providing explanations in understandable terms to humans. This requirement for explanations in human-understandable terms is a consistent theme in the literature on interpretable methods. Furthermore, this concept aligns with a fundamental principle in journalism and problem-solving: the importance of clear and understandable communication. This principle is often encapsulated in the five Ws (Who, What, When, Where, Why) [25]. Building on this, we propose that practical explanations or interpretations should include three critical elements: 'what' (the nature of the decision or outcome), 'where' (the context or specific region of the input that affects the decision-making process), and 'why' (the reasoning or factors behind a decision). This approach aims to make the interpretations more accessible and relevant to human understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "From this point of view, former works can be re-grouped into several groups which those works aim to solve. For ex-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10968" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "ample, a series of works in feature attribution [10, 23, 28] can be classified as works mainly focusing on 'where'. These methods explain model decisions by identifying highly contributed input regions. However, identifying highly related input regions primarily focuses on 'where' only, not enough explanation for 'what' and 'why'." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 286, + 289 + ], + "type": "text", + "content": "On the other hand, concept-based explanations mainly focus on 'why'. Concept-based explanations [6, 7, 10, 17] explain model decisions by decomposing model responses into smaller units called concepts. Kim et al. proposed TCAV [17], understanding and interpreting model decisions into a form of concept activation vectors. Recently, Fel et al. proposed CRAFT [7], which utilizes non-negative matrix factorization for identifying concept vectors. The works mentioned above mainly decompose activations into the form of vectors, which gives a great advantage for finding out 'why' but often fails to match or annotate the human-understandable term (i.e., name) that is needed for 'what'." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 286, + 458 + ], + "type": "text", + "content": "Other streams of work, called Neuron-concept association, aim to name what each neuron (e.g. convolution filters, layer outputs) represents, which answers 'what' the model cares about. Bau et al. proposed Network Dissection [3], which identifies each neuron's representing concept from Broaden dataset [3]. Recently, CLIP-Dissect [20] and FALCON [14] have shown that representing the concept of neurons can be matched automatically by using a pre-trained CLIP model [21]. The aforementioned methods are suitable for explaining 'what' but partially explain or have limitations for explaining 'where' and 'why'. Furthermore, these methods generate global explanations focusing on each neuron's role in the network, not the local explanations for each sample input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 639 + ], + "type": "text", + "content": "To address these issues, we propose a novel framework that can explain 'What', 'Where', and 'Why' (WWW) at once. WWW introduces adaptive selection for discovering each neuron's concept and interpreting each neuron to explain 'what'. By leveraging adaptive cosine similarity (ACS) and adaptive selection techniques, we achieve advanced performance compared to competitive methods. Moreover, we combined neuron activation map (NAM) and Shapley value [2, 16, 24] to generate class and sample concept maps and heatmaps to explain 'where' and 'why'. We also conducted various objective evaluations to assess the performance of the proposed method. In the experiments, WWW achieves better results in both quantitative and qualitative evaluations. Our key contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 286, + 712 + ], + "type": "text", + "content": "- We introduce a novel and effective way to generate high-quality explanations that explain 'what', 'where', and 'why' at once. Due to the powerful performance of the adaptive selection for concept discovery, WWW is able to achieve higher quantitative and better qualitative results in various metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "- We introduce a way to generate localized explanations from global neural network interpretation. By the novel combination of Shapley value and neuron activation maps, WWW is able to generate localized explanations with concept annotations for sample input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 132, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 545, + 179 + ], + "type": "text", + "content": "- WWW can be attached to various target models with different architectures, from conventional convolution neural networks to the recent attention-based Vision transformers, in a plug-and-play manner." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 196, + 397, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 196, + 397, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 397, + 209 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 217, + 462, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 462, + 230 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 462, + 230 + ], + "type": "text", + "content": "2.1. Neuron-Concept Association" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 237, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 536 + ], + "type": "text", + "content": "Bau et al. introduced Network Dissection [3], using the Broden dataset to identify which concepts individual neurons in a network represent. They use overlap between segmentation masks and feature maps to annotate concepts for neurons. Fong and Vedaldi expanded on this with Net2Vec [9], which looks at individual neurons and their combinations. Mu and Andreas further extended these ideas with Compositional Explanation [19], aiming to generate more complex and detailed explanations. These methods primarily focused on understanding 'what' a neuron represents. However, the aforementioned methods fell short in explaining the 'where' and 'why' of neuron representations. Additionally, they relied heavily on image-concept-matched datasets like Broden, which are often costly and hard to collect due to the need for pixel-wise labels. To address these limitations, recent approaches like HINT [27] and MILAN [12] have been developed. These methods train concept classifiers or models that reduce the dependency on image-concept-matched datasets. Moreover, approaches like CLIP-Dissect [20] and FALCON [14] leverage the pretrained CLIP model to use separate sets of image datasets and concept datasets. Despite these advancements in understanding 'what' neurons represent, there remains a gap in fully explaining the 'where' and 'why' of the neuron representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 549, + 449, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 449, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 449, + 562 + ], + "type": "text", + "content": "2.2. Vector-based Explanation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "The field of interpretability has significantly advanced, particularly in understanding and interpreting model decisions using Concept Activation Vectors (CAVs). Kim et al. introduced TCAV [17], understanding and interpreting model decisions into the form of CAVs. Recently, CRAFT [7] was introduced, leveraging non-negative matrix factorization to identify CAVs and localize the most relevant input regions for each CAV. Also, Achtibat et al. [1] and Fel et al. [6] proposed methods to interpret each CAV's role in decision-making. Despite these developments, a key challenge remains: translating CAVs into human-understandable terms." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10969" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 527, + 340 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 527, + 340 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 527, + 340 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 527, + 340 + ], + "type": "image", + "image_path": "bf758565ed0917fa187b3806594c33d01b812147cbee75e8c2ca58e3bf17fb5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "lines": [ + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "text", + "content": "Figure 1. Overall flow of Concept Discovery module identifying concepts for a single neuron. We first calculate the cosine similarity of CLIP features between images and concepts with the template from the selected high-activating images. Then, we subtract the cosine similarity of CLIP features between images and the base template by only considering the similarity between the concept and image. From calculated adaptive cosine similarity (ACS), we generate concept score " + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "text", + "content": " by the average similarity of images. Note that concept score " + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "inline_equation", + "content": "S = \\{s_1, s_2, \\dots, s_m\\}" + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "text", + "content": " are a group of scores, not a single scalar. From the calculated concept scores " + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 348, + 547, + 416 + ], + "type": "text", + "content": ", we select major concepts by adaptive selection. We also discover minor concepts using the same process but with crop images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 434, + 104, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 104, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 104, + 445 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 154, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 154, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 154, + 465 + ], + "type": "text", + "content": "3.1. Method Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 290, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 290, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 290, + 663 + ], + "type": "text", + "content": "WWW consists of three modules: Concept discovery, Localization, and Reasoning. The concept discovery module identifies each neuron's concept, which explains part 'what'. It selects neuron concepts from the concept set " + }, + { + "bbox": [ + 46, + 472, + 290, + 663 + ], + "type": "inline_equation", + "content": "(D_{\\text{concept}})" + }, + { + "bbox": [ + 46, + 472, + 290, + 663 + ], + "type": "text", + "content": " by leveraging adaptive cosine similarity and adaptive selection. The localization module is to identify highly contributed input regions of the test sample, which concept is present at 'where'. Also, the combination of the neuron activation map leveraging Shapley value helps identify 'where' the concept is and tells the 'why' of the individual predictions. The reasoning module identifies important neurons of the test sample and predicted class. By comparing the differences between the sample and class explanations, users can understand the 'why' of the model prediction and, even more, whether the prediction is reliable or not." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 197, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 197, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 197, + 684 + ], + "type": "text", + "content": "3.2. Concept Discovery Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "The concept discovery module aims to identify proper major and minor concepts from " + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": " that match an" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": "example-based representation of each neuron. Let target model " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " d, and let " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": "-th neuron in layer " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " of the target model. From the images in probing dataset " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": ", we select high activating images for neuron " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{rep}^{(l,i)}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": ". With selected images in " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{rep}^{(l,i)}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " and concepts in " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{concept}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": ", we calculate CLIP visual feature of " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{rep}^{(l,i)}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "V^{(l,i)} = \\left[v_1^{(l,i)}, v_2^{(l,i)}, \\dots, v_n^{(l,i)}\\right]" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " and CLIP text feature of " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{concept}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "T = [t_1, t_2, \\dots, t_m]" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " denotes number of example images for neuron " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": " stands for the number of concept in concept set " + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "inline_equation", + "content": "D_{concept}" + }, + { + "bbox": [ + 304, + 435, + 545, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "content": "We calculate concept score " + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "inline_equation", + "content": "s^{(l,i)}" + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "content": " by calculating adaptive cosine similarity (ACS), which allows us to reduce the effect of the base template and only consider the similarity between the image and the concept itself. Concept score " + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "inline_equation", + "content": "s_{j}^{(l,i)}" + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "content": "-th concept " + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "content": " of neuron " + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 304, + 567, + 545, + 640 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 643, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 643, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 315, + 643, + 545, + 673 + ], + "type": "interline_equation", + "content": "s _ {j} ^ {(l, i)} = \\frac {1}{n} \\sum_ {o = 1} ^ {n} \\left\\{\\cos \\left(v _ {o} ^ {(l, i)}, t _ {j}\\right) - \\cos \\left(v _ {o} ^ {(l, i)}, t _ {t e m}\\right) \\right\\}, \\tag {1}", + "image_path": "1bb6a750797c7b99373abfbce8073eb5f55211ac42fcf1ffcfc2baa8dbdb3c6d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq m" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "t_{tem}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " denotes CLIP text feature of base template(e.g. 'a photo of') and " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\cos(x, y)" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " denotes cosine similarity between " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10970" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 358 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 358 + ], + "type": "image", + "image_path": "d04ee3b5409b6957433008bf52fbadf8ad0cb6c743557fbf5ee0b869389d07bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 364, + 547, + 420 + ], + "lines": [ + { + "bbox": [ + 46, + 364, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 547, + 420 + ], + "type": "text", + "content": "Figure 2. Illustration of overall test time flow of WWW. In the test time (i.e., inference), class explanation selects important neurons with pre-computed Shapley value of the predicted class. On the other hand, the sample explanation selects important neurons with a Shapley value for the input sample. With selected important neurons, pre-computed concepts are annotated. After the concept annotation, a class heatmap is generated with the pre-computed Shapley value of the predicted class. On the other hand, a sample heatmap is generated with the Shapley value of the sample input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "text", + "content": "From the calculated concept scores " + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "inline_equation", + "content": "s^{(l,i)} = \\left[s_1^{(l,i)},\\dots ,s_m^{(l,i)}\\right]" + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "text", + "content": ", we select concepts where " + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "inline_equation", + "content": "s_j^{(l,i)} > \\delta^{(l,i)}" + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "text", + "content": ". Adaptive selection threshold " + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\delta^{(l,i)}" + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "text", + "content": " for neuron " + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 47, + 439, + 287, + 495 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 504, + 287, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 504, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 114, + 504, + 287, + 519 + ], + "type": "interline_equation", + "content": "\\delta^ {(l, i)} = \\alpha \\times \\max (s ^ {(l, i)}), \\tag {2}", + "image_path": "6051b62c70b64558bb720c44c60c1b3bef39c06e4a837a3e2c4cfedbaa8ce99b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "text", + "content": " denotes concept sensitivity for major and minor concepts. For discovering minor concepts, we select " + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "inline_equation", + "content": "D_{rep}^{(l,i)}" + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "text", + "content": " with cropped images of probing dataset " + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 46, + 529, + 287, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 167, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 167, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 167, + 586 + ], + "type": "text", + "content": "3.3. Localization Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "content": "The localization module aims to identify highly contributed input regions of each concept and generate a concept map and concept heatmap. We select important neurons with Taylor approximation of Shapley value introduced in [16] for generating a concept region map. For input image sample " + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "content": ", Neuron contribution " + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "inline_equation", + "content": "w^{(l,i)}" + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "content": " of neuron " + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 683, + 291, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 291, + 714 + ], + "type": "interline_equation", + "content": "w ^ {(l, i)} (x) = \\left| f (x) - f (x; a ^ {(l, i)} \\leftarrow 0) \\right| = \\left| a ^ {(l, i)} \\nabla_ {a ^ {(l, i)}} f (x) \\right|. \\tag {3}", + "image_path": "d030a533cba734e5c1e87d9bb31f4ca7c70aa977f72e35741e59e9cc8ca944b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "inline_equation", + "content": "a^{(l,i)}" + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "content": " denotes activation of neuron " + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "inline_equation", + "content": "(l,i)" + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "content": ". After calculating each layer's Neuron contribution, we rank neurons by calculated contribution (i.e., Shapley value) and select the top- " + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "content": " important neurons for the sample. The concept heatmap " + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "content": " is calculated with the weighted sum of important neurons' neuron activation maps (NAMs). Concept heatmap " + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "inline_equation", + "content": "M^l(x)" + }, + { + "bbox": [ + 304, + 440, + 545, + 534 + ], + "type": "text", + "content": " of important neurons is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 536, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 536, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 359, + 536, + 545, + 559 + ], + "type": "interline_equation", + "content": "M ^ {l} (x) = \\sum^ {u} w ^ {(l, u)} (x) A ^ {(l, u)} (x) \\tag {4}", + "image_path": "c8141b6d0386f35894e392ede50777f5d2f67ae2bb99a8b77016fe30223d2d1e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " denotes the index of important neurons, " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " denotes the neuron contribution, and " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " denotes the neuron activation map of the neuron. Note that " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " shows only related regions of a single or combination of important neurons, not the whole network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 416, + 647 + ], + "type": "text", + "content": "3.4. Reasoning Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "The reasoning module is designed to help users understand the 'why' of the model output. This not only explains the result but also includes 'why' this prediction is reliable or not. By leveraging the class-wise Shapley value introduced in [2], we can understand each neuron's class-wise contri" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10971" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "content": "bution and identify important neurons for each class. From Ahn et al. [2], the class-wise contribution of a neuron can be calculated by the average contribution of a neuron in class samples. From the class-wise contribution, we rank top-" + }, + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "content": " important neurons for each class. We can also generate a class-wise concept region map and concept heatmap with the index of important neurons. Class-wise maps can be used as a guideline for understanding the general case of the prediction. By comparing Class-wise maps and sample maps, users can identify which concept and region of the sample differs from general cases. This naturally helps users understand the 'why' of the model output and also 'why' the results are reliable or not. Figure 2 for further understanding the WWW flow for generating class and sample explanations. Figure 5 shows the generated explanation example." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 270, + 232, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 232, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 232, + 283 + ], + "type": "text", + "content": "3.5. Overall flow and Output of WWW" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 289, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 289, + 601 + ], + "type": "text", + "content": "As described above, WWW has three main modules explaining three 'w' of 'what', 'where', and 'why', respectively. In Section 3.5, we are going to follow the overall flow of our method by time sequence. Before starting inference, WWW needs pre-computing for discovering major and minor concepts with the concept discovery module (Sec 3.2) and also a class-wise contribution for class-wise analysis in the Reasoning Module (Sec 3.4). After pre-computing concepts and contributions, WWW is ready to generate test sample explanations. Figure 2 shows the overall flow of explaining the generation of WWW in the inference time. In the test time (i.e., inference), WWW generates a class-wise explanation of the predicted class and sample explanation into two tracks. WWW leverages the predicted class and class-wise contribution matrix to generate a class-wise explanation to select the most critical neurons for the predicted class. After identifying important neurons, WWW leverages the localization module (Sec 3.3) to generate a class concept region map and class concept combination heatmap. On the other hand, for the sample explanation, WWW calculates the Taylor-approximated Shapley value of the sample for the predicted class and identifies critical neurons for the prediction of the sample. After identifying important neurons for the sample, WWW leverages the sample Shapley value to generate the sample concept region map and sample concept combination heat map." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 123, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 123, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 123, + 624 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 715 + ], + "type": "text", + "content": "Comprehensive experiments have been conducted to evaluate our method. In section 4.1, we evaluate the performance of the concept discovery module both qualitatively (Sec. 4.1.1) and quantitatively (Sec. 4.1.2). Section 4.2 is an ablation study for the concept discovery module. In Section 4.3, we analyze generated explanations of both correct and wrong predictions with examples." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 87, + 332, + 109 + ], + "blocks": [ + { + "bbox": [ + 308, + 87, + 332, + 109 + ], + "lines": [ + { + "bbox": [ + 308, + 87, + 332, + 109 + ], + "spans": [ + { + "bbox": [ + 308, + 87, + 332, + 109 + ], + "type": "image", + "image_path": "753df40db424fd15b1f841e8e88684930ffa708dcf10ae125ef6c8c63c9fcabd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 333, + 87, + 352, + 109 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 381, + 77 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 381, + 77 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 381, + 77 + ], + "type": "text", + "content": "ResNet - 50 Layer 4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 80, + 367, + 87 + ], + "lines": [ + { + "bbox": [ + 340, + 80, + 367, + 87 + ], + "spans": [ + { + "bbox": [ + 340, + 80, + 367, + 87 + ], + "type": "text", + "content": "Neuron 1152" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 87, + 352, + 109 + ], + "lines": [ + { + "bbox": [ + 333, + 87, + 352, + 109 + ], + "spans": [ + { + "bbox": [ + 333, + 87, + 352, + 109 + ], + "type": "image", + "image_path": "b1833a0838b902a2981c64443502d4cc113aa2959862ad74f7930757081e718a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 88, + 423, + 95 + ], + "lines": [ + { + "bbox": [ + 353, + 88, + 423, + 95 + ], + "spans": [ + { + "bbox": [ + 353, + 88, + 423, + 95 + ], + "type": "text", + "content": "WWW (Ours): goldfish - flame fish" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 95, + 403, + 100 + ], + "lines": [ + { + "bbox": [ + 355, + 95, + 403, + 100 + ], + "spans": [ + { + "bbox": [ + 355, + 95, + 403, + 100 + ], + "type": "text", + "content": "CLIP-Dissect: swordtail" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 101, + 402, + 106 + ], + "lines": [ + { + "bbox": [ + 355, + 101, + 402, + 106 + ], + "spans": [ + { + "bbox": [ + 355, + 101, + 402, + 106 + ], + "type": "text", + "content": "MILAN: ORANGE objects FALCON: TRICK, CALE" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 106, + 417, + 110 + ], + "lines": [ + { + "bbox": [ + 355, + 106, + 417, + 110 + ], + "spans": [ + { + "bbox": [ + 355, + 106, + 417, + 110 + ], + "type": "text", + "content": "FALCON:Insh,sale,series,fare" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 115, + 366, + 121 + ], + "lines": [ + { + "bbox": [ + 341, + 115, + 366, + 121 + ], + "spans": [ + { + "bbox": [ + 341, + 115, + 366, + 121 + ], + "type": "text", + "content": "Neuron 816" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 128, + 420, + 133 + ], + "lines": [ + { + "bbox": [ + 355, + 128, + 420, + 133 + ], + "spans": [ + { + "bbox": [ + 355, + 128, + 420, + 133 + ], + "type": "text", + "content": "CLIP-Dissect: genus Chaetodon" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 121, + 333, + 141 + ], + "blocks": [ + { + "bbox": [ + 308, + 121, + 333, + 141 + ], + "lines": [ + { + "bbox": [ + 308, + 121, + 333, + 141 + ], + "spans": [ + { + "bbox": [ + 308, + 121, + 333, + 141 + ], + "type": "image", + "image_path": "b5ce624862242000be52437a649e115e43d3d4fe5ca4ddfde5852d1f553ef214.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 121, + 432, + 127 + ], + "lines": [ + { + "bbox": [ + 334, + 121, + 432, + 127 + ], + "spans": [ + { + "bbox": [ + 334, + 121, + 432, + 127 + ], + "type": "text", + "content": "WWW (Ours): saltwater fish - swordtail" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 133, + 395, + 138 + ], + "lines": [ + { + "bbox": [ + 355, + 133, + 395, + 138 + ], + "spans": [ + { + "bbox": [ + 355, + 133, + 395, + 138 + ], + "type": "text", + "content": "MILAN: Aquatic life" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 442, + 103, + 486, + 125 + ], + "blocks": [ + { + "bbox": [ + 355, + 139, + 437, + 143 + ], + "lines": [ + { + "bbox": [ + 355, + 139, + 437, + 143 + ], + "spans": [ + { + "bbox": [ + 355, + 139, + 437, + 143 + ], + "type": "text", + "content": "FALCON: cover, black, white, home, blue" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 439, + 70, + 501, + 78 + ], + "lines": [ + { + "bbox": [ + 439, + 70, + 501, + 78 + ], + "spans": [ + { + "bbox": [ + 439, + 70, + 501, + 78 + ], + "type": "text", + "content": "ResNet - 50 Final layer" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 455, + 92, + 474, + 98 + ], + "lines": [ + { + "bbox": [ + 455, + 92, + 474, + 98 + ], + "spans": [ + { + "bbox": [ + 455, + 92, + 474, + 98 + ], + "type": "text", + "content": "Neuron 1" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 99, + 486, + 103 + ], + "lines": [ + { + "bbox": [ + 442, + 99, + 486, + 103 + ], + "spans": [ + { + "bbox": [ + 442, + 99, + 486, + 103 + ], + "type": "text", + "content": "Ground Truth: goldfish" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 103, + 486, + 125 + ], + "lines": [ + { + "bbox": [ + 442, + 103, + 486, + 125 + ], + "spans": [ + { + "bbox": [ + 442, + 103, + 486, + 125 + ], + "type": "image", + "image_path": "a722818b2a6fcdedf84eccc43634f97f7b2039eb7cabe1612166beb09f42a729.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 103, + 501, + 109 + ], + "lines": [ + { + "bbox": [ + 487, + 103, + 501, + 109 + ], + "spans": [ + { + "bbox": [ + 487, + 103, + 501, + 109 + ], + "type": "text", + "content": "www" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 109, + 534, + 114 + ], + "lines": [ + { + "bbox": [ + 487, + 109, + 534, + 114 + ], + "spans": [ + { + "bbox": [ + 487, + 109, + 534, + 114 + ], + "type": "text", + "content": "CLIP-Dissect: swordtail" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 114, + 517, + 119 + ], + "lines": [ + { + "bbox": [ + 487, + 114, + 517, + 119 + ], + "spans": [ + { + "bbox": [ + 487, + 114, + 517, + 119 + ], + "type": "text", + "content": "MILAN: Nature" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 120, + 545, + 130 + ], + "lines": [ + { + "bbox": [ + 487, + 120, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 487, + 120, + 545, + 130 + ], + "type": "text", + "content": "FALCON: trish, size, storage assorted, gap" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 308, + 155, + 332, + 177 + ], + "blocks": [ + { + "bbox": [ + 340, + 147, + 367, + 153 + ], + "lines": [ + { + "bbox": [ + 340, + 147, + 367, + 153 + ], + "spans": [ + { + "bbox": [ + 340, + 147, + 367, + 153 + ], + "type": "text", + "content": "Neuron 2018" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 155, + 332, + 177 + ], + "lines": [ + { + "bbox": [ + 308, + 155, + 332, + 177 + ], + "spans": [ + { + "bbox": [ + 308, + 155, + 332, + 177 + ], + "type": "image", + "image_path": "f58cb48ab48c3973fe85d6f598d284e057f1b1d629b63eefc161804b9c3fe750.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 156, + 436, + 162 + ], + "lines": [ + { + "bbox": [ + 333, + 156, + 436, + 162 + ], + "spans": [ + { + "bbox": [ + 333, + 156, + 436, + 162 + ], + "type": "text", + "content": "WWW (Ours): Dutch oven - stewing pan" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 163, + 419, + 167 + ], + "lines": [ + { + "bbox": [ + 355, + 163, + 419, + 167 + ], + "spans": [ + { + "bbox": [ + 355, + 163, + 419, + 167 + ], + "type": "text", + "content": "CLIP-Dissect: electric frying pan" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 182, + 367, + 188 + ], + "lines": [ + { + "bbox": [ + 340, + 182, + 367, + 188 + ], + "spans": [ + { + "bbox": [ + 340, + 182, + 367, + 188 + ], + "type": "text", + "content": "Neuron 1911" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 196, + 405, + 201 + ], + "lines": [ + { + "bbox": [ + 355, + 196, + 405, + 201 + ], + "spans": [ + { + "bbox": [ + 355, + 196, + 405, + 201 + ], + "type": "text", + "content": "CLIP-Dissect: cookstove" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 443, + 175, + 488, + 198 + ], + "blocks": [ + { + "bbox": [ + 355, + 167, + 425, + 172 + ], + "lines": [ + { + "bbox": [ + 355, + 167, + 425, + 172 + ], + "spans": [ + { + "bbox": [ + 355, + 167, + 425, + 172 + ], + "type": "text", + "content": "MILAN: Rounded edges in pictures" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 172, + 441, + 177 + ], + "lines": [ + { + "bbox": [ + 355, + 172, + 441, + 177 + ], + "spans": [ + { + "bbox": [ + 355, + 172, + 441, + 177 + ], + "type": "text", + "content": "FALCON: black, card, woman, home, white" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 163, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 457, + 163, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 457, + 163, + 482, + 169 + ], + "type": "text", + "content": "Neuron 544" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 169, + 488, + 175 + ], + "lines": [ + { + "bbox": [ + 443, + 169, + 488, + 175 + ], + "spans": [ + { + "bbox": [ + 443, + 169, + 488, + 175 + ], + "type": "text", + "content": "Ground Truth: Dutch o" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 175, + 488, + 198 + ], + "lines": [ + { + "bbox": [ + 443, + 175, + 488, + 198 + ], + "spans": [ + { + "bbox": [ + 443, + 175, + 488, + 198 + ], + "type": "image", + "image_path": "cac259dfc7409f347023906ad41d2053160759e097e650a6dab4ba1ef3d200c9.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 175, + 543, + 181 + ], + "lines": [ + { + "bbox": [ + 488, + 175, + 543, + 181 + ], + "spans": [ + { + "bbox": [ + 488, + 175, + 543, + 181 + ], + "type": "text", + "content": "WWW (Ours): Dutch oven" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 181, + 543, + 186 + ], + "lines": [ + { + "bbox": [ + 490, + 181, + 543, + 186 + ], + "spans": [ + { + "bbox": [ + 490, + 181, + 543, + 186 + ], + "type": "text", + "content": "CLIP-Dissect: stewing pan" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 186, + 537, + 190 + ], + "lines": [ + { + "bbox": [ + 490, + 186, + 537, + 190 + ], + "spans": [ + { + "bbox": [ + 490, + 186, + 537, + 190 + ], + "type": "text", + "content": "MILAN: Circular objects" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 190, + 543, + 195 + ], + "lines": [ + { + "bbox": [ + 490, + 190, + 543, + 195 + ], + "spans": [ + { + "bbox": [ + 490, + 190, + 543, + 195 + ], + "type": "text", + "content": "FACLCON:men top game" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 196, + 517, + 201 + ], + "lines": [ + { + "bbox": [ + 490, + 196, + 517, + 201 + ], + "spans": [ + { + "bbox": [ + 490, + 196, + 517, + 201 + ], + "type": "text", + "content": "award, guard" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 190, + 420, + 195 + ], + "lines": [ + { + "bbox": [ + 355, + 190, + 420, + 195 + ], + "spans": [ + { + "bbox": [ + 355, + 190, + 420, + 195 + ], + "type": "text", + "content": "WWW (Ours): cookfire - cookfire" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 201, + 428, + 206 + ], + "lines": [ + { + "bbox": [ + 355, + 201, + 428, + 206 + ], + "spans": [ + { + "bbox": [ + 355, + 201, + 428, + 206 + ], + "type": "text", + "content": "MILAN: Animals and circular objects" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 206, + 442, + 211 + ], + "lines": [ + { + "bbox": [ + 355, + 206, + 442, + 211 + ], + "spans": [ + { + "bbox": [ + 355, + 206, + 442, + 211 + ], + "type": "text", + "content": "FALCON: shirt, school, mini, woman, house" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of WWW with other baselines. We compared WWW with three competitive baselines (CLIP-Dissect [20], MILAN [12], FALCON [14]) in two final layer neurons and four penultimate layer (i.e., layer 4) neurons with each neuron's highly activating images. layer-4 neurons are top-2 important neurons of the final layer class. We have colored the descriptions green if they match the images, yellow if they match but are too generic or similar, and red if they do not match." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 305, + 331, + 539, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 539, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 539, + 344 + ], + "type": "text", + "content": "4.1. Performance Evaluation for Concept Module" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 350, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 458 + ], + "type": "text", + "content": "In this experiment, we evaluate the concept matching performance of the concept discovery module with four other baselines, Network Dissection [3], MILAN [12], CLIP-Dissect [20] and FALCON [14]. We evaluate the performance of methods on the various models (e.g., ResNet-18 [11], ResNet-50 [11], and ViT-B/16 [5]), various probing datasets (e.g.,Imagenet, Places365) and various concept sets (e.g., Wordnet nouns, and labels of Places365, Broaden, and ImageNet)" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 306, + 477, + 420, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 420, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 420, + 489 + ], + "type": "text", + "content": "4.1.1 Qualitative Results" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "Settings. We compared WWW with the three most comparable methods (CLIP-Dissect [20], MILAN [12], FALCON [14]) in the penultimate layer and final layer of the model. We do not compare with Network Dissection [3] due to the limitation that probe image data " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " and concept set " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "D_{concept}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " is fixed to Broaden. We used a ResNet-50 [11] model pre-trained in the ImageNet-1k [22] dataset. For probe image data " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": ", we used the ImageNet-1k validation set, and we extracted all nouns in Wordnet[8] (about 80k) dataset for a concept set " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "D_{concept}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Results. Figure 3 shows examples of descriptions for hidden neurons in the penultimate and final layers. Neurons in the penultimate layer are top-2 important neurons of the final layer neuron's ground truth label class. We observed that WWW not only interpreted each neuron well but also showed robust interpretation that the most important neuron of the class in the penultimate layer represents the same major concept as the final layer neuron." + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10972" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 113, + 533, + 243 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on final layer concept matching performance of ResNet-50 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for " + }, + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": ". Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 113, + 533, + 243 + ], + "lines": [ + { + "bbox": [ + 60, + 113, + 533, + 243 + ], + "spans": [ + { + "bbox": [ + 60, + 113, + 533, + 243 + ], + "type": "table", + "html": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
Network Dissection [3]BrodenBroden(1.2k)0.7229 ± 0.0030.2989 ± 0.0050.0010 ± 0.0010.001
MILAN(b) [12]ImageNet val-0.7300 ± 0.0030.2485 ± 0.0050.0005 ± 0.0000.001
FALCON [14]ImageNet valLAION-400m0.7065 ± 0.0030.1790 ± 0.0010.0002 ± 0.0000.001
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9340 ± 0.0030.8376 ± 0.0060.7286 ± 0.0090.933
ImageNet valBroden (1.2k)0.7369 ± 0.0030.3432 ± 0.0040.0328 ± 0.0030.108
ImageNet valWordnet (80k)0.8689 ± 0.0040.6846 ± 0.0080.3647 ± 0.0140.456
WWW (Ours)ImageNet valImageNet (1k)0.9325 ± 0.0030.8327 ± 0.0060.7719 ± 0.0090.955
ImageNet valBroden (1.2k)0.7758 ± 0.0040.4414 ± 0.0070.0645 ± 0.0070.091
ImageNet valWordnet (80k)0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.0120.645
", + "image_path": "c23793173ec8aef1fcf56021743112b26f950bf910c4fe157e31c58e7657083b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 294, + 523, + 388 + ], + "blocks": [ + { + "bbox": [ + 46, + 251, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 288 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on final layer concept matching performance of ViT-B/16 trained on ImageNet. We compared predicted neuron concepts with ground truth labels of ImageNet. We used the Imagenet-1k validation set for " + }, + { + "bbox": [ + 46, + 251, + 547, + 288 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 46, + 251, + 547, + 288 + ], + "type": "text", + "content": ". Bold numbers represent the best scores between the same settings. The average score and standard errors of the 1000 final layer neurons are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 294, + 523, + 388 + ], + "lines": [ + { + "bbox": [ + 69, + 294, + 523, + 388 + ], + "spans": [ + { + "bbox": [ + 69, + 294, + 523, + 388 + ], + "type": "table", + "html": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]ImageNet valImageNet (1k)0.9337 ± 0.0030.8375 ± 0.0060.7289 ± 0.0090.933
ImageNet valBroden (1.2k)0.7365 ± 0.0030.3416 ± 0.0040.0319 ± 0.0030.106
ImageNet valWordnet (80k)0.8700 ± 0.0040.6886 ± 0.0080.3679 ± 0.0140.460
WWW (Ours)ImageNet valImageNet (1k)0.9331 ± 0.0030.8347 ± 0.0060.7718 ± 0.0090.955
ImageNet valBroden (1.2k)0.7754 ± 0.0040.4407 ± 0.0070.0651 ± 0.0070.091
ImageNet valWordnet (80k)0.8857 ± 0.0030.6970 ± 0.0080.4166 ± 0.0120.634
", + "image_path": "1457275cca632cf28803e1f7df6c56aa7bb8dd7b60009ad5102edfc38dab5a65.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 168, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 168, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 168, + 418 + ], + "type": "text", + "content": "4.1.2 Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 510 + ], + "type": "text", + "content": "In this section, we compare the performance of our methods with baselines. As introduced in [20], we evaluate final layer neuron concepts with the class labels with various metrics. By comparing generated explanations with the class labels, we can objectively evaluate the quality of the generated neuron labels with what each neuron is trained to represent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 593 + ], + "type": "text", + "content": "Metrics. CLIP cos and mpnet cos are measured as cosine similarities between the encoded feature of the class label and selected concepts with CLIP [21] model and mpnet [26] model, respectively. F1-score is measured to evaluate the discovered concept's balance of exactness and flexibility. Also, the hit rate is calculated as a rate of selected concepts that exactly match the class label." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Quantitative comparison on ResNet-50. Table 2 compares WWW with Network Dissection [3], MILAN(b) [12], FALCON [14], and CLIP-Dissect [20]. We evaluate the final layer neuron concepts of ResNet-50 with the class labels of ImageNet-1k. In table 2, WWW showed better performance as the concept set " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " gets larger and outperformed all other baselines when concept set " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " is set to Broden and Wordnet nouns. In the results of Wordnet nouns (" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "), comparison between WWW and CLIP-Dissect [20] are statistically significant across all metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 407, + 354, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 354, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 354, + 418 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "text", + "content": "Quantitative comparison on ViT-B/16. We compared predicted concepts of the final layer to ground truth labels of ViT-B/16 pre-trained on ImageNet. In Table 3, WWW showed better performance as the concept set " + }, + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "text", + "content": " gets larger and outperformed other baselines when concept set " + }, + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "text", + "content": " is set to Broden and Wordnet nouns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "type": "text", + "content": "Quantitative comparison on ResNet-18 pre-trained in Places365. We compared predicted labels to ground truth labels in final layer neurons of ResNet-18 pre-trained on Places365. In table 4, WWW showed better performance as the concept set " + }, + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "type": "text", + "content": " gets larger and outperformed other baselines when concept set " + }, + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "type": "inline_equation", + "content": "D_{\\text{concept}}" + }, + { + "bbox": [ + 304, + 492, + 547, + 576 + ], + "type": "text", + "content": " is set to Broden and Wordnet nouns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 586, + 398, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 398, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 398, + 600 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Evaluation on the effect of leveraging base template and ACS. Table 5 shows an ablation study over various parts used in WWW. Without using a base template, WWW showed the lowest performance overall. With the use of a base template, WWW shows slightly increased performance in CLIP cos and F1-score but not much advance overall. However, performance was significantly increased, and the highest scores in all three metrics were achieved when leveraging ACS. The findings suggest that while us-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10973" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 113, + 529, + 206 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on final layer concept matching performance of ResNet-18 trained on Places365. We compared predicted neuron concepts with ground truth labels of Places365. We used the Places365 test set for " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "D_{probe}" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": ". Bold numbers represent the best scores between the same settings. The average score and standard errors of the 365 final layer neurons are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 113, + 529, + 206 + ], + "lines": [ + { + "bbox": [ + 64, + 113, + 529, + 206 + ], + "spans": [ + { + "bbox": [ + 64, + 113, + 529, + 206 + ], + "type": "table", + "html": "
MethodDprobeDconceptCLIP cosmpnet cosF1-scoreHit Rate
CLIP-Dissect [20]Places365 testPlaces365 (0.4k)0.9562 ± 0.0040.8687 ± 0.0120.7233 ± 0.0230.723
Places365 testBroden (1.2k)0.8304 ± 0.0030.4678 ± 0.0070.1096 ± 0.0080.329
Places365 testWordnet (80k)0.8378 ± 0.0060.5107 ± 0.0140.1041 ± 0.0160.104
WWW (Ours)Places365 testPlaces365 (0.4k)0.9402 ± 0.0040.8204 ± 0.0130.6361 ± 0.0240.660
Places365 testBroden (1.2k)0.8925 ± 0.0050.6415 ± 0.0130.2242 ± 0.0210.255
Places365 testWordnet (80k)0.8492 ± 0.0040.5000 ± 0.0130.1106 ± 0.0150.142
", + "image_path": "7e25587ff550ae5010ee71add0b48c9e2060a67e3000f5fc14602f2e29da9809.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 276, + 286, + 322 + ], + "blocks": [ + { + "bbox": [ + 47, + 223, + 287, + 267 + ], + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 267 + ], + "type": "text", + "content": "Table 5. Ablation study on the use of the template and ACS. '√' in the Template represents WWW using a template (e.g., 'a photo of word') '√' in the ACS represents WWW using adaptive cosine similarity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 276, + 286, + 322 + ], + "lines": [ + { + "bbox": [ + 50, + 276, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 50, + 276, + 286, + 322 + ], + "type": "table", + "html": "
TemplateACSCLIP cosmpnet cosF1-score
0.8499 ± 0.0030.6123 ± 0.0070.3265 ± 0.009
✓0.8547 ± 0.0030.6075 ± 0.0070.3361 ± 0.009
✓✓0.8858 ± 0.0030.6945 ± 0.0080.4197 ± 0.012
", + "image_path": "7424b59fa03525a8ab5bd6efc26db2d2277dc271b5a9b461f751d903a2028945.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "text", + "content": "ing a base template in the CLIP model offers some benefits, it also has its limitations, particularly in terms of embedding concepts within a similar CLIP feature space. In contrast, ACS appears to reduce the uniformity between concepts, allowing the concept discovery module to discover more distinct and accurate concepts for each neuron representation. This highlights the effectiveness of ACS in enhancing the overall performance of the concept discovery module. Ablation study on Concept Sensitivity " + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "text", + "content": ". The ablation study depicted on the left side of Figure 4 examined the impact on the F1-score by varying levels of major concept sensitivity (denoted as " + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "text", + "content": "). We found that as concept sensitivity decreases, the F1-score initially increases " + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "inline_equation", + "content": "(\\alpha > 0.95)" + }, + { + "bbox": [ + 46, + 343, + 289, + 606 + ], + "type": "text", + "content": " and then decreases. This pattern is due to a trade-off. Higher concept sensitivity leads to more precise concept identification but at the cost of identifying fewer concepts. Conversely, lower sensitivity results in more concepts being identified that may be less similar to the target concept. The point at which the F1-score is maximized can be seen as the optimal balance in this trade-off, providing a guideline for setting the most effective level of concept sensitivity for the major concept discovery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 615, + 119, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 119, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 119, + 625 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 260, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 260, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 260, + 646 + ], + "type": "text", + "content": "4.3.1 Overall Explanation Generated by WWW" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Figure 5 illustrates an explanation example generated by WWW, where the left side of the figure provides a class explanation for a test sample by highlighting important neurons for the predicted class, and the right side depicts a sample explanation showing the important neurons identi" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 223, + 425, + 314 + ], + "blocks": [ + { + "bbox": [ + 307, + 223, + 425, + 314 + ], + "lines": [ + { + "bbox": [ + 307, + 223, + 425, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 425, + 314 + ], + "type": "image", + "image_path": "41abbd05d7ff74a9a13265c79cf6a52df4e98f8828677d3ec3efc0840011f7c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "type": "text", + "content": "Figure 4. Ablation of concept sensitivity and heatmap similarity feasibility result. Left figure illustrates the F1 score with respect to Concept Sensitivity " + }, + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "type": "text", + "content": " changes. Concept sensitivity " + }, + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 305, + 322, + 547, + 399 + ], + "type": "text", + "content": " that maximizes the F1-score is illustrated as the red line. The right figure illustrates the rejection test result of heatmap similarity and maximum softmax probability (MSP). # of hit denotes the number of correctly detected samples as a misprediction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 223, + 545, + 314 + ], + "blocks": [ + { + "bbox": [ + 425, + 223, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 425, + 223, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 425, + 223, + 545, + 314 + ], + "type": "image", + "image_path": "557ea332b44f6e5abb6c279d82309619a8e74eed22ebfb0dd5cc8e67bd31be50.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 411, + 545, + 557 + ], + "blocks": [ + { + "bbox": [ + 306, + 411, + 545, + 557 + ], + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 557 + ], + "type": "image", + "image_path": "fcd3a6304ac012c6e51a47bb15f0a9de8ece2ad445a2e3ac85310b692351500a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 565, + 547, + 643 + ], + "lines": [ + { + "bbox": [ + 304, + 565, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 547, + 643 + ], + "type": "text", + "content": "Figure 5. Example of generated explanation by WWW. From top to bottom, important neurons are displayed in the order of importance. Images in each neuron are examples of a major and minor concept, respectively. Colors in the top localization image show highly related regions for each concept. The bottom localization image is a weighted sum of important neuron activation maps displayed as a heatmap." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "fied for the test sample image prediction. This explanation target neurons from the penultimate layer of a ResNet-50 model pre-trained on ImageNet, using the ImageNet validation set " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "(D_{probe})" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " and Wordnet nouns " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "(D_{concept})" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " for the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10974" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "explanation. Notably, neuron 30 is consistently identified as the most crucial in both explanations. It represents the 'tench' concept, which matches the prediction and the test sample's ground truth. Interestingly, while neuron 1246 is the third most important for the concept, it is the second most significant for the test sample's prediction, and neuron 1480, not highlighted in the class explanation, emerges as the third important neuron in the sample explanation. Despite these neuron selection and ranking variations, the model accurately predicts the correct class. The overall heatmap of the class and sample explanations can explain this. The heatmap explanation reveals a high cosine similarity between the two heatmaps, indicating that the calculated weighted sum of NAMs is similar. Despite the difference in individual neuron importance, the highly related input region is remarkably similar in both cases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 280, + 248, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 248, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 248, + 292 + ], + "type": "text", + "content": "4.3.2 Anlaysis of Explanation of Failure Case" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 299, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 289, + 586 + ], + "type": "text", + "content": "Figure 6 provides an example of a mispredicted sample explanation generated by the WWW. In this particular case of failure, not only do the selected important neurons differ between the class and sample explanations, but the cosine similarity between their respective heatmaps is relatively low, measuring at 0.19. Interestingly, even though the ground-truth class explanations and the sample explanations highlight different important neurons, they both localize to similar regions in the heatmap, showing a relatively high similarity score of 0.47. From this observation, we explore the potential utility of heatmap similarity to predict uncertainty. On the right side of Figure 4, we present the results of a rejection test that evaluates heatmap similarity and the maximum softmax probability (MSP). When the rejection method detects the mispredicted sample correctly, we consider that as a hit. The number of hit is measured with respect to the rejected sample rate. As the rejection rate increases based on their uncertainty level, it becomes evident that heatmap similarity outperforms MSP regarding the detection of misprediction samples. This suggests that heatmap similarity can serve as a more effective measure of uncertainty compared to MSP. These findings indicate that heatmap similarity can be used as a tool for predicting uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 597, + 119, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 119, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 119, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "We proposed WWW, a unified framework that provides comprehensive explanations for the 'what', 'where', and 'why' of neural network decisions. WWW demonstrates superior performance in both quantitative and qualitative measures, offering a deeper and more detailed understanding of neural network behavior. This is achieved through a novel integration of adaptive selection for concept discovery, neuron activation maps, and Shapley values. WWW's" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 69, + 541, + 274 + ], + "blocks": [ + { + "bbox": [ + 313, + 69, + 541, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 69, + 541, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 69, + 541, + 274 + ], + "type": "image", + "image_path": "831385207359ed7e0d6f249e7eb3bc7cd0824bb87729be4ffe7505aa3f878c67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 283, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 283, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 545, + 361 + ], + "type": "text", + "content": "Figure 6. Example of failure case explanation by WWW. The explanations of the predicted label are presented on the left side. On the right side, the explanation of the ground-truth label is shown. In the upper half, we displayed concept attribution maps, which show important concepts and their respective regions. In the bottom half, we showed the overall heatmap which shows important regions for the model decision." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 382, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 501 + ], + "type": "text", + "content": "adaptability is also shown across various neural network architectures, including convolutional networks and attention-based Vision Transformers. Additionally, our approach to predicting uncertainty through heatmap similarity analysis introduces a new way to obtain the reliability of their predictions. By offering localized explanations with concept annotations for individual inputs, WWW enhances the transparency of the model's decision-making process, contributing to the broader goal of making AI more reliable and trustworthy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 514, + 409, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 409, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 409, + 527 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Research Foundation of Korea (NRF) Grant funded by the Korea Government (MSIT) under Grant 2021R1G1A1094990, by the Institute of Information and Communications Technology Planning and Evaluation (IITP) Grant funded by the Korea Government (MSIT) under Grant 2022-0-00078 (Explainable Logical Reasoning for Medical Knowledge Generation), Grant IITP-2024-RS-2023-00258649 (ITRC(Information Technology Research Center) support program), Grant 2021-0-02068 (Artificial Intelligence Innovation Hub), Grant RS-2022-00155911 (Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)), and by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10975" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Yong Hyun Ahn, Gyeong-Moon Park, and Seong Tae Kim. Line: Out-of-distribution detection by leveraging important neurons. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19852-19862, 2023. 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "text", + "content": "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 301 + ], + "type": "text", + "content": "[4] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 367 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[6] Thomas Fel, Victor Boutin, Mazda Moayeri, Rémi Cadène, Louis Bethune, Léo andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. arXiv preprint arXiv:2306.07304, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 490 + ], + "type": "text", + "content": "[7] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 513 + ], + "type": "text", + "content": "[8] Christiane Fellbaum. Wordnet and wordnets. encyclopedia of language and linguistics, 2005. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "text", + "content": "[9] Ruth Fong and Andrea Vedaldi. Net2vec: Quantifying and explaining how concepts are encoded by filters in deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8730-8738, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 613 + ], + "type": "text", + "content": "[10] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proceedings of the European conference on computer vision, pages 630-645. Springer, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[12] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2, 5, 6, 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[13] Alon Jacovi, Ana Marasović, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "text", + "content": "[14] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. International Conference on Machine Learning, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 196 + ], + "type": "text", + "content": "[15] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 262 + ], + "type": "text", + "content": "[16] Ashkan Khakzar, Soroosh Baselizadeh, Saurabh Khanduja, Christian Rupprecht, Seong Tae Kim, and Nassir Navab. Neural response interpretation through the lens of critical pathways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13528-13538, 2021. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 319 + ], + "type": "text", + "content": "[17] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 546, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 353 + ], + "type": "text", + "content": "[18] Mauritz Kop. Eu artificial intelligence act: The european approach to ai. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust ..., 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 546, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 546, + 387 + ], + "type": "text", + "content": "[19] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 546, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 432 + ], + "type": "text", + "content": "[20] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-dissect: Automatic description of neuron representations in deep vision networks. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 434, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 547, + 498 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 547, + 555 + ], + "type": "text", + "content": "[22] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 556, + 546, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 546, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 546, + 621 + ], + "type": "text", + "content": "[23] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 645 + ], + "type": "text", + "content": "[24] Lloyd S Shapley. A value for n-person games. Classics in game theory, 69, 1997. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 678 + ], + "type": "text", + "content": "[25] Michael C Sloan. Aristotle's nicomachean ethics as the original locus for the septem circumstantiae. Classical Philology, 105(3):236-251, 2010. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 712 + ], + "type": "text", + "content": "[26] K. Song et al. Mpnet: Masked and permuted pre-training for language understanding. NeurIPS, 33:16857-16867, 2020. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10976" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Andong Wang, Wei-Ning Lee, and Xiaojuan Qi. Hint: Hierarchical neuron concept explainer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10254–10264, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[28] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[29] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 1" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10977" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_content_list.json b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4803340adf613b666eee371ab800901bacf9ffe9 --- /dev/null +++ b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_content_list.json @@ -0,0 +1,1735 @@ +[ + { + "type": "text", + "text": "WaterRF: Robust Watermarks in Radiance Fields for Protection of Copyrights", + "text_level": 1, + "bbox": [ + 93, + 130, + 877, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngdong Jang $^{1}$ Dong In Lee $^{1}$ MinHyuk Jang $^{1}$ Jong Wook Kim $^{1}$ Feng Yang $^{2}$ Sangpil Kim $^{1*}$ $^{1}$ Korea University $^{2}$ Google Research", + "bbox": [ + 192, + 178, + 774, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advances in the Neural Radiance Fields (NeRF) research offer extensive applications in diverse domains, but protecting their copyrights has not yet been researched in depth. Recently, NeRF watermarking has been considered one of the pivotal solutions for safely deploying NeRF-based 3D representations. However, existing methods are designed to apply only to implicit or explicit NeRF representations. In this work, we introduce an innovative watermarking method that can be employed in both representations of NeRF. This is achieved by fine-tuning NeRF to embed binary messages in the rendering process. In detail, we propose utilizing the discrete wavelet transform in the NeRF space for watermarking. Furthermore, we adopt a deferred back-propagation technique and introduce a combination with the patch-wise loss to improve rendering quality and bit accuracy with minimum trade-offs. We evaluate our method in three different aspects: capacity, invisibility, and robustness of the embedded watermarks in the 2D-rendered images. Our method achieves state-of-the-art performance with faster training speed over the compared state-of-the-art methods. Project page: https://kuai-lab.github.io/cvpr2024waterf/", + "bbox": [ + 75, + 305, + 473, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Digital watermarking plays a pivotal role in reinforcing the copyright of digital assets, e.g., text, image, audio, video, and 3D content. Digital assets are easily misused without permission from the creators and owners of the digital assets. One way to protect the copyright is by encoding invisible watermarks in the digital assets.", + "bbox": [ + 75, + 679, + 468, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [35] have emerged into the spotlight in 3D content creation and modeling since NeRF can represent 3D objects or scenes in a compact way. With increasing interest in 3D content for the Metaverse, virtual reality, and augmented reality, NeRF has become important in digital watermarking research.", + "bbox": [ + 75, + 770, + 468, + 862 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A straightforward approach for protecting rendered 3D", + "bbox": [ + 96, + 862, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ca51836b73624f0772d63f0f8ae808cd9047232350cef2002b3168eccfce436.jpg", + "image_caption": [ + "Figure 1. The NeRF model can be fine-tuned to embed the watermark into the images of the novel view. The figure above shows the model owner Alice and her NeRF model is stolen by Bob. Alice can identify the copyright by using her watermark decoder. Our method embeds the watermark into all of the rendered images." + ], + "image_footnote": [], + "bbox": [ + 506, + 273, + 890, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images from the NeRF model is using the post-generation watermarking method that embeds the watermark into the 2D rendered image from NeRF with existing watermarking methods. However, this method only protects copyright for the rendered images but not the copyright of the NeRF model. A fundamental solution to protecting the model and rendered image is fine-tuning the NeRF model to encode the watermark into the NeRF model itself. The watermark will be encoded into the rendered image from the fine-tuned NeRF every time during the rendering process.", + "bbox": [ + 496, + 530, + 893, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NeRF has two primary representations: implicit and explicit representations. The implicit representation [1, 35, 37] uses a multi-layer perceptron (MLP) to represent a 3D scene and the explicit representation [3, 10, 58] uses a traditional 3D structure, such as voxels. Prior methods [22, 30] were limited to applying to only one type of NeRF representation. However, in order to overcome this limitation, we explore a new question: How can we embed the watermark into two distinct NeRF representations(implicit and explicit)?", + "bbox": [ + 496, + 685, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose an innovative watermarking method that can be applied to both types of NeRF representations. Our method integrates the watermarking process into the rendering process without changing the model architecture. It adjusts the pre-trained NeRF model so that all", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12087", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of the rendered images have same watermark embedded. We pre-train the decoder with the conventional deep learning method HiDDeN [59] to extract the watermark from the rendered image. Our method has several advantages. First, since it is a fine-tuning process, we can deal with the implicit and the explicit NeRF representations. Second, we do not need any extra processes for concealing watermarks in the rendered images. Therefore, less computational cost is required. Third, our method can protect the model and the rendered images simultaneously.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Invisibility and robustness are very important aspects in the digital watermarking domain. Concealing watermarks in the rendered image without degrading the original image quality and ensuring identification of the watermark is challenging. For instance, an embedded watermark in the image is better to be as invisible as possible, and also watermark should be identified after various distortion attacks. We intend to hide the watermark message in the low-level subband in the frequency domain by carefully designing the loss function that evaluates the loss value in the frequency domain. To enhance the image quality, we consider the local structure of the object by introducing patch-wise loss. We subdivided the rendered image and cached the gradient in the grid to encode the watermark locally in the image.", + "bbox": [ + 75, + 243, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments demonstrate that our method successfully encodes watermarks by fine-tuning the pretrained NeRF models and identifying the watermark message using a simplified decoder from HiDDeN [59]. Moreover, we show that our method is robust from diverse watermark attacks. We evaluate our method by measuring the capacity and robustness under diverse watermark attacks. The training time of our method is approximately six times faster than the CopyRNeRF [30]. Since our method can be applied to both implicit [35] and explicit [3] NeRF, our method can be used in more general applications than the other watermarking methods. Our method outperforms the other state-of-the-art watermarking methods within all metrics. In summary, our contributions are as follows:", + "bbox": [ + 75, + 455, + 472, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our method can be applied to implicit and explicit NeRF models, unlike other existing watermarking methods.", + "- We propose a novel watermarking method for NeRF that fine-tunes the NeRF model by minimizing the loss function which is evaluated in the frequency domain.", + "- We propose a patch-wise loss to improve rendering quality and bit accuracy and enable encoding the watermark locally in the image, reducing the color artifacts.", + "- The proposed watermarking method achieves state-of-the-art performance, and we show that our methods are robust in diverse watermark attacks." + ], + "bbox": [ + 76, + 667, + 468, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields. Due to the success of Neural Radiance Field (NeRF) [35], a highly photo-realistic", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "view synthesis of complex scenes has been achieved. Recently, NeRF has been used in various research and applications, including faster inference [26, 52, 54], 3D reconstruction [32, 39, 53], image processing [19, 31, 47], dynamic scenes [24, 42] and generative models [25, 33, 40]. Therefore, NeRF become the dominant 3D representation and is widely used. Several commercial products [36, 45] utilized NeRF 3D representations. Thus, the management of the copyright of NeRF-based 3D representations has been emerging as a crucial aspect. As all of those applications are based on implicit, explicit, or both representations, we explore a versatile watermarking method that can be trained on both representations.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Digital watermarking. Watermarking methods have been evolving over the decades [2, 43]. There are two main ways to recover watermarks [23]: multi-bit watermark and zero-bit watermark. The multi-bit watermark allows for the encoding and decoding of multi-bit messages from media. The zero-bit watermark can be used for ownership authentication. In this paper, we focus on the zero-bit watermark, which involves encoding and decoding multi-bit messages. Recently, with the great success of deep learning watermark methods like HiDDeN [59], considerable research has been conducted to embed watermarks into carriers, especially into 2D images [8, 27, 28], video [7, 29] and 3D [9, 49, 51].", + "bbox": [ + 496, + 296, + 893, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the NeRF field, the watermarks do not survive during the volume rendering. To overcome this issue, several methods [22, 30] have emerged. StegaNeRF [22] designed a steganography model on explicit NeRF that hides natural images in 3D scene representation. CopyRNeRF [30] secured the copyright of images rendered from an implicit 3D model by embedding a watermark into the rendered color representation. However, these methods have two big limitations. First, they could only be applied to one of the implicit or explicit representations, while both representations of NeRF greatly impact the advancement of NeRF. Second, these watermarks were not robust enough. Since StegaNeRF required the original image for message extraction, this method was not robust to distortion-like cropping, which made a lot of difference between the original and rendered images and CopyRNeRF did not resist cropping. To address these issues, we design a robust watermarking NeRF model that not only can be utilized on both implicit and explicit representation but is also robust to diverse attacks.", + "bbox": [ + 496, + 479, + 893, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Frequency Domain. It is known that embedding a watermark into the spatial domain is vulnerable to attacks such as cropping and compression because it directly modifies the image pixels [46]. For this reason, many studies have been conducted based on frequency domains [44, 57]. The following are the frequency transformation methods.", + "bbox": [ + 496, + 776, + 893, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Discrete Fourier Transform (DFT) expresses images in terms of phase and amplitude. When a watermark is em", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12088", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bedded into the frequency domain converted by DFT, the magnitude is invariant, making it robust against geometric attacks such as rotation and scaling [14, 20, 41].", + "bbox": [ + 89, + 90, + 467, + 136 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Discrete Cosine Transform (DCT) decomposes the energy of image data into a sum of cosine functions by representing signals in the frequency domain. It is useful for image compression because most of the energy of the image is concentrated in the top-left corner coefficient. Therefore, DCT is extensively used in watermarking methods that energy compaction. [5, 17, 38]", + "- Discrete Wavelet Transform (DWT) decomposes the image into four subbands, LL(Low-Low), LH(Low-High), HL(High-Low), and HH(High-High). The LL band has the most energy and contains low-frequency information. Additionally, the LL band can be recursively decomposed into n-levels. DWT is used in several watermarking methods [6, 50] and shows significant differences in performance depending on the level [11, 21, 48]. In this paper, we use DWT and the LL subband for fine-tuning NeRF to embed the watermark." + ], + "bbox": [ + 76, + 136, + 467, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: Representation of NeRF", + "text_level": 1, + "bbox": [ + 76, + 405, + 426, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we focus on both implicit and explicit representations of a NeRF model. The implicit NeRF [35] represents a scene using a multi-layer perceptrs (MLP) $\\Phi$ whose input is 3D location $\\pmb{x} = (x,y,z)\\in \\mathbb{R}^3$ and 2D viewing direction $d\\in \\mathbb{R}^2$ and output is color $c\\in \\mathbb{R}^3$ , volume density $\\sigma \\in \\mathbb{R}^{+}$ . These MLP-based radiance fields can be written as:", + "bbox": [ + 76, + 430, + 467, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma , c = \\Phi (x, d) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 531, + 467, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The explicit NeRF, such as TensoRF [3], models a scene as a 4D tensor, which represents a 3D voxel grid. These grid-based radiance fields can be written as:", + "bbox": [ + 76, + 544, + 467, + 588 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma , \\boldsymbol {c} = \\mathcal {T} _ {\\sigma} (\\boldsymbol {x}), \\mathcal {S} \\left(\\mathcal {T} _ {c} (\\boldsymbol {x}), \\boldsymbol {d}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 590, + 467, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{T}_{\\sigma} \\in \\mathbb{R}^{I \\times J \\times K}$ is a geometry tensor and $\\mathcal{T}_c^{I \\times J \\times K \\times C}$ is an appearance tensor for channel $C$ and the voxel grid resolution $I, J, K$ . $S$ is a decoding function including MLP and spherical harmonics function.", + "bbox": [ + 76, + 608, + 467, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both representations are followed by a volumetric rendering equation, which is used to synthesize novel views:", + "bbox": [ + 76, + 669, + 467, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} C (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\boldsymbol {r} (t)) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) d t, \\tag {3} \\\\ T (t) = e x p \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (\\boldsymbol {r} (s)) d s\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 704, + 467, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{r}(t)$ is a camera ray with near and far bounds $t_n$ and $t_f$ . $C(r)$ is an expected color of the ray $\\pmb{r}(t)$ .", + "bbox": [ + 76, + 779, + 467, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the NeRF representation type like the above, how can we embed a watermark in the various types, not just one type? Since these types of NeRF share a common property, rendering images through the same volume rendering function, we use the rendered images to embed watermarks into the models.", + "bbox": [ + 75, + 810, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 90, + 589, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a process of fine-tuning the NeRF that does not involve alterations to the model's architecture to embed the watermark message. Our method aims to embed the watermark into the weights $\\theta$ of the NeRF model in the frequency domain of rendered images. Our method stands apart from traditional watermarking methods, which focusing on training encoders and decoders. The difference lies in the fine-tuning process, which embeds the watermark without using an encoder. There are two phases: (1) Pre-training the watermark decoder $D$ , (2) Fine-tuning the NeRF model $F_{\\theta}$ to embed the message. Our method is illustrated in Fig. 2 and described in detail below.", + "bbox": [ + 498, + 117, + 890, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Pre-training the watermark decoder", + "text_level": 1, + "bbox": [ + 500, + 314, + 813, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We select HiDDeN [59] architecture as our watermark decoder. HiDDeN comprises two convolution networks for data hiding: a watermark encoder $E$ and a watermark decoder $D$ . For robustness, it includes a noise layer $N$ . However, in this training phase, where we focus solely on the decoder's performance, we have excluded an adversarial loss responsible for improving visual quality. After training the HiDDeN model, watermark encoder $E$ was not utilized in the second phase.", + "bbox": [ + 498, + 339, + 890, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder $E$ takes in a cover image $I_{o}\\in \\mathbb{R}^{H\\times W\\times 3}$ and a binary message $M\\in \\{0,1\\} ^L$ with length $L$ , as input. Then $E$ embeds $M$ into $I_{o}$ and produces an encoded image $I_w$ . In order to make the decoder resistant to various distortions such as rotation and JPEG compression, $I_{w}$ is transformed with a noise layer $N$ . The decoder $D$ , made of several convolution layers, receives $I_{w}$ as input and extracts a message $M^{\\prime}$ .", + "bbox": [ + 498, + 476, + 890, + 598 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {\\prime} = D \\left(N \\left(I _ {w}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 599, + 890, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We utilize a sigmoid function to set the range of the extracted message $M'$ between [0, 1]. The message loss is calculated with Binary Cross Entropy (BCE) between $M_L$ and sigmoid $sg(M_L')$ .", + "bbox": [ + 498, + 632, + 890, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m e s s a g e}} = - \\sum_ {i = 1} ^ {L} M _ {i} \\cdot \\log s g \\left(M _ {i} ^ {\\prime}\\right) \\\\ + \\left(1 - M _ {i}\\right) \\cdot \\log \\left(1 - s g \\left(M _ {i} ^ {\\prime}\\right)\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 700, + 890, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The decoder is trained to detect watermarks in images that have passed through the trained encoder. However, we do not use the encoder in the second stage. We find that when the decoder receives a vanilla-rendered image, there is a bias between the extracted message bits. Thus, after training the decoder, we conduct PCA whitening to a linear decoder layer to remove the bias without reducing the extraction ability.", + "bbox": [ + 498, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12089", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ec1956eca51d7a48ffcfcdabaa3a64a0e8b423bc0f07c678c384b7030dfc946c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 88, + 839, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/29f81458ee65570aaea6e4b94bfdfcf97b576b7da73d2c994978ca3f9cb0d1d1.jpg", + "image_caption": [ + "Figure 2. WaterF overview. Phase 1: We train the encoder and the decoder to extract messages. After phase 1, we do not use the encoder. Phase 2: We fine-tune the NeRF to embed the messages into the rendered images. (a) We disable auto-differentiation and render a full-resolution image to save memory. (b) We use DWT for the rendered images and choose the LL subband as the input of the pre-trained decoder. (c) We enable auto-differentiation and render the images patch by patch. Then the NeRF is optimized using Eq 7 and Eq 8." + ], + "image_footnote": [], + "bbox": [ + 133, + 180, + 838, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Embedding and extracting watermark on DWT", + "text_level": 1, + "bbox": [ + 75, + 476, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, a fine-tuning watermarking method for NeRF [22] in the spatial domain has emerged. Although the fine-tuning method of embedding a message in the spatial domain shows incomparable invisibility and message extraction ability, it is vulnerable to attacks that distort the spatial domain, such as cropping. Directly applying spatial domain techniques from the latent diffusion model [8] does not allow for effective adjustment of NeRF's weights.", + "bbox": [ + 75, + 500, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle these problems, we propose a fine-tuning method in a frequency domain instead of a spatial domain. Various watermarking techniques for images use the frequency domain have seen development and improvement over the years. We find that DWT is the appropriate domain for encoding the message into the weights of a NeRF model.", + "bbox": [ + 75, + 621, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The NeRF model renders diverse views of the 3D model given corresponding camera parameters. We transform the pixels of the rendered images, denoted by $X = (x_{c},y_{c})\\in \\mathbb{R}^{H\\times W\\times 3}$ , into wavelet forms, with $c$ representing the channel. The DWT is defined as [12]:", + "bbox": [ + 75, + 710, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} W _ {\\varphi} (j _ {0}, m, n) = \\frac {1}{\\sqrt {M N}} \\sum_ {x _ {c} = 0} ^ {M - 1} \\sum_ {y _ {c} = 0} ^ {N - 1} f (x _ {c}, y _ {c}) \\varphi_ {j _ {0}, m, n} (x _ {c}, y _ {c}), \\\\ W _ {\\psi} ^ {i} (j, m, n) = \\frac {1}{\\sqrt {M N}} \\sum_ {x _ {c} = 0} ^ {M - 1} \\sum_ {y _ {c} = 0} ^ {N - 1} f \\left(x _ {c}, y _ {c}\\right) \\psi_ {j, m, n} ^ {i} \\left(x _ {c}, y _ {c}\\right) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 799, + 477, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi (x,y)$ is a scaling function and $\\psi (x,y)$ is a wavelet function. $W_{\\varphi}(j_0,m,n)$ is called by an LL subband, which is an approximation of the image at scale $j_0$ $W_{\\psi}^{i}$ represents LH, HL, HH subbands, where $i = \\{\\mathrm{H},\\mathrm{V},\\mathrm{D}\\}$ and each denotes horizontal, vertical and diagonal coefficients.", + "bbox": [ + 498, + 476, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous studies [13, 15, 18] have selected $LH$ , $HL$ , and $HH$ subbands for embedding watermarks because the $LL$ subband contains significant information about the image. However, we choose the LL subband as an input of our decoder $D$ and get the extract message with $M' = D(W_{\\varphi})$ . We experimentally discover that embedding the watermark in the LL subband is more robust and effective than other subbands for the HiDDeN decoder.", + "bbox": [ + 498, + 551, + 890, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The DWT is characterized by its subbands being computed across different levels; therefore, choosing an optimal level for our purpose is necessary. The 1-level separates the images to 4 subbands $(LL_{1}, LH_{1}, HL_{1}, HH_{1})$ , then 2-level separate the $LL_{1}$ subband into 4 subbands $(LL_{2}, LH_{2}, HL_{2}, HH_{2})$ . We experimentally discover that selecting a level too high decreases visual quality due to excessive adjustment of crucial visual elements. Thus, we use the 2-level DWT because it best maintains a good balance between bit accuracy and reconstruction quality.", + "bbox": [ + 498, + 672, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Fine-tuning the NeRF model", + "text_level": 1, + "bbox": [ + 500, + 832, + 754, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF has two different representations, implicit and explicit representations (see Sec. 3). Previous studies [22, 30] focused on applying their methods to just one form of NeRF", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12090", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representation. In contrast, we propose a method applicable to both NeRF representations: implicit and explicit. Our approach involves fine-tuning the NeRF to ensure that novel view images include an embedded message.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method starts with preparing each pre-trained decoder for different lengths of message bits and also preparing an initial NeRF model $F_{\\theta_0}$ . Then we establish fixed binary messages $M = (m_1, \\dots, m_L) \\in \\{0, 1\\}^L$ .", + "bbox": [ + 75, + 151, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since rendering with NeRF requires a lot of memory, we turn off the auto-differentiation at first. The rendered images $X \\in \\mathbb{R}^{H \\times W \\times 3}$ at full resolution are rendered by $F_{\\theta_0}$ . Following this, the full resolution images $X$ are transformed into seven groups of wavelet subbands: $\\{LL_2, LH_2, HL_2, HH_2, LL_1, LH_1, HL_1, HH_1\\}$ . Our pre-trained decoder $D$ takes the $LL_2$ subband as an input and extracts a message $M' = D(W_{\\varphi}(j_0, m, n))$ from it. The message loss is calculated by comparing $M$ and $M'$ : $L_m = BCE(M, sg(M'))$ . We choose the Waston-VGG [4] to calculate the full image perceptual loss between the image without watermark $X_0$ and the image with a watermark $X$ : $L_i(X_0, X)$ . Full resolution loss function is a sum of the full image perceptual loss and the message loss.", + "bbox": [ + 75, + 212, + 470, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f u l l}} = \\lambda_ {i} \\mathcal {L} _ {\\text {p e r c e p t}} + \\lambda_ {m} \\mathcal {L} _ {\\text {m e s s a g e}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 433, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We cache the gradient calculated by Eq. 7 prior to the patch-wise rendering process to update the parameters. Then, we turn on the auto-differentiation so that the NeRF parameters enable patch-wise rendering. Our patch loss $L_{patch}$ is calculated by combining mean absolute error (MAE) across rendered pixel colors, total variation (TV) regularization, and SSIM loss to maintain the balance between bit-accuracy and visual quality.", + "bbox": [ + 75, + 457, + 468, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p a t c h}} = \\lambda_ {M A E} \\mathcal {L} _ {M A E} + \\lambda_ {T V} \\mathcal {L} _ {T V} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 587, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we optimize from the initial NeRF model $F_{\\theta_0}$ to watermarked $F_{\\theta}$ using Eq. 7 and Eq. 8.", + "bbox": [ + 75, + 612, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Deferred back-propagation with patch loss", + "text_level": 1, + "bbox": [ + 75, + 650, + 441, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In conventional NeRF models, rendering at full resolution is inefficient regarding memory usage. To solve the memory-inefficient problem, pixels are often sampled and rendered randomly to compute the L2 loss for training. However, the main limitation of random pixel-wise rendering is in its incompatibility with using CNN-based perceptual loss and our watermark decoder for extracting message bits, both of which require a full-resolution image as input.", + "bbox": [ + 75, + 674, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method implements a full-resolution rendering derived from our pre-trained decoder during the fine-tuning process. Each iteration presents a considerable challenge due to the extensive memory consumption. In order to relieve the memory constraints, we adopt the deferred back-propagation [55], specifically designed for memory-efficient, patch-wise optimization within NeRF.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method proposes a novel patch loss to be applied to the deferred back-propagation as detailed in Eq. 8. When the model renders the full-resolution image with autodifferentiation disabled, the image and message loss are computed by the rendered image, and the derived gradients are cached (see Sec. 4.3 for more details). Then, the model renders the full-resolution image with image patches that are re-rendered with auto-differentiation enabled and back-propagated using the corresponding gradients from the cached set.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our innovation is not solely relying on cached gradients for back-propagation during re-rendering. Instead, we calculate the loss between the re-rendered patch and the ground truth (GT) patch and carry out an additional backpropagation process. The patch loss consists of a sum of L1 loss (MAE), a Total Variation loss, and an SSIM loss, as defined in Eq. 8. By utilizing back-propagation, once for the combined loss from the full-resolution image and the message loss and once for the patch loss, we are able to enhance both image quality and message bit accuracy at once.", + "bbox": [ + 496, + 242, + 890, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 409, + 632, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 433, + 594, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method using datasets commonly used by NeRF: the Blender dataset [35], which consists of synthetic bounded scenes, and the LLFF dataset [34], which consists of real unbounded, forward-facing scenes. We follow the conventional literature in NeRF papers, which evaluate and compare using eight scenes each from the Blender dataset (including chair, drums, ficus, hotdog, lego, materials, mic) and the LLFF dataset (including fern, flower, fortress, horns, leaves, orchids, room, trex).", + "bbox": [ + 496, + 457, + 890, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 604, + 718, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vanilla NeRF [35] and TensoRF [3] employ similar training methods, with some specifics as detailed below. The training process is carried out on an image-by-image basis, with a batch size set to one. Ray tracing is implemented along with the camera pose, and the Adam optimizer [16] is used for the learning process. The training is completed with epochs ranging from 5 to 10. Since our pretrained decoder requires full-resolution images, NeRF renders full-resolution images during the fine-tuning process. However, NeRF requires a great amount of memory usage. We adopt the deferred back-propagation in ARF [55] to solve the memory issue and combine it with the patchwise loss to improve rendering quality and bit accuracy.", + "bbox": [ + 496, + 628, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The initial $\\lambda_{i}$ , $\\lambda_{m}$ was set to 0.1 and 0.9 for the LLFF dataset and 0.05 and 0.95 for the Blender dataset, respectively. Additionally, within the patch loss equation (Eq. 8), the parameters $\\lambda_{MAE}$ , $\\lambda_{TV}$ , and $\\lambda_{SSIM}$ were set to 0.1, 0.06, and 0.02.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12091", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c66902d6b4cca992d8ded814a2ccbe937ef378d68e7ea1ebc1b372de13a7192b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Bit Accuracy(%) ↑
4 bits8 bits16 bits32 bits48 bits
HiDDeN [59]+NeRF [35]50.3150.2550.1950.1150.04
CopyRNeRF [30]100.00100.0091.1678.0860.06
Ours (w/ NeRF [35])100.00100.0094.2486.8170.43
Ours (w/ TensoRF [3])100.00100.0095.6788.5885.82
", + "bbox": [ + 81, + 89, + 459, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 260, + 196, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method with three key essential aspects in watermarking techniques: 1) capacity, measuring the length of the messages that can be embedded into the image; 2) invisibility, which involves evaluating the difficulty in detecting embedded watermarks in images by peoples; 3) robustness, evaluating the robustness of our watermarking method under various image distortions such as Gaussian Noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. We employ Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [56] for the evaluation metrics to assess invisibility while utilizing bit accuracy to evaluate capacity and robustness. We conduct experiments on the invisibility and robustness of the message length of $M_L \\in \\{4,8,16,32,48\\}$ .", + "bbox": [ + 75, + 284, + 470, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline To the best of our knowledge, CopyRNeRF [30] stands out alone in its ability to embed bits with the implicit NeRF model, which can be converted into various modality data such as messages or images. The CopyRNeRF performs state-of-the-art on bit-accuracy and visual quality. As a result, we conduct our evaluations utilizing the CopyRNeRF. We select the following relevant comparative models:: HiDDeN [59] $^+$ NeRF [35], which employ the classical 2D watermarking method HiDDeN [59] on training images prior to training the NeRF model.", + "bbox": [ + 75, + 534, + 468, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the applicability of our method, we apply our method on two distinct NeRF models: Vanilla NeRF [35], symbolizing implicit representation NeRF, and TensoRF [3], representing explicit representation NeRF based on voxel grids. We select the Vanilla NeRF [35] since it is the first NeRF model to be introduced and can be considered the foundational and ancestral model for all subsequent NeRFs. TensoRF [3] is chosen because it is an influential NeRF model with significantly improved performance and training speed.", + "bbox": [ + 75, + 686, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Experimental results", + "text_level": 1, + "bbox": [ + 76, + 847, + 272, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Capacity As shown in Tab. 1, there is a clear trade-off between bit accuracy and message bit length. We conduct bit", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3b58ed33950e7d809bee54d9b15ab2699d8fed367606b06397e6e1dd7699c21e.jpg", + "table_caption": [ + "Table 1. Bit accuracies with different message lengths are compared with the state-of-the-art method. (HiDDeN [59] + NeRF [35] is pre-embedded before training). The results are the average of LLFF and Blender datasets." + ], + "table_footnote": [], + "table_body": "
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
HiDDeN [59]+NeRF [35]50.1926.530.9170.035
CopyRNeRF [30]91.1626.290.9100.038
Ours (w/ NeRF [35])94.2428.810.9540.025
Ours (w/ TensoRF [3])95.6732.790.9480.033
", + "bbox": [ + 498, + 89, + 901, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Bit accuracies and reconstruction qualities comparison with baselines. $\\uparrow (\\downarrow)$ means higher (lower) is better. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold.", + "bbox": [ + 498, + 186, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "accuracy tests on messages of lengths 4, 8, 16, 32, and 48 bits. Models implemented with our method show a gradual decrease in bit accuracy as the length of the message increases. However, when compared to CopyRNeRF [30], the decline in bit accuracy is less dramatic and less steep. Notably, CopyRNeRF [30] shows a modest $60.6\\%$ bit accuracy for 48 bits, whereas Vanilla NeRF [35] and TensoRF [3] using our method demonstrates a respectable $85.82\\%$ and $70.43\\%$ accuracy, respectively. Additionally, to demonstrate the stability of our method across a range of messages, we fine-tune 500 different messages, each corresponding to unique, randomly generated bits. We rigorously evaluate the bit accuracy for all images rendered by Vanilla NeRF [35] and TensoRF [3]. As illustrated in Fig. 3, the plot shows a minor descending trend that corresponds with the increasing number of unique messages, and it shows that our method is robust with using many different messages, which is a key value to identify unique NeRF model.", + "bbox": [ + 496, + 260, + 892, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Invisibility Our method demonstrates a trade-off between bit accuracy and the quality of reconstructed data, as detailed in Tab. 2. This trade-off is analyzed using several metrics, including bit accuracy and visual quality metrics. With TensoRF [3], our method achieves the best bit accuracy and PSNR results. Conversely, when applied to Vanilla", + "bbox": [ + 496, + 542, + 890, + 633 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c593b16d8e0edacbafacebd5ef85a62007188989298e4a486bbd0ce2636a1037.jpg", + "image_caption": [ + "Figure 3. Identification results. We average the bit accuracy for each 500 messages. Our method not only works efficiently with just one key but also works well with arbitrarily generated messages. We show the results on $M_L = 16$ bits." + ], + "image_footnote": [], + "bbox": [ + 532, + 660, + 856, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12092", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cbacb24a57e965f27c215c9b826172568ba40470738e04e4116755f31ac328eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No DistortionGaussian Noise (v = 0.1)Rotation (±π/6)Scaling (25%)Bit Accuracy(%) ↑
Gaussian Blur (deviation = 0.1)Crop (40%)Brightness (2.0)JPEG Compression (10% quality)Combined (Crop, Brightness, JPEG)
CopyRNeRF [30]91.1690.0488.1389.3390.06----
Ours (w/ NeRF [35])94.2494.0685.0291.3594.1283.4884.1486.8873.64
Ours (w/ TensoRF [3])95.6795.3693.1393.2995.2595.4090.9186.9984.12
", + "bbox": [ + 78, + 89, + 890, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e61dfb0f4c098e3e483b9b712c56131d0864efa55ba8f044b8af6795639d45ed.jpg", + "image_caption": [ + "CopyRNeRF" + ], + "image_footnote": [], + "bbox": [ + 112, + 229, + 200, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01c442fd356a006cb7b3e0aed43f4fe37a88c72f280cb53e6c04084b3e0128e4.jpg", + "image_caption": [ + "PSNR=32.69", + "Bit Acc=100%" + ], + "image_footnote": [], + "bbox": [ + 109, + 287, + 200, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/468834f85bffd511f7697ab38a26ecf5a7be37538af641cea89a7812cd2aa00b.jpg", + "image_caption": [ + "Ours + NeRF" + ], + "image_footnote": [], + "bbox": [ + 225, + 229, + 313, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/765ea5dec9083185de5dfff7d62210510bd684a0e70a4a22b4e16a2d121eeedc.jpg", + "image_caption": [ + "PSNR=32.78", + "Bit Acc=100%", + "Figure 4. Qualitative comparisons We show the differences $(\\times 10)$ between the rendered images and the ground truth. Our method achieves higher PSNR and bit accuracy than CopyRNeRF." + ], + "image_footnote": [], + "bbox": [ + 223, + 286, + 315, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1cb826a7901b0604f37127643e0dd5e479f92d7950e17e0b6d0224257f514e4d.jpg", + "image_caption": [ + "Ours + TensoRF" + ], + "image_footnote": [], + "bbox": [ + 334, + 229, + 421, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4222fcd486cd2647ad16e37b4fa00b4c175e2c130ac06e6e1ebb6a24fa011108.jpg", + "image_caption": [ + "PSNR=33.44", + "Bit Acc=100%" + ], + "image_footnote": [], + "bbox": [ + 334, + 286, + 426, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NeRF [35], it yields the highest SSIM and the lowest LPIPS (noting that higher SSIM and lower LPIPS are desirable). Although the SSIM and LPIPS scores for TensoRF [3] using our method are not the highest, they are comparable to those achieved with Vanilla NeRF [35]. In summary, models trained with our method attain superior bit accuracy, PSNR, SSIM, and the lowest LPIPS, indicating a well-balanced trade-off between bit accuracy and reconstruction quality.", + "bbox": [ + 75, + 441, + 468, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, to verify the invisibility of our Result, We compute the differences between ground truth and our rendered results. As illustrated in Fig. 4, the results of applying our method to both Vanilla NeRF [35] and TensoRF [3] achieve a better balance between reconstruction quality and bit accuracy compared to CopyRNeRF [30]. This indicates that our method can embed watermarks with minimal impact on the reconstruction quality.", + "bbox": [ + 75, + 565, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness We evaluate the robustness of the watermarked image to different attacks by applying Gaussian noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. As shown in Tab. 3, We experiment on both Vanilla NeRF [35] and TensorRF [3] when applied with our method, resulting in robustness against various attacks compared to CopyRNeRF [30].", + "bbox": [ + 75, + 695, + 468, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 815, + 225, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Patch Loss and Frequency Domain This section delves into the effectiveness of integrating patch loss and utilizing the frequency domain. Fig. 5 presents a comparative analysis of our method, applied to Vanilla NeRF [35], both with", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d6aadba5836318a6d847adfcb7e83558daf0db9a9dcb025a5f1b5468e77b4fa.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 506, + 244, + 594, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5160e12230c71f7ecad8f4b8d4a77b9e1cc72187ee0c5bb95e141b7da0aeb2bb.jpg", + "image_caption": [ + "Full method", + "Bit Acc=94.78%" + ], + "image_footnote": [], + "bbox": [ + 604, + 247, + 689, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dce74852b610b2e6636b96f71ef60c2bb01730e10e7ba33b82b0cf0e016ef264.jpg", + "image_caption": [ + "Without DWT With Patch", + "Bit Acc=82.47%" + ], + "image_footnote": [], + "bbox": [ + 699, + 244, + 787, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7de8cbc396a242e86a49210cdc1eabafd79296cd0a7eacd762d01076e18cddff.jpg", + "image_caption": [ + "With DWT without patch", + "Bit Acc=90.78%" + ], + "image_footnote": [], + "bbox": [ + 795, + 244, + 883, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7362d29fe760912c1b7b3b34a502e0933bbbdc76200e0098a3db621cdb5c9c83.jpg", + "image_caption": [ + "Figure 5. Reconstruction quality comparisons We evaluate our full method, our method without patch loss and our method without frequency domain about 16 bits with NeRF [35]." + ], + "image_footnote": [], + "bbox": [ + 506, + 306, + 599, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc70d8cfb9cb8f6189a6d33a1558abe19cce24f44a94e5e86226f74c62db673b.jpg", + "image_caption": [ + "Bit Acc=100.0%" + ], + "image_footnote": [], + "bbox": [ + 604, + 306, + 694, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/576c25827d98305969518040202cc79c8c932af5f7ed45247d7df19718304628.jpg", + "image_caption": [ + "Bit Acc=43.54%" + ], + "image_footnote": [], + "bbox": [ + 699, + 306, + 790, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/518b753a6a76183bbea86f73a8bd333faea75d132328211f34ca7e5edd16f83e.jpg", + "image_caption": [ + "Bit Acc=100.0%" + ], + "image_footnote": [], + "bbox": [ + 795, + 306, + 887, + 357 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/28f8c3a4ce88cda36c2243fb427023bd98cd2bf1c4c3e1f52cece2c2fff81757.jpg", + "table_caption": [ + "Table 3. Robustness under diverse attacks compared with the state-of-the-art method. The bit accuracy results are the average of LLFF and Blender datasets. We show the results on $M_{L} = 16$ bits. The best performances are highlighted in **bold**." + ], + "table_footnote": [], + "table_body": "
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
DFT71.0619.040.8220.1719
DCT43.7535.330.9670.0205
No frequency72.6832.800.9440.0392
DWT-Level 191.1633.040.9470.0332
DWT-Level 295.6732.790.9480.0336
DWT-Level 390.9632.450.9500.0331
DWT-Level 485.0231.820.9520.0329
", + "bbox": [ + 500, + 441, + 888, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Bit accuracies and reconstruction qualities compared with spatial and frequency domains (DFT, DCT) and DWT levels 1, 2, 3, and 4. We focus on low frequency as our assumption. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold.", + "bbox": [ + 498, + 590, + 890, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and without the incorporation of patch loss and use in the frequency domain. As shown in Fig. 5, the absence of DWT significantly decreases bit accuracy. Without the patch-wise loss, there is a decrease in reconstruction quality.", + "bbox": [ + 498, + 678, + 890, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DWT Level During our experiments with DWT, we evaluate performance across commonly used levels: 1, 2, 3, and 4. As shown in Tab. 4, it is indicated that the results from levels 1, 3, and 4 do not perform as well as the outcomes achieved at level 2, both in terms of bit accuracy and reconstruction quality balance. Additionally, we conduct applying DFT, DCT, and no frequency instead of DWT. As a result, our findings show that the DWT at level 2 is the most effective. Thus, we use the DWT at level 2 based on the results from the ablation experiments.", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce5bb067fd97030d896e26362147ec39387415e31a84bc170040ccc846a2d90b.jpg", + "image_caption": [ + "Figure 6. Bit accuracy comparison between DWT Subbands We evaluate the bit accuracy on level 1,2,3 and all subbands. We show the results on $M_L = 16$ bits." + ], + "image_footnote": [], + "bbox": [ + 101, + 80, + 423, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DWT Subband The choice of DWT subband is as critical as the level of DWT itself. Typically, subbands such as LH, HL, and HH, which are composed of high-frequency components less perceptible to the people, are employed for watermark embedding in conventional watermarking techniques. Conversely, in our experimental approach, we opt for the LL subband representing the image's coarse approximation and containing essential low-frequency details. This decision is based on the watermark decoder's pre-training process to extract watermarks specifically from the image's spatial domain, which is visually similar to the LL subband and its demonstrated robustness to JPEG compression. In Fig. 6, our empirical investigations confirm that the LL subband yields the highest bit accuracy.", + "bbox": [ + 80, + 347, + 467, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of Training time. Even though our proposed method necessitates the use of a pre-trained model's weights as a starting point for fine-tuning NeRF, the time requirement for training is significantly less compared to that of CopyRNeRF [30]. As depicted in Fig. 7, it is evident that the model utilizing our techniques substantially outperforms CopyRNeRF [30] regarding training efficiency. Our method, when applied to an explicit NeRF architecture, such as TensoRF [3], achieves a remarkable speed increase, performing up to six times faster than CopyRNeRF [30]. Similarly, implementing our techniques within the Implicit NeRF framework, like Vanilla NeRF [35], yields a significant speed improvement, with our model operating six times faster than CopyRNeRF [30] but it also surpasses the bit accuracy of the compared model.", + "bbox": [ + 80, + 569, + 467, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 80, + 813, + 192, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a neural 3D watermarking method for the NeRF model. Our method trains a 2D watermark decoder and NeRF model separately. Therefore, our pipeline only requires training the decoder once and re-use it on differ", + "bbox": [ + 80, + 839, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c655b89c58c463fd005f137661a1bc1732dc864c4a02f96728f5fa712b1859d9.jpg", + "image_caption": [ + "Figure 7. Analysis of train time for reconstruction from blender dataset at $160 \\times 160$ resolution. Note the introduction of a scale break within the graph to accommodate the significant disparities in training durations, with time represented in minutes on the left axis and in days on the right axis." + ], + "image_footnote": [], + "bbox": [ + 503, + 77, + 888, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ent NeRF watermark models. We adopt conventional watermark techniques in image watermarking, which transfer the image from the spatial domain to the frequency domain to encode the watermark into the image efficiently. We find that DWT and patch-wise loss enhance the overall reconstruction quality while performing high message bit accuracy. Through extensive experiments, we demonstrate that our method outperforms the state-of-the-art method and is six times faster than the previous work.", + "bbox": [ + 503, + 345, + 888, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our method shows outstanding performance in the digital watermark of NeRF, but training a decoder that extracts watermark from rendered images is time-consuming, taking approximately 12 hours on a single RTX 3090. Our method only encodes a unique message per model, requiring fine-tuning for message insertion every time. We observe that the bit accuracy drops when we apply numerous messages (see Fig. 3). Although previous studies, including us, report robustness against various distortions of images, there is no consideration about cases where the model is attacked. Future work may explore watermarks resilient to cases of model attacks.", + "bbox": [ + 503, + 492, + 888, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research was supported by the Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture(Project : 4D Content Generation and Copyright Protection with Artificial Intelligence, R2022020068, $80\\%$ ), the Korea Research Institute for Defense Technology Planning and Advancement (KRIT) Grant funded by the Korean Government(Grant KRIT-CT-23-021, $10\\%$ ), the National Research Foundation of Korea grant (NRF2022R1F1A1074334, $9\\%$ ), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), $1\\%$ ).", + "bbox": [ + 503, + 688, + 888, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12094", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 1", + "[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 2", + "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 1, 2, 3, 5, 6, 7, 8", + "[4] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 5", + "[5] Chinmayee Das, Swetalina Panigrahi, Vijay K Sharma, and KK Mahapatra. A novel blind robust image watermarking in dct domain using inter-block coefficient correlation. AEU-International Journal of Electronics and Communications, 68(3):244–253, 2014. 3", + "[6] Xin Deng, Chao Gao, and Mai Xu. Pirnet: Privacy-preserving image restoration network via wavelet lifting. In ICCV, pages 22368-22377, 2023. 3", + "[7] Elhameh Farri and Peyman Ayubi. A robust digital video watermarking based on ct-svd domain and chaotic dna sequences for copyright protection. Journal of Ambient Intelligence and Humanized Computing, 14(10):13113-13137, 2023. 2", + "[8] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. ICCV, 2023. 2, 4", + "[9] Felipe ABS Ferreira and Juliano B Lima. A robust 3d point cloud watermarking method based on the graph fourier transform. Multimedia Tools and Applications, 79(3):1921-1950, 2020. 2", + "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 1", + "[11] Youcai Gao, Jinwei Wang, and Yun-Qing Shi. Dynamic multi-watermarking and detecting in dwt domain. Journal of Real-Time Image Processing, 16:565-576, 2019. 3", + "[12] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4", + "[13] A Hernández-Joaquín, G Melendez-Melendez, and R Cumplido. A secure dwt-based dual watermarking scheme for image authentication and copyright protection. Multimedia Tools and Applications, 82(27):42739-42761, 2023. 4", + "[14] Ningombam Jimson and K Hemachandran. Dft based coefficient exchange digital image watermarking. In 2018 Second International Conference on Intelligent Computing and Control Systems (ICICCS), pages 567-571. IEEE, 2018. 3", + "[15] Priyank Khare and Vinay Kumar Srivastava. A reliable and secure image watermarking algorithm using homomorphic transform in dwt domain. Multidimensional Systems and Signal Processing, 32:131-160, 2021. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 5", + "[17] Hung-Jui Ko, Cheng-Ta Huang, Gwoboa Horng, and WANG Shiuh-Jeng. Robust and blind image watermarking in dct domain using inter-block coefficient correlation. Information Sciences, 517:128-147, 2020. 3", + "[18] Sanjay Kumar and Binod Kumar Singh. Dwt based color image watermarking using maximum entropy. Multimedia Tools and Applications, 80:15487-15510, 2021. 4", + "[19] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2", + "[20] Guezouli Larbi, Hachani Roumaissa, and Herri Sara. Embedding watermark in the magnitude matrix of the dft of image. In Proceedings of the 2018 International Conference on Computing and Pattern Recognition, pages 106-110, 2018. 3", + "[21] Yong-Seok Lee, Seo Young-Ho, and Kim Dong-Wook. Blind image watermarking based on adaptive data spreading in n-level dwt subbands. Security and communication Networks, 2019, 2019. 3", + "[22] Chenxin Li, Brandon Y Feng, Zhiwen Fan, Panwang Pan, and Zhangyang Wang. Steganerf: Embedding invisible information within neural radiance fields. In ICCV, pages 441-453, 2023. 1, 2, 4", + "[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2", + "[24] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, pages 4273-4284, 2023. 2", + "[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 2", + "[26] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2", + "[27] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR, pages 13548-13557, 2020. 2", + "[28] Xiyang Luo, Michael Goebel, Elnaz Barshan, and Feng Yang. Leca: A learned approach for efficient cover-agnostic watermarking. In Media Watermarking, Security, and Forensics, 2022. 2", + "[29] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing, 2023. 2", + "[30] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12095", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "radiance fields. In ICCV, pages 22401-22411, 2023. 1, 2, 4, 6, 7, 8", + "[31] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 2", + "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2", + "[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 2", + "[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 5", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 5, 6, 7, 8", + "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[37] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. ACM Trans. Graph., 41(4), 2022. 1", + "[38] Shabir A Parah, Javaid A Sheikh, Nazir A Loan, and Ghulam M Bhat. Robust and blind watermarking technique in dct domain using inter-block coefficient differencing. Digital Signal Processing, 53:11-24, 2016. 3", + "[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2", + "[40] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[41] SN Prajwalasimha, S Chethan Suputhra, and C Mohan. Performance analysis of combined discrete fourier transformation (dft) and successive division based image watermarking scheme. International Journal of Recent Technology and Engineering, 8(3):34-39, 2019. 3", + "[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318-10327, 2021. 2", + "[43] Deepti Shukla, Nirupama Tiwari, and Deepika Dubey. Survey on digital watermarking techniques. International Journal of Signal Processing, Image Processing and Pattern Recognition, 9(1):239-244, 2016. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] S Suthaharan and S Sathananthan. Transform domain technique: robust watermarking for digital images. In Proceedings of the IEEE SoutheastCon 2000. Preparing for The New Millennium' (Cat. No. 00CH37105), pages 407-412. IEEE, 2000. 2", + "[45] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[46] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2", + "[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2", + "[48] Jinwei Wang, Shiguo Lian, and Yun-Qing Shi. Hybrid multiplicative multi-watermarking in dwt domain. Multidimensional Systems and Signal Processing, 28:617-636, 2017. 3", + "[49] Yinghui Wang, Jing Liu, Yajie Yang, Douli Ma, and Ruijiao Liu. 3d model watermarking algorithm robust to geometric attacks. IET Image Processing, 11(10):822-832, 2017. 2", + "[50] Muyu Xu, Fangneng Zhan, Jiahui Zhang, Yingchen Yu, Xiaojin Zhang, Christian Theobalt, Ling Shao, and Shijian Lu. Wavenerf: Wavelet-based generalizable neural radiance fields. In ICCV, pages 18195-18204, 2023. 3", + "[51] Innfarn Yoo, Huiwen Chang, Xiyang Luo, Ondrej Stava, Ce Liu, Peyman Milanfar, and Feng Yang. Deep 3d-to-2d watermarking: embedding messages in 3d meshes and extracting them from 2d renderings. In CVPR, pages 10031–10040, 2022. 2", + "[52] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2", + "[53] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2", + "[54] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In CVPR, pages 12397-12406, 2023. 2", + "[55] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, pages 717-733. Springer, 2022. 5", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6", + "[57] Qiumei Zheng, Nan Liu, Baoqin Cao, Fenghua Wang, and Yanan Yang. Zero-watermarking algorithm in transform domain based on rgb channel and voting strategy. Journal of Information Processing Systems, 16(6):1391-1406, 2020. 2", + "[58] Haoyi Zhu. X-nerf: Explicit neural radiance field for multiscene 360deg insufficient rgb-d views. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5766–5775, 2023. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12096", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[59] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, pages 657–672, 2018. 2, 3, 6", + "bbox": [ + 78, + 90, + 470, + 135 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12097", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_model.json b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f30d3bae22c6c14922ed4e84e35d254ed088d5ae --- /dev/null +++ b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.131, + 0.878, + 0.155 + ], + "angle": 0, + "content": "WaterRF: Robust Watermarks in Radiance Fields for Protection of Copyrights" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.179, + 0.775, + 0.245 + ], + "angle": 0, + "content": "Youngdong Jang\\(^{1}\\) Dong In Lee\\(^{1}\\) MinHyuk Jang\\(^{1}\\) Jong Wook Kim\\(^{1}\\) Feng Yang\\(^{2}\\) Sangpil Kim\\(^{1*}\\) \n\\(^{1}\\)Korea University \\(^{2}\\)Google Research" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.474, + 0.64 + ], + "angle": 0, + "content": "The advances in the Neural Radiance Fields (NeRF) research offer extensive applications in diverse domains, but protecting their copyrights has not yet been researched in depth. Recently, NeRF watermarking has been considered one of the pivotal solutions for safely deploying NeRF-based 3D representations. However, existing methods are designed to apply only to implicit or explicit NeRF representations. In this work, we introduce an innovative watermarking method that can be employed in both representations of NeRF. This is achieved by fine-tuning NeRF to embed binary messages in the rendering process. In detail, we propose utilizing the discrete wavelet transform in the NeRF space for watermarking. Furthermore, we adopt a deferred back-propagation technique and introduce a combination with the patch-wise loss to improve rendering quality and bit accuracy with minimum trade-offs. We evaluate our method in three different aspects: capacity, invisibility, and robustness of the embedded watermarks in the 2D-rendered images. Our method achieves state-of-the-art performance with faster training speed over the compared state-of-the-art methods. Project page: https://kuai-lab.github.io/cvpr2024waterf/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.67 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.771 + ], + "angle": 0, + "content": "Digital watermarking plays a pivotal role in reinforcing the copyright of digital assets, e.g., text, image, audio, video, and 3D content. Digital assets are easily misused without permission from the creators and owners of the digital assets. One way to protect the copyright is by encoding invisible watermarks in the digital assets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.47, + 0.863 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [35] have emerged into the spotlight in 3D content creation and modeling since NeRF can represent 3D objects or scenes in a compact way. With increasing interest in 3D content for the Metaverse, virtual reality, and augmented reality, NeRF has become important in digital watermarking research." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.863, + 0.47, + 0.878 + ], + "angle": 0, + "content": "A straightforward approach for protecting rendered 3D" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.274, + 0.891, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.518 + ], + "angle": 0, + "content": "Figure 1. The NeRF model can be fine-tuned to embed the watermark into the images of the novel view. The figure above shows the model owner Alice and her NeRF model is stolen by Bob. Alice can identify the copyright by using her watermark decoder. Our method embeds the watermark into all of the rendered images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.684 + ], + "angle": 0, + "content": "images from the NeRF model is using the post-generation watermarking method that embeds the watermark into the 2D rendered image from NeRF with existing watermarking methods. However, this method only protects copyright for the rendered images but not the copyright of the NeRF model. A fundamental solution to protecting the model and rendered image is fine-tuning the NeRF model to encode the watermark into the NeRF model itself. The watermark will be encoded into the rendered image from the fine-tuned NeRF every time during the rendering process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.895, + 0.824 + ], + "angle": 0, + "content": "NeRF has two primary representations: implicit and explicit representations. The implicit representation [1, 35, 37] uses a multi-layer perceptron (MLP) to represent a 3D scene and the explicit representation [3, 10, 58] uses a traditional 3D structure, such as voxels. Prior methods [22, 30] were limited to applying to only one type of NeRF representation. However, in order to overcome this limitation, we explore a new question: How can we embed the watermark into two distinct NeRF representations(implicit and explicit)?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose an innovative watermarking method that can be applied to both types of NeRF representations. Our method integrates the watermarking process into the rendering process without changing the model architecture. It adjusts the pre-trained NeRF model so that all" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12087" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "of the rendered images have same watermark embedded. We pre-train the decoder with the conventional deep learning method HiDDeN [59] to extract the watermark from the rendered image. Our method has several advantages. First, since it is a fine-tuning process, we can deal with the implicit and the explicit NeRF representations. Second, we do not need any extra processes for concealing watermarks in the rendered images. Therefore, less computational cost is required. Third, our method can protect the model and the rendered images simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Invisibility and robustness are very important aspects in the digital watermarking domain. Concealing watermarks in the rendered image without degrading the original image quality and ensuring identification of the watermark is challenging. For instance, an embedded watermark in the image is better to be as invisible as possible, and also watermark should be identified after various distortion attacks. We intend to hide the watermark message in the low-level subband in the frequency domain by carefully designing the loss function that evaluates the loss value in the frequency domain. To enhance the image quality, we consider the local structure of the object by introducing patch-wise loss. We subdivided the rendered image and cached the gradient in the grid to encode the watermark locally in the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Our extensive experiments demonstrate that our method successfully encodes watermarks by fine-tuning the pretrained NeRF models and identifying the watermark message using a simplified decoder from HiDDeN [59]. Moreover, we show that our method is robust from diverse watermark attacks. We evaluate our method by measuring the capacity and robustness under diverse watermark attacks. The training time of our method is approximately six times faster than the CopyRNeRF [30]. Since our method can be applied to both implicit [35] and explicit [3] NeRF, our method can be used in more general applications than the other watermarking methods. Our method outperforms the other state-of-the-art watermarking methods within all metrics. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.668, + 0.47, + 0.699 + ], + "angle": 0, + "content": "- Our method can be applied to implicit and explicit NeRF models, unlike other existing watermarking methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.699, + 0.47, + 0.744 + ], + "angle": 0, + "content": "- We propose a novel watermarking method for NeRF that fine-tunes the NeRF model by minimizing the loss function which is evaluated in the frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.745, + 0.469, + 0.789 + ], + "angle": 0, + "content": "- We propose a patch-wise loss to improve rendering quality and bit accuracy and enable encoding the watermark locally in the image, reducing the color artifacts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.79, + 0.469, + 0.834 + ], + "angle": 0, + "content": "- The proposed watermarking method achieves state-of-the-art performance, and we show that our methods are robust in diverse watermark attacks." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.668, + 0.47, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Neural Radiance Fields. Due to the success of Neural Radiance Field (NeRF) [35], a highly photo-realistic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "view synthesis of complex scenes has been achieved. Recently, NeRF has been used in various research and applications, including faster inference [26, 52, 54], 3D reconstruction [32, 39, 53], image processing [19, 31, 47], dynamic scenes [24, 42] and generative models [25, 33, 40]. Therefore, NeRF become the dominant 3D representation and is widely used. Several commercial products [36, 45] utilized NeRF 3D representations. Thus, the management of the copyright of NeRF-based 3D representations has been emerging as a crucial aspect. As all of those applications are based on implicit, explicit, or both representations, we explore a versatile watermarking method that can be trained on both representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.895, + 0.479 + ], + "angle": 0, + "content": "Digital watermarking. Watermarking methods have been evolving over the decades [2, 43]. There are two main ways to recover watermarks [23]: multi-bit watermark and zero-bit watermark. The multi-bit watermark allows for the encoding and decoding of multi-bit messages from media. The zero-bit watermark can be used for ownership authentication. In this paper, we focus on the zero-bit watermark, which involves encoding and decoding multi-bit messages. Recently, with the great success of deep learning watermark methods like HiDDeN [59], considerable research has been conducted to embed watermarks into carriers, especially into 2D images [8, 27, 28], video [7, 29] and 3D [9, 49, 51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.895, + 0.769 + ], + "angle": 0, + "content": "On the NeRF field, the watermarks do not survive during the volume rendering. To overcome this issue, several methods [22, 30] have emerged. StegaNeRF [22] designed a steganography model on explicit NeRF that hides natural images in 3D scene representation. CopyRNeRF [30] secured the copyright of images rendered from an implicit 3D model by embedding a watermark into the rendered color representation. However, these methods have two big limitations. First, they could only be applied to one of the implicit or explicit representations, while both representations of NeRF greatly impact the advancement of NeRF. Second, these watermarks were not robust enough. Since StegaNeRF required the original image for message extraction, this method was not robust to distortion-like cropping, which made a lot of difference between the original and rendered images and CopyRNeRF did not resist cropping. To address these issues, we design a robust watermarking NeRF model that not only can be utilized on both implicit and explicit representation but is also robust to diverse attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.894, + 0.868 + ], + "angle": 0, + "content": "Frequency Domain. It is known that embedding a watermark into the spatial domain is vulnerable to attacks such as cropping and compression because it directly modifies the image pixels [46]. For this reason, many studies have been conducted based on frequency domains [44, 57]. The following are the frequency transformation methods." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "- Discrete Fourier Transform (DFT) expresses images in terms of phase and amplitude. When a watermark is em" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12088" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "bedded into the frequency domain converted by DFT, the magnitude is invariant, making it robust against geometric attacks such as rotation and scaling [14, 20, 41]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.137, + 0.468, + 0.242 + ], + "angle": 0, + "content": "- Discrete Cosine Transform (DCT) decomposes the energy of image data into a sum of cosine functions by representing signals in the frequency domain. It is useful for image compression because most of the energy of the image is concentrated in the top-left corner coefficient. Therefore, DCT is extensively used in watermarking methods that energy compaction. [5, 17, 38]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.243, + 0.468, + 0.392 + ], + "angle": 0, + "content": "- Discrete Wavelet Transform (DWT) decomposes the image into four subbands, LL(Low-Low), LH(Low-High), HL(High-Low), and HH(High-High). The LL band has the most energy and contains low-frequency information. Additionally, the LL band can be recursively decomposed into n-levels. DWT is used in several watermarking methods [6, 50] and shows significant differences in performance depending on the level [11, 21, 48]. In this paper, we use DWT and the LL subband for fine-tuning NeRF to embed the watermark." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.137, + 0.468, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.406, + 0.428, + 0.423 + ], + "angle": 0, + "content": "3. Preliminaries: Representation of NeRF" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.468, + 0.534 + ], + "angle": 0, + "content": "In this paper, we focus on both implicit and explicit representations of a NeRF model. The implicit NeRF [35] represents a scene using a multi-layer perceptrs (MLP) \\(\\Phi\\) whose input is 3D location \\(\\pmb{x} = (x,y,z)\\in \\mathbb{R}^3\\) and 2D viewing direction \\(d\\in \\mathbb{R}^2\\) and output is color \\(c\\in \\mathbb{R}^3\\), volume density \\(\\sigma \\in \\mathbb{R}^{+}\\). These MLP-based radiance fields can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.532, + 0.468, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\sigma , c = \\Phi (x, d) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.468, + 0.589 + ], + "angle": 0, + "content": "The explicit NeRF, such as TensoRF [3], models a scene as a 4D tensor, which represents a 3D voxel grid. These grid-based radiance fields can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.592, + 0.468, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\sigma , \\boldsymbol {c} = \\mathcal {T} _ {\\sigma} (\\boldsymbol {x}), \\mathcal {S} \\left(\\mathcal {T} _ {c} (\\boldsymbol {x}), \\boldsymbol {d}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.609, + 0.468, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}_{\\sigma} \\in \\mathbb{R}^{I \\times J \\times K}\\) is a geometry tensor and \\(\\mathcal{T}_c^{I \\times J \\times K \\times C}\\) is an appearance tensor for channel \\(C\\) and the voxel grid resolution \\(I, J, K\\). \\(S\\) is a decoding function including MLP and spherical harmonics function." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.67, + 0.468, + 0.7 + ], + "angle": 0, + "content": "Both representations are followed by a volumetric rendering equation, which is used to synthesize novel views:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.705, + 0.468, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} C (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\boldsymbol {r} (t)) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) d t, \\tag {3} \\\\ T (t) = e x p \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (\\boldsymbol {r} (s)) d s\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.468, + 0.811 + ], + "angle": 0, + "content": "where \\(\\pmb{r}(t)\\) is a camera ray with near and far bounds \\(t_n\\) and \\(t_f\\). \\(C(r)\\) is an expected color of the ray \\(\\pmb{r}(t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Considering the NeRF representation type like the above, how can we embed a watermark in the various types, not just one type? Since these types of NeRF share a common property, rendering images through the same volume rendering function, we use the rendered images to embed watermarks into the models." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.59, + 0.105 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.118, + 0.892, + 0.297 + ], + "angle": 0, + "content": "We propose a process of fine-tuning the NeRF that does not involve alterations to the model's architecture to embed the watermark message. Our method aims to embed the watermark into the weights \\(\\theta\\) of the NeRF model in the frequency domain of rendered images. Our method stands apart from traditional watermarking methods, which focusing on training encoders and decoders. The difference lies in the fine-tuning process, which embeds the watermark without using an encoder. There are two phases: (1) Pre-training the watermark decoder \\(D\\), (2) Fine-tuning the NeRF model \\(F_{\\theta}\\) to embed the message. Our method is illustrated in Fig. 2 and described in detail below." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.315, + 0.815, + 0.33 + ], + "angle": 0, + "content": "4.1. Pre-training the watermark decoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.34, + 0.892, + 0.476 + ], + "angle": 0, + "content": "We select HiDDeN [59] architecture as our watermark decoder. HiDDeN comprises two convolution networks for data hiding: a watermark encoder \\( E \\) and a watermark decoder \\( D \\). For robustness, it includes a noise layer \\( N \\). However, in this training phase, where we focus solely on the decoder's performance, we have excluded an adversarial loss responsible for improving visual quality. After training the HiDDeN model, watermark encoder \\( E \\) was not utilized in the second phase." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.599 + ], + "angle": 0, + "content": "The encoder \\(E\\) takes in a cover image \\(I_{o}\\in \\mathbb{R}^{H\\times W\\times 3}\\) and a binary message \\(M\\in \\{0,1\\} ^L\\) with length \\(L\\), as input. Then \\(E\\) embeds \\(M\\) into \\(I_{o}\\) and produces an encoded image \\(I_w\\). In order to make the decoder resistant to various distortions such as rotation and JPEG compression, \\(I_{w}\\) is transformed with a noise layer \\(N\\). The decoder \\(D\\), made of several convolution layers, receives \\(I_{w}\\) as input and extracts a message \\(M^{\\prime}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.601, + 0.891, + 0.618 + ], + "angle": 0, + "content": "\\[\nM ^ {\\prime} = D \\left(N \\left(I _ {w}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.633, + 0.891, + 0.695 + ], + "angle": 0, + "content": "We utilize a sigmoid function to set the range of the extracted message \\( M' \\) between [0, 1]. The message loss is calculated with Binary Cross Entropy (BCE) between \\( M_L \\) and sigmoid \\( sg(M_L') \\)." + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.702, + 0.891, + 0.762 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m e s s a g e}} = - \\sum_ {i = 1} ^ {L} M _ {i} \\cdot \\log s g \\left(M _ {i} ^ {\\prime}\\right) \\\\ + \\left(1 - M _ {i}\\right) \\cdot \\log \\left(1 - s g \\left(M _ {i} ^ {\\prime}\\right)\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The decoder is trained to detect watermarks in images that have passed through the trained encoder. However, we do not use the encoder in the second stage. We find that when the decoder receives a vanilla-rendered image, there is a bias between the extracted message bits. Thus, after training the decoder, we conduct PCA whitening to a linear decoder layer to remove the bias without reducing the extraction ability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12089" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.84, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.181, + 0.839, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.407, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Figure 2. WaterF overview. Phase 1: We train the encoder and the decoder to extract messages. After phase 1, we do not use the encoder. Phase 2: We fine-tune the NeRF to embed the messages into the rendered images. (a) We disable auto-differentiation and render a full-resolution image to save memory. (b) We use DWT for the rendered images and choose the LL subband as the input of the pre-trained decoder. (c) We enable auto-differentiation and render the images patch by patch. Then the NeRF is optimized using Eq 7 and Eq 8." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.477, + 0.469, + 0.492 + ], + "angle": 0, + "content": "4.2. Embedding and extracting watermark on DWT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.469, + 0.621 + ], + "angle": 0, + "content": "Recently, a fine-tuning watermarking method for NeRF [22] in the spatial domain has emerged. Although the fine-tuning method of embedding a message in the spatial domain shows incomparable invisibility and message extraction ability, it is vulnerable to attacks that distort the spatial domain, such as cropping. Directly applying spatial domain techniques from the latent diffusion model [8] does not allow for effective adjustment of NeRF's weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.469, + 0.712 + ], + "angle": 0, + "content": "To tackle these problems, we propose a fine-tuning method in a frequency domain instead of a spatial domain. Various watermarking techniques for images use the frequency domain have seen development and improvement over the years. We find that DWT is the appropriate domain for encoding the message into the weights of a NeRF model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.788 + ], + "angle": 0, + "content": "The NeRF model renders diverse views of the 3D model given corresponding camera parameters. We transform the pixels of the rendered images, denoted by \\(X = (x_{c},y_{c})\\in \\mathbb{R}^{H\\times W\\times 3}\\), into wavelet forms, with \\(c\\) representing the channel. The DWT is defined as [12]:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.8, + 0.478, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} W _ {\\varphi} (j _ {0}, m, n) = \\frac {1}{\\sqrt {M N}} \\sum_ {x _ {c} = 0} ^ {M - 1} \\sum_ {y _ {c} = 0} ^ {N - 1} f (x _ {c}, y _ {c}) \\varphi_ {j _ {0}, m, n} (x _ {c}, y _ {c}), \\\\ W _ {\\psi} ^ {i} (j, m, n) = \\frac {1}{\\sqrt {M N}} \\sum_ {x _ {c} = 0} ^ {M - 1} \\sum_ {y _ {c} = 0} ^ {N - 1} f \\left(x _ {c}, y _ {c}\\right) \\psi_ {j, m, n} ^ {i} \\left(x _ {c}, y _ {c}\\right) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.552 + ], + "angle": 0, + "content": "where \\(\\varphi (x,y)\\) is a scaling function and \\(\\psi (x,y)\\) is a wavelet function. \\(W_{\\varphi}(j_0,m,n)\\) is called by an LL subband, which is an approximation of the image at scale \\(j_0\\) \\(W_{\\psi}^{i}\\) represents LH, HL, HH subbands, where \\(i = \\{\\mathrm{H},\\mathrm{V},\\mathrm{D}\\}\\) and each denotes horizontal, vertical and diagonal coefficients." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Previous studies [13, 15, 18] have selected \\( LH \\), \\( HL \\), and \\( HH \\) subbands for embedding watermarks because the \\( LL \\) subband contains significant information about the image. However, we choose the LL subband as an input of our decoder \\( D \\) and get the extract message with \\( M' = D(W_{\\varphi}) \\). We experimentally discover that embedding the watermark in the LL subband is more robust and effective than other subbands for the HiDDeN decoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.673, + 0.892, + 0.826 + ], + "angle": 0, + "content": "The DWT is characterized by its subbands being computed across different levels; therefore, choosing an optimal level for our purpose is necessary. The 1-level separates the images to 4 subbands \\((LL_{1}, LH_{1}, HL_{1}, HH_{1})\\), then 2-level separate the \\(LL_{1}\\) subband into 4 subbands \\((LL_{2}, LH_{2}, HL_{2}, HH_{2})\\). We experimentally discover that selecting a level too high decreases visual quality due to excessive adjustment of crucial visual elements. Thus, we use the 2-level DWT because it best maintains a good balance between bit accuracy and reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.756, + 0.849 + ], + "angle": 0, + "content": "4.3. Fine-tuning the NeRF model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "NeRF has two different representations, implicit and explicit representations (see Sec. 3). Previous studies [22, 30] focused on applying their methods to just one form of NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12090" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "representation. In contrast, we propose a method applicable to both NeRF representations: implicit and explicit. Our approach involves fine-tuning the NeRF to ensure that novel view images include an embedded message." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.469, + 0.213 + ], + "angle": 0, + "content": "Our method starts with preparing each pre-trained decoder for different lengths of message bits and also preparing an initial NeRF model \\(F_{\\theta_0}\\). Then we establish fixed binary messages \\(M = (m_1, \\dots, m_L) \\in \\{0, 1\\}^L\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.424 + ], + "angle": 0, + "content": "Since rendering with NeRF requires a lot of memory, we turn off the auto-differentiation at first. The rendered images \\(X \\in \\mathbb{R}^{H \\times W \\times 3}\\) at full resolution are rendered by \\(F_{\\theta_0}\\). Following this, the full resolution images \\(X\\) are transformed into seven groups of wavelet subbands: \\(\\{LL_2, LH_2, HL_2, HH_2, LL_1, LH_1, HL_1, HH_1\\}\\). Our pre-trained decoder \\(D\\) takes the \\(LL_2\\) subband as an input and extracts a message \\(M' = D(W_{\\varphi}(j_0, m, n))\\) from it. The message loss is calculated by comparing \\(M\\) and \\(M'\\): \\(L_m = BCE(M, sg(M'))\\). We choose the Waston-VGG [4] to calculate the full image perceptual loss between the image without watermark \\(X_0\\) and the image with a watermark \\(X\\): \\(L_i(X_0, X)\\). Full resolution loss function is a sum of the full image perceptual loss and the message loss." + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.434, + 0.469, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f u l l}} = \\lambda_ {i} \\mathcal {L} _ {\\text {p e r c e p t}} + \\lambda_ {m} \\mathcal {L} _ {\\text {m e s s a g e}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.579 + ], + "angle": 0, + "content": "We cache the gradient calculated by Eq. 7 prior to the patch-wise rendering process to update the parameters. Then, we turn on the auto-differentiation so that the NeRF parameters enable patch-wise rendering. Our patch loss \\( L_{patch} \\) is calculated by combining mean absolute error (MAE) across rendered pixel colors, total variation (TV) regularization, and SSIM loss to maintain the balance between bit-accuracy and visual quality." + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.588, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p a t c h}} = \\lambda_ {M A E} \\mathcal {L} _ {M A E} + \\lambda_ {T V} \\mathcal {L} _ {T V} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Then we optimize from the initial NeRF model \\(F_{\\theta_0}\\) to watermarked \\(F_{\\theta}\\) using Eq. 7 and Eq. 8." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.651, + 0.442, + 0.668 + ], + "angle": 0, + "content": "4.4. Deferred back-propagation with patch loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.795 + ], + "angle": 0, + "content": "In conventional NeRF models, rendering at full resolution is inefficient regarding memory usage. To solve the memory-inefficient problem, pixels are often sampled and rendered randomly to compute the L2 loss for training. However, the main limitation of random pixel-wise rendering is in its incompatibility with using CNN-based perceptual loss and our watermark decoder for extracting message bits, both of which require a full-resolution image as input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our method implements a full-resolution rendering derived from our pre-trained decoder during the fine-tuning process. Each iteration presents a considerable challenge due to the extensive memory consumption. In order to relieve the memory constraints, we adopt the deferred back-propagation [55], specifically designed for memory-efficient, patch-wise optimization within NeRF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Our method proposes a novel patch loss to be applied to the deferred back-propagation as detailed in Eq. 8. When the model renders the full-resolution image with autodifferentiation disabled, the image and message loss are computed by the rendered image, and the derived gradients are cached (see Sec. 4.3 for more details). Then, the model renders the full-resolution image with image patches that are re-rendered with auto-differentiation enabled and back-propagated using the corresponding gradients from the cached set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Our innovation is not solely relying on cached gradients for back-propagation during re-rendering. Instead, we calculate the loss between the re-rendered patch and the ground truth (GT) patch and carry out an additional backpropagation process. The patch loss consists of a sum of L1 loss (MAE), a Total Variation loss, and an SSIM loss, as defined in Eq. 8. By utilizing back-propagation, once for the combined loss from the full-resolution image and the message loss and once for the patch loss, we are able to enhance both image quality and message bit accuracy at once." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.633, + 0.426 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.434, + 0.596, + 0.449 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.594 + ], + "angle": 0, + "content": "We evaluate our method using datasets commonly used by NeRF: the Blender dataset [35], which consists of synthetic bounded scenes, and the LLFF dataset [34], which consists of real unbounded, forward-facing scenes. We follow the conventional literature in NeRF papers, which evaluate and compare using eight scenes each from the Blender dataset (including chair, drums, ficus, hotdog, lego, materials, mic) and the LLFF dataset (including fern, flower, fortress, horns, leaves, orchids, room, trex)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.719, + 0.621 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Vanilla NeRF [35] and TensoRF [3] employ similar training methods, with some specifics as detailed below. The training process is carried out on an image-by-image basis, with a batch size set to one. Ray tracing is implemented along with the camera pose, and the Adam optimizer [16] is used for the learning process. The training is completed with epochs ranging from 5 to 10. Since our pretrained decoder requires full-resolution images, NeRF renders full-resolution images during the fine-tuning process. However, NeRF requires a great amount of memory usage. We adopt the deferred back-propagation in ARF [55] to solve the memory issue and combine it with the patchwise loss to improve rendering quality and bit accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The initial \\(\\lambda_{i}\\), \\(\\lambda_{m}\\) was set to 0.1 and 0.9 for the LLFF dataset and 0.05 and 0.95 for the Blender dataset, respectively. Additionally, within the patch loss equation (Eq. 8), the parameters \\(\\lambda_{MAE}\\), \\(\\lambda_{TV}\\), and \\(\\lambda_{SSIM}\\) were set to 0.1, 0.06, and 0.02." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12091" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.46, + 0.184 + ], + "angle": 0, + "content": "
Bit Accuracy(%) ↑
4 bits8 bits16 bits32 bits48 bits
HiDDeN [59]+NeRF [35]50.3150.2550.1950.1150.04
CopyRNeRF [30]100.00100.0091.1678.0860.06
Ours (w/ NeRF [35])100.00100.0094.2486.8170.43
Ours (w/ TensoRF [3])100.00100.0095.6788.5885.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.47, + 0.245 + ], + "angle": 0, + "content": "Table 1. Bit accuracies with different message lengths are compared with the state-of-the-art method. (HiDDeN [59] + NeRF [35] is pre-embedded before training). The results are the average of LLFF and Blender datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.261, + 0.197, + 0.276 + ], + "angle": 0, + "content": "5.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.471, + 0.526 + ], + "angle": 0, + "content": "We evaluate our method with three key essential aspects in watermarking techniques: 1) capacity, measuring the length of the messages that can be embedded into the image; 2) invisibility, which involves evaluating the difficulty in detecting embedded watermarks in images by peoples; 3) robustness, evaluating the robustness of our watermarking method under various image distortions such as Gaussian Noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. We employ Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [56] for the evaluation metrics to assess invisibility while utilizing bit accuracy to evaluate capacity and robustness. We conduct experiments on the invisibility and robustness of the message length of \\( M_L \\in \\{4,8,16,32,48\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Baseline To the best of our knowledge, CopyRNeRF [30] stands out alone in its ability to embed bits with the implicit NeRF model, which can be converted into various modality data such as messages or images. The CopyRNeRF performs state-of-the-art on bit-accuracy and visual quality. As a result, we conduct our evaluations utilizing the CopyRNeRF. We select the following relevant comparative models:: HiDDeN [59] \\(^+\\) NeRF [35], which employ the classical 2D watermarking method HiDDeN [59] on training images prior to training the NeRF model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.839 + ], + "angle": 0, + "content": "To demonstrate the applicability of our method, we apply our method on two distinct NeRF models: Vanilla NeRF [35], symbolizing implicit representation NeRF, and TensoRF [3], representing explicit representation NeRF based on voxel grids. We select the Vanilla NeRF [35] since it is the first NeRF model to be introduced and can be considered the foundational and ancestral model for all subsequent NeRFs. TensoRF [3] is chosen because it is an influential NeRF model with significantly improved performance and training speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.274, + 0.864 + ], + "angle": 0, + "content": "5.4. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Capacity As shown in Tab. 1, there is a clear trade-off between bit accuracy and message bit length. We conduct bit" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.09, + 0.902, + 0.181 + ], + "angle": 0, + "content": "
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
HiDDeN [59]+NeRF [35]50.1926.530.9170.035
CopyRNeRF [30]91.1626.290.9100.038
Ours (w/ NeRF [35])94.2428.810.9540.025
Ours (w/ TensoRF [3])95.6732.790.9480.033
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.187, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Table 2. Bit accuracies and reconstruction qualities comparison with baselines. \\(\\uparrow (\\downarrow)\\) means higher (lower) is better. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.893, + 0.533 + ], + "angle": 0, + "content": "accuracy tests on messages of lengths 4, 8, 16, 32, and 48 bits. Models implemented with our method show a gradual decrease in bit accuracy as the length of the message increases. However, when compared to CopyRNeRF [30], the decline in bit accuracy is less dramatic and less steep. Notably, CopyRNeRF [30] shows a modest \\(60.6\\%\\) bit accuracy for 48 bits, whereas Vanilla NeRF [35] and TensoRF [3] using our method demonstrates a respectable \\(85.82\\%\\) and \\(70.43\\%\\) accuracy, respectively. Additionally, to demonstrate the stability of our method across a range of messages, we fine-tune 500 different messages, each corresponding to unique, randomly generated bits. We rigorously evaluate the bit accuracy for all images rendered by Vanilla NeRF [35] and TensoRF [3]. As illustrated in Fig. 3, the plot shows a minor descending trend that corresponds with the increasing number of unique messages, and it shows that our method is robust with using many different messages, which is a key value to identify unique NeRF model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Invisibility Our method demonstrates a trade-off between bit accuracy and the quality of reconstructed data, as detailed in Tab. 2. This trade-off is analyzed using several metrics, including bit accuracy and visual quality metrics. With TensoRF [3], our method achieves the best bit accuracy and PSNR results. Conversely, when applied to Vanilla" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.661, + 0.857, + 0.85 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.854, + 0.892, + 0.909 + ], + "angle": 0, + "content": "Figure 3. Identification results. We average the bit accuracy for each 500 messages. Our method not only works efficiently with just one key but also works well with arbitrarily generated messages. We show the results on \\( M_L = 16 \\) bits." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12092" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.891, + 0.171 + ], + "angle": 0, + "content": "
No DistortionGaussian Noise (v = 0.1)Rotation (±π/6)Scaling (25%)Bit Accuracy(%) ↑
Gaussian Blur (deviation = 0.1)Crop (40%)Brightness (2.0)JPEG Compression (10% quality)Combined (Crop, Brightness, JPEG)
CopyRNeRF [30]91.1690.0488.1389.3390.06----
Ours (w/ NeRF [35])94.2494.0685.0291.3594.1283.4884.1486.8873.64
Ours (w/ TensoRF [3])95.6795.3693.1393.2995.2595.4090.9186.9984.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.177, + 0.893, + 0.204 + ], + "angle": 0, + "content": "Table 3. Robustness under diverse attacks compared with the state-of-the-art method. The bit accuracy results are the average of LLFF and Blender datasets. We show the results on \\( M_{L} = 16 \\) bits. The best performances are highlighted in **bold**." + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.216, + 0.196, + 0.23 + ], + "angle": 0, + "content": "CopyRNeRF" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.231, + 0.201, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.288, + 0.201, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.358, + 0.19, + 0.367 + ], + "angle": 0, + "content": "PSNR=32.69" + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.368, + 0.192, + 0.377 + ], + "angle": 0, + "content": "Bit Acc=100%" + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.217, + 0.311, + 0.23 + ], + "angle": 0, + "content": "Ours + NeRF" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.231, + 0.314, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.287, + 0.316, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.358, + 0.303, + 0.367 + ], + "angle": 0, + "content": "PSNR=32.78" + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.368, + 0.306, + 0.377 + ], + "angle": 0, + "content": "Bit Acc=100%" + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.217, + 0.429, + 0.23 + ], + "angle": 0, + "content": "Ours + TensoRF" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.231, + 0.422, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.287, + 0.427, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.358, + 0.413, + 0.367 + ], + "angle": 0, + "content": "PSNR=33.44" + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.368, + 0.417, + 0.377 + ], + "angle": 0, + "content": "Bit Acc=100%" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons We show the differences \\((\\times 10)\\) between the rendered images and the ground truth. Our method achieves higher PSNR and bit accuracy than CopyRNeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.565 + ], + "angle": 0, + "content": "NeRF [35], it yields the highest SSIM and the lowest LPIPS (noting that higher SSIM and lower LPIPS are desirable). Although the SSIM and LPIPS scores for TensoRF [3] using our method are not the highest, they are comparable to those achieved with Vanilla NeRF [35]. In summary, models trained with our method attain superior bit accuracy, PSNR, SSIM, and the lowest LPIPS, indicating a well-balanced trade-off between bit accuracy and reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Additionally, to verify the invisibility of our Result, We compute the differences between ground truth and our rendered results. As illustrated in Fig. 4, the results of applying our method to both Vanilla NeRF [35] and TensoRF [3] achieve a better balance between reconstruction quality and bit accuracy compared to CopyRNeRF [30]. This indicates that our method can embed watermarks with minimal impact on the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.469, + 0.803 + ], + "angle": 0, + "content": "Robustness We evaluate the robustness of the watermarked image to different attacks by applying Gaussian noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. As shown in Tab. 3, We experiment on both Vanilla NeRF [35] and TensorRF [3] when applied with our method, resulting in robustness against various attacks compared to CopyRNeRF [30]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.227, + 0.832 + ], + "angle": 0, + "content": "5.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Patch Loss and Frequency Domain This section delves into the effectiveness of integrating patch loss and utilizing the frequency domain. Fig. 5 presents a comparative analysis of our method, applied to Vanilla NeRF [35], both with" + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.225, + 0.593, + 0.237 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.226, + 0.684, + 0.236 + ], + "angle": 0, + "content": "Full method" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.246, + 0.596, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.248, + 0.691, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.29, + 0.692, + 0.3 + ], + "angle": 0, + "content": "Bit Acc=94.78%" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.22, + 0.783, + 0.242 + ], + "angle": 0, + "content": "Without DWT With Patch" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.246, + 0.788, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.29, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Bit Acc=82.47%" + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.22, + 0.882, + 0.242 + ], + "angle": 0, + "content": "With DWT without patch" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.246, + 0.884, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.29, + 0.884, + 0.3 + ], + "angle": 0, + "content": "Bit Acc=90.78%" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.307, + 0.6, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.307, + 0.696, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.359, + 0.692, + 0.368 + ], + "angle": 0, + "content": "Bit Acc=100.0%" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.307, + 0.792, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.359, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Bit Acc=43.54%" + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.307, + 0.888, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.359, + 0.884, + 0.368 + ], + "angle": 0, + "content": "Bit Acc=100.0%" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.383, + 0.893, + 0.425 + ], + "angle": 0, + "content": "Figure 5. Reconstruction quality comparisons We evaluate our full method, our method without patch loss and our method without frequency domain about 16 bits with NeRF [35]." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.442, + 0.889, + 0.586 + ], + "angle": 0, + "content": "
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
DFT71.0619.040.8220.1719
DCT43.7535.330.9670.0205
No frequency72.6832.800.9440.0392
DWT-Level 191.1633.040.9470.0332
DWT-Level 295.6732.790.9480.0336
DWT-Level 390.9632.450.9500.0331
DWT-Level 485.0231.820.9520.0329
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.661 + ], + "angle": 0, + "content": "Table 4. Bit accuracies and reconstruction qualities compared with spatial and frequency domains (DFT, DCT) and DWT levels 1, 2, 3, and 4. We focus on low frequency as our assumption. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.679, + 0.892, + 0.741 + ], + "angle": 0, + "content": "and without the incorporation of patch loss and use in the frequency domain. As shown in Fig. 5, the absence of DWT significantly decreases bit accuracy. Without the patch-wise loss, there is a decrease in reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "DWT Level During our experiments with DWT, we evaluate performance across commonly used levels: 1, 2, 3, and 4. As shown in Tab. 4, it is indicated that the results from levels 1, 3, and 4 do not perform as well as the outcomes achieved at level 2, both in terms of bit accuracy and reconstruction quality balance. Additionally, we conduct applying DFT, DCT, and no frequency instead of DWT. As a result, our findings show that the DWT at level 2 is the most effective. Thus, we use the DWT at level 2 based on the results from the ablation experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12093" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.082, + 0.424, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.286, + 0.47, + 0.328 + ], + "angle": 0, + "content": "Figure 6. Bit accuracy comparison between DWT Subbands We evaluate the bit accuracy on level 1,2,3 and all subbands. We show the results on \\(M_L = 16\\) bits." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.56 + ], + "angle": 0, + "content": "DWT Subband The choice of DWT subband is as critical as the level of DWT itself. Typically, subbands such as LH, HL, and HH, which are composed of high-frequency components less perceptible to the people, are employed for watermark embedding in conventional watermarking techniques. Conversely, in our experimental approach, we opt for the LL subband representing the image's coarse approximation and containing essential low-frequency details. This decision is based on the watermark decoder's pre-training process to extract watermarks specifically from the image's spatial domain, which is visually similar to the LL subband and its demonstrated robustness to JPEG compression. In Fig. 6, our empirical investigations confirm that the LL subband yields the highest bit accuracy." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.57, + 0.468, + 0.796 + ], + "angle": 0, + "content": "Comparison of Training time. Even though our proposed method necessitates the use of a pre-trained model's weights as a starting point for fine-tuning NeRF, the time requirement for training is significantly less compared to that of CopyRNeRF [30]. As depicted in Fig. 7, it is evident that the model utilizing our techniques substantially outperforms CopyRNeRF [30] regarding training efficiency. Our method, when applied to an explicit NeRF architecture, such as TensoRF [3], achieves a remarkable speed increase, performing up to six times faster than CopyRNeRF [30]. Similarly, implementing our techniques within the Implicit NeRF framework, like Vanilla NeRF [35], yields a significant speed improvement, with our model operating six times faster than CopyRNeRF [30] but it also surpasses the bit accuracy of the compared model." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.814, + 0.194, + 0.83 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We propose a neural 3D watermarking method for the NeRF model. Our method trains a 2D watermark decoder and NeRF model separately. Therefore, our pipeline only requires training the decoder once and re-use it on differ" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.078, + 0.89, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.262, + 0.89, + 0.331 + ], + "angle": 0, + "content": "Figure 7. Analysis of train time for reconstruction from blender dataset at \\(160 \\times 160\\) resolution. Note the introduction of a scale break within the graph to accommodate the significant disparities in training durations, with time represented in minutes on the left axis and in days on the right axis." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.346, + 0.89, + 0.481 + ], + "angle": 0, + "content": "ent NeRF watermark models. We adopt conventional watermark techniques in image watermarking, which transfer the image from the spatial domain to the frequency domain to encode the watermark into the image efficiently. We find that DWT and patch-wise loss enhance the overall reconstruction quality while performing high message bit accuracy. Through extensive experiments, we demonstrate that our method outperforms the state-of-the-art method and is six times faster than the previous work." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.493, + 0.89, + 0.673 + ], + "angle": 0, + "content": "Limitations. Our method shows outstanding performance in the digital watermark of NeRF, but training a decoder that extracts watermark from rendered images is time-consuming, taking approximately 12 hours on a single RTX 3090. Our method only encodes a unique message per model, requiring fine-tuning for message insertion every time. We observe that the bit accuracy drops when we apply numerous messages (see Fig. 3). Although previous studies, including us, report robustness against various distortions of images, there is no consideration about cases where the model is attacked. Future work may explore watermarks resilient to cases of model attacks." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.689, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. This research was supported by the Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture(Project : 4D Content Generation and Copyright Protection with Artificial Intelligence, R2022020068, \\(80\\%\\) ), the Korea Research Institute for Defense Technology Planning and Advancement (KRIT) Grant funded by the Korean Government(Grant KRIT-CT-23-021, \\(10\\%\\) ), the National Research Foundation of Korea grant (NRF2022R1F1A1074334, \\(9\\%\\) ), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), \\(1\\%\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12094" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[4] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[5] Chinmayee Das, Swetalina Panigrahi, Vijay K Sharma, and KK Mahapatra. A novel blind robust image watermarking in dct domain using inter-block coefficient correlation. AEU-International Journal of Electronics and Communications, 68(3):244–253, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[6] Xin Deng, Chao Gao, and Mai Xu. Pirnet: Privacy-preserving image restoration network via wavelet lifting. In ICCV, pages 22368-22377, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.492 + ], + "angle": 0, + "content": "[7] Elhameh Farri and Peyman Ayubi. A robust digital video watermarking based on ct-svd domain and chaotic dna sequences for copyright protection. Journal of Ambient Intelligence and Humanized Computing, 14(10):13113-13137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.493, + 0.471, + 0.548 + ], + "angle": 0, + "content": "[8] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. ICCV, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[9] Felipe ABS Ferreira and Juliano B Lima. A robust 3d point cloud watermarking method based on the graph fourier transform. Multimedia Tools and Applications, 79(3):1921-1950, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.606, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.662, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[11] Youcai Gao, Jinwei Wang, and Yun-Qing Shi. Dynamic multi-watermarking and detecting in dwt domain. Journal of Real-Time Image Processing, 16:565-576, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.705, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[12] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[13] A Hernández-Joaquín, G Melendez-Melendez, and R Cumplido. A secure dwt-based dual watermarking scheme for image authentication and copyright protection. Multimedia Tools and Applications, 82(27):42739-42761, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[14] Ningombam Jimson and K Hemachandran. Dft based coefficient exchange digital image watermarking. In 2018 Second International Conference on Intelligent Computing and Control Systems (ICICCS), pages 567-571. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Priyank Khare and Vinay Kumar Srivastava. A reliable and secure image watermarking algorithm using homomorphic transform in dwt domain. Multidimensional Systems and Signal Processing, 32:131-160, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[16] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[17] Hung-Jui Ko, Cheng-Ta Huang, Gwoboa Horng, and WANG Shiuh-Jeng. Robust and blind image watermarking in dct domain using inter-block coefficient correlation. Information Sciences, 517:128-147, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[18] Sanjay Kumar and Binod Kumar Singh. Dwt based color image watermarking using maximum entropy. Multimedia Tools and Applications, 80:15487-15510, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.319 + ], + "angle": 0, + "content": "[19] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[20] Guezouli Larbi, Hachani Roumaissa, and Herri Sara. Embedding watermark in the magnitude matrix of the dft of image. In Proceedings of the 2018 International Conference on Computing and Pattern Recognition, pages 106-110, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.893, + 0.445 + ], + "angle": 0, + "content": "[21] Yong-Seok Lee, Seo Young-Ho, and Kim Dong-Wook. Blind image watermarking based on adaptive data spreading in n-level dwt subbands. Security and communication Networks, 2019, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.893, + 0.502 + ], + "angle": 0, + "content": "[22] Chenxin Li, Brandon Y Feng, Zhiwen Fan, Panwang Pan, and Zhangyang Wang. Steganerf: Embedding invisible information within neural radiance fields. In ICCV, pages 441-453, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.893, + 0.589 + ], + "angle": 0, + "content": "[24] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, pages 4273-4284, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.893, + 0.657 + ], + "angle": 0, + "content": "[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[26] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.893, + 0.759 + ], + "angle": 0, + "content": "[27] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR, pages 13548-13557, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[28] Xiyang Luo, Michael Goebel, Elnaz Barshan, and Feng Yang. Leca: A learned approach for efficient cover-agnostic watermarking. In Media Watermarking, Security, and Forensics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[29] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[30] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12095" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "radiance fields. In ICCV, pages 22401-22411, 2023. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[31] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[37] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. ACM Trans. Graph., 41(4), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[38] Shabir A Parah, Javaid A Sheikh, Nazir A Loan, and Ghulam M Bhat. Robust and blind watermarking technique in dct domain using inter-block coefficient differencing. Digital Signal Processing, 53:11-24, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[40] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[41] SN Prajwalasimha, S Chethan Suputhra, and C Mohan. Performance analysis of combined discrete fourier transformation (dft) and successive division based image watermarking scheme. International Journal of Recent Technology and Engineering, 8(3):34-39, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318-10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] Deepti Shukla, Nirupama Tiwari, and Deepika Dubey. Survey on digital watermarking techniques. International Journal of Signal Processing, Image Processing and Pattern Recognition, 9(1):239-244, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[44] S Suthaharan and S Sathananthan. Transform domain technique: robust watermarking for digital images. In Proceedings of the IEEE SoutheastCon 2000. Preparing for The New Millennium' (Cat. No. 00CH37105), pages 407-412. IEEE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[45] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[46] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[48] Jinwei Wang, Shiguo Lian, and Yun-Qing Shi. Hybrid multiplicative multi-watermarking in dwt domain. Multidimensional Systems and Signal Processing, 28:617-636, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[49] Yinghui Wang, Jing Liu, Yajie Yang, Douli Ma, and Ruijiao Liu. 3d model watermarking algorithm robust to geometric attacks. IET Image Processing, 11(10):822-832, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[50] Muyu Xu, Fangneng Zhan, Jiahui Zhang, Yingchen Yu, Xiaojin Zhang, Christian Theobalt, Ling Shao, and Shijian Lu. Wavenerf: Wavelet-based generalizable neural radiance fields. In ICCV, pages 18195-18204, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[51] Innfarn Yoo, Huiwen Chang, Xiyang Luo, Ondrej Stava, Ce Liu, Peyman Milanfar, and Feng Yang. Deep 3d-to-2d watermarking: embedding messages in 3d meshes and extracting them from 2d renderings. In CVPR, pages 10031–10040, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[52] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[53] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[54] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In CVPR, pages 12397-12406, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[55] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, pages 717-733. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[57] Qiumei Zheng, Nan Liu, Baoqin Cao, Fenghua Wang, and Yanan Yang. Zero-watermarking algorithm in transform domain based on rgb channel and voting strategy. Journal of Information Processing Systems, 16(6):1391-1406, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[58] Haoyi Zhu. X-nerf: Explicit neural radiance field for multiscene 360deg insufficient rgb-d views. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5766–5775, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12096" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.136 + ], + "angle": 0, + "content": "[59] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, pages 657–672, 2018. 2, 3, 6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12097" + } + ] +] \ No newline at end of file diff --git a/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_origin.pdf b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e9b5edaa96ab7ef511bd95f5659b64660fab310 --- /dev/null +++ b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/36a5a6f2-e177-4dfb-ad96-1aaeb75afb72_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c07a1d35b963ddcefd3cccfe64cd1b74126590af2ce14d16a55c9c0cb6a03532 +size 1501375 diff --git a/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/full.md b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/full.md new file mode 100644 index 0000000000000000000000000000000000000000..483b3c120d84dcdeba6b0fe8c53227cbc4f5f1ac --- /dev/null +++ b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/full.md @@ -0,0 +1,354 @@ +# WaterRF: Robust Watermarks in Radiance Fields for Protection of Copyrights + +Youngdong Jang $^{1}$ Dong In Lee $^{1}$ MinHyuk Jang $^{1}$ Jong Wook Kim $^{1}$ Feng Yang $^{2}$ Sangpil Kim $^{1*}$ $^{1}$ Korea University $^{2}$ Google Research + +# Abstract + +The advances in the Neural Radiance Fields (NeRF) research offer extensive applications in diverse domains, but protecting their copyrights has not yet been researched in depth. Recently, NeRF watermarking has been considered one of the pivotal solutions for safely deploying NeRF-based 3D representations. However, existing methods are designed to apply only to implicit or explicit NeRF representations. In this work, we introduce an innovative watermarking method that can be employed in both representations of NeRF. This is achieved by fine-tuning NeRF to embed binary messages in the rendering process. In detail, we propose utilizing the discrete wavelet transform in the NeRF space for watermarking. Furthermore, we adopt a deferred back-propagation technique and introduce a combination with the patch-wise loss to improve rendering quality and bit accuracy with minimum trade-offs. We evaluate our method in three different aspects: capacity, invisibility, and robustness of the embedded watermarks in the 2D-rendered images. Our method achieves state-of-the-art performance with faster training speed over the compared state-of-the-art methods. Project page: https://kuai-lab.github.io/cvpr2024waterf/ + +# 1. Introduction + +Digital watermarking plays a pivotal role in reinforcing the copyright of digital assets, e.g., text, image, audio, video, and 3D content. Digital assets are easily misused without permission from the creators and owners of the digital assets. One way to protect the copyright is by encoding invisible watermarks in the digital assets. + +Neural Radiance Fields (NeRF) [35] have emerged into the spotlight in 3D content creation and modeling since NeRF can represent 3D objects or scenes in a compact way. With increasing interest in 3D content for the Metaverse, virtual reality, and augmented reality, NeRF has become important in digital watermarking research. + +A straightforward approach for protecting rendered 3D + +![](images/7ca51836b73624f0772d63f0f8ae808cd9047232350cef2002b3168eccfce436.jpg) +Figure 1. The NeRF model can be fine-tuned to embed the watermark into the images of the novel view. The figure above shows the model owner Alice and her NeRF model is stolen by Bob. Alice can identify the copyright by using her watermark decoder. Our method embeds the watermark into all of the rendered images. + +images from the NeRF model is using the post-generation watermarking method that embeds the watermark into the 2D rendered image from NeRF with existing watermarking methods. However, this method only protects copyright for the rendered images but not the copyright of the NeRF model. A fundamental solution to protecting the model and rendered image is fine-tuning the NeRF model to encode the watermark into the NeRF model itself. The watermark will be encoded into the rendered image from the fine-tuned NeRF every time during the rendering process. + +NeRF has two primary representations: implicit and explicit representations. The implicit representation [1, 35, 37] uses a multi-layer perceptron (MLP) to represent a 3D scene and the explicit representation [3, 10, 58] uses a traditional 3D structure, such as voxels. Prior methods [22, 30] were limited to applying to only one type of NeRF representation. However, in order to overcome this limitation, we explore a new question: How can we embed the watermark into two distinct NeRF representations(implicit and explicit)? + +In this paper, we propose an innovative watermarking method that can be applied to both types of NeRF representations. Our method integrates the watermarking process into the rendering process without changing the model architecture. It adjusts the pre-trained NeRF model so that all + +of the rendered images have same watermark embedded. We pre-train the decoder with the conventional deep learning method HiDDeN [59] to extract the watermark from the rendered image. Our method has several advantages. First, since it is a fine-tuning process, we can deal with the implicit and the explicit NeRF representations. Second, we do not need any extra processes for concealing watermarks in the rendered images. Therefore, less computational cost is required. Third, our method can protect the model and the rendered images simultaneously. + +Invisibility and robustness are very important aspects in the digital watermarking domain. Concealing watermarks in the rendered image without degrading the original image quality and ensuring identification of the watermark is challenging. For instance, an embedded watermark in the image is better to be as invisible as possible, and also watermark should be identified after various distortion attacks. We intend to hide the watermark message in the low-level subband in the frequency domain by carefully designing the loss function that evaluates the loss value in the frequency domain. To enhance the image quality, we consider the local structure of the object by introducing patch-wise loss. We subdivided the rendered image and cached the gradient in the grid to encode the watermark locally in the image. + +Our extensive experiments demonstrate that our method successfully encodes watermarks by fine-tuning the pretrained NeRF models and identifying the watermark message using a simplified decoder from HiDDeN [59]. Moreover, we show that our method is robust from diverse watermark attacks. We evaluate our method by measuring the capacity and robustness under diverse watermark attacks. The training time of our method is approximately six times faster than the CopyRNeRF [30]. Since our method can be applied to both implicit [35] and explicit [3] NeRF, our method can be used in more general applications than the other watermarking methods. Our method outperforms the other state-of-the-art watermarking methods within all metrics. In summary, our contributions are as follows: + +- Our method can be applied to implicit and explicit NeRF models, unlike other existing watermarking methods. +- We propose a novel watermarking method for NeRF that fine-tunes the NeRF model by minimizing the loss function which is evaluated in the frequency domain. +- We propose a patch-wise loss to improve rendering quality and bit accuracy and enable encoding the watermark locally in the image, reducing the color artifacts. +- The proposed watermarking method achieves state-of-the-art performance, and we show that our methods are robust in diverse watermark attacks. + +# 2. Related Work + +Neural Radiance Fields. Due to the success of Neural Radiance Field (NeRF) [35], a highly photo-realistic + +view synthesis of complex scenes has been achieved. Recently, NeRF has been used in various research and applications, including faster inference [26, 52, 54], 3D reconstruction [32, 39, 53], image processing [19, 31, 47], dynamic scenes [24, 42] and generative models [25, 33, 40]. Therefore, NeRF become the dominant 3D representation and is widely used. Several commercial products [36, 45] utilized NeRF 3D representations. Thus, the management of the copyright of NeRF-based 3D representations has been emerging as a crucial aspect. As all of those applications are based on implicit, explicit, or both representations, we explore a versatile watermarking method that can be trained on both representations. + +Digital watermarking. Watermarking methods have been evolving over the decades [2, 43]. There are two main ways to recover watermarks [23]: multi-bit watermark and zero-bit watermark. The multi-bit watermark allows for the encoding and decoding of multi-bit messages from media. The zero-bit watermark can be used for ownership authentication. In this paper, we focus on the zero-bit watermark, which involves encoding and decoding multi-bit messages. Recently, with the great success of deep learning watermark methods like HiDDeN [59], considerable research has been conducted to embed watermarks into carriers, especially into 2D images [8, 27, 28], video [7, 29] and 3D [9, 49, 51]. + +On the NeRF field, the watermarks do not survive during the volume rendering. To overcome this issue, several methods [22, 30] have emerged. StegaNeRF [22] designed a steganography model on explicit NeRF that hides natural images in 3D scene representation. CopyRNeRF [30] secured the copyright of images rendered from an implicit 3D model by embedding a watermark into the rendered color representation. However, these methods have two big limitations. First, they could only be applied to one of the implicit or explicit representations, while both representations of NeRF greatly impact the advancement of NeRF. Second, these watermarks were not robust enough. Since StegaNeRF required the original image for message extraction, this method was not robust to distortion-like cropping, which made a lot of difference between the original and rendered images and CopyRNeRF did not resist cropping. To address these issues, we design a robust watermarking NeRF model that not only can be utilized on both implicit and explicit representation but is also robust to diverse attacks. + +Frequency Domain. It is known that embedding a watermark into the spatial domain is vulnerable to attacks such as cropping and compression because it directly modifies the image pixels [46]. For this reason, many studies have been conducted based on frequency domains [44, 57]. The following are the frequency transformation methods. + +- Discrete Fourier Transform (DFT) expresses images in terms of phase and amplitude. When a watermark is em + +bedded into the frequency domain converted by DFT, the magnitude is invariant, making it robust against geometric attacks such as rotation and scaling [14, 20, 41]. + +- Discrete Cosine Transform (DCT) decomposes the energy of image data into a sum of cosine functions by representing signals in the frequency domain. It is useful for image compression because most of the energy of the image is concentrated in the top-left corner coefficient. Therefore, DCT is extensively used in watermarking methods that energy compaction. [5, 17, 38] +- Discrete Wavelet Transform (DWT) decomposes the image into four subbands, LL(Low-Low), LH(Low-High), HL(High-Low), and HH(High-High). The LL band has the most energy and contains low-frequency information. Additionally, the LL band can be recursively decomposed into n-levels. DWT is used in several watermarking methods [6, 50] and shows significant differences in performance depending on the level [11, 21, 48]. In this paper, we use DWT and the LL subband for fine-tuning NeRF to embed the watermark. + +# 3. Preliminaries: Representation of NeRF + +In this paper, we focus on both implicit and explicit representations of a NeRF model. The implicit NeRF [35] represents a scene using a multi-layer perceptrs (MLP) $\Phi$ whose input is 3D location $\pmb{x} = (x,y,z)\in \mathbb{R}^3$ and 2D viewing direction $d\in \mathbb{R}^2$ and output is color $c\in \mathbb{R}^3$ , volume density $\sigma \in \mathbb{R}^{+}$ . These MLP-based radiance fields can be written as: + +$$ +\sigma , c = \Phi (x, d) \tag {1} +$$ + +The explicit NeRF, such as TensoRF [3], models a scene as a 4D tensor, which represents a 3D voxel grid. These grid-based radiance fields can be written as: + +$$ +\sigma , \boldsymbol {c} = \mathcal {T} _ {\sigma} (\boldsymbol {x}), \mathcal {S} \left(\mathcal {T} _ {c} (\boldsymbol {x}), \boldsymbol {d}\right), \tag {2} +$$ + +where $\mathcal{T}_{\sigma} \in \mathbb{R}^{I \times J \times K}$ is a geometry tensor and $\mathcal{T}_c^{I \times J \times K \times C}$ is an appearance tensor for channel $C$ and the voxel grid resolution $I, J, K$ . $S$ is a decoding function including MLP and spherical harmonics function. + +Both representations are followed by a volumetric rendering equation, which is used to synthesize novel views: + +$$ +\begin{array}{l} C (r) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\boldsymbol {r} (t)) \boldsymbol {c} (\boldsymbol {r} (t), \boldsymbol {d}) d t, \tag {3} \\ T (t) = e x p \left(- \int_ {t _ {n}} ^ {t} \sigma (\boldsymbol {r} (s)) d s\right), \\ \end{array} +$$ + +where $\pmb{r}(t)$ is a camera ray with near and far bounds $t_n$ and $t_f$ . $C(r)$ is an expected color of the ray $\pmb{r}(t)$ . + +Considering the NeRF representation type like the above, how can we embed a watermark in the various types, not just one type? Since these types of NeRF share a common property, rendering images through the same volume rendering function, we use the rendered images to embed watermarks into the models. + +# 4. Method + +We propose a process of fine-tuning the NeRF that does not involve alterations to the model's architecture to embed the watermark message. Our method aims to embed the watermark into the weights $\theta$ of the NeRF model in the frequency domain of rendered images. Our method stands apart from traditional watermarking methods, which focusing on training encoders and decoders. The difference lies in the fine-tuning process, which embeds the watermark without using an encoder. There are two phases: (1) Pre-training the watermark decoder $D$ , (2) Fine-tuning the NeRF model $F_{\theta}$ to embed the message. Our method is illustrated in Fig. 2 and described in detail below. + +# 4.1. Pre-training the watermark decoder + +We select HiDDeN [59] architecture as our watermark decoder. HiDDeN comprises two convolution networks for data hiding: a watermark encoder $E$ and a watermark decoder $D$ . For robustness, it includes a noise layer $N$ . However, in this training phase, where we focus solely on the decoder's performance, we have excluded an adversarial loss responsible for improving visual quality. After training the HiDDeN model, watermark encoder $E$ was not utilized in the second phase. + +The encoder $E$ takes in a cover image $I_{o}\in \mathbb{R}^{H\times W\times 3}$ and a binary message $M\in \{0,1\} ^L$ with length $L$ , as input. Then $E$ embeds $M$ into $I_{o}$ and produces an encoded image $I_w$ . In order to make the decoder resistant to various distortions such as rotation and JPEG compression, $I_{w}$ is transformed with a noise layer $N$ . The decoder $D$ , made of several convolution layers, receives $I_{w}$ as input and extracts a message $M^{\prime}$ . + +$$ +M ^ {\prime} = D \left(N \left(I _ {w}\right)\right) \tag {4} +$$ + +We utilize a sigmoid function to set the range of the extracted message $M'$ between [0, 1]. The message loss is calculated with Binary Cross Entropy (BCE) between $M_L$ and sigmoid $sg(M_L')$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m e s s a g e}} = - \sum_ {i = 1} ^ {L} M _ {i} \cdot \log s g \left(M _ {i} ^ {\prime}\right) \\ + \left(1 - M _ {i}\right) \cdot \log \left(1 - s g \left(M _ {i} ^ {\prime}\right)\right) \tag {5} \\ \end{array} +$$ + +The decoder is trained to detect watermarks in images that have passed through the trained encoder. However, we do not use the encoder in the second stage. We find that when the decoder receives a vanilla-rendered image, there is a bias between the extracted message bits. Thus, after training the decoder, we conduct PCA whitening to a linear decoder layer to remove the bias without reducing the extraction ability. + +![](images/ec1956eca51d7a48ffcfcdabaa3a64a0e8b423bc0f07c678c384b7030dfc946c.jpg) + +![](images/29f81458ee65570aaea6e4b94bfdfcf97b576b7da73d2c994978ca3f9cb0d1d1.jpg) +Figure 2. WaterF overview. Phase 1: We train the encoder and the decoder to extract messages. After phase 1, we do not use the encoder. Phase 2: We fine-tune the NeRF to embed the messages into the rendered images. (a) We disable auto-differentiation and render a full-resolution image to save memory. (b) We use DWT for the rendered images and choose the LL subband as the input of the pre-trained decoder. (c) We enable auto-differentiation and render the images patch by patch. Then the NeRF is optimized using Eq 7 and Eq 8. + +# 4.2. Embedding and extracting watermark on DWT + +Recently, a fine-tuning watermarking method for NeRF [22] in the spatial domain has emerged. Although the fine-tuning method of embedding a message in the spatial domain shows incomparable invisibility and message extraction ability, it is vulnerable to attacks that distort the spatial domain, such as cropping. Directly applying spatial domain techniques from the latent diffusion model [8] does not allow for effective adjustment of NeRF's weights. + +To tackle these problems, we propose a fine-tuning method in a frequency domain instead of a spatial domain. Various watermarking techniques for images use the frequency domain have seen development and improvement over the years. We find that DWT is the appropriate domain for encoding the message into the weights of a NeRF model. + +The NeRF model renders diverse views of the 3D model given corresponding camera parameters. We transform the pixels of the rendered images, denoted by $X = (x_{c},y_{c})\in \mathbb{R}^{H\times W\times 3}$ , into wavelet forms, with $c$ representing the channel. The DWT is defined as [12]: + +$$ +\begin{array}{l} W _ {\varphi} (j _ {0}, m, n) = \frac {1}{\sqrt {M N}} \sum_ {x _ {c} = 0} ^ {M - 1} \sum_ {y _ {c} = 0} ^ {N - 1} f (x _ {c}, y _ {c}) \varphi_ {j _ {0}, m, n} (x _ {c}, y _ {c}), \\ W _ {\psi} ^ {i} (j, m, n) = \frac {1}{\sqrt {M N}} \sum_ {x _ {c} = 0} ^ {M - 1} \sum_ {y _ {c} = 0} ^ {N - 1} f \left(x _ {c}, y _ {c}\right) \psi_ {j, m, n} ^ {i} \left(x _ {c}, y _ {c}\right) \tag {6} \\ \end{array} +$$ + +where $\varphi (x,y)$ is a scaling function and $\psi (x,y)$ is a wavelet function. $W_{\varphi}(j_0,m,n)$ is called by an LL subband, which is an approximation of the image at scale $j_0$ $W_{\psi}^{i}$ represents LH, HL, HH subbands, where $i = \{\mathrm{H},\mathrm{V},\mathrm{D}\}$ and each denotes horizontal, vertical and diagonal coefficients. + +Previous studies [13, 15, 18] have selected $LH$ , $HL$ , and $HH$ subbands for embedding watermarks because the $LL$ subband contains significant information about the image. However, we choose the LL subband as an input of our decoder $D$ and get the extract message with $M' = D(W_{\varphi})$ . We experimentally discover that embedding the watermark in the LL subband is more robust and effective than other subbands for the HiDDeN decoder. + +The DWT is characterized by its subbands being computed across different levels; therefore, choosing an optimal level for our purpose is necessary. The 1-level separates the images to 4 subbands $(LL_{1}, LH_{1}, HL_{1}, HH_{1})$ , then 2-level separate the $LL_{1}$ subband into 4 subbands $(LL_{2}, LH_{2}, HL_{2}, HH_{2})$ . We experimentally discover that selecting a level too high decreases visual quality due to excessive adjustment of crucial visual elements. Thus, we use the 2-level DWT because it best maintains a good balance between bit accuracy and reconstruction quality. + +# 4.3. Fine-tuning the NeRF model + +NeRF has two different representations, implicit and explicit representations (see Sec. 3). Previous studies [22, 30] focused on applying their methods to just one form of NeRF + +representation. In contrast, we propose a method applicable to both NeRF representations: implicit and explicit. Our approach involves fine-tuning the NeRF to ensure that novel view images include an embedded message. + +Our method starts with preparing each pre-trained decoder for different lengths of message bits and also preparing an initial NeRF model $F_{\theta_0}$ . Then we establish fixed binary messages $M = (m_1, \dots, m_L) \in \{0, 1\}^L$ . + +Since rendering with NeRF requires a lot of memory, we turn off the auto-differentiation at first. The rendered images $X \in \mathbb{R}^{H \times W \times 3}$ at full resolution are rendered by $F_{\theta_0}$ . Following this, the full resolution images $X$ are transformed into seven groups of wavelet subbands: $\{LL_2, LH_2, HL_2, HH_2, LL_1, LH_1, HL_1, HH_1\}$ . Our pre-trained decoder $D$ takes the $LL_2$ subband as an input and extracts a message $M' = D(W_{\varphi}(j_0, m, n))$ from it. The message loss is calculated by comparing $M$ and $M'$ : $L_m = BCE(M, sg(M'))$ . We choose the Waston-VGG [4] to calculate the full image perceptual loss between the image without watermark $X_0$ and the image with a watermark $X$ : $L_i(X_0, X)$ . Full resolution loss function is a sum of the full image perceptual loss and the message loss. + +$$ +\mathcal {L} _ {\text {f u l l}} = \lambda_ {i} \mathcal {L} _ {\text {p e r c e p t}} + \lambda_ {m} \mathcal {L} _ {\text {m e s s a g e}} \tag {7} +$$ + +We cache the gradient calculated by Eq. 7 prior to the patch-wise rendering process to update the parameters. Then, we turn on the auto-differentiation so that the NeRF parameters enable patch-wise rendering. Our patch loss $L_{patch}$ is calculated by combining mean absolute error (MAE) across rendered pixel colors, total variation (TV) regularization, and SSIM loss to maintain the balance between bit-accuracy and visual quality. + +$$ +\mathcal {L} _ {\text {p a t c h}} = \lambda_ {M A E} \mathcal {L} _ {M A E} + \lambda_ {T V} \mathcal {L} _ {T V} + \lambda_ {S S I M} \mathcal {L} _ {S S I M} \tag {8} +$$ + +Then we optimize from the initial NeRF model $F_{\theta_0}$ to watermarked $F_{\theta}$ using Eq. 7 and Eq. 8. + +# 4.4. Deferred back-propagation with patch loss + +In conventional NeRF models, rendering at full resolution is inefficient regarding memory usage. To solve the memory-inefficient problem, pixels are often sampled and rendered randomly to compute the L2 loss for training. However, the main limitation of random pixel-wise rendering is in its incompatibility with using CNN-based perceptual loss and our watermark decoder for extracting message bits, both of which require a full-resolution image as input. + +Our method implements a full-resolution rendering derived from our pre-trained decoder during the fine-tuning process. Each iteration presents a considerable challenge due to the extensive memory consumption. In order to relieve the memory constraints, we adopt the deferred back-propagation [55], specifically designed for memory-efficient, patch-wise optimization within NeRF. + +Our method proposes a novel patch loss to be applied to the deferred back-propagation as detailed in Eq. 8. When the model renders the full-resolution image with autodifferentiation disabled, the image and message loss are computed by the rendered image, and the derived gradients are cached (see Sec. 4.3 for more details). Then, the model renders the full-resolution image with image patches that are re-rendered with auto-differentiation enabled and back-propagated using the corresponding gradients from the cached set. + +Our innovation is not solely relying on cached gradients for back-propagation during re-rendering. Instead, we calculate the loss between the re-rendered patch and the ground truth (GT) patch and carry out an additional backpropagation process. The patch loss consists of a sum of L1 loss (MAE), a Total Variation loss, and an SSIM loss, as defined in Eq. 8. By utilizing back-propagation, once for the combined loss from the full-resolution image and the message loss and once for the patch loss, we are able to enhance both image quality and message bit accuracy at once. + +# 5. Experiments + +# 5.1. Dataset + +We evaluate our method using datasets commonly used by NeRF: the Blender dataset [35], which consists of synthetic bounded scenes, and the LLFF dataset [34], which consists of real unbounded, forward-facing scenes. We follow the conventional literature in NeRF papers, which evaluate and compare using eight scenes each from the Blender dataset (including chair, drums, ficus, hotdog, lego, materials, mic) and the LLFF dataset (including fern, flower, fortress, horns, leaves, orchids, room, trex). + +# 5.2. Implementation Details + +Vanilla NeRF [35] and TensoRF [3] employ similar training methods, with some specifics as detailed below. The training process is carried out on an image-by-image basis, with a batch size set to one. Ray tracing is implemented along with the camera pose, and the Adam optimizer [16] is used for the learning process. The training is completed with epochs ranging from 5 to 10. Since our pretrained decoder requires full-resolution images, NeRF renders full-resolution images during the fine-tuning process. However, NeRF requires a great amount of memory usage. We adopt the deferred back-propagation in ARF [55] to solve the memory issue and combine it with the patchwise loss to improve rendering quality and bit accuracy. + +The initial $\lambda_{i}$ , $\lambda_{m}$ was set to 0.1 and 0.9 for the LLFF dataset and 0.05 and 0.95 for the Blender dataset, respectively. Additionally, within the patch loss equation (Eq. 8), the parameters $\lambda_{MAE}$ , $\lambda_{TV}$ , and $\lambda_{SSIM}$ were set to 0.1, 0.06, and 0.02. + +
Bit Accuracy(%) ↑
4 bits8 bits16 bits32 bits48 bits
HiDDeN [59]+NeRF [35]50.3150.2550.1950.1150.04
CopyRNeRF [30]100.00100.0091.1678.0860.06
Ours (w/ NeRF [35])100.00100.0094.2486.8170.43
Ours (w/ TensoRF [3])100.00100.0095.6788.5885.82
+ +# 5.3. Evaluation + +We evaluate our method with three key essential aspects in watermarking techniques: 1) capacity, measuring the length of the messages that can be embedded into the image; 2) invisibility, which involves evaluating the difficulty in detecting embedded watermarks in images by peoples; 3) robustness, evaluating the robustness of our watermarking method under various image distortions such as Gaussian Noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. We employ Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [56] for the evaluation metrics to assess invisibility while utilizing bit accuracy to evaluate capacity and robustness. We conduct experiments on the invisibility and robustness of the message length of $M_L \in \{4,8,16,32,48\}$ . + +Baseline To the best of our knowledge, CopyRNeRF [30] stands out alone in its ability to embed bits with the implicit NeRF model, which can be converted into various modality data such as messages or images. The CopyRNeRF performs state-of-the-art on bit-accuracy and visual quality. As a result, we conduct our evaluations utilizing the CopyRNeRF. We select the following relevant comparative models:: HiDDeN [59] $^+$ NeRF [35], which employ the classical 2D watermarking method HiDDeN [59] on training images prior to training the NeRF model. + +To demonstrate the applicability of our method, we apply our method on two distinct NeRF models: Vanilla NeRF [35], symbolizing implicit representation NeRF, and TensoRF [3], representing explicit representation NeRF based on voxel grids. We select the Vanilla NeRF [35] since it is the first NeRF model to be introduced and can be considered the foundational and ancestral model for all subsequent NeRFs. TensoRF [3] is chosen because it is an influential NeRF model with significantly improved performance and training speed. + +# 5.4. Experimental results + +Capacity As shown in Tab. 1, there is a clear trade-off between bit accuracy and message bit length. We conduct bit + +Table 1. Bit accuracies with different message lengths are compared with the state-of-the-art method. (HiDDeN [59] + NeRF [35] is pre-embedded before training). The results are the average of LLFF and Blender datasets. + +
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
HiDDeN [59]+NeRF [35]50.1926.530.9170.035
CopyRNeRF [30]91.1626.290.9100.038
Ours (w/ NeRF [35])94.2428.810.9540.025
Ours (w/ TensoRF [3])95.6732.790.9480.033
+ +Table 2. Bit accuracies and reconstruction qualities comparison with baselines. $\uparrow (\downarrow)$ means higher (lower) is better. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold. + +accuracy tests on messages of lengths 4, 8, 16, 32, and 48 bits. Models implemented with our method show a gradual decrease in bit accuracy as the length of the message increases. However, when compared to CopyRNeRF [30], the decline in bit accuracy is less dramatic and less steep. Notably, CopyRNeRF [30] shows a modest $60.6\%$ bit accuracy for 48 bits, whereas Vanilla NeRF [35] and TensoRF [3] using our method demonstrates a respectable $85.82\%$ and $70.43\%$ accuracy, respectively. Additionally, to demonstrate the stability of our method across a range of messages, we fine-tune 500 different messages, each corresponding to unique, randomly generated bits. We rigorously evaluate the bit accuracy for all images rendered by Vanilla NeRF [35] and TensoRF [3]. As illustrated in Fig. 3, the plot shows a minor descending trend that corresponds with the increasing number of unique messages, and it shows that our method is robust with using many different messages, which is a key value to identify unique NeRF model. + +Invisibility Our method demonstrates a trade-off between bit accuracy and the quality of reconstructed data, as detailed in Tab. 2. This trade-off is analyzed using several metrics, including bit accuracy and visual quality metrics. With TensoRF [3], our method achieves the best bit accuracy and PSNR results. Conversely, when applied to Vanilla + +![](images/c593b16d8e0edacbafacebd5ef85a62007188989298e4a486bbd0ce2636a1037.jpg) +Figure 3. Identification results. We average the bit accuracy for each 500 messages. Our method not only works efficiently with just one key but also works well with arbitrarily generated messages. We show the results on $M_L = 16$ bits. + +
No DistortionGaussian Noise (v = 0.1)Rotation (±π/6)Scaling (25%)Bit Accuracy(%) ↑
Gaussian Blur (deviation = 0.1)Crop (40%)Brightness (2.0)JPEG Compression (10% quality)Combined (Crop, Brightness, JPEG)
CopyRNeRF [30]91.1690.0488.1389.3390.06----
Ours (w/ NeRF [35])94.2494.0685.0291.3594.1283.4884.1486.8873.64
Ours (w/ TensoRF [3])95.6795.3693.1393.2995.2595.4090.9186.9984.12
+ +![](images/e61dfb0f4c098e3e483b9b712c56131d0864efa55ba8f044b8af6795639d45ed.jpg) +CopyRNeRF + +![](images/01c442fd356a006cb7b3e0aed43f4fe37a88c72f280cb53e6c04084b3e0128e4.jpg) +PSNR=32.69 +Bit Acc=100% + +![](images/468834f85bffd511f7697ab38a26ecf5a7be37538af641cea89a7812cd2aa00b.jpg) +Ours + NeRF + +![](images/765ea5dec9083185de5dfff7d62210510bd684a0e70a4a22b4e16a2d121eeedc.jpg) +PSNR=32.78 +Bit Acc=100% +Figure 4. Qualitative comparisons We show the differences $(\times 10)$ between the rendered images and the ground truth. Our method achieves higher PSNR and bit accuracy than CopyRNeRF. + +![](images/1cb826a7901b0604f37127643e0dd5e479f92d7950e17e0b6d0224257f514e4d.jpg) +Ours + TensoRF + +![](images/4222fcd486cd2647ad16e37b4fa00b4c175e2c130ac06e6e1ebb6a24fa011108.jpg) +PSNR=33.44 +Bit Acc=100% + +NeRF [35], it yields the highest SSIM and the lowest LPIPS (noting that higher SSIM and lower LPIPS are desirable). Although the SSIM and LPIPS scores for TensoRF [3] using our method are not the highest, they are comparable to those achieved with Vanilla NeRF [35]. In summary, models trained with our method attain superior bit accuracy, PSNR, SSIM, and the lowest LPIPS, indicating a well-balanced trade-off between bit accuracy and reconstruction quality. + +Additionally, to verify the invisibility of our Result, We compute the differences between ground truth and our rendered results. As illustrated in Fig. 4, the results of applying our method to both Vanilla NeRF [35] and TensoRF [3] achieve a better balance between reconstruction quality and bit accuracy compared to CopyRNeRF [30]. This indicates that our method can embed watermarks with minimal impact on the reconstruction quality. + +Robustness We evaluate the robustness of the watermarked image to different attacks by applying Gaussian noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. As shown in Tab. 3, We experiment on both Vanilla NeRF [35] and TensorRF [3] when applied with our method, resulting in robustness against various attacks compared to CopyRNeRF [30]. + +# 5.5. Ablation study + +Patch Loss and Frequency Domain This section delves into the effectiveness of integrating patch loss and utilizing the frequency domain. Fig. 5 presents a comparative analysis of our method, applied to Vanilla NeRF [35], both with + +![](images/6d6aadba5836318a6d847adfcb7e83558daf0db9a9dcb025a5f1b5468e77b4fa.jpg) +Ground Truth + +![](images/5160e12230c71f7ecad8f4b8d4a77b9e1cc72187ee0c5bb95e141b7da0aeb2bb.jpg) +Full method +Bit Acc=94.78% + +![](images/dce74852b610b2e6636b96f71ef60c2bb01730e10e7ba33b82b0cf0e016ef264.jpg) +Without DWT With Patch +Bit Acc=82.47% + +![](images/7de8cbc396a242e86a49210cdc1eabafd79296cd0a7eacd762d01076e18cddff.jpg) +With DWT without patch +Bit Acc=90.78% + +![](images/7362d29fe760912c1b7b3b34a502e0933bbbdc76200e0098a3db621cdb5c9c83.jpg) +Figure 5. Reconstruction quality comparisons We evaluate our full method, our method without patch loss and our method without frequency domain about 16 bits with NeRF [35]. + +![](images/cc70d8cfb9cb8f6189a6d33a1558abe19cce24f44a94e5e86226f74c62db673b.jpg) +Bit Acc=100.0% + +![](images/576c25827d98305969518040202cc79c8c932af5f7ed45247d7df19718304628.jpg) +Bit Acc=43.54% + +![](images/518b753a6a76183bbea86f73a8bd333faea75d132328211f34ca7e5edd16f83e.jpg) +Bit Acc=100.0% + +Table 3. Robustness under diverse attacks compared with the state-of-the-art method. The bit accuracy results are the average of LLFF and Blender datasets. We show the results on $M_{L} = 16$ bits. The best performances are highlighted in **bold**. + +
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
DFT71.0619.040.8220.1719
DCT43.7535.330.9670.0205
No frequency72.6832.800.9440.0392
DWT-Level 191.1633.040.9470.0332
DWT-Level 295.6732.790.9480.0336
DWT-Level 390.9632.450.9500.0331
DWT-Level 485.0231.820.9520.0329
+ +Table 4. Bit accuracies and reconstruction qualities compared with spatial and frequency domains (DFT, DCT) and DWT levels 1, 2, 3, and 4. We focus on low frequency as our assumption. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold. + +and without the incorporation of patch loss and use in the frequency domain. As shown in Fig. 5, the absence of DWT significantly decreases bit accuracy. Without the patch-wise loss, there is a decrease in reconstruction quality. + +DWT Level During our experiments with DWT, we evaluate performance across commonly used levels: 1, 2, 3, and 4. As shown in Tab. 4, it is indicated that the results from levels 1, 3, and 4 do not perform as well as the outcomes achieved at level 2, both in terms of bit accuracy and reconstruction quality balance. Additionally, we conduct applying DFT, DCT, and no frequency instead of DWT. As a result, our findings show that the DWT at level 2 is the most effective. Thus, we use the DWT at level 2 based on the results from the ablation experiments. + +![](images/ce5bb067fd97030d896e26362147ec39387415e31a84bc170040ccc846a2d90b.jpg) +Figure 6. Bit accuracy comparison between DWT Subbands We evaluate the bit accuracy on level 1,2,3 and all subbands. We show the results on $M_L = 16$ bits. + +DWT Subband The choice of DWT subband is as critical as the level of DWT itself. Typically, subbands such as LH, HL, and HH, which are composed of high-frequency components less perceptible to the people, are employed for watermark embedding in conventional watermarking techniques. Conversely, in our experimental approach, we opt for the LL subband representing the image's coarse approximation and containing essential low-frequency details. This decision is based on the watermark decoder's pre-training process to extract watermarks specifically from the image's spatial domain, which is visually similar to the LL subband and its demonstrated robustness to JPEG compression. In Fig. 6, our empirical investigations confirm that the LL subband yields the highest bit accuracy. + +Comparison of Training time. Even though our proposed method necessitates the use of a pre-trained model's weights as a starting point for fine-tuning NeRF, the time requirement for training is significantly less compared to that of CopyRNeRF [30]. As depicted in Fig. 7, it is evident that the model utilizing our techniques substantially outperforms CopyRNeRF [30] regarding training efficiency. Our method, when applied to an explicit NeRF architecture, such as TensoRF [3], achieves a remarkable speed increase, performing up to six times faster than CopyRNeRF [30]. Similarly, implementing our techniques within the Implicit NeRF framework, like Vanilla NeRF [35], yields a significant speed improvement, with our model operating six times faster than CopyRNeRF [30] but it also surpasses the bit accuracy of the compared model. + +# 6. Conclusion + +We propose a neural 3D watermarking method for the NeRF model. Our method trains a 2D watermark decoder and NeRF model separately. Therefore, our pipeline only requires training the decoder once and re-use it on differ + +![](images/c655b89c58c463fd005f137661a1bc1732dc864c4a02f96728f5fa712b1859d9.jpg) +Figure 7. Analysis of train time for reconstruction from blender dataset at $160 \times 160$ resolution. Note the introduction of a scale break within the graph to accommodate the significant disparities in training durations, with time represented in minutes on the left axis and in days on the right axis. + +ent NeRF watermark models. We adopt conventional watermark techniques in image watermarking, which transfer the image from the spatial domain to the frequency domain to encode the watermark into the image efficiently. We find that DWT and patch-wise loss enhance the overall reconstruction quality while performing high message bit accuracy. Through extensive experiments, we demonstrate that our method outperforms the state-of-the-art method and is six times faster than the previous work. + +Limitations. Our method shows outstanding performance in the digital watermark of NeRF, but training a decoder that extracts watermark from rendered images is time-consuming, taking approximately 12 hours on a single RTX 3090. Our method only encodes a unique message per model, requiring fine-tuning for message insertion every time. We observe that the bit accuracy drops when we apply numerous messages (see Fig. 3). Although previous studies, including us, report robustness against various distortions of images, there is no consideration about cases where the model is attacked. Future work may explore watermarks resilient to cases of model attacks. + +Acknowledgments. This research was supported by the Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture(Project : 4D Content Generation and Copyright Protection with Artificial Intelligence, R2022020068, $80\%$ ), the Korea Research Institute for Defense Technology Planning and Advancement (KRIT) Grant funded by the Korean Government(Grant KRIT-CT-23-021, $10\%$ ), the National Research Foundation of Korea grant (NRF2022R1F1A1074334, $9\%$ ), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), $1\%$ ). + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 1 +[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 2 +[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 1, 2, 3, 5, 6, 7, 8 +[4] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 5 +[5] Chinmayee Das, Swetalina Panigrahi, Vijay K Sharma, and KK Mahapatra. A novel blind robust image watermarking in dct domain using inter-block coefficient correlation. AEU-International Journal of Electronics and Communications, 68(3):244–253, 2014. 3 +[6] Xin Deng, Chao Gao, and Mai Xu. Pirnet: Privacy-preserving image restoration network via wavelet lifting. In ICCV, pages 22368-22377, 2023. 3 +[7] Elhameh Farri and Peyman Ayubi. A robust digital video watermarking based on ct-svd domain and chaotic dna sequences for copyright protection. Journal of Ambient Intelligence and Humanized Computing, 14(10):13113-13137, 2023. 2 +[8] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. ICCV, 2023. 2, 4 +[9] Felipe ABS Ferreira and Juliano B Lima. A robust 3d point cloud watermarking method based on the graph fourier transform. Multimedia Tools and Applications, 79(3):1921-1950, 2020. 2 +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 1 +[11] Youcai Gao, Jinwei Wang, and Yun-Qing Shi. Dynamic multi-watermarking and detecting in dwt domain. Journal of Real-Time Image Processing, 16:565-576, 2019. 3 +[12] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4 +[13] A Hernández-Joaquín, G Melendez-Melendez, and R Cumplido. A secure dwt-based dual watermarking scheme for image authentication and copyright protection. Multimedia Tools and Applications, 82(27):42739-42761, 2023. 4 +[14] Ningombam Jimson and K Hemachandran. Dft based coefficient exchange digital image watermarking. In 2018 Second International Conference on Intelligent Computing and Control Systems (ICICCS), pages 567-571. IEEE, 2018. 3 +[15] Priyank Khare and Vinay Kumar Srivastava. A reliable and secure image watermarking algorithm using homomorphic transform in dwt domain. Multidimensional Systems and Signal Processing, 32:131-160, 2021. 4 + +[16] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 5 +[17] Hung-Jui Ko, Cheng-Ta Huang, Gwoboa Horng, and WANG Shiuh-Jeng. Robust and blind image watermarking in dct domain using inter-block coefficient correlation. Information Sciences, 517:128-147, 2020. 3 +[18] Sanjay Kumar and Binod Kumar Singh. Dwt based color image watermarking using maximum entropy. Multimedia Tools and Applications, 80:15487-15510, 2021. 4 +[19] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2 +[20] Guezouli Larbi, Hachani Roumaissa, and Herri Sara. Embedding watermark in the magnitude matrix of the dft of image. In Proceedings of the 2018 International Conference on Computing and Pattern Recognition, pages 106-110, 2018. 3 +[21] Yong-Seok Lee, Seo Young-Ho, and Kim Dong-Wook. Blind image watermarking based on adaptive data spreading in n-level dwt subbands. Security and communication Networks, 2019, 2019. 3 +[22] Chenxin Li, Brandon Y Feng, Zhiwen Fan, Panwang Pan, and Zhangyang Wang. Steganerf: Embedding invisible information within neural radiance fields. In ICCV, pages 441-453, 2023. 1, 2, 4 +[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2 +[24] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, pages 4273-4284, 2023. 2 +[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 2 +[26] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2 +[27] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR, pages 13548-13557, 2020. 2 +[28] Xiyang Luo, Michael Goebel, Elnaz Barshan, and Feng Yang. Leca: A learned approach for efficient cover-agnostic watermarking. In Media Watermarking, Security, and Forensics, 2022. 2 +[29] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing, 2023. 2 +[30] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural + +radiance fields. In ICCV, pages 22401-22411, 2023. 1, 2, 4, 6, 7, 8 +[31] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 2 +[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2 +[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 2 +[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 5 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 5, 6, 7, 8 +[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[37] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. ACM Trans. Graph., 41(4), 2022. 1 +[38] Shabir A Parah, Javaid A Sheikh, Nazir A Loan, and Ghulam M Bhat. Robust and blind watermarking technique in dct domain using inter-block coefficient differencing. Digital Signal Processing, 53:11-24, 2016. 3 +[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2 +[40] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 2 +[41] SN Prajwalasimha, S Chethan Suputhra, and C Mohan. Performance analysis of combined discrete fourier transformation (dft) and successive division based image watermarking scheme. International Journal of Recent Technology and Engineering, 8(3):34-39, 2019. 3 +[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318-10327, 2021. 2 +[43] Deepti Shukla, Nirupama Tiwari, and Deepika Dubey. Survey on digital watermarking techniques. International Journal of Signal Processing, Image Processing and Pattern Recognition, 9(1):239-244, 2016. 2 + +[44] S Suthaharan and S Sathananthan. Transform domain technique: robust watermarking for digital images. In Proceedings of the IEEE SoutheastCon 2000. Preparing for The New Millennium' (Cat. No. 00CH37105), pages 407-412. IEEE, 2000. 2 +[45] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[46] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2 +[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2 +[48] Jinwei Wang, Shiguo Lian, and Yun-Qing Shi. Hybrid multiplicative multi-watermarking in dwt domain. Multidimensional Systems and Signal Processing, 28:617-636, 2017. 3 +[49] Yinghui Wang, Jing Liu, Yajie Yang, Douli Ma, and Ruijiao Liu. 3d model watermarking algorithm robust to geometric attacks. IET Image Processing, 11(10):822-832, 2017. 2 +[50] Muyu Xu, Fangneng Zhan, Jiahui Zhang, Yingchen Yu, Xiaojin Zhang, Christian Theobalt, Ling Shao, and Shijian Lu. Wavenerf: Wavelet-based generalizable neural radiance fields. In ICCV, pages 18195-18204, 2023. 3 +[51] Innfarn Yoo, Huiwen Chang, Xiyang Luo, Ondrej Stava, Ce Liu, Peyman Milanfar, and Feng Yang. Deep 3d-to-2d watermarking: embedding messages in 3d meshes and extracting them from 2d renderings. In CVPR, pages 10031–10040, 2022. 2 +[52] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2 +[53] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2 +[54] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In CVPR, pages 12397-12406, 2023. 2 +[55] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, pages 717-733. Springer, 2022. 5 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6 +[57] Qiumei Zheng, Nan Liu, Baoqin Cao, Fenghua Wang, and Yanan Yang. Zero-watermarking algorithm in transform domain based on rgb channel and voting strategy. Journal of Information Processing Systems, 16(6):1391-1406, 2020. 2 +[58] Haoyi Zhu. X-nerf: Explicit neural radiance field for multiscene 360deg insufficient rgb-d views. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5766–5775, 2023. 1 + +[59] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, pages 657–672, 2018. 2, 3, 6 \ No newline at end of file diff --git a/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/images.zip b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4fa8ef781483400c5032f50880f64f618f14846e --- /dev/null +++ b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b37749506a110f13d7934a500fb238c05795bbbf4602863f4d2ca65464f5cab6 +size 456161 diff --git a/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/layout.json b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25e73ce340f6bd0ef10bc7ea4c41ecaf51090c7e --- /dev/null +++ b/2024/WateRF_ Robust Watermarks in Radiance Fields for Protection of Copyrights/layout.json @@ -0,0 +1,9969 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 57, + 103, + 537, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 103, + 537, + 122 + ], + "spans": [ + { + "bbox": [ + 57, + 103, + 537, + 122 + ], + "type": "text", + "content": "WaterRF: Robust Watermarks in Radiance Fields for Protection of Copyrights" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": "Youngdong Jang" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": " Dong In Lee" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": " MinHyuk Jang" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": " Jong Wook Kim" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": " Feng Yang" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": " Sangpil Kim" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": "Korea University " + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 474, + 194 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 192, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 242, + 290, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 290, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 290, + 506 + ], + "type": "text", + "content": "The advances in the Neural Radiance Fields (NeRF) research offer extensive applications in diverse domains, but protecting their copyrights has not yet been researched in depth. Recently, NeRF watermarking has been considered one of the pivotal solutions for safely deploying NeRF-based 3D representations. However, existing methods are designed to apply only to implicit or explicit NeRF representations. In this work, we introduce an innovative watermarking method that can be employed in both representations of NeRF. This is achieved by fine-tuning NeRF to embed binary messages in the rendering process. In detail, we propose utilizing the discrete wavelet transform in the NeRF space for watermarking. Furthermore, we adopt a deferred back-propagation technique and introduce a combination with the patch-wise loss to improve rendering quality and bit accuracy with minimum trade-offs. We evaluate our method in three different aspects: capacity, invisibility, and robustness of the embedded watermarks in the 2D-rendered images. Our method achieves state-of-the-art performance with faster training speed over the compared state-of-the-art methods. Project page: https://kuai-lab.github.io/cvpr2024waterf/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 538, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 610 + ], + "type": "text", + "content": "Digital watermarking plays a pivotal role in reinforcing the copyright of digital assets, e.g., text, image, audio, video, and 3D content. Digital assets are easily misused without permission from the creators and owners of the digital assets. One way to protect the copyright is by encoding invisible watermarks in the digital assets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 610, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 287, + 683 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [35] have emerged into the spotlight in 3D content creation and modeling since NeRF can represent 3D objects or scenes in a compact way. With increasing interest in 3D content for the Metaverse, virtual reality, and augmented reality, NeRF has become important in digital watermarking research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 683, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 683, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 59, + 683, + 287, + 695 + ], + "type": "text", + "content": "A straightforward approach for protecting rendered 3D" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 217, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 310, + 217, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 310, + 217, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 310, + 217, + 545, + 345 + ], + "type": "image", + "image_path": "7ca51836b73624f0772d63f0f8ae808cd9047232350cef2002b3168eccfce436.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 354, + 547, + 410 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 410 + ], + "type": "text", + "content": "Figure 1. The NeRF model can be fine-tuned to embed the watermark into the images of the novel view. The figure above shows the model owner Alice and her NeRF model is stolen by Bob. Alice can identify the copyright by using her watermark decoder. Our method embeds the watermark into all of the rendered images." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 420, + 547, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 541 + ], + "type": "text", + "content": "images from the NeRF model is using the post-generation watermarking method that embeds the watermark into the 2D rendered image from NeRF with existing watermarking methods. However, this method only protects copyright for the rendered images but not the copyright of the NeRF model. A fundamental solution to protecting the model and rendered image is fine-tuning the NeRF model to encode the watermark into the NeRF model itself. The watermark will be encoded into the rendered image from the fine-tuned NeRF every time during the rendering process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 543, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 652 + ], + "type": "text", + "content": "NeRF has two primary representations: implicit and explicit representations. The implicit representation [1, 35, 37] uses a multi-layer perceptron (MLP) to represent a 3D scene and the explicit representation [3, 10, 58] uses a traditional 3D structure, such as voxels. Prior methods [22, 30] were limited to applying to only one type of NeRF representation. However, in order to overcome this limitation, we explore a new question: How can we embed the watermark into two distinct NeRF representations(implicit and explicit)?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose an innovative watermarking method that can be applied to both types of NeRF representations. Our method integrates the watermarking process into the rendering process without changing the model architecture. It adjusts the pre-trained NeRF model so that all" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12087" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "of the rendered images have same watermark embedded. We pre-train the decoder with the conventional deep learning method HiDDeN [59] to extract the watermark from the rendered image. Our method has several advantages. First, since it is a fine-tuning process, we can deal with the implicit and the explicit NeRF representations. Second, we do not need any extra processes for concealing watermarks in the rendered images. Therefore, less computational cost is required. Third, our method can protect the model and the rendered images simultaneously." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "type": "text", + "content": "Invisibility and robustness are very important aspects in the digital watermarking domain. Concealing watermarks in the rendered image without degrading the original image quality and ensuring identification of the watermark is challenging. For instance, an embedded watermark in the image is better to be as invisible as possible, and also watermark should be identified after various distortion attacks. We intend to hide the watermark message in the low-level subband in the frequency domain by carefully designing the loss function that evaluates the loss value in the frequency domain. To enhance the image quality, we consider the local structure of the object by introducing patch-wise loss. We subdivided the rendered image and cached the gradient in the grid to encode the watermark locally in the image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 529 + ], + "type": "text", + "content": "Our extensive experiments demonstrate that our method successfully encodes watermarks by fine-tuning the pretrained NeRF models and identifying the watermark message using a simplified decoder from HiDDeN [59]. Moreover, we show that our method is robust from diverse watermark attacks. We evaluate our method by measuring the capacity and robustness under diverse watermark attacks. The training time of our method is approximately six times faster than the CopyRNeRF [30]. Since our method can be applied to both implicit [35] and explicit [3] NeRF, our method can be used in more general applications than the other watermarking methods. Our method outperforms the other state-of-the-art watermarking methods within all metrics. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 529, + 287, + 660 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 529, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 553 + ], + "type": "text", + "content": "- Our method can be applied to implicit and explicit NeRF models, unlike other existing watermarking methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 553, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 287, + 589 + ], + "type": "text", + "content": "- We propose a novel watermarking method for NeRF that fine-tunes the NeRF model by minimizing the loss function which is evaluated in the frequency domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 590, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 624 + ], + "type": "text", + "content": "- We propose a patch-wise loss to improve rendering quality and bit accuracy and enable encoding the watermark locally in the image, reducing the color artifacts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 625, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 287, + 660 + ], + "type": "text", + "content": "- The proposed watermarking method achieves state-of-the-art performance, and we show that our methods are robust in diverse watermark attacks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Neural Radiance Fields. Due to the success of Neural Radiance Field (NeRF) [35], a highly photo-realistic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "view synthesis of complex scenes has been achieved. Recently, NeRF has been used in various research and applications, including faster inference [26, 52, 54], 3D reconstruction [32, 39, 53], image processing [19, 31, 47], dynamic scenes [24, 42] and generative models [25, 33, 40]. Therefore, NeRF become the dominant 3D representation and is widely used. Several commercial products [36, 45] utilized NeRF 3D representations. Thus, the management of the copyright of NeRF-based 3D representations has been emerging as a crucial aspect. As all of those applications are based on implicit, explicit, or both representations, we explore a versatile watermarking method that can be trained on both representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 235, + 547, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 547, + 379 + ], + "type": "text", + "content": "Digital watermarking. Watermarking methods have been evolving over the decades [2, 43]. There are two main ways to recover watermarks [23]: multi-bit watermark and zero-bit watermark. The multi-bit watermark allows for the encoding and decoding of multi-bit messages from media. The zero-bit watermark can be used for ownership authentication. In this paper, we focus on the zero-bit watermark, which involves encoding and decoding multi-bit messages. Recently, with the great success of deep learning watermark methods like HiDDeN [59], considerable research has been conducted to embed watermarks into carriers, especially into 2D images [8, 27, 28], video [7, 29] and 3D [9, 49, 51]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 609 + ], + "type": "text", + "content": "On the NeRF field, the watermarks do not survive during the volume rendering. To overcome this issue, several methods [22, 30] have emerged. StegaNeRF [22] designed a steganography model on explicit NeRF that hides natural images in 3D scene representation. CopyRNeRF [30] secured the copyright of images rendered from an implicit 3D model by embedding a watermark into the rendered color representation. However, these methods have two big limitations. First, they could only be applied to one of the implicit or explicit representations, while both representations of NeRF greatly impact the advancement of NeRF. Second, these watermarks were not robust enough. Since StegaNeRF required the original image for message extraction, this method was not robust to distortion-like cropping, which made a lot of difference between the original and rendered images and CopyRNeRF did not resist cropping. To address these issues, we design a robust watermarking NeRF model that not only can be utilized on both implicit and explicit representation but is also robust to diverse attacks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 615, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 547, + 687 + ], + "type": "text", + "content": "Frequency Domain. It is known that embedding a watermark into the spatial domain is vulnerable to attacks such as cropping and compression because it directly modifies the image pixels [46]. For this reason, many studies have been conducted based on frequency domains [44, 57]. The following are the frequency transformation methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "- Discrete Fourier Transform (DFT) expresses images in terms of phase and amplitude. When a watermark is em" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12088" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 108 + ], + "type": "text", + "content": "bedded into the frequency domain converted by DFT, the magnitude is invariant, making it robust against geometric attacks such as rotation and scaling [14, 20, 41]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 286, + 310 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 108, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 286, + 191 + ], + "type": "text", + "content": "- Discrete Cosine Transform (DCT) decomposes the energy of image data into a sum of cosine functions by representing signals in the frequency domain. It is useful for image compression because most of the energy of the image is concentrated in the top-left corner coefficient. Therefore, DCT is extensively used in watermarking methods that energy compaction. [5, 17, 38]" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 192, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 286, + 310 + ], + "type": "text", + "content": "- Discrete Wavelet Transform (DWT) decomposes the image into four subbands, LL(Low-Low), LH(Low-High), HL(High-Low), and HH(High-High). The LL band has the most energy and contains low-frequency information. Additionally, the LL band can be recursively decomposed into n-levels. DWT is used in several watermarking methods [6, 50] and shows significant differences in performance depending on the level [11, 21, 48]. In this paper, we use DWT and the LL subband for fine-tuning NeRF to embed the watermark." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 321, + 261, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 261, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 261, + 335 + ], + "type": "text", + "content": "3. Preliminaries: Representation of NeRF" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "content": "In this paper, we focus on both implicit and explicit representations of a NeRF model. The implicit NeRF [35] represents a scene using a multi-layer perceptrs (MLP) " + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "content": " whose input is 3D location " + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "inline_equation", + "content": "\\pmb{x} = (x,y,z)\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "content": " and 2D viewing direction " + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "inline_equation", + "content": "d\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "content": " and output is color " + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "content": ", volume density " + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 47, + 341, + 286, + 422 + ], + "type": "text", + "content": ". These MLP-based radiance fields can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 421, + 286, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 421, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 135, + 421, + 286, + 430 + ], + "type": "interline_equation", + "content": "\\sigma , c = \\Phi (x, d) \\tag {1}", + "image_path": "c2d9ae2ac352f9bd9bc714cfbed20824c13cf9a30ed4e0d20525863c345bd8ad.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "content": "The explicit NeRF, such as TensoRF [3], models a scene as a 4D tensor, which represents a 3D voxel grid. These grid-based radiance fields can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 468, + 286, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 468, + 286, + 480 + ], + "spans": [ + { + "bbox": [ + 110, + 468, + 286, + 480 + ], + "type": "interline_equation", + "content": "\\sigma , \\boldsymbol {c} = \\mathcal {T} _ {\\sigma} (\\boldsymbol {x}), \\mathcal {S} \\left(\\mathcal {T} _ {c} (\\boldsymbol {x}), \\boldsymbol {d}\\right), \\tag {2}", + "image_path": "52d7231cc5880252398f119462cbe86400b85dec307abe5d38787bb6516fea16.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\sigma} \\in \\mathbb{R}^{I \\times J \\times K}" + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "content": " is a geometry tensor and " + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c^{I \\times J \\times K \\times C}" + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "content": " is an appearance tensor for channel " + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "content": " and the voxel grid resolution " + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "inline_equation", + "content": "I, J, K" + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 482, + 286, + 529 + ], + "type": "text", + "content": " is a decoding function including MLP and spherical harmonics function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 530, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 286, + 554 + ], + "type": "text", + "content": "Both representations are followed by a volumetric rendering equation, which is used to synthesize novel views:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 558, + 286, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 558, + 286, + 615 + ], + "spans": [ + { + "bbox": [ + 86, + 558, + 286, + 615 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} C (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\boldsymbol {r} (t)) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) d t, \\tag {3} \\\\ T (t) = e x p \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (\\boldsymbol {r} (s)) d s\\right), \\\\ \\end{array}", + "image_path": "fc9c5301bf1859b254c739d06e217650bec0315e6698f667909e4b1bae1a4bc1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "inline_equation", + "content": "\\pmb{r}(t)" + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "content": " is a camera ray with near and far bounds " + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "inline_equation", + "content": "C(r)" + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "content": " is an expected color of the ray " + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "inline_equation", + "content": "\\pmb{r}(t)" + }, + { + "bbox": [ + 47, + 617, + 286, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": "Considering the NeRF representation type like the above, how can we embed a watermark in the various types, not just one type? Since these types of NeRF share a common property, rendering images through the same volume rendering function, we use the rendered images to embed watermarks into the models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 361, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 361, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 361, + 83 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "text", + "content": "We propose a process of fine-tuning the NeRF that does not involve alterations to the model's architecture to embed the watermark message. Our method aims to embed the watermark into the weights " + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "text", + "content": " of the NeRF model in the frequency domain of rendered images. Our method stands apart from traditional watermarking methods, which focusing on training encoders and decoders. The difference lies in the fine-tuning process, which embeds the watermark without using an encoder. There are two phases: (1) Pre-training the watermark decoder " + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "text", + "content": ", (2) Fine-tuning the NeRF model " + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 305, + 93, + 545, + 235 + ], + "type": "text", + "content": " to embed the message. Our method is illustrated in Fig. 2 and described in detail below." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 249, + 498, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 498, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 498, + 261 + ], + "type": "text", + "content": "4.1. Pre-training the watermark decoder" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "text", + "content": "We select HiDDeN [59] architecture as our watermark decoder. HiDDeN comprises two convolution networks for data hiding: a watermark encoder " + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "text", + "content": " and a watermark decoder " + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "text", + "content": ". For robustness, it includes a noise layer " + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "text", + "content": ". However, in this training phase, where we focus solely on the decoder's performance, we have excluded an adversarial loss responsible for improving visual quality. After training the HiDDeN model, watermark encoder " + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 269, + 545, + 376 + ], + "type": "text", + "content": " was not utilized in the second phase." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " takes in a cover image " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "I_{o}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " and a binary message " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "M\\in \\{0,1\\} ^L" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": ", as input. Then " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " embeds " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " and produces an encoded image " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "I_w" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": ". In order to make the decoder resistant to various distortions such as rotation and JPEG compression, " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "I_{w}" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " is transformed with a noise layer " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": ". The decoder " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": ", made of several convolution layers, receives " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "I_{w}" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": " as input and extracts a message " + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "inline_equation", + "content": "M^{\\prime}" + }, + { + "bbox": [ + 305, + 377, + 545, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 475, + 545, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 475, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 389, + 475, + 545, + 489 + ], + "type": "interline_equation", + "content": "M ^ {\\prime} = D \\left(N \\left(I _ {w}\\right)\\right) \\tag {4}", + "image_path": "cf0cfff9e19386a59137b6be8f75639f56ab15cdba2fecac208fff53ca2b9e15.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "text", + "content": "We utilize a sigmoid function to set the range of the extracted message " + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "text", + "content": " between [0, 1]. The message loss is calculated with Binary Cross Entropy (BCE) between " + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "inline_equation", + "content": "M_L" + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "text", + "content": " and sigmoid " + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "inline_equation", + "content": "sg(M_L')" + }, + { + "bbox": [ + 305, + 501, + 545, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 555, + 545, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 545, + 603 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m e s s a g e}} = - \\sum_ {i = 1} ^ {L} M _ {i} \\cdot \\log s g \\left(M _ {i} ^ {\\prime}\\right) \\\\ + \\left(1 - M _ {i}\\right) \\cdot \\log \\left(1 - s g \\left(M _ {i} ^ {\\prime}\\right)\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "7ec092116929f417302341f8c9d546e7b869b67926c6baebf5e5b97b8d1c48a1.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": "The decoder is trained to detect watermarks in images that have passed through the trained encoder. However, we do not use the encoder in the second stage. We find that when the decoder receives a vanilla-rendered image, there is a bias between the extracted message bits. Thus, after training the decoder, we conduct PCA whitening to a linear decoder layer to remove the bias without reducing the extraction ability." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12089" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 514, + 138 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 514, + 138 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 514, + 138 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 514, + 138 + ], + "type": "image", + "image_path": "ec1956eca51d7a48ffcfcdabaa3a64a0e8b423bc0f07c678c384b7030dfc946c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 143, + 513, + 318 + ], + "blocks": [ + { + "bbox": [ + 82, + 143, + 513, + 318 + ], + "lines": [ + { + "bbox": [ + 82, + 143, + 513, + 318 + ], + "spans": [ + { + "bbox": [ + 82, + 143, + 513, + 318 + ], + "type": "image", + "image_path": "29f81458ee65570aaea6e4b94bfdfcf97b576b7da73d2c994978ca3f9cb0d1d1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 322, + 545, + 368 + ], + "lines": [ + { + "bbox": [ + 45, + 322, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 45, + 322, + 545, + 368 + ], + "type": "text", + "content": "Figure 2. WaterF overview. Phase 1: We train the encoder and the decoder to extract messages. After phase 1, we do not use the encoder. Phase 2: We fine-tune the NeRF to embed the messages into the rendered images. (a) We disable auto-differentiation and render a full-resolution image to save memory. (b) We use DWT for the rendered images and choose the LL subband as the input of the pre-trained decoder. (c) We enable auto-differentiation and render the images patch by patch. Then the NeRF is optimized using Eq 7 and Eq 8." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 377, + 287, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 389 + ], + "type": "text", + "content": "4.2. Embedding and extracting watermark on DWT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "type": "text", + "content": "Recently, a fine-tuning watermarking method for NeRF [22] in the spatial domain has emerged. Although the fine-tuning method of embedding a message in the spatial domain shows incomparable invisibility and message extraction ability, it is vulnerable to attacks that distort the spatial domain, such as cropping. Directly applying spatial domain techniques from the latent diffusion model [8] does not allow for effective adjustment of NeRF's weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 492, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 563 + ], + "type": "text", + "content": "To tackle these problems, we propose a fine-tuning method in a frequency domain instead of a spatial domain. Various watermarking techniques for images use the frequency domain have seen development and improvement over the years. We find that DWT is the appropriate domain for encoding the message into the weights of a NeRF model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "text", + "content": "The NeRF model renders diverse views of the 3D model given corresponding camera parameters. We transform the pixels of the rendered images, denoted by " + }, + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "X = (x_{c},y_{c})\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "text", + "content": ", into wavelet forms, with " + }, + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "text", + "content": " representing the channel. The DWT is defined as [12]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 292, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 292, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 292, + 714 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} W _ {\\varphi} (j _ {0}, m, n) = \\frac {1}{\\sqrt {M N}} \\sum_ {x _ {c} = 0} ^ {M - 1} \\sum_ {y _ {c} = 0} ^ {N - 1} f (x _ {c}, y _ {c}) \\varphi_ {j _ {0}, m, n} (x _ {c}, y _ {c}), \\\\ W _ {\\psi} ^ {i} (j, m, n) = \\frac {1}{\\sqrt {M N}} \\sum_ {x _ {c} = 0} ^ {M - 1} \\sum_ {y _ {c} = 0} ^ {N - 1} f \\left(x _ {c}, y _ {c}\\right) \\psi_ {j, m, n} ^ {i} \\left(x _ {c}, y _ {c}\\right) \\tag {6} \\\\ \\end{array}", + "image_path": "affc6456fdbf7e715a5a0dff0f7ac9f9ef861cd7f043b6ae5d1c5e3b6cdd5176.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\varphi (x,y)" + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "content": " is a scaling function and " + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\psi (x,y)" + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "content": " is a wavelet function. " + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "inline_equation", + "content": "W_{\\varphi}(j_0,m,n)" + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "content": " is called by an LL subband, which is an approximation of the image at scale " + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "inline_equation", + "content": "W_{\\psi}^{i}" + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "content": " represents LH, HL, HH subbands, where " + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "inline_equation", + "content": "i = \\{\\mathrm{H},\\mathrm{V},\\mathrm{D}\\}" + }, + { + "bbox": [ + 305, + 377, + 545, + 437 + ], + "type": "text", + "content": " and each denotes horizontal, vertical and diagonal coefficients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": "Previous studies [13, 15, 18] have selected " + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "inline_equation", + "content": "LH" + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "inline_equation", + "content": "HL" + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "inline_equation", + "content": "HH" + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": " subbands for embedding watermarks because the " + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "inline_equation", + "content": "LL" + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": " subband contains significant information about the image. However, we choose the LL subband as an input of our decoder " + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": " and get the extract message with " + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "inline_equation", + "content": "M' = D(W_{\\varphi})" + }, + { + "bbox": [ + 305, + 437, + 545, + 532 + ], + "type": "text", + "content": ". We experimentally discover that embedding the watermark in the LL subband is more robust and effective than other subbands for the HiDDeN decoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "text", + "content": "The DWT is characterized by its subbands being computed across different levels; therefore, choosing an optimal level for our purpose is necessary. The 1-level separates the images to 4 subbands " + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "inline_equation", + "content": "(LL_{1}, LH_{1}, HL_{1}, HH_{1})" + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "text", + "content": ", then 2-level separate the " + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "inline_equation", + "content": "LL_{1}" + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "text", + "content": " subband into 4 subbands " + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "inline_equation", + "content": "(LL_{2}, LH_{2}, HL_{2}, HH_{2})" + }, + { + "bbox": [ + 305, + 533, + 545, + 654 + ], + "type": "text", + "content": ". We experimentally discover that selecting a level too high decreases visual quality due to excessive adjustment of crucial visual elements. Thus, we use the 2-level DWT because it best maintains a good balance between bit accuracy and reconstruction quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 659, + 462, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 462, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 462, + 672 + ], + "type": "text", + "content": "4.3. Fine-tuning the NeRF model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "NeRF has two different representations, implicit and explicit representations (see Sec. 3). Previous studies [22, 30] focused on applying their methods to just one form of NeRF" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12090" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "representation. In contrast, we propose a method applicable to both NeRF representations: implicit and explicit. Our approach involves fine-tuning the NeRF to ensure that novel view images include an embedded message." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "type": "text", + "content": "Our method starts with preparing each pre-trained decoder for different lengths of message bits and also preparing an initial NeRF model " + }, + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "type": "inline_equation", + "content": "F_{\\theta_0}" + }, + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "type": "text", + "content": ". Then we establish fixed binary messages " + }, + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "type": "inline_equation", + "content": "M = (m_1, \\dots, m_L) \\in \\{0, 1\\}^L" + }, + { + "bbox": [ + 46, + 120, + 287, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": "Since rendering with NeRF requires a lot of memory, we turn off the auto-differentiation at first. The rendered images " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " at full resolution are rendered by " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "F_{\\theta_0}" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": ". Following this, the full resolution images " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " are transformed into seven groups of wavelet subbands: " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "\\{LL_2, LH_2, HL_2, HH_2, LL_1, LH_1, HL_1, HH_1\\}" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": ". Our pre-trained decoder " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " takes the " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "LL_2" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " subband as an input and extracts a message " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "M' = D(W_{\\varphi}(j_0, m, n))" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " from it. The message loss is calculated by comparing " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "L_m = BCE(M, sg(M'))" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": ". We choose the Waston-VGG [4] to calculate the full image perceptual loss between the image without watermark " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": " and the image with a watermark " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "inline_equation", + "content": "L_i(X_0, X)" + }, + { + "bbox": [ + 46, + 168, + 288, + 335 + ], + "type": "text", + "content": ". Full resolution loss function is a sum of the full image perceptual loss and the message loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 343, + 287, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 343, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 94, + 343, + 287, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f u l l}} = \\lambda_ {i} \\mathcal {L} _ {\\text {p e r c e p t}} + \\lambda_ {m} \\mathcal {L} _ {\\text {m e s s a g e}} \\tag {7}", + "image_path": "b2ac8243a97ca500b102bb232494e71f0f14f492a4fb508bddd37fe2c5a6ecca.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "type": "text", + "content": "We cache the gradient calculated by Eq. 7 prior to the patch-wise rendering process to update the parameters. Then, we turn on the auto-differentiation so that the NeRF parameters enable patch-wise rendering. Our patch loss " + }, + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "type": "inline_equation", + "content": "L_{patch}" + }, + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "type": "text", + "content": " is calculated by combining mean absolute error (MAE) across rendered pixel colors, total variation (TV) regularization, and SSIM loss to maintain the balance between bit-accuracy and visual quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 465, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 465, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 52, + 465, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p a t c h}} = \\lambda_ {M A E} \\mathcal {L} _ {M A E} + \\lambda_ {T V} \\mathcal {L} _ {T V} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} \\tag {8}", + "image_path": "c01be8e0e34d3c691375892ac187154b1583a11fcd56c5113146baf84cdd3421.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "content": "Then we optimize from the initial NeRF model " + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "inline_equation", + "content": "F_{\\theta_0}" + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "content": " to watermarked " + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "content": " using Eq. 7 and Eq. 8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 515, + 270, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 270, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 270, + 529 + ], + "type": "text", + "content": "4.4. Deferred back-propagation with patch loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 629 + ], + "type": "text", + "content": "In conventional NeRF models, rendering at full resolution is inefficient regarding memory usage. To solve the memory-inefficient problem, pixels are often sampled and rendered randomly to compute the L2 loss for training. However, the main limitation of random pixel-wise rendering is in its incompatibility with using CNN-based perceptual loss and our watermark decoder for extracting message bits, both of which require a full-resolution image as input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Our method implements a full-resolution rendering derived from our pre-trained decoder during the fine-tuning process. Each iteration presents a considerable challenge due to the extensive memory consumption. In order to relieve the memory constraints, we adopt the deferred back-propagation [55], specifically designed for memory-efficient, patch-wise optimization within NeRF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "Our method proposes a novel patch loss to be applied to the deferred back-propagation as detailed in Eq. 8. When the model renders the full-resolution image with autodifferentiation disabled, the image and message loss are computed by the rendered image, and the derived gradients are cached (see Sec. 4.3 for more details). Then, the model renders the full-resolution image with image patches that are re-rendered with auto-differentiation enabled and back-propagated using the corresponding gradients from the cached set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "content": "Our innovation is not solely relying on cached gradients for back-propagation during re-rendering. Instead, we calculate the loss between the re-rendered patch and the ground truth (GT) patch and carry out an additional backpropagation process. The patch loss consists of a sum of L1 loss (MAE), a Total Variation loss, and an SSIM loss, as defined in Eq. 8. By utilizing back-propagation, once for the combined loss from the full-resolution image and the message loss and once for the patch loss, we are able to enhance both image quality and message bit accuracy at once." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 324, + 387, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 387, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 387, + 337 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 343, + 364, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 343, + 364, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 364, + 355 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 362, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 470 + ], + "type": "text", + "content": "We evaluate our method using datasets commonly used by NeRF: the Blender dataset [35], which consists of synthetic bounded scenes, and the LLFF dataset [34], which consists of real unbounded, forward-facing scenes. We follow the conventional literature in NeRF papers, which evaluate and compare using eight scenes each from the Blender dataset (including chair, drums, ficus, hotdog, lego, materials, mic) and the LLFF dataset (including fern, flower, fortress, horns, leaves, orchids, room, trex)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 479, + 440, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 440, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 440, + 491 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": "Vanilla NeRF [35] and TensoRF [3] employ similar training methods, with some specifics as detailed below. The training process is carried out on an image-by-image basis, with a batch size set to one. Ray tracing is implemented along with the camera pose, and the Adam optimizer [16] is used for the learning process. The training is completed with epochs ranging from 5 to 10. Since our pretrained decoder requires full-resolution images, NeRF renders full-resolution images during the fine-tuning process. However, NeRF requires a great amount of memory usage. We adopt the deferred back-propagation in ARF [55] to solve the memory issue and combine it with the patchwise loss to improve rendering quality and bit accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "The initial " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " was set to 0.1 and 0.9 for the LLFF dataset and 0.05 and 0.95 for the Blender dataset, respectively. Additionally, within the patch loss equation (Eq. 8), the parameters " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{MAE}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{TV}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{SSIM}" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " were set to 0.1, 0.06, and 0.02." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12091" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 281, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 281, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 281, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 281, + 145 + ], + "type": "table", + "html": "
Bit Accuracy(%) ↑
4 bits8 bits16 bits32 bits48 bits
HiDDeN [59]+NeRF [35]50.3150.2550.1950.1150.04
CopyRNeRF [30]100.00100.0091.1678.0860.06
Ours (w/ NeRF [35])100.00100.0094.2486.8170.43
Ours (w/ TensoRF [3])100.00100.0095.6788.5885.82
", + "image_path": "c66902d6b4cca992d8ded814a2ccbe937ef378d68e7ea1ebc1b372de13a7192b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 206, + 120, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 120, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 120, + 218 + ], + "type": "text", + "content": "5.3. Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 225, + 288, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 288, + 416 + ], + "type": "text", + "content": "We evaluate our method with three key essential aspects in watermarking techniques: 1) capacity, measuring the length of the messages that can be embedded into the image; 2) invisibility, which involves evaluating the difficulty in detecting embedded watermarks in images by peoples; 3) robustness, evaluating the robustness of our watermarking method under various image distortions such as Gaussian Noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. We employ Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [56] for the evaluation metrics to assess invisibility while utilizing bit accuracy to evaluate capacity and robustness. We conduct experiments on the invisibility and robustness of the message length of " + }, + { + "bbox": [ + 46, + 225, + 288, + 416 + ], + "type": "inline_equation", + "content": "M_L \\in \\{4,8,16,32,48\\}" + }, + { + "bbox": [ + 46, + 225, + 288, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 543 + ], + "type": "text", + "content": "Baseline To the best of our knowledge, CopyRNeRF [30] stands out alone in its ability to embed bits with the implicit NeRF model, which can be converted into various modality data such as messages or images. The CopyRNeRF performs state-of-the-art on bit-accuracy and visual quality. As a result, we conduct our evaluations utilizing the CopyRNeRF. We select the following relevant comparative models:: HiDDeN [59] " + }, + { + "bbox": [ + 46, + 423, + 287, + 543 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 423, + 287, + 543 + ], + "type": "text", + "content": " NeRF [35], which employ the classical 2D watermarking method HiDDeN [59] on training images prior to training the NeRF model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": "To demonstrate the applicability of our method, we apply our method on two distinct NeRF models: Vanilla NeRF [35], symbolizing implicit representation NeRF, and TensoRF [3], representing explicit representation NeRF based on voxel grids. We select the Vanilla NeRF [35] since it is the first NeRF model to be introduced and can be considered the foundational and ancestral model for all subsequent NeRFs. TensoRF [3] is chosen because it is an influential NeRF model with significantly improved performance and training speed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 167, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 167, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 167, + 684 + ], + "type": "text", + "content": "5.4. Experimental results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Capacity As shown in Tab. 1, there is a clear trade-off between bit accuracy and message bit length. We conduct bit" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 305, + 71, + 552, + 143 + ], + "blocks": [ + { + "bbox": [ + 46, + 149, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 194 + ], + "type": "text", + "content": "Table 1. Bit accuracies with different message lengths are compared with the state-of-the-art method. (HiDDeN [59] + NeRF [35] is pre-embedded before training). The results are the average of LLFF and Blender datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 71, + 552, + 143 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 552, + 143 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 552, + 143 + ], + "type": "table", + "html": "
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
HiDDeN [59]+NeRF [35]50.1926.530.9170.035
CopyRNeRF [30]91.1626.290.9100.038
Ours (w/ NeRF [35])94.2428.810.9540.025
Ours (w/ TensoRF [3])95.6732.790.9480.033
", + "image_path": "3b58ed33950e7d809bee54d9b15ab2699d8fed367606b06397e6e1dd7699c21e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "type": "text", + "content": "Table 2. Bit accuracies and reconstruction qualities comparison with baselines. " + }, + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "type": "text", + "content": " means higher (lower) is better. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "text", + "content": "accuracy tests on messages of lengths 4, 8, 16, 32, and 48 bits. Models implemented with our method show a gradual decrease in bit accuracy as the length of the message increases. However, when compared to CopyRNeRF [30], the decline in bit accuracy is less dramatic and less steep. Notably, CopyRNeRF [30] shows a modest " + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "text", + "content": " bit accuracy for 48 bits, whereas Vanilla NeRF [35] and TensoRF [3] using our method demonstrates a respectable " + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "inline_equation", + "content": "85.82\\%" + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "inline_equation", + "content": "70.43\\%" + }, + { + "bbox": [ + 304, + 206, + 546, + 422 + ], + "type": "text", + "content": " accuracy, respectively. Additionally, to demonstrate the stability of our method across a range of messages, we fine-tune 500 different messages, each corresponding to unique, randomly generated bits. We rigorously evaluate the bit accuracy for all images rendered by Vanilla NeRF [35] and TensoRF [3]. As illustrated in Fig. 3, the plot shows a minor descending trend that corresponds with the increasing number of unique messages, and it shows that our method is robust with using many different messages, which is a key value to identify unique NeRF model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "type": "text", + "content": "Invisibility Our method demonstrates a trade-off between bit accuracy and the quality of reconstructed data, as detailed in Tab. 2. This trade-off is analyzed using several metrics, including bit accuracy and visual quality metrics. With TensoRF [3], our method achieves the best bit accuracy and PSNR results. Conversely, when applied to Vanilla" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 523, + 524, + 673 + ], + "blocks": [ + { + "bbox": [ + 326, + 523, + 524, + 673 + ], + "lines": [ + { + "bbox": [ + 326, + 523, + 524, + 673 + ], + "spans": [ + { + "bbox": [ + 326, + 523, + 524, + 673 + ], + "type": "image", + "image_path": "c593b16d8e0edacbafacebd5ef85a62007188989298e4a486bbd0ce2636a1037.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 676, + 545, + 719 + ], + "lines": [ + { + "bbox": [ + 304, + 676, + 545, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 545, + 719 + ], + "type": "text", + "content": "Figure 3. Identification results. We average the bit accuracy for each 500 messages. Our method not only works efficiently with just one key but also works well with arbitrarily generated messages. We show the results on " + }, + { + "bbox": [ + 304, + 676, + 545, + 719 + ], + "type": "inline_equation", + "content": "M_L = 16" + }, + { + "bbox": [ + 304, + 676, + 545, + 719 + ], + "type": "text", + "content": " bits." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12092" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 545, + 135 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 545, + 135 + ], + "type": "table", + "html": "
No DistortionGaussian Noise (v = 0.1)Rotation (±π/6)Scaling (25%)Bit Accuracy(%) ↑
Gaussian Blur (deviation = 0.1)Crop (40%)Brightness (2.0)JPEG Compression (10% quality)Combined (Crop, Brightness, JPEG)
CopyRNeRF [30]91.1690.0488.1389.3390.06----
Ours (w/ NeRF [35])94.2494.0685.0291.3594.1283.4884.1486.8873.64
Ours (w/ TensoRF [3])95.6795.3693.1393.2995.2595.4090.9186.9984.12
", + "image_path": "cbacb24a57e965f27c215c9b826172568ba40470738e04e4116755f31ac328eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 182, + 123, + 226 + ], + "blocks": [ + { + "bbox": [ + 71, + 171, + 119, + 182 + ], + "lines": [ + { + "bbox": [ + 71, + 171, + 119, + 182 + ], + "spans": [ + { + "bbox": [ + 71, + 171, + 119, + 182 + ], + "type": "text", + "content": "CopyRNeRF" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 182, + 123, + 226 + ], + "lines": [ + { + "bbox": [ + 69, + 182, + 123, + 226 + ], + "spans": [ + { + "bbox": [ + 69, + 182, + 123, + 226 + ], + "type": "image", + "image_path": "e61dfb0f4c098e3e483b9b712c56131d0864efa55ba8f044b8af6795639d45ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 67, + 228, + 123, + 281 + ], + "blocks": [ + { + "bbox": [ + 67, + 228, + 123, + 281 + ], + "lines": [ + { + "bbox": [ + 67, + 228, + 123, + 281 + ], + "spans": [ + { + "bbox": [ + 67, + 228, + 123, + 281 + ], + "type": "image", + "image_path": "01c442fd356a006cb7b3e0aed43f4fe37a88c72f280cb53e6c04084b3e0128e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 283, + 116, + 290 + ], + "lines": [ + { + "bbox": [ + 75, + 283, + 116, + 290 + ], + "spans": [ + { + "bbox": [ + 75, + 283, + 116, + 290 + ], + "type": "text", + "content": "PSNR=32.69" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 291, + 117, + 298 + ], + "lines": [ + { + "bbox": [ + 75, + 291, + 117, + 298 + ], + "spans": [ + { + "bbox": [ + 75, + 291, + 117, + 298 + ], + "type": "text", + "content": "Bit Acc=100%" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 182, + 192, + 226 + ], + "blocks": [ + { + "bbox": [ + 140, + 171, + 190, + 182 + ], + "lines": [ + { + "bbox": [ + 140, + 171, + 190, + 182 + ], + "spans": [ + { + "bbox": [ + 140, + 171, + 190, + 182 + ], + "type": "text", + "content": "Ours + NeRF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 182, + 192, + 226 + ], + "lines": [ + { + "bbox": [ + 138, + 182, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 192, + 226 + ], + "type": "image", + "image_path": "468834f85bffd511f7697ab38a26ecf5a7be37538af641cea89a7812cd2aa00b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 227, + 193, + 282 + ], + "blocks": [ + { + "bbox": [ + 137, + 227, + 193, + 282 + ], + "lines": [ + { + "bbox": [ + 137, + 227, + 193, + 282 + ], + "spans": [ + { + "bbox": [ + 137, + 227, + 193, + 282 + ], + "type": "image", + "image_path": "765ea5dec9083185de5dfff7d62210510bd684a0e70a4a22b4e16a2d121eeedc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 283, + 185, + 290 + ], + "lines": [ + { + "bbox": [ + 144, + 283, + 185, + 290 + ], + "spans": [ + { + "bbox": [ + 144, + 283, + 185, + 290 + ], + "type": "text", + "content": "PSNR=32.78" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 291, + 187, + 298 + ], + "lines": [ + { + "bbox": [ + 144, + 291, + 187, + 298 + ], + "spans": [ + { + "bbox": [ + 144, + 291, + 187, + 298 + ], + "type": "text", + "content": "Bit Acc=100%" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 304, + 287, + 337 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 337 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons We show the differences " + }, + { + "bbox": [ + 46, + 304, + 287, + 337 + ], + "type": "inline_equation", + "content": "(\\times 10)" + }, + { + "bbox": [ + 46, + 304, + 287, + 337 + ], + "type": "text", + "content": " between the rendered images and the ground truth. Our method achieves higher PSNR and bit accuracy than CopyRNeRF." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 205, + 182, + 258, + 226 + ], + "blocks": [ + { + "bbox": [ + 202, + 171, + 262, + 182 + ], + "lines": [ + { + "bbox": [ + 202, + 171, + 262, + 182 + ], + "spans": [ + { + "bbox": [ + 202, + 171, + 262, + 182 + ], + "type": "text", + "content": "Ours + TensoRF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 182, + 258, + 226 + ], + "lines": [ + { + "bbox": [ + 205, + 182, + 258, + 226 + ], + "spans": [ + { + "bbox": [ + 205, + 182, + 258, + 226 + ], + "type": "image", + "image_path": "1cb826a7901b0604f37127643e0dd5e479f92d7950e17e0b6d0224257f514e4d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 205, + 227, + 261, + 282 + ], + "blocks": [ + { + "bbox": [ + 205, + 227, + 261, + 282 + ], + "lines": [ + { + "bbox": [ + 205, + 227, + 261, + 282 + ], + "spans": [ + { + "bbox": [ + 205, + 227, + 261, + 282 + ], + "type": "image", + "image_path": "4222fcd486cd2647ad16e37b4fa00b4c175e2c130ac06e6e1ebb6a24fa011108.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 283, + 252, + 290 + ], + "lines": [ + { + "bbox": [ + 212, + 283, + 252, + 290 + ], + "spans": [ + { + "bbox": [ + 212, + 283, + 252, + 290 + ], + "type": "text", + "content": "PSNR=33.44" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 291, + 255, + 298 + ], + "lines": [ + { + "bbox": [ + 212, + 291, + 255, + 298 + ], + "spans": [ + { + "bbox": [ + 212, + 291, + 255, + 298 + ], + "type": "text", + "content": "Bit Acc=100%" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 350, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 447 + ], + "type": "text", + "content": "NeRF [35], it yields the highest SSIM and the lowest LPIPS (noting that higher SSIM and lower LPIPS are desirable). Although the SSIM and LPIPS scores for TensoRF [3] using our method are not the highest, they are comparable to those achieved with Vanilla NeRF [35]. In summary, models trained with our method attain superior bit accuracy, PSNR, SSIM, and the lowest LPIPS, indicating a well-balanced trade-off between bit accuracy and reconstruction quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 448, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 544 + ], + "type": "text", + "content": "Additionally, to verify the invisibility of our Result, We compute the differences between ground truth and our rendered results. As illustrated in Fig. 4, the results of applying our method to both Vanilla NeRF [35] and TensoRF [3] achieve a better balance between reconstruction quality and bit accuracy compared to CopyRNeRF [30]. This indicates that our method can embed watermarks with minimal impact on the reconstruction quality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 551, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 635 + ], + "type": "text", + "content": "Robustness We evaluate the robustness of the watermarked image to different attacks by applying Gaussian noise, Rotation, Scaling, Gaussian Blur, Crop, Brightness, JPEG compression and combination of these distortions. As shown in Tab. 3, We experiment on both Vanilla NeRF [35] and TensorRF [3] when applied with our method, resulting in robustness against various attacks compared to CopyRNeRF [30]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 646, + 138, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 138, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 138, + 658 + ], + "type": "text", + "content": "5.5. Ablation study" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Patch Loss and Frequency Domain This section delves into the effectiveness of integrating patch loss and utilizing the frequency domain. Fig. 5 presents a comparative analysis of our method, applied to Vanilla NeRF [35], both with" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 310, + 194, + 364, + 229 + ], + "blocks": [ + { + "bbox": [ + 315, + 178, + 362, + 187 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 362, + 187 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 362, + 187 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 194, + 364, + 229 + ], + "lines": [ + { + "bbox": [ + 310, + 194, + 364, + 229 + ], + "spans": [ + { + "bbox": [ + 310, + 194, + 364, + 229 + ], + "type": "image", + "image_path": "6d6aadba5836318a6d847adfcb7e83558daf0db9a9dcb025a5f1b5468e77b4fa.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 370, + 196, + 422, + 228 + ], + "blocks": [ + { + "bbox": [ + 376, + 178, + 418, + 186 + ], + "lines": [ + { + "bbox": [ + 376, + 178, + 418, + 186 + ], + "spans": [ + { + "bbox": [ + 376, + 178, + 418, + 186 + ], + "type": "text", + "content": "Full method" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 196, + 422, + 228 + ], + "lines": [ + { + "bbox": [ + 370, + 196, + 422, + 228 + ], + "spans": [ + { + "bbox": [ + 370, + 196, + 422, + 228 + ], + "type": "image", + "image_path": "5160e12230c71f7ecad8f4b8d4a77b9e1cc72187ee0c5bb95e141b7da0aeb2bb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 229, + 423, + 237 + ], + "lines": [ + { + "bbox": [ + 370, + 229, + 423, + 237 + ], + "spans": [ + { + "bbox": [ + 370, + 229, + 423, + 237 + ], + "type": "text", + "content": "Bit Acc=94.78%" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 428, + 194, + 482, + 228 + ], + "blocks": [ + { + "bbox": [ + 432, + 174, + 479, + 191 + ], + "lines": [ + { + "bbox": [ + 432, + 174, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 432, + 174, + 479, + 191 + ], + "type": "text", + "content": "Without DWT With Patch" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 194, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 428, + 194, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 428, + 194, + 482, + 228 + ], + "type": "image", + "image_path": "dce74852b610b2e6636b96f71ef60c2bb01730e10e7ba33b82b0cf0e016ef264.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 229, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 429, + 229, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 429, + 229, + 482, + 237 + ], + "type": "text", + "content": "Bit Acc=82.47%" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 487, + 194, + 541, + 228 + ], + "blocks": [ + { + "bbox": [ + 496, + 174, + 539, + 191 + ], + "lines": [ + { + "bbox": [ + 496, + 174, + 539, + 191 + ], + "spans": [ + { + "bbox": [ + 496, + 174, + 539, + 191 + ], + "type": "text", + "content": "With DWT without patch" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 194, + 541, + 228 + ], + "lines": [ + { + "bbox": [ + 487, + 194, + 541, + 228 + ], + "spans": [ + { + "bbox": [ + 487, + 194, + 541, + 228 + ], + "type": "image", + "image_path": "7de8cbc396a242e86a49210cdc1eabafd79296cd0a7eacd762d01076e18cddff.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 229, + 541, + 237 + ], + "lines": [ + { + "bbox": [ + 488, + 229, + 541, + 237 + ], + "spans": [ + { + "bbox": [ + 488, + 229, + 541, + 237 + ], + "type": "text", + "content": "Bit Acc=90.78%" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 310, + 243, + 367, + 283 + ], + "blocks": [ + { + "bbox": [ + 310, + 243, + 367, + 283 + ], + "lines": [ + { + "bbox": [ + 310, + 243, + 367, + 283 + ], + "spans": [ + { + "bbox": [ + 310, + 243, + 367, + 283 + ], + "type": "image", + "image_path": "7362d29fe760912c1b7b3b34a502e0933bbbdc76200e0098a3db621cdb5c9c83.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 303, + 546, + 336 + ], + "lines": [ + { + "bbox": [ + 305, + 303, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 546, + 336 + ], + "type": "text", + "content": "Figure 5. Reconstruction quality comparisons We evaluate our full method, our method without patch loss and our method without frequency domain about 16 bits with NeRF [35]." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 370, + 243, + 425, + 283 + ], + "blocks": [ + { + "bbox": [ + 370, + 243, + 425, + 283 + ], + "lines": [ + { + "bbox": [ + 370, + 243, + 425, + 283 + ], + "spans": [ + { + "bbox": [ + 370, + 243, + 425, + 283 + ], + "type": "image", + "image_path": "cc70d8cfb9cb8f6189a6d33a1558abe19cce24f44a94e5e86226f74c62db673b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 284, + 423, + 291 + ], + "lines": [ + { + "bbox": [ + 371, + 284, + 423, + 291 + ], + "spans": [ + { + "bbox": [ + 371, + 284, + 423, + 291 + ], + "type": "text", + "content": "Bit Acc=100.0%" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 428, + 243, + 484, + 283 + ], + "blocks": [ + { + "bbox": [ + 428, + 243, + 484, + 283 + ], + "lines": [ + { + "bbox": [ + 428, + 243, + 484, + 283 + ], + "spans": [ + { + "bbox": [ + 428, + 243, + 484, + 283 + ], + "type": "image", + "image_path": "576c25827d98305969518040202cc79c8c932af5f7ed45247d7df19718304628.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 284, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 429, + 284, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 429, + 284, + 482, + 291 + ], + "type": "text", + "content": "Bit Acc=43.54%" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 487, + 243, + 543, + 283 + ], + "blocks": [ + { + "bbox": [ + 487, + 243, + 543, + 283 + ], + "lines": [ + { + "bbox": [ + 487, + 243, + 543, + 283 + ], + "spans": [ + { + "bbox": [ + 487, + 243, + 543, + 283 + ], + "type": "image", + "image_path": "518b753a6a76183bbea86f73a8bd333faea75d132328211f34ca7e5edd16f83e.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 284, + 541, + 291 + ], + "lines": [ + { + "bbox": [ + 488, + 284, + 541, + 291 + ], + "spans": [ + { + "bbox": [ + 488, + 284, + 541, + 291 + ], + "type": "text", + "content": "Bit Acc=100.0%" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "table", + "bbox": [ + 306, + 350, + 544, + 464 + ], + "blocks": [ + { + "bbox": [ + 47, + 140, + 546, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 546, + 161 + ], + "type": "text", + "content": "Table 3. Robustness under diverse attacks compared with the state-of-the-art method. The bit accuracy results are the average of LLFF and Blender datasets. We show the results on " + }, + { + "bbox": [ + 47, + 140, + 546, + 161 + ], + "type": "inline_equation", + "content": "M_{L} = 16" + }, + { + "bbox": [ + 47, + 140, + 546, + 161 + ], + "type": "text", + "content": " bits. The best performances are highlighted in **bold**." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 350, + 544, + 464 + ], + "lines": [ + { + "bbox": [ + 306, + 350, + 544, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 544, + 464 + ], + "type": "table", + "html": "
Bit Acc(%)↑PSNR ↑SSIM ↑LPIPS ↓
DFT71.0619.040.8220.1719
DCT43.7535.330.9670.0205
No frequency72.6832.800.9440.0392
DWT-Level 191.1633.040.9470.0332
DWT-Level 295.6732.790.9480.0336
DWT-Level 390.9632.450.9500.0331
DWT-Level 485.0231.820.9520.0329
", + "image_path": "28f8c3a4ce88cda36c2243fb427023bd98cd2bf1c4c3e1f52cece2c2fff81757.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "type": "text", + "content": "Table 4. Bit accuracies and reconstruction qualities compared with spatial and frequency domains (DFT, DCT) and DWT levels 1, 2, 3, and 4. We focus on low frequency as our assumption. We show the results on 16 bits. The results are evaluated in the same way as baselines. The best performances are highlighted in bold." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 537, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 537, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 537, + 545, + 586 + ], + "type": "text", + "content": "and without the incorporation of patch loss and use in the frequency domain. As shown in Fig. 5, the absence of DWT significantly decreases bit accuracy. Without the patch-wise loss, there is a decrease in reconstruction quality." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "DWT Level During our experiments with DWT, we evaluate performance across commonly used levels: 1, 2, 3, and 4. As shown in Tab. 4, it is indicated that the results from levels 1, 3, and 4 do not perform as well as the outcomes achieved at level 2, both in terms of bit accuracy and reconstruction quality balance. Additionally, we conduct applying DFT, DCT, and no frequency instead of DWT. As a result, our findings show that the DWT at level 2 is the most effective. Thus, we use the DWT at level 2 based on the results from the ablation experiments." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12093" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 64, + 259, + 223 + ], + "blocks": [ + { + "bbox": [ + 62, + 64, + 259, + 223 + ], + "lines": [ + { + "bbox": [ + 62, + 64, + 259, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 64, + 259, + 223 + ], + "type": "image", + "image_path": "ce5bb067fd97030d896e26362147ec39387415e31a84bc170040ccc846a2d90b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 226, + 287, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 226, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 226, + 287, + 259 + ], + "type": "text", + "content": "Figure 6. Bit accuracy comparison between DWT Subbands We evaluate the bit accuracy on level 1,2,3 and all subbands. We show the results on " + }, + { + "bbox": [ + 48, + 226, + 287, + 259 + ], + "type": "inline_equation", + "content": "M_L = 16" + }, + { + "bbox": [ + 48, + 226, + 287, + 259 + ], + "type": "text", + "content": " bits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 275, + 286, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 443 + ], + "type": "text", + "content": "DWT Subband The choice of DWT subband is as critical as the level of DWT itself. Typically, subbands such as LH, HL, and HH, which are composed of high-frequency components less perceptible to the people, are employed for watermark embedding in conventional watermarking techniques. Conversely, in our experimental approach, we opt for the LL subband representing the image's coarse approximation and containing essential low-frequency details. This decision is based on the watermark decoder's pre-training process to extract watermarks specifically from the image's spatial domain, which is visually similar to the LL subband and its demonstrated robustness to JPEG compression. In Fig. 6, our empirical investigations confirm that the LL subband yields the highest bit accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 451, + 286, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 451, + 286, + 630 + ], + "spans": [ + { + "bbox": [ + 49, + 451, + 286, + 630 + ], + "type": "text", + "content": "Comparison of Training time. Even though our proposed method necessitates the use of a pre-trained model's weights as a starting point for fine-tuning NeRF, the time requirement for training is significantly less compared to that of CopyRNeRF [30]. As depicted in Fig. 7, it is evident that the model utilizing our techniques substantially outperforms CopyRNeRF [30] regarding training efficiency. Our method, when applied to an explicit NeRF architecture, such as TensoRF [3], achieves a remarkable speed increase, performing up to six times faster than CopyRNeRF [30]. Similarly, implementing our techniques within the Implicit NeRF framework, like Vanilla NeRF [35], yields a significant speed improvement, with our model operating six times faster than CopyRNeRF [30] but it also surpasses the bit accuracy of the compared model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 644, + 118, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 644, + 118, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 644, + 118, + 657 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 665, + 286, + 712 + ], + "type": "text", + "content": "We propose a neural 3D watermarking method for the NeRF model. Our method trains a 2D watermark decoder and NeRF model separately. Therefore, our pipeline only requires training the decoder once and re-use it on differ" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 61, + 544, + 203 + ], + "blocks": [ + { + "bbox": [ + 308, + 61, + 544, + 203 + ], + "lines": [ + { + "bbox": [ + 308, + 61, + 544, + 203 + ], + "spans": [ + { + "bbox": [ + 308, + 61, + 544, + 203 + ], + "type": "image", + "image_path": "c655b89c58c463fd005f137661a1bc1732dc864c4a02f96728f5fa712b1859d9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 207, + 544, + 262 + ], + "lines": [ + { + "bbox": [ + 308, + 207, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 544, + 262 + ], + "type": "text", + "content": "Figure 7. Analysis of train time for reconstruction from blender dataset at " + }, + { + "bbox": [ + 308, + 207, + 544, + 262 + ], + "type": "inline_equation", + "content": "160 \\times 160" + }, + { + "bbox": [ + 308, + 207, + 544, + 262 + ], + "type": "text", + "content": " resolution. Note the introduction of a scale break within the graph to accommodate the significant disparities in training durations, with time represented in minutes on the left axis and in days on the right axis." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 274, + 544, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 544, + 380 + ], + "type": "text", + "content": "ent NeRF watermark models. We adopt conventional watermark techniques in image watermarking, which transfer the image from the spatial domain to the frequency domain to encode the watermark into the image efficiently. We find that DWT and patch-wise loss enhance the overall reconstruction quality while performing high message bit accuracy. Through extensive experiments, we demonstrate that our method outperforms the state-of-the-art method and is six times faster than the previous work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 390, + 544, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 390, + 544, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 390, + 544, + 533 + ], + "type": "text", + "content": "Limitations. Our method shows outstanding performance in the digital watermark of NeRF, but training a decoder that extracts watermark from rendered images is time-consuming, taking approximately 12 hours on a single RTX 3090. Our method only encodes a unique message per model, requiring fine-tuning for message insertion every time. We observe that the bit accuracy drops when we apply numerous messages (see Fig. 3). Although previous studies, including us, report robustness against various distortions of images, there is no consideration about cases where the model is attacked. Future work may explore watermarks resilient to cases of model attacks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "text", + "content": "Acknowledgments. This research was supported by the Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture(Project : 4D Content Generation and Copyright Protection with Artificial Intelligence, R2022020068, " + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "text", + "content": " ), the Korea Research Institute for Defense Technology Planning and Advancement (KRIT) Grant funded by the Korean Government(Grant KRIT-CT-23-021, " + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "text", + "content": " ), the National Research Foundation of Korea grant (NRF2022R1F1A1074334, " + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "text", + "content": " ), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), " + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 308, + 545, + 544, + 712 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "12094" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "text", + "content": "[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 203 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "text", + "content": "[4] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 301 + ], + "type": "text", + "content": "[5] Chinmayee Das, Swetalina Panigrahi, Vijay K Sharma, and KK Mahapatra. A novel blind robust image watermarking in dct domain using inter-block coefficient correlation. AEU-International Journal of Electronics and Communications, 68(3):244–253, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 335 + ], + "type": "text", + "content": "[6] Xin Deng, Chao Gao, and Mai Xu. Pirnet: Privacy-preserving image restoration network via wavelet lifting. In ICCV, pages 22368-22377, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 389 + ], + "type": "text", + "content": "[7] Elhameh Farri and Peyman Ayubi. A robust digital video watermarking based on ct-svd domain and chaotic dna sequences for copyright protection. Journal of Ambient Intelligence and Humanized Computing, 14(10):13113-13137, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 390, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 390, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 288, + 434 + ], + "type": "text", + "content": "[8] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. ICCV, 2023. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 479 + ], + "type": "text", + "content": "[9] Felipe ABS Ferreira and Juliano B Lima. A robust 3d point cloud watermarking method based on the graph fourier transform. Multimedia Tools and Applications, 79(3):1921-1950, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 523 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 524, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 288, + 557 + ], + "type": "text", + "content": "[11] Youcai Gao, Jinwei Wang, and Yun-Qing Shi. Dynamic multi-watermarking and detecting in dwt domain. Journal of Real-Time Image Processing, 16:565-576, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 558, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 579 + ], + "type": "text", + "content": "[12] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "type": "text", + "content": "[13] A Hernández-Joaquín, G Melendez-Melendez, and R Cumplido. A secure dwt-based dual watermarking scheme for image authentication and copyright protection. Multimedia Tools and Applications, 82(27):42739-42761, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "text", + "content": "[14] Ningombam Jimson and K Hemachandran. Dft based coefficient exchange digital image watermarking. In 2018 Second International Conference on Intelligent Computing and Control Systems (ICICCS), pages 567-571. IEEE, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[15] Priyank Khare and Vinay Kumar Srivastava. A reliable and secure image watermarking algorithm using homomorphic transform in dwt domain. Multidimensional Systems and Signal Processing, 32:131-160, 2021. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[16] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[17] Hung-Jui Ko, Cheng-Ta Huang, Gwoboa Horng, and WANG Shiuh-Jeng. Robust and blind image watermarking in dct domain using inter-block coefficient correlation. Information Sciences, 517:128-147, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 196 + ], + "type": "text", + "content": "[18] Sanjay Kumar and Binod Kumar Singh. Dwt based color image watermarking using maximum entropy. Multimedia Tools and Applications, 80:15487-15510, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 546, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 252 + ], + "type": "text", + "content": "[19] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 253, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 546, + 306 + ], + "type": "text", + "content": "[20] Guezouli Larbi, Hachani Roumaissa, and Herri Sara. Embedding watermark in the magnitude matrix of the dft of image. In Proceedings of the 2018 International Conference on Computing and Pattern Recognition, pages 106-110, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 308, + 546, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 352 + ], + "type": "text", + "content": "[21] Yong-Seok Lee, Seo Young-Ho, and Kim Dong-Wook. Blind image watermarking based on adaptive data spreading in n-level dwt subbands. Security and communication Networks, 2019, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 546, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 546, + 397 + ], + "type": "text", + "content": "[22] Chenxin Li, Brandon Y Feng, Zhiwen Fan, Panwang Pan, and Zhangyang Wang. Steganerf: Embedding invisible information within neural radiance fields. In ICCV, pages 441-453, 2023. 1, 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 398, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 546, + 431 + ], + "type": "text", + "content": "[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "type": "text", + "content": "[24] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, pages 4273-4284, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 467, + 546, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 546, + 520 + ], + "type": "text", + "content": "[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "type": "text", + "content": "[26] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 546, + 601 + ], + "type": "text", + "content": "[27] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR, pages 13548-13557, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "type": "text", + "content": "[28] Xiyang Luo, Michael Goebel, Elnaz Barshan, and Feng Yang. Leca: A learned approach for efficient cover-agnostic watermarking. In Media Watermarking, Security, and Forensics, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 646, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 546, + 690 + ], + "type": "text", + "content": "[29] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "text", + "content": "[30] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12095" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "radiance fields. In ICCV, pages 22401-22411, 2023. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[31] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 196 + ], + "type": "text", + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "type": "text", + "content": "[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 300 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 302, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 287, + 357 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "text", + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 437 + ], + "type": "text", + "content": "[37] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. ACM Trans. Graph., 41(4), 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "text", + "content": "[38] Shabir A Parah, Javaid A Sheikh, Nazir A Loan, and Ghulam M Bhat. Robust and blind watermarking technique in dct domain using inter-block coefficient differencing. Digital Signal Processing, 53:11-24, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 486, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 529 + ], + "type": "text", + "content": "[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "type": "text", + "content": "[40] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "text", + "content": "[41] SN Prajwalasimha, S Chethan Suputhra, and C Mohan. Performance analysis of combined discrete fourier transformation (dft) and successive division based image watermarking scheme. International Journal of Recent Technology and Engineering, 8(3):34-39, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318-10327, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[43] Deepti Shukla, Nirupama Tiwari, and Deepika Dubey. Survey on digital watermarking techniques. International Journal of Signal Processing, Image Processing and Pattern Recognition, 9(1):239-244, 2016. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[44] S Suthaharan and S Sathananthan. Transform domain technique: robust watermarking for digital images. In Proceedings of the IEEE SoutheastCon 2000. Preparing for The New Millennium' (Cat. No. 00CH37105), pages 407-412. IEEE, 2000. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "text", + "content": "[45] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[46] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "text", + "content": "[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "type": "text", + "content": "[48] Jinwei Wang, Shiguo Lian, and Yun-Qing Shi. Hybrid multiplicative multi-watermarking in dwt domain. Multidimensional Systems and Signal Processing, 28:617-636, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[49] Yinghui Wang, Jing Liu, Yajie Yang, Douli Ma, and Ruijiao Liu. 3d model watermarking algorithm robust to geometric attacks. IET Image Processing, 11(10):822-832, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[50] Muyu Xu, Fangneng Zhan, Jiahui Zhang, Yingchen Yu, Xiaojin Zhang, Christian Theobalt, Ling Shao, and Shijian Lu. Wavenerf: Wavelet-based generalizable neural radiance fields. In ICCV, pages 18195-18204, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 446 + ], + "type": "text", + "content": "[51] Innfarn Yoo, Huiwen Chang, Xiyang Luo, Ondrej Stava, Ce Liu, Peyman Milanfar, and Feng Yang. Deep 3d-to-2d watermarking: embedding messages in 3d meshes and extracting them from 2d renderings. In CVPR, pages 10031–10040, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[52] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[53] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "text", + "content": "[54] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In CVPR, pages 12397-12406, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[55] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, pages 717-733. Springer, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[57] Qiumei Zheng, Nan Liu, Baoqin Cao, Fenghua Wang, and Yanan Yang. Zero-watermarking algorithm in transform domain based on rgb channel and voting strategy. Journal of Information Processing Systems, 16(6):1391-1406, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[58] Haoyi Zhu. X-nerf: Explicit neural radiance field for multiscene 360deg insufficient rgb-d views. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5766–5775, 2023. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12096" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "type": "text", + "content": "[59] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, pages 657–672, 2018. 2, 3, 6" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12097" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_content_list.json b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cf03de774c058e9dba8c7f66c5854f0ea36a7cc --- /dev/null +++ b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_content_list.json @@ -0,0 +1,1552 @@ +[ + { + "type": "text", + "text": "Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models", + "text_level": 1, + "bbox": [ + 238, + 130, + 733, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peifei Zhu, Tsubasa Takahashi, Hirokatsu Kataoka \nLY Corporation", + "bbox": [ + 282, + 203, + 684, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peifei.zhu, tsubasa.takahashi, jpz4219,}@lycorp.co.jp", + "bbox": [ + 241, + 241, + 727, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion Models (DMs) have shown remarkable capabilities in various image-generation tasks. However, there are growing concerns that DMs could be used to imitate unauthorized creations and thus raise copyright issues. To address this issue, we propose a novel framework that embeds personal watermarks in the generation of adversarial examples. Such examples can force DMs to generate images with visible watermarks and prevent DMs from imitating unauthorized images. We construct a generator based on conditional adversarial networks and design three losses (adversarial loss, GAN loss, and perturbation loss) to generate adversarial examples that have subtle perturbation but can effectively attack DMs to prevent copyright violations. Training a generator for a personal watermark by our method only requires 5-10 samples within 2-3 minutes, and once the generator is trained, it can generate adversarial examples with that watermark significantly fast (0.2s per image). We conduct extensive experiments in various conditional image-generation scenarios. Compared to existing methods that generate images with chaotic textures, our method adds visible watermarks on the generated images, which is a more straightforward way to indicate copyright violations. We also observe that our adversarial examples exhibit good transferability across unknown generative models. Therefore, this work provides a simple yet powerful way to protect copyright from DM-based imitation.", + "bbox": [ + 75, + 325, + 473, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 752, + 209, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [19, 42, 46] have garnered significant attention in both the computer vision industry and academia. Compared to previous generative models such as generative adversarial networks (GANs) [15, 23, 32], DMs have demonstrated a remarkable ability to generate realistic images with higher resolution and diversity. Despite the positive aspects of DMs, there are public concerns about potential copyright violations when unauthorized images are", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7428e315389c5e19d3151d63d1b87971882ce6cbe95f36d2b569ef564ab4b89a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 295, + 883, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/933c572476350cfb0bcda88ebf9012a425dbc8eeebfbf9b97adc338b724b50a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 373, + 883, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52aa3d2ff349e20f34806b12e0a16c0178e34baa24b09c1578ac72a4c70f4ae9.jpg", + "image_caption": [ + "Figure 1. Copyright issues of DMs and adversarial example-based methods for copyright protection. Without protection, DMs can easily imitate original images under different image generation scenarios. A previous method (AdvDM [28]) generates adversarial examples by optimization against DMs to prevent DMs from extracting the feature of the original images, resulting in the generation of chaotic images. Our method goes one step further to build a generator that embeds personal watermarks into the generation of adversarial images. Such examples force DMs to generate images with visible watermarks for tracing copyright. Our method is fast, simple yet powerful in protecting copyrights against DMs." + ], + "image_footnote": [], + "bbox": [ + 504, + 455, + 883, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "used to train DMs or to generate new creations by DMs [55]. For example, an entity could use copyrighted paintings shared on the Internet to generate images with a similar style and appearance, potentially earning illegal revenue. Therefore, it is necessary to develop techniques to prevent the imitation of unauthorized creations by DMs.", + "bbox": [ + 496, + 748, + 893, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several previous works have been proposed to address these concerns. Watermarking [2, 33] is a common solution that embeds an invisible message into the image, which can then be extracted to identify the existence of copyright.", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent techniques typically integrate watermarking into the generative process of DMs [7, 12, 38]. However, these methods have several limitations. First, they are primarily designed to protect the copyright of DMs or to distinguish generated images from natural ones, which differs from our goal of protecting the copyright of human creations. Second, embedding watermarks into the generative process requires re-training or at least fine-tuning DMs, and a post-process to extract the watermark is needed to identify it, which can be costly. On the other hand, methods that use adversarial examples for DMs to protect copyright have been proposed recently [27, 28, 50]. The idea is to generate adversarial examples to hinder DMs from extracting the features of the original images. This generation process does not require any re-training of DMs or any post-processing for the generated image. However, there are still several limitations. First, even though these adversarial examples could add chaotic textures to the generated images, such changes are difficult to comprehend. For instance, users may wonder if the problem originates from the DM itself rather than considering the copyright issue. Second, the original image's copyright is not traceable since no copyright-related information is embedded. Third, each adversarial example needs to be optimized separately against DMs, which makes the generation time-consuming.", + "bbox": [ + 76, + 90, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a simple yet powerful method to protect copyrighted images from imitation by DMs. Different from previous adversarial example-based methods, we go one step further by embedding personal watermarks into the generation of adversarial examples for copyright tracing. Such examples could force DMs to generate images with visible watermarks as well as chaotic textures. In addition, instead of using iterative optimization against DMs to generate examples, we train a generator beforehand using only a few samples. Once the generator is trained, it can be used to generate adversarial examples significantly fast. As shown in Figure 1, instead of posting the original image on the Internet, posting adversarial examples with only subtle changes could prevent DMs from imitation. Compared to previous methods, our method provides a more straightforward way to warn of potential copyright violations.", + "bbox": [ + 76, + 472, + 472, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 94, + 715, + 421, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel framework that embeds personal watermarks into the generation of adversarial examples to prevent copyright violations caused by DM-based imitation. Our adversarial examples can force DMs to generate images with visible watermarks for tracing copyright.", + "- We train a generator based on a conditional GAN architecture to generate adversarial examples (Figure 2). We design three losses: adversarial loss, GAN loss, and weighted perturbation loss. These losses aim to improve the attack ability of the adversarial examples while making the perturbation invisible to the human eye." + ], + "bbox": [ + 76, + 734, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conduct extensive experiments in various image generation scenarios, along with robustness evaluations against defenses and assessments of transferability to other models. The experiments demonstrate that our adversarial examples can prevent unauthorized images from being learned and can generate visible watermarks for copyright tracing. Our generation process is significantly fast (0.2s per image), and the generated examples also exhibit good transferability across other generative models.", + "bbox": [ + 500, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 241, + 650, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Generative Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 267, + 756, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, DMs have made significant advancements in various fields [9, 37, 45, 55]. DMs are a type of generative model that simulates a diffusion process to generate new data samples. DMs consist of a forward process and a reverse process. The forward process involves gradually adding Gaussian noise to the data over a series of time steps until it becomes completely random. The reverse process gradually removes noise through a series of learned denoising steps, reconstructing the data back to its original form or generating new samples from the same distribution.", + "bbox": [ + 496, + 291, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DMs have been applied to various generation tasks such as image synthesis [9, 43], text-to-image generation [1, 41, 58, 59], and text-guided image editing [4, 34, 54, 61]. Among these, latent diffusion models [42], which perform the diffusion steps in the latent image space, have demonstrated a remarkable ability to generate high-resolution and realistic images. Despite the impressive performance, some concerns that generating new creations by DMs based on unauthorized images could lead to copyright issues.", + "bbox": [ + 496, + 441, + 893, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Watermarking", + "text_level": 1, + "bbox": [ + 500, + 589, + 702, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image watermarking is a technique that embeds a piece of information into an image to protect the image from copyright infringement. Traditional watermarking methods typically involve altering pixel values [6] or manipulating the frequency domain of an image [35] to embed watermarks. More recently, deep-learning-based methods have been developed. A common approach is using encoder-extractor networks, where an encoder embeds the watermark into the image, and an extractor is trained to retrieve the watermark from the marked image [49, 60, 63].", + "bbox": [ + 496, + 613, + 893, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In terms of watermarking for generative models, most existing works focus on protecting the copyright of generative models [11, 30, 38, 51, 57] or distinguishing generative images from natural images [7, 12, 56, 62]. These methods involve reformulating the training process of DMs to inject watermarks into the images, as well as developing a decoder to identify the embedded messages from the generated images. However, the re-training can be costly and it is impractical to check every generated image with a decoder.", + "bbox": [ + 496, + 765, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24421", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Adversarial Examples for Generative models", + "text_level": 1, + "bbox": [ + 76, + 90, + 459, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial examples are created by adding small but intentional perturbations to input samples such that the perturbed inputs cause the model to produce incorrect answers. Various adversarial attack methods have been proposed to confuse classification models [10, 16, 31, 52] or segmentation models [13, 53, 65]. There are also several existing works studying adversarial examples for generative models such as GANs [24], variational autoencoders (VAEs) [47], and flow-based generative models [39].", + "bbox": [ + 75, + 113, + 467, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, methods that generate adversarial examples for DMs [27, 28, 50] have been proposed. These methods maximize the diffusion loss to obtain a perturbation, and adding such a perturbation to the original image enables DMs to generate images with chaotic content. However, these methods have several drawbacks: 1) such chaotic content is not straightforward to indicate copyright violation, 2) the copyright of the original image is not traceable, and 3) optimization against DMs is time-consuming. We propose a simple yet powerful method that directly incorporates personal watermarks into the generation of adversarial examples, and we build a generator that can generate adversarial examples significantly faster than existing methods.", + "bbox": [ + 75, + 250, + 468, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 458, + 250, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 76, + 483, + 261, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an original image $x$ , we have a diffusion model $\\theta$ that can learn the data distribution of $x$ and generate images with similar style and appearance. In this work, we aim to generate an adversarial example $x'$ that can prevent $\\theta$ from extracting the feature of $x$ and can directly show copyright information on the generated image. For example, for imaged-to-image generation, the generated image from $x'$ by using $\\theta$ can be denoted as $\\theta(x')$ . We design a way to provide the copyright information by using a watermark that contains the ownership details (e.g. artist name) of the original image. Therefore, given a watermark image $m$ , the adversarial example $x'$ that embeds $m$ can be calculated by", + "bbox": [ + 75, + 506, + 468, + 688 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x ^ {\\prime} := \\underset {x ^ {\\prime}} {\\operatorname {a r g m i n}} \\| \\theta \\left(x ^ {\\prime}\\right) - m \\|, \\tag {1} \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\| x - x ^ {\\prime} \\| \\leq \\sigma , \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 696, + 468, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma$ is a constant to control the range of the perturbation. This is a targeted attack setting where we compel the image generated by DMs to be closer to the watermark, while ensuring that the adversarial example remains as close to the original one as possible.", + "bbox": [ + 75, + 750, + 467, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, several issues are associated with generating adversarial examples using Equation 1. First, directly optimizing $x'$ against DMs for each $x$ could be time-consuming, so we need to find a way to speed up the generation process. Second, since there are several image-generation sce", + "bbox": [ + 75, + 825, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "narios related to imitating the original image, such as text-guided image-to-image generation [4], text-to-image generation under textual inversion [14] or DreamBooth [43], we aim to generate an adversarial example $x'$ that can be applicable to all these scenarios. Meanwhile, we also expect that $x'$ could have good transferability on other generative models. Third, since our target attack aims to add personal watermarks to the generated images, the perturbation naturally becomes larger in the watermark regions. We need a solution to make the perturbation invisible to human eyes.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the first issue, we train a generator beforehand instead of implementing optimization every time. To address the second and third issues, we design an adversarial loss, a GAN loss, and a perturbation loss to improve the attack ability across various tasks and models while keeping the perturbation invisible.", + "bbox": [ + 496, + 242, + 890, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture Overview", + "text_level": 1, + "bbox": [ + 498, + 340, + 710, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a conditional GAN architecture for generating watermark-embedded adversarial examples. The architecture of our method contains three components: a generator $G$ , a discriminator $D$ , and a target DM $\\theta$ , shown in Figure 2. The generator $G$ has an encoder to extract features from the inputs and a decoder to generate perturbation using the features. The inputs of $G$ are the original image $x$ and the watermark $m$ , and the perturbation is generated conditioned on $m$ . We perform the conditioning by concatenating $x$ and $m$ along the channel dimension and feeding them into $G$ . The adversarial example $x'$ can be obtained by adding $x$ and the generated perturbation. The discriminator $D$ is a classifier that distinguishes $x'$ from $x$ , and the target DM $\\theta$ is always kept frozen.", + "bbox": [ + 496, + 364, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design three losses to optimize $G$ and $D$ . 1) A GAN loss aims to compel $x'$ to be closer to $x$ . By using the GAN loss, we can train $G$ to produce images that $D$ cannot distinguish from input images. 2) A perturbation loss is designed to bound the magnitude of the perturbation and make it invisible. 3) An adversarial loss is used for $G$ to generate perturbation that can attack DMs under various scenarios. The details are described in Section 3.3. Once the generator $G$ is optimized, it can be used directly to generate adversarial examples significantly fast.", + "bbox": [ + 496, + 575, + 890, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Loss functions", + "text_level": 1, + "bbox": [ + 500, + 736, + 645, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GAN loss. $\\mathcal{L}_{GAN}$ is used to quantify the difference between the adversarial example and the original image. For a set of images, $\\mathcal{L}_{GAN}$ can be written as:", + "bbox": [ + 496, + 758, + 890, + 805 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G A N} = \\mathbb {E} _ {x} \\log D (x) + \\mathbb {E} _ {x} \\log \\left\\{1 - D \\left(x ^ {\\prime}\\right) \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 815, + 890, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The adversarial image $x'$ is generated under the condition of the watermark $m$ , which can be written as:", + "bbox": [ + 496, + 843, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = x + G (x \\mid m). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 883, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24422", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e67fe78b717252a06df7e6f67ab1b44dea6a6340d416f4aea703959dfb42086.jpg", + "image_caption": [ + "Figure 2. Architecture overview. $G$ generates perturbation for $x$ conditioned on $m$ . $D$ and $G$ produce $\\mathcal{L}_{GAN}$ to compel $x'$ to be closer to $x$ . $\\mathcal{L}_{adv}$ aims to force the image generated by $\\theta$ to display a visible watermark $m$ , and $\\mathcal{L}_{pert}$ further bounds the magnitude of the perturbation." + ], + "image_footnote": [], + "bbox": [ + 145, + 92, + 816, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perturbation loss. $\\mathcal{L}_{pert}$ aims to limit the magnitude of the perturbation. We design this loss based on the soft hinge loss used in previous adversarial attacks [3, 29, 52]. However, the soft hinge loss assigns equal importance to all perturbations within the image. In our case, the perturbation in the watermark region will be larger than that in other regions, making the watermark visible even in the adversarial example. To better conceal the watermark as well as bound the perturbation in the whole image, we design a weighted perturbation loss, where we assign larger weights $w$ to the watermark region. This encourages the perturbation in the watermark region to be smaller and invisible. For a set of images, this loss can be expressed as:", + "bbox": [ + 75, + 337, + 472, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p e r t} = \\mathbb {E} _ {x} \\max (0, \\| G (x \\mid m) (1 + w \\cdot m) \\| _ {2} - c), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 542, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is a constant denoting a user-specified bound.", + "bbox": [ + 75, + 568, + 434, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial loss. $\\mathcal{L}_{adv}$ is used to generate a perturbation that can attack DMs. As we expect our adversarial examples to have transferability across different image generation scenarios and various generative models, we design the loss based on the latent representation of latent diffusion models (LDM). We choose LDM for two reasons. First, LDM has outperformed other models in both high-quality image generation and sampling efficiency, and it has been used in various generative models. Generating adversarial examples against LDM naturally increases their transferability across various models. Second, LDM maps an image $x$ to a latent representation $\\varepsilon(x)$ which is the output of the Variational Auto-Encoder (VAE). This representation is widely used in conditional image generation scenarios, and it provides a simple way to control the image generation of DMs. Therefore, instead of directly exploiting the generated image, we minimize the distance between the representation of the adversarial example $\\varepsilon(x')$ and that of the watermark $\\varepsilon(m)$ , formulated as:", + "bbox": [ + 75, + 583, + 472, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = \\mathbb {E} _ {x} \\| \\varepsilon \\left(x ^ {\\prime}\\right) - \\varepsilon (m) \\| _ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization. Finally, we combine the above three losses to formulate the objective function for training, written as:", + "bbox": [ + 498, + 337, + 890, + 368 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G} \\max _ {D} V (D, G) = \\mathcal {L} _ {a d v} + \\alpha \\mathcal {L} _ {G A N} + \\beta \\mathcal {L} _ {p e r t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 380, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are weights to control the balance of the three losses. By playing a minimax game between the $G$ and $D$ , the optimal parameters of the model can be obtained.", + "bbox": [ + 498, + 412, + 890, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Image Generation under Different Settings", + "text_level": 1, + "bbox": [ + 500, + 468, + 864, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our target is to prevent copyright violations, we focus on conditional image generation scenarios where DMs sample the distribution of the original images to generate new ones. Such scenarios include text-guided image-to-image generation [36] and text-to-image generation under textual inversion [14] described in this section, and other scenarios including DreamBooth [43], LoRA [20], and Custom Diffusion [25] discussed in Appendix A.", + "bbox": [ + 496, + 492, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-guide image-to-image generation. In this setting, we can pass a text prompt and an input image to condition the image generation. Our method can be directly applied to this setting, as shown in Figure 2. During the training of the generator and discriminator, we simply use the same prompt (e.g. \"A painting\") for all samples. For the inference, any prompt can be used for generating new images.", + "bbox": [ + 496, + 613, + 890, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Textual inversion. Textual inversion is a method for capturing new concepts from a small number of images. These new concepts can then be used to control image generation in text-to-image settings. The steps to apply our method in this setting are as follows. First, given a small set of images within a similar category (e.g. same artist), we extract a text $S_{*}$ to represent our new concept. We then replace the vector associated with the tokenized string with an optimized embedding $v_{*}$ . The optimization can be applied by minimizing the LDM loss over the images from the small set. During the image generation by DMs, the text $S_{*}$ is used as a condition to generate new images. We conduct this generation", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24423", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for both original images and adversarial images. The adversarial images are generated through the image-to-image generation setting as shown in Figure 2.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 148, + 209, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our method for generating adversarial examples for DMs. We conduct experiments under text-guided image-to-image generation and textual inversion, performing various analyses to assess the effectiveness, robustness, and transferability of our method. We perform the evaluation on WikiArt [48] and ImageNet [44] datasets. All experiments are run on an NVIDIA A100 80GB GPU. Training the generator using 100 samples with 200 epochs takes about 20 minutes.", + "bbox": [ + 75, + 174, + 468, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 318, + 284, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. The WikiArt dataset contains 50k paintings from 195 artists. We randomly select 3000 paintings from 50 artists. We create different watermarks for different artists. These watermarks are binary images with a black background and white artist names such as \"VAN_GOGH\", \"CLAUDE_MONNET\". Since only a small number of samples are needed to train our generator, we select 10 images from each artist (500 images in total) for training and the remaining 2500 images are for evaluation. For the ImageNet dataset, we randomly select 2000 images (100 for training, 1900 for evaluation) from the goldfish, tiger shark, peacock, goose, Eskimo dog, and tabby cat category. The watermarks are designed as \"IMAGENET CAT\", \"IMAGENET FISH\", \"IMAGENETDOG\", etc.", + "bbox": [ + 75, + 340, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate our method from two perspectives. For the adversarial example, we use MSE, PSNR, and SSIM to measure the image quality and the difference from the original image. For the image generated by DMs, we use Fréchet Inception Distance (FID) [18] and precision (prec.) [26] to measure the similarity between the generated and original images. Since we expect DMs to generate images with visible watermarks, we use Normalized Cross-Correlation (NCC) [21] to measure the similarity between the generated image and the watermark. Please note that, unlike traditional watermark methods, we do not extract the watermark from the generated image, and thus NCC is smaller in our case. However, to reduce the influence of the background, we first calculate a difference image between the image generated from the original image and that from the adversarial example, and then use this difference image and the watermark to obtain NCC.", + "bbox": [ + 75, + 553, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. For the generator $G$ , we adopt an architecture similar to that in [22]. For the discriminator $D$ , we use an architecture similar to the discriminator in AdvGAN [52]. Further details regarding the implementation of our model are provided in Appendix B. We set the weight of GAN loss $\\alpha = 1.0$ , the weight of perturbation", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cb277a58a8d311f287a80e34c438110ce965f2cd0acfd545b1bd703e9b6e59cb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNCC ↑FID ↑prec. ↓recall
Original0120.70.850.28
AdvDM0228.20.030.25
Mist0.09205.80.040.33
Ours0.31245.50.030.29
", + "bbox": [ + 535, + 87, + 857, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The performance of the generated image under text-guided image-to-image generation on WikiArt. Recall measures the diversity of the images and is only listed as a reference.", + "bbox": [ + 496, + 191, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss $\\beta = 10$ , the weight for watermark region $w = 4$ , and the bound $c = 10 / 255$ . During the model training, we set the batch size to 8 and the learning rate to 0.001. For comparison, we use AdvDM [28] and Mist (fuse mode) [27], which also use adversarial examples to protect copyright. We set their sampling step to 100 and the maximum perturbation to $10 / 255$ .", + "bbox": [ + 496, + 260, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Text-guided Image-to-Image Generation", + "text_level": 1, + "bbox": [ + 498, + 377, + 846, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of our method under the setting described in Section 3.4. In this setting, a strength parameter is used to control the level of noise added to the original image during the generation of new images. A strength value close to 0 produces an image nearly identical to the original, while a value close to 1 results in an image that significantly differs from the original. We set the strength to 0.3 to simulate realistic imitation, and additional results with different strength values can be found in the Appendix C.1. For the text prompt, we uniformly use \"A painting\" for all images for convenience, and experiments using varied prompts are also included in the Appendix C.2.", + "bbox": [ + 496, + 401, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first show the evaluation result of the image generated by LDM based on either the original image or the adversarial example. A high FID and a low precision mean that the generated image is far from the original one, which further indicates that LDM fails to capture the feature of the original image. As shown in Table 1, our adversarial examples largely increase FID and decrease precision. We also calculate NCC between the generated image and the watermark. For Mist, since it is possible to input a target image into the generation, we apply our idea to use different watermarks as the target images and use the same method to calculate NCC. Results show that our method has a significantly higher NCC, which means our adversarial examples can force LDM to generate images with more visible watermarks. These results suggest our method has great potential to prevent copyright violations caused by LDM imitation.", + "bbox": [ + 496, + 583, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also evaluate the quality of the adversarial example generated by different methods. As shown in Table 2, the image quality of our adversarial examples is slightly higher than that of the existing methods. Moreover, the generation is significantly faster than the existing methods since we", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24424", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/509cc43fbc35210ef2448b2fec15510eb28d0df191e34b7be428627d46093328.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMSE ↓PSNR ↑SSIM ↑Run time
AdvDM0.003829.10.8032s
Mist0.004029.00.8135s
Ours0.003730.10.800.2s
", + "bbox": [ + 89, + 88, + 455, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52edcfdfa9c73d131df21cba66b5313fcd596772c880e84fbe1f8450e1783599.jpg", + "image_caption": [ + "Figure 3. Comparison of different methods under text-guided image-to-image generation. Source is the image input to the LDM, and generated image is the output of the LDM (strength 0.3). The watermarks of our method are designed according to the artist name (from left to right: REMBRANDT, AUGUSTE_REN)." + ], + "image_footnote": [], + "bbox": [ + 86, + 237, + 467, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "directly use the trained generator instead of iterative optimization against DMs. Our method has an advantage when generating a large number of examples.", + "bbox": [ + 76, + 666, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For qualitative evaluation, we visualize two examples generated by different methods in Figure 3. More examples are shown in Appendix D.1. For the adversarial examples, all methods seem to introduce only subtle perturbations that are almost invisible to human eyes. For the generated images based on the adversarial examples, the existing methods succeed in adding chaotic textures to the generated image. However, such changes can be difficult to understand and evaluate. On the other hand, our adversarial examples succeed in generating images with visible watermarks as well as chaotic textures. Our method provides a more straightforward way to indicate copyright violations.", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c806d6c02d7cf4af38e39e4feb417563d8186154a8bef4d629f2b10712481f3d.jpg", + "table_caption": [ + "Table 2. The image quality and generation time (per image) of the adversarial examples on WikiArt." + ], + "table_footnote": [], + "table_body": "
MethodGenerated imageAdversarial example
NCC↑FID ↑PSNR ↑SSIM ↑
Original0120.7--
AdvDM0327.326.30.73
Mist0.15318.426.70.74
Ours0.40412.927.10.75
", + "bbox": [ + 516, + 88, + 879, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. The performance of the generated image and adversarial examples on ImageNet under textual inversion.", + "bbox": [ + 498, + 205, + 890, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Textual Inversion", + "text_level": 1, + "bbox": [ + 500, + 262, + 669, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Textual inversion is another important scenario related to copyright violations. In this scenario, given a small number of images, we first learn to represent the concept of the images in a new word $S_{*}$ , then use $S_{*}$ to guide image generation. Following the method of textual inversion [14], we separate our evaluation images into 4-image groups for each artist or ImageNet category. For each 4-image group, we set the maximum steps to optimize $S_{*}$ to 5000. After the optimization, we generate 10 images for each $S_{*}$ under 10 prompt templates. This process is implemented for both original images and our adversarial examples.", + "bbox": [ + 496, + 287, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluation, we compare our method with existing methods, and the results are shown in Table 3. For the generated image, our method has higher NCC and FID, which indicates our method successfully injects the watermark information in the generation of the adversarial example. In addition, our adversarial example also has higher PSNR and SSIM, which means the generated perturbation is subtle. An example of textual inversion is shown in Figure 4, and examples comparing with the previous methods are shown in Appendix D.2. The features of the original images can be learned and used to guide image generation. On the other hand, when applying our method, images are generated with very clear watermarks, which can be useful for protecting copyrighted images.", + "bbox": [ + 496, + 454, + 892, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 678, + 653, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weights of the losses. This experiment aims to study the mechanism of our method. In equation 6, we design three losses to train the generator. As our target is to attack DMs, the adversarial loss $\\mathcal{L}_{adv}$ is necessary, and we set its weight to 1. We then evaluate the effect of the GAN loss $\\mathcal{L}_{GAN}$ , the perturbation loss $\\mathcal{L}_{pert}$ , and the weight used to control the perturbation of the watermark region $w$ by setting their weights $\\alpha$ , $\\beta$ , and $w$ to 0 or their default value.", + "bbox": [ + 496, + 703, + 890, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 4. The first row is the result of only using $\\mathcal{L}_{adv}$ and setting other losses to 0. In this case, although the attack ability for DMs is high, the generated adversarial examples are full of noise. By comparing the first and second rows, we observe that $\\mathcal{L}_{GAN}$", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24425", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68349c4fdd95264dd252bb6227a6140ed0173626065bbf72428ae736cc7a4839.jpg", + "image_caption": [ + "Figure 4. Examples of textual inversion on ImageNet. The watermarks used in the examples are IMAGENET CAT and IMAGENETDOG." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 864, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aaa8b6f25d4a2725ab33ceaecd1553a43cd5f985d8e5a83cfe228cedb71ae836.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
WeightsGenerated imageAdversarial example
αβwNCC ↑FID ↑PSNR ↑SSIM ↑
×××0.60565.313.10.12
✓××0.41350.622.00.65
×✓×0.35318.224.20.70
×✓✓0.34292.227.80.76
✓✓✓0.31245.530.10.80
", + "bbox": [ + 78, + 335, + 467, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significantly improves the image quality of the adversarial example, even though the attack ability decreases. On the other hand, $\\mathcal{L}_{pert}$ which aims to bound the magnitude of the perturbation, also has an effect similar to that of $\\mathcal{L}_{GAN}$ (as seen when comparing first and third rows). In addition, adding weight term to $\\mathcal{L}_{pert}$ further improves the quality of the adversarial examples without sacrificing too much attack ability (as seen when comparing third and fourth rows). Finally, we combine all these terms to achieve a balanced performance between image quality and attack ability.", + "bbox": [ + 75, + 537, + 468, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Maximum perturbation bound. In the case of adversarial examples, there is always a trade-off between image quality and the success rate of the attack. To explore this trade-off, we conducted experiments with different bounds. We used PSNR to measure the image quality of the adversarial examples, and NCC to assess the visibility of the watermark in the generated image which indicates the attack success rate. Table 5 shows that as the bound increases, the watermark on the generated image becomes more distinct. However, this may also cause the watermark on the adversarial image to become more visible, which can reduce the image quality. In real-world applications, users can adjust the bounds to balance between achieving high-quality adversarial images and ensuring watermark visibility to protect their work.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f964a1979d3498b967d755b55ef0bbc37b33af7b1c22c8438ef02940621201cd.jpg", + "table_caption": [ + "Table 4. The effect of the GAN loss $(\\alpha)$ , the perturbation loss $(\\beta)$ , and the weight of the watermark region $(w)$ on generating adversarial examples (WikiArt) under image-to-image generation." + ], + "table_footnote": [], + "table_body": "
Bound/25526101520
PSNR ↑40.134.230.124.217.5
NCC ↑0.130.220.310.370.41
", + "bbox": [ + 527, + 335, + 864, + 385 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27d246381ef2fef072dd4bef04227b5271cb53e48cd8f0252b9ae81474f7bd48.jpg", + "table_caption": [ + "Table 5. The effect of using different perturbation bound. The experiment uses WikiArt under image-to-image generation." + ], + "table_footnote": [], + "table_body": "
SamplesNCC ↑FID ↑prec. ↓train time
10.12177.80.3220 sec
50.28235.30.041 min
100.30240.60.032 min
1000.31243.50.0220 min
", + "bbox": [ + 521, + 452, + 870, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. The effect and training time of using different number of samples to train a generator for a specific watermark. The experiment is conducted under image-to-image generation on WikiArt.", + "bbox": [ + 498, + 556, + 890, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of samples for training. To generate adversarial examples that contain a specific watermark, we need to include samples with that watermark in the training data. In this experiment, we analyze how many samples we need for a watermark to be learned by the generator. We choose a different number of samples to train a generator for a specific watermark. Please note that different watermarks can be trained together within the same generator, and this is the default setting for our other experiments. The results of training one personal watermark are in Table 6. We observe there is no large performance difference between using 10 samples and 100 samples for training, which shows that our method only needs a small number of samples to train to embed a specific watermark. Therefore, one use case of our method is to create a personal watermark generator using a few samples within $2 - 3\\mathrm{min}$ , and then this generator can be used to embed the personal watermark for any of that user's creations to protect copyright.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24426", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06b2645255f05d97384282410bca69fb0fa15555a568dd076c66a03828d73582.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DefenseNo attack (Origin)Attack (Ours)
NCC ↑FID ↑NCC ↑FID ↑
No defense0120.70.31245.5
JPEG0124.10.21202.2
RS0123.20.20195.6
TVM0132.50.16166.8
", + "bbox": [ + 93, + 88, + 452, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. The effects of applying defenses on adversarial examples under image-to-image generation.", + "bbox": [ + 76, + 205, + 468, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/262dcc52c79a31b0c9c2483d4d4b297eefc1d993eb535d18884507cf1bf7ed7e.jpg", + "image_caption": [ + "Figure 5. Examples of the generated image from original image, attack image, and attack image with defense." + ], + "image_footnote": [], + "bbox": [ + 80, + 253, + 459, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Robustness of adversarial examples", + "text_level": 1, + "bbox": [ + 76, + 463, + 383, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the robustness of the adversarial examples generated by our method, we apply several widely used adversarial defenses including JPEG compression [8], Randomized Smoothing (RS) [5], and Total Variance Minimization (TVM) [17]. Such defenses are directly processed on the adversarial examples to eliminate the perturbations.", + "bbox": [ + 75, + 487, + 468, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of applying attacks and defenses are shown in Table 7. The \"No attack\" column means applying defense on the original image to generate images. There is no watermark being embedded so NCC almost equals 0, and applying defenses slightly increases FID since the quality of the original image decreases. On the other hand, the \"Attack\" column means we first generate the adversarial examples, apply defenses to them (except the first row), and then use them to generate new images. We observe that NCC and FID both decrease when adding defenses to our adversarial examples. However, compared with \"No attack\", the \"Attack\" column still has higher NCC and FID which means the defenses cannot completely eliminate the effect of our attack. In Figure 5, we show some examples under defenses, and the watermark is still visible for all cases.", + "bbox": [ + 75, + 579, + 468, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Transferability on Other Generative Models", + "text_level": 1, + "bbox": [ + 76, + 816, + 450, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we explore the transferability of our method on other generative models. This is also known as the black-box attack, where one target DM is used to train the generator and obtain the adversarial examples, and then these", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca6ab38915a1c85aa02549dbffa4013df5390a5f4809492121a5df2dc13ab456.jpg", + "image_caption": [ + "Figure 6. Black-box attack to other models under image-to-image generation. The first row: generated images from the original image. Second row: generated images from our adversarial example." + ], + "image_footnote": [], + "bbox": [ + 501, + 93, + 890, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "adversarial examples are directly used to attack other models. We generate adversarial examples on Stable Diffusion 1.5 (SD 1.5) and test on other models or tools including Dreamshaper $8^{2}$ , NovelAI, and Runway AI magic tools (image variation) under image-to-image generation. The details of the settings are described in the Appendix E.", + "bbox": [ + 498, + 314, + 890, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are shown in Figure 6, and more examples can be found in Appendix D.3. Compared to images generated from the original image (first row), the images generated from our adversarial examples (second row) contain chaotic textures and a visible watermark for most cases. For the result of the Runway tool, although there is no obvious watermark on the generated image (as no parameter can be adjusted), our method still generates an image that differs significantly from the original one. Even though the watermarks are not as strong as they are in the white-box setting, our method still demonstrates good transferability. These results show that our method has the potential to prevent copyright violations from various image generative models.", + "bbox": [ + 496, + 407, + 892, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 625, + 617, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel method for copyright prevention against DMs. We build a generator to embed personal watermarks into the adversarial examples, and such examples can force DMs to generate images with visible watermarks and chaotic textures. We use a conditional GAN architecture with three carefully designed losses to optimize the generator. Experiments show our method performs well in various image generation scenarios and exhibits good transferability to attack other models. Additionally, our generation is significantly faster than those of previous methods. Therefore, this work provides a simple yet powerful way to protect image copyright against DMs.", + "bbox": [ + 496, + 651, + 890, + 832 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://huggingface.co/runwayml/stable-diffusion-v1-5", + "bbox": [ + 514, + 849, + 813, + 862 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2https://civitai.com/models/4384/dreamshaper", + "bbox": [ + 519, + 862, + 763, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3https://novelai.net/", + "bbox": [ + 519, + 875, + 625, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4https://runwayml.com/ai-magic-tools/image-to-image/", + "bbox": [ + 519, + 887, + 810, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 1", + "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 4", + "[4] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3, 1", + "[5] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019. 8", + "[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, and Chris Honsinger. Digital watermarking. Journal of Electronic Imaging, 11(3):414-414, 2002. 2", + "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2", + "[8] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 8", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3", + "[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2", + "[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2", + "[13] Volker Fischer, Mummadi Chaithanya Kumar, Jan Hendrik Metzen, and Thomas Brox. Adversarial examples for semantic image segmentation. arXiv preprint arXiv:1703.01101, 2017. 3", + "[14] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3, 4, 6", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3", + "[17] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2017. 8", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 1", + "[21] Bernd Jähne. Digital image processing. Springer Science & Business Media, 2005. 5", + "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 1", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1", + "[24] Jernej Kos, Ian Fischer, and Dawn Song. Adversarial examples for generative models. In 2018 IEEE security and privacy workshops (spw), pages 36-42. IEEE, 2018. 3", + "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4", + "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5", + "[27] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2, 3, 5", + "[28] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, XUE Zhengui, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. 2023. 1, 2, 3, 5" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24428", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. arXiv preprint arXiv:1611.02770, 2016. 4", + "[30] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2", + "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3", + "[32] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1", + "[33] Saraju P Mohanty. Digital watermarking: A tutorial review. URL: http://www.csee.usf.edu/smohanty/research/Reports/WMSurvey1999Mohanty.pdf, 1999.", + "[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2", + "[35] KA Navas, Mathews Cheriyan Ajay, M Lekshmi, Tampy S Archana, and M Sasikumar. Dwt-dct-svd based watermarking. In 2008 3rd International Conference on Communication Systems Software and Middleware and Workshops (COMSWARE'08), pages 271-274. IEEE, 2008. 2", + "[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 4", + "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[38] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2", + "[39] Phillip Pope, Yogesh Balaji, and Soheil Feizi. Adversarial robustness of flow-based generative models. In International Conference on Artificial Intelligence and Statistics, pages 3795-3805. PMLR, 2020. 3", + "[40] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2", + "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2", + "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 1", + "[44] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5", + "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1", + "[47] Pedro Tabacof, Julia Tavares, and Eduardo Valle. Adversarial images for variational autoencoders. arXiv preprint arXiv:1612.00155, 2016. 3", + "[48] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5", + "[49] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 2", + "[50] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3", + "[51] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2", + "[52] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610, 2018. 3, 4, 5", + "[53] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3", + "[54] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2", + "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24429", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "methods and applications. ACM Computing Surveys, 2022. 1, 2", + "[56] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2", + "[57] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2", + "[58] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909, 2023. 2", + "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[60] Ru Zhang, Shiqi Dong, and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia tools and applications, 78:8559-8575, 2019. 2", + "[61] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2", + "[62] Yuan Zhao, Bo Liu, Ming Ding, Baoping Liu, Tianqing Zhu, and Xin Yu. Proactive deepfake defence via identity watermarking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4602-4611, 2023. 2", + "[63] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 2", + "[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1", + "[65] Peifei Zhu, Genki Osada, Hirokatsu Kataoka, and Tsubasa Takahashi. Frequency-aware gan for adversarial manipulation generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4315-4324, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 739 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24430", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_model.json b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f6dbf8950e668c06c4c9ae3e2135c1881fcb8ed2 --- /dev/null +++ b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.239, + 0.131, + 0.734, + 0.177 + ], + "angle": 0, + "content": "Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.204, + 0.686, + 0.24 + ], + "angle": 0, + "content": "Peifei Zhu, Tsubasa Takahashi, Hirokatsu Kataoka \nLY Corporation" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.242, + 0.728, + 0.259 + ], + "angle": 0, + "content": "{peifei.zhu, tsubasa.takahashi, jpz4219,}@lycorp.co.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.719 + ], + "angle": 0, + "content": "Diffusion Models (DMs) have shown remarkable capabilities in various image-generation tasks. However, there are growing concerns that DMs could be used to imitate unauthorized creations and thus raise copyright issues. To address this issue, we propose a novel framework that embeds personal watermarks in the generation of adversarial examples. Such examples can force DMs to generate images with visible watermarks and prevent DMs from imitating unauthorized images. We construct a generator based on conditional adversarial networks and design three losses (adversarial loss, GAN loss, and perturbation loss) to generate adversarial examples that have subtle perturbation but can effectively attack DMs to prevent copyright violations. Training a generator for a personal watermark by our method only requires 5-10 samples within 2-3 minutes, and once the generator is trained, it can generate adversarial examples with that watermark significantly fast (0.2s per image). We conduct extensive experiments in various conditional image-generation scenarios. Compared to existing methods that generate images with chaotic textures, our method adds visible watermarks on the generated images, which is a more straightforward way to indicate copyright violations. We also observe that our adversarial examples exhibit good transferability across unknown generative models. Therefore, this work provides a simple yet powerful way to protect copyright from DM-based imitation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.21, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Diffusion models (DMs) [19, 42, 46] have garnered significant attention in both the computer vision industry and academia. Compared to previous generative models such as generative adversarial networks (GANs) [15, 23, 32], DMs have demonstrated a remarkable ability to generate realistic images with higher resolution and diversity. Despite the positive aspects of DMs, there are public concerns about potential copyright violations when unauthorized images are" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.296, + 0.885, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.374, + 0.884, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.457, + 0.885, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Figure 1. Copyright issues of DMs and adversarial example-based methods for copyright protection. Without protection, DMs can easily imitate original images under different image generation scenarios. A previous method (AdvDM [28]) generates adversarial examples by optimization against DMs to prevent DMs from extracting the feature of the original images, resulting in the generation of chaotic images. Our method goes one step further to build a generator that embeds personal watermarks into the generation of adversarial images. Such examples force DMs to generate images with visible watermarks for tracing copyright. Our method is fast, simple yet powerful in protecting copyrights against DMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.894, + 0.839 + ], + "angle": 0, + "content": "used to train DMs or to generate new creations by DMs [55]. For example, an entity could use copyrighted paintings shared on the Internet to generate images with a similar style and appearance, potentially earning illegal revenue. Therefore, it is necessary to develop techniques to prevent the imitation of unauthorized creations by DMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Several previous works have been proposed to address these concerns. Watermarking [2, 33] is a common solution that embeds an invisible message into the image, which can then be extracted to identify the existence of copyright." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.469 + ], + "angle": 0, + "content": "Recent techniques typically integrate watermarking into the generative process of DMs [7, 12, 38]. However, these methods have several limitations. First, they are primarily designed to protect the copyright of DMs or to distinguish generated images from natural ones, which differs from our goal of protecting the copyright of human creations. Second, embedding watermarks into the generative process requires re-training or at least fine-tuning DMs, and a post-process to extract the watermark is needed to identify it, which can be costly. On the other hand, methods that use adversarial examples for DMs to protect copyright have been proposed recently [27, 28, 50]. The idea is to generate adversarial examples to hinder DMs from extracting the features of the original images. This generation process does not require any re-training of DMs or any post-processing for the generated image. However, there are still several limitations. First, even though these adversarial examples could add chaotic textures to the generated images, such changes are difficult to comprehend. For instance, users may wonder if the problem originates from the DM itself rather than considering the copyright issue. Second, the original image's copyright is not traceable since no copyright-related information is embedded. Third, each adversarial example needs to be optimized separately against DMs, which makes the generation time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.473, + 0.473, + 0.715 + ], + "angle": 0, + "content": "In this work, we propose a simple yet powerful method to protect copyrighted images from imitation by DMs. Different from previous adversarial example-based methods, we go one step further by embedding personal watermarks into the generation of adversarial examples for copyright tracing. Such examples could force DMs to generate images with visible watermarks as well as chaotic textures. In addition, instead of using iterative optimization against DMs to generate examples, we train a generator beforehand using only a few samples. Once the generator is trained, it can be used to generate adversarial examples significantly fast. As shown in Figure 1, instead of posting the original image on the Internet, posting adversarial examples with only subtle changes could prevent DMs from imitation. Compared to previous methods, our method provides a more straightforward way to warn of potential copyright violations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.717, + 0.423, + 0.731 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.81 + ], + "angle": 0, + "content": "- We design a novel framework that embeds personal watermarks into the generation of adversarial examples to prevent copyright violations caused by DM-based imitation. Our adversarial examples can force DMs to generate images with visible watermarks for tracing copyright." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We train a generator based on a conditional GAN architecture to generate adversarial examples (Figure 2). We design three losses: adversarial loss, GAN loss, and weighted perturbation loss. These losses aim to improve the attack ability of the adversarial examples while making the perturbation invisible to the human eye." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "- We conduct extensive experiments in various image generation scenarios, along with robustness evaluations against defenses and assessments of transferability to other models. The experiments demonstrate that our adversarial examples can prevent unauthorized images from being learned and can generate visible watermarks for copyright tracing. Our generation process is significantly fast (0.2s per image), and the generated examples also exhibit good transferability across other generative models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.242, + 0.651, + 0.258 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.757, + 0.283 + ], + "angle": 0, + "content": "2.1. Generative Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.442 + ], + "angle": 0, + "content": "In recent years, DMs have made significant advancements in various fields [9, 37, 45, 55]. DMs are a type of generative model that simulates a diffusion process to generate new data samples. DMs consist of a forward process and a reverse process. The forward process involves gradually adding Gaussian noise to the data over a series of time steps until it becomes completely random. The reverse process gradually removes noise through a series of learned denoising steps, reconstructing the data back to its original form or generating new samples from the same distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.895, + 0.581 + ], + "angle": 0, + "content": "DMs have been applied to various generation tasks such as image synthesis [9, 43], text-to-image generation [1, 41, 58, 59], and text-guided image editing [4, 34, 54, 61]. Among these, latent diffusion models [42], which perform the diffusion steps in the latent image space, have demonstrated a remarkable ability to generate high-resolution and realistic images. Despite the impressive performance, some concerns that generating new creations by DMs based on unauthorized images could lead to copyright issues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.704, + 0.607 + ], + "angle": 0, + "content": "2.2. Image Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Image watermarking is a technique that embeds a piece of information into an image to protect the image from copyright infringement. Traditional watermarking methods typically involve altering pixel values [6] or manipulating the frequency domain of an image [35] to embed watermarks. More recently, deep-learning-based methods have been developed. A common approach is using encoder-extractor networks, where an encoder embeds the watermark into the image, and an extractor is trained to retrieve the watermark from the marked image [49, 60, 63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In terms of watermarking for generative models, most existing works focus on protecting the copyright of generative models [11, 30, 38, 51, 57] or distinguishing generative images from natural images [7, 12, 56, 62]. These methods involve reformulating the training process of DMs to inject watermarks into the images, as well as developing a decoder to identify the embedded messages from the generated images. However, the re-training can be costly and it is impractical to check every generated image with a decoder." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "24421" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.46, + 0.108 + ], + "angle": 0, + "content": "2.3. Adversarial Examples for Generative models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.468, + 0.25 + ], + "angle": 0, + "content": "Adversarial examples are created by adding small but intentional perturbations to input samples such that the perturbed inputs cause the model to produce incorrect answers. Various adversarial attack methods have been proposed to confuse classification models [10, 16, 31, 52] or segmentation models [13, 53, 65]. There are also several existing works studying adversarial examples for generative models such as GANs [24], variational autoencoders (VAEs) [47], and flow-based generative models [39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Recently, methods that generate adversarial examples for DMs [27, 28, 50] have been proposed. These methods maximize the diffusion loss to obtain a perturbation, and adding such a perturbation to the original image enables DMs to generate images with chaotic content. However, these methods have several drawbacks: 1) such chaotic content is not straightforward to indicate copyright violation, 2) the copyright of the original image is not traceable, and 3) optimization against DMs is time-consuming. We propose a simple yet powerful method that directly incorporates personal watermarks into the generation of adversarial examples, and we build a generator that can generate adversarial examples significantly faster than existing methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.459, + 0.251, + 0.475 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.262, + 0.499 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Given an original image \\( x \\), we have a diffusion model \\( \\theta \\) that can learn the data distribution of \\( x \\) and generate images with similar style and appearance. In this work, we aim to generate an adversarial example \\( x' \\) that can prevent \\( \\theta \\) from extracting the feature of \\( x \\) and can directly show copyright information on the generated image. For example, for imaged-to-image generation, the generated image from \\( x' \\) by using \\( \\theta \\) can be denoted as \\( \\theta(x') \\). We design a way to provide the copyright information by using a watermark that contains the ownership details (e.g. artist name) of the original image. Therefore, given a watermark image \\( m \\), the adversarial example \\( x' \\) that embeds \\( m \\) can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.697, + 0.469, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x ^ {\\prime} := \\underset {x ^ {\\prime}} {\\operatorname {a r g m i n}} \\| \\theta \\left(x ^ {\\prime}\\right) - m \\|, \\tag {1} \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\| x - x ^ {\\prime} \\| \\leq \\sigma , \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.826 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a constant to control the range of the perturbation. This is a targeted attack setting where we compel the image generated by DMs to be closer to the watermark, while ensuring that the adversarial example remains as close to the original one as possible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.902 + ], + "angle": 0, + "content": "However, several issues are associated with generating adversarial examples using Equation 1. First, directly optimizing \\( x' \\) against DMs for each \\( x \\) could be time-consuming, so we need to find a way to speed up the generation process. Second, since there are several image-generation sce" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "narios related to imitating the original image, such as text-guided image-to-image generation [4], text-to-image generation under textual inversion [14] or DreamBooth [43], we aim to generate an adversarial example \\( x' \\) that can be applicable to all these scenarios. Meanwhile, we also expect that \\( x' \\) could have good transferability on other generative models. Third, since our target attack aims to add personal watermarks to the generated images, the perturbation naturally becomes larger in the watermark regions. We need a solution to make the perturbation invisible to human eyes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.333 + ], + "angle": 0, + "content": "To address the first issue, we train a generator beforehand instead of implementing optimization every time. To address the second and third issues, we design an adversarial loss, a GAN loss, and a perturbation loss to improve the attack ability across various tasks and models while keeping the perturbation invisible." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.342, + 0.712, + 0.357 + ], + "angle": 0, + "content": "3.2. Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.576 + ], + "angle": 0, + "content": "We propose a conditional GAN architecture for generating watermark-embedded adversarial examples. The architecture of our method contains three components: a generator \\(G\\), a discriminator \\(D\\), and a target DM \\(\\theta\\), shown in Figure 2. The generator \\(G\\) has an encoder to extract features from the inputs and a decoder to generate perturbation using the features. The inputs of \\(G\\) are the original image \\(x\\) and the watermark \\(m\\), and the perturbation is generated conditioned on \\(m\\). We perform the conditioning by concatenating \\(x\\) and \\(m\\) along the channel dimension and feeding them into \\(G\\). The adversarial example \\(x'\\) can be obtained by adding \\(x\\) and the generated perturbation. The discriminator \\(D\\) is a classifier that distinguishes \\(x'\\) from \\(x\\), and the target DM \\(\\theta\\) is always kept frozen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.729 + ], + "angle": 0, + "content": "We design three losses to optimize \\( G \\) and \\( D \\). 1) A GAN loss aims to compel \\( x' \\) to be closer to \\( x \\). By using the GAN loss, we can train \\( G \\) to produce images that \\( D \\) cannot distinguish from input images. 2) A perturbation loss is designed to bound the magnitude of the perturbation and make it invisible. 3) An adversarial loss is used for \\( G \\) to generate perturbation that can attack DMs under various scenarios. The details are described in Section 3.3. Once the generator \\( G \\) is optimized, it can be used directly to generate adversarial examples significantly fast." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.647, + 0.752 + ], + "angle": 0, + "content": "3.3. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.806 + ], + "angle": 0, + "content": "GAN loss. \\(\\mathcal{L}_{GAN}\\) is used to quantify the difference between the adversarial example and the original image. For a set of images, \\(\\mathcal{L}_{GAN}\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.816, + 0.892, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G A N} = \\mathbb {E} _ {x} \\log D (x) + \\mathbb {E} _ {x} \\log \\left\\{1 - D \\left(x ^ {\\prime}\\right) \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.844, + 0.892, + 0.873 + ], + "angle": 0, + "content": "The adversarial image \\( x' \\) is generated under the condition of the watermark \\( m \\), which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = x + G (x \\mid m). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24422" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.093, + 0.817, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.283, + 0.895, + 0.313 + ], + "angle": 0, + "content": "Figure 2. Architecture overview. \\( G \\) generates perturbation for \\( x \\) conditioned on \\( m \\). \\( D \\) and \\( G \\) produce \\( \\mathcal{L}_{GAN} \\) to compel \\( x' \\) to be closer to \\( x \\). \\( \\mathcal{L}_{adv} \\) aims to force the image generated by \\( \\theta \\) to display a visible watermark \\( m \\), and \\( \\mathcal{L}_{pert} \\) further bounds the magnitude of the perturbation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.535 + ], + "angle": 0, + "content": "Perturbation loss. \\(\\mathcal{L}_{pert}\\) aims to limit the magnitude of the perturbation. We design this loss based on the soft hinge loss used in previous adversarial attacks [3, 29, 52]. However, the soft hinge loss assigns equal importance to all perturbations within the image. In our case, the perturbation in the watermark region will be larger than that in other regions, making the watermark visible even in the adversarial example. To better conceal the watermark as well as bound the perturbation in the whole image, we design a weighted perturbation loss, where we assign larger weights \\(w\\) to the watermark region. This encourages the perturbation in the watermark region to be smaller and invisible. For a set of images, this loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.544, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p e r t} = \\mathbb {E} _ {x} \\max (0, \\| G (x \\mid m) (1 + w \\cdot m) \\| _ {2} - c), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.435, + 0.584 + ], + "angle": 0, + "content": "where \\(c\\) is a constant denoting a user-specified bound." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.871 + ], + "angle": 0, + "content": "Adversarial loss. \\(\\mathcal{L}_{adv}\\) is used to generate a perturbation that can attack DMs. As we expect our adversarial examples to have transferability across different image generation scenarios and various generative models, we design the loss based on the latent representation of latent diffusion models (LDM). We choose LDM for two reasons. First, LDM has outperformed other models in both high-quality image generation and sampling efficiency, and it has been used in various generative models. Generating adversarial examples against LDM naturally increases their transferability across various models. Second, LDM maps an image \\(x\\) to a latent representation \\(\\varepsilon(x)\\) which is the output of the Variational Auto-Encoder (VAE). This representation is widely used in conditional image generation scenarios, and it provides a simple way to control the image generation of DMs. Therefore, instead of directly exploiting the generated image, we minimize the distance between the representation of the adversarial example \\(\\varepsilon(x')\\) and that of the watermark \\(\\varepsilon(m)\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.885, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = \\mathbb {E} _ {x} \\| \\varepsilon \\left(x ^ {\\prime}\\right) - \\varepsilon (m) \\| _ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.338, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Optimization. Finally, we combine the above three losses to formulate the objective function for training, written as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.381, + 0.892, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\min _ {G} \\max _ {D} V (D, G) = \\mathcal {L} _ {a d v} + \\alpha \\mathcal {L} _ {G A N} + \\beta \\mathcal {L} _ {p e r t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.46 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are weights to control the balance of the three losses. By playing a minimax game between the \\(G\\) and \\(D\\), the optimal parameters of the model can be obtained." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.866, + 0.485 + ], + "angle": 0, + "content": "3.4. Image Generation under Different Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Since our target is to prevent copyright violations, we focus on conditional image generation scenarios where DMs sample the distribution of the original images to generate new ones. Such scenarios include text-guided image-to-image generation [36] and text-to-image generation under textual inversion [14] described in this section, and other scenarios including DreamBooth [43], LoRA [20], and Custom Diffusion [25] discussed in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.72 + ], + "angle": 0, + "content": "Text-guide image-to-image generation. In this setting, we can pass a text prompt and an input image to condition the image generation. Our method can be directly applied to this setting, as shown in Figure 2. During the training of the generator and discriminator, we simply use the same prompt (e.g. \"A painting\") for all samples. For the inference, any prompt can be used for generating new images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Textual inversion. Textual inversion is a method for capturing new concepts from a small number of images. These new concepts can then be used to control image generation in text-to-image settings. The steps to apply our method in this setting are as follows. First, given a small set of images within a similar category (e.g. same artist), we extract a text \\( S_{*} \\) to represent our new concept. We then replace the vector associated with the tokenized string with an optimized embedding \\( v_{*} \\). The optimization can be applied by minimizing the LDM loss over the images from the small set. During the image generation by DMs, the text \\( S_{*} \\) is used as a condition to generate new images. We conduct this generation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24423" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.138 + ], + "angle": 0, + "content": "for both original images and adversarial images. The adversarial images are generated through the image-to-image generation setting as shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.15, + 0.21, + 0.167 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.175, + 0.47, + 0.311 + ], + "angle": 0, + "content": "In this section, we evaluate our method for generating adversarial examples for DMs. We conduct experiments under text-guided image-to-image generation and textual inversion, performing various analyses to assess the effectiveness, robustness, and transferability of our method. We perform the evaluation on WikiArt [48] and ImageNet [44] datasets. All experiments are run on an NVIDIA A100 80GB GPU. Training the generator using 100 samples with 200 epochs takes about 20 minutes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.319, + 0.285, + 0.336 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Datasets. The WikiArt dataset contains 50k paintings from 195 artists. We randomly select 3000 paintings from 50 artists. We create different watermarks for different artists. These watermarks are binary images with a black background and white artist names such as \"VAN_GOGH\", \"CLAUDE_MONNET\". Since only a small number of samples are needed to train our generator, we select 10 images from each artist (500 images in total) for training and the remaining 2500 images are for evaluation. For the ImageNet dataset, we randomly select 2000 images (100 for training, 1900 for evaluation) from the goldfish, tiger shark, peacock, goose, Eskimo dog, and tabby cat category. The watermarks are designed as \"IMAGENET CAT\", \"IMAGENET FISH\", \"IMAGENETDOG\", etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate our method from two perspectives. For the adversarial example, we use MSE, PSNR, and SSIM to measure the image quality and the difference from the original image. For the image generated by DMs, we use Fréchet Inception Distance (FID) [18] and precision (prec.) [26] to measure the similarity between the generated and original images. Since we expect DMs to generate images with visible watermarks, we use Normalized Cross-Correlation (NCC) [21] to measure the similarity between the generated image and the watermark. Please note that, unlike traditional watermark methods, we do not extract the watermark from the generated image, and thus NCC is smaller in our case. However, to reduce the influence of the background, we first calculate a difference image between the image generated from the original image and that from the adversarial example, and then use this difference image and the watermark to obtain NCC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Implementation details. For the generator \\( G \\), we adopt an architecture similar to that in [22]. For the discriminator \\( D \\), we use an architecture similar to the discriminator in AdvGAN [52]. Further details regarding the implementation of our model are provided in Appendix B. We set the weight of GAN loss \\( \\alpha = 1.0 \\), the weight of perturbation" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.088, + 0.858, + 0.169 + ], + "angle": 0, + "content": "
MethodNCC ↑FID ↑prec. ↓recall
Original0120.70.850.28
AdvDM0228.20.030.25
Mist0.09205.80.040.33
Ours0.31245.50.030.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Table 1. The performance of the generated image under text-guided image-to-image generation on WikiArt. Recall measures the diversity of the images and is only listed as a reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.368 + ], + "angle": 0, + "content": "loss \\(\\beta = 10\\), the weight for watermark region \\(w = 4\\), and the bound \\(c = 10 / 255\\). During the model training, we set the batch size to 8 and the learning rate to 0.001. For comparison, we use AdvDM [28] and Mist (fuse mode) [27], which also use adversarial examples to protect copyright. We set their sampling step to 100 and the maximum perturbation to \\(10 / 255\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.378, + 0.847, + 0.394 + ], + "angle": 0, + "content": "4.2. Text-guided Image-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.583 + ], + "angle": 0, + "content": "We evaluate the performance of our method under the setting described in Section 3.4. In this setting, a strength parameter is used to control the level of noise added to the original image during the generation of new images. A strength value close to 0 produces an image nearly identical to the original, while a value close to 1 results in an image that significantly differs from the original. We set the strength to 0.3 to simulate realistic imitation, and additional results with different strength values can be found in the Appendix C.1. For the text prompt, we uniformly use \"A painting\" for all images for convenience, and experiments using varied prompts are also included in the Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We first show the evaluation result of the image generated by LDM based on either the original image or the adversarial example. A high FID and a low precision mean that the generated image is far from the original one, which further indicates that LDM fails to capture the feature of the original image. As shown in Table 1, our adversarial examples largely increase FID and decrease precision. We also calculate NCC between the generated image and the watermark. For Mist, since it is possible to input a target image into the generation, we apply our idea to use different watermarks as the target images and use the same method to calculate NCC. Results show that our method has a significantly higher NCC, which means our adversarial examples can force LDM to generate images with more visible watermarks. These results suggest our method has great potential to prevent copyright violations caused by LDM imitation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also evaluate the quality of the adversarial example generated by different methods. As shown in Table 2, the image quality of our adversarial examples is slightly higher than that of the existing methods. Moreover, the generation is significantly faster than the existing methods since we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24424" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.457, + 0.155 + ], + "angle": 0, + "content": "
MethodMSE ↓PSNR ↑SSIM ↑Run time
AdvDM0.003829.10.8032s
Mist0.004029.00.8135s
Ours0.003730.10.800.2s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.177, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Table 2. The image quality and generation time (per image) of the adversarial examples on WikiArt." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.238, + 0.468, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Figure 3. Comparison of different methods under text-guided image-to-image generation. Source is the image input to the LDM, and generated image is the output of the LDM (strength 0.3). The watermarks of our method are designed according to the artist name (from left to right: REMBRANDT, AUGUSTE_REN)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.667, + 0.47, + 0.713 + ], + "angle": 0, + "content": "directly use the trained generator instead of iterative optimization against DMs. Our method has an advantage when generating a large number of examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For qualitative evaluation, we visualize two examples generated by different methods in Figure 3. More examples are shown in Appendix D.1. For the adversarial examples, all methods seem to introduce only subtle perturbations that are almost invisible to human eyes. For the generated images based on the adversarial examples, the existing methods succeed in adding chaotic textures to the generated image. However, such changes can be difficult to understand and evaluate. On the other hand, our adversarial examples succeed in generating images with visible watermarks as well as chaotic textures. Our method provides a more straightforward way to indicate copyright violations." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.88, + 0.185 + ], + "angle": 0, + "content": "
MethodGenerated imageAdversarial example
NCC↑FID ↑PSNR ↑SSIM ↑
Original0120.7--
AdvDM0327.326.30.73
Mist0.15318.426.70.74
Ours0.40412.927.10.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.207, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Table 3. The performance of the generated image and adversarial examples on ImageNet under textual inversion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.263, + 0.671, + 0.278 + ], + "angle": 0, + "content": "4.3. Textual Inversion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Textual inversion is another important scenario related to copyright violations. In this scenario, given a small number of images, we first learn to represent the concept of the images in a new word \\( S_{*} \\), then use \\( S_{*} \\) to guide image generation. Following the method of textual inversion [14], we separate our evaluation images into 4-image groups for each artist or ImageNet category. For each 4-image group, we set the maximum steps to optimize \\( S_{*} \\) to 5000. After the optimization, we generate 10 images for each \\( S_{*} \\) under 10 prompt templates. This process is implemented for both original images and our adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.668 + ], + "angle": 0, + "content": "For evaluation, we compare our method with existing methods, and the results are shown in Table 3. For the generated image, our method has higher NCC and FID, which indicates our method successfully injects the watermark information in the generation of the adversarial example. In addition, our adversarial example also has higher PSNR and SSIM, which means the generated perturbation is subtle. An example of textual inversion is shown in Figure 4, and examples comparing with the previous methods are shown in Appendix D.2. The features of the original images can be learned and used to guide image generation. On the other hand, when applying our method, images are generated with very clear watermarks, which can be useful for protecting copyrighted images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.655, + 0.696 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Weights of the losses. This experiment aims to study the mechanism of our method. In equation 6, we design three losses to train the generator. As our target is to attack DMs, the adversarial loss \\(\\mathcal{L}_{adv}\\) is necessary, and we set its weight to 1. We then evaluate the effect of the GAN loss \\(\\mathcal{L}_{GAN}\\), the perturbation loss \\(\\mathcal{L}_{pert}\\), and the weight used to control the perturbation of the watermark region \\(w\\) by setting their weights \\(\\alpha\\), \\(\\beta\\), and \\(w\\) to 0 or their default value." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "The results are shown in Table 4. The first row is the result of only using \\(\\mathcal{L}_{adv}\\) and setting other losses to 0. In this case, although the attack ability for DMs is high, the generated adversarial examples are full of noise. By comparing the first and second rows, we observe that \\(\\mathcal{L}_{GAN}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.09, + 0.865, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.298, + 0.891, + 0.312 + ], + "angle": 0, + "content": "Figure 4. Examples of textual inversion on ImageNet. The watermarks used in the examples are IMAGENET CAT and IMAGENETDOG." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.336, + 0.468, + 0.446 + ], + "angle": 0, + "content": "
WeightsGenerated imageAdversarial example
αβwNCC ↑FID ↑PSNR ↑SSIM ↑
×××0.60565.313.10.12
✓××0.41350.622.00.65
×✓×0.35318.224.20.70
×✓✓0.34292.227.80.76
✓✓✓0.31245.530.10.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.469, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Table 4. The effect of the GAN loss \\((\\alpha)\\), the perturbation loss \\((\\beta)\\), and the weight of the watermark region \\((w)\\) on generating adversarial examples (WikiArt) under image-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.689 + ], + "angle": 0, + "content": "significantly improves the image quality of the adversarial example, even though the attack ability decreases. On the other hand, \\(\\mathcal{L}_{pert}\\) which aims to bound the magnitude of the perturbation, also has an effect similar to that of \\(\\mathcal{L}_{GAN}\\) (as seen when comparing first and third rows). In addition, adding weight term to \\(\\mathcal{L}_{pert}\\) further improves the quality of the adversarial examples without sacrificing too much attack ability (as seen when comparing third and fourth rows). Finally, we combine all these terms to achieve a balanced performance between image quality and attack ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Maximum perturbation bound. In the case of adversarial examples, there is always a trade-off between image quality and the success rate of the attack. To explore this trade-off, we conducted experiments with different bounds. We used PSNR to measure the image quality of the adversarial examples, and NCC to assess the visibility of the watermark in the generated image which indicates the attack success rate. Table 5 shows that as the bound increases, the watermark on the generated image becomes more distinct. However, this may also cause the watermark on the adversarial image to become more visible, which can reduce the image quality. In real-world applications, users can adjust the bounds to balance between achieving high-quality adversarial images and ensuring watermark visibility to protect their work." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.336, + 0.866, + 0.386 + ], + "angle": 0, + "content": "
Bound/25526101520
PSNR ↑40.134.230.124.217.5
NCC ↑0.130.220.310.370.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Table 5. The effect of using different perturbation bound. The experiment uses WikiArt under image-to-image generation." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.453, + 0.871, + 0.535 + ], + "angle": 0, + "content": "
SamplesNCC ↑FID ↑prec. ↓train time
10.12177.80.3220 sec
50.28235.30.041 min
100.30240.60.032 min
1000.31243.50.0220 min
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Table 6. The effect and training time of using different number of samples to train a generator for a specific watermark. The experiment is conducted under image-to-image generation on WikiArt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The number of samples for training. To generate adversarial examples that contain a specific watermark, we need to include samples with that watermark in the training data. In this experiment, we analyze how many samples we need for a watermark to be learned by the generator. We choose a different number of samples to train a generator for a specific watermark. Please note that different watermarks can be trained together within the same generator, and this is the default setting for our other experiments. The results of training one personal watermark are in Table 6. We observe there is no large performance difference between using 10 samples and 100 samples for training, which shows that our method only needs a small number of samples to train to embed a specific watermark. Therefore, one use case of our method is to create a personal watermark generator using a few samples within \\(2 - 3\\mathrm{min}\\), and then this generator can be used to embed the personal watermark for any of that user's creations to protect copyright." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24426" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.454, + 0.184 + ], + "angle": 0, + "content": "
DefenseNo attack (Origin)Attack (Ours)
NCC ↑FID ↑NCC ↑FID ↑
No defense0120.70.31245.5
JPEG0124.10.21202.2
RS0123.20.20195.6
TVM0132.50.16166.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.206, + 0.47, + 0.234 + ], + "angle": 0, + "content": "Table 7. The effects of applying defenses on adversarial examples under image-to-image generation." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.254, + 0.46, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.408, + 0.468, + 0.437 + ], + "angle": 0, + "content": "Figure 5. Examples of the generated image from original image, attack image, and attack image with defense." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.464, + 0.385, + 0.481 + ], + "angle": 0, + "content": "4.5. Robustness of adversarial examples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.469, + 0.578 + ], + "angle": 0, + "content": "To evaluate the robustness of the adversarial examples generated by our method, we apply several widely used adversarial defenses including JPEG compression [8], Randomized Smoothing (RS) [5], and Total Variance Minimization (TVM) [17]. Such defenses are directly processed on the adversarial examples to eliminate the perturbations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.805 + ], + "angle": 0, + "content": "The results of applying attacks and defenses are shown in Table 7. The \"No attack\" column means applying defense on the original image to generate images. There is no watermark being embedded so NCC almost equals 0, and applying defenses slightly increases FID since the quality of the original image decreases. On the other hand, the \"Attack\" column means we first generate the adversarial examples, apply defenses to them (except the first row), and then use them to generate new images. We observe that NCC and FID both decrease when adding defenses to our adversarial examples. However, compared with \"No attack\", the \"Attack\" column still has higher NCC and FID which means the defenses cannot completely eliminate the effect of our attack. In Figure 5, we show some examples under defenses, and the watermark is still visible for all cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.452, + 0.833 + ], + "angle": 0, + "content": "4.6. Transferability on Other Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we explore the transferability of our method on other generative models. This is also known as the black-box attack, where one target DM is used to train the generator and obtain the adversarial examples, and then these" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.094, + 0.892, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.285 + ], + "angle": 0, + "content": "Figure 6. Black-box attack to other models under image-to-image generation. The first row: generated images from the original image. Second row: generated images from our adversarial example." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.315, + 0.892, + 0.406 + ], + "angle": 0, + "content": "adversarial examples are directly used to attack other models. We generate adversarial examples on Stable Diffusion 1.5 (SD 1.5) and test on other models or tools including Dreamshaper \\(8^{2}\\), NovelAI, and Runway AI magic tools (image variation) under image-to-image generation. The details of the settings are described in the Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.893, + 0.605 + ], + "angle": 0, + "content": "The results are shown in Figure 6, and more examples can be found in Appendix D.3. Compared to images generated from the original image (first row), the images generated from our adversarial examples (second row) contain chaotic textures and a visible watermark for most cases. For the result of the Runway tool, although there is no obvious watermark on the generated image (as no parameter can be adjusted), our method still generates an image that differs significantly from the original one. Even though the watermarks are not as strong as they are in the white-box setting, our method still demonstrates good transferability. These results show that our method has the potential to prevent copyright violations from various image generative models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.625, + 0.618, + 0.64 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.833 + ], + "angle": 0, + "content": "In this work, we propose a novel method for copyright prevention against DMs. We build a generator to embed personal watermarks into the adversarial examples, and such examples can force DMs to generate images with visible watermarks and chaotic textures. We use a conditional GAN architecture with three carefully designed losses to optimize the generator. Experiments show our method performs well in various image generation scenarios and exhibits good transferability to attack other models. Additionally, our generation is significantly faster than those of previous methods. Therefore, this work provides a simple yet powerful way to protect image copyright against DMs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.85, + 0.814, + 0.863 + ], + "angle": 0, + "content": "\\(^{1}\\)https://huggingface.co/runwayml/stable-diffusion-v1-5" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.863, + 0.764, + 0.876 + ], + "angle": 0, + "content": "2https://civitai.com/models/4384/dreamshaper" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.876, + 0.626, + 0.887 + ], + "angle": 0, + "content": "3https://novelai.net/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.812, + 0.9 + ], + "angle": 0, + "content": "4https://runwayml.com/ai-magic-tools/image-to-image/" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.85, + 0.814, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24427" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.225 + ], + "angle": 0, + "content": "[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[4] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[5] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, and Chris Honsinger. Digital watermarking. Journal of Electronic Imaging, 11(3):414-414, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.576 + ], + "angle": 0, + "content": "[8] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] Volker Fischer, Mummadi Chaithanya Kumar, Jan Hendrik Metzen, and Thomas Brox. Adversarial examples for semantic image segmentation. arXiv preprint arXiv:1703.01101, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[17] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[21] Bernd Jähne. Digital image processing. Springer Science & Business Media, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[24] Jernej Kos, Ian Fischer, and Dawn Song. Adversarial examples for generative models. In 2018 IEEE security and privacy workshops (spw), pages 36-42. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[27] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, XUE Zhengui, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. 2023. 1, 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. arXiv preprint arXiv:1611.02770, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.176 + ], + "angle": 0, + "content": "[30] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[32] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[33] Saraju P Mohanty. Digital watermarking: A tutorial review. URL: http://www.csee.usf.edu/smohanty/research/Reports/WMSurvey1999Mohanty.pdf, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[35] KA Navas, Mathews Cheriyan Ajay, M Lekshmi, Tampy S Archana, and M Sasikumar. Dwt-dct-svd based watermarking. In 2008 3rd International Conference on Communication Systems Software and Middleware and Workshops (COMSWARE'08), pages 271-274. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[38] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[39] Phillip Pope, Yogesh Balaji, and Soheil Feizi. Adversarial robustness of flow-based generative models. In International Conference on Artificial Intelligence and Statistics, pages 3795-3805. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[40] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[44] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.302 + ], + "angle": 0, + "content": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[47] Pedro Tabacof, Julia Tavares, and Eduardo Valle. Adversarial images for variational autoencoders. arXiv preprint arXiv:1612.00155, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[48] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.518 + ], + "angle": 0, + "content": "[49] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.893, + 0.589 + ], + "angle": 0, + "content": "[50] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[51] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.893, + 0.702 + ], + "angle": 0, + "content": "[52] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610, 2018. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[53] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[54] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24429" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "methods and applications. ACM Computing Surveys, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[56] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[57] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[58] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[60] Ru Zhang, Shiqi Dong, and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia tools and applications, 78:8559-8575, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[61] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[62] Yuan Zhao, Bo Liu, Ming Ding, Baoping Liu, Tianqing Zhu, and Xin Yu. Proactive deepfake defence via identity watermarking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4602-4611, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[63] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.47, + 0.74 + ], + "angle": 0, + "content": "[65] Peifei Zhu, Genki Osada, Hirokatsu Kataoka, and Tsubasa Takahashi. Frequency-aware gan for adversarial manipulation generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4315-4324, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24430" + } + ] +] \ No newline at end of file diff --git a/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_origin.pdf b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e13fbde3f9b45c61564ccb67db818202d6f76450 --- /dev/null +++ b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/9f5ff66f-0862-4163-9fbc-9a609665d28d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b48d182333af32bfb48ccdc435bfd6a21ea36249c55cf38c9d7bd3cdc9178f9 +size 8098464 diff --git a/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/full.md b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a124e6f8dddea9e0c7ec8e33fab91f6c9d0b011 --- /dev/null +++ b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/full.md @@ -0,0 +1,306 @@ +# Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models + +Peifei Zhu, Tsubasa Takahashi, Hirokatsu Kataoka +LY Corporation + +{peifei.zhu, tsubasa.takahashi, jpz4219,}@lycorp.co.jp + +# Abstract + +Diffusion Models (DMs) have shown remarkable capabilities in various image-generation tasks. However, there are growing concerns that DMs could be used to imitate unauthorized creations and thus raise copyright issues. To address this issue, we propose a novel framework that embeds personal watermarks in the generation of adversarial examples. Such examples can force DMs to generate images with visible watermarks and prevent DMs from imitating unauthorized images. We construct a generator based on conditional adversarial networks and design three losses (adversarial loss, GAN loss, and perturbation loss) to generate adversarial examples that have subtle perturbation but can effectively attack DMs to prevent copyright violations. Training a generator for a personal watermark by our method only requires 5-10 samples within 2-3 minutes, and once the generator is trained, it can generate adversarial examples with that watermark significantly fast (0.2s per image). We conduct extensive experiments in various conditional image-generation scenarios. Compared to existing methods that generate images with chaotic textures, our method adds visible watermarks on the generated images, which is a more straightforward way to indicate copyright violations. We also observe that our adversarial examples exhibit good transferability across unknown generative models. Therefore, this work provides a simple yet powerful way to protect copyright from DM-based imitation. + +# 1. Introduction + +Diffusion models (DMs) [19, 42, 46] have garnered significant attention in both the computer vision industry and academia. Compared to previous generative models such as generative adversarial networks (GANs) [15, 23, 32], DMs have demonstrated a remarkable ability to generate realistic images with higher resolution and diversity. Despite the positive aspects of DMs, there are public concerns about potential copyright violations when unauthorized images are + +![](images/7428e315389c5e19d3151d63d1b87971882ce6cbe95f36d2b569ef564ab4b89a.jpg) + +![](images/933c572476350cfb0bcda88ebf9012a425dbc8eeebfbf9b97adc338b724b50a6.jpg) + +![](images/52aa3d2ff349e20f34806b12e0a16c0178e34baa24b09c1578ac72a4c70f4ae9.jpg) +Figure 1. Copyright issues of DMs and adversarial example-based methods for copyright protection. Without protection, DMs can easily imitate original images under different image generation scenarios. A previous method (AdvDM [28]) generates adversarial examples by optimization against DMs to prevent DMs from extracting the feature of the original images, resulting in the generation of chaotic images. Our method goes one step further to build a generator that embeds personal watermarks into the generation of adversarial images. Such examples force DMs to generate images with visible watermarks for tracing copyright. Our method is fast, simple yet powerful in protecting copyrights against DMs. + +used to train DMs or to generate new creations by DMs [55]. For example, an entity could use copyrighted paintings shared on the Internet to generate images with a similar style and appearance, potentially earning illegal revenue. Therefore, it is necessary to develop techniques to prevent the imitation of unauthorized creations by DMs. + +Several previous works have been proposed to address these concerns. Watermarking [2, 33] is a common solution that embeds an invisible message into the image, which can then be extracted to identify the existence of copyright. + +Recent techniques typically integrate watermarking into the generative process of DMs [7, 12, 38]. However, these methods have several limitations. First, they are primarily designed to protect the copyright of DMs or to distinguish generated images from natural ones, which differs from our goal of protecting the copyright of human creations. Second, embedding watermarks into the generative process requires re-training or at least fine-tuning DMs, and a post-process to extract the watermark is needed to identify it, which can be costly. On the other hand, methods that use adversarial examples for DMs to protect copyright have been proposed recently [27, 28, 50]. The idea is to generate adversarial examples to hinder DMs from extracting the features of the original images. This generation process does not require any re-training of DMs or any post-processing for the generated image. However, there are still several limitations. First, even though these adversarial examples could add chaotic textures to the generated images, such changes are difficult to comprehend. For instance, users may wonder if the problem originates from the DM itself rather than considering the copyright issue. Second, the original image's copyright is not traceable since no copyright-related information is embedded. Third, each adversarial example needs to be optimized separately against DMs, which makes the generation time-consuming. + +In this work, we propose a simple yet powerful method to protect copyrighted images from imitation by DMs. Different from previous adversarial example-based methods, we go one step further by embedding personal watermarks into the generation of adversarial examples for copyright tracing. Such examples could force DMs to generate images with visible watermarks as well as chaotic textures. In addition, instead of using iterative optimization against DMs to generate examples, we train a generator beforehand using only a few samples. Once the generator is trained, it can be used to generate adversarial examples significantly fast. As shown in Figure 1, instead of posting the original image on the Internet, posting adversarial examples with only subtle changes could prevent DMs from imitation. Compared to previous methods, our method provides a more straightforward way to warn of potential copyright violations. + +Our contributions can be summarized as follows: + +- We design a novel framework that embeds personal watermarks into the generation of adversarial examples to prevent copyright violations caused by DM-based imitation. Our adversarial examples can force DMs to generate images with visible watermarks for tracing copyright. +- We train a generator based on a conditional GAN architecture to generate adversarial examples (Figure 2). We design three losses: adversarial loss, GAN loss, and weighted perturbation loss. These losses aim to improve the attack ability of the adversarial examples while making the perturbation invisible to the human eye. + +- We conduct extensive experiments in various image generation scenarios, along with robustness evaluations against defenses and assessments of transferability to other models. The experiments demonstrate that our adversarial examples can prevent unauthorized images from being learned and can generate visible watermarks for copyright tracing. Our generation process is significantly fast (0.2s per image), and the generated examples also exhibit good transferability across other generative models. + +# 2. Related Works + +# 2.1. Generative Diffusion Models + +In recent years, DMs have made significant advancements in various fields [9, 37, 45, 55]. DMs are a type of generative model that simulates a diffusion process to generate new data samples. DMs consist of a forward process and a reverse process. The forward process involves gradually adding Gaussian noise to the data over a series of time steps until it becomes completely random. The reverse process gradually removes noise through a series of learned denoising steps, reconstructing the data back to its original form or generating new samples from the same distribution. + +DMs have been applied to various generation tasks such as image synthesis [9, 43], text-to-image generation [1, 41, 58, 59], and text-guided image editing [4, 34, 54, 61]. Among these, latent diffusion models [42], which perform the diffusion steps in the latent image space, have demonstrated a remarkable ability to generate high-resolution and realistic images. Despite the impressive performance, some concerns that generating new creations by DMs based on unauthorized images could lead to copyright issues. + +# 2.2. Image Watermarking + +Image watermarking is a technique that embeds a piece of information into an image to protect the image from copyright infringement. Traditional watermarking methods typically involve altering pixel values [6] or manipulating the frequency domain of an image [35] to embed watermarks. More recently, deep-learning-based methods have been developed. A common approach is using encoder-extractor networks, where an encoder embeds the watermark into the image, and an extractor is trained to retrieve the watermark from the marked image [49, 60, 63]. + +In terms of watermarking for generative models, most existing works focus on protecting the copyright of generative models [11, 30, 38, 51, 57] or distinguishing generative images from natural images [7, 12, 56, 62]. These methods involve reformulating the training process of DMs to inject watermarks into the images, as well as developing a decoder to identify the embedded messages from the generated images. However, the re-training can be costly and it is impractical to check every generated image with a decoder. + +# 2.3. Adversarial Examples for Generative models + +Adversarial examples are created by adding small but intentional perturbations to input samples such that the perturbed inputs cause the model to produce incorrect answers. Various adversarial attack methods have been proposed to confuse classification models [10, 16, 31, 52] or segmentation models [13, 53, 65]. There are also several existing works studying adversarial examples for generative models such as GANs [24], variational autoencoders (VAEs) [47], and flow-based generative models [39]. + +Recently, methods that generate adversarial examples for DMs [27, 28, 50] have been proposed. These methods maximize the diffusion loss to obtain a perturbation, and adding such a perturbation to the original image enables DMs to generate images with chaotic content. However, these methods have several drawbacks: 1) such chaotic content is not straightforward to indicate copyright violation, 2) the copyright of the original image is not traceable, and 3) optimization against DMs is time-consuming. We propose a simple yet powerful method that directly incorporates personal watermarks into the generation of adversarial examples, and we build a generator that can generate adversarial examples significantly faster than existing methods. + +# 3. Proposed Method + +# 3.1. Problem Statement + +Given an original image $x$ , we have a diffusion model $\theta$ that can learn the data distribution of $x$ and generate images with similar style and appearance. In this work, we aim to generate an adversarial example $x'$ that can prevent $\theta$ from extracting the feature of $x$ and can directly show copyright information on the generated image. For example, for imaged-to-image generation, the generated image from $x'$ by using $\theta$ can be denoted as $\theta(x')$ . We design a way to provide the copyright information by using a watermark that contains the ownership details (e.g. artist name) of the original image. Therefore, given a watermark image $m$ , the adversarial example $x'$ that embeds $m$ can be calculated by + +$$ +\begin{array}{l} x ^ {\prime} := \underset {x ^ {\prime}} {\operatorname {a r g m i n}} \| \theta \left(x ^ {\prime}\right) - m \|, \tag {1} \\ \begin{array}{l} \text {s . t .} \quad \| x - x ^ {\prime} \| \leq \sigma , \end{array} \\ \end{array} +$$ + +where $\sigma$ is a constant to control the range of the perturbation. This is a targeted attack setting where we compel the image generated by DMs to be closer to the watermark, while ensuring that the adversarial example remains as close to the original one as possible. + +However, several issues are associated with generating adversarial examples using Equation 1. First, directly optimizing $x'$ against DMs for each $x$ could be time-consuming, so we need to find a way to speed up the generation process. Second, since there are several image-generation sce + +narios related to imitating the original image, such as text-guided image-to-image generation [4], text-to-image generation under textual inversion [14] or DreamBooth [43], we aim to generate an adversarial example $x'$ that can be applicable to all these scenarios. Meanwhile, we also expect that $x'$ could have good transferability on other generative models. Third, since our target attack aims to add personal watermarks to the generated images, the perturbation naturally becomes larger in the watermark regions. We need a solution to make the perturbation invisible to human eyes. + +To address the first issue, we train a generator beforehand instead of implementing optimization every time. To address the second and third issues, we design an adversarial loss, a GAN loss, and a perturbation loss to improve the attack ability across various tasks and models while keeping the perturbation invisible. + +# 3.2. Architecture Overview + +We propose a conditional GAN architecture for generating watermark-embedded adversarial examples. The architecture of our method contains three components: a generator $G$ , a discriminator $D$ , and a target DM $\theta$ , shown in Figure 2. The generator $G$ has an encoder to extract features from the inputs and a decoder to generate perturbation using the features. The inputs of $G$ are the original image $x$ and the watermark $m$ , and the perturbation is generated conditioned on $m$ . We perform the conditioning by concatenating $x$ and $m$ along the channel dimension and feeding them into $G$ . The adversarial example $x'$ can be obtained by adding $x$ and the generated perturbation. The discriminator $D$ is a classifier that distinguishes $x'$ from $x$ , and the target DM $\theta$ is always kept frozen. + +We design three losses to optimize $G$ and $D$ . 1) A GAN loss aims to compel $x'$ to be closer to $x$ . By using the GAN loss, we can train $G$ to produce images that $D$ cannot distinguish from input images. 2) A perturbation loss is designed to bound the magnitude of the perturbation and make it invisible. 3) An adversarial loss is used for $G$ to generate perturbation that can attack DMs under various scenarios. The details are described in Section 3.3. Once the generator $G$ is optimized, it can be used directly to generate adversarial examples significantly fast. + +# 3.3. Loss functions + +GAN loss. $\mathcal{L}_{GAN}$ is used to quantify the difference between the adversarial example and the original image. For a set of images, $\mathcal{L}_{GAN}$ can be written as: + +$$ +\mathcal {L} _ {G A N} = \mathbb {E} _ {x} \log D (x) + \mathbb {E} _ {x} \log \left\{1 - D \left(x ^ {\prime}\right) \right\}. \tag {2} +$$ + +The adversarial image $x'$ is generated under the condition of the watermark $m$ , which can be written as: + +$$ +x ^ {\prime} = x + G (x \mid m). \tag {3} +$$ + +![](images/7e67fe78b717252a06df7e6f67ab1b44dea6a6340d416f4aea703959dfb42086.jpg) +Figure 2. Architecture overview. $G$ generates perturbation for $x$ conditioned on $m$ . $D$ and $G$ produce $\mathcal{L}_{GAN}$ to compel $x'$ to be closer to $x$ . $\mathcal{L}_{adv}$ aims to force the image generated by $\theta$ to display a visible watermark $m$ , and $\mathcal{L}_{pert}$ further bounds the magnitude of the perturbation. + +Perturbation loss. $\mathcal{L}_{pert}$ aims to limit the magnitude of the perturbation. We design this loss based on the soft hinge loss used in previous adversarial attacks [3, 29, 52]. However, the soft hinge loss assigns equal importance to all perturbations within the image. In our case, the perturbation in the watermark region will be larger than that in other regions, making the watermark visible even in the adversarial example. To better conceal the watermark as well as bound the perturbation in the whole image, we design a weighted perturbation loss, where we assign larger weights $w$ to the watermark region. This encourages the perturbation in the watermark region to be smaller and invisible. For a set of images, this loss can be expressed as: + +$$ +\mathcal {L} _ {p e r t} = \mathbb {E} _ {x} \max (0, \| G (x \mid m) (1 + w \cdot m) \| _ {2} - c), \tag {4} +$$ + +where $c$ is a constant denoting a user-specified bound. + +Adversarial loss. $\mathcal{L}_{adv}$ is used to generate a perturbation that can attack DMs. As we expect our adversarial examples to have transferability across different image generation scenarios and various generative models, we design the loss based on the latent representation of latent diffusion models (LDM). We choose LDM for two reasons. First, LDM has outperformed other models in both high-quality image generation and sampling efficiency, and it has been used in various generative models. Generating adversarial examples against LDM naturally increases their transferability across various models. Second, LDM maps an image $x$ to a latent representation $\varepsilon(x)$ which is the output of the Variational Auto-Encoder (VAE). This representation is widely used in conditional image generation scenarios, and it provides a simple way to control the image generation of DMs. Therefore, instead of directly exploiting the generated image, we minimize the distance between the representation of the adversarial example $\varepsilon(x')$ and that of the watermark $\varepsilon(m)$ , formulated as: + +$$ +\mathcal {L} _ {a d v} = \mathbb {E} _ {x} \| \varepsilon \left(x ^ {\prime}\right) - \varepsilon (m) \| _ {2}. \tag {5} +$$ + +Optimization. Finally, we combine the above three losses to formulate the objective function for training, written as: + +$$ +\min _ {G} \max _ {D} V (D, G) = \mathcal {L} _ {a d v} + \alpha \mathcal {L} _ {G A N} + \beta \mathcal {L} _ {p e r t}, \tag {6} +$$ + +where $\alpha$ and $\beta$ are weights to control the balance of the three losses. By playing a minimax game between the $G$ and $D$ , the optimal parameters of the model can be obtained. + +# 3.4. Image Generation under Different Settings + +Since our target is to prevent copyright violations, we focus on conditional image generation scenarios where DMs sample the distribution of the original images to generate new ones. Such scenarios include text-guided image-to-image generation [36] and text-to-image generation under textual inversion [14] described in this section, and other scenarios including DreamBooth [43], LoRA [20], and Custom Diffusion [25] discussed in Appendix A. + +Text-guide image-to-image generation. In this setting, we can pass a text prompt and an input image to condition the image generation. Our method can be directly applied to this setting, as shown in Figure 2. During the training of the generator and discriminator, we simply use the same prompt (e.g. "A painting") for all samples. For the inference, any prompt can be used for generating new images. + +Textual inversion. Textual inversion is a method for capturing new concepts from a small number of images. These new concepts can then be used to control image generation in text-to-image settings. The steps to apply our method in this setting are as follows. First, given a small set of images within a similar category (e.g. same artist), we extract a text $S_{*}$ to represent our new concept. We then replace the vector associated with the tokenized string with an optimized embedding $v_{*}$ . The optimization can be applied by minimizing the LDM loss over the images from the small set. During the image generation by DMs, the text $S_{*}$ is used as a condition to generate new images. We conduct this generation + +for both original images and adversarial images. The adversarial images are generated through the image-to-image generation setting as shown in Figure 2. + +# 4. Experiments + +In this section, we evaluate our method for generating adversarial examples for DMs. We conduct experiments under text-guided image-to-image generation and textual inversion, performing various analyses to assess the effectiveness, robustness, and transferability of our method. We perform the evaluation on WikiArt [48] and ImageNet [44] datasets. All experiments are run on an NVIDIA A100 80GB GPU. Training the generator using 100 samples with 200 epochs takes about 20 minutes. + +# 4.1. Experimental Settings + +Datasets. The WikiArt dataset contains 50k paintings from 195 artists. We randomly select 3000 paintings from 50 artists. We create different watermarks for different artists. These watermarks are binary images with a black background and white artist names such as "VAN_GOGH", "CLAUDE_MONNET". Since only a small number of samples are needed to train our generator, we select 10 images from each artist (500 images in total) for training and the remaining 2500 images are for evaluation. For the ImageNet dataset, we randomly select 2000 images (100 for training, 1900 for evaluation) from the goldfish, tiger shark, peacock, goose, Eskimo dog, and tabby cat category. The watermarks are designed as "IMAGENET CAT", "IMAGENET FISH", "IMAGENETDOG", etc. + +Evaluation metrics. We evaluate our method from two perspectives. For the adversarial example, we use MSE, PSNR, and SSIM to measure the image quality and the difference from the original image. For the image generated by DMs, we use Fréchet Inception Distance (FID) [18] and precision (prec.) [26] to measure the similarity between the generated and original images. Since we expect DMs to generate images with visible watermarks, we use Normalized Cross-Correlation (NCC) [21] to measure the similarity between the generated image and the watermark. Please note that, unlike traditional watermark methods, we do not extract the watermark from the generated image, and thus NCC is smaller in our case. However, to reduce the influence of the background, we first calculate a difference image between the image generated from the original image and that from the adversarial example, and then use this difference image and the watermark to obtain NCC. + +Implementation details. For the generator $G$ , we adopt an architecture similar to that in [22]. For the discriminator $D$ , we use an architecture similar to the discriminator in AdvGAN [52]. Further details regarding the implementation of our model are provided in Appendix B. We set the weight of GAN loss $\alpha = 1.0$ , the weight of perturbation + +
MethodNCC ↑FID ↑prec. ↓recall
Original0120.70.850.28
AdvDM0228.20.030.25
Mist0.09205.80.040.33
Ours0.31245.50.030.29
+ +Table 1. The performance of the generated image under text-guided image-to-image generation on WikiArt. Recall measures the diversity of the images and is only listed as a reference. + +loss $\beta = 10$ , the weight for watermark region $w = 4$ , and the bound $c = 10 / 255$ . During the model training, we set the batch size to 8 and the learning rate to 0.001. For comparison, we use AdvDM [28] and Mist (fuse mode) [27], which also use adversarial examples to protect copyright. We set their sampling step to 100 and the maximum perturbation to $10 / 255$ . + +# 4.2. Text-guided Image-to-Image Generation + +We evaluate the performance of our method under the setting described in Section 3.4. In this setting, a strength parameter is used to control the level of noise added to the original image during the generation of new images. A strength value close to 0 produces an image nearly identical to the original, while a value close to 1 results in an image that significantly differs from the original. We set the strength to 0.3 to simulate realistic imitation, and additional results with different strength values can be found in the Appendix C.1. For the text prompt, we uniformly use "A painting" for all images for convenience, and experiments using varied prompts are also included in the Appendix C.2. + +We first show the evaluation result of the image generated by LDM based on either the original image or the adversarial example. A high FID and a low precision mean that the generated image is far from the original one, which further indicates that LDM fails to capture the feature of the original image. As shown in Table 1, our adversarial examples largely increase FID and decrease precision. We also calculate NCC between the generated image and the watermark. For Mist, since it is possible to input a target image into the generation, we apply our idea to use different watermarks as the target images and use the same method to calculate NCC. Results show that our method has a significantly higher NCC, which means our adversarial examples can force LDM to generate images with more visible watermarks. These results suggest our method has great potential to prevent copyright violations caused by LDM imitation. + +We also evaluate the quality of the adversarial example generated by different methods. As shown in Table 2, the image quality of our adversarial examples is slightly higher than that of the existing methods. Moreover, the generation is significantly faster than the existing methods since we + +
MethodMSE ↓PSNR ↑SSIM ↑Run time
AdvDM0.003829.10.8032s
Mist0.004029.00.8135s
Ours0.003730.10.800.2s
+ +![](images/52edcfdfa9c73d131df21cba66b5313fcd596772c880e84fbe1f8450e1783599.jpg) +Figure 3. Comparison of different methods under text-guided image-to-image generation. Source is the image input to the LDM, and generated image is the output of the LDM (strength 0.3). The watermarks of our method are designed according to the artist name (from left to right: REMBRANDT, AUGUSTE_REN). + +directly use the trained generator instead of iterative optimization against DMs. Our method has an advantage when generating a large number of examples. + +For qualitative evaluation, we visualize two examples generated by different methods in Figure 3. More examples are shown in Appendix D.1. For the adversarial examples, all methods seem to introduce only subtle perturbations that are almost invisible to human eyes. For the generated images based on the adversarial examples, the existing methods succeed in adding chaotic textures to the generated image. However, such changes can be difficult to understand and evaluate. On the other hand, our adversarial examples succeed in generating images with visible watermarks as well as chaotic textures. Our method provides a more straightforward way to indicate copyright violations. + +Table 2. The image quality and generation time (per image) of the adversarial examples on WikiArt. + +
MethodGenerated imageAdversarial example
NCC↑FID ↑PSNR ↑SSIM ↑
Original0120.7--
AdvDM0327.326.30.73
Mist0.15318.426.70.74
Ours0.40412.927.10.75
+ +Table 3. The performance of the generated image and adversarial examples on ImageNet under textual inversion. + +# 4.3. Textual Inversion + +Textual inversion is another important scenario related to copyright violations. In this scenario, given a small number of images, we first learn to represent the concept of the images in a new word $S_{*}$ , then use $S_{*}$ to guide image generation. Following the method of textual inversion [14], we separate our evaluation images into 4-image groups for each artist or ImageNet category. For each 4-image group, we set the maximum steps to optimize $S_{*}$ to 5000. After the optimization, we generate 10 images for each $S_{*}$ under 10 prompt templates. This process is implemented for both original images and our adversarial examples. + +For evaluation, we compare our method with existing methods, and the results are shown in Table 3. For the generated image, our method has higher NCC and FID, which indicates our method successfully injects the watermark information in the generation of the adversarial example. In addition, our adversarial example also has higher PSNR and SSIM, which means the generated perturbation is subtle. An example of textual inversion is shown in Figure 4, and examples comparing with the previous methods are shown in Appendix D.2. The features of the original images can be learned and used to guide image generation. On the other hand, when applying our method, images are generated with very clear watermarks, which can be useful for protecting copyrighted images. + +# 4.4. Ablation Study + +Weights of the losses. This experiment aims to study the mechanism of our method. In equation 6, we design three losses to train the generator. As our target is to attack DMs, the adversarial loss $\mathcal{L}_{adv}$ is necessary, and we set its weight to 1. We then evaluate the effect of the GAN loss $\mathcal{L}_{GAN}$ , the perturbation loss $\mathcal{L}_{pert}$ , and the weight used to control the perturbation of the watermark region $w$ by setting their weights $\alpha$ , $\beta$ , and $w$ to 0 or their default value. + +The results are shown in Table 4. The first row is the result of only using $\mathcal{L}_{adv}$ and setting other losses to 0. In this case, although the attack ability for DMs is high, the generated adversarial examples are full of noise. By comparing the first and second rows, we observe that $\mathcal{L}_{GAN}$ + +![](images/68349c4fdd95264dd252bb6227a6140ed0173626065bbf72428ae736cc7a4839.jpg) +Figure 4. Examples of textual inversion on ImageNet. The watermarks used in the examples are IMAGENET CAT and IMAGENETDOG. + +
WeightsGenerated imageAdversarial example
αβwNCC ↑FID ↑PSNR ↑SSIM ↑
×××0.60565.313.10.12
✓××0.41350.622.00.65
×✓×0.35318.224.20.70
×✓✓0.34292.227.80.76
✓✓✓0.31245.530.10.80
+ +significantly improves the image quality of the adversarial example, even though the attack ability decreases. On the other hand, $\mathcal{L}_{pert}$ which aims to bound the magnitude of the perturbation, also has an effect similar to that of $\mathcal{L}_{GAN}$ (as seen when comparing first and third rows). In addition, adding weight term to $\mathcal{L}_{pert}$ further improves the quality of the adversarial examples without sacrificing too much attack ability (as seen when comparing third and fourth rows). Finally, we combine all these terms to achieve a balanced performance between image quality and attack ability. + +Maximum perturbation bound. In the case of adversarial examples, there is always a trade-off between image quality and the success rate of the attack. To explore this trade-off, we conducted experiments with different bounds. We used PSNR to measure the image quality of the adversarial examples, and NCC to assess the visibility of the watermark in the generated image which indicates the attack success rate. Table 5 shows that as the bound increases, the watermark on the generated image becomes more distinct. However, this may also cause the watermark on the adversarial image to become more visible, which can reduce the image quality. In real-world applications, users can adjust the bounds to balance between achieving high-quality adversarial images and ensuring watermark visibility to protect their work. + +Table 4. The effect of the GAN loss $(\alpha)$ , the perturbation loss $(\beta)$ , and the weight of the watermark region $(w)$ on generating adversarial examples (WikiArt) under image-to-image generation. + +
Bound/25526101520
PSNR ↑40.134.230.124.217.5
NCC ↑0.130.220.310.370.41
+ +Table 5. The effect of using different perturbation bound. The experiment uses WikiArt under image-to-image generation. + +
SamplesNCC ↑FID ↑prec. ↓train time
10.12177.80.3220 sec
50.28235.30.041 min
100.30240.60.032 min
1000.31243.50.0220 min
+ +Table 6. The effect and training time of using different number of samples to train a generator for a specific watermark. The experiment is conducted under image-to-image generation on WikiArt. + +The number of samples for training. To generate adversarial examples that contain a specific watermark, we need to include samples with that watermark in the training data. In this experiment, we analyze how many samples we need for a watermark to be learned by the generator. We choose a different number of samples to train a generator for a specific watermark. Please note that different watermarks can be trained together within the same generator, and this is the default setting for our other experiments. The results of training one personal watermark are in Table 6. We observe there is no large performance difference between using 10 samples and 100 samples for training, which shows that our method only needs a small number of samples to train to embed a specific watermark. Therefore, one use case of our method is to create a personal watermark generator using a few samples within $2 - 3\mathrm{min}$ , and then this generator can be used to embed the personal watermark for any of that user's creations to protect copyright. + +
DefenseNo attack (Origin)Attack (Ours)
NCC ↑FID ↑NCC ↑FID ↑
No defense0120.70.31245.5
JPEG0124.10.21202.2
RS0123.20.20195.6
TVM0132.50.16166.8
+ +Table 7. The effects of applying defenses on adversarial examples under image-to-image generation. + +![](images/262dcc52c79a31b0c9c2483d4d4b297eefc1d993eb535d18884507cf1bf7ed7e.jpg) +Figure 5. Examples of the generated image from original image, attack image, and attack image with defense. + +# 4.5. Robustness of adversarial examples + +To evaluate the robustness of the adversarial examples generated by our method, we apply several widely used adversarial defenses including JPEG compression [8], Randomized Smoothing (RS) [5], and Total Variance Minimization (TVM) [17]. Such defenses are directly processed on the adversarial examples to eliminate the perturbations. + +The results of applying attacks and defenses are shown in Table 7. The "No attack" column means applying defense on the original image to generate images. There is no watermark being embedded so NCC almost equals 0, and applying defenses slightly increases FID since the quality of the original image decreases. On the other hand, the "Attack" column means we first generate the adversarial examples, apply defenses to them (except the first row), and then use them to generate new images. We observe that NCC and FID both decrease when adding defenses to our adversarial examples. However, compared with "No attack", the "Attack" column still has higher NCC and FID which means the defenses cannot completely eliminate the effect of our attack. In Figure 5, we show some examples under defenses, and the watermark is still visible for all cases. + +# 4.6. Transferability on Other Generative Models + +In this section, we explore the transferability of our method on other generative models. This is also known as the black-box attack, where one target DM is used to train the generator and obtain the adversarial examples, and then these + +![](images/ca6ab38915a1c85aa02549dbffa4013df5390a5f4809492121a5df2dc13ab456.jpg) +Figure 6. Black-box attack to other models under image-to-image generation. The first row: generated images from the original image. Second row: generated images from our adversarial example. + +adversarial examples are directly used to attack other models. We generate adversarial examples on Stable Diffusion 1.5 (SD 1.5) and test on other models or tools including Dreamshaper $8^{2}$ , NovelAI, and Runway AI magic tools (image variation) under image-to-image generation. The details of the settings are described in the Appendix E. + +The results are shown in Figure 6, and more examples can be found in Appendix D.3. Compared to images generated from the original image (first row), the images generated from our adversarial examples (second row) contain chaotic textures and a visible watermark for most cases. For the result of the Runway tool, although there is no obvious watermark on the generated image (as no parameter can be adjusted), our method still generates an image that differs significantly from the original one. Even though the watermarks are not as strong as they are in the white-box setting, our method still demonstrates good transferability. These results show that our method has the potential to prevent copyright violations from various image generative models. + +# 5. Conclusion + +In this work, we propose a novel method for copyright prevention against DMs. We build a generator to embed personal watermarks into the adversarial examples, and such examples can force DMs to generate images with visible watermarks and chaotic textures. We use a conditional GAN architecture with three carefully designed losses to optimize the generator. Experiments show our method performs well in various image generation scenarios and exhibits good transferability to attack other models. Additionally, our generation is significantly faster than those of previous methods. Therefore, this work provides a simple yet powerful way to protect image copyright against DMs. + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 1 +[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 4 +[4] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3, 1 +[5] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019. 8 +[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, and Chris Honsinger. Digital watermarking. Journal of Electronic Imaging, 11(3):414-414, 2002. 2 +[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2 +[8] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 8 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3 +[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2 +[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2 +[13] Volker Fischer, Mummadi Chaithanya Kumar, Jan Hendrik Metzen, and Thomas Brox. Adversarial examples for semantic image segmentation. arXiv preprint arXiv:1703.01101, 2017. 3 +[14] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen + +Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3, 4, 6 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3 +[17] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2017. 8 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 1 +[21] Bernd Jähne. Digital image processing. Springer Science & Business Media, 2005. 5 +[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 1 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1 +[24] Jernej Kos, Ian Fischer, and Dawn Song. Adversarial examples for generative models. In 2018 IEEE security and privacy workshops (spw), pages 36-42. IEEE, 2018. 3 +[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4 +[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5 +[27] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2, 3, 5 +[28] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, XUE Zhengui, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. 2023. 1, 2, 3, 5 + +[29] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. arXiv preprint arXiv:1611.02770, 2016. 4 +[30] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2 +[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3 +[32] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1 +[33] Saraju P Mohanty. Digital watermarking: A tutorial review. URL: http://www.csee.usf.edu/smohanty/research/Reports/WMSurvey1999Mohanty.pdf, 1999. +[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2 +[35] KA Navas, Mathews Cheriyan Ajay, M Lekshmi, Tampy S Archana, and M Sasikumar. Dwt-dct-svd based watermarking. In 2008 3rd International Conference on Communication Systems Software and Middleware and Workshops (COMSWARE'08), pages 271-274. IEEE, 2008. 2 +[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 4 +[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[38] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2 +[39] Phillip Pope, Yogesh Balaji, and Soheil Feizi. Adversarial robustness of flow-based generative models. In International Conference on Artificial Intelligence and Statistics, pages 3795-3805. PMLR, 2020. 3 +[40] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2 +[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2 +[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 1 +[44] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5 +[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1 +[47] Pedro Tabacof, Julia Tavares, and Eduardo Valle. Adversarial images for variational autoencoders. arXiv preprint arXiv:1612.00155, 2016. 3 +[48] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5 +[49] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 2 +[50] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3 +[51] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2 +[52] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610, 2018. 3, 4, 5 +[53] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3 +[54] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2 +[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of + +methods and applications. ACM Computing Surveys, 2022. 1, 2 +[56] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2 +[57] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2 +[58] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909, 2023. 2 +[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[60] Ru Zhang, Shiqi Dong, and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia tools and applications, 78:8559-8575, 2019. 2 +[61] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2 +[62] Yuan Zhao, Bo Liu, Ming Ding, Baoping Liu, Tianqing Zhu, and Xin Yu. Proactive deepfake defence via identity watermarking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4602-4611, 2023. 2 +[63] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 2 +[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1 +[65] Peifei Zhu, Genki Osada, Hirokatsu Kataoka, and Tsubasa Takahashi. Frequency-aware gan for adversarial manipulation generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4315-4324, 2023. 3 \ No newline at end of file diff --git a/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/images.zip b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d7b77b26c47b2dcb63b77ac494d92608550cf25 --- /dev/null +++ b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5c6771efa86296f853f10e65b7edfdf708221efeac859a8293178cb9e732954 +size 561562 diff --git a/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/layout.json b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7575a665b5b85d5d444e3b5627000f7f7bb5e09 --- /dev/null +++ b/2024/Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models/layout.json @@ -0,0 +1,8991 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "text", + "content": "Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 161, + 419, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 161, + 419, + 190 + ], + "spans": [ + { + "bbox": [ + 173, + 161, + 419, + 190 + ], + "type": "text", + "content": "Peifei Zhu, Tsubasa Takahashi, Hirokatsu Kataoka \nLY Corporation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 191, + 445, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 191, + 445, + 205 + ], + "spans": [ + { + "bbox": [ + 148, + 191, + 445, + 205 + ], + "type": "text", + "content": "{peifei.zhu, tsubasa.takahashi, jpz4219,}@lycorp.co.jp" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 569 + ], + "type": "text", + "content": "Diffusion Models (DMs) have shown remarkable capabilities in various image-generation tasks. However, there are growing concerns that DMs could be used to imitate unauthorized creations and thus raise copyright issues. To address this issue, we propose a novel framework that embeds personal watermarks in the generation of adversarial examples. Such examples can force DMs to generate images with visible watermarks and prevent DMs from imitating unauthorized images. We construct a generator based on conditional adversarial networks and design three losses (adversarial loss, GAN loss, and perturbation loss) to generate adversarial examples that have subtle perturbation but can effectively attack DMs to prevent copyright violations. Training a generator for a personal watermark by our method only requires 5-10 samples within 2-3 minutes, and once the generator is trained, it can generate adversarial examples with that watermark significantly fast (0.2s per image). We conduct extensive experiments in various conditional image-generation scenarios. Compared to existing methods that generate images with chaotic textures, our method adds visible watermarks on the generated images, which is a more straightforward way to indicate copyright violations. We also observe that our adversarial examples exhibit good transferability across unknown generative models. Therefore, this work provides a simple yet powerful way to protect copyright from DM-based imitation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Diffusion models (DMs) [19, 42, 46] have garnered significant attention in both the computer vision industry and academia. Compared to previous generative models such as generative adversarial networks (GANs) [15, 23, 32], DMs have demonstrated a remarkable ability to generate realistic images with higher resolution and diversity. Despite the positive aspects of DMs, there are public concerns about potential copyright violations when unauthorized images are" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 234, + 541, + 293 + ], + "blocks": [ + { + "bbox": [ + 310, + 234, + 541, + 293 + ], + "lines": [ + { + "bbox": [ + 310, + 234, + 541, + 293 + ], + "spans": [ + { + "bbox": [ + 310, + 234, + 541, + 293 + ], + "type": "image", + "image_path": "7428e315389c5e19d3151d63d1b87971882ce6cbe95f36d2b569ef564ab4b89a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 296, + 541, + 357 + ], + "blocks": [ + { + "bbox": [ + 309, + 296, + 541, + 357 + ], + "lines": [ + { + "bbox": [ + 309, + 296, + 541, + 357 + ], + "spans": [ + { + "bbox": [ + 309, + 296, + 541, + 357 + ], + "type": "image", + "image_path": "933c572476350cfb0bcda88ebf9012a425dbc8eeebfbf9b97adc338b724b50a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 361, + 541, + 426 + ], + "blocks": [ + { + "bbox": [ + 309, + 361, + 541, + 426 + ], + "lines": [ + { + "bbox": [ + 309, + 361, + 541, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 361, + 541, + 426 + ], + "type": "image", + "image_path": "52aa3d2ff349e20f34806b12e0a16c0178e34baa24b09c1578ac72a4c70f4ae9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 448, + 545, + 569 + ], + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 569 + ], + "type": "text", + "content": "Figure 1. Copyright issues of DMs and adversarial example-based methods for copyright protection. Without protection, DMs can easily imitate original images under different image generation scenarios. A previous method (AdvDM [28]) generates adversarial examples by optimization against DMs to prevent DMs from extracting the feature of the original images, resulting in the generation of chaotic images. Our method goes one step further to build a generator that embeds personal watermarks into the generation of adversarial images. Such examples force DMs to generate images with visible watermarks for tracing copyright. Our method is fast, simple yet powerful in protecting copyrights against DMs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 593, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 547, + 664 + ], + "type": "text", + "content": "used to train DMs or to generate new creations by DMs [55]. For example, an entity could use copyrighted paintings shared on the Internet to generate images with a similar style and appearance, potentially earning illegal revenue. Therefore, it is necessary to develop techniques to prevent the imitation of unauthorized creations by DMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Several previous works have been proposed to address these concerns. Watermarking [2, 33] is a common solution that embeds an invisible message into the image, which can then be extracted to identify the existence of copyright." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24420" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "content": "Recent techniques typically integrate watermarking into the generative process of DMs [7, 12, 38]. However, these methods have several limitations. First, they are primarily designed to protect the copyright of DMs or to distinguish generated images from natural ones, which differs from our goal of protecting the copyright of human creations. Second, embedding watermarks into the generative process requires re-training or at least fine-tuning DMs, and a post-process to extract the watermark is needed to identify it, which can be costly. On the other hand, methods that use adversarial examples for DMs to protect copyright have been proposed recently [27, 28, 50]. The idea is to generate adversarial examples to hinder DMs from extracting the features of the original images. This generation process does not require any re-training of DMs or any post-processing for the generated image. However, there are still several limitations. First, even though these adversarial examples could add chaotic textures to the generated images, such changes are difficult to comprehend. For instance, users may wonder if the problem originates from the DM itself rather than considering the copyright issue. Second, the original image's copyright is not traceable since no copyright-related information is embedded. Third, each adversarial example needs to be optimized separately against DMs, which makes the generation time-consuming." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 374, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 289, + 566 + ], + "type": "text", + "content": "In this work, we propose a simple yet powerful method to protect copyrighted images from imitation by DMs. Different from previous adversarial example-based methods, we go one step further by embedding personal watermarks into the generation of adversarial examples for copyright tracing. Such examples could force DMs to generate images with visible watermarks as well as chaotic textures. In addition, instead of using iterative optimization against DMs to generate examples, we train a generator beforehand using only a few samples. Once the generator is trained, it can be used to generate adversarial examples significantly fast. As shown in Figure 1, instead of posting the original image on the Internet, posting adversarial examples with only subtle changes could prevent DMs from imitation. Compared to previous methods, our method provides a more straightforward way to warn of potential copyright violations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 567, + 258, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 567, + 258, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 567, + 258, + 578 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 641 + ], + "type": "text", + "content": "- We design a novel framework that embeds personal watermarks into the generation of adversarial examples to prevent copyright violations caused by DM-based imitation. Our adversarial examples can force DMs to generate images with visible watermarks for tracing copyright." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "content": "- We train a generator based on a conditional GAN architecture to generate adversarial examples (Figure 2). We design three losses: adversarial loss, GAN loss, and weighted perturbation loss. These losses aim to improve the attack ability of the adversarial examples while making the perturbation invisible to the human eye." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 180 + ], + "type": "text", + "content": "- We conduct extensive experiments in various image generation scenarios, along with robustness evaluations against defenses and assessments of transferability to other models. The experiments demonstrate that our adversarial examples can prevent unauthorized images from being learned and can generate visible watermarks for copyright tracing. Our generation process is significantly fast (0.2s per image), and the generated examples also exhibit good transferability across other generative models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 191, + 398, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 398, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 398, + 204 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 212, + 463, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 463, + 224 + ], + "type": "text", + "content": "2.1. Generative Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 231, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 350 + ], + "type": "text", + "content": "In recent years, DMs have made significant advancements in various fields [9, 37, 45, 55]. DMs are a type of generative model that simulates a diffusion process to generate new data samples. DMs consist of a forward process and a reverse process. The forward process involves gradually adding Gaussian noise to the data over a series of time steps until it becomes completely random. The reverse process gradually removes noise through a series of learned denoising steps, reconstructing the data back to its original form or generating new samples from the same distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 350, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 460 + ], + "type": "text", + "content": "DMs have been applied to various generation tasks such as image synthesis [9, 43], text-to-image generation [1, 41, 58, 59], and text-guided image editing [4, 34, 54, 61]. Among these, latent diffusion models [42], which perform the diffusion steps in the latent image space, have demonstrated a remarkable ability to generate high-resolution and realistic images. Despite the impressive performance, some concerns that generating new creations by DMs based on unauthorized images could lead to copyright issues." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 467, + 430, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 430, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 430, + 480 + ], + "type": "text", + "content": "2.2. Image Watermarking" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 605 + ], + "type": "text", + "content": "Image watermarking is a technique that embeds a piece of information into an image to protect the image from copyright infringement. Traditional watermarking methods typically involve altering pixel values [6] or manipulating the frequency domain of an image [35] to embed watermarks. More recently, deep-learning-based methods have been developed. A common approach is using encoder-extractor networks, where an encoder embeds the watermark into the image, and an extractor is trained to retrieve the watermark from the marked image [49, 60, 63]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": "In terms of watermarking for generative models, most existing works focus on protecting the copyright of generative models [11, 30, 38, 51, 57] or distinguishing generative images from natural images [7, 12, 56, 62]. These methods involve reformulating the training process of DMs to inject watermarks into the images, as well as developing a decoder to identify the embedded messages from the generated images. However, the re-training can be costly and it is impractical to check every generated image with a decoder." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "24421" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 281, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 281, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 281, + 85 + ], + "type": "text", + "content": "2.3. Adversarial Examples for Generative models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 286, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 286, + 198 + ], + "type": "text", + "content": "Adversarial examples are created by adding small but intentional perturbations to input samples such that the perturbed inputs cause the model to produce incorrect answers. Various adversarial attack methods have been proposed to confuse classification models [10, 16, 31, 52] or segmentation models [13, 53, 65]. There are also several existing works studying adversarial examples for generative models such as GANs [24], variational autoencoders (VAEs) [47], and flow-based generative models [39]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 354 + ], + "type": "text", + "content": "Recently, methods that generate adversarial examples for DMs [27, 28, 50] have been proposed. These methods maximize the diffusion loss to obtain a perturbation, and adding such a perturbation to the original image enables DMs to generate images with chaotic content. However, these methods have several drawbacks: 1) such chaotic content is not straightforward to indicate copyright violation, 2) the copyright of the original image is not traceable, and 3) optimization against DMs is time-consuming. We propose a simple yet powerful method that directly incorporates personal watermarks into the generation of adversarial examples, and we build a generator that can generate adversarial examples significantly faster than existing methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 363, + 153, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 153, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 153, + 376 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 383, + 160, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 160, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 160, + 395 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": "Given an original image " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": ", we have a diffusion model " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " that can learn the data distribution of " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " and generate images with similar style and appearance. In this work, we aim to generate an adversarial example " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " that can prevent " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " from extracting the feature of " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " and can directly show copyright information on the generated image. For example, for imaged-to-image generation, the generated image from " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " by using " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " can be denoted as " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\theta(x')" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": ". We design a way to provide the copyright information by using a watermark that contains the ownership details (e.g. artist name) of the original image. Therefore, given a watermark image " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": ", the adversarial example " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " that embeds " + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 401, + 287, + 545 + ], + "type": "text", + "content": " can be calculated by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 552, + 287, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 552, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 111, + 552, + 287, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x ^ {\\prime} := \\underset {x ^ {\\prime}} {\\operatorname {a r g m i n}} \\| \\theta \\left(x ^ {\\prime}\\right) - m \\|, \\tag {1} \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\| x - x ^ {\\prime} \\| \\leq \\sigma , \\end{array} \\\\ \\end{array}", + "image_path": "ef751ae35e489f2482424350e2e8f074678eae75ed5f1ce651f914a970bc3b09.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 286, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 286, + 654 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 594, + 286, + 654 + ], + "type": "text", + "content": " is a constant to control the range of the perturbation. This is a targeted attack setting where we compel the image generated by DMs to be closer to the watermark, while ensuring that the adversarial example remains as close to the original one as possible." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "type": "text", + "content": "However, several issues are associated with generating adversarial examples using Equation 1. First, directly optimizing " + }, + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "type": "text", + "content": " against DMs for each " + }, + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 654, + 286, + 714 + ], + "type": "text", + "content": " could be time-consuming, so we need to find a way to speed up the generation process. Second, since there are several image-generation sce" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "narios related to imitating the original image, such as text-guided image-to-image generation [4], text-to-image generation under textual inversion [14] or DreamBooth [43], we aim to generate an adversarial example " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " that can be applicable to all these scenarios. Meanwhile, we also expect that " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " could have good transferability on other generative models. Third, since our target attack aims to add personal watermarks to the generated images, the perturbation naturally becomes larger in the watermark regions. We need a solution to make the perturbation invisible to human eyes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 192, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 263 + ], + "type": "text", + "content": "To address the first issue, we train a generator beforehand instead of implementing optimization every time. To address the second and third issues, we design an adversarial loss, a GAN loss, and a perturbation loss to improve the attack ability across various tasks and models while keeping the perturbation invisible." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 270, + 435, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 435, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 435, + 282 + ], + "type": "text", + "content": "3.2. Architecture Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": "We propose a conditional GAN architecture for generating watermark-embedded adversarial examples. The architecture of our method contains three components: a generator " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ", a discriminator " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ", and a target DM " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ", shown in Figure 2. The generator " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " has an encoder to extract features from the inputs and a decoder to generate perturbation using the features. The inputs of " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " are the original image " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " and the watermark " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ", and the perturbation is generated conditioned on " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ". We perform the conditioning by concatenating " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " along the channel dimension and feeding them into " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ". The adversarial example " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " can be obtained by adding " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " and the generated perturbation. The discriminator " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " is a classifier that distinguishes " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": ", and the target DM " + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 289, + 545, + 456 + ], + "type": "text", + "content": " is always kept frozen." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": "We design three losses to optimize " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": ". 1) A GAN loss aims to compel " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": " to be closer to " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": ". By using the GAN loss, we can train " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": " to produce images that " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": " cannot distinguish from input images. 2) A perturbation loss is designed to bound the magnitude of the perturbation and make it invisible. 3) An adversarial loss is used for " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": " to generate perturbation that can attack DMs under various scenarios. The details are described in Section 3.3. Once the generator " + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 456, + 545, + 577 + ], + "type": "text", + "content": " is optimized, it can be used directly to generate adversarial examples significantly fast." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 583, + 395, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 395, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 395, + 595 + ], + "type": "text", + "content": "3.3. Loss functions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "content": "GAN loss. " + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "content": " is used to quantify the difference between the adversarial example and the original image. For a set of images, " + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 646, + 545, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 646, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 330, + 646, + 545, + 660 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G A N} = \\mathbb {E} _ {x} \\log D (x) + \\mathbb {E} _ {x} \\log \\left\\{1 - D \\left(x ^ {\\prime}\\right) \\right\\}. \\tag {2}", + "image_path": "796ec6ef43af36d7a72c41384675b7ef0d181d7895011e5fb92c19529b9bb809.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "type": "text", + "content": "The adversarial image " + }, + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "type": "text", + "content": " is generated under the condition of the watermark " + }, + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 668, + 545, + 691 + ], + "type": "text", + "content": ", which can be written as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 381, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 381, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = x + G (x \\mid m). \\tag {3}", + "image_path": "ce53b65122f94ecf6847f292c2423eaa11baafd30fe6502290c684d8c0daf49d.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24422" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 73, + 500, + 207 + ], + "blocks": [ + { + "bbox": [ + 89, + 73, + 500, + 207 + ], + "lines": [ + { + "bbox": [ + 89, + 73, + 500, + 207 + ], + "spans": [ + { + "bbox": [ + 89, + 73, + 500, + 207 + ], + "type": "image", + "image_path": "7e67fe78b717252a06df7e6f67ab1b44dea6a6340d416f4aea703959dfb42086.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": "Figure 2. Architecture overview. " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " generates perturbation for " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " produce " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " to compel " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " to be closer to " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " aims to force the image generated by " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " to display a visible watermark " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pert}" + }, + { + "bbox": [ + 46, + 224, + 547, + 247 + ], + "type": "text", + "content": " further bounds the magnitude of the perturbation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "type": "text", + "content": "Perturbation loss. " + }, + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pert}" + }, + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "type": "text", + "content": " aims to limit the magnitude of the perturbation. We design this loss based on the soft hinge loss used in previous adversarial attacks [3, 29, 52]. However, the soft hinge loss assigns equal importance to all perturbations within the image. In our case, the perturbation in the watermark region will be larger than that in other regions, making the watermark visible even in the adversarial example. To better conceal the watermark as well as bound the perturbation in the whole image, we design a weighted perturbation loss, where we assign larger weights " + }, + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 267, + 289, + 423 + ], + "type": "text", + "content": " to the watermark region. This encourages the perturbation in the watermark region to be smaller and invisible. For a set of images, this loss can be expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 287, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p e r t} = \\mathbb {E} _ {x} \\max (0, \\| G (x \\mid m) (1 + w \\cdot m) \\| _ {2} - c), \\tag {4}", + "image_path": "ad3744c001e74f33d7e85300456b3ee474dc6f52a3aee9bb3c6e696a8e8527c3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 266, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 266, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 266, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 266, + 462 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 450, + 266, + 462 + ], + "type": "text", + "content": " is a constant denoting a user-specified bound." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "content": "Adversarial loss. " + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "content": " is used to generate a perturbation that can attack DMs. As we expect our adversarial examples to have transferability across different image generation scenarios and various generative models, we design the loss based on the latent representation of latent diffusion models (LDM). We choose LDM for two reasons. First, LDM has outperformed other models in both high-quality image generation and sampling efficiency, and it has been used in various generative models. Generating adversarial examples against LDM naturally increases their transferability across various models. Second, LDM maps an image " + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "content": " to a latent representation " + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "inline_equation", + "content": "\\varepsilon(x)" + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "content": " which is the output of the Variational Auto-Encoder (VAE). This representation is widely used in conditional image generation scenarios, and it provides a simple way to control the image generation of DMs. Therefore, instead of directly exploiting the generated image, we minimize the distance between the representation of the adversarial example " + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "inline_equation", + "content": "\\varepsilon(x')" + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "content": " and that of the watermark " + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "inline_equation", + "content": "\\varepsilon(m)" + }, + { + "bbox": [ + 46, + 462, + 289, + 689 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = \\mathbb {E} _ {x} \\| \\varepsilon \\left(x ^ {\\prime}\\right) - \\varepsilon (m) \\| _ {2}. \\tag {5}", + "image_path": "09832c3644a619d46c69802b5af0b48145e2f5db7ee1c12c67835cd1ee0b388a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 267, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 267, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 545, + 292 + ], + "type": "text", + "content": "Optimization. Finally, we combine the above three losses to formulate the objective function for training, written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 301, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 301, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 319, + 301, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\min _ {G} \\max _ {D} V (D, G) = \\mathcal {L} _ {a d v} + \\alpha \\mathcal {L} _ {G A N} + \\beta \\mathcal {L} _ {p e r t}, \\tag {6}", + "image_path": "5fe5790fcd03caa38c6c9065556be3d3a5eb439cd317e28eaf60a8c83f17375d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": " are weights to control the balance of the three losses. By playing a minimax game between the " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": ", the optimal parameters of the model can be obtained." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 371, + 529, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 529, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 529, + 384 + ], + "type": "text", + "content": "3.4. Image Generation under Different Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "type": "text", + "content": "Since our target is to prevent copyright violations, we focus on conditional image generation scenarios where DMs sample the distribution of the original images to generate new ones. Such scenarios include text-guided image-to-image generation [36] and text-to-image generation under textual inversion [14] described in this section, and other scenarios including DreamBooth [43], LoRA [20], and Custom Diffusion [25] discussed in Appendix A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "content": "Text-guide image-to-image generation. In this setting, we can pass a text prompt and an input image to condition the image generation. Our method can be directly applied to this setting, as shown in Figure 2. During the training of the generator and discriminator, we simply use the same prompt (e.g. \"A painting\") for all samples. For the inference, any prompt can be used for generating new images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Textual inversion. Textual inversion is a method for capturing new concepts from a small number of images. These new concepts can then be used to control image generation in text-to-image settings. The steps to apply our method in this setting are as follows. First, given a small set of images within a similar category (e.g. same artist), we extract a text " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "S_{*}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " to represent our new concept. We then replace the vector associated with the tokenized string with an optimized embedding " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "v_{*}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": ". The optimization can be applied by minimizing the LDM loss over the images from the small set. During the image generation by DMs, the text " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "S_{*}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " is used as a condition to generate new images. We conduct this generation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24423" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "for both original images and adversarial images. The adversarial images are generated through the image-to-image generation setting as shown in Figure 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 118, + 128, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 128, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 128, + 132 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 138, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 246 + ], + "type": "text", + "content": "In this section, we evaluate our method for generating adversarial examples for DMs. We conduct experiments under text-guided image-to-image generation and textual inversion, performing various analyses to assess the effectiveness, robustness, and transferability of our method. We perform the evaluation on WikiArt [48] and ImageNet [44] datasets. All experiments are run on an NVIDIA A100 80GB GPU. Training the generator using 100 samples with 200 epochs takes about 20 minutes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 252, + 174, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 174, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 174, + 266 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "content": "Datasets. The WikiArt dataset contains 50k paintings from 195 artists. We randomly select 3000 paintings from 50 artists. We create different watermarks for different artists. These watermarks are binary images with a black background and white artist names such as \"VAN_GOGH\", \"CLAUDE_MONNET\". Since only a small number of samples are needed to train our generator, we select 10 images from each artist (500 images in total) for training and the remaining 2500 images are for evaluation. For the ImageNet dataset, we randomly select 2000 images (100 for training, 1900 for evaluation) from the goldfish, tiger shark, peacock, goose, Eskimo dog, and tabby cat category. The watermarks are designed as \"IMAGENET CAT\", \"IMAGENET FISH\", \"IMAGENETDOG\", etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 640 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate our method from two perspectives. For the adversarial example, we use MSE, PSNR, and SSIM to measure the image quality and the difference from the original image. For the image generated by DMs, we use Fréchet Inception Distance (FID) [18] and precision (prec.) [26] to measure the similarity between the generated and original images. Since we expect DMs to generate images with visible watermarks, we use Normalized Cross-Correlation (NCC) [21] to measure the similarity between the generated image and the watermark. Please note that, unlike traditional watermark methods, we do not extract the watermark from the generated image, and thus NCC is smaller in our case. However, to reduce the influence of the background, we first calculate a difference image between the image generated from the original image and that from the adversarial example, and then use this difference image and the watermark to obtain NCC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Implementation details. For the generator " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ", we adopt an architecture similar to that in [22]. For the discriminator " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ", we use an architecture similar to the discriminator in AdvGAN [52]. Further details regarding the implementation of our model are provided in Appendix B. We set the weight of GAN loss " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ", the weight of perturbation" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 69, + 525, + 133 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 525, + 133 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 525, + 133 + ], + "type": "table", + "html": "
MethodNCC ↑FID ↑prec. ↓recall
Original0120.70.850.28
AdvDM0228.20.030.25
Mist0.09205.80.040.33
Ours0.31245.50.030.29
", + "image_path": "cb277a58a8d311f287a80e34c438110ce965f2cd0acfd545b1bd703e9b6e59cb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 152, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 185 + ], + "type": "text", + "content": "Table 1. The performance of the generated image under text-guided image-to-image generation on WikiArt. Recall measures the diversity of the images and is only listed as a reference." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\beta = 10" + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "text", + "content": ", the weight for watermark region " + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "inline_equation", + "content": "w = 4" + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "text", + "content": ", and the bound " + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "inline_equation", + "content": "c = 10 / 255" + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "text", + "content": ". During the model training, we set the batch size to 8 and the learning rate to 0.001. For comparison, we use AdvDM [28] and Mist (fuse mode) [27], which also use adversarial examples to protect copyright. We set their sampling step to 100 and the maximum perturbation to " + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "inline_equation", + "content": "10 / 255" + }, + { + "bbox": [ + 304, + 206, + 545, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 299, + 518, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 518, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 518, + 312 + ], + "type": "text", + "content": "4.2. Text-guided Image-to-Image Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 461 + ], + "type": "text", + "content": "We evaluate the performance of our method under the setting described in Section 3.4. In this setting, a strength parameter is used to control the level of noise added to the original image during the generation of new images. A strength value close to 0 produces an image nearly identical to the original, while a value close to 1 results in an image that significantly differs from the original. We set the strength to 0.3 to simulate realistic imitation, and additional results with different strength values can be found in the Appendix C.1. For the text prompt, we uniformly use \"A painting\" for all images for convenience, and experiments using varied prompts are also included in the Appendix C.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "type": "text", + "content": "We first show the evaluation result of the image generated by LDM based on either the original image or the adversarial example. A high FID and a low precision mean that the generated image is far from the original one, which further indicates that LDM fails to capture the feature of the original image. As shown in Table 1, our adversarial examples largely increase FID and decrease precision. We also calculate NCC between the generated image and the watermark. For Mist, since it is possible to input a target image into the generation, we apply our idea to use different watermarks as the target images and use the same method to calculate NCC. Results show that our method has a significantly higher NCC, which means our adversarial examples can force LDM to generate images with more visible watermarks. These results suggest our method has great potential to prevent copyright violations caused by LDM imitation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We also evaluate the quality of the adversarial example generated by different methods. As shown in Table 2, the image quality of our adversarial examples is slightly higher than that of the existing methods. Moreover, the generation is significantly faster than the existing methods since we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24424" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 279, + 122 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 279, + 122 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 279, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 279, + 122 + ], + "type": "table", + "html": "
MethodMSE ↓PSNR ↑SSIM ↑Run time
AdvDM0.003829.10.8032s
Mist0.004029.00.8135s
Ours0.003730.10.800.2s
", + "image_path": "509cc43fbc35210ef2448b2fec15510eb28d0df191e34b7be428627d46093328.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 188, + 286, + 420 + ], + "blocks": [ + { + "bbox": [ + 53, + 188, + 286, + 420 + ], + "lines": [ + { + "bbox": [ + 53, + 188, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 286, + 420 + ], + "type": "image", + "image_path": "52edcfdfa9c73d131df21cba66b5313fcd596772c880e84fbe1f8450e1783599.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 438, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 495 + ], + "type": "text", + "content": "Figure 3. Comparison of different methods under text-guided image-to-image generation. Source is the image input to the LDM, and generated image is the output of the LDM (strength 0.3). The watermarks of our method are designed according to the artist name (from left to right: REMBRANDT, AUGUSTE_REN)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 287, + 564 + ], + "type": "text", + "content": "directly use the trained generator instead of iterative optimization against DMs. Our method has an advantage when generating a large number of examples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "For qualitative evaluation, we visualize two examples generated by different methods in Figure 3. More examples are shown in Appendix D.1. For the adversarial examples, all methods seem to introduce only subtle perturbations that are almost invisible to human eyes. For the generated images based on the adversarial examples, the existing methods succeed in adding chaotic textures to the generated image. However, such changes can be difficult to understand and evaluate. On the other hand, our adversarial examples succeed in generating images with visible watermarks as well as chaotic textures. Our method provides a more straightforward way to indicate copyright violations." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 538, + 146 + ], + "blocks": [ + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "type": "text", + "content": "Table 2. The image quality and generation time (per image) of the adversarial examples on WikiArt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 538, + 146 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 538, + 146 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 538, + 146 + ], + "type": "table", + "html": "
MethodGenerated imageAdversarial example
NCC↑FID ↑PSNR ↑SSIM ↑
Original0120.7--
AdvDM0327.326.30.73
Mist0.15318.426.70.74
Ours0.40412.927.10.75
", + "image_path": "c806d6c02d7cf4af38e39e4feb417563d8186154a8bef4d629f2b10712481f3d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 163, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 545, + 185 + ], + "type": "text", + "content": "Table 3. The performance of the generated image and adversarial examples on ImageNet under textual inversion." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 208, + 410, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 208, + 410, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 410, + 220 + ], + "type": "text", + "content": "4.3. Textual Inversion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "text", + "content": "Textual inversion is another important scenario related to copyright violations. In this scenario, given a small number of images, we first learn to represent the concept of the images in a new word " + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "inline_equation", + "content": "S_{*}" + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "text", + "content": ", then use " + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "inline_equation", + "content": "S_{*}" + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "text", + "content": " to guide image generation. Following the method of textual inversion [14], we separate our evaluation images into 4-image groups for each artist or ImageNet category. For each 4-image group, we set the maximum steps to optimize " + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "inline_equation", + "content": "S_{*}" + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "text", + "content": " to 5000. After the optimization, we generate 10 images for each " + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "inline_equation", + "content": "S_{*}" + }, + { + "bbox": [ + 304, + 228, + 545, + 358 + ], + "type": "text", + "content": " under 10 prompt templates. This process is implemented for both original images and our adversarial examples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 529 + ], + "type": "text", + "content": "For evaluation, we compare our method with existing methods, and the results are shown in Table 3. For the generated image, our method has higher NCC and FID, which indicates our method successfully injects the watermark information in the generation of the adversarial example. In addition, our adversarial example also has higher PSNR and SSIM, which means the generated perturbation is subtle. An example of textual inversion is shown in Figure 4, and examples comparing with the previous methods are shown in Appendix D.2. The features of the original images can be learned and used to guide image generation. On the other hand, when applying our method, images are generated with very clear watermarks, which can be useful for protecting copyrighted images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": "Weights of the losses. This experiment aims to study the mechanism of our method. In equation 6, we design three losses to train the generator. As our target is to attack DMs, the adversarial loss " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": " is necessary, and we set its weight to 1. We then evaluate the effect of the GAN loss " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": ", the perturbation loss " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pert}" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": ", and the weight used to control the perturbation of the watermark region " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": " by setting their weights " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 557, + 545, + 652 + ], + "type": "text", + "content": " to 0 or their default value." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "The results are shown in Table 4. The first row is the result of only using " + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": " and setting other losses to 0. In this case, although the attack ability for DMs is high, the generated adversarial examples are full of noise. By comparing the first and second rows, we observe that " + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24425" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 529, + 220 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 529, + 220 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 529, + 220 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 529, + 220 + ], + "type": "image", + "image_path": "68349c4fdd95264dd252bb6227a6140ed0173626065bbf72428ae736cc7a4839.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 236, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 545, + 247 + ], + "type": "text", + "content": "Figure 4. Examples of textual inversion on ImageNet. The watermarks used in the examples are IMAGENET CAT and IMAGENETDOG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 266, + 286, + 353 + ], + "blocks": [ + { + "bbox": [ + 48, + 266, + 286, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 266, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 286, + 353 + ], + "type": "table", + "html": "
WeightsGenerated imageAdversarial example
αβwNCC ↑FID ↑PSNR ↑SSIM ↑
×××0.60565.313.10.12
✓××0.41350.622.00.65
×✓×0.35318.224.20.70
×✓✓0.34292.227.80.76
✓✓✓0.31245.530.10.80
", + "image_path": "aaa8b6f25d4a2725ab33ceaecd1553a43cd5f985d8e5a83cfe228cedb71ae836.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": "significantly improves the image quality of the adversarial example, even though the attack ability decreases. On the other hand, " + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pert}" + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": " which aims to bound the magnitude of the perturbation, also has an effect similar to that of " + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": " (as seen when comparing first and third rows). In addition, adding weight term to " + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pert}" + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": " further improves the quality of the adversarial examples without sacrificing too much attack ability (as seen when comparing third and fourth rows). Finally, we combine all these terms to achieve a balanced performance between image quality and attack ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Maximum perturbation bound. In the case of adversarial examples, there is always a trade-off between image quality and the success rate of the attack. To explore this trade-off, we conducted experiments with different bounds. We used PSNR to measure the image quality of the adversarial examples, and NCC to assess the visibility of the watermark in the generated image which indicates the attack success rate. Table 5 shows that as the bound increases, the watermark on the generated image becomes more distinct. However, this may also cause the watermark on the adversarial image to become more visible, which can reduce the image quality. In real-world applications, users can adjust the bounds to balance between achieving high-quality adversarial images and ensuring watermark visibility to protect their work." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 266, + 529, + 305 + ], + "blocks": [ + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "lines": [ + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "text", + "content": "Table 4. The effect of the GAN loss " + }, + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "text", + "content": ", the perturbation loss " + }, + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "text", + "content": ", and the weight of the watermark region " + }, + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "inline_equation", + "content": "(w)" + }, + { + "bbox": [ + 47, + 371, + 287, + 405 + ], + "type": "text", + "content": " on generating adversarial examples (WikiArt) under image-to-image generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 266, + 529, + 305 + ], + "lines": [ + { + "bbox": [ + 323, + 266, + 529, + 305 + ], + "spans": [ + { + "bbox": [ + 323, + 266, + 529, + 305 + ], + "type": "table", + "html": "
Bound/25526101520
PSNR ↑40.134.230.124.217.5
NCC ↑0.130.220.310.370.41
", + "image_path": "f964a1979d3498b967d755b55ef0bbc37b33af7b1c22c8438ef02940621201cd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 358, + 533, + 423 + ], + "blocks": [ + { + "bbox": [ + 305, + 323, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 346 + ], + "type": "text", + "content": "Table 5. The effect of using different perturbation bound. The experiment uses WikiArt under image-to-image generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 358, + 533, + 423 + ], + "lines": [ + { + "bbox": [ + 319, + 358, + 533, + 423 + ], + "spans": [ + { + "bbox": [ + 319, + 358, + 533, + 423 + ], + "type": "table", + "html": "
SamplesNCC ↑FID ↑prec. ↓train time
10.12177.80.3220 sec
50.28235.30.041 min
100.30240.60.032 min
1000.31243.50.0220 min
", + "image_path": "27d246381ef2fef072dd4bef04227b5271cb53e48cd8f0252b9ae81474f7bd48.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 441, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 475 + ], + "type": "text", + "content": "Table 6. The effect and training time of using different number of samples to train a generator for a specific watermark. The experiment is conducted under image-to-image generation on WikiArt." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "The number of samples for training. To generate adversarial examples that contain a specific watermark, we need to include samples with that watermark in the training data. In this experiment, we analyze how many samples we need for a watermark to be learned by the generator. We choose a different number of samples to train a generator for a specific watermark. Please note that different watermarks can be trained together within the same generator, and this is the default setting for our other experiments. The results of training one personal watermark are in Table 6. We observe there is no large performance difference between using 10 samples and 100 samples for training, which shows that our method only needs a small number of samples to train to embed a specific watermark. Therefore, one use case of our method is to create a personal watermark generator using a few samples within " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "2 - 3\\mathrm{min}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", and then this generator can be used to embed the personal watermark for any of that user's creations to protect copyright." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24426" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 277, + 145 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 277, + 145 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 277, + 145 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 277, + 145 + ], + "type": "table", + "html": "
DefenseNo attack (Origin)Attack (Ours)
NCC ↑FID ↑NCC ↑FID ↑
No defense0120.70.31245.5
JPEG0124.10.21202.2
RS0123.20.20195.6
TVM0132.50.16166.8
", + "image_path": "06b2645255f05d97384282410bca69fb0fa15555a568dd076c66a03828d73582.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 287, + 185 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 185 + ], + "type": "text", + "content": "Table 7. The effects of applying defenses on adversarial examples under image-to-image generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 201, + 281, + 305 + ], + "blocks": [ + { + "bbox": [ + 49, + 201, + 281, + 305 + ], + "lines": [ + { + "bbox": [ + 49, + 201, + 281, + 305 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 281, + 305 + ], + "type": "image", + "image_path": "262dcc52c79a31b0c9c2483d4d4b297eefc1d993eb535d18884507cf1bf7ed7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 323, + 286, + 346 + ], + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 346 + ], + "type": "text", + "content": "Figure 5. Examples of the generated image from original image, attack image, and attack image with defense." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 367, + 235, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 235, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 235, + 380 + ], + "type": "text", + "content": "4.5. Robustness of adversarial examples" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 386, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 457 + ], + "type": "text", + "content": "To evaluate the robustness of the adversarial examples generated by our method, we apply several widely used adversarial defenses including JPEG compression [8], Randomized Smoothing (RS) [5], and Total Variance Minimization (TVM) [17]. Such defenses are directly processed on the adversarial examples to eliminate the perturbations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 459, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 637 + ], + "type": "text", + "content": "The results of applying attacks and defenses are shown in Table 7. The \"No attack\" column means applying defense on the original image to generate images. There is no watermark being embedded so NCC almost equals 0, and applying defenses slightly increases FID since the quality of the original image decreases. On the other hand, the \"Attack\" column means we first generate the adversarial examples, apply defenses to them (except the first row), and then use them to generate new images. We observe that NCC and FID both decrease when adding defenses to our adversarial examples. However, compared with \"No attack\", the \"Attack\" column still has higher NCC and FID which means the defenses cannot completely eliminate the effect of our attack. In Figure 5, we show some examples under defenses, and the watermark is still visible for all cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "type": "text", + "content": "4.6. Transferability on Other Generative Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "In this section, we explore the transferability of our method on other generative models. This is also known as the black-box attack, where one target DM is used to train the generator and obtain the adversarial examples, and then these" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 74, + 545, + 174 + ], + "blocks": [ + { + "bbox": [ + 307, + 74, + 545, + 174 + ], + "lines": [ + { + "bbox": [ + 307, + 74, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 74, + 545, + 174 + ], + "type": "image", + "image_path": "ca6ab38915a1c85aa02549dbffa4013df5390a5f4809492121a5df2dc13ab456.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 192, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 225 + ], + "type": "text", + "content": "Figure 6. Black-box attack to other models under image-to-image generation. The first row: generated images from the original image. Second row: generated images from our adversarial example." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 249, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 545, + 321 + ], + "type": "text", + "content": "adversarial examples are directly used to attack other models. We generate adversarial examples on Stable Diffusion 1.5 (SD 1.5) and test on other models or tools including Dreamshaper " + }, + { + "bbox": [ + 305, + 249, + 545, + 321 + ], + "type": "inline_equation", + "content": "8^{2}" + }, + { + "bbox": [ + 305, + 249, + 545, + 321 + ], + "type": "text", + "content": ", NovelAI, and Runway AI magic tools (image variation) under image-to-image generation. The details of the settings are described in the Appendix E." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 323, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 479 + ], + "type": "text", + "content": "The results are shown in Figure 6, and more examples can be found in Appendix D.3. Compared to images generated from the original image (first row), the images generated from our adversarial examples (second row) contain chaotic textures and a visible watermark for most cases. For the result of the Runway tool, although there is no obvious watermark on the generated image (as no parameter can be adjusted), our method still generates an image that differs significantly from the original one. Even though the watermarks are not as strong as they are in the white-box setting, our method still demonstrates good transferability. These results show that our method has the potential to prevent copyright violations from various image generative models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 495, + 378, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 378, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 378, + 506 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 516, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 659 + ], + "type": "text", + "content": "In this work, we propose a novel method for copyright prevention against DMs. We build a generator to embed personal watermarks into the adversarial examples, and such examples can force DMs to generate images with visible watermarks and chaotic textures. We use a conditional GAN architecture with three carefully designed losses to optimize the generator. Experiments show our method performs well in various image generation scenarios and exhibits good transferability to attack other models. Additionally, our generation is significantly faster than those of previous methods. Therefore, this work provides a simple yet powerful way to protect image copyright against DMs." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 673, + 498, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 673, + 498, + 683 + ], + "spans": [ + { + "bbox": [ + 315, + 673, + 498, + 683 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 315, + 673, + 498, + 683 + ], + "type": "text", + "content": "https://huggingface.co/runwayml/stable-diffusion-v1-5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 683, + 467, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 683, + 467, + 693 + ], + "spans": [ + { + "bbox": [ + 318, + 683, + 467, + 693 + ], + "type": "text", + "content": "2https://civitai.com/models/4384/dreamshaper" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 693, + 383, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 693, + 383, + 702 + ], + "spans": [ + { + "bbox": [ + 318, + 693, + 383, + 702 + ], + "type": "text", + "content": "3https://novelai.net/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 703, + 496, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 496, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 496, + 712 + ], + "type": "text", + "content": "4https://runwayml.com/ai-magic-tools/image-to-image/" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24427" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 178 + ], + "type": "text", + "content": "[2] Mahbuba Begum and Mohammad Shorif Uddin. Digital image watermarking techniques: a review. Information, 11(2): 110, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "text", + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 256 + ], + "type": "text", + "content": "[4] Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 2, 3, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[5] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, and Chris Honsinger. Digital watermarking. Journal of Electronic Imaging, 11(3):414-414, 2002. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 379 + ], + "type": "text", + "content": "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 456 + ], + "type": "text", + "content": "[8] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 457, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 287, + 491 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "text", + "content": "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 600 + ], + "type": "text", + "content": "[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] Volker Fischer, Mummadi Chaithanya Kumar, Jan Hendrik Metzen, and Thomas Brox. Adversarial examples for semantic image segmentation. arXiv preprint arXiv:1703.01101, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "text", + "content": "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 219 + ], + "type": "text", + "content": "[17] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2017. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 355, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 375 + ], + "type": "text", + "content": "[21] Bernd Jähne. Digital image processing. Springer Science & Business Media, 2005. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 377, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 442 + ], + "type": "text", + "content": "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "text", + "content": "[24] Jernej Kos, Ian Fischer, and Dawn Song. Adversarial examples for generative models. In 2018 IEEE security and privacy workshops (spw), pages 36-42. IEEE, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "type": "text", + "content": "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 580, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 623 + ], + "type": "text", + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "text", + "content": "[27] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2, 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[28] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, XUE Zhengui, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24428" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. arXiv preprint arXiv:1611.02770, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "type": "text", + "content": "[30] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 205 + ], + "type": "text", + "content": "[32] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 249 + ], + "type": "text", + "content": "[33] Saraju P Mohanty. Digital watermarking: A tutorial review. URL: http://www.csee.usf.edu/smohanty/research/Reports/WMSurvey1999Mohanty.pdf, 1999." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "text", + "content": "[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "text", + "content": "[35] KA Navas, Mathews Cheriyan Ajay, M Lekshmi, Tampy S Archana, and M Sasikumar. Dwt-dct-svd based watermarking. In 2008 3rd International Conference on Communication Systems Software and Middleware and Workshops (COMSWARE'08), pages 271-274. IEEE, 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "text", + "content": "[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 502 + ], + "type": "text", + "content": "[38] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 547 + ], + "type": "text", + "content": "[39] Phillip Pope, Yogesh Balaji, and Soheil Feizi. Adversarial robustness of flow-based generative models. In International Conference on Artificial Intelligence and Statistics, pages 3795-3805. PMLR, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "text", + "content": "[40] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 116 + ], + "type": "text", + "content": "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 184 + ], + "type": "text", + "content": "[44] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "text", + "content": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 240, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 285 + ], + "type": "text", + "content": "[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 285, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 546, + 319 + ], + "type": "text", + "content": "[47] Pedro Tabacof, Julia Tavares, and Eduardo Valle. Adversarial images for variational autoencoders. arXiv preprint arXiv:1612.00155, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 364 + ], + "type": "text", + "content": "[48] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 546, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 410 + ], + "type": "text", + "content": "[49] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 546, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 546, + 466 + ], + "type": "text", + "content": "[50] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 467, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 546, + 510 + ], + "type": "text", + "content": "[51] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 512, + 546, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 546, + 555 + ], + "type": "text", + "content": "[52] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610, 2018. 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 556, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 546, + 611 + ], + "type": "text", + "content": "[53] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 678 + ], + "type": "text", + "content": "[54] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 679, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 546, + 714 + ], + "type": "text", + "content": "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24429" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "methods and applications. ACM Computing Surveys, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[56] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "type": "text", + "content": "[57] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pages 14448-14457, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "type": "text", + "content": "[58] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "text", + "content": "[60] Ru Zhang, Shiqi Dong, and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia tools and applications, 78:8559-8575, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[61] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "type": "text", + "content": "[62] Yuan Zhao, Bo Liu, Ming Ding, Baoping Liu, Tianqing Zhu, and Xin Yu. Proactive deepfake defence via identity watermarking. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4602-4611, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 474 + ], + "type": "text", + "content": "[63] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 476, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 530 + ], + "type": "text", + "content": "[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "text", + "content": "[65] Peifei Zhu, Genki Osada, Hirokatsu Kataoka, and Tsubasa Takahashi. Frequency-aware gan for adversarial manipulation generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4315-4324, 2023. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24430" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_content_list.json b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0dbd61d0e7f297e0be360a3c827330eaaf3a870c --- /dev/null +++ b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "WaveFace: Authentic Face Restoration with Efficient Frequency Recovery", + "text_level": 1, + "bbox": [ + 111, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunqi Miao \nUniversity of Warwick \nYunqi.Miao.1@warwick.ac.uk", + "bbox": [ + 125, + 180, + 357, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiankang Deng* \nImperial College London \njiankangdeng@gmail.com", + "bbox": [ + 398, + 180, + 599, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jungong Han* \nUniversity of Sheffield \njungonghan77@gmail.com", + "bbox": [ + 642, + 180, + 839, + 232 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17afac7aaab6dd92cbbefd6eb67a843ccb63900741fd73dc675906c66d3a8544.jpg", + "image_caption": [ + "Figure 1. Left: Illustration of our frequency-aware BFR scheme. Restoration is performed in the frequency domain instead of the pixel domain. Right: Comparisons with state-of-the-art face restoration methods on degraded images. Previous methods struggle to restore facial details or the original identity while our WaveFace achieves a good balance of realism and fidelity with fewer artifacts." + ], + "image_footnote": [], + "bbox": [ + 84, + 257, + 450, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b424b1262fcce6b66248cf1572b667955d956d962ed383b8759a6974b5c9e05a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 257, + 916, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 588, + 313, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although diffusion models are rising as a powerful solution for blind face restoration, they are criticized for two problems: 1) slow training and inference speed, and 2) failure in preserving identity and recovering fine-grained facial details. In this work, we propose WaveFace to solve the problems in the frequency domain, where low- and high-frequency components decomposed by wavelet transformation are considered individually to maximize authenticity as well as efficiency. The diffusion model is applied to recover the low-frequency component only, which presents general information of the original image but 1/16 in size. To preserve the original identity, the generation is conditioned on the low-frequency component of low-quality images at each denoising step. Meanwhile, high-frequency components at multiple decomposition levels are handled by a unified network, which recovers complex facial details in a single step. Evaluations on four benchmark datasets show that: 1) WaveFace outperforms state-of-the-art methods in authen", + "bbox": [ + 73, + 628, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ticity, especially in terms of identity preservation, and 2) authentic images are restored with the efficiency $10 \\times$ faster than existing diffusion model-based BFR methods.", + "bbox": [ + 498, + 589, + 893, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 651, + 632, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Blind face restoration (BFR) aims to recover high-quality (HQ) facial images from various degradations, including down-sampling, blurriness, noise, and compression artifact [1, 26, 35]. BFR is challenging since the types of degradation are generally unknown in real-world scenarios. Improvements in the restoration quality over the past few years mainly benefit from the usage of multiple facial priors. Geometric priors, including facial landmarks [3] and parsing maps [2, 24], are employed to provide explicit facial structure information. Reference priors, such as HQ images/features [8, 29, 38], are used as guidance during facial restoration. Generative priors are generally obtained by pre-trained generation models such as StyleGAN [13],", + "bbox": [ + 496, + 676, + 893, + 876 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 524, + 887, + 653, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6583", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which facilitate the recovery of realistic textures [2, 27, 32]. Inspired by the superior generative ability of diffusion models (DMs) [10, 25], how to unleash its potential in authentic face restoration has gained much attention.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite plausible restoration results being achieved, existing DMs based methods [30, 34] generally suffer from two problems: 1) DMs are trained on large images $(512 \\times 512)$ , which requires massive computational resources for both training and inference, as thousands of iterations are required to obtain adequate outputs. Difface [34], for example, takes about 25 seconds to sample an image from noise. 2) Low-quality (LQ) images are firstly dif- fused to an intermediate step and then denoised by an unconditional diffusion model to obtain HQ counterparts [34]. However, as shown in Fig. 1, such an unconditional generation will lead to great uncertainty in restored images, which means the original identity and facial details such as wrinkles cannot be well preserved.", + "bbox": [ + 75, + 154, + 470, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above problems, we transfer the BFR task from the pixel domain to the frequency domain via Discrete Wavelet Transform (DWT). As shown in Fig. 1, an image can be decomposed and reconstructed by its four quarter-sized sub-bands: low- and high-frequency components without sacrificing information. Low-frequency sub-band mainly contains general information such as face structure while high-frequency ones contain rich facial details. By exponentially shrinking the size of input images, restoration within the frequency domain not only speeds up both training and inference but also maintains the authenticity of restoration. In the paper, we devise an efficient BFR method WaveFace that restores authentic HQ images by recovering its frequency components.", + "bbox": [ + 75, + 367, + 468, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "WaveFace consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Recovery (HFR) module. A diffusion model is used in LCD to restore the low-frequency sub-band of HQ images whose size is only 1/16 of the original image. In addition to smaller inputs, LCD leverages LQ counterparts as a condition throughout the generation to preserve the original identity. Meanwhile, HFR recovers high-frequency sub-bands decomposed at multiple DWT levels simultaneously within one step. With the frequency components restored by two modules, authentic images can be reconstructed via discrete inverse wavelet transform (IWT) within 1 ms. The contributions can be summarized as follows:", + "bbox": [ + 75, + 580, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an efficient blind face restoration approach, WaveFace, that restores authentic images by recovering their frequency components individually.", + "- A conditional diffusion model is adopted to restore the low-frequency component, which is 1/16 the size of the original image.", + "- A one-pass network is used to recover high-frequency sub-bands decomposed at multiple DWT levels simul" + ], + "bbox": [ + 91, + 780, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "taneously.", + "bbox": [ + 527, + 92, + 599, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Comprehensive experiments demonstrate the superiority of methods in both efficiency and authenticity.", + "bbox": [ + 514, + 106, + 888, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 151, + 640, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion models", + "text_level": 1, + "bbox": [ + 500, + 176, + 666, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models (DMs) are emerging generative models that corrupt the data with the successive addition of Gaussian noise during the diffusion process and then learn to recover the data during denoising. State-of-the-art DMs [10] have revealed the potential in CV tasks, such as image deblurring [4, 15] and image super-resolution [22, 23].", + "bbox": [ + 496, + 199, + 890, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accelerating DMs. Despite being powerful in generation, DDPM [10] has the downside of low inference speed, which requires thousands of steps for sampling. To accelerate inference, DDIM [19] adopts a non-Markovian diffusion process that allows step skipping during sampling.", + "bbox": [ + 496, + 291, + 890, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional DMs. Unconditional DM-based generation leads to great uncertainty in outputs, which means they generally fail to preserve the original identity and fine-grained facial details. Therefore, conditions are injected by cross-attention layer [22], adaptive normalization layer [21], or concatenation operation [23] to control the characteristic of the synthesized images.", + "bbox": [ + 496, + 367, + 890, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Blind face restoration", + "text_level": 1, + "bbox": [ + 500, + 483, + 702, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Facial priors exploited in blind face restoration (BFR) can be categorized into three types: geometric priors, reference priors, and generative priors.", + "bbox": [ + 496, + 506, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometric priors based methods leverage knowledge from facial landmark [3, 6], facial parsing maps [2, 24, 37], and facial component heatmaps [33]. However, they tend to show inferior performance since degraded images fail to provide accurate and adequate structural information.", + "bbox": [ + 496, + 551, + 890, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reference priors based methods either leverage a high-quality (HQ) reference image sharing the same identity as the degraded one [17] or a pre-constructed dictionary storing HQ facial features [8, 29, 38]. Firstly, a vectorquantized (VQ) codebook is pre-trained on HQ faces via VQ-GAN [7] to provide rich facial details. Features of degraded inputs are then fused with the prior either at image [29] or latent [8, 38] level for the generation of HQ counterparts. However, reference prior is restricted by the size of the codebook, which limits the diversity and richness of generated images.", + "bbox": [ + 496, + 628, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative priors encapsulated in pretrained face models [13, 14] are also leveraged to restore faithful faces [2, 27, 32]. PSFRGAN [2] modulates features at different scales progressively with facial parsing maps to achieve semantic-aware style transformation. GFP-GAN [27] and GPEN [32] adopt the pre-trained StyleGAN as a decoder and achieve a good balance between visual quality and fidelity of re", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6584", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stored images. Recently, diffusion models [10, 25] have shown the powerful generative ability in restoring HQ content from noisy images. Inspired by the ability, DifFace [34] firstly maps low-quality (LQ) inputs into an intermediate step of the denoising process, from which its HQ counterpart is recursively sampled. DR2 [30] employs the diffusion model to remove degradations, followed by a super-resolution model to obtain HQ counterparts.", + "bbox": [ + 75, + 90, + 467, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, previous DM-based BFR methods are performed in the pixel domain, where large inputs require massive computational resources for both training and inference. Besides, they both apply the unconditional scheme that leads to significant uncertainty in restored results.", + "bbox": [ + 75, + 212, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 300, + 202, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discrete Wavelet Transformation (DWT). DWT decomposes an image into low-frequency and high-frequency subbands. The low-frequency component presents general information while high-frequency ones express facial details in the vertical, horizontal, and diagonal directions. In the paper, we use Haar wavelet for DWT, which is widely used in real-world applications due to simplicity [20].", + "bbox": [ + 75, + 325, + 467, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image $\\pmb{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ , its low-frequency subband $\\pmb{x}_{ll}^{1} \\in \\mathbb{R}^{H/2 \\times W/2 \\times 3}$ and high-frequency sub-bands $\\pmb{x}_{lh}^{1}, \\pmb{x}_{hl}^{1}$ , and $\\pmb{x}_{hh}^{1} \\in \\mathbb{R}^{H/2 \\times W/2 \\times 3}$ can be decomposed by:", + "bbox": [ + 75, + 430, + 467, + 477 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {l l} ^ {1}, \\boldsymbol {x} _ {l h} ^ {1}, \\boldsymbol {x} _ {h l} ^ {1}, \\boldsymbol {x} _ {h h} ^ {1} = D W T (\\boldsymbol {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 484, + 468, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $DWT(\\cdot)$ refers to the DWT operation. Despite that the input size is decreased by a factor of four $(H / 2\\times W / 2)$ after a single DWT decomposition, DWT can be performed on the low-frequency component to further reduce the computational cost and expedite the inference:", + "bbox": [ + 75, + 510, + 467, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {l l} ^ {J + 1}, \\boldsymbol {x} _ {l h} ^ {J + 1}, \\boldsymbol {x} _ {h l} ^ {J + 1}, \\boldsymbol {x} _ {h h} ^ {J + 1} = D W T (\\boldsymbol {x} _ {l l} ^ {J}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 592, + 468, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J$ is level of DWT decomposition and $\\pmb{x}_{ll}^{J + 1},\\pmb{x}_{lh}^{J + 1},\\pmb{x}_{hl}^{J + 1},\\pmb{x}_{hh}^{J + 1}\\in \\mathbb{R}^{H / 2^{(J + 1)}\\times W / 2^{(J + 1)}\\times 3}$ . Reversibly, given the frequency sub-bands, an image can be reconstructed via discrete inverse wavelet transform (IWT):", + "bbox": [ + 75, + 619, + 467, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} = I W T \\left(\\boldsymbol {x} _ {l l} ^ {1}, \\boldsymbol {x} _ {l h} ^ {1}, \\boldsymbol {x} _ {h l} ^ {1}, \\boldsymbol {x} _ {h h} ^ {1}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 696, + 467, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. Training of diffusion models (DMs) consists of a diffusion process and a denoising process. Diffusion process transforms an image from the real data distribution $\\pmb{y}_0\\sim q(\\pmb{y}_0)$ into a pure Gaussian noise $\\pmb{y}_T$ by successively applying the following Markov diffusion kernel:", + "bbox": [ + 75, + 719, + 467, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {y} _ {t} \\mid \\boldsymbol {y} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t}; \\sqrt {1 - \\beta_ {t}} \\boldsymbol {y} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 806, + 467, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{\\beta_t\\}_{t=1}^T$ is a pre-defined or learned noise variance schedule. The marginal distribution at arbitrary timestep $t$ can be denoted as:", + "bbox": [ + 75, + 829, + 467, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {y} _ {t} \\mid \\boldsymbol {y} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {y} _ {0}, (1 - \\alpha_ {t}) \\boldsymbol {I}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 883, + 467, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = \\prod_{s=1}^{t}(1 - \\beta_{s})$ . Given $\\pmb{y}_{t}$ , the denoising process aims to recover $\\pmb{y}_{0}$ by recursively learning the transition from $\\pmb{y}_{t-1}$ to $\\pmb{y}_{t}$ , which is defined as the following Gaussian distribution:", + "bbox": [ + 498, + 90, + 890, + 148 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 162, + 890, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where parameters $\\theta$ are optimized by a denoising network $\\epsilon_{\\theta}$ that predicts $\\pmb{\\mu}_{\\theta}(\\pmb{y}_t,t)$ and $\\pmb{\\Sigma}_{\\theta}(\\pmb{y}_t,t)$ .", + "bbox": [ + 498, + 191, + 890, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 236, + 633, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the paper, WaveFace is proposed to handle blind face restoration (BFR) in the frequency domain. The framework is depicted in Fig. 2. First, degraded images are mapped to the frequency domain via Discrete Wavelet Transformation (DWT) (Sec. 3). The proposed Low-frequency Conditional Denoising (LCD) module (Sec. 4.1) and High-Frequency Recovery (HFR) module (Sec. 4.2) are adopted on the low- and high-frequency components respectively to remove degradations and restore facial details. Recovered frequency components are used for the image reconstruction.", + "bbox": [ + 498, + 261, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Low-frequency Conditional Denoising", + "text_level": 1, + "bbox": [ + 498, + 436, + 828, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the low-frequency component is similar to the down-sampled version of the original image, which largely determines the restoration quality. Due to its powerful generative ability from noisy inputs, a diffusion model (DM) is adopted in LCD to restore the low-frequency subband of high-quality (HQ) images.", + "bbox": [ + 498, + 460, + 890, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conditional DM. We denote the low-frequency sub-band of a pair of LQ and HQ images as $(\\pmb{x}_{ll_0}^j,\\pmb{y}_{ll_0}^j)$ . Since the low-frequency sub-band only contains a single map after DWT, the DWT levels $j$ and $ll$ are omitted in this section for simplicity. According to the diffusion process defined in Eq. (5), images are successively destroyed by Gaussian noise as timestep $t$ increases, which means the distribution $q(\\pmb{x}_T|\\pmb{x}_0)\\approx q(\\pmb{y}_T|\\pmb{y}_0)\\approx \\mathcal{N}(0,\\pmb {I})$ after a large timestep $T$ . Given this assumption, our LCD aims to learn the posterior distribution $p(\\pmb{y}_0|\\pmb{x}_0)$ :", + "bbox": [ + 498, + 551, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {y} _ {0} \\mid \\boldsymbol {x} _ {0}\\right) = \\int q \\left(\\boldsymbol {y} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}\\right) \\mathrm {d} \\boldsymbol {y} _ {1: T}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 714, + 890, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_{\\theta}(\\boldsymbol{y}_{t-1}|\\boldsymbol{y}_t)$ refers to the unconditional denoising process defined in Eq. (6). However, the unconditional scheme fails to preserve the original identity due to the lack of guidance during generation. To solve the problem, the low-frequency sub-band of LQ images $\\boldsymbol{x}_0$ are injected as the condition for the denoising process:", + "bbox": [ + 498, + 766, + 890, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t, \\boldsymbol {x} _ {0}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t, \\boldsymbol {x} _ {0}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6585", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5f8ce459f818d7d6b4db2e4fe1b7817f56c1a3bb7f0098d8ee84f65f4b5e003.jpg", + "image_caption": [ + "Figure 2. Overall framework of WaveFace. It consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Refinement (HFR) module. LCD (Sec. 4.1) predicts clean samples $\\pmb{x}_0$ from noise conditioned on LQ inputs through $T$ steps. Meanwhile, high-frequency sub-bands are concatenated as HFR (Sec. 4.2) inputs to recover vivid facial details. The predicted frequency components are projected back to the image via IWT." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 885, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0$ is injected by concatenating with the input $y_{t}$ along the channel dimension.", + "bbox": [ + 75, + 400, + 467, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objectives. Following DDPM [10], the denoising network $\\epsilon_{\\theta}$ is trained to predict noise vectors with the objective:", + "bbox": [ + 76, + 430, + 467, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L C D} = \\mathbb {E} _ {\\boldsymbol {y} _ {0}, t, \\boldsymbol {\\epsilon} _ {t} \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\left[ | | \\boldsymbol {\\epsilon} _ {t} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}, t) | | ^ {2} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 472, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, with the learned parameterized Gaussian transitions $p_{\\theta}(\\pmb{y}_{t-1}|\\pmb{y}_t,\\pmb{x}_0)$ , the low-frequency sub-band of HQ images can be recovered from a random Gaussian noise $\\pmb{y}_T \\sim \\mathcal{N}(\\pmb{0},\\mathbf{I})$ by recursively applying:", + "bbox": [ + 76, + 508, + 468, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {y} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {z}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 580, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}),\\alpha_{t} = 1 - \\beta_{t}$ , and $\\bar{\\alpha}_t = \\prod_{i = 1}^t\\alpha_i$", + "bbox": [ + 75, + 627, + 431, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trade-off between efficiency and quality. Training a DM on large inputs ( $\\times 512$ or more) requires massive computational resources and the evaluation takes thousands of steps to sample from the noise. Although this problem can be alleviated by adopting low-frequency sub-bands at higher DWT levels, multiple times of decomposition will also lead to information reduction. How to set DWT level wisely is the key to balance between efficiency and authenticity.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To illustrate how DWT level $J$ affects the authenticity, we visualize the image reconstructed by the low-frequency component of HQ images and the corresponding high-frequency ones of its LQ counterpart at different DWT levels in Fig. 3. Although noisy at the high-frequency part, the quality of images reconstructed at $J \\leq 2$ is acceptable, with key facial features such as wrinkles still preserved. However, the mosaic effect starts to emerge when $J$ keeps increasing. Ablation studies about the effect of DWT level on", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "training/inference are shown in Sec. 5.2.1. In the paper, $J$ is set as 2 empirically.", + "bbox": [ + 498, + 400, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. High-frequency Recovery", + "text_level": 1, + "bbox": [ + 498, + 438, + 732, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the general face information is restored by LCD, rich facial details are generally embedded in the high-frequency component, as shown in Fig. 1, which improves the authenticity of restored images. High-frequency sub-bands at multiple DWT levels also vary in size, which means more than one diffusion model is required for the recovery. To avoid the extravagant computational cost, we adopt a U-shaped network that can recover high-frequency sub-bands at multiple DWT levels at the same time and takes only one step for inference.", + "bbox": [ + 496, + 460, + 890, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a LQ and HQ image pair $(\\pmb{x},\\pmb{y})$ and their high-frequency sub-bands: $\\pmb{x}_H^j = \\{\\pmb{x}_{lh}^j,\\pmb{x}_{hl}^j,\\pmb{x}_{hh}^j\\}$ and $\\pmb{y}_H^j = \\{\\pmb{y}_{lh}^j,\\pmb{y}_{hl}^j,\\pmb{y}_{hh}^j\\}$ , where $j\\in \\{1,2\\}$ refers to DWT level, the framework our high-frequency recovery (HFR) module is illustrated in Fig. 4. HFR takes high-frequency sub-bands of LQ images that are channel-wise concatenated as inputs and outputs the recovered ones $\\hat{\\pmb{x}}_H^j$ as follows:", + "bbox": [ + 498, + 612, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {H} ^ {1}, \\boldsymbol {F} _ {H} ^ {2} = \\mathcal {U} _ {\\phi_ {u}} \\left(\\mathcal {H} _ {\\phi_ {i n}} \\left(\\boldsymbol {x} _ {H} ^ {1}\\right), \\mathcal {F} _ {\\phi_ {f}} \\left(\\boldsymbol {x} _ {H} ^ {2}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 728, + 890, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {H} ^ {j} = \\mathcal {H} _ {\\phi_ {o u t} ^ {j}} \\left(\\boldsymbol {F} _ {H} ^ {j}\\right) + \\boldsymbol {x} _ {H} ^ {j}, \\quad j \\in \\{1, 2 \\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 750, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{U}(\\cdot),\\mathcal{H}(\\cdot),\\mathcal{F}(\\cdot)$ denotes the U-shaped network, the convolutional layer, and the fusion module with the parameter $\\phi_u,\\phi_{in / out}^j$ , and $\\phi_f$ , respectively. $F_H^1,F_H^2$ represents feature maps of the last and penultimate decoding layers of the U-shaped network. Considering efficiency, the fusion module only contains two stacked ConvBlocks that apply $3\\times 3$ and $1\\times 1$ convolutional layers with ReLU in between, followed by concatenation with the input.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6586", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/114fde351da94fa3f60faf34eea20eda9db80c421ed7b7cac1b61241e8320b30.jpg", + "image_caption": [ + "(a) $J = 3$ $(\\times 64)$" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 210, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3d5598ee903e6a11b0340d42609b1f08093e46f8fc8922d38234f8d8b6499da.jpg", + "image_caption": [ + "Figure 3. Visualization of DWT frequency components and images reconstructed by the low-frequency sub-band of an HQ image and high-frequency sub-bands of its LQ counterpart. DWT level $J$ and resolution of low- / high-frequency sub-bands $(\\times N)$ are reported." + ], + "image_footnote": [], + "bbox": [ + 212, + 88, + 346, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d1f9177ad887155a1b1d36b68469e731074220b7c0423160ec11e88627f56a1.jpg", + "image_caption": [ + "(b) $J = 2$ （ $\\times 128)$" + ], + "image_footnote": [], + "bbox": [ + 361, + 88, + 493, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd60efaf13450e8c213049271a62f5611cd57ccda4045978fa6e9b88b341449b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 629, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$(\\times 128)$", + "bbox": [ + 496, + 191, + 539, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/613ca2f7a939901f92019de08fcfa15a9b3b42cfca1ec63143fd13d839fda834.jpg", + "image_caption": [ + "(c) $J = 1$ （20 $(\\times 256)$" + ], + "image_footnote": [], + "bbox": [ + 643, + 88, + 777, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ab4f1e55665da1f37183e70eaef116ff997bc4237becd24ee06c621fe2b33a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 88, + 911, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6ec98b1931a5f4267ce0a7a5b30497351412b4d6d2941605e8388e1ad980ccf.jpg", + "image_caption": [ + "Figure 4. Illustration of high-frequency recovery (HFR) module." + ], + "image_footnote": [], + "bbox": [ + 81, + 261, + 472, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Objectives. We adopt two loss terms for the training of HFR: recovery loss $\\mathcal{L}_r$ and content loss [28] $\\mathcal{L}_c$ . The recovery loss is applied not only to restore HQ high-frequency sub-bands $\\pmb{y}_H^j$ at each level but also to the image $\\pmb{y}^{j - 1}$ reconstructed via inverse wavelet transform (IWT):", + "bbox": [ + 75, + 435, + 468, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\sum_ {j} \\left( \\right.\\left|\\left| \\hat {\\boldsymbol {x}} _ {H} ^ {j} - \\boldsymbol {y} _ {H} ^ {j} \\right|\\right| + \\alpha \\left. \\right\\rvert\\left| I W T \\left(\\boldsymbol {y} _ {l l} ^ {j}, \\hat {\\boldsymbol {x}} _ {H} ^ {j}\\right) - \\boldsymbol {y} ^ {j - 1} \\right|\\left. \\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 521, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $IWT(\\cdot)$ denotes IWT operation and $\\alpha = 1$ is the parameter balancing losses between the frequency domain and the pixel domain. In addition to minimizing the pixel-wise distance, the content loss is applied to maximize the luminance, contrast, and structural similarity of the restored and original HQ image:", + "bbox": [ + 75, + 568, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = 1 - S S I M \\left(I W T \\left(\\boldsymbol {y} _ {l l} ^ {2}, \\hat {\\boldsymbol {x}} _ {H} ^ {2}, \\hat {\\boldsymbol {x}} _ {H} ^ {1}\\right), \\boldsymbol {y}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 669, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall objective of HFR is $\\mathcal{L}_{HFR} = \\mathcal{L}_r + \\lambda \\mathcal{L}_c$ , where the weighting parameter $\\lambda$ is set as 10.", + "bbox": [ + 75, + 699, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 743, + 209, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Settings and Datasets", + "text_level": 1, + "bbox": [ + 76, + 768, + 277, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Dataset. FFHQ [13] is used as the training set, which contains 70,000 high-quality (HQ) face images. Following BFR benchmark works [8, 29], we resize HQ images to the resolution of $512 \\times 512$ , and then synthesize low-quality (LQ) counterparts as follows:", + "bbox": [ + 75, + 791, + 468, + 868 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} = \\left\\{\\left[ (\\boldsymbol {y} \\otimes \\boldsymbol {k} _ {\\sigma}) \\downarrow_ {s} + \\boldsymbol {n} _ {\\delta} \\right] _ {\\mathrm {J P E G} _ {q}} \\right\\} \\uparrow_ {s}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 878, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where a HQ image $\\pmb{y}$ is firstly blurred by a Gaussian kernel $k_{\\sigma}$ , followed by a downsampling of scale $s$ . Afterward, Gaussian noise $\\pmb{n}_{\\delta}$ and JPEG compression with quality factor $q$ are applied to the image, which is then upsampled back to the original size to obtain its LQ counterpart $\\pmb{x}$ . The hyper-parameters $\\sigma, s, \\delta$ , and $q$ are uniformly sampled from [0.1, 15], [0.8, 32], [0, 20], and [30, 95] respectively.", + "bbox": [ + 496, + 262, + 892, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing Dataset. We evaluate WaveFace on a synthetic dataset: CelebA-Test and three real-world datasets: LFW-Test [11], WebPhoto-Test [27], and WIDER-Test [38]. CelebA-Test contains 3000 HQ images from CelebAHQ [12], and LQ counterparts are synthesized via Eq. (15) with the same degradation setting. In terms of three realworld datasets, LFW-Test contains 1711 mildly degraded face images in the wild, which comprises the first image for each person in LFW [11]. WebPhoto-Test includes 407 images crawled from the internet, some of which are old photos with severe degradation. WIDER-Test consists of 970 images with severe degradations from the WIDER dataset [31].", + "bbox": [ + 496, + 369, + 892, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For evaluation, we adopt two pixelwise metrics (PSNR and SSIM), a reference perceptual metric (LPIPS [36]), and a non-reference perceptual metric (FID [9]). To measure the consistency of identity, the angle between embeddings extracted by ArcFace (\"Deg.\") [5] is used. All metrics are used for the evaluation of synthetic data while only FID is used for real-world datasets due to the lack of referential HQ images.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The DWT decomposition level is set as 2, which reduces the input size of the diffusion model from $512 \\times 512$ to $128 \\times 128$ . We train LCD and HFR individually with Adam [16] optimizer. LCD is trained for 200K iterations with the learning rate $1e-4$ and batch size 32. For HFR, the training takes 70K iterations. The learning rate gradually decays from $1e-3$ to $1e-5$ by a factor of 0.1.", + "bbox": [ + 496, + 688, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 805, + 665, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.1 Low-frequency Conditional Denoising", + "text_level": 1, + "bbox": [ + 500, + 829, + 818, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DWT Levels. We conduct an experiment to investigate how the level of wavelet decomposition (DWT) affects the efficiency and authenticity. DWT is applied 0, 1, 2, 3 times on", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6587", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1ba426f93fb5203bb6c55dddcad9edac28769f17d732db7888467856bcae863a.jpg", + "table_caption": [ + "Table 1. Comparison between state-of-the-art diffusion model-based BFR methods and ours at different DWT levels in terms of efficiency and authenticity. The best and the second best performances are highlighted and underlined." + ], + "table_footnote": [], + "table_body": "
MethodLevelResolution#Param. (M)Time (s)PSNR↑SSIM↑
DifFace [34]-× 512175.3825.0419.060.46
DR2 [30]-× 25693.568.7622.890.57
WaveFace0× 512109.0619.3726.420.65
1× 25671.418.1227.080.73
2× 12825.361.9726.970.71
3× 643.91.4624.810.46
", + "bbox": [ + 81, + 156, + 467, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the training set, where the resolution of the diffusion model (DM)'s input becomes $512 \\times 512$ , $256 \\times 256$ , $128 \\times 128$ , and $64 \\times 64$ , respectively. The number of parameters and the inference time are used to evaluate the efficiency while PSNR and SSIM are used to evaluate authenticity. The comparison results are reported in Tab. 1. In comparison with the model trained on original images ( $\\times 512$ ), our frequency-aware scheme applies DM on the low-frequency sub-band only, which avoids some unknown noise on high-frequency components. Thus, the restoration quality improves after the wavelet decomposition is applied once. The quality can be maintained until the DWT level increases to 2, where the inference process is accelerated by $10 \\times$ . However, the performance is significantly harmed when the DWT level further increases (3 or higher) as too many details in the low-frequency component have been reduced. DWT level is set as 2 throughout the experiments to achieve the tradeoff between the efficiency and the quality of restoration.", + "bbox": [ + 75, + 258, + 468, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare the efficiency with two DM-based BFR methods: Difface [34] and DR2 [30]. To make a fair comparison, the inference time refers to the time of the whole denoising process, which starts from the pure Gaussian noise. The quality of generated images cannot be compared with that reported in Tab. 3 as images are generated with pre-/post-processing and fewer sampling steps in their methods. As can be seen, our method achieves better restoration quality at a lower computational cost.", + "bbox": [ + 75, + 532, + 468, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conditions. To investigate how the condition facilitates the restoration, we first remove the condition from the input of our LCD module and train an unconditional diffusion model for 200K iterations for a fair comparison. The restored face is denoted by \"Uncond.\" in Fig. 5. We notice that not only the identity is not preserved but also the image quality drops significantly. This could be attributed to the slow convergence of the unconditional training manner [21].", + "bbox": [ + 75, + 671, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we compare with two widely-used conditioning schemes: adaptive group normalization layers (AdaGN) [21] and cross-attention [22]. We replace the low-frequency sub-band of LQ images with the corresponding identity embedding extracted by Arcface [5], which are subsequently injected via AdaGN. Furthermore, a cross-attention layer is added to condition the denoised image", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbbe31da901b1875dd50f21c649ad29f723fc671c25ae9e198af80e6046e6c27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 93, + 633, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44b71f2dc83ba4a498e4520dcccb4005e7dd1a58f792aa0046729979c563c423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 93, + 758, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2504fc39c96a2f24c25d36881b1778a9abffbafdcb406a0c30a2443d0f0eac3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 93, + 883, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4fe4353810bbc3becc9869695d331900b2f7e69496f5c0f824a6184c275e4d8f.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of different conditioning schemes on CelebA-Test. We adopt \"Concat.\" in WaveFace. PSNR(↑) / Deg.(↓) are reported." + ], + "image_footnote": [], + "bbox": [ + 509, + 196, + 635, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ac1b2ed86629e988223fb8c18fa65f974a84bcad6295f801712f1af72a2d11b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 196, + 761, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43d1c414e1d67135e0dbebd5aecfcbff4c7bceda8fa08ff17b410586f0e05b9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 196, + 883, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "at each step with the LQ low-frequency information. Experimental details of the two schemes will be discussed in supplementary materials. The restored images are shown in Fig. 5 as \"AdaGN\" and \"Cross-Att.\", respectively. Although the original identity is preserved by the AdaGN scheme, the background is not well restored. A possible reason is that the identity-related condition is performed at the latent level only, which fails to provide pixel-wise constraints. On the other hand, despite the cross-attention scheme achieves a comparable restoration result with ours (\"Concat.), the extra layer requires more computational resources.", + "bbox": [ + 496, + 359, + 890, + 526 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/01aecce2480d68212747c792ae729b70203ade36083e10ba7a2994f8e41cf81b.jpg", + "table_caption": [ + "Table 2. Ablation studies of the effectiveness of our high-frequency recovery (HFR) module. $\\mathbf{x}_H$ denotes high-frequency sub-bands of LQ images and 1, 2 stands for the level of wavelet decomposition. Best performance is indicated by Red." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓FID↓
LCD + xH1 + xH224.3530.4300.53717.264
LCD + HFR1 + xH226.0810.6090.48114.020
LCD + xH1 + HFR224.8190.4600.53615.293
LCD + HFR1 + HFR2 (Ours)26.9670.7110.34313.062
", + "bbox": [ + 504, + 604, + 887, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2 High-frequency Recovery", + "text_level": 1, + "bbox": [ + 500, + 696, + 732, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To investigate the effectiveness of our High-frequency Recovery (HFR) module, we replace the recovered high-frequency sub-bands at different DWT levels with that of low-quality (LQ) images for image reconstruction. Quantitative comparisons on the restoration quality of different schemes are reported in Tab. 2. The restoration quality drops significantly when HFR is removed with a decline of $2.62\\mathrm{dB}$ in PSNR. Besides, the lack of recovered high-frequency components at different DWT levels deteriorates the restoration quality to varying extents. In comparison with those at level 2 ( $\\times 128$ ), high-frequency sub-bands at level 1 ( $\\times 256$ ) contribute more to the restoration quality.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6588", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71d4eb0dfc89f77df6ea96380a7f57f7246846329453f8597aa00520ca8bc64d.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison between images reconstructed at different DWT levels with and without (w/o) HFR module on CelebA-Test. $\\pmb{x}$ , $\\hat{\\pmb{x}}$ and $\\pmb{y}$ denotes LQ images, restored images, and HQ images. $\\pmb{x}^j (j\\in \\{1,2\\})$ denotes frequency components at $j$ -th DWT level. PSNR(↑) of the restored images are reported." + ], + "image_footnote": [], + "bbox": [ + 76, + 95, + 460, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The exemplar illustrated in Fig. 6 helps to understand intuitively. Before our HFR is applied (Row 2), the restored image contains cluttered noise and lacks details such as hairstyle due to messy high-frequency sub-bands of LQ images. The quality improves when the image is constructed with the refined high-frequency component (Row 3), which is less noisy and presents fine-grained details.", + "bbox": [ + 75, + 607, + 468, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 724, + 452, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "State-of-the-art (SOTA) methods used for comparison include generative prior based methods (GPEN [32] and GFPGAN [27]), reference prior based methods (VQFR [8] and RestoreFormer [29]) and recent diffusion model-based methods (DifFace [34] and DR2 [30]). Evaluations are conducted on both synthetic and real-world datasets.", + "bbox": [ + 75, + 748, + 467, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic dataset. Quantitative comparison on CelebA-Test [12] are illustrated in Tab. 3. WaveFace achieves the best scores on reference-based metrics and ranks second on FID. Apart from the better restoration quality, our method", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f3977a9bfe191e403abfbd48a0bacc4ddfc5f7bc8ad5dd92058837f9b75f2a3.jpg", + "table_caption": [ + "Table 3. Quantitative comparison on CelebA-Test for blind face restoration. \"Deg.\" refers to the angle between identity embeddings of restored images and HQ counterparts. The best and the second best performances are highlighted and underlined." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑SSIM↑LPIPS↓FID↓Deg↓
Input23.0380.3890.58667.50554.958
GPEN [32]24.3190.6030.44022.25138.473
GFP-GAN [27]24.7710.6740.36114.50136.958
VQFR [8]23.7350.6140.35814.04838.684
RestoreFormer [29]24.1910.6360.36412.00034.752
CodeFormer [38]25.0710.6720.35914.38537.677
DR2 [30]23.5830.6130.40214.67151.030
Difface [34]24.1550.6670.39013.37949.958
WaveFace26.967↑1.900.7110.34313.06228.263↑6.49
", + "bbox": [ + 504, + 143, + 895, + 296 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/422a93f008cc973ebf5c03fca6b2bd6a4123acb72f558f2eb750b438bbff9a43.jpg", + "table_caption": [ + "Table 4. Quantitative comparisons on three real-world datasets (-Test) in terms of FID. The best and the second best performances are highlighted and underlined." + ], + "table_footnote": [], + "table_body": "
MethodsLFWWebPhotoWIDER
Input124.974170.112199.961
GPEN [32]50.79280.57246.340
GFP-GAN [27]49.56087.58439.499
VQFR [8]50.86775.34844.107
RestoreFormer [29]47.75077.33049.817
CodeFormer [38]51.86383.19338.784
DR2 [30]45.298112.34445.348
Difface [34]45.22787.81137.112
WaveFace43.17581.52536.913
", + "bbox": [ + 524, + 359, + 867, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "faithfully preserves the identity with the minimum angle between identity embeddings and their HQ counterparts, outperforming the second-best method by 6.5 degrees.", + "bbox": [ + 496, + 535, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we present the qualitative comparison with SOTA methods on images with increasing degradations in Fig. 7. GFPN and GFPGAN generally produce oversmoothed results. VQFR and RestoreFormer introduce obvious artifacts especially when the inputs are corrupted by mild or severe degradation. Besides, the above methods fail to preserve the identity since they are highly dependent on the StyleGAN prior, where the pre-defined latent space limits the diversity in restored results. Diffusion prior-based methods (DifFace and DR2) tend to generate faithful images with minor artifacts. However, due to the lack of condition during generation, both methods still suffer from the failure in identity preservation. On the contrary, our method, with the help of diffusion prior as well as the injected condition, is capable of generating high-quality facial images with fewer artifacts and, meanwhile, is faithful to the original identity.", + "bbox": [ + 496, + 582, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world datasets. We compare with SOTA methods on three real-world datasets in terms of FID score, which measures the KL divergence between the distribution of restored images and that of HQ images in FFHQ [13]. To", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6589", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbc65886e4c29313ea0ef29234c66e271dd66cfcd2d02542e44137ac495e12f8.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with state-of-the-art BFR methods on CelebA-Test. Our method achieves a better restoration quality with the original identity and facial details well preserved. (Zoom in for best view)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 885, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/416d300d67e0ad204adef5a921a2582887f3811fe53647558c22a702d7a9de81.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison with state-of-the-art BFR methods on real-world datasets, including LFW-Test (first row), WebPhoto-Test (second row), and WIDER-Test (third row). (Zoom in for best view)." + ], + "image_footnote": [], + "bbox": [ + 78, + 362, + 885, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reduce the large gap between the synthetic degradations and real-world ones, following previous diffusion prior-based methods [18, 34], a pre-denoising network [18] is applied before our method to handle the unknown degradations. Quantitative results with SOTA methods are reported in Tab. 4. WaveFace outperforms state-of-the-art (SOTA) methods on LFW-Test [11] and WIDER-Test [38] and beats previous diffusion model based BFR methods on WebPhoto-Test [27]. As mentioned in DiffFace [34], the FID score may not be the most suitable evaluation metric for WebPhoto-Test, as a total of 407 images is significantly distant from representing the data distribution.", + "bbox": [ + 75, + 641, + 468, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Three typical examples are shown in Fig. 8 with the increasing degradation level. It is observed that previous BFR methods tend to generate either hazy or unnatural faces. Some methods even fail to generate adequate restored results when handling severe degradations (Row 3). In con", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trast, our approach provides much more natural and realistic results with rich details such as wrinkles.", + "bbox": [ + 498, + 641, + 890, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 681, + 619, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose WaveFace to solve the task in the frequency domain to achieve the trade-off between efficiency and authenticity. Efficiency is achieved by applying a diffusion model only on the low-frequency sub-band whose size is 1/16 of the original image. Meanwhile, high-frequency components decomposed at multiple DWT levels are recovered simultaneously by a one-forward framework, which ensures the preservation of facial details.", + "bbox": [ + 496, + 704, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. There is a considerable gap between real-world degradations and the simulated ones (Eq. (15)), leading to inferior results in some cases. Our future work will focus on how to simulate degradations that fit real-world scenarios.", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6590", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 1", + "[2] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In CVPR, 2021. 1, 2", + "[3] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. FSRNet: End-to-end learning face super-resolution with facial priors. In CVPR, 2018. 1, 2", + "[4] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 2", + "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5, 6", + "[6] Jiankang Deng, Anastasios Roussos, Grigorios Chrysos, Evangelos Ververas, Irene Kotsia, Jie Shen, and Stefanos Zafeiriou. The menpo benchmark for multi-pose 2d and 3d facial landmark localisation and tracking. IJCV, 2019. 2", + "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[8] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In ECCV, 2022. 1, 2, 5, 7, 8", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4", + "[11] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Report, 2008. 5, 8", + "[12] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7", + "[13] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 5, 7", + "[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2", + "[15] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 2", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[17] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In ECCV, 2018. 2" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In ICCV, 2021. 8", + "[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2", + "[20] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 3", + "[21] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 2, 6", + "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 6", + "[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2", + "[24] Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming-Hsuan Yang. Deep semantic face deblurring. In CVPR, 2018. 1, 2", + "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3", + "[26] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 1", + "[27] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 7, 8", + "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 2004. 5", + "[29] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In CVPR, 2022. 1, 2, 5, 7, 8", + "[30] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. DR2: Diffusion-based robust degradation remover for blind face restoration. In CVPR, 2023. 2, 3, 6, 7, 8", + "[31] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In CVPR, 2016. 5", + "[32] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 2, 7, 8", + "[33] Xin Yu, Basura Fernando, Bernard Ghanem, Fatih Porikli, and Richard Hartley. Face super-resolution guided by facial component heatmaps. In ECCV, 2018. 2", + "[34] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv:2212.06512, 2022. 2, 3, 6, 7, 8", + "[35] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6591", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[37] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In CVPR, 2022. 2", + "[38] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022. 1, 2, 5, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 232 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6592", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_model.json b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d555c17f4de8076a25b098565c94714ca3a5d3b7 --- /dev/null +++ b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.131, + 0.859, + 0.154 + ], + "angle": 0, + "content": "WaveFace: Authentic Face Restoration with Efficient Frequency Recovery" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.181, + 0.359, + 0.233 + ], + "angle": 0, + "content": "Yunqi Miao \nUniversity of Warwick \nYunqi.Miao.1@warwick.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.181, + 0.601, + 0.233 + ], + "angle": 0, + "content": "Jiankang Deng* \nImperial College London \njiankangdeng@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.643, + 0.181, + 0.84, + 0.233 + ], + "angle": 0, + "content": "Jungong Han* \nUniversity of Sheffield \njungonghan77@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.258, + 0.452, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.258, + 0.918, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.531, + 0.895, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Left: Illustration of our frequency-aware BFR scheme. Restoration is performed in the frequency domain instead of the pixel domain. Right: Comparisons with state-of-the-art face restoration methods on degraded images. Previous methods struggle to restore facial details or the original identity while our WaveFace achieves a good balance of realism and fidelity with fewer artifacts." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.589, + 0.314, + 0.604 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Although diffusion models are rising as a powerful solution for blind face restoration, they are criticized for two problems: 1) slow training and inference speed, and 2) failure in preserving identity and recovering fine-grained facial details. In this work, we propose WaveFace to solve the problems in the frequency domain, where low- and high-frequency components decomposed by wavelet transformation are considered individually to maximize authenticity as well as efficiency. The diffusion model is applied to recover the low-frequency component only, which presents general information of the original image but 1/16 in size. To preserve the original identity, the generation is conditioned on the low-frequency component of low-quality images at each denoising step. Meanwhile, high-frequency components at multiple decomposition levels are handled by a unified network, which recovers complex facial details in a single step. Evaluations on four benchmark datasets show that: 1) WaveFace outperforms state-of-the-art methods in authen" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.895, + 0.636 + ], + "angle": 0, + "content": "ticity, especially in terms of identity preservation, and 2) authentic images are restored with the efficiency \\(10 \\times\\) faster than existing diffusion model-based BFR methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.652, + 0.633, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.895, + 0.877 + ], + "angle": 0, + "content": "Blind face restoration (BFR) aims to recover high-quality (HQ) facial images from various degradations, including down-sampling, blurriness, noise, and compression artifact [1, 26, 35]. BFR is challenging since the types of degradation are generally unknown in real-world scenarios. Improvements in the restoration quality over the past few years mainly benefit from the usage of multiple facial priors. Geometric priors, including facial landmarks [3] and parsing maps [2, 24], are employed to provide explicit facial structure information. Reference priors, such as HQ images/features [8, 29, 38], are used as guidance during facial restoration. Generative priors are generally obtained by pre-trained generation models such as StyleGAN [13]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.525, + 0.888, + 0.655, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "which facilitate the recovery of realistic textures [2, 27, 32]. Inspired by the superior generative ability of diffusion models (DMs) [10, 25], how to unleash its potential in authentic face restoration has gained much attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.365 + ], + "angle": 0, + "content": "Despite plausible restoration results being achieved, existing DMs based methods [30, 34] generally suffer from two problems: 1) DMs are trained on large images \\((512 \\times 512)\\), which requires massive computational resources for both training and inference, as thousands of iterations are required to obtain adequate outputs. Difface [34], for example, takes about 25 seconds to sample an image from noise. 2) Low-quality (LQ) images are firstly dif- fused to an intermediate step and then denoised by an unconditional diffusion model to obtain HQ counterparts [34]. However, as shown in Fig. 1, such an unconditional generation will lead to great uncertainty in restored images, which means the original identity and facial details such as wrinkles cannot be well preserved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.58 + ], + "angle": 0, + "content": "To solve the above problems, we transfer the BFR task from the pixel domain to the frequency domain via Discrete Wavelet Transform (DWT). As shown in Fig. 1, an image can be decomposed and reconstructed by its four quarter-sized sub-bands: low- and high-frequency components without sacrificing information. Low-frequency sub-band mainly contains general information such as face structure while high-frequency ones contain rich facial details. By exponentially shrinking the size of input images, restoration within the frequency domain not only speeds up both training and inference but also maintains the authenticity of restoration. In the paper, we devise an efficient BFR method WaveFace that restores authentic HQ images by recovering its frequency components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.779 + ], + "angle": 0, + "content": "WaveFace consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Recovery (HFR) module. A diffusion model is used in LCD to restore the low-frequency sub-band of HQ images whose size is only 1/16 of the original image. In addition to smaller inputs, LCD leverages LQ counterparts as a condition throughout the generation to preserve the original identity. Meanwhile, HFR recovers high-frequency sub-bands decomposed at multiple DWT levels simultaneously within one step. With the frequency components restored by two modules, authentic images can be reconstructed via discrete inverse wavelet transform (IWT) within 1 ms. The contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.781, + 0.468, + 0.825 + ], + "angle": 0, + "content": "- We propose an efficient blind face restoration approach, WaveFace, that restores authentic images by recovering their frequency components individually." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.826, + 0.468, + 0.87 + ], + "angle": 0, + "content": "- A conditional diffusion model is adopted to restore the low-frequency component, which is 1/16 the size of the original image." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- A one-pass network is used to recover high-frequency sub-bands decomposed at multiple DWT levels simul" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.093, + 0.6, + 0.107 + ], + "angle": 0, + "content": "taneously." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.107, + 0.89, + 0.137 + ], + "angle": 0, + "content": "- Comprehensive experiments demonstrate the superiority of methods in both efficiency and authenticity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.152, + 0.642, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.177, + 0.667, + 0.192 + ], + "angle": 0, + "content": "2.1. Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Diffusion Models (DMs) are emerging generative models that corrupt the data with the successive addition of Gaussian noise during the diffusion process and then learn to recover the data during denoising. State-of-the-art DMs [10] have revealed the potential in CV tasks, such as image deblurring [4, 15] and image super-resolution [22, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Accelerating DMs. Despite being powerful in generation, DDPM [10] has the downside of low inference speed, which requires thousands of steps for sampling. To accelerate inference, DDIM [19] adopts a non-Markovian diffusion process that allows step skipping during sampling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Conditional DMs. Unconditional DM-based generation leads to great uncertainty in outputs, which means they generally fail to preserve the original identity and fine-grained facial details. Therefore, conditions are injected by cross-attention layer [22], adaptive normalization layer [21], or concatenation operation [23] to control the characteristic of the synthesized images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.704, + 0.498 + ], + "angle": 0, + "content": "2.2. Blind face restoration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Facial priors exploited in blind face restoration (BFR) can be categorized into three types: geometric priors, reference priors, and generative priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Geometric priors based methods leverage knowledge from facial landmark [3, 6], facial parsing maps [2, 24, 37], and facial component heatmaps [33]. However, they tend to show inferior performance since degraded images fail to provide accurate and adequate structural information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Reference priors based methods either leverage a high-quality (HQ) reference image sharing the same identity as the degraded one [17] or a pre-constructed dictionary storing HQ facial features [8, 29, 38]. Firstly, a vectorquantized (VQ) codebook is pre-trained on HQ faces via VQ-GAN [7] to provide rich facial details. Features of degraded inputs are then fused with the prior either at image [29] or latent [8, 38] level for the generation of HQ counterparts. However, reference prior is restricted by the size of the codebook, which limits the diversity and richness of generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generative priors encapsulated in pretrained face models [13, 14] are also leveraged to restore faithful faces [2, 27, 32]. PSFRGAN [2] modulates features at different scales progressively with facial parsing maps to achieve semantic-aware style transformation. GFP-GAN [27] and GPEN [32] adopt the pre-trained StyleGAN as a decoder and achieve a good balance between visual quality and fidelity of re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6584" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.212 + ], + "angle": 0, + "content": "stored images. Recently, diffusion models [10, 25] have shown the powerful generative ability in restoring HQ content from noisy images. Inspired by the ability, DifFace [34] firstly maps low-quality (LQ) inputs into an intermediate step of the denoising process, from which its HQ counterpart is recursively sampled. DR2 [30] employs the diffusion model to remove degradations, followed by a super-resolution model to obtain HQ counterparts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.468, + 0.288 + ], + "angle": 0, + "content": "Nevertheless, previous DM-based BFR methods are performed in the pixel domain, where large inputs require massive computational resources for both training and inference. Besides, they both apply the unconditional scheme that leads to significant uncertainty in restored results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.301, + 0.203, + 0.317 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.468, + 0.431 + ], + "angle": 0, + "content": "Discrete Wavelet Transformation (DWT). DWT decomposes an image into low-frequency and high-frequency subbands. The low-frequency component presents general information while high-frequency ones express facial details in the vertical, horizontal, and diagonal directions. In the paper, we use Haar wavelet for DWT, which is widely used in real-world applications due to simplicity [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.468, + 0.478 + ], + "angle": 0, + "content": "Given an image \\(\\pmb{x} \\in \\mathbb{R}^{H \\times W \\times 3}\\), its low-frequency subband \\(\\pmb{x}_{ll}^{1} \\in \\mathbb{R}^{H/2 \\times W/2 \\times 3}\\) and high-frequency sub-bands \\(\\pmb{x}_{lh}^{1}, \\pmb{x}_{hl}^{1}\\), and \\(\\pmb{x}_{hh}^{1} \\in \\mathbb{R}^{H/2 \\times W/2 \\times 3}\\) can be decomposed by:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.485, + 0.469, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {l l} ^ {1}, \\boldsymbol {x} _ {l h} ^ {1}, \\boldsymbol {x} _ {h l} ^ {1}, \\boldsymbol {x} _ {h h} ^ {1} = D W T (\\boldsymbol {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.468, + 0.586 + ], + "angle": 0, + "content": "where \\(DWT(\\cdot)\\) refers to the DWT operation. Despite that the input size is decreased by a factor of four \\((H / 2\\times W / 2)\\) after a single DWT decomposition, DWT can be performed on the low-frequency component to further reduce the computational cost and expedite the inference:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.593, + 0.469, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {l l} ^ {J + 1}, \\boldsymbol {x} _ {l h} ^ {J + 1}, \\boldsymbol {x} _ {h l} ^ {J + 1}, \\boldsymbol {x} _ {h h} ^ {J + 1} = D W T (\\boldsymbol {x} _ {l l} ^ {J}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.468, + 0.682 + ], + "angle": 0, + "content": "where \\(J\\) is level of DWT decomposition and \\(\\pmb{x}_{ll}^{J + 1},\\pmb{x}_{lh}^{J + 1},\\pmb{x}_{hl}^{J + 1},\\pmb{x}_{hh}^{J + 1}\\in \\mathbb{R}^{H / 2^{(J + 1)}\\times W / 2^{(J + 1)}\\times 3}\\). Reversibly, given the frequency sub-bands, an image can be reconstructed via discrete inverse wavelet transform (IWT):" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.697, + 0.468, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} = I W T \\left(\\boldsymbol {x} _ {l l} ^ {1}, \\boldsymbol {x} _ {l h} ^ {1}, \\boldsymbol {x} _ {h l} ^ {1}, \\boldsymbol {x} _ {h h} ^ {1}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.795 + ], + "angle": 0, + "content": "Diffusion Models. Training of diffusion models (DMs) consists of a diffusion process and a denoising process. Diffusion process transforms an image from the real data distribution \\(\\pmb{y}_0\\sim q(\\pmb{y}_0)\\) into a pure Gaussian noise \\(\\pmb{y}_T\\) by successively applying the following Markov diffusion kernel:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.808, + 0.468, + 0.827 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {y} _ {t} \\mid \\boldsymbol {y} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t}; \\sqrt {1 - \\beta_ {t}} \\boldsymbol {y} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.83, + 0.468, + 0.874 + ], + "angle": 0, + "content": "where \\(\\{\\beta_t\\}_{t=1}^T\\) is a pre-defined or learned noise variance schedule. The marginal distribution at arbitrary timestep \\(t\\) can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.885, + 0.468, + 0.903 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {y} _ {t} \\mid \\boldsymbol {y} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {y} _ {0}, (1 - \\alpha_ {t}) \\boldsymbol {I}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.15 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = \\prod_{s=1}^{t}(1 - \\beta_{s})\\). Given \\(\\pmb{y}_{t}\\), the denoising process aims to recover \\(\\pmb{y}_{0}\\) by recursively learning the transition from \\(\\pmb{y}_{t-1}\\) to \\(\\pmb{y}_{t}\\), which is defined as the following Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.164, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.193, + 0.892, + 0.224 + ], + "angle": 0, + "content": "where parameters \\(\\theta\\) are optimized by a denoising network \\(\\epsilon_{\\theta}\\) that predicts \\(\\pmb{\\mu}_{\\theta}(\\pmb{y}_t,t)\\) and \\(\\pmb{\\Sigma}_{\\theta}(\\pmb{y}_t,t)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.237, + 0.635, + 0.254 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.427 + ], + "angle": 0, + "content": "In the paper, WaveFace is proposed to handle blind face restoration (BFR) in the frequency domain. The framework is depicted in Fig. 2. First, degraded images are mapped to the frequency domain via Discrete Wavelet Transformation (DWT) (Sec. 3). The proposed Low-frequency Conditional Denoising (LCD) module (Sec. 4.1) and High-Frequency Recovery (HFR) module (Sec. 4.2) are adopted on the low- and high-frequency components respectively to remove degradations and restore facial details. Recovered frequency components are used for the image reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.438, + 0.829, + 0.455 + ], + "angle": 0, + "content": "4.1. Low-frequency Conditional Denoising" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.551 + ], + "angle": 0, + "content": "As shown in Fig. 1, the low-frequency component is similar to the down-sampled version of the original image, which largely determines the restoration quality. Due to its powerful generative ability from noisy inputs, a diffusion model (DM) is adopted in LCD to restore the low-frequency subband of high-quality (HQ) images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Conditional DM. We denote the low-frequency sub-band of a pair of LQ and HQ images as \\((\\pmb{x}_{ll_0}^j,\\pmb{y}_{ll_0}^j)\\). Since the low-frequency sub-band only contains a single map after DWT, the DWT levels \\(j\\) and \\(ll\\) are omitted in this section for simplicity. According to the diffusion process defined in Eq. (5), images are successively destroyed by Gaussian noise as timestep \\(t\\) increases, which means the distribution \\(q(\\pmb{x}_T|\\pmb{x}_0)\\approx q(\\pmb{y}_T|\\pmb{y}_0)\\approx \\mathcal{N}(0,\\pmb {I})\\) after a large timestep \\(T\\). Given this assumption, our LCD aims to learn the posterior distribution \\(p(\\pmb{y}_0|\\pmb{x}_0)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.715, + 0.892, + 0.756 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {y} _ {0} \\mid \\boldsymbol {x} _ {0}\\right) = \\int q \\left(\\boldsymbol {y} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}\\right) \\mathrm {d} \\boldsymbol {y} _ {1: T}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.767, + 0.892, + 0.858 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(\\boldsymbol{y}_{t-1}|\\boldsymbol{y}_t)\\) refers to the unconditional denoising process defined in Eq. (6). However, the unconditional scheme fails to preserve the original identity due to the lack of guidance during generation. To solve the problem, the low-frequency sub-band of LQ images \\(\\boldsymbol{x}_0\\) are injected as the condition for the denoising process:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t, \\boldsymbol {x} _ {0}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t, \\boldsymbol {x} _ {0}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6585" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Overall framework of WaveFace. It consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Refinement (HFR) module. LCD (Sec. 4.1) predicts clean samples \\( \\pmb{x}_0 \\) from noise conditioned on LQ inputs through \\( T \\) steps. Meanwhile, high-frequency sub-bands are concatenated as HFR (Sec. 4.2) inputs to recover vivid facial details. The predicted frequency components are projected back to the image via IWT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.468, + 0.43 + ], + "angle": 0, + "content": "where \\( \\pmb{x}_0 \\) is injected by concatenating with the input \\( y_{t} \\) along the channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.468, + 0.462 + ], + "angle": 0, + "content": "Objectives. Following DDPM [10], the denoising network \\(\\epsilon_{\\theta}\\) is trained to predict noise vectors with the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.473, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\nL _ {L C D} = \\mathbb {E} _ {\\boldsymbol {y} _ {0}, t, \\boldsymbol {\\epsilon} _ {t} \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\left[ | | \\boldsymbol {\\epsilon} _ {t} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}, t) | | ^ {2} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.47, + 0.571 + ], + "angle": 0, + "content": "During inference, with the learned parameterized Gaussian transitions \\( p_{\\theta}(\\pmb{y}_{t-1}|\\pmb{y}_t,\\pmb{x}_0) \\), the low-frequency sub-band of HQ images can be recovered from a random Gaussian noise \\( \\pmb{y}_T \\sim \\mathcal{N}(\\pmb{0},\\mathbf{I}) \\) by recursively applying:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.581, + 0.47, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {y} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {z}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.433, + 0.645 + ], + "angle": 0, + "content": "where \\(z\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}),\\alpha_{t} = 1 - \\beta_{t}\\) , and \\(\\bar{\\alpha}_t = \\prod_{i = 1}^t\\alpha_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Trade-off between efficiency and quality. Training a DM on large inputs (\\( \\times 512 \\) or more) requires massive computational resources and the evaluation takes thousands of steps to sample from the noise. Although this problem can be alleviated by adopting low-frequency sub-bands at higher DWT levels, multiple times of decomposition will also lead to information reduction. How to set DWT level wisely is the key to balance between efficiency and authenticity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To illustrate how DWT level \\( J \\) affects the authenticity, we visualize the image reconstructed by the low-frequency component of HQ images and the corresponding high-frequency ones of its LQ counterpart at different DWT levels in Fig. 3. Although noisy at the high-frequency part, the quality of images reconstructed at \\( J \\leq 2 \\) is acceptable, with key facial features such as wrinkles still preserved. However, the mosaic effect starts to emerge when \\( J \\) keeps increasing. Ablation studies about the effect of DWT level on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.401, + 0.892, + 0.431 + ], + "angle": 0, + "content": "training/inference are shown in Sec. 5.2.1. In the paper, \\( J \\) is set as 2 empirically." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.439, + 0.733, + 0.456 + ], + "angle": 0, + "content": "4.2. High-frequency Recovery" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Although the general face information is restored by LCD, rich facial details are generally embedded in the high-frequency component, as shown in Fig. 1, which improves the authenticity of restored images. High-frequency sub-bands at multiple DWT levels also vary in size, which means more than one diffusion model is required for the recovery. To avoid the extravagant computational cost, we adopt a U-shaped network that can recover high-frequency sub-bands at multiple DWT levels at the same time and takes only one step for inference." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.892, + 0.722 + ], + "angle": 0, + "content": "Given a LQ and HQ image pair \\((\\pmb{x},\\pmb{y})\\) and their high-frequency sub-bands: \\(\\pmb{x}_H^j = \\{\\pmb{x}_{lh}^j,\\pmb{x}_{hl}^j,\\pmb{x}_{hh}^j\\}\\) and \\(\\pmb{y}_H^j = \\{\\pmb{y}_{lh}^j,\\pmb{y}_{hl}^j,\\pmb{y}_{hh}^j\\}\\), where \\(j\\in \\{1,2\\}\\) refers to DWT level, the framework our high-frequency recovery (HFR) module is illustrated in Fig. 4. HFR takes high-frequency sub-bands of LQ images that are channel-wise concatenated as inputs and outputs the recovered ones \\(\\hat{\\pmb{x}}_H^j\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.729, + 0.891, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {H} ^ {1}, \\boldsymbol {F} _ {H} ^ {2} = \\mathcal {U} _ {\\phi_ {u}} \\left(\\mathcal {H} _ {\\phi_ {i n}} \\left(\\boldsymbol {x} _ {H} ^ {1}\\right), \\mathcal {F} _ {\\phi_ {f}} \\left(\\boldsymbol {x} _ {H} ^ {2}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.75, + 0.891, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {H} ^ {j} = \\mathcal {H} _ {\\phi_ {o u t} ^ {j}} \\left(\\boldsymbol {F} _ {H} ^ {j}\\right) + \\boldsymbol {x} _ {H} ^ {j}, \\quad j \\in \\{1, 2 \\}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{U}(\\cdot),\\mathcal{H}(\\cdot),\\mathcal{F}(\\cdot)\\) denotes the U-shaped network, the convolutional layer, and the fusion module with the parameter \\(\\phi_u,\\phi_{in / out}^j\\) , and \\(\\phi_f\\) , respectively. \\(F_H^1,F_H^2\\) represents feature maps of the last and penultimate decoding layers of the U-shaped network. Considering efficiency, the fusion module only contains two stacked ConvBlocks that apply \\(3\\times 3\\) and \\(1\\times 1\\) convolutional layers with ReLU in between, followed by concatenation with the input." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6586" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.212, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.193, + 0.253, + 0.207 + ], + "angle": 0, + "content": "(a) \\(J = 3\\) \\((\\times 64)\\)" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.089, + 0.348, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.495, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.193, + 0.541, + 0.207 + ], + "angle": 0, + "content": "(b) \\(J = 2\\) （ \\(\\times 128)\\)" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.089, + 0.63, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.193, + 0.54, + 0.206 + ], + "angle": 0, + "content": "\\((\\times 128)\\)" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.089, + 0.778, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.193, + 0.823, + 0.207 + ], + "angle": 0, + "content": "(c) \\(J = 1\\) （20 \\((\\times 256)\\)" + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.089, + 0.913, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Figure 3. Visualization of DWT frequency components and images reconstructed by the low-frequency sub-band of an HQ image and high-frequency sub-bands of its LQ counterpart. DWT level \\( J \\) and resolution of low- / high-frequency sub-bands \\( (\\times N) \\) are reported." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.262, + 0.473, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.405, + 0.466, + 0.421 + ], + "angle": 0, + "content": "Figure 4. Illustration of high-frequency recovery (HFR) module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Objectives. We adopt two loss terms for the training of HFR: recovery loss \\(\\mathcal{L}_r\\) and content loss [28] \\(\\mathcal{L}_c\\). The recovery loss is applied not only to restore HQ high-frequency sub-bands \\(\\pmb{y}_H^j\\) at each level but also to the image \\(\\pmb{y}^{j - 1}\\) reconstructed via inverse wavelet transform (IWT):" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.522, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\sum_ {j} \\left( \\right.\\left|\\left| \\hat {\\boldsymbol {x}} _ {H} ^ {j} - \\boldsymbol {y} _ {H} ^ {j} \\right|\\right| + \\alpha \\left. \\right\\rvert\\left| I W T \\left(\\boldsymbol {y} _ {l l} ^ {j}, \\hat {\\boldsymbol {x}} _ {H} ^ {j}\\right) - \\boldsymbol {y} ^ {j - 1} \\right|\\left. \\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.66 + ], + "angle": 0, + "content": "where \\( IWT(\\cdot) \\) denotes IWT operation and \\( \\alpha = 1 \\) is the parameter balancing losses between the frequency domain and the pixel domain. In addition to minimizing the pixel-wise distance, the content loss is applied to maximize the luminance, contrast, and structural similarity of the restored and original HQ image:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.67, + 0.469, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = 1 - S S I M \\left(I W T \\left(\\boldsymbol {y} _ {l l} ^ {2}, \\hat {\\boldsymbol {x}} _ {H} ^ {2}, \\hat {\\boldsymbol {x}} _ {H} ^ {1}\\right), \\boldsymbol {y}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.731 + ], + "angle": 0, + "content": "The overall objective of HFR is \\(\\mathcal{L}_{HFR} = \\mathcal{L}_r + \\lambda \\mathcal{L}_c\\), where the weighting parameter \\(\\lambda\\) is set as 10." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.744, + 0.21, + 0.761 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.279, + 0.785 + ], + "angle": 0, + "content": "5.1. Settings and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Training Dataset. FFHQ [13] is used as the training set, which contains 70,000 high-quality (HQ) face images. Following BFR benchmark works [8, 29], we resize HQ images to the resolution of \\(512 \\times 512\\), and then synthesize low-quality (LQ) counterparts as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.879, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} = \\left\\{\\left[ (\\boldsymbol {y} \\otimes \\boldsymbol {k} _ {\\sigma}) \\downarrow_ {s} + \\boldsymbol {n} _ {\\delta} \\right] _ {\\mathrm {J P E G} _ {q}} \\right\\} \\uparrow_ {s}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.893, + 0.369 + ], + "angle": 0, + "content": "where a HQ image \\(\\pmb{y}\\) is firstly blurred by a Gaussian kernel \\(k_{\\sigma}\\), followed by a downsampling of scale \\(s\\). Afterward, Gaussian noise \\(\\pmb{n}_{\\delta}\\) and JPEG compression with quality factor \\(q\\) are applied to the image, which is then upsampled back to the original size to obtain its LQ counterpart \\(\\pmb{x}\\). The hyper-parameters \\(\\sigma, s, \\delta\\), and \\(q\\) are uniformly sampled from [0.1, 15], [0.8, 32], [0, 20], and [30, 95] respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.893, + 0.566 + ], + "angle": 0, + "content": "Testing Dataset. We evaluate WaveFace on a synthetic dataset: CelebA-Test and three real-world datasets: LFW-Test [11], WebPhoto-Test [27], and WIDER-Test [38]. CelebA-Test contains 3000 HQ images from CelebAHQ [12], and LQ counterparts are synthesized via Eq. (15) with the same degradation setting. In terms of three realworld datasets, LFW-Test contains 1711 mildly degraded face images in the wild, which comprises the first image for each person in LFW [11]. WebPhoto-Test includes 407 images crawled from the internet, some of which are old photos with severe degradation. WIDER-Test consists of 970 images with severe degradations from the WIDER dataset [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Evaluation Metrics. For evaluation, we adopt two pixelwise metrics (PSNR and SSIM), a reference perceptual metric (LPIPS [36]), and a non-reference perceptual metric (FID [9]). To measure the consistency of identity, the angle between embeddings extracted by ArcFace (\"Deg.\") [5] is used. All metrics are used for the evaluation of synthetic data while only FID is used for real-world datasets due to the lack of referential HQ images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Implementation Details. The DWT decomposition level is set as 2, which reduces the input size of the diffusion model from \\(512 \\times 512\\) to \\(128 \\times 128\\). We train LCD and HFR individually with Adam [16] optimizer. LCD is trained for 200K iterations with the learning rate \\(1e-4\\) and batch size 32. For HFR, the training takes 70K iterations. The learning rate gradually decays from \\(1e-3\\) to \\(1e-5\\) by a factor of 0.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.806, + 0.666, + 0.821 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.82, + 0.846 + ], + "angle": 0, + "content": "5.2.1 Low-frequency Conditional Denoising" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "DWT Levels. We conduct an experiment to investigate how the level of wavelet decomposition (DWT) affects the efficiency and authenticity. DWT is applied 0, 1, 2, 3 times on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6587" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison between state-of-the-art diffusion model-based BFR methods and ours at different DWT levels in terms of efficiency and authenticity. The best and the second best performances are highlighted and underlined." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.157, + 0.468, + 0.242 + ], + "angle": 0, + "content": "
MethodLevelResolution#Param. (M)Time (s)PSNR↑SSIM↑
DifFace [34]-× 512175.3825.0419.060.46
DR2 [30]-× 25693.568.7622.890.57
WaveFace0× 512109.0619.3726.420.65
1× 25671.418.1227.080.73
2× 12825.361.9726.970.71
3× 643.91.4624.810.46
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.531 + ], + "angle": 0, + "content": "the training set, where the resolution of the diffusion model (DM)'s input becomes \\(512 \\times 512\\), \\(256 \\times 256\\), \\(128 \\times 128\\), and \\(64 \\times 64\\), respectively. The number of parameters and the inference time are used to evaluate the efficiency while PSNR and SSIM are used to evaluate authenticity. The comparison results are reported in Tab. 1. In comparison with the model trained on original images (\\(\\times 512\\)), our frequency-aware scheme applies DM on the low-frequency sub-band only, which avoids some unknown noise on high-frequency components. Thus, the restoration quality improves after the wavelet decomposition is applied once. The quality can be maintained until the DWT level increases to 2, where the inference process is accelerated by \\(10 \\times\\). However, the performance is significantly harmed when the DWT level further increases (3 or higher) as too many details in the low-frequency component have been reduced. DWT level is set as 2 throughout the experiments to achieve the tradeoff between the efficiency and the quality of restoration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.67 + ], + "angle": 0, + "content": "We also compare the efficiency with two DM-based BFR methods: Difface [34] and DR2 [30]. To make a fair comparison, the inference time refers to the time of the whole denoising process, which starts from the pure Gaussian noise. The quality of generated images cannot be compared with that reported in Tab. 3 as images are generated with pre-/post-processing and fewer sampling steps in their methods. As can be seen, our method achieves better restoration quality at a lower computational cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Conditions. To investigate how the condition facilitates the restoration, we first remove the condition from the input of our LCD module and train an unconditional diffusion model for 200K iterations for a fair comparison. The restored face is denoted by \"Uncond.\" in Fig. 5. We notice that not only the identity is not preserved but also the image quality drops significantly. This could be attributed to the slow convergence of the unconditional training manner [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Additionally, we compare with two widely-used conditioning schemes: adaptive group normalization layers (AdaGN) [21] and cross-attention [22]. We replace the low-frequency sub-band of LQ images with the corresponding identity embedding extracted by Arcface [5], which are subsequently injected via AdaGN. Furthermore, a cross-attention layer is added to condition the denoised image" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.094, + 0.635, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.094, + 0.759, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.094, + 0.884, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.197, + 0.637, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.197, + 0.762, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.197, + 0.884, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of different conditioning schemes on CelebA-Test. We adopt \"Concat.\" in WaveFace. PSNR(↑) / Deg.(↓) are reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.527 + ], + "angle": 0, + "content": "at each step with the LQ low-frequency information. Experimental details of the two schemes will be discussed in supplementary materials. The restored images are shown in Fig. 5 as \"AdaGN\" and \"Cross-Att.\", respectively. Although the original identity is preserved by the AdaGN scheme, the background is not well restored. A possible reason is that the identity-related condition is performed at the latent level only, which fails to provide pixel-wise constraints. On the other hand, despite the cross-attention scheme achieves a comparable restoration result with ours (\"Concat.), the extra layer requires more computational resources." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.539, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Table 2. Ablation studies of the effectiveness of our high-frequency recovery (HFR) module. \\( \\mathbf{x}_H \\) denotes high-frequency sub-bands of LQ images and 1, 2 stands for the level of wavelet decomposition. Best performance is indicated by Red." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.606, + 0.888, + 0.678 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓FID↓
LCD + xH1 + xH224.3530.4300.53717.264
LCD + HFR1 + xH226.0810.6090.48114.020
LCD + xH1 + HFR224.8190.4600.53615.293
LCD + HFR1 + HFR2 (Ours)26.9670.7110.34313.062
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.733, + 0.712 + ], + "angle": 0, + "content": "5.2.2 High-frequency Recovery" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To investigate the effectiveness of our High-frequency Recovery (HFR) module, we replace the recovered high-frequency sub-bands at different DWT levels with that of low-quality (LQ) images for image reconstruction. Quantitative comparisons on the restoration quality of different schemes are reported in Tab. 2. The restoration quality drops significantly when HFR is removed with a decline of \\(2.62\\mathrm{dB}\\) in PSNR. Besides, the lack of recovered high-frequency components at different DWT levels deteriorates the restoration quality to varying extents. In comparison with those at level 2 (\\(\\times 128\\)), high-frequency sub-bands at level 1 (\\(\\times 256\\)) contribute more to the restoration quality." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.096, + 0.462, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.598 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison between images reconstructed at different DWT levels with and without (w/o) HFR module on CelebA-Test. \\( \\pmb{x} \\), \\( \\hat{\\pmb{x}} \\) and \\( \\pmb{y} \\) denotes LQ images, restored images, and HQ images. \\( \\pmb{x}^j (j\\in \\{1,2\\}) \\) denotes frequency components at \\( j \\)-th DWT level. PSNR(↑) of the restored images are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.714 + ], + "angle": 0, + "content": "The exemplar illustrated in Fig. 6 helps to understand intuitively. Before our HFR is applied (Row 2), the restored image contains cluttered noise and lacks details such as hairstyle due to messy high-frequency sub-bands of LQ images. The quality improves when the image is constructed with the refined high-frequency component (Row 3), which is less noisy and presents fine-grained details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.453, + 0.741 + ], + "angle": 0, + "content": "5.3. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.468, + 0.839 + ], + "angle": 0, + "content": "State-of-the-art (SOTA) methods used for comparison include generative prior based methods (GPEN [32] and GFPGAN [27]), reference prior based methods (VQFR [8] and RestoreFormer [29]) and recent diffusion model-based methods (DifFace [34] and DR2 [30]). Evaluations are conducted on both synthetic and real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Synthetic dataset. Quantitative comparison on CelebA-Test [12] are illustrated in Tab. 3. WaveFace achieves the best scores on reference-based metrics and ranks second on FID. Apart from the better restoration quality, our method" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.079, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on CelebA-Test for blind face restoration. \"Deg.\" refers to the angle between identity embeddings of restored images and HQ counterparts. The best and the second best performances are highlighted and underlined." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.145, + 0.897, + 0.297 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑LPIPS↓FID↓Deg↓
Input23.0380.3890.58667.50554.958
GPEN [32]24.3190.6030.44022.25138.473
GFP-GAN [27]24.7710.6740.36114.50136.958
VQFR [8]23.7350.6140.35814.04838.684
RestoreFormer [29]24.1910.6360.36412.00034.752
CodeFormer [38]25.0710.6720.35914.38537.677
DR2 [30]23.5830.6130.40214.67151.030
Difface [34]24.1550.6670.39013.37949.958
WaveFace26.967↑1.900.7110.34313.06228.263↑6.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Table 4. Quantitative comparisons on three real-world datasets (-Test) in terms of FID. The best and the second best performances are highlighted and underlined." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.36, + 0.869, + 0.519 + ], + "angle": 0, + "content": "
MethodsLFWWebPhotoWIDER
Input124.974170.112199.961
GPEN [32]50.79280.57246.340
GFP-GAN [27]49.56087.58439.499
VQFR [8]50.86775.34844.107
RestoreFormer [29]47.75077.33049.817
CodeFormer [38]51.86383.19338.784
DR2 [30]45.298112.34445.348
Difface [34]45.22787.81137.112
WaveFace43.17581.52536.913
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.891, + 0.581 + ], + "angle": 0, + "content": "faithfully preserves the identity with the minimum angle between identity embeddings and their HQ counterparts, outperforming the second-best method by 6.5 degrees." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Additionally, we present the qualitative comparison with SOTA methods on images with increasing degradations in Fig. 7. GFPN and GFPGAN generally produce oversmoothed results. VQFR and RestoreFormer introduce obvious artifacts especially when the inputs are corrupted by mild or severe degradation. Besides, the above methods fail to preserve the identity since they are highly dependent on the StyleGAN prior, where the pre-defined latent space limits the diversity in restored results. Diffusion prior-based methods (DifFace and DR2) tend to generate faithful images with minor artifacts. However, due to the lack of condition during generation, both methods still suffer from the failure in identity preservation. On the contrary, our method, with the help of diffusion prior as well as the injected condition, is capable of generating high-quality facial images with fewer artifacts and, meanwhile, is faithful to the original identity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Real-world datasets. We compare with SOTA methods on three real-world datasets in terms of FID score, which measures the KL divergence between the distribution of restored images and that of HQ images in FFHQ [13]. To" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6589" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.887, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.322, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with state-of-the-art BFR methods on CelebA-Test. Our method achieves a better restoration quality with the original identity and facial details well preserved. (Zoom in for best view)." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.363, + 0.887, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.601, + 0.89, + 0.629 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison with state-of-the-art BFR methods on real-world datasets, including LFW-Test (first row), WebPhoto-Test (second row), and WIDER-Test (third row). (Zoom in for best view)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.822 + ], + "angle": 0, + "content": "reduce the large gap between the synthetic degradations and real-world ones, following previous diffusion prior-based methods [18, 34], a pre-denoising network [18] is applied before our method to handle the unknown degradations. Quantitative results with SOTA methods are reported in Tab. 4. WaveFace outperforms state-of-the-art (SOTA) methods on LFW-Test [11] and WIDER-Test [38] and beats previous diffusion model based BFR methods on WebPhoto-Test [27]. As mentioned in DiffFace [34], the FID score may not be the most suitable evaluation metric for WebPhoto-Test, as a total of 407 images is significantly distant from representing the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Three typical examples are shown in Fig. 8 with the increasing degradation level. It is observed that previous BFR methods tend to generate either hazy or unnatural faces. Some methods even fail to generate adequate restored results when handling severe degradations (Row 3). In con" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.671 + ], + "angle": 0, + "content": "trast, our approach provides much more natural and realistic results with rich details such as wrinkles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.683, + 0.62, + 0.699 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We propose WaveFace to solve the task in the frequency domain to achieve the trade-off between efficiency and authenticity. Efficiency is achieved by applying a diffusion model only on the low-frequency sub-band whose size is 1/16 of the original image. Meanwhile, high-frequency components decomposed at multiple DWT levels are recovered simultaneously by a one-forward framework, which ensures the preservation of facial details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Limitations. There is a considerable gap between real-world degradations and the simulated ones (Eq. (15)), leading to inferior results in some cases. Our future work will focus on how to simulate degradations that fit real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6590" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[3] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. FSRNet: End-to-end learning face super-resolution with facial priors. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.469, + 0.313 + ], + "angle": 0, + "content": "[4] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[6] Jiankang Deng, Anastasios Roussos, Grigorios Chrysos, Evangelos Ververas, Irene Kotsia, Jie Shen, and Stefanos Zafeiriou. The menpo benchmark for multi-pose 2d and 3d facial landmark localisation and tracking. IJCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[8] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In ECCV, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.469, + 0.57 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[11] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Report, 2008. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[12] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[13] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[15] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In ECCV, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[20] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[21] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[24] Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming-Hsuan Yang. Deep semantic face deblurring. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[26] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[27] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[29] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In CVPR, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[30] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. DR2: Diffusion-based robust degradation remover for blind face restoration. In CVPR, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[31] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[32] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[33] Xin Yu, Basura Fernando, Bernard Ghanem, Fatih Porikli, and Richard Hartley. Face super-resolution guided by facial component heatmaps. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[34] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv:2212.06512, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6591" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[37] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[38] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6592" + } + ] +] \ No newline at end of file diff --git a/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_origin.pdf b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a3c3f3c826c2c98a05e35427eb27a685883183c --- /dev/null +++ b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/8a8efe74-fca8-4300-90b6-5878a71afa4d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a37bb56c96d2a10d3ce597d6fef460460f910fcc172dc705560c2890e87d210 +size 8147580 diff --git a/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/full.md b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c908af2f7b5fe26f5599ee6dc170b3f1d4ba967f --- /dev/null +++ b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/full.md @@ -0,0 +1,351 @@ +# WaveFace: Authentic Face Restoration with Efficient Frequency Recovery + +Yunqi Miao +University of Warwick +Yunqi.Miao.1@warwick.ac.uk + +Jiankang Deng* +Imperial College London +jiankangdeng@gmail.com + +Jungong Han* +University of Sheffield +jungonghan77@gmail.com + +![](images/17afac7aaab6dd92cbbefd6eb67a843ccb63900741fd73dc675906c66d3a8544.jpg) +Figure 1. Left: Illustration of our frequency-aware BFR scheme. Restoration is performed in the frequency domain instead of the pixel domain. Right: Comparisons with state-of-the-art face restoration methods on degraded images. Previous methods struggle to restore facial details or the original identity while our WaveFace achieves a good balance of realism and fidelity with fewer artifacts. + +![](images/b424b1262fcce6b66248cf1572b667955d956d962ed383b8759a6974b5c9e05a.jpg) + +# Abstract + +Although diffusion models are rising as a powerful solution for blind face restoration, they are criticized for two problems: 1) slow training and inference speed, and 2) failure in preserving identity and recovering fine-grained facial details. In this work, we propose WaveFace to solve the problems in the frequency domain, where low- and high-frequency components decomposed by wavelet transformation are considered individually to maximize authenticity as well as efficiency. The diffusion model is applied to recover the low-frequency component only, which presents general information of the original image but 1/16 in size. To preserve the original identity, the generation is conditioned on the low-frequency component of low-quality images at each denoising step. Meanwhile, high-frequency components at multiple decomposition levels are handled by a unified network, which recovers complex facial details in a single step. Evaluations on four benchmark datasets show that: 1) WaveFace outperforms state-of-the-art methods in authen + +ticity, especially in terms of identity preservation, and 2) authentic images are restored with the efficiency $10 \times$ faster than existing diffusion model-based BFR methods. + +# 1. Introduction + +Blind face restoration (BFR) aims to recover high-quality (HQ) facial images from various degradations, including down-sampling, blurriness, noise, and compression artifact [1, 26, 35]. BFR is challenging since the types of degradation are generally unknown in real-world scenarios. Improvements in the restoration quality over the past few years mainly benefit from the usage of multiple facial priors. Geometric priors, including facial landmarks [3] and parsing maps [2, 24], are employed to provide explicit facial structure information. Reference priors, such as HQ images/features [8, 29, 38], are used as guidance during facial restoration. Generative priors are generally obtained by pre-trained generation models such as StyleGAN [13], + +which facilitate the recovery of realistic textures [2, 27, 32]. Inspired by the superior generative ability of diffusion models (DMs) [10, 25], how to unleash its potential in authentic face restoration has gained much attention. + +Despite plausible restoration results being achieved, existing DMs based methods [30, 34] generally suffer from two problems: 1) DMs are trained on large images $(512 \times 512)$ , which requires massive computational resources for both training and inference, as thousands of iterations are required to obtain adequate outputs. Difface [34], for example, takes about 25 seconds to sample an image from noise. 2) Low-quality (LQ) images are firstly dif- fused to an intermediate step and then denoised by an unconditional diffusion model to obtain HQ counterparts [34]. However, as shown in Fig. 1, such an unconditional generation will lead to great uncertainty in restored images, which means the original identity and facial details such as wrinkles cannot be well preserved. + +To solve the above problems, we transfer the BFR task from the pixel domain to the frequency domain via Discrete Wavelet Transform (DWT). As shown in Fig. 1, an image can be decomposed and reconstructed by its four quarter-sized sub-bands: low- and high-frequency components without sacrificing information. Low-frequency sub-band mainly contains general information such as face structure while high-frequency ones contain rich facial details. By exponentially shrinking the size of input images, restoration within the frequency domain not only speeds up both training and inference but also maintains the authenticity of restoration. In the paper, we devise an efficient BFR method WaveFace that restores authentic HQ images by recovering its frequency components. + +WaveFace consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Recovery (HFR) module. A diffusion model is used in LCD to restore the low-frequency sub-band of HQ images whose size is only 1/16 of the original image. In addition to smaller inputs, LCD leverages LQ counterparts as a condition throughout the generation to preserve the original identity. Meanwhile, HFR recovers high-frequency sub-bands decomposed at multiple DWT levels simultaneously within one step. With the frequency components restored by two modules, authentic images can be reconstructed via discrete inverse wavelet transform (IWT) within 1 ms. The contributions can be summarized as follows: + +- We propose an efficient blind face restoration approach, WaveFace, that restores authentic images by recovering their frequency components individually. +- A conditional diffusion model is adopted to restore the low-frequency component, which is 1/16 the size of the original image. +- A one-pass network is used to recover high-frequency sub-bands decomposed at multiple DWT levels simul + +taneously. + +- Comprehensive experiments demonstrate the superiority of methods in both efficiency and authenticity. + +# 2. Related Work + +# 2.1. Diffusion models + +Diffusion Models (DMs) are emerging generative models that corrupt the data with the successive addition of Gaussian noise during the diffusion process and then learn to recover the data during denoising. State-of-the-art DMs [10] have revealed the potential in CV tasks, such as image deblurring [4, 15] and image super-resolution [22, 23]. + +Accelerating DMs. Despite being powerful in generation, DDPM [10] has the downside of low inference speed, which requires thousands of steps for sampling. To accelerate inference, DDIM [19] adopts a non-Markovian diffusion process that allows step skipping during sampling. + +Conditional DMs. Unconditional DM-based generation leads to great uncertainty in outputs, which means they generally fail to preserve the original identity and fine-grained facial details. Therefore, conditions are injected by cross-attention layer [22], adaptive normalization layer [21], or concatenation operation [23] to control the characteristic of the synthesized images. + +# 2.2. Blind face restoration + +Facial priors exploited in blind face restoration (BFR) can be categorized into three types: geometric priors, reference priors, and generative priors. + +Geometric priors based methods leverage knowledge from facial landmark [3, 6], facial parsing maps [2, 24, 37], and facial component heatmaps [33]. However, they tend to show inferior performance since degraded images fail to provide accurate and adequate structural information. + +Reference priors based methods either leverage a high-quality (HQ) reference image sharing the same identity as the degraded one [17] or a pre-constructed dictionary storing HQ facial features [8, 29, 38]. Firstly, a vectorquantized (VQ) codebook is pre-trained on HQ faces via VQ-GAN [7] to provide rich facial details. Features of degraded inputs are then fused with the prior either at image [29] or latent [8, 38] level for the generation of HQ counterparts. However, reference prior is restricted by the size of the codebook, which limits the diversity and richness of generated images. + +Generative priors encapsulated in pretrained face models [13, 14] are also leveraged to restore faithful faces [2, 27, 32]. PSFRGAN [2] modulates features at different scales progressively with facial parsing maps to achieve semantic-aware style transformation. GFP-GAN [27] and GPEN [32] adopt the pre-trained StyleGAN as a decoder and achieve a good balance between visual quality and fidelity of re + +stored images. Recently, diffusion models [10, 25] have shown the powerful generative ability in restoring HQ content from noisy images. Inspired by the ability, DifFace [34] firstly maps low-quality (LQ) inputs into an intermediate step of the denoising process, from which its HQ counterpart is recursively sampled. DR2 [30] employs the diffusion model to remove degradations, followed by a super-resolution model to obtain HQ counterparts. + +Nevertheless, previous DM-based BFR methods are performed in the pixel domain, where large inputs require massive computational resources for both training and inference. Besides, they both apply the unconditional scheme that leads to significant uncertainty in restored results. + +# 3. Preliminary + +Discrete Wavelet Transformation (DWT). DWT decomposes an image into low-frequency and high-frequency subbands. The low-frequency component presents general information while high-frequency ones express facial details in the vertical, horizontal, and diagonal directions. In the paper, we use Haar wavelet for DWT, which is widely used in real-world applications due to simplicity [20]. + +Given an image $\pmb{x} \in \mathbb{R}^{H \times W \times 3}$ , its low-frequency subband $\pmb{x}_{ll}^{1} \in \mathbb{R}^{H/2 \times W/2 \times 3}$ and high-frequency sub-bands $\pmb{x}_{lh}^{1}, \pmb{x}_{hl}^{1}$ , and $\pmb{x}_{hh}^{1} \in \mathbb{R}^{H/2 \times W/2 \times 3}$ can be decomposed by: + +$$ +\boldsymbol {x} _ {l l} ^ {1}, \boldsymbol {x} _ {l h} ^ {1}, \boldsymbol {x} _ {h l} ^ {1}, \boldsymbol {x} _ {h h} ^ {1} = D W T (\boldsymbol {x}), \tag {1} +$$ + +where $DWT(\cdot)$ refers to the DWT operation. Despite that the input size is decreased by a factor of four $(H / 2\times W / 2)$ after a single DWT decomposition, DWT can be performed on the low-frequency component to further reduce the computational cost and expedite the inference: + +$$ +\boldsymbol {x} _ {l l} ^ {J + 1}, \boldsymbol {x} _ {l h} ^ {J + 1}, \boldsymbol {x} _ {h l} ^ {J + 1}, \boldsymbol {x} _ {h h} ^ {J + 1} = D W T (\boldsymbol {x} _ {l l} ^ {J}), \tag {2} +$$ + +where $J$ is level of DWT decomposition and $\pmb{x}_{ll}^{J + 1},\pmb{x}_{lh}^{J + 1},\pmb{x}_{hl}^{J + 1},\pmb{x}_{hh}^{J + 1}\in \mathbb{R}^{H / 2^{(J + 1)}\times W / 2^{(J + 1)}\times 3}$ . Reversibly, given the frequency sub-bands, an image can be reconstructed via discrete inverse wavelet transform (IWT): + +$$ +\boldsymbol {x} = I W T \left(\boldsymbol {x} _ {l l} ^ {1}, \boldsymbol {x} _ {l h} ^ {1}, \boldsymbol {x} _ {h l} ^ {1}, \boldsymbol {x} _ {h h} ^ {1}\right). \tag {3} +$$ + +Diffusion Models. Training of diffusion models (DMs) consists of a diffusion process and a denoising process. Diffusion process transforms an image from the real data distribution $\pmb{y}_0\sim q(\pmb{y}_0)$ into a pure Gaussian noise $\pmb{y}_T$ by successively applying the following Markov diffusion kernel: + +$$ +q \left(\boldsymbol {y} _ {t} \mid \boldsymbol {y} _ {t - 1}\right) = \mathcal {N} \left(\boldsymbol {y} _ {t}; \sqrt {1 - \beta_ {t}} \boldsymbol {y} _ {t - 1}, \beta_ {t} \boldsymbol {I}\right), \tag {4} +$$ + +where $\{\beta_t\}_{t=1}^T$ is a pre-defined or learned noise variance schedule. The marginal distribution at arbitrary timestep $t$ can be denoted as: + +$$ +q \left(\boldsymbol {y} _ {t} \mid \boldsymbol {y} _ {0}\right) = \mathcal {N} \left(\boldsymbol {y} _ {t}; \sqrt {\alpha_ {t}} \boldsymbol {y} _ {0}, (1 - \alpha_ {t}) \boldsymbol {I}\right), \tag {5} +$$ + +where $\alpha_{t} = \prod_{s=1}^{t}(1 - \beta_{s})$ . Given $\pmb{y}_{t}$ , the denoising process aims to recover $\pmb{y}_{0}$ by recursively learning the transition from $\pmb{y}_{t-1}$ to $\pmb{y}_{t}$ , which is defined as the following Gaussian distribution: + +$$ +p _ {\theta} \left(\boldsymbol {y} _ {t - 1} \mid \boldsymbol {y} _ {t}\right) = \mathcal {N} \left(\boldsymbol {y} _ {t - 1}; \boldsymbol {\mu} _ {\theta} \left(\boldsymbol {y} _ {t}, t\right), \boldsymbol {\Sigma} _ {\theta} \left(\boldsymbol {y} _ {t}, t\right)\right), \tag {6} +$$ + +where parameters $\theta$ are optimized by a denoising network $\epsilon_{\theta}$ that predicts $\pmb{\mu}_{\theta}(\pmb{y}_t,t)$ and $\pmb{\Sigma}_{\theta}(\pmb{y}_t,t)$ . + +# 4. Methodology + +In the paper, WaveFace is proposed to handle blind face restoration (BFR) in the frequency domain. The framework is depicted in Fig. 2. First, degraded images are mapped to the frequency domain via Discrete Wavelet Transformation (DWT) (Sec. 3). The proposed Low-frequency Conditional Denoising (LCD) module (Sec. 4.1) and High-Frequency Recovery (HFR) module (Sec. 4.2) are adopted on the low- and high-frequency components respectively to remove degradations and restore facial details. Recovered frequency components are used for the image reconstruction. + +# 4.1. Low-frequency Conditional Denoising + +As shown in Fig. 1, the low-frequency component is similar to the down-sampled version of the original image, which largely determines the restoration quality. Due to its powerful generative ability from noisy inputs, a diffusion model (DM) is adopted in LCD to restore the low-frequency subband of high-quality (HQ) images. + +Conditional DM. We denote the low-frequency sub-band of a pair of LQ and HQ images as $(\pmb{x}_{ll_0}^j,\pmb{y}_{ll_0}^j)$ . Since the low-frequency sub-band only contains a single map after DWT, the DWT levels $j$ and $ll$ are omitted in this section for simplicity. According to the diffusion process defined in Eq. (5), images are successively destroyed by Gaussian noise as timestep $t$ increases, which means the distribution $q(\pmb{x}_T|\pmb{x}_0)\approx q(\pmb{y}_T|\pmb{y}_0)\approx \mathcal{N}(0,\pmb {I})$ after a large timestep $T$ . Given this assumption, our LCD aims to learn the posterior distribution $p(\pmb{y}_0|\pmb{x}_0)$ : + +$$ +p \left(\boldsymbol {y} _ {0} \mid \boldsymbol {x} _ {0}\right) = \int q \left(\boldsymbol {y} _ {T} \mid \boldsymbol {x} _ {0}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\boldsymbol {y} _ {t - 1} \mid \boldsymbol {y} _ {t}\right) \mathrm {d} \boldsymbol {y} _ {1: T}, \tag {7} +$$ + +where $p_{\theta}(\boldsymbol{y}_{t-1}|\boldsymbol{y}_t)$ refers to the unconditional denoising process defined in Eq. (6). However, the unconditional scheme fails to preserve the original identity due to the lack of guidance during generation. To solve the problem, the low-frequency sub-band of LQ images $\boldsymbol{x}_0$ are injected as the condition for the denoising process: + +$$ +p _ {\theta} \left(\boldsymbol {y} _ {t - 1} \mid \boldsymbol {y} _ {t}, \boldsymbol {x} _ {0}\right) = \mathcal {N} \left(\boldsymbol {y} _ {t - 1}; \boldsymbol {\mu} _ {\theta} \left(\boldsymbol {y} _ {t}, t, \boldsymbol {x} _ {0}\right), \boldsymbol {\Sigma} _ {\theta} \left(\boldsymbol {y} _ {t}, t, \boldsymbol {x} _ {0}\right)\right), \tag {8} +$$ + +![](images/b5f8ce459f818d7d6b4db2e4fe1b7817f56c1a3bb7f0098d8ee84f65f4b5e003.jpg) +Figure 2. Overall framework of WaveFace. It consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Refinement (HFR) module. LCD (Sec. 4.1) predicts clean samples $\pmb{x}_0$ from noise conditioned on LQ inputs through $T$ steps. Meanwhile, high-frequency sub-bands are concatenated as HFR (Sec. 4.2) inputs to recover vivid facial details. The predicted frequency components are projected back to the image via IWT. + +where $\pmb{x}_0$ is injected by concatenating with the input $y_{t}$ along the channel dimension. + +Objectives. Following DDPM [10], the denoising network $\epsilon_{\theta}$ is trained to predict noise vectors with the objective: + +$$ +L _ {L C D} = \mathbb {E} _ {\boldsymbol {y} _ {0}, t, \boldsymbol {\epsilon} _ {t} \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I})} \left[ | | \boldsymbol {\epsilon} _ {t} - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {y} _ {t}, \boldsymbol {x} _ {0}, t) | | ^ {2} \right]. \tag {9} +$$ + +During inference, with the learned parameterized Gaussian transitions $p_{\theta}(\pmb{y}_{t-1}|\pmb{y}_t,\pmb{x}_0)$ , the low-frequency sub-band of HQ images can be recovered from a random Gaussian noise $\pmb{y}_T \sim \mathcal{N}(\pmb{0},\mathbf{I})$ by recursively applying: + +$$ +\boldsymbol {y} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {y} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {y} _ {t}, \boldsymbol {x} _ {0}, t\right)\right) + \sigma_ {t} \boldsymbol {z}, \tag {10} +$$ + +where $z\sim \mathcal{N}(\mathbf{0},\mathbf{I}),\alpha_{t} = 1 - \beta_{t}$ , and $\bar{\alpha}_t = \prod_{i = 1}^t\alpha_i$ + +Trade-off between efficiency and quality. Training a DM on large inputs ( $\times 512$ or more) requires massive computational resources and the evaluation takes thousands of steps to sample from the noise. Although this problem can be alleviated by adopting low-frequency sub-bands at higher DWT levels, multiple times of decomposition will also lead to information reduction. How to set DWT level wisely is the key to balance between efficiency and authenticity. + +To illustrate how DWT level $J$ affects the authenticity, we visualize the image reconstructed by the low-frequency component of HQ images and the corresponding high-frequency ones of its LQ counterpart at different DWT levels in Fig. 3. Although noisy at the high-frequency part, the quality of images reconstructed at $J \leq 2$ is acceptable, with key facial features such as wrinkles still preserved. However, the mosaic effect starts to emerge when $J$ keeps increasing. Ablation studies about the effect of DWT level on + +training/inference are shown in Sec. 5.2.1. In the paper, $J$ is set as 2 empirically. + +# 4.2. High-frequency Recovery + +Although the general face information is restored by LCD, rich facial details are generally embedded in the high-frequency component, as shown in Fig. 1, which improves the authenticity of restored images. High-frequency sub-bands at multiple DWT levels also vary in size, which means more than one diffusion model is required for the recovery. To avoid the extravagant computational cost, we adopt a U-shaped network that can recover high-frequency sub-bands at multiple DWT levels at the same time and takes only one step for inference. + +Given a LQ and HQ image pair $(\pmb{x},\pmb{y})$ and their high-frequency sub-bands: $\pmb{x}_H^j = \{\pmb{x}_{lh}^j,\pmb{x}_{hl}^j,\pmb{x}_{hh}^j\}$ and $\pmb{y}_H^j = \{\pmb{y}_{lh}^j,\pmb{y}_{hl}^j,\pmb{y}_{hh}^j\}$ , where $j\in \{1,2\}$ refers to DWT level, the framework our high-frequency recovery (HFR) module is illustrated in Fig. 4. HFR takes high-frequency sub-bands of LQ images that are channel-wise concatenated as inputs and outputs the recovered ones $\hat{\pmb{x}}_H^j$ as follows: + +$$ +\boldsymbol {F} _ {H} ^ {1}, \boldsymbol {F} _ {H} ^ {2} = \mathcal {U} _ {\phi_ {u}} \left(\mathcal {H} _ {\phi_ {i n}} \left(\boldsymbol {x} _ {H} ^ {1}\right), \mathcal {F} _ {\phi_ {f}} \left(\boldsymbol {x} _ {H} ^ {2}\right)\right), \tag {11} +$$ + +$$ +\hat {\boldsymbol {x}} _ {H} ^ {j} = \mathcal {H} _ {\phi_ {o u t} ^ {j}} \left(\boldsymbol {F} _ {H} ^ {j}\right) + \boldsymbol {x} _ {H} ^ {j}, \quad j \in \{1, 2 \}, \tag {12} +$$ + +where $\mathcal{U}(\cdot),\mathcal{H}(\cdot),\mathcal{F}(\cdot)$ denotes the U-shaped network, the convolutional layer, and the fusion module with the parameter $\phi_u,\phi_{in / out}^j$ , and $\phi_f$ , respectively. $F_H^1,F_H^2$ represents feature maps of the last and penultimate decoding layers of the U-shaped network. Considering efficiency, the fusion module only contains two stacked ConvBlocks that apply $3\times 3$ and $1\times 1$ convolutional layers with ReLU in between, followed by concatenation with the input. + +![](images/114fde351da94fa3f60faf34eea20eda9db80c421ed7b7cac1b61241e8320b30.jpg) +(a) $J = 3$ $(\times 64)$ + +![](images/c3d5598ee903e6a11b0340d42609b1f08093e46f8fc8922d38234f8d8b6499da.jpg) +Figure 3. Visualization of DWT frequency components and images reconstructed by the low-frequency sub-band of an HQ image and high-frequency sub-bands of its LQ counterpart. DWT level $J$ and resolution of low- / high-frequency sub-bands $(\times N)$ are reported. + +![](images/6d1f9177ad887155a1b1d36b68469e731074220b7c0423160ec11e88627f56a1.jpg) +(b) $J = 2$ （ $\times 128)$ + +![](images/bd60efaf13450e8c213049271a62f5611cd57ccda4045978fa6e9b88b341449b.jpg) + +$(\times 128)$ + +![](images/613ca2f7a939901f92019de08fcfa15a9b3b42cfca1ec63143fd13d839fda834.jpg) +(c) $J = 1$ （20 $(\times 256)$ + +![](images/ab4f1e55665da1f37183e70eaef116ff997bc4237becd24ee06c621fe2b33a26.jpg) + +![](images/f6ec98b1931a5f4267ce0a7a5b30497351412b4d6d2941605e8388e1ad980ccf.jpg) +Figure 4. Illustration of high-frequency recovery (HFR) module. + +Objectives. We adopt two loss terms for the training of HFR: recovery loss $\mathcal{L}_r$ and content loss [28] $\mathcal{L}_c$ . The recovery loss is applied not only to restore HQ high-frequency sub-bands $\pmb{y}_H^j$ at each level but also to the image $\pmb{y}^{j - 1}$ reconstructed via inverse wavelet transform (IWT): + +$$ +\mathcal {L} _ {r} = \sum_ {j} \left( \right.\left|\left| \hat {\boldsymbol {x}} _ {H} ^ {j} - \boldsymbol {y} _ {H} ^ {j} \right|\right| + \alpha \left. \right\rvert\left| I W T \left(\boldsymbol {y} _ {l l} ^ {j}, \hat {\boldsymbol {x}} _ {H} ^ {j}\right) - \boldsymbol {y} ^ {j - 1} \right|\left. \right), \tag {13} +$$ + +where $IWT(\cdot)$ denotes IWT operation and $\alpha = 1$ is the parameter balancing losses between the frequency domain and the pixel domain. In addition to minimizing the pixel-wise distance, the content loss is applied to maximize the luminance, contrast, and structural similarity of the restored and original HQ image: + +$$ +\mathcal {L} _ {c} = 1 - S S I M \left(I W T \left(\boldsymbol {y} _ {l l} ^ {2}, \hat {\boldsymbol {x}} _ {H} ^ {2}, \hat {\boldsymbol {x}} _ {H} ^ {1}\right), \boldsymbol {y}\right). \tag {14} +$$ + +The overall objective of HFR is $\mathcal{L}_{HFR} = \mathcal{L}_r + \lambda \mathcal{L}_c$ , where the weighting parameter $\lambda$ is set as 10. + +# 5. Experiments + +# 5.1. Settings and Datasets + +Training Dataset. FFHQ [13] is used as the training set, which contains 70,000 high-quality (HQ) face images. Following BFR benchmark works [8, 29], we resize HQ images to the resolution of $512 \times 512$ , and then synthesize low-quality (LQ) counterparts as follows: + +$$ +\boldsymbol {x} = \left\{\left[ (\boldsymbol {y} \otimes \boldsymbol {k} _ {\sigma}) \downarrow_ {s} + \boldsymbol {n} _ {\delta} \right] _ {\mathrm {J P E G} _ {q}} \right\} \uparrow_ {s}, \tag {15} +$$ + +where a HQ image $\pmb{y}$ is firstly blurred by a Gaussian kernel $k_{\sigma}$ , followed by a downsampling of scale $s$ . Afterward, Gaussian noise $\pmb{n}_{\delta}$ and JPEG compression with quality factor $q$ are applied to the image, which is then upsampled back to the original size to obtain its LQ counterpart $\pmb{x}$ . The hyper-parameters $\sigma, s, \delta$ , and $q$ are uniformly sampled from [0.1, 15], [0.8, 32], [0, 20], and [30, 95] respectively. + +Testing Dataset. We evaluate WaveFace on a synthetic dataset: CelebA-Test and three real-world datasets: LFW-Test [11], WebPhoto-Test [27], and WIDER-Test [38]. CelebA-Test contains 3000 HQ images from CelebAHQ [12], and LQ counterparts are synthesized via Eq. (15) with the same degradation setting. In terms of three realworld datasets, LFW-Test contains 1711 mildly degraded face images in the wild, which comprises the first image for each person in LFW [11]. WebPhoto-Test includes 407 images crawled from the internet, some of which are old photos with severe degradation. WIDER-Test consists of 970 images with severe degradations from the WIDER dataset [31]. + +Evaluation Metrics. For evaluation, we adopt two pixelwise metrics (PSNR and SSIM), a reference perceptual metric (LPIPS [36]), and a non-reference perceptual metric (FID [9]). To measure the consistency of identity, the angle between embeddings extracted by ArcFace ("Deg.") [5] is used. All metrics are used for the evaluation of synthetic data while only FID is used for real-world datasets due to the lack of referential HQ images. + +Implementation Details. The DWT decomposition level is set as 2, which reduces the input size of the diffusion model from $512 \times 512$ to $128 \times 128$ . We train LCD and HFR individually with Adam [16] optimizer. LCD is trained for 200K iterations with the learning rate $1e-4$ and batch size 32. For HFR, the training takes 70K iterations. The learning rate gradually decays from $1e-3$ to $1e-5$ by a factor of 0.1. + +# 5.2. Ablation Studies + +# 5.2.1 Low-frequency Conditional Denoising + +DWT Levels. We conduct an experiment to investigate how the level of wavelet decomposition (DWT) affects the efficiency and authenticity. DWT is applied 0, 1, 2, 3 times on + +Table 1. Comparison between state-of-the-art diffusion model-based BFR methods and ours at different DWT levels in terms of efficiency and authenticity. The best and the second best performances are highlighted and underlined. + +
MethodLevelResolution#Param. (M)Time (s)PSNR↑SSIM↑
DifFace [34]-× 512175.3825.0419.060.46
DR2 [30]-× 25693.568.7622.890.57
WaveFace0× 512109.0619.3726.420.65
1× 25671.418.1227.080.73
2× 12825.361.9726.970.71
3× 643.91.4624.810.46
+ +the training set, where the resolution of the diffusion model (DM)'s input becomes $512 \times 512$ , $256 \times 256$ , $128 \times 128$ , and $64 \times 64$ , respectively. The number of parameters and the inference time are used to evaluate the efficiency while PSNR and SSIM are used to evaluate authenticity. The comparison results are reported in Tab. 1. In comparison with the model trained on original images ( $\times 512$ ), our frequency-aware scheme applies DM on the low-frequency sub-band only, which avoids some unknown noise on high-frequency components. Thus, the restoration quality improves after the wavelet decomposition is applied once. The quality can be maintained until the DWT level increases to 2, where the inference process is accelerated by $10 \times$ . However, the performance is significantly harmed when the DWT level further increases (3 or higher) as too many details in the low-frequency component have been reduced. DWT level is set as 2 throughout the experiments to achieve the tradeoff between the efficiency and the quality of restoration. + +We also compare the efficiency with two DM-based BFR methods: Difface [34] and DR2 [30]. To make a fair comparison, the inference time refers to the time of the whole denoising process, which starts from the pure Gaussian noise. The quality of generated images cannot be compared with that reported in Tab. 3 as images are generated with pre-/post-processing and fewer sampling steps in their methods. As can be seen, our method achieves better restoration quality at a lower computational cost. + +Conditions. To investigate how the condition facilitates the restoration, we first remove the condition from the input of our LCD module and train an unconditional diffusion model for 200K iterations for a fair comparison. The restored face is denoted by "Uncond." in Fig. 5. We notice that not only the identity is not preserved but also the image quality drops significantly. This could be attributed to the slow convergence of the unconditional training manner [21]. + +Additionally, we compare with two widely-used conditioning schemes: adaptive group normalization layers (AdaGN) [21] and cross-attention [22]. We replace the low-frequency sub-band of LQ images with the corresponding identity embedding extracted by Arcface [5], which are subsequently injected via AdaGN. Furthermore, a cross-attention layer is added to condition the denoised image + +![](images/bbbe31da901b1875dd50f21c649ad29f723fc671c25ae9e198af80e6046e6c27.jpg) + +![](images/44b71f2dc83ba4a498e4520dcccb4005e7dd1a58f792aa0046729979c563c423.jpg) + +![](images/e2504fc39c96a2f24c25d36881b1778a9abffbafdcb406a0c30a2443d0f0eac3.jpg) + +![](images/4fe4353810bbc3becc9869695d331900b2f7e69496f5c0f824a6184c275e4d8f.jpg) +Figure 5. Qualitative comparison of different conditioning schemes on CelebA-Test. We adopt "Concat." in WaveFace. PSNR(↑) / Deg.(↓) are reported. + +![](images/7ac1b2ed86629e988223fb8c18fa65f974a84bcad6295f801712f1af72a2d11b.jpg) + +![](images/43d1c414e1d67135e0dbebd5aecfcbff4c7bceda8fa08ff17b410586f0e05b9c.jpg) + +at each step with the LQ low-frequency information. Experimental details of the two schemes will be discussed in supplementary materials. The restored images are shown in Fig. 5 as "AdaGN" and "Cross-Att.", respectively. Although the original identity is preserved by the AdaGN scheme, the background is not well restored. A possible reason is that the identity-related condition is performed at the latent level only, which fails to provide pixel-wise constraints. On the other hand, despite the cross-attention scheme achieves a comparable restoration result with ours ("Concat.), the extra layer requires more computational resources. + +Table 2. Ablation studies of the effectiveness of our high-frequency recovery (HFR) module. $\mathbf{x}_H$ denotes high-frequency sub-bands of LQ images and 1, 2 stands for the level of wavelet decomposition. Best performance is indicated by Red. + +
PSNR↑SSIM↑LPIPS↓FID↓
LCD + xH1 + xH224.3530.4300.53717.264
LCD + HFR1 + xH226.0810.6090.48114.020
LCD + xH1 + HFR224.8190.4600.53615.293
LCD + HFR1 + HFR2 (Ours)26.9670.7110.34313.062
+ +# 5.2.2 High-frequency Recovery + +To investigate the effectiveness of our High-frequency Recovery (HFR) module, we replace the recovered high-frequency sub-bands at different DWT levels with that of low-quality (LQ) images for image reconstruction. Quantitative comparisons on the restoration quality of different schemes are reported in Tab. 2. The restoration quality drops significantly when HFR is removed with a decline of $2.62\mathrm{dB}$ in PSNR. Besides, the lack of recovered high-frequency components at different DWT levels deteriorates the restoration quality to varying extents. In comparison with those at level 2 ( $\times 128$ ), high-frequency sub-bands at level 1 ( $\times 256$ ) contribute more to the restoration quality. + +![](images/71d4eb0dfc89f77df6ea96380a7f57f7246846329453f8597aa00520ca8bc64d.jpg) +Figure 6. Qualitative comparison between images reconstructed at different DWT levels with and without (w/o) HFR module on CelebA-Test. $\pmb{x}$ , $\hat{\pmb{x}}$ and $\pmb{y}$ denotes LQ images, restored images, and HQ images. $\pmb{x}^j (j\in \{1,2\})$ denotes frequency components at $j$ -th DWT level. PSNR(↑) of the restored images are reported. + +The exemplar illustrated in Fig. 6 helps to understand intuitively. Before our HFR is applied (Row 2), the restored image contains cluttered noise and lacks details such as hairstyle due to messy high-frequency sub-bands of LQ images. The quality improves when the image is constructed with the refined high-frequency component (Row 3), which is less noisy and presents fine-grained details. + +# 5.3. Comparisons with State-of-the-Art Methods + +State-of-the-art (SOTA) methods used for comparison include generative prior based methods (GPEN [32] and GFPGAN [27]), reference prior based methods (VQFR [8] and RestoreFormer [29]) and recent diffusion model-based methods (DifFace [34] and DR2 [30]). Evaluations are conducted on both synthetic and real-world datasets. + +Synthetic dataset. Quantitative comparison on CelebA-Test [12] are illustrated in Tab. 3. WaveFace achieves the best scores on reference-based metrics and ranks second on FID. Apart from the better restoration quality, our method + +Table 3. Quantitative comparison on CelebA-Test for blind face restoration. "Deg." refers to the angle between identity embeddings of restored images and HQ counterparts. The best and the second best performances are highlighted and underlined. + +
MethodsPSNR↑SSIM↑LPIPS↓FID↓Deg↓
Input23.0380.3890.58667.50554.958
GPEN [32]24.3190.6030.44022.25138.473
GFP-GAN [27]24.7710.6740.36114.50136.958
VQFR [8]23.7350.6140.35814.04838.684
RestoreFormer [29]24.1910.6360.36412.00034.752
CodeFormer [38]25.0710.6720.35914.38537.677
DR2 [30]23.5830.6130.40214.67151.030
Difface [34]24.1550.6670.39013.37949.958
WaveFace26.967↑1.900.7110.34313.06228.263↑6.49
+ +Table 4. Quantitative comparisons on three real-world datasets (-Test) in terms of FID. The best and the second best performances are highlighted and underlined. + +
MethodsLFWWebPhotoWIDER
Input124.974170.112199.961
GPEN [32]50.79280.57246.340
GFP-GAN [27]49.56087.58439.499
VQFR [8]50.86775.34844.107
RestoreFormer [29]47.75077.33049.817
CodeFormer [38]51.86383.19338.784
DR2 [30]45.298112.34445.348
Difface [34]45.22787.81137.112
WaveFace43.17581.52536.913
+ +faithfully preserves the identity with the minimum angle between identity embeddings and their HQ counterparts, outperforming the second-best method by 6.5 degrees. + +Additionally, we present the qualitative comparison with SOTA methods on images with increasing degradations in Fig. 7. GFPN and GFPGAN generally produce oversmoothed results. VQFR and RestoreFormer introduce obvious artifacts especially when the inputs are corrupted by mild or severe degradation. Besides, the above methods fail to preserve the identity since they are highly dependent on the StyleGAN prior, where the pre-defined latent space limits the diversity in restored results. Diffusion prior-based methods (DifFace and DR2) tend to generate faithful images with minor artifacts. However, due to the lack of condition during generation, both methods still suffer from the failure in identity preservation. On the contrary, our method, with the help of diffusion prior as well as the injected condition, is capable of generating high-quality facial images with fewer artifacts and, meanwhile, is faithful to the original identity. + +Real-world datasets. We compare with SOTA methods on three real-world datasets in terms of FID score, which measures the KL divergence between the distribution of restored images and that of HQ images in FFHQ [13]. To + +![](images/cbc65886e4c29313ea0ef29234c66e271dd66cfcd2d02542e44137ac495e12f8.jpg) +Figure 7. Qualitative comparison with state-of-the-art BFR methods on CelebA-Test. Our method achieves a better restoration quality with the original identity and facial details well preserved. (Zoom in for best view). + +![](images/416d300d67e0ad204adef5a921a2582887f3811fe53647558c22a702d7a9de81.jpg) +Figure 8. Qualitative comparison with state-of-the-art BFR methods on real-world datasets, including LFW-Test (first row), WebPhoto-Test (second row), and WIDER-Test (third row). (Zoom in for best view). + +reduce the large gap between the synthetic degradations and real-world ones, following previous diffusion prior-based methods [18, 34], a pre-denoising network [18] is applied before our method to handle the unknown degradations. Quantitative results with SOTA methods are reported in Tab. 4. WaveFace outperforms state-of-the-art (SOTA) methods on LFW-Test [11] and WIDER-Test [38] and beats previous diffusion model based BFR methods on WebPhoto-Test [27]. As mentioned in DiffFace [34], the FID score may not be the most suitable evaluation metric for WebPhoto-Test, as a total of 407 images is significantly distant from representing the data distribution. + +Three typical examples are shown in Fig. 8 with the increasing degradation level. It is observed that previous BFR methods tend to generate either hazy or unnatural faces. Some methods even fail to generate adequate restored results when handling severe degradations (Row 3). In con + +trast, our approach provides much more natural and realistic results with rich details such as wrinkles. + +# 6. Conclusion + +We propose WaveFace to solve the task in the frequency domain to achieve the trade-off between efficiency and authenticity. Efficiency is achieved by applying a diffusion model only on the low-frequency sub-band whose size is 1/16 of the original image. Meanwhile, high-frequency components decomposed at multiple DWT levels are recovered simultaneously by a one-forward framework, which ensures the preservation of facial details. + +Limitations. There is a considerable gap between real-world degradations and the simulated ones (Eq. (15)), leading to inferior results in some cases. Our future work will focus on how to simulate degradations that fit real-world scenarios. + +# References + +[1] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 1 +[2] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In CVPR, 2021. 1, 2 +[3] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. FSRNet: End-to-end learning face super-resolution with facial priors. In CVPR, 2018. 1, 2 +[4] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 2 +[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5, 6 +[6] Jiankang Deng, Anastasios Roussos, Grigorios Chrysos, Evangelos Ververas, Irene Kotsia, Jie Shen, and Stefanos Zafeiriou. The menpo benchmark for multi-pose 2d and 3d facial landmark localisation and tracking. IJCV, 2019. 2 +[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[8] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In ECCV, 2022. 1, 2, 5, 7, 8 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4 +[11] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Report, 2008. 5, 8 +[12] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7 +[13] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 5, 7 +[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2 +[15] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 2 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[17] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In ECCV, 2018. 2 + +[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In ICCV, 2021. 8 +[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2 +[20] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 3 +[21] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 2, 6 +[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 6 +[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2 +[24] Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming-Hsuan Yang. Deep semantic face deblurring. In CVPR, 2018. 1, 2 +[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3 +[26] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 1 +[27] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 7, 8 +[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 2004. 5 +[29] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In CVPR, 2022. 1, 2, 5, 7, 8 +[30] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. DR2: Diffusion-based robust degradation remover for blind face restoration. In CVPR, 2023. 2, 3, 6, 7, 8 +[31] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In CVPR, 2016. 5 +[32] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 2, 7, 8 +[33] Xin Yu, Basura Fernando, Bernard Ghanem, Fatih Porikli, and Richard Hartley. Face super-resolution guided by facial component heatmaps. In ECCV, 2018. 2 +[34] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv:2212.06512, 2022. 2, 3, 6, 7, 8 +[35] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 1 + +[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[37] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In CVPR, 2022. 2 +[38] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022. 1, 2, 5, 7, 8 \ No newline at end of file diff --git a/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/images.zip b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e659af47da400d110fc28e2f7c880bf3e3947084 --- /dev/null +++ b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd4904f89bb06ef346fb760631b433cfb4c5898a0234af1a1890ce9bb70ffb0b +size 893973 diff --git a/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/layout.json b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc78af0072db8f2b29d7f45739e7ca09d39fe5ff --- /dev/null +++ b/2024/WaveFace_ Authentic Face Restoration with Efficient Frequency Recovery/layout.json @@ -0,0 +1,8892 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 68, + 103, + 525, + 121 + ], + "type": "text", + "content": "WaveFace: Authentic Face Restoration with Efficient Frequency Recovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 219, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 219, + 184 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 219, + 184 + ], + "type": "text", + "content": "Yunqi Miao \nUniversity of Warwick \nYunqi.Miao.1@warwick.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 143, + 367, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 143, + 367, + 184 + ], + "spans": [ + { + "bbox": [ + 244, + 143, + 367, + 184 + ], + "type": "text", + "content": "Jiankang Deng* \nImperial College London \njiankangdeng@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 393, + 143, + 514, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 143, + 514, + 184 + ], + "spans": [ + { + "bbox": [ + 393, + 143, + 514, + 184 + ], + "type": "text", + "content": "Jungong Han* \nUniversity of Sheffield \njungonghan77@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 204, + 276, + 413 + ], + "blocks": [ + { + "bbox": [ + 52, + 204, + 276, + 413 + ], + "lines": [ + { + "bbox": [ + 52, + 204, + 276, + 413 + ], + "spans": [ + { + "bbox": [ + 52, + 204, + 276, + 413 + ], + "type": "image", + "image_path": "17afac7aaab6dd92cbbefd6eb67a843ccb63900741fd73dc675906c66d3a8544.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 420, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 45, + 420, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 45, + 420, + 547, + 455 + ], + "type": "text", + "content": "Figure 1. Left: Illustration of our frequency-aware BFR scheme. Restoration is performed in the frequency domain instead of the pixel domain. Right: Comparisons with state-of-the-art face restoration methods on degraded images. Previous methods struggle to restore facial details or the original identity while our WaveFace achieves a good balance of realism and fidelity with fewer artifacts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 282, + 204, + 561, + 413 + ], + "blocks": [ + { + "bbox": [ + 282, + 204, + 561, + 413 + ], + "lines": [ + { + "bbox": [ + 282, + 204, + 561, + 413 + ], + "spans": [ + { + "bbox": [ + 282, + 204, + 561, + 413 + ], + "type": "image", + "image_path": "b424b1262fcce6b66248cf1572b667955d956d962ed383b8759a6974b5c9e05a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 466, + 192, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 466, + 192, + 478 + ], + "spans": [ + { + "bbox": [ + 143, + 466, + 192, + 478 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 289, + 714 + ], + "type": "text", + "content": "Although diffusion models are rising as a powerful solution for blind face restoration, they are criticized for two problems: 1) slow training and inference speed, and 2) failure in preserving identity and recovering fine-grained facial details. In this work, we propose WaveFace to solve the problems in the frequency domain, where low- and high-frequency components decomposed by wavelet transformation are considered individually to maximize authenticity as well as efficiency. The diffusion model is applied to recover the low-frequency component only, which presents general information of the original image but 1/16 in size. To preserve the original identity, the generation is conditioned on the low-frequency component of low-quality images at each denoising step. Meanwhile, high-frequency components at multiple decomposition levels are handled by a unified network, which recovers complex facial details in a single step. Evaluations on four benchmark datasets show that: 1) WaveFace outperforms state-of-the-art methods in authen" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 467, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 547, + 503 + ], + "type": "text", + "content": "ticity, especially in terms of identity preservation, and 2) authentic images are restored with the efficiency " + }, + { + "bbox": [ + 305, + 467, + 547, + 503 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 305, + 467, + 547, + 503 + ], + "type": "text", + "content": " faster than existing diffusion model-based BFR methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 516, + 387, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 387, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 387, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 536, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 547, + 694 + ], + "type": "text", + "content": "Blind face restoration (BFR) aims to recover high-quality (HQ) facial images from various degradations, including down-sampling, blurriness, noise, and compression artifact [1, 26, 35]. BFR is challenging since the types of degradation are generally unknown in real-world scenarios. Improvements in the restoration quality over the past few years mainly benefit from the usage of multiple facial priors. Geometric priors, including facial landmarks [3] and parsing maps [2, 24], are employed to provide explicit facial structure information. Reference priors, such as HQ images/features [8, 29, 38], are used as guidance during facial restoration. Generative priors are generally obtained by pre-trained generation models such as StyleGAN [13]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 321, + 703, + 400, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 703, + 400, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 703, + 400, + 713 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6583" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "which facilitate the recovery of realistic textures [2, 27, 32]. Inspired by the superior generative ability of diffusion models (DMs) [10, 25], how to unleash its potential in authentic face restoration has gained much attention." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 289 + ], + "type": "text", + "content": "Despite plausible restoration results being achieved, existing DMs based methods [30, 34] generally suffer from two problems: 1) DMs are trained on large images " + }, + { + "bbox": [ + 46, + 122, + 288, + 289 + ], + "type": "inline_equation", + "content": "(512 \\times 512)" + }, + { + "bbox": [ + 46, + 122, + 288, + 289 + ], + "type": "text", + "content": ", which requires massive computational resources for both training and inference, as thousands of iterations are required to obtain adequate outputs. Difface [34], for example, takes about 25 seconds to sample an image from noise. 2) Low-quality (LQ) images are firstly dif- fused to an intermediate step and then denoised by an unconditional diffusion model to obtain HQ counterparts [34]. However, as shown in Fig. 1, such an unconditional generation will lead to great uncertainty in restored images, which means the original identity and facial details such as wrinkles cannot be well preserved." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 459 + ], + "type": "text", + "content": "To solve the above problems, we transfer the BFR task from the pixel domain to the frequency domain via Discrete Wavelet Transform (DWT). As shown in Fig. 1, an image can be decomposed and reconstructed by its four quarter-sized sub-bands: low- and high-frequency components without sacrificing information. Low-frequency sub-band mainly contains general information such as face structure while high-frequency ones contain rich facial details. By exponentially shrinking the size of input images, restoration within the frequency domain not only speeds up both training and inference but also maintains the authenticity of restoration. In the paper, we devise an efficient BFR method WaveFace that restores authentic HQ images by recovering its frequency components." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": "WaveFace consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Recovery (HFR) module. A diffusion model is used in LCD to restore the low-frequency sub-band of HQ images whose size is only 1/16 of the original image. In addition to smaller inputs, LCD leverages LQ counterparts as a condition throughout the generation to preserve the original identity. Meanwhile, HFR recovers high-frequency sub-bands decomposed at multiple DWT levels simultaneously within one step. With the frequency components restored by two modules, authentic images can be reconstructed via discrete inverse wavelet transform (IWT) within 1 ms. The contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 618, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 618, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 286, + 653 + ], + "type": "text", + "content": "- We propose an efficient blind face restoration approach, WaveFace, that restores authentic images by recovering their frequency components individually." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 654, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 286, + 689 + ], + "type": "text", + "content": "- A conditional diffusion model is adopted to restore the low-frequency component, which is 1/16 the size of the original image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 286, + 713 + ], + "type": "text", + "content": "- A one-pass network is used to recover high-frequency sub-bands decomposed at multiple DWT levels simul" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 73, + 367, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 73, + 367, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 367, + 84 + ], + "type": "text", + "content": "taneously." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 84, + 544, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 84, + 544, + 108 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 544, + 108 + ], + "type": "text", + "content": "- Comprehensive experiments demonstrate the superiority of methods in both efficiency and authenticity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 120, + 392, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 392, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 392, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 140, + 408, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 408, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 408, + 152 + ], + "type": "text", + "content": "2.1. Diffusion models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 158, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 230 + ], + "type": "text", + "content": "Diffusion Models (DMs) are emerging generative models that corrupt the data with the successive addition of Gaussian noise during the diffusion process and then learn to recover the data during denoising. State-of-the-art DMs [10] have revealed the potential in CV tasks, such as image deblurring [4, 15] and image super-resolution [22, 23]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 231, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 291 + ], + "type": "text", + "content": "Accelerating DMs. Despite being powerful in generation, DDPM [10] has the downside of low inference speed, which requires thousands of steps for sampling. To accelerate inference, DDIM [19] adopts a non-Markovian diffusion process that allows step skipping during sampling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "content": "Conditional DMs. Unconditional DM-based generation leads to great uncertainty in outputs, which means they generally fail to preserve the original identity and fine-grained facial details. Therefore, conditions are injected by cross-attention layer [22], adaptive normalization layer [21], or concatenation operation [23] to control the characteristic of the synthesized images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 383, + 430, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 430, + 394 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 430, + 394 + ], + "type": "text", + "content": "2.2. Blind face restoration" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 401, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 437 + ], + "type": "text", + "content": "Facial priors exploited in blind face restoration (BFR) can be categorized into three types: geometric priors, reference priors, and generative priors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 437, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 497 + ], + "type": "text", + "content": "Geometric priors based methods leverage knowledge from facial landmark [3, 6], facial parsing maps [2, 24, 37], and facial component heatmaps [33]. However, they tend to show inferior performance since degraded images fail to provide accurate and adequate structural information." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": "Reference priors based methods either leverage a high-quality (HQ) reference image sharing the same identity as the degraded one [17] or a pre-constructed dictionary storing HQ facial features [8, 29, 38]. Firstly, a vectorquantized (VQ) codebook is pre-trained on HQ faces via VQ-GAN [7] to provide rich facial details. Features of degraded inputs are then fused with the prior either at image [29] or latent [8, 38] level for the generation of HQ counterparts. However, reference prior is restricted by the size of the codebook, which limits the diversity and richness of generated images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Generative priors encapsulated in pretrained face models [13, 14] are also leveraged to restore faithful faces [2, 27, 32]. PSFRGAN [2] modulates features at different scales progressively with facial parsing maps to achieve semantic-aware style transformation. GFP-GAN [27] and GPEN [32] adopt the pre-trained StyleGAN as a decoder and achieve a good balance between visual quality and fidelity of re" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6584" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "content": "stored images. Recently, diffusion models [10, 25] have shown the powerful generative ability in restoring HQ content from noisy images. Inspired by the ability, DifFace [34] firstly maps low-quality (LQ) inputs into an intermediate step of the denoising process, from which its HQ counterpart is recursively sampled. DR2 [30] employs the diffusion model to remove degradations, followed by a super-resolution model to obtain HQ counterparts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "content": "Nevertheless, previous DM-based BFR methods are performed in the pixel domain, where large inputs require massive computational resources for both training and inference. Besides, they both apply the unconditional scheme that leads to significant uncertainty in restored results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 238, + 124, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 124, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 124, + 251 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 286, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 341 + ], + "type": "text", + "content": "Discrete Wavelet Transformation (DWT). DWT decomposes an image into low-frequency and high-frequency subbands. The low-frequency component presents general information while high-frequency ones express facial details in the vertical, horizontal, and diagonal directions. In the paper, we use Haar wavelet for DWT, which is widely used in real-world applications due to simplicity [20]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "text", + "content": ", its low-frequency subband " + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{ll}^{1} \\in \\mathbb{R}^{H/2 \\times W/2 \\times 3}" + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "text", + "content": " and high-frequency sub-bands " + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{lh}^{1}, \\pmb{x}_{hl}^{1}" + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{hh}^{1} \\in \\mathbb{R}^{H/2 \\times W/2 \\times 3}" + }, + { + "bbox": [ + 46, + 341, + 286, + 378 + ], + "type": "text", + "content": " can be decomposed by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 384, + 287, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 384, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 101, + 384, + 287, + 398 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {l l} ^ {1}, \\boldsymbol {x} _ {l h} ^ {1}, \\boldsymbol {x} _ {h l} ^ {1}, \\boldsymbol {x} _ {h h} ^ {1} = D W T (\\boldsymbol {x}), \\tag {1}", + "image_path": "a9756042267f57b4a78eecf42ba272721db4cb1513cf72fb0b8a422e751635a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "type": "inline_equation", + "content": "DWT(\\cdot)" + }, + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "type": "text", + "content": " refers to the DWT operation. Despite that the input size is decreased by a factor of four " + }, + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "type": "inline_equation", + "content": "(H / 2\\times W / 2)" + }, + { + "bbox": [ + 46, + 404, + 286, + 464 + ], + "type": "text", + "content": " after a single DWT decomposition, DWT can be performed on the low-frequency component to further reduce the computational cost and expedite the inference:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 469, + 287, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 469, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 82, + 469, + 287, + 485 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {l l} ^ {J + 1}, \\boldsymbol {x} _ {l h} ^ {J + 1}, \\boldsymbol {x} _ {h l} ^ {J + 1}, \\boldsymbol {x} _ {h h} ^ {J + 1} = D W T (\\boldsymbol {x} _ {l l} ^ {J}), \\tag {2}", + "image_path": "cec4e097df0c2dfc5304cce787a304311f2390c744c43070709f23dea9cfd35a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "type": "text", + "content": " is level of DWT decomposition and " + }, + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{ll}^{J + 1},\\pmb{x}_{lh}^{J + 1},\\pmb{x}_{hl}^{J + 1},\\pmb{x}_{hh}^{J + 1}\\in \\mathbb{R}^{H / 2^{(J + 1)}\\times W / 2^{(J + 1)}\\times 3}" + }, + { + "bbox": [ + 46, + 491, + 286, + 540 + ], + "type": "text", + "content": ". Reversibly, given the frequency sub-bands, an image can be reconstructed via discrete inverse wavelet transform (IWT):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 552, + 286, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 552, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 102, + 552, + 286, + 566 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} = I W T \\left(\\boldsymbol {x} _ {l l} ^ {1}, \\boldsymbol {x} _ {l h} ^ {1}, \\boldsymbol {x} _ {h l} ^ {1}, \\boldsymbol {x} _ {h h} ^ {1}\\right). \\tag {3}", + "image_path": "5ce4d34a43bd632fc009e4e324a978303c2d0df8a86b8fcf1b87afe8beae09dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "type": "text", + "content": "Diffusion Models. Training of diffusion models (DMs) consists of a diffusion process and a denoising process. Diffusion process transforms an image from the real data distribution " + }, + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{y}_0\\sim q(\\pmb{y}_0)" + }, + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "type": "text", + "content": " into a pure Gaussian noise " + }, + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{y}_T" + }, + { + "bbox": [ + 46, + 570, + 286, + 629 + ], + "type": "text", + "content": " by successively applying the following Markov diffusion kernel:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 80, + 639, + 286, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 639, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 80, + 639, + 286, + 654 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {y} _ {t} \\mid \\boldsymbol {y} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t}; \\sqrt {1 - \\beta_ {t}} \\boldsymbol {y} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right), \\tag {4}", + "image_path": "c4169e4c93626e50b3cabe97aa6f9eb0850383b9779be00a5455056fe415bf79.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "spans": [ + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "type": "text", + "content": " is a pre-defined or learned noise variance schedule. The marginal distribution at arbitrary timestep " + }, + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 657, + 286, + 692 + ], + "type": "text", + "content": " can be denoted as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 700, + 286, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 700, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 86, + 700, + 286, + 715 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {y} _ {t} \\mid \\boldsymbol {y} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {y} _ {0}, (1 - \\alpha_ {t}) \\boldsymbol {I}\\right), \\tag {5}", + "image_path": "324581cb0ea9dbdc61f6a8cef0934d2759bf515fad05d887a5b5454dc92416e8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\prod_{s=1}^{t}(1 - \\beta_{s})" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t}" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": ", the denoising process aims to recover " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{0}" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": " by recursively learning the transition from " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t-1}" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t}" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": ", which is defined as the following Gaussian distribution:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 329, + 129, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 129, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 329, + 129, + 545, + 143 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t\\right)\\right), \\tag {6}", + "image_path": "260f947a1e8047a80ae5de4c11542be6162bb524b5c1ecc10b30cff8ffcf34ed.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "text", + "content": "where parameters " + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "text", + "content": " are optimized by a denoising network " + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "text", + "content": " that predicts " + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{\\theta}(\\pmb{y}_t,t)" + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{\\theta}(\\pmb{y}_t,t)" + }, + { + "bbox": [ + 305, + 152, + 545, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 187, + 388, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 187, + 388, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 388, + 201 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 207, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 338 + ], + "type": "text", + "content": "In the paper, WaveFace is proposed to handle blind face restoration (BFR) in the frequency domain. The framework is depicted in Fig. 2. First, degraded images are mapped to the frequency domain via Discrete Wavelet Transformation (DWT) (Sec. 3). The proposed Low-frequency Conditional Denoising (LCD) module (Sec. 4.1) and High-Frequency Recovery (HFR) module (Sec. 4.2) are adopted on the low- and high-frequency components respectively to remove degradations and restore facial details. Recovered frequency components are used for the image reconstruction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 346, + 507, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 507, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 507, + 360 + ], + "type": "text", + "content": "4.1. Low-frequency Conditional Denoising" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 365, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 436 + ], + "type": "text", + "content": "As shown in Fig. 1, the low-frequency component is similar to the down-sampled version of the original image, which largely determines the restoration quality. Due to its powerful generative ability from noisy inputs, a diffusion model (DM) is adopted in LCD to restore the low-frequency subband of high-quality (HQ) images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": "Conditional DM. We denote the low-frequency sub-band of a pair of LQ and HQ images as " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_{ll_0}^j,\\pmb{y}_{ll_0}^j)" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": ". Since the low-frequency sub-band only contains a single map after DWT, the DWT levels " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "ll" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": " are omitted in this section for simplicity. According to the diffusion process defined in Eq. (5), images are successively destroyed by Gaussian noise as timestep " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": " increases, which means the distribution " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_T|\\pmb{x}_0)\\approx q(\\pmb{y}_T|\\pmb{y}_0)\\approx \\mathcal{N}(0,\\pmb {I})" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": " after a large timestep " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": ". Given this assumption, our LCD aims to learn the posterior distribution " + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "p(\\pmb{y}_0|\\pmb{x}_0)" + }, + { + "bbox": [ + 305, + 437, + 545, + 557 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 566, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 566, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 321, + 566, + 545, + 598 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {y} _ {0} \\mid \\boldsymbol {x} _ {0}\\right) = \\int q \\left(\\boldsymbol {y} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}\\right) \\mathrm {d} \\boldsymbol {y} _ {1: T}, \\tag {7}", + "image_path": "25bd950a2ecddaecddad002cffa60353d80313c900777d30ac71bb4489eded02.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{y}_{t-1}|\\boldsymbol{y}_t)" + }, + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "type": "text", + "content": " refers to the unconditional denoising process defined in Eq. (6). However, the unconditional scheme fails to preserve the original identity due to the lack of guidance during generation. To solve the problem, the low-frequency sub-band of LQ images " + }, + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 305, + 607, + 545, + 679 + ], + "type": "text", + "content": " are injected as the condition for the denoising process:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {y} _ {t - 1} \\mid \\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {y} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t, \\boldsymbol {x} _ {0}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, t, \\boldsymbol {x} _ {0}\\right)\\right), \\tag {8}", + "image_path": "17f166f79edded6c96602ea75babfa9a2479e469bdb14b340cecf8ab15e4bdf3.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6585" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 542, + 252 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 252 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 252 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 252 + ], + "type": "image", + "image_path": "b5f8ce459f818d7d6b4db2e4fe1b7817f56c1a3bb7f0098d8ee84f65f4b5e003.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "type": "text", + "content": "Figure 2. Overall framework of WaveFace. It consists of a Low-frequency Conditional Denoising (LCD) module and a High-Frequency Refinement (HFR) module. LCD (Sec. 4.1) predicts clean samples " + }, + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "type": "text", + "content": " from noise conditioned on LQ inputs through " + }, + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 259, + 547, + 304 + ], + "type": "text", + "content": " steps. Meanwhile, high-frequency sub-bands are concatenated as HFR (Sec. 4.2) inputs to recover vivid facial details. The predicted frequency components are projected back to the image via IWT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "type": "text", + "content": " is injected by concatenating with the input " + }, + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 46, + 317, + 286, + 340 + ], + "type": "text", + "content": " along the channel dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "type": "text", + "content": "Objectives. Following DDPM [10], the denoising network " + }, + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "type": "text", + "content": " is trained to predict noise vectors with the objective:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 374, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 374, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 60, + 374, + 287, + 396 + ], + "type": "interline_equation", + "content": "L _ {L C D} = \\mathbb {E} _ {\\boldsymbol {y} _ {0}, t, \\boldsymbol {\\epsilon} _ {t} \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\left[ | | \\boldsymbol {\\epsilon} _ {t} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}, t) | | ^ {2} \\right]. \\tag {9}", + "image_path": "1e922ddebecb48fd8fa6dbe42002a1baf7ff6e7d1e44b3f5368d069c6919e6fb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "type": "text", + "content": "During inference, with the learned parameterized Gaussian transitions " + }, + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{y}_{t-1}|\\pmb{y}_t,\\pmb{x}_0)" + }, + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "type": "text", + "content": ", the low-frequency sub-band of HQ images can be recovered from a random Gaussian noise " + }, + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{y}_T \\sim \\mathcal{N}(\\pmb{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 403, + 287, + 452 + ], + "type": "text", + "content": " by recursively applying:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 460, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 460, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 52, + 460, + 287, + 487 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {y} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {y} _ {t}, \\boldsymbol {x} _ {0}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {z}, \\tag {10}", + "image_path": "615069a83501e70d9d0c20aa7aef88387da13917f0bc86a045a74b094fcb7e94.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 497, + 264, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 264, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 264, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 497, + 264, + 510 + ], + "type": "inline_equation", + "content": "z\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}),\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 46, + 497, + 264, + 510 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 46, + 497, + 264, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i = 1}^t\\alpha_i" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "Trade-off between efficiency and quality. Training a DM on large inputs (" + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\times 512" + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": " or more) requires massive computational resources and the evaluation takes thousands of steps to sample from the noise. Although this problem can be alleviated by adopting low-frequency sub-bands at higher DWT levels, multiple times of decomposition will also lead to information reduction. How to set DWT level wisely is the key to balance between efficiency and authenticity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "To illustrate how DWT level " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " affects the authenticity, we visualize the image reconstructed by the low-frequency component of HQ images and the corresponding high-frequency ones of its LQ counterpart at different DWT levels in Fig. 3. Although noisy at the high-frequency part, the quality of images reconstructed at " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "J \\leq 2" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is acceptable, with key facial features such as wrinkles still preserved. However, the mosaic effect starts to emerge when " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " keeps increasing. Ablation studies about the effect of DWT level on" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 317, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 317, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 545, + 341 + ], + "type": "text", + "content": "training/inference are shown in Sec. 5.2.1. In the paper, " + }, + { + "bbox": [ + 305, + 317, + 545, + 341 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 305, + 317, + 545, + 341 + ], + "type": "text", + "content": " is set as 2 empirically." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 347, + 448, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 448, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 448, + 361 + ], + "type": "text", + "content": "4.2. High-frequency Recovery" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 484 + ], + "type": "text", + "content": "Although the general face information is restored by LCD, rich facial details are generally embedded in the high-frequency component, as shown in Fig. 1, which improves the authenticity of restored images. High-frequency sub-bands at multiple DWT levels also vary in size, which means more than one diffusion model is required for the recovery. To avoid the extravagant computational cost, we adopt a U-shaped network that can recover high-frequency sub-bands at multiple DWT levels at the same time and takes only one step for inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "content": "Given a LQ and HQ image pair " + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "content": " and their high-frequency sub-bands: " + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\pmb{x}_H^j = \\{\\pmb{x}_{lh}^j,\\pmb{x}_{hl}^j,\\pmb{x}_{hh}^j\\}" + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\pmb{y}_H^j = \\{\\pmb{y}_{lh}^j,\\pmb{y}_{hl}^j,\\pmb{y}_{hh}^j\\}" + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "inline_equation", + "content": "j\\in \\{1,2\\}" + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "content": " refers to DWT level, the framework our high-frequency recovery (HFR) module is illustrated in Fig. 4. HFR takes high-frequency sub-bands of LQ images that are channel-wise concatenated as inputs and outputs the recovered ones " + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_H^j" + }, + { + "bbox": [ + 305, + 485, + 545, + 571 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 577, + 545, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 577, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 344, + 577, + 545, + 592 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {H} ^ {1}, \\boldsymbol {F} _ {H} ^ {2} = \\mathcal {U} _ {\\phi_ {u}} \\left(\\mathcal {H} _ {\\phi_ {i n}} \\left(\\boldsymbol {x} _ {H} ^ {1}\\right), \\mathcal {F} _ {\\phi_ {f}} \\left(\\boldsymbol {x} _ {H} ^ {2}\\right)\\right), \\tag {11}", + "image_path": "83990d3c08de9d2a7d0612fd0751a42a537a7309772088074545ce59c342ca23.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 594, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 594, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 344, + 594, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {H} ^ {j} = \\mathcal {H} _ {\\phi_ {o u t} ^ {j}} \\left(\\boldsymbol {F} _ {H} ^ {j}\\right) + \\boldsymbol {x} _ {H} ^ {j}, \\quad j \\in \\{1, 2 \\}, \\tag {12}", + "image_path": "05e5c78bb4890523bd01260c2ff27825eba3a2f7808a7eede52dd73e2678d38b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\cdot),\\mathcal{H}(\\cdot),\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " denotes the U-shaped network, the convolutional layer, and the fusion module with the parameter " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_u,\\phi_{in / out}^j" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_f" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " , respectively. " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_H^1,F_H^2" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " represents feature maps of the last and penultimate decoding layers of the U-shaped network. Considering efficiency, the fusion module only contains two stacked ConvBlocks that apply " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " convolutional layers with ReLU in between, followed by concatenation with the input." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6586" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 129, + 151 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 129, + 151 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 129, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 129, + 151 + ], + "type": "image", + "image_path": "114fde351da94fa3f60faf34eea20eda9db80c421ed7b7cac1b61241e8320b30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 152, + 154, + 163 + ], + "lines": [ + { + "bbox": [ + 92, + 152, + 154, + 163 + ], + "spans": [ + { + "bbox": [ + 92, + 152, + 154, + 163 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 92, + 152, + 154, + 163 + ], + "type": "inline_equation", + "content": "J = 3" + }, + { + "bbox": [ + 92, + 152, + 154, + 163 + ], + "type": "inline_equation", + "content": "(\\times 64)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 70, + 212, + 151 + ], + "blocks": [ + { + "bbox": [ + 130, + 70, + 212, + 151 + ], + "lines": [ + { + "bbox": [ + 130, + 70, + 212, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 70, + 212, + 151 + ], + "type": "image", + "image_path": "c3d5598ee903e6a11b0340d42609b1f08093e46f8fc8922d38234f8d8b6499da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "type": "text", + "content": "Figure 3. Visualization of DWT frequency components and images reconstructed by the low-frequency sub-band of an HQ image and high-frequency sub-bands of its LQ counterpart. DWT level " + }, + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "type": "text", + "content": " and resolution of low- / high-frequency sub-bands " + }, + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "type": "inline_equation", + "content": "(\\times N)" + }, + { + "bbox": [ + 46, + 170, + 547, + 194 + ], + "type": "text", + "content": " are reported." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 302, + 151 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 302, + 151 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 302, + 151 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 302, + 151 + ], + "type": "image", + "image_path": "6d1f9177ad887155a1b1d36b68469e731074220b7c0423160ec11e88627f56a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 152, + 331, + 163 + ], + "lines": [ + { + "bbox": [ + 262, + 152, + 331, + 163 + ], + "spans": [ + { + "bbox": [ + 262, + 152, + 331, + 163 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 262, + 152, + 331, + 163 + ], + "type": "inline_equation", + "content": "J = 2" + }, + { + "bbox": [ + 262, + 152, + 331, + 163 + ], + "type": "text", + "content": " （ " + }, + { + "bbox": [ + 262, + 152, + 331, + 163 + ], + "type": "inline_equation", + "content": "\\times 128)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 385, + 151 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 385, + 151 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 385, + 151 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 385, + 151 + ], + "type": "image", + "image_path": "bd60efaf13450e8c213049271a62f5611cd57ccda4045978fa6e9b88b341449b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 152, + 330, + 163 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 330, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 330, + 163 + ], + "type": "inline_equation", + "content": "(\\times 128)" + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 394, + 70, + 476, + 151 + ], + "blocks": [ + { + "bbox": [ + 394, + 70, + 476, + 151 + ], + "lines": [ + { + "bbox": [ + 394, + 70, + 476, + 151 + ], + "spans": [ + { + "bbox": [ + 394, + 70, + 476, + 151 + ], + "type": "image", + "image_path": "613ca2f7a939901f92019de08fcfa15a9b3b42cfca1ec63143fd13d839fda834.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 152, + 503, + 163 + ], + "lines": [ + { + "bbox": [ + 436, + 152, + 503, + 163 + ], + "spans": [ + { + "bbox": [ + 436, + 152, + 503, + 163 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 436, + 152, + 503, + 163 + ], + "type": "inline_equation", + "content": "J = 1" + }, + { + "bbox": [ + 436, + 152, + 503, + 163 + ], + "type": "text", + "content": " （20 " + }, + { + "bbox": [ + 436, + 152, + 503, + 163 + ], + "type": "inline_equation", + "content": "(\\times 256)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 477, + 70, + 558, + 151 + ], + "blocks": [ + { + "bbox": [ + 477, + 70, + 558, + 151 + ], + "lines": [ + { + "bbox": [ + 477, + 70, + 558, + 151 + ], + "spans": [ + { + "bbox": [ + 477, + 70, + 558, + 151 + ], + "type": "image", + "image_path": "ab4f1e55665da1f37183e70eaef116ff997bc4237becd24ee06c621fe2b33a26.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 207, + 289, + 310 + ], + "blocks": [ + { + "bbox": [ + 50, + 207, + 289, + 310 + ], + "lines": [ + { + "bbox": [ + 50, + 207, + 289, + 310 + ], + "spans": [ + { + "bbox": [ + 50, + 207, + 289, + 310 + ], + "type": "image", + "image_path": "f6ec98b1931a5f4267ce0a7a5b30497351412b4d6d2941605e8388e1ad980ccf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 320, + 285, + 333 + ], + "lines": [ + { + "bbox": [ + 49, + 320, + 285, + 333 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 285, + 333 + ], + "type": "text", + "content": "Figure 4. Illustration of high-frequency recovery (HFR) module." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "text", + "content": "Objectives. We adopt two loss terms for the training of HFR: recovery loss " + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "text", + "content": " and content loss [28] " + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "text", + "content": ". The recovery loss is applied not only to restore HQ high-frequency sub-bands " + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{y}_H^j" + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "text", + "content": " at each level but also to the image " + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{j - 1}" + }, + { + "bbox": [ + 46, + 345, + 287, + 405 + ], + "type": "text", + "content": " reconstructed via inverse wavelet transform (IWT):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 413, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\sum_ {j} \\left( \\right.\\left|\\left| \\hat {\\boldsymbol {x}} _ {H} ^ {j} - \\boldsymbol {y} _ {H} ^ {j} \\right|\\right| + \\alpha \\left. \\right\\rvert\\left| I W T \\left(\\boldsymbol {y} _ {l l} ^ {j}, \\hat {\\boldsymbol {x}} _ {H} ^ {j}\\right) - \\boldsymbol {y} ^ {j - 1} \\right|\\left. \\right), \\tag {13}", + "image_path": "376844e6d9f94af6c8c46c565ee0d7457c6ad3444a382972f02b0bff19d613bf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "type": "inline_equation", + "content": "IWT(\\cdot)" + }, + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "type": "text", + "content": " denotes IWT operation and " + }, + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 450, + 287, + 522 + ], + "type": "text", + "content": " is the parameter balancing losses between the frequency domain and the pixel domain. In addition to minimizing the pixel-wise distance, the content loss is applied to maximize the luminance, contrast, and structural similarity of the restored and original HQ image:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 71, + 530, + 287, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 530, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 71, + 530, + 287, + 545 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = 1 - S S I M \\left(I W T \\left(\\boldsymbol {y} _ {l l} ^ {2}, \\hat {\\boldsymbol {x}} _ {H} ^ {2}, \\hat {\\boldsymbol {x}} _ {H} ^ {1}\\right), \\boldsymbol {y}\\right). \\tag {14}", + "image_path": "69d7d4e80d6db0a4c1439149d4618f45cc6b4bd8c7c577f82d4282fd4ae3a9cb.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "text", + "content": "The overall objective of HFR is " + }, + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{HFR} = \\mathcal{L}_r + \\lambda \\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "text", + "content": ", where the weighting parameter " + }, + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "text", + "content": " is set as 10." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 589, + 128, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 128, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 128, + 602 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 609, + 170, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 170, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 170, + 621 + ], + "type": "text", + "content": "5.1. Settings and Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 627, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 688 + ], + "type": "text", + "content": "Training Dataset. FFHQ [13] is used as the training set, which contains 70,000 high-quality (HQ) face images. Following BFR benchmark works [8, 29], we resize HQ images to the resolution of " + }, + { + "bbox": [ + 46, + 627, + 287, + 688 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 627, + 287, + 688 + ], + "type": "text", + "content": ", and then synthesize low-quality (LQ) counterparts as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 92, + 696, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 696, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 92, + 696, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} = \\left\\{\\left[ (\\boldsymbol {y} \\otimes \\boldsymbol {k} _ {\\sigma}) \\downarrow_ {s} + \\boldsymbol {n} _ {\\delta} \\right] _ {\\mathrm {J P E G} _ {q}} \\right\\} \\uparrow_ {s}, \\tag {15}", + "image_path": "2dc0134eef2b91fb714bb25f7cf23f9b98bbbbcbd971c6c6a1840cabdc85fd1f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": "where a HQ image " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": " is firstly blurred by a Gaussian kernel " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "k_{\\sigma}" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": ", followed by a downsampling of scale " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": ". Afterward, Gaussian noise " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{n}_{\\delta}" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": " and JPEG compression with quality factor " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": " are applied to the image, which is then upsampled back to the original size to obtain its LQ counterpart " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": ". The hyper-parameters " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "\\sigma, s, \\delta" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 208, + 546, + 292 + ], + "type": "text", + "content": " are uniformly sampled from [0.1, 15], [0.8, 32], [0, 20], and [30, 95] respectively." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "content": "Testing Dataset. We evaluate WaveFace on a synthetic dataset: CelebA-Test and three real-world datasets: LFW-Test [11], WebPhoto-Test [27], and WIDER-Test [38]. CelebA-Test contains 3000 HQ images from CelebAHQ [12], and LQ counterparts are synthesized via Eq. (15) with the same degradation setting. In terms of three realworld datasets, LFW-Test contains 1711 mildly degraded face images in the wild, which comprises the first image for each person in LFW [11]. WebPhoto-Test includes 407 images crawled from the internet, some of which are old photos with severe degradation. WIDER-Test consists of 970 images with severe degradations from the WIDER dataset [31]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "Evaluation Metrics. For evaluation, we adopt two pixelwise metrics (PSNR and SSIM), a reference perceptual metric (LPIPS [36]), and a non-reference perceptual metric (FID [9]). To measure the consistency of identity, the angle between embeddings extracted by ArcFace (\"Deg.\") [5] is used. All metrics are used for the evaluation of synthetic data while only FID is used for real-world datasets due to the lack of referential HQ images." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": "Implementation Details. The DWT decomposition level is set as 2, which reduces the input size of the diffusion model from " + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": ". We train LCD and HFR individually with Adam [16] optimizer. LCD is trained for 200K iterations with the learning rate " + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "inline_equation", + "content": "1e-4" + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": " and batch size 32. For HFR, the training takes 70K iterations. The learning rate gradually decays from " + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "inline_equation", + "content": "1e-3" + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "inline_equation", + "content": "1e-5" + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": " by a factor of 0.1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 638, + 407, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 407, + 650 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 407, + 650 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 657, + 501, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 501, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 501, + 670 + ], + "type": "text", + "content": "5.2.1 Low-frequency Conditional Denoising" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "DWT Levels. We conduct an experiment to investigate how the level of wavelet decomposition (DWT) affects the efficiency and authenticity. DWT is applied 0, 1, 2, 3 times on" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6587" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 124, + 286, + 191 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "type": "text", + "content": "Table 1. Comparison between state-of-the-art diffusion model-based BFR methods and ours at different DWT levels in terms of efficiency and authenticity. The best and the second best performances are highlighted and underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 124, + 286, + 191 + ], + "lines": [ + { + "bbox": [ + 50, + 124, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 50, + 124, + 286, + 191 + ], + "type": "table", + "html": "
MethodLevelResolution#Param. (M)Time (s)PSNR↑SSIM↑
DifFace [34]-× 512175.3825.0419.060.46
DR2 [30]-× 25693.568.7622.890.57
WaveFace0× 512109.0619.3726.420.65
1× 25671.418.1227.080.73
2× 12825.361.9726.970.71
3× 643.91.4624.810.46
", + "image_path": "1ba426f93fb5203bb6c55dddcad9edac28769f17d732db7888467856bcae863a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": "the training set, where the resolution of the diffusion model (DM)'s input becomes " + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": ", respectively. The number of parameters and the inference time are used to evaluate the efficiency while PSNR and SSIM are used to evaluate authenticity. The comparison results are reported in Tab. 1. In comparison with the model trained on original images (" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\times 512" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": "), our frequency-aware scheme applies DM on the low-frequency sub-band only, which avoids some unknown noise on high-frequency components. Thus, the restoration quality improves after the wavelet decomposition is applied once. The quality can be maintained until the DWT level increases to 2, where the inference process is accelerated by " + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 205, + 287, + 420 + ], + "type": "text", + "content": ". However, the performance is significantly harmed when the DWT level further increases (3 or higher) as too many details in the low-frequency component have been reduced. DWT level is set as 2 throughout the experiments to achieve the tradeoff between the efficiency and the quality of restoration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "content": "We also compare the efficiency with two DM-based BFR methods: Difface [34] and DR2 [30]. To make a fair comparison, the inference time refers to the time of the whole denoising process, which starts from the pure Gaussian noise. The quality of generated images cannot be compared with that reported in Tab. 3 as images are generated with pre-/post-processing and fewer sampling steps in their methods. As can be seen, our method achieves better restoration quality at a lower computational cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 532, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 628 + ], + "type": "text", + "content": "Conditions. To investigate how the condition facilitates the restoration, we first remove the condition from the input of our LCD module and train an unconditional diffusion model for 200K iterations for a fair comparison. The restored face is denoted by \"Uncond.\" in Fig. 5. We notice that not only the identity is not preserved but also the image quality drops significantly. This could be attributed to the slow convergence of the unconditional training manner [21]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Additionally, we compare with two widely-used conditioning schemes: adaptive group normalization layers (AdaGN) [21] and cross-attention [22]. We replace the low-frequency sub-band of LQ images with the corresponding identity embedding extracted by Arcface [5], which are subsequently injected via AdaGN. Furthermore, a cross-attention layer is added to condition the denoised image" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 74, + 388, + 153 + ], + "blocks": [ + { + "bbox": [ + 317, + 74, + 388, + 153 + ], + "lines": [ + { + "bbox": [ + 317, + 74, + 388, + 153 + ], + "spans": [ + { + "bbox": [ + 317, + 74, + 388, + 153 + ], + "type": "image", + "image_path": "bbbe31da901b1875dd50f21c649ad29f723fc671c25ae9e198af80e6046e6c27.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 74, + 464, + 153 + ], + "blocks": [ + { + "bbox": [ + 391, + 74, + 464, + 153 + ], + "lines": [ + { + "bbox": [ + 391, + 74, + 464, + 153 + ], + "spans": [ + { + "bbox": [ + 391, + 74, + 464, + 153 + ], + "type": "image", + "image_path": "44b71f2dc83ba4a498e4520dcccb4005e7dd1a58f792aa0046729979c563c423.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 468, + 74, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 468, + 74, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 468, + 74, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 468, + 74, + 541, + 153 + ], + "type": "image", + "image_path": "e2504fc39c96a2f24c25d36881b1778a9abffbafdcb406a0c30a2443d0f0eac3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 156, + 389, + 234 + ], + "blocks": [ + { + "bbox": [ + 312, + 156, + 389, + 234 + ], + "lines": [ + { + "bbox": [ + 312, + 156, + 389, + 234 + ], + "spans": [ + { + "bbox": [ + 312, + 156, + 389, + 234 + ], + "type": "image", + "image_path": "4fe4353810bbc3becc9869695d331900b2f7e69496f5c0f824a6184c275e4d8f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 243, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 277 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of different conditioning schemes on CelebA-Test. We adopt \"Concat.\" in WaveFace. PSNR(↑) / Deg.(↓) are reported." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 156, + 466, + 235 + ], + "blocks": [ + { + "bbox": [ + 391, + 156, + 466, + 235 + ], + "lines": [ + { + "bbox": [ + 391, + 156, + 466, + 235 + ], + "spans": [ + { + "bbox": [ + 391, + 156, + 466, + 235 + ], + "type": "image", + "image_path": "7ac1b2ed86629e988223fb8c18fa65f974a84bcad6295f801712f1af72a2d11b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 468, + 156, + 541, + 235 + ], + "blocks": [ + { + "bbox": [ + 468, + 156, + 541, + 235 + ], + "lines": [ + { + "bbox": [ + 468, + 156, + 541, + 235 + ], + "spans": [ + { + "bbox": [ + 468, + 156, + 541, + 235 + ], + "type": "image", + "image_path": "43d1c414e1d67135e0dbebd5aecfcbff4c7bceda8fa08ff17b410586f0e05b9c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 285, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 417 + ], + "type": "text", + "content": "at each step with the LQ low-frequency information. Experimental details of the two schemes will be discussed in supplementary materials. The restored images are shown in Fig. 5 as \"AdaGN\" and \"Cross-Att.\", respectively. Although the original identity is preserved by the AdaGN scheme, the background is not well restored. A possible reason is that the identity-related condition is performed at the latent level only, which fails to provide pixel-wise constraints. On the other hand, despite the cross-attention scheme achieves a comparable restoration result with ours (\"Concat.), the extra layer requires more computational resources." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 479, + 543, + 536 + ], + "blocks": [ + { + "bbox": [ + 305, + 426, + 545, + 471 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 471 + ], + "type": "text", + "content": "Table 2. Ablation studies of the effectiveness of our high-frequency recovery (HFR) module. " + }, + { + "bbox": [ + 305, + 426, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_H" + }, + { + "bbox": [ + 305, + 426, + 545, + 471 + ], + "type": "text", + "content": " denotes high-frequency sub-bands of LQ images and 1, 2 stands for the level of wavelet decomposition. Best performance is indicated by Red." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 479, + 543, + 536 + ], + "lines": [ + { + "bbox": [ + 309, + 479, + 543, + 536 + ], + "spans": [ + { + "bbox": [ + 309, + 479, + 543, + 536 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓FID↓
LCD + xH1 + xH224.3530.4300.53717.264
LCD + HFR1 + xH226.0810.6090.48114.020
LCD + xH1 + HFR224.8190.4600.53615.293
LCD + HFR1 + HFR2 (Ours)26.9670.7110.34313.062
", + "image_path": "01aecce2480d68212747c792ae729b70203ade36083e10ba7a2994f8e41cf81b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "text", + "content": "5.2.2 High-frequency Recovery" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "To investigate the effectiveness of our High-frequency Recovery (HFR) module, we replace the recovered high-frequency sub-bands at different DWT levels with that of low-quality (LQ) images for image reconstruction. Quantitative comparisons on the restoration quality of different schemes are reported in Tab. 2. The restoration quality drops significantly when HFR is removed with a decline of " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.62\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " in PSNR. Besides, the lack of recovered high-frequency components at different DWT levels deteriorates the restoration quality to varying extents. In comparison with those at level 2 (" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\times 128" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "), high-frequency sub-bands at level 1 (" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\times 256" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ") contribute more to the restoration quality." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6588" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 76, + 282, + 411 + ], + "blocks": [ + { + "bbox": [ + 47, + 76, + 282, + 411 + ], + "lines": [ + { + "bbox": [ + 47, + 76, + 282, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 76, + 282, + 411 + ], + "type": "image", + "image_path": "71d4eb0dfc89f77df6ea96380a7f57f7246846329453f8597aa00520ca8bc64d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison between images reconstructed at different DWT levels with and without (w/o) HFR module on CelebA-Test. " + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "text", + "content": " denotes LQ images, restored images, and HQ images. " + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{x}^j (j\\in \\{1,2\\})" + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "text", + "content": " denotes frequency components at " + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 418, + 287, + 473 + ], + "type": "text", + "content": "-th DWT level. PSNR(↑) of the restored images are reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 481, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 565 + ], + "type": "text", + "content": "The exemplar illustrated in Fig. 6 helps to understand intuitively. Before our HFR is applied (Row 2), the restored image contains cluttered noise and lacks details such as hairstyle due to messy high-frequency sub-bands of LQ images. The quality improves when the image is constructed with the refined high-frequency component (Row 3), which is less noisy and presents fine-grained details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 574, + 277, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 277, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 277, + 586 + ], + "type": "text", + "content": "5.3. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 593, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 286, + 664 + ], + "type": "text", + "content": "State-of-the-art (SOTA) methods used for comparison include generative prior based methods (GPEN [32] and GFPGAN [27]), reference prior based methods (VQFR [8] and RestoreFormer [29]) and recent diffusion model-based methods (DifFace [34] and DR2 [30]). Evaluations are conducted on both synthetic and real-world datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Synthetic dataset. Quantitative comparison on CelebA-Test [12] are illustrated in Tab. 3. WaveFace achieves the best scores on reference-based metrics and ranks second on FID. Apart from the better restoration quality, our method" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 114, + 548, + 235 + ], + "blocks": [ + { + "bbox": [ + 305, + 62, + 545, + 106 + ], + "lines": [ + { + "bbox": [ + 305, + 62, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 305, + 62, + 545, + 106 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on CelebA-Test for blind face restoration. \"Deg.\" refers to the angle between identity embeddings of restored images and HQ counterparts. The best and the second best performances are highlighted and underlined." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 114, + 548, + 235 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 548, + 235 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 548, + 235 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑LPIPS↓FID↓Deg↓
Input23.0380.3890.58667.50554.958
GPEN [32]24.3190.6030.44022.25138.473
GFP-GAN [27]24.7710.6740.36114.50136.958
VQFR [8]23.7350.6140.35814.04838.684
RestoreFormer [29]24.1910.6360.36412.00034.752
CodeFormer [38]25.0710.6720.35914.38537.677
DR2 [30]23.5830.6130.40214.67151.030
Difface [34]24.1550.6670.39013.37949.958
WaveFace26.967↑1.900.7110.34313.06228.263↑6.49
", + "image_path": "3f3977a9bfe191e403abfbd48a0bacc4ddfc5f7bc8ad5dd92058837f9b75f2a3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 285, + 531, + 411 + ], + "blocks": [ + { + "bbox": [ + 305, + 243, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 276 + ], + "type": "text", + "content": "Table 4. Quantitative comparisons on three real-world datasets (-Test) in terms of FID. The best and the second best performances are highlighted and underlined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 285, + 531, + 411 + ], + "lines": [ + { + "bbox": [ + 321, + 285, + 531, + 411 + ], + "spans": [ + { + "bbox": [ + 321, + 285, + 531, + 411 + ], + "type": "table", + "html": "
MethodsLFWWebPhotoWIDER
Input124.974170.112199.961
GPEN [32]50.79280.57246.340
GFP-GAN [27]49.56087.58439.499
VQFR [8]50.86775.34844.107
RestoreFormer [29]47.75077.33049.817
CodeFormer [38]51.86383.19338.784
DR2 [30]45.298112.34445.348
Difface [34]45.22787.81137.112
WaveFace43.17581.52536.913
", + "image_path": "422a93f008cc973ebf5c03fca6b2bd6a4123acb72f558f2eb750b438bbff9a43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "content": "faithfully preserves the identity with the minimum angle between identity embeddings and their HQ counterparts, outperforming the second-best method by 6.5 degrees." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 664 + ], + "type": "text", + "content": "Additionally, we present the qualitative comparison with SOTA methods on images with increasing degradations in Fig. 7. GFPN and GFPGAN generally produce oversmoothed results. VQFR and RestoreFormer introduce obvious artifacts especially when the inputs are corrupted by mild or severe degradation. Besides, the above methods fail to preserve the identity since they are highly dependent on the StyleGAN prior, where the pre-defined latent space limits the diversity in restored results. Diffusion prior-based methods (DifFace and DR2) tend to generate faithful images with minor artifacts. However, due to the lack of condition during generation, both methods still suffer from the failure in identity preservation. On the contrary, our method, with the help of diffusion prior as well as the injected condition, is capable of generating high-quality facial images with fewer artifacts and, meanwhile, is faithful to the original identity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Real-world datasets. We compare with SOTA methods on three real-world datasets in terms of FID score, which measures the KL divergence between the distribution of restored images and that of HQ images in FFHQ [13]. To" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6589" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 542, + 254 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 542, + 254 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 542, + 254 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 542, + 254 + ], + "type": "image", + "image_path": "cbc65886e4c29313ea0ef29234c66e271dd66cfcd2d02542e44137ac495e12f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 255, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 545, + 277 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with state-of-the-art BFR methods on CelebA-Test. Our method achieves a better restoration quality with the original identity and facial details well preserved. (Zoom in for best view)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 287, + 542, + 472 + ], + "blocks": [ + { + "bbox": [ + 48, + 287, + 542, + 472 + ], + "lines": [ + { + "bbox": [ + 48, + 287, + 542, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 542, + 472 + ], + "type": "image", + "image_path": "416d300d67e0ad204adef5a921a2582887f3811fe53647558c22a702d7a9de81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 475, + 544, + 498 + ], + "lines": [ + { + "bbox": [ + 46, + 475, + 544, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 544, + 498 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison with state-of-the-art BFR methods on real-world datasets, including LFW-Test (first row), WebPhoto-Test (second row), and WIDER-Test (third row). (Zoom in for best view)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 651 + ], + "type": "text", + "content": "reduce the large gap between the synthetic degradations and real-world ones, following previous diffusion prior-based methods [18, 34], a pre-denoising network [18] is applied before our method to handle the unknown degradations. Quantitative results with SOTA methods are reported in Tab. 4. WaveFace outperforms state-of-the-art (SOTA) methods on LFW-Test [11] and WIDER-Test [38] and beats previous diffusion model based BFR methods on WebPhoto-Test [27]. As mentioned in DiffFace [34], the FID score may not be the most suitable evaluation metric for WebPhoto-Test, as a total of 407 images is significantly distant from representing the data distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Three typical examples are shown in Fig. 8 with the increasing degradation level. It is observed that previous BFR methods tend to generate either hazy or unnatural faces. Some methods even fail to generate adequate restored results when handling severe degradations (Row 3). In con" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 508, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 531 + ], + "type": "text", + "content": "trast, our approach provides much more natural and realistic results with rich details such as wrinkles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 540, + 379, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 379, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 379, + 553 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "type": "text", + "content": "We propose WaveFace to solve the task in the frequency domain to achieve the trade-off between efficiency and authenticity. Efficiency is achieved by applying a diffusion model only on the low-frequency sub-band whose size is 1/16 of the original image. Meanwhile, high-frequency components decomposed at multiple DWT levels are recovered simultaneously by a one-forward framework, which ensures the preservation of facial details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Limitations. There is a considerable gap between real-world degradations and the simulated ones (Eq. (15)), leading to inferior results in some cases. Our future work will focus on how to simulate degradations that fit real-world scenarios." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6590" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "text", + "content": "[2] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "text", + "content": "[3] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. FSRNet: End-to-end learning face super-resolution with facial priors. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 247 + ], + "type": "text", + "content": "[4] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[6] Jiankang Deng, Anastasios Roussos, Grigorios Chrysos, Evangelos Ververas, Irene Kotsia, Jie Shen, and Stefanos Zafeiriou. The menpo benchmark for multi-pose 2d and 3d facial landmark localisation and tracking. IJCV, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 328, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 360 + ], + "type": "text", + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 363, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 406 + ], + "type": "text", + "content": "[8] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In ECCV, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 453, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 474 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 475, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 520 + ], + "type": "text", + "content": "[11] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Report, 2008. 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "text", + "content": "[12] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[13] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 589, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 622 + ], + "type": "text", + "content": "[14] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "text", + "content": "[15] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 657, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 679 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[17] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In ECCV, 2018. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In ICCV, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "text", + "content": "[20] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "text", + "content": "[21] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "text", + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "text", + "content": "[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[24] Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming-Hsuan Yang. Deep semantic face deblurring. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "type": "text", + "content": "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[26] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 393, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 426 + ], + "type": "text", + "content": "[27] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 2004. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[29] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In CVPR, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "text", + "content": "[30] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. DR2: Diffusion-based robust degradation remover for blind face restoration. In CVPR, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 548, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 569 + ], + "type": "text", + "content": "[31] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In CVPR, 2016. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[32] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 2, 7, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "text", + "content": "[33] Xin Yu, Basura Fernando, Bernard Ghanem, Fatih Porikli, and Richard Hartley. Face super-resolution guided by facial component heatmaps. In ECCV, 2018. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "text", + "content": "[34] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv:2212.06512, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[35] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 1" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6591" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 184 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[37] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[38] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6592" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_content_list.json b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..55f55e5b58f948dae5f2a02b963711ed75be37b6 --- /dev/null +++ b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_content_list.json @@ -0,0 +1,2740 @@ +[ + { + "type": "text", + "text": "WaveMo: Learning Wavefront Modulations to See Through Scattering", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingyang Xie $^{1*}$ Haiyun Guo $^{2*}$ Brandon Y. Feng $^{3}$ Lingbo Jin $^{2}$ Ashok Veeraraghavan $^{2}$ Christopher A. Metzler $^{1}$", + "bbox": [ + 212, + 179, + 754, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Maryland", + "bbox": [ + 120, + 232, + 318, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Rice University", + "bbox": [ + 356, + 233, + 491, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Massachusetts Institute of Technology", + "bbox": [ + 529, + 233, + 843, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imaging through scattering media is a fundamental and pervasive challenge in fields ranging from medical diagnostics to astronomy. A promising strategy to overcome this challenge is wavefront modulation, which induces measurement diversity during image acquisition. Despite its importance, designing optimal wavefront modulations to image through scattering remains under-explored. This paper introduces a novel learning-based framework to address the gap. Our approach jointly optimizes wavefront modulations and a computationally lightweight feedforward \"proxy\" reconstruction network. This network is trained to recover scenes obscured by scattering, using measurements that are modified by these modulations. The learned modulations produced by our framework generalize effectively to unseen scattering scenarios and exhibit remarkable versatility. During deployment, the learned modulations can be decoupled from the proxy network to augment other more computationally expensive restoration algorithms. Through extensive experiments, we demonstrate our approach significantly advances the state of the art in imaging through scattering media. Our project webpage is at https://wavemo2024.github.io/.", + "bbox": [ + 76, + 318, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 681, + 209, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imaging through scattering media presents a significant challenge across diverse scenarios, ranging from navigating with fog [2, 28, 30], rain [9, 48, 49], or murky water to recovering intricate structures through human skin and tissue [10, 42, 47]. The core of this challenge are the irregular phase delays light experiences as it scatters. These phase delays blur and warp any images captured through the scattering media. Effectively addressing this issue is crucial for unlocking new computer vision capabilities in fields such as medical imaging and astronomy.", + "bbox": [ + 75, + 708, + 468, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From a frequency domain perspective, the effect of a", + "bbox": [ + 94, + 859, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bf9bf293682b61dd5a272e1c0d2960a5b835780691dc2254bbf65f4eeeb6ea1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 286, + 883, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/304bc80251eb51cc6c05a9c89f2fd7979305a2a4b3813d9ee7cfb8079200cb39.jpg", + "image_caption": [ + "Figure 1. Learned Wavefront Modulations. Top: During acquisition, we can modulate the wavefront of scattered light by using a spatial light modulator (SLM), and capture a set of image measurements useful for scene reconstruction. Bottom left: We propose learning modulations that enhance our ability to recover the scattered scene. Bottom right: Our learned modulations drastically improve the reconstruction quality of a state-of-the-art method [8] that previously applies randomly chosen modulations." + ], + "image_footnote": [], + "bbox": [ + 506, + 385, + 736, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/200f2d47f9cc1f78dccb64b5e02846836fa8cf306d9b3d76bcf15b88264558c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 386, + 883, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene passing through scattering is the destruction or filtering of frequency content. Recovering these lost frequencies is very difficult using purely computational methods, which are effectively forced to speculate on the missing frequencies based on those that have been preserved. A question naturally arises: Can we modify the measurement process to prevent the loss of information during acquisition?", + "bbox": [ + 496, + 612, + 890, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A promising solution is to actively modulate the scattered wavefront during acquisition, in hopes that applying multiple modulations increases the chance that a frequency component is preserved. This approach is related to phase diversity imaging [11, 15, 17, 25, 27, 37, 41, 45], which typically captures multiple scattered images of a static scene with different modulations applied using deformable mirrors or spatial light modulators (SLM). However, phase diversity imaging traditionally was limited to imaging simple scenes with mild optical aberrations, and the phase modulations were often selected based on simple heuristics. The recent development of neural wavefront shaping (NeuWS) [8]", + "bbox": [ + 496, + 719, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 94, + 887, + 209, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25276", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c47af18287b17a17f98b7d80316709f7af528a3b7cbaa6de49cf4471c5abf8c.jpg", + "image_caption": [ + "Figure 2. Overview of the Proposed End-to-end Learning Framework. During training, we jointly optimize the implicit neural representation for the wavefront modulations and the proxy reconstruction network in an end-to-end fashion. During inference, one can apply the learned wavefront modulations to any reconstruction algorithms for imaging through scattering, either a trained feed-forward reconstruction network or an unsupervised iterative optimization algorithm [8]. The former approach is far faster and performs better when test data and training data fall into similar distributions, while the latter generalizes better to unseen distributions of target scenes." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 854, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "brings a significant breakthrough by leveraging recent innovations in machine learning and enables imaging high-resolution and dynamic scenes through severe aberrations. Still, the wavefront modulations applied by NeuWS during acquisition are chosen randomly without explicit consideration of their impact on frequency preservation.", + "bbox": [ + 75, + 532, + 470, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we take a significant step towards a more systematic and principled approach in designing effective modulations that better preserve frequencies against scattering during image acquisition. Moving beyond the traditional reliance on simple heuristics for determining modulation patterns, our method combines the strengths of differentiable optimization and data-driven learning. Central to our approach is the novel integration of the optical model of wavefront modulations with a proxy reconstruction network, resulting in a fully differentiable system. This integration allows for the simultaneous optimization of both the proxy reconstruction network and the modulation patterns, which are optimized end-to-end based on a large, simulated training dataset. While imaging through scattering often requires iterative optimization algorithms because data-driven feedforward networks often fail to generalize outside their training domain, this paper demonstrates the remarkable finding that training a hard-to-generalize network can", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "become a useful proxy allowing us to learn generalizable modulations. Crucially, this proxy feedforward network allows us to bypass the expensive backpropagation computation through an iterative reconstruction algorithm, which would be necessary, but hopelessly impractical, if we were to naively optimize the modulations: each iteration would require us to finish an entire iterative reconstruction algorithm. As will be shown in our simulated and real data experiments, our learned modulations significantly enhance the reconstruction capability of both our data-driven proxy network and a state-of-the-art iterative optimization-based method [8] for imaging through scattering.", + "bbox": [ + 496, + 532, + 893, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are:", + "bbox": [ + 517, + 715, + 668, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel end-to-end learning framework to optimize acquisition-time wavefront modulations to enhance our abilities to see through scattering.", + "- We demonstrate in both simulated and real experiments that our learned modulations substantially improve image reconstruction quality and effectively generalize to unseen targets and scattering media.", + "- We show that the learned modulations can be decoupled from the jointly-trained proxy reconstruction network and significantly enhance the reconstruction quality of state-of-the-art unsupervised approaches." + ], + "bbox": [ + 514, + 732, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25277", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Imaging Through Scattering Media. A detailed review of imaging through scattering media can be found in [10]. We highlight a few recent learning-based approaches. Various data-driven approaches can be applied to image through scattering [5, 13, 20, 22, 35, 36]. They generally capture or synthesize large quantities of training data and then learn a mapping from corrupted measurements to reconstructed scenes. However, they tend to struggle with out-of-distribution data. There have also been unsupervised algorithms by fitting network-based parameterizations of the scene to a collection of measurements [8, 15], which can generalize across different distributions of aberrations and target scenes. Both of these methods take advantage of phase diversity.", + "bbox": [ + 75, + 114, + 472, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "End-to-End Learning of Optical Systems. End-to-end learning is a flexible learning-based system design framework that describes optical systems using parameterized differentiable optical models which can then be optimized with training data [31]. End-to-end learning has been applied successfully for extended depth-of-field imaging [14, 21, 29, 31], depth estimation [4, 14, 29, 43], high-dynamic-range imaging [23, 33], seeing through near-lens occlusions [30], and many other applications. To facilitate end-to-end learning, prior works have also incorporated the usage of a proxy network to mimic the response of a specific optical system or black-box algorithm [26, 39].", + "bbox": [ + 75, + 344, + 472, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Phase Diversity Imaging. Phase diversity imaging (PDI) can recover a clear image of an aberrated scene by obtaining images with deliberately introduced known aberrations. PDI does not pose strong assumptions on the target [45] or the aberration [16], and the optics and calibration process required are relatively simple compared to other adaptive optics imaging [11]. While early applications of PDI are often constrained to simple aberrations due to computation complexities [32], the introduction of coded diffraction patterns has enhanced its capacity for reconstructing intricate scenes [3]. Recent PDI developments include improved wavefront sensing [45, 50], phase retrieval [27, 44], adaptive optics [7, 15], and imaging through scattering [6, 8, 46].", + "bbox": [ + 75, + 542, + 472, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Learned Wavefront Modulations", + "text_level": 1, + "bbox": [ + 76, + 752, + 377, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we describe our approach for learning wavefront modulations through training a proxy reconstruction network.", + "bbox": [ + 75, + 777, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Scattering Problem", + "text_level": 1, + "bbox": [ + 76, + 816, + 263, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper focuses on the scenario where capturing an image through scattering media can be modeled as", + "bbox": [ + 75, + 839, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = h (\\phi) * x + \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 881, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is the captured measurement, $x$ is the target scene that we aim to reconstruct, $\\epsilon$ is noise, and $h(\\phi)$ is the unknown, spatially invariant point spread function (PSF) describing the optical scattering, which manifests as unknown phase delays $\\phi$ to the wavefront. Assuming the target object is illuminated with spatially incoherent monochromatic light, the PSF $h$ is related to the phase error $\\phi$ by $h(\\phi) = |\\mathcal{F}(m\\circ e^{j(\\phi)})|^2$ , where $\\mathcal{F}$ is the 2D Fourier transform, $\\circ$ is the hadamard product and $m$ is the aperture mask [12]. Our goal is to recover $x$ from $y$ captured through $\\phi$ .", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Phase Diversity", + "text_level": 1, + "bbox": [ + 498, + 253, + 655, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Phase diversity imaging seeks additional information about the phase error $\\phi$ by capturing a sequence of measurements $y_{1}, y_{2}, \\ldots, y_{K}$ , each with additional known phase errors $\\gamma_{1}, \\gamma_{2}, \\ldots, \\gamma_{K}$ . As a result of these additional modulations, the imaging forward model becomes $y_{i} = h(\\phi + \\gamma_{i}) * x + \\epsilon_{i}$ . In traditional approaches, the phase modulations $\\{\\gamma_{i}\\}_{i=1}^{n}$ were typically either simple defocus sweeps or randomly chosen patterns generated from Zernike polynomials [8, 25], and their reconstruction objective seeks to minimize", + "bbox": [ + 496, + 277, + 893, + 414 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P D I} (\\hat {x}, \\phi) = \\sum_ {i = 1} ^ {K} \\left| \\left| y _ {i} - h \\left(\\phi + \\gamma_ {i}\\right) * \\hat {x} \\right| \\right| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 426, + 890, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "over $\\hat{x}$ and $\\phi$ , where $\\hat{x}$ is the estimate of scene $x$ , assuming $\\epsilon_{i}$ follows an i.i.d. additive Gaussian distribution.", + "bbox": [ + 496, + 479, + 890, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Learned Modulations", + "text_level": 1, + "bbox": [ + 498, + 521, + 702, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While previous works have leveraged differentiable proxies of physical systems to design better algorithms [26, 39], this work leverages a differentiable proxy, $\\mathcal{P}$ , of non-differentiable algorithms to design better physical systems—specifically the phase modulation $\\Gamma = \\{\\gamma_i\\}_{i=1}^K$ used to image through scattering.", + "bbox": [ + 496, + 545, + 890, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We sample a training image $x_{n}$ and simulate passing it through an optical aberration drawn from some known distribution while we apply $K$ learnable modulations to the scattered wavefront. This model allows us to simulate a collection of $K$ images $Y_{n}^{\\Gamma} = \\{y_{1}^{n}, y_{2}^{n}, \\ldots, y_{K}^{n}\\}$ , which are different scattered observations of $x_{n}$ based on Eq. (1).", + "bbox": [ + 496, + 637, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then optimize both the proxy reconstruction algorithm $\\mathcal{P}$ and the modulation patterns $\\Gamma$ to minimize", + "bbox": [ + 496, + 728, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l e a r n e d}} (\\mathcal {P}, \\Gamma) = \\sum_ {n = 1} ^ {N} \\| \\mathcal {P} \\left(Y _ {n} ^ {\\Gamma}\\right) - x _ {n} \\| ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 771, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the ultimate goal of the learning is not to design a set of modulations specific to the network $\\mathcal{P}$ . Rather, we use the performance of $\\mathcal{P}$ as a proxy to probe how effective the modulations $\\Gamma$ are for the scattering problem. The differentiability of $\\mathcal{P}$ allows for back-propagation to guide", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25278", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the optimization of the modulations. As we will demonstrate, while $\\mathcal{P}$ itself may not generalize outside of its training data domain, the learned modulations $\\Gamma$ are compatible with other more generalizable reconstruction algorithms.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our end-to-end training pipeline is illustrated in Figure 2. For the proxy reconstruction network, we choose a U-Net with skip connections and self-attention modules [24]. To regularize the highly non-convex problem of optimizing modulations, we use an implicit neural representation for the modulations $\\Gamma$ . Specifically, the MLP $G$ takes a fixed vector $Z$ with 28 channels, which corresponds to the first 28 Zernike polynomials [18]; it outputs a 16-channel vector corresponding to 16 modulation patterns: $\\Gamma_{G} = G(Z)$ .", + "bbox": [ + 75, + 152, + 470, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incorporating the implicit neural representation for the modulations, our final loss function becomes", + "bbox": [ + 76, + 287, + 470, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n a l}} (\\mathcal {P}, G) = \\sum_ {n = 1} ^ {N} \\left\\| \\mathcal {P} \\left(Y _ {n} ^ {\\Gamma_ {G}}\\right) - x _ {n} \\right\\| ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 325, + 468, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Y_{n}^{\\Gamma_{G}} = \\{h(\\phi_{n} + G(Z)_{i})*x_{n}\\}_{i = 1}^{K}$ .", + "bbox": [ + 76, + 375, + 357, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Applying Learned Modulations", + "text_level": 1, + "bbox": [ + 76, + 398, + 354, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the trained proxy network $\\mathcal{P}$ demonstrates considerable capability in recovering scenes from scattering in our experiments, supervised data-driven methods often fail to generalize to novel or cross-domain scenes, particularly when compared to unsupervised approaches.", + "bbox": [ + 75, + 422, + 468, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, after obtaining the learned modulations $\\Gamma_{G}$ and applying them during real-world acquisition, we send the resulting modulated measurements to an unsupervised iterative optimization-based reconstruction algorithm [8] which does not suffer from the generalization issue of data-driven methods. In effect, our method offers a best-of-both-worlds scenario: the modulations $\\Gamma_{G}$ are effectively learned thanks to the joint supervised training with the proxy network $\\mathcal{P}$ , but their enhanced data acquisition quality is transferrable to other reconstruction algorithms, thus allowing us to benefit from the generalization and domain adaptability inherent in unsupervised reconstruction methods.", + "bbox": [ + 75, + 498, + 470, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Analysis of Learned Modulations", + "text_level": 1, + "bbox": [ + 76, + 691, + 380, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before we show the performance of the learned modulations with real-world experiments in Section 6, we first analyze the effectiveness of learned modulations in simulation.", + "bbox": [ + 75, + 717, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Impact on Frequency", + "text_level": 1, + "bbox": [ + 76, + 771, + 277, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the scattering media scrambles the incoming wavefront, it destroys some of the target scene's frequencies, either by destructively interfering with them or by scattering them outside the imaging sensor. If those frequencies are not captured in the measurement, it becomes difficult to faithfully recover them. By applying multiple learned wavefront modulation patterns, we increase the chance that each", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4261602f59e9f5979a37357f0aefcc7dd2e380c08a7fc2324d923bb286037c05.jpg", + "image_caption": [ + "Figure 3. MTF Comparison. We compare the MTF of learned wavefront modulations vs. the MTF of modulations randomly drawn from the same distribution of the scattering media. The X-axis represents spatial frequency, and the Y-axis represents the modulation transfer (the higher the better). Our learned wavefront modulations exhibit a higher MTF compared to unoptimized ones, especially over higher frequency bands, suggesting the learned modulations preserve more high-frequency information." + ], + "image_footnote": [], + "bbox": [ + 544, + 92, + 834, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frequency is better preserved by at least one of the modulations. With the extra information from all modulated measurements, a better image reconstruction performance could be achieved regardless of the reconstruction algorithm.", + "bbox": [ + 496, + 428, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we assess the performance of our learned modulations by their combined ability to preserve frequencies. We choose to the commonly used Modulation Transfer Function (MTF) as our metric [1, 19, 34, 38, 40]. MTF measures the contrast produced by an imaging system at different spatial frequencies, defined as $\\mathrm{MTF}_i^\\omega (\\phi) = \\left|\\mathcal{F}[h_i(\\phi)]\\right|^{\\omega}$ , where $h_i$ is the PSF for the $i$ -th modulation and the superscript $\\omega$ denotes indexing the MTF at a specific frequency $\\omega$ . For the multiple modulations in our problem setup, we define the combined MTF as:", + "bbox": [ + 496, + 489, + 892, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M T F} _ {\\text {c o m b}} ^ {\\omega} \\left(\\left\\{\\gamma_ {i} \\right\\}, \\phi\\right) = \\max _ {i} \\mathrm {M T F} _ {i} ^ {\\omega} \\left(\\phi + \\gamma_ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 654, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where for each frequency $\\omega$ , we take the maximum over the MTFs calculated from all modulations.", + "bbox": [ + 498, + 688, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Figure 3, we plot the combined MTF for both the learned wavefront modulations and the random ones as baseline. The learned ones have a clear advantage at high frequencies (the higher the frequency is, the further away it is to the center of the X-axis), suggesting that the learned modulations are better at preserving finer details of the target scene. Note that during training, we never explicitly encourage the MTF to be higher — the only loss term we use is the mean square error between reconstruction and ground truth image. This implies that our end-to-end training implicitly encourages the modulations to preserve higher-frequency information.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25279", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae4e87b6a0096d2b85d16fb5f7643b297f82d5925d9807e5cfd4de11e4f582c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 184, + 154 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28dfa595812ef7095784887990f417e343e725e24eed3b9e237c7cd6941ecaff.jpg", + "image_caption": [ + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 186, + 88, + 271, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f6a105f7e6971295d2476e263a677718acd7cf9031db511b9dae349e1bbc542.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 88, + 356, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4bf19fab200664f039d44da3ace70f90fa1a9f7fe36032282e2d38064744f047.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 88, + 442, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b7eae1b38ff40e8b4237810bc2def67b1e982ebbda40c71c3ac48c8861f4645c.jpg", + "image_caption": [ + "Reconstructions with no modulations" + ], + "image_footnote": [], + "bbox": [ + 99, + 162, + 184, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a90d055f6c28e5e53d0060d5ead0525c35b2d673fe8fe179ed24111ba04613bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 162, + 271, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/07435ec1a0c0b0fe25f8ce719dac3fad15115e8e6be4308c61f5ba211e20d3c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 162, + 356, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ea33a876323ea203b88fe116e5bbbb111c9fdade108c5aac0ee6981746113eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 162, + 442, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a54ef6f114dc57e21a90260ebe699df4b313c61301812e6e9929f48109855043.jpg", + "image_caption": [ + "Reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 99, + 238, + 184, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c83fcedd195d827d19f67041e331862c349fc697c7cfe79b54662748a72ab367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 238, + 271, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e03cce180316ff1548af11d49c2e48c55efd038d4b400208f4cdcc2ac9da85a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 238, + 356, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73db0ee5fd56def69b19dd9935954cd437312fad5a09aef4a292d652528b8a20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 238, + 442, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a07707edb5bdbb42878c0cb04158088039bd01c9f31fe2a96a1aa905756b456.jpg", + "image_caption": [ + "Reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 99, + 313, + 184, + 377 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/765ea69ced7093f2a9666955b35415c4d88ce303cc154f0e19f4312994c90522.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 313, + 271, + 377 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0274790ee82953a8882b85180b0f99a4e63673a74fe3c300a1936ffd676d4a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 313, + 356, + 377 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59bfb4d26c9e4ab586d05dec843a3713ff5fe37d2906257ef48a085768018ecb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 313, + 442, + 377 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff9015397632b08ab254d48afd8da86cb26a708886cf99b10532fcffc21cd37a.jpg", + "image_caption": [ + "Figure 4. Proxy Network Reconstruction on Simulated Scattering. Simulated imaging results through different aberrations. Learned modulations lead to a better quality." + ], + "image_footnote": [], + "bbox": [ + 99, + 387, + 184, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dbbb70a0505a226bbf409309e93ea1b064454a72f7503830370a1c03f8f799d6.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 184, + 387, + 271, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/94dcc419b901c3191c3367c89c84f6ed6d094baae07e0f5a1093866d54a97b46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 387, + 356, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6c3338a0a8b51c74419e11295df554fd0cb74f5ada46d27e466bdcb9817fa0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 387, + 442, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Impact on Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 544, + 313, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our jointly learned modulations and proxy network both quantitatively and qualitatively. Both the training and test set are from the MIT Places 365 Dataset [51]. Training details are provided later in Section 5. Here, we use two baselines: (1) simply training a feed-forward reconstruction network without any wavefront modulations, and (2) training a feed-forward reconstruction network with randomly chosen, fixed wavefront modulations as used in prior PDI work. For a fair comparison, our approach and baselines use the same network architecture and are trained on the same data for the same number of iterations. Their only difference is the usage of wavefront modulations.", + "bbox": [ + 75, + 566, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 4, compared to baselines, reconstructions using learned modulations demonstrate finer details and higher consistency with the ground truth. Table 1 shows the quantitative comparison averaged over a test set of 1,000 images, where our approach outperforms the baselines by over $3.9\\mathrm{dB}$ . While the results above are all from the jointly trained proxy reconstruction network, we will show in Section 6 that our learned wavefront modulations can decouple from the proxy network and also improve reconstruction quality based on iterative optimization [8].", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f6a8b3ab373184aff93be4db4ead0e29cfd2c7f6b81fb083c2d806f07ccbd6a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricModulations
NoneRandomLearned
PSNR25.47626.43930.391
SSIM0.76400.79800.9082
", + "bbox": [ + 557, + 80, + 836, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. PSNR and SSIM of Simulation Results. We report the test-time performance of our jointly trained modulations and the proxy network. We compare it against two baselines where either we do not use wavefront modulations or use randomly chosen, unoptimized modulations. The metrics are averaged over the reconstruction of 1000 test images, each of which is scattered with a randomly sampled optic aberration that is unseen during training. The results demonstrate the huge performance improvement from the optimization of wavefront modulations.", + "bbox": [ + 498, + 165, + 893, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Implementation", + "text_level": 1, + "bbox": [ + 500, + 310, + 658, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Software Implementation & Training. For the proxy reconstruction network, we use an attention U-Net [24]; for the MLP that represents the 16 wavefront modulations, we use a two-layer MLP with leaky-ReLU activation. In each training iteration, the optical aberrations are sampled randomly from Zernike basis functions. The standard deviation of the coefficients for each Zernike basis is randomly generated from a uniform distribution from 5 to 6. We train our proposed framework for 2 million iterations on the MIT Places Dataset [51] with the Adam Optimizer and a learning rate of 0.001. Training took 12 hours on an NVIDIA RTX A6000 GPU. For experimental results, we finetune our trained model on the captured bio-tissue data under the same training settings except with a learning rate of 0.0001.", + "bbox": [ + 496, + 335, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hardware Implementation. Our optical configuration is depicted in Figure 6. A continuous-wave laser with a wavelength of $532\\mathrm{nm}$ was collimated and subsequently passed through a laser speckle reducer (Optotune LSR-4C-L). The resulting spatially incoherent light was then polarized and reflected by a digital micromirror device (DMD). The DMD (LightCrafter™ 4500) display is capable of rapidly presenting 8-bit target images. For ease of assembly, a reflecting mirror was incorporated after the DMD. Then, the light is focused by lens L0 ( $f = 100\\mathrm{mm}$ ) onto the front focal plane of L1, a $10\\times$ microscopic objective. Our optical aberration—resin painted on glass—was placed at the same location. L1 and a tube lens L2 ( $f = 200\\mathrm{mm}$ ) constitute a $4f$ system, and thereby the SLM is manipulating the Fourier plane. We used a HOLOEYE LETO-3 phase-only SLM to modulate the wavefront. Finally, the light reflected by the SLM was imaged on the camera ( $1384\\times 1036$ pixel Grasshopper3) via L3 ( $f = 500\\mathrm{mm}$ ). Our capture speed is limited to 30 fps by the camera frame rate.", + "bbox": [ + 496, + 558, + 892, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Dataset. Our experimental data are from human pathological slides, imaged with optical microscopy (Zeiss Microscope Axio Imager.A2) and a $10 \\times$ objective.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25280", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/efb0def738077b443752410085a3f2ff235ac1f2fa9b6b0359563b79b30ff3e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 87, + 222, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a084187dec390a84202560764ea945d0e773c5655ff109dda528d9ff465100d.jpg", + "image_caption": [ + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 230, + 88, + 354, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0901dc6b4518e7728ccd61e687c56212ac5ed65bb477e5e033ce5c80f4bde7e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 89, + 486, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99fb54fef4243cba83d8ea5a83c5b3f99cdde1d1822ecbb0ff83f4c15e3ae0ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 146, + 222, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/017253ac0be3850a7dc8436278138eda3f0876b69c0866a24795d687c9d94ef5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 146, + 354, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/678c56530a8d64c3cc9866631ac7033802f18935bcdb51a7394b9622f6419919.jpg", + "image_caption": [ + "Proxy reconstructions with no modulations" + ], + "image_footnote": [], + "bbox": [ + 364, + 147, + 486, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b09bdd01b6600f5fff9170102afac143eb9a046fa5b03da78c1c0b32c2d3b979.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 205, + 222, + 252 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77829edce75a451cca4a3d9b90a25e70a7031629ab24cbe804c4c9f4b0da8346.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 205, + 354, + 252 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd70babec234fc43193b89a0fe4201e9d76549eaa7058233d24868d85c795333.jpg", + "image_caption": [ + "Proxy reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 364, + 205, + 486, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2eb7d0536d619d6957595f3b10efbee998f9a9f79113f83f0e4e97c07fcf55a5.jpg", + "image_caption": [ + "Proxy reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 98, + 265, + 223, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67e363f3c957ac5054b98fd49222ec208a0a19f6d7397a28e157d8f154b5f017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 265, + 354, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cdff0b9ff06fe6e244506cc6fcd7062080e16f58099d718e97a05f3e5781b72a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 265, + 486, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5102bc387ccfd1feb69e400d35b427ba3633987512a0d5bd5382d12f890bd2ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 323, + 223, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1ee28985945547b92e629f6c73ac5dd412b1a2e7b8adc4acc0f463b5efdc71d.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 230, + 323, + 354, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff614c8f3bb02c66dc9b220227f512a8a8f62d0121dcea219c803ed67eb2cb2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 323, + 486, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae6a1c1e8d7890f605985e7c8c7ae79c2b082c7ce3963205b759a04754b0a18c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 88, + 602, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f1b36c7ceb4aa164ecfa481216f4e957b1d0baf68e8f6f0bde496a3fedd1f5a.jpg", + "image_caption": [ + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 609, + 89, + 692, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5af787bd1e651aeb6942d512bb902ad990971de211f7174729b04e8d90c075cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 147, + 602, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/490b90633c61a91c1aa522829af3f8ef72f23d35977c614db54f2ceb0c3b6260.jpg", + "image_caption": [ + "Proxy reconstructions with no modulations" + ], + "image_footnote": [], + "bbox": [ + 611, + 146, + 692, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2380e34cbc0e12fce18d88eccb74ceae3d84017738f4adc66c116ec269c4c34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 779, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abafee8c49e3d8efb787059a909e2457441a33f022cadcd4fdc50ae1bb2d75a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 89, + 870, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/48d4de392fb768b2887b55b34cc2bca452b886134584198785e83729bfd733c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 205, + 602, + 252 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bd1eedbebac7de60e8e56739ea831b282de7bd7b755b93b460b6d7d979e60d4.jpg", + "image_caption": [ + "Proxy reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 611, + 205, + 692, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99aedd39cd4cf623c81d3843ad4244c93bee540e4f8618c57e0cfa79332c339d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 205, + 779, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/466bb3b2d837ca7cb59b9a29f5b7dc1cf0f4a76dffef0c20da7ebc132cb1ba4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 205, + 870, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aec3bc163077b361c53ee3e625baa4dcbd35eea14cd50a050b2a5d6d5ba34a82.jpg", + "image_caption": [ + "Proxy reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 521, + 265, + 602, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/985a9ad8bd7e4fa9006b817352bde2ec0bc92f345c3eb163b7c86ae8d2b2c922.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 266, + 692, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c6dd77a2df95c671cfb79dbdc1e764551ccd140cbc7e70bfe9117e014a106d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 266, + 779, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9df6e123028dd705da029dcde65ab68eb86c9fb2830df4c786dfc9c69fdcaca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 266, + 870, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/986f856fa960b339196195b5633521a6a9b7e16ab8916038979e390b0e2ef008.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 323, + 602, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75902cf97aa51d39031adcf85e1add4f175b673b2628a5121d0ff79e5aaee632.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 611, + 323, + 692, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eadc92a5c3c55b9cc5083e46762f2f05cfde3b4fe7ab931da9acd7a7b7b47767.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 323, + 779, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6f11c72e1e045c163da6f62efd2c824c3a9413d9fc549a6f495f7fc4595a875.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 323, + 870, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/368f466ffc48bb6881c0e2ac3a055e520f6ac42c36ad68692da3ffdf6ceda76b.jpg", + "image_caption": [ + "Figure 5. Proxy Network Reconstruction on Physical Scattering. Experimental results of imaging objects through scattering media by using our proxy reconstruction network. The left columns are in-distribution adipose tissue slides (zoomed-in region labeled with red boxes); the right columns are out-of-distribution targets. Learned modulations yield superior imaging quality for both in-distribution and out-of-distribution scenes, with the former out-performing the latter.", + "Figure 6. Optical Setup. A spatially incoherent light source illuminates the DMD and passes through thick scattering. The DMD displays the target images for training and testing. The SLM is placed onto the Fourier plane and projects learned patterns. The wavefront is modulated and subsequently imaged onto a camera." + ], + "image_footnote": [], + "bbox": [ + 101, + 478, + 455, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each mage is $256 \\times 256$ pixels. Our dataset includes 1,000 sectional images of adipose tissues. 40 of these images were utilized for testing and the rest were used for training. To evaluate the generalizability across different human tissues, we also imaged sections of human stomach and glandular epithelium, trachea, and finger. Additionally, we tested with targets outside the domain of the training data, such as numbers and letters.", + "bbox": [ + 75, + 779, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 479, + 705, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employed the optical system depicted in Figure 6. Our learned modulations significantly enhance the imaging capabilities through scattering media. We tested on datasets from both similar and distinctly different distributions from our experimental training data. Our learned modulations consistently yield superior visual reconstruction results. Furthermore, we evaluated these modulations on both static and dynamic scenes. We show that our learned modulations significantly enhance reconstruction in both cases, which suggests a broad applicability of our approach.", + "bbox": [ + 496, + 503, + 890, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Proxy Reconstruction Network", + "text_level": 1, + "bbox": [ + 498, + 665, + 772, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proxy network, fine-tuned on 960 adipose tissue images through random scattering, was evaluated on a test set from the same tissue type, as shown in the left of Figure 5. Reconstructions driven by measurements of 16 learned SLM patterns preserved the original contrast and high-frequency details of the dyed sections. The close-ups show remarkable detail accuracy. By comparison, reconstructions trained on measurements of random or no modulation struggled, particularly in lower-intensity areas, yielding blurry outputs. Beyond achieving good reconstruction fidelity on objects similar to our training dataset, we extended our evaluation to vastly different data distributions, like numbers and letters. The results, as shown on the right of Figure 5, clearly indicate that the learned modulations yield bet", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25281", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afb40c5c8aa46924fa4adef9c37a095b00dc2eb993d3696ef89ffd0e33708bd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 87, + 243, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbdbcdccf6a040f07ba99f2b32deb2b255721c6ee5f5e14f2963428003b90221.jpg", + "image_caption": [ + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 245, + 88, + 321, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5ec909d020f9af775c4f060389c9b9e0f23a4af85d22c75de3b9362fee2152d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 88, + 400, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/677f83b23404db4457714b2069b4195407d322f26be3632688bb62706f54b9f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 88, + 478, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88f52109252054f3ea82412818902bb7390773bc866df84d9aa2f1cd2f985628.jpg", + "image_caption": [ + "Reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 166, + 143, + 243, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/685d1e10ecb46f8061a6ef648a991a0200732dec64bb1b70f71b078b9f6c0db2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 143, + 321, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebc2a86cbdde8bd93fa091212e4a008b06d39b315b90e9861ea501a82d940532.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 143, + 400, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59d4e481cc39a2cacb930b594a70258db07a37de09efdf5c4eb4698d4c98c7af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 143, + 478, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81dbc391823d09fe51174ee9302c1c9eb645042fb1edc57c321a10962628141e.jpg", + "image_caption": [ + "Reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 166, + 199, + 243, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8959e7e1197f1ab2ebd39e666329379b6b5cda0927e7842ceda5e69fd4219f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 199, + 321, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b100a7ce931ab83f45f9714728e9d4754c6782a594fc99bcc2f3960e85d2aeeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 199, + 400, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1ae4d57b0d99422323689dcb71fc78d08030e32974a254d1fc5bbe41e8a2e40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 199, + 478, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/459f212daa3216bf97e44e9a3c02aeb086265fce8a51cdab05e6cd5a6c2b4b1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 256, + 243, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f09e314b5549379163781f9c03487c4e0707557bb5f5bd853dddb818eab382cb.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 245, + 256, + 321, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83994e5c0ec333b4073d6c541913d769097b7c880663ca1d0f1ff831d74e993c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 256, + 400, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf1a1f53e06f9600a11991bd0bd3ab3b5a51ec7ecc8ff6618552c608adfcb847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 256, + 478, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d06de3fcda707bab0908c51b786bdf60a10fa5c6a5a407760af2dc6f3b30da85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 571, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80047a784f6a8ad7ac5a5f6c131068c08faf85484025ea771690641e873f8f56.jpg", + "image_caption": [ + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 573, + 88, + 650, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b8259b8e9594a449e2d2c67bbbd71b3ac1cc86e7cc2776684b20d51846d36d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 88, + 728, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/805ea4f44c7eaec65598ea262307186eb035c3b6c7efbfc9cbc65043e639a9d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 88, + 807, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/900a0a3fd09e77f5b8550753df53f335732515fc8811d6e67fe298e048bd8275.jpg", + "image_caption": [ + "Reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 496, + 143, + 571, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d85aee4487be29256b2d8c5985b0ac6404a2197ad76bf305688cd8413398f728.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 143, + 650, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b22cb5527af65eb55f7636ec03d05a3fedbd0df023a6bd346ade4e8b2e046598.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 143, + 728, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ad583a53f7dfcfd2b902248894ac1943f1faaf12ec1d06e6eb4a98bf1e8f65c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 143, + 807, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b44daed79ff0f7c13cdca1bf51e374f783eb258fdd5a5ba3a095b71ec07adaa9.jpg", + "image_caption": [ + "Reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 496, + 199, + 571, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfdd2253d7bd23affd6347131956487e034518993de674471dce2567a5d9a2c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 199, + 650, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fb49de8fd2ce4a659af6c1c0a33e874fb426ea12f409cc4ba0df3faba889d6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 199, + 728, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c2b839b53c3bbd24ad825baa0f2f05cc0847ff18b79a0323ab4207bfcf787a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 199, + 807, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a71cc18bce7e9b3c2734761990901eff097d4c9ef411b6c4cfdcd4ac0e4f3f84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 256, + 571, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97082a8b6a172ee907c63785d1cafd76ec07b2f1ffcb4603b224238d20df8ca6.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 573, + 256, + 650, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78198196e6a652de8ff787cc51d1ec06bc55d84f0417c93aa4f56562527722c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 256, + 728, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a8cdfc74378419aa2f61616121353414daa0ca0dd633a20c22c30926d1c11ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 256, + 807, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2347cf0cc00e9110c516fe7114238ce0c5fe57be53bf01b965d8c2893f0295d3.jpg", + "image_caption": [ + "$\\mathrm{t} = 1$", + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 197, + 400, + 279, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e23e5819f8af33754f6da9a94d589d3642447fa5de361a8b5e0fd199cfda7f7b.jpg", + "image_caption": [ + "$\\mathrm{t} = 24$" + ], + "image_footnote": [], + "bbox": [ + 292, + 400, + 372, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a1d637fc435272aa3988bc650a541b33ed575127b9071ff385bdfb2a0a4cac1.jpg", + "image_caption": [ + "t=48" + ], + "image_footnote": [], + "bbox": [ + 388, + 400, + 468, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/898cd7280dbf47e87506ebd82496d00e0d7811947692d7028155b94109aac1d8.jpg", + "image_caption": [ + "Reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 199, + 459, + 279, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eca66562396695e35a5c1a745418ef9a4e34f7270138d648164c703501429961.jpg", + "image_caption": [ + "Reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 292, + 517, + 372, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32b7b0085eccf72f4f45d1ac0b35b5ea88be5485dee7e9ba1305eaaa176870ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 517, + 468, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40e91686a32669be00b49a90291cca80527730f79f722a1ef070578b062713f5.jpg", + "image_caption": [ + "Figure 8. Unsupervised Reconstruction on Physical Scattering with Dynamic Scenes. Experimental results of imaging an opening camera aperture through severe aberrations (left) and imaging non-regional motions (number 6 rotating clockwise and number 5 translating to the right) through severe aberrations. We use the unsupervised iterative reconstruction approach [8]. Learned modulations highly improve the quality of the reconstructed dynamic scene." + ], + "image_footnote": [], + "bbox": [ + 197, + 575, + 279, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/768429f557bd6682e1a10496dffee32d2d4deebac11f2f57b8f0dba019f65dc9.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 292, + 575, + 372, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d42a20717d879276fc69fc5bec855e2cf7672ddb16e6800911e5a0fcdfd8845.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 575, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/995e62438ef49fbb381dd2f429efe7f390dbd3fa60892ac2ce2d89e90947b56f.jpg", + "image_caption": [ + "Figure 7. Unsupervised Reconstruction on Physical Scattering with Static Scenes. Experimental results of imaging different static targets through scattering media using learned wavefront modulations with unsupervised iterative optimization [8]. Learned modulations enhance reconstruction quality." + ], + "image_footnote": [], + "bbox": [ + 496, + 392, + 549, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e44a2e8ea23e66318ec015d8da5a12d9d41378030507c2f520b6e1e5f29147c8.jpg", + "image_caption": [ + "t=24", + "Uncorrected measurements" + ], + "image_footnote": [], + "bbox": [ + 589, + 400, + 669, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b18b1fe9d0af0117ef0b20f6fc02f2cf7844c81a1586e940f4ad9a0e366edded.jpg", + "image_caption": [ + "t=48" + ], + "image_footnote": [], + "bbox": [ + 684, + 400, + 764, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62ac0fb1a45541ab57325828e9e7528b25312aa67de7f2bb9a8a4049fd98be4a.jpg", + "image_caption": [ + "Reconstructions with random modulations" + ], + "image_footnote": [], + "bbox": [ + 496, + 459, + 576, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee77cc7b512e9e3fda6b3faa34c010b6c9f264082e45e2fa5b98de6af4e2fb85.jpg", + "image_caption": [ + "Reconstructions with learned modulations" + ], + "image_footnote": [], + "bbox": [ + 495, + 517, + 576, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/841e13f6c5cbc8eeda4eb41e9a457686b6dcfca959c301b8df6423552c59a0b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 517, + 669, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55fc6a1e498b0883cf57f60195fedce4231db92d95a33e6f2402acabd54f3456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 517, + 764, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9d760e1e9631c5a53c581134b33481a139e43ab85b6a5a8cedd2045c291e9f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 575, + 576, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/892d8c8b2ea90da5b1bc70f7f45b8ded2b9c9d6a9f42ef77a33d0ff1db361b8c.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 594, + 575, + 669, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3cda9da38125a9b7d00e45e8105753ee730200d3502775d8be4358d676c69cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 575, + 764, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ter reconstruction quality, particularly in the USAF target and stars. PSNR and SSIM are computed in Tables 2 and 3. Among the table cells, \"None\" means that the measurements for each target only comprised the unmodulated image (no modulation patterns are applied); \"Random\" means that the modulation patterns are randomly sampled from the same distribution as the aberrations.", + "bbox": [ + 75, + 734, + 468, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Unsupervised Iterative Approaches", + "text_level": 1, + "bbox": [ + 500, + 734, + 805, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While the proxy network $\\mathcal{P}$ exhibits considerable effectiveness in real-world experiments, its performance remains unsatisfactory for out-of-distribution data, which is expected. Nevertheless, our learned modulations are designed to further augment generalizable, untrained reconstruction algorithms based on iterative optimizations. To validate this strategy, we deploy our learned modulations to a recently developed unsupervised iterative approach for imaging through scattering, NeuWS [8], in our benchmark study.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25282", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f423b72b6403d05b9d206bf9bdac0ff44c32dfbb89166a41b07e1062d2cc876.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodData typeModulation
NoneRandomLearned
ProxyTissue16.5317.5719.06
ProxyOut-of-dist.9.349.9010.71
Iterative [8]StaticN/A11.2614.61
Iterative [8]DynamicN/A8.9012.89
", + "bbox": [ + 76, + 82, + 470, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9f9fdcc5bcc3d3d167b5f2939d4b847d180886e2ee654c1d91d03caa249aafa.jpg", + "table_caption": [ + "Table 2. PSNR of Experimental Results. For our jointly trained feed-forward proxy reconstruction network (\"proxy\"), we tested on 40 tissue samples and 8 out-of-distribution scenes, all of which are static. For the iterative method [8], we tested on the same 8 out-of-distribution static scenes. We also tested [8] on 2 dynamic scenes, each with 48 frames. The iterative method relies on multiple wavefront modulations and therefore cannot recover objects with a single measurement, hence the \"N/A\". Compared against random modulations or no modulations, our learned modulations lead to better reconstruction performance for both the proxy network and the unsupervised iterative approach." + ], + "table_footnote": [], + "table_body": "
MethodData typeModulations
NoneRandomLearned
ProxyTissue0.440.480.58
ProxyOut-of-dist.0.290.300.32
Iterative [8]StaticN/A0.210.38
Iterative [8]DynamicN/A0.230.33
", + "bbox": [ + 76, + 357, + 468, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. SSIM of Experimental Results. Same as Table 2 but showing SSIM. Our learned modulations performs better.", + "bbox": [ + 75, + 474, + 468, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 shows the reconstruction performance of NeuWS [8] on 8 static objects, each using only 16 measurements. The superiority of using learned modulations is visually evident in the results. We excluded experiments with no modulations here, due to the algorithm's inability to recover objects using a single unmodulated measurement.", + "bbox": [ + 75, + 525, + 468, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For dynamic scenes, each scene involves 48 frames captured by cycling through the 16 SLM patterns. Figure 8 demonstrates the reconstruction results on dynamic scenes: the result on the left illustrates a camera aperture gradually enlarging; the result on the right shows two different motions to solve the non-regional dynamic reconstruction (the number 6 on the left rotates clockwise at $0.5^{\\circ}$ per frame, while the number 5 translates from the center to the right). Quantitative metrics on the dynamic results are included in Table 2 and 3. Full video results are in the project webpage.", + "bbox": [ + 75, + 617, + 468, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion & Conclusion", + "text_level": 1, + "bbox": [ + 76, + 782, + 310, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Role of the Proxy Reconstruction Network. The proxy network not only aims to reconstruct the images; its performance also indicates how good the wavefront modulations are for the restoration task. Therefore, joint differentiable optimization of both the proxy network and the modulations works in our favor, and the experimental results indeed in", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dicate that this is an effective way of learning useful wavefront modulations. Without this proxy network, we would have to somehow connect the modulation update with the performance of an iterative optimization method, which is impractical due to time and memory constraints.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularization Perspective. To recover missing signal frequencies due to scattering, most existing methods rely on prior information of the scene, either hand-crafted or data-driven. This work rethinks the strategy of incorporating prior knowledge: can we impose a prior over wavefront modulations? The goal here is to let the resulting imaging system capture more information about the target scene through scattering. Our results validate the effectiveness of learning a prior over the wavefront modulation domain.", + "bbox": [ + 496, + 196, + 893, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Applicability to Other Vision Tasks. The end-to-end learning approach can potentially benefit other vision tasks through scattering media, using a task-specific proxy network. For instance, if we were to do object detection or semantic segmentation through scattering media, we can replace the proxy reconstruction U-Net with a network tailored for these two tasks, and perform end-to-end training in the same way as before. Note that we validated our proposed method on a monochromatic imaging system, which is prevalent in medical and scientific imaging. Extending it to broadband imaging tasks, e.g., outdoor navigation, requires the optimization of wavelength-dependent modulation patterns, which we leave for future work.", + "bbox": [ + 496, + 361, + 893, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We proposed a robust end-to-end learning framework integrating the optical scattering model with a proxy image reconstruction network. The learned wavefront modulations can both work with the proxy reconstruction network and augment a generalizable, optimization-based algorithm. We conducted intensive experiments to validate our approach. This work shows the synergy of advanced wavefront modulation techniques with cutting-edge machine learning methods, representing a significant leap forward in computer vision and optical imaging.", + "bbox": [ + 496, + 587, + 893, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 753, + 653, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by AFOSR Young Investigator Program award no. FA9550-22-1-0208, ONR award no. N00014-23-1-2752, seed grant from SAAB, Inc., a seed grant from the UMD Brain and Behavior Institute, NSF Expeditions award no. CCF-1730574, ONR Scattering media award no. N00014-23-1-2714, and NSF PaThs up award no. EEC-1648451. We thank Sachin Shah for helpful discussions and Kevin F. Kelly for providing the DMD.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25283", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jesse K Adams, Dong Yan, Jimin Wu, Vivek Boominathan, Sibo Gao, Alex V Rodriguez, Soonyoung Kim, Jennifer Carns, Rebecca Richards-Kortum, Caleb Kemere, et al. In vivo lensless microscopy via a phase mask generating diffraction patterns with high-contrast contours. Nature Biomedical Engineering, 6(5):617-628, 2022. 4", + "[2] Codruta Orniana Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1798-1805, 2020. 1", + "[3] Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase Retrieval Via Wirtinger Flow: Theory and Algorithms. IEEE Transactions on Information Theory, 61(4):1985-2007, 2015. 3", + "[4] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 3", + "[5] Dongdong Chen, Mike Davies, Matthias J Ehrhardt, Carola-Bibiane Schonlieb, Ferdia Sherry, and Julián Tachella. Imaging with equivariant deep learning: From unrolled network design to fully unsupervised learning. IEEE Signal Processing Magazine, 40(1):134-147, 2023. 3", + "[6] Wei-Yu Chen, Matthew O'Toole, Aswin C Sankaranarayanan, and Anat Levin. Enhancing Speckle Statistics for Imaging Inside Scattering Media. Optica, 9(12):1408-1416, 2022. 3", + "[7] Santiago Echeverri-Chacón, René Restrepo, Carlos Cuartas-Vélez, and Néstor Uribe-Patarroyo. Vortex-enhanced coherent-illumination phase diversity for phase retrieval in coherent imaging systems. Optics letters, 41(8):1817-1820, 2016. 3", + "[8] Brandon Y Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K Sharma, Ashok Veeraraghavan, and Christopher A Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023. 1, 2, 3, 4, 5, 7, 8", + "[9] Xueyang Fu, Jia-Bin Huang, Delu Zeng, Yue Huang, Xing-hao Ding, and John William Paisley. Removing rain from single images via a deep detail network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1715-1723, 2017. 1", + "[10] Sylvain Gigan, Ori Katz, Hilton B de Aguiar, Esben Ravn Andresen, Alexandre Aubry, Jacopo Bertolotti, Emmanuel Bossy, Dorian Bouchet, Joshua Brake, Sophie Brasselet, et al. Roadmap on Wavefront Shaping and Deep Imaging in Complex Media. Journal of Physics: Photonics, 4(4): 042501, 2022. 1, 3", + "[11] Robert A Gonsalves. Phase retrieval and diversity in adaptive optics. Optical Engineering, 21(5):829-832, 1982. 1, 3", + "[12] Joseph W. Goodman. Introduction to Fourier Optics. Roberts & Co., Englewood, Colo, 3rd ed. edition, 2005. 3", + "[13] Xiaowen Hu, Jian Zhao, Jose Enrique Antonio-Lopez, Stefan Gausmann, Rodrigo Amezcua Correa, and Axel Schulzgen. Adaptive inverse mapping: a model-free" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semi-supervised learning approach towards robust imaging through dynamic scattering media. Optics Express, 31(9): 14343-14357, 2023. 3", + "[14] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 3", + "[15] Iksung Kang, Qinrong Zhang, Stella X Yu, and Na Ji. Coordinate-based neural representations for computational adaptive optics in widefield microscopy. arXiv preprint arXiv:2307.03812, 2023. 1, 3", + "[16] Richard L Kendrick, Daniel S Acton, and AL Duncan. Phase-diversity wave-front sensor for imaging systems. Applied Optics, 33(27):6533-6546, 1994. 3", + "[17] Peter Kner. Phase diversity for three-dimensional imaging. JOSA A, 30(10):1980-1987, 2013. 1", + "[18] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 4", + "[19] Juhyun Lee, Jinsoo Jeong, Jaebum Cho, Dongheon Yoo, Byounghyo Lee, and Byoungho Lee. Deep neural network for multi-depth hologram generation and its training strategy. Optics Express, 28(18):27137-27154, 2020. 4", + "[20] Yunzhe Li, Yujia Xue, and Lei Tian. Deep speckle correlation: a deep learning approach toward scalable imaging through scattering media. Optica, 5(10):1181-1190, 2018. 3", + "[21] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, 2022. 3", + "[22] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-Inverse Correlography: Towards Real-Time High-Resolution Non-Line-of-Sight Imaging. Optica, 7(1):63–71, 2020. 3", + "[23] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 3", + "[24] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, M. J. Lee, Mattias P. Heinrich, Kazunari Misawa, Kensaku Mori, Steven G. McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. ArXiv, abs/1804.03999, 2018. 4, 5", + "[25] Richard G Paxman, Timothy J Schulz, and James R Fienup. Joint estimation of object and aberrations by using phase diversity. JOSA A, 9(7):1072-1085, 1992. 1, 3", + "[26] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural Holography With Camera-in-the-Loop Training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 3", + "[27] Nikolaj Reiser, Min Guo, Hari Shroff, and Patrick J La Riviere. Phase diverse phase retrieval for microscopy: Com" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25284", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "parison of gaussian and poisson approaches. arXiv preprint arXiv:2308.00734, 2023. 1, 3", + "[28] Wenqi Ren, Sibo Liu, Hua Zhang, Jin shan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks. In European Conference on Computer Vision, 2016. 1", + "[29] Sachin Shah, Sakshum Kulshrestha, and Christopher A Metzler. Tidy-psfs: Computational imaging with time-averaged dynamic point-spread-functions. ICCV, 2023. 3", + "[30] Zheng Shi, Yuval Bahat, Seung-Hwan Baek, Qiang Fu, Hadi Amata, Xiao Li, Praneeth Chakravarthula, Wolfgang Heidrich, and Felix Heide. Seeing through obstructions with diffractive cloaking. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 1, 3", + "[31] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 3", + "[32] Carlas S Smith, Arnold J den Dekker, Raluca Andrei, Rufus Fraanje, and Michel Verhaegen. Fast phase diversity wavefront sensing using object independent metrics. In Adaptive Optics Systems III, pages 2250-2258. SPIE, 2012. 3", + "[33] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386–1396, 2020. 3", + "[34] Qilin Sun, Congli Wang, Fu Qiang, Dun Xiong, and Heidrich Wolfgang. End-to-end complex lens design with differentiable ray tracing. ACM Trans. Graph, 40(4):1-13, 2021. 4", + "[35] Yu Sun, Zhihao Xia, and Ulugbek S Kamilov. Efficient and Accurate Inversion of Multiple Scattering with Deep Learning. Optics Express, 26(11):14678-14688, 2018. 3", + "[36] Waleed Tahir, Hao Wang, and Lei Tian. Adaptive 3d scattering with a dynamic synthesis network. Light: Science & Applications, 11(1):42, 2022. 3", + "[37] Brian J Thelen, Richard G Paxman, David A Carcarara, and John H Seldin. Maximum a posteriori estimation of fixed aberrations, dynamic aberrations, and the object from phase-diverse speckle data. JOsA A, 16(5):1016-1025, 1999. 1", + "[38] Lei Tian and Laura Waller. Quantitative differential phase contrast imaging in an led array microscope. Optics express, 23(9):11394-11403, 2015. 4", + "[39] Ethan Tseng, Felix Yu, Yuting Yang, Fahim Mannan, Karl St. Arnaud, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Hyperparameter optimization in black-box image processing using differentiable proxies. ACM Transactions on Graphics (TOG), 38(4), 2019. 3", + "[40] Hakan Urey, Ned Nestorovic, Baldwin S Ng, and Abraham A Gross. Optics designs and system mtf for laser scanning displays. In Helmet-and Head-Mounted Displays IV, pages 238-248. SPIE, 1999. 4", + "[41] Curtis R Vogel, Tony F Chan, and Robert J Plemmons. Fast algorithms for phase-diversity-based blind deconvolution. In Adaptive Optical System Technologies, pages 994-1005. SPIE, 1998. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Abbie T Watnik and Dennis F Gardner. Wavefront Sensing in Deep Turbulence. Optics and Photonics News, 29(10): 38-45, 2018. 1", + "[43] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 3", + "[44] Meng Xiang, An Pan, Jinpeng Liu, Teli Xi, Xin Guo, Fei Liu, and Xiaopeng Shao. Phase diversity-based fourier psychography for varying aberration correction. Frontiers in Physics, 10:129, 2022. 3", + "[45] Qi Xin, Guohao Ju, Chunyue Zhang, and Shuyan Xu. Object-independent image-based wavefront sensing approach using phase diversity images and deep learning. Optics express, 27(18):26102-26119, 2019. 1, 3", + "[46] Tomer Yeminy and Ori Katz. Guidestar-Free Image-Guided Wavefront Shaping. Science Advances, 7(21), 2021. 3", + "[47] Seokchan Yoon, Moonseok Kim, Mooseok Jang, Youngwoon Choi, Wonjun Choi, Sungsam Kang, and Wonshik Choi. Deep Optical Imaging Within Complex Scattering Media. Nature Reviews Physics, 2(3):141-158, 2020. 1", + "[48] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022. 1", + "[49] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 1", + "[50] Peiguang Zhang, Chengliang Yang, Zihao Xu, Zhaoliang Cao, Quanquan Mu, and Li Xuan. High-accuracy wavefront sensing by phase diversity technique with bisymmetric defocuses diversity phase. Scientific Reports, 7(1):15361, 2017. 3", + "[51] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40:1452-1464, 2018. 5" + ], + "bbox": [ + 501, + 92, + 890, + 657 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25285", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_model.json b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5e1776c8b4593fc0dab4e1762d436b65c2bed91 --- /dev/null +++ b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_model.json @@ -0,0 +1,3465 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.843, + 0.154 + ], + "angle": 0, + "content": "WaveMo: Learning Wavefront Modulations to See Through Scattering" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.18, + 0.756, + 0.218 + ], + "angle": 0, + "content": "Mingyang Xie\\(^{1*}\\) Haiyun Guo\\(^{2*}\\) Brandon Y. Feng\\(^{3}\\) Lingbo Jin\\(^{2}\\) Ashok Veeraraghavan\\(^{2}\\) Christopher A. Metzler\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.233, + 0.32, + 0.253 + ], + "angle": 0, + "content": "1University of Maryland" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.234, + 0.492, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Rice University" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.234, + 0.844, + 0.253 + ], + "angle": 0, + "content": "3Massachusetts Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Imaging through scattering media is a fundamental and pervasive challenge in fields ranging from medical diagnostics to astronomy. A promising strategy to overcome this challenge is wavefront modulation, which induces measurement diversity during image acquisition. Despite its importance, designing optimal wavefront modulations to image through scattering remains under-explored. This paper introduces a novel learning-based framework to address the gap. Our approach jointly optimizes wavefront modulations and a computationally lightweight feedforward \"proxy\" reconstruction network. This network is trained to recover scenes obscured by scattering, using measurements that are modified by these modulations. The learned modulations produced by our framework generalize effectively to unseen scattering scenarios and exhibit remarkable versatility. During deployment, the learned modulations can be decoupled from the proxy network to augment other more computationally expensive restoration algorithms. Through extensive experiments, we demonstrate our approach significantly advances the state of the art in imaging through scattering media. Our project webpage is at https://wavemo2024.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.683, + 0.21, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.47, + 0.86 + ], + "angle": 0, + "content": "Imaging through scattering media presents a significant challenge across diverse scenarios, ranging from navigating with fog [2, 28, 30], rain [9, 48, 49], or murky water to recovering intricate structures through human skin and tissue [10, 42, 47]. The core of this challenge are the irregular phase delays light experiences as it scatters. These phase delays blur and warp any images captured through the scattering media. Effectively addressing this issue is crucial for unlocking new computer vision capabilities in fields such as medical imaging and astronomy." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.861, + 0.47, + 0.876 + ], + "angle": 0, + "content": "From a frequency domain perspective, the effect of a" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.287, + 0.885, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.386, + 0.738, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.387, + 0.884, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Figure 1. Learned Wavefront Modulations. Top: During acquisition, we can modulate the wavefront of scattered light by using a spatial light modulator (SLM), and capture a set of image measurements useful for scene reconstruction. Bottom left: We propose learning modulations that enhance our ability to recover the scattered scene. Bottom right: Our learned modulations drastically improve the reconstruction quality of a state-of-the-art method [8] that previously applies randomly chosen modulations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.719 + ], + "angle": 0, + "content": "scene passing through scattering is the destruction or filtering of frequency content. Recovering these lost frequencies is very difficult using purely computational methods, which are effectively forced to speculate on the missing frequencies based on those that have been preserved. A question naturally arises: Can we modify the measurement process to prevent the loss of information during acquisition?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.902 + ], + "angle": 0, + "content": "A promising solution is to actively modulate the scattered wavefront during acquisition, in hopes that applying multiple modulations increases the chance that a frequency component is preserved. This approach is related to phase diversity imaging [11, 15, 17, 25, 27, 37, 41, 45], which typically captures multiple scattered images of a static scene with different modulations applied using deformable mirrors or spatial light modulators (SLM). However, phase diversity imaging traditionally was limited to imaging simple scenes with mild optical aberrations, and the phase modulations were often selected based on simple heuristics. The recent development of neural wavefront shaping (NeuWS) [8]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.21, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25276" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.094, + 0.856, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.896, + 0.507 + ], + "angle": 0, + "content": "Figure 2. Overview of the Proposed End-to-end Learning Framework. During training, we jointly optimize the implicit neural representation for the wavefront modulations and the proxy reconstruction network in an end-to-end fashion. During inference, one can apply the learned wavefront modulations to any reconstruction algorithms for imaging through scattering, either a trained feed-forward reconstruction network or an unsupervised iterative optimization algorithm [8]. The former approach is far faster and performs better when test data and training data fall into similar distributions, while the latter generalizes better to unseen distributions of target scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.471, + 0.625 + ], + "angle": 0, + "content": "brings a significant breakthrough by leveraging recent innovations in machine learning and enables imaging high-resolution and dynamic scenes through severe aberrations. Still, the wavefront modulations applied by NeuWS during acquisition are chosen randomly without explicit consideration of their impact on frequency preservation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we take a significant step towards a more systematic and principled approach in designing effective modulations that better preserve frequencies against scattering during image acquisition. Moving beyond the traditional reliance on simple heuristics for determining modulation patterns, our method combines the strengths of differentiable optimization and data-driven learning. Central to our approach is the novel integration of the optical model of wavefront modulations with a proxy reconstruction network, resulting in a fully differentiable system. This integration allows for the simultaneous optimization of both the proxy reconstruction network and the modulation patterns, which are optimized end-to-end based on a large, simulated training dataset. While imaging through scattering often requires iterative optimization algorithms because data-driven feedforward networks often fail to generalize outside their training domain, this paper demonstrates the remarkable finding that training a hard-to-generalize network can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.715 + ], + "angle": 0, + "content": "become a useful proxy allowing us to learn generalizable modulations. Crucially, this proxy feedforward network allows us to bypass the expensive backpropagation computation through an iterative reconstruction algorithm, which would be necessary, but hopelessly impractical, if we were to naively optimize the modulations: each iteration would require us to finish an entire iterative reconstruction algorithm. As will be shown in our simulated and real data experiments, our learned modulations significantly enhance the reconstruction capability of both our data-driven proxy network and a state-of-the-art iterative optimization-based method [8] for imaging through scattering." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.716, + 0.669, + 0.73 + ], + "angle": 0, + "content": "Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.733, + 0.892, + 0.778 + ], + "angle": 0, + "content": "- We present a novel end-to-end learning framework to optimize acquisition-time wavefront modulations to enhance our abilities to see through scattering." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.779, + 0.892, + 0.838 + ], + "angle": 0, + "content": "- We demonstrate in both simulated and real experiments that our learned modulations substantially improve image reconstruction quality and effectively generalize to unseen targets and scattering media." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.839, + 0.892, + 0.899 + ], + "angle": 0, + "content": "- We show that the learned modulations can be decoupled from the jointly-trained proxy reconstruction network and significantly enhance the reconstruction quality of state-of-the-art unsupervised approaches." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.733, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25277" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.328 + ], + "angle": 0, + "content": "Learning-based Imaging Through Scattering Media. A detailed review of imaging through scattering media can be found in [10]. We highlight a few recent learning-based approaches. Various data-driven approaches can be applied to image through scattering [5, 13, 20, 22, 35, 36]. They generally capture or synthesize large quantities of training data and then learn a mapping from corrupted measurements to reconstructed scenes. However, they tend to struggle with out-of-distribution data. There have also been unsupervised algorithms by fitting network-based parameterizations of the scene to a collection of measurements [8, 15], which can generalize across different distributions of aberrations and target scenes. Both of these methods take advantage of phase diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.473, + 0.527 + ], + "angle": 0, + "content": "End-to-End Learning of Optical Systems. End-to-end learning is a flexible learning-based system design framework that describes optical systems using parameterized differentiable optical models which can then be optimized with training data [31]. End-to-end learning has been applied successfully for extended depth-of-field imaging [14, 21, 29, 31], depth estimation [4, 14, 29, 43], high-dynamic-range imaging [23, 33], seeing through near-lens occlusions [30], and many other applications. To facilitate end-to-end learning, prior works have also incorporated the usage of a proxy network to mimic the response of a specific optical system or black-box algorithm [26, 39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.473, + 0.742 + ], + "angle": 0, + "content": "Phase Diversity Imaging. Phase diversity imaging (PDI) can recover a clear image of an aberrated scene by obtaining images with deliberately introduced known aberrations. PDI does not pose strong assumptions on the target [45] or the aberration [16], and the optics and calibration process required are relatively simple compared to other adaptive optics imaging [11]. While early applications of PDI are often constrained to simple aberrations due to computation complexities [32], the introduction of coded diffraction patterns has enhanced its capacity for reconstructing intricate scenes [3]. Recent PDI developments include improved wavefront sensing [45, 50], phase retrieval [27, 44], adaptive optics [7, 15], and imaging through scattering [6, 8, 46]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.378, + 0.769 + ], + "angle": 0, + "content": "3. Learned Wavefront Modulations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Here we describe our approach for learning wavefront modulations through training a proxy reconstruction network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.264, + 0.833 + ], + "angle": 0, + "content": "3.1. Scattering Problem" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.871 + ], + "angle": 0, + "content": "This paper focuses on the scenario where capturing an image through scattering media can be modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.882, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\ny = h (\\phi) * x + \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "where \\(y\\) is the captured measurement, \\(x\\) is the target scene that we aim to reconstruct, \\(\\epsilon\\) is noise, and \\(h(\\phi)\\) is the unknown, spatially invariant point spread function (PSF) describing the optical scattering, which manifests as unknown phase delays \\(\\phi\\) to the wavefront. Assuming the target object is illuminated with spatially incoherent monochromatic light, the PSF \\(h\\) is related to the phase error \\(\\phi\\) by \\(h(\\phi) = |\\mathcal{F}(m\\circ e^{j(\\phi)})|^2\\), where \\(\\mathcal{F}\\) is the 2D Fourier transform, \\(\\circ\\) is the hadamard product and \\(m\\) is the aperture mask [12]. Our goal is to recover \\(x\\) from \\(y\\) captured through \\(\\phi\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.254, + 0.656, + 0.27 + ], + "angle": 0, + "content": "3.2. Phase Diversity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Phase diversity imaging seeks additional information about the phase error \\(\\phi\\) by capturing a sequence of measurements \\(y_{1}, y_{2}, \\ldots, y_{K}\\), each with additional known phase errors \\(\\gamma_{1}, \\gamma_{2}, \\ldots, \\gamma_{K}\\). As a result of these additional modulations, the imaging forward model becomes \\(y_{i} = h(\\phi + \\gamma_{i}) * x + \\epsilon_{i}\\). In traditional approaches, the phase modulations \\(\\{\\gamma_{i}\\}_{i=1}^{n}\\) were typically either simple defocus sweeps or randomly chosen patterns generated from Zernike polynomials [8, 25], and their reconstruction objective seeks to minimize" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.427, + 0.892, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P D I} (\\hat {x}, \\phi) = \\sum_ {i = 1} ^ {K} \\left| \\left| y _ {i} - h \\left(\\phi + \\gamma_ {i}\\right) * \\hat {x} \\right| \\right| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.51 + ], + "angle": 0, + "content": "over \\(\\hat{x}\\) and \\(\\phi\\), where \\(\\hat{x}\\) is the estimate of scene \\(x\\), assuming \\(\\epsilon_{i}\\) follows an i.i.d. additive Gaussian distribution." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.522, + 0.704, + 0.537 + ], + "angle": 0, + "content": "3.3. Learned Modulations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.637 + ], + "angle": 0, + "content": "While previous works have leveraged differentiable proxies of physical systems to design better algorithms [26, 39], this work leverages a differentiable proxy, \\(\\mathcal{P}\\), of non-differentiable algorithms to design better physical systems—specifically the phase modulation \\(\\Gamma = \\{\\gamma_i\\}_{i=1}^K\\) used to image through scattering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.728 + ], + "angle": 0, + "content": "We sample a training image \\( x_{n} \\) and simulate passing it through an optical aberration drawn from some known distribution while we apply \\( K \\) learnable modulations to the scattered wavefront. This model allows us to simulate a collection of \\( K \\) images \\( Y_{n}^{\\Gamma} = \\{y_{1}^{n}, y_{2}^{n}, \\ldots, y_{K}^{n}\\} \\), which are different scattered observations of \\( x_{n} \\) based on Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.76 + ], + "angle": 0, + "content": "We then optimize both the proxy reconstruction algorithm \\(\\mathcal{P}\\) and the modulation patterns \\(\\Gamma\\) to minimize" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.772, + 0.892, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l e a r n e d}} (\\mathcal {P}, \\Gamma) = \\sum_ {n = 1} ^ {N} \\| \\mathcal {P} \\left(Y _ {n} ^ {\\Gamma}\\right) - x _ {n} \\| ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Note that the ultimate goal of the learning is not to design a set of modulations specific to the network \\(\\mathcal{P}\\). Rather, we use the performance of \\(\\mathcal{P}\\) as a proxy to probe how effective the modulations \\(\\Gamma\\) are for the scattering problem. The differentiability of \\(\\mathcal{P}\\) allows for back-propagation to guide" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25278" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "the optimization of the modulations. As we will demonstrate, while \\(\\mathcal{P}\\) itself may not generalize outside of its training data domain, the learned modulations \\(\\Gamma\\) are compatible with other more generalizable reconstruction algorithms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.288 + ], + "angle": 0, + "content": "Our end-to-end training pipeline is illustrated in Figure 2. For the proxy reconstruction network, we choose a U-Net with skip connections and self-attention modules [24]. To regularize the highly non-convex problem of optimizing modulations, we use an implicit neural representation for the modulations \\(\\Gamma\\). Specifically, the MLP \\(G\\) takes a fixed vector \\(Z\\) with 28 channels, which corresponds to the first 28 Zernike polynomials [18]; it outputs a 16-channel vector corresponding to 16 modulation patterns: \\(\\Gamma_{G} = G(Z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.288, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Incorporating the implicit neural representation for the modulations, our final loss function becomes" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.326, + 0.47, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n a l}} (\\mathcal {P}, G) = \\sum_ {n = 1} ^ {N} \\left\\| \\mathcal {P} \\left(Y _ {n} ^ {\\Gamma_ {G}}\\right) - x _ {n} \\right\\| ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.375, + 0.359, + 0.394 + ], + "angle": 0, + "content": "where \\(Y_{n}^{\\Gamma_{G}} = \\{h(\\phi_{n} + G(Z)_{i})*x_{n}\\}_{i = 1}^{K}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.356, + 0.417 + ], + "angle": 0, + "content": "3.4. Applying Learned Modulations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.499 + ], + "angle": 0, + "content": "While the trained proxy network \\(\\mathcal{P}\\) demonstrates considerable capability in recovering scenes from scattering in our experiments, supervised data-driven methods often fail to generalize to novel or cross-domain scenes, particularly when compared to unsupervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.471, + 0.681 + ], + "angle": 0, + "content": "Therefore, after obtaining the learned modulations \\(\\Gamma_{G}\\) and applying them during real-world acquisition, we send the resulting modulated measurements to an unsupervised iterative optimization-based reconstruction algorithm [8] which does not suffer from the generalization issue of data-driven methods. In effect, our method offers a best-of-both-worlds scenario: the modulations \\(\\Gamma_{G}\\) are effectively learned thanks to the joint supervised training with the proxy network \\(\\mathcal{P}\\), but their enhanced data acquisition quality is transferrable to other reconstruction algorithms, thus allowing us to benefit from the generalization and domain adaptability inherent in unsupervised reconstruction methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.381, + 0.71 + ], + "angle": 0, + "content": "4. Analysis of Learned Modulations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Before we show the performance of the learned modulations with real-world experiments in Section 6, we first analyze the effectiveness of learned modulations in simulation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.279, + 0.789 + ], + "angle": 0, + "content": "4.1. Impact on Frequency" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "When the scattering media scrambles the incoming wavefront, it destroys some of the target scene's frequencies, either by destructively interfering with them or by scattering them outside the imaging sensor. If those frequencies are not captured in the measurement, it becomes difficult to faithfully recover them. By applying multiple learned wavefront modulation patterns, we increase the chance that each" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.093, + 0.835, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.289, + 0.895, + 0.402 + ], + "angle": 0, + "content": "Figure 3. MTF Comparison. We compare the MTF of learned wavefront modulations vs. the MTF of modulations randomly drawn from the same distribution of the scattering media. The X-axis represents spatial frequency, and the Y-axis represents the modulation transfer (the higher the better). Our learned wavefront modulations exhibit a higher MTF compared to unoptimized ones, especially over higher frequency bands, suggesting the learned modulations preserve more high-frequency information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.49 + ], + "angle": 0, + "content": "frequency is better preserved by at least one of the modulations. With the extra information from all modulated measurements, a better image reconstruction performance could be achieved regardless of the reconstruction algorithm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.641 + ], + "angle": 0, + "content": "Thus, we assess the performance of our learned modulations by their combined ability to preserve frequencies. We choose to the commonly used Modulation Transfer Function (MTF) as our metric [1, 19, 34, 38, 40]. MTF measures the contrast produced by an imaging system at different spatial frequencies, defined as \\(\\mathrm{MTF}_i^\\omega (\\phi) = \\left|\\mathcal{F}[h_i(\\phi)]\\right|^{\\omega}\\), where \\(h_i\\) is the PSF for the \\(i\\)-th modulation and the superscript \\(\\omega\\) denotes indexing the MTF at a specific frequency \\(\\omega\\). For the multiple modulations in our problem setup, we define the combined MTF as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.655, + 0.892, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M T F} _ {\\text {c o m b}} ^ {\\omega} \\left(\\left\\{\\gamma_ {i} \\right\\}, \\phi\\right) = \\max _ {i} \\mathrm {M T F} _ {i} ^ {\\omega} \\left(\\phi + \\gamma_ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.892, + 0.718 + ], + "angle": 0, + "content": "where for each frequency \\(\\omega\\), we take the maximum over the MTFs calculated from all modulations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In Figure 3, we plot the combined MTF for both the learned wavefront modulations and the random ones as baseline. The learned ones have a clear advantage at high frequencies (the higher the frequency is, the further away it is to the center of the X-axis), suggesting that the learned modulations are better at preserving finer details of the target scene. Note that during training, we never explicitly encourage the MTF to be higher — the only loss term we use is the mean square error between reconstruction and ground truth image. This implies that our end-to-end training implicitly encourages the modulations to preserve higher-frequency information." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25279" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.186, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.089, + 0.272, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.357, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.089, + 0.444, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.154, + 0.345, + 0.163 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.164, + 0.185, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.164, + 0.272, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.164, + 0.357, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.164, + 0.444, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.229, + 0.371, + 0.237 + ], + "angle": 0, + "content": "Reconstructions with no modulations" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.239, + 0.185, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.239, + 0.272, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.239, + 0.357, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.239, + 0.444, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.304, + 0.386, + 0.313 + ], + "angle": 0, + "content": "Reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.314, + 0.185, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.314, + 0.272, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.314, + 0.357, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.314, + 0.444, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.378, + 0.382, + 0.387 + ], + "angle": 0, + "content": "Reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.388, + 0.185, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.388, + 0.272, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.388, + 0.357, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.388, + 0.444, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.453, + 0.307, + 0.462 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.517 + ], + "angle": 0, + "content": "Figure 4. Proxy Network Reconstruction on Simulated Scattering. Simulated imaging results through different aberrations. Learned modulations lead to a better quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.314, + 0.56 + ], + "angle": 0, + "content": "4.2. Impact on Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We evaluate our jointly learned modulations and proxy network both quantitatively and qualitatively. Both the training and test set are from the MIT Places 365 Dataset [51]. Training details are provided later in Section 5. Here, we use two baselines: (1) simply training a feed-forward reconstruction network without any wavefront modulations, and (2) training a feed-forward reconstruction network with randomly chosen, fixed wavefront modulations as used in prior PDI work. For a fair comparison, our approach and baselines use the same network architecture and are trained on the same data for the same number of iterations. Their only difference is the usage of wavefront modulations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 4, compared to baselines, reconstructions using learned modulations demonstrate finer details and higher consistency with the ground truth. Table 1 shows the quantitative comparison averaged over a test set of 1,000 images, where our approach outperforms the baselines by over \\(3.9\\mathrm{dB}\\). While the results above are all from the jointly trained proxy reconstruction network, we will show in Section 6 that our learned wavefront modulations can decouple from the proxy network and also improve reconstruction quality based on iterative optimization [8]." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.082, + 0.837, + 0.162 + ], + "angle": 0, + "content": "
MetricModulations
NoneRandomLearned
PSNR25.47626.43930.391
SSIM0.76400.79800.9082
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.166, + 0.894, + 0.29 + ], + "angle": 0, + "content": "Table 1. PSNR and SSIM of Simulation Results. We report the test-time performance of our jointly trained modulations and the proxy network. We compare it against two baselines where either we do not use wavefront modulations or use randomly chosen, unoptimized modulations. The metrics are averaged over the reconstruction of 1000 test images, each of which is scattered with a randomly sampled optic aberration that is unseen during training. The results demonstrate the huge performance improvement from the optimization of wavefront modulations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.311, + 0.659, + 0.328 + ], + "angle": 0, + "content": "5. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Software Implementation & Training. For the proxy reconstruction network, we use an attention U-Net [24]; for the MLP that represents the 16 wavefront modulations, we use a two-layer MLP with leaky-ReLU activation. In each training iteration, the optical aberrations are sampled randomly from Zernike basis functions. The standard deviation of the coefficients for each Zernike basis is randomly generated from a uniform distribution from 5 to 6. We train our proposed framework for 2 million iterations on the MIT Places Dataset [51] with the Adam Optimizer and a learning rate of 0.001. Training took 12 hours on an NVIDIA RTX A6000 GPU. For experimental results, we finetune our trained model on the captured bio-tissue data under the same training settings except with a learning rate of 0.0001." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.893, + 0.846 + ], + "angle": 0, + "content": "Hardware Implementation. Our optical configuration is depicted in Figure 6. A continuous-wave laser with a wavelength of \\(532\\mathrm{nm}\\) was collimated and subsequently passed through a laser speckle reducer (Optotune LSR-4C-L). The resulting spatially incoherent light was then polarized and reflected by a digital micromirror device (DMD). The DMD (LightCrafter™ 4500) display is capable of rapidly presenting 8-bit target images. For ease of assembly, a reflecting mirror was incorporated after the DMD. Then, the light is focused by lens L0 (\\(f = 100\\mathrm{mm}\\)) onto the front focal plane of L1, a \\(10\\times\\) microscopic objective. Our optical aberration—resin painted on glass—was placed at the same location. L1 and a tube lens L2 (\\(f = 200\\mathrm{mm}\\)) constitute a \\(4f\\) system, and thereby the SLM is manipulating the Fourier plane. We used a HOLOEYE LETO-3 phase-only SLM to modulate the wavefront. Finally, the light reflected by the SLM was imaged on the camera (\\(1384\\times 1036\\) pixel Grasshopper3) via L3 (\\(f = 500\\mathrm{mm}\\)). Our capture speed is limited to 30 fps by the camera frame rate." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experimental Dataset. Our experimental data are from human pathological slides, imaged with optical microscopy (Zeiss Microscope Axio Imager.A2) and a \\(10 \\times\\) objective." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.223, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.089, + 0.356, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.09, + 0.488, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.137, + 0.364, + 0.145 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.147, + 0.223, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.147, + 0.356, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.148, + 0.488, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.195, + 0.402, + 0.204 + ], + "angle": 0, + "content": "Proxy reconstructions with no modulations" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.206, + 0.223, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.207, + 0.356, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.207, + 0.488, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.254, + 0.416, + 0.264 + ], + "angle": 0, + "content": "Proxy reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.266, + 0.224, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.266, + 0.356, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.266, + 0.488, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.313, + 0.415, + 0.322 + ], + "angle": 0, + "content": "Proxy reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.324, + 0.225, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.324, + 0.356, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.324, + 0.488, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.371, + 0.328, + 0.38 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.089, + 0.603, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.09, + 0.693, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.137, + 0.764, + 0.145 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.148, + 0.603, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.147, + 0.694, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.781, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.09, + 0.871, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.195, + 0.803, + 0.204 + ], + "angle": 0, + "content": "Proxy reconstructions with no modulations" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.206, + 0.603, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.207, + 0.693, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.207, + 0.781, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.206, + 0.871, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.254, + 0.816, + 0.264 + ], + "angle": 0, + "content": "Proxy reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.266, + 0.603, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.267, + 0.693, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.267, + 0.781, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.267, + 0.871, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.313, + 0.816, + 0.323 + ], + "angle": 0, + "content": "Proxy reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.324, + 0.603, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.324, + 0.693, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.324, + 0.781, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.324, + 0.871, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.371, + 0.729, + 0.38 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.399, + 0.894, + 0.455 + ], + "angle": 0, + "content": "Figure 5. Proxy Network Reconstruction on Physical Scattering. Experimental results of imaging objects through scattering media by using our proxy reconstruction network. The left columns are in-distribution adipose tissue slides (zoomed-in region labeled with red boxes); the right columns are out-of-distribution targets. Learned modulations yield superior imaging quality for both in-distribution and out-of-distribution scenes, with the former out-performing the latter." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.479, + 0.457, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Figure 6. Optical Setup. A spatially incoherent light source illuminates the DMD and passes through thick scattering. The DMD displays the target images for training and testing. The SLM is placed onto the Fourier plane and projects learned patterns. The wavefront is modulated and subsequently imaged onto a camera." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Each mage is \\(256 \\times 256\\) pixels. Our dataset includes 1,000 sectional images of adipose tissues. 40 of these images were utilized for testing and the rest were used for training. To evaluate the generalizability across different human tissues, we also imaged sections of human stomach and glandular epithelium, trachea, and finger. Additionally, we tested with targets outside the domain of the training data, such as numbers and letters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.706, + 0.497 + ], + "angle": 0, + "content": "6. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.657 + ], + "angle": 0, + "content": "We employed the optical system depicted in Figure 6. Our learned modulations significantly enhance the imaging capabilities through scattering media. We tested on datasets from both similar and distinctly different distributions from our experimental training data. Our learned modulations consistently yield superior visual reconstruction results. Furthermore, we evaluated these modulations on both static and dynamic scenes. We show that our learned modulations significantly enhance reconstruction in both cases, which suggests a broad applicability of our approach." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.774, + 0.682 + ], + "angle": 0, + "content": "6.1. Proxy Reconstruction Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proxy network, fine-tuned on 960 adipose tissue images through random scattering, was evaluated on a test set from the same tissue type, as shown in the left of Figure 5. Reconstructions driven by measurements of 16 learned SLM patterns preserved the original contrast and high-frequency details of the dyed sections. The close-ups show remarkable detail accuracy. By comparison, reconstructions trained on measurements of random or no modulation struggled, particularly in lower-intensity areas, yielding blurry outputs. Beyond achieving good reconstruction fidelity on objects similar to our training dataset, we extended our evaluation to vastly different data distributions, like numbers and letters. The results, as shown on the right of Figure 5, clearly indicate that the learned modulations yield bet" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25281" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.088, + 0.244, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.089, + 0.322, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.089, + 0.401, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.089, + 0.479, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.135, + 0.388, + 0.142 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.145, + 0.244, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.145, + 0.322, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.145, + 0.401, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.145, + 0.479, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.19, + 0.425, + 0.198 + ], + "angle": 0, + "content": "Reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.2, + 0.244, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.2, + 0.322, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.2, + 0.401, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.2, + 0.479, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.246, + 0.425, + 0.255 + ], + "angle": 0, + "content": "Reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.257, + 0.244, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.257, + 0.322, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.257, + 0.401, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.257, + 0.479, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.302, + 0.357, + 0.311 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.089, + 0.573, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.089, + 0.651, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.089, + 0.729, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.089, + 0.808, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.135, + 0.717, + 0.142 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.145, + 0.573, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.145, + 0.651, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.145, + 0.729, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.145, + 0.808, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.19, + 0.753, + 0.198 + ], + "angle": 0, + "content": "Reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.2, + 0.573, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.2, + 0.651, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.2, + 0.729, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.2, + 0.808, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.246, + 0.753, + 0.254 + ], + "angle": 0, + "content": "Reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.257, + 0.573, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.257, + 0.651, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.257, + 0.729, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.257, + 0.808, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.302, + 0.684, + 0.312 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Figure 7. Unsupervised Reconstruction on Physical Scattering with Static Scenes. Experimental results of imaging different static targets through scattering media using learned wavefront modulations with unsupervised iterative optimization [8]. Learned modulations enhance reconstruction quality." + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.392, + 0.258, + 0.401 + ], + "angle": 0, + "content": "\\(\\mathrm{t} = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.401, + 0.28, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.392, + 0.35, + 0.401 + ], + "angle": 0, + "content": "\\(\\mathrm{t} = 24\\)" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.401, + 0.374, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.392, + 0.443, + 0.4 + ], + "angle": 0, + "content": "t=48" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.401, + 0.47, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.449, + 0.402, + 0.457 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.46, + 0.28, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.508, + 0.441, + 0.516 + ], + "angle": 0, + "content": "Reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.518, + 0.374, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.518, + 0.469, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.566, + 0.441, + 0.574 + ], + "angle": 0, + "content": "Reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.577, + 0.28, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.577, + 0.374, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.577, + 0.469, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.626, + 0.368, + 0.634 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.393, + 0.55, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.393, + 0.646, + 0.401 + ], + "angle": 0, + "content": "t=24" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.401, + 0.671, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.393, + 0.736, + 0.401 + ], + "angle": 0, + "content": "t=48" + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.401, + 0.766, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.449, + 0.702, + 0.457 + ], + "angle": 0, + "content": "Uncorrected measurements" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.46, + 0.577, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.507, + 0.74, + 0.516 + ], + "angle": 0, + "content": "Reconstructions with random modulations" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.518, + 0.577, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.518, + 0.671, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.518, + 0.766, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.566, + 0.74, + 0.574 + ], + "angle": 0, + "content": "Reconstructions with learned modulations" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.577, + 0.577, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.577, + 0.671, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.577, + 0.765, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.626, + 0.667, + 0.635 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.654, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Figure 8. Unsupervised Reconstruction on Physical Scattering with Dynamic Scenes. Experimental results of imaging an opening camera aperture through severe aberrations (left) and imaging non-regional motions (number 6 rotating clockwise and number 5 translating to the right) through severe aberrations. We use the unsupervised iterative reconstruction approach [8]. Learned modulations highly improve the quality of the reconstructed dynamic scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.841 + ], + "angle": 0, + "content": "ter reconstruction quality, particularly in the USAF target and stars. PSNR and SSIM are computed in Tables 2 and 3. Among the table cells, \"None\" means that the measurements for each target only comprised the unmodulated image (no modulation patterns are applied); \"Random\" means that the modulation patterns are randomly sampled from the same distribution as the aberrations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.806, + 0.752 + ], + "angle": 0, + "content": "6.2. Unsupervised Iterative Approaches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While the proxy network \\(\\mathcal{P}\\) exhibits considerable effectiveness in real-world experiments, its performance remains unsatisfactory for out-of-distribution data, which is expected. Nevertheless, our learned modulations are designed to further augment generalizable, untrained reconstruction algorithms based on iterative optimizations. To validate this strategy, we deploy our learned modulations to a recently developed unsupervised iterative approach for imaging through scattering, NeuWS [8], in our benchmark study." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25282" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.083, + 0.472, + 0.198 + ], + "angle": 0, + "content": "
MethodData typeModulation
NoneRandomLearned
ProxyTissue16.5317.5719.06
ProxyOut-of-dist.9.349.9010.71
Iterative [8]StaticN/A11.2614.61
Iterative [8]DynamicN/A8.9012.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.201, + 0.47, + 0.355 + ], + "angle": 0, + "content": "Table 2. PSNR of Experimental Results. For our jointly trained feed-forward proxy reconstruction network (\"proxy\"), we tested on 40 tissue samples and 8 out-of-distribution scenes, all of which are static. For the iterative method [8], we tested on the same 8 out-of-distribution static scenes. We also tested [8] on 2 dynamic scenes, each with 48 frames. The iterative method relies on multiple wavefront modulations and therefore cannot recover objects with a single measurement, hence the \"N/A\". Compared against random modulations or no modulations, our learned modulations lead to better reconstruction performance for both the proxy network and the unsupervised iterative approach." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.358, + 0.47, + 0.472 + ], + "angle": 0, + "content": "
MethodData typeModulations
NoneRandomLearned
ProxyTissue0.440.480.58
ProxyOut-of-dist.0.290.300.32
Iterative [8]StaticN/A0.210.38
Iterative [8]DynamicN/A0.230.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.505 + ], + "angle": 0, + "content": "Table 3. SSIM of Experimental Results. Same as Table 2 but showing SSIM. Our learned modulations performs better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Figure 7 shows the reconstruction performance of NeuWS [8] on 8 static objects, each using only 16 measurements. The superiority of using learned modulations is visually evident in the results. We excluded experiments with no modulations here, due to the algorithm's inability to recover objects using a single unmodulated measurement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.77 + ], + "angle": 0, + "content": "For dynamic scenes, each scene involves 48 frames captured by cycling through the 16 SLM patterns. Figure 8 demonstrates the reconstruction results on dynamic scenes: the result on the left illustrates a camera aperture gradually enlarging; the result on the right shows two different motions to solve the non-regional dynamic reconstruction (the number 6 on the left rotates clockwise at \\(0.5^{\\circ}\\) per frame, while the number 5 translates from the center to the right). Quantitative metrics on the dynamic results are included in Table 2 and 3. Full video results are in the project webpage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.312, + 0.801 + ], + "angle": 0, + "content": "7. Discussion & Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Role of the Proxy Reconstruction Network. The proxy network not only aims to reconstruct the images; its performance also indicates how good the wavefront modulations are for the restoration task. Therefore, joint differentiable optimization of both the proxy network and the modulations works in our favor, and the experimental results indeed in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "dicate that this is an effective way of learning useful wavefront modulations. Without this proxy network, we would have to somehow connect the modulation update with the performance of an iterative optimization method, which is impractical due to time and memory constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Regularization Perspective. To recover missing signal frequencies due to scattering, most existing methods rely on prior information of the scene, either hand-crafted or data-driven. This work rethinks the strategy of incorporating prior knowledge: can we impose a prior over wavefront modulations? The goal here is to let the resulting imaging system capture more information about the target scene through scattering. Our results validate the effectiveness of learning a prior over the wavefront modulation domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.895, + 0.56 + ], + "angle": 0, + "content": "Applicability to Other Vision Tasks. The end-to-end learning approach can potentially benefit other vision tasks through scattering media, using a task-specific proxy network. For instance, if we were to do object detection or semantic segmentation through scattering media, we can replace the proxy reconstruction U-Net with a network tailored for these two tasks, and perform end-to-end training in the same way as before. Note that we validated our proposed method on a monochromatic imaging system, which is prevalent in medical and scientific imaging. Extending it to broadband imaging tasks, e.g., outdoor navigation, requires the optimization of wavelength-dependent modulation patterns, which we leave for future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.895, + 0.741 + ], + "angle": 0, + "content": "Conclusion. We proposed a robust end-to-end learning framework integrating the optical scattering model with a proxy image reconstruction network. The learned wavefront modulations can both work with the proxy reconstruction network and augment a generalizable, optimization-based algorithm. We conducted intensive experiments to validate our approach. This work shows the synergy of advanced wavefront modulation techniques with cutting-edge machine learning methods, representing a significant leap forward in computer vision and optical imaging." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.655, + 0.771 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by AFOSR Young Investigator Program award no. FA9550-22-1-0208, ONR award no. N00014-23-1-2752, seed grant from SAAB, Inc., a seed grant from the UMD Brain and Behavior Institute, NSF Expeditions award no. CCF-1730574, ONR Scattering media award no. N00014-23-1-2714, and NSF PaThs up award no. EEC-1648451. We thank Sachin Shah for helpful discussions and Kevin F. Kelly for providing the DMD." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25283" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Jesse K Adams, Dong Yan, Jimin Wu, Vivek Boominathan, Sibo Gao, Alex V Rodriguez, Soonyoung Kim, Jennifer Carns, Rebecca Richards-Kortum, Caleb Kemere, et al. In vivo lensless microscopy via a phase mask generating diffraction patterns with high-contrast contours. Nature Biomedical Engineering, 6(5):617-628, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Codruta Orniana Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1798-1805, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase Retrieval Via Wirtinger Flow: Theory and Algorithms. IEEE Transactions on Information Theory, 61(4):1985-2007, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Dongdong Chen, Mike Davies, Matthias J Ehrhardt, Carola-Bibiane Schonlieb, Ferdia Sherry, and Julián Tachella. Imaging with equivariant deep learning: From unrolled network design to fully unsupervised learning. IEEE Signal Processing Magazine, 40(1):134-147, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.491 + ], + "angle": 0, + "content": "[6] Wei-Yu Chen, Matthew O'Toole, Aswin C Sankaranarayanan, and Anat Levin. Enhancing Speckle Statistics for Imaging Inside Scattering Media. Optica, 9(12):1408-1416, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.562 + ], + "angle": 0, + "content": "[7] Santiago Echeverri-Chacón, René Restrepo, Carlos Cuartas-Vélez, and Néstor Uribe-Patarroyo. Vortex-enhanced coherent-illumination phase diversity for phase retrieval in coherent imaging systems. Optics letters, 41(8):1817-1820, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[8] Brandon Y Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K Sharma, Ashok Veeraraghavan, and Christopher A Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[9] Xueyang Fu, Jia-Bin Huang, Delu Zeng, Yue Huang, Xing-hao Ding, and John William Paisley. Removing rain from single images via a deep detail network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1715-1723, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[10] Sylvain Gigan, Ori Katz, Hilton B de Aguiar, Esben Ravn Andresen, Alexandre Aubry, Jacopo Bertolotti, Emmanuel Bossy, Dorian Bouchet, Joshua Brake, Sophie Brasselet, et al. Roadmap on Wavefront Shaping and Deep Imaging in Complex Media. Journal of Physics: Photonics, 4(4): 042501, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Robert A Gonsalves. Phase retrieval and diversity in adaptive optics. Optical Engineering, 21(5):829-832, 1982. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[12] Joseph W. Goodman. Introduction to Fourier Optics. Roberts & Co., Englewood, Colo, 3rd ed. edition, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Xiaowen Hu, Jian Zhao, Jose Enrique Antonio-Lopez, Stefan Gausmann, Rodrigo Amezcua Correa, and Axel Schulzgen. Adaptive inverse mapping: a model-free" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "semi-supervised learning approach towards robust imaging through dynamic scattering media. Optics Express, 31(9): 14343-14357, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[15] Iksung Kang, Qinrong Zhang, Stella X Yu, and Na Ji. Coordinate-based neural representations for computational adaptive optics in widefield microscopy. arXiv preprint arXiv:2307.03812, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[16] Richard L Kendrick, Daniel S Acton, and AL Duncan. Phase-diversity wave-front sensor for imaging systems. Applied Optics, 33(27):6533-6546, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[17] Peter Kner. Phase diversity for three-dimensional imaging. JOSA A, 30(10):1980-1987, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[19] Juhyun Lee, Jinsoo Jeong, Jaebum Cho, Dongheon Yoo, Byounghyo Lee, and Byoungho Lee. Deep neural network for multi-depth hologram generation and its training strategy. Optics Express, 28(18):27137-27154, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[20] Yunzhe Li, Yujia Xue, and Lei Tian. Deep speckle correlation: a deep learning approach toward scalable imaging through scattering media. Optica, 5(10):1181-1190, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[21] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[22] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-Inverse Correlography: Towards Real-Time High-Resolution Non-Line-of-Sight Imaging. Optica, 7(1):63–71, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[23] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[24] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, M. J. Lee, Mattias P. Heinrich, Kazunari Misawa, Kensaku Mori, Steven G. McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. ArXiv, abs/1804.03999, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[25] Richard G Paxman, Timothy J Schulz, and James R Fienup. Joint estimation of object and aberrations by using phase diversity. JOSA A, 9(7):1072-1085, 1992. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[26] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural Holography With Camera-in-the-Loop Training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Nikolaj Reiser, Min Guo, Hari Shroff, and Patrick J La Riviere. Phase diverse phase retrieval for microscopy: Com" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25284" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "parison of gaussian and poisson approaches. arXiv preprint arXiv:2308.00734, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[28] Wenqi Ren, Sibo Liu, Hua Zhang, Jin shan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks. In European Conference on Computer Vision, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[29] Sachin Shah, Sakshum Kulshrestha, and Christopher A Metzler. Tidy-psfs: Computational imaging with time-averaged dynamic point-spread-functions. ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[30] Zheng Shi, Yuval Bahat, Seung-Hwan Baek, Qiang Fu, Hadi Amata, Xiao Li, Praneeth Chakravarthula, Wolfgang Heidrich, and Felix Heide. Seeing through obstructions with diffractive cloaking. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[31] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[32] Carlas S Smith, Arnold J den Dekker, Raluca Andrei, Rufus Fraanje, and Michel Verhaegen. Fast phase diversity wavefront sensing using object independent metrics. In Adaptive Optics Systems III, pages 2250-2258. SPIE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[33] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386–1396, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[34] Qilin Sun, Congli Wang, Fu Qiang, Dun Xiong, and Heidrich Wolfgang. End-to-end complex lens design with differentiable ray tracing. ACM Trans. Graph, 40(4):1-13, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[35] Yu Sun, Zhihao Xia, and Ulugbek S Kamilov. Efficient and Accurate Inversion of Multiple Scattering with Deep Learning. Optics Express, 26(11):14678-14688, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[36] Waleed Tahir, Hao Wang, and Lei Tian. Adaptive 3d scattering with a dynamic synthesis network. Light: Science & Applications, 11(1):42, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[37] Brian J Thelen, Richard G Paxman, David A Carcarara, and John H Seldin. Maximum a posteriori estimation of fixed aberrations, dynamic aberrations, and the object from phase-diverse speckle data. JOsA A, 16(5):1016-1025, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[38] Lei Tian and Laura Waller. Quantitative differential phase contrast imaging in an led array microscope. Optics express, 23(9):11394-11403, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[39] Ethan Tseng, Felix Yu, Yuting Yang, Fahim Mannan, Karl St. Arnaud, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Hyperparameter optimization in black-box image processing using differentiable proxies. ACM Transactions on Graphics (TOG), 38(4), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[40] Hakan Urey, Ned Nestorovic, Baldwin S Ng, and Abraham A Gross. Optics designs and system mtf for laser scanning displays. In Helmet-and Head-Mounted Displays IV, pages 238-248. SPIE, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Curtis R Vogel, Tony F Chan, and Robert J Plemmons. Fast algorithms for phase-diversity-based blind deconvolution. In Adaptive Optical System Technologies, pages 994-1005. SPIE, 1998. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[42] Abbie T Watnik and Dennis F Gardner. Wavefront Sensing in Deep Turbulence. Optics and Photonics News, 29(10): 38-45, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[43] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[44] Meng Xiang, An Pan, Jinpeng Liu, Teli Xi, Xin Guo, Fei Liu, and Xiaopeng Shao. Phase diversity-based fourier psychography for varying aberration correction. Frontiers in Physics, 10:129, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[45] Qi Xin, Guohao Ju, Chunyue Zhang, and Shuyan Xu. Object-independent image-based wavefront sensing approach using phase diversity images and deep learning. Optics express, 27(18):26102-26119, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[46] Tomer Yeminy and Ori Katz. Guidestar-Free Image-Guided Wavefront Shaping. Science Advances, 7(21), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[47] Seokchan Yoon, Moonseok Kim, Mooseok Jang, Youngwoon Choi, Wonjun Choi, Sungsam Kang, and Wonshik Choi. Deep Optical Imaging Within Complex Scattering Media. Nature Reviews Physics, 2(3):141-158, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[48] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[49] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[50] Peiguang Zhang, Chengliang Yang, Zihao Xu, Zhaoliang Cao, Quanquan Mu, and Li Xuan. High-accuracy wavefront sensing by phase diversity technique with bisymmetric defocuses diversity phase. Scientific Reports, 7(1):15361, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[51] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40:1452-1464, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25285" + } + ] +] \ No newline at end of file diff --git a/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_origin.pdf b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75435d78c9a5449b9039cbeb5a964b857cda579c --- /dev/null +++ b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/7f09b732-d1d1-4da5-806d-7ae1886c2551_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a0b427b39a5b291faf291722c2c2a7f8526d1c7f7d16cafcd05796dcbc30599 +size 6200925 diff --git a/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/full.md b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b94911fe1f0395da340b238fe7ea2eb5ba090256 --- /dev/null +++ b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/full.md @@ -0,0 +1,504 @@ +# WaveMo: Learning Wavefront Modulations to See Through Scattering + +Mingyang Xie $^{1*}$ Haiyun Guo $^{2*}$ Brandon Y. Feng $^{3}$ Lingbo Jin $^{2}$ Ashok Veeraraghavan $^{2}$ Christopher A. Metzler $^{1}$ + +1University of Maryland + +$^{2}$ Rice University + +3Massachusetts Institute of Technology + +# Abstract + +Imaging through scattering media is a fundamental and pervasive challenge in fields ranging from medical diagnostics to astronomy. A promising strategy to overcome this challenge is wavefront modulation, which induces measurement diversity during image acquisition. Despite its importance, designing optimal wavefront modulations to image through scattering remains under-explored. This paper introduces a novel learning-based framework to address the gap. Our approach jointly optimizes wavefront modulations and a computationally lightweight feedforward "proxy" reconstruction network. This network is trained to recover scenes obscured by scattering, using measurements that are modified by these modulations. The learned modulations produced by our framework generalize effectively to unseen scattering scenarios and exhibit remarkable versatility. During deployment, the learned modulations can be decoupled from the proxy network to augment other more computationally expensive restoration algorithms. Through extensive experiments, we demonstrate our approach significantly advances the state of the art in imaging through scattering media. Our project webpage is at https://wavemo2024.github.io/. + +# 1. Introduction + +Imaging through scattering media presents a significant challenge across diverse scenarios, ranging from navigating with fog [2, 28, 30], rain [9, 48, 49], or murky water to recovering intricate structures through human skin and tissue [10, 42, 47]. The core of this challenge are the irregular phase delays light experiences as it scatters. These phase delays blur and warp any images captured through the scattering media. Effectively addressing this issue is crucial for unlocking new computer vision capabilities in fields such as medical imaging and astronomy. + +From a frequency domain perspective, the effect of a + +![](images/6bf9bf293682b61dd5a272e1c0d2960a5b835780691dc2254bbf65f4eeeb6ea1.jpg) + +![](images/304bc80251eb51cc6c05a9c89f2fd7979305a2a4b3813d9ee7cfb8079200cb39.jpg) +Figure 1. Learned Wavefront Modulations. Top: During acquisition, we can modulate the wavefront of scattered light by using a spatial light modulator (SLM), and capture a set of image measurements useful for scene reconstruction. Bottom left: We propose learning modulations that enhance our ability to recover the scattered scene. Bottom right: Our learned modulations drastically improve the reconstruction quality of a state-of-the-art method [8] that previously applies randomly chosen modulations. + +![](images/200f2d47f9cc1f78dccb64b5e02846836fa8cf306d9b3d76bcf15b88264558c9.jpg) + +scene passing through scattering is the destruction or filtering of frequency content. Recovering these lost frequencies is very difficult using purely computational methods, which are effectively forced to speculate on the missing frequencies based on those that have been preserved. A question naturally arises: Can we modify the measurement process to prevent the loss of information during acquisition? + +A promising solution is to actively modulate the scattered wavefront during acquisition, in hopes that applying multiple modulations increases the chance that a frequency component is preserved. This approach is related to phase diversity imaging [11, 15, 17, 25, 27, 37, 41, 45], which typically captures multiple scattered images of a static scene with different modulations applied using deformable mirrors or spatial light modulators (SLM). However, phase diversity imaging traditionally was limited to imaging simple scenes with mild optical aberrations, and the phase modulations were often selected based on simple heuristics. The recent development of neural wavefront shaping (NeuWS) [8] + +![](images/6c47af18287b17a17f98b7d80316709f7af528a3b7cbaa6de49cf4471c5abf8c.jpg) +Figure 2. Overview of the Proposed End-to-end Learning Framework. During training, we jointly optimize the implicit neural representation for the wavefront modulations and the proxy reconstruction network in an end-to-end fashion. During inference, one can apply the learned wavefront modulations to any reconstruction algorithms for imaging through scattering, either a trained feed-forward reconstruction network or an unsupervised iterative optimization algorithm [8]. The former approach is far faster and performs better when test data and training data fall into similar distributions, while the latter generalizes better to unseen distributions of target scenes. + +brings a significant breakthrough by leveraging recent innovations in machine learning and enables imaging high-resolution and dynamic scenes through severe aberrations. Still, the wavefront modulations applied by NeuWS during acquisition are chosen randomly without explicit consideration of their impact on frequency preservation. + +In this paper, we take a significant step towards a more systematic and principled approach in designing effective modulations that better preserve frequencies against scattering during image acquisition. Moving beyond the traditional reliance on simple heuristics for determining modulation patterns, our method combines the strengths of differentiable optimization and data-driven learning. Central to our approach is the novel integration of the optical model of wavefront modulations with a proxy reconstruction network, resulting in a fully differentiable system. This integration allows for the simultaneous optimization of both the proxy reconstruction network and the modulation patterns, which are optimized end-to-end based on a large, simulated training dataset. While imaging through scattering often requires iterative optimization algorithms because data-driven feedforward networks often fail to generalize outside their training domain, this paper demonstrates the remarkable finding that training a hard-to-generalize network can + +become a useful proxy allowing us to learn generalizable modulations. Crucially, this proxy feedforward network allows us to bypass the expensive backpropagation computation through an iterative reconstruction algorithm, which would be necessary, but hopelessly impractical, if we were to naively optimize the modulations: each iteration would require us to finish an entire iterative reconstruction algorithm. As will be shown in our simulated and real data experiments, our learned modulations significantly enhance the reconstruction capability of both our data-driven proxy network and a state-of-the-art iterative optimization-based method [8] for imaging through scattering. + +Our contributions are: + +- We present a novel end-to-end learning framework to optimize acquisition-time wavefront modulations to enhance our abilities to see through scattering. +- We demonstrate in both simulated and real experiments that our learned modulations substantially improve image reconstruction quality and effectively generalize to unseen targets and scattering media. +- We show that the learned modulations can be decoupled from the jointly-trained proxy reconstruction network and significantly enhance the reconstruction quality of state-of-the-art unsupervised approaches. + +# 2. Related Work + +Learning-based Imaging Through Scattering Media. A detailed review of imaging through scattering media can be found in [10]. We highlight a few recent learning-based approaches. Various data-driven approaches can be applied to image through scattering [5, 13, 20, 22, 35, 36]. They generally capture or synthesize large quantities of training data and then learn a mapping from corrupted measurements to reconstructed scenes. However, they tend to struggle with out-of-distribution data. There have also been unsupervised algorithms by fitting network-based parameterizations of the scene to a collection of measurements [8, 15], which can generalize across different distributions of aberrations and target scenes. Both of these methods take advantage of phase diversity. + +End-to-End Learning of Optical Systems. End-to-end learning is a flexible learning-based system design framework that describes optical systems using parameterized differentiable optical models which can then be optimized with training data [31]. End-to-end learning has been applied successfully for extended depth-of-field imaging [14, 21, 29, 31], depth estimation [4, 14, 29, 43], high-dynamic-range imaging [23, 33], seeing through near-lens occlusions [30], and many other applications. To facilitate end-to-end learning, prior works have also incorporated the usage of a proxy network to mimic the response of a specific optical system or black-box algorithm [26, 39]. + +Phase Diversity Imaging. Phase diversity imaging (PDI) can recover a clear image of an aberrated scene by obtaining images with deliberately introduced known aberrations. PDI does not pose strong assumptions on the target [45] or the aberration [16], and the optics and calibration process required are relatively simple compared to other adaptive optics imaging [11]. While early applications of PDI are often constrained to simple aberrations due to computation complexities [32], the introduction of coded diffraction patterns has enhanced its capacity for reconstructing intricate scenes [3]. Recent PDI developments include improved wavefront sensing [45, 50], phase retrieval [27, 44], adaptive optics [7, 15], and imaging through scattering [6, 8, 46]. + +# 3. Learned Wavefront Modulations + +Here we describe our approach for learning wavefront modulations through training a proxy reconstruction network. + +# 3.1. Scattering Problem + +This paper focuses on the scenario where capturing an image through scattering media can be modeled as + +$$ +y = h (\phi) * x + \epsilon , \tag {1} +$$ + +where $y$ is the captured measurement, $x$ is the target scene that we aim to reconstruct, $\epsilon$ is noise, and $h(\phi)$ is the unknown, spatially invariant point spread function (PSF) describing the optical scattering, which manifests as unknown phase delays $\phi$ to the wavefront. Assuming the target object is illuminated with spatially incoherent monochromatic light, the PSF $h$ is related to the phase error $\phi$ by $h(\phi) = |\mathcal{F}(m\circ e^{j(\phi)})|^2$ , where $\mathcal{F}$ is the 2D Fourier transform, $\circ$ is the hadamard product and $m$ is the aperture mask [12]. Our goal is to recover $x$ from $y$ captured through $\phi$ . + +# 3.2. Phase Diversity + +Phase diversity imaging seeks additional information about the phase error $\phi$ by capturing a sequence of measurements $y_{1}, y_{2}, \ldots, y_{K}$ , each with additional known phase errors $\gamma_{1}, \gamma_{2}, \ldots, \gamma_{K}$ . As a result of these additional modulations, the imaging forward model becomes $y_{i} = h(\phi + \gamma_{i}) * x + \epsilon_{i}$ . In traditional approaches, the phase modulations $\{\gamma_{i}\}_{i=1}^{n}$ were typically either simple defocus sweeps or randomly chosen patterns generated from Zernike polynomials [8, 25], and their reconstruction objective seeks to minimize + +$$ +\mathcal {L} _ {P D I} (\hat {x}, \phi) = \sum_ {i = 1} ^ {K} \left| \left| y _ {i} - h \left(\phi + \gamma_ {i}\right) * \hat {x} \right| \right| ^ {2}, \tag {2} +$$ + +over $\hat{x}$ and $\phi$ , where $\hat{x}$ is the estimate of scene $x$ , assuming $\epsilon_{i}$ follows an i.i.d. additive Gaussian distribution. + +# 3.3. Learned Modulations + +While previous works have leveraged differentiable proxies of physical systems to design better algorithms [26, 39], this work leverages a differentiable proxy, $\mathcal{P}$ , of non-differentiable algorithms to design better physical systems—specifically the phase modulation $\Gamma = \{\gamma_i\}_{i=1}^K$ used to image through scattering. + +We sample a training image $x_{n}$ and simulate passing it through an optical aberration drawn from some known distribution while we apply $K$ learnable modulations to the scattered wavefront. This model allows us to simulate a collection of $K$ images $Y_{n}^{\Gamma} = \{y_{1}^{n}, y_{2}^{n}, \ldots, y_{K}^{n}\}$ , which are different scattered observations of $x_{n}$ based on Eq. (1). + +We then optimize both the proxy reconstruction algorithm $\mathcal{P}$ and the modulation patterns $\Gamma$ to minimize + +$$ +\mathcal {L} _ {\text {l e a r n e d}} (\mathcal {P}, \Gamma) = \sum_ {n = 1} ^ {N} \| \mathcal {P} \left(Y _ {n} ^ {\Gamma}\right) - x _ {n} \| ^ {2}. \tag {3} +$$ + +Note that the ultimate goal of the learning is not to design a set of modulations specific to the network $\mathcal{P}$ . Rather, we use the performance of $\mathcal{P}$ as a proxy to probe how effective the modulations $\Gamma$ are for the scattering problem. The differentiability of $\mathcal{P}$ allows for back-propagation to guide + +the optimization of the modulations. As we will demonstrate, while $\mathcal{P}$ itself may not generalize outside of its training data domain, the learned modulations $\Gamma$ are compatible with other more generalizable reconstruction algorithms. + +Our end-to-end training pipeline is illustrated in Figure 2. For the proxy reconstruction network, we choose a U-Net with skip connections and self-attention modules [24]. To regularize the highly non-convex problem of optimizing modulations, we use an implicit neural representation for the modulations $\Gamma$ . Specifically, the MLP $G$ takes a fixed vector $Z$ with 28 channels, which corresponds to the first 28 Zernike polynomials [18]; it outputs a 16-channel vector corresponding to 16 modulation patterns: $\Gamma_{G} = G(Z)$ . + +Incorporating the implicit neural representation for the modulations, our final loss function becomes + +$$ +\mathcal {L} _ {\text {f i n a l}} (\mathcal {P}, G) = \sum_ {n = 1} ^ {N} \left\| \mathcal {P} \left(Y _ {n} ^ {\Gamma_ {G}}\right) - x _ {n} \right\| ^ {2}, \tag {4} +$$ + +where $Y_{n}^{\Gamma_{G}} = \{h(\phi_{n} + G(Z)_{i})*x_{n}\}_{i = 1}^{K}$ . + +# 3.4. Applying Learned Modulations + +While the trained proxy network $\mathcal{P}$ demonstrates considerable capability in recovering scenes from scattering in our experiments, supervised data-driven methods often fail to generalize to novel or cross-domain scenes, particularly when compared to unsupervised approaches. + +Therefore, after obtaining the learned modulations $\Gamma_{G}$ and applying them during real-world acquisition, we send the resulting modulated measurements to an unsupervised iterative optimization-based reconstruction algorithm [8] which does not suffer from the generalization issue of data-driven methods. In effect, our method offers a best-of-both-worlds scenario: the modulations $\Gamma_{G}$ are effectively learned thanks to the joint supervised training with the proxy network $\mathcal{P}$ , but their enhanced data acquisition quality is transferrable to other reconstruction algorithms, thus allowing us to benefit from the generalization and domain adaptability inherent in unsupervised reconstruction methods. + +# 4. Analysis of Learned Modulations + +Before we show the performance of the learned modulations with real-world experiments in Section 6, we first analyze the effectiveness of learned modulations in simulation. + +# 4.1. Impact on Frequency + +When the scattering media scrambles the incoming wavefront, it destroys some of the target scene's frequencies, either by destructively interfering with them or by scattering them outside the imaging sensor. If those frequencies are not captured in the measurement, it becomes difficult to faithfully recover them. By applying multiple learned wavefront modulation patterns, we increase the chance that each + +![](images/4261602f59e9f5979a37357f0aefcc7dd2e380c08a7fc2324d923bb286037c05.jpg) +Figure 3. MTF Comparison. We compare the MTF of learned wavefront modulations vs. the MTF of modulations randomly drawn from the same distribution of the scattering media. The X-axis represents spatial frequency, and the Y-axis represents the modulation transfer (the higher the better). Our learned wavefront modulations exhibit a higher MTF compared to unoptimized ones, especially over higher frequency bands, suggesting the learned modulations preserve more high-frequency information. + +frequency is better preserved by at least one of the modulations. With the extra information from all modulated measurements, a better image reconstruction performance could be achieved regardless of the reconstruction algorithm. + +Thus, we assess the performance of our learned modulations by their combined ability to preserve frequencies. We choose to the commonly used Modulation Transfer Function (MTF) as our metric [1, 19, 34, 38, 40]. MTF measures the contrast produced by an imaging system at different spatial frequencies, defined as $\mathrm{MTF}_i^\omega (\phi) = \left|\mathcal{F}[h_i(\phi)]\right|^{\omega}$ , where $h_i$ is the PSF for the $i$ -th modulation and the superscript $\omega$ denotes indexing the MTF at a specific frequency $\omega$ . For the multiple modulations in our problem setup, we define the combined MTF as: + +$$ +\mathrm {M T F} _ {\text {c o m b}} ^ {\omega} \left(\left\{\gamma_ {i} \right\}, \phi\right) = \max _ {i} \mathrm {M T F} _ {i} ^ {\omega} \left(\phi + \gamma_ {i}\right), \tag {5} +$$ + +where for each frequency $\omega$ , we take the maximum over the MTFs calculated from all modulations. + +In Figure 3, we plot the combined MTF for both the learned wavefront modulations and the random ones as baseline. The learned ones have a clear advantage at high frequencies (the higher the frequency is, the further away it is to the center of the X-axis), suggesting that the learned modulations are better at preserving finer details of the target scene. Note that during training, we never explicitly encourage the MTF to be higher — the only loss term we use is the mean square error between reconstruction and ground truth image. This implies that our end-to-end training implicitly encourages the modulations to preserve higher-frequency information. + +![](images/ae4e87b6a0096d2b85d16fb5f7643b297f82d5925d9807e5cfd4de11e4f582c5.jpg) + +![](images/28dfa595812ef7095784887990f417e343e725e24eed3b9e237c7cd6941ecaff.jpg) +Uncorrected measurements + +![](images/7f6a105f7e6971295d2476e263a677718acd7cf9031db511b9dae349e1bbc542.jpg) + +![](images/4bf19fab200664f039d44da3ace70f90fa1a9f7fe36032282e2d38064744f047.jpg) + +![](images/b7eae1b38ff40e8b4237810bc2def67b1e982ebbda40c71c3ac48c8861f4645c.jpg) +Reconstructions with no modulations + +![](images/a90d055f6c28e5e53d0060d5ead0525c35b2d673fe8fe179ed24111ba04613bb.jpg) + +![](images/07435ec1a0c0b0fe25f8ce719dac3fad15115e8e6be4308c61f5ba211e20d3c5.jpg) + +![](images/8ea33a876323ea203b88fe116e5bbbb111c9fdade108c5aac0ee6981746113eb.jpg) + +![](images/a54ef6f114dc57e21a90260ebe699df4b313c61301812e6e9929f48109855043.jpg) +Reconstructions with random modulations + +![](images/c83fcedd195d827d19f67041e331862c349fc697c7cfe79b54662748a72ab367.jpg) + +![](images/e03cce180316ff1548af11d49c2e48c55efd038d4b400208f4cdcc2ac9da85a4.jpg) + +![](images/73db0ee5fd56def69b19dd9935954cd437312fad5a09aef4a292d652528b8a20.jpg) + +![](images/2a07707edb5bdbb42878c0cb04158088039bd01c9f31fe2a96a1aa905756b456.jpg) +Reconstructions with learned modulations + +![](images/765ea69ced7093f2a9666955b35415c4d88ce303cc154f0e19f4312994c90522.jpg) + +![](images/b0274790ee82953a8882b85180b0f99a4e63673a74fe3c300a1936ffd676d4a5.jpg) + +![](images/59bfb4d26c9e4ab586d05dec843a3713ff5fe37d2906257ef48a085768018ecb.jpg) + +![](images/ff9015397632b08ab254d48afd8da86cb26a708886cf99b10532fcffc21cd37a.jpg) +Figure 4. Proxy Network Reconstruction on Simulated Scattering. Simulated imaging results through different aberrations. Learned modulations lead to a better quality. + +![](images/dbbb70a0505a226bbf409309e93ea1b064454a72f7503830370a1c03f8f799d6.jpg) +Ground truth + +![](images/94dcc419b901c3191c3367c89c84f6ed6d094baae07e0f5a1093866d54a97b46.jpg) + +![](images/f6c3338a0a8b51c74419e11295df554fd0cb74f5ada46d27e466bdcb9817fa0a.jpg) + +# 4.2. Impact on Reconstruction + +We evaluate our jointly learned modulations and proxy network both quantitatively and qualitatively. Both the training and test set are from the MIT Places 365 Dataset [51]. Training details are provided later in Section 5. Here, we use two baselines: (1) simply training a feed-forward reconstruction network without any wavefront modulations, and (2) training a feed-forward reconstruction network with randomly chosen, fixed wavefront modulations as used in prior PDI work. For a fair comparison, our approach and baselines use the same network architecture and are trained on the same data for the same number of iterations. Their only difference is the usage of wavefront modulations. + +As shown in Figure 4, compared to baselines, reconstructions using learned modulations demonstrate finer details and higher consistency with the ground truth. Table 1 shows the quantitative comparison averaged over a test set of 1,000 images, where our approach outperforms the baselines by over $3.9\mathrm{dB}$ . While the results above are all from the jointly trained proxy reconstruction network, we will show in Section 6 that our learned wavefront modulations can decouple from the proxy network and also improve reconstruction quality based on iterative optimization [8]. + +
MetricModulations
NoneRandomLearned
PSNR25.47626.43930.391
SSIM0.76400.79800.9082
+ +Table 1. PSNR and SSIM of Simulation Results. We report the test-time performance of our jointly trained modulations and the proxy network. We compare it against two baselines where either we do not use wavefront modulations or use randomly chosen, unoptimized modulations. The metrics are averaged over the reconstruction of 1000 test images, each of which is scattered with a randomly sampled optic aberration that is unseen during training. The results demonstrate the huge performance improvement from the optimization of wavefront modulations. + +# 5. Implementation + +Software Implementation & Training. For the proxy reconstruction network, we use an attention U-Net [24]; for the MLP that represents the 16 wavefront modulations, we use a two-layer MLP with leaky-ReLU activation. In each training iteration, the optical aberrations are sampled randomly from Zernike basis functions. The standard deviation of the coefficients for each Zernike basis is randomly generated from a uniform distribution from 5 to 6. We train our proposed framework for 2 million iterations on the MIT Places Dataset [51] with the Adam Optimizer and a learning rate of 0.001. Training took 12 hours on an NVIDIA RTX A6000 GPU. For experimental results, we finetune our trained model on the captured bio-tissue data under the same training settings except with a learning rate of 0.0001. + +Hardware Implementation. Our optical configuration is depicted in Figure 6. A continuous-wave laser with a wavelength of $532\mathrm{nm}$ was collimated and subsequently passed through a laser speckle reducer (Optotune LSR-4C-L). The resulting spatially incoherent light was then polarized and reflected by a digital micromirror device (DMD). The DMD (LightCrafter™ 4500) display is capable of rapidly presenting 8-bit target images. For ease of assembly, a reflecting mirror was incorporated after the DMD. Then, the light is focused by lens L0 ( $f = 100\mathrm{mm}$ ) onto the front focal plane of L1, a $10\times$ microscopic objective. Our optical aberration—resin painted on glass—was placed at the same location. L1 and a tube lens L2 ( $f = 200\mathrm{mm}$ ) constitute a $4f$ system, and thereby the SLM is manipulating the Fourier plane. We used a HOLOEYE LETO-3 phase-only SLM to modulate the wavefront. Finally, the light reflected by the SLM was imaged on the camera ( $1384\times 1036$ pixel Grasshopper3) via L3 ( $f = 500\mathrm{mm}$ ). Our capture speed is limited to 30 fps by the camera frame rate. + +Experimental Dataset. Our experimental data are from human pathological slides, imaged with optical microscopy (Zeiss Microscope Axio Imager.A2) and a $10 \times$ objective. + +![](images/efb0def738077b443752410085a3f2ff235ac1f2fa9b6b0359563b79b30ff3e8.jpg) + +![](images/6a084187dec390a84202560764ea945d0e773c5655ff109dda528d9ff465100d.jpg) +Uncorrected measurements + +![](images/0901dc6b4518e7728ccd61e687c56212ac5ed65bb477e5e033ce5c80f4bde7e4.jpg) + +![](images/99fb54fef4243cba83d8ea5a83c5b3f99cdde1d1822ecbb0ff83f4c15e3ae0ea.jpg) + +![](images/017253ac0be3850a7dc8436278138eda3f0876b69c0866a24795d687c9d94ef5.jpg) + +![](images/678c56530a8d64c3cc9866631ac7033802f18935bcdb51a7394b9622f6419919.jpg) +Proxy reconstructions with no modulations + +![](images/b09bdd01b6600f5fff9170102afac143eb9a046fa5b03da78c1c0b32c2d3b979.jpg) + +![](images/77829edce75a451cca4a3d9b90a25e70a7031629ab24cbe804c4c9f4b0da8346.jpg) + +![](images/fd70babec234fc43193b89a0fe4201e9d76549eaa7058233d24868d85c795333.jpg) +Proxy reconstructions with random modulations + +![](images/2eb7d0536d619d6957595f3b10efbee998f9a9f79113f83f0e4e97c07fcf55a5.jpg) +Proxy reconstructions with learned modulations + +![](images/67e363f3c957ac5054b98fd49222ec208a0a19f6d7397a28e157d8f154b5f017.jpg) + +![](images/cdff0b9ff06fe6e244506cc6fcd7062080e16f58099d718e97a05f3e5781b72a.jpg) + +![](images/5102bc387ccfd1feb69e400d35b427ba3633987512a0d5bd5382d12f890bd2ed.jpg) + +![](images/d1ee28985945547b92e629f6c73ac5dd412b1a2e7b8adc4acc0f463b5efdc71d.jpg) +Ground truth + +![](images/ff614c8f3bb02c66dc9b220227f512a8a8f62d0121dcea219c803ed67eb2cb2b.jpg) + +![](images/ae6a1c1e8d7890f605985e7c8c7ae79c2b082c7ce3963205b759a04754b0a18c.jpg) + +![](images/4f1b36c7ceb4aa164ecfa481216f4e957b1d0baf68e8f6f0bde496a3fedd1f5a.jpg) +Uncorrected measurements + +![](images/5af787bd1e651aeb6942d512bb902ad990971de211f7174729b04e8d90c075cd.jpg) + +![](images/490b90633c61a91c1aa522829af3f8ef72f23d35977c614db54f2ceb0c3b6260.jpg) +Proxy reconstructions with no modulations + +![](images/d2380e34cbc0e12fce18d88eccb74ceae3d84017738f4adc66c116ec269c4c34.jpg) + +![](images/abafee8c49e3d8efb787059a909e2457441a33f022cadcd4fdc50ae1bb2d75a8.jpg) + +![](images/48d4de392fb768b2887b55b34cc2bca452b886134584198785e83729bfd733c2.jpg) + +![](images/1bd1eedbebac7de60e8e56739ea831b282de7bd7b755b93b460b6d7d979e60d4.jpg) +Proxy reconstructions with random modulations + +![](images/99aedd39cd4cf623c81d3843ad4244c93bee540e4f8618c57e0cfa79332c339d.jpg) + +![](images/466bb3b2d837ca7cb59b9a29f5b7dc1cf0f4a76dffef0c20da7ebc132cb1ba4f.jpg) + +![](images/aec3bc163077b361c53ee3e625baa4dcbd35eea14cd50a050b2a5d6d5ba34a82.jpg) +Proxy reconstructions with learned modulations + +![](images/985a9ad8bd7e4fa9006b817352bde2ec0bc92f345c3eb163b7c86ae8d2b2c922.jpg) + +![](images/5c6dd77a2df95c671cfb79dbdc1e764551ccd140cbc7e70bfe9117e014a106d3.jpg) + +![](images/b9df6e123028dd705da029dcde65ab68eb86c9fb2830df4c786dfc9c69fdcaca.jpg) + +![](images/986f856fa960b339196195b5633521a6a9b7e16ab8916038979e390b0e2ef008.jpg) + +![](images/75902cf97aa51d39031adcf85e1add4f175b673b2628a5121d0ff79e5aaee632.jpg) +Ground truth + +![](images/eadc92a5c3c55b9cc5083e46762f2f05cfde3b4fe7ab931da9acd7a7b7b47767.jpg) + +![](images/d6f11c72e1e045c163da6f62efd2c824c3a9413d9fc549a6f495f7fc4595a875.jpg) + +![](images/368f466ffc48bb6881c0e2ac3a055e520f6ac42c36ad68692da3ffdf6ceda76b.jpg) +Figure 5. Proxy Network Reconstruction on Physical Scattering. Experimental results of imaging objects through scattering media by using our proxy reconstruction network. The left columns are in-distribution adipose tissue slides (zoomed-in region labeled with red boxes); the right columns are out-of-distribution targets. Learned modulations yield superior imaging quality for both in-distribution and out-of-distribution scenes, with the former out-performing the latter. +Figure 6. Optical Setup. A spatially incoherent light source illuminates the DMD and passes through thick scattering. The DMD displays the target images for training and testing. The SLM is placed onto the Fourier plane and projects learned patterns. The wavefront is modulated and subsequently imaged onto a camera. + +Each mage is $256 \times 256$ pixels. Our dataset includes 1,000 sectional images of adipose tissues. 40 of these images were utilized for testing and the rest were used for training. To evaluate the generalizability across different human tissues, we also imaged sections of human stomach and glandular epithelium, trachea, and finger. Additionally, we tested with targets outside the domain of the training data, such as numbers and letters. + +# 6. Experimental Results + +We employed the optical system depicted in Figure 6. Our learned modulations significantly enhance the imaging capabilities through scattering media. We tested on datasets from both similar and distinctly different distributions from our experimental training data. Our learned modulations consistently yield superior visual reconstruction results. Furthermore, we evaluated these modulations on both static and dynamic scenes. We show that our learned modulations significantly enhance reconstruction in both cases, which suggests a broad applicability of our approach. + +# 6.1. Proxy Reconstruction Network + +The proxy network, fine-tuned on 960 adipose tissue images through random scattering, was evaluated on a test set from the same tissue type, as shown in the left of Figure 5. Reconstructions driven by measurements of 16 learned SLM patterns preserved the original contrast and high-frequency details of the dyed sections. The close-ups show remarkable detail accuracy. By comparison, reconstructions trained on measurements of random or no modulation struggled, particularly in lower-intensity areas, yielding blurry outputs. Beyond achieving good reconstruction fidelity on objects similar to our training dataset, we extended our evaluation to vastly different data distributions, like numbers and letters. The results, as shown on the right of Figure 5, clearly indicate that the learned modulations yield bet + +![](images/afb40c5c8aa46924fa4adef9c37a095b00dc2eb993d3696ef89ffd0e33708bd3.jpg) + +![](images/dbdbcdccf6a040f07ba99f2b32deb2b255721c6ee5f5e14f2963428003b90221.jpg) +Uncorrected measurements + +![](images/e5ec909d020f9af775c4f060389c9b9e0f23a4af85d22c75de3b9362fee2152d.jpg) + +![](images/677f83b23404db4457714b2069b4195407d322f26be3632688bb62706f54b9f3.jpg) + +![](images/88f52109252054f3ea82412818902bb7390773bc866df84d9aa2f1cd2f985628.jpg) +Reconstructions with random modulations + +![](images/685d1e10ecb46f8061a6ef648a991a0200732dec64bb1b70f71b078b9f6c0db2.jpg) + +![](images/ebc2a86cbdde8bd93fa091212e4a008b06d39b315b90e9861ea501a82d940532.jpg) + +![](images/59d4e481cc39a2cacb930b594a70258db07a37de09efdf5c4eb4698d4c98c7af.jpg) + +![](images/81dbc391823d09fe51174ee9302c1c9eb645042fb1edc57c321a10962628141e.jpg) +Reconstructions with learned modulations + +![](images/d8959e7e1197f1ab2ebd39e666329379b6b5cda0927e7842ceda5e69fd4219f4.jpg) + +![](images/b100a7ce931ab83f45f9714728e9d4754c6782a594fc99bcc2f3960e85d2aeeb.jpg) + +![](images/d1ae4d57b0d99422323689dcb71fc78d08030e32974a254d1fc5bbe41e8a2e40.jpg) + +![](images/459f212daa3216bf97e44e9a3c02aeb086265fce8a51cdab05e6cd5a6c2b4b1d.jpg) + +![](images/f09e314b5549379163781f9c03487c4e0707557bb5f5bd853dddb818eab382cb.jpg) +Ground truth + +![](images/83994e5c0ec333b4073d6c541913d769097b7c880663ca1d0f1ff831d74e993c.jpg) + +![](images/cf1a1f53e06f9600a11991bd0bd3ab3b5a51ec7ecc8ff6618552c608adfcb847.jpg) + +![](images/d06de3fcda707bab0908c51b786bdf60a10fa5c6a5a407760af2dc6f3b30da85.jpg) + +![](images/80047a784f6a8ad7ac5a5f6c131068c08faf85484025ea771690641e873f8f56.jpg) +Uncorrected measurements + +![](images/8b8259b8e9594a449e2d2c67bbbd71b3ac1cc86e7cc2776684b20d51846d36d7.jpg) + +![](images/805ea4f44c7eaec65598ea262307186eb035c3b6c7efbfc9cbc65043e639a9d0.jpg) + +![](images/900a0a3fd09e77f5b8550753df53f335732515fc8811d6e67fe298e048bd8275.jpg) +Reconstructions with random modulations + +![](images/d85aee4487be29256b2d8c5985b0ac6404a2197ad76bf305688cd8413398f728.jpg) + +![](images/b22cb5527af65eb55f7636ec03d05a3fedbd0df023a6bd346ade4e8b2e046598.jpg) + +![](images/6ad583a53f7dfcfd2b902248894ac1943f1faaf12ec1d06e6eb4a98bf1e8f65c.jpg) + +![](images/b44daed79ff0f7c13cdca1bf51e374f783eb258fdd5a5ba3a095b71ec07adaa9.jpg) +Reconstructions with learned modulations + +![](images/cfdd2253d7bd23affd6347131956487e034518993de674471dce2567a5d9a2c3.jpg) + +![](images/3fb49de8fd2ce4a659af6c1c0a33e874fb426ea12f409cc4ba0df3faba889d6b.jpg) + +![](images/7c2b839b53c3bbd24ad825baa0f2f05cc0847ff18b79a0323ab4207bfcf787a1.jpg) + +![](images/a71cc18bce7e9b3c2734761990901eff097d4c9ef411b6c4cfdcd4ac0e4f3f84.jpg) + +![](images/97082a8b6a172ee907c63785d1cafd76ec07b2f1ffcb4603b224238d20df8ca6.jpg) +Ground truth + +![](images/78198196e6a652de8ff787cc51d1ec06bc55d84f0417c93aa4f56562527722c8.jpg) + +![](images/3a8cdfc74378419aa2f61616121353414daa0ca0dd633a20c22c30926d1c11ee.jpg) + +![](images/2347cf0cc00e9110c516fe7114238ce0c5fe57be53bf01b965d8c2893f0295d3.jpg) +$\mathrm{t} = 1$ +Uncorrected measurements + +![](images/e23e5819f8af33754f6da9a94d589d3642447fa5de361a8b5e0fd199cfda7f7b.jpg) +$\mathrm{t} = 24$ + +![](images/2a1d637fc435272aa3988bc650a541b33ed575127b9071ff385bdfb2a0a4cac1.jpg) +t=48 + +![](images/898cd7280dbf47e87506ebd82496d00e0d7811947692d7028155b94109aac1d8.jpg) +Reconstructions with random modulations + +![](images/eca66562396695e35a5c1a745418ef9a4e34f7270138d648164c703501429961.jpg) +Reconstructions with learned modulations + +![](images/32b7b0085eccf72f4f45d1ac0b35b5ea88be5485dee7e9ba1305eaaa176870ef.jpg) + +![](images/40e91686a32669be00b49a90291cca80527730f79f722a1ef070578b062713f5.jpg) +Figure 8. Unsupervised Reconstruction on Physical Scattering with Dynamic Scenes. Experimental results of imaging an opening camera aperture through severe aberrations (left) and imaging non-regional motions (number 6 rotating clockwise and number 5 translating to the right) through severe aberrations. We use the unsupervised iterative reconstruction approach [8]. Learned modulations highly improve the quality of the reconstructed dynamic scene. + +![](images/768429f557bd6682e1a10496dffee32d2d4deebac11f2f57b8f0dba019f65dc9.jpg) +Ground truth + +![](images/1d42a20717d879276fc69fc5bec855e2cf7672ddb16e6800911e5a0fcdfd8845.jpg) + +![](images/995e62438ef49fbb381dd2f429efe7f390dbd3fa60892ac2ce2d89e90947b56f.jpg) +Figure 7. Unsupervised Reconstruction on Physical Scattering with Static Scenes. Experimental results of imaging different static targets through scattering media using learned wavefront modulations with unsupervised iterative optimization [8]. Learned modulations enhance reconstruction quality. + +![](images/e44a2e8ea23e66318ec015d8da5a12d9d41378030507c2f520b6e1e5f29147c8.jpg) +t=24 +Uncorrected measurements + +![](images/b18b1fe9d0af0117ef0b20f6fc02f2cf7844c81a1586e940f4ad9a0e366edded.jpg) +t=48 + +![](images/62ac0fb1a45541ab57325828e9e7528b25312aa67de7f2bb9a8a4049fd98be4a.jpg) +Reconstructions with random modulations + +![](images/ee77cc7b512e9e3fda6b3faa34c010b6c9f264082e45e2fa5b98de6af4e2fb85.jpg) +Reconstructions with learned modulations + +![](images/841e13f6c5cbc8eeda4eb41e9a457686b6dcfca959c301b8df6423552c59a0b1.jpg) + +![](images/55fc6a1e498b0883cf57f60195fedce4231db92d95a33e6f2402acabd54f3456.jpg) + +![](images/b9d760e1e9631c5a53c581134b33481a139e43ab85b6a5a8cedd2045c291e9f8.jpg) + +![](images/892d8c8b2ea90da5b1bc70f7f45b8ded2b9c9d6a9f42ef77a33d0ff1db361b8c.jpg) +Ground truth + +![](images/e3cda9da38125a9b7d00e45e8105753ee730200d3502775d8be4358d676c69cd.jpg) + +ter reconstruction quality, particularly in the USAF target and stars. PSNR and SSIM are computed in Tables 2 and 3. Among the table cells, "None" means that the measurements for each target only comprised the unmodulated image (no modulation patterns are applied); "Random" means that the modulation patterns are randomly sampled from the same distribution as the aberrations. + +# 6.2. Unsupervised Iterative Approaches + +While the proxy network $\mathcal{P}$ exhibits considerable effectiveness in real-world experiments, its performance remains unsatisfactory for out-of-distribution data, which is expected. Nevertheless, our learned modulations are designed to further augment generalizable, untrained reconstruction algorithms based on iterative optimizations. To validate this strategy, we deploy our learned modulations to a recently developed unsupervised iterative approach for imaging through scattering, NeuWS [8], in our benchmark study. + +
MethodData typeModulation
NoneRandomLearned
ProxyTissue16.5317.5719.06
ProxyOut-of-dist.9.349.9010.71
Iterative [8]StaticN/A11.2614.61
Iterative [8]DynamicN/A8.9012.89
+ +Table 2. PSNR of Experimental Results. For our jointly trained feed-forward proxy reconstruction network ("proxy"), we tested on 40 tissue samples and 8 out-of-distribution scenes, all of which are static. For the iterative method [8], we tested on the same 8 out-of-distribution static scenes. We also tested [8] on 2 dynamic scenes, each with 48 frames. The iterative method relies on multiple wavefront modulations and therefore cannot recover objects with a single measurement, hence the "N/A". Compared against random modulations or no modulations, our learned modulations lead to better reconstruction performance for both the proxy network and the unsupervised iterative approach. + +
MethodData typeModulations
NoneRandomLearned
ProxyTissue0.440.480.58
ProxyOut-of-dist.0.290.300.32
Iterative [8]StaticN/A0.210.38
Iterative [8]DynamicN/A0.230.33
+ +Table 3. SSIM of Experimental Results. Same as Table 2 but showing SSIM. Our learned modulations performs better. + +Figure 7 shows the reconstruction performance of NeuWS [8] on 8 static objects, each using only 16 measurements. The superiority of using learned modulations is visually evident in the results. We excluded experiments with no modulations here, due to the algorithm's inability to recover objects using a single unmodulated measurement. + +For dynamic scenes, each scene involves 48 frames captured by cycling through the 16 SLM patterns. Figure 8 demonstrates the reconstruction results on dynamic scenes: the result on the left illustrates a camera aperture gradually enlarging; the result on the right shows two different motions to solve the non-regional dynamic reconstruction (the number 6 on the left rotates clockwise at $0.5^{\circ}$ per frame, while the number 5 translates from the center to the right). Quantitative metrics on the dynamic results are included in Table 2 and 3. Full video results are in the project webpage. + +# 7. Discussion & Conclusion + +Role of the Proxy Reconstruction Network. The proxy network not only aims to reconstruct the images; its performance also indicates how good the wavefront modulations are for the restoration task. Therefore, joint differentiable optimization of both the proxy network and the modulations works in our favor, and the experimental results indeed in + +dicate that this is an effective way of learning useful wavefront modulations. Without this proxy network, we would have to somehow connect the modulation update with the performance of an iterative optimization method, which is impractical due to time and memory constraints. + +Regularization Perspective. To recover missing signal frequencies due to scattering, most existing methods rely on prior information of the scene, either hand-crafted or data-driven. This work rethinks the strategy of incorporating prior knowledge: can we impose a prior over wavefront modulations? The goal here is to let the resulting imaging system capture more information about the target scene through scattering. Our results validate the effectiveness of learning a prior over the wavefront modulation domain. + +Applicability to Other Vision Tasks. The end-to-end learning approach can potentially benefit other vision tasks through scattering media, using a task-specific proxy network. For instance, if we were to do object detection or semantic segmentation through scattering media, we can replace the proxy reconstruction U-Net with a network tailored for these two tasks, and perform end-to-end training in the same way as before. Note that we validated our proposed method on a monochromatic imaging system, which is prevalent in medical and scientific imaging. Extending it to broadband imaging tasks, e.g., outdoor navigation, requires the optimization of wavelength-dependent modulation patterns, which we leave for future work. + +Conclusion. We proposed a robust end-to-end learning framework integrating the optical scattering model with a proxy image reconstruction network. The learned wavefront modulations can both work with the proxy reconstruction network and augment a generalizable, optimization-based algorithm. We conducted intensive experiments to validate our approach. This work shows the synergy of advanced wavefront modulation techniques with cutting-edge machine learning methods, representing a significant leap forward in computer vision and optical imaging. + +# Acknowledgements + +This work was supported in part by AFOSR Young Investigator Program award no. FA9550-22-1-0208, ONR award no. N00014-23-1-2752, seed grant from SAAB, Inc., a seed grant from the UMD Brain and Behavior Institute, NSF Expeditions award no. CCF-1730574, ONR Scattering media award no. N00014-23-1-2714, and NSF PaThs up award no. EEC-1648451. We thank Sachin Shah for helpful discussions and Kevin F. Kelly for providing the DMD. + +# References + +[1] Jesse K Adams, Dong Yan, Jimin Wu, Vivek Boominathan, Sibo Gao, Alex V Rodriguez, Soonyoung Kim, Jennifer Carns, Rebecca Richards-Kortum, Caleb Kemere, et al. In vivo lensless microscopy via a phase mask generating diffraction patterns with high-contrast contours. Nature Biomedical Engineering, 6(5):617-628, 2022. 4 +[2] Codruta Orniana Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1798-1805, 2020. 1 +[3] Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase Retrieval Via Wirtinger Flow: Theory and Algorithms. IEEE Transactions on Information Theory, 61(4):1985-2007, 2015. 3 +[4] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 3 +[5] Dongdong Chen, Mike Davies, Matthias J Ehrhardt, Carola-Bibiane Schonlieb, Ferdia Sherry, and Julián Tachella. Imaging with equivariant deep learning: From unrolled network design to fully unsupervised learning. IEEE Signal Processing Magazine, 40(1):134-147, 2023. 3 +[6] Wei-Yu Chen, Matthew O'Toole, Aswin C Sankaranarayanan, and Anat Levin. Enhancing Speckle Statistics for Imaging Inside Scattering Media. Optica, 9(12):1408-1416, 2022. 3 +[7] Santiago Echeverri-Chacón, René Restrepo, Carlos Cuartas-Vélez, and Néstor Uribe-Patarroyo. Vortex-enhanced coherent-illumination phase diversity for phase retrieval in coherent imaging systems. Optics letters, 41(8):1817-1820, 2016. 3 +[8] Brandon Y Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K Sharma, Ashok Veeraraghavan, and Christopher A Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023. 1, 2, 3, 4, 5, 7, 8 +[9] Xueyang Fu, Jia-Bin Huang, Delu Zeng, Yue Huang, Xing-hao Ding, and John William Paisley. Removing rain from single images via a deep detail network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1715-1723, 2017. 1 +[10] Sylvain Gigan, Ori Katz, Hilton B de Aguiar, Esben Ravn Andresen, Alexandre Aubry, Jacopo Bertolotti, Emmanuel Bossy, Dorian Bouchet, Joshua Brake, Sophie Brasselet, et al. Roadmap on Wavefront Shaping and Deep Imaging in Complex Media. Journal of Physics: Photonics, 4(4): 042501, 2022. 1, 3 +[11] Robert A Gonsalves. Phase retrieval and diversity in adaptive optics. Optical Engineering, 21(5):829-832, 1982. 1, 3 +[12] Joseph W. Goodman. Introduction to Fourier Optics. Roberts & Co., Englewood, Colo, 3rd ed. edition, 2005. 3 +[13] Xiaowen Hu, Jian Zhao, Jose Enrique Antonio-Lopez, Stefan Gausmann, Rodrigo Amezcua Correa, and Axel Schulzgen. Adaptive inverse mapping: a model-free + +semi-supervised learning approach towards robust imaging through dynamic scattering media. Optics Express, 31(9): 14343-14357, 2023. 3 +[14] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 3 +[15] Iksung Kang, Qinrong Zhang, Stella X Yu, and Na Ji. Coordinate-based neural representations for computational adaptive optics in widefield microscopy. arXiv preprint arXiv:2307.03812, 2023. 1, 3 +[16] Richard L Kendrick, Daniel S Acton, and AL Duncan. Phase-diversity wave-front sensor for imaging systems. Applied Optics, 33(27):6533-6546, 1994. 3 +[17] Peter Kner. Phase diversity for three-dimensional imaging. JOSA A, 30(10):1980-1987, 2013. 1 +[18] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 4 +[19] Juhyun Lee, Jinsoo Jeong, Jaebum Cho, Dongheon Yoo, Byounghyo Lee, and Byoungho Lee. Deep neural network for multi-depth hologram generation and its training strategy. Optics Express, 28(18):27137-27154, 2020. 4 +[20] Yunzhe Li, Yujia Xue, and Lei Tian. Deep speckle correlation: a deep learning approach toward scalable imaging through scattering media. Optica, 5(10):1181-1190, 2018. 3 +[21] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, 2022. 3 +[22] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-Inverse Correlography: Towards Real-Time High-Resolution Non-Line-of-Sight Imaging. Optica, 7(1):63–71, 2020. 3 +[23] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 3 +[24] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, M. J. Lee, Mattias P. Heinrich, Kazunari Misawa, Kensaku Mori, Steven G. McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. ArXiv, abs/1804.03999, 2018. 4, 5 +[25] Richard G Paxman, Timothy J Schulz, and James R Fienup. Joint estimation of object and aberrations by using phase diversity. JOSA A, 9(7):1072-1085, 1992. 1, 3 +[26] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural Holography With Camera-in-the-Loop Training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 3 +[27] Nikolaj Reiser, Min Guo, Hari Shroff, and Patrick J La Riviere. Phase diverse phase retrieval for microscopy: Com + +parison of gaussian and poisson approaches. arXiv preprint arXiv:2308.00734, 2023. 1, 3 +[28] Wenqi Ren, Sibo Liu, Hua Zhang, Jin shan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks. In European Conference on Computer Vision, 2016. 1 +[29] Sachin Shah, Sakshum Kulshrestha, and Christopher A Metzler. Tidy-psfs: Computational imaging with time-averaged dynamic point-spread-functions. ICCV, 2023. 3 +[30] Zheng Shi, Yuval Bahat, Seung-Hwan Baek, Qiang Fu, Hadi Amata, Xiao Li, Praneeth Chakravarthula, Wolfgang Heidrich, and Felix Heide. Seeing through obstructions with diffractive cloaking. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 1, 3 +[31] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 3 +[32] Carlas S Smith, Arnold J den Dekker, Raluca Andrei, Rufus Fraanje, and Michel Verhaegen. Fast phase diversity wavefront sensing using object independent metrics. In Adaptive Optics Systems III, pages 2250-2258. SPIE, 2012. 3 +[33] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386–1396, 2020. 3 +[34] Qilin Sun, Congli Wang, Fu Qiang, Dun Xiong, and Heidrich Wolfgang. End-to-end complex lens design with differentiable ray tracing. ACM Trans. Graph, 40(4):1-13, 2021. 4 +[35] Yu Sun, Zhihao Xia, and Ulugbek S Kamilov. Efficient and Accurate Inversion of Multiple Scattering with Deep Learning. Optics Express, 26(11):14678-14688, 2018. 3 +[36] Waleed Tahir, Hao Wang, and Lei Tian. Adaptive 3d scattering with a dynamic synthesis network. Light: Science & Applications, 11(1):42, 2022. 3 +[37] Brian J Thelen, Richard G Paxman, David A Carcarara, and John H Seldin. Maximum a posteriori estimation of fixed aberrations, dynamic aberrations, and the object from phase-diverse speckle data. JOsA A, 16(5):1016-1025, 1999. 1 +[38] Lei Tian and Laura Waller. Quantitative differential phase contrast imaging in an led array microscope. Optics express, 23(9):11394-11403, 2015. 4 +[39] Ethan Tseng, Felix Yu, Yuting Yang, Fahim Mannan, Karl St. Arnaud, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Hyperparameter optimization in black-box image processing using differentiable proxies. ACM Transactions on Graphics (TOG), 38(4), 2019. 3 +[40] Hakan Urey, Ned Nestorovic, Baldwin S Ng, and Abraham A Gross. Optics designs and system mtf for laser scanning displays. In Helmet-and Head-Mounted Displays IV, pages 238-248. SPIE, 1999. 4 +[41] Curtis R Vogel, Tony F Chan, and Robert J Plemmons. Fast algorithms for phase-diversity-based blind deconvolution. In Adaptive Optical System Technologies, pages 994-1005. SPIE, 1998. 1 + +[42] Abbie T Watnik and Dennis F Gardner. Wavefront Sensing in Deep Turbulence. Optics and Photonics News, 29(10): 38-45, 2018. 1 +[43] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 3 +[44] Meng Xiang, An Pan, Jinpeng Liu, Teli Xi, Xin Guo, Fei Liu, and Xiaopeng Shao. Phase diversity-based fourier psychography for varying aberration correction. Frontiers in Physics, 10:129, 2022. 3 +[45] Qi Xin, Guohao Ju, Chunyue Zhang, and Shuyan Xu. Object-independent image-based wavefront sensing approach using phase diversity images and deep learning. Optics express, 27(18):26102-26119, 2019. 1, 3 +[46] Tomer Yeminy and Ori Katz. Guidestar-Free Image-Guided Wavefront Shaping. Science Advances, 7(21), 2021. 3 +[47] Seokchan Yoon, Moonseok Kim, Mooseok Jang, Youngwoon Choi, Wonjun Choi, Sungsam Kang, and Wonshik Choi. Deep Optical Imaging Within Complex Scattering Media. Nature Reviews Physics, 2(3):141-158, 2020. 1 +[48] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022. 1 +[49] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 1 +[50] Peiguang Zhang, Chengliang Yang, Zihao Xu, Zhaoliang Cao, Quanquan Mu, and Li Xuan. High-accuracy wavefront sensing by phase diversity technique with bisymmetric defocuses diversity phase. Scientific Reports, 7(1):15361, 2017. 3 +[51] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40:1452-1464, 2018. 5 \ No newline at end of file diff --git a/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/images.zip b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb235eb4dea0130ea41fc087363912440539c298 --- /dev/null +++ b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:628d130235fda05734a215015dfd811d6d2f10d7dfc204110ade6ba3b4eaa013 +size 600410 diff --git a/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/layout.json b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4ad23cd5f4c36bd9ba960e9fb6424aafce9e994f --- /dev/null +++ b/2024/WaveMo_ Learning Wavefront Modulations to See Through Scattering/layout.json @@ -0,0 +1,13154 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "WaveMo: Learning Wavefront Modulations to See Through Scattering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": "Mingyang Xie" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": " Haiyun Guo" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": " Brandon Y. Feng" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": " Lingbo Jin" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": " Ashok Veeraraghavan" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": " Christopher A. Metzler" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 184, + 195, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 184, + 195, + 200 + ], + "spans": [ + { + "bbox": [ + 74, + 184, + 195, + 200 + ], + "type": "text", + "content": "1University of Maryland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 218, + 185, + 301, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 185, + 301, + 200 + ], + "spans": [ + { + "bbox": [ + 218, + 185, + 301, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 218, + 185, + 301, + 200 + ], + "type": "text", + "content": "Rice University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 185, + 516, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 185, + 516, + 200 + ], + "spans": [ + { + "bbox": [ + 324, + 185, + 516, + 200 + ], + "type": "text", + "content": "3Massachusetts Institute of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 252, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 515 + ], + "type": "text", + "content": "Imaging through scattering media is a fundamental and pervasive challenge in fields ranging from medical diagnostics to astronomy. A promising strategy to overcome this challenge is wavefront modulation, which induces measurement diversity during image acquisition. Despite its importance, designing optimal wavefront modulations to image through scattering remains under-explored. This paper introduces a novel learning-based framework to address the gap. Our approach jointly optimizes wavefront modulations and a computationally lightweight feedforward \"proxy\" reconstruction network. This network is trained to recover scenes obscured by scattering, using measurements that are modified by these modulations. The learned modulations produced by our framework generalize effectively to unseen scattering scenarios and exhibit remarkable versatility. During deployment, the learned modulations can be decoupled from the proxy network to augment other more computationally expensive restoration algorithms. Through extensive experiments, we demonstrate our approach significantly advances the state of the art in imaging through scattering media. Our project webpage is at https://wavemo2024.github.io/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 540, + 128, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 128, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 128, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 561, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 681 + ], + "type": "text", + "content": "Imaging through scattering media presents a significant challenge across diverse scenarios, ranging from navigating with fog [2, 28, 30], rain [9, 48, 49], or murky water to recovering intricate structures through human skin and tissue [10, 42, 47]. The core of this challenge are the irregular phase delays light experiences as it scatters. These phase delays blur and warp any images captured through the scattering media. Effectively addressing this issue is crucial for unlocking new computer vision capabilities in fields such as medical imaging and astronomy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "type": "text", + "content": "From a frequency domain perspective, the effect of a" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 227, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 310, + 227, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 310, + 227, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 310, + 227, + 541, + 302 + ], + "type": "image", + "image_path": "6bf9bf293682b61dd5a272e1c0d2960a5b835780691dc2254bbf65f4eeeb6ea1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 305, + 451, + 365 + ], + "blocks": [ + { + "bbox": [ + 310, + 305, + 451, + 365 + ], + "lines": [ + { + "bbox": [ + 310, + 305, + 451, + 365 + ], + "spans": [ + { + "bbox": [ + 310, + 305, + 451, + 365 + ], + "type": "image", + "image_path": "304bc80251eb51cc6c05a9c89f2fd7979305a2a4b3813d9ee7cfb8079200cb39.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 378, + 545, + 467 + ], + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 467 + ], + "type": "text", + "content": "Figure 1. Learned Wavefront Modulations. Top: During acquisition, we can modulate the wavefront of scattered light by using a spatial light modulator (SLM), and capture a set of image measurements useful for scene reconstruction. Bottom left: We propose learning modulations that enhance our ability to recover the scattered scene. Bottom right: Our learned modulations drastically improve the reconstruction quality of a state-of-the-art method [8] that previously applies randomly chosen modulations." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 457, + 306, + 541, + 365 + ], + "blocks": [ + { + "bbox": [ + 457, + 306, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 457, + 306, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 457, + 306, + 541, + 365 + ], + "type": "image", + "image_path": "200f2d47f9cc1f78dccb64b5e02846836fa8cf306d9b3d76bcf15b88264558c9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 569 + ], + "type": "text", + "content": "scene passing through scattering is the destruction or filtering of frequency content. Recovering these lost frequencies is very difficult using purely computational methods, which are effectively forced to speculate on the missing frequencies based on those that have been preserved. A question naturally arises: Can we modify the measurement process to prevent the loss of information during acquisition?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": "A promising solution is to actively modulate the scattered wavefront during acquisition, in hopes that applying multiple modulations increases the chance that a frequency component is preserved. This approach is related to phase diversity imaging [11, 15, 17, 25, 27, 37, 41, 45], which typically captures multiple scattered images of a static scene with different modulations applied using deformable mirrors or spatial light modulators (SLM). However, phase diversity imaging traditionally was limited to imaging simple scenes with mild optical aberrations, and the phase modulations were often selected based on simple heuristics. The recent development of neural wavefront shaping (NeuWS) [8]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 128, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 128, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 128, + 712 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25276" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 523, + 335 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 523, + 335 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 523, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 523, + 335 + ], + "type": "image", + "image_path": "6c47af18287b17a17f98b7d80316709f7af528a3b7cbaa6de49cf4471c5abf8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 345, + 548, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 548, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 548, + 401 + ], + "type": "text", + "content": "Figure 2. Overview of the Proposed End-to-end Learning Framework. During training, we jointly optimize the implicit neural representation for the wavefront modulations and the proxy reconstruction network in an end-to-end fashion. During inference, one can apply the learned wavefront modulations to any reconstruction algorithms for imaging through scattering, either a trained feed-forward reconstruction network or an unsupervised iterative optimization algorithm [8]. The former approach is far faster and performs better when test data and training data fall into similar distributions, while the latter generalizes better to unseen distributions of target scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": "brings a significant breakthrough by leveraging recent innovations in machine learning and enables imaging high-resolution and dynamic scenes through severe aberrations. Still, the wavefront modulations applied by NeuWS during acquisition are chosen randomly without explicit consideration of their impact on frequency preservation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we take a significant step towards a more systematic and principled approach in designing effective modulations that better preserve frequencies against scattering during image acquisition. Moving beyond the traditional reliance on simple heuristics for determining modulation patterns, our method combines the strengths of differentiable optimization and data-driven learning. Central to our approach is the novel integration of the optical model of wavefront modulations with a proxy reconstruction network, resulting in a fully differentiable system. This integration allows for the simultaneous optimization of both the proxy reconstruction network and the modulation patterns, which are optimized end-to-end based on a large, simulated training dataset. While imaging through scattering often requires iterative optimization algorithms because data-driven feedforward networks often fail to generalize outside their training domain, this paper demonstrates the remarkable finding that training a hard-to-generalize network can" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 422, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 566 + ], + "type": "text", + "content": "become a useful proxy allowing us to learn generalizable modulations. Crucially, this proxy feedforward network allows us to bypass the expensive backpropagation computation through an iterative reconstruction algorithm, which would be necessary, but hopelessly impractical, if we were to naively optimize the modulations: each iteration would require us to finish an entire iterative reconstruction algorithm. As will be shown in our simulated and real data experiments, our learned modulations significantly enhance the reconstruction capability of both our data-driven proxy network and a state-of-the-art iterative optimization-based method [8] for imaging through scattering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 567, + 409, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 409, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 409, + 578 + ], + "type": "text", + "content": "Our contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 580, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 580, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 580, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 315, + 580, + 545, + 616 + ], + "type": "text", + "content": "- We present a novel end-to-end learning framework to optimize acquisition-time wavefront modulations to enhance our abilities to see through scattering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 616, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 616, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 315, + 616, + 545, + 663 + ], + "type": "text", + "content": "- We demonstrate in both simulated and real experiments that our learned modulations substantially improve image reconstruction quality and effectively generalize to unseen targets and scattering media." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 664, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 664, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 664, + 545, + 712 + ], + "type": "text", + "content": "- We show that the learned modulations can be decoupled from the jointly-trained proxy reconstruction network and significantly enhance the reconstruction quality of state-of-the-art unsupervised approaches." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25277" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "content": "Learning-based Imaging Through Scattering Media. A detailed review of imaging through scattering media can be found in [10]. We highlight a few recent learning-based approaches. Various data-driven approaches can be applied to image through scattering [5, 13, 20, 22, 35, 36]. They generally capture or synthesize large quantities of training data and then learn a mapping from corrupted measurements to reconstructed scenes. However, they tend to struggle with out-of-distribution data. There have also been unsupervised algorithms by fitting network-based parameterizations of the scene to a collection of measurements [8, 15], which can generalize across different distributions of aberrations and target scenes. Both of these methods take advantage of phase diversity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 273, + 289, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 289, + 417 + ], + "type": "text", + "content": "End-to-End Learning of Optical Systems. End-to-end learning is a flexible learning-based system design framework that describes optical systems using parameterized differentiable optical models which can then be optimized with training data [31]. End-to-end learning has been applied successfully for extended depth-of-field imaging [14, 21, 29, 31], depth estimation [4, 14, 29, 43], high-dynamic-range imaging [23, 33], seeing through near-lens occlusions [30], and many other applications. To facilitate end-to-end learning, prior works have also incorporated the usage of a proxy network to mimic the response of a specific optical system or black-box algorithm [26, 39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 289, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 587 + ], + "type": "text", + "content": "Phase Diversity Imaging. Phase diversity imaging (PDI) can recover a clear image of an aberrated scene by obtaining images with deliberately introduced known aberrations. PDI does not pose strong assumptions on the target [45] or the aberration [16], and the optics and calibration process required are relatively simple compared to other adaptive optics imaging [11]. While early applications of PDI are often constrained to simple aberrations due to computation complexities [32], the introduction of coded diffraction patterns has enhanced its capacity for reconstructing intricate scenes [3]. Recent PDI developments include improved wavefront sensing [45, 50], phase retrieval [27, 44], adaptive optics [7, 15], and imaging through scattering [6, 8, 46]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 596, + 231, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 231, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 231, + 609 + ], + "type": "text", + "content": "3. Learned Wavefront Modulations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "type": "text", + "content": "Here we describe our approach for learning wavefront modulations through training a proxy reconstruction network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 161, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 161, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 161, + 659 + ], + "type": "text", + "content": "3.1. Scattering Problem" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "content": "This paper focuses on the scenario where capturing an image through scattering media can be modeled as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 698, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 698, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 129, + 698, + 287, + 711 + ], + "type": "interline_equation", + "content": "y = h (\\phi) * x + \\epsilon , \\tag {1}", + "image_path": "5dc5b75626b030ab938c923fd0968b0ccdf41317f7b10cb4f512ea341db7d184.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is the captured measurement, " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is the target scene that we aim to reconstruct, " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is noise, and " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "h(\\phi)" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is the unknown, spatially invariant point spread function (PSF) describing the optical scattering, which manifests as unknown phase delays " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " to the wavefront. Assuming the target object is illuminated with spatially incoherent monochromatic light, the PSF " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is related to the phase error " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "h(\\phi) = |\\mathcal{F}(m\\circ e^{j(\\phi)})|^2" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is the 2D Fourier transform, " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is the hadamard product and " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " is the aperture mask [12]. Our goal is to recover " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " captured through " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 201, + 401, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 401, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 401, + 213 + ], + "type": "text", + "content": "3.2. Phase Diversity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "content": "Phase diversity imaging seeks additional information about the phase error " + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "content": " by capturing a sequence of measurements " + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "inline_equation", + "content": "y_{1}, y_{2}, \\ldots, y_{K}" + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "content": ", each with additional known phase errors " + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\gamma_{1}, \\gamma_{2}, \\ldots, \\gamma_{K}" + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "content": ". As a result of these additional modulations, the imaging forward model becomes " + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "inline_equation", + "content": "y_{i} = h(\\phi + \\gamma_{i}) * x + \\epsilon_{i}" + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "content": ". In traditional approaches, the phase modulations " + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\{\\gamma_{i}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 304, + 220, + 547, + 328 + ], + "type": "text", + "content": " were typically either simple defocus sweeps or randomly chosen patterns generated from Zernike polynomials [8, 25], and their reconstruction objective seeks to minimize" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 338, + 338, + 545, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 338, + 338, + 545, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P D I} (\\hat {x}, \\phi) = \\sum_ {i = 1} ^ {K} \\left| \\left| y _ {i} - h \\left(\\phi + \\gamma_ {i}\\right) * \\hat {x} \\right| \\right| ^ {2}, \\tag {2}", + "image_path": "656ae62da52e2e78c8f0a3c525df9b23c65af80c3aafa6bcbf2dcc5c3ab12f20.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "content": "over " + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "content": " is the estimate of scene " + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "content": ", assuming " + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}" + }, + { + "bbox": [ + 304, + 380, + 545, + 403 + ], + "type": "text", + "content": " follows an i.i.d. additive Gaussian distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 413, + 430, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 430, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 430, + 425 + ], + "type": "text", + "content": "3.3. Learned Modulations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "text", + "content": "While previous works have leveraged differentiable proxies of physical systems to design better algorithms [26, 39], this work leverages a differentiable proxy, " + }, + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "text", + "content": ", of non-differentiable algorithms to design better physical systems—specifically the phase modulation " + }, + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\Gamma = \\{\\gamma_i\\}_{i=1}^K" + }, + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "text", + "content": " used to image through scattering." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "content": "We sample a training image " + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "content": " and simulate passing it through an optical aberration drawn from some known distribution while we apply " + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "content": " learnable modulations to the scattered wavefront. This model allows us to simulate a collection of " + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "inline_equation", + "content": "Y_{n}^{\\Gamma} = \\{y_{1}^{n}, y_{2}^{n}, \\ldots, y_{K}^{n}\\}" + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "content": ", which are different scattered observations of " + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 304, + 505, + 545, + 576 + ], + "type": "text", + "content": " based on Eq. (1)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "type": "text", + "content": "We then optimize both the proxy reconstruction algorithm " + }, + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "type": "text", + "content": " and the modulation patterns " + }, + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 577, + 545, + 601 + ], + "type": "text", + "content": " to minimize" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 611, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 611, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 343, + 611, + 545, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l e a r n e d}} (\\mathcal {P}, \\Gamma) = \\sum_ {n = 1} ^ {N} \\| \\mathcal {P} \\left(Y _ {n} ^ {\\Gamma}\\right) - x _ {n} \\| ^ {2}. \\tag {3}", + "image_path": "3129f418c4fef58dfc5124e384e63e9d87913def850cc66c84a90b8dd1ba5393.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "Note that the ultimate goal of the learning is not to design a set of modulations specific to the network " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": ". Rather, we use the performance of " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " as a proxy to probe how effective the modulations " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " are for the scattering problem. The differentiability of " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " allows for back-propagation to guide" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25278" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "the optimization of the modulations. As we will demonstrate, while " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " itself may not generalize outside of its training data domain, the learned modulations " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " are compatible with other more generalizable reconstruction algorithms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": "Our end-to-end training pipeline is illustrated in Figure 2. For the proxy reconstruction network, we choose a U-Net with skip connections and self-attention modules [24]. To regularize the highly non-convex problem of optimizing modulations, we use an implicit neural representation for the modulations " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": ". Specifically, the MLP " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": " takes a fixed vector " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": " with 28 channels, which corresponds to the first 28 Zernike polynomials [18]; it outputs a 16-channel vector corresponding to 16 modulation patterns: " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\Gamma_{G} = G(Z)" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 288, + 251 + ], + "type": "text", + "content": "Incorporating the implicit neural representation for the modulations, our final loss function becomes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 258, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 258, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 85, + 258, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n a l}} (\\mathcal {P}, G) = \\sum_ {n = 1} ^ {N} \\left\\| \\mathcal {P} \\left(Y _ {n} ^ {\\Gamma_ {G}}\\right) - x _ {n} \\right\\| ^ {2}, \\tag {4}", + "image_path": "d4e4f7d117287ccb886d8bdde474c9c106baa39342161447500bf48d25675e7b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 219, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 219, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 219, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 297, + 219, + 312 + ], + "type": "inline_equation", + "content": "Y_{n}^{\\Gamma_{G}} = \\{h(\\phi_{n} + G(Z)_{i})*x_{n}\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 47, + 297, + 219, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 316, + 217, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 217, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 217, + 330 + ], + "type": "text", + "content": "3.4. Applying Learned Modulations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "type": "text", + "content": "While the trained proxy network " + }, + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "type": "text", + "content": " demonstrates considerable capability in recovering scenes from scattering in our experiments, supervised data-driven methods often fail to generalize to novel or cross-domain scenes, particularly when compared to unsupervised approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "text", + "content": "Therefore, after obtaining the learned modulations " + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\Gamma_{G}" + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "text", + "content": " and applying them during real-world acquisition, we send the resulting modulated measurements to an unsupervised iterative optimization-based reconstruction algorithm [8] which does not suffer from the generalization issue of data-driven methods. In effect, our method offers a best-of-both-worlds scenario: the modulations " + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\Gamma_{G}" + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "text", + "content": " are effectively learned thanks to the joint supervised training with the proxy network " + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 395, + 288, + 539 + ], + "type": "text", + "content": ", but their enhanced data acquisition quality is transferrable to other reconstruction algorithms, thus allowing us to benefit from the generalization and domain adaptability inherent in unsupervised reconstruction methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 548, + 233, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 233, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 233, + 562 + ], + "type": "text", + "content": "4. Analysis of Learned Modulations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "type": "text", + "content": "Before we show the performance of the learned modulations with real-world experiments in Section 6, we first analyze the effectiveness of learned modulations in simulation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 611, + 170, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 170, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 170, + 624 + ], + "type": "text", + "content": "4.1. Impact on Frequency" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "When the scattering media scrambles the incoming wavefront, it destroys some of the target scene's frequencies, either by destructively interfering with them or by scattering them outside the imaging sensor. If those frequencies are not captured in the measurement, it becomes difficult to faithfully recover them. By applying multiple learned wavefront modulation patterns, we increase the chance that each" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 333, + 73, + 511, + 220 + ], + "blocks": [ + { + "bbox": [ + 333, + 73, + 511, + 220 + ], + "lines": [ + { + "bbox": [ + 333, + 73, + 511, + 220 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 511, + 220 + ], + "type": "image", + "image_path": "4261602f59e9f5979a37357f0aefcc7dd2e380c08a7fc2324d923bb286037c05.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 228, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 547, + 318 + ], + "type": "text", + "content": "Figure 3. MTF Comparison. We compare the MTF of learned wavefront modulations vs. the MTF of modulations randomly drawn from the same distribution of the scattering media. The X-axis represents spatial frequency, and the Y-axis represents the modulation transfer (the higher the better). Our learned wavefront modulations exhibit a higher MTF compared to unoptimized ones, especially over higher frequency bands, suggesting the learned modulations preserve more high-frequency information." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 339, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 388 + ], + "type": "text", + "content": "frequency is better preserved by at least one of the modulations. With the extra information from all modulated measurements, a better image reconstruction performance could be achieved regardless of the reconstruction algorithm." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "content": "Thus, we assess the performance of our learned modulations by their combined ability to preserve frequencies. We choose to the commonly used Modulation Transfer Function (MTF) as our metric [1, 19, 34, 38, 40]. MTF measures the contrast produced by an imaging system at different spatial frequencies, defined as " + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{MTF}_i^\\omega (\\phi) = \\left|\\mathcal{F}[h_i(\\phi)]\\right|^{\\omega}" + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "content": " is the PSF for the " + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "content": "-th modulation and the superscript " + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "content": " denotes indexing the MTF at a specific frequency " + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 388, + 546, + 507 + ], + "type": "text", + "content": ". For the multiple modulations in our problem setup, we define the combined MTF as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 518, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 518, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 339, + 518, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\mathrm {M T F} _ {\\text {c o m b}} ^ {\\omega} \\left(\\left\\{\\gamma_ {i} \\right\\}, \\phi\\right) = \\max _ {i} \\mathrm {M T F} _ {i} ^ {\\omega} \\left(\\phi + \\gamma_ {i}\\right), \\tag {5}", + "image_path": "2569862edc29f624f830fdc442f2eb07058f6d3e76323af344d6071da3f8a54e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 545, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 568 + ], + "type": "text", + "content": "where for each frequency " + }, + { + "bbox": [ + 305, + 545, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 305, + 545, + 545, + 568 + ], + "type": "text", + "content": ", we take the maximum over the MTFs calculated from all modulations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In Figure 3, we plot the combined MTF for both the learned wavefront modulations and the random ones as baseline. The learned ones have a clear advantage at high frequencies (the higher the frequency is, the further away it is to the center of the X-axis), suggesting that the learned modulations are better at preserving finer details of the target scene. Note that during training, we never explicitly encourage the MTF to be higher — the only loss term we use is the mean square error between reconstruction and ground truth image. This implies that our end-to-end training implicitly encourages the modulations to preserve higher-frequency information." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25279" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 113, + 122 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 113, + 122 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 113, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 113, + 122 + ], + "type": "image", + "image_path": "ae4e87b6a0096d2b85d16fb5f7643b297f82d5925d9807e5cfd4de11e4f582c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 114, + 70, + 166, + 121 + ], + "blocks": [ + { + "bbox": [ + 114, + 70, + 166, + 121 + ], + "lines": [ + { + "bbox": [ + 114, + 70, + 166, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 70, + 166, + 121 + ], + "type": "image", + "image_path": "28dfa595812ef7095784887990f417e343e725e24eed3b9e237c7cd6941ecaff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 121, + 211, + 129 + ], + "lines": [ + { + "bbox": [ + 122, + 121, + 211, + 129 + ], + "spans": [ + { + "bbox": [ + 122, + 121, + 211, + 129 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 218, + 121 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 218, + 121 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 218, + 121 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 218, + 121 + ], + "type": "image", + "image_path": "7f6a105f7e6971295d2476e263a677718acd7cf9031db511b9dae349e1bbc542.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 70, + 271, + 121 + ], + "blocks": [ + { + "bbox": [ + 219, + 70, + 271, + 121 + ], + "lines": [ + { + "bbox": [ + 219, + 70, + 271, + 121 + ], + "spans": [ + { + "bbox": [ + 219, + 70, + 271, + 121 + ], + "type": "image", + "image_path": "4bf19fab200664f039d44da3ace70f90fa1a9f7fe36032282e2d38064744f047.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 129, + 113, + 180 + ], + "blocks": [ + { + "bbox": [ + 61, + 129, + 113, + 180 + ], + "lines": [ + { + "bbox": [ + 61, + 129, + 113, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 113, + 180 + ], + "type": "image", + "image_path": "b7eae1b38ff40e8b4237810bc2def67b1e982ebbda40c71c3ac48c8861f4645c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 181, + 227, + 187 + ], + "lines": [ + { + "bbox": [ + 107, + 181, + 227, + 187 + ], + "spans": [ + { + "bbox": [ + 107, + 181, + 227, + 187 + ], + "type": "text", + "content": "Reconstructions with no modulations" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 113, + 129, + 166, + 180 + ], + "blocks": [ + { + "bbox": [ + 113, + 129, + 166, + 180 + ], + "lines": [ + { + "bbox": [ + 113, + 129, + 166, + 180 + ], + "spans": [ + { + "bbox": [ + 113, + 129, + 166, + 180 + ], + "type": "image", + "image_path": "a90d055f6c28e5e53d0060d5ead0525c35b2d673fe8fe179ed24111ba04613bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 167, + 129, + 218, + 180 + ], + "blocks": [ + { + "bbox": [ + 167, + 129, + 218, + 180 + ], + "lines": [ + { + "bbox": [ + 167, + 129, + 218, + 180 + ], + "spans": [ + { + "bbox": [ + 167, + 129, + 218, + 180 + ], + "type": "image", + "image_path": "07435ec1a0c0b0fe25f8ce719dac3fad15115e8e6be4308c61f5ba211e20d3c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 219, + 129, + 271, + 180 + ], + "blocks": [ + { + "bbox": [ + 219, + 129, + 271, + 180 + ], + "lines": [ + { + "bbox": [ + 219, + 129, + 271, + 180 + ], + "spans": [ + { + "bbox": [ + 219, + 129, + 271, + 180 + ], + "type": "image", + "image_path": "8ea33a876323ea203b88fe116e5bbbb111c9fdade108c5aac0ee6981746113eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 61, + 189, + 113, + 240 + ], + "blocks": [ + { + "bbox": [ + 61, + 189, + 113, + 240 + ], + "lines": [ + { + "bbox": [ + 61, + 189, + 113, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 189, + 113, + 240 + ], + "type": "image", + "image_path": "a54ef6f114dc57e21a90260ebe699df4b313c61301812e6e9929f48109855043.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 240, + 236, + 247 + ], + "lines": [ + { + "bbox": [ + 99, + 240, + 236, + 247 + ], + "spans": [ + { + "bbox": [ + 99, + 240, + 236, + 247 + ], + "type": "text", + "content": "Reconstructions with random modulations" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 113, + 189, + 166, + 239 + ], + "blocks": [ + { + "bbox": [ + 113, + 189, + 166, + 239 + ], + "lines": [ + { + "bbox": [ + 113, + 189, + 166, + 239 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 166, + 239 + ], + "type": "image", + "image_path": "c83fcedd195d827d19f67041e331862c349fc697c7cfe79b54662748a72ab367.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 167, + 189, + 218, + 239 + ], + "blocks": [ + { + "bbox": [ + 167, + 189, + 218, + 239 + ], + "lines": [ + { + "bbox": [ + 167, + 189, + 218, + 239 + ], + "spans": [ + { + "bbox": [ + 167, + 189, + 218, + 239 + ], + "type": "image", + "image_path": "e03cce180316ff1548af11d49c2e48c55efd038d4b400208f4cdcc2ac9da85a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 219, + 189, + 271, + 239 + ], + "blocks": [ + { + "bbox": [ + 219, + 189, + 271, + 239 + ], + "lines": [ + { + "bbox": [ + 219, + 189, + 271, + 239 + ], + "spans": [ + { + "bbox": [ + 219, + 189, + 271, + 239 + ], + "type": "image", + "image_path": "73db0ee5fd56def69b19dd9935954cd437312fad5a09aef4a292d652528b8a20.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 61, + 248, + 113, + 299 + ], + "blocks": [ + { + "bbox": [ + 61, + 248, + 113, + 299 + ], + "lines": [ + { + "bbox": [ + 61, + 248, + 113, + 299 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 113, + 299 + ], + "type": "image", + "image_path": "2a07707edb5bdbb42878c0cb04158088039bd01c9f31fe2a96a1aa905756b456.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 299, + 233, + 306 + ], + "lines": [ + { + "bbox": [ + 97, + 299, + 233, + 306 + ], + "spans": [ + { + "bbox": [ + 97, + 299, + 233, + 306 + ], + "type": "text", + "content": "Reconstructions with learned modulations" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 113, + 248, + 166, + 299 + ], + "blocks": [ + { + "bbox": [ + 113, + 248, + 166, + 299 + ], + "lines": [ + { + "bbox": [ + 113, + 248, + 166, + 299 + ], + "spans": [ + { + "bbox": [ + 113, + 248, + 166, + 299 + ], + "type": "image", + "image_path": "765ea69ced7093f2a9666955b35415c4d88ce303cc154f0e19f4312994c90522.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 167, + 248, + 218, + 299 + ], + "blocks": [ + { + "bbox": [ + 167, + 248, + 218, + 299 + ], + "lines": [ + { + "bbox": [ + 167, + 248, + 218, + 299 + ], + "spans": [ + { + "bbox": [ + 167, + 248, + 218, + 299 + ], + "type": "image", + "image_path": "b0274790ee82953a8882b85180b0f99a4e63673a74fe3c300a1936ffd676d4a5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 219, + 248, + 271, + 299 + ], + "blocks": [ + { + "bbox": [ + 219, + 248, + 271, + 299 + ], + "lines": [ + { + "bbox": [ + 219, + 248, + 271, + 299 + ], + "spans": [ + { + "bbox": [ + 219, + 248, + 271, + 299 + ], + "type": "image", + "image_path": "59bfb4d26c9e4ab586d05dec843a3713ff5fe37d2906257ef48a085768018ecb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 61, + 307, + 113, + 358 + ], + "blocks": [ + { + "bbox": [ + 61, + 307, + 113, + 358 + ], + "lines": [ + { + "bbox": [ + 61, + 307, + 113, + 358 + ], + "spans": [ + { + "bbox": [ + 61, + 307, + 113, + 358 + ], + "type": "image", + "image_path": "ff9015397632b08ab254d48afd8da86cb26a708886cf99b10532fcffc21cd37a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 409 + ], + "type": "text", + "content": "Figure 4. Proxy Network Reconstruction on Simulated Scattering. Simulated imaging results through different aberrations. Learned modulations lead to a better quality." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 113, + 307, + 166, + 358 + ], + "blocks": [ + { + "bbox": [ + 113, + 307, + 166, + 358 + ], + "lines": [ + { + "bbox": [ + 113, + 307, + 166, + 358 + ], + "spans": [ + { + "bbox": [ + 113, + 307, + 166, + 358 + ], + "type": "image", + "image_path": "dbbb70a0505a226bbf409309e93ea1b064454a72f7503830370a1c03f8f799d6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 358, + 187, + 365 + ], + "lines": [ + { + "bbox": [ + 144, + 358, + 187, + 365 + ], + "spans": [ + { + "bbox": [ + 144, + 358, + 187, + 365 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 167, + 307, + 218, + 358 + ], + "blocks": [ + { + "bbox": [ + 167, + 307, + 218, + 358 + ], + "lines": [ + { + "bbox": [ + 167, + 307, + 218, + 358 + ], + "spans": [ + { + "bbox": [ + 167, + 307, + 218, + 358 + ], + "type": "image", + "image_path": "94dcc419b901c3191c3367c89c84f6ed6d094baae07e0f5a1093866d54a97b46.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 219, + 307, + 271, + 358 + ], + "blocks": [ + { + "bbox": [ + 219, + 307, + 271, + 358 + ], + "lines": [ + { + "bbox": [ + 219, + 307, + 271, + 358 + ], + "spans": [ + { + "bbox": [ + 219, + 307, + 271, + 358 + ], + "type": "image", + "image_path": "f6c3338a0a8b51c74419e11295df554fd0cb74f5ada46d27e466bdcb9817fa0a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 431, + 192, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 192, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 192, + 443 + ], + "type": "text", + "content": "4.2. Impact on Reconstruction" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 449, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 593 + ], + "type": "text", + "content": "We evaluate our jointly learned modulations and proxy network both quantitatively and qualitatively. Both the training and test set are from the MIT Places 365 Dataset [51]. Training details are provided later in Section 5. Here, we use two baselines: (1) simply training a feed-forward reconstruction network without any wavefront modulations, and (2) training a feed-forward reconstruction network with randomly chosen, fixed wavefront modulations as used in prior PDI work. For a fair comparison, our approach and baselines use the same network architecture and are trained on the same data for the same number of iterations. Their only difference is the usage of wavefront modulations." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "As shown in Figure 4, compared to baselines, reconstructions using learned modulations demonstrate finer details and higher consistency with the ground truth. Table 1 shows the quantitative comparison averaged over a test set of 1,000 images, where our approach outperforms the baselines by over " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "3.9\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". While the results above are all from the jointly trained proxy reconstruction network, we will show in Section 6 that our learned wavefront modulations can decouple from the proxy network and also improve reconstruction quality based on iterative optimization [8]." + } + ] + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 341, + 64, + 512, + 128 + ], + "blocks": [ + { + "bbox": [ + 341, + 64, + 512, + 128 + ], + "lines": [ + { + "bbox": [ + 341, + 64, + 512, + 128 + ], + "spans": [ + { + "bbox": [ + 341, + 64, + 512, + 128 + ], + "type": "table", + "html": "
MetricModulations
NoneRandomLearned
PSNR25.47626.43930.391
SSIM0.76400.79800.9082
", + "image_path": "f6a8b3ab373184aff93be4db4ead0e29cfd2c7f6b81fb083c2d806f07ccbd6a9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 131, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 305, + 131, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 547, + 229 + ], + "type": "text", + "content": "Table 1. PSNR and SSIM of Simulation Results. We report the test-time performance of our jointly trained modulations and the proxy network. We compare it against two baselines where either we do not use wavefront modulations or use randomly chosen, unoptimized modulations. The metrics are averaged over the reconstruction of 1000 test images, each of which is scattered with a randomly sampled optic aberration that is unseen during training. The results demonstrate the huge performance improvement from the optimization of wavefront modulations." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 246, + 403, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 403, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 403, + 259 + ], + "type": "text", + "content": "5. Implementation" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 266, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 434 + ], + "type": "text", + "content": "Software Implementation & Training. For the proxy reconstruction network, we use an attention U-Net [24]; for the MLP that represents the 16 wavefront modulations, we use a two-layer MLP with leaky-ReLU activation. In each training iteration, the optical aberrations are sampled randomly from Zernike basis functions. The standard deviation of the coefficients for each Zernike basis is randomly generated from a uniform distribution from 5 to 6. We train our proposed framework for 2 million iterations on the MIT Places Dataset [51] with the Adam Optimizer and a learning rate of 0.001. Training took 12 hours on an NVIDIA RTX A6000 GPU. For experimental results, we finetune our trained model on the captured bio-tissue data under the same training settings except with a learning rate of 0.0001." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": "Hardware Implementation. Our optical configuration is depicted in Figure 6. A continuous-wave laser with a wavelength of " + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "532\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": " was collimated and subsequently passed through a laser speckle reducer (Optotune LSR-4C-L). The resulting spatially incoherent light was then polarized and reflected by a digital micromirror device (DMD). The DMD (LightCrafter™ 4500) display is capable of rapidly presenting 8-bit target images. For ease of assembly, a reflecting mirror was incorporated after the DMD. Then, the light is focused by lens L0 (" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "f = 100\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": ") onto the front focal plane of L1, a " + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": " microscopic objective. Our optical aberration—resin painted on glass—was placed at the same location. L1 and a tube lens L2 (" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "f = 200\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": ") constitute a " + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "4f" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": " system, and thereby the SLM is manipulating the Fourier plane. We used a HOLOEYE LETO-3 phase-only SLM to modulate the wavefront. Finally, the light reflected by the SLM was imaged on the camera (" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "1384\\times 1036" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": " pixel Grasshopper3) via L3 (" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "inline_equation", + "content": "f = 500\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 442, + 546, + 670 + ], + "type": "text", + "content": "). Our capture speed is limited to 30 fps by the camera frame rate." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Experimental Dataset. Our experimental data are from human pathological slides, imaged with optical microscopy (Zeiss Microscope Axio Imager.A2) and a " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " objective." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25280" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 136, + 107 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 136, + 107 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 136, + 107 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 136, + 107 + ], + "type": "image", + "image_path": "efb0def738077b443752410085a3f2ff235ac1f2fa9b6b0359563b79b30ff3e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 70, + 217, + 108 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 217, + 108 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 217, + 108 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 217, + 108 + ], + "type": "image", + "image_path": "6a084187dec390a84202560764ea945d0e773c5655ff109dda528d9ff465100d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 108, + 222, + 114 + ], + "lines": [ + { + "bbox": [ + 138, + 108, + 222, + 114 + ], + "spans": [ + { + "bbox": [ + 138, + 108, + 222, + 114 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 223, + 71, + 298, + 108 + ], + "blocks": [ + { + "bbox": [ + 223, + 71, + 298, + 108 + ], + "lines": [ + { + "bbox": [ + 223, + 71, + 298, + 108 + ], + "spans": [ + { + "bbox": [ + 223, + 71, + 298, + 108 + ], + "type": "image", + "image_path": "0901dc6b4518e7728ccd61e687c56212ac5ed65bb477e5e033ce5c80f4bde7e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 116, + 136, + 153 + ], + "blocks": [ + { + "bbox": [ + 60, + 116, + 136, + 153 + ], + "lines": [ + { + "bbox": [ + 60, + 116, + 136, + 153 + ], + "spans": [ + { + "bbox": [ + 60, + 116, + 136, + 153 + ], + "type": "image", + "image_path": "99fb54fef4243cba83d8ea5a83c5b3f99cdde1d1822ecbb0ff83f4c15e3ae0ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 141, + 116, + 217, + 153 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 217, + 153 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 217, + 153 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 217, + 153 + ], + "type": "image", + "image_path": "017253ac0be3850a7dc8436278138eda3f0876b69c0866a24795d687c9d94ef5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 223, + 117, + 298, + 154 + ], + "blocks": [ + { + "bbox": [ + 223, + 117, + 298, + 154 + ], + "lines": [ + { + "bbox": [ + 223, + 117, + 298, + 154 + ], + "spans": [ + { + "bbox": [ + 223, + 117, + 298, + 154 + ], + "type": "image", + "image_path": "678c56530a8d64c3cc9866631ac7033802f18935bcdb51a7394b9622f6419919.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 154, + 246, + 161 + ], + "lines": [ + { + "bbox": [ + 113, + 154, + 246, + 161 + ], + "spans": [ + { + "bbox": [ + 113, + 154, + 246, + 161 + ], + "type": "text", + "content": "Proxy reconstructions with no modulations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 163, + 136, + 200 + ], + "blocks": [ + { + "bbox": [ + 60, + 163, + 136, + 200 + ], + "lines": [ + { + "bbox": [ + 60, + 163, + 136, + 200 + ], + "spans": [ + { + "bbox": [ + 60, + 163, + 136, + 200 + ], + "type": "image", + "image_path": "b09bdd01b6600f5fff9170102afac143eb9a046fa5b03da78c1c0b32c2d3b979.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 141, + 163, + 217, + 200 + ], + "blocks": [ + { + "bbox": [ + 141, + 163, + 217, + 200 + ], + "lines": [ + { + "bbox": [ + 141, + 163, + 217, + 200 + ], + "spans": [ + { + "bbox": [ + 141, + 163, + 217, + 200 + ], + "type": "image", + "image_path": "77829edce75a451cca4a3d9b90a25e70a7031629ab24cbe804c4c9f4b0da8346.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 163, + 298, + 201 + ], + "blocks": [ + { + "bbox": [ + 223, + 163, + 298, + 201 + ], + "lines": [ + { + "bbox": [ + 223, + 163, + 298, + 201 + ], + "spans": [ + { + "bbox": [ + 223, + 163, + 298, + 201 + ], + "type": "image", + "image_path": "fd70babec234fc43193b89a0fe4201e9d76549eaa7058233d24868d85c795333.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 201, + 254, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 254, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 254, + 209 + ], + "type": "text", + "content": "Proxy reconstructions with random modulations" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 60, + 210, + 137, + 247 + ], + "blocks": [ + { + "bbox": [ + 60, + 210, + 137, + 247 + ], + "lines": [ + { + "bbox": [ + 60, + 210, + 137, + 247 + ], + "spans": [ + { + "bbox": [ + 60, + 210, + 137, + 247 + ], + "type": "image", + "image_path": "2eb7d0536d619d6957595f3b10efbee998f9a9f79113f83f0e4e97c07fcf55a5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 247, + 253, + 255 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 253, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 253, + 255 + ], + "type": "text", + "content": "Proxy reconstructions with learned modulations" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 210, + 217, + 247 + ], + "blocks": [ + { + "bbox": [ + 141, + 210, + 217, + 247 + ], + "lines": [ + { + "bbox": [ + 141, + 210, + 217, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 210, + 217, + 247 + ], + "type": "image", + "image_path": "67e363f3c957ac5054b98fd49222ec208a0a19f6d7397a28e157d8f154b5f017.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 223, + 210, + 298, + 247 + ], + "blocks": [ + { + "bbox": [ + 223, + 210, + 298, + 247 + ], + "lines": [ + { + "bbox": [ + 223, + 210, + 298, + 247 + ], + "spans": [ + { + "bbox": [ + 223, + 210, + 298, + 247 + ], + "type": "image", + "image_path": "cdff0b9ff06fe6e244506cc6fcd7062080e16f58099d718e97a05f3e5781b72a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 60, + 256, + 137, + 293 + ], + "blocks": [ + { + "bbox": [ + 60, + 256, + 137, + 293 + ], + "lines": [ + { + "bbox": [ + 60, + 256, + 137, + 293 + ], + "spans": [ + { + "bbox": [ + 60, + 256, + 137, + 293 + ], + "type": "image", + "image_path": "5102bc387ccfd1feb69e400d35b427ba3633987512a0d5bd5382d12f890bd2ed.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 141, + 256, + 217, + 293 + ], + "blocks": [ + { + "bbox": [ + 141, + 256, + 217, + 293 + ], + "lines": [ + { + "bbox": [ + 141, + 256, + 217, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 217, + 293 + ], + "type": "image", + "image_path": "d1ee28985945547b92e629f6c73ac5dd412b1a2e7b8adc4acc0f463b5efdc71d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 293, + 200, + 300 + ], + "lines": [ + { + "bbox": [ + 159, + 293, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 159, + 293, + 200, + 300 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 223, + 256, + 298, + 293 + ], + "blocks": [ + { + "bbox": [ + 223, + 256, + 298, + 293 + ], + "lines": [ + { + "bbox": [ + 223, + 256, + 298, + 293 + ], + "spans": [ + { + "bbox": [ + 223, + 256, + 298, + 293 + ], + "type": "image", + "image_path": "ff614c8f3bb02c66dc9b220227f512a8a8f62d0121dcea219c803ed67eb2cb2b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 369, + 108 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 369, + 108 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 369, + 108 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 369, + 108 + ], + "type": "image", + "image_path": "ae6a1c1e8d7890f605985e7c8c7ae79c2b082c7ce3963205b759a04754b0a18c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 373, + 71, + 424, + 108 + ], + "blocks": [ + { + "bbox": [ + 373, + 71, + 424, + 108 + ], + "lines": [ + { + "bbox": [ + 373, + 71, + 424, + 108 + ], + "spans": [ + { + "bbox": [ + 373, + 71, + 424, + 108 + ], + "type": "image", + "image_path": "4f1b36c7ceb4aa164ecfa481216f4e957b1d0baf68e8f6f0bde496a3fedd1f5a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 108, + 467, + 114 + ], + "lines": [ + { + "bbox": [ + 383, + 108, + 467, + 114 + ], + "spans": [ + { + "bbox": [ + 383, + 108, + 467, + 114 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 318, + 117, + 369, + 152 + ], + "blocks": [ + { + "bbox": [ + 318, + 117, + 369, + 152 + ], + "lines": [ + { + "bbox": [ + 318, + 117, + 369, + 152 + ], + "spans": [ + { + "bbox": [ + 318, + 117, + 369, + 152 + ], + "type": "image", + "image_path": "5af787bd1e651aeb6942d512bb902ad990971de211f7174729b04e8d90c075cd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 374, + 116, + 424, + 153 + ], + "blocks": [ + { + "bbox": [ + 374, + 116, + 424, + 153 + ], + "lines": [ + { + "bbox": [ + 374, + 116, + 424, + 153 + ], + "spans": [ + { + "bbox": [ + 374, + 116, + 424, + 153 + ], + "type": "image", + "image_path": "490b90633c61a91c1aa522829af3f8ef72f23d35977c614db54f2ceb0c3b6260.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 154, + 491, + 161 + ], + "lines": [ + { + "bbox": [ + 359, + 154, + 491, + 161 + ], + "spans": [ + { + "bbox": [ + 359, + 154, + 491, + 161 + ], + "type": "text", + "content": "Proxy reconstructions with no modulations" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 477, + 108 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 477, + 108 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 477, + 108 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 477, + 108 + ], + "type": "image", + "image_path": "d2380e34cbc0e12fce18d88eccb74ceae3d84017738f4adc66c116ec269c4c34.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 483, + 71, + 533, + 108 + ], + "blocks": [ + { + "bbox": [ + 483, + 71, + 533, + 108 + ], + "lines": [ + { + "bbox": [ + 483, + 71, + 533, + 108 + ], + "spans": [ + { + "bbox": [ + 483, + 71, + 533, + 108 + ], + "type": "image", + "image_path": "abafee8c49e3d8efb787059a909e2457441a33f022cadcd4fdc50ae1bb2d75a8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 318, + 163, + 369, + 200 + ], + "blocks": [ + { + "bbox": [ + 318, + 163, + 369, + 200 + ], + "lines": [ + { + "bbox": [ + 318, + 163, + 369, + 200 + ], + "spans": [ + { + "bbox": [ + 318, + 163, + 369, + 200 + ], + "type": "image", + "image_path": "48d4de392fb768b2887b55b34cc2bca452b886134584198785e83729bfd733c2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 374, + 163, + 424, + 201 + ], + "blocks": [ + { + "bbox": [ + 374, + 163, + 424, + 201 + ], + "lines": [ + { + "bbox": [ + 374, + 163, + 424, + 201 + ], + "spans": [ + { + "bbox": [ + 374, + 163, + 424, + 201 + ], + "type": "image", + "image_path": "1bd1eedbebac7de60e8e56739ea831b282de7bd7b755b93b460b6d7d979e60d4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 201, + 499, + 209 + ], + "lines": [ + { + "bbox": [ + 351, + 201, + 499, + 209 + ], + "spans": [ + { + "bbox": [ + 351, + 201, + 499, + 209 + ], + "type": "text", + "content": "Proxy reconstructions with random modulations" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 427, + 163, + 477, + 201 + ], + "blocks": [ + { + "bbox": [ + 427, + 163, + 477, + 201 + ], + "lines": [ + { + "bbox": [ + 427, + 163, + 477, + 201 + ], + "spans": [ + { + "bbox": [ + 427, + 163, + 477, + 201 + ], + "type": "image", + "image_path": "99aedd39cd4cf623c81d3843ad4244c93bee540e4f8618c57e0cfa79332c339d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 483, + 163, + 533, + 201 + ], + "blocks": [ + { + "bbox": [ + 483, + 163, + 533, + 201 + ], + "lines": [ + { + "bbox": [ + 483, + 163, + 533, + 201 + ], + "spans": [ + { + "bbox": [ + 483, + 163, + 533, + 201 + ], + "type": "image", + "image_path": "466bb3b2d837ca7cb59b9a29f5b7dc1cf0f4a76dffef0c20da7ebc132cb1ba4f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 319, + 210, + 369, + 247 + ], + "blocks": [ + { + "bbox": [ + 319, + 210, + 369, + 247 + ], + "lines": [ + { + "bbox": [ + 319, + 210, + 369, + 247 + ], + "spans": [ + { + "bbox": [ + 319, + 210, + 369, + 247 + ], + "type": "image", + "image_path": "aec3bc163077b361c53ee3e625baa4dcbd35eea14cd50a050b2a5d6d5ba34a82.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 247, + 499, + 255 + ], + "lines": [ + { + "bbox": [ + 351, + 247, + 499, + 255 + ], + "spans": [ + { + "bbox": [ + 351, + 247, + 499, + 255 + ], + "type": "text", + "content": "Proxy reconstructions with learned modulations" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 374, + 211, + 424, + 247 + ], + "blocks": [ + { + "bbox": [ + 374, + 211, + 424, + 247 + ], + "lines": [ + { + "bbox": [ + 374, + 211, + 424, + 247 + ], + "spans": [ + { + "bbox": [ + 374, + 211, + 424, + 247 + ], + "type": "image", + "image_path": "985a9ad8bd7e4fa9006b817352bde2ec0bc92f345c3eb163b7c86ae8d2b2c922.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 427, + 211, + 477, + 247 + ], + "blocks": [ + { + "bbox": [ + 427, + 211, + 477, + 247 + ], + "lines": [ + { + "bbox": [ + 427, + 211, + 477, + 247 + ], + "spans": [ + { + "bbox": [ + 427, + 211, + 477, + 247 + ], + "type": "image", + "image_path": "5c6dd77a2df95c671cfb79dbdc1e764551ccd140cbc7e70bfe9117e014a106d3.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 483, + 211, + 533, + 247 + ], + "blocks": [ + { + "bbox": [ + 483, + 211, + 533, + 247 + ], + "lines": [ + { + "bbox": [ + 483, + 211, + 533, + 247 + ], + "spans": [ + { + "bbox": [ + 483, + 211, + 533, + 247 + ], + "type": "image", + "image_path": "b9df6e123028dd705da029dcde65ab68eb86c9fb2830df4c786dfc9c69fdcaca.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 319, + 256, + 369, + 293 + ], + "blocks": [ + { + "bbox": [ + 319, + 256, + 369, + 293 + ], + "lines": [ + { + "bbox": [ + 319, + 256, + 369, + 293 + ], + "spans": [ + { + "bbox": [ + 319, + 256, + 369, + 293 + ], + "type": "image", + "image_path": "986f856fa960b339196195b5633521a6a9b7e16ab8916038979e390b0e2ef008.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 374, + 256, + 424, + 293 + ], + "blocks": [ + { + "bbox": [ + 374, + 256, + 424, + 293 + ], + "lines": [ + { + "bbox": [ + 374, + 256, + 424, + 293 + ], + "spans": [ + { + "bbox": [ + 374, + 256, + 424, + 293 + ], + "type": "image", + "image_path": "75902cf97aa51d39031adcf85e1add4f175b673b2628a5121d0ff79e5aaee632.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 293, + 446, + 300 + ], + "lines": [ + { + "bbox": [ + 404, + 293, + 446, + 300 + ], + "spans": [ + { + "bbox": [ + 404, + 293, + 446, + 300 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 427, + 256, + 477, + 293 + ], + "blocks": [ + { + "bbox": [ + 427, + 256, + 477, + 293 + ], + "lines": [ + { + "bbox": [ + 427, + 256, + 477, + 293 + ], + "spans": [ + { + "bbox": [ + 427, + 256, + 477, + 293 + ], + "type": "image", + "image_path": "eadc92a5c3c55b9cc5083e46762f2f05cfde3b4fe7ab931da9acd7a7b7b47767.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 483, + 256, + 533, + 293 + ], + "blocks": [ + { + "bbox": [ + 483, + 256, + 533, + 293 + ], + "lines": [ + { + "bbox": [ + 483, + 256, + 533, + 293 + ], + "spans": [ + { + "bbox": [ + 483, + 256, + 533, + 293 + ], + "type": "image", + "image_path": "d6f11c72e1e045c163da6f62efd2c824c3a9413d9fc549a6f495f7fc4595a875.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 62, + 379, + 279, + 491 + ], + "blocks": [ + { + "bbox": [ + 46, + 316, + 547, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 547, + 360 + ], + "type": "text", + "content": "Figure 5. Proxy Network Reconstruction on Physical Scattering. Experimental results of imaging objects through scattering media by using our proxy reconstruction network. The left columns are in-distribution adipose tissue slides (zoomed-in region labeled with red boxes); the right columns are out-of-distribution targets. Learned modulations yield superior imaging quality for both in-distribution and out-of-distribution scenes, with the former out-performing the latter." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 379, + 279, + 491 + ], + "lines": [ + { + "bbox": [ + 62, + 379, + 279, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 279, + 491 + ], + "type": "image", + "image_path": "368f466ffc48bb6881c0e2ac3a055e520f6ac42c36ad68692da3ffdf6ceda76b.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 502, + 287, + 559 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 559 + ], + "type": "text", + "content": "Figure 6. Optical Setup. A spatially incoherent light source illuminates the DMD and passes through thick scattering. The DMD displays the target images for training and testing. The SLM is placed onto the Fourier plane and projects learned patterns. The wavefront is modulated and subsequently imaged onto a camera." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 46, + 617, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 712 + ], + "type": "text", + "content": "Each mage is " + }, + { + "bbox": [ + 46, + 617, + 288, + 712 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 617, + 288, + 712 + ], + "type": "text", + "content": " pixels. Our dataset includes 1,000 sectional images of adipose tissues. 40 of these images were utilized for testing and the rest were used for training. To evaluate the generalizability across different human tissues, we also imaged sections of human stomach and glandular epithelium, trachea, and finger. Additionally, we tested with targets outside the domain of the training data, such as numbers and letters." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 306, + 380, + 432, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 432, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 432, + 393 + ], + "type": "text", + "content": "6. Experimental Results" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 399, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 520 + ], + "type": "text", + "content": "We employed the optical system depicted in Figure 6. Our learned modulations significantly enhance the imaging capabilities through scattering media. We tested on datasets from both similar and distinctly different distributions from our experimental training data. Our learned modulations consistently yield superior visual reconstruction results. Furthermore, we evaluated these modulations on both static and dynamic scenes. We show that our learned modulations significantly enhance reconstruction in both cases, which suggests a broad applicability of our approach." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 305, + 527, + 473, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 473, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 473, + 540 + ], + "type": "text", + "content": "6.1. Proxy Reconstruction Network" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "The proxy network, fine-tuned on 960 adipose tissue images through random scattering, was evaluated on a test set from the same tissue type, as shown in the left of Figure 5. Reconstructions driven by measurements of 16 learned SLM patterns preserved the original contrast and high-frequency details of the dyed sections. The close-ups show remarkable detail accuracy. By comparison, reconstructions trained on measurements of random or no modulation struggled, particularly in lower-intensity areas, yielding blurry outputs. Beyond achieving good reconstruction fidelity on objects similar to our training dataset, we extended our evaluation to vastly different data distributions, like numbers and letters. The results, as shown on the right of Figure 5, clearly indicate that the learned modulations yield bet" + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25281" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 69, + 149, + 106 + ], + "blocks": [ + { + "bbox": [ + 102, + 69, + 149, + 106 + ], + "lines": [ + { + "bbox": [ + 102, + 69, + 149, + 106 + ], + "spans": [ + { + "bbox": [ + 102, + 69, + 149, + 106 + ], + "type": "image", + "image_path": "afb40c5c8aa46924fa4adef9c37a095b00dc2eb993d3696ef89ffd0e33708bd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 150, + 70, + 197, + 106 + ], + "blocks": [ + { + "bbox": [ + 150, + 70, + 197, + 106 + ], + "lines": [ + { + "bbox": [ + 150, + 70, + 197, + 106 + ], + "spans": [ + { + "bbox": [ + 150, + 70, + 197, + 106 + ], + "type": "image", + "image_path": "dbdbcdccf6a040f07ba99f2b32deb2b255721c6ee5f5e14f2963428003b90221.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 106, + 237, + 112 + ], + "lines": [ + { + "bbox": [ + 157, + 106, + 237, + 112 + ], + "spans": [ + { + "bbox": [ + 157, + 106, + 237, + 112 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 197, + 70, + 245, + 106 + ], + "blocks": [ + { + "bbox": [ + 197, + 70, + 245, + 106 + ], + "lines": [ + { + "bbox": [ + 197, + 70, + 245, + 106 + ], + "spans": [ + { + "bbox": [ + 197, + 70, + 245, + 106 + ], + "type": "image", + "image_path": "e5ec909d020f9af775c4f060389c9b9e0f23a4af85d22c75de3b9362fee2152d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 70, + 293, + 106 + ], + "blocks": [ + { + "bbox": [ + 246, + 70, + 293, + 106 + ], + "lines": [ + { + "bbox": [ + 246, + 70, + 293, + 106 + ], + "spans": [ + { + "bbox": [ + 246, + 70, + 293, + 106 + ], + "type": "image", + "image_path": "677f83b23404db4457714b2069b4195407d322f26be3632688bb62706f54b9f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 102, + 114, + 149, + 149 + ], + "blocks": [ + { + "bbox": [ + 102, + 114, + 149, + 149 + ], + "lines": [ + { + "bbox": [ + 102, + 114, + 149, + 149 + ], + "spans": [ + { + "bbox": [ + 102, + 114, + 149, + 149 + ], + "type": "image", + "image_path": "88f52109252054f3ea82412818902bb7390773bc866df84d9aa2f1cd2f985628.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 150, + 260, + 156 + ], + "lines": [ + { + "bbox": [ + 135, + 150, + 260, + 156 + ], + "spans": [ + { + "bbox": [ + 135, + 150, + 260, + 156 + ], + "type": "text", + "content": "Reconstructions with random modulations" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 150, + 114, + 197, + 149 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 197, + 149 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 197, + 149 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 197, + 149 + ], + "type": "image", + "image_path": "685d1e10ecb46f8061a6ef648a991a0200732dec64bb1b70f71b078b9f6c0db2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 197, + 114, + 245, + 149 + ], + "blocks": [ + { + "bbox": [ + 197, + 114, + 245, + 149 + ], + "lines": [ + { + "bbox": [ + 197, + 114, + 245, + 149 + ], + "spans": [ + { + "bbox": [ + 197, + 114, + 245, + 149 + ], + "type": "image", + "image_path": "ebc2a86cbdde8bd93fa091212e4a008b06d39b315b90e9861ea501a82d940532.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 246, + 114, + 293, + 149 + ], + "blocks": [ + { + "bbox": [ + 246, + 114, + 293, + 149 + ], + "lines": [ + { + "bbox": [ + 246, + 114, + 293, + 149 + ], + "spans": [ + { + "bbox": [ + 246, + 114, + 293, + 149 + ], + "type": "image", + "image_path": "59d4e481cc39a2cacb930b594a70258db07a37de09efdf5c4eb4698d4c98c7af.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 102, + 158, + 149, + 194 + ], + "blocks": [ + { + "bbox": [ + 102, + 158, + 149, + 194 + ], + "lines": [ + { + "bbox": [ + 102, + 158, + 149, + 194 + ], + "spans": [ + { + "bbox": [ + 102, + 158, + 149, + 194 + ], + "type": "image", + "image_path": "81dbc391823d09fe51174ee9302c1c9eb645042fb1edc57c321a10962628141e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 194, + 260, + 201 + ], + "lines": [ + { + "bbox": [ + 136, + 194, + 260, + 201 + ], + "spans": [ + { + "bbox": [ + 136, + 194, + 260, + 201 + ], + "type": "text", + "content": "Reconstructions with learned modulations" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 150, + 158, + 197, + 194 + ], + "blocks": [ + { + "bbox": [ + 150, + 158, + 197, + 194 + ], + "lines": [ + { + "bbox": [ + 150, + 158, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 150, + 158, + 197, + 194 + ], + "type": "image", + "image_path": "d8959e7e1197f1ab2ebd39e666329379b6b5cda0927e7842ceda5e69fd4219f4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 197, + 158, + 245, + 194 + ], + "blocks": [ + { + "bbox": [ + 197, + 158, + 245, + 194 + ], + "lines": [ + { + "bbox": [ + 197, + 158, + 245, + 194 + ], + "spans": [ + { + "bbox": [ + 197, + 158, + 245, + 194 + ], + "type": "image", + "image_path": "b100a7ce931ab83f45f9714728e9d4754c6782a594fc99bcc2f3960e85d2aeeb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 246, + 158, + 293, + 194 + ], + "blocks": [ + { + "bbox": [ + 246, + 158, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 246, + 158, + 293, + 194 + ], + "spans": [ + { + "bbox": [ + 246, + 158, + 293, + 194 + ], + "type": "image", + "image_path": "d1ae4d57b0d99422323689dcb71fc78d08030e32974a254d1fc5bbe41e8a2e40.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 102, + 203, + 149, + 239 + ], + "blocks": [ + { + "bbox": [ + 102, + 203, + 149, + 239 + ], + "lines": [ + { + "bbox": [ + 102, + 203, + 149, + 239 + ], + "spans": [ + { + "bbox": [ + 102, + 203, + 149, + 239 + ], + "type": "image", + "image_path": "459f212daa3216bf97e44e9a3c02aeb086265fce8a51cdab05e6cd5a6c2b4b1d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 150, + 203, + 197, + 239 + ], + "blocks": [ + { + "bbox": [ + 150, + 203, + 197, + 239 + ], + "lines": [ + { + "bbox": [ + 150, + 203, + 197, + 239 + ], + "spans": [ + { + "bbox": [ + 150, + 203, + 197, + 239 + ], + "type": "image", + "image_path": "f09e314b5549379163781f9c03487c4e0707557bb5f5bd853dddb818eab382cb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 239, + 218, + 246 + ], + "lines": [ + { + "bbox": [ + 178, + 239, + 218, + 246 + ], + "spans": [ + { + "bbox": [ + 178, + 239, + 218, + 246 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 197, + 203, + 245, + 239 + ], + "blocks": [ + { + "bbox": [ + 197, + 203, + 245, + 239 + ], + "lines": [ + { + "bbox": [ + 197, + 203, + 245, + 239 + ], + "spans": [ + { + "bbox": [ + 197, + 203, + 245, + 239 + ], + "type": "image", + "image_path": "83994e5c0ec333b4073d6c541913d769097b7c880663ca1d0f1ff831d74e993c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 246, + 203, + 293, + 239 + ], + "blocks": [ + { + "bbox": [ + 246, + 203, + 293, + 239 + ], + "lines": [ + { + "bbox": [ + 246, + 203, + 293, + 239 + ], + "spans": [ + { + "bbox": [ + 246, + 203, + 293, + 239 + ], + "type": "image", + "image_path": "cf1a1f53e06f9600a11991bd0bd3ab3b5a51ec7ecc8ff6618552c608adfcb847.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 350, + 106 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 350, + 106 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 350, + 106 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 350, + 106 + ], + "type": "image", + "image_path": "d06de3fcda707bab0908c51b786bdf60a10fa5c6a5a407760af2dc6f3b30da85.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 351, + 70, + 398, + 106 + ], + "blocks": [ + { + "bbox": [ + 351, + 70, + 398, + 106 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 398, + 106 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 398, + 106 + ], + "type": "image", + "image_path": "80047a784f6a8ad7ac5a5f6c131068c08faf85484025ea771690641e873f8f56.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 106, + 438, + 112 + ], + "lines": [ + { + "bbox": [ + 358, + 106, + 438, + 112 + ], + "spans": [ + { + "bbox": [ + 358, + 106, + 438, + 112 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 399, + 70, + 446, + 106 + ], + "blocks": [ + { + "bbox": [ + 399, + 70, + 446, + 106 + ], + "lines": [ + { + "bbox": [ + 399, + 70, + 446, + 106 + ], + "spans": [ + { + "bbox": [ + 399, + 70, + 446, + 106 + ], + "type": "image", + "image_path": "8b8259b8e9594a449e2d2c67bbbd71b3ac1cc86e7cc2776684b20d51846d36d7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 447, + 70, + 494, + 106 + ], + "blocks": [ + { + "bbox": [ + 447, + 70, + 494, + 106 + ], + "lines": [ + { + "bbox": [ + 447, + 70, + 494, + 106 + ], + "spans": [ + { + "bbox": [ + 447, + 70, + 494, + 106 + ], + "type": "image", + "image_path": "805ea4f44c7eaec65598ea262307186eb035c3b6c7efbfc9cbc65043e639a9d0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 350, + 150 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 350, + 150 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 350, + 150 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 350, + 150 + ], + "type": "image", + "image_path": "900a0a3fd09e77f5b8550753df53f335732515fc8811d6e67fe298e048bd8275.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 150, + 460, + 156 + ], + "lines": [ + { + "bbox": [ + 337, + 150, + 460, + 156 + ], + "spans": [ + { + "bbox": [ + 337, + 150, + 460, + 156 + ], + "type": "text", + "content": "Reconstructions with random modulations" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 351, + 114, + 398, + 150 + ], + "blocks": [ + { + "bbox": [ + 351, + 114, + 398, + 150 + ], + "lines": [ + { + "bbox": [ + 351, + 114, + 398, + 150 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 398, + 150 + ], + "type": "image", + "image_path": "d85aee4487be29256b2d8c5985b0ac6404a2197ad76bf305688cd8413398f728.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 446, + 150 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 446, + 150 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 446, + 150 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 446, + 150 + ], + "type": "image", + "image_path": "b22cb5527af65eb55f7636ec03d05a3fedbd0df023a6bd346ade4e8b2e046598.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 447, + 114, + 494, + 150 + ], + "blocks": [ + { + "bbox": [ + 447, + 114, + 494, + 150 + ], + "lines": [ + { + "bbox": [ + 447, + 114, + 494, + 150 + ], + "spans": [ + { + "bbox": [ + 447, + 114, + 494, + 150 + ], + "type": "image", + "image_path": "6ad583a53f7dfcfd2b902248894ac1943f1faaf12ec1d06e6eb4a98bf1e8f65c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 304, + 158, + 350, + 194 + ], + "blocks": [ + { + "bbox": [ + 304, + 158, + 350, + 194 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 350, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 350, + 194 + ], + "type": "image", + "image_path": "b44daed79ff0f7c13cdca1bf51e374f783eb258fdd5a5ba3a095b71ec07adaa9.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 194, + 460, + 201 + ], + "lines": [ + { + "bbox": [ + 337, + 194, + 460, + 201 + ], + "spans": [ + { + "bbox": [ + 337, + 194, + 460, + 201 + ], + "type": "text", + "content": "Reconstructions with learned modulations" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 351, + 158, + 398, + 194 + ], + "blocks": [ + { + "bbox": [ + 351, + 158, + 398, + 194 + ], + "lines": [ + { + "bbox": [ + 351, + 158, + 398, + 194 + ], + "spans": [ + { + "bbox": [ + 351, + 158, + 398, + 194 + ], + "type": "image", + "image_path": "cfdd2253d7bd23affd6347131956487e034518993de674471dce2567a5d9a2c3.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 399, + 158, + 446, + 194 + ], + "blocks": [ + { + "bbox": [ + 399, + 158, + 446, + 194 + ], + "lines": [ + { + "bbox": [ + 399, + 158, + 446, + 194 + ], + "spans": [ + { + "bbox": [ + 399, + 158, + 446, + 194 + ], + "type": "image", + "image_path": "3fb49de8fd2ce4a659af6c1c0a33e874fb426ea12f409cc4ba0df3faba889d6b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 447, + 158, + 494, + 194 + ], + "blocks": [ + { + "bbox": [ + 447, + 158, + 494, + 194 + ], + "lines": [ + { + "bbox": [ + 447, + 158, + 494, + 194 + ], + "spans": [ + { + "bbox": [ + 447, + 158, + 494, + 194 + ], + "type": "image", + "image_path": "7c2b839b53c3bbd24ad825baa0f2f05cc0847ff18b79a0323ab4207bfcf787a1.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 304, + 203, + 350, + 239 + ], + "blocks": [ + { + "bbox": [ + 304, + 203, + 350, + 239 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 350, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 350, + 239 + ], + "type": "image", + "image_path": "a71cc18bce7e9b3c2734761990901eff097d4c9ef411b6c4cfdcd4ac0e4f3f84.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 351, + 203, + 398, + 239 + ], + "blocks": [ + { + "bbox": [ + 351, + 203, + 398, + 239 + ], + "lines": [ + { + "bbox": [ + 351, + 203, + 398, + 239 + ], + "spans": [ + { + "bbox": [ + 351, + 203, + 398, + 239 + ], + "type": "image", + "image_path": "97082a8b6a172ee907c63785d1cafd76ec07b2f1ffcb4603b224238d20df8ca6.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 239, + 418, + 247 + ], + "lines": [ + { + "bbox": [ + 378, + 239, + 418, + 247 + ], + "spans": [ + { + "bbox": [ + 378, + 239, + 418, + 247 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 399, + 203, + 446, + 239 + ], + "blocks": [ + { + "bbox": [ + 399, + 203, + 446, + 239 + ], + "lines": [ + { + "bbox": [ + 399, + 203, + 446, + 239 + ], + "spans": [ + { + "bbox": [ + 399, + 203, + 446, + 239 + ], + "type": "image", + "image_path": "78198196e6a652de8ff787cc51d1ec06bc55d84f0417c93aa4f56562527722c8.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 447, + 203, + 494, + 239 + ], + "blocks": [ + { + "bbox": [ + 447, + 203, + 494, + 239 + ], + "lines": [ + { + "bbox": [ + 447, + 203, + 494, + 239 + ], + "spans": [ + { + "bbox": [ + 447, + 203, + 494, + 239 + ], + "type": "image", + "image_path": "3a8cdfc74378419aa2f61616121353414daa0ca0dd633a20c22c30926d1c11ee.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 121, + 317, + 171, + 354 + ], + "blocks": [ + { + "bbox": [ + 141, + 310, + 157, + 317 + ], + "lines": [ + { + "bbox": [ + 141, + 310, + 157, + 317 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 157, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{t} = 1" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 121, + 317, + 171, + 354 + ], + "lines": [ + { + "bbox": [ + 121, + 317, + 171, + 354 + ], + "spans": [ + { + "bbox": [ + 121, + 317, + 171, + 354 + ], + "type": "image", + "image_path": "2347cf0cc00e9110c516fe7114238ce0c5fe57be53bf01b965d8c2893f0295d3.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 355, + 246, + 361 + ], + "lines": [ + { + "bbox": [ + 162, + 355, + 246, + 361 + ], + "spans": [ + { + "bbox": [ + 162, + 355, + 246, + 361 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 179, + 317, + 228, + 354 + ], + "blocks": [ + { + "bbox": [ + 194, + 310, + 214, + 317 + ], + "lines": [ + { + "bbox": [ + 194, + 310, + 214, + 317 + ], + "spans": [ + { + "bbox": [ + 194, + 310, + 214, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{t} = 24" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 317, + 228, + 354 + ], + "lines": [ + { + "bbox": [ + 179, + 317, + 228, + 354 + ], + "spans": [ + { + "bbox": [ + 179, + 317, + 228, + 354 + ], + "type": "image", + "image_path": "e23e5819f8af33754f6da9a94d589d3642447fa5de361a8b5e0fd199cfda7f7b.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 238, + 317, + 287, + 354 + ], + "blocks": [ + { + "bbox": [ + 252, + 310, + 271, + 316 + ], + "lines": [ + { + "bbox": [ + 252, + 310, + 271, + 316 + ], + "spans": [ + { + "bbox": [ + 252, + 310, + 271, + 316 + ], + "type": "text", + "content": "t=48" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 317, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 238, + 317, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 238, + 317, + 287, + 354 + ], + "type": "image", + "image_path": "2a1d637fc435272aa3988bc650a541b33ed575127b9071ff385bdfb2a0a4cac1.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 122, + 364, + 171, + 401 + ], + "blocks": [ + { + "bbox": [ + 122, + 364, + 171, + 401 + ], + "lines": [ + { + "bbox": [ + 122, + 364, + 171, + 401 + ], + "spans": [ + { + "bbox": [ + 122, + 364, + 171, + 401 + ], + "type": "image", + "image_path": "898cd7280dbf47e87506ebd82496d00e0d7811947692d7028155b94109aac1d8.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 402, + 269, + 408 + ], + "lines": [ + { + "bbox": [ + 138, + 402, + 269, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 269, + 408 + ], + "type": "text", + "content": "Reconstructions with random modulations" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 179, + 410, + 228, + 447 + ], + "blocks": [ + { + "bbox": [ + 179, + 410, + 228, + 447 + ], + "lines": [ + { + "bbox": [ + 179, + 410, + 228, + 447 + ], + "spans": [ + { + "bbox": [ + 179, + 410, + 228, + 447 + ], + "type": "image", + "image_path": "eca66562396695e35a5c1a745418ef9a4e34f7270138d648164c703501429961.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 448, + 269, + 454 + ], + "lines": [ + { + "bbox": [ + 138, + 448, + 269, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 269, + 454 + ], + "type": "text", + "content": "Reconstructions with learned modulations" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 237, + 410, + 287, + 447 + ], + "blocks": [ + { + "bbox": [ + 237, + 410, + 287, + 447 + ], + "lines": [ + { + "bbox": [ + 237, + 410, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 237, + 410, + 287, + 447 + ], + "type": "image", + "image_path": "32b7b0085eccf72f4f45d1ac0b35b5ea88be5485dee7e9ba1305eaaa176870ef.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 121, + 456, + 171, + 495 + ], + "blocks": [ + { + "bbox": [ + 121, + 456, + 171, + 495 + ], + "lines": [ + { + "bbox": [ + 121, + 456, + 171, + 495 + ], + "spans": [ + { + "bbox": [ + 121, + 456, + 171, + 495 + ], + "type": "image", + "image_path": "40e91686a32669be00b49a90291cca80527730f79f722a1ef070578b062713f5.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 517, + 545, + 561 + ], + "lines": [ + { + "bbox": [ + 46, + 517, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 545, + 561 + ], + "type": "text", + "content": "Figure 8. Unsupervised Reconstruction on Physical Scattering with Dynamic Scenes. Experimental results of imaging an opening camera aperture through severe aberrations (left) and imaging non-regional motions (number 6 rotating clockwise and number 5 translating to the right) through severe aberrations. We use the unsupervised iterative reconstruction approach [8]. Learned modulations highly improve the quality of the reconstructed dynamic scene." + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 179, + 456, + 228, + 495 + ], + "blocks": [ + { + "bbox": [ + 179, + 456, + 228, + 495 + ], + "lines": [ + { + "bbox": [ + 179, + 456, + 228, + 495 + ], + "spans": [ + { + "bbox": [ + 179, + 456, + 228, + 495 + ], + "type": "image", + "image_path": "768429f557bd6682e1a10496dffee32d2d4deebac11f2f57b8f0dba019f65dc9.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 495, + 225, + 502 + ], + "lines": [ + { + "bbox": [ + 182, + 495, + 225, + 502 + ], + "spans": [ + { + "bbox": [ + 182, + 495, + 225, + 502 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 237, + 456, + 287, + 495 + ], + "blocks": [ + { + "bbox": [ + 237, + 456, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 237, + 456, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 237, + 456, + 287, + 495 + ], + "type": "image", + "image_path": "1d42a20717d879276fc69fc5bec855e2cf7672ddb16e6800911e5a0fcdfd8845.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 304, + 311, + 336, + 355 + ], + "blocks": [ + { + "bbox": [ + 46, + 261, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 545, + 294 + ], + "type": "text", + "content": "Figure 7. Unsupervised Reconstruction on Physical Scattering with Static Scenes. Experimental results of imaging different static targets through scattering media using learned wavefront modulations with unsupervised iterative optimization [8]. Learned modulations enhance reconstruction quality." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 311, + 336, + 355 + ], + "lines": [ + { + "bbox": [ + 304, + 311, + 336, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 336, + 355 + ], + "type": "image", + "image_path": "995e62438ef49fbb381dd2f429efe7f390dbd3fa60892ac2ce2d89e90947b56f.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 361, + 317, + 410, + 354 + ], + "blocks": [ + { + "bbox": [ + 376, + 311, + 395, + 317 + ], + "lines": [ + { + "bbox": [ + 376, + 311, + 395, + 317 + ], + "spans": [ + { + "bbox": [ + 376, + 311, + 395, + 317 + ], + "type": "text", + "content": "t=24" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 317, + 410, + 354 + ], + "lines": [ + { + "bbox": [ + 361, + 317, + 410, + 354 + ], + "spans": [ + { + "bbox": [ + 361, + 317, + 410, + 354 + ], + "type": "image", + "image_path": "e44a2e8ea23e66318ec015d8da5a12d9d41378030507c2f520b6e1e5f29147c8.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 355, + 429, + 361 + ], + "lines": [ + { + "bbox": [ + 345, + 355, + 429, + 361 + ], + "spans": [ + { + "bbox": [ + 345, + 355, + 429, + 361 + ], + "type": "text", + "content": "Uncorrected measurements" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 419, + 317, + 468, + 354 + ], + "blocks": [ + { + "bbox": [ + 431, + 311, + 450, + 317 + ], + "lines": [ + { + "bbox": [ + 431, + 311, + 450, + 317 + ], + "spans": [ + { + "bbox": [ + 431, + 311, + 450, + 317 + ], + "type": "text", + "content": "t=48" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 317, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 419, + 317, + 468, + 354 + ], + "spans": [ + { + "bbox": [ + 419, + 317, + 468, + 354 + ], + "type": "image", + "image_path": "b18b1fe9d0af0117ef0b20f6fc02f2cf7844c81a1586e940f4ad9a0e366edded.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 304, + 364, + 353, + 400 + ], + "blocks": [ + { + "bbox": [ + 304, + 364, + 353, + 400 + ], + "lines": [ + { + "bbox": [ + 304, + 364, + 353, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 353, + 400 + ], + "type": "image", + "image_path": "62ac0fb1a45541ab57325828e9e7528b25312aa67de7f2bb9a8a4049fd98be4a.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 401, + 452, + 408 + ], + "lines": [ + { + "bbox": [ + 321, + 401, + 452, + 408 + ], + "spans": [ + { + "bbox": [ + 321, + 401, + 452, + 408 + ], + "type": "text", + "content": "Reconstructions with random modulations" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 303, + 410, + 353, + 447 + ], + "blocks": [ + { + "bbox": [ + 303, + 410, + 353, + 447 + ], + "lines": [ + { + "bbox": [ + 303, + 410, + 353, + 447 + ], + "spans": [ + { + "bbox": [ + 303, + 410, + 353, + 447 + ], + "type": "image", + "image_path": "ee77cc7b512e9e3fda6b3faa34c010b6c9f264082e45e2fa5b98de6af4e2fb85.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 448, + 452, + 454 + ], + "lines": [ + { + "bbox": [ + 321, + 448, + 452, + 454 + ], + "spans": [ + { + "bbox": [ + 321, + 448, + 452, + 454 + ], + "type": "text", + "content": "Reconstructions with learned modulations" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 361, + 410, + 410, + 447 + ], + "blocks": [ + { + "bbox": [ + 361, + 410, + 410, + 447 + ], + "lines": [ + { + "bbox": [ + 361, + 410, + 410, + 447 + ], + "spans": [ + { + "bbox": [ + 361, + 410, + 410, + 447 + ], + "type": "image", + "image_path": "841e13f6c5cbc8eeda4eb41e9a457686b6dcfca959c301b8df6423552c59a0b1.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 419, + 410, + 468, + 447 + ], + "blocks": [ + { + "bbox": [ + 419, + 410, + 468, + 447 + ], + "lines": [ + { + "bbox": [ + 419, + 410, + 468, + 447 + ], + "spans": [ + { + "bbox": [ + 419, + 410, + 468, + 447 + ], + "type": "image", + "image_path": "55fc6a1e498b0883cf57f60195fedce4231db92d95a33e6f2402acabd54f3456.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 303, + 456, + 353, + 495 + ], + "blocks": [ + { + "bbox": [ + 303, + 456, + 353, + 495 + ], + "lines": [ + { + "bbox": [ + 303, + 456, + 353, + 495 + ], + "spans": [ + { + "bbox": [ + 303, + 456, + 353, + 495 + ], + "type": "image", + "image_path": "b9d760e1e9631c5a53c581134b33481a139e43ab85b6a5a8cedd2045c291e9f8.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 364, + 456, + 410, + 495 + ], + "blocks": [ + { + "bbox": [ + 364, + 456, + 410, + 495 + ], + "lines": [ + { + "bbox": [ + 364, + 456, + 410, + 495 + ], + "spans": [ + { + "bbox": [ + 364, + 456, + 410, + 495 + ], + "type": "image", + "image_path": "892d8c8b2ea90da5b1bc70f7f45b8ded2b9c9d6a9f42ef77a33d0ff1db361b8c.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 495, + 408, + 502 + ], + "lines": [ + { + "bbox": [ + 365, + 495, + 408, + 502 + ], + "spans": [ + { + "bbox": [ + 365, + 495, + 408, + 502 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 419, + 456, + 468, + 495 + ], + "blocks": [ + { + "bbox": [ + 419, + 456, + 468, + 495 + ], + "lines": [ + { + "bbox": [ + 419, + 456, + 468, + 495 + ], + "spans": [ + { + "bbox": [ + 419, + 456, + 468, + 495 + ], + "type": "image", + "image_path": "e3cda9da38125a9b7d00e45e8105753ee730200d3502775d8be4358d676c69cd.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "type": "text", + "content": "ter reconstruction quality, particularly in the USAF target and stars. PSNR and SSIM are computed in Tables 2 and 3. Among the table cells, \"None\" means that the measurements for each target only comprised the unmodulated image (no modulation patterns are applied); \"Random\" means that the modulation patterns are randomly sampled from the same distribution as the aberrations." + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 306, + 582, + 493, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 493, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 493, + 595 + ], + "type": "text", + "content": "6.2. Unsupervised Iterative Approaches" + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "While the proxy network " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " exhibits considerable effectiveness in real-world experiments, its performance remains unsatisfactory for out-of-distribution data, which is expected. Nevertheless, our learned modulations are designed to further augment generalizable, untrained reconstruction algorithms based on iterative optimizations. To validate this strategy, we deploy our learned modulations to a recently developed unsupervised iterative approach for imaging through scattering, NeuWS [8], in our benchmark study." + } + ] + } + ], + "index": 76 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25282" + } + ] + } + ], + "index": 77 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 65, + 288, + 156 + ], + "blocks": [ + { + "bbox": [ + 47, + 65, + 288, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 65, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 65, + 288, + 156 + ], + "type": "table", + "html": "
MethodData typeModulation
NoneRandomLearned
ProxyTissue16.5317.5719.06
ProxyOut-of-dist.9.349.9010.71
Iterative [8]StaticN/A11.2614.61
Iterative [8]DynamicN/A8.9012.89
", + "image_path": "9f423b72b6403d05b9d206bf9bdac0ff44c32dfbb89166a41b07e1062d2cc876.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 283, + 287, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 281 + ], + "type": "text", + "content": "Table 2. PSNR of Experimental Results. For our jointly trained feed-forward proxy reconstruction network (\"proxy\"), we tested on 40 tissue samples and 8 out-of-distribution scenes, all of which are static. For the iterative method [8], we tested on the same 8 out-of-distribution static scenes. We also tested [8] on 2 dynamic scenes, each with 48 frames. The iterative method relies on multiple wavefront modulations and therefore cannot recover objects with a single measurement, hence the \"N/A\". Compared against random modulations or no modulations, our learned modulations lead to better reconstruction performance for both the proxy network and the unsupervised iterative approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 283, + 287, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 283, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 287, + 373 + ], + "type": "table", + "html": "
MethodData typeModulations
NoneRandomLearned
ProxyTissue0.440.480.58
ProxyOut-of-dist.0.290.300.32
Iterative [8]StaticN/A0.210.38
Iterative [8]DynamicN/A0.230.33
", + "image_path": "e9f9fdcc5bcc3d3d167b5f2939d4b847d180886e2ee654c1d91d03caa249aafa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 399 + ], + "type": "text", + "content": "Table 3. SSIM of Experimental Results. Same as Table 2 but showing SSIM. Our learned modulations performs better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "text", + "content": "Figure 7 shows the reconstruction performance of NeuWS [8] on 8 static objects, each using only 16 measurements. The superiority of using learned modulations is visually evident in the results. We excluded experiments with no modulations here, due to the algorithm's inability to recover objects using a single unmodulated measurement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "type": "text", + "content": "For dynamic scenes, each scene involves 48 frames captured by cycling through the 16 SLM patterns. Figure 8 demonstrates the reconstruction results on dynamic scenes: the result on the left illustrates a camera aperture gradually enlarging; the result on the right shows two different motions to solve the non-regional dynamic reconstruction (the number 6 on the left rotates clockwise at " + }, + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "type": "inline_equation", + "content": "0.5^{\\circ}" + }, + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "type": "text", + "content": " per frame, while the number 5 translates from the center to the right). Quantitative metrics on the dynamic results are included in Table 2 and 3. Full video results are in the project webpage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 620, + 190, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 190, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 190, + 634 + ], + "type": "text", + "content": "7. Discussion & Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Role of the Proxy Reconstruction Network. The proxy network not only aims to reconstruct the images; its performance also indicates how good the wavefront modulations are for the restoration task. Therefore, joint differentiable optimization of both the proxy network and the modulations works in our favor, and the experimental results indeed in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "dicate that this is an effective way of learning useful wavefront modulations. Without this proxy network, we would have to somehow connect the modulation update with the performance of an iterative optimization method, which is impractical due to time and memory constraints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 156, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 264 + ], + "type": "text", + "content": "Regularization Perspective. To recover missing signal frequencies due to scattering, most existing methods rely on prior information of the scene, either hand-crafted or data-driven. This work rethinks the strategy of incorporating prior knowledge: can we impose a prior over wavefront modulations? The goal here is to let the resulting imaging system capture more information about the target scene through scattering. Our results validate the effectiveness of learning a prior over the wavefront modulation domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 286, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 547, + 443 + ], + "type": "text", + "content": "Applicability to Other Vision Tasks. The end-to-end learning approach can potentially benefit other vision tasks through scattering media, using a task-specific proxy network. For instance, if we were to do object detection or semantic segmentation through scattering media, we can replace the proxy reconstruction U-Net with a network tailored for these two tasks, and perform end-to-end training in the same way as before. Note that we validated our proposed method on a monochromatic imaging system, which is prevalent in medical and scientific imaging. Extending it to broadband imaging tasks, e.g., outdoor navigation, requires the optimization of wavelength-dependent modulation patterns, which we leave for future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 465, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 547, + 586 + ], + "type": "text", + "content": "Conclusion. We proposed a robust end-to-end learning framework integrating the optical scattering model with a proxy image reconstruction network. The learned wavefront modulations can both work with the proxy reconstruction network and augment a generalizable, optimization-based algorithm. We conducted intensive experiments to validate our approach. This work shows the synergy of advanced wavefront modulation techniques with cutting-edge machine learning methods, representing a significant leap forward in computer vision and optical imaging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 597, + 400, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 400, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 400, + 610 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "This work was supported in part by AFOSR Young Investigator Program award no. FA9550-22-1-0208, ONR award no. N00014-23-1-2752, seed grant from SAAB, Inc., a seed grant from the UMD Brain and Behavior Institute, NSF Expeditions award no. CCF-1730574, ONR Scattering media award no. N00014-23-1-2714, and NSF PaThs up award no. EEC-1648451. We thank Sachin Shah for helpful discussions and Kevin F. Kelly for providing the DMD." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25283" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jesse K Adams, Dong Yan, Jimin Wu, Vivek Boominathan, Sibo Gao, Alex V Rodriguez, Soonyoung Kim, Jennifer Carns, Rebecca Richards-Kortum, Caleb Kemere, et al. In vivo lensless microscopy via a phase mask generating diffraction patterns with high-contrast contours. Nature Biomedical Engineering, 6(5):617-628, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Codruta Orniana Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1798-1805, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "text", + "content": "[3] Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase Retrieval Via Wirtinger Flow: Theory and Algorithms. IEEE Transactions on Information Theory, 61(4):1985-2007, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 290 + ], + "type": "text", + "content": "[4] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "text", + "content": "[5] Dongdong Chen, Mike Davies, Matthias J Ehrhardt, Carola-Bibiane Schonlieb, Ferdia Sherry, and Julián Tachella. Imaging with equivariant deep learning: From unrolled network design to fully unsupervised learning. IEEE Signal Processing Magazine, 40(1):134-147, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "type": "text", + "content": "[6] Wei-Yu Chen, Matthew O'Toole, Aswin C Sankaranarayanan, and Anat Levin. Enhancing Speckle Statistics for Imaging Inside Scattering Media. Optica, 9(12):1408-1416, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "type": "text", + "content": "[7] Santiago Echeverri-Chacón, René Restrepo, Carlos Cuartas-Vélez, and Néstor Uribe-Patarroyo. Vortex-enhanced coherent-illumination phase diversity for phase retrieval in coherent imaging systems. Optics letters, 41(8):1817-1820, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 511 + ], + "type": "text", + "content": "[8] Brandon Y Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K Sharma, Ashok Veeraraghavan, and Christopher A Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "text", + "content": "[9] Xueyang Fu, Jia-Bin Huang, Delu Zeng, Yue Huang, Xing-hao Ding, and John William Paisley. Removing rain from single images via a deep detail network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1715-1723, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 633 + ], + "type": "text", + "content": "[10] Sylvain Gigan, Ori Katz, Hilton B de Aguiar, Esben Ravn Andresen, Alexandre Aubry, Jacopo Bertolotti, Emmanuel Bossy, Dorian Bouchet, Joshua Brake, Sophie Brasselet, et al. Roadmap on Wavefront Shaping and Deep Imaging in Complex Media. Journal of Physics: Photonics, 4(4): 042501, 2022. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "text", + "content": "[11] Robert A Gonsalves. Phase retrieval and diversity in adaptive optics. Optical Engineering, 21(5):829-832, 1982. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 679 + ], + "type": "text", + "content": "[12] Joseph W. Goodman. Introduction to Fourier Optics. Roberts & Co., Englewood, Colo, 3rd ed. edition, 2005. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Xiaowen Hu, Jian Zhao, Jose Enrique Antonio-Lopez, Stefan Gausmann, Rodrigo Amezcua Correa, and Axel Schulzgen. Adaptive inverse mapping: a model-free" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "semi-supervised learning approach towards robust imaging through dynamic scattering media. Optics Express, 31(9): 14343-14357, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "text", + "content": "[14] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[15] Iksung Kang, Qinrong Zhang, Stella X Yu, and Na Ji. Coordinate-based neural representations for computational adaptive optics in widefield microscopy. arXiv preprint arXiv:2307.03812, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "type": "text", + "content": "[16] Richard L Kendrick, Daniel S Acton, and AL Duncan. Phase-diversity wave-front sensor for imaging systems. Applied Optics, 33(27):6533-6546, 1994. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 263 + ], + "type": "text", + "content": "[17] Peter Kner. Phase diversity for three-dimensional imaging. JOSA A, 30(10):1980-1987, 2013. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "type": "text", + "content": "[18] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "text", + "content": "[19] Juhyun Lee, Jinsoo Jeong, Jaebum Cho, Dongheon Yoo, Byounghyo Lee, and Byoungho Lee. Deep neural network for multi-depth hologram generation and its training strategy. Optics Express, 28(18):27137-27154, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 344, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 386 + ], + "type": "text", + "content": "[20] Yunzhe Li, Yujia Xue, and Lei Tian. Deep speckle correlation: a deep learning approach toward scalable imaging through scattering media. Optica, 5(10):1181-1190, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 432 + ], + "type": "text", + "content": "[21] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 434, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 488 + ], + "type": "text", + "content": "[22] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-Inverse Correlography: Towards Real-Time High-Resolution Non-Line-of-Sight Imaging. Optica, 7(1):63–71, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "type": "text", + "content": "[23] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "type": "text", + "content": "[24] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, M. J. Lee, Mattias P. Heinrich, Kazunari Misawa, Kensaku Mori, Steven G. McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. ArXiv, abs/1804.03999, 2018. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "text", + "content": "[25] Richard G Paxman, Timothy J Schulz, and James R Fienup. Joint estimation of object and aberrations by using phase diversity. JOSA A, 9(7):1072-1085, 1992. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural Holography With Camera-in-the-Loop Training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Nikolaj Reiser, Min Guo, Hari Shroff, and Patrick J La Riviere. Phase diverse phase retrieval for microscopy: Com" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25284" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "parison of gaussian and poisson approaches. arXiv preprint arXiv:2308.00734, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[28] Wenqi Ren, Sibo Liu, Hua Zhang, Jin shan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks. In European Conference on Computer Vision, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[29] Sachin Shah, Sakshum Kulshrestha, and Christopher A Metzler. Tidy-psfs: Computational imaging with time-averaged dynamic point-spread-functions. ICCV, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[30] Zheng Shi, Yuval Bahat, Seung-Hwan Baek, Qiang Fu, Hadi Amata, Xiao Li, Praneeth Chakravarthula, Wolfgang Heidrich, and Felix Heide. Seeing through obstructions with diffractive cloaking. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[31] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 331 + ], + "type": "text", + "content": "[32] Carlas S Smith, Arnold J den Dekker, Raluca Andrei, Rufus Fraanje, and Michel Verhaegen. Fast phase diversity wavefront sensing using object independent metrics. In Adaptive Optics Systems III, pages 2250-2258. SPIE, 2012. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "text", + "content": "[33] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386–1396, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[34] Qilin Sun, Congli Wang, Fu Qiang, Dun Xiong, and Heidrich Wolfgang. End-to-end complex lens design with differentiable ray tracing. ACM Trans. Graph, 40(4):1-13, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "text", + "content": "[35] Yu Sun, Zhihao Xia, and Ulugbek S Kamilov. Efficient and Accurate Inversion of Multiple Scattering with Deep Learning. Optics Express, 26(11):14678-14688, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[36] Waleed Tahir, Hao Wang, and Lei Tian. Adaptive 3d scattering with a dynamic synthesis network. Light: Science & Applications, 11(1):42, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[37] Brian J Thelen, Richard G Paxman, David A Carcarara, and John H Seldin. Maximum a posteriori estimation of fixed aberrations, dynamic aberrations, and the object from phase-diverse speckle data. JOsA A, 16(5):1016-1025, 1999. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "text", + "content": "[38] Lei Tian and Laura Waller. Quantitative differential phase contrast imaging in an led array microscope. Optics express, 23(9):11394-11403, 2015. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "text", + "content": "[39] Ethan Tseng, Felix Yu, Yuting Yang, Fahim Mannan, Karl St. Arnaud, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Hyperparameter optimization in black-box image processing using differentiable proxies. ACM Transactions on Graphics (TOG), 38(4), 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[40] Hakan Urey, Ned Nestorovic, Baldwin S Ng, and Abraham A Gross. Optics designs and system mtf for laser scanning displays. In Helmet-and Head-Mounted Displays IV, pages 238-248. SPIE, 1999. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[41] Curtis R Vogel, Tony F Chan, and Robert J Plemmons. Fast algorithms for phase-diversity-based blind deconvolution. In Adaptive Optical System Technologies, pages 994-1005. SPIE, 1998. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 521 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[42] Abbie T Watnik and Dennis F Gardner. Wavefront Sensing in Deep Turbulence. Optics and Photonics News, 29(10): 38-45, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[43] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[44] Meng Xiang, An Pan, Jinpeng Liu, Teli Xi, Xin Guo, Fei Liu, and Xiaopeng Shao. Phase diversity-based fourier psychography for varying aberration correction. Frontiers in Physics, 10:129, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[45] Qi Xin, Guohao Ju, Chunyue Zhang, and Shuyan Xu. Object-independent image-based wavefront sensing approach using phase diversity images and deep learning. Optics express, 27(18):26102-26119, 2019. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 274 + ], + "type": "text", + "content": "[46] Tomer Yeminy and Ori Katz. Guidestar-Free Image-Guided Wavefront Shaping. Science Advances, 7(21), 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[47] Seokchan Yoon, Moonseok Kim, Mooseok Jang, Youngwoon Choi, Wonjun Choi, Sungsam Kang, and Wonshik Choi. Deep Optical Imaging Within Complex Scattering Media. Nature Reviews Physics, 2(3):141-158, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "text", + "content": "[48] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[49] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 473 + ], + "type": "text", + "content": "[50] Peiguang Zhang, Chengliang Yang, Zihao Xu, Zhaoliang Cao, Quanquan Mu, and Li Xuan. High-accuracy wavefront sensing by phase diversity technique with bisymmetric defocuses diversity phase. Scientific Reports, 7(1):15361, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 476, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 521 + ], + "type": "text", + "content": "[51] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40:1452-1464, 2018. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25285" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_content_list.json b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c100384e806641979a8c60e1fdea874f0e51bb1 --- /dev/null +++ b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_content_list.json @@ -0,0 +1,1766 @@ +[ + { + "type": "text", + "text": "Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration", + "text_level": 1, + "bbox": [ + 116, + 130, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Zhao, Weiling Cai*, Chenyu Dong, Chengwei Hu \nSchool of Artificial Intelligence, Nanjing Normal University", + "bbox": [ + 241, + 203, + 720, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 310, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Underwater images are subject to intricate and diverse degradation, inevitably affecting the effectiveness of underwater visual tasks. However, most approaches primarily operate in the raw pixel space of images, which limits the exploration of the frequency characteristics of underwater images, leading to an inadequate utilization of deep models' representational capabilities in producing high-quality images. In this paper, we introduce a novel Underwater Image Enhancement (UIE) framework, named WF-Diff, designed to fully leverage the characteristics of frequency domain information and diffusion models. WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). With our full exploration of the frequency domain information, WFI2-net aims to achieve preliminary enhancement of frequency information in the wavelet space. Our proposed FRDAM can further refine the high- and low-frequency information of the initial enhanced images, which can be viewed as a plug-and-play universal module to adjust the detail of the underwater images. With the above techniques, our algorithm can show SOTA performance on real-world underwater image datasets, and achieves competitive performance in visual quality. The code is available at https://github.com/zhihefang/WF-Diff.", + "bbox": [ + 75, + 287, + 473, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 672, + 209, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Underwater image restoration is a practical but challenging technology in the field of underwater vision, widely used for tasks, such as underwater robotics[26] and underwater object tracking[6]. Due to light refraction, absorption, and scattering in underwater scenes, underwater images are usually severely distorted, with low contrast and blurriness [2]. Therefore, clear underwater images play a critical role in fields that need to interact with the underwater environment. The main goal of underwater image enhancement (UIE) is to obtain high-quality images by removing scattering and correcting color distortion in degraded images. UIE is crucial for vision-related underwater tasks.", + "bbox": [ + 75, + 698, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c2150c3ba66b3750da5912ea5029f6940c1e7309121d304299829a7ddac99ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 272, + 883, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6996da4f8487b3b1d29fd489b0beca75cc692d0ad32c9730e1ebf611aba35ad1.jpg", + "image_caption": [ + "Figure 1. Our motivations. The amplitude and phase are produced by Fast Fourier Transform (FFT) and the recombined images are obtained by Inverse FFT (IFFT). We further explore the frequency properties for underwater images in Wavelet space." + ], + "image_footnote": [], + "bbox": [ + 503, + 338, + 883, + 455 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/8d9290a9dca574e80730b3882e3972d8be9bff79ac954d7b8354f6f8f06829ae.jpg", + "table_caption": [ + "Table 1. Evaluation of using different frequency domain transformation strategies on the UIEBD [28]. S1 refers to swapping the amplitude in original images, S2 refers to only swapping the amplitude of low-frequency sub-images in the wavelet space, and S3 refers to swapping the amplitude of low-frequency sub-images and high-frequency sub-images in the wavelet space." + ], + "table_footnote": [], + "table_body": "
StrategyPSNR ↑SSIM ↑LPIPS↓FID↓
S128.680.90270.095728.70
S227.100.88130.102333.04
S329.970.93430.082023.55
", + "bbox": [ + 545, + 623, + 849, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this problem, traditional UIE methods based on the physical properties of the underwater images were proposed [15, 17, 29-31]. These methods investigate the physical mechanism of the degradation caused by color cast or scattering and compensate for them to enhance the underwater images. However, these physics-based models with limited representation capacity cannot address all the complex physical and optical factors underlying the underwater scenes, which leads to poor enhancement results under highly complex and diverse underwater scenes. Recently, some learning-based methods [7, 18, 28, 36] for UIE can produce better results, since neural networks have powerful feature representation and nonlinear mapping capabilities.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. E-mail: caiwl@njnu.edu.cn", + "bbox": [ + 94, + 886, + 406, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8281", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ties. It can learn the mapping of an image from degenerate to clear from a substantial quantity of paired training data. However, most previous methods are based on the raw pixel space of images, with limited exploration of the properties of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models for generating high-quality images.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on insights from previous Fourier-based works [12, 48], we explore the properties of the Fourier frequency information for the UIE task, as illustrated in Figure 1. Given two images (an underwater image and its corresponding ground-truth), we swap their amplitude components and combine them with corresponding phase components in the Fourier space. The recombined results show that the visual appearance is swapped following the amplitude swapping, which indicates the degradation information of underwater images is mainly contained in the amplitude component. We further explore the properties of the amplitude components in the Wavelet space. Specifically, the images can be decomposed into low-frequency sub-images and high-frequency sub-images using discrete wavelet transformations (DWT), and then we swap amplitude components of low-frequency sub-images. From visual results, we can find a similar phenomenon, which means the color degradation information is mainly contained in low-frequency sub-images, and the texture and detail degradation information is mainly contained in high-frequency sub-images. Table 1 shows the quantitative evaluation of the different frequency domain strategies, proving that our discovery is objective. Consequently, how to adequately exploit the properties of frequency domain information and effectively incorporate them into a unified image enhancement network is a crucial issue.", + "bbox": [ + 75, + 215, + 472, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion-based methods [10, 35] have garnered significant attention due to their outstanding performance in image synthesis [23, 24, 32, 34, 52] and restoration tasks [5, 40, 46, 51]. These methods rely on a hierarchical denoising autoencoder architecture, enabling them to iteratively reverse a diffusion process and achieve high-quality mapping from randomly sampled Gaussian noise to target images or latent distributions [10]. Tang et al. [36] present an image enhancement approach with diffusion model in underwater scenes. While standard diffusion models exhibit sufficient capability, unforeseen artifacts may arise as a result of the diversity introduced during the sampling process from randomly generated Gaussian noise to images [45]. Furthermore, the diffusion model needs to recover both the high and low-frequency information of images, which limits their ability to focus on fine-grained information, missing out on texture and details. Thereby, it is very crucial that the powerful representation capabilities of diffusion models can be fully utilized.", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we develop a novel UIE framework to fully exploit the properties of frequency domain information and diffusion models, called WF-Diff, which mainly consists of two stages: frequency preliminary enhancement and frequency diffusion adjustment. The first stage aims to preliminarily enhance the high-frequency and low-frequency components of underwater images by utilizing the frequency domain characteristics. Specifically, we first convert the input images into the wavelet space using discrete wavelet transformations (DWT), obtaining an average coefficient that represents the low-frequency content information of the input image and three high-frequency coefficients that represent sparse vertical, horizontal, and diagonal details of the input image. Then, we design a wavelet-based Fourier information interaction network (WFI2-net), fully integrating the characteristics of Transformer [22] and Fourier prior information to enhance high- and low-frequency content, respectively. Moreover, to achieve interaction of high- and low-frequency information, we propose a cross-frequency conditioner (CFC) to further improve the generation quality. The target of the second stage is to make adjustment to the initial enhanced coarse results in terms of details and textures via diffusion model. Consequently, we propose a frequency residual diffusion adjustment module (FRDAM). Unlike previous diffusion-based work, FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models in the wavelet space, which can not only increase the model's focus on fine-grained information but also mitigate the adverse effects of the diversity of the sampling process.", + "bbox": [ + 496, + 90, + 893, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of our method are as follows:", + "bbox": [ + 498, + 571, + 890, + 599 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore in depth the properties of the frequency domain for underwater images. Based on the properties and diffusion model, we propose a novel UIE framework, named WF-Diff, with the goal of achieving frequency enhancement and diffusion adjustment.", + "- We propose a frequency residual diffusion adjustment module (FRDAM) to further refine the high- and low-frequency information of the initial enhanced images. FRDAM can be viewed as a plug-and-play universal module to adjust the detail of the underwater images.", + "- We propose a cross-frequency conditioner (CFC) to achieve the cross-frequency interaction of high- and low-frequency information.", + "- Experimental results compared with SOTAs considerably show that our developed WF-Diff performs the superiority against previous UIE approaches, and extensive ablation experiments can demonstrate the effectiveness of our contributions." + ], + "bbox": [ + 500, + 628, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8282", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52e27cbb5d2ffc143f2fe469b95c68d819141d717404f60deb1e5cd2dc613def.jpg", + "image_caption": [ + "Figure 2. Overall framework of WF-Diff. It contains two detachable networks, Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Furthermore, the proposed cross-frequency conditioner (CFC) aims to achieve the cross-frequency interaction of high- and low-frequency information." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 877, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 507, + 227, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Underwater Image Enhancement", + "text_level": 1, + "bbox": [ + 76, + 537, + 369, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, existing UID methods can be briefly categorized into the physical and deep model-based approaches [15, 18, 28-30, 36]. Most UID methods based on the physical model utilize prior knowledge to establish models, such as water dark channel priors [29], attenuation curve priors [38], fuzzy priors [4]. In addition, Akkaynak and Treibitz [1] proposed a method based on the revised physical imaging model. However, the depth map of the underwater scene is difficult to obtain. This leads to unstable performance, which usually suffers from severe color cast and artifacts. Therefore, the manually established priors restrain the model's robustness and scalability under the complicated and varied circumstances. Recently, deep learning-based methods [18, 28, 36] have achieved acceptable performance. To alleviate the need for real-world underwater paired training data, many methods introduced GAN-based framework for UIE [7, 14, 21, 49], such as WaterGAN [21], UGAN [7] and UIE-DAL [37]. Recently, some complex frameworks were proposed and achieve the-state-of-the-art performance [15, 29]. Ucolor [19] combined the underwater physical imaging model in the raw space and designed a medium transmission guided model. Yang et al.", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[43] proposed a reflected light-aware multi-scale progressive restoration network to obtain images with both color equalization and rich texture in various underwater scenes. Huang et al. [13] proposed a mean teacher based semi-supervised network, which effectively leverages the knowledge from unlabeled data. However, most previous methods are based on the spatial domain, with limited exploration of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models.", + "bbox": [ + 496, + 508, + 890, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion Model", + "text_level": 1, + "bbox": [ + 500, + 676, + 661, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, Diffusion Probabilistic Models (DPMs) [10, 35] have been widely adopted for conditional image generation [5, 40, 42, 46, 50]. Sahara et al. [33] proposed Palette, which has demonstrated the excellent performance of diffusion models in the field of conditional image generation, including colorization, in-painting and JPEG restoration. Tang et al. [36] presented an image enhancement approach with diffusion model in underwater scenes. However, the reverse process starts from randomly sampled Gaussian noise to full images [45], which can lead to unexpected artifacts due to the diversity of the sampling process. Furthermore, the diffusion model needs to recover both the high and low-frequency information in images, which lim", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8283", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "its their ability to focus on fine-grained information. Consequently, how to incorporate diffusion models into a unified underwater image enhancement network is a vital issue.", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 152, + 210, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 76, + 178, + 264, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an underwater image as input, our goal is to learn a network to generate an output that eliminates the color cast from input while enhancing the image details. The overall framework of WF-Diff is shown in Figure 2. WF-Diff is designed to fully leverage the characteristics of frequency domain information and powerful ability of diffusion models. Specifically, WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). We first convert the input into the wavelet space using discrete wavelet transformations (DWT), obtaining a low-frequency coefficient and three high-frequency coefficients. WFI2-net is dedicated to achieving preliminary enhancement of frequency information. We fully integrate the characteristics of Transformer and Fourier prior Information, and design wide transformer block (WTB) and spatial-frequency fusion block (SFFB) to enhance high- and low-frequency content respectively. FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Note that, our proposed FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models, respectively. Additionally, the proposed cross-frequency conditioner (CFC) strives to achieve cross-frequency interaction between high- and low-frequency information.", + "bbox": [ + 76, + 200, + 472, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Discrete Wavelet and Fourier Transform", + "text_level": 1, + "bbox": [ + 76, + 636, + 429, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discrete Wavelet transform (DWT) has been widely applied to low-level vision tasks [11, 16]. We firstly use DWT to decompose an input into multiple frequency sub-bands so that we can achieve the color correction of low-frequency information and detail enhancement of high-frequency information, respectively. Given an underwater image as input $I \\in \\mathbb{R}^{H \\times W \\times c}$ , we use DWT with Haar wavelets to decompose the input. Haar wavelets consist of the low-pass filter $L$ , and the high-pass filter $H$ , as follows:", + "bbox": [ + 76, + 660, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = \\frac {1}{\\sqrt {2}} [ 1, 1 ] ^ {T}, H = \\frac {1}{\\sqrt {2}} [ 1, - 1 ] ^ {T}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 808, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can obtain four sub-bands, which can be expressed as:", + "bbox": [ + 76, + 852, + 468, + 880 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {L L}, \\left\\{I _ {L H}, I _ {H L}, I _ {H H} \\right\\} = \\mathrm {D W T} (I), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{LL},\\{I_{LH},I_{HL},I_{HH}\\} \\in \\mathbb{R}^{\\frac{H}{2}\\times \\frac{W}{2}\\times c}$ represent the low-frequency component of the input and high-frequency components in the vertical, horizontal, and diagonal directions, respectively. More specifically, the low-frequency component contains the content and color information of the input image, and the other three high-frequency coefficients contain details information of global structures and textures [31]. The sub-bands are downsampled to half-resolution of the input but do not result in information loss due to the biorthogonal property of DWT. For low-frequency component $I_{LL}$ , we will explore its properties in Fourier space.", + "bbox": [ + 496, + 88, + 890, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we introduce the operation of the Fourier transform [48]. Given a image $x \\in \\mathbb{R}^{H \\times W \\times 1}$ , whose shape is $H \\times W$ , the Fourier transform $\\mathcal{F}$ which converts $x$ to the Fourier space $X$ can be expressed as:", + "bbox": [ + 498, + 272, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (x) (u, v) = X (u, v) = \\frac {1}{\\sqrt {H W}} \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} x (h, w) e ^ {- j 2 \\pi \\left(\\frac {h}{H} u + \\frac {w}{W} v\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 339, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h, w$ are the coordinates in the spatial space and $u, v$ are the coordinates in the Fourier space. $\\mathcal{F}^{-1}$ denotes the inverse transform of $\\mathcal{F}$ . Complex component $X(u, v)$ in the Fourier space can be represented by a amplitude component $\\mathcal{A}(X(u, v))$ and a phase component $\\mathcal{P}(X(u, v))$ as follows:", + "bbox": [ + 498, + 381, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} (X (u, v)) = \\sqrt {R ^ {2} (X (u , v)) + I ^ {2} (X (u , v))}, \\tag {4} \\\\ \\mathcal {P} (X (u, v)) = \\arctan [ \\frac {I (X (u , v))}{R (X (u , v))} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 463, + 890, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R(x)$ and $I(x)$ represent the real and imaginary parts of $X(u, v)$ , respectively. Note that, the Fourier operation can be computed alone in each channel for feature maps.", + "bbox": [ + 496, + 507, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Figure 1 and Table 1 (our motivation), we conclude that the color degradation information of underwater images is mainly contained in the amplitude component of low-frequency sub-band, and the texture and detail degradation information is mainly contained in high-frequency sub-bands.", + "bbox": [ + 496, + 551, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Frequency Preliminary Enhancement", + "text_level": 1, + "bbox": [ + 500, + 650, + 831, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above analysis, in frequency preliminary enhancement stage, we design a simple but effective WFI2-net with a parallel encoder-decoder (U-Net-like) format to restore the amplitude component of low-frequency information and high-frequency components, respectively. We also utilize skip connections to connect the features at the same level in the encoder and decoder. For high-frequency branch, we utilize advantage of transformer modeling global information to enhance high-frequency coefficients. We design wide transformer block (WTB) using multi-scale information, aiming to model long range dependencies. Our low-frequency branch aims to restore the amplitude component in Fourier space. In order to obtain rich frequency and spatial information, we design spatial-frequency fusion block (SFFB).", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8284", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/07874ab2c66a9096f5c0e889f10ace59987f12d06afcb6dc0d31d288e9b4ee2a.jpg", + "image_caption": [ + "Figure 3. The architecture of (a) Wide Transformer Block, (b) Spatial-Frequency Fusion Block and (c) Cross-Frequency Conditioner." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 369, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb4f1f13acda6bcadf098ccf3ecd3da65d1d627b237d70e1fba648a071ef9a6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 88, + 581, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef6e42c87c0984218dcdfbe2c25bb5a1752db5a16a1f33f71d4edb2bf1741b80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 88, + 879, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Wide Transformer Block. Unlike low-frequency coefficients, high-frequency components contain global structure and texture details. Consequently, the high-frequency branch focuses on modeling global and local features. WTB is shown in Figure 3 (a). Given $I_{LH}, I_{HL}, I_{HH} \\in \\mathbb{R}^{\\frac{H}{2} \\times \\frac{W}{2} \\times c}$ , WTB firstly obtains their embedding features $T_{in} \\in \\mathbb{R}^{3 \\times \\frac{H}{2} \\times \\frac{W}{2} \\times C}$ through convolution projection. To be specific, WTB are composed of an attention (Atten) module and a feed-forward network (FFN) module, and the computation can be denoted in the WTB as:", + "bbox": [ + 76, + 253, + 472, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ, K, V, L = \\operatorname {S p l i t} \\left(W _ {d} W _ {p} \\left(\\operatorname {N o r m} \\left(T _ {i - 1}\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 411, + 468, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {T} _ {i} = S A (Q, K, V) + C A (L) + T _ {i - 1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 434, + 468, + 452 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = F F N \\left(\\operatorname {N o r m} \\left(\\hat {T} _ {i}\\right)\\right) + \\hat {T} _ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 455, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $SA$ and $CA$ refer to self-attention and channel attention, respectively. Norm refers to normalization. $T_{i - 1}$ represents the input embeddings of the current WTB. $W_{d}$ and $W_{p}$ denote $1 \\times 1$ point-wise convolution and multi-scale kernel depth-wise convolution, respectively; Split refers to the split operation. $L$ aims to focus on local information.", + "bbox": [ + 75, + 477, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial-Frequency Fusion Block. We show the structure of SFFB in Figure 3 (b), which has a spatial domain unit (SDU) and a frequency domain unit (FDU) for interaction of dual domain representations. In spatial domain unit, we employ multi-scale convolution kernels in order to enlarge the limited spatial receptive field. After obtaining the spatial embeddings $F_{s}$ , we firstly utilize the FFT to obtain the amplitude $\\mathcal{A}(F_s)$ and phase $\\mathcal{P}(F_s)$ components. Then, the $\\mathcal{A}(F_s)$ and $\\mathcal{P}(F_s)$ are fed into two layers $1^{*}1$ conv to obtain $\\mathcal{A}'(F_s)$ and $\\mathcal{P}'(F_s)$ . Finally, we use the IFFT algorithm to map $\\mathcal{A}'(F_s)$ and $\\mathcal{P}'(F_s)$ to image space and obtain the frequency embeddings $F_{f}$ . The fusion embeddings of spatial domain and frequency domain can be expressed as:", + "bbox": [ + 75, + 568, + 470, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s f} = F _ {s} + F _ {f}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 771, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. We denote $I_{LL}^{\\prime}$ as output of low-frequency branch, and $I_{LH}^{\\prime}, I_{HL}^{\\prime}$ and $I_{HH}^{\\prime}$ as output of high-frequency branch. Ground-truth (G) can be decomposed $G_{LL}, G_{LH}, G_{HL}, G_{HH}$ by DWT. The high-frequency loss can be expressed as:", + "bbox": [ + 75, + 795, + 470, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} = \\left\\| I _ {(i)} ^ {\\prime} - G _ {(i)} \\right\\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 877, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i \\in \\{LH, HL, HH\\}$ . For low-frequency information, we only constrain the amplitude components. Consequently, the low-frequency loss can be expressed as:", + "bbox": [ + 496, + 253, + 890, + 300 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a} = \\left| \\left| \\mathcal {A} \\left(I _ {L L} ^ {\\prime}\\right)\\right) - \\mathcal {A} \\left(G _ {L L}\\right) \\right| | _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 308, + 890, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}()$ refers to the amplitude component in Fourier transform. Finally, We further use an adversarial loss in Wasserstein GAN as reconstruction Loss $\\mathcal{L}_{rec}$ .", + "bbox": [ + 496, + 334, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Cross-Frequency Conditioner", + "text_level": 1, + "bbox": [ + 498, + 387, + 769, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The detailed structure of CFC is shown in Figure 3 (c). The purpose of CFC is to facilitate information interaction between high and low frequency features, achieving cross-frequency mutual reinforcement between high and low frequency branches, thus enhancing overall representation power. We denote $T_{in}$ and $F_{in}$ as the input features of CFC, which represent high- and low-frequency embeddings. For the high-frequency embedding features $T_{in} \\in \\mathbb{R}^{3 \\times \\frac{H}{2} \\times \\frac{W}{2} \\times C}$ , we can obtain $T_{LH}, T_{HL}, T_{HH} \\in \\mathbb{R}^{\\frac{H}{2} \\times \\frac{W}{2} \\times c}$ via split operation. By adding these extracted coefficients, we obtain the aggregated high-frequency embeddings. We use different linear projections to construct $Q$ and $K$ in CFC:", + "bbox": [ + 496, + 410, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\operatorname {C o n v} _ {1 \\times 1} \\left(T _ {L H} + T _ {H L} + T _ {H H}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 616, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = \\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {i n}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 643, + 890, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, $V_{T}$ of high-frequency embeddings and $V_{F}$ of low-frequency embeddings can be obtained:", + "bbox": [ + 498, + 662, + 890, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {T} = \\operatorname {C o n v} _ {1 \\times 1} \\left(T _ {L H} + T _ {H L} + T _ {H H}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 703, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {F} = \\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {i n}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 729, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output feature map $T_{out}$ and $F_{out}$ can then be obtained from the formula:", + "bbox": [ + 498, + 751, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {o u t} = R \\left(\\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V _ {T}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 786, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o u t} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V _ {F}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 830, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R$ denotes a replication operation, and $\\sqrt{d_k}$ is the number of columns of matrix $Q$ .", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8285", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Frequency Diffusion Adjustment", + "text_level": 1, + "bbox": [ + 76, + 90, + 370, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FRDAM aims to further adjust the high- and low-frequency information using the powerful representation of diffusion model. Generally, FRDAM can be divided into two branch, namely low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB). We adopt the diffusion process proposed in DDPM [10] to construct the residual distribution of high- and low-frequency information for each branch, which can be described as a forward diffusion process and a reverse diffusion process.", + "bbox": [ + 75, + 113, + 468, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Forward Diffusion Process. The forward diffusion process can be viewed as a Markov chain progressively adding Gaussian noise to the data. Given the initial enhanced frequency components $I_{i}^{\\prime}$ and its ground truth $G_{i}$ , $i \\in \\{LL, LH, HL, HH\\}$ , we calculate their residual distribution $x_{0} = G_{i} - I_{i}^{\\prime}$ , then introduce Gaussian noise based on the time step, as follows:", + "bbox": [ + 75, + 250, + 468, + 357 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 371, + 468, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ is a variable controlling the variance of the noise. Introducing $\\alpha_{t} = 1 - \\beta_{t}$ , this process can be described as:", + "bbox": [ + 76, + 392, + 467, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t - 1}, \\quad \\epsilon_ {t - 1} \\sim \\mathcal {N} (0, \\mathcal {Z}). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 428, + 468, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With Gaussian distributions are merged, We can obtain :", + "bbox": [ + 96, + 452, + 468, + 467 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 476, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reverse Diffusion Process. The reverse diffusion process aims to restore the residual distribution from the Gaussian noise. The reverse diffusion can be expressed as:", + "bbox": [ + 76, + 496, + 468, + 541 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, x _ {c} ^ {(l)}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right), \\sigma_ {t} ^ {2} \\mathcal {Z}\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 546, + 468, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we take the LDFB as an example, and $x_{c}^{(l)}$ refers to the conditional image $I_{LL}^{\\prime}$ . $\\pmb{\\mu}_{\\theta}(x_t,x_c^{(l)},t)$ and $\\sigma_t^2$ are the mean and variance from the estimate of step t, respectively. In LDFB and HDFB, we follow the setup of [35], they can be expressed as:", + "bbox": [ + 76, + 571, + 468, + 650 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right) = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\left(1 - \\bar {\\alpha} _ {t}\\right)} \\epsilon_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right)\\right), (21) \\\\ \\sigma_ {t} ^ {2} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}, (22) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 656, + 468, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(x_t,x_c^{(l)},t)$ is the estimated value with a Unet.", + "bbox": [ + 76, + 724, + 439, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We optimize an objective function for the noise estimated by the network and the noise $\\epsilon^{(l)}$ actually added in LDFB. Therefore, the diffusion loss process is:", + "bbox": [ + 76, + 743, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d m} (\\theta) = \\left\\| \\epsilon^ {(l)} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon^ {(l)}, x _ {c} ^ {(l)}, t\\right) \\right\\|. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 792, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generally, the frequency diffusion adjustment process is to refine the high- and low-frequency component of the initial enhancement. The whole diffusion process can be formulated:", + "bbox": [ + 76, + 816, + 468, + 876 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {(i)} = \\mathcal {F} _ {H D F B} \\left(\\epsilon_ {s} ^ {(h)}, I _ {(i)} ^ {\\prime}\\right), i \\in \\{L H, H L, H H \\}, \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 882, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {L L} = \\mathcal {F} _ {L D F B} \\left(\\epsilon_ {s} ^ {(l)}, I _ {L L} ^ {\\prime}\\right), \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 92, + 890, + 109 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_s^{(h)}\\in \\mathbb{R}^{3\\times \\frac{H}{2}\\times \\frac{W}{2}\\times 3}$ and $\\epsilon_s^{(l)}\\in \\mathbb{R}^{\\frac{H}{2}\\times \\frac{W}{2}\\times 3}$ are Gaussian noise.", + "bbox": [ + 498, + 116, + 890, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ultimately, the refined frequency components are obtained as the addition of the diffusion generative residual distribution and the initial enhanced frequency components. Then, we employ IDWT to obtain the final generated image:", + "bbox": [ + 498, + 151, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {f i n a l} = I D W T \\left(I _ {(i)} ^ {\\prime} + \\hat {I} _ {(i)}, I _ {L L} ^ {\\prime} + \\hat {I} _ {L L}\\right), i \\in \\{L H, H L, H H \\} \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 224, + 901, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 262, + 630, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 287, + 581, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our network, implemented using PyTorch 1.7, underwent training and testing on an NVIDIA GeForce RTX 3090 GPU. We employed the Adam optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ . The patch size was configured as $256 \\times 256$ , and the batch size was set to 2. The diffusion model's total time steps, denoted as T, were set to 1000, and the number of training iterations reached one million. The initial learning rate was established at 0.0001.", + "bbox": [ + 498, + 310, + 890, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We utilize the real-world UIEBD dataset [28] and the LSUI dataset [18] for training and evaluating our model. The UIEBD dataset comprises 890 underwater images with corresponding labels. Out of these, 700 images are allocated for training, and the remaining 190 are designated for testing. The LSUI dataset is randomly partitioned into 4500 images for training and 504 images for testing. In addition, to verify the generalization of WF-Diff, we use non-reference benchmarks U45 [20], which contains 45 underwater images for testing.", + "bbox": [ + 498, + 431, + 890, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison methods. We conduct a comparative analysis between WF-Diff and eight state-of-the-art (SOTA) UIE methods, namely UIEC^2-Net [39], Water-Net [28], UWCNN [3], SCNet [8], UIEWD [25], U-color [19], U-shape [18], and DM-water [36]. To ensure a fair and rigorous comparison, we utilize the provided source codes from the respective authors and adhere strictly to the identical experimental settings across all evaluations.", + "bbox": [ + 498, + 583, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We primarily utilize well-established full-reference image quality assessment metrics: PSNR and SSIM [41]. PSNR and SSIM offer quantitative comparisons of our method with other approaches at both pixel and structural levels. Higher PSNR and SSIM values signify superior quality of the generated images. Additionally, we incorporate the LPIPS and FID metrics for full-reference image evaluation. LPIPS [47] is a deep neural network-based image quality metric that assesses the perceptual similarity between an image and a reference image. FID [9] measures the distance between the distributions of real and generated images. A lower LPIPS and FID score indicates a more effective UIE approach. For non-reference benchmarks U45,", + "bbox": [ + 498, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8286", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bf471245e283cb23dca867690861b8ba5ed652cc830a89f9ce0ef336795dab07.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of different UIE methods on the UIEBD, LSUI and U45 datasets. The best results are highlighted in bold and the second best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsUIEWDUWCNNUIEC^2-NetWater-NetSCNetU-colorU-shapeDM-waterOurs
UIBDFID↓85.1294.4435.0637.4833.6638.2546.1131.0727.85
LPIPS↓0.39560.35250.20330.21160.24970.23370.22640.14360.1248
PSNR↑14.6515.4020.1419.3520.4120.7121.2521.8823.86
SSIM↑0.72650.77490.82150.83210.82350.84110.84530.81940.8730
LSUIFID↓98.49100.534.5138.90158.9945.0628.5627.9126.75
LPIPS↓0.39620.34500.14320.16780.2830.1230.10280.11380.1096
PSNR↑15.4318.2420.8619.7322.6322.9124.1627.6527.26
SSIM↑0.78020.84650.88670.82260.91760.89020.93220.88670.9437
U45UIQM↑2.4582.3792.7802.9572.8563.1043.1513.0863.181
UCIQE↑0.5830.5670.5910.6010.5940.5860.5920.6340.619
", + "bbox": [ + 76, + 122, + 893, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bfa9fdf1354aa95921dc1c522da95e1d126ca8b39acaedad87f907dbe181caf.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with other SOTA methods on real underwater images." + ], + "image_footnote": [], + "bbox": [ + 83, + 343, + 885, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we introduce UIQM [27] and UCIQE [44] to evaluate our method.", + "bbox": [ + 76, + 691, + 470, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results and Comparisons", + "text_level": 1, + "bbox": [ + 76, + 738, + 310, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the quantitative results compared with different baselines on UIEBD, LSUI and U45 datasets, including with UIEC^2-Net, Water-Net, UIEWD, UWCNN, SCNet, U-color, U-shape, and DM-water. We mainly use PSNR, SSIM, LPIPS and FID as our quantitative indices for UIEBD and LSUI datasets, and UIQM and UCIQE for non-reference dataset U45. The results in Table 2 show that our algorithm outperforms state-of-the-art methods obviously, and achieve state-of-the-art performance in terms of image", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "quality evaluation metrics on UIE task, which verifies the robustness of the proposed WF-Diff. To better validate the superiority of our methods, in Figure 4, we show the visual results comparison with state-of-the-art methods on real underwater images. The six examples are randomly selected on the UIEBD and LSUI datasets. Our methods consistently generate natural and better visual results on testing images, strongly proving that WF-Diff has good generalization performance for real-world applications.", + "bbox": [ + 496, + 691, + 893, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 844, + 653, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study with WFI2-net. In order to evaluate the effectiveness of each part in WFI2-net, we conduct two ab", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8287", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a5351001393d3ea50984e02a541a4c4eb5ab59a80c32f715dd4b3eeb499d67a.jpg", + "table_caption": [ + "Table 3. Ablation study with network structure of WFI2-net on UIEBD dataset." + ], + "table_footnote": [], + "table_body": "
WTBSFFBUIEBD
SACASDUFDUPSNR ↑SSIM↑
×✓✓✓20.940.8541
✓×✓✓20.820.8473
✓✓×✓21.110.8586
✓✓✓×20.230.8346
✓✓✓✓21.870.8622
", + "bbox": [ + 109, + 128, + 439, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/72e4a556e824321f432f678f8a97356819f7e8fd91213b88ce69438e22c80196.jpg", + "table_caption": [ + "Table 4. Ablation study with loss of WFI2-net on UIEBD dataset." + ], + "table_footnote": [], + "table_body": "
MethodsUIEBD
LhLrecLaCFCPSNR ↑SSIM↑
×✓✓✓20.460.8274
✓×✓✓20.650.8399
✓✓×✓19.810.8213
✓✓✓×20.970.8425
✓✓✓✓21.870.8622
", + "bbox": [ + 109, + 305, + 439, + 444 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a1e9c99034a20db128705ef6ccc836d5f5129eb9fe3665a02909924a27a73cb.jpg", + "table_caption": [ + "Table 5. Ablation study with FRDAM on UIEBD dataset. DM refers to diffusion model (DM), RDM refers to residual DM. D-L refers to refining low-frequency component in wavelet space, and D-H refers to refining high-frequency components in wavelet space." + ], + "table_footnote": [], + "table_body": "
MethodDMRDMD-LD-HCFCPSNR↑
A✓××××20.86
B×✓×××22.37
C×✓✓××22.32
D×✓×✓×22.58
E×✓✓✓×23.44
F×✓✓✓✓23.86
", + "bbox": [ + 76, + 541, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ce8f5b83440314316edfeb324a054b8a0f8e33615e89a923271832e831b0ca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 685, + 173, + 729 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a42bb405fb55c40cf68d32fafc8dc8e634dc63d81671a490de5722991eb1c0da.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 178, + 685, + 269, + 731 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e0035f3d8970625de9e5a3ea51832ef0be936d4aca831c9161c3bfc1afd595a.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 274, + 685, + 364, + 731 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec570d0ab29c9a8cb0af6105ae1e6865ce883378840e48e17e6748ecc45b4bb3.jpg", + "image_caption": [ + "C" + ], + "image_footnote": [], + "bbox": [ + 372, + 685, + 464, + 731 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86d01bf4be972e6bae92526334c2cc007e285d6a93d59a57dca0dec27fcde04c.jpg", + "image_caption": [ + "input", + "D", + "Figure 5. Visual results of ablation study with FRDAM." + ], + "image_footnote": [], + "bbox": [ + 83, + 742, + 174, + 787 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31b24dcea85bb04a297863eafe656f8b463997c378e726ce1b611a4eb3987200.jpg", + "image_caption": [ + "E" + ], + "image_footnote": [], + "bbox": [ + 178, + 742, + 269, + 787 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59256db715b1dc3377efcc9f4cd5b3f4acce04bf12127dbe294dbbcc2617ad21.jpg", + "image_caption": [ + "F" + ], + "image_footnote": [], + "bbox": [ + 274, + 742, + 366, + 787 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0515a0e928e77986f2670c5a2d47c600afbfa759645df275dfbe996e8ebcc833.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 372, + 742, + 464, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lation experiments with WFI2-net on the UIEBD dataset in Table 3 and 4. CFC is cross-frequency conditioner. Note that, we do not discuss the effect of our proposed FRDAM", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fec30c812c6baee906888bed4283e7bc0f9200b64bfab615470328da486bd690.jpg", + "table_caption": [ + "Table 6. Comparison results of inference speed." + ], + "table_footnote": [], + "table_body": "
MethodSCNetUshapeDM-waterWFI2-netWF-Diff
Inference time (s)0.01490.03530.14410.07390.2816
", + "bbox": [ + 513, + 108, + 877, + 142 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2816a0de9395b0184aaef7b2285708c0540a15e6452ddcd993f5d29527d8bf4f.jpg", + "table_caption": [ + "Table 7. Comparison results of FLOPs." + ], + "table_footnote": [], + "table_body": "
MethodSCNetWaterNetUcolorUshapeWFI2-net
FLOPs (G)5.88193.7443.866.294.1
", + "bbox": [ + 513, + 167, + 877, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "here. We regularly remove one component to each configuration at one time, and our strategy achieves the best performance by using all loss functions and blocks, proving that each part of WFI2-net is useful for UIE task.", + "bbox": [ + 496, + 226, + 890, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study with FRDAM. In this section, we will discuss the effectiveness of FRDAM. Table 5 shows the quantitative results on the UIEBD dataset, and Figure 5 shows the visual results. Note that, Model A and B achieve diffusion process in the pixel level, and model C, D, E and E achieve diffusion process in wavelet space. Model A obtains relatively poor results in PSNR and generates images with color distortion or artifacts in Figure 5 due to the diversity of the sampling process. Model B does not yield fully satisfactory results, because it needs to adjust both the high and low-frequency information in images, which limits their ability to focus on fine-grained information. Compared to model C, model D achieves better results, suggesting that the degradation information is mainly in the high-frequency information in frequency diffusion adjustment stage. Model F achieves the best performance, proving that our designed FRDAM is best for UIE task.", + "bbox": [ + 496, + 290, + 892, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 561, + 617, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we develop a novel UIE framework, namely WF-Diff. With fully utilizing the frequency domain characteristics and diffusion model, WFI2-net can achieve enhancement and adjustment of frequency information. Our proposed FRDAM is a plug-and-play universal module to adjust the details of the underwater images. WF-Diff shows SOTA performance on UIE task, and extensive ablation experiments prove that each of our contributions is effective.", + "bbox": [ + 496, + 590, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. As a result of employing two diffusion models, our approach doesn't confer an advantage in terms of inference speed. Table 6 presents the actual values of inference speed for WF-Diff, indicating that our model does not outperform recent approaches in terms of inference time. Furthermore, Table 7 provides a comparison of FLOPs for WFI2-net. It should be mentioned that the implicit sampling step is set to 10 for WF-Diff. Hence, in the future, we will delve into methods to expedite the sampling process.", + "bbox": [ + 496, + 715, + 890, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62276138 and 62371232).", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8288", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Derya Akkaynak and Tali Treibitz. Sea-thru: A method for removing water from underwater images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1682-1691. Computer Vision Foundation / IEEE, 2019. 3", + "[2] Derya Akkaynak, Tali Treibitz, Tom Shlesinger, Yossi Loya, Raz Tamir, and David Iluz. What is the space of attenuation coefficients in underwater computer vision? In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 568-577, 2017. 1", + "[3] Saeed Anwar, Chongyi Li, and Fatih Porikli. Deep underwater image enhancement. CoRR, abs/1807.03528, 2018. 6", + "[4] John Yi-Wu Chiang and Ying-Ching Chen. Underwater image enhancement by wavelength compensation and dehazing. IEEE Trans. Image Process., 21(4):1756-1769, 2012. 3", + "[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14347-14356. IEEE, 2021. 2, 3", + "[6] Karin de Langis and Junaed Sattar. Realtime multi-diver tracking and re-identification for underwater human-robot collaboration. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 11140-11146, 2020. 1", + "[7] Cameron Fabbri, Md Jahidul Islam, and Junaed Sattar. Enhancing underwater imagery using generative adversarial networks. In 2018 IEEE International Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May 21-25, 2018, pages 7159-7165. IEEE, 2018. 1, 3", + "[8] Zhenqi Fu, Xiaopeng Lin, Wu Wang, Yue Huang, and Xinghao Ding. Underwater image enhancement via learning water type desensitized representations. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2764-2768. IEEE, 2022. 6", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6", + "[11] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based CNN for multi-scale face super resolution. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1698-1706, 2017. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. pages 163-180, 2022. 2", + "[13] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18145-18155. IEEE, 2023. 3", + "[14] Md Jahidul Islam, Youya Xia, and Junaed Sattar. Fast underwater image enhancement for improved visual perception. IEEE Robotics Autom. Lett., 5(2):3227-3234, 2020. 3", + "[15] Paulo Drews Jr., Erickson Rangel do Nascimento, F. Moraes, Silvia S. C. Botelho, and Mario F. M. Campos. Transmission estimation in underwater single images. In 2013 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 825-830, 2013. 1, 3", + "[16] Eunhee Kang, Won Chang, Jae Jun Yoo, and Jong Chul Ye. Deep convolutional framelet denoising for low-dose CT via wavelet residual network. IEEE Trans. Medical Imaging, 37 (6):1358-1369, 2018. 4", + "[17] Chongyi Li, Jichang Guo, Runmin Cong, Yanwei Pang, and Bo Wang. Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior. IEEE Trans. Image Process., 25(12):5664-5677, 2016. 1", + "[18] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Trans. Image Process., 29:4376-4389, 2020. 1, 3, 6", + "[19] Chongyi Li, Saeed Anwar, Junhui Hou, Runmin Cong, Chunle Guo, and Wenqi Ren. Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE Trans. Image Process., 30:4985-5000, 2021. 3, 6", + "[20] Hanyu Li, Jingjing Li, and Wei Wang. A fusion adversarial underwater image enhancement network with a public test dataset. arXiv preprint arXiv:1906.06819, 2019. 6", + "[21] Jie Li, Katherine A. Skinner, Ryan M. Eustice, and Matthew Johnson-Roberson. Watergan: Unsupervised generative network to enable real-time color correction of monocular underwater images. IEEE Robotics Autom. Lett., 3(1):387-394, 2018. 3", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 2", + "[23] Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-Icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 2", + "[24] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8289", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ziyin Ma and Changjae Oh. A wavelet-based dual-stream network for underwater image enhancement. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2769-2773. IEEE, 2022. 6", + "[26] James McMahon and Erion Plaku. Autonomous data collection with timed communication constraints for unmanned underwater vehicles. IEEE Robotics Autom. Lett., 6(2): 1832-1839, 2021. 1", + "[27] Karen Panetta, Chen Gao, and Sos Agaian. Human-visual-system-inspired underwater image quality measures. IEEE Journal of Oceanic Engineering, 41(3):541-551, 2015. 7", + "[28] Lintao Peng, Chunli Zhu, and Liheng Bian. U-shape transformer for underwater image enhancement. IEEE Trans. Image Process., 32:3066-3079, 2023. 1, 3, 6", + "[29] Yan-Tsung Peng and Pamela C. Cosman. Underwater image restoration based on image blurriness and light absorption. IEEE Trans. Image Process., 26(4):1579-1594, 2017. 1, 3", + "[30] Yan-Tsung Peng, Keming Cao, and Pamela C. Cosman. Generalization of the dark channel prior for single image restoration. IEEE Trans. Image Process., 27(6):2856-2868, 2018. 3", + "[31] Priyadharsini Ravisankar, T. Sree Sharmila, and V. Rajendran. A wavelet transform based contrast enhancement method for underwater acoustic images. Multidimens. Syst. Signal Process., 29(4):1845-1859, 2018. 1, 4", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2", + "[33] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 15:1-15:10. ACM, 2022. 3", + "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021. 2, 3, 6", + "[36] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 5419-5427. ACM, 2023. 1, 2, 3, 6", + "[37] Pritish M. Uplavikar, Zhenyu Wu, and Zhangyang Wang. All-in-one underwater image enhancement using domain" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1-8, 2019. 3", + "[38] Yi Wang, Hui Liu, and Lap-Pui Chau. Single underwater image restoration using adaptive attenuation-curve prior. IEEE Trans. Circuits Syst. I Regul. Pap., 65-I(3):992-1002, 2018. 3", + "[39] Yudong Wang, Jichang Guo, Huan Gao, and HuiHui Yue. Uiec $^2$ -net: Cnn-based underwater image enhancement using two color space. Signal Process. Image Commun., 96: 116250, 2021. 6", + "[40] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 2, 3", + "[41] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6", + "[42] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16272-16282. IEEE, 2022. 3", + "[43] Jian Yang, Chen Li, and Xuelong Li. Underwater image restoration with light-aware progressive network. In IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023, pages 1-5. IEEE, 2023. 3", + "[44] Miao Yang and Arcot Sowmya. An underwater color image quality evaluation metric. IEEE Transactions on Image Processing, 24(12):6062-6071, 2015. 7", + "[45] Zongyuan Yang, Baolin Liu, Yongping Xiong, Lan Yi, Guibin Wu, Xiaojun Tang, Ziqi Liu, Junjie Zhou, and Xing Zhang. Docdiff: Document enhancement via residual diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 2795-2806. ACM, 2023. 2, 3", + "[46] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with A generative diffusion model. CoRR, abs/2308.13164, 2023. 2, 3", + "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595, 2018. 6", + "[48] Chen Zhao, Wei-Ling Cai, Chenyu Dong, and Ziqi Zeng. Toward sufficient spatial-frequency interaction for gradient-aware underwater image enhancement. arXiv preprint arXiv:2202.08537, 2023. 2, 4", + "[49] Chen Zhao, Wei-Ling Cai, and Zheng Yuan. Spectral normalization and dual contrastive regularization for image-to" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8290", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image translation. The Visual Computer, pages 1-12, 2024. 3", + "[50] Chen Zhao, Chenyu Dong, and Weiling Cai. Learning a physical-aware diffusion model based on transformer for underwater image enhancement. arXiv preprint arXiv:2403.01497, 2024. 3", + "[51] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 2", + "[52] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. arXiv preprint arXiv:2402.05408, 2024. 2" + ], + "bbox": [ + 78, + 90, + 468, + 261 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8291", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_model.json b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fdc1433bc1e5382e9072914ffb41055c2221689b --- /dev/null +++ b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_model.json @@ -0,0 +1,2444 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.131, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.204, + 0.721, + 0.241 + ], + "angle": 0, + "content": "Chen Zhao, Weiling Cai*, Chenyu Dong, Chengwei Hu \nSchool of Artificial Intelligence, Nanjing Normal University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.275, + 0.312, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.474, + 0.665 + ], + "angle": 0, + "content": "Underwater images are subject to intricate and diverse degradation, inevitably affecting the effectiveness of underwater visual tasks. However, most approaches primarily operate in the raw pixel space of images, which limits the exploration of the frequency characteristics of underwater images, leading to an inadequate utilization of deep models' representational capabilities in producing high-quality images. In this paper, we introduce a novel Underwater Image Enhancement (UIE) framework, named WF-Diff, designed to fully leverage the characteristics of frequency domain information and diffusion models. WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). With our full exploration of the frequency domain information, WFI2-net aims to achieve preliminary enhancement of frequency information in the wavelet space. Our proposed FRDAM can further refine the high- and low-frequency information of the initial enhanced images, which can be viewed as a plug-and-play universal module to adjust the detail of the underwater images. With the above techniques, our algorithm can show SOTA performance on real-world underwater image datasets, and achieves competitive performance in visual quality. The code is available at https://github.com/zhihefang/WF-Diff." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.673, + 0.21, + 0.689 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Underwater image restoration is a practical but challenging technology in the field of underwater vision, widely used for tasks, such as underwater robotics[26] and underwater object tracking[6]. Due to light refraction, absorption, and scattering in underwater scenes, underwater images are usually severely distorted, with low contrast and blurriness [2]. Therefore, clear underwater images play a critical role in fields that need to interact with the underwater environment. The main goal of underwater image enhancement (UIE) is to obtain high-quality images by removing scattering and correcting color distortion in degraded images. UIE is crucial for vision-related underwater tasks." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.273, + 0.885, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.339, + 0.885, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.467, + 0.893, + 0.524 + ], + "angle": 0, + "content": "Figure 1. Our motivations. The amplitude and phase are produced by Fast Fourier Transform (FFT) and the recombined images are obtained by Inverse FFT (IFFT). We further explore the frequency properties for underwater images in Wavelet space." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.539, + 0.894, + 0.622 + ], + "angle": 0, + "content": "Table 1. Evaluation of using different frequency domain transformation strategies on the UIEBD [28]. S1 refers to swapping the amplitude in original images, S2 refers to only swapping the amplitude of low-frequency sub-images in the wavelet space, and S3 refers to swapping the amplitude of low-frequency sub-images and high-frequency sub-images in the wavelet space." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.624, + 0.851, + 0.69 + ], + "angle": 0, + "content": "
StrategyPSNR ↑SSIM ↑LPIPS↓FID↓
S128.680.90270.095728.70
S227.100.88130.102333.04
S329.970.93430.082023.55
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To address this problem, traditional UIE methods based on the physical properties of the underwater images were proposed [15, 17, 29-31]. These methods investigate the physical mechanism of the degradation caused by color cast or scattering and compensate for them to enhance the underwater images. However, these physics-based models with limited representation capacity cannot address all the complex physical and optical factors underlying the underwater scenes, which leads to poor enhancement results under highly complex and diverse underwater scenes. Recently, some learning-based methods [7, 18, 28, 36] for UIE can produce better results, since neural networks have powerful feature representation and nonlinear mapping capabilities." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.407, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author. E-mail: caiwl@njnu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8281" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "ties. It can learn the mapping of an image from degenerate to clear from a substantial quantity of paired training data. However, most previous methods are based on the raw pixel space of images, with limited exploration of the properties of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models for generating high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.609 + ], + "angle": 0, + "content": "Building on insights from previous Fourier-based works [12, 48], we explore the properties of the Fourier frequency information for the UIE task, as illustrated in Figure 1. Given two images (an underwater image and its corresponding ground-truth), we swap their amplitude components and combine them with corresponding phase components in the Fourier space. The recombined results show that the visual appearance is swapped following the amplitude swapping, which indicates the degradation information of underwater images is mainly contained in the amplitude component. We further explore the properties of the amplitude components in the Wavelet space. Specifically, the images can be decomposed into low-frequency sub-images and high-frequency sub-images using discrete wavelet transformations (DWT), and then we swap amplitude components of low-frequency sub-images. From visual results, we can find a similar phenomenon, which means the color degradation information is mainly contained in low-frequency sub-images, and the texture and detail degradation information is mainly contained in high-frequency sub-images. Table 1 shows the quantitative evaluation of the different frequency domain strategies, proving that our discovery is objective. Consequently, how to adequately exploit the properties of frequency domain information and effectively incorporate them into a unified image enhancement network is a crucial issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Recently, diffusion-based methods [10, 35] have garnered significant attention due to their outstanding performance in image synthesis [23, 24, 32, 34, 52] and restoration tasks [5, 40, 46, 51]. These methods rely on a hierarchical denoising autoencoder architecture, enabling them to iteratively reverse a diffusion process and achieve high-quality mapping from randomly sampled Gaussian noise to target images or latent distributions [10]. Tang et al. [36] present an image enhancement approach with diffusion model in underwater scenes. While standard diffusion models exhibit sufficient capability, unforeseen artifacts may arise as a result of the diversity introduced during the sampling process from randomly generated Gaussian noise to images [45]. Furthermore, the diffusion model needs to recover both the high and low-frequency information of images, which limits their ability to focus on fine-grained information, missing out on texture and details. Thereby, it is very crucial that the powerful representation capabilities of diffusion models can be fully utilized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.561 + ], + "angle": 0, + "content": "In this paper, we develop a novel UIE framework to fully exploit the properties of frequency domain information and diffusion models, called WF-Diff, which mainly consists of two stages: frequency preliminary enhancement and frequency diffusion adjustment. The first stage aims to preliminarily enhance the high-frequency and low-frequency components of underwater images by utilizing the frequency domain characteristics. Specifically, we first convert the input images into the wavelet space using discrete wavelet transformations (DWT), obtaining an average coefficient that represents the low-frequency content information of the input image and three high-frequency coefficients that represent sparse vertical, horizontal, and diagonal details of the input image. Then, we design a wavelet-based Fourier information interaction network (WFI2-net), fully integrating the characteristics of Transformer [22] and Fourier prior information to enhance high- and low-frequency content, respectively. Moreover, to achieve interaction of high- and low-frequency information, we propose a cross-frequency conditioner (CFC) to further improve the generation quality. The target of the second stage is to make adjustment to the initial enhanced coarse results in terms of details and textures via diffusion model. Consequently, we propose a frequency residual diffusion adjustment module (FRDAM). Unlike previous diffusion-based work, FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models in the wavelet space, which can not only increase the model's focus on fine-grained information but also mitigate the adverse effects of the diversity of the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.601 + ], + "angle": 0, + "content": "In summary, the main contributions of our method are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.629, + 0.891, + 0.704 + ], + "angle": 0, + "content": "- We explore in depth the properties of the frequency domain for underwater images. Based on the properties and diffusion model, we propose a novel UIE framework, named WF-Diff, with the goal of achieving frequency enhancement and diffusion adjustment." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.705, + 0.892, + 0.78 + ], + "angle": 0, + "content": "- We propose a frequency residual diffusion adjustment module (FRDAM) to further refine the high- and low-frequency information of the initial enhanced images. FRDAM can be viewed as a plug-and-play universal module to adjust the detail of the underwater images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.825 + ], + "angle": 0, + "content": "- We propose a cross-frequency conditioner (CFC) to achieve the cross-frequency interaction of high- and low-frequency information." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Experimental results compared with SOTAs considerably show that our developed WF-Diff performs the superiority against previous UIE approaches, and extensive ablation experiments can demonstrate the effectiveness of our contributions." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.629, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8282" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.878, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.431, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Figure 2. Overall framework of WF-Diff. It contains two detachable networks, Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Furthermore, the proposed cross-frequency conditioner (CFC) aims to achieve the cross-frequency interaction of high- and low-frequency information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.228, + 0.523 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.539, + 0.37, + 0.556 + ], + "angle": 0, + "content": "2.1. Underwater Image Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Currently, existing UID methods can be briefly categorized into the physical and deep model-based approaches [15, 18, 28-30, 36]. Most UID methods based on the physical model utilize prior knowledge to establish models, such as water dark channel priors [29], attenuation curve priors [38], fuzzy priors [4]. In addition, Akkaynak and Treibitz [1] proposed a method based on the revised physical imaging model. However, the depth map of the underwater scene is difficult to obtain. This leads to unstable performance, which usually suffers from severe color cast and artifacts. Therefore, the manually established priors restrain the model's robustness and scalability under the complicated and varied circumstances. Recently, deep learning-based methods [18, 28, 36] have achieved acceptable performance. To alleviate the need for real-world underwater paired training data, many methods introduced GAN-based framework for UIE [7, 14, 21, 49], such as WaterGAN [21], UGAN [7] and UIE-DAL [37]. Recently, some complex frameworks were proposed and achieve the-state-of-the-art performance [15, 29]. Ucolor [19] combined the underwater physical imaging model in the raw space and designed a medium transmission guided model. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[43] proposed a reflected light-aware multi-scale progressive restoration network to obtain images with both color equalization and rich texture in various underwater scenes. Huang et al. [13] proposed a mean teacher based semi-supervised network, which effectively leverages the knowledge from unlabeled data. However, most previous methods are based on the spatial domain, with limited exploration of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.662, + 0.693 + ], + "angle": 0, + "content": "2.2. Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, Diffusion Probabilistic Models (DPMs) [10, 35] have been widely adopted for conditional image generation [5, 40, 42, 46, 50]. Sahara et al. [33] proposed Palette, which has demonstrated the excellent performance of diffusion models in the field of conditional image generation, including colorization, in-painting and JPEG restoration. Tang et al. [36] presented an image enhancement approach with diffusion model in underwater scenes. However, the reverse process starts from randomly sampled Gaussian noise to full images [45], which can lead to unexpected artifacts due to the diversity of the sampling process. Furthermore, the diffusion model needs to recover both the high and low-frequency information in images, which lim" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8283" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "its their ability to focus on fine-grained information. Consequently, how to incorporate diffusion models into a unified underwater image enhancement network is a vital issue." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.153, + 0.212, + 0.17 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.179, + 0.266, + 0.193 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.202, + 0.473, + 0.626 + ], + "angle": 0, + "content": "Given an underwater image as input, our goal is to learn a network to generate an output that eliminates the color cast from input while enhancing the image details. The overall framework of WF-Diff is shown in Figure 2. WF-Diff is designed to fully leverage the characteristics of frequency domain information and powerful ability of diffusion models. Specifically, WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). We first convert the input into the wavelet space using discrete wavelet transformations (DWT), obtaining a low-frequency coefficient and three high-frequency coefficients. WFI2-net is dedicated to achieving preliminary enhancement of frequency information. We fully integrate the characteristics of Transformer and Fourier prior Information, and design wide transformer block (WTB) and spatial-frequency fusion block (SFFB) to enhance high- and low-frequency content respectively. FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Note that, our proposed FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models, respectively. Additionally, the proposed cross-frequency conditioner (CFC) strives to achieve cross-frequency interaction between high- and low-frequency information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.637, + 0.431, + 0.652 + ], + "angle": 0, + "content": "3.2. Discrete Wavelet and Fourier Transform" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.661, + 0.47, + 0.797 + ], + "angle": 0, + "content": "Discrete Wavelet transform (DWT) has been widely applied to low-level vision tasks [11, 16]. We firstly use DWT to decompose an input into multiple frequency sub-bands so that we can achieve the color correction of low-frequency information and detail enhancement of high-frequency information, respectively. Given an underwater image as input \\( I \\in \\mathbb{R}^{H \\times W \\times c} \\), we use DWT with Haar wavelets to decompose the input. Haar wavelets consist of the low-pass filter \\( L \\), and the high-pass filter \\( H \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.809, + 0.469, + 0.842 + ], + "angle": 0, + "content": "\\[\nL = \\frac {1}{\\sqrt {2}} [ 1, 1 ] ^ {T}, H = \\frac {1}{\\sqrt {2}} [ 1, - 1 ] ^ {T}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.853, + 0.469, + 0.881 + ], + "angle": 0, + "content": "We can obtain four sub-bands, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nI _ {L L}, \\left\\{I _ {L H}, I _ {H L}, I _ {H H} \\right\\} = \\mathrm {D W T} (I), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.272 + ], + "angle": 0, + "content": "where \\(I_{LL},\\{I_{LH},I_{HL},I_{HH}\\} \\in \\mathbb{R}^{\\frac{H}{2}\\times \\frac{W}{2}\\times c}\\) represent the low-frequency component of the input and high-frequency components in the vertical, horizontal, and diagonal directions, respectively. More specifically, the low-frequency component contains the content and color information of the input image, and the other three high-frequency coefficients contain details information of global structures and textures [31]. The sub-bands are downsampled to half-resolution of the input but do not result in information loss due to the biorthogonal property of DWT. For low-frequency component \\(I_{LL}\\), we will explore its properties in Fourier space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Then, we introduce the operation of the Fourier transform [48]. Given a image \\( x \\in \\mathbb{R}^{H \\times W \\times 1} \\), whose shape is \\( H \\times W \\), the Fourier transform \\( \\mathcal{F} \\) which converts \\( x \\) to the Fourier space \\( X \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.34, + 0.892, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (x) (u, v) = X (u, v) = \\frac {1}{\\sqrt {H W}} \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} x (h, w) e ^ {- j 2 \\pi \\left(\\frac {h}{H} u + \\frac {w}{W} v\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.457 + ], + "angle": 0, + "content": "where \\(h, w\\) are the coordinates in the spatial space and \\(u, v\\) are the coordinates in the Fourier space. \\(\\mathcal{F}^{-1}\\) denotes the inverse transform of \\(\\mathcal{F}\\). Complex component \\(X(u, v)\\) in the Fourier space can be represented by a amplitude component \\(\\mathcal{A}(X(u, v))\\) and a phase component \\(\\mathcal{P}(X(u, v))\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.464, + 0.892, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} (X (u, v)) = \\sqrt {R ^ {2} (X (u , v)) + I ^ {2} (X (u , v))}, \\tag {4} \\\\ \\mathcal {P} (X (u, v)) = \\arctan [ \\frac {I (X (u , v))}{R (X (u , v))} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.553 + ], + "angle": 0, + "content": "where \\( R(x) \\) and \\( I(x) \\) represent the real and imaginary parts of \\( X(u, v) \\), respectively. Note that, the Fourier operation can be computed alone in each channel for feature maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.644 + ], + "angle": 0, + "content": "According to Figure 1 and Table 1 (our motivation), we conclude that the color degradation information of underwater images is mainly contained in the amplitude component of low-frequency sub-band, and the texture and detail degradation information is mainly contained in high-frequency sub-bands." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.833, + 0.667 + ], + "angle": 0, + "content": "3.3. Frequency Preliminary Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Based on the above analysis, in frequency preliminary enhancement stage, we design a simple but effective WFI2-net with a parallel encoder-decoder (U-Net-like) format to restore the amplitude component of low-frequency information and high-frequency components, respectively. We also utilize skip connections to connect the features at the same level in the encoder and decoder. For high-frequency branch, we utilize advantage of transformer modeling global information to enhance high-frequency coefficients. We design wide transformer block (WTB) using multi-scale information, aiming to model long range dependencies. Our low-frequency branch aims to restore the amplitude component in Fourier space. In order to obtain rich frequency and spatial information, we design spatial-frequency fusion block (SFFB)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8284" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.37, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.583, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.089, + 0.88, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.877, + 0.246 + ], + "angle": 0, + "content": "Figure 3. The architecture of (a) Wide Transformer Block, (b) Spatial-Frequency Fusion Block and (c) Cross-Frequency Conditioner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.254, + 0.473, + 0.405 + ], + "angle": 0, + "content": "Wide Transformer Block. Unlike low-frequency coefficients, high-frequency components contain global structure and texture details. Consequently, the high-frequency branch focuses on modeling global and local features. WTB is shown in Figure 3 (a). Given \\( I_{LH}, I_{HL}, I_{HH} \\in \\mathbb{R}^{\\frac{H}{2} \\times \\frac{W}{2} \\times c} \\), WTB firstly obtains their embedding features \\( T_{in} \\in \\mathbb{R}^{3 \\times \\frac{H}{2} \\times \\frac{W}{2} \\times C} \\) through convolution projection. To be specific, WTB are composed of an attention (Atten) module and a feed-forward network (FFN) module, and the computation can be denoted in the WTB as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.412, + 0.47, + 0.43 + ], + "angle": 0, + "content": "\\[\nQ, K, V, L = \\operatorname {S p l i t} \\left(W _ {d} W _ {p} \\left(\\operatorname {N o r m} \\left(T _ {i - 1}\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.435, + 0.47, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\hat {T} _ {i} = S A (Q, K, V) + C A (L) + T _ {i - 1}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.457, + 0.469, + 0.475 + ], + "angle": 0, + "content": "\\[\nT _ {i} = F F N \\left(\\operatorname {N o r m} \\left(\\hat {T} _ {i}\\right)\\right) + \\hat {T} _ {i}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.568 + ], + "angle": 0, + "content": "where \\( SA \\) and \\( CA \\) refer to self-attention and channel attention, respectively. Norm refers to normalization. \\( T_{i - 1} \\) represents the input embeddings of the current WTB. \\( W_{d} \\) and \\( W_{p} \\) denote \\( 1 \\times 1 \\) point-wise convolution and multi-scale kernel depth-wise convolution, respectively; Split refers to the split operation. \\( L \\) aims to focus on local information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.766 + ], + "angle": 0, + "content": "Spatial-Frequency Fusion Block. We show the structure of SFFB in Figure 3 (b), which has a spatial domain unit (SDU) and a frequency domain unit (FDU) for interaction of dual domain representations. In spatial domain unit, we employ multi-scale convolution kernels in order to enlarge the limited spatial receptive field. After obtaining the spatial embeddings \\( F_{s} \\), we firstly utilize the FFT to obtain the amplitude \\( \\mathcal{A}(F_s) \\) and phase \\( \\mathcal{P}(F_s) \\) components. Then, the \\( \\mathcal{A}(F_s) \\) and \\( \\mathcal{P}(F_s) \\) are fed into two layers \\( 1^{*}1 \\) conv to obtain \\( \\mathcal{A}'(F_s) \\) and \\( \\mathcal{P}'(F_s) \\). Finally, we use the IFFT algorithm to map \\( \\mathcal{A}'(F_s) \\) and \\( \\mathcal{P}'(F_s) \\) to image space and obtain the frequency embeddings \\( F_{f} \\). The fusion embeddings of spatial domain and frequency domain can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.772, + 0.469, + 0.789 + ], + "angle": 0, + "content": "\\[\nF _ {s f} = F _ {s} + F _ {f}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.874 + ], + "angle": 0, + "content": "Loss Function. We denote \\( I_{LL}^{\\prime} \\) as output of low-frequency branch, and \\( I_{LH}^{\\prime}, I_{HL}^{\\prime} \\) and \\( I_{HH}^{\\prime} \\) as output of high-frequency branch. Ground-truth (G) can be decomposed \\( G_{LL}, G_{LH}, G_{HL}, G_{HH} \\) by DWT. The high-frequency loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.878, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} = \\left\\| I _ {(i)} ^ {\\prime} - G _ {(i)} \\right\\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.301 + ], + "angle": 0, + "content": "where \\( i \\in \\{LH, HL, HH\\} \\). For low-frequency information, we only constrain the amplitude components. Consequently, the low-frequency loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.309, + 0.892, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a} = \\left| \\left| \\mathcal {A} \\left(I _ {L L} ^ {\\prime}\\right)\\right) - \\mathcal {A} \\left(G _ {L L}\\right) \\right| | _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.38 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}()\\) refers to the amplitude component in Fourier transform. Finally, We further use an adversarial loss in Wasserstein GAN as reconstruction Loss \\(\\mathcal{L}_{rec}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.388, + 0.771, + 0.404 + ], + "angle": 0, + "content": "3.4. Cross-Frequency Conditioner" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.606 + ], + "angle": 0, + "content": "The detailed structure of CFC is shown in Figure 3 (c). The purpose of CFC is to facilitate information interaction between high and low frequency features, achieving cross-frequency mutual reinforcement between high and low frequency branches, thus enhancing overall representation power. We denote \\( T_{in} \\) and \\( F_{in} \\) as the input features of CFC, which represent high- and low-frequency embeddings. For the high-frequency embedding features \\( T_{in} \\in \\mathbb{R}^{3 \\times \\frac{H}{2} \\times \\frac{W}{2} \\times C} \\), we can obtain \\( T_{LH}, T_{HL}, T_{HH} \\in \\mathbb{R}^{\\frac{H}{2} \\times \\frac{W}{2} \\times c} \\) via split operation. By adding these extracted coefficients, we obtain the aggregated high-frequency embeddings. We use different linear projections to construct \\( Q \\) and \\( K \\) in CFC:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.617, + 0.891, + 0.634 + ], + "angle": 0, + "content": "\\[\nQ = \\operatorname {C o n v} _ {1 \\times 1} \\left(T _ {L H} + T _ {H L} + T _ {H H}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.644, + 0.891, + 0.66 + ], + "angle": 0, + "content": "\\[\nK = \\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {i n}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.664, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Similarly, \\( V_{T} \\) of high-frequency embeddings and \\( V_{F} \\) of low-frequency embeddings can be obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.704, + 0.891, + 0.721 + ], + "angle": 0, + "content": "\\[\nV _ {T} = \\operatorname {C o n v} _ {1 \\times 1} \\left(T _ {L H} + T _ {H L} + T _ {H H}\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.731, + 0.891, + 0.747 + ], + "angle": 0, + "content": "\\[\nV _ {F} = \\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {i n}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.752, + 0.891, + 0.78 + ], + "angle": 0, + "content": "The output feature map \\( T_{out} \\) and \\( F_{out} \\) can then be obtained from the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.787, + 0.891, + 0.822 + ], + "angle": 0, + "content": "\\[\nT _ {o u t} = R \\left(\\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V _ {T}\\right), \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.832, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\[\nF _ {o u t} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V _ {F}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\( R \\) denotes a replication operation, and \\( \\sqrt{d_k} \\) is the number of columns of matrix \\( Q \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8285" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.372, + 0.108 + ], + "angle": 0, + "content": "3.5. Frequency Diffusion Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.469, + 0.249 + ], + "angle": 0, + "content": "FRDAM aims to further adjust the high- and low-frequency information using the powerful representation of diffusion model. Generally, FRDAM can be divided into two branch, namely low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB). We adopt the diffusion process proposed in DDPM [10] to construct the residual distribution of high- and low-frequency information for each branch, which can be described as a forward diffusion process and a reverse diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.469, + 0.358 + ], + "angle": 0, + "content": "Forward Diffusion Process. The forward diffusion process can be viewed as a Markov chain progressively adding Gaussian noise to the data. Given the initial enhanced frequency components \\( I_{i}^{\\prime} \\) and its ground truth \\( G_{i} \\), \\( i \\in \\{LL, LH, HL, HH\\} \\), we calculate their residual distribution \\( x_{0} = G_{i} - I_{i}^{\\prime} \\), then introduce Gaussian noise based on the time step, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.372, + 0.469, + 0.39 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.393, + 0.468, + 0.424 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) is a variable controlling the variance of the noise. Introducing \\(\\alpha_{t} = 1 - \\beta_{t}\\), this process can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.429, + 0.469, + 0.447 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t - 1}, \\quad \\epsilon_ {t - 1} \\sim \\mathcal {N} (0, \\mathcal {Z}). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.453, + 0.469, + 0.468 + ], + "angle": 0, + "content": "With Gaussian distributions are merged, We can obtain :" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.477, + 0.469, + 0.493 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.469, + 0.542 + ], + "angle": 0, + "content": "Reverse Diffusion Process. The reverse diffusion process aims to restore the residual distribution from the Gaussian noise. The reverse diffusion can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.547, + 0.469, + 0.565 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, x _ {c} ^ {(l)}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right), \\sigma_ {t} ^ {2} \\mathcal {Z}\\right), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.573, + 0.469, + 0.651 + ], + "angle": 0, + "content": "where we take the LDFB as an example, and \\(x_{c}^{(l)}\\) refers to the conditional image \\(I_{LL}^{\\prime}\\). \\(\\pmb{\\mu}_{\\theta}(x_t,x_c^{(l)},t)\\) and \\(\\sigma_t^2\\) are the mean and variance from the estimate of step t, respectively. In LDFB and HDFB, we follow the setup of [35], they can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.657, + 0.469, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right) = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\left(1 - \\bar {\\alpha} _ {t}\\right)} \\epsilon_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right)\\right), (21) \\\\ \\sigma_ {t} ^ {2} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}, (22) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.441, + 0.743 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(x_t,x_c^{(l)},t)\\) is the estimated value with a Unet." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.744, + 0.469, + 0.788 + ], + "angle": 0, + "content": "We optimize an objective function for the noise estimated by the network and the noise \\(\\epsilon^{(l)}\\) actually added in LDFB. Therefore, the diffusion loss process is:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.794, + 0.469, + 0.812 + ], + "angle": 0, + "content": "\\[\nL _ {d m} (\\theta) = \\left\\| \\epsilon^ {(l)} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon^ {(l)}, x _ {c} ^ {(l)}, t\\right) \\right\\|. \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.818, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Generally, the frequency diffusion adjustment process is to refine the high- and low-frequency component of the initial enhancement. The whole diffusion process can be formulated:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.883, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {(i)} = \\mathcal {F} _ {H D F B} \\left(\\epsilon_ {s} ^ {(h)}, I _ {(i)} ^ {\\prime}\\right), i \\in \\{L H, H L, H H \\}, \\tag {24}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.093, + 0.891, + 0.111 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {L L} = \\mathcal {F} _ {L D F B} \\left(\\epsilon_ {s} ^ {(l)}, I _ {L L} ^ {\\prime}\\right), \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.117, + 0.892, + 0.15 + ], + "angle": 0, + "content": "where \\(\\epsilon_s^{(h)}\\in \\mathbb{R}^{3\\times \\frac{H}{2}\\times \\frac{W}{2}\\times 3}\\) and \\(\\epsilon_s^{(l)}\\in \\mathbb{R}^{\\frac{H}{2}\\times \\frac{W}{2}\\times 3}\\) are Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.152, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Ultimately, the refined frequency components are obtained as the addition of the diffusion generative residual distribution and the initial enhanced frequency components. Then, we employ IDWT to obtain the final generated image:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.225, + 0.903, + 0.258 + ], + "angle": 0, + "content": "\\[\nI _ {f i n a l} = I D W T \\left(I _ {(i)} ^ {\\prime} + \\hat {I} _ {(i)}, I _ {L L} ^ {\\prime} + \\hat {I} _ {L L}\\right), i \\in \\{L H, H L, H H \\} \\tag {26}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.263, + 0.632, + 0.279 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.582, + 0.304 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Implementation details. Our network, implemented using PyTorch 1.7, underwent training and testing on an NVIDIA GeForce RTX 3090 GPU. We employed the Adam optimizer with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\). The patch size was configured as \\(256 \\times 256\\), and the batch size was set to 2. The diffusion model's total time steps, denoted as T, were set to 1000, and the number of training iterations reached one million. The initial learning rate was established at 0.0001." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.432, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Datasets. We utilize the real-world UIEBD dataset [28] and the LSUI dataset [18] for training and evaluating our model. The UIEBD dataset comprises 890 underwater images with corresponding labels. Out of these, 700 images are allocated for training, and the remaining 190 are designated for testing. The LSUI dataset is randomly partitioned into 4500 images for training and 504 images for testing. In addition, to verify the generalization of WF-Diff, we use non-reference benchmarks U45 [20], which contains 45 underwater images for testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Comparison methods. We conduct a comparative analysis between WF-Diff and eight state-of-the-art (SOTA) UIE methods, namely UIEC^2-Net [39], Water-Net [28], UWCNN [3], SCNet [8], UIEWD [25], U-color [19], U-shape [18], and DM-water [36]. To ensure a fair and rigorous comparison, we utilize the provided source codes from the respective authors and adhere strictly to the identical experimental settings across all evaluations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We primarily utilize well-established full-reference image quality assessment metrics: PSNR and SSIM [41]. PSNR and SSIM offer quantitative comparisons of our method with other approaches at both pixel and structural levels. Higher PSNR and SSIM values signify superior quality of the generated images. Additionally, we incorporate the LPIPS and FID metrics for full-reference image evaluation. LPIPS [47] is a deep neural network-based image quality metric that assesses the perceptual similarity between an image and a reference image. FID [9] measures the distance between the distributions of real and generated images. A lower LPIPS and FID score indicates a more effective UIE approach. For non-reference benchmarks U45," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8286" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of different UIE methods on the UIEBD, LSUI and U45 datasets. The best results are highlighted in bold and the second best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.123, + 0.895, + 0.336 + ], + "angle": 0, + "content": "
MethodsUIEWDUWCNNUIEC^2-NetWater-NetSCNetU-colorU-shapeDM-waterOurs
UIBDFID↓85.1294.4435.0637.4833.6638.2546.1131.0727.85
LPIPS↓0.39560.35250.20330.21160.24970.23370.22640.14360.1248
PSNR↑14.6515.4020.1419.3520.4120.7121.2521.8823.86
SSIM↑0.72650.77490.82150.83210.82350.84110.84530.81940.8730
LSUIFID↓98.49100.534.5138.90158.9945.0628.5627.9126.75
LPIPS↓0.39620.34500.14320.16780.2830.1230.10280.11380.1096
PSNR↑15.4318.2420.8619.7322.6322.9124.1627.6527.26
SSIM↑0.78020.84650.88670.82260.91760.89020.93220.88670.9437
U45UIQM↑2.4582.3792.7802.9572.8563.1043.1513.0863.181
UCIQE↑0.5830.5670.5910.6010.5940.5860.5920.6340.619
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.344, + 0.887, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.652, + 0.746, + 0.668 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with other SOTA methods on real underwater images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.693, + 0.471, + 0.722 + ], + "angle": 0, + "content": "we introduce UIQM [27] and UCIQE [44] to evaluate our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.312, + 0.757 + ], + "angle": 0, + "content": "4.2. Results and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows the quantitative results compared with different baselines on UIEBD, LSUI and U45 datasets, including with UIEC^2-Net, Water-Net, UIEWD, UWCNN, SCNet, U-color, U-shape, and DM-water. We mainly use PSNR, SSIM, LPIPS and FID as our quantitative indices for UIEBD and LSUI datasets, and UIQM and UCIQE for non-reference dataset U45. The results in Table 2 show that our algorithm outperforms state-of-the-art methods obviously, and achieve state-of-the-art performance in terms of image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.895, + 0.83 + ], + "angle": 0, + "content": "quality evaluation metrics on UIE task, which verifies the robustness of the proposed WF-Diff. To better validate the superiority of our methods, in Figure 4, we show the visual results comparison with state-of-the-art methods on real underwater images. The six examples are randomly selected on the UIEBD and LSUI datasets. Our methods consistently generate natural and better visual results on testing images, strongly proving that WF-Diff has good generalization performance for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.655, + 0.862 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Ablation study with WFI2-net. In order to evaluate the effectiveness of each part in WFI2-net, we conduct two ab" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8287" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 3. Ablation study with network structure of WFI2-net on UIEBD dataset." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.129, + 0.44, + 0.264 + ], + "angle": 0, + "content": "
WTBSFFBUIEBD
SACASDUFDUPSNR ↑SSIM↑
×✓✓✓20.940.8541
✓×✓✓20.820.8473
✓✓×✓21.110.8586
✓✓✓×20.230.8346
✓✓✓✓21.870.8622
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.28, + 0.468, + 0.295 + ], + "angle": 0, + "content": "Table 4. Ablation study with loss of WFI2-net on UIEBD dataset." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.306, + 0.44, + 0.445 + ], + "angle": 0, + "content": "
MethodsUIEBD
LhLrecLaCFCPSNR ↑SSIM↑
×✓✓✓20.460.8274
✓×✓✓20.650.8399
✓✓×✓19.810.8213
✓✓✓×20.970.8425
✓✓✓✓21.870.8622
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Table 5. Ablation study with FRDAM on UIEBD dataset. DM refers to diffusion model (DM), RDM refers to residual DM. D-L refers to refining low-frequency component in wavelet space, and D-H refers to refining high-frequency components in wavelet space." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.542, + 0.47, + 0.668 + ], + "angle": 0, + "content": "
MethodDMRDMD-LD-HCFCPSNR↑
A✓××××20.86
B×✓×××22.37
C×✓✓××22.32
D×✓×✓×22.58
E×✓✓✓×23.44
F×✓✓✓✓23.86
" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.686, + 0.174, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.733, + 0.149, + 0.743 + ], + "angle": 0, + "content": "input" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.686, + 0.271, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.733, + 0.235, + 0.742 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.686, + 0.366, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.733, + 0.331, + 0.742 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.686, + 0.465, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.733, + 0.429, + 0.742 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.743, + 0.176, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.79, + 0.137, + 0.798 + ], + "angle": 0, + "content": "D" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.743, + 0.271, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.79, + 0.233, + 0.798 + ], + "angle": 0, + "content": "E" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.743, + 0.367, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.79, + 0.329, + 0.798 + ], + "angle": 0, + "content": "F" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.743, + 0.465, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.79, + 0.431, + 0.798 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.812, + 0.439, + 0.827 + ], + "angle": 0, + "content": "Figure 5. Visual results of ablation study with FRDAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "lation experiments with WFI2-net on the UIEBD dataset in Table 3 and 4. CFC is cross-frequency conditioner. Note that, we do not discuss the effect of our proposed FRDAM" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.09, + 0.838, + 0.104 + ], + "angle": 0, + "content": "Table 6. Comparison results of inference speed." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.109, + 0.879, + 0.143 + ], + "angle": 0, + "content": "
MethodSCNetUshapeDM-waterWFI2-netWF-Diff
Inference time (s)0.01490.03530.14410.07390.2816
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.153, + 0.812, + 0.166 + ], + "angle": 0, + "content": "Table 7. Comparison results of FLOPs." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.168, + 0.879, + 0.205 + ], + "angle": 0, + "content": "
MethodSCNetWaterNetUcolorUshapeWFI2-net
FLOPs (G)5.88193.7443.866.294.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.227, + 0.892, + 0.287 + ], + "angle": 0, + "content": "here. We regularly remove one component to each configuration at one time, and our strategy achieves the best performance by using all loss functions and blocks, proving that each part of WFI2-net is useful for UIE task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.893, + 0.548 + ], + "angle": 0, + "content": "Ablation study with FRDAM. In this section, we will discuss the effectiveness of FRDAM. Table 5 shows the quantitative results on the UIEBD dataset, and Figure 5 shows the visual results. Note that, Model A and B achieve diffusion process in the pixel level, and model C, D, E and E achieve diffusion process in wavelet space. Model A obtains relatively poor results in PSNR and generates images with color distortion or artifacts in Figure 5 due to the diversity of the sampling process. Model B does not yield fully satisfactory results, because it needs to adjust both the high and low-frequency information in images, which limits their ability to focus on fine-grained information. Compared to model C, model D achieves better results, suggesting that the degradation information is mainly in the high-frequency information in frequency diffusion adjustment stage. Model F achieves the best performance, proving that our designed FRDAM is best for UIE task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.562, + 0.618, + 0.577 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.713 + ], + "angle": 0, + "content": "In this paper, we develop a novel UIE framework, namely WF-Diff. With fully utilizing the frequency domain characteristics and diffusion model, WFI2-net can achieve enhancement and adjustment of frequency information. Our proposed FRDAM is a plug-and-play universal module to adjust the details of the underwater images. WF-Diff shows SOTA performance on UIE task, and extensive ablation experiments prove that each of our contributions is effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Limitations. As a result of employing two diffusion models, our approach doesn't confer an advantage in terms of inference speed. Table 6 presents the actual values of inference speed for WF-Diff, indicating that our model does not outperform recent approaches in terms of inference time. Furthermore, Table 7 provides a comparison of FLOPs for WFI2-net. It should be mentioned that the implicit sampling step is set to 10 for WF-Diff. Hence, in the future, we will delve into methods to expedite the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62276138 and 62371232)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8288" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Derya Akkaynak and Tali Treibitz. Sea-thru: A method for removing water from underwater images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1682-1691. Computer Vision Foundation / IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Derya Akkaynak, Tali Treibitz, Tom Shlesinger, Yossi Loya, Raz Tamir, and David Iluz. What is the space of attenuation coefficients in underwater computer vision? In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 568-577, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Saeed Anwar, Chongyi Li, and Fatih Porikli. Deep underwater image enhancement. CoRR, abs/1807.03528, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] John Yi-Wu Chiang and Ying-Ching Chen. Underwater image enhancement by wavelength compensation and dehazing. IEEE Trans. Image Process., 21(4):1756-1769, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14347-14356. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[6] Karin de Langis and Junaed Sattar. Realtime multi-diver tracking and re-identification for underwater human-robot collaboration. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 11140-11146, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[7] Cameron Fabbri, Md Jahidul Islam, and Junaed Sattar. Enhancing underwater imagery using generative adversarial networks. In 2018 IEEE International Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May 21-25, 2018, pages 7159-7165. IEEE, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[8] Zhenqi Fu, Xiaopeng Lin, Wu Wang, Yue Huang, and Xinghao Ding. Underwater image enhancement via learning water type desensitized representations. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2764-2768. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based CNN for multi-scale face super resolution. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1698-1706, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[12] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. pages 163-180, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[13] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18145-18155. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[14] Md Jahidul Islam, Youya Xia, and Junaed Sattar. Fast underwater image enhancement for improved visual perception. IEEE Robotics Autom. Lett., 5(2):3227-3234, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[15] Paulo Drews Jr., Erickson Rangel do Nascimento, F. Moraes, Silvia S. C. Botelho, and Mario F. M. Campos. Transmission estimation in underwater single images. In 2013 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 825-830, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[16] Eunhee Kang, Won Chang, Jae Jun Yoo, and Jong Chul Ye. Deep convolutional framelet denoising for low-dose CT via wavelet residual network. IEEE Trans. Medical Imaging, 37 (6):1358-1369, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[17] Chongyi Li, Jichang Guo, Runmin Cong, Yanwei Pang, and Bo Wang. Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior. IEEE Trans. Image Process., 25(12):5664-5677, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[18] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Trans. Image Process., 29:4376-4389, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[19] Chongyi Li, Saeed Anwar, Junhui Hou, Runmin Cong, Chunle Guo, and Wenqi Ren. Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE Trans. Image Process., 30:4985-5000, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[20] Hanyu Li, Jingjing Li, and Wei Wang. A fusion adversarial underwater image enhancement network with a public test dataset. arXiv preprint arXiv:1906.06819, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[21] Jie Li, Katherine A. Skinner, Ryan M. Eustice, and Matthew Johnson-Roberson. Watergan: Unsupervised generative network to enable real-time color correction of monocular underwater images. IEEE Robotics Autom. Lett., 3(1):387-394, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[23] Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-Icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8289" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Ziyin Ma and Changjae Oh. A wavelet-based dual-stream network for underwater image enhancement. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2769-2773. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[26] James McMahon and Erion Plaku. Autonomous data collection with timed communication constraints for unmanned underwater vehicles. IEEE Robotics Autom. Lett., 6(2): 1832-1839, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[27] Karen Panetta, Chen Gao, and Sos Agaian. Human-visual-system-inspired underwater image quality measures. IEEE Journal of Oceanic Engineering, 41(3):541-551, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[28] Lintao Peng, Chunli Zhu, and Liheng Bian. U-shape transformer for underwater image enhancement. IEEE Trans. Image Process., 32:3066-3079, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[29] Yan-Tsung Peng and Pamela C. Cosman. Underwater image restoration based on image blurriness and light absorption. IEEE Trans. Image Process., 26(4):1579-1594, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[30] Yan-Tsung Peng, Keming Cao, and Pamela C. Cosman. Generalization of the dark channel prior for single image restoration. IEEE Trans. Image Process., 27(6):2856-2868, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.47, + 0.462 + ], + "angle": 0, + "content": "[31] Priyadharsini Ravisankar, T. Sree Sharmila, and V. Rajendran. A wavelet transform based contrast enhancement method for underwater acoustic images. Multidimens. Syst. Signal Process., 29(4):1845-1859, 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 15:1-15:10. ACM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[36] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 5419-5427. ACM, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[37] Pritish M. Uplavikar, Zhenyu Wu, and Zhangyang Wang. All-in-one underwater image enhancement using domain" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1-8, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[38] Yi Wang, Hui Liu, and Lap-Pui Chau. Single underwater image restoration using adaptive attenuation-curve prior. IEEE Trans. Circuits Syst. I Regul. Pap., 65-I(3):992-1002, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[39] Yudong Wang, Jichang Guo, Huan Gao, and HuiHui Yue. Uiec\\(^2\\) -net: Cnn-based underwater image enhancement using two color space. Signal Process. Image Commun., 96: 116250, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[40] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[41] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[42] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16272-16282. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[43] Jian Yang, Chen Li, and Xuelong Li. Underwater image restoration with light-aware progressive network. In IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023, pages 1-5. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.54, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[44] Miao Yang and Arcot Sowmya. An underwater color image quality evaluation metric. IEEE Transactions on Image Processing, 24(12):6062-6071, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[45] Zongyuan Yang, Baolin Liu, Yongping Xiong, Lan Yi, Guibin Wu, Xiaojun Tang, Ziqi Liu, Junjie Zhou, and Xing Zhang. Docdiff: Document enhancement via residual diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 2795-2806. ACM, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[46] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with A generative diffusion model. CoRR, abs/2308.13164, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[48] Chen Zhao, Wei-Ling Cai, Chenyu Dong, and Ziqi Zeng. Toward sufficient spatial-frequency interaction for gradient-aware underwater image enhancement. arXiv preprint arXiv:2202.08537, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Chen Zhao, Wei-Ling Cai, and Zheng Yuan. Spectral normalization and dual contrastive regularization for image-to" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8290" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "image translation. The Visual Computer, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[50] Chen Zhao, Chenyu Dong, and Weiling Cai. Learning a physical-aware diffusion model based on transformer for underwater image enhancement. arXiv preprint arXiv:2403.01497, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[51] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[52] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. arXiv preprint arXiv:2402.05408, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "8291" + } + ] +] \ No newline at end of file diff --git a/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_origin.pdf b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c953af543f87389d2a69c783c5089bc00baf1039 --- /dev/null +++ b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/f627431b-e7bf-434f-9f57-1c74972ca141_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a19bb69530343d7a8cce1d360faaf356e3c34b2752857229df747fedd4d05027 +size 9689237 diff --git a/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/full.md b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..065a10194a333105c5d3528e468aa78df592b3fd --- /dev/null +++ b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/full.md @@ -0,0 +1,385 @@ +# Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration + +Chen Zhao, Weiling Cai*, Chenyu Dong, Chengwei Hu +School of Artificial Intelligence, Nanjing Normal University + +# Abstract + +Underwater images are subject to intricate and diverse degradation, inevitably affecting the effectiveness of underwater visual tasks. However, most approaches primarily operate in the raw pixel space of images, which limits the exploration of the frequency characteristics of underwater images, leading to an inadequate utilization of deep models' representational capabilities in producing high-quality images. In this paper, we introduce a novel Underwater Image Enhancement (UIE) framework, named WF-Diff, designed to fully leverage the characteristics of frequency domain information and diffusion models. WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). With our full exploration of the frequency domain information, WFI2-net aims to achieve preliminary enhancement of frequency information in the wavelet space. Our proposed FRDAM can further refine the high- and low-frequency information of the initial enhanced images, which can be viewed as a plug-and-play universal module to adjust the detail of the underwater images. With the above techniques, our algorithm can show SOTA performance on real-world underwater image datasets, and achieves competitive performance in visual quality. The code is available at https://github.com/zhihefang/WF-Diff. + +# 1. Introduction + +Underwater image restoration is a practical but challenging technology in the field of underwater vision, widely used for tasks, such as underwater robotics[26] and underwater object tracking[6]. Due to light refraction, absorption, and scattering in underwater scenes, underwater images are usually severely distorted, with low contrast and blurriness [2]. Therefore, clear underwater images play a critical role in fields that need to interact with the underwater environment. The main goal of underwater image enhancement (UIE) is to obtain high-quality images by removing scattering and correcting color distortion in degraded images. UIE is crucial for vision-related underwater tasks. + +![](images/2c2150c3ba66b3750da5912ea5029f6940c1e7309121d304299829a7ddac99ef.jpg) + +![](images/6996da4f8487b3b1d29fd489b0beca75cc692d0ad32c9730e1ebf611aba35ad1.jpg) +Figure 1. Our motivations. The amplitude and phase are produced by Fast Fourier Transform (FFT) and the recombined images are obtained by Inverse FFT (IFFT). We further explore the frequency properties for underwater images in Wavelet space. + +Table 1. Evaluation of using different frequency domain transformation strategies on the UIEBD [28]. S1 refers to swapping the amplitude in original images, S2 refers to only swapping the amplitude of low-frequency sub-images in the wavelet space, and S3 refers to swapping the amplitude of low-frequency sub-images and high-frequency sub-images in the wavelet space. + +
StrategyPSNR ↑SSIM ↑LPIPS↓FID↓
S128.680.90270.095728.70
S227.100.88130.102333.04
S329.970.93430.082023.55
+ +To address this problem, traditional UIE methods based on the physical properties of the underwater images were proposed [15, 17, 29-31]. These methods investigate the physical mechanism of the degradation caused by color cast or scattering and compensate for them to enhance the underwater images. However, these physics-based models with limited representation capacity cannot address all the complex physical and optical factors underlying the underwater scenes, which leads to poor enhancement results under highly complex and diverse underwater scenes. Recently, some learning-based methods [7, 18, 28, 36] for UIE can produce better results, since neural networks have powerful feature representation and nonlinear mapping capabilities. + +ties. It can learn the mapping of an image from degenerate to clear from a substantial quantity of paired training data. However, most previous methods are based on the raw pixel space of images, with limited exploration of the properties of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models for generating high-quality images. + +Building on insights from previous Fourier-based works [12, 48], we explore the properties of the Fourier frequency information for the UIE task, as illustrated in Figure 1. Given two images (an underwater image and its corresponding ground-truth), we swap their amplitude components and combine them with corresponding phase components in the Fourier space. The recombined results show that the visual appearance is swapped following the amplitude swapping, which indicates the degradation information of underwater images is mainly contained in the amplitude component. We further explore the properties of the amplitude components in the Wavelet space. Specifically, the images can be decomposed into low-frequency sub-images and high-frequency sub-images using discrete wavelet transformations (DWT), and then we swap amplitude components of low-frequency sub-images. From visual results, we can find a similar phenomenon, which means the color degradation information is mainly contained in low-frequency sub-images, and the texture and detail degradation information is mainly contained in high-frequency sub-images. Table 1 shows the quantitative evaluation of the different frequency domain strategies, proving that our discovery is objective. Consequently, how to adequately exploit the properties of frequency domain information and effectively incorporate them into a unified image enhancement network is a crucial issue. + +Recently, diffusion-based methods [10, 35] have garnered significant attention due to their outstanding performance in image synthesis [23, 24, 32, 34, 52] and restoration tasks [5, 40, 46, 51]. These methods rely on a hierarchical denoising autoencoder architecture, enabling them to iteratively reverse a diffusion process and achieve high-quality mapping from randomly sampled Gaussian noise to target images or latent distributions [10]. Tang et al. [36] present an image enhancement approach with diffusion model in underwater scenes. While standard diffusion models exhibit sufficient capability, unforeseen artifacts may arise as a result of the diversity introduced during the sampling process from randomly generated Gaussian noise to images [45]. Furthermore, the diffusion model needs to recover both the high and low-frequency information of images, which limits their ability to focus on fine-grained information, missing out on texture and details. Thereby, it is very crucial that the powerful representation capabilities of diffusion models can be fully utilized. + +In this paper, we develop a novel UIE framework to fully exploit the properties of frequency domain information and diffusion models, called WF-Diff, which mainly consists of two stages: frequency preliminary enhancement and frequency diffusion adjustment. The first stage aims to preliminarily enhance the high-frequency and low-frequency components of underwater images by utilizing the frequency domain characteristics. Specifically, we first convert the input images into the wavelet space using discrete wavelet transformations (DWT), obtaining an average coefficient that represents the low-frequency content information of the input image and three high-frequency coefficients that represent sparse vertical, horizontal, and diagonal details of the input image. Then, we design a wavelet-based Fourier information interaction network (WFI2-net), fully integrating the characteristics of Transformer [22] and Fourier prior information to enhance high- and low-frequency content, respectively. Moreover, to achieve interaction of high- and low-frequency information, we propose a cross-frequency conditioner (CFC) to further improve the generation quality. The target of the second stage is to make adjustment to the initial enhanced coarse results in terms of details and textures via diffusion model. Consequently, we propose a frequency residual diffusion adjustment module (FRDAM). Unlike previous diffusion-based work, FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models in the wavelet space, which can not only increase the model's focus on fine-grained information but also mitigate the adverse effects of the diversity of the sampling process. + +In summary, the main contributions of our method are as follows: + +- We explore in depth the properties of the frequency domain for underwater images. Based on the properties and diffusion model, we propose a novel UIE framework, named WF-Diff, with the goal of achieving frequency enhancement and diffusion adjustment. +- We propose a frequency residual diffusion adjustment module (FRDAM) to further refine the high- and low-frequency information of the initial enhanced images. FRDAM can be viewed as a plug-and-play universal module to adjust the detail of the underwater images. +- We propose a cross-frequency conditioner (CFC) to achieve the cross-frequency interaction of high- and low-frequency information. +- Experimental results compared with SOTAs considerably show that our developed WF-Diff performs the superiority against previous UIE approaches, and extensive ablation experiments can demonstrate the effectiveness of our contributions. + +![](images/52e27cbb5d2ffc143f2fe469b95c68d819141d717404f60deb1e5cd2dc613def.jpg) +Figure 2. Overall framework of WF-Diff. It contains two detachable networks, Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Furthermore, the proposed cross-frequency conditioner (CFC) aims to achieve the cross-frequency interaction of high- and low-frequency information. + +# 2. Related Works + +# 2.1. Underwater Image Enhancement + +Currently, existing UID methods can be briefly categorized into the physical and deep model-based approaches [15, 18, 28-30, 36]. Most UID methods based on the physical model utilize prior knowledge to establish models, such as water dark channel priors [29], attenuation curve priors [38], fuzzy priors [4]. In addition, Akkaynak and Treibitz [1] proposed a method based on the revised physical imaging model. However, the depth map of the underwater scene is difficult to obtain. This leads to unstable performance, which usually suffers from severe color cast and artifacts. Therefore, the manually established priors restrain the model's robustness and scalability under the complicated and varied circumstances. Recently, deep learning-based methods [18, 28, 36] have achieved acceptable performance. To alleviate the need for real-world underwater paired training data, many methods introduced GAN-based framework for UIE [7, 14, 21, 49], such as WaterGAN [21], UGAN [7] and UIE-DAL [37]. Recently, some complex frameworks were proposed and achieve the-state-of-the-art performance [15, 29]. Ucolor [19] combined the underwater physical imaging model in the raw space and designed a medium transmission guided model. Yang et al. + +[43] proposed a reflected light-aware multi-scale progressive restoration network to obtain images with both color equalization and rich texture in various underwater scenes. Huang et al. [13] proposed a mean teacher based semi-supervised network, which effectively leverages the knowledge from unlabeled data. However, most previous methods are based on the spatial domain, with limited exploration of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models. + +# 2.2. Diffusion Model + +Recently, Diffusion Probabilistic Models (DPMs) [10, 35] have been widely adopted for conditional image generation [5, 40, 42, 46, 50]. Sahara et al. [33] proposed Palette, which has demonstrated the excellent performance of diffusion models in the field of conditional image generation, including colorization, in-painting and JPEG restoration. Tang et al. [36] presented an image enhancement approach with diffusion model in underwater scenes. However, the reverse process starts from randomly sampled Gaussian noise to full images [45], which can lead to unexpected artifacts due to the diversity of the sampling process. Furthermore, the diffusion model needs to recover both the high and low-frequency information in images, which lim + +its their ability to focus on fine-grained information. Consequently, how to incorporate diffusion models into a unified underwater image enhancement network is a vital issue. + +# 3. Methodology + +# 3.1. Overall Framework + +Given an underwater image as input, our goal is to learn a network to generate an output that eliminates the color cast from input while enhancing the image details. The overall framework of WF-Diff is shown in Figure 2. WF-Diff is designed to fully leverage the characteristics of frequency domain information and powerful ability of diffusion models. Specifically, WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). We first convert the input into the wavelet space using discrete wavelet transformations (DWT), obtaining a low-frequency coefficient and three high-frequency coefficients. WFI2-net is dedicated to achieving preliminary enhancement of frequency information. We fully integrate the characteristics of Transformer and Fourier prior Information, and design wide transformer block (WTB) and spatial-frequency fusion block (SFFB) to enhance high- and low-frequency content respectively. FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Note that, our proposed FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models, respectively. Additionally, the proposed cross-frequency conditioner (CFC) strives to achieve cross-frequency interaction between high- and low-frequency information. + +# 3.2. Discrete Wavelet and Fourier Transform + +Discrete Wavelet transform (DWT) has been widely applied to low-level vision tasks [11, 16]. We firstly use DWT to decompose an input into multiple frequency sub-bands so that we can achieve the color correction of low-frequency information and detail enhancement of high-frequency information, respectively. Given an underwater image as input $I \in \mathbb{R}^{H \times W \times c}$ , we use DWT with Haar wavelets to decompose the input. Haar wavelets consist of the low-pass filter $L$ , and the high-pass filter $H$ , as follows: + +$$ +L = \frac {1}{\sqrt {2}} [ 1, 1 ] ^ {T}, H = \frac {1}{\sqrt {2}} [ 1, - 1 ] ^ {T}. \tag {1} +$$ + +We can obtain four sub-bands, which can be expressed as: + +$$ +I _ {L L}, \left\{I _ {L H}, I _ {H L}, I _ {H H} \right\} = \mathrm {D W T} (I), \tag {2} +$$ + +where $I_{LL},\{I_{LH},I_{HL},I_{HH}\} \in \mathbb{R}^{\frac{H}{2}\times \frac{W}{2}\times c}$ represent the low-frequency component of the input and high-frequency components in the vertical, horizontal, and diagonal directions, respectively. More specifically, the low-frequency component contains the content and color information of the input image, and the other three high-frequency coefficients contain details information of global structures and textures [31]. The sub-bands are downsampled to half-resolution of the input but do not result in information loss due to the biorthogonal property of DWT. For low-frequency component $I_{LL}$ , we will explore its properties in Fourier space. + +Then, we introduce the operation of the Fourier transform [48]. Given a image $x \in \mathbb{R}^{H \times W \times 1}$ , whose shape is $H \times W$ , the Fourier transform $\mathcal{F}$ which converts $x$ to the Fourier space $X$ can be expressed as: + +$$ +\mathcal {F} (x) (u, v) = X (u, v) = \frac {1}{\sqrt {H W}} \sum_ {h = 0} ^ {H - 1} \sum_ {w = 0} ^ {W - 1} x (h, w) e ^ {- j 2 \pi \left(\frac {h}{H} u + \frac {w}{W} v\right)}, \tag {3} +$$ + +where $h, w$ are the coordinates in the spatial space and $u, v$ are the coordinates in the Fourier space. $\mathcal{F}^{-1}$ denotes the inverse transform of $\mathcal{F}$ . Complex component $X(u, v)$ in the Fourier space can be represented by a amplitude component $\mathcal{A}(X(u, v))$ and a phase component $\mathcal{P}(X(u, v))$ as follows: + +$$ +\begin{array}{l} \mathcal {A} (X (u, v)) = \sqrt {R ^ {2} (X (u , v)) + I ^ {2} (X (u , v))}, \tag {4} \\ \mathcal {P} (X (u, v)) = \arctan [ \frac {I (X (u , v))}{R (X (u , v))} ], \\ \end{array} +$$ + +where $R(x)$ and $I(x)$ represent the real and imaginary parts of $X(u, v)$ , respectively. Note that, the Fourier operation can be computed alone in each channel for feature maps. + +According to Figure 1 and Table 1 (our motivation), we conclude that the color degradation information of underwater images is mainly contained in the amplitude component of low-frequency sub-band, and the texture and detail degradation information is mainly contained in high-frequency sub-bands. + +# 3.3. Frequency Preliminary Enhancement + +Based on the above analysis, in frequency preliminary enhancement stage, we design a simple but effective WFI2-net with a parallel encoder-decoder (U-Net-like) format to restore the amplitude component of low-frequency information and high-frequency components, respectively. We also utilize skip connections to connect the features at the same level in the encoder and decoder. For high-frequency branch, we utilize advantage of transformer modeling global information to enhance high-frequency coefficients. We design wide transformer block (WTB) using multi-scale information, aiming to model long range dependencies. Our low-frequency branch aims to restore the amplitude component in Fourier space. In order to obtain rich frequency and spatial information, we design spatial-frequency fusion block (SFFB). + +![](images/07874ab2c66a9096f5c0e889f10ace59987f12d06afcb6dc0d31d288e9b4ee2a.jpg) +Figure 3. The architecture of (a) Wide Transformer Block, (b) Spatial-Frequency Fusion Block and (c) Cross-Frequency Conditioner. + +![](images/fb4f1f13acda6bcadf098ccf3ecd3da65d1d627b237d70e1fba648a071ef9a6c.jpg) + +![](images/ef6e42c87c0984218dcdfbe2c25bb5a1752db5a16a1f33f71d4edb2bf1741b80.jpg) + +Wide Transformer Block. Unlike low-frequency coefficients, high-frequency components contain global structure and texture details. Consequently, the high-frequency branch focuses on modeling global and local features. WTB is shown in Figure 3 (a). Given $I_{LH}, I_{HL}, I_{HH} \in \mathbb{R}^{\frac{H}{2} \times \frac{W}{2} \times c}$ , WTB firstly obtains their embedding features $T_{in} \in \mathbb{R}^{3 \times \frac{H}{2} \times \frac{W}{2} \times C}$ through convolution projection. To be specific, WTB are composed of an attention (Atten) module and a feed-forward network (FFN) module, and the computation can be denoted in the WTB as: + +$$ +Q, K, V, L = \operatorname {S p l i t} \left(W _ {d} W _ {p} \left(\operatorname {N o r m} \left(T _ {i - 1}\right)\right)\right), \tag {5} +$$ + +$$ +\hat {T} _ {i} = S A (Q, K, V) + C A (L) + T _ {i - 1}, \tag {6} +$$ + +$$ +T _ {i} = F F N \left(\operatorname {N o r m} \left(\hat {T} _ {i}\right)\right) + \hat {T} _ {i}, \tag {7} +$$ + +where $SA$ and $CA$ refer to self-attention and channel attention, respectively. Norm refers to normalization. $T_{i - 1}$ represents the input embeddings of the current WTB. $W_{d}$ and $W_{p}$ denote $1 \times 1$ point-wise convolution and multi-scale kernel depth-wise convolution, respectively; Split refers to the split operation. $L$ aims to focus on local information. + +Spatial-Frequency Fusion Block. We show the structure of SFFB in Figure 3 (b), which has a spatial domain unit (SDU) and a frequency domain unit (FDU) for interaction of dual domain representations. In spatial domain unit, we employ multi-scale convolution kernels in order to enlarge the limited spatial receptive field. After obtaining the spatial embeddings $F_{s}$ , we firstly utilize the FFT to obtain the amplitude $\mathcal{A}(F_s)$ and phase $\mathcal{P}(F_s)$ components. Then, the $\mathcal{A}(F_s)$ and $\mathcal{P}(F_s)$ are fed into two layers $1^{*}1$ conv to obtain $\mathcal{A}'(F_s)$ and $\mathcal{P}'(F_s)$ . Finally, we use the IFFT algorithm to map $\mathcal{A}'(F_s)$ and $\mathcal{P}'(F_s)$ to image space and obtain the frequency embeddings $F_{f}$ . The fusion embeddings of spatial domain and frequency domain can be expressed as: + +$$ +F _ {s f} = F _ {s} + F _ {f}. \tag {8} +$$ + +Loss Function. We denote $I_{LL}^{\prime}$ as output of low-frequency branch, and $I_{LH}^{\prime}, I_{HL}^{\prime}$ and $I_{HH}^{\prime}$ as output of high-frequency branch. Ground-truth (G) can be decomposed $G_{LL}, G_{LH}, G_{HL}, G_{HH}$ by DWT. The high-frequency loss can be expressed as: + +$$ +\mathcal {L} _ {h} = \left\| I _ {(i)} ^ {\prime} - G _ {(i)} \right\| _ {2}, \tag {9} +$$ + +where $i \in \{LH, HL, HH\}$ . For low-frequency information, we only constrain the amplitude components. Consequently, the low-frequency loss can be expressed as: + +$$ +\mathcal {L} _ {a} = \left| \left| \mathcal {A} \left(I _ {L L} ^ {\prime}\right)\right) - \mathcal {A} \left(G _ {L L}\right) \right| | _ {1}, \tag {10} +$$ + +where $\mathcal{A}()$ refers to the amplitude component in Fourier transform. Finally, We further use an adversarial loss in Wasserstein GAN as reconstruction Loss $\mathcal{L}_{rec}$ . + +# 3.4. Cross-Frequency Conditioner + +The detailed structure of CFC is shown in Figure 3 (c). The purpose of CFC is to facilitate information interaction between high and low frequency features, achieving cross-frequency mutual reinforcement between high and low frequency branches, thus enhancing overall representation power. We denote $T_{in}$ and $F_{in}$ as the input features of CFC, which represent high- and low-frequency embeddings. For the high-frequency embedding features $T_{in} \in \mathbb{R}^{3 \times \frac{H}{2} \times \frac{W}{2} \times C}$ , we can obtain $T_{LH}, T_{HL}, T_{HH} \in \mathbb{R}^{\frac{H}{2} \times \frac{W}{2} \times c}$ via split operation. By adding these extracted coefficients, we obtain the aggregated high-frequency embeddings. We use different linear projections to construct $Q$ and $K$ in CFC: + +$$ +Q = \operatorname {C o n v} _ {1 \times 1} \left(T _ {L H} + T _ {H L} + T _ {H H}\right), \tag {11} +$$ + +$$ +K = \operatorname {C o n v} _ {1 \times 1} \left(F _ {i n}\right). \tag {12} +$$ + +Similarly, $V_{T}$ of high-frequency embeddings and $V_{F}$ of low-frequency embeddings can be obtained: + +$$ +V _ {T} = \operatorname {C o n v} _ {1 \times 1} \left(T _ {L H} + T _ {H L} + T _ {H H}\right), \tag {13} +$$ + +$$ +V _ {F} = \operatorname {C o n v} _ {1 \times 1} \left(F _ {i n}\right). \tag {14} +$$ + +The output feature map $T_{out}$ and $F_{out}$ can then be obtained from the formula: + +$$ +T _ {o u t} = R \left(\operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V _ {T}\right), \tag {15} +$$ + +$$ +F _ {o u t} = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V _ {F}, \tag {16} +$$ + +where $R$ denotes a replication operation, and $\sqrt{d_k}$ is the number of columns of matrix $Q$ . + +# 3.5. Frequency Diffusion Adjustment + +FRDAM aims to further adjust the high- and low-frequency information using the powerful representation of diffusion model. Generally, FRDAM can be divided into two branch, namely low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB). We adopt the diffusion process proposed in DDPM [10] to construct the residual distribution of high- and low-frequency information for each branch, which can be described as a forward diffusion process and a reverse diffusion process. + +Forward Diffusion Process. The forward diffusion process can be viewed as a Markov chain progressively adding Gaussian noise to the data. Given the initial enhanced frequency components $I_{i}^{\prime}$ and its ground truth $G_{i}$ , $i \in \{LL, LH, HL, HH\}$ , we calculate their residual distribution $x_{0} = G_{i} - I_{i}^{\prime}$ , then introduce Gaussian noise based on the time step, as follows: + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} I\right), \tag {17} +$$ + +where $\beta_{t}$ is a variable controlling the variance of the noise. Introducing $\alpha_{t} = 1 - \beta_{t}$ , this process can be described as: + +$$ +x _ {t} = \sqrt {\alpha_ {t}} x _ {t - 1} + \sqrt {1 - \alpha_ {t}} \epsilon_ {t - 1}, \quad \epsilon_ {t - 1} \sim \mathcal {N} (0, \mathcal {Z}). \tag {18} +$$ + +With Gaussian distributions are merged, We can obtain : + +$$ +q \left(x _ {t} \mid x _ {0}\right) = \mathcal {N} \left(x _ {t}; \sqrt {\bar {\alpha} _ {t}} x _ {0}, (1 - \bar {\alpha} _ {t}) I\right). \tag {19} +$$ + +Reverse Diffusion Process. The reverse diffusion process aims to restore the residual distribution from the Gaussian noise. The reverse diffusion can be expressed as: + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, x _ {c} ^ {(l)}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, x _ {c} ^ {(l)}, t\right), \sigma_ {t} ^ {2} \mathcal {Z}\right), \tag {20} +$$ + +where we take the LDFB as an example, and $x_{c}^{(l)}$ refers to the conditional image $I_{LL}^{\prime}$ . $\pmb{\mu}_{\theta}(x_t,x_c^{(l)},t)$ and $\sigma_t^2$ are the mean and variance from the estimate of step t, respectively. In LDFB and HDFB, we follow the setup of [35], they can be expressed as: + +$$ +\begin{array}{l} \mu_ {\theta} \left(x _ {t}, x _ {c} ^ {(l)}, t\right) = \frac {1}{\sqrt {\alpha} _ {t}} \left(x _ {t} - \frac {\beta_ {t}}{\left(1 - \bar {\alpha} _ {t}\right)} \epsilon_ {\theta} \left(x _ {t}, x _ {c} ^ {(l)}, t\right)\right), (21) \\ \sigma_ {t} ^ {2} = \frac {1 - \bar {\alpha} _ {t - 1}}{1 - \bar {\alpha} _ {t}} \beta_ {t}, (22) \\ \end{array} +$$ + +where $\epsilon_{\theta}(x_t,x_c^{(l)},t)$ is the estimated value with a Unet. + +We optimize an objective function for the noise estimated by the network and the noise $\epsilon^{(l)}$ actually added in LDFB. Therefore, the diffusion loss process is: + +$$ +L _ {d m} (\theta) = \left\| \epsilon^ {(l)} - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon^ {(l)}, x _ {c} ^ {(l)}, t\right) \right\|. \tag {23} +$$ + +Generally, the frequency diffusion adjustment process is to refine the high- and low-frequency component of the initial enhancement. The whole diffusion process can be formulated: + +$$ +\hat {I} _ {(i)} = \mathcal {F} _ {H D F B} \left(\epsilon_ {s} ^ {(h)}, I _ {(i)} ^ {\prime}\right), i \in \{L H, H L, H H \}, \tag {24} +$$ + +$$ +\hat {I} _ {L L} = \mathcal {F} _ {L D F B} \left(\epsilon_ {s} ^ {(l)}, I _ {L L} ^ {\prime}\right), \tag {25} +$$ + +where $\epsilon_s^{(h)}\in \mathbb{R}^{3\times \frac{H}{2}\times \frac{W}{2}\times 3}$ and $\epsilon_s^{(l)}\in \mathbb{R}^{\frac{H}{2}\times \frac{W}{2}\times 3}$ are Gaussian noise. + +Ultimately, the refined frequency components are obtained as the addition of the diffusion generative residual distribution and the initial enhanced frequency components. Then, we employ IDWT to obtain the final generated image: + +$$ +I _ {f i n a l} = I D W T \left(I _ {(i)} ^ {\prime} + \hat {I} _ {(i)}, I _ {L L} ^ {\prime} + \hat {I} _ {L L}\right), i \in \{L H, H L, H H \} \tag {26} +$$ + +# 4. Experiments + +# 4.1. Setup + +Implementation details. Our network, implemented using PyTorch 1.7, underwent training and testing on an NVIDIA GeForce RTX 3090 GPU. We employed the Adam optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ . The patch size was configured as $256 \times 256$ , and the batch size was set to 2. The diffusion model's total time steps, denoted as T, were set to 1000, and the number of training iterations reached one million. The initial learning rate was established at 0.0001. + +Datasets. We utilize the real-world UIEBD dataset [28] and the LSUI dataset [18] for training and evaluating our model. The UIEBD dataset comprises 890 underwater images with corresponding labels. Out of these, 700 images are allocated for training, and the remaining 190 are designated for testing. The LSUI dataset is randomly partitioned into 4500 images for training and 504 images for testing. In addition, to verify the generalization of WF-Diff, we use non-reference benchmarks U45 [20], which contains 45 underwater images for testing. + +Comparison methods. We conduct a comparative analysis between WF-Diff and eight state-of-the-art (SOTA) UIE methods, namely UIEC^2-Net [39], Water-Net [28], UWCNN [3], SCNet [8], UIEWD [25], U-color [19], U-shape [18], and DM-water [36]. To ensure a fair and rigorous comparison, we utilize the provided source codes from the respective authors and adhere strictly to the identical experimental settings across all evaluations. + +Evaluation Metrics. We primarily utilize well-established full-reference image quality assessment metrics: PSNR and SSIM [41]. PSNR and SSIM offer quantitative comparisons of our method with other approaches at both pixel and structural levels. Higher PSNR and SSIM values signify superior quality of the generated images. Additionally, we incorporate the LPIPS and FID metrics for full-reference image evaluation. LPIPS [47] is a deep neural network-based image quality metric that assesses the perceptual similarity between an image and a reference image. FID [9] measures the distance between the distributions of real and generated images. A lower LPIPS and FID score indicates a more effective UIE approach. For non-reference benchmarks U45, + +Table 2. Quantitative comparison of different UIE methods on the UIEBD, LSUI and U45 datasets. The best results are highlighted in bold and the second best results are underlined. + +
MethodsUIEWDUWCNNUIEC^2-NetWater-NetSCNetU-colorU-shapeDM-waterOurs
UIBDFID↓85.1294.4435.0637.4833.6638.2546.1131.0727.85
LPIPS↓0.39560.35250.20330.21160.24970.23370.22640.14360.1248
PSNR↑14.6515.4020.1419.3520.4120.7121.2521.8823.86
SSIM↑0.72650.77490.82150.83210.82350.84110.84530.81940.8730
LSUIFID↓98.49100.534.5138.90158.9945.0628.5627.9126.75
LPIPS↓0.39620.34500.14320.16780.2830.1230.10280.11380.1096
PSNR↑15.4318.2420.8619.7322.6322.9124.1627.6527.26
SSIM↑0.78020.84650.88670.82260.91760.89020.93220.88670.9437
U45UIQM↑2.4582.3792.7802.9572.8563.1043.1513.0863.181
UCIQE↑0.5830.5670.5910.6010.5940.5860.5920.6340.619
+ +![](images/1bfa9fdf1354aa95921dc1c522da95e1d126ca8b39acaedad87f907dbe181caf.jpg) +Figure 4. Qualitative comparison with other SOTA methods on real underwater images. + +we introduce UIQM [27] and UCIQE [44] to evaluate our method. + +# 4.2. Results and Comparisons + +Table 2 shows the quantitative results compared with different baselines on UIEBD, LSUI and U45 datasets, including with UIEC^2-Net, Water-Net, UIEWD, UWCNN, SCNet, U-color, U-shape, and DM-water. We mainly use PSNR, SSIM, LPIPS and FID as our quantitative indices for UIEBD and LSUI datasets, and UIQM and UCIQE for non-reference dataset U45. The results in Table 2 show that our algorithm outperforms state-of-the-art methods obviously, and achieve state-of-the-art performance in terms of image + +quality evaluation metrics on UIE task, which verifies the robustness of the proposed WF-Diff. To better validate the superiority of our methods, in Figure 4, we show the visual results comparison with state-of-the-art methods on real underwater images. The six examples are randomly selected on the UIEBD and LSUI datasets. Our methods consistently generate natural and better visual results on testing images, strongly proving that WF-Diff has good generalization performance for real-world applications. + +# 4.3. Ablation Study + +Ablation study with WFI2-net. In order to evaluate the effectiveness of each part in WFI2-net, we conduct two ab + +Table 3. Ablation study with network structure of WFI2-net on UIEBD dataset. + +
WTBSFFBUIEBD
SACASDUFDUPSNR ↑SSIM↑
×✓✓✓20.940.8541
✓×✓✓20.820.8473
✓✓×✓21.110.8586
✓✓✓×20.230.8346
✓✓✓✓21.870.8622
+ +Table 4. Ablation study with loss of WFI2-net on UIEBD dataset. + +
MethodsUIEBD
LhLrecLaCFCPSNR ↑SSIM↑
×✓✓✓20.460.8274
✓×✓✓20.650.8399
✓✓×✓19.810.8213
✓✓✓×20.970.8425
✓✓✓✓21.870.8622
+ +Table 5. Ablation study with FRDAM on UIEBD dataset. DM refers to diffusion model (DM), RDM refers to residual DM. D-L refers to refining low-frequency component in wavelet space, and D-H refers to refining high-frequency components in wavelet space. + +
MethodDMRDMD-LD-HCFCPSNR↑
A✓××××20.86
B×✓×××22.37
C×✓✓××22.32
D×✓×✓×22.58
E×✓✓✓×23.44
F×✓✓✓✓23.86
+ +![](images/1ce8f5b83440314316edfeb324a054b8a0f8e33615e89a923271832e831b0ca3.jpg) + +![](images/a42bb405fb55c40cf68d32fafc8dc8e634dc63d81671a490de5722991eb1c0da.jpg) +A + +![](images/8e0035f3d8970625de9e5a3ea51832ef0be936d4aca831c9161c3bfc1afd595a.jpg) +B + +![](images/ec570d0ab29c9a8cb0af6105ae1e6865ce883378840e48e17e6748ecc45b4bb3.jpg) +C + +![](images/86d01bf4be972e6bae92526334c2cc007e285d6a93d59a57dca0dec27fcde04c.jpg) +input +D +Figure 5. Visual results of ablation study with FRDAM. + +![](images/31b24dcea85bb04a297863eafe656f8b463997c378e726ce1b611a4eb3987200.jpg) +E + +![](images/59256db715b1dc3377efcc9f4cd5b3f4acce04bf12127dbe294dbbcc2617ad21.jpg) +F + +![](images/0515a0e928e77986f2670c5a2d47c600afbfa759645df275dfbe996e8ebcc833.jpg) +GT + +lation experiments with WFI2-net on the UIEBD dataset in Table 3 and 4. CFC is cross-frequency conditioner. Note that, we do not discuss the effect of our proposed FRDAM + +Table 6. Comparison results of inference speed. + +
MethodSCNetUshapeDM-waterWFI2-netWF-Diff
Inference time (s)0.01490.03530.14410.07390.2816
+ +Table 7. Comparison results of FLOPs. + +
MethodSCNetWaterNetUcolorUshapeWFI2-net
FLOPs (G)5.88193.7443.866.294.1
+ +here. We regularly remove one component to each configuration at one time, and our strategy achieves the best performance by using all loss functions and blocks, proving that each part of WFI2-net is useful for UIE task. + +Ablation study with FRDAM. In this section, we will discuss the effectiveness of FRDAM. Table 5 shows the quantitative results on the UIEBD dataset, and Figure 5 shows the visual results. Note that, Model A and B achieve diffusion process in the pixel level, and model C, D, E and E achieve diffusion process in wavelet space. Model A obtains relatively poor results in PSNR and generates images with color distortion or artifacts in Figure 5 due to the diversity of the sampling process. Model B does not yield fully satisfactory results, because it needs to adjust both the high and low-frequency information in images, which limits their ability to focus on fine-grained information. Compared to model C, model D achieves better results, suggesting that the degradation information is mainly in the high-frequency information in frequency diffusion adjustment stage. Model F achieves the best performance, proving that our designed FRDAM is best for UIE task. + +# 5. Conclusion + +In this paper, we develop a novel UIE framework, namely WF-Diff. With fully utilizing the frequency domain characteristics and diffusion model, WFI2-net can achieve enhancement and adjustment of frequency information. Our proposed FRDAM is a plug-and-play universal module to adjust the details of the underwater images. WF-Diff shows SOTA performance on UIE task, and extensive ablation experiments prove that each of our contributions is effective. + +Limitations. As a result of employing two diffusion models, our approach doesn't confer an advantage in terms of inference speed. Table 6 presents the actual values of inference speed for WF-Diff, indicating that our model does not outperform recent approaches in terms of inference time. Furthermore, Table 7 provides a comparison of FLOPs for WFI2-net. It should be mentioned that the implicit sampling step is set to 10 for WF-Diff. Hence, in the future, we will delve into methods to expedite the sampling process. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62276138 and 62371232). + +# References + +[1] Derya Akkaynak and Tali Treibitz. Sea-thru: A method for removing water from underwater images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1682-1691. Computer Vision Foundation / IEEE, 2019. 3 +[2] Derya Akkaynak, Tali Treibitz, Tom Shlesinger, Yossi Loya, Raz Tamir, and David Iluz. What is the space of attenuation coefficients in underwater computer vision? In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 568-577, 2017. 1 +[3] Saeed Anwar, Chongyi Li, and Fatih Porikli. Deep underwater image enhancement. CoRR, abs/1807.03528, 2018. 6 +[4] John Yi-Wu Chiang and Ying-Ching Chen. Underwater image enhancement by wavelength compensation and dehazing. IEEE Trans. Image Process., 21(4):1756-1769, 2012. 3 +[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14347-14356. IEEE, 2021. 2, 3 +[6] Karin de Langis and Junaed Sattar. Realtime multi-diver tracking and re-identification for underwater human-robot collaboration. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 11140-11146, 2020. 1 +[7] Cameron Fabbri, Md Jahidul Islam, and Junaed Sattar. Enhancing underwater imagery using generative adversarial networks. In 2018 IEEE International Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May 21-25, 2018, pages 7159-7165. IEEE, 2018. 1, 3 +[8] Zhenqi Fu, Xiaopeng Lin, Wu Wang, Yue Huang, and Xinghao Ding. Underwater image enhancement via learning water type desensitized representations. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2764-2768. IEEE, 2022. 6 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6 +[11] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based CNN for multi-scale face super resolution. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1698-1706, 2017. 4 + +[12] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. pages 163-180, 2022. 2 +[13] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18145-18155. IEEE, 2023. 3 +[14] Md Jahidul Islam, Youya Xia, and Junaed Sattar. Fast underwater image enhancement for improved visual perception. IEEE Robotics Autom. Lett., 5(2):3227-3234, 2020. 3 +[15] Paulo Drews Jr., Erickson Rangel do Nascimento, F. Moraes, Silvia S. C. Botelho, and Mario F. M. Campos. Transmission estimation in underwater single images. In 2013 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 825-830, 2013. 1, 3 +[16] Eunhee Kang, Won Chang, Jae Jun Yoo, and Jong Chul Ye. Deep convolutional framelet denoising for low-dose CT via wavelet residual network. IEEE Trans. Medical Imaging, 37 (6):1358-1369, 2018. 4 +[17] Chongyi Li, Jichang Guo, Runmin Cong, Yanwei Pang, and Bo Wang. Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior. IEEE Trans. Image Process., 25(12):5664-5677, 2016. 1 +[18] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Trans. Image Process., 29:4376-4389, 2020. 1, 3, 6 +[19] Chongyi Li, Saeed Anwar, Junhui Hou, Runmin Cong, Chunle Guo, and Wenqi Ren. Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE Trans. Image Process., 30:4985-5000, 2021. 3, 6 +[20] Hanyu Li, Jingjing Li, and Wei Wang. A fusion adversarial underwater image enhancement network with a public test dataset. arXiv preprint arXiv:1906.06819, 2019. 6 +[21] Jie Li, Katherine A. Skinner, Ryan M. Eustice, and Matthew Johnson-Roberson. Watergan: Unsupervised generative network to enable real-time color correction of monocular underwater images. IEEE Robotics Autom. Lett., 3(1):387-394, 2018. 3 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 2 +[23] Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-Icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 2 +[24] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 2 + +[25] Ziyin Ma and Changjae Oh. A wavelet-based dual-stream network for underwater image enhancement. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2769-2773. IEEE, 2022. 6 +[26] James McMahon and Erion Plaku. Autonomous data collection with timed communication constraints for unmanned underwater vehicles. IEEE Robotics Autom. Lett., 6(2): 1832-1839, 2021. 1 +[27] Karen Panetta, Chen Gao, and Sos Agaian. Human-visual-system-inspired underwater image quality measures. IEEE Journal of Oceanic Engineering, 41(3):541-551, 2015. 7 +[28] Lintao Peng, Chunli Zhu, and Liheng Bian. U-shape transformer for underwater image enhancement. IEEE Trans. Image Process., 32:3066-3079, 2023. 1, 3, 6 +[29] Yan-Tsung Peng and Pamela C. Cosman. Underwater image restoration based on image blurriness and light absorption. IEEE Trans. Image Process., 26(4):1579-1594, 2017. 1, 3 +[30] Yan-Tsung Peng, Keming Cao, and Pamela C. Cosman. Generalization of the dark channel prior for single image restoration. IEEE Trans. Image Process., 27(6):2856-2868, 2018. 3 +[31] Priyadharsini Ravisankar, T. Sree Sharmila, and V. Rajendran. A wavelet transform based contrast enhancement method for underwater acoustic images. Multidimens. Syst. Signal Process., 29(4):1845-1859, 2018. 1, 4 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2 +[33] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 15:1-15:10. ACM, 2022. 3 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021. 2, 3, 6 +[36] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 5419-5427. ACM, 2023. 1, 2, 3, 6 +[37] Pritish M. Uplavikar, Zhenyu Wu, and Zhangyang Wang. All-in-one underwater image enhancement using domain + +adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1-8, 2019. 3 +[38] Yi Wang, Hui Liu, and Lap-Pui Chau. Single underwater image restoration using adaptive attenuation-curve prior. IEEE Trans. Circuits Syst. I Regul. Pap., 65-I(3):992-1002, 2018. 3 +[39] Yudong Wang, Jichang Guo, Huan Gao, and HuiHui Yue. Uiec $^2$ -net: Cnn-based underwater image enhancement using two color space. Signal Process. Image Commun., 96: 116250, 2021. 6 +[40] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 2, 3 +[41] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6 +[42] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16272-16282. IEEE, 2022. 3 +[43] Jian Yang, Chen Li, and Xuelong Li. Underwater image restoration with light-aware progressive network. In IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023, pages 1-5. IEEE, 2023. 3 +[44] Miao Yang and Arcot Sowmya. An underwater color image quality evaluation metric. IEEE Transactions on Image Processing, 24(12):6062-6071, 2015. 7 +[45] Zongyuan Yang, Baolin Liu, Yongping Xiong, Lan Yi, Guibin Wu, Xiaojun Tang, Ziqi Liu, Junjie Zhou, and Xing Zhang. Docdiff: Document enhancement via residual diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 2795-2806. ACM, 2023. 2, 3 +[46] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with A generative diffusion model. CoRR, abs/2308.13164, 2023. 2, 3 +[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595, 2018. 6 +[48] Chen Zhao, Wei-Ling Cai, Chenyu Dong, and Ziqi Zeng. Toward sufficient spatial-frequency interaction for gradient-aware underwater image enhancement. arXiv preprint arXiv:2202.08537, 2023. 2, 4 +[49] Chen Zhao, Wei-Ling Cai, and Zheng Yuan. Spectral normalization and dual contrastive regularization for image-to + +image translation. The Visual Computer, pages 1-12, 2024. 3 +[50] Chen Zhao, Chenyu Dong, and Weiling Cai. Learning a physical-aware diffusion model based on transformer for underwater image enhancement. arXiv preprint arXiv:2403.01497, 2024. 3 +[51] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 2 +[52] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. arXiv preprint arXiv:2402.05408, 2024. 2 \ No newline at end of file diff --git a/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/images.zip b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..364e026eb44c259c3c56b684556acc3326edbd1c --- /dev/null +++ b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d7b18b1e3e9c96c2a17ba87941ce3a73e84e1662076f3f216caa42e4942628c +size 805757 diff --git a/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/layout.json b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09e1dbae11ea756bfaac36c45ac51fe324122b1e --- /dev/null +++ b/2024/Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration/layout.json @@ -0,0 +1,9129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 103, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 103, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 71, + 103, + 523, + 140 + ], + "type": "text", + "content": "Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 161, + 441, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 161, + 441, + 190 + ], + "spans": [ + { + "bbox": [ + 148, + 161, + 441, + 190 + ], + "type": "text", + "content": "Chen Zhao, Weiling Cai*, Chenyu Dong, Chengwei Hu \nSchool of Artificial Intelligence, Nanjing Normal University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 190, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 190, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 190, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 290, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 526 + ], + "type": "text", + "content": "Underwater images are subject to intricate and diverse degradation, inevitably affecting the effectiveness of underwater visual tasks. However, most approaches primarily operate in the raw pixel space of images, which limits the exploration of the frequency characteristics of underwater images, leading to an inadequate utilization of deep models' representational capabilities in producing high-quality images. In this paper, we introduce a novel Underwater Image Enhancement (UIE) framework, named WF-Diff, designed to fully leverage the characteristics of frequency domain information and diffusion models. WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). With our full exploration of the frequency domain information, WFI2-net aims to achieve preliminary enhancement of frequency information in the wavelet space. Our proposed FRDAM can further refine the high- and low-frequency information of the initial enhanced images, which can be viewed as a plug-and-play universal module to adjust the detail of the underwater images. With the above techniques, our algorithm can show SOTA performance on real-world underwater image datasets, and achieves competitive performance in visual quality. The code is available at https://github.com/zhihefang/WF-Diff." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 553, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 696 + ], + "type": "text", + "content": "Underwater image restoration is a practical but challenging technology in the field of underwater vision, widely used for tasks, such as underwater robotics[26] and underwater object tracking[6]. Due to light refraction, absorption, and scattering in underwater scenes, underwater images are usually severely distorted, with low contrast and blurriness [2]. Therefore, clear underwater images play a critical role in fields that need to interact with the underwater environment. The main goal of underwater image enhancement (UIE) is to obtain high-quality images by removing scattering and correcting color distortion in degraded images. UIE is crucial for vision-related underwater tasks." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 216, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 310, + 216, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 310, + 216, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 310, + 216, + 541, + 266 + ], + "type": "image", + "image_path": "2c2150c3ba66b3750da5912ea5029f6940c1e7309121d304299829a7ddac99ef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 268, + 541, + 361 + ], + "blocks": [ + { + "bbox": [ + 308, + 268, + 541, + 361 + ], + "lines": [ + { + "bbox": [ + 308, + 268, + 541, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 268, + 541, + 361 + ], + "type": "image", + "image_path": "6996da4f8487b3b1d29fd489b0beca75cc692d0ad32c9730e1ebf611aba35ad1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 369, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 546, + 415 + ], + "type": "text", + "content": "Figure 1. Our motivations. The amplitude and phase are produced by Fast Fourier Transform (FFT) and the recombined images are obtained by Inverse FFT (IFFT). We further explore the frequency properties for underwater images in Wavelet space." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 334, + 494, + 520, + 546 + ], + "blocks": [ + { + "bbox": [ + 305, + 426, + 547, + 492 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 492 + ], + "type": "text", + "content": "Table 1. Evaluation of using different frequency domain transformation strategies on the UIEBD [28]. S1 refers to swapping the amplitude in original images, S2 refers to only swapping the amplitude of low-frequency sub-images in the wavelet space, and S3 refers to swapping the amplitude of low-frequency sub-images and high-frequency sub-images in the wavelet space." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 494, + 520, + 546 + ], + "lines": [ + { + "bbox": [ + 334, + 494, + 520, + 546 + ], + "spans": [ + { + "bbox": [ + 334, + 494, + 520, + 546 + ], + "type": "table", + "html": "
StrategyPSNR ↑SSIM ↑LPIPS↓FID↓
S128.680.90270.095728.70
S227.100.88130.102333.04
S329.970.93430.082023.55
", + "image_path": "8d9290a9dca574e80730b3882e3972d8be9bff79ac954d7b8354f6f8f06829ae.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "To address this problem, traditional UIE methods based on the physical properties of the underwater images were proposed [15, 17, 29-31]. These methods investigate the physical mechanism of the degradation caused by color cast or scattering and compensate for them to enhance the underwater images. However, these physics-based models with limited representation capacity cannot address all the complex physical and optical factors underlying the underwater scenes, which leads to poor enhancement results under highly complex and diverse underwater scenes. Recently, some learning-based methods [7, 18, 28, 36] for UIE can produce better results, since neural networks have powerful feature representation and nonlinear mapping capabilities." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 249, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 249, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 249, + 712 + ], + "type": "text", + "content": "*Corresponding Author. E-mail: caiwl@njnu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8281" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "ties. It can learn the mapping of an image from degenerate to clear from a substantial quantity of paired training data. However, most previous methods are based on the raw pixel space of images, with limited exploration of the properties of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models for generating high-quality images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 482 + ], + "type": "text", + "content": "Building on insights from previous Fourier-based works [12, 48], we explore the properties of the Fourier frequency information for the UIE task, as illustrated in Figure 1. Given two images (an underwater image and its corresponding ground-truth), we swap their amplitude components and combine them with corresponding phase components in the Fourier space. The recombined results show that the visual appearance is swapped following the amplitude swapping, which indicates the degradation information of underwater images is mainly contained in the amplitude component. We further explore the properties of the amplitude components in the Wavelet space. Specifically, the images can be decomposed into low-frequency sub-images and high-frequency sub-images using discrete wavelet transformations (DWT), and then we swap amplitude components of low-frequency sub-images. From visual results, we can find a similar phenomenon, which means the color degradation information is mainly contained in low-frequency sub-images, and the texture and detail degradation information is mainly contained in high-frequency sub-images. Table 1 shows the quantitative evaluation of the different frequency domain strategies, proving that our discovery is objective. Consequently, how to adequately exploit the properties of frequency domain information and effectively incorporate them into a unified image enhancement network is a crucial issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Recently, diffusion-based methods [10, 35] have garnered significant attention due to their outstanding performance in image synthesis [23, 24, 32, 34, 52] and restoration tasks [5, 40, 46, 51]. These methods rely on a hierarchical denoising autoencoder architecture, enabling them to iteratively reverse a diffusion process and achieve high-quality mapping from randomly sampled Gaussian noise to target images or latent distributions [10]. Tang et al. [36] present an image enhancement approach with diffusion model in underwater scenes. While standard diffusion models exhibit sufficient capability, unforeseen artifacts may arise as a result of the diversity introduced during the sampling process from randomly generated Gaussian noise to images [45]. Furthermore, the diffusion model needs to recover both the high and low-frequency information of images, which limits their ability to focus on fine-grained information, missing out on texture and details. Thereby, it is very crucial that the powerful representation capabilities of diffusion models can be fully utilized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 444 + ], + "type": "text", + "content": "In this paper, we develop a novel UIE framework to fully exploit the properties of frequency domain information and diffusion models, called WF-Diff, which mainly consists of two stages: frequency preliminary enhancement and frequency diffusion adjustment. The first stage aims to preliminarily enhance the high-frequency and low-frequency components of underwater images by utilizing the frequency domain characteristics. Specifically, we first convert the input images into the wavelet space using discrete wavelet transformations (DWT), obtaining an average coefficient that represents the low-frequency content information of the input image and three high-frequency coefficients that represent sparse vertical, horizontal, and diagonal details of the input image. Then, we design a wavelet-based Fourier information interaction network (WFI2-net), fully integrating the characteristics of Transformer [22] and Fourier prior information to enhance high- and low-frequency content, respectively. Moreover, to achieve interaction of high- and low-frequency information, we propose a cross-frequency conditioner (CFC) to further improve the generation quality. The target of the second stage is to make adjustment to the initial enhanced coarse results in terms of details and textures via diffusion model. Consequently, we propose a frequency residual diffusion adjustment module (FRDAM). Unlike previous diffusion-based work, FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models in the wavelet space, which can not only increase the model's focus on fine-grained information but also mitigate the adverse effects of the diversity of the sampling process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 453, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 475 + ], + "type": "text", + "content": "In summary, the main contributions of our method are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 498, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 498, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 545, + 557 + ], + "type": "text", + "content": "- We explore in depth the properties of the frequency domain for underwater images. Based on the properties and diffusion model, we propose a novel UIE framework, named WF-Diff, with the goal of achieving frequency enhancement and diffusion adjustment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 558, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 545, + 617 + ], + "type": "text", + "content": "- We propose a frequency residual diffusion adjustment module (FRDAM) to further refine the high- and low-frequency information of the initial enhanced images. FRDAM can be viewed as a plug-and-play universal module to adjust the detail of the underwater images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 618, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 653 + ], + "type": "text", + "content": "- We propose a cross-frequency conditioner (CFC) to achieve the cross-frequency interaction of high- and low-frequency information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 545, + 712 + ], + "type": "text", + "content": "- Experimental results compared with SOTAs considerably show that our developed WF-Diff performs the superiority against previous UIE approaches, and extensive ablation experiments can demonstrate the effectiveness of our contributions." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8282" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 537, + 329 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 537, + 329 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 537, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 537, + 329 + ], + "type": "image", + "image_path": "52e27cbb5d2ffc143f2fe469b95c68d819141d717404f60deb1e5cd2dc613def.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 341, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 46, + 341, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 545, + 397 + ], + "type": "text", + "content": "Figure 2. Overall framework of WF-Diff. It contains two detachable networks, Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Furthermore, the proposed cross-frequency conditioner (CFC) aims to achieve the cross-frequency interaction of high- and low-frequency information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 402, + 139, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 139, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 139, + 414 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 426, + 226, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 226, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 226, + 440 + ], + "type": "text", + "content": "2.1. Underwater Image Enhancement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "Currently, existing UID methods can be briefly categorized into the physical and deep model-based approaches [15, 18, 28-30, 36]. Most UID methods based on the physical model utilize prior knowledge to establish models, such as water dark channel priors [29], attenuation curve priors [38], fuzzy priors [4]. In addition, Akkaynak and Treibitz [1] proposed a method based on the revised physical imaging model. However, the depth map of the underwater scene is difficult to obtain. This leads to unstable performance, which usually suffers from severe color cast and artifacts. Therefore, the manually established priors restrain the model's robustness and scalability under the complicated and varied circumstances. Recently, deep learning-based methods [18, 28, 36] have achieved acceptable performance. To alleviate the need for real-world underwater paired training data, many methods introduced GAN-based framework for UIE [7, 14, 21, 49], such as WaterGAN [21], UGAN [7] and UIE-DAL [37]. Recently, some complex frameworks were proposed and achieve the-state-of-the-art performance [15, 29]. Ucolor [19] combined the underwater physical imaging model in the raw space and designed a medium transmission guided model. Yang et al." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 403, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 523 + ], + "type": "text", + "content": "[43] proposed a reflected light-aware multi-scale progressive restoration network to obtain images with both color equalization and rich texture in various underwater scenes. Huang et al. [13] proposed a mean teacher based semi-supervised network, which effectively leverages the knowledge from unlabeled data. However, most previous methods are based on the spatial domain, with limited exploration of the frequency space for underwater images, which results in an inability to effectively harness the representation power of the deep models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 536, + 405, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 405, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 405, + 548 + ], + "type": "text", + "content": "2.2. Diffusion Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Recently, Diffusion Probabilistic Models (DPMs) [10, 35] have been widely adopted for conditional image generation [5, 40, 42, 46, 50]. Sahara et al. [33] proposed Palette, which has demonstrated the excellent performance of diffusion models in the field of conditional image generation, including colorization, in-painting and JPEG restoration. Tang et al. [36] presented an image enhancement approach with diffusion model in underwater scenes. However, the reverse process starts from randomly sampled Gaussian noise to full images [45], which can lead to unexpected artifacts due to the diversity of the sampling process. Furthermore, the diffusion model needs to recover both the high and low-frequency information in images, which lim" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8283" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "its their ability to focus on fine-grained information. Consequently, how to incorporate diffusion models into a unified underwater image enhancement network is a vital issue." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 129, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 129, + 134 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 129, + 134 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 141, + 162, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 162, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 162, + 152 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 159, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 159, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 289, + 495 + ], + "type": "text", + "content": "Given an underwater image as input, our goal is to learn a network to generate an output that eliminates the color cast from input while enhancing the image details. The overall framework of WF-Diff is shown in Figure 2. WF-Diff is designed to fully leverage the characteristics of frequency domain information and powerful ability of diffusion models. Specifically, WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). We first convert the input into the wavelet space using discrete wavelet transformations (DWT), obtaining a low-frequency coefficient and three high-frequency coefficients. WFI2-net is dedicated to achieving preliminary enhancement of frequency information. We fully integrate the characteristics of Transformer and Fourier prior Information, and design wide transformer block (WTB) and spatial-frequency fusion block (SFFB) to enhance high- and low-frequency content respectively. FRDAM consists of low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB), which aims to further adjust the high- and low-frequency information of the initial enhanced images. Note that, our proposed FRDAM learns the residual distribution of high- and low-frequency information between the ground-truth and the initial enhanced results using two diffusion models, respectively. Additionally, the proposed cross-frequency conditioner (CFC) strives to achieve cross-frequency interaction between high- and low-frequency information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 504, + 263, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 263, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 263, + 516 + ], + "type": "text", + "content": "3.2. Discrete Wavelet and Fourier Transform" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "text", + "content": "Discrete Wavelet transform (DWT) has been widely applied to low-level vision tasks [11, 16]. We firstly use DWT to decompose an input into multiple frequency sub-bands so that we can achieve the color correction of low-frequency information and detail enhancement of high-frequency information, respectively. Given an underwater image as input " + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times c}" + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "text", + "content": ", we use DWT with Haar wavelets to decompose the input. Haar wavelets consist of the low-pass filter " + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "text", + "content": ", and the high-pass filter " + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 523, + 287, + 631 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 640, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 640, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 94, + 640, + 287, + 666 + ], + "type": "interline_equation", + "content": "L = \\frac {1}{\\sqrt {2}} [ 1, 1 ] ^ {T}, H = \\frac {1}{\\sqrt {2}} [ 1, - 1 ] ^ {T}. \\tag {1}", + "image_path": "4f2303d88f10b04d8e5be3295f945d253c3a438f098b4dea272b59110852a9d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 675, + 287, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 287, + 697 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 287, + 697 + ], + "type": "text", + "content": "We can obtain four sub-bands, which can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "I _ {L L}, \\left\\{I _ {L H}, I _ {H L}, I _ {H H} \\right\\} = \\mathrm {D W T} (I), \\tag {2}", + "image_path": "05e175dcc41cf7033939b732783ee9006077d8fc3b8cdbe1c063804b54633842.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "type": "inline_equation", + "content": "I_{LL},\\{I_{LH},I_{HL},I_{HH}\\} \\in \\mathbb{R}^{\\frac{H}{2}\\times \\frac{W}{2}\\times c}" + }, + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "type": "text", + "content": " represent the low-frequency component of the input and high-frequency components in the vertical, horizontal, and diagonal directions, respectively. More specifically, the low-frequency component contains the content and color information of the input image, and the other three high-frequency coefficients contain details information of global structures and textures [31]. The sub-bands are downsampled to half-resolution of the input but do not result in information loss due to the biorthogonal property of DWT. For low-frequency component " + }, + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "type": "inline_equation", + "content": "I_{LL}" + }, + { + "bbox": [ + 304, + 70, + 545, + 215 + ], + "type": "text", + "content": ", we will explore its properties in Fourier space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "content": "Then, we introduce the operation of the Fourier transform [48]. Given a image " + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "content": ", whose shape is " + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "content": ", the Fourier transform " + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "content": " which converts " + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "content": " to the Fourier space " + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 216, + 545, + 264 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 269, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 269, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 269, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (x) (u, v) = X (u, v) = \\frac {1}{\\sqrt {H W}} \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} x (h, w) e ^ {- j 2 \\pi \\left(\\frac {h}{H} u + \\frac {w}{W} v\\right)}, \\tag {3}", + "image_path": "6b579bcdf1ca92b77d7ab9b271e3e6c6170e7ae48f56fd9e62f8ec4d8ae5ceb7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": " are the coordinates in the spatial space and " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": " are the coordinates in the Fourier space. " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": " denotes the inverse transform of " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": ". Complex component " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "X(u, v)" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": " in the Fourier space can be represented by a amplitude component " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(X(u, v))" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": " and a phase component " + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(X(u, v))" + }, + { + "bbox": [ + 305, + 302, + 545, + 361 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 367, + 545, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 367, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 323, + 367, + 545, + 397 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} (X (u, v)) = \\sqrt {R ^ {2} (X (u , v)) + I ^ {2} (X (u , v))}, \\tag {4} \\\\ \\mathcal {P} (X (u, v)) = \\arctan [ \\frac {I (X (u , v))}{R (X (u , v))} ], \\\\ \\end{array}", + "image_path": "559cf005c959e5541393ceb326e4f555d877aded81759733210e575d1176925b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "inline_equation", + "content": "R(x)" + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "inline_equation", + "content": "I(x)" + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "content": " represent the real and imaginary parts of " + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "inline_equation", + "content": "X(u, v)" + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "content": ", respectively. Note that, the Fourier operation can be computed alone in each channel for feature maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "content": "According to Figure 1 and Table 1 (our motivation), we conclude that the color degradation information of underwater images is mainly contained in the amplitude component of low-frequency sub-band, and the texture and detail degradation information is mainly contained in high-frequency sub-bands." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 515, + 509, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 509, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 509, + 528 + ], + "type": "text", + "content": "3.3. Frequency Preliminary Enhancement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Based on the above analysis, in frequency preliminary enhancement stage, we design a simple but effective WFI2-net with a parallel encoder-decoder (U-Net-like) format to restore the amplitude component of low-frequency information and high-frequency components, respectively. We also utilize skip connections to connect the features at the same level in the encoder and decoder. For high-frequency branch, we utilize advantage of transformer modeling global information to enhance high-frequency coefficients. We design wide transformer block (WTB) using multi-scale information, aiming to model long range dependencies. Our low-frequency branch aims to restore the amplitude component in Fourier space. In order to obtain rich frequency and spatial information, we design spatial-frequency fusion block (SFFB)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8284" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 226, + 173 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 226, + 173 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 226, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 226, + 173 + ], + "type": "image", + "image_path": "07874ab2c66a9096f5c0e889f10ace59987f12d06afcb6dc0d31d288e9b4ee2a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 536, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 536, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 536, + 194 + ], + "type": "text", + "content": "Figure 3. The architecture of (a) Wide Transformer Block, (b) Spatial-Frequency Fusion Block and (c) Cross-Frequency Conditioner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 356, + 174 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 356, + 174 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 356, + 174 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 356, + 174 + ], + "type": "image", + "image_path": "fb4f1f13acda6bcadf098ccf3ecd3da65d1d627b237d70e1fba648a071ef9a6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 538, + 175 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 538, + 175 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 538, + 175 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 538, + 175 + ], + "type": "image", + "image_path": "ef6e42c87c0984218dcdfbe2c25bb5a1752db5a16a1f33f71d4edb2bf1741b80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "type": "text", + "content": "Wide Transformer Block. Unlike low-frequency coefficients, high-frequency components contain global structure and texture details. Consequently, the high-frequency branch focuses on modeling global and local features. WTB is shown in Figure 3 (a). Given " + }, + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "type": "inline_equation", + "content": "I_{LH}, I_{HL}, I_{HH} \\in \\mathbb{R}^{\\frac{H}{2} \\times \\frac{W}{2} \\times c}" + }, + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "type": "text", + "content": ", WTB firstly obtains their embedding features " + }, + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "type": "inline_equation", + "content": "T_{in} \\in \\mathbb{R}^{3 \\times \\frac{H}{2} \\times \\frac{W}{2} \\times C}" + }, + { + "bbox": [ + 47, + 201, + 289, + 320 + ], + "type": "text", + "content": " through convolution projection. To be specific, WTB are composed of an attention (Atten) module and a feed-forward network (FFN) module, and the computation can be denoted in the WTB as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 326, + 287, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 326, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 77, + 326, + 287, + 340 + ], + "type": "interline_equation", + "content": "Q, K, V, L = \\operatorname {S p l i t} \\left(W _ {d} W _ {p} \\left(\\operatorname {N o r m} \\left(T _ {i - 1}\\right)\\right)\\right), \\tag {5}", + "image_path": "70e4a52c151c579a4b9779720976a183633f42388e81afc7e214ceead747ff52.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 344, + 287, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 344, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 88, + 344, + 287, + 358 + ], + "type": "interline_equation", + "content": "\\hat {T} _ {i} = S A (Q, K, V) + C A (L) + T _ {i - 1}, \\tag {6}", + "image_path": "77c1680fa8a79d9938f008078ae4352bd2b0ac39c66f618a18f7f4c2c8fcd3f1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 361, + 287, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 361, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 287, + 376 + ], + "type": "interline_equation", + "content": "T _ {i} = F F N \\left(\\operatorname {N o r m} \\left(\\hat {T} _ {i}\\right)\\right) + \\hat {T} _ {i}, \\tag {7}", + "image_path": "51ea597af17644e479ca5eaadf9dd415d0056a25374753e8545e3f82178a6304.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "SA" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "CA" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " refer to self-attention and channel attention, respectively. Norm refers to normalization. " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "T_{i - 1}" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " represents the input embeddings of the current WTB. " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "W_{d}" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " point-wise convolution and multi-scale kernel depth-wise convolution, respectively; Split refers to the split operation. " + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 378, + 287, + 449 + ], + "type": "text", + "content": " aims to focus on local information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": "Spatial-Frequency Fusion Block. We show the structure of SFFB in Figure 3 (b), which has a spatial domain unit (SDU) and a frequency domain unit (FDU) for interaction of dual domain representations. In spatial domain unit, we employ multi-scale convolution kernels in order to enlarge the limited spatial receptive field. After obtaining the spatial embeddings " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": ", we firstly utilize the FFT to obtain the amplitude " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " and phase " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " components. Then, the " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " are fed into two layers " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "1^{*}1" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " conv to obtain " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{A}'(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": ". Finally, we use the IFFT algorithm to map " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{A}'(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'(F_s)" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": " to image space and obtain the frequency embeddings " + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "inline_equation", + "content": "F_{f}" + }, + { + "bbox": [ + 46, + 450, + 288, + 606 + ], + "type": "text", + "content": ". The fusion embeddings of spatial domain and frequency domain can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 611, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 287, + 624 + ], + "type": "interline_equation", + "content": "F _ {s f} = F _ {s} + F _ {f}. \\tag {8}", + "image_path": "a8e360b8f9cedb9506aa33fad10fbb82b1df06d15455c659b8a66bf0af083bd4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "text", + "content": "Loss Function. We denote " + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "inline_equation", + "content": "I_{LL}^{\\prime}" + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "text", + "content": " as output of low-frequency branch, and " + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "inline_equation", + "content": "I_{LH}^{\\prime}, I_{HL}^{\\prime}" + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "inline_equation", + "content": "I_{HH}^{\\prime}" + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "text", + "content": " as output of high-frequency branch. Ground-truth (G) can be decomposed " + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "inline_equation", + "content": "G_{LL}, G_{LH}, G_{HL}, G_{HH}" + }, + { + "bbox": [ + 46, + 630, + 288, + 692 + ], + "type": "text", + "content": " by DWT. The high-frequency loss can be expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 121, + 695, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 695, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 121, + 695, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} = \\left\\| I _ {(i)} ^ {\\prime} - G _ {(i)} \\right\\| _ {2}, \\tag {9}", + "image_path": "5034d3e152059de1d947556e89e2fed0d108a3adcf3fe5b474c2e5ee180c6da2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 201, + 545, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 238 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 201, + 545, + 238 + ], + "type": "inline_equation", + "content": "i \\in \\{LH, HL, HH\\}" + }, + { + "bbox": [ + 304, + 201, + 545, + 238 + ], + "type": "text", + "content": ". For low-frequency information, we only constrain the amplitude components. Consequently, the low-frequency loss can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 244, + 545, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 244, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 361, + 244, + 545, + 259 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a} = \\left| \\left| \\mathcal {A} \\left(I _ {L L} ^ {\\prime}\\right)\\right) - \\mathcal {A} \\left(G _ {L L}\\right) \\right| | _ {1}, \\tag {10}", + "image_path": "733dae235992b60fd1e4c09475fc62799462057a844c00baa14dc06396eaa8ca.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{A}()" + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "content": " refers to the amplitude component in Fourier transform. Finally, We further use an adversarial loss in Wasserstein GAN as reconstruction Loss " + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 307, + 471, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 471, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 471, + 319 + ], + "type": "text", + "content": "3.4. Cross-Frequency Conditioner" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": "The detailed structure of CFC is shown in Figure 3 (c). The purpose of CFC is to facilitate information interaction between high and low frequency features, achieving cross-frequency mutual reinforcement between high and low frequency branches, thus enhancing overall representation power. We denote " + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "inline_equation", + "content": "T_{in}" + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": " as the input features of CFC, which represent high- and low-frequency embeddings. For the high-frequency embedding features " + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "inline_equation", + "content": "T_{in} \\in \\mathbb{R}^{3 \\times \\frac{H}{2} \\times \\frac{W}{2} \\times C}" + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": ", we can obtain " + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "inline_equation", + "content": "T_{LH}, T_{HL}, T_{HH} \\in \\mathbb{R}^{\\frac{H}{2} \\times \\frac{W}{2} \\times c}" + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": " via split operation. By adding these extracted coefficients, we obtain the aggregated high-frequency embeddings. We use different linear projections to construct " + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 325, + 545, + 479 + ], + "type": "text", + "content": " in CFC:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 488, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 488, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 347, + 488, + 545, + 502 + ], + "type": "interline_equation", + "content": "Q = \\operatorname {C o n v} _ {1 \\times 1} \\left(T _ {L H} + T _ {H L} + T _ {H H}\\right), \\tag {11}", + "image_path": "142defa9cfdbdb68c7703fcf9f169ef0392159520361d91e894d277515fdb834.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 381, + 510, + 545, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 510, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 381, + 510, + 545, + 522 + ], + "type": "interline_equation", + "content": "K = \\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {i n}\\right). \\tag {12}", + "image_path": "7e47af4f889e2d9e04c5d5a4284b54fe9f6415518942a775824cfeda15816477.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "inline_equation", + "content": "V_{T}" + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "content": " of high-frequency embeddings and " + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "content": " of low-frequency embeddings can be obtained:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 346, + 557, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 557, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 346, + 557, + 545, + 571 + ], + "type": "interline_equation", + "content": "V _ {T} = \\operatorname {C o n v} _ {1 \\times 1} \\left(T _ {L H} + T _ {H L} + T _ {H H}\\right), \\tag {13}", + "image_path": "02be3c1d856f41d69f71d17c2d5dbf322e9480b277cd0366d637af849cd1853f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 380, + 578, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 578, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 380, + 578, + 545, + 591 + ], + "type": "interline_equation", + "content": "V _ {F} = \\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {i n}\\right). \\tag {14}", + "image_path": "2a1bcf77e20fd9ec9658fd9d0bb754a6f9aa24c5415e645e84898def09ac0b14.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "type": "text", + "content": "The output feature map " + }, + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "type": "inline_equation", + "content": "T_{out}" + }, + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 305, + 595, + 545, + 617 + ], + "type": "text", + "content": " can then be obtained from the formula:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 355, + 623, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 623, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 355, + 623, + 545, + 651 + ], + "type": "interline_equation", + "content": "T _ {o u t} = R \\left(\\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V _ {T}\\right), \\tag {15}", + "image_path": "3c346883d67cc4a097805b24d29b86453b5df7557c9a0a59c94eeea245d7181d.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 363, + 658, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 658, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 363, + 658, + 545, + 685 + ], + "type": "interline_equation", + "content": "F _ {o u t} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V _ {F}, \\tag {16}", + "image_path": "3a3f95e88fe921f5b2f6c365b3c6317b0bb413a466d281ea59e2c8c24d568c80.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " denotes a replication operation, and " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sqrt{d_k}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the number of columns of matrix " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8285" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 227, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 227, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 227, + 85 + ], + "type": "text", + "content": "3.5. Frequency Diffusion Adjustment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 197 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 197 + ], + "type": "text", + "content": "FRDAM aims to further adjust the high- and low-frequency information using the powerful representation of diffusion model. Generally, FRDAM can be divided into two branch, namely low-frequency diffusion branch (LDFB) and high-frequency diffusion branch (HDFB). We adopt the diffusion process proposed in DDPM [10] to construct the residual distribution of high- and low-frequency information for each branch, which can be described as a forward diffusion process and a reverse diffusion process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "text", + "content": "Forward Diffusion Process. The forward diffusion process can be viewed as a Markov chain progressively adding Gaussian noise to the data. Given the initial enhanced frequency components " + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "inline_equation", + "content": "I_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "text", + "content": " and its ground truth " + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "inline_equation", + "content": "i \\in \\{LL, LH, HL, HH\\}" + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "text", + "content": ", we calculate their residual distribution " + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "inline_equation", + "content": "x_{0} = G_{i} - I_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 198, + 287, + 283 + ], + "type": "text", + "content": ", then introduce Gaussian noise based on the time step, as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 294, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 294, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 82, + 294, + 287, + 308 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {17}", + "image_path": "6dd6e2815c0ea91e45751ec0d738e90fc296b02fbae9a64544624918d2b9f76d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "type": "text", + "content": " is a variable controlling the variance of the noise. Introducing " + }, + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 47, + 311, + 286, + 335 + ], + "type": "text", + "content": ", this process can be described as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 339, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 339, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 52, + 339, + 287, + 354 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t - 1}, \\quad \\epsilon_ {t - 1} \\sim \\mathcal {N} (0, \\mathcal {Z}). \\tag {18}", + "image_path": "6a472ab34fa25a888c4a6e03922ae0c1170eed853477eb33928fe131db0982d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 358, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 358, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 358, + 287, + 370 + ], + "type": "text", + "content": "With Gaussian distributions are merged, We can obtain :" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 377, + 287, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 377, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 88, + 377, + 287, + 390 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {19}", + "image_path": "93381a7b718e669e8bfbfe2ed964829966c76aad1d28bd74018cfac6e6477898.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 393, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 287, + 429 + ], + "type": "text", + "content": "Reverse Diffusion Process. The reverse diffusion process aims to restore the residual distribution from the Gaussian noise. The reverse diffusion can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 433, + 287, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 287, + 447 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, x _ {c} ^ {(l)}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right), \\sigma_ {t} ^ {2} \\mathcal {Z}\\right), \\tag {20}", + "image_path": "b9b256874bfd72b6259b24808835345b1d86cba94f448397e4e89a3adff33769.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "text", + "content": "where we take the LDFB as an example, and " + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "inline_equation", + "content": "x_{c}^{(l)}" + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "text", + "content": " refers to the conditional image " + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "inline_equation", + "content": "I_{LL}^{\\prime}" + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{\\theta}(x_t,x_c^{(l)},t)" + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 47, + 453, + 287, + 515 + ], + "type": "text", + "content": " are the mean and variance from the estimate of step t, respectively. In LDFB and HDFB, we follow the setup of [35], they can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 520, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 520, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 52, + 520, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right) = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\left(1 - \\bar {\\alpha} _ {t}\\right)} \\epsilon_ {\\theta} \\left(x _ {t}, x _ {c} ^ {(l)}, t\\right)\\right), (21) \\\\ \\sigma_ {t} ^ {2} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}, (22) \\\\ \\end{array}", + "image_path": "694568a7ee6b89cfe8ea7c84ea79db2493a304e5181cc3e3b25520f088806c41.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 574, + 269, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 269, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 269, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 574, + 269, + 588 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,x_c^{(l)},t)" + }, + { + "bbox": [ + 47, + 574, + 269, + 588 + ], + "type": "text", + "content": " is the estimated value with a Unet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 589, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 287, + 624 + ], + "type": "text", + "content": "We optimize an objective function for the noise estimated by the network and the noise " + }, + { + "bbox": [ + 47, + 589, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon^{(l)}" + }, + { + "bbox": [ + 47, + 589, + 287, + 624 + ], + "type": "text", + "content": " actually added in LDFB. Therefore, the diffusion loss process is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 52, + 628, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 628, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 52, + 628, + 287, + 643 + ], + "type": "interline_equation", + "content": "L _ {d m} (\\theta) = \\left\\| \\epsilon^ {(l)} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon^ {(l)}, x _ {c} ^ {(l)}, t\\right) \\right\\|. \\tag {23}", + "image_path": "76202d95627122e1c491c09998da82abdf9958f30f87f698ce06a7df86a155d1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "text", + "content": "Generally, the frequency diffusion adjustment process is to refine the high- and low-frequency component of the initial enhancement. The whole diffusion process can be formulated:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 60, + 699, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 699, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 60, + 699, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {(i)} = \\mathcal {F} _ {H D F B} \\left(\\epsilon_ {s} ^ {(h)}, I _ {(i)} ^ {\\prime}\\right), i \\in \\{L H, H L, H H \\}, \\tag {24}", + "image_path": "6b7698419313671849d320ce474e19133dd48cdd6dccdd09ab46db8b6c6f073c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 371, + 73, + 545, + 87 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 73, + 545, + 87 + ], + "spans": [ + { + "bbox": [ + 371, + 73, + 545, + 87 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {L L} = \\mathcal {F} _ {L D F B} \\left(\\epsilon_ {s} ^ {(l)}, I _ {L L} ^ {\\prime}\\right), \\tag {25}", + "image_path": "b4039e7d09fe71bd3ccb7c2ba9a0c8ed816bb0f78878af3935b9c13e13734240.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\epsilon_s^{(h)}\\in \\mathbb{R}^{3\\times \\frac{H}{2}\\times \\frac{W}{2}\\times 3}" + }, + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\epsilon_s^{(l)}\\in \\mathbb{R}^{\\frac{H}{2}\\times \\frac{W}{2}\\times 3}" + }, + { + "bbox": [ + 305, + 92, + 545, + 118 + ], + "type": "text", + "content": " are Gaussian noise." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 120, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 545, + 168 + ], + "type": "text", + "content": "Ultimately, the refined frequency components are obtained as the addition of the diffusion generative residual distribution and the initial enhanced frequency components. Then, we employ IDWT to obtain the final generated image:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 178, + 552, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 552, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 552, + 204 + ], + "type": "interline_equation", + "content": "I _ {f i n a l} = I D W T \\left(I _ {(i)} ^ {\\prime} + \\hat {I} _ {(i)}, I _ {L L} ^ {\\prime} + \\hat {I} _ {L L}\\right), i \\in \\{L H, H L, H H \\} \\tag {26}", + "image_path": "3247f2195177150b36d8e7a06658efa9f9b43101898528cfaae7be9c425d5b7c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 208, + 386, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 208, + 386, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 386, + 220 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 228, + 356, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 356, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 356, + 240 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "text", + "content": "Implementation details. Our network, implemented using PyTorch 1.7, underwent training and testing on an NVIDIA GeForce RTX 3090 GPU. We employed the Adam optimizer with " + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "text", + "content": ". The patch size was configured as " + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 246, + 545, + 341 + ], + "type": "text", + "content": ", and the batch size was set to 2. The diffusion model's total time steps, denoted as T, were set to 1000, and the number of training iterations reached one million. The initial learning rate was established at 0.0001." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 342, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 461 + ], + "type": "text", + "content": "Datasets. We utilize the real-world UIEBD dataset [28] and the LSUI dataset [18] for training and evaluating our model. The UIEBD dataset comprises 890 underwater images with corresponding labels. Out of these, 700 images are allocated for training, and the remaining 190 are designated for testing. The LSUI dataset is randomly partitioned into 4500 images for training and 504 images for testing. In addition, to verify the generalization of WF-Diff, we use non-reference benchmarks U45 [20], which contains 45 underwater images for testing." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 462, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 545, + 557 + ], + "type": "text", + "content": "Comparison methods. We conduct a comparative analysis between WF-Diff and eight state-of-the-art (SOTA) UIE methods, namely UIEC^2-Net [39], Water-Net [28], UWCNN [3], SCNet [8], UIEWD [25], U-color [19], U-shape [18], and DM-water [36]. To ensure a fair and rigorous comparison, we utilize the provided source codes from the respective authors and adhere strictly to the identical experimental settings across all evaluations." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We primarily utilize well-established full-reference image quality assessment metrics: PSNR and SSIM [41]. PSNR and SSIM offer quantitative comparisons of our method with other approaches at both pixel and structural levels. Higher PSNR and SSIM values signify superior quality of the generated images. Additionally, we incorporate the LPIPS and FID metrics for full-reference image evaluation. LPIPS [47] is a deep neural network-based image quality metric that assesses the perceptual similarity between an image and a reference image. FID [9] measures the distance between the distributions of real and generated images. A lower LPIPS and FID score indicates a more effective UIE approach. For non-reference benchmarks U45," + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8286" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 97, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of different UIE methods on the UIEBD, LSUI and U45 datasets. The best results are highlighted in bold and the second best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 97, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 47, + 97, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 547, + 266 + ], + "type": "table", + "html": "
MethodsUIEWDUWCNNUIEC^2-NetWater-NetSCNetU-colorU-shapeDM-waterOurs
UIBDFID↓85.1294.4435.0637.4833.6638.2546.1131.0727.85
LPIPS↓0.39560.35250.20330.21160.24970.23370.22640.14360.1248
PSNR↑14.6515.4020.1419.3520.4120.7121.2521.8823.86
SSIM↑0.72650.77490.82150.83210.82350.84110.84530.81940.8730
LSUIFID↓98.49100.534.5138.90158.9945.0628.5627.9126.75
LPIPS↓0.39620.34500.14320.16780.2830.1230.10280.11380.1096
PSNR↑15.4318.2420.8619.7322.6322.9124.1627.6527.26
SSIM↑0.78020.84650.88670.82260.91760.89020.93220.88670.9437
U45UIQM↑2.4582.3792.7802.9572.8563.1043.1513.0863.181
UCIQE↑0.5830.5670.5910.6010.5940.5860.5920.6340.619
", + "image_path": "bf471245e283cb23dca867690861b8ba5ed652cc830a89f9ce0ef336795dab07.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 272, + 542, + 509 + ], + "blocks": [ + { + "bbox": [ + 51, + 272, + 542, + 509 + ], + "lines": [ + { + "bbox": [ + 51, + 272, + 542, + 509 + ], + "spans": [ + { + "bbox": [ + 51, + 272, + 542, + 509 + ], + "type": "image", + "image_path": "1bfa9fdf1354aa95921dc1c522da95e1d126ca8b39acaedad87f907dbe181caf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 516, + 456, + 529 + ], + "lines": [ + { + "bbox": [ + 137, + 516, + 456, + 529 + ], + "spans": [ + { + "bbox": [ + 137, + 516, + 456, + 529 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with other SOTA methods on real underwater images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 548, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 288, + 571 + ], + "type": "text", + "content": "we introduce UIQM [27] and UCIQE [44] to evaluate our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 585, + 190, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 190, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 190, + 599 + ], + "type": "text", + "content": "4.2. Results and Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Table 2 shows the quantitative results compared with different baselines on UIEBD, LSUI and U45 datasets, including with UIEC^2-Net, Water-Net, UIEWD, UWCNN, SCNet, U-color, U-shape, and DM-water. We mainly use PSNR, SSIM, LPIPS and FID as our quantitative indices for UIEBD and LSUI datasets, and UIQM and UCIQE for non-reference dataset U45. The results in Table 2 show that our algorithm outperforms state-of-the-art methods obviously, and achieve state-of-the-art performance in terms of image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "type": "text", + "content": "quality evaluation metrics on UIE task, which verifies the robustness of the proposed WF-Diff. To better validate the superiority of our methods, in Figure 4, we show the visual results comparison with state-of-the-art methods on real underwater images. The six examples are randomly selected on the UIEBD and LSUI datasets. Our methods consistently generate natural and better visual results on testing images, strongly proving that WF-Diff has good generalization performance for real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 669, + 400, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 400, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 400, + 682 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Ablation study with WFI2-net. In order to evaluate the effectiveness of each part in WFI2-net, we conduct two ab" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8287" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 102, + 269, + 209 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 3. Ablation study with network structure of WFI2-net on UIEBD dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 102, + 269, + 209 + ], + "lines": [ + { + "bbox": [ + 67, + 102, + 269, + 209 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 269, + 209 + ], + "type": "table", + "html": "
WTBSFFBUIEBD
SACASDUFDUPSNR ↑SSIM↑
×✓✓✓20.940.8541
✓×✓✓20.820.8473
✓✓×✓21.110.8586
✓✓✓×20.230.8346
✓✓✓✓21.870.8622
", + "image_path": "0a5351001393d3ea50984e02a541a4c4eb5ab59a80c32f715dd4b3eeb499d67a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 242, + 269, + 352 + ], + "blocks": [ + { + "bbox": [ + 47, + 221, + 286, + 233 + ], + "lines": [ + { + "bbox": [ + 47, + 221, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 286, + 233 + ], + "type": "text", + "content": "Table 4. Ablation study with loss of WFI2-net on UIEBD dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 242, + 269, + 352 + ], + "lines": [ + { + "bbox": [ + 67, + 242, + 269, + 352 + ], + "spans": [ + { + "bbox": [ + 67, + 242, + 269, + 352 + ], + "type": "table", + "html": "
MethodsUIEBD
LhLrecLaCFCPSNR ↑SSIM↑
×✓✓✓20.460.8274
✓×✓✓20.650.8399
✓✓×✓19.810.8213
✓✓✓×20.970.8425
✓✓✓✓21.870.8622
", + "image_path": "72e4a556e824321f432f678f8a97356819f7e8fd91213b88ce69438e22c80196.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 47, + 429, + 287, + 529 + ], + "blocks": [ + { + "bbox": [ + 46, + 364, + 287, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 419 + ], + "type": "text", + "content": "Table 5. Ablation study with FRDAM on UIEBD dataset. DM refers to diffusion model (DM), RDM refers to residual DM. D-L refers to refining low-frequency component in wavelet space, and D-H refers to refining high-frequency components in wavelet space." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 429, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 47, + 429, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 287, + 529 + ], + "type": "table", + "html": "
MethodDMRDMD-LD-HCFCPSNR↑
A✓××××20.86
B×✓×××22.37
C×✓✓××22.32
D×✓×✓×22.58
E×✓✓✓×23.44
F×✓✓✓✓23.86
", + "image_path": "1a1e9c99034a20db128705ef6ccc836d5f5129eb9fe3665a02909924a27a73cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 543, + 106, + 578 + ], + "blocks": [ + { + "bbox": [ + 50, + 543, + 106, + 578 + ], + "lines": [ + { + "bbox": [ + 50, + 543, + 106, + 578 + ], + "spans": [ + { + "bbox": [ + 50, + 543, + 106, + 578 + ], + "type": "image", + "image_path": "1ce8f5b83440314316edfeb324a054b8a0f8e33615e89a923271832e831b0ca3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 109, + 543, + 165, + 579 + ], + "blocks": [ + { + "bbox": [ + 109, + 543, + 165, + 579 + ], + "lines": [ + { + "bbox": [ + 109, + 543, + 165, + 579 + ], + "spans": [ + { + "bbox": [ + 109, + 543, + 165, + 579 + ], + "type": "image", + "image_path": "a42bb405fb55c40cf68d32fafc8dc8e634dc63d81671a490de5722991eb1c0da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 580, + 143, + 587 + ], + "lines": [ + { + "bbox": [ + 135, + 580, + 143, + 587 + ], + "spans": [ + { + "bbox": [ + 135, + 580, + 143, + 587 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 168, + 543, + 223, + 579 + ], + "blocks": [ + { + "bbox": [ + 168, + 543, + 223, + 579 + ], + "lines": [ + { + "bbox": [ + 168, + 543, + 223, + 579 + ], + "spans": [ + { + "bbox": [ + 168, + 543, + 223, + 579 + ], + "type": "image", + "image_path": "8e0035f3d8970625de9e5a3ea51832ef0be936d4aca831c9161c3bfc1afd595a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 580, + 202, + 587 + ], + "lines": [ + { + "bbox": [ + 195, + 580, + 202, + 587 + ], + "spans": [ + { + "bbox": [ + 195, + 580, + 202, + 587 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 228, + 543, + 284, + 579 + ], + "blocks": [ + { + "bbox": [ + 228, + 543, + 284, + 579 + ], + "lines": [ + { + "bbox": [ + 228, + 543, + 284, + 579 + ], + "spans": [ + { + "bbox": [ + 228, + 543, + 284, + 579 + ], + "type": "image", + "image_path": "ec570d0ab29c9a8cb0af6105ae1e6865ce883378840e48e17e6748ecc45b4bb3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 580, + 262, + 587 + ], + "lines": [ + { + "bbox": [ + 255, + 580, + 262, + 587 + ], + "spans": [ + { + "bbox": [ + 255, + 580, + 262, + 587 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 588, + 107, + 624 + ], + "blocks": [ + { + "bbox": [ + 72, + 580, + 91, + 588 + ], + "lines": [ + { + "bbox": [ + 72, + 580, + 91, + 588 + ], + "spans": [ + { + "bbox": [ + 72, + 580, + 91, + 588 + ], + "type": "text", + "content": "input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 588, + 107, + 624 + ], + "lines": [ + { + "bbox": [ + 51, + 588, + 107, + 624 + ], + "spans": [ + { + "bbox": [ + 51, + 588, + 107, + 624 + ], + "type": "image", + "image_path": "86d01bf4be972e6bae92526334c2cc007e285d6a93d59a57dca0dec27fcde04c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 625, + 83, + 632 + ], + "lines": [ + { + "bbox": [ + 76, + 625, + 83, + 632 + ], + "spans": [ + { + "bbox": [ + 76, + 625, + 83, + 632 + ], + "type": "text", + "content": "D" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 643, + 268, + 654 + ], + "lines": [ + { + "bbox": [ + 65, + 643, + 268, + 654 + ], + "spans": [ + { + "bbox": [ + 65, + 643, + 268, + 654 + ], + "type": "text", + "content": "Figure 5. Visual results of ablation study with FRDAM." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 109, + 588, + 165, + 624 + ], + "blocks": [ + { + "bbox": [ + 109, + 588, + 165, + 624 + ], + "lines": [ + { + "bbox": [ + 109, + 588, + 165, + 624 + ], + "spans": [ + { + "bbox": [ + 109, + 588, + 165, + 624 + ], + "type": "image", + "image_path": "31b24dcea85bb04a297863eafe656f8b463997c378e726ce1b611a4eb3987200.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 625, + 142, + 632 + ], + "lines": [ + { + "bbox": [ + 135, + 625, + 142, + 632 + ], + "spans": [ + { + "bbox": [ + 135, + 625, + 142, + 632 + ], + "type": "text", + "content": "E" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 168, + 588, + 224, + 624 + ], + "blocks": [ + { + "bbox": [ + 168, + 588, + 224, + 624 + ], + "lines": [ + { + "bbox": [ + 168, + 588, + 224, + 624 + ], + "spans": [ + { + "bbox": [ + 168, + 588, + 224, + 624 + ], + "type": "image", + "image_path": "59256db715b1dc3377efcc9f4cd5b3f4acce04bf12127dbe294dbbcc2617ad21.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 625, + 201, + 632 + ], + "lines": [ + { + "bbox": [ + 195, + 625, + 201, + 632 + ], + "spans": [ + { + "bbox": [ + 195, + 625, + 201, + 632 + ], + "type": "text", + "content": "F" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 228, + 588, + 284, + 624 + ], + "blocks": [ + { + "bbox": [ + 228, + 588, + 284, + 624 + ], + "lines": [ + { + "bbox": [ + 228, + 588, + 284, + 624 + ], + "spans": [ + { + "bbox": [ + 228, + 588, + 284, + 624 + ], + "type": "image", + "image_path": "0515a0e928e77986f2670c5a2d47c600afbfa759645df275dfbe996e8ebcc833.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 625, + 263, + 632 + ], + "lines": [ + { + "bbox": [ + 252, + 625, + 263, + 632 + ], + "spans": [ + { + "bbox": [ + 252, + 625, + 263, + 632 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "lation experiments with WFI2-net on the UIEBD dataset in Table 3 and 4. CFC is cross-frequency conditioner. Note that, we do not discuss the effect of our proposed FRDAM" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 314, + 86, + 537, + 113 + ], + "blocks": [ + { + "bbox": [ + 338, + 71, + 512, + 82 + ], + "lines": [ + { + "bbox": [ + 338, + 71, + 512, + 82 + ], + "spans": [ + { + "bbox": [ + 338, + 71, + 512, + 82 + ], + "type": "text", + "content": "Table 6. Comparison results of inference speed." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 86, + 537, + 113 + ], + "lines": [ + { + "bbox": [ + 314, + 86, + 537, + 113 + ], + "spans": [ + { + "bbox": [ + 314, + 86, + 537, + 113 + ], + "type": "table", + "html": "
MethodSCNetUshapeDM-waterWFI2-netWF-Diff
Inference time (s)0.01490.03530.14410.07390.2816
", + "image_path": "fec30c812c6baee906888bed4283e7bc0f9200b64bfab615470328da486bd690.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 314, + 133, + 537, + 162 + ], + "blocks": [ + { + "bbox": [ + 353, + 121, + 496, + 131 + ], + "lines": [ + { + "bbox": [ + 353, + 121, + 496, + 131 + ], + "spans": [ + { + "bbox": [ + 353, + 121, + 496, + 131 + ], + "type": "text", + "content": "Table 7. Comparison results of FLOPs." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 133, + 537, + 162 + ], + "lines": [ + { + "bbox": [ + 314, + 133, + 537, + 162 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 537, + 162 + ], + "type": "table", + "html": "
MethodSCNetWaterNetUcolorUshapeWFI2-net
FLOPs (G)5.88193.7443.866.294.1
", + "image_path": "2816a0de9395b0184aaef7b2285708c0540a15e6452ddcd993f5d29527d8bf4f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 179, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 545, + 227 + ], + "type": "text", + "content": "here. We regularly remove one component to each configuration at one time, and our strategy achieves the best performance by using all loss functions and blocks, proving that each part of WFI2-net is useful for UIE task." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": "Ablation study with FRDAM. In this section, we will discuss the effectiveness of FRDAM. Table 5 shows the quantitative results on the UIEBD dataset, and Figure 5 shows the visual results. Note that, Model A and B achieve diffusion process in the pixel level, and model C, D, E and E achieve diffusion process in wavelet space. Model A obtains relatively poor results in PSNR and generates images with color distortion or artifacts in Figure 5 due to the diversity of the sampling process. Model B does not yield fully satisfactory results, because it needs to adjust both the high and low-frequency information in images, which limits their ability to focus on fine-grained information. Compared to model C, model D achieves better results, suggesting that the degradation information is mainly in the high-frequency information in frequency diffusion adjustment stage. Model F achieves the best performance, proving that our designed FRDAM is best for UIE task." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 445, + 378, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 378, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 378, + 456 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 468, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 564 + ], + "type": "text", + "content": "In this paper, we develop a novel UIE framework, namely WF-Diff. With fully utilizing the frequency domain characteristics and diffusion model, WFI2-net can achieve enhancement and adjustment of frequency information. Our proposed FRDAM is a plug-and-play universal module to adjust the details of the underwater images. WF-Diff shows SOTA performance on UIE task, and extensive ablation experiments prove that each of our contributions is effective." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "content": "Limitations. As a result of employing two diffusion models, our approach doesn't confer an advantage in terms of inference speed. Table 6 presents the actual values of inference speed for WF-Diff, indicating that our model does not outperform recent approaches in terms of inference time. Furthermore, Table 7 provides a comparison of FLOPs for WFI2-net. It should be mentioned that the implicit sampling step is set to 10 for WF-Diff. Hence, in the future, we will delve into methods to expedite the sampling process." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant No. 62276138 and 62371232)." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8288" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Derya Akkaynak and Tali Treibitz. Sea-thru: A method for removing water from underwater images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1682-1691. Computer Vision Foundation / IEEE, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "type": "text", + "content": "[2] Derya Akkaynak, Tali Treibitz, Tom Shlesinger, Yossi Loya, Raz Tamir, and David Iluz. What is the space of attenuation coefficients in underwater computer vision? In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 568-577, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 235 + ], + "type": "text", + "content": "[3] Saeed Anwar, Chongyi Li, and Fatih Porikli. Deep underwater image enhancement. CoRR, abs/1807.03528, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "text", + "content": "[4] John Yi-Wu Chiang and Ying-Ching Chen. Underwater image enhancement by wavelength compensation and dehazing. IEEE Trans. Image Process., 21(4):1756-1769, 2012. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "type": "text", + "content": "[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14347-14356. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 403 + ], + "type": "text", + "content": "[6] Karin de Langis and Junaed Sattar. Realtime multi-diver tracking and re-identification for underwater human-robot collaboration. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 11140-11146, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 458 + ], + "type": "text", + "content": "[7] Cameron Fabbri, Md Jahidul Islam, and Junaed Sattar. Enhancing underwater imagery using generative adversarial networks. In 2018 IEEE International Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May 21-25, 2018, pages 7159-7165. IEEE, 2018. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 525 + ], + "type": "text", + "content": "[8] Zhenqi Fu, Xiaopeng Lin, Wu Wang, Yue Huang, and Xinghao Ding. Underwater image enhancement via learning water type desensitized representations. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2764-2768. IEEE, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 602 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[11] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based CNN for multi-scale face super resolution. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1698-1706, 2017. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "text", + "content": "[12] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. pages 163-180, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "text", + "content": "[13] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18145-18155. IEEE, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "text", + "content": "[14] Md Jahidul Islam, Youya Xia, and Junaed Sattar. Fast underwater image enhancement for improved visual perception. IEEE Robotics Autom. Lett., 5(2):3227-3234, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 285 + ], + "type": "text", + "content": "[15] Paulo Drews Jr., Erickson Rangel do Nascimento, F. Moraes, Silvia S. C. Botelho, and Mario F. M. Campos. Transmission estimation in underwater single images. In 2013 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 825-830, 2013. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[16] Eunhee Kang, Won Chang, Jae Jun Yoo, and Jong Chul Ye. Deep convolutional framelet denoising for low-dose CT via wavelet residual network. IEEE Trans. Medical Imaging, 37 (6):1358-1369, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 376 + ], + "type": "text", + "content": "[17] Chongyi Li, Jichang Guo, Runmin Cong, Yanwei Pang, and Bo Wang. Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior. IEEE Trans. Image Process., 25(12):5664-5677, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "text", + "content": "[18] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Trans. Image Process., 29:4376-4389, 2020. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[19] Chongyi Li, Saeed Anwar, Junhui Hou, Runmin Cong, Chunle Guo, and Wenqi Ren. Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE Trans. Image Process., 30:4985-5000, 2021. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "text", + "content": "[20] Hanyu Li, Jingjing Li, and Wei Wang. A fusion adversarial underwater image enhancement network with a public test dataset. arXiv preprint arXiv:1906.06819, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[21] Jie Li, Katherine A. Skinner, Ryan M. Eustice, and Matthew Johnson-Roberson. Watergan: Unsupervised generative network to enable real-time color correction of monocular underwater images. IEEE Robotics Autom. Lett., 3(1):387-394, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 568, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 634 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[23] Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-Icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[24] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8289" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Ziyin Ma and Changjae Oh. A wavelet-based dual-stream network for underwater image enhancement. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022, pages 2769-2773. IEEE, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[26] James McMahon and Erion Plaku. Autonomous data collection with timed communication constraints for unmanned underwater vehicles. IEEE Robotics Autom. Lett., 6(2): 1832-1839, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[27] Karen Panetta, Chen Gao, and Sos Agaian. Human-visual-system-inspired underwater image quality measures. IEEE Journal of Oceanic Engineering, 41(3):541-551, 2015. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[28] Lintao Peng, Chunli Zhu, and Liheng Bian. U-shape transformer for underwater image enhancement. IEEE Trans. Image Process., 32:3066-3079, 2023. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[29] Yan-Tsung Peng and Pamela C. Cosman. Underwater image restoration based on image blurriness and light absorption. IEEE Trans. Image Process., 26(4):1579-1594, 2017. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[30] Yan-Tsung Peng, Keming Cao, and Pamela C. Cosman. Generalization of the dark channel prior for single image restoration. IEEE Trans. Image Process., 27(6):2856-2868, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "text", + "content": "[31] Priyadharsini Ravisankar, T. Sree Sharmila, and V. Rajendran. A wavelet transform based contrast enhancement method for underwater acoustic images. Multidimens. Syst. Signal Process., 29(4):1845-1859, 2018. 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 367, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 432 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 510 + ], + "type": "text", + "content": "[33] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 15:1-15:10. ACM, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[36] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 5419-5427. ACM, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[37] Pritish M. Uplavikar, Zhenyu Wu, and Zhangyang Wang. All-in-one underwater image enhancement using domain" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1-8, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[38] Yi Wang, Hui Liu, and Lap-Pui Chau. Single underwater image restoration using adaptive attenuation-curve prior. IEEE Trans. Circuits Syst. I Regul. Pap., 65-I(3):992-1002, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "text", + "content": "[39] Yudong Wang, Jichang Guo, Huan Gao, and HuiHui Yue. Uiec" + }, + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "text", + "content": " -net: Cnn-based underwater image enhancement using two color space. Signal Process. Image Commun., 96: 116250, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "text", + "content": "[40] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "text", + "content": "[41] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 305, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 370 + ], + "type": "text", + "content": "[42] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16272-16282. IEEE, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "type": "text", + "content": "[43] Jian Yang, Chen Li, and Xuelong Li. Underwater image restoration with light-aware progressive network. In IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023, pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 427, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 545, + 459 + ], + "type": "text", + "content": "[44] Miao Yang and Arcot Sowmya. An underwater color image quality evaluation metric. IEEE Transactions on Image Processing, 24(12):6062-6071, 2015. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 535 + ], + "type": "text", + "content": "[45] Zongyuan Yang, Baolin Liu, Yongping Xiong, Lan Yi, Guibin Wu, Xiaojun Tang, Ziqi Liu, Junjie Zhou, and Xing Zhang. Docdiff: Document enhancement via residual diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 2795-2806. ACM, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 580 + ], + "type": "text", + "content": "[46] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with A generative diffusion model. CoRR, abs/2308.13164, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 582, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 645 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[48] Chen Zhao, Wei-Ling Cai, Chenyu Dong, and Ziqi Zeng. Toward sufficient spatial-frequency interaction for gradient-aware underwater image enhancement. arXiv preprint arXiv:2202.08537, 2023. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[49] Chen Zhao, Wei-Ling Cai, and Zheng Yuan. Spectral normalization and dual contrastive regularization for image-to" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8290" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 207 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "image translation. The Visual Computer, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[50] Chen Zhao, Chenyu Dong, and Weiling Cai. Learning a physical-aware diffusion model based on transformer for underwater image enhancement. arXiv preprint arXiv:2403.01497, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "text", + "content": "[51] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[52] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. arXiv preprint arXiv:2402.05408, 2024. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "8291" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_content_list.json b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5edc3811472b8508309c2dbb466fbccabf910d12 --- /dev/null +++ b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_content_list.json @@ -0,0 +1,1610 @@ +[ + { + "type": "text", + "text": "Weak-to-Strong 3D Object Detection with X-Ray Distillation", + "text_level": 1, + "bbox": [ + 179, + 130, + 789, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexander Gambashidze\\*1,2", + "bbox": [ + 158, + 179, + 380, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aleksandr Dadukin", + "bbox": [ + 429, + 180, + 591, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maxim Golyadkin1,2", + "bbox": [ + 640, + 180, + 807, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maria Razzhivina", + "bbox": [ + 308, + 205, + 459, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ilya Makarov1,3", + "bbox": [ + 532, + 205, + 658, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 258, + 312, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper addresses the critical challenges of sparsity and occlusion in LiDAR-based 3D object detection. Current methods often rely on supplementary modules or specific architectural designs, potentially limiting their applicability to new and evolving architectures. To our knowledge, we are the first to propose a versatile technique that seamlessly integrates into any existing framework for 3D Object Detection, marking the first instance of Weak-to-Strong generalization in 3D computer vision. We introduce a novel framework, X-Ray Distillation with Object-Complete Frames, suitable for both supervised and semi-supervised settings, that leverages the temporal aspect of point cloud sequences. This method extracts crucial information from both previous and subsequent LiDAR frames, creating Object-Complete frames that represent objects from multiple viewpoints, thus addressing occlusion and sparsity. Given the limitation of not being able to generate Object-Complete frames during online inference, we utilize Knowledge Distillation within a Teacher-Student framework. This technique encourages the strong Student model to emulate the behavior of the weaker Teacher, which processes simple and informative Object-Complete frames, effectively offering a comprehensive view of objects as if seen through X-ray vision. Our proposed methods surpass state-of-the-art in semi-supervised learning by 1-1.5 mAP and enhance the performance of five established supervised models by 1-2 mAP on standard autonomous driving datasets, even with default hyperparameters. Code for Object-Complete frames is available here: https://github.com/sakharok13/X-Ray-Teacher-Patching-Tools.", + "bbox": [ + 75, + 309, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00632d629e00b25464ef5870158fe326a1d02ef4d2b4313bdcc4341c9666ed51.jpg", + "image_caption": [ + "Figure 1. 3D object detection directly from sparse LiDAR data (top left) provides noisy predictions (bottom left). Adding object completion stage (top right) helps to train 3D object detection X-Ray Teacher, which is robust and can be distilled to baseline model. Red, Yellow and Blue colors of bounding boxes are related to classical LiDAR-based object detection, Our model on Object-Complete frames predictions and Ground Truth labels, respectively." + ], + "image_footnote": [], + "bbox": [ + 511, + 306, + 877, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "You're just not thinking fourth dimensionally... the bridge will exist.", + "bbox": [ + 560, + 707, + 864, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dr. Emmett Brown, Back to the Future", + "bbox": [ + 660, + 743, + 890, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 784, + 630, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection is a fundamental task in the field of computer vision and autonomous systems, playing a key role in the advancement of self-driving technology [1] and contributing significantly to the robotics industry [38, 51]. Currently, LiDAR-based 3D object detection [6, 17, 36] demonstrates superior performance compared to camera-", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author, email: alexandergambashidze@gmail.com", + "bbox": [ + 93, + 849, + 447, + 862 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Artificial Intelligence Research Institute", + "bbox": [ + 98, + 862, + 313, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ HSE University", + "bbox": [ + 98, + 875, + 187, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3ISP RAS", + "bbox": [ + 98, + 887, + 151, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based [19, 39, 40, 45] and radar-based [3, 8, 22, 25] approaches. Furthermore, LiDAR point clouds are the key ingredient of multimodal fusion-based approaches [14, 21, 37], so LiDAR-based 3D detection continues to be a strong focus of the research community.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The point cloud-based 3D object detection challenges the following issues: sparsity, occlusions, and the complexity of 3D data annotation. Sparsity: large point clouds are sparse due to the LiDAR sensing process's inherent characteristics that leads to an imprecise representation of the captured scene. Additionally, there is an imbalance in point density. In particular, the point cloud is sparser in the far range and contains less spatial information affecting feature representation and box prediction. Occlusions: another problem arises from the frequent occurrence of partial occlusion in LiDAR frames. This primarily happens due to the fact that the frames are acquired from a single fixed point of view, making them essentially 2.5D. To detect and accurately locate a highly occluded object, a detector must recognize the hidden shapes of the object even when a significant portion of its parts is missing. Since the absence of certain shapes inevitably affects object perception, this becomes a critical detection challenge. Our previous approaches focused on depth completion [23], depth inpainting [30] or self-supervised depth pretraining [15, 16] could not address occlusion problems. Data annotation: finally, data annotation is a formidable challenge in 3D object detection due to the complexity of annotating objects in three-dimensional space. For example, a skilled annotator can spend weeks annotating just one hour of LiDAR data [9, 24]. This problem is partially addressed by semi-supervised learning approaches in a teacher-student framework [33, 35, 46, 48]. However, the quality of pseudo labels and the performance of these methods are limited due to the aforementioned sparsity and occlusion problems. Therefore, overcoming these inherent challenges is essential to improve the accuracy and reliability of 3D detection systems.", + "bbox": [ + 75, + 170, + 472, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The challenges posed by sparsity and occlusion have led to the formulation of several methodologies. Some of them [42] are aimed at improving the computational efficiency of processing sparse data without improving the quality of the predictions. Others, including our work, focus on improving detection performance on sparse occluded data [18, 20, 26, 41]. Li et al. [18] enforce shape constraints to improve object localization with explicit shape priors obtained from a database of CAD models. Najibi et al. [26] and Li et al. [20] do the same by implicitly introducing priors with novel modules pretrained for point cloud completion and SDF approximation. Xu et al. [41] propose a Shape Occupancy Probability estimation module to refine predicted bounding boxes for occluded objects. However, all of these mentioned approaches utilize new modules and", + "bbox": [ + 75, + 674, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adapt the model architecture to take advantage of them. As a result, these methods have limited applicability to new architectures that will emerge as the field evolves.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address all three major challenges in 3D object detection with our novel framework called X-Ray Distillation with Object-Complete Frames that is easily plugged in existing approaches and new architectures. It is designed for universal application to any LiDAR-based detector, improving performance on sparse and occluded objects. Our approach exploits the properties of existing large-scale autonomous driving datasets, which consist of sequences of LiDAR frames. Such property makes it possible to reconstruct complete shapes for occluded objects using other occurrences of these objects in the sequence, ensuring that all objects are equipped with points from all available viewpoints in a scene. We then use this completed data in the Teacher-Student framework for both semi-supervised learning and knowledge distillation in a supervised setting. We train our Teacher on extremely informative Object-Complete frames thus making it a weaker model [4]. Then, we use it to extract features from such simple Object-Complete frames and distills this knowledge to a stronger Student, which operates with the original data, to guide him on how to extract rich features from occluded objects. To generate Object-Complete frames, we leverage ground truth object tracking labels. Since there are no labels for the majority of data in the semi-supervised setting, we propose an Objects Temporal Fusion block to detect, track, and use point cloud registration techniques to construct Object-Complete frames.", + "bbox": [ + 496, + 146, + 892, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate the proposed X-Ray Teacher framework on nuScenes [5] and Waymo Open Dataset [32] 3D object detection benchmarks for SECOND [42], CenterPoint [47], and DSVT [36] models in a supervised learning paradigm. Experiments show a steady improvement by 1-2 mAP with minimal or no impact on time and computational resources during the inference stage. For semi-supervised performance evaluation, we use ONCE [24] benchmark.", + "bbox": [ + 496, + 561, + 890, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Applying these novel ideas for 3D object detection, we provide the following contributions:", + "bbox": [ + 496, + 691, + 890, + 722 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose the X-Ray Teacher framework for semi-supervised learning, which achieves state-of-the-art performance on the ONCE benchmark.", + "2. We show that our approach improves the quality of four supervised learning models, including the current state-of-the-art model, and demonstrates the potential to improve the performance of any supervised model trained on sequential data.", + "3. We suggest the Objects Temporal Fusion block to generate Object-Complete frames for data that lacks ground truth tracking labels." + ], + "bbox": [ + 500, + 731, + 890, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15056", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 90, + 215, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Object Detection for point clouds", + "text_level": 1, + "bbox": [ + 76, + 114, + 367, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modern 3D detectors predominantly use point clouds, voxel representations, or a combination of both. The pioneering PointNet model family [28, 29] provided a significant step forward in 3D recognition. However, direct processing of point clouds poses several challenges: it requires point sampling, grouping, and computation of point-wise features, which can be computationally intensive. To integrate insights from 2D computer vision, a transition to a pixel equivalent, namely voxels, is required. Methods such as [41, 43, 47, 49] first convert point clouds to voxels, followed by the use of 3D sparse convolutions [12]. Recent advances also include the integration of modified self-attention layers, achieving state-of-the-art results [36].", + "bbox": [ + 76, + 138, + 468, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Semi-supervised 3D detection", + "text_level": 1, + "bbox": [ + 76, + 344, + 339, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although Semi-supervised 3D object detection is not as extensively researched as 2D detection, it has seen some significant contributions. Among these, SESS [48] stands out for using the general Mean Teacher approach with data augmentations and consistency loss. Similarly, 3D IoUMatch [35] is notable for its unique localization strategies and 3D IoU-guided techniques for box filtering. Proficient Teachers [46] is notable for its box voting and contrastive losses. We do not propose yet another standalone method for Semi-supervised 3D Object Detection; rather, our emphasis is on creating a plug-in technique designed to augment and enhance the performance of existing methods.", + "bbox": [ + 76, + 367, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Knowledge Distillation", + "text_level": 1, + "bbox": [ + 76, + 559, + 289, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The concept of Knowledge Distillation (KD) was first introduced by Hinton et al. [10]. KD describes a learning approach where a larger teacher network guides the training of a smaller student network for various tasks [13]. Broadly, KD methodologies are categorized into two types: logits/regression distillation and feature map distillation. Our focus is on a hybrid approach, combining these methods: matching feature maps, regression, and classification heads with pseudo labels generated by the teacher model.", + "bbox": [ + 76, + 580, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In 3D object detection, Knowledge Distillation is mainly utilized to minimize parameters and FLOPS while aiming to preserve box prediction quality [7, 44]. Yet, methods that aim to outperform state-of-the-art models are uncommon. In contrast to this trend, our X-Ray Teacher model challenges the norm by providing the quality improvement.", + "bbox": [ + 76, + 717, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 821, + 210, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of X-Ray Teacher", + "text_level": 1, + "bbox": [ + 76, + 847, + 326, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a novel training framework to address the challenges of sparsity and occlusion in 3D Object Detec", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion based on LiDAR data. This framework is not limited to any specific object detection model and has the potential for applications across various deep learning architectures. Our method is designed to process the LiDAR data structured as a sequence of frames.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The two core elements of our approach are Object Complete Frames Generation and Teacher-Student Knowledge Distillation. Our approach for 3D object detection can be applied in both, supervised and semi-supervised settings, with minor differences in the elements implementation.", + "bbox": [ + 496, + 169, + 890, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Complete Frames Generation. In this step, we reconstruct the complete shapes of objects presented in the scene by utilizing information from the other frames within the same sequence. Given that autonomous driving datasets are composed of sequential data, we can efficiently leverage their temporal nature: we add points from both the future and the past when objects are observed from different viewpoints. It allows us to reconstruct the complete shapes of objects without shape databases or reconstruction modules.", + "bbox": [ + 496, + 244, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to verify the validity of our approach, we trained a CenterPoint [47] model on both, the original and object-complete NuScenes [5] datasets, and then evaluated their performance on the respective validation sets. The models trained on the original and the object-complete frames achieved mAP scores of $59.2\\%$ and $79.5\\%$ , respectively. This difference of $20~\\mathrm{mAP}$ suggests that 1) it would be beneficial to transform unlabeled original point clouds into Object-Complete ones and to annotate them with a corresponding X-Ray Teacher pretrained on such informative frames 2) the features extracted by a weaker X-Ray Teacher might be distilled to a stronger student to share the knowledge of complete shapes.", + "bbox": [ + 496, + 383, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Teacher-Student Knowledge Distillation. The necessity for this step arises because we cannot generate Object Complete Frames during the online inference stage, as it is not possible to access future data. Therefore, we need to encourage the model to behave as if it were observing shape-complete objects, even when dealing with occluded ones. A well-known method for enabling a deep learning model to imitate another model's behavior involves using Knowledge Distillation within a Teacher-Student framework. However, for conventional Knowledge Distillation, both the Teacher and Student models typically process data of the same complexity (the only difference may lie in the complexity of augmentations [31]). In contrast to the standard knowledge distillation, we enrich the data for training Teacher, which significantly improves its performance in 3D object detection. Then, we teach the Student to extract important information from less detailed data by distilling knowledge from the Teacher model.", + "bbox": [ + 496, + 580, + 890, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of simplifying the Student model, as it is usually done in standard Knowledge Distillation, we take an opposite approach and design the Student to be more complex", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15057", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c60898a961fecfaffa9e7f9afbd0deff0743ee49aa7e95fa41691618e859d675.jpg", + "image_caption": [ + "Figure 2. Overall X-Ray Knowledge Distillation for Supervised Learning. X-Ray Teacher is frozen and pretrained on Object-Complete frames which are taken as input. The Student is guided to mimic the Teacher's behaviour through Knowledge Distillation losses: $L_{feat}$ for intermediate embeddings matching, $L_{reg}$ for bounding box regression, $L_{det}$ for basic detection, and $L_{cls}$ for classification." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 851, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "than the Teacher. It helps to extract high-quality information from more intricate and ambiguous data and demand a Student model to be more robust and to have a more complex receptive field capacity.", + "bbox": [ + 75, + 340, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In what follows, we provide detailed descriptions of how we implement Object Complete Frames Generation and Teacher-Student Knowledge Distillation steps in both supervised and semi-supervised settings.", + "bbox": [ + 75, + 402, + 468, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Supervised X-Ray Teacher", + "text_level": 1, + "bbox": [ + 76, + 469, + 316, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the supervised setting of 3D Object Detection, models are trained and evaluated using datasets that provide labeled data, including precise bounding boxes and instance IDs. Object-Complete Frame Generation for labeled data involves aggregating objects based on their instance IDs and merging different views into a unified point cloud (see Section 4.2.2 for details).", + "bbox": [ + 75, + 492, + 468, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the distillation process (see Figure 2), we train Teacher model on Object-Complete frames and then freeze it. Then, we train baseline model (playing Student role) to directly minimize Knowledge Distillation losses inspired by [44]. The distillation is done by matching Teacher and Student backbone encoders' embeddings, output labels for bounding box regression, classes distribution for classification task (objects like pedestrians, cars, cyclists, etc.), and intermediate features obtained from the outputs of regression and classification heads before postprocessing (assigning labels). Specifically, we define the following losses:", + "bbox": [ + 75, + 599, + 468, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {h e a d s}} ^ {K D} = \\alpha_ {1} \\mathcal {L} _ {\\text {r e g}} ^ {K D} + \\alpha_ {2} \\mathcal {L} _ {\\text {c l s}} ^ {K D} (1) \\\\ = \\alpha_ {1} D _ {K L} \\left(S _ {\\mathrm {c l s}} \\| T _ {\\mathrm {c l s}}\\right) + \\alpha_ {2} \\mathrm {M S E} \\left(S _ {\\mathrm {r e g}}, T _ {\\mathrm {r e g}}\\right) (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 782, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t}} ^ {K D} = \\operatorname {M S E} \\left(T _ {\\text {b a c k}}, \\phi \\left(\\omega \\left(S _ {\\text {b a c k}}\\right)\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 825, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\det } ^ {K D} = \\mathcal {L} _ {\\text {d e t e c t i o n}} \\left(S _ {\\text {p r e d s}}, \\tilde {T} _ {\\text {b o x e s}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 845, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$T$ and $S$ are the outputs of our Teacher and Student models, respectively. The Student takes the original frame $F$", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as input, while the Teacher receives the Object-Complete Frame $\\tilde{F}$ , so $F \\subset \\tilde{F}$ . $S_{\\text{back}}$ and $T_{\\text{back}}$ are referred to the output of the backbone module and $S_{\\text{reg}}, T_{\\text{reg}}, S_{\\text{cls}}, T_{\\text{cls}}$ are the outputs of regression and classifications heads. $\\tilde{T}_{\\text{boxes}}$ are the X-Ray Teacher's predicted boxes after postprocessing. $S_{\\text{preds}}$ is an overall Student's output; $\\alpha_1, \\alpha_2$ are non-negative hyper-parameters. $L_{\\text{detection}}$ is a basic detection loss that is used for training 3D object detection models. $\\phi$ is a 1x1 Convolution to better match the Teacher's feature maps. $\\omega$ refers to some extra convolutions to make the Student more flexible. We discovered that X-Ray Distillation does not need extra convolutions in the case of the NuScenes dataset, so their usage for encoder feature adjustment is also a hyperparameter of the model. By MSE, we mean Mean Squared Error.", + "bbox": [ + 496, + 340, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the training objective can be written as follows:", + "bbox": [ + 517, + 568, + 885, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {h e a d s}} ^ {K D} + \\lambda_ {2} \\mathcal {L} _ {\\text {f e a t}} ^ {K D} + \\lambda_ {3} \\mathcal {L} _ {\\det } ^ {K D} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 603, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3$ are non-negative hyper-parameters balancing the contribution of each term.", + "bbox": [ + 498, + 628, + 888, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semi-supervised X-Ray Teacher", + "text_level": 1, + "bbox": [ + 500, + 665, + 782, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-supervised learning is characterized by the availability of a small amount of labeled data and a larger pool of unlabeled data, making the use of the Object Complete Frame generation approach proposed for supervised settings infeasible.", + "bbox": [ + 496, + 688, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to overcome this limitation, we introduce the Objects Temporal Fusion block (as shown in Figure 3), which is designed to enable Object Complete Frame generation in situations where ground truth labeling is missing. This block leverages a model pretrained on labeled data to detect and track objects across unlabeled sequences. Subsequently, it employs Point Cloud Registration (PCR) to merge objects points from different views for all detected objects.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2e12ca954dcad7fe353ec5147fffbb5f8d7c5adba8b04d357ea3bd16c206911.jpg", + "image_caption": [ + "Figure 3. Object-Complete Frame generation process for semi-supervised setting. It consists of tracking and Point Cloud Registration. We track objects across all frames in the whole sequence, then we use Point Cloud Registration to merge points that represent the same object from different views, and finally we replace the original object with the new, complete one." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 470, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The provided steps outline a precise and detailed procedure for Objects Temporal Fusion:", + "bbox": [ + 76, + 273, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. to label all LiDAR frames using a pretrained model;", + "2. to greedily track objects across all frames in each sequence using predicted bounding boxes; assign unique IDs to every object instance to facilitate the identification of recurring objects within the scene and to organize them in a sequential manner;", + "3. for each object within each sequence, to merge different views of the same object by applying a deep learning model for Point Cloud Registration; this process generates a complete point cloud for each object, which is then used to replace the occluded one in each frame;", + "4. to fine-tune the base model on Object-Complete labeled frames; this refined model forms our X-Ray Teacher." + ], + "bbox": [ + 76, + 303, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More details for greedy tracking and Point Cloud Registration models can be found in Section 4.2.2.", + "bbox": [ + 76, + 500, + 467, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Knowledge Distillation step presented in Figure 4 can be integrated with any semi-supervised 3D Object Detection approach that uses pseudo labeling as a form of self-distillation. All methods in the domain usually follow this paradigm, which underlines the high universality of our approach. From this perspective, the use of Object-Complete frames for pseudo label prediction refines the 3D bounding boxes and improves the quality of self-distillation. The sole limitation is that existing semi-supervised methods [33, 46] typically update the Teacher model's weights using the exponential moving average of the Student's weights. We refrain from this practice because our Teacher is finetuned on Object-Complete frames and thus predicts higher-quality labels (see proofs in Section 4.3).", + "bbox": [ + 75, + 529, + 468, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section covers the datasets, implementation details, experiment results, and a detailed analysis of the different components that affect performance. It starts with Section 4.1, which defines the datasets and metrics used to prove the validity of our ideas. Section 4.2 delves into the network architectures and training parameters, while Section 4.4 analyzes the obtained metric values. Finally, Section 4.3 examines various parts of our approach and their influence", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on the overall performance.", + "bbox": [ + 498, + 90, + 684, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data", + "text_level": 1, + "bbox": [ + 500, + 113, + 573, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate our models, we use three large-scale autonomous driving datasets: NuScenes and Waymo Open Dataset for the supervised setting and ONCE for the semi-supervised setting.", + "bbox": [ + 496, + 136, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NuScenes [5] is a popular outdoor dataset with diverse annotations for different tasks. It has 40,157 annotated samples. For 3D object detection, we provide NuScenes Detection Score (NDS) and mean Average Precision (mAP).", + "bbox": [ + 496, + 196, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Waymo Open [32] is also one of the most popular outdoor 3D perception datasets. It contains 1150 point cloud sequences and has more than 200K total frames. All results are evaluated with 3D mean Average Precision (mAP) and its weighted variant (mAPH)", + "bbox": [ + 496, + 257, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ONCE [24] is a large-scale dataset with 1 million pout cloud samples from LiDAR and only 15K annotated frames that are divided into train, val, and test with 5k, 3k, and 8k samples, respectively. This dataset is designed exactly for semi-supervised learning tasks and simulates real life: annotations are expensive and time-consuming. All not labeled frames are divided into Small (70 sequences), Medium (321 sequences), and Large (560 sequences) parts. We follow the ONCE Benchmark and use mAP over all classes with the 3D IoU thresholds 0.7, 0.3, and 0.5 for classes \"Vehicle\", \"Pedestrian\", and \"Cyclist\", respectively.", + "bbox": [ + 496, + 333, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 506, + 718, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Network architectures", + "text_level": 1, + "bbox": [ + 500, + 530, + 710, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For supervised setting, we use SECOND [50], CenterPointVoxel [47], CBGS [50], DSVT [36] within our framework. The implementations of these models are based on the OpenPCDet [34] library, and we adhere to the default configurations suggested by this library for both training and inference. In addition, we present scaled versions of these models, which are essentially the original networks augmented with five additional convolutional layers stacked on top of the BEVEncoder with the following parameters 1x Conv(512, 128), 3x Conv(128, 128), and 1x Conv(128, 512) with Batch Normalizations and ReLU activations after each convolution. With the help of light Grid Search, we choose the following distillation hyperparameters: $\\alpha_{1} = 2$ , $\\alpha_{2} = 1$ , $\\lambda_{1} = 0.7$ , $\\lambda_{2} = 0.3$ , $\\lambda_{3} = 1$ .", + "bbox": [ + 496, + 551, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For semi-supervised setting, we follow the previous works [35, 46] and use SECOND and CenterPointVoxel models from OpenPCDet for validation of comparison. For training and inference, we use recently proposed refined configurations from work [11]. X-Ray Teacher model is fine-tuned on object-complete frames with the same hyperparameters for 10 epochs.", + "bbox": [ + 496, + 765, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For computations, we use 4x A100 40GB GPU and AMD EPYC 7702 CPU.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15059", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f6861b1f660e21e5761897ab98ffa2cfcf28b07bdbdf9d363141f1dabdaee63.jpg", + "image_caption": [ + "Figure 4. Semi-supervised X-Ray Teacher pipeline for 3D object detection task. Unlabeled sequences are processed by the Objects Temporal Fusion block to create more complete object representations by aggregating information over time. The Student model learns from both pseudo-labeled predictions and actual labeled data with ground truth annotations." + ], + "image_footnote": [], + "bbox": [ + 140, + 90, + 826, + 324 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cf0c2dd2d6928a6325ccc23e78d03f85a54522546b75f9b709fe055ae4f2fb85.jpg", + "table_caption": [ + "Table 1. Evaluation of the impact of teacher fine-tuning on ObjectComplete frames in semi-supervised setting. The results indicate that teacher fine-tuning is essential." + ], + "table_footnote": [], + "table_body": "
MethodmAP
X-Ray MT SECOND60.12
X-Ray PT SECOND61.48
X-Ray MT SECOND* (ours)65.26
X-Ray PT SECOND* (ours)68.43
X-Ray MT CenterPoint59.35
X-Ray PT CenterPoint62.17
X-Ray MT CenterPoint* (ours)67.60
X-Ray PT CenterPoint* (ours)70.55
", + "bbox": [ + 130, + 455, + 415, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Object-Complete Frame Generation", + "text_level": 1, + "bbox": [ + 76, + 626, + 383, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The key concept of this paper - Object-Complete Frame Generation - combines diverse ideas to achieve optimal Object Completion, encompassing facets such as detection, tracking, and the registration of point clouds.", + "bbox": [ + 75, + 654, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The detection phase is construed as an elective stage, exclusively implemented on unlabeled data. Within this stage discerned instances are encapsulated into generated 3D bounding boxes. E.g., a substantial proportion of instances within the ONCE dataset lacks accompanying labels.", + "bbox": [ + 75, + 717, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, the greedy tracking procedure traverses the entire set of frames, associating the appearance of instances at the i-th frame with potential candidate appearances in the following frame. The list of candidate instances is built by including all instances from the succeeding frame that fall within a prescribed radius, defined", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as twice the maximum dimension across their respective bounding boxes. The nearest instance is selected from the list, while the remaining instances are discarded. The series of matches made using this algorithm are combined into a single sequence, called track. When there are no more potential matches in the next frame, the track is considered as terminated. For the given instance in some specific frame there is the only corresponding track across the entire scene.", + "bbox": [ + 496, + 405, + 890, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Those prepared instances are then used for Object Completion, a process executed through several sequential steps:", + "bbox": [ + 498, + 527, + 890, + 558 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Point clouds associated with instances are extracted from their respective frames, they are translated back to the zero-point of the global basis. The rotation of the instances are also reset to identity.", + "2. The point clouds corresponding to instances within a common track undergo a merging process, constituting the Point Cloud Registration phase, wherein diverse set of merging approaches is used to glue them into a larger, densely populated point cloud.", + "3. Later on the corresponding point clouds are replaced with their respective densely populated point cloud from the previous step, restoring their original translation and rotation in a specific frame." + ], + "bbox": [ + 500, + 559, + 890, + 753 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/da3ded3e077c1aec6d65b3d1488c3da0e4804bbe514f74779043a679a7c71571.jpg", + "table_caption": [ + "Table 2. Comparison of three distillation strategies: using only classification and regression heads matching, BEV features and heads matching, and the final pipeline that is the SECOND-Scaled model; experiment performed using Waymo Validation set." + ], + "table_footnote": [], + "table_body": "
TechniquemAP/mAPH L1mAP/mAPH L2
Heads match67.5/63.461.1/57.2
BEV & Heads67.8/63.561.6/57.4
Full Pipeline (ours)68.3/64.361.9/58.0
", + "bbox": [ + 504, + 837, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15060", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Various merging strategies were empirically tested in generating Object-Complete point clouds:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Geometry: assumes the objects (with boxes) are best aligned by the detected bounding box, performs both inverse translation and rotation to clear their geometric transformations and then merges intact point clouds.", + "2. GeDi: uses GeDi Point Cloud Registration method [27].", + "3. Greedy Grid: uses Greedy Grid implementation [2]." + ], + "bbox": [ + 76, + 125, + 467, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object-Complete frames have the potential to become exceedingly large, so we employ a point subsampling strategy.", + "bbox": [ + 76, + 218, + 467, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/57880af228bbee1933748e7bced994f2ceb99dcc79dbd81be7c750f43f620a65.jpg", + "table_caption": [ + "Table 3. Registration methods comparison for Object Complete Frame Generation on the ONCE Small split. We trained our best SECOND model in the semi-supervised setting with X-Ray Proficient Teacher on three different types of preprocessed data and compared results on ONCE validation set." + ], + "table_footnote": [], + "table_body": "
MethodmAP
Box Geometry67.88
Greedy Grid [2]68.17
GeDi [27]68.43
", + "bbox": [ + 181, + 343, + 366, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e692a4c9b11b0e4783969ca607ba285d69a9825ed2fca86901b70bda2af9de6.jpg", + "table_caption": [ + "Table 4. Comparison for supervised 3D Object Detection task on Waymo Open Dataset. We compare baseline models, their scaled versions and X-Ray distillation with default hyperparameters." + ], + "table_footnote": [], + "table_body": "
ModelmAP/mAPH L1mAP/mAPH L2#params
SECOND X-Ray Teacher*85.1/70.375.1/64.75.3m
SECOND67.2/63.161.0/57.25.3m
X-Ray SECOND67.0/62.860.4/56.75.3m
SECOND-Scaled66.8/62.759.4/56.16.2m
X-Ray SECOND-Scaled68.3/64.361.9/58.06.2m
CenterPoint X-Ray Teacher*88.3/78.676.4/72.98.3m
CenterPoint74.4/71.768.2/65.88.3m
X-Ray CenterPoint73.2/69.767.1/64.58.3m
CenterPoint-Scaled74.1/71.567.9/65.39.2m
X-Ray CenterPoint-Scaled75.2 /72.168.9/66.39.2m
DSVT Pillar X-Ray Teacher*89.3/79.779.1/73.48.6m
DSVT Pillar79.5/77.173.2/71.08.6m
X-Ray DSVT Pillar79.2/76.772.6/70.38.6m
DSVT Pillar-Scaled79.6/77.273.3/71.29.5m
X-Ray DSVT Pillar-Scaled80.1/77.973.7/71.49.5m
", + "bbox": [ + 78, + 491, + 468, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 724, + 240, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide the comparison of different components of our approach and show how they affect the performance. Specifically, we prove the usefulness of teacher fine-tuning and compare Point Cloud Registration and distillation methods. We also analyse if dealing with sparsity and incomplete shapes really improves the performance.", + "bbox": [ + 75, + 750, + 467, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we performed a detailed comparison of the semi-supervised performance of our X-Ray method, both with and without teacher fine-tuning using Object-Complete frames. We use Mean Teacher and Proficient Teacher with", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/300a4c7e097f7720962997b69d717caac67e81ff3c19730ff0aca0b95d13eb4c.jpg", + "image_caption": [ + "Baseline Model Predictions", + "X-Ray Teacher Predictions", + "Figure 5. Visual comparison between noisy and poor baseline (original) 3D detector SECOND (left column) and our X-Ray Teacher that perceives Object-Complete frames. We compare two identical timestamps and view angles. The Baseline model fails to detect some objects while X-Ray Teacher does not. This explains why knowledge distillation is indeed beneficial and should improve models." + ], + "image_footnote": [], + "bbox": [ + 531, + 88, + 862, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SECOND and CenterPointVoxel models trained on ONCE Small split. Table 1 illustrates the substantial impact of teacher fine-tuning on performance, showing that neglecting this step results in a noticeable performance decline.", + "bbox": [ + 496, + 426, + 890, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we noted before, the reconstructed objects will not be perfect in the semi-supervised setting, which is precisely why we included Table 3. PCR models trained on domain objects like cars, pedestrians, cyclists, etc., should improve our method even more.", + "bbox": [ + 496, + 488, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we mentioned earlier, we combine several techniques for the knowledge distillation: BEV features matching, heads output matching with simple regression and KL divergence and detection loss on the teachers predictions. We also compare partial solutions on the Waymo validation set with a SECOND model in Table 2", + "bbox": [ + 496, + 564, + 890, + 652 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/473a7ce838ecd983e78f6b2298686e85e02cc77c3929c654555f0fdcbbc5b500.jpg", + "table_caption": [ + "Table 5. Comparison on NuScenes dataset. Our method improves baselines without scaling models due to the fact that NuScene's object-complete frames are less informative compared to Waymo." + ], + "table_footnote": [], + "table_body": "
ModelmAPNDS
X-Ray Teacher* CBGS77.172.4
CBGS50.059.2
X-Ray CBGS (ours)50.860.4
X-Ray Teacher* CenterPoint-Voxel79.577.1
CenterPoint-Voxel53.461.3
X-Ray CenterPoint-Voxel (ours)54.362.9
X-Ray Teacher* Transfusion-L81.378.2
Transfusion-L56.166.3
X-Ray Transfusion-L (ours)56.866.9
", + "bbox": [ + 514, + 719, + 877, + 882 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f18231aa4da5715a1be7f66cf2b9116c93046494e32ab3d68c7c27dcb9039e39.jpg", + "table_caption": [ + "Table 6. Performance of X-Ray-powered Mean Teacher and Proficient Teacher methods in the semi-supervised setting using SECOND and CenterPoint baselines, ONCE validation set. Our approach consistently outperforms the state-of-the-art for Semi-Supervised 3D Object Detection in terms of mAP across all splits." + ], + "table_footnote": [], + "table_body": "
MethodSECONDCenterPoint
Train (5k labeled samples)
Pretraining63.2264.41
Small (5k labeled + 100k unlabeled samples)
Mean Teacher64.3166.47
X-Ray MT (ours)65.26 (+0.95)67.60 (+1.13)
Proficient Teacher67.0667.72
X-Ray PT (ours)68.43 (+1.37)68.76 (+1.04)
Medium (5k labeled + 500k unlabeled samples)
Mean Teacher64.7366.87
X-Ray MT (ours)65.62 (+0.89)67.78 (+0.91)
Proficient Teacher67.4968.54
X-Ray PT (ours)68.75 (+1.26)69.96 (+1.42)
Large (5k labeled + 1M unlabeled samples)
Mean Teacher65.0367.45
X-Ray MT (ours)65.97 (+0.94)68.17 (+0.72)
Proficient Teacher67.8969.68
X-Ray PT (ours)69.10 (+1.21)70.55 (+0.87)
", + "bbox": [ + 93, + 169, + 454, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate various Point Cloud Registration (PCR) methods used in the Object Complete Frame Generation process. This analysis, detailed in Table 3, is conducted on a ONCE Small split. The results indicate that superior PCR leads to the creation of less noisy objects, which in turn contributes to improved overall quality. However, it's important to note that methods, such as the Greedy Grid method [2] and GeDi [27], are computationally more expensive. This introduces a trade-off between computational efficiency and the quality of the results, highlighting the need for a balanced approach in the selection of PCR methods.", + "bbox": [ + 75, + 476, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Model comparison", + "text_level": 1, + "bbox": [ + 76, + 654, + 256, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1 Supervised Learning", + "text_level": 1, + "bbox": [ + 76, + 678, + 276, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform model comparisons using SECOND [42], CenterPoint-Voxel [47], and DSVT Pillar [36] on the Waymo dataset, and CBGS [50] and CenterPoint-Voxel on the NuScenes dataset. Additionally, we train scaled versions of these models without the X-Ray Teacher to show that the improvements in detection quality are due to the effectiveness of our method, not just an increase in the number of parameters. We scale Waymo students because of the extremely dense and complete point clouds, unlike NuScenes (see supplementary materials), where simpler data requires fewer parameters to learn meaningful feature representations. The results, presented in Tables 4 and 5, demonstrate that our approach consistently outperforms", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "baseline models by 1-2 mAP.", + "bbox": [ + 498, + 90, + 696, + 106 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5cc732965e49a234694564c22908bcb4c9cff2b29253ffeca5864204b236204b.jpg", + "table_caption": [ + "Table 7. Student performance on ONCE validation small split with different tracking methods used in Objects Temporal Fusion" + ], + "table_footnote": [], + "table_body": "
Tracking methodSECONDCenterPoint
Greedy68.4368.76
Kalman Filter + IoU68.5768.91
ReID + Kalman Filter68.7969.12
", + "bbox": [ + 539, + 157, + 857, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2 Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 500, + 250, + 741, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method in the context of pseudo-label-based semi-supervised learning, we perform a comparative analysis with the Mean Teacher [33] and Proficient Teacher [46] methods, which use the SECOND and CenterPointVoxel models. We compare the results obtained with and without the use of the X-Ray Teacher, as detailed in Table 6. Our results show that the application of our approach consistently improves performance, yielding an improvement of $0.8\\mathrm{-}1.4\\mathrm{mAP}$ .", + "bbox": [ + 496, + 273, + 893, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 422, + 619, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our research, we proposed the innovative X-Ray Teacher framework, tailored to improve 3D Object Detection models in supervised and semi-supervised settings. Our extensive results have shown that this approach not only achieves state-of-the-art performance in the semi-supervised setting on the ONCE benchmark, but also consistently improves the quality of supervised models on NuScenes and Waymo Open Dataset. The main contributions of our work include the design of the X-Ray Teacher framework, the development of the Objects Temporal Fusion block for generating Object-Complete frames for data lacking ground truth tracking labels, and the demonstration of the potential of our method to improve the performance of any supervised model trained on sequential data. Future work will focus on optimizing the Objects Temporal Fusion block for more complex environments and exploring the integration of our framework with a broader range of model architectures and applications.", + "bbox": [ + 496, + 448, + 893, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 734, + 687, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported in part through computational resources of HPC facilities at HSE University.", + "bbox": [ + 496, + 758, + 890, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work of I. Makarov on 3D object detection related work was supported by a grant for research centers in the field of artificial intelligence, provided by the Analytical Center in accordance with the subsidy agreement (agreement identifier 000000D730321P5Q0002) and the agreement with the Ivannikov Institute for System Programming of dated November 2, 2021 No. 70-2021-00142.", + "bbox": [ + 496, + 789, + 890, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15062", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018.", + "[2] David Bojanić, Kristijan Bartol, Josep Forest, Stefan Gumhold, Tomislav Petković, and Tomislav Pribanić. Challenging the universal representation of deep models for 3d point cloud registration. In BMVC 2022 Workshop Universal Representations for Computer Vision, 2022.", + "[3] Daniel Brodeski, Igal Bilik, and Raja Giryes. Deep radar detector. In 2019 IEEE Radar Conference (RadarConf), pages 1-6, 2019.", + "[4] Collin Burns et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. arXiv preprint arXiv:2312.09390, 2023.", + "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[6] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M. Alvarez. Focalformer3d: Focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8394-8405, October 2023.", + "[7] Hyeon Cho, Junyong Choi, Geonwoo Baek, and Wonjun Hwang. itkd: Interchange transfer-based knowledge distillation for 3d object detection. In CVPR, 2023.", + "[8] Xu Dong, Pengluo Wang, Pengyue Zhang, and Langechuan Liu. Probabilistic oriented object detection in automotive radar. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 458-467, 2020.", + "[9] J. Fang, D. Zhou, F. Yan, T. Zhao, F. Zhang, Y. Ma, L. Wang, and R. Yang. Augmented lidar simulator for autonomous driving. In IEEE Robotics and Automation, 2020.", + "[10] Jeff Dean Geoffrey Hinton, Oriol Vinyals. Distilling the knowledge in a neural network. In arXiv preprint arXiv:1503.02531, 2015.", + "[11] Maksim Golyadkin, Alexander Gambashidze, Ildar Nurgaliev, and Ilya Makarov. Refining the once benchmark with hyperparameter tuning. IEEE Access, 12:3805-3814, 2024.", + "[12] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017.", + "[13] Chengming Hu, Xuan Li, Dan Liu, Haolun Wu, Xi Chen, Ju Wang, and Xue Liu. Teacher-student architecture for knowledge distillation: A survey. In arXiv preprint arXiv:2210.17332, 2022.", + "[14] Haotian Hu, Fanyi Wang, Jingwen Su, Yaonong Wang, Laifeng Hu, Weiye Fang, Jingwei Xu, and Zhiwang Zhang. Ea-iss: Edge-aware lift-splat-shot framework for 3d bev object detection. arXiv preprint arXiv:2303.17895, 2023.", + "[15] Ilia Indyk and Ilya Makarov. Monovan: Visual attention for self-supervised monocular depth estimation. In 2023 IEEE International Symposium on Mixed and Augmented Reality" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "(ISMAR), pages 1211-1220. IEEE, 2023.", + "[16] Aleksei Karpov and Ilya Makarov. Exploring efficiency of vision transformers for self-supervised monocular depth estimation. In 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 711-719, 2022.", + "[17] Junho Koh, Junhyung Lee, Youngwoo Lee, Jaekyum Kim, and Jun Won Choi. Mgtanet: Encoding sequential lidar points using long short-term motion-guided temporal attention for 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):1179-1187, Jun. 2023.", + "[18] Haoang Li, Jinhu Dong, Binghui Wen, Ming Gao, Tianyu Huang, Yun-Hui Liu, and Daniel Cremers. Ddit: Semantic scene completion via deformable deep implicit templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21894-21904, October 2023.", + "[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[20] Ziyu Li, Yuncong Yao, Zhibin Quan, Jin Xie, and Wankou Yang. Spatial information enhancement network for 3d object detection from point cloud. Pattern Recognition, page 108684, 2022.", + "[21] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023.", + "[22] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 924–932, 2019.", + "[23] Ilya Makarov, Vladimir Aliev, and Olga Gerasimova. Semidense depth interpolation using deep convolutional neural networks. In Proceedings of the 25th ACM International Conference on Multimedia, MM '17, page 1407-1415, New York, NY, USA, 2017. Association for Computing Machinery.", + "[24] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021.", + "[25] Michael Meyer, Georg Kuschk, and Sven Tomforde. Graph convolutional networks for 3d object detection on radar data. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3053-3062, 2021.", + "[26] Mahyar Najibi, Guangda Lai, Abhijit Kundu, Zhichao Lu, Vivek Rathod, Thomas Funkhouser, Caroline Pantofaru, David Ross, Larry S. Davis, and Alireza Fathi. Dops: Learning to detect 3d objects and predict their 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[27] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud regis" + ], + "bbox": [ + 503, + 92, + 892, + 893 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15063", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tration. In IEEE Trans. on Pattern Analysis and Machine Intelligence, (early access) 2022.", + "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017.", + "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017.", + "[30] Ilia Semenkov, Aleksei Karpov, Andrey V Savchenko, and Ilya Makarov. Inpainting semantic and depth features to improve visual place recognition in the wild. IEEE Access, 2024.", + "[31] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020.", + "[32] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017.", + "[34] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.", + "[35] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J. Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2020.", + "[36] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, June 2023.", + "[37] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023.", + "[38] Haiyang Wang, Wenguan Wang, Xizhou Zhu, Jifeng Dai, and Liwei Wang. Collaborative visual navigation. arXiv" + ], + "bbox": [ + 78, + 92, + 468, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "preprint arXiv:2107.01151, 2021.", + "[39] Yan Wang, Bin Yang, Rui Hu, Ming Liang, and Raquel Urtasun. Plumenet: Efficient 3d object detection from stereo images. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3383-3390, 2021.", + "[40] Xinshuo Weng and Kris Kitani. Monocular 3d object detection with pseudo-lidar point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019.", + "[41] Qiangeng Xu, Yiqi Zhong, and Ulrich Neumann. Behind the curtain: Learning occluded shapes for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2893-2901, 2022.", + "[42] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018.", + "[43] Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018.", + "[44] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. In NeurIPS, 2022.", + "[45] Lehan Yang and Kele Xu. Cross modality knowledge distillation for multi-modal aerial view object classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 382-387, June 2021.", + "[46] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022.", + "[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021.", + "[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11087, 2020.", + "[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In European Conference on Computer Vision, pages 496-513. Springer, 2022.", + "[50] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. In arXiv preprint arXiv:1908.09492, 2019.", + "[51] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3357–3364, 2017." + ], + "bbox": [ + 501, + 92, + 893, + 809 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15064", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_model.json b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a499441df4edb05b00c56b39851e7d05c0b88021 --- /dev/null +++ b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.79, + 0.152 + ], + "angle": 0, + "content": "Weak-to-Strong 3D Object Detection with X-Ray Distillation" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.382, + 0.198 + ], + "angle": 0, + "content": "Alexander Gambashidze\\*1,2" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.181, + 0.593, + 0.199 + ], + "angle": 0, + "content": "Aleksandr Dadukin" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.181, + 0.808, + 0.199 + ], + "angle": 0, + "content": "Maxim Golyadkin1,2" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.206, + 0.46, + 0.223 + ], + "angle": 0, + "content": "Maria Razzhivina" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.206, + 0.659, + 0.224 + ], + "angle": 0, + "content": "Ilya Makarov1,3" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.259, + 0.313, + 0.274 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.474, + 0.765 + ], + "angle": 0, + "content": "This paper addresses the critical challenges of sparsity and occlusion in LiDAR-based 3D object detection. Current methods often rely on supplementary modules or specific architectural designs, potentially limiting their applicability to new and evolving architectures. To our knowledge, we are the first to propose a versatile technique that seamlessly integrates into any existing framework for 3D Object Detection, marking the first instance of Weak-to-Strong generalization in 3D computer vision. We introduce a novel framework, X-Ray Distillation with Object-Complete Frames, suitable for both supervised and semi-supervised settings, that leverages the temporal aspect of point cloud sequences. This method extracts crucial information from both previous and subsequent LiDAR frames, creating Object-Complete frames that represent objects from multiple viewpoints, thus addressing occlusion and sparsity. Given the limitation of not being able to generate Object-Complete frames during online inference, we utilize Knowledge Distillation within a Teacher-Student framework. This technique encourages the strong Student model to emulate the behavior of the weaker Teacher, which processes simple and informative Object-Complete frames, effectively offering a comprehensive view of objects as if seen through X-ray vision. Our proposed methods surpass state-of-the-art in semi-supervised learning by 1-1.5 mAP and enhance the performance of five established supervised models by 1-2 mAP on standard autonomous driving datasets, even with default hyperparameters. Code for Object-Complete frames is available here: https://github.com/sakharok13/X-Ray-Teacher-Patching-Tools." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.308, + 0.878, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Figure 1. 3D object detection directly from sparse LiDAR data (top left) provides noisy predictions (bottom left). Adding object completion stage (top right) helps to train 3D object detection X-Ray Teacher, which is robust and can be distilled to baseline model. Red, Yellow and Blue colors of bounding boxes are related to classical LiDAR-based object detection, Our model on Object-Complete frames predictions and Ground Truth labels, respectively." + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.708, + 0.866, + 0.734 + ], + "angle": 0, + "content": "You're just not thinking fourth dimensionally... the bridge will exist." + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.744, + 0.892, + 0.757 + ], + "angle": 0, + "content": "Dr. Emmett Brown, Back to the Future" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.785, + 0.631, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D object detection is a fundamental task in the field of computer vision and autonomous systems, playing a key role in the advancement of self-driving technology [1] and contributing significantly to the robotics industry [38, 51]. Currently, LiDAR-based 3D object detection [6, 17, 36] demonstrates superior performance compared to camera-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.85, + 0.448, + 0.863 + ], + "angle": 0, + "content": "*Corresponding author, email: alexandergambashidze@gmail.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.863, + 0.315, + 0.875 + ], + "angle": 0, + "content": "1Artificial Intelligence Research Institute" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.875, + 0.188, + 0.888 + ], + "angle": 0, + "content": "\\(^{2}\\)HSE University" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.153, + 0.899 + ], + "angle": 0, + "content": "3ISP RAS" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.85, + 0.448, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15055" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.168 + ], + "angle": 0, + "content": "based [19, 39, 40, 45] and radar-based [3, 8, 22, 25] approaches. Furthermore, LiDAR point clouds are the key ingredient of multimodal fusion-based approaches [14, 21, 37], so LiDAR-based 3D detection continues to be a strong focus of the research community." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.473, + 0.669 + ], + "angle": 0, + "content": "The point cloud-based 3D object detection challenges the following issues: sparsity, occlusions, and the complexity of 3D data annotation. Sparsity: large point clouds are sparse due to the LiDAR sensing process's inherent characteristics that leads to an imprecise representation of the captured scene. Additionally, there is an imbalance in point density. In particular, the point cloud is sparser in the far range and contains less spatial information affecting feature representation and box prediction. Occlusions: another problem arises from the frequent occurrence of partial occlusion in LiDAR frames. This primarily happens due to the fact that the frames are acquired from a single fixed point of view, making them essentially 2.5D. To detect and accurately locate a highly occluded object, a detector must recognize the hidden shapes of the object even when a significant portion of its parts is missing. Since the absence of certain shapes inevitably affects object perception, this becomes a critical detection challenge. Our previous approaches focused on depth completion [23], depth inpainting [30] or self-supervised depth pretraining [15, 16] could not address occlusion problems. Data annotation: finally, data annotation is a formidable challenge in 3D object detection due to the complexity of annotating objects in three-dimensional space. For example, a skilled annotator can spend weeks annotating just one hour of LiDAR data [9, 24]. This problem is partially addressed by semi-supervised learning approaches in a teacher-student framework [33, 35, 46, 48]. However, the quality of pseudo labels and the performance of these methods are limited due to the aforementioned sparsity and occlusion problems. Therefore, overcoming these inherent challenges is essential to improve the accuracy and reliability of 3D detection systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The challenges posed by sparsity and occlusion have led to the formulation of several methodologies. Some of them [42] are aimed at improving the computational efficiency of processing sparse data without improving the quality of the predictions. Others, including our work, focus on improving detection performance on sparse occluded data [18, 20, 26, 41]. Li et al. [18] enforce shape constraints to improve object localization with explicit shape priors obtained from a database of CAD models. Najibi et al. [26] and Li et al. [20] do the same by implicitly introducing priors with novel modules pretrained for point cloud completion and SDF approximation. Xu et al. [41] propose a Shape Occupancy Probability estimation module to refine predicted bounding boxes for occluded objects. However, all of these mentioned approaches utilize new modules and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "adapt the model architecture to take advantage of them. As a result, these methods have limited applicability to new architectures that will emerge as the field evolves." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.147, + 0.893, + 0.555 + ], + "angle": 0, + "content": "In this work, we address all three major challenges in 3D object detection with our novel framework called X-Ray Distillation with Object-Complete Frames that is easily plugged in existing approaches and new architectures. It is designed for universal application to any LiDAR-based detector, improving performance on sparse and occluded objects. Our approach exploits the properties of existing large-scale autonomous driving datasets, which consist of sequences of LiDAR frames. Such property makes it possible to reconstruct complete shapes for occluded objects using other occurrences of these objects in the sequence, ensuring that all objects are equipped with points from all available viewpoints in a scene. We then use this completed data in the Teacher-Student framework for both semi-supervised learning and knowledge distillation in a supervised setting. We train our Teacher on extremely informative Object-Complete frames thus making it a weaker model [4]. Then, we use it to extract features from such simple Object-Complete frames and distills this knowledge to a stronger Student, which operates with the original data, to guide him on how to extract rich features from occluded objects. To generate Object-Complete frames, we leverage ground truth object tracking labels. Since there are no labels for the majority of data in the semi-supervised setting, we propose an Objects Temporal Fusion block to detect, track, and use point cloud registration techniques to construct Object-Complete frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.684 + ], + "angle": 0, + "content": "We validate the proposed X-Ray Teacher framework on nuScenes [5] and Waymo Open Dataset [32] 3D object detection benchmarks for SECOND [42], CenterPoint [47], and DSVT [36] models in a supervised learning paradigm. Experiments show a steady improvement by 1-2 mAP with minimal or no impact on time and computational resources during the inference stage. For semi-supervised performance evaluation, we use ONCE [24] benchmark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.723 + ], + "angle": 0, + "content": "Applying these novel ideas for 3D object detection, we provide the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.732, + 0.891, + 0.775 + ], + "angle": 0, + "content": "1. We propose the X-Ray Teacher framework for semi-supervised learning, which achieves state-of-the-art performance on the ONCE benchmark." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.778, + 0.892, + 0.852 + ], + "angle": 0, + "content": "2. We show that our approach improves the quality of four supervised learning models, including the current state-of-the-art model, and demonstrates the potential to improve the performance of any supervised model trained on sequential data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.853, + 0.892, + 0.898 + ], + "angle": 0, + "content": "3. We suggest the Objects Temporal Fusion block to generate Object-Complete frames for data that lacks ground truth tracking labels." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.732, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15056" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.216, + 0.105 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.368, + 0.131 + ], + "angle": 0, + "content": "2.1. Object Detection for point clouds" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.139, + 0.47, + 0.335 + ], + "angle": 0, + "content": "Modern 3D detectors predominantly use point clouds, voxel representations, or a combination of both. The pioneering PointNet model family [28, 29] provided a significant step forward in 3D recognition. However, direct processing of point clouds poses several challenges: it requires point sampling, grouping, and computation of point-wise features, which can be computationally intensive. To integrate insights from 2D computer vision, a transition to a pixel equivalent, namely voxels, is required. Methods such as [41, 43, 47, 49] first convert point clouds to voxels, followed by the use of 3D sparse convolutions [12]. Recent advances also include the integration of modified self-attention layers, achieving state-of-the-art results [36]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.345, + 0.341, + 0.36 + ], + "angle": 0, + "content": "2.2. Semi-supervised 3D detection" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Although Semi-supervised 3D object detection is not as extensively researched as 2D detection, it has seen some significant contributions. Among these, SESS [48] stands out for using the general Mean Teacher approach with data augmentations and consistency loss. Similarly, 3D IoUMatch [35] is notable for its unique localization strategies and 3D IoU-guided techniques for box filtering. Proficient Teachers [46] is notable for its box voting and contrastive losses. We do not propose yet another standalone method for Semi-supervised 3D Object Detection; rather, our emphasis is on creating a plug-in technique designed to augment and enhance the performance of existing methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.56, + 0.29, + 0.574 + ], + "angle": 0, + "content": "2.3. Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.47, + 0.717 + ], + "angle": 0, + "content": "The concept of Knowledge Distillation (KD) was first introduced by Hinton et al. [10]. KD describes a learning approach where a larger teacher network guides the training of a smaller student network for various tasks [13]. Broadly, KD methodologies are categorized into two types: logits/regression distillation and feature map distillation. Our focus is on a hybrid approach, combining these methods: matching feature maps, regression, and classification heads with pseudo labels generated by the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.47, + 0.81 + ], + "angle": 0, + "content": "In 3D object detection, Knowledge Distillation is mainly utilized to minimize parameters and FLOPS while aiming to preserve box prediction quality [7, 44]. Yet, methods that aim to outperform state-of-the-art models are uncommon. In contrast to this trend, our X-Ray Teacher model challenges the norm by providing the quality improvement." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.212, + 0.839 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.327, + 0.864 + ], + "angle": 0, + "content": "3.1. Overview of X-Ray Teacher" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We introduce a novel training framework to address the challenges of sparsity and occlusion in 3D Object Detec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "tion based on LiDAR data. This framework is not limited to any specific object detection model and has the potential for applications across various deep learning architectures. Our method is designed to process the LiDAR data structured as a sequence of frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.892, + 0.244 + ], + "angle": 0, + "content": "The two core elements of our approach are Object Complete Frames Generation and Teacher-Student Knowledge Distillation. Our approach for 3D object detection can be applied in both, supervised and semi-supervised settings, with minor differences in the elements implementation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Object Complete Frames Generation. In this step, we reconstruct the complete shapes of objects presented in the scene by utilizing information from the other frames within the same sequence. Given that autonomous driving datasets are composed of sequential data, we can efficiently leverage their temporal nature: we add points from both the future and the past when objects are observed from different viewpoints. It allows us to reconstruct the complete shapes of objects without shape databases or reconstruction modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.581 + ], + "angle": 0, + "content": "In order to verify the validity of our approach, we trained a CenterPoint [47] model on both, the original and object-complete NuScenes [5] datasets, and then evaluated their performance on the respective validation sets. The models trained on the original and the object-complete frames achieved mAP scores of \\(59.2\\%\\) and \\(79.5\\%\\), respectively. This difference of \\(20~\\mathrm{mAP}\\) suggests that 1) it would be beneficial to transform unlabeled original point clouds into Object-Complete ones and to annotate them with a corresponding X-Ray Teacher pretrained on such informative frames 2) the features extracted by a weaker X-Ray Teacher might be distilled to a stronger student to share the knowledge of complete shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Teacher-Student Knowledge Distillation. The necessity for this step arises because we cannot generate Object Complete Frames during the online inference stage, as it is not possible to access future data. Therefore, we need to encourage the model to behave as if it were observing shape-complete objects, even when dealing with occluded ones. A well-known method for enabling a deep learning model to imitate another model's behavior involves using Knowledge Distillation within a Teacher-Student framework. However, for conventional Knowledge Distillation, both the Teacher and Student models typically process data of the same complexity (the only difference may lie in the complexity of augmentations [31]). In contrast to the standard knowledge distillation, we enrich the data for training Teacher, which significantly improves its performance in 3D object detection. Then, we teach the Student to extract important information from less detailed data by distilling knowledge from the Teacher model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Instead of simplifying the Student model, as it is usually done in standard Knowledge Distillation, we take an opposite approach and design the Student to be more complex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15057" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.852, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overall X-Ray Knowledge Distillation for Supervised Learning. X-Ray Teacher is frozen and pretrained on Object-Complete frames which are taken as input. The Student is guided to mimic the Teacher's behaviour through Knowledge Distillation losses: \\( L_{feat} \\) for intermediate embeddings matching, \\( L_{reg} \\) for bounding box regression, \\( L_{det} \\) for basic detection, and \\( L_{cls} \\) for classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.403 + ], + "angle": 0, + "content": "than the Teacher. It helps to extract high-quality information from more intricate and ambiguous data and demand a Student model to be more robust and to have a more complex receptive field capacity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.464 + ], + "angle": 0, + "content": "In what follows, we provide detailed descriptions of how we implement Object Complete Frames Generation and Teacher-Student Knowledge Distillation steps in both supervised and semi-supervised settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.318, + 0.485 + ], + "angle": 0, + "content": "3.2. Supervised X-Ray Teacher" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.599 + ], + "angle": 0, + "content": "In the supervised setting of 3D Object Detection, models are trained and evaluated using datasets that provide labeled data, including precise bounding boxes and instance IDs. Object-Complete Frame Generation for labeled data involves aggregating objects based on their instance IDs and merging different views into a unified point cloud (see Section 4.2.2 for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.766 + ], + "angle": 0, + "content": "For the distillation process (see Figure 2), we train Teacher model on Object-Complete frames and then freeze it. Then, we train baseline model (playing Student role) to directly minimize Knowledge Distillation losses inspired by [44]. The distillation is done by matching Teacher and Student backbone encoders' embeddings, output labels for bounding box regression, classes distribution for classification task (objects like pedestrians, cars, cyclists, etc.), and intermediate features obtained from the outputs of regression and classification heads before postprocessing (assigning labels). Specifically, we define the following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.784, + 0.47, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {h e a d s}} ^ {K D} = \\alpha_ {1} \\mathcal {L} _ {\\text {r e g}} ^ {K D} + \\alpha_ {2} \\mathcal {L} _ {\\text {c l s}} ^ {K D} (1) \\\\ = \\alpha_ {1} D _ {K L} \\left(S _ {\\mathrm {c l s}} \\| T _ {\\mathrm {c l s}}\\right) + \\alpha_ {2} \\mathrm {M S E} \\left(S _ {\\mathrm {r e g}}, T _ {\\mathrm {r e g}}\\right) (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.826, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f e a t}} ^ {K D} = \\operatorname {M S E} \\left(T _ {\\text {b a c k}}, \\phi \\left(\\omega \\left(S _ {\\text {b a c k}}\\right)\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.847, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\det } ^ {K D} = \\mathcal {L} _ {\\text {d e t e c t i o n}} \\left(S _ {\\text {p r e d s}}, \\tilde {T} _ {\\text {b o x e s}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(T\\) and \\(S\\) are the outputs of our Teacher and Student models, respectively. The Student takes the original frame \\(F\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.568 + ], + "angle": 0, + "content": "as input, while the Teacher receives the Object-Complete Frame \\(\\tilde{F}\\), so \\(F \\subset \\tilde{F}\\). \\(S_{\\text{back}}\\) and \\(T_{\\text{back}}\\) are referred to the output of the backbone module and \\(S_{\\text{reg}}, T_{\\text{reg}}, S_{\\text{cls}}, T_{\\text{cls}}\\) are the outputs of regression and classifications heads. \\(\\tilde{T}_{\\text{boxes}}\\) are the X-Ray Teacher's predicted boxes after postprocessing. \\(S_{\\text{preds}}\\) is an overall Student's output; \\(\\alpha_1, \\alpha_2\\) are non-negative hyper-parameters. \\(L_{\\text{detection}}\\) is a basic detection loss that is used for training 3D object detection models. \\(\\phi\\) is a 1x1 Convolution to better match the Teacher's feature maps. \\(\\omega\\) refers to some extra convolutions to make the Student more flexible. We discovered that X-Ray Distillation does not need extra convolutions in the case of the NuScenes dataset, so their usage for encoder feature adjustment is also a hyperparameter of the model. By MSE, we mean Mean Squared Error." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.569, + 0.886, + 0.584 + ], + "angle": 0, + "content": "Finally, the training objective can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.604, + 0.891, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {h e a d s}} ^ {K D} + \\lambda_ {2} \\mathcal {L} _ {\\text {f e a t}} ^ {K D} + \\lambda_ {3} \\mathcal {L} _ {\\det } ^ {K D} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.89, + 0.659 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3\\) are non-negative hyper-parameters balancing the contribution of each term." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.783, + 0.683 + ], + "angle": 0, + "content": "3.3. Semi-supervised X-Ray Teacher" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.891, + 0.763 + ], + "angle": 0, + "content": "Semi-supervised learning is characterized by the availability of a small amount of labeled data and a larger pool of unlabeled data, making the use of the Object Complete Frame generation approach proposed for supervised settings infeasible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In order to overcome this limitation, we introduce the Objects Temporal Fusion block (as shown in Figure 3), which is designed to enable Object Complete Frame generation in situations where ground truth labeling is missing. This block leverages a model pretrained on labeled data to detect and track objects across unlabeled sequences. Subsequently, it employs Point Cloud Registration (PCR) to merge objects points from different views for all detected objects." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15058" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.164, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Figure 3. Object-Complete Frame generation process for semi-supervised setting. It consists of tracking and Point Cloud Registration. We track objects across all frames in the whole sequence, then we use Point Cloud Registration to merge points that represent the same object from different views, and finally we replace the original object with the new, complete one." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.469, + 0.304 + ], + "angle": 0, + "content": "The provided steps outline a precise and detailed procedure for Objects Temporal Fusion:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.304, + 0.442, + 0.318 + ], + "angle": 0, + "content": "1. to label all LiDAR frames using a pretrained model;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.32, + 0.469, + 0.393 + ], + "angle": 0, + "content": "2. to greedily track objects across all frames in each sequence using predicted bounding boxes; assign unique IDs to every object instance to facilitate the identification of recurring objects within the scene and to organize them in a sequential manner;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.395, + 0.469, + 0.469 + ], + "angle": 0, + "content": "3. for each object within each sequence, to merge different views of the same object by applying a deep learning model for Point Cloud Registration; this process generates a complete point cloud for each object, which is then used to replace the occluded one in each frame;" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.469, + 0.5 + ], + "angle": 0, + "content": "4. to fine-tune the base model on Object-Complete labeled frames; this refined model forms our X-Ray Teacher." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.304, + 0.469, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.501, + 0.468, + 0.529 + ], + "angle": 0, + "content": "More details for greedy tracking and Point Cloud Registration models can be found in Section 4.2.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.742 + ], + "angle": 0, + "content": "The Knowledge Distillation step presented in Figure 4 can be integrated with any semi-supervised 3D Object Detection approach that uses pseudo labeling as a form of self-distillation. All methods in the domain usually follow this paradigm, which underlines the high universality of our approach. From this perspective, the use of Object-Complete frames for pseudo label prediction refines the 3D bounding boxes and improves the quality of self-distillation. The sole limitation is that existing semi-supervised methods [33, 46] typically update the Teacher model's weights using the exponential moving average of the Student's weights. We refrain from this practice because our Teacher is finetuned on Object-Complete frames and thus predicts higher-quality labels (see proofs in Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.21, + 0.772 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This section covers the datasets, implementation details, experiment results, and a detailed analysis of the different components that affect performance. It starts with Section 4.1, which defines the datasets and metrics used to prove the validity of our ideas. Section 4.2 delves into the network architectures and training parameters, while Section 4.4 analyzes the obtained metric values. Finally, Section 4.3 examines various parts of our approach and their influence" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.685, + 0.107 + ], + "angle": 0, + "content": "on the overall performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.114, + 0.575, + 0.128 + ], + "angle": 0, + "content": "4.1. Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.198 + ], + "angle": 0, + "content": "To evaluate our models, we use three large-scale autonomous driving datasets: NuScenes and Waymo Open Dataset for the supervised setting and ONCE for the semi-supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.258 + ], + "angle": 0, + "content": "NuScenes [5] is a popular outdoor dataset with diverse annotations for different tasks. It has 40,157 annotated samples. For 3D object detection, we provide NuScenes Detection Score (NDS) and mean Average Precision (mAP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Waymo Open [32] is also one of the most popular outdoor 3D perception datasets. It contains 1150 point cloud sequences and has more than 200K total frames. All results are evaluated with 3D mean Average Precision (mAP) and its weighted variant (mAPH)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.501 + ], + "angle": 0, + "content": "ONCE [24] is a large-scale dataset with 1 million pout cloud samples from LiDAR and only 15K annotated frames that are divided into train, val, and test with 5k, 3k, and 8k samples, respectively. This dataset is designed exactly for semi-supervised learning tasks and simulates real life: annotations are expensive and time-consuming. All not labeled frames are divided into Small (70 sequences), Medium (321 sequences), and Large (560 sequences) parts. We follow the ONCE Benchmark and use mAP over all classes with the 3D IoU thresholds 0.7, 0.3, and 0.5 for classes \"Vehicle\", \"Pedestrian\", and \"Cyclist\", respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.719, + 0.523 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.712, + 0.544 + ], + "angle": 0, + "content": "4.2.1 Network architectures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.765 + ], + "angle": 0, + "content": "For supervised setting, we use SECOND [50], CenterPointVoxel [47], CBGS [50], DSVT [36] within our framework. The implementations of these models are based on the OpenPCDet [34] library, and we adhere to the default configurations suggested by this library for both training and inference. In addition, we present scaled versions of these models, which are essentially the original networks augmented with five additional convolutional layers stacked on top of the BEVEncoder with the following parameters 1x Conv(512, 128), 3x Conv(128, 128), and 1x Conv(128, 512) with Batch Normalizations and ReLU activations after each convolution. With the help of light Grid Search, we choose the following distillation hyperparameters: \\(\\alpha_{1} = 2\\), \\(\\alpha_{2} = 1\\), \\(\\lambda_{1} = 0.7\\), \\(\\lambda_{2} = 0.3\\), \\(\\lambda_{3} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.871 + ], + "angle": 0, + "content": "For semi-supervised setting, we follow the previous works [35, 46] and use SECOND and CenterPointVoxel models from OpenPCDet for validation of comparison. For training and inference, we use recently proposed refined configurations from work [11]. X-Ray Teacher model is fine-tuned on object-complete frames with the same hyperparameters for 10 epochs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "For computations, we use 4x A100 40GB GPU and AMD EPYC 7702 CPU." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15059" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.092, + 0.827, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Semi-supervised X-Ray Teacher pipeline for 3D object detection task. Unlabeled sequences are processed by the Objects Temporal Fusion block to create more complete object representations by aggregating information over time. The Student model learns from both pseudo-labeled predictions and actual labeled data with ground truth annotations." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.404, + 0.471, + 0.446 + ], + "angle": 0, + "content": "Table 1. Evaluation of the impact of teacher fine-tuning on ObjectComplete frames in semi-supervised setting. The results indicate that teacher fine-tuning is essential." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.457, + 0.416, + 0.598 + ], + "angle": 0, + "content": "
MethodmAP
X-Ray MT SECOND60.12
X-Ray PT SECOND61.48
X-Ray MT SECOND* (ours)65.26
X-Ray PT SECOND* (ours)68.43
X-Ray MT CenterPoint59.35
X-Ray PT CenterPoint62.17
X-Ray MT CenterPoint* (ours)67.60
X-Ray PT CenterPoint* (ours)70.55
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.385, + 0.643 + ], + "angle": 0, + "content": "4.2.2 Object-Complete Frame Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.469, + 0.716 + ], + "angle": 0, + "content": "The key concept of this paper - Object-Complete Frame Generation - combines diverse ideas to achieve optimal Object Completion, encompassing facets such as detection, tracking, and the registration of point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.807 + ], + "angle": 0, + "content": "The detection phase is construed as an elective stage, exclusively implemented on unlabeled data. Within this stage discerned instances are encapsulated into generated 3D bounding boxes. E.g., a substantial proportion of instances within the ONCE dataset lacks accompanying labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Subsequently, the greedy tracking procedure traverses the entire set of frames, associating the appearance of instances at the i-th frame with potential candidate appearances in the following frame. The list of candidate instances is built by including all instances from the succeeding frame that fall within a prescribed radius, defined" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.527 + ], + "angle": 0, + "content": "as twice the maximum dimension across their respective bounding boxes. The nearest instance is selected from the list, while the remaining instances are discarded. The series of matches made using this algorithm are combined into a single sequence, called track. When there are no more potential matches in the next frame, the track is considered as terminated. For the given instance in some specific frame there is the only corresponding track across the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.559 + ], + "angle": 0, + "content": "Those prepared instances are then used for Object Completion, a process executed through several sequential steps:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.56, + 0.892, + 0.619 + ], + "angle": 0, + "content": "1. Point clouds associated with instances are extracted from their respective frames, they are translated back to the zero-point of the global basis. The rotation of the instances are also reset to identity." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.62, + 0.892, + 0.695 + ], + "angle": 0, + "content": "2. The point clouds corresponding to instances within a common track undergo a merging process, constituting the Point Cloud Registration phase, wherein diverse set of merging approaches is used to glue them into a larger, densely populated point cloud." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.695, + 0.892, + 0.755 + ], + "angle": 0, + "content": "3. Later on the corresponding point clouds are replaced with their respective densely populated point cloud from the previous step, restoring their original translation and rotation in a specific frame." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.56, + 0.892, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.771, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Table 2. Comparison of three distillation strategies: using only classification and regression heads matching, BEV features and heads matching, and the final pipeline that is the SECOND-Scaled model; experiment performed using Waymo Validation set." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.838, + 0.892, + 0.899 + ], + "angle": 0, + "content": "
TechniquemAP/mAPH L1mAP/mAPH L2
Heads match67.5/63.461.1/57.2
BEV & Heads67.8/63.561.6/57.4
Full Pipeline (ours)68.3/64.361.9/58.0
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15060" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.469, + 0.122 + ], + "angle": 0, + "content": "Various merging strategies were empirically tested in generating Object-Complete point clouds:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.125, + 0.468, + 0.184 + ], + "angle": 0, + "content": "1. Geometry: assumes the objects (with boxes) are best aligned by the detected bounding box, performs both inverse translation and rotation to clear their geometric transformations and then merges intact point clouds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.185, + 0.468, + 0.2 + ], + "angle": 0, + "content": "2. GeDi: uses GeDi Point Cloud Registration method [27]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.201, + 0.442, + 0.216 + ], + "angle": 0, + "content": "3. Greedy Grid: uses Greedy Grid implementation [2]." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.125, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.468, + 0.251 + ], + "angle": 0, + "content": "Object-Complete frames have the potential to become exceedingly large, so we employ a point subsampling strategy." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.263, + 0.469, + 0.333 + ], + "angle": 0, + "content": "Table 3. Registration methods comparison for Object Complete Frame Generation on the ONCE Small split. We trained our best SECOND model in the semi-supervised setting with X-Ray Proficient Teacher on three different types of preprocessed data and compared results on ONCE validation set." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.344, + 0.367, + 0.41 + ], + "angle": 0, + "content": "
MethodmAP
Box Geometry67.88
Greedy Grid [2]68.17
GeDi [27]68.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.438, + 0.469, + 0.482 + ], + "angle": 0, + "content": "Table 4. Comparison for supervised 3D Object Detection task on Waymo Open Dataset. We compare baseline models, their scaled versions and X-Ray distillation with default hyperparameters." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.699 + ], + "angle": 0, + "content": "
ModelmAP/mAPH L1mAP/mAPH L2#params
SECOND X-Ray Teacher*85.1/70.375.1/64.75.3m
SECOND67.2/63.161.0/57.25.3m
X-Ray SECOND67.0/62.860.4/56.75.3m
SECOND-Scaled66.8/62.759.4/56.16.2m
X-Ray SECOND-Scaled68.3/64.361.9/58.06.2m
CenterPoint X-Ray Teacher*88.3/78.676.4/72.98.3m
CenterPoint74.4/71.768.2/65.88.3m
X-Ray CenterPoint73.2/69.767.1/64.58.3m
CenterPoint-Scaled74.1/71.567.9/65.39.2m
X-Ray CenterPoint-Scaled75.2 /72.168.9/66.39.2m
DSVT Pillar X-Ray Teacher*89.3/79.779.1/73.48.6m
DSVT Pillar79.5/77.173.2/71.08.6m
X-Ray DSVT Pillar79.2/76.772.6/70.38.6m
DSVT Pillar-Scaled79.6/77.273.3/71.29.5m
X-Ray DSVT Pillar-Scaled80.1/77.973.7/71.49.5m
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.241, + 0.74 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.84 + ], + "angle": 0, + "content": "In this section, we provide the comparison of different components of our approach and show how they affect the performance. Specifically, we prove the usefulness of teacher fine-tuning and compare Point Cloud Registration and distillation methods. We also analyse if dealing with sparsity and incomplete shapes really improves the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.899 + ], + "angle": 0, + "content": "First, we performed a detailed comparison of the semi-supervised performance of our X-Ray method, both with and without teacher fine-tuning using Object-Complete frames. We use Mean Teacher and Proficient Teacher with" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.863, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.281, + 0.662, + 0.288 + ], + "angle": 0, + "content": "Baseline Model Predictions" + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.281, + 0.825, + 0.288 + ], + "angle": 0, + "content": "X-Ray Teacher Predictions" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.303, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Figure 5. Visual comparison between noisy and poor baseline (original) 3D detector SECOND (left column) and our X-Ray Teacher that perceives Object-Complete frames. We compare two identical timestamps and view angles. The Baseline model fails to detect some objects while X-Ray Teacher does not. This explains why knowledge distillation is indeed beneficial and should improve models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.488 + ], + "angle": 0, + "content": "SECOND and CenterPointVoxel models trained on ONCE Small split. Table 1 illustrates the substantial impact of teacher fine-tuning on performance, showing that neglecting this step results in a noticeable performance decline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.891, + 0.563 + ], + "angle": 0, + "content": "As we noted before, the reconstructed objects will not be perfect in the semi-supervised setting, which is precisely why we included Table 3. PCR models trained on domain objects like cars, pedestrians, cyclists, etc., should improve our method even more." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.654 + ], + "angle": 0, + "content": "As we mentioned earlier, we combine several techniques for the knowledge distillation: BEV features matching, heads output matching with simple regression and KL divergence and detection loss on the teachers predictions. We also compare partial solutions on the Waymo validation set with a SECOND model in Table 2" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Table 5. Comparison on NuScenes dataset. Our method improves baselines without scaling models due to the fact that NuScene's object-complete frames are less informative compared to Waymo." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.72, + 0.878, + 0.883 + ], + "angle": 0, + "content": "
ModelmAPNDS
X-Ray Teacher* CBGS77.172.4
CBGS50.059.2
X-Ray CBGS (ours)50.860.4
X-Ray Teacher* CenterPoint-Voxel79.577.1
CenterPoint-Voxel53.461.3
X-Ray CenterPoint-Voxel (ours)54.362.9
X-Ray Teacher* Transfusion-L81.378.2
Transfusion-L56.166.3
X-Ray Transfusion-L (ours)56.866.9
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15061" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.16 + ], + "angle": 0, + "content": "Table 6. Performance of X-Ray-powered Mean Teacher and Proficient Teacher methods in the semi-supervised setting using SECOND and CenterPoint baselines, ONCE validation set. Our approach consistently outperforms the state-of-the-art for Semi-Supervised 3D Object Detection in terms of mAP across all splits." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.17, + 0.455, + 0.451 + ], + "angle": 0, + "content": "
MethodSECONDCenterPoint
Train (5k labeled samples)
Pretraining63.2264.41
Small (5k labeled + 100k unlabeled samples)
Mean Teacher64.3166.47
X-Ray MT (ours)65.26 (+0.95)67.60 (+1.13)
Proficient Teacher67.0667.72
X-Ray PT (ours)68.43 (+1.37)68.76 (+1.04)
Medium (5k labeled + 500k unlabeled samples)
Mean Teacher64.7366.87
X-Ray MT (ours)65.62 (+0.89)67.78 (+0.91)
Proficient Teacher67.4968.54
X-Ray PT (ours)68.75 (+1.26)69.96 (+1.42)
Large (5k labeled + 1M unlabeled samples)
Mean Teacher65.0367.45
X-Ray MT (ours)65.97 (+0.94)68.17 (+0.72)
Proficient Teacher67.8969.68
X-Ray PT (ours)69.10 (+1.21)70.55 (+0.87)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.645 + ], + "angle": 0, + "content": "We evaluate various Point Cloud Registration (PCR) methods used in the Object Complete Frame Generation process. This analysis, detailed in Table 3, is conducted on a ONCE Small split. The results indicate that superior PCR leads to the creation of less noisy objects, which in turn contributes to improved overall quality. However, it's important to note that methods, such as the Greedy Grid method [2] and GeDi [27], are computationally more expensive. This introduces a trade-off between computational efficiency and the quality of the results, highlighting the need for a balanced approach in the selection of PCR methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.258, + 0.672 + ], + "angle": 0, + "content": "4.4. Model comparison" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.277, + 0.696 + ], + "angle": 0, + "content": "4.4.1 Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We perform model comparisons using SECOND [42], CenterPoint-Voxel [47], and DSVT Pillar [36] on the Waymo dataset, and CBGS [50] and CenterPoint-Voxel on the NuScenes dataset. Additionally, we train scaled versions of these models without the X-Ray Teacher to show that the improvements in detection quality are due to the effectiveness of our method, not just an increase in the number of parameters. We scale Waymo students because of the extremely dense and complete point clouds, unlike NuScenes (see supplementary materials), where simpler data requires fewer parameters to learn meaningful feature representations. The results, presented in Tables 4 and 5, demonstrate that our approach consistently outperforms" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.697, + 0.107 + ], + "angle": 0, + "content": "baseline models by 1-2 mAP." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.119, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Table 7. Student performance on ONCE validation small split with different tracking methods used in Objects Temporal Fusion" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.159, + 0.858, + 0.219 + ], + "angle": 0, + "content": "
Tracking methodSECONDCenterPoint
Greedy68.4368.76
Kalman Filter + IoU68.5768.91
ReID + Kalman Filter68.7969.12
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.251, + 0.742, + 0.267 + ], + "angle": 0, + "content": "4.4.2 Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.895, + 0.412 + ], + "angle": 0, + "content": "To validate the effectiveness of our method in the context of pseudo-label-based semi-supervised learning, we perform a comparative analysis with the Mean Teacher [33] and Proficient Teacher [46] methods, which use the SECOND and CenterPointVoxel models. We compare the results obtained with and without the use of the X-Ray Teacher, as detailed in Table 6. Our results show that the application of our approach consistently improves performance, yielding an improvement of \\(0.8\\mathrm{-}1.4\\mathrm{mAP}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.424, + 0.62, + 0.44 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.895, + 0.723 + ], + "angle": 0, + "content": "In our research, we proposed the innovative X-Ray Teacher framework, tailored to improve 3D Object Detection models in supervised and semi-supervised settings. Our extensive results have shown that this approach not only achieves state-of-the-art performance in the semi-supervised setting on the ONCE benchmark, but also consistently improves the quality of supervised models on NuScenes and Waymo Open Dataset. The main contributions of our work include the design of the X-Ray Teacher framework, the development of the Objects Temporal Fusion block for generating Object-Complete frames for data lacking ground truth tracking labels, and the demonstration of the potential of our method to improve the performance of any supervised model trained on sequential data. Future work will focus on optimizing the Objects Temporal Fusion block for more complex environments and exploring the integration of our framework with a broader range of model architectures and applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.688, + 0.753 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.79 + ], + "angle": 0, + "content": "This research was supported in part through computational resources of HPC facilities at HSE University." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.896 + ], + "angle": 0, + "content": "The work of I. Makarov on 3D object detection related work was supported by a grant for research centers in the field of artificial intelligence, provided by the Analytical Center in accordance with the subsidy agreement (agreement identifier 000000D730321P5Q0002) and the agreement with the Ivannikov Institute for System Programming of dated November 2, 2021 No. 70-2021-00142." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15062" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.47, + 0.238 + ], + "angle": 0, + "content": "[2] David Bojanić, Kristijan Bartol, Josep Forest, Stefan Gumhold, Tomislav Petković, and Tomislav Pribanić. Challenging the universal representation of deep models for 3d point cloud registration. In BMVC 2022 Workshop Universal Representations for Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.239, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[3] Daniel Brodeski, Igal Bilik, and Raja Giryes. Deep radar detector. In 2019 IEEE Radar Conference (RadarConf), pages 1-6, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[4] Collin Burns et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. arXiv preprint arXiv:2312.09390, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[6] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M. Alvarez. Focalformer3d: Focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8394-8405, October 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Hyeon Cho, Junyong Choi, Geonwoo Baek, and Wonjun Hwang. itkd: Interchange transfer-based knowledge distillation for 3d object detection. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[8] Xu Dong, Pengluo Wang, Pengyue Zhang, and Langechuan Liu. Probabilistic oriented object detection in automotive radar. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 458-467, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[9] J. Fang, D. Zhou, F. Yan, T. Zhao, F. Zhang, Y. Ma, L. Wang, and R. Yang. Augmented lidar simulator for autonomous driving. In IEEE Robotics and Automation, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.631, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[10] Jeff Dean Geoffrey Hinton, Oriol Vinyals. Distilling the knowledge in a neural network. In arXiv preprint arXiv:1503.02531, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.671, + 0.47, + 0.711 + ], + "angle": 0, + "content": "[11] Maksim Golyadkin, Alexander Gambashidze, Ildar Nurgaliev, and Ilya Makarov. Refining the once benchmark with hyperparameter tuning. IEEE Access, 12:3805-3814, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.712, + 0.47, + 0.75 + ], + "angle": 0, + "content": "[12] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.751, + 0.47, + 0.805 + ], + "angle": 0, + "content": "[13] Chengming Hu, Xuan Li, Dan Liu, Haolun Wu, Xi Chen, Ju Wang, and Xue Liu. Teacher-student architecture for knowledge distillation: A survey. In arXiv preprint arXiv:2210.17332, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.806, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[14] Haotian Hu, Fanyi Wang, Jingwen Su, Yaonong Wang, Laifeng Hu, Weiye Fang, Jingwei Xu, and Zhiwang Zhang. Ea-iss: Edge-aware lift-splat-shot framework for 3d bev object detection. arXiv preprint arXiv:2303.17895, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Ilia Indyk and Ilya Makarov. Monovan: Visual attention for self-supervised monocular depth estimation. In 2023 IEEE International Symposium on Mixed and Augmented Reality" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.78, + 0.106 + ], + "angle": 0, + "content": "(ISMAR), pages 1211-1220. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[16] Aleksei Karpov and Ilya Makarov. Exploring efficiency of vision transformers for self-supervised monocular depth estimation. In 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 711-719, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.16, + 0.893, + 0.228 + ], + "angle": 0, + "content": "[17] Junho Koh, Junhyung Lee, Youngwoo Lee, Jaekyum Kim, and Jun Won Choi. Mgtanet: Encoding sequential lidar points using long short-term motion-guided temporal attention for 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):1179-1187, Jun. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.228, + 0.893, + 0.308 + ], + "angle": 0, + "content": "[18] Haoang Li, Jinhu Dong, Binghui Wen, Ming Gao, Tianyu Huang, Yun-Hui Liu, and Daniel Cremers. Ddit: Semantic scene completion via deformable deep implicit templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21894-21904, October 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.893, + 0.364 + ], + "angle": 0, + "content": "[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.365, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[20] Ziyu Li, Yuncong Yao, Zhibin Quan, Jin Xie, and Wankou Yang. Spatial information enhancement network for 3d object detection from point cloud. Pattern Recognition, page 108684, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.893, + 0.486 + ], + "angle": 0, + "content": "[21] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.486, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[22] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 924–932, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[23] Ilya Makarov, Vladimir Aliev, and Olga Gerasimova. Semidense depth interpolation using deep convolutional neural networks. In Proceedings of the 25th ACM International Conference on Multimedia, MM '17, page 1407-1415, New York, NY, USA, 2017. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.663, + 0.893, + 0.731 + ], + "angle": 0, + "content": "[24] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.893, + 0.786 + ], + "angle": 0, + "content": "[25] Michael Meyer, Georg Kuschk, and Sven Tomforde. Graph convolutional networks for 3d object detection on radar data. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3053-3062, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.893, + 0.867 + ], + "angle": 0, + "content": "[26] Mahyar Najibi, Guangda Lai, Abhijit Kundu, Zhichao Lu, Vivek Rathod, Thomas Funkhouser, Caroline Pantofaru, David Ross, Larry S. Davis, and Alireza Fathi. Dops: Learning to detect 3d objects and predict their 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.868, + 0.893, + 0.895 + ], + "angle": 0, + "content": "[27] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud regis" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.893, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15063" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "tration. In IEEE Trans. on Pattern Analysis and Machine Intelligence, (early access) 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.187, + 0.469, + 0.242 + ], + "angle": 0, + "content": "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[30] Ilia Semenkov, Aleksei Karpov, Andrey V Savchenko, and Ilya Makarov. Inpainting semantic and depth features to improve visual place recognition in the wild. IEEE Access, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[31] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[32] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.543, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.639, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[34] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[35] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J. Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[36] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.469, + 0.868 + ], + "angle": 0, + "content": "[37] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.869, + 0.469, + 0.895 + ], + "angle": 0, + "content": "[38] Haiyang Wang, Wenguan Wang, Xizhou Zhu, Jifeng Dai, and Liwei Wang. Collaborative visual navigation. arXiv" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.093, + 0.736, + 0.106 + ], + "angle": 0, + "content": "preprint arXiv:2107.01151, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[39] Yan Wang, Bin Yang, Rui Hu, Ming Liang, and Raquel Urtasun. Plumenet: Efficient 3d object detection from stereo images. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3383-3390, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.16, + 0.894, + 0.214 + ], + "angle": 0, + "content": "[40] Xinshuo Weng and Kris Kitani. Monocular 3d object detection with pseudo-lidar point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.892, + 0.269 + ], + "angle": 0, + "content": "[41] Qiangeng Xu, Yiqi Zhong, and Ulrich Neumann. Behind the curtain: Learning occluded shapes for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2893-2901, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.269, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[42] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[43] Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[44] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.389, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[45] Lehan Yang and Kele Xu. Cross modality knowledge distillation for multi-modal aerial view object classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 382-387, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[46] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.579, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11087, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In European Conference on Computer Vision, pages 496-513. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[50] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. In arXiv preprint arXiv:1908.09492, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.892, + 0.81 + ], + "angle": 0, + "content": "[51] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3357–3364, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15064" + } + ] +] \ No newline at end of file diff --git a/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_origin.pdf b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d64368509a67d55e194036e2c941d517b7e706af --- /dev/null +++ b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/ecb5d87e-b1e5-43e0-b246-7e30c3a3bbcd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d542e58b9960359e73d25defd08e0ad287c8446be5d2396b5182bb00f23951c4 +size 1016105 diff --git a/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/full.md b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..746439264f5ec6c4c9851510f8c6206a53e6dc2d --- /dev/null +++ b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/full.md @@ -0,0 +1,317 @@ +# Weak-to-Strong 3D Object Detection with X-Ray Distillation + +Alexander Gambashidze\*1,2 + +Aleksandr Dadukin + +Maxim Golyadkin1,2 + +Maria Razzhivina + +Ilya Makarov1,3 + +# Abstract + +This paper addresses the critical challenges of sparsity and occlusion in LiDAR-based 3D object detection. Current methods often rely on supplementary modules or specific architectural designs, potentially limiting their applicability to new and evolving architectures. To our knowledge, we are the first to propose a versatile technique that seamlessly integrates into any existing framework for 3D Object Detection, marking the first instance of Weak-to-Strong generalization in 3D computer vision. We introduce a novel framework, X-Ray Distillation with Object-Complete Frames, suitable for both supervised and semi-supervised settings, that leverages the temporal aspect of point cloud sequences. This method extracts crucial information from both previous and subsequent LiDAR frames, creating Object-Complete frames that represent objects from multiple viewpoints, thus addressing occlusion and sparsity. Given the limitation of not being able to generate Object-Complete frames during online inference, we utilize Knowledge Distillation within a Teacher-Student framework. This technique encourages the strong Student model to emulate the behavior of the weaker Teacher, which processes simple and informative Object-Complete frames, effectively offering a comprehensive view of objects as if seen through X-ray vision. Our proposed methods surpass state-of-the-art in semi-supervised learning by 1-1.5 mAP and enhance the performance of five established supervised models by 1-2 mAP on standard autonomous driving datasets, even with default hyperparameters. Code for Object-Complete frames is available here: https://github.com/sakharok13/X-Ray-Teacher-Patching-Tools. + +![](images/00632d629e00b25464ef5870158fe326a1d02ef4d2b4313bdcc4341c9666ed51.jpg) +Figure 1. 3D object detection directly from sparse LiDAR data (top left) provides noisy predictions (bottom left). Adding object completion stage (top right) helps to train 3D object detection X-Ray Teacher, which is robust and can be distilled to baseline model. Red, Yellow and Blue colors of bounding boxes are related to classical LiDAR-based object detection, Our model on Object-Complete frames predictions and Ground Truth labels, respectively. + +You're just not thinking fourth dimensionally... the bridge will exist. + +Dr. Emmett Brown, Back to the Future + +# 1. Introduction + +3D object detection is a fundamental task in the field of computer vision and autonomous systems, playing a key role in the advancement of self-driving technology [1] and contributing significantly to the robotics industry [38, 51]. Currently, LiDAR-based 3D object detection [6, 17, 36] demonstrates superior performance compared to camera- + +based [19, 39, 40, 45] and radar-based [3, 8, 22, 25] approaches. Furthermore, LiDAR point clouds are the key ingredient of multimodal fusion-based approaches [14, 21, 37], so LiDAR-based 3D detection continues to be a strong focus of the research community. + +The point cloud-based 3D object detection challenges the following issues: sparsity, occlusions, and the complexity of 3D data annotation. Sparsity: large point clouds are sparse due to the LiDAR sensing process's inherent characteristics that leads to an imprecise representation of the captured scene. Additionally, there is an imbalance in point density. In particular, the point cloud is sparser in the far range and contains less spatial information affecting feature representation and box prediction. Occlusions: another problem arises from the frequent occurrence of partial occlusion in LiDAR frames. This primarily happens due to the fact that the frames are acquired from a single fixed point of view, making them essentially 2.5D. To detect and accurately locate a highly occluded object, a detector must recognize the hidden shapes of the object even when a significant portion of its parts is missing. Since the absence of certain shapes inevitably affects object perception, this becomes a critical detection challenge. Our previous approaches focused on depth completion [23], depth inpainting [30] or self-supervised depth pretraining [15, 16] could not address occlusion problems. Data annotation: finally, data annotation is a formidable challenge in 3D object detection due to the complexity of annotating objects in three-dimensional space. For example, a skilled annotator can spend weeks annotating just one hour of LiDAR data [9, 24]. This problem is partially addressed by semi-supervised learning approaches in a teacher-student framework [33, 35, 46, 48]. However, the quality of pseudo labels and the performance of these methods are limited due to the aforementioned sparsity and occlusion problems. Therefore, overcoming these inherent challenges is essential to improve the accuracy and reliability of 3D detection systems. + +The challenges posed by sparsity and occlusion have led to the formulation of several methodologies. Some of them [42] are aimed at improving the computational efficiency of processing sparse data without improving the quality of the predictions. Others, including our work, focus on improving detection performance on sparse occluded data [18, 20, 26, 41]. Li et al. [18] enforce shape constraints to improve object localization with explicit shape priors obtained from a database of CAD models. Najibi et al. [26] and Li et al. [20] do the same by implicitly introducing priors with novel modules pretrained for point cloud completion and SDF approximation. Xu et al. [41] propose a Shape Occupancy Probability estimation module to refine predicted bounding boxes for occluded objects. However, all of these mentioned approaches utilize new modules and + +adapt the model architecture to take advantage of them. As a result, these methods have limited applicability to new architectures that will emerge as the field evolves. + +In this work, we address all three major challenges in 3D object detection with our novel framework called X-Ray Distillation with Object-Complete Frames that is easily plugged in existing approaches and new architectures. It is designed for universal application to any LiDAR-based detector, improving performance on sparse and occluded objects. Our approach exploits the properties of existing large-scale autonomous driving datasets, which consist of sequences of LiDAR frames. Such property makes it possible to reconstruct complete shapes for occluded objects using other occurrences of these objects in the sequence, ensuring that all objects are equipped with points from all available viewpoints in a scene. We then use this completed data in the Teacher-Student framework for both semi-supervised learning and knowledge distillation in a supervised setting. We train our Teacher on extremely informative Object-Complete frames thus making it a weaker model [4]. Then, we use it to extract features from such simple Object-Complete frames and distills this knowledge to a stronger Student, which operates with the original data, to guide him on how to extract rich features from occluded objects. To generate Object-Complete frames, we leverage ground truth object tracking labels. Since there are no labels for the majority of data in the semi-supervised setting, we propose an Objects Temporal Fusion block to detect, track, and use point cloud registration techniques to construct Object-Complete frames. + +We validate the proposed X-Ray Teacher framework on nuScenes [5] and Waymo Open Dataset [32] 3D object detection benchmarks for SECOND [42], CenterPoint [47], and DSVT [36] models in a supervised learning paradigm. Experiments show a steady improvement by 1-2 mAP with minimal or no impact on time and computational resources during the inference stage. For semi-supervised performance evaluation, we use ONCE [24] benchmark. + +Applying these novel ideas for 3D object detection, we provide the following contributions: + +1. We propose the X-Ray Teacher framework for semi-supervised learning, which achieves state-of-the-art performance on the ONCE benchmark. +2. We show that our approach improves the quality of four supervised learning models, including the current state-of-the-art model, and demonstrates the potential to improve the performance of any supervised model trained on sequential data. +3. We suggest the Objects Temporal Fusion block to generate Object-Complete frames for data that lacks ground truth tracking labels. + +# 2. Related work + +# 2.1. Object Detection for point clouds + +Modern 3D detectors predominantly use point clouds, voxel representations, or a combination of both. The pioneering PointNet model family [28, 29] provided a significant step forward in 3D recognition. However, direct processing of point clouds poses several challenges: it requires point sampling, grouping, and computation of point-wise features, which can be computationally intensive. To integrate insights from 2D computer vision, a transition to a pixel equivalent, namely voxels, is required. Methods such as [41, 43, 47, 49] first convert point clouds to voxels, followed by the use of 3D sparse convolutions [12]. Recent advances also include the integration of modified self-attention layers, achieving state-of-the-art results [36]. + +# 2.2. Semi-supervised 3D detection + +Although Semi-supervised 3D object detection is not as extensively researched as 2D detection, it has seen some significant contributions. Among these, SESS [48] stands out for using the general Mean Teacher approach with data augmentations and consistency loss. Similarly, 3D IoUMatch [35] is notable for its unique localization strategies and 3D IoU-guided techniques for box filtering. Proficient Teachers [46] is notable for its box voting and contrastive losses. We do not propose yet another standalone method for Semi-supervised 3D Object Detection; rather, our emphasis is on creating a plug-in technique designed to augment and enhance the performance of existing methods. + +# 2.3. Knowledge Distillation + +The concept of Knowledge Distillation (KD) was first introduced by Hinton et al. [10]. KD describes a learning approach where a larger teacher network guides the training of a smaller student network for various tasks [13]. Broadly, KD methodologies are categorized into two types: logits/regression distillation and feature map distillation. Our focus is on a hybrid approach, combining these methods: matching feature maps, regression, and classification heads with pseudo labels generated by the teacher model. + +In 3D object detection, Knowledge Distillation is mainly utilized to minimize parameters and FLOPS while aiming to preserve box prediction quality [7, 44]. Yet, methods that aim to outperform state-of-the-art models are uncommon. In contrast to this trend, our X-Ray Teacher model challenges the norm by providing the quality improvement. + +# 3. Methodology + +# 3.1. Overview of X-Ray Teacher + +We introduce a novel training framework to address the challenges of sparsity and occlusion in 3D Object Detec + +tion based on LiDAR data. This framework is not limited to any specific object detection model and has the potential for applications across various deep learning architectures. Our method is designed to process the LiDAR data structured as a sequence of frames. + +The two core elements of our approach are Object Complete Frames Generation and Teacher-Student Knowledge Distillation. Our approach for 3D object detection can be applied in both, supervised and semi-supervised settings, with minor differences in the elements implementation. + +Object Complete Frames Generation. In this step, we reconstruct the complete shapes of objects presented in the scene by utilizing information from the other frames within the same sequence. Given that autonomous driving datasets are composed of sequential data, we can efficiently leverage their temporal nature: we add points from both the future and the past when objects are observed from different viewpoints. It allows us to reconstruct the complete shapes of objects without shape databases or reconstruction modules. + +In order to verify the validity of our approach, we trained a CenterPoint [47] model on both, the original and object-complete NuScenes [5] datasets, and then evaluated their performance on the respective validation sets. The models trained on the original and the object-complete frames achieved mAP scores of $59.2\%$ and $79.5\%$ , respectively. This difference of $20~\mathrm{mAP}$ suggests that 1) it would be beneficial to transform unlabeled original point clouds into Object-Complete ones and to annotate them with a corresponding X-Ray Teacher pretrained on such informative frames 2) the features extracted by a weaker X-Ray Teacher might be distilled to a stronger student to share the knowledge of complete shapes. + +Teacher-Student Knowledge Distillation. The necessity for this step arises because we cannot generate Object Complete Frames during the online inference stage, as it is not possible to access future data. Therefore, we need to encourage the model to behave as if it were observing shape-complete objects, even when dealing with occluded ones. A well-known method for enabling a deep learning model to imitate another model's behavior involves using Knowledge Distillation within a Teacher-Student framework. However, for conventional Knowledge Distillation, both the Teacher and Student models typically process data of the same complexity (the only difference may lie in the complexity of augmentations [31]). In contrast to the standard knowledge distillation, we enrich the data for training Teacher, which significantly improves its performance in 3D object detection. Then, we teach the Student to extract important information from less detailed data by distilling knowledge from the Teacher model. + +Instead of simplifying the Student model, as it is usually done in standard Knowledge Distillation, we take an opposite approach and design the Student to be more complex + +![](images/c60898a961fecfaffa9e7f9afbd0deff0743ee49aa7e95fa41691618e859d675.jpg) +Figure 2. Overall X-Ray Knowledge Distillation for Supervised Learning. X-Ray Teacher is frozen and pretrained on Object-Complete frames which are taken as input. The Student is guided to mimic the Teacher's behaviour through Knowledge Distillation losses: $L_{feat}$ for intermediate embeddings matching, $L_{reg}$ for bounding box regression, $L_{det}$ for basic detection, and $L_{cls}$ for classification. + +than the Teacher. It helps to extract high-quality information from more intricate and ambiguous data and demand a Student model to be more robust and to have a more complex receptive field capacity. + +In what follows, we provide detailed descriptions of how we implement Object Complete Frames Generation and Teacher-Student Knowledge Distillation steps in both supervised and semi-supervised settings. + +# 3.2. Supervised X-Ray Teacher + +In the supervised setting of 3D Object Detection, models are trained and evaluated using datasets that provide labeled data, including precise bounding boxes and instance IDs. Object-Complete Frame Generation for labeled data involves aggregating objects based on their instance IDs and merging different views into a unified point cloud (see Section 4.2.2 for details). + +For the distillation process (see Figure 2), we train Teacher model on Object-Complete frames and then freeze it. Then, we train baseline model (playing Student role) to directly minimize Knowledge Distillation losses inspired by [44]. The distillation is done by matching Teacher and Student backbone encoders' embeddings, output labels for bounding box regression, classes distribution for classification task (objects like pedestrians, cars, cyclists, etc.), and intermediate features obtained from the outputs of regression and classification heads before postprocessing (assigning labels). Specifically, we define the following losses: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {h e a d s}} ^ {K D} = \alpha_ {1} \mathcal {L} _ {\text {r e g}} ^ {K D} + \alpha_ {2} \mathcal {L} _ {\text {c l s}} ^ {K D} (1) \\ = \alpha_ {1} D _ {K L} \left(S _ {\mathrm {c l s}} \| T _ {\mathrm {c l s}}\right) + \alpha_ {2} \mathrm {M S E} \left(S _ {\mathrm {r e g}}, T _ {\mathrm {r e g}}\right) (2) \\ \end{array} +$$ + +$$ +\mathcal {L} _ {\text {f e a t}} ^ {K D} = \operatorname {M S E} \left(T _ {\text {b a c k}}, \phi \left(\omega \left(S _ {\text {b a c k}}\right)\right)\right) \tag {3} +$$ + +$$ +\mathcal {L} _ {\det } ^ {K D} = \mathcal {L} _ {\text {d e t e c t i o n}} \left(S _ {\text {p r e d s}}, \tilde {T} _ {\text {b o x e s}}\right) \tag {4} +$$ + +$T$ and $S$ are the outputs of our Teacher and Student models, respectively. The Student takes the original frame $F$ + +as input, while the Teacher receives the Object-Complete Frame $\tilde{F}$ , so $F \subset \tilde{F}$ . $S_{\text{back}}$ and $T_{\text{back}}$ are referred to the output of the backbone module and $S_{\text{reg}}, T_{\text{reg}}, S_{\text{cls}}, T_{\text{cls}}$ are the outputs of regression and classifications heads. $\tilde{T}_{\text{boxes}}$ are the X-Ray Teacher's predicted boxes after postprocessing. $S_{\text{preds}}$ is an overall Student's output; $\alpha_1, \alpha_2$ are non-negative hyper-parameters. $L_{\text{detection}}$ is a basic detection loss that is used for training 3D object detection models. $\phi$ is a 1x1 Convolution to better match the Teacher's feature maps. $\omega$ refers to some extra convolutions to make the Student more flexible. We discovered that X-Ray Distillation does not need extra convolutions in the case of the NuScenes dataset, so their usage for encoder feature adjustment is also a hyperparameter of the model. By MSE, we mean Mean Squared Error. + +Finally, the training objective can be written as follows: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\text {h e a d s}} ^ {K D} + \lambda_ {2} \mathcal {L} _ {\text {f e a t}} ^ {K D} + \lambda_ {3} \mathcal {L} _ {\det } ^ {K D} \tag {5} +$$ + +where $\lambda_1, \lambda_2, \lambda_3$ are non-negative hyper-parameters balancing the contribution of each term. + +# 3.3. Semi-supervised X-Ray Teacher + +Semi-supervised learning is characterized by the availability of a small amount of labeled data and a larger pool of unlabeled data, making the use of the Object Complete Frame generation approach proposed for supervised settings infeasible. + +In order to overcome this limitation, we introduce the Objects Temporal Fusion block (as shown in Figure 3), which is designed to enable Object Complete Frame generation in situations where ground truth labeling is missing. This block leverages a model pretrained on labeled data to detect and track objects across unlabeled sequences. Subsequently, it employs Point Cloud Registration (PCR) to merge objects points from different views for all detected objects. + +![](images/b2e12ca954dcad7fe353ec5147fffbb5f8d7c5adba8b04d357ea3bd16c206911.jpg) +Figure 3. Object-Complete Frame generation process for semi-supervised setting. It consists of tracking and Point Cloud Registration. We track objects across all frames in the whole sequence, then we use Point Cloud Registration to merge points that represent the same object from different views, and finally we replace the original object with the new, complete one. + +The provided steps outline a precise and detailed procedure for Objects Temporal Fusion: + +1. to label all LiDAR frames using a pretrained model; +2. to greedily track objects across all frames in each sequence using predicted bounding boxes; assign unique IDs to every object instance to facilitate the identification of recurring objects within the scene and to organize them in a sequential manner; +3. for each object within each sequence, to merge different views of the same object by applying a deep learning model for Point Cloud Registration; this process generates a complete point cloud for each object, which is then used to replace the occluded one in each frame; +4. to fine-tune the base model on Object-Complete labeled frames; this refined model forms our X-Ray Teacher. + +More details for greedy tracking and Point Cloud Registration models can be found in Section 4.2.2. + +The Knowledge Distillation step presented in Figure 4 can be integrated with any semi-supervised 3D Object Detection approach that uses pseudo labeling as a form of self-distillation. All methods in the domain usually follow this paradigm, which underlines the high universality of our approach. From this perspective, the use of Object-Complete frames for pseudo label prediction refines the 3D bounding boxes and improves the quality of self-distillation. The sole limitation is that existing semi-supervised methods [33, 46] typically update the Teacher model's weights using the exponential moving average of the Student's weights. We refrain from this practice because our Teacher is finetuned on Object-Complete frames and thus predicts higher-quality labels (see proofs in Section 4.3). + +# 4. Experiments + +This section covers the datasets, implementation details, experiment results, and a detailed analysis of the different components that affect performance. It starts with Section 4.1, which defines the datasets and metrics used to prove the validity of our ideas. Section 4.2 delves into the network architectures and training parameters, while Section 4.4 analyzes the obtained metric values. Finally, Section 4.3 examines various parts of our approach and their influence + +on the overall performance. + +# 4.1. Data + +To evaluate our models, we use three large-scale autonomous driving datasets: NuScenes and Waymo Open Dataset for the supervised setting and ONCE for the semi-supervised setting. + +NuScenes [5] is a popular outdoor dataset with diverse annotations for different tasks. It has 40,157 annotated samples. For 3D object detection, we provide NuScenes Detection Score (NDS) and mean Average Precision (mAP). + +Waymo Open [32] is also one of the most popular outdoor 3D perception datasets. It contains 1150 point cloud sequences and has more than 200K total frames. All results are evaluated with 3D mean Average Precision (mAP) and its weighted variant (mAPH) + +ONCE [24] is a large-scale dataset with 1 million pout cloud samples from LiDAR and only 15K annotated frames that are divided into train, val, and test with 5k, 3k, and 8k samples, respectively. This dataset is designed exactly for semi-supervised learning tasks and simulates real life: annotations are expensive and time-consuming. All not labeled frames are divided into Small (70 sequences), Medium (321 sequences), and Large (560 sequences) parts. We follow the ONCE Benchmark and use mAP over all classes with the 3D IoU thresholds 0.7, 0.3, and 0.5 for classes "Vehicle", "Pedestrian", and "Cyclist", respectively. + +# 4.2. Implementation Details + +# 4.2.1 Network architectures + +For supervised setting, we use SECOND [50], CenterPointVoxel [47], CBGS [50], DSVT [36] within our framework. The implementations of these models are based on the OpenPCDet [34] library, and we adhere to the default configurations suggested by this library for both training and inference. In addition, we present scaled versions of these models, which are essentially the original networks augmented with five additional convolutional layers stacked on top of the BEVEncoder with the following parameters 1x Conv(512, 128), 3x Conv(128, 128), and 1x Conv(128, 512) with Batch Normalizations and ReLU activations after each convolution. With the help of light Grid Search, we choose the following distillation hyperparameters: $\alpha_{1} = 2$ , $\alpha_{2} = 1$ , $\lambda_{1} = 0.7$ , $\lambda_{2} = 0.3$ , $\lambda_{3} = 1$ . + +For semi-supervised setting, we follow the previous works [35, 46] and use SECOND and CenterPointVoxel models from OpenPCDet for validation of comparison. For training and inference, we use recently proposed refined configurations from work [11]. X-Ray Teacher model is fine-tuned on object-complete frames with the same hyperparameters for 10 epochs. + +For computations, we use 4x A100 40GB GPU and AMD EPYC 7702 CPU. + +![](images/6f6861b1f660e21e5761897ab98ffa2cfcf28b07bdbdf9d363141f1dabdaee63.jpg) +Figure 4. Semi-supervised X-Ray Teacher pipeline for 3D object detection task. Unlabeled sequences are processed by the Objects Temporal Fusion block to create more complete object representations by aggregating information over time. The Student model learns from both pseudo-labeled predictions and actual labeled data with ground truth annotations. + +Table 1. Evaluation of the impact of teacher fine-tuning on ObjectComplete frames in semi-supervised setting. The results indicate that teacher fine-tuning is essential. + +
MethodmAP
X-Ray MT SECOND60.12
X-Ray PT SECOND61.48
X-Ray MT SECOND* (ours)65.26
X-Ray PT SECOND* (ours)68.43
X-Ray MT CenterPoint59.35
X-Ray PT CenterPoint62.17
X-Ray MT CenterPoint* (ours)67.60
X-Ray PT CenterPoint* (ours)70.55
+ +# 4.2.2 Object-Complete Frame Generation + +The key concept of this paper - Object-Complete Frame Generation - combines diverse ideas to achieve optimal Object Completion, encompassing facets such as detection, tracking, and the registration of point clouds. + +The detection phase is construed as an elective stage, exclusively implemented on unlabeled data. Within this stage discerned instances are encapsulated into generated 3D bounding boxes. E.g., a substantial proportion of instances within the ONCE dataset lacks accompanying labels. + +Subsequently, the greedy tracking procedure traverses the entire set of frames, associating the appearance of instances at the i-th frame with potential candidate appearances in the following frame. The list of candidate instances is built by including all instances from the succeeding frame that fall within a prescribed radius, defined + +as twice the maximum dimension across their respective bounding boxes. The nearest instance is selected from the list, while the remaining instances are discarded. The series of matches made using this algorithm are combined into a single sequence, called track. When there are no more potential matches in the next frame, the track is considered as terminated. For the given instance in some specific frame there is the only corresponding track across the entire scene. + +Those prepared instances are then used for Object Completion, a process executed through several sequential steps: + +1. Point clouds associated with instances are extracted from their respective frames, they are translated back to the zero-point of the global basis. The rotation of the instances are also reset to identity. +2. The point clouds corresponding to instances within a common track undergo a merging process, constituting the Point Cloud Registration phase, wherein diverse set of merging approaches is used to glue them into a larger, densely populated point cloud. +3. Later on the corresponding point clouds are replaced with their respective densely populated point cloud from the previous step, restoring their original translation and rotation in a specific frame. + +Table 2. Comparison of three distillation strategies: using only classification and regression heads matching, BEV features and heads matching, and the final pipeline that is the SECOND-Scaled model; experiment performed using Waymo Validation set. + +
TechniquemAP/mAPH L1mAP/mAPH L2
Heads match67.5/63.461.1/57.2
BEV & Heads67.8/63.561.6/57.4
Full Pipeline (ours)68.3/64.361.9/58.0
+ +Various merging strategies were empirically tested in generating Object-Complete point clouds: + +1. Geometry: assumes the objects (with boxes) are best aligned by the detected bounding box, performs both inverse translation and rotation to clear their geometric transformations and then merges intact point clouds. +2. GeDi: uses GeDi Point Cloud Registration method [27]. +3. Greedy Grid: uses Greedy Grid implementation [2]. + +Object-Complete frames have the potential to become exceedingly large, so we employ a point subsampling strategy. + +Table 3. Registration methods comparison for Object Complete Frame Generation on the ONCE Small split. We trained our best SECOND model in the semi-supervised setting with X-Ray Proficient Teacher on three different types of preprocessed data and compared results on ONCE validation set. + +
MethodmAP
Box Geometry67.88
Greedy Grid [2]68.17
GeDi [27]68.43
+ +Table 4. Comparison for supervised 3D Object Detection task on Waymo Open Dataset. We compare baseline models, their scaled versions and X-Ray distillation with default hyperparameters. + +
ModelmAP/mAPH L1mAP/mAPH L2#params
SECOND X-Ray Teacher*85.1/70.375.1/64.75.3m
SECOND67.2/63.161.0/57.25.3m
X-Ray SECOND67.0/62.860.4/56.75.3m
SECOND-Scaled66.8/62.759.4/56.16.2m
X-Ray SECOND-Scaled68.3/64.361.9/58.06.2m
CenterPoint X-Ray Teacher*88.3/78.676.4/72.98.3m
CenterPoint74.4/71.768.2/65.88.3m
X-Ray CenterPoint73.2/69.767.1/64.58.3m
CenterPoint-Scaled74.1/71.567.9/65.39.2m
X-Ray CenterPoint-Scaled75.2 /72.168.9/66.39.2m
DSVT Pillar X-Ray Teacher*89.3/79.779.1/73.48.6m
DSVT Pillar79.5/77.173.2/71.08.6m
X-Ray DSVT Pillar79.2/76.772.6/70.38.6m
DSVT Pillar-Scaled79.6/77.273.3/71.29.5m
X-Ray DSVT Pillar-Scaled80.1/77.973.7/71.49.5m
+ +# 4.3. Ablation Studies + +In this section, we provide the comparison of different components of our approach and show how they affect the performance. Specifically, we prove the usefulness of teacher fine-tuning and compare Point Cloud Registration and distillation methods. We also analyse if dealing with sparsity and incomplete shapes really improves the performance. + +First, we performed a detailed comparison of the semi-supervised performance of our X-Ray method, both with and without teacher fine-tuning using Object-Complete frames. We use Mean Teacher and Proficient Teacher with + +![](images/300a4c7e097f7720962997b69d717caac67e81ff3c19730ff0aca0b95d13eb4c.jpg) +Baseline Model Predictions +X-Ray Teacher Predictions +Figure 5. Visual comparison between noisy and poor baseline (original) 3D detector SECOND (left column) and our X-Ray Teacher that perceives Object-Complete frames. We compare two identical timestamps and view angles. The Baseline model fails to detect some objects while X-Ray Teacher does not. This explains why knowledge distillation is indeed beneficial and should improve models. + +SECOND and CenterPointVoxel models trained on ONCE Small split. Table 1 illustrates the substantial impact of teacher fine-tuning on performance, showing that neglecting this step results in a noticeable performance decline. + +As we noted before, the reconstructed objects will not be perfect in the semi-supervised setting, which is precisely why we included Table 3. PCR models trained on domain objects like cars, pedestrians, cyclists, etc., should improve our method even more. + +As we mentioned earlier, we combine several techniques for the knowledge distillation: BEV features matching, heads output matching with simple regression and KL divergence and detection loss on the teachers predictions. We also compare partial solutions on the Waymo validation set with a SECOND model in Table 2 + +Table 5. Comparison on NuScenes dataset. Our method improves baselines without scaling models due to the fact that NuScene's object-complete frames are less informative compared to Waymo. + +
ModelmAPNDS
X-Ray Teacher* CBGS77.172.4
CBGS50.059.2
X-Ray CBGS (ours)50.860.4
X-Ray Teacher* CenterPoint-Voxel79.577.1
CenterPoint-Voxel53.461.3
X-Ray CenterPoint-Voxel (ours)54.362.9
X-Ray Teacher* Transfusion-L81.378.2
Transfusion-L56.166.3
X-Ray Transfusion-L (ours)56.866.9
+ +Table 6. Performance of X-Ray-powered Mean Teacher and Proficient Teacher methods in the semi-supervised setting using SECOND and CenterPoint baselines, ONCE validation set. Our approach consistently outperforms the state-of-the-art for Semi-Supervised 3D Object Detection in terms of mAP across all splits. + +
MethodSECONDCenterPoint
Train (5k labeled samples)
Pretraining63.2264.41
Small (5k labeled + 100k unlabeled samples)
Mean Teacher64.3166.47
X-Ray MT (ours)65.26 (+0.95)67.60 (+1.13)
Proficient Teacher67.0667.72
X-Ray PT (ours)68.43 (+1.37)68.76 (+1.04)
Medium (5k labeled + 500k unlabeled samples)
Mean Teacher64.7366.87
X-Ray MT (ours)65.62 (+0.89)67.78 (+0.91)
Proficient Teacher67.4968.54
X-Ray PT (ours)68.75 (+1.26)69.96 (+1.42)
Large (5k labeled + 1M unlabeled samples)
Mean Teacher65.0367.45
X-Ray MT (ours)65.97 (+0.94)68.17 (+0.72)
Proficient Teacher67.8969.68
X-Ray PT (ours)69.10 (+1.21)70.55 (+0.87)
+ +We evaluate various Point Cloud Registration (PCR) methods used in the Object Complete Frame Generation process. This analysis, detailed in Table 3, is conducted on a ONCE Small split. The results indicate that superior PCR leads to the creation of less noisy objects, which in turn contributes to improved overall quality. However, it's important to note that methods, such as the Greedy Grid method [2] and GeDi [27], are computationally more expensive. This introduces a trade-off between computational efficiency and the quality of the results, highlighting the need for a balanced approach in the selection of PCR methods. + +# 4.4. Model comparison + +# 4.4.1 Supervised Learning + +We perform model comparisons using SECOND [42], CenterPoint-Voxel [47], and DSVT Pillar [36] on the Waymo dataset, and CBGS [50] and CenterPoint-Voxel on the NuScenes dataset. Additionally, we train scaled versions of these models without the X-Ray Teacher to show that the improvements in detection quality are due to the effectiveness of our method, not just an increase in the number of parameters. We scale Waymo students because of the extremely dense and complete point clouds, unlike NuScenes (see supplementary materials), where simpler data requires fewer parameters to learn meaningful feature representations. The results, presented in Tables 4 and 5, demonstrate that our approach consistently outperforms + +baseline models by 1-2 mAP. + +Table 7. Student performance on ONCE validation small split with different tracking methods used in Objects Temporal Fusion + +
Tracking methodSECONDCenterPoint
Greedy68.4368.76
Kalman Filter + IoU68.5768.91
ReID + Kalman Filter68.7969.12
+ +# 4.4.2 Semi-Supervised Learning + +To validate the effectiveness of our method in the context of pseudo-label-based semi-supervised learning, we perform a comparative analysis with the Mean Teacher [33] and Proficient Teacher [46] methods, which use the SECOND and CenterPointVoxel models. We compare the results obtained with and without the use of the X-Ray Teacher, as detailed in Table 6. Our results show that the application of our approach consistently improves performance, yielding an improvement of $0.8\mathrm{-}1.4\mathrm{mAP}$ . + +# 5. Conclusion + +In our research, we proposed the innovative X-Ray Teacher framework, tailored to improve 3D Object Detection models in supervised and semi-supervised settings. Our extensive results have shown that this approach not only achieves state-of-the-art performance in the semi-supervised setting on the ONCE benchmark, but also consistently improves the quality of supervised models on NuScenes and Waymo Open Dataset. The main contributions of our work include the design of the X-Ray Teacher framework, the development of the Objects Temporal Fusion block for generating Object-Complete frames for data lacking ground truth tracking labels, and the demonstration of the potential of our method to improve the performance of any supervised model trained on sequential data. Future work will focus on optimizing the Objects Temporal Fusion block for more complex environments and exploring the integration of our framework with a broader range of model architectures and applications. + +# 6. Acknowledgements + +This research was supported in part through computational resources of HPC facilities at HSE University. + +The work of I. Makarov on 3D object detection related work was supported by a grant for research centers in the field of artificial intelligence, provided by the Analytical Center in accordance with the subsidy agreement (agreement identifier 000000D730321P5Q0002) and the agreement with the Ivannikov Institute for System Programming of dated November 2, 2021 No. 70-2021-00142. + +# References + +[1] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018. +[2] David Bojanić, Kristijan Bartol, Josep Forest, Stefan Gumhold, Tomislav Petković, and Tomislav Pribanić. Challenging the universal representation of deep models for 3d point cloud registration. In BMVC 2022 Workshop Universal Representations for Computer Vision, 2022. +[3] Daniel Brodeski, Igal Bilik, and Raja Giryes. Deep radar detector. In 2019 IEEE Radar Conference (RadarConf), pages 1-6, 2019. +[4] Collin Burns et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. arXiv preprint arXiv:2312.09390, 2023. +[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[6] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M. Alvarez. Focalformer3d: Focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8394-8405, October 2023. +[7] Hyeon Cho, Junyong Choi, Geonwoo Baek, and Wonjun Hwang. itkd: Interchange transfer-based knowledge distillation for 3d object detection. In CVPR, 2023. +[8] Xu Dong, Pengluo Wang, Pengyue Zhang, and Langechuan Liu. Probabilistic oriented object detection in automotive radar. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 458-467, 2020. +[9] J. Fang, D. Zhou, F. Yan, T. Zhao, F. Zhang, Y. Ma, L. Wang, and R. Yang. Augmented lidar simulator for autonomous driving. In IEEE Robotics and Automation, 2020. +[10] Jeff Dean Geoffrey Hinton, Oriol Vinyals. Distilling the knowledge in a neural network. In arXiv preprint arXiv:1503.02531, 2015. +[11] Maksim Golyadkin, Alexander Gambashidze, Ildar Nurgaliev, and Ilya Makarov. Refining the once benchmark with hyperparameter tuning. IEEE Access, 12:3805-3814, 2024. +[12] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. +[13] Chengming Hu, Xuan Li, Dan Liu, Haolun Wu, Xi Chen, Ju Wang, and Xue Liu. Teacher-student architecture for knowledge distillation: A survey. In arXiv preprint arXiv:2210.17332, 2022. +[14] Haotian Hu, Fanyi Wang, Jingwen Su, Yaonong Wang, Laifeng Hu, Weiye Fang, Jingwei Xu, and Zhiwang Zhang. Ea-iss: Edge-aware lift-splat-shot framework for 3d bev object detection. arXiv preprint arXiv:2303.17895, 2023. +[15] Ilia Indyk and Ilya Makarov. Monovan: Visual attention for self-supervised monocular depth estimation. In 2023 IEEE International Symposium on Mixed and Augmented Reality + +(ISMAR), pages 1211-1220. IEEE, 2023. +[16] Aleksei Karpov and Ilya Makarov. Exploring efficiency of vision transformers for self-supervised monocular depth estimation. In 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 711-719, 2022. +[17] Junho Koh, Junhyung Lee, Youngwoo Lee, Jaekyum Kim, and Jun Won Choi. Mgtanet: Encoding sequential lidar points using long short-term motion-guided temporal attention for 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):1179-1187, Jun. 2023. +[18] Haoang Li, Jinhu Dong, Binghui Wen, Ming Gao, Tianyu Huang, Yun-Hui Liu, and Daniel Cremers. Ddit: Semantic scene completion via deformable deep implicit templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21894-21904, October 2023. +[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[20] Ziyu Li, Yuncong Yao, Zhibin Quan, Jin Xie, and Wankou Yang. Spatial information enhancement network for 3d object detection from point cloud. Pattern Recognition, page 108684, 2022. +[21] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. +[22] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 924–932, 2019. +[23] Ilya Makarov, Vladimir Aliev, and Olga Gerasimova. Semidense depth interpolation using deep convolutional neural networks. In Proceedings of the 25th ACM International Conference on Multimedia, MM '17, page 1407-1415, New York, NY, USA, 2017. Association for Computing Machinery. +[24] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. +[25] Michael Meyer, Georg Kuschk, and Sven Tomforde. Graph convolutional networks for 3d object detection on radar data. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3053-3062, 2021. +[26] Mahyar Najibi, Guangda Lai, Abhijit Kundu, Zhichao Lu, Vivek Rathod, Thomas Funkhouser, Caroline Pantofaru, David Ross, Larry S. Davis, and Alireza Fathi. Dops: Learning to detect 3d objects and predict their 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[27] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud regis + +tration. In IEEE Trans. on Pattern Analysis and Machine Intelligence, (early access) 2022. +[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. +[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. +[30] Ilia Semenkov, Aleksei Karpov, Andrey V Savchenko, and Ilya Makarov. Inpainting semantic and depth features to improve visual place recognition in the wild. IEEE Access, 2024. +[31] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020. +[32] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. +[34] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020. +[35] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J. Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2020. +[36] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, June 2023. +[37] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023. +[38] Haiyang Wang, Wenguan Wang, Xizhou Zhu, Jifeng Dai, and Liwei Wang. Collaborative visual navigation. arXiv + +preprint arXiv:2107.01151, 2021. +[39] Yan Wang, Bin Yang, Rui Hu, Ming Liang, and Raquel Urtasun. Plumenet: Efficient 3d object detection from stereo images. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3383-3390, 2021. +[40] Xinshuo Weng and Kris Kitani. Monocular 3d object detection with pseudo-lidar point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019. +[41] Qiangeng Xu, Yiqi Zhong, and Ulrich Neumann. Behind the curtain: Learning occluded shapes for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2893-2901, 2022. +[42] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. +[43] Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018. +[44] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. In NeurIPS, 2022. +[45] Lehan Yang and Kele Xu. Cross modality knowledge distillation for multi-modal aerial view object classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 382-387, June 2021. +[46] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022. +[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. +[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11087, 2020. +[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In European Conference on Computer Vision, pages 496-513. Springer, 2022. +[50] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. In arXiv preprint arXiv:1908.09492, 2019. +[51] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3357–3364, 2017. \ No newline at end of file diff --git a/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/images.zip b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f51eb939e2963f90c806f6bf06ed373513407455 --- /dev/null +++ b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05f29e795124d72be868f1215a4a5dd09c56595b7f4b9814e27fa48195a0cc9b +size 511499 diff --git a/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/layout.json b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c45c7efbff26b1c68aa646608534566f0b3a8de7 --- /dev/null +++ b/2024/Weak-to-Strong 3D Object Detection with X-Ray Distillation/layout.json @@ -0,0 +1,7665 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 483, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 483, + 120 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 483, + 120 + ], + "type": "text", + "content": "Weak-to-Strong 3D Object Detection with X-Ray Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 233, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 233, + 156 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 233, + 156 + ], + "type": "text", + "content": "Alexander Gambashidze\\*1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 263, + 143, + 362, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 143, + 362, + 157 + ], + "spans": [ + { + "bbox": [ + 263, + 143, + 362, + 157 + ], + "type": "text", + "content": "Aleksandr Dadukin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 392, + 143, + 494, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 143, + 494, + 157 + ], + "spans": [ + { + "bbox": [ + 392, + 143, + 494, + 157 + ], + "type": "text", + "content": "Maxim Golyadkin1,2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 189, + 163, + 281, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 163, + 281, + 176 + ], + "spans": [ + { + "bbox": [ + 189, + 163, + 281, + 176 + ], + "type": "text", + "content": "Maria Razzhivina" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 163, + 403, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 163, + 403, + 177 + ], + "spans": [ + { + "bbox": [ + 326, + 163, + 403, + 177 + ], + "type": "text", + "content": "Ilya Makarov1,3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 205, + 191, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 205, + 191, + 217 + ], + "spans": [ + { + "bbox": [ + 143, + 205, + 191, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 245, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 290, + 605 + ], + "type": "text", + "content": "This paper addresses the critical challenges of sparsity and occlusion in LiDAR-based 3D object detection. Current methods often rely on supplementary modules or specific architectural designs, potentially limiting their applicability to new and evolving architectures. To our knowledge, we are the first to propose a versatile technique that seamlessly integrates into any existing framework for 3D Object Detection, marking the first instance of Weak-to-Strong generalization in 3D computer vision. We introduce a novel framework, X-Ray Distillation with Object-Complete Frames, suitable for both supervised and semi-supervised settings, that leverages the temporal aspect of point cloud sequences. This method extracts crucial information from both previous and subsequent LiDAR frames, creating Object-Complete frames that represent objects from multiple viewpoints, thus addressing occlusion and sparsity. Given the limitation of not being able to generate Object-Complete frames during online inference, we utilize Knowledge Distillation within a Teacher-Student framework. This technique encourages the strong Student model to emulate the behavior of the weaker Teacher, which processes simple and informative Object-Complete frames, effectively offering a comprehensive view of objects as if seen through X-ray vision. Our proposed methods surpass state-of-the-art in semi-supervised learning by 1-1.5 mAP and enhance the performance of five established supervised models by 1-2 mAP on standard autonomous driving datasets, even with default hyperparameters. Code for Object-Complete frames is available here: https://github.com/sakharok13/X-Ray-Teacher-Patching-Tools." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 243, + 537, + 430 + ], + "blocks": [ + { + "bbox": [ + 313, + 243, + 537, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 537, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 537, + 430 + ], + "type": "image", + "image_path": "00632d629e00b25464ef5870158fe326a1d02ef4d2b4313bdcc4341c9666ed51.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 445, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 535 + ], + "type": "text", + "content": "Figure 1. 3D object detection directly from sparse LiDAR data (top left) provides noisy predictions (bottom left). Adding object completion stage (top right) helps to train 3D object detection X-Ray Teacher, which is robust and can be distilled to baseline model. Red, Yellow and Blue colors of bounding boxes are related to classical LiDAR-based object detection, Our model on Object-Complete frames predictions and Ground Truth labels, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 560, + 529, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 560, + 529, + 581 + ], + "spans": [ + { + "bbox": [ + 343, + 560, + 529, + 581 + ], + "type": "text", + "content": "You're just not thinking fourth dimensionally... the bridge will exist." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 404, + 589, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 589, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 404, + 589, + 545, + 599 + ], + "type": "text", + "content": "Dr. Emmett Brown, Back to the Future" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 621, + 386, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 386, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 386, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "3D object detection is a fundamental task in the field of computer vision and autonomous systems, playing a key role in the advancement of self-driving technology [1] and contributing significantly to the robotics industry [38, 51]. Currently, LiDAR-based 3D object detection [6, 17, 36] demonstrates superior performance compared to camera-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 673, + 274, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 673, + 274, + 683 + ], + "spans": [ + { + "bbox": [ + 57, + 673, + 274, + 683 + ], + "type": "text", + "content": "*Corresponding author, email: alexandergambashidze@gmail.com" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 60, + 683, + 192, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 683, + 192, + 693 + ], + "spans": [ + { + "bbox": [ + 60, + 683, + 192, + 693 + ], + "type": "text", + "content": "1Artificial Intelligence Research Institute" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 60, + 693, + 115, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 115, + 703 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 115, + 703 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 693, + 115, + 703 + ], + "type": "text", + "content": "HSE University" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 60, + 703, + 93, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 93, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 93, + 712 + ], + "type": "text", + "content": "3ISP RAS" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15055" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "based [19, 39, 40, 45] and radar-based [3, 8, 22, 25] approaches. Furthermore, LiDAR point clouds are the key ingredient of multimodal fusion-based approaches [14, 21, 37], so LiDAR-based 3D detection continues to be a strong focus of the research community." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 135, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 289, + 529 + ], + "type": "text", + "content": "The point cloud-based 3D object detection challenges the following issues: sparsity, occlusions, and the complexity of 3D data annotation. Sparsity: large point clouds are sparse due to the LiDAR sensing process's inherent characteristics that leads to an imprecise representation of the captured scene. Additionally, there is an imbalance in point density. In particular, the point cloud is sparser in the far range and contains less spatial information affecting feature representation and box prediction. Occlusions: another problem arises from the frequent occurrence of partial occlusion in LiDAR frames. This primarily happens due to the fact that the frames are acquired from a single fixed point of view, making them essentially 2.5D. To detect and accurately locate a highly occluded object, a detector must recognize the hidden shapes of the object even when a significant portion of its parts is missing. Since the absence of certain shapes inevitably affects object perception, this becomes a critical detection challenge. Our previous approaches focused on depth completion [23], depth inpainting [30] or self-supervised depth pretraining [15, 16] could not address occlusion problems. Data annotation: finally, data annotation is a formidable challenge in 3D object detection due to the complexity of annotating objects in three-dimensional space. For example, a skilled annotator can spend weeks annotating just one hour of LiDAR data [9, 24]. This problem is partially addressed by semi-supervised learning approaches in a teacher-student framework [33, 35, 46, 48]. However, the quality of pseudo labels and the performance of these methods are limited due to the aforementioned sparsity and occlusion problems. Therefore, overcoming these inherent challenges is essential to improve the accuracy and reliability of 3D detection systems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": "The challenges posed by sparsity and occlusion have led to the formulation of several methodologies. Some of them [42] are aimed at improving the computational efficiency of processing sparse data without improving the quality of the predictions. Others, including our work, focus on improving detection performance on sparse occluded data [18, 20, 26, 41]. Li et al. [18] enforce shape constraints to improve object localization with explicit shape priors obtained from a database of CAD models. Najibi et al. [26] and Li et al. [20] do the same by implicitly introducing priors with novel modules pretrained for point cloud completion and SDF approximation. Xu et al. [41] propose a Shape Occupancy Probability estimation module to refine predicted bounding boxes for occluded objects. However, all of these mentioned approaches utilize new modules and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "adapt the model architecture to take advantage of them. As a result, these methods have limited applicability to new architectures that will emerge as the field evolves." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 116, + 546, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 116, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 546, + 439 + ], + "type": "text", + "content": "In this work, we address all three major challenges in 3D object detection with our novel framework called X-Ray Distillation with Object-Complete Frames that is easily plugged in existing approaches and new architectures. It is designed for universal application to any LiDAR-based detector, improving performance on sparse and occluded objects. Our approach exploits the properties of existing large-scale autonomous driving datasets, which consist of sequences of LiDAR frames. Such property makes it possible to reconstruct complete shapes for occluded objects using other occurrences of these objects in the sequence, ensuring that all objects are equipped with points from all available viewpoints in a scene. We then use this completed data in the Teacher-Student framework for both semi-supervised learning and knowledge distillation in a supervised setting. We train our Teacher on extremely informative Object-Complete frames thus making it a weaker model [4]. Then, we use it to extract features from such simple Object-Complete frames and distills this knowledge to a stronger Student, which operates with the original data, to guide him on how to extract rich features from occluded objects. To generate Object-Complete frames, we leverage ground truth object tracking labels. Since there are no labels for the majority of data in the semi-supervised setting, we propose an Objects Temporal Fusion block to detect, track, and use point cloud registration techniques to construct Object-Complete frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 445, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 541 + ], + "type": "text", + "content": "We validate the proposed X-Ray Teacher framework on nuScenes [5] and Waymo Open Dataset [32] 3D object detection benchmarks for SECOND [42], CenterPoint [47], and DSVT [36] models in a supervised learning paradigm. Experiments show a steady improvement by 1-2 mAP with minimal or no impact on time and computational resources during the inference stage. For semi-supervised performance evaluation, we use ONCE [24] benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "content": "Applying these novel ideas for 3D object detection, we provide the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 579, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 579, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 579, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 579, + 545, + 613 + ], + "type": "text", + "content": "1. We propose the X-Ray Teacher framework for semi-supervised learning, which achieves state-of-the-art performance on the ONCE benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 616, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 545, + 674 + ], + "type": "text", + "content": "2. We show that our approach improves the quality of four supervised learning models, including the current state-of-the-art model, and demonstrates the potential to improve the performance of any supervised model trained on sequential data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 675, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 675, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 306, + 675, + 545, + 711 + ], + "type": "text", + "content": "3. We suggest the Objects Temporal Fusion block to generate Object-Complete frames for data that lacks ground truth tracking labels." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15056" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 225, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 225, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 225, + 103 + ], + "type": "text", + "content": "2.1. Object Detection for point clouds" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 287, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 287, + 265 + ], + "type": "text", + "content": "Modern 3D detectors predominantly use point clouds, voxel representations, or a combination of both. The pioneering PointNet model family [28, 29] provided a significant step forward in 3D recognition. However, direct processing of point clouds poses several challenges: it requires point sampling, grouping, and computation of point-wise features, which can be computationally intensive. To integrate insights from 2D computer vision, a transition to a pixel equivalent, namely voxels, is required. Methods such as [41, 43, 47, 49] first convert point clouds to voxels, followed by the use of 3D sparse convolutions [12]. Recent advances also include the integration of modified self-attention layers, achieving state-of-the-art results [36]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 273, + 208, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 208, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 208, + 285 + ], + "type": "text", + "content": "2.2. Semi-supervised 3D detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 291, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 287, + 434 + ], + "type": "text", + "content": "Although Semi-supervised 3D object detection is not as extensively researched as 2D detection, it has seen some significant contributions. Among these, SESS [48] stands out for using the general Mean Teacher approach with data augmentations and consistency loss. Similarly, 3D IoUMatch [35] is notable for its unique localization strategies and 3D IoU-guided techniques for box filtering. Proficient Teachers [46] is notable for its box voting and contrastive losses. We do not propose yet another standalone method for Semi-supervised 3D Object Detection; rather, our emphasis is on creating a plug-in technique designed to augment and enhance the performance of existing methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 443, + 177, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 177, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 177, + 454 + ], + "type": "text", + "content": "2.3. Knowledge Distillation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 460, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 567 + ], + "type": "text", + "content": "The concept of Knowledge Distillation (KD) was first introduced by Hinton et al. [10]. KD describes a learning approach where a larger teacher network guides the training of a smaller student network for various tasks [13]. Broadly, KD methodologies are categorized into two types: logits/regression distillation and feature map distillation. Our focus is on a hybrid approach, combining these methods: matching feature maps, regression, and classification heads with pseudo labels generated by the teacher model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 568, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 641 + ], + "type": "text", + "content": "In 3D object detection, Knowledge Distillation is mainly utilized to minimize parameters and FLOPS while aiming to preserve box prediction quality [7, 44]. Yet, methods that aim to outperform state-of-the-art models are uncommon. In contrast to this trend, our X-Ray Teacher model challenges the norm by providing the quality improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 129, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 129, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 129, + 664 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 200, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 200, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 200, + 684 + ], + "type": "text", + "content": "3.1. Overview of X-Ray Teacher" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We introduce a novel training framework to address the challenges of sparsity and occlusion in 3D Object Detec" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "tion based on LiDAR data. This framework is not limited to any specific object detection model and has the potential for applications across various deep learning architectures. Our method is designed to process the LiDAR data structured as a sequence of frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 134, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 545, + 193 + ], + "type": "text", + "content": "The two core elements of our approach are Object Complete Frames Generation and Teacher-Student Knowledge Distillation. Our approach for 3D object detection can be applied in both, supervised and semi-supervised settings, with minor differences in the elements implementation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 194, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 302 + ], + "type": "text", + "content": "Object Complete Frames Generation. In this step, we reconstruct the complete shapes of objects presented in the scene by utilizing information from the other frames within the same sequence. Given that autonomous driving datasets are composed of sequential data, we can efficiently leverage their temporal nature: we add points from both the future and the past when objects are observed from different viewpoints. It allows us to reconstruct the complete shapes of objects without shape databases or reconstruction modules." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": "In order to verify the validity of our approach, we trained a CenterPoint [47] model on both, the original and object-complete NuScenes [5] datasets, and then evaluated their performance on the respective validation sets. The models trained on the original and the object-complete frames achieved mAP scores of " + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "inline_equation", + "content": "59.2\\%" + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "inline_equation", + "content": "79.5\\%" + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": ", respectively. This difference of " + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "inline_equation", + "content": "20~\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": " suggests that 1) it would be beneficial to transform unlabeled original point clouds into Object-Complete ones and to annotate them with a corresponding X-Ray Teacher pretrained on such informative frames 2) the features extracted by a weaker X-Ray Teacher might be distilled to a stronger student to share the knowledge of complete shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 460, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 674 + ], + "type": "text", + "content": "Teacher-Student Knowledge Distillation. The necessity for this step arises because we cannot generate Object Complete Frames during the online inference stage, as it is not possible to access future data. Therefore, we need to encourage the model to behave as if it were observing shape-complete objects, even when dealing with occluded ones. A well-known method for enabling a deep learning model to imitate another model's behavior involves using Knowledge Distillation within a Teacher-Student framework. However, for conventional Knowledge Distillation, both the Teacher and Student models typically process data of the same complexity (the only difference may lie in the complexity of augmentations [31]). In contrast to the standard knowledge distillation, we enrich the data for training Teacher, which significantly improves its performance in 3D object detection. Then, we teach the Student to extract important information from less detailed data by distilling knowledge from the Teacher model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Instead of simplifying the Student model, as it is usually done in standard Knowledge Distillation, we take an opposite approach and design the Student to be more complex" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15057" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 521, + 207 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 521, + 207 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 521, + 207 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 521, + 207 + ], + "type": "image", + "image_path": "c60898a961fecfaffa9e7f9afbd0deff0743ee49aa7e95fa41691618e859d675.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. Overall X-Ray Knowledge Distillation for Supervised Learning. X-Ray Teacher is frozen and pretrained on Object-Complete frames which are taken as input. The Student is guided to mimic the Teacher's behaviour through Knowledge Distillation losses: " + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "inline_equation", + "content": "L_{feat}" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": " for intermediate embeddings matching, " + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": " for bounding box regression, " + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": " for basic detection, and " + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": " for classification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 319 + ], + "type": "text", + "content": "than the Teacher. It helps to extract high-quality information from more intricate and ambiguous data and demand a Student model to be more robust and to have a more complex receptive field capacity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 367 + ], + "type": "text", + "content": "In what follows, we provide detailed descriptions of how we implement Object Complete Frames Generation and Teacher-Student Knowledge Distillation steps in both supervised and semi-supervised settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 194, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 194, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 194, + 384 + ], + "type": "text", + "content": "3.2. Supervised X-Ray Teacher" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 390, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 474 + ], + "type": "text", + "content": "In the supervised setting of 3D Object Detection, models are trained and evaluated using datasets that provide labeled data, including precise bounding boxes and instance IDs. Object-Complete Frame Generation for labeled data involves aggregating objects based on their instance IDs and merging different views into a unified point cloud (see Section 4.2.2 for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 475, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 606 + ], + "type": "text", + "content": "For the distillation process (see Figure 2), we train Teacher model on Object-Complete frames and then freeze it. Then, we train baseline model (playing Student role) to directly minimize Knowledge Distillation losses inspired by [44]. The distillation is done by matching Teacher and Student backbone encoders' embeddings, output labels for bounding box regression, classes distribution for classification task (objects like pedestrians, cars, cyclists, etc.), and intermediate features obtained from the outputs of regression and classification heads before postprocessing (assigning labels). Specifically, we define the following losses:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 620, + 287, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 620, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 61, + 620, + 287, + 652 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {h e a d s}} ^ {K D} = \\alpha_ {1} \\mathcal {L} _ {\\text {r e g}} ^ {K D} + \\alpha_ {2} \\mathcal {L} _ {\\text {c l s}} ^ {K D} (1) \\\\ = \\alpha_ {1} D _ {K L} \\left(S _ {\\mathrm {c l s}} \\| T _ {\\mathrm {c l s}}\\right) + \\alpha_ {2} \\mathrm {M S E} \\left(S _ {\\mathrm {r e g}}, T _ {\\mathrm {r e g}}\\right) (2) \\\\ \\end{array}", + "image_path": "dc2de9c087b1cf6414326fd9bbc6889770e129b60daa65121e360af85aae1b88.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 654, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 654, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 64, + 654, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f e a t}} ^ {K D} = \\operatorname {M S E} \\left(T _ {\\text {b a c k}}, \\phi \\left(\\omega \\left(S _ {\\text {b a c k}}\\right)\\right)\\right) \\tag {3}", + "image_path": "819f5b6a816ff5b2590b0ff1d49114547b023ff80df7e9bd07a56b3cf2535c13.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 670, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 670, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 64, + 670, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\det } ^ {K D} = \\mathcal {L} _ {\\text {d e t e c t i o n}} \\left(S _ {\\text {p r e d s}}, \\tilde {T} _ {\\text {b o x e s}}\\right) \\tag {4}", + "image_path": "7100f4170ef1df336ebac2cc3f35539408376e988aea92c47924e7be1408ff13.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " are the outputs of our Teacher and Student models, respectively. The Student takes the original frame " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "F" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": "as input, while the Teacher receives the Object-Complete Frame " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": ", so " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "F \\subset \\tilde{F}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "S_{\\text{back}}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "T_{\\text{back}}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " are referred to the output of the backbone module and " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "S_{\\text{reg}}, T_{\\text{reg}}, S_{\\text{cls}}, T_{\\text{cls}}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " are the outputs of regression and classifications heads. " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\tilde{T}_{\\text{boxes}}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " are the X-Ray Teacher's predicted boxes after postprocessing. " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "S_{\\text{preds}}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " is an overall Student's output; " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\alpha_1, \\alpha_2" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " are non-negative hyper-parameters. " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "L_{\\text{detection}}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " is a basic detection loss that is used for training 3D object detection models. " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " is a 1x1 Convolution to better match the Teacher's feature maps. " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " refers to some extra convolutions to make the Student more flexible. We discovered that X-Ray Distillation does not need extra convolutions in the case of the NuScenes dataset, so their usage for encoder feature adjustment is also a hyperparameter of the model. By MSE, we mean Mean Squared Error." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 450, + 542, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 450, + 542, + 462 + ], + "spans": [ + { + "bbox": [ + 317, + 450, + 542, + 462 + ], + "type": "text", + "content": "Finally, the training objective can be written as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 478, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 478, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 353, + 478, + 545, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {h e a d s}} ^ {K D} + \\lambda_ {2} \\mathcal {L} _ {\\text {f e a t}} ^ {K D} + \\lambda_ {3} \\mathcal {L} _ {\\det } ^ {K D} \\tag {5}", + "image_path": "9080f29c74b6dac262a8a2143e34ae93e45462ac4aecf6e8879215767b68ea78.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 498, + 544, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 544, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 544, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 498, + 544, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 305, + 498, + 544, + 521 + ], + "type": "text", + "content": " are non-negative hyper-parameters balancing the contribution of each term." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 527, + 479, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 479, + 540 + ], + "type": "text", + "content": "3.3. Semi-supervised X-Ray Teacher" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 604 + ], + "type": "text", + "content": "Semi-supervised learning is characterized by the availability of a small amount of labeled data and a larger pool of unlabeled data, making the use of the Object Complete Frame generation approach proposed for supervised settings infeasible." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In order to overcome this limitation, we introduce the Objects Temporal Fusion block (as shown in Figure 3), which is designed to enable Object Complete Frame generation in situations where ground truth labeling is missing. This block leverages a model pretrained on labeled data to detect and track objects across unlabeled sequences. Subsequently, it employs Point Cloud Registration (PCR) to merge objects points from different views for all detected objects." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15058" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 288, + 119 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 119 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 119 + ], + "type": "image", + "image_path": "b2e12ca954dcad7fe353ec5147fffbb5f8d7c5adba8b04d357ea3bd16c206911.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 129, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 129, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 287, + 196 + ], + "type": "text", + "content": "Figure 3. Object-Complete Frame generation process for semi-supervised setting. It consists of tracking and Point Cloud Registration. We track objects across all frames in the whole sequence, then we use Point Cloud Registration to merge points that represent the same object from different views, and finally we replace the original object with the new, complete one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 287, + 240 + ], + "type": "text", + "content": "The provided steps outline a precise and detailed procedure for Objects Temporal Fusion:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 240, + 287, + 396 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 240, + 270, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 270, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 270, + 251 + ], + "type": "text", + "content": "1. to label all LiDAR frames using a pretrained model;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 253, + 287, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 287, + 311 + ], + "type": "text", + "content": "2. to greedily track objects across all frames in each sequence using predicted bounding boxes; assign unique IDs to every object instance to facilitate the identification of recurring objects within the scene and to organize them in a sequential manner;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 312, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 287, + 371 + ], + "type": "text", + "content": "3. for each object within each sequence, to merge different views of the same object by applying a deep learning model for Point Cloud Registration; this process generates a complete point cloud for each object, which is then used to replace the occluded one in each frame;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": "4. to fine-tune the base model on Object-Complete labeled frames; this refined model forms our X-Ray Teacher." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "type": "text", + "content": "More details for greedy tracking and Point Cloud Registration models can be found in Section 4.2.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 419, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 587 + ], + "type": "text", + "content": "The Knowledge Distillation step presented in Figure 4 can be integrated with any semi-supervised 3D Object Detection approach that uses pseudo labeling as a form of self-distillation. All methods in the domain usually follow this paradigm, which underlines the high universality of our approach. From this perspective, the use of Object-Complete frames for pseudo label prediction refines the 3D bounding boxes and improves the quality of self-distillation. The sole limitation is that existing semi-supervised methods [33, 46] typically update the Teacher model's weights using the exponential moving average of the Student's weights. We refrain from this practice because our Teacher is finetuned on Object-Complete frames and thus predicts higher-quality labels (see proofs in Section 4.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 597, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "This section covers the datasets, implementation details, experiment results, and a detailed analysis of the different components that affect performance. It starts with Section 4.1, which defines the datasets and metrics used to prove the validity of our ideas. Section 4.2 delves into the network architectures and training parameters, while Section 4.4 analyzes the obtained metric values. Finally, Section 4.3 examines various parts of our approach and their influence" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 419, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 419, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 419, + 84 + ], + "type": "text", + "content": "on the overall performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 90, + 351, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 90, + 351, + 101 + ], + "spans": [ + { + "bbox": [ + 306, + 90, + 351, + 101 + ], + "type": "text", + "content": "4.1. Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "content": "To evaluate our models, we use three large-scale autonomous driving datasets: NuScenes and Waymo Open Dataset for the supervised setting and ONCE for the semi-supervised setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "type": "text", + "content": "NuScenes [5] is a popular outdoor dataset with diverse annotations for different tasks. It has 40,157 annotated samples. For 3D object detection, we provide NuScenes Detection Score (NDS) and mean Average Precision (mAP)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 204, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 263 + ], + "type": "text", + "content": "Waymo Open [32] is also one of the most popular outdoor 3D perception datasets. It contains 1150 point cloud sequences and has more than 200K total frames. All results are evaluated with 3D mean Average Precision (mAP) and its weighted variant (mAPH)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "content": "ONCE [24] is a large-scale dataset with 1 million pout cloud samples from LiDAR and only 15K annotated frames that are divided into train, val, and test with 5k, 3k, and 8k samples, respectively. This dataset is designed exactly for semi-supervised learning tasks and simulates real life: annotations are expensive and time-consuming. All not labeled frames are divided into Small (70 sequences), Medium (321 sequences), and Large (560 sequences) parts. We follow the ONCE Benchmark and use mAP over all classes with the 3D IoU thresholds 0.7, 0.3, and 0.5 for classes \"Vehicle\", \"Pedestrian\", and \"Cyclist\", respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 401, + 440, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 440, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 440, + 414 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 420, + 435, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 435, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 435, + 430 + ], + "type": "text", + "content": "4.2.1 Network architectures" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": "For supervised setting, we use SECOND [50], CenterPointVoxel [47], CBGS [50], DSVT [36] within our framework. The implementations of these models are based on the OpenPCDet [34] library, and we adhere to the default configurations suggested by this library for both training and inference. In addition, we present scaled versions of these models, which are essentially the original networks augmented with five additional convolutional layers stacked on top of the BEVEncoder with the following parameters 1x Conv(512, 128), 3x Conv(128, 128), and 1x Conv(128, 512) with Batch Normalizations and ReLU activations after each convolution. With the help of light Grid Search, we choose the following distillation hyperparameters: " + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 2" + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 1" + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.7" + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.3" + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{3} = 1" + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": "For semi-supervised setting, we follow the previous works [35, 46] and use SECOND and CenterPointVoxel models from OpenPCDet for validation of comparison. For training and inference, we use recently proposed refined configurations from work [11]. X-Ray Teacher model is fine-tuned on object-complete frames with the same hyperparameters for 10 epochs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "For computations, we use 4x A100 40GB GPU and AMD EPYC 7702 CPU." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15059" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 72, + 506, + 257 + ], + "blocks": [ + { + "bbox": [ + 86, + 72, + 506, + 257 + ], + "lines": [ + { + "bbox": [ + 86, + 72, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 86, + 72, + 506, + 257 + ], + "type": "image", + "image_path": "6f6861b1f660e21e5761897ab98ffa2cfcf28b07bdbdf9d363141f1dabdaee63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 301 + ], + "type": "text", + "content": "Figure 4. Semi-supervised X-Ray Teacher pipeline for 3D object detection task. Unlabeled sequences are processed by the Objects Temporal Fusion block to create more complete object representations by aggregating information over time. The Student model learns from both pseudo-labeled predictions and actual labeled data with ground truth annotations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 361, + 254, + 473 + ], + "blocks": [ + { + "bbox": [ + 46, + 319, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 353 + ], + "type": "text", + "content": "Table 1. Evaluation of the impact of teacher fine-tuning on ObjectComplete frames in semi-supervised setting. The results indicate that teacher fine-tuning is essential." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 361, + 254, + 473 + ], + "lines": [ + { + "bbox": [ + 80, + 361, + 254, + 473 + ], + "spans": [ + { + "bbox": [ + 80, + 361, + 254, + 473 + ], + "type": "table", + "html": "
MethodmAP
X-Ray MT SECOND60.12
X-Ray PT SECOND61.48
X-Ray MT SECOND* (ours)65.26
X-Ray PT SECOND* (ours)68.43
X-Ray MT CenterPoint59.35
X-Ray PT CenterPoint62.17
X-Ray MT CenterPoint* (ours)67.60
X-Ray PT CenterPoint* (ours)70.55
", + "image_path": "cf0c2dd2d6928a6325ccc23e78d03f85a54522546b75f9b709fe055ae4f2fb85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 496, + 235, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 235, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 235, + 509 + ], + "type": "text", + "content": "4.2.2 Object-Complete Frame Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 567 + ], + "type": "text", + "content": "The key concept of this paper - Object-Complete Frame Generation - combines diverse ideas to achieve optimal Object Completion, encompassing facets such as detection, tracking, and the registration of point clouds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 639 + ], + "type": "text", + "content": "The detection phase is construed as an elective stage, exclusively implemented on unlabeled data. Within this stage discerned instances are encapsulated into generated 3D bounding boxes. E.g., a substantial proportion of instances within the ONCE dataset lacks accompanying labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Subsequently, the greedy tracking procedure traverses the entire set of frames, associating the appearance of instances at the i-th frame with potential candidate appearances in the following frame. The list of candidate instances is built by including all instances from the succeeding frame that fall within a prescribed radius, defined" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "type": "text", + "content": "as twice the maximum dimension across their respective bounding boxes. The nearest instance is selected from the list, while the remaining instances are discarded. The series of matches made using this algorithm are combined into a single sequence, called track. When there are no more potential matches in the next frame, the track is considered as terminated. For the given instance in some specific frame there is the only corresponding track across the entire scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 418, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 442 + ], + "type": "text", + "content": "Those prepared instances are then used for Object Completion, a process executed through several sequential steps:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 443, + 545, + 597 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 443, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 545, + 490 + ], + "type": "text", + "content": "1. Point clouds associated with instances are extracted from their respective frames, they are translated back to the zero-point of the global basis. The rotation of the instances are also reset to identity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 491, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 545, + 550 + ], + "type": "text", + "content": "2. The point clouds corresponding to instances within a common track undergo a merging process, constituting the Point Cloud Registration phase, wherein diverse set of merging approaches is used to glue them into a larger, densely populated point cloud." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 550, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 545, + 597 + ], + "type": "text", + "content": "3. Later on the corresponding point clouds are replaced with their respective densely populated point cloud from the previous step, restoring their original translation and rotation in a specific frame." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 309, + 663, + 545, + 712 + ], + "blocks": [ + { + "bbox": [ + 305, + 610, + 545, + 654 + ], + "lines": [ + { + "bbox": [ + 305, + 610, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 545, + 654 + ], + "type": "text", + "content": "Table 2. Comparison of three distillation strategies: using only classification and regression heads matching, BEV features and heads matching, and the final pipeline that is the SECOND-Scaled model; experiment performed using Waymo Validation set." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 663, + 545, + 712 + ], + "lines": [ + { + "bbox": [ + 309, + 663, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 309, + 663, + 545, + 712 + ], + "type": "table", + "html": "
TechniquemAP/mAPH L1mAP/mAPH L2
Heads match67.5/63.461.1/57.2
BEV & Heads67.8/63.561.6/57.4
Full Pipeline (ours)68.3/64.361.9/58.0
", + "image_path": "da3ded3e077c1aec6d65b3d1488c3da0e4804bbe514f74779043a679a7c71571.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15060" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "Various merging strategies were empirically tested in generating Object-Complete point clouds:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 286, + 171 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 99, + 286, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 286, + 145 + ], + "type": "text", + "content": "1. Geometry: assumes the objects (with boxes) are best aligned by the detected bounding box, performs both inverse translation and rotation to clear their geometric transformations and then merges intact point clouds." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 146, + 286, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 286, + 158 + ], + "type": "text", + "content": "2. GeDi: uses GeDi Point Cloud Registration method [27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 159, + 270, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 159, + 270, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 270, + 171 + ], + "type": "text", + "content": "3. Greedy Grid: uses Greedy Grid implementation [2]." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 173, + 286, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 286, + 198 + ], + "type": "text", + "content": "Object-Complete frames have the potential to become exceedingly large, so we employ a point subsampling strategy." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 111, + 272, + 224, + 324 + ], + "blocks": [ + { + "bbox": [ + 46, + 208, + 287, + 263 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 263 + ], + "type": "text", + "content": "Table 3. Registration methods comparison for Object Complete Frame Generation on the ONCE Small split. We trained our best SECOND model in the semi-supervised setting with X-Ray Proficient Teacher on three different types of preprocessed data and compared results on ONCE validation set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 272, + 224, + 324 + ], + "lines": [ + { + "bbox": [ + 111, + 272, + 224, + 324 + ], + "spans": [ + { + "bbox": [ + 111, + 272, + 224, + 324 + ], + "type": "table", + "html": "
MethodmAP
Box Geometry67.88
Greedy Grid [2]68.17
GeDi [27]68.43
", + "image_path": "57880af228bbee1933748e7bced994f2ceb99dcc79dbd81be7c750f43f620a65.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 48, + 389, + 287, + 553 + ], + "blocks": [ + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 381 + ], + "type": "text", + "content": "Table 4. Comparison for supervised 3D Object Detection task on Waymo Open Dataset. We compare baseline models, their scaled versions and X-Ray distillation with default hyperparameters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 389, + 287, + 553 + ], + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 553 + ], + "type": "table", + "html": "
ModelmAP/mAPH L1mAP/mAPH L2#params
SECOND X-Ray Teacher*85.1/70.375.1/64.75.3m
SECOND67.2/63.161.0/57.25.3m
X-Ray SECOND67.0/62.860.4/56.75.3m
SECOND-Scaled66.8/62.759.4/56.16.2m
X-Ray SECOND-Scaled68.3/64.361.9/58.06.2m
CenterPoint X-Ray Teacher*88.3/78.676.4/72.98.3m
CenterPoint74.4/71.768.2/65.88.3m
X-Ray CenterPoint73.2/69.767.1/64.58.3m
CenterPoint-Scaled74.1/71.567.9/65.39.2m
X-Ray CenterPoint-Scaled75.2 /72.168.9/66.39.2m
DSVT Pillar X-Ray Teacher*89.3/79.779.1/73.48.6m
DSVT Pillar79.5/77.173.2/71.08.6m
X-Ray DSVT Pillar79.2/76.772.6/70.38.6m
DSVT Pillar-Scaled79.6/77.273.3/71.29.5m
X-Ray DSVT Pillar-Scaled80.1/77.973.7/71.49.5m
", + "image_path": "4e692a4c9b11b0e4783969ca607ba285d69a9825ed2fca86901b70bda2af9de6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 574, + 147, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 147, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 147, + 586 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "content": "In this section, we provide the comparison of different components of our approach and show how they affect the performance. Specifically, we prove the usefulness of teacher fine-tuning and compare Point Cloud Registration and distillation methods. We also analyse if dealing with sparsity and incomplete shapes really improves the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "First, we performed a detailed comparison of the semi-supervised performance of our X-Ray method, both with and without teacher fine-tuning using Object-Complete frames. We use Mean Teacher and Proficient Teacher with" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 528, + 221 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 528, + 221 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 528, + 221 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 528, + 221 + ], + "type": "image", + "image_path": "300a4c7e097f7720962997b69d717caac67e81ff3c19730ff0aca0b95d13eb4c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 222, + 405, + 228 + ], + "lines": [ + { + "bbox": [ + 349, + 222, + 405, + 228 + ], + "spans": [ + { + "bbox": [ + 349, + 222, + 405, + 228 + ], + "type": "text", + "content": "Baseline Model Predictions" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 222, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 449, + 222, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 449, + 222, + 504, + 228 + ], + "type": "text", + "content": "X-Ray Teacher Predictions" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 239, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 316 + ], + "type": "text", + "content": "Figure 5. Visual comparison between noisy and poor baseline (original) 3D detector SECOND (left column) and our X-Ray Teacher that perceives Object-Complete frames. We compare two identical timestamps and view angles. The Baseline model fails to detect some objects while X-Ray Teacher does not. This explains why knowledge distillation is indeed beneficial and should improve models." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "text", + "content": "SECOND and CenterPointVoxel models trained on ONCE Small split. Table 1 illustrates the substantial impact of teacher fine-tuning on performance, showing that neglecting this step results in a noticeable performance decline." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 387, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 445 + ], + "type": "text", + "content": "As we noted before, the reconstructed objects will not be perfect in the semi-supervised setting, which is precisely why we included Table 3. PCR models trained on domain objects like cars, pedestrians, cyclists, etc., should improve our method even more." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 447, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 517 + ], + "type": "text", + "content": "As we mentioned earlier, we combine several techniques for the knowledge distillation: BEV features matching, heads output matching with simple regression and KL divergence and detection loss on the teachers predictions. We also compare partial solutions on the Waymo validation set with a SECOND model in Table 2" + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 315, + 570, + 537, + 699 + ], + "blocks": [ + { + "bbox": [ + 304, + 528, + 545, + 561 + ], + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 561 + ], + "type": "text", + "content": "Table 5. Comparison on NuScenes dataset. Our method improves baselines without scaling models due to the fact that NuScene's object-complete frames are less informative compared to Waymo." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 570, + 537, + 699 + ], + "lines": [ + { + "bbox": [ + 315, + 570, + 537, + 699 + ], + "spans": [ + { + "bbox": [ + 315, + 570, + 537, + 699 + ], + "type": "table", + "html": "
ModelmAPNDS
X-Ray Teacher* CBGS77.172.4
CBGS50.059.2
X-Ray CBGS (ours)50.860.4
X-Ray Teacher* CenterPoint-Voxel79.577.1
CenterPoint-Voxel53.461.3
X-Ray CenterPoint-Voxel (ours)54.362.9
X-Ray Teacher* Transfusion-L81.378.2
Transfusion-L56.166.3
X-Ray Transfusion-L (ours)56.866.9
", + "image_path": "473a7ce838ecd983e78f6b2298686e85e02cc77c3929c654555f0fdcbbc5b500.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15061" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 134, + 278, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 126 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 126 + ], + "type": "text", + "content": "Table 6. Performance of X-Ray-powered Mean Teacher and Proficient Teacher methods in the semi-supervised setting using SECOND and CenterPoint baselines, ONCE validation set. Our approach consistently outperforms the state-of-the-art for Semi-Supervised 3D Object Detection in terms of mAP across all splits." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 134, + 278, + 357 + ], + "lines": [ + { + "bbox": [ + 57, + 134, + 278, + 357 + ], + "spans": [ + { + "bbox": [ + 57, + 134, + 278, + 357 + ], + "type": "table", + "html": "
MethodSECONDCenterPoint
Train (5k labeled samples)
Pretraining63.2264.41
Small (5k labeled + 100k unlabeled samples)
Mean Teacher64.3166.47
X-Ray MT (ours)65.26 (+0.95)67.60 (+1.13)
Proficient Teacher67.0667.72
X-Ray PT (ours)68.43 (+1.37)68.76 (+1.04)
Medium (5k labeled + 500k unlabeled samples)
Mean Teacher64.7366.87
X-Ray MT (ours)65.62 (+0.89)67.78 (+0.91)
Proficient Teacher67.4968.54
X-Ray PT (ours)68.75 (+1.26)69.96 (+1.42)
Large (5k labeled + 1M unlabeled samples)
Mean Teacher65.0367.45
X-Ray MT (ours)65.97 (+0.94)68.17 (+0.72)
Proficient Teacher67.8969.68
X-Ray PT (ours)69.10 (+1.21)70.55 (+0.87)
", + "image_path": "f18231aa4da5715a1be7f66cf2b9116c93046494e32ab3d68c7c27dcb9039e39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 377, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 510 + ], + "type": "text", + "content": "We evaluate various Point Cloud Registration (PCR) methods used in the Object Complete Frame Generation process. This analysis, detailed in Table 3, is conducted on a ONCE Small split. The results indicate that superior PCR leads to the creation of less noisy objects, which in turn contributes to improved overall quality. However, it's important to note that methods, such as the Greedy Grid method [2] and GeDi [27], are computationally more expensive. This introduces a trade-off between computational efficiency and the quality of the results, highlighting the need for a balanced approach in the selection of PCR methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 518, + 157, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 157, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 157, + 532 + ], + "type": "text", + "content": "4.4. Model comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 537, + 169, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 169, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 169, + 551 + ], + "type": "text", + "content": "4.4.1 Supervised Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "We perform model comparisons using SECOND [42], CenterPoint-Voxel [47], and DSVT Pillar [36] on the Waymo dataset, and CBGS [50] and CenterPoint-Voxel on the NuScenes dataset. Additionally, we train scaled versions of these models without the X-Ray Teacher to show that the improvements in detection quality are due to the effectiveness of our method, not just an increase in the number of parameters. We scale Waymo students because of the extremely dense and complete point clouds, unlike NuScenes (see supplementary materials), where simpler data requires fewer parameters to learn meaningful feature representations. The results, presented in Tables 4 and 5, demonstrate that our approach consistently outperforms" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 426, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 426, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 426, + 84 + ], + "type": "text", + "content": "baseline models by 1-2 mAP." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 330, + 125, + 525, + 173 + ], + "blocks": [ + { + "bbox": [ + 305, + 94, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 305, + 94, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 94, + 545, + 118 + ], + "type": "text", + "content": "Table 7. Student performance on ONCE validation small split with different tracking methods used in Objects Temporal Fusion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 125, + 525, + 173 + ], + "lines": [ + { + "bbox": [ + 330, + 125, + 525, + 173 + ], + "spans": [ + { + "bbox": [ + 330, + 125, + 525, + 173 + ], + "type": "table", + "html": "
Tracking methodSECONDCenterPoint
Greedy68.4368.76
Kalman Filter + IoU68.5768.91
ReID + Kalman Filter68.7969.12
", + "image_path": "5cc732965e49a234694564c22908bcb4c9cff2b29253ffeca5864204b236204b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 198, + 454, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 198, + 454, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 454, + 211 + ], + "type": "text", + "content": "4.4.2 Semi-Supervised Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 326 + ], + "type": "text", + "content": "To validate the effectiveness of our method in the context of pseudo-label-based semi-supervised learning, we perform a comparative analysis with the Mean Teacher [33] and Proficient Teacher [46] methods, which use the SECOND and CenterPointVoxel models. We compare the results obtained with and without the use of the X-Ray Teacher, as detailed in Table 6. Our results show that the application of our approach consistently improves performance, yielding an improvement of " + }, + { + "bbox": [ + 304, + 217, + 547, + 326 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{-}1.4\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 217, + 547, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 335, + 379, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 379, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 379, + 348 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 355, + 547, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 572 + ], + "type": "text", + "content": "In our research, we proposed the innovative X-Ray Teacher framework, tailored to improve 3D Object Detection models in supervised and semi-supervised settings. Our extensive results have shown that this approach not only achieves state-of-the-art performance in the semi-supervised setting on the ONCE benchmark, but also consistently improves the quality of supervised models on NuScenes and Waymo Open Dataset. The main contributions of our work include the design of the X-Ray Teacher framework, the development of the Objects Temporal Fusion block for generating Object-Complete frames for data lacking ground truth tracking labels, and the demonstration of the potential of our method to improve the performance of any supervised model trained on sequential data. Future work will focus on optimizing the Objects Temporal Fusion block for more complex environments and exploring the integration of our framework with a broader range of model architectures and applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 582, + 421, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 421, + 596 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 421, + 596 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "content": "This research was supported in part through computational resources of HPC facilities at HSE University." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 625, + 545, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 709 + ], + "type": "text", + "content": "The work of I. Makarov on 3D object detection related work was supported by a grant for research centers in the field of artificial intelligence, provided by the Analytical Center in accordance with the subsidy agreement (agreement identifier 000000D730321P5Q0002) and the agreement with the Ivannikov Institute for System Programming of dated November 2, 2021 No. 70-2021-00142." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15062" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 287, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 287, + 188 + ], + "type": "text", + "content": "[2] David Bojanić, Kristijan Bartol, Josep Forest, Stefan Gumhold, Tomislav Petković, and Tomislav Pribanić. Challenging the universal representation of deep models for 3d point cloud registration. In BMVC 2022 Workshop Universal Representations for Computer Vision, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 189, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 189, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 189, + 287, + 219 + ], + "type": "text", + "content": "[3] Daniel Brodeski, Igal Bilik, and Raja Giryes. Deep radar detector. In 2019 IEEE Radar Conference (RadarConf), pages 1-6, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 251 + ], + "type": "text", + "content": "[4] Collin Burns et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. arXiv preprint arXiv:2312.09390, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 316 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 380 + ], + "type": "text", + "content": "[6] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M. Alvarez. Focalformer3d: Focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8394-8405, October 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 381, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 287, + 413 + ], + "type": "text", + "content": "[7] Hyeon Cho, Junyong Choi, Geonwoo Baek, and Wonjun Hwang. itkd: Interchange transfer-based knowledge distillation for 3d object detection. In CVPR, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 466 + ], + "type": "text", + "content": "[8] Xu Dong, Pengluo Wang, Pengyue Zhang, and Langechuan Liu. Probabilistic oriented object detection in automotive radar. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 458-467, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 287, + 499 + ], + "type": "text", + "content": "[9] J. Fang, D. Zhou, F. Yan, T. Zhao, F. Zhang, Y. Ma, L. Wang, and R. Yang. Augmented lidar simulator for autonomous driving. In IEEE Robotics and Automation, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 530 + ], + "type": "text", + "content": "[10] Jeff Dean Geoffrey Hinton, Oriol Vinyals. Distilling the knowledge in a neural network. In arXiv preprint arXiv:1503.02531, 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 531, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 563 + ], + "type": "text", + "content": "[11] Maksim Golyadkin, Alexander Gambashidze, Ildar Nurgaliev, and Ilya Makarov. Refining the once benchmark with hyperparameter tuning. IEEE Access, 12:3805-3814, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 563, + 287, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 287, + 594 + ], + "type": "text", + "content": "[12] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 594, + 287, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 594, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 48, + 594, + 287, + 637 + ], + "type": "text", + "content": "[13] Chengming Hu, Xuan Li, Dan Liu, Haolun Wu, Xi Chen, Ju Wang, and Xue Liu. Teacher-student architecture for knowledge distillation: A survey. In arXiv preprint arXiv:2210.17332, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 638, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 287, + 680 + ], + "type": "text", + "content": "[14] Haotian Hu, Fanyi Wang, Jingwen Su, Yaonong Wang, Laifeng Hu, Weiye Fang, Jingwei Xu, and Zhiwang Zhang. Ea-iss: Edge-aware lift-splat-shot framework for 3d bev object detection. arXiv preprint arXiv:2303.17895, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[15] Ilia Indyk and Ilya Makarov. Monovan: Visual attention for self-supervised monocular depth estimation. In 2023 IEEE International Symposium on Mixed and Augmented Reality" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 546, + 708 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 477, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 477, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 477, + 83 + ], + "type": "text", + "content": "(ISMAR), pages 1211-1220. IEEE, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "text", + "content": "[16] Aleksei Karpov and Ilya Makarov. Exploring efficiency of vision transformers for self-supervised monocular depth estimation. In 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 711-719, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 126, + 546, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 126, + 546, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 546, + 180 + ], + "type": "text", + "content": "[17] Junho Koh, Junhyung Lee, Youngwoo Lee, Jaekyum Kim, and Jun Won Choi. Mgtanet: Encoding sequential lidar points using long short-term motion-guided temporal attention for 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):1179-1187, Jun. 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 180, + 546, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 180, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 546, + 243 + ], + "type": "text", + "content": "[18] Haoang Li, Jinhu Dong, Binghui Wen, Ming Gao, Tianyu Huang, Yun-Hui Liu, and Daniel Cremers. Ddit: Semantic scene completion via deformable deep implicit templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21894-21904, October 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 245, + 546, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 546, + 288 + ], + "type": "text", + "content": "[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 289, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 546, + 330 + ], + "type": "text", + "content": "[20] Ziyu Li, Yuncong Yao, Zhibin Quan, Jin Xie, and Wankou Yang. Spatial information enhancement network for 3d object detection from point cloud. Pattern Recognition, page 108684, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 331, + 546, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 384 + ], + "type": "text", + "content": "[21] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 384, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 546, + 459 + ], + "type": "text", + "content": "[22] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 924–932, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 546, + 525 + ], + "type": "text", + "content": "[23] Ilya Makarov, Vladimir Aliev, and Olga Gerasimova. Semidense depth interpolation using deep convolutional neural networks. In Proceedings of the 25th ACM International Conference on Multimedia, MM '17, page 1407-1415, New York, NY, USA, 2017. Association for Computing Machinery." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 525, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 546, + 578 + ], + "type": "text", + "content": "[24] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 578, + 546, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 546, + 622 + ], + "type": "text", + "content": "[25] Michael Meyer, Georg Kuschk, and Sven Tomforde. Graph convolutional networks for 3d object detection on radar data. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3053-3062, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 622, + 546, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 546, + 686 + ], + "type": "text", + "content": "[26] Mahyar Najibi, Guangda Lai, Abhijit Kundu, Zhichao Lu, Vivek Rathod, Thomas Funkhouser, Caroline Pantofaru, David Ross, Larry S. Davis, and Alireza Fathi. Dops: Learning to detect 3d objects and predict their 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 687, + 546, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 687, + 546, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 687, + 546, + 708 + ], + "type": "text", + "content": "[27] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud regis" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15063" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 708 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "tration. In IEEE Trans. on Pattern Analysis and Machine Intelligence, (early access) 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 147 + ], + "type": "text", + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 191 + ], + "type": "text", + "content": "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 191, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 191, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 191, + 287, + 233 + ], + "type": "text", + "content": "[30] Ilia Semenkov, Aleksei Karpov, Andrey V Savchenko, and Ilya Makarov. Inpainting semantic and depth features to improve visual place recognition in the wild. IEEE Access, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 319 + ], + "type": "text", + "content": "[31] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 429 + ], + "type": "text", + "content": "[32] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 430, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 505 + ], + "type": "text", + "content": "[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 506, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 547 + ], + "type": "text", + "content": "[34] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[35] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J. Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 580, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 643 + ], + "type": "text", + "content": "[36] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, June 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 644, + 287, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 287, + 687 + ], + "type": "text", + "content": "[37] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 688, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 688, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 688, + 287, + 708 + ], + "type": "text", + "content": "[38] Haiyang Wang, Wenguan Wang, Xizhou Zhu, Jifeng Dai, and Liwei Wang. Collaborative visual navigation. arXiv" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 641 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 450, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 450, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 450, + 83 + ], + "type": "text", + "content": "preprint arXiv:2107.01151, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 83, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 545, + 126 + ], + "type": "text", + "content": "[39] Yan Wang, Bin Yang, Rui Hu, Ming Liang, and Raquel Urtasun. Plumenet: Efficient 3d object detection from stereo images. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3383-3390, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 126, + 547, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 126, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 547, + 169 + ], + "type": "text", + "content": "[40] Xinshuo Weng and Kris Kitani. Monocular 3d object detection with pseudo-lidar point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 170, + 545, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 545, + 213 + ], + "type": "text", + "content": "[41] Qiangeng Xu, Yiqi Zhong, and Ulrich Neumann. Behind the curtain: Learning occluded shapes for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2893-2901, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 213, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 213, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 213, + 545, + 233 + ], + "type": "text", + "content": "[42] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "type": "text", + "content": "[43] Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[44] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. In NeurIPS, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "type": "text", + "content": "[45] Lehan Yang and Kele Xu. Cross modality knowledge distillation for multi-modal aerial view object classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 382-387, June 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[46] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In European Conference on Computer Vision, pages 727-743. Springer, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 458, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 458, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 308, + 458, + 545, + 501 + ], + "type": "text", + "content": "[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11087, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In European Conference on Computer Vision, pages 496-513. Springer, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "text", + "content": "[50] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. In arXiv preprint arXiv:1908.09492, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 586, + 545, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 545, + 641 + ], + "type": "text", + "content": "[51] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3357–3364, 2017." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15064" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_content_list.json b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..421b9ef03c0dd3ee5dc16bc93678ad1134302019 --- /dev/null +++ b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_content_list.json @@ -0,0 +1,1445 @@ +[ + { + "type": "text", + "text": "Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection", + "text_level": 1, + "bbox": [ + 124, + 128, + 846, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Chen Jiahao Qi Xingyue Liu Kangcheng Bin Ruigang Fu Xikun Hu Ping Zhong* National University of Defense Technology, China", + "bbox": [ + 86, + 203, + 890, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenchen21c@nudt.edu.cn, qijiahao1996@nudt.edu.cn, liuxingyuel8@nudt.edu.cn, binkc21@nudt.edu.cn, furuigang08@nudt.edu.cn, xikun@nudt.edu.cn, zhongping@nudt.edu.cn", + "bbox": [ + 99, + 241, + 862, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visible-infrared (RGB-IR) image fusion has shown great potentials in object detection based on unmanned aerial vehicles (UAVs). However, the weakly misalignment problem between multimodal image pairs limits its performance in object detection. Most existing methods often ignore the modality gap and emphasize a strict alignment, resulting in an upper bound of alignment quality and an increase of implementation costs. To address these challenges, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA), which could adaptively adjust the relative positions between multimodal features. Considering the impact of modality gap on the cross-modality spatial matching, a Cross-modality Spatial Offset Modeling (CSOM) module is designed to establish a common subspace to estimate the precise feature-level offsets. Then, an Offset-guided Deformable Alignment and Fusion (ODAF) module is utilized to implicitly capture optimal fusion positions for detection task rather than conducting a strict alignment. Comprehensive experiments demonstrate that our method not only achieves state-of-the-art performance in the UAVs-based object detection task but also shows strong robustness to the weakly misalignment problem.", + "bbox": [ + 75, + 342, + 473, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visible-infrared (RGB-IR) object detection on unmanned aerial vehicles (UAVs) has attracted extensive attention [2, 27, 39], due to its ability to make full use of complementary information and achieve robust around-the-clock detection. However, the collected RGB-IR images are typically mismatched in resolutions and field of views owing to disparate imaging principles. Thus, it is challenging to fuse the corresponding information of target between multimodal images.", + "bbox": [ + 75, + 734, + 468, + 872 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52d5000f81f66fa61293d1f6c6fcd59c797f57cead50e577bba722d9cd0f30ae.jpg", + "image_caption": [ + "(a) Image-level alignment method" + ], + "image_footnote": [], + "bbox": [ + 501, + 306, + 696, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9e37ab502cb93b5f1a75269cee6c2b42d1fb070dcc4e98601546f23db00fb74.jpg", + "image_caption": [ + "(b) Feature-level alignment method" + ], + "image_footnote": [ + ":Multimodal Features" + ], + "bbox": [ + 718, + 308, + 890, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/734ae5bdfa091c01da8da7917ea0f75ae7f619750ef3858c94c257b5f71a79a6.jpg", + "image_caption": [ + "(c) Our OFA", + "Figure 1. Previous multimodal object detection architectures under weakly misalignment conditions vs. our OAFA. (a) Image-level alignment method: executing a pixel-level strict alignment by explicitly extracting the global deformation fields from multimodal images. (b) Feature-level alignment method: explicitly estimating feature offsets through multimodal features to achieve strict feature alignment. It can be further classified into global and local methods based on whether alignment is exclusively performed on target regions. (c) Our OAFA: the basic global offsets are estimated by the modality-invariant features to eliminate the impact of modality gap. Based on this offset, we perform adaptive feature alignment without strict alignment." + ], + "image_footnote": [ + "(F) : Fusion Module", + "Modality-specific Features", + "Modality-invariant Features", + "Deformation Field" + ], + "bbox": [ + 514, + 398, + 754, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To increase the quality of fusion representations, researchers always apply pre-registration techniques, involving image cropping [43] and affine transformation [47] before the fusing process. Nevertheless, since distinct imaging time in multimodal sensors and intricate movements in targets, it is difficult, or even impossible, to achieve precise UAVs image alignment [29]. It inevitably leads to local inconsistencies in multimodal spatial distribution, named the weakly misalignment problem [41], resulting in a mismatch of the multimodal target features in corresponding positions [40]. Since the detection task is sensitive to spatial locations, the weakly misalignment problem would lead to significant performance degradation. Several studies have been carried out to deal with this issue by accomplishing", + "bbox": [ + 496, + 689, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 325, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26836", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a strict alignment based on offset estimation, as shown in Fig. 1a and Fig. 1b. Although these methods have made significant progress, there still remain two main challenges:", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inaccurate offset estimation. Conventional methods devote to predicting spatial offsets by identifying the distribution and appearance discrepancies in image or feature level. However, in fact, the discrepancies arise not only from spatial offsets but also from modality gap [31], which stems from the differences in imaging principles and wavelengths between RGB and IR sensors. This gap could make it hard to correctly match cross-modality representations in spatial, leading to misleading offset estimation. Hence, the first challenge is how to bridge the modality gap to achieve accurate cross-modal offset estimation.", + "bbox": [ + 75, + 138, + 468, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "High dependency on strict alignment. The strict alignment is necessary for most existing methods. Unfortunately, the strict alignment always requires additional annotation information, such as the ground truth offsets between multimodal representations [45]. It will lead to higher implementation costs that limit real-world applications. Moreover, the detection performance relies heavily on the quality of alignment. However, even the state-of-the-art alignment methods have the risk of failure in some scenarios [18], which may cause the error delivery to the detection task. Consequently, the second challenge is how to enhance the robustness of the detection model to the weakly misalignment problem without strict alignment.", + "bbox": [ + 75, + 306, + 468, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, the purpose of this paper is to achieve adaptive alignment between RGB (sensed modality) and IR (reference modality) images based on spatial offsets for superior detection performance instead of strict alignment. Technically, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA) for UAVs-based multimodal object detection under weakly misalignment conditions. It mainly consists of a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module for dealing with the two challenges, respectively. Considering the first challenge, CSOM establishes a cross-modal common subspace to obtain the modality-invariant features for eliminating modality gap. As distributional aligned mappings, the modality-invariant features could be utilized to evaluate spatial consistency between multimodal features for more accurate cross-modal offsets estimation.", + "bbox": [ + 75, + 506, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for the second challenge, ODAF is designed to adaptively adjust feature positions by implicit offset compensation, aiming to make full use of the offsets and achieve superior detection performance without strict alignment. In specific, ODAF involves a learnable deformation component that dynamically predicts the offsets of sampling convolution kernels. With the help of the offsets provided by CSOM, ODAF captures optimal sampling positions in sensed modality features to attain comprehensive target rep", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "resentations after fusion. This module predicts the corresponding sampling positions for diverse regions in different image pairs, contributing to a strong flexibility of OAFA in handling diverse and complex offsets. Besides, the training stability of deformable components is improved since the explicit guidance of the offsets reduces the burden of the adaptive alignment.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: (1) We propose a robust RGB-IR object detection method for UAVs images under weakly misalignment conditions, effectively implementing adaptive feature alignment instead of strict alignment. (2) Considering the difficulty of estimating cross-modality offset, we build a multimodal common subspace to reduce the impact of modality gap on multimodal spatial matching. (3) An offset-guided deformable alignment module is exploited to achieve adaptive alignment by implicitly capturing optimal fusion positions, enhancing the accuracy of the detection task. (4) Extensive experiments on public dataset show that our OAFA is superior to state-of-the-art models only with a simple fusion strategy. Moreover, it proves the effectiveness in addressing the weakly misalignment problem.", + "bbox": [ + 496, + 198, + 892, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 436, + 640, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal object detection under weakly misalignment conditions. A general solution to address the weakly misalignment problem is to conduct pixel-level prealignment [24, 46, 52], which learn a global mapping function from multimodal images to correct the pixel coordinates. Many of these methods [3, 16, 26] are supervised with the deformable fields in order to achieve strict prealignment. To achieve end-to-end training, Zhang et al. [45] first introduced a data-driven region feature alignment module to predict proposal regions offsets from multimodal features, which was inexpensively constrained by the target locations in RGB-IR images. On this basis, Yuan et al. [41] made full use of the oriented multimodal annotations to realize multimodal target alignment in position, size, and angle. Due to the requirement of extra supervisions for the strict alignment process, the above methods are labor-intensive and costly to implement. Besides, the existing modality gap at both the pixel-level and feature-level leads to incorrect estimation of spatial offsets. Consequently, it is indeed significant to propose a novel method for RGB-IR object detection under weakly misalignment conditions.", + "bbox": [ + 496, + 462, + 892, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Common subspace learning. Common subspace learning aims to exploit shared information among multiple data. In field of multimodal learning, it can be used to bridge the modality gap by extracting the unified latent representations in the inter-modality samples. Works can be broadly grouped into transformation-based [9, 23, 32, 37], weight-sharing-based [12, 14, 15], and decoupling-based methods [13, 20, 38]. The transformation-based methods", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26837", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a58246ba63d96996c9784b32e94c2344362e96ea3525ababf77e20b3249e5b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e040a3eeb21d4a404706801326bac3d5f93911841f87e41535c7eb0b0c4b4eb4.jpg", + "image_caption": [ + "Figure 2. An Overview of OAFA. It consists of two components: a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module. In our implementation, we choose the RGB images as the sensed modality whose features are shifted to achieve alignment, and the IR images as the reference modality. The CSOM can decouple the original multimodal features to modality-invariant and modality-specific features and predict a superior spatial offset in a common subspace. The ODAF utilizes this basic offset to adaptively align the sample positions of the RGB features with IR features, and then fuses the aligned RGB and original IR features. Finally, the fused features are conducted to generate the ultimate detection results." + ], + "image_footnote": [], + "bbox": [ + 80, + 353, + 885, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generally transform one modality (images/features) into another modality by domain transfer [9] or auto-encoder [23, 32, 37]. It is hard to achieve an end-to-end training due to the requirement of a pre-trained transformation network. The weight-sharing-based methods devote to designing the weight-shared network within a two-stream framework to only capture modality-invariant features [12, 14, 15]. Nevertheless, overemphasis on common features may result in the underutilization of complementary information. The decoupling-based methods overcome this obstacle by trying to separate multimodal features into modality-invariant and modality-specific features with a series of constraint functions [13, 20, 38]. Inspired by the third category methods, we develop a decoupled multimodal learning network with three constraints for the detection task. Moreover, due to the complexity of UAVs-based multimodal features, we design the network with partial weight sharing and custom encoder architecture to achieve effective disentanglement.", + "bbox": [ + 75, + 508, + 472, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformable convolution. Convolutional neural networks (CNNs) are inherently limited in modeling geometric transformations because of their fixed kernel structures. To enhance the transformation modeling capability of CNNs, Dai et al. [8] designed the deformable convolution, which has shown superior performance in several high-level vision tasks [1, 28, 51]. In recent years, it has also been", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "used in the field of video super-resolution [5, 30, 33, 34]. As a pioneer work, Tian et al. [30] adopted the deformable convolution to align the features of adjacent frames in the temporal domain. It demonstrates that deformable convolution can effectively capture the motion cues of the targets. Nonetheless, Chan et al. [4] has proved that the deformable alignment is difficult to train. Training instability frequently leads to offset overflow, degrading the final performance in downstream task. Therefore, we construct a basic offset as explicit guidance to reduce the burden of deformable alignment training.", + "bbox": [ + 496, + 508, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 686, + 589, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 710, + 609, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method aims to enhance the accuracy and robustness of the RGB-IR object detection model under weakly misalignment conditions. As shown in Fig. 2, a pair of weakly misaligned RGB-IR images are fed into a two-stream network to acquire the multiscale multimodal representations. Subsequently, a decoupled learning network is designed to develop a common subspace to figure out the modality-invariant and modality-specific features. To mitigate the influence of modality gap on offset estimation, we exclusively employ the modality-invariant features to predict spatial offsets. Then, this basic offsets are utilized to implica", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26838", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity adjust the spatial positions of the modality-invariant and modality-specific features. It accomplishes adaptive feature alignment that is tailored for RGB-IR object detection task without any extra supervisions. Finally, the aligned RGB and original IR features are fused with a concatenation operation to generate final detection results.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cross-modality Spatial Offset Modeling Module", + "text_level": 1, + "bbox": [ + 76, + 188, + 468, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the RGB features $F_{rgb}$ and IR features $F_{ir}$ , CSOM predicts the spatial offsets $\\phi_c$ as basic knowledge for the subsequent alignment module. The CSOM consists of a decoupled multimodal learning submodule and a spatial offset modeling submodule.", + "bbox": [ + 75, + 227, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoupled multimodal learning. This submodule mainly builds a common subspace $\\tilde{\\mathbb{C}} = [\\mathbb{C}_{rgb},\\mathbb{C}_{ir}]$ and a specific subspace $\\tilde{\\mathbb{S}} = [\\mathbb{S}_{rgb},\\mathbb{S}_{ir}]$ , as shown in Fig. 2a. For each input RGB-IR image pair $\\mathcal{I}_{rgb}\\in \\mathbb{R}^{H\\times W\\times 3}$ and $\\mathcal{I}_{ir}\\in \\mathbb{R}^{H\\times W\\times 1}$ , we map it to the multimodal representations $\\mathcal{F}_m(m\\in \\{rgb,ir\\})$ by a basic pyramid feature extractor $\\mathcal{B}_m$ . The $\\mathcal{F}_m$ is the bounded random samples on the interval $[a,b]^N$ . Then, the modality-invariant features $\\mathcal{F}_m^c$ and modality-specific features $\\mathcal{F}_m^s$ projected from $\\tilde{\\mathbb{C}}$ and $\\mathbb{S}$ are figured out by the encoding functions:", + "bbox": [ + 75, + 303, + 470, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {m} ^ {s} = \\mathcal {S} _ {m} \\left(\\mathcal {F} _ {m}; \\theta_ {m} ^ {s}\\right), \\mathcal {F} _ {m} ^ {c} = \\mathcal {C} _ {m} \\left(\\mathcal {F} _ {m}; \\theta^ {c}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 462, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the invariant feature encoder $\\mathcal{C}_m$ shares the parameters $\\theta^c$ across two modalities, whereas the specific feature encoder $\\mathcal{S}_m$ assigns separate parameters $\\theta_m^s$ for each modality. Considering the diversity information in UAVs images, we not only design the decoupled network in objective function but also in network architecture.", + "bbox": [ + 75, + 486, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In terms of the network architecture, an accepted assumption is that the input features are more related in the low layers, while the high-layer features tend to be uncorrelated [48]. Thus, $\\mathcal{C}_m$ is constructed with a convolutional layer followed by the SiLU activation function. In contrast, $S_{m}$ contains a convolutional layer and a C3 module with three convolutional and a series of bottleneck layers to obtain deeper features.", + "bbox": [ + 75, + 577, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for the objective functions, inspired by [13], we apply three types of constraints for decoupled multimodal learning: Similarity Loss, Difference Loss, and Semantic Loss. First, we employ a similarity loss $\\mathcal{L}_{sim}$ to minimize the distance between $\\mathcal{F}_{rgb}^{c}$ and $\\mathcal{F}_{ir}^{c}$ to learn common subspace $\\tilde{\\mathbb{C}}$ . The $\\mathcal{L}_{sim}$ is defined by a state-of-the-art distance metric named central moment discrepancy [42] to measure the distribution dissimilarity between two features:", + "bbox": [ + 75, + 698, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s i m} = \\frac {1}{| b - a |} \\left\\| E \\left(\\mathcal {F} _ {r g b} ^ {c}\\right) - E \\left(\\mathcal {F} _ {i r} ^ {c}\\right) \\right\\| _ {2} \\\\ + \\sum_ {k = 2} ^ {K} \\frac {1}{| b - a | ^ {k}} \\left\\| C _ {K} \\left(\\mathcal {F} _ {r g b} ^ {c}\\right) - C _ {K} \\left(\\mathcal {F} _ {i r} ^ {c}\\right) \\right\\| _ {2}, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 827, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E(\\cdot)$ is the empirical expectation vector, $C_K(\\cdot)$ is the vector of $k$ -th order sample central moments, and $\\| \\cdot \\| _2$ represents the Euclidean norm.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we present a difference loss $\\mathcal{L}_{dif}$ for highlighting the dissimilarity between $\\mathcal{F}_{rgb}^{s}$ and $\\mathcal{F}_{ir}^{s}$ to capture complementary information. Additionally, as $\\mathcal{F}_m^c$ and $\\mathcal{F}_m^s$ are expected to be mutually independent, $\\mathcal{L}_{dif}$ should be extended to them as well. These discrepancies are achieved by the squared Frobenius norm:", + "bbox": [ + 496, + 136, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d i f} = \\sum_ {m \\in \\{r g b, i r \\}} \\sqrt {\\operatorname {T r} \\left(\\left(\\left(F _ {m} ^ {s}\\right) ^ {T} F _ {m} ^ {c}\\right) ^ {T} \\left(\\left(F _ {m} ^ {s}\\right) ^ {T} F _ {m} ^ {c}\\right)\\right)} \\tag {3} \\\\ + \\sqrt {T r ((\\left(F _ {r g b} ^ {s}\\right) ^ {T} F _ {i r} ^ {s}) ^ {T} (\\left(F _ {r g b} ^ {s}\\right) ^ {T} F _ {i r} ^ {s}))}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 234, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Tr(\\cdot)$ is trace of the matrix.", + "bbox": [ + 500, + 306, + 728, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Third, only restrained by the above objectives, there exists a potential risk that $\\mathcal{C}_m$ learns useless noisy cues, whereas $S_m$ acquires almost all modality information. To address this issue, we add a semantic loss $\\mathcal{L}_{sem}$ to enforce $\\mathcal{C}_m$ to learn discriminative features:", + "bbox": [ + 496, + 323, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} = \\sum_ {m \\in \\{r g b, i r \\}} \\left(\\mathcal {L} _ {c l s} \\left(\\mathcal {F} _ {m} ^ {c}\\right) + \\mathcal {L} _ {r e g} \\left(\\mathcal {F} _ {m} ^ {c}\\right) + \\mathcal {L} _ {o b j} \\left(\\mathcal {F} _ {m} ^ {c}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 404, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{cls}$ , $\\mathcal{L}_{reg}$ , and $\\mathcal{L}_{obj}$ are the detection loss functions of the YOLOv5 [17]. It should be noted that this object detection branch, serving as an auxiliary task, is not incorporated during the inference stage to enhance the efficiency of our OAFA. Finally, the total loss of the decoupled multimodal learning submodule can be defined as:", + "bbox": [ + 496, + 452, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e c} = \\lambda_ {1} \\mathcal {L} _ {s i m} + \\lambda_ {2} \\mathcal {L} _ {s i m} + \\lambda_ {3} \\mathcal {L} _ {s e m}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 551, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{i}$ is the trade-off weights. In this study, we set $\\lambda_{1}$ , $\\lambda_{2}$ , and $\\lambda_{3}$ to 0.03, 0.01, and 0.03 respectively.", + "bbox": [ + 496, + 575, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial offset modeling. We model the spatial offsets $\\phi_c$ by the estimating spatial differences between $\\mathcal{F}_{rgb}^c$ and $\\mathcal{F}_{ir}^c$ . In the field of change detection [49], the differences are calculated directly by subtracting or concatenating two features. However, it would introduce significant noises when dealing with intricate backgrounds in UAVs images. To solve this problem, a feature difference enhancement module is presented to filter out irrelevant changes and concentrate on shift objects. As shown in Fig. 2b, $\\mathcal{F}_m^c$ is firstly fed into a spatial attention module to capture the representations of the target and crucial background areas in each modality. The structure of the spatial attention module is the same as [35]. In this case, the spatially enhanced unimodal feature $F_m^{c,p}$ can be computed as:", + "bbox": [ + 496, + 604, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {m} ^ {c, p} = \\sigma \\left(f _ {7 \\times 7} \\left(\\operatorname {C a t} \\left(\\operatorname {M a x} \\left(\\mathcal {F} _ {m} ^ {c}\\right), \\operatorname {M e a n} \\left(\\mathcal {F} _ {m} ^ {c}\\right)\\right)\\right)\\right) \\odot \\mathcal {F} _ {m} ^ {c}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 823, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid function, $f_{i\\times j}$ denotes the $i\\times j$ convolution layer, $Cat(\\cdot)$ refers to the concatenation operation, and $\\odot$ indicates the Hadamard product operation.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26839", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct a concatenation instead of subtraction strategy to extract the changing features $F_{dif}$ to mitigate the influence of complex background. Then, a channel difference enhanced module is introduced to further suppress channel noise. Specifically, we employ simultaneous average-pooling and max-pooling to generate distinct spatial context descriptors. Following this, channel attention can be derived by transmitting the spatial context descriptors to a shared multilayer perceptron network $f_{mlp}$ . $\\mathcal{F}_m^{c,p}$ are combined with the channel attention map to get the spatial-channel enhanced features $\\mathcal{F}_m^{c,e}$ . The above operation is represented as:", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} \\mathcal {F} _ {m} ^ {c, e} = \\sigma (f _ {m l p} (A v g P o o l (\\mathcal {F} _ {d i f})) + \\\\ f _ {m l p} (M a x P o o l (\\mathcal {F} _ {d i f}))) \\times \\mathcal {F} _ {m} ^ {c, p} + \\mathcal {F} _ {m} ^ {c, p}. \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 282, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, $\\phi_c$ is obtained through a fusion of the $\\mathcal{F}_{rgb}^{c,e}$ and $\\mathcal{F}_{ir}^{c,e}$ along with a nonlinear layer.", + "bbox": [ + 76, + 330, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Offset-guided Deformable Alignment and Fusion Module", + "text_level": 1, + "bbox": [ + 76, + 372, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have obtained the offsets $\\phi_c$ from CSOM, which represents the spatial differences between $\\mathcal{F}_{rgb}$ and $\\mathcal{F}_{ir}$ . Inspired by [5], the $\\phi_c$ can be an effectively initialized offset to unsupervised alignment task. Hence, ODAF is designed to adaptively capture optimal fusion location with the help of the basic knowledge $\\phi_c$ , incorporating desired complementary information for the subsequent detection task. It is composed of two submodules: offset-guided deformable alignment and decoupled feature fusion.", + "bbox": [ + 75, + 411, + 468, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Offset-guided deformable alignment. We mainly apply a deformable convolution to achieve implicit offset compensation and adaptive alignment for the detection task. Unlike the default deformable convolution that learns offsets of convolution kernels from their original features, the offsets generated from our module are mapped from the basic offsets $\\phi_c$ , which allows for more effective feature correction and reliable model training. Considering different contributions from each offset region for the detection task, we add a modulation scalar $\\Delta m_k$ that is learned from $\\phi_c$ to dynamically aggregate information around the corresponding position $p$ . As shown in Fig. 2c, given the center sampling value $x(p)$ in the sensed features $\\mathcal{F}_{rgb}^n (n\\in \\{s,c\\})$ and the basic offsets $\\phi_c$ , each counterpart $y(p)$ in aligned features $\\mathcal{F}_{rgb}^{n,a}$ can be obtained as follows:", + "bbox": [ + 75, + 547, + 470, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny (p) = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\cdot x \\left(p + \\phi_ {c} + p _ {k} + \\Delta p _ {k}\\right) \\cdot \\Delta m _ {k}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 787, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K$ , $p_k$ , and $w_k$ denote the number of kernel weights, the $k$ -th fixed offset, and the $k$ -th kernel weight, respectively. For instance, $p_k \\in \\{(-1,1), (-1,0), \\dots, (1,1)\\}$ donates a regular grid of a $3 \\times 3$ kernel defined with $K = 9$ .", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\Delta p_{k}$ denotes the implicit compensation to $\\phi_c$ of the $k$ -th location. It is estimated with $\\operatorname{conv}(x)[p + \\phi_c]$ where $\\phi_c$ can be regarded as a prior of $\\Delta p_{k}$ in the standard deformable convolution. Besides, since the location $p + \\phi_c + p_k + \\Delta p_k$ may be fractional, we adopt bilinear interpolation to obtain the final sampling values.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoupled feature fusion. In traditional fusion process, $\\mathcal{F}_m^c$ and $\\mathcal{F}_m^s$ are directly concatenated to obtain the fused features $\\mathcal{F}_f$ [6]. Unfortunately, the similarity constraint between $\\mathcal{F}_{rgb}^{c,a}$ and $\\mathcal{F}_{ir}^{c}$ leads to much redundancy. To eliminate duplicate information and increase discriminative representations, $\\mathcal{F}_{rgb}^{c,a}$ and $\\mathcal{F}_{ir}^{c}$ should be combined and optimized before fusion. Thus, the fused features $F_{f}$ are defined as:", + "bbox": [ + 496, + 181, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {f} = C a t \\left(f _ {1 \\times 1} \\left(C a t \\left(\\mathcal {F} _ {r g b} ^ {c, a}, \\mathcal {F} _ {i r} ^ {c}\\right)\\right), \\mathcal {F} _ {r g b} ^ {s, a}, \\mathcal {F} _ {i r} ^ {s}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 297, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 500, + 324, + 656, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training loss. The overall method is based on a supervised detection framework. The purpose of our multimodal alignment is not to form a strict alignment but to achieve optimal adjustment between multimodal features to improve the performance of the subsequent detection task. Thus, the alignment subtask is integrated into the main detection task instead of adding extra alignment loss. The total loss function of our OAFA can be represented as follows:", + "bbox": [ + 496, + 348, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {d e c}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 479, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-stage training strategy. A premise for effective feature alignment is that $\\mathcal{F}_m^c$ contains sufficient spatial information, thereby estimating more accurate basic offsets. To this end, we adopt a two-stage learning strategy to train our OAFA end-to-end. In stage I, to gain effective $\\mathcal{F}_m^c$ , we try to ensure that pre-decoupling features $\\mathcal{F}_m$ contain more comprehensive target representations. Technically, we only train the feature extraction and fusion module with $\\mathcal{L}_{cls}$ , $\\mathcal{L}_{reg}$ , and $\\mathcal{L}_{obj}$ . Specifically, a pair of $\\mathcal{I}_{rgb}$ and $\\mathcal{I}_{ir}$ are fed into a two-stream network to extract multiscale features $\\mathcal{F}_m$ and gain fusion features $\\mathcal{F}_f$ with the concatenation operation. Then, the $\\mathcal{F}_f$ is put into the detection network to obtain the detection results. In stage II, we apply all the proposed modules to our network. In this process, the unimodal feature extraction network $\\mathcal{B}_m$ is initialized with the weights from stage I. Then, the joint optimization of the unimodal feature extraction and multimodal feature alignment network is carried out effectively with $\\mathcal{L}_{total}$ .", + "bbox": [ + 496, + 507, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 792, + 633, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 816, + 705, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. We evaluate our method on DroneVehicle dataset [29], which is a large-scale multimodal dataset for vehicle detection in UAVs images. It contains 28,439 image pairs with 953,087 vehicles, covering", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26840", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd41082a33f17763e39345eb3181f5ac4ff14cc73f3dc3e9d63feb82d70709a4.jpg", + "image_caption": [ + "Figure 3. Four Examples of detection results on the test set of DroneVehicle dataset under weakly misalignment conditions. The confidence threshold is set to 0.25. The results of fusion methods are visual in IR images to correspond to the supervisory label. Red, green, and pink rectangles represent car, freight-car, and truck targets, respectively. Objects correctly detected are represented in green dashed circles, while incorrectly detected objects are represented in red dashed circles." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 892, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "different viewing angles, heights, lighting conditions, and scenarios. The dense-oriented annotations are made for five categories: car, truck, bus, van, and freight-car. Though the DroneVehicle performs pre-registration on each RGB-IR image pair, the weakly misalignment problem still exists. As shown in Fig. 4, about $35\\%$ bounding boxes exhibit deviation issues in DroneVehicle dataset, and the shift distance mostly ranges from 0 to 15 pixels.", + "bbox": [ + 75, + 510, + 470, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This dataset is divided into 17,990 image pairs for training, 1,469 image pairs for validating, and the remaining 8,980 image pairs for testing. Following the standard evaluation metrics, we report the mean average precision (mAP) with intersection over union (IoU) threshold of 0.4 on the validation set.", + "bbox": [ + 75, + 637, + 468, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. The experiments are carried out on a single NVIDIA RTX A6000 GPU with 48 GB of memory. We implement our algorithm with Pytorch toolbox and SGD optimizer with a momentum of 0.937 and a weight decay of 0.0005. The initial learning rate is set to 0.01 and eventually reduced to 0.002 by cosine annealing [22]. The input size of our model is set to $640 \\times 640$ in the preprocessing stage and the batch size is 16. The training epoch is set to 150 with 50 and 100 epochs in the first and second stages, respectively. We use the ground truth of the IR images as the training label, owning to the more comprehensive target", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/01aa374e69d58cd5caafd5adafb1d331df9e6bc81dc5ebb2a123a52f0f4137bc.jpg", + "image_caption": [ + "(a) Statistics of misalignment ratio" + ], + "image_footnote": [], + "bbox": [ + 511, + 508, + 681, + 625 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db56953588ad45889834f1bad95b0f82c6270f50f4d8a93064b93921e6e34bf8.jpg", + "image_caption": [ + "(b) Statistics of shift value", + "Figure 4. The statistics of ground truth bounding-boxes misalignment ratio and shift value within RGB-IR image pairs on DroneVehicle dataset. annotations available in the IR modality." + ], + "image_footnote": [], + "bbox": [ + 683, + 521, + 895, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results Comparisons", + "text_level": 1, + "bbox": [ + 500, + 710, + 697, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare our method with the state-of-the-art unimodal and multimodal object detection methods. Unimodal methods: We choose a series of oriented object detection methods for comparison, including the one-stage methods: RetinaNet [19], $\\mathbf{S}^2\\mathbf{A}$ Net [11], and YOLOv5s [17]; and the two-stage methods: Faster R-CNN [25], Oriented R-CNN [36], and RoITransformer [10]. RGB and IR images are separately used as training data to evaluate the detection performance of each algorithm. Multimodal fusion methods: We also compare our method with eight recent fusion methods:", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8ae8b7be05ca3ff6a760f3291a40a37fb8f298f615befec69c4044d7b72c2f3.jpg", + "table_caption": [ + "Table 1. Detection results (mAP, in %) on DroneVehicle dataset. Note that all detectors locate and classify vehicles with OBB heads. The speed refers to the speed of network inference without post-processing (batch size = 1). In speed, our method is only compared with the fusion methods. Best results are highlighted in bold. And the second one is marked with underline." + ], + "table_footnote": [], + "table_body": "
DetectorsModalityCarTruckFreight-carBusVanmAP (%) ↑Speed (fps) ↑
RetinaNet [19]78.534.424.169.828.847.1-
Faster R-CNN [25]79.049.037.277.037.055.9-
Oriented R-CNN [36]RGB80.153.841.685.443.360.8-
S²A-Net [11]80.054.242.284.943.861.0-
RoITransformer [10]61.655.142.385.544.861.6-
YOLOv5s [17]78.655.343.887.146.062.1-
RetinaNet [19]88.835.439.576.532.154.5-
Faster R-CNN [25]89.453.548.387.042.664.2-
Oriented R-CNN [36]IR89.857.453.189.345.467.0-
S²A-Net [11]89.954.555.888.948.467.5-
RoITransformer [10]90.160.458.989.752.270.3-
YOLOv5s [17]90.059.560.889.553.870.7-
UA-CMDet [29]87.560.746.887.138.064.09.1
Halfway Fusion [21]90.162.358.589.149.870.020.4
CIAN [44]90.163.860.789.150.370.821.7
AR-CNN [45]90.164.862.189.451.571.618.2
MBNet [50]RGB+IR90.164.462.488.853.671.921.7
TSFADet [41]89.967.963.789.854.073.118.6
C²Former [40]90.268.364.489.858.574.2-
SLBAF-Net [7]90.272.068.689.959.976.163.2
Ours90.376.873.390.366.079.433.1
", + "bbox": [ + 80, + 133, + 888, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UA-CMDet [29], Halfway Fusion [21], CIAN [44], MBNet [50], AR-CNN [45], TSFADet [41], $C^2$ Former [40], and SLBAF-Net [7]. Among them, UA-CMDet, MBNet, TSFADet, and $C^2$ Former all consider the weakly misalignment problem during the fusion process, which affirms the universality and significance of this problem. For each compared method, we use their original experimental settings to ensure equity.", + "bbox": [ + 75, + 455, + 468, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. The results are shown in Tab. 1. In the unimodal methods, benefitting from a well-designed multiscale feature extraction network, YOLOv5s has comparable detection accuracy (62.1% in RGB images and 70.7% in IR images), even better than some fusion methods. Correspondingly, the multimodal methods SLBAF-Net, constructed upon the two-stream YOLOv5 framework, has also yielded excellent results. Compared with all methods, we can find that our OAFA achieves the best average performance on the validation set, which is 3.3% higher than the second place. It is worth noting that our method has obvious advantages in distinguishing confusing categories (truck and freight-car). The reason may stem from the fact that OAFA excels in identifying the optimal spatial correspondences of multimodal features, obtaining reliable features of the target after fusion.", + "bbox": [ + 75, + 580, + 468, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Our baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branches are fused with a concatenation operation in multiscale features. Among the compared methods mentioned above, we select the optimal unimodal", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YOLOv5s and the baseline model as the qualitative comparative model. We provide visual detection results of the compared methods in Fig. 3. It can be observed that, given the constraint of the IR labels, the objects are successfully detected by OAFA when the target positions are shifted, while both unimodal and baseline model experience classification and localization errors or fail to detect, especially for occluded and few-shot object detection. The underlying causes can be speculated that the weak misalignment problem confuses the fusing model to learn discriminative features. In contrast, our method has the capability to capture aligned and reliable features, improving the performance in classification and localization processes.", + "bbox": [ + 496, + 455, + 890, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speed comparison. We compare the speed of OAFA with other fusion detection methods to prove the efficiency of our method. For a fair comparison, all the detection models are tested with batch size 1, and the input size is set to $640 \\times 640$ . As shown in Tab. 1, our proposed OAFA achieves a speed of 33.1 FPS, outperforming most existing multimodal object detection methods. Although our method requires more computational time compared to SLBAF-Net, it would be worthwhile to invest this additional time to achieve superior performance. Furthermore, it can be seen that OAFA can accomplish real-time detection on the NVIDIA RTX A6000 platform.", + "bbox": [ + 496, + 654, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 845, + 653, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct several ablation studies to validate the effectiveness of the different components in our", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26842", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proposed method. The results are shown in Tab. 2.", + "bbox": [ + 76, + 90, + 410, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of CSOM module. We first analyze the effect of the CSOM module as it provides the fundamental offset estimation, which determines the stability of the model training. We perform ablation experiments on both the decoupled multimodal learning (DML) module and the spatial offset modeling (SOM) module separately. As presented in the first to third rows of Tab. 2, the performance of the baseline model is improved by $0.4\\%$ and $1.6\\%$ after incorporating the SOM and DML modules, respectively. It verifies that both coarse-grained feature alignment and increasing fusion flexibility through decouple learning are effective for the detection task. Besides, when integrating these two submodules, the performance improvement of model over baseline is $2.1\\%$ , which indicates that the quality of the feature alignment can be enhanced by mitigating modal gap.", + "bbox": [ + 75, + 107, + 467, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of ODAF module. We then verify the necessity of the ODAF module in addressing the weak misalignment problem. The results demonstrate that despite ODAF has poor stability in model training without dependable offset guidance, incorporating ODAF into the baseline can provide $+1.7\\%$ significant improvement. It is owing to adaptively adjust the optimal sampling position for acquiring reliable information in RGB images.", + "bbox": [ + 75, + 334, + 467, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of two-stage training. The two-stage training strategy provides comprehensive unimodal representations for multimodal alignment and fusion. We can observe that the accuracy decreases to $77.3\\%$ if we directly train the whole network without two-stage training. It is $0.9\\%$ and $0.5\\%$ lower than exclusively incorporating CSOM and ODAF, respectively. The reason could be that the inappropriate initialization in feature extraction network leads to training instability and makes it hard to obtain abundant representations in spatial and semantic. After undergoing a two-stage training process, our method attains a performance level of $79.4\\%$ , surpassing the capabilities of any individual module. It demonstrates that the two-stage training strategy effectively enhances training accuracy and stability. The superior method to fully exploit the strengths of CSOM and ODAF will be further studied in future work.", + "bbox": [ + 75, + 455, + 467, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All in all, not only each independent module but also the entire proposed framework can contribute to the multimodal object detection under weakly misalignment conditions.", + "bbox": [ + 75, + 698, + 467, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Robustness to Position Shift", + "text_level": 1, + "bbox": [ + 76, + 755, + 326, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To quantitatively assess the robustness of our method to the weakly misalignment issue, we conduct experiments on DroneVehicle by manually simulating positional offsets. The tested model is trained on the original training image pairs and validated on the shifted image pairs. In our experiments, we hold the IR images constant while introducing spatial offsets along the $x$ -axis and $y$ -axis for RGB images. The variations in pixel values are defined within the", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/643e614f8ae8852afac8c1958f64b0ffda1286d729845fe2d7669887296c9180.jpg", + "table_caption": [ + "Table 2. Ablation study on DroneVehicle dataset. The baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branch are fused with a multiscale concatenation operation. The D denotes DML, S denotes SOM, C denotes CSOM, and O denotes ODAF." + ], + "table_footnote": [], + "table_body": "
MethodCSOMODAFTwo-stage TrainingmAP (%) ↑
DMLSOM
Baseline76.1
Baseline+S✓76.5
Baseline+D✓77.7
Baseline+C✓✓78.2
Baseline+O✓77.8
Baseline+C+O✓✓✓77.3
OAFA✓✓✓✓79.4
", + "bbox": [ + 504, + 162, + 888, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67de58df0de66f1a6722fb231b8b6fbc1826868a1f3cfad115393880f9a7b0f2.jpg", + "image_caption": [ + "(a) SLBAF-Net (Baseline)" + ], + "image_footnote": [], + "bbox": [ + 509, + 282, + 686, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fbc57013de2708b7741397311d829d0152a1d067604b295a0748ffd272f0167d.jpg", + "image_caption": [ + "(b) Our method", + "Figure 5. Surface plot visualization of the position shift experiments. Horizontal coordinates indicate various shift steps along the $x$ -axis and $y$ -axis. Vertical coordinate denotes the detection performances in terms of mAP." + ], + "image_footnote": [], + "bbox": [ + 705, + 282, + 883, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "range $\\{(\\Delta x, \\Delta y) \\mid \\Delta x, \\Delta y \\in [15, -15]; \\Delta x, \\Delta y \\in \\mathbb{Z}\\}$ . The quantitative outcomes of this experiment in baseline and our model are discernible from Fig. 5. It can be seen that in response to larger position shifts, our method consistently exhibits minor fluctuations in performance, whereas the baseline model experiences a noticeable performance decline. This result demonstrates the robustness of OAFA under weakly misalignment conditions.", + "bbox": [ + 496, + 465, + 890, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 602, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a robust RGB-IR object detection method OAFA for UAVs images under weakly misalignment conditions. Different from previous methods, OAFA can achieve adaptive multimodal feature alignment without strict alignment. To mitigate the impact of the modality gap on multimodal spatial matching, the modality-invariant features are acquired in a common subspace to estimate accurate offsets. Taking the offsets as basic knowledge, the adaptive alignment implicitly captures optimal fusion positions and steadily carries out feature-level corrections with a deformable alignment module. Experimental results on DroneVehicle dataset demonstrate the superior performances of OAFA compared with the state-of-the-art methods. Moreover, our method shows robustness against the weakly misalignment problem in position shift experiments.", + "bbox": [ + 496, + 628, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586).", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26843", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gedas Bertasius, Lorenzo Torresani, and Jianbo Shi. Object detection in video with spatiotemporal sampling networks. In ECCV, pages 342-357, 2018. 3", + "[2] Ilker Bozcan and Erdal Kayacan. AU-AIR: A multi-modal unmanned aerial vehicle dataset for low altitude traffic surveillance. In ICRA, pages 8504-8510, 2020. 1", + "[3] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Dong Nie, Minjeong Kim, Qian Wang, and Dinggang Shen. Deformable image registration based on similarity-steered CNN regression. In MICCAI, pages 300-308, 2017. 2", + "[4] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In AAAI, pages 973-981, 2021. 3", + "[5] Kelvin C. K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5962-5971, 2022. 3, 5", + "[6] Hao Chen, Yongjian Deng, Youfu Li, Tzu-Yi Hung, and Guosheng Lin. RGBD salient object detection via disentangled cross-modal fusion. IEEE TIP, 29:8407-8416, 2020. 5", + "[7] Xiaolong Cheng, Keke Geng, Ziwei Wang, Jinhu Wang, Yuxiao Sun, and Pengbo Ding. Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment. Multimedia Tools and Applications, pages 1-20, 2023. 7", + "[8] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017. 3", + "[9] Chaitanya Devaguptapu, Ninad Akolekar, Manuj M. Sharma, and Vineeth N. Balasubramanian. Borrow from anywhere: Pseudo multi-modal object detection in thermal imagery. In CVPR, pages 1029-1038, 2019. 2, 3", + "[10] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In CVPR, pages 2849-2858, 2019. 6, 7", + "[11] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE TGRS, 60: 1-11, 2022. 6, 7", + "[12] Renlong Hang, Zhu Li, Pedram Ghamisi, Danfeng Hong, Guiyu Xia, and Qingshan Liu. Classification of hyperspectral and lidar data using coupled cnns. IEEE TGRS, 58(7): 4939-4950, 2020. 2, 3", + "[13] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. MISA: modality-invariant and -specific representations for multimodal sentiment analysis. In ACM MM, pages 1122-1131, 2020. 2, 3, 4", + "[14] Danfeng Hong, Naoto Yokoya, Jocelyn Chanussot, and Xiao Xiang Zhu. Cospace: Common subspace learning from hyperspectral-multispectral correspondences. IEEE TGRS, 57(7):4349-4359, 2019. 2, 3", + "[15] Danfeng Hong, Jocelyn Chanussot, Naoto Yokoya, Jian Kang, and Xiao Xiang Zhu. Learning-shared cross-modality representation using multispectral-lidar and hyperspectral data. IEEE GRSL, 17(8):1470-1474, 2020. 2, 3", + "[16] Lloyd Haydn Hughes, Diego Marcos, Sylvain Lobry, Devis Tuia, and Michael Schmitt. A deep learning framework for" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matching of sar and optical imagery. ISPRS, 169:166-179, 2020. 2", + "[17] Glenn Jocher. ultralytics/yolov5. https://github.com/ultralytics/yolov5, Oct. 2020. 4, 6, 7", + "[18] Huafeng Li, Junzhi Zhao, Jinxing Li, Zhengtao Yu, and Guangming Lu. Feature dynamic alignment and refinement for infrared-visible image fusion: Translation robust fusion. Information Fusion, 95:26-41, 2023. 2", + "[19] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 6, 7", + "[20] Fan Liu, Zhiyong Cheng, Huilin Chen, Anan Liu, Liqiang Nie, and Mohan S. Kankanhalli. Disentangled multimodal representation learning for recommendation. arXiv:2203.05406, 2022. 2, 3", + "[21] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Multispectral deep neural networks for pedestrian detection. In BMVC, 2016. 7", + "[22] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6", + "[23] Sijie Mai, Haifeng Hu, and Songlong Xing. Modality to modality translation: An adversarial representation learning and graph fusion network for multimodal fusion. In AAAI, pages 164-172, 2020. 2, 3", + "[24] Sayan Nag. Image registration techniques: A survey. arXiv:1712.07540, 2017. 2", + "[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2017. 6, 7", + "[26] Marc-Michel Rohé, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: Learning deformable image registration using shape matching. In MIC-CAI, pages 266–274, 2017. 2", + "[27] Hazim Shakhatreh, Ahmad Sawalmeh, Ala I. Al-Fuqaha, Zuochao Dou, Eyad Almaita, Issa Khalil, Noor Shamsiah Othman, Abdallah Khreishah, and Mohsen Guizani. Unmanned aerial vehicles (uavs): A survey on civil applications and key research challenges. IEEE Access, 7:48572-48634, 2019. 1", + "[28] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 536-553, 2018. 3", + "[29] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning. IEEE TCSVT, 32(10):6700-6713, 2022. 1, 5, 7", + "[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally-deformable alignment network for video super-resolution. In CVPR, pages 3357-3366, 2020. 3", + "[31] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In IJCAI, pages 3508-3515, 2022. 2", + "[32] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In ICCV, pages 3622–3631, 2019. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26844", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Hua Wang, Dewei Su, Chuangchuang Liu, Longcun Jin, Xianfang Sun, and Xinyi Peng. Deformable non-local network for video super-resolution. IEEE Access, 7:177734-177744, 2019. 3", + "[34] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. In CVPR, pages 1954-1963, 2019. 3", + "[35] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In ECCV, pages 3–19, 2018. 4", + "[36] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented R-CNN for object detection. In ICCV, pages 3500-3509, 2021. 6, 7", + "[37] Han Xu and Jiayi Ma. Emfusion: An unsupervised enhanced medical image fusion network. Information Fusion, 76:177-186, 2021. 2, 3", + "[38] Xing Xu, Kaiyi Lin, Lianli Gao, Huimin Lu, Heng Tao Shen, and Xuelong Li. Learning cross-modal common representations by private-shared subspaces separation. IEEE Trans. Cybern., 52(5):3261-3275, 2022. 2, 3", + "[39] Huang Yao, Rongjun Qin, and Xiaoyu Chen. Unmanned aerial vehicle for remote sensing applications - A review. Remote Sensing, 11(12):1443, 2019. 1", + "[40] Maoxun Yuan and Xingxing Wei. $\\mathbf{C}^2$ former: Calibrated and complementary transformer forrgb-infrared object detection. arXiv:2306.16175, 2023. 1, 7", + "[41] Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In ECCV, pages 509-525, 2022. 1, 2, 7", + "[42] Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschlager, and Susanne Saminger-Platz. Central moment discrepancy (CMD) for domain-invariant representation learning. In ICLR, 2017. 4", + "[43] Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Grid anchor based image cropping: A new benchmark and an efficient model. IEEE TPAMI, 44(3):1304-1319, 2022. 1", + "[44] Lu Zhang, Zhiyong Liu, Shifeng Zhang, Xu Yang, Hong Qiao, Kaizhu Huang, and Amir Hussain. Cross-modality interactive attention network for multispectral pedestrian detection. Information Fusion, 50:20-29, 2019. 7", + "[45] Lu Zhang, Zhiyong Liu, Xiangyu Zhu, Zhan Song, Xu Yang, Zhen Lei, and Hong Qiao. Weakly aligned feature fusion for multimodal object detection. arXiv:2204.09848, 2022. 2, 7", + "[46] Xinyue Zhang, Chengcai Leng, Yameng Hong, Zhao Pei, Irene Cheng, and Anup Basu. Multimodal remote sensing image registration methods and advancements: A survey. Remote Sensing, 13(24):5128, 2021. 2", + "[47] Zhimeng Zhang and Yu Ding. Adaptive affine transformation: A simple and effective operation for spatial misaligned image generation. In ACM MM, pages 1167-1176, 2022. 1", + "[48] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, 2023. 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Jiaxiang Zheng, Yichen Tian, Chao Yuan, Kai Yin, Feifei Zhang, Fangmiao Chen, and Qiang Chen. Mdesnet: Multitask difference-enhanced siamese network for building change detection in high-resolution remote sensing images. Remote Sensing, 14(15):3775, 2022. 4", + "[50] Kailai Zhou, Linsen Chen, and Xun Cao. Improving multispectral pedestrian detection by addressing modality imbalance problems. In ECCV, pages 787-803, 2020. 7", + "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3", + "[52] Barbara Zitová and Jan Flusser. Image registration methods: a survey. Image Vis. Comput., 21(11):977-1000, 2003. 2" + ], + "bbox": [ + 501, + 92, + 890, + 275 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26845", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_model.json b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0aa5034cacc095e2311672f59f532e5a4c203a3 --- /dev/null +++ b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.326, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.848, + 0.177 + ], + "angle": 0, + "content": "Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.204, + 0.891, + 0.24 + ], + "angle": 0, + "content": "Chen Chen Jiahao Qi Xingyue Liu Kangcheng Bin Ruigang Fu Xikun Hu Ping Zhong* National University of Defense Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.242, + 0.864, + 0.275 + ], + "angle": 0, + "content": "chenchen21c@nudt.edu.cn, qijiahao1996@nudt.edu.cn, liuxingyuel8@nudt.edu.cn, binkc21@nudt.edu.cn, furuigang08@nudt.edu.cn, xikun@nudt.edu.cn, zhongping@nudt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.678 + ], + "angle": 0, + "content": "Visible-infrared (RGB-IR) image fusion has shown great potentials in object detection based on unmanned aerial vehicles (UAVs). However, the weakly misalignment problem between multimodal image pairs limits its performance in object detection. Most existing methods often ignore the modality gap and emphasize a strict alignment, resulting in an upper bound of alignment quality and an increase of implementation costs. To address these challenges, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA), which could adaptively adjust the relative positions between multimodal features. Considering the impact of modality gap on the cross-modality spatial matching, a Cross-modality Spatial Offset Modeling (CSOM) module is designed to establish a common subspace to estimate the precise feature-level offsets. Then, an Offset-guided Deformable Alignment and Fusion (ODAF) module is utilized to implicitly capture optimal fusion positions for detection task rather than conducting a strict alignment. Comprehensive experiments demonstrate that our method not only achieves state-of-the-art performance in the UAVs-based object detection task but also shows strong robustness to the weakly misalignment problem." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.21, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.873 + ], + "angle": 0, + "content": "Visible-infrared (RGB-IR) object detection on unmanned aerial vehicles (UAVs) has attracted extensive attention [2, 27, 39], due to its ability to make full use of complementary information and achieve robust around-the-clock detection. However, the collected RGB-IR images are typically mismatched in resolutions and field of views owing to disparate imaging principles. Thus, it is challenging to fuse the corresponding information of target between multimodal images." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.308, + 0.697, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.382, + 0.674, + 0.392 + ], + "angle": 0, + "content": "(a) Image-level alignment method" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.309, + 0.892, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.381, + 0.879, + 0.391 + ], + "angle": 0, + "content": "(b) Feature-level alignment method" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.4, + 0.755, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.487, + 0.671, + 0.496 + ], + "angle": 0, + "content": "(c) Our OFA" + }, + { + "type": "image_footnote", + "bbox": [ + 0.77, + 0.405, + 0.862, + 0.417 + ], + "angle": 0, + "content": "(F) : Fusion Module" + }, + { + "type": "image_footnote", + "bbox": [ + 0.776, + 0.421, + 0.885, + 0.432 + ], + "angle": 0, + "content": ":Multimodal Features" + }, + { + "type": "image_footnote", + "bbox": [ + 0.776, + 0.437, + 0.872, + 0.454 + ], + "angle": 0, + "content": "Modality-specific Features" + }, + { + "type": "image_footnote", + "bbox": [ + 0.776, + 0.457, + 0.879, + 0.48 + ], + "angle": 0, + "content": "Modality-invariant Features" + }, + { + "type": "image_footnote", + "bbox": [ + 0.797, + 0.481, + 0.874, + 0.49 + ], + "angle": 0, + "content": "Deformation Field" + }, + { + "type": "list", + "bbox": [ + 0.77, + 0.405, + 0.885, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.676 + ], + "angle": 0, + "content": "Figure 1. Previous multimodal object detection architectures under weakly misalignment conditions vs. our OAFA. (a) Image-level alignment method: executing a pixel-level strict alignment by explicitly extracting the global deformation fields from multimodal images. (b) Feature-level alignment method: explicitly estimating feature offsets through multimodal features to achieve strict feature alignment. It can be further classified into global and local methods based on whether alignment is exclusively performed on target regions. (c) Our OAFA: the basic global offsets are estimated by the modality-invariant features to eliminate the impact of modality gap. Based on this offset, we perform adaptive feature alignment without strict alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To increase the quality of fusion representations, researchers always apply pre-registration techniques, involving image cropping [43] and affine transformation [47] before the fusing process. Nevertheless, since distinct imaging time in multimodal sensors and intricate movements in targets, it is difficult, or even impossible, to achieve precise UAVs image alignment [29]. It inevitably leads to local inconsistencies in multimodal spatial distribution, named the weakly misalignment problem [41], resulting in a mismatch of the multimodal target features in corresponding positions [40]. Since the detection task is sensitive to spatial locations, the weakly misalignment problem would lead to significant performance degradation. Several studies have been carried out to deal with this issue by accomplishing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26836" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "a strict alignment based on offset estimation, as shown in Fig. 1a and Fig. 1b. Although these methods have made significant progress, there still remain two main challenges:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Inaccurate offset estimation. Conventional methods devote to predicting spatial offsets by identifying the distribution and appearance discrepancies in image or feature level. However, in fact, the discrepancies arise not only from spatial offsets but also from modality gap [31], which stems from the differences in imaging principles and wavelengths between RGB and IR sensors. This gap could make it hard to correctly match cross-modality representations in spatial, leading to misleading offset estimation. Hence, the first challenge is how to bridge the modality gap to achieve accurate cross-modal offset estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.504 + ], + "angle": 0, + "content": "High dependency on strict alignment. The strict alignment is necessary for most existing methods. Unfortunately, the strict alignment always requires additional annotation information, such as the ground truth offsets between multimodal representations [45]. It will lead to higher implementation costs that limit real-world applications. Moreover, the detection performance relies heavily on the quality of alignment. However, even the state-of-the-art alignment methods have the risk of failure in some scenarios [18], which may cause the error delivery to the detection task. Consequently, the second challenge is how to enhance the robustness of the detection model to the weakly misalignment problem without strict alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.763 + ], + "angle": 0, + "content": "To this end, the purpose of this paper is to achieve adaptive alignment between RGB (sensed modality) and IR (reference modality) images based on spatial offsets for superior detection performance instead of strict alignment. Technically, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA) for UAVs-based multimodal object detection under weakly misalignment conditions. It mainly consists of a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module for dealing with the two challenges, respectively. Considering the first challenge, CSOM establishes a cross-modal common subspace to obtain the modality-invariant features for eliminating modality gap. As distributional aligned mappings, the modality-invariant features could be utilized to evaluate spatial consistency between multimodal features for more accurate cross-modal offsets estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As for the second challenge, ODAF is designed to adaptively adjust feature positions by implicit offset compensation, aiming to make full use of the offsets and achieve superior detection performance without strict alignment. In specific, ODAF involves a learnable deformation component that dynamically predicts the offsets of sampling convolution kernels. With the help of the offsets provided by CSOM, ODAF captures optimal sampling positions in sensed modality features to attain comprehensive target rep" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "resentations after fusion. This module predicts the corresponding sampling positions for diverse regions in different image pairs, contributing to a strong flexibility of OAFA in handling diverse and complex offsets. Besides, the training stability of deformable components is improved since the explicit guidance of the offsets reduces the burden of the adaptive alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: (1) We propose a robust RGB-IR object detection method for UAVs images under weakly misalignment conditions, effectively implementing adaptive feature alignment instead of strict alignment. (2) Considering the difficulty of estimating cross-modality offset, we build a multimodal common subspace to reduce the impact of modality gap on multimodal spatial matching. (3) An offset-guided deformable alignment module is exploited to achieve adaptive alignment by implicitly capturing optimal fusion positions, enhancing the accuracy of the detection task. (4) Extensive experiments on public dataset show that our OAFA is superior to state-of-the-art models only with a simple fusion strategy. Moreover, it proves the effectiveness in addressing the weakly misalignment problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.642, + 0.453 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Multimodal object detection under weakly misalignment conditions. A general solution to address the weakly misalignment problem is to conduct pixel-level prealignment [24, 46, 52], which learn a global mapping function from multimodal images to correct the pixel coordinates. Many of these methods [3, 16, 26] are supervised with the deformable fields in order to achieve strict prealignment. To achieve end-to-end training, Zhang et al. [45] first introduced a data-driven region feature alignment module to predict proposal regions offsets from multimodal features, which was inexpensively constrained by the target locations in RGB-IR images. On this basis, Yuan et al. [41] made full use of the oriented multimodal annotations to realize multimodal target alignment in position, size, and angle. Due to the requirement of extra supervisions for the strict alignment process, the above methods are labor-intensive and costly to implement. Besides, the existing modality gap at both the pixel-level and feature-level leads to incorrect estimation of spatial offsets. Consequently, it is indeed significant to propose a novel method for RGB-IR object detection under weakly misalignment conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Common subspace learning. Common subspace learning aims to exploit shared information among multiple data. In field of multimodal learning, it can be used to bridge the modality gap by extracting the unified latent representations in the inter-modality samples. Works can be broadly grouped into transformation-based [9, 23, 32, 37], weight-sharing-based [12, 14, 15], and decoupling-based methods [13, 20, 38]. The transformation-based methods" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.354, + 0.887, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.417, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Figure 2. An Overview of OAFA. It consists of two components: a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module. In our implementation, we choose the RGB images as the sensed modality whose features are shifted to achieve alignment, and the IR images as the reference modality. The CSOM can decouple the original multimodal features to modality-invariant and modality-specific features and predict a superior spatial offset in a common subspace. The ODAF utilizes this basic offset to adaptively align the sample positions of the RGB features with IR features, and then fuses the aligned RGB and original IR features. Finally, the fused features are conducted to generate the ultimate detection results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.782 + ], + "angle": 0, + "content": "generally transform one modality (images/features) into another modality by domain transfer [9] or auto-encoder [23, 32, 37]. It is hard to achieve an end-to-end training due to the requirement of a pre-trained transformation network. The weight-sharing-based methods devote to designing the weight-shared network within a two-stream framework to only capture modality-invariant features [12, 14, 15]. Nevertheless, overemphasis on common features may result in the underutilization of complementary information. The decoupling-based methods overcome this obstacle by trying to separate multimodal features into modality-invariant and modality-specific features with a series of constraint functions [13, 20, 38]. Inspired by the third category methods, we develop a decoupled multimodal learning network with three constraints for the detection task. Moreover, due to the complexity of UAVs-based multimodal features, we design the network with partial weight sharing and custom encoder architecture to achieve effective disentanglement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Deformable convolution. Convolutional neural networks (CNNs) are inherently limited in modeling geometric transformations because of their fixed kernel structures. To enhance the transformation modeling capability of CNNs, Dai et al. [8] designed the deformable convolution, which has shown superior performance in several high-level vision tasks [1, 28, 51]. In recent years, it has also been" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.678 + ], + "angle": 0, + "content": "used in the field of video super-resolution [5, 30, 33, 34]. As a pioneer work, Tian et al. [30] adopted the deformable convolution to align the features of adjacent frames in the temporal domain. It demonstrates that deformable convolution can effectively capture the motion cues of the targets. Nonetheless, Chan et al. [4] has proved that the deformable alignment is difficult to train. Training instability frequently leads to offset overflow, degrading the final performance in downstream task. Therefore, we construct a basic offset as explicit guidance to reduce the burden of deformable alignment training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.591, + 0.702 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.61, + 0.726 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method aims to enhance the accuracy and robustness of the RGB-IR object detection model under weakly misalignment conditions. As shown in Fig. 2, a pair of weakly misaligned RGB-IR images are fed into a two-stream network to acquire the multiscale multimodal representations. Subsequently, a decoupled learning network is designed to develop a common subspace to figure out the modality-invariant and modality-specific features. To mitigate the influence of modality gap on offset estimation, we exclusively employ the modality-invariant features to predict spatial offsets. Then, this basic offsets are utilized to implica" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26838" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "ity adjust the spatial positions of the modality-invariant and modality-specific features. It accomplishes adaptive feature alignment that is tailored for RGB-IR object detection task without any extra supervisions. Finally, the aligned RGB and original IR features are fused with a concatenation operation to generate final detection results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.189, + 0.47, + 0.22 + ], + "angle": 0, + "content": "3.2. Cross-modality Spatial Offset Modeling Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Given the RGB features \\( F_{rgb} \\) and IR features \\( F_{ir} \\), CSOM predicts the spatial offsets \\( \\phi_c \\) as basic knowledge for the subsequent alignment module. The CSOM consists of a decoupled multimodal learning submodule and a spatial offset modeling submodule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.456 + ], + "angle": 0, + "content": "Decoupled multimodal learning. This submodule mainly builds a common subspace \\(\\tilde{\\mathbb{C}} = [\\mathbb{C}_{rgb},\\mathbb{C}_{ir}]\\) and a specific subspace \\(\\tilde{\\mathbb{S}} = [\\mathbb{S}_{rgb},\\mathbb{S}_{ir}]\\), as shown in Fig. 2a. For each input RGB-IR image pair \\(\\mathcal{I}_{rgb}\\in \\mathbb{R}^{H\\times W\\times 3}\\) and \\(\\mathcal{I}_{ir}\\in \\mathbb{R}^{H\\times W\\times 1}\\), we map it to the multimodal representations \\(\\mathcal{F}_m(m\\in \\{rgb,ir\\})\\) by a basic pyramid feature extractor \\(\\mathcal{B}_m\\). The \\(\\mathcal{F}_m\\) is the bounded random samples on the interval \\([a,b]^N\\). Then, the modality-invariant features \\(\\mathcal{F}_m^c\\) and modality-specific features \\(\\mathcal{F}_m^s\\) projected from \\(\\tilde{\\mathbb{C}}\\) and \\(\\mathbb{S}\\) are figured out by the encoding functions:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.463, + 0.47, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {m} ^ {s} = \\mathcal {S} _ {m} \\left(\\mathcal {F} _ {m}; \\theta_ {m} ^ {s}\\right), \\mathcal {F} _ {m} ^ {c} = \\mathcal {C} _ {m} \\left(\\mathcal {F} _ {m}; \\theta^ {c}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.576 + ], + "angle": 0, + "content": "where the invariant feature encoder \\(\\mathcal{C}_m\\) shares the parameters \\(\\theta^c\\) across two modalities, whereas the specific feature encoder \\(\\mathcal{S}_m\\) assigns separate parameters \\(\\theta_m^s\\) for each modality. Considering the diversity information in UAVs images, we not only design the decoupled network in objective function but also in network architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.698 + ], + "angle": 0, + "content": "In terms of the network architecture, an accepted assumption is that the input features are more related in the low layers, while the high-layer features tend to be uncorrelated [48]. Thus, \\(\\mathcal{C}_m\\) is constructed with a convolutional layer followed by the SiLU activation function. In contrast, \\(S_{m}\\) contains a convolutional layer and a C3 module with three convolutional and a series of bottleneck layers to obtain deeper features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.822 + ], + "angle": 0, + "content": "As for the objective functions, inspired by [13], we apply three types of constraints for decoupled multimodal learning: Similarity Loss, Difference Loss, and Semantic Loss. First, we employ a similarity loss \\(\\mathcal{L}_{sim}\\) to minimize the distance between \\(\\mathcal{F}_{rgb}^{c}\\) and \\(\\mathcal{F}_{ir}^{c}\\) to learn common subspace \\(\\tilde{\\mathbb{C}}\\). The \\(\\mathcal{L}_{sim}\\) is defined by a state-of-the-art distance metric named central moment discrepancy [42] to measure the distribution dissimilarity between two features:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.828, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s i m} = \\frac {1}{| b - a |} \\left\\| E \\left(\\mathcal {F} _ {r g b} ^ {c}\\right) - E \\left(\\mathcal {F} _ {i r} ^ {c}\\right) \\right\\| _ {2} \\\\ + \\sum_ {k = 2} ^ {K} \\frac {1}{| b - a | ^ {k}} \\left\\| C _ {K} \\left(\\mathcal {F} _ {r g b} ^ {c}\\right) - C _ {K} \\left(\\mathcal {F} _ {i r} ^ {c}\\right) \\right\\| _ {2}, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "where \\(E(\\cdot)\\) is the empirical expectation vector, \\(C_K(\\cdot)\\) is the vector of \\(k\\)-th order sample central moments, and \\(\\| \\cdot \\| _2\\) represents the Euclidean norm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Second, we present a difference loss \\(\\mathcal{L}_{dif}\\) for highlighting the dissimilarity between \\(\\mathcal{F}_{rgb}^{s}\\) and \\(\\mathcal{F}_{ir}^{s}\\) to capture complementary information. Additionally, as \\(\\mathcal{F}_m^c\\) and \\(\\mathcal{F}_m^s\\) are expected to be mutually independent, \\(\\mathcal{L}_{dif}\\) should be extended to them as well. These discrepancies are achieved by the squared Frobenius norm:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.235, + 0.892, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d i f} = \\sum_ {m \\in \\{r g b, i r \\}} \\sqrt {\\operatorname {T r} \\left(\\left(\\left(F _ {m} ^ {s}\\right) ^ {T} F _ {m} ^ {c}\\right) ^ {T} \\left(\\left(F _ {m} ^ {s}\\right) ^ {T} F _ {m} ^ {c}\\right)\\right)} \\tag {3} \\\\ + \\sqrt {T r ((\\left(F _ {r g b} ^ {s}\\right) ^ {T} F _ {i r} ^ {s}) ^ {T} (\\left(F _ {r g b} ^ {s}\\right) ^ {T} F _ {i r} ^ {s}))}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.308, + 0.729, + 0.322 + ], + "angle": 0, + "content": "where \\(Tr(\\cdot)\\) is trace of the matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Third, only restrained by the above objectives, there exists a potential risk that \\( \\mathcal{C}_m \\) learns useless noisy cues, whereas \\( S_m \\) acquires almost all modality information. To address this issue, we add a semantic loss \\( \\mathcal{L}_{sem} \\) to enforce \\( \\mathcal{C}_m \\) to learn discriminative features:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.405, + 0.891, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} = \\sum_ {m \\in \\{r g b, i r \\}} \\left(\\mathcal {L} _ {c l s} \\left(\\mathcal {F} _ {m} ^ {c}\\right) + \\mathcal {L} _ {r e g} \\left(\\mathcal {F} _ {m} ^ {c}\\right) + \\mathcal {L} _ {o b j} \\left(\\mathcal {F} _ {m} ^ {c}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.542 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{cls}\\), \\(\\mathcal{L}_{reg}\\), and \\(\\mathcal{L}_{obj}\\) are the detection loss functions of the YOLOv5 [17]. It should be noted that this object detection branch, serving as an auxiliary task, is not incorporated during the inference stage to enhance the efficiency of our OAFA. Finally, the total loss of the decoupled multimodal learning submodule can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.552, + 0.891, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e c} = \\lambda_ {1} \\mathcal {L} _ {s i m} + \\lambda_ {2} \\mathcal {L} _ {s i m} + \\lambda_ {3} \\mathcal {L} _ {s e m}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.891, + 0.605 + ], + "angle": 0, + "content": "where \\(\\lambda_{i}\\) is the trade-off weights. In this study, we set \\(\\lambda_{1}\\), \\(\\lambda_{2}\\), and \\(\\lambda_{3}\\) to 0.03, 0.01, and 0.03 respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Spatial offset modeling. We model the spatial offsets \\(\\phi_c\\) by the estimating spatial differences between \\(\\mathcal{F}_{rgb}^c\\) and \\(\\mathcal{F}_{ir}^c\\). In the field of change detection [49], the differences are calculated directly by subtracting or concatenating two features. However, it would introduce significant noises when dealing with intricate backgrounds in UAVs images. To solve this problem, a feature difference enhancement module is presented to filter out irrelevant changes and concentrate on shift objects. As shown in Fig. 2b, \\(\\mathcal{F}_m^c\\) is firstly fed into a spatial attention module to capture the representations of the target and crucial background areas in each modality. The structure of the spatial attention module is the same as [35]. In this case, the spatially enhanced unimodal feature \\(F_m^{c,p}\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.824, + 0.892, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {m} ^ {c, p} = \\sigma \\left(f _ {7 \\times 7} \\left(\\operatorname {C a t} \\left(\\operatorname {M a x} \\left(\\mathcal {F} _ {m} ^ {c}\\right), \\operatorname {M e a n} \\left(\\mathcal {F} _ {m} ^ {c}\\right)\\right)\\right)\\right) \\odot \\mathcal {F} _ {m} ^ {c}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid function, \\(f_{i\\times j}\\) denotes the \\(i\\times j\\) convolution layer, \\(Cat(\\cdot)\\) refers to the concatenation operation, and \\(\\odot\\) indicates the Hadamard product operation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26839" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "We conduct a concatenation instead of subtraction strategy to extract the changing features \\( F_{dif} \\) to mitigate the influence of complex background. Then, a channel difference enhanced module is introduced to further suppress channel noise. Specifically, we employ simultaneous average-pooling and max-pooling to generate distinct spatial context descriptors. Following this, channel attention can be derived by transmitting the spatial context descriptors to a shared multilayer perceptron network \\( f_{mlp} \\). \\( \\mathcal{F}_m^{c,p} \\) are combined with the channel attention map to get the spatial-channel enhanced features \\( \\mathcal{F}_m^{c,e} \\). The above operation is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.284, + 0.47, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} \\mathcal {F} _ {m} ^ {c, e} = \\sigma (f _ {m l p} (A v g P o o l (\\mathcal {F} _ {d i f})) + \\\\ f _ {m l p} (M a x P o o l (\\mathcal {F} _ {d i f}))) \\times \\mathcal {F} _ {m} ^ {c, p} + \\mathcal {F} _ {m} ^ {c, p}. \\end{array} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.332, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Finally, \\(\\phi_c\\) is obtained through a fusion of the \\(\\mathcal{F}_{rgb}^{c,e}\\) and \\(\\mathcal{F}_{ir}^{c,e}\\) along with a nonlinear layer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.373, + 0.47, + 0.403 + ], + "angle": 0, + "content": "3.3. Offset-guided Deformable Alignment and Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.547 + ], + "angle": 0, + "content": "We have obtained the offsets \\(\\phi_c\\) from CSOM, which represents the spatial differences between \\(\\mathcal{F}_{rgb}\\) and \\(\\mathcal{F}_{ir}\\). Inspired by [5], the \\(\\phi_c\\) can be an effectively initialized offset to unsupervised alignment task. Hence, ODAF is designed to adaptively capture optimal fusion location with the help of the basic knowledge \\(\\phi_c\\), incorporating desired complementary information for the subsequent detection task. It is composed of two submodules: offset-guided deformable alignment and decoupled feature fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Offset-guided deformable alignment. We mainly apply a deformable convolution to achieve implicit offset compensation and adaptive alignment for the detection task. Unlike the default deformable convolution that learns offsets of convolution kernels from their original features, the offsets generated from our module are mapped from the basic offsets \\(\\phi_c\\), which allows for more effective feature correction and reliable model training. Considering different contributions from each offset region for the detection task, we add a modulation scalar \\(\\Delta m_k\\) that is learned from \\(\\phi_c\\) to dynamically aggregate information around the corresponding position \\(p\\). As shown in Fig. 2c, given the center sampling value \\(x(p)\\) in the sensed features \\(\\mathcal{F}_{rgb}^n (n\\in \\{s,c\\})\\) and the basic offsets \\(\\phi_c\\), each counterpart \\(y(p)\\) in aligned features \\(\\mathcal{F}_{rgb}^{n,a}\\) can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.788, + 0.47, + 0.831 + ], + "angle": 0, + "content": "\\[\ny (p) = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\cdot x \\left(p + \\phi_ {c} + p _ {k} + \\Delta p _ {k}\\right) \\cdot \\Delta m _ {k}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(K\\), \\(p_k\\), and \\(w_k\\) denote the number of kernel weights, the \\(k\\)-th fixed offset, and the \\(k\\)-th kernel weight, respectively. For instance, \\(p_k \\in \\{(-1,1), (-1,0), \\dots, (1,1)\\}\\) donates a regular grid of a \\(3 \\times 3\\) kernel defined with \\(K = 9\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\(\\Delta p_{k}\\) denotes the implicit compensation to \\(\\phi_c\\) of the \\(k\\)-th location. It is estimated with \\(\\operatorname{conv}(x)[p + \\phi_c]\\) where \\(\\phi_c\\) can be regarded as a prior of \\(\\Delta p_{k}\\) in the standard deformable convolution. Besides, since the location \\(p + \\phi_c + p_k + \\Delta p_k\\) may be fractional, we adopt bilinear interpolation to obtain the final sampling values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Decoupled feature fusion. In traditional fusion process, \\(\\mathcal{F}_m^c\\) and \\(\\mathcal{F}_m^s\\) are directly concatenated to obtain the fused features \\(\\mathcal{F}_f\\) [6]. Unfortunately, the similarity constraint between \\(\\mathcal{F}_{rgb}^{c,a}\\) and \\(\\mathcal{F}_{ir}^{c}\\) leads to much redundancy. To eliminate duplicate information and increase discriminative representations, \\(\\mathcal{F}_{rgb}^{c,a}\\) and \\(\\mathcal{F}_{ir}^{c}\\) should be combined and optimized before fusion. Thus, the fused features \\(F_{f}\\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.298, + 0.892, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {f} = C a t \\left(f _ {1 \\times 1} \\left(C a t \\left(\\mathcal {F} _ {r g b} ^ {c, a}, \\mathcal {F} _ {i r} ^ {c}\\right)\\right), \\mathcal {F} _ {r g b} ^ {s, a}, \\mathcal {F} _ {i r} ^ {s}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.325, + 0.658, + 0.341 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Training loss. The overall method is based on a supervised detection framework. The purpose of our multimodal alignment is not to form a strict alignment but to achieve optimal adjustment between multimodal features to improve the performance of the subsequent detection task. Thus, the alignment subtask is integrated into the main detection task instead of adding extra alignment loss. The total loss function of our OAFA can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.481, + 0.892, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {d e c}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Two-stage training strategy. A premise for effective feature alignment is that \\(\\mathcal{F}_m^c\\) contains sufficient spatial information, thereby estimating more accurate basic offsets. To this end, we adopt a two-stage learning strategy to train our OAFA end-to-end. In stage I, to gain effective \\(\\mathcal{F}_m^c\\), we try to ensure that pre-decoupling features \\(\\mathcal{F}_m\\) contain more comprehensive target representations. Technically, we only train the feature extraction and fusion module with \\(\\mathcal{L}_{cls}\\), \\(\\mathcal{L}_{reg}\\), and \\(\\mathcal{L}_{obj}\\). Specifically, a pair of \\(\\mathcal{I}_{rgb}\\) and \\(\\mathcal{I}_{ir}\\) are fed into a two-stream network to extract multiscale features \\(\\mathcal{F}_m\\) and gain fusion features \\(\\mathcal{F}_f\\) with the concatenation operation. Then, the \\(\\mathcal{F}_f\\) is put into the detection network to obtain the detection results. In stage II, we apply all the proposed modules to our network. In this process, the unimodal feature extraction network \\(\\mathcal{B}_m\\) is initialized with the weights from stage I. Then, the joint optimization of the unimodal feature extraction and multimodal feature alignment network is carried out effectively with \\(\\mathcal{L}_{total}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.634, + 0.81 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.707, + 0.834 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Datasets and metrics. We evaluate our method on DroneVehicle dataset [29], which is a large-scale multimodal dataset for vehicle detection in UAVs images. It contains 28,439 image pairs with 953,087 vehicles, covering" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.893, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 3. Four Examples of detection results on the test set of DroneVehicle dataset under weakly misalignment conditions. The confidence threshold is set to 0.25. The results of fusion methods are visual in IR images to correspond to the supervisory label. Red, green, and pink rectangles represent car, freight-car, and truck targets, respectively. Objects correctly detected are represented in green dashed circles, while incorrectly detected objects are represented in red dashed circles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.471, + 0.633 + ], + "angle": 0, + "content": "different viewing angles, heights, lighting conditions, and scenarios. The dense-oriented annotations are made for five categories: car, truck, bus, van, and freight-car. Though the DroneVehicle performs pre-registration on each RGB-IR image pair, the weakly misalignment problem still exists. As shown in Fig. 4, about \\(35\\%\\) bounding boxes exhibit deviation issues in DroneVehicle dataset, and the shift distance mostly ranges from 0 to 15 pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.729 + ], + "angle": 0, + "content": "This dataset is divided into 17,990 image pairs for training, 1,469 image pairs for validating, and the remaining 8,980 image pairs for testing. Following the standard evaluation metrics, we report the mean average precision (mAP) with intersection over union (IoU) threshold of 0.4 on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Training details. The experiments are carried out on a single NVIDIA RTX A6000 GPU with 48 GB of memory. We implement our algorithm with Pytorch toolbox and SGD optimizer with a momentum of 0.937 and a weight decay of 0.0005. The initial learning rate is set to 0.01 and eventually reduced to 0.002 by cosine annealing [22]. The input size of our model is set to \\(640 \\times 640\\) in the preprocessing stage and the batch size is 16. The training epoch is set to 150 with 50 and 100 epochs in the first and second stages, respectively. We use the ground truth of the IR images as the training label, owning to the more comprehensive target" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.509, + 0.682, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.629, + 0.684, + 0.642 + ], + "angle": 0, + "content": "(a) Statistics of misalignment ratio" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.522, + 0.896, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.629, + 0.858, + 0.641 + ], + "angle": 0, + "content": "(b) Statistics of shift value" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Figure 4. The statistics of ground truth bounding-boxes misalignment ratio and shift value within RGB-IR image pairs on DroneVehicle dataset. annotations available in the IR modality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.698, + 0.728 + ], + "angle": 0, + "content": "4.2. Results Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Compared methods. We compare our method with the state-of-the-art unimodal and multimodal object detection methods. Unimodal methods: We choose a series of oriented object detection methods for comparison, including the one-stage methods: RetinaNet [19], \\(\\mathbf{S}^2\\mathbf{A}\\) Net [11], and YOLOv5s [17]; and the two-stage methods: Faster R-CNN [25], Oriented R-CNN [36], and RoITransformer [10]. RGB and IR images are separately used as training data to evaluate the detection performance of each algorithm. Multimodal fusion methods: We also compare our method with eight recent fusion methods:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26841" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. Detection results (mAP, in %) on DroneVehicle dataset. Note that all detectors locate and classify vehicles with OBB heads. The speed refers to the speed of network inference without post-processing (batch size = 1). In speed, our method is only compared with the fusion methods. Best results are highlighted in bold. And the second one is marked with underline." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.134, + 0.89, + 0.452 + ], + "angle": 0, + "content": "
DetectorsModalityCarTruckFreight-carBusVanmAP (%) ↑Speed (fps) ↑
RetinaNet [19]78.534.424.169.828.847.1-
Faster R-CNN [25]79.049.037.277.037.055.9-
Oriented R-CNN [36]RGB80.153.841.685.443.360.8-
S²A-Net [11]80.054.242.284.943.861.0-
RoITransformer [10]61.655.142.385.544.861.6-
YOLOv5s [17]78.655.343.887.146.062.1-
RetinaNet [19]88.835.439.576.532.154.5-
Faster R-CNN [25]89.453.548.387.042.664.2-
Oriented R-CNN [36]IR89.857.453.189.345.467.0-
S²A-Net [11]89.954.555.888.948.467.5-
RoITransformer [10]90.160.458.989.752.270.3-
YOLOv5s [17]90.059.560.889.553.870.7-
UA-CMDet [29]87.560.746.887.138.064.09.1
Halfway Fusion [21]90.162.358.589.149.870.020.4
CIAN [44]90.163.860.789.150.370.821.7
AR-CNN [45]90.164.862.189.451.571.618.2
MBNet [50]RGB+IR90.164.462.488.853.671.921.7
TSFADet [41]89.967.963.789.854.073.118.6
C²Former [40]90.268.364.489.858.574.2-
SLBAF-Net [7]90.272.068.689.959.976.163.2
Ours90.376.873.390.366.079.433.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.578 + ], + "angle": 0, + "content": "UA-CMDet [29], Halfway Fusion [21], CIAN [44], MBNet [50], AR-CNN [45], TSFADet [41], \\(C^2\\)Former [40], and SLBAF-Net [7]. Among them, UA-CMDet, MBNet, TSFADet, and \\(C^2\\)Former all consider the weakly misalignment problem during the fusion process, which affirms the universality and significance of this problem. For each compared method, we use their original experimental settings to ensure equity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Quantitative comparison. The results are shown in Tab. 1. In the unimodal methods, benefitting from a well-designed multiscale feature extraction network, YOLOv5s has comparable detection accuracy (62.1% in RGB images and 70.7% in IR images), even better than some fusion methods. Correspondingly, the multimodal methods SLBAF-Net, constructed upon the two-stream YOLOv5 framework, has also yielded excellent results. Compared with all methods, we can find that our OAFA achieves the best average performance on the validation set, which is 3.3% higher than the second place. It is worth noting that our method has obvious advantages in distinguishing confusing categories (truck and freight-car). The reason may stem from the fact that OAFA excels in identifying the optimal spatial correspondences of multimodal features, obtaining reliable features of the target after fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparison. Our baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branches are fused with a concatenation operation in multiscale features. Among the compared methods mentioned above, we select the optimal unimodal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.653 + ], + "angle": 0, + "content": "YOLOv5s and the baseline model as the qualitative comparative model. We provide visual detection results of the compared methods in Fig. 3. It can be observed that, given the constraint of the IR labels, the objects are successfully detected by OAFA when the target positions are shifted, while both unimodal and baseline model experience classification and localization errors or fail to detect, especially for occluded and few-shot object detection. The underlying causes can be speculated that the weak misalignment problem confuses the fusing model to learn discriminative features. In contrast, our method has the capability to capture aligned and reliable features, improving the performance in classification and localization processes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Speed comparison. We compare the speed of OAFA with other fusion detection methods to prove the efficiency of our method. For a fair comparison, all the detection models are tested with batch size 1, and the input size is set to \\(640 \\times 640\\). As shown in Tab. 1, our proposed OAFA achieves a speed of 33.1 FPS, outperforming most existing multimodal object detection methods. Although our method requires more computational time compared to SLBAF-Net, it would be worthwhile to invest this additional time to achieve superior performance. Furthermore, it can be seen that OAFA can accomplish real-time detection on the NVIDIA RTX A6000 platform." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.655, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct several ablation studies to validate the effectiveness of the different components in our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.411, + 0.106 + ], + "angle": 0, + "content": "proposed method. The results are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.334 + ], + "angle": 0, + "content": "Effect of CSOM module. We first analyze the effect of the CSOM module as it provides the fundamental offset estimation, which determines the stability of the model training. We perform ablation experiments on both the decoupled multimodal learning (DML) module and the spatial offset modeling (SOM) module separately. As presented in the first to third rows of Tab. 2, the performance of the baseline model is improved by \\(0.4\\%\\) and \\(1.6\\%\\) after incorporating the SOM and DML modules, respectively. It verifies that both coarse-grained feature alignment and increasing fusion flexibility through decouple learning are effective for the detection task. Besides, when integrating these two submodules, the performance improvement of model over baseline is \\(2.1\\%\\), which indicates that the quality of the feature alignment can be enhanced by mitigating modal gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.468, + 0.456 + ], + "angle": 0, + "content": "Effect of ODAF module. We then verify the necessity of the ODAF module in addressing the weak misalignment problem. The results demonstrate that despite ODAF has poor stability in model training without dependable offset guidance, incorporating ODAF into the baseline can provide \\(+1.7\\%\\) significant improvement. It is owing to adaptively adjust the optimal sampling position for acquiring reliable information in RGB images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.468, + 0.697 + ], + "angle": 0, + "content": "Effect of two-stage training. The two-stage training strategy provides comprehensive unimodal representations for multimodal alignment and fusion. We can observe that the accuracy decreases to \\(77.3\\%\\) if we directly train the whole network without two-stage training. It is \\(0.9\\%\\) and \\(0.5\\%\\) lower than exclusively incorporating CSOM and ODAF, respectively. The reason could be that the inappropriate initialization in feature extraction network leads to training instability and makes it hard to obtain abundant representations in spatial and semantic. After undergoing a two-stage training process, our method attains a performance level of \\(79.4\\%\\), surpassing the capabilities of any individual module. It demonstrates that the two-stage training strategy effectively enhances training accuracy and stability. The superior method to fully exploit the strengths of CSOM and ODAF will be further studied in future work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.468, + 0.745 + ], + "angle": 0, + "content": "All in all, not only each independent module but also the entire proposed framework can contribute to the multimodal object detection under weakly misalignment conditions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.327, + 0.771 + ], + "angle": 0, + "content": "4.4. Robustness to Position Shift" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To quantitatively assess the robustness of our method to the weakly misalignment issue, we conduct experiments on DroneVehicle by manually simulating positional offsets. The tested model is trained on the original training image pairs and validated on the shifted image pairs. In our experiments, we hold the IR images constant while introducing spatial offsets along the \\(x\\)-axis and \\(y\\)-axis for RGB images. The variations in pixel values are defined within the" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.159 + ], + "angle": 0, + "content": "Table 2. Ablation study on DroneVehicle dataset. The baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branch are fused with a multiscale concatenation operation. The D denotes DML, S denotes SOM, C denotes CSOM, and O denotes ODAF." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.163, + 0.89, + 0.274 + ], + "angle": 0, + "content": "
MethodCSOMODAFTwo-stage TrainingmAP (%) ↑
DMLSOM
Baseline76.1
Baseline+S✓76.5
Baseline+D✓77.7
Baseline+C✓✓78.2
Baseline+O✓77.8
Baseline+C+O✓✓✓77.3
OAFA✓✓✓✓79.4
" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.283, + 0.687, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.391, + 0.672, + 0.404 + ], + "angle": 0, + "content": "(a) SLBAF-Net (Baseline)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.283, + 0.885, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.392, + 0.842, + 0.403 + ], + "angle": 0, + "content": "(b) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Figure 5. Surface plot visualization of the position shift experiments. Horizontal coordinates indicate various shift steps along the \\(x\\)-axis and \\(y\\)-axis. Vertical coordinate denotes the detection performances in terms of mAP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.588 + ], + "angle": 0, + "content": "range \\(\\{(\\Delta x, \\Delta y) \\mid \\Delta x, \\Delta y \\in [15, -15]; \\Delta x, \\Delta y \\in \\mathbb{Z}\\}\\). The quantitative outcomes of this experiment in baseline and our model are discernible from Fig. 5. It can be seen that in response to larger position shifts, our method consistently exhibits minor fluctuations in performance, whereas the baseline model experiences a noticeable performance decline. This result demonstrates the robustness of OAFA under weakly misalignment conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.619, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.869 + ], + "angle": 0, + "content": "In this paper, we propose a robust RGB-IR object detection method OAFA for UAVs images under weakly misalignment conditions. Different from previous methods, OAFA can achieve adaptive multimodal feature alignment without strict alignment. To mitigate the impact of the modality gap on multimodal spatial matching, the modality-invariant features are acquired in a common subspace to estimate accurate offsets. Taking the offsets as basic knowledge, the adaptive alignment implicitly captures optimal fusion positions and steadily carries out feature-level corrections with a deformable alignment module. Experimental results on DroneVehicle dataset demonstrate the superior performances of OAFA compared with the state-of-the-art methods. Moreover, our method shows robustness against the weakly misalignment problem in position shift experiments." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26843" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Gedas Bertasius, Lorenzo Torresani, and Jianbo Shi. Object detection in video with spatiotemporal sampling networks. In ECCV, pages 342-357, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] Ilker Bozcan and Erdal Kayacan. AU-AIR: A multi-modal unmanned aerial vehicle dataset for low altitude traffic surveillance. In ICRA, pages 8504-8510, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.255 + ], + "angle": 0, + "content": "[3] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Dong Nie, Minjeong Kim, Qian Wang, and Dinggang Shen. Deformable image registration based on similarity-steered CNN regression. In MICCAI, pages 300-308, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[4] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In AAAI, pages 973-981, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[5] Kelvin C. K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5962-5971, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[6] Hao Chen, Yongjian Deng, Youfu Li, Tzu-Yi Hung, and Guosheng Lin. RGBD salient object detection via disentangled cross-modal fusion. IEEE TIP, 29:8407-8416, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[7] Xiaolong Cheng, Keke Geng, Ziwei Wang, Jinhu Wang, Yuxiao Sun, and Pengbo Ding. Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment. Multimedia Tools and Applications, pages 1-20, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[8] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[9] Chaitanya Devaguptapu, Ninad Akolekar, Manuj M. Sharma, and Vineeth N. Balasubramanian. Borrow from anywhere: Pseudo multi-modal object detection in thermal imagery. In CVPR, pages 1029-1038, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[10] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In CVPR, pages 2849-2858, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[11] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE TGRS, 60: 1-11, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[12] Renlong Hang, Zhu Li, Pedram Ghamisi, Danfeng Hong, Guiyu Xia, and Qingshan Liu. Classification of hyperspectral and lidar data using coupled cnns. IEEE TGRS, 58(7): 4939-4950, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[13] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. MISA: modality-invariant and -specific representations for multimodal sentiment analysis. In ACM MM, pages 1122-1131, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[14] Danfeng Hong, Naoto Yokoya, Jocelyn Chanussot, and Xiao Xiang Zhu. Cospace: Common subspace learning from hyperspectral-multispectral correspondences. IEEE TGRS, 57(7):4349-4359, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[15] Danfeng Hong, Jocelyn Chanussot, Naoto Yokoya, Jian Kang, and Xiao Xiang Zhu. Learning-shared cross-modality representation using multispectral-lidar and hyperspectral data. IEEE GRSL, 17(8):1470-1474, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Lloyd Haydn Hughes, Diego Marcos, Sylvain Lobry, Devis Tuia, and Michael Schmitt. A deep learning framework for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "matching of sar and optical imagery. ISPRS, 169:166-179, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.89, + 0.148 + ], + "angle": 0, + "content": "[17] Glenn Jocher. ultralytics/yolov5. https://github.com/ultralytics/yolov5, Oct. 2020. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[18] Huafeng Li, Junzhi Zhao, Jinxing Li, Zhengtao Yu, and Guangming Lu. Feature dynamic alignment and refinement for infrared-visible image fusion: Translation robust fusion. Information Fusion, 95:26-41, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[20] Fan Liu, Zhiyong Cheng, Huilin Chen, Anan Liu, Liqiang Nie, and Mohan S. Kankanhalli. Disentangled multimodal representation learning for recommendation. arXiv:2203.05406, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[21] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Multispectral deep neural networks for pedestrian detection. In BMVC, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[22] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[23] Sijie Mai, Haifeng Hu, and Songlong Xing. Modality to modality translation: An adversarial representation learning and graph fusion network for multimodal fusion. In AAAI, pages 164-172, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.454 + ], + "angle": 0, + "content": "[24] Sayan Nag. Image registration techniques: A survey. arXiv:1712.07540, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[26] Marc-Michel Rohé, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: Learning deformable image registration using shape matching. In MIC-CAI, pages 266–274, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[27] Hazim Shakhatreh, Ahmad Sawalmeh, Ala I. Al-Fuqaha, Zuochao Dou, Eyad Almaita, Issa Khalil, Noor Shamsiah Othman, Abdallah Khreishah, and Mohsen Guizani. Unmanned aerial vehicles (uavs): A survey on civil applications and key research challenges. IEEE Access, 7:48572-48634, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[28] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 536-553, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[29] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning. IEEE TCSVT, 32(10):6700-6713, 2022. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally-deformable alignment network for video super-resolution. In CVPR, pages 3357-3366, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[31] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In IJCAI, pages 3508-3515, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In ICCV, pages 3622–3631, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[33] Hua Wang, Dewei Su, Chuangchuang Liu, Longcun Jin, Xianfang Sun, and Xinyi Peng. Deformable non-local network for video super-resolution. IEEE Access, 7:177734-177744, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[34] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. In CVPR, pages 1954-1963, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[35] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In ECCV, pages 3–19, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[36] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented R-CNN for object detection. In ICCV, pages 3500-3509, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[37] Han Xu and Jiayi Ma. Emfusion: An unsupervised enhanced medical image fusion network. Information Fusion, 76:177-186, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[38] Xing Xu, Kaiyi Lin, Lianli Gao, Huimin Lu, Heng Tao Shen, and Xuelong Li. Learning cross-modal common representations by private-shared subspaces separation. IEEE Trans. Cybern., 52(5):3261-3275, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[39] Huang Yao, Rongjun Qin, and Xiaoyu Chen. Unmanned aerial vehicle for remote sensing applications - A review. Remote Sensing, 11(12):1443, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[40] Maoxun Yuan and Xingxing Wei. \\(\\mathbf{C}^2\\)former: Calibrated and complementary transformer forrgb-infrared object detection. arXiv:2306.16175, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[41] Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In ECCV, pages 509-525, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[42] Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschlager, and Susanne Saminger-Platz. Central moment discrepancy (CMD) for domain-invariant representation learning. In ICLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[43] Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Grid anchor based image cropping: A new benchmark and an efficient model. IEEE TPAMI, 44(3):1304-1319, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[44] Lu Zhang, Zhiyong Liu, Shifeng Zhang, Xu Yang, Hong Qiao, Kaizhu Huang, and Amir Hussain. Cross-modality interactive attention network for multispectral pedestrian detection. Information Fusion, 50:20-29, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[45] Lu Zhang, Zhiyong Liu, Xiangyu Zhu, Zhan Song, Xu Yang, Zhen Lei, and Hong Qiao. Weakly aligned feature fusion for multimodal object detection. arXiv:2204.09848, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[46] Xinyue Zhang, Chengcai Leng, Yameng Hong, Zhao Pei, Irene Cheng, and Anup Basu. Multimodal remote sensing image registration methods and advancements: A survey. Remote Sensing, 13(24):5128, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[47] Zhimeng Zhang and Yu Ding. Adaptive affine transformation: A simple and effective operation for spatial misaligned image generation. In ACM MM, pages 1167-1176, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[48] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[49] Jiaxiang Zheng, Yichen Tian, Chao Yuan, Kai Yin, Feifei Zhang, Fangmiao Chen, and Qiang Chen. Mdesnet: Multitask difference-enhanced siamese network for building change detection in high-resolution remote sensing images. Remote Sensing, 14(15):3775, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[50] Kailai Zhou, Linsen Chen, and Xun Cao. Improving multispectral pedestrian detection by addressing modality imbalance problems. In ECCV, pages 787-803, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[52] Barbara Zitová and Jan Flusser. Image registration methods: a survey. Image Vis. Comput., 21(11):977-1000, 2003. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26845" + } + ] +] \ No newline at end of file diff --git a/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_origin.pdf b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..21884a6107b7d26e58506d316f3ae2b01e8d97fc --- /dev/null +++ b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/7c6f4239-84d8-42d9-a2a3-2ef9f81fe732_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74ea6a76e9f86fc29d92c843cca3855fc4b3f132a34cb2aaea154d864ca41d1d +size 9718187 diff --git a/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/full.md b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b7df69352e411b141d20d55eb0548f9ca7ebb30 --- /dev/null +++ b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/full.md @@ -0,0 +1,295 @@ +# Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection + +Chen Chen Jiahao Qi Xingyue Liu Kangcheng Bin Ruigang Fu Xikun Hu Ping Zhong* National University of Defense Technology, China + +chenchen21c@nudt.edu.cn, qijiahao1996@nudt.edu.cn, liuxingyuel8@nudt.edu.cn, binkc21@nudt.edu.cn, furuigang08@nudt.edu.cn, xikun@nudt.edu.cn, zhongping@nudt.edu.cn + +# Abstract + +Visible-infrared (RGB-IR) image fusion has shown great potentials in object detection based on unmanned aerial vehicles (UAVs). However, the weakly misalignment problem between multimodal image pairs limits its performance in object detection. Most existing methods often ignore the modality gap and emphasize a strict alignment, resulting in an upper bound of alignment quality and an increase of implementation costs. To address these challenges, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA), which could adaptively adjust the relative positions between multimodal features. Considering the impact of modality gap on the cross-modality spatial matching, a Cross-modality Spatial Offset Modeling (CSOM) module is designed to establish a common subspace to estimate the precise feature-level offsets. Then, an Offset-guided Deformable Alignment and Fusion (ODAF) module is utilized to implicitly capture optimal fusion positions for detection task rather than conducting a strict alignment. Comprehensive experiments demonstrate that our method not only achieves state-of-the-art performance in the UAVs-based object detection task but also shows strong robustness to the weakly misalignment problem. + +# 1. Introduction + +Visible-infrared (RGB-IR) object detection on unmanned aerial vehicles (UAVs) has attracted extensive attention [2, 27, 39], due to its ability to make full use of complementary information and achieve robust around-the-clock detection. However, the collected RGB-IR images are typically mismatched in resolutions and field of views owing to disparate imaging principles. Thus, it is challenging to fuse the corresponding information of target between multimodal images. + +![](images/52d5000f81f66fa61293d1f6c6fcd59c797f57cead50e577bba722d9cd0f30ae.jpg) +(a) Image-level alignment method + +(b) Feature-level alignment method +![](images/a9e37ab502cb93b5f1a75269cee6c2b42d1fb070dcc4e98601546f23db00fb74.jpg) +:Multimodal Features + +(c) Our OFA +Figure 1. Previous multimodal object detection architectures under weakly misalignment conditions vs. our OAFA. (a) Image-level alignment method: executing a pixel-level strict alignment by explicitly extracting the global deformation fields from multimodal images. (b) Feature-level alignment method: explicitly estimating feature offsets through multimodal features to achieve strict feature alignment. It can be further classified into global and local methods based on whether alignment is exclusively performed on target regions. (c) Our OAFA: the basic global offsets are estimated by the modality-invariant features to eliminate the impact of modality gap. Based on this offset, we perform adaptive feature alignment without strict alignment. +![](images/734ae5bdfa091c01da8da7917ea0f75ae7f619750ef3858c94c257b5f71a79a6.jpg) +(F) : Fusion Module +Modality-specific Features +Modality-invariant Features +Deformation Field + +To increase the quality of fusion representations, researchers always apply pre-registration techniques, involving image cropping [43] and affine transformation [47] before the fusing process. Nevertheless, since distinct imaging time in multimodal sensors and intricate movements in targets, it is difficult, or even impossible, to achieve precise UAVs image alignment [29]. It inevitably leads to local inconsistencies in multimodal spatial distribution, named the weakly misalignment problem [41], resulting in a mismatch of the multimodal target features in corresponding positions [40]. Since the detection task is sensitive to spatial locations, the weakly misalignment problem would lead to significant performance degradation. Several studies have been carried out to deal with this issue by accomplishing + +a strict alignment based on offset estimation, as shown in Fig. 1a and Fig. 1b. Although these methods have made significant progress, there still remain two main challenges: + +Inaccurate offset estimation. Conventional methods devote to predicting spatial offsets by identifying the distribution and appearance discrepancies in image or feature level. However, in fact, the discrepancies arise not only from spatial offsets but also from modality gap [31], which stems from the differences in imaging principles and wavelengths between RGB and IR sensors. This gap could make it hard to correctly match cross-modality representations in spatial, leading to misleading offset estimation. Hence, the first challenge is how to bridge the modality gap to achieve accurate cross-modal offset estimation. + +High dependency on strict alignment. The strict alignment is necessary for most existing methods. Unfortunately, the strict alignment always requires additional annotation information, such as the ground truth offsets between multimodal representations [45]. It will lead to higher implementation costs that limit real-world applications. Moreover, the detection performance relies heavily on the quality of alignment. However, even the state-of-the-art alignment methods have the risk of failure in some scenarios [18], which may cause the error delivery to the detection task. Consequently, the second challenge is how to enhance the robustness of the detection model to the weakly misalignment problem without strict alignment. + +To this end, the purpose of this paper is to achieve adaptive alignment between RGB (sensed modality) and IR (reference modality) images based on spatial offsets for superior detection performance instead of strict alignment. Technically, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA) for UAVs-based multimodal object detection under weakly misalignment conditions. It mainly consists of a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module for dealing with the two challenges, respectively. Considering the first challenge, CSOM establishes a cross-modal common subspace to obtain the modality-invariant features for eliminating modality gap. As distributional aligned mappings, the modality-invariant features could be utilized to evaluate spatial consistency between multimodal features for more accurate cross-modal offsets estimation. + +As for the second challenge, ODAF is designed to adaptively adjust feature positions by implicit offset compensation, aiming to make full use of the offsets and achieve superior detection performance without strict alignment. In specific, ODAF involves a learnable deformation component that dynamically predicts the offsets of sampling convolution kernels. With the help of the offsets provided by CSOM, ODAF captures optimal sampling positions in sensed modality features to attain comprehensive target rep + +resentations after fusion. This module predicts the corresponding sampling positions for diverse regions in different image pairs, contributing to a strong flexibility of OAFA in handling diverse and complex offsets. Besides, the training stability of deformable components is improved since the explicit guidance of the offsets reduces the burden of the adaptive alignment. + +Our contributions can be summarized as follows: (1) We propose a robust RGB-IR object detection method for UAVs images under weakly misalignment conditions, effectively implementing adaptive feature alignment instead of strict alignment. (2) Considering the difficulty of estimating cross-modality offset, we build a multimodal common subspace to reduce the impact of modality gap on multimodal spatial matching. (3) An offset-guided deformable alignment module is exploited to achieve adaptive alignment by implicitly capturing optimal fusion positions, enhancing the accuracy of the detection task. (4) Extensive experiments on public dataset show that our OAFA is superior to state-of-the-art models only with a simple fusion strategy. Moreover, it proves the effectiveness in addressing the weakly misalignment problem. + +# 2. Related Work + +Multimodal object detection under weakly misalignment conditions. A general solution to address the weakly misalignment problem is to conduct pixel-level prealignment [24, 46, 52], which learn a global mapping function from multimodal images to correct the pixel coordinates. Many of these methods [3, 16, 26] are supervised with the deformable fields in order to achieve strict prealignment. To achieve end-to-end training, Zhang et al. [45] first introduced a data-driven region feature alignment module to predict proposal regions offsets from multimodal features, which was inexpensively constrained by the target locations in RGB-IR images. On this basis, Yuan et al. [41] made full use of the oriented multimodal annotations to realize multimodal target alignment in position, size, and angle. Due to the requirement of extra supervisions for the strict alignment process, the above methods are labor-intensive and costly to implement. Besides, the existing modality gap at both the pixel-level and feature-level leads to incorrect estimation of spatial offsets. Consequently, it is indeed significant to propose a novel method for RGB-IR object detection under weakly misalignment conditions. + +Common subspace learning. Common subspace learning aims to exploit shared information among multiple data. In field of multimodal learning, it can be used to bridge the modality gap by extracting the unified latent representations in the inter-modality samples. Works can be broadly grouped into transformation-based [9, 23, 32, 37], weight-sharing-based [12, 14, 15], and decoupling-based methods [13, 20, 38]. The transformation-based methods + +![](images/3a58246ba63d96996c9784b32e94c2344362e96ea3525ababf77e20b3249e5b5.jpg) + +![](images/e040a3eeb21d4a404706801326bac3d5f93911841f87e41535c7eb0b0c4b4eb4.jpg) +Figure 2. An Overview of OAFA. It consists of two components: a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module. In our implementation, we choose the RGB images as the sensed modality whose features are shifted to achieve alignment, and the IR images as the reference modality. The CSOM can decouple the original multimodal features to modality-invariant and modality-specific features and predict a superior spatial offset in a common subspace. The ODAF utilizes this basic offset to adaptively align the sample positions of the RGB features with IR features, and then fuses the aligned RGB and original IR features. Finally, the fused features are conducted to generate the ultimate detection results. + +generally transform one modality (images/features) into another modality by domain transfer [9] or auto-encoder [23, 32, 37]. It is hard to achieve an end-to-end training due to the requirement of a pre-trained transformation network. The weight-sharing-based methods devote to designing the weight-shared network within a two-stream framework to only capture modality-invariant features [12, 14, 15]. Nevertheless, overemphasis on common features may result in the underutilization of complementary information. The decoupling-based methods overcome this obstacle by trying to separate multimodal features into modality-invariant and modality-specific features with a series of constraint functions [13, 20, 38]. Inspired by the third category methods, we develop a decoupled multimodal learning network with three constraints for the detection task. Moreover, due to the complexity of UAVs-based multimodal features, we design the network with partial weight sharing and custom encoder architecture to achieve effective disentanglement. + +Deformable convolution. Convolutional neural networks (CNNs) are inherently limited in modeling geometric transformations because of their fixed kernel structures. To enhance the transformation modeling capability of CNNs, Dai et al. [8] designed the deformable convolution, which has shown superior performance in several high-level vision tasks [1, 28, 51]. In recent years, it has also been + +used in the field of video super-resolution [5, 30, 33, 34]. As a pioneer work, Tian et al. [30] adopted the deformable convolution to align the features of adjacent frames in the temporal domain. It demonstrates that deformable convolution can effectively capture the motion cues of the targets. Nonetheless, Chan et al. [4] has proved that the deformable alignment is difficult to train. Training instability frequently leads to offset overflow, degrading the final performance in downstream task. Therefore, we construct a basic offset as explicit guidance to reduce the burden of deformable alignment training. + +# 3. Method + +# 3.1. Overview + +Our method aims to enhance the accuracy and robustness of the RGB-IR object detection model under weakly misalignment conditions. As shown in Fig. 2, a pair of weakly misaligned RGB-IR images are fed into a two-stream network to acquire the multiscale multimodal representations. Subsequently, a decoupled learning network is designed to develop a common subspace to figure out the modality-invariant and modality-specific features. To mitigate the influence of modality gap on offset estimation, we exclusively employ the modality-invariant features to predict spatial offsets. Then, this basic offsets are utilized to implica + +ity adjust the spatial positions of the modality-invariant and modality-specific features. It accomplishes adaptive feature alignment that is tailored for RGB-IR object detection task without any extra supervisions. Finally, the aligned RGB and original IR features are fused with a concatenation operation to generate final detection results. + +# 3.2. Cross-modality Spatial Offset Modeling Module + +Given the RGB features $F_{rgb}$ and IR features $F_{ir}$ , CSOM predicts the spatial offsets $\phi_c$ as basic knowledge for the subsequent alignment module. The CSOM consists of a decoupled multimodal learning submodule and a spatial offset modeling submodule. + +Decoupled multimodal learning. This submodule mainly builds a common subspace $\tilde{\mathbb{C}} = [\mathbb{C}_{rgb},\mathbb{C}_{ir}]$ and a specific subspace $\tilde{\mathbb{S}} = [\mathbb{S}_{rgb},\mathbb{S}_{ir}]$ , as shown in Fig. 2a. For each input RGB-IR image pair $\mathcal{I}_{rgb}\in \mathbb{R}^{H\times W\times 3}$ and $\mathcal{I}_{ir}\in \mathbb{R}^{H\times W\times 1}$ , we map it to the multimodal representations $\mathcal{F}_m(m\in \{rgb,ir\})$ by a basic pyramid feature extractor $\mathcal{B}_m$ . The $\mathcal{F}_m$ is the bounded random samples on the interval $[a,b]^N$ . Then, the modality-invariant features $\mathcal{F}_m^c$ and modality-specific features $\mathcal{F}_m^s$ projected from $\tilde{\mathbb{C}}$ and $\mathbb{S}$ are figured out by the encoding functions: + +$$ +\mathcal {F} _ {m} ^ {s} = \mathcal {S} _ {m} \left(\mathcal {F} _ {m}; \theta_ {m} ^ {s}\right), \mathcal {F} _ {m} ^ {c} = \mathcal {C} _ {m} \left(\mathcal {F} _ {m}; \theta^ {c}\right), \tag {1} +$$ + +where the invariant feature encoder $\mathcal{C}_m$ shares the parameters $\theta^c$ across two modalities, whereas the specific feature encoder $\mathcal{S}_m$ assigns separate parameters $\theta_m^s$ for each modality. Considering the diversity information in UAVs images, we not only design the decoupled network in objective function but also in network architecture. + +In terms of the network architecture, an accepted assumption is that the input features are more related in the low layers, while the high-layer features tend to be uncorrelated [48]. Thus, $\mathcal{C}_m$ is constructed with a convolutional layer followed by the SiLU activation function. In contrast, $S_{m}$ contains a convolutional layer and a C3 module with three convolutional and a series of bottleneck layers to obtain deeper features. + +As for the objective functions, inspired by [13], we apply three types of constraints for decoupled multimodal learning: Similarity Loss, Difference Loss, and Semantic Loss. First, we employ a similarity loss $\mathcal{L}_{sim}$ to minimize the distance between $\mathcal{F}_{rgb}^{c}$ and $\mathcal{F}_{ir}^{c}$ to learn common subspace $\tilde{\mathbb{C}}$ . The $\mathcal{L}_{sim}$ is defined by a state-of-the-art distance metric named central moment discrepancy [42] to measure the distribution dissimilarity between two features: + +$$ +\begin{array}{l} \mathcal {L} _ {s i m} = \frac {1}{| b - a |} \left\| E \left(\mathcal {F} _ {r g b} ^ {c}\right) - E \left(\mathcal {F} _ {i r} ^ {c}\right) \right\| _ {2} \\ + \sum_ {k = 2} ^ {K} \frac {1}{| b - a | ^ {k}} \left\| C _ {K} \left(\mathcal {F} _ {r g b} ^ {c}\right) - C _ {K} \left(\mathcal {F} _ {i r} ^ {c}\right) \right\| _ {2}, \tag {2} \\ \end{array} +$$ + +where $E(\cdot)$ is the empirical expectation vector, $C_K(\cdot)$ is the vector of $k$ -th order sample central moments, and $\| \cdot \| _2$ represents the Euclidean norm. + +Second, we present a difference loss $\mathcal{L}_{dif}$ for highlighting the dissimilarity between $\mathcal{F}_{rgb}^{s}$ and $\mathcal{F}_{ir}^{s}$ to capture complementary information. Additionally, as $\mathcal{F}_m^c$ and $\mathcal{F}_m^s$ are expected to be mutually independent, $\mathcal{L}_{dif}$ should be extended to them as well. These discrepancies are achieved by the squared Frobenius norm: + +$$ +\begin{array}{l} \mathcal {L} _ {d i f} = \sum_ {m \in \{r g b, i r \}} \sqrt {\operatorname {T r} \left(\left(\left(F _ {m} ^ {s}\right) ^ {T} F _ {m} ^ {c}\right) ^ {T} \left(\left(F _ {m} ^ {s}\right) ^ {T} F _ {m} ^ {c}\right)\right)} \tag {3} \\ + \sqrt {T r ((\left(F _ {r g b} ^ {s}\right) ^ {T} F _ {i r} ^ {s}) ^ {T} (\left(F _ {r g b} ^ {s}\right) ^ {T} F _ {i r} ^ {s}))}, \\ \end{array} +$$ + +where $Tr(\cdot)$ is trace of the matrix. + +Third, only restrained by the above objectives, there exists a potential risk that $\mathcal{C}_m$ learns useless noisy cues, whereas $S_m$ acquires almost all modality information. To address this issue, we add a semantic loss $\mathcal{L}_{sem}$ to enforce $\mathcal{C}_m$ to learn discriminative features: + +$$ +\mathcal {L} _ {s e m} = \sum_ {m \in \{r g b, i r \}} \left(\mathcal {L} _ {c l s} \left(\mathcal {F} _ {m} ^ {c}\right) + \mathcal {L} _ {r e g} \left(\mathcal {F} _ {m} ^ {c}\right) + \mathcal {L} _ {o b j} \left(\mathcal {F} _ {m} ^ {c}\right)\right), \tag {4} +$$ + +where $\mathcal{L}_{cls}$ , $\mathcal{L}_{reg}$ , and $\mathcal{L}_{obj}$ are the detection loss functions of the YOLOv5 [17]. It should be noted that this object detection branch, serving as an auxiliary task, is not incorporated during the inference stage to enhance the efficiency of our OAFA. Finally, the total loss of the decoupled multimodal learning submodule can be defined as: + +$$ +\mathcal {L} _ {d e c} = \lambda_ {1} \mathcal {L} _ {s i m} + \lambda_ {2} \mathcal {L} _ {s i m} + \lambda_ {3} \mathcal {L} _ {s e m}, \tag {5} +$$ + +where $\lambda_{i}$ is the trade-off weights. In this study, we set $\lambda_{1}$ , $\lambda_{2}$ , and $\lambda_{3}$ to 0.03, 0.01, and 0.03 respectively. + +Spatial offset modeling. We model the spatial offsets $\phi_c$ by the estimating spatial differences between $\mathcal{F}_{rgb}^c$ and $\mathcal{F}_{ir}^c$ . In the field of change detection [49], the differences are calculated directly by subtracting or concatenating two features. However, it would introduce significant noises when dealing with intricate backgrounds in UAVs images. To solve this problem, a feature difference enhancement module is presented to filter out irrelevant changes and concentrate on shift objects. As shown in Fig. 2b, $\mathcal{F}_m^c$ is firstly fed into a spatial attention module to capture the representations of the target and crucial background areas in each modality. The structure of the spatial attention module is the same as [35]. In this case, the spatially enhanced unimodal feature $F_m^{c,p}$ can be computed as: + +$$ +\mathcal {F} _ {m} ^ {c, p} = \sigma \left(f _ {7 \times 7} \left(\operatorname {C a t} \left(\operatorname {M a x} \left(\mathcal {F} _ {m} ^ {c}\right), \operatorname {M e a n} \left(\mathcal {F} _ {m} ^ {c}\right)\right)\right)\right) \odot \mathcal {F} _ {m} ^ {c}, \tag {6} +$$ + +where $\sigma$ is the sigmoid function, $f_{i\times j}$ denotes the $i\times j$ convolution layer, $Cat(\cdot)$ refers to the concatenation operation, and $\odot$ indicates the Hadamard product operation. + +We conduct a concatenation instead of subtraction strategy to extract the changing features $F_{dif}$ to mitigate the influence of complex background. Then, a channel difference enhanced module is introduced to further suppress channel noise. Specifically, we employ simultaneous average-pooling and max-pooling to generate distinct spatial context descriptors. Following this, channel attention can be derived by transmitting the spatial context descriptors to a shared multilayer perceptron network $f_{mlp}$ . $\mathcal{F}_m^{c,p}$ are combined with the channel attention map to get the spatial-channel enhanced features $\mathcal{F}_m^{c,e}$ . The above operation is represented as: + +$$ +\begin{array}{r} \mathcal {F} _ {m} ^ {c, e} = \sigma (f _ {m l p} (A v g P o o l (\mathcal {F} _ {d i f})) + \\ f _ {m l p} (M a x P o o l (\mathcal {F} _ {d i f}))) \times \mathcal {F} _ {m} ^ {c, p} + \mathcal {F} _ {m} ^ {c, p}. \end{array} \tag {7} +$$ + +Finally, $\phi_c$ is obtained through a fusion of the $\mathcal{F}_{rgb}^{c,e}$ and $\mathcal{F}_{ir}^{c,e}$ along with a nonlinear layer. + +# 3.3. Offset-guided Deformable Alignment and Fusion Module + +We have obtained the offsets $\phi_c$ from CSOM, which represents the spatial differences between $\mathcal{F}_{rgb}$ and $\mathcal{F}_{ir}$ . Inspired by [5], the $\phi_c$ can be an effectively initialized offset to unsupervised alignment task. Hence, ODAF is designed to adaptively capture optimal fusion location with the help of the basic knowledge $\phi_c$ , incorporating desired complementary information for the subsequent detection task. It is composed of two submodules: offset-guided deformable alignment and decoupled feature fusion. + +Offset-guided deformable alignment. We mainly apply a deformable convolution to achieve implicit offset compensation and adaptive alignment for the detection task. Unlike the default deformable convolution that learns offsets of convolution kernels from their original features, the offsets generated from our module are mapped from the basic offsets $\phi_c$ , which allows for more effective feature correction and reliable model training. Considering different contributions from each offset region for the detection task, we add a modulation scalar $\Delta m_k$ that is learned from $\phi_c$ to dynamically aggregate information around the corresponding position $p$ . As shown in Fig. 2c, given the center sampling value $x(p)$ in the sensed features $\mathcal{F}_{rgb}^n (n\in \{s,c\})$ and the basic offsets $\phi_c$ , each counterpart $y(p)$ in aligned features $\mathcal{F}_{rgb}^{n,a}$ can be obtained as follows: + +$$ +y (p) = \sum_ {k = 1} ^ {K} \omega_ {k} \cdot x \left(p + \phi_ {c} + p _ {k} + \Delta p _ {k}\right) \cdot \Delta m _ {k}, \tag {8} +$$ + +where $K$ , $p_k$ , and $w_k$ denote the number of kernel weights, the $k$ -th fixed offset, and the $k$ -th kernel weight, respectively. For instance, $p_k \in \{(-1,1), (-1,0), \dots, (1,1)\}$ donates a regular grid of a $3 \times 3$ kernel defined with $K = 9$ . + +$\Delta p_{k}$ denotes the implicit compensation to $\phi_c$ of the $k$ -th location. It is estimated with $\operatorname{conv}(x)[p + \phi_c]$ where $\phi_c$ can be regarded as a prior of $\Delta p_{k}$ in the standard deformable convolution. Besides, since the location $p + \phi_c + p_k + \Delta p_k$ may be fractional, we adopt bilinear interpolation to obtain the final sampling values. + +Decoupled feature fusion. In traditional fusion process, $\mathcal{F}_m^c$ and $\mathcal{F}_m^s$ are directly concatenated to obtain the fused features $\mathcal{F}_f$ [6]. Unfortunately, the similarity constraint between $\mathcal{F}_{rgb}^{c,a}$ and $\mathcal{F}_{ir}^{c}$ leads to much redundancy. To eliminate duplicate information and increase discriminative representations, $\mathcal{F}_{rgb}^{c,a}$ and $\mathcal{F}_{ir}^{c}$ should be combined and optimized before fusion. Thus, the fused features $F_{f}$ are defined as: + +$$ +\mathcal {F} _ {f} = C a t \left(f _ {1 \times 1} \left(C a t \left(\mathcal {F} _ {r g b} ^ {c, a}, \mathcal {F} _ {i r} ^ {c}\right)\right), \mathcal {F} _ {r g b} ^ {s, a}, \mathcal {F} _ {i r} ^ {s}\right). \tag {9} +$$ + +# 3.4. Model Training + +Training loss. The overall method is based on a supervised detection framework. The purpose of our multimodal alignment is not to form a strict alignment but to achieve optimal adjustment between multimodal features to improve the performance of the subsequent detection task. Thus, the alignment subtask is integrated into the main detection task instead of adding extra alignment loss. The total loss function of our OAFA can be represented as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {\text {o b j}} + \mathcal {L} _ {\text {d e c}}. \tag {10} +$$ + +Two-stage training strategy. A premise for effective feature alignment is that $\mathcal{F}_m^c$ contains sufficient spatial information, thereby estimating more accurate basic offsets. To this end, we adopt a two-stage learning strategy to train our OAFA end-to-end. In stage I, to gain effective $\mathcal{F}_m^c$ , we try to ensure that pre-decoupling features $\mathcal{F}_m$ contain more comprehensive target representations. Technically, we only train the feature extraction and fusion module with $\mathcal{L}_{cls}$ , $\mathcal{L}_{reg}$ , and $\mathcal{L}_{obj}$ . Specifically, a pair of $\mathcal{I}_{rgb}$ and $\mathcal{I}_{ir}$ are fed into a two-stream network to extract multiscale features $\mathcal{F}_m$ and gain fusion features $\mathcal{F}_f$ with the concatenation operation. Then, the $\mathcal{F}_f$ is put into the detection network to obtain the detection results. In stage II, we apply all the proposed modules to our network. In this process, the unimodal feature extraction network $\mathcal{B}_m$ is initialized with the weights from stage I. Then, the joint optimization of the unimodal feature extraction and multimodal feature alignment network is carried out effectively with $\mathcal{L}_{total}$ . + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets and metrics. We evaluate our method on DroneVehicle dataset [29], which is a large-scale multimodal dataset for vehicle detection in UAVs images. It contains 28,439 image pairs with 953,087 vehicles, covering + +![](images/bd41082a33f17763e39345eb3181f5ac4ff14cc73f3dc3e9d63feb82d70709a4.jpg) +Figure 3. Four Examples of detection results on the test set of DroneVehicle dataset under weakly misalignment conditions. The confidence threshold is set to 0.25. The results of fusion methods are visual in IR images to correspond to the supervisory label. Red, green, and pink rectangles represent car, freight-car, and truck targets, respectively. Objects correctly detected are represented in green dashed circles, while incorrectly detected objects are represented in red dashed circles. + +different viewing angles, heights, lighting conditions, and scenarios. The dense-oriented annotations are made for five categories: car, truck, bus, van, and freight-car. Though the DroneVehicle performs pre-registration on each RGB-IR image pair, the weakly misalignment problem still exists. As shown in Fig. 4, about $35\%$ bounding boxes exhibit deviation issues in DroneVehicle dataset, and the shift distance mostly ranges from 0 to 15 pixels. + +This dataset is divided into 17,990 image pairs for training, 1,469 image pairs for validating, and the remaining 8,980 image pairs for testing. Following the standard evaluation metrics, we report the mean average precision (mAP) with intersection over union (IoU) threshold of 0.4 on the validation set. + +Training details. The experiments are carried out on a single NVIDIA RTX A6000 GPU with 48 GB of memory. We implement our algorithm with Pytorch toolbox and SGD optimizer with a momentum of 0.937 and a weight decay of 0.0005. The initial learning rate is set to 0.01 and eventually reduced to 0.002 by cosine annealing [22]. The input size of our model is set to $640 \times 640$ in the preprocessing stage and the batch size is 16. The training epoch is set to 150 with 50 and 100 epochs in the first and second stages, respectively. We use the ground truth of the IR images as the training label, owning to the more comprehensive target + +![](images/01aa374e69d58cd5caafd5adafb1d331df9e6bc81dc5ebb2a123a52f0f4137bc.jpg) +(a) Statistics of misalignment ratio + +![](images/db56953588ad45889834f1bad95b0f82c6270f50f4d8a93064b93921e6e34bf8.jpg) +(b) Statistics of shift value +Figure 4. The statistics of ground truth bounding-boxes misalignment ratio and shift value within RGB-IR image pairs on DroneVehicle dataset. annotations available in the IR modality. + +# 4.2. Results Comparisons + +Compared methods. We compare our method with the state-of-the-art unimodal and multimodal object detection methods. Unimodal methods: We choose a series of oriented object detection methods for comparison, including the one-stage methods: RetinaNet [19], $\mathbf{S}^2\mathbf{A}$ Net [11], and YOLOv5s [17]; and the two-stage methods: Faster R-CNN [25], Oriented R-CNN [36], and RoITransformer [10]. RGB and IR images are separately used as training data to evaluate the detection performance of each algorithm. Multimodal fusion methods: We also compare our method with eight recent fusion methods: + +Table 1. Detection results (mAP, in %) on DroneVehicle dataset. Note that all detectors locate and classify vehicles with OBB heads. The speed refers to the speed of network inference without post-processing (batch size = 1). In speed, our method is only compared with the fusion methods. Best results are highlighted in bold. And the second one is marked with underline. + +
DetectorsModalityCarTruckFreight-carBusVanmAP (%) ↑Speed (fps) ↑
RetinaNet [19]78.534.424.169.828.847.1-
Faster R-CNN [25]79.049.037.277.037.055.9-
Oriented R-CNN [36]RGB80.153.841.685.443.360.8-
S²A-Net [11]80.054.242.284.943.861.0-
RoITransformer [10]61.655.142.385.544.861.6-
YOLOv5s [17]78.655.343.887.146.062.1-
RetinaNet [19]88.835.439.576.532.154.5-
Faster R-CNN [25]89.453.548.387.042.664.2-
Oriented R-CNN [36]IR89.857.453.189.345.467.0-
S²A-Net [11]89.954.555.888.948.467.5-
RoITransformer [10]90.160.458.989.752.270.3-
YOLOv5s [17]90.059.560.889.553.870.7-
UA-CMDet [29]87.560.746.887.138.064.09.1
Halfway Fusion [21]90.162.358.589.149.870.020.4
CIAN [44]90.163.860.789.150.370.821.7
AR-CNN [45]90.164.862.189.451.571.618.2
MBNet [50]RGB+IR90.164.462.488.853.671.921.7
TSFADet [41]89.967.963.789.854.073.118.6
C²Former [40]90.268.364.489.858.574.2-
SLBAF-Net [7]90.272.068.689.959.976.163.2
Ours90.376.873.390.366.079.433.1
+ +UA-CMDet [29], Halfway Fusion [21], CIAN [44], MBNet [50], AR-CNN [45], TSFADet [41], $C^2$ Former [40], and SLBAF-Net [7]. Among them, UA-CMDet, MBNet, TSFADet, and $C^2$ Former all consider the weakly misalignment problem during the fusion process, which affirms the universality and significance of this problem. For each compared method, we use their original experimental settings to ensure equity. + +Quantitative comparison. The results are shown in Tab. 1. In the unimodal methods, benefitting from a well-designed multiscale feature extraction network, YOLOv5s has comparable detection accuracy (62.1% in RGB images and 70.7% in IR images), even better than some fusion methods. Correspondingly, the multimodal methods SLBAF-Net, constructed upon the two-stream YOLOv5 framework, has also yielded excellent results. Compared with all methods, we can find that our OAFA achieves the best average performance on the validation set, which is 3.3% higher than the second place. It is worth noting that our method has obvious advantages in distinguishing confusing categories (truck and freight-car). The reason may stem from the fact that OAFA excels in identifying the optimal spatial correspondences of multimodal features, obtaining reliable features of the target after fusion. + +Qualitative comparison. Our baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branches are fused with a concatenation operation in multiscale features. Among the compared methods mentioned above, we select the optimal unimodal + +YOLOv5s and the baseline model as the qualitative comparative model. We provide visual detection results of the compared methods in Fig. 3. It can be observed that, given the constraint of the IR labels, the objects are successfully detected by OAFA when the target positions are shifted, while both unimodal and baseline model experience classification and localization errors or fail to detect, especially for occluded and few-shot object detection. The underlying causes can be speculated that the weak misalignment problem confuses the fusing model to learn discriminative features. In contrast, our method has the capability to capture aligned and reliable features, improving the performance in classification and localization processes. + +Speed comparison. We compare the speed of OAFA with other fusion detection methods to prove the efficiency of our method. For a fair comparison, all the detection models are tested with batch size 1, and the input size is set to $640 \times 640$ . As shown in Tab. 1, our proposed OAFA achieves a speed of 33.1 FPS, outperforming most existing multimodal object detection methods. Although our method requires more computational time compared to SLBAF-Net, it would be worthwhile to invest this additional time to achieve superior performance. Furthermore, it can be seen that OAFA can accomplish real-time detection on the NVIDIA RTX A6000 platform. + +# 4.3. Ablation Study + +In this section, we conduct several ablation studies to validate the effectiveness of the different components in our + +proposed method. The results are shown in Tab. 2. + +Effect of CSOM module. We first analyze the effect of the CSOM module as it provides the fundamental offset estimation, which determines the stability of the model training. We perform ablation experiments on both the decoupled multimodal learning (DML) module and the spatial offset modeling (SOM) module separately. As presented in the first to third rows of Tab. 2, the performance of the baseline model is improved by $0.4\%$ and $1.6\%$ after incorporating the SOM and DML modules, respectively. It verifies that both coarse-grained feature alignment and increasing fusion flexibility through decouple learning are effective for the detection task. Besides, when integrating these two submodules, the performance improvement of model over baseline is $2.1\%$ , which indicates that the quality of the feature alignment can be enhanced by mitigating modal gap. + +Effect of ODAF module. We then verify the necessity of the ODAF module in addressing the weak misalignment problem. The results demonstrate that despite ODAF has poor stability in model training without dependable offset guidance, incorporating ODAF into the baseline can provide $+1.7\%$ significant improvement. It is owing to adaptively adjust the optimal sampling position for acquiring reliable information in RGB images. + +Effect of two-stage training. The two-stage training strategy provides comprehensive unimodal representations for multimodal alignment and fusion. We can observe that the accuracy decreases to $77.3\%$ if we directly train the whole network without two-stage training. It is $0.9\%$ and $0.5\%$ lower than exclusively incorporating CSOM and ODAF, respectively. The reason could be that the inappropriate initialization in feature extraction network leads to training instability and makes it hard to obtain abundant representations in spatial and semantic. After undergoing a two-stage training process, our method attains a performance level of $79.4\%$ , surpassing the capabilities of any individual module. It demonstrates that the two-stage training strategy effectively enhances training accuracy and stability. The superior method to fully exploit the strengths of CSOM and ODAF will be further studied in future work. + +All in all, not only each independent module but also the entire proposed framework can contribute to the multimodal object detection under weakly misalignment conditions. + +# 4.4. Robustness to Position Shift + +To quantitatively assess the robustness of our method to the weakly misalignment issue, we conduct experiments on DroneVehicle by manually simulating positional offsets. The tested model is trained on the original training image pairs and validated on the shifted image pairs. In our experiments, we hold the IR images constant while introducing spatial offsets along the $x$ -axis and $y$ -axis for RGB images. The variations in pixel values are defined within the + +Table 2. Ablation study on DroneVehicle dataset. The baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branch are fused with a multiscale concatenation operation. The D denotes DML, S denotes SOM, C denotes CSOM, and O denotes ODAF. + +
MethodCSOMODAFTwo-stage TrainingmAP (%) ↑
DMLSOM
Baseline76.1
Baseline+S✓76.5
Baseline+D✓77.7
Baseline+C✓✓78.2
Baseline+O✓77.8
Baseline+C+O✓✓✓77.3
OAFA✓✓✓✓79.4
+ +![](images/67de58df0de66f1a6722fb231b8b6fbc1826868a1f3cfad115393880f9a7b0f2.jpg) +(a) SLBAF-Net (Baseline) + +![](images/fbc57013de2708b7741397311d829d0152a1d067604b295a0748ffd272f0167d.jpg) +(b) Our method +Figure 5. Surface plot visualization of the position shift experiments. Horizontal coordinates indicate various shift steps along the $x$ -axis and $y$ -axis. Vertical coordinate denotes the detection performances in terms of mAP. + +range $\{(\Delta x, \Delta y) \mid \Delta x, \Delta y \in [15, -15]; \Delta x, \Delta y \in \mathbb{Z}\}$ . The quantitative outcomes of this experiment in baseline and our model are discernible from Fig. 5. It can be seen that in response to larger position shifts, our method consistently exhibits minor fluctuations in performance, whereas the baseline model experiences a noticeable performance decline. This result demonstrates the robustness of OAFA under weakly misalignment conditions. + +# 5. Conclusion + +In this paper, we propose a robust RGB-IR object detection method OAFA for UAVs images under weakly misalignment conditions. Different from previous methods, OAFA can achieve adaptive multimodal feature alignment without strict alignment. To mitigate the impact of the modality gap on multimodal spatial matching, the modality-invariant features are acquired in a common subspace to estimate accurate offsets. Taking the offsets as basic knowledge, the adaptive alignment implicitly captures optimal fusion positions and steadily carries out feature-level corrections with a deformable alignment module. Experimental results on DroneVehicle dataset demonstrate the superior performances of OAFA compared with the state-of-the-art methods. Moreover, our method shows robustness against the weakly misalignment problem in position shift experiments. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586). + +# References + +[1] Gedas Bertasius, Lorenzo Torresani, and Jianbo Shi. Object detection in video with spatiotemporal sampling networks. In ECCV, pages 342-357, 2018. 3 +[2] Ilker Bozcan and Erdal Kayacan. AU-AIR: A multi-modal unmanned aerial vehicle dataset for low altitude traffic surveillance. In ICRA, pages 8504-8510, 2020. 1 +[3] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Dong Nie, Minjeong Kim, Qian Wang, and Dinggang Shen. Deformable image registration based on similarity-steered CNN regression. In MICCAI, pages 300-308, 2017. 2 +[4] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In AAAI, pages 973-981, 2021. 3 +[5] Kelvin C. K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5962-5971, 2022. 3, 5 +[6] Hao Chen, Yongjian Deng, Youfu Li, Tzu-Yi Hung, and Guosheng Lin. RGBD salient object detection via disentangled cross-modal fusion. IEEE TIP, 29:8407-8416, 2020. 5 +[7] Xiaolong Cheng, Keke Geng, Ziwei Wang, Jinhu Wang, Yuxiao Sun, and Pengbo Ding. Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment. Multimedia Tools and Applications, pages 1-20, 2023. 7 +[8] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017. 3 +[9] Chaitanya Devaguptapu, Ninad Akolekar, Manuj M. Sharma, and Vineeth N. Balasubramanian. Borrow from anywhere: Pseudo multi-modal object detection in thermal imagery. In CVPR, pages 1029-1038, 2019. 2, 3 +[10] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In CVPR, pages 2849-2858, 2019. 6, 7 +[11] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE TGRS, 60: 1-11, 2022. 6, 7 +[12] Renlong Hang, Zhu Li, Pedram Ghamisi, Danfeng Hong, Guiyu Xia, and Qingshan Liu. Classification of hyperspectral and lidar data using coupled cnns. IEEE TGRS, 58(7): 4939-4950, 2020. 2, 3 +[13] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. MISA: modality-invariant and -specific representations for multimodal sentiment analysis. In ACM MM, pages 1122-1131, 2020. 2, 3, 4 +[14] Danfeng Hong, Naoto Yokoya, Jocelyn Chanussot, and Xiao Xiang Zhu. Cospace: Common subspace learning from hyperspectral-multispectral correspondences. IEEE TGRS, 57(7):4349-4359, 2019. 2, 3 +[15] Danfeng Hong, Jocelyn Chanussot, Naoto Yokoya, Jian Kang, and Xiao Xiang Zhu. Learning-shared cross-modality representation using multispectral-lidar and hyperspectral data. IEEE GRSL, 17(8):1470-1474, 2020. 2, 3 +[16] Lloyd Haydn Hughes, Diego Marcos, Sylvain Lobry, Devis Tuia, and Michael Schmitt. A deep learning framework for + +matching of sar and optical imagery. ISPRS, 169:166-179, 2020. 2 +[17] Glenn Jocher. ultralytics/yolov5. https://github.com/ultralytics/yolov5, Oct. 2020. 4, 6, 7 +[18] Huafeng Li, Junzhi Zhao, Jinxing Li, Zhengtao Yu, and Guangming Lu. Feature dynamic alignment and refinement for infrared-visible image fusion: Translation robust fusion. Information Fusion, 95:26-41, 2023. 2 +[19] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 6, 7 +[20] Fan Liu, Zhiyong Cheng, Huilin Chen, Anan Liu, Liqiang Nie, and Mohan S. Kankanhalli. Disentangled multimodal representation learning for recommendation. arXiv:2203.05406, 2022. 2, 3 +[21] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Multispectral deep neural networks for pedestrian detection. In BMVC, 2016. 7 +[22] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6 +[23] Sijie Mai, Haifeng Hu, and Songlong Xing. Modality to modality translation: An adversarial representation learning and graph fusion network for multimodal fusion. In AAAI, pages 164-172, 2020. 2, 3 +[24] Sayan Nag. Image registration techniques: A survey. arXiv:1712.07540, 2017. 2 +[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2017. 6, 7 +[26] Marc-Michel Rohé, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: Learning deformable image registration using shape matching. In MIC-CAI, pages 266–274, 2017. 2 +[27] Hazim Shakhatreh, Ahmad Sawalmeh, Ala I. Al-Fuqaha, Zuochao Dou, Eyad Almaita, Issa Khalil, Noor Shamsiah Othman, Abdallah Khreishah, and Mohsen Guizani. Unmanned aerial vehicles (uavs): A survey on civil applications and key research challenges. IEEE Access, 7:48572-48634, 2019. 1 +[28] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 536-553, 2018. 3 +[29] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning. IEEE TCSVT, 32(10):6700-6713, 2022. 1, 5, 7 +[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally-deformable alignment network for video super-resolution. In CVPR, pages 3357-3366, 2020. 3 +[31] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In IJCAI, pages 3508-3515, 2022. 2 +[32] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In ICCV, pages 3622–3631, 2019. 2, 3 + +[33] Hua Wang, Dewei Su, Chuangchuang Liu, Longcun Jin, Xianfang Sun, and Xinyi Peng. Deformable non-local network for video super-resolution. IEEE Access, 7:177734-177744, 2019. 3 +[34] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. In CVPR, pages 1954-1963, 2019. 3 +[35] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In ECCV, pages 3–19, 2018. 4 +[36] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented R-CNN for object detection. In ICCV, pages 3500-3509, 2021. 6, 7 +[37] Han Xu and Jiayi Ma. Emfusion: An unsupervised enhanced medical image fusion network. Information Fusion, 76:177-186, 2021. 2, 3 +[38] Xing Xu, Kaiyi Lin, Lianli Gao, Huimin Lu, Heng Tao Shen, and Xuelong Li. Learning cross-modal common representations by private-shared subspaces separation. IEEE Trans. Cybern., 52(5):3261-3275, 2022. 2, 3 +[39] Huang Yao, Rongjun Qin, and Xiaoyu Chen. Unmanned aerial vehicle for remote sensing applications - A review. Remote Sensing, 11(12):1443, 2019. 1 +[40] Maoxun Yuan and Xingxing Wei. $\mathbf{C}^2$ former: Calibrated and complementary transformer forrgb-infrared object detection. arXiv:2306.16175, 2023. 1, 7 +[41] Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In ECCV, pages 509-525, 2022. 1, 2, 7 +[42] Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschlager, and Susanne Saminger-Platz. Central moment discrepancy (CMD) for domain-invariant representation learning. In ICLR, 2017. 4 +[43] Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Grid anchor based image cropping: A new benchmark and an efficient model. IEEE TPAMI, 44(3):1304-1319, 2022. 1 +[44] Lu Zhang, Zhiyong Liu, Shifeng Zhang, Xu Yang, Hong Qiao, Kaizhu Huang, and Amir Hussain. Cross-modality interactive attention network for multispectral pedestrian detection. Information Fusion, 50:20-29, 2019. 7 +[45] Lu Zhang, Zhiyong Liu, Xiangyu Zhu, Zhan Song, Xu Yang, Zhen Lei, and Hong Qiao. Weakly aligned feature fusion for multimodal object detection. arXiv:2204.09848, 2022. 2, 7 +[46] Xinyue Zhang, Chengcai Leng, Yameng Hong, Zhao Pei, Irene Cheng, and Anup Basu. Multimodal remote sensing image registration methods and advancements: A survey. Remote Sensing, 13(24):5128, 2021. 2 +[47] Zhimeng Zhang and Yu Ding. Adaptive affine transformation: A simple and effective operation for spatial misaligned image generation. In ACM MM, pages 1167-1176, 2022. 1 +[48] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, 2023. 4 + +[49] Jiaxiang Zheng, Yichen Tian, Chao Yuan, Kai Yin, Feifei Zhang, Fangmiao Chen, and Qiang Chen. Mdesnet: Multitask difference-enhanced siamese network for building change detection in high-resolution remote sensing images. Remote Sensing, 14(15):3775, 2022. 4 +[50] Kailai Zhou, Linsen Chen, and Xun Cao. Improving multispectral pedestrian detection by addressing modality imbalance problems. In ECCV, pages 787-803, 2020. 7 +[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3 +[52] Barbara Zitová and Jan Flusser. Image registration methods: a survey. Image Vis. Comput., 21(11):977-1000, 2003. 2 \ No newline at end of file diff --git a/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/images.zip b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a862f8794f60fa892d12d0edbe3acda21999cc3b --- /dev/null +++ b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0ca93e3e5a57a98cb358e825a6695d858cff0570c4a1a007b83f74be1d83386 +size 703272 diff --git a/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/layout.json b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae496e1c0f1aedc2466e8ed4e029005c244d5109 --- /dev/null +++ b/2024/Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection/layout.json @@ -0,0 +1,9156 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 518, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 518, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 518, + 140 + ], + "type": "text", + "content": "Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 161, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 545, + 190 + ], + "type": "text", + "content": "Chen Chen Jiahao Qi Xingyue Liu Kangcheng Bin Ruigang Fu Xikun Hu Ping Zhong* National University of Defense Technology, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 191, + 528, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 528, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 528, + 217 + ], + "type": "text", + "content": "chenchen21c@nudt.edu.cn, qijiahao1996@nudt.edu.cn, liuxingyuel8@nudt.edu.cn, binkc21@nudt.edu.cn, furuigang08@nudt.edu.cn, xikun@nudt.edu.cn, zhongping@nudt.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 271, + 290, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 536 + ], + "type": "text", + "content": "Visible-infrared (RGB-IR) image fusion has shown great potentials in object detection based on unmanned aerial vehicles (UAVs). However, the weakly misalignment problem between multimodal image pairs limits its performance in object detection. Most existing methods often ignore the modality gap and emphasize a strict alignment, resulting in an upper bound of alignment quality and an increase of implementation costs. To address these challenges, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA), which could adaptively adjust the relative positions between multimodal features. Considering the impact of modality gap on the cross-modality spatial matching, a Cross-modality Spatial Offset Modeling (CSOM) module is designed to establish a common subspace to estimate the precise feature-level offsets. Then, an Offset-guided Deformable Alignment and Fusion (ODAF) module is utilized to implicitly capture optimal fusion positions for detection task rather than conducting a strict alignment. Comprehensive experiments demonstrate that our method not only achieves state-of-the-art performance in the UAVs-based object detection task but also shows strong robustness to the weakly misalignment problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 691 + ], + "type": "text", + "content": "Visible-infrared (RGB-IR) object detection on unmanned aerial vehicles (UAVs) has attracted extensive attention [2, 27, 39], due to its ability to make full use of complementary information and achieve robust around-the-clock detection. However, the collected RGB-IR images are typically mismatched in resolutions and field of views owing to disparate imaging principles. Thus, it is challenging to fuse the corresponding information of target between multimodal images." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 243, + 426, + 299 + ], + "blocks": [ + { + "bbox": [ + 307, + 243, + 426, + 299 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 426, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 426, + 299 + ], + "type": "image", + "image_path": "52d5000f81f66fa61293d1f6c6fcd59c797f57cead50e577bba722d9cd0f30ae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 302, + 412, + 310 + ], + "lines": [ + { + "bbox": [ + 326, + 302, + 412, + 310 + ], + "spans": [ + { + "bbox": [ + 326, + 302, + 412, + 310 + ], + "type": "text", + "content": "(a) Image-level alignment method" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 244, + 545, + 299 + ], + "blocks": [ + { + "bbox": [ + 440, + 244, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 440, + 244, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 440, + 244, + 545, + 299 + ], + "type": "image", + "image_path": "a9e37ab502cb93b5f1a75269cee6c2b42d1fb070dcc4e98601546f23db00fb74.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 301, + 537, + 309 + ], + "lines": [ + { + "bbox": [ + 446, + 301, + 537, + 309 + ], + "spans": [ + { + "bbox": [ + 446, + 301, + 537, + 309 + ], + "type": "text", + "content": "(b) Feature-level alignment method" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 333, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 474, + 333, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 474, + 333, + 541, + 342 + ], + "type": "text", + "content": ":Multimodal Features" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 316, + 462, + 382 + ], + "blocks": [ + { + "bbox": [ + 315, + 316, + 462, + 382 + ], + "lines": [ + { + "bbox": [ + 315, + 316, + 462, + 382 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 462, + 382 + ], + "type": "image", + "image_path": "734ae5bdfa091c01da8da7917ea0f75ae7f619750ef3858c94c257b5f71a79a6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 385, + 410, + 392 + ], + "lines": [ + { + "bbox": [ + 372, + 385, + 410, + 392 + ], + "spans": [ + { + "bbox": [ + 372, + 385, + 410, + 392 + ], + "type": "text", + "content": "(c) Our OFA" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 320, + 527, + 330 + ], + "lines": [ + { + "bbox": [ + 471, + 320, + 527, + 330 + ], + "spans": [ + { + "bbox": [ + 471, + 320, + 527, + 330 + ], + "type": "text", + "content": "(F) : Fusion Module" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 474, + 346, + 533, + 359 + ], + "lines": [ + { + "bbox": [ + 474, + 346, + 533, + 359 + ], + "spans": [ + { + "bbox": [ + 474, + 346, + 533, + 359 + ], + "type": "text", + "content": "Modality-specific Features" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 474, + 361, + 537, + 380 + ], + "lines": [ + { + "bbox": [ + 474, + 361, + 537, + 380 + ], + "spans": [ + { + "bbox": [ + 474, + 361, + 537, + 380 + ], + "type": "text", + "content": "Modality-invariant Features" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 487, + 380, + 534, + 388 + ], + "lines": [ + { + "bbox": [ + 487, + 380, + 534, + 388 + ], + "spans": [ + { + "bbox": [ + 487, + 380, + 534, + 388 + ], + "type": "text", + "content": "Deformation Field" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 304, + 402, + 547, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 535 + ], + "type": "text", + "content": "Figure 1. Previous multimodal object detection architectures under weakly misalignment conditions vs. our OAFA. (a) Image-level alignment method: executing a pixel-level strict alignment by explicitly extracting the global deformation fields from multimodal images. (b) Feature-level alignment method: explicitly estimating feature offsets through multimodal features to achieve strict feature alignment. It can be further classified into global and local methods based on whether alignment is exclusively performed on target regions. (c) Our OAFA: the basic global offsets are estimated by the modality-invariant features to eliminate the impact of modality gap. Based on this offset, we perform adaptive feature alignment without strict alignment." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": "To increase the quality of fusion representations, researchers always apply pre-registration techniques, involving image cropping [43] and affine transformation [47] before the fusing process. Nevertheless, since distinct imaging time in multimodal sensors and intricate movements in targets, it is difficult, or even impossible, to achieve precise UAVs image alignment [29]. It inevitably leads to local inconsistencies in multimodal spatial distribution, named the weakly misalignment problem [41], resulting in a mismatch of the multimodal target features in corresponding positions [40]. Since the detection task is sensitive to spatial locations, the weakly misalignment problem would lead to significant performance degradation. Several studies have been carried out to deal with this issue by accomplishing" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26836" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "a strict alignment based on offset estimation, as shown in Fig. 1a and Fig. 1b. Although these methods have made significant progress, there still remain two main challenges:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 242 + ], + "type": "text", + "content": "Inaccurate offset estimation. Conventional methods devote to predicting spatial offsets by identifying the distribution and appearance discrepancies in image or feature level. However, in fact, the discrepancies arise not only from spatial offsets but also from modality gap [31], which stems from the differences in imaging principles and wavelengths between RGB and IR sensors. This gap could make it hard to correctly match cross-modality representations in spatial, leading to misleading offset estimation. Hence, the first challenge is how to bridge the modality gap to achieve accurate cross-modal offset estimation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 399 + ], + "type": "text", + "content": "High dependency on strict alignment. The strict alignment is necessary for most existing methods. Unfortunately, the strict alignment always requires additional annotation information, such as the ground truth offsets between multimodal representations [45]. It will lead to higher implementation costs that limit real-world applications. Moreover, the detection performance relies heavily on the quality of alignment. However, even the state-of-the-art alignment methods have the risk of failure in some scenarios [18], which may cause the error delivery to the detection task. Consequently, the second challenge is how to enhance the robustness of the detection model to the weakly misalignment problem without strict alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 401, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 604 + ], + "type": "text", + "content": "To this end, the purpose of this paper is to achieve adaptive alignment between RGB (sensed modality) and IR (reference modality) images based on spatial offsets for superior detection performance instead of strict alignment. Technically, we propose a novel method named Offset-guided Adaptive Feature Alignment (OAFA) for UAVs-based multimodal object detection under weakly misalignment conditions. It mainly consists of a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module for dealing with the two challenges, respectively. Considering the first challenge, CSOM establishes a cross-modal common subspace to obtain the modality-invariant features for eliminating modality gap. As distributional aligned mappings, the modality-invariant features could be utilized to evaluate spatial consistency between multimodal features for more accurate cross-modal offsets estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "As for the second challenge, ODAF is designed to adaptively adjust feature positions by implicit offset compensation, aiming to make full use of the offsets and achieve superior detection performance without strict alignment. In specific, ODAF involves a learnable deformation component that dynamically predicts the offsets of sampling convolution kernels. With the help of the offsets provided by CSOM, ODAF captures optimal sampling positions in sensed modality features to attain comprehensive target rep" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "resentations after fusion. This module predicts the corresponding sampling positions for diverse regions in different image pairs, contributing to a strong flexibility of OAFA in handling diverse and complex offsets. Besides, the training stability of deformable components is improved since the explicit guidance of the offsets reduces the burden of the adaptive alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 157, + 546, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 335 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: (1) We propose a robust RGB-IR object detection method for UAVs images under weakly misalignment conditions, effectively implementing adaptive feature alignment instead of strict alignment. (2) Considering the difficulty of estimating cross-modality offset, we build a multimodal common subspace to reduce the impact of modality gap on multimodal spatial matching. (3) An offset-guided deformable alignment module is exploited to achieve adaptive alignment by implicitly capturing optimal fusion positions, enhancing the accuracy of the detection task. (4) Extensive experiments on public dataset show that our OAFA is superior to state-of-the-art models only with a simple fusion strategy. Moreover, it proves the effectiveness in addressing the weakly misalignment problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 346, + 392, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 392, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 392, + 358 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 366, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 617 + ], + "type": "text", + "content": "Multimodal object detection under weakly misalignment conditions. A general solution to address the weakly misalignment problem is to conduct pixel-level prealignment [24, 46, 52], which learn a global mapping function from multimodal images to correct the pixel coordinates. Many of these methods [3, 16, 26] are supervised with the deformable fields in order to achieve strict prealignment. To achieve end-to-end training, Zhang et al. [45] first introduced a data-driven region feature alignment module to predict proposal regions offsets from multimodal features, which was inexpensively constrained by the target locations in RGB-IR images. On this basis, Yuan et al. [41] made full use of the oriented multimodal annotations to realize multimodal target alignment in position, size, and angle. Due to the requirement of extra supervisions for the strict alignment process, the above methods are labor-intensive and costly to implement. Besides, the existing modality gap at both the pixel-level and feature-level leads to incorrect estimation of spatial offsets. Consequently, it is indeed significant to propose a novel method for RGB-IR object detection under weakly misalignment conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Common subspace learning. Common subspace learning aims to exploit shared information among multiple data. In field of multimodal learning, it can be used to bridge the modality gap by extracting the unified latent representations in the inter-modality samples. Works can be broadly grouped into transformation-based [9, 23, 32, 37], weight-sharing-based [12, 14, 15], and decoupling-based methods [13, 20, 38]. The transformation-based methods" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26837" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 278 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 278 + ], + "type": "image", + "image_path": "3a58246ba63d96996c9784b32e94c2344362e96ea3525ababf77e20b3249e5b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 280, + 542, + 321 + ], + "blocks": [ + { + "bbox": [ + 49, + 280, + 542, + 321 + ], + "lines": [ + { + "bbox": [ + 49, + 280, + 542, + 321 + ], + "spans": [ + { + "bbox": [ + 49, + 280, + 542, + 321 + ], + "type": "image", + "image_path": "e040a3eeb21d4a404706801326bac3d5f93911841f87e41535c7eb0b0c4b4eb4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 330, + 546, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 546, + 396 + ], + "type": "text", + "content": "Figure 2. An Overview of OAFA. It consists of two components: a Cross-modality Spatial Offset Modeling (CSOM) module and an Offset-guided Deformable Alignment and Fusion (ODAF) module. In our implementation, we choose the RGB images as the sensed modality whose features are shifted to achieve alignment, and the IR images as the reference modality. The CSOM can decouple the original multimodal features to modality-invariant and modality-specific features and predict a superior spatial offset in a common subspace. The ODAF utilizes this basic offset to adaptively align the sample positions of the RGB features with IR features, and then fuses the aligned RGB and original IR features. Finally, the fused features are conducted to generate the ultimate detection results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 403, + 289, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 619 + ], + "type": "text", + "content": "generally transform one modality (images/features) into another modality by domain transfer [9] or auto-encoder [23, 32, 37]. It is hard to achieve an end-to-end training due to the requirement of a pre-trained transformation network. The weight-sharing-based methods devote to designing the weight-shared network within a two-stream framework to only capture modality-invariant features [12, 14, 15]. Nevertheless, overemphasis on common features may result in the underutilization of complementary information. The decoupling-based methods overcome this obstacle by trying to separate multimodal features into modality-invariant and modality-specific features with a series of constraint functions [13, 20, 38]. Inspired by the third category methods, we develop a decoupled multimodal learning network with three constraints for the detection task. Moreover, due to the complexity of UAVs-based multimodal features, we design the network with partial weight sharing and custom encoder architecture to achieve effective disentanglement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Deformable convolution. Convolutional neural networks (CNNs) are inherently limited in modeling geometric transformations because of their fixed kernel structures. To enhance the transformation modeling capability of CNNs, Dai et al. [8] designed the deformable convolution, which has shown superior performance in several high-level vision tasks [1, 28, 51]. In recent years, it has also been" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 403, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 536 + ], + "type": "text", + "content": "used in the field of video super-resolution [5, 30, 33, 34]. As a pioneer work, Tian et al. [30] adopted the deformable convolution to align the features of adjacent frames in the temporal domain. It demonstrates that deformable convolution can effectively capture the motion cues of the targets. Nonetheless, Chan et al. [4] has proved that the deformable alignment is difficult to train. Training instability frequently leads to offset overflow, degrading the final performance in downstream task. Therefore, we construct a basic offset as explicit guidance to reduce the burden of deformable alignment training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 544, + 361, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 361, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 361, + 555 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 563, + 373, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 373, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 373, + 574 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Our method aims to enhance the accuracy and robustness of the RGB-IR object detection model under weakly misalignment conditions. As shown in Fig. 2, a pair of weakly misaligned RGB-IR images are fed into a two-stream network to acquire the multiscale multimodal representations. Subsequently, a decoupled learning network is designed to develop a common subspace to figure out the modality-invariant and modality-specific features. To mitigate the influence of modality gap on offset estimation, we exclusively employ the modality-invariant features to predict spatial offsets. Then, this basic offsets are utilized to implica" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26838" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "ity adjust the spatial positions of the modality-invariant and modality-specific features. It accomplishes adaptive feature alignment that is tailored for RGB-IR object detection task without any extra supervisions. Finally, the aligned RGB and original IR features are fused with a concatenation operation to generate final detection results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 149, + 287, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 287, + 174 + ], + "type": "text", + "content": "3.2. Cross-modality Spatial Offset Modeling Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": "Given the RGB features " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "F_{rgb}" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": " and IR features " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "F_{ir}" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": ", CSOM predicts the spatial offsets " + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 180, + 287, + 239 + ], + "type": "text", + "content": " as basic knowledge for the subsequent alignment module. The CSOM consists of a decoupled multimodal learning submodule and a spatial offset modeling submodule." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": "Decoupled multimodal learning. This submodule mainly builds a common subspace " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbb{C}} = [\\mathbb{C}_{rgb},\\mathbb{C}_{ir}]" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " and a specific subspace " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbb{S}} = [\\mathbb{S}_{rgb},\\mathbb{S}_{ir}]" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": ", as shown in Fig. 2a. For each input RGB-IR image pair " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{rgb}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ir}\\in \\mathbb{R}^{H\\times W\\times 1}" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": ", we map it to the multimodal representations " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m(m\\in \\{rgb,ir\\})" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " by a basic pyramid feature extractor " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_m" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " is the bounded random samples on the interval " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "[a,b]^N" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": ". Then, the modality-invariant features " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^c" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " and modality-specific features " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^s" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " projected from " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbb{C}}" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 46, + 240, + 288, + 361 + ], + "type": "text", + "content": " are figured out by the encoding functions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 366, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 366, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 81, + 366, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {m} ^ {s} = \\mathcal {S} _ {m} \\left(\\mathcal {F} _ {m}; \\theta_ {m} ^ {s}\\right), \\mathcal {F} _ {m} ^ {c} = \\mathcal {C} _ {m} \\left(\\mathcal {F} _ {m}; \\theta^ {c}\\right), \\tag {1}", + "image_path": "f52041ee5734cec2f8a52618e44437b8383eff7c28f084e441cbc48ffd0a4a98.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": "where the invariant feature encoder " + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_m" + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": " shares the parameters " + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": " across two modalities, whereas the specific feature encoder " + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_m" + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": " assigns separate parameters " + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\theta_m^s" + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": " for each modality. Considering the diversity information in UAVs images, we not only design the decoupled network in objective function but also in network architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "content": "In terms of the network architecture, an accepted assumption is that the input features are more related in the low layers, while the high-layer features tend to be uncorrelated [48]. Thus, " + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_m" + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "content": " is constructed with a convolutional layer followed by the SiLU activation function. In contrast, " + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "content": " contains a convolutional layer and a C3 module with three convolutional and a series of bottleneck layers to obtain deeper features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "content": "As for the objective functions, inspired by [13], we apply three types of constraints for decoupled multimodal learning: Similarity Loss, Difference Loss, and Semantic Loss. First, we employ a similarity loss " + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "content": " to minimize the distance between " + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^{c}" + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}^{c}" + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "content": " to learn common subspace " + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbb{C}}" + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 46, + 553, + 287, + 651 + ], + "type": "text", + "content": " is defined by a state-of-the-art distance metric named central moment discrepancy [42] to measure the distribution dissimilarity between two features:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 655, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 655, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 655, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s i m} = \\frac {1}{| b - a |} \\left\\| E \\left(\\mathcal {F} _ {r g b} ^ {c}\\right) - E \\left(\\mathcal {F} _ {i r} ^ {c}\\right) \\right\\| _ {2} \\\\ + \\sum_ {k = 2} ^ {K} \\frac {1}{| b - a | ^ {k}} \\left\\| C _ {K} \\left(\\mathcal {F} _ {r g b} ^ {c}\\right) - C _ {K} \\left(\\mathcal {F} _ {i r} ^ {c}\\right) \\right\\| _ {2}, \\tag {2} \\\\ \\end{array}", + "image_path": "ea2071edb7ecdb49f267496e04325ea32c7547197fdca7b7541c335163207438.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " is the empirical expectation vector, " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "C_K(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " is the vector of " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "-th order sample central moments, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " represents the Euclidean norm." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": "Second, we present a difference loss " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dif}" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " for highlighting the dissimilarity between " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^{s}" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}^{s}" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " to capture complementary information. Additionally, as " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^c" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^s" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " are expected to be mutually independent, " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dif}" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " should be extended to them as well. These discrepancies are achieved by the squared Frobenius norm:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 186, + 545, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 545, + 239 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d i f} = \\sum_ {m \\in \\{r g b, i r \\}} \\sqrt {\\operatorname {T r} \\left(\\left(\\left(F _ {m} ^ {s}\\right) ^ {T} F _ {m} ^ {c}\\right) ^ {T} \\left(\\left(F _ {m} ^ {s}\\right) ^ {T} F _ {m} ^ {c}\\right)\\right)} \\tag {3} \\\\ + \\sqrt {T r ((\\left(F _ {r g b} ^ {s}\\right) ^ {T} F _ {i r} ^ {s}) ^ {T} (\\left(F _ {r g b} ^ {s}\\right) ^ {T} F _ {i r} ^ {s}))}, \\\\ \\end{array}", + "image_path": "2791dee25b71f544a3d41283532c1c66162269af6b12c44dbf464342be20a40d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 243, + 446, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 446, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 446, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 243, + 446, + 255 + ], + "type": "inline_equation", + "content": "Tr(\\cdot)" + }, + { + "bbox": [ + 306, + 243, + 446, + 255 + ], + "type": "text", + "content": " is trace of the matrix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "text", + "content": "Third, only restrained by the above objectives, there exists a potential risk that " + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_m" + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "text", + "content": " learns useless noisy cues, whereas " + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "inline_equation", + "content": "S_m" + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "text", + "content": " acquires almost all modality information. To address this issue, we add a semantic loss " + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "text", + "content": " to enforce " + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_m" + }, + { + "bbox": [ + 304, + 256, + 545, + 315 + ], + "type": "text", + "content": " to learn discriminative features:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 320, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} = \\sum_ {m \\in \\{r g b, i r \\}} \\left(\\mathcal {L} _ {c l s} \\left(\\mathcal {F} _ {m} ^ {c}\\right) + \\mathcal {L} _ {r e g} \\left(\\mathcal {F} _ {m} ^ {c}\\right) + \\mathcal {L} _ {o b j} \\left(\\mathcal {F} _ {m} ^ {c}\\right)\\right), \\tag {4}", + "image_path": "a4b285487af57560942b55afccf6ff2674c0c2fe000f15fc9673d784411ac2db.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 304, + 358, + 545, + 429 + ], + "type": "text", + "content": " are the detection loss functions of the YOLOv5 [17]. It should be noted that this object detection branch, serving as an auxiliary task, is not incorporated during the inference stage to enhance the efficiency of our OAFA. Finally, the total loss of the decoupled multimodal learning submodule can be defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 437, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 437, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 347, + 437, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e c} = \\lambda_ {1} \\mathcal {L} _ {s i m} + \\lambda_ {2} \\mathcal {L} _ {s i m} + \\lambda_ {3} \\mathcal {L} _ {s e m}, \\tag {5}", + "image_path": "e5eef92315bbee0dc55c7e33b6adf690b9c04057448a68ccc6fe67d9b4cbd0f7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "text", + "content": " is the trade-off weights. In this study, we set " + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 304, + 456, + 545, + 479 + ], + "type": "text", + "content": " to 0.03, 0.01, and 0.03 respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "content": "Spatial offset modeling. We model the spatial offsets " + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "content": " by the estimating spatial differences between " + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^c" + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}^c" + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "content": ". In the field of change detection [49], the differences are calculated directly by subtracting or concatenating two features. However, it would introduce significant noises when dealing with intricate backgrounds in UAVs images. To solve this problem, a feature difference enhancement module is presented to filter out irrelevant changes and concentrate on shift objects. As shown in Fig. 2b, " + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^c" + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "content": " is firstly fed into a spatial attention module to capture the representations of the target and crucial background areas in each modality. The structure of the spatial attention module is the same as [35]. In this case, the spatially enhanced unimodal feature " + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "inline_equation", + "content": "F_m^{c,p}" + }, + { + "bbox": [ + 304, + 479, + 545, + 647 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 652, + 545, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 652, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 311, + 652, + 545, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {m} ^ {c, p} = \\sigma \\left(f _ {7 \\times 7} \\left(\\operatorname {C a t} \\left(\\operatorname {M a x} \\left(\\mathcal {F} _ {m} ^ {c}\\right), \\operatorname {M e a n} \\left(\\mathcal {F} _ {m} ^ {c}\\right)\\right)\\right)\\right) \\odot \\mathcal {F} _ {m} ^ {c}, \\tag {6}", + "image_path": "081da7df7747f188fc1a2f2ba6bcdc4fa6ff842df9d2df6edb901718e2a6e90e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the sigmoid function, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{i\\times j}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "i\\times j" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " convolution layer, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "Cat(\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " refers to the concatenation operation, and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " indicates the Hadamard product operation." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26839" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "We conduct a concatenation instead of subtraction strategy to extract the changing features " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F_{dif}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " to mitigate the influence of complex background. Then, a channel difference enhanced module is introduced to further suppress channel noise. Specifically, we employ simultaneous average-pooling and max-pooling to generate distinct spatial context descriptors. Following this, channel attention can be derived by transmitting the spatial context descriptors to a shared multilayer perceptron network " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "f_{mlp}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^{c,p}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " are combined with the channel attention map to get the spatial-channel enhanced features " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^{c,e}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". The above operation is represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 224, + 287, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 224, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 287, + 255 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} \\mathcal {F} _ {m} ^ {c, e} = \\sigma (f _ {m l p} (A v g P o o l (\\mathcal {F} _ {d i f})) + \\\\ f _ {m l p} (M a x P o o l (\\mathcal {F} _ {d i f}))) \\times \\mathcal {F} _ {m} ^ {c, p} + \\mathcal {F} _ {m} ^ {c, p}. \\end{array} \\tag {7}", + "image_path": "20ce668b253ca97aea1e0d10a5036e83dbf5669d2cfd796213db341120aa613c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "text", + "content": " is obtained through a fusion of the " + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^{c,e}" + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}^{c,e}" + }, + { + "bbox": [ + 47, + 262, + 287, + 289 + ], + "type": "text", + "content": " along with a nonlinear layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "text", + "content": "3.3. Offset-guided Deformable Alignment and Fusion Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "content": "We have obtained the offsets " + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "content": " from CSOM, which represents the spatial differences between " + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}" + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}" + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "content": ". Inspired by [5], the " + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "content": " can be an effectively initialized offset to unsupervised alignment task. Hence, ODAF is designed to adaptively capture optimal fusion location with the help of the basic knowledge " + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 326, + 287, + 433 + ], + "type": "text", + "content": ", incorporating desired complementary information for the subsequent detection task. It is composed of two submodules: offset-guided deformable alignment and decoupled feature fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": "Offset-guided deformable alignment. We mainly apply a deformable convolution to achieve implicit offset compensation and adaptive alignment for the detection task. Unlike the default deformable convolution that learns offsets of convolution kernels from their original features, the offsets generated from our module are mapped from the basic offsets " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": ", which allows for more effective feature correction and reliable model training. Considering different contributions from each offset region for the detection task, we add a modulation scalar " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "\\Delta m_k" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": " that is learned from " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": " to dynamically aggregate information around the corresponding position " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": ". As shown in Fig. 2c, given the center sampling value " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "x(p)" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": " in the sensed features " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^n (n\\in \\{s,c\\})" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": " and the basic offsets " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": ", each counterpart " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "y(p)" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": " in aligned features " + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^{n,a}" + }, + { + "bbox": [ + 46, + 434, + 288, + 616 + ], + "type": "text", + "content": " can be obtained as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 624, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 624, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 63, + 624, + 287, + 658 + ], + "type": "interline_equation", + "content": "y (p) = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\cdot x \\left(p + \\phi_ {c} + p _ {k} + \\Delta p _ {k}\\right) \\cdot \\Delta m _ {k}, \\tag {8}", + "image_path": "41b4546e0d8f4f11949b5219298232daa204997f4535baf2ba518cba2e852ff4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_k" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " denote the number of kernel weights, the " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "-th fixed offset, and the " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "-th kernel weight, respectively. For instance, " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_k \\in \\{(-1,1), (-1,0), \\dots, (1,1)\\}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " donates a regular grid of a " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " kernel defined with " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "K = 9" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Delta p_{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " denotes the implicit compensation to " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "-th location. It is estimated with " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\operatorname{conv}(x)[p + \\phi_c]" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " can be regarded as a prior of " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Delta p_{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " in the standard deformable convolution. Besides, since the location " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "p + \\phi_c + p_k + \\Delta p_k" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " may be fractional, we adopt bilinear interpolation to obtain the final sampling values." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": "Decoupled feature fusion. In traditional fusion process, " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^c" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^s" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " are directly concatenated to obtain the fused features " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_f" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " [6]. Unfortunately, the similarity constraint between " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^{c,a}" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}^{c}" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " leads to much redundancy. To eliminate duplicate information and increase discriminative representations, " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}^{c,a}" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ir}^{c}" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " should be combined and optimized before fusion. Thus, the fused features " + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "inline_equation", + "content": "F_{f}" + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 236, + 545, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 236, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 331, + 236, + 545, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {f} = C a t \\left(f _ {1 \\times 1} \\left(C a t \\left(\\mathcal {F} _ {r g b} ^ {c, a}, \\mathcal {F} _ {i r} ^ {c}\\right)\\right), \\mathcal {F} _ {r g b} ^ {s, a}, \\mathcal {F} _ {i r} ^ {s}\\right). \\tag {9}", + "image_path": "0e1851b6eb999831607c27c950b5fb3ae4e262a13964b0098e9a5d9e63f2572a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 257, + 402, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 402, + 270 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 276, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 371 + ], + "type": "text", + "content": "Training loss. The overall method is based on a supervised detection framework. The purpose of our multimodal alignment is not to form a strict alignment but to achieve optimal adjustment between multimodal features to improve the performance of the subsequent detection task. Thus, the alignment subtask is integrated into the main detection task instead of adding extra alignment loss. The total loss function of our OAFA can be represented as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 380, + 545, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 380, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 348, + 380, + 545, + 394 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {d e c}}. \\tag {10}", + "image_path": "9eda755004300859d59012b1ccc44a5a511afc803d3cdbd80c13b919bf066fec.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": "Two-stage training strategy. A premise for effective feature alignment is that " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^c" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " contains sufficient spatial information, thereby estimating more accurate basic offsets. To this end, we adopt a two-stage learning strategy to train our OAFA end-to-end. In stage I, to gain effective " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m^c" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": ", we try to ensure that pre-decoupling features " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " contain more comprehensive target representations. Technically, we only train the feature extraction and fusion module with " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": ". Specifically, a pair of " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{rgb}" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ir}" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " are fed into a two-stream network to extract multiscale features " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " and gain fusion features " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_f" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " with the concatenation operation. Then, the " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_f" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " is put into the detection network to obtain the detection results. In stage II, we apply all the proposed modules to our network. In this process, the unimodal feature extraction network " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_m" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": " is initialized with the weights from stage I. Then, the joint optimization of the unimodal feature extraction and multimodal feature alignment network is carried out effectively with " + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 628, + 388, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 388, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 388, + 641 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 432, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 432, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 432, + 660 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Datasets and metrics. We evaluate our method on DroneVehicle dataset [29], which is a large-scale multimodal dataset for vehicle detection in UAVs images. It contains 28,439 image pairs with 953,087 vehicles, covering" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26840" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 546, + 354 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 546, + 354 + ], + "type": "image", + "image_path": "bd41082a33f17763e39345eb3181f5ac4ff14cc73f3dc3e9d63feb82d70709a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 356, + 546, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 546, + 401 + ], + "type": "text", + "content": "Figure 3. Four Examples of detection results on the test set of DroneVehicle dataset under weakly misalignment conditions. The confidence threshold is set to 0.25. The results of fusion methods are visual in IR images to correspond to the supervisory label. Red, green, and pink rectangles represent car, freight-car, and truck targets, respectively. Objects correctly detected are represented in green dashed circles, while incorrectly detected objects are represented in red dashed circles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "text", + "content": "different viewing angles, heights, lighting conditions, and scenarios. The dense-oriented annotations are made for five categories: car, truck, bus, van, and freight-car. Though the DroneVehicle performs pre-registration on each RGB-IR image pair, the weakly misalignment problem still exists. As shown in Fig. 4, about " + }, + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "text", + "content": " bounding boxes exhibit deviation issues in DroneVehicle dataset, and the shift distance mostly ranges from 0 to 15 pixels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 505, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 577 + ], + "type": "text", + "content": "This dataset is divided into 17,990 image pairs for training, 1,469 image pairs for validating, and the remaining 8,980 image pairs for testing. Following the standard evaluation metrics, we report the mean average precision (mAP) with intersection over union (IoU) threshold of 0.4 on the validation set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Training details. The experiments are carried out on a single NVIDIA RTX A6000 GPU with 48 GB of memory. We implement our algorithm with Pytorch toolbox and SGD optimizer with a momentum of 0.937 and a weight decay of 0.0005. The initial learning rate is set to 0.01 and eventually reduced to 0.002 by cosine annealing [22]. The input size of our model is set to " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " in the preprocessing stage and the batch size is 16. The training epoch is set to 150 with 50 and 100 epochs in the first and second stages, respectively. We use the ground truth of the IR images as the training label, owning to the more comprehensive target" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 403, + 417, + 495 + ], + "blocks": [ + { + "bbox": [ + 313, + 403, + 417, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 417, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 417, + 495 + ], + "type": "image", + "image_path": "01aa374e69d58cd5caafd5adafb1d331df9e6bc81dc5ebb2a123a52f0f4137bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 498, + 418, + 508 + ], + "lines": [ + { + "bbox": [ + 307, + 498, + 418, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 418, + 508 + ], + "type": "text", + "content": "(a) Statistics of misalignment ratio" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 418, + 413, + 548, + 495 + ], + "blocks": [ + { + "bbox": [ + 418, + 413, + 548, + 495 + ], + "lines": [ + { + "bbox": [ + 418, + 413, + 548, + 495 + ], + "spans": [ + { + "bbox": [ + 418, + 413, + 548, + 495 + ], + "type": "image", + "image_path": "db56953588ad45889834f1bad95b0f82c6270f50f4d8a93064b93921e6e34bf8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 498, + 525, + 507 + ], + "lines": [ + { + "bbox": [ + 439, + 498, + 525, + 507 + ], + "spans": [ + { + "bbox": [ + 439, + 498, + 525, + 507 + ], + "type": "text", + "content": "(b) Statistics of shift value" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 512, + 545, + 556 + ], + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 556 + ], + "type": "text", + "content": "Figure 4. The statistics of ground truth bounding-boxes misalignment ratio and shift value within RGB-IR image pairs on DroneVehicle dataset. annotations available in the IR modality." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 563, + 427, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 427, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 427, + 576 + ], + "type": "text", + "content": "4.2. Results Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Compared methods. We compare our method with the state-of-the-art unimodal and multimodal object detection methods. Unimodal methods: We choose a series of oriented object detection methods for comparison, including the one-stage methods: RetinaNet [19], " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathbf{A}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " Net [11], and YOLOv5s [17]; and the two-stage methods: Faster R-CNN [25], Oriented R-CNN [36], and RoITransformer [10]. RGB and IR images are separately used as training data to evaluate the detection performance of each algorithm. Multimodal fusion methods: We also compare our method with eight recent fusion methods:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26841" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 106, + 544, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Detection results (mAP, in %) on DroneVehicle dataset. Note that all detectors locate and classify vehicles with OBB heads. The speed refers to the speed of network inference without post-processing (batch size = 1). In speed, our method is only compared with the fusion methods. Best results are highlighted in bold. And the second one is marked with underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 106, + 544, + 357 + ], + "lines": [ + { + "bbox": [ + 49, + 106, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 544, + 357 + ], + "type": "table", + "html": "
DetectorsModalityCarTruckFreight-carBusVanmAP (%) ↑Speed (fps) ↑
RetinaNet [19]78.534.424.169.828.847.1-
Faster R-CNN [25]79.049.037.277.037.055.9-
Oriented R-CNN [36]RGB80.153.841.685.443.360.8-
S²A-Net [11]80.054.242.284.943.861.0-
RoITransformer [10]61.655.142.385.544.861.6-
YOLOv5s [17]78.655.343.887.146.062.1-
RetinaNet [19]88.835.439.576.532.154.5-
Faster R-CNN [25]89.453.548.387.042.664.2-
Oriented R-CNN [36]IR89.857.453.189.345.467.0-
S²A-Net [11]89.954.555.888.948.467.5-
RoITransformer [10]90.160.458.989.752.270.3-
YOLOv5s [17]90.059.560.889.553.870.7-
UA-CMDet [29]87.560.746.887.138.064.09.1
Halfway Fusion [21]90.162.358.589.149.870.020.4
CIAN [44]90.163.860.789.150.370.821.7
AR-CNN [45]90.164.862.189.451.571.618.2
MBNet [50]RGB+IR90.164.462.488.853.671.921.7
TSFADet [41]89.967.963.789.854.073.118.6
C²Former [40]90.268.364.489.858.574.2-
SLBAF-Net [7]90.272.068.689.959.976.163.2
Ours90.376.873.390.366.079.433.1
", + "image_path": "f8ae8b7be05ca3ff6a760f3291a40a37fb8f298f615befec69c4044d7b72c2f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": "UA-CMDet [29], Halfway Fusion [21], CIAN [44], MBNet [50], AR-CNN [45], TSFADet [41], " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "C^2" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": "Former [40], and SLBAF-Net [7]. Among them, UA-CMDet, MBNet, TSFADet, and " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "C^2" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": "Former all consider the weakly misalignment problem during the fusion process, which affirms the universality and significance of this problem. For each compared method, we use their original experimental settings to ensure equity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 651 + ], + "type": "text", + "content": "Quantitative comparison. The results are shown in Tab. 1. In the unimodal methods, benefitting from a well-designed multiscale feature extraction network, YOLOv5s has comparable detection accuracy (62.1% in RGB images and 70.7% in IR images), even better than some fusion methods. Correspondingly, the multimodal methods SLBAF-Net, constructed upon the two-stream YOLOv5 framework, has also yielded excellent results. Compared with all methods, we can find that our OAFA achieves the best average performance on the validation set, which is 3.3% higher than the second place. It is worth noting that our method has obvious advantages in distinguishing confusing categories (truck and freight-car). The reason may stem from the fact that OAFA excels in identifying the optimal spatial correspondences of multimodal features, obtaining reliable features of the target after fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Qualitative comparison. Our baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branches are fused with a concatenation operation in multiscale features. Among the compared methods mentioned above, we select the optimal unimodal" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 361, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 517 + ], + "type": "text", + "content": "YOLOv5s and the baseline model as the qualitative comparative model. We provide visual detection results of the compared methods in Fig. 3. It can be observed that, given the constraint of the IR labels, the objects are successfully detected by OAFA when the target positions are shifted, while both unimodal and baseline model experience classification and localization errors or fail to detect, especially for occluded and few-shot object detection. The underlying causes can be speculated that the weak misalignment problem confuses the fusing model to learn discriminative features. In contrast, our method has the capability to capture aligned and reliable features, improving the performance in classification and localization processes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "type": "text", + "content": "Speed comparison. We compare the speed of OAFA with other fusion detection methods to prove the efficiency of our method. For a fair comparison, all the detection models are tested with batch size 1, and the input size is set to " + }, + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "type": "text", + "content": ". As shown in Tab. 1, our proposed OAFA achieves a speed of 33.1 FPS, outperforming most existing multimodal object detection methods. Although our method requires more computational time compared to SLBAF-Net, it would be worthwhile to invest this additional time to achieve superior performance. Furthermore, it can be seen that OAFA can accomplish real-time detection on the NVIDIA RTX A6000 platform." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 670, + 400, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 400, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 400, + 684 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "In this section, we conduct several ablation studies to validate the effectiveness of the different components in our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26842" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 251, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 251, + 83 + ], + "type": "text", + "content": "proposed method. The results are shown in Tab. 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "text", + "content": "Effect of CSOM module. We first analyze the effect of the CSOM module as it provides the fundamental offset estimation, which determines the stability of the model training. We perform ablation experiments on both the decoupled multimodal learning (DML) module and the spatial offset modeling (SOM) module separately. As presented in the first to third rows of Tab. 2, the performance of the baseline model is improved by " + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "text", + "content": " after incorporating the SOM and DML modules, respectively. It verifies that both coarse-grained feature alignment and increasing fusion flexibility through decouple learning are effective for the detection task. Besides, when integrating these two submodules, the performance improvement of model over baseline is " + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 85, + 286, + 264 + ], + "type": "text", + "content": ", which indicates that the quality of the feature alignment can be enhanced by mitigating modal gap." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": "Effect of ODAF module. We then verify the necessity of the ODAF module in addressing the weak misalignment problem. The results demonstrate that despite ODAF has poor stability in model training without dependable offset guidance, incorporating ODAF into the baseline can provide " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": " significant improvement. It is owing to adaptively adjust the optimal sampling position for acquiring reliable information in RGB images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "text", + "content": "Effect of two-stage training. The two-stage training strategy provides comprehensive unimodal representations for multimodal alignment and fusion. We can observe that the accuracy decreases to " + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "inline_equation", + "content": "77.3\\%" + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "text", + "content": " if we directly train the whole network without two-stage training. It is " + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "text", + "content": " lower than exclusively incorporating CSOM and ODAF, respectively. The reason could be that the inappropriate initialization in feature extraction network leads to training instability and makes it hard to obtain abundant representations in spatial and semantic. After undergoing a two-stage training process, our method attains a performance level of " + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 46, + 361, + 286, + 552 + ], + "type": "text", + "content": ", surpassing the capabilities of any individual module. It demonstrates that the two-stage training strategy effectively enhances training accuracy and stability. The superior method to fully exploit the strengths of CSOM and ODAF will be further studied in future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 553, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 286, + 590 + ], + "type": "text", + "content": "All in all, not only each independent module but also the entire proposed framework can contribute to the multimodal object detection under weakly misalignment conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 598, + 200, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 200, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 200, + 610 + ], + "type": "text", + "content": "4.4. Robustness to Position Shift" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "To quantitatively assess the robustness of our method to the weakly misalignment issue, we conduct experiments on DroneVehicle by manually simulating positional offsets. The tested model is trained on the original training image pairs and validated on the shifted image pairs. In our experiments, we hold the IR images constant while introducing spatial offsets along the " + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "-axis and " + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "-axis for RGB images. The variations in pixel values are defined within the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 129, + 544, + 217 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 125 + ], + "type": "text", + "content": "Table 2. Ablation study on DroneVehicle dataset. The baseline model is SLBAF-Net, a two-stream framework based on YOLOv5s where RGB and IR branch are fused with a multiscale concatenation operation. The D denotes DML, S denotes SOM, C denotes CSOM, and O denotes ODAF." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 129, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 309, + 129, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 309, + 129, + 544, + 217 + ], + "type": "table", + "html": "
MethodCSOMODAFTwo-stage TrainingmAP (%) ↑
DMLSOM
Baseline76.1
Baseline+S✓76.5
Baseline+D✓77.7
Baseline+C✓✓78.2
Baseline+O✓77.8
Baseline+C+O✓✓✓77.3
OAFA✓✓✓✓79.4
", + "image_path": "643e614f8ae8852afac8c1958f64b0ffda1286d729845fe2d7669887296c9180.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 224, + 420, + 304 + ], + "blocks": [ + { + "bbox": [ + 312, + 224, + 420, + 304 + ], + "lines": [ + { + "bbox": [ + 312, + 224, + 420, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 224, + 420, + 304 + ], + "type": "image", + "image_path": "67de58df0de66f1a6722fb231b8b6fbc1826868a1f3cfad115393880f9a7b0f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 309, + 411, + 319 + ], + "lines": [ + { + "bbox": [ + 325, + 309, + 411, + 319 + ], + "spans": [ + { + "bbox": [ + 325, + 309, + 411, + 319 + ], + "type": "text", + "content": "(a) SLBAF-Net (Baseline)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 224, + 541, + 304 + ], + "blocks": [ + { + "bbox": [ + 432, + 224, + 541, + 304 + ], + "lines": [ + { + "bbox": [ + 432, + 224, + 541, + 304 + ], + "spans": [ + { + "bbox": [ + 432, + 224, + 541, + 304 + ], + "type": "image", + "image_path": "fbc57013de2708b7741397311d829d0152a1d067604b295a0748ffd272f0167d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 310, + 515, + 319 + ], + "lines": [ + { + "bbox": [ + 463, + 310, + 515, + 319 + ], + "spans": [ + { + "bbox": [ + 463, + 310, + 515, + 319 + ], + "type": "text", + "content": "(b) Our method" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "type": "text", + "content": "Figure 5. Surface plot visualization of the position shift experiments. Horizontal coordinates indicate various shift steps along the " + }, + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "type": "text", + "content": "-axis and " + }, + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 323, + 545, + 367 + ], + "type": "text", + "content": "-axis. Vertical coordinate denotes the detection performances in terms of mAP." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 369, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 465 + ], + "type": "text", + "content": "range " + }, + { + "bbox": [ + 304, + 369, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\{(\\Delta x, \\Delta y) \\mid \\Delta x, \\Delta y \\in [15, -15]; \\Delta x, \\Delta y \\in \\mathbb{Z}\\}" + }, + { + "bbox": [ + 304, + 369, + 545, + 465 + ], + "type": "text", + "content": ". The quantitative outcomes of this experiment in baseline and our model are discernible from Fig. 5. It can be seen that in response to larger position shifts, our method consistently exhibits minor fluctuations in performance, whereas the baseline model experiences a noticeable performance decline. This result demonstrates the robustness of OAFA under weakly misalignment conditions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 688 + ], + "type": "text", + "content": "In this paper, we propose a robust RGB-IR object detection method OAFA for UAVs images under weakly misalignment conditions. Different from previous methods, OAFA can achieve adaptive multimodal feature alignment without strict alignment. To mitigate the impact of the modality gap on multimodal spatial matching, the modality-invariant features are acquired in a common subspace to estimate accurate offsets. Taking the offsets as basic knowledge, the adaptive alignment implicitly captures optimal fusion positions and steadily carries out feature-level corrections with a deformable alignment module. Experimental results on DroneVehicle dataset demonstrate the superior performances of OAFA compared with the state-of-the-art methods. Moreover, our method shows robustness against the weakly misalignment problem in position shift experiments." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26843" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Gedas Bertasius, Lorenzo Torresani, and Jianbo Shi. Object detection in video with spatiotemporal sampling networks. In ECCV, pages 342-357, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Ilker Bozcan and Erdal Kayacan. AU-AIR: A multi-modal unmanned aerial vehicle dataset for low altitude traffic surveillance. In ICRA, pages 8504-8510, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[3] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Dong Nie, Minjeong Kim, Qian Wang, and Dinggang Shen. Deformable image registration based on similarity-steered CNN regression. In MICCAI, pages 300-308, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "type": "text", + "content": "[4] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In AAAI, pages 973-981, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[5] Kelvin C. K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5962-5971, 2022. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "type": "text", + "content": "[6] Hao Chen, Yongjian Deng, Youfu Li, Tzu-Yi Hung, and Guosheng Lin. RGBD salient object detection via disentangled cross-modal fusion. IEEE TIP, 29:8407-8416, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 368 + ], + "type": "text", + "content": "[7] Xiaolong Cheng, Keke Geng, Ziwei Wang, Jinhu Wang, Yuxiao Sun, and Pengbo Ding. Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment. Multimedia Tools and Applications, pages 1-20, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 402 + ], + "type": "text", + "content": "[8] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[9] Chaitanya Devaguptapu, Ninad Akolekar, Manuj M. Sharma, and Vineeth N. Balasubramanian. Borrow from anywhere: Pseudo multi-modal object detection in thermal imagery. In CVPR, pages 1029-1038, 2019. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "text", + "content": "[10] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In CVPR, pages 2849-2858, 2019. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "text", + "content": "[11] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE TGRS, 60: 1-11, 2022. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[12] Renlong Hang, Zhu Li, Pedram Ghamisi, Danfeng Hong, Guiyu Xia, and Qingshan Liu. Classification of hyperspectral and lidar data using coupled cnns. IEEE TGRS, 58(7): 4939-4950, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[13] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. MISA: modality-invariant and -specific representations for multimodal sentiment analysis. In ACM MM, pages 1122-1131, 2020. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "text", + "content": "[14] Danfeng Hong, Naoto Yokoya, Jocelyn Chanussot, and Xiao Xiang Zhu. Cospace: Common subspace learning from hyperspectral-multispectral correspondences. IEEE TGRS, 57(7):4349-4359, 2019. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[15] Danfeng Hong, Jocelyn Chanussot, Naoto Yokoya, Jian Kang, and Xiao Xiang Zhu. Learning-shared cross-modality representation using multispectral-lidar and hyperspectral data. IEEE GRSL, 17(8):1470-1474, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Lloyd Haydn Hughes, Diego Marcos, Sylvain Lobry, Devis Tuia, and Michael Schmitt. A deep learning framework for" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "matching of sar and optical imagery. ISPRS, 169:166-179, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 95, + 544, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 544, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 544, + 117 + ], + "type": "text", + "content": "[17] Glenn Jocher. ultralytics/yolov5. https://github.com/ultralytics/yolov5, Oct. 2020. 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[18] Huafeng Li, Junzhi Zhao, Jinxing Li, Zhengtao Yu, and Guangming Lu. Feature dynamic alignment and refinement for infrared-visible image fusion: Translation robust fusion. Information Fusion, 95:26-41, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "text", + "content": "[20] Fan Liu, Zhiyong Cheng, Huilin Chen, Anan Liu, Liqiang Nie, and Mohan S. Kankanhalli. Disentangled multimodal representation learning for recommendation. arXiv:2203.05406, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[21] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Multispectral deep neural networks for pedestrian detection. In BMVC, 2016. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "type": "text", + "content": "[22] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[23] Sijie Mai, Haifeng Hu, and Songlong Xing. Modality to modality translation: An adversarial representation learning and graph fusion network for multimodal fusion. In AAAI, pages 164-172, 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 338, + 545, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 359 + ], + "type": "text", + "content": "[24] Sayan Nag. Image registration techniques: A survey. arXiv:1712.07540, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "text", + "content": "[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2017. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "type": "text", + "content": "[26] Marc-Michel Rohé, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: Learning deformable image registration using shape matching. In MIC-CAI, pages 266–274, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "type": "text", + "content": "[27] Hazim Shakhatreh, Ahmad Sawalmeh, Ala I. Al-Fuqaha, Zuochao Dou, Eyad Almaita, Issa Khalil, Noor Shamsiah Othman, Abdallah Khreishah, and Mohsen Guizani. Unmanned aerial vehicles (uavs): A survey on civil applications and key research challenges. IEEE Access, 7:48572-48634, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[28] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 536-553, 2018. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "text", + "content": "[29] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning. IEEE TCSVT, 32(10):6700-6713, 2022. 1, 5, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "text", + "content": "[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally-deformable alignment network for video super-resolution. In CVPR, pages 3357-3366, 2020. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[31] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In IJCAI, pages 3508-3515, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[32] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In ICCV, pages 3622–3631, 2019. 2, 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26844" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[33] Hua Wang, Dewei Su, Chuangchuang Liu, Longcun Jin, Xianfang Sun, and Xinyi Peng. Deformable non-local network for video super-resolution. IEEE Access, 7:177734-177744, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[34] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. In CVPR, pages 1954-1963, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "text", + "content": "[35] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In ECCV, pages 3–19, 2018. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "text", + "content": "[36] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented R-CNN for object detection. In ICCV, pages 3500-3509, 2021. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 262 + ], + "type": "text", + "content": "[37] Han Xu and Jiayi Ma. Emfusion: An unsupervised enhanced medical image fusion network. Information Fusion, 76:177-186, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 286, + 308 + ], + "type": "text", + "content": "[38] Xing Xu, Kaiyi Lin, Lianli Gao, Huimin Lu, Heng Tao Shen, and Xuelong Li. Learning cross-modal common representations by private-shared subspaces separation. IEEE Trans. Cybern., 52(5):3261-3275, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[39] Huang Yao, Rongjun Qin, and Xiaoyu Chen. Unmanned aerial vehicle for remote sensing applications - A review. Remote Sensing, 11(12):1443, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "type": "text", + "content": "[40] Maoxun Yuan and Xingxing Wei. " + }, + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^2" + }, + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "type": "text", + "content": "former: Calibrated and complementary transformer forrgb-infrared object detection. arXiv:2306.16175, 2023. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "text", + "content": "[41] Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In ECCV, pages 509-525, 2022. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[42] Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschlager, and Susanne Saminger-Platz. Central moment discrepancy (CMD) for domain-invariant representation learning. In ICLR, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "text", + "content": "[43] Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Grid anchor based image cropping: A new benchmark and an efficient model. IEEE TPAMI, 44(3):1304-1319, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[44] Lu Zhang, Zhiyong Liu, Shifeng Zhang, Xu Yang, Hong Qiao, Kaizhu Huang, and Amir Hussain. Cross-modality interactive attention network for multispectral pedestrian detection. Information Fusion, 50:20-29, 2019. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[45] Lu Zhang, Zhiyong Liu, Xiangyu Zhu, Zhan Song, Xu Yang, Zhen Lei, and Hong Qiao. Weakly aligned feature fusion for multimodal object detection. arXiv:2204.09848, 2022. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "text", + "content": "[46] Xinyue Zhang, Chengcai Leng, Yameng Hong, Zhao Pei, Irene Cheng, and Anup Basu. Multimodal remote sensing image registration methods and advancements: A survey. Remote Sensing, 13(24):5128, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[47] Zhimeng Zhang and Yu Ding. Adaptive affine transformation: A simple and effective operation for spatial misaligned image generation. In ACM MM, pages 1167-1176, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[48] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, 2023. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 218 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[49] Jiaxiang Zheng, Yichen Tian, Chao Yuan, Kai Yin, Feifei Zhang, Fangmiao Chen, and Qiang Chen. Mdesnet: Multitask difference-enhanced siamese network for building change detection in high-resolution remote sensing images. Remote Sensing, 14(15):3775, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[50] Kailai Zhou, Linsen Chen, and Xun Cao. Improving multispectral pedestrian detection by addressing modality imbalance problems. In ECCV, pages 787-803, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 218 + ], + "type": "text", + "content": "[52] Barbara Zitová and Jan Flusser. Image registration methods: a survey. Image Vis. Comput., 21(11):977-1000, 2003. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26845" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_content_list.json b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..15163458088ea1880baa46bd51dc8846401fe58f --- /dev/null +++ b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_content_list.json @@ -0,0 +1,1598 @@ +[ + { + "type": "text", + "text": "Weakly Supervised Monocular 3D Detection with a Single-View Image", + "text_level": 1, + "bbox": [ + 130, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xueying Jiang $^{1}$ Sheng Jin $^{1}$ Lewei Lu $^{2}$ Xiaoqin Zhang $^{3}$ Shijian Lu $^{1*}$ $^{1}$ S-Lab, Nanyang Technological University \n $^{2}$ Sensetime Research \n $^{3}$ College of Computer Science and Technology, Zhejiang University of Technology", + "bbox": [ + 155, + 179, + 812, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D detection (M3D) aims for precise 3D object localization from a single-view image which usually involves labor-intensive annotation of 3D detection boxes. Weakly supervised M3D has recently been studied to obviate the 3D annotation process by leveraging many existing 2D annotations, but it often requires extra training data such as LiDAR point clouds or multi-view images which greatly degrades its applicability and usability in various applications. We propose SKD-WM3D, a weakly supervised monocular 3D detection framework that exploits depth information to achieve M3D with a single-view image exclusively without any 3D annotations or other training data. One key design in SKD-WM3D is a self-knowledge distillation framework, which transforms image features into 3D-like representations by fusing depth information and effectively mitigates the inherent depth ambiguity in monocular scenarios with little computational overhead in inference. In addition, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Extensive experiments show that SKD-WM3D surpasses the state-of-the-art clearly and is even on par with many fully supervised methods.", + "bbox": [ + 75, + 318, + 472, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D detection (M3D) has emerged as one key component in the area of autonomous driving and computer vision. Its primary target is to recognize objects and obtain their 3D localization from single-view images. Thanks to its low deployment cost, M3D [5, 34] has attracted increasing attention in both academic and industrial sectors, achieving very impressive progress in recent years. On the other hand, most existing studies [21, 33, 37, 39] adopt a fully supervised setup which have been facing increasing scalability concern as large-scale 3D boxes are often labor-intensive", + "bbox": [ + 75, + 723, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to collect. Effective M3D training without 3D annotations has become a critical issue while handling M3D problems in various research and practical tasks.", + "bbox": [ + 500, + 287, + 890, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly supervised M3D (WM3D) [34] has recently been explored for learning effective 3D detectors without 3D box annotations, aiming to exploit 2D annotations to make up for the absence of 3D information. For example, WeakM3D [34] exploits LiDAR point clouds to infer 3D information as illustrated in Figure 1(a). However, it requires costly and complicated LiDAR sensors to collect point clouds which limits its applicability and usability greatly. WeakMono3D [42] employs 2D information only by either leveraging multi-view stereo with images from multiple cameras or constructing pseudo-multi-view perspective from sequential video frames as illustrated in Figure 1(b). However, collecting multi-view images is complicated, and resorting to a pseudo multi-view perspective degrades the detection performance clearly. With the advance of single-view depth estimation, WM3D with depth from a single-view image presents a potential solution for compensating the absence of 3D annotations. On the other hand, direct integration of such depth into existing frameworks often necessitates complex network architectures which further incurs significant computational costs. This gives rise to a pertinent question: When not using additional LiDAR point clouds or multi-view image pairs, is it possible to harness the depth from off-the-shelf depth estimators without introducing much computational overhead in inference?", + "bbox": [ + 496, + 337, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We design SKD-WM3D, a novel weakly supervised monocular 3D object detection method that is exclusively grounded on single-view images. One key design in SKD-WM3D is a self-knowledge distillation framework which consists of a Depth-guided Self-teaching Network (DSN) and a Monocular 3D Detection Network (MDN). As illustrated in Figure 1(c), SKD-WM3D utilizes depth information obtained from an off-the-shelf depth estimator [17] to enhance the 3D localization ability of DSN and transfers such ability to MDN via self-knowledge distillation. Such self-distillation design enables MDN to unearth the intrinsic depth information from single-view images independently,", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10508", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23679a6c6316d3da32c47fa05c8d92c85f8467f36c213b18077975dcf47d17b4.jpg", + "image_caption": [ + "(a) LiDAR-based WM3D" + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 241, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ed648161116fb8ded3accfde6bf0be8ca6a3431bf6fd1b8dab2175cbd86a119.jpg", + "image_caption": [ + "(b) Multi-view WM3D" + ], + "image_footnote": [], + "bbox": [ + 267, + 87, + 553, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3fbf79a99a0aa1dd25ae988d0b22e581abfce7194ff3c1165be3fa1cfaac853.jpg", + "image_caption": [ + "(c) Single-view WM3D (Ours)", + "Figure 1. Different paradigms in weakly supervised monocular 3D detection. Our approach in (c) leverages Pseudo Depth Labels from a single-view image to achieve weakly supervised monocular 3D detection, requiring no extra training data like LiDAR point clouds or multi-view images as in (a) and (b). It improves usability and applicability greatly. The Pseudo Depth Labels are obtained with an off-the-shelf depth estimator [17] without extra training and ground-truth depth labels. Data in red denotes extra data in network training." + ], + "image_footnote": [], + "bbox": [ + 594, + 88, + 890, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bypassing additional modules such as pre-trained depth estimation networks and leading to precise and efficient 3D localization with little computational overhead during inference. On top of DSN and MDN, we design an uncertainty-aware distillation loss to optimize the utilization of the transferred 3D localization knowledge by weighting up more certain knowledge while weighting down less certain knowledge. In addition, we design a gradient-targeted transfer modulation strategy to synchronize the learning paces of DSN and MDN during the process of learning 3D localization knowledge, by prioritizing MDN learning at the initial stage when MDN lags behind DSN and enabling it to provide more feedback to DSN when MDN is better trained at late stages.", + "bbox": [ + 75, + 349, + 468, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution can be summarized in three aspects. First, we design a novel framework that achieves weakly supervised monocular 3D detection by distilling knowledge between a depth-guided self-teaching network and a monocular 3D detection network. Without any extra training data like LiDAR point clouds or multi-view images, the framework exploits depth exclusively from a single image with little computational overhead in inference. Second, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Third, the proposed approach clearly outperforms the state-of-the-art in weakly supervised monocular 3D detection, and its performance is even on par with several fully supervised methods.", + "bbox": [ + 75, + 563, + 468, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 805, + 218, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Monocular 3D Detection", + "text_level": 1, + "bbox": [ + 76, + 830, + 300, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 3D object detection aims to predict 3D object localization from single-view images. Standard monocular detectors [1, 6, 14, 53, 56] operate solely on single images,", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "without utilizing additional data. However, the inherent depth ambiguity of monocular detection significantly hinders its performance compared to its stereo counterparts. To address this limitation, various approaches seek solutions with the help of extra data, such as LiDAR point clouds [4, 7, 21, 27], video sequences [2], 3D CAD models [5, 24, 31], and depth estimation [10, 35, 45, 48]. Specifically, MonoRUn [4] adopts an uncertainty-aware regional reconstruction network for regressing pixel-associated 3D object coordinates with LiDAR point clouds as extra supervision. MonoDistill [7] introduces an effective distillation-based approach that incorporates spatial information from LiDAR signals into monocular 3D detection. Additionally, pseudo-LiDAR-based methods [45, 48] convert estimated depth maps to simulate the real LiDAR point clouds to utilize the well-designed LiDAR-based 3D detector. During inference, compared with methods using depth estimation, our method eliminates the need for pseudo depth labels and complex network architectures, with little computational overhead. Besides, existing fully supervised methods require large-scale 3D box ground truth, which is labor-intensive to collect and annotate.", + "bbox": [ + 496, + 349, + 890, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Weakly Supervised 3D Object Detection", + "text_level": 1, + "bbox": [ + 500, + 695, + 841, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the high cost of annotating 3D boxes in the 3D object detection task, various weakly supervised approaches [26, 29, 34, 36] have been proposed. For example, WS3D [29] presents a weakly supervised method for 3D LiDAR object detection, which requires only a limited number of weakly annotated scenes with center-annotated BEV maps. VS3D [36] introduces a cross-model knowledge distillation strategy to transfer the knowledge from the RGB domain to the point cloud domain, using LiDAR point clouds as weak supervision. Recent research on weakly supervised 3D object detection has turned to explore the monocular setting. For example, WeakM3D [34] generates 2D boxes to select", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10509", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/141d4e32d52dadcc203433a5afe29fc63c2ec21bd73a2590abedcb929b8fbcfe.jpg", + "image_caption": [ + "Figure 2. The framework of the proposed self-knowledge distillation network. The framework consists of a depth-guided self-teaching network and a monocular 3D detection network. The depth-guided self-teaching network acquires comprehensive 3D localization knowledge by leveraging depth information and transfers its learned expertise to the monocular 3D detection network via soft label distillation to enhance its performance. We design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy to facilitate the knowledge transfer between the two networks effectively. During inference, the monocular 3D detection network extracts intrinsic depth information from single-view images independently with little computational overhead." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 895, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RoI LiDAR point clouds as weak supervision and then predicts 3D boxes that closely align with the selected RoI LiDAR point clouds. More recently, WeakMono3D [42] eliminates the need for LiDAR, offering both multi-view and single-view yet multi-frame versions. While the former acquires stereo image inputs from multiple cameras, the latter constructs a pseudo-multi-view perspective using multiple video frames. The multi-frame version exhibits inferior 3D scene comprehension compared with the multi-view approach due to its smaller inter-frame disparity, leading to degraded performance. Instead of requiring extra training data like LiDAR point clouds or multi-view images, we tackle the challenge of weakly supervised monocular 3D detection by leveraging a single-view image exclusively.", + "bbox": [ + 75, + 489, + 472, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Self-Knowledge Distillation", + "text_level": 1, + "bbox": [ + 76, + 710, + 323, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge distillation [8, 9, 11, 16, 18, 23, 32, 38, 43, 54] aims to transfer knowledge from a pre-trained teacher network to a student network for improving its performance. Self-knowledge distillation [30, 41, 47], distinct from traditional knowledge distillation, leverages the information within the student network to facilitate its learning without the pre-trained teacher network. Specifically, data augmentation approach [15, 46, 50] transfers knowledge through different distortions of the same training data. However, they are susceptible to inappropriate augmentations, such as improper instance rotation or distortion, potentially intro", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ducing noise that hampers network learning. Another typical approach exploits auxiliary networks [52, 57]. For example, DKS [40] introduces auxiliary supervision branches and pairwise knowledge alignments, while FRSKD [19] adds a new branch supervised by the original features and utilizes both soft-label and feature-map distillation. Our work is the first that introduces self-knowledge distillation with auxiliary networks for weakly supervised monocular 3D detection. It effectively exploits depth information from single-view images with little computational overhead during inference.", + "bbox": [ + 496, + 489, + 893, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 671, + 635, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the proposed SKD-WM3D. First, the problem definition and overview are presented in Sec. 3.1. Then detailed designs of SKD-WM3D are introduced, including the self-knowledge distillation framework in Sec. 3.2, the uncertainty-aware distillation loss in Sec. 3.3 and the gradient-targeted transfer modulation strategy in Sec. 3.4. Finally, loss functions are presented in Sec. 3.5.", + "bbox": [ + 496, + 698, + 893, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition and Overview", + "text_level": 1, + "bbox": [ + 500, + 830, + 794, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly supervised monocular 3D detection takes an RGB image and the corresponding 2D bounding boxes as supervision, aiming to classify objects and determine their", + "bbox": [ + 496, + 854, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10510", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bounding boxes in 3D space without involving any 3D annotations in training. The prediction of each object is composed of the object category $C$ , a 2D bounding box $B^{2D}$ and a 3D bounding box $B^{3D}$ . Specifically, the 3D box $B^{3D}$ can be further decomposed to the object 3D location $(x_{3D}, y_{3D}, z_{3D})$ , the object dimension with height, width and length $(h_{3D}, w_{3D}, l_{3D})$ , as well as orientation $\\theta$ .", + "bbox": [ + 76, + 90, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design a self-knowledge distillation framework to tackle the challenge of weakly supervised monocular 3D detection from a single-view image. As Figure 2 shows, the framework consists of two subnetworks including a Depth-Guided Self-Teaching Network and a Monocular 3D Detection Network. In the Depth-Guided Self-Teaching Network, the global features $F_{G}$ extracted by the backbone are fed into a Depth Head to obtain depth features. Next, the global features $F_{G}$ and the extracted depth features are fed into a Fusion Layer to obtain 3D-like features $F_{3D}$ . The 3D-like features of each object are then obtained via RoIAlign, and further fed to a Depth-Aware 3D Head to predict 3D box $\\widehat{B}_p^{3D}$ and uncertainty $\\widehat{U}$ . In the Monocular 3D Detection Network, We first use RoIAlign to generate object-level features from the global features $F_{G}$ , and then feed them to a 2D-to-3D Head to predict 3D box $\\widehat{B}_p^{2D}$ and uncertainty $U$ . The 3D boxes predicted by both networks are further projected into 2D boxes. Moreover, we design an uncertainty-aware distillation loss $L_{ud}$ to obtain low-uncertainty knowledge, and a gradient-targeted transfer modulation strategy to synchronize the learning paces between the two networks by controlling gradients $\\widehat{G}$ and $G$ of $L_{ud}$ .", + "bbox": [ + 76, + 198, + 467, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Self-Knowledge Distillation Framework", + "text_level": 1, + "bbox": [ + 76, + 537, + 416, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The self-knowledge distillation framework enhances the 3D localization ability of the depth-guided self-teaching network by utilizing depth information from an off-the-shelf depth estimator and then transfers the ability to the monocular 3D detection network via self-knowledge distillation.", + "bbox": [ + 76, + 560, + 467, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth-Guided Self-Teaching Network. To equip the self-teaching network with 3D localization ability, we propose to learn from global features $F_{G}$ and depth information from an off-the-shelf depth estimator to acquire comprehensive 3D knowledge. The depth information is exploited via two major designs. Firstly, we introduce a depth head $\\mathcal{D}$ that extracts depth features $F_{D}$ as follows:", + "bbox": [ + 76, + 652, + 467, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {D} = \\mathcal {D} \\left(F _ {G}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 773, + 467, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The depth features $F_{D}$ are exploited to generate depth maps $D_{p}$ , where the depth map generation is supervised by the pseudo ground truth of the depth map $D_{gt}$ that is predicted by an off-the-shelf depth estimator by using the focal loss [22] as depth loss $L_{dep}$ . Hence, the depth features can be acquired by the depth-guided self-teaching network effectively.", + "bbox": [ + 76, + 795, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 1. We generate depth pseudo labels using an off-the-shelf depth estimator [17] with frozen weights, eliminating the need for additional training and ground-truth depth labels. Adopting an off-the-shelf depth estimator incurs negligible costs as compared with prior studies that require either point clouds [34] or multi-view images [42].", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, we obtain 3D-like features $F_{G3D}$ by integrating the depth features $F_{D}$ that provide information along the depth dimension, as well as the global features $F_{G}$ that capture knowledge about the 2D image plane. Specifically, we design a fusion layer that fuses the depth features $F_{D}$ with the global features $F_{G}$ to derive the $F_{G3D}$ as follows:", + "bbox": [ + 498, + 183, + 890, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {G 3 D} = F F N \\left(C A \\left(S A \\left(F _ {D}\\right), F _ {G}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 291, + 890, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $FFN$ is the feed-forward network, and $CA$ , $SA$ denote CrossAttention, SelfAttention, respectively. The structures of CrossAttention and SelfAttention employ the standard transformer architecture [44]. The obtained 3D comprehension improves the network's ability to precisely locate objects, effectively mitigating depth ambiguity arising from single-view image input.", + "bbox": [ + 498, + 316, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular 3D Detection Network. The monocular 3D detection network acquires the 3D localization knowledge from the depth-guided self-teaching network. By distilling soft labels generated by the depth-guided self-teaching network, the monocular 3D detection network can extract intrinsic depth information from images independently during inference. This kills the need for additional complex modules such as pre-trained depth estimation networks or depth fusion modules, facilitating the inference with little computational overhead.", + "bbox": [ + 498, + 446, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Uncertainty-Aware Distillation Loss", + "text_level": 1, + "bbox": [ + 500, + 609, + 808, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the knowledge distillation process, uncertain knowledge could affect the network training negatively if all transferred knowledge is treated equally. To benefit more from certain knowledge and weaken the effect of uncertain knowledge, we design an uncertainty-aware distillation loss between the 3D boxes that are predicted by the two networks in the self-knowledge distillation framework. The uncertainty-aware distillation loss exploits the prediction uncertainty to modulate the distillation loss magnitude as follows:", + "bbox": [ + 498, + 633, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {u d} = \\frac {L _ {d}}{\\operatorname* {m i n} ((\\widehat {U} + U) / 2 , \\alpha)} + \\left\\| \\operatorname * {m i n} (\\frac {\\widehat {U} + U}{2}, \\alpha) \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 810, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widehat{U}$ and $U$ denote the uncertainty of the 3D boxes that are predicted by the two networks, respectively. Here", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we assume the 3D box predictions have the Laplace distribution, and we adopt the standard deviations as the uncertainties, inspired by [20, 25, 33]. $\\left\\| \\min (\\frac{\\widehat{U} + U}{2},\\alpha)\\right\\| ^2$ is the L2 regularization, and $\\alpha$ is a fixed value set to 0.1. $L_{d}$ denotes the basic distillations loss, and we employ the commonly used SmoothL1 [13] loss to enforce the consistency between the 3D boxes predicted by the two networks. The SmoothL1 loss leaves a soft margin when computing the difference between the two 3D boxes:", + "bbox": [ + 75, + 90, + 472, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d} = \\left\\{ \\begin{array}{l l} 0. 5 \\times \\left(\\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D}\\right) ^ {2}, & \\text {i f} \\left| \\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \\right| < 1. 0 \\\\ \\left| \\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \\right| - 0. 5, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 255, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widehat{B}_p^{3D}$ and $B_p^{3D}$ are the predicted 3D boxes from the depth-guided self-teaching network and the monocular 3D detection network, respectively.", + "bbox": [ + 76, + 303, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 2. The proposed uncertainty-aware distillation loss $L_{ud}$ integrates average uncertainty $\\frac{\\widehat{U} + U}{2}$ as regularization and a weighted component for the basic distillation loss $L_{d}$ , allowing adaptive learning adjustments based on the knowledge's uncertainty level. Specifically, when dealing with uncertain knowledge, a smaller weight is assigned to the basic distillation loss $L_{d}$ to mitigate potential adverse effects on network learning. Consequently, the network prioritizes optimizing uncertainty reduction in such scenarios. When dealing with certain knowledge, the network emphasizes optimizing the basic distillation loss $L_{d}$ due to its higher weight. Notably, the basic distillation loss $L_{d}$ simply considers box consistency, while integrating uncertainty is beneficial for enhancing the knowledge distillation process.", + "bbox": [ + 75, + 367, + 472, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Transfer Modulation Strategy", + "text_level": 1, + "bbox": [ + 76, + 590, + 344, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The depth-guided self-teaching network, which leverages depth information to predict 3D boxes, transfers its learned 3D knowledge to the monocular 3D detection network. The asynchronous learning paces of the two networks pose potential challenges to effective 3D knowledge transfer.", + "bbox": [ + 75, + 613, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design a gradient-targeted transfer modulation strategy to synchronize the learning pace of the depth-guided self-teaching network and the monocular 3D detection network. We modulate the knowledge transfer dynamically, by controlling the gradients from the uncertainty-aware distillation loss $L_{ud}$ . Specifically, we adapt the gradients based on the 2D projection performance of each network, assigning smaller backward gradients for the good-performing network and higher backward gradients for the bad-performing network. The gradient-targeted transfer modulation strategy is formulated as follows:", + "bbox": [ + 75, + 689, + 470, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {G} ^ {\\prime} = \\frac {2 \\times \\widehat {L} _ {p r o j}}{\\widehat {L} _ {p r o j} + L _ {p r o j}} \\times \\widehat {G}, G ^ {\\prime} = \\frac {2 \\times L _ {p r o j}}{\\widehat {L} _ {p r o j} + L _ {p r o j}} \\times G, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 864, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\widehat{G}$ and $G$ are the original gradients of the two networks, $\\widehat{G}^{\\prime}$ and $G^{\\prime}$ are the modified gradients, $\\widehat{L}_{proj}$ and $L_{proj}$ are projection losses, computed between the projected 2D boxes from 3D box predictions and 2D box annotations.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The gradient-targeted transfer modulation strategy prioritizes training the monocular 3D detection network when its learning lags behind the depth-guided self-teaching network at the early training stage. As the monocular 3D detection network learns and improves gradually, it is enabled to provide more feedback progressively to the depth-guided self-teaching network.", + "bbox": [ + 496, + 167, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 282, + 651, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall objective consists of three losses including $L_{ud}$ , $L_{dep}$ and $L_{base}$ . $L_{ud}$ is the uncertainty-aware distillation loss as defined in Sec. 3.3. $L_{dep}$ is the depth loss for supervising the predicted depth map. $L_{base}$ includes losses for supervising 2D boxes prediction by 2D heads and the 3D box predictions, which has been adopted in prior CenterNet [55] and WeakMono3D [42]. We set the weight for each loss item to 1.0, and the overall loss function can be formulated as follows:", + "bbox": [ + 496, + 306, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {u d} + L _ {d e p} + L _ {b a s e}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 455, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 484, + 633, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 510, + 602, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments over the KITTI 3D dataset [12] and the nuScenes dataset [3] that have been widely adopted for benchmarking of 3D object detection methods. The KITTI 3D dataset consists of 7,481 images for training and 7,518 images for testing. The labels of the train set are publicly available and the labels of the test set are stored on a test server for evaluation. For ablation studies, we follow [5] which divides the 7,481 training samples into a new train set with 3,712 images and a validation set with 3,769 images. The nuScenes dataset comprises 1,000 video scenes, including RGB images captured by 6 surround-view cameras. The dataset is split into a training set (700 scenes), a validation set (150 scenes), and a test set (150 scenes). Following [34], the performance on the validation set is reported.", + "bbox": [ + 496, + 534, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Protocols", + "text_level": 1, + "bbox": [ + 500, + 756, + 696, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the KITTI 3D dataset, following [39], we adopt the evaluation metric $\\mathrm{AP}|_{R_{40}}$ which is the average of the AP of 40 recall points. We report the average precision on bird's eye view and 3D object detection as $\\mathrm{AP}_{BEV}|_{R_{40}}$ and $\\mathrm{AP}_{3D}|_{R_{40}}$ . In addition, as most weakly supervised 3D object detection methods apply IoU threshold of 0.7 for the test set and 0.5 for the validation set, we adopt the same thresholds for fair benchmarking. We adopt four metrics", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10512", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b0734fa605a937ef21ab45bb4ed871bea73f02636aac209d8f20214e6c65676e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.7)|R40
EasyModerateHard
WeakM3D [34]ResNet-50Weak11.82/5.035.66/2.264.08/1.63
WeakMono3D [42]DLA-3412.31/6.988.80/4.857.81/4.45
SKD-WM3D (Ours)DLA-3415.71/8.9510.15/5.548.08/4.53
", + "bbox": [ + 81, + 88, + 895, + 183 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/29d18220b2e52e9740655ff64e77a0128c9a1026235f706e78b0198de44635b6.jpg", + "table_caption": [ + "Table 1. Comparison on the performance of the Car category on KITTI test set. For all results, we use $\\mathrm{AP}|_{R_{40}}$ metrics with IoU threshold equal to 0.7. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
CenterNet [55]DLA-3434.36/20.0027.91/17.5024.65/15.57
MonoGRNet [35]VGG-1652.13/47.5935.99/32.2828.72/25.50
M3D-RPN [1]DenseNet-12153.35/48.5339.60/35.9431.76/28.59
MonoPair [6]DLA-34Full61.06/55.3847.63/42.3941.92/37.99
MonoDLE [28]DLA-3460.73/55.4146.87/43.4241.89/37.81
GUPNet [25]DLA-3461.78/57.6247.06/42.3340.88/37.59
Kinematic [2]DenseNet-12161.79/55.4444.68/39.4734.56/31.26
MonoDistill [7]DLA-3471.45/65.6953.11/49.3546.94/43.49
MonoDETR [53]*ResNet-5072.34/68.0551.97/48.4246.94/43.48
VS3D [36]VGG-16Weak31.59/22.6220.59/14.4316.28/10.91
Autolabels [51]ResNeXt10150.51/38.3130.97/19.9023.72/14.83
WeakM3D [34]ResNet-5058.20/50.1638.02/29.9430.17/23.11
WeakMono3D [42]DLA-3454.32/49.3742.83/39.0140.07/36.34
SKD-WM3D (Ours)DLA-3455.47/50.2144.35/41.5741.86/36.92
", + "bbox": [ + 83, + 234, + 888, + 500 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c36259030da3efc0c5c8fc03e9c3ac3163f0424e08174e5425a26e39193c4ca.jpg", + "table_caption": [ + "Table 2. Comparison on the performance of the Car category on KITTI val set. For all results, we use $\\mathrm{AP}|_{R_{40}}$ metric with IoU threshold equal to 0.5. * denotes this performance is reproduced from the official code. The best results of weakly supervised 3D object detection approaches are in bold." + ], + "table_footnote": [], + "table_body": "
MethodAP↑ATE↓ASE↓AAE↓
WeakM3D [34]0.2140.8140.2340.682
SKD-WM3D (Ours)0.2420.7950.2310.659
", + "bbox": [ + 81, + 575, + 467, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison on the performance of the Car category on nuScenes val set. The best results are in **bold**.", + "bbox": [ + 75, + 640, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the evaluation on the nuScenes dataset, namely, AP (Average Precision), ATE (Average Translation Error), ASE (Average Scale Error), and AAE (Average Attribute Error). Following [34], AVE (Average Velocity Error) and AOE (Average Orientation Error) are not reported due to the lack of supervision for velocity and movement direction in the weakly supervised approach.", + "bbox": [ + 75, + 696, + 468, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 815, + 294, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on 2 NVIDIA V100 GPUs with batch size of 16, and train the framework with 150 epochs. We use the Adam optimizer with the initial learning rate $1e^{-5}$ , which is gradually increased to $1e^{-3}$ for the first 5", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "epochs and decayed with rate 0.1 at the 90 and 120 epochs. We employ DLA-34 [49] as the detector's backbone. The pseudo ground truth of the depth map is generated with an off-the-shelf depth estimator [17] without using the ground truth of depth label.", + "bbox": [ + 496, + 578, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 665, + 867, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with several state-of-the-art weakly supervised monocular 3D detection methods on the KITTI test set. As Table 1 shows, our method achieves superior detection performance across all metrics. This superior performance is largely attributed to our designed self-knowledge distillation framework that extracts and exploits intrinsic depth information from a single-view image effectively. It should be highlighted that our method employs a single-view image exclusively without involving additional training data such as LiDAR point clouds [34] or multi-view image pairs [42].", + "bbox": [ + 496, + 688, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the benchmarking on the KITTI validation set. Specifically, we compare our method against both state-of-the-art weakly supervised monocular 3D de", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/057c72841b314a35e0ac0a50c0a326cbf5dc7208239f4aa06276bd42bbd1f46a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 85, + 346, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0600ff562e8ac152b44fa46026027a87d9e8aab8f50c6450e65fafad4e6accc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 87, + 620, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08cfb4d05d59cecb5ea2b04bbfa5b9cb8de5a9a444455febc5696e50c09c708b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 892, + 142 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d65e0f934483acce78f0b92473e426bec7c015c5c74af43a06d15a82baaadf27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 145, + 348, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60304c8f5a19a03ad701811eb6601a39cfff78425974ad0bdc67d00ce53dad3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 145, + 586, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e8ce927cd9f6f264a05af9ee17148a05d81c2859c7600b5711df746721c292b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 145, + 893, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df3e0f5292b7c4c7430b564b130d2f26fa936f378e06f1dd171a092f90d7119a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 263, + 348, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/876f9e572f579b8bf2508db789aed052dfcd62f342d6eefe691d6abcee33100b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 263, + 619, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b9f731d9955b215ed737a365cb6480da93be966318806fd5790d32ff911a7db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 263, + 892, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/314e618824552a66c1f8f7393644dd56a82856be675d3e1ecbf390dcb615c8c6.jpg", + "image_caption": [ + "Figure 3. Qualitative illustration on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The ground truth of LiDAR point clouds is utilized for visualization purposes only. Best viewed with zoom-in." + ], + "image_footnote": [], + "bbox": [ + 76, + 321, + 348, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aece145fc69eb5683cfab827127466604d765005f0773e5253e5be49c1cedfd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 321, + 586, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dba897aa40fcd90d30c7227385a89e2390837d76e9ab9f31fb19def9bf007c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 321, + 892, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/058031bb8fea94bea61f020151aad422aca6d5bc02058c9294da65d9587a4af5.jpg", + "image_caption": [ + "Uncertainty: 0.24" + ], + "image_footnote": [], + "bbox": [ + 78, + 488, + 236, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/098fbf338e7868bddc59d92105513f091bbe932038be96b13f210d2d39624f22.jpg", + "image_caption": [ + "Uncertainty: 0.41" + ], + "image_footnote": [], + "bbox": [ + 243, + 488, + 400, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75163d5cb3fa6b78eedcb513a2d195e35852eeb84dd28f0d7e163fdb73878a88.jpg", + "image_caption": [ + "Uncertainty: 0.67" + ], + "image_footnote": [], + "bbox": [ + 406, + 488, + 563, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1614e1879ea16e5919c26b747e9a341e92830d223a75c5835cad2ac7c9bbd06a.jpg", + "image_caption": [ + "Uncertainty: 0.75", + "Figure 4. Qualitative illustration of object detection and the corresponding detection uncertainties on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The detection accuracy is closely correlated with the detection uncertainty. Best viewed with zoom-in." + ], + "image_footnote": [], + "bbox": [ + 571, + 488, + 727, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25642c57775beb7f4493c77949fcb90743213f073fddaf535d8a530f15703460.jpg", + "image_caption": [ + "Uncertainty: 0.82" + ], + "image_footnote": [], + "bbox": [ + 733, + 488, + 890, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tection methods and fully supervised methods. It can be seen that our method achieves superior performance compared to WeakM3D [34] and WeakMono3D [42] across most metrics, even without using LiDAR point clouds or multi-view images. Notably, our method significantly outperforms WeakM3D [34] in the Moderate and Hard categories, primarily due to the sparse nature of distant LiDAR point clouds adversely affecting its performance. Additionally, its performance is even on par with several fully supervised methods [1, 35, 55].", + "bbox": [ + 75, + 674, + 468, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 shows the results on the nuScenes validation set. It can be seen that our proposed method outperforms WeakM3D [34] across all four evaluation metrics, validating the effectiveness of our approach.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results Figure 3 shows qualitative results with both 2D RGB images and 3D point clouds. In simple scenarios, our model achieves great prediction accuracy, which is largely attributed to the proposed self-knowledge distillation framework as well as the uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy, all working together to facilitate comprehensive 3D information extraction effectively. However, for heavily occluded or distant objects, the accuracy of orientation and depth estimation tends to drop, which is common for monocular 3D detection due to its ill-posed nature. In addition, Figure 4 shows the visualization of predicted bounding boxes and their corresponding uncertainties. It can be observed that the prediction accuracy of bounding boxes has a close correlation with the prediction uncertainty.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10514", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/89f750aa98c0b83629882ef2bcfe05d68a9b16c1e1e97429eda4451fff0a6a5b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IndexMDNDSNAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
1✓0.00/0.000.00/0.000.00/0.00
2✓45.23/40.9634.27/31.0230.17/26.27
3✓✓55.47/50.2144.35/41.5741.86/36.92
", + "bbox": [ + 81, + 88, + 467, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/622d00918e15729b8075cb642107b120747a0a8a1a3f8b5bac69ccfbf7493bfd.jpg", + "table_caption": [ + "Table 4. Ablation study of the proposed self-knowledge distillation framework. The best results are in bold. MDN denotes the Monocular 3D Detection Network, while DSN denotes the Depth-Guided Self-Teaching Network." + ], + "table_footnote": [], + "table_body": "
IndexLudTMSAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
149.95/44.6138.24/35.7437.28/34.82
2✓53.16/48.1341.85/39.0240.14/35.70
3✓52.35/46.3041.45/38.9139.73/35.44
4✓✓55.47/50.2144.35/41.5741.86/36.92
", + "bbox": [ + 81, + 241, + 465, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2024d358e5197623cec60688a03b31283cf287cfe9eb24e08bfb234cf683655e.jpg", + "table_caption": [ + "Table 5. Ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. The best results are in **bold**. $L_{ud}$ denotes the Uncertainty-Aware Distillation Loss, while TMS denotes the gradient-targeted transfer modulation strategy." + ], + "table_footnote": [], + "table_body": "
MethodWeakM3D [34]MonoDistill [7]*SKD-WM3D (Ours)*
FPS13.925.030.3
", + "bbox": [ + 81, + 422, + 465, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparison on inference speed of M3D methods. * denotes this method utilizes dense depth maps.", + "bbox": [ + 76, + 465, + 467, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 522, + 228, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct extensive ablation studies on the KITTI validation dataset to evaluate our designs. Specifically, we evaluated the efficacy of the two individual networks in the proposed self-knowledge distillation framework. Additionally, we examine the effect of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. Lastly, we evaluated the efficiency of our monocular 3D detection framework.", + "bbox": [ + 75, + 546, + 467, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-Knowledge Distillation Framework. We train two models to assess the contributions of the two networks in our proposed self-knowledge distillation framework. As Table 4 shows, training the monocular 3D detection network alone produces few meaningful detection results as the absence of depth information leads to ambiguous object localization along the depth dimension. As a comparison, training the depth-guided self-teaching network alone can produce reasonable detection results thanks to the depth map pseudo labels. In addition, training both subnetworks concurrently produces the best 3D detection, validating the effectiveness of extracting 3D information from a single image. We can also see that including the self-knowledge distillation on top of the depth-guided self-teaching network", + "bbox": [ + 75, + 688, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "greatly improves the detection by reducing the adverse effects of uncertain knowledge and enabling communication between the two subnetworks during training.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Uncertainty-Aware Distillation Loss and Gradient-Targeted Transfer Modulation Strategy. Table 5 shows the ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. It can be observed that the baseline does not perform well due to the adverse effect of uncertain knowledge and the asynchronous learning paces of the two subnetworks. On top of the baseline, including either the uncertainty-aware distillation loss or the gradient-targeted transfer modulation strategy improves the detection performance significantly, underscoring the importance of attaining high-certainty knowledge and synchronizing the learning paces of the two networks. In addition, combining the two designs achieves the best performance, highlighting their complementary nature and collaborative roles in knowledge acquisition and knowledge transfer.", + "bbox": [ + 496, + 154, + 890, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference speed comparison. Table 6 compares the inference speed on the KITTI validation set. Our method demonstrates superior efficiency thanks to our designed self-knowledge distillation framework, without utilizing complex network architectures during inference.", + "bbox": [ + 496, + 415, + 890, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 503, + 617, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we point out that previous weakly supervised monocular 3D detection methods either require additional LiDAR point clouds or paired images from multiple viewpoints or temporal sequences. To overcome these constraints, we propose a weakly supervised monocular 3D object detection approach that is exclusively grounded on single-view image inputs. Central to our approach is a self-knowledge distillation framework, which effectively harnesses the depth information within a single-view image with little computational overhead during inference. We further introduce an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy, facilitating knowledge acquisition and knowledge transfer, respectively. Finally, extensive experiments demonstrate the effectiveness of our method. Moving forward, we plan to further generalize our work to diverse and challenging scenarios, such as occlusions, varying lighting, and weather conditions, thereby enhancing its practical applicability.", + "bbox": [ + 496, + 529, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 815, + 658, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2, 6, 7", + "[2] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3d object detection in monocular video. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 135-152. Springer, 2020. 2, 6", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5", + "[4] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10379-10388, 2021. 2", + "[5] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2147-2156, 2016. 1, 2, 5", + "[6] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 2, 6", + "[7] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. International Conference on Learning Representations, 2022. 2, 6, 8", + "[8] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In International Conference on Machine Learning, pages 2006-2015. PMLR, 2020. 3", + "[9] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3872-3882, 2023. 3", + "[10] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1000-1001, 2020. 2", + "[11] Huan-ang Gao, Beiwen Tian, Pengfei Li, Hao Zhao, and Guyue Zhou. Dqs3d: Densely-matched quantization-aware semi-supervised 3d detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 5", + "[13] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1440-1448, 2015. 5", + "[14] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2", + "[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3", + "[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[17] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In International Conference on Robotics and Automation, pages 13656-13662. IEEE, 2021. 1, 2, 4, 6", + "[18] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. Advances in Neural Information Processing Systems, 35:33716-33727, 2022. 3", + "[19] Mingi Ji, Seungjae Shin, Seunghyun Hwang, Gibeom Park, and Il-Chul Moon. Refine myself by teaching myself: Feature refinement via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10664–10673, 2021. 3", + "[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in Neural Information Processing Systems, 30, 2017. 5", + "[21] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11867-11876, 2019. 1, 2", + "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2980-2988, 2017. 4", + "[23] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7096-7104, 2019. 3", + "[24] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2", + "[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 5, 6" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhipeng Luo, Zhongang Cai, Changqing Zhou, Gongjie Zhang, Haiyu Zhao, Shuai Yi, Shijian Lu, Hongsheng Li, Shanghang Zhang, and Ziwei Liu. Unsupervised domain adaptive 3d detection with multi-level consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8866-8875, 2021. 2", + "[27] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6851–6860, 2019. 2", + "[28] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 6", + "[29] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 515-531. Springer, 2020. 2", + "[30] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in Neural Information Processing Systems, 32, 2019. 3", + "[31] J Krishna Murthy, GV Sai Krishna, Falak Chhaya, and K Madhava Krishna. Reconstructing vehicles from a single image: Shape priors for road scene understanding. In International Conference on Robotics and Automation, pages 724-731. IEEE, 2017. 2", + "[32] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3", + "[33] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 71-88. Springer, 2022. 1, 5", + "[34] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. International Conference on Learning Representations, 2022. 1, 2, 4, 5, 6, 7, 8", + "[35] Zengyi Qin, Jinglu Wang, and Yan Lu. Monognet: A geometric reasoning network for monocular 3d object localization. Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 2, 6, 7", + "[36] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 6", + "[37] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 1", + "[38] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets:" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3", + "[39] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection: From single to multiclass recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1219-1231, 2020. 1, 5", + "[40] Dawei Sun, Anbang Yao, Aojun Zhou, and Hao Zhao. Deeply-supervised knowledge synergy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2019. 3", + "[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 3", + "[42] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular 3d object detection using multi-view projection and direction consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7", + "[43] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 3", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4", + "[45] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2", + "[46] Ting-Bing Xu and Cheng-Lin Liu. Data-distortion guided self-distillation for deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5565-5572, 2019. 3", + "[47] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[48] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. International Conference on Learning Representations, 2020. 2", + "[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2403-2412, 2018. 6", + "[50] Sukmin Yun, Jongjin Park, Kimin Lee, and Jinwoo Shin. Regularizing class-wise predictions via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10517", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 13876-13885, 2020. 3", + "[51] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 6", + "[52] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3713-3722, 2019. 3", + "[53] Renrui Zhang, Han Qiu, Tai Wang, Xuanzhuo Xu, Ziyu Guo, Yu Qiao, Peng Gao, and Hongsheng Li. Monodetr: Depth-aware transformer for monocular 3d object detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 6", + "[54] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 3", + "[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5, 6, 7", + "[56] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[57] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Advances in Neural Information Processing Systems, 31, 2018. 3" + ], + "bbox": [ + 78, + 90, + 468, + 544 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10518", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_model.json b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3761df1351835defb3e401e239b498c71ff5f85e --- /dev/null +++ b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "Weakly Supervised Monocular 3D Detection with a Single-View Image" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.813, + 0.235 + ], + "angle": 0, + "content": "Xueying Jiang\\(^{1}\\) Sheng Jin\\(^{1}\\) Lewei Lu\\(^{2}\\) Xiaoqin Zhang\\(^{3}\\) Shijian Lu\\(^{1*}\\) \n\\(^{1}\\)S-Lab, Nanyang Technological University \n\\(^{2}\\)Sensetime Research \n\\(^{3}\\)College of Computer Science and Technology, Zhejiang University of Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Monocular 3D detection (M3D) aims for precise 3D object localization from a single-view image which usually involves labor-intensive annotation of 3D detection boxes. Weakly supervised M3D has recently been studied to obviate the 3D annotation process by leveraging many existing 2D annotations, but it often requires extra training data such as LiDAR point clouds or multi-view images which greatly degrades its applicability and usability in various applications. We propose SKD-WM3D, a weakly supervised monocular 3D detection framework that exploits depth information to achieve M3D with a single-view image exclusively without any 3D annotations or other training data. One key design in SKD-WM3D is a self-knowledge distillation framework, which transforms image features into 3D-like representations by fusing depth information and effectively mitigates the inherent depth ambiguity in monocular scenarios with little computational overhead in inference. In addition, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Extensive experiments show that SKD-WM3D surpasses the state-of-the-art clearly and is even on par with many fully supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Monocular 3D detection (M3D) has emerged as one key component in the area of autonomous driving and computer vision. Its primary target is to recognize objects and obtain their 3D localization from single-view images. Thanks to its low deployment cost, M3D [5, 34] has attracted increasing attention in both academic and industrial sectors, achieving very impressive progress in recent years. On the other hand, most existing studies [21, 33, 37, 39] adopt a fully supervised setup which have been facing increasing scalability concern as large-scale 3D boxes are often labor-intensive" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.334 + ], + "angle": 0, + "content": "to collect. Effective M3D training without 3D annotations has become a critical issue while handling M3D problems in various research and practical tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.893, + 0.716 + ], + "angle": 0, + "content": "Weakly supervised M3D (WM3D) [34] has recently been explored for learning effective 3D detectors without 3D box annotations, aiming to exploit 2D annotations to make up for the absence of 3D information. For example, WeakM3D [34] exploits LiDAR point clouds to infer 3D information as illustrated in Figure 1(a). However, it requires costly and complicated LiDAR sensors to collect point clouds which limits its applicability and usability greatly. WeakMono3D [42] employs 2D information only by either leveraging multi-view stereo with images from multiple cameras or constructing pseudo-multi-view perspective from sequential video frames as illustrated in Figure 1(b). However, collecting multi-view images is complicated, and resorting to a pseudo multi-view perspective degrades the detection performance clearly. With the advance of single-view depth estimation, WM3D with depth from a single-view image presents a potential solution for compensating the absence of 3D annotations. On the other hand, direct integration of such depth into existing frameworks often necessitates complex network architectures which further incurs significant computational costs. This gives rise to a pertinent question: When not using additional LiDAR point clouds or multi-view image pairs, is it possible to harness the depth from off-the-shelf depth estimators without introducing much computational overhead in inference?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We design SKD-WM3D, a novel weakly supervised monocular 3D object detection method that is exclusively grounded on single-view images. One key design in SKD-WM3D is a self-knowledge distillation framework which consists of a Depth-guided Self-teaching Network (DSN) and a Monocular 3D Detection Network (MDN). As illustrated in Figure 1(c), SKD-WM3D utilizes depth information obtained from an off-the-shelf depth estimator [17] to enhance the 3D localization ability of DSN and transfers such ability to MDN via self-knowledge distillation. Such self-distillation design enables MDN to unearth the intrinsic depth information from single-view images independently," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10508" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.242, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.233, + 0.251 + ], + "angle": 0, + "content": "(a) LiDAR-based WM3D" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.088, + 0.554, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.238, + 0.469, + 0.251 + ], + "angle": 0, + "content": "(b) Multi-view WM3D" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.089, + 0.891, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.238, + 0.829, + 0.251 + ], + "angle": 0, + "content": "(c) Single-view WM3D (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.893, + 0.325 + ], + "angle": 0, + "content": "Figure 1. Different paradigms in weakly supervised monocular 3D detection. Our approach in (c) leverages Pseudo Depth Labels from a single-view image to achieve weakly supervised monocular 3D detection, requiring no extra training data like LiDAR point clouds or multi-view images as in (a) and (b). It improves usability and applicability greatly. The Pseudo Depth Labels are obtained with an off-the-shelf depth estimator [17] without extra training and ground-truth depth labels. Data in red denotes extra data in network training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.562 + ], + "angle": 0, + "content": "bypassing additional modules such as pre-trained depth estimation networks and leading to precise and efficient 3D localization with little computational overhead during inference. On top of DSN and MDN, we design an uncertainty-aware distillation loss to optimize the utilization of the transferred 3D localization knowledge by weighting up more certain knowledge while weighting down less certain knowledge. In addition, we design a gradient-targeted transfer modulation strategy to synchronize the learning paces of DSN and MDN during the process of learning 3D localization knowledge, by prioritizing MDN learning at the initial stage when MDN lags behind DSN and enabling it to provide more feedback to DSN when MDN is better trained at late stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.469, + 0.79 + ], + "angle": 0, + "content": "Our contribution can be summarized in three aspects. First, we design a novel framework that achieves weakly supervised monocular 3D detection by distilling knowledge between a depth-guided self-teaching network and a monocular 3D detection network. Without any extra training data like LiDAR point clouds or multi-view images, the framework exploits depth exclusively from a single image with little computational overhead in inference. Second, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Third, the proposed approach clearly outperforms the state-of-the-art in weakly supervised monocular 3D detection, and its performance is even on par with several fully supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.219, + 0.821 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.302, + 0.846 + ], + "angle": 0, + "content": "2.1. Monocular 3D Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Monocular 3D object detection aims to predict 3D object localization from single-view images. Standard monocular detectors [1, 6, 14, 53, 56] operate solely on single images," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.683 + ], + "angle": 0, + "content": "without utilizing additional data. However, the inherent depth ambiguity of monocular detection significantly hinders its performance compared to its stereo counterparts. To address this limitation, various approaches seek solutions with the help of extra data, such as LiDAR point clouds [4, 7, 21, 27], video sequences [2], 3D CAD models [5, 24, 31], and depth estimation [10, 35, 45, 48]. Specifically, MonoRUn [4] adopts an uncertainty-aware regional reconstruction network for regressing pixel-associated 3D object coordinates with LiDAR point clouds as extra supervision. MonoDistill [7] introduces an effective distillation-based approach that incorporates spatial information from LiDAR signals into monocular 3D detection. Additionally, pseudo-LiDAR-based methods [45, 48] convert estimated depth maps to simulate the real LiDAR point clouds to utilize the well-designed LiDAR-based 3D detector. During inference, compared with methods using depth estimation, our method eliminates the need for pseudo depth labels and complex network architectures, with little computational overhead. Besides, existing fully supervised methods require large-scale 3D box ground truth, which is labor-intensive to collect and annotate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.842, + 0.712 + ], + "angle": 0, + "content": "2.2. Weakly Supervised 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Due to the high cost of annotating 3D boxes in the 3D object detection task, various weakly supervised approaches [26, 29, 34, 36] have been proposed. For example, WS3D [29] presents a weakly supervised method for 3D LiDAR object detection, which requires only a limited number of weakly annotated scenes with center-annotated BEV maps. VS3D [36] introduces a cross-model knowledge distillation strategy to transfer the knowledge from the RGB domain to the point cloud domain, using LiDAR point clouds as weak supervision. Recent research on weakly supervised 3D object detection has turned to explore the monocular setting. For example, WeakM3D [34] generates 2D boxes to select" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10509" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.896, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.379, + 0.893, + 0.465 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed self-knowledge distillation network. The framework consists of a depth-guided self-teaching network and a monocular 3D detection network. The depth-guided self-teaching network acquires comprehensive 3D localization knowledge by leveraging depth information and transfers its learned expertise to the monocular 3D detection network via soft label distillation to enhance its performance. We design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy to facilitate the knowledge transfer between the two networks effectively. During inference, the monocular 3D detection network extracts intrinsic depth information from single-view images independently with little computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.473, + 0.703 + ], + "angle": 0, + "content": "RoI LiDAR point clouds as weak supervision and then predicts 3D boxes that closely align with the selected RoI LiDAR point clouds. More recently, WeakMono3D [42] eliminates the need for LiDAR, offering both multi-view and single-view yet multi-frame versions. While the former acquires stereo image inputs from multiple cameras, the latter constructs a pseudo-multi-view perspective using multiple video frames. The multi-frame version exhibits inferior 3D scene comprehension compared with the multi-view approach due to its smaller inter-frame disparity, leading to degraded performance. Instead of requiring extra training data like LiDAR point clouds or multi-view images, we tackle the challenge of weakly supervised monocular 3D detection by leveraging a single-view image exclusively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.325, + 0.727 + ], + "angle": 0, + "content": "2.3. Self-Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Knowledge distillation [8, 9, 11, 16, 18, 23, 32, 38, 43, 54] aims to transfer knowledge from a pre-trained teacher network to a student network for improving its performance. Self-knowledge distillation [30, 41, 47], distinct from traditional knowledge distillation, leverages the information within the student network to facilitate its learning without the pre-trained teacher network. Specifically, data augmentation approach [15, 46, 50] transfers knowledge through different distortions of the same training data. However, they are susceptible to inappropriate augmentations, such as improper instance rotation or distortion, potentially intro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.657 + ], + "angle": 0, + "content": "ducing noise that hampers network learning. Another typical approach exploits auxiliary networks [52, 57]. For example, DKS [40] introduces auxiliary supervision branches and pairwise knowledge alignments, while FRSKD [19] adds a new branch supervised by the original features and utilizes both soft-label and feature-map distillation. Our work is the first that introduces self-knowledge distillation with auxiliary networks for weakly supervised monocular 3D detection. It effectively exploits depth information from single-view images with little computational overhead during inference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.637, + 0.69 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.894, + 0.819 + ], + "angle": 0, + "content": "This section presents the proposed SKD-WM3D. First, the problem definition and overview are presented in Sec. 3.1. Then detailed designs of SKD-WM3D are introduced, including the self-knowledge distillation framework in Sec. 3.2, the uncertainty-aware distillation loss in Sec. 3.3 and the gradient-targeted transfer modulation strategy in Sec. 3.4. Finally, loss functions are presented in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.795, + 0.848 + ], + "angle": 0, + "content": "3.1. Problem Definition and Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Weakly supervised monocular 3D detection takes an RGB image and the corresponding 2D bounding boxes as supervision, aiming to classify objects and determine their" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "bounding boxes in 3D space without involving any 3D annotations in training. The prediction of each object is composed of the object category \\( C \\), a 2D bounding box \\( B^{2D} \\) and a 3D bounding box \\( B^{3D} \\). Specifically, the 3D box \\( B^{3D} \\) can be further decomposed to the object 3D location \\( (x_{3D}, y_{3D}, z_{3D}) \\), the object dimension with height, width and length \\( (h_{3D}, w_{3D}, l_{3D}) \\), as well as orientation \\( \\theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.199, + 0.468, + 0.53 + ], + "angle": 0, + "content": "We design a self-knowledge distillation framework to tackle the challenge of weakly supervised monocular 3D detection from a single-view image. As Figure 2 shows, the framework consists of two subnetworks including a Depth-Guided Self-Teaching Network and a Monocular 3D Detection Network. In the Depth-Guided Self-Teaching Network, the global features \\( F_{G} \\) extracted by the backbone are fed into a Depth Head to obtain depth features. Next, the global features \\( F_{G} \\) and the extracted depth features are fed into a Fusion Layer to obtain 3D-like features \\( F_{3D} \\). The 3D-like features of each object are then obtained via RoIAlign, and further fed to a Depth-Aware 3D Head to predict 3D box \\( \\widehat{B}_p^{3D} \\) and uncertainty \\( \\widehat{U} \\). In the Monocular 3D Detection Network, We first use RoIAlign to generate object-level features from the global features \\( F_{G} \\), and then feed them to a 2D-to-3D Head to predict 3D box \\( \\widehat{B}_p^{2D} \\) and uncertainty \\( U \\). The 3D boxes predicted by both networks are further projected into 2D boxes. Moreover, we design an uncertainty-aware distillation loss \\( L_{ud} \\) to obtain low-uncertainty knowledge, and a gradient-targeted transfer modulation strategy to synchronize the learning paces between the two networks by controlling gradients \\( \\widehat{G} \\) and \\( G \\) of \\( L_{ud} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.538, + 0.418, + 0.552 + ], + "angle": 0, + "content": "3.2. Self-Knowledge Distillation Framework" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.561, + 0.468, + 0.636 + ], + "angle": 0, + "content": "The self-knowledge distillation framework enhances the 3D localization ability of the depth-guided self-teaching network by utilizing depth information from an off-the-shelf depth estimator and then transfers the ability to the monocular 3D detection network via self-knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.654, + 0.468, + 0.759 + ], + "angle": 0, + "content": "Depth-Guided Self-Teaching Network. To equip the self-teaching network with 3D localization ability, we propose to learn from global features \\( F_{G} \\) and depth information from an off-the-shelf depth estimator to acquire comprehensive 3D knowledge. The depth information is exploited via two major designs. Firstly, we introduce a depth head \\( \\mathcal{D} \\) that extracts depth features \\( F_{D} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.774, + 0.468, + 0.79 + ], + "angle": 0, + "content": "\\[\nF _ {D} = \\mathcal {D} \\left(F _ {G}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The depth features \\( F_{D} \\) are exploited to generate depth maps \\( D_{p} \\), where the depth map generation is supervised by the pseudo ground truth of the depth map \\( D_{gt} \\) that is predicted by an off-the-shelf depth estimator by using the focal loss [22] as depth loss \\( L_{dep} \\). Hence, the depth features can be acquired by the depth-guided self-teaching network effectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.182 + ], + "angle": 0, + "content": "Remark 1. We generate depth pseudo labels using an off-the-shelf depth estimator [17] with frozen weights, eliminating the need for additional training and ground-truth depth labels. Adopting an off-the-shelf depth estimator incurs negligible costs as compared with prior studies that require either point clouds [34] or multi-view images [42]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Secondly, we obtain 3D-like features \\( F_{G3D} \\) by integrating the depth features \\( F_{D} \\) that provide information along the depth dimension, as well as the global features \\( F_{G} \\) that capture knowledge about the 2D image plane. Specifically, we design a fusion layer that fuses the depth features \\( F_{D} \\) with the global features \\( F_{G} \\) to derive the \\( F_{G3D} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.292, + 0.891, + 0.308 + ], + "angle": 0, + "content": "\\[\nF _ {G 3 D} = F F N \\left(C A \\left(S A \\left(F _ {D}\\right), F _ {G}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.318, + 0.891, + 0.424 + ], + "angle": 0, + "content": "where the \\(FFN\\) is the feed-forward network, and \\(CA\\), \\(SA\\) denote CrossAttention, SelfAttention, respectively. The structures of CrossAttention and SelfAttention employ the standard transformer architecture [44]. The obtained 3D comprehension improves the network's ability to precisely locate objects, effectively mitigating depth ambiguity arising from single-view image input." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Monocular 3D Detection Network. The monocular 3D detection network acquires the 3D localization knowledge from the depth-guided self-teaching network. By distilling soft labels generated by the depth-guided self-teaching network, the monocular 3D detection network can extract intrinsic depth information from images independently during inference. This kills the need for additional complex modules such as pre-trained depth estimation networks or depth fusion modules, facilitating the inference with little computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.81, + 0.625 + ], + "angle": 0, + "content": "3.3. Uncertainty-Aware Distillation Loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.634, + 0.891, + 0.784 + ], + "angle": 0, + "content": "During the knowledge distillation process, uncertain knowledge could affect the network training negatively if all transferred knowledge is treated equally. To benefit more from certain knowledge and weaken the effect of uncertain knowledge, we design an uncertainty-aware distillation loss between the 3D boxes that are predicted by the two networks in the self-knowledge distillation framework. The uncertainty-aware distillation loss exploits the prediction uncertainty to modulate the distillation loss magnitude as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.811, + 0.891, + 0.855 + ], + "angle": 0, + "content": "\\[\nL _ {u d} = \\frac {L _ {d}}{\\operatorname* {m i n} ((\\widehat {U} + U) / 2 , \\alpha)} + \\left\\| \\operatorname * {m i n} (\\frac {\\widehat {U} + U}{2}, \\alpha) \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(\\widehat{U}\\) and \\(U\\) denote the uncertainty of the 3D boxes that are predicted by the two networks, respectively. Here" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10511" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.236 + ], + "angle": 0, + "content": "we assume the 3D box predictions have the Laplace distribution, and we adopt the standard deviations as the uncertainties, inspired by [20, 25, 33]. \\(\\left\\| \\min (\\frac{\\widehat{U} + U}{2},\\alpha)\\right\\| ^2\\) is the L2 regularization, and \\(\\alpha\\) is a fixed value set to 0.1. \\(L_{d}\\) denotes the basic distillations loss, and we employ the commonly used SmoothL1 [13] loss to enforce the consistency between the 3D boxes predicted by the two networks. The SmoothL1 loss leaves a soft margin when computing the difference between the two 3D boxes:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.256, + 0.47, + 0.304 + ], + "angle": 0, + "content": "\\[\nL _ {d} = \\left\\{ \\begin{array}{l l} 0. 5 \\times \\left(\\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D}\\right) ^ {2}, & \\text {i f} \\left| \\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \\right| < 1. 0 \\\\ \\left| \\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \\right| - 0. 5, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.304, + 0.47, + 0.351 + ], + "angle": 0, + "content": "where \\(\\widehat{B}_p^{3D}\\) and \\(B_p^{3D}\\) are the predicted 3D boxes from the depth-guided self-teaching network and the monocular 3D detection network, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.584 + ], + "angle": 0, + "content": "Remark 2. The proposed uncertainty-aware distillation loss \\( L_{ud} \\) integrates average uncertainty \\( \\frac{\\widehat{U} + U}{2} \\) as regularization and a weighted component for the basic distillation loss \\( L_{d} \\), allowing adaptive learning adjustments based on the knowledge's uncertainty level. Specifically, when dealing with uncertain knowledge, a smaller weight is assigned to the basic distillation loss \\( L_{d} \\) to mitigate potential adverse effects on network learning. Consequently, the network prioritizes optimizing uncertainty reduction in such scenarios. When dealing with certain knowledge, the network emphasizes optimizing the basic distillation loss \\( L_{d} \\) due to its higher weight. Notably, the basic distillation loss \\( L_{d} \\) simply considers box consistency, while integrating uncertainty is beneficial for enhancing the knowledge distillation process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.345, + 0.608 + ], + "angle": 0, + "content": "3.4. Transfer Modulation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.689 + ], + "angle": 0, + "content": "The depth-guided self-teaching network, which leverages depth information to predict 3D boxes, transfers its learned 3D knowledge to the monocular 3D detection network. The asynchronous learning paces of the two networks pose potential challenges to effective 3D knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.856 + ], + "angle": 0, + "content": "We design a gradient-targeted transfer modulation strategy to synchronize the learning pace of the depth-guided self-teaching network and the monocular 3D detection network. We modulate the knowledge transfer dynamically, by controlling the gradients from the uncertainty-aware distillation loss \\( L_{ud} \\). Specifically, we adapt the gradients based on the 2D projection performance of each network, assigning smaller backward gradients for the good-performing network and higher backward gradients for the bad-performing network. The gradient-targeted transfer modulation strategy is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.866, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\widehat {G} ^ {\\prime} = \\frac {2 \\times \\widehat {L} _ {p r o j}}{\\widehat {L} _ {p r o j} + L _ {p r o j}} \\times \\widehat {G}, G ^ {\\prime} = \\frac {2 \\times L _ {p r o j}}{\\widehat {L} _ {p r o j} + L _ {p r o j}} \\times G, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Where \\(\\widehat{G}\\) and \\(G\\) are the original gradients of the two networks, \\(\\widehat{G}^{\\prime}\\) and \\(G^{\\prime}\\) are the modified gradients, \\(\\widehat{L}_{proj}\\) and \\(L_{proj}\\) are projection losses, computed between the projected 2D boxes from 3D box predictions and 2D box annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.273 + ], + "angle": 0, + "content": "The gradient-targeted transfer modulation strategy prioritizes training the monocular 3D detection network when its learning lags behind the depth-guided self-teaching network at the early training stage. As the monocular 3D detection network learns and improves gradually, it is enabled to provide more feedback progressively to the depth-guided self-teaching network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.283, + 0.653, + 0.299 + ], + "angle": 0, + "content": "3.5. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.442 + ], + "angle": 0, + "content": "The overall objective consists of three losses including \\( L_{ud} \\), \\( L_{dep} \\) and \\( L_{base} \\). \\( L_{ud} \\) is the uncertainty-aware distillation loss as defined in Sec. 3.3. \\( L_{dep} \\) is the depth loss for supervising the predicted depth map. \\( L_{base} \\) includes losses for supervising 2D boxes prediction by 2D heads and the 3D box predictions, which has been adopted in prior CenterNet [55] and WeakMono3D [42]. We set the weight for each loss item to 1.0, and the overall loss function can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.456, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\nL = L _ {u d} + L _ {d e p} + L _ {b a s e}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.486, + 0.634, + 0.504 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.603, + 0.526 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.747 + ], + "angle": 0, + "content": "We conduct experiments over the KITTI 3D dataset [12] and the nuScenes dataset [3] that have been widely adopted for benchmarking of 3D object detection methods. The KITTI 3D dataset consists of 7,481 images for training and 7,518 images for testing. The labels of the train set are publicly available and the labels of the test set are stored on a test server for evaluation. For ablation studies, we follow [5] which divides the 7,481 training samples into a new train set with 3,712 images and a validation set with 3,769 images. The nuScenes dataset comprises 1,000 video scenes, including RGB images captured by 6 surround-view cameras. The dataset is split into a training set (700 scenes), a validation set (150 scenes), and a test set (150 scenes). Following [34], the performance on the validation set is reported." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.697, + 0.772 + ], + "angle": 0, + "content": "4.2. Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For the KITTI 3D dataset, following [39], we adopt the evaluation metric \\(\\mathrm{AP}|_{R_{40}}\\) which is the average of the AP of 40 recall points. We report the average precision on bird's eye view and 3D object detection as \\(\\mathrm{AP}_{BEV}|_{R_{40}}\\) and \\(\\mathrm{AP}_{3D}|_{R_{40}}\\). In addition, as most weakly supervised 3D object detection methods apply IoU threshold of 0.7 for the test set and 0.5 for the validation set, we adopt the same thresholds for fair benchmarking. We adopt four metrics" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10512" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.897, + 0.184 + ], + "angle": 0, + "content": "
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.7)|R40
EasyModerateHard
WeakM3D [34]ResNet-50Weak11.82/5.035.66/2.264.08/1.63
WeakMono3D [42]DLA-3412.31/6.988.80/4.857.81/4.45
SKD-WM3D (Ours)DLA-3415.71/8.9510.15/5.548.08/4.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.194, + 0.895, + 0.224 + ], + "angle": 0, + "content": "Table 1. Comparison on the performance of the Car category on KITTI test set. For all results, we use \\( \\mathrm{AP}|_{R_{40}} \\) metrics with IoU threshold equal to 0.7. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.235, + 0.889, + 0.501 + ], + "angle": 0, + "content": "
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
CenterNet [55]DLA-3434.36/20.0027.91/17.5024.65/15.57
MonoGRNet [35]VGG-1652.13/47.5935.99/32.2828.72/25.50
M3D-RPN [1]DenseNet-12153.35/48.5339.60/35.9431.76/28.59
MonoPair [6]DLA-34Full61.06/55.3847.63/42.3941.92/37.99
MonoDLE [28]DLA-3460.73/55.4146.87/43.4241.89/37.81
GUPNet [25]DLA-3461.78/57.6247.06/42.3340.88/37.59
Kinematic [2]DenseNet-12161.79/55.4444.68/39.4734.56/31.26
MonoDistill [7]DLA-3471.45/65.6953.11/49.3546.94/43.49
MonoDETR [53]*ResNet-5072.34/68.0551.97/48.4246.94/43.48
VS3D [36]VGG-16Weak31.59/22.6220.59/14.4316.28/10.91
Autolabels [51]ResNeXt10150.51/38.3130.97/19.9023.72/14.83
WeakM3D [34]ResNet-5058.20/50.1638.02/29.9430.17/23.11
WeakMono3D [42]DLA-3454.32/49.3742.83/39.0140.07/36.34
SKD-WM3D (Ours)DLA-3455.47/50.2144.35/41.5741.86/36.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.51, + 0.893, + 0.555 + ], + "angle": 0, + "content": "Table 2. Comparison on the performance of the Car category on KITTI val set. For all results, we use \\(\\mathrm{AP}|_{R_{40}}\\) metric with IoU threshold equal to 0.5. * denotes this performance is reproduced from the official code. The best results of weakly supervised 3D object detection approaches are in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.576, + 0.468, + 0.632 + ], + "angle": 0, + "content": "
MethodAP↑ATE↓ASE↓AAE↓
WeakM3D [34]0.2140.8140.2340.682
SKD-WM3D (Ours)0.2420.7950.2310.659
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Table 3. Comparison on the performance of the Car category on nuScenes val set. The best results are in **bold**." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.804 + ], + "angle": 0, + "content": "for the evaluation on the nuScenes dataset, namely, AP (Average Precision), ATE (Average Translation Error), ASE (Average Scale Error), and AAE (Average Attribute Error). Following [34], AVE (Average Velocity Error) and AOE (Average Orientation Error) are not reported due to the lack of supervision for velocity and movement direction in the weakly supervised approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.295, + 0.833 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We conduct experiments on 2 NVIDIA V100 GPUs with batch size of 16, and train the framework with 150 epochs. We use the Adam optimizer with the initial learning rate \\( 1e^{-5} \\), which is gradually increased to \\( 1e^{-3} \\) for the first 5" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.656 + ], + "angle": 0, + "content": "epochs and decayed with rate 0.1 at the 90 and 120 epochs. We employ DLA-34 [49] as the detector's backbone. The pseudo ground truth of the depth map is generated with an off-the-shelf depth estimator [17] without using the ground truth of depth label." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.869, + 0.683 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We compare our method with several state-of-the-art weakly supervised monocular 3D detection methods on the KITTI test set. As Table 1 shows, our method achieves superior detection performance across all metrics. This superior performance is largely attributed to our designed self-knowledge distillation framework that extracts and exploits intrinsic depth information from a single-view image effectively. It should be highlighted that our method employs a single-view image exclusively without involving additional training data such as LiDAR point clouds [34] or multi-view image pairs [42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows the benchmarking on the KITTI validation set. Specifically, we compare our method against both state-of-the-art weakly supervised monocular 3D de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10513" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.348, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.088, + 0.621, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.893, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.146, + 0.349, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.146, + 0.587, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.146, + 0.894, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.264, + 0.349, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.264, + 0.62, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.264, + 0.893, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.323, + 0.349, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.323, + 0.587, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.323, + 0.893, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.448, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Figure 3. Qualitative illustration on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The ground truth of LiDAR point clouds is utilized for visualization purposes only. Best viewed with zoom-in." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.489, + 0.237, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.575, + 0.215, + 0.588 + ], + "angle": 0, + "content": "Uncertainty: 0.24" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.489, + 0.401, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.575, + 0.377, + 0.588 + ], + "angle": 0, + "content": "Uncertainty: 0.41" + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.489, + 0.565, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.575, + 0.542, + 0.588 + ], + "angle": 0, + "content": "Uncertainty: 0.67" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.489, + 0.728, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.575, + 0.706, + 0.588 + ], + "angle": 0, + "content": "Uncertainty: 0.75" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.489, + 0.892, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.575, + 0.869, + 0.588 + ], + "angle": 0, + "content": "Uncertainty: 0.82" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.606, + 0.893, + 0.648 + ], + "angle": 0, + "content": "Figure 4. Qualitative illustration of object detection and the corresponding detection uncertainties on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The detection accuracy is closely correlated with the detection uncertainty. Best viewed with zoom-in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.826 + ], + "angle": 0, + "content": "tection methods and fully supervised methods. It can be seen that our method achieves superior performance compared to WeakM3D [34] and WeakMono3D [42] across most metrics, even without using LiDAR point clouds or multi-view images. Notably, our method significantly outperforms WeakM3D [34] in the Moderate and Hard categories, primarily due to the sparse nature of distant LiDAR point clouds adversely affecting its performance. Additionally, its performance is even on par with several fully supervised methods [1, 35, 55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Table 3 shows the results on the nuScenes validation set. It can be seen that our proposed method outperforms WeakM3D [34] across all four evaluation metrics, validating the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results Figure 3 shows qualitative results with both 2D RGB images and 3D point clouds. In simple scenarios, our model achieves great prediction accuracy, which is largely attributed to the proposed self-knowledge distillation framework as well as the uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy, all working together to facilitate comprehensive 3D information extraction effectively. However, for heavily occluded or distant objects, the accuracy of orientation and depth estimation tends to drop, which is common for monocular 3D detection due to its ill-posed nature. In addition, Figure 4 shows the visualization of predicted bounding boxes and their corresponding uncertainties. It can be observed that the prediction accuracy of bounding boxes has a close correlation with the prediction uncertainty." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10514" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.161 + ], + "angle": 0, + "content": "
IndexMDNDSNAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
1✓0.00/0.000.00/0.000.00/0.00
2✓45.23/40.9634.27/31.0230.17/26.27
3✓✓55.47/50.2144.35/41.5741.86/36.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.171, + 0.47, + 0.228 + ], + "angle": 0, + "content": "Table 4. Ablation study of the proposed self-knowledge distillation framework. The best results are in bold. MDN denotes the Monocular 3D Detection Network, while DSN denotes the Depth-Guided Self-Teaching Network." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.242, + 0.466, + 0.328 + ], + "angle": 0, + "content": "
IndexLudTMSAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
149.95/44.6138.24/35.7437.28/34.82
2✓53.16/48.1341.85/39.0240.14/35.70
3✓52.35/46.3041.45/38.9139.73/35.44
4✓✓55.47/50.2144.35/41.5741.86/36.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.338, + 0.468, + 0.409 + ], + "angle": 0, + "content": "Table 5. Ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. The best results are in **bold**. \\( L_{ud} \\) denotes the Uncertainty-Aware Distillation Loss, while TMS denotes the gradient-targeted transfer modulation strategy." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.423, + 0.466, + 0.457 + ], + "angle": 0, + "content": "
MethodWeakM3D [34]MonoDistill [7]*SKD-WM3D (Ours)*
FPS13.925.030.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.467, + 0.468, + 0.497 + ], + "angle": 0, + "content": "Table 6. Comparison on inference speed of M3D methods. * denotes this method utilizes dense depth maps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.523, + 0.23, + 0.54 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.468, + 0.668 + ], + "angle": 0, + "content": "We conduct extensive ablation studies on the KITTI validation dataset to evaluate our designs. Specifically, we evaluated the efficacy of the two individual networks in the proposed self-knowledge distillation framework. Additionally, we examine the effect of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. Lastly, we evaluated the efficiency of our monocular 3D detection framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Self-Knowledge Distillation Framework. We train two models to assess the contributions of the two networks in our proposed self-knowledge distillation framework. As Table 4 shows, training the monocular 3D detection network alone produces few meaningful detection results as the absence of depth information leads to ambiguous object localization along the depth dimension. As a comparison, training the depth-guided self-teaching network alone can produce reasonable detection results thanks to the depth map pseudo labels. In addition, training both subnetworks concurrently produces the best 3D detection, validating the effectiveness of extracting 3D information from a single image. We can also see that including the self-knowledge distillation on top of the depth-guided self-teaching network" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "greatly improves the detection by reducing the adverse effects of uncertain knowledge and enabling communication between the two subnetworks during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Uncertainty-Aware Distillation Loss and Gradient-Targeted Transfer Modulation Strategy. Table 5 shows the ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. It can be observed that the baseline does not perform well due to the adverse effect of uncertain knowledge and the asynchronous learning paces of the two subnetworks. On top of the baseline, including either the uncertainty-aware distillation loss or the gradient-targeted transfer modulation strategy improves the detection performance significantly, underscoring the importance of attaining high-certainty knowledge and synchronizing the learning paces of the two networks. In addition, combining the two designs achieves the best performance, highlighting their complementary nature and collaborative roles in knowledge acquisition and knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Inference speed comparison. Table 6 compares the inference speed on the KITTI validation set. Our method demonstrates superior efficiency thanks to our designed self-knowledge distillation framework, without utilizing complex network architectures during inference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.619, + 0.52 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.802 + ], + "angle": 0, + "content": "In this paper, we point out that previous weakly supervised monocular 3D detection methods either require additional LiDAR point clouds or paired images from multiple viewpoints or temporal sequences. To overcome these constraints, we propose a weakly supervised monocular 3D object detection approach that is exclusively grounded on single-view image inputs. Central to our approach is a self-knowledge distillation framework, which effectively harnesses the depth information within a single-view image with little computational overhead during inference. We further introduce an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy, facilitating knowledge acquisition and knowledge transfer, respectively. Finally, extensive experiments demonstrate the effectiveness of our method. Moving forward, we plan to further generalize our work to diverse and challenging scenarios, such as occlusions, varying lighting, and weather conditions, thereby enhancing its practical applicability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.66, + 0.833 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10515" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3d object detection in monocular video. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 135-152. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[4] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10379-10388, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2147-2156, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[6] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.472, + 0.599 + ], + "angle": 0, + "content": "[7] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. International Conference on Learning Representations, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[8] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In International Conference on Machine Learning, pages 2006-2015. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[9] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3872-3882, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[10] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1000-1001, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[11] Huan-ang Gao, Beiwen Tian, Pengfei Li, Hao Zhao, and Guyue Zhou. Dqs3d: Densely-matched quantization-aware semi-supervised 3d detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[13] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1440-1448, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[14] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In International Conference on Robotics and Automation, pages 13656-13662. IEEE, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. Advances in Neural Information Processing Systems, 35:33716-33727, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[19] Mingi Ji, Seungjae Shin, Seunghyun Hwang, Gibeom Park, and Il-Chul Moon. Refine myself by teaching myself: Feature refinement via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10664–10673, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in Neural Information Processing Systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11867-11876, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2980-2988, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[23] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7096-7104, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10516" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[26] Zhipeng Luo, Zhongang Cai, Changqing Zhou, Gongjie Zhang, Haiyu Zhao, Shuai Yi, Shijian Lu, Hongsheng Li, Shanghang Zhang, and Ziwei Liu. Unsupervised domain adaptive 3d detection with multi-level consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8866-8875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[27] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6851–6860, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[28] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[29] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 515-531. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[30] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[31] J Krishna Murthy, GV Sai Krishna, Falak Chhaya, and K Madhava Krishna. Reconstructing vehicles from a single image: Shape priors for road scene understanding. In International Conference on Robotics and Automation, pages 724-731. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[32] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[33] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 71-88. Springer, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[34] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. International Conference on Learning Representations, 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[35] Zengyi Qin, Jinglu Wang, and Yan Lu. Monognet: A geometric reasoning network for monocular 3d object localization. Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[36] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[37] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[38] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[39] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection: From single to multiclass recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1219-1231, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[40] Dawei Sun, Anbang Yao, Aojun Zhou, and Hao Zhao. Deeply-supervised knowledge synergy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[42] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular 3d object detection using multi-view projection and direction consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[43] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[45] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[46] Ting-Bing Xu and Cheng-Lin Liu. Data-distortion guided self-distillation for deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5565-5572, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[47] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[48] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2403-2412, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[50] Sukmin Yun, Jongjin Park, Kimin Lee, and Jinwoo Shin. Regularizing class-wise predictions via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10517" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 13876-13885, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[51] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[52] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3713-3722, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[53] Renrui Zhang, Han Qiu, Tai Wang, Xuanzhuo Xu, Ziyu Guo, Yu Qiao, Peng Gao, and Hongsheng Li. Monodetr: Depth-aware transformer for monocular 3d object detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[54] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[56] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[57] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Advances in Neural Information Processing Systems, 31, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10518" + } + ] +] \ No newline at end of file diff --git a/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_origin.pdf b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..272d704e788f90e6a8f38b94963ce64021b1c309 --- /dev/null +++ b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/e8def396-7aa9-4446-a151-70f8d14a292d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc508d9ba2b3a866f0e38513c1f313f48c4e9b6f1be8614a4201e8ef8fa3e79d +size 3036202 diff --git a/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/full.md b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..816cd7dabd64af7c7c96b69f698fb7d1ee6f3b55 --- /dev/null +++ b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/full.md @@ -0,0 +1,316 @@ +# Weakly Supervised Monocular 3D Detection with a Single-View Image + +Xueying Jiang $^{1}$ Sheng Jin $^{1}$ Lewei Lu $^{2}$ Xiaoqin Zhang $^{3}$ Shijian Lu $^{1*}$ $^{1}$ S-Lab, Nanyang Technological University + $^{2}$ Sensetime Research + $^{3}$ College of Computer Science and Technology, Zhejiang University of Technology + +# Abstract + +Monocular 3D detection (M3D) aims for precise 3D object localization from a single-view image which usually involves labor-intensive annotation of 3D detection boxes. Weakly supervised M3D has recently been studied to obviate the 3D annotation process by leveraging many existing 2D annotations, but it often requires extra training data such as LiDAR point clouds or multi-view images which greatly degrades its applicability and usability in various applications. We propose SKD-WM3D, a weakly supervised monocular 3D detection framework that exploits depth information to achieve M3D with a single-view image exclusively without any 3D annotations or other training data. One key design in SKD-WM3D is a self-knowledge distillation framework, which transforms image features into 3D-like representations by fusing depth information and effectively mitigates the inherent depth ambiguity in monocular scenarios with little computational overhead in inference. In addition, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Extensive experiments show that SKD-WM3D surpasses the state-of-the-art clearly and is even on par with many fully supervised methods. + +# 1. Introduction + +Monocular 3D detection (M3D) has emerged as one key component in the area of autonomous driving and computer vision. Its primary target is to recognize objects and obtain their 3D localization from single-view images. Thanks to its low deployment cost, M3D [5, 34] has attracted increasing attention in both academic and industrial sectors, achieving very impressive progress in recent years. On the other hand, most existing studies [21, 33, 37, 39] adopt a fully supervised setup which have been facing increasing scalability concern as large-scale 3D boxes are often labor-intensive + +to collect. Effective M3D training without 3D annotations has become a critical issue while handling M3D problems in various research and practical tasks. + +Weakly supervised M3D (WM3D) [34] has recently been explored for learning effective 3D detectors without 3D box annotations, aiming to exploit 2D annotations to make up for the absence of 3D information. For example, WeakM3D [34] exploits LiDAR point clouds to infer 3D information as illustrated in Figure 1(a). However, it requires costly and complicated LiDAR sensors to collect point clouds which limits its applicability and usability greatly. WeakMono3D [42] employs 2D information only by either leveraging multi-view stereo with images from multiple cameras or constructing pseudo-multi-view perspective from sequential video frames as illustrated in Figure 1(b). However, collecting multi-view images is complicated, and resorting to a pseudo multi-view perspective degrades the detection performance clearly. With the advance of single-view depth estimation, WM3D with depth from a single-view image presents a potential solution for compensating the absence of 3D annotations. On the other hand, direct integration of such depth into existing frameworks often necessitates complex network architectures which further incurs significant computational costs. This gives rise to a pertinent question: When not using additional LiDAR point clouds or multi-view image pairs, is it possible to harness the depth from off-the-shelf depth estimators without introducing much computational overhead in inference? + +We design SKD-WM3D, a novel weakly supervised monocular 3D object detection method that is exclusively grounded on single-view images. One key design in SKD-WM3D is a self-knowledge distillation framework which consists of a Depth-guided Self-teaching Network (DSN) and a Monocular 3D Detection Network (MDN). As illustrated in Figure 1(c), SKD-WM3D utilizes depth information obtained from an off-the-shelf depth estimator [17] to enhance the 3D localization ability of DSN and transfers such ability to MDN via self-knowledge distillation. Such self-distillation design enables MDN to unearth the intrinsic depth information from single-view images independently, + +![](images/23679a6c6316d3da32c47fa05c8d92c85f8467f36c213b18077975dcf47d17b4.jpg) +(a) LiDAR-based WM3D + +![](images/9ed648161116fb8ded3accfde6bf0be8ca6a3431bf6fd1b8dab2175cbd86a119.jpg) +(b) Multi-view WM3D + +![](images/e3fbf79a99a0aa1dd25ae988d0b22e581abfce7194ff3c1165be3fa1cfaac853.jpg) +(c) Single-view WM3D (Ours) +Figure 1. Different paradigms in weakly supervised monocular 3D detection. Our approach in (c) leverages Pseudo Depth Labels from a single-view image to achieve weakly supervised monocular 3D detection, requiring no extra training data like LiDAR point clouds or multi-view images as in (a) and (b). It improves usability and applicability greatly. The Pseudo Depth Labels are obtained with an off-the-shelf depth estimator [17] without extra training and ground-truth depth labels. Data in red denotes extra data in network training. + +bypassing additional modules such as pre-trained depth estimation networks and leading to precise and efficient 3D localization with little computational overhead during inference. On top of DSN and MDN, we design an uncertainty-aware distillation loss to optimize the utilization of the transferred 3D localization knowledge by weighting up more certain knowledge while weighting down less certain knowledge. In addition, we design a gradient-targeted transfer modulation strategy to synchronize the learning paces of DSN and MDN during the process of learning 3D localization knowledge, by prioritizing MDN learning at the initial stage when MDN lags behind DSN and enabling it to provide more feedback to DSN when MDN is better trained at late stages. + +Our contribution can be summarized in three aspects. First, we design a novel framework that achieves weakly supervised monocular 3D detection by distilling knowledge between a depth-guided self-teaching network and a monocular 3D detection network. Without any extra training data like LiDAR point clouds or multi-view images, the framework exploits depth exclusively from a single image with little computational overhead in inference. Second, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Third, the proposed approach clearly outperforms the state-of-the-art in weakly supervised monocular 3D detection, and its performance is even on par with several fully supervised methods. + +# 2. Related Work + +# 2.1. Monocular 3D Detection + +Monocular 3D object detection aims to predict 3D object localization from single-view images. Standard monocular detectors [1, 6, 14, 53, 56] operate solely on single images, + +without utilizing additional data. However, the inherent depth ambiguity of monocular detection significantly hinders its performance compared to its stereo counterparts. To address this limitation, various approaches seek solutions with the help of extra data, such as LiDAR point clouds [4, 7, 21, 27], video sequences [2], 3D CAD models [5, 24, 31], and depth estimation [10, 35, 45, 48]. Specifically, MonoRUn [4] adopts an uncertainty-aware regional reconstruction network for regressing pixel-associated 3D object coordinates with LiDAR point clouds as extra supervision. MonoDistill [7] introduces an effective distillation-based approach that incorporates spatial information from LiDAR signals into monocular 3D detection. Additionally, pseudo-LiDAR-based methods [45, 48] convert estimated depth maps to simulate the real LiDAR point clouds to utilize the well-designed LiDAR-based 3D detector. During inference, compared with methods using depth estimation, our method eliminates the need for pseudo depth labels and complex network architectures, with little computational overhead. Besides, existing fully supervised methods require large-scale 3D box ground truth, which is labor-intensive to collect and annotate. + +# 2.2. Weakly Supervised 3D Object Detection + +Due to the high cost of annotating 3D boxes in the 3D object detection task, various weakly supervised approaches [26, 29, 34, 36] have been proposed. For example, WS3D [29] presents a weakly supervised method for 3D LiDAR object detection, which requires only a limited number of weakly annotated scenes with center-annotated BEV maps. VS3D [36] introduces a cross-model knowledge distillation strategy to transfer the knowledge from the RGB domain to the point cloud domain, using LiDAR point clouds as weak supervision. Recent research on weakly supervised 3D object detection has turned to explore the monocular setting. For example, WeakM3D [34] generates 2D boxes to select + +![](images/141d4e32d52dadcc203433a5afe29fc63c2ec21bd73a2590abedcb929b8fbcfe.jpg) +Figure 2. The framework of the proposed self-knowledge distillation network. The framework consists of a depth-guided self-teaching network and a monocular 3D detection network. The depth-guided self-teaching network acquires comprehensive 3D localization knowledge by leveraging depth information and transfers its learned expertise to the monocular 3D detection network via soft label distillation to enhance its performance. We design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy to facilitate the knowledge transfer between the two networks effectively. During inference, the monocular 3D detection network extracts intrinsic depth information from single-view images independently with little computational overhead. + +RoI LiDAR point clouds as weak supervision and then predicts 3D boxes that closely align with the selected RoI LiDAR point clouds. More recently, WeakMono3D [42] eliminates the need for LiDAR, offering both multi-view and single-view yet multi-frame versions. While the former acquires stereo image inputs from multiple cameras, the latter constructs a pseudo-multi-view perspective using multiple video frames. The multi-frame version exhibits inferior 3D scene comprehension compared with the multi-view approach due to its smaller inter-frame disparity, leading to degraded performance. Instead of requiring extra training data like LiDAR point clouds or multi-view images, we tackle the challenge of weakly supervised monocular 3D detection by leveraging a single-view image exclusively. + +# 2.3. Self-Knowledge Distillation + +Knowledge distillation [8, 9, 11, 16, 18, 23, 32, 38, 43, 54] aims to transfer knowledge from a pre-trained teacher network to a student network for improving its performance. Self-knowledge distillation [30, 41, 47], distinct from traditional knowledge distillation, leverages the information within the student network to facilitate its learning without the pre-trained teacher network. Specifically, data augmentation approach [15, 46, 50] transfers knowledge through different distortions of the same training data. However, they are susceptible to inappropriate augmentations, such as improper instance rotation or distortion, potentially intro + +ducing noise that hampers network learning. Another typical approach exploits auxiliary networks [52, 57]. For example, DKS [40] introduces auxiliary supervision branches and pairwise knowledge alignments, while FRSKD [19] adds a new branch supervised by the original features and utilizes both soft-label and feature-map distillation. Our work is the first that introduces self-knowledge distillation with auxiliary networks for weakly supervised monocular 3D detection. It effectively exploits depth information from single-view images with little computational overhead during inference. + +# 3. Methodology + +This section presents the proposed SKD-WM3D. First, the problem definition and overview are presented in Sec. 3.1. Then detailed designs of SKD-WM3D are introduced, including the self-knowledge distillation framework in Sec. 3.2, the uncertainty-aware distillation loss in Sec. 3.3 and the gradient-targeted transfer modulation strategy in Sec. 3.4. Finally, loss functions are presented in Sec. 3.5. + +# 3.1. Problem Definition and Overview + +Weakly supervised monocular 3D detection takes an RGB image and the corresponding 2D bounding boxes as supervision, aiming to classify objects and determine their + +bounding boxes in 3D space without involving any 3D annotations in training. The prediction of each object is composed of the object category $C$ , a 2D bounding box $B^{2D}$ and a 3D bounding box $B^{3D}$ . Specifically, the 3D box $B^{3D}$ can be further decomposed to the object 3D location $(x_{3D}, y_{3D}, z_{3D})$ , the object dimension with height, width and length $(h_{3D}, w_{3D}, l_{3D})$ , as well as orientation $\theta$ . + +We design a self-knowledge distillation framework to tackle the challenge of weakly supervised monocular 3D detection from a single-view image. As Figure 2 shows, the framework consists of two subnetworks including a Depth-Guided Self-Teaching Network and a Monocular 3D Detection Network. In the Depth-Guided Self-Teaching Network, the global features $F_{G}$ extracted by the backbone are fed into a Depth Head to obtain depth features. Next, the global features $F_{G}$ and the extracted depth features are fed into a Fusion Layer to obtain 3D-like features $F_{3D}$ . The 3D-like features of each object are then obtained via RoIAlign, and further fed to a Depth-Aware 3D Head to predict 3D box $\widehat{B}_p^{3D}$ and uncertainty $\widehat{U}$ . In the Monocular 3D Detection Network, We first use RoIAlign to generate object-level features from the global features $F_{G}$ , and then feed them to a 2D-to-3D Head to predict 3D box $\widehat{B}_p^{2D}$ and uncertainty $U$ . The 3D boxes predicted by both networks are further projected into 2D boxes. Moreover, we design an uncertainty-aware distillation loss $L_{ud}$ to obtain low-uncertainty knowledge, and a gradient-targeted transfer modulation strategy to synchronize the learning paces between the two networks by controlling gradients $\widehat{G}$ and $G$ of $L_{ud}$ . + +# 3.2. Self-Knowledge Distillation Framework + +The self-knowledge distillation framework enhances the 3D localization ability of the depth-guided self-teaching network by utilizing depth information from an off-the-shelf depth estimator and then transfers the ability to the monocular 3D detection network via self-knowledge distillation. + +Depth-Guided Self-Teaching Network. To equip the self-teaching network with 3D localization ability, we propose to learn from global features $F_{G}$ and depth information from an off-the-shelf depth estimator to acquire comprehensive 3D knowledge. The depth information is exploited via two major designs. Firstly, we introduce a depth head $\mathcal{D}$ that extracts depth features $F_{D}$ as follows: + +$$ +F _ {D} = \mathcal {D} \left(F _ {G}\right), \tag {1} +$$ + +The depth features $F_{D}$ are exploited to generate depth maps $D_{p}$ , where the depth map generation is supervised by the pseudo ground truth of the depth map $D_{gt}$ that is predicted by an off-the-shelf depth estimator by using the focal loss [22] as depth loss $L_{dep}$ . Hence, the depth features can be acquired by the depth-guided self-teaching network effectively. + +Remark 1. We generate depth pseudo labels using an off-the-shelf depth estimator [17] with frozen weights, eliminating the need for additional training and ground-truth depth labels. Adopting an off-the-shelf depth estimator incurs negligible costs as compared with prior studies that require either point clouds [34] or multi-view images [42]. + +Secondly, we obtain 3D-like features $F_{G3D}$ by integrating the depth features $F_{D}$ that provide information along the depth dimension, as well as the global features $F_{G}$ that capture knowledge about the 2D image plane. Specifically, we design a fusion layer that fuses the depth features $F_{D}$ with the global features $F_{G}$ to derive the $F_{G3D}$ as follows: + +$$ +F _ {G 3 D} = F F N \left(C A \left(S A \left(F _ {D}\right), F _ {G}\right)\right), \tag {2} +$$ + +where the $FFN$ is the feed-forward network, and $CA$ , $SA$ denote CrossAttention, SelfAttention, respectively. The structures of CrossAttention and SelfAttention employ the standard transformer architecture [44]. The obtained 3D comprehension improves the network's ability to precisely locate objects, effectively mitigating depth ambiguity arising from single-view image input. + +Monocular 3D Detection Network. The monocular 3D detection network acquires the 3D localization knowledge from the depth-guided self-teaching network. By distilling soft labels generated by the depth-guided self-teaching network, the monocular 3D detection network can extract intrinsic depth information from images independently during inference. This kills the need for additional complex modules such as pre-trained depth estimation networks or depth fusion modules, facilitating the inference with little computational overhead. + +# 3.3. Uncertainty-Aware Distillation Loss + +During the knowledge distillation process, uncertain knowledge could affect the network training negatively if all transferred knowledge is treated equally. To benefit more from certain knowledge and weaken the effect of uncertain knowledge, we design an uncertainty-aware distillation loss between the 3D boxes that are predicted by the two networks in the self-knowledge distillation framework. The uncertainty-aware distillation loss exploits the prediction uncertainty to modulate the distillation loss magnitude as follows: + +$$ +L _ {u d} = \frac {L _ {d}}{\operatorname* {m i n} ((\widehat {U} + U) / 2 , \alpha)} + \left\| \operatorname * {m i n} (\frac {\widehat {U} + U}{2}, \alpha) \right\| ^ {2}, \tag {3} +$$ + +where $\widehat{U}$ and $U$ denote the uncertainty of the 3D boxes that are predicted by the two networks, respectively. Here + +we assume the 3D box predictions have the Laplace distribution, and we adopt the standard deviations as the uncertainties, inspired by [20, 25, 33]. $\left\| \min (\frac{\widehat{U} + U}{2},\alpha)\right\| ^2$ is the L2 regularization, and $\alpha$ is a fixed value set to 0.1. $L_{d}$ denotes the basic distillations loss, and we employ the commonly used SmoothL1 [13] loss to enforce the consistency between the 3D boxes predicted by the two networks. The SmoothL1 loss leaves a soft margin when computing the difference between the two 3D boxes: + +$$ +L _ {d} = \left\{ \begin{array}{l l} 0. 5 \times \left(\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D}\right) ^ {2}, & \text {i f} \left| \widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \right| < 1. 0 \\ \left| \widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \right| - 0. 5, & \text {o t h e r w i s e} \end{array} , \right. \tag {4} +$$ + +where $\widehat{B}_p^{3D}$ and $B_p^{3D}$ are the predicted 3D boxes from the depth-guided self-teaching network and the monocular 3D detection network, respectively. + +Remark 2. The proposed uncertainty-aware distillation loss $L_{ud}$ integrates average uncertainty $\frac{\widehat{U} + U}{2}$ as regularization and a weighted component for the basic distillation loss $L_{d}$ , allowing adaptive learning adjustments based on the knowledge's uncertainty level. Specifically, when dealing with uncertain knowledge, a smaller weight is assigned to the basic distillation loss $L_{d}$ to mitigate potential adverse effects on network learning. Consequently, the network prioritizes optimizing uncertainty reduction in such scenarios. When dealing with certain knowledge, the network emphasizes optimizing the basic distillation loss $L_{d}$ due to its higher weight. Notably, the basic distillation loss $L_{d}$ simply considers box consistency, while integrating uncertainty is beneficial for enhancing the knowledge distillation process. + +# 3.4. Transfer Modulation Strategy + +The depth-guided self-teaching network, which leverages depth information to predict 3D boxes, transfers its learned 3D knowledge to the monocular 3D detection network. The asynchronous learning paces of the two networks pose potential challenges to effective 3D knowledge transfer. + +We design a gradient-targeted transfer modulation strategy to synchronize the learning pace of the depth-guided self-teaching network and the monocular 3D detection network. We modulate the knowledge transfer dynamically, by controlling the gradients from the uncertainty-aware distillation loss $L_{ud}$ . Specifically, we adapt the gradients based on the 2D projection performance of each network, assigning smaller backward gradients for the good-performing network and higher backward gradients for the bad-performing network. The gradient-targeted transfer modulation strategy is formulated as follows: + +$$ +\widehat {G} ^ {\prime} = \frac {2 \times \widehat {L} _ {p r o j}}{\widehat {L} _ {p r o j} + L _ {p r o j}} \times \widehat {G}, G ^ {\prime} = \frac {2 \times L _ {p r o j}}{\widehat {L} _ {p r o j} + L _ {p r o j}} \times G, \tag {5} +$$ + +Where $\widehat{G}$ and $G$ are the original gradients of the two networks, $\widehat{G}^{\prime}$ and $G^{\prime}$ are the modified gradients, $\widehat{L}_{proj}$ and $L_{proj}$ are projection losses, computed between the projected 2D boxes from 3D box predictions and 2D box annotations. + +The gradient-targeted transfer modulation strategy prioritizes training the monocular 3D detection network when its learning lags behind the depth-guided self-teaching network at the early training stage. As the monocular 3D detection network learns and improves gradually, it is enabled to provide more feedback progressively to the depth-guided self-teaching network. + +# 3.5. Loss Functions + +The overall objective consists of three losses including $L_{ud}$ , $L_{dep}$ and $L_{base}$ . $L_{ud}$ is the uncertainty-aware distillation loss as defined in Sec. 3.3. $L_{dep}$ is the depth loss for supervising the predicted depth map. $L_{base}$ includes losses for supervising 2D boxes prediction by 2D heads and the 3D box predictions, which has been adopted in prior CenterNet [55] and WeakMono3D [42]. We set the weight for each loss item to 1.0, and the overall loss function can be formulated as follows: + +$$ +L = L _ {u d} + L _ {d e p} + L _ {b a s e}. \tag {6} +$$ + +# 4. Experiments + +# 4.1. Datasets + +We conduct experiments over the KITTI 3D dataset [12] and the nuScenes dataset [3] that have been widely adopted for benchmarking of 3D object detection methods. The KITTI 3D dataset consists of 7,481 images for training and 7,518 images for testing. The labels of the train set are publicly available and the labels of the test set are stored on a test server for evaluation. For ablation studies, we follow [5] which divides the 7,481 training samples into a new train set with 3,712 images and a validation set with 3,769 images. The nuScenes dataset comprises 1,000 video scenes, including RGB images captured by 6 surround-view cameras. The dataset is split into a training set (700 scenes), a validation set (150 scenes), and a test set (150 scenes). Following [34], the performance on the validation set is reported. + +# 4.2. Evaluation Protocols + +For the KITTI 3D dataset, following [39], we adopt the evaluation metric $\mathrm{AP}|_{R_{40}}$ which is the average of the AP of 40 recall points. We report the average precision on bird's eye view and 3D object detection as $\mathrm{AP}_{BEV}|_{R_{40}}$ and $\mathrm{AP}_{3D}|_{R_{40}}$ . In addition, as most weakly supervised 3D object detection methods apply IoU threshold of 0.7 for the test set and 0.5 for the validation set, we adopt the same thresholds for fair benchmarking. We adopt four metrics + +
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.7)|R40
EasyModerateHard
WeakM3D [34]ResNet-50Weak11.82/5.035.66/2.264.08/1.63
WeakMono3D [42]DLA-3412.31/6.988.80/4.857.81/4.45
SKD-WM3D (Ours)DLA-3415.71/8.9510.15/5.548.08/4.53
+ +Table 1. Comparison on the performance of the Car category on KITTI test set. For all results, we use $\mathrm{AP}|_{R_{40}}$ metrics with IoU threshold equal to 0.7. The best results are in bold. + +
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
CenterNet [55]DLA-3434.36/20.0027.91/17.5024.65/15.57
MonoGRNet [35]VGG-1652.13/47.5935.99/32.2828.72/25.50
M3D-RPN [1]DenseNet-12153.35/48.5339.60/35.9431.76/28.59
MonoPair [6]DLA-34Full61.06/55.3847.63/42.3941.92/37.99
MonoDLE [28]DLA-3460.73/55.4146.87/43.4241.89/37.81
GUPNet [25]DLA-3461.78/57.6247.06/42.3340.88/37.59
Kinematic [2]DenseNet-12161.79/55.4444.68/39.4734.56/31.26
MonoDistill [7]DLA-3471.45/65.6953.11/49.3546.94/43.49
MonoDETR [53]*ResNet-5072.34/68.0551.97/48.4246.94/43.48
VS3D [36]VGG-16Weak31.59/22.6220.59/14.4316.28/10.91
Autolabels [51]ResNeXt10150.51/38.3130.97/19.9023.72/14.83
WeakM3D [34]ResNet-5058.20/50.1638.02/29.9430.17/23.11
WeakMono3D [42]DLA-3454.32/49.3742.83/39.0140.07/36.34
SKD-WM3D (Ours)DLA-3455.47/50.2144.35/41.5741.86/36.92
+ +Table 2. Comparison on the performance of the Car category on KITTI val set. For all results, we use $\mathrm{AP}|_{R_{40}}$ metric with IoU threshold equal to 0.5. * denotes this performance is reproduced from the official code. The best results of weakly supervised 3D object detection approaches are in bold. + +
MethodAP↑ATE↓ASE↓AAE↓
WeakM3D [34]0.2140.8140.2340.682
SKD-WM3D (Ours)0.2420.7950.2310.659
+ +Table 3. Comparison on the performance of the Car category on nuScenes val set. The best results are in **bold**. + +for the evaluation on the nuScenes dataset, namely, AP (Average Precision), ATE (Average Translation Error), ASE (Average Scale Error), and AAE (Average Attribute Error). Following [34], AVE (Average Velocity Error) and AOE (Average Orientation Error) are not reported due to the lack of supervision for velocity and movement direction in the weakly supervised approach. + +# 4.3. Implementation Details + +We conduct experiments on 2 NVIDIA V100 GPUs with batch size of 16, and train the framework with 150 epochs. We use the Adam optimizer with the initial learning rate $1e^{-5}$ , which is gradually increased to $1e^{-3}$ for the first 5 + +epochs and decayed with rate 0.1 at the 90 and 120 epochs. We employ DLA-34 [49] as the detector's backbone. The pseudo ground truth of the depth map is generated with an off-the-shelf depth estimator [17] without using the ground truth of depth label. + +# 4.4. Comparison with State-of-the-Art Methods + +We compare our method with several state-of-the-art weakly supervised monocular 3D detection methods on the KITTI test set. As Table 1 shows, our method achieves superior detection performance across all metrics. This superior performance is largely attributed to our designed self-knowledge distillation framework that extracts and exploits intrinsic depth information from a single-view image effectively. It should be highlighted that our method employs a single-view image exclusively without involving additional training data such as LiDAR point clouds [34] or multi-view image pairs [42]. + +Table 2 shows the benchmarking on the KITTI validation set. Specifically, we compare our method against both state-of-the-art weakly supervised monocular 3D de + +![](images/057c72841b314a35e0ac0a50c0a326cbf5dc7208239f4aa06276bd42bbd1f46a.jpg) + +![](images/0600ff562e8ac152b44fa46026027a87d9e8aab8f50c6450e65fafad4e6accc4.jpg) + +![](images/08cfb4d05d59cecb5ea2b04bbfa5b9cb8de5a9a444455febc5696e50c09c708b.jpg) + +![](images/d65e0f934483acce78f0b92473e426bec7c015c5c74af43a06d15a82baaadf27.jpg) + +![](images/60304c8f5a19a03ad701811eb6601a39cfff78425974ad0bdc67d00ce53dad3f.jpg) + +![](images/9e8ce927cd9f6f264a05af9ee17148a05d81c2859c7600b5711df746721c292b.jpg) + +![](images/df3e0f5292b7c4c7430b564b130d2f26fa936f378e06f1dd171a092f90d7119a.jpg) + +![](images/876f9e572f579b8bf2508db789aed052dfcd62f342d6eefe691d6abcee33100b.jpg) + +![](images/4b9f731d9955b215ed737a365cb6480da93be966318806fd5790d32ff911a7db.jpg) + +![](images/314e618824552a66c1f8f7393644dd56a82856be675d3e1ecbf390dcb615c8c6.jpg) +Figure 3. Qualitative illustration on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The ground truth of LiDAR point clouds is utilized for visualization purposes only. Best viewed with zoom-in. + +![](images/aece145fc69eb5683cfab827127466604d765005f0773e5253e5be49c1cedfd5.jpg) + +![](images/6dba897aa40fcd90d30c7227385a89e2390837d76e9ab9f31fb19def9bf007c0.jpg) + +![](images/058031bb8fea94bea61f020151aad422aca6d5bc02058c9294da65d9587a4af5.jpg) +Uncertainty: 0.24 + +![](images/098fbf338e7868bddc59d92105513f091bbe932038be96b13f210d2d39624f22.jpg) +Uncertainty: 0.41 + +![](images/75163d5cb3fa6b78eedcb513a2d195e35852eeb84dd28f0d7e163fdb73878a88.jpg) +Uncertainty: 0.67 + +![](images/1614e1879ea16e5919c26b747e9a341e92830d223a75c5835cad2ac7c9bbd06a.jpg) +Uncertainty: 0.75 +Figure 4. Qualitative illustration of object detection and the corresponding detection uncertainties on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The detection accuracy is closely correlated with the detection uncertainty. Best viewed with zoom-in. + +![](images/25642c57775beb7f4493c77949fcb90743213f073fddaf535d8a530f15703460.jpg) +Uncertainty: 0.82 + +tection methods and fully supervised methods. It can be seen that our method achieves superior performance compared to WeakM3D [34] and WeakMono3D [42] across most metrics, even without using LiDAR point clouds or multi-view images. Notably, our method significantly outperforms WeakM3D [34] in the Moderate and Hard categories, primarily due to the sparse nature of distant LiDAR point clouds adversely affecting its performance. Additionally, its performance is even on par with several fully supervised methods [1, 35, 55]. + +Table 3 shows the results on the nuScenes validation set. It can be seen that our proposed method outperforms WeakM3D [34] across all four evaluation metrics, validating the effectiveness of our approach. + +Qualitative Results Figure 3 shows qualitative results with both 2D RGB images and 3D point clouds. In simple scenarios, our model achieves great prediction accuracy, which is largely attributed to the proposed self-knowledge distillation framework as well as the uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy, all working together to facilitate comprehensive 3D information extraction effectively. However, for heavily occluded or distant objects, the accuracy of orientation and depth estimation tends to drop, which is common for monocular 3D detection due to its ill-posed nature. In addition, Figure 4 shows the visualization of predicted bounding boxes and their corresponding uncertainties. It can be observed that the prediction accuracy of bounding boxes has a close correlation with the prediction uncertainty. + +
IndexMDNDSNAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
1✓0.00/0.000.00/0.000.00/0.00
2✓45.23/40.9634.27/31.0230.17/26.27
3✓✓55.47/50.2144.35/41.5741.86/36.92
+ +Table 4. Ablation study of the proposed self-knowledge distillation framework. The best results are in bold. MDN denotes the Monocular 3D Detection Network, while DSN denotes the Depth-Guided Self-Teaching Network. + +
IndexLudTMSAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
149.95/44.6138.24/35.7437.28/34.82
2✓53.16/48.1341.85/39.0240.14/35.70
3✓52.35/46.3041.45/38.9139.73/35.44
4✓✓55.47/50.2144.35/41.5741.86/36.92
+ +Table 5. Ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. The best results are in **bold**. $L_{ud}$ denotes the Uncertainty-Aware Distillation Loss, while TMS denotes the gradient-targeted transfer modulation strategy. + +
MethodWeakM3D [34]MonoDistill [7]*SKD-WM3D (Ours)*
FPS13.925.030.3
+ +Table 6. Comparison on inference speed of M3D methods. * denotes this method utilizes dense depth maps. + +# 4.5. Ablation Study + +We conduct extensive ablation studies on the KITTI validation dataset to evaluate our designs. Specifically, we evaluated the efficacy of the two individual networks in the proposed self-knowledge distillation framework. Additionally, we examine the effect of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. Lastly, we evaluated the efficiency of our monocular 3D detection framework. + +Self-Knowledge Distillation Framework. We train two models to assess the contributions of the two networks in our proposed self-knowledge distillation framework. As Table 4 shows, training the monocular 3D detection network alone produces few meaningful detection results as the absence of depth information leads to ambiguous object localization along the depth dimension. As a comparison, training the depth-guided self-teaching network alone can produce reasonable detection results thanks to the depth map pseudo labels. In addition, training both subnetworks concurrently produces the best 3D detection, validating the effectiveness of extracting 3D information from a single image. We can also see that including the self-knowledge distillation on top of the depth-guided self-teaching network + +greatly improves the detection by reducing the adverse effects of uncertain knowledge and enabling communication between the two subnetworks during training. + +Uncertainty-Aware Distillation Loss and Gradient-Targeted Transfer Modulation Strategy. Table 5 shows the ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. It can be observed that the baseline does not perform well due to the adverse effect of uncertain knowledge and the asynchronous learning paces of the two subnetworks. On top of the baseline, including either the uncertainty-aware distillation loss or the gradient-targeted transfer modulation strategy improves the detection performance significantly, underscoring the importance of attaining high-certainty knowledge and synchronizing the learning paces of the two networks. In addition, combining the two designs achieves the best performance, highlighting their complementary nature and collaborative roles in knowledge acquisition and knowledge transfer. + +Inference speed comparison. Table 6 compares the inference speed on the KITTI validation set. Our method demonstrates superior efficiency thanks to our designed self-knowledge distillation framework, without utilizing complex network architectures during inference. + +# 5. Conclusion + +In this paper, we point out that previous weakly supervised monocular 3D detection methods either require additional LiDAR point clouds or paired images from multiple viewpoints or temporal sequences. To overcome these constraints, we propose a weakly supervised monocular 3D object detection approach that is exclusively grounded on single-view image inputs. Central to our approach is a self-knowledge distillation framework, which effectively harnesses the depth information within a single-view image with little computational overhead during inference. We further introduce an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy, facilitating knowledge acquisition and knowledge transfer, respectively. Finally, extensive experiments demonstrate the effectiveness of our method. Moving forward, we plan to further generalize our work to diverse and challenging scenarios, such as occlusions, varying lighting, and weather conditions, thereby enhancing its practical applicability. + +# Acknowledgement + +This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2, 6, 7 +[2] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3d object detection in monocular video. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 135-152. Springer, 2020. 2, 6 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5 +[4] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10379-10388, 2021. 2 +[5] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2147-2156, 2016. 1, 2, 5 +[6] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 2, 6 +[7] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. International Conference on Learning Representations, 2022. 2, 6, 8 +[8] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In International Conference on Machine Learning, pages 2006-2015. PMLR, 2020. 3 +[9] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3872-3882, 2023. 3 +[10] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1000-1001, 2020. 2 +[11] Huan-ang Gao, Beiwen Tian, Pengfei Li, Hao Zhao, and Guyue Zhou. Dqs3d: Densely-matched quantization-aware semi-supervised 3d detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 5 +[13] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1440-1448, 2015. 5 +[14] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2 +[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3 +[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[17] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In International Conference on Robotics and Automation, pages 13656-13662. IEEE, 2021. 1, 2, 4, 6 +[18] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. Advances in Neural Information Processing Systems, 35:33716-33727, 2022. 3 +[19] Mingi Ji, Seungjae Shin, Seunghyun Hwang, Gibeom Park, and Il-Chul Moon. Refine myself by teaching myself: Feature refinement via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10664–10673, 2021. 3 +[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in Neural Information Processing Systems, 30, 2017. 5 +[21] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11867-11876, 2019. 1, 2 +[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2980-2988, 2017. 4 +[23] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7096-7104, 2019. 3 +[24] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2 +[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 5, 6 + +[26] Zhipeng Luo, Zhongang Cai, Changqing Zhou, Gongjie Zhang, Haiyu Zhao, Shuai Yi, Shijian Lu, Hongsheng Li, Shanghang Zhang, and Ziwei Liu. Unsupervised domain adaptive 3d detection with multi-level consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8866-8875, 2021. 2 +[27] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6851–6860, 2019. 2 +[28] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 6 +[29] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 515-531. Springer, 2020. 2 +[30] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in Neural Information Processing Systems, 32, 2019. 3 +[31] J Krishna Murthy, GV Sai Krishna, Falak Chhaya, and K Madhava Krishna. Reconstructing vehicles from a single image: Shape priors for road scene understanding. In International Conference on Robotics and Automation, pages 724-731. IEEE, 2017. 2 +[32] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3 +[33] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 71-88. Springer, 2022. 1, 5 +[34] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. International Conference on Learning Representations, 2022. 1, 2, 4, 5, 6, 7, 8 +[35] Zengyi Qin, Jinglu Wang, and Yan Lu. Monognet: A geometric reasoning network for monocular 3d object localization. Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 2, 6, 7 +[36] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 6 +[37] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 1 +[38] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: + +Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3 +[39] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection: From single to multiclass recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1219-1231, 2020. 1, 5 +[40] Dawei Sun, Anbang Yao, Aojun Zhou, and Hao Zhao. Deeply-supervised knowledge synergy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2019. 3 +[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 3 +[42] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular 3d object detection using multi-view projection and direction consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7 +[43] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 3 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4 +[45] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2 +[46] Ting-Bing Xu and Cheng-Lin Liu. Data-distortion guided self-distillation for deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5565-5572, 2019. 3 +[47] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[48] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. International Conference on Learning Representations, 2020. 2 +[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2403-2412, 2018. 6 +[50] Sukmin Yun, Jongjin Park, Kimin Lee, and Jinwoo Shin. Regularizing class-wise predictions via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 13876-13885, 2020. 3 +[51] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 6 +[52] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3713-3722, 2019. 3 +[53] Renrui Zhang, Han Qiu, Tai Wang, Xuanzhuo Xu, Ziyu Guo, Yu Qiao, Peng Gao, and Hongsheng Li. Monodetr: Depth-aware transformer for monocular 3d object detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 6 +[54] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 3 +[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5, 6, 7 +[56] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[57] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Advances in Neural Information Processing Systems, 31, 2018. 3 \ No newline at end of file diff --git a/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/images.zip b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4a395e54475fd4c0d410359aac5fd56257a71aaf --- /dev/null +++ b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c22c48ee96a95a55278ddcce2b4356248914464a7f2b93e1db87d8e1f411190d +size 596749 diff --git a/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/layout.json b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe5ce7f460643f678ffb56f91f93e59dc27861b1 --- /dev/null +++ b/2024/Weakly Supervised Monocular 3D Detection with a Single-View Image/layout.json @@ -0,0 +1,8630 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 515, + 121 + ], + "type": "text", + "content": "Weakly Supervised Monocular 3D Detection with a Single-View Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": "Xueying Jiang" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": " Sheng Jin" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": " Lewei Lu" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": " Xiaoqin Zhang" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": " Shijian Lu" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University \n" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": "Sensetime Research \n" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 142, + 497, + 186 + ], + "type": "text", + "content": "College of Computer Science and Technology, Zhejiang University of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "type": "text", + "content": "Monocular 3D detection (M3D) aims for precise 3D object localization from a single-view image which usually involves labor-intensive annotation of 3D detection boxes. Weakly supervised M3D has recently been studied to obviate the 3D annotation process by leveraging many existing 2D annotations, but it often requires extra training data such as LiDAR point clouds or multi-view images which greatly degrades its applicability and usability in various applications. We propose SKD-WM3D, a weakly supervised monocular 3D detection framework that exploits depth information to achieve M3D with a single-view image exclusively without any 3D annotations or other training data. One key design in SKD-WM3D is a self-knowledge distillation framework, which transforms image features into 3D-like representations by fusing depth information and effectively mitigates the inherent depth ambiguity in monocular scenarios with little computational overhead in inference. In addition, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Extensive experiments show that SKD-WM3D surpasses the state-of-the-art clearly and is even on par with many fully supervised methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "type": "text", + "content": "Monocular 3D detection (M3D) has emerged as one key component in the area of autonomous driving and computer vision. Its primary target is to recognize objects and obtain their 3D localization from single-view images. Thanks to its low deployment cost, M3D [5, 34] has attracted increasing attention in both academic and industrial sectors, achieving very impressive progress in recent years. On the other hand, most existing studies [21, 33, 37, 39] adopt a fully supervised setup which have been facing increasing scalability concern as large-scale 3D boxes are often labor-intensive" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "content": "to collect. Effective M3D training without 3D annotations has become a critical issue while handling M3D problems in various research and practical tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 267, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 546, + 567 + ], + "type": "text", + "content": "Weakly supervised M3D (WM3D) [34] has recently been explored for learning effective 3D detectors without 3D box annotations, aiming to exploit 2D annotations to make up for the absence of 3D information. For example, WeakM3D [34] exploits LiDAR point clouds to infer 3D information as illustrated in Figure 1(a). However, it requires costly and complicated LiDAR sensors to collect point clouds which limits its applicability and usability greatly. WeakMono3D [42] employs 2D information only by either leveraging multi-view stereo with images from multiple cameras or constructing pseudo-multi-view perspective from sequential video frames as illustrated in Figure 1(b). However, collecting multi-view images is complicated, and resorting to a pseudo multi-view perspective degrades the detection performance clearly. With the advance of single-view depth estimation, WM3D with depth from a single-view image presents a potential solution for compensating the absence of 3D annotations. On the other hand, direct integration of such depth into existing frameworks often necessitates complex network architectures which further incurs significant computational costs. This gives rise to a pertinent question: When not using additional LiDAR point clouds or multi-view image pairs, is it possible to harness the depth from off-the-shelf depth estimators without introducing much computational overhead in inference?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "We design SKD-WM3D, a novel weakly supervised monocular 3D object detection method that is exclusively grounded on single-view images. One key design in SKD-WM3D is a self-knowledge distillation framework which consists of a Depth-guided Self-teaching Network (DSN) and a Monocular 3D Detection Network (MDN). As illustrated in Figure 1(c), SKD-WM3D utilizes depth information obtained from an off-the-shelf depth estimator [17] to enhance the 3D localization ability of DSN and transfers such ability to MDN via self-knowledge distillation. Such self-distillation design enables MDN to unearth the intrinsic depth information from single-view images independently," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10508" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 148, + 185 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 148, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 148, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 148, + 185 + ], + "type": "image", + "image_path": "23679a6c6316d3da32c47fa05c8d92c85f8467f36c213b18077975dcf47d17b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 142, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 142, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 142, + 198 + ], + "type": "text", + "content": "(a) LiDAR-based WM3D" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 69, + 339, + 185 + ], + "blocks": [ + { + "bbox": [ + 164, + 69, + 339, + 185 + ], + "lines": [ + { + "bbox": [ + 164, + 69, + 339, + 185 + ], + "spans": [ + { + "bbox": [ + 164, + 69, + 339, + 185 + ], + "type": "image", + "image_path": "9ed648161116fb8ded3accfde6bf0be8ca6a3431bf6fd1b8dab2175cbd86a119.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 188, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 208, + 188, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 208, + 188, + 287, + 198 + ], + "type": "text", + "content": "(b) Multi-view WM3D" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 364, + 70, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 364, + 70, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 364, + 70, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 364, + 70, + 545, + 185 + ], + "type": "image", + "image_path": "e3fbf79a99a0aa1dd25ae988d0b22e581abfce7194ff3c1165be3fa1cfaac853.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 188, + 507, + 198 + ], + "lines": [ + { + "bbox": [ + 403, + 188, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 403, + 188, + 507, + 198 + ], + "type": "text", + "content": "(c) Single-view WM3D (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 213, + 546, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 546, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 546, + 257 + ], + "type": "text", + "content": "Figure 1. Different paradigms in weakly supervised monocular 3D detection. Our approach in (c) leverages Pseudo Depth Labels from a single-view image to achieve weakly supervised monocular 3D detection, requiring no extra training data like LiDAR point clouds or multi-view images as in (a) and (b). It improves usability and applicability greatly. The Pseudo Depth Labels are obtained with an off-the-shelf depth estimator [17] without extra training and ground-truth depth labels. Data in red denotes extra data in network training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 277, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 445 + ], + "type": "text", + "content": "bypassing additional modules such as pre-trained depth estimation networks and leading to precise and efficient 3D localization with little computational overhead during inference. On top of DSN and MDN, we design an uncertainty-aware distillation loss to optimize the utilization of the transferred 3D localization knowledge by weighting up more certain knowledge while weighting down less certain knowledge. In addition, we design a gradient-targeted transfer modulation strategy to synchronize the learning paces of DSN and MDN during the process of learning 3D localization knowledge, by prioritizing MDN learning at the initial stage when MDN lags behind DSN and enabling it to provide more feedback to DSN when MDN is better trained at late stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "type": "text", + "content": "Our contribution can be summarized in three aspects. First, we design a novel framework that achieves weakly supervised monocular 3D detection by distilling knowledge between a depth-guided self-teaching network and a monocular 3D detection network. Without any extra training data like LiDAR point clouds or multi-view images, the framework exploits depth exclusively from a single image with little computational overhead in inference. Second, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Third, the proposed approach clearly outperforms the state-of-the-art in weakly supervised monocular 3D detection, and its performance is even on par with several fully supervised methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 638, + 134, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 134, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 134, + 650 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 184, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 184, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 184, + 670 + ], + "type": "text", + "content": "2.1. Monocular 3D Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Monocular 3D object detection aims to predict 3D object localization from single-view images. Standard monocular detectors [1, 6, 14, 53, 56] operate solely on single images," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 540 + ], + "type": "text", + "content": "without utilizing additional data. However, the inherent depth ambiguity of monocular detection significantly hinders its performance compared to its stereo counterparts. To address this limitation, various approaches seek solutions with the help of extra data, such as LiDAR point clouds [4, 7, 21, 27], video sequences [2], 3D CAD models [5, 24, 31], and depth estimation [10, 35, 45, 48]. Specifically, MonoRUn [4] adopts an uncertainty-aware regional reconstruction network for regressing pixel-associated 3D object coordinates with LiDAR point clouds as extra supervision. MonoDistill [7] introduces an effective distillation-based approach that incorporates spatial information from LiDAR signals into monocular 3D detection. Additionally, pseudo-LiDAR-based methods [45, 48] convert estimated depth maps to simulate the real LiDAR point clouds to utilize the well-designed LiDAR-based 3D detector. During inference, compared with methods using depth estimation, our method eliminates the need for pseudo depth labels and complex network architectures, with little computational overhead. Besides, existing fully supervised methods require large-scale 3D box ground truth, which is labor-intensive to collect and annotate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 551, + 515, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 515, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 515, + 563 + ], + "type": "text", + "content": "2.2. Weakly Supervised 3D Object Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Due to the high cost of annotating 3D boxes in the 3D object detection task, various weakly supervised approaches [26, 29, 34, 36] have been proposed. For example, WS3D [29] presents a weakly supervised method for 3D LiDAR object detection, which requires only a limited number of weakly annotated scenes with center-annotated BEV maps. VS3D [36] introduces a cross-model knowledge distillation strategy to transfer the knowledge from the RGB domain to the point cloud domain, using LiDAR point clouds as weak supervision. Recent research on weakly supervised 3D object detection has turned to explore the monocular setting. For example, WeakM3D [34] generates 2D boxes to select" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10509" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 548, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 548, + 291 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 548, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 548, + 291 + ], + "type": "image", + "image_path": "141d4e32d52dadcc203433a5afe29fc63c2ec21bd73a2590abedcb929b8fbcfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 300, + 546, + 368 + ], + "lines": [ + { + "bbox": [ + 45, + 300, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 45, + 300, + 546, + 368 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed self-knowledge distillation network. The framework consists of a depth-guided self-teaching network and a monocular 3D detection network. The depth-guided self-teaching network acquires comprehensive 3D localization knowledge by leveraging depth information and transfers its learned expertise to the monocular 3D detection network via soft label distillation to enhance its performance. We design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy to facilitate the knowledge transfer between the two networks effectively. During inference, the monocular 3D detection network extracts intrinsic depth information from single-view images independently with little computational overhead." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 388, + 289, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 289, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 289, + 556 + ], + "type": "text", + "content": "RoI LiDAR point clouds as weak supervision and then predicts 3D boxes that closely align with the selected RoI LiDAR point clouds. More recently, WeakMono3D [42] eliminates the need for LiDAR, offering both multi-view and single-view yet multi-frame versions. While the former acquires stereo image inputs from multiple cameras, the latter constructs a pseudo-multi-view perspective using multiple video frames. The multi-frame version exhibits inferior 3D scene comprehension compared with the multi-view approach due to its smaller inter-frame disparity, leading to degraded performance. Instead of requiring extra training data like LiDAR point clouds or multi-view images, we tackle the challenge of weakly supervised monocular 3D detection by leveraging a single-view image exclusively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 563, + 198, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 198, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 198, + 575 + ], + "type": "text", + "content": "2.3. Self-Knowledge Distillation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Knowledge distillation [8, 9, 11, 16, 18, 23, 32, 38, 43, 54] aims to transfer knowledge from a pre-trained teacher network to a student network for improving its performance. Self-knowledge distillation [30, 41, 47], distinct from traditional knowledge distillation, leverages the information within the student network to facilitate its learning without the pre-trained teacher network. Specifically, data augmentation approach [15, 46, 50] transfers knowledge through different distortions of the same training data. However, they are susceptible to inappropriate augmentations, such as improper instance rotation or distortion, potentially intro" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": "ducing noise that hampers network learning. Another typical approach exploits auxiliary networks [52, 57]. For example, DKS [40] introduces auxiliary supervision branches and pairwise knowledge alignments, while FRSKD [19] adds a new branch supervised by the original features and utilizes both soft-label and feature-map distillation. Our work is the first that introduces self-knowledge distillation with auxiliary networks for weakly supervised monocular 3D detection. It effectively exploits depth information from single-view images with little computational overhead during inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 532, + 389, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 389, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 389, + 546 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 553, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 547, + 648 + ], + "type": "text", + "content": "This section presents the proposed SKD-WM3D. First, the problem definition and overview are presented in Sec. 3.1. Then detailed designs of SKD-WM3D are introduced, including the self-knowledge distillation framework in Sec. 3.2, the uncertainty-aware distillation loss in Sec. 3.3 and the gradient-targeted transfer modulation strategy in Sec. 3.4. Finally, loss functions are presented in Sec. 3.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 486, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 486, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 486, + 671 + ], + "type": "text", + "content": "3.1. Problem Definition and Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": "Weakly supervised monocular 3D detection takes an RGB image and the corresponding 2D bounding boxes as supervision, aiming to classify objects and determine their" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10510" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": "bounding boxes in 3D space without involving any 3D annotations in training. The prediction of each object is composed of the object category " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": ", a 2D bounding box " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "B^{2D}" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": " and a 3D bounding box " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "B^{3D}" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": ". Specifically, the 3D box " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "B^{3D}" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": " can be further decomposed to the object 3D location " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "(x_{3D}, y_{3D}, z_{3D})" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": ", the object dimension with height, width and length " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "(h_{3D}, w_{3D}, l_{3D})" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": ", as well as orientation " + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 72, + 286, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": "We design a self-knowledge distillation framework to tackle the challenge of weakly supervised monocular 3D detection from a single-view image. As Figure 2 shows, the framework consists of two subnetworks including a Depth-Guided Self-Teaching Network and a Monocular 3D Detection Network. In the Depth-Guided Self-Teaching Network, the global features " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " extracted by the backbone are fed into a Depth Head to obtain depth features. Next, the global features " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " and the extracted depth features are fed into a Fusion Layer to obtain 3D-like features " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "F_{3D}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": ". The 3D-like features of each object are then obtained via RoIAlign, and further fed to a Depth-Aware 3D Head to predict 3D box " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "\\widehat{B}_p^{3D}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " and uncertainty " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "\\widehat{U}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": ". In the Monocular 3D Detection Network, We first use RoIAlign to generate object-level features from the global features " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": ", and then feed them to a 2D-to-3D Head to predict 3D box " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "\\widehat{B}_p^{2D}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " and uncertainty " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": ". The 3D boxes predicted by both networks are further projected into 2D boxes. Moreover, we design an uncertainty-aware distillation loss " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " to obtain low-uncertainty knowledge, and a gradient-targeted transfer modulation strategy to synchronize the learning paces between the two networks by controlling gradients " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "\\widehat{G}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 47, + 157, + 286, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 426, + 255, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 255, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 255, + 437 + ], + "type": "text", + "content": "3.2. Self-Knowledge Distillation Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 444, + 286, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 286, + 503 + ], + "type": "text", + "content": "The self-knowledge distillation framework enhances the 3D localization ability of the depth-guided self-teaching network by utilizing depth information from an off-the-shelf depth estimator and then transfers the ability to the monocular 3D detection network via self-knowledge distillation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "text", + "content": "Depth-Guided Self-Teaching Network. To equip the self-teaching network with 3D localization ability, we propose to learn from global features " + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "text", + "content": " and depth information from an off-the-shelf depth estimator to acquire comprehensive 3D knowledge. The depth information is exploited via two major designs. Firstly, we introduce a depth head " + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "text", + "content": " that extracts depth features " + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "inline_equation", + "content": "F_{D}" + }, + { + "bbox": [ + 47, + 517, + 286, + 601 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 613, + 286, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 613, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 136, + 613, + 286, + 625 + ], + "type": "interline_equation", + "content": "F _ {D} = \\mathcal {D} \\left(F _ {G}\\right), \\tag {1}", + "image_path": "f6f633afdd652508b18a8b00a025d2c17226b2d073807edd939b4214aa8436d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": "The depth features " + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "inline_equation", + "content": "F_{D}" + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": " are exploited to generate depth maps " + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": ", where the depth map generation is supervised by the pseudo ground truth of the depth map " + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "inline_equation", + "content": "D_{gt}" + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": " that is predicted by an off-the-shelf depth estimator by using the focal loss [22] as depth loss " + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": ". Hence, the depth features can be acquired by the depth-guided self-teaching network effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "Remark 1. We generate depth pseudo labels using an off-the-shelf depth estimator [17] with frozen weights, eliminating the need for additional training and ground-truth depth labels. Adopting an off-the-shelf depth estimator incurs negligible costs as compared with prior studies that require either point clouds [34] or multi-view images [42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": "Secondly, we obtain 3D-like features " + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{G3D}" + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": " by integrating the depth features " + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{D}" + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": " that provide information along the depth dimension, as well as the global features " + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": " that capture knowledge about the 2D image plane. Specifically, we design a fusion layer that fuses the depth features " + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{D}" + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": " with the global features " + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": " to derive the " + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{G3D}" + }, + { + "bbox": [ + 305, + 145, + 545, + 217 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 349, + 231, + 545, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 231, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 349, + 231, + 545, + 243 + ], + "type": "interline_equation", + "content": "F _ {G 3 D} = F F N \\left(C A \\left(S A \\left(F _ {D}\\right), F _ {G}\\right)\\right), \\tag {2}", + "image_path": "72e2bff494ce7aef96b167151141d5f2d273bd3b14c0c6df10e221bb4e65da01.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "inline_equation", + "content": "FFN" + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "text", + "content": " is the feed-forward network, and " + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "inline_equation", + "content": "CA" + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "inline_equation", + "content": "SA" + }, + { + "bbox": [ + 305, + 251, + 545, + 335 + ], + "type": "text", + "content": " denote CrossAttention, SelfAttention, respectively. The structures of CrossAttention and SelfAttention employ the standard transformer architecture [44]. The obtained 3D comprehension improves the network's ability to precisely locate objects, effectively mitigating depth ambiguity arising from single-view image input." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 354, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 472 + ], + "type": "text", + "content": "Monocular 3D Detection Network. The monocular 3D detection network acquires the 3D localization knowledge from the depth-guided self-teaching network. By distilling soft labels generated by the depth-guided self-teaching network, the monocular 3D detection network can extract intrinsic depth information from images independently during inference. This kills the need for additional complex modules such as pre-trained depth estimation networks or depth fusion modules, facilitating the inference with little computational overhead." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 483, + 495, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 495, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 495, + 495 + ], + "type": "text", + "content": "3.3. Uncertainty-Aware Distillation Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 502, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 620 + ], + "type": "text", + "content": "During the knowledge distillation process, uncertain knowledge could affect the network training negatively if all transferred knowledge is treated equally. To benefit more from certain knowledge and weaken the effect of uncertain knowledge, we design an uncertainty-aware distillation loss between the 3D boxes that are predicted by the two networks in the self-knowledge distillation framework. The uncertainty-aware distillation loss exploits the prediction uncertainty to modulate the distillation loss magnitude as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 642, + 545, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 642, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 311, + 642, + 545, + 677 + ], + "type": "interline_equation", + "content": "L _ {u d} = \\frac {L _ {d}}{\\operatorname* {m i n} ((\\widehat {U} + U) / 2 , \\alpha)} + \\left\\| \\operatorname * {m i n} (\\frac {\\widehat {U} + U}{2}, \\alpha) \\right\\| ^ {2}, \\tag {3}", + "image_path": "f30a0d5cae09cae43da67d64796b1ca14715ebd40aa0a31a0717c217d65bc9cb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{U}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " denote the uncertainty of the 3D boxes that are predicted by the two networks, respectively. Here" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10511" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "text", + "content": "we assume the 3D box predictions have the Laplace distribution, and we adopt the standard deviations as the uncertainties, inspired by [20, 25, 33]. " + }, + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "inline_equation", + "content": "\\left\\| \\min (\\frac{\\widehat{U} + U}{2},\\alpha)\\right\\| ^2" + }, + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "text", + "content": " is the L2 regularization, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "text", + "content": " is a fixed value set to 0.1. " + }, + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 186 + ], + "type": "text", + "content": " denotes the basic distillations loss, and we employ the commonly used SmoothL1 [13] loss to enforce the consistency between the 3D boxes predicted by the two networks. The SmoothL1 loss leaves a soft margin when computing the difference between the two 3D boxes:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 202, + 287, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 202, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 52, + 202, + 287, + 240 + ], + "type": "interline_equation", + "content": "L _ {d} = \\left\\{ \\begin{array}{l l} 0. 5 \\times \\left(\\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D}\\right) ^ {2}, & \\text {i f} \\left| \\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \\right| < 1. 0 \\\\ \\left| \\widehat {B} _ {p} ^ {3 D} - B _ {p} ^ {3 D} \\right| - 0. 5, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {4}", + "image_path": "0c298dea297f820b2c703da4fea362668156ffde46843e5f269598122915bdf0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\widehat{B}_p^{3D}" + }, + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "type": "inline_equation", + "content": "B_p^{3D}" + }, + { + "bbox": [ + 47, + 240, + 287, + 277 + ], + "type": "text", + "content": " are the predicted 3D boxes from the depth-guided self-teaching network and the monocular 3D detection network, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": "Remark 2. The proposed uncertainty-aware distillation loss " + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": " integrates average uncertainty " + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "inline_equation", + "content": "\\frac{\\widehat{U} + U}{2}" + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": " as regularization and a weighted component for the basic distillation loss " + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": ", allowing adaptive learning adjustments based on the knowledge's uncertainty level. Specifically, when dealing with uncertain knowledge, a smaller weight is assigned to the basic distillation loss " + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": " to mitigate potential adverse effects on network learning. Consequently, the network prioritizes optimizing uncertainty reduction in such scenarios. When dealing with certain knowledge, the network emphasizes optimizing the basic distillation loss " + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": " due to its higher weight. Notably, the basic distillation loss " + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 291, + 289, + 462 + ], + "type": "text", + "content": " simply considers box consistency, while integrating uncertainty is beneficial for enhancing the knowledge distillation process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 468, + 211, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 211, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 211, + 481 + ], + "type": "text", + "content": "3.4. Transfer Modulation Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": "The depth-guided self-teaching network, which leverages depth information to predict 3D boxes, transfers its learned 3D knowledge to the monocular 3D detection network. The asynchronous learning paces of the two networks pose potential challenges to effective 3D knowledge transfer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "type": "text", + "content": "We design a gradient-targeted transfer modulation strategy to synchronize the learning pace of the depth-guided self-teaching network and the monocular 3D detection network. We modulate the knowledge transfer dynamically, by controlling the gradients from the uncertainty-aware distillation loss " + }, + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "type": "text", + "content": ". Specifically, we adapt the gradients based on the 2D projection performance of each network, assigning smaller backward gradients for the good-performing network and higher backward gradients for the bad-performing network. The gradient-targeted transfer modulation strategy is formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 685, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 685, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 52, + 685, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\widehat {G} ^ {\\prime} = \\frac {2 \\times \\widehat {L} _ {p r o j}}{\\widehat {L} _ {p r o j} + L _ {p r o j}} \\times \\widehat {G}, G ^ {\\prime} = \\frac {2 \\times L _ {p r o j}}{\\widehat {L} _ {p r o j} + L _ {p r o j}} \\times G, \\tag {5}", + "image_path": "9f69c2d44371c75a1bc643c4dc9abf8cfc18fab51cc563a968a2cd571717acc0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\widehat{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " are the original gradients of the two networks, " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\widehat{G}^{\\prime}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "G^{\\prime}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " are the modified gradients, " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\widehat{L}_{proj}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "L_{proj}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " are projection losses, computed between the projected 2D boxes from 3D box predictions and 2D box annotations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": "The gradient-targeted transfer modulation strategy prioritizes training the monocular 3D detection network when its learning lags behind the depth-guided self-teaching network at the early training stage. As the monocular 3D detection network learns and improves gradually, it is enabled to provide more feedback progressively to the depth-guided self-teaching network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 224, + 399, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 399, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 399, + 236 + ], + "type": "text", + "content": "3.5. Loss Functions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": "The overall objective consists of three losses including " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "L_{base}" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": " is the uncertainty-aware distillation loss as defined in Sec. 3.3. " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": " is the depth loss for supervising the predicted depth map. " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "L_{base}" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": " includes losses for supervising 2D boxes prediction by 2D heads and the 3D box predictions, which has been adopted in prior CenterNet [55] and WeakMono3D [42]. We set the weight for each loss item to 1.0, and the overall loss function can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 371, + 361, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 361, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 371, + 361, + 545, + 374 + ], + "type": "interline_equation", + "content": "L = L _ {u d} + L _ {d e p} + L _ {b a s e}. \\tag {6}", + "image_path": "49b96de24ca210ac0b79159b50d2282ec09b777afc9344139d208845cef0e191.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 384, + 388, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 388, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 388, + 399 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 404, + 369, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 369, + 416 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 369, + 416 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 423, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 591 + ], + "type": "text", + "content": "We conduct experiments over the KITTI 3D dataset [12] and the nuScenes dataset [3] that have been widely adopted for benchmarking of 3D object detection methods. The KITTI 3D dataset consists of 7,481 images for training and 7,518 images for testing. The labels of the train set are publicly available and the labels of the test set are stored on a test server for evaluation. For ablation studies, we follow [5] which divides the 7,481 training samples into a new train set with 3,712 images and a validation set with 3,769 images. The nuScenes dataset comprises 1,000 video scenes, including RGB images captured by 6 surround-view cameras. The dataset is split into a training set (700 scenes), a validation set (150 scenes), and a test set (150 scenes). Following [34], the performance on the validation set is reported." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 599, + 426, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 426, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 426, + 611 + ], + "type": "text", + "content": "4.2. Evaluation Protocols" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "For the KITTI 3D dataset, following [39], we adopt the evaluation metric " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}|_{R_{40}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " which is the average of the AP of 40 recall points. We report the average precision on bird's eye view and 3D object detection as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{BEV}|_{R_{40}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D}|_{R_{40}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". In addition, as most weakly supervised 3D object detection methods apply IoU threshold of 0.7 for the test set and 0.5 for the validation set, we adopt the same thresholds for fair benchmarking. We adopt four metrics" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10512" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 548, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 548, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 548, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 548, + 145 + ], + "type": "table", + "html": "
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.7)|R40
EasyModerateHard
WeakM3D [34]ResNet-50Weak11.82/5.035.66/2.264.08/1.63
WeakMono3D [42]DLA-3412.31/6.988.80/4.857.81/4.45
SKD-WM3D (Ours)DLA-3415.71/8.9510.15/5.548.08/4.53
", + "image_path": "b0734fa605a937ef21ab45bb4ed871bea73f02636aac209d8f20214e6c65676e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 186, + 544, + 396 + ], + "blocks": [ + { + "bbox": [ + 46, + 153, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 547, + 177 + ], + "type": "text", + "content": "Table 1. Comparison on the performance of the Car category on KITTI test set. For all results, we use " + }, + { + "bbox": [ + 46, + 153, + 547, + 177 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}|_{R_{40}}" + }, + { + "bbox": [ + 46, + 153, + 547, + 177 + ], + "type": "text", + "content": " metrics with IoU threshold equal to 0.7. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 186, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 51, + 186, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 51, + 186, + 544, + 396 + ], + "type": "table", + "html": "
MethodBackboneSupervisionAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
CenterNet [55]DLA-3434.36/20.0027.91/17.5024.65/15.57
MonoGRNet [35]VGG-1652.13/47.5935.99/32.2828.72/25.50
M3D-RPN [1]DenseNet-12153.35/48.5339.60/35.9431.76/28.59
MonoPair [6]DLA-34Full61.06/55.3847.63/42.3941.92/37.99
MonoDLE [28]DLA-3460.73/55.4146.87/43.4241.89/37.81
GUPNet [25]DLA-3461.78/57.6247.06/42.3340.88/37.59
Kinematic [2]DenseNet-12161.79/55.4444.68/39.4734.56/31.26
MonoDistill [7]DLA-3471.45/65.6953.11/49.3546.94/43.49
MonoDETR [53]*ResNet-5072.34/68.0551.97/48.4246.94/43.48
VS3D [36]VGG-16Weak31.59/22.6220.59/14.4316.28/10.91
Autolabels [51]ResNeXt10150.51/38.3130.97/19.9023.72/14.83
WeakM3D [34]ResNet-5058.20/50.1638.02/29.9430.17/23.11
WeakMono3D [42]DLA-3454.32/49.3742.83/39.0140.07/36.34
SKD-WM3D (Ours)DLA-3455.47/50.2144.35/41.5741.86/36.92
", + "image_path": "29d18220b2e52e9740655ff64e77a0128c9a1026235f706e78b0198de44635b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 456, + 286, + 500 + ], + "blocks": [ + { + "bbox": [ + 46, + 403, + 546, + 439 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 546, + 439 + ], + "type": "text", + "content": "Table 2. Comparison on the performance of the Car category on KITTI val set. For all results, we use " + }, + { + "bbox": [ + 46, + 403, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}|_{R_{40}}" + }, + { + "bbox": [ + 46, + 403, + 546, + 439 + ], + "type": "text", + "content": " metric with IoU threshold equal to 0.5. * denotes this performance is reproduced from the official code. The best results of weakly supervised 3D object detection approaches are in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 456, + 286, + 500 + ], + "lines": [ + { + "bbox": [ + 50, + 456, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 50, + 456, + 286, + 500 + ], + "type": "table", + "html": "
MethodAP↑ATE↓ASE↓AAE↓
WeakM3D [34]0.2140.8140.2340.682
SKD-WM3D (Ours)0.2420.7950.2310.659
", + "image_path": "2c36259030da3efc0c5c8fc03e9c3ac3163f0424e08174e5425a26e39193c4ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 529 + ], + "type": "text", + "content": "Table 3. Comparison on the performance of the Car category on nuScenes val set. The best results are in **bold**." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 552, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 636 + ], + "type": "text", + "content": "for the evaluation on the nuScenes dataset, namely, AP (Average Precision), ATE (Average Translation Error), ASE (Average Scale Error), and AAE (Average Attribute Error). Following [34], AVE (Average Velocity Error) and AOE (Average Orientation Error) are not reported due to the lack of supervision for velocity and movement direction in the weakly supervised approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 646, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 180, + 659 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We conduct experiments on 2 NVIDIA V100 GPUs with batch size of 16, and train the framework with 150 epochs. We use the Adam optimizer with the initial learning rate " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", which is gradually increased to " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " for the first 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "content": "epochs and decayed with rate 0.1 at the 90 and 120 epochs. We employ DLA-34 [49] as the detector's backbone. The pseudo ground truth of the depth map is generated with an off-the-shelf depth estimator [17] without using the ground truth of depth label." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 527, + 531, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 531, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 531, + 540 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": "We compare our method with several state-of-the-art weakly supervised monocular 3D detection methods on the KITTI test set. As Table 1 shows, our method achieves superior detection performance across all metrics. This superior performance is largely attributed to our designed self-knowledge distillation framework that extracts and exploits intrinsic depth information from a single-view image effectively. It should be highlighted that our method employs a single-view image exclusively without involving additional training data such as LiDAR point clouds [34] or multi-view image pairs [42]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Table 2 shows the benchmarking on the KITTI validation set. Specifically, we compare our method against both state-of-the-art weakly supervised monocular 3D de" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10513" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 212, + 114 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 212, + 114 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 212, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 212, + 114 + ], + "type": "image", + "image_path": "057c72841b314a35e0ac0a50c0a326cbf5dc7208239f4aa06276bd42bbd1f46a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 69, + 380, + 114 + ], + "blocks": [ + { + "bbox": [ + 214, + 69, + 380, + 114 + ], + "lines": [ + { + "bbox": [ + 214, + 69, + 380, + 114 + ], + "spans": [ + { + "bbox": [ + 214, + 69, + 380, + 114 + ], + "type": "image", + "image_path": "0600ff562e8ac152b44fa46026027a87d9e8aab8f50c6450e65fafad4e6accc4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 546, + 113 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 546, + 113 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 546, + 113 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 546, + 113 + ], + "type": "image", + "image_path": "08cfb4d05d59cecb5ea2b04bbfa5b9cb8de5a9a444455febc5696e50c09c708b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 115, + 213, + 207 + ], + "blocks": [ + { + "bbox": [ + 47, + 115, + 213, + 207 + ], + "lines": [ + { + "bbox": [ + 47, + 115, + 213, + 207 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 213, + 207 + ], + "type": "image", + "image_path": "d65e0f934483acce78f0b92473e426bec7c015c5c74af43a06d15a82baaadf27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 213, + 115, + 359, + 207 + ], + "blocks": [ + { + "bbox": [ + 213, + 115, + 359, + 207 + ], + "lines": [ + { + "bbox": [ + 213, + 115, + 359, + 207 + ], + "spans": [ + { + "bbox": [ + 213, + 115, + 359, + 207 + ], + "type": "image", + "image_path": "60304c8f5a19a03ad701811eb6601a39cfff78425974ad0bdc67d00ce53dad3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 115, + 547, + 207 + ], + "blocks": [ + { + "bbox": [ + 380, + 115, + 547, + 207 + ], + "lines": [ + { + "bbox": [ + 380, + 115, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 380, + 115, + 547, + 207 + ], + "type": "image", + "image_path": "9e8ce927cd9f6f264a05af9ee17148a05d81c2859c7600b5711df746721c292b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 209, + 213, + 254 + ], + "blocks": [ + { + "bbox": [ + 48, + 209, + 213, + 254 + ], + "lines": [ + { + "bbox": [ + 48, + 209, + 213, + 254 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 213, + 254 + ], + "type": "image", + "image_path": "df3e0f5292b7c4c7430b564b130d2f26fa936f378e06f1dd171a092f90d7119a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 214, + 209, + 379, + 254 + ], + "blocks": [ + { + "bbox": [ + 214, + 209, + 379, + 254 + ], + "lines": [ + { + "bbox": [ + 214, + 209, + 379, + 254 + ], + "spans": [ + { + "bbox": [ + 214, + 209, + 379, + 254 + ], + "type": "image", + "image_path": "876f9e572f579b8bf2508db789aed052dfcd62f342d6eefe691d6abcee33100b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 381, + 209, + 546, + 254 + ], + "blocks": [ + { + "bbox": [ + 381, + 209, + 546, + 254 + ], + "lines": [ + { + "bbox": [ + 381, + 209, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 381, + 209, + 546, + 254 + ], + "type": "image", + "image_path": "4b9f731d9955b215ed737a365cb6480da93be966318806fd5790d32ff911a7db.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 47, + 255, + 213, + 347 + ], + "blocks": [ + { + "bbox": [ + 47, + 255, + 213, + 347 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 213, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 213, + 347 + ], + "type": "image", + "image_path": "314e618824552a66c1f8f7393644dd56a82856be675d3e1ecbf390dcb615c8c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 546, + 376 + ], + "type": "text", + "content": "Figure 3. Qualitative illustration on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The ground truth of LiDAR point clouds is utilized for visualization purposes only. Best viewed with zoom-in." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 213, + 255, + 359, + 347 + ], + "blocks": [ + { + "bbox": [ + 213, + 255, + 359, + 347 + ], + "lines": [ + { + "bbox": [ + 213, + 255, + 359, + 347 + ], + "spans": [ + { + "bbox": [ + 213, + 255, + 359, + 347 + ], + "type": "image", + "image_path": "aece145fc69eb5683cfab827127466604d765005f0773e5253e5be49c1cedfd5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 380, + 255, + 546, + 347 + ], + "blocks": [ + { + "bbox": [ + 380, + 255, + 546, + 347 + ], + "lines": [ + { + "bbox": [ + 380, + 255, + 546, + 347 + ], + "spans": [ + { + "bbox": [ + 380, + 255, + 546, + 347 + ], + "type": "image", + "image_path": "6dba897aa40fcd90d30c7227385a89e2390837d76e9ab9f31fb19def9bf007c0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 48, + 387, + 145, + 453 + ], + "blocks": [ + { + "bbox": [ + 48, + 387, + 145, + 453 + ], + "lines": [ + { + "bbox": [ + 48, + 387, + 145, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 145, + 453 + ], + "type": "image", + "image_path": "058031bb8fea94bea61f020151aad422aca6d5bc02058c9294da65d9587a4af5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 455, + 131, + 465 + ], + "lines": [ + { + "bbox": [ + 62, + 455, + 131, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 131, + 465 + ], + "type": "text", + "content": "Uncertainty: 0.24" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 149, + 387, + 245, + 453 + ], + "blocks": [ + { + "bbox": [ + 149, + 387, + 245, + 453 + ], + "lines": [ + { + "bbox": [ + 149, + 387, + 245, + 453 + ], + "spans": [ + { + "bbox": [ + 149, + 387, + 245, + 453 + ], + "type": "image", + "image_path": "098fbf338e7868bddc59d92105513f091bbe932038be96b13f210d2d39624f22.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 455, + 230, + 465 + ], + "lines": [ + { + "bbox": [ + 162, + 455, + 230, + 465 + ], + "spans": [ + { + "bbox": [ + 162, + 455, + 230, + 465 + ], + "type": "text", + "content": "Uncertainty: 0.41" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 249, + 387, + 345, + 453 + ], + "blocks": [ + { + "bbox": [ + 249, + 387, + 345, + 453 + ], + "lines": [ + { + "bbox": [ + 249, + 387, + 345, + 453 + ], + "spans": [ + { + "bbox": [ + 249, + 387, + 345, + 453 + ], + "type": "image", + "image_path": "75163d5cb3fa6b78eedcb513a2d195e35852eeb84dd28f0d7e163fdb73878a88.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 455, + 331, + 465 + ], + "lines": [ + { + "bbox": [ + 262, + 455, + 331, + 465 + ], + "spans": [ + { + "bbox": [ + 262, + 455, + 331, + 465 + ], + "type": "text", + "content": "Uncertainty: 0.67" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 350, + 387, + 445, + 453 + ], + "blocks": [ + { + "bbox": [ + 350, + 387, + 445, + 453 + ], + "lines": [ + { + "bbox": [ + 350, + 387, + 445, + 453 + ], + "spans": [ + { + "bbox": [ + 350, + 387, + 445, + 453 + ], + "type": "image", + "image_path": "1614e1879ea16e5919c26b747e9a341e92830d223a75c5835cad2ac7c9bbd06a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 455, + 432, + 465 + ], + "lines": [ + { + "bbox": [ + 362, + 455, + 432, + 465 + ], + "spans": [ + { + "bbox": [ + 362, + 455, + 432, + 465 + ], + "type": "text", + "content": "Uncertainty: 0.75" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 479, + 546, + 513 + ], + "lines": [ + { + "bbox": [ + 46, + 479, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 546, + 513 + ], + "type": "text", + "content": "Figure 4. Qualitative illustration of object detection and the corresponding detection uncertainties on KITTI val set. Red boxes denote ground-truth annotations and Green boxes denote our predictions. The detection accuracy is closely correlated with the detection uncertainty. Best viewed with zoom-in." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 449, + 387, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 449, + 387, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 449, + 387, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 449, + 387, + 545, + 453 + ], + "type": "image", + "image_path": "25642c57775beb7f4493c77949fcb90743213f073fddaf535d8a530f15703460.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 455, + 531, + 465 + ], + "lines": [ + { + "bbox": [ + 463, + 455, + 531, + 465 + ], + "spans": [ + { + "bbox": [ + 463, + 455, + 531, + 465 + ], + "type": "text", + "content": "Uncertainty: 0.82" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 534, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 654 + ], + "type": "text", + "content": "tection methods and fully supervised methods. It can be seen that our method achieves superior performance compared to WeakM3D [34] and WeakMono3D [42] across most metrics, even without using LiDAR point clouds or multi-view images. Notably, our method significantly outperforms WeakM3D [34] in the Moderate and Hard categories, primarily due to the sparse nature of distant LiDAR point clouds adversely affecting its performance. Additionally, its performance is even on par with several fully supervised methods [1, 35, 55]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Table 3 shows the results on the nuScenes validation set. It can be seen that our proposed method outperforms WeakM3D [34] across all four evaluation metrics, validating the effectiveness of our approach." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Qualitative Results Figure 3 shows qualitative results with both 2D RGB images and 3D point clouds. In simple scenarios, our model achieves great prediction accuracy, which is largely attributed to the proposed self-knowledge distillation framework as well as the uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy, all working together to facilitate comprehensive 3D information extraction effectively. However, for heavily occluded or distant objects, the accuracy of orientation and depth estimation tends to drop, which is common for monocular 3D detection due to its ill-posed nature. In addition, Figure 4 shows the visualization of predicted bounding boxes and their corresponding uncertainties. It can be observed that the prediction accuracy of bounding boxes has a close correlation with the prediction uncertainty." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10514" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 127 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 127 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 127 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 127 + ], + "type": "table", + "html": "
IndexMDNDSNAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
1✓0.00/0.000.00/0.000.00/0.00
2✓45.23/40.9634.27/31.0230.17/26.27
3✓✓55.47/50.2144.35/41.5741.86/36.92
", + "image_path": "89f750aa98c0b83629882ef2bcfe05d68a9b16c1e1e97429eda4451fff0a6a5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 191, + 285, + 259 + ], + "blocks": [ + { + "bbox": [ + 47, + 135, + 287, + 180 + ], + "lines": [ + { + "bbox": [ + 47, + 135, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 135, + 287, + 180 + ], + "type": "text", + "content": "Table 4. Ablation study of the proposed self-knowledge distillation framework. The best results are in bold. MDN denotes the Monocular 3D Detection Network, while DSN denotes the Depth-Guided Self-Teaching Network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 191, + 285, + 259 + ], + "lines": [ + { + "bbox": [ + 50, + 191, + 285, + 259 + ], + "spans": [ + { + "bbox": [ + 50, + 191, + 285, + 259 + ], + "type": "table", + "html": "
IndexLudTMSAPBEV/AP3D(IoU=0.5)|R40
EasyModerateHard
149.95/44.6138.24/35.7437.28/34.82
2✓53.16/48.1341.85/39.0240.14/35.70
3✓52.35/46.3041.45/38.9139.73/35.44
4✓✓55.47/50.2144.35/41.5741.86/36.92
", + "image_path": "622d00918e15729b8075cb642107b120747a0a8a1a3f8b5bac69ccfbf7493bfd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 335, + 285, + 361 + ], + "blocks": [ + { + "bbox": [ + 47, + 267, + 286, + 323 + ], + "lines": [ + { + "bbox": [ + 47, + 267, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 286, + 323 + ], + "type": "text", + "content": "Table 5. Ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. The best results are in **bold**. " + }, + { + "bbox": [ + 47, + 267, + 286, + 323 + ], + "type": "inline_equation", + "content": "L_{ud}" + }, + { + "bbox": [ + 47, + 267, + 286, + 323 + ], + "type": "text", + "content": " denotes the Uncertainty-Aware Distillation Loss, while TMS denotes the gradient-targeted transfer modulation strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 335, + 285, + 361 + ], + "lines": [ + { + "bbox": [ + 50, + 335, + 285, + 361 + ], + "spans": [ + { + "bbox": [ + 50, + 335, + 285, + 361 + ], + "type": "table", + "html": "
MethodWeakM3D [34]MonoDistill [7]*SKD-WM3D (Ours)*
FPS13.925.030.3
", + "image_path": "2024d358e5197623cec60688a03b31283cf287cfe9eb24e08bfb234cf683655e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 286, + 393 + ], + "lines": [ + { + "bbox": [ + 47, + 369, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 286, + 393 + ], + "type": "text", + "content": "Table 6. Comparison on inference speed of M3D methods. * denotes this method utilizes dense depth maps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 414, + 140, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 140, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 140, + 427 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 433, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 286, + 529 + ], + "type": "text", + "content": "We conduct extensive ablation studies on the KITTI validation dataset to evaluate our designs. Specifically, we evaluated the efficacy of the two individual networks in the proposed self-knowledge distillation framework. Additionally, we examine the effect of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. Lastly, we evaluated the efficiency of our monocular 3D detection framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": "Self-Knowledge Distillation Framework. We train two models to assess the contributions of the two networks in our proposed self-knowledge distillation framework. As Table 4 shows, training the monocular 3D detection network alone produces few meaningful detection results as the absence of depth information leads to ambiguous object localization along the depth dimension. As a comparison, training the depth-guided self-teaching network alone can produce reasonable detection results thanks to the depth map pseudo labels. In addition, training both subnetworks concurrently produces the best 3D detection, validating the effectiveness of extracting 3D information from a single image. We can also see that including the self-knowledge distillation on top of the depth-guided self-teaching network" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "greatly improves the detection by reducing the adverse effects of uncertain knowledge and enabling communication between the two subnetworks during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 122, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 545, + 315 + ], + "type": "text", + "content": "Uncertainty-Aware Distillation Loss and Gradient-Targeted Transfer Modulation Strategy. Table 5 shows the ablation study of the proposed uncertainty-aware distillation loss and the gradient-targeted transfer modulation strategy. It can be observed that the baseline does not perform well due to the adverse effect of uncertain knowledge and the asynchronous learning paces of the two subnetworks. On top of the baseline, including either the uncertainty-aware distillation loss or the gradient-targeted transfer modulation strategy improves the detection performance significantly, underscoring the importance of attaining high-certainty knowledge and synchronizing the learning paces of the two networks. In addition, combining the two designs achieves the best performance, highlighting their complementary nature and collaborative roles in knowledge acquisition and knowledge transfer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 329, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 389 + ], + "type": "text", + "content": "Inference speed comparison. Table 6 compares the inference speed on the KITTI validation set. Our method demonstrates superior efficiency thanks to our designed self-knowledge distillation framework, without utilizing complex network architectures during inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 399, + 378, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 378, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 378, + 411 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 419, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 635 + ], + "type": "text", + "content": "In this paper, we point out that previous weakly supervised monocular 3D detection methods either require additional LiDAR point clouds or paired images from multiple viewpoints or temporal sequences. To overcome these constraints, we propose a weakly supervised monocular 3D object detection approach that is exclusively grounded on single-view image inputs. Central to our approach is a self-knowledge distillation framework, which effectively harnesses the depth information within a single-view image with little computational overhead during inference. We further introduce an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy, facilitating knowledge acquisition and knowledge transfer, respectively. Finally, extensive experiments demonstrate the effectiveness of our method. Moving forward, we plan to further generalize our work to diverse and challenging scenarios, such as occlusions, varying lighting, and weather conditions, thereby enhancing its practical applicability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10515" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3d object detection in monocular video. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 135-152. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 248 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "type": "text", + "content": "[4] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10379-10388, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 361 + ], + "type": "text", + "content": "[5] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2147-2156, 2016. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 288, + 418 + ], + "type": "text", + "content": "[6] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "type": "text", + "content": "[7] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. International Conference on Learning Representations, 2022. 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "type": "text", + "content": "[8] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In International Conference on Machine Learning, pages 2006-2015. PMLR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 577 + ], + "type": "text", + "content": "[9] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3872-3882, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "text", + "content": "[10] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1000-1001, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 634, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 288, + 679 + ], + "type": "text", + "content": "[11] Huan-ang Gao, Beiwen Tian, Pengfei Li, Hao Zhao, and Guyue Zhou. Dqs3d: Densely-matched quantization-aware semi-supervised 3d detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[13] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1440-1448, 2015. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[14] Tong He and Stefano Soatto. Mono3d++: Monocular 3d vehicle detection with two-scale 3d hypotheses and task priors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8409-8416, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "type": "text", + "content": "[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "text", + "content": "[17] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In International Conference on Robotics and Automation, pages 13656-13662. IEEE, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "text", + "content": "[18] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. Advances in Neural Information Processing Systems, 35:33716-33727, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "text", + "content": "[19] Mingi Ji, Seungjae Shin, Seunghyun Hwang, Gibeom Park, and Il-Chul Moon. Refine myself by teaching myself: Feature refinement via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10664–10673, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 449 + ], + "type": "text", + "content": "[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in Neural Information Processing Systems, 30, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "text", + "content": "[21] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11867-11876, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2980-2988, 2017. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "text", + "content": "[23] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7096-7104, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "text", + "content": "[24] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "text", + "content": "[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. 5, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10516" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[26] Zhipeng Luo, Zhongang Cai, Changqing Zhou, Gongjie Zhang, Haiyu Zhao, Shuai Yi, Shijian Lu, Hongsheng Li, Shanghang Zhang, and Ziwei Liu. Unsupervised domain adaptive 3d detection with multi-level consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8866-8875, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 196 + ], + "type": "text", + "content": "[27] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6851–6860, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "text", + "content": "[28] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 252, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 288, + 308 + ], + "type": "text", + "content": "[29] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 515-531. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 342 + ], + "type": "text", + "content": "[30] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "text", + "content": "[31] J Krishna Murthy, GV Sai Krishna, Falak Chhaya, and K Madhava Krishna. Reconstructing vehicles from a single image: Shape priors for road scene understanding. In International Conference on Robotics and Automation, pages 724-731. IEEE, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[32] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "text", + "content": "[33] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In Proceedings of the IEEE/CVF European Conference on Computer Vision, pages 71-88. Springer, 2022. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[34] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. International Conference on Learning Representations, 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[35] Zengyi Qin, Jinglu Wang, and Yan Lu. Monognet: A geometric reasoning network for monocular 3d object localization. Proceedings of the AAAI Conference on Artificial Intelligence, 2019. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[36] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[37] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[38] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "text", + "content": "[39] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection: From single to multiclass recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1219-1231, 2020. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[40] Dawei Sun, Anbang Yao, Aojun Zhou, and Hao Zhao. Deeply-supervised knowledge synergy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 252 + ], + "type": "text", + "content": "[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 319 + ], + "type": "text", + "content": "[42] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular 3d object detection using multi-view projection and direction consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "text", + "content": "[43] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "type": "text", + "content": "[45] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "text", + "content": "[46] Ting-Bing Xu and Cheng-Lin Liu. Data-distortion guided self-distillation for deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5565-5572, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "type": "text", + "content": "[47] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "text", + "content": "[48] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2403-2412, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "type": "text", + "content": "[50] Sukmin Yun, Jongjin Park, Kimin Lee, and Jinwoo Shin. Regularizing class-wise predictions via self-knowledge distillation. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10517" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 431 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 68, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 287, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 13876-13885, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[51] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[52] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3713-3722, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "text", + "content": "[53] Renrui Zhang, Han Qiu, Tai Wang, Xuanzhuo Xu, Ziyu Guo, Yu Qiao, Peng Gao, and Hongsheng Li. Monodetr: Depth-aware transformer for monocular 3d object detection. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "type": "text", + "content": "[54] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 340 + ], + "type": "text", + "content": "[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "text", + "content": "[56] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 431 + ], + "type": "text", + "content": "[57] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Advances in Neural Information Processing Systems, 31, 2018. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10518" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_content_list.json b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d00962822fbc86c4c0522331d175f0bb856a2843 --- /dev/null +++ b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_content_list.json @@ -0,0 +1,1612 @@ +[ + { + "type": "text", + "text": "Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle", + "text_level": 1, + "bbox": [ + 99, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeokjun Kweon* KAIST", + "bbox": [ + 181, + 181, + 331, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "0327june@kaist.ac.kr", + "bbox": [ + 168, + 219, + 346, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jihun Kim* KAIST", + "bbox": [ + 444, + 181, + 535, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jihun1998@kaist.ac.kr", + "bbox": [ + 398, + 219, + 584, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kuk-Jin Yoon KAIST", + "bbox": [ + 661, + 181, + 772, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kjyoon@kaist.ac.kr", + "bbox": [ + 633, + 219, + 795, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 310, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Manual annotation of every point in a point cloud is a costly and labor-intensive process. While weakly supervised point cloud semantic segmentation (WSPCSS) with sparse annotation shows promise, the limited information from initial sparse labels can place an upper bound on performance. As a new research direction for WSPCSS, we propose a novel Region Exploration via Artificial Labeling (REAL) framework. It leverages a foundational image model as an artificial oracle within the active learning context, eliminating the need for manual annotation by a human oracle. To integrate the 2D model into the 3D domain, we first introduce a Projection-based Point-to-Segment (PP2S) module, designed to enable prompt segmentation of 3D data without additional training. The REAL framework samples query points based on model predictions and requests annotations from PP2S, dynamically refining labels and improving model training. Furthermore, to overcome several challenges of employing an artificial model as an oracle, we formulate effective query sampling and label updating strategies. Our comprehensive experiments and comparisons demonstrate that the REAL framework significantly outperforms existing methods across various benchmarks. The code is available at https://github.com/jihun1998/AO.", + "bbox": [ + 75, + 299, + 472, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 207, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud semantic segmentation has been extensively researched for robotics or autonomous driving applications. While fully supervised approaches [7, 10, 14, 28, 29, 35] utilizing deep learning have demonstrated remarkable progress, the densely annotated point-wise labels required for these methods pose a significant challenge due to the labor-intensive and expensive annotation process. This obstacle hinders the practical application of existing methods.", + "bbox": [ + 75, + 688, + 468, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to this challenge, Weakly Supervised Point Cloud Semantic Segmentation (WSPCSS) has gained attention, leveraging more affordable weak labels [15, 23, 24, 26, 40, 42, 47, 48]. Among various types of weak labels, sparse annotations have become a widely adopted setting, where only a tiny subset (e.g., $0.02\\%$ ) of the entire point cloud is", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7276b774f940b61be70e989ed1b975016f2a7e9949e50e1f5473060fb200d42d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 268, + 883, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f338933edb540490bbef37c53cd2dfcea9d95ddca585a919971bae58ef1bce7.jpg", + "image_caption": [ + "Figure 1. Visualization of the proposed REAL framework. We tackle the challenge of WSPCSS, leveraging a foundational imagery model as an artificial oracle within the context of active learning. The below examples illustrate the evolving labels as training proceeds." + ], + "image_footnote": [], + "bbox": [ + 504, + 405, + 885, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "labeled. Existing WSPCSS works have proposed remedies focusing on two key aspects: 1) fully utilizing the limited information from sparse labels and 2) learning the features beneficial for segmentation from the data. The methods based on self-labeling use the regions confidently predicted by the model as pseudo-labels for training the model itself. Meanwhile, various studies have proposed self-supervised learning methods, such as consistency against data augmentation, contrastive learning, or masked modeling.", + "bbox": [ + 496, + 564, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While these approaches have shown substantial results, the overall amount of information provided is constrained by the initial sparse annotation, potentially placing an upper limit on performance. In this paper, instead of solely focusing on how to exploit the given limited information, we embark on a new and more challenging research direction. Specifically, we propose a novel WSPCSS method using an additional publicly available source of information—leveraging advancements in the 2D domain with the Segment Anything Model (SAM) [19].", + "bbox": [ + 496, + 702, + 890, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We begin by utilizing the promptable segmentation capability of SAM to enhance initially provided sparse annotations. To bridge the gap between the 2D model and 3D", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3721", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "point cloud data, we introduce the Projection-based Point-to-Segment (PP2S) module. This module comprises the projection of weakly annotated points onto images, acquiring 2D masks using SAM, and back-projecting these masks into 3D, facilitating segmentation in the point cloud. The proposed PP2S-based preprocessing effectively improves label quality, resulting in significantly enhanced semantic segmentation performance.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the labels preprocessed by PP2S remain incomplete due to potential noises from calibration errors or SAM ambiguity. To further enhance label quality during training, we propose the Region Exploration via Artificial Labeling (REAL) framework, utilizing PP2S as an artificial oracle within the context of active learning, as depicted in Fig. 1. The REAL framework samples query points according to model predictions and subsequently requests annotations from PP2S to dynamically enhance training labels. Here, unlike general active learning methods relying on a human oracle, our artificial oracle cannot directly assign the classes of the requested query. Therefore, we design specialized strategies for query sampling and label refinement to unlock the segmentation capability of the PP2S.", + "bbox": [ + 75, + 213, + 468, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that the main novelty of our approach lies in how we effectively integrate the benefits of SAM into the realm of WSPCSS. We analyze the REAL framework through extensive experiments, including quantitative and qualitative ablation studies. Furthermore, our method achieves new state-of-the-art (SoTA) results on S3DIS [3] and ScanNetV2 [8] under all the tested settings, surpassing the existing WSPCSS methods. Remarkably, even with the $0.004\\%$ setting, our method outperforms conventional methods using the $0.02\\%$ setting, underscoring the superiority of the proposed approach.", + "bbox": [ + 75, + 425, + 468, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 608, + 218, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Weakly Supervised Approaches", + "text_level": 1, + "bbox": [ + 76, + 633, + 356, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Originating from weakly supervised learning in imagery domain [2, 21, 22, 31, 39, 45], weakly supervised approach has emerged as a cost-effective approach to mitigate the challenges associated with acquiring fully labeled data [15, 18, 23, 24, 26, 40, 42, 47, 48]. Xu and Lee [40] first propose weakly supervised configurations involving annotations of $10\\%$ of the data and a single point label for each category. PSD [48] introduces a perturbed branch and constraint self-distillation loss between the perturbed and original branches. OTOC [26] utilizes pseudo-label methods and iteratively updates pseudo-labels through a self-training strategy. MIL [42] proposes the transformer model derived from multiple instance learning and integrates adaptive global weighted pooling to their model. HybridCR [23] introduces an architecture that leverages pseudo-label methods and consistency regularization between the original and", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "augmented branches. CPCM [24] introduces a region-wise masking strategy and contextual masked training method to integrate the benefit of the masked autoencoder. Despite the significant progress in WSPCSS, in this paper, we would like to explore a new and more challenging research direction. In this light, we suggest a novel WSPCSS method, using a publicly available additional source of information.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Point Cloud Segmentation Using 2D Models", + "text_level": 1, + "bbox": [ + 498, + 207, + 872, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There have been extensive efforts to achieve point cloud segmentation in an image-based manner as in [12, 16, 20, 30, 43]. With the rapid advancement in the 2D domain, multi-modal self-supervised approaches [1, 6, 13] and employing language-based model [9, 27, 34] have also been particularly prevalent. Recently, there has been a surge in attempts to address various point cloud tasks using the Segment Anything Model (SAM) [19] The zero-shot approach [44, 46], which directly applies SAM to the image corresponding to the point cloud, has demonstrated promising segmentation results. However, SAM's lack of explicit semantics limits its applicability to high-level perception tasks. Adaptation-based approaches [11] can make SAM directly learn 3D tasks; however, they necessitate dense ground truth data during the learning process, rendering them unsuitable for WSPCSS. To tackle the weakly supervised setting, this paper presents a novel WSPCSS method that effectively refines weak labels using the foundational 2D model as an artificial oracle.", + "bbox": [ + 496, + 229, + 890, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Active Learning for Point Cloud Segmentation", + "text_level": 1, + "bbox": [ + 498, + 527, + 890, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Active Learning (AL) aims to identify a subset of instances that an oracle manually annotates during training. As in [17, 32, 36, 41], AL has also been extensively explored for WSPCSS. Given that the REAL framework actively employs SAM as an artificial oracle, it may share a conceptual scheme with conventional AL-based methods to some extent. Nevertheless, note that our weakly-supervised setting significantly differs from them. Our framework solely relies on initial sparse labels and does not involve any additional manual annotation within the training loop.", + "bbox": [ + 496, + 550, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 715, + 591, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 741, + 640, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem setting. We denote an input point cloud as a set of $N$ points: $X = \\{x_{1},\\ldots ,x_{N}\\}$ . In typical fully-supervised point cloud semantic segmentation, the labels for all points are provided as $Y = \\{y_{1},\\dots ,y_{N}\\}$ where $y_{i}\\in \\{1,\\dots ,C\\}$ and $C$ is the number of classes. On the other hand, in weakly-supervised point cloud semantic segmentation, only a tiny subset of the points is labeled, and the other points are unlabelled. We formulate this setting as $(X,Y) = (X^{L},Y^{L})\\cup (X^{U},Y^{U})$ . Here, $X^{L} = \\{x_{1}^{L},\\ldots ,x_{n}^{L}\\}$ is the", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3722", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "points labeled with sparse annotations $Y^{L} = \\{y_{1}^{L},\\dots ,y_{n}^{L}\\}$ Conversely, $X^{U} = \\{x_{1}^{U},\\ldots ,x_{N - n}^{U}\\}$ is not labeled, and we represent it as $Y^{U} = \\{y_{1}^{U},\\dots ,y_{N - n}^{U}\\}$ , where $y_{i}^{U} = 0$", + "bbox": [ + 76, + 89, + 467, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segment Anything Model. SAM is a foundational model for promptable segmentation, comprising 1) an image encoder, 2) a prompt encoder, and 3) a decoder predicting the mask using the embeddings from encoders. We denote the inference process of SAM as", + "bbox": [ + 76, + 136, + 467, + 210 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {2 D} = \\operatorname {S A M} (I; Z), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 217, + 467, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I$ and $Z$ are an input image and prompt, respectively. $M^{2D}$ denotes a set of the pixels of the predicted mask. In this paper, we use a point prompt (i.e., a pixel coordinate).", + "bbox": [ + 76, + 242, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Projection-based Point-to-Segment", + "text_level": 1, + "bbox": [ + 76, + 294, + 380, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by enhancing the initially given weak annotations using SAM. However, SAM is a vision foundation model designed and trained for 2D imagery. Therefore, to directly process 3D point cloud data using SAM, it is necessary to conduct additional adaptation or fine-tuning.", + "bbox": [ + 76, + 316, + 467, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead, we propose to utilize image data that captures the same scene as the provided point cloud data. We believe that incorporating image data is not overly restrictive in practice. In fact, point cloud data is usually created from RGBD frames using Structure from Motion (SfM) [3, 8] or captured by LiDAR paired with RGBD cameras [4, 5, 33]. Hence, most point cloud datasets naturally contain RGB images, depth maps, and associated camera parameters. Under this philosophy, we introduce a Projection-based Point-to-Segment (PP2S) module, enabling promptable segmentation of 3D point cloud data using SAM via projection. It is noteworthy that our method utilizes image data during the training phase only and can perform semantic segmentation using input point cloud data alone during testing.", + "bbox": [ + 75, + 393, + 467, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates the visualization of PP2S. Formally, for each point cloud $X$ , we assume that the model can access data from $J$ cameras: images, depth maps, and projection matrices. Initially, we project the set of weakly annotated points $X^{L} = \\{x_{i}^{L}\\}$ onto the image $I_{j}$ as:", + "bbox": [ + 76, + 604, + 467, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} = \\operatorname {p r o j} _ {j} x _ {i} ^ {L} \\text {w h e r e} i \\in \\{1, \\dots , n \\}, j \\in \\{1, \\dots , J \\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 686, + 467, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{proj}_j$ denotes the projection matrix, and $p_{i,j}$ represents the 2D coordinate of the $i$ th weakly annotated points on the $j$ th image. After the projection, we discard points that fall outside the image boundaries. We also filter out occluded points using the depth map $D_j$ , though alternative visibility testing algorithms can be employed.", + "bbox": [ + 76, + 718, + 467, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, SAM performs segmentation using the coordinates of projected points as individual input prompts as", + "bbox": [ + 76, + 809, + 467, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} ^ {2 D} = \\operatorname {S A M} \\left(I _ {j}; p _ {i, j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 844, + 467, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{i,j}^{2D}$ is a set of the pixels of predicted 2D mask on $I_j$ , given input point prompt $p_{i,j}$ . Since $p_{i,j}$ corresponds to", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9e3ef5efdabdc8bac842395a92c9afc8483a8c9e92b9eb59862ec8dbbefde46f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 883, + 176 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5d5f5d52c42afc0921b170db5e0742956aa933f46913fd0f03e72d716cdade6.jpg", + "image_caption": [ + "Figure 2. Visualization of the proposed Projection-based Point-to-Segment (PP2S) module. The process begins by projecting the weak labels onto images through projection matrices. Using these projected points as point prompts, SAM predicts 2D masks, which are subsequently back-projected onto the point cloud." + ], + "image_footnote": [], + "bbox": [ + 503, + 176, + 883, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$x_{i}^{L}$ and is labeled as $y_{i}^{L}$ , the pixels in $M_{i,j}^{2D}$ are also assigned to $y_{i}^{L}$ . As multiple cameras exist per a single point cloud, we conduct the above process for all images in parallel.", + "bbox": [ + 498, + 412, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, we back-project the acquired 2D masks to 3D coordinates. Initially, similar to Equ. 2, we project the point cloud onto the $j$ th image as $p_{k,j} = \\mathrm{proj}_j x_k$ , where $x_k$ is the $k$ th point in the point cloud and $p_{k,j}$ denotes the projection of $x_k$ . If $p_{k,j}$ is an element of $M_{i,j}^{2D}$ , it means that the projection result of $x_k$ belongs to the mask of the weakly annotated point $x_i^L$ , within the image $I_j$ . Therefore, we define the 3D mask $M_{i,j}^{3D}$ as the set of $x_k$ such that its projection $p_{k,j}$ is in $M_{i,j}^{2D}$ .", + "bbox": [ + 498, + 460, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above process can be formulated as", + "bbox": [ + 517, + 598, + 784, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} ^ {3 D} = \\left\\{x _ {k} \\mid p _ {k, j} \\in M _ {i, j} ^ {2 D} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 619, + 890, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{i,j}^{3D}$ denotes the set of points predicted by SAM to be the same segment with $x_{i}^{L}$ from the perspective of $j$ th image. Accordingly, $y_{i,j}^{3D}$ , the class of $M_{i,j}^{3D}$ , is $y_{i}^{L}$ .", + "bbox": [ + 498, + 650, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the 3D masks are not mutually exclusive. Given that a single 3D point could be observed from multiple images, it may be shared by more than one 3D mask. To address this, we establish a consensus among the 3D masks through a mask-wise voting system. In our system, each 3D mask $M_{i,j}^{3D}$ casts a vote for its points being set to $y_{i,j}^{3D}$ . Therefore, from the perspective of each point, its class is determined by the voting of the 3D masks that include the point. The above process can be formulated as", + "bbox": [ + 496, + 699, + 890, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv _ {k} = \\left\\{y _ {i, j} ^ {3 D} \\right\\} \\forall i, j \\text {s u c h t h a t} x _ {k} \\in M _ {i, j} ^ {3 D}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 842, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $v_{k}$ is a multiset of the voted classes. Finally, we assign the class with the highest number of votes among $v_{k}$ as", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3723", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/821b1a39497bff83c1f2cd91ca6131edfc0671f3d224bcf3f8ad433e0f583861.jpg", + "image_caption": [ + "Figure 3. Possible sources of noise in PP2S. (A): The 3D door point is projected onto the chair in the image, resulting in an incorrect 2D mask. (B): Erroneous back-projection causes wall points in 3D to be mislabeled as chair. (C) Above: The 2D mask of the ceiling covers only a small portion of the entire object. (C) Below: The 2D mask of the wall encroaches into the regions of column." + ], + "image_footnote": [], + "bbox": [ + 78, + 92, + 459, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the artificial label $a_{k}$ for $k$ th point $x_{k}$ :", + "bbox": [ + 76, + 409, + 328, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {k} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} v _ {k} = \\emptyset \\\\ \\operatorname {M o d e} \\left(v _ {k}\\right) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 435, + 468, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the Mode operator returns the most frequent element.", + "bbox": [ + 76, + 488, + 467, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed PP2S module effectively leverages the segmentation capability of SAM to expand the sparsely annotated initial labels. These resulting artificial labels, denoted as $A = \\{a_k\\}$ , can directly serve as a supervised signal for training. Notably, we have observed that training with $A$ yields significant performance improvements compared to using initial $Y$ . The detailed experimental results are demonstrated in Section 4.", + "bbox": [ + 75, + 503, + 468, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Region Exploration via Artificial Labeling", + "text_level": 1, + "bbox": [ + 76, + 635, + 436, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned, the proposed PP2S effectively enhances the initial sparse annotations, resulting in a substantial improvement in segmentation performance. However, we also have recognized several limitations associated with this preprocessing approach. The PP2S-enhanced labels can serve as better guidance than the initial sparse labels. Nevertheless, they are still fixed during the entire training phase and thereby impose an upper bound on performance. Notably, we cannot guide the regions not covered by the 2D masks obtained within PP2S. Furthermore, even if we assume that the initially provided sparse annotations are perfectly correct, errors stemming from factors such as noises of camera projection matrices (Fig. 3 A, B) or inherent ambiguity of SAM (Fig. 3 C) are inevitable. While the voting system can mitigate some of these, it cannot completely eliminate them.", + "bbox": [ + 75, + 659, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Region Exploring via Artificial Labeling" + ], + "code_body": "Require: Segmentation model $f_{\\theta}$ , Training dataset $\\mathcal{D} =$ $\\{(X,Y)\\}$ , Learning rate $\\eta$ , Number of epochs $T$ \nEnsure: Optimized $f_{\\theta}$ \n1: for $(X,Y)$ in $\\mathcal{D}$ do \n2: $Y\\gets \\mathrm{PP2S}(X;X^L)$ \n3: end for \n4: Initialize the model parameter $\\theta$ \n5: for $t\\gets 1,2,\\ldots ,T$ do \n6: for $(X,Y)$ in $\\mathcal{D}$ do \n7: $Y^{t - 1}\\gets Y$ \n8: $S\\gets \\mathrm{softmax}(f_{\\theta}(X))$ \n9: $\\hat{Y}\\gets \\mathrm{argmax}(S)$ \n10: $Q\\gets \\mathrm{QuerySampling}(X,S,Y^{t - 1})$ \n11: $A\\gets \\mathrm{PP2S}(X;Q)$ \n12: $Y^{t}\\gets \\mathrm{LabelUpdate}(Y^{t - 1},A,\\hat{Y})$ \n13: $Y\\gets Y^{t}$ \n14: $\\theta \\leftarrow \\theta -\\eta \\nabla_{\\theta}\\mathcal{L}(S,Y)$ \n15: end for \n16: end for", + "bbox": [ + 501, + 109, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, how can we further improve label quality, especially during training? We draw inspiration from the concept of active learning. In active labeling, the model identifies a set of instances, known as a query, during training and requests human annotators (i.e., oracle) to label these points. Typically, the query consists of instances that confuse the model at a given timestep, making annotating the query highly effective for the model to learn decision boundaries. However, active learning requires manual annotation within the training loop, potentially making it even more expensive than WSPCSS.", + "bbox": [ + 496, + 410, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we present the Region Exploration via Artificial Labeling (REAL) framework—a novel WSPCSS method utilizing the PP2S module as an artificial oracle, as depicted in Fig. 4. Specifically, the proposed PP2S module serves as an oracle in the REAL framework, responding to requests to annotate the query. Given that the artificial oracle is, in fact, a pretrained network, the REAL framework is completely relieved of the necessity for manual annotation, except for the initial sparse labels.", + "bbox": [ + 496, + 580, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, unlike a human oracle in conventional methods, our artificial oracle lacks explicit semantic knowledge. Therefore, we cannot expect the PP2S module to directly label the classes for the confusing points. Instead, we focus on leveraging what PP2S can provide to us. PP2S cannot directly annotate the given points; however, it can precisely predict which points should be grouped into a segment. To fully utilize this attribute, we define the query $Q = \\{q_i\\}$ to consist of the confidently identified points, instead of ambiguous ones from the perspective of the model. Trusting the prediction of the model on these points, we can self-label the query with minimal risk. We represent this pro", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3724", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e3da3fe03aceecc7f9da1adc9efa253e4ec16ddbcbecec297306a825b13b677.jpg", + "image_caption": [ + "Region Exploration via Artificial Labeling (REAL, Sec. 3.3)", + "Figure 4. Visualization of the proposed Region Exploration via Artificial Labeling (REAL) framework. The objective is to dynamically enhance the quality of labels for weakly supervised point cloud semantic segmentation. In the REAL framework, the model $f_{\\theta}$ takes the point cloud $X$ as input and predicts semantic segmentation logits $S$ . We then sample the most confident points from the unlabeled regions of the previous label $Y^{t-1}$ based on the predictions $\\hat{Y}$ and reliability scores $R$ . The sampled points then serve as prompts for the proposed PP2S module, producing artificial labels $A$ . Finally, by combining the model prediction and $A$ , we refine $Y^{t-1}$ into the updated labels $Y^t$ which serve as pseudo-ground truth for training the logits." + ], + "image_footnote": [], + "bbox": [ + 83, + 116, + 885, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cess as QuerySampling in Algorithm 1, which will be thoroughly discussed in Section 3.4.", + "bbox": [ + 75, + 525, + 467, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we request the PP2S module to act as an artificial oracle, annotating the unknown regions using the given query. For this, the self-labeled query serves as an input prompt for the PP2S module as follows:", + "bbox": [ + 75, + 556, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {P P 2 S} (X; Q), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 628, + 468, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ represents the artificial label obtained by PP2S from the given query $Q$ . This approach is akin to using a set of sparsely annotated points $(X^L)$ for PP2S in the preprocessing phase, as described in Section 3.2. The primary distinction in the REAL framework is that the input prompt is self-labeled, not manually annotated.", + "bbox": [ + 75, + 657, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the obtained artificial label $A$ , we update the label of the previous step, $Y^{t-1}$ , into $Y^t$ . However, $A$ may not be entirely reliable, as it could be more prone to noise compared to labels provided by a human oracle. This potential for noise could arise from inaccuracies in the self-labeled input query or the issues highlighted in Fig. 3.", + "bbox": [ + 75, + 750, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the risk of incorporating errors during the label-updating process, we devise a cautious approach using the prediction of the model. Specifically, if a discrepancy exists between the class assigned by the artificial label", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$A$ and the model's prediction $\\hat{Y}$ for a specific point, we exclude that point from the updating process. This conservative update process can be formulated as:", + "bbox": [ + 498, + 525, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} a _ {i} & \\text {i f} a _ {i} = \\hat {y} _ {i} \\\\ y _ {i} ^ {t - 1} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 580, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the model $f_{\\theta}$ is optimized by minimizing the cross-entropy loss between $\\hat{Y}$ and $Y^{t}$ .", + "bbox": [ + 498, + 631, + 888, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Query Sampling Strategy for REAL", + "text_level": 1, + "bbox": [ + 500, + 671, + 812, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the REAL framework, query sampling is a critical component, as the quality of the artificial label depends on the quality of the input query. We posit that an effective query must meet two key criteria: 1) it should exhibit a high level of confidence, and 2) it should aid the model in acquiring new knowledge. To address the first criterion, we design a margin-based reliability metric for evaluating the confidence of the model's prediction for each point as follows:", + "bbox": [ + 496, + 694, + 890, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} = \\operatorname {M a x} \\left(s _ {i}\\right) - \\operatorname {M a x} 2 \\left(s _ {i}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 828, + 890, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_i \\in \\mathbb{R}^C$ is the predicted logit of the $i$ th point $x_i$ , and $r_i$ denotes its reliability. Max and Max2 are operators that return the maximum and the second maximum el", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3725", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce20e8baf2ab7ef32c9826296e236d86c41d1390f15d7c285d93d24d00022ecf.jpg", + "image_caption": [ + "Figure 5. Verification of the reliability metric. The points presenting higher reliability show higher semantic segmentation accuracy." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 457, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ements of the input vector, respectively. We verify a robust correlation between the devised reliability metric and the accuracy of model predictions, as illustrated in Fig. 5. For instance, among points with reliability scores exceeding 0.5, the model correctly predicts the class for $84.5\\%$ of them. We set the threshold for query sampling at 0.95, corresponding to around $94\\%$ accuracy.", + "bbox": [ + 75, + 229, + 470, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the contrary, directly quantifying the utility of a query for the second criterion is inherently elusive. In this context, we propose an intuitive hypothesis: if the model has successfully leveraged the labels from the previous timestep, $Y^{t-1}$ , the regions previously annotated in $Y^{t-1}$ are expected to contribute less novel information during the current timestep, $t$ . Therefore, the focus of new queries should be on effectively exploring the regions not previously or incorrectly labeled in $Y^{t-1}$ .", + "bbox": [ + 75, + 335, + 468, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, we narrow the sampling range of the query to the points that are predicted by the model differently from the previous label. We formulate the above constraint as", + "bbox": [ + 75, + 470, + 468, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {c} = \\left\\{x _ {i} \\right\\} \\forall i \\text {s u c h t h a t} \\hat {y} _ {i} = c \\text {a n d} y _ {i} ^ {t - 1} \\neq c. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 523, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $Z_{c}$ is the mother set for sampling $q_{c}$ , the query point of the class $c$ . Considering the first criterion, we sample $q_{c}$ as the most reliable point among $Z_{c}$ , and $Q$ is the set of $q_{c}$ . We formulate the above processes as", + "bbox": [ + 75, + 547, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\operatorname {Q u e r y S a m p l i n g} (X, S, Y ^ {t - 1}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 614, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 641, + 209, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 76, + 665, + 281, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on S3DIS [3] and ScanNetV2 [8]. S3DIS comprises 6 areas with 272 rooms and 13 categories. We use the area 5 for evaluation following previous studies. ScanNetV2 includes 1613 3D scenes with 20 categories. We follow the official split (1201 training, 312 validation, and 100 test scenes). In addition to the point cloud data, the proposed REAL framework requires the data from cameras (i.e., images, depth maps, and camera projection matrices) to leverage SAM. The datasets inherently include such data since they are reconstructed from the sequence of RGBD frames. During training, we utilize the data from 48 and 17 cameras per scene on average for S3DIS and ScanNetV2, respectively. Note that our framework does not require 2D data at inference time.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. A quantitative comparison between the initially sparse annotations (Initial) and the artificial labels from PP2S (PP2S). Pre, Rec, and # denote precision, recall, and the proportion of annotated points relative to the total number of points. The performance of the point cloud semantic segmentation model trained by each model is represented as Model mIoU. Every metric is in %.", + "bbox": [ + 498, + 89, + 892, + 172 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a8c7729b83e96010200762660c55b67a91460db50e949684482c083cafdf8138.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LabelsmIoUPreRec#Model mIoU
Initial0.0041000.0040.00439.7
PP2S30.483.132.438.453.0
", + "bbox": [ + 524, + 176, + 864, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73ed7a6e9f304e8b527c4cfce59c7f270aed69a63f85c610edbae5ff17741ba5.jpg", + "image_caption": [ + "(A) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 522, + 236, + 692, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b14a1b2f96268334df4edd870f80282eb8a837132e98706f11d130ffc8e1b5ba.jpg", + "image_caption": [ + "(B) Labels of the previous step", + "(D) Explored regions (setting II)" + ], + "image_footnote": [], + "bbox": [ + 694, + 237, + 864, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/407bce03c378fee676c01e3a5955b31913472da376d66c02347faaa2ad656b8a.jpg", + "image_caption": [ + "(C) Explored regions (setting 1)", + "Prec: -0.09%", + "Rec: +11.92%", + "Prec: -0.03%", + "Rec: +5.58%", + "Figure 6. Validation of our query sampling strategy. Setting I samples the queries following our strategy while setting II samples from the entire point set. We present the improvement obtained by artificial labels in terms of precision (Prec) and recall (Rec)." + ], + "image_footnote": [], + "bbox": [ + 522, + 339, + 692, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocols. Previous studies [23, 24, 42, 48] have conducted experiments with various levels of sparsity. We believe that the most cost-efficient yet practical setting involves labeling only one point per object. Accordingly, our experiments mainly target the 1pt weak setting, which corresponds to $0.004\\%$ of points in S3DIS. For a fair comparison, we also conduct experiments under the $0.02\\%$ setting, which is the most widely used. For ScanNetV2, we follow the conventional works, using 20pt per scene. The overall performance is assessed on all points within the test set. Following the standard practice, we employ the mean Intersection over Union (mIoU) as our quantitative metric. Implementation details. We choose PTv2 [37] network as our main backbone. To verify the effect of the backbone, we additionally test Closer [25] backbone for the $0.004\\%$ setting on S3DIS. Training details are the same as standard settings of the backbones. We employ a pretrained ViT-H model for SAM. Further details are in the Supp.", + "bbox": [ + 496, + 517, + 892, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Analysis on REAL", + "text_level": 1, + "bbox": [ + 500, + 801, + 679, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PP2S-based preprocessing. We first validate the preprocessing strategy built upon the proposed PP2S. This strategy aims to enhance the initial sparse annotations into artificial labels by harnessing the promptable segmentation capability of SAM via projection. Table 1 provides various met", + "bbox": [ + 496, + 824, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3726", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af5fcc7a63c0192ddb3a1e7c1190f840d721cba417e78fb5f07c0b23494c455c.jpg", + "image_caption": [ + "Figure 7. Progressive visualization of the behavior of the proposed REAL framework throughout training. $A$ is the artificial label, and $Y$ denotes the subsequent label updated at timestep $t$ . The highlighted boxes indicate newly explored regions by artificial oracle, which were unknown (red, blue) or wrongly labeled (green) in earlier timesteps." + ], + "image_footnote": [], + "bbox": [ + 102, + 87, + 851, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rics of the initial labels and those of the labels enhanced by PP2S. Furthermore, we compare the performance of the point cloud semantic segmentation model when trained with the original sparse annotations and artificial labels enhanced by PP2S. The results demonstrate that our PP2S significantly improves the quality of the label, consequently boosting the model's performance. This analysis supports the validity of PP2S as a bridge for transferring the segmentation capability of SAM into the realm of weakly supervised point cloud semantic segmentation.", + "bbox": [ + 75, + 428, + 470, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query sampling strategy. We conduct ablation studies regarding the proposed query sampling strategy introduced in Section 3.4. In this section, we compare two settings: (I) using the proposed strategy sampling from the narrowed mother set of Equ. 10 and (II) sampling the most reliable point among every point without any constraint. Figure 6 provides both qualitative and quantitative comparisons between (I) and (II), displaying the improvement achieved by the artificial labels of each setting. (I) outperforms (II) in terms of recall while maintaining precision. This distinction primarily arises from the fact that most query points of (II) are already labeled in $Y^{t-1}$ , limiting their potential for exploring into unknown regions. On the other hand, the query of (I) is explicitly restricted to represent genuinely novel information compared with previous labels $Y^{t-1}$ , even without compromising precision. This implies that the proposed query sampling strategy yields better labels, facilitating the model to learn segmentation effectively. Accordingly, the performance with our query sampling strategy (62.7%) is significantly higher than without using it (59.8%). Additional results are in Supp.", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Active label enhancement. To demonstrate the behavior of the proposed REAL framework intuitively, Fig. 7 depicts progressive visualization throughout training. The labels are significantly improved as training proceeds, thanks to the artificial labels from our artificial oracle. Notably, the highlighted boxes indicate newly explored regions, which were unknown (red, blue) or wrongly labeled (green) by the labels in earlier timesteps. The REAL framework effectively integrates these regions into the current label.", + "bbox": [ + 496, + 428, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 498, + 574, + 802, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with SoTA methods on S3DIS, as shown in Table 2. The proposed REAL significantly enhances performance across all settings. Specifically, REAL with PTv2 [37] outperforms CPCM by $2.9\\%$ in mIoU under the $0.02\\%$ setting, which is even comparable to the SoTA results of the $1\\%$ setting. Furthermore, REAL with PTv2 surpasses the other SoTA methods under the $0.02\\%$ setting, using only 1pt for each object $(0.004\\%)$ . We also observe a significant performance improvement when employing the REAL in Closer [25]. These results imply that the gain of REAL is not restricted by the backbone.", + "bbox": [ + 496, + 598, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, Fig. 8 provides a qualitative comparison. Notably, the semantic confusion in the baseline is remarkably improved in ours. Furthermore, REAL demonstrates a meaningful enhancement in its ability to group the points of each object precisely. Consequently, REAL achieves remarkable semantic segmentation results even in complicated scenes, using only $0.004\\%$ of annotations.", + "bbox": [ + 496, + 763, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we evaluate the performance of our method against SoTA methods on ScanNetV2 [8], as detailed in Ta", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3727", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ea0e42bfd15338e2df459265b27929a742d14610a0b5b4ef7ad97cbb1c34360.jpg", + "table_caption": [ + "Table 2. Comparisons of the proposed methods with SoTA methods on S3DIS test set. Every metric is in mIoU (%). Bold number represents the best result. † denotes the results from our reimplementation." + ], + "table_footnote": [], + "table_body": "
SettingsMethodsmIoUceil.floorwallbeamcol.wind.doorchairtablebook.sofaboardclutter
FullyRandLA [14]62.491.295.780.10.025.262.347.475.883.260.870.865.254.0
KPConv [35]67.192.897.382.40.023.958.069.091.081.575.375.466.758.9
RFCR [10]68.794.298.384.30.028.562.471.292.082.676.171.171.661.3
Closer† [25]66.093.298.281.20.024.849.969.390.480.472.074.668.055.8
PTv2† [38]70.993.898.586.10.029.560.578.093.281.876.679.283.061.7
1%SPT [47]61.891.596.980.60.018.258.147.275.885.765.368.965.050.2
PSD [48]63.592.397.780.70.027.856.262.578.784.163.170.458.953.2
HybridCR [23]65.392.593.982.60.024.264.463.278.381.769.074.468.256.5
0.03%PSD [48]48.287.996.062.10.020.649.340.955.161.943.950.727.331.1
HybridCR [23]51.585.491.965.90.018.051.434.263.878.352.459.629.939.0
0.02%MIL [42]51.486.693.275.00.029.345.346.760.562.356.547.533.732.2
CPCM [24]62.392.695.679.40.017.849.359.485.775.669.160.768.255.8
PTv2† [38]55.180.594.773.90.021.742.938.981.459.262.552.966.142.1
+PP2S60.079.093.177.50.131.549.761.470.073.769.956.971.045.6
+REAL65.286.995.880.20.027.060.377.872.379.970.864.777.454.1
0.004%Closer† [25]38.782.592.469.40.115.023.335.650.947.80.319.137.329.1
+PP2S44.962.988.867.90.011.729.244.746.463.253.525.750.738.6
+REAL57.879.097.770.80.124.450.460.184.564.664.866.839.948.7
PTv2† [38]39.776.687.265.40.09.430.622.361.750.341.714.622.434.1
+PP2S53.065.993.168.90.027.342.755.868.463.758.252.550.541.8
+REAL62.784.095.780.30.031.157.463.775.576.968.770.065.446.8
", + "bbox": [ + 78, + 121, + 890, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de220b6e7ad44dc8499b83f51eeacbfe0570ec27bd784535fe63ec087127cb53.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons between the semantic segmentation results of the baseline and ours on S3DIS under $0.004\\%$ setting. The displayed boxes indicate the regions where significant improvements are achieved by the proposed method." + ], + "image_footnote": [], + "bbox": [ + 81, + 402, + 462, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble 3. Our method demonstrates superior performance on both the validation and test splits. Notably, REAL outperforms CPCM by $3.6\\%$ in the validation set and $2.7\\%$ in the test set. These results imply that our method effectively utilizes the potential of weak labels. The qualitative results for ScanNetV2 can be found in Supp.", + "bbox": [ + 76, + 603, + 468, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 708, + 194, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Expensive and labor-intensive point-wise annotation hinders the practical application of point cloud semantic segmentation. While conventional WSPCSS methods have exhibited promising results, the fixed input sparse label potentially constrains the achievable performance. In pursuit of a novel research direction, we propose the utilization of an additional source of information, introducing the Region Exploration via Artificial Labeling (REAL) framework. This framework aims to harness the power of the image foundational model, SAM, within the context of active learning. To establish the connection between the 2D model", + "bbox": [ + 76, + 734, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f539e440992a0f7893e31922b83d38fb99152b66684ac493a50097bc3fb74c5d.jpg", + "table_caption": [ + "Table 3. Comparisons of the proposed methods with SoTA methods on ScanNetV2 val/test set. The results are in mIoU." + ], + "table_footnote": [], + "table_body": "
SettingsMethodsVal.Test
FullyKPConv [35]-68.4
RFCR [10]-70.2
1%SPT [47]-51.1
PSD [48]-54.7
HybridCR [23]56.956.8
20 ptsMIL [42]57.854.4
CPCM [24]62.762.8
Ours (REAL)66.365.5
", + "bbox": [ + 560, + 435, + 834, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and 3D data, we designed the Projection-based Point-to-Segment (PP2S) module. The PP2S is employed as an artificial oracle in the REAL framework, dynamically enhancing labels during training. Although using an artificial oracle eliminates the burden of manual annotation, it introduces some drawbacks due to the model's imperfections. As a remedy, we devised several strategies for query sampling and label updating. Through extensive experiments, we demonstrated the working logic of our framework in detail. Finally, the proposed REAL framework achieved new SoTA on various datasets, surpassing existing works. We believe that our approach pioneers the effective integration of the foundational image model into the realm of WSPCSS.", + "bbox": [ + 496, + 578, + 890, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3728", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 2", + "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4981-4990, 2018. 2", + "[3] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 3, 6", + "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 3", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3", + "[6] Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5291-5301, 2023. 2", + "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 3, 6, 7", + "[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2", + "[10] Jingyu Gong, Jiachen Xu, Xin Tan, Haichuan Song, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Omni-supervised point cloud segmentation via gradual receptive field component reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11673-11682, 2021. 1, 8", + "[11] Shizhan Gong, Yuan Zhong, Wenao Ma, Jinpeng Li, Zhao Wang, Jingyang Zhang, Pheng-Ann Heng, and Qi Dou. 3dsam-adapter: Holistic adaptation of sam from 2d to 3d" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for promptable medical image segmentation. arXiv preprint arXiv:2306.13465, 2023. 2", + "[12] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 2", + "[13] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2", + "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 8", + "[15] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In European Conference on Computer Vision, pages 600-619. Springer, 2022. 1, 2", + "[16] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021. 2", + "[17] Zeyu Hu, Xuyang Bai, Runze Zhang, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Lidal: Inter-frame uncertainty based active learning for 3d lidar semantic segmentation. In European Conference on Computer Vision, pages 248-265. Springer, 2022. 2", + "[18] Jihun Kim, Hyeokjun Kwon, Yunseo Yang, and Kuk-Jin Yoon. Learning point cloud completion without complete point clouds: A pose-aware approach. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14157-14167. IEEE, 2023. 2", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2", + "[20] Hyeokjun Kweon and Kuk-Jin Yoon. Joint learning of 2d-3d weakly supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:30499–30511, 2022. 2", + "[21] Hyeokjun Kweon, Sung-Hoon Yoon, Hyeonseong Kim, Daehe Park, and Kuk-Jin Yoon. Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6994-7003, 2021. 2", + "[22] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 2" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3729", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14930-14939, 2022. 1, 2, 6, 8", + "[24] Lizhao Liu, Zhuangwei Zhuang, Shangxin Huang, Xunlong Xiao, Tianhang Xiang, Cen Chen, Jingdong Wang, and Mingkui Tan. Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18413-18422, 2023. 1, 2, 6, 8", + "[25] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 326-342. Springer, 2020. 6, 7, 8", + "[26] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021. 1, 2", + "[27] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1190-1199, 2023. 2", + "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1", + "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1", + "[30] Damien Robert, Bruno Vallet, and Loic Landrieu. Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5575-5584, 2022. 2", + "[31] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 2", + "[32] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022. 2", + "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceed-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3", + "[34] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2", + "[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 8", + "[36] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2", + "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 6, 7", + "[38] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 8", + "[39] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2", + "[40] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020. 1, 2", + "[41] Zongyi Xu, Bo Yuan, Shanshan Zhao, Qianni Zhang, and Xinbo Gao. Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18098-18108, 2023. 2", + "[42] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022. 1, 2, 6, 8", + "[43] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2d-3d interlaced transformer for point cloud segmentation with scene-level supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 977–987, 2023. 2", + "[44] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 2", + "[45] Sung-Hoon Yoon, Hyeokjun Kweon, Jegyeong Cho, Shinjeong Kim, and Kuk-Jin Yoon. Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3730", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "October 23-27, 2022, Proceedings, Part XXIX, pages 326-344. Springer Nature Switzerland Cham, 2022. 2", + "[46] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: Zero-shot 3d object detection via segment anything model. arXiv preprint arXiv:2306.02245, 2023. 2", + "[47] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3421-3429, 2021. 1, 2, 8", + "[48] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15520-15528, 2021. 1, 2, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3731", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_model.json b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab83d629b594cd348939e69e43f6c25b7e9a03c9 --- /dev/null +++ b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.182, + 0.332, + 0.215 + ], + "angle": 0, + "content": "Hyeokjun Kweon* KAIST" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.22, + 0.348, + 0.233 + ], + "angle": 0, + "content": "0327june@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.182, + 0.537, + 0.215 + ], + "angle": 0, + "content": "Jihun Kim* KAIST" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.22, + 0.585, + 0.233 + ], + "angle": 0, + "content": "jihun1998@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.663, + 0.182, + 0.774, + 0.215 + ], + "angle": 0, + "content": "Kuk-Jin Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.635, + 0.22, + 0.796, + 0.233 + ], + "angle": 0, + "content": "kjyoon@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.312, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.473, + 0.633 + ], + "angle": 0, + "content": "Manual annotation of every point in a point cloud is a costly and labor-intensive process. While weakly supervised point cloud semantic segmentation (WSPCSS) with sparse annotation shows promise, the limited information from initial sparse labels can place an upper bound on performance. As a new research direction for WSPCSS, we propose a novel Region Exploration via Artificial Labeling (REAL) framework. It leverages a foundational image model as an artificial oracle within the active learning context, eliminating the need for manual annotation by a human oracle. To integrate the 2D model into the 3D domain, we first introduce a Projection-based Point-to-Segment (PP2S) module, designed to enable prompt segmentation of 3D data without additional training. The REAL framework samples query points based on model predictions and requests annotations from PP2S, dynamically refining labels and improving model training. Furthermore, to overcome several challenges of employing an artificial model as an oracle, we formulate effective query sampling and label updating strategies. Our comprehensive experiments and comparisons demonstrate that the REAL framework significantly outperforms existing methods across various benchmarks. The code is available at https://github.com/jihun1998/AO." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.208, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Point cloud semantic segmentation has been extensively researched for robotics or autonomous driving applications. While fully supervised approaches [7, 10, 14, 28, 29, 35] utilizing deep learning have demonstrated remarkable progress, the densely annotated point-wise labels required for these methods pose a significant challenge due to the labor-intensive and expensive annotation process. This obstacle hinders the practical application of existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In response to this challenge, Weakly Supervised Point Cloud Semantic Segmentation (WSPCSS) has gained attention, leveraging more affordable weak labels [15, 23, 24, 26, 40, 42, 47, 48]. Among various types of weak labels, sparse annotations have become a widely adopted setting, where only a tiny subset (e.g., \\(0.02\\%\\)) of the entire point cloud is" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.27, + 0.885, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.406, + 0.887, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Figure 1. Visualization of the proposed REAL framework. We tackle the challenge of WSPCSS, leveraging a foundational imagery model as an artificial oracle within the context of active learning. The below examples illustrate the evolving labels as training proceeds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.702 + ], + "angle": 0, + "content": "labeled. Existing WSPCSS works have proposed remedies focusing on two key aspects: 1) fully utilizing the limited information from sparse labels and 2) learning the features beneficial for segmentation from the data. The methods based on self-labeling use the regions confidently predicted by the model as pseudo-labels for training the model itself. Meanwhile, various studies have proposed self-supervised learning methods, such as consistency against data augmentation, contrastive learning, or masked modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.854 + ], + "angle": 0, + "content": "While these approaches have shown substantial results, the overall amount of information provided is constrained by the initial sparse annotation, potentially placing an upper limit on performance. In this paper, instead of solely focusing on how to exploit the given limited information, we embark on a new and more challenging research direction. Specifically, we propose a novel WSPCSS method using an additional publicly available source of information—leveraging advancements in the 2D domain with the Segment Anything Model (SAM) [19]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We begin by utilizing the promptable segmentation capability of SAM to enhance initially provided sparse annotations. To bridge the gap between the 2D model and 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3721" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "point cloud data, we introduce the Projection-based Point-to-Segment (PP2S) module. This module comprises the projection of weakly annotated points onto images, acquiring 2D masks using SAM, and back-projecting these masks into 3D, facilitating segmentation in the point cloud. The proposed PP2S-based preprocessing effectively improves label quality, resulting in significantly enhanced semantic segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.47, + 0.426 + ], + "angle": 0, + "content": "However, the labels preprocessed by PP2S remain incomplete due to potential noises from calibration errors or SAM ambiguity. To further enhance label quality during training, we propose the Region Exploration via Artificial Labeling (REAL) framework, utilizing PP2S as an artificial oracle within the context of active learning, as depicted in Fig. 1. The REAL framework samples query points according to model predictions and subsequently requests annotations from PP2S to dynamically enhance training labels. Here, unlike general active learning methods relying on a human oracle, our artificial oracle cannot directly assign the classes of the requested query. Therefore, we design specialized strategies for query sampling and label refinement to unlock the segmentation capability of the PP2S." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.593 + ], + "angle": 0, + "content": "Note that the main novelty of our approach lies in how we effectively integrate the benefits of SAM into the realm of WSPCSS. We analyze the REAL framework through extensive experiments, including quantitative and qualitative ablation studies. Furthermore, our method achieves new state-of-the-art (SoTA) results on S3DIS [3] and ScanNetV2 [8] under all the tested settings, surpassing the existing WSPCSS methods. Remarkably, even with the \\(0.004\\%\\) setting, our method outperforms conventional methods using the \\(0.02\\%\\) setting, underscoring the superiority of the proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.609, + 0.22, + 0.625 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.357, + 0.652 + ], + "angle": 0, + "content": "2.1. Weakly Supervised Approaches" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Originating from weakly supervised learning in imagery domain [2, 21, 22, 31, 39, 45], weakly supervised approach has emerged as a cost-effective approach to mitigate the challenges associated with acquiring fully labeled data [15, 18, 23, 24, 26, 40, 42, 47, 48]. Xu and Lee [40] first propose weakly supervised configurations involving annotations of \\(10\\%\\) of the data and a single point label for each category. PSD [48] introduces a perturbed branch and constraint self-distillation loss between the perturbed and original branches. OTOC [26] utilizes pseudo-label methods and iteratively updates pseudo-labels through a self-training strategy. MIL [42] proposes the transformer model derived from multiple instance learning and integrates adaptive global weighted pooling to their model. HybridCR [23] introduces an architecture that leverages pseudo-label methods and consistency regularization between the original and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "augmented branches. CPCM [24] introduces a region-wise masking strategy and contextual masked training method to integrate the benefit of the masked autoencoder. Despite the significant progress in WSPCSS, in this paper, we would like to explore a new and more challenging research direction. In this light, we suggest a novel WSPCSS method, using a publicly available additional source of information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.208, + 0.874, + 0.223 + ], + "angle": 0, + "content": "2.2. Point Cloud Segmentation Using 2D Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.518 + ], + "angle": 0, + "content": "There have been extensive efforts to achieve point cloud segmentation in an image-based manner as in [12, 16, 20, 30, 43]. With the rapid advancement in the 2D domain, multi-modal self-supervised approaches [1, 6, 13] and employing language-based model [9, 27, 34] have also been particularly prevalent. Recently, there has been a surge in attempts to address various point cloud tasks using the Segment Anything Model (SAM) [19] The zero-shot approach [44, 46], which directly applies SAM to the image corresponding to the point cloud, has demonstrated promising segmentation results. However, SAM's lack of explicit semantics limits its applicability to high-level perception tasks. Adaptation-based approaches [11] can make SAM directly learn 3D tasks; however, they necessitate dense ground truth data during the learning process, rendering them unsuitable for WSPCSS. To tackle the weakly supervised setting, this paper presents a novel WSPCSS method that effectively refines weak labels using the foundational 2D model as an artificial oracle." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.544 + ], + "angle": 0, + "content": "2.3. Active Learning for Point Cloud Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Active Learning (AL) aims to identify a subset of instances that an oracle manually annotates during training. As in [17, 32, 36, 41], AL has also been extensively explored for WSPCSS. Given that the REAL framework actively employs SAM as an artificial oracle, it may share a conceptual scheme with conventional AL-based methods to some extent. Nevertheless, note that our weakly-supervised setting significantly differs from them. Our framework solely relies on initial sparse labels and does not involve any additional manual annotation within the training loop." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.593, + 0.732 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.642, + 0.756 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Problem setting. We denote an input point cloud as a set of \\(N\\) points: \\(X = \\{x_{1},\\ldots ,x_{N}\\}\\). In typical fully-supervised point cloud semantic segmentation, the labels for all points are provided as \\(Y = \\{y_{1},\\dots ,y_{N}\\}\\) where \\(y_{i}\\in \\{1,\\dots ,C\\}\\) and \\(C\\) is the number of classes. On the other hand, in weakly-supervised point cloud semantic segmentation, only a tiny subset of the points is labeled, and the other points are unlabelled. We formulate this setting as \\((X,Y) = (X^{L},Y^{L})\\cup (X^{U},Y^{U})\\). Here, \\(X^{L} = \\{x_{1}^{L},\\ldots ,x_{n}^{L}\\}\\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.468, + 0.138 + ], + "angle": 0, + "content": "points labeled with sparse annotations \\( Y^{L} = \\{y_{1}^{L},\\dots ,y_{n}^{L}\\} \\) Conversely, \\( X^{U} = \\{x_{1}^{U},\\ldots ,x_{N - n}^{U}\\} \\) is not labeled, and we represent it as \\( Y^{U} = \\{y_{1}^{U},\\dots ,y_{N - n}^{U}\\} \\), where \\( y_{i}^{U} = 0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.468, + 0.211 + ], + "angle": 0, + "content": "Segment Anything Model. SAM is a foundational model for promptable segmentation, comprising 1) an image encoder, 2) a prompt encoder, and 3) a decoder predicting the mask using the embeddings from encoders. We denote the inference process of SAM as" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.218, + 0.468, + 0.235 + ], + "angle": 0, + "content": "\\[\nM ^ {2 D} = \\operatorname {S A M} (I; Z), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.243, + 0.468, + 0.288 + ], + "angle": 0, + "content": "where \\(I\\) and \\(Z\\) are an input image and prompt, respectively. \\(M^{2D}\\) denotes a set of the pixels of the predicted mask. In this paper, we use a point prompt (i.e., a pixel coordinate)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.295, + 0.381, + 0.311 + ], + "angle": 0, + "content": "3.2. Projection-based Point-to-Segment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.468, + 0.393 + ], + "angle": 0, + "content": "We begin by enhancing the initially given weak annotations using SAM. However, SAM is a vision foundation model designed and trained for 2D imagery. Therefore, to directly process 3D point cloud data using SAM, it is necessary to conduct additional adaptation or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.468, + 0.605 + ], + "angle": 0, + "content": "Instead, we propose to utilize image data that captures the same scene as the provided point cloud data. We believe that incorporating image data is not overly restrictive in practice. In fact, point cloud data is usually created from RGBD frames using Structure from Motion (SfM) [3, 8] or captured by LiDAR paired with RGBD cameras [4, 5, 33]. Hence, most point cloud datasets naturally contain RGB images, depth maps, and associated camera parameters. Under this philosophy, we introduce a Projection-based Point-to-Segment (PP2S) module, enabling promptable segmentation of 3D point cloud data using SAM via projection. It is noteworthy that our method utilizes image data during the training phase only and can perform semantic segmentation using input point cloud data alone during testing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.606, + 0.468, + 0.681 + ], + "angle": 0, + "content": "Figure 2 illustrates the visualization of PP2S. Formally, for each point cloud \\(X\\), we assume that the model can access data from \\(J\\) cameras: images, depth maps, and projection matrices. Initially, we project the set of weakly annotated points \\(X^{L} = \\{x_{i}^{L}\\}\\) onto the image \\(I_{j}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.687, + 0.468, + 0.717 + ], + "angle": 0, + "content": "\\[\np _ {i, j} = \\operatorname {p r o j} _ {j} x _ {i} ^ {L} \\text {w h e r e} i \\in \\{1, \\dots , n \\}, j \\in \\{1, \\dots , J \\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.719, + 0.468, + 0.809 + ], + "angle": 0, + "content": "where \\(\\mathrm{proj}_j\\) denotes the projection matrix, and \\(p_{i,j}\\) represents the 2D coordinate of the \\(i\\)th weakly annotated points on the \\(j\\)th image. After the projection, we discard points that fall outside the image boundaries. We also filter out occluded points using the depth map \\(D_j\\), though alternative visibility testing algorithms can be employed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Then, SAM performs segmentation using the coordinates of projected points as individual input prompts as" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.845, + 0.468, + 0.863 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} ^ {2 D} = \\operatorname {S A M} \\left(I _ {j}; p _ {i, j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.87, + 0.468, + 0.902 + ], + "angle": 0, + "content": "where \\( M_{i,j}^{2D} \\) is a set of the pixels of predicted 2D mask on \\( I_j \\), given input point prompt \\( p_{i,j} \\). Since \\( p_{i,j} \\) corresponds to" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.885, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.178, + 0.885, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Figure 2. Visualization of the proposed Projection-based Point-to-Segment (PP2S) module. The process begins by projecting the weak labels onto images through projection matrices. Using these projected points as point prompts, SAM predicts 2D masks, which are subsequently back-projected onto the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.413, + 0.891, + 0.46 + ], + "angle": 0, + "content": "\\(x_{i}^{L}\\) and is labeled as \\(y_{i}^{L}\\), the pixels in \\(M_{i,j}^{2D}\\) are also assigned to \\(y_{i}^{L}\\). As multiple cameras exist per a single point cloud, we conduct the above process for all images in parallel." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Subsequently, we back-project the acquired 2D masks to 3D coordinates. Initially, similar to Equ. 2, we project the point cloud onto the \\(j\\)th image as \\(p_{k,j} = \\mathrm{proj}_j x_k\\), where \\(x_k\\) is the \\(k\\)th point in the point cloud and \\(p_{k,j}\\) denotes the projection of \\(x_k\\). If \\(p_{k,j}\\) is an element of \\(M_{i,j}^{2D}\\), it means that the projection result of \\(x_k\\) belongs to the mask of the weakly annotated point \\(x_i^L\\), within the image \\(I_j\\). Therefore, we define the 3D mask \\(M_{i,j}^{3D}\\) as the set of \\(x_k\\) such that its projection \\(p_{k,j}\\) is in \\(M_{i,j}^{2D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.599, + 0.785, + 0.613 + ], + "angle": 0, + "content": "The above process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.621, + 0.891, + 0.642 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} ^ {3 D} = \\left\\{x _ {k} \\mid p _ {k, j} \\in M _ {i, j} ^ {2 D} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.651, + 0.891, + 0.699 + ], + "angle": 0, + "content": "where \\( M_{i,j}^{3D} \\) denotes the set of points predicted by SAM to be the same segment with \\( x_{i}^{L} \\) from the perspective of \\( j \\)th image. Accordingly, \\( y_{i,j}^{3D} \\), the class of \\( M_{i,j}^{3D} \\), is \\( y_{i}^{L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.834 + ], + "angle": 0, + "content": "However, the 3D masks are not mutually exclusive. Given that a single 3D point could be observed from multiple images, it may be shared by more than one 3D mask. To address this, we establish a consensus among the 3D masks through a mask-wise voting system. In our system, each 3D mask \\( M_{i,j}^{3D} \\) casts a vote for its points being set to \\( y_{i,j}^{3D} \\). Therefore, from the perspective of each point, its class is determined by the voting of the 3D masks that include the point. The above process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.843, + 0.891, + 0.862 + ], + "angle": 0, + "content": "\\[\nv _ {k} = \\left\\{y _ {i, j} ^ {3 D} \\right\\} \\forall i, j \\text {s u c h t h a t} x _ {k} \\in M _ {i, j} ^ {3 D}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\( v_{k} \\) is a multiset of the voted classes. Finally, we assign the class with the highest number of votes among \\( v_{k} \\) as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.093, + 0.46, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Figure 3. Possible sources of noise in PP2S. (A): The 3D door point is projected onto the chair in the image, resulting in an incorrect 2D mask. (B): Erroneous back-projection causes wall points in 3D to be mislabeled as chair. (C) Above: The 2D mask of the ceiling covers only a small portion of the entire object. (C) Below: The 2D mask of the wall encroaches into the regions of column." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.33, + 0.424 + ], + "angle": 0, + "content": "the artificial label \\(a_{k}\\) for \\(k\\)th point \\(x_{k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.436, + 0.469, + 0.477 + ], + "angle": 0, + "content": "\\[\na _ {k} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} v _ {k} = \\emptyset \\\\ \\operatorname {M o d e} \\left(v _ {k}\\right) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.468, + 0.503 + ], + "angle": 0, + "content": "where the Mode operator returns the most frequent element." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.469, + 0.624 + ], + "angle": 0, + "content": "The proposed PP2S module effectively leverages the segmentation capability of SAM to expand the sparsely annotated initial labels. These resulting artificial labels, denoted as \\( A = \\{a_k\\} \\), can directly serve as a supervised signal for training. Notably, we have observed that training with \\( A \\) yields significant performance improvements compared to using initial \\( Y \\). The detailed experimental results are demonstrated in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.437, + 0.652 + ], + "angle": 0, + "content": "3.3. Region Exploration via Artificial Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.9 + ], + "angle": 0, + "content": "As mentioned, the proposed PP2S effectively enhances the initial sparse annotations, resulting in a substantial improvement in segmentation performance. However, we also have recognized several limitations associated with this preprocessing approach. The PP2S-enhanced labels can serve as better guidance than the initial sparse labels. Nevertheless, they are still fixed during the entire training phase and thereby impose an upper bound on performance. Notably, we cannot guide the regions not covered by the 2D masks obtained within PP2S. Furthermore, even if we assume that the initially provided sparse annotations are perfectly correct, errors stemming from factors such as noises of camera projection matrices (Fig. 3 A, B) or inherent ambiguity of SAM (Fig. 3 C) are inevitable. While the voting system can mitigate some of these, it cannot completely eliminate them." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.86, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Region Exploring via Artificial Labeling" + }, + { + "type": "algorithm", + "bbox": [ + 0.502, + 0.111, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Require: Segmentation model \\(f_{\\theta}\\) , Training dataset \\(\\mathcal{D} =\\) \\(\\{(X,Y)\\}\\) , Learning rate \\(\\eta\\) , Number of epochs \\(T\\) \nEnsure: Optimized \\(f_{\\theta}\\) \n1: for \\((X,Y)\\) in \\(\\mathcal{D}\\) do \n2: \\(Y\\gets \\mathrm{PP2S}(X;X^L)\\) \n3: end for \n4: Initialize the model parameter \\(\\theta\\) \n5: for \\(t\\gets 1,2,\\ldots ,T\\) do \n6: for \\((X,Y)\\) in \\(\\mathcal{D}\\) do \n7: \\(Y^{t - 1}\\gets Y\\) \n8: \\(S\\gets \\mathrm{softmax}(f_{\\theta}(X))\\) \n9: \\(\\hat{Y}\\gets \\mathrm{argmax}(S)\\) \n10: \\(Q\\gets \\mathrm{QuerySampling}(X,S,Y^{t - 1})\\) \n11: \\(A\\gets \\mathrm{PP2S}(X;Q)\\) \n12: \\(Y^{t}\\gets \\mathrm{LabelUpdate}(Y^{t - 1},A,\\hat{Y})\\) \n13: \\(Y\\gets Y^{t}\\) \n14: \\(\\theta \\leftarrow \\theta -\\eta \\nabla_{\\theta}\\mathcal{L}(S,Y)\\) \n15: end for \n16: end for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Then, how can we further improve label quality, especially during training? We draw inspiration from the concept of active learning. In active labeling, the model identifies a set of instances, known as a query, during training and requests human annotators (i.e., oracle) to label these points. Typically, the query consists of instances that confuse the model at a given timestep, making annotating the query highly effective for the model to learn decision boundaries. However, active learning requires manual annotation within the training loop, potentially making it even more expensive than WSPCSS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.716 + ], + "angle": 0, + "content": "To address this challenge, we present the Region Exploration via Artificial Labeling (REAL) framework—a novel WSPCSS method utilizing the PP2S module as an artificial oracle, as depicted in Fig. 4. Specifically, the proposed PP2S module serves as an oracle in the REAL framework, responding to requests to annotate the query. Given that the artificial oracle is, in fact, a pretrained network, the REAL framework is completely relieved of the necessity for manual annotation, except for the initial sparse labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "However, unlike a human oracle in conventional methods, our artificial oracle lacks explicit semantic knowledge. Therefore, we cannot expect the PP2S module to directly label the classes for the confusing points. Instead, we focus on leveraging what PP2S can provide to us. PP2S cannot directly annotate the given points; however, it can precisely predict which points should be grouped into a segment. To fully utilize this attribute, we define the query \\( Q = \\{q_i\\} \\) to consist of the confidently identified points, instead of ambiguous ones from the perspective of the model. Trusting the prediction of the model on these points, we can self-label the query with minimal risk. We represent this pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3724" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.096, + 0.816, + 0.113 + ], + "angle": 0, + "content": "Region Exploration via Artificial Labeling (REAL, Sec. 3.3)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.117, + 0.887, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 4. Visualization of the proposed Region Exploration via Artificial Labeling (REAL) framework. The objective is to dynamically enhance the quality of labels for weakly supervised point cloud semantic segmentation. In the REAL framework, the model \\( f_{\\theta} \\) takes the point cloud \\( X \\) as input and predicts semantic segmentation logits \\( S \\). We then sample the most confident points from the unlabeled regions of the previous label \\( Y^{t-1} \\) based on the predictions \\( \\hat{Y} \\) and reliability scores \\( R \\). The sampled points then serve as prompts for the proposed PP2S module, producing artificial labels \\( A \\). Finally, by combining the model prediction and \\( A \\), we refine \\( Y^{t-1} \\) into the updated labels \\( Y^t \\) which serve as pseudo-ground truth for training the logits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.468, + 0.556 + ], + "angle": 0, + "content": "cess as QuerySampling in Algorithm 1, which will be thoroughly discussed in Section 3.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.469, + 0.618 + ], + "angle": 0, + "content": "Subsequently, we request the PP2S module to act as an artificial oracle, annotating the unknown regions using the given query. For this, the self-labeled query serves as an input prompt for the PP2S module as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.63, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {P P 2 S} (X; Q), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.748 + ], + "angle": 0, + "content": "where \\( A \\) represents the artificial label obtained by PP2S from the given query \\( Q \\). This approach is akin to using a set of sparsely annotated points \\( (X^L) \\) for PP2S in the preprocessing phase, as described in Section 3.2. The primary distinction in the REAL framework is that the input prompt is self-labeled, not manually annotated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.84 + ], + "angle": 0, + "content": "With the obtained artificial label \\( A \\), we update the label of the previous step, \\( Y^{t-1} \\), into \\( Y^t \\). However, \\( A \\) may not be entirely reliable, as it could be more prone to noise compared to labels provided by a human oracle. This potential for noise could arise from inaccuracies in the self-labeled input query or the issues highlighted in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To mitigate the risk of incorporating errors during the label-updating process, we devise a cautious approach using the prediction of the model. Specifically, if a discrepancy exists between the class assigned by the artificial label" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.572 + ], + "angle": 0, + "content": "\\(A\\) and the model's prediction \\(\\hat{Y}\\) for a specific point, we exclude that point from the updating process. This conservative update process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.582, + 0.891, + 0.623 + ], + "angle": 0, + "content": "\\[\ny _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} a _ {i} & \\text {i f} a _ {i} = \\hat {y} _ {i} \\\\ y _ {i} ^ {t - 1} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.632, + 0.89, + 0.663 + ], + "angle": 0, + "content": "Finally, the model \\( f_{\\theta} \\) is optimized by minimizing the cross-entropy loss between \\( \\hat{Y} \\) and \\( Y^{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.813, + 0.689 + ], + "angle": 0, + "content": "3.4. Query Sampling Strategy for REAL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.816 + ], + "angle": 0, + "content": "In the REAL framework, query sampling is a critical component, as the quality of the artificial label depends on the quality of the input query. We posit that an effective query must meet two key criteria: 1) it should exhibit a high level of confidence, and 2) it should aid the model in acquiring new knowledge. To address the first criterion, we design a margin-based reliability metric for evaluating the confidence of the model's prediction for each point as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.829, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\nr _ {i} = \\operatorname {M a x} \\left(s _ {i}\\right) - \\operatorname {M a x} 2 \\left(s _ {i}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\( s_i \\in \\mathbb{R}^C \\) is the predicted logit of the \\( i \\)th point \\( x_i \\), and \\( r_i \\) denotes its reliability. Max and Max2 are operators that return the maximum and the second maximum el" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3725" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.458, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.19, + 0.47, + 0.22 + ], + "angle": 0, + "content": "Figure 5. Verification of the reliability metric. The points presenting higher reliability show higher semantic segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.471, + 0.336 + ], + "angle": 0, + "content": "ements of the input vector, respectively. We verify a robust correlation between the devised reliability metric and the accuracy of model predictions, as illustrated in Fig. 5. For instance, among points with reliability scores exceeding 0.5, the model correctly predicts the class for \\(84.5\\%\\) of them. We set the threshold for query sampling at 0.95, corresponding to around \\(94\\%\\) accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.472 + ], + "angle": 0, + "content": "On the contrary, directly quantifying the utility of a query for the second criterion is inherently elusive. In this context, we propose an intuitive hypothesis: if the model has successfully leveraged the labels from the previous timestep, \\( Y^{t-1} \\), the regions previously annotated in \\( Y^{t-1} \\) are expected to contribute less novel information during the current timestep, \\( t \\). Therefore, the focus of new queries should be on effectively exploring the regions not previously or incorrectly labeled in \\( Y^{t-1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.517 + ], + "angle": 0, + "content": "To this end, we narrow the sampling range of the query to the points that are predicted by the model differently from the previous label. We formulate the above constraint as" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.524, + 0.47, + 0.543 + ], + "angle": 0, + "content": "\\[\nZ _ {c} = \\left\\{x _ {i} \\right\\} \\forall i \\text {s u c h t h a t} \\hat {y} _ {i} = c \\text {a n d} y _ {i} ^ {t - 1} \\neq c. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Here, \\( Z_{c} \\) is the mother set for sampling \\( q_{c} \\), the query point of the class \\( c \\). Considering the first criterion, we sample \\( q_{c} \\) as the most reliable point among \\( Z_{c} \\), and \\( Q \\) is the set of \\( q_{c} \\). We formulate the above processes as" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.616, + 0.47, + 0.635 + ], + "angle": 0, + "content": "\\[\nQ = \\operatorname {Q u e r y S a m p l i n g} (X, S, Y ^ {t - 1}). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.21, + 0.659 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.282, + 0.683 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on S3DIS [3] and ScanNetV2 [8]. S3DIS comprises 6 areas with 272 rooms and 13 categories. We use the area 5 for evaluation following previous studies. ScanNetV2 includes 1613 3D scenes with 20 categories. We follow the official split (1201 training, 312 validation, and 100 test scenes). In addition to the point cloud data, the proposed REAL framework requires the data from cameras (i.e., images, depth maps, and camera projection matrices) to leverage SAM. The datasets inherently include such data since they are reconstructed from the sequence of RGBD frames. During training, we utilize the data from 48 and 17 cameras per scene on average for S3DIS and ScanNetV2, respectively. Note that our framework does not require 2D data at inference time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.09, + 0.893, + 0.173 + ], + "angle": 0, + "content": "Table 1. A quantitative comparison between the initially sparse annotations (Initial) and the artificial labels from PP2S (PP2S). Pre, Rec, and # denote precision, recall, and the proportion of annotated points relative to the total number of points. The performance of the point cloud semantic segmentation model trained by each model is represented as Model mIoU. Every metric is in %." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.178, + 0.866, + 0.219 + ], + "angle": 0, + "content": "
LabelsmIoUPreRec#Model mIoU
Initial0.0041000.0040.00439.7
PP2S30.483.132.438.453.0
" + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.225, + 0.65, + 0.236 + ], + "angle": 0, + "content": "(A) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.237, + 0.693, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.328, + 0.686, + 0.34 + ], + "angle": 0, + "content": "(C) Explored regions (setting 1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.225, + 0.856, + 0.237 + ], + "angle": 0, + "content": "(B) Labels of the previous step" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.238, + 0.866, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.328, + 0.862, + 0.34 + ], + "angle": 0, + "content": "(D) Explored regions (setting II)" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.34, + 0.693, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.431, + 0.597, + 0.44 + ], + "angle": 0, + "content": "Prec: -0.09%" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.431, + 0.681, + 0.439 + ], + "angle": 0, + "content": "Rec: +11.92%" + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.431, + 0.77, + 0.439 + ], + "angle": 0, + "content": "Prec: -0.03%" + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.431, + 0.851, + 0.439 + ], + "angle": 0, + "content": "Rec: +5.58%" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.451, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 6. Validation of our query sampling strategy. Setting I samples the queries following our strategy while setting II samples from the entire point set. We present the improvement obtained by artificial labels in terms of precision (Prec) and recall (Rec)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.791 + ], + "angle": 0, + "content": "Evaluation protocols. Previous studies [23, 24, 42, 48] have conducted experiments with various levels of sparsity. We believe that the most cost-efficient yet practical setting involves labeling only one point per object. Accordingly, our experiments mainly target the 1pt weak setting, which corresponds to \\(0.004\\%\\) of points in S3DIS. For a fair comparison, we also conduct experiments under the \\(0.02\\%\\) setting, which is the most widely used. For ScanNetV2, we follow the conventional works, using 20pt per scene. The overall performance is assessed on all points within the test set. Following the standard practice, we employ the mean Intersection over Union (mIoU) as our quantitative metric. Implementation details. We choose PTv2 [37] network as our main backbone. To verify the effect of the backbone, we additionally test Closer [25] backbone for the \\(0.004\\%\\) setting on S3DIS. Training details are the same as standard settings of the backbones. We employ a pretrained ViT-H model for SAM. Further details are in the Supp." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.681, + 0.818 + ], + "angle": 0, + "content": "4.2. Analysis on REAL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.893, + 0.903 + ], + "angle": 0, + "content": "PP2S-based preprocessing. We first validate the preprocessing strategy built upon the proposed PP2S. This strategy aims to enhance the initial sparse annotations into artificial labels by harnessing the promptable segmentation capability of SAM via projection. Table 1 provides various met" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.088, + 0.852, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 7. Progressive visualization of the behavior of the proposed REAL framework throughout training. \\( A \\) is the artificial label, and \\( Y \\) denotes the subsequent label updated at timestep \\( t \\). The highlighted boxes indicate newly explored regions by artificial oracle, which were unknown (red, blue) or wrongly labeled (green) in earlier timesteps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.471, + 0.581 + ], + "angle": 0, + "content": "rics of the initial labels and those of the labels enhanced by PP2S. Furthermore, we compare the performance of the point cloud semantic segmentation model when trained with the original sparse annotations and artificial labels enhanced by PP2S. The results demonstrate that our PP2S significantly improves the quality of the label, consequently boosting the model's performance. This analysis supports the validity of PP2S as a bridge for transferring the segmentation capability of SAM into the realm of weakly supervised point cloud semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Query sampling strategy. We conduct ablation studies regarding the proposed query sampling strategy introduced in Section 3.4. In this section, we compare two settings: (I) using the proposed strategy sampling from the narrowed mother set of Equ. 10 and (II) sampling the most reliable point among every point without any constraint. Figure 6 provides both qualitative and quantitative comparisons between (I) and (II), displaying the improvement achieved by the artificial labels of each setting. (I) outperforms (II) in terms of recall while maintaining precision. This distinction primarily arises from the fact that most query points of (II) are already labeled in \\( Y^{t-1} \\), limiting their potential for exploring into unknown regions. On the other hand, the query of (I) is explicitly restricted to represent genuinely novel information compared with previous labels \\( Y^{t-1} \\), even without compromising precision. This implies that the proposed query sampling strategy yields better labels, facilitating the model to learn segmentation effectively. Accordingly, the performance with our query sampling strategy (62.7%) is significantly higher than without using it (59.8%). Additional results are in Supp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Active label enhancement. To demonstrate the behavior of the proposed REAL framework intuitively, Fig. 7 depicts progressive visualization throughout training. The labels are significantly improved as training proceeds, thanks to the artificial labels from our artificial oracle. Notably, the highlighted boxes indicate newly explored regions, which were unknown (red, blue) or wrongly labeled (green) by the labels in earlier timesteps. The REAL framework effectively integrates these regions into the current label." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.575, + 0.803, + 0.592 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.764 + ], + "angle": 0, + "content": "We compare our method with SoTA methods on S3DIS, as shown in Table 2. The proposed REAL significantly enhances performance across all settings. Specifically, REAL with PTv2 [37] outperforms CPCM by \\(2.9\\%\\) in mIoU under the \\(0.02\\%\\) setting, which is even comparable to the SoTA results of the \\(1\\%\\) setting. Furthermore, REAL with PTv2 surpasses the other SoTA methods under the \\(0.02\\%\\) setting, using only 1pt for each object \\((0.004\\%)\\). We also observe a significant performance improvement when employing the REAL in Closer [25]. These results imply that the gain of REAL is not restricted by the backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Besides, Fig. 8 provides a qualitative comparison. Notably, the semantic confusion in the baseline is remarkably improved in ours. Furthermore, REAL demonstrates a meaningful enhancement in its ability to group the points of each object precisely. Consequently, REAL achieves remarkable semantic segmentation results even in complicated scenes, using only \\(0.004\\%\\) of annotations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we evaluate the performance of our method against SoTA methods on ScanNetV2 [8], as detailed in Ta" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3727" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Comparisons of the proposed methods with SoTA methods on S3DIS test set. Every metric is in mIoU (%). Bold number represents the best result. † denotes the results from our reimplementation." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.122, + 0.892, + 0.398 + ], + "angle": 0, + "content": "
SettingsMethodsmIoUceil.floorwallbeamcol.wind.doorchairtablebook.sofaboardclutter
FullyRandLA [14]62.491.295.780.10.025.262.347.475.883.260.870.865.254.0
KPConv [35]67.192.897.382.40.023.958.069.091.081.575.375.466.758.9
RFCR [10]68.794.298.384.30.028.562.471.292.082.676.171.171.661.3
Closer† [25]66.093.298.281.20.024.849.969.390.480.472.074.668.055.8
PTv2† [38]70.993.898.586.10.029.560.578.093.281.876.679.283.061.7
1%SPT [47]61.891.596.980.60.018.258.147.275.885.765.368.965.050.2
PSD [48]63.592.397.780.70.027.856.262.578.784.163.170.458.953.2
HybridCR [23]65.392.593.982.60.024.264.463.278.381.769.074.468.256.5
0.03%PSD [48]48.287.996.062.10.020.649.340.955.161.943.950.727.331.1
HybridCR [23]51.585.491.965.90.018.051.434.263.878.352.459.629.939.0
0.02%MIL [42]51.486.693.275.00.029.345.346.760.562.356.547.533.732.2
CPCM [24]62.392.695.679.40.017.849.359.485.775.669.160.768.255.8
PTv2† [38]55.180.594.773.90.021.742.938.981.459.262.552.966.142.1
+PP2S60.079.093.177.50.131.549.761.470.073.769.956.971.045.6
+REAL65.286.995.880.20.027.060.377.872.379.970.864.777.454.1
0.004%Closer† [25]38.782.592.469.40.115.023.335.650.947.80.319.137.329.1
+PP2S44.962.988.867.90.011.729.244.746.463.253.525.750.738.6
+REAL57.879.097.770.80.124.450.460.184.564.664.866.839.948.7
PTv2† [38]39.776.687.265.40.09.430.622.361.750.341.714.622.434.1
+PP2S53.065.993.168.90.027.342.755.868.463.758.252.550.541.8
+REAL62.784.095.780.30.031.157.463.775.576.968.770.065.446.8
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.404, + 0.464, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.537, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons between the semantic segmentation results of the baseline and ours on S3DIS under \\(0.004\\%\\) setting. The displayed boxes indicate the regions where significant improvements are achieved by the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.604, + 0.47, + 0.695 + ], + "angle": 0, + "content": "ble 3. Our method demonstrates superior performance on both the validation and test splits. Notably, REAL outperforms CPCM by \\(3.6\\%\\) in the validation set and \\(2.7\\%\\) in the test set. These results imply that our method effectively utilizes the potential of weak labels. The qualitative results for ScanNetV2 can be found in Supp." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.196, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Expensive and labor-intensive point-wise annotation hinders the practical application of point cloud semantic segmentation. While conventional WSPCSS methods have exhibited promising results, the fixed input sparse label potentially constrains the achievable performance. In pursuit of a novel research direction, we propose the utilization of an additional source of information, introducing the Region Exploration via Artificial Labeling (REAL) framework. This framework aims to harness the power of the image foundational model, SAM, within the context of active learning. To establish the connection between the 2D model" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Table 3. Comparisons of the proposed methods with SoTA methods on ScanNetV2 val/test set. The results are in mIoU." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.436, + 0.835, + 0.569 + ], + "angle": 0, + "content": "
SettingsMethodsVal.Test
FullyKPConv [35]-68.4
RFCR [10]-70.2
1%SPT [47]-51.1
PSD [48]-54.7
HybridCR [23]56.956.8
20 ptsMIL [42]57.854.4
CPCM [24]62.762.8
Ours (REAL)66.365.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.775 + ], + "angle": 0, + "content": "and 3D data, we designed the Projection-based Point-to-Segment (PP2S) module. The PP2S is employed as an artificial oracle in the REAL framework, dynamically enhancing labels during training. Although using an artificial oracle eliminates the burden of manual annotation, it introduces some drawbacks due to the model's imperfections. As a remedy, we devised several strategies for query sampling and label updating. Through extensive experiments, we demonstrated the working logic of our framework in detail. Finally, the proposed REAL framework achieved new SoTA on various datasets, surpassing existing works. We believe that our approach pioneers the effective integration of the foundational image model into the realm of WSPCSS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3728" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4981-4990, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[3] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.563 + ], + "angle": 0, + "content": "[6] Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5291-5301, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.706, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[10] Jingyu Gong, Jiachen Xu, Xin Tan, Haichuan Song, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Omni-supervised point cloud segmentation via gradual receptive field component reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11673-11682, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[11] Shizhan Gong, Yuan Zhong, Wenao Ma, Jinpeng Li, Zhao Wang, Jingyang Zhang, Pheng-Ann Heng, and Qi Dou. 3dsam-adapter: Holistic adaptation of sam from 2d to 3d" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "for promptable medical image segmentation. arXiv preprint arXiv:2306.13465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[12] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[13] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[15] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In European Conference on Computer Vision, pages 600-619. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[16] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[17] Zeyu Hu, Xuyang Bai, Runze Zhang, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Lidal: Inter-frame uncertainty based active learning for 3d lidar semantic segmentation. In European Conference on Computer Vision, pages 248-265. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[18] Jihun Kim, Hyeokjun Kwon, Yunseo Yang, and Kuk-Jin Yoon. Learning point cloud completion without complete point clouds: A pose-aware approach. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14157-14167. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[20] Hyeokjun Kweon and Kuk-Jin Yoon. Joint learning of 2d-3d weakly supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:30499–30511, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[21] Hyeokjun Kweon, Sung-Hoon Yoon, Hyeonseong Kim, Daehe Park, and Kuk-Jin Yoon. Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6994-7003, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[22] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3729" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[23] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14930-14939, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[24] Lizhao Liu, Zhuangwei Zhuang, Shangxin Huang, Xunlong Xiao, Tianhang Xiang, Cen Chen, Jingdong Wang, and Mingkui Tan. Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18413-18422, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.345 + ], + "angle": 0, + "content": "[25] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 326-342. Springer, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[26] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.501 + ], + "angle": 0, + "content": "[27] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1190-1199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[30] Damien Robert, Bruno Vallet, and Loic Landrieu. Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5575-5584, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[31] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[32] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[34] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[36] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[38] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[39] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[40] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[41] Zongyi Xu, Bo Yuan, Shanshan Zhao, Qianni Zhang, and Xinbo Gao. Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18098-18108, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[42] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[43] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2d-3d interlaced transformer for point cloud segmentation with scene-level supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 977–987, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[44] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[45] Sung-Hoon Yoon, Hyeokjun Kweon, Jegyeong Cho, Shinjeong Kim, and Kuk-Jin Yoon. Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "October 23-27, 2022, Proceedings, Part XXIX, pages 326-344. Springer Nature Switzerland Cham, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[46] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: Zero-shot 3d object detection via segment anything model. arXiv preprint arXiv:2306.02245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[47] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3421-3429, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[48] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15520-15528, 2021. 1, 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3731" + } + ] +] \ No newline at end of file diff --git a/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_origin.pdf b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4fd5b6519e40c05b57d4d6f84f33113997732e86 --- /dev/null +++ b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/310f76a4-fe34-49c0-9b81-fed65524e491_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73748906846b193e2db84ed82966bb8b23e0257709af5550d4f653475e804433 +size 4805089 diff --git a/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/full.md b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2c4835901d5a740bc31184585bdf64b6257664f --- /dev/null +++ b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/full.md @@ -0,0 +1,344 @@ +# Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle + +Hyeokjun Kweon* KAIST + +0327june@kaist.ac.kr + +Jihun Kim* KAIST + +jihun1998@kaist.ac.kr + +Kuk-Jin Yoon KAIST + +kjyoon@kaist.ac.kr + +# Abstract + +Manual annotation of every point in a point cloud is a costly and labor-intensive process. While weakly supervised point cloud semantic segmentation (WSPCSS) with sparse annotation shows promise, the limited information from initial sparse labels can place an upper bound on performance. As a new research direction for WSPCSS, we propose a novel Region Exploration via Artificial Labeling (REAL) framework. It leverages a foundational image model as an artificial oracle within the active learning context, eliminating the need for manual annotation by a human oracle. To integrate the 2D model into the 3D domain, we first introduce a Projection-based Point-to-Segment (PP2S) module, designed to enable prompt segmentation of 3D data without additional training. The REAL framework samples query points based on model predictions and requests annotations from PP2S, dynamically refining labels and improving model training. Furthermore, to overcome several challenges of employing an artificial model as an oracle, we formulate effective query sampling and label updating strategies. Our comprehensive experiments and comparisons demonstrate that the REAL framework significantly outperforms existing methods across various benchmarks. The code is available at https://github.com/jihun1998/AO. + +# 1. Introduction + +Point cloud semantic segmentation has been extensively researched for robotics or autonomous driving applications. While fully supervised approaches [7, 10, 14, 28, 29, 35] utilizing deep learning have demonstrated remarkable progress, the densely annotated point-wise labels required for these methods pose a significant challenge due to the labor-intensive and expensive annotation process. This obstacle hinders the practical application of existing methods. + +In response to this challenge, Weakly Supervised Point Cloud Semantic Segmentation (WSPCSS) has gained attention, leveraging more affordable weak labels [15, 23, 24, 26, 40, 42, 47, 48]. Among various types of weak labels, sparse annotations have become a widely adopted setting, where only a tiny subset (e.g., $0.02\%$ ) of the entire point cloud is + +![](images/7276b774f940b61be70e989ed1b975016f2a7e9949e50e1f5473060fb200d42d.jpg) + +![](images/8f338933edb540490bbef37c53cd2dfcea9d95ddca585a919971bae58ef1bce7.jpg) +Figure 1. Visualization of the proposed REAL framework. We tackle the challenge of WSPCSS, leveraging a foundational imagery model as an artificial oracle within the context of active learning. The below examples illustrate the evolving labels as training proceeds. + +labeled. Existing WSPCSS works have proposed remedies focusing on two key aspects: 1) fully utilizing the limited information from sparse labels and 2) learning the features beneficial for segmentation from the data. The methods based on self-labeling use the regions confidently predicted by the model as pseudo-labels for training the model itself. Meanwhile, various studies have proposed self-supervised learning methods, such as consistency against data augmentation, contrastive learning, or masked modeling. + +While these approaches have shown substantial results, the overall amount of information provided is constrained by the initial sparse annotation, potentially placing an upper limit on performance. In this paper, instead of solely focusing on how to exploit the given limited information, we embark on a new and more challenging research direction. Specifically, we propose a novel WSPCSS method using an additional publicly available source of information—leveraging advancements in the 2D domain with the Segment Anything Model (SAM) [19]. + +We begin by utilizing the promptable segmentation capability of SAM to enhance initially provided sparse annotations. To bridge the gap between the 2D model and 3D + +point cloud data, we introduce the Projection-based Point-to-Segment (PP2S) module. This module comprises the projection of weakly annotated points onto images, acquiring 2D masks using SAM, and back-projecting these masks into 3D, facilitating segmentation in the point cloud. The proposed PP2S-based preprocessing effectively improves label quality, resulting in significantly enhanced semantic segmentation performance. + +However, the labels preprocessed by PP2S remain incomplete due to potential noises from calibration errors or SAM ambiguity. To further enhance label quality during training, we propose the Region Exploration via Artificial Labeling (REAL) framework, utilizing PP2S as an artificial oracle within the context of active learning, as depicted in Fig. 1. The REAL framework samples query points according to model predictions and subsequently requests annotations from PP2S to dynamically enhance training labels. Here, unlike general active learning methods relying on a human oracle, our artificial oracle cannot directly assign the classes of the requested query. Therefore, we design specialized strategies for query sampling and label refinement to unlock the segmentation capability of the PP2S. + +Note that the main novelty of our approach lies in how we effectively integrate the benefits of SAM into the realm of WSPCSS. We analyze the REAL framework through extensive experiments, including quantitative and qualitative ablation studies. Furthermore, our method achieves new state-of-the-art (SoTA) results on S3DIS [3] and ScanNetV2 [8] under all the tested settings, surpassing the existing WSPCSS methods. Remarkably, even with the $0.004\%$ setting, our method outperforms conventional methods using the $0.02\%$ setting, underscoring the superiority of the proposed approach. + +# 2. Related Work + +# 2.1. Weakly Supervised Approaches + +Originating from weakly supervised learning in imagery domain [2, 21, 22, 31, 39, 45], weakly supervised approach has emerged as a cost-effective approach to mitigate the challenges associated with acquiring fully labeled data [15, 18, 23, 24, 26, 40, 42, 47, 48]. Xu and Lee [40] first propose weakly supervised configurations involving annotations of $10\%$ of the data and a single point label for each category. PSD [48] introduces a perturbed branch and constraint self-distillation loss between the perturbed and original branches. OTOC [26] utilizes pseudo-label methods and iteratively updates pseudo-labels through a self-training strategy. MIL [42] proposes the transformer model derived from multiple instance learning and integrates adaptive global weighted pooling to their model. HybridCR [23] introduces an architecture that leverages pseudo-label methods and consistency regularization between the original and + +augmented branches. CPCM [24] introduces a region-wise masking strategy and contextual masked training method to integrate the benefit of the masked autoencoder. Despite the significant progress in WSPCSS, in this paper, we would like to explore a new and more challenging research direction. In this light, we suggest a novel WSPCSS method, using a publicly available additional source of information. + +# 2.2. Point Cloud Segmentation Using 2D Models + +There have been extensive efforts to achieve point cloud segmentation in an image-based manner as in [12, 16, 20, 30, 43]. With the rapid advancement in the 2D domain, multi-modal self-supervised approaches [1, 6, 13] and employing language-based model [9, 27, 34] have also been particularly prevalent. Recently, there has been a surge in attempts to address various point cloud tasks using the Segment Anything Model (SAM) [19] The zero-shot approach [44, 46], which directly applies SAM to the image corresponding to the point cloud, has demonstrated promising segmentation results. However, SAM's lack of explicit semantics limits its applicability to high-level perception tasks. Adaptation-based approaches [11] can make SAM directly learn 3D tasks; however, they necessitate dense ground truth data during the learning process, rendering them unsuitable for WSPCSS. To tackle the weakly supervised setting, this paper presents a novel WSPCSS method that effectively refines weak labels using the foundational 2D model as an artificial oracle. + +# 2.3. Active Learning for Point Cloud Segmentation + +Active Learning (AL) aims to identify a subset of instances that an oracle manually annotates during training. As in [17, 32, 36, 41], AL has also been extensively explored for WSPCSS. Given that the REAL framework actively employs SAM as an artificial oracle, it may share a conceptual scheme with conventional AL-based methods to some extent. Nevertheless, note that our weakly-supervised setting significantly differs from them. Our framework solely relies on initial sparse labels and does not involve any additional manual annotation within the training loop. + +# 3. Method + +# 3.1. Preliminaries + +Problem setting. We denote an input point cloud as a set of $N$ points: $X = \{x_{1},\ldots ,x_{N}\}$ . In typical fully-supervised point cloud semantic segmentation, the labels for all points are provided as $Y = \{y_{1},\dots ,y_{N}\}$ where $y_{i}\in \{1,\dots ,C\}$ and $C$ is the number of classes. On the other hand, in weakly-supervised point cloud semantic segmentation, only a tiny subset of the points is labeled, and the other points are unlabelled. We formulate this setting as $(X,Y) = (X^{L},Y^{L})\cup (X^{U},Y^{U})$ . Here, $X^{L} = \{x_{1}^{L},\ldots ,x_{n}^{L}\}$ is the + +points labeled with sparse annotations $Y^{L} = \{y_{1}^{L},\dots ,y_{n}^{L}\}$ Conversely, $X^{U} = \{x_{1}^{U},\ldots ,x_{N - n}^{U}\}$ is not labeled, and we represent it as $Y^{U} = \{y_{1}^{U},\dots ,y_{N - n}^{U}\}$ , where $y_{i}^{U} = 0$ + +Segment Anything Model. SAM is a foundational model for promptable segmentation, comprising 1) an image encoder, 2) a prompt encoder, and 3) a decoder predicting the mask using the embeddings from encoders. We denote the inference process of SAM as + +$$ +M ^ {2 D} = \operatorname {S A M} (I; Z), \tag {1} +$$ + +where $I$ and $Z$ are an input image and prompt, respectively. $M^{2D}$ denotes a set of the pixels of the predicted mask. In this paper, we use a point prompt (i.e., a pixel coordinate). + +# 3.2. Projection-based Point-to-Segment + +We begin by enhancing the initially given weak annotations using SAM. However, SAM is a vision foundation model designed and trained for 2D imagery. Therefore, to directly process 3D point cloud data using SAM, it is necessary to conduct additional adaptation or fine-tuning. + +Instead, we propose to utilize image data that captures the same scene as the provided point cloud data. We believe that incorporating image data is not overly restrictive in practice. In fact, point cloud data is usually created from RGBD frames using Structure from Motion (SfM) [3, 8] or captured by LiDAR paired with RGBD cameras [4, 5, 33]. Hence, most point cloud datasets naturally contain RGB images, depth maps, and associated camera parameters. Under this philosophy, we introduce a Projection-based Point-to-Segment (PP2S) module, enabling promptable segmentation of 3D point cloud data using SAM via projection. It is noteworthy that our method utilizes image data during the training phase only and can perform semantic segmentation using input point cloud data alone during testing. + +Figure 2 illustrates the visualization of PP2S. Formally, for each point cloud $X$ , we assume that the model can access data from $J$ cameras: images, depth maps, and projection matrices. Initially, we project the set of weakly annotated points $X^{L} = \{x_{i}^{L}\}$ onto the image $I_{j}$ as: + +$$ +p _ {i, j} = \operatorname {p r o j} _ {j} x _ {i} ^ {L} \text {w h e r e} i \in \{1, \dots , n \}, j \in \{1, \dots , J \} \tag {2} +$$ + +where $\mathrm{proj}_j$ denotes the projection matrix, and $p_{i,j}$ represents the 2D coordinate of the $i$ th weakly annotated points on the $j$ th image. After the projection, we discard points that fall outside the image boundaries. We also filter out occluded points using the depth map $D_j$ , though alternative visibility testing algorithms can be employed. + +Then, SAM performs segmentation using the coordinates of projected points as individual input prompts as + +$$ +M _ {i, j} ^ {2 D} = \operatorname {S A M} \left(I _ {j}; p _ {i, j}\right), \tag {3} +$$ + +where $M_{i,j}^{2D}$ is a set of the pixels of predicted 2D mask on $I_j$ , given input point prompt $p_{i,j}$ . Since $p_{i,j}$ corresponds to + +![](images/9e3ef5efdabdc8bac842395a92c9afc8483a8c9e92b9eb59862ec8dbbefde46f.jpg) + +![](images/f5d5f5d52c42afc0921b170db5e0742956aa933f46913fd0f03e72d716cdade6.jpg) +Figure 2. Visualization of the proposed Projection-based Point-to-Segment (PP2S) module. The process begins by projecting the weak labels onto images through projection matrices. Using these projected points as point prompts, SAM predicts 2D masks, which are subsequently back-projected onto the point cloud. + +$x_{i}^{L}$ and is labeled as $y_{i}^{L}$ , the pixels in $M_{i,j}^{2D}$ are also assigned to $y_{i}^{L}$ . As multiple cameras exist per a single point cloud, we conduct the above process for all images in parallel. + +Subsequently, we back-project the acquired 2D masks to 3D coordinates. Initially, similar to Equ. 2, we project the point cloud onto the $j$ th image as $p_{k,j} = \mathrm{proj}_j x_k$ , where $x_k$ is the $k$ th point in the point cloud and $p_{k,j}$ denotes the projection of $x_k$ . If $p_{k,j}$ is an element of $M_{i,j}^{2D}$ , it means that the projection result of $x_k$ belongs to the mask of the weakly annotated point $x_i^L$ , within the image $I_j$ . Therefore, we define the 3D mask $M_{i,j}^{3D}$ as the set of $x_k$ such that its projection $p_{k,j}$ is in $M_{i,j}^{2D}$ . + +The above process can be formulated as + +$$ +M _ {i, j} ^ {3 D} = \left\{x _ {k} \mid p _ {k, j} \in M _ {i, j} ^ {2 D} \right\}, \tag {4} +$$ + +where $M_{i,j}^{3D}$ denotes the set of points predicted by SAM to be the same segment with $x_{i}^{L}$ from the perspective of $j$ th image. Accordingly, $y_{i,j}^{3D}$ , the class of $M_{i,j}^{3D}$ , is $y_{i}^{L}$ . + +However, the 3D masks are not mutually exclusive. Given that a single 3D point could be observed from multiple images, it may be shared by more than one 3D mask. To address this, we establish a consensus among the 3D masks through a mask-wise voting system. In our system, each 3D mask $M_{i,j}^{3D}$ casts a vote for its points being set to $y_{i,j}^{3D}$ . Therefore, from the perspective of each point, its class is determined by the voting of the 3D masks that include the point. The above process can be formulated as + +$$ +v _ {k} = \left\{y _ {i, j} ^ {3 D} \right\} \forall i, j \text {s u c h t h a t} x _ {k} \in M _ {i, j} ^ {3 D}, \tag {5} +$$ + +where $v_{k}$ is a multiset of the voted classes. Finally, we assign the class with the highest number of votes among $v_{k}$ as + +![](images/821b1a39497bff83c1f2cd91ca6131edfc0671f3d224bcf3f8ad433e0f583861.jpg) +Figure 3. Possible sources of noise in PP2S. (A): The 3D door point is projected onto the chair in the image, resulting in an incorrect 2D mask. (B): Erroneous back-projection causes wall points in 3D to be mislabeled as chair. (C) Above: The 2D mask of the ceiling covers only a small portion of the entire object. (C) Below: The 2D mask of the wall encroaches into the regions of column. + +the artificial label $a_{k}$ for $k$ th point $x_{k}$ : + +$$ +a _ {k} = \left\{ \begin{array}{l l} 0 & \text {i f} v _ {k} = \emptyset \\ \operatorname {M o d e} \left(v _ {k}\right) & \text {o t h e r w i s e ,} \end{array} \right. \tag {6} +$$ + +where the Mode operator returns the most frequent element. + +The proposed PP2S module effectively leverages the segmentation capability of SAM to expand the sparsely annotated initial labels. These resulting artificial labels, denoted as $A = \{a_k\}$ , can directly serve as a supervised signal for training. Notably, we have observed that training with $A$ yields significant performance improvements compared to using initial $Y$ . The detailed experimental results are demonstrated in Section 4. + +# 3.3. Region Exploration via Artificial Labeling + +As mentioned, the proposed PP2S effectively enhances the initial sparse annotations, resulting in a substantial improvement in segmentation performance. However, we also have recognized several limitations associated with this preprocessing approach. The PP2S-enhanced labels can serve as better guidance than the initial sparse labels. Nevertheless, they are still fixed during the entire training phase and thereby impose an upper bound on performance. Notably, we cannot guide the regions not covered by the 2D masks obtained within PP2S. Furthermore, even if we assume that the initially provided sparse annotations are perfectly correct, errors stemming from factors such as noises of camera projection matrices (Fig. 3 A, B) or inherent ambiguity of SAM (Fig. 3 C) are inevitable. While the voting system can mitigate some of these, it cannot completely eliminate them. + +Algorithm 1 Region Exploring via Artificial Labeling +Require: Segmentation model $f_{\theta}$ , Training dataset $\mathcal{D} =$ $\{(X,Y)\}$ , Learning rate $\eta$ , Number of epochs $T$ +Ensure: Optimized $f_{\theta}$ +1: for $(X,Y)$ in $\mathcal{D}$ do +2: $Y\gets \mathrm{PP2S}(X;X^L)$ +3: end for +4: Initialize the model parameter $\theta$ +5: for $t\gets 1,2,\ldots ,T$ do +6: for $(X,Y)$ in $\mathcal{D}$ do +7: $Y^{t - 1}\gets Y$ +8: $S\gets \mathrm{softmax}(f_{\theta}(X))$ +9: $\hat{Y}\gets \mathrm{argmax}(S)$ +10: $Q\gets \mathrm{QuerySampling}(X,S,Y^{t - 1})$ +11: $A\gets \mathrm{PP2S}(X;Q)$ +12: $Y^{t}\gets \mathrm{LabelUpdate}(Y^{t - 1},A,\hat{Y})$ +13: $Y\gets Y^{t}$ +14: $\theta \leftarrow \theta -\eta \nabla_{\theta}\mathcal{L}(S,Y)$ +15: end for +16: end for + +Then, how can we further improve label quality, especially during training? We draw inspiration from the concept of active learning. In active labeling, the model identifies a set of instances, known as a query, during training and requests human annotators (i.e., oracle) to label these points. Typically, the query consists of instances that confuse the model at a given timestep, making annotating the query highly effective for the model to learn decision boundaries. However, active learning requires manual annotation within the training loop, potentially making it even more expensive than WSPCSS. + +To address this challenge, we present the Region Exploration via Artificial Labeling (REAL) framework—a novel WSPCSS method utilizing the PP2S module as an artificial oracle, as depicted in Fig. 4. Specifically, the proposed PP2S module serves as an oracle in the REAL framework, responding to requests to annotate the query. Given that the artificial oracle is, in fact, a pretrained network, the REAL framework is completely relieved of the necessity for manual annotation, except for the initial sparse labels. + +However, unlike a human oracle in conventional methods, our artificial oracle lacks explicit semantic knowledge. Therefore, we cannot expect the PP2S module to directly label the classes for the confusing points. Instead, we focus on leveraging what PP2S can provide to us. PP2S cannot directly annotate the given points; however, it can precisely predict which points should be grouped into a segment. To fully utilize this attribute, we define the query $Q = \{q_i\}$ to consist of the confidently identified points, instead of ambiguous ones from the perspective of the model. Trusting the prediction of the model on these points, we can self-label the query with minimal risk. We represent this pro + +![](images/9e3da3fe03aceecc7f9da1adc9efa253e4ec16ddbcbecec297306a825b13b677.jpg) +Region Exploration via Artificial Labeling (REAL, Sec. 3.3) +Figure 4. Visualization of the proposed Region Exploration via Artificial Labeling (REAL) framework. The objective is to dynamically enhance the quality of labels for weakly supervised point cloud semantic segmentation. In the REAL framework, the model $f_{\theta}$ takes the point cloud $X$ as input and predicts semantic segmentation logits $S$ . We then sample the most confident points from the unlabeled regions of the previous label $Y^{t-1}$ based on the predictions $\hat{Y}$ and reliability scores $R$ . The sampled points then serve as prompts for the proposed PP2S module, producing artificial labels $A$ . Finally, by combining the model prediction and $A$ , we refine $Y^{t-1}$ into the updated labels $Y^t$ which serve as pseudo-ground truth for training the logits. + +cess as QuerySampling in Algorithm 1, which will be thoroughly discussed in Section 3.4. + +Subsequently, we request the PP2S module to act as an artificial oracle, annotating the unknown regions using the given query. For this, the self-labeled query serves as an input prompt for the PP2S module as follows: + +$$ +A = \operatorname {P P 2 S} (X; Q), \tag {7} +$$ + +where $A$ represents the artificial label obtained by PP2S from the given query $Q$ . This approach is akin to using a set of sparsely annotated points $(X^L)$ for PP2S in the preprocessing phase, as described in Section 3.2. The primary distinction in the REAL framework is that the input prompt is self-labeled, not manually annotated. + +With the obtained artificial label $A$ , we update the label of the previous step, $Y^{t-1}$ , into $Y^t$ . However, $A$ may not be entirely reliable, as it could be more prone to noise compared to labels provided by a human oracle. This potential for noise could arise from inaccuracies in the self-labeled input query or the issues highlighted in Fig. 3. + +To mitigate the risk of incorporating errors during the label-updating process, we devise a cautious approach using the prediction of the model. Specifically, if a discrepancy exists between the class assigned by the artificial label + +$A$ and the model's prediction $\hat{Y}$ for a specific point, we exclude that point from the updating process. This conservative update process can be formulated as: + +$$ +y _ {i} ^ {t} = \left\{ \begin{array}{l l} a _ {i} & \text {i f} a _ {i} = \hat {y} _ {i} \\ y _ {i} ^ {t - 1} & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +Finally, the model $f_{\theta}$ is optimized by minimizing the cross-entropy loss between $\hat{Y}$ and $Y^{t}$ . + +# 3.4. Query Sampling Strategy for REAL + +In the REAL framework, query sampling is a critical component, as the quality of the artificial label depends on the quality of the input query. We posit that an effective query must meet two key criteria: 1) it should exhibit a high level of confidence, and 2) it should aid the model in acquiring new knowledge. To address the first criterion, we design a margin-based reliability metric for evaluating the confidence of the model's prediction for each point as follows: + +$$ +r _ {i} = \operatorname {M a x} \left(s _ {i}\right) - \operatorname {M a x} 2 \left(s _ {i}\right), \tag {9} +$$ + +where $s_i \in \mathbb{R}^C$ is the predicted logit of the $i$ th point $x_i$ , and $r_i$ denotes its reliability. Max and Max2 are operators that return the maximum and the second maximum el + +![](images/ce20e8baf2ab7ef32c9826296e236d86c41d1390f15d7c285d93d24d00022ecf.jpg) +Figure 5. Verification of the reliability metric. The points presenting higher reliability show higher semantic segmentation accuracy. + +ements of the input vector, respectively. We verify a robust correlation between the devised reliability metric and the accuracy of model predictions, as illustrated in Fig. 5. For instance, among points with reliability scores exceeding 0.5, the model correctly predicts the class for $84.5\%$ of them. We set the threshold for query sampling at 0.95, corresponding to around $94\%$ accuracy. + +On the contrary, directly quantifying the utility of a query for the second criterion is inherently elusive. In this context, we propose an intuitive hypothesis: if the model has successfully leveraged the labels from the previous timestep, $Y^{t-1}$ , the regions previously annotated in $Y^{t-1}$ are expected to contribute less novel information during the current timestep, $t$ . Therefore, the focus of new queries should be on effectively exploring the regions not previously or incorrectly labeled in $Y^{t-1}$ . + +To this end, we narrow the sampling range of the query to the points that are predicted by the model differently from the previous label. We formulate the above constraint as + +$$ +Z _ {c} = \left\{x _ {i} \right\} \forall i \text {s u c h t h a t} \hat {y} _ {i} = c \text {a n d} y _ {i} ^ {t - 1} \neq c. \tag {10} +$$ + +Here, $Z_{c}$ is the mother set for sampling $q_{c}$ , the query point of the class $c$ . Considering the first criterion, we sample $q_{c}$ as the most reliable point among $Z_{c}$ , and $Q$ is the set of $q_{c}$ . We formulate the above processes as + +$$ +Q = \operatorname {Q u e r y S a m p l i n g} (X, S, Y ^ {t - 1}). \tag {11} +$$ + +# 4. Experiments + +# 4.1. Experimental settings + +Datasets. We conduct experiments on S3DIS [3] and ScanNetV2 [8]. S3DIS comprises 6 areas with 272 rooms and 13 categories. We use the area 5 for evaluation following previous studies. ScanNetV2 includes 1613 3D scenes with 20 categories. We follow the official split (1201 training, 312 validation, and 100 test scenes). In addition to the point cloud data, the proposed REAL framework requires the data from cameras (i.e., images, depth maps, and camera projection matrices) to leverage SAM. The datasets inherently include such data since they are reconstructed from the sequence of RGBD frames. During training, we utilize the data from 48 and 17 cameras per scene on average for S3DIS and ScanNetV2, respectively. Note that our framework does not require 2D data at inference time. + +Table 1. A quantitative comparison between the initially sparse annotations (Initial) and the artificial labels from PP2S (PP2S). Pre, Rec, and # denote precision, recall, and the proportion of annotated points relative to the total number of points. The performance of the point cloud semantic segmentation model trained by each model is represented as Model mIoU. Every metric is in %. + +
LabelsmIoUPreRec#Model mIoU
Initial0.0041000.0040.00439.7
PP2S30.483.132.438.453.0
+ +![](images/73ed7a6e9f304e8b527c4cfce59c7f270aed69a63f85c610edbae5ff17741ba5.jpg) +(A) Ground Truth + +![](images/b14a1b2f96268334df4edd870f80282eb8a837132e98706f11d130ffc8e1b5ba.jpg) +(B) Labels of the previous step +(D) Explored regions (setting II) + +![](images/407bce03c378fee676c01e3a5955b31913472da376d66c02347faaa2ad656b8a.jpg) +(C) Explored regions (setting 1) +Prec: -0.09% +Rec: +11.92% +Prec: -0.03% +Rec: +5.58% +Figure 6. Validation of our query sampling strategy. Setting I samples the queries following our strategy while setting II samples from the entire point set. We present the improvement obtained by artificial labels in terms of precision (Prec) and recall (Rec). + +Evaluation protocols. Previous studies [23, 24, 42, 48] have conducted experiments with various levels of sparsity. We believe that the most cost-efficient yet practical setting involves labeling only one point per object. Accordingly, our experiments mainly target the 1pt weak setting, which corresponds to $0.004\%$ of points in S3DIS. For a fair comparison, we also conduct experiments under the $0.02\%$ setting, which is the most widely used. For ScanNetV2, we follow the conventional works, using 20pt per scene. The overall performance is assessed on all points within the test set. Following the standard practice, we employ the mean Intersection over Union (mIoU) as our quantitative metric. Implementation details. We choose PTv2 [37] network as our main backbone. To verify the effect of the backbone, we additionally test Closer [25] backbone for the $0.004\%$ setting on S3DIS. Training details are the same as standard settings of the backbones. We employ a pretrained ViT-H model for SAM. Further details are in the Supp. + +# 4.2. Analysis on REAL + +PP2S-based preprocessing. We first validate the preprocessing strategy built upon the proposed PP2S. This strategy aims to enhance the initial sparse annotations into artificial labels by harnessing the promptable segmentation capability of SAM via projection. Table 1 provides various met + +![](images/af5fcc7a63c0192ddb3a1e7c1190f840d721cba417e78fb5f07c0b23494c455c.jpg) +Figure 7. Progressive visualization of the behavior of the proposed REAL framework throughout training. $A$ is the artificial label, and $Y$ denotes the subsequent label updated at timestep $t$ . The highlighted boxes indicate newly explored regions by artificial oracle, which were unknown (red, blue) or wrongly labeled (green) in earlier timesteps. + +rics of the initial labels and those of the labels enhanced by PP2S. Furthermore, we compare the performance of the point cloud semantic segmentation model when trained with the original sparse annotations and artificial labels enhanced by PP2S. The results demonstrate that our PP2S significantly improves the quality of the label, consequently boosting the model's performance. This analysis supports the validity of PP2S as a bridge for transferring the segmentation capability of SAM into the realm of weakly supervised point cloud semantic segmentation. + +Query sampling strategy. We conduct ablation studies regarding the proposed query sampling strategy introduced in Section 3.4. In this section, we compare two settings: (I) using the proposed strategy sampling from the narrowed mother set of Equ. 10 and (II) sampling the most reliable point among every point without any constraint. Figure 6 provides both qualitative and quantitative comparisons between (I) and (II), displaying the improvement achieved by the artificial labels of each setting. (I) outperforms (II) in terms of recall while maintaining precision. This distinction primarily arises from the fact that most query points of (II) are already labeled in $Y^{t-1}$ , limiting their potential for exploring into unknown regions. On the other hand, the query of (I) is explicitly restricted to represent genuinely novel information compared with previous labels $Y^{t-1}$ , even without compromising precision. This implies that the proposed query sampling strategy yields better labels, facilitating the model to learn segmentation effectively. Accordingly, the performance with our query sampling strategy (62.7%) is significantly higher than without using it (59.8%). Additional results are in Supp. + +Active label enhancement. To demonstrate the behavior of the proposed REAL framework intuitively, Fig. 7 depicts progressive visualization throughout training. The labels are significantly improved as training proceeds, thanks to the artificial labels from our artificial oracle. Notably, the highlighted boxes indicate newly explored regions, which were unknown (red, blue) or wrongly labeled (green) by the labels in earlier timesteps. The REAL framework effectively integrates these regions into the current label. + +# 4.3. Comparison with State-of-the-arts + +We compare our method with SoTA methods on S3DIS, as shown in Table 2. The proposed REAL significantly enhances performance across all settings. Specifically, REAL with PTv2 [37] outperforms CPCM by $2.9\%$ in mIoU under the $0.02\%$ setting, which is even comparable to the SoTA results of the $1\%$ setting. Furthermore, REAL with PTv2 surpasses the other SoTA methods under the $0.02\%$ setting, using only 1pt for each object $(0.004\%)$ . We also observe a significant performance improvement when employing the REAL in Closer [25]. These results imply that the gain of REAL is not restricted by the backbone. + +Besides, Fig. 8 provides a qualitative comparison. Notably, the semantic confusion in the baseline is remarkably improved in ours. Furthermore, REAL demonstrates a meaningful enhancement in its ability to group the points of each object precisely. Consequently, REAL achieves remarkable semantic segmentation results even in complicated scenes, using only $0.004\%$ of annotations. + +Furthermore, we evaluate the performance of our method against SoTA methods on ScanNetV2 [8], as detailed in Ta + +Table 2. Comparisons of the proposed methods with SoTA methods on S3DIS test set. Every metric is in mIoU (%). Bold number represents the best result. † denotes the results from our reimplementation. + +
SettingsMethodsmIoUceil.floorwallbeamcol.wind.doorchairtablebook.sofaboardclutter
FullyRandLA [14]62.491.295.780.10.025.262.347.475.883.260.870.865.254.0
KPConv [35]67.192.897.382.40.023.958.069.091.081.575.375.466.758.9
RFCR [10]68.794.298.384.30.028.562.471.292.082.676.171.171.661.3
Closer† [25]66.093.298.281.20.024.849.969.390.480.472.074.668.055.8
PTv2† [38]70.993.898.586.10.029.560.578.093.281.876.679.283.061.7
1%SPT [47]61.891.596.980.60.018.258.147.275.885.765.368.965.050.2
PSD [48]63.592.397.780.70.027.856.262.578.784.163.170.458.953.2
HybridCR [23]65.392.593.982.60.024.264.463.278.381.769.074.468.256.5
0.03%PSD [48]48.287.996.062.10.020.649.340.955.161.943.950.727.331.1
HybridCR [23]51.585.491.965.90.018.051.434.263.878.352.459.629.939.0
0.02%MIL [42]51.486.693.275.00.029.345.346.760.562.356.547.533.732.2
CPCM [24]62.392.695.679.40.017.849.359.485.775.669.160.768.255.8
PTv2† [38]55.180.594.773.90.021.742.938.981.459.262.552.966.142.1
+PP2S60.079.093.177.50.131.549.761.470.073.769.956.971.045.6
+REAL65.286.995.880.20.027.060.377.872.379.970.864.777.454.1
0.004%Closer† [25]38.782.592.469.40.115.023.335.650.947.80.319.137.329.1
+PP2S44.962.988.867.90.011.729.244.746.463.253.525.750.738.6
+REAL57.879.097.770.80.124.450.460.184.564.664.866.839.948.7
PTv2† [38]39.776.687.265.40.09.430.622.361.750.341.714.622.434.1
+PP2S53.065.993.168.90.027.342.755.868.463.758.252.550.541.8
+REAL62.784.095.780.30.031.157.463.775.576.968.770.065.446.8
+ +![](images/de220b6e7ad44dc8499b83f51eeacbfe0570ec27bd784535fe63ec087127cb53.jpg) +Figure 8. Qualitative comparisons between the semantic segmentation results of the baseline and ours on S3DIS under $0.004\%$ setting. The displayed boxes indicate the regions where significant improvements are achieved by the proposed method. + +ble 3. Our method demonstrates superior performance on both the validation and test splits. Notably, REAL outperforms CPCM by $3.6\%$ in the validation set and $2.7\%$ in the test set. These results imply that our method effectively utilizes the potential of weak labels. The qualitative results for ScanNetV2 can be found in Supp. + +# 5. Conclusion + +Expensive and labor-intensive point-wise annotation hinders the practical application of point cloud semantic segmentation. While conventional WSPCSS methods have exhibited promising results, the fixed input sparse label potentially constrains the achievable performance. In pursuit of a novel research direction, we propose the utilization of an additional source of information, introducing the Region Exploration via Artificial Labeling (REAL) framework. This framework aims to harness the power of the image foundational model, SAM, within the context of active learning. To establish the connection between the 2D model + +Table 3. Comparisons of the proposed methods with SoTA methods on ScanNetV2 val/test set. The results are in mIoU. + +
SettingsMethodsVal.Test
FullyKPConv [35]-68.4
RFCR [10]-70.2
1%SPT [47]-51.1
PSD [48]-54.7
HybridCR [23]56.956.8
20 ptsMIL [42]57.854.4
CPCM [24]62.762.8
Ours (REAL)66.365.5
+ +and 3D data, we designed the Projection-based Point-to-Segment (PP2S) module. The PP2S is employed as an artificial oracle in the REAL framework, dynamically enhancing labels during training. Although using an artificial oracle eliminates the burden of manual annotation, it introduces some drawbacks due to the model's imperfections. As a remedy, we devised several strategies for query sampling and label updating. Through extensive experiments, we demonstrated the working logic of our framework in detail. Finally, the proposed REAL framework achieved new SoTA on various datasets, surpassing existing works. We believe that our approach pioneers the effective integration of the foundational image model into the realm of WSPCSS. + +Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601). + +# References + +[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 2 +[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4981-4990, 2018. 2 +[3] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 3, 6 +[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 3 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3 +[6] Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5291-5301, 2023. 2 +[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 3, 6, 7 +[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2 +[10] Jingyu Gong, Jiachen Xu, Xin Tan, Haichuan Song, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Omni-supervised point cloud segmentation via gradual receptive field component reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11673-11682, 2021. 1, 8 +[11] Shizhan Gong, Yuan Zhong, Wenao Ma, Jinpeng Li, Zhao Wang, Jingyang Zhang, Pheng-Ann Heng, and Qi Dou. 3dsam-adapter: Holistic adaptation of sam from 2d to 3d + +for promptable medical image segmentation. arXiv preprint arXiv:2306.13465, 2023. 2 +[12] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 2 +[13] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2 +[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 8 +[15] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In European Conference on Computer Vision, pages 600-619. Springer, 2022. 1, 2 +[16] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021. 2 +[17] Zeyu Hu, Xuyang Bai, Runze Zhang, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Lidal: Inter-frame uncertainty based active learning for 3d lidar semantic segmentation. In European Conference on Computer Vision, pages 248-265. Springer, 2022. 2 +[18] Jihun Kim, Hyeokjun Kwon, Yunseo Yang, and Kuk-Jin Yoon. Learning point cloud completion without complete point clouds: A pose-aware approach. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14157-14167. IEEE, 2023. 2 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2 +[20] Hyeokjun Kweon and Kuk-Jin Yoon. Joint learning of 2d-3d weakly supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:30499–30511, 2022. 2 +[21] Hyeokjun Kweon, Sung-Hoon Yoon, Hyeonseong Kim, Daehe Park, and Kuk-Jin Yoon. Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6994-7003, 2021. 2 +[22] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 2 + +[23] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14930-14939, 2022. 1, 2, 6, 8 +[24] Lizhao Liu, Zhuangwei Zhuang, Shangxin Huang, Xunlong Xiao, Tianhang Xiang, Cen Chen, Jingdong Wang, and Mingkui Tan. Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18413-18422, 2023. 1, 2, 6, 8 +[25] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 326-342. Springer, 2020. 6, 7, 8 +[26] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021. 1, 2 +[27] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1190-1199, 2023. 2 +[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1 +[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1 +[30] Damien Robert, Bruno Vallet, and Loic Landrieu. Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5575-5584, 2022. 2 +[31] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 2 +[32] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022. 2 +[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceed- + +ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3 +[34] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2 +[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 8 +[36] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2 +[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 6, 7 +[38] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 8 +[39] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2 +[40] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020. 1, 2 +[41] Zongyi Xu, Bo Yuan, Shanshan Zhao, Qianni Zhang, and Xinbo Gao. Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18098-18108, 2023. 2 +[42] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022. 1, 2, 6, 8 +[43] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2d-3d interlaced transformer for point cloud segmentation with scene-level supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 977–987, 2023. 2 +[44] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 2 +[45] Sung-Hoon Yoon, Hyeokjun Kweon, Jegyeong Cho, Shinjeong Kim, and Kuk-Jin Yoon. Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, + +October 23-27, 2022, Proceedings, Part XXIX, pages 326-344. Springer Nature Switzerland Cham, 2022. 2 +[46] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: Zero-shot 3d object detection via segment anything model. arXiv preprint arXiv:2306.02245, 2023. 2 +[47] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3421-3429, 2021. 1, 2, 8 +[48] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15520-15528, 2021. 1, 2, 6, 8 \ No newline at end of file diff --git a/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/images.zip b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..92561724618d9741be14fd2f650637724f6e78bd --- /dev/null +++ b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0963e11c1699da120c110ef582009748e7dcc055d0e76510757d1d2832037929 +size 842105 diff --git a/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/layout.json b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aeb193fa7e7d1ec1e04117570aa3480c454f18b9 --- /dev/null +++ b/2024/Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle/layout.json @@ -0,0 +1,9778 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "text", + "content": "Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 144, + 203, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 144, + 203, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 144, + 203, + 170 + ], + "type": "text", + "content": "Hyeokjun Kweon* KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 174, + 212, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 174, + 212, + 184 + ], + "spans": [ + { + "bbox": [ + 103, + 174, + 212, + 184 + ], + "type": "text", + "content": "0327june@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 272, + 144, + 328, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 144, + 328, + 170 + ], + "spans": [ + { + "bbox": [ + 272, + 144, + 328, + 170 + ], + "type": "text", + "content": "Jihun Kim* KAIST" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 174, + 358, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 174, + 358, + 184 + ], + "spans": [ + { + "bbox": [ + 244, + 174, + 358, + 184 + ], + "type": "text", + "content": "jihun1998@kaist.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 405, + 144, + 473, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 144, + 473, + 170 + ], + "spans": [ + { + "bbox": [ + 405, + 144, + 473, + 170 + ], + "type": "text", + "content": "Kuk-Jin Yoon KAIST" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 388, + 174, + 487, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 174, + 487, + 184 + ], + "spans": [ + { + "bbox": [ + 388, + 174, + 487, + 184 + ], + "type": "text", + "content": "kjyoon@kaist.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 237, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 289, + 501 + ], + "type": "text", + "content": "Manual annotation of every point in a point cloud is a costly and labor-intensive process. While weakly supervised point cloud semantic segmentation (WSPCSS) with sparse annotation shows promise, the limited information from initial sparse labels can place an upper bound on performance. As a new research direction for WSPCSS, we propose a novel Region Exploration via Artificial Labeling (REAL) framework. It leverages a foundational image model as an artificial oracle within the active learning context, eliminating the need for manual annotation by a human oracle. To integrate the 2D model into the 3D domain, we first introduce a Projection-based Point-to-Segment (PP2S) module, designed to enable prompt segmentation of 3D data without additional training. The REAL framework samples query points based on model predictions and requests annotations from PP2S, dynamically refining labels and improving model training. Furthermore, to overcome several challenges of employing an artificial model as an oracle, we formulate effective query sampling and label updating strategies. Our comprehensive experiments and comparisons demonstrate that the REAL framework significantly outperforms existing methods across various benchmarks. The code is available at https://github.com/jihun1998/AO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 525, + 127, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 127, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 127, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 545, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 641 + ], + "type": "text", + "content": "Point cloud semantic segmentation has been extensively researched for robotics or autonomous driving applications. While fully supervised approaches [7, 10, 14, 28, 29, 35] utilizing deep learning have demonstrated remarkable progress, the densely annotated point-wise labels required for these methods pose a significant challenge due to the labor-intensive and expensive annotation process. This obstacle hinders the practical application of existing methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "In response to this challenge, Weakly Supervised Point Cloud Semantic Segmentation (WSPCSS) has gained attention, leveraging more affordable weak labels [15, 23, 24, 26, 40, 42, 47, 48]. Among various types of weak labels, sparse annotations have become a widely adopted setting, where only a tiny subset (e.g., " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ") of the entire point cloud is" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 213, + 541, + 319 + ], + "blocks": [ + { + "bbox": [ + 310, + 213, + 541, + 319 + ], + "lines": [ + { + "bbox": [ + 310, + 213, + 541, + 319 + ], + "spans": [ + { + "bbox": [ + 310, + 213, + 541, + 319 + ], + "type": "image", + "image_path": "7276b774f940b61be70e989ed1b975016f2a7e9949e50e1f5473060fb200d42d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 321, + 542, + 369 + ], + "blocks": [ + { + "bbox": [ + 309, + 321, + 542, + 369 + ], + "lines": [ + { + "bbox": [ + 309, + 321, + 542, + 369 + ], + "spans": [ + { + "bbox": [ + 309, + 321, + 542, + 369 + ], + "type": "image", + "image_path": "8f338933edb540490bbef37c53cd2dfcea9d95ddca585a919971bae58ef1bce7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 379, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 434 + ], + "type": "text", + "content": "Figure 1. Visualization of the proposed REAL framework. We tackle the challenge of WSPCSS, leveraging a foundational imagery model as an artificial oracle within the context of active learning. The below examples illustrate the evolving labels as training proceeds." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 447, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 555 + ], + "type": "text", + "content": "labeled. Existing WSPCSS works have proposed remedies focusing on two key aspects: 1) fully utilizing the limited information from sparse labels and 2) learning the features beneficial for segmentation from the data. The methods based on self-labeling use the regions confidently predicted by the model as pseudo-labels for training the model itself. Meanwhile, various studies have proposed self-supervised learning methods, such as consistency against data augmentation, contrastive learning, or masked modeling." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "type": "text", + "content": "While these approaches have shown substantial results, the overall amount of information provided is constrained by the initial sparse annotation, potentially placing an upper limit on performance. In this paper, instead of solely focusing on how to exploit the given limited information, we embark on a new and more challenging research direction. Specifically, we propose a novel WSPCSS method using an additional publicly available source of information—leveraging advancements in the 2D domain with the Segment Anything Model (SAM) [19]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We begin by utilizing the promptable segmentation capability of SAM to enhance initially provided sparse annotations. To bridge the gap between the 2D model and 3D" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "3721" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "point cloud data, we introduce the Projection-based Point-to-Segment (PP2S) module. This module comprises the projection of weakly annotated points onto images, acquiring 2D masks using SAM, and back-projecting these masks into 3D, facilitating segmentation in the point cloud. The proposed PP2S-based preprocessing effectively improves label quality, resulting in significantly enhanced semantic segmentation performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 287, + 337 + ], + "type": "text", + "content": "However, the labels preprocessed by PP2S remain incomplete due to potential noises from calibration errors or SAM ambiguity. To further enhance label quality during training, we propose the Region Exploration via Artificial Labeling (REAL) framework, utilizing PP2S as an artificial oracle within the context of active learning, as depicted in Fig. 1. The REAL framework samples query points according to model predictions and subsequently requests annotations from PP2S to dynamically enhance training labels. Here, unlike general active learning methods relying on a human oracle, our artificial oracle cannot directly assign the classes of the requested query. Therefore, we design specialized strategies for query sampling and label refinement to unlock the segmentation capability of the PP2S." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "type": "text", + "content": "Note that the main novelty of our approach lies in how we effectively integrate the benefits of SAM into the realm of WSPCSS. We analyze the REAL framework through extensive experiments, including quantitative and qualitative ablation studies. Furthermore, our method achieves new state-of-the-art (SoTA) results on S3DIS [3] and ScanNetV2 [8] under all the tested settings, surpassing the existing WSPCSS methods. Remarkably, even with the " + }, + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "type": "inline_equation", + "content": "0.004\\%" + }, + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "type": "text", + "content": " setting, our method outperforms conventional methods using the " + }, + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 46, + 337, + 287, + 469 + ], + "type": "text", + "content": " setting, underscoring the superiority of the proposed approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 482, + 134, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 134, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 134, + 495 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 218, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 218, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 218, + 516 + ], + "type": "text", + "content": "2.1. Weakly Supervised Approaches" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Originating from weakly supervised learning in imagery domain [2, 21, 22, 31, 39, 45], weakly supervised approach has emerged as a cost-effective approach to mitigate the challenges associated with acquiring fully labeled data [15, 18, 23, 24, 26, 40, 42, 47, 48]. Xu and Lee [40] first propose weakly supervised configurations involving annotations of " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " of the data and a single point label for each category. PSD [48] introduces a perturbed branch and constraint self-distillation loss between the perturbed and original branches. OTOC [26] utilizes pseudo-label methods and iteratively updates pseudo-labels through a self-training strategy. MIL [42] proposes the transformer model derived from multiple instance learning and integrates adaptive global weighted pooling to their model. HybridCR [23] introduces an architecture that leverages pseudo-label methods and consistency regularization between the original and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "augmented branches. CPCM [24] introduces a region-wise masking strategy and contextual masked training method to integrate the benefit of the masked autoencoder. Despite the significant progress in WSPCSS, in this paper, we would like to explore a new and more challenging research direction. In this light, we suggest a novel WSPCSS method, using a publicly available additional source of information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 164, + 534, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 534, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 534, + 176 + ], + "type": "text", + "content": "2.2. Point Cloud Segmentation Using 2D Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 410 + ], + "type": "text", + "content": "There have been extensive efforts to achieve point cloud segmentation in an image-based manner as in [12, 16, 20, 30, 43]. With the rapid advancement in the 2D domain, multi-modal self-supervised approaches [1, 6, 13] and employing language-based model [9, 27, 34] have also been particularly prevalent. Recently, there has been a surge in attempts to address various point cloud tasks using the Segment Anything Model (SAM) [19] The zero-shot approach [44, 46], which directly applies SAM to the image corresponding to the point cloud, has demonstrated promising segmentation results. However, SAM's lack of explicit semantics limits its applicability to high-level perception tasks. Adaptation-based approaches [11] can make SAM directly learn 3D tasks; however, they necessitate dense ground truth data during the learning process, rendering them unsuitable for WSPCSS. To tackle the weakly supervised setting, this paper presents a novel WSPCSS method that effectively refines weak labels using the foundational 2D model as an artificial oracle." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 418, + 545, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 430 + ], + "type": "text", + "content": "2.3. Active Learning for Point Cloud Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 436, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 557 + ], + "type": "text", + "content": "Active Learning (AL) aims to identify a subset of instances that an oracle manually annotates during training. As in [17, 32, 36, 41], AL has also been extensively explored for WSPCSS. Given that the REAL framework actively employs SAM as an artificial oracle, it may share a conceptual scheme with conventional AL-based methods to some extent. Nevertheless, note that our weakly-supervised setting significantly differs from them. Our framework solely relies on initial sparse labels and does not involve any additional manual annotation within the training loop." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 567, + 362, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 362, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 362, + 579 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 587, + 392, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 392, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 392, + 598 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Problem setting. We denote an input point cloud as a set of " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " points: " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "X = \\{x_{1},\\ldots ,x_{N}\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ". In typical fully-supervised point cloud semantic segmentation, the labels for all points are provided as " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "Y = \\{y_{1},\\dots ,y_{N}\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "y_{i}\\in \\{1,\\dots ,C\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " is the number of classes. On the other hand, in weakly-supervised point cloud semantic segmentation, only a tiny subset of the points is labeled, and the other points are unlabelled. We formulate this setting as " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "(X,Y) = (X^{L},Y^{L})\\cup (X^{U},Y^{U})" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "X^{L} = \\{x_{1}^{L},\\ldots ,x_{n}^{L}\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3722" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "text", + "content": "points labeled with sparse annotations " + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "inline_equation", + "content": "Y^{L} = \\{y_{1}^{L},\\dots ,y_{n}^{L}\\}" + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "text", + "content": " Conversely, " + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "inline_equation", + "content": "X^{U} = \\{x_{1}^{U},\\ldots ,x_{N - n}^{U}\\}" + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "text", + "content": " is not labeled, and we represent it as " + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "inline_equation", + "content": "Y^{U} = \\{y_{1}^{U},\\dots ,y_{N - n}^{U}\\}" + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 71, + 286, + 109 + ], + "type": "inline_equation", + "content": "y_{i}^{U} = 0" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 286, + 167 + ], + "type": "text", + "content": "Segment Anything Model. SAM is a foundational model for promptable segmentation, comprising 1) an image encoder, 2) a prompt encoder, and 3) a decoder predicting the mask using the embeddings from encoders. We denote the inference process of SAM as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 172, + 286, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 172, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 123, + 172, + 286, + 186 + ], + "type": "interline_equation", + "content": "M ^ {2 D} = \\operatorname {S A M} (I; Z), \\tag {1}", + "image_path": "7e8b69feaeea7c999f5c9d8e3da9e9c3bad4ff20647c62d0924e38f55a0a1ceb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "text", + "content": " are an input image and prompt, respectively. " + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "inline_equation", + "content": "M^{2D}" + }, + { + "bbox": [ + 47, + 192, + 286, + 228 + ], + "type": "text", + "content": " denotes a set of the pixels of the predicted mask. In this paper, we use a point prompt (i.e., a pixel coordinate)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 233, + 233, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 233, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 233, + 246 + ], + "type": "text", + "content": "3.2. Projection-based Point-to-Segment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 286, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 286, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 286, + 311 + ], + "type": "text", + "content": "We begin by enhancing the initially given weak annotations using SAM. However, SAM is a vision foundation model designed and trained for 2D imagery. Therefore, to directly process 3D point cloud data using SAM, it is necessary to conduct additional adaptation or fine-tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 312, + 286, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 286, + 479 + ], + "type": "text", + "content": "Instead, we propose to utilize image data that captures the same scene as the provided point cloud data. We believe that incorporating image data is not overly restrictive in practice. In fact, point cloud data is usually created from RGBD frames using Structure from Motion (SfM) [3, 8] or captured by LiDAR paired with RGBD cameras [4, 5, 33]. Hence, most point cloud datasets naturally contain RGB images, depth maps, and associated camera parameters. Under this philosophy, we introduce a Projection-based Point-to-Segment (PP2S) module, enabling promptable segmentation of 3D point cloud data using SAM via projection. It is noteworthy that our method utilizes image data during the training phase only and can perform semantic segmentation using input point cloud data alone during testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "text", + "content": "Figure 2 illustrates the visualization of PP2S. Formally, for each point cloud " + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "text", + "content": ", we assume that the model can access data from " + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "text", + "content": " cameras: images, depth maps, and projection matrices. Initially, we project the set of weakly annotated points " + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "inline_equation", + "content": "X^{L} = \\{x_{i}^{L}\\}" + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "text", + "content": " onto the image " + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 47, + 479, + 286, + 539 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 286, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 286, + 567 + ], + "type": "interline_equation", + "content": "p _ {i, j} = \\operatorname {p r o j} _ {j} x _ {i} ^ {L} \\text {w h e r e} i \\in \\{1, \\dots , n \\}, j \\in \\{1, \\dots , J \\} \\tag {2}", + "image_path": "b47d468782ec50093756aeaff1f1d927b5e75105e7bd90be7ffdd422bc427eee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{proj}_j" + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "content": " denotes the projection matrix, and " + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "content": " represents the 2D coordinate of the " + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "content": "th weakly annotated points on the " + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "content": "th image. After the projection, we discard points that fall outside the image boundaries. We also filter out occluded points using the depth map " + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "inline_equation", + "content": "D_j" + }, + { + "bbox": [ + 47, + 569, + 286, + 640 + ], + "type": "text", + "content": ", though alternative visibility testing algorithms can be employed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 641, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 286, + 664 + ], + "type": "text", + "content": "Then, SAM performs segmentation using the coordinates of projected points as individual input prompts as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 669, + 286, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 669, + 286, + 683 + ], + "spans": [ + { + "bbox": [ + 118, + 669, + 286, + 683 + ], + "type": "interline_equation", + "content": "M _ {i, j} ^ {2 D} = \\operatorname {S A M} \\left(I _ {j}; p _ {i, j}\\right), \\tag {3}", + "image_path": "41321f66a305c6743859ff969ad69e9f2fe898477b4e6b38e8a596890df5daf9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "M_{i,j}^{2D}" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": " is a set of the pixels of predicted 2D mask on " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": ", given input point prompt " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": " corresponds to" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 541, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 541, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 541, + 140 + ], + "type": "image", + "image_path": "9e3ef5efdabdc8bac842395a92c9afc8483a8c9e92b9eb59862ec8dbbefde46f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 140, + 541, + 257 + ], + "blocks": [ + { + "bbox": [ + 308, + 140, + 541, + 257 + ], + "lines": [ + { + "bbox": [ + 308, + 140, + 541, + 257 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 541, + 257 + ], + "type": "image", + "image_path": "f5d5f5d52c42afc0921b170db5e0742956aa933f46913fd0f03e72d716cdade6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 262, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 318 + ], + "type": "text", + "content": "Figure 2. Visualization of the proposed Projection-based Point-to-Segment (PP2S) module. The process begins by projecting the weak labels onto images through projection matrices. Using these projected points as point prompts, SAM predicts 2D masks, which are subsequently back-projected onto the point cloud." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "x_{i}^{L}" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": " and is labeled as " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "y_{i}^{L}" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": ", the pixels in " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "M_{i,j}^{2D}" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": " are also assigned to " + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "inline_equation", + "content": "y_{i}^{L}" + }, + { + "bbox": [ + 305, + 327, + 545, + 364 + ], + "type": "text", + "content": ". As multiple cameras exist per a single point cloud, we conduct the above process for all images in parallel." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": "Subsequently, we back-project the acquired 2D masks to 3D coordinates. Initially, similar to Equ. 2, we project the point cloud onto the " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": "th image as " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "p_{k,j} = \\mathrm{proj}_j x_k" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "x_k" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": "th point in the point cloud and " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "p_{k,j}" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " denotes the projection of " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "x_k" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "p_{k,j}" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " is an element of " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "M_{i,j}^{2D}" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": ", it means that the projection result of " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "x_k" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " belongs to the mask of the weakly annotated point " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "x_i^L" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": ", within the image " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": ". Therefore, we define the 3D mask " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "M_{i,j}^{3D}" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " as the set of " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "x_k" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " such that its projection " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "p_{k,j}" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": " is in " + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "inline_equation", + "content": "M_{i,j}^{2D}" + }, + { + "bbox": [ + 305, + 365, + 545, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 474, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 474, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 474, + 480, + 485 + ], + "type": "text", + "content": "The above process can be formulated as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 368, + 491, + 545, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 491, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 368, + 491, + 545, + 508 + ], + "type": "interline_equation", + "content": "M _ {i, j} ^ {3 D} = \\left\\{x _ {k} \\mid p _ {k, j} \\in M _ {i, j} ^ {2 D} \\right\\}, \\tag {4}", + "image_path": "06ac9f37a851a511f21fd97952fdc9b1a650381d0ece89601f7ee49c4d9fa778.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "inline_equation", + "content": "M_{i,j}^{3D}" + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": " denotes the set of points predicted by SAM to be the same segment with " + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "inline_equation", + "content": "x_{i}^{L}" + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": " from the perspective of " + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": "th image. Accordingly, " + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "inline_equation", + "content": "y_{i,j}^{3D}" + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": ", the class of " + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "inline_equation", + "content": "M_{i,j}^{3D}" + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": ", is " + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "inline_equation", + "content": "y_{i}^{L}" + }, + { + "bbox": [ + 305, + 515, + 545, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "type": "text", + "content": "However, the 3D masks are not mutually exclusive. Given that a single 3D point could be observed from multiple images, it may be shared by more than one 3D mask. To address this, we establish a consensus among the 3D masks through a mask-wise voting system. In our system, each 3D mask " + }, + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "type": "inline_equation", + "content": "M_{i,j}^{3D}" + }, + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "type": "text", + "content": " casts a vote for its points being set to " + }, + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "type": "inline_equation", + "content": "y_{i,j}^{3D}" + }, + { + "bbox": [ + 304, + 554, + 545, + 660 + ], + "type": "text", + "content": ". Therefore, from the perspective of each point, its class is determined by the voting of the 3D masks that include the point. The above process can be formulated as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 340, + 667, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 667, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 340, + 667, + 545, + 682 + ], + "type": "interline_equation", + "content": "v _ {k} = \\left\\{y _ {i, j} ^ {3 D} \\right\\} \\forall i, j \\text {s u c h t h a t} x _ {k} \\in M _ {i, j} ^ {3 D}, \\tag {5}", + "image_path": "7b79c4bdc201f626569fcd628ad93bffe817f35621f125e31ec3cf828534012b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is a multiset of the voted classes. Finally, we assign the class with the highest number of votes among " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3723" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 73, + 281, + 243 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 281, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 281, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 281, + 243 + ], + "type": "image", + "image_path": "821b1a39497bff83c1f2cd91ca6131edfc0671f3d224bcf3f8ad433e0f583861.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 287, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 314 + ], + "type": "text", + "content": "Figure 3. Possible sources of noise in PP2S. (A): The 3D door point is projected onto the chair in the image, resulting in an incorrect 2D mask. (B): Erroneous back-projection causes wall points in 3D to be mislabeled as chair. (C) Above: The 2D mask of the ceiling covers only a small portion of the entire object. (C) Below: The 2D mask of the wall encroaches into the regions of column." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "text", + "content": "the artificial label " + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "text", + "content": "th point " + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 47, + 324, + 201, + 335 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 345, + 287, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 345, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 102, + 345, + 287, + 377 + ], + "type": "interline_equation", + "content": "a _ {k} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} v _ {k} = \\emptyset \\\\ \\operatorname {M o d e} \\left(v _ {k}\\right) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {6}", + "image_path": "a31ee3c8d9ffeadfafa74ae13d9c18f5939bfe1324f96e84db45c099c10a9c13.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 387, + 286, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 286, + 398 + ], + "type": "text", + "content": "where the Mode operator returns the most frequent element." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "text", + "content": "The proposed PP2S module effectively leverages the segmentation capability of SAM to expand the sparsely annotated initial labels. These resulting artificial labels, denoted as " + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "inline_equation", + "content": "A = \\{a_k\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "text", + "content": ", can directly serve as a supervised signal for training. Notably, we have observed that training with " + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "text", + "content": " yields significant performance improvements compared to using initial " + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 399, + 287, + 494 + ], + "type": "text", + "content": ". The detailed experimental results are demonstrated in Section 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 267, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 267, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 267, + 516 + ], + "type": "text", + "content": "3.3. Region Exploration via Artificial Labeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 712 + ], + "type": "text", + "content": "As mentioned, the proposed PP2S effectively enhances the initial sparse annotations, resulting in a substantial improvement in segmentation performance. However, we also have recognized several limitations associated with this preprocessing approach. The PP2S-enhanced labels can serve as better guidance than the initial sparse labels. Nevertheless, they are still fixed during the entire training phase and thereby impose an upper bound on performance. Notably, we cannot guide the regions not covered by the 2D masks obtained within PP2S. Furthermore, even if we assume that the initially provided sparse annotations are perfectly correct, errors stemming from factors such as noises of camera projection matrices (Fig. 3 A, B) or inherent ambiguity of SAM (Fig. 3 C) are inevitable. While the voting system can mitigate some of these, it cannot completely eliminate them." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 307, + 87, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 526, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 526, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 526, + 84 + ], + "type": "text", + "content": "Algorithm 1 Region Exploring via Artificial Labeling" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": "Require: Segmentation model " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " , Training dataset " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{D} =" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\{(X,Y)\\}" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " , Learning rate " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " , Number of epochs " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \nEnsure: Optimized " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n1: for " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "(X,Y)" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "Y\\gets \\mathrm{PP2S}(X;X^L)" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n3: end for \n4: Initialize the model parameter " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n5: for " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "t\\gets 1,2,\\ldots ,T" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " do \n6: for " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "(X,Y)" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " do \n7: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "Y^{t - 1}\\gets Y" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "S\\gets \\mathrm{softmax}(f_{\\theta}(X))" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\hat{Y}\\gets \\mathrm{argmax}(S)" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "Q\\gets \\mathrm{QuerySampling}(X,S,Y^{t - 1})" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "A\\gets \\mathrm{PP2S}(X;Q)" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "Y^{t}\\gets \\mathrm{LabelUpdate}(Y^{t - 1},A,\\hat{Y})" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "Y\\gets Y^{t}" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n14: " + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\theta \\leftarrow \\theta -\\eta \\nabla_{\\theta}\\mathcal{L}(S,Y)" + }, + { + "bbox": [ + 307, + 87, + 545, + 315 + ], + "type": "text", + "content": " \n15: end for \n16: end for" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 325, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 456 + ], + "type": "text", + "content": "Then, how can we further improve label quality, especially during training? We draw inspiration from the concept of active learning. In active labeling, the model identifies a set of instances, known as a query, during training and requests human annotators (i.e., oracle) to label these points. Typically, the query consists of instances that confuse the model at a given timestep, making annotating the query highly effective for the model to learn decision boundaries. However, active learning requires manual annotation within the training loop, potentially making it even more expensive than WSPCSS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": "To address this challenge, we present the Region Exploration via Artificial Labeling (REAL) framework—a novel WSPCSS method utilizing the PP2S module as an artificial oracle, as depicted in Fig. 4. Specifically, the proposed PP2S module serves as an oracle in the REAL framework, responding to requests to annotate the query. Given that the artificial oracle is, in fact, a pretrained network, the REAL framework is completely relieved of the necessity for manual annotation, except for the initial sparse labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "However, unlike a human oracle in conventional methods, our artificial oracle lacks explicit semantic knowledge. Therefore, we cannot expect the PP2S module to directly label the classes for the confusing points. Instead, we focus on leveraging what PP2S can provide to us. PP2S cannot directly annotate the given points; however, it can precisely predict which points should be grouped into a segment. To fully utilize this attribute, we define the query " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "Q = \\{q_i\\}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " to consist of the confidently identified points, instead of ambiguous ones from the perspective of the model. Trusting the prediction of the model on these points, we can self-label the query with minimal risk. We represent this pro" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3724" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 92, + 542, + 328 + ], + "blocks": [ + { + "bbox": [ + 225, + 76, + 499, + 89 + ], + "lines": [ + { + "bbox": [ + 225, + 76, + 499, + 89 + ], + "spans": [ + { + "bbox": [ + 225, + 76, + 499, + 89 + ], + "type": "text", + "content": "Region Exploration via Artificial Labeling (REAL, Sec. 3.3)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 92, + 542, + 328 + ], + "lines": [ + { + "bbox": [ + 51, + 92, + 542, + 328 + ], + "spans": [ + { + "bbox": [ + 51, + 92, + 542, + 328 + ], + "type": "image", + "image_path": "9e3da3fe03aceecc7f9da1adc9efa253e4ec16ddbcbecec297306a825b13b677.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": "Figure 4. Visualization of the proposed Region Exploration via Artificial Labeling (REAL) framework. The objective is to dynamically enhance the quality of labels for weakly supervised point cloud semantic segmentation. In the REAL framework, the model " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": " takes the point cloud " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": " as input and predicts semantic segmentation logits " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": ". We then sample the most confident points from the unlabeled regions of the previous label " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": " based on the predictions " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": " and reliability scores " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": ". The sampled points then serve as prompts for the proposed PP2S module, producing artificial labels " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": ". Finally, by combining the model prediction and " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": ", we refine " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": " into the updated labels " + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "inline_equation", + "content": "Y^t" + }, + { + "bbox": [ + 46, + 334, + 545, + 401 + ], + "type": "text", + "content": " which serve as pseudo-ground truth for training the logits." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 416, + 286, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 286, + 440 + ], + "type": "text", + "content": "cess as QuerySampling in Algorithm 1, which will be thoroughly discussed in Section 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "content": "Subsequently, we request the PP2S module to act as an artificial oracle, annotating the unknown regions using the given query. For this, the self-labeled query serves as an input prompt for the PP2S module as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 498, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 498, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 128, + 498, + 287, + 512 + ], + "type": "interline_equation", + "content": "A = \\operatorname {P P 2 S} (X; Q), \\tag {7}", + "image_path": "7323c7437a23a48c4310826e009af292a9951adee0694ca0e2b69fb5a9ec53c8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "content": " represents the artificial label obtained by PP2S from the given query " + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "content": ". This approach is akin to using a set of sparsely annotated points " + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "(X^L)" + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "content": " for PP2S in the preprocessing phase, as described in Section 3.2. The primary distinction in the REAL framework is that the input prompt is self-labeled, not manually annotated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": "With the obtained artificial label " + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": ", we update the label of the previous step, " + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": ", into " + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "Y^t" + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": " may not be entirely reliable, as it could be more prone to noise compared to labels provided by a human oracle. This potential for noise could arise from inaccuracies in the self-labeled input query or the issues highlighted in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "To mitigate the risk of incorporating errors during the label-updating process, we devise a cautious approach using the prediction of the model. Specifically, if a discrepancy exists between the class assigned by the artificial label" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "content": " and the model's prediction " + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "content": " for a specific point, we exclude that point from the updating process. This conservative update process can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 460, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 460, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 372, + 460, + 545, + 493 + ], + "type": "interline_equation", + "content": "y _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} a _ {i} & \\text {i f} a _ {i} = \\hat {y} _ {i} \\\\ y _ {i} ^ {t - 1} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "47a1697f3f8a95ebd7fa8202f45fb42983e91a9f5d5aab02a624817cb74262d3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "text", + "content": "Finally, the model " + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "text", + "content": " is optimized by minimizing the cross-entropy loss between " + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "inline_equation", + "content": "Y^{t}" + }, + { + "bbox": [ + 305, + 500, + 544, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 532, + 497, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 497, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 497, + 545 + ], + "type": "text", + "content": "3.4. Query Sampling Strategy for REAL" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "type": "text", + "content": "In the REAL framework, query sampling is a critical component, as the quality of the artificial label depends on the quality of the input query. We posit that an effective query must meet two key criteria: 1) it should exhibit a high level of confidence, and 2) it should aid the model in acquiring new knowledge. To address the first criterion, we design a margin-based reliability metric for evaluating the confidence of the model's prediction for each point as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 656, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 656, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 368, + 656, + 545, + 669 + ], + "type": "interline_equation", + "content": "r _ {i} = \\operatorname {M a x} \\left(s _ {i}\\right) - \\operatorname {M a x} 2 \\left(s _ {i}\\right), \\tag {9}", + "image_path": "2f6850b49704be34b849f67426eef3379d3509d9513229dcac5c797ce33a015a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "s_i \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " is the predicted logit of the " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "th point " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " denotes its reliability. Max and Max2 are operators that return the maximum and the second maximum el" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3725" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 280, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 280, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 280, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 280, + 148 + ], + "type": "image", + "image_path": "ce20e8baf2ab7ef32c9826296e236d86c41d1390f15d7c285d93d24d00022ecf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 150, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 46, + 150, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 287, + 174 + ], + "type": "text", + "content": "Figure 5. Verification of the reliability metric. The points presenting higher reliability show higher semantic segmentation accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "type": "text", + "content": "ements of the input vector, respectively. We verify a robust correlation between the devised reliability metric and the accuracy of model predictions, as illustrated in Fig. 5. For instance, among points with reliability scores exceeding 0.5, the model correctly predicts the class for " + }, + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "type": "inline_equation", + "content": "84.5\\%" + }, + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "type": "text", + "content": " of them. We set the threshold for query sampling at 0.95, corresponding to around " + }, + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 46, + 182, + 288, + 266 + ], + "type": "text", + "content": " accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": "On the contrary, directly quantifying the utility of a query for the second criterion is inherently elusive. In this context, we propose an intuitive hypothesis: if the model has successfully leveraged the labels from the previous timestep, " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": ", the regions previously annotated in " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": " are expected to contribute less novel information during the current timestep, " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": ". Therefore, the focus of new queries should be on effectively exploring the regions not previously or incorrectly labeled in " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 409 + ], + "type": "text", + "content": "To this end, we narrow the sampling range of the query to the points that are predicted by the model differently from the previous label. We formulate the above constraint as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 415, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 415, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 61, + 415, + 287, + 430 + ], + "type": "interline_equation", + "content": "Z _ {c} = \\left\\{x _ {i} \\right\\} \\forall i \\text {s u c h t h a t} \\hat {y} _ {i} = c \\text {a n d} y _ {i} ^ {t - 1} \\neq c. \\tag {10}", + "image_path": "7efa0d3143bca6810422fac5a8f8b8061cd326edf6ec04321809a69376e84c1f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "Z_{c}" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": " is the mother set for sampling " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "q_{c}" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": ", the query point of the class " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": ". Considering the first criterion, we sample " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "q_{c}" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": " as the most reliable point among " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "Z_{c}" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": " is the set of " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "q_{c}" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": ". We formulate the above processes as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 487, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 487, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 94, + 487, + 287, + 502 + ], + "type": "interline_equation", + "content": "Q = \\operatorname {Q u e r y S a m p l i n g} (X, S, Y ^ {t - 1}). \\tag {11}", + "image_path": "4928d078394906a22eddae491f81f5d8e84ab8eb13bf6ca7213f71e52ac808df.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 527, + 172, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 172, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 172, + 540 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Datasets. We conduct experiments on S3DIS [3] and ScanNetV2 [8]. S3DIS comprises 6 areas with 272 rooms and 13 categories. We use the area 5 for evaluation following previous studies. ScanNetV2 includes 1613 3D scenes with 20 categories. We follow the official split (1201 training, 312 validation, and 100 test scenes). In addition to the point cloud data, the proposed REAL framework requires the data from cameras (i.e., images, depth maps, and camera projection matrices) to leverage SAM. The datasets inherently include such data since they are reconstructed from the sequence of RGBD frames. During training, we utilize the data from 48 and 17 cameras per scene on average for S3DIS and ScanNetV2, respectively. Note that our framework does not require 2D data at inference time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 71, + 546, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 546, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 546, + 137 + ], + "type": "text", + "content": "Table 1. A quantitative comparison between the initially sparse annotations (Initial) and the artificial labels from PP2S (PP2S). Pre, Rec, and # denote precision, recall, and the proportion of annotated points relative to the total number of points. The performance of the point cloud semantic segmentation model trained by each model is represented as Model mIoU. Every metric is in %." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 321, + 140, + 529, + 173 + ], + "blocks": [ + { + "bbox": [ + 321, + 140, + 529, + 173 + ], + "lines": [ + { + "bbox": [ + 321, + 140, + 529, + 173 + ], + "spans": [ + { + "bbox": [ + 321, + 140, + 529, + 173 + ], + "type": "table", + "html": "
LabelsmIoUPreRec#Model mIoU
Initial0.0041000.0040.00439.7
PP2S30.483.132.438.453.0
", + "image_path": "a8c7729b83e96010200762660c55b67a91460db50e949684482c083cafdf8138.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 187, + 424, + 258 + ], + "blocks": [ + { + "bbox": [ + 342, + 178, + 397, + 186 + ], + "lines": [ + { + "bbox": [ + 342, + 178, + 397, + 186 + ], + "spans": [ + { + "bbox": [ + 342, + 178, + 397, + 186 + ], + "type": "text", + "content": "(A) Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 187, + 424, + 258 + ], + "lines": [ + { + "bbox": [ + 320, + 187, + 424, + 258 + ], + "spans": [ + { + "bbox": [ + 320, + 187, + 424, + 258 + ], + "type": "image", + "image_path": "73ed7a6e9f304e8b527c4cfce59c7f270aed69a63f85c610edbae5ff17741ba5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 188, + 529, + 258 + ], + "blocks": [ + { + "bbox": [ + 427, + 178, + 523, + 187 + ], + "lines": [ + { + "bbox": [ + 427, + 178, + 523, + 187 + ], + "spans": [ + { + "bbox": [ + 427, + 178, + 523, + 187 + ], + "type": "text", + "content": "(B) Labels of the previous step" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 188, + 529, + 258 + ], + "lines": [ + { + "bbox": [ + 425, + 188, + 529, + 258 + ], + "spans": [ + { + "bbox": [ + 425, + 188, + 529, + 258 + ], + "type": "image", + "image_path": "b14a1b2f96268334df4edd870f80282eb8a837132e98706f11d130ffc8e1b5ba.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 259, + 527, + 269 + ], + "lines": [ + { + "bbox": [ + 427, + 259, + 527, + 269 + ], + "spans": [ + { + "bbox": [ + 427, + 259, + 527, + 269 + ], + "type": "text", + "content": "(D) Explored regions (setting II)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 320, + 269, + 424, + 339 + ], + "blocks": [ + { + "bbox": [ + 323, + 259, + 419, + 269 + ], + "lines": [ + { + "bbox": [ + 323, + 259, + 419, + 269 + ], + "spans": [ + { + "bbox": [ + 323, + 259, + 419, + 269 + ], + "type": "text", + "content": "(C) Explored regions (setting 1)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 269, + 424, + 339 + ], + "lines": [ + { + "bbox": [ + 320, + 269, + 424, + 339 + ], + "spans": [ + { + "bbox": [ + 320, + 269, + 424, + 339 + ], + "type": "image", + "image_path": "407bce03c378fee676c01e3a5955b31913472da376d66c02347faaa2ad656b8a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 341, + 365, + 348 + ], + "lines": [ + { + "bbox": [ + 329, + 341, + 365, + 348 + ], + "spans": [ + { + "bbox": [ + 329, + 341, + 365, + 348 + ], + "type": "text", + "content": "Prec: -0.09%" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 341, + 416, + 347 + ], + "lines": [ + { + "bbox": [ + 377, + 341, + 416, + 347 + ], + "spans": [ + { + "bbox": [ + 377, + 341, + 416, + 347 + ], + "type": "text", + "content": "Rec: +11.92%" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 341, + 471, + 347 + ], + "lines": [ + { + "bbox": [ + 433, + 341, + 471, + 347 + ], + "spans": [ + { + "bbox": [ + 433, + 341, + 471, + 347 + ], + "type": "text", + "content": "Prec: -0.03%" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 341, + 520, + 347 + ], + "lines": [ + { + "bbox": [ + 485, + 341, + 520, + 347 + ], + "spans": [ + { + "bbox": [ + 485, + 341, + 520, + 347 + ], + "type": "text", + "content": "Rec: +5.58%" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 357, + 546, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 357, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 546, + 401 + ], + "type": "text", + "content": "Figure 6. Validation of our query sampling strategy. Setting I samples the queries following our strategy while setting II samples from the entire point set. We present the improvement obtained by artificial labels in terms of precision (Prec) and recall (Rec)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": "Evaluation protocols. Previous studies [23, 24, 42, 48] have conducted experiments with various levels of sparsity. We believe that the most cost-efficient yet practical setting involves labeling only one point per object. Accordingly, our experiments mainly target the 1pt weak setting, which corresponds to " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "0.004\\%" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " of points in S3DIS. For a fair comparison, we also conduct experiments under the " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " setting, which is the most widely used. For ScanNetV2, we follow the conventional works, using 20pt per scene. The overall performance is assessed on all points within the test set. Following the standard practice, we employ the mean Intersection over Union (mIoU) as our quantitative metric. Implementation details. We choose PTv2 [37] network as our main backbone. To verify the effect of the backbone, we additionally test Closer [25] backbone for the " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "0.004\\%" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " setting on S3DIS. Training details are the same as standard settings of the backbones. We employ a pretrained ViT-H model for SAM. Further details are in the Supp." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 635, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 416, + 647 + ], + "type": "text", + "content": "4.2. Analysis on REAL" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "text", + "content": "PP2S-based preprocessing. We first validate the preprocessing strategy built upon the proposed PP2S. This strategy aims to enhance the initial sparse annotations into artificial labels by harnessing the promptable segmentation capability of SAM via projection. Table 1 provides various met" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3726" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 69, + 521, + 292 + ], + "blocks": [ + { + "bbox": [ + 63, + 69, + 521, + 292 + ], + "lines": [ + { + "bbox": [ + 63, + 69, + 521, + 292 + ], + "spans": [ + { + "bbox": [ + 63, + 69, + 521, + 292 + ], + "type": "image", + "image_path": "af5fcc7a63c0192ddb3a1e7c1190f840d721cba417e78fb5f07c0b23494c455c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "text", + "content": "Figure 7. Progressive visualization of the behavior of the proposed REAL framework throughout training. " + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "text", + "content": " is the artificial label, and " + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "text", + "content": " denotes the subsequent label updated at timestep " + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 295, + 547, + 329 + ], + "type": "text", + "content": ". The highlighted boxes indicate newly explored regions by artificial oracle, which were unknown (red, blue) or wrongly labeled (green) in earlier timesteps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 339, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 288, + 460 + ], + "type": "text", + "content": "rics of the initial labels and those of the labels enhanced by PP2S. Furthermore, we compare the performance of the point cloud semantic segmentation model when trained with the original sparse annotations and artificial labels enhanced by PP2S. The results demonstrate that our PP2S significantly improves the quality of the label, consequently boosting the model's performance. This analysis supports the validity of PP2S as a bridge for transferring the segmentation capability of SAM into the realm of weakly supervised point cloud semantic segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "Query sampling strategy. We conduct ablation studies regarding the proposed query sampling strategy introduced in Section 3.4. In this section, we compare two settings: (I) using the proposed strategy sampling from the narrowed mother set of Equ. 10 and (II) sampling the most reliable point among every point without any constraint. Figure 6 provides both qualitative and quantitative comparisons between (I) and (II), displaying the improvement achieved by the artificial labels of each setting. (I) outperforms (II) in terms of recall while maintaining precision. This distinction primarily arises from the fact that most query points of (II) are already labeled in " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", limiting their potential for exploring into unknown regions. On the other hand, the query of (I) is explicitly restricted to represent genuinely novel information compared with previous labels " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "Y^{t-1}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", even without compromising precision. This implies that the proposed query sampling strategy yields better labels, facilitating the model to learn segmentation effectively. Accordingly, the performance with our query sampling strategy (62.7%) is significantly higher than without using it (59.8%). Additional results are in Supp." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 339, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 448 + ], + "type": "text", + "content": "Active label enhancement. To demonstrate the behavior of the proposed REAL framework intuitively, Fig. 7 depicts progressive visualization throughout training. The labels are significantly improved as training proceeds, thanks to the artificial labels from our artificial oracle. Notably, the highlighted boxes indicate newly explored regions, which were unknown (red, blue) or wrongly labeled (green) by the labels in earlier timesteps. The REAL framework effectively integrates these regions into the current label." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 455, + 491, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 491, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 491, + 468 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": "We compare our method with SoTA methods on S3DIS, as shown in Table 2. The proposed REAL significantly enhances performance across all settings. Specifically, REAL with PTv2 [37] outperforms CPCM by " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " in mIoU under the " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " setting, which is even comparable to the SoTA results of the " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " setting. Furthermore, REAL with PTv2 surpasses the other SoTA methods under the " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " setting, using only 1pt for each object " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "(0.004\\%)" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": ". We also observe a significant performance improvement when employing the REAL in Closer [25]. These results imply that the gain of REAL is not restricted by the backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "content": "Besides, Fig. 8 provides a qualitative comparison. Notably, the semantic confusion in the baseline is remarkably improved in ours. Furthermore, REAL demonstrates a meaningful enhancement in its ability to group the points of each object precisely. Consequently, REAL achieves remarkable semantic segmentation results even in complicated scenes, using only " + }, + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "inline_equation", + "content": "0.004\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "content": " of annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Furthermore, we evaluate the performance of our method against SoTA methods on ScanNetV2 [8], as detailed in Ta" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3727" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 96, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Comparisons of the proposed methods with SoTA methods on S3DIS test set. Every metric is in mIoU (%). Bold number represents the best result. † denotes the results from our reimplementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 96, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 48, + 96, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 545, + 315 + ], + "type": "table", + "html": "
SettingsMethodsmIoUceil.floorwallbeamcol.wind.doorchairtablebook.sofaboardclutter
FullyRandLA [14]62.491.295.780.10.025.262.347.475.883.260.870.865.254.0
KPConv [35]67.192.897.382.40.023.958.069.091.081.575.375.466.758.9
RFCR [10]68.794.298.384.30.028.562.471.292.082.676.171.171.661.3
Closer† [25]66.093.298.281.20.024.849.969.390.480.472.074.668.055.8
PTv2† [38]70.993.898.586.10.029.560.578.093.281.876.679.283.061.7
1%SPT [47]61.891.596.980.60.018.258.147.275.885.765.368.965.050.2
PSD [48]63.592.397.780.70.027.856.262.578.784.163.170.458.953.2
HybridCR [23]65.392.593.982.60.024.264.463.278.381.769.074.468.256.5
0.03%PSD [48]48.287.996.062.10.020.649.340.955.161.943.950.727.331.1
HybridCR [23]51.585.491.965.90.018.051.434.263.878.352.459.629.939.0
0.02%MIL [42]51.486.693.275.00.029.345.346.760.562.356.547.533.732.2
CPCM [24]62.392.695.679.40.017.849.359.485.775.669.160.768.255.8
PTv2† [38]55.180.594.773.90.021.742.938.981.459.262.552.966.142.1
+PP2S60.079.093.177.50.131.549.761.470.073.769.956.971.045.6
+REAL65.286.995.880.20.027.060.377.872.379.970.864.777.454.1
0.004%Closer† [25]38.782.592.469.40.115.023.335.650.947.80.319.137.329.1
+PP2S44.962.988.867.90.011.729.244.746.463.253.525.750.738.6
+REAL57.879.097.770.80.124.450.460.184.564.664.866.839.948.7
PTv2† [38]39.776.687.265.40.09.430.622.361.750.341.714.622.434.1
+PP2S53.065.993.168.90.027.342.755.868.463.758.252.550.541.8
+REAL62.784.095.780.30.031.157.463.775.576.968.770.065.446.8
", + "image_path": "1ea0e42bfd15338e2df459265b27929a742d14610a0b5b4ef7ad97cbb1c34360.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 319, + 283, + 420 + ], + "blocks": [ + { + "bbox": [ + 50, + 319, + 283, + 420 + ], + "lines": [ + { + "bbox": [ + 50, + 319, + 283, + 420 + ], + "spans": [ + { + "bbox": [ + 50, + 319, + 283, + 420 + ], + "type": "image", + "image_path": "de220b6e7ad44dc8499b83f51eeacbfe0570ec27bd784535fe63ec087127cb53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 425, + 287, + 468 + ], + "lines": [ + { + "bbox": [ + 47, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 287, + 468 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons between the semantic segmentation results of the baseline and ours on S3DIS under " + }, + { + "bbox": [ + 47, + 425, + 287, + 468 + ], + "type": "inline_equation", + "content": "0.004\\%" + }, + { + "bbox": [ + 47, + 425, + 287, + 468 + ], + "type": "text", + "content": " setting. The displayed boxes indicate the regions where significant improvements are achieved by the proposed method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "type": "text", + "content": "ble 3. Our method demonstrates superior performance on both the validation and test splits. Notably, REAL outperforms CPCM by " + }, + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "type": "text", + "content": " in the validation set and " + }, + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 47, + 478, + 287, + 550 + ], + "type": "text", + "content": " in the test set. These results imply that our method effectively utilizes the potential of weak labels. The qualitative results for ScanNetV2 can be found in Supp." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 561, + 119, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 119, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 119, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 288, + 713 + ], + "type": "text", + "content": "Expensive and labor-intensive point-wise annotation hinders the practical application of point cloud semantic segmentation. While conventional WSPCSS methods have exhibited promising results, the fixed input sparse label potentially constrains the achievable performance. In pursuit of a novel research direction, we propose the utilization of an additional source of information, introducing the Region Exploration via Artificial Labeling (REAL) framework. This framework aims to harness the power of the image foundational model, SAM, within the context of active learning. To establish the connection between the 2D model" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 343, + 345, + 511, + 450 + ], + "blocks": [ + { + "bbox": [ + 305, + 319, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 340 + ], + "type": "text", + "content": "Table 3. Comparisons of the proposed methods with SoTA methods on ScanNetV2 val/test set. The results are in mIoU." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 345, + 511, + 450 + ], + "lines": [ + { + "bbox": [ + 343, + 345, + 511, + 450 + ], + "spans": [ + { + "bbox": [ + 343, + 345, + 511, + 450 + ], + "type": "table", + "html": "
SettingsMethodsVal.Test
FullyKPConv [35]-68.4
RFCR [10]-70.2
1%SPT [47]-51.1
PSD [48]-54.7
HybridCR [23]56.956.8
20 ptsMIL [42]57.854.4
CPCM [24]62.762.8
Ours (REAL)66.365.5
", + "image_path": "f539e440992a0f7893e31922b83d38fb99152b66684ac493a50097bc3fb74c5d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 458, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 613 + ], + "type": "text", + "content": "and 3D data, we designed the Projection-based Point-to-Segment (PP2S) module. The PP2S is employed as an artificial oracle in the REAL framework, dynamically enhancing labels during training. Although using an artificial oracle eliminates the burden of manual annotation, it introduces some drawbacks due to the model's imperfections. As a remedy, we devised several strategies for query sampling and label updating. Through extensive experiments, we demonstrated the working logic of our framework in detail. Finally, the proposed REAL framework achieved new SoTA on various datasets, surpassing existing works. We believe that our approach pioneers the effective integration of the foundational image model into the realm of WSPCSS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3728" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4981-4990, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 286, + 246 + ], + "type": "text", + "content": "[3] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "type": "text", + "content": "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 379 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 286, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 445 + ], + "type": "text", + "content": "[6] Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5291-5301, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 286, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 500 + ], + "type": "text", + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 286, + 557 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 612 + ], + "type": "text", + "content": "[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 677 + ], + "type": "text", + "content": "[10] Jingyu Gong, Jiachen Xu, Xin Tan, Haichuan Song, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Omni-supervised point cloud segmentation via gradual receptive field component reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11673-11682, 2021. 1, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[11] Shizhan Gong, Yuan Zhong, Wenao Ma, Jinpeng Li, Zhao Wang, Jingyang Zhang, Pheng-Ann Heng, and Qi Dou. 3dsam-adapter: Holistic adaptation of sam from 2d to 3d" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "text", + "content": "for promptable medical image segmentation. arXiv preprint arXiv:2306.13465, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[12] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "text", + "content": "[13] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "text", + "content": "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 276, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 331 + ], + "type": "text", + "content": "[15] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In European Conference on Computer Vision, pages 600-619. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 333, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 387 + ], + "type": "text", + "content": "[16] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "text", + "content": "[17] Zeyu Hu, Xuyang Bai, Runze Zhang, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Lidal: Inter-frame uncertainty based active learning for 3d lidar semantic segmentation. In European Conference on Computer Vision, pages 248-265. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[18] Jihun Kim, Hyeokjun Kwon, Yunseo Yang, and Kuk-Jin Yoon. Learning point cloud completion without complete point clouds: A pose-aware approach. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14157-14167. IEEE, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 546, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 589 + ], + "type": "text", + "content": "[20] Hyeokjun Kweon and Kuk-Jin Yoon. Joint learning of 2d-3d weakly supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:30499–30511, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "type": "text", + "content": "[21] Hyeokjun Kweon, Sung-Hoon Yoon, Hyeonseong Kim, Daehe Park, and Kuk-Jin Yoon. Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6994-7003, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[22] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3729" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[23] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14930-14939, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 206 + ], + "type": "text", + "content": "[24] Lizhao Liu, Zhuangwei Zhuang, Shangxin Huang, Xunlong Xiao, Tianhang Xiang, Cen Chen, Jingdong Wang, and Mingkui Tan. Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18413-18422, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 208, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 273 + ], + "type": "text", + "content": "[25] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 326-342. Springer, 2020. 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 274, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 330 + ], + "type": "text", + "content": "[26] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[27] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1190-1199, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "type": "text", + "content": "[29] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[30] Damien Robert, Bruno Vallet, and Loic Landrieu. Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5575-5584, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "text", + "content": "[31] Shenghai Rong, Bohai Tu, Zilei Wang, and Junjie Li. Boundary-enhanced co-training for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19574-19584, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[32] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceed-" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "text", + "content": "[34] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "text", + "content": "[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "text", + "content": "[36] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "text", + "content": "[38] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 355, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 410 + ], + "type": "text", + "content": "[39] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "type": "text", + "content": "[40] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "type": "text", + "content": "[41] Zongyi Xu, Bo Yuan, Shanshan Zhao, Qianni Zhang, and Xinbo Gao. Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18098-18108, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "text", + "content": "[42] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "text", + "content": "[43] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2d-3d interlaced transformer for point cloud segmentation with scene-level supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 977–987, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 657 + ], + "type": "text", + "content": "[44] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[45] Sung-Hoon Yoon, Hyeokjun Kweon, Jegyeong Cho, Shinjeong Kim, and Kuk-Jin Yoon. Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "3730" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "October 23-27, 2022, Proceedings, Part XXIX, pages 326-344. Springer Nature Switzerland Cham, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[46] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: Zero-shot 3d object detection via segment anything model. arXiv preprint arXiv:2306.02245, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[47] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3421-3429, 2021. 1, 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[48] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15520-15528, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "3731" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_content_list.json b/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e3528154ded7e2f58e16f8f45f54a3d92e599b4a --- /dev/null +++ b/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_content_list.json @@ -0,0 +1,1704 @@ +[ + { + "type": "text", + "text": "Weakly Supervised Video Individual Counting", + "text_level": 1, + "bbox": [ + 250, + 130, + 720, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyan Liu1 Guorong Li1,3* Yuankai Qi2 Ziheng Yan1 \nZhenjun Han1 Anton van den Hengel5 Ming-Hsuan Yang6 Qingming Huang1,3,4 \n1 University of Chinese Academy of Science; 2 Macquarie University;", + "bbox": [ + 148, + 179, + 839, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Lab of Big Data Mining and Knowledge Management, UCAS, China; 4 Key Lab of Intell. Info. Process., Inst. of Comput. Tech., CAS;", + "bbox": [ + 184, + 233, + 784, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Australian Institute for Machine Learning, The University of Adelaide; $^{6}$ University of California at Merced.", + "bbox": [ + 197, + 268, + 769, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Individual Counting (VIC) aims to predict the number of unique individuals in a single video. Existing methods learn representations based on trajectory labels for individuals, which are annotation-expensive. To provide a more realistic reflection of the underlying practical challenge, we introduce a weakly supervised VIC task, wherein trajectory labels are not provided. Instead, two types of labels are provided to indicate traffic entering the field of view (inflow) and leaving the field view (outflow). We also propose the first solution as a baseline that formulates the task as a weakly supervised contrastive learning problem under group-level matching. In doing so, we devise an end-to-end trainable soft contrastive loss to drive the network to distinguish inflow, outflow, and the remaining. To facilitate future study in this direction, we generate annotations from the existing VIC datasets SenseCrowd and CroHD and also build a new dataset, UAVVIC. Extensive results show that our baseline weakly supervised method outperforms supervised methods, and thus, little information is lost in the transition to the more practically relevant weakly supervised task. The code and trained model can be found at CGNet.", + "bbox": [ + 75, + 371, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 719, + 209, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Crowd Counting (VCC) has garnered much interest due to its broad range of practical applications, particularly in crowd safety management. This task requires a model", + "bbox": [ + 75, + 744, + 468, + 792 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec0c13051ce0b6be7d462d48dfc100d9760e3f42a1a825e80a15f20e219a8714.jpg", + "image_caption": [ + "(a) Annotation of VIC method" + ], + "image_footnote": [], + "bbox": [ + 501, + 340, + 687, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e46e3d19e326bba2e088530a1881144304acb8c6648f7b78c3016a1e913fdfd4.jpg", + "image_caption": [ + "(b) Annotation of our method", + "Figure 1. Existing VIC method requires a unique label indicating the position of each human in each frame, and these labels are consistent across frames. Weakly supervised VIC only requires labels indicating each human position and whether they are an inflow/outflow pedestrian (purple/yellow in (b)) in the current frame. The transition from individual-level labels to identity-agnostic group-level labels represents a significant reduction in the labeling effort." + ], + "image_footnote": [], + "bbox": [ + 707, + 340, + 893, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to count the number of people in each frame of a video. A limitation of VCC [12, 51, 53] is that it gives an imprecise estimate of the number of unique individuals appearing in a video sequence, as people are counted multiple times if they appear in several frames. To overcome this drawback, Video Individual Counting (VIC) was introduced, wherein a method must count the total number of people with unique identities appearing in a video sequence.", + "bbox": [ + 496, + 609, + 893, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The apparent approach to crowd counting is to count the people who appear in the first frame and add the number of people who come into the camera's field of view in later frames (the inflow). Following this principle, Han et al. [14] devised DRNet, the only existing published method for VIC, identifying repeated observations of individuals across consecutive frames based on their appearance, simultaneously predicting inflow and outflow. Although the number of outflows does not contribute to the final count, DRNet finds it helps determine individual associations. Han et al. [14] also explored using conventional multi-object track", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 803, + 222, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants 62272438, 62236008, U21B2038, and 61931008, in part by the Key Deployment Program of the Chinese Academy of Sciences under Grant KGFZD145-23-18, and in part by the Fundamental Research Funds for Central Universities (E2ET1104). Yuankai Qi, Anton van den Hengel, and Ming-Hsuan Yang are not supported by the aforementioned fundings.", + "bbox": [ + 75, + 815, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19228", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing (MOT) methods [7, 41, 44, 54, 56] to tackle the VIC problem. Unfortunately, it turns out MOT methods suffer from poor accuracy. More notably, DRNet and MOT-based methods require trajectory labels (or similar individual association labels) to supervise identity association, which is highly annotation-expensive.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is that counting people in the previous and current frames does not require accurate identity associations for those appearing in both frames. For example, as long as we can predict individual $15\\sim 17$ as entering in Fig. 1a, even if we arbitrarily associate individuals $0\\sim 14$ in frame $i$ to individuals $0\\sim 14$ in frame $i$ , we can still correctly infer the crowd counts. We can still count crowds if we relax individual-level identity associations to group-level associations and do not require trajectory annotations. Thus, we just need to annotate inflow and outflow people (then we can derive the individual exists in two frames), which reduces annotation costs compared to creating individual pairwise target associations for each observed pedestrian between neighboring frames. We name crowd counting with such annotations as Weakly supervised Video Individual Counting (WVIC).", + "bbox": [ + 75, + 191, + 470, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the WVIC task, we propose a benchmark method based on Contrastive learning with Group-level matching, namely CGNet. We design a soft contrastive loss to drive the network to learn discriminative representations that can facilitate identifying the required group associations and, thus, the inflow. Moreover, to better represent each individual, we design a memory-based individual count predictor, where historical templates of individuals are stored and updated in memory to enhance the robustness of association during inference. Considering that the two existing datasets for VIC are all captured by static cameras, we collect a weakly supervised VIC dataset based on moving UAVs, named UAVVIC. This dataset provides about 400,000 inflow/outflow and bounding box labels for four categories: pedestrian, car, bus, and van. Although devised for crowd counting, our proposed baseline method also performs well in counting other objects.", + "bbox": [ + 75, + 441, + 468, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 96, + 710, + 351, + 724 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose WVIC, a weakly supervised video individual counting task. This task does not require expensive per-target trajectory annotations and only requires two types of identity-agnostic annotations.", + "- We automatically reannounce two existing datasets, CroHD and SenseCrowd, and collect a new dataset, UAVVIC, to pave the way for future studies.", + "- We propose a strong baseline, CGNet, equipped with a newly designed group level matching soft contrastive loss, performing favorably against the supervised methods on the three datasets mentioned above." + ], + "bbox": [ + 78, + 734, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 648, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Crowd Counting (VCC) estimates the number of people in each video frame. Most of the existing VCC methods [2, 5, 6, 11, 12, 30, 51, 53, 58] can be divided into two categories according to their solved problems: region of interest (ROI) and line of interest (LOI). ROI methods [11, 12, 51, 53] detect pedestrians within a specific region (or the whole image), and they focus on leveraging temporal context information to improve the prediction of the current frame. LSTN [12] models the group flow of crowds in local regions. Xiong et al. in [53] fuse history frame features using a ConvLSTM. TAN [51] explores context information from adjacent density maps. CLRNet proposes a local self-attention module to help the model focus on highly related regions in adjacent frames. ROI methods can not be used in the VIC task. LOI methods [2, 5, 6, 30, 57, 58] counts people passing through a specific line. Most of the existing LOI methods [2, 5, 6] apply blobs to crowds, counting objects when the whole blobs cross the line. Ma et al. in [30] sample fixed line width areas within temporal image slices and accumulated crowds in this area. Zheng et al. [58] instead sum the number of people within neighboring blocks near a line based on local velocity. Zhao et al. in [57] train a crowd velocity map predictor with the help of trajectory labels. In practice, LOI methods perform poorly on the VIC task as people can enter or leave the field of view in any direction, making it difficult to find a suitable virtual line.", + "bbox": [ + 496, + 114, + 890, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple Object Tracking (MOT) aims to predict the trajectories for multiple targets in a video. Object association is a main challenge in MOT, establishing correspondences between objects across frames to generate targets' trajectories. Existing object association methods in MOT include Probabilistic Data Association (PDA) [36, 37], Joint Probabilistic Data Association (JPDA) [16, 23, 29, 40], and graph matching methods like the Hungarian algorithm [3, 50]. Recent research has started to use more powerful feature representations to enhance the robustness of data association. Deep learning-based methods like LSTM can model temporal information to handle long-term occlusion [20, 34, 45]. Many works [4, 15, 46] also focus on the association of tracklets. Specifically, they associate multiple tracklets over longer periods using a graph model to combine the final trajectory, which can relieve the effects of lost tracked targets. These methods perform poorly on the VIC task, however, due to the inevitable ID switching problem [14]. Additionally, they require trajectory labels to train their models.", + "bbox": [ + 496, + 523, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Individual Counting (VIC). To the best of our knowledge, DRNet [14] is the only method for the VIC task. It predicts the number of initial pedestrians and matches pedestrians in adjacent frames by adapting a differentiable optimal transport loss. Although the VIC task only requires the total count in the video, DRNet needs individual", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19229", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matching between frames. WVIC avoids the requirement for such matching annotation, which enables easy expansion of dataset scales and thus facilitates future large deep learning models.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. A Benchmark Method", + "text_level": 1, + "bbox": [ + 76, + 162, + 289, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of Weakly-supervised Video Individual Counting (WVIC) is to predict the number of unique individuals $(\\hat{N})$ in a video provided a series of its frames with $\\delta$ as the sampling rate. Given a video $\\mathbf{V}$ consisting of $T$ frames $\\mathbf{V}^1, \\dots, \\mathbf{V}^T$ , the label information comprises coordinate/inflow/outflow data for each individual. Specifically, letting $\\mathbf{P}^i = [p_1^i, \\dots, p_{n_i}^i]$ denote the annotation of the coordinates of the center of the people in frame $\\mathbf{V}^i$ , we sample the frame pairs $\\{(\\mathbf{V}^1, \\mathbf{V}^{\\delta + 1}), (\\mathbf{V}^{\\delta + 1}, \\mathbf{V}^{2\\delta + 1}), \\dots, (\\mathbf{V}^{T - \\delta}, \\mathbf{V}^T)\\}$ , where $\\delta$ is an integer frame offset. The weakly supervised inflow label $\\mathbf{I}^i = [\\mathbf{I}_1^i, \\dots, \\mathbf{I}_{n_i}^i]$ and outflow label $\\mathbf{O}^i = [\\mathbf{O}_1^i, \\dots, \\mathbf{O}_{n_i}^i]$ . If the $u$ -th person in frame $i$ doesn't appear in $\\mathbf{V}^{i - \\delta}$ , then $\\mathbf{I}_u^i = 1$ ; otherwise, $\\mathbf{I}_u^i = 0$ . For $\\mathbf{O}^i$ , if the $u$ -th person in frame $V^i$ doesn't appear in $\\mathbf{V}^{i + \\delta}$ , then $\\mathbf{O}_u^i = 1$ ; otherwise $\\mathbf{O}_u^i = 0$ . If, by a slight abuse of notation, we refer to people by their coordinates in particular images, we see that the set $\\{p_u^i | \\mathbf{O}_u^i = 0\\}$ and $\\{p_u^{i + \\delta} | \\mathbf{I}_u^i = 0\\}$ represent the people that appear both in $\\mathbf{V}^i$ and $\\mathbf{V}^{i + \\delta}$ .", + "bbox": [ + 75, + 188, + 472, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Weakly Supervised Representation Learning", + "text_level": 1, + "bbox": [ + 76, + 483, + 455, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the inference pipeline of our CGNet consists of three components: an image-level locator to generate coordinates for pedestrians, an encoder to produce representations based on pedestrians' coordinates, and a Memory-based individual Count Predictor (MCP) to predict the final inflow count based on the representations. The locator is trained independently with the coordinate annotations, and existing image crowd localization networks such as FIDT [25] can be used. The encoder is trained with our Weakly Supervised Representation Learning method (WSRL) to extract the discriminative features of each individual. It can be any feature extractor, such as ConvNext [26]. MCP does not need training.", + "bbox": [ + 75, + 506, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Provided two $\\delta$ -adjacent frames $\\mathbf{V}^i$ and $\\mathbf{V}^j$ and the corresponding annotations $\\mathbf{P}^i$ , $\\mathbf{P}^j$ , $\\mathbf{O}^i$ and $\\mathbf{I}^j$ , we group the individuals into four sets: $\\mathbf{X} = \\{p_u^i | \\mathbf{O}_u^i = 0\\}$ , $\\mathbf{Y} = \\{p_u^j | \\mathbf{I}_u^j = 0\\}$ , denoting shared individuals in the previous frame and the current frame, respectively; $\\hat{\\mathbf{X}} = \\mathbf{P}^i - \\mathbf{X}$ and $\\hat{\\mathbf{Y}} = \\mathbf{P}^j - \\mathbf{Y}$ , denoting outflow individuals and inflow individuals, respectively. An encoder learns representations $\\mathbf{F}$ for each pedestrian based on the crops around those coordinates. Specifically, to generate the $u$ -th feature $f_u^i$ in $\\mathbf{F}$ , if the locator outputs points, we extract a rectangle patch of size $96 \\times 64$ centered at the predicted coordinate $p_u^i$ , and if the outputs are bounding boxes, we directly crop the bounding box from the original picture. Each cropped patch is", + "bbox": [ + 75, + 703, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resized to $224 \\times 224$ and fed into the encoder, a ConvNext-S [26], generating a feature of $7 \\times 7 \\times 768$ , which is flattened and normalized into a 1D vector $f_{u}^{i}$ to represent the final feature for the individual:", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {F} _ {\\mathbf {X}} ^ {i} = \\operatorname {E n c o d e r} (\\mathbf {X}), \\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i} = \\operatorname {E n c o d e r} (\\hat {\\mathbf {X}}) \\\\ \\bar {\\mathbf {F}} ^ {i} = \\bar {\\mathbf {F}} _ {\\hat {\\mathbf {X}}} + \\left(\\bar {\\mathbf {H}}\\right) \\bar {\\mathbf {F}} ^ {i} = \\bar {\\mathbf {F}} _ {\\hat {\\mathbf {H}}} + \\left(\\hat {\\mathbf {H}}\\right) \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 154, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathbf {Y}} ^ {j} = \\operatorname {E n c o d e r} (\\mathbf {Y}), \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} = \\operatorname {E n c o d e r} (\\hat {\\mathbf {Y}})\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 176, + 834, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, a similarity matrix $\\mathbf{S} \\in \\mathbb{R}^{n_i \\times n_j}$ between $\\mathbf{F}^i = \\left[\\mathbf{F}_{\\mathbf{X}}^i, \\mathbf{F}_{\\hat{\\mathbf{X}}}^i\\right]$ and $\\mathbf{F}^j = \\left[\\mathbf{F}_{\\mathbf{Y}}^j, \\mathbf{F}_{\\hat{\\mathbf{Y}}}^j\\right]$ is computed by:", + "bbox": [ + 500, + 202, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\left[ \\begin{array}{l l} \\left(\\mathbf {F} _ {\\mathbf {X}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\mathbf {Y}} ^ {j} & \\left(\\mathbf {F} _ {\\mathbf {X}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} \\\\ \\left(\\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\mathbf {Y}} ^ {j} & \\left(\\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} \\end{array} \\right] = \\left[ \\begin{array}{l l} \\mathbf {S} _ {0} & \\mathbf {S} _ {1} \\\\ \\mathbf {S} _ {2} & \\mathbf {S} _ {3} \\end{array} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 253, + 890, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The matrix $\\mathbf{S}$ is divided into four parts: $\\mathbf{S}_0\\in \\mathbb{R}^{m_{i,j}\\times m_{i,j}}$ $\\mathbf{S}_1\\in \\mathbb{R}^{m_{i,j}\\times (n_j - m_{i,j})}$ $\\mathbf{S}_2\\in \\mathbb{R}^{(n_i - m_{i,j})\\times m_{i,j}}$ $\\mathbf{S}_3\\in$ $\\mathbb{R}^{(n_i - m_{i,j})\\times (n_j - m_{i,j})}$ , where $m_{i,j} = (\\mathbf{1} - \\mathbf{I}^{\\mathbf{j}})^T(\\mathbf{1} - \\mathbf{O}^i)$ denotes the number of shared individuals in $\\mathbf{V}^i$ and $\\mathbf{V}^j$ . $\\mathbf{S}_0$ is the similarity matrix of $\\mathbf{X}$ and $\\mathbf{Y}$ , while $\\mathbf{S}_1$ , $\\mathbf{S}_2$ and $\\mathbf{S}_3$ are the similarity matrix of $\\mathbf{X}$ and $\\hat{\\mathbf{Y}}$ , of $\\hat{\\mathbf{X}}$ and $\\mathbf{Y}$ , of $\\hat{\\mathbf{X}}$ and $\\hat{\\mathbf{Y}}$ , respectively.", + "bbox": [ + 496, + 299, + 890, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To pull the matched groups (X and Y) closer and push away individual pairs from unmatched groups, we propose a weakly supervised Group-level Matching Loss (GML) to constrain different parts of S.", + "bbox": [ + 496, + 406, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Constraint for $\\mathbf{S}_0, \\mathbf{S}_1, \\mathbf{S}_2$ . $\\mathbf{S}_0$ denotes the similarity among individuals shared between two frames, and it does not matter whether the individuals are correctly matched as long as we can assign them a one-to-one match. As for $\\mathbf{S}_1$ and $\\mathbf{S}_2$ , they are the similarities of different identities and thus should be zero.", + "bbox": [ + 496, + 469, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we introduce a latent variable matrix $\\Omega \\in \\mathbb{R}^{m_{i,j}\\times m_{i,j}}$ , where the $(u,v)$ -th element $\\Omega_{u,v}$ denote the probability that the $u$ -th pedestrian in $\\mathbf{X}$ is matched with $v$ -th pedestrian in $\\mathbf{Y}$ . To increase the similarity between the pairs that have a higher matching probability and decrease the similarity between the pairs that have a lower matching probability, inspired by [32, 39], we define the soft contrastive loss as:", + "bbox": [ + 496, + 560, + 890, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s c o n} (i) = \\min _ {\\Omega} - \\sum_ {u = 1} ^ {m _ {i, j}} \\sum_ {v = 1} ^ {m _ {i, j}} \\Omega_ {u, v} C _ {u, v}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 683, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad \\mathbf {1} _ {m _ {i, j}} ^ {T} \\Omega = \\mathbf {1}, \\quad \\Omega \\mathbf {1} _ {m _ {i, j}} = \\mathbf {1}, 0 \\leq \\Omega_ {u, v} \\leq 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 726, + 859, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C_{u,v}$ denotes the contrastive similarity [10], and is calculated as,", + "bbox": [ + 500, + 750, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {u, v} = \\frac {e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v}}}{e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v}} + \\sum_ {1 \\leq u ^ {\\prime} \\leq n _ {i}} ^ {u ^ {\\prime} \\neq u} e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u ^ {\\prime} , v}} + \\sum_ {1 \\leq v ^ {\\prime} \\leq n _ {j}} ^ {v ^ {\\prime} \\neq v} e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v ^ {\\prime}}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 784, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ is a temperature hyper-parameter [52] and $\\mathbf{S}_{u,v} = \\langle f_u^i,f_v^j\\rangle$ is the $(u,v)$ -th element of $\\mathbf{S}$ , calculated by the $u$ -th feature in $\\mathbf{F}^i$ and the $v$ -th feature in $\\mathbf{F}^j$ . As $\\Omega_{u,v}$ describes the matching probability, our soft contrastive loss", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/899ece5f0975dd10996ad6b68d31afe5e4cd176984664a9df7e592366177ed50.jpg", + "image_caption": [ + "Figure 2. The inference pipeline of our CGNet and the weakly supervised representation learning method (WSRL). The pipeline comprises a frame-level crowd locator, an encoder, and a Memory-based individual count predictor (MCP). The locator predicts the coordinates for pedestrians. The encoder generates representations for each individual, and MCP predicts inflow counts and updates the individual templates stored in the memory. To pull the matched groups $(\\mathbf{X}$ and $\\mathbf{Y})$ closer and push away individual pairs from unmatched groups, WSRL exploits inflow and outflow labels to optimize the encoder with a novel Group level Matching Loss (GML), which consists of a soft contrastive loss $(\\mathcal{L}_{scon})$ and a hinge loss $(\\mathcal{L}_{hinge})$ ." + ], + "image_footnote": [], + "bbox": [ + 122, + 94, + 849, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "evaluates the expectation of contrastive loss. It should be noting that in $C_{u,v}$ , $(f_u^i, f_v^j)$ is considered as positive pairs, while $\\{(f_u^i, f_{v'}^j) | v' \\neq v\\}$ and $\\{(f_{u'}, f_v^j) | u' \\neq u\\}$ are considered as negative pairs. Thus, $\\mathcal{L}_{scon}$ can encourage $S_0$ to be a permutation matrix, which indicates that there exists a one-to-one match between $X$ and $\\hat{Y}$ , and drive $S_1, S_2$ to be zero matrices.", + "bbox": [ + 75, + 450, + 472, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The problem defined by Eq. (3) is a typical balanced Optimal Transport (OT [35]) problem which transports $\\mathbf{1}_{m_{i,j}}$ to $\\mathbf{1}_{m_{i,j}}$ using $\\mathbf{1} - C$ as the cost matrix and $\\Omega$ is the transport matrix. Therefore, we consider $\\mathcal{L}_{scon}(i)$ as an OT loss, which can be solved with the Sinkhorn algorithm [38].", + "bbox": [ + 75, + 563, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constraint for $\\mathbf{S}_3$ . Obviously, the elements in $\\mathbf{S}_3$ measure the similarity of different individuals, so they should be as small as possible. However, directly constraining $\\mathbf{S}_3$ to zero matrix will make features from $\\hat{\\mathbf{X}}, \\hat{\\mathbf{Y}}$ easily collapse to zero-vector. To this end, the widely used Hinge L1 loss [18] is adopted to constrain $\\mathbf{S}_3$ :", + "bbox": [ + 75, + 642, + 470, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h i n g e} (i) = \\frac {\\sum R e l u (\\mathbf {S _ {3}} - \\theta)}{(n _ {i} - m _ {i , j}) (n _ {j} - m _ {i , j})}, \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 751, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is a threshold used to ignore small values in $\\mathbf{S}_3$ .", + "bbox": [ + 76, + 803, + 447, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the group-level matching loss of the training set $\\mathbf{V}$ is formulated as,", + "bbox": [ + 76, + 821, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G M L} (\\mathbf {V}) = \\sum_ {0 < i < T, i = k \\delta + 1} \\left(\\mathcal {L} _ {s c o n} (i) + \\mathcal {L} _ {h i n g e} (i)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Memory-based Count Predictor", + "text_level": 1, + "bbox": [ + 500, + 450, + 782, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle individuals' appearance variance and re-entering cases, inspired by [22, 33, 55], we propose a memory-based count predictor (MCP) to reason the final count for the video, which stores templates of recently appeared individuals in a flexible memory. Specifically, when processing frame $\\mathbf{V}^{i + \\delta}$ , the memory is denoted as $\\mathcal{G}^i = \\{\\mathcal{G}_1^i,\\mathcal{G}_2^i,\\dots ,\\mathcal{G}_{\\mathbf{N}^i}^i\\}$ , where $\\mathbf{N}^i$ is the memory size at time $i$ . $\\mathcal{G}_u^i = \\{g_{u,0}^i,\\dots,g_{u,k}^i\\}$ is a set of templates for $u$ -th pedestrian in the memory, and $k$ is the number of templates stored for $u$ -th pedestrian.", + "bbox": [ + 496, + 473, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first use the crowd locator to predict the coordinates of pedestrians in $\\mathbf{V}^{i + \\delta}$ , and then extract their corresponding features, denoted as $\\{f_1^{i + \\delta}, f_2^{i + \\delta}, \\dots, f_{n_{i + \\delta}}^{i + \\delta}\\}$ . To associate the individuals in $\\mathbf{V}^{i + \\delta}$ with templates in memory, we first define the cost $\\hat{C}_{u,u'}$ of matching the $u$ -th pedestrian with the $u'$ -th template in $\\mathcal{G}^i$ as", + "bbox": [ + 496, + 625, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} _ {u, u ^ {\\prime}} = \\max _ {v} \\left(1 - \\left\\langle f _ {u} ^ {i + \\delta}, g _ {u ^ {\\prime}, v} ^ {i} \\right\\rangle\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 723, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we generate an optimal one-to-one match $\\pi \\in \\{0,1\\}^{n_{i + \\delta}\\times \\mathbf{N}^i}$ (assuming $\\mathbf{N}^i\\geq n_i$ ) by solving the following problem with Hungarian Algorithm [21]:", + "bbox": [ + 498, + 753, + 893, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi = \\arg \\min _ {\\pi} \\sum_ {u, u ^ {\\prime}} \\pi_ {u, u ^ {\\prime}} \\hat {C} _ {u, u ^ {\\prime}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 806, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad \\forall u ^ {\\prime}, \\sum_ {u} \\pi_ {u, u ^ {\\prime}} = 1, \\quad \\forall u, \\sum_ {u} \\pi_ {u, u ^ {\\prime}} \\leq 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 843, + 849, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If the matching cost for $u$ meets $\\sum_{u' = 1}^{\\mathbf{N}^i}\\pi_{u,u'}\\hat{C}_{u,u'} > \\zeta$", + "bbox": [ + 500, + 882, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19231", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$u$ will be judged as an inflow pedestrian. Otherwise, if $\\pi_{u,u'} = 1$ , $u$ is associated with $g_{u'}^i$ .", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Template update. Similar to [22], $\\mathcal{G}_u^i$ has a time-to-live factor (ttl). $ttl_u$ will decrease one if $\\mathcal{G}_u^i$ has no mapping to any pedestrian feature in $\\mathbf{F}^{i + \\delta}$ . If $ttl_u$ decreases to zero, the template $\\mathcal{G}_u^i$ will be dropped. Meanwhile, whenever $\\mathcal{G}_u^i$ is associated with a pedestrian in the current frame, $ttl_u$ will be reset to $ttlmax$ . This time-to-live threshold allows us to save some recently appeared targets while discarding those that have not appeared for a long time. If pedestrian $u$ is associated with $u'$ -th template, we add $f_{u}^{i + \\delta}$ to $\\mathcal{G}_{u'}^i$ to update the template", + "bbox": [ + 75, + 121, + 468, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {u ^ {\\prime}} ^ {i + \\delta} \\leftarrow \\mathcal {G} _ {u ^ {\\prime}} ^ {i} \\cup \\left\\{f _ {u} ^ {i + \\delta} \\mid \\pi_ {u, u ^ {\\prime}} = 1, \\sum_ {u ^ {\\prime} = 1} ^ {\\mathbf {N} ^ {i}} \\pi_ {u, u ^ {\\prime}} \\hat {C} _ {u, u ^ {\\prime}} \\leq \\zeta \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 284, + 468, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If pedestrian $u$ is an inflow pedestrian, $\\{f_u^{i + \\delta}\\}$ will be added to the memory $\\mathcal{G}^{i + \\delta}$ as a new template.", + "bbox": [ + 76, + 313, + 468, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output number of the total individuals in video $I$ is then calculated via", + "bbox": [ + 76, + 345, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {N} = n _ {1} + \\sum_ {\\delta + 1 \\leq i \\leq T, i = k \\delta + 1} \\hat {n} _ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 383, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{n}_i$ is the number of inflow pedestrians at time $i$ .", + "bbox": [ + 76, + 428, + 437, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 457, + 209, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 482, + 173, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We test our CGNet on three datasets: CroHD [42], SenseCrowd [24], and UAVVIC. CroHD has four training videos and five testing videos. SenseCrowd contains 634 videos, and we split the train, validation, and test dataset following DRNet [14]. UAVVIC is our proposed dataset collected by a moving UAV camera in various scenes, including campus, beach, car park, highway, city road, and square. UAVVIC consists of 221 videos (100 for training, 100 for testing, and 21 for validation), and 5,396 frames are sampled with 3s as the interval. Annotations consist of 398,158 bounding boxes, and group-level matching labels in neighbor frames are provided. The resolutions of UAVVIC are 4K and 1080P for better capture of tiny pedestrians from drone view. A detailed comparison of UAVVIC with existing video crowd counting datasets [8, 9, 13, 24, 42] is shown in Tab. 1. UAVVIC provided a moving shot with a larger range of pedestrians. More details about UAVVIC are provided in the supplementary materials.", + "bbox": [ + 75, + 505, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 76, + 786, + 173, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mean Absolute Error (MAE), Mean Square Error (MSE), and Weighted Relative Absolute Errors (WRAE) are used for evaluation. The first two metrics are common metrics applied in VCC. However, unlike VCC schemes, we only count the same person once in the video. WRAE [14] is proposed to balance the performance on videos with", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e48405731f81ba1343facac742cd503dc9f60af03ac01bd197a82319db80479d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NameResolutionRangeMoving shotPointBoxTrajectoryIn-Out
Mall [9]640*48013-53X✓XXX
UCSD [8]238*15811-46X✓XXX
FDST [13]1920*10809-57X✓XXX
1280*720
SenseCrowd [24]-1-296X✓✓✓X
CroHD [42]-25-346X✓✓✓X
UAVVIC3840*21600-735✓✓✓X✓
1920*1080
", + "bbox": [ + 504, + 88, + 888, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of different video crowd counting datasets.", + "bbox": [ + 501, + 186, + 885, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different lengths and pedestrian numbers: $WRAE = \\sum_{i=1}^{K} \\frac{T_i}{\\sum_{j=1}^{K} T_j} \\frac{|N_i - \\hat{N}_i|}{N_i}$ , where $K$ is the total number of videos, and $T_i$ is the length of video $i$ , $N_i / \\hat{N}_i$ denotes the number of ground truth/predicted counts in $i$ -th video.", + "bbox": [ + 498, + 210, + 893, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 295, + 718, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. If there are no special instructions, the loca-tor we applied is a well-known crowd localization model, FIDT [25]. All parameters are official from FIDT, except that we replaced the backbone from HRNet-W48 [47] to HRNet-W18 [47] for efficient inference. The learning rate is set as $1e^{-4}$ along with AdamW [28] as the optimizer and applying the pre-train weight from Timm [49]. All related models can be trained on one RTX3090 (24G memory).", + "bbox": [ + 496, + 318, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing. In the testing phase, the time interval $\\delta$ is set as $3s$ following the setting in [14], the threshold $\\zeta$ is set as $0.7$ , $\\gamma$ in Eq. (4) is set as 10, and the max time-to-live (ttlmax) is set as 3, the max size of memory (memmax) is set as 5.", + "bbox": [ + 496, + 441, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Overall Performance", + "text_level": 1, + "bbox": [ + 498, + 513, + 697, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. To the best of our knowledge, only DRNet[14] has been specifically designed for the VIC task. To evaluate the effectiveness of our proposed CGNet, we also tested two categories of approaches that could be applied to VIC: 1) Multi-object tracking (MOT) methods, including HeadHunter-T [43], FairMOT [56], PHDTT [44], SMILE [48], SparseTrack [27], Deep-OC-SORT [31], and BoT-SORT [1], where we consider the total number of indices in the entire video as individual count. We set the frame rate to 1 FPS for HeadHunter-T and 0.33 FPS for the other MOT methods to obtain better VIC performance. 2) A recent cross-line video crowd-counting, LOI [57], which also requires trajectory labels in training.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on SenseCrowd: As shown in Tab. 2, it is evident that our CGNet exhibits superior performance on this larger dataset, SenseCrowd. Particularly, CGNet improved the overall MAE by about $28\\%$ compared to DRNet. Furthermore, we achieve the lowest MAE across all density levels except for D0. Compared to the existing lowest MAE on D1, D2, D3, and D4, the performance improvements of our CGNet are about $27\\%$ , $63\\%$ , $37\\%$ , and $17\\%$ , respectively. In Fig. 3, we show qualitative results on three representative scenes,i.e., mall, station, and outdoor plaza. Although the inflows in the current flow are not concentrated, CGNet", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19232", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e5b7e7e501e5f533ddca8baee52582ff7025253a685864959c262df4ba75b8ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 87, + 372, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c99610530f15a593083609005a31cb0ea38e2016e75a6676865517b0b668667.jpg", + "image_caption": [ + "(a) Previous frame (0s)" + ], + "image_footnote": [], + "bbox": [ + 377, + 88, + 589, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fbb32e315b8ecd072b065c520434f9980f6ac9f1b666bd2a7058b683018c8010.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 88, + 810, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7a94857b819b4959248b6767268bdd292414e3fac4548a45ebc63fc1f3324fdb.jpg", + "image_caption": [ + "GT inflow: 4", + "Predict inflow: 4" + ], + "image_footnote": [], + "bbox": [ + 161, + 196, + 372, + 290 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9d22f4482517acb43e4a085fe28e1f5180b6371669a9502d1f61c71cb616879.jpg", + "image_caption": [ + "GT inflow: 11", + "(b) Current frame (3s)", + "Figure 3. Results on the SenseCrowd dataset. Red dots are predictions of the locator. Blue/green/yellow circles are correct predicted inflow, error predicted inflow, and missed inflow, respectively." + ], + "image_footnote": [], + "bbox": [ + 377, + 196, + 589, + 290 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3e6f205947c67ee45c2337fe81a5d29cca46f4b22e4a61144c4f53badd9593a.jpg", + "image_caption": [ + "GT inflow: 15", + "Predict inflow: 14" + ], + "image_footnote": [], + "bbox": [ + 591, + 196, + 805, + 290 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c0d38cb62a97a0ac175e5ed75f8d35d4eae4e3092decf53a5cb3391c6a51f1ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
FairMOT [56]35.462.348.913.522.467.984.4145.8
HeadHunter-T [43]30.050.638.611.825.756.092.6131.4
SMILE [48]27.2236.732.59.221.033.876.5203.5
SparseTrack [27]28.3643.830.89.921.542.189.2189.5
Deep-OC-SORT [31]26.0443.629.59.525.129.160.3156.5
BoT-SORT [1]24.6732.428.38.720.531.170.2165.1
LOI [57]24.733.137.412.525.439.339.686.7
DRNet [14]12.324.712.74.18.023.350.077.0
CGNet (Ours)8.8617.6912.65.05.88.525.063.4
", + "bbox": [ + 114, + 359, + 852, + 523 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9012d7641343fde91633a4c50d17fdf921d3a3e19cf0f34ebd9638a1715f0310.jpg", + "table_caption": [ + "Table 2. Performance comparison on SenseCrowd. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and $[200, +\\infty)$ , respectively. The best values are highlighted in blue font." + ], + "table_footnote": [], + "table_body": "
MethodMAEMSEWRAE(%)MAE on five testing scenes
CroHD11CroHD12CroHD13CroHD14CroHD15
PHDTT [44]2130.42808.3401.6247379347944911327
FairMOT [56]256.2300.844.1114272844081000
HeadHunter-T [43]253.2351.732.7651015155821003
SMILE [48]257.6334.540.516231156649246
SparseTrack [27]176.6208.627.62917225833688
Deep-OC-SORT [31]165.2195.933.16835118616160
BoT-SORT [1]154.8176.4227.44921013828691
LOI [57]305.0371.146.060243458630131
DRNet [14]141.1192.327.4313381825561
CGNet (Ours)75.095.114.577214144138
", + "bbox": [ + 112, + 569, + 854, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Performance comparison on CroHD dataset. CroHD11-CroHD15 are five test scenes labeled in [14]. The best values are highlighted in blue font.", + "bbox": [ + 75, + 758, + 892, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "can accurately predict them. More visualization results can be found in supplementary material.", + "bbox": [ + 75, + 801, + 468, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on CroHD. Comparison with existing methods on CroHD is shown in Tab. 3. Our CGNet achieved the lowest MAE, MSE, and WRAE. Especially, except for CroHD15, our method obtained the lowest MAE across the other four", + "bbox": [ + 75, + 839, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "testing scenes. As CGNet is trained with only point annotations, it cannot crop the patch of optimal size to extract features on CroHD15. It should be noted that on CroHD13 and CroHD14, which have the highest average density (245.9 and 259.6 person/frame), the performances of all the compared methods except for DRNet degrade sig-", + "bbox": [ + 496, + 801, + 892, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a26c5f9df63f7b9f86dd10e61a89a9220be79118c97f56cd1b9240505d72c015.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMAE/MSE/WRAEMAE/MSE/WRAE on four different categories
PedestrianCarBusVan
BoT-SORT [1]49.9/228.2/29.992.4/398.2/49.059.1/259.1/39.223.0/109.6/57.125.1/146.2/77.3
Deep-OC-SORT [31]37.1/115.4/35.285.2/213.5/22.133.4/113.1/21.019.1/ 78.2/26.210.7/56.6/32.8
DRNet [14]18.4/ 41.0/15.234.5/ 92.1/19.119.0/ 32.7/17.810.4/ 22.1/32.29.9/17.2/29.4
CGNet (Ours)12.9/ 37.4/ 12.031.2/ 98.4/14.416.9/ 45.2/ 7.32.1/ 4.1/19.21.3/ 2.0/25.1
", + "bbox": [ + 119, + 88, + 851, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a6bbde685b6fa7962f4477b10000351331ffa4855987c8295d63e69fd98c01f4.jpg", + "table_caption": [ + "Table 4. Performance comparison on UAVVIC. All the methods are trained independently with four classes. The values of each entry are MAE/MSE/WRAE. The best values are highlighted in blue font." + ], + "table_footnote": [ + "Table 5. Ablation study on the main components. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and $[200, +\\infty)$ , respectively. $①$ , No constrain on $\\mathbf{S}_1$ , $\\mathbf{S}_2$ ; $②$ , w/o Eq. (5); $③$ , w/o MCP; $④$ , full CGNet. The best values are highlighted in blue font." + ], + "table_body": "
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
①13.3532.322.89.712.315.326.543.4
②9.8420.013.85.36.312.028.867.7
③10.7117.4016.96.98.911.724.350.6
④8.8617.6912.65.05.88.525.063.4
", + "bbox": [ + 81, + 244, + 467, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nificantly. In contrast, the performance of our CGNet is relatively stable in these two scenes. Furthermore, even without individual-level matching labels, our CGNet performs favorably against the state-of-the-art approach, DRNet, by a large margin (about $46\\%$ on the MAE), demonstrating the effectiveness of our method.", + "bbox": [ + 75, + 421, + 468, + 512 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eba599d1e565fbc096ce4f4681298ae1b78e4dbba74f2bdcdc85ceba407b63cc.jpg", + "image_caption": [ + "(a) Previous frame (0s)", + "Figure 4. Qualitative visualization of our CGNet on UAVVIC." + ], + "image_footnote": [], + "bbox": [ + 89, + 521, + 271, + 612 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6706da357701209329e8b5b7ad86308250a97e53c3680723cfa6df129ee09c33.jpg", + "image_caption": [ + "(b) Current frame (3s)" + ], + "image_footnote": [], + "bbox": [ + 272, + 522, + 452, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on UAVVIC. The results on our proposed UAVVIC dataset are shown in Tab. 4. All methods are trained and tested on each category, respectively. Since there is no trajectory annotation on UAVVIC, we train the compared methods [1, 14, 31] on VisDrone [59], a UAV dataset for MOT. Our CGNet shows favorable performance in all categories. The average MAE is improved by about $30\\%$ compared to DRNet. In Fig. 4, we show the quality result of our CGNet on the UAVVIC dataset. Although the UAV moves, CGNet can still detect inflow objects accurately.", + "bbox": [ + 75, + 641, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 801, + 240, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted several ablation studies on the SenseCrowd. 1). We verify the effectiveness of the main components of our CGNet. 2). We analyze the effect of two main parameters. 3). We report the performance of our CGNet with different locators. 4). We use the trajectories generated by", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23d33fb0d1353c0e5af4cb555f60bb704ace775dad287b5aeef8cc6cdbd127d9.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 514, + 247, + 694, + 410 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bd1d33c6e964bea8c7791d00f4c4f20b8120e9b0dfc146bc0b9d2cac17ad65e.jpg", + "image_caption": [ + "(b)", + "Figure 5. Effectiveness of MCP on SenseCrowd. (a) The average MAE with different lengths of videos with and w/o MCP. (b) Performance with different max time-to-live factor tllmax of MCP." + ], + "image_footnote": [], + "bbox": [ + 699, + 243, + 875, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our CGNet as pseudo-labels for existing VIC methods. 5). We compare the time cost of the WVIC and VIC labels.", + "bbox": [ + 498, + 500, + 888, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Constrain for $\\mathbf{S}_1, \\mathbf{S}_2$ . Eq. (4) constrains $\\mathbf{S}_0, \\mathbf{S}_1, \\mathbf{S}_2$ together. Without considering un-matched pairs in $\\mathbf{X} \\times \\hat{\\mathbf{Y}}$ and $\\hat{\\mathbf{X}} \\times \\mathbf{Y}$ , it can be changed to constrain only $\\mathbf{S}_0$ by remove elements in $\\mathbf{S}_1, \\mathbf{S}_2$ . As shown in Tab. 5, MAE dropped by about 4.51 without constraints on $\\mathbf{S}_1$ and $\\mathbf{S}_2$ (① VS ④). This is because elements in $\\mathbf{S}_1$ and $\\mathbf{S}_2$ are related to pedestrians that exist only in one frame. Considering $\\mathbf{X} \\times \\hat{\\mathbf{Y}}$ and $\\hat{\\mathbf{X}} \\times \\mathbf{Y}$ as negative pairs, our soft contrastive loss can push apart individuals in $\\hat{\\mathbf{X}} (\\hat{\\mathbf{Y}})$ with that in $\\mathbf{Y} (\\mathbf{X})$ , thus enhancing the distinguishing ability of the representations.", + "bbox": [ + 498, + 531, + 892, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Eq. (5). We replace the hinge MAE loss in Eq. (5) to normal L1 loss and derive the method ② in Tab. 5. Compared to L1 loss, our $\\mathcal{L}_{hinge}(\\cdot)$ gives a soft margin to the inner product of two features other than encouraging them to be 0. Optimizing the inner product of all features toward 0 makes it easier to cause either side to collapse a 0 vector, which is not conducive to learning different representations. Compared the performance of ② with that of ④ in Tab. 5, $\\mathcal{L}_{hinge}(\\cdot)$ improves MAE in all density levels.", + "bbox": [ + 498, + 686, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of MCP. As shown in Tab. 5, compared $③$ with $④$ , MCP brings about 4.1 improvements on WRAE. Specifically, we evaluate the average MAE of our CGNet on videos with different lengths. As shown in Fig. 5(a), the performance gap between $③$ and $④$ becomes larger as", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19234", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae7cfd1c73575eed1861557f6a1de15577f42c0f437bb25e5da38115d222f75e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
δMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
510.9620.2619.56.56.910.429.279.9
48.7114.3312.94.24.910.227.870.1
38.8617.6912.65.05.88.525.063.4
29.7720.0117.37.18.812.417.452.6
19.5615.3416.66.87.410.523.145.5
", + "bbox": [ + 81, + 88, + 464, + 179 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8eb5fc7fbef258fba022843e9036fd6f2b239ec782d1165e5aa52ce5b7890316.jpg", + "table_caption": [ + "Table 6. Ablation study on interval $\\delta$ . The best values are highlighted in blue font." + ], + "table_footnote": [], + "table_body": "
MethodLocatorEncoderMAEMSEWRAE
①FIDTConvNext-S8.8617.699.27
②Yolov8 nanoConvNext-S8.0116.428.85
③Yolov8 smallConvNext-S7.5615.198.05
④VGG16+FPNConvNext-S9.2017.459.55
⑤VGG16+FPNPrRoiIPooling10.0118.229.70
⑥VGG16+FPNPrRoiIPooling12.5923.3212.58
", + "bbox": [ + 81, + 220, + 464, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the length of the video increases. This is because, with the stored templates of individuals, MCP can handle appearance variations and individuals' re-entry, alleviating error accumulation during inflow reasoning.", + "bbox": [ + 75, + 386, + 467, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of ttlmax. The performance with different max long-time-live factor ttlmax is shown in Fig. 5(b). Compared with no memory (i.e., ttlmax = 0), adding memory is helpful for all density levels except for extremely high-density levels (D4). It may be because high-density crowds result in too many candidate individual-level matches, making it difficult for our CGNet to choose the right one. Meanwhile, the performance degrades when ttlmax increases from 3 to 6 for that larger ttlmax = 6 leads to the templates of individuals that appeared a long time ago being stored in the memory. As the probability of these people re-entering the field is low, the recorded templates become distractions.", + "bbox": [ + 75, + 446, + 467, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Interval $\\delta$ . The performance with different intervals $\\delta$ is shown in Tab. 6. The performance on higher-density video is better when $\\delta$ is smaller. This is because crowds flow faster in those videos, and long intervals may lead to missing pedestrians in sampled frames. The overall best performance is achieved when $\\delta$ is set as $4s$ .", + "bbox": [ + 75, + 628, + 467, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Locator. We replaced FIDT in our CGNet with Yolo V8 [19] nano and small and the localization branch (VGG16+FPN) used in DRNet, respectively. Different from FIDT, these three locators are trained with box annotations. For Yolo V8, the predicted bounding boxes are fed to our feature extractor to generate representations of individuals. For a fair comparison with DRNet, we also replace our feature extractor ConvNext-S with PrRoIPooling [17] used in DRNet [14] and remove MCP. As shown in Tab. 7, changing locators can further boost the performance of our method. With the same locator and feature extractor as DRNet (⑤), the MAE of our method is 10.01, 2.99 lower", + "bbox": [ + 75, + 719, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a599c73c507469a23de8512fb30407c6e3bd3b7f2d34ca6de6290a45aa7ce85.jpg", + "table_caption": [ + "Table 7. Performance on SenseCrowd of our scheme with different locators and feature extractors. In $⑥$ , MCP is replaced with the inflow reasoning method in DRNet. The best values are highlighted in blue font." + ], + "table_footnote": [], + "table_body": "
LocatorGround Truth LabelsPseudo Labels
MAEMSEWRAEMAEMSEWRAE
FairMOT35.462.348.937.264.849.8
HeadHunter-T30.050.638.632.554.239.9
SMILE27.236.732.527.537.035.2
SparseTrack28.443.830.830.545.031.5
Deep-OC-SORT26.043.629.529.545.331.7
BoT-SORT24.732.428.325.035.933.8
LOI24.733.137.426.935.239.9
DRNet12.324.712.714.525.513.3
", + "bbox": [ + 504, + 88, + 887, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61a49e0cdf5c19ecb8e0c8e36cfd5c132271ab2594c38a5853757de87e00eb6e.jpg", + "table_caption": [ + "Table 8. Performance of existing VIC methods trained with different trajectory labels." + ], + "table_footnote": [], + "table_body": "
(ni, ni+δ)(9,8)(60,53)(116,115)(289,292)(645,724)
Association [14]23.8 s104.5 s174.4 s586.9 s1663.1 s
In-Out4.4 s42.6 s53.7 s95.0 s571.9 s
", + "bbox": [ + 501, + 263, + 895, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Time cost for a pair of frames with two types of labeling methods. $n_i$ denotes the number of pedestrians in frame $i$ .", + "bbox": [ + 498, + 321, + 890, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "than that of DRNet. Even without MCP (⑥), the MAE of our method is only slightly larger than that of DRNet (12.59 v.s. 12.3), demonstrating that our method can learn effective representations without trajectory labels.", + "bbox": [ + 496, + 354, + 890, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performs as a Pseudo Trajectory Generator. The latent variable $\\Omega$ is obtained by solving a problem defined by Eq. (3). The Hungarian Algorithm generates a one-to-one match using $\\Omega$ as the cost matrix. Based on the matching, we generate pseudo trajectory labels, which are used to train existing VIC methods. We show the results on SenseCrowd in Tab. 8. With the generated pseudo label, the performance of all the methods degrades slightly than using the ground truth trajectory labels, demonstrating that exact individual-level matching annotations may not be necessary for VIC.", + "bbox": [ + 496, + 416, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Label time cost. We compare the time cost of annotating our weakly supervised label and association label [14] on five frame pairs with various densities in Tab. 9. The annotation time for our weakly supervised label is only about $30\\%$ of that of the association label, making it easier to build a larger dataset for this VIC.", + "bbox": [ + 496, + 569, + 890, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 679, + 625, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a new task, Weakly supervised Video Individual Counting (WVIC), in this paper. Unlike conventional video individual counting tasks, which require expensive trajectory labels, our WVIC task just requires two types of easily available identity-agnostic annotations. This largely reduces the annotation cost and facilitates future works to upscale data size. Furthermore, we propose a strong baseline for the WVIC task. The key idea is a Group-level Matching Loss (GML), which makes it easier to distinguish inflow individuals from all objects. Last, we have extended two existing datasets and built a totally new dataset for the WVIC task. Our method achieves favorable performance on all three datasets, even compared to supervised methods.", + "bbox": [ + 496, + 705, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19235", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 5, 6, 7", + "[2] Antonio Albiol, Inmaculada Mora, and Valery Naranjo. Real-time high density people counter using morphological tools. IEEE TITS, 2(4):204-218, 2001. 2", + "[3] Marco Allodi, Alberto Broggi, Domenico Giaquinto, Marco Patander, and Antonio Prioletti. Machine learning in tracking associations with stereo vision and lidar observations for an autonomous vehicle. In IVS, pages 648-653, 2016. 2", + "[4] Seung Hwan Bae and Kuk-Jin Yoon. Confidence-based data association and discriminative deep appearance learning for robust online multi-object tracking. IEEE TPAMI, 40(3):595-610, 2018. 2", + "[5] Javier Barandiarán, Berta Murguia, and Fernando Boto. Real-time people counting using multiple lines. In WIAMIS, pages 159-162, 2008. 2", + "[6] Jesús Bescós, José M. Menéndez, and Narciso García. DCT based segmentation applied to a scalable zenithal people counter. In ICIP, pages 1005-1008, 2003. 2", + "[7] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, pages 1-6, 2017. 2", + "[8] Antoni B. Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE TPAMI, 30(5):909-926, 2008. 5", + "[9] Ke Chen, Chen Change Loy, Shaogang Gong, and Tony Xiang. Feature mining for localised crowd counting. In BMVC, pages 3-14, 2012. 5", + "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3", + "[11] Li Dong, Haijun Zhang, Jianghong Ma, Xiaofei Xu, Yimin Yang, and Q. M. Jonathan Wu. Clrnet: A cross locality relation network for crowd counting in videos. IEEE TNNLS, pages 1-15, 2022. 2", + "[12] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. In ICME, pages 814-819, 2019. 1, 2", + "[13] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. arXiv preprint arXiv:1907.07911, 2019.5", + "[14] Tao Han, Lei Bai, Junyu Gao, Qi Wang, and Wanli Ouyang. DR.VIC: decomposition and reasoning for video individual counting. In CVPR, pages 3073-3082, 2022. 1, 2, 5, 6, 7, 8", + "[15] Yuhang He, Jie Han, Wentao Yu, Xiaopeng Hong, Xing Wei, and Yihong Gong. City-scale multi-camera vehicle tracking by semantic attribute parsing and cross-camera tracklet matching. In CVPRW, pages 2456-2465, 2020. 2", + "[16] Andinet Hunde and Beshah Ayalew. Automated multi-target tracking in public traffic in the presence of data association uncertainty. In ACC, pages 300-306, 2018. 2", + "[17] Borui Jiang, Ruixuan Luo, Jiayuan Mao, Tiantian Xiao, and Yuning Jiang. Acquisition of localization confidence for accurate object detection. In ECCV, pages 784-799, 2018. 8", + "[18] Thorsten Joachims. Support vector method for multivariate performance measures. In ICML, pages 377-384, 2005. 4" + ], + "bbox": [ + 78, + 114, + 470, + 896 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.8", + "[20] Chanho Kim, Fuxin Li, Mazen Alotaibi, and James M. Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, pages 9553-9562, 2021. 2", + "[21] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 4", + "[22] L.J. Latecki and R. Miezianko. Object tracking with dynamic template update and occlusion detec. In ICPR, volume 1, pages 556-560, 2006. 4, 5", + "[23] Eui-Hyuk Lee, Qian Zhang, and Taek Lyul Song. Markov chain realization of joint integrated probabilistic data association. Sensors, 17(12):2865, 2017. 2", + "[24] Haopeng Li, Lingbo Liu, Kunlin Yang, Shinan Liu, Junyu Gao, Bin Zhao, Rui Zhang, and Jun Hou. Video crowd localization with multifocus gaussian neighborhood attention and a large-scale benchmark. IEEE TIP, 31:6032-6047, 2022. 5", + "[25] Dingkang Liang, Wei Xu, Yingying Zhu, and Yu Zhou. Focal inverse distance transform maps for crowd localization. IEEE TMM, 25:6040-6052, 2023. 3, 5", + "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 3", + "[27] Zelin Liu, Xinggang Wang, Cheng Wang, Wenyu Liu, and Xiang Bai. Sparsetrack: Multi-object tracking by performing scene decomposition based on pseudo-depth. arXiv preprint arXiv:2306.05238, 2023. 5, 6", + "[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In *ICLR*, 2018. 5", + "[29] Zhongzhen Luo, Mina Attari, Saeid R. Habibi, and Martin von Mohrenschild. Online multiple maneuvering vehicle tracking system based on multi-model smooth variable structure filter. IEEE TITS, 21(2):603-616, 2020. 2", + "[30] Zheng Ma and Antoni B. Chan. Counting people crossing a line using integer programming and local features. IEEE TCSVT, 26(10):1955-1969, 2016. 2", + "[31] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 5, 6, 7", + "[32] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3", + "[33] Zhuojin Pan and Xiujuan Wang. Correlation tracking algorithm based on adaptive template update. In ICISP, volume 1, pages 98-101, 2010. 4", + "[34] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, pages 164-173, 2021. 2", + "[35] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning, 11(5-6):355-607, 2019. 4", + "[36] Christopher Rasmussen and Gregory D. Hager. Probabilistic data association methods for tracking complex visual objects. IEEE TPAMI, 23(6):560-576, 2001. 2", + "[37] Christian Ritter, Andrea Imle, Ji Young Lee, Barbara Muller," + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19236", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Oliver T. Fackler, Ralf Bartenschlager, and Karl Rohr. Two-filter probabilistic data association for tracking of virus particles in fluorescence microscopy images. In *ISBI*, pages 957–960, 2018. 2", + "[38] Richard Sinkhorn. A Relationship Between Arbitrary Positive Matrices and Doubly Stochastic Matrices. The Annals of Mathematical Statistics, 35(2):876 - 879, 1964. 4", + "[39] Rakshith Sharma Srinivasa, Jaejin Cho, Chouchang Yang, Yashas Malur Saidutta, Ching-Hua Lee, Yilin Shen, and Hongxia Jin. CWCL: Cross-modal transfer with continuously weighted contrastive loss. In NeurIPS, 2023. 3", + "[40] Jason Stauch, Travis Bessell, Mark Rutten, Jason Baldwin, Moriba Jah, and Keric Hill. Joint probabilistic data association and smoothing applied to multiple space object tracking. JGCD, 41(1):19-33, 2018. 2", + "[41] Ramana Sundararaman, Cedric De Almeida Braga, Éric Marchand, and Julien Pettré. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 2", + "[42] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 5", + "[43] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, June 2021. 5, 6", + "[44] Xuan-Thuy Vo, Van-Dung Hoang, Duy-Linh Nguyen, and Kang-Hyun Jo. Pedestrian head detection and tracking via global vision transformer. In IFCVW, pages 155-167, 2022. 2, 5, 6", + "[45] Xingyu Wan, Jinjun Wang, Zhifeng Kong, Qing Zhao, and Shunming Deng. Multi-object tracking using online metric learning with long short-term memory. In ICIP, pages 788-792, 2018. 2", + "[46] Bing Wang, Gang Wang, Kap Luk Chan, and Li Wang. Tracklet association by online target-specific metric learning and coherent dynamics estimation. IEEE TPAMI, 39(3):589-602, 2017. 2", + "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2020. 5", + "[48] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking. arXiv preprint arXiv:2211.08824, 2023. 5, 6", + "[49] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5", + "[50] Hefeng Wu, Yafei Hu, Keze Wang, Hanhui Li, Lin Nie, and Hui Cheng. Instance-aware representation learning and association for online multi-person tracking. PR, 94:25-34, 2019. 2", + "[51] Xingjiao Wu, Baohan Xu, Yingbin Zheng, Hao Ye, Jing Yang, and Liang He. Fast video crowd counting with a temporal aware network. Neurocomputing, 403:13-20, 2020. 1, 2", + "[52] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 3" + ], + "bbox": [ + 78, + 90, + 468, + 890 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Feng Xiong, Xingjian Shi, and Dit-Yan Yeung. Spatiotemporal modeling for crowd counting in videos. In ICCV, pages 5161-5169, 2017. 1, 2", + "[54] Bo Yang, Chang Huang, and Ram Nevatia. Learning affinities and dependencies for multi-target tracking using a CRF model. In CVPR, pages 1233-1240, 2011. 2", + "[55] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE TPAMI, 43(1):360-374, 2021. 4", + "[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129:3069–3087, 2021. 2, 5, 6", + "[57] Zhuoyi Zhao, Hongsheng Li, Rui Zhao, and Xiaogang Wang. Crossing-line crowd counting with two-phase deep neural networks. In ECCV, pages 712-726, 2016. 2, 5, 6", + "[58] Huicheng Zheng, Zijian Lin, Jiepeng Cen, Zeyu Wu, and Yadan Zhao. Cross-line pedestrian counting based on spatially-consistent two-stage local crowd density estimation and accumulation. IEEE TCSVT, 29(3):787-799, 2018. 2", + "[59] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE TPAMI, 44(11):7380-7399, 2021. 7" + ], + "bbox": [ + 501, + 92, + 890, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19237", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_model.json b/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4c758d1959ea819eb4481988ec5cef732b4b484a --- /dev/null +++ b/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_model.json @@ -0,0 +1,2497 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.131, + 0.722, + 0.154 + ], + "angle": 0, + "content": "Weakly Supervised Video Individual Counting" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.18, + 0.841, + 0.229 + ], + "angle": 0, + "content": "Xinyan Liu1 Guorong Li1,3* Yuankai Qi2 Ziheng Yan1 \nZhenjun Han1 Anton van den Hengel5 Ming-Hsuan Yang6 Qingming Huang1,3,4 \n1 University of Chinese Academy of Science; 2 Macquarie University;" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.234, + 0.785, + 0.269 + ], + "angle": 0, + "content": "3 Key Lab of Big Data Mining and Knowledge Management, UCAS, China; 4 Key Lab of Intell. Info. Process., Inst. of Comput. Tech., CAS;" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.27, + 0.771, + 0.305 + ], + "angle": 0, + "content": "\\(^{5}\\) Australian Institute for Machine Learning, The University of Adelaide; \\(^{6}\\) University of California at Merced." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Video Individual Counting (VIC) aims to predict the number of unique individuals in a single video. Existing methods learn representations based on trajectory labels for individuals, which are annotation-expensive. To provide a more realistic reflection of the underlying practical challenge, we introduce a weakly supervised VIC task, wherein trajectory labels are not provided. Instead, two types of labels are provided to indicate traffic entering the field of view (inflow) and leaving the field view (outflow). We also propose the first solution as a baseline that formulates the task as a weakly supervised contrastive learning problem under group-level matching. In doing so, we devise an end-to-end trainable soft contrastive loss to drive the network to distinguish inflow, outflow, and the remaining. To facilitate future study in this direction, we generate annotations from the existing VIC datasets SenseCrowd and CroHD and also build a new dataset, UAVVIC. Extensive results show that our baseline weakly supervised method outperforms supervised methods, and thus, little information is lost in the transition to the more practically relevant weakly supervised task. The code and trained model can be found at CGNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.72, + 0.21, + 0.736 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Video Crowd Counting (VCC) has garnered much interest due to its broad range of practical applications, particularly in crowd safety management. This task requires a model" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.342, + 0.688, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.457, + 0.674, + 0.469 + ], + "angle": 0, + "content": "(a) Annotation of VIC method" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.342, + 0.895, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.457, + 0.88, + 0.469 + ], + "angle": 0, + "content": "(b) Annotation of our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.481, + 0.895, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Existing VIC method requires a unique label indicating the position of each human in each frame, and these labels are consistent across frames. Weakly supervised VIC only requires labels indicating each human position and whether they are an inflow/outflow pedestrian (purple/yellow in (b)) in the current frame. The transition from individual-level labels to identity-agnostic group-level labels represents a significant reduction in the labeling effort." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.732 + ], + "angle": 0, + "content": "to count the number of people in each frame of a video. A limitation of VCC [12, 51, 53] is that it gives an imprecise estimate of the number of unique individuals appearing in a video sequence, as people are counted multiple times if they appear in several frames. To overcome this drawback, Video Individual Counting (VIC) was introduced, wherein a method must count the total number of people with unique identities appearing in a video sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "The apparent approach to crowd counting is to count the people who appear in the first frame and add the number of people who come into the camera's field of view in later frames (the inflow). Following this principle, Han et al. [14] devised DRNet, the only existing published method for VIC, identifying repeated observations of individuals across consecutive frames based on their appearance, simultaneously predicting inflow and outflow. Although the number of outflows does not contribute to the final count, DRNet finds it helps determine individual associations. Han et al. [14] also explored using conventional multi-object track" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.804, + 0.223, + 0.817 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.816, + 0.47, + 0.902 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62272438, 62236008, U21B2038, and 61931008, in part by the Key Deployment Program of the Chinese Academy of Sciences under Grant KGFZD145-23-18, and in part by the Fundamental Research Funds for Central Universities (E2ET1104). Yuankai Qi, Anton van den Hengel, and Ming-Hsuan Yang are not supported by the aforementioned fundings." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.804, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19228" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.182 + ], + "angle": 0, + "content": "ing (MOT) methods [7, 41, 44, 54, 56] to tackle the VIC problem. Unfortunately, it turns out MOT methods suffer from poor accuracy. More notably, DRNet and MOT-based methods require trajectory labels (or similar individual association labels) to supervise identity association, which is highly annotation-expensive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.192, + 0.471, + 0.433 + ], + "angle": 0, + "content": "Our key insight is that counting people in the previous and current frames does not require accurate identity associations for those appearing in both frames. For example, as long as we can predict individual \\(15\\sim 17\\) as entering in Fig. 1a, even if we arbitrarily associate individuals \\(0\\sim 14\\) in frame \\(i\\) to individuals \\(0\\sim 14\\) in frame \\(i\\), we can still correctly infer the crowd counts. We can still count crowds if we relax individual-level identity associations to group-level associations and do not require trajectory annotations. Thus, we just need to annotate inflow and outflow people (then we can derive the individual exists in two frames), which reduces annotation costs compared to creating individual pairwise target associations for each observed pedestrian between neighboring frames. We name crowd counting with such annotations as Weakly supervised Video Individual Counting (WVIC)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.699 + ], + "angle": 0, + "content": "To address the WVIC task, we propose a benchmark method based on Contrastive learning with Group-level matching, namely CGNet. We design a soft contrastive loss to drive the network to learn discriminative representations that can facilitate identifying the required group associations and, thus, the inflow. Moreover, to better represent each individual, we design a memory-based individual count predictor, where historical templates of individuals are stored and updated in memory to enhance the robustness of association during inference. Considering that the two existing datasets for VIC are all captured by static cameras, we collect a weakly supervised VIC dataset based on moving UAVs, named UAVVIC. This dataset provides about 400,000 inflow/outflow and bounding box labels for four categories: pedestrian, car, bus, and van. Although devised for crowd counting, our proposed baseline method also performs well in counting other objects." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.711, + 0.352, + 0.725 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.735, + 0.468, + 0.794 + ], + "angle": 0, + "content": "- We propose WVIC, a weakly supervised video individual counting task. This task does not require expensive per-target trajectory annotations and only requires two types of identity-agnostic annotations." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.796, + 0.468, + 0.84 + ], + "angle": 0, + "content": "- We automatically reannounce two existing datasets, CroHD and SenseCrowd, and collect a new dataset, UAVVIC, to pave the way for future studies." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.84, + 0.468, + 0.899 + ], + "angle": 0, + "content": "- We propose a strong baseline, CGNet, equipped with a newly designed group level matching soft contrastive loss, performing favorably against the supervised methods on the three datasets mentioned above." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.735, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.65, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Video Crowd Counting (VCC) estimates the number of people in each video frame. Most of the existing VCC methods [2, 5, 6, 11, 12, 30, 51, 53, 58] can be divided into two categories according to their solved problems: region of interest (ROI) and line of interest (LOI). ROI methods [11, 12, 51, 53] detect pedestrians within a specific region (or the whole image), and they focus on leveraging temporal context information to improve the prediction of the current frame. LSTN [12] models the group flow of crowds in local regions. Xiong et al. in [53] fuse history frame features using a ConvLSTM. TAN [51] explores context information from adjacent density maps. CLRNet proposes a local self-attention module to help the model focus on highly related regions in adjacent frames. ROI methods can not be used in the VIC task. LOI methods [2, 5, 6, 30, 57, 58] counts people passing through a specific line. Most of the existing LOI methods [2, 5, 6] apply blobs to crowds, counting objects when the whole blobs cross the line. Ma et al. in [30] sample fixed line width areas within temporal image slices and accumulated crowds in this area. Zheng et al. [58] instead sum the number of people within neighboring blocks near a line based on local velocity. Zhao et al. in [57] train a crowd velocity map predictor with the help of trajectory labels. In practice, LOI methods perform poorly on the VIC task as people can enter or leave the field of view in any direction, making it difficult to find a suitable virtual line." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Multiple Object Tracking (MOT) aims to predict the trajectories for multiple targets in a video. Object association is a main challenge in MOT, establishing correspondences between objects across frames to generate targets' trajectories. Existing object association methods in MOT include Probabilistic Data Association (PDA) [36, 37], Joint Probabilistic Data Association (JPDA) [16, 23, 29, 40], and graph matching methods like the Hungarian algorithm [3, 50]. Recent research has started to use more powerful feature representations to enhance the robustness of data association. Deep learning-based methods like LSTM can model temporal information to handle long-term occlusion [20, 34, 45]. Many works [4, 15, 46] also focus on the association of tracklets. Specifically, they associate multiple tracklets over longer periods using a graph model to combine the final trajectory, which can relieve the effects of lost tracked targets. These methods perform poorly on the VIC task, however, due to the inevitable ID switching problem [14]. Additionally, they require trajectory labels to train their models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video Individual Counting (VIC). To the best of our knowledge, DRNet [14] is the only method for the VIC task. It predicts the number of initial pedestrians and matches pedestrians in adjacent frames by adapting a differentiable optimal transport loss. Although the VIC task only requires the total count in the video, DRNet needs individual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "matching between frames. WVIC avoids the requirement for such matching annotation, which enables easy expansion of dataset scales and thus facilitates future large deep learning models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.164, + 0.29, + 0.18 + ], + "angle": 0, + "content": "3. A Benchmark Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.473, + 0.477 + ], + "angle": 0, + "content": "The goal of Weakly-supervised Video Individual Counting (WVIC) is to predict the number of unique individuals \\((\\hat{N})\\) in a video provided a series of its frames with \\(\\delta\\) as the sampling rate. Given a video \\(\\mathbf{V}\\) consisting of \\(T\\) frames \\(\\mathbf{V}^1, \\dots, \\mathbf{V}^T\\), the label information comprises coordinate/inflow/outflow data for each individual. Specifically, letting \\(\\mathbf{P}^i = [p_1^i, \\dots, p_{n_i}^i]\\) denote the annotation of the coordinates of the center of the people in frame \\(\\mathbf{V}^i\\), we sample the frame pairs \\(\\{(\\mathbf{V}^1, \\mathbf{V}^{\\delta + 1}), (\\mathbf{V}^{\\delta + 1}, \\mathbf{V}^{2\\delta + 1}), \\dots, (\\mathbf{V}^{T - \\delta}, \\mathbf{V}^T)\\}\\), where \\(\\delta\\) is an integer frame offset. The weakly supervised inflow label \\(\\mathbf{I}^i = [\\mathbf{I}_1^i, \\dots, \\mathbf{I}_{n_i}^i]\\) and outflow label \\(\\mathbf{O}^i = [\\mathbf{O}_1^i, \\dots, \\mathbf{O}_{n_i}^i]\\). If the \\(u\\)-th person in frame \\(i\\) doesn't appear in \\(\\mathbf{V}^{i - \\delta}\\), then \\(\\mathbf{I}_u^i = 1\\); otherwise, \\(\\mathbf{I}_u^i = 0\\). For \\(\\mathbf{O}^i\\), if the \\(u\\)-th person in frame \\(V^i\\) doesn't appear in \\(\\mathbf{V}^{i + \\delta}\\), then \\(\\mathbf{O}_u^i = 1\\); otherwise \\(\\mathbf{O}_u^i = 0\\). If, by a slight abuse of notation, we refer to people by their coordinates in particular images, we see that the set \\(\\{p_u^i | \\mathbf{O}_u^i = 0\\}\\) and \\(\\{p_u^{i + \\delta} | \\mathbf{I}_u^i = 0\\}\\) represent the people that appear both in \\(\\mathbf{V}^i\\) and \\(\\mathbf{V}^{i + \\delta}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.457, + 0.501 + ], + "angle": 0, + "content": "3.1. Weakly Supervised Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.703 + ], + "angle": 0, + "content": "As shown in Fig. 2, the inference pipeline of our CGNet consists of three components: an image-level locator to generate coordinates for pedestrians, an encoder to produce representations based on pedestrians' coordinates, and a Memory-based individual Count Predictor (MCP) to predict the final inflow count based on the representations. The locator is trained independently with the coordinate annotations, and existing image crowd localization networks such as FIDT [25] can be used. The encoder is trained with our Weakly Supervised Representation Learning method (WSRL) to extract the discriminative features of each individual. It can be any feature extractor, such as ConvNext [26]. MCP does not need training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Provided two \\(\\delta\\)-adjacent frames \\(\\mathbf{V}^i\\) and \\(\\mathbf{V}^j\\) and the corresponding annotations \\(\\mathbf{P}^i\\), \\(\\mathbf{P}^j\\), \\(\\mathbf{O}^i\\) and \\(\\mathbf{I}^j\\), we group the individuals into four sets: \\(\\mathbf{X} = \\{p_u^i | \\mathbf{O}_u^i = 0\\}\\), \\(\\mathbf{Y} = \\{p_u^j | \\mathbf{I}_u^j = 0\\}\\), denoting shared individuals in the previous frame and the current frame, respectively; \\(\\hat{\\mathbf{X}} = \\mathbf{P}^i - \\mathbf{X}\\) and \\(\\hat{\\mathbf{Y}} = \\mathbf{P}^j - \\mathbf{Y}\\), denoting outflow individuals and inflow individuals, respectively. An encoder learns representations \\(\\mathbf{F}\\) for each pedestrian based on the crops around those coordinates. Specifically, to generate the \\(u\\)-th feature \\(f_u^i\\) in \\(\\mathbf{F}\\), if the locator outputs points, we extract a rectangle patch of size \\(96 \\times 64\\) centered at the predicted coordinate \\(p_u^i\\), and if the outputs are bounding boxes, we directly crop the bounding box from the original picture. Each cropped patch is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "resized to \\(224 \\times 224\\) and fed into the encoder, a ConvNext-S [26], generating a feature of \\(7 \\times 7 \\times 768\\), which is flattened and normalized into a 1D vector \\(f_{u}^{i}\\) to represent the final feature for the individual:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.155, + 0.892, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {F} _ {\\mathbf {X}} ^ {i} = \\operatorname {E n c o d e r} (\\mathbf {X}), \\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i} = \\operatorname {E n c o d e r} (\\hat {\\mathbf {X}}) \\\\ \\bar {\\mathbf {F}} ^ {i} = \\bar {\\mathbf {F}} _ {\\hat {\\mathbf {X}}} + \\left(\\bar {\\mathbf {H}}\\right) \\bar {\\mathbf {F}} ^ {i} = \\bar {\\mathbf {F}} _ {\\hat {\\mathbf {H}}} + \\left(\\hat {\\mathbf {H}}\\right) \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.178, + 0.836, + 0.198 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathbf {Y}} ^ {j} = \\operatorname {E n c o d e r} (\\mathbf {Y}), \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} = \\operatorname {E n c o d e r} (\\hat {\\mathbf {Y}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.203, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Then, a similarity matrix \\(\\mathbf{S} \\in \\mathbb{R}^{n_i \\times n_j}\\) between \\(\\mathbf{F}^i = \\left[\\mathbf{F}_{\\mathbf{X}}^i, \\mathbf{F}_{\\hat{\\mathbf{X}}}^i\\right]\\) and \\(\\mathbf{F}^j = \\left[\\mathbf{F}_{\\mathbf{Y}}^j, \\mathbf{F}_{\\hat{\\mathbf{Y}}}^j\\right]\\) is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.254, + 0.892, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\left[ \\begin{array}{l l} \\left(\\mathbf {F} _ {\\mathbf {X}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\mathbf {Y}} ^ {j} & \\left(\\mathbf {F} _ {\\mathbf {X}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} \\\\ \\left(\\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\mathbf {Y}} ^ {j} & \\left(\\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} \\end{array} \\right] = \\left[ \\begin{array}{l l} \\mathbf {S} _ {0} & \\mathbf {S} _ {1} \\\\ \\mathbf {S} _ {2} & \\mathbf {S} _ {3} \\end{array} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.406 + ], + "angle": 0, + "content": "The matrix \\(\\mathbf{S}\\) is divided into four parts: \\(\\mathbf{S}_0\\in \\mathbb{R}^{m_{i,j}\\times m_{i,j}}\\) \\(\\mathbf{S}_1\\in \\mathbb{R}^{m_{i,j}\\times (n_j - m_{i,j})}\\) \\(\\mathbf{S}_2\\in \\mathbb{R}^{(n_i - m_{i,j})\\times m_{i,j}}\\) \\(\\mathbf{S}_3\\in\\) \\(\\mathbb{R}^{(n_i - m_{i,j})\\times (n_j - m_{i,j})}\\), where \\(m_{i,j} = (\\mathbf{1} - \\mathbf{I}^{\\mathbf{j}})^T(\\mathbf{1} - \\mathbf{O}^i)\\) denotes the number of shared individuals in \\(\\mathbf{V}^i\\) and \\(\\mathbf{V}^j\\). \\(\\mathbf{S}_0\\) is the similarity matrix of \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\), while \\(\\mathbf{S}_1\\), \\(\\mathbf{S}_2\\) and \\(\\mathbf{S}_3\\) are the similarity matrix of \\(\\mathbf{X}\\) and \\(\\hat{\\mathbf{Y}}\\), of \\(\\hat{\\mathbf{X}}\\) and \\(\\mathbf{Y}\\), of \\(\\hat{\\mathbf{X}}\\) and \\(\\hat{\\mathbf{Y}}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.467 + ], + "angle": 0, + "content": "To pull the matched groups (X and Y) closer and push away individual pairs from unmatched groups, we propose a weakly supervised Group-level Matching Loss (GML) to constrain different parts of S." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Constraint for \\(\\mathbf{S}_0, \\mathbf{S}_1, \\mathbf{S}_2\\). \\(\\mathbf{S}_0\\) denotes the similarity among individuals shared between two frames, and it does not matter whether the individuals are correctly matched as long as we can assign them a one-to-one match. As for \\(\\mathbf{S}_1\\) and \\(\\mathbf{S}_2\\), they are the similarities of different identities and thus should be zero." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.68 + ], + "angle": 0, + "content": "To this end, we introduce a latent variable matrix \\(\\Omega \\in \\mathbb{R}^{m_{i,j}\\times m_{i,j}}\\), where the \\((u,v)\\)-th element \\(\\Omega_{u,v}\\) denote the probability that the \\(u\\)-th pedestrian in \\(\\mathbf{X}\\) is matched with \\(v\\)-th pedestrian in \\(\\mathbf{Y}\\). To increase the similarity between the pairs that have a higher matching probability and decrease the similarity between the pairs that have a lower matching probability, inspired by [32, 39], we define the soft contrastive loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.684, + 0.892, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s c o n} (i) = \\min _ {\\Omega} - \\sum_ {u = 1} ^ {m _ {i, j}} \\sum_ {v = 1} ^ {m _ {i, j}} \\Omega_ {u, v} C _ {u, v}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.727, + 0.86, + 0.747 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad \\mathbf {1} _ {m _ {i, j}} ^ {T} \\Omega = \\mathbf {1}, \\quad \\Omega \\mathbf {1} _ {m _ {i, j}} = \\mathbf {1}, 0 \\leq \\Omega_ {u, v} \\leq 1,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.751, + 0.892, + 0.78 + ], + "angle": 0, + "content": "where \\(C_{u,v}\\) denotes the contrastive similarity [10], and is calculated as," + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.785, + 0.892, + 0.836 + ], + "angle": 0, + "content": "\\[\nC _ {u, v} = \\frac {e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v}}}{e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v}} + \\sum_ {1 \\leq u ^ {\\prime} \\leq n _ {i}} ^ {u ^ {\\prime} \\neq u} e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u ^ {\\prime} , v}} + \\sum_ {1 \\leq v ^ {\\prime} \\leq n _ {j}} ^ {v ^ {\\prime} \\neq v} e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v ^ {\\prime}}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a temperature hyper-parameter [52] and \\(\\mathbf{S}_{u,v} = \\langle f_u^i,f_v^j\\rangle\\) is the \\((u,v)\\)-th element of \\(\\mathbf{S}\\), calculated by the \\(u\\)-th feature in \\(\\mathbf{F}^i\\) and the \\(v\\)-th feature in \\(\\mathbf{F}^j\\). As \\(\\Omega_{u,v}\\) describes the matching probability, our soft contrastive loss" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.095, + 0.85, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.895, + 0.445 + ], + "angle": 0, + "content": "Figure 2. The inference pipeline of our CGNet and the weakly supervised representation learning method (WSRL). The pipeline comprises a frame-level crowd locator, an encoder, and a Memory-based individual count predictor (MCP). The locator predicts the coordinates for pedestrians. The encoder generates representations for each individual, and MCP predicts inflow counts and updates the individual templates stored in the memory. To pull the matched groups \\((\\mathbf{X}\\) and \\(\\mathbf{Y})\\) closer and push away individual pairs from unmatched groups, WSRL exploits inflow and outflow labels to optimize the encoder with a novel Group level Matching Loss (GML), which consists of a soft contrastive loss \\((\\mathcal{L}_{scon})\\) and a hinge loss \\((\\mathcal{L}_{hinge})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.473, + 0.558 + ], + "angle": 0, + "content": "evaluates the expectation of contrastive loss. It should be noting that in \\( C_{u,v} \\), \\( (f_u^i, f_v^j) \\) is considered as positive pairs, while \\( \\{(f_u^i, f_{v'}^j) | v' \\neq v\\} \\) and \\( \\{(f_{u'}, f_v^j) | u' \\neq u\\} \\) are considered as negative pairs. Thus, \\( \\mathcal{L}_{scon} \\) can encourage \\( S_0 \\) to be a permutation matrix, which indicates that there exists a one-to-one match between \\( X \\) and \\( \\hat{Y} \\), and drive \\( S_1, S_2 \\) to be zero matrices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.64 + ], + "angle": 0, + "content": "The problem defined by Eq. (3) is a typical balanced Optimal Transport (OT [35]) problem which transports \\( \\mathbf{1}_{m_{i,j}} \\) to \\( \\mathbf{1}_{m_{i,j}} \\) using \\( \\mathbf{1} - C \\) as the cost matrix and \\( \\Omega \\) is the transport matrix. Therefore, we consider \\( \\mathcal{L}_{scon}(i) \\) as an OT loss, which can be solved with the Sinkhorn algorithm [38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.471, + 0.735 + ], + "angle": 0, + "content": "Constraint for \\(\\mathbf{S}_3\\). Obviously, the elements in \\(\\mathbf{S}_3\\) measure the similarity of different individuals, so they should be as small as possible. However, directly constraining \\(\\mathbf{S}_3\\) to zero matrix will make features from \\(\\hat{\\mathbf{X}}, \\hat{\\mathbf{Y}}\\) easily collapse to zero-vector. To this end, the widely used Hinge L1 loss [18] is adopted to constrain \\(\\mathbf{S}_3\\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.752, + 0.47, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h i n g e} (i) = \\frac {\\sum R e l u (\\mathbf {S _ {3}} - \\theta)}{(n _ {i} - m _ {i , j}) (n _ {j} - m _ {i , j})}, \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.804, + 0.449, + 0.82 + ], + "angle": 0, + "content": "where \\(\\theta\\) is a threshold used to ignore small values in \\(\\mathbf{S}_3\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.823, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Finally, the group-level matching loss of the training set \\(\\mathbf{V}\\) is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.87, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G M L} (\\mathbf {V}) = \\sum_ {0 < i < T, i = k \\delta + 1} \\left(\\mathcal {L} _ {s c o n} (i) + \\mathcal {L} _ {h i n g e} (i)\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.452, + 0.783, + 0.467 + ], + "angle": 0, + "content": "3.2. Memory-based Count Predictor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.625 + ], + "angle": 0, + "content": "To handle individuals' appearance variance and re-entering cases, inspired by [22, 33, 55], we propose a memory-based count predictor (MCP) to reason the final count for the video, which stores templates of recently appeared individuals in a flexible memory. Specifically, when processing frame \\(\\mathbf{V}^{i + \\delta}\\), the memory is denoted as \\(\\mathcal{G}^i = \\{\\mathcal{G}_1^i,\\mathcal{G}_2^i,\\dots ,\\mathcal{G}_{\\mathbf{N}^i}^i\\}\\), where \\(\\mathbf{N}^i\\) is the memory size at time \\(i\\). \\(\\mathcal{G}_u^i = \\{g_{u,0}^i,\\dots,g_{u,k}^i\\}\\) is a set of templates for \\(u\\)-th pedestrian in the memory, and \\(k\\) is the number of templates stored for \\(u\\)-th pedestrian." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.719 + ], + "angle": 0, + "content": "We first use the crowd locator to predict the coordinates of pedestrians in \\(\\mathbf{V}^{i + \\delta}\\), and then extract their corresponding features, denoted as \\(\\{f_1^{i + \\delta}, f_2^{i + \\delta}, \\dots, f_{n_{i + \\delta}}^{i + \\delta}\\}\\). To associate the individuals in \\(\\mathbf{V}^{i + \\delta}\\) with templates in memory, we first define the cost \\(\\hat{C}_{u,u'}\\) of matching the \\(u\\)-th pedestrian with the \\(u'\\)-th template in \\(\\mathcal{G}^i\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.724, + 0.892, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\hat {C} _ {u, u ^ {\\prime}} = \\max _ {v} \\left(1 - \\left\\langle f _ {u} ^ {i + \\delta}, g _ {u ^ {\\prime}, v} ^ {i} \\right\\rangle\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.754, + 0.894, + 0.801 + ], + "angle": 0, + "content": "Then, we generate an optimal one-to-one match \\(\\pi \\in \\{0,1\\}^{n_{i + \\delta}\\times \\mathbf{N}^i}\\) (assuming \\(\\mathbf{N}^i\\geq n_i\\)) by solving the following problem with Hungarian Algorithm [21]:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.807, + 0.892, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\pi = \\arg \\min _ {\\pi} \\sum_ {u, u ^ {\\prime}} \\pi_ {u, u ^ {\\prime}} \\hat {C} _ {u, u ^ {\\prime}} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.844, + 0.85, + 0.874 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad \\forall u ^ {\\prime}, \\sum_ {u} \\pi_ {u, u ^ {\\prime}} = 1, \\quad \\forall u, \\sum_ {u} \\pi_ {u, u ^ {\\prime}} \\leq 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.883, + 0.892, + 0.904 + ], + "angle": 0, + "content": "If the matching cost for \\(u\\) meets \\(\\sum_{u' = 1}^{\\mathbf{N}^i}\\pi_{u,u'}\\hat{C}_{u,u'} > \\zeta\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19231" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "\\(u\\) will be judged as an inflow pedestrian. Otherwise, if \\(\\pi_{u,u'} = 1\\), \\(u\\) is associated with \\(g_{u'}^i\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Template update. Similar to [22], \\(\\mathcal{G}_u^i\\) has a time-to-live factor (ttl). \\(ttl_u\\) will decrease one if \\(\\mathcal{G}_u^i\\) has no mapping to any pedestrian feature in \\(\\mathbf{F}^{i + \\delta}\\). If \\(ttl_u\\) decreases to zero, the template \\(\\mathcal{G}_u^i\\) will be dropped. Meanwhile, whenever \\(\\mathcal{G}_u^i\\) is associated with a pedestrian in the current frame, \\(ttl_u\\) will be reset to \\(ttlmax\\). This time-to-live threshold allows us to save some recently appeared targets while discarding those that have not appeared for a long time. If pedestrian \\(u\\) is associated with \\(u'\\)-th template, we add \\(f_{u}^{i + \\delta}\\) to \\(\\mathcal{G}_{u'}^i\\) to update the template" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.285, + 0.47, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {u ^ {\\prime}} ^ {i + \\delta} \\leftarrow \\mathcal {G} _ {u ^ {\\prime}} ^ {i} \\cup \\left\\{f _ {u} ^ {i + \\delta} \\mid \\pi_ {u, u ^ {\\prime}} = 1, \\sum_ {u ^ {\\prime} = 1} ^ {\\mathbf {N} ^ {i}} \\pi_ {u, u ^ {\\prime}} \\hat {C} _ {u, u ^ {\\prime}} \\leq \\zeta \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.469, + 0.345 + ], + "angle": 0, + "content": "If pedestrian \\( u \\) is an inflow pedestrian, \\( \\{f_u^{i + \\delta}\\} \\) will be added to the memory \\( \\mathcal{G}^{i + \\delta} \\) as a new template." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.346, + 0.469, + 0.375 + ], + "angle": 0, + "content": "The output number of the total individuals in video \\( I \\) is then calculated via" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.385, + 0.469, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\hat {N} = n _ {1} + \\sum_ {\\delta + 1 \\leq i \\leq T, i = k \\delta + 1} \\hat {n} _ {i}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.429, + 0.438, + 0.445 + ], + "angle": 0, + "content": "where \\(\\hat{n}_i\\) is the number of inflow pedestrians at time \\(i\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.458, + 0.21, + 0.475 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.174, + 0.497 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We test our CGNet on three datasets: CroHD [42], SenseCrowd [24], and UAVVIC. CroHD has four training videos and five testing videos. SenseCrowd contains 634 videos, and we split the train, validation, and test dataset following DRNet [14]. UAVVIC is our proposed dataset collected by a moving UAV camera in various scenes, including campus, beach, car park, highway, city road, and square. UAVVIC consists of 221 videos (100 for training, 100 for testing, and 21 for validation), and 5,396 frames are sampled with 3s as the interval. Annotations consist of 398,158 bounding boxes, and group-level matching labels in neighbor frames are provided. The resolutions of UAVVIC are 4K and 1080P for better capture of tiny pedestrians from drone view. A detailed comparison of UAVVIC with existing video crowd counting datasets [8, 9, 13, 24, 42] is shown in Tab. 1. UAVVIC provided a moving shot with a larger range of pedestrians. More details about UAVVIC are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.174, + 0.802 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Mean Absolute Error (MAE), Mean Square Error (MSE), and Weighted Relative Absolute Errors (WRAE) are used for evaluation. The first two metrics are common metrics applied in VCC. However, unlike VCC schemes, we only count the same person once in the video. WRAE [14] is proposed to balance the performance on videos with" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.178 + ], + "angle": 0, + "content": "
NameResolutionRangeMoving shotPointBoxTrajectoryIn-Out
Mall [9]640*48013-53X✓XXX
UCSD [8]238*15811-46X✓XXX
FDST [13]1920*10809-57X✓XXX
1280*720
SenseCrowd [24]-1-296X✓✓✓X
CroHD [42]-25-346X✓✓✓X
UAVVIC3840*21600-735✓✓✓X✓
1920*1080
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.187, + 0.887, + 0.202 + ], + "angle": 0, + "content": "Table 1. Comparison of different video crowd counting datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.212, + 0.894, + 0.284 + ], + "angle": 0, + "content": "different lengths and pedestrian numbers: \\( WRAE = \\sum_{i=1}^{K} \\frac{T_i}{\\sum_{j=1}^{K} T_j} \\frac{|N_i - \\hat{N}_i|}{N_i} \\), where \\( K \\) is the total number of videos, and \\( T_i \\) is the length of video \\( i \\), \\( N_i / \\hat{N}_i \\) denotes the number of ground truth/predicted counts in \\( i \\)-th video." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.296, + 0.719, + 0.312 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Training. If there are no special instructions, the loca-tor we applied is a well-known crowd localization model, FIDT [25]. All parameters are official from FIDT, except that we replaced the backbone from HRNet-W48 [47] to HRNet-W18 [47] for efficient inference. The learning rate is set as \\( 1e^{-4} \\) along with AdamW [28] as the optimizer and applying the pre-train weight from Timm [49]. All related models can be trained on one RTX3090 (24G memory)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Testing. In the testing phase, the time interval \\(\\delta\\) is set as \\(3s\\) following the setting in [14], the threshold \\(\\zeta\\) is set as \\(0.7\\), \\(\\gamma\\) in Eq. (4) is set as 10, and the max time-to-live (ttlmax) is set as 3, the max size of memory (memmax) is set as 5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.698, + 0.529 + ], + "angle": 0, + "content": "4.4. Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Comparison Methods. To the best of our knowledge, only DRNet[14] has been specifically designed for the VIC task. To evaluate the effectiveness of our proposed CGNet, we also tested two categories of approaches that could be applied to VIC: 1) Multi-object tracking (MOT) methods, including HeadHunter-T [43], FairMOT [56], PHDTT [44], SMILE [48], SparseTrack [27], Deep-OC-SORT [31], and BoT-SORT [1], where we consider the total number of indices in the entire video as individual count. We set the frame rate to 1 FPS for HeadHunter-T and 0.33 FPS for the other MOT methods to obtain better VIC performance. 2) A recent cross-line video crowd-counting, LOI [57], which also requires trajectory labels in training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on SenseCrowd: As shown in Tab. 2, it is evident that our CGNet exhibits superior performance on this larger dataset, SenseCrowd. Particularly, CGNet improved the overall MAE by about \\(28\\%\\) compared to DRNet. Furthermore, we achieve the lowest MAE across all density levels except for D0. Compared to the existing lowest MAE on D1, D2, D3, and D4, the performance improvements of our CGNet are about \\(27\\%\\), \\(63\\%\\), \\(37\\%\\), and \\(17\\%\\), respectively. In Fig. 3, we show qualitative results on three representative scenes,i.e., mall, station, and outdoor plaza. Although the inflows in the current flow are not concentrated, CGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19232" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.088, + 0.373, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.089, + 0.591, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.089, + 0.811, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.185, + 0.546, + 0.196 + ], + "angle": 0, + "content": "(a) Previous frame (0s)" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.197, + 0.374, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.292, + 0.227, + 0.301 + ], + "angle": 0, + "content": "GT inflow: 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.292, + 0.372, + 0.3 + ], + "angle": 0, + "content": "Predict inflow: 4" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.197, + 0.591, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.292, + 0.443, + 0.3 + ], + "angle": 0, + "content": "GT inflow: 11" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.197, + 0.807, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.292, + 0.662, + 0.301 + ], + "angle": 0, + "content": "GT inflow: 15" + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.292, + 0.802, + 0.3 + ], + "angle": 0, + "content": "Predict inflow: 14" + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.309, + 0.545, + 0.321 + ], + "angle": 0, + "content": "(b) Current frame (3s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.894, + 0.359 + ], + "angle": 0, + "content": "Figure 3. Results on the SenseCrowd dataset. Red dots are predictions of the locator. Blue/green/yellow circles are correct predicted inflow, error predicted inflow, and missed inflow, respectively." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.361, + 0.853, + 0.525 + ], + "angle": 0, + "content": "
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
FairMOT [56]35.462.348.913.522.467.984.4145.8
HeadHunter-T [43]30.050.638.611.825.756.092.6131.4
SMILE [48]27.2236.732.59.221.033.876.5203.5
SparseTrack [27]28.3643.830.89.921.542.189.2189.5
Deep-OC-SORT [31]26.0443.629.59.525.129.160.3156.5
BoT-SORT [1]24.6732.428.38.720.531.170.2165.1
LOI [57]24.733.137.412.525.439.339.686.7
DRNet [14]12.324.712.74.18.023.350.077.0
CGNet (Ours)8.8617.6912.65.05.88.525.063.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.536, + 0.893, + 0.565 + ], + "angle": 0, + "content": "Table 2. Performance comparison on SenseCrowd. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and \\([200, +\\infty)\\), respectively. The best values are highlighted in blue font." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.57, + 0.855, + 0.749 + ], + "angle": 0, + "content": "
MethodMAEMSEWRAE(%)MAE on five testing scenes
CroHD11CroHD12CroHD13CroHD14CroHD15
PHDTT [44]2130.42808.3401.6247379347944911327
FairMOT [56]256.2300.844.1114272844081000
HeadHunter-T [43]253.2351.732.7651015155821003
SMILE [48]257.6334.540.516231156649246
SparseTrack [27]176.6208.627.62917225833688
Deep-OC-SORT [31]165.2195.933.16835118616160
BoT-SORT [1]154.8176.4227.44921013828691
LOI [57]305.0371.146.060243458630131
DRNet [14]141.1192.327.4313381825561
CGNet (Ours)75.095.114.577214144138
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.759, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Table 3. Performance comparison on CroHD dataset. CroHD11-CroHD15 are five test scenes labeled in [14]. The best values are highlighted in blue font." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.833 + ], + "angle": 0, + "content": "can accurately predict them. More visualization results can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Results on CroHD. Comparison with existing methods on CroHD is shown in Tab. 3. Our CGNet achieved the lowest MAE, MSE, and WRAE. Especially, except for CroHD15, our method obtained the lowest MAE across the other four" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.893, + 0.895 + ], + "angle": 0, + "content": "testing scenes. As CGNet is trained with only point annotations, it cannot crop the patch of optimal size to extract features on CroHD15. It should be noted that on CroHD13 and CroHD14, which have the highest average density (245.9 and 259.6 person/frame), the performances of all the compared methods except for DRNet degrade sig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19233" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.852, + 0.183 + ], + "angle": 0, + "content": "
MethodMAE/MSE/WRAEMAE/MSE/WRAE on four different categories
PedestrianCarBusVan
BoT-SORT [1]49.9/228.2/29.992.4/398.2/49.059.1/259.1/39.223.0/109.6/57.125.1/146.2/77.3
Deep-OC-SORT [31]37.1/115.4/35.285.2/213.5/22.133.4/113.1/21.019.1/ 78.2/26.210.7/56.6/32.8
DRNet [14]18.4/ 41.0/15.234.5/ 92.1/19.119.0/ 32.7/17.810.4/ 22.1/32.29.9/17.2/29.4
CGNet (Ours)12.9/ 37.4/ 12.031.2/ 98.4/14.416.9/ 45.2/ 7.32.1/ 4.1/19.21.3/ 2.0/25.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.193, + 0.894, + 0.222 + ], + "angle": 0, + "content": "Table 4. Performance comparison on UAVVIC. All the methods are trained independently with four classes. The values of each entry are MAE/MSE/WRAE. The best values are highlighted in blue font." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.245, + 0.468, + 0.319 + ], + "angle": 0, + "content": "
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
①13.3532.322.89.712.315.326.543.4
②9.8420.013.85.36.312.028.867.7
③10.7117.4016.96.98.911.724.350.6
④8.8617.6912.65.05.88.525.063.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Table 5. Ablation study on the main components. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and \\([200, +\\infty)\\), respectively. \\(①\\), No constrain on \\(\\mathbf{S}_1\\), \\(\\mathbf{S}_2\\); \\(②\\), w/o Eq. (5); \\(③\\), w/o MCP; \\(④\\), full CGNet. The best values are highlighted in blue font." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.513 + ], + "angle": 0, + "content": "nificantly. In contrast, the performance of our CGNet is relatively stable in these two scenes. Furthermore, even without individual-level matching labels, our CGNet performs favorably against the state-of-the-art approach, DRNet, by a large margin (about \\(46\\%\\) on the MAE), demonstrating the effectiveness of our method." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.522, + 0.272, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.615, + 0.242, + 0.626 + ], + "angle": 0, + "content": "(a) Previous frame (0s)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.523, + 0.454, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.615, + 0.422, + 0.626 + ], + "angle": 0, + "content": "(b) Current frame (3s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.629, + 0.458, + 0.642 + ], + "angle": 0, + "content": "Figure 4. Qualitative visualization of our CGNet on UAVVIC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Results on UAVVIC. The results on our proposed UAVVIC dataset are shown in Tab. 4. All methods are trained and tested on each category, respectively. Since there is no trajectory annotation on UAVVIC, we train the compared methods [1, 14, 31] on VisDrone [59], a UAV dataset for MOT. Our CGNet shows favorable performance in all categories. The average MAE is improved by about \\(30\\%\\) compared to DRNet. In Fig. 4, we show the quality result of our CGNet on the UAVVIC dataset. Although the UAV moves, CGNet can still detect inflow objects accurately." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.241, + 0.817 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We conducted several ablation studies on the SenseCrowd. 1). We verify the effectiveness of the main components of our CGNet. 2). We analyze the effect of two main parameters. 3). We report the performance of our CGNet with different locators. 4). We use the trajectories generated by" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.248, + 0.695, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.416, + 0.615, + 0.428 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.244, + 0.877, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.416, + 0.797, + 0.428 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.44, + 0.893, + 0.482 + ], + "angle": 0, + "content": "Figure 5. Effectiveness of MCP on SenseCrowd. (a) The average MAE with different lengths of videos with and w/o MCP. (b) Performance with different max time-to-live factor tllmax of MCP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.5, + 0.89, + 0.531 + ], + "angle": 0, + "content": "our CGNet as pseudo-labels for existing VIC methods. 5). We compare the time cost of the WVIC and VIC labels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.532, + 0.893, + 0.685 + ], + "angle": 0, + "content": "Effectiveness of Constrain for \\(\\mathbf{S}_1, \\mathbf{S}_2\\). Eq. (4) constrains \\(\\mathbf{S}_0, \\mathbf{S}_1, \\mathbf{S}_2\\) together. Without considering un-matched pairs in \\(\\mathbf{X} \\times \\hat{\\mathbf{Y}}\\) and \\(\\hat{\\mathbf{X}} \\times \\mathbf{Y}\\), it can be changed to constrain only \\(\\mathbf{S}_0\\) by remove elements in \\(\\mathbf{S}_1, \\mathbf{S}_2\\). As shown in Tab. 5, MAE dropped by about 4.51 without constraints on \\(\\mathbf{S}_1\\) and \\(\\mathbf{S}_2\\) (① VS ④). This is because elements in \\(\\mathbf{S}_1\\) and \\(\\mathbf{S}_2\\) are related to pedestrians that exist only in one frame. Considering \\(\\mathbf{X} \\times \\hat{\\mathbf{Y}}\\) and \\(\\hat{\\mathbf{X}} \\times \\mathbf{Y}\\) as negative pairs, our soft contrastive loss can push apart individuals in \\(\\hat{\\mathbf{X}} (\\hat{\\mathbf{Y}})\\) with that in \\(\\mathbf{Y} (\\mathbf{X})\\), thus enhancing the distinguishing ability of the representations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Effectiveness of Eq. (5). We replace the hinge MAE loss in Eq. (5) to normal L1 loss and derive the method ② in Tab. 5. Compared to L1 loss, our \\(\\mathcal{L}_{hinge}(\\cdot)\\) gives a soft margin to the inner product of two features other than encouraging them to be 0. Optimizing the inner product of all features toward 0 makes it easier to cause either side to collapse a 0 vector, which is not conducive to learning different representations. Compared the performance of ② with that of ④ in Tab. 5, \\(\\mathcal{L}_{hinge}(\\cdot)\\) improves MAE in all density levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of MCP. As shown in Tab. 5, compared \\(③\\) with \\(④\\), MCP brings about 4.1 improvements on WRAE. Specifically, we evaluate the average MAE of our CGNet on videos with different lengths. As shown in Fig. 5(a), the performance gap between \\(③\\) and \\(④\\) becomes larger as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19234" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.465, + 0.18 + ], + "angle": 0, + "content": "
δMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
510.9620.2619.56.56.910.429.279.9
48.7114.3312.94.24.910.227.870.1
38.8617.6912.65.05.88.525.063.4
29.7720.0117.37.18.812.417.452.6
19.5615.3416.66.87.410.523.145.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.19, + 0.468, + 0.218 + ], + "angle": 0, + "content": "Table 6. Ablation study on interval \\( \\delta \\) . The best values are highlighted in blue font." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.222, + 0.465, + 0.311 + ], + "angle": 0, + "content": "
MethodLocatorEncoderMAEMSEWRAE
①FIDTConvNext-S8.8617.699.27
②Yolov8 nanoConvNext-S8.0116.428.85
③Yolov8 smallConvNext-S7.5615.198.05
④VGG16+FPNConvNext-S9.2017.459.55
⑤VGG16+FPNPrRoiIPooling10.0118.229.70
⑥VGG16+FPNPrRoiIPooling12.5923.3212.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.322, + 0.468, + 0.377 + ], + "angle": 0, + "content": "Table 7. Performance on SenseCrowd of our scheme with different locators and feature extractors. In \\(⑥\\), MCP is replaced with the inflow reasoning method in DRNet. The best values are highlighted in blue font." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.468, + 0.447 + ], + "angle": 0, + "content": "the length of the video increases. This is because, with the stored templates of individuals, MCP can handle appearance variations and individuals' re-entry, alleviating error accumulation during inflow reasoning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.468, + 0.627 + ], + "angle": 0, + "content": "Effect of ttlmax. The performance with different max long-time-live factor ttlmax is shown in Fig. 5(b). Compared with no memory (i.e., ttlmax = 0), adding memory is helpful for all density levels except for extremely high-density levels (D4). It may be because high-density crowds result in too many candidate individual-level matches, making it difficult for our CGNet to choose the right one. Meanwhile, the performance degrades when ttlmax increases from 3 to 6 for that larger ttlmax = 6 leads to the templates of individuals that appeared a long time ago being stored in the memory. As the probability of these people re-entering the field is low, the recorded templates become distractions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Effect of Interval \\(\\delta\\). The performance with different intervals \\(\\delta\\) is shown in Tab. 6. The performance on higher-density video is better when \\(\\delta\\) is smaller. This is because crowds flow faster in those videos, and long intervals may lead to missing pedestrians in sampled frames. The overall best performance is achieved when \\(\\delta\\) is set as \\(4s\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Effect of the Locator. We replaced FIDT in our CGNet with Yolo V8 [19] nano and small and the localization branch (VGG16+FPN) used in DRNet, respectively. Different from FIDT, these three locators are trained with box annotations. For Yolo V8, the predicted bounding boxes are fed to our feature extractor to generate representations of individuals. For a fair comparison with DRNet, we also replace our feature extractor ConvNext-S with PrRoIPooling [17] used in DRNet [14] and remove MCP. As shown in Tab. 7, changing locators can further boost the performance of our method. With the same locator and feature extractor as DRNet (⑤), the MAE of our method is 10.01, 2.99 lower" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.888, + 0.216 + ], + "angle": 0, + "content": "
LocatorGround Truth LabelsPseudo Labels
MAEMSEWRAEMAEMSEWRAE
FairMOT35.462.348.937.264.849.8
HeadHunter-T30.050.638.632.554.239.9
SMILE27.236.732.527.537.035.2
SparseTrack28.443.830.830.545.031.5
Deep-OC-SORT26.043.629.529.545.331.7
BoT-SORT24.732.428.325.035.933.8
LOI24.733.137.426.935.239.9
DRNet12.324.712.714.525.513.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.226, + 0.89, + 0.254 + ], + "angle": 0, + "content": "Table 8. Performance of existing VIC methods trained with different trajectory labels." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.265, + 0.897, + 0.311 + ], + "angle": 0, + "content": "
(ni, ni+δ)(9,8)(60,53)(116,115)(289,292)(645,724)
Association [14]23.8 s104.5 s174.4 s586.9 s1663.1 s
In-Out4.4 s42.6 s53.7 s95.0 s571.9 s
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.322, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Table 9. Time cost for a pair of frames with two types of labeling methods. \\( n_i \\) denotes the number of pedestrians in frame \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.415 + ], + "angle": 0, + "content": "than that of DRNet. Even without MCP (⑥), the MAE of our method is only slightly larger than that of DRNet (12.59 v.s. 12.3), demonstrating that our method can learn effective representations without trajectory labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Performs as a Pseudo Trajectory Generator. The latent variable \\(\\Omega\\) is obtained by solving a problem defined by Eq. (3). The Hungarian Algorithm generates a one-to-one match using \\(\\Omega\\) as the cost matrix. Based on the matching, we generate pseudo trajectory labels, which are used to train existing VIC methods. We show the results on SenseCrowd in Tab. 8. With the generated pseudo label, the performance of all the methods degrades slightly than using the ground truth trajectory labels, demonstrating that exact individual-level matching annotations may not be necessary for VIC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Label time cost. We compare the time cost of annotating our weakly supervised label and association label [14] on five frame pairs with various densities in Tab. 9. The annotation time for our weakly supervised label is only about \\(30\\%\\) of that of the association label, making it easier to build a larger dataset for this VIC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.627, + 0.696 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.887 + ], + "angle": 0, + "content": "We propose a new task, Weakly supervised Video Individual Counting (WVIC), in this paper. Unlike conventional video individual counting tasks, which require expensive trajectory labels, our WVIC task just requires two types of easily available identity-agnostic annotations. This largely reduces the annotation cost and facilitates future works to upscale data size. Furthermore, we propose a strong baseline for the WVIC task. The key idea is a Group-level Matching Loss (GML), which makes it easier to distinguish inflow individuals from all objects. Last, we have extended two existing datasets and built a totally new dataset for the WVIC task. Our method achieves favorable performance on all three datasets, even compared to supervised methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19235" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.196 + ], + "angle": 0, + "content": "[2] Antonio Albiol, Inmaculada Mora, and Valery Naranjo. Real-time high density people counter using morphological tools. IEEE TITS, 2(4):204-218, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.197, + 0.471, + 0.25 + ], + "angle": 0, + "content": "[3] Marco Allodi, Alberto Broggi, Domenico Giaquinto, Marco Patander, and Antonio Prioletti. Machine learning in tracking associations with stereo vision and lidar observations for an autonomous vehicle. In IVS, pages 648-653, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[4] Seung Hwan Bae and Kuk-Jin Yoon. Confidence-based data association and discriminative deep appearance learning for robust online multi-object tracking. IEEE TPAMI, 40(3):595-610, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.471, + 0.345 + ], + "angle": 0, + "content": "[5] Javier Barandiarán, Berta Murguia, and Fernando Boto. Real-time people counting using multiple lines. In WIAMIS, pages 159-162, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[6] Jesús Bescós, José M. Menéndez, and Narciso García. DCT based segmentation applied to a scalable zenithal people counter. In ICIP, pages 1005-1008, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[7] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, pages 1-6, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[8] Antoni B. Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE TPAMI, 30(5):909-926, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.506 + ], + "angle": 0, + "content": "[9] Ke Chen, Chen Change Loy, Shaogang Gong, and Tony Xiang. Feature mining for localised crowd counting. In BMVC, pages 3-14, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[11] Li Dong, Haijun Zhang, Jianghong Ma, Xiaofei Xu, Yimin Yang, and Q. M. Jonathan Wu. Clrnet: A cross locality relation network for crowd counting in videos. IEEE TNNLS, pages 1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.47, + 0.64 + ], + "angle": 0, + "content": "[12] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. In ICME, pages 814-819, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.47, + 0.693 + ], + "angle": 0, + "content": "[13] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. arXiv preprint arXiv:1907.07911, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.694, + 0.47, + 0.735 + ], + "angle": 0, + "content": "[14] Tao Han, Lei Bai, Junyu Gao, Qi Wang, and Wanli Ouyang. DR.VIC: decomposition and reasoning for video individual counting. In CVPR, pages 3073-3082, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[15] Yuhang He, Jie Han, Wentao Yu, Xiaopeng Hong, Xing Wei, and Yihong Gong. City-scale multi-camera vehicle tracking by semantic attribute parsing and cross-camera tracklet matching. In CVPRW, pages 2456-2465, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[16] Andinet Hunde and Beshah Ayalew. Automated multi-target tracking in public traffic in the presence of data association uncertainty. In ACC, pages 300-306, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.47, + 0.869 + ], + "angle": 0, + "content": "[17] Borui Jiang, Ruixuan Luo, Jiayuan Mao, Tiantian Xiao, and Yuning Jiang. Acquisition of localization confidence for accurate object detection. In ECCV, pages 784-799, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.87, + 0.47, + 0.897 + ], + "angle": 0, + "content": "[18] Thorsten Joachims. Support vector method for multivariate performance measures. In ICML, pages 377-384, 2005. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[20] Chanho Kim, Fuxin Li, Mazen Alotaibi, and James M. Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, pages 9553-9562, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.227 + ], + "angle": 0, + "content": "[21] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.228, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[22] L.J. Latecki and R. Miezianko. Object tracking with dynamic template update and occlusion detec. In ICPR, volume 1, pages 556-560, 2006. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.269, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[23] Eui-Hyuk Lee, Qian Zhang, and Taek Lyul Song. Markov chain realization of joint integrated probabilistic data association. Sensors, 17(12):2865, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[24] Haopeng Li, Lingbo Liu, Kunlin Yang, Shinan Liu, Junyu Gao, Bin Zhao, Rui Zhang, and Jun Hou. Video crowd localization with multifocus gaussian neighborhood attention and a large-scale benchmark. IEEE TIP, 31:6032-6047, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[25] Dingkang Liang, Wei Xu, Yingying Zhu, and Yu Zhou. Focal inverse distance transform maps for crowd localization. IEEE TMM, 25:6040-6052, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.444, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[27] Zelin Liu, Xinggang Wang, Cheng Wang, Wenyu Liu, and Xiang Bai. Sparsetrack: Multi-object tracking by performing scene decomposition based on pseudo-depth. arXiv preprint arXiv:2306.05238, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In *ICLR*, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.524, + 0.892, + 0.577 + ], + "angle": 0, + "content": "[29] Zhongzhen Luo, Mina Attari, Saeid R. Habibi, and Martin von Mohrenschild. Online multiple maneuvering vehicle tracking system based on multi-model smooth variable structure filter. IEEE TITS, 21(2):603-616, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.578, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[30] Zheng Ma and Antoni B. Chan. Counting people crossing a line using integer programming and local features. IEEE TCSVT, 26(10):1955-1969, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[31] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[32] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[33] Zhuojin Pan and Xiujuan Wang. Correlation tracking algorithm based on adaptive template update. In ICISP, volume 1, pages 98-101, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.753, + 0.892, + 0.805 + ], + "angle": 0, + "content": "[34] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, pages 164-173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.806, + 0.892, + 0.847 + ], + "angle": 0, + "content": "[35] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning, 11(5-6):355-607, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.848, + 0.892, + 0.887 + ], + "angle": 0, + "content": "[36] Christopher Rasmussen and Gregory D. Hager. Probabilistic data association methods for tracking complex visual objects. IEEE TPAMI, 23(6):560-576, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.888, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[37] Christian Ritter, Andrea Imle, Ji Young Lee, Barbara Muller," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19236" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Oliver T. Fackler, Ralf Bartenschlager, and Karl Rohr. Two-filter probabilistic data association for tracking of virus particles in fluorescence microscopy images. In *ISBI*, pages 957–960, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[38] Richard Sinkhorn. A Relationship Between Arbitrary Positive Matrices and Doubly Stochastic Matrices. The Annals of Mathematical Statistics, 35(2):876 - 879, 1964. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.188, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[39] Rakshith Sharma Srinivasa, Jaejin Cho, Chouchang Yang, Yashas Malur Saidutta, Ching-Hua Lee, Yilin Shen, and Hongxia Jin. CWCL: Cross-modal transfer with continuously weighted contrastive loss. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[40] Jason Stauch, Travis Bessell, Mark Rutten, Jason Baldwin, Moriba Jah, and Keric Hill. Joint probabilistic data association and smoothing applied to multiple space object tracking. JGCD, 41(1):19-33, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.337 + ], + "angle": 0, + "content": "[41] Ramana Sundararaman, Cedric De Almeida Braga, Éric Marchand, and Julien Pettré. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.336, + 0.469, + 0.377 + ], + "angle": 0, + "content": "[42] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[43] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, June 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.417, + 0.469, + 0.47 + ], + "angle": 0, + "content": "[44] Xuan-Thuy Vo, Van-Dung Hoang, Duy-Linh Nguyen, and Kang-Hyun Jo. Pedestrian head detection and tracking via global vision transformer. In IFCVW, pages 155-167, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[45] Xingyu Wan, Jinjun Wang, Zhifeng Kong, Qing Zhao, and Shunming Deng. Multi-object tracking using online metric learning with long short-term memory. In ICIP, pages 788-792, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[46] Bing Wang, Gang Wang, Kap Luk Chan, and Li Wang. Tracklet association by online target-specific metric learning and coherent dynamics estimation. IEEE TPAMI, 39(3):589-602, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[48] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking. arXiv preprint arXiv:2211.08824, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[49] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.469, + 0.794 + ], + "angle": 0, + "content": "[50] Hefeng Wu, Yafei Hu, Keze Wang, Hanhui Li, Lin Nie, and Hui Cheng. Instance-aware representation learning and association for online multi-person tracking. PR, 94:25-34, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.795, + 0.469, + 0.848 + ], + "angle": 0, + "content": "[51] Xingjiao Wu, Baohan Xu, Yingbin Zheng, Hao Ye, Jing Yang, and Liang He. Fast video crowd counting with a temporal aware network. Neurocomputing, 403:13-20, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.849, + 0.469, + 0.891 + ], + "angle": 0, + "content": "[52] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[53] Feng Xiong, Xingjian Shi, and Dit-Yan Yeung. Spatiotemporal modeling for crowd counting in videos. In ICCV, pages 5161-5169, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[54] Bo Yang, Chang Huang, and Ram Nevatia. Learning affinities and dependencies for multi-target tracking using a CRF model. In CVPR, pages 1233-1240, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.213 + ], + "angle": 0, + "content": "[55] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE TPAMI, 43(1):360-374, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.213, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129:3069–3087, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[57] Zhuoyi Zhao, Hongsheng Li, Rui Zhao, and Xiaogang Wang. Crossing-line crowd counting with two-phase deep neural networks. In ECCV, pages 712-726, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[58] Huicheng Zheng, Zijian Lin, Jiepeng Cen, Zeyu Wu, and Yadan Zhao. Cross-line pedestrian counting based on spatially-consistent two-stage local crowd density estimation and accumulation. IEEE TCSVT, 29(3):787-799, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[59] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE TPAMI, 44(11):7380-7399, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19237" + } + ] +] \ No newline at end of file diff --git a/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_origin.pdf b/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9f8d045b9a423348e3df03f39c2dfa5353ac78c --- /dev/null +++ b/2024/Weakly Supervised Video Individual Counting/357e37d5-1951-48de-9fbc-83246528162f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:343efe2e8957962d072caf37e05b84758d530022faff8c03c477f769da83e551 +size 3917401 diff --git a/2024/Weakly Supervised Video Individual Counting/full.md b/2024/Weakly Supervised Video Individual Counting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..14bab886de2c3a5d246d0b3812787fce88eba746 --- /dev/null +++ b/2024/Weakly Supervised Video Individual Counting/full.md @@ -0,0 +1,361 @@ +# Weakly Supervised Video Individual Counting + +Xinyan Liu1 Guorong Li1,3* Yuankai Qi2 Ziheng Yan1 +Zhenjun Han1 Anton van den Hengel5 Ming-Hsuan Yang6 Qingming Huang1,3,4 +1 University of Chinese Academy of Science; 2 Macquarie University; + +3 Key Lab of Big Data Mining and Knowledge Management, UCAS, China; 4 Key Lab of Intell. Info. Process., Inst. of Comput. Tech., CAS; + +$^{5}$ Australian Institute for Machine Learning, The University of Adelaide; $^{6}$ University of California at Merced. + +# Abstract + +Video Individual Counting (VIC) aims to predict the number of unique individuals in a single video. Existing methods learn representations based on trajectory labels for individuals, which are annotation-expensive. To provide a more realistic reflection of the underlying practical challenge, we introduce a weakly supervised VIC task, wherein trajectory labels are not provided. Instead, two types of labels are provided to indicate traffic entering the field of view (inflow) and leaving the field view (outflow). We also propose the first solution as a baseline that formulates the task as a weakly supervised contrastive learning problem under group-level matching. In doing so, we devise an end-to-end trainable soft contrastive loss to drive the network to distinguish inflow, outflow, and the remaining. To facilitate future study in this direction, we generate annotations from the existing VIC datasets SenseCrowd and CroHD and also build a new dataset, UAVVIC. Extensive results show that our baseline weakly supervised method outperforms supervised methods, and thus, little information is lost in the transition to the more practically relevant weakly supervised task. The code and trained model can be found at CGNet. + +# 1. Introduction + +Video Crowd Counting (VCC) has garnered much interest due to its broad range of practical applications, particularly in crowd safety management. This task requires a model + +![](images/ec0c13051ce0b6be7d462d48dfc100d9760e3f42a1a825e80a15f20e219a8714.jpg) +(a) Annotation of VIC method + +![](images/e46e3d19e326bba2e088530a1881144304acb8c6648f7b78c3016a1e913fdfd4.jpg) +(b) Annotation of our method +Figure 1. Existing VIC method requires a unique label indicating the position of each human in each frame, and these labels are consistent across frames. Weakly supervised VIC only requires labels indicating each human position and whether they are an inflow/outflow pedestrian (purple/yellow in (b)) in the current frame. The transition from individual-level labels to identity-agnostic group-level labels represents a significant reduction in the labeling effort. + +to count the number of people in each frame of a video. A limitation of VCC [12, 51, 53] is that it gives an imprecise estimate of the number of unique individuals appearing in a video sequence, as people are counted multiple times if they appear in several frames. To overcome this drawback, Video Individual Counting (VIC) was introduced, wherein a method must count the total number of people with unique identities appearing in a video sequence. + +The apparent approach to crowd counting is to count the people who appear in the first frame and add the number of people who come into the camera's field of view in later frames (the inflow). Following this principle, Han et al. [14] devised DRNet, the only existing published method for VIC, identifying repeated observations of individuals across consecutive frames based on their appearance, simultaneously predicting inflow and outflow. Although the number of outflows does not contribute to the final count, DRNet finds it helps determine individual associations. Han et al. [14] also explored using conventional multi-object track + +ing (MOT) methods [7, 41, 44, 54, 56] to tackle the VIC problem. Unfortunately, it turns out MOT methods suffer from poor accuracy. More notably, DRNet and MOT-based methods require trajectory labels (or similar individual association labels) to supervise identity association, which is highly annotation-expensive. + +Our key insight is that counting people in the previous and current frames does not require accurate identity associations for those appearing in both frames. For example, as long as we can predict individual $15\sim 17$ as entering in Fig. 1a, even if we arbitrarily associate individuals $0\sim 14$ in frame $i$ to individuals $0\sim 14$ in frame $i$ , we can still correctly infer the crowd counts. We can still count crowds if we relax individual-level identity associations to group-level associations and do not require trajectory annotations. Thus, we just need to annotate inflow and outflow people (then we can derive the individual exists in two frames), which reduces annotation costs compared to creating individual pairwise target associations for each observed pedestrian between neighboring frames. We name crowd counting with such annotations as Weakly supervised Video Individual Counting (WVIC). + +To address the WVIC task, we propose a benchmark method based on Contrastive learning with Group-level matching, namely CGNet. We design a soft contrastive loss to drive the network to learn discriminative representations that can facilitate identifying the required group associations and, thus, the inflow. Moreover, to better represent each individual, we design a memory-based individual count predictor, where historical templates of individuals are stored and updated in memory to enhance the robustness of association during inference. Considering that the two existing datasets for VIC are all captured by static cameras, we collect a weakly supervised VIC dataset based on moving UAVs, named UAVVIC. This dataset provides about 400,000 inflow/outflow and bounding box labels for four categories: pedestrian, car, bus, and van. Although devised for crowd counting, our proposed baseline method also performs well in counting other objects. + +Our main contributions are as follows: + +- We propose WVIC, a weakly supervised video individual counting task. This task does not require expensive per-target trajectory annotations and only requires two types of identity-agnostic annotations. +- We automatically reannounce two existing datasets, CroHD and SenseCrowd, and collect a new dataset, UAVVIC, to pave the way for future studies. +- We propose a strong baseline, CGNet, equipped with a newly designed group level matching soft contrastive loss, performing favorably against the supervised methods on the three datasets mentioned above. + +# 2. Related Works + +Video Crowd Counting (VCC) estimates the number of people in each video frame. Most of the existing VCC methods [2, 5, 6, 11, 12, 30, 51, 53, 58] can be divided into two categories according to their solved problems: region of interest (ROI) and line of interest (LOI). ROI methods [11, 12, 51, 53] detect pedestrians within a specific region (or the whole image), and they focus on leveraging temporal context information to improve the prediction of the current frame. LSTN [12] models the group flow of crowds in local regions. Xiong et al. in [53] fuse history frame features using a ConvLSTM. TAN [51] explores context information from adjacent density maps. CLRNet proposes a local self-attention module to help the model focus on highly related regions in adjacent frames. ROI methods can not be used in the VIC task. LOI methods [2, 5, 6, 30, 57, 58] counts people passing through a specific line. Most of the existing LOI methods [2, 5, 6] apply blobs to crowds, counting objects when the whole blobs cross the line. Ma et al. in [30] sample fixed line width areas within temporal image slices and accumulated crowds in this area. Zheng et al. [58] instead sum the number of people within neighboring blocks near a line based on local velocity. Zhao et al. in [57] train a crowd velocity map predictor with the help of trajectory labels. In practice, LOI methods perform poorly on the VIC task as people can enter or leave the field of view in any direction, making it difficult to find a suitable virtual line. + +Multiple Object Tracking (MOT) aims to predict the trajectories for multiple targets in a video. Object association is a main challenge in MOT, establishing correspondences between objects across frames to generate targets' trajectories. Existing object association methods in MOT include Probabilistic Data Association (PDA) [36, 37], Joint Probabilistic Data Association (JPDA) [16, 23, 29, 40], and graph matching methods like the Hungarian algorithm [3, 50]. Recent research has started to use more powerful feature representations to enhance the robustness of data association. Deep learning-based methods like LSTM can model temporal information to handle long-term occlusion [20, 34, 45]. Many works [4, 15, 46] also focus on the association of tracklets. Specifically, they associate multiple tracklets over longer periods using a graph model to combine the final trajectory, which can relieve the effects of lost tracked targets. These methods perform poorly on the VIC task, however, due to the inevitable ID switching problem [14]. Additionally, they require trajectory labels to train their models. + +Video Individual Counting (VIC). To the best of our knowledge, DRNet [14] is the only method for the VIC task. It predicts the number of initial pedestrians and matches pedestrians in adjacent frames by adapting a differentiable optimal transport loss. Although the VIC task only requires the total count in the video, DRNet needs individual + +matching between frames. WVIC avoids the requirement for such matching annotation, which enables easy expansion of dataset scales and thus facilitates future large deep learning models. + +# 3. A Benchmark Method + +The goal of Weakly-supervised Video Individual Counting (WVIC) is to predict the number of unique individuals $(\hat{N})$ in a video provided a series of its frames with $\delta$ as the sampling rate. Given a video $\mathbf{V}$ consisting of $T$ frames $\mathbf{V}^1, \dots, \mathbf{V}^T$ , the label information comprises coordinate/inflow/outflow data for each individual. Specifically, letting $\mathbf{P}^i = [p_1^i, \dots, p_{n_i}^i]$ denote the annotation of the coordinates of the center of the people in frame $\mathbf{V}^i$ , we sample the frame pairs $\{(\mathbf{V}^1, \mathbf{V}^{\delta + 1}), (\mathbf{V}^{\delta + 1}, \mathbf{V}^{2\delta + 1}), \dots, (\mathbf{V}^{T - \delta}, \mathbf{V}^T)\}$ , where $\delta$ is an integer frame offset. The weakly supervised inflow label $\mathbf{I}^i = [\mathbf{I}_1^i, \dots, \mathbf{I}_{n_i}^i]$ and outflow label $\mathbf{O}^i = [\mathbf{O}_1^i, \dots, \mathbf{O}_{n_i}^i]$ . If the $u$ -th person in frame $i$ doesn't appear in $\mathbf{V}^{i - \delta}$ , then $\mathbf{I}_u^i = 1$ ; otherwise, $\mathbf{I}_u^i = 0$ . For $\mathbf{O}^i$ , if the $u$ -th person in frame $V^i$ doesn't appear in $\mathbf{V}^{i + \delta}$ , then $\mathbf{O}_u^i = 1$ ; otherwise $\mathbf{O}_u^i = 0$ . If, by a slight abuse of notation, we refer to people by their coordinates in particular images, we see that the set $\{p_u^i | \mathbf{O}_u^i = 0\}$ and $\{p_u^{i + \delta} | \mathbf{I}_u^i = 0\}$ represent the people that appear both in $\mathbf{V}^i$ and $\mathbf{V}^{i + \delta}$ . + +# 3.1. Weakly Supervised Representation Learning + +As shown in Fig. 2, the inference pipeline of our CGNet consists of three components: an image-level locator to generate coordinates for pedestrians, an encoder to produce representations based on pedestrians' coordinates, and a Memory-based individual Count Predictor (MCP) to predict the final inflow count based on the representations. The locator is trained independently with the coordinate annotations, and existing image crowd localization networks such as FIDT [25] can be used. The encoder is trained with our Weakly Supervised Representation Learning method (WSRL) to extract the discriminative features of each individual. It can be any feature extractor, such as ConvNext [26]. MCP does not need training. + +Provided two $\delta$ -adjacent frames $\mathbf{V}^i$ and $\mathbf{V}^j$ and the corresponding annotations $\mathbf{P}^i$ , $\mathbf{P}^j$ , $\mathbf{O}^i$ and $\mathbf{I}^j$ , we group the individuals into four sets: $\mathbf{X} = \{p_u^i | \mathbf{O}_u^i = 0\}$ , $\mathbf{Y} = \{p_u^j | \mathbf{I}_u^j = 0\}$ , denoting shared individuals in the previous frame and the current frame, respectively; $\hat{\mathbf{X}} = \mathbf{P}^i - \mathbf{X}$ and $\hat{\mathbf{Y}} = \mathbf{P}^j - \mathbf{Y}$ , denoting outflow individuals and inflow individuals, respectively. An encoder learns representations $\mathbf{F}$ for each pedestrian based on the crops around those coordinates. Specifically, to generate the $u$ -th feature $f_u^i$ in $\mathbf{F}$ , if the locator outputs points, we extract a rectangle patch of size $96 \times 64$ centered at the predicted coordinate $p_u^i$ , and if the outputs are bounding boxes, we directly crop the bounding box from the original picture. Each cropped patch is + +resized to $224 \times 224$ and fed into the encoder, a ConvNext-S [26], generating a feature of $7 \times 7 \times 768$ , which is flattened and normalized into a 1D vector $f_{u}^{i}$ to represent the final feature for the individual: + +$$ +\begin{array}{l} \mathbf {F} _ {\mathbf {X}} ^ {i} = \operatorname {E n c o d e r} (\mathbf {X}), \mathbf {F} _ {\hat {\mathbf {X}}} ^ {i} = \operatorname {E n c o d e r} (\hat {\mathbf {X}}) \\ \bar {\mathbf {F}} ^ {i} = \bar {\mathbf {F}} _ {\hat {\mathbf {X}}} + \left(\bar {\mathbf {H}}\right) \bar {\mathbf {F}} ^ {i} = \bar {\mathbf {F}} _ {\hat {\mathbf {H}}} + \left(\hat {\mathbf {H}}\right) \end{array} \tag {1} +$$ + +$$ +\mathbf {F} _ {\mathbf {Y}} ^ {j} = \operatorname {E n c o d e r} (\mathbf {Y}), \mathbf {F} _ {\hat {\mathbf {Y}}} ^ {j} = \operatorname {E n c o d e r} (\hat {\mathbf {Y}}) +$$ + +Then, a similarity matrix $\mathbf{S} \in \mathbb{R}^{n_i \times n_j}$ between $\mathbf{F}^i = \left[\mathbf{F}_{\mathbf{X}}^i, \mathbf{F}_{\hat{\mathbf{X}}}^i\right]$ and $\mathbf{F}^j = \left[\mathbf{F}_{\mathbf{Y}}^j, \mathbf{F}_{\hat{\mathbf{Y}}}^j\right]$ is computed by: + +$$ +\mathbf {S} = \left[ \begin{array}{l l} \left(\mathbf {F} _ {\mathbf {X}} ^ {i}\right) ^ {\top} \mathbf {F} _ {\mathbf {Y}} ^ {j} & \left(\mathbf {F} _ {\mathbf {X}} ^ {i}\right) ^ {\top} \mathbf {F} _ {\hat {\mathbf {Y}}} ^ {j} \\ \left(\mathbf {F} _ {\hat {\mathbf {X}}} ^ {i}\right) ^ {\top} \mathbf {F} _ {\mathbf {Y}} ^ {j} & \left(\mathbf {F} _ {\hat {\mathbf {X}}} ^ {i}\right) ^ {\top} \mathbf {F} _ {\hat {\mathbf {Y}}} ^ {j} \end{array} \right] = \left[ \begin{array}{l l} \mathbf {S} _ {0} & \mathbf {S} _ {1} \\ \mathbf {S} _ {2} & \mathbf {S} _ {3} \end{array} \right]. \tag {2} +$$ + +The matrix $\mathbf{S}$ is divided into four parts: $\mathbf{S}_0\in \mathbb{R}^{m_{i,j}\times m_{i,j}}$ $\mathbf{S}_1\in \mathbb{R}^{m_{i,j}\times (n_j - m_{i,j})}$ $\mathbf{S}_2\in \mathbb{R}^{(n_i - m_{i,j})\times m_{i,j}}$ $\mathbf{S}_3\in$ $\mathbb{R}^{(n_i - m_{i,j})\times (n_j - m_{i,j})}$ , where $m_{i,j} = (\mathbf{1} - \mathbf{I}^{\mathbf{j}})^T(\mathbf{1} - \mathbf{O}^i)$ denotes the number of shared individuals in $\mathbf{V}^i$ and $\mathbf{V}^j$ . $\mathbf{S}_0$ is the similarity matrix of $\mathbf{X}$ and $\mathbf{Y}$ , while $\mathbf{S}_1$ , $\mathbf{S}_2$ and $\mathbf{S}_3$ are the similarity matrix of $\mathbf{X}$ and $\hat{\mathbf{Y}}$ , of $\hat{\mathbf{X}}$ and $\mathbf{Y}$ , of $\hat{\mathbf{X}}$ and $\hat{\mathbf{Y}}$ , respectively. + +To pull the matched groups (X and Y) closer and push away individual pairs from unmatched groups, we propose a weakly supervised Group-level Matching Loss (GML) to constrain different parts of S. + +Constraint for $\mathbf{S}_0, \mathbf{S}_1, \mathbf{S}_2$ . $\mathbf{S}_0$ denotes the similarity among individuals shared between two frames, and it does not matter whether the individuals are correctly matched as long as we can assign them a one-to-one match. As for $\mathbf{S}_1$ and $\mathbf{S}_2$ , they are the similarities of different identities and thus should be zero. + +To this end, we introduce a latent variable matrix $\Omega \in \mathbb{R}^{m_{i,j}\times m_{i,j}}$ , where the $(u,v)$ -th element $\Omega_{u,v}$ denote the probability that the $u$ -th pedestrian in $\mathbf{X}$ is matched with $v$ -th pedestrian in $\mathbf{Y}$ . To increase the similarity between the pairs that have a higher matching probability and decrease the similarity between the pairs that have a lower matching probability, inspired by [32, 39], we define the soft contrastive loss as: + +$$ +\mathcal {L} _ {s c o n} (i) = \min _ {\Omega} - \sum_ {u = 1} ^ {m _ {i, j}} \sum_ {v = 1} ^ {m _ {i, j}} \Omega_ {u, v} C _ {u, v}, \tag {3} +$$ + +$$ +s. t. \quad \mathbf {1} _ {m _ {i, j}} ^ {T} \Omega = \mathbf {1}, \quad \Omega \mathbf {1} _ {m _ {i, j}} = \mathbf {1}, 0 \leq \Omega_ {u, v} \leq 1, +$$ + +where $C_{u,v}$ denotes the contrastive similarity [10], and is calculated as, + +$$ +C _ {u, v} = \frac {e ^ {\frac {1}{\gamma} \mathbf {S} _ {u , v}}}{e ^ {\frac {1}{\gamma} \mathbf {S} _ {u , v}} + \sum_ {1 \leq u ^ {\prime} \leq n _ {i}} ^ {u ^ {\prime} \neq u} e ^ {\frac {1}{\gamma} \mathbf {S} _ {u ^ {\prime} , v}} + \sum_ {1 \leq v ^ {\prime} \leq n _ {j}} ^ {v ^ {\prime} \neq v} e ^ {\frac {1}{\gamma} \mathbf {S} _ {u , v ^ {\prime}}}}, \tag {4} +$$ + +where $\gamma$ is a temperature hyper-parameter [52] and $\mathbf{S}_{u,v} = \langle f_u^i,f_v^j\rangle$ is the $(u,v)$ -th element of $\mathbf{S}$ , calculated by the $u$ -th feature in $\mathbf{F}^i$ and the $v$ -th feature in $\mathbf{F}^j$ . As $\Omega_{u,v}$ describes the matching probability, our soft contrastive loss + +![](images/899ece5f0975dd10996ad6b68d31afe5e4cd176984664a9df7e592366177ed50.jpg) +Figure 2. The inference pipeline of our CGNet and the weakly supervised representation learning method (WSRL). The pipeline comprises a frame-level crowd locator, an encoder, and a Memory-based individual count predictor (MCP). The locator predicts the coordinates for pedestrians. The encoder generates representations for each individual, and MCP predicts inflow counts and updates the individual templates stored in the memory. To pull the matched groups $(\mathbf{X}$ and $\mathbf{Y})$ closer and push away individual pairs from unmatched groups, WSRL exploits inflow and outflow labels to optimize the encoder with a novel Group level Matching Loss (GML), which consists of a soft contrastive loss $(\mathcal{L}_{scon})$ and a hinge loss $(\mathcal{L}_{hinge})$ . + +evaluates the expectation of contrastive loss. It should be noting that in $C_{u,v}$ , $(f_u^i, f_v^j)$ is considered as positive pairs, while $\{(f_u^i, f_{v'}^j) | v' \neq v\}$ and $\{(f_{u'}, f_v^j) | u' \neq u\}$ are considered as negative pairs. Thus, $\mathcal{L}_{scon}$ can encourage $S_0$ to be a permutation matrix, which indicates that there exists a one-to-one match between $X$ and $\hat{Y}$ , and drive $S_1, S_2$ to be zero matrices. + +The problem defined by Eq. (3) is a typical balanced Optimal Transport (OT [35]) problem which transports $\mathbf{1}_{m_{i,j}}$ to $\mathbf{1}_{m_{i,j}}$ using $\mathbf{1} - C$ as the cost matrix and $\Omega$ is the transport matrix. Therefore, we consider $\mathcal{L}_{scon}(i)$ as an OT loss, which can be solved with the Sinkhorn algorithm [38]. + +Constraint for $\mathbf{S}_3$ . Obviously, the elements in $\mathbf{S}_3$ measure the similarity of different individuals, so they should be as small as possible. However, directly constraining $\mathbf{S}_3$ to zero matrix will make features from $\hat{\mathbf{X}}, \hat{\mathbf{Y}}$ easily collapse to zero-vector. To this end, the widely used Hinge L1 loss [18] is adopted to constrain $\mathbf{S}_3$ : + +$$ +\mathcal {L} _ {h i n g e} (i) = \frac {\sum R e l u (\mathbf {S _ {3}} - \theta)}{(n _ {i} - m _ {i , j}) (n _ {j} - m _ {i , j})}, \qquad (5) +$$ + +where $\theta$ is a threshold used to ignore small values in $\mathbf{S}_3$ . + +Finally, the group-level matching loss of the training set $\mathbf{V}$ is formulated as, + +$$ +\mathbf {G M L} (\mathbf {V}) = \sum_ {0 < i < T, i = k \delta + 1} \left(\mathcal {L} _ {s c o n} (i) + \mathcal {L} _ {h i n g e} (i)\right). \tag {6} +$$ + +# 3.2. Memory-based Count Predictor + +To handle individuals' appearance variance and re-entering cases, inspired by [22, 33, 55], we propose a memory-based count predictor (MCP) to reason the final count for the video, which stores templates of recently appeared individuals in a flexible memory. Specifically, when processing frame $\mathbf{V}^{i + \delta}$ , the memory is denoted as $\mathcal{G}^i = \{\mathcal{G}_1^i,\mathcal{G}_2^i,\dots ,\mathcal{G}_{\mathbf{N}^i}^i\}$ , where $\mathbf{N}^i$ is the memory size at time $i$ . $\mathcal{G}_u^i = \{g_{u,0}^i,\dots,g_{u,k}^i\}$ is a set of templates for $u$ -th pedestrian in the memory, and $k$ is the number of templates stored for $u$ -th pedestrian. + +We first use the crowd locator to predict the coordinates of pedestrians in $\mathbf{V}^{i + \delta}$ , and then extract their corresponding features, denoted as $\{f_1^{i + \delta}, f_2^{i + \delta}, \dots, f_{n_{i + \delta}}^{i + \delta}\}$ . To associate the individuals in $\mathbf{V}^{i + \delta}$ with templates in memory, we first define the cost $\hat{C}_{u,u'}$ of matching the $u$ -th pedestrian with the $u'$ -th template in $\mathcal{G}^i$ as + +$$ +\hat {C} _ {u, u ^ {\prime}} = \max _ {v} \left(1 - \left\langle f _ {u} ^ {i + \delta}, g _ {u ^ {\prime}, v} ^ {i} \right\rangle\right). \tag {7} +$$ + +Then, we generate an optimal one-to-one match $\pi \in \{0,1\}^{n_{i + \delta}\times \mathbf{N}^i}$ (assuming $\mathbf{N}^i\geq n_i$ ) by solving the following problem with Hungarian Algorithm [21]: + +$$ +\pi = \arg \min _ {\pi} \sum_ {u, u ^ {\prime}} \pi_ {u, u ^ {\prime}} \hat {C} _ {u, u ^ {\prime}} \tag {8} +$$ + +$$ +s. t. \quad \forall u ^ {\prime}, \sum_ {u} \pi_ {u, u ^ {\prime}} = 1, \quad \forall u, \sum_ {u} \pi_ {u, u ^ {\prime}} \leq 1. +$$ + +If the matching cost for $u$ meets $\sum_{u' = 1}^{\mathbf{N}^i}\pi_{u,u'}\hat{C}_{u,u'} > \zeta$ + +$u$ will be judged as an inflow pedestrian. Otherwise, if $\pi_{u,u'} = 1$ , $u$ is associated with $g_{u'}^i$ . + +Template update. Similar to [22], $\mathcal{G}_u^i$ has a time-to-live factor (ttl). $ttl_u$ will decrease one if $\mathcal{G}_u^i$ has no mapping to any pedestrian feature in $\mathbf{F}^{i + \delta}$ . If $ttl_u$ decreases to zero, the template $\mathcal{G}_u^i$ will be dropped. Meanwhile, whenever $\mathcal{G}_u^i$ is associated with a pedestrian in the current frame, $ttl_u$ will be reset to $ttlmax$ . This time-to-live threshold allows us to save some recently appeared targets while discarding those that have not appeared for a long time. If pedestrian $u$ is associated with $u'$ -th template, we add $f_{u}^{i + \delta}$ to $\mathcal{G}_{u'}^i$ to update the template + +$$ +\mathcal {G} _ {u ^ {\prime}} ^ {i + \delta} \leftarrow \mathcal {G} _ {u ^ {\prime}} ^ {i} \cup \left\{f _ {u} ^ {i + \delta} \mid \pi_ {u, u ^ {\prime}} = 1, \sum_ {u ^ {\prime} = 1} ^ {\mathbf {N} ^ {i}} \pi_ {u, u ^ {\prime}} \hat {C} _ {u, u ^ {\prime}} \leq \zeta \right\}. \tag {9} +$$ + +If pedestrian $u$ is an inflow pedestrian, $\{f_u^{i + \delta}\}$ will be added to the memory $\mathcal{G}^{i + \delta}$ as a new template. + +The output number of the total individuals in video $I$ is then calculated via + +$$ +\hat {N} = n _ {1} + \sum_ {\delta + 1 \leq i \leq T, i = k \delta + 1} \hat {n} _ {i}, \tag {10} +$$ + +where $\hat{n}_i$ is the number of inflow pedestrians at time $i$ . + +# 4. Experiments + +# 4.1. Dataset + +We test our CGNet on three datasets: CroHD [42], SenseCrowd [24], and UAVVIC. CroHD has four training videos and five testing videos. SenseCrowd contains 634 videos, and we split the train, validation, and test dataset following DRNet [14]. UAVVIC is our proposed dataset collected by a moving UAV camera in various scenes, including campus, beach, car park, highway, city road, and square. UAVVIC consists of 221 videos (100 for training, 100 for testing, and 21 for validation), and 5,396 frames are sampled with 3s as the interval. Annotations consist of 398,158 bounding boxes, and group-level matching labels in neighbor frames are provided. The resolutions of UAVVIC are 4K and 1080P for better capture of tiny pedestrians from drone view. A detailed comparison of UAVVIC with existing video crowd counting datasets [8, 9, 13, 24, 42] is shown in Tab. 1. UAVVIC provided a moving shot with a larger range of pedestrians. More details about UAVVIC are provided in the supplementary materials. + +# 4.2. Metrics + +Mean Absolute Error (MAE), Mean Square Error (MSE), and Weighted Relative Absolute Errors (WRAE) are used for evaluation. The first two metrics are common metrics applied in VCC. However, unlike VCC schemes, we only count the same person once in the video. WRAE [14] is proposed to balance the performance on videos with + +
NameResolutionRangeMoving shotPointBoxTrajectoryIn-Out
Mall [9]640*48013-53X✓XXX
UCSD [8]238*15811-46X✓XXX
FDST [13]1920*10809-57X✓XXX
1280*720
SenseCrowd [24]-1-296X✓✓✓X
CroHD [42]-25-346X✓✓✓X
UAVVIC3840*21600-735✓✓✓X✓
1920*1080
+ +Table 1. Comparison of different video crowd counting datasets. + +different lengths and pedestrian numbers: $WRAE = \sum_{i=1}^{K} \frac{T_i}{\sum_{j=1}^{K} T_j} \frac{|N_i - \hat{N}_i|}{N_i}$ , where $K$ is the total number of videos, and $T_i$ is the length of video $i$ , $N_i / \hat{N}_i$ denotes the number of ground truth/predicted counts in $i$ -th video. + +# 4.3. Implementation Details + +Training. If there are no special instructions, the loca-tor we applied is a well-known crowd localization model, FIDT [25]. All parameters are official from FIDT, except that we replaced the backbone from HRNet-W48 [47] to HRNet-W18 [47] for efficient inference. The learning rate is set as $1e^{-4}$ along with AdamW [28] as the optimizer and applying the pre-train weight from Timm [49]. All related models can be trained on one RTX3090 (24G memory). + +Testing. In the testing phase, the time interval $\delta$ is set as $3s$ following the setting in [14], the threshold $\zeta$ is set as $0.7$ , $\gamma$ in Eq. (4) is set as 10, and the max time-to-live (ttlmax) is set as 3, the max size of memory (memmax) is set as 5. + +# 4.4. Overall Performance + +Comparison Methods. To the best of our knowledge, only DRNet[14] has been specifically designed for the VIC task. To evaluate the effectiveness of our proposed CGNet, we also tested two categories of approaches that could be applied to VIC: 1) Multi-object tracking (MOT) methods, including HeadHunter-T [43], FairMOT [56], PHDTT [44], SMILE [48], SparseTrack [27], Deep-OC-SORT [31], and BoT-SORT [1], where we consider the total number of indices in the entire video as individual count. We set the frame rate to 1 FPS for HeadHunter-T and 0.33 FPS for the other MOT methods to obtain better VIC performance. 2) A recent cross-line video crowd-counting, LOI [57], which also requires trajectory labels in training. + +Results on SenseCrowd: As shown in Tab. 2, it is evident that our CGNet exhibits superior performance on this larger dataset, SenseCrowd. Particularly, CGNet improved the overall MAE by about $28\%$ compared to DRNet. Furthermore, we achieve the lowest MAE across all density levels except for D0. Compared to the existing lowest MAE on D1, D2, D3, and D4, the performance improvements of our CGNet are about $27\%$ , $63\%$ , $37\%$ , and $17\%$ , respectively. In Fig. 3, we show qualitative results on three representative scenes,i.e., mall, station, and outdoor plaza. Although the inflows in the current flow are not concentrated, CGNet + +![](images/e5b7e7e501e5f533ddca8baee52582ff7025253a685864959c262df4ba75b8ea.jpg) + +![](images/6c99610530f15a593083609005a31cb0ea38e2016e75a6676865517b0b668667.jpg) +(a) Previous frame (0s) + +![](images/fbb32e315b8ecd072b065c520434f9980f6ac9f1b666bd2a7058b683018c8010.jpg) + +![](images/7a94857b819b4959248b6767268bdd292414e3fac4548a45ebc63fc1f3324fdb.jpg) +GT inflow: 4 +Predict inflow: 4 + +![](images/f9d22f4482517acb43e4a085fe28e1f5180b6371669a9502d1f61c71cb616879.jpg) +GT inflow: 11 +(b) Current frame (3s) +Figure 3. Results on the SenseCrowd dataset. Red dots are predictions of the locator. Blue/green/yellow circles are correct predicted inflow, error predicted inflow, and missed inflow, respectively. + +![](images/e3e6f205947c67ee45c2337fe81a5d29cca46f4b22e4a61144c4f53badd9593a.jpg) +GT inflow: 15 +Predict inflow: 14 + +
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
FairMOT [56]35.462.348.913.522.467.984.4145.8
HeadHunter-T [43]30.050.638.611.825.756.092.6131.4
SMILE [48]27.2236.732.59.221.033.876.5203.5
SparseTrack [27]28.3643.830.89.921.542.189.2189.5
Deep-OC-SORT [31]26.0443.629.59.525.129.160.3156.5
BoT-SORT [1]24.6732.428.38.720.531.170.2165.1
LOI [57]24.733.137.412.525.439.339.686.7
DRNet [14]12.324.712.74.18.023.350.077.0
CGNet (Ours)8.8617.6912.65.05.88.525.063.4
+ +Table 2. Performance comparison on SenseCrowd. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and $[200, +\infty)$ , respectively. The best values are highlighted in blue font. + +
MethodMAEMSEWRAE(%)MAE on five testing scenes
CroHD11CroHD12CroHD13CroHD14CroHD15
PHDTT [44]2130.42808.3401.6247379347944911327
FairMOT [56]256.2300.844.1114272844081000
HeadHunter-T [43]253.2351.732.7651015155821003
SMILE [48]257.6334.540.516231156649246
SparseTrack [27]176.6208.627.62917225833688
Deep-OC-SORT [31]165.2195.933.16835118616160
BoT-SORT [1]154.8176.4227.44921013828691
LOI [57]305.0371.146.060243458630131
DRNet [14]141.1192.327.4313381825561
CGNet (Ours)75.095.114.577214144138
+ +Table 3. Performance comparison on CroHD dataset. CroHD11-CroHD15 are five test scenes labeled in [14]. The best values are highlighted in blue font. + +can accurately predict them. More visualization results can be found in supplementary material. + +Results on CroHD. Comparison with existing methods on CroHD is shown in Tab. 3. Our CGNet achieved the lowest MAE, MSE, and WRAE. Especially, except for CroHD15, our method obtained the lowest MAE across the other four + +testing scenes. As CGNet is trained with only point annotations, it cannot crop the patch of optimal size to extract features on CroHD15. It should be noted that on CroHD13 and CroHD14, which have the highest average density (245.9 and 259.6 person/frame), the performances of all the compared methods except for DRNet degrade sig- + +
MethodMAE/MSE/WRAEMAE/MSE/WRAE on four different categories
PedestrianCarBusVan
BoT-SORT [1]49.9/228.2/29.992.4/398.2/49.059.1/259.1/39.223.0/109.6/57.125.1/146.2/77.3
Deep-OC-SORT [31]37.1/115.4/35.285.2/213.5/22.133.4/113.1/21.019.1/ 78.2/26.210.7/56.6/32.8
DRNet [14]18.4/ 41.0/15.234.5/ 92.1/19.119.0/ 32.7/17.810.4/ 22.1/32.29.9/17.2/29.4
CGNet (Ours)12.9/ 37.4/ 12.031.2/ 98.4/14.416.9/ 45.2/ 7.32.1/ 4.1/19.21.3/ 2.0/25.1
+ +Table 4. Performance comparison on UAVVIC. All the methods are trained independently with four classes. The values of each entry are MAE/MSE/WRAE. The best values are highlighted in blue font. + +
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
①13.3532.322.89.712.315.326.543.4
②9.8420.013.85.36.312.028.867.7
③10.7117.4016.96.98.911.724.350.6
④8.8617.6912.65.05.88.525.063.4
+ +Table 5. Ablation study on the main components. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and $[200, +\infty)$ , respectively. $①$ , No constrain on $\mathbf{S}_1$ , $\mathbf{S}_2$ ; $②$ , w/o Eq. (5); $③$ , w/o MCP; $④$ , full CGNet. The best values are highlighted in blue font. + +nificantly. In contrast, the performance of our CGNet is relatively stable in these two scenes. Furthermore, even without individual-level matching labels, our CGNet performs favorably against the state-of-the-art approach, DRNet, by a large margin (about $46\%$ on the MAE), demonstrating the effectiveness of our method. + +![](images/eba599d1e565fbc096ce4f4681298ae1b78e4dbba74f2bdcdc85ceba407b63cc.jpg) +(a) Previous frame (0s) +Figure 4. Qualitative visualization of our CGNet on UAVVIC. + +![](images/6706da357701209329e8b5b7ad86308250a97e53c3680723cfa6df129ee09c33.jpg) +(b) Current frame (3s) + +Results on UAVVIC. The results on our proposed UAVVIC dataset are shown in Tab. 4. All methods are trained and tested on each category, respectively. Since there is no trajectory annotation on UAVVIC, we train the compared methods [1, 14, 31] on VisDrone [59], a UAV dataset for MOT. Our CGNet shows favorable performance in all categories. The average MAE is improved by about $30\%$ compared to DRNet. In Fig. 4, we show the quality result of our CGNet on the UAVVIC dataset. Although the UAV moves, CGNet can still detect inflow objects accurately. + +# 4.5. Ablation Studies + +We conducted several ablation studies on the SenseCrowd. 1). We verify the effectiveness of the main components of our CGNet. 2). We analyze the effect of two main parameters. 3). We report the performance of our CGNet with different locators. 4). We use the trajectories generated by + +![](images/23d33fb0d1353c0e5af4cb555f60bb704ace775dad287b5aeef8cc6cdbd127d9.jpg) +(a) + +![](images/0bd1d33c6e964bea8c7791d00f4c4f20b8120e9b0dfc146bc0b9d2cac17ad65e.jpg) +(b) +Figure 5. Effectiveness of MCP on SenseCrowd. (a) The average MAE with different lengths of videos with and w/o MCP. (b) Performance with different max time-to-live factor tllmax of MCP. + +our CGNet as pseudo-labels for existing VIC methods. 5). We compare the time cost of the WVIC and VIC labels. + +Effectiveness of Constrain for $\mathbf{S}_1, \mathbf{S}_2$ . Eq. (4) constrains $\mathbf{S}_0, \mathbf{S}_1, \mathbf{S}_2$ together. Without considering un-matched pairs in $\mathbf{X} \times \hat{\mathbf{Y}}$ and $\hat{\mathbf{X}} \times \mathbf{Y}$ , it can be changed to constrain only $\mathbf{S}_0$ by remove elements in $\mathbf{S}_1, \mathbf{S}_2$ . As shown in Tab. 5, MAE dropped by about 4.51 without constraints on $\mathbf{S}_1$ and $\mathbf{S}_2$ (① VS ④). This is because elements in $\mathbf{S}_1$ and $\mathbf{S}_2$ are related to pedestrians that exist only in one frame. Considering $\mathbf{X} \times \hat{\mathbf{Y}}$ and $\hat{\mathbf{X}} \times \mathbf{Y}$ as negative pairs, our soft contrastive loss can push apart individuals in $\hat{\mathbf{X}} (\hat{\mathbf{Y}})$ with that in $\mathbf{Y} (\mathbf{X})$ , thus enhancing the distinguishing ability of the representations. + +Effectiveness of Eq. (5). We replace the hinge MAE loss in Eq. (5) to normal L1 loss and derive the method ② in Tab. 5. Compared to L1 loss, our $\mathcal{L}_{hinge}(\cdot)$ gives a soft margin to the inner product of two features other than encouraging them to be 0. Optimizing the inner product of all features toward 0 makes it easier to cause either side to collapse a 0 vector, which is not conducive to learning different representations. Compared the performance of ② with that of ④ in Tab. 5, $\mathcal{L}_{hinge}(\cdot)$ improves MAE in all density levels. + +Effectiveness of MCP. As shown in Tab. 5, compared $③$ with $④$ , MCP brings about 4.1 improvements on WRAE. Specifically, we evaluate the average MAE of our CGNet on videos with different lengths. As shown in Fig. 5(a), the performance gap between $③$ and $④$ becomes larger as + +
δMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
510.9620.2619.56.56.910.429.279.9
48.7114.3312.94.24.910.227.870.1
38.8617.6912.65.05.88.525.063.4
29.7720.0117.37.18.812.417.452.6
19.5615.3416.66.87.410.523.145.5
+ +Table 6. Ablation study on interval $\delta$ . The best values are highlighted in blue font. + +
MethodLocatorEncoderMAEMSEWRAE
①FIDTConvNext-S8.8617.699.27
②Yolov8 nanoConvNext-S8.0116.428.85
③Yolov8 smallConvNext-S7.5615.198.05
④VGG16+FPNConvNext-S9.2017.459.55
⑤VGG16+FPNPrRoiIPooling10.0118.229.70
⑥VGG16+FPNPrRoiIPooling12.5923.3212.58
+ +the length of the video increases. This is because, with the stored templates of individuals, MCP can handle appearance variations and individuals' re-entry, alleviating error accumulation during inflow reasoning. + +Effect of ttlmax. The performance with different max long-time-live factor ttlmax is shown in Fig. 5(b). Compared with no memory (i.e., ttlmax = 0), adding memory is helpful for all density levels except for extremely high-density levels (D4). It may be because high-density crowds result in too many candidate individual-level matches, making it difficult for our CGNet to choose the right one. Meanwhile, the performance degrades when ttlmax increases from 3 to 6 for that larger ttlmax = 6 leads to the templates of individuals that appeared a long time ago being stored in the memory. As the probability of these people re-entering the field is low, the recorded templates become distractions. + +Effect of Interval $\delta$ . The performance with different intervals $\delta$ is shown in Tab. 6. The performance on higher-density video is better when $\delta$ is smaller. This is because crowds flow faster in those videos, and long intervals may lead to missing pedestrians in sampled frames. The overall best performance is achieved when $\delta$ is set as $4s$ . + +Effect of the Locator. We replaced FIDT in our CGNet with Yolo V8 [19] nano and small and the localization branch (VGG16+FPN) used in DRNet, respectively. Different from FIDT, these three locators are trained with box annotations. For Yolo V8, the predicted bounding boxes are fed to our feature extractor to generate representations of individuals. For a fair comparison with DRNet, we also replace our feature extractor ConvNext-S with PrRoIPooling [17] used in DRNet [14] and remove MCP. As shown in Tab. 7, changing locators can further boost the performance of our method. With the same locator and feature extractor as DRNet (⑤), the MAE of our method is 10.01, 2.99 lower + +Table 7. Performance on SenseCrowd of our scheme with different locators and feature extractors. In $⑥$ , MCP is replaced with the inflow reasoning method in DRNet. The best values are highlighted in blue font. + +
LocatorGround Truth LabelsPseudo Labels
MAEMSEWRAEMAEMSEWRAE
FairMOT35.462.348.937.264.849.8
HeadHunter-T30.050.638.632.554.239.9
SMILE27.236.732.527.537.035.2
SparseTrack28.443.830.830.545.031.5
Deep-OC-SORT26.043.629.529.545.331.7
BoT-SORT24.732.428.325.035.933.8
LOI24.733.137.426.935.239.9
DRNet12.324.712.714.525.513.3
+ +Table 8. Performance of existing VIC methods trained with different trajectory labels. + +
(ni, ni+δ)(9,8)(60,53)(116,115)(289,292)(645,724)
Association [14]23.8 s104.5 s174.4 s586.9 s1663.1 s
In-Out4.4 s42.6 s53.7 s95.0 s571.9 s
+ +Table 9. Time cost for a pair of frames with two types of labeling methods. $n_i$ denotes the number of pedestrians in frame $i$ . + +than that of DRNet. Even without MCP (⑥), the MAE of our method is only slightly larger than that of DRNet (12.59 v.s. 12.3), demonstrating that our method can learn effective representations without trajectory labels. + +Performs as a Pseudo Trajectory Generator. The latent variable $\Omega$ is obtained by solving a problem defined by Eq. (3). The Hungarian Algorithm generates a one-to-one match using $\Omega$ as the cost matrix. Based on the matching, we generate pseudo trajectory labels, which are used to train existing VIC methods. We show the results on SenseCrowd in Tab. 8. With the generated pseudo label, the performance of all the methods degrades slightly than using the ground truth trajectory labels, demonstrating that exact individual-level matching annotations may not be necessary for VIC. + +Label time cost. We compare the time cost of annotating our weakly supervised label and association label [14] on five frame pairs with various densities in Tab. 9. The annotation time for our weakly supervised label is only about $30\%$ of that of the association label, making it easier to build a larger dataset for this VIC. + +# 5. Conclusions + +We propose a new task, Weakly supervised Video Individual Counting (WVIC), in this paper. Unlike conventional video individual counting tasks, which require expensive trajectory labels, our WVIC task just requires two types of easily available identity-agnostic annotations. This largely reduces the annotation cost and facilitates future works to upscale data size. Furthermore, we propose a strong baseline for the WVIC task. The key idea is a Group-level Matching Loss (GML), which makes it easier to distinguish inflow individuals from all objects. Last, we have extended two existing datasets and built a totally new dataset for the WVIC task. Our method achieves favorable performance on all three datasets, even compared to supervised methods. + +# References + +[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 5, 6, 7 +[2] Antonio Albiol, Inmaculada Mora, and Valery Naranjo. Real-time high density people counter using morphological tools. IEEE TITS, 2(4):204-218, 2001. 2 +[3] Marco Allodi, Alberto Broggi, Domenico Giaquinto, Marco Patander, and Antonio Prioletti. Machine learning in tracking associations with stereo vision and lidar observations for an autonomous vehicle. In IVS, pages 648-653, 2016. 2 +[4] Seung Hwan Bae and Kuk-Jin Yoon. Confidence-based data association and discriminative deep appearance learning for robust online multi-object tracking. IEEE TPAMI, 40(3):595-610, 2018. 2 +[5] Javier Barandiarán, Berta Murguia, and Fernando Boto. Real-time people counting using multiple lines. In WIAMIS, pages 159-162, 2008. 2 +[6] Jesús Bescós, José M. Menéndez, and Narciso García. DCT based segmentation applied to a scalable zenithal people counter. In ICIP, pages 1005-1008, 2003. 2 +[7] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, pages 1-6, 2017. 2 +[8] Antoni B. Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE TPAMI, 30(5):909-926, 2008. 5 +[9] Ke Chen, Chen Change Loy, Shaogang Gong, and Tony Xiang. Feature mining for localised crowd counting. In BMVC, pages 3-14, 2012. 5 +[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3 +[11] Li Dong, Haijun Zhang, Jianghong Ma, Xiaofei Xu, Yimin Yang, and Q. M. Jonathan Wu. Clrnet: A cross locality relation network for crowd counting in videos. IEEE TNNLS, pages 1-15, 2022. 2 +[12] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. In ICME, pages 814-819, 2019. 1, 2 +[13] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. arXiv preprint arXiv:1907.07911, 2019.5 +[14] Tao Han, Lei Bai, Junyu Gao, Qi Wang, and Wanli Ouyang. DR.VIC: decomposition and reasoning for video individual counting. In CVPR, pages 3073-3082, 2022. 1, 2, 5, 6, 7, 8 +[15] Yuhang He, Jie Han, Wentao Yu, Xiaopeng Hong, Xing Wei, and Yihong Gong. City-scale multi-camera vehicle tracking by semantic attribute parsing and cross-camera tracklet matching. In CVPRW, pages 2456-2465, 2020. 2 +[16] Andinet Hunde and Beshah Ayalew. Automated multi-target tracking in public traffic in the presence of data association uncertainty. In ACC, pages 300-306, 2018. 2 +[17] Borui Jiang, Ruixuan Luo, Jiayuan Mao, Tiantian Xiao, and Yuning Jiang. Acquisition of localization confidence for accurate object detection. In ECCV, pages 784-799, 2018. 8 +[18] Thorsten Joachims. Support vector method for multivariate performance measures. In ICML, pages 377-384, 2005. 4 + +[19] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.8 +[20] Chanho Kim, Fuxin Li, Mazen Alotaibi, and James M. Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, pages 9553-9562, 2021. 2 +[21] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 4 +[22] L.J. Latecki and R. Miezianko. Object tracking with dynamic template update and occlusion detec. In ICPR, volume 1, pages 556-560, 2006. 4, 5 +[23] Eui-Hyuk Lee, Qian Zhang, and Taek Lyul Song. Markov chain realization of joint integrated probabilistic data association. Sensors, 17(12):2865, 2017. 2 +[24] Haopeng Li, Lingbo Liu, Kunlin Yang, Shinan Liu, Junyu Gao, Bin Zhao, Rui Zhang, and Jun Hou. Video crowd localization with multifocus gaussian neighborhood attention and a large-scale benchmark. IEEE TIP, 31:6032-6047, 2022. 5 +[25] Dingkang Liang, Wei Xu, Yingying Zhu, and Yu Zhou. Focal inverse distance transform maps for crowd localization. IEEE TMM, 25:6040-6052, 2023. 3, 5 +[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 3 +[27] Zelin Liu, Xinggang Wang, Cheng Wang, Wenyu Liu, and Xiang Bai. Sparsetrack: Multi-object tracking by performing scene decomposition based on pseudo-depth. arXiv preprint arXiv:2306.05238, 2023. 5, 6 +[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In *ICLR*, 2018. 5 +[29] Zhongzhen Luo, Mina Attari, Saeid R. Habibi, and Martin von Mohrenschild. Online multiple maneuvering vehicle tracking system based on multi-model smooth variable structure filter. IEEE TITS, 21(2):603-616, 2020. 2 +[30] Zheng Ma and Antoni B. Chan. Counting people crossing a line using integer programming and local features. IEEE TCSVT, 26(10):1955-1969, 2016. 2 +[31] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 5, 6, 7 +[32] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3 +[33] Zhuojin Pan and Xiujuan Wang. Correlation tracking algorithm based on adaptive template update. In ICISP, volume 1, pages 98-101, 2010. 4 +[34] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, pages 164-173, 2021. 2 +[35] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning, 11(5-6):355-607, 2019. 4 +[36] Christopher Rasmussen and Gregory D. Hager. Probabilistic data association methods for tracking complex visual objects. IEEE TPAMI, 23(6):560-576, 2001. 2 +[37] Christian Ritter, Andrea Imle, Ji Young Lee, Barbara Muller, + +Oliver T. Fackler, Ralf Bartenschlager, and Karl Rohr. Two-filter probabilistic data association for tracking of virus particles in fluorescence microscopy images. In *ISBI*, pages 957–960, 2018. 2 +[38] Richard Sinkhorn. A Relationship Between Arbitrary Positive Matrices and Doubly Stochastic Matrices. The Annals of Mathematical Statistics, 35(2):876 - 879, 1964. 4 +[39] Rakshith Sharma Srinivasa, Jaejin Cho, Chouchang Yang, Yashas Malur Saidutta, Ching-Hua Lee, Yilin Shen, and Hongxia Jin. CWCL: Cross-modal transfer with continuously weighted contrastive loss. In NeurIPS, 2023. 3 +[40] Jason Stauch, Travis Bessell, Mark Rutten, Jason Baldwin, Moriba Jah, and Keric Hill. Joint probabilistic data association and smoothing applied to multiple space object tracking. JGCD, 41(1):19-33, 2018. 2 +[41] Ramana Sundararaman, Cedric De Almeida Braga, Éric Marchand, and Julien Pettré. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 2 +[42] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 5 +[43] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, June 2021. 5, 6 +[44] Xuan-Thuy Vo, Van-Dung Hoang, Duy-Linh Nguyen, and Kang-Hyun Jo. Pedestrian head detection and tracking via global vision transformer. In IFCVW, pages 155-167, 2022. 2, 5, 6 +[45] Xingyu Wan, Jinjun Wang, Zhifeng Kong, Qing Zhao, and Shunming Deng. Multi-object tracking using online metric learning with long short-term memory. In ICIP, pages 788-792, 2018. 2 +[46] Bing Wang, Gang Wang, Kap Luk Chan, and Li Wang. Tracklet association by online target-specific metric learning and coherent dynamics estimation. IEEE TPAMI, 39(3):589-602, 2017. 2 +[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2020. 5 +[48] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking. arXiv preprint arXiv:2211.08824, 2023. 5, 6 +[49] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5 +[50] Hefeng Wu, Yafei Hu, Keze Wang, Hanhui Li, Lin Nie, and Hui Cheng. Instance-aware representation learning and association for online multi-person tracking. PR, 94:25-34, 2019. 2 +[51] Xingjiao Wu, Baohan Xu, Yingbin Zheng, Hao Ye, Jing Yang, and Liang He. Fast video crowd counting with a temporal aware network. Neurocomputing, 403:13-20, 2020. 1, 2 +[52] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 3 + +[53] Feng Xiong, Xingjian Shi, and Dit-Yan Yeung. Spatiotemporal modeling for crowd counting in videos. In ICCV, pages 5161-5169, 2017. 1, 2 +[54] Bo Yang, Chang Huang, and Ram Nevatia. Learning affinities and dependencies for multi-target tracking using a CRF model. In CVPR, pages 1233-1240, 2011. 2 +[55] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE TPAMI, 43(1):360-374, 2021. 4 +[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129:3069–3087, 2021. 2, 5, 6 +[57] Zhuoyi Zhao, Hongsheng Li, Rui Zhao, and Xiaogang Wang. Crossing-line crowd counting with two-phase deep neural networks. In ECCV, pages 712-726, 2016. 2, 5, 6 +[58] Huicheng Zheng, Zijian Lin, Jiepeng Cen, Zeyu Wu, and Yadan Zhao. Cross-line pedestrian counting based on spatially-consistent two-stage local crowd density estimation and accumulation. IEEE TCSVT, 29(3):787-799, 2018. 2 +[59] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE TPAMI, 44(11):7380-7399, 2021. 7 \ No newline at end of file diff --git a/2024/Weakly Supervised Video Individual Counting/images.zip b/2024/Weakly Supervised Video Individual Counting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ca53fde8c2ec5c57ff26bd026d9fc1ec102db09c --- /dev/null +++ b/2024/Weakly Supervised Video Individual Counting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a399b3a674e802eb745141be9f6d337cb08513014a3f970585bade4841cbc69 +size 721322 diff --git a/2024/Weakly Supervised Video Individual Counting/layout.json b/2024/Weakly Supervised Video Individual Counting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..02dfad6b58d0747f0cc78f7027b503f1d8dcb480 --- /dev/null +++ b/2024/Weakly Supervised Video Individual Counting/layout.json @@ -0,0 +1,12129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 103, + 441, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 103, + 441, + 121 + ], + "spans": [ + { + "bbox": [ + 153, + 103, + 441, + 121 + ], + "type": "text", + "content": "Weakly Supervised Video Individual Counting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 514, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 514, + 181 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 514, + 181 + ], + "type": "text", + "content": "Xinyan Liu1 Guorong Li1,3* Yuankai Qi2 Ziheng Yan1 \nZhenjun Han1 Anton van den Hengel5 Ming-Hsuan Yang6 Qingming Huang1,3,4 \n1 University of Chinese Academy of Science; 2 Macquarie University;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 185, + 480, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 185, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 113, + 185, + 480, + 213 + ], + "type": "text", + "content": "3 Key Lab of Big Data Mining and Knowledge Management, UCAS, China; 4 Key Lab of Intell. Info. Process., Inst. of Comput. Tech., CAS;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 213, + 471, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 213, + 471, + 241 + ], + "spans": [ + { + "bbox": [ + 121, + 213, + 471, + 241 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 121, + 213, + 471, + 241 + ], + "type": "text", + "content": " Australian Institute for Machine Learning, The University of Adelaide; " + }, + { + "bbox": [ + 121, + 213, + 471, + 241 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 121, + 213, + 471, + 241 + ], + "type": "text", + "content": " University of California at Merced." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 294, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 290, + 547 + ], + "type": "text", + "content": "Video Individual Counting (VIC) aims to predict the number of unique individuals in a single video. Existing methods learn representations based on trajectory labels for individuals, which are annotation-expensive. To provide a more realistic reflection of the underlying practical challenge, we introduce a weakly supervised VIC task, wherein trajectory labels are not provided. Instead, two types of labels are provided to indicate traffic entering the field of view (inflow) and leaving the field view (outflow). We also propose the first solution as a baseline that formulates the task as a weakly supervised contrastive learning problem under group-level matching. In doing so, we devise an end-to-end trainable soft contrastive loss to drive the network to distinguish inflow, outflow, and the remaining. To facilitate future study in this direction, we generate annotations from the existing VIC datasets SenseCrowd and CroHD and also build a new dataset, UAVVIC. Extensive results show that our baseline weakly supervised method outperforms supervised methods, and thus, little information is lost in the transition to the more practically relevant weakly supervised task. The code and trained model can be found at CGNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 570, + 128, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 128, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 128, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "content": "Video Crowd Counting (VCC) has garnered much interest due to its broad range of practical applications, particularly in crowd safety management. This task requires a model" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 270, + 421, + 356 + ], + "blocks": [ + { + "bbox": [ + 307, + 270, + 421, + 356 + ], + "lines": [ + { + "bbox": [ + 307, + 270, + 421, + 356 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 421, + 356 + ], + "type": "image", + "image_path": "ec0c13051ce0b6be7d462d48dfc100d9760e3f42a1a825e80a15f20e219a8714.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 361, + 412, + 371 + ], + "lines": [ + { + "bbox": [ + 312, + 361, + 412, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 412, + 371 + ], + "type": "text", + "content": "(a) Annotation of VIC method" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 433, + 270, + 547, + 356 + ], + "blocks": [ + { + "bbox": [ + 433, + 270, + 547, + 356 + ], + "lines": [ + { + "bbox": [ + 433, + 270, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 433, + 270, + 547, + 356 + ], + "type": "image", + "image_path": "e46e3d19e326bba2e088530a1881144304acb8c6648f7b78c3016a1e913fdfd4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 361, + 538, + 371 + ], + "lines": [ + { + "bbox": [ + 442, + 361, + 538, + 371 + ], + "spans": [ + { + "bbox": [ + 442, + 361, + 538, + 371 + ], + "type": "text", + "content": "(b) Annotation of our method" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 380, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 470 + ], + "type": "text", + "content": "Figure 1. Existing VIC method requires a unique label indicating the position of each human in each frame, and these labels are consistent across frames. Weakly supervised VIC only requires labels indicating each human position and whether they are an inflow/outflow pedestrian (purple/yellow in (b)) in the current frame. The transition from individual-level labels to identity-agnostic group-level labels represents a significant reduction in the labeling effort." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "type": "text", + "content": "to count the number of people in each frame of a video. A limitation of VCC [12, 51, 53] is that it gives an imprecise estimate of the number of unique individuals appearing in a video sequence, as people are counted multiple times if they appear in several frames. To overcome this drawback, Video Individual Counting (VIC) was introduced, wherein a method must count the total number of people with unique identities appearing in a video sequence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "The apparent approach to crowd counting is to count the people who appear in the first frame and add the number of people who come into the camera's field of view in later frames (the inflow). Following this principle, Han et al. [14] devised DRNet, the only existing published method for VIC, identifying repeated observations of individuals across consecutive frames based on their appearance, simultaneously predicting inflow and outflow. Although the number of outflows does not contribute to the final count, DRNet finds it helps determine individual associations. Han et al. [14] also explored using conventional multi-object track" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 636, + 136, + 647 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 636, + 136, + 647 + ], + "spans": [ + { + "bbox": [ + 58, + 636, + 136, + 647 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 646, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 287, + 714 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62272438, 62236008, U21B2038, and 61931008, in part by the Key Deployment Program of the Chinese Academy of Sciences under Grant KGFZD145-23-18, and in part by the Fundamental Research Funds for Central Universities (E2ET1104). Yuankai Qi, Anton van den Hengel, and Ming-Hsuan Yang are not supported by the aforementioned fundings." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19228" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "ing (MOT) methods [7, 41, 44, 54, 56] to tackle the VIC problem. Unfortunately, it turns out MOT methods suffer from poor accuracy. More notably, DRNet and MOT-based methods require trajectory labels (or similar individual association labels) to supervise identity association, which is highly annotation-expensive." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "content": "Our key insight is that counting people in the previous and current frames does not require accurate identity associations for those appearing in both frames. For example, as long as we can predict individual " + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "inline_equation", + "content": "15\\sim 17" + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "content": " as entering in Fig. 1a, even if we arbitrarily associate individuals " + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "inline_equation", + "content": "0\\sim 14" + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "content": " to individuals " + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "inline_equation", + "content": "0\\sim 14" + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 152, + 288, + 342 + ], + "type": "text", + "content": ", we can still correctly infer the crowd counts. We can still count crowds if we relax individual-level identity associations to group-level associations and do not require trajectory annotations. Thus, we just need to annotate inflow and outflow people (then we can derive the individual exists in two frames), which reduces annotation costs compared to creating individual pairwise target associations for each observed pedestrian between neighboring frames. We name crowd counting with such annotations as Weakly supervised Video Individual Counting (WVIC)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 553 + ], + "type": "text", + "content": "To address the WVIC task, we propose a benchmark method based on Contrastive learning with Group-level matching, namely CGNet. We design a soft contrastive loss to drive the network to learn discriminative representations that can facilitate identifying the required group associations and, thus, the inflow. Moreover, to better represent each individual, we design a memory-based individual count predictor, where historical templates of individuals are stored and updated in memory to enhance the robustness of association during inference. Considering that the two existing datasets for VIC are all captured by static cameras, we collect a weakly supervised VIC dataset based on moving UAVs, named UAVVIC. This dataset provides about 400,000 inflow/outflow and bounding box labels for four categories: pedestrian, car, bus, and van. Although devised for crowd counting, our proposed baseline method also performs well in counting other objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 563, + 215, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 563, + 215, + 574 + ], + "spans": [ + { + "bbox": [ + 59, + 563, + 215, + 574 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 582, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 582, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 286, + 628 + ], + "type": "text", + "content": "- We propose WVIC, a weakly supervised video individual counting task. This task does not require expensive per-target trajectory annotations and only requires two types of identity-agnostic annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 630, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 630, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 630, + 286, + 665 + ], + "type": "text", + "content": "- We automatically reannounce two existing datasets, CroHD and SenseCrowd, and collect a new dataset, UAVVIC, to pave the way for future studies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 286, + 712 + ], + "type": "text", + "content": "- We propose a strong baseline, CGNet, equipped with a newly designed group level matching soft contrastive loss, performing favorably against the supervised methods on the three datasets mentioned above." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 397, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 397, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 397, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 413 + ], + "type": "text", + "content": "Video Crowd Counting (VCC) estimates the number of people in each video frame. Most of the existing VCC methods [2, 5, 6, 11, 12, 30, 51, 53, 58] can be divided into two categories according to their solved problems: region of interest (ROI) and line of interest (LOI). ROI methods [11, 12, 51, 53] detect pedestrians within a specific region (or the whole image), and they focus on leveraging temporal context information to improve the prediction of the current frame. LSTN [12] models the group flow of crowds in local regions. Xiong et al. in [53] fuse history frame features using a ConvLSTM. TAN [51] explores context information from adjacent density maps. CLRNet proposes a local self-attention module to help the model focus on highly related regions in adjacent frames. ROI methods can not be used in the VIC task. LOI methods [2, 5, 6, 30, 57, 58] counts people passing through a specific line. Most of the existing LOI methods [2, 5, 6] apply blobs to crowds, counting objects when the whole blobs cross the line. Ma et al. in [30] sample fixed line width areas within temporal image slices and accumulated crowds in this area. Zheng et al. [58] instead sum the number of people within neighboring blocks near a line based on local velocity. Zhao et al. in [57] train a crowd velocity map predictor with the help of trajectory labels. In practice, LOI methods perform poorly on the VIC task as people can enter or leave the field of view in any direction, making it difficult to find a suitable virtual line." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 641 + ], + "type": "text", + "content": "Multiple Object Tracking (MOT) aims to predict the trajectories for multiple targets in a video. Object association is a main challenge in MOT, establishing correspondences between objects across frames to generate targets' trajectories. Existing object association methods in MOT include Probabilistic Data Association (PDA) [36, 37], Joint Probabilistic Data Association (JPDA) [16, 23, 29, 40], and graph matching methods like the Hungarian algorithm [3, 50]. Recent research has started to use more powerful feature representations to enhance the robustness of data association. Deep learning-based methods like LSTM can model temporal information to handle long-term occlusion [20, 34, 45]. Many works [4, 15, 46] also focus on the association of tracklets. Specifically, they associate multiple tracklets over longer periods using a graph model to combine the final trajectory, which can relieve the effects of lost tracked targets. These methods perform poorly on the VIC task, however, due to the inevitable ID switching problem [14]. Additionally, they require trajectory labels to train their models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Video Individual Counting (VIC). To the best of our knowledge, DRNet [14] is the only method for the VIC task. It predicts the number of initial pedestrians and matches pedestrians in adjacent frames by adapting a differentiable optimal transport loss. Although the VIC task only requires the total count in the video, DRNet needs individual" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19229" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "matching between frames. WVIC avoids the requirement for such matching annotation, which enables easy expansion of dataset scales and thus facilitates future large deep learning models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 177, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 177, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 177, + 142 + ], + "type": "text", + "content": "3. A Benchmark Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": "The goal of Weakly-supervised Video Individual Counting (WVIC) is to predict the number of unique individuals " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "(\\hat{N})" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " in a video provided a series of its frames with " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " as the sampling rate. Given a video " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^1, \\dots, \\mathbf{V}^T" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ", the label information comprises coordinate/inflow/outflow data for each individual. Specifically, letting " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^i = [p_1^i, \\dots, p_{n_i}^i]" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " denote the annotation of the coordinates of the center of the people in frame " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ", we sample the frame pairs " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{V}^1, \\mathbf{V}^{\\delta + 1}), (\\mathbf{V}^{\\delta + 1}, \\mathbf{V}^{2\\delta + 1}), \\dots, (\\mathbf{V}^{T - \\delta}, \\mathbf{V}^T)\\}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " is an integer frame offset. The weakly supervised inflow label " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^i = [\\mathbf{I}_1^i, \\dots, \\mathbf{I}_{n_i}^i]" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " and outflow label " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^i = [\\mathbf{O}_1^i, \\dots, \\mathbf{O}_{n_i}^i]" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ". If the " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": "-th person in frame " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " doesn't appear in " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{i - \\delta}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_u^i = 1" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": "; otherwise, " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_u^i = 0" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^i" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ", if the " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": "-th person in frame " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "V^i" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " doesn't appear in " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{i + \\delta}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_u^i = 1" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": "; otherwise " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_u^i = 0" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": ". If, by a slight abuse of notation, we refer to people by their coordinates in particular images, we see that the set " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\{p_u^i | \\mathbf{O}_u^i = 0\\}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\{p_u^{i + \\delta} | \\mathbf{I}_u^i = 0\\}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " represent the people that appear both in " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{i + \\delta}" + }, + { + "bbox": [ + 46, + 149, + 289, + 377 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 383, + 279, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 279, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 279, + 396 + ], + "type": "text", + "content": "3.1. Weakly Supervised Representation Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 556 + ], + "type": "text", + "content": "As shown in Fig. 2, the inference pipeline of our CGNet consists of three components: an image-level locator to generate coordinates for pedestrians, an encoder to produce representations based on pedestrians' coordinates, and a Memory-based individual Count Predictor (MCP) to predict the final inflow count based on the representations. The locator is trained independently with the coordinate annotations, and existing image crowd localization networks such as FIDT [25] can be used. The encoder is trained with our Weakly Supervised Representation Learning method (WSRL) to extract the discriminative features of each individual. It can be any feature extractor, such as ConvNext [26]. MCP does not need training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": "Provided two " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": "-adjacent frames " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^j" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " and the corresponding annotations " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^i" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^i" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^j" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", we group the individuals into four sets: " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{p_u^i | \\mathbf{O}_u^i = 0\\}" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{p_u^j | \\mathbf{I}_u^j = 0\\}" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", denoting shared individuals in the previous frame and the current frame, respectively; " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = \\mathbf{P}^i - \\mathbf{X}" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} = \\mathbf{P}^j - \\mathbf{Y}" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", denoting outflow individuals and inflow individuals, respectively. An encoder learns representations " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " for each pedestrian based on the crops around those coordinates. Specifically, to generate the " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": "-th feature " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_u^i" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", if the locator outputs points, we extract a rectangle patch of size " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "96 \\times 64" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": " centered at the predicted coordinate " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "p_u^i" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ", and if the outputs are bounding boxes, we directly crop the bounding box from the original picture. Each cropped patch is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "resized to " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " and fed into the encoder, a ConvNext-S [26], generating a feature of " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "7 \\times 7 \\times 768" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": ", which is flattened and normalized into a 1D vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "f_{u}^{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " to represent the final feature for the individual:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 340, + 122, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 122, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 340, + 122, + 545, + 144 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {F} _ {\\mathbf {X}} ^ {i} = \\operatorname {E n c o d e r} (\\mathbf {X}), \\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i} = \\operatorname {E n c o d e r} (\\hat {\\mathbf {X}}) \\\\ \\bar {\\mathbf {F}} ^ {i} = \\bar {\\mathbf {F}} _ {\\hat {\\mathbf {X}}} + \\left(\\bar {\\mathbf {H}}\\right) \\bar {\\mathbf {F}} ^ {i} = \\bar {\\mathbf {F}} _ {\\hat {\\mathbf {H}}} + \\left(\\hat {\\mathbf {H}}\\right) \\end{array} \\tag {1}", + "image_path": "ec54cbb053a37fddeb9a2aaca57539d2b767a81542d5b6058c7e7cfcd40f1377.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 340, + 140, + 511, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 140, + 511, + 156 + ], + "spans": [ + { + "bbox": [ + 340, + 140, + 511, + 156 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathbf {Y}} ^ {j} = \\operatorname {E n c o d e r} (\\mathbf {Y}), \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} = \\operatorname {E n c o d e r} (\\hat {\\mathbf {Y}})", + "image_path": "f5528f118993e7805c1dd34338b2a494afcba3adb4ad1eacbaef6f9612e5d8a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "text", + "content": "Then, a similarity matrix " + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{n_i \\times n_j}" + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^i = \\left[\\mathbf{F}_{\\mathbf{X}}^i, \\mathbf{F}_{\\hat{\\mathbf{X}}}^i\\right]" + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^j = \\left[\\mathbf{F}_{\\mathbf{Y}}^j, \\mathbf{F}_{\\hat{\\mathbf{Y}}}^j\\right]" + }, + { + "bbox": [ + 306, + 160, + 545, + 192 + ], + "type": "text", + "content": " is computed by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 201, + 545, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 201, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 318, + 201, + 545, + 234 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\left[ \\begin{array}{l l} \\left(\\mathbf {F} _ {\\mathbf {X}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\mathbf {Y}} ^ {j} & \\left(\\mathbf {F} _ {\\mathbf {X}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} \\\\ \\left(\\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\mathbf {Y}} ^ {j} & \\left(\\mathbf {F} _ {\\hat {\\mathbf {X}}} ^ {i}\\right) ^ {\\top} \\mathbf {F} _ {\\hat {\\mathbf {Y}}} ^ {j} \\end{array} \\right] = \\left[ \\begin{array}{l l} \\mathbf {S} _ {0} & \\mathbf {S} _ {1} \\\\ \\mathbf {S} _ {2} & \\mathbf {S} _ {3} \\end{array} \\right]. \\tag {2}", + "image_path": "cb7f5d858cbac3197d8c950543a1745485f0e126376e211e1b0703aa710aadc2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " is divided into four parts: " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0\\in \\mathbb{R}^{m_{i,j}\\times m_{i,j}}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1\\in \\mathbb{R}^{m_{i,j}\\times (n_j - m_{i,j})}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2\\in \\mathbb{R}^{(n_i - m_{i,j})\\times m_{i,j}}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3\\in" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(n_i - m_{i,j})\\times (n_j - m_{i,j})}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "m_{i,j} = (\\mathbf{1} - \\mathbf{I}^{\\mathbf{j}})^T(\\mathbf{1} - \\mathbf{O}^i)" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " denotes the number of shared individuals in " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^j" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " is the similarity matrix of " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " are the similarity matrix of " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ", of " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ", of " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 322, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 369 + ], + "type": "text", + "content": "To pull the matched groups (X and Y) closer and push away individual pairs from unmatched groups, we propose a weakly supervised Group-level Matching Loss (GML) to constrain different parts of S." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "text", + "content": "Constraint for " + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0, \\mathbf{S}_1, \\mathbf{S}_2" + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0" + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "text", + "content": " denotes the similarity among individuals shared between two frames, and it does not matter whether the individuals are correctly matched as long as we can assign them a one-to-one match. As for " + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2" + }, + { + "bbox": [ + 304, + 372, + 545, + 443 + ], + "type": "text", + "content": ", they are the similarities of different identities and thus should be zero." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": "To this end, we introduce a latent variable matrix " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\Omega \\in \\mathbb{R}^{m_{i,j}\\times m_{i,j}}" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": "-th element " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\Omega_{u,v}" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": " denote the probability that the " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": "-th pedestrian in " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": " is matched with " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": "-th pedestrian in " + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 444, + 545, + 538 + ], + "type": "text", + "content": ". To increase the similarity between the pairs that have a higher matching probability and decrease the similarity between the pairs that have a lower matching probability, inspired by [32, 39], we define the soft contrastive loss as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 541, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 541, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 541, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s c o n} (i) = \\min _ {\\Omega} - \\sum_ {u = 1} ^ {m _ {i, j}} \\sum_ {v = 1} ^ {m _ {i, j}} \\Omega_ {u, v} C _ {u, v}, \\tag {3}", + "image_path": "f1035b9f44db288de06aa8cd6a1f9b6c5b27a04e10c19d21a6867b77ad60f147.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 575, + 526, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 575, + 526, + 591 + ], + "spans": [ + { + "bbox": [ + 325, + 575, + 526, + 591 + ], + "type": "interline_equation", + "content": "s. t. \\quad \\mathbf {1} _ {m _ {i, j}} ^ {T} \\Omega = \\mathbf {1}, \\quad \\Omega \\mathbf {1} _ {m _ {i, j}} = \\mathbf {1}, 0 \\leq \\Omega_ {u, v} \\leq 1,", + "image_path": "52a9621e43eba0e77ab1cca51f3514768bccc0df11ef000551a675e3b73f97b7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 594, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 545, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 594, + 545, + 617 + ], + "type": "inline_equation", + "content": "C_{u,v}" + }, + { + "bbox": [ + 306, + 594, + 545, + 617 + ], + "type": "text", + "content": " denotes the contrastive similarity [10], and is calculated as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 621, + 545, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 621, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 311, + 621, + 545, + 662 + ], + "type": "interline_equation", + "content": "C _ {u, v} = \\frac {e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v}}}{e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v}} + \\sum_ {1 \\leq u ^ {\\prime} \\leq n _ {i}} ^ {u ^ {\\prime} \\neq u} e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u ^ {\\prime} , v}} + \\sum_ {1 \\leq v ^ {\\prime} \\leq n _ {j}} ^ {v ^ {\\prime} \\neq v} e ^ {\\frac {1}{\\gamma} \\mathbf {S} _ {u , v ^ {\\prime}}}}, \\tag {4}", + "image_path": "b82353cafe30c742e5f181206c36d0cebde8a0717d824c594b6a26e6c1903ec7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is a temperature hyper-parameter [52] and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{u,v} = \\langle f_u^i,f_v^j\\rangle" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", calculated by the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-th feature in " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^i" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-th feature in " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^j" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Omega_{u,v}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " describes the matching probability, our soft contrastive loss" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19230" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 75, + 520, + 274 + ], + "blocks": [ + { + "bbox": [ + 75, + 75, + 520, + 274 + ], + "lines": [ + { + "bbox": [ + 75, + 75, + 520, + 274 + ], + "spans": [ + { + "bbox": [ + 75, + 75, + 520, + 274 + ], + "type": "image", + "image_path": "899ece5f0975dd10996ad6b68d31afe5e4cd176984664a9df7e592366177ed50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": "Figure 2. The inference pipeline of our CGNet and the weakly supervised representation learning method (WSRL). The pipeline comprises a frame-level crowd locator, an encoder, and a Memory-based individual count predictor (MCP). The locator predicts the coordinates for pedestrians. The encoder generates representations for each individual, and MCP predicts inflow counts and updates the individual templates stored in the memory. To pull the matched groups " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{Y})" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " closer and push away individual pairs from unmatched groups, WSRL exploits inflow and outflow labels to optimize the encoder with a novel Group level Matching Loss (GML), which consists of a soft contrastive loss " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{scon})" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " and a hinge loss " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{hinge})" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": "evaluates the expectation of contrastive loss. It should be noting that in " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "C_{u,v}" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "(f_u^i, f_v^j)" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " is considered as positive pairs, while " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\{(f_u^i, f_{v'}^j) | v' \\neq v\\}" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\{(f_{u'}, f_v^j) | u' \\neq u\\}" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " are considered as negative pairs. Thus, " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{scon}" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " can encourage " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " to be a permutation matrix, which indicates that there exists a one-to-one match between " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": ", and drive " + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "inline_equation", + "content": "S_1, S_2" + }, + { + "bbox": [ + 46, + 357, + 289, + 441 + ], + "type": "text", + "content": " to be zero matrices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "content": "The problem defined by Eq. (3) is a typical balanced Optimal Transport (OT [35]) problem which transports " + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{m_{i,j}}" + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{m_{i,j}}" + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{1} - C" + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "content": " as the cost matrix and " + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "content": " is the transport matrix. Therefore, we consider " + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{scon}(i)" + }, + { + "bbox": [ + 46, + 446, + 287, + 506 + ], + "type": "text", + "content": " as an OT loss, which can be solved with the Sinkhorn algorithm [38]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "content": "Constraint for " + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3" + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "content": ". Obviously, the elements in " + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3" + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "content": " measure the similarity of different individuals, so they should be as small as possible. However, directly constraining " + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3" + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "content": " to zero matrix will make features from " + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}, \\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "content": " easily collapse to zero-vector. To this end, the widely used Hinge L1 loss [18] is adopted to constrain " + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3" + }, + { + "bbox": [ + 46, + 509, + 288, + 582 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 595, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 595, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 88, + 595, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h i n g e} (i) = \\frac {\\sum R e l u (\\mathbf {S _ {3}} - \\theta)}{(n _ {i} - m _ {i , j}) (n _ {j} - m _ {i , j})}, \\qquad (5)", + "image_path": "70803c820138513e4ae8108125684efdcdb25853f938a864196e8ce2513eb46b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "type": "text", + "content": " is a threshold used to ignore small values in " + }, + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_3" + }, + { + "bbox": [ + 47, + 636, + 274, + 649 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": "Finally, the group-level matching loss of the training set " + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": " is formulated as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {G M L} (\\mathbf {V}) = \\sum_ {0 < i < T, i = k \\delta + 1} \\left(\\mathcal {L} _ {s c o n} (i) + \\mathcal {L} _ {h i n g e} (i)\\right). \\tag {6}", + "image_path": "2a6ed0bfd9886baf2de697ab52ba7a9607d8e253a1dea95c1de9f0491ca9b522.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 357, + 479, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 479, + 369 + ], + "type": "text", + "content": "3.2. Memory-based Count Predictor" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": "To handle individuals' appearance variance and re-entering cases, inspired by [22, 33, 55], we propose a memory-based count predictor (MCP) to reason the final count for the video, which stores templates of recently appeared individuals in a flexible memory. Specifically, when processing frame " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{i + \\delta}" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": ", the memory is denoted as " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^i = \\{\\mathcal{G}_1^i,\\mathcal{G}_2^i,\\dots ,\\mathcal{G}_{\\mathbf{N}^i}^i\\}" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^i" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": " is the memory size at time " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_u^i = \\{g_{u,0}^i,\\dots,g_{u,k}^i\\}" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": " is a set of templates for " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": "-th pedestrian in the memory, and " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": " is the number of templates stored for " + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 375, + 545, + 495 + ], + "type": "text", + "content": "-th pedestrian." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": "We first use the crowd locator to predict the coordinates of pedestrians in " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{i + \\delta}" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": ", and then extract their corresponding features, denoted as " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\{f_1^{i + \\delta}, f_2^{i + \\delta}, \\dots, f_{n_{i + \\delta}}^{i + \\delta}\\}" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": ". To associate the individuals in " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{i + \\delta}" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": " with templates in memory, we first define the cost " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\hat{C}_{u,u'}" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": " of matching the " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": "-th pedestrian with the " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "u'" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": "-th template in " + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^i" + }, + { + "bbox": [ + 304, + 495, + 545, + 569 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 573, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 573, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 358, + 573, + 545, + 591 + ], + "type": "interline_equation", + "content": "\\hat {C} _ {u, u ^ {\\prime}} = \\max _ {v} \\left(1 - \\left\\langle f _ {u} ^ {i + \\delta}, g _ {u ^ {\\prime}, v} ^ {i} \\right\\rangle\\right). \\tag {7}", + "image_path": "30ac97713781997b0aedd1c7615b959dc206a4326f364e2cfc73a78f4e7062cf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "type": "text", + "content": "Then, we generate an optimal one-to-one match " + }, + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "type": "inline_equation", + "content": "\\pi \\in \\{0,1\\}^{n_{i + \\delta}\\times \\mathbf{N}^i}" + }, + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "type": "text", + "content": " (assuming " + }, + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^i\\geq n_i" + }, + { + "bbox": [ + 305, + 597, + 547, + 634 + ], + "type": "text", + "content": ") by solving the following problem with Hungarian Algorithm [21]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 639, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 639, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 348, + 639, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\pi = \\arg \\min _ {\\pi} \\sum_ {u, u ^ {\\prime}} \\pi_ {u, u ^ {\\prime}} \\hat {C} _ {u, u ^ {\\prime}} \\tag {8}", + "image_path": "79c8aafb6a42c14e8263c559c8137fd900a296c478542d1d848c8b63d1e7e8cd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 668, + 520, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 668, + 520, + 692 + ], + "spans": [ + { + "bbox": [ + 332, + 668, + 520, + 692 + ], + "type": "interline_equation", + "content": "s. t. \\quad \\forall u ^ {\\prime}, \\sum_ {u} \\pi_ {u, u ^ {\\prime}} = 1, \\quad \\forall u, \\sum_ {u} \\pi_ {u, u ^ {\\prime}} \\leq 1.", + "image_path": "b3d2f9fa3867236daf1fe64f6d42dd2fde2be9e8076d99752146f94459945a83.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 699, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 699, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 699, + 545, + 715 + ], + "type": "text", + "content": "If the matching cost for " + }, + { + "bbox": [ + 306, + 699, + 545, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 306, + 699, + 545, + 715 + ], + "type": "text", + "content": " meets " + }, + { + "bbox": [ + 306, + 699, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{u' = 1}^{\\mathbf{N}^i}\\pi_{u,u'}\\hat{C}_{u,u'} > \\zeta" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19231" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " will be judged as an inflow pedestrian. Otherwise, if " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\pi_{u,u'} = 1" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is associated with " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "g_{u'}^i" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": "Template update. Similar to [22], " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_u^i" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " has a time-to-live factor (ttl). " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "ttl_u" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " will decrease one if " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_u^i" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " has no mapping to any pedestrian feature in " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{i + \\delta}" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "ttl_u" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " decreases to zero, the template " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_u^i" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " will be dropped. Meanwhile, whenever " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_u^i" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " is associated with a pedestrian in the current frame, " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "ttl_u" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " will be reset to " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "ttlmax" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": ". This time-to-live threshold allows us to save some recently appeared targets while discarding those that have not appeared for a long time. If pedestrian " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " is associated with " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "u'" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": "-th template, we add " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "f_{u}^{i + \\delta}" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{u'}^i" + }, + { + "bbox": [ + 46, + 96, + 287, + 216 + ], + "type": "text", + "content": " to update the template" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 225, + 287, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 287, + 242 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {u ^ {\\prime}} ^ {i + \\delta} \\leftarrow \\mathcal {G} _ {u ^ {\\prime}} ^ {i} \\cup \\left\\{f _ {u} ^ {i + \\delta} \\mid \\pi_ {u, u ^ {\\prime}} = 1, \\sum_ {u ^ {\\prime} = 1} ^ {\\mathbf {N} ^ {i}} \\pi_ {u, u ^ {\\prime}} \\hat {C} _ {u, u ^ {\\prime}} \\leq \\zeta \\right\\}. \\tag {9}", + "image_path": "efdfe434e3b50d79b7e7569e2042199fcd381dd6cdb495e77f9de19363f7f1d4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "text", + "content": "If pedestrian " + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "text", + "content": " is an inflow pedestrian, " + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "inline_equation", + "content": "\\{f_u^{i + \\delta}\\}" + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "text", + "content": " will be added to the memory " + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{i + \\delta}" + }, + { + "bbox": [ + 47, + 248, + 287, + 273 + ], + "type": "text", + "content": " as a new template." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 274, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 287, + 297 + ], + "type": "text", + "content": "The output number of the total individuals in video " + }, + { + "bbox": [ + 47, + 274, + 287, + 297 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 274, + 287, + 297 + ], + "type": "text", + "content": " is then calculated via" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 304, + 287, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 304, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 103, + 304, + 287, + 331 + ], + "type": "interline_equation", + "content": "\\hat {N} = n _ {1} + \\sum_ {\\delta + 1 \\leq i \\leq T, i = k \\delta + 1} \\hat {n} _ {i}, \\tag {10}", + "image_path": "11cf9f0a65ee0502aa13272210b631af897abb3083aa387045b55abab0993d13.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "type": "inline_equation", + "content": "\\hat{n}_i" + }, + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "type": "text", + "content": " is the number of inflow pedestrians at time " + }, + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 339, + 268, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 382, + 106, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 106, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 106, + 393 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "content": "We test our CGNet on three datasets: CroHD [42], SenseCrowd [24], and UAVVIC. CroHD has four training videos and five testing videos. SenseCrowd contains 634 videos, and we split the train, validation, and test dataset following DRNet [14]. UAVVIC is our proposed dataset collected by a moving UAV camera in various scenes, including campus, beach, car park, highway, city road, and square. UAVVIC consists of 221 videos (100 for training, 100 for testing, and 21 for validation), and 5,396 frames are sampled with 3s as the interval. Annotations consist of 398,158 bounding boxes, and group-level matching labels in neighbor frames are provided. The resolutions of UAVVIC are 4K and 1080P for better capture of tiny pedestrians from drone view. A detailed comparison of UAVVIC with existing video crowd counting datasets [8, 9, 13, 24, 42] is shown in Tab. 1. UAVVIC provided a moving shot with a larger range of pedestrians. More details about UAVVIC are provided in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 623, + 106, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 106, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 106, + 635 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Mean Absolute Error (MAE), Mean Square Error (MSE), and Weighted Relative Absolute Errors (WRAE) are used for evaluation. The first two metrics are common metrics applied in VCC. However, unlike VCC schemes, we only count the same person once in the video. WRAE [14] is proposed to balance the performance on videos with" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "type": "table", + "html": "
NameResolutionRangeMoving shotPointBoxTrajectoryIn-Out
Mall [9]640*48013-53X✓XXX
UCSD [8]238*15811-46X✓XXX
FDST [13]1920*10809-57X✓XXX
1280*720
SenseCrowd [24]-1-296X✓✓✓X
CroHD [42]-25-346X✓✓✓X
UAVVIC3840*21600-735✓✓✓X✓
1920*1080
", + "image_path": "e48405731f81ba1343facac742cd503dc9f60af03ac01bd197a82319db80479d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 148, + 542, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 148, + 542, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 542, + 159 + ], + "type": "text", + "content": "Table 1. Comparison of different video crowd counting datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": "different lengths and pedestrian numbers: " + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "inline_equation", + "content": "WRAE = \\sum_{i=1}^{K} \\frac{T_i}{\\sum_{j=1}^{K} T_j} \\frac{|N_i - \\hat{N}_i|}{N_i}" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": " is the total number of videos, and " + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": " is the length of video " + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "inline_equation", + "content": "N_i / \\hat{N}_i" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": " denotes the number of ground truth/predicted counts in " + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 167, + 547, + 224 + ], + "type": "text", + "content": "-th video." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 234, + 440, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 440, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 440, + 247 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "content": "Training. If there are no special instructions, the loca-tor we applied is a well-known crowd localization model, FIDT [25]. All parameters are official from FIDT, except that we replaced the backbone from HRNet-W48 [47] to HRNet-W18 [47] for efficient inference. The learning rate is set as " + }, + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "content": " along with AdamW [28] as the optimizer and applying the pre-train weight from Timm [49]. All related models can be trained on one RTX3090 (24G memory)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "content": "Testing. In the testing phase, the time interval " + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "inline_equation", + "content": "3s" + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "content": " following the setting in [14], the threshold " + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "inline_equation", + "content": "0.7" + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 350, + 545, + 397 + ], + "type": "text", + "content": " in Eq. (4) is set as 10, and the max time-to-live (ttlmax) is set as 3, the max size of memory (memmax) is set as 5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 407, + 427, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 427, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 427, + 418 + ], + "type": "text", + "content": "4.4. Overall Performance" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "Comparison Methods. To the best of our knowledge, only DRNet[14] has been specifically designed for the VIC task. To evaluate the effectiveness of our proposed CGNet, we also tested two categories of approaches that could be applied to VIC: 1) Multi-object tracking (MOT) methods, including HeadHunter-T [43], FairMOT [56], PHDTT [44], SMILE [48], SparseTrack [27], Deep-OC-SORT [31], and BoT-SORT [1], where we consider the total number of indices in the entire video as individual count. We set the frame rate to 1 FPS for HeadHunter-T and 0.33 FPS for the other MOT methods to obtain better VIC performance. 2) A recent cross-line video crowd-counting, LOI [57], which also requires trajectory labels in training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Results on SenseCrowd: As shown in Tab. 2, it is evident that our CGNet exhibits superior performance on this larger dataset, SenseCrowd. Particularly, CGNet improved the overall MAE by about " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " compared to DRNet. Furthermore, we achieve the lowest MAE across all density levels except for D0. Compared to the existing lowest MAE on D1, D2, D3, and D4, the performance improvements of our CGNet are about " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", respectively. In Fig. 3, we show qualitative results on three representative scenes,i.e., mall, station, and outdoor plaza. Although the inflows in the current flow are not concentrated, CGNet" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19232" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 69, + 228, + 144 + ], + "blocks": [ + { + "bbox": [ + 97, + 69, + 228, + 144 + ], + "lines": [ + { + "bbox": [ + 97, + 69, + 228, + 144 + ], + "spans": [ + { + "bbox": [ + 97, + 69, + 228, + 144 + ], + "type": "image", + "image_path": "e5b7e7e501e5f533ddca8baee52582ff7025253a685864959c262df4ba75b8ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 70, + 361, + 144 + ], + "blocks": [ + { + "bbox": [ + 231, + 70, + 361, + 144 + ], + "lines": [ + { + "bbox": [ + 231, + 70, + 361, + 144 + ], + "spans": [ + { + "bbox": [ + 231, + 70, + 361, + 144 + ], + "type": "image", + "image_path": "6c99610530f15a593083609005a31cb0ea38e2016e75a6676865517b0b668667.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 146, + 334, + 155 + ], + "lines": [ + { + "bbox": [ + 260, + 146, + 334, + 155 + ], + "spans": [ + { + "bbox": [ + 260, + 146, + 334, + 155 + ], + "type": "text", + "content": "(a) Previous frame (0s)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 365, + 70, + 496, + 144 + ], + "blocks": [ + { + "bbox": [ + 365, + 70, + 496, + 144 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 496, + 144 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 496, + 144 + ], + "type": "image", + "image_path": "fbb32e315b8ecd072b065c520434f9980f6ac9f1b666bd2a7058b683018c8010.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 99, + 156, + 228, + 230 + ], + "blocks": [ + { + "bbox": [ + 99, + 156, + 228, + 230 + ], + "lines": [ + { + "bbox": [ + 99, + 156, + 228, + 230 + ], + "spans": [ + { + "bbox": [ + 99, + 156, + 228, + 230 + ], + "type": "image", + "image_path": "7a94857b819b4959248b6767268bdd292414e3fac4548a45ebc63fc1f3324fdb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 231, + 138, + 238 + ], + "lines": [ + { + "bbox": [ + 102, + 231, + 138, + 238 + ], + "spans": [ + { + "bbox": [ + 102, + 231, + 138, + 238 + ], + "type": "text", + "content": "GT inflow: 4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 231, + 227, + 237 + ], + "lines": [ + { + "bbox": [ + 181, + 231, + 227, + 237 + ], + "spans": [ + { + "bbox": [ + 181, + 231, + 227, + 237 + ], + "type": "text", + "content": "Predict inflow: 4" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 231, + 156, + 361, + 230 + ], + "blocks": [ + { + "bbox": [ + 231, + 156, + 361, + 230 + ], + "lines": [ + { + "bbox": [ + 231, + 156, + 361, + 230 + ], + "spans": [ + { + "bbox": [ + 231, + 156, + 361, + 230 + ], + "type": "image", + "image_path": "f9d22f4482517acb43e4a085fe28e1f5180b6371669a9502d1f61c71cb616879.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 231, + 271, + 237 + ], + "lines": [ + { + "bbox": [ + 233, + 231, + 271, + 237 + ], + "spans": [ + { + "bbox": [ + 233, + 231, + 271, + 237 + ], + "type": "text", + "content": "GT inflow: 11" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 244, + 333, + 254 + ], + "lines": [ + { + "bbox": [ + 260, + 244, + 333, + 254 + ], + "spans": [ + { + "bbox": [ + 260, + 244, + 333, + 254 + ], + "type": "text", + "content": "(b) Current frame (3s)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 261, + 547, + 284 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 284 + ], + "type": "text", + "content": "Figure 3. Results on the SenseCrowd dataset. Red dots are predictions of the locator. Blue/green/yellow circles are correct predicted inflow, error predicted inflow, and missed inflow, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 156, + 493, + 230 + ], + "blocks": [ + { + "bbox": [ + 362, + 156, + 493, + 230 + ], + "lines": [ + { + "bbox": [ + 362, + 156, + 493, + 230 + ], + "spans": [ + { + "bbox": [ + 362, + 156, + 493, + 230 + ], + "type": "image", + "image_path": "e3e6f205947c67ee45c2337fe81a5d29cca46f4b22e4a61144c4f53badd9593a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 231, + 405, + 238 + ], + "lines": [ + { + "bbox": [ + 365, + 231, + 405, + 238 + ], + "spans": [ + { + "bbox": [ + 365, + 231, + 405, + 238 + ], + "type": "text", + "content": "GT inflow: 15" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 231, + 490, + 237 + ], + "lines": [ + { + "bbox": [ + 440, + 231, + 490, + 237 + ], + "spans": [ + { + "bbox": [ + 440, + 231, + 490, + 237 + ], + "type": "text", + "content": "Predict inflow: 14" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 70, + 285, + 522, + 415 + ], + "blocks": [ + { + "bbox": [ + 70, + 285, + 522, + 415 + ], + "lines": [ + { + "bbox": [ + 70, + 285, + 522, + 415 + ], + "spans": [ + { + "bbox": [ + 70, + 285, + 522, + 415 + ], + "type": "table", + "html": "
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
FairMOT [56]35.462.348.913.522.467.984.4145.8
HeadHunter-T [43]30.050.638.611.825.756.092.6131.4
SMILE [48]27.2236.732.59.221.033.876.5203.5
SparseTrack [27]28.3643.830.89.921.542.189.2189.5
Deep-OC-SORT [31]26.0443.629.59.525.129.160.3156.5
BoT-SORT [1]24.6732.428.38.720.531.170.2165.1
LOI [57]24.733.137.412.525.439.339.686.7
DRNet [14]12.324.712.74.18.023.350.077.0
CGNet (Ours)8.8617.6912.65.05.88.525.063.4
", + "image_path": "c0d38cb62a97a0ac175e5ed75f8d35d4eae4e3092decf53a5cb3391c6a51f1ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 69, + 451, + 523, + 593 + ], + "blocks": [ + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "type": "text", + "content": "Table 2. Performance comparison on SenseCrowd. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and " + }, + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "type": "inline_equation", + "content": "[200, +\\infty)" + }, + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "type": "text", + "content": ", respectively. The best values are highlighted in blue font." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 451, + 523, + 593 + ], + "lines": [ + { + "bbox": [ + 69, + 451, + 523, + 593 + ], + "spans": [ + { + "bbox": [ + 69, + 451, + 523, + 593 + ], + "type": "table", + "html": "
MethodMAEMSEWRAE(%)MAE on five testing scenes
CroHD11CroHD12CroHD13CroHD14CroHD15
PHDTT [44]2130.42808.3401.6247379347944911327
FairMOT [56]256.2300.844.1114272844081000
HeadHunter-T [43]253.2351.732.7651015155821003
SMILE [48]257.6334.540.516231156649246
SparseTrack [27]176.6208.627.62917225833688
Deep-OC-SORT [31]165.2195.933.16835118616160
BoT-SORT [1]154.8176.4227.44921013828691
LOI [57]305.0371.146.060243458630131
DRNet [14]141.1192.327.4313381825561
CGNet (Ours)75.095.114.577214144138
", + "image_path": "9012d7641343fde91633a4c50d17fdf921d3a3e19cf0f34ebd9638a1715f0310.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 601, + 546, + 624 + ], + "lines": [ + { + "bbox": [ + 46, + 601, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 546, + 624 + ], + "type": "text", + "content": "Table 3. Performance comparison on CroHD dataset. CroHD11-CroHD15 are five test scenes labeled in [14]. The best values are highlighted in blue font." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 635, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 659 + ], + "type": "text", + "content": "can accurately predict them. More visualization results can be found in supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": "Results on CroHD. Comparison with existing methods on CroHD is shown in Tab. 3. Our CGNet achieved the lowest MAE, MSE, and WRAE. Especially, except for CroHD15, our method obtained the lowest MAE across the other four" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 635, + 546, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 546, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 546, + 708 + ], + "type": "text", + "content": "testing scenes. As CGNet is trained with only point annotations, it cannot crop the patch of optimal size to extract features on CroHD15. It should be noted that on CroHD13 and CroHD14, which have the highest average density (245.9 and 259.6 person/frame), the performances of all the compared methods except for DRNet degrade sig-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19233" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 521, + 144 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 521, + 144 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 521, + 144 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 521, + 144 + ], + "type": "table", + "html": "
MethodMAE/MSE/WRAEMAE/MSE/WRAE on four different categories
PedestrianCarBusVan
BoT-SORT [1]49.9/228.2/29.992.4/398.2/49.059.1/259.1/39.223.0/109.6/57.125.1/146.2/77.3
Deep-OC-SORT [31]37.1/115.4/35.285.2/213.5/22.133.4/113.1/21.019.1/ 78.2/26.210.7/56.6/32.8
DRNet [14]18.4/ 41.0/15.234.5/ 92.1/19.119.0/ 32.7/17.810.4/ 22.1/32.29.9/17.2/29.4
CGNet (Ours)12.9/ 37.4/ 12.031.2/ 98.4/14.416.9/ 45.2/ 7.32.1/ 4.1/19.21.3/ 2.0/25.1
", + "image_path": "a26c5f9df63f7b9f86dd10e61a89a9220be79118c97f56cd1b9240505d72c015.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 194, + 286, + 252 + ], + "blocks": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "type": "text", + "content": "Table 4. Performance comparison on UAVVIC. All the methods are trained independently with four classes. The values of each entry are MAE/MSE/WRAE. The best values are highlighted in blue font." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 194, + 286, + 252 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 286, + 252 + ], + "type": "table", + "html": "
MethodMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
①13.3532.322.89.712.315.326.543.4
②9.8420.013.85.36.312.028.867.7
③10.7117.4016.96.98.911.724.350.6
④8.8617.6912.65.05.88.525.063.4
", + "image_path": "a6bbde685b6fa7962f4477b10000351331ffa4855987c8295d63e69fd98c01f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": "Table 5. Ablation study on the main components. The density labels D0-D4 are defined in [14], where D0, D1, D2, D3, and D4 correspond to the count range [0, 50), [50, 100), [100, 150), [150, 200) and " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "[200, +\\infty)" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": ", No constrain on " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": ", w/o Eq. (5); " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": ", w/o MCP; " + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 46, + 260, + 287, + 326 + ], + "type": "text", + "content": ", full CGNet. The best values are highlighted in blue font." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 406 + ], + "type": "text", + "content": "nificantly. In contrast, the performance of our CGNet is relatively stable in these two scenes. Furthermore, even without individual-level matching labels, our CGNet performs favorably against the state-of-the-art approach, DRNet, by a large margin (about " + }, + { + "bbox": [ + 46, + 334, + 287, + 406 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 46, + 334, + 287, + 406 + ], + "type": "text", + "content": " on the MAE), demonstrating the effectiveness of our method." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 413, + 166, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 413, + 166, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 166, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 166, + 485 + ], + "type": "image", + "image_path": "eba599d1e565fbc096ce4f4681298ae1b78e4dbba74f2bdcdc85ceba407b63cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 487, + 148, + 495 + ], + "lines": [ + { + "bbox": [ + 72, + 487, + 148, + 495 + ], + "spans": [ + { + "bbox": [ + 72, + 487, + 148, + 495 + ], + "type": "text", + "content": "(a) Previous frame (0s)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 498, + 280, + 508 + ], + "lines": [ + { + "bbox": [ + 53, + 498, + 280, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 280, + 508 + ], + "type": "text", + "content": "Figure 4. Qualitative visualization of our CGNet on UAVVIC." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 167, + 414, + 277, + 485 + ], + "blocks": [ + { + "bbox": [ + 167, + 414, + 277, + 485 + ], + "lines": [ + { + "bbox": [ + 167, + 414, + 277, + 485 + ], + "spans": [ + { + "bbox": [ + 167, + 414, + 277, + 485 + ], + "type": "image", + "image_path": "6706da357701209329e8b5b7ad86308250a97e53c3680723cfa6df129ee09c33.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 487, + 258, + 495 + ], + "lines": [ + { + "bbox": [ + 186, + 487, + 258, + 495 + ], + "spans": [ + { + "bbox": [ + 186, + 487, + 258, + 495 + ], + "type": "text", + "content": "(b) Current frame (3s)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "content": "Results on UAVVIC. The results on our proposed UAVVIC dataset are shown in Tab. 4. All methods are trained and tested on each category, respectively. Since there is no trajectory annotation on UAVVIC, we train the compared methods [1, 14, 31] on VisDrone [59], a UAV dataset for MOT. Our CGNet shows favorable performance in all categories. The average MAE is improved by about " + }, + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "content": " compared to DRNet. In Fig. 4, we show the quality result of our CGNet on the UAVVIC dataset. Although the UAV moves, CGNet can still detect inflow objects accurately." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 635, + 147, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 147, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 147, + 647 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "We conducted several ablation studies on the SenseCrowd. 1). We verify the effectiveness of the main components of our CGNet. 2). We analyze the effect of two main parameters. 3). We report the performance of our CGNet with different locators. 4). We use the trajectories generated by" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 196, + 425, + 325 + ], + "blocks": [ + { + "bbox": [ + 315, + 196, + 425, + 325 + ], + "lines": [ + { + "bbox": [ + 315, + 196, + 425, + 325 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 425, + 325 + ], + "type": "image", + "image_path": "23d33fb0d1353c0e5af4cb555f60bb704ace775dad287b5aeef8cc6cdbd127d9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 329, + 376, + 338 + ], + "lines": [ + { + "bbox": [ + 365, + 329, + 376, + 338 + ], + "spans": [ + { + "bbox": [ + 365, + 329, + 376, + 338 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 428, + 193, + 536, + 325 + ], + "blocks": [ + { + "bbox": [ + 428, + 193, + 536, + 325 + ], + "lines": [ + { + "bbox": [ + 428, + 193, + 536, + 325 + ], + "spans": [ + { + "bbox": [ + 428, + 193, + 536, + 325 + ], + "type": "image", + "image_path": "0bd1d33c6e964bea8c7791d00f4c4f20b8120e9b0dfc146bc0b9d2cac17ad65e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 329, + 487, + 338 + ], + "lines": [ + { + "bbox": [ + 476, + 329, + 487, + 338 + ], + "spans": [ + { + "bbox": [ + 476, + 329, + 487, + 338 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 348, + 546, + 381 + ], + "lines": [ + { + "bbox": [ + 305, + 348, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 546, + 381 + ], + "type": "text", + "content": "Figure 5. Effectiveness of MCP on SenseCrowd. (a) The average MAE with different lengths of videos with and w/o MCP. (b) Performance with different max time-to-live factor tllmax of MCP." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 396, + 544, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 544, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 544, + 420 + ], + "type": "text", + "content": "our CGNet as pseudo-labels for existing VIC methods. 5). We compare the time cost of the WVIC and VIC labels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": "Effectiveness of Constrain for " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1, \\mathbf{S}_2" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": ". Eq. (4) constrains " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0, \\mathbf{S}_1, \\mathbf{S}_2" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " together. Without considering un-matched pairs in " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\times \\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} \\times \\mathbf{Y}" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": ", it can be changed to constrain only " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " by remove elements in " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1, \\mathbf{S}_2" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": ". As shown in Tab. 5, MAE dropped by about 4.51 without constraints on " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " (① VS ④). This is because elements in " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " are related to pedestrians that exist only in one frame. Considering " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\times \\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} \\times \\mathbf{Y}" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " as negative pairs, our soft contrastive loss can push apart individuals in " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} (\\hat{\\mathbf{Y}})" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": " with that in " + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} (\\mathbf{X})" + }, + { + "bbox": [ + 305, + 421, + 546, + 542 + ], + "type": "text", + "content": ", thus enhancing the distinguishing ability of the representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "type": "text", + "content": "Effectiveness of Eq. (5). We replace the hinge MAE loss in Eq. (5) to normal L1 loss and derive the method ② in Tab. 5. Compared to L1 loss, our " + }, + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hinge}(\\cdot)" + }, + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "type": "text", + "content": " gives a soft margin to the inner product of two features other than encouraging them to be 0. Optimizing the inner product of all features toward 0 makes it easier to cause either side to collapse a 0 vector, which is not conducive to learning different representations. Compared the performance of ② with that of ④ in Tab. 5, " + }, + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hinge}(\\cdot)" + }, + { + "bbox": [ + 305, + 544, + 545, + 652 + ], + "type": "text", + "content": " improves MAE in all density levels." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "Effectiveness of MCP. As shown in Tab. 5, compared " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": ", MCP brings about 4.1 improvements on WRAE. Specifically, we evaluate the average MAE of our CGNet on videos with different lengths. As shown in Fig. 5(a), the performance gap between " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " becomes larger as" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19234" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 284, + 142 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 284, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 284, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 284, + 142 + ], + "type": "table", + "html": "
δMAEMSEWRAE(%)MAE on five different density levels
D0D1D2D3D4
510.9620.2619.56.56.910.429.279.9
48.7114.3312.94.24.910.227.870.1
38.8617.6912.65.05.88.525.063.4
29.7720.0117.37.18.812.417.452.6
19.5615.3416.66.87.410.523.145.5
", + "image_path": "ae7cfd1c73575eed1861557f6a1de15577f42c0f437bb25e5da38115d222f75e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 175, + 284, + 246 + ], + "blocks": [ + { + "bbox": [ + 47, + 150, + 286, + 172 + ], + "lines": [ + { + "bbox": [ + 47, + 150, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 150, + 286, + 172 + ], + "type": "text", + "content": "Table 6. Ablation study on interval " + }, + { + "bbox": [ + 47, + 150, + 286, + 172 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 150, + 286, + 172 + ], + "type": "text", + "content": " . The best values are highlighted in blue font." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 175, + 284, + 246 + ], + "lines": [ + { + "bbox": [ + 50, + 175, + 284, + 246 + ], + "spans": [ + { + "bbox": [ + 50, + 175, + 284, + 246 + ], + "type": "table", + "html": "
MethodLocatorEncoderMAEMSEWRAE
①FIDTConvNext-S8.8617.699.27
②Yolov8 nanoConvNext-S8.0116.428.85
③Yolov8 smallConvNext-S7.5615.198.05
④VGG16+FPNConvNext-S9.2017.459.55
⑤VGG16+FPNPrRoiIPooling10.0118.229.70
⑥VGG16+FPNPrRoiIPooling12.5923.3212.58
", + "image_path": "8eb5fc7fbef258fba022843e9036fd6f2b239ec782d1165e5aa52ce5b7890316.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 286, + 354 + ], + "type": "text", + "content": "the length of the video increases. This is because, with the stored templates of individuals, MCP can handle appearance variations and individuals' re-entry, alleviating error accumulation during inflow reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 354, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 286, + 496 + ], + "type": "text", + "content": "Effect of ttlmax. The performance with different max long-time-live factor ttlmax is shown in Fig. 5(b). Compared with no memory (i.e., ttlmax = 0), adding memory is helpful for all density levels except for extremely high-density levels (D4). It may be because high-density crowds result in too many candidate individual-level matches, making it difficult for our CGNet to choose the right one. Meanwhile, the performance degrades when ttlmax increases from 3 to 6 for that larger ttlmax = 6 leads to the templates of individuals that appeared a long time ago being stored in the memory. As the probability of these people re-entering the field is low, the recorded templates become distractions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "content": "Effect of Interval " + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "content": ". The performance with different intervals " + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "content": " is shown in Tab. 6. The performance on higher-density video is better when " + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "content": " is smaller. This is because crowds flow faster in those videos, and long intervals may lead to missing pedestrians in sampled frames. The overall best performance is achieved when " + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "inline_equation", + "content": "4s" + }, + { + "bbox": [ + 46, + 498, + 286, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": "Effect of the Locator. We replaced FIDT in our CGNet with Yolo V8 [19] nano and small and the localization branch (VGG16+FPN) used in DRNet, respectively. Different from FIDT, these three locators are trained with box annotations. For Yolo V8, the predicted bounding boxes are fed to our feature extractor to generate representations of individuals. For a fair comparison with DRNet, we also replace our feature extractor ConvNext-S with PrRoIPooling [17] used in DRNet [14] and remove MCP. As shown in Tab. 7, changing locators can further boost the performance of our method. With the same locator and feature extractor as DRNet (⑤), the MAE of our method is 10.01, 2.99 lower" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 171 + ], + "blocks": [ + { + "bbox": [ + 46, + 255, + 286, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 298 + ], + "type": "text", + "content": "Table 7. Performance on SenseCrowd of our scheme with different locators and feature extractors. In " + }, + { + "bbox": [ + 46, + 255, + 286, + 298 + ], + "type": "inline_equation", + "content": "⑥" + }, + { + "bbox": [ + 46, + 255, + 286, + 298 + ], + "type": "text", + "content": ", MCP is replaced with the inflow reasoning method in DRNet. The best values are highlighted in blue font." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 543, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 171 + ], + "type": "table", + "html": "
LocatorGround Truth LabelsPseudo Labels
MAEMSEWRAEMAEMSEWRAE
FairMOT35.462.348.937.264.849.8
HeadHunter-T30.050.638.632.554.239.9
SMILE27.236.732.527.537.035.2
SparseTrack28.443.830.830.545.031.5
Deep-OC-SORT26.043.629.529.545.331.7
BoT-SORT24.732.428.325.035.933.8
LOI24.733.137.426.935.239.9
DRNet12.324.712.714.525.513.3
", + "image_path": "5a599c73c507469a23de8512fb30407c6e3bd3b7f2d34ca6de6290a45aa7ce85.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 209, + 548, + 246 + ], + "blocks": [ + { + "bbox": [ + 306, + 178, + 544, + 201 + ], + "lines": [ + { + "bbox": [ + 306, + 178, + 544, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 544, + 201 + ], + "type": "text", + "content": "Table 8. Performance of existing VIC methods trained with different trajectory labels." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 209, + 548, + 246 + ], + "lines": [ + { + "bbox": [ + 307, + 209, + 548, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 548, + 246 + ], + "type": "table", + "html": "
(ni, ni+δ)(9,8)(60,53)(116,115)(289,292)(645,724)
Association [14]23.8 s104.5 s174.4 s586.9 s1663.1 s
In-Out4.4 s42.6 s53.7 s95.0 s571.9 s
", + "image_path": "61a49e0cdf5c19ecb8e0c8e36cfd5c132271ab2594c38a5853757de87e00eb6e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "type": "text", + "content": "Table 9. Time cost for a pair of frames with two types of labeling methods. " + }, + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "type": "text", + "content": " denotes the number of pedestrians in frame " + }, + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 255, + 545, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 281, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 328 + ], + "type": "text", + "content": "than that of DRNet. Even without MCP (⑥), the MAE of our method is only slightly larger than that of DRNet (12.59 v.s. 12.3), demonstrating that our method can learn effective representations without trajectory labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "type": "text", + "content": "Performs as a Pseudo Trajectory Generator. The latent variable " + }, + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "type": "text", + "content": " is obtained by solving a problem defined by Eq. (3). The Hungarian Algorithm generates a one-to-one match using " + }, + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 330, + 545, + 449 + ], + "type": "text", + "content": " as the cost matrix. Based on the matching, we generate pseudo trajectory labels, which are used to train existing VIC methods. We show the results on SenseCrowd in Tab. 8. With the generated pseudo label, the performance of all the methods degrades slightly than using the ground truth trajectory labels, demonstrating that exact individual-level matching annotations may not be necessary for VIC." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 451, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 522 + ], + "type": "text", + "content": "Label time cost. We compare the time cost of annotating our weakly supervised label and association label [14] on five frame pairs with various densities in Tab. 9. The annotation time for our weakly supervised label is only about " + }, + { + "bbox": [ + 304, + 451, + 545, + 522 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 451, + 545, + 522 + ], + "type": "text", + "content": " of that of the association label, making it easier to build a larger dataset for this VIC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 538, + 383, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 383, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 383, + 551 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 559, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 702 + ], + "type": "text", + "content": "We propose a new task, Weakly supervised Video Individual Counting (WVIC), in this paper. Unlike conventional video individual counting tasks, which require expensive trajectory labels, our WVIC task just requires two types of easily available identity-agnostic annotations. This largely reduces the annotation cost and facilitates future works to upscale data size. Furthermore, we propose a strong baseline for the WVIC task. The key idea is a Group-level Matching Loss (GML), which makes it easier to distinguish inflow individuals from all objects. Last, we have extended two existing datasets and built a totally new dataset for the WVIC task. Our method achieves favorable performance on all three datasets, even compared to supervised methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19235" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 710 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 155 + ], + "type": "text", + "content": "[2] Antonio Albiol, Inmaculada Mora, and Valery Naranjo. Real-time high density people counter using morphological tools. IEEE TITS, 2(4):204-218, 2001. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 288, + 198 + ], + "type": "text", + "content": "[3] Marco Allodi, Alberto Broggi, Domenico Giaquinto, Marco Patander, and Antonio Prioletti. Machine learning in tracking associations with stereo vision and lidar observations for an autonomous vehicle. In IVS, pages 648-653, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 240 + ], + "type": "text", + "content": "[4] Seung Hwan Bae and Kuk-Jin Yoon. Confidence-based data association and discriminative deep appearance learning for robust online multi-object tracking. IEEE TPAMI, 40(3):595-610, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 241, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 273 + ], + "type": "text", + "content": "[5] Javier Barandiarán, Berta Murguia, and Fernando Boto. Real-time people counting using multiple lines. In WIAMIS, pages 159-162, 2008. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 304 + ], + "type": "text", + "content": "[6] Jesús Bescós, José M. Menéndez, and Narciso García. DCT based segmentation applied to a scalable zenithal people counter. In ICIP, pages 1005-1008, 2003. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 305, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 288, + 337 + ], + "type": "text", + "content": "[7] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, pages 1-6, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 337, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 368 + ], + "type": "text", + "content": "[8] Antoni B. Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE TPAMI, 30(5):909-926, 2008. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 288, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 400 + ], + "type": "text", + "content": "[9] Ke Chen, Chen Change Loy, Shaogang Gong, and Tony Xiang. Feature mining for localised crowd counting. In BMVC, pages 3-14, 2012. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 401, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 287, + 432 + ], + "type": "text", + "content": "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "text", + "content": "[11] Li Dong, Haijun Zhang, Jianghong Ma, Xiaofei Xu, Yimin Yang, and Q. M. Jonathan Wu. Clrnet: A cross locality relation network for crowd counting in videos. IEEE TNNLS, pages 1-15, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "type": "text", + "content": "[12] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. In ICME, pages 814-819, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 507, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 507, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 507, + 287, + 548 + ], + "type": "text", + "content": "[13] Yanyan Fang, Biyun Zhan, Wandi Cai, Shenghua Gao, and Bo Hu. Locality-constrained spatial transformer network for video crowd counting. arXiv preprint arXiv:1907.07911, 2019.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 549, + 287, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 549, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 48, + 549, + 287, + 582 + ], + "type": "text", + "content": "[14] Tao Han, Lei Bai, Junyu Gao, Qi Wang, and Wanli Ouyang. DR.VIC: decomposition and reasoning for video individual counting. In CVPR, pages 3073-3082, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[15] Yuhang He, Jie Han, Wentao Yu, Xiaopeng Hong, Xing Wei, and Yihong Gong. City-scale multi-camera vehicle tracking by semantic attribute parsing and cross-camera tracklet matching. In CVPRW, pages 2456-2465, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "text", + "content": "[16] Andinet Hunde and Beshah Ayalew. Automated multi-target tracking in public traffic in the presence of data association uncertainty. In ACC, pages 300-306, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 657, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 688 + ], + "type": "text", + "content": "[17] Borui Jiang, Ruixuan Luo, Jiayuan Mao, Tiantian Xiao, and Yuning Jiang. Acquisition of localization confidence for accurate object detection. In ECCV, pages 784-799, 2018. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 689, + 287, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 287, + 710 + ], + "type": "text", + "content": "[18] Thorsten Joachims. Support vector method for multivariate performance measures. In ICML, pages 377-384, 2005. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "text", + "content": "[20] Chanho Kim, Fuxin Li, Mazen Alotaibi, and James M. Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, pages 9553-9562, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 148, + 545, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 179 + ], + "type": "text", + "content": "[21] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 180, + 545, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 180, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 545, + 212 + ], + "type": "text", + "content": "[22] L.J. Latecki and R. Miezianko. Object tracking with dynamic template update and occlusion detec. In ICPR, volume 1, pages 556-560, 2006. 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 213, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 243 + ], + "type": "text", + "content": "[23] Eui-Hyuk Lee, Qian Zhang, and Taek Lyul Song. Markov chain realization of joint integrated probabilistic data association. Sensors, 17(12):2865, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 244, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 286 + ], + "type": "text", + "content": "[24] Haopeng Li, Lingbo Liu, Kunlin Yang, Shinan Liu, Junyu Gao, Bin Zhao, Rui Zhang, and Jun Hou. Video crowd localization with multifocus gaussian neighborhood attention and a large-scale benchmark. IEEE TIP, 31:6032-6047, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "text", + "content": "[25] Dingkang Liang, Wei Xu, Yingying Zhu, and Yu Zhou. Focal inverse distance transform maps for crowd localization. IEEE TMM, 25:6040-6052, 2023. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 319, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 350 + ], + "type": "text", + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 351, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 351, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 351, + 545, + 392 + ], + "type": "text", + "content": "[27] Zelin Liu, Xinggang Wang, Cheng Wang, Wenyu Liu, and Xiang Bai. Sparsetrack: Multi-object tracking by performing scene decomposition based on pseudo-depth. arXiv preprint arXiv:2306.05238, 2023. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 393, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 414 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In *ICLR*, 2018. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 415, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 456 + ], + "type": "text", + "content": "[29] Zhongzhen Luo, Mina Attari, Saeid R. Habibi, and Martin von Mohrenschild. Online multiple maneuvering vehicle tracking system based on multi-model smooth variable structure filter. IEEE TITS, 21(2):603-616, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 457, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 457, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 457, + 545, + 489 + ], + "type": "text", + "content": "[30] Zheng Ma and Antoni B. Chan. Counting people crossing a line using integer programming and local features. IEEE TCSVT, 26(10):1955-1969, 2016. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 490, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 531 + ], + "type": "text", + "content": "[31] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 532, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 563 + ], + "type": "text", + "content": "[32] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 564, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 596 + ], + "type": "text", + "content": "[33] Zhuojin Pan and Xiujuan Wang. Correlation tracking algorithm based on adaptive template update. In ICISP, volume 1, pages 98-101, 2010. 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 596, + 545, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 596, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 307, + 596, + 545, + 637 + ], + "type": "text", + "content": "[34] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, pages 164-173, 2021. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 638, + 545, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 638, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 307, + 638, + 545, + 670 + ], + "type": "text", + "content": "[35] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning, 11(5-6):355-607, 2019. 4" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 671, + 545, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 671, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 307, + 671, + 545, + 702 + ], + "type": "text", + "content": "[36] Christopher Rasmussen and Gregory D. Hager. Probabilistic data association methods for tracking complex visual objects. IEEE TPAMI, 23(6):560-576, 2001. 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 703, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 703, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 703, + 545, + 712 + ], + "type": "text", + "content": "[37] Christian Ritter, Andrea Imle, Ji Young Lee, Barbara Muller," + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19236" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 705 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Oliver T. Fackler, Ralf Bartenschlager, and Karl Rohr. Two-filter probabilistic data association for tracking of virus particles in fluorescence microscopy images. In *ISBI*, pages 957–960, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 148 + ], + "type": "text", + "content": "[38] Richard Sinkhorn. A Relationship Between Arbitrary Positive Matrices and Doubly Stochastic Matrices. The Annals of Mathematical Statistics, 35(2):876 - 879, 1964. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 191 + ], + "type": "text", + "content": "[39] Rakshith Sharma Srinivasa, Jaejin Cho, Chouchang Yang, Yashas Malur Saidutta, Ching-Hua Lee, Yilin Shen, and Hongxia Jin. CWCL: Cross-modal transfer with continuously weighted contrastive loss. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 191, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 191, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 191, + 287, + 233 + ], + "type": "text", + "content": "[40] Jason Stauch, Travis Bessell, Mark Rutten, Jason Baldwin, Moriba Jah, and Keric Hill. Joint probabilistic data association and smoothing applied to multiple space object tracking. JGCD, 41(1):19-33, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 266 + ], + "type": "text", + "content": "[41] Ramana Sundararaman, Cedric De Almeida Braga, Éric Marchand, and Julien Pettré. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 298 + ], + "type": "text", + "content": "[42] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "text", + "content": "[43] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pettre. Tracking pedestrian heads in dense crowd. In CVPR, pages 3865-3875, June 2021. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 330, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 287, + 372 + ], + "type": "text", + "content": "[44] Xuan-Thuy Vo, Van-Dung Hoang, Duy-Linh Nguyen, and Kang-Hyun Jo. Pedestrian head detection and tracking via global vision transformer. In IFCVW, pages 155-167, 2022. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 373, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 415 + ], + "type": "text", + "content": "[45] Xingyu Wan, Jinjun Wang, Zhifeng Kong, Qing Zhao, and Shunming Deng. Multi-object tracking using online metric learning with long short-term memory. In ICIP, pages 788-792, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[46] Bing Wang, Gang Wang, Kap Luk Chan, and Li Wang. Tracklet association by online target-specific metric learning and coherent dynamics estimation. IEEE TPAMI, 39(3):589-602, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 512 + ], + "type": "text", + "content": "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[48] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking. arXiv preprint arXiv:2211.08824, 2023. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "text", + "content": "[49] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "text", + "content": "[50] Hefeng Wu, Yafei Hu, Keze Wang, Hanhui Li, Lin Nie, and Hui Cheng. Instance-aware representation learning and association for online multi-person tracking. PR, 94:25-34, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 629, + 287, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 629, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 48, + 629, + 287, + 671 + ], + "type": "text", + "content": "[51] Xingjiao Wu, Baohan Xu, Yingbin Zheng, Hao Ye, Jing Yang, and Liang He. Fast video crowd counting with a temporal aware network. Neurocomputing, 403:13-20, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 672, + 287, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 672, + 287, + 705 + ], + "spans": [ + { + "bbox": [ + 48, + 672, + 287, + 705 + ], + "type": "text", + "content": "[52] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 329 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[53] Feng Xiong, Xingjian Shi, and Dit-Yan Yeung. Spatiotemporal modeling for crowd counting in videos. In ICCV, pages 5161-5169, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "type": "text", + "content": "[54] Bo Yang, Chang Huang, and Ram Nevatia. Learning affinities and dependencies for multi-target tracking using a CRF model. In CVPR, pages 1233-1240, 2011. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 137, + 545, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 168 + ], + "type": "text", + "content": "[55] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE TPAMI, 43(1):360-374, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 168, + 545, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 168, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 307, + 168, + 545, + 212 + ], + "type": "text", + "content": "[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129:3069–3087, 2021. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 212, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 244 + ], + "type": "text", + "content": "[57] Zhuoyi Zhao, Hongsheng Li, Rui Zhao, and Xiaogang Wang. Crossing-line crowd counting with two-phase deep neural networks. In ECCV, pages 712-726, 2016. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 244, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 286 + ], + "type": "text", + "content": "[58] Huicheng Zheng, Zijian Lin, Jiepeng Cen, Zeyu Wu, and Yadan Zhao. Cross-line pedestrian counting based on spatially-consistent two-stage local crowd density estimation and accumulation. IEEE TCSVT, 29(3):787-799, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[59] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE TPAMI, 44(11):7380-7399, 2021. 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19237" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_content_list.json b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6cf729f3be5d53d2a9431336692b48198c564a2 --- /dev/null +++ b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kranthi Kumar Rachavarapu IIT Madras", + "bbox": [ + 140, + 204, + 370, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kranthi.rachavarapu@gmail.com", + "bbox": [ + 127, + 242, + 385, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kalyan Ramakrishnan \nUniversity of Oxford", + "bbox": [ + 442, + 204, + 620, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kalyanr@robots.ox.ac.uk", + "bbox": [ + 428, + 242, + 633, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rajagopalan A. N. IIT Madras", + "bbox": [ + 684, + 204, + 831, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "raju@ee.iitm.ac.in", + "bbox": [ + 676, + 242, + 836, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we address the weakly-supervised Audio-Visual Video Parsing (AVVP) problem, which aims at labeling events in a video as audible, visible, or both, and temporally localizing and classifying them into known categories. This is challenging since we only have access to video-level (weak) event labels when training but need to predict event labels at the segment (frame) level at test time. Recent methods employ multiple-instance learning (MIL) techniques that tend to focus solely on the most discriminative segments, resulting in frequent misclassifications. Our idea is to first construct several \"prototype\" features for each event class by clustering key segments identified for the event in the training data. We then assign pseudo labels to all training segments based on their feature similarities with these prototypes and re-train the model under weak and strong supervision. We facilitate this by structuring the feature space with contrastive learning using pseudo labels. Experiments show that we outperform existing methods for weakly-supervised AVVP. We also show that learning with weak and iteratively re-estimated pseudo labels can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure.", + "bbox": [ + 75, + 325, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 207, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the problem of weakly-supervised Audio-Visual Video Parsing (AVVP) [48], where the goal is to classify events in a video and localize them over time and modalities, given only video-level (weak) event labels for training. Such a formulation is attractive since it forgoes the need for expensive and tedious fine-grained labeling. Moreover, other weakly-supervised video-based temporal prediction tasks can be posed as special cases of AVVP. For instance, Audio-Visual Event Localization (AVEL) [47] considers events that are simultaneously audible and visible, whereas Temporal Action Localization (TAL) [10] considers only visible events. AVVP is challenging as an event", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c60c9561d79fd0886536a15974b8956a7f3bae7f9a2fd150691401d3c3d834a7.jpg", + "image_caption": [ + "Figure 1. Example predictions of a model trained using MIL under weak supervision [48] and our method. For clarity, we only show visual events. The MIL model detects the event \"Guitar\" correctly in the beginning, but there are some mistakes (red dotted boxes) later on. Moreover, although the last three frames are similar, they are labeled inconsistently by the model. Our method consistently localizes well. \"GT\" is the ground truth." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 883, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "could occur in just the audio stream, visual stream, or both at the same time. Audio-visual data is ubiquitous, with highly variable events in terms of appearance and duration.", + "bbox": [ + 496, + 598, + 890, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods for weakly-supervised AVVP [30, 48] use multiple-instance learning (MIL) [7] techniques that optimize for video-level labels during training and make frame-level predictions during evaluation. Here, video-level predictions are obtained by pooling (aggregating) the frame-level predictions. A drawback of MIL approaches is that a model may learn to focus solely on the most discriminative (\"prominent\") instances (here, audio/visual frames in a video) and still minimize the training loss successfully. This means such models may not reliably detect the full temporal extent of events. Fig. 1 shows an example where an MIL model [48] for AVVP fails to localize events accurately. Moreover, the model labels frames that are visually similar differently. Another reason for misclassifications in MIL models is that they typically contain a frame-level classifier with a single weight vector per event class. However, this may be insufficient to capture the intra-class variation", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18952", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for the event across time, modalities, and examples while training with weak labels.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to address these issues with the following ideas. First, given that the MIL model can find discriminative frames that trigger the video-level labels, we can identify such frames and propagate their predicted labels to other frames in the training data based on similarities in the feature space, thus creating frame-level pseudo labels on the training data. Second, we can use their features to construct multiple prototypes per event class, ensuring a more robust representation to propagate pseudo labels with. Once this is done for the training data, we can re-train the model with strong supervision from pseudo labels and weak supervision from video-level labels. Fig. 1 shows an example where our method produces more accurate predictions than the MIL model. Because our method relies on feature similarities, we further enforce structure in the feature space via contrastive learning using pseudo labels.", + "bbox": [ + 75, + 121, + 468, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We re-estimate pseudo labels on the training data after every epoch and train with them over the next epoch. We show that such a training procedure can be viewed through an expectation-maximization (EM) lens with the unknown frame-level labels as latent variables. This interpretation provides convergence guarantees and further insight. Our method achieves state-of-the-art performance for weakly-supervised AVVP and also works well on the related weakly-supervised Temporal Action Localization (TAL) task. We summarize our contributions below:", + "bbox": [ + 75, + 363, + 468, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a prototype-based pseudo-labeling method for weakly-supervised AVVP that exploits similarities in feature space.", + "2. We show how to perform contrastive learning using pseudo labels to help semantically structure the feature space.", + "3. We present an expectation-maximization view of our general training procedure.", + "4. We show state-of-the-art results for weakly-supervised AVVP, and on several metrics for weakly-supervised TAL." + ], + "bbox": [ + 76, + 513, + 468, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 693, + 218, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-Visual Video Parsing (AVVP). Tian et al. [48] formulated AVVP as a multimodal multiple-instance learning (MMIL) problem and proposed a hybrid attention network with a learned pooling function (HAN) to capture unimodal and cross-modal contexts. Most subsequent work has used the HAN architecture with other techniques to improve performance. Lamba et al. [30] utilized cross-modal information to learn better representations with adversarial and self-supervised losses. Wu et al. [52] estimated modality-level labels by exchanging the audio/visual streams between pairs of videos, followed by retraining under stronger supervision. Lin et al. [32] used cross-video and cross-modality", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "signals, such as event co-occurrence across modalities, as additional sources of supervision. Mo et al. [36] leveraged multi-modal grouping to learn discriminative subspaces. Gao et al. [11] extract the \"presence\" and \"absence\" evidence for events from uni- and cross-modal information. While Fu et al. [9] focused on learning rate imbalance between modalities and Zhou et al. [60] utilized CLIP [41] to get segment-labels, our work leverages in-model feature similarities for pseudo labeling. PoiBin [40] modeled the number of positive frames for an event as a latent variable to improve localization. Different from previous methods, we estimate segment-level pseudo labels and re-train under full supervision.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pseudo-Labeling refers to estimating labels for unlabeled data using the predictions of a trained model. This has been used to improve performance on several weakly-supervised tasks including object detection [1, 46, 59] and image classification [2, 13, 21]. A few works [34, 39, 58] have employed pseudo-labeling to improve performance on Temporal Action Localization (TAL), generating labels from model outputs or attention weights. Luo et al. [34] posed the TAL task as expectation-maximization, using pseudolabels to approximate the E and M steps. Contrary to these, we propose a pseudo-labeling strategy using feature similarities between segments guided by the model predictions and weak labels for a video.", + "bbox": [ + 496, + 287, + 890, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prototype-based Classification. Prototype learning with a nearest-neighbor (non-parametric) classifier has inspired various machine learning techniques [5, 12, 15, 43, 44]. Recently, there has been increased interest in integrating this into deep learning methods, e.g., zero-shot [25], few-shot [45], unsupervised [54, 55], and supervised [16, 35, 53, 56] learning. Prototype learning has also been widely used for image segmentation [8, 50, 51, 61]. Recently, prototype-based methods have also been used for semi-supervised image [37] and text [57] classification. A few works have used prototype-based methods with multiple-instance learning. E.g., Rymarczyk et al. [42] processed large medical images as a bag of patches, and Huang et al. [22] learned prototypes for the TAL task with graph convolutional networks. These works adopt an embedding-level MIL approach ([24]) with one prototype per class, while we adopt instance-level MIL and construct multiple prototypes per event class.", + "bbox": [ + 496, + 486, + 890, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 760, + 627, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 785, + 681, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the audio-visual video parsing (AVVP) problem, we aim to identify audible or visible events in an unconstrained video and simultaneously localize them in time. Specifically, given a video $x$ consisting $T$ non-overlapping temporal segments $\\{x_{t} = (x_{t}^{a},x_{t}^{v})\\}_{t = 1}^{T}$ of the audio $(A)$ and visual $(V)$ streams, our objective is to classify each segment", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into $C$ possible events (e.g., singing, vehicle moving, etc.). At test time, we need to predict segment-level event labels $y_{t} = (\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}}, \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})$ for each audio and visual segment, where $\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}}, \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}} \\in \\{0,1\\}^{C}$ are segment-level audio and visual event labels, respectively.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weak Supervision. In weakly-supervised AVVP, for each video $x$ , we only have access to the corresponding video-level event label vector $\\mathbf{w} = (w_{1},\\dots ,w_{C})\\in \\{0,1\\}^{C}$ where $w_{c} = 1$ if any of the segments in the video contains the $c$ -th event, and $w_{c} = 0$ otherwise. Such weak, video-level labels only indicate whether an event occurs in the video or not. During evaluation, we still need to predict segment-level labels $\\{(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}},\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\}_{t = 1}^{T}$ . Note that multiple events can simultaneously occur in a video, i.e., $\\sum_{c}w_{c}\\geq 1$ , and each segment could have more than one event.", + "bbox": [ + 75, + 167, + 467, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. MIL Framework", + "text_level": 1, + "bbox": [ + 76, + 325, + 243, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt an instance-level MIL approach [24] as done by prior work [30, 32, 36, 48, 52]: First, each instance (here, an audio/visual segment) is assigned a classification score (here, a $C$ -dimensional vector). Then, the scores are aggregated by a pooling operation, and the aggregate is used in the loss. In our case, this looks like", + "bbox": [ + 75, + 349, + 467, + 439 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ \\left\\{\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {a}} \\right\\} _ {t = 1} ^ {T}, \\left\\{\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {v}} \\right\\} _ {t = 1} ^ {T} \\right] = f _ {\\theta} (x) (1) \\\\ \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} = g _ {\\phi} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}}\\right) \\in [ 0, 1 ] ^ {C}, \\forall t \\in [ 1, T ], m \\in \\{a, v \\} (2) \\\\ \\hat {\\mathbf {w}} = \\sigma_ {\\theta , x} \\left(\\left\\{\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {a}} \\right\\} _ {t = 1} ^ {T}, \\left\\{\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {v}} \\right\\} _ {t = 1} ^ {T}\\right) \\in [ 0, 1 ] ^ {C}, (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 450, + 467, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\theta}$ is a feature extractor (encoder), $g_{\\phi}$ is a feature classifier, $\\sigma_{\\theta,x}$ is a MIL-pooling operator. The predicted video-level probability vector $\\hat{\\mathbf{w}}$ is computed from segment-level classification scores $\\{\\mathbf{p}_t^m\\}$ , which are computed from segment features $\\{\\mathbf{f}_{\\mathbf{t}}^{\\mathbf{m}}\\}$ . We adopt the Hybrid Attention Network (HAN) [48] architecture to implement $f_{\\theta}, g_{\\phi}$ , and $\\sigma$ . Here, $f_{\\theta}$ consists of a transformer-based encoder with cross-modal attention and $g_{\\phi}$ is a linear classifier with the sigmoid activation. Attention-based averaging is used for $\\sigma$ . We provide implementation details in the Supplementary.", + "bbox": [ + 75, + 521, + 467, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This model is trained end-to-end with supervision at the video level by minimizing a weighted cross-entropy loss", + "bbox": [ + 76, + 686, + 467, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M I L}} = \\operatorname {C E} (\\hat {\\mathbf {w}}, \\mathbf {w}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 729, + 467, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with weights inversely proportional to the event distribution in the dataset. At test time, we get segment-level predictions from Eq. (2).", + "bbox": [ + 75, + 756, + 467, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Prototype-based Pseudo-Labeling", + "text_level": 1, + "bbox": [ + 76, + 814, + 382, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While instance-level MIL enables finding key segments that trigger the video-level labels [24, 33], they often miss out on the full temporal extent of events in a video. However, if we identify such discriminative segments for each label", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in a training video, we could use them to build a set of prototype features that represent each class. A high degree of similarity of a segment with the prototype features for an event class means the segment probably contains the event. We use this idea to generate pseudo labels for all segments in a training video and then re-train the model under strong supervision. Fig. 2 shows an overview of our approach. We provide pseudocode in Algorithm 1.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Constructing Prototype Features", + "text_level": 1, + "bbox": [ + 498, + 220, + 785, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The segment-level linear classifier $g_{\\phi}$ (Eq. (2)) can itself be considered a prototype-based classifier, where the prototype for each class is the corresponding weight vector in $\\phi$ . However, this requires that an event have similar representations across modalities and examples in the training data. We instead construct multiple prototypes for each class in the following two steps.", + "bbox": [ + 496, + 244, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segment Selection. For each training video $x$ , we find positive segments $\\mathcal{P}_x(c)$ for a target event $c$ (if $w_c = 1$ ) as those the model is highly confident contain the event. I.e.,", + "bbox": [ + 496, + 349, + 890, + 395 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {x} (c) = \\left\\{ \\begin{array}{l l} \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) \\geq \\gamma p ^ {*} (c) \\right\\} & \\text {i f} w _ {c} = 1 \\\\ \\phi & \\text {i f} w _ {c} = 0, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 406, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma \\in (0,1]$ is a hyperparameter and $p^*(c) = \\max_{t,m} \\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}(c)$ is the maximum predicted probability for the event over time and modalities for a given video. Clearly, $\\mathcal{P}_X(c)$ will contain at least one audio/visual segment from video $x$ when $w_c = 1$ . We assume these segments contain event $c$ with high probability (so are likely true positives). We also find negative segments $\\mathcal{N}_x(c)$ for a target event $c$ as those the model is highly confident do not contain the event when $w_c = 1$ (likely true negatives), or is highly confident to contain the event when $w_c = 0$ (definite false positives):", + "bbox": [ + 496, + 455, + 890, + 622 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} _ {x} (c) = \\left\\{ \\begin{array}{l l} \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) \\leq \\beta p ^ {*} (c) \\right\\} & \\text {i f} w _ {c} = 1 \\\\ \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = p ^ {*} (c) > = 0. 5 \\right\\} & \\text {i f} w _ {c} = 0, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 633, + 890, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta \\in (0,1)$ is a hyperparameter set to a small value to retrieve segments without the event $c$ (if any). Note that $\\mathcal{N}_x(c)$ may contain no segments at all even when $w_{c} = 1$ , i.e., be a null set.", + "bbox": [ + 496, + 685, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we aggregate the positive and negative sets over all videos in training data $\\mathcal{D}$ to get prototypical segments as", + "bbox": [ + 500, + 744, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} (c) = \\bigcup_ {x \\in \\mathcal {D}} \\mathcal {P} _ {x} (c), \\quad \\mathcal {N} (c) = \\bigcup_ {x \\in \\mathcal {D}} \\mathcal {N} _ {x} (c). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 787, + 890, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype Generation. We now generate multiple prototypes (feature vectors) for each event class by hypothesizing that having more than one prototype would help better capture the intra-class variation across modalities and training examples, and create a more robust representation.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18954", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af2047dd15287e6043608b15af802cf7050fae0e8f2582d8bc37a28b035c7955.jpg", + "image_caption": [ + "Figure 2. An overview of our method. The top branch is trained using MIL with weak labels and using full supervision with pseudo labels, which are generated by the bottom branch. This is done by first selecting key segments for each event from the training data and clustering them to construct a set of prototype features for the event, used to guide pseudo-label generation." + ], + "image_footnote": [], + "bbox": [ + 125, + 85, + 790, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we extract features for all segments in $\\mathcal{P}(c)$ and $\\mathcal{N}(c)$ using Eq. (1), and call the resulting feature sets $\\mathcal{F}_{\\mathcal{P}}(c)$ and $\\mathcal{F}_{\\mathcal{N}}(c)$ . Here, instead of $f_{\\theta}$ , we use a momentum encoder [19] $f_{\\tilde{\\theta}}$ having the same architecture but whose weights are set (after each update to $\\theta$ ) as an exponential moving average (EMA) of $\\theta$ over training steps: $\\tilde{\\theta} \\gets (1 - \\eta)\\tilde{\\theta} + \\eta\\theta$ , where $\\eta \\in (0,1)$ . The momentum encoder is more stable and less sensitive to noise from weak supervision, leading to slightly improved performance.", + "bbox": [ + 75, + 356, + 472, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we perform $k$ -means clustering separately on $\\mathcal{F}_{\\mathcal{P}}(c)$ and $\\mathcal{F}_{\\mathcal{N}}(c)$ to get two sets of cluster centroids. We call the centroids prototypes:", + "bbox": [ + 76, + 497, + 470, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {\\mathcal {P}} (c) = \\left\\{\\mathbf {f} _ {\\mathbf {i}} ^ {+} \\right\\} _ {i = 1} ^ {k _ {p}}, \\quad \\mathcal {C} _ {\\mathcal {N}} (c) = \\left\\{\\mathbf {f} _ {\\mathbf {j}} ^ {-} \\right\\} _ {j = 1} ^ {k _ {n}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 564, + 468, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clustering for finding prototypes imposes a natural bottleneck [26], helping discard irrelevant information in the segments. We repeat this for each event class.", + "bbox": [ + 75, + 604, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Estimating Pseudo Labels", + "text_level": 1, + "bbox": [ + 76, + 675, + 313, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assign pseudo labels to each segment in a training video by considering its feature similarities with the prototypes for each class. We assume that if an audio/visual segment contains event $c$ , it would be on average closer to $\\mathcal{P}(c)$ in feature space than to $\\mathcal{N}(c)$ . For this, we explore two approaches: (i) hard labeling, which assigns the label of the closest-on-average set among the prototype sets for class $c$ :", + "bbox": [ + 75, + 704, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = w _ {c} \\cdot \\mathbb {1} \\left[ \\frac {\\sum_ {\\mathbf {f} ^ {+} \\in \\mathcal {C} _ {\\mathcal {P}} (c)} e ^ {(\\mathbf {f} ^ {+}) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}}}}{| \\mathcal {C} _ {\\mathcal {P}} |} > \\frac {\\sum_ {\\mathbf {f} ^ {-} \\in \\mathcal {C} _ {\\mathcal {N}} (c)}}{| \\mathcal {C} _ {\\mathcal {N}} |} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 840, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and (ii) soft labeling, which instead assigns a probability score:", + "bbox": [ + 500, + 356, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = w _ {c} \\cdot \\frac {\\frac {1}{| \\mathcal {C} _ {\\mathcal {P}} |} \\sum_ {\\mathbf {f} ^ {+} \\in \\mathcal {C} _ {\\mathcal {P}} (c)} e ^ {\\left(\\mathbf {f} ^ {+}\\right) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}} / \\tau}}{\\sum_ {\\mathcal {S} \\in \\left\\{\\mathcal {C} _ {\\mathcal {P}}, \\mathcal {C} _ {\\mathcal {N}} \\right\\}} \\frac {1}{| \\mathcal {S} |} \\sum_ {\\mathbf {f} ^ {\\prime} \\in \\mathcal {S} (c)} e ^ {\\left(\\mathbf {f} ^ {\\prime}\\right) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}} / \\tau}} \\in (0, 1), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 387, + 890, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau > 0$ is a temperature hyperparameter.", + "bbox": [ + 500, + 464, + 805, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When training our model, we generate pseudo labels $\\{\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\}$ after every epoch for each training video and use them to fully supervise the segment-level probabilities $\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\}$ in Eq. (2) using the binary cross-entropy loss", + "bbox": [ + 500, + 479, + 890, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P L}} = \\operatorname {C E} \\left(\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}}, \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 545, + 890, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "averaged over segments and training examples.", + "bbox": [ + 500, + 566, + 812, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Contrastive Learning with Pseudo Labels", + "text_level": 1, + "bbox": [ + 500, + 589, + 854, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that our method exploits feature similarities between segments based on the events they may contain, we use contrastive learning based on pseudo labels to help semantically structure the feature space. Since multiple events can occur simultaneously, for each audio/visual segment, we consider a similar segment as one estimated to have some common event(s) and a dissimilar segment as one estimated to have no common events. Formally, given a batch of training examples, let $B = \\{\\mathbf{f}_{\\mathbf{t}}^{\\mathbf{m}} \\to \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\}$ be a dictionary from features to the assigned pseudo labels for all segments and over all videos in the batch. For every feature $\\mathbf{f} \\in B$ , the similar feature-set is then $S(\\mathbf{f}) = \\{\\mathbf{g} \\in B \\mid B(\\mathbf{g}) \\cap B(\\mathbf{f}) \\neq \\phi\\}$ and the dissimilar set is $D(\\mathbf{f}) = \\{\\mathbf{g} \\in B \\mid B(\\mathbf{g}) \\cap B(\\mathbf{f}) = \\phi\\}$ . To pull segments with a common predicted event closer in the feature space, we use the contrastive loss [38]", + "bbox": [ + 496, + 612, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L}} = - \\log \\frac {\\sum_ {\\mathbf {g} \\in \\mathcal {S} (\\mathbf {f})} e ^ {\\mathbf {g} ^ {T} \\mathbf {f} / \\tau_ {c}}}{\\sum_ {\\mathbf {g} \\in \\mathcal {S} (\\mathbf {f})} e ^ {\\mathbf {g} ^ {T} \\mathbf {f} / \\tau_ {c}} + \\sum_ {\\mathbf {h} \\in \\mathcal {D} (\\mathbf {f})} e ^ {\\mathbf {h} ^ {T} \\mathbf {f} / \\tau_ {c}}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 843, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/31193cd3a085fd82a89a5e246a2b1e19338e255ca60b9581e709aeb257553b04.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
AVE* [47]47.240.437.134.735.431.639.935.541.636.5
AVSDN* [31]47.834.152.046.337.126.545.735.650.837.7
HAN [48]60.151.352.948.948.943.054.047.755.448.0
MA [52]60.353.660.056.455.149.058.953.057.950.6
CVCM-MA [32]60.853.863.558.957.049.560.554.059.552.1
MGN-MA [36]60.250.961.959.755.549.659.253.458.749.9
JoMoLD [4]61.353.963.859.957.249.660.854.559.952.5
PoiBin [40]63.154.163.560.357.751.561.455.260.652.3
CMPAE [11]64.256.666.463.759.251.863.357.462.855.7
Ours65.957.366.764.361.954.364.859.963.757.9
", + "bbox": [ + 151, + 88, + 820, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results (% F1-scores) on all metrics defined in Tian et al. [48]. \\* indicates reproduced performance. The best and second best results are in bold and underline, respectively.", + "bbox": [ + 75, + 291, + 892, + 320 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Pseudocode for our method" + ], + "code_body": "1 Init: model params $\\psi = \\{\\theta ,\\phi \\}$ , EMA params $\\psi_{E} = \\{\\theta_{E},\\phi_{E}\\}$ \n2 while $\\psi$ has not converged do \n3 #estimate segment - level pseudo labels \n4 for $(x,\\mathbf{w})$ in train set do \n5 $\\left|\\{\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\gets p_{\\theta}(\\cdot |\\mathbf{w},x)\\right.;$ \n6 end \n7 #train \n8 for $(x,\\mathbf{w})$ in train set do \n9 $\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\gets p_{\\theta ,\\phi}(\\cdot |x);$ \n10 $\\hat{\\mathbf{w}} = \\sigma_{\\theta ,x}(\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\});$ \n11 $\\mathcal{L} = \\mathcal{L}_{\\mathrm{MIL}} + \\lambda_{\\mathrm{PL}}\\mathcal{L}_{\\mathrm{PL}} + \\lambda_{\\mathrm{CL}}\\mathcal{L}_{\\mathrm{CL}};$ \n12 $\\psi \\gets \\psi -\\alpha \\nabla_{\\psi}\\mathcal{L};$ \n13 $\\psi_{E}\\gets (1 - \\eta)\\psi_{E} + \\eta \\psi ;$ \n14 end \n15 end", + "bbox": [ + 78, + 353, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "averaged over all $\\mathbf{f} \\in B$ , where $\\tau_{c} > 0$ is a temperature hyperparameter.", + "bbox": [ + 75, + 641, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 694, + 290, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model using the following loss in an end-to-end fashion:", + "bbox": [ + 75, + 722, + 468, + 751 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M I L}} + \\lambda_ {\\mathrm {P L}} \\mathcal {L} _ {\\mathrm {P L}} + \\lambda_ {\\mathrm {C L}} \\mathcal {L} _ {\\mathrm {C L}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 773, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{PL}}, \\lambda_{\\mathrm{CL}} > 0$ are hyperparameters. During training, we generate pseudo labels once every epoch for all training examples using the best model so far. During inference, we predict the segment-level probabilities using the non-parametric prototypical pseudo-labeling (Eq. (10)) with video-level predictions from Eq. (3).", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 330, + 630, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 356, + 581, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLP Dataset. We conduct experiments on the Look, Listen and Parse (LLP) dataset [48] which consists of 11849 YouTube videos, each of 10s duration, labeled into 25 event categories. These videos are unconstrained with a wide variety of scene content, including daily activities, music performances, and vehicle sounds. We use 10000 videos with weak (video-level) labels for training. The remaining 1849 fully-annotated videos (with segment-level labels) are used for validation and testing. The train/val/test split has been provided in the LLP dataset.", + "bbox": [ + 496, + 378, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous work, we use F1-scores on audio, visual, and audio-visual events as evaluation metrics. These are computed both at the segment and event level. We also include the aggregate metrics \"Type@AV\" and \"Event@AV\", again computed at the segment and event level. See Tian et al. [48] for a full explanation of metrics.", + "bbox": [ + 496, + 531, + 890, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use ResNet-152 [18] pretrained on ImageNet [6] and R(2+1)D-18 [49] pre-trained on Kinetics-400 [27] as visual feature extractors to generate a 512-d visual feature per segment. We use VGGish [20] pre-trained on AudioSet [14] to generate a 128-d audio feature per segment. We use the Adam optimizer [28] with a batch size of 16 and a learning rate $\\alpha = 3e - 4$ for 50 epochs. We take $\\eta = 0.999$ for the momentum encoder, $\\lambda_{\\mathrm{PL}} = \\lambda_{\\mathrm{CL}} = 0.1$ in Eq. (13), $\\gamma = 1$ in Eq. (5), $\\beta = 0.2$ in Eq. (6), $\\tau = 0.1$ in Eq. (10), and $\\tau_{c} = 0.2$ in Eq. (12). We set $k_{p} = k_{n} = 10$ in Eq. (8). Hyperparameter values are based on validation performance.", + "bbox": [ + 496, + 638, + 890, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparisons with Prior Work", + "text_level": 1, + "bbox": [ + 500, + 830, + 769, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fair comparison, all methods considered use the same pre-trained feature extractors and train/val/test split.", + "bbox": [ + 496, + 854, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative. We compare methods in Table 1 on all the", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18956", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e504539dcc566fdaa74ed92febaad0af3d6b78d86b3045246478c4fb7b6d264.jpg", + "image_caption": [ + "Figure 3. Audio-Visual Video Parsing results of our methods with HAN [48], MA [52], JoMoLD [4] and CMPAE [11] on two videos." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 477, + 391 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ab05ab4e5bc3d13898ad66819dcb2888e7b518db9b16f2918442eb6334277a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 146, + 880, + 390 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9a7911985318fc4b46e860f4ae7da95a612b8ce74820e1f272259cbd21b9fcc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
Base+NPLH58.952.158.549.855.748.257.850.155.746.6
Base+NPILS59.352.559.051.156.048.758.250.856.247.6
Base+PPLH62.557.462.959.659.654.061.956.360.355.7
Base+PPLS65.957.366.764.361.954.364.859.963.757.9
", + "bbox": [ + 81, + 425, + 467, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance for different pseudo labeling strategies.", + "bbox": [ + 89, + 517, + 454, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metrics listed earlier. We see that our method outperforms others on audio, audio-visual, Type@AV, and Event@AV scores at both the segment and event levels. In particular, we achieve an average improvement of 2.6 points for audio-visual events. Averaged over all metrics, our method achieves a 1.86 point improvement over the previous best method CMPAE [11]. Since we do not require additional learnable parameters to generate pseudo labels, this improvement is with a comparable number of parameters to previous methods.", + "bbox": [ + 75, + 550, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative. We qualitatively compare our method with some previous methods on two examples in Fig. 3. Here, Fig. 3 (left) contains events \"Singing\" and \"Banjo\" in both modalities. Fig. 3 (Right) shows another scene with the event \"Speech\" in only the audio modality, and \"Violin\" occurring in both modalities. As evident, the common failure mode of previous methods is that they detect only a portion of the duration of events, while ours more commonly detects the full extent of events. As discussed in Sec. 1, this could be because of multiple-instance learning encouraging the model to pick out only the most discriminative segments containing an event, as opposed to all segments.", + "bbox": [ + 75, + 704, + 468, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computational Cost. When training, our approach takes", + "bbox": [ + 76, + 886, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "67.4s per epoch (17.1s for forward+backward propagation and 50.3s for prototype-based pseudo labeling). Inference time is 5.9ms on average per video, which is comparable to the baselines (2.9ms).", + "bbox": [ + 498, + 428, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 500, + 498, + 601, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pseudo-Labeling Strategies. We report an ablation study in Table 2 to validate our specific pseudo-labeling method. Here, Base is the baseline MIL model explained in Sec. 3.2. Base+NPL refers to the base model trained with naive pseudo labeling, i.e., by directly using the segment-level predictions of the model as pseudo labels for retraining. In this case, H and S refer to using hard (binarized) and soft (continuous) pseudo labels, respectively. We observe no consistent improvement with naive pseudo labeling, likely due to the MIL model producing incorrect predictions for the non-discriminative segments for an event, leading to less reliable pseudo labels. Base+PPL refers to the base model trained with prototype-based pseudo labels as described in Sec. 4.2. In this case, H and S refer to using hard and soft pseudo labels, as defined in Eqs. (9) and (10), respectively. Our method significantly improves performance by generating more reliable pseudo labels throughout a training video. We also see that soft labeling is better than hard labeling. Since segment-level predictions may be noisy, we expect mistakes to affect performance more with hard rather than soft labels.", + "bbox": [ + 496, + 522, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reliability of Pseudo Labels. We now assess if the pseudo labels generated by our method are reliable for re-training. For this, we take the pre-trained baselines HAN [48] and MA [52], generate prototypes for each event from the train-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6cf0e70fde01999da2897c2e1d22efaec43694c25ecf60ff317077738b9f814c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
HAN60.151.352.948.948.943.054.047.755.448.0
+PPLTest62.555.455.351.152.346.956.050.958.350.6
MA60.353.660.056.455.149.058.953.057.950.6
+PPLTest61.755.461.857.957.551.660.655.059.452.6
", + "bbox": [ + 81, + 89, + 467, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb7db5efc731b7e1a61941d5d00977d2b01df0930cf6536e7539110cc1ca8d8c.jpg", + "image_caption": [ + "Figure 4. Performance for different $\\gamma$ and $\\beta$ choices." + ], + "image_footnote": [], + "bbox": [ + 83, + 209, + 261, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/280184cfa4828bba9fb6a9b25a9cf6dde190c3fe77f86a6adab73f7476b677a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 208, + 449, + 324 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a73e3ce0e950aa6dc663a597d3ab2e34270a1df6c6a79eb14188d98497aca51.jpg", + "table_caption": [ + "Table 3. Evaluating prototype-based pseudo labels on the test set." + ], + "table_footnote": [], + "table_body": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
No-CL64.156.459.657.357.149.661.256.460.856.8
With-CL65.957.366.764.361.954.364.859.963.757.9
", + "bbox": [ + 81, + 359, + 467, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing data, and simply predict the prototype-based pseudo labels (Eq. (9)) on the test set for evaluation (i.e., discard the classifiers at test time). We call the resulting classifiers $\\mathrm{HAN} + \\mathrm{PPL}_{\\mathrm{Test}}$ and $\\mathrm{MA} + \\mathrm{PPL}_{\\mathrm{Test}}$ and show results in Table 3. With no other change to the models other than the classifiers, we see significant improvements in all metrics, showing that our pseudo labels are indeed reliable estimates of the true labels. Note that for the results in Table 1, we recompute pseudo labels on the training data every epoch, with a continually improving base model.", + "bbox": [ + 75, + 460, + 467, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of $\\gamma$ and $\\beta$ . The hyperparameters $\\gamma$ and $\\beta$ in Eqs. (5) and (6) roughly control the proportion of segments that will be selected as positives and negatives for assigning pseudo labels. We experiment with different values and plot the average of all segment/event-level metrics in Fig. 4. Performance is susceptible to $\\gamma$ , which controls positives, and is tolerant to a slight variation in $\\beta$ . The best performance is for $\\gamma = 1$ and $\\beta = 0.2$ .", + "bbox": [ + 75, + 612, + 467, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Contrastive Learning. We analyze the impact of contrastive learning using pseudo labels (Eq. (12)) by simply setting $\\lambda_{\\mathrm{CL}} = 0$ in Eq. (13). Results are reported in Tab. 4. While only training with prototype-based pseudo labels $(\\mathsf{No - CL})$ significantly improves the baseline, adding a contrastive loss based on our pseudo labels (With-CL) further yields a large improvement.", + "bbox": [ + 75, + 734, + 467, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Clusters. We take $k_{p} = k_{n}$ and try values in $\\{1,2,\\dots,100\\}$ . We plot the average of all segment/event-level metrics in Fig. 5. The best performance was for $k = 10$ clusters per class. Having fewer clusters helps cap", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57b6c0c53107decadc8ad13f2930dc7c42130d5eef81b2fd1eeb06c059496e3f.jpg", + "image_caption": [ + "Figure 5. Performance for different choices of the number of clusters/prototypes per event." + ], + "image_footnote": [], + "bbox": [ + 516, + 93, + 859, + 226 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2a441e86dadc310039b2e7807e15d63822f7635fa0eebf4e2e17b4fcbd111f5e.jpg", + "table_caption": [ + "Table 4. Effect of contrastive learning (CL) with pseudo labels." + ], + "table_footnote": [], + "table_body": "
MethodtIoUAvg.
0.10.20.30.40.50.60.7[0.1:0.5][0.3:0.7][0.1:0.7]
EM-MIL [34]59.152.745.536.830.522.716.444.930.437.7
ASM-Loc [17]71.265.557.146.836.625.213.455.435.845.1
DELU [3]71.566.256.547.740.527.215.356.537.446.4
Ours72.766.957.946.937.426.820.156.437.846.9
", + "bbox": [ + 503, + 281, + 888, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results (\\% mAP) for the TAL task at different tIoU thresholds.", + "bbox": [ + 500, + 361, + 890, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ture more general features for an event instead of instance-specific details, but having too few hurts performance by discarding intra-class variation.", + "bbox": [ + 498, + 417, + 890, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results for a Different Task", + "text_level": 1, + "bbox": [ + 500, + 473, + 746, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct preliminary experiments for the weakly-supervised Temporal Action Localization (TAL) task [10] on the THUMOS14 dataset [23] to further validate our method. TAL aims to temporally localize and classify visual events in videos into known categories. We report mean average precisions (mAP) at different temporal intersection over union (tIoU) thresholds and the results are in Table 5. We achieve better or comparable performance w.r.t. the current state-of-the-art method DELU [3], and outperform previous methods. Check the supplementary for a full comparison.", + "bbox": [ + 498, + 497, + 890, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Interpretation as Expectation-Maximization", + "text_level": 1, + "bbox": [ + 500, + 678, + 890, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now show that iteratively re-estimating soft pseudo labels and training the model under weak $+$ strong supervision (Eq. (13)) can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure. For this, we view the segment-level labels $y = \\{(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}},\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\}_{t = 1}^{T}$ for each video as latent variables since they are not observed when training under weak supervision. The observed variables are the video segments $x = \\{(x_{t}^{a},x_{t}^{v})\\}_{t = 1}^{T}$ and weak labels $\\mathbf{w}$ . For clarity, assume a single event class, i.e., $\\mathbf{w},\\mathbf{y}_{\\mathbf{m}}^{\\mathbf{t}}\\in \\{0,1\\}$ .", + "bbox": [ + 498, + 703, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our goal is to maximize the data log-likelihood $\\mathcal{L}(\\theta) = \\ln p_{\\theta}(\\mathbf{w}|x)$ over model parameters $\\theta$ (assume just one data point $x$ for clarity). Instead of maximizing this directly, we", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18958", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can attempt to maximize the variational lower bound [29] $\\mathcal{L}_{\\mathrm{VLB}}$ for this likelihood (random variables are capitalized):", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V L B}} (\\theta , \\phi) = \\underset {Y \\sim q _ {\\phi} (\\cdot | \\mathbf {w}, x)} {\\mathbb {E}} [ \\ln p _ {\\theta} (\\mathbf {w} | Y, x) ] \\\\ - \\mathrm {D} _ {\\mathrm {K L}} \\left[ q _ {\\phi} (\\cdot | \\mathbf {w}, x) \\| p _ {\\theta} (\\cdot | x) \\right], \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 148, + 468, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $q_{\\phi}(y|\\mathbf{w},x)$ is any conditional distribution over labels $y$ , with parameters $\\phi$ , meant to approximate the true posterior $p_{\\theta}(y|x)$ . The EM algorithm alternates between two steps, first approximating the posterior by maximizing $\\mathcal{L}_{\\mathrm{VLB}}$ over $\\phi$ (E step), and then improving the data likelihood by maximizing it over $\\theta$ (M step). Let $j \\geq 1$ index the iterations (note, $\\theta$ and $\\phi$ here are not the same as in Sec. 3.2).", + "bbox": [ + 76, + 205, + 468, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "E Step. The $j$ -th update is given by", + "bbox": [ + 76, + 311, + 315, + 325 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\phi_ {j}} (y | \\mathbf {w}, x) = p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x), \\quad \\forall y. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 339, + 468, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assuming conditional independence of the segment-level labels given the video and its weak labels, we can express the right-hand side as $p_{\\theta}(y|\\mathbf{w},x) = \\prod_{t,m}p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)$ , computed using", + "bbox": [ + 76, + 367, + 468, + 428 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} \\mid \\mathbf {w}, x\\right) := \\left\\{ \\begin{array}{l l} p _ {\\theta} \\left(\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} \\mid x\\right) & \\text {i f} \\mathbf {w} = 1 \\\\ 1 - \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} & \\text {i f} \\mathbf {w} = 0, \\end{array} \\right. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 439, + 468, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|x)$ represents the prediction of the model for modality $m$ and segment $t$ in video $x$ . Thus, $p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)$ becomes a soft pseudo label for the segment $x_{m}^{t}$ .", + "bbox": [ + 76, + 489, + 468, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "M Step. The $j$ -th update finds $\\theta_{j}$ by maximizing $\\mathcal{L}_{\\mathrm{VLB}}(\\theta, \\phi_j)$ , which, due to Eq. (15), is equivalent to", + "bbox": [ + 76, + 536, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (y | x) \\\\ - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (\\mathbf {w} | y, x), \\tag {17} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 578, + 468, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{\\theta}(y|\\mathbf{w},x)$ and $p_{\\theta}(y|x)$ are computed as in Eq. (16). The first term in Eq. (17) can be expanded as the total cross-entropy between the segment-level predictions $p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|x)$ and the pseudo-label targets $p_{\\theta_{j - 1}}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)$ and is therefore equal to $\\mathcal{L}_{\\mathrm{PL}}$ (see Eq. (13)). To understand the second term in Eq. (17), note that in our setting, $p_{\\theta}(\\mathbf{w}|y,x) = p_{\\theta}(\\mathbf{w}|x)$ since we make video-level predictions using MIL-pooling without access to the segment-level labels $(y)$ . The second term then reduces to $-\\ln p_{\\theta}(\\mathbf{w}|x)$ , which is simply the negative log-likelihood of the video-level label $(\\mathbf{w})$ . This means the second term is equal to $\\mathcal{L}_{\\mathrm{MIL}}$ (see Eq. (13)). I.e., the M step is equivalently", + "bbox": [ + 76, + 657, + 468, + 840 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (y | x) - \\ln p _ {\\theta} (\\mathbf {w} | x). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 864, + 468, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "If we simply maximized the data log-likelihood $\\mathcal{L}(\\theta)$ directly (no pseudo labels), this could converge to a solution that picks out just the most discriminative segments in each video, which may suffice to reduce the video-level loss $\\mathcal{L}_{\\mathrm{MIL}}$ . As we showed in this paper, using pseudo labels can help mitigate this, but the quality of the estimated pseudo labels has a crucial role. E.g., hard labeling is inferior to soft labeling since it does not attempt to compute the E step (Eq. (15)) exactly. While soft labeling does attempt an exact E step, better inductive biases (such as prototype-based labeling with contrastive learning) here can help reach a favorable solution, as reported in Sec. 5.3 and Table 2. We also compare with an off-the-shelf pseudo-labeling method [34] in the supplementary material and find that we significantly outperform it, showing that it is not pseudo-labeling per se that works for AVVP - how we estimate the pseudo labels is paramount.", + "bbox": [ + 498, + 90, + 890, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 378, + 617, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a prototype-based pseudo-labeling method for weakly-supervised AVVP that outperforms existing methods. Some interesting findings are that (i) naive pseudolabeling does not produce reliable training targets, (ii) contrastive learning using pseudo labels is an effective strategy, and (iii) training with weak and re-estimated soft pseudo labels is an EM algorithm.", + "bbox": [ + 498, + 409, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We end with some limitations and suggestions for future work. First, performance is sensitive to hyperparameter $\\beta$ as shown in Fig. 4. This could be because identifying negatives for a class from the predictions of an MIL model is less likely to be accurate than identifying positives. Moreover, negative segments for an event would have a much larger variation in content. Second, events with larger intra-class variation may need more prototypes to represent, while we assume the same number to be optimal for all events for convenience. Third, while we can interpret our method as an EM algorithm, EM may only converge to a locally optimal solution. As discussed in Sec. 6, future work could explore inductive biases in the pseudolabeling step that favor better solutions. Finally, it is not clear how many related video-/image-based prediction tasks our method can generalize to. While we provided results for TAL in Sec. 5.4, more evidence of adapting our method to related tasks would be useful in future work.", + "bbox": [ + 498, + 521, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 824, + 658, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Support from Institute of Eminence (IoE) project No. SB22231269EEETWO005001 for Research Centre in Computer Vision is gratefully acknowledged.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18959", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 2", + "[2] Ricardo Cabral, Fernando De la Torre, Joao Paulo Costeira, and Alexandre Bernardino. Matrix Completion for Weakly-Supervised Multi-Label Image Classification. IEEE transactions on pattern analysis and machine intelligence, 37(1): 121-135, 2014. 2", + "[3] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 192-208. Springer, 2022. 7", + "[4] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In Computer Vision - ECCV 2022, pages 431-448, Cham, 2022. Springer Nature Switzerland. 5, 6", + "[5] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE transactions on information theory, 13(1): 21-27, 1967. 2", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5", + "[7] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the Multiple Instance Problem with Axis-Parallel Rectangles. Artificial Intelligence, 89(1):31-71, 1997. 1", + "[8] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, 2018. 2", + "[9] Jie Fu, Junyu Gao, Bing-Kun Bao, and Changsheng Xu. Multimodal imbalance-aware gradient modulation for weakly-supervised audio-visual video parsing. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2", + "[10] Adrien Gaidon, Zaid Harchaoui, and Cordelia Schmid. Temporal localization of actions with actons. IEEE transactions on pattern analysis and machine intelligence, 35(11):2782-2795, 2013. 1, 7", + "[11] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2, 5, 6", + "[12] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE transactions on pattern analysis and machine intelligence, 34(3):417-435, 2012. 2", + "[13] Weifeng Ge, Xiangru Lin, and Yizhou Yu. Weakly Supervised Complementary Parts Models for Fine-Grained Image Classification from the Bottom Up. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3034-3043, 2019. 2", + "[14] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780. IEEE, 2017. 5", + "[15] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2", + "[16] Samantha Guerriero, Barbara Caputo, and Thomas Mensink. DeepNCM: Deep Nearest Class Mean Classifiers, 2018. 2", + "[17] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13925–13935, 2022. 7", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4", + "[20] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 iieme international conference on acoustics, speech and signal processing (icassp), pages 131-135. IEEE, 2017. 5", + "[21] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly Supervised Image Classification through Noise Regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11517-11525, 2019. 2", + "[22] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Relational prototypical network for weakly supervised temporal action localization. In Proceedings of the AAAI conference on artificial intelligence, pages 11053-11060, 2020. 2", + "[23] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding, 155:1-23, 2017. 7", + "[24] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 3", + "[25] Saumya Jetley, Bernardino Romera-Paredes, Sadeep Jayasumana, and Philip Torr. Prototypical priors: From improving classification to zero-shot learning. arXiv preprint arXiv:1512.01192, 2015. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18960", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 4", + "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5", + "[28] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR (Poster), 2015. 5", + "[29] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes, 2022. 8", + "[30] Jatin Lamba, Jayaprakash Akula, Rishabh Dabral, Preethi Jyothi, Ganesh Ramakrishnan, et al. Cross-modal learning for audio-visual video parsing. arXiv preprint arXiv:2104.04598, 2021. 1, 2, 3", + "[31] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002-2006. IEEE, 2019. 5", + "[32] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5", + "[33] Guoqing Liu, Jianxin Wu, and Zhi-Hua Zhou. Key Instance Detection in Multi-Instance Learning. In Proceedings of the Asian Conference on Machine Learning, pages 253–268, Singapore Management University, Singapore, 2012. PMLR. 3", + "[34] Zhekun Luo, Devin Guillory, Baifeng Shi, Wei Ke, Fang Wan, Trevor Darrell, and Huijuan Xu. Weakly-supervised action localization with expectation-maximization multi-instance learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 729-745. Springer, 2020. 2, 7, 8", + "[35] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 2", + "[36] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In Advances in Neural Information Processing Systems, 2022. 2, 3, 5", + "[37] Islam Nassar, Munawar Hayat, Ehsan Abbasnejad, Hamid Rezatofighi, and Gholamreza Haffari. ProtoCon: Pseudo-Label Refinement via Online Clustering and Prototypical Consistency for Efficient Semi-Supervised Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11641–11650, 2023. 2", + "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding, 2018. 4" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Alejandro Pardo, Human Alwassel, Fabian Caba Heilbron, Ali Thabet, and Bernard Ghanem. RefineLoc: Iterative Refinement for Weakly-Supervised Action Localization. In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3318-3327, 2021. 2", + "[40] Kranthi Kumar Rachavarapu and A N Rajagopalan. Boosting positive segments for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10192–10202, 2023. 2, 5", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[42] Dawid Rymarczyk, Aneta Kaczyńska, Jarosław Kraus, Adam Pardyl, and Bartosz Zielinski. Protomil: Multiple instance learning with prototypical parts for fine-grained interpretability. arXiv e-prints, pages arXiv-2108, 2021. 2", + "[43] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood structure. In Artificial intelligence and statistics, pages 412-419. PMLR, 2007. 2", + "[44] Atsushi Sato and Keiji Yamada. Generalized learning vector quantization. Advances in neural information processing systems, 8, 1995. 2", + "[45] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2", + "[46] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple Instance Detection Network with Online Instance Classifier Refinement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2843-2851, 2017. 2", + "[47] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, pages 247–263, 2018. 1, 5", + "[48] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 436–454. Springer, 2020. 1, 2, 3, 5, 6", + "[49] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 5", + "[50] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 2", + "[51] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18961", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1326–1335, 2021. 2, 3, 5, 6", + "[53] Zhirong Wu, Alexei A Efros, and Stella X Yu. Improving generalization via scalable neighborhood component analysis. In Proceedings of the European conference on computer vision (ECCV), pages 685-701, 2018. 2", + "[54] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2", + "[55] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 2", + "[56] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 2", + "[57] Weiyi Yang, Richong Zhang, Junfan Chen, Lihong Wang, and Jaein Kim. Prototype-Guided Pseudo Labeling for Semi-Supervised Text Classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16369-16382, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[58] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In European conference on computer vision, pages 37-54. Springer, 2020. 2", + "[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning Deep Features for Discriminative Localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2", + "[60] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2", + "[61] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2582-2593, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 752 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18962", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_model.json b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4c0b4584b5b3f8009dbc13626e8cdc9f968dd654 --- /dev/null +++ b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_model.json @@ -0,0 +1,2444 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.834, + 0.177 + ], + "angle": 0, + "content": "Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.205, + 0.371, + 0.238 + ], + "angle": 0, + "content": "Kranthi Kumar Rachavarapu IIT Madras" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.243, + 0.386, + 0.256 + ], + "angle": 0, + "content": "kranthi.rachavarapu@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.205, + 0.621, + 0.239 + ], + "angle": 0, + "content": "Kalyan Ramakrishnan \nUniversity of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.243, + 0.634, + 0.256 + ], + "angle": 0, + "content": "kalyanr@robots.ox.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.205, + 0.833, + 0.238 + ], + "angle": 0, + "content": "Rajagopalan A. N. IIT Madras" + }, + { + "type": "text", + "bbox": [ + 0.677, + 0.243, + 0.838, + 0.256 + ], + "angle": 0, + "content": "raju@ee.iitm.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.659 + ], + "angle": 0, + "content": "In this paper, we address the weakly-supervised Audio-Visual Video Parsing (AVVP) problem, which aims at labeling events in a video as audible, visible, or both, and temporally localizing and classifying them into known categories. This is challenging since we only have access to video-level (weak) event labels when training but need to predict event labels at the segment (frame) level at test time. Recent methods employ multiple-instance learning (MIL) techniques that tend to focus solely on the most discriminative segments, resulting in frequent misclassifications. Our idea is to first construct several \"prototype\" features for each event class by clustering key segments identified for the event in the training data. We then assign pseudo labels to all training segments based on their feature similarities with these prototypes and re-train the model under weak and strong supervision. We facilitate this by structuring the feature space with contrastive learning using pseudo labels. Experiments show that we outperform existing methods for weakly-supervised AVVP. We also show that learning with weak and iteratively re-estimated pseudo labels can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.208, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this paper, we explore the problem of weakly-supervised Audio-Visual Video Parsing (AVVP) [48], where the goal is to classify events in a video and localize them over time and modalities, given only video-level (weak) event labels for training. Such a formulation is attractive since it forgoes the need for expensive and tedious fine-grained labeling. Moreover, other weakly-supervised video-based temporal prediction tasks can be posed as special cases of AVVP. For instance, Audio-Visual Event Localization (AVEL) [47] considers events that are simultaneously audible and visible, whereas Temporal Action Localization (TAL) [10] considers only visible events. AVVP is challenging as an event" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.885, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Figure 1. Example predictions of a model trained using MIL under weak supervision [48] and our method. For clarity, we only show visual events. The MIL model detects the event \"Guitar\" correctly in the beginning, but there are some mistakes (red dotted boxes) later on. Moreover, although the last three frames are similar, they are labeled inconsistently by the model. Our method consistently localizes well. \"GT\" is the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.891, + 0.644 + ], + "angle": 0, + "content": "could occur in just the audio stream, visual stream, or both at the same time. Audio-visual data is ubiquitous, with highly variable events in terms of appearance and duration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Recent methods for weakly-supervised AVVP [30, 48] use multiple-instance learning (MIL) [7] techniques that optimize for video-level labels during training and make frame-level predictions during evaluation. Here, video-level predictions are obtained by pooling (aggregating) the frame-level predictions. A drawback of MIL approaches is that a model may learn to focus solely on the most discriminative (\"prominent\") instances (here, audio/visual frames in a video) and still minimize the training loss successfully. This means such models may not reliably detect the full temporal extent of events. Fig. 1 shows an example where an MIL model [48] for AVVP fails to localize events accurately. Moreover, the model labels frames that are visually similar differently. Another reason for misclassifications in MIL models is that they typically contain a frame-level classifier with a single weight vector per event class. However, this may be insufficient to capture the intra-class variation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "for the event across time, modalities, and examples while training with weak labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.363 + ], + "angle": 0, + "content": "We propose to address these issues with the following ideas. First, given that the MIL model can find discriminative frames that trigger the video-level labels, we can identify such frames and propagate their predicted labels to other frames in the training data based on similarities in the feature space, thus creating frame-level pseudo labels on the training data. Second, we can use their features to construct multiple prototypes per event class, ensuring a more robust representation to propagate pseudo labels with. Once this is done for the training data, we can re-train the model with strong supervision from pseudo labels and weak supervision from video-level labels. Fig. 1 shows an example where our method produces more accurate predictions than the MIL model. Because our method relies on feature similarities, we further enforce structure in the feature space via contrastive learning using pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.514 + ], + "angle": 0, + "content": "We re-estimate pseudo labels on the training data after every epoch and train with them over the next epoch. We show that such a training procedure can be viewed through an expectation-maximization (EM) lens with the unknown frame-level labels as latent variables. This interpretation provides convergence guarantees and further insight. Our method achieves state-of-the-art performance for weakly-supervised AVVP and also works well on the related weakly-supervised Temporal Action Localization (TAL) task. We summarize our contributions below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.515, + 0.469, + 0.559 + ], + "angle": 0, + "content": "1. We introduce a prototype-based pseudo-labeling method for weakly-supervised AVVP that exploits similarities in feature space." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.56, + 0.469, + 0.605 + ], + "angle": 0, + "content": "2. We show how to perform contrastive learning using pseudo labels to help semantically structure the feature space." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.469, + 0.636 + ], + "angle": 0, + "content": "3. We present an expectation-maximization view of our general training procedure." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.636, + 0.469, + 0.679 + ], + "angle": 0, + "content": "4. We show state-of-the-art results for weakly-supervised AVVP, and on several metrics for weakly-supervised TAL." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.515, + 0.469, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.219, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Audio-Visual Video Parsing (AVVP). Tian et al. [48] formulated AVVP as a multimodal multiple-instance learning (MMIL) problem and proposed a hybrid attention network with a learned pooling function (HAN) to capture unimodal and cross-modal contexts. Most subsequent work has used the HAN architecture with other techniques to improve performance. Lamba et al. [30] utilized cross-modal information to learn better representations with adversarial and self-supervised losses. Wu et al. [52] estimated modality-level labels by exchanging the audio/visual streams between pairs of videos, followed by retraining under stronger supervision. Lin et al. [32] used cross-video and cross-modality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "signals, such as event co-occurrence across modalities, as additional sources of supervision. Mo et al. [36] leveraged multi-modal grouping to learn discriminative subspaces. Gao et al. [11] extract the \"presence\" and \"absence\" evidence for events from uni- and cross-modal information. While Fu et al. [9] focused on learning rate imbalance between modalities and Zhou et al. [60] utilized CLIP [41] to get segment-labels, our work leverages in-model feature similarities for pseudo labeling. PoiBin [40] modeled the number of positive frames for an event as a latent variable to improve localization. Different from previous methods, we estimate segment-level pseudo labels and re-train under full supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Pseudo-Labeling refers to estimating labels for unlabeled data using the predictions of a trained model. This has been used to improve performance on several weakly-supervised tasks including object detection [1, 46, 59] and image classification [2, 13, 21]. A few works [34, 39, 58] have employed pseudo-labeling to improve performance on Temporal Action Localization (TAL), generating labels from model outputs or attention weights. Luo et al. [34] posed the TAL task as expectation-maximization, using pseudolabels to approximate the E and M steps. Contrary to these, we propose a pseudo-labeling strategy using feature similarities between segments guided by the model predictions and weak labels for a video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Prototype-based Classification. Prototype learning with a nearest-neighbor (non-parametric) classifier has inspired various machine learning techniques [5, 12, 15, 43, 44]. Recently, there has been increased interest in integrating this into deep learning methods, e.g., zero-shot [25], few-shot [45], unsupervised [54, 55], and supervised [16, 35, 53, 56] learning. Prototype learning has also been widely used for image segmentation [8, 50, 51, 61]. Recently, prototype-based methods have also been used for semi-supervised image [37] and text [57] classification. A few works have used prototype-based methods with multiple-instance learning. E.g., Rymarczyk et al. [42] processed large medical images as a bag of patches, and Huang et al. [22] learned prototypes for the TAL task with graph convolutional networks. These works adopt an embedding-level MIL approach ([24]) with one prototype per class, while we adopt instance-level MIL and construct multiple prototypes per event class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.761, + 0.628, + 0.777 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.682, + 0.801 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the audio-visual video parsing (AVVP) problem, we aim to identify audible or visible events in an unconstrained video and simultaneously localize them in time. Specifically, given a video \\( x \\) consisting \\( T \\) non-overlapping temporal segments \\( \\{x_{t} = (x_{t}^{a},x_{t}^{v})\\}_{t = 1}^{T} \\) of the audio \\( (A) \\) and visual \\( (V) \\) streams, our objective is to classify each segment" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18953" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "into \\(C\\) possible events (e.g., singing, vehicle moving, etc.). At test time, we need to predict segment-level event labels \\(y_{t} = (\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}}, \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\) for each audio and visual segment, where \\(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}}, \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}} \\in \\{0,1\\}^{C}\\) are segment-level audio and visual event labels, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.318 + ], + "angle": 0, + "content": "Weak Supervision. In weakly-supervised AVVP, for each video \\( x \\), we only have access to the corresponding video-level event label vector \\( \\mathbf{w} = (w_{1},\\dots ,w_{C})\\in \\{0,1\\}^{C} \\) where \\( w_{c} = 1 \\) if any of the segments in the video contains the \\( c \\)-th event, and \\( w_{c} = 0 \\) otherwise. Such weak, video-level labels only indicate whether an event occurs in the video or not. During evaluation, we still need to predict segment-level labels \\( \\{(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}},\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\}_{t = 1}^{T} \\). Note that multiple events can simultaneously occur in a video, i.e., \\( \\sum_{c}w_{c}\\geq 1 \\), and each segment could have more than one event." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.327, + 0.245, + 0.342 + ], + "angle": 0, + "content": "3.2. MIL Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.468, + 0.44 + ], + "angle": 0, + "content": "We adopt an instance-level MIL approach [24] as done by prior work [30, 32, 36, 48, 52]: First, each instance (here, an audio/visual segment) is assigned a classification score (here, a \\(C\\)-dimensional vector). Then, the scores are aggregated by a pooling operation, and the aggregate is used in the loss. In our case, this looks like" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.452, + 0.468, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ \\left\\{\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {a}} \\right\\} _ {t = 1} ^ {T}, \\left\\{\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {v}} \\right\\} _ {t = 1} ^ {T} \\right] = f _ {\\theta} (x) (1) \\\\ \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} = g _ {\\phi} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}}\\right) \\in [ 0, 1 ] ^ {C}, \\forall t \\in [ 1, T ], m \\in \\{a, v \\} (2) \\\\ \\hat {\\mathbf {w}} = \\sigma_ {\\theta , x} \\left(\\left\\{\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {a}} \\right\\} _ {t = 1} ^ {T}, \\left\\{\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {v}} \\right\\} _ {t = 1} ^ {T}\\right) \\in [ 0, 1 ] ^ {C}, (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.687 + ], + "angle": 0, + "content": "where \\( f_{\\theta} \\) is a feature extractor (encoder), \\( g_{\\phi} \\) is a feature classifier, \\( \\sigma_{\\theta,x} \\) is a MIL-pooling operator. The predicted video-level probability vector \\( \\hat{\\mathbf{w}} \\) is computed from segment-level classification scores \\( \\{\\mathbf{p}_t^m\\} \\), which are computed from segment features \\( \\{\\mathbf{f}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\). We adopt the Hybrid Attention Network (HAN) [48] architecture to implement \\( f_{\\theta}, g_{\\phi} \\), and \\( \\sigma \\). Here, \\( f_{\\theta} \\) consists of a transformer-based encoder with cross-modal attention and \\( g_{\\phi} \\) is a linear classifier with the sigmoid activation. Attention-based averaging is used for \\( \\sigma \\). We provide implementation details in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.688, + 0.468, + 0.718 + ], + "angle": 0, + "content": "This model is trained end-to-end with supervision at the video level by minimizing a weighted cross-entropy loss" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.73, + 0.468, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M I L}} = \\operatorname {C E} (\\hat {\\mathbf {w}}, \\mathbf {w}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.468, + 0.802 + ], + "angle": 0, + "content": "with weights inversely proportional to the event distribution in the dataset. At test time, we get segment-level predictions from Eq. (2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.383, + 0.833 + ], + "angle": 0, + "content": "4. Prototype-based Pseudo-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "While instance-level MIL enables finding key segments that trigger the video-level labels [24, 33], they often miss out on the full temporal extent of events in a video. However, if we identify such discriminative segments for each label" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "in a training video, we could use them to build a set of prototype features that represent each class. A high degree of similarity of a segment with the prototype features for an event class means the segment probably contains the event. We use this idea to generate pseudo labels for all segments in a training video and then re-train the model under strong supervision. Fig. 2 shows an overview of our approach. We provide pseudocode in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.222, + 0.787, + 0.238 + ], + "angle": 0, + "content": "4.1. Constructing Prototype Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.35 + ], + "angle": 0, + "content": "The segment-level linear classifier \\( g_{\\phi} \\) (Eq. (2)) can itself be considered a prototype-based classifier, where the prototype for each class is the corresponding weight vector in \\( \\phi \\). However, this requires that an event have similar representations across modalities and examples in the training data. We instead construct multiple prototypes for each class in the following two steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Segment Selection. For each training video \\( x \\), we find positive segments \\( \\mathcal{P}_x(c) \\) for a target event \\( c \\) (if \\( w_c = 1 \\)) as those the model is highly confident contain the event. I.e.," + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.407, + 0.891, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {x} (c) = \\left\\{ \\begin{array}{l l} \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) \\geq \\gamma p ^ {*} (c) \\right\\} & \\text {i f} w _ {c} = 1 \\\\ \\phi & \\text {i f} w _ {c} = 0, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.623 + ], + "angle": 0, + "content": "where \\(\\gamma \\in (0,1]\\) is a hyperparameter and \\(p^*(c) = \\max_{t,m} \\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}(c)\\) is the maximum predicted probability for the event over time and modalities for a given video. Clearly, \\(\\mathcal{P}_X(c)\\) will contain at least one audio/visual segment from video \\(x\\) when \\(w_c = 1\\). We assume these segments contain event \\(c\\) with high probability (so are likely true positives). We also find negative segments \\(\\mathcal{N}_x(c)\\) for a target event \\(c\\) as those the model is highly confident do not contain the event when \\(w_c = 1\\) (likely true negatives), or is highly confident to contain the event when \\(w_c = 0\\) (definite false positives):" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.634, + 0.891, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} _ {x} (c) = \\left\\{ \\begin{array}{l l} \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) \\leq \\beta p ^ {*} (c) \\right\\} & \\text {i f} w _ {c} = 1 \\\\ \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = p ^ {*} (c) > = 0. 5 \\right\\} & \\text {i f} w _ {c} = 0, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.891, + 0.745 + ], + "angle": 0, + "content": "where \\(\\beta \\in (0,1)\\) is a hyperparameter set to a small value to retrieve segments without the event \\(c\\) (if any). Note that \\(\\mathcal{N}_x(c)\\) may contain no segments at all even when \\(w_{c} = 1\\), i.e., be a null set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.746, + 0.891, + 0.777 + ], + "angle": 0, + "content": "Finally, we aggregate the positive and negative sets over all videos in training data \\(\\mathcal{D}\\) to get prototypical segments as" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.788, + 0.891, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} (c) = \\bigcup_ {x \\in \\mathcal {D}} \\mathcal {P} _ {x} (c), \\quad \\mathcal {N} (c) = \\bigcup_ {x \\in \\mathcal {D}} \\mathcal {N} _ {x} (c). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Prototype Generation. We now generate multiple prototypes (feature vectors) for each event class by hypothesizing that having more than one prototype would help better capture the intra-class variation across modalities and training examples, and create a more robust representation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.087, + 0.791, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 2. An overview of our method. The top branch is trained using MIL with weak labels and using full supervision with pseudo labels, which are generated by the bottom branch. This is done by first selecting key segments for each event from the training data and clustering them to construct a set of prototype features for the event, used to guide pseudo-label generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.473, + 0.493 + ], + "angle": 0, + "content": "First, we extract features for all segments in \\(\\mathcal{P}(c)\\) and \\(\\mathcal{N}(c)\\) using Eq. (1), and call the resulting feature sets \\(\\mathcal{F}_{\\mathcal{P}}(c)\\) and \\(\\mathcal{F}_{\\mathcal{N}}(c)\\). Here, instead of \\(f_{\\theta}\\), we use a momentum encoder [19] \\(f_{\\tilde{\\theta}}\\) having the same architecture but whose weights are set (after each update to \\(\\theta\\)) as an exponential moving average (EMA) of \\(\\theta\\) over training steps: \\(\\tilde{\\theta} \\gets (1 - \\eta)\\tilde{\\theta} + \\eta\\theta\\), where \\(\\eta \\in (0,1)\\). The momentum encoder is more stable and less sensitive to noise from weak supervision, leading to slightly improved performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.498, + 0.471, + 0.543 + ], + "angle": 0, + "content": "Next, we perform \\( k \\)-means clustering separately on \\( \\mathcal{F}_{\\mathcal{P}}(c) \\) and \\( \\mathcal{F}_{\\mathcal{N}}(c) \\) to get two sets of cluster centroids. We call the centroids prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.565, + 0.47, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {\\mathcal {P}} (c) = \\left\\{\\mathbf {f} _ {\\mathbf {i}} ^ {+} \\right\\} _ {i = 1} ^ {k _ {p}}, \\quad \\mathcal {C} _ {\\mathcal {N}} (c) = \\left\\{\\mathbf {f} _ {\\mathbf {j}} ^ {-} \\right\\} _ {j = 1} ^ {k _ {n}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Clustering for finding prototypes imposes a natural bottleneck [26], helping discard irrelevant information in the segments. We repeat this for each event class." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.676, + 0.314, + 0.691 + ], + "angle": 0, + "content": "4.2. Estimating Pseudo Labels" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.81 + ], + "angle": 0, + "content": "We assign pseudo labels to each segment in a training video by considering its feature similarities with the prototypes for each class. We assume that if an audio/visual segment contains event \\( c \\), it would be on average closer to \\( \\mathcal{P}(c) \\) in feature space than to \\( \\mathcal{N}(c) \\). For this, we explore two approaches: (i) hard labeling, which assigns the label of the closest-on-average set among the prototype sets for class \\( c \\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.842, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = w _ {c} \\cdot \\mathbb {1} \\left[ \\frac {\\sum_ {\\mathbf {f} ^ {+} \\in \\mathcal {C} _ {\\mathcal {P}} (c)} e ^ {(\\mathbf {f} ^ {+}) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}}}}{| \\mathcal {C} _ {\\mathcal {P}} |} > \\frac {\\sum_ {\\mathbf {f} ^ {-} \\in \\mathcal {C} _ {\\mathcal {N}} (c)}}{| \\mathcal {C} _ {\\mathcal {N}} |} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.357, + 0.892, + 0.385 + ], + "angle": 0, + "content": "and (ii) soft labeling, which instead assigns a probability score:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.388, + 0.892, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = w _ {c} \\cdot \\frac {\\frac {1}{| \\mathcal {C} _ {\\mathcal {P}} |} \\sum_ {\\mathbf {f} ^ {+} \\in \\mathcal {C} _ {\\mathcal {P}} (c)} e ^ {\\left(\\mathbf {f} ^ {+}\\right) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}} / \\tau}}{\\sum_ {\\mathcal {S} \\in \\left\\{\\mathcal {C} _ {\\mathcal {P}}, \\mathcal {C} _ {\\mathcal {N}} \\right\\}} \\frac {1}{| \\mathcal {S} |} \\sum_ {\\mathbf {f} ^ {\\prime} \\in \\mathcal {S} (c)} e ^ {\\left(\\mathbf {f} ^ {\\prime}\\right) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}} / \\tau}} \\in (0, 1), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.465, + 0.807, + 0.479 + ], + "angle": 0, + "content": "where \\(\\tau > 0\\) is a temperature hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.48, + 0.892, + 0.541 + ], + "angle": 0, + "content": "When training our model, we generate pseudo labels \\(\\{\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\}\\) after every epoch for each training video and use them to fully supervise the segment-level probabilities \\(\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\}\\) in Eq. (2) using the binary cross-entropy loss" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.546, + 0.892, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P L}} = \\operatorname {C E} \\left(\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}}, \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.568, + 0.813, + 0.584 + ], + "angle": 0, + "content": "averaged over segments and training examples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.855, + 0.606 + ], + "angle": 0, + "content": "4.3. Contrastive Learning with Pseudo Labels" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Given that our method exploits feature similarities between segments based on the events they may contain, we use contrastive learning based on pseudo labels to help semantically structure the feature space. Since multiple events can occur simultaneously, for each audio/visual segment, we consider a similar segment as one estimated to have some common event(s) and a dissimilar segment as one estimated to have no common events. Formally, given a batch of training examples, let \\( B = \\{\\mathbf{f}_{\\mathbf{t}}^{\\mathbf{m}} \\to \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\) be a dictionary from features to the assigned pseudo labels for all segments and over all videos in the batch. For every feature \\( \\mathbf{f} \\in B \\), the similar feature-set is then \\( S(\\mathbf{f}) = \\{\\mathbf{g} \\in B \\mid B(\\mathbf{g}) \\cap B(\\mathbf{f}) \\neq \\phi\\} \\) and the dissimilar set is \\( D(\\mathbf{f}) = \\{\\mathbf{g} \\in B \\mid B(\\mathbf{g}) \\cap B(\\mathbf{f}) = \\phi\\} \\). To pull segments with a common predicted event closer in the feature space, we use the contrastive loss [38]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.844, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L}} = - \\log \\frac {\\sum_ {\\mathbf {g} \\in \\mathcal {S} (\\mathbf {f})} e ^ {\\mathbf {g} ^ {T} \\mathbf {f} / \\tau_ {c}}}{\\sum_ {\\mathbf {g} \\in \\mathcal {S} (\\mathbf {f})} e ^ {\\mathbf {g} ^ {T} \\mathbf {f} / \\tau_ {c}} + \\sum_ {\\mathbf {h} \\in \\mathcal {D} (\\mathbf {f})} e ^ {\\mathbf {h} ^ {T} \\mathbf {f} / \\tau_ {c}}}, \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18955" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.821, + 0.281 + ], + "angle": 0, + "content": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
AVE* [47]47.240.437.134.735.431.639.935.541.636.5
AVSDN* [31]47.834.152.046.337.126.545.735.650.837.7
HAN [48]60.151.352.948.948.943.054.047.755.448.0
MA [52]60.353.660.056.455.149.058.953.057.950.6
CVCM-MA [32]60.853.863.558.957.049.560.554.059.552.1
MGN-MA [36]60.250.961.959.755.549.659.253.458.749.9
JoMoLD [4]61.353.963.859.957.249.660.854.559.952.5
PoiBin [40]63.154.163.560.357.751.561.455.260.652.3
CMPAE [11]64.256.666.463.759.251.863.357.462.855.7
Ours65.957.366.764.361.954.364.859.963.757.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.321 + ], + "angle": 0, + "content": "Table 1. Results (% F1-scores) on all metrics defined in Tian et al. [48]. \\* indicates reproduced performance. The best and second best results are in bold and underline, respectively." + }, + { + "type": "code_caption", + "bbox": [ + 0.083, + 0.336, + 0.367, + 0.351 + ], + "angle": 0, + "content": "Algorithm 1: Pseudocode for our method" + }, + { + "type": "algorithm", + "bbox": [ + 0.08, + 0.354, + 0.47, + 0.601 + ], + "angle": 0, + "content": "1 Init: model params \\(\\psi = \\{\\theta ,\\phi \\}\\) , EMA params \\(\\psi_{E} = \\{\\theta_{E},\\phi_{E}\\}\\) \n2 while \\(\\psi\\) has not converged do \n3 #estimate segment - level pseudo labels \n4 for \\((x,\\mathbf{w})\\) in train set do \n5 \\(\\left|\\{\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\gets p_{\\theta}(\\cdot |\\mathbf{w},x)\\right.;\\) \n6 end \n7 #train \n8 for \\((x,\\mathbf{w})\\) in train set do \n9 \\(\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\gets p_{\\theta ,\\phi}(\\cdot |x);\\) \n10 \\(\\hat{\\mathbf{w}} = \\sigma_{\\theta ,x}(\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\});\\) \n11 \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{MIL}} + \\lambda_{\\mathrm{PL}}\\mathcal{L}_{\\mathrm{PL}} + \\lambda_{\\mathrm{CL}}\\mathcal{L}_{\\mathrm{CL}};\\) \n12 \\(\\psi \\gets \\psi -\\alpha \\nabla_{\\psi}\\mathcal{L};\\) \n13 \\(\\psi_{E}\\gets (1 - \\eta)\\psi_{E} + \\eta \\psi ;\\) \n14 end \n15 end" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.673 + ], + "angle": 0, + "content": "averaged over all \\(\\mathbf{f} \\in B\\), where \\(\\tau_{c} > 0\\) is a temperature hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.291, + 0.71 + ], + "angle": 0, + "content": "4.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.469, + 0.752 + ], + "angle": 0, + "content": "We train our model using the following loss in an end-to-end fashion:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.774, + 0.469, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M I L}} + \\lambda_ {\\mathrm {P L}} \\mathcal {L} _ {\\mathrm {P L}} + \\lambda_ {\\mathrm {C L}} \\mathcal {L} _ {\\mathrm {C L}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{PL}}, \\lambda_{\\mathrm{CL}} > 0\\) are hyperparameters. During training, we generate pseudo labels once every epoch for all training examples using the best model so far. During inference, we predict the segment-level probabilities using the non-parametric prototypical pseudo-labeling (Eq. (10)) with video-level predictions from Eq. (3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.331, + 0.632, + 0.349 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.582, + 0.373 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.53 + ], + "angle": 0, + "content": "LLP Dataset. We conduct experiments on the Look, Listen and Parse (LLP) dataset [48] which consists of 11849 YouTube videos, each of 10s duration, labeled into 25 event categories. These videos are unconstrained with a wide variety of scene content, including daily activities, music performances, and vehicle sounds. We use 10000 videos with weak (video-level) labels for training. The remaining 1849 fully-annotated videos (with segment-level labels) are used for validation and testing. The train/val/test split has been provided in the LLP dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous work, we use F1-scores on audio, visual, and audio-visual events as evaluation metrics. These are computed both at the segment and event level. We also include the aggregate metrics \"Type@AV\" and \"Event@AV\", again computed at the segment and event level. See Tian et al. [48] for a full explanation of metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.82 + ], + "angle": 0, + "content": "Implementation Details. We use ResNet-152 [18] pretrained on ImageNet [6] and R(2+1)D-18 [49] pre-trained on Kinetics-400 [27] as visual feature extractors to generate a 512-d visual feature per segment. We use VGGish [20] pre-trained on AudioSet [14] to generate a 128-d audio feature per segment. We use the Adam optimizer [28] with a batch size of 16 and a learning rate \\(\\alpha = 3e - 4\\) for 50 epochs. We take \\(\\eta = 0.999\\) for the momentum encoder, \\(\\lambda_{\\mathrm{PL}} = \\lambda_{\\mathrm{CL}} = 0.1\\) in Eq. (13), \\(\\gamma = 1\\) in Eq. (5), \\(\\beta = 0.2\\) in Eq. (6), \\(\\tau = 0.1\\) in Eq. (10), and \\(\\tau_{c} = 0.2\\) in Eq. (12). We set \\(k_{p} = k_{n} = 10\\) in Eq. (8). Hyperparameter values are based on validation performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.771, + 0.847 + ], + "angle": 0, + "content": "5.2. Comparisons with Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.891, + 0.884 + ], + "angle": 0, + "content": "For a fair comparison, all methods considered use the same pre-trained feature extractors and train/val/test split." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Quantitative. We compare methods in Table 1 on all the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.478, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.147, + 0.882, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.401, + 0.88, + 0.417 + ], + "angle": 0, + "content": "Figure 3. Audio-Visual Video Parsing results of our methods with HAN [48], MA [52], JoMoLD [4] and CMPAE [11] on two videos." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.426, + 0.468, + 0.507 + ], + "angle": 0, + "content": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
Base+NPLH58.952.158.549.855.748.257.850.155.746.6
Base+NPILS59.352.559.051.156.048.758.250.856.247.6
Base+PPLH62.557.462.959.659.654.061.956.360.355.7
Base+PPLS65.957.366.764.361.954.364.859.963.757.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.518, + 0.455, + 0.532 + ], + "angle": 0, + "content": "Table 2. Performance for different pseudo labeling strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.701 + ], + "angle": 0, + "content": "metrics listed earlier. We see that our method outperforms others on audio, audio-visual, Type@AV, and Event@AV scores at both the segment and event levels. In particular, we achieve an average improvement of 2.6 points for audio-visual events. Averaged over all metrics, our method achieves a 1.86 point improvement over the previous best method CMPAE [11]. Since we do not require additional learnable parameters to generate pseudo labels, this improvement is with a comparable number of parameters to previous methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Qualitative. We qualitatively compare our method with some previous methods on two examples in Fig. 3. Here, Fig. 3 (left) contains events \"Singing\" and \"Banjo\" in both modalities. Fig. 3 (Right) shows another scene with the event \"Speech\" in only the audio modality, and \"Violin\" occurring in both modalities. As evident, the common failure mode of previous methods is that they detect only a portion of the duration of events, while ours more commonly detects the full extent of events. As discussed in Sec. 1, this could be because of multiple-instance learning encouraging the model to pick out only the most discriminative segments containing an event, as opposed to all segments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.887, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Computational Cost. When training, our approach takes" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.489 + ], + "angle": 0, + "content": "67.4s per epoch (17.1s for forward+backward propagation and 50.3s for prototype-based pseudo labeling). Inference time is 5.9ms on average per video, which is comparable to the baselines (2.9ms)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.602, + 0.515 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Pseudo-Labeling Strategies. We report an ablation study in Table 2 to validate our specific pseudo-labeling method. Here, Base is the baseline MIL model explained in Sec. 3.2. Base+NPL refers to the base model trained with naive pseudo labeling, i.e., by directly using the segment-level predictions of the model as pseudo labels for retraining. In this case, H and S refer to using hard (binarized) and soft (continuous) pseudo labels, respectively. We observe no consistent improvement with naive pseudo labeling, likely due to the MIL model producing incorrect predictions for the non-discriminative segments for an event, leading to less reliable pseudo labels. Base+PPL refers to the base model trained with prototype-based pseudo labels as described in Sec. 4.2. In this case, H and S refer to using hard and soft pseudo labels, as defined in Eqs. (9) and (10), respectively. Our method significantly improves performance by generating more reliable pseudo labels throughout a training video. We also see that soft labeling is better than hard labeling. Since segment-level predictions may be noisy, we expect mistakes to affect performance more with hard rather than soft labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Reliability of Pseudo Labels. We now assess if the pseudo labels generated by our method are reliable for re-training. For this, we take the pre-trained baselines HAN [48] and MA [52], generate prototypes for each event from the train-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18957" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.162 + ], + "angle": 0, + "content": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
HAN60.151.352.948.948.943.054.047.755.448.0
+PPLTest62.555.455.351.152.346.956.050.958.350.6
MA60.353.660.056.455.149.058.953.057.950.6
+PPLTest61.755.461.857.957.551.660.655.059.452.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.173, + 0.468, + 0.187 + ], + "angle": 0, + "content": "Table 3. Evaluating prototype-based pseudo labels on the test set." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.21, + 0.262, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.209, + 0.45, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.339, + 0.429, + 0.354 + ], + "angle": 0, + "content": "Figure 4. Performance for different \\(\\gamma\\) and \\(\\beta\\) choices." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.361, + 0.468, + 0.424 + ], + "angle": 0, + "content": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
No-CL64.156.459.657.357.149.661.256.460.856.8
With-CL65.957.366.764.361.954.364.859.963.757.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.434, + 0.46, + 0.448 + ], + "angle": 0, + "content": "Table 4. Effect of contrastive learning (CL) with pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.468, + 0.612 + ], + "angle": 0, + "content": "ing data, and simply predict the prototype-based pseudo labels (Eq. (9)) on the test set for evaluation (i.e., discard the classifiers at test time). We call the resulting classifiers \\(\\mathrm{HAN} + \\mathrm{PPL}_{\\mathrm{Test}}\\) and \\(\\mathrm{MA} + \\mathrm{PPL}_{\\mathrm{Test}}\\) and show results in Table 3. With no other change to the models other than the classifiers, we see significant improvements in all metrics, showing that our pseudo labels are indeed reliable estimates of the true labels. Note that for the results in Table 1, we recompute pseudo labels on the training data every epoch, with a continually improving base model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Effect of \\(\\gamma\\) and \\(\\beta\\). The hyperparameters \\(\\gamma\\) and \\(\\beta\\) in Eqs. (5) and (6) roughly control the proportion of segments that will be selected as positives and negatives for assigning pseudo labels. We experiment with different values and plot the average of all segment/event-level metrics in Fig. 4. Performance is susceptible to \\(\\gamma\\), which controls positives, and is tolerant to a slight variation in \\(\\beta\\). The best performance is for \\(\\gamma = 1\\) and \\(\\beta = 0.2\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Effect of Contrastive Learning. We analyze the impact of contrastive learning using pseudo labels (Eq. (12)) by simply setting \\(\\lambda_{\\mathrm{CL}} = 0\\) in Eq. (13). Results are reported in Tab. 4. While only training with prototype-based pseudo labels \\((\\mathsf{No - CL})\\) significantly improves the baseline, adding a contrastive loss based on our pseudo labels (With-CL) further yields a large improvement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Number of Clusters. We take \\( k_{p} = k_{n} \\) and try values in \\( \\{1,2,\\dots,100\\} \\). We plot the average of all segment/event-level metrics in Fig. 5. The best performance was for \\( k = 10 \\) clusters per class. Having fewer clusters helps cap" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.094, + 0.861, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.238, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Performance for different choices of the number of clusters/prototypes per event." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.282, + 0.89, + 0.352 + ], + "angle": 0, + "content": "
MethodtIoUAvg.
0.10.20.30.40.50.60.7[0.1:0.5][0.3:0.7][0.1:0.7]
EM-MIL [34]59.152.745.536.830.522.716.444.930.437.7
ASM-Loc [17]71.265.557.146.836.625.213.455.435.845.1
DELU [3]71.566.256.547.740.527.215.356.537.446.4
Ours72.766.957.946.937.426.820.156.437.846.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.362, + 0.892, + 0.389 + ], + "angle": 0, + "content": "Table 5. Results (\\% mAP) for the TAL task at different tIoU thresholds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.463 + ], + "angle": 0, + "content": "ture more general features for an event instead of instance-specific details, but having too few hurts performance by discarding intra-class variation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.747, + 0.488 + ], + "angle": 0, + "content": "5.4. Results for a Different Task" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.663 + ], + "angle": 0, + "content": "We conduct preliminary experiments for the weakly-supervised Temporal Action Localization (TAL) task [10] on the THUMOS14 dataset [23] to further validate our method. TAL aims to temporally localize and classify visual events in videos into known categories. We report mean average precisions (mAP) at different temporal intersection over union (tIoU) thresholds and the results are in Table 5. We achieve better or comparable performance w.r.t. the current state-of-the-art method DELU [3], and outperform previous methods. Check the supplementary for a full comparison." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.891, + 0.696 + ], + "angle": 0, + "content": "6. Interpretation as Expectation-Maximization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.856 + ], + "angle": 0, + "content": "We now show that iteratively re-estimating soft pseudo labels and training the model under weak \\(+\\) strong supervision (Eq. (13)) can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure. For this, we view the segment-level labels \\(y = \\{(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}},\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\}_{t = 1}^{T}\\) for each video as latent variables since they are not observed when training under weak supervision. The observed variables are the video segments \\(x = \\{(x_{t}^{a},x_{t}^{v})\\}_{t = 1}^{T}\\) and weak labels \\(\\mathbf{w}\\). For clarity, assume a single event class, i.e., \\(\\mathbf{w},\\mathbf{y}_{\\mathbf{m}}^{\\mathbf{t}}\\in \\{0,1\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our goal is to maximize the data log-likelihood \\(\\mathcal{L}(\\theta) = \\ln p_{\\theta}(\\mathbf{w}|x)\\) over model parameters \\(\\theta\\) (assume just one data point \\(x\\) for clarity). Instead of maximizing this directly, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18958" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "can attempt to maximize the variational lower bound [29] \\(\\mathcal{L}_{\\mathrm{VLB}}\\) for this likelihood (random variables are capitalized):" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.149, + 0.469, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V L B}} (\\theta , \\phi) = \\underset {Y \\sim q _ {\\phi} (\\cdot | \\mathbf {w}, x)} {\\mathbb {E}} [ \\ln p _ {\\theta} (\\mathbf {w} | Y, x) ] \\\\ - \\mathrm {D} _ {\\mathrm {K L}} \\left[ q _ {\\phi} (\\cdot | \\mathbf {w}, x) \\| p _ {\\theta} (\\cdot | x) \\right], \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.469, + 0.311 + ], + "angle": 0, + "content": "where \\(q_{\\phi}(y|\\mathbf{w},x)\\) is any conditional distribution over labels \\(y\\), with parameters \\(\\phi\\), meant to approximate the true posterior \\(p_{\\theta}(y|x)\\). The EM algorithm alternates between two steps, first approximating the posterior by maximizing \\(\\mathcal{L}_{\\mathrm{VLB}}\\) over \\(\\phi\\) (E step), and then improving the data likelihood by maximizing it over \\(\\theta\\) (M step). Let \\(j \\geq 1\\) index the iterations (note, \\(\\theta\\) and \\(\\phi\\) here are not the same as in Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.312, + 0.316, + 0.327 + ], + "angle": 0, + "content": "E Step. The \\(j\\)-th update is given by" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.34, + 0.469, + 0.357 + ], + "angle": 0, + "content": "\\[\nq _ {\\phi_ {j}} (y | \\mathbf {w}, x) = p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x), \\quad \\forall y. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.469, + 0.429 + ], + "angle": 0, + "content": "Assuming conditional independence of the segment-level labels given the video and its weak labels, we can express the right-hand side as \\( p_{\\theta}(y|\\mathbf{w},x) = \\prod_{t,m}p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x) \\), computed using" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.44, + 0.469, + 0.48 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} \\mid \\mathbf {w}, x\\right) := \\left\\{ \\begin{array}{l l} p _ {\\theta} \\left(\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} \\mid x\\right) & \\text {i f} \\mathbf {w} = 1 \\\\ 1 - \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} & \\text {i f} \\mathbf {w} = 0, \\end{array} \\right. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.469, + 0.538 + ], + "angle": 0, + "content": "where \\( p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|x) \\) represents the prediction of the model for modality \\( m \\) and segment \\( t \\) in video \\( x \\). Thus, \\( p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x) \\) becomes a soft pseudo label for the segment \\( x_{m}^{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.537, + 0.469, + 0.569 + ], + "angle": 0, + "content": "M Step. The \\(j\\)-th update finds \\(\\theta_{j}\\) by maximizing \\(\\mathcal{L}_{\\mathrm{VLB}}(\\theta, \\phi_j)\\), which, due to Eq. (15), is equivalent to" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.579, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (y | x) \\\\ - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (\\mathbf {w} | y, x), \\tag {17} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.469, + 0.842 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(y|\\mathbf{w},x)\\) and \\(p_{\\theta}(y|x)\\) are computed as in Eq. (16). The first term in Eq. (17) can be expanded as the total cross-entropy between the segment-level predictions \\(p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|x)\\) and the pseudo-label targets \\(p_{\\theta_{j - 1}}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)\\) and is therefore equal to \\(\\mathcal{L}_{\\mathrm{PL}}\\) (see Eq. (13)). To understand the second term in Eq. (17), note that in our setting, \\(p_{\\theta}(\\mathbf{w}|y,x) = p_{\\theta}(\\mathbf{w}|x)\\) since we make video-level predictions using MIL-pooling without access to the segment-level labels \\((y)\\). The second term then reduces to \\(-\\ln p_{\\theta}(\\mathbf{w}|x)\\), which is simply the negative log-likelihood of the video-level label \\((\\mathbf{w})\\). This means the second term is equal to \\(\\mathcal{L}_{\\mathrm{MIL}}\\) (see Eq. (13)). I.e., the M step is equivalently" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.865, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (y | x) - \\ln p _ {\\theta} (\\mathbf {w} | x). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.349 + ], + "angle": 0, + "content": "If we simply maximized the data log-likelihood \\(\\mathcal{L}(\\theta)\\) directly (no pseudo labels), this could converge to a solution that picks out just the most discriminative segments in each video, which may suffice to reduce the video-level loss \\(\\mathcal{L}_{\\mathrm{MIL}}\\). As we showed in this paper, using pseudo labels can help mitigate this, but the quality of the estimated pseudo labels has a crucial role. E.g., hard labeling is inferior to soft labeling since it does not attempt to compute the E step (Eq. (15)) exactly. While soft labeling does attempt an exact E step, better inductive biases (such as prototype-based labeling with contrastive learning) here can help reach a favorable solution, as reported in Sec. 5.3 and Table 2. We also compare with an off-the-shelf pseudo-labeling method [34] in the supplementary material and find that we significantly outperform it, showing that it is not pseudo-labeling per se that works for AVVP - how we estimate the pseudo labels is paramount." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.379, + 0.619, + 0.395 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.516 + ], + "angle": 0, + "content": "We proposed a prototype-based pseudo-labeling method for weakly-supervised AVVP that outperforms existing methods. Some interesting findings are that (i) naive pseudolabeling does not produce reliable training targets, (ii) contrastive learning using pseudo labels is an effective strategy, and (iii) training with weak and re-estimated soft pseudo labels is an EM algorithm." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.793 + ], + "angle": 0, + "content": "We end with some limitations and suggestions for future work. First, performance is sensitive to hyperparameter \\(\\beta\\) as shown in Fig. 4. This could be because identifying negatives for a class from the predictions of an MIL model is less likely to be accurate than identifying positives. Moreover, negative segments for an event would have a much larger variation in content. Second, events with larger intra-class variation may need more prototypes to represent, while we assume the same number to be optimal for all events for convenience. Third, while we can interpret our method as an EM algorithm, EM may only converge to a locally optimal solution. As discussed in Sec. 6, future work could explore inductive biases in the pseudolabeling step that favor better solutions. Finally, it is not clear how many related video-/image-based prediction tasks our method can generalize to. While we provided results for TAL in Sec. 5.4, more evidence of adapting our method to related tasks would be useful in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.825, + 0.66, + 0.842 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Support from Institute of Eminence (IoE) project No. SB22231269EEETWO005001 for Research Centre in Computer Vision is gratefully acknowledged." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18959" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Ricardo Cabral, Fernando De la Torre, Joao Paulo Costeira, and Alexandre Bernardino. Matrix Completion for Weakly-Supervised Multi-Label Image Classification. IEEE transactions on pattern analysis and machine intelligence, 37(1): 121-135, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[3] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 192-208. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[4] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In Computer Vision - ECCV 2022, pages 431-448, Cham, 2022. Springer Nature Switzerland. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[5] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE transactions on information theory, 13(1): 21-27, 1967. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[7] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the Multiple Instance Problem with Axis-Parallel Rectangles. Artificial Intelligence, 89(1):31-71, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[8] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[9] Jie Fu, Junyu Gao, Bing-Kun Bao, and Changsheng Xu. Multimodal imbalance-aware gradient modulation for weakly-supervised audio-visual video parsing. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[10] Adrien Gaidon, Zaid Harchaoui, and Cordelia Schmid. Temporal localization of actions with actons. IEEE transactions on pattern analysis and machine intelligence, 35(11):2782-2795, 2013. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[11] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[12] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE transactions on pattern analysis and machine intelligence, 34(3):417-435, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Weifeng Ge, Xiangru Lin, and Yizhou Yu. Weakly Supervised Complementary Parts Models for Fine-Grained Image Classification from the Bottom Up. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3034-3043, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[14] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[15] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Samantha Guerriero, Barbara Caputo, and Thomas Mensink. DeepNCM: Deep Nearest Class Mean Classifiers, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[17] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13925–13935, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[20] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 iieme international conference on acoustics, speech and signal processing (icassp), pages 131-135. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[21] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly Supervised Image Classification through Noise Regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11517-11525, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[22] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Relational prototypical network for weakly supervised temporal action localization. In Proceedings of the AAAI conference on artificial intelligence, pages 11053-11060, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding, 155:1-23, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[24] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Saumya Jetley, Bernardino Romera-Paredes, Sadeep Jayasumana, and Philip Torr. Prototypical priors: From improving classification to zero-shot learning. arXiv preprint arXiv:1512.01192, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18960" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.265 + ], + "angle": 0, + "content": "[28] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR (Poster), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.47, + 0.295 + ], + "angle": 0, + "content": "[29] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[30] Jatin Lamba, Jayaprakash Akula, Rishabh Dabral, Preethi Jyothi, Ganesh Ramakrishnan, et al. Cross-modal learning for audio-visual video parsing. arXiv preprint arXiv:2104.04598, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.355, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[31] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002-2006. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[32] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.499, + 0.471, + 0.568 + ], + "angle": 0, + "content": "[33] Guoqing Liu, Jianxin Wu, and Zhi-Hua Zhou. Key Instance Detection in Multi-Instance Learning. In Proceedings of the Asian Conference on Machine Learning, pages 253–268, Singapore Management University, Singapore, 2012. PMLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.47, + 0.668 + ], + "angle": 0, + "content": "[34] Zhekun Luo, Devin Guillory, Baifeng Shi, Wei Ke, Fang Wan, Trevor Darrell, and Huijuan Xu. Weakly-supervised action localization with expectation-maximization multi-instance learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 729-745. Springer, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[35] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[36] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In Advances in Neural Information Processing Systems, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.869 + ], + "angle": 0, + "content": "[37] Islam Nassar, Munawar Hayat, Ehsan Abbasnejad, Hamid Rezatofighi, and Gholamreza Haffari. ProtoCon: Pseudo-Label Refinement via Online Clustering and Prototypical Consistency for Efficient Semi-Supervised Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11641–11650, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[39] Alejandro Pardo, Human Alwassel, Fabian Caba Heilbron, Ali Thabet, and Bernard Ghanem. RefineLoc: Iterative Refinement for Weakly-Supervised Action Localization. In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3318-3327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.23 + ], + "angle": 0, + "content": "[40] Kranthi Kumar Rachavarapu and A N Rajagopalan. Boosting positive segments for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10192–10202, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.893, + 0.372 + ], + "angle": 0, + "content": "[42] Dawid Rymarczyk, Aneta Kaczyńska, Jarosław Kraus, Adam Pardyl, and Bartosz Zielinski. Protomil: Multiple instance learning with prototypical parts for fine-grained interpretability. arXiv e-prints, pages arXiv-2108, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.893, + 0.426 + ], + "angle": 0, + "content": "[43] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood structure. In Artificial intelligence and statistics, pages 412-419. PMLR, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.893, + 0.468 + ], + "angle": 0, + "content": "[44] Atsushi Sato and Keiji Yamada. Generalized learning vector quantization. Advances in neural information processing systems, 8, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[45] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[46] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple Instance Detection Network with Online Instance Classifier Refinement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2843-2851, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.893, + 0.622 + ], + "angle": 0, + "content": "[47] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, pages 247–263, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.893, + 0.704 + ], + "angle": 0, + "content": "[48] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 436–454. Springer, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[49] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[50] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[51] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[52] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1326–1335, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[53] Zhirong Wu, Alexei A Efros, and Stella X Yu. Improving generalization via scalable neighborhood component analysis. In Proceedings of the European conference on computer vision (ECCV), pages 685-701, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[54] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[55] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.469, + 0.4 + ], + "angle": 0, + "content": "[56] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.403, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[57] Weiyi Yang, Richong Zhang, Junfan Chen, Lihong Wang, and Jaein Kim. Prototype-Guided Pseudo Labeling for Semi-Supervised Text Classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16369-16382, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.502, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[58] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In European conference on computer vision, pages 37-54. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.572, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning Deep Features for Discriminative Localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.643, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[60] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.686, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[61] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2582-2593, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18962" + } + ] +] \ No newline at end of file diff --git a/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_origin.pdf b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..77cda8cb09e357a1526efb0dd6577fbb53269184 --- /dev/null +++ b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/9ca95c3c-3225-4d8b-a7dd-d7e7b707e693_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9541e87165a748db14f5e403e4649cbb053ef4610bfed19b7fae9af7502cd2f1 +size 2158987 diff --git a/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/full.md b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c561e24f4c5d792e253fe5084b27c4ba7845d910 --- /dev/null +++ b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/full.md @@ -0,0 +1,382 @@ +# Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling + +Kranthi Kumar Rachavarapu IIT Madras + +kranthi.rachavarapu@gmail.com + +Kalyan Ramakrishnan +University of Oxford + +kalyanr@robots.ox.ac.uk + +Rajagopalan A. N. IIT Madras + +raju@ee.iitm.ac.in + +# Abstract + +In this paper, we address the weakly-supervised Audio-Visual Video Parsing (AVVP) problem, which aims at labeling events in a video as audible, visible, or both, and temporally localizing and classifying them into known categories. This is challenging since we only have access to video-level (weak) event labels when training but need to predict event labels at the segment (frame) level at test time. Recent methods employ multiple-instance learning (MIL) techniques that tend to focus solely on the most discriminative segments, resulting in frequent misclassifications. Our idea is to first construct several "prototype" features for each event class by clustering key segments identified for the event in the training data. We then assign pseudo labels to all training segments based on their feature similarities with these prototypes and re-train the model under weak and strong supervision. We facilitate this by structuring the feature space with contrastive learning using pseudo labels. Experiments show that we outperform existing methods for weakly-supervised AVVP. We also show that learning with weak and iteratively re-estimated pseudo labels can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure. + +# 1. Introduction + +In this paper, we explore the problem of weakly-supervised Audio-Visual Video Parsing (AVVP) [48], where the goal is to classify events in a video and localize them over time and modalities, given only video-level (weak) event labels for training. Such a formulation is attractive since it forgoes the need for expensive and tedious fine-grained labeling. Moreover, other weakly-supervised video-based temporal prediction tasks can be posed as special cases of AVVP. For instance, Audio-Visual Event Localization (AVEL) [47] considers events that are simultaneously audible and visible, whereas Temporal Action Localization (TAL) [10] considers only visible events. AVVP is challenging as an event + +![](images/c60c9561d79fd0886536a15974b8956a7f3bae7f9a2fd150691401d3c3d834a7.jpg) +Figure 1. Example predictions of a model trained using MIL under weak supervision [48] and our method. For clarity, we only show visual events. The MIL model detects the event "Guitar" correctly in the beginning, but there are some mistakes (red dotted boxes) later on. Moreover, although the last three frames are similar, they are labeled inconsistently by the model. Our method consistently localizes well. "GT" is the ground truth. + +could occur in just the audio stream, visual stream, or both at the same time. Audio-visual data is ubiquitous, with highly variable events in terms of appearance and duration. + +Recent methods for weakly-supervised AVVP [30, 48] use multiple-instance learning (MIL) [7] techniques that optimize for video-level labels during training and make frame-level predictions during evaluation. Here, video-level predictions are obtained by pooling (aggregating) the frame-level predictions. A drawback of MIL approaches is that a model may learn to focus solely on the most discriminative ("prominent") instances (here, audio/visual frames in a video) and still minimize the training loss successfully. This means such models may not reliably detect the full temporal extent of events. Fig. 1 shows an example where an MIL model [48] for AVVP fails to localize events accurately. Moreover, the model labels frames that are visually similar differently. Another reason for misclassifications in MIL models is that they typically contain a frame-level classifier with a single weight vector per event class. However, this may be insufficient to capture the intra-class variation + +for the event across time, modalities, and examples while training with weak labels. + +We propose to address these issues with the following ideas. First, given that the MIL model can find discriminative frames that trigger the video-level labels, we can identify such frames and propagate their predicted labels to other frames in the training data based on similarities in the feature space, thus creating frame-level pseudo labels on the training data. Second, we can use their features to construct multiple prototypes per event class, ensuring a more robust representation to propagate pseudo labels with. Once this is done for the training data, we can re-train the model with strong supervision from pseudo labels and weak supervision from video-level labels. Fig. 1 shows an example where our method produces more accurate predictions than the MIL model. Because our method relies on feature similarities, we further enforce structure in the feature space via contrastive learning using pseudo labels. + +We re-estimate pseudo labels on the training data after every epoch and train with them over the next epoch. We show that such a training procedure can be viewed through an expectation-maximization (EM) lens with the unknown frame-level labels as latent variables. This interpretation provides convergence guarantees and further insight. Our method achieves state-of-the-art performance for weakly-supervised AVVP and also works well on the related weakly-supervised Temporal Action Localization (TAL) task. We summarize our contributions below: + +1. We introduce a prototype-based pseudo-labeling method for weakly-supervised AVVP that exploits similarities in feature space. +2. We show how to perform contrastive learning using pseudo labels to help semantically structure the feature space. +3. We present an expectation-maximization view of our general training procedure. +4. We show state-of-the-art results for weakly-supervised AVVP, and on several metrics for weakly-supervised TAL. + +# 2. Related Work + +Audio-Visual Video Parsing (AVVP). Tian et al. [48] formulated AVVP as a multimodal multiple-instance learning (MMIL) problem and proposed a hybrid attention network with a learned pooling function (HAN) to capture unimodal and cross-modal contexts. Most subsequent work has used the HAN architecture with other techniques to improve performance. Lamba et al. [30] utilized cross-modal information to learn better representations with adversarial and self-supervised losses. Wu et al. [52] estimated modality-level labels by exchanging the audio/visual streams between pairs of videos, followed by retraining under stronger supervision. Lin et al. [32] used cross-video and cross-modality + +signals, such as event co-occurrence across modalities, as additional sources of supervision. Mo et al. [36] leveraged multi-modal grouping to learn discriminative subspaces. Gao et al. [11] extract the "presence" and "absence" evidence for events from uni- and cross-modal information. While Fu et al. [9] focused on learning rate imbalance between modalities and Zhou et al. [60] utilized CLIP [41] to get segment-labels, our work leverages in-model feature similarities for pseudo labeling. PoiBin [40] modeled the number of positive frames for an event as a latent variable to improve localization. Different from previous methods, we estimate segment-level pseudo labels and re-train under full supervision. + +Pseudo-Labeling refers to estimating labels for unlabeled data using the predictions of a trained model. This has been used to improve performance on several weakly-supervised tasks including object detection [1, 46, 59] and image classification [2, 13, 21]. A few works [34, 39, 58] have employed pseudo-labeling to improve performance on Temporal Action Localization (TAL), generating labels from model outputs or attention weights. Luo et al. [34] posed the TAL task as expectation-maximization, using pseudolabels to approximate the E and M steps. Contrary to these, we propose a pseudo-labeling strategy using feature similarities between segments guided by the model predictions and weak labels for a video. + +Prototype-based Classification. Prototype learning with a nearest-neighbor (non-parametric) classifier has inspired various machine learning techniques [5, 12, 15, 43, 44]. Recently, there has been increased interest in integrating this into deep learning methods, e.g., zero-shot [25], few-shot [45], unsupervised [54, 55], and supervised [16, 35, 53, 56] learning. Prototype learning has also been widely used for image segmentation [8, 50, 51, 61]. Recently, prototype-based methods have also been used for semi-supervised image [37] and text [57] classification. A few works have used prototype-based methods with multiple-instance learning. E.g., Rymarczyk et al. [42] processed large medical images as a bag of patches, and Huang et al. [22] learned prototypes for the TAL task with graph convolutional networks. These works adopt an embedding-level MIL approach ([24]) with one prototype per class, while we adopt instance-level MIL and construct multiple prototypes per event class. + +# 3. Background + +# 3.1. Problem Definition + +In the audio-visual video parsing (AVVP) problem, we aim to identify audible or visible events in an unconstrained video and simultaneously localize them in time. Specifically, given a video $x$ consisting $T$ non-overlapping temporal segments $\{x_{t} = (x_{t}^{a},x_{t}^{v})\}_{t = 1}^{T}$ of the audio $(A)$ and visual $(V)$ streams, our objective is to classify each segment + +into $C$ possible events (e.g., singing, vehicle moving, etc.). At test time, we need to predict segment-level event labels $y_{t} = (\mathbf{y}_{\mathbf{t}}^{\mathbf{a}}, \mathbf{y}_{\mathbf{t}}^{\mathbf{v}})$ for each audio and visual segment, where $\mathbf{y}_{\mathbf{t}}^{\mathbf{a}}, \mathbf{y}_{\mathbf{t}}^{\mathbf{v}} \in \{0,1\}^{C}$ are segment-level audio and visual event labels, respectively. + +Weak Supervision. In weakly-supervised AVVP, for each video $x$ , we only have access to the corresponding video-level event label vector $\mathbf{w} = (w_{1},\dots ,w_{C})\in \{0,1\}^{C}$ where $w_{c} = 1$ if any of the segments in the video contains the $c$ -th event, and $w_{c} = 0$ otherwise. Such weak, video-level labels only indicate whether an event occurs in the video or not. During evaluation, we still need to predict segment-level labels $\{(\mathbf{y}_{\mathbf{t}}^{\mathbf{a}},\mathbf{y}_{\mathbf{t}}^{\mathbf{v}})\}_{t = 1}^{T}$ . Note that multiple events can simultaneously occur in a video, i.e., $\sum_{c}w_{c}\geq 1$ , and each segment could have more than one event. + +# 3.2. MIL Framework + +We adopt an instance-level MIL approach [24] as done by prior work [30, 32, 36, 48, 52]: First, each instance (here, an audio/visual segment) is assigned a classification score (here, a $C$ -dimensional vector). Then, the scores are aggregated by a pooling operation, and the aggregate is used in the loss. In our case, this looks like + +$$ +\begin{array}{l} \left[ \left\{\mathbf {f} _ {\mathbf {t}} ^ {\mathbf {a}} \right\} _ {t = 1} ^ {T}, \left\{\mathbf {f} _ {\mathbf {t}} ^ {\mathbf {v}} \right\} _ {t = 1} ^ {T} \right] = f _ {\theta} (x) (1) \\ \mathbf {p} _ {\mathbf {t}} ^ {\mathbf {m}} = g _ {\phi} \left(\mathbf {f} _ {\mathbf {t}} ^ {\mathbf {m}}\right) \in [ 0, 1 ] ^ {C}, \forall t \in [ 1, T ], m \in \{a, v \} (2) \\ \hat {\mathbf {w}} = \sigma_ {\theta , x} \left(\left\{\mathbf {p} _ {\mathbf {t}} ^ {\mathbf {a}} \right\} _ {t = 1} ^ {T}, \left\{\mathbf {p} _ {\mathbf {t}} ^ {\mathbf {v}} \right\} _ {t = 1} ^ {T}\right) \in [ 0, 1 ] ^ {C}, (3) \\ \end{array} +$$ + +where $f_{\theta}$ is a feature extractor (encoder), $g_{\phi}$ is a feature classifier, $\sigma_{\theta,x}$ is a MIL-pooling operator. The predicted video-level probability vector $\hat{\mathbf{w}}$ is computed from segment-level classification scores $\{\mathbf{p}_t^m\}$ , which are computed from segment features $\{\mathbf{f}_{\mathbf{t}}^{\mathbf{m}}\}$ . We adopt the Hybrid Attention Network (HAN) [48] architecture to implement $f_{\theta}, g_{\phi}$ , and $\sigma$ . Here, $f_{\theta}$ consists of a transformer-based encoder with cross-modal attention and $g_{\phi}$ is a linear classifier with the sigmoid activation. Attention-based averaging is used for $\sigma$ . We provide implementation details in the Supplementary. + +This model is trained end-to-end with supervision at the video level by minimizing a weighted cross-entropy loss + +$$ +\mathcal {L} _ {\mathrm {M I L}} = \operatorname {C E} (\hat {\mathbf {w}}, \mathbf {w}) \tag {4} +$$ + +with weights inversely proportional to the event distribution in the dataset. At test time, we get segment-level predictions from Eq. (2). + +# 4. Prototype-based Pseudo-Labeling + +While instance-level MIL enables finding key segments that trigger the video-level labels [24, 33], they often miss out on the full temporal extent of events in a video. However, if we identify such discriminative segments for each label + +in a training video, we could use them to build a set of prototype features that represent each class. A high degree of similarity of a segment with the prototype features for an event class means the segment probably contains the event. We use this idea to generate pseudo labels for all segments in a training video and then re-train the model under strong supervision. Fig. 2 shows an overview of our approach. We provide pseudocode in Algorithm 1. + +# 4.1. Constructing Prototype Features + +The segment-level linear classifier $g_{\phi}$ (Eq. (2)) can itself be considered a prototype-based classifier, where the prototype for each class is the corresponding weight vector in $\phi$ . However, this requires that an event have similar representations across modalities and examples in the training data. We instead construct multiple prototypes for each class in the following two steps. + +Segment Selection. For each training video $x$ , we find positive segments $\mathcal{P}_x(c)$ for a target event $c$ (if $w_c = 1$ ) as those the model is highly confident contain the event. I.e., + +$$ +\mathcal {P} _ {x} (c) = \left\{ \begin{array}{l l} \left\{x _ {t} ^ {m} \in x \mid \mathbf {p} _ {\mathbf {t}} ^ {\mathbf {m}} (c) \geq \gamma p ^ {*} (c) \right\} & \text {i f} w _ {c} = 1 \\ \phi & \text {i f} w _ {c} = 0, \end{array} \right. \tag {5} +$$ + +where $\gamma \in (0,1]$ is a hyperparameter and $p^*(c) = \max_{t,m} \mathbf{p}_{\mathbf{t}}^{\mathbf{m}}(c)$ is the maximum predicted probability for the event over time and modalities for a given video. Clearly, $\mathcal{P}_X(c)$ will contain at least one audio/visual segment from video $x$ when $w_c = 1$ . We assume these segments contain event $c$ with high probability (so are likely true positives). We also find negative segments $\mathcal{N}_x(c)$ for a target event $c$ as those the model is highly confident do not contain the event when $w_c = 1$ (likely true negatives), or is highly confident to contain the event when $w_c = 0$ (definite false positives): + +$$ +\mathcal {N} _ {x} (c) = \left\{ \begin{array}{l l} \left\{x _ {t} ^ {m} \in x \mid \mathbf {p} _ {\mathbf {t}} ^ {\mathbf {m}} (c) \leq \beta p ^ {*} (c) \right\} & \text {i f} w _ {c} = 1 \\ \left\{x _ {t} ^ {m} \in x \mid \mathbf {p} _ {\mathbf {t}} ^ {\mathbf {m}} (c) = p ^ {*} (c) > = 0. 5 \right\} & \text {i f} w _ {c} = 0, \end{array} \right. \tag {6} +$$ + +where $\beta \in (0,1)$ is a hyperparameter set to a small value to retrieve segments without the event $c$ (if any). Note that $\mathcal{N}_x(c)$ may contain no segments at all even when $w_{c} = 1$ , i.e., be a null set. + +Finally, we aggregate the positive and negative sets over all videos in training data $\mathcal{D}$ to get prototypical segments as + +$$ +\mathcal {P} (c) = \bigcup_ {x \in \mathcal {D}} \mathcal {P} _ {x} (c), \quad \mathcal {N} (c) = \bigcup_ {x \in \mathcal {D}} \mathcal {N} _ {x} (c). \tag {7} +$$ + +Prototype Generation. We now generate multiple prototypes (feature vectors) for each event class by hypothesizing that having more than one prototype would help better capture the intra-class variation across modalities and training examples, and create a more robust representation. + +![](images/af2047dd15287e6043608b15af802cf7050fae0e8f2582d8bc37a28b035c7955.jpg) +Figure 2. An overview of our method. The top branch is trained using MIL with weak labels and using full supervision with pseudo labels, which are generated by the bottom branch. This is done by first selecting key segments for each event from the training data and clustering them to construct a set of prototype features for the event, used to guide pseudo-label generation. + +First, we extract features for all segments in $\mathcal{P}(c)$ and $\mathcal{N}(c)$ using Eq. (1), and call the resulting feature sets $\mathcal{F}_{\mathcal{P}}(c)$ and $\mathcal{F}_{\mathcal{N}}(c)$ . Here, instead of $f_{\theta}$ , we use a momentum encoder [19] $f_{\tilde{\theta}}$ having the same architecture but whose weights are set (after each update to $\theta$ ) as an exponential moving average (EMA) of $\theta$ over training steps: $\tilde{\theta} \gets (1 - \eta)\tilde{\theta} + \eta\theta$ , where $\eta \in (0,1)$ . The momentum encoder is more stable and less sensitive to noise from weak supervision, leading to slightly improved performance. + +Next, we perform $k$ -means clustering separately on $\mathcal{F}_{\mathcal{P}}(c)$ and $\mathcal{F}_{\mathcal{N}}(c)$ to get two sets of cluster centroids. We call the centroids prototypes: + +$$ +\mathcal {C} _ {\mathcal {P}} (c) = \left\{\mathbf {f} _ {\mathbf {i}} ^ {+} \right\} _ {i = 1} ^ {k _ {p}}, \quad \mathcal {C} _ {\mathcal {N}} (c) = \left\{\mathbf {f} _ {\mathbf {j}} ^ {-} \right\} _ {j = 1} ^ {k _ {n}}. \tag {8} +$$ + +Clustering for finding prototypes imposes a natural bottleneck [26], helping discard irrelevant information in the segments. We repeat this for each event class. + +# 4.2. Estimating Pseudo Labels + +We assign pseudo labels to each segment in a training video by considering its feature similarities with the prototypes for each class. We assume that if an audio/visual segment contains event $c$ , it would be on average closer to $\mathcal{P}(c)$ in feature space than to $\mathcal{N}(c)$ . For this, we explore two approaches: (i) hard labeling, which assigns the label of the closest-on-average set among the prototype sets for class $c$ : + +$$ +\mathbf {y} _ {\mathbf {t}} ^ {\mathbf {m}} (c) = w _ {c} \cdot \mathbb {1} \left[ \frac {\sum_ {\mathbf {f} ^ {+} \in \mathcal {C} _ {\mathcal {P}} (c)} e ^ {(\mathbf {f} ^ {+}) ^ {T} \mathbf {f} _ {\mathbf {t}} ^ {\mathbf {m}}}}{| \mathcal {C} _ {\mathcal {P}} |} > \frac {\sum_ {\mathbf {f} ^ {-} \in \mathcal {C} _ {\mathcal {N}} (c)}}{| \mathcal {C} _ {\mathcal {N}} |} \right], \tag {9} +$$ + +and (ii) soft labeling, which instead assigns a probability score: + +$$ +\mathbf {y} _ {\mathbf {t}} ^ {\mathbf {m}} (c) = w _ {c} \cdot \frac {\frac {1}{| \mathcal {C} _ {\mathcal {P}} |} \sum_ {\mathbf {f} ^ {+} \in \mathcal {C} _ {\mathcal {P}} (c)} e ^ {\left(\mathbf {f} ^ {+}\right) ^ {T} \mathbf {f} _ {\mathbf {t}} ^ {\mathbf {m}} / \tau}}{\sum_ {\mathcal {S} \in \left\{\mathcal {C} _ {\mathcal {P}}, \mathcal {C} _ {\mathcal {N}} \right\}} \frac {1}{| \mathcal {S} |} \sum_ {\mathbf {f} ^ {\prime} \in \mathcal {S} (c)} e ^ {\left(\mathbf {f} ^ {\prime}\right) ^ {T} \mathbf {f} _ {\mathbf {t}} ^ {\mathbf {m}} / \tau}} \in (0, 1), \tag {10} +$$ + +where $\tau > 0$ is a temperature hyperparameter. + +When training our model, we generate pseudo labels $\{\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}\}$ after every epoch for each training video and use them to fully supervise the segment-level probabilities $\{\mathbf{p}_{\mathbf{t}}^{\mathbf{m}}\}$ in Eq. (2) using the binary cross-entropy loss + +$$ +\mathcal {L} _ {\mathrm {P L}} = \operatorname {C E} \left(\mathbf {p} _ {\mathbf {t}} ^ {\mathbf {m}}, \mathbf {y} _ {\mathbf {t}} ^ {\mathbf {m}}\right), \tag {11} +$$ + +averaged over segments and training examples. + +# 4.3. Contrastive Learning with Pseudo Labels + +Given that our method exploits feature similarities between segments based on the events they may contain, we use contrastive learning based on pseudo labels to help semantically structure the feature space. Since multiple events can occur simultaneously, for each audio/visual segment, we consider a similar segment as one estimated to have some common event(s) and a dissimilar segment as one estimated to have no common events. Formally, given a batch of training examples, let $B = \{\mathbf{f}_{\mathbf{t}}^{\mathbf{m}} \to \mathbf{y}_{\mathbf{t}}^{\mathbf{m}}\}$ be a dictionary from features to the assigned pseudo labels for all segments and over all videos in the batch. For every feature $\mathbf{f} \in B$ , the similar feature-set is then $S(\mathbf{f}) = \{\mathbf{g} \in B \mid B(\mathbf{g}) \cap B(\mathbf{f}) \neq \phi\}$ and the dissimilar set is $D(\mathbf{f}) = \{\mathbf{g} \in B \mid B(\mathbf{g}) \cap B(\mathbf{f}) = \phi\}$ . To pull segments with a common predicted event closer in the feature space, we use the contrastive loss [38] + +$$ +\mathcal {L} _ {\mathrm {C L}} = - \log \frac {\sum_ {\mathbf {g} \in \mathcal {S} (\mathbf {f})} e ^ {\mathbf {g} ^ {T} \mathbf {f} / \tau_ {c}}}{\sum_ {\mathbf {g} \in \mathcal {S} (\mathbf {f})} e ^ {\mathbf {g} ^ {T} \mathbf {f} / \tau_ {c}} + \sum_ {\mathbf {h} \in \mathcal {D} (\mathbf {f})} e ^ {\mathbf {h} ^ {T} \mathbf {f} / \tau_ {c}}}, \tag {12} +$$ + +
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
AVE* [47]47.240.437.134.735.431.639.935.541.636.5
AVSDN* [31]47.834.152.046.337.126.545.735.650.837.7
HAN [48]60.151.352.948.948.943.054.047.755.448.0
MA [52]60.353.660.056.455.149.058.953.057.950.6
CVCM-MA [32]60.853.863.558.957.049.560.554.059.552.1
MGN-MA [36]60.250.961.959.755.549.659.253.458.749.9
JoMoLD [4]61.353.963.859.957.249.660.854.559.952.5
PoiBin [40]63.154.163.560.357.751.561.455.260.652.3
CMPAE [11]64.256.666.463.759.251.863.357.462.855.7
Ours65.957.366.764.361.954.364.859.963.757.9
+ +Table 1. Results (% F1-scores) on all metrics defined in Tian et al. [48]. \* indicates reproduced performance. The best and second best results are in bold and underline, respectively. + +Algorithm 1: Pseudocode for our method +1 Init: model params $\psi = \{\theta ,\phi \}$ , EMA params $\psi_{E} = \{\theta_{E},\phi_{E}\}$ +2 while $\psi$ has not converged do +3 #estimate segment - level pseudo labels +4 for $(x,\mathbf{w})$ in train set do +5 $\left|\{\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}\} \gets p_{\theta}(\cdot |\mathbf{w},x)\right.;$ +6 end +7 #train +8 for $(x,\mathbf{w})$ in train set do +9 $\{\mathbf{p}_{\mathbf{t}}^{\mathbf{m}}\} \gets p_{\theta ,\phi}(\cdot |x);$ +10 $\hat{\mathbf{w}} = \sigma_{\theta ,x}(\{\mathbf{p}_{\mathbf{t}}^{\mathbf{m}}\});$ +11 $\mathcal{L} = \mathcal{L}_{\mathrm{MIL}} + \lambda_{\mathrm{PL}}\mathcal{L}_{\mathrm{PL}} + \lambda_{\mathrm{CL}}\mathcal{L}_{\mathrm{CL}};$ +12 $\psi \gets \psi -\alpha \nabla_{\psi}\mathcal{L};$ +13 $\psi_{E}\gets (1 - \eta)\psi_{E} + \eta \psi ;$ +14 end +15 end + +averaged over all $\mathbf{f} \in B$ , where $\tau_{c} > 0$ is a temperature hyperparameter. + +# 4.4. Training and Inference + +We train our model using the following loss in an end-to-end fashion: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {M I L}} + \lambda_ {\mathrm {P L}} \mathcal {L} _ {\mathrm {P L}} + \lambda_ {\mathrm {C L}} \mathcal {L} _ {\mathrm {C L}}, \tag {13} +$$ + +where $\lambda_{\mathrm{PL}}, \lambda_{\mathrm{CL}} > 0$ are hyperparameters. During training, we generate pseudo labels once every epoch for all training examples using the best model so far. During inference, we predict the segment-level probabilities using the non-parametric prototypical pseudo-labeling (Eq. (10)) with video-level predictions from Eq. (3). + +# 5. Experiments + +# 5.1. Setup + +LLP Dataset. We conduct experiments on the Look, Listen and Parse (LLP) dataset [48] which consists of 11849 YouTube videos, each of 10s duration, labeled into 25 event categories. These videos are unconstrained with a wide variety of scene content, including daily activities, music performances, and vehicle sounds. We use 10000 videos with weak (video-level) labels for training. The remaining 1849 fully-annotated videos (with segment-level labels) are used for validation and testing. The train/val/test split has been provided in the LLP dataset. + +Evaluation Metrics. Following previous work, we use F1-scores on audio, visual, and audio-visual events as evaluation metrics. These are computed both at the segment and event level. We also include the aggregate metrics "Type@AV" and "Event@AV", again computed at the segment and event level. See Tian et al. [48] for a full explanation of metrics. + +Implementation Details. We use ResNet-152 [18] pretrained on ImageNet [6] and R(2+1)D-18 [49] pre-trained on Kinetics-400 [27] as visual feature extractors to generate a 512-d visual feature per segment. We use VGGish [20] pre-trained on AudioSet [14] to generate a 128-d audio feature per segment. We use the Adam optimizer [28] with a batch size of 16 and a learning rate $\alpha = 3e - 4$ for 50 epochs. We take $\eta = 0.999$ for the momentum encoder, $\lambda_{\mathrm{PL}} = \lambda_{\mathrm{CL}} = 0.1$ in Eq. (13), $\gamma = 1$ in Eq. (5), $\beta = 0.2$ in Eq. (6), $\tau = 0.1$ in Eq. (10), and $\tau_{c} = 0.2$ in Eq. (12). We set $k_{p} = k_{n} = 10$ in Eq. (8). Hyperparameter values are based on validation performance. + +# 5.2. Comparisons with Prior Work + +For a fair comparison, all methods considered use the same pre-trained feature extractors and train/val/test split. + +Quantitative. We compare methods in Table 1 on all the + +![](images/8e504539dcc566fdaa74ed92febaad0af3d6b78d86b3045246478c4fb7b6d264.jpg) +Figure 3. Audio-Visual Video Parsing results of our methods with HAN [48], MA [52], JoMoLD [4] and CMPAE [11] on two videos. + +![](images/7ab05ab4e5bc3d13898ad66819dcb2888e7b518db9b16f2918442eb6334277a3.jpg) + +
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
Base+NPLH58.952.158.549.855.748.257.850.155.746.6
Base+NPILS59.352.559.051.156.048.758.250.856.247.6
Base+PPLH62.557.462.959.659.654.061.956.360.355.7
Base+PPLS65.957.366.764.361.954.364.859.963.757.9
+ +Table 2. Performance for different pseudo labeling strategies. + +metrics listed earlier. We see that our method outperforms others on audio, audio-visual, Type@AV, and Event@AV scores at both the segment and event levels. In particular, we achieve an average improvement of 2.6 points for audio-visual events. Averaged over all metrics, our method achieves a 1.86 point improvement over the previous best method CMPAE [11]. Since we do not require additional learnable parameters to generate pseudo labels, this improvement is with a comparable number of parameters to previous methods. + +Qualitative. We qualitatively compare our method with some previous methods on two examples in Fig. 3. Here, Fig. 3 (left) contains events "Singing" and "Banjo" in both modalities. Fig. 3 (Right) shows another scene with the event "Speech" in only the audio modality, and "Violin" occurring in both modalities. As evident, the common failure mode of previous methods is that they detect only a portion of the duration of events, while ours more commonly detects the full extent of events. As discussed in Sec. 1, this could be because of multiple-instance learning encouraging the model to pick out only the most discriminative segments containing an event, as opposed to all segments. + +Computational Cost. When training, our approach takes + +67.4s per epoch (17.1s for forward+backward propagation and 50.3s for prototype-based pseudo labeling). Inference time is 5.9ms on average per video, which is comparable to the baselines (2.9ms). + +# 5.3. Analysis + +Pseudo-Labeling Strategies. We report an ablation study in Table 2 to validate our specific pseudo-labeling method. Here, Base is the baseline MIL model explained in Sec. 3.2. Base+NPL refers to the base model trained with naive pseudo labeling, i.e., by directly using the segment-level predictions of the model as pseudo labels for retraining. In this case, H and S refer to using hard (binarized) and soft (continuous) pseudo labels, respectively. We observe no consistent improvement with naive pseudo labeling, likely due to the MIL model producing incorrect predictions for the non-discriminative segments for an event, leading to less reliable pseudo labels. Base+PPL refers to the base model trained with prototype-based pseudo labels as described in Sec. 4.2. In this case, H and S refer to using hard and soft pseudo labels, as defined in Eqs. (9) and (10), respectively. Our method significantly improves performance by generating more reliable pseudo labels throughout a training video. We also see that soft labeling is better than hard labeling. Since segment-level predictions may be noisy, we expect mistakes to affect performance more with hard rather than soft labels. + +Reliability of Pseudo Labels. We now assess if the pseudo labels generated by our method are reliable for re-training. For this, we take the pre-trained baselines HAN [48] and MA [52], generate prototypes for each event from the train- + +
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
HAN60.151.352.948.948.943.054.047.755.448.0
+PPLTest62.555.455.351.152.346.956.050.958.350.6
MA60.353.660.056.455.149.058.953.057.950.6
+PPLTest61.755.461.857.957.551.660.655.059.452.6
+ +![](images/eb7db5efc731b7e1a61941d5d00977d2b01df0930cf6536e7539110cc1ca8d8c.jpg) +Figure 4. Performance for different $\gamma$ and $\beta$ choices. + +![](images/280184cfa4828bba9fb6a9b25a9cf6dde190c3fe77f86a6adab73f7476b677a5.jpg) + +Table 3. Evaluating prototype-based pseudo labels on the test set. + +
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
No-CL64.156.459.657.357.149.661.256.460.856.8
With-CL65.957.366.764.361.954.364.859.963.757.9
+ +ing data, and simply predict the prototype-based pseudo labels (Eq. (9)) on the test set for evaluation (i.e., discard the classifiers at test time). We call the resulting classifiers $\mathrm{HAN} + \mathrm{PPL}_{\mathrm{Test}}$ and $\mathrm{MA} + \mathrm{PPL}_{\mathrm{Test}}$ and show results in Table 3. With no other change to the models other than the classifiers, we see significant improvements in all metrics, showing that our pseudo labels are indeed reliable estimates of the true labels. Note that for the results in Table 1, we recompute pseudo labels on the training data every epoch, with a continually improving base model. + +Effect of $\gamma$ and $\beta$ . The hyperparameters $\gamma$ and $\beta$ in Eqs. (5) and (6) roughly control the proportion of segments that will be selected as positives and negatives for assigning pseudo labels. We experiment with different values and plot the average of all segment/event-level metrics in Fig. 4. Performance is susceptible to $\gamma$ , which controls positives, and is tolerant to a slight variation in $\beta$ . The best performance is for $\gamma = 1$ and $\beta = 0.2$ . + +Effect of Contrastive Learning. We analyze the impact of contrastive learning using pseudo labels (Eq. (12)) by simply setting $\lambda_{\mathrm{CL}} = 0$ in Eq. (13). Results are reported in Tab. 4. While only training with prototype-based pseudo labels $(\mathsf{No - CL})$ significantly improves the baseline, adding a contrastive loss based on our pseudo labels (With-CL) further yields a large improvement. + +Number of Clusters. We take $k_{p} = k_{n}$ and try values in $\{1,2,\dots,100\}$ . We plot the average of all segment/event-level metrics in Fig. 5. The best performance was for $k = 10$ clusters per class. Having fewer clusters helps cap + +![](images/57b6c0c53107decadc8ad13f2930dc7c42130d5eef81b2fd1eeb06c059496e3f.jpg) +Figure 5. Performance for different choices of the number of clusters/prototypes per event. + +Table 4. Effect of contrastive learning (CL) with pseudo labels. + +
MethodtIoUAvg.
0.10.20.30.40.50.60.7[0.1:0.5][0.3:0.7][0.1:0.7]
EM-MIL [34]59.152.745.536.830.522.716.444.930.437.7
ASM-Loc [17]71.265.557.146.836.625.213.455.435.845.1
DELU [3]71.566.256.547.740.527.215.356.537.446.4
Ours72.766.957.946.937.426.820.156.437.846.9
+ +Table 5. Results (\% mAP) for the TAL task at different tIoU thresholds. + +ture more general features for an event instead of instance-specific details, but having too few hurts performance by discarding intra-class variation. + +# 5.4. Results for a Different Task + +We conduct preliminary experiments for the weakly-supervised Temporal Action Localization (TAL) task [10] on the THUMOS14 dataset [23] to further validate our method. TAL aims to temporally localize and classify visual events in videos into known categories. We report mean average precisions (mAP) at different temporal intersection over union (tIoU) thresholds and the results are in Table 5. We achieve better or comparable performance w.r.t. the current state-of-the-art method DELU [3], and outperform previous methods. Check the supplementary for a full comparison. + +# 6. Interpretation as Expectation-Maximization + +We now show that iteratively re-estimating soft pseudo labels and training the model under weak $+$ strong supervision (Eq. (13)) can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure. For this, we view the segment-level labels $y = \{(\mathbf{y}_{\mathbf{t}}^{\mathbf{a}},\mathbf{y}_{\mathbf{t}}^{\mathbf{v}})\}_{t = 1}^{T}$ for each video as latent variables since they are not observed when training under weak supervision. The observed variables are the video segments $x = \{(x_{t}^{a},x_{t}^{v})\}_{t = 1}^{T}$ and weak labels $\mathbf{w}$ . For clarity, assume a single event class, i.e., $\mathbf{w},\mathbf{y}_{\mathbf{m}}^{\mathbf{t}}\in \{0,1\}$ . + +Our goal is to maximize the data log-likelihood $\mathcal{L}(\theta) = \ln p_{\theta}(\mathbf{w}|x)$ over model parameters $\theta$ (assume just one data point $x$ for clarity). Instead of maximizing this directly, we + +can attempt to maximize the variational lower bound [29] $\mathcal{L}_{\mathrm{VLB}}$ for this likelihood (random variables are capitalized): + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {V L B}} (\theta , \phi) = \underset {Y \sim q _ {\phi} (\cdot | \mathbf {w}, x)} {\mathbb {E}} [ \ln p _ {\theta} (\mathbf {w} | Y, x) ] \\ - \mathrm {D} _ {\mathrm {K L}} \left[ q _ {\phi} (\cdot | \mathbf {w}, x) \| p _ {\theta} (\cdot | x) \right], \tag {14} \\ \end{array} +$$ + +where $q_{\phi}(y|\mathbf{w},x)$ is any conditional distribution over labels $y$ , with parameters $\phi$ , meant to approximate the true posterior $p_{\theta}(y|x)$ . The EM algorithm alternates between two steps, first approximating the posterior by maximizing $\mathcal{L}_{\mathrm{VLB}}$ over $\phi$ (E step), and then improving the data likelihood by maximizing it over $\theta$ (M step). Let $j \geq 1$ index the iterations (note, $\theta$ and $\phi$ here are not the same as in Sec. 3.2). + +E Step. The $j$ -th update is given by + +$$ +q _ {\phi_ {j}} (y | \mathbf {w}, x) = p _ {\theta_ {j - 1}} (y | \mathbf {w}, x), \quad \forall y. \tag {15} +$$ + +Assuming conditional independence of the segment-level labels given the video and its weak labels, we can express the right-hand side as $p_{\theta}(y|\mathbf{w},x) = \prod_{t,m}p_{\theta}(\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}|\mathbf{w},x)$ , computed using + +$$ +p _ {\theta} \left(\mathbf {y} _ {\mathbf {t}} ^ {\mathbf {m}} \mid \mathbf {w}, x\right) := \left\{ \begin{array}{l l} p _ {\theta} \left(\mathbf {y} _ {\mathbf {t}} ^ {\mathbf {m}} \mid x\right) & \text {i f} \mathbf {w} = 1 \\ 1 - \mathbf {y} _ {\mathbf {t}} ^ {\mathbf {m}} & \text {i f} \mathbf {w} = 0, \end{array} \right. \tag {16} +$$ + +where $p_{\theta}(\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}|x)$ represents the prediction of the model for modality $m$ and segment $t$ in video $x$ . Thus, $p_{\theta}(\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}|\mathbf{w},x)$ becomes a soft pseudo label for the segment $x_{m}^{t}$ . + +M Step. The $j$ -th update finds $\theta_{j}$ by maximizing $\mathcal{L}_{\mathrm{VLB}}(\theta, \phi_j)$ , which, due to Eq. (15), is equivalent to + +$$ +\begin{array}{l} \min _ {\theta} - \sum_ {y} p _ {\theta_ {j - 1}} (y | \mathbf {w}, x) \ln p _ {\theta} (y | x) \\ - \sum_ {y} p _ {\theta_ {j - 1}} (y | \mathbf {w}, x) \ln p _ {\theta} (\mathbf {w} | y, x), \tag {17} \\ \end{array} +$$ + +where $p_{\theta}(y|\mathbf{w},x)$ and $p_{\theta}(y|x)$ are computed as in Eq. (16). The first term in Eq. (17) can be expanded as the total cross-entropy between the segment-level predictions $p_{\theta}(\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}|x)$ and the pseudo-label targets $p_{\theta_{j - 1}}(\mathbf{y}_{\mathbf{t}}^{\mathbf{m}}|\mathbf{w},x)$ and is therefore equal to $\mathcal{L}_{\mathrm{PL}}$ (see Eq. (13)). To understand the second term in Eq. (17), note that in our setting, $p_{\theta}(\mathbf{w}|y,x) = p_{\theta}(\mathbf{w}|x)$ since we make video-level predictions using MIL-pooling without access to the segment-level labels $(y)$ . The second term then reduces to $-\ln p_{\theta}(\mathbf{w}|x)$ , which is simply the negative log-likelihood of the video-level label $(\mathbf{w})$ . This means the second term is equal to $\mathcal{L}_{\mathrm{MIL}}$ (see Eq. (13)). I.e., the M step is equivalently + +$$ +\min _ {\theta} - \sum_ {y} p _ {\theta_ {j - 1}} (y | \mathbf {w}, x) \ln p _ {\theta} (y | x) - \ln p _ {\theta} (\mathbf {w} | x). \tag {18} +$$ + +If we simply maximized the data log-likelihood $\mathcal{L}(\theta)$ directly (no pseudo labels), this could converge to a solution that picks out just the most discriminative segments in each video, which may suffice to reduce the video-level loss $\mathcal{L}_{\mathrm{MIL}}$ . As we showed in this paper, using pseudo labels can help mitigate this, but the quality of the estimated pseudo labels has a crucial role. E.g., hard labeling is inferior to soft labeling since it does not attempt to compute the E step (Eq. (15)) exactly. While soft labeling does attempt an exact E step, better inductive biases (such as prototype-based labeling with contrastive learning) here can help reach a favorable solution, as reported in Sec. 5.3 and Table 2. We also compare with an off-the-shelf pseudo-labeling method [34] in the supplementary material and find that we significantly outperform it, showing that it is not pseudo-labeling per se that works for AVVP - how we estimate the pseudo labels is paramount. + +# 7. Conclusion + +We proposed a prototype-based pseudo-labeling method for weakly-supervised AVVP that outperforms existing methods. Some interesting findings are that (i) naive pseudolabeling does not produce reliable training targets, (ii) contrastive learning using pseudo labels is an effective strategy, and (iii) training with weak and re-estimated soft pseudo labels is an EM algorithm. + +We end with some limitations and suggestions for future work. First, performance is sensitive to hyperparameter $\beta$ as shown in Fig. 4. This could be because identifying negatives for a class from the predictions of an MIL model is less likely to be accurate than identifying positives. Moreover, negative segments for an event would have a much larger variation in content. Second, events with larger intra-class variation may need more prototypes to represent, while we assume the same number to be optimal for all events for convenience. Third, while we can interpret our method as an EM algorithm, EM may only converge to a locally optimal solution. As discussed in Sec. 6, future work could explore inductive biases in the pseudolabeling step that favor better solutions. Finally, it is not clear how many related video-/image-based prediction tasks our method can generalize to. While we provided results for TAL in Sec. 5.4, more evidence of adapting our method to related tasks would be useful in future work. + +# Acknowledgement + +Support from Institute of Eminence (IoE) project No. SB22231269EEETWO005001 for Research Centre in Computer Vision is gratefully acknowledged. + +# References + +[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 2 +[2] Ricardo Cabral, Fernando De la Torre, Joao Paulo Costeira, and Alexandre Bernardino. Matrix Completion for Weakly-Supervised Multi-Label Image Classification. IEEE transactions on pattern analysis and machine intelligence, 37(1): 121-135, 2014. 2 +[3] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 192-208. Springer, 2022. 7 +[4] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In Computer Vision - ECCV 2022, pages 431-448, Cham, 2022. Springer Nature Switzerland. 5, 6 +[5] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE transactions on information theory, 13(1): 21-27, 1967. 2 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5 +[7] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the Multiple Instance Problem with Axis-Parallel Rectangles. Artificial Intelligence, 89(1):31-71, 1997. 1 +[8] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, 2018. 2 +[9] Jie Fu, Junyu Gao, Bing-Kun Bao, and Changsheng Xu. Multimodal imbalance-aware gradient modulation for weakly-supervised audio-visual video parsing. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[10] Adrien Gaidon, Zaid Harchaoui, and Cordelia Schmid. Temporal localization of actions with actons. IEEE transactions on pattern analysis and machine intelligence, 35(11):2782-2795, 2013. 1, 7 +[11] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2, 5, 6 +[12] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE transactions on pattern analysis and machine intelligence, 34(3):417-435, 2012. 2 +[13] Weifeng Ge, Xiangru Lin, and Yizhou Yu. Weakly Supervised Complementary Parts Models for Fine-Grained Image Classification from the Bottom Up. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3034-3043, 2019. 2 +[14] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780. IEEE, 2017. 5 +[15] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2 +[16] Samantha Guerriero, Barbara Caputo, and Thomas Mensink. DeepNCM: Deep Nearest Class Mean Classifiers, 2018. 2 +[17] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13925–13935, 2022. 7 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4 +[20] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 iieme international conference on acoustics, speech and signal processing (icassp), pages 131-135. IEEE, 2017. 5 +[21] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly Supervised Image Classification through Noise Regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11517-11525, 2019. 2 +[22] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Relational prototypical network for weakly supervised temporal action localization. In Proceedings of the AAAI conference on artificial intelligence, pages 11053-11060, 2020. 2 +[23] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos "in the wild". Computer Vision and Image Understanding, 155:1-23, 2017. 7 +[24] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 3 +[25] Saumya Jetley, Bernardino Romera-Paredes, Sadeep Jayasumana, and Philip Torr. Prototypical priors: From improving classification to zero-shot learning. arXiv preprint arXiv:1512.01192, 2015. 2 + +[26] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 4 +[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5 +[28] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR (Poster), 2015. 5 +[29] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes, 2022. 8 +[30] Jatin Lamba, Jayaprakash Akula, Rishabh Dabral, Preethi Jyothi, Ganesh Ramakrishnan, et al. Cross-modal learning for audio-visual video parsing. arXiv preprint arXiv:2104.04598, 2021. 1, 2, 3 +[31] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002-2006. IEEE, 2019. 5 +[32] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5 +[33] Guoqing Liu, Jianxin Wu, and Zhi-Hua Zhou. Key Instance Detection in Multi-Instance Learning. In Proceedings of the Asian Conference on Machine Learning, pages 253–268, Singapore Management University, Singapore, 2012. PMLR. 3 +[34] Zhekun Luo, Devin Guillory, Baifeng Shi, Wei Ke, Fang Wan, Trevor Darrell, and Huijuan Xu. Weakly-supervised action localization with expectation-maximization multi-instance learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 729-745. Springer, 2020. 2, 7, 8 +[35] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 2 +[36] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In Advances in Neural Information Processing Systems, 2022. 2, 3, 5 +[37] Islam Nassar, Munawar Hayat, Ehsan Abbasnejad, Hamid Rezatofighi, and Gholamreza Haffari. ProtoCon: Pseudo-Label Refinement via Online Clustering and Prototypical Consistency for Efficient Semi-Supervised Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11641–11650, 2023. 2 +[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding, 2018. 4 + +[39] Alejandro Pardo, Human Alwassel, Fabian Caba Heilbron, Ali Thabet, and Bernard Ghanem. RefineLoc: Iterative Refinement for Weakly-Supervised Action Localization. In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3318-3327, 2021. 2 +[40] Kranthi Kumar Rachavarapu and A N Rajagopalan. Boosting positive segments for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10192–10202, 2023. 2, 5 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[42] Dawid Rymarczyk, Aneta Kaczyńska, Jarosław Kraus, Adam Pardyl, and Bartosz Zielinski. Protomil: Multiple instance learning with prototypical parts for fine-grained interpretability. arXiv e-prints, pages arXiv-2108, 2021. 2 +[43] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood structure. In Artificial intelligence and statistics, pages 412-419. PMLR, 2007. 2 +[44] Atsushi Sato and Keiji Yamada. Generalized learning vector quantization. Advances in neural information processing systems, 8, 1995. 2 +[45] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2 +[46] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple Instance Detection Network with Online Instance Classifier Refinement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2843-2851, 2017. 2 +[47] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, pages 247–263, 2018. 1, 5 +[48] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 436–454. Springer, 2020. 1, 2, 3, 5, 6 +[49] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 5 +[50] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 2 +[51] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 2 + +[52] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1326–1335, 2021. 2, 3, 5, 6 +[53] Zhirong Wu, Alexei A Efros, and Stella X Yu. Improving generalization via scalable neighborhood component analysis. In Proceedings of the European conference on computer vision (ECCV), pages 685-701, 2018. 2 +[54] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2 +[55] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 2 +[56] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 2 +[57] Weiyi Yang, Richong Zhang, Junfan Chen, Lihong Wang, and Jaein Kim. Prototype-Guided Pseudo Labeling for Semi-Supervised Text Classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16369-16382, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[58] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In European conference on computer vision, pages 37-54. Springer, 2020. 2 +[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning Deep Features for Discriminative Localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2 +[60] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2 +[61] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2582-2593, 2022. 2 \ No newline at end of file diff --git a/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/images.zip b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c6f4b777e915d7dfdf9ecca67978ace6dd98b949 --- /dev/null +++ b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce4ce66e48d25b0e31097b1ef0d81895886293ce94cfdf105cc8e101e76efa76 +size 573301 diff --git a/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/layout.json b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..81f0f7963864405b64d4c39382d8b2beb51a161f --- /dev/null +++ b/2024/Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling/layout.json @@ -0,0 +1,10531 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "type": "text", + "content": "Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 162, + 227, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 162, + 227, + 188 + ], + "spans": [ + { + "bbox": [ + 86, + 162, + 227, + 188 + ], + "type": "text", + "content": "Kranthi Kumar Rachavarapu IIT Madras" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 192, + 236, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 192, + 236, + 202 + ], + "spans": [ + { + "bbox": [ + 78, + 192, + 236, + 202 + ], + "type": "text", + "content": "kranthi.rachavarapu@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 271, + 162, + 380, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 162, + 380, + 189 + ], + "spans": [ + { + "bbox": [ + 271, + 162, + 380, + 189 + ], + "type": "text", + "content": "Kalyan Ramakrishnan \nUniversity of Oxford" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 262, + 192, + 388, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 192, + 388, + 202 + ], + "spans": [ + { + "bbox": [ + 262, + 192, + 388, + 202 + ], + "type": "text", + "content": "kalyanr@robots.ox.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 419, + 162, + 509, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 162, + 509, + 188 + ], + "spans": [ + { + "bbox": [ + 419, + 162, + 509, + 188 + ], + "type": "text", + "content": "Rajagopalan A. N. IIT Madras" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 414, + 192, + 512, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 192, + 512, + 202 + ], + "spans": [ + { + "bbox": [ + 414, + 192, + 512, + 202 + ], + "type": "text", + "content": "raju@ee.iitm.ac.in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "type": "text", + "content": "In this paper, we address the weakly-supervised Audio-Visual Video Parsing (AVVP) problem, which aims at labeling events in a video as audible, visible, or both, and temporally localizing and classifying them into known categories. This is challenging since we only have access to video-level (weak) event labels when training but need to predict event labels at the segment (frame) level at test time. Recent methods employ multiple-instance learning (MIL) techniques that tend to focus solely on the most discriminative segments, resulting in frequent misclassifications. Our idea is to first construct several \"prototype\" features for each event class by clustering key segments identified for the event in the training data. We then assign pseudo labels to all training segments based on their feature similarities with these prototypes and re-train the model under weak and strong supervision. We facilitate this by structuring the feature space with contrastive learning using pseudo labels. Experiments show that we outperform existing methods for weakly-supervised AVVP. We also show that learning with weak and iteratively re-estimated pseudo labels can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 548, + 127, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 127, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 127, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "In this paper, we explore the problem of weakly-supervised Audio-Visual Video Parsing (AVVP) [48], where the goal is to classify events in a video and localize them over time and modalities, given only video-level (weak) event labels for training. Such a formulation is attractive since it forgoes the need for expensive and tedious fine-grained labeling. Moreover, other weakly-supervised video-based temporal prediction tasks can be posed as special cases of AVVP. For instance, Audio-Visual Event Localization (AVEL) [47] considers events that are simultaneously audible and visible, whereas Temporal Action Localization (TAL) [10] considers only visible events. AVVP is challenging as an event" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 541, + 381 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 541, + 381 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 541, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 541, + 381 + ], + "type": "image", + "image_path": "c60c9561d79fd0886536a15974b8956a7f3bae7f9a2fd150691401d3c3d834a7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 389, + 545, + 467 + ], + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 467 + ], + "type": "text", + "content": "Figure 1. Example predictions of a model trained using MIL under weak supervision [48] and our method. For clarity, we only show visual events. The MIL model detects the event \"Guitar\" correctly in the beginning, but there are some mistakes (red dotted boxes) later on. Moreover, although the last three frames are similar, they are labeled inconsistently by the model. Our method consistently localizes well. \"GT\" is the ground truth." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "type": "text", + "content": "could occur in just the audio stream, visual stream, or both at the same time. Audio-visual data is ubiquitous, with highly variable events in terms of appearance and duration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Recent methods for weakly-supervised AVVP [30, 48] use multiple-instance learning (MIL) [7] techniques that optimize for video-level labels during training and make frame-level predictions during evaluation. Here, video-level predictions are obtained by pooling (aggregating) the frame-level predictions. A drawback of MIL approaches is that a model may learn to focus solely on the most discriminative (\"prominent\") instances (here, audio/visual frames in a video) and still minimize the training loss successfully. This means such models may not reliably detect the full temporal extent of events. Fig. 1 shows an example where an MIL model [48] for AVVP fails to localize events accurately. Moreover, the model labels frames that are visually similar differently. Another reason for misclassifications in MIL models is that they typically contain a frame-level classifier with a single weight vector per event class. However, this may be insufficient to capture the intra-class variation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18952" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "for the event across time, modalities, and examples while training with weak labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 287 + ], + "type": "text", + "content": "We propose to address these issues with the following ideas. First, given that the MIL model can find discriminative frames that trigger the video-level labels, we can identify such frames and propagate their predicted labels to other frames in the training data based on similarities in the feature space, thus creating frame-level pseudo labels on the training data. Second, we can use their features to construct multiple prototypes per event class, ensuring a more robust representation to propagate pseudo labels with. Once this is done for the training data, we can re-train the model with strong supervision from pseudo labels and weak supervision from video-level labels. Fig. 1 shows an example where our method produces more accurate predictions than the MIL model. Because our method relies on feature similarities, we further enforce structure in the feature space via contrastive learning using pseudo labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 407 + ], + "type": "text", + "content": "We re-estimate pseudo labels on the training data after every epoch and train with them over the next epoch. We show that such a training procedure can be viewed through an expectation-maximization (EM) lens with the unknown frame-level labels as latent variables. This interpretation provides convergence guarantees and further insight. Our method achieves state-of-the-art performance for weakly-supervised AVVP and also works well on the related weakly-supervised Temporal Action Localization (TAL) task. We summarize our contributions below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 287, + 537 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 407, + 287, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 287, + 442 + ], + "type": "text", + "content": "1. We introduce a prototype-based pseudo-labeling method for weakly-supervised AVVP that exploits similarities in feature space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "type": "text", + "content": "2. We show how to perform contrastive learning using pseudo labels to help semantically structure the feature space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "type": "text", + "content": "3. We present an expectation-maximization view of our general training procedure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 287, + 537 + ], + "type": "text", + "content": "4. We show state-of-the-art results for weakly-supervised AVVP, and on several metrics for weakly-supervised TAL." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 549, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 134, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Audio-Visual Video Parsing (AVVP). Tian et al. [48] formulated AVVP as a multimodal multiple-instance learning (MMIL) problem and proposed a hybrid attention network with a learned pooling function (HAN) to capture unimodal and cross-modal contexts. Most subsequent work has used the HAN architecture with other techniques to improve performance. Lamba et al. [30] utilized cross-modal information to learn better representations with adversarial and self-supervised losses. Wu et al. [52] estimated modality-level labels by exchanging the audio/visual streams between pairs of videos, followed by retraining under stronger supervision. Lin et al. [32] used cross-video and cross-modality" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "signals, such as event co-occurrence across modalities, as additional sources of supervision. Mo et al. [36] leveraged multi-modal grouping to learn discriminative subspaces. Gao et al. [11] extract the \"presence\" and \"absence\" evidence for events from uni- and cross-modal information. While Fu et al. [9] focused on learning rate imbalance between modalities and Zhou et al. [60] utilized CLIP [41] to get segment-labels, our work leverages in-model feature similarities for pseudo labeling. PoiBin [40] modeled the number of positive frames for an event as a latent variable to improve localization. Different from previous methods, we estimate segment-level pseudo labels and re-train under full supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 228, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 382 + ], + "type": "text", + "content": "Pseudo-Labeling refers to estimating labels for unlabeled data using the predictions of a trained model. This has been used to improve performance on several weakly-supervised tasks including object detection [1, 46, 59] and image classification [2, 13, 21]. A few works [34, 39, 58] have employed pseudo-labeling to improve performance on Temporal Action Localization (TAL), generating labels from model outputs or attention weights. Luo et al. [34] posed the TAL task as expectation-maximization, using pseudolabels to approximate the E and M steps. Contrary to these, we propose a pseudo-labeling strategy using feature similarities between segments guided by the model predictions and weak labels for a video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 385, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 589 + ], + "type": "text", + "content": "Prototype-based Classification. Prototype learning with a nearest-neighbor (non-parametric) classifier has inspired various machine learning techniques [5, 12, 15, 43, 44]. Recently, there has been increased interest in integrating this into deep learning methods, e.g., zero-shot [25], few-shot [45], unsupervised [54, 55], and supervised [16, 35, 53, 56] learning. Prototype learning has also been widely used for image segmentation [8, 50, 51, 61]. Recently, prototype-based methods have also been used for semi-supervised image [37] and text [57] classification. A few works have used prototype-based methods with multiple-instance learning. E.g., Rymarczyk et al. [42] processed large medical images as a bag of patches, and Huang et al. [22] learned prototypes for the TAL task with graph convolutional networks. These works adopt an embedding-level MIL approach ([24]) with one prototype per class, while we adopt instance-level MIL and construct multiple prototypes per event class." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 602, + 384, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 602, + 384, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 602, + 384, + 615 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 622, + 417, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 417, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 417, + 634 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "In the audio-visual video parsing (AVVP) problem, we aim to identify audible or visible events in an unconstrained video and simultaneously localize them in time. Specifically, given a video " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " consisting " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " non-overlapping temporal segments " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{x_{t} = (x_{t}^{a},x_{t}^{v})\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " of the audio " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " and visual " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " streams, our objective is to classify each segment" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18953" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "into " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": " possible events (e.g., singing, vehicle moving, etc.). At test time, we need to predict segment-level event labels " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "y_{t} = (\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}}, \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": " for each audio and visual segment, where " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}}, \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}} \\in \\{0,1\\}^{C}" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": " are segment-level audio and visual event labels, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": "Weak Supervision. In weakly-supervised AVVP, for each video " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": ", we only have access to the corresponding video-level event label vector " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = (w_{1},\\dots ,w_{C})\\in \\{0,1\\}^{C}" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "w_{c} = 1" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": " if any of the segments in the video contains the " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": "-th event, and " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "w_{c} = 0" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": " otherwise. Such weak, video-level labels only indicate whether an event occurs in the video or not. During evaluation, we still need to predict segment-level labels " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}},\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": ". Note that multiple events can simultaneously occur in a video, i.e., " + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "inline_equation", + "content": "\\sum_{c}w_{c}\\geq 1" + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": ", and each segment could have more than one event." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 149, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 149, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 149, + 270 + ], + "type": "text", + "content": "3.2. MIL Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 286, + 348 + ], + "type": "text", + "content": "We adopt an instance-level MIL approach [24] as done by prior work [30, 32, 36, 48, 52]: First, each instance (here, an audio/visual segment) is assigned a classification score (here, a " + }, + { + "bbox": [ + 46, + 277, + 286, + 348 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 277, + 286, + 348 + ], + "type": "text", + "content": "-dimensional vector). Then, the scores are aggregated by a pooling operation, and the aggregate is used in the loss. In our case, this looks like" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 357, + 286, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 357, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 59, + 357, + 286, + 405 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ \\left\\{\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {a}} \\right\\} _ {t = 1} ^ {T}, \\left\\{\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {v}} \\right\\} _ {t = 1} ^ {T} \\right] = f _ {\\theta} (x) (1) \\\\ \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} = g _ {\\phi} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}}\\right) \\in [ 0, 1 ] ^ {C}, \\forall t \\in [ 1, T ], m \\in \\{a, v \\} (2) \\\\ \\hat {\\mathbf {w}} = \\sigma_ {\\theta , x} \\left(\\left\\{\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {a}} \\right\\} _ {t = 1} ^ {T}, \\left\\{\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {v}} \\right\\} _ {t = 1} ^ {T}\\right) \\in [ 0, 1 ] ^ {C}, (3) \\\\ \\end{array}", + "image_path": "ac8b05f19cff5249852f798003ff5aadd85b2c39d36db88f8908b393117bd92a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": " is a feature extractor (encoder), " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": " is a feature classifier, " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta,x}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": " is a MIL-pooling operator. The predicted video-level probability vector " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": " is computed from segment-level classification scores " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_t^m\\}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": ", which are computed from segment features " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_{\\mathbf{t}}^{\\mathbf{m}}\\}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": ". We adopt the Hybrid Attention Network (HAN) [48] architecture to implement " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}, g_{\\phi}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": " consists of a transformer-based encoder with cross-modal attention and " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": " is a linear classifier with the sigmoid activation. Attention-based averaging is used for " + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 413, + 286, + 544 + ], + "type": "text", + "content": ". We provide implementation details in the Supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 544, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 286, + 568 + ], + "type": "text", + "content": "This model is trained end-to-end with supervision at the video level by minimizing a weighted cross-entropy loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 578, + 286, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 578, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 129, + 578, + 286, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M I L}} = \\operatorname {C E} (\\hat {\\mathbf {w}}, \\mathbf {w}) \\tag {4}", + "image_path": "d3be40ccc0bce6f17f8899daa7c5124cbed7d11821b4485397f958adc1de1493.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 599, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 286, + 635 + ], + "type": "text", + "content": "with weights inversely proportional to the event distribution in the dataset. At test time, we get segment-level predictions from Eq. (2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 645, + 234, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 234, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 234, + 659 + ], + "type": "text", + "content": "4. Prototype-based Pseudo-Labeling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "While instance-level MIL enables finding key segments that trigger the video-level labels [24, 33], they often miss out on the full temporal extent of events in a video. However, if we identify such discriminative segments for each label" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "in a training video, we could use them to build a set of prototype features that represent each class. A high degree of similarity of a segment with the prototype features for an event class means the segment probably contains the event. We use this idea to generate pseudo labels for all segments in a training video and then re-train the model under strong supervision. Fig. 2 shows an overview of our approach. We provide pseudocode in Algorithm 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 175, + 481, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 481, + 188 + ], + "type": "text", + "content": "4.1. Constructing Prototype Features" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": "The segment-level linear classifier " + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": " (Eq. (2)) can itself be considered a prototype-based classifier, where the prototype for each class is the corresponding weight vector in " + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": ". However, this requires that an event have similar representations across modalities and examples in the training data. We instead construct multiple prototypes for each class in the following two steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "text", + "content": "Segment Selection. For each training video " + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "text", + "content": ", we find positive segments " + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_x(c)" + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "text", + "content": " for a target event " + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "text", + "content": " (if " + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "inline_equation", + "content": "w_c = 1" + }, + { + "bbox": [ + 304, + 277, + 545, + 313 + ], + "type": "text", + "content": ") as those the model is highly confident contain the event. I.e.," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 322, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 322, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 322, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {x} (c) = \\left\\{ \\begin{array}{l l} \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) \\geq \\gamma p ^ {*} (c) \\right\\} & \\text {i f} w _ {c} = 1 \\\\ \\phi & \\text {i f} w _ {c} = 0, \\end{array} \\right. \\tag {5}", + "image_path": "69a06ab8f039dcca6c3fa4534278ce585f994d087d8980cdd5df5ab761162f32.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\gamma \\in (0,1]" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " is a hyperparameter and " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "p^*(c) = \\max_{t,m} \\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}(c)" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " is the maximum predicted probability for the event over time and modalities for a given video. Clearly, " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_X(c)" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " will contain at least one audio/visual segment from video " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "w_c = 1" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": ". We assume these segments contain event " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " with high probability (so are likely true positives). We also find negative segments " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_x(c)" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " for a target event " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " as those the model is highly confident do not contain the event when " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "w_c = 1" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " (likely true negatives), or is highly confident to contain the event when " + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "inline_equation", + "content": "w_c = 0" + }, + { + "bbox": [ + 304, + 361, + 545, + 493 + ], + "type": "text", + "content": " (definite false positives):" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 502, + 545, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 545, + 542 + ], + "type": "interline_equation", + "content": "\\mathcal {N} _ {x} (c) = \\left\\{ \\begin{array}{l l} \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) \\leq \\beta p ^ {*} (c) \\right\\} & \\text {i f} w _ {c} = 1 \\\\ \\left\\{x _ {t} ^ {m} \\in x \\mid \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = p ^ {*} (c) > = 0. 5 \\right\\} & \\text {i f} w _ {c} = 0, \\end{array} \\right. \\tag {6}", + "image_path": "3ded314c300b251d16de9bea56d62680ea1e4946709ad246de8bd5cc6907cdde.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,1)" + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "text", + "content": " is a hyperparameter set to a small value to retrieve segments without the event " + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "text", + "content": " (if any). Note that " + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_x(c)" + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "text", + "content": " may contain no segments at all even when " + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "inline_equation", + "content": "w_{c} = 1" + }, + { + "bbox": [ + 304, + 543, + 545, + 590 + ], + "type": "text", + "content": ", i.e., be a null set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 590, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 545, + 615 + ], + "type": "text", + "content": "Finally, we aggregate the positive and negative sets over all videos in training data " + }, + { + "bbox": [ + 306, + 590, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 306, + 590, + 545, + 615 + ], + "type": "text", + "content": " to get prototypical segments as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 340, + 624, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 624, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 340, + 624, + 545, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {P} (c) = \\bigcup_ {x \\in \\mathcal {D}} \\mathcal {P} _ {x} (c), \\quad \\mathcal {N} (c) = \\bigcup_ {x \\in \\mathcal {D}} \\mathcal {N} _ {x} (c). \\tag {7}", + "image_path": "e298ec15d75b0cc992e38f4b7ab5d9805874439ea5f2ee7cffcc6d73bdd0d47f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Prototype Generation. We now generate multiple prototypes (feature vectors) for each event class by hypothesizing that having more than one prototype would help better capture the intra-class variation across modalities and training examples, and create a more robust representation." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18954" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 68, + 484, + 220 + ], + "blocks": [ + { + "bbox": [ + 77, + 68, + 484, + 220 + ], + "lines": [ + { + "bbox": [ + 77, + 68, + 484, + 220 + ], + "spans": [ + { + "bbox": [ + 77, + 68, + 484, + 220 + ], + "type": "image", + "image_path": "af2047dd15287e6043608b15af802cf7050fae0e8f2582d8bc37a28b035c7955.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "type": "text", + "content": "Figure 2. An overview of our method. The top branch is trained using MIL with weak labels and using full supervision with pseudo labels, which are generated by the bottom branch. This is done by first selecting key segments for each event from the training data and clustering them to construct a set of prototype features for the event, used to guide pseudo-label generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": "First, we extract features for all segments in " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(c)" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(c)" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": " using Eq. (1), and call the resulting feature sets " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{P}}(c)" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{N}}(c)" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": ". Here, instead of " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": ", we use a momentum encoder [19] " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "f_{\\tilde{\\theta}}" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": " having the same architecture but whose weights are set (after each update to " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": ") as an exponential moving average (EMA) of " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": " over training steps: " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta} \\gets (1 - \\eta)\\tilde{\\theta} + \\eta\\theta" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\eta \\in (0,1)" + }, + { + "bbox": [ + 46, + 282, + 289, + 390 + ], + "type": "text", + "content": ". The momentum encoder is more stable and less sensitive to noise from weak supervision, leading to slightly improved performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "text", + "content": "Next, we perform " + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "text", + "content": "-means clustering separately on " + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{P}}(c)" + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{N}}(c)" + }, + { + "bbox": [ + 47, + 394, + 288, + 430 + ], + "type": "text", + "content": " to get two sets of cluster centroids. We call the centroids prototypes:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 447, + 287, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 447, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 84, + 447, + 287, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {\\mathcal {P}} (c) = \\left\\{\\mathbf {f} _ {\\mathbf {i}} ^ {+} \\right\\} _ {i = 1} ^ {k _ {p}}, \\quad \\mathcal {C} _ {\\mathcal {N}} (c) = \\left\\{\\mathbf {f} _ {\\mathbf {j}} ^ {-} \\right\\} _ {j = 1} ^ {k _ {n}}. \\tag {8}", + "image_path": "f4abb2a37018bd4652adfd182ce7ab22b3d2cc59d9f5c1ed8a269c647f108683.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 515 + ], + "type": "text", + "content": "Clustering for finding prototypes imposes a natural bottleneck [26], helping discard irrelevant information in the segments. We repeat this for each event class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 535, + 192, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 192, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 192, + 547 + ], + "type": "text", + "content": "4.2. Estimating Pseudo Labels" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "text", + "content": "We assign pseudo labels to each segment in a training video by considering its feature similarities with the prototypes for each class. We assume that if an audio/visual segment contains event " + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "text", + "content": ", it would be on average closer to " + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(c)" + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "text", + "content": " in feature space than to " + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(c)" + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "text", + "content": ". For this, we explore two approaches: (i) hard labeling, which assigns the label of the closest-on-average set among the prototype sets for class " + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 558, + 287, + 641 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = w _ {c} \\cdot \\mathbb {1} \\left[ \\frac {\\sum_ {\\mathbf {f} ^ {+} \\in \\mathcal {C} _ {\\mathcal {P}} (c)} e ^ {(\\mathbf {f} ^ {+}) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}}}}{| \\mathcal {C} _ {\\mathcal {P}} |} > \\frac {\\sum_ {\\mathbf {f} ^ {-} \\in \\mathcal {C} _ {\\mathcal {N}} (c)}}{| \\mathcal {C} _ {\\mathcal {N}} |} \\right], \\tag {9}", + "image_path": "8689933f93c10ed3155d2664cecb527f761836f799cebbfd2ffc9b041b0c2e7c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 282, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 545, + 304 + ], + "type": "text", + "content": "and (ii) soft labeling, which instead assigns a probability score:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 307, + 545, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 307, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 310, + 307, + 545, + 367 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} (c) = w _ {c} \\cdot \\frac {\\frac {1}{| \\mathcal {C} _ {\\mathcal {P}} |} \\sum_ {\\mathbf {f} ^ {+} \\in \\mathcal {C} _ {\\mathcal {P}} (c)} e ^ {\\left(\\mathbf {f} ^ {+}\\right) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}} / \\tau}}{\\sum_ {\\mathcal {S} \\in \\left\\{\\mathcal {C} _ {\\mathcal {P}}, \\mathcal {C} _ {\\mathcal {N}} \\right\\}} \\frac {1}{| \\mathcal {S} |} \\sum_ {\\mathbf {f} ^ {\\prime} \\in \\mathcal {S} (c)} e ^ {\\left(\\mathbf {f} ^ {\\prime}\\right) ^ {T} \\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {m}} / \\tau}} \\in (0, 1), \\tag {10}", + "image_path": "df20a9ff42afe38ee1c32d73868cca6a5e57c04fe589d1b6f02e9b95925e57a5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 368, + 493, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 493, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 493, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 368, + 493, + 379 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 306, + 368, + 493, + 379 + ], + "type": "text", + "content": " is a temperature hyperparameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "type": "text", + "content": "When training our model, we generate pseudo labels " + }, + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\}" + }, + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "type": "text", + "content": " after every epoch for each training video and use them to fully supervise the segment-level probabilities " + }, + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\}" + }, + { + "bbox": [ + 306, + 380, + 545, + 428 + ], + "type": "text", + "content": " in Eq. (2) using the binary cross-entropy loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 432, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 432, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 381, + 432, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P L}} = \\operatorname {C E} \\left(\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {m}}, \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}}\\right), \\tag {11}", + "image_path": "d6635db5529e54cc8cebca8579a2c594784cb73e7eb01e6711a49a7792708c38.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 449, + 497, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 497, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 497, + 462 + ], + "type": "text", + "content": "averaged over segments and training examples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 467, + 523, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 523, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 523, + 479 + ], + "type": "text", + "content": "4.3. Contrastive Learning with Pseudo Labels" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "content": "Given that our method exploits feature similarities between segments based on the events they may contain, we use contrastive learning based on pseudo labels to help semantically structure the feature space. Since multiple events can occur simultaneously, for each audio/visual segment, we consider a similar segment as one estimated to have some common event(s) and a dissimilar segment as one estimated to have no common events. Formally, given a batch of training examples, let " + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "inline_equation", + "content": "B = \\{\\mathbf{f}_{\\mathbf{t}}^{\\mathbf{m}} \\to \\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\}" + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "content": " be a dictionary from features to the assigned pseudo labels for all segments and over all videos in the batch. For every feature " + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in B" + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "content": ", the similar feature-set is then " + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "inline_equation", + "content": "S(\\mathbf{f}) = \\{\\mathbf{g} \\in B \\mid B(\\mathbf{g}) \\cap B(\\mathbf{f}) \\neq \\phi\\}" + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "content": " and the dissimilar set is " + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "inline_equation", + "content": "D(\\mathbf{f}) = \\{\\mathbf{g} \\in B \\mid B(\\mathbf{g}) \\cap B(\\mathbf{f}) = \\phi\\}" + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "content": ". To pull segments with a common predicted event closer in the feature space, we use the contrastive loss [38]" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 668, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 668, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 323, + 668, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L}} = - \\log \\frac {\\sum_ {\\mathbf {g} \\in \\mathcal {S} (\\mathbf {f})} e ^ {\\mathbf {g} ^ {T} \\mathbf {f} / \\tau_ {c}}}{\\sum_ {\\mathbf {g} \\in \\mathcal {S} (\\mathbf {f})} e ^ {\\mathbf {g} ^ {T} \\mathbf {f} / \\tau_ {c}} + \\sum_ {\\mathbf {h} \\in \\mathcal {D} (\\mathbf {f})} e ^ {\\mathbf {h} ^ {T} \\mathbf {f} / \\tau_ {c}}}, \\tag {12}", + "image_path": "14f504b160bc163887590005bd5798da8b35369c2ac89a9660de95a93c1e0d5f.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18955" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 502, + 222 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 502, + 222 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 502, + 222 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 502, + 222 + ], + "type": "table", + "html": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
AVE* [47]47.240.437.134.735.431.639.935.541.636.5
AVSDN* [31]47.834.152.046.337.126.545.735.650.837.7
HAN [48]60.151.352.948.948.943.054.047.755.448.0
MA [52]60.353.660.056.455.149.058.953.057.950.6
CVCM-MA [32]60.853.863.558.957.049.560.554.059.552.1
MGN-MA [36]60.250.961.959.755.549.659.253.458.749.9
JoMoLD [4]61.353.963.859.957.249.660.854.559.952.5
PoiBin [40]63.154.163.560.357.751.561.455.260.652.3
CMPAE [11]64.256.666.463.759.251.863.357.462.855.7
Ours65.957.366.764.361.954.364.859.963.757.9
", + "image_path": "31193cd3a085fd82a89a5e246a2b1e19338e255ca60b9581e709aeb257553b04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "type": "text", + "content": "Table 1. Results (% F1-scores) on all metrics defined in Tian et al. [48]. \\* indicates reproduced performance. The best and second best results are in bold and underline, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "code", + "bbox": [ + 48, + 280, + 287, + 475 + ], + "blocks": [ + { + "bbox": [ + 50, + 266, + 224, + 277 + ], + "lines": [ + { + "bbox": [ + 50, + 266, + 224, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 266, + 224, + 277 + ], + "type": "text", + "content": "Algorithm 1: Pseudocode for our method" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": "1 Init: model params " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\psi = \\{\\theta ,\\phi \\}" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " , EMA params " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\psi_{E} = \\{\\theta_{E},\\phi_{E}\\}" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n2 while " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " has not converged do \n3 #estimate segment - level pseudo labels \n4 for " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "(x,\\mathbf{w})" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " in train set do \n5 " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\left|\\{\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\gets p_{\\theta}(\\cdot |\\mathbf{w},x)\\right.;" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n6 end \n7 #train \n8 for " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "(x,\\mathbf{w})" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " in train set do \n9 " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\} \\gets p_{\\theta ,\\phi}(\\cdot |x);" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n10 " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}} = \\sigma_{\\theta ,x}(\\{\\mathbf{p}_{\\mathbf{t}}^{\\mathbf{m}}\\});" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n11 " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{MIL}} + \\lambda_{\\mathrm{PL}}\\mathcal{L}_{\\mathrm{PL}} + \\lambda_{\\mathrm{CL}}\\mathcal{L}_{\\mathrm{CL}};" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n12 " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\psi \\gets \\psi -\\alpha \\nabla_{\\psi}\\mathcal{L};" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n13 " + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\psi_{E}\\gets (1 - \\eta)\\psi_{E} + \\eta \\psi ;" + }, + { + "bbox": [ + 48, + 280, + 287, + 475 + ], + "type": "text", + "content": " \n14 end \n15 end" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "type": "text", + "content": "averaged over all " + }, + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in B" + }, + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\tau_{c} > 0" + }, + { + "bbox": [ + 46, + 508, + 287, + 533 + ], + "type": "text", + "content": " is a temperature hyperparameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 550, + 178, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 178, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 178, + 562 + ], + "type": "text", + "content": "4.4. Training and Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 572, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 595 + ], + "type": "text", + "content": "We train our model using the following loss in an end-to-end fashion:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 613, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 613, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 98, + 613, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M I L}} + \\lambda_ {\\mathrm {P L}} \\mathcal {L} _ {\\mathrm {P L}} + \\lambda_ {\\mathrm {C L}} \\mathcal {L} _ {\\mathrm {C L}}, \\tag {13}", + "image_path": "b169f9a7046199de91c9213a4eb18d89422345c7c1dbe9fdf8ccc551a03670dd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{PL}}, \\lambda_{\\mathrm{CL}} > 0" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " are hyperparameters. During training, we generate pseudo labels once every epoch for all training examples using the best model so far. During inference, we predict the segment-level probabilities using the non-parametric prototypical pseudo-labeling (Eq. (10)) with video-level predictions from Eq. (3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 262, + 386, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 386, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 386, + 276 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 282, + 356, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 356, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 356, + 295 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 419 + ], + "type": "text", + "content": "LLP Dataset. We conduct experiments on the Look, Listen and Parse (LLP) dataset [48] which consists of 11849 YouTube videos, each of 10s duration, labeled into 25 event categories. These videos are unconstrained with a wide variety of scene content, including daily activities, music performances, and vehicle sounds. We use 10000 videos with weak (video-level) labels for training. The remaining 1849 fully-annotated videos (with segment-level labels) are used for validation and testing. The train/val/test split has been provided in the LLP dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 421, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 503 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous work, we use F1-scores on audio, visual, and audio-visual events as evaluation metrics. These are computed both at the segment and event level. We also include the aggregate metrics \"Type@AV\" and \"Event@AV\", again computed at the segment and event level. See Tian et al. [48] for a full explanation of metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": "Implementation Details. We use ResNet-152 [18] pretrained on ImageNet [6] and R(2+1)D-18 [49] pre-trained on Kinetics-400 [27] as visual feature extractors to generate a 512-d visual feature per segment. We use VGGish [20] pre-trained on AudioSet [14] to generate a 128-d audio feature per segment. We use the Adam optimizer [28] with a batch size of 16 and a learning rate " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\alpha = 3e - 4" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " for 50 epochs. We take " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\eta = 0.999" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " for the momentum encoder, " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{PL}} = \\lambda_{\\mathrm{CL}} = 0.1" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (13), " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (5), " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\beta = 0.2" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (6), " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (10), and " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\tau_{c} = 0.2" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (12). We set " + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "inline_equation", + "content": "k_{p} = k_{n} = 10" + }, + { + "bbox": [ + 304, + 506, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (8). Hyperparameter values are based on validation performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 658, + 471, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 471, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 471, + 670 + ], + "type": "text", + "content": "5.2. Comparisons with Prior Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 700 + ], + "type": "text", + "content": "For a fair comparison, all methods considered use the same pre-trained feature extractors and train/val/test split." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Quantitative. We compare methods in Table 1 on all the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18956" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 292, + 310 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 292, + 310 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 292, + 310 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 292, + 310 + ], + "type": "image", + "image_path": "8e504539dcc566fdaa74ed92febaad0af3d6b78d86b3045246478c4fb7b6d264.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 317, + 538, + 330 + ], + "lines": [ + { + "bbox": [ + 54, + 317, + 538, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 538, + 330 + ], + "type": "text", + "content": "Figure 3. Audio-Visual Video Parsing results of our methods with HAN [48], MA [52], JoMoLD [4] and CMPAE [11] on two videos." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 116, + 539, + 309 + ], + "blocks": [ + { + "bbox": [ + 298, + 116, + 539, + 309 + ], + "lines": [ + { + "bbox": [ + 298, + 116, + 539, + 309 + ], + "spans": [ + { + "bbox": [ + 298, + 116, + 539, + 309 + ], + "type": "image", + "image_path": "7ab05ab4e5bc3d13898ad66819dcb2888e7b518db9b16f2918442eb6334277a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 337, + 286, + 401 + ], + "blocks": [ + { + "bbox": [ + 50, + 337, + 286, + 401 + ], + "lines": [ + { + "bbox": [ + 50, + 337, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 50, + 337, + 286, + 401 + ], + "type": "table", + "html": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
Base+NPLH58.952.158.549.855.748.257.850.155.746.6
Base+NPILS59.352.559.051.156.048.758.250.856.247.6
Base+PPLH62.557.462.959.659.654.061.956.360.355.7
Base+PPLS65.957.366.764.361.954.364.859.963.757.9
", + "image_path": "c9a7911985318fc4b46e860f4ae7da95a612b8ce74820e1f272259cbd21b9fcc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 278, + 421 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 278, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 278, + 421 + ], + "type": "text", + "content": "Table 2. Performance for different pseudo labeling strategies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 436, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 555 + ], + "type": "text", + "content": "metrics listed earlier. We see that our method outperforms others on audio, audio-visual, Type@AV, and Event@AV scores at both the segment and event levels. In particular, we achieve an average improvement of 2.6 points for audio-visual events. Averaged over all metrics, our method achieves a 1.86 point improvement over the previous best method CMPAE [11]. Since we do not require additional learnable parameters to generate pseudo labels, this improvement is with a comparable number of parameters to previous methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 701 + ], + "type": "text", + "content": "Qualitative. We qualitatively compare our method with some previous methods on two examples in Fig. 3. Here, Fig. 3 (left) contains events \"Singing\" and \"Banjo\" in both modalities. Fig. 3 (Right) shows another scene with the event \"Speech\" in only the audio modality, and \"Violin\" occurring in both modalities. As evident, the common failure mode of previous methods is that they detect only a portion of the duration of events, while ours more commonly detects the full extent of events. As discussed in Sec. 1, this could be because of multiple-instance learning encouraging the model to pick out only the most discriminative segments containing an event, as opposed to all segments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "text", + "content": "Computational Cost. When training, our approach takes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 339, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 387 + ], + "type": "text", + "content": "67.4s per epoch (17.1s for forward+backward propagation and 50.3s for prototype-based pseudo labeling). Inference time is 5.9ms on average per video, which is comparable to the baselines (2.9ms)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 395, + 368, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 368, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 368, + 407 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 414, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 665 + ], + "type": "text", + "content": "Pseudo-Labeling Strategies. We report an ablation study in Table 2 to validate our specific pseudo-labeling method. Here, Base is the baseline MIL model explained in Sec. 3.2. Base+NPL refers to the base model trained with naive pseudo labeling, i.e., by directly using the segment-level predictions of the model as pseudo labels for retraining. In this case, H and S refer to using hard (binarized) and soft (continuous) pseudo labels, respectively. We observe no consistent improvement with naive pseudo labeling, likely due to the MIL model producing incorrect predictions for the non-discriminative segments for an event, leading to less reliable pseudo labels. Base+PPL refers to the base model trained with prototype-based pseudo labels as described in Sec. 4.2. In this case, H and S refer to using hard and soft pseudo labels, as defined in Eqs. (9) and (10), respectively. Our method significantly improves performance by generating more reliable pseudo labels throughout a training video. We also see that soft labeling is better than hard labeling. Since segment-level predictions may be noisy, we expect mistakes to affect performance more with hard rather than soft labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Reliability of Pseudo Labels. We now assess if the pseudo labels generated by our method are reliable for re-training. For this, we take the pre-trained baselines HAN [48] and MA [52], generate prototypes for each event from the train-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18957" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 128 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 128 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 128 + ], + "type": "table", + "html": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
HAN60.151.352.948.948.943.054.047.755.448.0
+PPLTest62.555.455.351.152.346.956.050.958.350.6
MA60.353.660.056.455.149.058.953.057.950.6
+PPLTest61.755.461.857.957.551.660.655.059.452.6
", + "image_path": "6cf0e70fde01999da2897c2e1d22efaec43694c25ecf60ff317077738b9f814c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 166, + 160, + 258 + ], + "blocks": [ + { + "bbox": [ + 51, + 166, + 160, + 258 + ], + "lines": [ + { + "bbox": [ + 51, + 166, + 160, + 258 + ], + "spans": [ + { + "bbox": [ + 51, + 166, + 160, + 258 + ], + "type": "image", + "image_path": "eb7db5efc731b7e1a61941d5d00977d2b01df0930cf6536e7539110cc1ca8d8c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "lines": [ + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "spans": [ + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "type": "text", + "content": "Figure 4. Performance for different " + }, + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 70, + 268, + 262, + 280 + ], + "type": "text", + "content": " choices." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 165, + 275, + 257 + ], + "blocks": [ + { + "bbox": [ + 169, + 165, + 275, + 257 + ], + "lines": [ + { + "bbox": [ + 169, + 165, + 275, + 257 + ], + "spans": [ + { + "bbox": [ + 169, + 165, + 275, + 257 + ], + "type": "image", + "image_path": "280184cfa4828bba9fb6a9b25a9cf6dde190c3fe77f86a6adab73f7476b677a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 285, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 48, + 137, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 137, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 286, + 148 + ], + "type": "text", + "content": "Table 3. Evaluating prototype-based pseudo labels on the test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 285, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 50, + 285, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 50, + 285, + 286, + 335 + ], + "type": "table", + "html": "
MethodAudioVisualAudio-VisualType@AVEvent@AV
Seg.EventSeg.EventSeg.EventSeg.EventSeg.Event
Base59.551.158.753.454.547.457.650.656.548.5
No-CL64.156.459.657.357.149.661.256.460.856.8
With-CL65.957.366.764.361.954.364.859.963.757.9
", + "image_path": "9a73e3ce0e950aa6dc663a597d3ab2e34270a1df6c6a79eb14188d98497aca51.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "content": "ing data, and simply predict the prototype-based pseudo labels (Eq. (9)) on the test set for evaluation (i.e., discard the classifiers at test time). We call the resulting classifiers " + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{HAN} + \\mathrm{PPL}_{\\mathrm{Test}}" + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{MA} + \\mathrm{PPL}_{\\mathrm{Test}}" + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "content": " and show results in Table 3. With no other change to the models other than the classifiers, we see significant improvements in all metrics, showing that our pseudo labels are indeed reliable estimates of the true labels. Note that for the results in Table 1, we recompute pseudo labels on the training data every epoch, with a continually improving base model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": ". The hyperparameters " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": " in Eqs. (5) and (6) roughly control the proportion of segments that will be selected as positives and negatives for assigning pseudo labels. We experiment with different values and plot the average of all segment/event-level metrics in Fig. 4. Performance is susceptible to " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": ", which controls positives, and is tolerant to a slight variation in " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": ". The best performance is for " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\beta = 0.2" + }, + { + "bbox": [ + 46, + 485, + 286, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "type": "text", + "content": "Effect of Contrastive Learning. We analyze the impact of contrastive learning using pseudo labels (Eq. (12)) by simply setting " + }, + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CL}} = 0" + }, + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "type": "text", + "content": " in Eq. (13). Results are reported in Tab. 4. While only training with prototype-based pseudo labels " + }, + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "type": "inline_equation", + "content": "(\\mathsf{No - CL})" + }, + { + "bbox": [ + 46, + 582, + 286, + 665 + ], + "type": "text", + "content": " significantly improves the baseline, adding a contrastive loss based on our pseudo labels (With-CL) further yields a large improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "Number of Clusters. We take " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "k_{p} = k_{n}" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " and try values in " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots,100\\}" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": ". We plot the average of all segment/event-level metrics in Fig. 5. The best performance was for " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " clusters per class. Having fewer clusters helps cap" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 74, + 526, + 179 + ], + "blocks": [ + { + "bbox": [ + 316, + 74, + 526, + 179 + ], + "lines": [ + { + "bbox": [ + 316, + 74, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 316, + 74, + 526, + 179 + ], + "type": "image", + "image_path": "57b6c0c53107decadc8ad13f2930dc7c42130d5eef81b2fd1eeb06c059496e3f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 188, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 306, + 188, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 545, + 212 + ], + "type": "text", + "content": "Figure 5. Performance for different choices of the number of clusters/prototypes per event." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 223, + 544, + 278 + ], + "blocks": [ + { + "bbox": [ + 52, + 343, + 281, + 354 + ], + "lines": [ + { + "bbox": [ + 52, + 343, + 281, + 354 + ], + "spans": [ + { + "bbox": [ + 52, + 343, + 281, + 354 + ], + "type": "text", + "content": "Table 4. Effect of contrastive learning (CL) with pseudo labels." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 223, + 544, + 278 + ], + "lines": [ + { + "bbox": [ + 308, + 223, + 544, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 544, + 278 + ], + "type": "table", + "html": "
MethodtIoUAvg.
0.10.20.30.40.50.60.7[0.1:0.5][0.3:0.7][0.1:0.7]
EM-MIL [34]59.152.745.536.830.522.716.444.930.437.7
ASM-Loc [17]71.265.557.146.836.625.213.455.435.845.1
DELU [3]71.566.256.547.740.527.215.356.537.446.4
Ours72.766.957.946.937.426.820.156.437.846.9
", + "image_path": "2a441e86dadc310039b2e7807e15d63822f7635fa0eebf4e2e17b4fcbd111f5e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 286, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 306, + 286, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 545, + 308 + ], + "type": "text", + "content": "Table 5. Results (\\% mAP) for the TAL task at different tIoU thresholds." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 331, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 366 + ], + "type": "text", + "content": "ture more general features for an event instead of instance-specific details, but having too few hurts performance by discarding intra-class variation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 375, + 457, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 457, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 457, + 386 + ], + "type": "text", + "content": "5.4. Results for a Different Task" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 394, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 525 + ], + "type": "text", + "content": "We conduct preliminary experiments for the weakly-supervised Temporal Action Localization (TAL) task [10] on the THUMOS14 dataset [23] to further validate our method. TAL aims to temporally localize and classify visual events in videos into known categories. We report mean average precisions (mAP) at different temporal intersection over union (tIoU) thresholds and the results are in Table 5. We achieve better or comparable performance w.r.t. the current state-of-the-art method DELU [3], and outperform previous methods. Check the supplementary for a full comparison." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 537, + 545, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 545, + 551 + ], + "type": "text", + "content": "6. Interpretation as Expectation-Maximization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "content": "We now show that iteratively re-estimating soft pseudo labels and training the model under weak " + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "content": " strong supervision (Eq. (13)) can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure. For this, we view the segment-level labels " + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "y = \\{(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{a}},\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{v}})\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "content": " for each video as latent variables since they are not observed when training under weak supervision. The observed variables are the video segments " + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "x = \\{(x_{t}^{a},x_{t}^{v})\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "content": " and weak labels " + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "content": ". For clarity, assume a single event class, i.e., " + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{w},\\mathbf{y}_{\\mathbf{m}}^{\\mathbf{t}}\\in \\{0,1\\}" + }, + { + "bbox": [ + 305, + 557, + 545, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Our goal is to maximize the data log-likelihood " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta) = \\ln p_{\\theta}(\\mathbf{w}|x)" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " over model parameters " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " (assume just one data point " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " for clarity). Instead of maximizing this directly, we" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18958" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "can attempt to maximize the variational lower bound [29] " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VLB}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " for this likelihood (random variables are capitalized):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 118, + 287, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 118, + 287, + 154 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 287, + 154 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V L B}} (\\theta , \\phi) = \\underset {Y \\sim q _ {\\phi} (\\cdot | \\mathbf {w}, x)} {\\mathbb {E}} [ \\ln p _ {\\theta} (\\mathbf {w} | Y, x) ] \\\\ - \\mathrm {D} _ {\\mathrm {K L}} \\left[ q _ {\\phi} (\\cdot | \\mathbf {w}, x) \\| p _ {\\theta} (\\cdot | x) \\right], \\tag {14} \\\\ \\end{array}", + "image_path": "2073bceccb6b82e7f982dd71c59c4bb50e4407de0cc06cd5ed92a3202b9f801b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "q_{\\phi}(y|\\mathbf{w},x)" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " is any conditional distribution over labels " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": ", with parameters " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": ", meant to approximate the true posterior " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|x)" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": ". The EM algorithm alternates between two steps, first approximating the posterior by maximizing " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VLB}}" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " (E step), and then improving the data likelihood by maximizing it over " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " (M step). Let " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "j \\geq 1" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " index the iterations (note, " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 163, + 287, + 246 + ], + "type": "text", + "content": " here are not the same as in Sec. 3.2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 193, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 193, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 193, + 258 + ], + "type": "text", + "content": "E Step. The " + }, + { + "bbox": [ + 47, + 247, + 193, + 258 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 247, + 193, + 258 + ], + "type": "text", + "content": "-th update is given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 269, + 287, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 269, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 93, + 269, + 287, + 282 + ], + "type": "interline_equation", + "content": "q _ {\\phi_ {j}} (y | \\mathbf {w}, x) = p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x), \\quad \\forall y. \\tag {15}", + "image_path": "ee5186c93c72899d0138319a16a2ff28de2b0902645003cbfde910db825a43b1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "type": "text", + "content": "Assuming conditional independence of the segment-level labels given the video and its weak labels, we can express the right-hand side as " + }, + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|\\mathbf{w},x) = \\prod_{t,m}p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)" + }, + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "type": "text", + "content": ", computed using" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 348, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 348, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 81, + 348, + 287, + 380 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} \\mid \\mathbf {w}, x\\right) := \\left\\{ \\begin{array}{l l} p _ {\\theta} \\left(\\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} \\mid x\\right) & \\text {i f} \\mathbf {w} = 1 \\\\ 1 - \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {m}} & \\text {i f} \\mathbf {w} = 0, \\end{array} \\right. \\tag {16}", + "image_path": "59be135de6d005c14e7f892274f635ae980d99ee06353a7c8f21b063c66e4be6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|x)" + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": " represents the prediction of the model for modality " + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": " and segment " + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": " in video " + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)" + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": " becomes a soft pseudo label for the segment " + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "inline_equation", + "content": "x_{m}^{t}" + }, + { + "bbox": [ + 47, + 388, + 287, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "text", + "content": "M Step. The " + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "text", + "content": "-th update finds " + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "text", + "content": " by maximizing " + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VLB}}(\\theta, \\phi_j)" + }, + { + "bbox": [ + 47, + 425, + 287, + 450 + ], + "type": "text", + "content": ", which, due to Eq. (15), is equivalent to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 458, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 458, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 64, + 458, + 287, + 512 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (y | x) \\\\ - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (\\mathbf {w} | y, x), \\tag {17} \\\\ \\end{array}", + "image_path": "d1c8912e324609d967f0d7e2e21f27ed3386fedd87c2b9cf2978c9126df9eabf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|\\mathbf{w},x)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|x)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " are computed as in Eq. (16). The first term in Eq. (17) can be expanded as the total cross-entropy between the segment-level predictions " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|x)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " and the pseudo-label targets " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta_{j - 1}}(\\mathbf{y}_{\\mathbf{t}}^{\\mathbf{m}}|\\mathbf{w},x)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " and is therefore equal to " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PL}}" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " (see Eq. (13)). To understand the second term in Eq. (17), note that in our setting, " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{w}|y,x) = p_{\\theta}(\\mathbf{w}|x)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " since we make video-level predictions using MIL-pooling without access to the segment-level labels " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "(y)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": ". The second term then reduces to " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "-\\ln p_{\\theta}(\\mathbf{w}|x)" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": ", which is simply the negative log-likelihood of the video-level label " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{w})" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": ". This means the second term is equal to " + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MIL}}" + }, + { + "bbox": [ + 47, + 521, + 287, + 666 + ], + "type": "text", + "content": " (see Eq. (13)). I.e., the M step is equivalently" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 685, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 685, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 52, + 685, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} - \\sum_ {y} p _ {\\theta_ {j - 1}} (y | \\mathbf {w}, x) \\ln p _ {\\theta} (y | x) - \\ln p _ {\\theta} (\\mathbf {w} | x). \\tag {18}", + "image_path": "614b7e3dd3bca83947c5a28098270795310c0bcd6700c7cc0b903351f1287ec6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "type": "text", + "content": "If we simply maximized the data log-likelihood " + }, + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta)" + }, + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "type": "text", + "content": " directly (no pseudo labels), this could converge to a solution that picks out just the most discriminative segments in each video, which may suffice to reduce the video-level loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MIL}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 276 + ], + "type": "text", + "content": ". As we showed in this paper, using pseudo labels can help mitigate this, but the quality of the estimated pseudo labels has a crucial role. E.g., hard labeling is inferior to soft labeling since it does not attempt to compute the E step (Eq. (15)) exactly. While soft labeling does attempt an exact E step, better inductive biases (such as prototype-based labeling with contrastive learning) here can help reach a favorable solution, as reported in Sec. 5.3 and Table 2. We also compare with an off-the-shelf pseudo-labeling method [34] in the supplementary material and find that we significantly outperform it, showing that it is not pseudo-labeling per se that works for AVVP - how we estimate the pseudo labels is paramount." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 300, + 378, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 378, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 378, + 312 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "type": "text", + "content": "We proposed a prototype-based pseudo-labeling method for weakly-supervised AVVP that outperforms existing methods. Some interesting findings are that (i) naive pseudolabeling does not produce reliable training targets, (ii) contrastive learning using pseudo labels is an effective strategy, and (iii) training with weak and re-estimated soft pseudo labels is an EM algorithm." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 413, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 628 + ], + "type": "text", + "content": "We end with some limitations and suggestions for future work. First, performance is sensitive to hyperparameter " + }, + { + "bbox": [ + 305, + 413, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 413, + 545, + 628 + ], + "type": "text", + "content": " as shown in Fig. 4. This could be because identifying negatives for a class from the predictions of an MIL model is less likely to be accurate than identifying positives. Moreover, negative segments for an event would have a much larger variation in content. Second, events with larger intra-class variation may need more prototypes to represent, while we assume the same number to be optimal for all events for convenience. Third, while we can interpret our method as an EM algorithm, EM may only converge to a locally optimal solution. As discussed in Sec. 6, future work could explore inductive biases in the pseudolabeling step that favor better solutions. Finally, it is not clear how many related video-/image-based prediction tasks our method can generalize to. While we provided results for TAL in Sec. 5.4, more evidence of adapting our method to related tasks would be useful in future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 653, + 403, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 403, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 403, + 666 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Support from Institute of Eminence (IoE) project No. SB22231269EEETWO005001 for Research Centre in Computer Vision is gratefully acknowledged." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18959" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "text", + "content": "[2] Ricardo Cabral, Fernando De la Torre, Joao Paulo Costeira, and Alexandre Bernardino. Matrix Completion for Weakly-Supervised Multi-Label Image Classification. IEEE transactions on pattern analysis and machine intelligence, 37(1): 121-135, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 257 + ], + "type": "text", + "content": "[3] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 192-208. Springer, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "text", + "content": "[4] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In Computer Vision - ECCV 2022, pages 431-448, Cham, 2022. Springer Nature Switzerland. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 348 + ], + "type": "text", + "content": "[5] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE transactions on information theory, 13(1): 21-27, 1967. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 440 + ], + "type": "text", + "content": "[7] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the Multiple Instance Problem with Axis-Parallel Rectangles. Artificial Intelligence, 89(1):31-71, 1997. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 464 + ], + "type": "text", + "content": "[8] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 518 + ], + "type": "text", + "content": "[9] Jie Fu, Junyu Gao, Bing-Kun Bao, and Changsheng Xu. Multimodal imbalance-aware gradient modulation for weakly-supervised audio-visual video parsing. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "text", + "content": "[10] Adrien Gaidon, Zaid Harchaoui, and Cordelia Schmid. Temporal localization of actions with actons. IEEE transactions on pattern analysis and machine intelligence, 35(11):2782-2795, 2013. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "text", + "content": "[11] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[12] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE transactions on pattern analysis and machine intelligence, 34(3):417-435, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Weifeng Ge, Xiangru Lin, and Yizhou Yu. Weakly Supervised Complementary Parts Models for Fine-Grained Image Classification from the Bottom Up. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3034-3043, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "text", + "content": "[14] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780. IEEE, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[15] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 227 + ], + "type": "text", + "content": "[16] Samantha Guerriero, Barbara Caputo, and Thomas Mensink. DeepNCM: Deep Nearest Class Mean Classifiers, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "type": "text", + "content": "[17] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13925–13935, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "type": "text", + "content": "[20] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 iieme international conference on acoustics, speech and signal processing (icassp), pages 131-135. IEEE, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[21] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly Supervised Image Classification through Noise Regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11517-11525, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "text", + "content": "[22] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Relational prototypical network for weakly supervised temporal action localization. In Proceedings of the AAAI conference on artificial intelligence, pages 11053-11060, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[23] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding, 155:1-23, 2017. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[24] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[25] Saumya Jetley, Bernardino Romera-Paredes, Sadeep Jayasumana, and Philip Torr. Prototypical priors: From improving classification to zero-shot learning. arXiv preprint arXiv:1512.01192, 2015. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18960" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 184 + ], + "type": "text", + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 209 + ], + "type": "text", + "content": "[28] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In ICLR (Poster), 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 212, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 287, + 233 + ], + "type": "text", + "content": "[29] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 48, + 236, + 288, + 280 + ], + "type": "text", + "content": "[30] Jatin Lamba, Jayaprakash Akula, Rishabh Dabral, Preethi Jyothi, Ganesh Ramakrishnan, et al. Cross-modal learning for audio-visual video parsing. arXiv preprint arXiv:2104.04598, 2021. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 281, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 281, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 288, + 335 + ], + "type": "text", + "content": "[31] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002-2006. IEEE, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "text", + "content": "[32] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 395, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 395, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 395, + 288, + 449 + ], + "type": "text", + "content": "[33] Guoqing Liu, Jianxin Wu, and Zhi-Hua Zhou. Key Instance Detection in Multi-Instance Learning. In Proceedings of the Asian Conference on Machine Learning, pages 253–268, Singapore Management University, Singapore, 2012. PMLR. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 452, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 529 + ], + "type": "text", + "content": "[34] Zhekun Luo, Devin Guillory, Baifeng Shi, Wei Ke, Fang Wan, Trevor Darrell, and Huijuan Xu. Weakly-supervised action localization with expectation-maximization multi-instance learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 729-745. Springer, 2020. 2, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 531, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 564 + ], + "type": "text", + "content": "[35] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[36] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In Advances in Neural Information Processing Systems, 2022. 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 688 + ], + "type": "text", + "content": "[37] Islam Nassar, Munawar Hayat, Ehsan Abbasnejad, Hamid Rezatofighi, and Gholamreza Haffari. ProtoCon: Pseudo-Label Refinement via Online Clustering and Prototypical Consistency for Efficient Semi-Supervised Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11641–11650, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding, 2018. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[39] Alejandro Pardo, Human Alwassel, Fabian Caba Heilbron, Ali Thabet, and Bernard Ghanem. RefineLoc: Iterative Refinement for Weakly-Supervised Action Localization. In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3318-3327, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 182 + ], + "type": "text", + "content": "[40] Kranthi Kumar Rachavarapu and A N Rajagopalan. Boosting positive segments for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10192–10202, 2023. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 183, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 547, + 248 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 250, + 546, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 546, + 294 + ], + "type": "text", + "content": "[42] Dawid Rymarczyk, Aneta Kaczyńska, Jarosław Kraus, Adam Pardyl, and Bartosz Zielinski. Protomil: Multiple instance learning with prototypical parts for fine-grained interpretability. arXiv e-prints, pages arXiv-2108, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 294, + 546, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 546, + 337 + ], + "type": "text", + "content": "[43] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood structure. In Artificial intelligence and statistics, pages 412-419. PMLR, 2007. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 338, + 546, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 546, + 370 + ], + "type": "text", + "content": "[44] Atsushi Sato and Keiji Yamada. Generalized learning vector quantization. Advances in neural information processing systems, 8, 1995. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 546, + 404 + ], + "type": "text", + "content": "[45] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "type": "text", + "content": "[46] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple Instance Detection Network with Online Instance Classifier Refinement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2843-2851, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 459, + 546, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 546, + 492 + ], + "type": "text", + "content": "[47] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, pages 247–263, 2018. 1, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 492, + 546, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 546, + 557 + ], + "type": "text", + "content": "[48] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 436–454. Springer, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "text", + "content": "[49] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 614, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 657 + ], + "type": "text", + "content": "[50] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "type": "text", + "content": "[51] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18961" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 596 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[52] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1326–1335, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[53] Zhirong Wu, Alexei A Efros, and Stella X Yu. Improving generalization via scalable neighborhood component analysis. In Proceedings of the European conference on computer vision (ECCV), pages 685-701, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 216 + ], + "type": "text", + "content": "[54] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 261 + ], + "type": "text", + "content": "[55] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 287, + 316 + ], + "type": "text", + "content": "[56] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 287, + 395 + ], + "type": "text", + "content": "[57] Weiyi Yang, Richong Zhang, Junfan Chen, Lihong Wang, and Jaein Kim. Prototype-Guided Pseudo Labeling for Semi-Supervised Text Classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16369-16382, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 397, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 397, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 49, + 397, + 287, + 450 + ], + "type": "text", + "content": "[58] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In European conference on computer vision, pages 37-54. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 453, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 453, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 453, + 287, + 506 + ], + "type": "text", + "content": "[59] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning Deep Features for Discriminative Localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 509, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 509, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 49, + 509, + 287, + 540 + ], + "type": "text", + "content": "[60] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 543, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 543, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 49, + 543, + 287, + 596 + ], + "type": "text", + "content": "[61] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2582-2593, 2022. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18962" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_content_list.json b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1119f128cf0f4dc4bfa5f9444960ca585b123320 --- /dev/null +++ b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_content_list.json @@ -0,0 +1,1446 @@ +[ + { + "type": "text", + "text": "Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingqun Qi $^{1}$ , Jiahao Pan $^{1}$ , Peng Li $^{1}$ , Ruibin Yuan $^{1}$ , Xiaowei Chi $^{1}$ , Mengfei Li $^{1}$ , Wenhan Luo $^{1}$ , Wei Xue $^{1}$ , Shanghang Zhang $^{2}$ , Qifeng Liu $^{1,\\boxtimes}$ , Yike Guo $^{1,\\boxtimes}$ , $^{1}$ The Hong Kong University of Science and Technology $^{2}$ Peking University", + "bbox": [ + 171, + 202, + 795, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xingqun.qi@connect.ust.hk, {liuqifeng, yikeguo}@ust.hk", + "bbox": [ + 240, + 277, + 718, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de4fc90de6e2baa5b3d0bc2d0986790c72f3606ad31cdf70f1f53f2540899ba6.jpg", + "image_caption": [ + "Figure 1. Diverse exemplary clips sampled by our method from our newly collected BEAT Emotion Transition Dataset. The vital frames are visualized to demonstrate that the upper body gestures change with the emotion transition of human speech, synchronously. From top to bottom: the input speech audio, the corresponding transcript, and two sampled clips. Best view on screen." + ], + "image_footnote": [], + "bbox": [ + 104, + 304, + 864, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 580, + 313, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating vivid and emotional 3D co-speech gestures is crucial for virtual avatar animation in human-machine interaction applications. While the existing methods enable generating the gestures to follow a single emotion label, they overlook that long gesture sequence modeling with emotion transition is more practical in real scenes. In addition, the lack of large-scale available datasets with emotional transition speech and corresponding 3D human gestures also limits the addressing of this task. To fulfill this goal, we first incorporate the ChatGPT-4 and an audio inpainting approach to construct the high-fidelity emotion transition human speeches. Considering obtaining the realistic 3D pose annotations corresponding to the dynamically inpainted emotion transition audio is extremely difficult, we propose a novel weakly supervised training strategy to encourage authority gesture transitions. Specifically, to enhance the coordination of transition gestures w.r.t. dif", + "bbox": [ + 75, + 614, + 472, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferent emotional ones, we model the temporal association representation between two different emotional gesture sequences as style guidance and infuse it into the transition generation. We further devise an emotion mixture mechanism that provides weak supervision based on a learnable mixed emotion label for transition gestures. Last, we present a keyframe sampler to supply effective initial posture cues in long sequences, enabling us to generate diverse gestures. Extensive experiments demonstrate that our method outperforms the state-of-the-art models constructed by adapting single emotion-conditioned counterparts on our newly defined emotion transition task and datasets. Our code and dataset will be released on the project page: https://xingqunqi-lab.github.io/Emo-Transition-Gesture/.", + "bbox": [ + 496, + 583, + 893, + 795 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 828, + 632, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Co-speech gesture generation aims to synthesize vivid and emotional human postures coordinated with the audio input. These non-verbal behaviors serve as a key factor dur-", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 99, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10424", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing human conversations that significantly facilitates the delivery of speech content [5, 14, 16]. Meanwhile, modeling co-speech gestures has a wide range of embodied AI applications in human-machine interaction [22, 27], robot assistants [9], and virtual/augmented reality (AR/VR) [11]. Conventionally, many researchers usually focus on synthesizing human upper body gestures consistent with speech audio [32, 47, 48, 50, 53].", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, except for a few recent works that generate the co-speech gestures of a single emotion category [2, 3, 28, 52], previous works mostly focus on emotion-agnostic generation [4, 48, 50]. Most of them overlook synthesizing the long sequence co-speech gestures with the emotion transitions, which are more practical in real-world scenes. For example, a person may not maintain a single emotion forever when communicating with others or in speech talking. In this work, we therefore introduce the task of speech-driven emotion transition for generating vivid and diverse 3D co-speech gestures, displayed in Figure 1. There are two main challenges in this task: 1) Datasets of 3D co-speech gestures synchronized with emotion transition speech audios are scarce. Creating such a dataset containing 3D human postures is difficult due to the lack of guidance from emotional experts and complex motion capture systems. 2) Modeling the plausible and temporal coherent co-speech gestures from one emotion to another in long sequences is difficult, especially in the transition duration.", + "bbox": [ + 75, + 214, + 473, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the issue of data scarcity, we newly present two datasets containing emotion-transition human speech that are built upon previous single-emotion ones [28, 32]. In particular, thanks to the developed language model ChatGPT-4 [34], we first leverage it to generate text transcripts of the transition based on speech context. Then, by employing the audioLDM2 technique [29, 30] for audio inpainting, we ensure the inpainted transition's timbral consistency with its adjacent contexts and a smooth emotional transition throughout. Afterwards, to support our insight on modeling co-speech gestures coherent with emotion-transition speech, corresponding 3D human postures of transition are required.", + "bbox": [ + 75, + 505, + 470, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to the dynamically generated transition transcripts, it is infeasible to construct the aligned realistic 3D pose annotation of human bodies. Hence, we solve the challenges of vivid co-speech gesture generation in a novel weakly supervised pattern, containing a motion transition infusion mechanism and an emotion mixture strategy. Specifically, in the motion transition infusion mechanism, we model the temporal correlation between the generated head and tail gesture features as style guidance representation. The style guidance representation provides motion transition cues that are infused into the transition embedding via an adaptive instance normalization (AdaIN) layer [18]. Along with this manner, we can effectively en", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hance the coordination of transition gestures w.r.t. two different emotional ones.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, to alleviate the lack of supervision during the transition between two emotions, the emotion mixture strategy is built to provide weak emotional supervision of the generated transition gestures. Concretely, we learn a joint embedding of two different emotional gesture sequences using a temporal aggregation encoder. Then, we pre-train an emotion classifier based on the annotated human 3D poses with single emotion labels in the dataset. Here, this joint embedding is leveraged as an emotion mixture weight for the pre-trained classifier to facilitate high-fidelity transition gesture synthesis with desirable properties. Finally, considering the generated 3D postures should be non-deterministic given the human speech, we devise a keyframe sampler to produce diverse initial poses as reference. In this fashion, our method enables diverse co-speech gesture generation with emotion transitions. Extensive experiments conducted on our newly constructed two datasets verify the effectiveness of our methods, displaying vivid and emotional 3D co-speech gestures.", + "bbox": [ + 496, + 121, + 892, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions are summarized as follows:", + "bbox": [ + 500, + 409, + 854, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new task of emotion transition co-speech gesture generation cooperating with two newly constructed datasets named BEAT Emotion Transition (BEAT-ETrans) and TED Emotion Transition (TED-ETrans), significantly facilitating research on 3D human motion modeling.", + "- We design a motion transition infusion mechanism to ensure the temporal coordination of transition gestures w.r.t. two different emotional ones and a weakly supervised emotion mixture strategy to enable high-fidelity transition gesture synthesis with desirable properties.", + "- Extensive experiments show that our method outperforms state-of-the-art counterparts on both datasets, displaying realistic and vivid co-speech gestures given emotion transition human audiios." + ], + "bbox": [ + 500, + 431, + 890, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 685, + 640, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Co-speech Gesture Synthesis", + "text_level": 1, + "bbox": [ + 500, + 710, + 756, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthesizing human co-speech gestures plays a significant role in various applications [17, 22, 37, 46]. Numerous studies have been proposed to address these issues that are roughly divided into rule-based approaches [20, 21], machine-learning-based approaches [23, 41], and deep learning-based ones [2, 28, 32, 36, 48, 50, 52, 53]. Traditional researchers follow the rule-based workflow, leveraging the speech-gesture pairs as guidance to generate co-speech gestures pre-defined by linguistic experts. Other early works integrate the manually defined gesture features with machine learning techniques to synthesize the co", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "speech gestures. In the aforementioned two manners, the researchers usually focus on the optimization of the matching process between human speech and pre-defined gestures. It may require experts much expensive effort in the speech-gesture pair construction.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently works focus on building the mapping directly from the input human speech and sequential gestures by exploiting the deep neural networks. They usually leverage multi-modality cues to facilitate the generation of co-speech gestures, associating with the speech audio [25, 48, 53], speaker identity [50], emotion [2, 28, 31], and text transcript [32]. However, they overlook that emotion transition of the long sequential co-speech gesture modeling is much more practical in the real scenes. Moreover, since the lack of annotated 3D gestures corresponding to dynamically constructed transition speech, few of the above methods could be directly adapted to this new thought.", + "bbox": [ + 75, + 167, + 467, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D Human Motion Modeling", + "text_level": 1, + "bbox": [ + 76, + 357, + 334, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human motion modeling aims to generate realistic and smooth human motions with various multi-modality conditions [8, 19, 33], including co-speech gesture generation as a sub-task. One of the hottest topics is synthesizing human motion from text prompts with a few past postures as the seed [35, 45, 51]. These methods usually engage in forcing human motion to represent the semantic expression aligned with the text. Literally, the task most closely related to ours is AI choreographer [24, 26, 42] which generates the motion from music signals. However, the AI choreographer works mainly focus on the rhythmic-coherent motion of the whole human body but without subtle finger gestures. While sharing a similar goal with the approaches mentioned above, our newly defined work differs from them significantly. We take the emotion-transition long sequence co-speech gesture without corresponding 3D human pose annotation into consideration, thus motivating us to utilize the motion transition of two annotated emotional gestures and coherent the overall sequence.", + "bbox": [ + 75, + 378, + 467, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 680, + 250, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Emotion Transition Dataset Construction", + "text_level": 1, + "bbox": [ + 76, + 704, + 429, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to address the emotion transition co-speech gesture generation in a weakly supervised manner. Due to the existing paired speech-gesture datasets [28, 36], we could focus on synthesizing the high-fidelity transition human audios. Synthesizing datasets conducive to our task focuses on ensuring semantic coherence, smooth emotional audio transitions, consistent timbre, and audio fidelity.", + "bbox": [ + 75, + 728, + 467, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary: Considering our key insight to modeling the co-speech gesture with emotion transition, we first split the existing aligned speech-gesture pairs [28, 36] into four-second clips. To diversify the datasets, we randomly splice", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "two clips from the same speaker to construct an emotion-transition candidate pair. The clip with neutral emotion is leveraged as head speech, and the other with various emotions is represented as tail speech. We leverage the dynamically synthesized two-second audio as a transition to combine the head and tail speeches. In the following, we briefly summarize our efforts in constructing the dynamic transition speech audio.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Textual Inpainting of Transition: To ensure the semantic coherence of transition w.r.t. head/tail speeches, we exploit the advanced language generation model ChatGPT-4 [34] to complete the transcript according to the context. Literally, we follow the conventional estimation that people usually talk 30 phonemes [40] in a two-second speech as the prompt to guide transcript generation by ChatGPT-4.", + "bbox": [ + 496, + 218, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthesis of Transition Audio: Once we obtain the transcript of transitions, we employ a superior text-to-speech model, audioLDM2 [29, 30], to generate corresponding speech audio. Here, we leverage the speaker embeddings extracted via SpeechBrain's ECAPA-TDNN [7, 39] as prior guidance to maintain the identity consistency of the generated transition audio. Then, we adopt Whisper [38] to restrict the word error rate, thereby ensuring the accuracy and clarity of the speech content. In this fashion, the synthesized transition audio realizes controlled duration while the natural smooth tonality is well preserved. The datasets will be released in the future. More details are provided in the Section 4.1.", + "bbox": [ + 496, + 330, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 539, + 700, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of audio signal $A = \\{a_{1},\\dots,a_{N}\\}$ as input to model $\\mathcal{G}$ , our goal is to generate vivid and emotional 3D co-speech gestures $P = \\{p_1,\\dots,p_N\\}$ of the upper human body. $N$ represents the number of synthesized postures corresponding to the audio $A$ . Each $p_i$ is denoted as $J$ joints with 3D representation. In particular, we define the audio signal as consisting of a head speech, a dynamically inpainted transition speech, and a tail speech. Here, the head and tail speeches are randomly selected from previous co-speech gesture datasets [28, 32] with a single emotion label, respectively. Following the conventions of previous works [28, 32, 50, 53], we invoke $M$ frame poses as the initial seed to guide generation. The overall objective is expressed as", + "bbox": [ + 496, + 563, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathcal {G}} {\\arg \\min } \\| P - \\mathcal {G} (A, \\{p _ {1}, \\dots , p _ {M} \\}) \\|. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 787, + 890, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that only the generated gestures with head and tail speeches are supervised with ground truth coming from existing datasets [28, 32]. The transition gestures with $L$ frames, where $L \\ll N$ , will be weakly supervised through the following processes within a motion transition infusion", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10426", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/662578f6ec4cc05fcf2d0fec7f6f331fd96aa0dfcb47d4764f04e953c20b9580.jpg", + "image_caption": [ + "Figure 2. The overview of our proposed method. The middle part (blue) displays the overall pipeline for 3D co-speech gesture generation from emotion transition human speech. The left part (green) depicts our proposed Motion Transition Infusion Mechanism that enhances the coordination of transition gestures w.r.t. head/tail ones. The right part (orange) shows the designed Emotion Mixture Strategy to provide weak supervision of the generated transition gestures, thereby achieving authority producing." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 357, + 359 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb9764782a5de94261561685a9bd967e28e50cc24601da8b5754068c517bf27b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 87, + 604, + 359 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/749191ea0de3bdfb80e452fc3fb24446d9ab2f5055bc03d0fe16a35985c5b9a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 88, + 857, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mechanism and an emotion mixture strategy. The audio signal is fed into an audio encoder for feature extraction. Our overall workflow is shown in Figure 2.", + "bbox": [ + 75, + 446, + 470, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Weakly-supervised Emotion Transition", + "text_level": 1, + "bbox": [ + 76, + 500, + 413, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Transition Infusion Mechanism: To ensure the temporal coherence of the transition gestures w.r.t. head/tail ones, we propose a motion transition infusion mechanism to model the temporal association representation between different emotional gestures. As depicted in Figure 2 (left), the total sequential features $\\mathbf{f}^{total}$ consist of the features of head $\\mathbf{f}^{head}$ , transition $\\mathbf{f}^{tran}$ , and tail $\\mathbf{f}^{tail}$ . Inspired by [33], we nominate a head chunk $\\mathbf{f}_{chunk}^{head}$ composed of the last $L$ frames of head gesture embeddings and a tail chunk $\\mathbf{f}_{chunk}^{tail}$ consisting of the first $L$ ones of the tail gesture embeddings. Here, the dimension of each frame representing gesture embedding is $\\mathbb{R}^{1\\times D}$ .", + "bbox": [ + 75, + 523, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, we first calculate the temporal correlation matrix $\\mathbf{C}_{trans}^{head} \\in \\mathbb{R}^{L \\times L}$ between the head chunk embedding and transition embedding. Here, the temporal correlation matrix represents the temporal variations in the gestures from the head to the transition. Similarly, we obtain the correlation matrix $\\mathbf{C}_{trans}^{tail}$ from tail chunk embedding and transition embedding. Then, the global temporal dependency from head to tail is computed via matrix multiplication between $\\mathbf{C}_{trans}^{head}$ and $\\mathbf{C}_{trans}^{tail}$ . Further, we develop a motion encoder to acquire the sequence-aware style guidance representation based on the global temporal dependency. The style guidance representation is exploited to boost the transition embeddings via an adaptive instance normalization", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(AdaIN) layer [18]. By doing so, we derive the transition $\\mathbf{f}^{tran}$ as", + "bbox": [ + 498, + 446, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {t r a n} = \\operatorname {A d a I N} \\left\\{\\mathbf {f} ^ {t r a n}, \\operatorname {E n c} \\left(\\mathbf {C} _ {t r a n s} ^ {h e a d} \\otimes \\mathbf {C} _ {t r a n s} ^ {t a i l}\\right) \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 481, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ indicates matrix multiplication, and Enc denotes the motion encoder.", + "bbox": [ + 498, + 506, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Emotion Mixture Strategy:** Considering obtaining the realistic 3D pose annotations corresponding to dynamically inpainted transition audio is quite difficult, we design an emotion mixture strategy to provide weak supervision of the generated gestures. Our key insight is built upon the fact that different emotions in the head/tail would lead to different gesture motions, thereby the emotion represented by transition gestures would be a mixture of the head and tail ones. As shown in Figure 2 (right), we utilize two learnable parameters as soft emotion labels of the transition gestures.", + "bbox": [ + 496, + 536, + 892, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, through computing the correlation matrix between the embeddings of head gestures and tail gestures, we obtain the motion deformation representation $S$ from one emotion to another (e.g., neutral-to-anger). This motion deformation is formulated as:", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {\\text {h e a d}} ^ {\\text {t a i l}} = \\operatorname {S o f t m a x} \\left(\\mathbf {f} ^ {\\text {h e a d}} \\otimes \\mathbf {f} ^ {\\text {t a i l} ^ {\\prime}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 765, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where ${}^{\\prime }$ indicates the transpose operation. With the help of the motion deformation representation, we obtain the two learnable emotion embeddings $\\left( {{\\mathbf{e}}^{\\text{head }},{\\mathbf{e}}^{\\text{tail }}}\\right)$ from the head and tail,respectively,as", + "bbox": [ + 496, + 792, + 893, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} ^ {h e a d} = \\operatorname {M a x P o o l} \\left\\{\\mathbf {W} _ {\\text {l i n e a r}} (\\mathbf {f} ^ {h e a d}) \\odot \\mathbf {S} _ {\\text {h e a d}} ^ {t a i l} \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 857, + 852, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} ^ {t a i l} = \\operatorname {M a x P o o l} \\left\\{\\mathbf {W} _ {t a i l} \\left(\\mathbf {f} ^ {h e a d}\\right) \\odot \\mathbf {S} _ {h e a d} ^ {t a i l} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 878, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10427", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{linear}$ denotes linear matrix, and $\\odot$ means dot product. Maxpool indicates the AdaptiveMaxPooling operation. $\\mathbf{e}^{tail}$ is calculated in a similar way. Then, we obtain the two learnable parameters represented by emotion weights of head and tail gestures, as follows", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n(\\sigma , 1 - \\sigma) = \\operatorname {S o f t m a x} \\left(\\mathbf {e} ^ {\\text {h e a d}}, \\mathbf {e} ^ {\\text {t a i l}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 180, + 468, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once we acquire the learnable emotion weights, we leverage a pre-trained pose-based emotion classifier to provide weak supervision. In this fashion, the authority of generated transition gestures is well-preserved. The training details of the pre-trained pose-based emotion classifier are provided in the supplementary material.", + "bbox": [ + 76, + 212, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Keyframe Posture Sampling: Conventionally, researchers [32, 50, 53] directly leverage the padded initial poses as conditional seeds to guide the co-speech gesture generation. In long-sequence modeling, an intuitive manner is to extend the initial pose length $M$ for adapting the overall synthesized gestures' length. However, this would lead to the poor generalizability of the network (i.e., the performance gains degradation from the reduction of initial poses). Therefore, we propose a simple yet effective VAE-based [43] keyframe sampler to provide high-fidelity posture prior conditions while enabling diverse results. Evenly, we split the annotated sequence into several chunks keeping the same length with the transition of length $L$ . The sampler is trained with the keyframe randomly selected to reconstruct the corresponding chunk. In the inference phase, the keyframe sampler samples diverse chunks, blending as the initial postures for gesture generation. Since the transition sequence lacks the pose annotation, we randomly select the frame from head or tail sequences for posture sampling.", + "bbox": [ + 76, + 309, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further enhance the sequence-aware correspondence of the generated co-speech gestures, we leverage the diverse initial postures as the query $Q$ to match the key features $K$ and value features $V$ in the transformer-based backbone [44]. Similar to [37], we adopt a motion discriminator to ensure the temporal smoothness of the generated results. For more details about network architecture please refer to supplementary material.", + "bbox": [ + 76, + 598, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Objective Functions", + "text_level": 1, + "bbox": [ + 76, + 738, + 266, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction Loss: We leverage the ground truth 3D pose annotation of the head and tail to constrain the generated co-speech gestures as:", + "bbox": [ + 76, + 768, + 468, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left\\| P _ {\\{h e a d, t a i l \\}} - \\hat {P} _ {\\{h e a d, t a i l \\}} \\right\\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 827, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{P}_{\\{\\text{head}, \\text{tail}\\}}$ denotes generated gestures of head and tail speeches.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Learning Loss: To ensure the realism of the generated gestures, we further exploit the adversarial training loss, expressed as:", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a d v} = \\mathbb {E} _ {P} [ \\log \\mathcal {D} (P) ] \\\\ + \\mathbb {E} _ {A} [ \\log (1 - (\\mathcal {G} (A, \\{p _ {1}, \\dots , p _ {M} \\})) ], \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 143, + 890, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ denotes the motion discriminator and $\\mathcal{G}$ means gesture generator.", + "bbox": [ + 498, + 186, + 890, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Weakly Supervision Loss: We leverage the pre-trained pose-based emotion classifier to provide weak supervision of the transition gestures upon the learnable emotion weights:", + "bbox": [ + 498, + 223, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e m o t i o n}} = - y \\log \\mathcal {F} (\\hat {P} _ {\\text {t r a n s}}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 290, + 890, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ is the learnable emotion label, $\\mathcal{F}$ is the emotion classifier, $\\hat{P}_{trans}$ is the generated transition gestures.", + "bbox": [ + 498, + 315, + 890, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the overall objective is:", + "bbox": [ + 500, + 345, + 709, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G} \\max _ {D} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {r} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\mathcal {L} _ {\\text {e m o t i o n}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 368, + 890, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $\\lambda_r$ , and $\\lambda_{adv}$ are weight coefficients.", + "bbox": [ + 500, + 397, + 772, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 422, + 630, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Setting", + "text_level": 1, + "bbox": [ + 500, + 448, + 802, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BEAT Emotion Transition Dataset (BEAT-ETrans): Since there are only single emotion labels of aligned speech-gesture corpus in the original BEAT dataset [28], to satisfy our insight on emotion transition co-speech gesture generation modeling, we recollect a BEAT Emotion Transition Dataset (dubbed BEAT-ETrans). In particular, we resample the motion FPS as 15 and intercept the continuous 60 frames with stride 30 as the head/tail clips. Here, the head clips are all annotated as neutral, and tail clips are denoted with the other seven emotions: anger, happiness, fear, disgust, sadness, contempt, and surprise. As for one head speech, we randomly select two or three tails with different emotions to construct the head-tail pairs. By leveraging the two-second (i.e., corresponding 30-frame postures) transition to blend the heads and tails, we obtain the 10-second human speech clips in our BEAT-ETrans. We obtain 58,077 clips, including a total of 161.3 hours reported in Table 1. Then the clip numbers of training/validation/testing sets are randomly split as 41,908/4,077/12,092. In all of our experiments, we utilize the upper body with 71 joints.", + "bbox": [ + 498, + 470, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TED Emotion Transition Dataset (TED-ETrans): Inspired by [32, 36], we further newly collect a TED Emotion Transition Dataset (dubbed TED-ETrans) based on more than 1.7K speakers from in-the-wild TED talk show videos, demonstrated in Table 1. Due to the lack of emotional labels, we first leverage the annotated BEAT dataset to pretrain an audio-based emotion classifier for labeling TED audi-", + "bbox": [ + 498, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10428", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/256577836683683bad762129682403823f5c440ca775c0c053521c894b2d65c4.jpg", + "table_caption": [ + "Table 1. Statistics comparison of existing 3D co-speech gesture datasets with ours. Our BEAT-ETrans and TED-ETrans are built upon the existing BEAT [28] and TED-Expressive [32], respectively. To the best of our knowledge, we are the first to present two large datasets with emotion transition human speech." + ], + "table_footnote": [], + "table_body": "
DatasetJoint AnnotationModalityDuration (hours)
BodyHandAudioTextSpeakersSingle EmotionEmotion Transition
TED [50]pseudo label9✘✓✓1,766✘✘106.1
SCG [15]pseudo label1424✓✘6✘✘33
Trinity [10]mo-cap2438✓✓1✘✘4
ZeroEGGS [12]mo-cap2748✓✓1✘✘2
BEAT [28]mo-cap2748✓✓30✓✘35
TED-Expressive [32]pseudo label1330✓✓1,764✘✘100.8
BEAT-ETrans (ours)mo-cap2748✓✓308✓161.3
TED-ETrans (ours)pseudo label1330✓✓1,7646✓59.8
", + "bbox": [ + 76, + 142, + 890, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/304c76c8b14819d76fd137a816c54a3b442e0483eaba4d11e184defcedafc218.jpg", + "table_caption": [ + "Table 2. Comparison with the start-of-the-art methods on our newly collected BEAT-ETrans and TED-ETrans datasets. $\\uparrow$ denotes the higher the better, and $\\downarrow$ indicates the lower the better. $\\pm$ means ${95}\\%$ confidence interval." + ], + "table_footnote": [], + "table_body": "
ModelsBEAT-ETransTED-ETrans
FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑
Seq2Seq [49]ICRA'1940.9547.930.14196.66±2.1629.6049.470.26572.81±1.99
S2G [13]CVPR'1925.5637.040.67198.26±2.0418.1641.630.82476.82±2.32
Trimodal [50]TOG'2014.0942.500.764100.87±2.1221.0633.200.75882.87±1.86
CAMN [28]ECCV'229.0327.530.794118.46±2.3319.2841.040.78579.03±1.49
HA2G [32]CVPR'227.2825.790.779121.77±2.3116.7240.380.78780.14±1.65
DiffGesture [53]CVPR'236.6825.030.788122.29±2.0118.6925.130.81892.01±2.07
Ours4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
", + "bbox": [ + 86, + 344, + 883, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classification threshold as $\\geq 0.95$ , and the two uncommon emotions (i.e. fear, disgust) are dropped. Then we maintain the same data pre-processing strategy with our BEAT-ETrans to obtain a total of 21,515 clips with 59.8 hours. The final clip division criteria of the TED-ETrans dataset are training/validation/testing with 15,061/2,152/4,302 respectively. In practice, the 43 upper body joints are leveraged in the experiments.", + "bbox": [ + 75, + 510, + 470, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: We set the total generated cospeech gesture length as $N = 150$ , and the transition and chunk lengths are $L = 30$ . Conventionally, we leverage $M = 4$ frames as the reference initial poses. The feature dimension $D = 512$ in practice. The raw audio of human speech is converted to mel-spectrograms with FFT window size 1024, and hop length 512. The audio encoder takes the ResNetSE34 [6] as the backbone. Empirically, we set $\\lambda_r = 20$ , and $\\lambda_{adv} = 2$ . Our models are implemented on the Pytorch platform with a single NVIDIA Tesla V100 GPU. The initial learning rate is set to 0.0003 by utilizing Adam Optimizer. The whole training takes 100 epochs with a batch size of 96.", + "bbox": [ + 75, + 637, + 468, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: To fully evaluate the realism and diversity of the generated co-speech gestures, we introduce various metrics:", + "bbox": [ + 75, + 839, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- FGD: Fréchet Gesture Distance (FGD) [50] is utilized to", + "bbox": [ + 76, + 885, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "measure the distribution distance between the realistic sequential gestures and generated ones. Since we only have the 3D joint annotation of the head/tail, we take the network architecture provided in [32, 50] to train the autoencoder for distance computing on the two datasets, respectively. The FGD of transition gestures is calculated as the average value between the distribution distance of transition and head/tail, indicated as $\\mathrm{FGD}_{\\text {trans }}$ . Similarly, $\\mathrm{FGD}_{h + t}$ means the distance between the generated head/tail gestures and ground truth.", + "bbox": [ + 509, + 510, + 890, + 660 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- BC: Beat Consistency Score (BC) [28, 32] measures the speech audio alignment degree with the generated co-speech gestures.", + "- Diversity: Similar to [32, 53], we exploit the same feature extractor in FGD to obtain the feature embeddings of the generated gestures. The diversity reflects the average distance between 500 random combination pairs in the testing set of 60 speech audios." + ], + "bbox": [ + 500, + 661, + 890, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 789, + 720, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with SOTA Methods: To the best of our knowledge, we are the first to explore the co-speech gesture generation with emotion transition human audios. To fully verify the superiority of our method, we implement various state-of-the-art single-emotion-based counterparts: Seq2Seq [49], S2G [13], Trimodal [50], CAMN [28],", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10429", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e2b839b96d9dde43712d7a432ce3cd36af2ffa2bcf1d58974e9644383f1a35a5.jpg", + "table_caption": [ + "Table 3. Ablation study on different components of the proposed method. $\\checkmark$ indicates the employment of a certain module. $\\uparrow$ denotes the higher the better, and $\\downarrow$ indicates the lower the better. $\\pm$ means $95\\%$ confidence interval. MTIM: Motion Transition Infusion Mechanism; EMS: Emotion Mixture Strategy; FKS: Keyframe Sampler." + ], + "table_footnote": [], + "table_body": "
Model VariationsBEAT-ETransTED-ETrans
BaselineMTIMEMSKFSFGDh+t↓FGDtrans↓BC↑Diversity ↑FGDh+t↓FGDtrans↓BC↑Diversity ↑
✓21.2156.330.70188.89±1.5826.4447.870.71274.19±2.75
✓✓10.7527.940.866108.69±2.6320.9536.380.82782.42±2.24
✓✓✓5.6921.700.878112.80±1.9514.0226.210.90087.21±2.12
✓✓✓✓4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
", + "bbox": [ + 84, + 142, + 883, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HA2G [32], and DiffGesture [53]. For a fair comparison, all the models are implemented by the source code released by the authors. Note that since we only have the ground truth of head and tail gestures, several recent VAE-based works [1, 2, 48, 52] cannot be directly applied in the experiments (or they have not released codes so far).", + "bbox": [ + 75, + 265, + 468, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As reported in Table 2, we adopt the $\\mathrm{FGD}_{h + t}$ , $\\mathrm{FGD}_{trans}$ , BC, and Diversity for a well-rounded view of comparisons. Our method outperforms all the competitors by a large margin on both two datasets. Remarkably, on the TED-ETrans dataset, our method even achieves $34.8\\%$ (i.e., $(18.69 - 12.19) / 18.69 \\approx 34.8\\%$ ) improvement over the sub-optimal counterparts in $\\mathrm{FGD}_{h + t}$ . We observe both the DiffGesture [53] and ours synthesize the high-fidelity gestures of head/tail speech with much lower $\\mathrm{FGD}_{h + t}$ than others. However, the DiffGesture shows worse performance on $\\mathrm{FGD}_{trans}$ due to lack of supervision. In terms of diversity, our simple yet effective keyframe sampler provides authority and diverse initial postures as the reference, thus enabling us to demonstrate diverse gesture styles compared to other counterparts. Moreover, we find that the diversity scores on BEAT-ETrans are much higher than those on TED-ETrans dataset. This can be attributed to the more complex human joints in the BEAT-ETrans dataset.", + "bbox": [ + 75, + 359, + 470, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study: To further verify the effectiveness of our proposed methods, we conduct the ablation study of different components as variations, reported in Table 3. The baseline model is implemented by a simple transformer-based pipeline with stacking three times blocks in the encoder-decoder. Obviously, all the combinations of our proposed components have positive impacts on the generated results. Specifically, by adding the motion transition infusion mechanism to the baseline, the indicator BC has achieved significant improvement (e.g., $0.701 \\rightarrow 0.866$ in the BEAT-Etrans). This result verifies that our motion transition infusion mechanism effectively models the temporal correlation between the transitions w.r.t. head/tail gestures, thus leading to the generated results preserving rhythm coherency with given speech, globally.", + "bbox": [ + 75, + 640, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, adopting the emotion mixture strategy ideally improves the performance of $\\mathrm{FGD}_{trans}$ on both two", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "datasets. This indicates that the learnable emotion mixture wights can provide effective weak supervision by leveraging the pre-trained pose-based emotion classifier. Besides, we have observed significant improvement in the performance of $\\mathrm{FGD}_{h + t}$ during this phase compared to the previous version. This aligns with our transformer backbone's emphasis on modeling the sequential temporal correlations as a whole. The better transition gestures encourage our model to maintain better temporal consistency, thereby the head/tail gestures achieve better results.", + "bbox": [ + 496, + 265, + 890, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, after additionally employing the keyframe sampler to produce authority initial postures as the reference, our method obtains the best performance. Although the properties of BC and FGD on both datasets just have slightly improvement, the diversity realizes a noticeably better achievement (e.g., $112.80 \\rightarrow 124.93$ in the BEAT-ETrans dataset). This highly supports our insight into keyframe-driven diversification strategy.", + "bbox": [ + 496, + 419, + 892, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 556, + 710, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization: To fully demonstrate the performance of our method, we show the visualized keyframes generated from ours compared with counterparts on our newly collected TED-ETrans and BEAT-ETrans datasets, respectively. As depicted in Figure 3, our method displays vivid and diverse results against others. In particular, we observe that Seq2Seq and Trimodal tend to synthesize unreasonable and stiff results (e.g., the blue rectangle in the right BEAT-ETrans dataset). Although CAMN and HA2G can generate natural upper-body postures, we find that they sometimes produce unreliable subtle fingers (e.g., the red rectangle in the left TED-ETrans dataset). Both our method and DiffGesture create reasonable gestures. However, the results synthesized by DiffuGesture are mismatched with the emotion transitions. In contrast, our method can synthesize the synchronous motions (e.g., in the BEAT-ETrans, the arms become droopy as emotion turns to sadness). Meanwhile, we further verify the diversification results as shown in Figure 1. Given the same input audio, our method generates diverse and vivid co-speech gestures. Please refer to the supplementary material for more visualization results.", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10430", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76cca928db845133bbc5a0af9f746f3de94363a40259366ce6dc6f317dd9af4d.jpg", + "image_caption": [ + "Figure 3. Visualization of our generated 3D co-speech gestures against various state-of-the-art methods. The samples of the left part are from our newly collected TED-ETrans dataset, and the samples of the right part are from our BEAT-ETrans dataset. Best view on screen." + ], + "image_footnote": [], + "bbox": [ + 122, + 87, + 849, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study: To further analyze the quality of results by various counterparts and ours, we conduct a user study by visualizing the results and inviting 15 subjects. For each emotion in both datasets, we randomly select 12 samples for each participant (7 emotions in BEAT-ETrans, 5 emotions in TED-ETrans). All the participants are recruited anonymously from schools including various majors. In particular, the subjects are required to rate the generated co-speech gestures from 0 to 5 (the higher, the better) in terms of naturalness, motion smoothness, and speech-gesture coherency. The results are demonstrated in Figure 4. Our method showcases the best performance compared with all the competitors. Especially in terms of motion smoothness, our method achieves noticeable advantages, indicating the effectiveness of our proposed motion transition infusion mechanism and the emotion mixture strategy.", + "bbox": [ + 75, + 508, + 472, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 767, + 194, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a new task of 3D co-speech gesture generation given emotion transition human speech. We therefore newly collected two datasets named the BEAT-ETrans and the TED-ETrans to fulfill this goal while significantly facilitating the research on 3D human motion modeling. Then, we fully take advantage of the sequential temporal correlation via a motion transition infusion mechanism", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f33d9cd1a3b6d664abb605ba5c08d7a67132d8284afbbf4724235a8e1af55d33.jpg", + "image_caption": [ + "Figure 4. User study on gesture naturalness, motion smoothness, and speech-gesture coherency." + ], + "image_footnote": [], + "bbox": [ + 517, + 508, + 875, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to ensure the generated gestures preserve temporal coherence. Furthermore, we design an emotion mixture strategy to supply emotional weak supervision of the synthesized transitions. Extensive experiments conducted on our two newly collected datasets show the superiority of the method. As our method intends to generate diverse and vivid emotion transition gestures, we will investigate diversifying the 3D gesture with temporal smooth sampling, instead of the keyframe-wise manner.", + "bbox": [ + 496, + 709, + 890, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is funded in part by the Theme-based Research Scheme grant (No.T45-205/21-N) and the InnoHK funding, Hong Kong SAR.", + "bbox": [ + 496, + 851, + 890, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 7", + "[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2, 3, 7", + "[3] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2027-2036, 2021. 2", + "[4] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 2", + "[5] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 2", + "[6] Joon Son Chung, Jaesung Huh, Seongkyu Mun, Minjae Lee, Hee Soo Heo, Soyeon Choe, Chiheon Ham, Sunghwan Jung, Bong-Jin Lee, and Icksang Han. In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982, 2020. 6", + "[7] Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. arXiv preprint arXiv:2005.07143, 2020. 3", + "[8] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. arXiv preprint arXiv:2305.08953, 2023. 3", + "[9] Maged Farouk. Studying human robot interaction and its characteristics. International Journal of Computations, Information and Manufacturing (IJCIM), 2(1), 2022. 2", + "[10] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 6", + "[11] Yu Fu, Yan Hu, and Veronica Sundstedt. A systematic literature review of virtual, augmented, and mixed reality game applications in healthcare. ACM Transactions on Computing for Healthcare (HEALTH), 3(2):1-27, 2022. 2", + "[12] Saeed Ghorbani, Ylva Fe2rstl, Daniel Holden, Nikolaus F Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. In Computer Graphics Forum, pages 206–216. Wiley Online Library, 2023. 6", + "[13] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Susan Goldin-Meadow and Martha Wagner Alibali. Gesture's role in speaking, learning, and creating language. Annual review of psychology, 64:257-283, 2013. 2", + "[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 6", + "[16] Autumn B Hostetter and Martha W Alibali. Visible embodiment: Gestures as simulated action. Psychonomic bulletin & review, 15:495-514, 2008. 2", + "[17] Ann Huang, Pascal Knierim, Francesco Chiossi, Lewis L Chuang, and Robin Welsch. Proxemics for human-agent interaction in augmented reality. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-13, 2022. 2", + "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4", + "[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 3", + "[20] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2", + "[21] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thórisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2", + "[22] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities for reactive robotic response. In IROS, page 2071. Tokyo, 2013. 2", + "[23] Sergey Levine, Philipp Krahenbuhl, Sebastian Thrun, and Vladlen Koltun. Gesture controllers. In Acm siggraph 2010 papers, pages 1-11. 2010. 2", + "[24] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020.3", + "[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 3", + "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10432", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7590–7598, 2023. 2", + "[28] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6", + "[29] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. AudioLDM: Text-to-audio generation with latent diffusion models. Proceedings of the International Conference on Machine Learning, 2023. 2, 3", + "[30] Haohe Liu, Qiao Tian, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D. Plumbley. AudioLDM 2: Learning holistic audio generation with self-supervised pretraining. arXiv preprint arXiv:2308.05734, 2023. 2, 3", + "[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via masked audio gesture modeling. arXiv preprint arXiv:2401.00374, 2023. 3", + "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 5, 6, 7", + "[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 3, 4", + "[34] OpenAI. Gpt-4 technical report, 2023. 2, 3", + "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3", + "[36] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2, 3, 5", + "[37] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3d hand gesture prediction from body dynamics by bilateral hand disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4616-4626, 2023. 2, 5", + "[38] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 3", + "[39] Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan," + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, et al. Speechbrain: A general-purpose speech toolkit. arXiv preprint arXiv:2106.04624, 2021. 3", + "[40] Peter Roach. Some languages are spoken more quickly than others. 1998. 3", + "[41] Mehmet E Sargin, Yucel Yemez, Engin Erzin, and Ahmet M Tekalp. Analysis of head gesture and prosody patterns for prosody-driven head-gesture animation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(8):1330-1345, 2008. 2", + "[42] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 3", + "[43] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 5", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[45] Jordan Voas, Yili Wang, Qixing Huang, and Raymond Mooney. What is the best automated metric for text to motion generation? arXiv preprint arXiv:2309.10248, 2023. 3", + "[46] Zijian Wang, Xingqun Qi, Kun Yuan, and Muyi Sun. Self-supervised correlation mining network for person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7703-7712, 2022. 2", + "[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 5860–5868, 2023. 2", + "[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 2, 3, 7", + "[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 6", + "[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 2, 3, 5, 6", + "[51] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In International Cconference on Multimodal Interaction, pages 179-185. 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10433", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 7", + "[53] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10434", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_model.json b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3ef70b9afb21f516fd7f6ec68249a74c312daef --- /dev/null +++ b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_model.json @@ -0,0 +1,2059 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.131, + 0.753, + 0.176 + ], + "angle": 0, + "content": "Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.203, + 0.796, + 0.275 + ], + "angle": 0, + "content": "Xingqun Qi\\(^{1}\\), Jiahao Pan\\(^{1}\\), Peng Li\\(^{1}\\), Ruibin Yuan\\(^{1}\\), Xiaowei Chi\\(^{1}\\), Mengfei Li\\(^{1}\\), Wenhan Luo\\(^{1}\\), Wei Xue\\(^{1}\\), Shanghang Zhang\\(^{2}\\), Qifeng Liu\\(^{1,\\boxtimes}\\), Yike Guo\\(^{1,\\boxtimes}\\), \\(^{1}\\) The Hong Kong University of Science and Technology \\(^{2}\\) Peking University" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.278, + 0.72, + 0.293 + ], + "angle": 0, + "content": "xingqun.qi@connect.ust.hk, {liuqifeng, yikeguo}@ust.hk" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.305, + 0.866, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.527, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Figure 1. Diverse exemplary clips sampled by our method from our newly collected BEAT Emotion Transition Dataset. The vital frames are visualized to demonstrate that the upper body gestures change with the emotion transition of human speech, synchronously. From top to bottom: the input speech audio, the corresponding transcript, and two sampled clips. Best view on screen." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.582, + 0.314, + 0.599 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.473, + 0.873 + ], + "angle": 0, + "content": "Generating vivid and emotional 3D co-speech gestures is crucial for virtual avatar animation in human-machine interaction applications. While the existing methods enable generating the gestures to follow a single emotion label, they overlook that long gesture sequence modeling with emotion transition is more practical in real scenes. In addition, the lack of large-scale available datasets with emotional transition speech and corresponding 3D human gestures also limits the addressing of this task. To fulfill this goal, we first incorporate the ChatGPT-4 and an audio inpainting approach to construct the high-fidelity emotion transition human speeches. Considering obtaining the realistic 3D pose annotations corresponding to the dynamically inpainted emotion transition audio is extremely difficult, we propose a novel weakly supervised training strategy to encourage authority gesture transitions. Specifically, to enhance the coordination of transition gestures w.r.t. dif" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.796 + ], + "angle": 0, + "content": "ferent emotional ones, we model the temporal association representation between two different emotional gesture sequences as style guidance and infuse it into the transition generation. We further devise an emotion mixture mechanism that provides weak supervision based on a learnable mixed emotion label for transition gestures. Last, we present a keyframe sampler to supply effective initial posture cues in long sequences, enabling us to generate diverse gestures. Extensive experiments demonstrate that our method outperforms the state-of-the-art models constructed by adapting single emotion-conditioned counterparts on our newly defined emotion transition task and datasets. Our code and dataset will be released on the project page: https://xingqunqi-lab.github.io/Emo-Transition-Gesture/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.829, + 0.633, + 0.844 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Co-speech gesture generation aims to synthesize vivid and emotional human postures coordinated with the audio input. These non-verbal behaviors serve as a key factor dur-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.245, + 0.901 + ], + "angle": 0, + "content": "Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10424" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "ing human conversations that significantly facilitates the delivery of speech content [5, 14, 16]. Meanwhile, modeling co-speech gestures has a wide range of embodied AI applications in human-machine interaction [22, 27], robot assistants [9], and virtual/augmented reality (AR/VR) [11]. Conventionally, many researchers usually focus on synthesizing human upper body gestures consistent with speech audio [32, 47, 48, 50, 53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.474, + 0.502 + ], + "angle": 0, + "content": "Nevertheless, except for a few recent works that generate the co-speech gestures of a single emotion category [2, 3, 28, 52], previous works mostly focus on emotion-agnostic generation [4, 48, 50]. Most of them overlook synthesizing the long sequence co-speech gestures with the emotion transitions, which are more practical in real-world scenes. For example, a person may not maintain a single emotion forever when communicating with others or in speech talking. In this work, we therefore introduce the task of speech-driven emotion transition for generating vivid and diverse 3D co-speech gestures, displayed in Figure 1. There are two main challenges in this task: 1) Datasets of 3D co-speech gestures synchronized with emotion transition speech audios are scarce. Creating such a dataset containing 3D human postures is difficult due to the lack of guidance from emotional experts and complex motion capture systems. 2) Modeling the plausible and temporal coherent co-speech gestures from one emotion to another in long sequences is difficult, especially in the transition duration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.702 + ], + "angle": 0, + "content": "To overcome the issue of data scarcity, we newly present two datasets containing emotion-transition human speech that are built upon previous single-emotion ones [28, 32]. In particular, thanks to the developed language model ChatGPT-4 [34], we first leverage it to generate text transcripts of the transition based on speech context. Then, by employing the audioLDM2 technique [29, 30] for audio inpainting, we ensure the inpainted transition's timbral consistency with its adjacent contexts and a smooth emotional transition throughout. Afterwards, to support our insight on modeling co-speech gestures coherent with emotion-transition speech, corresponding 3D human postures of transition are required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "However, due to the dynamically generated transition transcripts, it is infeasible to construct the aligned realistic 3D pose annotation of human bodies. Hence, we solve the challenges of vivid co-speech gesture generation in a novel weakly supervised pattern, containing a motion transition infusion mechanism and an emotion mixture strategy. Specifically, in the motion transition infusion mechanism, we model the temporal correlation between the generated head and tail gesture features as style guidance representation. The style guidance representation provides motion transition cues that are infused into the transition embedding via an adaptive instance normalization (AdaIN) layer [18]. Along with this manner, we can effectively en" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "hance the coordination of transition gestures w.r.t. two different emotional ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.893, + 0.409 + ], + "angle": 0, + "content": "Moreover, to alleviate the lack of supervision during the transition between two emotions, the emotion mixture strategy is built to provide weak emotional supervision of the generated transition gestures. Concretely, we learn a joint embedding of two different emotional gesture sequences using a temporal aggregation encoder. Then, we pre-train an emotion classifier based on the annotated human 3D poses with single emotion labels in the dataset. Here, this joint embedding is leveraged as an emotion mixture weight for the pre-trained classifier to facilitate high-fidelity transition gesture synthesis with desirable properties. Finally, considering the generated 3D postures should be non-deterministic given the human speech, we devise a keyframe sampler to produce diverse initial poses as reference. In this fashion, our method enables diverse co-speech gesture generation with emotion transitions. Extensive experiments conducted on our newly constructed two datasets verify the effectiveness of our methods, displaying vivid and emotional 3D co-speech gestures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.41, + 0.855, + 0.424 + ], + "angle": 0, + "content": "Overall, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.432, + 0.892, + 0.522 + ], + "angle": 0, + "content": "- We introduce a new task of emotion transition co-speech gesture generation cooperating with two newly constructed datasets named BEAT Emotion Transition (BEAT-ETrans) and TED Emotion Transition (TED-ETrans), significantly facilitating research on 3D human motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.529, + 0.892, + 0.605 + ], + "angle": 0, + "content": "- We design a motion transition infusion mechanism to ensure the temporal coordination of transition gestures w.r.t. two different emotional ones and a weakly supervised emotion mixture strategy to enable high-fidelity transition gesture synthesis with desirable properties." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.611, + 0.892, + 0.67 + ], + "angle": 0, + "content": "- Extensive experiments show that our method outperforms state-of-the-art counterparts on both datasets, displaying realistic and vivid co-speech gestures given emotion transition human audiios." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.432, + 0.892, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.686, + 0.642, + 0.701 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.757, + 0.727 + ], + "angle": 0, + "content": "2.1. Co-speech Gesture Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Synthesizing human co-speech gestures plays a significant role in various applications [17, 22, 37, 46]. Numerous studies have been proposed to address these issues that are roughly divided into rule-based approaches [20, 21], machine-learning-based approaches [23, 41], and deep learning-based ones [2, 28, 32, 36, 48, 50, 52, 53]. Traditional researchers follow the rule-based workflow, leveraging the speech-gesture pairs as guidance to generate co-speech gestures pre-defined by linguistic experts. Other early works integrate the manually defined gesture features with machine learning techniques to synthesize the co" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "speech gestures. In the aforementioned two manners, the researchers usually focus on the optimization of the matching process between human speech and pre-defined gestures. It may require experts much expensive effort in the speech-gesture pair construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.348 + ], + "angle": 0, + "content": "Recently works focus on building the mapping directly from the input human speech and sequential gestures by exploiting the deep neural networks. They usually leverage multi-modality cues to facilitate the generation of co-speech gestures, associating with the speech audio [25, 48, 53], speaker identity [50], emotion [2, 28, 31], and text transcript [32]. However, they overlook that emotion transition of the long sequential co-speech gesture modeling is much more practical in the real scenes. Moreover, since the lack of annotated 3D gestures corresponding to dynamically constructed transition speech, few of the above methods could be directly adapted to this new thought." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.358, + 0.336, + 0.373 + ], + "angle": 0, + "content": "2.2. 3D Human Motion Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.468, + 0.668 + ], + "angle": 0, + "content": "Human motion modeling aims to generate realistic and smooth human motions with various multi-modality conditions [8, 19, 33], including co-speech gesture generation as a sub-task. One of the hottest topics is synthesizing human motion from text prompts with a few past postures as the seed [35, 45, 51]. These methods usually engage in forcing human motion to represent the semantic expression aligned with the text. Literally, the task most closely related to ours is AI choreographer [24, 26, 42] which generates the motion from music signals. However, the AI choreographer works mainly focus on the rhythmic-coherent motion of the whole human body but without subtle finger gestures. While sharing a similar goal with the approaches mentioned above, our newly defined work differs from them significantly. We take the emotion-transition long sequence co-speech gesture without corresponding 3D human pose annotation into consideration, thus motivating us to utilize the motion transition of two annotated emotional gestures and coherent the overall sequence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.25, + 0.697 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.705, + 0.43, + 0.72 + ], + "angle": 0, + "content": "3.1. Emotion Transition Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.468, + 0.834 + ], + "angle": 0, + "content": "We aim to address the emotion transition co-speech gesture generation in a weakly supervised manner. Due to the existing paired speech-gesture datasets [28, 36], we could focus on synthesizing the high-fidelity transition human audios. Synthesizing datasets conducive to our task focuses on ensuring semantic coherence, smooth emotional audio transitions, consistent timbre, and audio fidelity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Preliminary: Considering our key insight to modeling the co-speech gesture with emotion transition, we first split the existing aligned speech-gesture pairs [28, 36] into four-second clips. To diversify the datasets, we randomly splice" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "two clips from the same speaker to construct an emotion-transition candidate pair. The clip with neutral emotion is leveraged as head speech, and the other with various emotions is represented as tail speech. We leverage the dynamically synthesized two-second audio as a transition to combine the head and tail speeches. In the following, we briefly summarize our efforts in constructing the dynamic transition speech audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Textual Inpainting of Transition: To ensure the semantic coherence of transition w.r.t. head/tail speeches, we exploit the advanced language generation model ChatGPT-4 [34] to complete the transcript according to the context. Literally, we follow the conventional estimation that people usually talk 30 phonemes [40] in a two-second speech as the prompt to guide transcript generation by ChatGPT-4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Synthesis of Transition Audio: Once we obtain the transcript of transitions, we employ a superior text-to-speech model, audioLDM2 [29, 30], to generate corresponding speech audio. Here, we leverage the speaker embeddings extracted via SpeechBrain's ECAPA-TDNN [7, 39] as prior guidance to maintain the identity consistency of the generated transition audio. Then, we adopt Whisper [38] to restrict the word error rate, thereby ensuring the accuracy and clarity of the speech content. In this fashion, the synthesized transition audio realizes controlled duration while the natural smooth tonality is well preserved. The datasets will be released in the future. More details are provided in the Section 4.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.702, + 0.554 + ], + "angle": 0, + "content": "3.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Given a sequence of audio signal \\( A = \\{a_{1},\\dots,a_{N}\\} \\) as input to model \\( \\mathcal{G} \\), our goal is to generate vivid and emotional 3D co-speech gestures \\( P = \\{p_1,\\dots,p_N\\} \\) of the upper human body. \\( N \\) represents the number of synthesized postures corresponding to the audio \\( A \\). Each \\( p_i \\) is denoted as \\( J \\) joints with 3D representation. In particular, we define the audio signal as consisting of a head speech, a dynamically inpainted transition speech, and a tail speech. Here, the head and tail speeches are randomly selected from previous co-speech gesture datasets [28, 32] with a single emotion label, respectively. Following the conventions of previous works [28, 32, 50, 53], we invoke \\( M \\) frame poses as the initial seed to guide generation. The overall objective is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.788, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathcal {G}} {\\arg \\min } \\| P - \\mathcal {G} (A, \\{p _ {1}, \\dots , p _ {M} \\}) \\|. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Note that only the generated gestures with head and tail speeches are supervised with ground truth coming from existing datasets [28, 32]. The transition gestures with \\( L \\) frames, where \\( L \\ll N \\), will be weakly supervised through the following processes within a motion transition infusion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.359, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.088, + 0.605, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.089, + 0.858, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 2. The overview of our proposed method. The middle part (blue) displays the overall pipeline for 3D co-speech gesture generation from emotion transition human speech. The left part (green) depicts our proposed Motion Transition Infusion Mechanism that enhances the coordination of transition gestures w.r.t. head/tail ones. The right part (orange) shows the designed Emotion Mixture Strategy to provide weak supervision of the generated transition gestures, thereby achieving authority producing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.495 + ], + "angle": 0, + "content": "mechanism and an emotion mixture strategy. The audio signal is fed into an audio encoder for feature extraction. Our overall workflow is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.414, + 0.516 + ], + "angle": 0, + "content": "3.3. Weakly-supervised Emotion Transition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.705 + ], + "angle": 0, + "content": "Motion Transition Infusion Mechanism: To ensure the temporal coherence of the transition gestures w.r.t. head/tail ones, we propose a motion transition infusion mechanism to model the temporal association representation between different emotional gestures. As depicted in Figure 2 (left), the total sequential features \\(\\mathbf{f}^{total}\\) consist of the features of head \\(\\mathbf{f}^{head}\\), transition \\(\\mathbf{f}^{tran}\\), and tail \\(\\mathbf{f}^{tail}\\). Inspired by [33], we nominate a head chunk \\(\\mathbf{f}_{chunk}^{head}\\) composed of the last \\(L\\) frames of head gesture embeddings and a tail chunk \\(\\mathbf{f}_{chunk}^{tail}\\) consisting of the first \\(L\\) ones of the tail gesture embeddings. Here, the dimension of each frame representing gesture embedding is \\(\\mathbb{R}^{1\\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In particular, we first calculate the temporal correlation matrix \\(\\mathbf{C}_{trans}^{head} \\in \\mathbb{R}^{L \\times L}\\) between the head chunk embedding and transition embedding. Here, the temporal correlation matrix represents the temporal variations in the gestures from the head to the transition. Similarly, we obtain the correlation matrix \\(\\mathbf{C}_{trans}^{tail}\\) from tail chunk embedding and transition embedding. Then, the global temporal dependency from head to tail is computed via matrix multiplication between \\(\\mathbf{C}_{trans}^{head}\\) and \\(\\mathbf{C}_{trans}^{tail}\\). Further, we develop a motion encoder to acquire the sequence-aware style guidance representation based on the global temporal dependency. The style guidance representation is exploited to boost the transition embeddings via an adaptive instance normalization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.448, + 0.892, + 0.478 + ], + "angle": 0, + "content": "(AdaIN) layer [18]. By doing so, we derive the transition \\(\\mathbf{f}^{tran}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.482, + 0.892, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {t r a n} = \\operatorname {A d a I N} \\left\\{\\mathbf {f} ^ {t r a n}, \\operatorname {E n c} \\left(\\mathbf {C} _ {t r a n s} ^ {h e a d} \\otimes \\mathbf {C} _ {t r a n s} ^ {t a i l}\\right) \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.892, + 0.536 + ], + "angle": 0, + "content": "where \\(\\otimes\\) indicates matrix multiplication, and Enc denotes the motion encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.688 + ], + "angle": 0, + "content": "**Emotion Mixture Strategy:** Considering obtaining the realistic 3D pose annotations corresponding to dynamically inpainted transition audio is quite difficult, we design an emotion mixture strategy to provide weak supervision of the generated gestures. Our key insight is built upon the fact that different emotions in the head/tail would lead to different gesture motions, thereby the emotion represented by transition gestures would be a mixture of the head and tail ones. As shown in Figure 2 (right), we utilize two learnable parameters as soft emotion labels of the transition gestures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Specifically, through computing the correlation matrix between the embeddings of head gestures and tail gestures, we obtain the motion deformation representation \\( S \\) from one emotion to another (e.g., neutral-to-anger). This motion deformation is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.766, + 0.892, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {\\text {h e a d}} ^ {\\text {t a i l}} = \\operatorname {S o f t m a x} \\left(\\mathbf {f} ^ {\\text {h e a d}} \\otimes \\mathbf {f} ^ {\\text {t a i l} ^ {\\prime}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.894, + 0.854 + ], + "angle": 0, + "content": "where \\( {}^{\\prime } \\) indicates the transpose operation. With the help of the motion deformation representation, we obtain the two learnable emotion embeddings \\( \\left( {{\\mathbf{e}}^{\\text{head }},{\\mathbf{e}}^{\\text{tail }}}\\right) \\) from the head and tail,respectively,as" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.858, + 0.854, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} ^ {h e a d} = \\operatorname {M a x P o o l} \\left\\{\\mathbf {W} _ {\\text {l i n e a r}} (\\mathbf {f} ^ {h e a d}) \\odot \\mathbf {S} _ {\\text {h e a d}} ^ {t a i l} \\right\\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.879, + 0.891, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} ^ {t a i l} = \\operatorname {M a x P o o l} \\left\\{\\mathbf {W} _ {t a i l} \\left(\\mathbf {f} ^ {h e a d}\\right) \\odot \\mathbf {S} _ {h e a d} ^ {t a i l} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10427" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{linear}\\) denotes linear matrix, and \\(\\odot\\) means dot product. Maxpool indicates the AdaptiveMaxPooling operation. \\(\\mathbf{e}^{tail}\\) is calculated in a similar way. Then, we obtain the two learnable parameters represented by emotion weights of head and tail gestures, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.181, + 0.469, + 0.199 + ], + "angle": 0, + "content": "\\[\n(\\sigma , 1 - \\sigma) = \\operatorname {S o f t m a x} \\left(\\mathbf {e} ^ {\\text {h e a d}}, \\mathbf {e} ^ {\\text {t a i l}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.469, + 0.304 + ], + "angle": 0, + "content": "Once we acquire the learnable emotion weights, we leverage a pre-trained pose-based emotion classifier to provide weak supervision. In this fashion, the authority of generated transition gestures is well-preserved. The training details of the pre-trained pose-based emotion classifier are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.31, + 0.469, + 0.597 + ], + "angle": 0, + "content": "Keyframe Posture Sampling: Conventionally, researchers [32, 50, 53] directly leverage the padded initial poses as conditional seeds to guide the co-speech gesture generation. In long-sequence modeling, an intuitive manner is to extend the initial pose length \\(M\\) for adapting the overall synthesized gestures' length. However, this would lead to the poor generalizability of the network (i.e., the performance gains degradation from the reduction of initial poses). Therefore, we propose a simple yet effective VAE-based [43] keyframe sampler to provide high-fidelity posture prior conditions while enabling diverse results. Evenly, we split the annotated sequence into several chunks keeping the same length with the transition of length \\(L\\). The sampler is trained with the keyframe randomly selected to reconstruct the corresponding chunk. In the inference phase, the keyframe sampler samples diverse chunks, blending as the initial postures for gesture generation. Since the transition sequence lacks the pose annotation, we randomly select the frame from head or tail sequences for posture sampling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.469, + 0.719 + ], + "angle": 0, + "content": "To further enhance the sequence-aware correspondence of the generated co-speech gestures, we leverage the diverse initial postures as the query \\(Q\\) to match the key features \\(K\\) and value features \\(V\\) in the transformer-based backbone [44]. Similar to [37], we adopt a motion discriminator to ensure the temporal smoothness of the generated results. For more details about network architecture please refer to supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.267, + 0.755 + ], + "angle": 0, + "content": "3.4. Objective Functions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.469, + 0.815 + ], + "angle": 0, + "content": "Reconstruction Loss: We leverage the ground truth 3D pose annotation of the head and tail to constrain the generated co-speech gestures as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.828, + 0.469, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left\\| P _ {\\{h e a d, t a i l \\}} - \\hat {P} _ {\\{h e a d, t a i l \\}} \\right\\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\hat{P}_{\\{\\text{head}, \\text{tail}\\}}\\) denotes generated gestures of head and tail speeches." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "Adversarial Learning Loss: To ensure the realism of the generated gestures, we further exploit the adversarial training loss, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.144, + 0.892, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a d v} = \\mathbb {E} _ {P} [ \\log \\mathcal {D} (P) ] \\\\ + \\mathbb {E} _ {A} [ \\log (1 - (\\mathcal {G} (A, \\{p _ {1}, \\dots , p _ {M} \\})) ], \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.187, + 0.892, + 0.218 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) denotes the motion discriminator and \\(\\mathcal{G}\\) means gesture generator." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.224, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Weakly Supervision Loss: We leverage the pre-trained pose-based emotion classifier to provide weak supervision of the transition gestures upon the learnable emotion weights:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.291, + 0.892, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e m o t i o n}} = - y \\log \\mathcal {F} (\\hat {P} _ {\\text {t r a n s}}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.346 + ], + "angle": 0, + "content": "where \\(y\\) is the learnable emotion label, \\(\\mathcal{F}\\) is the emotion classifier, \\(\\hat{P}_{trans}\\) is the generated transition gestures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.347, + 0.71, + 0.361 + ], + "angle": 0, + "content": "Finally, the overall objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.369, + 0.892, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\min _ {G} \\max _ {D} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {r} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\mathcal {L} _ {\\text {e m o t i o n}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.398, + 0.773, + 0.412 + ], + "angle": 0, + "content": "The \\(\\lambda_r\\), and \\(\\lambda_{adv}\\) are weight coefficients." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.424, + 0.632, + 0.44 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.449, + 0.803, + 0.466 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.774 + ], + "angle": 0, + "content": "BEAT Emotion Transition Dataset (BEAT-ETrans): Since there are only single emotion labels of aligned speech-gesture corpus in the original BEAT dataset [28], to satisfy our insight on emotion transition co-speech gesture generation modeling, we recollect a BEAT Emotion Transition Dataset (dubbed BEAT-ETrans). In particular, we resample the motion FPS as 15 and intercept the continuous 60 frames with stride 30 as the head/tail clips. Here, the head clips are all annotated as neutral, and tail clips are denoted with the other seven emotions: anger, happiness, fear, disgust, sadness, contempt, and surprise. As for one head speech, we randomly select two or three tails with different emotions to construct the head-tail pairs. By leveraging the two-second (i.e., corresponding 30-frame postures) transition to blend the heads and tails, we obtain the 10-second human speech clips in our BEAT-ETrans. We obtain 58,077 clips, including a total of 161.3 hours reported in Table 1. Then the clip numbers of training/validation/testing sets are randomly split as 41,908/4,077/12,092. In all of our experiments, we utilize the upper body with 71 joints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "TED Emotion Transition Dataset (TED-ETrans): Inspired by [32, 36], we further newly collect a TED Emotion Transition Dataset (dubbed TED-ETrans) based on more than 1.7K speakers from in-the-wild TED talk show videos, demonstrated in Table 1. Due to the lack of emotional labels, we first leverage the annotated BEAT dataset to pretrain an audio-based emotion classifier for labeling TED audi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10428" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Statistics comparison of existing 3D co-speech gesture datasets with ours. Our BEAT-ETrans and TED-ETrans are built upon the existing BEAT [28] and TED-Expressive [32], respectively. To the best of our knowledge, we are the first to present two large datasets with emotion transition human speech." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.143, + 0.891, + 0.294 + ], + "angle": 0, + "content": "
DatasetJoint AnnotationModalityDuration (hours)
BodyHandAudioTextSpeakersSingle EmotionEmotion Transition
TED [50]pseudo label9✘✓✓1,766✘✘106.1
SCG [15]pseudo label1424✓✘6✘✘33
Trinity [10]mo-cap2438✓✓1✘✘4
ZeroEGGS [12]mo-cap2748✓✓1✘✘2
BEAT [28]mo-cap2748✓✓30✓✘35
TED-Expressive [32]pseudo label1330✓✓1,764✘✘100.8
BEAT-ETrans (ours)mo-cap2748✓✓308✓161.3
TED-ETrans (ours)pseudo label1330✓✓1,7646✓59.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Table 2. Comparison with the start-of-the-art methods on our newly collected BEAT-ETrans and TED-ETrans datasets. \\( \\uparrow \\) denotes the higher the better, and \\( \\downarrow \\) indicates the lower the better. \\( \\pm \\) means \\( {95}\\% \\) confidence interval." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.345, + 0.885, + 0.487 + ], + "angle": 0, + "content": "
ModelsBEAT-ETransTED-ETrans
FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑
Seq2Seq [49]ICRA'1940.9547.930.14196.66±2.1629.6049.470.26572.81±1.99
S2G [13]CVPR'1925.5637.040.67198.26±2.0418.1641.630.82476.82±2.32
Trimodal [50]TOG'2014.0942.500.764100.87±2.1221.0633.200.75882.87±1.86
CAMN [28]ECCV'229.0327.530.794118.46±2.3319.2841.040.78579.03±1.49
HA2G [32]CVPR'227.2825.790.779121.77±2.3116.7240.380.78780.14±1.65
DiffGesture [53]CVPR'236.6825.030.788122.29±2.0118.6925.130.81892.01±2.07
Ours4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.471, + 0.633 + ], + "angle": 0, + "content": "classification threshold as \\(\\geq 0.95\\), and the two uncommon emotions (i.e. fear, disgust) are dropped. Then we maintain the same data pre-processing strategy with our BEAT-ETrans to obtain a total of 21,515 clips with 59.8 hours. The final clip division criteria of the TED-ETrans dataset are training/validation/testing with 15,061/2,152/4,302 respectively. In practice, the 43 upper body joints are leveraged in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Implementation Details: We set the total generated cospeech gesture length as \\( N = 150 \\), and the transition and chunk lengths are \\( L = 30 \\). Conventionally, we leverage \\( M = 4 \\) frames as the reference initial poses. The feature dimension \\( D = 512 \\) in practice. The raw audio of human speech is converted to mel-spectrograms with FFT window size 1024, and hop length 512. The audio encoder takes the ResNetSE34 [6] as the backbone. Empirically, we set \\( \\lambda_r = 20 \\), and \\( \\lambda_{adv} = 2 \\). Our models are implemented on the Pytorch platform with a single NVIDIA Tesla V100 GPU. The initial learning rate is set to 0.0003 by utilizing Adam Optimizer. The whole training takes 100 epochs with a batch size of 96." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Evaluation Metrics: To fully evaluate the realism and diversity of the generated co-speech gestures, we introduce various metrics:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.471, + 0.901 + ], + "angle": 0, + "content": "- FGD: Fréchet Gesture Distance (FGD) [50] is utilized to" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.511, + 0.892, + 0.661 + ], + "angle": 0, + "content": "measure the distribution distance between the realistic sequential gestures and generated ones. Since we only have the 3D joint annotation of the head/tail, we take the network architecture provided in [32, 50] to train the autoencoder for distance computing on the two datasets, respectively. The FGD of transition gestures is calculated as the average value between the distribution distance of transition and head/tail, indicated as \\(\\mathrm{FGD}_{\\text {trans }}\\). Similarly, \\(\\mathrm{FGD}_{h + t}\\) means the distance between the generated head/tail gestures and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.662, + 0.891, + 0.706 + ], + "angle": 0, + "content": "- BC: Beat Consistency Score (BC) [28, 32] measures the speech audio alignment degree with the generated co-speech gestures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.708, + 0.891, + 0.782 + ], + "angle": 0, + "content": "- Diversity: Similar to [32, 53], we exploit the same feature extractor in FGD to obtain the feature embeddings of the generated gestures. The diversity reflects the average distance between 500 random combination pairs in the testing set of 60 speech audios." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.662, + 0.891, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.79, + 0.722, + 0.807 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Comparisons with SOTA Methods: To the best of our knowledge, we are the first to explore the co-speech gesture generation with emotion transition human audios. To fully verify the superiority of our method, we implement various state-of-the-art single-emotion-based counterparts: Seq2Seq [49], S2G [13], Trimodal [50], CAMN [28]," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10429" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 3. Ablation study on different components of the proposed method. \\(\\checkmark\\) indicates the employment of a certain module. \\(\\uparrow\\) denotes the higher the better, and \\(\\downarrow\\) indicates the lower the better. \\(\\pm\\) means \\(95\\%\\) confidence interval. MTIM: Motion Transition Infusion Mechanism; EMS: Emotion Mixture Strategy; FKS: Keyframe Sampler." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.143, + 0.885, + 0.242 + ], + "angle": 0, + "content": "
Model VariationsBEAT-ETransTED-ETrans
BaselineMTIMEMSKFSFGDh+t↓FGDtrans↓BC↑Diversity ↑FGDh+t↓FGDtrans↓BC↑Diversity ↑
✓21.2156.330.70188.89±1.5826.4447.870.71274.19±2.75
✓✓10.7527.940.866108.69±2.6320.9536.380.82782.42±2.24
✓✓✓5.6921.700.878112.80±1.9514.0226.210.90087.21±2.12
✓✓✓✓4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.356 + ], + "angle": 0, + "content": "HA2G [32], and DiffGesture [53]. For a fair comparison, all the models are implemented by the source code released by the authors. Note that since we only have the ground truth of head and tail gestures, several recent VAE-based works [1, 2, 48, 52] cannot be directly applied in the experiments (or they have not released codes so far)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.471, + 0.631 + ], + "angle": 0, + "content": "As reported in Table 2, we adopt the \\(\\mathrm{FGD}_{h + t}\\), \\(\\mathrm{FGD}_{trans}\\), BC, and Diversity for a well-rounded view of comparisons. Our method outperforms all the competitors by a large margin on both two datasets. Remarkably, on the TED-ETrans dataset, our method even achieves \\(34.8\\%\\) (i.e., \\((18.69 - 12.19) / 18.69 \\approx 34.8\\%\\) ) improvement over the sub-optimal counterparts in \\(\\mathrm{FGD}_{h + t}\\). We observe both the DiffGesture [53] and ours synthesize the high-fidelity gestures of head/tail speech with much lower \\(\\mathrm{FGD}_{h + t}\\) than others. However, the DiffGesture shows worse performance on \\(\\mathrm{FGD}_{trans}\\) due to lack of supervision. In terms of diversity, our simple yet effective keyframe sampler provides authority and diverse initial postures as the reference, thus enabling us to demonstrate diverse gesture styles compared to other counterparts. Moreover, we find that the diversity scores on BEAT-ETrans are much higher than those on TED-ETrans dataset. This can be attributed to the more complex human joints in the BEAT-ETrans dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Ablation Study: To further verify the effectiveness of our proposed methods, we conduct the ablation study of different components as variations, reported in Table 3. The baseline model is implemented by a simple transformer-based pipeline with stacking three times blocks in the encoder-decoder. Obviously, all the combinations of our proposed components have positive impacts on the generated results. Specifically, by adding the motion transition infusion mechanism to the baseline, the indicator BC has achieved significant improvement (e.g., \\(0.701 \\rightarrow 0.866\\) in the BEAT-Etrans). This result verifies that our motion transition infusion mechanism effectively models the temporal correlation between the transitions w.r.t. head/tail gestures, thus leading to the generated results preserving rhythm coherency with given speech, globally." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Moreover, adopting the emotion mixture strategy ideally improves the performance of \\(\\mathrm{FGD}_{trans}\\) on both two" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.417 + ], + "angle": 0, + "content": "datasets. This indicates that the learnable emotion mixture wights can provide effective weak supervision by leveraging the pre-trained pose-based emotion classifier. Besides, we have observed significant improvement in the performance of \\(\\mathrm{FGD}_{h + t}\\) during this phase compared to the previous version. This aligns with our transformer backbone's emphasis on modeling the sequential temporal correlations as a whole. The better transition gestures encourage our model to maintain better temporal consistency, thereby the head/tail gestures achieve better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Finally, after additionally employing the keyframe sampler to produce authority initial postures as the reference, our method obtains the best performance. Although the properties of BC and FGD on both datasets just have slightly improvement, the diversity realizes a noticeably better achievement (e.g., \\(112.80 \\rightarrow 124.93\\) in the BEAT-ETrans dataset). This highly supports our insight into keyframe-driven diversification strategy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.712, + 0.574 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Visualization: To fully demonstrate the performance of our method, we show the visualized keyframes generated from ours compared with counterparts on our newly collected TED-ETrans and BEAT-ETrans datasets, respectively. As depicted in Figure 3, our method displays vivid and diverse results against others. In particular, we observe that Seq2Seq and Trimodal tend to synthesize unreasonable and stiff results (e.g., the blue rectangle in the right BEAT-ETrans dataset). Although CAMN and HA2G can generate natural upper-body postures, we find that they sometimes produce unreliable subtle fingers (e.g., the red rectangle in the left TED-ETrans dataset). Both our method and DiffGesture create reasonable gestures. However, the results synthesized by DiffuGesture are mismatched with the emotion transitions. In contrast, our method can synthesize the synchronous motions (e.g., in the BEAT-ETrans, the arms become droopy as emotion turns to sadness). Meanwhile, we further verify the diversification results as shown in Figure 1. Given the same input audio, our method generates diverse and vivid co-speech gestures. Please refer to the supplementary material for more visualization results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10430" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.088, + 0.85, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.895, + 0.491 + ], + "angle": 0, + "content": "Figure 3. Visualization of our generated 3D co-speech gestures against various state-of-the-art methods. The samples of the left part are from our newly collected TED-ETrans dataset, and the samples of the right part are from our BEAT-ETrans dataset. Best view on screen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.752 + ], + "angle": 0, + "content": "User Study: To further analyze the quality of results by various counterparts and ours, we conduct a user study by visualizing the results and inviting 15 subjects. For each emotion in both datasets, we randomly select 12 samples for each participant (7 emotions in BEAT-ETrans, 5 emotions in TED-ETrans). All the participants are recruited anonymously from schools including various majors. In particular, the subjects are required to rate the generated co-speech gestures from 0 to 5 (the higher, the better) in terms of naturalness, motion smoothness, and speech-gesture coherency. The results are demonstrated in Figure 4. Our method showcases the best performance compared with all the competitors. Especially in terms of motion smoothness, our method achieves noticeable advantages, indicating the effectiveness of our proposed motion transition infusion mechanism and the emotion mixture strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.196, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a new task of 3D co-speech gesture generation given emotion transition human speech. We therefore newly collected two datasets named the BEAT-ETrans and the TED-ETrans to fulfill this goal while significantly facilitating the research on 3D human motion modeling. Then, we fully take advantage of the sequential temporal correlation via a motion transition infusion mechanism" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.51, + 0.876, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Figure 4. User study on gesture naturalness, motion smoothness, and speech-gesture coherency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.892, + 0.846 + ], + "angle": 0, + "content": "to ensure the generated gestures preserve temporal coherence. Furthermore, we design an emotion mixture strategy to supply emotional weak supervision of the synthesized transitions. Extensive experiments conducted on our two newly collected datasets show the superiority of the method. As our method intends to generate diverse and vivid emotion transition gestures, we will investigate diversifying the 3D gesture with temporal smooth sampling, instead of the keyframe-wise manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Acknowledgements. This work is funded in part by the Theme-based Research Scheme grant (No.T45-205/21-N) and the InnoHK funding, Hong Kong SAR." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10431" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2027-2036, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Joon Son Chung, Jaesung Huh, Seongkyu Mun, Minjae Lee, Hee Soo Heo, Soyeon Choe, Chiheon Ham, Sunghwan Jung, Bong-Jin Lee, and Icksang Han. In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. arXiv preprint arXiv:2005.07143, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. arXiv preprint arXiv:2305.08953, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Maged Farouk. Studying human robot interaction and its characteristics. International Journal of Computations, Information and Manufacturing (IJCIM), 2(1), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[10] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[11] Yu Fu, Yan Hu, and Veronica Sundstedt. A systematic literature review of virtual, augmented, and mixed reality game applications in healthcare. ACM Transactions on Computing for Healthcare (HEALTH), 3(2):1-27, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[12] Saeed Ghorbani, Ylva Fe2rstl, Daniel Holden, Nikolaus F Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. In Computer Graphics Forum, pages 206–216. Wiley Online Library, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Susan Goldin-Meadow and Martha Wagner Alibali. Gesture's role in speaking, learning, and creating language. Annual review of psychology, 64:257-283, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.264 + ], + "angle": 0, + "content": "[16] Autumn B Hostetter and Martha W Alibali. Visible embodiment: Gestures as simulated action. Psychonomic bulletin & review, 15:495-514, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[17] Ann Huang, Pascal Knierim, Francesco Chiossi, Lewis L Chuang, and Robin Welsch. Proxemics for human-agent interaction in augmented reality. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.338, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.397, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[21] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thórisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.598, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[22] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities for reactive robotic response. In IROS, page 2071. Tokyo, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.642, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[23] Sergey Levine, Philipp Krahenbuhl, Sebastian Thrun, and Vladlen Koltun. Gesture controllers. In Acm siggraph 2010 papers, pages 1-11. 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[24] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10432" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7590–7598, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.244 + ], + "angle": 0, + "content": "[28] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[29] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. AudioLDM: Text-to-audio generation with latent diffusion models. Proceedings of the International Conference on Machine Learning, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[30] Haohe Liu, Qiao Tian, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D. Plumbley. AudioLDM 2: Learning holistic audio generation with self-supervised pretraining. arXiv preprint arXiv:2308.05734, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via masked audio gesture modeling. arXiv preprint arXiv:2401.00374, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.37, + 0.621 + ], + "angle": 0, + "content": "[34] OpenAI. Gpt-4 technical report, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[36] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[37] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3d hand gesture prediction from body dynamics by bilateral hand disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4616-4626, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, et al. Speechbrain: A general-purpose speech toolkit. arXiv preprint arXiv:2106.04624, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[40] Peter Roach. Some languages are spoken more quickly than others. 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[41] Mehmet E Sargin, Yucel Yemez, Engin Erzin, and Ahmet M Tekalp. Analysis of head gesture and prosody patterns for prosody-driven head-gesture animation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(8):1330-1345, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[42] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[43] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[45] Jordan Voas, Yili Wang, Qixing Huang, and Raymond Mooney. What is the best automated metric for text to motion generation? arXiv preprint arXiv:2309.10248, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[46] Zijian Wang, Xingqun Qi, Kun Yuan, and Muyi Sun. Self-supervised correlation mining network for person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7703-7712, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 5860–5868, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In International Cconference on Multimodal Interaction, pages 179-185. 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10433" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[53] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10434" + } + ] +] \ No newline at end of file diff --git a/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_origin.pdf b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db7c981fed380b5a55a15e11aa2bff4b0d641245 --- /dev/null +++ b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/82233940-816a-43f0-81c1-3331488ee413_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ee9d4310f85efc35ebdf98e9e4ee2ea30e4372f3bc6576f398238da14deddd3 +size 1863670 diff --git a/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/full.md b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..248d97fb301e3c061761e9c6300004aa2e4ef197 --- /dev/null +++ b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/full.md @@ -0,0 +1,291 @@ +# Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation + +Xingqun Qi $^{1}$ , Jiahao Pan $^{1}$ , Peng Li $^{1}$ , Ruibin Yuan $^{1}$ , Xiaowei Chi $^{1}$ , Mengfei Li $^{1}$ , Wenhan Luo $^{1}$ , Wei Xue $^{1}$ , Shanghang Zhang $^{2}$ , Qifeng Liu $^{1,\boxtimes}$ , Yike Guo $^{1,\boxtimes}$ , $^{1}$ The Hong Kong University of Science and Technology $^{2}$ Peking University + +xingqun.qi@connect.ust.hk, {liuqifeng, yikeguo}@ust.hk + +![](images/de4fc90de6e2baa5b3d0bc2d0986790c72f3606ad31cdf70f1f53f2540899ba6.jpg) +Figure 1. Diverse exemplary clips sampled by our method from our newly collected BEAT Emotion Transition Dataset. The vital frames are visualized to demonstrate that the upper body gestures change with the emotion transition of human speech, synchronously. From top to bottom: the input speech audio, the corresponding transcript, and two sampled clips. Best view on screen. + +# Abstract + +Generating vivid and emotional 3D co-speech gestures is crucial for virtual avatar animation in human-machine interaction applications. While the existing methods enable generating the gestures to follow a single emotion label, they overlook that long gesture sequence modeling with emotion transition is more practical in real scenes. In addition, the lack of large-scale available datasets with emotional transition speech and corresponding 3D human gestures also limits the addressing of this task. To fulfill this goal, we first incorporate the ChatGPT-4 and an audio inpainting approach to construct the high-fidelity emotion transition human speeches. Considering obtaining the realistic 3D pose annotations corresponding to the dynamically inpainted emotion transition audio is extremely difficult, we propose a novel weakly supervised training strategy to encourage authority gesture transitions. Specifically, to enhance the coordination of transition gestures w.r.t. dif + +ferent emotional ones, we model the temporal association representation between two different emotional gesture sequences as style guidance and infuse it into the transition generation. We further devise an emotion mixture mechanism that provides weak supervision based on a learnable mixed emotion label for transition gestures. Last, we present a keyframe sampler to supply effective initial posture cues in long sequences, enabling us to generate diverse gestures. Extensive experiments demonstrate that our method outperforms the state-of-the-art models constructed by adapting single emotion-conditioned counterparts on our newly defined emotion transition task and datasets. Our code and dataset will be released on the project page: https://xingqunqi-lab.github.io/Emo-Transition-Gesture/. + +# 1. Introduction + +Co-speech gesture generation aims to synthesize vivid and emotional human postures coordinated with the audio input. These non-verbal behaviors serve as a key factor dur- + +ing human conversations that significantly facilitates the delivery of speech content [5, 14, 16]. Meanwhile, modeling co-speech gestures has a wide range of embodied AI applications in human-machine interaction [22, 27], robot assistants [9], and virtual/augmented reality (AR/VR) [11]. Conventionally, many researchers usually focus on synthesizing human upper body gestures consistent with speech audio [32, 47, 48, 50, 53]. + +Nevertheless, except for a few recent works that generate the co-speech gestures of a single emotion category [2, 3, 28, 52], previous works mostly focus on emotion-agnostic generation [4, 48, 50]. Most of them overlook synthesizing the long sequence co-speech gestures with the emotion transitions, which are more practical in real-world scenes. For example, a person may not maintain a single emotion forever when communicating with others or in speech talking. In this work, we therefore introduce the task of speech-driven emotion transition for generating vivid and diverse 3D co-speech gestures, displayed in Figure 1. There are two main challenges in this task: 1) Datasets of 3D co-speech gestures synchronized with emotion transition speech audios are scarce. Creating such a dataset containing 3D human postures is difficult due to the lack of guidance from emotional experts and complex motion capture systems. 2) Modeling the plausible and temporal coherent co-speech gestures from one emotion to another in long sequences is difficult, especially in the transition duration. + +To overcome the issue of data scarcity, we newly present two datasets containing emotion-transition human speech that are built upon previous single-emotion ones [28, 32]. In particular, thanks to the developed language model ChatGPT-4 [34], we first leverage it to generate text transcripts of the transition based on speech context. Then, by employing the audioLDM2 technique [29, 30] for audio inpainting, we ensure the inpainted transition's timbral consistency with its adjacent contexts and a smooth emotional transition throughout. Afterwards, to support our insight on modeling co-speech gestures coherent with emotion-transition speech, corresponding 3D human postures of transition are required. + +However, due to the dynamically generated transition transcripts, it is infeasible to construct the aligned realistic 3D pose annotation of human bodies. Hence, we solve the challenges of vivid co-speech gesture generation in a novel weakly supervised pattern, containing a motion transition infusion mechanism and an emotion mixture strategy. Specifically, in the motion transition infusion mechanism, we model the temporal correlation between the generated head and tail gesture features as style guidance representation. The style guidance representation provides motion transition cues that are infused into the transition embedding via an adaptive instance normalization (AdaIN) layer [18]. Along with this manner, we can effectively en + +hance the coordination of transition gestures w.r.t. two different emotional ones. + +Moreover, to alleviate the lack of supervision during the transition between two emotions, the emotion mixture strategy is built to provide weak emotional supervision of the generated transition gestures. Concretely, we learn a joint embedding of two different emotional gesture sequences using a temporal aggregation encoder. Then, we pre-train an emotion classifier based on the annotated human 3D poses with single emotion labels in the dataset. Here, this joint embedding is leveraged as an emotion mixture weight for the pre-trained classifier to facilitate high-fidelity transition gesture synthesis with desirable properties. Finally, considering the generated 3D postures should be non-deterministic given the human speech, we devise a keyframe sampler to produce diverse initial poses as reference. In this fashion, our method enables diverse co-speech gesture generation with emotion transitions. Extensive experiments conducted on our newly constructed two datasets verify the effectiveness of our methods, displaying vivid and emotional 3D co-speech gestures. + +Overall, our contributions are summarized as follows: + +- We introduce a new task of emotion transition co-speech gesture generation cooperating with two newly constructed datasets named BEAT Emotion Transition (BEAT-ETrans) and TED Emotion Transition (TED-ETrans), significantly facilitating research on 3D human motion modeling. +- We design a motion transition infusion mechanism to ensure the temporal coordination of transition gestures w.r.t. two different emotional ones and a weakly supervised emotion mixture strategy to enable high-fidelity transition gesture synthesis with desirable properties. +- Extensive experiments show that our method outperforms state-of-the-art counterparts on both datasets, displaying realistic and vivid co-speech gestures given emotion transition human audiios. + +# 2. Related Work + +# 2.1. Co-speech Gesture Synthesis + +Synthesizing human co-speech gestures plays a significant role in various applications [17, 22, 37, 46]. Numerous studies have been proposed to address these issues that are roughly divided into rule-based approaches [20, 21], machine-learning-based approaches [23, 41], and deep learning-based ones [2, 28, 32, 36, 48, 50, 52, 53]. Traditional researchers follow the rule-based workflow, leveraging the speech-gesture pairs as guidance to generate co-speech gestures pre-defined by linguistic experts. Other early works integrate the manually defined gesture features with machine learning techniques to synthesize the co + +speech gestures. In the aforementioned two manners, the researchers usually focus on the optimization of the matching process between human speech and pre-defined gestures. It may require experts much expensive effort in the speech-gesture pair construction. + +Recently works focus on building the mapping directly from the input human speech and sequential gestures by exploiting the deep neural networks. They usually leverage multi-modality cues to facilitate the generation of co-speech gestures, associating with the speech audio [25, 48, 53], speaker identity [50], emotion [2, 28, 31], and text transcript [32]. However, they overlook that emotion transition of the long sequential co-speech gesture modeling is much more practical in the real scenes. Moreover, since the lack of annotated 3D gestures corresponding to dynamically constructed transition speech, few of the above methods could be directly adapted to this new thought. + +# 2.2. 3D Human Motion Modeling + +Human motion modeling aims to generate realistic and smooth human motions with various multi-modality conditions [8, 19, 33], including co-speech gesture generation as a sub-task. One of the hottest topics is synthesizing human motion from text prompts with a few past postures as the seed [35, 45, 51]. These methods usually engage in forcing human motion to represent the semantic expression aligned with the text. Literally, the task most closely related to ours is AI choreographer [24, 26, 42] which generates the motion from music signals. However, the AI choreographer works mainly focus on the rhythmic-coherent motion of the whole human body but without subtle finger gestures. While sharing a similar goal with the approaches mentioned above, our newly defined work differs from them significantly. We take the emotion-transition long sequence co-speech gesture without corresponding 3D human pose annotation into consideration, thus motivating us to utilize the motion transition of two annotated emotional gestures and coherent the overall sequence. + +# 3. Proposed Method + +# 3.1. Emotion Transition Dataset Construction + +We aim to address the emotion transition co-speech gesture generation in a weakly supervised manner. Due to the existing paired speech-gesture datasets [28, 36], we could focus on synthesizing the high-fidelity transition human audios. Synthesizing datasets conducive to our task focuses on ensuring semantic coherence, smooth emotional audio transitions, consistent timbre, and audio fidelity. + +Preliminary: Considering our key insight to modeling the co-speech gesture with emotion transition, we first split the existing aligned speech-gesture pairs [28, 36] into four-second clips. To diversify the datasets, we randomly splice + +two clips from the same speaker to construct an emotion-transition candidate pair. The clip with neutral emotion is leveraged as head speech, and the other with various emotions is represented as tail speech. We leverage the dynamically synthesized two-second audio as a transition to combine the head and tail speeches. In the following, we briefly summarize our efforts in constructing the dynamic transition speech audio. + +Textual Inpainting of Transition: To ensure the semantic coherence of transition w.r.t. head/tail speeches, we exploit the advanced language generation model ChatGPT-4 [34] to complete the transcript according to the context. Literally, we follow the conventional estimation that people usually talk 30 phonemes [40] in a two-second speech as the prompt to guide transcript generation by ChatGPT-4. + +Synthesis of Transition Audio: Once we obtain the transcript of transitions, we employ a superior text-to-speech model, audioLDM2 [29, 30], to generate corresponding speech audio. Here, we leverage the speaker embeddings extracted via SpeechBrain's ECAPA-TDNN [7, 39] as prior guidance to maintain the identity consistency of the generated transition audio. Then, we adopt Whisper [38] to restrict the word error rate, thereby ensuring the accuracy and clarity of the speech content. In this fashion, the synthesized transition audio realizes controlled duration while the natural smooth tonality is well preserved. The datasets will be released in the future. More details are provided in the Section 4.1. + +# 3.2. Problem Formulation + +Given a sequence of audio signal $A = \{a_{1},\dots,a_{N}\}$ as input to model $\mathcal{G}$ , our goal is to generate vivid and emotional 3D co-speech gestures $P = \{p_1,\dots,p_N\}$ of the upper human body. $N$ represents the number of synthesized postures corresponding to the audio $A$ . Each $p_i$ is denoted as $J$ joints with 3D representation. In particular, we define the audio signal as consisting of a head speech, a dynamically inpainted transition speech, and a tail speech. Here, the head and tail speeches are randomly selected from previous co-speech gesture datasets [28, 32] with a single emotion label, respectively. Following the conventions of previous works [28, 32, 50, 53], we invoke $M$ frame poses as the initial seed to guide generation. The overall objective is expressed as + +$$ +\underset {\mathcal {G}} {\arg \min } \| P - \mathcal {G} (A, \{p _ {1}, \dots , p _ {M} \}) \|. \tag {1} +$$ + +Note that only the generated gestures with head and tail speeches are supervised with ground truth coming from existing datasets [28, 32]. The transition gestures with $L$ frames, where $L \ll N$ , will be weakly supervised through the following processes within a motion transition infusion + +![](images/662578f6ec4cc05fcf2d0fec7f6f331fd96aa0dfcb47d4764f04e953c20b9580.jpg) +Figure 2. The overview of our proposed method. The middle part (blue) displays the overall pipeline for 3D co-speech gesture generation from emotion transition human speech. The left part (green) depicts our proposed Motion Transition Infusion Mechanism that enhances the coordination of transition gestures w.r.t. head/tail ones. The right part (orange) shows the designed Emotion Mixture Strategy to provide weak supervision of the generated transition gestures, thereby achieving authority producing. + +![](images/cb9764782a5de94261561685a9bd967e28e50cc24601da8b5754068c517bf27b.jpg) + +![](images/749191ea0de3bdfb80e452fc3fb24446d9ab2f5055bc03d0fe16a35985c5b9a6.jpg) + +mechanism and an emotion mixture strategy. The audio signal is fed into an audio encoder for feature extraction. Our overall workflow is shown in Figure 2. + +# 3.3. Weakly-supervised Emotion Transition + +Motion Transition Infusion Mechanism: To ensure the temporal coherence of the transition gestures w.r.t. head/tail ones, we propose a motion transition infusion mechanism to model the temporal association representation between different emotional gestures. As depicted in Figure 2 (left), the total sequential features $\mathbf{f}^{total}$ consist of the features of head $\mathbf{f}^{head}$ , transition $\mathbf{f}^{tran}$ , and tail $\mathbf{f}^{tail}$ . Inspired by [33], we nominate a head chunk $\mathbf{f}_{chunk}^{head}$ composed of the last $L$ frames of head gesture embeddings and a tail chunk $\mathbf{f}_{chunk}^{tail}$ consisting of the first $L$ ones of the tail gesture embeddings. Here, the dimension of each frame representing gesture embedding is $\mathbb{R}^{1\times D}$ . + +In particular, we first calculate the temporal correlation matrix $\mathbf{C}_{trans}^{head} \in \mathbb{R}^{L \times L}$ between the head chunk embedding and transition embedding. Here, the temporal correlation matrix represents the temporal variations in the gestures from the head to the transition. Similarly, we obtain the correlation matrix $\mathbf{C}_{trans}^{tail}$ from tail chunk embedding and transition embedding. Then, the global temporal dependency from head to tail is computed via matrix multiplication between $\mathbf{C}_{trans}^{head}$ and $\mathbf{C}_{trans}^{tail}$ . Further, we develop a motion encoder to acquire the sequence-aware style guidance representation based on the global temporal dependency. The style guidance representation is exploited to boost the transition embeddings via an adaptive instance normalization + +(AdaIN) layer [18]. By doing so, we derive the transition $\mathbf{f}^{tran}$ as + +$$ +\mathbf {f} ^ {t r a n} = \operatorname {A d a I N} \left\{\mathbf {f} ^ {t r a n}, \operatorname {E n c} \left(\mathbf {C} _ {t r a n s} ^ {h e a d} \otimes \mathbf {C} _ {t r a n s} ^ {t a i l}\right) \right\}, \tag {2} +$$ + +where $\otimes$ indicates matrix multiplication, and Enc denotes the motion encoder. + +**Emotion Mixture Strategy:** Considering obtaining the realistic 3D pose annotations corresponding to dynamically inpainted transition audio is quite difficult, we design an emotion mixture strategy to provide weak supervision of the generated gestures. Our key insight is built upon the fact that different emotions in the head/tail would lead to different gesture motions, thereby the emotion represented by transition gestures would be a mixture of the head and tail ones. As shown in Figure 2 (right), we utilize two learnable parameters as soft emotion labels of the transition gestures. + +Specifically, through computing the correlation matrix between the embeddings of head gestures and tail gestures, we obtain the motion deformation representation $S$ from one emotion to another (e.g., neutral-to-anger). This motion deformation is formulated as: + +$$ +\mathbf {S} _ {\text {h e a d}} ^ {\text {t a i l}} = \operatorname {S o f t m a x} \left(\mathbf {f} ^ {\text {h e a d}} \otimes \mathbf {f} ^ {\text {t a i l} ^ {\prime}}\right), \tag {3} +$$ + +where ${}^{\prime }$ indicates the transpose operation. With the help of the motion deformation representation, we obtain the two learnable emotion embeddings $\left( {{\mathbf{e}}^{\text{head }},{\mathbf{e}}^{\text{tail }}}\right)$ from the head and tail,respectively,as + +$$ +\mathbf {e} ^ {h e a d} = \operatorname {M a x P o o l} \left\{\mathbf {W} _ {\text {l i n e a r}} (\mathbf {f} ^ {h e a d}) \odot \mathbf {S} _ {\text {h e a d}} ^ {t a i l} \right\}, +$$ + +$$ +\mathbf {e} ^ {t a i l} = \operatorname {M a x P o o l} \left\{\mathbf {W} _ {t a i l} \left(\mathbf {f} ^ {h e a d}\right) \odot \mathbf {S} _ {h e a d} ^ {t a i l} \right\}, \tag {4} +$$ + +where $\mathbf{W}_{linear}$ denotes linear matrix, and $\odot$ means dot product. Maxpool indicates the AdaptiveMaxPooling operation. $\mathbf{e}^{tail}$ is calculated in a similar way. Then, we obtain the two learnable parameters represented by emotion weights of head and tail gestures, as follows + +$$ +(\sigma , 1 - \sigma) = \operatorname {S o f t m a x} \left(\mathbf {e} ^ {\text {h e a d}}, \mathbf {e} ^ {\text {t a i l}}\right). \tag {5} +$$ + +Once we acquire the learnable emotion weights, we leverage a pre-trained pose-based emotion classifier to provide weak supervision. In this fashion, the authority of generated transition gestures is well-preserved. The training details of the pre-trained pose-based emotion classifier are provided in the supplementary material. + +Keyframe Posture Sampling: Conventionally, researchers [32, 50, 53] directly leverage the padded initial poses as conditional seeds to guide the co-speech gesture generation. In long-sequence modeling, an intuitive manner is to extend the initial pose length $M$ for adapting the overall synthesized gestures' length. However, this would lead to the poor generalizability of the network (i.e., the performance gains degradation from the reduction of initial poses). Therefore, we propose a simple yet effective VAE-based [43] keyframe sampler to provide high-fidelity posture prior conditions while enabling diverse results. Evenly, we split the annotated sequence into several chunks keeping the same length with the transition of length $L$ . The sampler is trained with the keyframe randomly selected to reconstruct the corresponding chunk. In the inference phase, the keyframe sampler samples diverse chunks, blending as the initial postures for gesture generation. Since the transition sequence lacks the pose annotation, we randomly select the frame from head or tail sequences for posture sampling. + +To further enhance the sequence-aware correspondence of the generated co-speech gestures, we leverage the diverse initial postures as the query $Q$ to match the key features $K$ and value features $V$ in the transformer-based backbone [44]. Similar to [37], we adopt a motion discriminator to ensure the temporal smoothness of the generated results. For more details about network architecture please refer to supplementary material. + +# 3.4. Objective Functions + +Reconstruction Loss: We leverage the ground truth 3D pose annotation of the head and tail to constrain the generated co-speech gestures as: + +$$ +\mathcal {L} _ {r e c} = \left\| P _ {\{h e a d, t a i l \}} - \hat {P} _ {\{h e a d, t a i l \}} \right\| _ {1}, \tag {6} +$$ + +where $\hat{P}_{\{\text{head}, \text{tail}\}}$ denotes generated gestures of head and tail speeches. + +Adversarial Learning Loss: To ensure the realism of the generated gestures, we further exploit the adversarial training loss, expressed as: + +$$ +\begin{array}{l} \mathcal {L} _ {a d v} = \mathbb {E} _ {P} [ \log \mathcal {D} (P) ] \\ + \mathbb {E} _ {A} [ \log (1 - (\mathcal {G} (A, \{p _ {1}, \dots , p _ {M} \})) ], \tag {7} \\ \end{array} +$$ + +where $\mathcal{D}$ denotes the motion discriminator and $\mathcal{G}$ means gesture generator. + +Weakly Supervision Loss: We leverage the pre-trained pose-based emotion classifier to provide weak supervision of the transition gestures upon the learnable emotion weights: + +$$ +\mathcal {L} _ {\text {e m o t i o n}} = - y \log \mathcal {F} (\hat {P} _ {\text {t r a n s}}), \tag {8} +$$ + +where $y$ is the learnable emotion label, $\mathcal{F}$ is the emotion classifier, $\hat{P}_{trans}$ is the generated transition gestures. + +Finally, the overall objective is: + +$$ +\min _ {G} \max _ {D} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {r} \mathcal {L} _ {\text {r e c}} + \lambda_ {\text {a d v}} \mathcal {L} _ {\text {a d v}} + \mathcal {L} _ {\text {e m o t i o n}}. \tag {9} +$$ + +The $\lambda_r$ , and $\lambda_{adv}$ are weight coefficients. + +# 4. Experiments + +# 4.1. Datasets and Experimental Setting + +BEAT Emotion Transition Dataset (BEAT-ETrans): Since there are only single emotion labels of aligned speech-gesture corpus in the original BEAT dataset [28], to satisfy our insight on emotion transition co-speech gesture generation modeling, we recollect a BEAT Emotion Transition Dataset (dubbed BEAT-ETrans). In particular, we resample the motion FPS as 15 and intercept the continuous 60 frames with stride 30 as the head/tail clips. Here, the head clips are all annotated as neutral, and tail clips are denoted with the other seven emotions: anger, happiness, fear, disgust, sadness, contempt, and surprise. As for one head speech, we randomly select two or three tails with different emotions to construct the head-tail pairs. By leveraging the two-second (i.e., corresponding 30-frame postures) transition to blend the heads and tails, we obtain the 10-second human speech clips in our BEAT-ETrans. We obtain 58,077 clips, including a total of 161.3 hours reported in Table 1. Then the clip numbers of training/validation/testing sets are randomly split as 41,908/4,077/12,092. In all of our experiments, we utilize the upper body with 71 joints. + +TED Emotion Transition Dataset (TED-ETrans): Inspired by [32, 36], we further newly collect a TED Emotion Transition Dataset (dubbed TED-ETrans) based on more than 1.7K speakers from in-the-wild TED talk show videos, demonstrated in Table 1. Due to the lack of emotional labels, we first leverage the annotated BEAT dataset to pretrain an audio-based emotion classifier for labeling TED audi- + +Table 1. Statistics comparison of existing 3D co-speech gesture datasets with ours. Our BEAT-ETrans and TED-ETrans are built upon the existing BEAT [28] and TED-Expressive [32], respectively. To the best of our knowledge, we are the first to present two large datasets with emotion transition human speech. + +
DatasetJoint AnnotationModalityDuration (hours)
BodyHandAudioTextSpeakersSingle EmotionEmotion Transition
TED [50]pseudo label9✘✓✓1,766✘✘106.1
SCG [15]pseudo label1424✓✘6✘✘33
Trinity [10]mo-cap2438✓✓1✘✘4
ZeroEGGS [12]mo-cap2748✓✓1✘✘2
BEAT [28]mo-cap2748✓✓30✓✘35
TED-Expressive [32]pseudo label1330✓✓1,764✘✘100.8
BEAT-ETrans (ours)mo-cap2748✓✓308✓161.3
TED-ETrans (ours)pseudo label1330✓✓1,7646✓59.8
+ +Table 2. Comparison with the start-of-the-art methods on our newly collected BEAT-ETrans and TED-ETrans datasets. $\uparrow$ denotes the higher the better, and $\downarrow$ indicates the lower the better. $\pm$ means ${95}\%$ confidence interval. + +
ModelsBEAT-ETransTED-ETrans
FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑
Seq2Seq [49]ICRA'1940.9547.930.14196.66±2.1629.6049.470.26572.81±1.99
S2G [13]CVPR'1925.5637.040.67198.26±2.0418.1641.630.82476.82±2.32
Trimodal [50]TOG'2014.0942.500.764100.87±2.1221.0633.200.75882.87±1.86
CAMN [28]ECCV'229.0327.530.794118.46±2.3319.2841.040.78579.03±1.49
HA2G [32]CVPR'227.2825.790.779121.77±2.3116.7240.380.78780.14±1.65
DiffGesture [53]CVPR'236.6825.030.788122.29±2.0118.6925.130.81892.01±2.07
Ours4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
+ +classification threshold as $\geq 0.95$ , and the two uncommon emotions (i.e. fear, disgust) are dropped. Then we maintain the same data pre-processing strategy with our BEAT-ETrans to obtain a total of 21,515 clips with 59.8 hours. The final clip division criteria of the TED-ETrans dataset are training/validation/testing with 15,061/2,152/4,302 respectively. In practice, the 43 upper body joints are leveraged in the experiments. + +Implementation Details: We set the total generated cospeech gesture length as $N = 150$ , and the transition and chunk lengths are $L = 30$ . Conventionally, we leverage $M = 4$ frames as the reference initial poses. The feature dimension $D = 512$ in practice. The raw audio of human speech is converted to mel-spectrograms with FFT window size 1024, and hop length 512. The audio encoder takes the ResNetSE34 [6] as the backbone. Empirically, we set $\lambda_r = 20$ , and $\lambda_{adv} = 2$ . Our models are implemented on the Pytorch platform with a single NVIDIA Tesla V100 GPU. The initial learning rate is set to 0.0003 by utilizing Adam Optimizer. The whole training takes 100 epochs with a batch size of 96. + +Evaluation Metrics: To fully evaluate the realism and diversity of the generated co-speech gestures, we introduce various metrics: + +- FGD: Fréchet Gesture Distance (FGD) [50] is utilized to + +measure the distribution distance between the realistic sequential gestures and generated ones. Since we only have the 3D joint annotation of the head/tail, we take the network architecture provided in [32, 50] to train the autoencoder for distance computing on the two datasets, respectively. The FGD of transition gestures is calculated as the average value between the distribution distance of transition and head/tail, indicated as $\mathrm{FGD}_{\text {trans }}$ . Similarly, $\mathrm{FGD}_{h + t}$ means the distance between the generated head/tail gestures and ground truth. + +- BC: Beat Consistency Score (BC) [28, 32] measures the speech audio alignment degree with the generated co-speech gestures. +- Diversity: Similar to [32, 53], we exploit the same feature extractor in FGD to obtain the feature embeddings of the generated gestures. The diversity reflects the average distance between 500 random combination pairs in the testing set of 60 speech audios. + +# 4.2. Quantitative Evaluation + +Comparisons with SOTA Methods: To the best of our knowledge, we are the first to explore the co-speech gesture generation with emotion transition human audios. To fully verify the superiority of our method, we implement various state-of-the-art single-emotion-based counterparts: Seq2Seq [49], S2G [13], Trimodal [50], CAMN [28], + +Table 3. Ablation study on different components of the proposed method. $\checkmark$ indicates the employment of a certain module. $\uparrow$ denotes the higher the better, and $\downarrow$ indicates the lower the better. $\pm$ means $95\%$ confidence interval. MTIM: Motion Transition Infusion Mechanism; EMS: Emotion Mixture Strategy; FKS: Keyframe Sampler. + +
Model VariationsBEAT-ETransTED-ETrans
BaselineMTIMEMSKFSFGDh+t↓FGDtrans↓BC↑Diversity ↑FGDh+t↓FGDtrans↓BC↑Diversity ↑
✓21.2156.330.70188.89±1.5826.4447.870.71274.19±2.75
✓✓10.7527.940.866108.69±2.6320.9536.380.82782.42±2.24
✓✓✓5.6921.700.878112.80±1.9514.0226.210.90087.21±2.12
✓✓✓✓4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
+ +HA2G [32], and DiffGesture [53]. For a fair comparison, all the models are implemented by the source code released by the authors. Note that since we only have the ground truth of head and tail gestures, several recent VAE-based works [1, 2, 48, 52] cannot be directly applied in the experiments (or they have not released codes so far). + +As reported in Table 2, we adopt the $\mathrm{FGD}_{h + t}$ , $\mathrm{FGD}_{trans}$ , BC, and Diversity for a well-rounded view of comparisons. Our method outperforms all the competitors by a large margin on both two datasets. Remarkably, on the TED-ETrans dataset, our method even achieves $34.8\%$ (i.e., $(18.69 - 12.19) / 18.69 \approx 34.8\%$ ) improvement over the sub-optimal counterparts in $\mathrm{FGD}_{h + t}$ . We observe both the DiffGesture [53] and ours synthesize the high-fidelity gestures of head/tail speech with much lower $\mathrm{FGD}_{h + t}$ than others. However, the DiffGesture shows worse performance on $\mathrm{FGD}_{trans}$ due to lack of supervision. In terms of diversity, our simple yet effective keyframe sampler provides authority and diverse initial postures as the reference, thus enabling us to demonstrate diverse gesture styles compared to other counterparts. Moreover, we find that the diversity scores on BEAT-ETrans are much higher than those on TED-ETrans dataset. This can be attributed to the more complex human joints in the BEAT-ETrans dataset. + +Ablation Study: To further verify the effectiveness of our proposed methods, we conduct the ablation study of different components as variations, reported in Table 3. The baseline model is implemented by a simple transformer-based pipeline with stacking three times blocks in the encoder-decoder. Obviously, all the combinations of our proposed components have positive impacts on the generated results. Specifically, by adding the motion transition infusion mechanism to the baseline, the indicator BC has achieved significant improvement (e.g., $0.701 \rightarrow 0.866$ in the BEAT-Etrans). This result verifies that our motion transition infusion mechanism effectively models the temporal correlation between the transitions w.r.t. head/tail gestures, thus leading to the generated results preserving rhythm coherency with given speech, globally. + +Moreover, adopting the emotion mixture strategy ideally improves the performance of $\mathrm{FGD}_{trans}$ on both two + +datasets. This indicates that the learnable emotion mixture wights can provide effective weak supervision by leveraging the pre-trained pose-based emotion classifier. Besides, we have observed significant improvement in the performance of $\mathrm{FGD}_{h + t}$ during this phase compared to the previous version. This aligns with our transformer backbone's emphasis on modeling the sequential temporal correlations as a whole. The better transition gestures encourage our model to maintain better temporal consistency, thereby the head/tail gestures achieve better results. + +Finally, after additionally employing the keyframe sampler to produce authority initial postures as the reference, our method obtains the best performance. Although the properties of BC and FGD on both datasets just have slightly improvement, the diversity realizes a noticeably better achievement (e.g., $112.80 \rightarrow 124.93$ in the BEAT-ETrans dataset). This highly supports our insight into keyframe-driven diversification strategy. + +# 4.3. Qualitative Evaluation + +Visualization: To fully demonstrate the performance of our method, we show the visualized keyframes generated from ours compared with counterparts on our newly collected TED-ETrans and BEAT-ETrans datasets, respectively. As depicted in Figure 3, our method displays vivid and diverse results against others. In particular, we observe that Seq2Seq and Trimodal tend to synthesize unreasonable and stiff results (e.g., the blue rectangle in the right BEAT-ETrans dataset). Although CAMN and HA2G can generate natural upper-body postures, we find that they sometimes produce unreliable subtle fingers (e.g., the red rectangle in the left TED-ETrans dataset). Both our method and DiffGesture create reasonable gestures. However, the results synthesized by DiffuGesture are mismatched with the emotion transitions. In contrast, our method can synthesize the synchronous motions (e.g., in the BEAT-ETrans, the arms become droopy as emotion turns to sadness). Meanwhile, we further verify the diversification results as shown in Figure 1. Given the same input audio, our method generates diverse and vivid co-speech gestures. Please refer to the supplementary material for more visualization results. + +![](images/76cca928db845133bbc5a0af9f746f3de94363a40259366ce6dc6f317dd9af4d.jpg) +Figure 3. Visualization of our generated 3D co-speech gestures against various state-of-the-art methods. The samples of the left part are from our newly collected TED-ETrans dataset, and the samples of the right part are from our BEAT-ETrans dataset. Best view on screen. + +User Study: To further analyze the quality of results by various counterparts and ours, we conduct a user study by visualizing the results and inviting 15 subjects. For each emotion in both datasets, we randomly select 12 samples for each participant (7 emotions in BEAT-ETrans, 5 emotions in TED-ETrans). All the participants are recruited anonymously from schools including various majors. In particular, the subjects are required to rate the generated co-speech gestures from 0 to 5 (the higher, the better) in terms of naturalness, motion smoothness, and speech-gesture coherency. The results are demonstrated in Figure 4. Our method showcases the best performance compared with all the competitors. Especially in terms of motion smoothness, our method achieves noticeable advantages, indicating the effectiveness of our proposed motion transition infusion mechanism and the emotion mixture strategy. + +# 5. Conclusion + +In this paper, we introduce a new task of 3D co-speech gesture generation given emotion transition human speech. We therefore newly collected two datasets named the BEAT-ETrans and the TED-ETrans to fulfill this goal while significantly facilitating the research on 3D human motion modeling. Then, we fully take advantage of the sequential temporal correlation via a motion transition infusion mechanism + +![](images/f33d9cd1a3b6d664abb605ba5c08d7a67132d8284afbbf4724235a8e1af55d33.jpg) +Figure 4. User study on gesture naturalness, motion smoothness, and speech-gesture coherency. + +to ensure the generated gestures preserve temporal coherence. Furthermore, we design an emotion mixture strategy to supply emotional weak supervision of the synthesized transitions. Extensive experiments conducted on our two newly collected datasets show the superiority of the method. As our method intends to generate diverse and vivid emotion transition gestures, we will investigate diversifying the 3D gesture with temporal smooth sampling, instead of the keyframe-wise manner. + +Acknowledgements. This work is funded in part by the Theme-based Research Scheme grant (No.T45-205/21-N) and the InnoHK funding, Hong Kong SAR. + +# References + +[1] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 7 +[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2, 3, 7 +[3] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2027-2036, 2021. 2 +[4] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 2 +[5] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 2 +[6] Joon Son Chung, Jaesung Huh, Seongkyu Mun, Minjae Lee, Hee Soo Heo, Soyeon Choe, Chiheon Ham, Sunghwan Jung, Bong-Jin Lee, and Icksang Han. In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982, 2020. 6 +[7] Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. arXiv preprint arXiv:2005.07143, 2020. 3 +[8] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. arXiv preprint arXiv:2305.08953, 2023. 3 +[9] Maged Farouk. Studying human robot interaction and its characteristics. International Journal of Computations, Information and Manufacturing (IJCIM), 2(1), 2022. 2 +[10] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 6 +[11] Yu Fu, Yan Hu, and Veronica Sundstedt. A systematic literature review of virtual, augmented, and mixed reality game applications in healthcare. ACM Transactions on Computing for Healthcare (HEALTH), 3(2):1-27, 2022. 2 +[12] Saeed Ghorbani, Ylva Fe2rstl, Daniel Holden, Nikolaus F Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. In Computer Graphics Forum, pages 206–216. Wiley Online Library, 2023. 6 +[13] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 6 + +[14] Susan Goldin-Meadow and Martha Wagner Alibali. Gesture's role in speaking, learning, and creating language. Annual review of psychology, 64:257-283, 2013. 2 +[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 6 +[16] Autumn B Hostetter and Martha W Alibali. Visible embodiment: Gestures as simulated action. Psychonomic bulletin & review, 15:495-514, 2008. 2 +[17] Ann Huang, Pascal Knierim, Francesco Chiossi, Lewis L Chuang, and Robin Welsch. Proxemics for human-agent interaction in augmented reality. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-13, 2022. 2 +[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4 +[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 3 +[20] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2 +[21] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thórisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2 +[22] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities for reactive robotic response. In IROS, page 2071. Tokyo, 2013. 2 +[23] Sergey Levine, Philipp Krahenbuhl, Sebastian Thrun, and Vladlen Koltun. Gesture controllers. In Acm siggraph 2010 papers, pages 1-11. 2010. 2 +[24] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020.3 +[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 3 +[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3 + +[27] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7590–7598, 2023. 2 +[28] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6 +[29] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. AudioLDM: Text-to-audio generation with latent diffusion models. Proceedings of the International Conference on Machine Learning, 2023. 2, 3 +[30] Haohe Liu, Qiao Tian, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D. Plumbley. AudioLDM 2: Learning holistic audio generation with self-supervised pretraining. arXiv preprint arXiv:2308.05734, 2023. 2, 3 +[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via masked audio gesture modeling. arXiv preprint arXiv:2401.00374, 2023. 3 +[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 5, 6, 7 +[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 3, 4 +[34] OpenAI. Gpt-4 technical report, 2023. 2, 3 +[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3 +[36] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2, 3, 5 +[37] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3d hand gesture prediction from body dynamics by bilateral hand disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4616-4626, 2023. 2, 5 +[38] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 3 +[39] Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan, + +Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, et al. Speechbrain: A general-purpose speech toolkit. arXiv preprint arXiv:2106.04624, 2021. 3 +[40] Peter Roach. Some languages are spoken more quickly than others. 1998. 3 +[41] Mehmet E Sargin, Yucel Yemez, Engin Erzin, and Ahmet M Tekalp. Analysis of head gesture and prosody patterns for prosody-driven head-gesture animation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(8):1330-1345, 2008. 2 +[42] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 3 +[43] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 5 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[45] Jordan Voas, Yili Wang, Qixing Huang, and Raymond Mooney. What is the best automated metric for text to motion generation? arXiv preprint arXiv:2309.10248, 2023. 3 +[46] Zijian Wang, Xingqun Qi, Kun Yuan, and Muyi Sun. Self-supervised correlation mining network for person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7703-7712, 2022. 2 +[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 5860–5868, 2023. 2 +[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 2, 3, 7 +[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 6 +[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 2, 3, 5, 6 +[51] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In International Cconference on Multimodal Interaction, pages 179-185. 2023. 3 + +[52] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 7 +[53] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 5, 6, 7 \ No newline at end of file diff --git a/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/images.zip b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86edff855482dbf9bad2b6e8697e8c645d91c0ee --- /dev/null +++ b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0af131a7ae16cb5baebb96929f8aa00d4fb8c4021507abff93583985a6a98609 +size 549454 diff --git a/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/layout.json b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2361fcb3e1e953ec546616427b801e0918e17120 --- /dev/null +++ b/2024/Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation/layout.json @@ -0,0 +1,7940 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 139 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 139 + ], + "type": "text", + "content": "Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": "Xingqun Qi" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Jiahao Pan" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Peng Li" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Ruibin Yuan" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Xiaowei Chi" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Mengfei Li" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Wenhan Luo" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Wei Xue" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Shanghang Zhang" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Qifeng Liu" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1,\\boxtimes}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", Yike Guo" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1,\\boxtimes}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": " The Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 160, + 487, + 217 + ], + "type": "text", + "content": " Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 220, + 440, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 220, + 440, + 232 + ], + "spans": [ + { + "bbox": [ + 147, + 220, + 440, + 232 + ], + "type": "text", + "content": "xingqun.qi@connect.ust.hk, {liuqifeng, yikeguo}@ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 64, + 241, + 529, + 411 + ], + "blocks": [ + { + "bbox": [ + 64, + 241, + 529, + 411 + ], + "lines": [ + { + "bbox": [ + 64, + 241, + 529, + 411 + ], + "spans": [ + { + "bbox": [ + 64, + 241, + 529, + 411 + ], + "type": "image", + "image_path": "de4fc90de6e2baa5b3d0bc2d0986790c72f3606ad31cdf70f1f53f2540899ba6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 417, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 417, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 547, + 452 + ], + "type": "text", + "content": "Figure 1. Diverse exemplary clips sampled by our method from our newly collected BEAT Emotion Transition Dataset. The vital frames are visualized to demonstrate that the upper body gestures change with the emotion transition of human speech, synchronously. From top to bottom: the input speech audio, the corresponding transcript, and two sampled clips. Best view on screen." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 460, + 192, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 460, + 192, + 474 + ], + "spans": [ + { + "bbox": [ + 143, + 460, + 192, + 474 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 487, + 289, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 289, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 289, + 691 + ], + "type": "text", + "content": "Generating vivid and emotional 3D co-speech gestures is crucial for virtual avatar animation in human-machine interaction applications. While the existing methods enable generating the gestures to follow a single emotion label, they overlook that long gesture sequence modeling with emotion transition is more practical in real scenes. In addition, the lack of large-scale available datasets with emotional transition speech and corresponding 3D human gestures also limits the addressing of this task. To fulfill this goal, we first incorporate the ChatGPT-4 and an audio inpainting approach to construct the high-fidelity emotion transition human speeches. Considering obtaining the realistic 3D pose annotations corresponding to the dynamically inpainted emotion transition audio is extremely difficult, we propose a novel weakly supervised training strategy to encourage authority gesture transitions. Specifically, to enhance the coordination of transition gestures w.r.t. dif" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 547, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 630 + ], + "type": "text", + "content": "ferent emotional ones, we model the temporal association representation between two different emotional gesture sequences as style guidance and infuse it into the transition generation. We further devise an emotion mixture mechanism that provides weak supervision based on a learnable mixed emotion label for transition gestures. Last, we present a keyframe sampler to supply effective initial posture cues in long sequences, enabling us to generate diverse gestures. Extensive experiments demonstrate that our method outperforms the state-of-the-art models constructed by adapting single emotion-conditioned counterparts on our newly defined emotion transition task and datasets. Our code and dataset will be released on the project page: https://xingqunqi-lab.github.io/Emo-Transition-Gesture/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 656, + 387, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 387, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 387, + 668 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Co-speech gesture generation aims to synthesize vivid and emotional human postures coordinated with the audio input. These non-verbal behaviors serve as a key factor dur-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 149, + 713 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10424" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "ing human conversations that significantly facilitates the delivery of speech content [5, 14, 16]. Meanwhile, modeling co-speech gestures has a wide range of embodied AI applications in human-machine interaction [22, 27], robot assistants [9], and virtual/augmented reality (AR/VR) [11]. Conventionally, many researchers usually focus on synthesizing human upper body gestures consistent with speech audio [32, 47, 48, 50, 53]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 290, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 290, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 290, + 397 + ], + "type": "text", + "content": "Nevertheless, except for a few recent works that generate the co-speech gestures of a single emotion category [2, 3, 28, 52], previous works mostly focus on emotion-agnostic generation [4, 48, 50]. Most of them overlook synthesizing the long sequence co-speech gestures with the emotion transitions, which are more practical in real-world scenes. For example, a person may not maintain a single emotion forever when communicating with others or in speech talking. In this work, we therefore introduce the task of speech-driven emotion transition for generating vivid and diverse 3D co-speech gestures, displayed in Figure 1. There are two main challenges in this task: 1) Datasets of 3D co-speech gestures synchronized with emotion transition speech audios are scarce. Creating such a dataset containing 3D human postures is difficult due to the lack of guidance from emotional experts and complex motion capture systems. 2) Modeling the plausible and temporal coherent co-speech gestures from one emotion to another in long sequences is difficult, especially in the transition duration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "content": "To overcome the issue of data scarcity, we newly present two datasets containing emotion-transition human speech that are built upon previous single-emotion ones [28, 32]. In particular, thanks to the developed language model ChatGPT-4 [34], we first leverage it to generate text transcripts of the transition based on speech context. Then, by employing the audioLDM2 technique [29, 30] for audio inpainting, we ensure the inpainted transition's timbral consistency with its adjacent contexts and a smooth emotional transition throughout. Afterwards, to support our insight on modeling co-speech gestures coherent with emotion-transition speech, corresponding 3D human postures of transition are required." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "However, due to the dynamically generated transition transcripts, it is infeasible to construct the aligned realistic 3D pose annotation of human bodies. Hence, we solve the challenges of vivid co-speech gesture generation in a novel weakly supervised pattern, containing a motion transition infusion mechanism and an emotion mixture strategy. Specifically, in the motion transition infusion mechanism, we model the temporal correlation between the generated head and tail gesture features as style guidance representation. The style guidance representation provides motion transition cues that are infused into the transition embedding via an adaptive instance normalization (AdaIN) layer [18]. Along with this manner, we can effectively en" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "hance the coordination of transition gestures w.r.t. two different emotional ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 96, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 546, + 323 + ], + "type": "text", + "content": "Moreover, to alleviate the lack of supervision during the transition between two emotions, the emotion mixture strategy is built to provide weak emotional supervision of the generated transition gestures. Concretely, we learn a joint embedding of two different emotional gesture sequences using a temporal aggregation encoder. Then, we pre-train an emotion classifier based on the annotated human 3D poses with single emotion labels in the dataset. Here, this joint embedding is leveraged as an emotion mixture weight for the pre-trained classifier to facilitate high-fidelity transition gesture synthesis with desirable properties. Finally, considering the generated 3D postures should be non-deterministic given the human speech, we devise a keyframe sampler to produce diverse initial poses as reference. In this fashion, our method enables diverse co-speech gesture generation with emotion transitions. Extensive experiments conducted on our newly constructed two datasets verify the effectiveness of our methods, displaying vivid and emotional 3D co-speech gestures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 324, + 523, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 523, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 523, + 335 + ], + "type": "text", + "content": "Overall, our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 342, + 545, + 530 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 342, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 545, + 413 + ], + "type": "text", + "content": "- We introduce a new task of emotion transition co-speech gesture generation cooperating with two newly constructed datasets named BEAT Emotion Transition (BEAT-ETrans) and TED Emotion Transition (TED-ETrans), significantly facilitating research on 3D human motion modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 418, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 545, + 479 + ], + "type": "text", + "content": "- We design a motion transition infusion mechanism to ensure the temporal coordination of transition gestures w.r.t. two different emotional ones and a weakly supervised emotion mixture strategy to enable high-fidelity transition gesture synthesis with desirable properties." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 483, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 545, + 530 + ], + "type": "text", + "content": "- Extensive experiments show that our method outperforms state-of-the-art counterparts on both datasets, displaying realistic and vivid co-speech gestures given emotion transition human audiios." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 543, + 392, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 543, + 392, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 543, + 392, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 563, + 463, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 463, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 463, + 575 + ], + "type": "text", + "content": "2.1. Co-speech Gesture Synthesis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Synthesizing human co-speech gestures plays a significant role in various applications [17, 22, 37, 46]. Numerous studies have been proposed to address these issues that are roughly divided into rule-based approaches [20, 21], machine-learning-based approaches [23, 41], and deep learning-based ones [2, 28, 32, 36, 48, 50, 52, 53]. Traditional researchers follow the rule-based workflow, leveraging the speech-gesture pairs as guidance to generate co-speech gestures pre-defined by linguistic experts. Other early works integrate the manually defined gesture features with machine learning techniques to synthesize the co" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10425" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "speech gestures. In the aforementioned two manners, the researchers usually focus on the optimization of the matching process between human speech and pre-defined gestures. It may require experts much expensive effort in the speech-gesture pair construction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 275 + ], + "type": "text", + "content": "Recently works focus on building the mapping directly from the input human speech and sequential gestures by exploiting the deep neural networks. They usually leverage multi-modality cues to facilitate the generation of co-speech gestures, associating with the speech audio [25, 48, 53], speaker identity [50], emotion [2, 28, 31], and text transcript [32]. However, they overlook that emotion transition of the long sequential co-speech gesture modeling is much more practical in the real scenes. Moreover, since the lack of annotated 3D gestures corresponding to dynamically constructed transition speech, few of the above methods could be directly adapted to this new thought." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 283, + 205, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 205, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 205, + 295 + ], + "type": "text", + "content": "2.2. 3D Human Motion Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 529 + ], + "type": "text", + "content": "Human motion modeling aims to generate realistic and smooth human motions with various multi-modality conditions [8, 19, 33], including co-speech gesture generation as a sub-task. One of the hottest topics is synthesizing human motion from text prompts with a few past postures as the seed [35, 45, 51]. These methods usually engage in forcing human motion to represent the semantic expression aligned with the text. Literally, the task most closely related to ours is AI choreographer [24, 26, 42] which generates the motion from music signals. However, the AI choreographer works mainly focus on the rhythmic-coherent motion of the whole human body but without subtle finger gestures. While sharing a similar goal with the approaches mentioned above, our newly defined work differs from them significantly. We take the emotion-transition long sequence co-speech gesture without corresponding 3D human pose annotation into consideration, thus motivating us to utilize the motion transition of two annotated emotional gestures and coherent the overall sequence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 539, + 153, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 153, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 153, + 552 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 558, + 263, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 263, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 263, + 570 + ], + "type": "text", + "content": "3.1. Emotion Transition Dataset Construction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 577, + 286, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 286, + 660 + ], + "type": "text", + "content": "We aim to address the emotion transition co-speech gesture generation in a weakly supervised manner. Due to the existing paired speech-gesture datasets [28, 36], we could focus on synthesizing the high-fidelity transition human audios. Synthesizing datasets conducive to our task focuses on ensuring semantic coherence, smooth emotional audio transitions, consistent timbre, and audio fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Preliminary: Considering our key insight to modeling the co-speech gesture with emotion transition, we first split the existing aligned speech-gesture pairs [28, 36] into four-second clips. To diversify the datasets, we randomly splice" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "two clips from the same speaker to construct an emotion-transition candidate pair. The clip with neutral emotion is leveraged as head speech, and the other with various emotions is represented as tail speech. We leverage the dynamically synthesized two-second audio as a transition to combine the head and tail speeches. In the following, we briefly summarize our efforts in constructing the dynamic transition speech audio." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 173, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 257 + ], + "type": "text", + "content": "Textual Inpainting of Transition: To ensure the semantic coherence of transition w.r.t. head/tail speeches, we exploit the advanced language generation model ChatGPT-4 [34] to complete the transcript according to the context. Literally, we follow the conventional estimation that people usually talk 30 phonemes [40] in a two-second speech as the prompt to guide transcript generation by ChatGPT-4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 262, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 417 + ], + "type": "text", + "content": "Synthesis of Transition Audio: Once we obtain the transcript of transitions, we employ a superior text-to-speech model, audioLDM2 [29, 30], to generate corresponding speech audio. Here, we leverage the speaker embeddings extracted via SpeechBrain's ECAPA-TDNN [7, 39] as prior guidance to maintain the identity consistency of the generated transition audio. Then, we adopt Whisper [38] to restrict the word error rate, thereby ensuring the accuracy and clarity of the speech content. In this fashion, the synthesized transition audio realizes controlled duration while the natural smooth tonality is well preserved. The datasets will be released in the future. More details are provided in the Section 4.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 427, + 429, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 429, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 429, + 438 + ], + "type": "text", + "content": "3.2. Problem Formulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": "Given a sequence of audio signal " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "A = \\{a_{1},\\dots,a_{N}\\}" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": " as input to model " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": ", our goal is to generate vivid and emotional 3D co-speech gestures " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "P = \\{p_1,\\dots,p_N\\}" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": " of the upper human body. " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": " represents the number of synthesized postures corresponding to the audio " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": " joints with 3D representation. In particular, we define the audio signal as consisting of a head speech, a dynamically inpainted transition speech, and a tail speech. Here, the head and tail speeches are randomly selected from previous co-speech gesture datasets [28, 32] with a single emotion label, respectively. Following the conventions of previous works [28, 32, 50, 53], we invoke " + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 446, + 545, + 613 + ], + "type": "text", + "content": " frame poses as the initial seed to guide generation. The overall objective is expressed as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 624, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 624, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 350, + 624, + 545, + 644 + ], + "type": "interline_equation", + "content": "\\underset {\\mathcal {G}} {\\arg \\min } \\| P - \\mathcal {G} (A, \\{p _ {1}, \\dots , p _ {M} \\}) \\|. \\tag {1}", + "image_path": "ffa1b09f0c521ca67899ff8f4b41a8c260fb257d66905d1a78933823c927c4d4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Note that only the generated gestures with head and tail speeches are supervised with ground truth coming from existing datasets [28, 32]. The transition gestures with " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " frames, where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "L \\ll N" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ", will be weakly supervised through the following processes within a motion transition infusion" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10426" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 219, + 285 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 219, + 285 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 219, + 285 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 219, + 285 + ], + "type": "image", + "image_path": "662578f6ec4cc05fcf2d0fec7f6f331fd96aa0dfcb47d4764f04e953c20b9580.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 547, + 340 + ], + "type": "text", + "content": "Figure 2. The overview of our proposed method. The middle part (blue) displays the overall pipeline for 3D co-speech gesture generation from emotion transition human speech. The left part (green) depicts our proposed Motion Transition Infusion Mechanism that enhances the coordination of transition gestures w.r.t. head/tail ones. The right part (orange) shows the designed Emotion Mixture Strategy to provide weak supervision of the generated transition gestures, thereby achieving authority producing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 69, + 370, + 285 + ], + "blocks": [ + { + "bbox": [ + 220, + 69, + 370, + 285 + ], + "lines": [ + { + "bbox": [ + 220, + 69, + 370, + 285 + ], + "spans": [ + { + "bbox": [ + 220, + 69, + 370, + 285 + ], + "type": "image", + "image_path": "cb9764782a5de94261561685a9bd967e28e50cc24601da8b5754068c517bf27b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 70, + 525, + 285 + ], + "blocks": [ + { + "bbox": [ + 370, + 70, + 525, + 285 + ], + "lines": [ + { + "bbox": [ + 370, + 70, + 525, + 285 + ], + "spans": [ + { + "bbox": [ + 370, + 70, + 525, + 285 + ], + "type": "image", + "image_path": "749191ea0de3bdfb80e452fc3fb24446d9ab2f5055bc03d0fe16a35985c5b9a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 392 + ], + "type": "text", + "content": "mechanism and an emotion mixture strategy. The audio signal is fed into an audio encoder for feature extraction. Our overall workflow is shown in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 253, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 253, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 253, + 408 + ], + "type": "text", + "content": "3.3. Weakly-supervised Emotion Transition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": "Motion Transition Infusion Mechanism: To ensure the temporal coherence of the transition gestures w.r.t. head/tail ones, we propose a motion transition infusion mechanism to model the temporal association representation between different emotional gestures. As depicted in Figure 2 (left), the total sequential features " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{total}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": " consist of the features of head " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{head}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": ", transition " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{tran}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": ", and tail " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{tail}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": ". Inspired by [33], we nominate a head chunk " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{chunk}^{head}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": " composed of the last " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": " frames of head gesture embeddings and a tail chunk " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{chunk}^{tail}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": " consisting of the first " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": " ones of the tail gesture embeddings. Here, the dimension of each frame representing gesture embedding is " + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 46, + 415, + 287, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "In particular, we first calculate the temporal correlation matrix " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{trans}^{head} \\in \\mathbb{R}^{L \\times L}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " between the head chunk embedding and transition embedding. Here, the temporal correlation matrix represents the temporal variations in the gestures from the head to the transition. Similarly, we obtain the correlation matrix " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{trans}^{tail}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " from tail chunk embedding and transition embedding. Then, the global temporal dependency from head to tail is computed via matrix multiplication between " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{trans}^{head}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{trans}^{tail}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ". Further, we develop a motion encoder to acquire the sequence-aware style guidance representation based on the global temporal dependency. The style guidance representation is exploited to boost the transition embeddings via an adaptive instance normalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "type": "text", + "content": "(AdaIN) layer [18]. By doing so, we derive the transition " + }, + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{tran}" + }, + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 381, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 381, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 381, + 545, + 398 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {t r a n} = \\operatorname {A d a I N} \\left\\{\\mathbf {f} ^ {t r a n}, \\operatorname {E n c} \\left(\\mathbf {C} _ {t r a n s} ^ {h e a d} \\otimes \\mathbf {C} _ {t r a n s} ^ {t a i l}\\right) \\right\\}, \\tag {2}", + "image_path": "df44cbb83729c74dd22dd034b4722d0c522fc79f019756e5af32ccff4c236450.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 401, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 545, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 401, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 305, + 401, + 545, + 424 + ], + "type": "text", + "content": " indicates matrix multiplication, and Enc denotes the motion encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 544 + ], + "type": "text", + "content": "**Emotion Mixture Strategy:** Considering obtaining the realistic 3D pose annotations corresponding to dynamically inpainted transition audio is quite difficult, we design an emotion mixture strategy to provide weak supervision of the generated gestures. Our key insight is built upon the fact that different emotions in the head/tail would lead to different gesture motions, thereby the emotion represented by transition gestures would be a mixture of the head and tail ones. As shown in Figure 2 (right), we utilize two learnable parameters as soft emotion labels of the transition gestures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "Specifically, through computing the correlation matrix between the embeddings of head gestures and tail gestures, we obtain the motion deformation representation " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " from one emotion to another (e.g., neutral-to-anger). This motion deformation is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 606, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 606, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 357, + 606, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {\\text {h e a d}} ^ {\\text {t a i l}} = \\operatorname {S o f t m a x} \\left(\\mathbf {f} ^ {\\text {h e a d}} \\otimes \\mathbf {f} ^ {\\text {t a i l} ^ {\\prime}}\\right), \\tag {3}", + "image_path": "64c384cd4b024ae898cd1fcfd6787e8afe4dd2a339d9f3fe3547cdc3ca325874.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "type": "inline_equation", + "content": "{}^{\\prime }" + }, + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "type": "text", + "content": " indicates the transpose operation. With the help of the motion deformation representation, we obtain the two learnable emotion embeddings " + }, + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "type": "inline_equation", + "content": "\\left( {{\\mathbf{e}}^{\\text{head }},{\\mathbf{e}}^{\\text{tail }}}\\right)" + }, + { + "bbox": [ + 304, + 628, + 547, + 676 + ], + "type": "text", + "content": " from the head and tail,respectively,as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 328, + 679, + 522, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 679, + 522, + 695 + ], + "spans": [ + { + "bbox": [ + 328, + 679, + 522, + 695 + ], + "type": "interline_equation", + "content": "\\mathbf {e} ^ {h e a d} = \\operatorname {M a x P o o l} \\left\\{\\mathbf {W} _ {\\text {l i n e a r}} (\\mathbf {f} ^ {h e a d}) \\odot \\mathbf {S} _ {\\text {h e a d}} ^ {t a i l} \\right\\},", + "image_path": "1ba27ab1f2bd984c493cea302b1bc5a97f2ea34ac693dc3e3b9ea0507fc81037.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 333, + 696, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 696, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 333, + 696, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {e} ^ {t a i l} = \\operatorname {M a x P o o l} \\left\\{\\mathbf {W} _ {t a i l} \\left(\\mathbf {f} ^ {h e a d}\\right) \\odot \\mathbf {S} _ {h e a d} ^ {t a i l} \\right\\}, \\tag {4}", + "image_path": "6d4e665fd1b1d34cb2b947cebb987de4365a54f39d3e3c4e6a411144a60c1dff.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10427" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{linear}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " denotes linear matrix, and " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " means dot product. Maxpool indicates the AdaptiveMaxPooling operation. " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{tail}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " is calculated in a similar way. Then, we obtain the two learnable parameters represented by emotion weights of head and tail gestures, as follows" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 143, + 287, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 143, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 94, + 143, + 287, + 157 + ], + "type": "interline_equation", + "content": "(\\sigma , 1 - \\sigma) = \\operatorname {S o f t m a x} \\left(\\mathbf {e} ^ {\\text {h e a d}}, \\mathbf {e} ^ {\\text {t a i l}}\\right). \\tag {5}", + "image_path": "52f9737f6dd7ec6cce50f8aad89562edae9e1060ab4ec8379c536a1aa4555cb1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": "Once we acquire the learnable emotion weights, we leverage a pre-trained pose-based emotion classifier to provide weak supervision. In this fashion, the authority of generated transition gestures is well-preserved. The training details of the pre-trained pose-based emotion classifier are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "type": "text", + "content": "Keyframe Posture Sampling: Conventionally, researchers [32, 50, 53] directly leverage the padded initial poses as conditional seeds to guide the co-speech gesture generation. In long-sequence modeling, an intuitive manner is to extend the initial pose length " + }, + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "type": "text", + "content": " for adapting the overall synthesized gestures' length. However, this would lead to the poor generalizability of the network (i.e., the performance gains degradation from the reduction of initial poses). Therefore, we propose a simple yet effective VAE-based [43] keyframe sampler to provide high-fidelity posture prior conditions while enabling diverse results. Evenly, we split the annotated sequence into several chunks keeping the same length with the transition of length " + }, + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 245, + 287, + 472 + ], + "type": "text", + "content": ". The sampler is trained with the keyframe randomly selected to reconstruct the corresponding chunk. In the inference phase, the keyframe sampler samples diverse chunks, blending as the initial postures for gesture generation. Since the transition sequence lacks the pose annotation, we randomly select the frame from head or tail sequences for posture sampling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "text", + "content": "To further enhance the sequence-aware correspondence of the generated co-speech gestures, we leverage the diverse initial postures as the query " + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "text", + "content": " to match the key features " + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "text", + "content": " and value features " + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 474, + 287, + 569 + ], + "type": "text", + "content": " in the transformer-based backbone [44]. Similar to [37], we adopt a motion discriminator to ensure the temporal smoothness of the generated results. For more details about network architecture please refer to supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 585, + 163, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 163, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 163, + 597 + ], + "type": "text", + "content": "3.4. Objective Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 609, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 287, + 645 + ], + "type": "text", + "content": "Reconstruction Loss: We leverage the ground truth 3D pose annotation of the head and tail to constrain the generated co-speech gestures as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 655, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 655, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 89, + 655, + 287, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left\\| P _ {\\{h e a d, t a i l \\}} - \\hat {P} _ {\\{h e a d, t a i l \\}} \\right\\| _ {1}, \\tag {6}", + "image_path": "8d6d87ba33d93299ab9d3c08f2f72f39713447577abe3211d4104f2c83526d4e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\{\\text{head}, \\text{tail}\\}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denotes generated gestures of head and tail speeches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "Adversarial Learning Loss: To ensure the realism of the generated gestures, we further exploit the adversarial training loss, expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 114, + 545, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 114, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 335, + 114, + 545, + 142 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a d v} = \\mathbb {E} _ {P} [ \\log \\mathcal {D} (P) ] \\\\ + \\mathbb {E} _ {A} [ \\log (1 - (\\mathcal {G} (A, \\{p _ {1}, \\dots , p _ {M} \\})) ], \\tag {7} \\\\ \\end{array}", + "image_path": "b8735bf07663de32dd448f922aab4f106f9fe576a2d280d8779cc37b0b04139f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "text", + "content": " denotes the motion discriminator and " + }, + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "text", + "content": " means gesture generator." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 177, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 177, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 177, + 545, + 224 + ], + "type": "text", + "content": "Weakly Supervision Loss: We leverage the pre-trained pose-based emotion classifier to provide weak supervision of the transition gestures upon the learnable emotion weights:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 230, + 545, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 230, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 361, + 230, + 545, + 244 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e m o t i o n}} = - y \\log \\mathcal {F} (\\hat {P} _ {\\text {t r a n s}}), \\tag {8}", + "image_path": "f242d39817a35caff8e3c23dae2a504aed467f5e60a22063cde2bde02944f997.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "text", + "content": " is the learnable emotion label, " + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "text", + "content": " is the emotion classifier, " + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\hat{P}_{trans}" + }, + { + "bbox": [ + 305, + 250, + 545, + 274 + ], + "type": "text", + "content": " is the generated transition gestures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 274, + 434, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 434, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 434, + 285 + ], + "type": "text", + "content": "Finally, the overall objective is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 292, + 545, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 545, + 310 + ], + "type": "interline_equation", + "content": "\\min _ {G} \\max _ {D} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {r} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\mathcal {L} _ {\\text {e m o t i o n}}. \\tag {9}", + "image_path": "2c77d4a2aded293fa572592c67975ac71eaed50d459873cee4d3ed5dac161611.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "type": "inline_equation", + "content": "\\lambda_{adv}" + }, + { + "bbox": [ + 306, + 315, + 473, + 326 + ], + "type": "text", + "content": " are weight coefficients." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 335, + 386, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 386, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 386, + 348 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 355, + 491, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 355, + 491, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 355, + 491, + 369 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Setting" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 373, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 613 + ], + "type": "text", + "content": "BEAT Emotion Transition Dataset (BEAT-ETrans): Since there are only single emotion labels of aligned speech-gesture corpus in the original BEAT dataset [28], to satisfy our insight on emotion transition co-speech gesture generation modeling, we recollect a BEAT Emotion Transition Dataset (dubbed BEAT-ETrans). In particular, we resample the motion FPS as 15 and intercept the continuous 60 frames with stride 30 as the head/tail clips. Here, the head clips are all annotated as neutral, and tail clips are denoted with the other seven emotions: anger, happiness, fear, disgust, sadness, contempt, and surprise. As for one head speech, we randomly select two or three tails with different emotions to construct the head-tail pairs. By leveraging the two-second (i.e., corresponding 30-frame postures) transition to blend the heads and tails, we obtain the 10-second human speech clips in our BEAT-ETrans. We obtain 58,077 clips, including a total of 161.3 hours reported in Table 1. Then the clip numbers of training/validation/testing sets are randomly split as 41,908/4,077/12,092. In all of our experiments, we utilize the upper body with 71 joints." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": "TED Emotion Transition Dataset (TED-ETrans): Inspired by [32, 36], we further newly collect a TED Emotion Transition Dataset (dubbed TED-ETrans) based on more than 1.7K speakers from in-the-wild TED talk show videos, demonstrated in Table 1. Due to the lack of emotional labels, we first leverage the annotated BEAT dataset to pretrain an audio-based emotion classifier for labeling TED audi-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10428" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 113, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Statistics comparison of existing 3D co-speech gesture datasets with ours. Our BEAT-ETrans and TED-ETrans are built upon the existing BEAT [28] and TED-Expressive [32], respectively. To the best of our knowledge, we are the first to present two large datasets with emotion transition human speech." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 113, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 113, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 113, + 545, + 232 + ], + "type": "table", + "html": "
DatasetJoint AnnotationModalityDuration (hours)
BodyHandAudioTextSpeakersSingle EmotionEmotion Transition
TED [50]pseudo label9✘✓✓1,766✘✘106.1
SCG [15]pseudo label1424✓✘6✘✘33
Trinity [10]mo-cap2438✓✓1✘✘4
ZeroEGGS [12]mo-cap2748✓✓1✘✘2
BEAT [28]mo-cap2748✓✓30✓✘35
TED-Expressive [32]pseudo label1330✓✓1,764✘✘100.8
BEAT-ETrans (ours)mo-cap2748✓✓308✓161.3
TED-ETrans (ours)pseudo label1330✓✓1,7646✓59.8
", + "image_path": "256577836683683bad762129682403823f5c440ca775c0c053521c894b2d65c4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 273, + 541, + 385 + ], + "blocks": [ + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "text", + "content": "Table 2. Comparison with the start-of-the-art methods on our newly collected BEAT-ETrans and TED-ETrans datasets. " + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "text", + "content": " denotes the higher the better, and " + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "text", + "content": " indicates the lower the better. " + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "text", + "content": " means " + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "inline_equation", + "content": "{95}\\%" + }, + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "text", + "content": " confidence interval." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 273, + 541, + 385 + ], + "lines": [ + { + "bbox": [ + 53, + 273, + 541, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 541, + 385 + ], + "type": "table", + "html": "
ModelsBEAT-ETransTED-ETrans
FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑FGDh+t ↓FGDtrans ↓BC ↑Diversity ↑
Seq2Seq [49]ICRA'1940.9547.930.14196.66±2.1629.6049.470.26572.81±1.99
S2G [13]CVPR'1925.5637.040.67198.26±2.0418.1641.630.82476.82±2.32
Trimodal [50]TOG'2014.0942.500.764100.87±2.1221.0633.200.75882.87±1.86
CAMN [28]ECCV'229.0327.530.794118.46±2.3319.2841.040.78579.03±1.49
HA2G [32]CVPR'227.2825.790.779121.77±2.3116.7240.380.78780.14±1.65
DiffGesture [53]CVPR'236.6825.030.788122.29±2.0118.6925.130.81892.01±2.07
Ours4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
", + "image_path": "304c76c8b14819d76fd137a816c54a3b442e0483eaba4d11e184defcedafc218.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "text", + "content": "classification threshold as " + }, + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\geq 0.95" + }, + { + "bbox": [ + 46, + 404, + 288, + 501 + ], + "type": "text", + "content": ", and the two uncommon emotions (i.e. fear, disgust) are dropped. Then we maintain the same data pre-processing strategy with our BEAT-ETrans to obtain a total of 21,515 clips with 59.8 hours. The final clip division criteria of the TED-ETrans dataset are training/validation/testing with 15,061/2,152/4,302 respectively. In practice, the 43 upper body joints are leveraged in the experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": "Implementation Details: We set the total generated cospeech gesture length as " + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "inline_equation", + "content": "N = 150" + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": ", and the transition and chunk lengths are " + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "inline_equation", + "content": "L = 30" + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": ". Conventionally, we leverage " + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": " frames as the reference initial poses. The feature dimension " + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "inline_equation", + "content": "D = 512" + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": " in practice. The raw audio of human speech is converted to mel-spectrograms with FFT window size 1024, and hop length 512. The audio encoder takes the ResNetSE34 [6] as the backbone. Empirically, we set " + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\lambda_r = 20" + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\lambda_{adv} = 2" + }, + { + "bbox": [ + 46, + 505, + 287, + 659 + ], + "type": "text", + "content": ". Our models are implemented on the Pytorch platform with a single NVIDIA Tesla V100 GPU. The initial learning rate is set to 0.0003 by utilizing Adam Optimizer. The whole training takes 100 epochs with a batch size of 96." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 700 + ], + "type": "text", + "content": "Evaluation Metrics: To fully evaluate the realism and diversity of the generated co-speech gestures, we introduce various metrics:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "type": "text", + "content": "- FGD: Fréchet Gesture Distance (FGD) [50] is utilized to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "type": "text", + "content": "measure the distribution distance between the realistic sequential gestures and generated ones. Since we only have the 3D joint annotation of the head/tail, we take the network architecture provided in [32, 50] to train the autoencoder for distance computing on the two datasets, respectively. The FGD of transition gestures is calculated as the average value between the distribution distance of transition and head/tail, indicated as " + }, + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{\\text {trans }}" + }, + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{h + t}" + }, + { + "bbox": [ + 312, + 404, + 545, + 523 + ], + "type": "text", + "content": " means the distance between the generated head/tail gestures and ground truth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 524, + 545, + 619 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 524, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 545, + 559 + ], + "type": "text", + "content": "- BC: Beat Consistency Score (BC) [28, 32] measures the speech audio alignment degree with the generated co-speech gestures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 560, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 545, + 619 + ], + "type": "text", + "content": "- Diversity: Similar to [32, 53], we exploit the same feature extractor in FGD to obtain the feature embeddings of the generated gestures. The diversity reflects the average distance between 500 random combination pairs in the testing set of 60 speech audios." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 625, + 441, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 625, + 441, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 625, + 441, + 639 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Comparisons with SOTA Methods: To the best of our knowledge, we are the first to explore the co-speech gesture generation with emotion transition human audios. To fully verify the superiority of our method, we implement various state-of-the-art single-emotion-based counterparts: Seq2Seq [49], S2G [13], Trimodal [50], CAMN [28]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "10429" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 113, + 541, + 191 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 3. Ablation study on different components of the proposed method. " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " indicates the employment of a certain module. " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " denotes the higher the better, and " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " indicates the lower the better. " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " means " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " confidence interval. MTIM: Motion Transition Infusion Mechanism; EMS: Emotion Mixture Strategy; FKS: Keyframe Sampler." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 113, + 541, + 191 + ], + "lines": [ + { + "bbox": [ + 52, + 113, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 52, + 113, + 541, + 191 + ], + "type": "table", + "html": "
Model VariationsBEAT-ETransTED-ETrans
BaselineMTIMEMSKFSFGDh+t↓FGDtrans↓BC↑Diversity ↑FGDh+t↓FGDtrans↓BC↑Diversity ↑
✓21.2156.330.70188.89±1.5826.4447.870.71274.19±2.75
✓✓10.7527.940.866108.69±2.6320.9536.380.82782.42±2.24
✓✓✓5.6921.700.878112.80±1.9514.0226.210.90087.21±2.12
✓✓✓✓4.4218.840.881124.93±2.1012.1923.540.90693.79±2.53
", + "image_path": "e2b839b96d9dde43712d7a432ce3cd36af2ffa2bcf1d58974e9644383f1a35a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 281 + ], + "type": "text", + "content": "HA2G [32], and DiffGesture [53]. For a fair comparison, all the models are implemented by the source code released by the authors. Note that since we only have the ground truth of head and tail gestures, several recent VAE-based works [1, 2, 48, 52] cannot be directly applied in the experiments (or they have not released codes so far)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": "As reported in Table 2, we adopt the " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{h + t}" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{trans}" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": ", BC, and Diversity for a well-rounded view of comparisons. Our method outperforms all the competitors by a large margin on both two datasets. Remarkably, on the TED-ETrans dataset, our method even achieves " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "34.8\\%" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "(18.69 - 12.19) / 18.69 \\approx 34.8\\%" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": " ) improvement over the sub-optimal counterparts in " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{h + t}" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": ". We observe both the DiffGesture [53] and ours synthesize the high-fidelity gestures of head/tail speech with much lower " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{h + t}" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": " than others. However, the DiffGesture shows worse performance on " + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{trans}" + }, + { + "bbox": [ + 46, + 285, + 288, + 499 + ], + "type": "text", + "content": " due to lack of supervision. In terms of diversity, our simple yet effective keyframe sampler provides authority and diverse initial postures as the reference, thus enabling us to demonstrate diverse gesture styles compared to other counterparts. Moreover, we find that the diversity scores on BEAT-ETrans are much higher than those on TED-ETrans dataset. This can be attributed to the more complex human joints in the BEAT-ETrans dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 507, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 686 + ], + "type": "text", + "content": "Ablation Study: To further verify the effectiveness of our proposed methods, we conduct the ablation study of different components as variations, reported in Table 3. The baseline model is implemented by a simple transformer-based pipeline with stacking three times blocks in the encoder-decoder. Obviously, all the combinations of our proposed components have positive impacts on the generated results. Specifically, by adding the motion transition infusion mechanism to the baseline, the indicator BC has achieved significant improvement (e.g., " + }, + { + "bbox": [ + 46, + 507, + 287, + 686 + ], + "type": "inline_equation", + "content": "0.701 \\rightarrow 0.866" + }, + { + "bbox": [ + 46, + 507, + 287, + 686 + ], + "type": "text", + "content": " in the BEAT-Etrans). This result verifies that our motion transition infusion mechanism effectively models the temporal correlation between the transitions w.r.t. head/tail gestures, thus leading to the generated results preserving rhythm coherency with given speech, globally." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Moreover, adopting the emotion mixture strategy ideally improves the performance of " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{trans}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " on both two" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 210, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 330 + ], + "type": "text", + "content": "datasets. This indicates that the learnable emotion mixture wights can provide effective weak supervision by leveraging the pre-trained pose-based emotion classifier. Besides, we have observed significant improvement in the performance of " + }, + { + "bbox": [ + 304, + 210, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD}_{h + t}" + }, + { + "bbox": [ + 304, + 210, + 545, + 330 + ], + "type": "text", + "content": " during this phase compared to the previous version. This aligns with our transformer backbone's emphasis on modeling the sequential temporal correlations as a whole. The better transition gestures encourage our model to maintain better temporal consistency, thereby the head/tail gestures achieve better results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 332, + 546, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 546, + 429 + ], + "type": "text", + "content": "Finally, after additionally employing the keyframe sampler to produce authority initial postures as the reference, our method obtains the best performance. Although the properties of BC and FGD on both datasets just have slightly improvement, the diversity realizes a noticeably better achievement (e.g., " + }, + { + "bbox": [ + 304, + 332, + 546, + 429 + ], + "type": "inline_equation", + "content": "112.80 \\rightarrow 124.93" + }, + { + "bbox": [ + 304, + 332, + 546, + 429 + ], + "type": "text", + "content": " in the BEAT-ETrans dataset). This highly supports our insight into keyframe-driven diversification strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 441, + 435, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 435, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 435, + 454 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "Visualization: To fully demonstrate the performance of our method, we show the visualized keyframes generated from ours compared with counterparts on our newly collected TED-ETrans and BEAT-ETrans datasets, respectively. As depicted in Figure 3, our method displays vivid and diverse results against others. In particular, we observe that Seq2Seq and Trimodal tend to synthesize unreasonable and stiff results (e.g., the blue rectangle in the right BEAT-ETrans dataset). Although CAMN and HA2G can generate natural upper-body postures, we find that they sometimes produce unreliable subtle fingers (e.g., the red rectangle in the left TED-ETrans dataset). Both our method and DiffGesture create reasonable gestures. However, the results synthesized by DiffuGesture are mismatched with the emotion transitions. In contrast, our method can synthesize the synchronous motions (e.g., in the BEAT-ETrans, the arms become droopy as emotion turns to sadness). Meanwhile, we further verify the diversification results as shown in Figure 1. Given the same input audio, our method generates diverse and vivid co-speech gestures. Please refer to the supplementary material for more visualization results." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10430" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 69, + 520, + 358 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 520, + 358 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 520, + 358 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 520, + 358 + ], + "type": "image", + "image_path": "76cca928db845133bbc5a0af9f746f3de94363a40259366ce6dc6f317dd9af4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 547, + 388 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 547, + 388 + ], + "type": "text", + "content": "Figure 3. Visualization of our generated 3D co-speech gestures against various state-of-the-art methods. The samples of the left part are from our newly collected TED-ETrans dataset, and the samples of the right part are from our BEAT-ETrans dataset. Best view on screen." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 289, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 595 + ], + "type": "text", + "content": "User Study: To further analyze the quality of results by various counterparts and ours, we conduct a user study by visualizing the results and inviting 15 subjects. For each emotion in both datasets, we randomly select 12 samples for each participant (7 emotions in BEAT-ETrans, 5 emotions in TED-ETrans). All the participants are recruited anonymously from schools including various majors. In particular, the subjects are required to rate the generated co-speech gestures from 0 to 5 (the higher, the better) in terms of naturalness, motion smoothness, and speech-gesture coherency. The results are demonstrated in Figure 4. Our method showcases the best performance compared with all the competitors. Especially in terms of motion smoothness, our method achieves noticeable advantages, indicating the effectiveness of our proposed motion transition infusion mechanism and the emotion mixture strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 608, + 119, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 119, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 119, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a new task of 3D co-speech gesture generation given emotion transition human speech. We therefore newly collected two datasets named the BEAT-ETrans and the TED-ETrans to fulfill this goal while significantly facilitating the research on 3D human motion modeling. Then, we fully take advantage of the sequential temporal correlation via a motion transition infusion mechanism" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 403, + 536, + 524 + ], + "blocks": [ + { + "bbox": [ + 317, + 403, + 536, + 524 + ], + "lines": [ + { + "bbox": [ + 317, + 403, + 536, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 403, + 536, + 524 + ], + "type": "image", + "image_path": "f33d9cd1a3b6d664abb605ba5c08d7a67132d8284afbbf4724235a8e1af55d33.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "content": "Figure 4. User study on gesture naturalness, motion smoothness, and speech-gesture coherency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 562, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 545, + 670 + ], + "type": "text", + "content": "to ensure the generated gestures preserve temporal coherence. Furthermore, we design an emotion mixture strategy to supply emotional weak supervision of the synthesized transitions. Extensive experiments conducted on our two newly collected datasets show the superiority of the method. As our method intends to generate diverse and vivid emotion transition gestures, we will investigate diversifying the 3D gesture with temporal smooth sampling, instead of the keyframe-wise manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 674, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 711 + ], + "type": "text", + "content": "Acknowledgements. This work is funded in part by the Theme-based Research Scheme grant (No.T45-205/21-N) and the InnoHK funding, Hong Kong SAR." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10431" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "text", + "content": "[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "text", + "content": "[3] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2027-2036, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "text", + "content": "[4] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 301, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 346 + ], + "type": "text", + "content": "[5] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "text", + "content": "[6] Joon Son Chung, Jaesung Huh, Seongkyu Mun, Minjae Lee, Hee Soo Heo, Soyeon Choe, Chiheon Ham, Sunghwan Jung, Bong-Jin Lee, and Icksang Han. In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[7] Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. arXiv preprint arXiv:2005.07143, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "text", + "content": "[8] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. arXiv preprint arXiv:2305.08953, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 513 + ], + "type": "text", + "content": "[9] Maged Farouk. Studying human robot interaction and its characteristics. International Journal of Computations, Information and Manufacturing (IJCIM), 2(1), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "text", + "content": "[10] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "text", + "content": "[11] Yu Fu, Yan Hu, and Veronica Sundstedt. A systematic literature review of virtual, augmented, and mixed reality game applications in healthcare. ACM Transactions on Computing for Healthcare (HEALTH), 3(2):1-27, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "text", + "content": "[12] Saeed Ghorbani, Ylva Fe2rstl, Daniel Holden, Nikolaus F Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. In Computer Graphics Forum, pages 206–216. Wiley Online Library, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[13] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Susan Goldin-Meadow and Martha Wagner Alibali. Gesture's role in speaking, learning, and creating language. Annual review of psychology, 64:257-283, 2013. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 174 + ], + "type": "text", + "content": "[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 209 + ], + "type": "text", + "content": "[16] Autumn B Hostetter and Martha W Alibali. Visible embodiment: Gestures as simulated action. Psychonomic bulletin & review, 15:495-514, 2008. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 266 + ], + "type": "text", + "content": "[17] Ann Huang, Pascal Knierim, Francesco Chiossi, Lewis L Chuang, and Robin Welsch. Proxemics for human-agent interaction in augmented reality. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-13, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 267, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 267, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 545, + 312 + ], + "type": "text", + "content": "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 314, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 314, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 314, + 545, + 346 + ], + "type": "text", + "content": "[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 348, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 381 + ], + "type": "text", + "content": "[20] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 384, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 471 + ], + "type": "text", + "content": "[21] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thórisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 473, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 473, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 473, + 545, + 507 + ], + "type": "text", + "content": "[22] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities for reactive robotic response. In IROS, page 2071. Tokyo, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 508, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 508, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 508, + 545, + 542 + ], + "type": "text", + "content": "[23] Sergey Levine, Philipp Krahenbuhl, Sebastian Thrun, and Vladlen Koltun. Gesture controllers. In Acm siggraph 2010 papers, pages 1-11. 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 587 + ], + "type": "text", + "content": "[24] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 589, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 655 + ], + "type": "text", + "content": "[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10432" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7590–7598, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "text", + "content": "[28] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "type": "text", + "content": "[29] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. AudioLDM: Text-to-audio generation with latent diffusion models. Proceedings of the International Conference on Machine Learning, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "text", + "content": "[30] Haohe Liu, Qiao Tian, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D. Plumbley. AudioLDM 2: Learning holistic audio generation with self-supervised pretraining. arXiv preprint arXiv:2308.05734, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "text", + "content": "[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via masked audio gesture modeling. arXiv preprint arXiv:2401.00374, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "type": "text", + "content": "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "text", + "content": "[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 481, + 226, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 226, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 226, + 491 + ], + "type": "text", + "content": "[34] OpenAI. Gpt-4 technical report, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 536 + ], + "type": "text", + "content": "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 580 + ], + "type": "text", + "content": "[36] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "text", + "content": "[37] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3d hand gesture prediction from body dynamics by bilateral hand disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4616-4626, 2023. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[39] Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, et al. Speechbrain: A general-purpose speech toolkit. arXiv preprint arXiv:2106.04624, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 126 + ], + "type": "text", + "content": "[40] Peter Roach. Some languages are spoken more quickly than others. 1998. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 128, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 545, + 182 + ], + "type": "text", + "content": "[41] Mehmet E Sargin, Yucel Yemez, Engin Erzin, and Ahmet M Tekalp. Analysis of head gesture and prosody patterns for prosody-driven head-gesture animation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(8):1330-1345, 2008. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 183, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 248 + ], + "type": "text", + "content": "[42] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "type": "text", + "content": "[43] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 371 + ], + "type": "text", + "content": "[45] Jordan Voas, Yili Wang, Qixing Huang, and Raymond Mooney. What is the best automated metric for text to motion generation? arXiv preprint arXiv:2309.10248, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "text", + "content": "[46] Zijian Wang, Xingqun Qi, Kun Yuan, and Muyi Sun. Self-supervised correlation mining network for person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7703-7712, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 491 + ], + "type": "text", + "content": "[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 5860–5868, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 492, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 548 + ], + "type": "text", + "content": "[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 612 + ], + "type": "text", + "content": "[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "text", + "content": "[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[51] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In International Cconference on Multimodal Interaction, pages 179-185. 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10433" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 183 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[53] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10434" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_content_list.json b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ec8613e3a16cf68fb4223c8246dda2f707a6fb5 --- /dev/null +++ b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_content_list.json @@ -0,0 +1,1683 @@ +[ + { + "type": "text", + "text": "What Do You See in Vehicle? Comprehensive Vision Solution for In-Vehicle Gaze Estimation", + "text_level": 1, + "bbox": [ + 176, + 130, + 794, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yihua Cheng $^{1}$ Yaning Zhu $^{2}$ Zongji Wang $^{3}$ Hongquan Hao $^{4}$ Yongwei Liu $^{4}$ Shiqing Cheng $^{4}$ Xi Wang $^{4}$ Hyung Jin Chang $^{1}$ $^{1}$ University of Birmingham $^{1}$ , Huazhong University of Science and Technology $^{2}$ $^{2}$ NIST, Chinese Academy of Sciences $^{3}$ , CalmCar $^{4}$ \ny.cheng.2@bham.ac.uk, h.j.chang@bham.ac.uk", + "bbox": [ + 158, + 202, + 810, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Driver's eye gaze holds a wealth of cognitive and intentional cues crucial for intelligent vehicles. Despite its significance, research on in-vehicle gaze estimation remains limited due to the scarcity of comprehensive and well-annotated datasets in real driving scenarios. In this paper, we present three novel elements to advance in-vehicle gaze research. Firstly, we introduce IVGaze, a pioneering dataset capturing in-vehicle gaze, collected from 125 subjects and covering a large range of gaze and head poses within vehicles. In this dataset, we propose a new vision-based solution for in-vehicle gaze collection, introducing a refined gaze target calibration method to tackle annotation challenges. Second, our research focuses on in-vehicle gaze estimation leveraging the IVGaze. In-vehicle face images often suffer from low resolution, prompting our introduction of a gaze pyramid transformer that leverages transformer-based multilevel features integration. Expanding upon this, we introduce the dual-stream gaze pyramid transformer (GazeDPTR). Employing perspective transformation, we rotate virtual cameras to normalize images, utilizing camera pose to merge normalized and original images for accurate gaze estimation. GazeDPTR shows state-of-the-art performance on the IVGaze dataset. Thirdly, we explore a novel strategy for gaze zone classification by extending the GazeDPTR. A foundational tri-plane and project gaze onto these planes are newly defined. Leveraging both positional features from the projection points and visual attributes from images, we achieve superior performance compared to relying solely on visual features, substantiating the advantage of gaze estimation. The project is available at https://yihua.zone/work/ivgaze.", + "bbox": [ + 75, + 359, + 473, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 843, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding driver intention and behavior based on driver gaze is in high demand in intelligent vehicles, facilitating", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3605938badceae03d6633d6fc61a398963c10a0f843780a50db3f42012ef2c82.jpg", + "image_caption": [ + "Figure 1. In-vehicle gaze estimation illustration. The driver's gaze direction is estimated based on the facial images captured by the camera behind the steering wheels." + ], + "image_footnote": [], + "bbox": [ + 511, + 325, + 880, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "diverse applications such as in-vehicle interaction [1, 2, 25] and driver monitor systems [17, 21, 22]. Recent advances in vehicle gaze estimation concentrate primarily on gaze zone estimation [16, 18, 31, 32]. These approaches define multiple coarse regions, such as side mirrors and windshields, and conduct classification based on face images.", + "bbox": [ + 496, + 526, + 892, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaze estimation serves as an upstream task of gaze zone estimation and can offer more precise information to understand driver attention. However, these methods typically require a large-scale dataset for training. Although there are numerous gaze datasets, collected in indoor [39, 40] or outdoor [20] environments, their applicability to the vehicle environment is limited due to the different environments and camera settings, resulting in suboptimal performance. Creating an in-vehicle gaze dataset proves challenging due to the confined and irregular nature of the vehicular environment. Constructing in-vehicle gaze collection systems remains an unsolved issue, as traditional gaze collection systems are impractical for use within vehicles. The absence of in-vehicle gaze datasets acts as a significant barrier to the progress of in-vehicle gaze estimation.", + "bbox": [ + 496, + 619, + 893, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a comprehensive vision-based", + "bbox": [ + 517, + 845, + 892, + 861 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our work focuses on gaze direction estimation. We abbreviate gaze direction as gaze in the rest.", + "bbox": [ + 500, + 875, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1556", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in-vehicle gaze estimation research: a novel vision-based gaze collection system for vehicles, offering a first-of-its-kind in-vehicle gaze dataset, a dual-stream gaze pyramid transformer for accurate in-vehicle gaze estimation, and its extension to gaze zone classification, showcasing its effectiveness in enhancing gaze estimation.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we introduce IVGaze, an in-vehicle gaze dataset collected from 125 subjects. IVGaze provides a dense distribution of gaze directions, covering a wide range within in-vehicle environments. It contains various conditions, including diverse head poses, eye movements, illumination variations, and the presence of face accessories such as glasses, sunglasses, and masks. To collect IVGaze, we propose a vision-based gaze collection system. The system does not require dedicated eye-tracking devices and is easy to reproduce. We use a single camera for facial appearance capture and paste stickers in vehicles as gaze targets. However, a significant challenge lies in calibrating the 3D position of gaze targets. This challenge arises because gaze targets are out of the camera's field of view (out-of-FoV). To address this issue, we present a refined gaze target calibration method, which leverages an auxiliary camera to capture gaze targets and a transparent chessboard for calibration.", + "bbox": [ + 75, + 181, + 468, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, we explore in-vehicle gaze estimation using the IVGaze. Face images often suffer from low resolution due to the inherent limitations of cameras in vehicles. We propose a gaze pyramid transformer that utilizes a transformer to integrate multilevel features. Expanding upon this, we propose a dual-stream gaze pyramid transformer (GazeDPTR). We rotate virtual cameras via perspective transformation to normalize images, and leverage camera pose to merge normalized and original images. GazeDPTR shows state-of-the-art results on IVGaze.", + "bbox": [ + 75, + 440, + 468, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thirdly, we extend GazeDPTR for the downstream gaze zone classification task. It is challenging to compute the intersection of gaze and the vehicle. We define a foundational tri-plane and project gaze to the tri-plane. We extract positional features from intersection points and predict gaze zones using both positional features and visual features from face images. Our experiment demonstrates that the gaze zone classification can be further enhanced by positional features, showing the advantage of gaze estimation.", + "bbox": [ + 75, + 592, + 468, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 744, + 225, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Gaze Data Collection", + "text_level": 1, + "bbox": [ + 76, + 770, + 276, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The human gaze is inherently implicit and poses a challenge for objective measurement, making gaze annotation complex in gaze data collection. Some methods capture the human gaze through intrusive devices such as eye-tracking glasses [13, 19]. However, the eye-tracking glasses have a notable impact on the quality of the captured facial images. Fischer et al. [13] attempted to mitigate this impact", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/cdfd582c673589eecd0a48900355166ea67a4828b6778d8aac64c6cd26a51fc6.jpg", + "table_caption": [ + "Table 1. The comparison of gaze estimation datasets. Gaze annotation is challenging in the vehicle environment, which results in existing in-vehicle datasets only providing gaze zone annotation. Our work addresses this issue and contributes the first in-vehicle gaze dataset containing gaze annotation and natural face images." + ], + "table_footnote": [], + "table_body": "
DatasetsEnv.# Sub.Annotation
GazeZone
Gaze360 [20]Outdoor238✓-
MPIIGaze [39]15✓-
EyeDiap [15]Indoor16✓-
EVE [27]54✓-
ETH-XGaze [40]110✓-
Vora et al. [31]10✗✓
Jha & Busso [18]16✗✓
Wang et al. [32]Vehicle3✗✓
Rangesh et al. [28]13✗✓
Ghosh et al. [16]338✗✓
IVGaze (Ours)125✓✓
", + "bbox": [ + 501, + 164, + 888, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "using GAN, but there are also some artifacts in the resulting face images. Vision-based gaze collection systems typically define gaze as direction vectors originating from facial centers towards specific gaze targets [27, 39, 40]. However, these gaze targets often remain outside the camera's field of view. This out-of-field view challenge complicates the calibration of gaze target positions. Zhang et al. [39, 40] sets screen points as gaze targets and uses a mirror to calibrate the screen plane. Kellnhofer et al. [20] uses $360^{\\circ}$ panoramic cameras to capture gaze targets and human faces simultaneously where the panoramic camera is pre-calibrated. However, these strategies are not applicable to vehicles.", + "bbox": [ + 496, + 383, + 890, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In-vehicle gaze dataset usually defines different region such as windshield and left/right mirror in vehicles and perform gaze zone classification [12, 14, 16, 18, 23, 30, 34]. Kasahara et al. [19] collects an in-vehicle gaze dataset but subjects are required to wear eye-tracking glasses, which means the dataset is not applicable in the real world. Our gaze collection system does not require dedicated devices and produces natural face images.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Appearance-based Gaze Estimation", + "text_level": 1, + "bbox": [ + 500, + 696, + 808, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Appearance-based gaze estimation directly learns mapping function from facial appearance to human gaze [10]. Conventional gaze estimation methods extract eye features from eye images [7, 9, 26, 36]. They concatenate the head pose vector with eye features for gaze estimation. Recent methods directly learn gaze from face images [3, 4, 11, 33, 35]. Face images provide both eye region and head pose information, resulting in superior performance compared to methods relying solely on eye images. However, the subtlety of the human eye in face images poses a challenge as it can be easily overlooked by the network. Zhang et al. [37] utilize a learnable attention map to guide the net", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1557", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05326171b8c2d985f03644cbebb4e804555a48e2d148099086e68565d0a34d94.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 462, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/53a7a6a1e467b62214bd05f80e3fbe76bf19fc501c5ba6588b97fdfc0dc42a7d.jpg", + "image_caption": [ + "(b)", + "(c)", + "(d)", + "Figure 2. We construct a vision-based in-vehicle gaze collection system comprising a DMS camera, a depth camera, and strategically placed gaze targets, as depicted in (c). The DMS camera is positioned behind the steering wheel to capture drivers' facial appearances, while the gaze targets, positioned beyond the DMS camera's field-of-view (FoV), such as on the windshield, remain unobserved. The depth camera, utilized for calibration purposes, is temporarily installed for capturing gaze target positions in 3D with respect to its own coordinates, and it is removed during data collection. To facilitate the calibration of the depth camera's pose relative to the DMS camera, we propose employing a transparent chessboard, which is placed between the two cameras." + ], + "image_footnote": [], + "bbox": [ + 464, + 88, + 890, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work's focus specifically on the eye. Cheng et al. [11] crop eye images from face images and construct a cascade network to leverage them. In recent developments, transformers have showcased remarkable capabilities in gaze estimation. Cheng et al. [5] demonstrate that transformers can achieve state-of-the-art performance across various benchmarks. They apply transformers to address dual-view gaze estimation and show that dual-view images outperform single-view images [6].", + "bbox": [ + 75, + 383, + 472, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. In-Vehicle Gaze Data Collection System", + "text_level": 1, + "bbox": [ + 76, + 536, + 433, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our vision-based system for in-vehicle gaze data collection, eliminating the need for dedicated eye-tracking devices. Our system includes a neat gaze target calibration method that effectively addresses the core challenge in gaze annotation. Leveraging this system, we collect the first in-vehicle gaze dataset from 125 subjects. The dataset spans a diverse range of gaze, facilitating and advancing future research in in-vehicle gaze estimation.", + "bbox": [ + 75, + 561, + 468, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. System Setup", + "text_level": 1, + "bbox": [ + 76, + 694, + 218, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We place gaze targets within the vehicle, and subjects are instructed to look at each designated target during the data collection process. Simultaneously, we record their facial appearance along with corresponding gaze targets.", + "bbox": [ + 75, + 718, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera. Our system uses one camera of a driver monitor system (DMS) for facial appearance capture. The DMS camera is an infrared camera with a capture resolution of $1280 \\times 800$ . The camera is located behind the steering wheel and directly points at the head region of drivers.", + "bbox": [ + 75, + 779, + 470, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaze targets. We mark gaze targets with red points on stickers, and each target is uniquely labeled with a distinct number printed on stickers to aid differentiation. We strate-", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gically position these stickers within various gaze regions within the vehicle, including the windshield, left and right-side mirrors, rear-view mirror, center console, speedometer, handbrake, and etc. These targets cover a large gaze region, satisfying the gaze estimation requirements in vehicles.", + "bbox": [ + 496, + 383, + 892, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Collection Procedure", + "text_level": 1, + "bbox": [ + 500, + 469, + 697, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We meticulously design the collection procedure to minimize error. During the collection process, participants are instructed to look at specific gaze targets in a predefined sequence. They are also required to speak the corresponding number associated with each target to confirm their focus accuracy. Participants must maintain focus on the target for 3 seconds to facilitate image capture. We conduct a preliminary check on the captured images and discard any erroneous ones. Once this verification is complete, the current data collection phase concludes, and participants are instructed to shift their gaze to the next gaze targets.", + "bbox": [ + 496, + 492, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design three postures for each participant. Participants should complete the collection process three times with different postures. The first posture requires participants not to change head pose. They should preserve the same head pose and only move eyeballs to focus a gaze target. We aim to collect sufficient eye movement data with this posture. We do not require participants to focus on the target where participants think it is hard to focus with only eye movement, e.g., the target in the side mirror. Participants can rotate their heads to look at these points but they are also required to preserve the new head pose until they cannot look at any targets with the current head pose. The second posture requires participants not to move their head position but they can rotate their heads to focus a gaze target. We do not set any constraints in the third posture. Participants can freely perform head movements in this pos", + "bbox": [ + 496, + 659, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1558", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2870d1c47d06ce53cf6858d51f7ccad31f2f5d7b3058ed26afe593d8380b2b4f.jpg", + "image_caption": [ + "Figure 3. Our dataset is collected using IR cameras in the vehicle environment. (a) We present image samples of IVGaze, highlighting the challenges posed by realistic in-vehicle conditions, including cases with sunglasses and reflections in glasses. (b) We categorize the image count based on their mean pixel value, showing the diversity of illumination conditions. (c) The image count is analyzed based on face accessories including glasses, sunglasses, and masks." + ], + "image_footnote": [], + "bbox": [ + 84, + 104, + 370, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a885f36b2846195ca82470208a05578eb83ec9da22241f7d01ef4fb96e2a0360.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 89, + 704, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c25bef82c6cf83982cf40d0a67ec6db2599fa66e99b69270499d00a6b301bf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 90, + 929, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture. The last two postures bring head pose variation for IVGaze. We do not use devices such as headrests to constrain participants. Participants only need to preserve the specific posture and focus on targets in a comfortable way.", + "bbox": [ + 75, + 325, + 468, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Gaze Annotation", + "text_level": 1, + "bbox": [ + 76, + 398, + 243, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-based gaze collection systems define gaze as unit direction vectors originating from the face center $\\mathbf{o}$ and extending toward gaze targets $\\mathbf{t}$ [10]. Our system follows this framework and the gaze annotation can be computed as $\\mathbf{g} = (\\mathbf{t} - \\mathbf{o}) / ||\\mathbf{t} - \\mathbf{o}||_2$ . The gaze annotation is decomposed into 3D face center estimation and gaze target calibration.", + "bbox": [ + 75, + 422, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D face center estimation has been effectively addressed in previous research. We first detect facial landmarks and fit a 3D morphable face model for 3D facial landmarks [24, 40]. We select the 3D position of the nose as gaze origin o. However, the calibration of gaze target positions presents a challenge because gaze targets are out-of-FoV. It means that the DMS camera cannot capture any images of the gaze targets, complicating the calibration problem.", + "bbox": [ + 75, + 515, + 468, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Gaze Target Calibration", + "text_level": 1, + "bbox": [ + 76, + 648, + 302, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We need the 3D gaze target position $\\mathbf{t}$ for gaze annotation. However, a key problem is all gaze targets are out-of-FoV. Conventional gaze datasets often present gaze targets on a screen and employ a mirror to solve the out-of-FoV problem, where the mirror reflects the content of the screen [36]. However, these methods are not suitable for vehicle scenarios and can lead to a substantial accumulation of errors.", + "bbox": [ + 75, + 672, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we propose a neat and efficient gaze target calibration method. Our basic idea is to employ an auxiliary camera to capture gaze targets. This allows a straightforward calculation of the 3D gaze target positions w.r.t. the auxiliary camera coordinate system. However, it also introduces a new challenge: how to calibrate extrinsic matrix of the auxiliary cameras w.r.t. DMS camera? Stereo calibration is typically used to compute the pose of two cam", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eras. However, it requires two captured images sharing corresponding points. In our system, gaze targets are located behind the DMS camera which means the DMS camera and auxiliary camera should be oriented in opposite directions, making traditional stereo calibration infeasible.", + "bbox": [ + 496, + 325, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve this problem, we propose to use a transparent chessboard for the pose calibration. We set a transparent chessboard between the two cameras and the two cameras respectively capture each side of the chessboard. We first calibrate the two camera poses w.r.t. the chessboard coordinate system and then compute the pose between two cameras. It is worth noting that there are two different chessboard coordinate systems corresponding to each side of the chessboard. We can derive the transformation matrix between the two chessboard coordinate systems where $\\mathbf{R}_{\\mathrm{chess}} = \\mathrm{diag}(0,0,-1)$ and $\\mathbf{t}_{\\mathrm{chess}} = (0,0,-\\mathrm{d})$ . The d is the thickness of the chessboard.", + "bbox": [ + 496, + 402, + 892, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In detail, we use an Intel RealSense D435 depth camera as the auxiliary camera since it can provide accurate 3D positions. We show the camera layout in Fig. 2(b). We can obtain the transformation matrix $\\mathbf{R}_{\\mathrm{dms}}$ , $\\mathbf{t}_{\\mathrm{dms}}$ and $\\mathbf{R}_{\\mathrm{depth}}$ , $\\mathbf{t}_{\\mathrm{depth}}$ which can transfer a point from chessboard coordinate systems to camera coordinate systems. We have", + "bbox": [ + 496, + 583, + 890, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {\\mathrm {r o t}} = \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {R} _ {\\mathrm {c h e s s}} \\mathbf {R} _ {\\mathrm {d e p t h}} ^ {- 1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 684, + 890, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {\\text {r o t}} = - \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {R} _ {\\text {c h e s s}} \\mathbf {R} _ {\\text {d e p t h}} ^ {- 1} \\mathbf {t} _ {\\text {d e p t h}} + \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {t} _ {\\text {c h e s s}} + \\mathbf {t} _ {\\mathrm {d m s}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 715, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use these matrices to convert points from the depth camera coordinate system into the DMS camera coordinate system. Please refer to the supplementary material for details.", + "bbox": [ + 496, + 747, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. In-Vehicle Gaze Dataset", + "text_level": 1, + "bbox": [ + 500, + 801, + 720, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect the first in-vehicle gaze dataset IVGaze which provides dense gaze annotation and natural face images. We show the dataset samples in Fig. 3 (a).", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Statistics. We collect 44,705 images from 125 subjects. The number of images per subject ranges from", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1559", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ac84c8c9ac9218a7dd53f0a820ebc628d22089999853487859601635fcce461.jpg", + "image_caption": [ + "Figure 4. We show the distribution of data for gaze (left) and head movements (right). Brighter regions denote higher data density." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 269, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2203024d1579751e1ce55940485aa68ef0c9a015915d53d07494b80feb46ddc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 88, + 450, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "169 to 946. Most subjects provide around 300 images per person, i.e., they repeat the collection procedure three times with three different postures. 40 subjects are female and 85 subjects are male, with ages spanning from 20 to 50 years.", + "bbox": [ + 76, + 270, + 467, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collection Conditions. The dataset was collected between 9 am and $7\\mathrm{pm}$ in outdoor environments, covering a wide range of lighting conditions. We also collect images in indoor environments, i.e., a garage. The dataset statistics across pixel intensity are shown in Fig. 3 (b).", + "bbox": [ + 76, + 332, + 467, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face Accessories. We consider two face accessories during the collection: glasses and masks. For subjects who did not wear glasses, we provided glasses and asked them to repeat the collection wearing glasses. We provided white and black masks for some subjects and asked them to repeat the collection wearing masks. We also required a few subjects to wear sunglasses to facilitate future research. We show the statistics in Fig. 3 (c).", + "bbox": [ + 75, + 411, + 467, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Distribution. We plot the angular distribution of gaze and head pose. Note that the data normalization method (introduced in the next section) changes gaze directions and head poses with a rotation matrix. Fig. 4 shows the distribution in normalization space. The horizontal gaze is from $-50^{\\circ}$ to $90^{\\circ}$ , and the vertical gaze is from $-40^{\\circ}$ to $40^{\\circ}$ . It is worth noting that the vehicle environment has limited space, and our dataset already covers almost all regions. Our dataset is collected in the real vehicle environment, and the driver seat in the vehicle is located on the left. Therefore, our dataset has a relatively small distribution on the left. The right figure in Fig. 4 shows the head pose distribution. Our dataset contains a large range of head poses in the yaw axis, as subjects need to look at gaze targets from the left-side mirror to the right-side mirror.", + "bbox": [ + 75, + 534, + 467, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. In-Vehicle Gaze Estimation", + "text_level": 1, + "bbox": [ + 76, + 782, + 331, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The in-vehicle environment brings new settings and challenges. In this section, we systematically explore in-vehicle gaze estimation. We introduce a revised data preprocessing method and describe a novel gaze estimation network. We also extend the network for an application and prove the advantage of in-vehicle gaze estimation.", + "bbox": [ + 76, + 810, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Revisiting Data Normalization in Vehicle", + "text_level": 1, + "bbox": [ + 500, + 90, + 846, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data normalization rotates and translates cameras via perspective transformation in face images to reduce head pose variation. Conventional methods usually define the rotated camera coordinate system $\\mathbf{C}_r$ based on head pose [29, 38]. However, these methods cannot bring performance improvement in the vehicle environment.", + "bbox": [ + 498, + 113, + 890, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $x$ -axis of $\\mathbf{C}_r$ is typically defined as the $x$ -axis of head coordinate systems. It means they will rotate images to keep the head straight. We observed that the designed $x$ -axis will produce unstable results, especially in the extreme head pose. Therefore, we propose that do not rotate virtual cameras based on the $x$ -axis of head coordinate systems. In detail, we compute the $z$ -axis of $\\mathbf{C}_r$ as direction vectors from cameras to face centers. We use the $x$ -axis of the virtual camera, i.e., $(1,0,0)$ rather than the head coordinate system. We compute the $y$ -axis as $y = z\\times x$ and also recompute the $x$ -axis as $x = y\\times z$ . Finally, we have a rotation matrix $\\mathbf{R} = [x;y;z]$ and use a scale matrix $S$ to maintain the user-camera distance. We warp the image based on matrix SR. The human gaze is also changed as $\\mathbf{g}^n = \\mathbf{R}\\mathbf{g}^o$ , where $\\mathbf{g}^n$ and $\\mathbf{g}^o$ denotes gaze in the normalization and origin space.", + "bbox": [ + 498, + 204, + 892, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Gaze Pyramid Transformer", + "text_level": 1, + "bbox": [ + 500, + 454, + 748, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaze estimation methods typically learn gaze from face images. However, face images captured in vehicles often suffer from low resolution owing to the limited capabilities of the camera. This low resolution can lead conventional methods to overlook subtle details, especially in the eye region. Some techniques address this challenge by cropping eye images and processing them separately [8], but these approaches encounter difficulties when faced with extreme head poses, as one of the eye images may not be visible.", + "bbox": [ + 498, + 477, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a gaze pyramid transformer (GazePTR) in this section. We build a feature pyramid and input multilevel features into a transformer for gaze estimation. As shown in Fig. 5, our network uses a convolution network to extract features from facial images. We collect feature maps from different levels rather than the last level. We use $1 \\times 1$ convolution layers and global average pooling layers to preserve the same feature dimension. Finally, we input the multi-level feature into a transformer. A learnable token is used to aggregate multi-level features for gaze estimation.", + "bbox": [ + 498, + 614, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the network has the same performance achieved by directly utilizing the last feature map for gaze estimation. This result can be attributed to the network's tendency to overlook low-level features. To address this issue, we introduce additional constraints by estimating gaze from each level of features and calculating corresponding loss. It can be understood that we first extract multi-level features, with each feature being potentially useful. A transformer is used to effectively ensemble these features.", + "bbox": [ + 498, + 765, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1560", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65ba7fefcea69103c4bd47f4963e9bfb6bc14be0a0e5033533b2547bc9d1c9c4.jpg", + "image_caption": [ + "Figure 5. The GazeDPTR directly crop face for origin images and rotates virtual cameras via perspective transformation for normalized images. It builds a dual-stream network to extract features from the two images based on the GazePTR for feature extraction which integrates multi-level features via transformers. To further merge the features from two streams, we leverage a transformer where camera pose is used as the positional feature in the transformer. We define the original camera pose as $\\mathbf{C} = \\text{diag}(1, 1, 1)$ and the camera pose in normalization space is RC. We also extend the network for gaze zone classification. We define a tri-plane and project gaze into them. We extract positional features from three intersection points via a transformer. We also extract visual features from images and predict the gaze zone based on both visual features and positional features. The whole network is trained in an end-to-end manner." + ], + "image_footnote": [], + "bbox": [ + 81, + 94, + 893, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Dual-Stream Gaze Pyramid Transformer", + "text_level": 1, + "bbox": [ + 76, + 390, + 429, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dual-view images have been demonstrated to outperform single-view images in gaze estimation [6]. In our work, we rotate cameras in data normalization through perspective transformation. This inspired us to formulate the dual-camera setting leveraging both original and normalized images. However, dual-view images improve accuracy since they contain more visual information. It remains uncertain whether the combination of normalized and original images can provide additional insights beyond what each offers.", + "bbox": [ + 75, + 412, + 468, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate our hypothesis, we executed an oracle baseline. We separately train GazePTR on the original and the normalized datasets, selecting the best result from each image pair. The result reveals a significant performance improvement, validating the inherent advantages of utilizing both images. Please refer to the supplementary for details.", + "bbox": [ + 75, + 549, + 468, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, we propose a dual-stream gaze pyramid transformer (GazeDPTR). Our method leverages both normalized images and original images for gaze estimation. We employ GazePTR to extract features from the two images and a transformer is used to integrate the two features for gaze estimation. Note that, the two features are in different camera coordinate systems and correspond to different gaze. To establish the connection, we use the camera pose as the positional information in the transformer. However, it is impossible the calibrate a virtual camera. We define the camera pose of original images as $\\mathbf{C} = \\text{diag}(1, 1, 1)$ and the camera pose of normalized images is RC. We extract the $z$ -axis and use positional encoding for camera pose [6].", + "bbox": [ + 75, + 641, + 470, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Strategy for Gaze Zone Classification", + "text_level": 1, + "bbox": [ + 76, + 845, + 400, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detecting driver attention is essential in driver monitoring systems. In this section, we introduce a novel strategy to", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "extend GazeDPTR for gaze zone classification. GazeDPTR estimates gaze direction from face images. While the conventional solution involves projecting gaze direction and computing intersections within the vehicle, it is not practical without a 3D vehicle model. Zhang et al. use a triplane for self-supervised loss function [35]. Our alternative strategy defines a foundational tri-plane and computes the intersection of gaze vectors with this tri-plane. We extract positional features from three intersection points to facilitate gaze zone classification.", + "bbox": [ + 496, + 390, + 890, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In detail, we define a tri-plane with normal vectors $(1,0,0)$ , $(0,1,0)$ , and $(0,0,1)$ , intersecting at the origin $(0,0,0)$ . Given the estimated gaze vector $\\mathbf{g}^n$ , we convert it back to the original space using $\\mathbf{g}^o = \\mathbf{R}^{-1}\\mathbf{g}^n$ . We then project $\\mathbf{g}^o$ onto the tri-plane with the face center $\\mathbf{o}$ and obtain three intersection points. To enhance the positional features, we apply positional encoding to these points and input them into a 2-layer transformer. The positional encoding not only increases feature dimensions but also normalizes them within the range of -1 to 1. Additionally, we use another learnable token to aggregate visual features from the original images. Gaze zone classification is performed using both visual features and positional features.", + "bbox": [ + 496, + 542, + 892, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the whole network in an end-to-end manner. We use L1 loss for gaze estimation and cross-entropy loss for gaze zone classification. Please refer to the supplementary for implementation details.", + "bbox": [ + 496, + 739, + 890, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 500, + 814, + 624, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: IVGaze contains 44,705 images of 125 subjects. To perform within-dataset evaluation, we divide the dataset into three subsets based on subjects. The image numbers of the three subsets are 15,165, 14,674, and 14,866. We", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1561", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8bb4fea72daa82d1c9c30fb59aca1ada6f5ea3b55c1da09b1b7aa1c3f064a2a5.jpg", + "table_caption": [ + "Table 2. Performance comparison in gaze estimation. Our methods achieve better performance than the compared methods." + ], + "table_footnote": [], + "table_body": "
Angular ErrorAverage Precision (AP)
<2°<4°<6°<8°
FullFace [37]13.67°2.3%8.8%17.8%28%
DWG [16]8.82°6.6%21.7%38.1%53.2%
Gaze360 [20]8.15°9.2%27.3%44.6%58.9%
FullFace+ [37]7.48°14.2%31.1%46.7%63.1%
GazeTR [5]7.33°17%32.8%47.5%64.7%
XGaze [40]7.06°11.7%32.7%51.5%66.7%
GazePTR7.04°17.6%34%49.3%66.7%
GazeDPTR6.71°22.1%36%50.3%68.4%
", + "bbox": [ + 76, + 121, + 475, + 300 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "perform three-fold cross-validation on our dataset.", + "bbox": [ + 75, + 319, + 410, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metric: We use the angular error as the evaluation metric of gaze estimation as most of the methods [10], where a smaller value represents a better model. However, we notice the angular error only shows the overall performance while cannot give deep insights. Therefore, we define the average precision (AP) where AP of $< k^{\\circ}$ means an estimation is considered correct if the angular error is lower than $k^{\\circ}$ . Regarding the gaze zone classification, average precision is used as a common multi-class classification.", + "bbox": [ + 75, + 335, + 468, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 76, + 479, + 370, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compared our methods with SOTA methods in the in-vehicle gaze estimation task. We compared our methods with FullFace [37], Gaze360 [20], XGaze [40], DWG [16] and GazeTR [5]. We replaced the backbone of the FullFace method from AlexNet to ResNet18 for a more convincing comparison. We denote the new method as FullFace+.", + "bbox": [ + 75, + 503, + 468, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The result is shown in Tab. 2. GazePTR and GazeDPTR both show better performance than the compared methods. GazePTR and GazeTR have the same backbone and transformer architecture. However, GazePTR outperforms $0.29^{\\circ}$ thanGazeTR since GazePTR uses multi-level feature maps. GazeDPTR builds a dual-stream network and further brings $0.33^{\\circ}$ improvement than GazePTR. These results show the advantage of our methods.", + "bbox": [ + 75, + 595, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interestingly, XGaze has a similar angular error to GazePTR. However, it is significantly worse than GazePTR in AP of $< 2^{\\circ}$ . It is because that XGaze uses ResNet50 as the backbone. The deep backbone enhances the feature extraction ability but also easily overlooks the small eye region, making it challenging to achieve precise gaze estimation.", + "bbox": [ + 75, + 715, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. The Impact of Face Accessories", + "text_level": 1, + "bbox": [ + 76, + 816, + 354, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "IVGaze provides rich samples of wearing face accessories, including glasses, masks, and sunglasses. We also evaluate their impacts on the accuracy of gaze estimation and show in Tab. 3. The result shows that eyeglasses have a", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/14744fc076aec7d025c13b12c6d6ec8f5883998b5f161a3ac42a3e6d3c6f61a2.jpg", + "table_caption": [ + "Table 3. We show the impacts of different face accessories on performance. The sunglasses bring a significant performance drop." + ], + "table_footnote": [], + "table_body": "
GlassesMaskSunglasses
withw/owithw/owith
FullFace [37]14.43°12.40°15.20°13.35°21.39°
DWG [16]9.20°8.19°9.43°8.69°17.43°
Gaze360[20]8.30°7.91°8.95°7.99°17.99°
FullFace+ [37]7.59°7.30°8.37°7.30°16.50°
XGaze [40]7.07°7.03°7.80°6.90°15.15°
GazeTR [5]7.40°7.22°8.12°7.17°17.49°
GazePTR7.13°6.90°7.78°6.89°16.54°
GazeDPTR6.77°6.63°7.44°6.57°16.41°
", + "bbox": [ + 501, + 121, + 903, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f791c54df84bfd40997e2f380edc5b5e5dc237cc0175e878e26be2b5d4d4fd85.jpg", + "image_caption": [ + "Figure 6. We compute the angular degree between gaze direction/head orientation with the frontal direction, i.e., $(0,0,-1)$ , and count the mean gaze estimation accuracy in different ranges. Interestingly, GazePTR brings significant performance improvement to GazeTR in the $0^{\\circ} - 20^{\\circ}$ gaze range. This is because subjects prefer to move their eyeball rather than their head when the gaze target is located at a close range. GazePTR utilizes multilevel features and can accurately capture eye movement." + ], + "image_footnote": [], + "bbox": [ + 504, + 311, + 696, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77c89d723cf5f7a5e4c5b674946d7f369db0b1838b8629bf259fccea5abfa316.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 311, + 890, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "large impact if the basic performance is not good. However, as the basic performance improves, the performance difference reduces. This is because most of glasses are transparent, enabling a robust model to extract features effectively. Sunglasses have a significant impact on performance, with average performance exceeding $15^{\\circ}$ . Masks have a substantial impact on performance, as they obscure facial regions, complicating the extraction of head-related features.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance Distribution", + "text_level": 1, + "bbox": [ + 500, + 696, + 733, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the performance distribution to provide deep insight. We first show the performance distribution in different gaze ranges. We introduce a novel concept, the gaze range of $a^{\\circ} - b^{\\circ}$ , which is defined as a set of samples satisfying the requirement that the angular degree between the gaze direction and the frontal gaze direction is in the range. The result is shown in Fig. 6. Interestingly, GazePTR and GazeDPTR have better performance in $0^{\\circ} - 20^{\\circ}$ . This is because subjects prefer to move their eye rather than head to look at objects in this range. Our methods leverage multi-level features and easily capture eye movement. We also show the performance distribution in different head ranges.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1562", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6629b29aa078134e28edec39f1105508bd8f8f18613995251f6bcf99aae64e6d.jpg", + "table_caption": [ + "Table 4. We present the performance of GazePTR at each level of the feature. The higher-level feature achieves better performance. Our method integrates features from all levels and achieves the best performance." + ], + "table_footnote": [], + "table_body": "
1st2nd3rd4thOurs
GazePTR10.33°8.59°7.61°7.35°7.04°
", + "bbox": [ + 83, + 148, + 460, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e8e441645049396b043793ff1c22d448516eb9d2dd73a6c1c7f0b8e7b1b5ba2.jpg", + "table_caption": [ + "Table 5. We present the performance of GazePTR separately training on the original images and normalized images. GazeDPTR leverages camera pose to establish a connection between the two images and use both two images for gaze estimation. We also removed the camera pose to perform an ablation study." + ], + "table_footnote": [], + "table_body": "
Original imagesNormalized imagesGazeDPTR (w/o camera pose)GazeDPTR
Acc7.44°7.04°7.03°6.71°
", + "bbox": [ + 81, + 273, + 467, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our methods are robust in different head ranges where the maximum performance difference is only $1^{\\circ}$ .", + "bbox": [ + 76, + 330, + 468, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 372, + 230, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GazePTR leverages multi-level features for gaze estimation. We first conduct an ablation study to demonstrate the advantage. We obtain the gaze performance of GazePTR at each level of feature and show it in Tab. 4. The result demonstrates a higher-level feature achieves better performance. GazePTR integrates multi-level features and achieves the best performance. The result proves the advantage of our design.", + "bbox": [ + 75, + 395, + 468, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GazeDPTR leverages both original and normalized images along with their corresponding camera poses. To demonstrate the effectiveness of our design, we perform an ablation study in Tab. 5. Training GazePTR separately on original and normalized images yields performances of $7.44^{\\circ}$ and $7.04^{\\circ}$ respectively. GazeDPTR, which incorporates both image types, achieves a performance improvement of $0.34^{\\circ}$ . We also experimented by excluding the positional encoding of the camera pose in GazeDPTR. Interestingly, the result is comparable to using only normalized images.", + "bbox": [ + 75, + 516, + 470, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Additional Experiments", + "text_level": 1, + "bbox": [ + 76, + 676, + 299, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Normalization. We compare our method with the data normalization method [38] in the left of Tab. 6. The previous method cannot work well in vehicle environments while our method brings $0.3^{\\circ}$ performance improvement.", + "bbox": [ + 75, + 700, + 468, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Case Study. We visual gt and our prediction in Fig. 7.", + "bbox": [ + 76, + 761, + 434, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Gaze Zone Classification", + "text_level": 1, + "bbox": [ + 76, + 786, + 303, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a basis tri-plane to acquire positional feature and combine both positional feature and visual features for gaze zone classification. We respectively evaluate each feature and show the result in the right of Tab. 6. The visual feature achieves better performance than the positional feature since the positional feature is obtained from gaze pro", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3fdac2ee66c01b82a04a81324af32814a284a38cf3099f4c291633f325028843.jpg", + "table_caption": [ + "Table 6. We evaluate different data normalization methods in the left table. Our method outperforms the previous method [38]. In the right table, we evaluate different features in the gaze zone classification. Our method uses both positional feature and visual features, achieving $2.4\\%$ improvement over visual features." + ], + "table_footnote": [], + "table_body": "
Normalization [38]oursGazeTRGazePTRVisual FeaturePositional FeatureAP
××7.77°7.44°✓79.4%
✓8.64°8.53°✓75.3%
✓7.33°7.04°✓✓81.8%
", + "bbox": [ + 504, + 166, + 888, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c74a154f013274f2bc0b275fa40e5817ef46509a6250cf9368aed0da587aca0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 250, + 599, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5deb96d370044fb41499b0012f948c4a7593447cedbf57c49d285d762735f14a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 250, + 691, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1688cd8b86a1f5c59544da0c7e3e9af73fe7a36d888b5a3eef616d3ff49b11e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 250, + 784, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e31a932506a0c0240de6a0e51f540b0d4a606164062623b93a4cf8417c1e09eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 250, + 875, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9647d81b316851b2ef573ca3f6839f399e2460907b3eed813c173f8d35ade647.jpg", + "image_caption": [ + "Figure 7. We illustrate GT with green lines and Prediction with red lines. The number shows the angular error in degree. The right two figures are failure cases since the eye is hard to capture." + ], + "image_footnote": [], + "bbox": [ + 511, + 319, + 599, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50a223a2104ebf1e8a768ca9c88c872432c8e42ad342478cccb04a0b585a34f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 319, + 691, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e27d54d41d9c2cabf8a652684d9e86a5f9c4b8ef67f66e838e929c33a6fa38e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 319, + 784, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4488432ef6d3d8cd112629dd8c585833d8c2679102c30b9459937fb21ddaf9ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 319, + 875, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "jection. The projection increases the model interpretability but decreases the fitting ability. Our method uses the two features and gets $2.4\\%$ performance improvement. The result demonstrates that the gaze zone classification task could be enhanced by gaze direction cues, highlighting the advantage of our gaze estimation work.", + "bbox": [ + 496, + 448, + 890, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 549, + 619, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we provide systematical research in in-vehicle gaze estimation, including dataset, algorithm, and extensive application. We first solve the data collection issue in a vehicle where a gaze target calibration method is proposed. We collect the first in-vehicle gaze dataset containing dense gaze annotation and natural face images. We further explore the algorithm for in-vehicle gaze estimation. Our work brings two deep insights 1) multi-level feature is useful to capture eye region information. 2) simultaneously leveraging original images and normalized images could achieve better performance. We also extend our work for the downstream gaze zone classification task. We demonstrate that the gaze direction cues could bring performance improvement for gaze zone classification.", + "bbox": [ + 496, + 574, + 890, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments:", + "text_level": 1, + "bbox": [ + 500, + 800, + 684, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00608, Artificial intelligence research about multi-modal interactions for empathetic conversations with humans)", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1563", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdul Rafey Aftab. Multimodal driver interaction with gesture, gaze and speech. In 2019 International Conference on Multimodal Interaction, pages 487-492, 2019. 1", + "[2] Aya Ataya, Won Kim, Ahmed Elsharkawy, and SeungJun Kim. Gaze-head input: Examining potential interaction with immediate experience sampling in an autonomous vehicle. Applied Sciences, 10(24), 2020. 1", + "[3] Yiwei Bao, Yunfei Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with rotation consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4207-4216, 2022. 2", + "[4] Xin Cai, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Source-free adaptive gaze estimation by uncertainty reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22035-22045, 2023. 2", + "[5] Yihua Cheng and Feng Lu. Gaze estimation using transformer. ICPR, 2022. 3, 7", + "[6] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20632-20641, 2023. 3, 6", + "[7] Yihua Cheng, Feng Lu, and Xucong Zhang. Appearance-based gaze estimation via evaluation-guided asymmetric regression. In The European Conference on Computer Vision, 2018. 2", + "[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 5", + "[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2", + "[10] Yihua Cheng, Haofei Wang, Yiwei Bao, and Feng Lu. Appearance-based gaze estimation with deep learning: A review and benchmark. arXiv preprint arXiv:2104.12668, 2021. 2, 4, 7", + "[11] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. AAAI, 2022. 2, 3", + "[12] In-Ho Choi, Sung Kyung Hong, and Yong-Guk Kim. Real-time categorization of driver's gaze zone using the deep learning techniques. In 2016 International conference on big data and smart computing (BigComp), pages 143-148. IEEE, 2016. 2", + "[13] Tobias Fischer, Hyung Jin Chang, and Yiannis Demiris. Rtgene: Real-time eye gaze estimation in natural environments. In The European Conference on Computer Vision, 2018. 2", + "[14] Lex Fridman, Philipp Langhans, Joonbum Lee, and Bryan Reimer. Driver gaze region estimation without use of eye movement. IEEE Intelligent Systems, 31(3):49-56, 2016. 2", + "[15] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2014. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Shreya Ghosh, Abhinav Dhall, Garima Sharma, Sarthak Gupta, and Nicu Sebe. Speak2label: Using domain knowledge for creating a large scale driver gaze zone estimation dataset. In The IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2896-2905, 2021. 1, 2, 7", + "[17] Amie C Hayley, Brook Shiferaw, Blair Aitken, Frederick Vinckenbosch, Timothy L Brown, and Luke A Downey. Driver monitoring systems (dms): The future of impaired driving management? Traffic injury prevention, 22(4):313-317, 2021. 1", + "[18] Sumit Jha and Carlos Busso. Probabilistic estimation of the gaze region of the driver using dense classification. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 697-702. IEEE, 2018. 1, 2", + "[19] Isaac Kasahara, Simon Stent, and Hyun Soo Park. Look both ways: Self-supervising driver gaze estimation and road scene saliency. In European Conference on Computer Vision, pages 126-142. Springer, 2022. 2", + "[20] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In The IEEE International Conference on Computer Vision, 2019. 1, 2, 7", + "[21] Muhammad Qasim Khan and Sukhan Lee. A comprehensive survey of driving monitoring and assistance systems. Sensors, 19(11), 2019. 1", + "[22] Muhammad Qasim Khan and Sukhan Lee. Gaze and eye tracking: Techniques and applications in adas. Sensors, 19 (24), 2019. 1", + "[23] Sung Joo Lee, Jaeik Jo, Ho Gi Jung, Kang Ryoung Park, and Jaihie Kim. Real-time gaze estimator based on driver's head orientation for forward collision warning system. IEEE Transactions on Intelligent Transportation Systems, 12(1): 254-267, 2011. 2", + "[24] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 4", + "[25] Prajval Kumar Murali, Mohsen Kaboli, and Ravinder Dahiya. Intelligent in-vehicle interaction technologies. Advanced Intelligent Systems, 4(2):2100122, 2022. 1", + "[26] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In The European Conference on Computer Vision, 2018. 2", + "[27] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In The European Conference on Computer Vision, pages 747-763. Springer, 2020. 2", + "[28] Akshay Rangesh, Bowen Zhang, and Mohan M Trivedi. Driver gaze estimation in the real world: Overcoming the eyeglass challenge. In The IEEE Intelligent Vehicles Symposium (IV), pages 1054-1059. IEEE, 2020. 2", + "[29] Yusuke Sugano, Yasuyuki Matsushita, and Yoichi Sato. Learning-by-synthesis for appearance-based 3d gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition, 2014. 5", + "[30] Ashish Tawari, Kuo Hao Chen, and Mohan M Trivedi. Where is the driver looking: Analysis of head, eye and" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1564", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "iris for robust gaze zone estimation. In 17th International IEEE conference on intelligent transportation systems (ITSC), pages 988-994. IEEE, 2014. 2", + "[31] Sourabh Vora, Akshay Rangesh, and Mohan Manubhai Trivedi. Driver gaze zone estimation using convolutional neural networks: A general framework and ablative analysis. IEEE Transactions on Intelligent Vehicles, 3(3):254-265, 2018. 1, 2", + "[32] Yafei Wang, Guoliang Yuan, Zetian Mi, Jinjia Peng, Xueyan Ding, Zheng Liang, and Xianping Fu. Continuous driver's gaze zone estimation using rgb-d camera. Sensors, 19(6): 1287, 2019. 1, 2", + "[33] Yaoming Wang, Yangzhou Jiang, Jin Li, Bingbing Ni, Wenrui Dai, Chenglin Li, Hongkai Xiong, and Teng Li. Contrastive regression for domain adaptation on gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19376-19385, 2022. 2", + "[34] Guoliang Yuan, Yafei Wang, Huizhu Yan, and Xianping Fu. Self-calibrated driver gaze estimation via gaze pattern learning. Knowledge-Based Systems, 235:107630, 2022. 2", + "[35] Mingfang Zhang, Yunfei Liu, and Feng Lu. Gazeonce: Real-time multi-person gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4197-4206, 2022. 2, 6", + "[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In The IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2, 4", + "[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2299-2308, 2017. 2, 7", + "[38] Xucong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2018. 5, 8", + "[39] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):162-175, 2019. 1, 2", + "[40] Xuong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In The European Conference on Computer Vision, 2020. 1, 2, 4, 7" + ], + "bbox": [ + 78, + 90, + 468, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1565", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_model.json b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b468f68764e0f7dcac0aa4c67c0c68ef6a85e9c2 --- /dev/null +++ b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.131, + 0.795, + 0.175 + ], + "angle": 0, + "content": "What Do You See in Vehicle? Comprehensive Vision Solution for In-Vehicle Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.811, + 0.294 + ], + "angle": 0, + "content": "Yihua Cheng\\(^{1}\\) Yaning Zhu\\(^{2}\\) Zongji Wang\\(^{3}\\) Hongquan Hao\\(^{4}\\) Yongwei Liu\\(^{4}\\) Shiqing Cheng\\(^{4}\\) Xi Wang\\(^{4}\\) Hyung Jin Chang\\(^{1}\\) \n\\(^{1}\\)University of Birmingham\\(^{1}\\), Huazhong University of Science and Technology\\(^{2}\\) \n\\(^{2}\\)NIST, Chinese Academy of Sciences\\(^{3}\\), CalmCar\\(^{4}\\) \ny.cheng.2@bham.ac.uk, h.j.chang@bham.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.474, + 0.83 + ], + "angle": 0, + "content": "Driver's eye gaze holds a wealth of cognitive and intentional cues crucial for intelligent vehicles. Despite its significance, research on in-vehicle gaze estimation remains limited due to the scarcity of comprehensive and well-annotated datasets in real driving scenarios. In this paper, we present three novel elements to advance in-vehicle gaze research. Firstly, we introduce IVGaze, a pioneering dataset capturing in-vehicle gaze, collected from 125 subjects and covering a large range of gaze and head poses within vehicles. In this dataset, we propose a new vision-based solution for in-vehicle gaze collection, introducing a refined gaze target calibration method to tackle annotation challenges. Second, our research focuses on in-vehicle gaze estimation leveraging the IVGaze. In-vehicle face images often suffer from low resolution, prompting our introduction of a gaze pyramid transformer that leverages transformer-based multilevel features integration. Expanding upon this, we introduce the dual-stream gaze pyramid transformer (GazeDPTR). Employing perspective transformation, we rotate virtual cameras to normalize images, utilizing camera pose to merge normalized and original images for accurate gaze estimation. GazeDPTR shows state-of-the-art performance on the IVGaze dataset. Thirdly, we explore a novel strategy for gaze zone classification by extending the GazeDPTR. A foundational tri-plane and project gaze onto these planes are newly defined. Leveraging both positional features from the projection points and visual attributes from images, we achieve superior performance compared to relying solely on visual features, substantiating the advantage of gaze estimation. The project is available at https://yihua.zone/work/ivgaze." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.844, + 0.21, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Understanding driver intention and behavior based on driver gaze is in high demand in intelligent vehicles, facilitating" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.326, + 0.882, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.478, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 1. In-vehicle gaze estimation illustration. The driver's gaze direction is estimated based on the facial images captured by the camera behind the steering wheels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.893, + 0.619 + ], + "angle": 0, + "content": "diverse applications such as in-vehicle interaction [1, 2, 25] and driver monitor systems [17, 21, 22]. Recent advances in vehicle gaze estimation concentrate primarily on gaze zone estimation [16, 18, 31, 32]. These approaches define multiple coarse regions, such as side mirrors and windshields, and conduct classification based on face images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.895, + 0.846 + ], + "angle": 0, + "content": "Gaze estimation serves as an upstream task of gaze zone estimation and can offer more precise information to understand driver attention. However, these methods typically require a large-scale dataset for training. Although there are numerous gaze datasets, collected in indoor [39, 40] or outdoor [20] environments, their applicability to the vehicle environment is limited due to the different environments and camera settings, resulting in suboptimal performance. Creating an in-vehicle gaze dataset proves challenging due to the confined and irregular nature of the vehicular environment. Constructing in-vehicle gaze collection systems remains an unsolved issue, as traditional gaze collection systems are impractical for use within vehicles. The absence of in-vehicle gaze datasets acts as a significant barrier to the progress of in-vehicle gaze estimation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.847, + 0.893, + 0.862 + ], + "angle": 0, + "content": "In this paper, we present a comprehensive vision-based" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.893, + 0.901 + ], + "angle": 0, + "content": "1Our work focuses on gaze direction estimation. We abbreviate gaze direction as gaze in the rest." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1556" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "in-vehicle gaze estimation research: a novel vision-based gaze collection system for vehicles, offering a first-of-its-kind in-vehicle gaze dataset, a dual-stream gaze pyramid transformer for accurate in-vehicle gaze estimation, and its extension to gaze zone classification, showcasing its effectiveness in enhancing gaze estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.44 + ], + "angle": 0, + "content": "First, we introduce IVGaze, an in-vehicle gaze dataset collected from 125 subjects. IVGaze provides a dense distribution of gaze directions, covering a wide range within in-vehicle environments. It contains various conditions, including diverse head poses, eye movements, illumination variations, and the presence of face accessories such as glasses, sunglasses, and masks. To collect IVGaze, we propose a vision-based gaze collection system. The system does not require dedicated eye-tracking devices and is easy to reproduce. We use a single camera for facial appearance capture and paste stickers in vehicles as gaze targets. However, a significant challenge lies in calibrating the 3D position of gaze targets. This challenge arises because gaze targets are out of the camera's field of view (out-of-FoV). To address this issue, we present a refined gaze target calibration method, which leverages an auxiliary camera to capture gaze targets and a transparent chessboard for calibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Secondly, we explore in-vehicle gaze estimation using the IVGaze. Face images often suffer from low resolution due to the inherent limitations of cameras in vehicles. We propose a gaze pyramid transformer that utilizes a transformer to integrate multilevel features. Expanding upon this, we propose a dual-stream gaze pyramid transformer (GazeDPTR). We rotate virtual cameras via perspective transformation to normalize images, and leverage camera pose to merge normalized and original images. GazeDPTR shows state-of-the-art results on IVGaze." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Thirdly, we extend GazeDPTR for the downstream gaze zone classification task. It is challenging to compute the intersection of gaze and the vehicle. We define a foundational tri-plane and project gaze to the tri-plane. We extract positional features from intersection points and predict gaze zones using both positional features and visual features from face images. Our experiment demonstrates that the gaze zone classification can be further enhanced by positional features, showing the advantage of gaze estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.745, + 0.227, + 0.76 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.277, + 0.785 + ], + "angle": 0, + "content": "2.1. Gaze Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The human gaze is inherently implicit and poses a challenge for objective measurement, making gaze annotation complex in gaze data collection. Some methods capture the human gaze through intrusive devices such as eye-tracking glasses [13, 19]. However, the eye-tracking glasses have a notable impact on the quality of the captured facial images. Fischer et al. [13] attempted to mitigate this impact" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.09, + 0.892, + 0.16 + ], + "angle": 0, + "content": "Table 1. The comparison of gaze estimation datasets. Gaze annotation is challenging in the vehicle environment, which results in existing in-vehicle datasets only providing gaze zone annotation. Our work addresses this issue and contributes the first in-vehicle gaze dataset containing gaze annotation and natural face images." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.165, + 0.89, + 0.375 + ], + "angle": 0, + "content": "
DatasetsEnv.# Sub.Annotation
GazeZone
Gaze360 [20]Outdoor238✓-
MPIIGaze [39]15✓-
EyeDiap [15]Indoor16✓-
EVE [27]54✓-
ETH-XGaze [40]110✓-
Vora et al. [31]10✗✓
Jha & Busso [18]16✗✓
Wang et al. [32]Vehicle3✗✓
Rangesh et al. [28]13✗✓
Ghosh et al. [16]338✗✓
IVGaze (Ours)125✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.566 + ], + "angle": 0, + "content": "using GAN, but there are also some artifacts in the resulting face images. Vision-based gaze collection systems typically define gaze as direction vectors originating from facial centers towards specific gaze targets [27, 39, 40]. However, these gaze targets often remain outside the camera's field of view. This out-of-field view challenge complicates the calibration of gaze target positions. Zhang et al. [39, 40] sets screen points as gaze targets and uses a mirror to calibrate the screen plane. Kellnhofer et al. [20] uses \\(360^{\\circ}\\) panoramic cameras to capture gaze targets and human faces simultaneously where the panoramic camera is pre-calibrated. However, these strategies are not applicable to vehicles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.687 + ], + "angle": 0, + "content": "In-vehicle gaze dataset usually defines different region such as windshield and left/right mirror in vehicles and perform gaze zone classification [12, 14, 16, 18, 23, 30, 34]. Kasahara et al. [19] collects an in-vehicle gaze dataset but subjects are required to wear eye-tracking glasses, which means the dataset is not applicable in the real world. Our gaze collection system does not require dedicated devices and produces natural face images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.81, + 0.712 + ], + "angle": 0, + "content": "2.2. Appearance-based Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Appearance-based gaze estimation directly learns mapping function from facial appearance to human gaze [10]. Conventional gaze estimation methods extract eye features from eye images [7, 9, 26, 36]. They concatenate the head pose vector with eye features for gaze estimation. Recent methods directly learn gaze from face images [3, 4, 11, 33, 35]. Face images provide both eye region and head pose information, resulting in superior performance compared to methods relying solely on eye images. However, the subtlety of the human eye in face images poses a challenge as it can be easily overlooked by the network. Zhang et al. [37] utilize a learnable attention map to guide the net" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1557" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.463, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.28, + 0.264, + 0.29 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.891, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.28, + 0.58, + 0.29 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.28, + 0.73, + 0.29 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.28, + 0.844, + 0.29 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.894, + 0.377 + ], + "angle": 0, + "content": "Figure 2. We construct a vision-based in-vehicle gaze collection system comprising a DMS camera, a depth camera, and strategically placed gaze targets, as depicted in (c). The DMS camera is positioned behind the steering wheel to capture drivers' facial appearances, while the gaze targets, positioned beyond the DMS camera's field-of-view (FoV), such as on the windshield, remain unobserved. The depth camera, utilized for calibration purposes, is temporarily installed for capturing gaze target positions in 3D with respect to its own coordinates, and it is removed during data collection. To facilitate the calibration of the depth camera's pose relative to the DMS camera, we propose employing a transparent chessboard, which is placed between the two cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.522 + ], + "angle": 0, + "content": "work's focus specifically on the eye. Cheng et al. [11] crop eye images from face images and construct a cascade network to leverage them. In recent developments, transformers have showcased remarkable capabilities in gaze estimation. Cheng et al. [5] demonstrate that transformers can achieve state-of-the-art performance across various benchmarks. They apply transformers to address dual-view gaze estimation and show that dual-view images outperform single-view images [6]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.537, + 0.434, + 0.554 + ], + "angle": 0, + "content": "3. In-Vehicle Gaze Data Collection System" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.684 + ], + "angle": 0, + "content": "In this section, we introduce our vision-based system for in-vehicle gaze data collection, eliminating the need for dedicated eye-tracking devices. Our system includes a neat gaze target calibration method that effectively addresses the core challenge in gaze annotation. Leveraging this system, we collect the first in-vehicle gaze dataset from 125 subjects. The dataset spans a diverse range of gaze, facilitating and advancing future research in in-vehicle gaze estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.219, + 0.71 + ], + "angle": 0, + "content": "3.1. System Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We place gaze targets within the vehicle, and subjects are instructed to look at each designated target during the data collection process. Simultaneously, we record their facial appearance along with corresponding gaze targets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Camera. Our system uses one camera of a driver monitor system (DMS) for facial appearance capture. The DMS camera is an infrared camera with a capture resolution of \\(1280 \\times 800\\). The camera is located behind the steering wheel and directly points at the head region of drivers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Gaze targets. We mark gaze targets with red points on stickers, and each target is uniquely labeled with a distinct number printed on stickers to aid differentiation. We strate-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.462 + ], + "angle": 0, + "content": "gically position these stickers within various gaze regions within the vehicle, including the windshield, left and right-side mirrors, rear-view mirror, center console, speedometer, handbrake, and etc. These targets cover a large gaze region, satisfying the gaze estimation requirements in vehicles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.698, + 0.485 + ], + "angle": 0, + "content": "3.2. Collection Procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.659 + ], + "angle": 0, + "content": "We meticulously design the collection procedure to minimize error. During the collection process, participants are instructed to look at specific gaze targets in a predefined sequence. They are also required to speak the corresponding number associated with each target to confirm their focus accuracy. Participants must maintain focus on the target for 3 seconds to facilitate image capture. We conduct a preliminary check on the captured images and discard any erroneous ones. Once this verification is complete, the current data collection phase concludes, and participants are instructed to shift their gaze to the next gaze targets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We design three postures for each participant. Participants should complete the collection process three times with different postures. The first posture requires participants not to change head pose. They should preserve the same head pose and only move eyeballs to focus a gaze target. We aim to collect sufficient eye movement data with this posture. We do not require participants to focus on the target where participants think it is hard to focus with only eye movement, e.g., the target in the side mirror. Participants can rotate their heads to look at these points but they are also required to preserve the new head pose until they cannot look at any targets with the current head pose. The second posture requires participants not to move their head position but they can rotate their heads to focus a gaze target. We do not set any constraints in the third posture. Participants can freely perform head movements in this pos" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1558" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.106, + 0.372, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.09, + 0.705, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.091, + 0.931, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Figure 3. Our dataset is collected using IR cameras in the vehicle environment. (a) We present image samples of IVGaze, highlighting the challenges posed by realistic in-vehicle conditions, including cases with sunglasses and reflections in glasses. (b) We categorize the image count based on their mean pixel value, showing the diversity of illumination conditions. (c) The image count is analyzed based on face accessories including glasses, sunglasses, and masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.388 + ], + "angle": 0, + "content": "ture. The last two postures bring head pose variation for IVGaze. We do not use devices such as headrests to constrain participants. Participants only need to preserve the specific posture and focus on targets in a comfortable way." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.245, + 0.415 + ], + "angle": 0, + "content": "3.3. Gaze Annotation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.469, + 0.514 + ], + "angle": 0, + "content": "Vision-based gaze collection systems define gaze as unit direction vectors originating from the face center \\(\\mathbf{o}\\) and extending toward gaze targets \\(\\mathbf{t}\\) [10]. Our system follows this framework and the gaze annotation can be computed as \\(\\mathbf{g} = (\\mathbf{t} - \\mathbf{o}) / ||\\mathbf{t} - \\mathbf{o}||_2\\). The gaze annotation is decomposed into 3D face center estimation and gaze target calibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.469, + 0.637 + ], + "angle": 0, + "content": "3D face center estimation has been effectively addressed in previous research. We first detect facial landmarks and fit a 3D morphable face model for 3D facial landmarks [24, 40]. We select the 3D position of the nose as gaze origin o. However, the calibration of gaze target positions presents a challenge because gaze targets are out-of-FoV. It means that the DMS camera cannot capture any images of the gaze targets, complicating the calibration problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.303, + 0.666 + ], + "angle": 0, + "content": "3.4. Gaze Target Calibration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.779 + ], + "angle": 0, + "content": "We need the 3D gaze target position \\(\\mathbf{t}\\) for gaze annotation. However, a key problem is all gaze targets are out-of-FoV. Conventional gaze datasets often present gaze targets on a screen and employ a mirror to solve the out-of-FoV problem, where the mirror reflects the content of the screen [36]. However, these methods are not suitable for vehicle scenarios and can lead to a substantial accumulation of errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we propose a neat and efficient gaze target calibration method. Our basic idea is to employ an auxiliary camera to capture gaze targets. This allows a straightforward calculation of the 3D gaze target positions w.r.t. the auxiliary camera coordinate system. However, it also introduces a new challenge: how to calibrate extrinsic matrix of the auxiliary cameras w.r.t. DMS camera? Stereo calibration is typically used to compute the pose of two cam" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.403 + ], + "angle": 0, + "content": "eras. However, it requires two captured images sharing corresponding points. In our system, gaze targets are located behind the DMS camera which means the DMS camera and auxiliary camera should be oriented in opposite directions, making traditional stereo calibration infeasible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.893, + 0.583 + ], + "angle": 0, + "content": "To solve this problem, we propose to use a transparent chessboard for the pose calibration. We set a transparent chessboard between the two cameras and the two cameras respectively capture each side of the chessboard. We first calibrate the two camera poses w.r.t. the chessboard coordinate system and then compute the pose between two cameras. It is worth noting that there are two different chessboard coordinate systems corresponding to each side of the chessboard. We can derive the transformation matrix between the two chessboard coordinate systems where \\(\\mathbf{R}_{\\mathrm{chess}} = \\mathrm{diag}(0,0,-1)\\) and \\(\\mathbf{t}_{\\mathrm{chess}} = (0,0,-\\mathrm{d})\\). The d is the thickness of the chessboard." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.675 + ], + "angle": 0, + "content": "In detail, we use an Intel RealSense D435 depth camera as the auxiliary camera since it can provide accurate 3D positions. We show the camera layout in Fig. 2(b). We can obtain the transformation matrix \\(\\mathbf{R}_{\\mathrm{dms}}\\), \\(\\mathbf{t}_{\\mathrm{dms}}\\) and \\(\\mathbf{R}_{\\mathrm{depth}}\\), \\(\\mathbf{t}_{\\mathrm{depth}}\\) which can transfer a point from chessboard coordinate systems to camera coordinate systems. We have" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.685, + 0.891, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {\\mathrm {r o t}} = \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {R} _ {\\mathrm {c h e s s}} \\mathbf {R} _ {\\mathrm {d e p t h}} ^ {- 1}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.716, + 0.891, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {\\text {r o t}} = - \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {R} _ {\\text {c h e s s}} \\mathbf {R} _ {\\text {d e p t h}} ^ {- 1} \\mathbf {t} _ {\\text {d e p t h}} + \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {t} _ {\\text {c h e s s}} + \\mathbf {t} _ {\\mathrm {d m s}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.891, + 0.793 + ], + "angle": 0, + "content": "We use these matrices to convert points from the depth camera coordinate system into the DMS camera coordinate system. Please refer to the supplementary material for details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.722, + 0.817 + ], + "angle": 0, + "content": "3.5. In-Vehicle Gaze Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.891, + 0.871 + ], + "angle": 0, + "content": "We collect the first in-vehicle gaze dataset IVGaze which provides dense gaze annotation and natural face images. We show the dataset samples in Fig. 3 (a)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset Statistics. We collect 44,705 images from 125 subjects. The number of images per subject ranges from" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1559" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.27, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.089, + 0.451, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.226, + 0.47, + 0.254 + ], + "angle": 0, + "content": "Figure 4. We show the distribution of data for gaze (left) and head movements (right). Brighter regions denote higher data density." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.271, + 0.468, + 0.332 + ], + "angle": 0, + "content": "169 to 946. Most subjects provide around 300 images per person, i.e., they repeat the collection procedure three times with three different postures. 40 subjects are female and 85 subjects are male, with ages spanning from 20 to 50 years." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.468, + 0.409 + ], + "angle": 0, + "content": "Collection Conditions. The dataset was collected between 9 am and \\(7\\mathrm{pm}\\) in outdoor environments, covering a wide range of lighting conditions. We also collect images in indoor environments, i.e., a garage. The dataset statistics across pixel intensity are shown in Fig. 3 (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.468, + 0.533 + ], + "angle": 0, + "content": "Face Accessories. We consider two face accessories during the collection: glasses and masks. For subjects who did not wear glasses, we provided glasses and asked them to repeat the collection wearing glasses. We provided white and black masks for some subjects and asked them to repeat the collection wearing masks. We also required a few subjects to wear sunglasses to facilitate future research. We show the statistics in Fig. 3 (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.468, + 0.762 + ], + "angle": 0, + "content": "Data Distribution. We plot the angular distribution of gaze and head pose. Note that the data normalization method (introduced in the next section) changes gaze directions and head poses with a rotation matrix. Fig. 4 shows the distribution in normalization space. The horizontal gaze is from \\(-50^{\\circ}\\) to \\(90^{\\circ}\\), and the vertical gaze is from \\(-40^{\\circ}\\) to \\(40^{\\circ}\\). It is worth noting that the vehicle environment has limited space, and our dataset already covers almost all regions. Our dataset is collected in the real vehicle environment, and the driver seat in the vehicle is located on the left. Therefore, our dataset has a relatively small distribution on the left. The right figure in Fig. 4 shows the head pose distribution. Our dataset contains a large range of head poses in the yaw axis, as subjects need to look at gaze targets from the left-side mirror to the right-side mirror." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.783, + 0.332, + 0.799 + ], + "angle": 0, + "content": "4. In-Vehicle Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The in-vehicle environment brings new settings and challenges. In this section, we systematically explore in-vehicle gaze estimation. We introduce a revised data preprocessing method and describe a novel gaze estimation network. We also extend the network for an application and prove the advantage of in-vehicle gaze estimation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.848, + 0.107 + ], + "angle": 0, + "content": "4.1. Revisiting Data Normalization in Vehicle" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.892, + 0.204 + ], + "angle": 0, + "content": "Data normalization rotates and translates cameras via perspective transformation in face images to reduce head pose variation. Conventional methods usually define the rotated camera coordinate system \\(\\mathbf{C}_r\\) based on head pose [29, 38]. However, these methods cannot bring performance improvement in the vehicle environment." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.205, + 0.893, + 0.448 + ], + "angle": 0, + "content": "The \\(x\\)-axis of \\(\\mathbf{C}_r\\) is typically defined as the \\(x\\)-axis of head coordinate systems. It means they will rotate images to keep the head straight. We observed that the designed \\(x\\)-axis will produce unstable results, especially in the extreme head pose. Therefore, we propose that do not rotate virtual cameras based on the \\(x\\)-axis of head coordinate systems. In detail, we compute the \\(z\\)-axis of \\(\\mathbf{C}_r\\) as direction vectors from cameras to face centers. We use the \\(x\\)-axis of the virtual camera, i.e., \\((1,0,0)\\) rather than the head coordinate system. We compute the \\(y\\)-axis as \\(y = z\\times x\\) and also recompute the \\(x\\)-axis as \\(x = y\\times z\\). Finally, we have a rotation matrix \\(\\mathbf{R} = [x;y;z]\\) and use a scale matrix \\(S\\) to maintain the user-camera distance. We warp the image based on matrix SR. The human gaze is also changed as \\(\\mathbf{g}^n = \\mathbf{R}\\mathbf{g}^o\\), where \\(\\mathbf{g}^n\\) and \\(\\mathbf{g}^o\\) denotes gaze in the normalization and origin space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.749, + 0.47 + ], + "angle": 0, + "content": "4.2. Gaze Pyramid Transformer" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Gaze estimation methods typically learn gaze from face images. However, face images captured in vehicles often suffer from low resolution owing to the limited capabilities of the camera. This low resolution can lead conventional methods to overlook subtle details, especially in the eye region. Some techniques address this challenge by cropping eye images and processing them separately [8], but these approaches encounter difficulties when faced with extreme head poses, as one of the eye images may not be visible." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.615, + 0.892, + 0.765 + ], + "angle": 0, + "content": "We propose a gaze pyramid transformer (GazePTR) in this section. We build a feature pyramid and input multilevel features into a transformer for gaze estimation. As shown in Fig. 5, our network uses a convolution network to extract features from facial images. We collect feature maps from different levels rather than the last level. We use \\(1 \\times 1\\) convolution layers and global average pooling layers to preserve the same feature dimension. Finally, we input the multi-level feature into a transformer. A learnable token is used to aggregate multi-level features for gaze estimation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, the network has the same performance achieved by directly utilizing the last feature map for gaze estimation. This result can be attributed to the network's tendency to overlook low-level features. To address this issue, we introduce additional constraints by estimating gaze from each level of features and calculating corresponding loss. It can be understood that we first extract multi-level features, with each feature being potentially useful. A transformer is used to effectively ensemble these features." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1560" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.095, + 0.895, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.896, + 0.379 + ], + "angle": 0, + "content": "Figure 5. The GazeDPTR directly crop face for origin images and rotates virtual cameras via perspective transformation for normalized images. It builds a dual-stream network to extract features from the two images based on the GazePTR for feature extraction which integrates multi-level features via transformers. To further merge the features from two streams, we leverage a transformer where camera pose is used as the positional feature in the transformer. We define the original camera pose as \\(\\mathbf{C} = \\text{diag}(1, 1, 1)\\) and the camera pose in normalization space is RC. We also extend the network for gaze zone classification. We define a tri-plane and project gaze into them. We extract positional features from three intersection points via a transformer. We also extract visual features from images and predict the gaze zone based on both visual features and positional features. The whole network is trained in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.391, + 0.43, + 0.407 + ], + "angle": 0, + "content": "4.3. Dual-Stream Gaze Pyramid Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Dual-view images have been demonstrated to outperform single-view images in gaze estimation [6]. In our work, we rotate cameras in data normalization through perspective transformation. This inspired us to formulate the dual-camera setting leveraging both original and normalized images. However, dual-view images improve accuracy since they contain more visual information. It remains uncertain whether the combination of normalized and original images can provide additional insights beyond what each offers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.641 + ], + "angle": 0, + "content": "To validate our hypothesis, we executed an oracle baseline. We separately train GazePTR on the original and the normalized datasets, selecting the best result from each image pair. The result reveals a significant performance improvement, validating the inherent advantages of utilizing both images. Please refer to the supplementary for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Therefore, we propose a dual-stream gaze pyramid transformer (GazeDPTR). Our method leverages both normalized images and original images for gaze estimation. We employ GazePTR to extract features from the two images and a transformer is used to integrate the two features for gaze estimation. Note that, the two features are in different camera coordinate systems and correspond to different gaze. To establish the connection, we use the camera pose as the positional information in the transformer. However, it is impossible the calibrate a virtual camera. We define the camera pose of original images as \\(\\mathbf{C} = \\text{diag}(1, 1, 1)\\) and the camera pose of normalized images is RC. We extract the \\(z\\)-axis and use positional encoding for camera pose [6]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.401, + 0.864 + ], + "angle": 0, + "content": "4.4. Strategy for Gaze Zone Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Detecting driver attention is essential in driver monitoring systems. In this section, we introduce a novel strategy to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.542 + ], + "angle": 0, + "content": "extend GazeDPTR for gaze zone classification. GazeDPTR estimates gaze direction from face images. While the conventional solution involves projecting gaze direction and computing intersections within the vehicle, it is not practical without a 3D vehicle model. Zhang et al. use a triplane for self-supervised loss function [35]. Our alternative strategy defines a foundational tri-plane and computes the intersection of gaze vectors with this tri-plane. We extract positional features from three intersection points to facilitate gaze zone classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.893, + 0.739 + ], + "angle": 0, + "content": "In detail, we define a tri-plane with normal vectors \\((1,0,0)\\), \\((0,1,0)\\), and \\((0,0,1)\\), intersecting at the origin \\((0,0,0)\\). Given the estimated gaze vector \\(\\mathbf{g}^n\\), we convert it back to the original space using \\(\\mathbf{g}^o = \\mathbf{R}^{-1}\\mathbf{g}^n\\). We then project \\(\\mathbf{g}^o\\) onto the tri-plane with the face center \\(\\mathbf{o}\\) and obtain three intersection points. To enhance the positional features, we apply positional encoding to these points and input them into a 2-layer transformer. The positional encoding not only increases feature dimensions but also normalizes them within the range of -1 to 1. Additionally, we use another learnable token to aggregate visual features from the original images. Gaze zone classification is performed using both visual features and positional features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.8 + ], + "angle": 0, + "content": "We train the whole network in an end-to-end manner. We use L1 loss for gaze estimation and cross-entropy loss for gaze zone classification. Please refer to the supplementary for implementation details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.625, + 0.832 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset: IVGaze contains 44,705 images of 125 subjects. To perform within-dataset evaluation, we divide the dataset into three subsets based on subjects. The image numbers of the three subsets are 15,165, 14,674, and 14,866. We" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1561" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 2. Performance comparison in gaze estimation. Our methods achieve better performance than the compared methods." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.122, + 0.477, + 0.301 + ], + "angle": 0, + "content": "
Angular ErrorAverage Precision (AP)
<2°<4°<6°<8°
FullFace [37]13.67°2.3%8.8%17.8%28%
DWG [16]8.82°6.6%21.7%38.1%53.2%
Gaze360 [20]8.15°9.2%27.3%44.6%58.9%
FullFace+ [37]7.48°14.2%31.1%46.7%63.1%
GazeTR [5]7.33°17%32.8%47.5%64.7%
XGaze [40]7.06°11.7%32.7%51.5%66.7%
GazePTR7.04°17.6%34%49.3%66.7%
GazeDPTR6.71°22.1%36%50.3%68.4%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.411, + 0.334 + ], + "angle": 0, + "content": "perform three-fold cross-validation on our dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.472 + ], + "angle": 0, + "content": "Evaluation Metric: We use the angular error as the evaluation metric of gaze estimation as most of the methods [10], where a smaller value represents a better model. However, we notice the angular error only shows the overall performance while cannot give deep insights. Therefore, we define the average precision (AP) where AP of \\( < k^{\\circ} \\) means an estimation is considered correct if the angular error is lower than \\( k^{\\circ} \\). Regarding the gaze zone classification, average precision is used as a common multi-class classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.371, + 0.497 + ], + "angle": 0, + "content": "5.1. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.595 + ], + "angle": 0, + "content": "We first compared our methods with SOTA methods in the in-vehicle gaze estimation task. We compared our methods with FullFace [37], Gaze360 [20], XGaze [40], DWG [16] and GazeTR [5]. We replaced the backbone of the FullFace method from AlexNet to ResNet18 for a more convincing comparison. We denote the new method as FullFace+." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.716 + ], + "angle": 0, + "content": "The result is shown in Tab. 2. GazePTR and GazeDPTR both show better performance than the compared methods. GazePTR and GazeTR have the same backbone and transformer architecture. However, GazePTR outperforms \\(0.29^{\\circ}\\) thanGazeTR since GazePTR uses multi-level feature maps. GazeDPTR builds a dual-stream network and further brings \\(0.33^{\\circ}\\) improvement than GazePTR. These results show the advantage of our methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Interestingly, XGaze has a similar angular error to GazePTR. However, it is significantly worse than GazePTR in AP of \\(< 2^{\\circ}\\). It is because that XGaze uses ResNet50 as the backbone. The deep backbone enhances the feature extraction ability but also easily overlooks the small eye region, making it challenging to achieve precise gaze estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.355, + 0.833 + ], + "angle": 0, + "content": "5.2. The Impact of Face Accessories" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "IVGaze provides rich samples of wearing face accessories, including glasses, masks, and sunglasses. We also evaluate their impacts on the accuracy of gaze estimation and show in Tab. 3. The result shows that eyeglasses have a" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 3. We show the impacts of different face accessories on performance. The sunglasses bring a significant performance drop." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.122, + 0.905, + 0.301 + ], + "angle": 0, + "content": "
GlassesMaskSunglasses
withw/owithw/owith
FullFace [37]14.43°12.40°15.20°13.35°21.39°
DWG [16]9.20°8.19°9.43°8.69°17.43°
Gaze360[20]8.30°7.91°8.95°7.99°17.99°
FullFace+ [37]7.59°7.30°8.37°7.30°16.50°
XGaze [40]7.07°7.03°7.80°6.90°15.15°
GazeTR [5]7.40°7.22°8.12°7.17°17.49°
GazePTR7.13°6.90°7.78°6.89°16.54°
GazeDPTR6.77°6.63°7.44°6.57°16.41°
" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.313, + 0.697, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.313, + 0.891, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.445, + 0.894, + 0.556 + ], + "angle": 0, + "content": "Figure 6. We compute the angular degree between gaze direction/head orientation with the frontal direction, i.e., \\((0,0,-1)\\), and count the mean gaze estimation accuracy in different ranges. Interestingly, GazePTR brings significant performance improvement to GazeTR in the \\(0^{\\circ} - 20^{\\circ}\\) gaze range. This is because subjects prefer to move their eyeball rather than their head when the gaze target is located at a close range. GazePTR utilizes multilevel features and can accurately capture eye movement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.688 + ], + "angle": 0, + "content": "large impact if the basic performance is not good. However, as the basic performance improves, the performance difference reduces. This is because most of glasses are transparent, enabling a robust model to extract features effectively. Sunglasses have a significant impact on performance, with average performance exceeding \\(15^{\\circ}\\). Masks have a substantial impact on performance, as they obscure facial regions, complicating the extraction of head-related features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.734, + 0.712 + ], + "angle": 0, + "content": "5.3. Performance Distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We show the performance distribution to provide deep insight. We first show the performance distribution in different gaze ranges. We introduce a novel concept, the gaze range of \\(a^{\\circ} - b^{\\circ}\\), which is defined as a set of samples satisfying the requirement that the angular degree between the gaze direction and the frontal gaze direction is in the range. The result is shown in Fig. 6. Interestingly, GazePTR and GazeDPTR have better performance in \\(0^{\\circ} - 20^{\\circ}\\). This is because subjects prefer to move their eye rather than head to look at objects in this range. Our methods leverage multi-level features and easily capture eye movement. We also show the performance distribution in different head ranges." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1562" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.146 + ], + "angle": 0, + "content": "Table 4. We present the performance of GazePTR at each level of the feature. The higher-level feature achieves better performance. Our method integrates features from all levels and achieves the best performance." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.15, + 0.462, + 0.187 + ], + "angle": 0, + "content": "
1st2nd3rd4thOurs
GazePTR10.33°8.59°7.61°7.35°7.04°
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.196, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Table 5. We present the performance of GazePTR separately training on the original images and normalized images. GazeDPTR leverages camera pose to establish a connection between the two images and use both two images for gaze estimation. We also removed the camera pose to perform an ablation study." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.274, + 0.468, + 0.324 + ], + "angle": 0, + "content": "
Original imagesNormalized imagesGazeDPTR (w/o camera pose)GazeDPTR
Acc7.44°7.04°7.03°6.71°
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.47, + 0.363 + ], + "angle": 0, + "content": "Our methods are robust in different head ranges where the maximum performance difference is only \\(1^{\\circ}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.373, + 0.231, + 0.389 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.517 + ], + "angle": 0, + "content": "GazePTR leverages multi-level features for gaze estimation. We first conduct an ablation study to demonstrate the advantage. We obtain the gaze performance of GazePTR at each level of feature and show it in Tab. 4. The result demonstrates a higher-level feature achieves better performance. GazePTR integrates multi-level features and achieves the best performance. The result proves the advantage of our design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.471, + 0.669 + ], + "angle": 0, + "content": "GazeDPTR leverages both original and normalized images along with their corresponding camera poses. To demonstrate the effectiveness of our design, we perform an ablation study in Tab. 5. Training GazePTR separately on original and normalized images yields performances of \\(7.44^{\\circ}\\) and \\(7.04^{\\circ}\\) respectively. GazeDPTR, which incorporates both image types, achieves a performance improvement of \\(0.34^{\\circ}\\). We also experimented by excluding the positional encoding of the camera pose in GazeDPTR. Interestingly, the result is comparable to using only normalized images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.3, + 0.695 + ], + "angle": 0, + "content": "5.5. Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.762 + ], + "angle": 0, + "content": "Data Normalization. We compare our method with the data normalization method [38] in the left of Tab. 6. The previous method cannot work well in vehicle environments while our method brings \\(0.3^{\\circ}\\) performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.762, + 0.436, + 0.777 + ], + "angle": 0, + "content": "Case Study. We visual gt and our prediction in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.305, + 0.802 + ], + "angle": 0, + "content": "5.6. Gaze Zone Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We propose a basis tri-plane to acquire positional feature and combine both positional feature and visual features for gaze zone classification. We respectively evaluate each feature and show the result in the right of Tab. 6. The visual feature achieves better performance than the positional feature since the positional feature is obtained from gaze pro" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.16 + ], + "angle": 0, + "content": "Table 6. We evaluate different data normalization methods in the left table. Our method outperforms the previous method [38]. In the right table, we evaluate different features in the gaze zone classification. Our method uses both positional feature and visual features, achieving \\(2.4\\%\\) improvement over visual features." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.167, + 0.89, + 0.247 + ], + "angle": 0, + "content": "
Normalization [38]oursGazeTRGazePTRVisual FeaturePositional FeatureAP
××7.77°7.44°✓79.4%
✓8.64°8.53°✓75.3%
✓7.33°7.04°✓✓81.8%
" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.25, + 0.601, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.25, + 0.692, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.25, + 0.785, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.25, + 0.876, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.32, + 0.6, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.32, + 0.692, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.32, + 0.785, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.32, + 0.876, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.392, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 7. We illustrate GT with green lines and Prediction with red lines. The number shows the angular error in degree. The right two figures are failure cases since the eye is hard to capture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.541 + ], + "angle": 0, + "content": "jection. The projection increases the model interpretability but decreases the fitting ability. Our method uses the two features and gets \\(2.4\\%\\) performance improvement. The result demonstrates that the gaze zone classification task could be enhanced by gaze direction cues, highlighting the advantage of our gaze estimation work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.62, + 0.566 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.787 + ], + "angle": 0, + "content": "In this work, we provide systematical research in in-vehicle gaze estimation, including dataset, algorithm, and extensive application. We first solve the data collection issue in a vehicle where a gaze target calibration method is proposed. We collect the first in-vehicle gaze dataset containing dense gaze annotation and natural face images. We further explore the algorithm for in-vehicle gaze estimation. Our work brings two deep insights 1) multi-level feature is useful to capture eye region information. 2) simultaneously leveraging original images and normalized images could achieve better performance. We also extend our work for the downstream gaze zone classification task. We demonstrate that the gaze direction cues could bring performance improvement for gaze zone classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.685, + 0.818 + ], + "angle": 0, + "content": "7. Acknowledgments:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00608, Artificial intelligence research about multi-modal interactions for empathetic conversations with humans)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1563" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Abdul Rafey Aftab. Multimodal driver interaction with gesture, gaze and speech. In 2019 International Conference on Multimodal Interaction, pages 487-492, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Aya Ataya, Won Kim, Ahmed Elsharkawy, and SeungJun Kim. Gaze-head input: Examining potential interaction with immediate experience sampling in an autonomous vehicle. Applied Sciences, 10(24), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Yiwei Bao, Yunfei Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with rotation consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4207-4216, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Xin Cai, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Source-free adaptive gaze estimation by uncertainty reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22035-22045, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[5] Yihua Cheng and Feng Lu. Gaze estimation using transformer. ICPR, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[6] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20632-20641, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[7] Yihua Cheng, Feng Lu, and Xucong Zhang. Appearance-based gaze estimation via evaluation-guided asymmetric regression. In The European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[10] Yihua Cheng, Haofei Wang, Yiwei Bao, and Feng Lu. Appearance-based gaze estimation with deep learning: A review and benchmark. arXiv preprint arXiv:2104.12668, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[11] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. AAAI, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[12] In-Ho Choi, Sung Kyung Hong, and Yong-Guk Kim. Real-time categorization of driver's gaze zone using the deep learning techniques. In 2016 International conference on big data and smart computing (BigComp), pages 143-148. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[13] Tobias Fischer, Hyung Jin Chang, and Yiannis Demiris. Rtgene: Real-time eye gaze estimation in natural environments. In The European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[14] Lex Fridman, Philipp Langhans, Joonbum Lee, and Bryan Reimer. Driver gaze region estimation without use of eye movement. IEEE Intelligent Systems, 31(3):49-56, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.175 + ], + "angle": 0, + "content": "[16] Shreya Ghosh, Abhinav Dhall, Garima Sharma, Sarthak Gupta, and Nicu Sebe. Speak2label: Using domain knowledge for creating a large scale driver gaze zone estimation dataset. In The IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2896-2905, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.893, + 0.244 + ], + "angle": 0, + "content": "[17] Amie C Hayley, Brook Shiferaw, Blair Aitken, Frederick Vinckenbosch, Timothy L Brown, and Luke A Downey. Driver monitoring systems (dms): The future of impaired driving management? Traffic injury prevention, 22(4):313-317, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.245, + 0.893, + 0.301 + ], + "angle": 0, + "content": "[18] Sumit Jha and Carlos Busso. Probabilistic estimation of the gaze region of the driver using dense classification. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 697-702. IEEE, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[19] Isaac Kasahara, Simon Stent, and Hyun Soo Park. Look both ways: Self-supervising driver gaze estimation and road scene saliency. In European Conference on Computer Vision, pages 126-142. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[20] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In The IEEE International Conference on Computer Vision, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.893, + 0.453 + ], + "angle": 0, + "content": "[21] Muhammad Qasim Khan and Sukhan Lee. A comprehensive survey of driving monitoring and assistance systems. Sensors, 19(11), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[22] Muhammad Qasim Khan and Sukhan Lee. Gaze and eye tracking: Techniques and applications in adas. Sensors, 19 (24), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.893, + 0.565 + ], + "angle": 0, + "content": "[23] Sung Joo Lee, Jaeik Jo, Ho Gi Jung, Kang Ryoung Park, and Jaihie Kim. Real-time gaze estimator based on driver's head orientation for forward collision warning system. IEEE Transactions on Intelligent Transportation Systems, 12(1): 254-267, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.566, + 0.893, + 0.62 + ], + "angle": 0, + "content": "[24] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.621, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[25] Prajval Kumar Murali, Mohsen Kaboli, and Ravinder Dahiya. Intelligent in-vehicle interaction technologies. Advanced Intelligent Systems, 4(2):2100122, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.663, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[26] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In The European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[27] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In The European Conference on Computer Vision, pages 747-763. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[28] Akshay Rangesh, Bowen Zhang, and Mohan M Trivedi. Driver gaze estimation in the real world: Overcoming the eyeglass challenge. In The IEEE Intelligent Vehicles Symposium (IV), pages 1054-1059. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[29] Yusuke Sugano, Yasuyuki Matsushita, and Yoichi Sato. Learning-by-synthesis for appearance-based 3d gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[30] Ashish Tawari, Kuo Hao Chen, and Mohan M Trivedi. Where is the driver looking: Analysis of head, eye and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1564" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "iris for robust gaze zone estimation. In 17th International IEEE conference on intelligent transportation systems (ITSC), pages 988-994. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[31] Sourabh Vora, Akshay Rangesh, and Mohan Manubhai Trivedi. Driver gaze zone estimation using convolutional neural networks: A general framework and ablative analysis. IEEE Transactions on Intelligent Vehicles, 3(3):254-265, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[32] Yafei Wang, Guoliang Yuan, Zetian Mi, Jinjia Peng, Xueyan Ding, Zheng Liang, and Xianping Fu. Continuous driver's gaze zone estimation using rgb-d camera. Sensors, 19(6): 1287, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[33] Yaoming Wang, Yangzhou Jiang, Jin Li, Bingbing Ni, Wenrui Dai, Chenglin Li, Hongkai Xiong, and Teng Li. Contrastive regression for domain adaptation on gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19376-19385, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.388 + ], + "angle": 0, + "content": "[34] Guoliang Yuan, Yafei Wang, Huizhu Yan, and Xianping Fu. Self-calibrated driver gaze estimation via gaze pattern learning. Knowledge-Based Systems, 235:107630, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[35] Mingfang Zhang, Yunfei Liu, and Feng Lu. Gazeonce: Real-time multi-person gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4197-4206, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In The IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2299-2308, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[38] Xucong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2018. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[39] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):162-175, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[40] Xuong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In The European Conference on Computer Vision, 2020. 1, 2, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1565" + } + ] +] \ No newline at end of file diff --git a/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_origin.pdf b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c11571bb13d374a7307ea471a5dbe4971060ec6 --- /dev/null +++ b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/d9745d3f-8ba4-4f6a-bd07-c7535881512b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28b0046780691c9f107729de5aa379f807a1e0a4cca5366926591e25d5867e79 +size 1710329 diff --git a/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/full.md b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a28cbe9137b6310004048a98553195a4f7964e1f --- /dev/null +++ b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/full.md @@ -0,0 +1,315 @@ +# What Do You See in Vehicle? Comprehensive Vision Solution for In-Vehicle Gaze Estimation + +Yihua Cheng $^{1}$ Yaning Zhu $^{2}$ Zongji Wang $^{3}$ Hongquan Hao $^{4}$ Yongwei Liu $^{4}$ Shiqing Cheng $^{4}$ Xi Wang $^{4}$ Hyung Jin Chang $^{1}$ $^{1}$ University of Birmingham $^{1}$ , Huazhong University of Science and Technology $^{2}$ $^{2}$ NIST, Chinese Academy of Sciences $^{3}$ , CalmCar $^{4}$ +y.cheng.2@bham.ac.uk, h.j.chang@bham.ac.uk + +# Abstract + +Driver's eye gaze holds a wealth of cognitive and intentional cues crucial for intelligent vehicles. Despite its significance, research on in-vehicle gaze estimation remains limited due to the scarcity of comprehensive and well-annotated datasets in real driving scenarios. In this paper, we present three novel elements to advance in-vehicle gaze research. Firstly, we introduce IVGaze, a pioneering dataset capturing in-vehicle gaze, collected from 125 subjects and covering a large range of gaze and head poses within vehicles. In this dataset, we propose a new vision-based solution for in-vehicle gaze collection, introducing a refined gaze target calibration method to tackle annotation challenges. Second, our research focuses on in-vehicle gaze estimation leveraging the IVGaze. In-vehicle face images often suffer from low resolution, prompting our introduction of a gaze pyramid transformer that leverages transformer-based multilevel features integration. Expanding upon this, we introduce the dual-stream gaze pyramid transformer (GazeDPTR). Employing perspective transformation, we rotate virtual cameras to normalize images, utilizing camera pose to merge normalized and original images for accurate gaze estimation. GazeDPTR shows state-of-the-art performance on the IVGaze dataset. Thirdly, we explore a novel strategy for gaze zone classification by extending the GazeDPTR. A foundational tri-plane and project gaze onto these planes are newly defined. Leveraging both positional features from the projection points and visual attributes from images, we achieve superior performance compared to relying solely on visual features, substantiating the advantage of gaze estimation. The project is available at https://yihua.zone/work/ivgaze. + +# 1. Introduction + +Understanding driver intention and behavior based on driver gaze is in high demand in intelligent vehicles, facilitating + +![](images/3605938badceae03d6633d6fc61a398963c10a0f843780a50db3f42012ef2c82.jpg) +Figure 1. In-vehicle gaze estimation illustration. The driver's gaze direction is estimated based on the facial images captured by the camera behind the steering wheels. + +diverse applications such as in-vehicle interaction [1, 2, 25] and driver monitor systems [17, 21, 22]. Recent advances in vehicle gaze estimation concentrate primarily on gaze zone estimation [16, 18, 31, 32]. These approaches define multiple coarse regions, such as side mirrors and windshields, and conduct classification based on face images. + +Gaze estimation serves as an upstream task of gaze zone estimation and can offer more precise information to understand driver attention. However, these methods typically require a large-scale dataset for training. Although there are numerous gaze datasets, collected in indoor [39, 40] or outdoor [20] environments, their applicability to the vehicle environment is limited due to the different environments and camera settings, resulting in suboptimal performance. Creating an in-vehicle gaze dataset proves challenging due to the confined and irregular nature of the vehicular environment. Constructing in-vehicle gaze collection systems remains an unsolved issue, as traditional gaze collection systems are impractical for use within vehicles. The absence of in-vehicle gaze datasets acts as a significant barrier to the progress of in-vehicle gaze estimation. + +In this paper, we present a comprehensive vision-based + +in-vehicle gaze estimation research: a novel vision-based gaze collection system for vehicles, offering a first-of-its-kind in-vehicle gaze dataset, a dual-stream gaze pyramid transformer for accurate in-vehicle gaze estimation, and its extension to gaze zone classification, showcasing its effectiveness in enhancing gaze estimation. + +First, we introduce IVGaze, an in-vehicle gaze dataset collected from 125 subjects. IVGaze provides a dense distribution of gaze directions, covering a wide range within in-vehicle environments. It contains various conditions, including diverse head poses, eye movements, illumination variations, and the presence of face accessories such as glasses, sunglasses, and masks. To collect IVGaze, we propose a vision-based gaze collection system. The system does not require dedicated eye-tracking devices and is easy to reproduce. We use a single camera for facial appearance capture and paste stickers in vehicles as gaze targets. However, a significant challenge lies in calibrating the 3D position of gaze targets. This challenge arises because gaze targets are out of the camera's field of view (out-of-FoV). To address this issue, we present a refined gaze target calibration method, which leverages an auxiliary camera to capture gaze targets and a transparent chessboard for calibration. + +Secondly, we explore in-vehicle gaze estimation using the IVGaze. Face images often suffer from low resolution due to the inherent limitations of cameras in vehicles. We propose a gaze pyramid transformer that utilizes a transformer to integrate multilevel features. Expanding upon this, we propose a dual-stream gaze pyramid transformer (GazeDPTR). We rotate virtual cameras via perspective transformation to normalize images, and leverage camera pose to merge normalized and original images. GazeDPTR shows state-of-the-art results on IVGaze. + +Thirdly, we extend GazeDPTR for the downstream gaze zone classification task. It is challenging to compute the intersection of gaze and the vehicle. We define a foundational tri-plane and project gaze to the tri-plane. We extract positional features from intersection points and predict gaze zones using both positional features and visual features from face images. Our experiment demonstrates that the gaze zone classification can be further enhanced by positional features, showing the advantage of gaze estimation. + +# 2. Related Works + +# 2.1. Gaze Data Collection + +The human gaze is inherently implicit and poses a challenge for objective measurement, making gaze annotation complex in gaze data collection. Some methods capture the human gaze through intrusive devices such as eye-tracking glasses [13, 19]. However, the eye-tracking glasses have a notable impact on the quality of the captured facial images. Fischer et al. [13] attempted to mitigate this impact + +Table 1. The comparison of gaze estimation datasets. Gaze annotation is challenging in the vehicle environment, which results in existing in-vehicle datasets only providing gaze zone annotation. Our work addresses this issue and contributes the first in-vehicle gaze dataset containing gaze annotation and natural face images. + +
DatasetsEnv.# Sub.Annotation
GazeZone
Gaze360 [20]Outdoor238✓-
MPIIGaze [39]15✓-
EyeDiap [15]Indoor16✓-
EVE [27]54✓-
ETH-XGaze [40]110✓-
Vora et al. [31]10✗✓
Jha & Busso [18]16✗✓
Wang et al. [32]Vehicle3✗✓
Rangesh et al. [28]13✗✓
Ghosh et al. [16]338✗✓
IVGaze (Ours)125✓✓
+ +using GAN, but there are also some artifacts in the resulting face images. Vision-based gaze collection systems typically define gaze as direction vectors originating from facial centers towards specific gaze targets [27, 39, 40]. However, these gaze targets often remain outside the camera's field of view. This out-of-field view challenge complicates the calibration of gaze target positions. Zhang et al. [39, 40] sets screen points as gaze targets and uses a mirror to calibrate the screen plane. Kellnhofer et al. [20] uses $360^{\circ}$ panoramic cameras to capture gaze targets and human faces simultaneously where the panoramic camera is pre-calibrated. However, these strategies are not applicable to vehicles. + +In-vehicle gaze dataset usually defines different region such as windshield and left/right mirror in vehicles and perform gaze zone classification [12, 14, 16, 18, 23, 30, 34]. Kasahara et al. [19] collects an in-vehicle gaze dataset but subjects are required to wear eye-tracking glasses, which means the dataset is not applicable in the real world. Our gaze collection system does not require dedicated devices and produces natural face images. + +# 2.2. Appearance-based Gaze Estimation + +Appearance-based gaze estimation directly learns mapping function from facial appearance to human gaze [10]. Conventional gaze estimation methods extract eye features from eye images [7, 9, 26, 36]. They concatenate the head pose vector with eye features for gaze estimation. Recent methods directly learn gaze from face images [3, 4, 11, 33, 35]. Face images provide both eye region and head pose information, resulting in superior performance compared to methods relying solely on eye images. However, the subtlety of the human eye in face images poses a challenge as it can be easily overlooked by the network. Zhang et al. [37] utilize a learnable attention map to guide the net + +![](images/05326171b8c2d985f03644cbebb4e804555a48e2d148099086e68565d0a34d94.jpg) +(a) + +![](images/53a7a6a1e467b62214bd05f80e3fbe76bf19fc501c5ba6588b97fdfc0dc42a7d.jpg) +(b) +(c) +(d) +Figure 2. We construct a vision-based in-vehicle gaze collection system comprising a DMS camera, a depth camera, and strategically placed gaze targets, as depicted in (c). The DMS camera is positioned behind the steering wheel to capture drivers' facial appearances, while the gaze targets, positioned beyond the DMS camera's field-of-view (FoV), such as on the windshield, remain unobserved. The depth camera, utilized for calibration purposes, is temporarily installed for capturing gaze target positions in 3D with respect to its own coordinates, and it is removed during data collection. To facilitate the calibration of the depth camera's pose relative to the DMS camera, we propose employing a transparent chessboard, which is placed between the two cameras. + +work's focus specifically on the eye. Cheng et al. [11] crop eye images from face images and construct a cascade network to leverage them. In recent developments, transformers have showcased remarkable capabilities in gaze estimation. Cheng et al. [5] demonstrate that transformers can achieve state-of-the-art performance across various benchmarks. They apply transformers to address dual-view gaze estimation and show that dual-view images outperform single-view images [6]. + +# 3. In-Vehicle Gaze Data Collection System + +In this section, we introduce our vision-based system for in-vehicle gaze data collection, eliminating the need for dedicated eye-tracking devices. Our system includes a neat gaze target calibration method that effectively addresses the core challenge in gaze annotation. Leveraging this system, we collect the first in-vehicle gaze dataset from 125 subjects. The dataset spans a diverse range of gaze, facilitating and advancing future research in in-vehicle gaze estimation. + +# 3.1. System Setup + +We place gaze targets within the vehicle, and subjects are instructed to look at each designated target during the data collection process. Simultaneously, we record their facial appearance along with corresponding gaze targets. + +Camera. Our system uses one camera of a driver monitor system (DMS) for facial appearance capture. The DMS camera is an infrared camera with a capture resolution of $1280 \times 800$ . The camera is located behind the steering wheel and directly points at the head region of drivers. + +Gaze targets. We mark gaze targets with red points on stickers, and each target is uniquely labeled with a distinct number printed on stickers to aid differentiation. We strate- + +gically position these stickers within various gaze regions within the vehicle, including the windshield, left and right-side mirrors, rear-view mirror, center console, speedometer, handbrake, and etc. These targets cover a large gaze region, satisfying the gaze estimation requirements in vehicles. + +# 3.2. Collection Procedure + +We meticulously design the collection procedure to minimize error. During the collection process, participants are instructed to look at specific gaze targets in a predefined sequence. They are also required to speak the corresponding number associated with each target to confirm their focus accuracy. Participants must maintain focus on the target for 3 seconds to facilitate image capture. We conduct a preliminary check on the captured images and discard any erroneous ones. Once this verification is complete, the current data collection phase concludes, and participants are instructed to shift their gaze to the next gaze targets. + +We design three postures for each participant. Participants should complete the collection process three times with different postures. The first posture requires participants not to change head pose. They should preserve the same head pose and only move eyeballs to focus a gaze target. We aim to collect sufficient eye movement data with this posture. We do not require participants to focus on the target where participants think it is hard to focus with only eye movement, e.g., the target in the side mirror. Participants can rotate their heads to look at these points but they are also required to preserve the new head pose until they cannot look at any targets with the current head pose. The second posture requires participants not to move their head position but they can rotate their heads to focus a gaze target. We do not set any constraints in the third posture. Participants can freely perform head movements in this pos + +![](images/2870d1c47d06ce53cf6858d51f7ccad31f2f5d7b3058ed26afe593d8380b2b4f.jpg) +Figure 3. Our dataset is collected using IR cameras in the vehicle environment. (a) We present image samples of IVGaze, highlighting the challenges posed by realistic in-vehicle conditions, including cases with sunglasses and reflections in glasses. (b) We categorize the image count based on their mean pixel value, showing the diversity of illumination conditions. (c) The image count is analyzed based on face accessories including glasses, sunglasses, and masks. + +![](images/a885f36b2846195ca82470208a05578eb83ec9da22241f7d01ef4fb96e2a0360.jpg) + +![](images/8c25bef82c6cf83982cf40d0a67ec6db2599fa66e99b69270499d00a6b301bf3.jpg) + +ture. The last two postures bring head pose variation for IVGaze. We do not use devices such as headrests to constrain participants. Participants only need to preserve the specific posture and focus on targets in a comfortable way. + +# 3.3. Gaze Annotation + +Vision-based gaze collection systems define gaze as unit direction vectors originating from the face center $\mathbf{o}$ and extending toward gaze targets $\mathbf{t}$ [10]. Our system follows this framework and the gaze annotation can be computed as $\mathbf{g} = (\mathbf{t} - \mathbf{o}) / ||\mathbf{t} - \mathbf{o}||_2$ . The gaze annotation is decomposed into 3D face center estimation and gaze target calibration. + +3D face center estimation has been effectively addressed in previous research. We first detect facial landmarks and fit a 3D morphable face model for 3D facial landmarks [24, 40]. We select the 3D position of the nose as gaze origin o. However, the calibration of gaze target positions presents a challenge because gaze targets are out-of-FoV. It means that the DMS camera cannot capture any images of the gaze targets, complicating the calibration problem. + +# 3.4. Gaze Target Calibration + +We need the 3D gaze target position $\mathbf{t}$ for gaze annotation. However, a key problem is all gaze targets are out-of-FoV. Conventional gaze datasets often present gaze targets on a screen and employ a mirror to solve the out-of-FoV problem, where the mirror reflects the content of the screen [36]. However, these methods are not suitable for vehicle scenarios and can lead to a substantial accumulation of errors. + +In this section, we propose a neat and efficient gaze target calibration method. Our basic idea is to employ an auxiliary camera to capture gaze targets. This allows a straightforward calculation of the 3D gaze target positions w.r.t. the auxiliary camera coordinate system. However, it also introduces a new challenge: how to calibrate extrinsic matrix of the auxiliary cameras w.r.t. DMS camera? Stereo calibration is typically used to compute the pose of two cam + +eras. However, it requires two captured images sharing corresponding points. In our system, gaze targets are located behind the DMS camera which means the DMS camera and auxiliary camera should be oriented in opposite directions, making traditional stereo calibration infeasible. + +To solve this problem, we propose to use a transparent chessboard for the pose calibration. We set a transparent chessboard between the two cameras and the two cameras respectively capture each side of the chessboard. We first calibrate the two camera poses w.r.t. the chessboard coordinate system and then compute the pose between two cameras. It is worth noting that there are two different chessboard coordinate systems corresponding to each side of the chessboard. We can derive the transformation matrix between the two chessboard coordinate systems where $\mathbf{R}_{\mathrm{chess}} = \mathrm{diag}(0,0,-1)$ and $\mathbf{t}_{\mathrm{chess}} = (0,0,-\mathrm{d})$ . The d is the thickness of the chessboard. + +In detail, we use an Intel RealSense D435 depth camera as the auxiliary camera since it can provide accurate 3D positions. We show the camera layout in Fig. 2(b). We can obtain the transformation matrix $\mathbf{R}_{\mathrm{dms}}$ , $\mathbf{t}_{\mathrm{dms}}$ and $\mathbf{R}_{\mathrm{depth}}$ , $\mathbf{t}_{\mathrm{depth}}$ which can transfer a point from chessboard coordinate systems to camera coordinate systems. We have + +$$ +\mathbf {R} _ {\mathrm {r o t}} = \mathbf {R} _ {\mathrm {d m s}} \mathbf {R} _ {\mathrm {c h e s s}} \mathbf {R} _ {\mathrm {d e p t h}} ^ {- 1}, \tag {1} +$$ + +$$ +\mathbf {t} _ {\text {r o t}} = - \mathbf {R} _ {\mathrm {d m s}} \mathbf {R} _ {\text {c h e s s}} \mathbf {R} _ {\text {d e p t h}} ^ {- 1} \mathbf {t} _ {\text {d e p t h}} + \mathbf {R} _ {\mathrm {d m s}} \mathbf {t} _ {\text {c h e s s}} + \mathbf {t} _ {\mathrm {d m s}}. \tag {2} +$$ + +We use these matrices to convert points from the depth camera coordinate system into the DMS camera coordinate system. Please refer to the supplementary material for details. + +# 3.5. In-Vehicle Gaze Dataset + +We collect the first in-vehicle gaze dataset IVGaze which provides dense gaze annotation and natural face images. We show the dataset samples in Fig. 3 (a). + +Dataset Statistics. We collect 44,705 images from 125 subjects. The number of images per subject ranges from + +![](images/6ac84c8c9ac9218a7dd53f0a820ebc628d22089999853487859601635fcce461.jpg) +Figure 4. We show the distribution of data for gaze (left) and head movements (right). Brighter regions denote higher data density. + +![](images/d2203024d1579751e1ce55940485aa68ef0c9a015915d53d07494b80feb46ddc.jpg) + +169 to 946. Most subjects provide around 300 images per person, i.e., they repeat the collection procedure three times with three different postures. 40 subjects are female and 85 subjects are male, with ages spanning from 20 to 50 years. + +Collection Conditions. The dataset was collected between 9 am and $7\mathrm{pm}$ in outdoor environments, covering a wide range of lighting conditions. We also collect images in indoor environments, i.e., a garage. The dataset statistics across pixel intensity are shown in Fig. 3 (b). + +Face Accessories. We consider two face accessories during the collection: glasses and masks. For subjects who did not wear glasses, we provided glasses and asked them to repeat the collection wearing glasses. We provided white and black masks for some subjects and asked them to repeat the collection wearing masks. We also required a few subjects to wear sunglasses to facilitate future research. We show the statistics in Fig. 3 (c). + +Data Distribution. We plot the angular distribution of gaze and head pose. Note that the data normalization method (introduced in the next section) changes gaze directions and head poses with a rotation matrix. Fig. 4 shows the distribution in normalization space. The horizontal gaze is from $-50^{\circ}$ to $90^{\circ}$ , and the vertical gaze is from $-40^{\circ}$ to $40^{\circ}$ . It is worth noting that the vehicle environment has limited space, and our dataset already covers almost all regions. Our dataset is collected in the real vehicle environment, and the driver seat in the vehicle is located on the left. Therefore, our dataset has a relatively small distribution on the left. The right figure in Fig. 4 shows the head pose distribution. Our dataset contains a large range of head poses in the yaw axis, as subjects need to look at gaze targets from the left-side mirror to the right-side mirror. + +# 4. In-Vehicle Gaze Estimation + +The in-vehicle environment brings new settings and challenges. In this section, we systematically explore in-vehicle gaze estimation. We introduce a revised data preprocessing method and describe a novel gaze estimation network. We also extend the network for an application and prove the advantage of in-vehicle gaze estimation. + +# 4.1. Revisiting Data Normalization in Vehicle + +Data normalization rotates and translates cameras via perspective transformation in face images to reduce head pose variation. Conventional methods usually define the rotated camera coordinate system $\mathbf{C}_r$ based on head pose [29, 38]. However, these methods cannot bring performance improvement in the vehicle environment. + +The $x$ -axis of $\mathbf{C}_r$ is typically defined as the $x$ -axis of head coordinate systems. It means they will rotate images to keep the head straight. We observed that the designed $x$ -axis will produce unstable results, especially in the extreme head pose. Therefore, we propose that do not rotate virtual cameras based on the $x$ -axis of head coordinate systems. In detail, we compute the $z$ -axis of $\mathbf{C}_r$ as direction vectors from cameras to face centers. We use the $x$ -axis of the virtual camera, i.e., $(1,0,0)$ rather than the head coordinate system. We compute the $y$ -axis as $y = z\times x$ and also recompute the $x$ -axis as $x = y\times z$ . Finally, we have a rotation matrix $\mathbf{R} = [x;y;z]$ and use a scale matrix $S$ to maintain the user-camera distance. We warp the image based on matrix SR. The human gaze is also changed as $\mathbf{g}^n = \mathbf{R}\mathbf{g}^o$ , where $\mathbf{g}^n$ and $\mathbf{g}^o$ denotes gaze in the normalization and origin space. + +# 4.2. Gaze Pyramid Transformer + +Gaze estimation methods typically learn gaze from face images. However, face images captured in vehicles often suffer from low resolution owing to the limited capabilities of the camera. This low resolution can lead conventional methods to overlook subtle details, especially in the eye region. Some techniques address this challenge by cropping eye images and processing them separately [8], but these approaches encounter difficulties when faced with extreme head poses, as one of the eye images may not be visible. + +We propose a gaze pyramid transformer (GazePTR) in this section. We build a feature pyramid and input multilevel features into a transformer for gaze estimation. As shown in Fig. 5, our network uses a convolution network to extract features from facial images. We collect feature maps from different levels rather than the last level. We use $1 \times 1$ convolution layers and global average pooling layers to preserve the same feature dimension. Finally, we input the multi-level feature into a transformer. A learnable token is used to aggregate multi-level features for gaze estimation. + +However, the network has the same performance achieved by directly utilizing the last feature map for gaze estimation. This result can be attributed to the network's tendency to overlook low-level features. To address this issue, we introduce additional constraints by estimating gaze from each level of features and calculating corresponding loss. It can be understood that we first extract multi-level features, with each feature being potentially useful. A transformer is used to effectively ensemble these features. + +![](images/65ba7fefcea69103c4bd47f4963e9bfb6bc14be0a0e5033533b2547bc9d1c9c4.jpg) +Figure 5. The GazeDPTR directly crop face for origin images and rotates virtual cameras via perspective transformation for normalized images. It builds a dual-stream network to extract features from the two images based on the GazePTR for feature extraction which integrates multi-level features via transformers. To further merge the features from two streams, we leverage a transformer where camera pose is used as the positional feature in the transformer. We define the original camera pose as $\mathbf{C} = \text{diag}(1, 1, 1)$ and the camera pose in normalization space is RC. We also extend the network for gaze zone classification. We define a tri-plane and project gaze into them. We extract positional features from three intersection points via a transformer. We also extract visual features from images and predict the gaze zone based on both visual features and positional features. The whole network is trained in an end-to-end manner. + +# 4.3. Dual-Stream Gaze Pyramid Transformer + +Dual-view images have been demonstrated to outperform single-view images in gaze estimation [6]. In our work, we rotate cameras in data normalization through perspective transformation. This inspired us to formulate the dual-camera setting leveraging both original and normalized images. However, dual-view images improve accuracy since they contain more visual information. It remains uncertain whether the combination of normalized and original images can provide additional insights beyond what each offers. + +To validate our hypothesis, we executed an oracle baseline. We separately train GazePTR on the original and the normalized datasets, selecting the best result from each image pair. The result reveals a significant performance improvement, validating the inherent advantages of utilizing both images. Please refer to the supplementary for details. + +Therefore, we propose a dual-stream gaze pyramid transformer (GazeDPTR). Our method leverages both normalized images and original images for gaze estimation. We employ GazePTR to extract features from the two images and a transformer is used to integrate the two features for gaze estimation. Note that, the two features are in different camera coordinate systems and correspond to different gaze. To establish the connection, we use the camera pose as the positional information in the transformer. However, it is impossible the calibrate a virtual camera. We define the camera pose of original images as $\mathbf{C} = \text{diag}(1, 1, 1)$ and the camera pose of normalized images is RC. We extract the $z$ -axis and use positional encoding for camera pose [6]. + +# 4.4. Strategy for Gaze Zone Classification + +Detecting driver attention is essential in driver monitoring systems. In this section, we introduce a novel strategy to + +extend GazeDPTR for gaze zone classification. GazeDPTR estimates gaze direction from face images. While the conventional solution involves projecting gaze direction and computing intersections within the vehicle, it is not practical without a 3D vehicle model. Zhang et al. use a triplane for self-supervised loss function [35]. Our alternative strategy defines a foundational tri-plane and computes the intersection of gaze vectors with this tri-plane. We extract positional features from three intersection points to facilitate gaze zone classification. + +In detail, we define a tri-plane with normal vectors $(1,0,0)$ , $(0,1,0)$ , and $(0,0,1)$ , intersecting at the origin $(0,0,0)$ . Given the estimated gaze vector $\mathbf{g}^n$ , we convert it back to the original space using $\mathbf{g}^o = \mathbf{R}^{-1}\mathbf{g}^n$ . We then project $\mathbf{g}^o$ onto the tri-plane with the face center $\mathbf{o}$ and obtain three intersection points. To enhance the positional features, we apply positional encoding to these points and input them into a 2-layer transformer. The positional encoding not only increases feature dimensions but also normalizes them within the range of -1 to 1. Additionally, we use another learnable token to aggregate visual features from the original images. Gaze zone classification is performed using both visual features and positional features. + +We train the whole network in an end-to-end manner. We use L1 loss for gaze estimation and cross-entropy loss for gaze zone classification. Please refer to the supplementary for implementation details. + +# 5. Experiment + +Dataset: IVGaze contains 44,705 images of 125 subjects. To perform within-dataset evaluation, we divide the dataset into three subsets based on subjects. The image numbers of the three subsets are 15,165, 14,674, and 14,866. We + +Table 2. Performance comparison in gaze estimation. Our methods achieve better performance than the compared methods. + +
Angular ErrorAverage Precision (AP)
<2°<4°<6°<8°
FullFace [37]13.67°2.3%8.8%17.8%28%
DWG [16]8.82°6.6%21.7%38.1%53.2%
Gaze360 [20]8.15°9.2%27.3%44.6%58.9%
FullFace+ [37]7.48°14.2%31.1%46.7%63.1%
GazeTR [5]7.33°17%32.8%47.5%64.7%
XGaze [40]7.06°11.7%32.7%51.5%66.7%
GazePTR7.04°17.6%34%49.3%66.7%
GazeDPTR6.71°22.1%36%50.3%68.4%
+ +perform three-fold cross-validation on our dataset. + +Evaluation Metric: We use the angular error as the evaluation metric of gaze estimation as most of the methods [10], where a smaller value represents a better model. However, we notice the angular error only shows the overall performance while cannot give deep insights. Therefore, we define the average precision (AP) where AP of $< k^{\circ}$ means an estimation is considered correct if the angular error is lower than $k^{\circ}$ . Regarding the gaze zone classification, average precision is used as a common multi-class classification. + +# 5.1. Comparison with SOTA Methods + +We first compared our methods with SOTA methods in the in-vehicle gaze estimation task. We compared our methods with FullFace [37], Gaze360 [20], XGaze [40], DWG [16] and GazeTR [5]. We replaced the backbone of the FullFace method from AlexNet to ResNet18 for a more convincing comparison. We denote the new method as FullFace+. + +The result is shown in Tab. 2. GazePTR and GazeDPTR both show better performance than the compared methods. GazePTR and GazeTR have the same backbone and transformer architecture. However, GazePTR outperforms $0.29^{\circ}$ thanGazeTR since GazePTR uses multi-level feature maps. GazeDPTR builds a dual-stream network and further brings $0.33^{\circ}$ improvement than GazePTR. These results show the advantage of our methods. + +Interestingly, XGaze has a similar angular error to GazePTR. However, it is significantly worse than GazePTR in AP of $< 2^{\circ}$ . It is because that XGaze uses ResNet50 as the backbone. The deep backbone enhances the feature extraction ability but also easily overlooks the small eye region, making it challenging to achieve precise gaze estimation. + +# 5.2. The Impact of Face Accessories + +IVGaze provides rich samples of wearing face accessories, including glasses, masks, and sunglasses. We also evaluate their impacts on the accuracy of gaze estimation and show in Tab. 3. The result shows that eyeglasses have a + +Table 3. We show the impacts of different face accessories on performance. The sunglasses bring a significant performance drop. + +
GlassesMaskSunglasses
withw/owithw/owith
FullFace [37]14.43°12.40°15.20°13.35°21.39°
DWG [16]9.20°8.19°9.43°8.69°17.43°
Gaze360[20]8.30°7.91°8.95°7.99°17.99°
FullFace+ [37]7.59°7.30°8.37°7.30°16.50°
XGaze [40]7.07°7.03°7.80°6.90°15.15°
GazeTR [5]7.40°7.22°8.12°7.17°17.49°
GazePTR7.13°6.90°7.78°6.89°16.54°
GazeDPTR6.77°6.63°7.44°6.57°16.41°
+ +![](images/f791c54df84bfd40997e2f380edc5b5e5dc237cc0175e878e26be2b5d4d4fd85.jpg) +Figure 6. We compute the angular degree between gaze direction/head orientation with the frontal direction, i.e., $(0,0,-1)$ , and count the mean gaze estimation accuracy in different ranges. Interestingly, GazePTR brings significant performance improvement to GazeTR in the $0^{\circ} - 20^{\circ}$ gaze range. This is because subjects prefer to move their eyeball rather than their head when the gaze target is located at a close range. GazePTR utilizes multilevel features and can accurately capture eye movement. + +![](images/77c89d723cf5f7a5e4c5b674946d7f369db0b1838b8629bf259fccea5abfa316.jpg) + +large impact if the basic performance is not good. However, as the basic performance improves, the performance difference reduces. This is because most of glasses are transparent, enabling a robust model to extract features effectively. Sunglasses have a significant impact on performance, with average performance exceeding $15^{\circ}$ . Masks have a substantial impact on performance, as they obscure facial regions, complicating the extraction of head-related features. + +# 5.3. Performance Distribution + +We show the performance distribution to provide deep insight. We first show the performance distribution in different gaze ranges. We introduce a novel concept, the gaze range of $a^{\circ} - b^{\circ}$ , which is defined as a set of samples satisfying the requirement that the angular degree between the gaze direction and the frontal gaze direction is in the range. The result is shown in Fig. 6. Interestingly, GazePTR and GazeDPTR have better performance in $0^{\circ} - 20^{\circ}$ . This is because subjects prefer to move their eye rather than head to look at objects in this range. Our methods leverage multi-level features and easily capture eye movement. We also show the performance distribution in different head ranges. + +Table 4. We present the performance of GazePTR at each level of the feature. The higher-level feature achieves better performance. Our method integrates features from all levels and achieves the best performance. + +
1st2nd3rd4thOurs
GazePTR10.33°8.59°7.61°7.35°7.04°
+ +Table 5. We present the performance of GazePTR separately training on the original images and normalized images. GazeDPTR leverages camera pose to establish a connection between the two images and use both two images for gaze estimation. We also removed the camera pose to perform an ablation study. + +
Original imagesNormalized imagesGazeDPTR (w/o camera pose)GazeDPTR
Acc7.44°7.04°7.03°6.71°
+ +Our methods are robust in different head ranges where the maximum performance difference is only $1^{\circ}$ . + +# 5.4. Ablation Study + +GazePTR leverages multi-level features for gaze estimation. We first conduct an ablation study to demonstrate the advantage. We obtain the gaze performance of GazePTR at each level of feature and show it in Tab. 4. The result demonstrates a higher-level feature achieves better performance. GazePTR integrates multi-level features and achieves the best performance. The result proves the advantage of our design. + +GazeDPTR leverages both original and normalized images along with their corresponding camera poses. To demonstrate the effectiveness of our design, we perform an ablation study in Tab. 5. Training GazePTR separately on original and normalized images yields performances of $7.44^{\circ}$ and $7.04^{\circ}$ respectively. GazeDPTR, which incorporates both image types, achieves a performance improvement of $0.34^{\circ}$ . We also experimented by excluding the positional encoding of the camera pose in GazeDPTR. Interestingly, the result is comparable to using only normalized images. + +# 5.5. Additional Experiments + +Data Normalization. We compare our method with the data normalization method [38] in the left of Tab. 6. The previous method cannot work well in vehicle environments while our method brings $0.3^{\circ}$ performance improvement. + +Case Study. We visual gt and our prediction in Fig. 7. + +# 5.6. Gaze Zone Classification + +We propose a basis tri-plane to acquire positional feature and combine both positional feature and visual features for gaze zone classification. We respectively evaluate each feature and show the result in the right of Tab. 6. The visual feature achieves better performance than the positional feature since the positional feature is obtained from gaze pro + +Table 6. We evaluate different data normalization methods in the left table. Our method outperforms the previous method [38]. In the right table, we evaluate different features in the gaze zone classification. Our method uses both positional feature and visual features, achieving $2.4\%$ improvement over visual features. + +
Normalization [38]oursGazeTRGazePTRVisual FeaturePositional FeatureAP
××7.77°7.44°✓79.4%
✓8.64°8.53°✓75.3%
✓7.33°7.04°✓✓81.8%
+ +![](images/7c74a154f013274f2bc0b275fa40e5817ef46509a6250cf9368aed0da587aca0.jpg) + +![](images/5deb96d370044fb41499b0012f948c4a7593447cedbf57c49d285d762735f14a.jpg) + +![](images/e1688cd8b86a1f5c59544da0c7e3e9af73fe7a36d888b5a3eef616d3ff49b11e.jpg) + +![](images/e31a932506a0c0240de6a0e51f540b0d4a606164062623b93a4cf8417c1e09eb.jpg) + +![](images/9647d81b316851b2ef573ca3f6839f399e2460907b3eed813c173f8d35ade647.jpg) +Figure 7. We illustrate GT with green lines and Prediction with red lines. The number shows the angular error in degree. The right two figures are failure cases since the eye is hard to capture. + +![](images/50a223a2104ebf1e8a768ca9c88c872432c8e42ad342478cccb04a0b585a34f4.jpg) + +![](images/7e27d54d41d9c2cabf8a652684d9e86a5f9c4b8ef67f66e838e929c33a6fa38e.jpg) + +![](images/4488432ef6d3d8cd112629dd8c585833d8c2679102c30b9459937fb21ddaf9ce.jpg) + +jection. The projection increases the model interpretability but decreases the fitting ability. Our method uses the two features and gets $2.4\%$ performance improvement. The result demonstrates that the gaze zone classification task could be enhanced by gaze direction cues, highlighting the advantage of our gaze estimation work. + +# 6. Conclusion + +In this work, we provide systematical research in in-vehicle gaze estimation, including dataset, algorithm, and extensive application. We first solve the data collection issue in a vehicle where a gaze target calibration method is proposed. We collect the first in-vehicle gaze dataset containing dense gaze annotation and natural face images. We further explore the algorithm for in-vehicle gaze estimation. Our work brings two deep insights 1) multi-level feature is useful to capture eye region information. 2) simultaneously leveraging original images and normalized images could achieve better performance. We also extend our work for the downstream gaze zone classification task. We demonstrate that the gaze direction cues could bring performance improvement for gaze zone classification. + +# 7. Acknowledgments: + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00608, Artificial intelligence research about multi-modal interactions for empathetic conversations with humans) + +# References + +[1] Abdul Rafey Aftab. Multimodal driver interaction with gesture, gaze and speech. In 2019 International Conference on Multimodal Interaction, pages 487-492, 2019. 1 +[2] Aya Ataya, Won Kim, Ahmed Elsharkawy, and SeungJun Kim. Gaze-head input: Examining potential interaction with immediate experience sampling in an autonomous vehicle. Applied Sciences, 10(24), 2020. 1 +[3] Yiwei Bao, Yunfei Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with rotation consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4207-4216, 2022. 2 +[4] Xin Cai, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Source-free adaptive gaze estimation by uncertainty reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22035-22045, 2023. 2 +[5] Yihua Cheng and Feng Lu. Gaze estimation using transformer. ICPR, 2022. 3, 7 +[6] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20632-20641, 2023. 3, 6 +[7] Yihua Cheng, Feng Lu, and Xucong Zhang. Appearance-based gaze estimation via evaluation-guided asymmetric regression. In The European Conference on Computer Vision, 2018. 2 +[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 5 +[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2 +[10] Yihua Cheng, Haofei Wang, Yiwei Bao, and Feng Lu. Appearance-based gaze estimation with deep learning: A review and benchmark. arXiv preprint arXiv:2104.12668, 2021. 2, 4, 7 +[11] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. AAAI, 2022. 2, 3 +[12] In-Ho Choi, Sung Kyung Hong, and Yong-Guk Kim. Real-time categorization of driver's gaze zone using the deep learning techniques. In 2016 International conference on big data and smart computing (BigComp), pages 143-148. IEEE, 2016. 2 +[13] Tobias Fischer, Hyung Jin Chang, and Yiannis Demiris. Rtgene: Real-time eye gaze estimation in natural environments. In The European Conference on Computer Vision, 2018. 2 +[14] Lex Fridman, Philipp Langhans, Joonbum Lee, and Bryan Reimer. Driver gaze region estimation without use of eye movement. IEEE Intelligent Systems, 31(3):49-56, 2016. 2 +[15] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2014. 2 + +[16] Shreya Ghosh, Abhinav Dhall, Garima Sharma, Sarthak Gupta, and Nicu Sebe. Speak2label: Using domain knowledge for creating a large scale driver gaze zone estimation dataset. In The IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2896-2905, 2021. 1, 2, 7 +[17] Amie C Hayley, Brook Shiferaw, Blair Aitken, Frederick Vinckenbosch, Timothy L Brown, and Luke A Downey. Driver monitoring systems (dms): The future of impaired driving management? Traffic injury prevention, 22(4):313-317, 2021. 1 +[18] Sumit Jha and Carlos Busso. Probabilistic estimation of the gaze region of the driver using dense classification. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 697-702. IEEE, 2018. 1, 2 +[19] Isaac Kasahara, Simon Stent, and Hyun Soo Park. Look both ways: Self-supervising driver gaze estimation and road scene saliency. In European Conference on Computer Vision, pages 126-142. Springer, 2022. 2 +[20] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In The IEEE International Conference on Computer Vision, 2019. 1, 2, 7 +[21] Muhammad Qasim Khan and Sukhan Lee. A comprehensive survey of driving monitoring and assistance systems. Sensors, 19(11), 2019. 1 +[22] Muhammad Qasim Khan and Sukhan Lee. Gaze and eye tracking: Techniques and applications in adas. Sensors, 19 (24), 2019. 1 +[23] Sung Joo Lee, Jaeik Jo, Ho Gi Jung, Kang Ryoung Park, and Jaihie Kim. Real-time gaze estimator based on driver's head orientation for forward collision warning system. IEEE Transactions on Intelligent Transportation Systems, 12(1): 254-267, 2011. 2 +[24] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 4 +[25] Prajval Kumar Murali, Mohsen Kaboli, and Ravinder Dahiya. Intelligent in-vehicle interaction technologies. Advanced Intelligent Systems, 4(2):2100122, 2022. 1 +[26] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In The European Conference on Computer Vision, 2018. 2 +[27] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In The European Conference on Computer Vision, pages 747-763. Springer, 2020. 2 +[28] Akshay Rangesh, Bowen Zhang, and Mohan M Trivedi. Driver gaze estimation in the real world: Overcoming the eyeglass challenge. In The IEEE Intelligent Vehicles Symposium (IV), pages 1054-1059. IEEE, 2020. 2 +[29] Yusuke Sugano, Yasuyuki Matsushita, and Yoichi Sato. Learning-by-synthesis for appearance-based 3d gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition, 2014. 5 +[30] Ashish Tawari, Kuo Hao Chen, and Mohan M Trivedi. Where is the driver looking: Analysis of head, eye and + +iris for robust gaze zone estimation. In 17th International IEEE conference on intelligent transportation systems (ITSC), pages 988-994. IEEE, 2014. 2 +[31] Sourabh Vora, Akshay Rangesh, and Mohan Manubhai Trivedi. Driver gaze zone estimation using convolutional neural networks: A general framework and ablative analysis. IEEE Transactions on Intelligent Vehicles, 3(3):254-265, 2018. 1, 2 +[32] Yafei Wang, Guoliang Yuan, Zetian Mi, Jinjia Peng, Xueyan Ding, Zheng Liang, and Xianping Fu. Continuous driver's gaze zone estimation using rgb-d camera. Sensors, 19(6): 1287, 2019. 1, 2 +[33] Yaoming Wang, Yangzhou Jiang, Jin Li, Bingbing Ni, Wenrui Dai, Chenglin Li, Hongkai Xiong, and Teng Li. Contrastive regression for domain adaptation on gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19376-19385, 2022. 2 +[34] Guoliang Yuan, Yafei Wang, Huizhu Yan, and Xianping Fu. Self-calibrated driver gaze estimation via gaze pattern learning. Knowledge-Based Systems, 235:107630, 2022. 2 +[35] Mingfang Zhang, Yunfei Liu, and Feng Lu. Gazeonce: Real-time multi-person gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4197-4206, 2022. 2, 6 +[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In The IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2, 4 +[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2299-2308, 2017. 2, 7 +[38] Xucong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2018. 5, 8 +[39] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):162-175, 2019. 1, 2 +[40] Xuong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In The European Conference on Computer Vision, 2020. 1, 2, 4, 7 \ No newline at end of file diff --git a/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/images.zip b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..70e583b09549939daf6809fe6d74835939426764 --- /dev/null +++ b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efa246bdf6ccb6c7b5a2e404b6a3426a0f1103d16ecbfe605c3fc098114bba3f +size 485173 diff --git a/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/layout.json b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac114034699f4e5b455cc345f41bb1eac27d1f02 --- /dev/null +++ b/2024/What Do You See in Vehicle_ Comprehensive Vision Solution for In-Vehicle Gaze Estimation/layout.json @@ -0,0 +1,8395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 486, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 486, + 138 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 486, + 138 + ], + "type": "text", + "content": "What Do You See in Vehicle? Comprehensive Vision Solution for In-Vehicle Gaze Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": "Yihua Cheng" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Yaning Zhu" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Zongji Wang" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Hongquan Hao" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Yongwei Liu" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Shiqing Cheng" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Xi Wang" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " Hyung Jin Chang" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": "University of Birmingham" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": ", Huazhong University of Science and Technology" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": "NIST, Chinese Academy of Sciences" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": ", CalmCar" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 496, + 232 + ], + "type": "text", + "content": " \ny.cheng.2@bham.ac.uk, h.j.chang@bham.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 285, + 290, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 290, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 290, + 657 + ], + "type": "text", + "content": "Driver's eye gaze holds a wealth of cognitive and intentional cues crucial for intelligent vehicles. Despite its significance, research on in-vehicle gaze estimation remains limited due to the scarcity of comprehensive and well-annotated datasets in real driving scenarios. In this paper, we present three novel elements to advance in-vehicle gaze research. Firstly, we introduce IVGaze, a pioneering dataset capturing in-vehicle gaze, collected from 125 subjects and covering a large range of gaze and head poses within vehicles. In this dataset, we propose a new vision-based solution for in-vehicle gaze collection, introducing a refined gaze target calibration method to tackle annotation challenges. Second, our research focuses on in-vehicle gaze estimation leveraging the IVGaze. In-vehicle face images often suffer from low resolution, prompting our introduction of a gaze pyramid transformer that leverages transformer-based multilevel features integration. Expanding upon this, we introduce the dual-stream gaze pyramid transformer (GazeDPTR). Employing perspective transformation, we rotate virtual cameras to normalize images, utilizing camera pose to merge normalized and original images for accurate gaze estimation. GazeDPTR shows state-of-the-art performance on the IVGaze dataset. Thirdly, we explore a novel strategy for gaze zone classification by extending the GazeDPTR. A foundational tri-plane and project gaze onto these planes are newly defined. Leveraging both positional features from the projection points and visual attributes from images, we achieve superior performance compared to relying solely on visual features, substantiating the advantage of gaze estimation. The project is available at https://yihua.zone/work/ivgaze." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 668, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Understanding driver intention and behavior based on driver gaze is in high demand in intelligent vehicles, facilitating" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 258, + 539, + 376 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 539, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 539, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 539, + 376 + ], + "type": "image", + "image_path": "3605938badceae03d6633d6fc61a398963c10a0f843780a50db3f42012ef2c82.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 378, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 378, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 546, + 411 + ], + "type": "text", + "content": "Figure 1. In-vehicle gaze estimation illustration. The driver's gaze direction is estimated based on the facial images captured by the camera behind the steering wheels." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 417, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 546, + 490 + ], + "type": "text", + "content": "diverse applications such as in-vehicle interaction [1, 2, 25] and driver monitor systems [17, 21, 22]. Recent advances in vehicle gaze estimation concentrate primarily on gaze zone estimation [16, 18, 31, 32]. These approaches define multiple coarse regions, such as side mirrors and windshields, and conduct classification based on face images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 491, + 547, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 670 + ], + "type": "text", + "content": "Gaze estimation serves as an upstream task of gaze zone estimation and can offer more precise information to understand driver attention. However, these methods typically require a large-scale dataset for training. Although there are numerous gaze datasets, collected in indoor [39, 40] or outdoor [20] environments, their applicability to the vehicle environment is limited due to the different environments and camera settings, resulting in suboptimal performance. Creating an in-vehicle gaze dataset proves challenging due to the confined and irregular nature of the vehicular environment. Constructing in-vehicle gaze collection systems remains an unsolved issue, as traditional gaze collection systems are impractical for use within vehicles. The absence of in-vehicle gaze datasets acts as a significant barrier to the progress of in-vehicle gaze estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 670, + 546, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 546, + 682 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 546, + 682 + ], + "type": "text", + "content": "In this paper, we present a comprehensive vision-based" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "text", + "content": "1Our work focuses on gaze direction estimation. We abbreviate gaze direction as gaze in the rest." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1556" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "in-vehicle gaze estimation research: a novel vision-based gaze collection system for vehicles, offering a first-of-its-kind in-vehicle gaze dataset, a dual-stream gaze pyramid transformer for accurate in-vehicle gaze estimation, and its extension to gaze zone classification, showcasing its effectiveness in enhancing gaze estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 348 + ], + "type": "text", + "content": "First, we introduce IVGaze, an in-vehicle gaze dataset collected from 125 subjects. IVGaze provides a dense distribution of gaze directions, covering a wide range within in-vehicle environments. It contains various conditions, including diverse head poses, eye movements, illumination variations, and the presence of face accessories such as glasses, sunglasses, and masks. To collect IVGaze, we propose a vision-based gaze collection system. The system does not require dedicated eye-tracking devices and is easy to reproduce. We use a single camera for facial appearance capture and paste stickers in vehicles as gaze targets. However, a significant challenge lies in calibrating the 3D position of gaze targets. This challenge arises because gaze targets are out of the camera's field of view (out-of-FoV). To address this issue, we present a refined gaze target calibration method, which leverages an auxiliary camera to capture gaze targets and a transparent chessboard for calibration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 349, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 467 + ], + "type": "text", + "content": "Secondly, we explore in-vehicle gaze estimation using the IVGaze. Face images often suffer from low resolution due to the inherent limitations of cameras in vehicles. We propose a gaze pyramid transformer that utilizes a transformer to integrate multilevel features. Expanding upon this, we propose a dual-stream gaze pyramid transformer (GazeDPTR). We rotate virtual cameras via perspective transformation to normalize images, and leverage camera pose to merge normalized and original images. GazeDPTR shows state-of-the-art results on IVGaze." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 469, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 577 + ], + "type": "text", + "content": "Thirdly, we extend GazeDPTR for the downstream gaze zone classification task. It is challenging to compute the intersection of gaze and the vehicle. We define a foundational tri-plane and project gaze to the tri-plane. We extract positional features from intersection points and predict gaze zones using both positional features and visual features from face images. Our experiment demonstrates that the gaze zone classification can be further enhanced by positional features, showing the advantage of gaze estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 590, + 138, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 138, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 138, + 601 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 169, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 169, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 169, + 621 + ], + "type": "text", + "content": "2.1. Gaze Data Collection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "The human gaze is inherently implicit and poses a challenge for objective measurement, making gaze annotation complex in gaze data collection. Some methods capture the human gaze through intrusive devices such as eye-tracking glasses [13, 19]. However, the eye-tracking glasses have a notable impact on the quality of the captured facial images. Fischer et al. [13] attempted to mitigate this impact" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 130, + 544, + 297 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 545, + 126 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 545, + 126 + ], + "type": "text", + "content": "Table 1. The comparison of gaze estimation datasets. Gaze annotation is challenging in the vehicle environment, which results in existing in-vehicle datasets only providing gaze zone annotation. Our work addresses this issue and contributes the first in-vehicle gaze dataset containing gaze annotation and natural face images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 130, + 544, + 297 + ], + "lines": [ + { + "bbox": [ + 307, + 130, + 544, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 544, + 297 + ], + "type": "table", + "html": "
DatasetsEnv.# Sub.Annotation
GazeZone
Gaze360 [20]Outdoor238✓-
MPIIGaze [39]15✓-
EyeDiap [15]Indoor16✓-
EVE [27]54✓-
ETH-XGaze [40]110✓-
Vora et al. [31]10✗✓
Jha & Busso [18]16✗✓
Wang et al. [32]Vehicle3✗✓
Rangesh et al. [28]13✗✓
Ghosh et al. [16]338✗✓
IVGaze (Ours)125✓✓
", + "image_path": "cdfd582c673589eecd0a48900355166ea67a4828b6778d8aac64c6cd26a51fc6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "content": "using GAN, but there are also some artifacts in the resulting face images. Vision-based gaze collection systems typically define gaze as direction vectors originating from facial centers towards specific gaze targets [27, 39, 40]. However, these gaze targets often remain outside the camera's field of view. This out-of-field view challenge complicates the calibration of gaze target positions. Zhang et al. [39, 40] sets screen points as gaze targets and uses a mirror to calibrate the screen plane. Kellnhofer et al. [20] uses " + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "content": " panoramic cameras to capture gaze targets and human faces simultaneously where the panoramic camera is pre-calibrated. However, these strategies are not applicable to vehicles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "In-vehicle gaze dataset usually defines different region such as windshield and left/right mirror in vehicles and perform gaze zone classification [12, 14, 16, 18, 23, 30, 34]. Kasahara et al. [19] collects an in-vehicle gaze dataset but subjects are required to wear eye-tracking glasses, which means the dataset is not applicable in the real world. Our gaze collection system does not require dedicated devices and produces natural face images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 552, + 495, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 495, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 495, + 563 + ], + "type": "text", + "content": "2.2. Appearance-based Gaze Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Appearance-based gaze estimation directly learns mapping function from facial appearance to human gaze [10]. Conventional gaze estimation methods extract eye features from eye images [7, 9, 26, 36]. They concatenate the head pose vector with eye features for gaze estimation. Recent methods directly learn gaze from face images [3, 4, 11, 33, 35]. Face images provide both eye region and head pose information, resulting in superior performance compared to methods relying solely on eye images. However, the subtlety of the human eye in face images poses a challenge as it can be easily overlooked by the network. Zhang et al. [37] utilize a learnable attention map to guide the net" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1557" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 283, + 220 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 283, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 283, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 283, + 220 + ], + "type": "image", + "image_path": "05326171b8c2d985f03644cbebb4e804555a48e2d148099086e68565d0a34d94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 221, + 161, + 229 + ], + "lines": [ + { + "bbox": [ + 151, + 221, + 161, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 221, + 161, + 229 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 545, + 220 + ], + "type": "image", + "image_path": "53a7a6a1e467b62214bd05f80e3fbe76bf19fc501c5ba6588b97fdfc0dc42a7d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 221, + 354, + 229 + ], + "lines": [ + { + "bbox": [ + 345, + 221, + 354, + 229 + ], + "spans": [ + { + "bbox": [ + 345, + 221, + 354, + 229 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 221, + 446, + 229 + ], + "lines": [ + { + "bbox": [ + 438, + 221, + 446, + 229 + ], + "spans": [ + { + "bbox": [ + 438, + 221, + 446, + 229 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 221, + 516, + 229 + ], + "lines": [ + { + "bbox": [ + 507, + 221, + 516, + 229 + ], + "spans": [ + { + "bbox": [ + 507, + 221, + 516, + 229 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 232, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 298 + ], + "type": "text", + "content": "Figure 2. We construct a vision-based in-vehicle gaze collection system comprising a DMS camera, a depth camera, and strategically placed gaze targets, as depicted in (c). The DMS camera is positioned behind the steering wheel to capture drivers' facial appearances, while the gaze targets, positioned beyond the DMS camera's field-of-view (FoV), such as on the windshield, remain unobserved. The depth camera, utilized for calibration purposes, is temporarily installed for capturing gaze target positions in 3D with respect to its own coordinates, and it is removed during data collection. To facilitate the calibration of the depth camera's pose relative to the DMS camera, we propose employing a transparent chessboard, which is placed between the two cameras." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 413 + ], + "type": "text", + "content": "work's focus specifically on the eye. Cheng et al. [11] crop eye images from face images and construct a cascade network to leverage them. In recent developments, transformers have showcased remarkable capabilities in gaze estimation. Cheng et al. [5] demonstrate that transformers can achieve state-of-the-art performance across various benchmarks. They apply transformers to address dual-view gaze estimation and show that dual-view images outperform single-view images [6]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 425, + 265, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 265, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 265, + 438 + ], + "type": "text", + "content": "3. In-Vehicle Gaze Data Collection System" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "content": "In this section, we introduce our vision-based system for in-vehicle gaze data collection, eliminating the need for dedicated eye-tracking devices. Our system includes a neat gaze target calibration method that effectively addresses the core challenge in gaze annotation. Leveraging this system, we collect the first in-vehicle gaze dataset from 125 subjects. The dataset spans a diverse range of gaze, facilitating and advancing future research in in-vehicle gaze estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 550, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 134, + 562 + ], + "type": "text", + "content": "3.1. System Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": "We place gaze targets within the vehicle, and subjects are instructed to look at each designated target during the data collection process. Simultaneously, we record their facial appearance along with corresponding gaze targets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "type": "text", + "content": "Camera. Our system uses one camera of a driver monitor system (DMS) for facial appearance capture. The DMS camera is an infrared camera with a capture resolution of " + }, + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "type": "inline_equation", + "content": "1280 \\times 800" + }, + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "type": "text", + "content": ". The camera is located behind the steering wheel and directly points at the head region of drivers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Gaze targets. We mark gaze targets with red points on stickers, and each target is uniquely labeled with a distinct number printed on stickers to aid differentiation. We strate-" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 546, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 365 + ], + "type": "text", + "content": "gically position these stickers within various gaze regions within the vehicle, including the windshield, left and right-side mirrors, rear-view mirror, center console, speedometer, handbrake, and etc. These targets cover a large gaze region, satisfying the gaze estimation requirements in vehicles." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 372, + 427, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 427, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 427, + 384 + ], + "type": "text", + "content": "3.2. Collection Procedure" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "content": "We meticulously design the collection procedure to minimize error. During the collection process, participants are instructed to look at specific gaze targets in a predefined sequence. They are also required to speak the corresponding number associated with each target to confirm their focus accuracy. Participants must maintain focus on the target for 3 seconds to facilitate image capture. We conduct a preliminary check on the captured images and discard any erroneous ones. Once this verification is complete, the current data collection phase concludes, and participants are instructed to shift their gaze to the next gaze targets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "content": "We design three postures for each participant. Participants should complete the collection process three times with different postures. The first posture requires participants not to change head pose. They should preserve the same head pose and only move eyeballs to focus a gaze target. We aim to collect sufficient eye movement data with this posture. We do not require participants to focus on the target where participants think it is hard to focus with only eye movement, e.g., the target in the side mirror. Participants can rotate their heads to look at these points but they are also required to preserve the new head pose until they cannot look at any targets with the current head pose. The second posture requires participants not to move their head position but they can rotate their heads to focus a gaze target. We do not set any constraints in the third posture. Participants can freely perform head movements in this pos" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1558" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 83, + 227, + 201 + ], + "blocks": [ + { + "bbox": [ + 52, + 83, + 227, + 201 + ], + "lines": [ + { + "bbox": [ + 52, + 83, + 227, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 83, + 227, + 201 + ], + "type": "image", + "image_path": "2870d1c47d06ce53cf6858d51f7ccad31f2f5d7b3058ed26afe593d8380b2b4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 545, + 248 + ], + "type": "text", + "content": "Figure 3. Our dataset is collected using IR cameras in the vehicle environment. (a) We present image samples of IVGaze, highlighting the challenges posed by realistic in-vehicle conditions, including cases with sunglasses and reflections in glasses. (b) We categorize the image count based on their mean pixel value, showing the diversity of illumination conditions. (c) The image count is analyzed based on face accessories including glasses, sunglasses, and masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 71, + 431, + 201 + ], + "blocks": [ + { + "bbox": [ + 231, + 71, + 431, + 201 + ], + "lines": [ + { + "bbox": [ + 231, + 71, + 431, + 201 + ], + "spans": [ + { + "bbox": [ + 231, + 71, + 431, + 201 + ], + "type": "image", + "image_path": "a885f36b2846195ca82470208a05578eb83ec9da22241f7d01ef4fb96e2a0360.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 569, + 201 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 569, + 201 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 569, + 201 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 569, + 201 + ], + "type": "image", + "image_path": "8c25bef82c6cf83982cf40d0a67ec6db2599fa66e99b69270499d00a6b301bf3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 307 + ], + "type": "text", + "content": "ture. The last two postures bring head pose variation for IVGaze. We do not use devices such as headrests to constrain participants. Participants only need to preserve the specific posture and focus on targets in a comfortable way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 316, + 149, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 149, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 149, + 328 + ], + "type": "text", + "content": "3.3. Gaze Annotation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": "Vision-based gaze collection systems define gaze as unit direction vectors originating from the face center " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " and extending toward gaze targets " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " [10]. Our system follows this framework and the gaze annotation can be computed as " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{g} = (\\mathbf{t} - \\mathbf{o}) / ||\\mathbf{t} - \\mathbf{o}||_2" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": ". The gaze annotation is decomposed into 3D face center estimation and gaze target calibration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "type": "text", + "content": "3D face center estimation has been effectively addressed in previous research. We first detect facial landmarks and fit a 3D morphable face model for 3D facial landmarks [24, 40]. We select the 3D position of the nose as gaze origin o. However, the calibration of gaze target positions presents a challenge because gaze targets are out-of-FoV. It means that the DMS camera cannot capture any images of the gaze targets, complicating the calibration problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 514, + 185, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 185, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 185, + 527 + ], + "type": "text", + "content": "3.4. Gaze Target Calibration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": "We need the 3D gaze target position " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " for gaze annotation. However, a key problem is all gaze targets are out-of-FoV. Conventional gaze datasets often present gaze targets on a screen and employ a mirror to solve the out-of-FoV problem, where the mirror reflects the content of the screen [36]. However, these methods are not suitable for vehicle scenarios and can lead to a substantial accumulation of errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "In this section, we propose a neat and efficient gaze target calibration method. Our basic idea is to employ an auxiliary camera to capture gaze targets. This allows a straightforward calculation of the 3D gaze target positions w.r.t. the auxiliary camera coordinate system. However, it also introduces a new challenge: how to calibrate extrinsic matrix of the auxiliary cameras w.r.t. DMS camera? Stereo calibration is typically used to compute the pose of two cam" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "content": "eras. However, it requires two captured images sharing corresponding points. In our system, gaze targets are located behind the DMS camera which means the DMS camera and auxiliary camera should be oriented in opposite directions, making traditional stereo calibration infeasible." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "type": "text", + "content": "To solve this problem, we propose to use a transparent chessboard for the pose calibration. We set a transparent chessboard between the two cameras and the two cameras respectively capture each side of the chessboard. We first calibrate the two camera poses w.r.t. the chessboard coordinate system and then compute the pose between two cameras. It is worth noting that there are two different chessboard coordinate systems corresponding to each side of the chessboard. We can derive the transformation matrix between the two chessboard coordinate systems where " + }, + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{chess}} = \\mathrm{diag}(0,0,-1)" + }, + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{chess}} = (0,0,-\\mathrm{d})" + }, + { + "bbox": [ + 304, + 319, + 546, + 461 + ], + "type": "text", + "content": ". The d is the thickness of the chessboard." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "text", + "content": "In detail, we use an Intel RealSense D435 depth camera as the auxiliary camera since it can provide accurate 3D positions. We show the camera layout in Fig. 2(b). We can obtain the transformation matrix " + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{dms}}" + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{dms}}" + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 304, + 462, + 545, + 534 + ], + "type": "text", + "content": " which can transfer a point from chessboard coordinate systems to camera coordinate systems. We have" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 366, + 542, + 545, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 542, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 366, + 542, + 545, + 558 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {\\mathrm {r o t}} = \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {R} _ {\\mathrm {c h e s s}} \\mathbf {R} _ {\\mathrm {d e p t h}} ^ {- 1}, \\tag {1}", + "image_path": "475ed0ce339c3e67ed6ae2d6184d3ef43f692f0ac7106ea3d42448e74f1dd979.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 567, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 591 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {\\text {r o t}} = - \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {R} _ {\\text {c h e s s}} \\mathbf {R} _ {\\text {d e p t h}} ^ {- 1} \\mathbf {t} _ {\\text {d e p t h}} + \\mathbf {R} _ {\\mathrm {d m s}} \\mathbf {t} _ {\\text {c h e s s}} + \\mathbf {t} _ {\\mathrm {d m s}}. \\tag {2}", + "image_path": "7c865cf4fc91da34ab3ac233720d40cfd8ebe932cbea3c9ebca333e28d887bdb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "content": "We use these matrices to convert points from the depth camera coordinate system into the DMS camera coordinate system. Please refer to the supplementary material for details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 441, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 441, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 441, + 647 + ], + "type": "text", + "content": "3.5. In-Vehicle Gaze Dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "We collect the first in-vehicle gaze dataset IVGaze which provides dense gaze annotation and natural face images. We show the dataset samples in Fig. 3 (a)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Dataset Statistics. We collect 44,705 images from 125 subjects. The number of images per subject ranges from" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1559" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 165, + 175 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 165, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 165, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 165, + 175 + ], + "type": "image", + "image_path": "6ac84c8c9ac9218a7dd53f0a820ebc628d22089999853487859601635fcce461.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "type": "text", + "content": "Figure 4. We show the distribution of data for gaze (left) and head movements (right). Brighter regions denote higher data density." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 70, + 276, + 175 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 276, + 175 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 276, + 175 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 276, + 175 + ], + "type": "image", + "image_path": "d2203024d1579751e1ce55940485aa68ef0c9a015915d53d07494b80feb46ddc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 214, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 286, + 262 + ], + "type": "text", + "content": "169 to 946. Most subjects provide around 300 images per person, i.e., they repeat the collection procedure three times with three different postures. 40 subjects are female and 85 subjects are male, with ages spanning from 20 to 50 years." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 263, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 286, + 323 + ], + "type": "text", + "content": "Collection Conditions. The dataset was collected between 9 am and " + }, + { + "bbox": [ + 47, + 263, + 286, + 323 + ], + "type": "inline_equation", + "content": "7\\mathrm{pm}" + }, + { + "bbox": [ + 47, + 263, + 286, + 323 + ], + "type": "text", + "content": " in outdoor environments, covering a wide range of lighting conditions. We also collect images in indoor environments, i.e., a garage. The dataset statistics across pixel intensity are shown in Fig. 3 (b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 326, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 286, + 422 + ], + "type": "text", + "content": "Face Accessories. We consider two face accessories during the collection: glasses and masks. For subjects who did not wear glasses, we provided glasses and asked them to repeat the collection wearing glasses. We provided white and black masks for some subjects and asked them to repeat the collection wearing masks. We also required a few subjects to wear sunglasses to facilitate future research. We show the statistics in Fig. 3 (c)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "text", + "content": "Data Distribution. We plot the angular distribution of gaze and head pose. Note that the data normalization method (introduced in the next section) changes gaze directions and head poses with a rotation matrix. Fig. 4 shows the distribution in normalization space. The horizontal gaze is from " + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "inline_equation", + "content": "-50^{\\circ}" + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "text", + "content": ", and the vertical gaze is from " + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "inline_equation", + "content": "-40^{\\circ}" + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "inline_equation", + "content": "40^{\\circ}" + }, + { + "bbox": [ + 46, + 423, + 286, + 603 + ], + "type": "text", + "content": ". It is worth noting that the vehicle environment has limited space, and our dataset already covers almost all regions. Our dataset is collected in the real vehicle environment, and the driver seat in the vehicle is located on the left. Therefore, our dataset has a relatively small distribution on the left. The right figure in Fig. 4 shows the head pose distribution. Our dataset contains a large range of head poses in the yaw axis, as subjects need to look at gaze targets from the left-side mirror to the right-side mirror." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 620, + 203, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 203, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 203, + 632 + ], + "type": "text", + "content": "4. In-Vehicle Gaze Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "text", + "content": "The in-vehicle environment brings new settings and challenges. In this section, we systematically explore in-vehicle gaze estimation. We introduce a revised data preprocessing method and describe a novel gaze estimation network. We also extend the network for an application and prove the advantage of in-vehicle gaze estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 518, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 518, + 84 + ], + "type": "text", + "content": "4.1. Revisiting Data Normalization in Vehicle" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 90, + 545, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 545, + 161 + ], + "type": "text", + "content": "Data normalization rotates and translates cameras via perspective transformation in face images to reduce head pose variation. Conventional methods usually define the rotated camera coordinate system " + }, + { + "bbox": [ + 305, + 90, + 545, + 161 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_r" + }, + { + "bbox": [ + 305, + 90, + 545, + 161 + ], + "type": "text", + "content": " based on head pose [29, 38]. However, these methods cannot bring performance improvement in the vehicle environment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis of " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_r" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " is typically defined as the " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis of head coordinate systems. It means they will rotate images to keep the head straight. We observed that the designed " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis will produce unstable results, especially in the extreme head pose. Therefore, we propose that do not rotate virtual cameras based on the " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis of head coordinate systems. In detail, we compute the " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis of " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_r" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " as direction vectors from cameras to face centers. We use the " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis of the virtual camera, i.e., " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "(1,0,0)" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " rather than the head coordinate system. We compute the " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis as " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "y = z\\times x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " and also recompute the " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": "-axis as " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "x = y\\times z" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": ". Finally, we have a rotation matrix " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = [x;y;z]" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " and use a scale matrix " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " to maintain the user-camera distance. We warp the image based on matrix SR. The human gaze is also changed as " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^n = \\mathbf{R}\\mathbf{g}^o" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^n" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^o" + }, + { + "bbox": [ + 305, + 162, + 546, + 354 + ], + "type": "text", + "content": " denotes gaze in the normalization and origin space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 360, + 458, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 458, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 458, + 372 + ], + "type": "text", + "content": "4.2. Gaze Pyramid Transformer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 378, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 486 + ], + "type": "text", + "content": "Gaze estimation methods typically learn gaze from face images. However, face images captured in vehicles often suffer from low resolution owing to the limited capabilities of the camera. This low resolution can lead conventional methods to overlook subtle details, especially in the eye region. Some techniques address this challenge by cropping eye images and processing them separately [8], but these approaches encounter difficulties when faced with extreme head poses, as one of the eye images may not be visible." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 487, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 545, + 605 + ], + "type": "text", + "content": "We propose a gaze pyramid transformer (GazePTR) in this section. We build a feature pyramid and input multilevel features into a transformer for gaze estimation. As shown in Fig. 5, our network uses a convolution network to extract features from facial images. We collect feature maps from different levels rather than the last level. We use " + }, + { + "bbox": [ + 305, + 487, + 545, + 605 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 487, + 545, + 605 + ], + "type": "text", + "content": " convolution layers and global average pooling layers to preserve the same feature dimension. Finally, we input the multi-level feature into a transformer. A learnable token is used to aggregate multi-level features for gaze estimation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 606, + 545, + 713 + ], + "type": "text", + "content": "However, the network has the same performance achieved by directly utilizing the last feature map for gaze estimation. This result can be attributed to the network's tendency to overlook low-level features. To address this issue, we introduce additional constraints by estimating gaze from each level of features and calculating corresponding loss. It can be understood that we first extract multi-level features, with each feature being potentially useful. A transformer is used to effectively ensemble these features." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1560" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 75, + 547, + 214 + ], + "blocks": [ + { + "bbox": [ + 50, + 75, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 50, + 75, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 547, + 214 + ], + "type": "image", + "image_path": "65ba7fefcea69103c4bd47f4963e9bfb6bc14be0a0e5033533b2547bc9d1c9c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 548, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 548, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 548, + 300 + ], + "type": "text", + "content": "Figure 5. The GazeDPTR directly crop face for origin images and rotates virtual cameras via perspective transformation for normalized images. It builds a dual-stream network to extract features from the two images based on the GazePTR for feature extraction which integrates multi-level features via transformers. To further merge the features from two streams, we leverage a transformer where camera pose is used as the positional feature in the transformer. We define the original camera pose as " + }, + { + "bbox": [ + 46, + 222, + 548, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\text{diag}(1, 1, 1)" + }, + { + "bbox": [ + 46, + 222, + 548, + 300 + ], + "type": "text", + "content": " and the camera pose in normalization space is RC. We also extend the network for gaze zone classification. We define a tri-plane and project gaze into them. We extract positional features from three intersection points via a transformer. We also extract visual features from images and predict the gaze zone based on both visual features and positional features. The whole network is trained in an end-to-end manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 309, + 263, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 263, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 263, + 322 + ], + "type": "text", + "content": "4.3. Dual-Stream Gaze Pyramid Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": "Dual-view images have been demonstrated to outperform single-view images in gaze estimation [6]. In our work, we rotate cameras in data normalization through perspective transformation. This inspired us to formulate the dual-camera setting leveraging both original and normalized images. However, dual-view images improve accuracy since they contain more visual information. It remains uncertain whether the combination of normalized and original images can provide additional insights beyond what each offers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 507 + ], + "type": "text", + "content": "To validate our hypothesis, we executed an oracle baseline. We separately train GazePTR on the original and the normalized datasets, selecting the best result from each image pair. The result reveals a significant performance improvement, validating the inherent advantages of utilizing both images. Please refer to the supplementary for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "type": "text", + "content": "Therefore, we propose a dual-stream gaze pyramid transformer (GazeDPTR). Our method leverages both normalized images and original images for gaze estimation. We employ GazePTR to extract features from the two images and a transformer is used to integrate the two features for gaze estimation. Note that, the two features are in different camera coordinate systems and correspond to different gaze. To establish the connection, we use the camera pose as the positional information in the transformer. However, it is impossible the calibrate a virtual camera. We define the camera pose of original images as " + }, + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\text{diag}(1, 1, 1)" + }, + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "type": "text", + "content": " and the camera pose of normalized images is RC. We extract the " + }, + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 508, + 288, + 664 + ], + "type": "text", + "content": "-axis and use positional encoding for camera pose [6]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 245, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 245, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 245, + 684 + ], + "type": "text", + "content": "4.4. Strategy for Gaze Zone Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Detecting driver attention is essential in driver monitoring systems. In this section, we introduce a novel strategy to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 309, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 429 + ], + "type": "text", + "content": "extend GazeDPTR for gaze zone classification. GazeDPTR estimates gaze direction from face images. While the conventional solution involves projecting gaze direction and computing intersections within the vehicle, it is not practical without a 3D vehicle model. Zhang et al. use a triplane for self-supervised loss function [35]. Our alternative strategy defines a foundational tri-plane and computes the intersection of gaze vectors with this tri-plane. We extract positional features from three intersection points to facilitate gaze zone classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": "In detail, we define a tri-plane with normal vectors " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "(1,0,0)" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "(0,1,0)" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "(0,0,1)" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": ", intersecting at the origin " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "(0,0,0)" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": ". Given the estimated gaze vector " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^n" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": ", we convert it back to the original space using " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^o = \\mathbf{R}^{-1}\\mathbf{g}^n" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": ". We then project " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^o" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": " onto the tri-plane with the face center " + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 304, + 430, + 546, + 585 + ], + "type": "text", + "content": " and obtain three intersection points. To enhance the positional features, we apply positional encoding to these points and input them into a 2-layer transformer. The positional encoding not only increases feature dimensions but also normalizes them within the range of -1 to 1. Additionally, we use another learnable token to aggregate visual features from the original images. Gaze zone classification is performed using both visual features and positional features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 586, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 633 + ], + "type": "text", + "content": "We train the whole network in an end-to-end manner. We use L1 loss for gaze estimation and cross-entropy loss for gaze zone classification. Please refer to the supplementary for implementation details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 645, + 382, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 382, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 382, + 658 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Dataset: IVGaze contains 44,705 images of 125 subjects. To perform within-dataset evaluation, we divide the dataset into three subsets based on subjects. The image numbers of the three subsets are 15,165, 14,674, and 14,866. We" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1561" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 96, + 291, + 238 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "type": "text", + "content": "Table 2. Performance comparison in gaze estimation. Our methods achieve better performance than the compared methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 96, + 291, + 238 + ], + "lines": [ + { + "bbox": [ + 47, + 96, + 291, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 291, + 238 + ], + "type": "table", + "html": "
Angular ErrorAverage Precision (AP)
<2°<4°<6°<8°
FullFace [37]13.67°2.3%8.8%17.8%28%
DWG [16]8.82°6.6%21.7%38.1%53.2%
Gaze360 [20]8.15°9.2%27.3%44.6%58.9%
FullFace+ [37]7.48°14.2%31.1%46.7%63.1%
GazeTR [5]7.33°17%32.8%47.5%64.7%
XGaze [40]7.06°11.7%32.7%51.5%66.7%
GazePTR7.04°17.6%34%49.3%66.7%
GazeDPTR6.71°22.1%36%50.3%68.4%
", + "image_path": "8bb4fea72daa82d1c9c30fb59aca1ada6f5ea3b55c1da09b1b7aa1c3f064a2a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 251, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 251, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 251, + 264 + ], + "type": "text", + "content": "perform three-fold cross-validation on our dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": "Evaluation Metric: We use the angular error as the evaluation metric of gaze estimation as most of the methods [10], where a smaller value represents a better model. However, we notice the angular error only shows the overall performance while cannot give deep insights. Therefore, we define the average precision (AP) where AP of " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "< k^{\\circ}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": " means an estimation is considered correct if the angular error is lower than " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "k^{\\circ}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": ". Regarding the gaze zone classification, average precision is used as a common multi-class classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 380, + 227, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 227, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 227, + 393 + ], + "type": "text", + "content": "5.1. Comparison with SOTA Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "type": "text", + "content": "We first compared our methods with SOTA methods in the in-vehicle gaze estimation task. We compared our methods with FullFace [37], Gaze360 [20], XGaze [40], DWG [16] and GazeTR [5]. We replaced the backbone of the FullFace method from AlexNet to ResNet18 for a more convincing comparison. We denote the new method as FullFace+." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": "The result is shown in Tab. 2. GazePTR and GazeDPTR both show better performance than the compared methods. GazePTR and GazeTR have the same backbone and transformer architecture. However, GazePTR outperforms " + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "0.29^{\\circ}" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": " thanGazeTR since GazePTR uses multi-level feature maps. GazeDPTR builds a dual-stream network and further brings " + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "0.33^{\\circ}" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": " improvement than GazePTR. These results show the advantage of our methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "type": "text", + "content": "Interestingly, XGaze has a similar angular error to GazePTR. However, it is significantly worse than GazePTR in AP of " + }, + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "type": "inline_equation", + "content": "< 2^{\\circ}" + }, + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "type": "text", + "content": ". It is because that XGaze uses ResNet50 as the backbone. The deep backbone enhances the feature extraction ability but also easily overlooks the small eye region, making it challenging to achieve precise gaze estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 217, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 217, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 217, + 659 + ], + "type": "text", + "content": "5.2. The Impact of Face Accessories" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "IVGaze provides rich samples of wearing face accessories, including glasses, masks, and sunglasses. We also evaluate their impacts on the accuracy of gaze estimation and show in Tab. 3. The result shows that eyeglasses have a" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 96, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 3. We show the impacts of different face accessories on performance. The sunglasses bring a significant performance drop." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 96, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 307, + 96, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 553, + 238 + ], + "type": "table", + "html": "
GlassesMaskSunglasses
withw/owithw/owith
FullFace [37]14.43°12.40°15.20°13.35°21.39°
DWG [16]9.20°8.19°9.43°8.69°17.43°
Gaze360[20]8.30°7.91°8.95°7.99°17.99°
FullFace+ [37]7.59°7.30°8.37°7.30°16.50°
XGaze [40]7.07°7.03°7.80°6.90°15.15°
GazeTR [5]7.40°7.22°8.12°7.17°17.49°
GazePTR7.13°6.90°7.78°6.89°16.54°
GazeDPTR6.77°6.63°7.44°6.57°16.41°
", + "image_path": "14744fc076aec7d025c13b12c6d6ec8f5883998b5f161a3ac42a3e6d3c6f61a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 247, + 426, + 344 + ], + "blocks": [ + { + "bbox": [ + 309, + 247, + 426, + 344 + ], + "lines": [ + { + "bbox": [ + 309, + 247, + 426, + 344 + ], + "spans": [ + { + "bbox": [ + 309, + 247, + 426, + 344 + ], + "type": "image", + "image_path": "f791c54df84bfd40997e2f380edc5b5e5dc237cc0175e878e26be2b5d4d4fd85.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "type": "text", + "content": "Figure 6. We compute the angular degree between gaze direction/head orientation with the frontal direction, i.e., " + }, + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "type": "inline_equation", + "content": "(0,0,-1)" + }, + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "type": "text", + "content": ", and count the mean gaze estimation accuracy in different ranges. Interestingly, GazePTR brings significant performance improvement to GazeTR in the " + }, + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "type": "inline_equation", + "content": "0^{\\circ} - 20^{\\circ}" + }, + { + "bbox": [ + 304, + 352, + 547, + 440 + ], + "type": "text", + "content": " gaze range. This is because subjects prefer to move their eyeball rather than their head when the gaze target is located at a close range. GazePTR utilizes multilevel features and can accurately capture eye movement." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 247, + 545, + 343 + ], + "blocks": [ + { + "bbox": [ + 427, + 247, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 427, + 247, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 427, + 247, + 545, + 343 + ], + "type": "image", + "image_path": "77c89d723cf5f7a5e4c5b674946d7f369db0b1838b8629bf259fccea5abfa316.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "large impact if the basic performance is not good. However, as the basic performance improves, the performance difference reduces. This is because most of glasses are transparent, enabling a robust model to extract features effectively. Sunglasses have a significant impact on performance, with average performance exceeding " + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": ". Masks have a substantial impact on performance, as they obscure facial regions, complicating the extraction of head-related features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 552, + 449, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 449, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 449, + 563 + ], + "type": "text", + "content": "5.3. Performance Distribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We show the performance distribution to provide deep insight. We first show the performance distribution in different gaze ranges. We introduce a novel concept, the gaze range of " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "a^{\\circ} - b^{\\circ}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", which is defined as a set of samples satisfying the requirement that the angular degree between the gaze direction and the frontal gaze direction is in the range. The result is shown in Fig. 6. Interestingly, GazePTR and GazeDPTR have better performance in " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "0^{\\circ} - 20^{\\circ}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". This is because subjects prefer to move their eye rather than head to look at objects in this range. Our methods leverage multi-level features and easily capture eye movement. We also show the performance distribution in different head ranges." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1562" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 118, + 282, + 148 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "type": "text", + "content": "Table 4. We present the performance of GazePTR at each level of the feature. The higher-level feature achieves better performance. Our method integrates features from all levels and achieves the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 118, + 282, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 118, + 282, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 118, + 282, + 148 + ], + "type": "table", + "html": "
1st2nd3rd4thOurs
GazePTR10.33°8.59°7.61°7.35°7.04°
", + "image_path": "6629b29aa078134e28edec39f1105508bd8f8f18613995251f6bcf99aae64e6d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 217, + 286, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 155, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 155, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 287, + 212 + ], + "type": "text", + "content": "Table 5. We present the performance of GazePTR separately training on the original images and normalized images. GazeDPTR leverages camera pose to establish a connection between the two images and use both two images for gaze estimation. We also removed the camera pose to perform an ablation study." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 217, + 286, + 256 + ], + "lines": [ + { + "bbox": [ + 50, + 217, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 286, + 256 + ], + "type": "table", + "html": "
Original imagesNormalized imagesGazeDPTR (w/o camera pose)GazeDPTR
Acc7.44°7.04°7.03°6.71°
", + "image_path": "5e8e441645049396b043793ff1c22d448516eb9d2dd73a6c1c7f0b8e7b1b5ba2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 262, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 287 + ], + "type": "text", + "content": "Our methods are robust in different head ranges where the maximum performance difference is only " + }, + { + "bbox": [ + 47, + 262, + 287, + 287 + ], + "type": "inline_equation", + "content": "1^{\\circ}" + }, + { + "bbox": [ + 47, + 262, + 287, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 295, + 141, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 141, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 141, + 308 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "type": "text", + "content": "GazePTR leverages multi-level features for gaze estimation. We first conduct an ablation study to demonstrate the advantage. We obtain the gaze performance of GazePTR at each level of feature and show it in Tab. 4. The result demonstrates a higher-level feature achieves better performance. GazePTR integrates multi-level features and achieves the best performance. The result proves the advantage of our design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "text", + "content": "GazeDPTR leverages both original and normalized images along with their corresponding camera poses. To demonstrate the effectiveness of our design, we perform an ablation study in Tab. 5. Training GazePTR separately on original and normalized images yields performances of " + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "inline_equation", + "content": "7.44^{\\circ}" + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "inline_equation", + "content": "7.04^{\\circ}" + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "text", + "content": " respectively. GazeDPTR, which incorporates both image types, achieves a performance improvement of " + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "inline_equation", + "content": "0.34^{\\circ}" + }, + { + "bbox": [ + 46, + 409, + 288, + 529 + ], + "type": "text", + "content": ". We also experimented by excluding the positional encoding of the camera pose in GazeDPTR. Interestingly, the result is comparable to using only normalized images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 183, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 183, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 183, + 550 + ], + "type": "text", + "content": "5.5. Additional Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "content": "Data Normalization. We compare our method with the data normalization method [38] in the left of Tab. 6. The previous method cannot work well in vehicle environments while our method brings " + }, + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "inline_equation", + "content": "0.3^{\\circ}" + }, + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "content": " performance improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 603, + 266, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 266, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 266, + 615 + ], + "type": "text", + "content": "Case Study. We visual gt and our prediction in Fig. 7." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 623, + 186, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 186, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 186, + 635 + ], + "type": "text", + "content": "5.6. Gaze Zone Classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "We propose a basis tri-plane to acquire positional feature and combine both positional feature and visual features for gaze zone classification. We respectively evaluate each feature and show the result in the right of Tab. 6. The visual feature achieves better performance than the positional feature since the positional feature is obtained from gaze pro" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 132, + 544, + 195 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 126 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 126 + ], + "type": "text", + "content": "Table 6. We evaluate different data normalization methods in the left table. Our method outperforms the previous method [38]. In the right table, we evaluate different features in the gaze zone classification. Our method uses both positional feature and visual features, achieving " + }, + { + "bbox": [ + 305, + 70, + 545, + 126 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 305, + 70, + 545, + 126 + ], + "type": "text", + "content": " improvement over visual features." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 132, + 544, + 195 + ], + "lines": [ + { + "bbox": [ + 309, + 132, + 544, + 195 + ], + "spans": [ + { + "bbox": [ + 309, + 132, + 544, + 195 + ], + "type": "table", + "html": "
Normalization [38]oursGazeTRGazePTRVisual FeaturePositional FeatureAP
××7.77°7.44°✓79.4%
✓8.64°8.53°✓75.3%
✓7.33°7.04°✓✓81.8%
", + "image_path": "3fdac2ee66c01b82a04a81324af32814a284a38cf3099f4c291633f325028843.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 313, + 198, + 367, + 252 + ], + "blocks": [ + { + "bbox": [ + 313, + 198, + 367, + 252 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 367, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 367, + 252 + ], + "type": "image", + "image_path": "7c74a154f013274f2bc0b275fa40e5817ef46509a6250cf9368aed0da587aca0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 370, + 198, + 423, + 252 + ], + "blocks": [ + { + "bbox": [ + 370, + 198, + 423, + 252 + ], + "lines": [ + { + "bbox": [ + 370, + 198, + 423, + 252 + ], + "spans": [ + { + "bbox": [ + 370, + 198, + 423, + 252 + ], + "type": "image", + "image_path": "5deb96d370044fb41499b0012f948c4a7593447cedbf57c49d285d762735f14a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 425, + 198, + 480, + 252 + ], + "blocks": [ + { + "bbox": [ + 425, + 198, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 425, + 198, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 425, + 198, + 480, + 252 + ], + "type": "image", + "image_path": "e1688cd8b86a1f5c59544da0c7e3e9af73fe7a36d888b5a3eef616d3ff49b11e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 482, + 198, + 536, + 252 + ], + "blocks": [ + { + "bbox": [ + 482, + 198, + 536, + 252 + ], + "lines": [ + { + "bbox": [ + 482, + 198, + 536, + 252 + ], + "spans": [ + { + "bbox": [ + 482, + 198, + 536, + 252 + ], + "type": "image", + "image_path": "e31a932506a0c0240de6a0e51f540b0d4a606164062623b93a4cf8417c1e09eb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 313, + 253, + 367, + 305 + ], + "blocks": [ + { + "bbox": [ + 313, + 253, + 367, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 253, + 367, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 367, + 305 + ], + "type": "image", + "image_path": "9647d81b316851b2ef573ca3f6839f399e2460907b3eed813c173f8d35ade647.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 310, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 305, + 310, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 545, + 344 + ], + "type": "text", + "content": "Figure 7. We illustrate GT with green lines and Prediction with red lines. The number shows the angular error in degree. The right two figures are failure cases since the eye is hard to capture." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 370, + 253, + 423, + 306 + ], + "blocks": [ + { + "bbox": [ + 370, + 253, + 423, + 306 + ], + "lines": [ + { + "bbox": [ + 370, + 253, + 423, + 306 + ], + "spans": [ + { + "bbox": [ + 370, + 253, + 423, + 306 + ], + "type": "image", + "image_path": "50a223a2104ebf1e8a768ca9c88c872432c8e42ad342478cccb04a0b585a34f4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 425, + 253, + 480, + 306 + ], + "blocks": [ + { + "bbox": [ + 425, + 253, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 425, + 253, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 425, + 253, + 480, + 306 + ], + "type": "image", + "image_path": "7e27d54d41d9c2cabf8a652684d9e86a5f9c4b8ef67f66e838e929c33a6fa38e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 482, + 253, + 536, + 306 + ], + "blocks": [ + { + "bbox": [ + 482, + 253, + 536, + 306 + ], + "lines": [ + { + "bbox": [ + 482, + 253, + 536, + 306 + ], + "spans": [ + { + "bbox": [ + 482, + 253, + 536, + 306 + ], + "type": "image", + "image_path": "4488432ef6d3d8cd112629dd8c585833d8c2679102c30b9459937fb21ddaf9ce.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 355, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 428 + ], + "type": "text", + "content": "jection. The projection increases the model interpretability but decreases the fitting ability. Our method uses the two features and gets " + }, + { + "bbox": [ + 304, + 355, + 545, + 428 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 355, + 545, + 428 + ], + "type": "text", + "content": " performance improvement. The result demonstrates that the gaze zone classification task could be enhanced by gaze direction cues, highlighting the advantage of our gaze estimation work." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 435, + 379, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 379, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 379, + 448 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 455, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 623 + ], + "type": "text", + "content": "In this work, we provide systematical research in in-vehicle gaze estimation, including dataset, algorithm, and extensive application. We first solve the data collection issue in a vehicle where a gaze target calibration method is proposed. We collect the first in-vehicle gaze dataset containing dense gaze annotation and natural face images. We further explore the algorithm for in-vehicle gaze estimation. Our work brings two deep insights 1) multi-level feature is useful to capture eye region information. 2) simultaneously leveraging original images and normalized images could achieve better performance. We also extend our work for the downstream gaze zone classification task. We demonstrate that the gaze direction cues could bring performance improvement for gaze zone classification." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 634, + 419, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 419, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 419, + 647 + ], + "type": "text", + "content": "7. Acknowledgments:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00608, Artificial intelligence research about multi-modal interactions for empathetic conversations with humans)" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1563" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Abdul Rafey Aftab. Multimodal driver interaction with gesture, gaze and speech. In 2019 International Conference on Multimodal Interaction, pages 487-492, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Aya Ataya, Won Kim, Ahmed Elsharkawy, and SeungJun Kim. Gaze-head input: Examining potential interaction with immediate experience sampling in an autonomous vehicle. Applied Sciences, 10(24), 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Yiwei Bao, Yunfei Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with rotation consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4207-4216, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "text", + "content": "[4] Xin Cai, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Source-free adaptive gaze estimation by uncertainty reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22035-22045, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 268, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 288, + 291 + ], + "type": "text", + "content": "[5] Yihua Cheng and Feng Lu. Gaze estimation using transformer. ICPR, 2022. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "text", + "content": "[6] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20632-20641, 2023. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "text", + "content": "[7] Yihua Cheng, Feng Lu, and Xucong Zhang. Appearance-based gaze estimation via evaluation-guided asymmetric regression. In The European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "text", + "content": "[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 424, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 457 + ], + "type": "text", + "content": "[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 457, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 501 + ], + "type": "text", + "content": "[10] Yihua Cheng, Haofei Wang, Yiwei Bao, and Feng Lu. Appearance-based gaze estimation with deep learning: A review and benchmark. arXiv preprint arXiv:2104.12668, 2021. 2, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "text", + "content": "[11] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. AAAI, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "text", + "content": "[12] In-Ho Choi, Sung Kyung Hong, and Yong-Guk Kim. Real-time categorization of driver's gaze zone using the deep learning techniques. In 2016 International conference on big data and smart computing (BigComp), pages 143-148. IEEE, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "text", + "content": "[13] Tobias Fischer, Hyung Jin Chang, and Yiannis Demiris. Rtgene: Real-time eye gaze estimation in natural environments. In The European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "type": "text", + "content": "[14] Lex Fridman, Philipp Langhans, Joonbum Lee, and Bryan Reimer. Driver gaze region estimation without use of eye movement. IEEE Intelligent Systems, 31(3):49-56, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[15] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2014. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "type": "text", + "content": "[16] Shreya Ghosh, Abhinav Dhall, Garima Sharma, Sarthak Gupta, and Nicu Sebe. Speak2label: Using domain knowledge for creating a large scale driver gaze zone estimation dataset. In The IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2896-2905, 2021. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 139, + 546, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 546, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 546, + 193 + ], + "type": "text", + "content": "[17] Amie C Hayley, Brook Shiferaw, Blair Aitken, Frederick Vinckenbosch, Timothy L Brown, and Luke A Downey. Driver monitoring systems (dms): The future of impaired driving management? Traffic injury prevention, 22(4):313-317, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 194, + 546, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 546, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 546, + 238 + ], + "type": "text", + "content": "[18] Sumit Jha and Carlos Busso. Probabilistic estimation of the gaze region of the driver using dense classification. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 697-702. IEEE, 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 238, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 546, + 282 + ], + "type": "text", + "content": "[19] Isaac Kasahara, Simon Stent, and Hyun Soo Park. Look both ways: Self-supervising driver gaze estimation and road scene saliency. In European Conference on Computer Vision, pages 126-142. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 282, + 546, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 546, + 327 + ], + "type": "text", + "content": "[20] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In The IEEE International Conference on Computer Vision, 2019. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 327, + 546, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 546, + 358 + ], + "type": "text", + "content": "[21] Muhammad Qasim Khan and Sukhan Lee. A comprehensive survey of driving monitoring and assistance systems. Sensors, 19(11), 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 358, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 546, + 392 + ], + "type": "text", + "content": "[22] Muhammad Qasim Khan and Sukhan Lee. Gaze and eye tracking: Techniques and applications in adas. Sensors, 19 (24), 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 392, + 546, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 546, + 447 + ], + "type": "text", + "content": "[23] Sung Joo Lee, Jaeik Jo, Ho Gi Jung, Kang Ryoung Park, and Jaihie Kim. Real-time gaze estimator based on driver's head orientation for forward collision warning system. IEEE Transactions on Intelligent Transportation Systems, 12(1): 254-267, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 448, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 448, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 448, + 546, + 491 + ], + "type": "text", + "content": "[24] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 491, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 546, + 525 + ], + "type": "text", + "content": "[25] Prajval Kumar Murali, Mohsen Kaboli, and Ravinder Dahiya. Intelligent in-vehicle interaction technologies. Advanced Intelligent Systems, 4(2):2100122, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 525, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 546, + 558 + ], + "type": "text", + "content": "[26] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In The European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 558, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 546, + 602 + ], + "type": "text", + "content": "[27] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In The European Conference on Computer Vision, pages 747-763. Springer, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 602, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 546, + 647 + ], + "type": "text", + "content": "[28] Akshay Rangesh, Bowen Zhang, and Mohan M Trivedi. Driver gaze estimation in the real world: Overcoming the eyeglass challenge. In The IEEE Intelligent Vehicles Symposium (IV), pages 1054-1059. IEEE, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "type": "text", + "content": "[29] Yusuke Sugano, Yasuyuki Matsushita, and Yoichi Sato. Learning-by-synthesis for appearance-based 3d gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition, 2014. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 690, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 546, + 713 + ], + "type": "text", + "content": "[30] Ashish Tawari, Kuo Hao Chen, and Mohan M Trivedi. Where is the driver looking: Analysis of head, eye and" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1564" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 597 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "iris for robust gaze zone estimation. In 17th International IEEE conference on intelligent transportation systems (ITSC), pages 988-994. IEEE, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 161 + ], + "type": "text", + "content": "[31] Sourabh Vora, Akshay Rangesh, and Mohan Manubhai Trivedi. Driver gaze zone estimation using convolutional neural networks: A general framework and ablative analysis. IEEE Transactions on Intelligent Vehicles, 3(3):254-265, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "text", + "content": "[32] Yafei Wang, Guoliang Yuan, Zetian Mi, Jinjia Peng, Xueyan Ding, Zheng Liang, and Xianping Fu. Continuous driver's gaze zone estimation using rgb-d camera. Sensors, 19(6): 1287, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "text", + "content": "[33] Yaoming Wang, Yangzhou Jiang, Jin Li, Bingbing Ni, Wenrui Dai, Chenglin Li, Hongkai Xiong, and Teng Li. Contrastive regression for domain adaptation on gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19376-19385, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "type": "text", + "content": "[34] Guoliang Yuan, Yafei Wang, Huizhu Yan, and Xianping Fu. Self-calibrated driver gaze estimation via gaze pattern learning. Knowledge-Based Systems, 235:107630, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[35] Mingfang Zhang, Yunfei Liu, and Feng Lu. Gazeonce: Real-time multi-person gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4197-4206, 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "text", + "content": "[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In The IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 452 + ], + "type": "text", + "content": "[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2299-2308, 2017. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "text", + "content": "[38] Xucong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the ACM Symposium on Eye Tracking Research & Applications, 2018. 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "text", + "content": "[39] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):162-175, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "text", + "content": "[40] Xuong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In The European Conference on Computer Vision, 2020. 1, 2, 4, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1565" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_content_list.json b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e7fa1a26c47e2085495bfb12020dba529c77b0b0 --- /dev/null +++ b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_content_list.json @@ -0,0 +1,1324 @@ +[ + { + "type": "text", + "text": "What, How, and When Should Object Detectors Update in Continually Changing Test Domains?", + "text_level": 1, + "bbox": [ + 204, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jayeon Yoo $^{1}$ Dongkwan Lee $^{1}$ Inseop Chung $^{1}$ Donghyun Kim $^{2*}$ Nojun Kwak $^{1*}$ Seoul National University Korea University", + "bbox": [ + 142, + 202, + 833, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1{jayeon.yoo, biancco, jis3613, nojunk}@snu.ac.kr, $^2\\mathrm{d}$ kim@korea.ac.kr", + "bbox": [ + 181, + 239, + 782, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is a well-known fact that the performance of deep learning models deteriorates when they encounter a distribution shift at test time. Test-time adaptation (TTA) algorithms have been proposed to adapt the model online while inferring test data. However, existing research predominantly focuses on classification tasks through the optimization of batch normalization layers or classification heads, but this approach limits its applicability to various model architectures like Transformers and makes it challenging to apply to other tasks, such as object detection. In this paper, we propose a novel online adaption approach for object detection in continually changing test domains, considering which part of the model to update, how to update it, and when to perform the update. By introducing architecture-agnostic and lightweight adaptor modules and only updating these while leaving the pre-trained backbone unchanged, we can rapidly adapt to new test domains in an efficient way and prevent catastrophic forgetting. Furthermore, we present a practical and straightforward class-wise feature aligning method for object detection to resolve domain shifts. Additionally, we enhance efficiency by determining when the model is sufficiently adapted or when additional adaptation is needed due to changes in the test distribution. Our approach surpasses baselines on widely used benchmarks, achieving improvements of up to $4.9\\% p$ and $7.9\\% p$ in mAP for COCO $\\rightarrow$ COCO-corrupted and SHIFT, respectively, while maintaining about 20 FPS or higher. The implementation code is available at https://github.com/natureyoo/ContinualTTA_ObjectDetection.", + "bbox": [ + 75, + 323, + 483, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 790, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models excel in many vision tasks but struggle with domain shifts when test data distributions diverge from training data [3, 27, 42]. These shifts, common in real-world settings, arise from natural variations, weather", + "bbox": [ + 75, + 816, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9dfbac57f05601a9cee12ea9481b3c07e19a607e1c687920a9668d2a48e2e0ad.jpg", + "image_caption": [ + "Figure 1. We propose an online adaptation method for object detection in continually changing test domains. Object detectors trained with clean images suffer from performance degradation due to various corruption, such as camera sensor degradation or environmental changes (Direct-Test). Updating full parameters for online adaptation require a large number of test samples and vulnerable to drastic domain changes (Full-Finetuning), while using only our lightweight adaptor is robust and quickly adapts within a few time steps (Ours). We can further improve efficiency by skipping unnecessary adaptation steps (Ours-Skip)." + ], + "image_footnote": [], + "bbox": [ + 490, + 289, + 901, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "changes (e.g., fog, rain), camera differences (e.g., pixelate, defocus blur), and various other factors. Test-Time Adaptation (TTA) [3, 27, 32, 42, 45, 49] addresses these shifts by adapting models to the target (test) distribution online. Considering the dynamic nature of real-world applications, it becomes crucial to extend this approach to Continual Test-Time Adaptation (CTA), which accounts for the ongoing evolution of test distributions. This is particularly vital in environments like autonomous driving, where conditions can change continuously (e.g., from clear to rainy, or day to night). Despite its significance, continual test-time adaptation for object detection remains underexplored.", + "bbox": [ + 496, + 672, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several TTA methods [6, 31, 38] tailored for object detection have been proposed. ActMAD [31] aligns all the output feature maps $(\\mathbb{R}^{C\\times H\\times W})$ after Batch Nor", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-corresponding authors", + "bbox": [ + 94, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23354", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "malization (BN) layers [15] to adapt the test domain to be similar to that of the training domain. However, this approach requires significant memory during adaptation and does not explicitly consider the objects present in the image. TeST [38] and STFAR [6] adapt to a test domain by utilizing weak and strong augmented test samples with a teacher-student network [39], but they significantly increase inference costs since they require additional forward passes and update steps. Also, these methods update all network parameters, making them highly inefficient in online adaptation and vulnerable to losing task-specific knowledge when the test domain experiences continual or drastic changes.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to develop an efficient continual test-time adaptation (CTA) method for object detection. We investigate the following three key aspects to improve efficiency; what to update: while previous TTA methods for object detection [6, 31, 38] use full fine-tuning, updating all parameters at test time, they are inefficient and prone to losing task-specific knowledge in relatively complex object detection tasks. Updating BN layers, as done in many TTA methods for classification [18, 32, 45, 49], is not as effective for object detection, given its smaller batch size compared to classification and the limitation in applying various backbones, such as Transformer [28, 43]. how to update: several previous TTA methods for object detection [6, 38] adapt the model by using teacher-student networks, resulting in a significant decrease in inference speed, which is detrimental during test time. While another existing method [31] aligns feature distributions for adaptation, it does not consider each object individually, focusing only on image features, making it less effective for object detection. when to update: most TTA or CTA methods update models using all incoming test samples. However, it is inefficient to update continuously the model if it is already sufficiently adapted when the change of the test domain is not significant.", + "bbox": [ + 75, + 276, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, (1) we propose an efficient continual test-time adaptation method for object detectors to adapt to continually changing test domains through the use of lightweight adaptors which require only $0.54\\% \\sim 0.89\\%$ additional parameters compared to the full model. It exhibits efficiency in parameters, memory usage, and adaptation time, along with robustness to continuous domain shifts without catastrophic forgetting. Additionally, it demonstrates wide applicability to various backbone types compared to BN-based TTA methods [18, 24, 32, 45, 49, 50]. (2) To enhance the adaptation effectiveness in the object detection task, we align the feature distribution of the test domain with that of the training domain at both the image-level and object-level using only the mean and variance of features. For estimating the mean of the test domain features, we employ Exponentially Moving Average (EMA) as we can leverage only the current incoming test samples, not the entire test domain data. Due to the unavailability of training", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data access, we utilize only the mean and variance of the features from a few training samples. (3) We also introduce two novel criteria that do not require additional resources to determine when the model needs adaptation to enhance efficiency in a continually changing test domain environment. As illustrated in Fig. 1, our approach Ours, employing adaptors, tends to adapt much faster to domain changes compared to full parameter updates. This enables efficient TTA by using only a few test samples to update the adaptor and skipping the rest of the updates as shown in Ours-Skip. Our main contributions are summarized as follows:", + "bbox": [ + 496, + 90, + 893, + 256 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an architecture-agnostic lightweight adaptor, constituting only a maximum of $0.89\\%$ of the total model parameters, into the backbone of the object detector to adapt the model in a continually changing test domain. This approach ensures efficiency in parameters, memory usage, and adaptation speed, demonstrating the robust preservation of task-specific knowledge owing to its inherent structural characteristics.", + "- We propose a straightforward and effective adaptation loss for CTA in object detection tasks. This is achieved by aligning the distribution of training and test domain features at both the image and object levels, utilizing only the mean and variance of a few training samples and EMA-updated mean features of the test domain.", + "- We also propose two criteria to determine when the model requires adaptation, enabling dynamic skipping or resuming adaptation as needed. This enhancement significantly boosts inference speed by up to about 2 times while maintaining adaptation performance.", + "- Our adaptation method proves effective for diverse types of domain shifts, including weather changes and sensor variations, regardless of whether the domain shift is drastic or continuous. In particular, our approach consistently improves the mAP by up to $7.9\\%$ in $\\mathrm{COCO} \\rightarrow \\mathrm{COCO-C}$ and SHIFT-Discrete/Continuous with higher than 20 FPS." + ], + "bbox": [ + 500, + 258, + 893, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 648, + 640, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time adaptation. Recently, there has been a surge of interest in research that adapts models online using unlabeled test samples while simultaneously inferring the test sample to address the domain shift problem, where the test data distribution differs from that of the training data. There are two lines for online adaptation to the test domain, Test-time Training (TTT) and Test-time Adaptation (TTA). TTT [1, 2, 27, 42] involves modifying the model architecture during training to train it with self-supervised loss, allowing adaptation to the test domain in the test time by applying this self-supervised loss to the unlabeled test samples. On the other hand, TTA aims to adapt the trained model directly to the test domain without specifically tailored model architectures or losses during training time. NORM [37] and DUA [30] address the domain shifts by", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23355", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adjusting the statistics of batch normalization (BN) layers using the current test samples, without updating other parameters, inspired by [23]. Following this, [24, 32, 45, 50] and [18] update the affine parameters of BN layers using unsupervised loss, entropy minimization loss to enhance the confidence of test data predictions, and feature distribution alignments loss, respectively. Several studies [16, 17] update the classifier head using the pseudo-prototypes from the test domain. However, these methods limit their applicability to architectures without BN layers or to object detection tasks that involve multiple objects in a single image. Others [31, 40, 49] update full parameters for online adaptation to the test domain in an online manner, but this approach is inefficient and susceptible to the noisy signal from the unsupervised loss. While existing TTA methods are oriented towards classification tasks, we aim to propose an effective and efficient method for online adaptation in the object detection task.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual test-time adaptation. Recent studies [33, 46] point out that existing TTA methods have primarily focused on adapting to test domains following an i.i.d assumption and may not perform well when the test data distribution deviates from this assumption. [46] introduces a ContinuaTTA (CTA) method designed for scenarios where the test domain continuously changes over time. This poses challenges in preventing the model from over-adapting to a particular domain shift and preserving the knowledge of the pre-trained model to avoid catastrophic forgetting. In the field of CTA, the self-training strategy adopting an Exponentially Moving Average (EMA) teacher-student structure is attracting interest as an effective algorithm enabling robust representation to be learned through self-knowledge distillation. In many studies, the EMA teacher-student structure and catastrophic restoration of source model weights have been proposed as a solution to achieve the goal of CTA [4, 46, 47]. Approaches using source replay [34], and anti-forgetting regularization [32] have also achieved good performances in robust continuous adaptation. Furthermore, there is growing attention on methods that mitigate the computational and memory challenges associated with CTA, such as [13], which relies on updates to batch normalization statistics.", + "bbox": [ + 75, + 367, + 472, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-time adaptive object detection. Research on TTA for Object Detection (TTAOD) is progressively emerging [6, 31, 38, 44]. Most existing TTAOD methods [6, 38, 44] exploit a teacher-student network to adapt to the test domain, following the self-training approach commonly employed in Unsupervised Domain Adaptation for object detection [7, 19, 21, 36]. However, it is inefficient for TTA due to data augmentation requirements and additional forward and backward steps, resulting in slower inference speeds and higher memory usage. Another approach, ActMAD [31], aligns the distributions of output feature maps", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "after all BN layers along the height, width, and channel axes to adapt to the test domain. However, this location-aware feature alignment is limited to datasets with fixed location priors, such as driving datasets, and is less effective for natural images like COCO. Additionally, CTA for Object Detection (CTAOD) have not been thoroughly explored. Therefore, there is a need for an effective CTAOD method considering memory and time efficiency.", + "bbox": [ + 496, + 90, + 893, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 227, + 591, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable the efficient and effective Continual Test-time Adaptation of Object Detectors (CTAOD), we introduce an approach that specifies which part of the model should be updated, describes how to update those using unlabeled test data, and determines whether we perform model updates or not to improve efficiency.", + "bbox": [ + 496, + 253, + 893, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 498, + 356, + 629, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume that we have an object detector $h \\circ g_{\\Theta}$ , here $h$ and $g$ are the RoI head and the backbone, respectively with their parameters being $\\Theta$ . The training dataset is denoted as $D_{train} = \\{(x_i, y_i)\\}_{i=1}^N$ , where $x_i \\sim P_{train}(x)$ and $y_i = (bbox_i, c_i)$ , containing information on the bounding box (bbox) and class label $c_i \\in \\mathcal{C}$ . Consider deploying the detector to the test environments where the test data at period $T$ is denoted as $x_j^T \\sim P_{test}^T(x)$ , $P_{test}^T \\neq P_{train}$ and $P_{test}^T$ deviates from the i.i.d. assumption. In addition, the domain of $P_{test}^T$ continually changes according to $T$ (i.e., $P_{test}^T \\neq P_{test}^{T-1}$ ). Our goal is to adapt the detector $h \\circ g$ to $P_{test}^T$ using only test data $x_j^T$ while making predictions.", + "bbox": [ + 496, + 378, + 893, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. What to update: Adaptation via an adaptor", + "text_level": 1, + "bbox": [ + 498, + 574, + 870, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous methods [6, 31, 38, 44] adapt the model to the test domain by updating all parameters $\\Theta$ , leading to inefficiency at test time and a high risk of losing task knowledge from the training data. In contrast, we adapt the model by introducing an adaptor with an extremely small set of parameters and updating only this module while freezing $\\Theta$ . We introduce a shallow adaptor in parallel for each block, inspired by [5, 14], where transformer-based models are fine-tuned for downstream tasks through parameter-efficient adaptors, as shown in Fig. 2. Each adaptor consists of down-projection layers $\\mathcal{W}_{down} \\in \\mathbb{R}^{d \\times \\frac{d}{r}}$ , up-projection layers $\\mathcal{W}_{up} \\in \\mathbb{R}^{\\frac{d}{r} \\times d}$ and RLUs, where $d$ denotes the input channel dimension and $r$ is the channel reduction ratio set to 32 for all adaptors. We use MLP layers for the Transformer block (Fig. 2a) and $1 \\times 1$ convolutional layers for the ResNet block (Fig. 2b) to introduce architecture-agnostic adaptors. The up-projection layer is initialized to 0 values so that the adaptor does not modify the output of the block, but as the adaptor is gradually updated, it adjusts the output of the block to adapt to the test domain. Even as the adaptor", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23356", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6699e81c9c72d975f0d533edc09ae91b2b4adc92fff6ef6d71389a27b98f553e.jpg", + "image_caption": [ + "Figure 2. We attach an adaptor, which is a shallow and low-rank MLP or CNN, to every $N$ block in parallel. We update only these adaptors while other parameters are frozen. Our approach can be applied to diverse architectures including CNNs and Transformers." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 277, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d40db235519bc4ba5f5f700896391ad47202e140b5235f9e0ebdee3490839c18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 88, + 478, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is updated in the test domain, the original backbone parameter $\\Theta$ remains frozen and fully preserved. This structural preservation, as evident in Ours in Fig. 1, enables robust and efficient adaptation to domain changes by maintaining relatively complex task knowledge in object detection and updating very few parameters.", + "bbox": [ + 75, + 352, + 470, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. How to update: EMA feature alignment", + "text_level": 1, + "bbox": [ + 76, + 455, + 419, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To adapt the object detector to the test domain, we align the feature distribution of the test domain with that of the training data, inspired by [18, 31, 40]. In contrast to these methods that solely align image feature distribution, we additionally align object-level features in a class-wise manner, considering class frequency, to enhance its effectiveness for object detection. As the training data is not accessible during test time, we pre-compute the first and second-order statistics, denoted as $\\mu_{tr} = \\mathbb{E}[F_{tr}]$ and $\\Sigma_{tr} = \\mathbb{Var}[F_{tr}]$ , where the operators $\\mathbb{E}$ and $\\mathbb{Var}$ represent the mean and variance respectively. The features $F_{tr} = \\{g_{\\Theta}(x_{tr})\\}$ are computed using only 2,000 training samples, a small subset of the training data. Since a sufficient amount of test domain data is not available at once, and only the current incoming test data, whose features are denoted as $F_{te}^{t}$ , is accessible at time step $t$ , we estimate the mean of test data features using an exponentially moving average (EMA) as follows:", + "bbox": [ + 75, + 479, + 470, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t e} ^ {t} = (1 - \\alpha) \\cdot \\mu_ {t e} ^ {t - 1} + \\alpha \\cdot \\mathbb {E} [ F _ {t e} ^ {t} ], \\quad \\mathrm {s . t .} \\mu_ {t e} ^ {0} = \\mu_ {t r}. (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 748, + 468, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the typically small batch size in object detection compared to classification, we approximate the variance of the test features as $\\Sigma_{te} \\simeq \\Sigma_{tr}$ to reduce instability.", + "bbox": [ + 75, + 779, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-level feature alignment. We estimate the training and test feature distributions as normal distributions and minimize the KL divergence between them as follows:", + "bbox": [ + 75, + 825, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i m g} = \\mathrm {D} _ {K L} (\\mathcal {N} (\\mu_ {t r}, \\Sigma_ {t r}), \\mathcal {N} (\\mu_ {t e} ^ {t}, \\Sigma_ {t r})). \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region-level class-wise feature alignment. In object detection, we deal with multiple objects within a single image, making it challenging to apply the class-wise feature alignment proposed in [40], a TTA method for classification. To handle region-level features that correspond to an object, we use ground truth bounding boxes for the training data and utilize the class predictions of RoI pooled features, $f_{te}^{t}$ , for unlabeled test data. In object detection, domain shifts often result in lower recall rates, as a significant number of proposals are predicted as background [22]. To mitigate this issue, we filter out features with background scores exceeding a specific threshold. Subsequently, we assign them to the foreground class with the highest probability, as follows:", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t e} ^ {k, t} = \\left\\{f _ {t e} ^ {t} \\mid \\underset {c} {\\operatorname {a r g m a x}} p _ {f g} = k, p _ {b g} < 0. 5 \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 296, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e h _ {c l s} \\left(f _ {t e} ^ {t}\\right) = \\left[ p _ {f g}, p _ {b g} \\right] = \\left[ p _ {0}, \\dots , p _ {C - 1}, p _ {b g} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 324, + 857, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We estimate the class-wise feature distribution of the test domain by exploiting $F_{te}^{k,t}$ and Eq.1. Furthermore, we introduce a weighting scheme for aligning features of less frequently appearing classes, taking into account the severe class imbalance where specific instance (e.g., person) may appear multiple times within a single image, as follows:", + "bbox": [ + 496, + 349, + 892, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN ^ {k, t} = N ^ {k, t - 1} + \\left\\| F _ {t e} ^ {k, t} \\right\\|, \\quad \\text {s . t .} \\quad N ^ {k, 0} = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 449, + 828, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {k, t} = \\log \\left(\\frac {\\operatorname* {m a x} _ {i} N ^ {i , t}}{N ^ {k , t}}\\right) + 0. 0 1 \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 470, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o b j} = \\sum_ {k} w ^ {k, t} \\cdot D _ {K L} (\\mathcal {N} (\\mu_ {t r} ^ {k}, \\Sigma_ {t r} ^ {k}), \\mathcal {N} (\\mu_ {t e} ^ {k, t}, \\Sigma_ {t r} ^ {k})).\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 508, + 861, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the class-wise mean $\\mu^k$ and variance $\\Sigma^k$ of the training and test data are obtained in the same way as the image-level features. We can effectively adapt the object detector by updating the model to align the feature distribution at both the image and object levels as $L = L_{img} + L_{obj}$ .", + "bbox": [ + 496, + 550, + 892, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. When to update: Adaptation on demand", + "text_level": 1, + "bbox": [ + 498, + 636, + 846, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1, Ours, which only updates the adaptor proposed in Sec. 3.2, efficiently adapts to changes in the test domain, even with a small subset of early test samples. We leverage its rapid adaptation characteristics to reduce computational costs by skipping model updates (i.e., skipping backward passes) when the model has already sufficiently adapted to the current test domain and resuming model updates when confronted with a new test domain. Therefore, we introduce two criteria to determine when to update the model or not as follows:", + "bbox": [ + 496, + 659, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Criterion 1) When the distribution gap exceeds the in-domain distribution gap. Recall that $L_{img}$ (Eq. 2) measures the distribution gap between the test and train distributions. We assume a model is well-adapted to the current test domain when $L_{img}$ is closer to the in-domain distribution gap. We measure the in-domain distribution gap by", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23357", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5fdcd17f8b52f5a481a48fea362a9847ce21371e8adb0922ab70cac0fc1f5447.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 89, + 444, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec62c3ff5daaf694be509d578aebb28b879890ce9a7e77535a31467b488a2301.jpg", + "image_caption": [ + "(a) The ratio of $L_{img}$ to $D_{KL}^{in}$", + "(b) The ratio of $L_{img}$ to $L_{ema}^{t}$", + "Figure 3. The test domain undergoes a shift every 4,000 time steps, and each metric reaches its peak at the same intervals." + ], + "image_footnote": [], + "bbox": [ + 78, + 195, + 444, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sampling two disjoint subsets, $x_{i}$ and $x_{j}$ , of training features $F_{tr}$ from Sec. 3.3 as follows:", + "bbox": [ + 76, + 348, + 467, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {K L} ^ {i n} = D _ {K L} \\left(\\mathcal {N} \\left(\\mu_ {t r} ^ {i}, \\Sigma_ {t r} ^ {i}\\right), \\mathcal {N} \\left(\\mu_ {t r} ^ {j}, \\Sigma_ {t r} ^ {j}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 388, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{tr}^{i},\\Sigma_{tr}^{i}$ are obtained from $x_{i}\\sim P_{train}(x)$ and $\\mu_{tr}^{j},\\Sigma_{tr}^{j}$ from $x_{j}\\sim P_{train}(x)$ . In other words, if $L_{img}$ is noticeably higher than the in-domain distribution gap $D_{KL}^{in}$ , we consider a model encountering a test domain whose distribution differs from $P_{train}(x)$ and needs to be updated. Based on this, we introduce a new index $\\frac{L_{img}}{D_{KL}^{in}}$ . Fig. 3a plots the trend of this index during the model adaptation to a continually changing test domain. It shows that the index has a large value in the early stages of a domain change, decreases rapidly, and then maintains a value close to 1. This index exhibits a similar trend regardless of the backbone type and dataset, as included in the appendix. Therefore, we establish the criterion that model updates are necessary when this index exceeds a certain threshold, $\\tau_{1}$ , as $\\frac{L_{img}}{D_{KL}^{in}} >\\tau_{1}$ .", + "bbox": [ + 75, + 417, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(Criterion 2) When the distribution gap suddenly increases. Additionally, we can determine when the test distribution changes and model updates are necessary by observing the trend of the distribution gap (i.e., $L_{img}$ ). The convergence of $L_{img}$ indicates that a model is well-adapted to the current test domain. To put it differently, $L_{img}$ will exhibit a sudden increase when the model encounters a new test domain. We introduce an additional index, denoted as $\\frac{L_{img}}{L_{ema}^t}$ , representing the ratio of the current $L_{img}$ to its exponentially moving average $L_{ema}^t$ at time $t$ . We calculate it using the following formula: $L_{ema}^t = 0.99 \\cdot L_{ema}^{t-1} + 0.01 \\cdot L_{img}$ . Fig. 3b illustrates the trend of the ratio of $L_{img}$ over the timesteps. It tends to reach a value of 1 as the loss stabilizes at a specific level. Nevertheless, when the model encounters shifts in the test distribution, the ratio experiences a sharp increase, indicating the necessity of a model update when it exceeds a specific threshold, $\\tau_2$ , as $\\frac{L_{img}}{L_{ema}^t} > \\tau_2$ .", + "bbox": [ + 75, + 638, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If at least one of the two criteria is satisfied, we conclude that the model requires adaptation and proceed to update it.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 133, + 630, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 4.1 presents the two object detection benchmark datasets with test distributions that change continuously, either in a drastic or gradual manner, and our implementation detail is in 4.2. Sec. 4.4 compares our method with other TTA baselines described in Secs. 4.3.. We present detailed ablation studies of our method analyzing the effectiveness and efficiency of our method in terms of what, how, and when to update the models for CTAOD in Sec. 4.5.", + "bbox": [ + 496, + 157, + 890, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 287, + 601, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We experiment with the following three scenarios.", + "bbox": [ + 498, + 311, + 831, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathrm{COCO} \\rightarrow \\mathrm{COCO-C}$ simulates continuous and drastic realistic test domain changes over a long sequence. MS-COCO [25] collects 80 classes of common objects in their natural context with 118k training images and 5k validation images. COCO-C is created by employing 15 types of realistic corruptions [29], such as image distortion and various weather conditions, to simulate test domain changes. In the experiments, the model is only trained on the COCO train set and sequentially evaluated on each corruption in the COCO-C validation set during test-time for reproducing continually changing test domains. Finally, the model is evaluated on the original COCO validation set to assess how well it preserves knowledge of the original domain (denoted as Org.). SHIFT-(Discrete / Continuous) [41] is a synthetic driving image dataset with 6 classes under different conditions using five weather attributes (clear, cloudy, overcast, fog, rain) and three time-of-day attributes (daytime, dawn, night). In SHIFT-Discrete, there are image sets for each attribute, and the model is sequentially evaluated on these attributes, cloudy $\\rightarrow$ overcast $\\rightarrow$ foggy $\\rightarrow$ rainy $\\rightarrow$ dawn $\\rightarrow$ night $\\rightarrow$ clear which contains 2.4k, 1.6k, 2.7k, 3.2k, 1.2k, 1.4k, and 2.8k validation images, respectively. This simulates scenarios where the domain undergoes drastic changes. In SHIFT-Continuous, the model is evaluated on four sequences, each consisting of 4k frames, continuously transitioning from clear to foggy (or rainy) and back to clear.", + "bbox": [ + 496, + 325, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Detail", + "text_level": 1, + "bbox": [ + 500, + 726, + 709, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on Faster-RCNN [35] with FPN [26], using ImageNet1k pre-trained ResNet50 [11] and ImageNet21k pre-trained Swin-Tiny [28] backbones. Specifically, Swin-Tiny is employed for its superior domain transfer [20] and detection efficacy over other advanced backbones [9]. For COCO $\\rightarrow$ COCO-C with ResNet50, we use a publicly available model trained with strong augmentation [48]. Other models are trained using detectron2 framework following [35] and [28]. For test-time adaptation, we always set the learning to 0.001 for the SGD opti", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23358", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/723e48cd2ad214241f6c38b1745b401a8010954b1c64198aa31dd590a26b7a5a.jpg", + "table_caption": [ + "Table 1. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on COCO $\\rightarrow$ COCO-C. Our model consistently outperforms baselines on the two different backbones. Furthermore, Ours-Skip with ResNet notably reduces backward passes by as much as $90.5\\%$ , leading to a significantly improved frames per second (FPS) rate by up to $109.9\\%$ ." + ], + "table_footnote": [], + "table_body": "
BackboneMethodNoiseBlurWeatherDigitalOrg.Avg.# step
GauShtImpDefGlsMtnZmSnwFrsFogBrtCntElsPxJpgFor.Back.FPS
Swin-T [28]Direct-Test9.711.410.013.47.512.15.220.724.836.136.012.919.14.915.843.017.780K021.5
ActMAD10.712.09.412.35.79.54.515.317.527.628.21.116.72.68.736.313.980K80K8.3
Mean-Teacher10.012.111.212.88.112.14.919.623.734.934.08.018.96.117.641.017.2160K80K6.9
Ours13.616.616.114.013.614.28.323.727.237.436.427.227.222.222.342.322.680K80K9.5
Ours-Skip13.315.315.114.012.813.96.522.025.435.534.926.525.923.420.241.221.680K9.7K17.7
ResNet50 [11]Direct-Test9.111.09.812.64.58.84.619.123.138.438.021.415.65.311.944.217.380K025.8
NORM9.911.911.012.65.29.15.119.423.538.237.622.417.25.710.343.417.580K025.8
DUA9.811.710.812.85.28.95.119.323.738.437.822.317.25.410.144.117.180K025.8
ActMAD9.19.67.011.03.26.13.312.814.027.727.83.912.92.37.234.310.580K80K9.6
Mean-Teacher9.612.512.04.02.94.83.116.223.535.134.021.816.68.212.740.314.5160K80K8.1
Ours12.717.817.512.411.511.36.622.826.938.638.528.025.121.222.241.822.280K80K10.1
Ours-Skip14.417.116.013.911.712.26.322.125.537.737.125.524.123.121.142.821.980K7.6K21.2
", + "bbox": [ + 76, + 132, + 893, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mizer, and $\\alpha$ of Eq. 1 to 0.01, while $\\tau_{1}$ and $\\tau_{2}$ are set to 1.1 and 1.05, respectively. We use the same hyper-parameters across all backbones and datasets, with a batch size of 4.", + "bbox": [ + 75, + 321, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Baselines", + "text_level": 1, + "bbox": [ + 76, + 376, + 184, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Direct-Test evaluates the model trained in the training domain without adaptation to the test domain. ActMAD [31] is a TTA method aligning the distribution of output features across all BN layers. To apply ActMAD to the Swin Transformer-based model, we align the output features of the LN layers. We implement Mean-Teacher using a teacher-student network framework to reproduce as close as possible to TeST [38], as its implementation is not publicly available. We follow the FixMatch [39] augmentation method and report results after tuning all hyper-parameters in our scenario. NORM [37] and DUA [30], TTA methods initially designed for classification, are directly applicable to detection tasks by either mixing a certain amount of current batch statistics or updating batch statistics via EMA. However, these are only compatible with architectures containing BN layers. Additional details are provided in Appendix.", + "bbox": [ + 73, + 398, + 470, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Main Results", + "text_level": 1, + "bbox": [ + 76, + 650, + 215, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the performance of each method using mAP and efficiency metrics, including the number of forward and backward passes, as well as FPS during test-time adaptation. Results of COCO and SHIFT are in Tab. 1 and 2, respectively.", + "bbox": [ + 75, + 672, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{COCO} \\rightarrow \\mathbf{COCO-C}$ . Tab. 1 demonstrates the effective adaptation performance of Ours in the challenging COCO benchmark with 80 classes due to object-level class-wise feature alignment. ActMAD also aligns feature distribution for TTA, but is not effective since it only aligns whole feature maps without considering specific classes in the image. NORM and DUA, applicable only to ResNet [11], show minimal performance improvement by adaptation as they are not specifically tailored for object detection and only modify batch statistics across the entire feature map. Ad", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ditionally, ActMAD and Mean-Teacher, updating full parameters, gradually lose task knowledge in the continually changing test distributions, resulting in much lower performance on Org., the domain identical to the training data, than that of Direct-Test. In contrast, Ours effectively prevents catastrophic forgetting by freezing the original parameters of the models and updating only the adaptor, obtaining performance on par with Direct-Test on the Org. domain and consistently high performance across corrupted domains, with an average mAP improvement of $4.9\\%$ compared to that of Direct-Test. Furthermore, leveraging the rapid adaptation ability of the adaptor, Ours-Skip, which skips unnecessary adaptation, allows using only a maximum of about $12\\%$ of the total samples for adaptation without significant performance loss. This leads to a substantial improvement in inference speed, more than doubling compared to other TTA methods, reaching over 17.7 FPS.", + "bbox": [ + 496, + 321, + 893, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SHIFT-Discrete. Ours is also effective in SHIFT, which simulates continuous changes in weather and time in driving scenarios according to the left section of Tab. 2. Especially, Ours shows significant improvements in mAP by $7 - 9\\% \\mathrm{p}$ , particularly for the foggy and dawn attributes where Direct-Test obtains lower performance due to severe domain shift. In contrast, with ActMAD, catastrophic forgetting takes place when adapting to the cloudy and overcast weather. This is due to the updating of the full parameters, despite that Direct-Test already shows proficient performance in these conditions. As a result, the performance in the later domains is worse than that of the Direct-Test. DUA, which updates batch statistics using EMA, shows a gradual decrease in performance as the domain continuously changes, resulting in much lower performance in the original clear domain (i.e., clear). On the other hand, NORM, which utilizes the statistics of the current batch samples, exhibits no catastrophic forgetting and relatively good adaptation, as SHIFT is a relatively easier task compared to COCO due to having only 6 classes. Compared to NORM, Ours shows better adaptation performance, and is", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23359", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/814cdaaff59f1f33347e296af4745cc9c7abffbff0584953a08fd87f19b24aee.jpg", + "table_caption": [ + "Table 2. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on SHIFT-Discrete and SHIFT-Continuous. Baselines perform effectively in a particular setting but lack generalizability across various settings. Our method consistently achieves results that are either better or on par with the best model in all settings, demonstrating its strong stability. Ours-Skip also effectively reduces the number of backward passes without compromising mAP performance, resulting in a higher FPS." + ], + "table_footnote": [], + "table_body": "
BackboneMethodSHIFT-DiscreteSHIFT-Continuous
mAP# stepmAP# Avg. step
cloudyoverc.fograindawnnightclearAvg.For.Back.FPSclear↔fogclear↔rainFor.Back.FPS
Swin-T [28]Direct-Test50.038.923.145.126.939.545.938.515.3K027.518.121.14K028.3
ActMAD49.838.421.443.119.032.044.835.515.3K15.3K9.315.616.34K4K9.8
Mean-Teacher50.039.225.745.426.037.542.238.030.6K15.3K7.820.424.38K4K6.5
Ours50.339.232.246.730.439.944.340.415.3K15.3K11.223.922.64K4K11.6
Ours-Skip50.339.729.147.130.241.545.940.615.3K6.1K20.025.123.84K0.83K19.2
ResNet50 [11]Direct-Test49.437.919.743.120.135.345.635.915.3K030.112.115.44K030.0
NORM49.738.622.944.725.137.445.537.715.3K030.116.919.44K030.0
DUA45.231.527.731.915.218.621.127.315.3K030.122.522.44K030.0
ActMAD49.237.718.040.616.032.944.334.115.3K15.3K11.312.716.34K4K11.2
Mean-Teacher49.638.426.843.426.633.141.637.130.6K15.3K9.916.020.88K4K9.8
Ours49.738.727.446.327.437.643.838.715.3K15.3K12.920.921.94K4K13.9
Ours-Skip49.738.826.946.227.638.845.039.015.3K8.9K21.520.022.54K0.75K21.3
", + "bbox": [ + 76, + 146, + 893, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ed825c1a1ac9af85f56aec0af0d9cdc8f6442c25e9d030011223b811f2b8920b.jpg", + "table_caption": [ + "Table 3. Comparison of adaptation performance (mAP), the number of trainable parameters (# Params), and memory usage (Cache) according to which part of the backbone is updated. SD / SC denotes SHIFT-Discrete/Continuous, respectively." + ], + "table_footnote": [], + "table_body": "
BackboneTrainable ParamsmAP# ParamsCache
SDSCNumRatioAvg.Max
Swin-TFull-params38.420.627.7M100%0.8611.0
LayerNorm38.520.00.03M0.1%0.657.49
adaptor (Ours)40.423.20.15M0.5%0.656.96
ResNet50Full-params37.620.423.7M100%1.659.29
BatchNorm37.920.20.05M0.2%1.479.11
adaptor (Ours)38.721.70.21M0.9%1.485.41
", + "bbox": [ + 78, + 406, + 470, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "also applicable to BN-layer-free Swin Transformers.", + "bbox": [ + 75, + 550, + 423, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SHIFT-Continuous. In scenarios where the test domain gradually changes across the entire sequence, Ours also demonstrates effectiveness, improving mAP by up to $7\\%$ p, as shown in the right section of Tab. 2. While DUA performs well in the clear to foggy transition, it is prone to catastrophic forgetting in situations where the sequence becomes longer, and the test domain changes more diversely, as seen in the left section. Our strategy for determining when model adaptation is necessary is particularly effective in SHIFT. It improves FPS by about 9, reaching about 20 FPS, while enhancing mAP. This is likely due to avoiding overfitting that can occur when adapting to all repetitive frames in SHIFT, which consists of continuous frames, leading to improvements in both inference speed and adaptation performance.", + "bbox": [ + 75, + 564, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Additional Analyses", + "text_level": 1, + "bbox": [ + 76, + 786, + 267, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We aim to demonstrate the effectiveness and detailed analysis of our proposed model in terms of 1) which parts of, 2) how, and 3) when the model should be updated.", + "bbox": [ + 75, + 809, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Which part to update? Tab. 3 shows how updating different parts of the backbone model affects the performance and the memory usage during continual test-time adapta", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/322eb45c74e3137811d3e87bd2f13255e7b2c33ee6ea57d1453667d53ab56705.jpg", + "table_caption": [ + "Table 4. Ablation on each component of our loss. SHIFT-D / C denotes SHIFT-Discrete / Continuous, respectively. The left and right value in each cell corresponds to the mAP for the Swin-T and ResNet50 backbone, respectively." + ], + "table_footnote": [], + "table_body": "
LimgLobjCOCOSHIFT-D.SHIFT-C.
--17.7/ 17.338.5/ 35.919.6/ 13.8
✓-16.7/ 18.136.6/ 37.019.1/ 16.0
✓no class weight17.8/ 18.939.7/ 38.025.1/ 23.4
✓class weight wk,t22.6/ 22.240.4/ 38.723.2/ 21.7
", + "bbox": [ + 506, + 406, + 879, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion. We compare (1) updating full parameters, (2) affine parameters of the normalization layer, and (3) our proposed adaptor for each backbone on the SHIFT dataset. Although our adaptor has fewer parameters, about $0.9\\%$ or less of the full parameters, it demonstrates the best adaptation performance. Updating only the affine parameters of the normalization layer, while having fewer parameters, seems less effective for adaptation in object detection compared to classification [32, 45]. Additionally, our adaptor requires only about $60\\%$ of the memory compared to updating the full parameters, making it memory-efficient.", + "bbox": [ + 496, + 505, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on each component in our loss. Tab. 4 presents the effects of image-level feature alignment, $L_{img}$ , object-level feature class-wise alignment $L_{obj}$ , and class frequency weighting $w^{k,t}$ proposed to address class imbalance. Aligning only the image-level feature distribution with $L_{img}$ (first row) leads to modest adaptation in the ResNet50 backbone, while performance in the SwinT backbone is even lower than without adaptation. Notably, aligning object-level features with $L_{obj}$ leads to a substantial improvement, with the mAP increasing by approximately $10\\%$ compared to the no-adaptation scenario. Introducing class-specific frequency-based weighting $w^{k,t}$ , despite a slight performance decrease in the SHIFT-Continuous setting, proves highly effective, particularly in scenarios with significant class imbalance, such as COCO", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23360", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0059f846b713f8181cc7e07ca39ecd830bfa65635687bdd667bc6c446e5734ee.jpg", + "image_caption": [ + "(a) Swin Transformer backbone" + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 276, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b060176513247b6cdd997985aa5eb06888b9970d1159a8e387ea328ae0ece65.jpg", + "image_caption": [ + "(b) ResNet50 backbone" + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 472, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/147bb4ac8b27f43f9b253acb0e1e9e798e1089eb73839c7e1be1d64b7283cf2b.jpg", + "image_caption": [ + "Figure 4. Comparison of mAP and FPS from Ours-Skip with varying values of $\\tau_{1}(\\diamond)$ and $\\tau_{2}(\\triangle)$ against Evenly-Skip $(\\times)$ , adapting every $N$ -th instances, on COCO $\\rightarrow$ COCO-C using both (a) Swin-T and (b) ResNet50. The upward and rightward movement indicates a better strategy with higher mAP and faster inference speed, showing that Ours-Skip is consistently better than Evenly-Skip." + ], + "image_footnote": [], + "bbox": [ + 96, + 335, + 465, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ece3d0df3f5442d9afaf91a80237d30caa564bd3664eb2cf442291defdacc59.jpg", + "image_caption": [ + "(a) Accumulated number of backward steps", + "(b) Number of backward steps and mAP of Direct-Test in each domain", + "Figure 5. Analysis of the adaptation of Ours-Skip." + ], + "image_footnote": [], + "bbox": [ + 98, + 446, + 467, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with 80 classes, where it enhances the mAP by around $5\\%$ p. Trade-off between adaptation performance and efficiency according to different skipping strategies. Fig. 4 presents mAP and FPS depending on the values of $\\tau_{1}$ and $\\tau_{2}$ in the Sec. 3.4 on $\\mathrm{COCO} \\rightarrow \\mathrm{COCO - C}$ , which are used for two criteria to determine when the adaptation is needed. We also show the simple baseline Evenly-Skip, which adapts every $N$ -th step and skips the rest. In Fig. 4, the blue lines ( $\\triangle$ ) show the results when $\\tau_{1}$ is changing from 1.0 to infinity, where only criterion 2 is used, while $\\tau_{2}$ is fixed at 1.05. As $\\tau_{1}$ decreases, more adaptation is required, leading to slower FPS but higher mAP. The green lines ( $\\diamond$ ) show the results of changing $\\tau_{2}$ , where ‘ $\\tau_{2} = \\inf$ ' denotes using only criterion 1, without criterion 2. For all main experiments, we set $\\tau_{1}$ and $\\tau_{2}$ as 1.1 and 1.05, respectively, considering the balance between mAP and FPS. Additionally, our skipping strategy consistently outperforms Evenly-Skip, achieving higher values in both mAP and FPS. This indicates that our criterion for deciding when to bypass model updates provides an effective balance between accuracy and speed.", + "bbox": [ + 75, + 595, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When do models actually update? We analyze when the model actually skips adaptation and only performs inference or actively utilizes test samples for model adaptation based on the two criteria we propose. This analysis is conducted in COCO to COCO-C with 15 corruption domains and 1 original domain. Fig. 5a plots the number of backward passes, i.e., the number of batches of test samples used for adaptation, with different values of $\\tau_{1}$ for the two backbones. The horizontal and vertical axes represent sequentially incoming test domains and the cumulative backward numbers, respectively. A steep slope in a region indicates frequent adaptation, while a gentle slope indicates skipping adaptation, performing only inference. Notably, even without explicit information about when the test domain changes, the model actively performs adaptation, especially right after the test domain changes. This trend is consistent regardless of changes in $\\tau$ value or backbone type. The number of backward passes largely depends on the value of $\\tau_{1}$ rather than the backbone type, suggesting that a consistent $\\tau_{1}$ value can be used irrespective of the backbone. Fig. 5b illustrates adaptation patterns by dividing backward steps for each domain in the case of Swin-T backbone with $\\tau_{1} = 1.1$ . More clearly, it shows that adaptation occurs actively around the points where each domain changes, and afterward, adaptation happens intermittently or almost not at all. The light pink bars represent the performance of Direct-Test, showing that domains with initially high model performance tend to have less adaptation, while domains with lower performance initially need more adaptation. In other words, the amount of skipping adaptation is proportional to the amount of the domain shift. Interestingly, the second domain, 'Shot Noise', shows almost no adaptation despite the lower performance of the Direct-Test. We conjecture that the preceding domain, 'Gaussian Noise', shares a similar nature of noise, leading the model to decide that additional adaptation steps may not be necessary. As a result, our skipping strategy enables the model to efficiently adapt, considering both the original domain the model is trained on and the previous domain the model has been adapted to.", + "bbox": [ + 496, + 90, + 893, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 693, + 619, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce an efficient Continual Test-time Adaptation (CTA) method for object detection in the continually changing domain. Our approach involves 1) lightweight adaptors, 2) class-wise object-level feature alignment, and 3) skipping unnecessary adaptation, offering highly efficient and effective solution to diverse domain shifts. It significantly improve mAP performance in various CTA scenarios without serious slowdown in the inference speed.", + "bbox": [ + 496, + 718, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by NRF grant (2021R1A2C3006659) and IITP grants (2021-0-01343, 2022-0-00320, 2019-0-00079), all funded by the Korea government (MSIT).", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander Bartler, Florian Bender, Felix Wiewel, and Bin Yang. Ttaps: Test-time adaption by aligning prototypes using self-supervision. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 2", + "[2] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2", + "[3] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1", + "[4] Dhanajit Brahma and Piyush Rai. A probabilistic framework for lifelong test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3582-3591, 2023. 3", + "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 3", + "[6] Yijin Chen, Xun Xu, Yongyi Su, and Kui Jia. Stfar: Improving object detection robustness at test-time by self-training with feature alignment regularization. arXiv preprint arXiv:2303.17937, 2023. 1, 2, 3", + "[7] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 3", + "[8] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 3", + "[9] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, et al. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[10] Shirsendu Sukanta Halder, Jean-François Lalonde, and Raoul de Charette. Physics-based rendering for improving robustness to rain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10203-10212, 2019. 3", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6, 7, 2", + "[12] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3", + "[13] Junyuan Hong, Lingjuan Lyu, Jiayu Zhou, and Michael Spranger. Mecta: Memory-economic continual test-time model adaptation. In The Eleventh International Conference on Learning Representations, 2022. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2", + "[16] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for modelagnostic domain generalization. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[17] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In International Conference on Learning Representations (ICLR), 2023. 3", + "[18] Sanghun Jung, Jungsoo Lee, Nanhee Kim, Amirreza Shaban, Byron Boots, and Jaegul Choo. Cafa: Class-aware feature alignment for test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4", + "[19] Mehran Khodabandeh, Arash Vahdat, Mani Ranjbar, and William G Macready. A robust learning approach to domain adaptive object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 480-490, 2019. 3", + "[20] Donghyun Kim, Kaihong Wang, Stan Sclaroff, and Kate Saenko. A broad study of pre-training for domain generalization and adaptation. In European Conference on Computer Vision, pages 621-638. Springer, 2022. 5", + "[21] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6092–6101, 2019. 3", + "[22] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8474-8481, 2021. 4", + "[23] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation, 2017. 3", + "[24] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test-time adaptation, 2023. 2, 3", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23362", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 1, 2", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6, 7", + "[29] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 5", + "[30] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2, 6, 1", + "[31] Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin, Mateusz Kozinski, Horst Possegger, and Horst Bischof. Actmad: Activation matching to align distributions for test-time-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24152-24161, 2023. 1, 2, 3, 4, 6", + "[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 3, 7", + "[33] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400, 2023. 3", + "[34] Mario obler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 3", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. 2016. 5", + "[36] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2, 6, 1", + "[38] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2759-2769, 2023. 1, 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 6", + "[40] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. Advances in Neural Information Processing Systems, 35:17543-17555, 2022. 3, 4", + "[41] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 5", + "[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2", + "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[44] Vibashan VS, Poojan Oza, and Vishal M Patel. Towards online domain adaptive object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 478-488, 2023. 3", + "[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 7", + "[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3", + "[47] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees GM Snoek. Energy-based test sample adaptation for domain generalization. arXiv preprint arXiv:2302.11215, 2023. 3", + "[48] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5", + "[49] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 1, 2, 3", + "[50] Bowen Zhao, Chen Chen, and Shu-Tao Xia1. Delta: Degradation-free fully test-time adaptation. In International Conference on Learning Representations (ICLR), 2023. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23363", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_model.json b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da5ca99b6a3635b026111ab9d6b49133d1933a1d --- /dev/null +++ b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "What, How, and When Should Object Detectors Update in Continually Changing Test Domains?" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.834, + 0.24 + ], + "angle": 0, + "content": "Jayeon Yoo\\(^{1}\\) Dongkwan Lee\\(^{1}\\) Inseop Chung\\(^{1}\\) Donghyun Kim\\(^{2*}\\) Nojun Kwak\\(^{1*}\\) Seoul National University Korea University" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.241, + 0.783, + 0.258 + ], + "angle": 0, + "content": "1{jayeon.yoo, biancco, jis3613, nojunk}@snu.ac.kr, \\(^2\\mathrm{d}\\) kim@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.485, + 0.764 + ], + "angle": 0, + "content": "It is a well-known fact that the performance of deep learning models deteriorates when they encounter a distribution shift at test time. Test-time adaptation (TTA) algorithms have been proposed to adapt the model online while inferring test data. However, existing research predominantly focuses on classification tasks through the optimization of batch normalization layers or classification heads, but this approach limits its applicability to various model architectures like Transformers and makes it challenging to apply to other tasks, such as object detection. In this paper, we propose a novel online adaption approach for object detection in continually changing test domains, considering which part of the model to update, how to update it, and when to perform the update. By introducing architecture-agnostic and lightweight adaptor modules and only updating these while leaving the pre-trained backbone unchanged, we can rapidly adapt to new test domains in an efficient way and prevent catastrophic forgetting. Furthermore, we present a practical and straightforward class-wise feature aligning method for object detection to resolve domain shifts. Additionally, we enhance efficiency by determining when the model is sufficiently adapted or when additional adaptation is needed due to changes in the test distribution. Our approach surpasses baselines on widely used benchmarks, achieving improvements of up to \\(4.9\\% p\\) and \\(7.9\\% p\\) in mAP for COCO \\(\\rightarrow\\) COCO-corrupted and SHIFT, respectively, while maintaining about 20 FPS or higher. The implementation code is available at https://github.com/natureyoo/ContinualTTA_ObjectDetection." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.791, + 0.21, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Deep learning models excel in many vision tasks but struggle with domain shifts when test data distributions diverge from training data [3, 27, 42]. These shifts, common in real-world settings, arise from natural variations, weather" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.29, + 0.903, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.653 + ], + "angle": 0, + "content": "Figure 1. We propose an online adaptation method for object detection in continually changing test domains. Object detectors trained with clean images suffer from performance degradation due to various corruption, such as camera sensor degradation or environmental changes (Direct-Test). Updating full parameters for online adaptation require a large number of test samples and vulnerable to drastic domain changes (Full-Finetuning), while using only our lightweight adaptor is robust and quickly adapts within a few time steps (Ours). We can further improve efficiency by skipping unnecessary adaptation steps (Ours-Skip)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.854 + ], + "angle": 0, + "content": "changes (e.g., fog, rain), camera differences (e.g., pixelate, defocus blur), and various other factors. Test-Time Adaptation (TTA) [3, 27, 32, 42, 45, 49] addresses these shifts by adapting models to the target (test) distribution online. Considering the dynamic nature of real-world applications, it becomes crucial to extend this approach to Continual Test-Time Adaptation (CTA), which accounts for the ongoing evolution of test distributions. This is particularly vital in environments like autonomous driving, where conditions can change continuously (e.g., from clear to rainy, or day to night). Despite its significance, continual test-time adaptation for object detection remains underexplored." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Recently, several TTA methods [6, 31, 38] tailored for object detection have been proposed. ActMAD [31] aligns all the output feature maps \\((\\mathbb{R}^{C\\times H\\times W})\\) after Batch Nor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Co-corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23354" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "malization (BN) layers [15] to adapt the test domain to be similar to that of the training domain. However, this approach requires significant memory during adaptation and does not explicitly consider the objects present in the image. TeST [38] and STFAR [6] adapt to a test domain by utilizing weak and strong augmented test samples with a teacher-student network [39], but they significantly increase inference costs since they require additional forward passes and update steps. Also, these methods update all network parameters, making them highly inefficient in online adaptation and vulnerable to losing task-specific knowledge when the test domain experiences continual or drastic changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.626 + ], + "angle": 0, + "content": "In this paper, we aim to develop an efficient continual test-time adaptation (CTA) method for object detection. We investigate the following three key aspects to improve efficiency; what to update: while previous TTA methods for object detection [6, 31, 38] use full fine-tuning, updating all parameters at test time, they are inefficient and prone to losing task-specific knowledge in relatively complex object detection tasks. Updating BN layers, as done in many TTA methods for classification [18, 32, 45, 49], is not as effective for object detection, given its smaller batch size compared to classification and the limitation in applying various backbones, such as Transformer [28, 43]. how to update: several previous TTA methods for object detection [6, 38] adapt the model by using teacher-student networks, resulting in a significant decrease in inference speed, which is detrimental during test time. While another existing method [31] aligns feature distributions for adaptation, it does not consider each object individually, focusing only on image features, making it less effective for object detection. when to update: most TTA or CTA methods update models using all incoming test samples. However, it is inefficient to update continuously the model if it is already sufficiently adapted when the change of the test domain is not significant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To this end, (1) we propose an efficient continual test-time adaptation method for object detectors to adapt to continually changing test domains through the use of lightweight adaptors which require only \\(0.54\\% \\sim 0.89\\%\\) additional parameters compared to the full model. It exhibits efficiency in parameters, memory usage, and adaptation time, along with robustness to continuous domain shifts without catastrophic forgetting. Additionally, it demonstrates wide applicability to various backbone types compared to BN-based TTA methods [18, 24, 32, 45, 49, 50]. (2) To enhance the adaptation effectiveness in the object detection task, we align the feature distribution of the test domain with that of the training domain at both the image-level and object-level using only the mean and variance of features. For estimating the mean of the test domain features, we employ Exponentially Moving Average (EMA) as we can leverage only the current incoming test samples, not the entire test domain data. Due to the unavailability of training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.257 + ], + "angle": 0, + "content": "data access, we utilize only the mean and variance of the features from a few training samples. (3) We also introduce two novel criteria that do not require additional resources to determine when the model needs adaptation to enhance efficiency in a continually changing test domain environment. As illustrated in Fig. 1, our approach Ours, employing adaptors, tends to adapt much faster to domain changes compared to full parameter updates. This enables efficient TTA by using only a few test samples to update the adaptor and skipping the rest of the updates as shown in Ours-Skip. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.259, + 0.895, + 0.379 + ], + "angle": 0, + "content": "- We introduce an architecture-agnostic lightweight adaptor, constituting only a maximum of \\(0.89\\%\\) of the total model parameters, into the backbone of the object detector to adapt the model in a continually changing test domain. This approach ensures efficiency in parameters, memory usage, and adaptation speed, demonstrating the robust preservation of task-specific knowledge owing to its inherent structural characteristics." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.381, + 0.894, + 0.47 + ], + "angle": 0, + "content": "- We propose a straightforward and effective adaptation loss for CTA in object detection tasks. This is achieved by aligning the distribution of training and test domain features at both the image and object levels, utilizing only the mean and variance of a few training samples and EMA-updated mean features of the test domain." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.471, + 0.894, + 0.547 + ], + "angle": 0, + "content": "- We also propose two criteria to determine when the model requires adaptation, enabling dynamic skipping or resuming adaptation as needed. This enhancement significantly boosts inference speed by up to about 2 times while maintaining adaptation performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.547, + 0.895, + 0.637 + ], + "angle": 0, + "content": "- Our adaptation method proves effective for diverse types of domain shifts, including weather changes and sensor variations, regardless of whether the domain shift is drastic or continuous. In particular, our approach consistently improves the mAP by up to \\(7.9\\%\\) in \\(\\mathrm{COCO} \\rightarrow \\mathrm{COCO-C}\\) and SHIFT-Discrete/Continuous with higher than 20 FPS." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.259, + 0.895, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.642, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Test-time adaptation. Recently, there has been a surge of interest in research that adapts models online using unlabeled test samples while simultaneously inferring the test sample to address the domain shift problem, where the test data distribution differs from that of the training data. There are two lines for online adaptation to the test domain, Test-time Training (TTT) and Test-time Adaptation (TTA). TTT [1, 2, 27, 42] involves modifying the model architecture during training to train it with self-supervised loss, allowing adaptation to the test domain in the test time by applying this self-supervised loss to the unlabeled test samples. On the other hand, TTA aims to adapt the trained model directly to the test domain without specifically tailored model architectures or losses during training time. NORM [37] and DUA [30] address the domain shifts by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23355" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "adjusting the statistics of batch normalization (BN) layers using the current test samples, without updating other parameters, inspired by [23]. Following this, [24, 32, 45, 50] and [18] update the affine parameters of BN layers using unsupervised loss, entropy minimization loss to enhance the confidence of test data predictions, and feature distribution alignments loss, respectively. Several studies [16, 17] update the classifier head using the pseudo-prototypes from the test domain. However, these methods limit their applicability to architectures without BN layers or to object detection tasks that involve multiple objects in a single image. Others [31, 40, 49] update full parameters for online adaptation to the test domain in an online manner, but this approach is inefficient and susceptible to the noisy signal from the unsupervised loss. While existing TTA methods are oriented towards classification tasks, we aim to propose an effective and efficient method for online adaptation in the object detection task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.732 + ], + "angle": 0, + "content": "Continual test-time adaptation. Recent studies [33, 46] point out that existing TTA methods have primarily focused on adapting to test domains following an i.i.d assumption and may not perform well when the test data distribution deviates from this assumption. [46] introduces a ContinuaTTA (CTA) method designed for scenarios where the test domain continuously changes over time. This poses challenges in preventing the model from over-adapting to a particular domain shift and preserving the knowledge of the pre-trained model to avoid catastrophic forgetting. In the field of CTA, the self-training strategy adopting an Exponentially Moving Average (EMA) teacher-student structure is attracting interest as an effective algorithm enabling robust representation to be learned through self-knowledge distillation. In many studies, the EMA teacher-student structure and catastrophic restoration of source model weights have been proposed as a solution to achieve the goal of CTA [4, 46, 47]. Approaches using source replay [34], and anti-forgetting regularization [32] have also achieved good performances in robust continuous adaptation. Furthermore, there is growing attention on methods that mitigate the computational and memory challenges associated with CTA, such as [13], which relies on updates to batch normalization statistics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Test-time adaptive object detection. Research on TTA for Object Detection (TTAOD) is progressively emerging [6, 31, 38, 44]. Most existing TTAOD methods [6, 38, 44] exploit a teacher-student network to adapt to the test domain, following the self-training approach commonly employed in Unsupervised Domain Adaptation for object detection [7, 19, 21, 36]. However, it is inefficient for TTA due to data augmentation requirements and additional forward and backward steps, resulting in slower inference speeds and higher memory usage. Another approach, ActMAD [31], aligns the distributions of output feature maps" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.214 + ], + "angle": 0, + "content": "after all BN layers along the height, width, and channel axes to adapt to the test domain. However, this location-aware feature alignment is limited to datasets with fixed location priors, such as driving datasets, and is less effective for natural images like COCO. Additionally, CTA for Object Detection (CTAOD) have not been thoroughly explored. Therefore, there is a need for an effective CTAOD method considering memory and time efficiency." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.228, + 0.593, + 0.243 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.895, + 0.346 + ], + "angle": 0, + "content": "To enable the efficient and effective Continual Test-time Adaptation of Object Detectors (CTAOD), we introduce an approach that specifies which part of the model should be updated, describes how to update those using unlabeled test data, and determines whether we perform model updates or not to improve efficiency." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.357, + 0.63, + 0.373 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Assume that we have an object detector \\( h \\circ g_{\\Theta} \\), here \\( h \\) and \\( g \\) are the RoI head and the backbone, respectively with their parameters being \\( \\Theta \\). The training dataset is denoted as \\( D_{train} = \\{(x_i, y_i)\\}_{i=1}^N \\), where \\( x_i \\sim P_{train}(x) \\) and \\( y_i = (bbox_i, c_i) \\), containing information on the bounding box (bbox) and class label \\( c_i \\in \\mathcal{C} \\). Consider deploying the detector to the test environments where the test data at period \\( T \\) is denoted as \\( x_j^T \\sim P_{test}^T(x) \\), \\( P_{test}^T \\neq P_{train} \\) and \\( P_{test}^T \\) deviates from the i.i.d. assumption. In addition, the domain of \\( P_{test}^T \\) continually changes according to \\( T \\) (i.e., \\( P_{test}^T \\neq P_{test}^{T-1} \\)). Our goal is to adapt the detector \\( h \\circ g \\) to \\( P_{test}^T \\) using only test data \\( x_j^T \\) while making predictions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.575, + 0.871, + 0.592 + ], + "angle": 0, + "content": "3.2. What to update: Adaptation via an adaptor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Previous methods [6, 31, 38, 44] adapt the model to the test domain by updating all parameters \\(\\Theta\\), leading to inefficiency at test time and a high risk of losing task knowledge from the training data. In contrast, we adapt the model by introducing an adaptor with an extremely small set of parameters and updating only this module while freezing \\(\\Theta\\). We introduce a shallow adaptor in parallel for each block, inspired by [5, 14], where transformer-based models are fine-tuned for downstream tasks through parameter-efficient adaptors, as shown in Fig. 2. Each adaptor consists of down-projection layers \\(\\mathcal{W}_{down} \\in \\mathbb{R}^{d \\times \\frac{d}{r}}\\), up-projection layers \\(\\mathcal{W}_{up} \\in \\mathbb{R}^{\\frac{d}{r} \\times d}\\) and RLUs, where \\(d\\) denotes the input channel dimension and \\(r\\) is the channel reduction ratio set to 32 for all adaptors. We use MLP layers for the Transformer block (Fig. 2a) and \\(1 \\times 1\\) convolutional layers for the ResNet block (Fig. 2b) to introduce architecture-agnostic adaptors. The up-projection layer is initialized to 0 values so that the adaptor does not modify the output of the block, but as the adaptor is gradually updated, it adjusts the output of the block to adapt to the test domain. Even as the adaptor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23356" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.09, + 0.278, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.48, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.471, + 0.334 + ], + "angle": 0, + "content": "Figure 2. We attach an adaptor, which is a shallow and low-rank MLP or CNN, to every \\(N\\) block in parallel. We update only these adaptors while other parameters are frozen. Our approach can be applied to diverse architectures including CNNs and Transformers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.471, + 0.445 + ], + "angle": 0, + "content": "is updated in the test domain, the original backbone parameter \\(\\Theta\\) remains frozen and fully preserved. This structural preservation, as evident in Ours in Fig. 1, enables robust and efficient adaptation to domain changes by maintaining relatively complex task knowledge in object detection and updating very few parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.456, + 0.42, + 0.472 + ], + "angle": 0, + "content": "3.3. How to update: EMA feature alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.471, + 0.737 + ], + "angle": 0, + "content": "To adapt the object detector to the test domain, we align the feature distribution of the test domain with that of the training data, inspired by [18, 31, 40]. In contrast to these methods that solely align image feature distribution, we additionally align object-level features in a class-wise manner, considering class frequency, to enhance its effectiveness for object detection. As the training data is not accessible during test time, we pre-compute the first and second-order statistics, denoted as \\(\\mu_{tr} = \\mathbb{E}[F_{tr}]\\) and \\(\\Sigma_{tr} = \\mathbb{Var}[F_{tr}]\\), where the operators \\(\\mathbb{E}\\) and \\(\\mathbb{Var}\\) represent the mean and variance respectively. The features \\(F_{tr} = \\{g_{\\Theta}(x_{tr})\\}\\) are computed using only 2,000 training samples, a small subset of the training data. Since a sufficient amount of test domain data is not available at once, and only the current incoming test data, whose features are denoted as \\(F_{te}^{t}\\), is accessible at time step \\(t\\), we estimate the mean of test data features using an exponentially moving average (EMA) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.749, + 0.47, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t e} ^ {t} = (1 - \\alpha) \\cdot \\mu_ {t e} ^ {t - 1} + \\alpha \\cdot \\mathbb {E} [ F _ {t e} ^ {t} ], \\quad \\mathrm {s . t .} \\mu_ {t e} ^ {0} = \\mu_ {t r}. (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.826 + ], + "angle": 0, + "content": "Considering the typically small batch size in object detection compared to classification, we approximate the variance of the test features as \\(\\Sigma_{te} \\simeq \\Sigma_{tr}\\) to reduce instability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Image-level feature alignment. We estimate the training and test feature distributions as normal distributions and minimize the KL divergence between them as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {i m g} = \\mathrm {D} _ {K L} (\\mathcal {N} (\\mu_ {t r}, \\Sigma_ {t r}), \\mathcal {N} (\\mu_ {t e} ^ {t}, \\Sigma_ {t r})). \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "Region-level class-wise feature alignment. In object detection, we deal with multiple objects within a single image, making it challenging to apply the class-wise feature alignment proposed in [40], a TTA method for classification. To handle region-level features that correspond to an object, we use ground truth bounding boxes for the training data and utilize the class predictions of RoI pooled features, \\( f_{te}^{t} \\), for unlabeled test data. In object detection, domain shifts often result in lower recall rates, as a significant number of proposals are predicted as background [22]. To mitigate this issue, we filter out features with background scores exceeding a specific threshold. Subsequently, we assign them to the foreground class with the highest probability, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.297, + 0.891, + 0.327 + ], + "angle": 0, + "content": "\\[\nF _ {t e} ^ {k, t} = \\left\\{f _ {t e} ^ {t} \\mid \\underset {c} {\\operatorname {a r g m a x}} p _ {f g} = k, p _ {b g} < 0. 5 \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.325, + 0.858, + 0.342 + ], + "angle": 0, + "content": "\\[\nw h e r e h _ {c l s} \\left(f _ {t e} ^ {t}\\right) = \\left[ p _ {f g}, p _ {b g} \\right] = \\left[ p _ {0}, \\dots , p _ {C - 1}, p _ {b g} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.893, + 0.442 + ], + "angle": 0, + "content": "We estimate the class-wise feature distribution of the test domain by exploiting \\( F_{te}^{k,t} \\) and Eq.1. Furthermore, we introduce a weighting scheme for aligning features of less frequently appearing classes, taking into account the severe class imbalance where specific instance (e.g., person) may appear multiple times within a single image, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.45, + 0.83, + 0.47 + ], + "angle": 0, + "content": "\\[\nN ^ {k, t} = N ^ {k, t - 1} + \\left\\| F _ {t e} ^ {k, t} \\right\\|, \\quad \\text {s . t .} \\quad N ^ {k, 0} = 0\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.472, + 0.892, + 0.507 + ], + "angle": 0, + "content": "\\[\nw ^ {k, t} = \\log \\left(\\frac {\\operatorname* {m a x} _ {i} N ^ {i , t}}{N ^ {k , t}}\\right) + 0. 0 1 \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.509, + 0.862, + 0.54 + ], + "angle": 0, + "content": "\\[\nL _ {o b j} = \\sum_ {k} w ^ {k, t} \\cdot D _ {K L} (\\mathcal {N} (\\mu_ {t r} ^ {k}, \\Sigma_ {t r} ^ {k}), \\mathcal {N} (\\mu_ {t e} ^ {k, t}, \\Sigma_ {t r} ^ {k})).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.63 + ], + "angle": 0, + "content": "Here, the class-wise mean \\(\\mu^k\\) and variance \\(\\Sigma^k\\) of the training and test data are obtained in the same way as the image-level features. We can effectively adapt the object detector by updating the model to align the feature distribution at both the image and object levels as \\(L = L_{img} + L_{obj}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.637, + 0.848, + 0.653 + ], + "angle": 0, + "content": "3.4. When to update: Adaptation on demand" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.809 + ], + "angle": 0, + "content": "As shown in Fig. 1, Ours, which only updates the adaptor proposed in Sec. 3.2, efficiently adapts to changes in the test domain, even with a small subset of early test samples. We leverage its rapid adaptation characteristics to reduce computational costs by skipping model updates (i.e., skipping backward passes) when the model has already sufficiently adapted to the current test domain and resuming model updates when confronted with a new test domain. Therefore, we introduce two criteria to determine when to update the model or not as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "(Criterion 1) When the distribution gap exceeds the in-domain distribution gap. Recall that \\(L_{img}\\) (Eq. 2) measures the distribution gap between the test and train distributions. We assume a model is well-adapted to the current test domain when \\(L_{img}\\) is closer to the in-domain distribution gap. We measure the in-domain distribution gap by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.445, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.181, + 0.345, + 0.196 + ], + "angle": 0, + "content": "(a) The ratio of \\( L_{img} \\) to \\( D_{KL}^{in} \\)" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.196, + 0.446, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.181, + 0.284, + 0.346, + 0.298 + ], + "angle": 0, + "content": "(b) The ratio of \\( L_{img} \\) to \\( L_{ema}^{t} \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.301, + 0.468, + 0.329 + ], + "angle": 0, + "content": "Figure 3. The test domain undergoes a shift every 4,000 time steps, and each metric reaches its peak at the same intervals." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.349, + 0.468, + 0.379 + ], + "angle": 0, + "content": "sampling two disjoint subsets, \\( x_{i} \\) and \\( x_{j} \\), of training features \\( F_{tr} \\) from Sec. 3.3 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.39, + 0.469, + 0.409 + ], + "angle": 0, + "content": "\\[\nD _ {K L} ^ {i n} = D _ {K L} \\left(\\mathcal {N} \\left(\\mu_ {t r} ^ {i}, \\Sigma_ {t r} ^ {i}\\right), \\mathcal {N} \\left(\\mu_ {t r} ^ {j}, \\Sigma_ {t r} ^ {j}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.469, + 0.641 + ], + "angle": 0, + "content": "where \\(\\mu_{tr}^{i},\\Sigma_{tr}^{i}\\) are obtained from \\(x_{i}\\sim P_{train}(x)\\) and \\(\\mu_{tr}^{j},\\Sigma_{tr}^{j}\\) from \\(x_{j}\\sim P_{train}(x)\\). In other words, if \\(L_{img}\\) is noticeably higher than the in-domain distribution gap \\(D_{KL}^{in}\\), we consider a model encountering a test domain whose distribution differs from \\(P_{train}(x)\\) and needs to be updated. Based on this, we introduce a new index \\(\\frac{L_{img}}{D_{KL}^{in}}\\). Fig. 3a plots the trend of this index during the model adaptation to a continually changing test domain. It shows that the index has a large value in the early stages of a domain change, decreases rapidly, and then maintains a value close to 1. This index exhibits a similar trend regardless of the backbone type and dataset, as included in the appendix. Therefore, we establish the criterion that model updates are necessary when this index exceeds a certain threshold, \\(\\tau_{1}\\), as \\(\\frac{L_{img}}{D_{KL}^{in}} >\\tau_{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.906 + ], + "angle": 0, + "content": "(Criterion 2) When the distribution gap suddenly increases. Additionally, we can determine when the test distribution changes and model updates are necessary by observing the trend of the distribution gap (i.e., \\( L_{img} \\)). The convergence of \\( L_{img} \\) indicates that a model is well-adapted to the current test domain. To put it differently, \\( L_{img} \\) will exhibit a sudden increase when the model encounters a new test domain. We introduce an additional index, denoted as \\( \\frac{L_{img}}{L_{ema}^t} \\), representing the ratio of the current \\( L_{img} \\) to its exponentially moving average \\( L_{ema}^t \\) at time \\( t \\). We calculate it using the following formula: \\( L_{ema}^t = 0.99 \\cdot L_{ema}^{t-1} + 0.01 \\cdot L_{img} \\). Fig. 3b illustrates the trend of the ratio of \\( L_{img} \\) over the timesteps. It tends to reach a value of 1 as the loss stabilizes at a specific level. Nevertheless, when the model encounters shifts in the test distribution, the ratio experiences a sharp increase, indicating the necessity of a model update when it exceeds a specific threshold, \\( \\tau_2 \\), as \\( \\frac{L_{img}}{L_{ema}^t} > \\tau_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "If at least one of the two criteria is satisfied, we conclude that the model requires adaptation and proceed to update it." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.134, + 0.632, + 0.151 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.159, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Sec. 4.1 presents the two object detection benchmark datasets with test distributions that change continuously, either in a drastic or gradual manner, and our implementation detail is in 4.2. Sec. 4.4 compares our method with other TTA baselines described in Secs. 4.3.. We present detailed ablation studies of our method analyzing the effectiveness and efficiency of our method in terms of what, how, and when to update the models for CTAOD in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.602, + 0.303 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.312, + 0.833, + 0.326 + ], + "angle": 0, + "content": "We experiment with the following three scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.718 + ], + "angle": 0, + "content": "\\(\\mathrm{COCO} \\rightarrow \\mathrm{COCO-C}\\) simulates continuous and drastic realistic test domain changes over a long sequence. MS-COCO [25] collects 80 classes of common objects in their natural context with 118k training images and 5k validation images. COCO-C is created by employing 15 types of realistic corruptions [29], such as image distortion and various weather conditions, to simulate test domain changes. In the experiments, the model is only trained on the COCO train set and sequentially evaluated on each corruption in the COCO-C validation set during test-time for reproducing continually changing test domains. Finally, the model is evaluated on the original COCO validation set to assess how well it preserves knowledge of the original domain (denoted as Org.). SHIFT-(Discrete / Continuous) [41] is a synthetic driving image dataset with 6 classes under different conditions using five weather attributes (clear, cloudy, overcast, fog, rain) and three time-of-day attributes (daytime, dawn, night). In SHIFT-Discrete, there are image sets for each attribute, and the model is sequentially evaluated on these attributes, cloudy \\(\\rightarrow\\) overcast \\(\\rightarrow\\) foggy \\(\\rightarrow\\) rainy \\(\\rightarrow\\) dawn \\(\\rightarrow\\) night \\(\\rightarrow\\) clear which contains 2.4k, 1.6k, 2.7k, 3.2k, 1.2k, 1.4k, and 2.8k validation images, respectively. This simulates scenarios where the domain undergoes drastic changes. In SHIFT-Continuous, the model is evaluated on four sequences, each consisting of 4k frames, continuously transitioning from clear to foggy (or rainy) and back to clear." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.71, + 0.743 + ], + "angle": 0, + "content": "4.2. Implementation Detail" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on Faster-RCNN [35] with FPN [26], using ImageNet1k pre-trained ResNet50 [11] and ImageNet21k pre-trained Swin-Tiny [28] backbones. Specifically, Swin-Tiny is employed for its superior domain transfer [20] and detection efficacy over other advanced backbones [9]. For COCO \\(\\rightarrow\\) COCO-C with ResNet50, we use a publicly available model trained with strong augmentation [48]. Other models are trained using detectron2 framework following [35] and [28]. For test-time adaptation, we always set the learning to 0.001 for the SGD opti" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23358" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.894, + 0.131 + ], + "angle": 0, + "content": "Table 1. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on COCO \\(\\rightarrow\\) COCO-C. Our model consistently outperforms baselines on the two different backbones. Furthermore, Ours-Skip with ResNet notably reduces backward passes by as much as \\(90.5\\%\\), leading to a significantly improved frames per second (FPS) rate by up to \\(109.9\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.133, + 0.895, + 0.304 + ], + "angle": 0, + "content": "
BackboneMethodNoiseBlurWeatherDigitalOrg.Avg.# step
GauShtImpDefGlsMtnZmSnwFrsFogBrtCntElsPxJpgFor.Back.FPS
Swin-T [28]Direct-Test9.711.410.013.47.512.15.220.724.836.136.012.919.14.915.843.017.780K021.5
ActMAD10.712.09.412.35.79.54.515.317.527.628.21.116.72.68.736.313.980K80K8.3
Mean-Teacher10.012.111.212.88.112.14.919.623.734.934.08.018.96.117.641.017.2160K80K6.9
Ours13.616.616.114.013.614.28.323.727.237.436.427.227.222.222.342.322.680K80K9.5
Ours-Skip13.315.315.114.012.813.96.522.025.435.534.926.525.923.420.241.221.680K9.7K17.7
ResNet50 [11]Direct-Test9.111.09.812.64.58.84.619.123.138.438.021.415.65.311.944.217.380K025.8
NORM9.911.911.012.65.29.15.119.423.538.237.622.417.25.710.343.417.580K025.8
DUA9.811.710.812.85.28.95.119.323.738.437.822.317.25.410.144.117.180K025.8
ActMAD9.19.67.011.03.26.13.312.814.027.727.83.912.92.37.234.310.580K80K9.6
Mean-Teacher9.612.512.04.02.94.83.116.223.535.134.021.816.68.212.740.314.5160K80K8.1
Ours12.717.817.512.411.511.36.622.826.938.638.528.025.121.222.241.822.280K80K10.1
Ours-Skip14.417.116.013.911.712.26.322.125.537.737.125.524.123.121.142.821.980K7.6K21.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.368 + ], + "angle": 0, + "content": "mizer, and \\(\\alpha\\) of Eq. 1 to 0.01, while \\(\\tau_{1}\\) and \\(\\tau_{2}\\) are set to 1.1 and 1.05, respectively. We use the same hyper-parameters across all backbones and datasets, with a batch size of 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.377, + 0.186, + 0.392 + ], + "angle": 0, + "content": "4.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.4, + 0.471, + 0.643 + ], + "angle": 0, + "content": "Direct-Test evaluates the model trained in the training domain without adaptation to the test domain. ActMAD [31] is a TTA method aligning the distribution of output features across all BN layers. To apply ActMAD to the Swin Transformer-based model, we align the output features of the LN layers. We implement Mean-Teacher using a teacher-student network framework to reproduce as close as possible to TeST [38], as its implementation is not publicly available. We follow the FixMatch [39] augmentation method and report results after tuning all hyper-parameters in our scenario. NORM [37] and DUA [30], TTA methods initially designed for classification, are directly applicable to detection tasks by either mixing a certain amount of current batch statistics or updating batch statistics via EMA. However, these are only compatible with architectures containing BN layers. Additional details are provided in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.216, + 0.667 + ], + "angle": 0, + "content": "4.4. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.75 + ], + "angle": 0, + "content": "We compare the performance of each method using mAP and efficiency metrics, including the number of forward and backward passes, as well as FPS during test-time adaptation. Results of COCO and SHIFT are in Tab. 1 and 2, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "\\(\\mathbf{COCO} \\rightarrow \\mathbf{COCO-C}\\). Tab. 1 demonstrates the effective adaptation performance of Ours in the challenging COCO benchmark with 80 classes due to object-level class-wise feature alignment. ActMAD also aligns feature distribution for TTA, but is not effective since it only aligns whole feature maps without considering specific classes in the image. NORM and DUA, applicable only to ResNet [11], show minimal performance improvement by adaptation as they are not specifically tailored for object detection and only modify batch statistics across the entire feature map. Ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.581 + ], + "angle": 0, + "content": "ditionally, ActMAD and Mean-Teacher, updating full parameters, gradually lose task knowledge in the continually changing test distributions, resulting in much lower performance on Org., the domain identical to the training data, than that of Direct-Test. In contrast, Ours effectively prevents catastrophic forgetting by freezing the original parameters of the models and updating only the adaptor, obtaining performance on par with Direct-Test on the Org. domain and consistently high performance across corrupted domains, with an average mAP improvement of \\(4.9\\%\\) compared to that of Direct-Test. Furthermore, leveraging the rapid adaptation ability of the adaptor, Ours-Skip, which skips unnecessary adaptation, allows using only a maximum of about \\(12\\%\\) of the total samples for adaptation without significant performance loss. This leads to a substantial improvement in inference speed, more than doubling compared to other TTA methods, reaching over 17.7 FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "SHIFT-Discrete. Ours is also effective in SHIFT, which simulates continuous changes in weather and time in driving scenarios according to the left section of Tab. 2. Especially, Ours shows significant improvements in mAP by \\(7 - 9\\% \\mathrm{p}\\), particularly for the foggy and dawn attributes where Direct-Test obtains lower performance due to severe domain shift. In contrast, with ActMAD, catastrophic forgetting takes place when adapting to the cloudy and overcast weather. This is due to the updating of the full parameters, despite that Direct-Test already shows proficient performance in these conditions. As a result, the performance in the later domains is worse than that of the Direct-Test. DUA, which updates batch statistics using EMA, shows a gradual decrease in performance as the domain continuously changes, resulting in much lower performance in the original clear domain (i.e., clear). On the other hand, NORM, which utilizes the statistics of the current batch samples, exhibits no catastrophic forgetting and relatively good adaptation, as SHIFT is a relatively easier task compared to COCO due to having only 6 classes. Compared to NORM, Ours shows better adaptation performance, and is" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23359" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.146 + ], + "angle": 0, + "content": "Table 2. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on SHIFT-Discrete and SHIFT-Continuous. Baselines perform effectively in a particular setting but lack generalizability across various settings. Our method consistently achieves results that are either better or on par with the best model in all settings, demonstrating its strong stability. Ours-Skip also effectively reduces the number of backward passes without compromising mAP performance, resulting in a higher FPS." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.147, + 0.895, + 0.332 + ], + "angle": 0, + "content": "
BackboneMethodSHIFT-DiscreteSHIFT-Continuous
mAP# stepmAP# Avg. step
cloudyoverc.fograindawnnightclearAvg.For.Back.FPSclear↔fogclear↔rainFor.Back.FPS
Swin-T [28]Direct-Test50.038.923.145.126.939.545.938.515.3K027.518.121.14K028.3
ActMAD49.838.421.443.119.032.044.835.515.3K15.3K9.315.616.34K4K9.8
Mean-Teacher50.039.225.745.426.037.542.238.030.6K15.3K7.820.424.38K4K6.5
Ours50.339.232.246.730.439.944.340.415.3K15.3K11.223.922.64K4K11.6
Ours-Skip50.339.729.147.130.241.545.940.615.3K6.1K20.025.123.84K0.83K19.2
ResNet50 [11]Direct-Test49.437.919.743.120.135.345.635.915.3K030.112.115.44K030.0
NORM49.738.622.944.725.137.445.537.715.3K030.116.919.44K030.0
DUA45.231.527.731.915.218.621.127.315.3K030.122.522.44K030.0
ActMAD49.237.718.040.616.032.944.334.115.3K15.3K11.312.716.34K4K11.2
Mean-Teacher49.638.426.843.426.633.141.637.130.6K15.3K9.916.020.88K4K9.8
Ours49.738.727.446.327.437.643.838.715.3K15.3K12.920.921.94K4K13.9
Ours-Skip49.738.826.946.227.638.845.039.015.3K8.9K21.520.022.54K0.75K21.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.405 + ], + "angle": 0, + "content": "Table 3. Comparison of adaptation performance (mAP), the number of trainable parameters (# Params), and memory usage (Cache) according to which part of the backbone is updated. SD / SC denotes SHIFT-Discrete/Continuous, respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.407, + 0.472, + 0.531 + ], + "angle": 0, + "content": "
BackboneTrainable ParamsmAP# ParamsCache
SDSCNumRatioAvg.Max
Swin-TFull-params38.420.627.7M100%0.8611.0
LayerNorm38.520.00.03M0.1%0.657.49
adaptor (Ours)40.423.20.15M0.5%0.656.96
ResNet50Full-params37.620.423.7M100%1.659.29
BatchNorm37.920.20.05M0.2%1.479.11
adaptor (Ours)38.721.70.21M0.9%1.485.41
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.424, + 0.565 + ], + "angle": 0, + "content": "also applicable to BN-layer-free Swin Transformers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.779 + ], + "angle": 0, + "content": "SHIFT-Continuous. In scenarios where the test domain gradually changes across the entire sequence, Ours also demonstrates effectiveness, improving mAP by up to \\(7\\%\\) p, as shown in the right section of Tab. 2. While DUA performs well in the clear to foggy transition, it is prone to catastrophic forgetting in situations where the sequence becomes longer, and the test domain changes more diversely, as seen in the left section. Our strategy for determining when model adaptation is necessary is particularly effective in SHIFT. It improves FPS by about 9, reaching about 20 FPS, while enhancing mAP. This is likely due to avoiding overfitting that can occur when adapting to all repetitive frames in SHIFT, which consists of continuous frames, leading to improvements in both inference speed and adaptation performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.268, + 0.803 + ], + "angle": 0, + "content": "4.5. Additional Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.856 + ], + "angle": 0, + "content": "We aim to demonstrate the effectiveness and detailed analysis of our proposed model in terms of 1) which parts of, 2) how, and 3) when the model should be updated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Which part to update? Tab. 3 shows how updating different parts of the backbone model affects the performance and the memory usage during continual test-time adapta" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.349, + 0.895, + 0.405 + ], + "angle": 0, + "content": "Table 4. Ablation on each component of our loss. SHIFT-D / C denotes SHIFT-Discrete / Continuous, respectively. The left and right value in each cell corresponds to the mAP for the Swin-T and ResNet50 backbone, respectively." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.407, + 0.88, + 0.484 + ], + "angle": 0, + "content": "
LimgLobjCOCOSHIFT-D.SHIFT-C.
--17.7/ 17.338.5/ 35.919.6/ 13.8
✓-16.7/ 18.136.6/ 37.019.1/ 16.0
✓no class weight17.8/ 18.939.7/ 38.025.1/ 23.4
✓class weight wk,t22.6/ 22.240.4/ 38.723.2/ 21.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.672 + ], + "angle": 0, + "content": "tion. We compare (1) updating full parameters, (2) affine parameters of the normalization layer, and (3) our proposed adaptor for each backbone on the SHIFT dataset. Although our adaptor has fewer parameters, about \\(0.9\\%\\) or less of the full parameters, it demonstrates the best adaptation performance. Updating only the affine parameters of the normalization layer, while having fewer parameters, seems less effective for adaptation in object detection compared to classification [32, 45]. Additionally, our adaptor requires only about \\(60\\%\\) of the memory compared to updating the full parameters, making it memory-efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation study on each component in our loss. Tab. 4 presents the effects of image-level feature alignment, \\( L_{img} \\), object-level feature class-wise alignment \\( L_{obj} \\), and class frequency weighting \\( w^{k,t} \\) proposed to address class imbalance. Aligning only the image-level feature distribution with \\( L_{img} \\) (first row) leads to modest adaptation in the ResNet50 backbone, while performance in the SwinT backbone is even lower than without adaptation. Notably, aligning object-level features with \\( L_{obj} \\) leads to a substantial improvement, with the mAP increasing by approximately \\( 10\\% \\) compared to the no-adaptation scenario. Introducing class-specific frequency-based weighting \\( w^{k,t} \\), despite a slight performance decrease in the SHIFT-Continuous setting, proves highly effective, particularly in scenarios with significant class imbalance, such as COCO" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.277, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.226, + 0.262, + 0.237 + ], + "angle": 0, + "content": "(a) Swin Transformer backbone" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.473, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.226, + 0.434, + 0.237 + ], + "angle": 0, + "content": "(b) ResNet50 backbone" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.334 + ], + "angle": 0, + "content": "Figure 4. Comparison of mAP and FPS from Ours-Skip with varying values of \\(\\tau_{1}(\\diamond)\\) and \\(\\tau_{2}(\\triangle)\\) against Evenly-Skip \\((\\times)\\), adapting every \\(N\\)-th instances, on COCO \\(\\rightarrow\\) COCO-C using both (a) Swin-T and (b) ResNet50. The upward and rightward movement indicates a better strategy with higher mAP and faster inference speed, showing that Ours-Skip is consistently better than Evenly-Skip." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.337, + 0.466, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.434, + 0.399, + 0.446 + ], + "angle": 0, + "content": "(a) Accumulated number of backward steps" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.447, + 0.468, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.537, + 0.47, + 0.55 + ], + "angle": 0, + "content": "(b) Number of backward steps and mAP of Direct-Test in each domain" + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.559, + 0.422, + 0.573 + ], + "angle": 0, + "content": "Figure 5. Analysis of the adaptation of Ours-Skip." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.902 + ], + "angle": 0, + "content": "with 80 classes, where it enhances the mAP by around \\(5\\%\\) p. Trade-off between adaptation performance and efficiency according to different skipping strategies. Fig. 4 presents mAP and FPS depending on the values of \\(\\tau_{1}\\) and \\(\\tau_{2}\\) in the Sec. 3.4 on \\(\\mathrm{COCO} \\rightarrow \\mathrm{COCO - C}\\), which are used for two criteria to determine when the adaptation is needed. We also show the simple baseline Evenly-Skip, which adapts every \\(N\\)-th step and skips the rest. In Fig. 4, the blue lines (\\(\\triangle\\)) show the results when \\(\\tau_{1}\\) is changing from 1.0 to infinity, where only criterion 2 is used, while \\(\\tau_{2}\\) is fixed at 1.05. As \\(\\tau_{1}\\) decreases, more adaptation is required, leading to slower FPS but higher mAP. The green lines (\\(\\diamond\\)) show the results of changing \\(\\tau_{2}\\), where ‘\\(\\tau_{2} = \\inf\\)' denotes using only criterion 1, without criterion 2. For all main experiments, we set \\(\\tau_{1}\\) and \\(\\tau_{2}\\) as 1.1 and 1.05, respectively, considering the balance between mAP and FPS. Additionally, our skipping strategy consistently outperforms Evenly-Skip, achieving higher values in both mAP and FPS. This indicates that our criterion for deciding when to bypass model updates provides an effective balance between accuracy and speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.683 + ], + "angle": 0, + "content": "When do models actually update? We analyze when the model actually skips adaptation and only performs inference or actively utilizes test samples for model adaptation based on the two criteria we propose. This analysis is conducted in COCO to COCO-C with 15 corruption domains and 1 original domain. Fig. 5a plots the number of backward passes, i.e., the number of batches of test samples used for adaptation, with different values of \\(\\tau_{1}\\) for the two backbones. The horizontal and vertical axes represent sequentially incoming test domains and the cumulative backward numbers, respectively. A steep slope in a region indicates frequent adaptation, while a gentle slope indicates skipping adaptation, performing only inference. Notably, even without explicit information about when the test domain changes, the model actively performs adaptation, especially right after the test domain changes. This trend is consistent regardless of changes in \\(\\tau\\) value or backbone type. The number of backward passes largely depends on the value of \\(\\tau_{1}\\) rather than the backbone type, suggesting that a consistent \\(\\tau_{1}\\) value can be used irrespective of the backbone. Fig. 5b illustrates adaptation patterns by dividing backward steps for each domain in the case of Swin-T backbone with \\(\\tau_{1} = 1.1\\). More clearly, it shows that adaptation occurs actively around the points where each domain changes, and afterward, adaptation happens intermittently or almost not at all. The light pink bars represent the performance of Direct-Test, showing that domains with initially high model performance tend to have less adaptation, while domains with lower performance initially need more adaptation. In other words, the amount of skipping adaptation is proportional to the amount of the domain shift. Interestingly, the second domain, 'Shot Noise', shows almost no adaptation despite the lower performance of the Direct-Test. We conjecture that the preceding domain, 'Gaussian Noise', shares a similar nature of noise, leading the model to decide that additional adaptation steps may not be necessary. As a result, our skipping strategy enables the model to efficiently adapt, considering both the original domain the model is trained on and the previous domain the model has been adapted to." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.62, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.84 + ], + "angle": 0, + "content": "We introduce an efficient Continual Test-time Adaptation (CTA) method for object detection in the continually changing domain. Our approach involves 1) lightweight adaptors, 2) class-wise object-level feature alignment, and 3) skipping unnecessary adaptation, offering highly efficient and effective solution to diverse domain shifts. It significantly improve mAP performance in various CTA scenarios without serious slowdown in the inference speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by NRF grant (2021R1A2C3006659) and IITP grants (2021-0-01343, 2022-0-00320, 2019-0-00079), all funded by the Korea government (MSIT)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23361" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Alexander Bartler, Florian Bender, Felix Wiewel, and Bin Yang. Ttaps: Test-time adaption by aligning prototypes using self-supervision. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.472, + 0.354 + ], + "angle": 0, + "content": "[4] Dhanajit Brahma and Piyush Rai. A probabilistic framework for lifelong test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3582-3591, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[6] Yijin Chen, Xun Xu, Yongyi Su, and Kui Jia. Stfar: Improving object detection robustness at test-time by self-training with feature alignment regularization. arXiv preprint arXiv:2303.17937, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[8] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, et al. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[10] Shirsendu Sukanta Halder, Jean-François Lalonde, and Raoul de Charette. Physics-based rendering for improving robustness to rain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10203-10212, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[12] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Junyuan Hong, Lingjuan Lyu, Jiayu Zhou, and Michael Spranger. Mecta: Memory-economic continual test-time model adaptation. In The Eleventh International Conference on Learning Representations, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for modelagnostic domain generalization. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[17] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Sanghun Jung, Jungsoo Lee, Nanhee Kim, Amirreza Shaban, Byron Boots, and Jaegul Choo. Cafa: Class-aware feature alignment for test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[19] Mehran Khodabandeh, Arash Vahdat, Mani Ranjbar, and William G Macready. A robust learning approach to domain adaptive object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 480-490, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[20] Donghyun Kim, Kaihong Wang, Stan Sclaroff, and Kate Saenko. A broad study of pre-training for domain generalization and adaptation. In European Conference on Computer Vision, pages 621-638. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[21] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6092–6101, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8474-8481, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[23] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[24] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test-time adaptation, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23362" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[29] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.371 + ], + "angle": 0, + "content": "[30] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[31] Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin, Mateusz Kozinski, Horst Possegger, and Horst Bischof. Actmad: Activation matching to align distributions for test-time-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24152-24161, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[33] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[34] Mario obler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[36] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[38] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2759-2769, 2023. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[40] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. Advances in Neural Information Processing Systems, 35:17543-17555, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.893, + 0.316 + ], + "angle": 0, + "content": "[41] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.893, + 0.444 + ], + "angle": 0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.893, + 0.5 + ], + "angle": 0, + "content": "[44] Vibashan VS, Poojan Oza, and Vishal M Patel. Towards online domain adaptive object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 478-488, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.501, + 0.893, + 0.556 + ], + "angle": 0, + "content": "[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.894, + 0.613 + ], + "angle": 0, + "content": "[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.893, + 0.657 + ], + "angle": 0, + "content": "[47] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees GM Snoek. Energy-based test sample adaptation for domain generalization. arXiv preprint arXiv:2302.11215, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.893, + 0.727 + ], + "angle": 0, + "content": "[48] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.893, + 0.782 + ], + "angle": 0, + "content": "[49] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.785, + 0.893, + 0.827 + ], + "angle": 0, + "content": "[50] Bowen Zhao, Chen Chen, and Shu-Tao Xia1. Delta: Degradation-free fully test-time adaptation. In International Conference on Learning Representations (ICLR), 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23363" + } + ] +] \ No newline at end of file diff --git a/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_origin.pdf b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..34be974d64a10fbd50d7bad35fe549523a560a08 --- /dev/null +++ b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/d778613f-28c0-4d7a-80d6-b7c15f243643_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0563b3732f38cf520cefcf6a722bcf3f850b0cad6e06721dc5dba0428b0a6a21 +size 4144132 diff --git a/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/full.md b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4d4a9551c3a9b1e7b0b93df26b4f839c9fd72c35 --- /dev/null +++ b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/full.md @@ -0,0 +1,271 @@ +# What, How, and When Should Object Detectors Update in Continually Changing Test Domains? + +Jayeon Yoo $^{1}$ Dongkwan Lee $^{1}$ Inseop Chung $^{1}$ Donghyun Kim $^{2*}$ Nojun Kwak $^{1*}$ Seoul National University Korea University + +1{jayeon.yoo, biancco, jis3613, nojunk}@snu.ac.kr, $^2\mathrm{d}$ kim@korea.ac.kr + +# Abstract + +It is a well-known fact that the performance of deep learning models deteriorates when they encounter a distribution shift at test time. Test-time adaptation (TTA) algorithms have been proposed to adapt the model online while inferring test data. However, existing research predominantly focuses on classification tasks through the optimization of batch normalization layers or classification heads, but this approach limits its applicability to various model architectures like Transformers and makes it challenging to apply to other tasks, such as object detection. In this paper, we propose a novel online adaption approach for object detection in continually changing test domains, considering which part of the model to update, how to update it, and when to perform the update. By introducing architecture-agnostic and lightweight adaptor modules and only updating these while leaving the pre-trained backbone unchanged, we can rapidly adapt to new test domains in an efficient way and prevent catastrophic forgetting. Furthermore, we present a practical and straightforward class-wise feature aligning method for object detection to resolve domain shifts. Additionally, we enhance efficiency by determining when the model is sufficiently adapted or when additional adaptation is needed due to changes in the test distribution. Our approach surpasses baselines on widely used benchmarks, achieving improvements of up to $4.9\% p$ and $7.9\% p$ in mAP for COCO $\rightarrow$ COCO-corrupted and SHIFT, respectively, while maintaining about 20 FPS or higher. The implementation code is available at https://github.com/natureyoo/ContinualTTA_ObjectDetection. + +# 1. Introduction + +Deep learning models excel in many vision tasks but struggle with domain shifts when test data distributions diverge from training data [3, 27, 42]. These shifts, common in real-world settings, arise from natural variations, weather + +![](images/9dfbac57f05601a9cee12ea9481b3c07e19a607e1c687920a9668d2a48e2e0ad.jpg) +Figure 1. We propose an online adaptation method for object detection in continually changing test domains. Object detectors trained with clean images suffer from performance degradation due to various corruption, such as camera sensor degradation or environmental changes (Direct-Test). Updating full parameters for online adaptation require a large number of test samples and vulnerable to drastic domain changes (Full-Finetuning), while using only our lightweight adaptor is robust and quickly adapts within a few time steps (Ours). We can further improve efficiency by skipping unnecessary adaptation steps (Ours-Skip). + +changes (e.g., fog, rain), camera differences (e.g., pixelate, defocus blur), and various other factors. Test-Time Adaptation (TTA) [3, 27, 32, 42, 45, 49] addresses these shifts by adapting models to the target (test) distribution online. Considering the dynamic nature of real-world applications, it becomes crucial to extend this approach to Continual Test-Time Adaptation (CTA), which accounts for the ongoing evolution of test distributions. This is particularly vital in environments like autonomous driving, where conditions can change continuously (e.g., from clear to rainy, or day to night). Despite its significance, continual test-time adaptation for object detection remains underexplored. + +Recently, several TTA methods [6, 31, 38] tailored for object detection have been proposed. ActMAD [31] aligns all the output feature maps $(\mathbb{R}^{C\times H\times W})$ after Batch Nor + +malization (BN) layers [15] to adapt the test domain to be similar to that of the training domain. However, this approach requires significant memory during adaptation and does not explicitly consider the objects present in the image. TeST [38] and STFAR [6] adapt to a test domain by utilizing weak and strong augmented test samples with a teacher-student network [39], but they significantly increase inference costs since they require additional forward passes and update steps. Also, these methods update all network parameters, making them highly inefficient in online adaptation and vulnerable to losing task-specific knowledge when the test domain experiences continual or drastic changes. + +In this paper, we aim to develop an efficient continual test-time adaptation (CTA) method for object detection. We investigate the following three key aspects to improve efficiency; what to update: while previous TTA methods for object detection [6, 31, 38] use full fine-tuning, updating all parameters at test time, they are inefficient and prone to losing task-specific knowledge in relatively complex object detection tasks. Updating BN layers, as done in many TTA methods for classification [18, 32, 45, 49], is not as effective for object detection, given its smaller batch size compared to classification and the limitation in applying various backbones, such as Transformer [28, 43]. how to update: several previous TTA methods for object detection [6, 38] adapt the model by using teacher-student networks, resulting in a significant decrease in inference speed, which is detrimental during test time. While another existing method [31] aligns feature distributions for adaptation, it does not consider each object individually, focusing only on image features, making it less effective for object detection. when to update: most TTA or CTA methods update models using all incoming test samples. However, it is inefficient to update continuously the model if it is already sufficiently adapted when the change of the test domain is not significant. + +To this end, (1) we propose an efficient continual test-time adaptation method for object detectors to adapt to continually changing test domains through the use of lightweight adaptors which require only $0.54\% \sim 0.89\%$ additional parameters compared to the full model. It exhibits efficiency in parameters, memory usage, and adaptation time, along with robustness to continuous domain shifts without catastrophic forgetting. Additionally, it demonstrates wide applicability to various backbone types compared to BN-based TTA methods [18, 24, 32, 45, 49, 50]. (2) To enhance the adaptation effectiveness in the object detection task, we align the feature distribution of the test domain with that of the training domain at both the image-level and object-level using only the mean and variance of features. For estimating the mean of the test domain features, we employ Exponentially Moving Average (EMA) as we can leverage only the current incoming test samples, not the entire test domain data. Due to the unavailability of training + +data access, we utilize only the mean and variance of the features from a few training samples. (3) We also introduce two novel criteria that do not require additional resources to determine when the model needs adaptation to enhance efficiency in a continually changing test domain environment. As illustrated in Fig. 1, our approach Ours, employing adaptors, tends to adapt much faster to domain changes compared to full parameter updates. This enables efficient TTA by using only a few test samples to update the adaptor and skipping the rest of the updates as shown in Ours-Skip. Our main contributions are summarized as follows: + +- We introduce an architecture-agnostic lightweight adaptor, constituting only a maximum of $0.89\%$ of the total model parameters, into the backbone of the object detector to adapt the model in a continually changing test domain. This approach ensures efficiency in parameters, memory usage, and adaptation speed, demonstrating the robust preservation of task-specific knowledge owing to its inherent structural characteristics. +- We propose a straightforward and effective adaptation loss for CTA in object detection tasks. This is achieved by aligning the distribution of training and test domain features at both the image and object levels, utilizing only the mean and variance of a few training samples and EMA-updated mean features of the test domain. +- We also propose two criteria to determine when the model requires adaptation, enabling dynamic skipping or resuming adaptation as needed. This enhancement significantly boosts inference speed by up to about 2 times while maintaining adaptation performance. +- Our adaptation method proves effective for diverse types of domain shifts, including weather changes and sensor variations, regardless of whether the domain shift is drastic or continuous. In particular, our approach consistently improves the mAP by up to $7.9\%$ in $\mathrm{COCO} \rightarrow \mathrm{COCO-C}$ and SHIFT-Discrete/Continuous with higher than 20 FPS. + +# 2. Related Work + +Test-time adaptation. Recently, there has been a surge of interest in research that adapts models online using unlabeled test samples while simultaneously inferring the test sample to address the domain shift problem, where the test data distribution differs from that of the training data. There are two lines for online adaptation to the test domain, Test-time Training (TTT) and Test-time Adaptation (TTA). TTT [1, 2, 27, 42] involves modifying the model architecture during training to train it with self-supervised loss, allowing adaptation to the test domain in the test time by applying this self-supervised loss to the unlabeled test samples. On the other hand, TTA aims to adapt the trained model directly to the test domain without specifically tailored model architectures or losses during training time. NORM [37] and DUA [30] address the domain shifts by + +adjusting the statistics of batch normalization (BN) layers using the current test samples, without updating other parameters, inspired by [23]. Following this, [24, 32, 45, 50] and [18] update the affine parameters of BN layers using unsupervised loss, entropy minimization loss to enhance the confidence of test data predictions, and feature distribution alignments loss, respectively. Several studies [16, 17] update the classifier head using the pseudo-prototypes from the test domain. However, these methods limit their applicability to architectures without BN layers or to object detection tasks that involve multiple objects in a single image. Others [31, 40, 49] update full parameters for online adaptation to the test domain in an online manner, but this approach is inefficient and susceptible to the noisy signal from the unsupervised loss. While existing TTA methods are oriented towards classification tasks, we aim to propose an effective and efficient method for online adaptation in the object detection task. + +Continual test-time adaptation. Recent studies [33, 46] point out that existing TTA methods have primarily focused on adapting to test domains following an i.i.d assumption and may not perform well when the test data distribution deviates from this assumption. [46] introduces a ContinuaTTA (CTA) method designed for scenarios where the test domain continuously changes over time. This poses challenges in preventing the model from over-adapting to a particular domain shift and preserving the knowledge of the pre-trained model to avoid catastrophic forgetting. In the field of CTA, the self-training strategy adopting an Exponentially Moving Average (EMA) teacher-student structure is attracting interest as an effective algorithm enabling robust representation to be learned through self-knowledge distillation. In many studies, the EMA teacher-student structure and catastrophic restoration of source model weights have been proposed as a solution to achieve the goal of CTA [4, 46, 47]. Approaches using source replay [34], and anti-forgetting regularization [32] have also achieved good performances in robust continuous adaptation. Furthermore, there is growing attention on methods that mitigate the computational and memory challenges associated with CTA, such as [13], which relies on updates to batch normalization statistics. + +Test-time adaptive object detection. Research on TTA for Object Detection (TTAOD) is progressively emerging [6, 31, 38, 44]. Most existing TTAOD methods [6, 38, 44] exploit a teacher-student network to adapt to the test domain, following the self-training approach commonly employed in Unsupervised Domain Adaptation for object detection [7, 19, 21, 36]. However, it is inefficient for TTA due to data augmentation requirements and additional forward and backward steps, resulting in slower inference speeds and higher memory usage. Another approach, ActMAD [31], aligns the distributions of output feature maps + +after all BN layers along the height, width, and channel axes to adapt to the test domain. However, this location-aware feature alignment is limited to datasets with fixed location priors, such as driving datasets, and is less effective for natural images like COCO. Additionally, CTA for Object Detection (CTAOD) have not been thoroughly explored. Therefore, there is a need for an effective CTAOD method considering memory and time efficiency. + +# 3. Method + +To enable the efficient and effective Continual Test-time Adaptation of Object Detectors (CTAOD), we introduce an approach that specifies which part of the model should be updated, describes how to update those using unlabeled test data, and determines whether we perform model updates or not to improve efficiency. + +# 3.1. Preliminary + +Assume that we have an object detector $h \circ g_{\Theta}$ , here $h$ and $g$ are the RoI head and the backbone, respectively with their parameters being $\Theta$ . The training dataset is denoted as $D_{train} = \{(x_i, y_i)\}_{i=1}^N$ , where $x_i \sim P_{train}(x)$ and $y_i = (bbox_i, c_i)$ , containing information on the bounding box (bbox) and class label $c_i \in \mathcal{C}$ . Consider deploying the detector to the test environments where the test data at period $T$ is denoted as $x_j^T \sim P_{test}^T(x)$ , $P_{test}^T \neq P_{train}$ and $P_{test}^T$ deviates from the i.i.d. assumption. In addition, the domain of $P_{test}^T$ continually changes according to $T$ (i.e., $P_{test}^T \neq P_{test}^{T-1}$ ). Our goal is to adapt the detector $h \circ g$ to $P_{test}^T$ using only test data $x_j^T$ while making predictions. + +# 3.2. What to update: Adaptation via an adaptor + +Previous methods [6, 31, 38, 44] adapt the model to the test domain by updating all parameters $\Theta$ , leading to inefficiency at test time and a high risk of losing task knowledge from the training data. In contrast, we adapt the model by introducing an adaptor with an extremely small set of parameters and updating only this module while freezing $\Theta$ . We introduce a shallow adaptor in parallel for each block, inspired by [5, 14], where transformer-based models are fine-tuned for downstream tasks through parameter-efficient adaptors, as shown in Fig. 2. Each adaptor consists of down-projection layers $\mathcal{W}_{down} \in \mathbb{R}^{d \times \frac{d}{r}}$ , up-projection layers $\mathcal{W}_{up} \in \mathbb{R}^{\frac{d}{r} \times d}$ and RLUs, where $d$ denotes the input channel dimension and $r$ is the channel reduction ratio set to 32 for all adaptors. We use MLP layers for the Transformer block (Fig. 2a) and $1 \times 1$ convolutional layers for the ResNet block (Fig. 2b) to introduce architecture-agnostic adaptors. The up-projection layer is initialized to 0 values so that the adaptor does not modify the output of the block, but as the adaptor is gradually updated, it adjusts the output of the block to adapt to the test domain. Even as the adaptor + +![](images/6699e81c9c72d975f0d533edc09ae91b2b4adc92fff6ef6d71389a27b98f553e.jpg) +Figure 2. We attach an adaptor, which is a shallow and low-rank MLP or CNN, to every $N$ block in parallel. We update only these adaptors while other parameters are frozen. Our approach can be applied to diverse architectures including CNNs and Transformers. + +![](images/d40db235519bc4ba5f5f700896391ad47202e140b5235f9e0ebdee3490839c18.jpg) + +is updated in the test domain, the original backbone parameter $\Theta$ remains frozen and fully preserved. This structural preservation, as evident in Ours in Fig. 1, enables robust and efficient adaptation to domain changes by maintaining relatively complex task knowledge in object detection and updating very few parameters. + +# 3.3. How to update: EMA feature alignment + +To adapt the object detector to the test domain, we align the feature distribution of the test domain with that of the training data, inspired by [18, 31, 40]. In contrast to these methods that solely align image feature distribution, we additionally align object-level features in a class-wise manner, considering class frequency, to enhance its effectiveness for object detection. As the training data is not accessible during test time, we pre-compute the first and second-order statistics, denoted as $\mu_{tr} = \mathbb{E}[F_{tr}]$ and $\Sigma_{tr} = \mathbb{Var}[F_{tr}]$ , where the operators $\mathbb{E}$ and $\mathbb{Var}$ represent the mean and variance respectively. The features $F_{tr} = \{g_{\Theta}(x_{tr})\}$ are computed using only 2,000 training samples, a small subset of the training data. Since a sufficient amount of test domain data is not available at once, and only the current incoming test data, whose features are denoted as $F_{te}^{t}$ , is accessible at time step $t$ , we estimate the mean of test data features using an exponentially moving average (EMA) as follows: + +$$ +\mu_ {t e} ^ {t} = (1 - \alpha) \cdot \mu_ {t e} ^ {t - 1} + \alpha \cdot \mathbb {E} [ F _ {t e} ^ {t} ], \quad \mathrm {s . t .} \mu_ {t e} ^ {0} = \mu_ {t r}. (1) +$$ + +Considering the typically small batch size in object detection compared to classification, we approximate the variance of the test features as $\Sigma_{te} \simeq \Sigma_{tr}$ to reduce instability. + +Image-level feature alignment. We estimate the training and test feature distributions as normal distributions and minimize the KL divergence between them as follows: + +$$ +L _ {i m g} = \mathrm {D} _ {K L} (\mathcal {N} (\mu_ {t r}, \Sigma_ {t r}), \mathcal {N} (\mu_ {t e} ^ {t}, \Sigma_ {t r})). \qquad (2) +$$ + +Region-level class-wise feature alignment. In object detection, we deal with multiple objects within a single image, making it challenging to apply the class-wise feature alignment proposed in [40], a TTA method for classification. To handle region-level features that correspond to an object, we use ground truth bounding boxes for the training data and utilize the class predictions of RoI pooled features, $f_{te}^{t}$ , for unlabeled test data. In object detection, domain shifts often result in lower recall rates, as a significant number of proposals are predicted as background [22]. To mitigate this issue, we filter out features with background scores exceeding a specific threshold. Subsequently, we assign them to the foreground class with the highest probability, as follows: + +$$ +F _ {t e} ^ {k, t} = \left\{f _ {t e} ^ {t} \mid \underset {c} {\operatorname {a r g m a x}} p _ {f g} = k, p _ {b g} < 0. 5 \right\}, \tag {3} +$$ + +$$ +w h e r e h _ {c l s} \left(f _ {t e} ^ {t}\right) = \left[ p _ {f g}, p _ {b g} \right] = \left[ p _ {0}, \dots , p _ {C - 1}, p _ {b g} \right]. +$$ + +We estimate the class-wise feature distribution of the test domain by exploiting $F_{te}^{k,t}$ and Eq.1. Furthermore, we introduce a weighting scheme for aligning features of less frequently appearing classes, taking into account the severe class imbalance where specific instance (e.g., person) may appear multiple times within a single image, as follows: + +$$ +N ^ {k, t} = N ^ {k, t - 1} + \left\| F _ {t e} ^ {k, t} \right\|, \quad \text {s . t .} \quad N ^ {k, 0} = 0 +$$ + +$$ +w ^ {k, t} = \log \left(\frac {\operatorname* {m a x} _ {i} N ^ {i , t}}{N ^ {k , t}}\right) + 0. 0 1 \tag {4} +$$ + +$$ +L _ {o b j} = \sum_ {k} w ^ {k, t} \cdot D _ {K L} (\mathcal {N} (\mu_ {t r} ^ {k}, \Sigma_ {t r} ^ {k}), \mathcal {N} (\mu_ {t e} ^ {k, t}, \Sigma_ {t r} ^ {k})). +$$ + +Here, the class-wise mean $\mu^k$ and variance $\Sigma^k$ of the training and test data are obtained in the same way as the image-level features. We can effectively adapt the object detector by updating the model to align the feature distribution at both the image and object levels as $L = L_{img} + L_{obj}$ . + +# 3.4. When to update: Adaptation on demand + +As shown in Fig. 1, Ours, which only updates the adaptor proposed in Sec. 3.2, efficiently adapts to changes in the test domain, even with a small subset of early test samples. We leverage its rapid adaptation characteristics to reduce computational costs by skipping model updates (i.e., skipping backward passes) when the model has already sufficiently adapted to the current test domain and resuming model updates when confronted with a new test domain. Therefore, we introduce two criteria to determine when to update the model or not as follows: + +(Criterion 1) When the distribution gap exceeds the in-domain distribution gap. Recall that $L_{img}$ (Eq. 2) measures the distribution gap between the test and train distributions. We assume a model is well-adapted to the current test domain when $L_{img}$ is closer to the in-domain distribution gap. We measure the in-domain distribution gap by + +![](images/5fdcd17f8b52f5a481a48fea362a9847ce21371e8adb0922ab70cac0fc1f5447.jpg) + +![](images/ec62c3ff5daaf694be509d578aebb28b879890ce9a7e77535a31467b488a2301.jpg) +(a) The ratio of $L_{img}$ to $D_{KL}^{in}$ +(b) The ratio of $L_{img}$ to $L_{ema}^{t}$ +Figure 3. The test domain undergoes a shift every 4,000 time steps, and each metric reaches its peak at the same intervals. + +sampling two disjoint subsets, $x_{i}$ and $x_{j}$ , of training features $F_{tr}$ from Sec. 3.3 as follows: + +$$ +D _ {K L} ^ {i n} = D _ {K L} \left(\mathcal {N} \left(\mu_ {t r} ^ {i}, \Sigma_ {t r} ^ {i}\right), \mathcal {N} \left(\mu_ {t r} ^ {j}, \Sigma_ {t r} ^ {j}\right)\right), \tag {5} +$$ + +where $\mu_{tr}^{i},\Sigma_{tr}^{i}$ are obtained from $x_{i}\sim P_{train}(x)$ and $\mu_{tr}^{j},\Sigma_{tr}^{j}$ from $x_{j}\sim P_{train}(x)$ . In other words, if $L_{img}$ is noticeably higher than the in-domain distribution gap $D_{KL}^{in}$ , we consider a model encountering a test domain whose distribution differs from $P_{train}(x)$ and needs to be updated. Based on this, we introduce a new index $\frac{L_{img}}{D_{KL}^{in}}$ . Fig. 3a plots the trend of this index during the model adaptation to a continually changing test domain. It shows that the index has a large value in the early stages of a domain change, decreases rapidly, and then maintains a value close to 1. This index exhibits a similar trend regardless of the backbone type and dataset, as included in the appendix. Therefore, we establish the criterion that model updates are necessary when this index exceeds a certain threshold, $\tau_{1}$ , as $\frac{L_{img}}{D_{KL}^{in}} >\tau_{1}$ . + +(Criterion 2) When the distribution gap suddenly increases. Additionally, we can determine when the test distribution changes and model updates are necessary by observing the trend of the distribution gap (i.e., $L_{img}$ ). The convergence of $L_{img}$ indicates that a model is well-adapted to the current test domain. To put it differently, $L_{img}$ will exhibit a sudden increase when the model encounters a new test domain. We introduce an additional index, denoted as $\frac{L_{img}}{L_{ema}^t}$ , representing the ratio of the current $L_{img}$ to its exponentially moving average $L_{ema}^t$ at time $t$ . We calculate it using the following formula: $L_{ema}^t = 0.99 \cdot L_{ema}^{t-1} + 0.01 \cdot L_{img}$ . Fig. 3b illustrates the trend of the ratio of $L_{img}$ over the timesteps. It tends to reach a value of 1 as the loss stabilizes at a specific level. Nevertheless, when the model encounters shifts in the test distribution, the ratio experiences a sharp increase, indicating the necessity of a model update when it exceeds a specific threshold, $\tau_2$ , as $\frac{L_{img}}{L_{ema}^t} > \tau_2$ . + +If at least one of the two criteria is satisfied, we conclude that the model requires adaptation and proceed to update it. + +# 4. Experiments + +Sec. 4.1 presents the two object detection benchmark datasets with test distributions that change continuously, either in a drastic or gradual manner, and our implementation detail is in 4.2. Sec. 4.4 compares our method with other TTA baselines described in Secs. 4.3.. We present detailed ablation studies of our method analyzing the effectiveness and efficiency of our method in terms of what, how, and when to update the models for CTAOD in Sec. 4.5. + +# 4.1. Datasets + +We experiment with the following three scenarios. + +$\mathrm{COCO} \rightarrow \mathrm{COCO-C}$ simulates continuous and drastic realistic test domain changes over a long sequence. MS-COCO [25] collects 80 classes of common objects in their natural context with 118k training images and 5k validation images. COCO-C is created by employing 15 types of realistic corruptions [29], such as image distortion and various weather conditions, to simulate test domain changes. In the experiments, the model is only trained on the COCO train set and sequentially evaluated on each corruption in the COCO-C validation set during test-time for reproducing continually changing test domains. Finally, the model is evaluated on the original COCO validation set to assess how well it preserves knowledge of the original domain (denoted as Org.). SHIFT-(Discrete / Continuous) [41] is a synthetic driving image dataset with 6 classes under different conditions using five weather attributes (clear, cloudy, overcast, fog, rain) and three time-of-day attributes (daytime, dawn, night). In SHIFT-Discrete, there are image sets for each attribute, and the model is sequentially evaluated on these attributes, cloudy $\rightarrow$ overcast $\rightarrow$ foggy $\rightarrow$ rainy $\rightarrow$ dawn $\rightarrow$ night $\rightarrow$ clear which contains 2.4k, 1.6k, 2.7k, 3.2k, 1.2k, 1.4k, and 2.8k validation images, respectively. This simulates scenarios where the domain undergoes drastic changes. In SHIFT-Continuous, the model is evaluated on four sequences, each consisting of 4k frames, continuously transitioning from clear to foggy (or rainy) and back to clear. + +# 4.2. Implementation Detail + +We conduct experiments on Faster-RCNN [35] with FPN [26], using ImageNet1k pre-trained ResNet50 [11] and ImageNet21k pre-trained Swin-Tiny [28] backbones. Specifically, Swin-Tiny is employed for its superior domain transfer [20] and detection efficacy over other advanced backbones [9]. For COCO $\rightarrow$ COCO-C with ResNet50, we use a publicly available model trained with strong augmentation [48]. Other models are trained using detectron2 framework following [35] and [28]. For test-time adaptation, we always set the learning to 0.001 for the SGD opti + +Table 1. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on COCO $\rightarrow$ COCO-C. Our model consistently outperforms baselines on the two different backbones. Furthermore, Ours-Skip with ResNet notably reduces backward passes by as much as $90.5\%$ , leading to a significantly improved frames per second (FPS) rate by up to $109.9\%$ . + +
BackboneMethodNoiseBlurWeatherDigitalOrg.Avg.# step
GauShtImpDefGlsMtnZmSnwFrsFogBrtCntElsPxJpgFor.Back.FPS
Swin-T [28]Direct-Test9.711.410.013.47.512.15.220.724.836.136.012.919.14.915.843.017.780K021.5
ActMAD10.712.09.412.35.79.54.515.317.527.628.21.116.72.68.736.313.980K80K8.3
Mean-Teacher10.012.111.212.88.112.14.919.623.734.934.08.018.96.117.641.017.2160K80K6.9
Ours13.616.616.114.013.614.28.323.727.237.436.427.227.222.222.342.322.680K80K9.5
Ours-Skip13.315.315.114.012.813.96.522.025.435.534.926.525.923.420.241.221.680K9.7K17.7
ResNet50 [11]Direct-Test9.111.09.812.64.58.84.619.123.138.438.021.415.65.311.944.217.380K025.8
NORM9.911.911.012.65.29.15.119.423.538.237.622.417.25.710.343.417.580K025.8
DUA9.811.710.812.85.28.95.119.323.738.437.822.317.25.410.144.117.180K025.8
ActMAD9.19.67.011.03.26.13.312.814.027.727.83.912.92.37.234.310.580K80K9.6
Mean-Teacher9.612.512.04.02.94.83.116.223.535.134.021.816.68.212.740.314.5160K80K8.1
Ours12.717.817.512.411.511.36.622.826.938.638.528.025.121.222.241.822.280K80K10.1
Ours-Skip14.417.116.013.911.712.26.322.125.537.737.125.524.123.121.142.821.980K7.6K21.2
+ +mizer, and $\alpha$ of Eq. 1 to 0.01, while $\tau_{1}$ and $\tau_{2}$ are set to 1.1 and 1.05, respectively. We use the same hyper-parameters across all backbones and datasets, with a batch size of 4. + +# 4.3. Baselines + +Direct-Test evaluates the model trained in the training domain without adaptation to the test domain. ActMAD [31] is a TTA method aligning the distribution of output features across all BN layers. To apply ActMAD to the Swin Transformer-based model, we align the output features of the LN layers. We implement Mean-Teacher using a teacher-student network framework to reproduce as close as possible to TeST [38], as its implementation is not publicly available. We follow the FixMatch [39] augmentation method and report results after tuning all hyper-parameters in our scenario. NORM [37] and DUA [30], TTA methods initially designed for classification, are directly applicable to detection tasks by either mixing a certain amount of current batch statistics or updating batch statistics via EMA. However, these are only compatible with architectures containing BN layers. Additional details are provided in Appendix. + +# 4.4. Main Results + +We compare the performance of each method using mAP and efficiency metrics, including the number of forward and backward passes, as well as FPS during test-time adaptation. Results of COCO and SHIFT are in Tab. 1 and 2, respectively. + +$\mathbf{COCO} \rightarrow \mathbf{COCO-C}$ . Tab. 1 demonstrates the effective adaptation performance of Ours in the challenging COCO benchmark with 80 classes due to object-level class-wise feature alignment. ActMAD also aligns feature distribution for TTA, but is not effective since it only aligns whole feature maps without considering specific classes in the image. NORM and DUA, applicable only to ResNet [11], show minimal performance improvement by adaptation as they are not specifically tailored for object detection and only modify batch statistics across the entire feature map. Ad + +ditionally, ActMAD and Mean-Teacher, updating full parameters, gradually lose task knowledge in the continually changing test distributions, resulting in much lower performance on Org., the domain identical to the training data, than that of Direct-Test. In contrast, Ours effectively prevents catastrophic forgetting by freezing the original parameters of the models and updating only the adaptor, obtaining performance on par with Direct-Test on the Org. domain and consistently high performance across corrupted domains, with an average mAP improvement of $4.9\%$ compared to that of Direct-Test. Furthermore, leveraging the rapid adaptation ability of the adaptor, Ours-Skip, which skips unnecessary adaptation, allows using only a maximum of about $12\%$ of the total samples for adaptation without significant performance loss. This leads to a substantial improvement in inference speed, more than doubling compared to other TTA methods, reaching over 17.7 FPS. + +SHIFT-Discrete. Ours is also effective in SHIFT, which simulates continuous changes in weather and time in driving scenarios according to the left section of Tab. 2. Especially, Ours shows significant improvements in mAP by $7 - 9\% \mathrm{p}$ , particularly for the foggy and dawn attributes where Direct-Test obtains lower performance due to severe domain shift. In contrast, with ActMAD, catastrophic forgetting takes place when adapting to the cloudy and overcast weather. This is due to the updating of the full parameters, despite that Direct-Test already shows proficient performance in these conditions. As a result, the performance in the later domains is worse than that of the Direct-Test. DUA, which updates batch statistics using EMA, shows a gradual decrease in performance as the domain continuously changes, resulting in much lower performance in the original clear domain (i.e., clear). On the other hand, NORM, which utilizes the statistics of the current batch samples, exhibits no catastrophic forgetting and relatively good adaptation, as SHIFT is a relatively easier task compared to COCO due to having only 6 classes. Compared to NORM, Ours shows better adaptation performance, and is + +Table 2. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on SHIFT-Discrete and SHIFT-Continuous. Baselines perform effectively in a particular setting but lack generalizability across various settings. Our method consistently achieves results that are either better or on par with the best model in all settings, demonstrating its strong stability. Ours-Skip also effectively reduces the number of backward passes without compromising mAP performance, resulting in a higher FPS. + +
BackboneMethodSHIFT-DiscreteSHIFT-Continuous
mAP# stepmAP# Avg. step
cloudyoverc.fograindawnnightclearAvg.For.Back.FPSclear↔fogclear↔rainFor.Back.FPS
Swin-T [28]Direct-Test50.038.923.145.126.939.545.938.515.3K027.518.121.14K028.3
ActMAD49.838.421.443.119.032.044.835.515.3K15.3K9.315.616.34K4K9.8
Mean-Teacher50.039.225.745.426.037.542.238.030.6K15.3K7.820.424.38K4K6.5
Ours50.339.232.246.730.439.944.340.415.3K15.3K11.223.922.64K4K11.6
Ours-Skip50.339.729.147.130.241.545.940.615.3K6.1K20.025.123.84K0.83K19.2
ResNet50 [11]Direct-Test49.437.919.743.120.135.345.635.915.3K030.112.115.44K030.0
NORM49.738.622.944.725.137.445.537.715.3K030.116.919.44K030.0
DUA45.231.527.731.915.218.621.127.315.3K030.122.522.44K030.0
ActMAD49.237.718.040.616.032.944.334.115.3K15.3K11.312.716.34K4K11.2
Mean-Teacher49.638.426.843.426.633.141.637.130.6K15.3K9.916.020.88K4K9.8
Ours49.738.727.446.327.437.643.838.715.3K15.3K12.920.921.94K4K13.9
Ours-Skip49.738.826.946.227.638.845.039.015.3K8.9K21.520.022.54K0.75K21.3
+ +Table 3. Comparison of adaptation performance (mAP), the number of trainable parameters (# Params), and memory usage (Cache) according to which part of the backbone is updated. SD / SC denotes SHIFT-Discrete/Continuous, respectively. + +
BackboneTrainable ParamsmAP# ParamsCache
SDSCNumRatioAvg.Max
Swin-TFull-params38.420.627.7M100%0.8611.0
LayerNorm38.520.00.03M0.1%0.657.49
adaptor (Ours)40.423.20.15M0.5%0.656.96
ResNet50Full-params37.620.423.7M100%1.659.29
BatchNorm37.920.20.05M0.2%1.479.11
adaptor (Ours)38.721.70.21M0.9%1.485.41
+ +also applicable to BN-layer-free Swin Transformers. + +SHIFT-Continuous. In scenarios where the test domain gradually changes across the entire sequence, Ours also demonstrates effectiveness, improving mAP by up to $7\%$ p, as shown in the right section of Tab. 2. While DUA performs well in the clear to foggy transition, it is prone to catastrophic forgetting in situations where the sequence becomes longer, and the test domain changes more diversely, as seen in the left section. Our strategy for determining when model adaptation is necessary is particularly effective in SHIFT. It improves FPS by about 9, reaching about 20 FPS, while enhancing mAP. This is likely due to avoiding overfitting that can occur when adapting to all repetitive frames in SHIFT, which consists of continuous frames, leading to improvements in both inference speed and adaptation performance. + +# 4.5. Additional Analyses + +We aim to demonstrate the effectiveness and detailed analysis of our proposed model in terms of 1) which parts of, 2) how, and 3) when the model should be updated. + +Which part to update? Tab. 3 shows how updating different parts of the backbone model affects the performance and the memory usage during continual test-time adapta + +Table 4. Ablation on each component of our loss. SHIFT-D / C denotes SHIFT-Discrete / Continuous, respectively. The left and right value in each cell corresponds to the mAP for the Swin-T and ResNet50 backbone, respectively. + +
LimgLobjCOCOSHIFT-D.SHIFT-C.
--17.7/ 17.338.5/ 35.919.6/ 13.8
✓-16.7/ 18.136.6/ 37.019.1/ 16.0
✓no class weight17.8/ 18.939.7/ 38.025.1/ 23.4
✓class weight wk,t22.6/ 22.240.4/ 38.723.2/ 21.7
+ +tion. We compare (1) updating full parameters, (2) affine parameters of the normalization layer, and (3) our proposed adaptor for each backbone on the SHIFT dataset. Although our adaptor has fewer parameters, about $0.9\%$ or less of the full parameters, it demonstrates the best adaptation performance. Updating only the affine parameters of the normalization layer, while having fewer parameters, seems less effective for adaptation in object detection compared to classification [32, 45]. Additionally, our adaptor requires only about $60\%$ of the memory compared to updating the full parameters, making it memory-efficient. + +Ablation study on each component in our loss. Tab. 4 presents the effects of image-level feature alignment, $L_{img}$ , object-level feature class-wise alignment $L_{obj}$ , and class frequency weighting $w^{k,t}$ proposed to address class imbalance. Aligning only the image-level feature distribution with $L_{img}$ (first row) leads to modest adaptation in the ResNet50 backbone, while performance in the SwinT backbone is even lower than without adaptation. Notably, aligning object-level features with $L_{obj}$ leads to a substantial improvement, with the mAP increasing by approximately $10\%$ compared to the no-adaptation scenario. Introducing class-specific frequency-based weighting $w^{k,t}$ , despite a slight performance decrease in the SHIFT-Continuous setting, proves highly effective, particularly in scenarios with significant class imbalance, such as COCO + +![](images/0059f846b713f8181cc7e07ca39ecd830bfa65635687bdd667bc6c446e5734ee.jpg) +(a) Swin Transformer backbone + +![](images/9b060176513247b6cdd997985aa5eb06888b9970d1159a8e387ea328ae0ece65.jpg) +(b) ResNet50 backbone + +![](images/147bb4ac8b27f43f9b253acb0e1e9e798e1089eb73839c7e1be1d64b7283cf2b.jpg) +Figure 4. Comparison of mAP and FPS from Ours-Skip with varying values of $\tau_{1}(\diamond)$ and $\tau_{2}(\triangle)$ against Evenly-Skip $(\times)$ , adapting every $N$ -th instances, on COCO $\rightarrow$ COCO-C using both (a) Swin-T and (b) ResNet50. The upward and rightward movement indicates a better strategy with higher mAP and faster inference speed, showing that Ours-Skip is consistently better than Evenly-Skip. + +![](images/7ece3d0df3f5442d9afaf91a80237d30caa564bd3664eb2cf442291defdacc59.jpg) +(a) Accumulated number of backward steps +(b) Number of backward steps and mAP of Direct-Test in each domain +Figure 5. Analysis of the adaptation of Ours-Skip. + +with 80 classes, where it enhances the mAP by around $5\%$ p. Trade-off between adaptation performance and efficiency according to different skipping strategies. Fig. 4 presents mAP and FPS depending on the values of $\tau_{1}$ and $\tau_{2}$ in the Sec. 3.4 on $\mathrm{COCO} \rightarrow \mathrm{COCO - C}$ , which are used for two criteria to determine when the adaptation is needed. We also show the simple baseline Evenly-Skip, which adapts every $N$ -th step and skips the rest. In Fig. 4, the blue lines ( $\triangle$ ) show the results when $\tau_{1}$ is changing from 1.0 to infinity, where only criterion 2 is used, while $\tau_{2}$ is fixed at 1.05. As $\tau_{1}$ decreases, more adaptation is required, leading to slower FPS but higher mAP. The green lines ( $\diamond$ ) show the results of changing $\tau_{2}$ , where ‘ $\tau_{2} = \inf$ ' denotes using only criterion 1, without criterion 2. For all main experiments, we set $\tau_{1}$ and $\tau_{2}$ as 1.1 and 1.05, respectively, considering the balance between mAP and FPS. Additionally, our skipping strategy consistently outperforms Evenly-Skip, achieving higher values in both mAP and FPS. This indicates that our criterion for deciding when to bypass model updates provides an effective balance between accuracy and speed. + +When do models actually update? We analyze when the model actually skips adaptation and only performs inference or actively utilizes test samples for model adaptation based on the two criteria we propose. This analysis is conducted in COCO to COCO-C with 15 corruption domains and 1 original domain. Fig. 5a plots the number of backward passes, i.e., the number of batches of test samples used for adaptation, with different values of $\tau_{1}$ for the two backbones. The horizontal and vertical axes represent sequentially incoming test domains and the cumulative backward numbers, respectively. A steep slope in a region indicates frequent adaptation, while a gentle slope indicates skipping adaptation, performing only inference. Notably, even without explicit information about when the test domain changes, the model actively performs adaptation, especially right after the test domain changes. This trend is consistent regardless of changes in $\tau$ value or backbone type. The number of backward passes largely depends on the value of $\tau_{1}$ rather than the backbone type, suggesting that a consistent $\tau_{1}$ value can be used irrespective of the backbone. Fig. 5b illustrates adaptation patterns by dividing backward steps for each domain in the case of Swin-T backbone with $\tau_{1} = 1.1$ . More clearly, it shows that adaptation occurs actively around the points where each domain changes, and afterward, adaptation happens intermittently or almost not at all. The light pink bars represent the performance of Direct-Test, showing that domains with initially high model performance tend to have less adaptation, while domains with lower performance initially need more adaptation. In other words, the amount of skipping adaptation is proportional to the amount of the domain shift. Interestingly, the second domain, 'Shot Noise', shows almost no adaptation despite the lower performance of the Direct-Test. We conjecture that the preceding domain, 'Gaussian Noise', shares a similar nature of noise, leading the model to decide that additional adaptation steps may not be necessary. As a result, our skipping strategy enables the model to efficiently adapt, considering both the original domain the model is trained on and the previous domain the model has been adapted to. + +# 5. Conclusion + +We introduce an efficient Continual Test-time Adaptation (CTA) method for object detection in the continually changing domain. Our approach involves 1) lightweight adaptors, 2) class-wise object-level feature alignment, and 3) skipping unnecessary adaptation, offering highly efficient and effective solution to diverse domain shifts. It significantly improve mAP performance in various CTA scenarios without serious slowdown in the inference speed. + +Acknowledgement This work was supported by NRF grant (2021R1A2C3006659) and IITP grants (2021-0-01343, 2022-0-00320, 2019-0-00079), all funded by the Korea government (MSIT). + +# References + +[1] Alexander Bartler, Florian Bender, Felix Wiewel, and Bin Yang. Ttaps: Test-time adaption by aligning prototypes using self-supervision. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 2 +[2] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2 +[3] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1 +[4] Dhanajit Brahma and Piyush Rai. A probabilistic framework for lifelong test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3582-3591, 2023. 3 +[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 3 +[6] Yijin Chen, Xun Xu, Yongyi Su, and Kui Jia. Stfar: Improving object detection robustness at test-time by self-training with feature alignment regularization. arXiv preprint arXiv:2303.17937, 2023. 1, 2, 3 +[7] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 3 +[8] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 3 +[9] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, et al. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 5 +[10] Shirsendu Sukanta Halder, Jean-François Lalonde, and Raoul de Charette. Physics-based rendering for improving robustness to rain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10203-10212, 2019. 3 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6, 7, 2 +[12] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3 +[13] Junyuan Hong, Lingjuan Lyu, Jiayu Zhou, and Michael Spranger. Mecta: Memory-economic continual test-time model adaptation. In The Eleventh International Conference on Learning Representations, 2022. 3 + +[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2 +[16] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for modelagnostic domain generalization. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[17] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In International Conference on Learning Representations (ICLR), 2023. 3 +[18] Sanghun Jung, Jungsoo Lee, Nanhee Kim, Amirreza Shaban, Byron Boots, and Jaegul Choo. Cafa: Class-aware feature alignment for test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4 +[19] Mehran Khodabandeh, Arash Vahdat, Mani Ranjbar, and William G Macready. A robust learning approach to domain adaptive object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 480-490, 2019. 3 +[20] Donghyun Kim, Kaihong Wang, Stan Sclaroff, and Kate Saenko. A broad study of pre-training for domain generalization and adaptation. In European Conference on Computer Vision, pages 621-638. Springer, 2022. 5 +[21] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6092–6101, 2019. 3 +[22] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8474-8481, 2021. 4 +[23] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation, 2017. 3 +[24] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test-time adaptation, 2023. 2, 3 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5 + +[27] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 1, 2 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6, 7 +[29] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 5 +[30] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2, 6, 1 +[31] Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin, Mateusz Kozinski, Horst Possegger, and Horst Bischof. Actmad: Activation matching to align distributions for test-time-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24152-24161, 2023. 1, 2, 3, 4, 6 +[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 3, 7 +[33] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400, 2023. 3 +[34] Mario obler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 3 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. 2016. 5 +[36] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2, 6, 1 +[38] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2759-2769, 2023. 1, 2, 3, 6 + +[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 6 +[40] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. Advances in Neural Information Processing Systems, 35:17543-17555, 2022. 3, 4 +[41] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 5 +[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2 +[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[44] Vibashan VS, Poojan Oza, and Vishal M Patel. Towards online domain adaptive object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 478-488, 2023. 3 +[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 7 +[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3 +[47] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees GM Snoek. Energy-based test sample adaptation for domain generalization. arXiv preprint arXiv:2302.11215, 2023. 3 +[48] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5 +[49] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 1, 2, 3 +[50] Bowen Zhao, Chen Chen, and Shu-Tao Xia1. Delta: Degradation-free fully test-time adaptation. In International Conference on Learning Representations (ICLR), 2023. 2, 3 \ No newline at end of file diff --git a/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/images.zip b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bcf0828e234d799c72c0600565f05c3cf8ddb079 --- /dev/null +++ b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e591b425ea8954c3a01b8270c67d40e40d305aa4e00c008e0f542101427ad82f +size 454475 diff --git a/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/layout.json b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..77400196c098d3bf4b2321a2e2d11293d826eaea --- /dev/null +++ b/2024/What How and When Should Object Detectors Update in Continually Changing Test Domains_/layout.json @@ -0,0 +1,8677 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 470, + 140 + ], + "type": "text", + "content": "What, How, and When Should Object Detectors Update in Continually Changing Test Domains?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "content": "Jayeon Yoo" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "content": " Dongkwan Lee" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "content": " Inseop Chung" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "content": " Donghyun Kim" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "content": " Nojun Kwak" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 160, + 510, + 190 + ], + "type": "text", + "content": " Seoul National University Korea University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 190, + 479, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 190, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 479, + 204 + ], + "type": "text", + "content": "1{jayeon.yoo, biancco, jis3613, nojunk}@snu.ac.kr, " + }, + { + "bbox": [ + 111, + 190, + 479, + 204 + ], + "type": "inline_equation", + "content": "^2\\mathrm{d}" + }, + { + "bbox": [ + 111, + 190, + 479, + 204 + ], + "type": "text", + "content": " kim@korea.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "text", + "content": "It is a well-known fact that the performance of deep learning models deteriorates when they encounter a distribution shift at test time. Test-time adaptation (TTA) algorithms have been proposed to adapt the model online while inferring test data. However, existing research predominantly focuses on classification tasks through the optimization of batch normalization layers or classification heads, but this approach limits its applicability to various model architectures like Transformers and makes it challenging to apply to other tasks, such as object detection. In this paper, we propose a novel online adaption approach for object detection in continually changing test domains, considering which part of the model to update, how to update it, and when to perform the update. By introducing architecture-agnostic and lightweight adaptor modules and only updating these while leaving the pre-trained backbone unchanged, we can rapidly adapt to new test domains in an efficient way and prevent catastrophic forgetting. Furthermore, we present a practical and straightforward class-wise feature aligning method for object detection to resolve domain shifts. Additionally, we enhance efficiency by determining when the model is sufficiently adapted or when additional adaptation is needed due to changes in the test distribution. Our approach surpasses baselines on widely used benchmarks, achieving improvements of up to " + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "inline_equation", + "content": "4.9\\% p" + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "inline_equation", + "content": "7.9\\% p" + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "text", + "content": " in mAP for COCO " + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 256, + 296, + 605 + ], + "type": "text", + "content": " COCO-corrupted and SHIFT, respectively, while maintaining about 20 FPS or higher. The implementation code is available at https://github.com/natureyoo/ContinualTTA_ObjectDetection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 647, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 288, + 696 + ], + "type": "text", + "content": "Deep learning models excel in many vision tasks but struggle with domain shifts when test data distributions diverge from training data [3, 27, 42]. These shifts, common in real-world settings, arise from natural variations, weather" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 300, + 229, + 552, + 403 + ], + "blocks": [ + { + "bbox": [ + 300, + 229, + 552, + 403 + ], + "lines": [ + { + "bbox": [ + 300, + 229, + 552, + 403 + ], + "spans": [ + { + "bbox": [ + 300, + 229, + 552, + 403 + ], + "type": "image", + "image_path": "9dfbac57f05601a9cee12ea9481b3c07e19a607e1c687920a9668d2a48e2e0ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 406, + 547, + 517 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 517 + ], + "type": "text", + "content": "Figure 1. We propose an online adaptation method for object detection in continually changing test domains. Object detectors trained with clean images suffer from performance degradation due to various corruption, such as camera sensor degradation or environmental changes (Direct-Test). Updating full parameters for online adaptation require a large number of test samples and vulnerable to drastic domain changes (Full-Finetuning), while using only our lightweight adaptor is robust and quickly adapts within a few time steps (Ours). We can further improve efficiency by skipping unnecessary adaptation steps (Ours-Skip)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": "changes (e.g., fog, rain), camera differences (e.g., pixelate, defocus blur), and various other factors. Test-Time Adaptation (TTA) [3, 27, 32, 42, 45, 49] addresses these shifts by adapting models to the target (test) distribution online. Considering the dynamic nature of real-world applications, it becomes crucial to extend this approach to Continual Test-Time Adaptation (CTA), which accounts for the ongoing evolution of test distributions. This is particularly vital in environments like autonomous driving, where conditions can change continuously (e.g., from clear to rainy, or day to night). Despite its significance, continual test-time adaptation for object detection remains underexplored." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Recently, several TTA methods [6, 31, 38] tailored for object detection have been proposed. ActMAD [31] aligns all the output feature maps " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{C\\times H\\times W})" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " after Batch Nor" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 148, + 712 + ], + "type": "text", + "content": "*Co-corresponding authors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23354" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "malization (BN) layers [15] to adapt the test domain to be similar to that of the training domain. However, this approach requires significant memory during adaptation and does not explicitly consider the objects present in the image. TeST [38] and STFAR [6] adapt to a test domain by utilizing weak and strong augmented test samples with a teacher-student network [39], but they significantly increase inference costs since they require additional forward passes and update steps. Also, these methods update all network parameters, making them highly inefficient in online adaptation and vulnerable to losing task-specific knowledge when the test domain experiences continual or drastic changes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 219, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 495 + ], + "type": "text", + "content": "In this paper, we aim to develop an efficient continual test-time adaptation (CTA) method for object detection. We investigate the following three key aspects to improve efficiency; what to update: while previous TTA methods for object detection [6, 31, 38] use full fine-tuning, updating all parameters at test time, they are inefficient and prone to losing task-specific knowledge in relatively complex object detection tasks. Updating BN layers, as done in many TTA methods for classification [18, 32, 45, 49], is not as effective for object detection, given its smaller batch size compared to classification and the limitation in applying various backbones, such as Transformer [28, 43]. how to update: several previous TTA methods for object detection [6, 38] adapt the model by using teacher-student networks, resulting in a significant decrease in inference speed, which is detrimental during test time. While another existing method [31] aligns feature distributions for adaptation, it does not consider each object individually, focusing only on image features, making it less effective for object detection. when to update: most TTA or CTA methods update models using all incoming test samples. However, it is inefficient to update continuously the model if it is already sufficiently adapted when the change of the test domain is not significant." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "To this end, (1) we propose an efficient continual test-time adaptation method for object detectors to adapt to continually changing test domains through the use of lightweight adaptors which require only " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "0.54\\% \\sim 0.89\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " additional parameters compared to the full model. It exhibits efficiency in parameters, memory usage, and adaptation time, along with robustness to continuous domain shifts without catastrophic forgetting. Additionally, it demonstrates wide applicability to various backbone types compared to BN-based TTA methods [18, 24, 32, 45, 49, 50]. (2) To enhance the adaptation effectiveness in the object detection task, we align the feature distribution of the test domain with that of the training domain at both the image-level and object-level using only the mean and variance of features. For estimating the mean of the test domain features, we employ Exponentially Moving Average (EMA) as we can leverage only the current incoming test samples, not the entire test domain data. Due to the unavailability of training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 203 + ], + "type": "text", + "content": "data access, we utilize only the mean and variance of the features from a few training samples. (3) We also introduce two novel criteria that do not require additional resources to determine when the model needs adaptation to enhance efficiency in a continually changing test domain environment. As illustrated in Fig. 1, our approach Ours, employing adaptors, tends to adapt much faster to domain changes compared to full parameter updates. This enables efficient TTA by using only a few test samples to update the adaptor and skipping the rest of the updates as shown in Ours-Skip. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 205, + 547, + 504 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 205, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 547, + 300 + ], + "type": "text", + "content": "- We introduce an architecture-agnostic lightweight adaptor, constituting only a maximum of " + }, + { + "bbox": [ + 306, + 205, + 547, + 300 + ], + "type": "inline_equation", + "content": "0.89\\%" + }, + { + "bbox": [ + 306, + 205, + 547, + 300 + ], + "type": "text", + "content": " of the total model parameters, into the backbone of the object detector to adapt the model in a continually changing test domain. This approach ensures efficiency in parameters, memory usage, and adaptation speed, demonstrating the robust preservation of task-specific knowledge owing to its inherent structural characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 301, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 547, + 372 + ], + "type": "text", + "content": "- We propose a straightforward and effective adaptation loss for CTA in object detection tasks. This is achieved by aligning the distribution of training and test domain features at both the image and object levels, utilizing only the mean and variance of a few training samples and EMA-updated mean features of the test domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 373, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 547, + 433 + ], + "type": "text", + "content": "- We also propose two criteria to determine when the model requires adaptation, enabling dynamic skipping or resuming adaptation as needed. This enhancement significantly boosts inference speed by up to about 2 times while maintaining adaptation performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "type": "text", + "content": "- Our adaptation method proves effective for diverse types of domain shifts, including weather changes and sensor variations, regardless of whether the domain shift is drastic or continuous. In particular, our approach consistently improves the mAP by up to " + }, + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO} \\rightarrow \\mathrm{COCO-C}" + }, + { + "bbox": [ + 306, + 433, + 547, + 504 + ], + "type": "text", + "content": " and SHIFT-Discrete/Continuous with higher than 20 FPS." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 514, + 392, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 392, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 392, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Test-time adaptation. Recently, there has been a surge of interest in research that adapts models online using unlabeled test samples while simultaneously inferring the test sample to address the domain shift problem, where the test data distribution differs from that of the training data. There are two lines for online adaptation to the test domain, Test-time Training (TTT) and Test-time Adaptation (TTA). TTT [1, 2, 27, 42] involves modifying the model architecture during training to train it with self-supervised loss, allowing adaptation to the test domain in the test time by applying this self-supervised loss to the unlabeled test samples. On the other hand, TTA aims to adapt the trained model directly to the test domain without specifically tailored model architectures or losses during training time. NORM [37] and DUA [30] address the domain shifts by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23355" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "adjusting the statistics of batch normalization (BN) layers using the current test samples, without updating other parameters, inspired by [23]. Following this, [24, 32, 45, 50] and [18] update the affine parameters of BN layers using unsupervised loss, entropy minimization loss to enhance the confidence of test data predictions, and feature distribution alignments loss, respectively. Several studies [16, 17] update the classifier head using the pseudo-prototypes from the test domain. However, these methods limit their applicability to architectures without BN layers or to object detection tasks that involve multiple objects in a single image. Others [31, 40, 49] update full parameters for online adaptation to the test domain in an online manner, but this approach is inefficient and susceptible to the noisy signal from the unsupervised loss. While existing TTA methods are oriented towards classification tasks, we aim to propose an effective and efficient method for online adaptation in the object detection task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 579 + ], + "type": "text", + "content": "Continual test-time adaptation. Recent studies [33, 46] point out that existing TTA methods have primarily focused on adapting to test domains following an i.i.d assumption and may not perform well when the test data distribution deviates from this assumption. [46] introduces a ContinuaTTA (CTA) method designed for scenarios where the test domain continuously changes over time. This poses challenges in preventing the model from over-adapting to a particular domain shift and preserving the knowledge of the pre-trained model to avoid catastrophic forgetting. In the field of CTA, the self-training strategy adopting an Exponentially Moving Average (EMA) teacher-student structure is attracting interest as an effective algorithm enabling robust representation to be learned through self-knowledge distillation. In many studies, the EMA teacher-student structure and catastrophic restoration of source model weights have been proposed as a solution to achieve the goal of CTA [4, 46, 47]. Approaches using source replay [34], and anti-forgetting regularization [32] have also achieved good performances in robust continuous adaptation. Furthermore, there is growing attention on methods that mitigate the computational and memory challenges associated with CTA, such as [13], which relies on updates to batch normalization statistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Test-time adaptive object detection. Research on TTA for Object Detection (TTAOD) is progressively emerging [6, 31, 38, 44]. Most existing TTAOD methods [6, 38, 44] exploit a teacher-student network to adapt to the test domain, following the self-training approach commonly employed in Unsupervised Domain Adaptation for object detection [7, 19, 21, 36]. However, it is inefficient for TTA due to data augmentation requirements and additional forward and backward steps, resulting in slower inference speeds and higher memory usage. Another approach, ActMAD [31], aligns the distributions of output feature maps" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": "after all BN layers along the height, width, and channel axes to adapt to the test domain. However, this location-aware feature alignment is limited to datasets with fixed location priors, such as driving datasets, and is less effective for natural images like COCO. Additionally, CTA for Object Detection (CTAOD) have not been thoroughly explored. Therefore, there is a need for an effective CTAOD method considering memory and time efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 180, + 362, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 362, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 362, + 192 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 201, + 547, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 274 + ], + "type": "text", + "content": "To enable the efficient and effective Continual Test-time Adaptation of Object Detectors (CTAOD), we introduce an approach that specifies which part of the model should be updated, describes how to update those using unlabeled test data, and determines whether we perform model updates or not to improve efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 282, + 385, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 385, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 385, + 295 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": "Assume that we have an object detector " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "h \\circ g_{\\Theta}" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": ", here " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " are the RoI head and the backbone, respectively with their parameters being " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": ". The training dataset is denoted as " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "D_{train} = \\{(x_i, y_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "x_i \\sim P_{train}(x)" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "y_i = (bbox_i, c_i)" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": ", containing information on the bounding box (bbox) and class label " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "c_i \\in \\mathcal{C}" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": ". Consider deploying the detector to the test environments where the test data at period " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "x_j^T \\sim P_{test}^T(x)" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "P_{test}^T \\neq P_{train}" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "P_{test}^T" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " deviates from the i.i.d. assumption. In addition, the domain of " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "P_{test}^T" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " continually changes according to " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "P_{test}^T \\neq P_{test}^{T-1}" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": "). Our goal is to adapt the detector " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "h \\circ g" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "P_{test}^T" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " using only test data " + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "inline_equation", + "content": "x_j^T" + }, + { + "bbox": [ + 304, + 300, + 547, + 449 + ], + "type": "text", + "content": " while making predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 455, + 533, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 533, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 533, + 468 + ], + "type": "text", + "content": "3.2. What to update: Adaptation via an adaptor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "Previous methods [6, 31, 38, 44] adapt the model to the test domain by updating all parameters " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": ", leading to inefficiency at test time and a high risk of losing task knowledge from the training data. In contrast, we adapt the model by introducing an adaptor with an extremely small set of parameters and updating only this module while freezing " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": ". We introduce a shallow adaptor in parallel for each block, inspired by [5, 14], where transformer-based models are fine-tuned for downstream tasks through parameter-efficient adaptors, as shown in Fig. 2. Each adaptor consists of down-projection layers " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{down} \\in \\mathbb{R}^{d \\times \\frac{d}{r}}" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": ", up-projection layers " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{up} \\in \\mathbb{R}^{\\frac{d}{r} \\times d}" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": " and RLUs, where " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": " denotes the input channel dimension and " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": " is the channel reduction ratio set to 32 for all adaptors. We use MLP layers for the Transformer block (Fig. 2a) and " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": " convolutional layers for the ResNet block (Fig. 2b) to introduce architecture-agnostic adaptors. The up-projection layer is initialized to 0 values so that the adaptor does not modify the output of the block, but as the adaptor is gradually updated, it adjusts the output of the block to adapt to the test domain. Even as the adaptor" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23356" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 170, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 170, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 170, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 170, + 203 + ], + "type": "image", + "image_path": "6699e81c9c72d975f0d533edc09ae91b2b4adc92fff6ef6d71389a27b98f553e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 288, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 288, + 264 + ], + "type": "text", + "content": "Figure 2. We attach an adaptor, which is a shallow and low-rank MLP or CNN, to every " + }, + { + "bbox": [ + 46, + 218, + 288, + 264 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 218, + 288, + 264 + ], + "type": "text", + "content": " block in parallel. We update only these adaptors while other parameters are frozen. Our approach can be applied to diverse architectures including CNNs and Transformers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 293, + 204 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 293, + 204 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 293, + 204 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 293, + 204 + ], + "type": "image", + "image_path": "d40db235519bc4ba5f5f700896391ad47202e140b5235f9e0ebdee3490839c18.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 288, + 352 + ], + "type": "text", + "content": "is updated in the test domain, the original backbone parameter " + }, + { + "bbox": [ + 46, + 279, + 288, + 352 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 279, + 288, + 352 + ], + "type": "text", + "content": " remains frozen and fully preserved. This structural preservation, as evident in Ours in Fig. 1, enables robust and efficient adaptation to domain changes by maintaining relatively complex task knowledge in object detection and updating very few parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 361, + 257, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 257, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 257, + 373 + ], + "type": "text", + "content": "3.3. How to update: EMA feature alignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": "To adapt the object detector to the test domain, we align the feature distribution of the test domain with that of the training data, inspired by [18, 31, 40]. In contrast to these methods that solely align image feature distribution, we additionally align object-level features in a class-wise manner, considering class frequency, to enhance its effectiveness for object detection. As the training data is not accessible during test time, we pre-compute the first and second-order statistics, denoted as " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\mu_{tr} = \\mathbb{E}[F_{tr}]" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\Sigma_{tr} = \\mathbb{Var}[F_{tr}]" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": ", where the operators " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{Var}" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": " represent the mean and variance respectively. The features " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "F_{tr} = \\{g_{\\Theta}(x_{tr})\\}" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": " are computed using only 2,000 training samples, a small subset of the training data. Since a sufficient amount of test domain data is not available at once, and only the current incoming test data, whose features are denoted as " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "F_{te}^{t}" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": ", is accessible at time step " + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 380, + 288, + 583 + ], + "type": "text", + "content": ", we estimate the mean of test data features using an exponentially moving average (EMA) as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 593, + 287, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 593, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 53, + 593, + 287, + 608 + ], + "type": "interline_equation", + "content": "\\mu_ {t e} ^ {t} = (1 - \\alpha) \\cdot \\mu_ {t e} ^ {t - 1} + \\alpha \\cdot \\mathbb {E} [ F _ {t e} ^ {t} ], \\quad \\mathrm {s . t .} \\mu_ {t e} ^ {0} = \\mu_ {t r}. (1)", + "image_path": "60c66fd145f4b2d96de9fa939d8cdf0d0893ee2d169decced451c5c65c779910.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 654 + ], + "type": "text", + "content": "Considering the typically small batch size in object detection compared to classification, we approximate the variance of the test features as " + }, + { + "bbox": [ + 46, + 617, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\Sigma_{te} \\simeq \\Sigma_{tr}" + }, + { + "bbox": [ + 46, + 617, + 287, + 654 + ], + "type": "text", + "content": " to reduce instability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "Image-level feature alignment. We estimate the training and test feature distributions as normal distributions and minimize the KL divergence between them as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 80, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "L _ {i m g} = \\mathrm {D} _ {K L} (\\mathcal {N} (\\mu_ {t r}, \\Sigma_ {t r}), \\mathcal {N} (\\mu_ {t e} ^ {t}, \\Sigma_ {t r})). \\qquad (2)", + "image_path": "1d7c424f93cfb6127057818409b54db40523d79e1375f9eefd934ef1eb0fcafd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "Region-level class-wise feature alignment. In object detection, we deal with multiple objects within a single image, making it challenging to apply the class-wise feature alignment proposed in [40], a TTA method for classification. To handle region-level features that correspond to an object, we use ground truth bounding boxes for the training data and utilize the class predictions of RoI pooled features, " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "f_{te}^{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ", for unlabeled test data. In object detection, domain shifts often result in lower recall rates, as a significant number of proposals are predicted as background [22]. To mitigate this issue, we filter out features with background scores exceeding a specific threshold. Subsequently, we assign them to the foreground class with the highest probability, as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 235, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 235, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 331, + 235, + 545, + 258 + ], + "type": "interline_equation", + "content": "F _ {t e} ^ {k, t} = \\left\\{f _ {t e} ^ {t} \\mid \\underset {c} {\\operatorname {a r g m a x}} p _ {f g} = k, p _ {b g} < 0. 5 \\right\\}, \\tag {3}", + "image_path": "b8b999f450657a6204025034c7efb35c34e4090f08f918b5e40231dde6f224a4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 257, + 525, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 257, + 525, + 270 + ], + "spans": [ + { + "bbox": [ + 315, + 257, + 525, + 270 + ], + "type": "interline_equation", + "content": "w h e r e h _ {c l s} \\left(f _ {t e} ^ {t}\\right) = \\left[ p _ {f g}, p _ {b g} \\right] = \\left[ p _ {0}, \\dots , p _ {C - 1}, p _ {b g} \\right].", + "image_path": "99abfbf97345d494982963538cf37a493ed44c95ede3cccdb36fd9cf8f8c739c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 546, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 546, + 350 + ], + "type": "text", + "content": "We estimate the class-wise feature distribution of the test domain by exploiting " + }, + { + "bbox": [ + 304, + 277, + 546, + 350 + ], + "type": "inline_equation", + "content": "F_{te}^{k,t}" + }, + { + "bbox": [ + 304, + 277, + 546, + 350 + ], + "type": "text", + "content": " and Eq.1. Furthermore, we introduce a weighting scheme for aligning features of less frequently appearing classes, taking into account the severe class imbalance where specific instance (e.g., person) may appear multiple times within a single image, as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 356, + 507, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 356, + 507, + 372 + ], + "spans": [ + { + "bbox": [ + 332, + 356, + 507, + 372 + ], + "type": "interline_equation", + "content": "N ^ {k, t} = N ^ {k, t - 1} + \\left\\| F _ {t e} ^ {k, t} \\right\\|, \\quad \\text {s . t .} \\quad N ^ {k, 0} = 0", + "image_path": "6a90b994a80f51fbb59a818b12271b40303fe8b3d181346285b570290e768187.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 373, + 545, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 373, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 352, + 373, + 545, + 401 + ], + "type": "interline_equation", + "content": "w ^ {k, t} = \\log \\left(\\frac {\\operatorname* {m a x} _ {i} N ^ {i , t}}{N ^ {k , t}}\\right) + 0. 0 1 \\tag {4}", + "image_path": "c9a0c4c9afc6862be8ff677488a348c26ed69800d51cff71368e5d08f696f83b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 403, + 527, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 403, + 527, + 427 + ], + "spans": [ + { + "bbox": [ + 312, + 403, + 527, + 427 + ], + "type": "interline_equation", + "content": "L _ {o b j} = \\sum_ {k} w ^ {k, t} \\cdot D _ {K L} (\\mathcal {N} (\\mu_ {t r} ^ {k}, \\Sigma_ {t r} ^ {k}), \\mathcal {N} (\\mu_ {t e} ^ {k, t}, \\Sigma_ {t r} ^ {k})).", + "image_path": "44b164a0c7231898b0a81fe57774f030e0c097e2c808bc3eb66bbc8507e6a356.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "text", + "content": "Here, the class-wise mean " + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\mu^k" + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\Sigma^k" + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "text", + "content": " of the training and test data are obtained in the same way as the image-level features. We can effectively adapt the object detector by updating the model to align the feature distribution at both the image and object levels as " + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "inline_equation", + "content": "L = L_{img} + L_{obj}" + }, + { + "bbox": [ + 304, + 436, + 546, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 504, + 518, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 504, + 518, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 518, + 517 + ], + "type": "text", + "content": "3.4. When to update: Adaptation on demand" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 522, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 640 + ], + "type": "text", + "content": "As shown in Fig. 1, Ours, which only updates the adaptor proposed in Sec. 3.2, efficiently adapts to changes in the test domain, even with a small subset of early test samples. We leverage its rapid adaptation characteristics to reduce computational costs by skipping model updates (i.e., skipping backward passes) when the model has already sufficiently adapted to the current test domain and resuming model updates when confronted with a new test domain. Therefore, we introduce two criteria to determine when to update the model or not as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "(Criterion 1) When the distribution gap exceeds the in-domain distribution gap. Recall that " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " (Eq. 2) measures the distribution gap between the test and train distributions. We assume a model is well-adapted to the current test domain when " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is closer to the in-domain distribution gap. We measure the in-domain distribution gap by" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23357" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 272, + 140 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 272, + 140 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 272, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 272, + 140 + ], + "type": "image", + "image_path": "5fdcd17f8b52f5a481a48fea362a9847ce21371e8adb0922ab70cac0fc1f5447.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 155, + 272, + 222 + ], + "blocks": [ + { + "bbox": [ + 111, + 143, + 211, + 155 + ], + "lines": [ + { + "bbox": [ + 111, + 143, + 211, + 155 + ], + "spans": [ + { + "bbox": [ + 111, + 143, + 211, + 155 + ], + "type": "text", + "content": "(a) The ratio of " + }, + { + "bbox": [ + 111, + 143, + 211, + 155 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 111, + 143, + 211, + 155 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 111, + 143, + 211, + 155 + ], + "type": "inline_equation", + "content": "D_{KL}^{in}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 155, + 272, + 222 + ], + "lines": [ + { + "bbox": [ + 48, + 155, + 272, + 222 + ], + "spans": [ + { + "bbox": [ + 48, + 155, + 272, + 222 + ], + "type": "image", + "image_path": "ec62c3ff5daaf694be509d578aebb28b879890ce9a7e77535a31467b488a2301.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 224, + 211, + 236 + ], + "lines": [ + { + "bbox": [ + 110, + 224, + 211, + 236 + ], + "spans": [ + { + "bbox": [ + 110, + 224, + 211, + 236 + ], + "type": "text", + "content": "(b) The ratio of " + }, + { + "bbox": [ + 110, + 224, + 211, + 236 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 110, + 224, + 211, + 236 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 110, + 224, + 211, + 236 + ], + "type": "inline_equation", + "content": "L_{ema}^{t}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 238, + 286, + 260 + ], + "lines": [ + { + "bbox": [ + 47, + 238, + 286, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 286, + 260 + ], + "type": "text", + "content": "Figure 3. The test domain undergoes a shift every 4,000 time steps, and each metric reaches its peak at the same intervals." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "text", + "content": "sampling two disjoint subsets, " + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "text", + "content": ", of training features " + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "inline_equation", + "content": "F_{tr}" + }, + { + "bbox": [ + 47, + 276, + 286, + 300 + ], + "type": "text", + "content": " from Sec. 3.3 as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 308, + 287, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 308, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 80, + 308, + 287, + 323 + ], + "type": "interline_equation", + "content": "D _ {K L} ^ {i n} = D _ {K L} \\left(\\mathcal {N} \\left(\\mu_ {t r} ^ {i}, \\Sigma_ {t r} ^ {i}\\right), \\mathcal {N} \\left(\\mu_ {t r} ^ {j}, \\Sigma_ {t r} ^ {j}\\right)\\right), \\tag {5}", + "image_path": "e6a83832c342a0c7c1d6ae9c76d4631445d6709045cea0aba83f126e8e55b4ef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mu_{tr}^{i},\\Sigma_{tr}^{i}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": " are obtained from " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "x_{i}\\sim P_{train}(x)" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mu_{tr}^{j},\\Sigma_{tr}^{j}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "x_{j}\\sim P_{train}(x)" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": ". In other words, if " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": " is noticeably higher than the in-domain distribution gap " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "D_{KL}^{in}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": ", we consider a model encountering a test domain whose distribution differs from " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "P_{train}(x)" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": " and needs to be updated. Based on this, we introduce a new index " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\frac{L_{img}}{D_{KL}^{in}}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": ". Fig. 3a plots the trend of this index during the model adaptation to a continually changing test domain. It shows that the index has a large value in the early stages of a domain change, decreases rapidly, and then maintains a value close to 1. This index exhibits a similar trend regardless of the backbone type and dataset, as included in the appendix. Therefore, we establish the criterion that model updates are necessary when this index exceeds a certain threshold, " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\frac{L_{img}}{D_{KL}^{in}} >\\tau_{1}" + }, + { + "bbox": [ + 46, + 331, + 287, + 507 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": "(Criterion 2) When the distribution gap suddenly increases. Additionally, we can determine when the test distribution changes and model updates are necessary by observing the trend of the distribution gap (i.e., " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": "). The convergence of " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": " indicates that a model is well-adapted to the current test domain. To put it differently, " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": " will exhibit a sudden increase when the model encounters a new test domain. We introduce an additional index, denoted as " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "\\frac{L_{img}}{L_{ema}^t}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": ", representing the ratio of the current " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": " to its exponentially moving average " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{ema}^t" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": ". We calculate it using the following formula: " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{ema}^t = 0.99 \\cdot L_{ema}^{t-1} + 0.01 \\cdot L_{img}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": ". Fig. 3b illustrates the trend of the ratio of " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": " over the timesteps. It tends to reach a value of 1 as the loss stabilizes at a specific level. Nevertheless, when the model encounters shifts in the test distribution, the ratio experiences a sharp increase, indicating the necessity of a model update when it exceeds a specific threshold, " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "\\tau_2" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "inline_equation", + "content": "\\frac{L_{img}}{L_{ema}^t} > \\tau_2" + }, + { + "bbox": [ + 46, + 506, + 287, + 717 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "If at least one of the two criteria is satisfied, we conclude that the model requires adaptation and proceed to update it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 106, + 386, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 386, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 386, + 119 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 125, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 221 + ], + "type": "text", + "content": "Sec. 4.1 presents the two object detection benchmark datasets with test distributions that change continuously, either in a drastic or gradual manner, and our implementation detail is in 4.2. Sec. 4.4 compares our method with other TTA baselines described in Secs. 4.3.. We present detailed ablation studies of our method analyzing the effectiveness and efficiency of our method in terms of what, how, and when to update the models for CTAOD in Sec. 4.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 228, + 368, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 368, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 368, + 239 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 247, + 509, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 509, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 509, + 258 + ], + "type": "text", + "content": "We experiment with the following three scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO} \\rightarrow \\mathrm{COCO-C}" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " simulates continuous and drastic realistic test domain changes over a long sequence. MS-COCO [25] collects 80 classes of common objects in their natural context with 118k training images and 5k validation images. COCO-C is created by employing 15 types of realistic corruptions [29], such as image distortion and various weather conditions, to simulate test domain changes. In the experiments, the model is only trained on the COCO train set and sequentially evaluated on each corruption in the COCO-C validation set during test-time for reproducing continually changing test domains. Finally, the model is evaluated on the original COCO validation set to assess how well it preserves knowledge of the original domain (denoted as Org.). SHIFT-(Discrete / Continuous) [41] is a synthetic driving image dataset with 6 classes under different conditions using five weather attributes (clear, cloudy, overcast, fog, rain) and three time-of-day attributes (daytime, dawn, night). In SHIFT-Discrete, there are image sets for each attribute, and the model is sequentially evaluated on these attributes, cloudy " + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " overcast " + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " foggy " + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " rainy " + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " dawn " + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " night " + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 258, + 545, + 568 + ], + "type": "text", + "content": " clear which contains 2.4k, 1.6k, 2.7k, 3.2k, 1.2k, 1.4k, and 2.8k validation images, respectively. This simulates scenarios where the domain undergoes drastic changes. In SHIFT-Continuous, the model is evaluated on four sequences, each consisting of 4k frames, continuously transitioning from clear to foggy (or rainy) and back to clear." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 575, + 434, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 434, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 434, + 588 + ], + "type": "text", + "content": "4.2. Implementation Detail" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We conduct experiments on Faster-RCNN [35] with FPN [26], using ImageNet1k pre-trained ResNet50 [11] and ImageNet21k pre-trained Swin-Tiny [28] backbones. Specifically, Swin-Tiny is employed for its superior domain transfer [20] and detection efficacy over other advanced backbones [9]. For COCO " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " COCO-C with ResNet50, we use a publicly available model trained with strong augmentation [48]. Other models are trained using detectron2 framework following [35] and [28]. For test-time adaptation, we always set the learning to 0.001 for the SGD opti" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23358" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 105, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on COCO " + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": " COCO-C. Our model consistently outperforms baselines on the two different backbones. Furthermore, Ours-Skip with ResNet notably reduces backward passes by as much as " + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "90.5\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": ", leading to a significantly improved frames per second (FPS) rate by up to " + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "109.9\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 105, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 105, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 547, + 240 + ], + "type": "table", + "html": "
BackboneMethodNoiseBlurWeatherDigitalOrg.Avg.# step
GauShtImpDefGlsMtnZmSnwFrsFogBrtCntElsPxJpgFor.Back.FPS
Swin-T [28]Direct-Test9.711.410.013.47.512.15.220.724.836.136.012.919.14.915.843.017.780K021.5
ActMAD10.712.09.412.35.79.54.515.317.527.628.21.116.72.68.736.313.980K80K8.3
Mean-Teacher10.012.111.212.88.112.14.919.623.734.934.08.018.96.117.641.017.2160K80K6.9
Ours13.616.616.114.013.614.28.323.727.237.436.427.227.222.222.342.322.680K80K9.5
Ours-Skip13.315.315.114.012.813.96.522.025.435.534.926.525.923.420.241.221.680K9.7K17.7
ResNet50 [11]Direct-Test9.111.09.812.64.58.84.619.123.138.438.021.415.65.311.944.217.380K025.8
NORM9.911.911.012.65.29.15.119.423.538.237.622.417.25.710.343.417.580K025.8
DUA9.811.710.812.85.28.95.119.323.738.437.822.317.25.410.144.117.180K025.8
ActMAD9.19.67.011.03.26.13.312.814.027.727.83.912.92.37.234.310.580K80K9.6
Mean-Teacher9.612.512.04.02.94.83.116.223.535.134.021.816.68.212.740.314.5160K80K8.1
Ours12.717.817.512.411.511.36.622.826.938.638.528.025.121.222.241.822.280K80K10.1
Ours-Skip14.417.116.013.911.712.26.322.125.537.737.125.524.123.121.142.821.980K7.6K21.2
", + "image_path": "723e48cd2ad214241f6c38b1745b401a8010954b1c64198aa31dd590a26b7a5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "text", + "content": "mizer, and " + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "text", + "content": " of Eq. 1 to 0.01, while " + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 46, + 255, + 287, + 291 + ], + "type": "text", + "content": " are set to 1.1 and 1.05, respectively. We use the same hyper-parameters across all backbones and datasets, with a batch size of 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 298, + 113, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 113, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 113, + 310 + ], + "type": "text", + "content": "4.3. Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 316, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 316, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 45, + 316, + 288, + 509 + ], + "type": "text", + "content": "Direct-Test evaluates the model trained in the training domain without adaptation to the test domain. ActMAD [31] is a TTA method aligning the distribution of output features across all BN layers. To apply ActMAD to the Swin Transformer-based model, we align the output features of the LN layers. We implement Mean-Teacher using a teacher-student network framework to reproduce as close as possible to TeST [38], as its implementation is not publicly available. We follow the FixMatch [39] augmentation method and report results after tuning all hyper-parameters in our scenario. NORM [37] and DUA [30], TTA methods initially designed for classification, are directly applicable to detection tasks by either mixing a certain amount of current batch statistics or updating batch statistics via EMA. However, these are only compatible with architectures containing BN layers. Additional details are provided in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 132, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 132, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 132, + 528 + ], + "type": "text", + "content": "4.4. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 594 + ], + "type": "text", + "content": "We compare the performance of each method using mAP and efficiency metrics, including the number of forward and backward passes, as well as FPS during test-time adaptation. Results of COCO and SHIFT are in Tab. 1 and 2, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{COCO} \\rightarrow \\mathbf{COCO-C}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Tab. 1 demonstrates the effective adaptation performance of Ours in the challenging COCO benchmark with 80 classes due to object-level class-wise feature alignment. ActMAD also aligns feature distribution for TTA, but is not effective since it only aligns whole feature maps without considering specific classes in the image. NORM and DUA, applicable only to ResNet [11], show minimal performance improvement by adaptation as they are not specifically tailored for object detection and only modify batch statistics across the entire feature map. Ad" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "type": "text", + "content": "ditionally, ActMAD and Mean-Teacher, updating full parameters, gradually lose task knowledge in the continually changing test distributions, resulting in much lower performance on Org., the domain identical to the training data, than that of Direct-Test. In contrast, Ours effectively prevents catastrophic forgetting by freezing the original parameters of the models and updating only the adaptor, obtaining performance on par with Direct-Test on the Org. domain and consistently high performance across corrupted domains, with an average mAP improvement of " + }, + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "type": "text", + "content": " compared to that of Direct-Test. Furthermore, leveraging the rapid adaptation ability of the adaptor, Ours-Skip, which skips unnecessary adaptation, allows using only a maximum of about " + }, + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 304, + 255, + 547, + 460 + ], + "type": "text", + "content": " of the total samples for adaptation without significant performance loss. This leads to a substantial improvement in inference speed, more than doubling compared to other TTA methods, reaching over 17.7 FPS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "SHIFT-Discrete. Ours is also effective in SHIFT, which simulates continuous changes in weather and time in driving scenarios according to the left section of Tab. 2. Especially, Ours shows significant improvements in mAP by " + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "7 - 9\\% \\mathrm{p}" + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": ", particularly for the foggy and dawn attributes where Direct-Test obtains lower performance due to severe domain shift. In contrast, with ActMAD, catastrophic forgetting takes place when adapting to the cloudy and overcast weather. This is due to the updating of the full parameters, despite that Direct-Test already shows proficient performance in these conditions. As a result, the performance in the later domains is worse than that of the Direct-Test. DUA, which updates batch statistics using EMA, shows a gradual decrease in performance as the domain continuously changes, resulting in much lower performance in the original clear domain (i.e., clear). On the other hand, NORM, which utilizes the statistics of the current batch samples, exhibits no catastrophic forgetting and relatively good adaptation, as SHIFT is a relatively easier task compared to COCO due to having only 6 classes. Compared to NORM, Ours shows better adaptation performance, and is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23359" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 116, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 115 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 115 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 115 + ], + "type": "text", + "content": "Table 2. Comparison of mAP, the number of backward and forward passes, and FPS between baselines and our model on SHIFT-Discrete and SHIFT-Continuous. Baselines perform effectively in a particular setting but lack generalizability across various settings. Our method consistently achieves results that are either better or on par with the best model in all settings, demonstrating its strong stability. Ours-Skip also effectively reduces the number of backward passes without compromising mAP performance, resulting in a higher FPS." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 116, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 116, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 547, + 262 + ], + "type": "table", + "html": "
BackboneMethodSHIFT-DiscreteSHIFT-Continuous
mAP# stepmAP# Avg. step
cloudyoverc.fograindawnnightclearAvg.For.Back.FPSclear↔fogclear↔rainFor.Back.FPS
Swin-T [28]Direct-Test50.038.923.145.126.939.545.938.515.3K027.518.121.14K028.3
ActMAD49.838.421.443.119.032.044.835.515.3K15.3K9.315.616.34K4K9.8
Mean-Teacher50.039.225.745.426.037.542.238.030.6K15.3K7.820.424.38K4K6.5
Ours50.339.232.246.730.439.944.340.415.3K15.3K11.223.922.64K4K11.6
Ours-Skip50.339.729.147.130.241.545.940.615.3K6.1K20.025.123.84K0.83K19.2
ResNet50 [11]Direct-Test49.437.919.743.120.135.345.635.915.3K030.112.115.44K030.0
NORM49.738.622.944.725.137.445.537.715.3K030.116.919.44K030.0
DUA45.231.527.731.915.218.621.127.315.3K030.122.522.44K030.0
ActMAD49.237.718.040.616.032.944.334.115.3K15.3K11.312.716.34K4K11.2
Mean-Teacher49.638.426.843.426.633.141.637.130.6K15.3K9.916.020.88K4K9.8
Ours49.738.727.446.327.437.643.838.715.3K15.3K12.920.921.94K4K13.9
Ours-Skip49.738.826.946.227.638.845.039.015.3K8.9K21.520.022.54K0.75K21.3
", + "image_path": "814cdaaff59f1f33347e296af4745cc9c7abffbff0584953a08fd87f19b24aee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 322, + 288, + 420 + ], + "blocks": [ + { + "bbox": [ + 46, + 276, + 287, + 320 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 320 + ], + "type": "text", + "content": "Table 3. Comparison of adaptation performance (mAP), the number of trainable parameters (# Params), and memory usage (Cache) according to which part of the backbone is updated. SD / SC denotes SHIFT-Discrete/Continuous, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 322, + 288, + 420 + ], + "lines": [ + { + "bbox": [ + 48, + 322, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 288, + 420 + ], + "type": "table", + "html": "
BackboneTrainable ParamsmAP# ParamsCache
SDSCNumRatioAvg.Max
Swin-TFull-params38.420.627.7M100%0.8611.0
LayerNorm38.520.00.03M0.1%0.657.49
adaptor (Ours)40.423.20.15M0.5%0.656.96
ResNet50Full-params37.620.423.7M100%1.659.29
BatchNorm37.920.20.05M0.2%1.479.11
adaptor (Ours)38.721.70.21M0.9%1.485.41
", + "image_path": "ed825c1a1ac9af85f56aec0af0d9cdc8f6442c25e9d030011223b811f2b8920b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 259, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 259, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 259, + 447 + ], + "type": "text", + "content": "also applicable to BN-layer-free Swin Transformers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 616 + ], + "type": "text", + "content": "SHIFT-Continuous. In scenarios where the test domain gradually changes across the entire sequence, Ours also demonstrates effectiveness, improving mAP by up to " + }, + { + "bbox": [ + 46, + 447, + 287, + 616 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 447, + 287, + 616 + ], + "type": "text", + "content": " p, as shown in the right section of Tab. 2. While DUA performs well in the clear to foggy transition, it is prone to catastrophic forgetting in situations where the sequence becomes longer, and the test domain changes more diversely, as seen in the left section. Our strategy for determining when model adaptation is necessary is particularly effective in SHIFT. It improves FPS by about 9, reaching about 20 FPS, while enhancing mAP. This is likely due to avoiding overfitting that can occur when adapting to all repetitive frames in SHIFT, which consists of continuous frames, leading to improvements in both inference speed and adaptation performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 164, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 164, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 164, + 635 + ], + "type": "text", + "content": "4.5. Additional Analyses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": "We aim to demonstrate the effectiveness and detailed analysis of our proposed model in terms of 1) which parts of, 2) how, and 3) when the model should be updated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Which part to update? Tab. 3 shows how updating different parts of the backbone model affects the performance and the memory usage during continual test-time adapta" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 322, + 538, + 383 + ], + "blocks": [ + { + "bbox": [ + 304, + 276, + 547, + 320 + ], + "lines": [ + { + "bbox": [ + 304, + 276, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 547, + 320 + ], + "type": "text", + "content": "Table 4. Ablation on each component of our loss. SHIFT-D / C denotes SHIFT-Discrete / Continuous, respectively. The left and right value in each cell corresponds to the mAP for the Swin-T and ResNet50 backbone, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 322, + 538, + 383 + ], + "lines": [ + { + "bbox": [ + 310, + 322, + 538, + 383 + ], + "spans": [ + { + "bbox": [ + 310, + 322, + 538, + 383 + ], + "type": "table", + "html": "
LimgLobjCOCOSHIFT-D.SHIFT-C.
--17.7/ 17.338.5/ 35.919.6/ 13.8
✓-16.7/ 18.136.6/ 37.019.1/ 16.0
✓no class weight17.8/ 18.939.7/ 38.025.1/ 23.4
✓class weight wk,t22.6/ 22.240.4/ 38.723.2/ 21.7
", + "image_path": "322eb45c74e3137811d3e87bd2f13255e7b2c33ee6ea57d1453667d53ab56705.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "type": "text", + "content": "tion. We compare (1) updating full parameters, (2) affine parameters of the normalization layer, and (3) our proposed adaptor for each backbone on the SHIFT dataset. Although our adaptor has fewer parameters, about " + }, + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "type": "text", + "content": " or less of the full parameters, it demonstrates the best adaptation performance. Updating only the affine parameters of the normalization layer, while having fewer parameters, seems less effective for adaptation in object detection compared to classification [32, 45]. Additionally, our adaptor requires only about " + }, + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 400, + 545, + 532 + ], + "type": "text", + "content": " of the memory compared to updating the full parameters, making it memory-efficient." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Ablation study on each component in our loss. Tab. 4 presents the effects of image-level feature alignment, " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", object-level feature class-wise alignment " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{obj}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", and class frequency weighting " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "w^{k,t}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " proposed to address class imbalance. Aligning only the image-level feature distribution with " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{img}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " (first row) leads to modest adaptation in the ResNet50 backbone, while performance in the SwinT backbone is even lower than without adaptation. Notably, aligning object-level features with " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{obj}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " leads to a substantial improvement, with the mAP increasing by approximately " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " compared to the no-adaptation scenario. Introducing class-specific frequency-based weighting " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "w^{k,t}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", despite a slight performance decrease in the SHIFT-Continuous setting, proves highly effective, particularly in scenarios with significant class imbalance, such as COCO" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23360" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 169, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 169, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 169, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 169, + 177 + ], + "type": "image", + "image_path": "0059f846b713f8181cc7e07ca39ecd830bfa65635687bdd667bc6c446e5734ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 178, + 160, + 187 + ], + "lines": [ + { + "bbox": [ + 56, + 178, + 160, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 160, + 187 + ], + "type": "text", + "content": "(a) Swin Transformer backbone" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 289, + 177 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 289, + 177 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 289, + 177 + ], + "type": "image", + "image_path": "9b060176513247b6cdd997985aa5eb06888b9970d1159a8e387ea328ae0ece65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 178, + 265, + 187 + ], + "lines": [ + { + "bbox": [ + 187, + 178, + 265, + 187 + ], + "spans": [ + { + "bbox": [ + 187, + 178, + 265, + 187 + ], + "type": "text", + "content": "(b) ResNet50 backbone" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 266, + 285, + 341 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": "Figure 4. Comparison of mAP and FPS from Ours-Skip with varying values of " + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "inline_equation", + "content": "\\tau_{1}(\\diamond)" + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "inline_equation", + "content": "\\tau_{2}(\\triangle)" + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": " against Evenly-Skip " + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "inline_equation", + "content": "(\\times)" + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": ", adapting every " + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": "-th instances, on COCO " + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": " COCO-C using both (a) Swin-T and (b) ResNet50. The upward and rightward movement indicates a better strategy with higher mAP and faster inference speed, showing that Ours-Skip is consistently better than Evenly-Skip." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 266, + 285, + 341 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 285, + 341 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 285, + 341 + ], + "type": "image", + "image_path": "147bb4ac8b27f43f9b253acb0e1e9e798e1089eb73839c7e1be1d64b7283cf2b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 60, + 354, + 286, + 423 + ], + "blocks": [ + { + "bbox": [ + 102, + 343, + 244, + 353 + ], + "lines": [ + { + "bbox": [ + 102, + 343, + 244, + 353 + ], + "spans": [ + { + "bbox": [ + 102, + 343, + 244, + 353 + ], + "type": "text", + "content": "(a) Accumulated number of backward steps" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 354, + 286, + 423 + ], + "lines": [ + { + "bbox": [ + 60, + 354, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 60, + 354, + 286, + 423 + ], + "type": "image", + "image_path": "7ece3d0df3f5442d9afaf91a80237d30caa564bd3664eb2cf442291defdacc59.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 425, + 287, + 435 + ], + "lines": [ + { + "bbox": [ + 59, + 425, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 59, + 425, + 287, + 435 + ], + "type": "text", + "content": "(b) Number of backward steps and mAP of Direct-Test in each domain" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 442, + 258, + 453 + ], + "lines": [ + { + "bbox": [ + 75, + 442, + 258, + 453 + ], + "spans": [ + { + "bbox": [ + 75, + 442, + 258, + 453 + ], + "type": "text", + "content": "Figure 5. Analysis of the adaptation of Ours-Skip." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": "with 80 classes, where it enhances the mAP by around " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " p. Trade-off between adaptation performance and efficiency according to different skipping strategies. Fig. 4 presents mAP and FPS depending on the values of " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " in the Sec. 3.4 on " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO} \\rightarrow \\mathrm{COCO - C}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": ", which are used for two criteria to determine when the adaptation is needed. We also show the simple baseline Evenly-Skip, which adapts every " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": "-th step and skips the rest. In Fig. 4, the blue lines (" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": ") show the results when " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " is changing from 1.0 to infinity, where only criterion 2 is used, while " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " is fixed at 1.05. As " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " decreases, more adaptation is required, leading to slower FPS but higher mAP. The green lines (" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": ") show the results of changing " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": ", where ‘" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{2} = \\inf" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": "' denotes using only criterion 1, without criterion 2. For all main experiments, we set " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 46, + 472, + 287, + 714 + ], + "type": "text", + "content": " as 1.1 and 1.05, respectively, considering the balance between mAP and FPS. Additionally, our skipping strategy consistently outperforms Evenly-Skip, achieving higher values in both mAP and FPS. This indicates that our criterion for deciding when to bypass model updates provides an effective balance between accuracy and speed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "content": "When do models actually update? We analyze when the model actually skips adaptation and only performs inference or actively utilizes test samples for model adaptation based on the two criteria we propose. This analysis is conducted in COCO to COCO-C with 15 corruption domains and 1 original domain. Fig. 5a plots the number of backward passes, i.e., the number of batches of test samples used for adaptation, with different values of " + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "content": " for the two backbones. The horizontal and vertical axes represent sequentially incoming test domains and the cumulative backward numbers, respectively. A steep slope in a region indicates frequent adaptation, while a gentle slope indicates skipping adaptation, performing only inference. Notably, even without explicit information about when the test domain changes, the model actively performs adaptation, especially right after the test domain changes. This trend is consistent regardless of changes in " + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "content": " value or backbone type. The number of backward passes largely depends on the value of " + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "content": " rather than the backbone type, suggesting that a consistent " + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "content": " value can be used irrespective of the backbone. Fig. 5b illustrates adaptation patterns by dividing backward steps for each domain in the case of Swin-T backbone with " + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\tau_{1} = 1.1" + }, + { + "bbox": [ + 304, + 72, + 547, + 540 + ], + "type": "text", + "content": ". More clearly, it shows that adaptation occurs actively around the points where each domain changes, and afterward, adaptation happens intermittently or almost not at all. The light pink bars represent the performance of Direct-Test, showing that domains with initially high model performance tend to have less adaptation, while domains with lower performance initially need more adaptation. In other words, the amount of skipping adaptation is proportional to the amount of the domain shift. Interestingly, the second domain, 'Shot Noise', shows almost no adaptation despite the lower performance of the Direct-Test. We conjecture that the preceding domain, 'Gaussian Noise', shares a similar nature of noise, leading the model to decide that additional adaptation steps may not be necessary. As a result, our skipping strategy enables the model to efficiently adapt, considering both the original domain the model is trained on and the previous domain the model has been adapted to." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "type": "text", + "content": "We introduce an efficient Continual Test-time Adaptation (CTA) method for object detection in the continually changing domain. Our approach involves 1) lightweight adaptors, 2) class-wise object-level feature alignment, and 3) skipping unnecessary adaptation, offering highly efficient and effective solution to diverse domain shifts. It significantly improve mAP performance in various CTA scenarios without serious slowdown in the inference speed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was supported by NRF grant (2021R1A2C3006659) and IITP grants (2021-0-01343, 2022-0-00320, 2019-0-00079), all funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23361" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Alexander Bartler, Florian Bender, Felix Wiewel, and Bin Yang. Ttaps: Test-time adaption by aligning prototypes using self-supervision. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Dhanajit Brahma and Piyush Rai. A probabilistic framework for lifelong test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3582-3591, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "text", + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "text", + "content": "[6] Yijin Chen, Xun Xu, Yongyi Su, and Kui Jia. Stfar: Improving object detection robustness at test-time by self-training with feature alignment regularization. arXiv preprint arXiv:2303.17937, 2023. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "text", + "content": "[7] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "text", + "content": "[8] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 534 + ], + "type": "text", + "content": "[9] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, et al. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[10] Shirsendu Sukanta Halder, Jean-François Lalonde, and Raoul de Charette. Physics-based rendering for improving robustness to rain. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10203-10212, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6, 7, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "text", + "content": "[12] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Junyuan Hong, Lingjuan Lyu, Jiayu Zhou, and Michael Spranger. Mecta: Memory-economic continual test-time model adaptation. In The Eleventh International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "text", + "content": "[16] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for modelagnostic domain generalization. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[17] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[18] Sanghun Jung, Jungsoo Lee, Nanhee Kim, Amirreza Shaban, Byron Boots, and Jaegul Choo. Cafa: Class-aware feature alignment for test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "text", + "content": "[19] Mehran Khodabandeh, Arash Vahdat, Mani Ranjbar, and William G Macready. A robust learning approach to domain adaptive object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 480-490, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "text", + "content": "[20] Donghyun Kim, Kaihong Wang, Stan Sclaroff, and Kate Saenko. A broad study of pre-training for domain generalization and adaptation. In European Conference on Computer Vision, pages 621-638. Springer, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "type": "text", + "content": "[21] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6092–6101, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "text", + "content": "[22] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8474-8481, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "text", + "content": "[23] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "type": "text", + "content": "[24] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test-time adaptation, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 657 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23362" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "text", + "content": "[29] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "text", + "content": "[30] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2, 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "type": "text", + "content": "[31] Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin, Mateusz Kozinski, Horst Possegger, and Horst Bischof. Actmad: Activation matching to align distributions for test-time-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24152-24161, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "type": "text", + "content": "[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[33] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "text", + "content": "[34] Mario obler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 602 + ], + "type": "text", + "content": "[36] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "text", + "content": "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2, 6, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[38] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2759-2769, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 654 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "text", + "content": "[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "text", + "content": "[40] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. Advances in Neural Information Processing Systems, 35:17543-17555, 2022. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 184, + 546, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 546, + 250 + ], + "type": "text", + "content": "[41] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 251, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 546, + 306 + ], + "type": "text", + "content": "[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 308, + 546, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 351 + ], + "type": "text", + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 352, + 546, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 546, + 396 + ], + "type": "text", + "content": "[44] Vibashan VS, Poojan Oza, and Vishal M Patel. Towards online domain adaptive object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 478-488, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 396, + 546, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 546, + 440 + ], + "type": "text", + "content": "[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 441, + 547, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 547, + 485 + ], + "type": "text", + "content": "[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 487, + 546, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 546, + 520 + ], + "type": "text", + "content": "[47] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees GM Snoek. Energy-based test sample adaptation for domain generalization. arXiv preprint arXiv:2302.11215, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 520, + 546, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 546, + 575 + ], + "type": "text", + "content": "[48] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 576, + 546, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 546, + 619 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 546, + 619 + ], + "type": "text", + "content": "[49] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 621, + 546, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 621, + 546, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 621, + 546, + 654 + ], + "type": "text", + "content": "[50] Bowen Zhao, Chen Chen, and Shu-Tao Xia1. Delta: Degradation-free fully test-time adaptation. In International Conference on Learning Representations (ICLR), 2023. 2, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23363" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_content_list.json b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb6f8a8130a38638842cbc4359f769d710a8edd1 --- /dev/null +++ b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_content_list.json @@ -0,0 +1,1410 @@ +[ + { + "type": "text", + "text": "What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Letian Zhang $^{1}$ Xiaotong Zhai $^{2}$ Zhongkai Zhao $^{3,6}$ Yongshuo Zong $^{4}$ Xin Wen $^{5,6}$ Bingchen Zhao $^{4,6}$", + "bbox": [ + 76, + 203, + 898, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tongji University $^{2}$ University of Warwick $^{3}$ National University of Singapore", + "bbox": [ + 215, + 228, + 776, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ University of Edinburgh $^{5}$ The University of Hong Kong $^{6}$ LunarAI", + "bbox": [ + 246, + 247, + 738, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Counterfactual reasoning, a fundamental aspect of human cognition, involves contemplating alternatives to established facts or past events, significantly enhancing our abilities in planning and decision-making. In light of the advancements in current multi-modal large language models, we explore their effectiveness in counterfactual reasoning. To facilitate this investigation, we introduce a novel dataset, C-VQA, specifically designed to examine the counterfactual reasoning capabilities of modern multi-modal large language models. This dataset is constructed by infusing original questions with counterfactual presuppositions, spanning various types such as numerical and boolean queries. It encompasses a mix of real and synthetic data, representing a wide range of difficulty levels. Our thorough evaluations of contemporary vision-language models using this dataset have revealed substantial performance drops, with some models showing up to a $40\\%$ decrease, highlighting a significant gap between current models and human-like vision reasoning capabilities. We hope our dataset will serve as a vital benchmark for evaluating the counterfactual reasoning capabilities of models. Code and dataset are publicly available at https://bzhao.me/C-VQA/.", + "bbox": [ + 75, + 330, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Counterfactuals are the building blocks of moral behavior as well as scientific thought.\"", + "bbox": [ + 120, + 719, + 439, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "— Judea Pearl, The Book of Why", + "bbox": [ + 215, + 750, + 434, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Counterfactual ability is a pivotal cognitive function in humans, enabling us to envision alternate realities and outcomes based on different choices or events. This capacity underpins our decision-making, moral reasoning, and problem-solving skills. Recent development of multi-modal large language models (MLLMs) [26, 42, 45, 48] have dramatically improved the capabilities of image recognition, image-based dialogue, and language grounding, etc. [4, 9,", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90f9414a2e20aef1b14dac3483756b636b888c2c3ca58904e03e8453ab307c9d.jpg", + "image_caption": [ + "Figure 1. Examples of $C$ -VQA(top), and performance comparison of LLaVA-1.5 [25] w/ and w/o counterfactuality (bottom). $C$ -VQA is constructed by adding counterfactual presuppositions to the questions. We observe that state-of-the-art models all exhibited significant performance drops on the counterfactual questions." + ], + "image_footnote": [], + "bbox": [ + 506, + 297, + 890, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "11, 12, 14, 20, 24, 26, 36]. These developments raise the possibility of achieving higher levels of artificial intelligence. Consequently, we pose a critical question: are contemporary MLLMs equipped for counterfactual reasoning?", + "bbox": [ + 498, + 731, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual question answering is one of the central tasks to evaluate the ability of vision-language models. Current benchmarks focus on evaluating different aspects of the abilities, such as visual recognition (e.g., VQAv2 [13]), external knowledge (e.g., OKVQA [29]), compositional reasoning (e.g., CLEVR [19]). They only require the model to understand the image contents by grounding the concept in", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21853", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "volved in the language questions to the image content. With the help of large language models, these tasks are relatively easy to solve, as evidenced by the high scores achieved by modern MLLMs [5, 8, 26]. The most relevant benchmarks are CF-VQA [30] and VQA-CP [1]. However, they focus on constructing the counterfactual questions between the training set and the test set, e.g., showing a green banana in the test set while most of the bananas in the training set are yellow. They do not evaluate the counterfactual ability of the MLLMs - the skill to imagine \"what if\" scenarios that differ from what actually happened.", + "bbox": [ + 73, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this, we propose a novel and challenging evaluation scenario where the language model not only needs to query for the correct visual representation of image content but also needs to be able to perform counterfactual reasoning on those representations. This cannot be solved by encoding world knowledge or object spatial relation alone, but also requires the model to further understand and imagine the given scenarios. Specifically, in our proposed benchmark, each visual question is modified by a counterfactual presupposition - making an assumption based on a scenario that did not actually happen, but could have. This requires the MLLMs to understand the visual content and then perform reasoning over the image contents accordingly to answer the questions. For example, in Fig. 1, to answer the question of \"How many cats are in the image?\", the model can directly query the image and count the occurrence of the visual concept \"cat\". However, with the counterfactual presupposition \"if the TV was off\", one of the cats on the screen will not be visible. Thus, it poses a more challenging scenario for the models as the model not only needs to identify the occurrence of the concept \"cat\", but also reason about its state after a counterfactual operation is taken.", + "bbox": [ + 75, + 258, + 472, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We build our dataset on top of the commonly used VQAv2 dataset [3], where we collect over 3,000 images suitable for asking counterfactual questions. In addition, we also created over 3,000 synthetic images and corresponding questions to further test the out-of-distribution ability of multi-modal large language models.", + "bbox": [ + 75, + 594, + 470, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate a wide array of contemporery state-of-the-art multi-modal language models on our dataset, such as LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], Instruct-BLIP [8], Qwen-VL [4] and CogVLM [37]. Our experiments show several interesting findings: (1) Neuro-symbolic models perform worse than end-to-end models on complex counterfactual reasoning; (2) No model family can consistently address counterfactual questions. All of the models suffer from a large performance drop from our counterfactual questions; (3) Even the strongest GPT-4V model [32] cannot solve our benchmark; (4) The MLLMs also demonstrate a systematic bias in answering gender-related counterfactual questions.", + "bbox": [ + 75, + 686, + 470, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the main contributions of this paper:", + "bbox": [ + 96, + 885, + 439, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ecf92ba4ce1f7aff1ec5213fabeec088e2fb42ba95188c2386b05d4eb04240f4.jpg", + "image_caption": [ + "Figure 2. Our annotation flow for $C$ -VQA-Real. We select images and questions from the VQAv2 dataset [3], and then utilize ChatGPT to add counterfactual presupposition to the questions and get the corresponding answers. All questions and answers are carefully inspected by human annotators." + ], + "image_footnote": [], + "bbox": [ + 509, + 90, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel and challenging dataset $C$ -VQA with both real and synthetic image-question pairs. The questions contain counterfactual presuppositions and are of various difficulties.", + "2. We conduct extensive evaluation of current state-of-the-art vision-language models on our dataset. We show that they struggle with counterfactual reasoning.", + "3. We summarize our findings from the experiments on proposed $C$ -VQA, hoping to provide valuable insights for future research in this field." + ], + "bbox": [ + 500, + 556, + 890, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 723, + 650, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Question Answering. Visual question answering (VQA) aims to evaluate machines' capabilities of visual understanding, visual reasoning, and the application of commonsense knowledge. Several datasets have been proposed for VQA [3, 6, 13, 18, 19, 23, 29, 34, 39]. COCOQA [34] and VQA [3] are the first to propose the task of VQA, and they contain enormous pictures and questions covering daily life objects. Builds on the VQA dataset. VQAv2 [13] presents a more rigorous evaluation for the VQA task by mitigating language biases and shortcuts in", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21854", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the VQA dataset. Further works [18, 19, 29] continue to extend VQA evaluation to different aspects of image understanding. CLEVR dataset [19] devises a pipeline for generating synthetic data for evaluating the compositional reasoning ability of VQA models. GQA dataset [18] leverages the scene graph structure to generate reasoning questions on real-world images to test the compositional reasoning ability. In this work, we build our dataset $C$ -VQA on the top of VQAv2 [13], where the questions and answers are modified by counterfactual presupposition. Different from previous VQA tasks, our proposed counterfactual dataset provides a new challenging scenario for complementary MLLMs. It further examines the models' abilities to parse the scene structure and reason about the observed world after the counterfactual presuppositions.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation of Reasoning Abilities. The evaluation of generative models has always been a difficult and challenging problem, as well as for large language models. There have been many efforts for the evaluation of the reasoning ability of LLMs in different aspects [1, 7, 16, 17, 30, 40, 43]. For example, GSM8k [7] and MATH [17] evaluate the mathematical reasoning of LLMs, while MMLU [16] aggregates a diverse range of subjects and tasks for evaluation. For counterfactual reasoning, IfQA [43] proposes the first QA dataset that specifically designed to assess the counterfactual reasoning capabilities of language models. In [40], the counterfactual reasoning abilities of state-of-the-art LLMs (GPT-4 [31], Claude [2]) are evaluated under eleven different tasks with counterfactual presuppositions. It is shown that current LLMs cannot reason with counterfactuals reliably. Most related to our work is CF-VQA [30] and VQA-CP [1] which focus on testing the ability of VQA models to answer \"counterfactual\" questions which are defined as having different property distribution in training and testing set, such as training with yellow bananas but test with green ones. In our work, we propose to build the first visual question-answering dataset with counterfactual presuppositions directly added to the question text, this enables the testing of current SOTA MLLMs which leverages internet-scale training data. Our evaluation results show current approaches for MLLMs do not facilitate counterfactual reasoning. Thus, further work should be done to create stronger MLLMs.", + "bbox": [ + 75, + 329, + 472, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Modal LLM Benchmarks With the increasing interest in training novel multi-modal LLMs that can perceive multi-modal inputs, various benchmarks for evaluating the performance of MLLMs are proposed. MME [10] measures the performance of multi-modal LLMs using both perception and cognition abilities on 14 subtasks. MM-Bench [27] covers over 20 different ability dimensions. A robust evaluation method is also proposed that leverages ChatGPT to match the model prediction to given choices. SEED", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bench [22] proposes a benchmark with over 19k questions, covering both image and video modality to evaluate the performance of multi-modal LLMs. MM-Vet [44] defines 16 emergent tasks of interest from 6 core vision-language capabilities, with an LLM-based evaluator, relative strengths and weaknesses of different system paradigms are identified. In this paper, we propose an orthogonal direction for evaluating the reasoning ability of current multi-modal LLMs: counterfactual reasoning.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 498, + 239, + 589, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the construction process of our proposed dataset $C$ -VQA, which consists of two parts: $C$ -VQA-Real and $C$ -VQA-Synthetic. $C$ -VQA-Real contains 3,144 image and question-answer pairs, where each question is not only related to the image content but also comes with a counterfactual presupposition. All images and original questions of $C$ -VQA-Real come from the VQAv2 dataset. $C$ -VQA-Synthetic contains 3,000 image and question-answer pairs, and all images and questions in $C$ -VQA-Synthetic are generated automatically. Each image corresponds to an original question and a counterfactual question. The counterfactual presuppositions in $C$ -VQA enable a new and challenging scenario for VQA models. Both $C$ -VQA-Real and $C$ -VQA-Synthetic contain numerical and boolean questions that can be answered with a number or a boolean value respectively. We illustrate the counterfactual presupposition generation and verification steps below.", + "bbox": [ + 496, + 266, + 893, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Annotation", + "text_level": 1, + "bbox": [ + 498, + 532, + 624, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Counterfactual presupposition type of C-VQA-Real. When designing counterfactual questions, we employ different types of counterfactual presuppositions as follows.", + "bbox": [ + 496, + 555, + 890, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Numerical questions are split into two groups: direct group and indirect group. In the numerical direct group, we add counterfactual presuppositions that change the original answers directly. For example, these questions typically have the form \"How many $X$ would there be if two $X$ were added/removed?\" In the numerical indirect group, the counterfactual presuppositions change the original answers indirectly. It requires more reasoning steps to get the new answers. For example, answering the sentence \"How many cats would be there if the TV was off?\" requires the model to recognize how many cats are images on the TV and understand that when the TV is off, those cats will no longer be there. For boolean questions, the counterfactual presuppositions are often designed to reverse the fact as well as the answer. For example, \"Would the cat be asleep if it was woken up?\" Examples of these counterfactual modifications are presented in Fig. 1.", + "bbox": [ + 496, + 601, + 893, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Question and answer annotation of C-VQA-Real. A two-stage annotate-and-prompt process is employed to cre", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21855", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3ad7688b665063f4ded6e3ea45dce36d24160ad7ebe19baf762edcc085f2a5b2.jpg", + "image_caption": [ + "Table 1. Three instances in $C$ -VQA-Synthetic. They are made from different templates." + ], + "image_footnote": [ + "How many flowers would be inside black polygons if all flowers in green polygons were removed? Select the correct answer: A:5 B:1 C:3 D:7" + ], + "bbox": [ + 101, + 104, + 338, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9f36f199c3274e2f8e798d7eaf4a7828045cf29fddd5af488473463427301c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 114, + 434, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c649122c9dbfd15fcb6dc8fd4a68eefd8dd08cadcfafd48a321a079c4d9aacab.jpg", + "image_caption": [], + "image_footnote": [ + "How many dots would there be in all the circles together if 24 dots were removed from the circles? Select the correct answer: A:7 B:11 C:3 D:9" + ], + "bbox": [ + 361, + 181, + 433, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/554a8b7f6949f73024a513f73bf937252696350ca3bcafdf83485d60c0bd43df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 116, + 516, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/84734a8e0e9311e60afadfffdc8a583d03c94846f24c5d9b35d4b9a7d52aea83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 181, + 516, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c47a3f4d0d669f5d8f382cdfb11de60bd0870138d7bcd10c8c39ed424089e8b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 116, + 598, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9874b6282d1171973d64c21626b94109ffb48164c987a822aca0679798db9b32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 181, + 598, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c160b7db5f23996e3a62c72dcdefdd01986c856a622f9654872e93485d610005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 116, + 694, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a01eceb8678aa5495c96601f60697470efca8deb05a94e2229f025f59a46d8be.jpg", + "image_caption": [], + "image_footnote": [ + "How many dots would a circle contain at most if one of the circles with most dots were removed? Select the correct answer: A:5 B:1 C:3 D:7" + ], + "bbox": [ + 622, + 181, + 694, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/97af5aedccded3dd5e3e696548fa678aed345d68f63e6219f7aaec27ddf315aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 116, + 776, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da45b890211115be0fc62b46163c4e7863646f80f0142390ae04e2313ed0bc78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 181, + 776, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9440e8c34a396bbaf8cf8a1d78316c61fe48fb5e43b6b639e22b78a6db2506ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 116, + 857, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b9bb009fe9cd35674dc7b39e47845630e42c9f776da2fb88461b69760adb208e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 181, + 857, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ate $C$ -VQA. First, we manually annotate 200 questions and answers for each of groups we have defined above. Then, these manually annotated questions and answers are used as in-context-examples [33] to prompt ChatGPT [31] to generate new counterfactual modified questions for the remaining samples. Additionally, to maximize the correctness of ChatGPT-generated questions, we leverage chain-of-thought [38] and divide the whole task into several smaller tasks: read the original question, figure out a proper counterfactual presupposition, figure out how the answer will change, and output the new question and answer. We also provide in-context examples to further ensure the sentence structure and to illustrate various counterfactual presuppositions. For numerical question groups, we first prompt ChatGPT to produce a counterfactual presupposition and then generate a new question with it, and the new answer is also calculated. The annotation flow is plotted in Fig. 2. However, this strategy cannot be directly applied to boolean questions. We notice the randomly generated counterfactual presuppositions often fail to flip the original answer, and thus a large proportion of answers remain unchanged or even cannot be determined. Therefore, we employ a prompting strategy that flips the original answer first and then generates a corresponding counterfactual presupposition.", + "bbox": [ + 76, + 332, + 472, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Verification", + "text_level": 1, + "bbox": [ + 76, + 720, + 204, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we leverage ChatGPT in the process of generating the annotation of $C$ -VQA-Real, we employ a rigorous verification process to ensure the correctness of the generated annotations in order to remove any hallucination or calculation mistakes1. Therefore, we further verify the questions and answers manually to address the errors caused by ChatGPT. Therefore, although the counterfactual questions generated by ChatGPT are idiomatic, they may be improper in the", + "bbox": [ + 76, + 743, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "context of the scene in the image. Thus we further verify the questions and answers manually to address any potential errors. Our verification mainly consists of two stages: (i) image-related verification and (ii) answer-reasonability verification. And each image is at least verified by two people to ensure correctness. We examine the following two key points:", + "bbox": [ + 500, + 332, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(i) Whether the new question is image-related? To make sure that the generated question is indeed asking the model to reason with the scene in the image, we manually examine all generated counterfactual questions and remove all questions that were modified to be not related to the image content.", + "bbox": [ + 498, + 455, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(ii) Whether the new answer is reasonable? Automatically generated answers may be wrong since ChatGPT may make calculation mistakes, reasoning mistakes, etc. We correct these answers manually. Furthermore, some questions may be ambiguous, with no deterministic answers. We remove these questions from the dataset. Since ChatGPT is text-only and cannot see the images, it has to access to the position or color of the objects. Thus, it tends to add counterfactual presuppositions that remove all objects rather than just manipulate a few of the objects. To ensure a rigorous evaluation, we manually annotate all numerical indirect questions. After completing all the annotations, we use ChatGPT again to inspect and polish all the questions again in order to make sure the questions are grammatically correct.", + "bbox": [ + 498, + 565, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Implementation of $C$ -VQA-Synthetic", + "text_level": 1, + "bbox": [ + 500, + 801, + 807, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the procedure we take to generate the $C$ -VQA-Synthetic which contains synthetic images with counterfactual questions and answers. The advantage of $C$ -VQA-Synthetic is that this enables the automatic generation of high-quality counterfactual image-related ques", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "At the time of curating the dataset, GPT-4V API was not publicly available yet [32].", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21856", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3788a415196f53856ce27b875557d8860b582b81c33446ccd6079c5c860deebb.jpg", + "table_caption": [ + "Table 2. Six question templates of $C$ -VQA-Synthetic. Each template has 500 instances with different images and questions. $\\{ \\mathrm{{Color}}\\}$ and $N$ are randomly generated for each instance." + ], + "table_footnote": [], + "table_body": "
TypeQuestion Template
Flower-countingHow many flowers would be outside the {Color} polygons if all polygons were {Color}?
How many flowers would be inside {Color} polygons if we removed N flowers in {Color} polygons?
How many flowers would be inside {Color1} polygons if all flowers in {Color2} polygons were removed?
Dot-countingHow many dots would there be in all the circles together if N dots were removed from the circles?
How many dots would there be in the top three circles together if the two rightmost circles and dots in them were removed from the circles?
How many dots would a circle contain at most if one of the circles with most dots were removed?
", + "bbox": [ + 78, + 132, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tions, saving the time and labor cost of manually annotating. We designed six abstract question types in total and generated 500 instances for each type, resulting in a total of 3,000 synthetic images with counterfactual questions. All images are generated via a predefined procedure in [6] randomly, and we use templates to produce original and counterfactual questions. Both original and counterfactual questions are in a multiple-choice question format. Tab. 1 shows three instances of C-VQA-Synthetic. In the following, we detail the design of our C-VQA-Synthetic by explaining the process of generating two sets of synthetic images and questions: Flower-counting and Dot-counting.", + "bbox": [ + 75, + 354, + 470, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Flower-Counting Puzzles. Three types of questions in $C$ -VQA-Synthetic are based on flower-counting tasks. We first randomly sample two polygons $S_{1}$ and $S_{2}$ in different colors $C_{1}$ and $C_{2}$ . The two polygons divide the whole image into four parts: $\\text{in}_1 \\& \\text{in}_2$ , $\\text{in}_1 \\& \\text{out}_2$ , $\\text{out}_1 \\& \\text{in}_2$ and $\\text{out}_1 \\& \\text{out}_2$ , where $\\text{in}_i$ means inside both $S_{i}$ , and $\\text{out}_i$ means outside $S_{i}$ . Then, we select several flower instances from the Icons-50 dataset [15] and insert them randomly into the four parts. The flowers are guaranteed not to intersect with the edge of polygons to ensure an accurate answer for reasoning questions. The number of flowers in part $p$ is denoted as $\\text{num}(p)$ . With the images procedurally generated, we design questions related to the image contents by filling in question templates. The question templates are shown in Tab. 2. To answer those questions well, a multimodal model not only needs to reason about the spatial relationship, but also needs to perform counterfactual reasoning. For example, the first question template would require the model to reason about the color and the spatial relationship of the image as well as the counterfactual change of color, and the third question template requires the model to reason about the intersection of regions and the counterfactual removal of items.", + "bbox": [ + 75, + 551, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fdcbcc21209dbd3e885121117a76d8ff671c035c239aec1910afc20f4c0290c.jpg", + "image_caption": [ + "Figure 3. Breakdown of answers in numerical groups of $C$ -VQA-Real. We show the percentage of answers in the numerical direct group and numerical indirect group. The share of 0, 1, and 2 in the indirect group are higher while the others are lower." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 694, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/beffa157f2bb7b133268a1be90178bedda6f0521cb89642a1d92294e72e29636.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 883, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dot-Counting Puzzles. The other three types of questions are based on dot-counting tasks. We arrange six circles in the images and then insert dots into these circles randomly. The questions are about the number of dots in the circles. For example, the first template is \"How many dots are there in all the circles together?\" for the original question and \"How many dots would there be in all the circles together if $N$ dots were removed from the circles?\" for the counterfactual question. Suppose the original answer is $\\text{ans}_1$ , we can calculate the counterfactual answer $\\text{ans}_2$ by the equation $\\text{ans}_2 = \\text{ans}_1 - N$ . For the third type of counterfactual question, \"How many dots would a circle contain at most if one of the circles with most dots were removed?\", in order to fully evaluate the reasoning ability of models, the circle with the most dots is designed to be unique to ensure the counterfactual answer varies from the original one.", + "bbox": [ + 496, + 330, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present all six question templates in Tab. 2. All templates are \"How many\" questions, and we insert random elements in images and templates. The answers are designed to have a wide range so that we can rigorously test the counting and reasoning ability of MLLMs. Additionally, the domain gap between these synthetic images poses a challenge for the current multi-modal models as C-VQA-Synthetic is asking for both domain generalization and counterfactual reasoning, which we argue is an important ability for multimodal models.", + "bbox": [ + 496, + 575, + 890, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Dataset Statistics", + "text_level": 1, + "bbox": [ + 500, + 739, + 668, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question Type and Length. C-VQA-Real contains 3144 questions in total, with 2014 numerical questions and 1130 boolean questions. In numerical questions, 1150 questions are from the direct group and 864 questions are from the indirect group. All the numerical questions are \"How many\" questions. Most of the questions in the boolean group start with \"Is\" or \"Are\" before they are changed into counterfactual ones. The average length of questions in C-VQA-Real is 13.15 words, much longer than that of original ques", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21857", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/69187a309f90ef54b7ec4688fb0d8025f816715d086c6d898253089721a79700.jpg", + "table_caption": [ + "Table 3. Evaluation results on three subgroups of C-VQA-Real by ViperGPT[9], VisProg[14], BLIP2 [24], InstructBLIP [8], MiniGPT4 [5, 47], LLaVA [26], CogVLM [37], and Qwen-VL [4]. All of them suffer from significant performance drop from counterfactual questions." + ], + "table_footnote": [], + "table_body": "
Model TypeModelNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
Neuro-symbolicVisProg40.339.9 (-0.4)38.416.8 (-21.6)75.429.5 (-45.9)
ViperGPT83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
End-to-endBLIP2 (FlanT5XXL)43.432.3 (-11.1)49.137.5 (-11.6)77.875.1 (-2.7)
InstructBLIP (FlanT5XXL)57.742.1 (-15.6)58.943.8 (-15.1)83.173.9 (-9.2)
InstructBLIP (Vicuna-7B)64.530.1 (-34.4)64.230.2 (-34.0)83.950.4 (-33.5)
InstructBLIP (Vicuna-13B)63.541.7 (-21.8)63.439.7 (-23.7)86.661.8 (-24.8)
MiniGPT-4 (Vicuna-7B)31.219.6 (-11.6)30.619.6 (-11.0)55.841.2 (-14.6)
MiniGPT-v2 (Llama2-Chat-7B)55.025.9 (-29.1)55.229.1 (-26.1)76.846.7 (-30.1)
LLaVA-7B (Vicuna-7B)38.838.8 (-0.0)42.037.5 (-4.5)60.755.9 (-4.8)
LLaVA-13B (Vicuna-13B)31.231.2 (-0.0)38.331.4 (-6.9)67.364.3 (-3.0)
LLaVA-1.5-7B (Vicuna-7B)60.436.5 (-23.9)62.237.4 (-24.8)86.258.5 (-27.7)
LLaVA-1.5-13B (Vicuna-13B)62.041.0 (-21.0)66.441.2 (-25.2)88.060.7 (-27.3)
CogVLM-Chat49.220.3 (-28.9)50.325.9 (-24.4)83.660.0 (-23.6)
Qwen-VL-Chat65.130.7 (-34.4)69.229.4 (-39.8)88.149.3 (-38.8)
", + "bbox": [ + 76, + 132, + 893, + 419 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4d91898e14f5b3ed6baa324c8c1a2389820fd998a0d930ed470ba546396ea54d.jpg", + "table_caption": [ + "Table 4. Evaluation results on C-VQA-Real by four end-to-end models when chain-of-thought prompt \"Let's think step by step:\" is post-pended to the questions." + ], + "table_footnote": [], + "table_body": "
ModelNum. DirectNum. IndirectBoolean
Ori.CounterfactOri.CounterfactOri.Counterfact
InstructBLIP-13B63.234.2 (-29.0)62.333.7 (-28.6)82.851.0 (-31.8)
MiniGPT-v2-7B46.821.4 (-25.4)44.024.5 (-19.5)72.946.3 (-26.6)
LLaVA-1.5-13B62.141.7 (-20.4)64.942.2 (-22.7)87.960.9 (-27.0)
Qwen-VL-Chat65.229.7 (-35.5)68.424.9 (-43.5)87.447.5 (-39.9)
", + "bbox": [ + 78, + 484, + 468, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions (5.88 words) [3]. $C$ -VQA-Synthetic contains 3000 questions with a multiple-choice answer selection setting. The average length of counterfactual questions in $C$ -VQA-Synthetic is 25.33 words, also longer than that of original questions (16.00 words).", + "bbox": [ + 75, + 590, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Answer Statistics. For $C$ -VQA-Real, each answer in the numerical type is an exact number, with no ambiguous answers such as \"a lot\" or \"many\". The distribution of the answers is shown in Fig. 3. Each answer in the boolean type is a single \"yes\" or a single \"no\". The percentage of \"no\" is $55.13\\%$ , while \"yes\" is $44.87\\%$ . For $C$ -VQA-Synthetic, the options are randomly generated, and we control the proportion of each option close to $25\\%$ .", + "bbox": [ + 75, + 690, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 828, + 209, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provide the evaluation of current state-of-the-art multi-modal LLMs on our proposed $C$ -VQA. Our evaluation covers both neuro-symbolic models like", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ViperGPT [9] and VisProg [14] and end-to-end models including LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], InstructBLIP [8], CogVLM [37], and Qwen-VL [4]. Our evaluation results reveal several interesting findings as detailed below. Implementation details are in the appendix.", + "bbox": [ + 496, + 441, + 893, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiments on C-VQA-Real", + "text_level": 1, + "bbox": [ + 498, + 527, + 754, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance of End-to-end Models. As shown in Tab. 3, end-to-end models perform significantly worse with counterfactual questions than the original ones in all three groups. We find that reasoning difficulties have a significant impact on the results. For the numerical direct group, models require only one simple reasoning step to get the answer. As a consequence, the difference between the evaluation accuracy of the original and counterfactual questions is smaller compared to other groups. We further evaluate four end-to-end models on C-VQA-Real with a prompt \"Let's think step by step\" for chain-of-thought reasoning [21, 38]. As shown in Tab. 4, the performance of the models remains constant or even drops with CoT, which indicates that the CoT strategy contributes little to solving our counterfactual questions. This highlights the need for novel prompting techniques for eliciting counterfactual reasoning ability or new paradigms for fine-tuning end-to-end multi-modal models to help with the counterfactual reasoning tasks.", + "bbox": [ + 496, + 549, + 893, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance of Neuro-symbolic Models. Similar to end-to-end models, the performance of neuro-symbolic models on counterfactual questions is worse than that on", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21858", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52ab6941db4939a1799eec3c2aa448b6a187e1fb04f666b1898fe94326240e7b.jpg", + "table_caption": [ + "Table 5. Evaluation on the GPT-4V model using a subset of C-VQA-Real. Note that the results are not directly comparable with other tables." + ], + "table_footnote": [], + "table_body": "
GPT-4VOriginalCounterfactual
Numerical Direct50.414.6 (-35.8)
Numerical Indirect56.840.7 (-16.1)
Boolean65.948.0 (-17.9)
", + "bbox": [ + 78, + 132, + 470, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the original ones. And it still holds that reasoning difficulty influences accuracy. In the evaluation of ViperGPT [9], the accuracy drops $12.4\\%$ when one-step reasoning is required (i.e., 'Numerical direct'), and it drops $48.4\\%$ when multi-step reasoning is needed (i.e., 'Numerical indirect'). We note that the gap of neuro-symbolic models between counterfactual performance and original performance is much larger than that of any end-to-end model on 'Numerical indirect' and 'Boolean' questions. This indicates that despite the fact that neuro-symbolic models like ViperGPT [9] and VisProg [14] directly generate codes, it cannot handle complex counterfactual reasoning tasks.", + "bbox": [ + 75, + 236, + 468, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 1: Performance of neuro-symbolic models are worse than end-to-end models on complex reasoning tasks.", + "bbox": [ + 91, + 431, + 452, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results of Neuro-symbolic Models With Different Code LLMs. Keeping the API tools, in-context prompts, and questions the same, we can test the coding abilities of different code-generation LLMs by replacing the ChatGPT model inside ViperGPT. The evaluation results are shown in Tab. 6. Notably, the counterfactual presupposition causes significant and consistent performance drops across different model families and scales. This observation emphasizes the inherent limitations of current language models in handling more complex reasoning scenarios. Within this general trend, individual model performance varies. GPT-3.5-turbo generally outperforms other models in the original setting but also suffers from substantial drops in the counterfactual scenarios. CodeLlama [35] shows a relatively moderate performance decrease when counterfactuals are introduced, suggesting some resilience but still a noticeable drop. WizardCoder [28] performs least effectively in both original and counterfactual contexts. Our results here highlight an important and urgent need for improvement in enhancing the models' abilities to handle counterfactual reasoning.", + "bbox": [ + 75, + 518, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the above experiments, we can see that no model families, neither end-to-end nor neuro-symbolic, can handle counterfactual questions well. This indicates that $C$ -VQA is challenging and calls for further explorations.", + "bbox": [ + 75, + 837, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 2: No model family can consistently address counterfactual questions.", + "bbox": [ + 514, + 92, + 875, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of GPT-4V. We also provide an additional evaluation of our proposed benchmark with the GPT-4V [32]. Due to the rate limit and the late date when the API becomes available, we only test GPT-4V on a $10\\%$ randomly selected subset of C-VQA-Real, results are presented in Tab. 5. We show that despite the strong performance of GPT-4V at various visual problems [41], our counterfactual questions still pose a challenge to GPT-4V as the model shows an over $10\\%$ drop in performance for all three types of counterfactual questions in our benchmark.", + "bbox": [ + 496, + 145, + 890, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 3: Even strongest models such as GPT-4V cannot solve our benchmark.", + "bbox": [ + 514, + 309, + 875, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experiments on C-VQA-Synthetic", + "text_level": 1, + "bbox": [ + 500, + 361, + 785, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test the synthetic questions in the same way as the manually annotated questions. However, we find that almost no models show normal performance in the synthetic dataset $C$ -VQA-Synthetic. As shown in Tab. 7, the performance of InstructBLIP (Vicuna-7B) is at a random level $25\\%$ and the performance of InstructBLIP (FlanT5XXL) and BLIP2 (FlanT5XXL) is worse than the random level. We further prompt InstructBLIP (Vicuna) to merely count the flowers and it gives random answers, which indicates that it cannot recognize flowers in the synthetic images. Similar to the observation of [49], we notice that the distribution of the options answered by these models is highly biased. For example, BLIP2 (FlanT5XXL) answers $(D)$ to most questions. It is important to emphasize that the prompt for multiple-choice questions must be meticulously designed; failure to do so may result in some models being unable to select any of the provided options. For neuro-symbolic models, they perform better in the reasoning part but fail to answer the synthetic questions because no available APIs can handle the queries to the synthetic data. For example, ViperGPT cannot check whether a flower is in a polygon because no tools are available for querying this. However, it is worth noting that it can often produce correct code even when counterfactual presuppositions are added. These results indicate that current models cannot generalize beyond the training domain of real images, let alone handling counterfactual reasoning in the OOD synthetic domain.", + "bbox": [ + 496, + 383, + 892, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Bias Analysis", + "text_level": 1, + "bbox": [ + 500, + 801, + 638, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As our dataset $C$ -VQA-Real is based on VQAv2 and the images were from the COCO dataset, we further study the bias of the COCO dataset in the counterfactual questions. Prior work [46] studies bias propagation pathways within image captioning on the COCO dataset by annotating the", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21859", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a81ed5de577cf0ff139924e100a0a1b5ba62d9ff8165b1733d314ef6595f751.jpg", + "table_caption": [ + "Table 6. Evaluation results of ViperGPT on $C$ -VQA-Real when combining with different code LLMs." + ], + "table_footnote": [], + "table_body": "
LLM of ViperGPTNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
GPT-3.5-turbo83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
CodeLlama-7B-Instruct78.050.8 (-27.2)72.220.4 (-51.8)70.531.9 (-38.6)
CodeLlama-13B-Instruct78.972.4 (-6.5)73.426.2 (-47.2)74.541.6 (-32.9)
CodeLlama-34B-Instruct78.368.3 (-10.0)73.520.9 (-52.6)65.036.5 (-28.5)
WizardCoder-Python-7B72.859.1 (-13.7)68.618.3 (-50.3)27.729.5 (+1.8)
WizardCoder-Python-13B73.762.3 (-11.4)67.618.8 (-48.8)56.334.4 (-21.9)
WizardCoder-15B78.055.4 (-22.6)72.618.8 (-53.8)57.715.6 (-42.1)
", + "bbox": [ + 80, + 104, + 893, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b0730beb7b2b791bab83025abb0d866aaa070ad15f5f7303bd6bd013904d809c.jpg", + "table_caption": [ + "Table 7. Evaluation results on C-VQA-Synthetic." + ], + "table_footnote": [], + "table_body": "
ModelOriginalCounterfactual
InstructBLIP (Vicuna-7B)26.925.7
InstructBLIP (FlanT5XXL)16.719.9
BLIP2 (FlanT5XXL)17.520.2
", + "bbox": [ + 78, + 291, + 468, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98f0f29ead01254ec6310bcd5fe106db2101536094b6f7284b46acb15e3207c9.jpg", + "image_caption": [ + "Figure 4. Qualitative example of biases in MLLMs. Given similar questions, InstructBLIP (Vicuna-7B) provides correct answers for the male instance but incorrect answers for the female instance." + ], + "image_footnote": [], + "bbox": [ + 78, + 383, + 264, + 556 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74c4dad494fa4e30432b3ce438e709478829f6c2f46e24ffcea48c62f59bef4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 383, + 468, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perceived gender and skin color of 28,315 of the depicted people. In virtue of this dataset, we could get the overlap between our dataset and the selected images and find that 985 images in $C$ -VQA-Real are annotated with gender and skin information. Among these images, we only keep those that contain only males or females, resulting in 717 images eventually (510 male images and 207 female images). We then evaluate the performance of different models by subgroups. The result is shown in Fig. 5.", + "bbox": [ + 75, + 626, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, for each model we obtain the performance of both original and counterfactual questions of different subgroups. We then compute the difference between the accuracy of original and counterfactual questions in each subgroup. Formally, we have the formula diff(male) = male(ori) - male(cf) and diff(female) = female(ori) - female(cf). As shown in Fig. 5, most end-to-end models have a larger diff(female) than diff(male), which indicates that when presenting", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78ebaa052cb12cbff3f6b96fbf4629a139723f43265882bb278915496a47fc99.jpg", + "image_caption": [ + "Figure 5. Performance difference of original and counterfactual questions on the male and female subgroup on $C$ -VQA-Real. We can see that end-to-end models are often biased toward the male subgroup, and neuro-symbolic models are biased toward the female subgroup. The larger the gap between the performance differences, the larger the bias." + ], + "image_footnote": [], + "bbox": [ + 501, + 276, + 890, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "counterfactual presuppositions, they struggle more with females images. Further studies are needed to understand the reasons and improve the fairness in the model reasoning process.", + "bbox": [ + 496, + 561, + 893, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding 4: MLLMs demonstrate systematic bias in answering gender-related counterfactual questions.", + "bbox": [ + 514, + 638, + 877, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 691, + 619, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study the ability of current multi-modal language models to handle counterfactuals - a core cognition ability of human intelligence. We build a novel dataset $C$ -VQA to test this counterfactual reasoning ability, the dataset is designed to have three types of counterfactual questions, with mixed real and synthetic images and questions. Our evaluations on $C$ -VQA reveal several findings, the most significant finding is that no model in the current multi-modal model literature can consistently handle our counterfactual questions. We have released our code and the dataset to help the community move forward on achieving human-level multi-modal intelligence.", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21860", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In CVPR, 2018. 2, 3", + "[2] Anthropic. Claude. https://claude.ai, 2023.3", + "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 2, 6", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1, 2, 6", + "[5] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 2, 6", + "[6] Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit, Kevin A Smith, and Joshua B Tenenbaum. Are deep neural networks smarter than second graders? In CVPR, 2023. 2, 5", + "[7] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv:2110.14168, 2021. 3", + "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. NeurIPS, 2023. 2, 6", + "[9] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In ICCV, 2023. 1, 6, 7, 14", + "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394, 2023. 3", + "[11] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1", + "[12] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. ICLR 2024, 2023. 1", + "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 1, 2, 3", + "[14] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2017. 1, 6, 7", + "[15] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and surface variations. arXiv:1807.01697, 2019. 5" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021. 3", + "[17] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In NeurIPS, 2021. 3", + "[18] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 2, 3", + "[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 1, 2, 3", + "[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023. 1", + "[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 6", + "[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125, 2023. 3", + "[23] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 2", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2, 6", + "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1", + "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 6, 14", + "[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 3", + "[28] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. arXiv:2306.08568, 2023. 7", + "[29] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 1, 2, 3", + "[30] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In CVPR, 2021. 2, 3", + "[31] OpenAI. Gpt-4 technical report. arXiv:2303.08774, 2023. 3, 4", + "[32] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2,4,7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21861", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Dong Qingxiu, Li Lei, Dai Damai, Zheng Ce, Wu Zhiyong, Chang Baobao, Sun Xu, Xu Jingjing, Li Lei, and Sui Zhi-fang. A survey on in-context learning. arXiv:2304.08485, 2023. 4, 11", + "[34] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In NeurIPS, 2015. 2", + "[35] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv:2308.12950, 2023. 7", + "[36] Haoqin Tu, Bingchen Zhao, Chen Wei, and Cihang Xie. Sight beyond text: Multi-modal training enhances llms in truthfulness and ethics. arXiv:2309.07120, 2023. 1", + "[37] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv:2311.03079, 2023. 2, 6", + "[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 4, 6, 11", + "[39] Te-Lin Wu, Zi-Yi Dou, Qingyuan Hu, Yu Hou, Nischal Chandra, Marjorie Freedman, Ralph Weischedel, and Nanyun Peng. ACQUIRED: A dataset for answering counterfactual questions in real-life videos. In Conference on Empirical Methods in Natural Language Processing, 2023. 2", + "[40] Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Akyurek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, and Yoon Kim. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. arXiv:2307.02477, 2023. 3", + "[41] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv:2309.17421, 9, 2023. 7", + "[42] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1", + "[43] Wenhao Yu, Meng Jiang, Peter Clark, and Ashish Sabharwal. Ifqa: A dataset for open-domain question answering under counterfactual presuppositions. arXiv:2305.14010, 2023. 3", + "[44] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 3", + "[45] Bingchen Zhao, Quan Cui, Hao Wu, Osamu Yoshie, and Cheng Yang. Vision learners meet web image-text pairs. arXiv:2301.07088, 2023. 1", + "[46] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 7", + "[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yongshuo Zong, Oisin Mac Aodha, and Timothy Hospedales. Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008, 2023. 1", + "[49] Yongshuo Zong, Tingyang Yu, Bingchen Zhao, Ruchika Chavhan, and Timothy Hospedales. Fool your (vision and) language model with embarrassingly simple permutations. arXiv:2310.01651, 2023. 7" + ], + "bbox": [ + 501, + 92, + 890, + 189 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21862", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_model.json b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93161244522c3f2cd444dcd2178ee890efa4e257 --- /dev/null +++ b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.204, + 0.899, + 0.224 + ], + "angle": 0, + "content": "Letian Zhang\\(^{1}\\) Xiaotong Zhai\\(^{2}\\) Zhongkai Zhao\\(^{3,6}\\) Yongshuo Zong\\(^{4}\\) Xin Wen\\(^{5,6}\\) Bingchen Zhao\\(^{4,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.229, + 0.777, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)Tongji University \\(^{2}\\)University of Warwick \\(^{3}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.248, + 0.74, + 0.265 + ], + "angle": 0, + "content": "\\(^{4}\\)University of Edinburgh \\(^{5}\\)The University of Hong Kong \\(^{6}\\)LunarAI" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.299, + 0.314, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Counterfactual reasoning, a fundamental aspect of human cognition, involves contemplating alternatives to established facts or past events, significantly enhancing our abilities in planning and decision-making. In light of the advancements in current multi-modal large language models, we explore their effectiveness in counterfactual reasoning. To facilitate this investigation, we introduce a novel dataset, C-VQA, specifically designed to examine the counterfactual reasoning capabilities of modern multi-modal large language models. This dataset is constructed by infusing original questions with counterfactual presuppositions, spanning various types such as numerical and boolean queries. It encompasses a mix of real and synthetic data, representing a wide range of difficulty levels. Our thorough evaluations of contemporary vision-language models using this dataset have revealed substantial performance drops, with some models showing up to a \\(40\\%\\) decrease, highlighting a significant gap between current models and human-like vision reasoning capabilities. We hope our dataset will serve as a vital benchmark for evaluating the counterfactual reasoning capabilities of models. Code and dataset are publicly available at https://bzhao.me/C-VQA/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.72, + 0.441, + 0.75 + ], + "angle": 0, + "content": "\"Counterfactuals are the building blocks of moral behavior as well as scientific thought.\"" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.751, + 0.436, + 0.765 + ], + "angle": 0, + "content": "— Judea Pearl, The Book of Why" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Counterfactual ability is a pivotal cognitive function in humans, enabling us to envision alternate realities and outcomes based on different choices or events. This capacity underpins our decision-making, moral reasoning, and problem-solving skills. Recent development of multi-modal large language models (MLLMs) [26, 42, 45, 48] have dramatically improved the capabilities of image recognition, image-based dialogue, and language grounding, etc. [4, 9," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.298, + 0.891, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.656, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Figure 1. Examples of \\( C \\)-VQA(top), and performance comparison of LLaVA-1.5 [25] w/ and w/o counterfactuality (bottom). \\( C \\)-VQA is constructed by adding counterfactual presuppositions to the questions. We observe that state-of-the-art models all exhibited significant performance drops on the counterfactual questions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.732, + 0.894, + 0.794 + ], + "angle": 0, + "content": "11, 12, 14, 20, 24, 26, 36]. These developments raise the possibility of achieving higher levels of artificial intelligence. Consequently, we pose a critical question: are contemporary MLLMs equipped for counterfactual reasoning?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Visual question answering is one of the central tasks to evaluate the ability of vision-language models. Current benchmarks focus on evaluating different aspects of the abilities, such as visual recognition (e.g., VQAv2 [13]), external knowledge (e.g., OKVQA [29]), compositional reasoning (e.g., CLEVR [19]). They only require the model to understand the image contents by grounding the concept in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "21853" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "volved in the language questions to the image content. With the help of large language models, these tasks are relatively easy to solve, as evidenced by the high scores achieved by modern MLLMs [5, 8, 26]. The most relevant benchmarks are CF-VQA [30] and VQA-CP [1]. However, they focus on constructing the counterfactual questions between the training set and the test set, e.g., showing a green banana in the test set while most of the bananas in the training set are yellow. They do not evaluate the counterfactual ability of the MLLMs - the skill to imagine \"what if\" scenarios that differ from what actually happened." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.593 + ], + "angle": 0, + "content": "In light of this, we propose a novel and challenging evaluation scenario where the language model not only needs to query for the correct visual representation of image content but also needs to be able to perform counterfactual reasoning on those representations. This cannot be solved by encoding world knowledge or object spatial relation alone, but also requires the model to further understand and imagine the given scenarios. Specifically, in our proposed benchmark, each visual question is modified by a counterfactual presupposition - making an assumption based on a scenario that did not actually happen, but could have. This requires the MLLMs to understand the visual content and then perform reasoning over the image contents accordingly to answer the questions. For example, in Fig. 1, to answer the question of \"How many cats are in the image?\", the model can directly query the image and count the occurrence of the visual concept \"cat\". However, with the counterfactual presupposition \"if the TV was off\", one of the cats on the screen will not be visible. Thus, it poses a more challenging scenario for the models as the model not only needs to identify the occurrence of the concept \"cat\", but also reason about its state after a counterfactual operation is taken." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.472, + 0.686 + ], + "angle": 0, + "content": "We build our dataset on top of the commonly used VQAv2 dataset [3], where we collect over 3,000 images suitable for asking counterfactual questions. In addition, we also created over 3,000 synthetic images and corresponding questions to further test the out-of-distribution ability of multi-modal large language models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.472, + 0.884 + ], + "angle": 0, + "content": "We evaluate a wide array of contemporery state-of-the-art multi-modal language models on our dataset, such as LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], Instruct-BLIP [8], Qwen-VL [4] and CogVLM [37]. Our experiments show several interesting findings: (1) Neuro-symbolic models perform worse than end-to-end models on complex counterfactual reasoning; (2) No model family can consistently address counterfactual questions. All of the models suffer from a large performance drop from our counterfactual questions; (3) Even the strongest GPT-4V model [32] cannot solve our benchmark; (4) The MLLMs also demonstrate a systematic bias in answering gender-related counterfactual questions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.441, + 0.902 + ], + "angle": 0, + "content": "We summarize the main contributions of this paper:" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.092, + 0.891, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.53 + ], + "angle": 0, + "content": "Figure 2. Our annotation flow for \\( C \\)-VQA-Real. We select images and questions from the VQAv2 dataset [3], and then utilize ChatGPT to add counterfactual presupposition to the questions and get the corresponding answers. All questions and answers are carefully inspected by human annotators." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.557, + 0.892, + 0.617 + ], + "angle": 0, + "content": "1. We propose a novel and challenging dataset \\( C \\)-VQA with both real and synthetic image-question pairs. The questions contain counterfactual presuppositions and are of various difficulties." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.618, + 0.892, + 0.663 + ], + "angle": 0, + "content": "2. We conduct extensive evaluation of current state-of-the-art vision-language models on our dataset. We show that they struggle with counterfactual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.664, + 0.892, + 0.707 + ], + "angle": 0, + "content": "3. We summarize our findings from the experiments on proposed \\(C\\)-VQA, hoping to provide valuable insights for future research in this field." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.557, + 0.892, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.651, + 0.74 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Visual Question Answering. Visual question answering (VQA) aims to evaluate machines' capabilities of visual understanding, visual reasoning, and the application of commonsense knowledge. Several datasets have been proposed for VQA [3, 6, 13, 18, 19, 23, 29, 34, 39]. COCOQA [34] and VQA [3] are the first to propose the task of VQA, and they contain enormous pictures and questions covering daily life objects. Builds on the VQA dataset. VQAv2 [13] presents a more rigorous evaluation for the VQA task by mitigating language biases and shortcuts in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21854" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "the VQA dataset. Further works [18, 19, 29] continue to extend VQA evaluation to different aspects of image understanding. CLEVR dataset [19] devises a pipeline for generating synthetic data for evaluating the compositional reasoning ability of VQA models. GQA dataset [18] leverages the scene graph structure to generate reasoning questions on real-world images to test the compositional reasoning ability. In this work, we build our dataset \\(C\\)-VQA on the top of VQAv2 [13], where the questions and answers are modified by counterfactual presupposition. Different from previous VQA tasks, our proposed counterfactual dataset provides a new challenging scenario for complementary MLLMs. It further examines the models' abilities to parse the scene structure and reason about the observed world after the counterfactual presuppositions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.473, + 0.755 + ], + "angle": 0, + "content": "Evaluation of Reasoning Abilities. The evaluation of generative models has always been a difficult and challenging problem, as well as for large language models. There have been many efforts for the evaluation of the reasoning ability of LLMs in different aspects [1, 7, 16, 17, 30, 40, 43]. For example, GSM8k [7] and MATH [17] evaluate the mathematical reasoning of LLMs, while MMLU [16] aggregates a diverse range of subjects and tasks for evaluation. For counterfactual reasoning, IfQA [43] proposes the first QA dataset that specifically designed to assess the counterfactual reasoning capabilities of language models. In [40], the counterfactual reasoning abilities of state-of-the-art LLMs (GPT-4 [31], Claude [2]) are evaluated under eleven different tasks with counterfactual presuppositions. It is shown that current LLMs cannot reason with counterfactuals reliably. Most related to our work is CF-VQA [30] and VQA-CP [1] which focus on testing the ability of VQA models to answer \"counterfactual\" questions which are defined as having different property distribution in training and testing set, such as training with yellow bananas but test with green ones. In our work, we propose to build the first visual question-answering dataset with counterfactual presuppositions directly added to the question text, this enables the testing of current SOTA MLLMs which leverages internet-scale training data. Our evaluation results show current approaches for MLLMs do not facilitate counterfactual reasoning. Thus, further work should be done to create stronger MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Multi-Modal LLM Benchmarks With the increasing interest in training novel multi-modal LLMs that can perceive multi-modal inputs, various benchmarks for evaluating the performance of MLLMs are proposed. MME [10] measures the performance of multi-modal LLMs using both perception and cognition abilities on 14 subtasks. MM-Bench [27] covers over 20 different ability dimensions. A robust evaluation method is also proposed that leverages ChatGPT to match the model prediction to given choices. SEED" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "Bench [22] proposes a benchmark with over 19k questions, covering both image and video modality to evaluate the performance of multi-modal LLMs. MM-Vet [44] defines 16 emergent tasks of interest from 6 core vision-language capabilities, with an LLM-based evaluator, relative strengths and weaknesses of different system paradigms are identified. In this paper, we propose an orthogonal direction for evaluating the reasoning ability of current multi-modal LLMs: counterfactual reasoning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.241, + 0.59, + 0.258 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.895, + 0.526 + ], + "angle": 0, + "content": "This section presents the construction process of our proposed dataset \\( C \\)-VQA, which consists of two parts: \\( C \\)-VQA-Real and \\( C \\)-VQA-Synthetic. \\( C \\)-VQA-Real contains 3,144 image and question-answer pairs, where each question is not only related to the image content but also comes with a counterfactual presupposition. All images and original questions of \\( C \\)-VQA-Real come from the VQAv2 dataset. \\( C \\)-VQA-Synthetic contains 3,000 image and question-answer pairs, and all images and questions in \\( C \\)-VQA-Synthetic are generated automatically. Each image corresponds to an original question and a counterfactual question. The counterfactual presuppositions in \\( C \\)-VQA enable a new and challenging scenario for VQA models. Both \\( C \\)-VQA-Real and \\( C \\)-VQA-Synthetic contain numerical and boolean questions that can be answered with a number or a boolean value respectively. We illustrate the counterfactual presupposition generation and verification steps below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.534, + 0.625, + 0.548 + ], + "angle": 0, + "content": "3.1. Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Counterfactual presupposition type of C-VQA-Real. When designing counterfactual questions, we employ different types of counterfactual presuppositions as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.895, + 0.86 + ], + "angle": 0, + "content": "Numerical questions are split into two groups: direct group and indirect group. In the numerical direct group, we add counterfactual presuppositions that change the original answers directly. For example, these questions typically have the form \"How many \\( X \\) would there be if two \\( X \\) were added/removed?\" In the numerical indirect group, the counterfactual presuppositions change the original answers indirectly. It requires more reasoning steps to get the new answers. For example, answering the sentence \"How many cats would be there if the TV was off?\" requires the model to recognize how many cats are images on the TV and understand that when the TV is off, those cats will no longer be there. For boolean questions, the counterfactual presuppositions are often designed to reverse the fact as well as the answer. For example, \"Would the cat be asleep if it was woken up?\" Examples of these counterfactual modifications are presented in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Question and answer annotation of C-VQA-Real. A two-stage annotate-and-prompt process is employed to cre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21855" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.089, + 0.75, + 0.104 + ], + "angle": 0, + "content": "Table 1. Three instances in \\( C \\) -VQA-Synthetic. They are made from different templates." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.105, + 0.339, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.1, + 0.251, + 0.348, + 0.305 + ], + "angle": 0, + "content": "How many flowers would be inside black polygons if all flowers in green polygons were removed? Select the correct answer: A:5 B:1 C:3 D:7" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.116, + 0.435, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.182, + 0.434, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.117, + 0.517, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.182, + 0.517, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.117, + 0.599, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.182, + 0.599, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.36, + 0.25, + 0.608, + 0.305 + ], + "angle": 0, + "content": "How many dots would there be in all the circles together if 24 dots were removed from the circles? Select the correct answer: A:7 B:11 C:3 D:9" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.117, + 0.695, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.182, + 0.695, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.117, + 0.777, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.182, + 0.777, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.117, + 0.859, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.182, + 0.859, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.621, + 0.25, + 0.869, + 0.305 + ], + "angle": 0, + "content": "How many dots would a circle contain at most if one of the circles with most dots were removed? Select the correct answer: A:5 B:1 C:3 D:7" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.473, + 0.71 + ], + "angle": 0, + "content": "ate \\(C\\)-VQA. First, we manually annotate 200 questions and answers for each of groups we have defined above. Then, these manually annotated questions and answers are used as in-context-examples [33] to prompt ChatGPT [31] to generate new counterfactual modified questions for the remaining samples. Additionally, to maximize the correctness of ChatGPT-generated questions, we leverage chain-of-thought [38] and divide the whole task into several smaller tasks: read the original question, figure out a proper counterfactual presupposition, figure out how the answer will change, and output the new question and answer. We also provide in-context examples to further ensure the sentence structure and to illustrate various counterfactual presuppositions. For numerical question groups, we first prompt ChatGPT to produce a counterfactual presupposition and then generate a new question with it, and the new answer is also calculated. The annotation flow is plotted in Fig. 2. However, this strategy cannot be directly applied to boolean questions. We notice the randomly generated counterfactual presuppositions often fail to flip the original answer, and thus a large proportion of answers remain unchanged or even cannot be determined. Therefore, we employ a prompting strategy that flips the original answer first and then generates a corresponding counterfactual presupposition." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.205, + 0.735 + ], + "angle": 0, + "content": "3.2. Verification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.744, + 0.47, + 0.866 + ], + "angle": 0, + "content": "As we leverage ChatGPT in the process of generating the annotation of \\( C \\)-VQA-Real, we employ a rigorous verification process to ensure the correctness of the generated annotations in order to remove any hallucination or calculation mistakes1. Therefore, we further verify the questions and answers manually to address the errors caused by ChatGPT. Therefore, although the counterfactual questions generated by ChatGPT are idiomatic, they may be improper in the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.333, + 0.892, + 0.438 + ], + "angle": 0, + "content": "context of the scene in the image. Thus we further verify the questions and answers manually to address any potential errors. Our verification mainly consists of two stages: (i) image-related verification and (ii) answer-reasonability verification. And each image is at least verified by two people to ensure correctness. We examine the following two key points:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.547 + ], + "angle": 0, + "content": "(i) Whether the new question is image-related? To make sure that the generated question is indeed asking the model to reason with the scene in the image, we manually examine all generated counterfactual questions and remove all questions that were modified to be not related to the image content." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.892, + 0.791 + ], + "angle": 0, + "content": "(ii) Whether the new answer is reasonable? Automatically generated answers may be wrong since ChatGPT may make calculation mistakes, reasoning mistakes, etc. We correct these answers manually. Furthermore, some questions may be ambiguous, with no deterministic answers. We remove these questions from the dataset. Since ChatGPT is text-only and cannot see the images, it has to access to the position or color of the objects. Thus, it tends to add counterfactual presuppositions that remove all objects rather than just manipulate a few of the objects. To ensure a rigorous evaluation, we manually annotate all numerical indirect questions. After completing all the annotations, we use ChatGPT again to inspect and polish all the questions again in order to make sure the questions are grammatically correct." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.808, + 0.818 + ], + "angle": 0, + "content": "3.3. Implementation of \\(C\\)-VQA-Synthetic" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the procedure we take to generate the \\(C\\)-VQA-Synthetic which contains synthetic images with counterfactual questions and answers. The advantage of \\(C\\)-VQA-Synthetic is that this enables the automatic generation of high-quality counterfactual image-related ques" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "At the time of curating the dataset, GPT-4V API was not publicly available yet [32]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21856" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 2. Six question templates of \\( C \\) -VQA-Synthetic. Each template has 500 instances with different images and questions. \\( \\{ \\mathrm{{Color}}\\} \\) and \\( N \\) are randomly generated for each instance." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.133, + 0.47, + 0.349 + ], + "angle": 0, + "content": "
TypeQuestion Template
Flower-countingHow many flowers would be outside the {Color} polygons if all polygons were {Color}?
How many flowers would be inside {Color} polygons if we removed N flowers in {Color} polygons?
How many flowers would be inside {Color1} polygons if all flowers in {Color2} polygons were removed?
Dot-countingHow many dots would there be in all the circles together if N dots were removed from the circles?
How many dots would there be in the top three circles together if the two rightmost circles and dots in them were removed from the circles?
How many dots would a circle contain at most if one of the circles with most dots were removed?
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.471, + 0.538 + ], + "angle": 0, + "content": "tions, saving the time and labor cost of manually annotating. We designed six abstract question types in total and generated 500 instances for each type, resulting in a total of 3,000 synthetic images with counterfactual questions. All images are generated via a predefined procedure in [6] randomly, and we use templates to produce original and counterfactual questions. Both original and counterfactual questions are in a multiple-choice question format. Tab. 1 shows three instances of C-VQA-Synthetic. In the following, we detail the design of our C-VQA-Synthetic by explaining the process of generating two sets of synthetic images and questions: Flower-counting and Dot-counting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Flower-Counting Puzzles. Three types of questions in \\( C \\)-VQA-Synthetic are based on flower-counting tasks. We first randomly sample two polygons \\( S_{1} \\) and \\( S_{2} \\) in different colors \\( C_{1} \\) and \\( C_{2} \\). The two polygons divide the whole image into four parts: \\( \\text{in}_1 \\& \\text{in}_2 \\), \\( \\text{in}_1 \\& \\text{out}_2 \\), \\( \\text{out}_1 \\& \\text{in}_2 \\) and \\( \\text{out}_1 \\& \\text{out}_2 \\), where \\( \\text{in}_i \\) means inside both \\( S_{i} \\), and \\( \\text{out}_i \\) means outside \\( S_{i} \\). Then, we select several flower instances from the Icons-50 dataset [15] and insert them randomly into the four parts. The flowers are guaranteed not to intersect with the edge of polygons to ensure an accurate answer for reasoning questions. The number of flowers in part \\( p \\) is denoted as \\( \\text{num}(p) \\). With the images procedurally generated, we design questions related to the image contents by filling in question templates. The question templates are shown in Tab. 2. To answer those questions well, a multimodal model not only needs to reason about the spatial relationship, but also needs to perform counterfactual reasoning. For example, the first question template would require the model to reason about the color and the spatial relationship of the image as well as the counterfactual change of color, and the third question template requires the model to reason about the intersection of regions and the counterfactual removal of items." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.695, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.885, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.247, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Breakdown of answers in numerical groups of \\( C \\)-VQA-Real. We show the percentage of answers in the numerical direct group and numerical indirect group. The share of 0, 1, and 2 in the indirect group are higher while the others are lower." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Dot-Counting Puzzles. The other three types of questions are based on dot-counting tasks. We arrange six circles in the images and then insert dots into these circles randomly. The questions are about the number of dots in the circles. For example, the first template is \"How many dots are there in all the circles together?\" for the original question and \"How many dots would there be in all the circles together if \\( N \\) dots were removed from the circles?\" for the counterfactual question. Suppose the original answer is \\( \\text{ans}_1 \\), we can calculate the counterfactual answer \\( \\text{ans}_2 \\) by the equation \\( \\text{ans}_2 = \\text{ans}_1 - N \\). For the third type of counterfactual question, \"How many dots would a circle contain at most if one of the circles with most dots were removed?\", in order to fully evaluate the reasoning ability of models, the circle with the most dots is designed to be unique to ensure the counterfactual answer varies from the original one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.726 + ], + "angle": 0, + "content": "We present all six question templates in Tab. 2. All templates are \"How many\" questions, and we insert random elements in images and templates. The answers are designed to have a wide range so that we can rigorously test the counting and reasoning ability of MLLMs. Additionally, the domain gap between these synthetic images poses a challenge for the current multi-modal models as C-VQA-Synthetic is asking for both domain generalization and counterfactual reasoning, which we argue is an important ability for multimodal models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.669, + 0.755 + ], + "angle": 0, + "content": "3.4. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Question Type and Length. C-VQA-Real contains 3144 questions in total, with 2014 numerical questions and 1130 boolean questions. In numerical questions, 1150 questions are from the direct group and 864 questions are from the indirect group. All the numerical questions are \"How many\" questions. Most of the questions in the boolean group start with \"Is\" or \"Are\" before they are changed into counterfactual ones. The average length of questions in C-VQA-Real is 13.15 words, much longer than that of original ques" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21857" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 3. Evaluation results on three subgroups of C-VQA-Real by ViperGPT[9], VisProg[14], BLIP2 [24], InstructBLIP [8], MiniGPT4 [5, 47], LLaVA [26], CogVLM [37], and Qwen-VL [4]. All of them suffer from significant performance drop from counterfactual questions." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.133, + 0.895, + 0.42 + ], + "angle": 0, + "content": "
Model TypeModelNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
Neuro-symbolicVisProg40.339.9 (-0.4)38.416.8 (-21.6)75.429.5 (-45.9)
ViperGPT83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
End-to-endBLIP2 (FlanT5XXL)43.432.3 (-11.1)49.137.5 (-11.6)77.875.1 (-2.7)
InstructBLIP (FlanT5XXL)57.742.1 (-15.6)58.943.8 (-15.1)83.173.9 (-9.2)
InstructBLIP (Vicuna-7B)64.530.1 (-34.4)64.230.2 (-34.0)83.950.4 (-33.5)
InstructBLIP (Vicuna-13B)63.541.7 (-21.8)63.439.7 (-23.7)86.661.8 (-24.8)
MiniGPT-4 (Vicuna-7B)31.219.6 (-11.6)30.619.6 (-11.0)55.841.2 (-14.6)
MiniGPT-v2 (Llama2-Chat-7B)55.025.9 (-29.1)55.229.1 (-26.1)76.846.7 (-30.1)
LLaVA-7B (Vicuna-7B)38.838.8 (-0.0)42.037.5 (-4.5)60.755.9 (-4.8)
LLaVA-13B (Vicuna-13B)31.231.2 (-0.0)38.331.4 (-6.9)67.364.3 (-3.0)
LLaVA-1.5-7B (Vicuna-7B)60.436.5 (-23.9)62.237.4 (-24.8)86.258.5 (-27.7)
LLaVA-1.5-13B (Vicuna-13B)62.041.0 (-21.0)66.441.2 (-25.2)88.060.7 (-27.3)
CogVLM-Chat49.220.3 (-28.9)50.325.9 (-24.4)83.660.0 (-23.6)
Qwen-VL-Chat65.130.7 (-34.4)69.229.4 (-39.8)88.149.3 (-38.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Table 4. Evaluation results on C-VQA-Real by four end-to-end models when chain-of-thought prompt \"Let's think step by step:\" is post-pended to the questions." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.485, + 0.47, + 0.584 + ], + "angle": 0, + "content": "
ModelNum. DirectNum. IndirectBoolean
Ori.CounterfactOri.CounterfactOri.Counterfact
InstructBLIP-13B63.234.2 (-29.0)62.333.7 (-28.6)82.851.0 (-31.8)
MiniGPT-v2-7B46.821.4 (-25.4)44.024.5 (-19.5)72.946.3 (-26.6)
LLaVA-1.5-13B62.141.7 (-20.4)64.942.2 (-22.7)87.960.9 (-27.0)
Qwen-VL-Chat65.229.7 (-35.5)68.424.9 (-43.5)87.447.5 (-39.9)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.669 + ], + "angle": 0, + "content": "tions (5.88 words) [3]. \\(C\\)-VQA-Synthetic contains 3000 questions with a multiple-choice answer selection setting. The average length of counterfactual questions in \\(C\\)-VQA-Synthetic is 25.33 words, also longer than that of original questions (16.00 words)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Answer Statistics. For \\( C \\)-VQA-Real, each answer in the numerical type is an exact number, with no ambiguous answers such as \"a lot\" or \"many\". The distribution of the answers is shown in Fig. 3. Each answer in the boolean type is a single \"yes\" or a single \"no\". The percentage of \"no\" is \\( 55.13\\% \\), while \"yes\" is \\( 44.87\\% \\). For \\( C \\)-VQA-Synthetic, the options are randomly generated, and we control the proportion of each option close to \\( 25\\% \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.21, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we provide the evaluation of current state-of-the-art multi-modal LLMs on our proposed \\(C\\)-VQA. Our evaluation covers both neuro-symbolic models like" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.895, + 0.522 + ], + "angle": 0, + "content": "ViperGPT [9] and VisProg [14] and end-to-end models including LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], InstructBLIP [8], CogVLM [37], and Qwen-VL [4]. Our evaluation results reveal several interesting findings as detailed below. Implementation details are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.755, + 0.544 + ], + "angle": 0, + "content": "4.1. Experiments on C-VQA-Real" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Performance of End-to-end Models. As shown in Tab. 3, end-to-end models perform significantly worse with counterfactual questions than the original ones in all three groups. We find that reasoning difficulties have a significant impact on the results. For the numerical direct group, models require only one simple reasoning step to get the answer. As a consequence, the difference between the evaluation accuracy of the original and counterfactual questions is smaller compared to other groups. We further evaluate four end-to-end models on C-VQA-Real with a prompt \"Let's think step by step\" for chain-of-thought reasoning [21, 38]. As shown in Tab. 4, the performance of the models remains constant or even drops with CoT, which indicates that the CoT strategy contributes little to solving our counterfactual questions. This highlights the need for novel prompting techniques for eliciting counterfactual reasoning ability or new paradigms for fine-tuning end-to-end multi-modal models to help with the counterfactual reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Performance of Neuro-symbolic Models. Similar to end-to-end models, the performance of neuro-symbolic models on counterfactual questions is worse than that on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21858" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 5. Evaluation on the GPT-4V model using a subset of C-VQA-Real. Note that the results are not directly comparable with other tables." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.133, + 0.472, + 0.212 + ], + "angle": 0, + "content": "
GPT-4VOriginalCounterfactual
Numerical Direct50.414.6 (-35.8)
Numerical Indirect56.840.7 (-16.1)
Boolean65.948.0 (-17.9)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.47, + 0.419 + ], + "angle": 0, + "content": "the original ones. And it still holds that reasoning difficulty influences accuracy. In the evaluation of ViperGPT [9], the accuracy drops \\(12.4\\%\\) when one-step reasoning is required (i.e., 'Numerical direct'), and it drops \\(48.4\\%\\) when multi-step reasoning is needed (i.e., 'Numerical indirect'). We note that the gap of neuro-symbolic models between counterfactual performance and original performance is much larger than that of any end-to-end model on 'Numerical indirect' and 'Boolean' questions. This indicates that despite the fact that neuro-symbolic models like ViperGPT [9] and VisProg [14] directly generate codes, it cannot handle complex counterfactual reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.433, + 0.454, + 0.48 + ], + "angle": 0, + "content": "Finding 1: Performance of neuro-symbolic models are worse than end-to-end models on complex reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Results of Neuro-symbolic Models With Different Code LLMs. Keeping the API tools, in-context prompts, and questions the same, we can test the coding abilities of different code-generation LLMs by replacing the ChatGPT model inside ViperGPT. The evaluation results are shown in Tab. 6. Notably, the counterfactual presupposition causes significant and consistent performance drops across different model families and scales. This observation emphasizes the inherent limitations of current language models in handling more complex reasoning scenarios. Within this general trend, individual model performance varies. GPT-3.5-turbo generally outperforms other models in the original setting but also suffers from substantial drops in the counterfactual scenarios. CodeLlama [35] shows a relatively moderate performance decrease when counterfactuals are introduced, suggesting some resilience but still a noticeable drop. WizardCoder [28] performs least effectively in both original and counterfactual contexts. Our results here highlight an important and urgent need for improvement in enhancing the models' abilities to handle counterfactual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.47, + 0.899 + ], + "angle": 0, + "content": "From the above experiments, we can see that no model families, neither end-to-end nor neuro-symbolic, can handle counterfactual questions well. This indicates that \\( C \\)-VQA is challenging and calls for further explorations." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.093, + 0.877, + 0.126 + ], + "angle": 0, + "content": "Finding 2: No model family can consistently address counterfactual questions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.146, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Evaluation of GPT-4V. We also provide an additional evaluation of our proposed benchmark with the GPT-4V [32]. Due to the rate limit and the late date when the API becomes available, we only test GPT-4V on a \\(10\\%\\) randomly selected subset of C-VQA-Real, results are presented in Tab. 5. We show that despite the strong performance of GPT-4V at various visual problems [41], our counterfactual questions still pose a challenge to GPT-4V as the model shows an over \\(10\\%\\) drop in performance for all three types of counterfactual questions in our benchmark." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.31, + 0.877, + 0.34 + ], + "angle": 0, + "content": "Finding 3: Even strongest models such as GPT-4V cannot solve our benchmark." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.787, + 0.379 + ], + "angle": 0, + "content": "4.2. Experiments on C-VQA-Synthetic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.793 + ], + "angle": 0, + "content": "We test the synthetic questions in the same way as the manually annotated questions. However, we find that almost no models show normal performance in the synthetic dataset \\( C \\)-VQA-Synthetic. As shown in Tab. 7, the performance of InstructBLIP (Vicuna-7B) is at a random level \\( 25\\% \\) and the performance of InstructBLIP (FlanT5XXL) and BLIP2 (FlanT5XXL) is worse than the random level. We further prompt InstructBLIP (Vicuna) to merely count the flowers and it gives random answers, which indicates that it cannot recognize flowers in the synthetic images. Similar to the observation of [49], we notice that the distribution of the options answered by these models is highly biased. For example, BLIP2 (FlanT5XXL) answers \\( (D) \\) to most questions. It is important to emphasize that the prompt for multiple-choice questions must be meticulously designed; failure to do so may result in some models being unable to select any of the provided options. For neuro-symbolic models, they perform better in the reasoning part but fail to answer the synthetic questions because no available APIs can handle the queries to the synthetic data. For example, ViperGPT cannot check whether a flower is in a polygon because no tools are available for querying this. However, it is worth noting that it can often produce correct code even when counterfactual presuppositions are added. These results indicate that current models cannot generalize beyond the training domain of real images, let alone handling counterfactual reasoning in the OOD synthetic domain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.64, + 0.818 + ], + "angle": 0, + "content": "4.3. Bias Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "As our dataset \\( C \\)-VQA-Real is based on VQAv2 and the images were from the COCO dataset, we further study the bias of the COCO dataset in the counterfactual questions. Prior work [46] studies bias propagation pathways within image captioning on the COCO dataset by annotating the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21859" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.183, + 0.089, + 0.787, + 0.104 + ], + "angle": 0, + "content": "Table 6. Evaluation results of ViperGPT on \\( C \\) -VQA-Real when combining with different code LLMs." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.105, + 0.895, + 0.266 + ], + "angle": 0, + "content": "
LLM of ViperGPTNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
GPT-3.5-turbo83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
CodeLlama-7B-Instruct78.050.8 (-27.2)72.220.4 (-51.8)70.531.9 (-38.6)
CodeLlama-13B-Instruct78.972.4 (-6.5)73.426.2 (-47.2)74.541.6 (-32.9)
CodeLlama-34B-Instruct78.368.3 (-10.0)73.520.9 (-52.6)65.036.5 (-28.5)
WizardCoder-Python-7B72.859.1 (-13.7)68.618.3 (-50.3)27.729.5 (+1.8)
WizardCoder-Python-13B73.762.3 (-11.4)67.618.8 (-48.8)56.334.4 (-21.9)
WizardCoder-15B78.055.4 (-22.6)72.618.8 (-53.8)57.715.6 (-42.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.276, + 0.42, + 0.29 + ], + "angle": 0, + "content": "Table 7. Evaluation results on C-VQA-Synthetic." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.369 + ], + "angle": 0, + "content": "
ModelOriginalCounterfactual
InstructBLIP (Vicuna-7B)26.925.7
InstructBLIP (FlanT5XXL)16.719.9
BLIP2 (FlanT5XXL)17.520.2
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.384, + 0.266, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.385, + 0.47, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.573, + 0.471, + 0.616 + ], + "angle": 0, + "content": "Figure 4. Qualitative example of biases in MLLMs. Given similar questions, InstructBLIP (Vicuna-7B) provides correct answers for the male instance but incorrect answers for the female instance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.764 + ], + "angle": 0, + "content": "perceived gender and skin color of 28,315 of the depicted people. In virtue of this dataset, we could get the overlap between our dataset and the selected images and find that 985 images in \\( C \\)-VQA-Real are annotated with gender and skin information. Among these images, we only keep those that contain only males or females, resulting in 717 images eventually (510 male images and 207 female images). We then evaluate the performance of different models by subgroups. The result is shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Specifically, for each model we obtain the performance of both original and counterfactual questions of different subgroups. We then compute the difference between the accuracy of original and counterfactual questions in each subgroup. Formally, we have the formula diff(male) = male(ori) - male(cf) and diff(female) = female(ori) - female(cf). As shown in Fig. 5, most end-to-end models have a larger diff(female) than diff(male), which indicates that when presenting" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.277, + 0.891, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.469, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Figure 5. Performance difference of original and counterfactual questions on the male and female subgroup on \\( C \\)-VQA-Real. We can see that end-to-end models are often biased toward the male subgroup, and neuro-symbolic models are biased toward the female subgroup. The larger the gap between the performance differences, the larger the bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.894, + 0.625 + ], + "angle": 0, + "content": "counterfactual presuppositions, they struggle more with females images. Further studies are needed to understand the reasons and improve the fairness in the model reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.639, + 0.878, + 0.672 + ], + "angle": 0, + "content": "Finding 4: MLLMs demonstrate systematic bias in answering gender-related counterfactual questions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.62, + 0.709 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we study the ability of current multi-modal language models to handle counterfactuals - a core cognition ability of human intelligence. We build a novel dataset \\( C \\)-VQA to test this counterfactual reasoning ability, the dataset is designed to have three types of counterfactual questions, with mixed real and synthetic images and questions. Our evaluations on \\( C \\)-VQA reveal several findings, the most significant finding is that no model in the current multi-modal model literature can consistently handle our counterfactual questions. We have released our code and the dataset to help the community move forward on achieving human-level multi-modal intelligence." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21860" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In CVPR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.437, + 0.185 + ], + "angle": 0, + "content": "[2] Anthropic. Claude. https://claude.ai, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[5] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[6] Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit, Kevin A Smith, and Joshua B Tenenbaum. Are deep neural networks smarter than second graders? In CVPR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.469, + 0.491 + ], + "angle": 0, + "content": "[7] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv:2110.14168, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. NeurIPS, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[9] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In ICCV, 2023. 1, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[11] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[12] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. ICLR 2024, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.818, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[14] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2017. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and surface variations. arXiv:1807.01697, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[17] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[18] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[23] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.624, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[28] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. arXiv:2306.08568, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[29] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[30] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] OpenAI. Gpt-4 technical report. arXiv:2303.08774, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2,4,7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21861" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Dong Qingxiu, Li Lei, Dai Damai, Zheng Ce, Wu Zhiyong, Chang Baobao, Sun Xu, Xu Jingjing, Li Lei, and Sui Zhi-fang. A survey on in-context learning. arXiv:2304.08485, 2023. 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[34] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[35] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv:2308.12950, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[36] Haoqin Tu, Bingchen Zhao, Chen Wei, and Cihang Xie. Sight beyond text: Multi-modal training enhances llms in truthfulness and ethics. arXiv:2309.07120, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[37] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv:2311.03079, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 4, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[39] Te-Lin Wu, Zi-Yi Dou, Qingyuan Hu, Yu Hou, Nischal Chandra, Marjorie Freedman, Ralph Weischedel, and Nanyun Peng. ACQUIRED: A dataset for answering counterfactual questions in real-life videos. In Conference on Empirical Methods in Natural Language Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[40] Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Akyurek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, and Yoon Kim. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. arXiv:2307.02477, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[41] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv:2309.17421, 9, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[42] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[43] Wenhao Yu, Meng Jiang, Peter Clark, and Ashish Sabharwal. Ifqa: A dataset for open-domain question answering under counterfactual presuppositions. arXiv:2305.14010, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[44] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.761, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[45] Bingchen Zhao, Quan Cui, Hao Wu, Osamu Yoshie, and Cheng Yang. Vision learners meet web image-text pairs. arXiv:2301.07088, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[46] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.847, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[48] Yongshuo Zong, Oisin Mac Aodha, and Timothy Hospedales. Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[49] Yongshuo Zong, Tingyang Yu, Bingchen Zhao, Ruchika Chavhan, and Timothy Hospedales. Fool your (vision and) language model with embarrassingly simple permutations. arXiv:2310.01651, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21862" + } + ] +] \ No newline at end of file diff --git a/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_origin.pdf b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4f1fe3811ea39658a92cb50b1c4127baf367c88f --- /dev/null +++ b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/9f3da693-f7a9-4a65-84ec-f6f22431f6d4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba4d5cc3757d9cb3a317f6b82915eaa9d982c74979ef106f7d00c9d89d76498b +size 1138277 diff --git a/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/full.md b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dbe3cb101ae2880cb261e840eb728491ee9f950f --- /dev/null +++ b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/full.md @@ -0,0 +1,267 @@ +# What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models + +Letian Zhang $^{1}$ Xiaotong Zhai $^{2}$ Zhongkai Zhao $^{3,6}$ Yongshuo Zong $^{4}$ Xin Wen $^{5,6}$ Bingchen Zhao $^{4,6}$ + +$^{1}$ Tongji University $^{2}$ University of Warwick $^{3}$ National University of Singapore + +$^{4}$ University of Edinburgh $^{5}$ The University of Hong Kong $^{6}$ LunarAI + +# Abstract + +Counterfactual reasoning, a fundamental aspect of human cognition, involves contemplating alternatives to established facts or past events, significantly enhancing our abilities in planning and decision-making. In light of the advancements in current multi-modal large language models, we explore their effectiveness in counterfactual reasoning. To facilitate this investigation, we introduce a novel dataset, C-VQA, specifically designed to examine the counterfactual reasoning capabilities of modern multi-modal large language models. This dataset is constructed by infusing original questions with counterfactual presuppositions, spanning various types such as numerical and boolean queries. It encompasses a mix of real and synthetic data, representing a wide range of difficulty levels. Our thorough evaluations of contemporary vision-language models using this dataset have revealed substantial performance drops, with some models showing up to a $40\%$ decrease, highlighting a significant gap between current models and human-like vision reasoning capabilities. We hope our dataset will serve as a vital benchmark for evaluating the counterfactual reasoning capabilities of models. Code and dataset are publicly available at https://bzhao.me/C-VQA/. + +# 1. Introduction + +"Counterfactuals are the building blocks of moral behavior as well as scientific thought." + +— Judea Pearl, The Book of Why + +Counterfactual ability is a pivotal cognitive function in humans, enabling us to envision alternate realities and outcomes based on different choices or events. This capacity underpins our decision-making, moral reasoning, and problem-solving skills. Recent development of multi-modal large language models (MLLMs) [26, 42, 45, 48] have dramatically improved the capabilities of image recognition, image-based dialogue, and language grounding, etc. [4, 9, + +![](images/90f9414a2e20aef1b14dac3483756b636b888c2c3ca58904e03e8453ab307c9d.jpg) +Figure 1. Examples of $C$ -VQA(top), and performance comparison of LLaVA-1.5 [25] w/ and w/o counterfactuality (bottom). $C$ -VQA is constructed by adding counterfactual presuppositions to the questions. We observe that state-of-the-art models all exhibited significant performance drops on the counterfactual questions. + +11, 12, 14, 20, 24, 26, 36]. These developments raise the possibility of achieving higher levels of artificial intelligence. Consequently, we pose a critical question: are contemporary MLLMs equipped for counterfactual reasoning? + +Visual question answering is one of the central tasks to evaluate the ability of vision-language models. Current benchmarks focus on evaluating different aspects of the abilities, such as visual recognition (e.g., VQAv2 [13]), external knowledge (e.g., OKVQA [29]), compositional reasoning (e.g., CLEVR [19]). They only require the model to understand the image contents by grounding the concept in + +volved in the language questions to the image content. With the help of large language models, these tasks are relatively easy to solve, as evidenced by the high scores achieved by modern MLLMs [5, 8, 26]. The most relevant benchmarks are CF-VQA [30] and VQA-CP [1]. However, they focus on constructing the counterfactual questions between the training set and the test set, e.g., showing a green banana in the test set while most of the bananas in the training set are yellow. They do not evaluate the counterfactual ability of the MLLMs - the skill to imagine "what if" scenarios that differ from what actually happened. + +In light of this, we propose a novel and challenging evaluation scenario where the language model not only needs to query for the correct visual representation of image content but also needs to be able to perform counterfactual reasoning on those representations. This cannot be solved by encoding world knowledge or object spatial relation alone, but also requires the model to further understand and imagine the given scenarios. Specifically, in our proposed benchmark, each visual question is modified by a counterfactual presupposition - making an assumption based on a scenario that did not actually happen, but could have. This requires the MLLMs to understand the visual content and then perform reasoning over the image contents accordingly to answer the questions. For example, in Fig. 1, to answer the question of "How many cats are in the image?", the model can directly query the image and count the occurrence of the visual concept "cat". However, with the counterfactual presupposition "if the TV was off", one of the cats on the screen will not be visible. Thus, it poses a more challenging scenario for the models as the model not only needs to identify the occurrence of the concept "cat", but also reason about its state after a counterfactual operation is taken. + +We build our dataset on top of the commonly used VQAv2 dataset [3], where we collect over 3,000 images suitable for asking counterfactual questions. In addition, we also created over 3,000 synthetic images and corresponding questions to further test the out-of-distribution ability of multi-modal large language models. + +We evaluate a wide array of contemporery state-of-the-art multi-modal language models on our dataset, such as LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], Instruct-BLIP [8], Qwen-VL [4] and CogVLM [37]. Our experiments show several interesting findings: (1) Neuro-symbolic models perform worse than end-to-end models on complex counterfactual reasoning; (2) No model family can consistently address counterfactual questions. All of the models suffer from a large performance drop from our counterfactual questions; (3) Even the strongest GPT-4V model [32] cannot solve our benchmark; (4) The MLLMs also demonstrate a systematic bias in answering gender-related counterfactual questions. + +We summarize the main contributions of this paper: + +![](images/ecf92ba4ce1f7aff1ec5213fabeec088e2fb42ba95188c2386b05d4eb04240f4.jpg) +Figure 2. Our annotation flow for $C$ -VQA-Real. We select images and questions from the VQAv2 dataset [3], and then utilize ChatGPT to add counterfactual presupposition to the questions and get the corresponding answers. All questions and answers are carefully inspected by human annotators. + +1. We propose a novel and challenging dataset $C$ -VQA with both real and synthetic image-question pairs. The questions contain counterfactual presuppositions and are of various difficulties. +2. We conduct extensive evaluation of current state-of-the-art vision-language models on our dataset. We show that they struggle with counterfactual reasoning. +3. We summarize our findings from the experiments on proposed $C$ -VQA, hoping to provide valuable insights for future research in this field. + +# 2. Related Works + +Visual Question Answering. Visual question answering (VQA) aims to evaluate machines' capabilities of visual understanding, visual reasoning, and the application of commonsense knowledge. Several datasets have been proposed for VQA [3, 6, 13, 18, 19, 23, 29, 34, 39]. COCOQA [34] and VQA [3] are the first to propose the task of VQA, and they contain enormous pictures and questions covering daily life objects. Builds on the VQA dataset. VQAv2 [13] presents a more rigorous evaluation for the VQA task by mitigating language biases and shortcuts in + +the VQA dataset. Further works [18, 19, 29] continue to extend VQA evaluation to different aspects of image understanding. CLEVR dataset [19] devises a pipeline for generating synthetic data for evaluating the compositional reasoning ability of VQA models. GQA dataset [18] leverages the scene graph structure to generate reasoning questions on real-world images to test the compositional reasoning ability. In this work, we build our dataset $C$ -VQA on the top of VQAv2 [13], where the questions and answers are modified by counterfactual presupposition. Different from previous VQA tasks, our proposed counterfactual dataset provides a new challenging scenario for complementary MLLMs. It further examines the models' abilities to parse the scene structure and reason about the observed world after the counterfactual presuppositions. + +Evaluation of Reasoning Abilities. The evaluation of generative models has always been a difficult and challenging problem, as well as for large language models. There have been many efforts for the evaluation of the reasoning ability of LLMs in different aspects [1, 7, 16, 17, 30, 40, 43]. For example, GSM8k [7] and MATH [17] evaluate the mathematical reasoning of LLMs, while MMLU [16] aggregates a diverse range of subjects and tasks for evaluation. For counterfactual reasoning, IfQA [43] proposes the first QA dataset that specifically designed to assess the counterfactual reasoning capabilities of language models. In [40], the counterfactual reasoning abilities of state-of-the-art LLMs (GPT-4 [31], Claude [2]) are evaluated under eleven different tasks with counterfactual presuppositions. It is shown that current LLMs cannot reason with counterfactuals reliably. Most related to our work is CF-VQA [30] and VQA-CP [1] which focus on testing the ability of VQA models to answer "counterfactual" questions which are defined as having different property distribution in training and testing set, such as training with yellow bananas but test with green ones. In our work, we propose to build the first visual question-answering dataset with counterfactual presuppositions directly added to the question text, this enables the testing of current SOTA MLLMs which leverages internet-scale training data. Our evaluation results show current approaches for MLLMs do not facilitate counterfactual reasoning. Thus, further work should be done to create stronger MLLMs. + +Multi-Modal LLM Benchmarks With the increasing interest in training novel multi-modal LLMs that can perceive multi-modal inputs, various benchmarks for evaluating the performance of MLLMs are proposed. MME [10] measures the performance of multi-modal LLMs using both perception and cognition abilities on 14 subtasks. MM-Bench [27] covers over 20 different ability dimensions. A robust evaluation method is also proposed that leverages ChatGPT to match the model prediction to given choices. SEED + +Bench [22] proposes a benchmark with over 19k questions, covering both image and video modality to evaluate the performance of multi-modal LLMs. MM-Vet [44] defines 16 emergent tasks of interest from 6 core vision-language capabilities, with an LLM-based evaluator, relative strengths and weaknesses of different system paradigms are identified. In this paper, we propose an orthogonal direction for evaluating the reasoning ability of current multi-modal LLMs: counterfactual reasoning. + +# 3. Dataset + +This section presents the construction process of our proposed dataset $C$ -VQA, which consists of two parts: $C$ -VQA-Real and $C$ -VQA-Synthetic. $C$ -VQA-Real contains 3,144 image and question-answer pairs, where each question is not only related to the image content but also comes with a counterfactual presupposition. All images and original questions of $C$ -VQA-Real come from the VQAv2 dataset. $C$ -VQA-Synthetic contains 3,000 image and question-answer pairs, and all images and questions in $C$ -VQA-Synthetic are generated automatically. Each image corresponds to an original question and a counterfactual question. The counterfactual presuppositions in $C$ -VQA enable a new and challenging scenario for VQA models. Both $C$ -VQA-Real and $C$ -VQA-Synthetic contain numerical and boolean questions that can be answered with a number or a boolean value respectively. We illustrate the counterfactual presupposition generation and verification steps below. + +# 3.1. Annotation + +Counterfactual presupposition type of C-VQA-Real. When designing counterfactual questions, we employ different types of counterfactual presuppositions as follows. + +Numerical questions are split into two groups: direct group and indirect group. In the numerical direct group, we add counterfactual presuppositions that change the original answers directly. For example, these questions typically have the form "How many $X$ would there be if two $X$ were added/removed?" In the numerical indirect group, the counterfactual presuppositions change the original answers indirectly. It requires more reasoning steps to get the new answers. For example, answering the sentence "How many cats would be there if the TV was off?" requires the model to recognize how many cats are images on the TV and understand that when the TV is off, those cats will no longer be there. For boolean questions, the counterfactual presuppositions are often designed to reverse the fact as well as the answer. For example, "Would the cat be asleep if it was woken up?" Examples of these counterfactual modifications are presented in Fig. 1. + +Question and answer annotation of C-VQA-Real. A two-stage annotate-and-prompt process is employed to cre + +Table 1. Three instances in $C$ -VQA-Synthetic. They are made from different templates. +![](images/3ad7688b665063f4ded6e3ea45dce36d24160ad7ebe19baf762edcc085f2a5b2.jpg) +How many flowers would be inside black polygons if all flowers in green polygons were removed? Select the correct answer: A:5 B:1 C:3 D:7 + +![](images/f9f36f199c3274e2f8e798d7eaf4a7828045cf29fddd5af488473463427301c9.jpg) + +![](images/c649122c9dbfd15fcb6dc8fd4a68eefd8dd08cadcfafd48a321a079c4d9aacab.jpg) +How many dots would there be in all the circles together if 24 dots were removed from the circles? Select the correct answer: A:7 B:11 C:3 D:9 + +![](images/554a8b7f6949f73024a513f73bf937252696350ca3bcafdf83485d60c0bd43df.jpg) + +![](images/84734a8e0e9311e60afadfffdc8a583d03c94846f24c5d9b35d4b9a7d52aea83.jpg) + +![](images/c47a3f4d0d669f5d8f382cdfb11de60bd0870138d7bcd10c8c39ed424089e8b0.jpg) + +![](images/9874b6282d1171973d64c21626b94109ffb48164c987a822aca0679798db9b32.jpg) + +![](images/c160b7db5f23996e3a62c72dcdefdd01986c856a622f9654872e93485d610005.jpg) + +![](images/a01eceb8678aa5495c96601f60697470efca8deb05a94e2229f025f59a46d8be.jpg) +How many dots would a circle contain at most if one of the circles with most dots were removed? Select the correct answer: A:5 B:1 C:3 D:7 + +![](images/97af5aedccded3dd5e3e696548fa678aed345d68f63e6219f7aaec27ddf315aa.jpg) + +![](images/da45b890211115be0fc62b46163c4e7863646f80f0142390ae04e2313ed0bc78.jpg) + +![](images/9440e8c34a396bbaf8cf8a1d78316c61fe48fb5e43b6b639e22b78a6db2506ae.jpg) + +![](images/b9bb009fe9cd35674dc7b39e47845630e42c9f776da2fb88461b69760adb208e.jpg) + +ate $C$ -VQA. First, we manually annotate 200 questions and answers for each of groups we have defined above. Then, these manually annotated questions and answers are used as in-context-examples [33] to prompt ChatGPT [31] to generate new counterfactual modified questions for the remaining samples. Additionally, to maximize the correctness of ChatGPT-generated questions, we leverage chain-of-thought [38] and divide the whole task into several smaller tasks: read the original question, figure out a proper counterfactual presupposition, figure out how the answer will change, and output the new question and answer. We also provide in-context examples to further ensure the sentence structure and to illustrate various counterfactual presuppositions. For numerical question groups, we first prompt ChatGPT to produce a counterfactual presupposition and then generate a new question with it, and the new answer is also calculated. The annotation flow is plotted in Fig. 2. However, this strategy cannot be directly applied to boolean questions. We notice the randomly generated counterfactual presuppositions often fail to flip the original answer, and thus a large proportion of answers remain unchanged or even cannot be determined. Therefore, we employ a prompting strategy that flips the original answer first and then generates a corresponding counterfactual presupposition. + +# 3.2. Verification + +As we leverage ChatGPT in the process of generating the annotation of $C$ -VQA-Real, we employ a rigorous verification process to ensure the correctness of the generated annotations in order to remove any hallucination or calculation mistakes1. Therefore, we further verify the questions and answers manually to address the errors caused by ChatGPT. Therefore, although the counterfactual questions generated by ChatGPT are idiomatic, they may be improper in the + +context of the scene in the image. Thus we further verify the questions and answers manually to address any potential errors. Our verification mainly consists of two stages: (i) image-related verification and (ii) answer-reasonability verification. And each image is at least verified by two people to ensure correctness. We examine the following two key points: + +(i) Whether the new question is image-related? To make sure that the generated question is indeed asking the model to reason with the scene in the image, we manually examine all generated counterfactual questions and remove all questions that were modified to be not related to the image content. + +(ii) Whether the new answer is reasonable? Automatically generated answers may be wrong since ChatGPT may make calculation mistakes, reasoning mistakes, etc. We correct these answers manually. Furthermore, some questions may be ambiguous, with no deterministic answers. We remove these questions from the dataset. Since ChatGPT is text-only and cannot see the images, it has to access to the position or color of the objects. Thus, it tends to add counterfactual presuppositions that remove all objects rather than just manipulate a few of the objects. To ensure a rigorous evaluation, we manually annotate all numerical indirect questions. After completing all the annotations, we use ChatGPT again to inspect and polish all the questions again in order to make sure the questions are grammatically correct. + +# 3.3. Implementation of $C$ -VQA-Synthetic + +In this section, we introduce the procedure we take to generate the $C$ -VQA-Synthetic which contains synthetic images with counterfactual questions and answers. The advantage of $C$ -VQA-Synthetic is that this enables the automatic generation of high-quality counterfactual image-related ques + +Table 2. Six question templates of $C$ -VQA-Synthetic. Each template has 500 instances with different images and questions. $\{ \mathrm{{Color}}\}$ and $N$ are randomly generated for each instance. + +
TypeQuestion Template
Flower-countingHow many flowers would be outside the {Color} polygons if all polygons were {Color}?
How many flowers would be inside {Color} polygons if we removed N flowers in {Color} polygons?
How many flowers would be inside {Color1} polygons if all flowers in {Color2} polygons were removed?
Dot-countingHow many dots would there be in all the circles together if N dots were removed from the circles?
How many dots would there be in the top three circles together if the two rightmost circles and dots in them were removed from the circles?
How many dots would a circle contain at most if one of the circles with most dots were removed?
+ +tions, saving the time and labor cost of manually annotating. We designed six abstract question types in total and generated 500 instances for each type, resulting in a total of 3,000 synthetic images with counterfactual questions. All images are generated via a predefined procedure in [6] randomly, and we use templates to produce original and counterfactual questions. Both original and counterfactual questions are in a multiple-choice question format. Tab. 1 shows three instances of C-VQA-Synthetic. In the following, we detail the design of our C-VQA-Synthetic by explaining the process of generating two sets of synthetic images and questions: Flower-counting and Dot-counting. + +Flower-Counting Puzzles. Three types of questions in $C$ -VQA-Synthetic are based on flower-counting tasks. We first randomly sample two polygons $S_{1}$ and $S_{2}$ in different colors $C_{1}$ and $C_{2}$ . The two polygons divide the whole image into four parts: $\text{in}_1 \& \text{in}_2$ , $\text{in}_1 \& \text{out}_2$ , $\text{out}_1 \& \text{in}_2$ and $\text{out}_1 \& \text{out}_2$ , where $\text{in}_i$ means inside both $S_{i}$ , and $\text{out}_i$ means outside $S_{i}$ . Then, we select several flower instances from the Icons-50 dataset [15] and insert them randomly into the four parts. The flowers are guaranteed not to intersect with the edge of polygons to ensure an accurate answer for reasoning questions. The number of flowers in part $p$ is denoted as $\text{num}(p)$ . With the images procedurally generated, we design questions related to the image contents by filling in question templates. The question templates are shown in Tab. 2. To answer those questions well, a multimodal model not only needs to reason about the spatial relationship, but also needs to perform counterfactual reasoning. For example, the first question template would require the model to reason about the color and the spatial relationship of the image as well as the counterfactual change of color, and the third question template requires the model to reason about the intersection of regions and the counterfactual removal of items. + +![](images/5fdcbcc21209dbd3e885121117a76d8ff671c035c239aec1910afc20f4c0290c.jpg) +Figure 3. Breakdown of answers in numerical groups of $C$ -VQA-Real. We show the percentage of answers in the numerical direct group and numerical indirect group. The share of 0, 1, and 2 in the indirect group are higher while the others are lower. + +![](images/beffa157f2bb7b133268a1be90178bedda6f0521cb89642a1d92294e72e29636.jpg) + +Dot-Counting Puzzles. The other three types of questions are based on dot-counting tasks. We arrange six circles in the images and then insert dots into these circles randomly. The questions are about the number of dots in the circles. For example, the first template is "How many dots are there in all the circles together?" for the original question and "How many dots would there be in all the circles together if $N$ dots were removed from the circles?" for the counterfactual question. Suppose the original answer is $\text{ans}_1$ , we can calculate the counterfactual answer $\text{ans}_2$ by the equation $\text{ans}_2 = \text{ans}_1 - N$ . For the third type of counterfactual question, "How many dots would a circle contain at most if one of the circles with most dots were removed?", in order to fully evaluate the reasoning ability of models, the circle with the most dots is designed to be unique to ensure the counterfactual answer varies from the original one. + +We present all six question templates in Tab. 2. All templates are "How many" questions, and we insert random elements in images and templates. The answers are designed to have a wide range so that we can rigorously test the counting and reasoning ability of MLLMs. Additionally, the domain gap between these synthetic images poses a challenge for the current multi-modal models as C-VQA-Synthetic is asking for both domain generalization and counterfactual reasoning, which we argue is an important ability for multimodal models. + +# 3.4. Dataset Statistics + +Question Type and Length. C-VQA-Real contains 3144 questions in total, with 2014 numerical questions and 1130 boolean questions. In numerical questions, 1150 questions are from the direct group and 864 questions are from the indirect group. All the numerical questions are "How many" questions. Most of the questions in the boolean group start with "Is" or "Are" before they are changed into counterfactual ones. The average length of questions in C-VQA-Real is 13.15 words, much longer than that of original ques + +Table 3. Evaluation results on three subgroups of C-VQA-Real by ViperGPT[9], VisProg[14], BLIP2 [24], InstructBLIP [8], MiniGPT4 [5, 47], LLaVA [26], CogVLM [37], and Qwen-VL [4]. All of them suffer from significant performance drop from counterfactual questions. + +
Model TypeModelNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
Neuro-symbolicVisProg40.339.9 (-0.4)38.416.8 (-21.6)75.429.5 (-45.9)
ViperGPT83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
End-to-endBLIP2 (FlanT5XXL)43.432.3 (-11.1)49.137.5 (-11.6)77.875.1 (-2.7)
InstructBLIP (FlanT5XXL)57.742.1 (-15.6)58.943.8 (-15.1)83.173.9 (-9.2)
InstructBLIP (Vicuna-7B)64.530.1 (-34.4)64.230.2 (-34.0)83.950.4 (-33.5)
InstructBLIP (Vicuna-13B)63.541.7 (-21.8)63.439.7 (-23.7)86.661.8 (-24.8)
MiniGPT-4 (Vicuna-7B)31.219.6 (-11.6)30.619.6 (-11.0)55.841.2 (-14.6)
MiniGPT-v2 (Llama2-Chat-7B)55.025.9 (-29.1)55.229.1 (-26.1)76.846.7 (-30.1)
LLaVA-7B (Vicuna-7B)38.838.8 (-0.0)42.037.5 (-4.5)60.755.9 (-4.8)
LLaVA-13B (Vicuna-13B)31.231.2 (-0.0)38.331.4 (-6.9)67.364.3 (-3.0)
LLaVA-1.5-7B (Vicuna-7B)60.436.5 (-23.9)62.237.4 (-24.8)86.258.5 (-27.7)
LLaVA-1.5-13B (Vicuna-13B)62.041.0 (-21.0)66.441.2 (-25.2)88.060.7 (-27.3)
CogVLM-Chat49.220.3 (-28.9)50.325.9 (-24.4)83.660.0 (-23.6)
Qwen-VL-Chat65.130.7 (-34.4)69.229.4 (-39.8)88.149.3 (-38.8)
+ +Table 4. Evaluation results on C-VQA-Real by four end-to-end models when chain-of-thought prompt "Let's think step by step:" is post-pended to the questions. + +
ModelNum. DirectNum. IndirectBoolean
Ori.CounterfactOri.CounterfactOri.Counterfact
InstructBLIP-13B63.234.2 (-29.0)62.333.7 (-28.6)82.851.0 (-31.8)
MiniGPT-v2-7B46.821.4 (-25.4)44.024.5 (-19.5)72.946.3 (-26.6)
LLaVA-1.5-13B62.141.7 (-20.4)64.942.2 (-22.7)87.960.9 (-27.0)
Qwen-VL-Chat65.229.7 (-35.5)68.424.9 (-43.5)87.447.5 (-39.9)
+ +tions (5.88 words) [3]. $C$ -VQA-Synthetic contains 3000 questions with a multiple-choice answer selection setting. The average length of counterfactual questions in $C$ -VQA-Synthetic is 25.33 words, also longer than that of original questions (16.00 words). + +Answer Statistics. For $C$ -VQA-Real, each answer in the numerical type is an exact number, with no ambiguous answers such as "a lot" or "many". The distribution of the answers is shown in Fig. 3. Each answer in the boolean type is a single "yes" or a single "no". The percentage of "no" is $55.13\%$ , while "yes" is $44.87\%$ . For $C$ -VQA-Synthetic, the options are randomly generated, and we control the proportion of each option close to $25\%$ . + +# 4. Experiments + +In this section, we provide the evaluation of current state-of-the-art multi-modal LLMs on our proposed $C$ -VQA. Our evaluation covers both neuro-symbolic models like + +ViperGPT [9] and VisProg [14] and end-to-end models including LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], InstructBLIP [8], CogVLM [37], and Qwen-VL [4]. Our evaluation results reveal several interesting findings as detailed below. Implementation details are in the appendix. + +# 4.1. Experiments on C-VQA-Real + +Performance of End-to-end Models. As shown in Tab. 3, end-to-end models perform significantly worse with counterfactual questions than the original ones in all three groups. We find that reasoning difficulties have a significant impact on the results. For the numerical direct group, models require only one simple reasoning step to get the answer. As a consequence, the difference between the evaluation accuracy of the original and counterfactual questions is smaller compared to other groups. We further evaluate four end-to-end models on C-VQA-Real with a prompt "Let's think step by step" for chain-of-thought reasoning [21, 38]. As shown in Tab. 4, the performance of the models remains constant or even drops with CoT, which indicates that the CoT strategy contributes little to solving our counterfactual questions. This highlights the need for novel prompting techniques for eliciting counterfactual reasoning ability or new paradigms for fine-tuning end-to-end multi-modal models to help with the counterfactual reasoning tasks. + +Performance of Neuro-symbolic Models. Similar to end-to-end models, the performance of neuro-symbolic models on counterfactual questions is worse than that on + +Table 5. Evaluation on the GPT-4V model using a subset of C-VQA-Real. Note that the results are not directly comparable with other tables. + +
GPT-4VOriginalCounterfactual
Numerical Direct50.414.6 (-35.8)
Numerical Indirect56.840.7 (-16.1)
Boolean65.948.0 (-17.9)
+ +the original ones. And it still holds that reasoning difficulty influences accuracy. In the evaluation of ViperGPT [9], the accuracy drops $12.4\%$ when one-step reasoning is required (i.e., 'Numerical direct'), and it drops $48.4\%$ when multi-step reasoning is needed (i.e., 'Numerical indirect'). We note that the gap of neuro-symbolic models between counterfactual performance and original performance is much larger than that of any end-to-end model on 'Numerical indirect' and 'Boolean' questions. This indicates that despite the fact that neuro-symbolic models like ViperGPT [9] and VisProg [14] directly generate codes, it cannot handle complex counterfactual reasoning tasks. + +Finding 1: Performance of neuro-symbolic models are worse than end-to-end models on complex reasoning tasks. + +Results of Neuro-symbolic Models With Different Code LLMs. Keeping the API tools, in-context prompts, and questions the same, we can test the coding abilities of different code-generation LLMs by replacing the ChatGPT model inside ViperGPT. The evaluation results are shown in Tab. 6. Notably, the counterfactual presupposition causes significant and consistent performance drops across different model families and scales. This observation emphasizes the inherent limitations of current language models in handling more complex reasoning scenarios. Within this general trend, individual model performance varies. GPT-3.5-turbo generally outperforms other models in the original setting but also suffers from substantial drops in the counterfactual scenarios. CodeLlama [35] shows a relatively moderate performance decrease when counterfactuals are introduced, suggesting some resilience but still a noticeable drop. WizardCoder [28] performs least effectively in both original and counterfactual contexts. Our results here highlight an important and urgent need for improvement in enhancing the models' abilities to handle counterfactual reasoning. + +From the above experiments, we can see that no model families, neither end-to-end nor neuro-symbolic, can handle counterfactual questions well. This indicates that $C$ -VQA is challenging and calls for further explorations. + +Finding 2: No model family can consistently address counterfactual questions. + +Evaluation of GPT-4V. We also provide an additional evaluation of our proposed benchmark with the GPT-4V [32]. Due to the rate limit and the late date when the API becomes available, we only test GPT-4V on a $10\%$ randomly selected subset of C-VQA-Real, results are presented in Tab. 5. We show that despite the strong performance of GPT-4V at various visual problems [41], our counterfactual questions still pose a challenge to GPT-4V as the model shows an over $10\%$ drop in performance for all three types of counterfactual questions in our benchmark. + +Finding 3: Even strongest models such as GPT-4V cannot solve our benchmark. + +# 4.2. Experiments on C-VQA-Synthetic + +We test the synthetic questions in the same way as the manually annotated questions. However, we find that almost no models show normal performance in the synthetic dataset $C$ -VQA-Synthetic. As shown in Tab. 7, the performance of InstructBLIP (Vicuna-7B) is at a random level $25\%$ and the performance of InstructBLIP (FlanT5XXL) and BLIP2 (FlanT5XXL) is worse than the random level. We further prompt InstructBLIP (Vicuna) to merely count the flowers and it gives random answers, which indicates that it cannot recognize flowers in the synthetic images. Similar to the observation of [49], we notice that the distribution of the options answered by these models is highly biased. For example, BLIP2 (FlanT5XXL) answers $(D)$ to most questions. It is important to emphasize that the prompt for multiple-choice questions must be meticulously designed; failure to do so may result in some models being unable to select any of the provided options. For neuro-symbolic models, they perform better in the reasoning part but fail to answer the synthetic questions because no available APIs can handle the queries to the synthetic data. For example, ViperGPT cannot check whether a flower is in a polygon because no tools are available for querying this. However, it is worth noting that it can often produce correct code even when counterfactual presuppositions are added. These results indicate that current models cannot generalize beyond the training domain of real images, let alone handling counterfactual reasoning in the OOD synthetic domain. + +# 4.3. Bias Analysis + +As our dataset $C$ -VQA-Real is based on VQAv2 and the images were from the COCO dataset, we further study the bias of the COCO dataset in the counterfactual questions. Prior work [46] studies bias propagation pathways within image captioning on the COCO dataset by annotating the + +Table 6. Evaluation results of ViperGPT on $C$ -VQA-Real when combining with different code LLMs. + +
LLM of ViperGPTNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
GPT-3.5-turbo83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
CodeLlama-7B-Instruct78.050.8 (-27.2)72.220.4 (-51.8)70.531.9 (-38.6)
CodeLlama-13B-Instruct78.972.4 (-6.5)73.426.2 (-47.2)74.541.6 (-32.9)
CodeLlama-34B-Instruct78.368.3 (-10.0)73.520.9 (-52.6)65.036.5 (-28.5)
WizardCoder-Python-7B72.859.1 (-13.7)68.618.3 (-50.3)27.729.5 (+1.8)
WizardCoder-Python-13B73.762.3 (-11.4)67.618.8 (-48.8)56.334.4 (-21.9)
WizardCoder-15B78.055.4 (-22.6)72.618.8 (-53.8)57.715.6 (-42.1)
+ +Table 7. Evaluation results on C-VQA-Synthetic. + +
ModelOriginalCounterfactual
InstructBLIP (Vicuna-7B)26.925.7
InstructBLIP (FlanT5XXL)16.719.9
BLIP2 (FlanT5XXL)17.520.2
+ +![](images/98f0f29ead01254ec6310bcd5fe106db2101536094b6f7284b46acb15e3207c9.jpg) +Figure 4. Qualitative example of biases in MLLMs. Given similar questions, InstructBLIP (Vicuna-7B) provides correct answers for the male instance but incorrect answers for the female instance. + +![](images/74c4dad494fa4e30432b3ce438e709478829f6c2f46e24ffcea48c62f59bef4a.jpg) + +perceived gender and skin color of 28,315 of the depicted people. In virtue of this dataset, we could get the overlap between our dataset and the selected images and find that 985 images in $C$ -VQA-Real are annotated with gender and skin information. Among these images, we only keep those that contain only males or females, resulting in 717 images eventually (510 male images and 207 female images). We then evaluate the performance of different models by subgroups. The result is shown in Fig. 5. + +Specifically, for each model we obtain the performance of both original and counterfactual questions of different subgroups. We then compute the difference between the accuracy of original and counterfactual questions in each subgroup. Formally, we have the formula diff(male) = male(ori) - male(cf) and diff(female) = female(ori) - female(cf). As shown in Fig. 5, most end-to-end models have a larger diff(female) than diff(male), which indicates that when presenting + +![](images/78ebaa052cb12cbff3f6b96fbf4629a139723f43265882bb278915496a47fc99.jpg) +Figure 5. Performance difference of original and counterfactual questions on the male and female subgroup on $C$ -VQA-Real. We can see that end-to-end models are often biased toward the male subgroup, and neuro-symbolic models are biased toward the female subgroup. The larger the gap between the performance differences, the larger the bias. + +counterfactual presuppositions, they struggle more with females images. Further studies are needed to understand the reasons and improve the fairness in the model reasoning process. + +Finding 4: MLLMs demonstrate systematic bias in answering gender-related counterfactual questions. + +# 5. Conclusion + +In this paper, we study the ability of current multi-modal language models to handle counterfactuals - a core cognition ability of human intelligence. We build a novel dataset $C$ -VQA to test this counterfactual reasoning ability, the dataset is designed to have three types of counterfactual questions, with mixed real and synthetic images and questions. Our evaluations on $C$ -VQA reveal several findings, the most significant finding is that no model in the current multi-modal model literature can consistently handle our counterfactual questions. We have released our code and the dataset to help the community move forward on achieving human-level multi-modal intelligence. + +# References + +[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In CVPR, 2018. 2, 3 +[2] Anthropic. Claude. https://claude.ai, 2023.3 +[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 2, 6 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1, 2, 6 +[5] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 2, 6 +[6] Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit, Kevin A Smith, and Joshua B Tenenbaum. Are deep neural networks smarter than second graders? In CVPR, 2023. 2, 5 +[7] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv:2110.14168, 2021. 3 +[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. NeurIPS, 2023. 2, 6 +[9] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In ICCV, 2023. 1, 6, 7, 14 +[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394, 2023. 3 +[11] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1 +[12] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. ICLR 2024, 2023. 1 +[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 1, 2, 3 +[14] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2017. 1, 6, 7 +[15] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and surface variations. arXiv:1807.01697, 2019. 5 + +[16] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021. 3 +[17] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In NeurIPS, 2021. 3 +[18] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 2, 3 +[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 1, 2, 3 +[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023. 1 +[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 6 +[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125, 2023. 3 +[23] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 2 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2, 6 +[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 6, 14 +[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 3 +[28] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. arXiv:2306.08568, 2023. 7 +[29] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 1, 2, 3 +[30] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In CVPR, 2021. 2, 3 +[31] OpenAI. Gpt-4 technical report. arXiv:2303.08774, 2023. 3, 4 +[32] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2,4,7 + +[33] Dong Qingxiu, Li Lei, Dai Damai, Zheng Ce, Wu Zhiyong, Chang Baobao, Sun Xu, Xu Jingjing, Li Lei, and Sui Zhi-fang. A survey on in-context learning. arXiv:2304.08485, 2023. 4, 11 +[34] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In NeurIPS, 2015. 2 +[35] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv:2308.12950, 2023. 7 +[36] Haoqin Tu, Bingchen Zhao, Chen Wei, and Cihang Xie. Sight beyond text: Multi-modal training enhances llms in truthfulness and ethics. arXiv:2309.07120, 2023. 1 +[37] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv:2311.03079, 2023. 2, 6 +[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 4, 6, 11 +[39] Te-Lin Wu, Zi-Yi Dou, Qingyuan Hu, Yu Hou, Nischal Chandra, Marjorie Freedman, Ralph Weischedel, and Nanyun Peng. ACQUIRED: A dataset for answering counterfactual questions in real-life videos. In Conference on Empirical Methods in Natural Language Processing, 2023. 2 +[40] Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Akyurek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, and Yoon Kim. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. arXiv:2307.02477, 2023. 3 +[41] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv:2309.17421, 9, 2023. 7 +[42] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1 +[43] Wenhao Yu, Meng Jiang, Peter Clark, and Ashish Sabharwal. Ifqa: A dataset for open-domain question answering under counterfactual presuppositions. arXiv:2305.14010, 2023. 3 +[44] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 3 +[45] Bingchen Zhao, Quan Cui, Hao Wu, Osamu Yoshie, and Cheng Yang. Vision learners meet web image-text pairs. arXiv:2301.07088, 2023. 1 +[46] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 7 +[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2, 6 + +[48] Yongshuo Zong, Oisin Mac Aodha, and Timothy Hospedales. Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008, 2023. 1 +[49] Yongshuo Zong, Tingyang Yu, Bingchen Zhao, Ruchika Chavhan, and Timothy Hospedales. Fool your (vision and) language model with embarrassingly simple permutations. arXiv:2310.01651, 2023. 7 \ No newline at end of file diff --git a/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/images.zip b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2c8a3f514fa7fb6d4a1bc75c3f864572e7101f87 --- /dev/null +++ b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a9dd3b9b01ad55569aa26379d8054f315ac3a03f0614c137c27d0d5a59ff389 +size 544438 diff --git a/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/layout.json b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56861412e884f4b2f66c895b354eba5d0c566771 --- /dev/null +++ b/2024/What If the TV Was Off_ Examining Counterfactual Reasoning Abilities of Multi-modal Language Models/layout.json @@ -0,0 +1,7889 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "text", + "content": "What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "content": "Letian Zhang" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "content": " Xiaotong Zhai" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "content": " Zhongkai Zhao" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "inline_equation", + "content": "^{3,6}" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "content": " Yongshuo Zong" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "content": " Xin Wen" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "inline_equation", + "content": "^{5,6}" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "text", + "content": " Bingchen Zhao" + }, + { + "bbox": [ + 47, + 161, + 550, + 177 + ], + "type": "inline_equation", + "content": "^{4,6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "text", + "content": "Tongji University " + }, + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "text", + "content": "University of Warwick " + }, + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 181, + 475, + 195 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "spans": [ + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "text", + "content": "University of Edinburgh " + }, + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 151, + 196, + 452, + 209 + ], + "type": "text", + "content": "LunarAI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "text", + "content": "Counterfactual reasoning, a fundamental aspect of human cognition, involves contemplating alternatives to established facts or past events, significantly enhancing our abilities in planning and decision-making. In light of the advancements in current multi-modal large language models, we explore their effectiveness in counterfactual reasoning. To facilitate this investigation, we introduce a novel dataset, C-VQA, specifically designed to examine the counterfactual reasoning capabilities of modern multi-modal large language models. This dataset is constructed by infusing original questions with counterfactual presuppositions, spanning various types such as numerical and boolean queries. It encompasses a mix of real and synthetic data, representing a wide range of difficulty levels. Our thorough evaluations of contemporary vision-language models using this dataset have revealed substantial performance drops, with some models showing up to a " + }, + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 262, + 290, + 525 + ], + "type": "text", + "content": " decrease, highlighting a significant gap between current models and human-like vision reasoning capabilities. We hope our dataset will serve as a vital benchmark for evaluating the counterfactual reasoning capabilities of models. Code and dataset are publicly available at https://bzhao.me/C-VQA/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 570, + 269, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 570, + 269, + 594 + ], + "spans": [ + { + "bbox": [ + 74, + 570, + 269, + 594 + ], + "type": "text", + "content": "\"Counterfactuals are the building blocks of moral behavior as well as scientific thought.\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 594, + 266, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 266, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 266, + 605 + ], + "type": "text", + "content": "— Judea Pearl, The Book of Why" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Counterfactual ability is a pivotal cognitive function in humans, enabling us to envision alternate realities and outcomes based on different choices or events. This capacity underpins our decision-making, moral reasoning, and problem-solving skills. Recent development of multi-modal large language models (MLLMs) [26, 42, 45, 48] have dramatically improved the capabilities of image recognition, image-based dialogue, and language grounding, etc. [4, 9," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 236, + 545, + 517 + ], + "blocks": [ + { + "bbox": [ + 310, + 236, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 310, + 236, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 310, + 236, + 545, + 517 + ], + "type": "image", + "image_path": "90f9414a2e20aef1b14dac3483756b636b888c2c3ca58904e03e8453ab307c9d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "lines": [ + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "type": "text", + "content": "Figure 1. Examples of " + }, + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "type": "text", + "content": "-VQA(top), and performance comparison of LLaVA-1.5 [25] w/ and w/o counterfactuality (bottom). " + }, + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 519, + 545, + 575 + ], + "type": "text", + "content": "-VQA is constructed by adding counterfactual presuppositions to the questions. We observe that state-of-the-art models all exhibited significant performance drops on the counterfactual questions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 579, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 579, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 579, + 547, + 628 + ], + "type": "text", + "content": "11, 12, 14, 20, 24, 26, 36]. These developments raise the possibility of achieving higher levels of artificial intelligence. Consequently, we pose a critical question: are contemporary MLLMs equipped for counterfactual reasoning?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Visual question answering is one of the central tasks to evaluate the ability of vision-language models. Current benchmarks focus on evaluating different aspects of the abilities, such as visual recognition (e.g., VQAv2 [13]), external knowledge (e.g., OKVQA [29]), compositional reasoning (e.g., CLEVR [19]). They only require the model to understand the image contents by grounding the concept in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21853" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 289, + 204 + ], + "type": "text", + "content": "volved in the language questions to the image content. With the help of large language models, these tasks are relatively easy to solve, as evidenced by the high scores achieved by modern MLLMs [5, 8, 26]. The most relevant benchmarks are CF-VQA [30] and VQA-CP [1]. However, they focus on constructing the counterfactual questions between the training set and the test set, e.g., showing a green banana in the test set while most of the bananas in the training set are yellow. They do not evaluate the counterfactual ability of the MLLMs - the skill to imagine \"what if\" scenarios that differ from what actually happened." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 469 + ], + "type": "text", + "content": "In light of this, we propose a novel and challenging evaluation scenario where the language model not only needs to query for the correct visual representation of image content but also needs to be able to perform counterfactual reasoning on those representations. This cannot be solved by encoding world knowledge or object spatial relation alone, but also requires the model to further understand and imagine the given scenarios. Specifically, in our proposed benchmark, each visual question is modified by a counterfactual presupposition - making an assumption based on a scenario that did not actually happen, but could have. This requires the MLLMs to understand the visual content and then perform reasoning over the image contents accordingly to answer the questions. For example, in Fig. 1, to answer the question of \"How many cats are in the image?\", the model can directly query the image and count the occurrence of the visual concept \"cat\". However, with the counterfactual presupposition \"if the TV was off\", one of the cats on the screen will not be visible. Thus, it poses a more challenging scenario for the models as the model not only needs to identify the occurrence of the concept \"cat\", but also reason about its state after a counterfactual operation is taken." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 471, + 288, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 288, + 543 + ], + "type": "text", + "content": "We build our dataset on top of the commonly used VQAv2 dataset [3], where we collect over 3,000 images suitable for asking counterfactual questions. In addition, we also created over 3,000 synthetic images and corresponding questions to further test the out-of-distribution ability of multi-modal large language models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 700 + ], + "type": "text", + "content": "We evaluate a wide array of contemporery state-of-the-art multi-modal language models on our dataset, such as LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], Instruct-BLIP [8], Qwen-VL [4] and CogVLM [37]. Our experiments show several interesting findings: (1) Neuro-symbolic models perform worse than end-to-end models on complex counterfactual reasoning; (2) No model family can consistently address counterfactual questions. All of the models suffer from a large performance drop from our counterfactual questions; (3) Even the strongest GPT-4V model [32] cannot solve our benchmark; (4) The MLLMs also demonstrate a systematic bias in answering gender-related counterfactual questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 269, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 269, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 269, + 714 + ], + "type": "text", + "content": "We summarize the main contributions of this paper:" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 361 + ], + "type": "image", + "image_path": "ecf92ba4ce1f7aff1ec5213fabeec088e2fb42ba95188c2386b05d4eb04240f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 363, + 547, + 419 + ], + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 419 + ], + "type": "text", + "content": "Figure 2. Our annotation flow for " + }, + { + "bbox": [ + 304, + 363, + 547, + 419 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 363, + 547, + 419 + ], + "type": "text", + "content": "-VQA-Real. We select images and questions from the VQAv2 dataset [3], and then utilize ChatGPT to add counterfactual presupposition to the questions and get the corresponding answers. All questions and answers are carefully inspected by human annotators." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 441, + 545, + 559 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 441, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 545, + 488 + ], + "type": "text", + "content": "1. We propose a novel and challenging dataset " + }, + { + "bbox": [ + 306, + 441, + 545, + 488 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 441, + 545, + 488 + ], + "type": "text", + "content": "-VQA with both real and synthetic image-question pairs. The questions contain counterfactual presuppositions and are of various difficulties." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 489, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 545, + 525 + ], + "type": "text", + "content": "2. We conduct extensive evaluation of current state-of-the-art vision-language models on our dataset. We show that they struggle with counterfactual reasoning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 525, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 559 + ], + "type": "text", + "content": "3. We summarize our findings from the experiments on proposed " + }, + { + "bbox": [ + 306, + 525, + 545, + 559 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 525, + 545, + 559 + ], + "type": "text", + "content": "-VQA, hoping to provide valuable insights for future research in this field." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 573, + 398, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 398, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 398, + 586 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Visual Question Answering. Visual question answering (VQA) aims to evaluate machines' capabilities of visual understanding, visual reasoning, and the application of commonsense knowledge. Several datasets have been proposed for VQA [3, 6, 13, 18, 19, 23, 29, 34, 39]. COCOQA [34] and VQA [3] are the first to propose the task of VQA, and they contain enormous pictures and questions covering daily life objects. Builds on the VQA dataset. VQAv2 [13] presents a more rigorous evaluation for the VQA task by mitigating language biases and shortcuts in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21854" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "the VQA dataset. Further works [18, 19, 29] continue to extend VQA evaluation to different aspects of image understanding. CLEVR dataset [19] devises a pipeline for generating synthetic data for evaluating the compositional reasoning ability of VQA models. GQA dataset [18] leverages the scene graph structure to generate reasoning questions on real-world images to test the compositional reasoning ability. In this work, we build our dataset " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "-VQA on the top of VQAv2 [13], where the questions and answers are modified by counterfactual presupposition. Different from previous VQA tasks, our proposed counterfactual dataset provides a new challenging scenario for complementary MLLMs. It further examines the models' abilities to parse the scene structure and reason about the observed world after the counterfactual presuppositions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 289, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 289, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 289, + 597 + ], + "type": "text", + "content": "Evaluation of Reasoning Abilities. The evaluation of generative models has always been a difficult and challenging problem, as well as for large language models. There have been many efforts for the evaluation of the reasoning ability of LLMs in different aspects [1, 7, 16, 17, 30, 40, 43]. For example, GSM8k [7] and MATH [17] evaluate the mathematical reasoning of LLMs, while MMLU [16] aggregates a diverse range of subjects and tasks for evaluation. For counterfactual reasoning, IfQA [43] proposes the first QA dataset that specifically designed to assess the counterfactual reasoning capabilities of language models. In [40], the counterfactual reasoning abilities of state-of-the-art LLMs (GPT-4 [31], Claude [2]) are evaluated under eleven different tasks with counterfactual presuppositions. It is shown that current LLMs cannot reason with counterfactuals reliably. Most related to our work is CF-VQA [30] and VQA-CP [1] which focus on testing the ability of VQA models to answer \"counterfactual\" questions which are defined as having different property distribution in training and testing set, such as training with yellow bananas but test with green ones. In our work, we propose to build the first visual question-answering dataset with counterfactual presuppositions directly added to the question text, this enables the testing of current SOTA MLLMs which leverages internet-scale training data. Our evaluation results show current approaches for MLLMs do not facilitate counterfactual reasoning. Thus, further work should be done to create stronger MLLMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Multi-Modal LLM Benchmarks With the increasing interest in training novel multi-modal LLMs that can perceive multi-modal inputs, various benchmarks for evaluating the performance of MLLMs are proposed. MME [10] measures the performance of multi-modal LLMs using both perception and cognition abilities on 14 subtasks. MM-Bench [27] covers over 20 different ability dimensions. A robust evaluation method is also proposed that leverages ChatGPT to match the model prediction to given choices. SEED" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "Bench [22] proposes a benchmark with over 19k questions, covering both image and video modality to evaluate the performance of multi-modal LLMs. MM-Vet [44] defines 16 emergent tasks of interest from 6 core vision-language capabilities, with an LLM-based evaluator, relative strengths and weaknesses of different system paradigms are identified. In this paper, we propose an orthogonal direction for evaluating the reasoning ability of current multi-modal LLMs: counterfactual reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 190, + 361, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 361, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 361, + 204 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "This section presents the construction process of our proposed dataset " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA, which consists of two parts: " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Real and " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Synthetic. " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Real contains 3,144 image and question-answer pairs, where each question is not only related to the image content but also comes with a counterfactual presupposition. All images and original questions of " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Real come from the VQAv2 dataset. " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Synthetic contains 3,000 image and question-answer pairs, and all images and questions in " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Synthetic are generated automatically. Each image corresponds to an original question and a counterfactual question. The counterfactual presuppositions in " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA enable a new and challenging scenario for VQA models. Both " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Real and " + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 211, + 547, + 416 + ], + "type": "text", + "content": "-VQA-Synthetic contain numerical and boolean questions that can be answered with a number or a boolean value respectively. We illustrate the counterfactual presupposition generation and verification steps below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 422, + 382, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 382, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 382, + 434 + ], + "type": "text", + "content": "3.1. Annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 440, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 476 + ], + "type": "text", + "content": "Counterfactual presupposition type of C-VQA-Real. When designing counterfactual questions, we employ different types of counterfactual presuppositions as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "type": "text", + "content": "Numerical questions are split into two groups: direct group and indirect group. In the numerical direct group, we add counterfactual presuppositions that change the original answers directly. For example, these questions typically have the form \"How many " + }, + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "type": "text", + "content": " would there be if two " + }, + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 476, + 547, + 681 + ], + "type": "text", + "content": " were added/removed?\" In the numerical indirect group, the counterfactual presuppositions change the original answers indirectly. It requires more reasoning steps to get the new answers. For example, answering the sentence \"How many cats would be there if the TV was off?\" requires the model to recognize how many cats are images on the TV and understand that when the TV is off, those cats will no longer be there. For boolean questions, the counterfactual presuppositions are often designed to reverse the fact as well as the answer. For example, \"Would the cat be asleep if it was woken up?\" Examples of these counterfactual modifications are presented in Fig. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Question and answer annotation of C-VQA-Real. A two-stage annotate-and-prompt process is employed to cre" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21855" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 83, + 207, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 459, + 82 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 459, + 82 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 459, + 82 + ], + "type": "text", + "content": "Table 1. Three instances in " + }, + { + "bbox": [ + 134, + 70, + 459, + 82 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 134, + 70, + 459, + 82 + ], + "type": "text", + "content": " -VQA-Synthetic. They are made from different templates." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 83, + 207, + 198 + ], + "lines": [ + { + "bbox": [ + 62, + 83, + 207, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 83, + 207, + 198 + ], + "type": "image", + "image_path": "3ad7688b665063f4ded6e3ea45dce36d24160ad7ebe19baf762edcc085f2a5b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 198, + 212, + 241 + ], + "lines": [ + { + "bbox": [ + 61, + 198, + 212, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 212, + 241 + ], + "type": "text", + "content": "How many flowers would be inside black polygons if all flowers in green polygons were removed? Select the correct answer: A:5 B:1 C:3 D:7" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 221, + 91, + 266, + 135 + ], + "blocks": [ + { + "bbox": [ + 221, + 91, + 266, + 135 + ], + "lines": [ + { + "bbox": [ + 221, + 91, + 266, + 135 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 266, + 135 + ], + "type": "image", + "image_path": "f9f36f199c3274e2f8e798d7eaf4a7828045cf29fddd5af488473463427301c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 144, + 265, + 186 + ], + "blocks": [ + { + "bbox": [ + 221, + 144, + 265, + 186 + ], + "lines": [ + { + "bbox": [ + 221, + 144, + 265, + 186 + ], + "spans": [ + { + "bbox": [ + 221, + 144, + 265, + 186 + ], + "type": "image", + "image_path": "c649122c9dbfd15fcb6dc8fd4a68eefd8dd08cadcfafd48a321a079c4d9aacab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 198, + 372, + 241 + ], + "lines": [ + { + "bbox": [ + 220, + 198, + 372, + 241 + ], + "spans": [ + { + "bbox": [ + 220, + 198, + 372, + 241 + ], + "type": "text", + "content": "How many dots would there be in all the circles together if 24 dots were removed from the circles? Select the correct answer: A:7 B:11 C:3 D:9" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 92, + 316, + 135 + ], + "blocks": [ + { + "bbox": [ + 272, + 92, + 316, + 135 + ], + "lines": [ + { + "bbox": [ + 272, + 92, + 316, + 135 + ], + "spans": [ + { + "bbox": [ + 272, + 92, + 316, + 135 + ], + "type": "image", + "image_path": "554a8b7f6949f73024a513f73bf937252696350ca3bcafdf83485d60c0bd43df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 144, + 316, + 186 + ], + "blocks": [ + { + "bbox": [ + 272, + 144, + 316, + 186 + ], + "lines": [ + { + "bbox": [ + 272, + 144, + 316, + 186 + ], + "spans": [ + { + "bbox": [ + 272, + 144, + 316, + 186 + ], + "type": "image", + "image_path": "84734a8e0e9311e60afadfffdc8a583d03c94846f24c5d9b35d4b9a7d52aea83.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 92, + 366, + 135 + ], + "blocks": [ + { + "bbox": [ + 322, + 92, + 366, + 135 + ], + "lines": [ + { + "bbox": [ + 322, + 92, + 366, + 135 + ], + "spans": [ + { + "bbox": [ + 322, + 92, + 366, + 135 + ], + "type": "image", + "image_path": "c47a3f4d0d669f5d8f382cdfb11de60bd0870138d7bcd10c8c39ed424089e8b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 144, + 366, + 186 + ], + "blocks": [ + { + "bbox": [ + 323, + 144, + 366, + 186 + ], + "lines": [ + { + "bbox": [ + 323, + 144, + 366, + 186 + ], + "spans": [ + { + "bbox": [ + 323, + 144, + 366, + 186 + ], + "type": "image", + "image_path": "9874b6282d1171973d64c21626b94109ffb48164c987a822aca0679798db9b32.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 380, + 92, + 425, + 135 + ], + "blocks": [ + { + "bbox": [ + 380, + 92, + 425, + 135 + ], + "lines": [ + { + "bbox": [ + 380, + 92, + 425, + 135 + ], + "spans": [ + { + "bbox": [ + 380, + 92, + 425, + 135 + ], + "type": "image", + "image_path": "c160b7db5f23996e3a62c72dcdefdd01986c856a622f9654872e93485d610005.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 381, + 144, + 425, + 186 + ], + "blocks": [ + { + "bbox": [ + 381, + 144, + 425, + 186 + ], + "lines": [ + { + "bbox": [ + 381, + 144, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 381, + 144, + 425, + 186 + ], + "type": "image", + "image_path": "a01eceb8678aa5495c96601f60697470efca8deb05a94e2229f025f59a46d8be.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 198, + 531, + 241 + ], + "lines": [ + { + "bbox": [ + 380, + 198, + 531, + 241 + ], + "spans": [ + { + "bbox": [ + 380, + 198, + 531, + 241 + ], + "type": "text", + "content": "How many dots would a circle contain at most if one of the circles with most dots were removed? Select the correct answer: A:5 B:1 C:3 D:7" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 432, + 92, + 475, + 135 + ], + "blocks": [ + { + "bbox": [ + 432, + 92, + 475, + 135 + ], + "lines": [ + { + "bbox": [ + 432, + 92, + 475, + 135 + ], + "spans": [ + { + "bbox": [ + 432, + 92, + 475, + 135 + ], + "type": "image", + "image_path": "97af5aedccded3dd5e3e696548fa678aed345d68f63e6219f7aaec27ddf315aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 144, + 475, + 186 + ], + "blocks": [ + { + "bbox": [ + 432, + 144, + 475, + 186 + ], + "lines": [ + { + "bbox": [ + 432, + 144, + 475, + 186 + ], + "spans": [ + { + "bbox": [ + 432, + 144, + 475, + 186 + ], + "type": "image", + "image_path": "da45b890211115be0fc62b46163c4e7863646f80f0142390ae04e2313ed0bc78.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 482, + 92, + 525, + 135 + ], + "blocks": [ + { + "bbox": [ + 482, + 92, + 525, + 135 + ], + "lines": [ + { + "bbox": [ + 482, + 92, + 525, + 135 + ], + "spans": [ + { + "bbox": [ + 482, + 92, + 525, + 135 + ], + "type": "image", + "image_path": "9440e8c34a396bbaf8cf8a1d78316c61fe48fb5e43b6b639e22b78a6db2506ae.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 482, + 144, + 525, + 186 + ], + "blocks": [ + { + "bbox": [ + 482, + 144, + 525, + 186 + ], + "lines": [ + { + "bbox": [ + 482, + 144, + 525, + 186 + ], + "spans": [ + { + "bbox": [ + 482, + 144, + 525, + 186 + ], + "type": "image", + "image_path": "b9bb009fe9cd35674dc7b39e47845630e42c9f776da2fb88461b69760adb208e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 263, + 289, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 289, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 289, + 562 + ], + "type": "text", + "content": "ate " + }, + { + "bbox": [ + 47, + 263, + 289, + 562 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 263, + 289, + 562 + ], + "type": "text", + "content": "-VQA. First, we manually annotate 200 questions and answers for each of groups we have defined above. Then, these manually annotated questions and answers are used as in-context-examples [33] to prompt ChatGPT [31] to generate new counterfactual modified questions for the remaining samples. Additionally, to maximize the correctness of ChatGPT-generated questions, we leverage chain-of-thought [38] and divide the whole task into several smaller tasks: read the original question, figure out a proper counterfactual presupposition, figure out how the answer will change, and output the new question and answer. We also provide in-context examples to further ensure the sentence structure and to illustrate various counterfactual presuppositions. For numerical question groups, we first prompt ChatGPT to produce a counterfactual presupposition and then generate a new question with it, and the new answer is also calculated. The annotation flow is plotted in Fig. 2. However, this strategy cannot be directly applied to boolean questions. We notice the randomly generated counterfactual presuppositions often fail to flip the original answer, and thus a large proportion of answers remain unchanged or even cannot be determined. Therefore, we employ a prompting strategy that flips the original answer first and then generates a corresponding counterfactual presupposition." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 571, + 125, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 125, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 125, + 582 + ], + "type": "text", + "content": "3.2. Verification" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 589, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 287, + 685 + ], + "type": "text", + "content": "As we leverage ChatGPT in the process of generating the annotation of " + }, + { + "bbox": [ + 47, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 589, + 287, + 685 + ], + "type": "text", + "content": "-VQA-Real, we employ a rigorous verification process to ensure the correctness of the generated annotations in order to remove any hallucination or calculation mistakes1. Therefore, we further verify the questions and answers manually to address the errors caused by ChatGPT. Therefore, although the counterfactual questions generated by ChatGPT are idiomatic, they may be improper in the" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 263, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 545, + 346 + ], + "type": "text", + "content": "context of the scene in the image. Thus we further verify the questions and answers manually to address any potential errors. Our verification mainly consists of two stages: (i) image-related verification and (ii) answer-reasonability verification. And each image is at least verified by two people to ensure correctness. We examine the following two key points:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 361, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 433 + ], + "type": "text", + "content": "(i) Whether the new question is image-related? To make sure that the generated question is indeed asking the model to reason with the scene in the image, we manually examine all generated counterfactual questions and remove all questions that were modified to be not related to the image content." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 448, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 626 + ], + "type": "text", + "content": "(ii) Whether the new answer is reasonable? Automatically generated answers may be wrong since ChatGPT may make calculation mistakes, reasoning mistakes, etc. We correct these answers manually. Furthermore, some questions may be ambiguous, with no deterministic answers. We remove these questions from the dataset. Since ChatGPT is text-only and cannot see the images, it has to access to the position or color of the objects. Thus, it tends to add counterfactual presuppositions that remove all objects rather than just manipulate a few of the objects. To ensure a rigorous evaluation, we manually annotate all numerical indirect questions. After completing all the annotations, we use ChatGPT again to inspect and polish all the questions again in order to make sure the questions are grammatically correct." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "text", + "content": "3.3. Implementation of " + }, + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "text", + "content": "-VQA-Synthetic" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this section, we introduce the procedure we take to generate the " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "-VQA-Synthetic which contains synthetic images with counterfactual questions and answers. The advantage of " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "-VQA-Synthetic is that this enables the automatic generation of high-quality counterfactual image-related ques" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "At the time of curating the dataset, GPT-4V API was not publicly available yet [32]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21856" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 105, + 287, + 276 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "text", + "content": "Table 2. Six question templates of " + }, + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "text", + "content": " -VQA-Synthetic. Each template has 500 instances with different images and questions. " + }, + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "inline_equation", + "content": "\\{ \\mathrm{{Color}}\\}" + }, + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "text", + "content": " are randomly generated for each instance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 105, + 287, + 276 + ], + "lines": [ + { + "bbox": [ + 48, + 105, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 105, + 287, + 276 + ], + "type": "table", + "html": "
TypeQuestion Template
Flower-countingHow many flowers would be outside the {Color} polygons if all polygons were {Color}?
How many flowers would be inside {Color} polygons if we removed N flowers in {Color} polygons?
How many flowers would be inside {Color1} polygons if all flowers in {Color2} polygons were removed?
Dot-countingHow many dots would there be in all the circles together if N dots were removed from the circles?
How many dots would there be in the top three circles together if the two rightmost circles and dots in them were removed from the circles?
How many dots would a circle contain at most if one of the circles with most dots were removed?
", + "image_path": "3788a415196f53856ce27b875557d8860b582b81c33446ccd6079c5c860deebb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 281, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 288, + 426 + ], + "type": "text", + "content": "tions, saving the time and labor cost of manually annotating. We designed six abstract question types in total and generated 500 instances for each type, resulting in a total of 3,000 synthetic images with counterfactual questions. All images are generated via a predefined procedure in [6] randomly, and we use templates to produce original and counterfactual questions. Both original and counterfactual questions are in a multiple-choice question format. Tab. 1 shows three instances of C-VQA-Synthetic. In the following, we detail the design of our C-VQA-Synthetic by explaining the process of generating two sets of synthetic images and questions: Flower-counting and Dot-counting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": "Flower-Counting Puzzles. Three types of questions in " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": "-VQA-Synthetic are based on flower-counting tasks. We first randomly sample two polygons " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " in different colors " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_{1}" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_{2}" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ". The two polygons divide the whole image into four parts: " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{in}_1 \\& \\text{in}_2" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{in}_1 \\& \\text{out}_2" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{out}_1 \\& \\text{in}_2" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{out}_1 \\& \\text{out}_2" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{in}_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " means inside both " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{out}_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " means outside " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ". Then, we select several flower instances from the Icons-50 dataset [15] and insert them randomly into the four parts. The flowers are guaranteed not to intersect with the edge of polygons to ensure an accurate answer for reasoning questions. The number of flowers in part " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\text{num}(p)" + }, + { + "bbox": [ + 46, + 437, + 287, + 713 + ], + "type": "text", + "content": ". With the images procedurally generated, we design questions related to the image contents by filling in question templates. The question templates are shown in Tab. 2. To answer those questions well, a multimodal model not only needs to reason about the spatial relationship, but also needs to perform counterfactual reasoning. For example, the first question template would require the model to reason about the color and the spatial relationship of the image as well as the counterfactual change of color, and the third question template requires the model to reason about the intersection of regions and the counterfactual removal of items." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 425, + 194 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 425, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 425, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 425, + 194 + ], + "type": "image", + "image_path": "5fdcbcc21209dbd3e885121117a76d8ff671c035c239aec1910afc20f4c0290c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 195, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 195, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 547, + 240 + ], + "type": "text", + "content": "Figure 3. Breakdown of answers in numerical groups of " + }, + { + "bbox": [ + 304, + 195, + 547, + 240 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 195, + 547, + 240 + ], + "type": "text", + "content": "-VQA-Real. We show the percentage of answers in the numerical direct group and numerical indirect group. The share of 0, 1, and 2 in the indirect group are higher while the others are lower." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 541, + 194 + ], + "type": "image", + "image_path": "beffa157f2bb7b133268a1be90178bedda6f0521cb89642a1d92294e72e29636.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "content": "Dot-Counting Puzzles. The other three types of questions are based on dot-counting tasks. We arrange six circles in the images and then insert dots into these circles randomly. The questions are about the number of dots in the circles. For example, the first template is \"How many dots are there in all the circles together?\" for the original question and \"How many dots would there be in all the circles together if " + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "content": " dots were removed from the circles?\" for the counterfactual question. Suppose the original answer is " + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\text{ans}_1" + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "content": ", we can calculate the counterfactual answer " + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\text{ans}_2" + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "content": " by the equation " + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\text{ans}_2 = \\text{ans}_1 - N" + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "content": ". For the third type of counterfactual question, \"How many dots would a circle contain at most if one of the circles with most dots were removed?\", in order to fully evaluate the reasoning ability of models, the circle with the most dots is designed to be unique to ensure the counterfactual answer varies from the original one." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 456, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 574 + ], + "type": "text", + "content": "We present all six question templates in Tab. 2. All templates are \"How many\" questions, and we insert random elements in images and templates. The answers are designed to have a wide range so that we can rigorously test the counting and reasoning ability of MLLMs. Additionally, the domain gap between these synthetic images poses a challenge for the current multi-modal models as C-VQA-Synthetic is asking for both domain generalization and counterfactual reasoning, which we argue is an important ability for multimodal models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 586, + 409, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 409, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 409, + 597 + ], + "type": "text", + "content": "3.4. Dataset Statistics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Question Type and Length. C-VQA-Real contains 3144 questions in total, with 2014 numerical questions and 1130 boolean questions. In numerical questions, 1150 questions are from the direct group and 864 questions are from the indirect group. All the numerical questions are \"How many\" questions. Most of the questions in the boolean group start with \"Is\" or \"Are\" before they are changed into counterfactual ones. The average length of questions in C-VQA-Real is 13.15 words, much longer than that of original ques" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21857" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 105, + 547, + 332 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 3. Evaluation results on three subgroups of C-VQA-Real by ViperGPT[9], VisProg[14], BLIP2 [24], InstructBLIP [8], MiniGPT4 [5, 47], LLaVA [26], CogVLM [37], and Qwen-VL [4]. All of them suffer from significant performance drop from counterfactual questions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 105, + 547, + 332 + ], + "lines": [ + { + "bbox": [ + 47, + 105, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 547, + 332 + ], + "type": "table", + "html": "
Model TypeModelNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
Neuro-symbolicVisProg40.339.9 (-0.4)38.416.8 (-21.6)75.429.5 (-45.9)
ViperGPT83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
End-to-endBLIP2 (FlanT5XXL)43.432.3 (-11.1)49.137.5 (-11.6)77.875.1 (-2.7)
InstructBLIP (FlanT5XXL)57.742.1 (-15.6)58.943.8 (-15.1)83.173.9 (-9.2)
InstructBLIP (Vicuna-7B)64.530.1 (-34.4)64.230.2 (-34.0)83.950.4 (-33.5)
InstructBLIP (Vicuna-13B)63.541.7 (-21.8)63.439.7 (-23.7)86.661.8 (-24.8)
MiniGPT-4 (Vicuna-7B)31.219.6 (-11.6)30.619.6 (-11.0)55.841.2 (-14.6)
MiniGPT-v2 (Llama2-Chat-7B)55.025.9 (-29.1)55.229.1 (-26.1)76.846.7 (-30.1)
LLaVA-7B (Vicuna-7B)38.838.8 (-0.0)42.037.5 (-4.5)60.755.9 (-4.8)
LLaVA-13B (Vicuna-13B)31.231.2 (-0.0)38.331.4 (-6.9)67.364.3 (-3.0)
LLaVA-1.5-7B (Vicuna-7B)60.436.5 (-23.9)62.237.4 (-24.8)86.258.5 (-27.7)
LLaVA-1.5-13B (Vicuna-13B)62.041.0 (-21.0)66.441.2 (-25.2)88.060.7 (-27.3)
CogVLM-Chat49.220.3 (-28.9)50.325.9 (-24.4)83.660.0 (-23.6)
Qwen-VL-Chat65.130.7 (-34.4)69.229.4 (-39.8)88.149.3 (-38.8)
", + "image_path": "69187a309f90ef54b7ec4688fb0d8025f816715d086c6d898253089721a79700.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 384, + 287, + 462 + ], + "blocks": [ + { + "bbox": [ + 46, + 349, + 287, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 383 + ], + "type": "text", + "content": "Table 4. Evaluation results on C-VQA-Real by four end-to-end models when chain-of-thought prompt \"Let's think step by step:\" is post-pended to the questions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 384, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 48, + 384, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 287, + 462 + ], + "type": "table", + "html": "
ModelNum. DirectNum. IndirectBoolean
Ori.CounterfactOri.CounterfactOri.Counterfact
InstructBLIP-13B63.234.2 (-29.0)62.333.7 (-28.6)82.851.0 (-31.8)
MiniGPT-v2-7B46.821.4 (-25.4)44.024.5 (-19.5)72.946.3 (-26.6)
LLaVA-1.5-13B62.141.7 (-20.4)64.942.2 (-22.7)87.960.9 (-27.0)
Qwen-VL-Chat65.229.7 (-35.5)68.424.9 (-43.5)87.447.5 (-39.9)
", + "image_path": "4d91898e14f5b3ed6baa324c8c1a2389820fd998a0d930ed470ba546396ea54d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "type": "text", + "content": "tions (5.88 words) [3]. " + }, + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "type": "text", + "content": "-VQA-Synthetic contains 3000 questions with a multiple-choice answer selection setting. The average length of counterfactual questions in " + }, + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 468, + 287, + 529 + ], + "type": "text", + "content": "-VQA-Synthetic is 25.33 words, also longer than that of original questions (16.00 words)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "content": "Answer Statistics. For " + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "content": "-VQA-Real, each answer in the numerical type is an exact number, with no ambiguous answers such as \"a lot\" or \"many\". The distribution of the answers is shown in Fig. 3. Each answer in the boolean type is a single \"yes\" or a single \"no\". The percentage of \"no\" is " + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "inline_equation", + "content": "55.13\\%" + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "content": ", while \"yes\" is " + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "inline_equation", + "content": "44.87\\%" + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "content": "-VQA-Synthetic, the options are randomly generated, and we control the proportion of each option close to " + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 547, + 287, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 656, + 128, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 128, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 128, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "In this section, we provide the evaluation of current state-of-the-art multi-modal LLMs on our proposed " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "-VQA. Our evaluation covers both neuro-symbolic models like" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 350, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 413 + ], + "type": "text", + "content": "ViperGPT [9] and VisProg [14] and end-to-end models including LLaVA [26], MiniGPT4 [5, 47], BLIP2 [24], InstructBLIP [8], CogVLM [37], and Qwen-VL [4]. Our evaluation results reveal several interesting findings as detailed below. Implementation details are in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 418, + 462, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 462, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 462, + 430 + ], + "type": "text", + "content": "4.1. Experiments on C-VQA-Real" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 435, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 665 + ], + "type": "text", + "content": "Performance of End-to-end Models. As shown in Tab. 3, end-to-end models perform significantly worse with counterfactual questions than the original ones in all three groups. We find that reasoning difficulties have a significant impact on the results. For the numerical direct group, models require only one simple reasoning step to get the answer. As a consequence, the difference between the evaluation accuracy of the original and counterfactual questions is smaller compared to other groups. We further evaluate four end-to-end models on C-VQA-Real with a prompt \"Let's think step by step\" for chain-of-thought reasoning [21, 38]. As shown in Tab. 4, the performance of the models remains constant or even drops with CoT, which indicates that the CoT strategy contributes little to solving our counterfactual questions. This highlights the need for novel prompting techniques for eliciting counterfactual reasoning ability or new paradigms for fine-tuning end-to-end multi-modal models to help with the counterfactual reasoning tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Performance of Neuro-symbolic Models. Similar to end-to-end models, the performance of neuro-symbolic models on counterfactual questions is worse than that on" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21858" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 105, + 288, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "text", + "content": "Table 5. Evaluation on the GPT-4V model using a subset of C-VQA-Real. Note that the results are not directly comparable with other tables." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 105, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 105, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 105, + 288, + 167 + ], + "type": "table", + "html": "
GPT-4VOriginalCounterfactual
Numerical Direct50.414.6 (-35.8)
Numerical Indirect56.840.7 (-16.1)
Boolean65.948.0 (-17.9)
", + "image_path": "52ab6941db4939a1799eec3c2aa448b6a187e1fb04f666b1898fe94326240e7b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "type": "text", + "content": "the original ones. And it still holds that reasoning difficulty influences accuracy. In the evaluation of ViperGPT [9], the accuracy drops " + }, + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "type": "text", + "content": " when one-step reasoning is required (i.e., 'Numerical direct'), and it drops " + }, + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "type": "inline_equation", + "content": "48.4\\%" + }, + { + "bbox": [ + 46, + 187, + 287, + 331 + ], + "type": "text", + "content": " when multi-step reasoning is needed (i.e., 'Numerical indirect'). We note that the gap of neuro-symbolic models between counterfactual performance and original performance is much larger than that of any end-to-end model on 'Numerical indirect' and 'Boolean' questions. This indicates that despite the fact that neuro-symbolic models like ViperGPT [9] and VisProg [14] directly generate codes, it cannot handle complex counterfactual reasoning tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 342, + 277, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 277, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 277, + 380 + ], + "type": "text", + "content": "Finding 1: Performance of neuro-symbolic models are worse than end-to-end models on complex reasoning tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 662 + ], + "type": "text", + "content": "Results of Neuro-symbolic Models With Different Code LLMs. Keeping the API tools, in-context prompts, and questions the same, we can test the coding abilities of different code-generation LLMs by replacing the ChatGPT model inside ViperGPT. The evaluation results are shown in Tab. 6. Notably, the counterfactual presupposition causes significant and consistent performance drops across different model families and scales. This observation emphasizes the inherent limitations of current language models in handling more complex reasoning scenarios. Within this general trend, individual model performance varies. GPT-3.5-turbo generally outperforms other models in the original setting but also suffers from substantial drops in the counterfactual scenarios. CodeLlama [35] shows a relatively moderate performance decrease when counterfactuals are introduced, suggesting some resilience but still a noticeable drop. WizardCoder [28] performs least effectively in both original and counterfactual contexts. Our results here highlight an important and urgent need for improvement in enhancing the models' abilities to handle counterfactual reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 663, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 287, + 712 + ], + "type": "text", + "content": "From the above experiments, we can see that no model families, neither end-to-end nor neuro-symbolic, can handle counterfactual questions well. This indicates that " + }, + { + "bbox": [ + 46, + 663, + 287, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 663, + 287, + 712 + ], + "type": "text", + "content": "-VQA is challenging and calls for further explorations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 73, + 536, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 536, + 99 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 536, + 99 + ], + "type": "text", + "content": "Finding 2: No model family can consistently address counterfactual questions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "type": "text", + "content": "Evaluation of GPT-4V. We also provide an additional evaluation of our proposed benchmark with the GPT-4V [32]. Due to the rate limit and the late date when the API becomes available, we only test GPT-4V on a " + }, + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "type": "text", + "content": " randomly selected subset of C-VQA-Real, results are presented in Tab. 5. We show that despite the strong performance of GPT-4V at various visual problems [41], our counterfactual questions still pose a challenge to GPT-4V as the model shows an over " + }, + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 115, + 545, + 236 + ], + "type": "text", + "content": " drop in performance for all three types of counterfactual questions in our benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 245, + 536, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 245, + 536, + 269 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 536, + 269 + ], + "type": "text", + "content": "Finding 3: Even strongest models such as GPT-4V cannot solve our benchmark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 286, + 481, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 481, + 300 + ], + "type": "text", + "content": "4.2. Experiments on C-VQA-Synthetic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "text", + "content": "We test the synthetic questions in the same way as the manually annotated questions. However, we find that almost no models show normal performance in the synthetic dataset " + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "text", + "content": "-VQA-Synthetic. As shown in Tab. 7, the performance of InstructBLIP (Vicuna-7B) is at a random level " + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "text", + "content": " and the performance of InstructBLIP (FlanT5XXL) and BLIP2 (FlanT5XXL) is worse than the random level. We further prompt InstructBLIP (Vicuna) to merely count the flowers and it gives random answers, which indicates that it cannot recognize flowers in the synthetic images. Similar to the observation of [49], we notice that the distribution of the options answered by these models is highly biased. For example, BLIP2 (FlanT5XXL) answers " + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 304, + 304, + 546, + 628 + ], + "type": "text", + "content": " to most questions. It is important to emphasize that the prompt for multiple-choice questions must be meticulously designed; failure to do so may result in some models being unable to select any of the provided options. For neuro-symbolic models, they perform better in the reasoning part but fail to answer the synthetic questions because no available APIs can handle the queries to the synthetic data. For example, ViperGPT cannot check whether a flower is in a polygon because no tools are available for querying this. However, it is worth noting that it can often produce correct code even when counterfactual presuppositions are added. These results indicate that current models cannot generalize beyond the training domain of real images, let alone handling counterfactual reasoning in the OOD synthetic domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 391, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 391, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 391, + 647 + ], + "type": "text", + "content": "4.3. Bias Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "As our dataset " + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "-VQA-Real is based on VQAv2 and the images were from the COCO dataset, we further study the bias of the COCO dataset in the counterfactual questions. Prior work [46] studies bias propagation pathways within image captioning on the COCO dataset by annotating the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21859" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 83, + 547, + 210 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 481, + 82 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 481, + 82 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 481, + 82 + ], + "type": "text", + "content": "Table 6. Evaluation results of ViperGPT on " + }, + { + "bbox": [ + 111, + 70, + 481, + 82 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 111, + 70, + 481, + 82 + ], + "type": "text", + "content": " -VQA-Real when combining with different code LLMs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 83, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 83, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 83, + 547, + 210 + ], + "type": "table", + "html": "
LLM of ViperGPTNumerical DirectNumerical IndirectBoolean
OriginalCounterfactualOriginalCounterfactualOriginalCounterfactual
GPT-3.5-turbo83.871.4 (-12.4)78.630.2 (-48.4)95.028.2 (-66.8)
CodeLlama-7B-Instruct78.050.8 (-27.2)72.220.4 (-51.8)70.531.9 (-38.6)
CodeLlama-13B-Instruct78.972.4 (-6.5)73.426.2 (-47.2)74.541.6 (-32.9)
CodeLlama-34B-Instruct78.368.3 (-10.0)73.520.9 (-52.6)65.036.5 (-28.5)
WizardCoder-Python-7B72.859.1 (-13.7)68.618.3 (-50.3)27.729.5 (+1.8)
WizardCoder-Python-13B73.762.3 (-11.4)67.618.8 (-48.8)56.334.4 (-21.9)
WizardCoder-15B78.055.4 (-22.6)72.618.8 (-53.8)57.715.6 (-42.1)
", + "image_path": "3a81ed5de577cf0ff139924e100a0a1b5ba62d9ff8165b1733d314ef6595f751.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 231, + 287, + 292 + ], + "blocks": [ + { + "bbox": [ + 77, + 218, + 257, + 229 + ], + "lines": [ + { + "bbox": [ + 77, + 218, + 257, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 218, + 257, + 229 + ], + "type": "text", + "content": "Table 7. Evaluation results on C-VQA-Synthetic." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 231, + 287, + 292 + ], + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 292 + ], + "type": "table", + "html": "
ModelOriginalCounterfactual
InstructBLIP (Vicuna-7B)26.925.7
InstructBLIP (FlanT5XXL)16.719.9
BLIP2 (FlanT5XXL)17.520.2
", + "image_path": "b0730beb7b2b791bab83025abb0d866aaa070ad15f5f7303bd6bd013904d809c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 304, + 162, + 441 + ], + "blocks": [ + { + "bbox": [ + 48, + 304, + 162, + 441 + ], + "lines": [ + { + "bbox": [ + 48, + 304, + 162, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 162, + 441 + ], + "type": "image", + "image_path": "98f0f29ead01254ec6310bcd5fe106db2101536094b6f7284b46acb15e3207c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 453, + 288, + 487 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 487 + ], + "type": "text", + "content": "Figure 4. Qualitative example of biases in MLLMs. Given similar questions, InstructBLIP (Vicuna-7B) provides correct answers for the male instance but incorrect answers for the female instance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 304, + 287, + 441 + ], + "blocks": [ + { + "bbox": [ + 173, + 304, + 287, + 441 + ], + "lines": [ + { + "bbox": [ + 173, + 304, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 173, + 304, + 287, + 441 + ], + "type": "image", + "image_path": "74c4dad494fa4e30432b3ce438e709478829f6c2f46e24ffcea48c62f59bef4a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 496, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 605 + ], + "type": "text", + "content": "perceived gender and skin color of 28,315 of the depicted people. In virtue of this dataset, we could get the overlap between our dataset and the selected images and find that 985 images in " + }, + { + "bbox": [ + 46, + 496, + 287, + 605 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 496, + 287, + 605 + ], + "type": "text", + "content": "-VQA-Real are annotated with gender and skin information. Among these images, we only keep those that contain only males or females, resulting in 717 images eventually (510 male images and 207 female images). We then evaluate the performance of different models by subgroups. The result is shown in Fig. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Specifically, for each model we obtain the performance of both original and counterfactual questions of different subgroups. We then compute the difference between the accuracy of original and counterfactual questions in each subgroup. Formally, we have the formula diff(male) = male(ori) - male(cf) and diff(female) = female(ori) - female(cf). As shown in Fig. 5, most end-to-end models have a larger diff(female) than diff(male), which indicates that when presenting" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 219, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 307, + 219, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 369 + ], + "type": "image", + "image_path": "78ebaa052cb12cbff3f6b96fbf4629a139723f43265882bb278915496a47fc99.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 371, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 304, + 371, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 547, + 437 + ], + "type": "text", + "content": "Figure 5. Performance difference of original and counterfactual questions on the male and female subgroup on " + }, + { + "bbox": [ + 304, + 371, + 547, + 437 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 371, + 547, + 437 + ], + "type": "text", + "content": "-VQA-Real. We can see that end-to-end models are often biased toward the male subgroup, and neuro-symbolic models are biased toward the female subgroup. The larger the gap between the performance differences, the larger the bias." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": "counterfactual presuppositions, they struggle more with females images. Further studies are needed to understand the reasons and improve the fairness in the model reasoning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 506, + 537, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 506, + 537, + 532 + ], + "spans": [ + { + "bbox": [ + 315, + 506, + 537, + 532 + ], + "type": "text", + "content": "Finding 4: MLLMs demonstrate systematic bias in answering gender-related counterfactual questions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 548, + 379, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 379, + 561 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we study the ability of current multi-modal language models to handle counterfactuals - a core cognition ability of human intelligence. We build a novel dataset " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "-VQA to test this counterfactual reasoning ability, the dataset is designed to have three types of counterfactual questions, with mixed real and synthetic images and questions. Our evaluations on " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "-VQA reveal several findings, the most significant finding is that no model in the current multi-modal model literature can consistently handle our counterfactual questions. We have released our code and the dataset to help the community move forward on achieving human-level multi-modal intelligence." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21860" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In CVPR, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 267, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 267, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 267, + 146 + ], + "type": "text", + "content": "[2] Anthropic. Claude. https://claude.ai, 2023.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "type": "text", + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "type": "text", + "content": "[5] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "text", + "content": "[6] Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit, Kevin A Smith, and Joshua B Tenenbaum. Are deep neural networks smarter than second graders? In CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 388 + ], + "type": "text", + "content": "[7] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv:2110.14168, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "type": "text", + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. NeurIPS, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "type": "text", + "content": "[9] Surís Dídac, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In ICCV, 2023. 1, 6, 7, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "text", + "content": "[11] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "text", + "content": "[12] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. ICLR 2024, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[14] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2017. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and surface variations. arXiv:1807.01697, 2019. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[17] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "text", + "content": "[18] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 195, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 248 + ], + "type": "text", + "content": "[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "type": "text", + "content": "[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 283, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 315 + ], + "type": "text", + "content": "[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 358 + ], + "type": "text", + "content": "[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "text", + "content": "[23] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 394, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 436 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 471, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 492 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 6, 14" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "text", + "content": "[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 580 + ], + "type": "text", + "content": "[28] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. arXiv:2306.08568, 2023. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[29] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 1, 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "text", + "content": "[30] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 658, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 678 + ], + "type": "text", + "content": "[31] OpenAI. Gpt-4 technical report. arXiv:2303.08774, 2023. 3, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[32] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2,4,7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21861" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Dong Qingxiu, Li Lei, Dai Damai, Zheng Ce, Wu Zhiyong, Chang Baobao, Sun Xu, Xu Jingjing, Li Lei, and Sui Zhi-fang. A survey on in-context learning. arXiv:2304.08485, 2023. 4, 11" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[34] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "text", + "content": "[35] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv:2308.12950, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "type": "text", + "content": "[36] Haoqin Tu, Bingchen Zhao, Chen Wei, and Cihang Xie. Sight beyond text: Multi-modal training enhances llms in truthfulness and ethics. arXiv:2309.07120, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 285 + ], + "type": "text", + "content": "[37] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv:2311.03079, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 286, + 330 + ], + "type": "text", + "content": "[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 4, 6, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 332, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 287, + 387 + ], + "type": "text", + "content": "[39] Te-Lin Wu, Zi-Yi Dou, Qingyuan Hu, Yu Hou, Nischal Chandra, Marjorie Freedman, Ralph Weischedel, and Nanyun Peng. ACQUIRED: A dataset for answering counterfactual questions in real-life videos. In Conference on Empirical Methods in Natural Language Processing, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 441 + ], + "type": "text", + "content": "[40] Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Akyurek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, and Yoon Kim. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. arXiv:2307.02477, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 487 + ], + "type": "text", + "content": "[41] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv:2309.17421, 9, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 489, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 287, + 520 + ], + "type": "text", + "content": "[42] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "text", + "content": "[43] Wenhao Yu, Meng Jiang, Peter Clark, and Ashish Sabharwal. Ifqa: A dataset for open-domain question answering under counterfactual presuppositions. arXiv:2305.14010, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "type": "text", + "content": "[44] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 602, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 602, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 602, + 287, + 633 + ], + "type": "text", + "content": "[45] Bingchen Zhao, Quan Cui, Hao Wu, Osamu Yoshie, and Cheng Yang. Vision learners meet web image-text pairs. arXiv:2301.07088, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 636, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 667 + ], + "type": "text", + "content": "[46] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592, 2023. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] Yongshuo Zong, Oisin Mac Aodha, and Timothy Hospedales. Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[49] Yongshuo Zong, Tingyang Yu, Bingchen Zhao, Ruchika Chavhan, and Timothy Hospedales. Fool your (vision and) language model with embarrassingly simple permutations. arXiv:2310.01651, 2023. 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21862" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_content_list.json b/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b66a04ac2f1d494f0211e4a9cbd5fdc6ba84e8d7 --- /dev/null +++ b/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_content_list.json @@ -0,0 +1,2128 @@ +[ + { + "type": "text", + "text": "What Sketch Explainability Really Means for Downstream Tasks?", + "text_level": 1, + "bbox": [ + 145, + 130, + 815, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hmrishav Bandyopadhyay1 Pinaki Nath Chowdhury1 Ayan Kumar Bhunia1 \nAneeshan Sain1 Tao Xiang1,2 Yi-Zhe Song1,2 \n1SketchX, CVSSP, University of Surrey, United Kingdom. \n2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence.", + "bbox": [ + 173, + 172, + 795, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{h.bandyopadhyay, p.chowdhury, a.bhunia, a.sain, t.xiang, y.song}@surrey.ac.uk", + "bbox": [ + 140, + 247, + 826, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the unique modality of sketch for explainability, emphasising the profound impact of human strokes compared to conventional pixel-oriented studies. Beyond explanations of network behavior, we discern the genuine implications of explainability across diverse downstream sketch-related tasks. We propose a lightweight and portable explainability solution - a seamless plugin that integrates effortlessly with any pre-trained model, eliminating the need for re-training. Demonstrating its adaptability, we present four applications: highly studied retrieval and generation, and completely novel assisted drawing and sketch adversarial attacks. The centrepiece to our solution is a stroke-level attribution map that takes different forms when linked with downstream tasks. By addressing the inherent non-differentiability of rasterisation, we enable explanations at both coarse stroke level (SLA) and partial stroke level (P-SLA), each with its advantages for specific downstream tasks.", + "bbox": [ + 73, + 316, + 473, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 603, + 207, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sketches, rooted in human expression [43], offer a distinctive modality for exploring explainability [61, 70]. In contrast to photos, where each pixel is independent and lacks inherent meaning, sketches are organised into strokes, with each stroke carrying subjective meaning assigned by the sketcher [44]. This paper explores sketch explainability, but with a unique perspective – aiming to provide explanations and unravel the true implications of explainability on various downstream sketch-related tasks.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With this perspective in mind, our approach champions an explainability solution that is (i) lightweight and portable - a plugin seamlessly integrating with multiple pre-trained models without necessitating re-training [98], and (ii) easily adaptable to a diverse array of downstream sketch-specific tasks, benefiting the broader community.", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our solution is exclusively centred on human strokes, aiming to attribute explanation on different stroke granularity: individual strokes (coarse) and their parts (fine). The", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/482f4d2870546db7edb97a1621470c70cadaa0d41a4106b3b4b329f57e112ef4.jpg", + "image_caption": [ + "Figure 1. We attribute explanations for individual strokes (stroke-level attribution) and their vector coordinate points (point-level attribution). Stroke-level attribution rasterises individual strokes (non-differentiably) to produce $n$ -stroke images. Next, we sum the stroke images to get the complete sketch image used for downstream tasks. Point-level Attribution computes distance transform from stroke coordinates and thresholds to get the sketch image. Our explainability solution works without re-training for existing tasks like SBIR and sketch-to-photo generation and novel tasks like filtering noisy strokes for assisted drawing and adversarial attack by removing a small stroke." + ], + "image_footnote": [], + "bbox": [ + 498, + 287, + 890, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "output of our model markedly differs from typical saliency maps [1] found in photo-based explainability models, where the emphasis is mostly on visualisation [90]. Ours is a task-driven attribution map that assigns stroke-level attributes capturing how altering stroke characteristics can impact model prediction. Depending on the downstream tasks, attributions can be grounded to, for example (i) importance of entire strokes, which is more suitable to filter noisy strokes [15] in assisted drawing, and remove small strokes for adversarial attacks on existing sketch encoders, and (ii) stroke shape and length, where a partial-stroke level attribution is beneficial for tasks like sketch-based image retrieval [77] and sketch-to-photo generation [55].", + "bbox": [ + 496, + 640, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To showcase the adaptability of our model, we carefully devise four applications: two well-studied tasks from existing literature (retrieval [12, 26] and generation [55, 102]), and two entirely novel tasks (assisted drawing and sketch", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10997", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adversarial attack). In retrieval, we evaluate reliability of model predictions by comparing predicted stroke order with the order in which a human draws them. For generation, we pinpoint strokes with the least influence, offering explicit feedback to end-users regarding which strokes the model prioritised and which it overlooked. In assisted drawing [4], we assist novice artists in faithfully sketching a particular photo by identifying strokes that do not match the target photo. Lastly, in adversarial attacks, we unveil the vulnerability of state-of-the-art sketch encoders by removing a small imperceptible stroke in any sketch, resulting in significant changes to the model's prediction.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The focal point connecting all downstream tasks is our proposed stroke-level attribution. The key question, therefore, is how to backpropagate information to strokes while addressing the inherent non-differentiability of rasterisation - strokes are most often represented as discrete coordinates and rasterised before feeding into downstream applications. We provide two solutions for non-differentiability: (i) coarse stroke level: we first rasterise individual strokes to produce raster stroke images. Then, we combine these stroke images to get the complete sketch image (see Fig. 1) - since this addition of stroke images is a differentiable operation, we can backpropagate information from the complete sketch to individual raster stroke images. (ii) fine partial-stroke level: we create a distance transform image from stroke coordinates (\"red dots\" in Fig. 1 vector sketch) by calculating the minimum distance of each pixel in the image from the coordinates. Then, we threshold the distance value to get the sketch image (white pixels for a high distance and black pixels for a low distance). Since the distance function and our threshold step are both differentiable, we can backpropagate information from the sketch images to stroke coordinates.", + "bbox": [ + 75, + 273, + 472, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarised as follows: (i) We explore sketch explainability, emphasising the importance of strokes in human-drawn sketches. (ii) We highlight the profound impact of explainability on various sketch-related domains, presenting applications in retrieval, generation, assisted Drawing, and adversarial attacks. (iii) We solve for the non-differentiability problem of rasterisation, and provide both stroke-level and partial-stroke level attribution.", + "bbox": [ + 75, + 606, + 472, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 741, + 227, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sketch for Visual Understanding: Having a high visual proximity [43] to real images and carrying human subjectivity [9, 76], amateur sketches or abstract line drawings [59] has been a popular modality for customised expression, thus driving extensive applications as a query for retrieval [26, 32, 76, 85] of object [29] or scene [24] images, 3D shapes [104], and even concepts like in 'dictionary-like' games [11]. As a canvas for creativity, sketch helps image-editing [62, 110, 114], or generation of objects [20, 21, 38, 99],", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scenes [109], and 3D shapes [10, 39, 105, 119]. Being easily editable, sketch enables interactive access to AI systems like image-segmentation [45, 122], object localisation [96], image-inpainting [110, 114], and incremental learning [14]. Being application-specific however, such works largely ignored explaining the 'how' of sketch-correspondence. The few who did, customised training pipelines [61, 70] for niche tasks. In this work, we thus make the first attempt at visualising salient sketch-regions (strokes), as an explainability-tool (like GradCAM [81] in photos) for existing pre-trained sketch-based downstream networks.", + "bbox": [ + 496, + 90, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explaining CNN Predictions: CNN explanations visually highlight regions of input having the maximum influence on a model's predictions [87]. This visualisation of 'salient' regions is either through an analysis [37] of a regular pre-trained network after it has completed training (post-hoc), or by designing and training explicitly interpretable (i.e. explain-and-predict) models [17, 19, 64]. Given a pre-trained CNN, a post-hoc algorithm either visualises (i) model attributes like feature and activation-maps [34, 65, 66, 82, 113, 120] that imply saliency of specific input (pixel) regions, or (ii) input attributes directly as pixels [91] or pixel-regions [51] coloured according to their relative importance. Visualisation of input attributes is facilitated through perturbation based algorithms [18, 36, 37] and gradient-based analysis [87, 89, 91]. Perturbation-based algorithms [27, 68, 113] detect saliency of pixel-regions by measuring impact of their absence on the prediction score. Whereas, gradient-based algorithms [8, 30, 86, 92, 117] measure the gradient of the prediction with respect to individual input-pixels (input-gradients [83]), attributing them based on this value. Unlike images however, sketch is a sparse-information modality [23] for pixels. As such, we explore explainability in sketches by computing stroke and point attribution for fine-grained sketch explanations.", + "bbox": [ + 496, + 261, + 893, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation of CNN Explanations: The evaluation of CNN explanations has evolved over time - from naive qualitative analysis of visualisations [82, 87, 91, 113] to standardised theoretical [5, 92] and empirical [72, 107, 113] baselines. Analysing explanations theoretically helps evaluate them in a model-agnostic environment, where their mathematical form is checked against pre-defined properties (axioms)[5, 54, 92]. Empirical evaluations, on the other hand, involve experiments measuring $(i)$ variance in explanations upon perturbations of inputs [107] and model weights [1, 2] (sanity checks), and $(ii)$ accuracy of explanations in locating important features [37, 72, 82, 117]. These features, when perturbed, influence the CNN maximally, as measured by perturbation-based metrics [6, 51, 68, 78, 82, 83, 113]. Recently, however, these evaluation protocols have met criticism [35] due to their detachment from humans. Instead, human studies [53, 58, 82, 88], and human-centered metrics [35] have been proposed for evaluating explanation in", + "bbox": [ + 496, + 628, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10998", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "terpretability. In this work, we attribute strokes by backpropagation [86, 87, 91, 92], evaluating attributions through sanity checks [1], empirical metrics [51] and human studies [53] on downstream sketch-based applications [112].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 159, + 207, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we provide a brief overview of some standard concepts, ubiquitous in explainability literature [1] to help formalise the question: \"what entails a good explanation?\"", + "bbox": [ + 75, + 176, + 468, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribution Algorithms: It highlights relevant regions (e.g., pixels in an image, see Fig. 1) that are responsible for the model's prediction. Despite its importance for safety-critical applications [49, 75, 101], making an attribution algorithm interpretable to humans remains an open problem. Surprisingly, a more faithful attribution is usually less interpretable and vice-versa [91, 113]. Prior works [16, 81] study this trade-off between faithfulness vs. interpretability as an answer to: \"What makes a good visual explanation?\"", + "bbox": [ + 75, + 226, + 468, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attribution map $\\mathbb{A} \\in \\mathbb{R}^{H \\times W}$ is typically calculated using gradients for an input $\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ for a classification model $\\hat{y} = F_{\\theta}(\\mathrm{X}) \\in \\mathbb{R}^{C}$ , pre-trained on $C$ categories. The gradients are a simple and good indicator of how much the model prediction changes for input $\\mathrm{X}$ as,", + "bbox": [ + 75, + 361, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} = \\partial F _ {\\theta} (\\mathrm {X}) / \\partial \\mathrm {X} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 440, + 468, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interpretability: It is the ability of an attribution algorithm to provide a qualitative \"understanding\" for a model [73]. This \"understanding\" depends on the target audience, e.g., a human expert may interpret a small Bayesian network [56], but a layman is more comfortable with a weighted attention (or feature) map that highlights salient regions [1]. To evaluate the interpretability of attribution maps, prior works either (i) perform downstream tasks [31, 81] that depend on the interpretation (e.g., object localisation – predict the bounding box and semantic segmentation for image region with the highest attribution) or (ii) human studies [53], typically conducted in two setups – class discriminative (given an attribution map, ask users to identify the category predicted by a model), and trustworthiness (compare attribution maps from a strong and a weak model, and ask users to identify the stronger model).", + "bbox": [ + 75, + 459, + 468, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Faithfulness: It is the ability of attribution algorithms to accurately \"explain\" the computation learned by a model. For example, in theory, a fully faithful attribution is the entire model (e.g., ResNet-18 [41]) but is not interpretable by a human. In practice, for an attribution to be meaningful, it is often impossible to be completely faithful. To balance this trade-off, prior works explore human-interpretable attributions that are locally faithful for a given model prediction. One approach is image occlusion [73], where the difference in the model scores is measured when masking different patches in an input image. Image patches that significantly change the model score are deemed important by the attribution algorithm.", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 89, + 671, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attribution map $\\mathbb{A}$ in Eq. (1) gives a faithful fine-grained explanation for each pixel in X. However, such pixel attribution does not provide meaningful explanations when interpreting sparse human-drawn sketches (many empty white pixels). Additionally, pixel attribution does not consider the sketch construction process - humans sketch a sequence of strokes, not pixels. Hence, it is most appropriate that sketch predictions should be attributed to strokes and not pixels. However, the key challenge to stroke attributions is that most sketch applications [77, 118] use raster sketches - a non-differentiable process to convert a sequence of strokes into pixels. In the following sections, we propose two stroke attribution algorithms that consider the sketch construction process by designing a differentiable rasterisation pipeline for a vector sequence of strokes.", + "bbox": [ + 496, + 114, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Non-differentiable Rasterisation" + ], + "code_body": "Data: V $\\leftarrow$ Vector Sketch of size $\\mathbb{R}^{T\\times 5}$ \nResult: X $\\leftarrow$ Blank (Zero) Canvas of size $\\mathbb{R}^{h\\times w\\times 3}$ . \n $\\mathbb{B}(\\cdot ,\\cdot)\\gets$ Bresenham Function ; \n $v_{0} = (x_{0},y_{0},q_{0}^{1},q_{0}^{2},q_{0}^{3})\\gets \\mathrm{V}[0]$ . \n $v_{\\mathrm{prev}}\\gets v_0$ . \n $q_{\\mathrm{prev}}\\gets q_0^1$ . \nfor $v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})\\gets \\mathrm{V}[1\\dots T]$ do if $q_{\\mathrm{prev}} = 1$ and $q_t^1 = 1$ then for $(p_x^i,p_y^i)\\gets \\mathbb{B}(v_{\\mathrm{prev}},v_t)$ do |X $(p_x^i,p_y^i)\\gets 255;$ end end $v_{\\mathrm{prev}}\\gets v_t$ . $q_{\\mathrm{prev}}\\gets q_t^1$ . if $q_{t}^{3} = 1$ then exit(); /\\* End of Drawing \\*/ end", + "bbox": [ + 516, + 371, + 867, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Sketch Representations", + "text_level": 1, + "bbox": [ + 500, + 666, + 715, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While sketches are used in several formats (or representations) like Raster [111], Vector [40], or Bézier [28], digital sketches are primarily captured in vector form, as a list of points traced on a drawing pad. These points are usually a five-element vector $v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})$ where, $(x_{t},y_{t})$ are absolute coordinates in a $(H\\times W)$ drawing canvas and the last three elements are one-hot encoding of pen-states: pen touching the paper $(1,0,0)$ , pen is lifted $(0,1,0)$ , and end of drawing $(0,0,1)$ . Hence, a vector sketch with $T$ points is represented as $\\mathbf{V}\\in \\mathbb{R}^{T\\times 5}$ .", + "bbox": [ + 496, + 688, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As most downstream sketch applications work on rasterised sketches $\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ , prior works [111] translate (non-differently) a vector sketch $\\mathrm{V}$ to its equivalent raster sketch $\\mathrm{X}$ using Algorithm 1. For this, the Bre", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10999", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97dc47e35b6d7b1ce1fc4bc686694c42bb19e462417b7bebceca4c9ff4c672ab.jpg", + "image_caption": [ + "Figure 2. Coarse Stroke-level Attribution. Backpropagate gradients from raster sketch X to raster strokes $S_{i}$ , with weight $\\omega_{i}$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 460, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "senham function $\\mathbb{B}(\\cdot)$ is used to connect two vector points $\\{v_{t - 1}, v_t\\} \\in \\mathrm{V}$ in the pixel space $\\{(p_x^1, p_y^1), \\ldots, (p_x^n, p_y^n)\\} \\in \\mathbb{B}(v_{t - 1}, v_t)$ via a continuous line. Next, we show how our sketch attributions overcome this non-differentiable rasterisation and backpropagate gradients (Eq. (1)) from pixels $\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ to strokes and points in $\\mathrm{V} \\in \\mathbb{R}^{T \\times 5}$ .", + "bbox": [ + 76, + 220, + 468, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Coarse Stroke-level Attribution (SLA)", + "text_level": 1, + "bbox": [ + 76, + 324, + 405, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we backpropagate gradients from pixel space in $\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ to strokes, defined as a continuous set of points $\\{v_t, v_{t+1}, \\ldots, v_{t+n}\\}$ from the first pen-down $(1,0,0)$ till the pen-up $(0,1,0)$ state. Algorithm 1 converts the vector stroke points into a raster stroke $S_i \\in \\mathbb{R}^{H \\times W \\times 3}$ . The final raster sketch is then a differentiable composition of $m$ raster strokes $\\mathrm{X} = \\sum_{k=1}^{m} S_k$ . We compute SLA as,", + "bbox": [ + 75, + 349, + 470, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathcal {S} _ {i}} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 468, + 468, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This, however, gives a degenerate solution where all strokes will have the same attribution $\\partial \\sum_{k=1}^{m} S_k / \\partial(S_i) = 1$ . To avoid this, we compute a weight factor $\\omega_i \\in \\mathbb{R}^{H \\times W \\times 3}$ for each stroke $S_i$ such that,", + "bbox": [ + 75, + 513, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {i} \\left(p _ {x}, p _ {y}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\left(p _ {x}, p _ {y}\\right) \\in \\mathbb {B} \\left(v _ {t - 1}, v _ {t}\\right) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 585, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, given consecutive vector sequence of points $(v_{t - 1}, v_t)$ in stroke $S_i$ , we find all points $(p_x, p_y)$ using Bresenham function $\\mathbb{B}(\\cdot)$ that lie \"on the stroke\" and assign $\\omega_i(p_x, p_y) = 1$ . Hence, longer strokes will have more 1's compared to shorter strokes. Finally, Eq. (2) is adapted as", + "bbox": [ + 75, + 638, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 726, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SLA makes the non-differentiable rasterisation $(\\mathrm{V} \\to \\mathrm{X})$ partially differentiable (stokes $S$ to sketch X). In other words, SLA answers \"Which strokes in a sketch are important\". Next, we make the rasterisation fully differentiable and backpropagate gradients to vector points V to answer \"Which point in a sketch is important\".", + "bbox": [ + 75, + 771, + 468, + 863 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6d2bf8074d1ead883ebc215632550611d23976834afd81b2473cbf07644f9a2.jpg", + "image_caption": [ + "Figure 3. Partial Stroke-level Attribution. Backpropagate gradients from raster sketch X to vector sequence of coordinates V." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 874, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Partial Stroke-level Attribution (P-SLA)", + "text_level": 1, + "bbox": [ + 498, + 222, + 844, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike SLA, which partially captures the sketch construction process (stroke-level), partial stroke-level attribution (P-SLA) can fully backpropagate gradients to the vector list of coordinates $\\mathrm{V} \\in \\mathbb{R}^{T \\times 5}$ traced on a drawing pad. Given a blank canvas $\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we (i) calculate the minimum distance of every pixel $(p_x, p_y)$ in $\\mathrm{X}$ from a line segment $(v_{t-1}, v_t)$ in $\\mathrm{V}$ , and (ii) compute the pixel intensity of $\\mathrm{X}(p_x, p_y)$ as function of the minimum distance as", + "bbox": [ + 498, + 244, + 890, + 367 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {X} \\left(p _ {x}, p _ {y}\\right) = \\sigma [ 2 - 5 \\min _ {t = 2, \\dots , T} (5) \\\\ \\left. \\mathrm {d i s t} ((p _ {x}, p _ {y}), v _ {t - 1}, v _ {t}) + (1 - q _ {t - 1} ^ {1}) 1 0 ^ {6}) \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 377, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the sigmoid function, and $\\mathrm{dist}(\\cdot)$ is a distance function (see Supp.) from a point $(p_x,p_y)$ to a line segment $(v_{t - 1},v_t)$ . For pen-up states $(q_{t - 1}^{1} = 0)$ , we blow up $(\\times 10^{6})$ the distance values that make the pixel intensities $\\mathrm{X}(p_x,p_y)\\to 0$ , i.e., not render strokes for $(v_{t - 1},v_t)$ . Finally, we compute P-SLA", + "bbox": [ + 498, + 434, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial v _ {t}} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 535, + 890, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Applications of Stroke Attributions", + "text_level": 1, + "bbox": [ + 500, + 585, + 818, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite its simplicity, designing stroke attribution algorithms that capture the sketch construction process unlocks insights into numerous existing downstream tasks like classification [40], robust sketch-based image retrieval (category-level [33] and fine-grained [79]) and enables some novel sketch applications like assisted drawing [15], interactive sketch to photo generation [55], adversarial attacks on human-drawn sketches, and discovering the \"arrow of time\" [100] in raster sketches.", + "bbox": [ + 496, + 611, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Robust Sketch Based Image Retrieval", + "text_level": 1, + "bbox": [ + 500, + 756, + 823, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a query sketch $\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ , category-level sketch-based image retrieval (SBIR) aims to fetch category-specific photos from a gallery of multi-category photos (e.g., given sketch of a 'shoe' retrieve any photo 'shoe' from a gallery of 'shoes+hats+cows'). Conversely, fine-grained SBIR aims to retrieve one instance from a gallery of the same category photos (e.g., given the sketch of a 'shoe' retrieve one photo shoe from a gallery of all shoes). Deep learning frameworks", + "bbox": [ + 496, + 777, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "For overlapping strokes in $S_{i}$ and $S_{j}$ , we clamp the maximum pixel value using differentiable functions like torch.clamp", + "bbox": [ + 76, + 875, + 467, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11000", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6c7cf5c6642548aed8d9977a02b4656b1294e1207f9dfcfd860d1983860e26fb.jpg", + "table_caption": [ + "Table 1. Stroke attribution (SLA, P-SLA) make SBIR systems reliable. Sketches with a high correlation $(Corr)$ of stroke saliency (predicted by SLA or P-SLA) with human-drawn temporal stroke order tend to have higher retrieval accuracy." + ], + "table_footnote": [], + "table_body": "
MetricsFull DatasetCorr ≥ 0.5Corr ≤ 0.1
SLAP-SLASLAP-SLA
CategorymAP53.155.357.651.750.1
LevelP@20065.966.768.564.661.5
FineAcc.@115.316.417.613.812.7
GrainedAcc.@534.236.939.431.128.3
", + "bbox": [ + 78, + 155, + 470, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "learn a joint sketch-photo manifold (for category and fine-grained) via a feature extractor [26, 29, 103] trained using triplet loss [112]. Recent adoption of foundation models for SBIR [77] shifts focus to robust deployment using the open-set generalisation of CLIP [71].", + "bbox": [ + 75, + 256, + 468, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Towards this goal of robust deployment, our sketch attribution algorithms $(\\mathbb{A}_i^R$ and $\\mathbb{A}_t^V)$ can predict which strokes the network focuses on when retrieving a photo (Fig. 4). Apart from interpreting SBIR models, sketch attribution can also help detect potential failures at runtime (inference). First, we use the attribution scores $\\mathbb{A}_i^R$ or $\\mathbb{A}_t^V$ to rearrange the strokes from highest to lowest. The attribution scores indicate the most salient to the least salient strokes that affect model prediction. Second, we calculate a correlation $(Corr)$ of our predictor stroke order with the ground-truth temporal stroke order drawn by a user (humans draw the most salient regions first and least salient areas last [33, 79]). A high correlation indicates that humans and our model prioritise strokes similarly, whereas a low $Corr$ denotes that the model and the user prioritise different strokes. We evaluate SBIR on a pre-trained SOTA model [77] using CLIP with prompt learning as a sketch and photo encoder.", + "bbox": [ + 75, + 335, + 472, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We use TU-Berlin [33] (for category-level SBIR) and Sketchy [79] (for fine-grained SBIR). TU-Berlin contains 250 categories, with 80 free-hand sketches in each, and 204, 489 images [116]. Sketchy [79] has 75, 471 sketches over 125 categories having 100 images in each [108].", + "bbox": [ + 75, + 599, + 468, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: Following [77], we use mean average precision (mAP) and precision for top 200 retrieved samples $(\\mathbb{P}@\\mathbb{200})$ for category-level SBIR. For fine-grained SBIR [23], we measure Acc.@q, i.e., the percentage of sketches whose true matched photo is in the top-q list.", + "bbox": [ + 75, + 684, + 470, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results: We divide the evaluation set into two sets: (i) those that have a high correlation $Corr \\geq 0.5$ , and (ii) those with a low correlation $Corr \\leq 0.1$ of ground-truth and predicted stroke order. Tab. 1 shows sketches with a high $Corr \\geq 0.5$ are 1.7/3.9 more accurate in Acc.@1/Acc.@5 than those with $Corr \\leq 0.1$ for fine-grained SBIR, and 3.3/1.7 better in mAP/P@200 for category-level SBIR. Full Dataset indicates the performance of the pre-trained model on the entire evaluation set.", + "bbox": [ + 75, + 767, + 472, + 905 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81299084bc63347c20a46cf747ab3e28cf0e2ecb551e4b192eab4f18863aff04.jpg", + "image_caption": [ + "Figure 4. Sketch attributions from stroke-level and point-level for image retrieval. High correlation of human-drawn stroke order with that from sketch-attributions (high $\\rightarrow$ low) indicate our sketch encoder gives more importance to salient strokes drawn early on." + ], + "image_footnote": [], + "bbox": [ + 509, + 92, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Assisted Drawing via Noisy Stroke Removal", + "text_level": 1, + "bbox": [ + 498, + 340, + 872, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although sketching has enabled many exciting applications [25, 55, 63, 118], the fear to sketch (i.e., \"I can't sketch\") has proven fatal for its widespread adoption. To solve this, prior works [15] used complex (and hard to train [12]) reinforcement learning [80] to predict the importance of each stroke in a sketch. Next, a stroke subset selector removes noisy (less important) strokes, leaving only those positively contributing to the downstream tasks.", + "bbox": [ + 496, + 358, + 890, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we focus on assisted drawing - given a photo, we use attribution scores $(\\mathbb{A}_i^R$ or $\\mathbb{A}_t^V)$ to help humans draw a faithful and clean sketch. Our method is significantly simpler than reinforcement learning alternatives [15].", + "bbox": [ + 496, + 479, + 890, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For SLA, an input sketch $\\mathrm{X} = \\sum_{k=1}^{m} \\omega_k S_k$ is composed of $m$ strokes. We calculate the cosine similarity (sim) of input sketch $\\mathrm{X}$ with its target photo $\\mathrm{P} \\in \\mathbb{R}^{H \\times W \\times 3}$ to measure \"how faithfully $\\mathrm{X}$ describes $\\mathrm{P}$ \". Next, we backpropagate gradients from the cosine similarity to strokes $S_i$ and calculate stroke-level attribution score $\\mathbb{A}_i^R$ as", + "bbox": [ + 496, + 539, + 893, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial \\operatorname {s i m} (F _ {\\theta} (\\mathrm {X}) , F _ {\\theta} (\\mathrm {P}))}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 641, + 890, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pre-trained sketch and photo encoder $F_{\\theta}(\\cdot)$ must be highly accurate to judge sketch-photo correspondence. Hence, we use pre-trained CLIP+prompts encoder from [77]. To remove noisy strokes, we only update the weights $\\omega_{i}\\in \\mathbb{R}^{H\\times W\\times 3}$ using a normalised attribution score $\\mathbb{A}_i^R$ as", + "bbox": [ + 496, + 686, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {i} ^ {*} = \\omega_ {i} \\cdot \\text {G u m b e l . S o f t m a x} \\left(\\frac {\\mathbb {A} _ {i} ^ {R}}{\\sum_ {\\forall i} \\mathbb {A} _ {i} ^ {R}} + \\Delta\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 773, + 890, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gumbel_Softmax(·) makes the output one-hot (discrete value) in the forward pass but differentiable with a probability distribution that sum to 1 in the backward pass [47]. The", + "bbox": [ + 496, + 819, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2The sketch and photo encoder $F_{\\theta}(\\cdot)$ could be a siamese-style shared network or two independent models with different network weights [77].", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11001", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02207721c41a183f11951d641c93908624ec77b6048c9f07c99dc64e6296e5a8.jpg", + "image_caption": [ + "Figure 5. Assisted drawing via sketch healing (or filtering noisy strokes) using stroke attributions from SLA and P-SLA. This helps users having fear-to-sketch (\"I can't sketch\")." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 433, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "modified sketch is constructed as $\\mathrm{X} = \\omega_1^* S_1 + \\dots + \\omega_m^* S_m$ . Intuitively, we keep strokes that contribute $(\\mathbb{A}_i^R)$ to a high cosine similarity matching human sketch $\\mathrm{X}$ and target photo $\\mathrm{P}$ and remove $S_i$ with normalised $\\mathbb{A}_i^R$ lower than $(0.5 - \\Delta)$ .", + "bbox": [ + 75, + 244, + 468, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to SLA, we can also use P-SLA to compute the attribution $\\mathbb{A}_t^V$ for each point $v_{t}$ in the vector sketch $\\mathrm{V}\\in \\mathbb{R}^{H\\times W\\times 3}$ , by measuring the cosine similarity between input X and target photo P as", + "bbox": [ + 75, + 306, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial \\operatorname {s i m} \\left(F _ {\\theta} (\\mathrm {X}) , F _ {\\theta} (\\mathrm {I})\\right)}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 372, + 468, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We remove noisy points $v_{t}$ by updating the pen-states in Eq. (5) from pen-down $(1,0,0)$ to pen-up $(0,1,0)$ depending on its attribution $\\mathbb{A}_t^V$ for point $v_{t}\\in \\mathrm{V}$ .", + "bbox": [ + 75, + 417, + 468, + 465 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq _ {t - 1} ^ {1 *} = q _ {t - 1} ^ {1} \\cdot \\text {G u m b e l . S o f t m a x} \\left(\\frac {\\mathbb {A} _ {t - 1} ^ {V}}{\\sum_ {\\forall t} \\mathbb {A} _ {t} ^ {V}}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 469, + 468, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using updated values for $q_{t-1}^{1*}$ , we update $q_{t-1}^{2*} = 1 - q_{t-1}^{1*}$ and recalculate pixel intensities for raster sketch $\\mathrm{X}(p_x, p_y)$ . The value of hyper-parameter $\\Delta$ significantly affects the stroke removal process. We found the optimal $\\Delta$ for SLA and P-SLA is 0.3 and 0.1, respectively. A higher $\\Delta$ for P-SLA gives broken lines with a drop in the visual quality of an input sketch. Next, we evaluate stroke filtering using SLA and P-SLA on popular human-drawn sketch datasets.", + "bbox": [ + 75, + 511, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: For a fair comparison with prior works [15], we evaluate on fine-grained SBIR datasets QMUL-ShoeV2 and QMUL-Chair-V2 [12, 67, 112]. It consists of 6,730/1,800 sketches and 2,000/400 photos from ShoeV2/Chair-V2. We evaluate on the standard test-split of 679/525 sketches and 200/100 photos.", + "bbox": [ + 75, + 636, + 468, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric: We measure the retrieval accuracy of the clean sketch with the target photo by computing Acc.@1 and Acc.@5. A high accuracy need not correspond to high visual quality to the human eye [84]. Hence, we conducted a small human study with 5 participants and reported the mean opinion score (MOS) [46]; each was asked to compare two sets of 50 sketch pairs (GT sketch vs SLA filtered) and (GT sketch vs P-SLA filtered).", + "bbox": [ + 75, + 729, + 468, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: Removing noisy strokes from human-drawn sketches is still a new topic; hence, to the best of our knowledge, there is only one work by Bhunia et al. [15]. From", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/083de7ee413bf5f9a739223370e5b26b32487fa956a14bfa67a5273c7eb59d63.jpg", + "table_caption": [ + "Table 2. Noisy stroke removal using SLA and P-SLA attribution." + ], + "table_footnote": [], + "table_body": "
MetricsGT SketchSLA filteredP-SLA filteredBhunia et al. [15]
Shoe-V2Acc.@133.436.136.543.7
Acc.@567.868.769.374.9
MOS28.685.757.1-
Chair-V2Acc.@153.354.956.564.8
Acc.@574.376.677.179.1
MOS35.871.457.1-
", + "bbox": [ + 501, + 107, + 890, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2, both Bhunia et al. [15], and ours improve fine-grained SBIR performance by $10.3\\%$ and $3.1\\%$ , respectively. However, [15] outperforms our SLA and P-SLA filtered methods by $7.6\\%$ and $7.2\\%$ , respectively. This performance gap is likely because [15] trains the baseline model [12] using actor-critic version of PPO [80], whereas our SLA and P-SLA work post-hoc [98] without training the baseline model [12]. Additionally, Bhunia et al. [15] aims to design a robust SBIR pipeline, whereas our SLA/P-SLA filtering aims to assist humans in drawing sketches. For human study (MOS): (i) users prefer SLA-filtered $78.5\\%$ vs. $21.5\\%$ for GT sketch, (ii) however, P-SLA-filtered are preferred only $57.1\\%$ vs. $42.9\\%$ for GT sketch. This is verified by Fig. 5 where P-SLA-filtered sketches have broken strokes that degrade their visual quality.", + "bbox": [ + 496, + 224, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Interactive Sketch To Photo Generation", + "text_level": 1, + "bbox": [ + 498, + 460, + 841, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The upsurge of large-scale image generation models (e.g., Stable-Diffusion [74], GigaGAN [50]) helped develop sketch-conditional image generation [63, 118]. However, a key limitation of conditional image generation is that these models do not always faithfully follow the input condition. This was resolved in two stages for text-to-image generation: (i) find word tokens with low influence on generated image, and (ii) iteratively update its activation until it reaches a minimum required value. In this section, we design a pipeline for faithful sketch-to-image generation.", + "bbox": [ + 496, + 484, + 890, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using sketch attributions from SLA and P-SLA, we design a post-hoc [98] method that gives feedback to the user - which strokes the model focuses on and which strokes are being ignored. Given this feedback, a user can interact with the system to ensure the model attends all salient regions.", + "bbox": [ + 496, + 635, + 890, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our interactive pipeline is built on top of a pre-trained sketch-to-photo generation model [55], comprising a modified ResNet-50 [41] as sketch encoder and StyleGAN [52] as image decoder. Given a raster sketch X, the modified sketch encoder computes a latent vector $z_{s2p}^{+} = F_{\\theta}(\\mathrm{X})$ where $z_{s2p}^{+} \\in \\mathbb{R}^{14 \\times 512}$ . Next, the StyleGAN [52] decoder generates the underlying image from $z_{s2p}^{+}$ . For a faithful sketch-to-image generation, we measure the \"influence\" of each stroke/vector point on the latent code $z_{s2p}^{+}$ . Particularly, prior works [106] suggest that $z_{s2p}^{+}$ is disentangled into 14-level semantic feature hierarchy, where $z_{s2p}^{+1} \\in \\mathbb{R}^{512}$ has coarse-level features controlling major semantic struc", + "bbox": [ + 496, + 710, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11002", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/daefada324147b350a086fe96ce19923fb2dcd72d509a6bd86e73f1b71ae4ea5.jpg", + "image_caption": [ + "Figure 6. Interactive Sketch to Photo Generation: Our stroke attribution algorithms make existing sketch-to-photo generation pipelines [55] more faithful. We achieve this by computing the stroke-level $\\mathbb{A}_i^S$ or coordinate-level $\\mathbb{A}_t^V$ attribution that has a maximal influence on the latent code $z_{s2p}^{+}$ used by the image decoder." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tures and $z_{s2p}^{+14}\\in \\mathbb{R}^{512}$ has fine-level features controlling colour schemes, etc. Since sketches primarily convey semantic structure, we use the sum of the first 7-layers to compute stroke attribution $\\mathbb{A}_i^R$ or $\\mathbb{A}_t^V$ as", + "bbox": [ + 75, + 306, + 470, + 369 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial \\sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 378, + 359, + 415 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial \\sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 412, + 468, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we qualitatively evaluate our iterative sketch to photo generation pipeline, as shown in Fig. 6.", + "bbox": [ + 75, + 472, + 468, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Adversarial Attacks on Human Sketches", + "text_level": 1, + "bbox": [ + 76, + 513, + 424, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Szegedy et al. [94] discovered that predictions by deep networks can be manipulated with extremely low-magnitude input perturbations. For images, these can be restricted to be imperceptible to human vision, but their effect can completely change the output prediction by a deep network. Such adversarial attacks are possible in image classification [94], semantic segmentation [7, 42], object detection [97, 115], object tracking [22, 48], etc. Studying these quirks is crucial as it can pose a real threat to deep learning as a pragmatic technology [3]. While major work has been dedicated to attacks on images, the recent surge of deployable sketch applications [77, 118] motivates us to present the first study on adversarial attacks for human-sketches.", + "bbox": [ + 75, + 537, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show how our stroke attribution algorithms (SLA and P-SLA) provide the necessary information for adversarial attacks. For brevity, we focus on adversarial attacks on sketch classification [40, 111]. Intuitively, we use sketch attribution to remove the smallest stroke (in SLA Attack) and minimum number of points (in P-SLA Attack), yet have the maximum impact on changing prediction of a pre-trained classifier $F_{\\theta}(\\mathrm{X}) = y_{\\mathrm{cls}}$ . Ours is (i) a white-box [69, 95] setting – we have access to network weights and gradients of our ResNet-18 [41] classifier, pre-trained on QuickDraw [40] or TU-Berlin [33]; (ii) untargeted attack – while tar", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca9e5a683715cb8de978f44d36fdeefde87dd0180fea9337af13f84881dd5dc7.jpg", + "table_caption": [ + "Table 3. Sketch Adversarial Attacks: Using stroke attributions, we remove a small stroke $(|S_i| \\leq \\epsilon)$ that misclassifies an input sketch." + ], + "table_footnote": [], + "table_body": "
No AttackSLA AttackP-SLA Attack
ε = 5ε = 15ε = 5ε = 15
QuickDraw67.265.764.565.163.7
TU-Berlin74.971.568.570.268.1
", + "bbox": [ + 501, + 125, + 890, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "geted attacks [60, 121] misclassify $F_{\\theta}(\\cdot)$ from $y_{\\mathrm{cls}}^{\\mathrm{GT}}$ to a specific target class $y_{\\mathrm{cls}}^{*}$ , untargeted attacks [93] aim to misclassify to any arbitrary class $y_{\\mathrm{cls}}^{\\mathrm{GT}} \\neq y_{\\mathrm{cls}}$ . For a neater description of SLA and P-SLA attacks, we define the rasterisation process using $\\mathcal{R}(\\cdot)$ as $\\mathrm{X} = \\omega_1\\mathcal{S}_1 + \\dots + \\omega_m\\mathcal{S}_m = \\mathcal{R}(\\mathcal{S})$ for SLA and following Eq. (5) for P-SLA we define $\\mathrm{X} = \\mathcal{R}(\\{v_1,\\ldots v_T\\}) = \\mathcal{R}(\\mathrm{V})$ . Next, we find a stroke $S_{adv}$ with stroke length $|S_i|$ less than some threshold $\\epsilon$ as", + "bbox": [ + 496, + 193, + 893, + 316 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {a d v} = \\arg \\max _ {| \\mathcal {S} _ {j} | \\leq \\epsilon} \\mathcal {L} _ {\\mathrm {c l s}} \\left(F _ {\\theta} \\left(\\mathcal {R} \\left(\\mathcal {S} - \\left\\{\\mathcal {S} _ {j} \\right\\}\\right)\\right), y _ {\\mathrm {c l s}} ^ {\\mathrm {G T}}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 325, + 890, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike typical adversarial attacks on images that add a small noise $(\\mathrm{X} + \\Delta x)$ with $||\\Delta x||_{\\infty}\\leq \\epsilon$ , for sketch adversarial attacks, we remove a small stroke $(\\mathrm{X} - S_{adv})$ such that the stroke length is less than $\\epsilon$ as $|\\mathcal{S}_{adv}|\\leq \\epsilon$ . For PSLA attack, we find a subset of $\\epsilon$ vector points $\\mathrm{V}_{adv} =$ $\\{v_1^{adv},\\dots v_\\epsilon^{adv}\\}$ from input sketch $\\mathrm{V}\\in \\mathbb{R}^{T\\times 5}$ which maximises the categorical cross-entropy loss $\\mathcal{L}_{\\mathrm{cls}}$ as", + "bbox": [ + 498, + 363, + 892, + 470 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} v _ {t} ^ {a d v} = \\mathcal {L} _ {\\mathrm {c l s}} \\left(F _ {\\theta} \\left(\\mathcal {R} (\\mathrm {V} - \\{v _ {t} \\})\\right), y _ {\\mathrm {c l s}} ^ {\\mathrm {G T}}\\right) \\tag {13} \\\\ \\mathrm {V} _ {a d v} = \\operatorname {t o p} @ k \\left(\\left\\{v _ {1} ^ {a d v}, v _ {2} ^ {a d v}, \\dots v _ {T} ^ {a d v} \\right\\}, \\epsilon\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 479, + 890, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $\\mathrm{top}@\\mathrm{k}(\\cdot ,\\epsilon)$ picks the highest $\\epsilon$ elements. Fig. 7 shows the adversarial strokes $S_{adv}$ and points $\\mathrm{V}_{adv}$ in red.", + "bbox": [ + 498, + 531, + 890, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset: We evaluate sketch adversarial attacks on QuickDraw [40] and TU-Berlin [33]. We use a subset [103] of $50M$ sketches in QuickDraw [40] as $3.8M$ samples across 345 categories split in $2.1M$ sketches for training, $0.3M$ for validation and $0.4M$ for evaluation. See Sec. 5.1 for details on TU-Berlin [33] dataset.", + "bbox": [ + 496, + 565, + 892, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation: We measure the drop in classification accuracy when using SLA and P-SLA attacks in Tab. 3. Unlike image-based adversarial attacks where $\\epsilon$ is a pixel intensity (non-integer, decimal value), our sketch attacks occur on stroke/point length, making $\\epsilon$ an integer. A higher $\\epsilon \\geq 20$ removes \"visible\" strokes in SLA and broken lines in P-SLA (Fig. 7). Hence, we evaluate accuracy drop for $\\epsilon = 5$ and $\\epsilon = 15$ in Tab. 3. We observe for $\\epsilon = 5$ P-SLA offers a better adversarial attack than SLA by a margin of $1.3\\%$ .", + "bbox": [ + 496, + 659, + 893, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Human Study", + "text_level": 1, + "bbox": [ + 500, + 806, + 640, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpretability aims to help humans understand a model's reasoning process (transparency), verify that its predictions are based on the right constraints (fairness), and evaluate its confidence (trustworthiness) [73, 81]. In Sec. 5, we evaluated interpretability using several automatic metrics (e.g.,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11003", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9cf77f56775305bf24959d400911878cad02868649a9647f3052cc029a52b0a8.jpg", + "image_caption": [ + "Figure 7. Adversarial attacks on human drawn sketches using SLA and P-SLA. The adversarial strokes are marked in RED." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 281, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91fd081462594502b2646ec0a3efcf609c8e57ca1bb300821047060e05535b6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 361, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b3a6af5e965fd9930020cf7cf5177a99aec8da7511cfef5f57fad78829c2440.jpg", + "image_caption": [ + "GT:Angel" + ], + "image_footnote": [], + "bbox": [ + 361, + 88, + 480, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcb3f9f886c5b283fd2482c167a0bd8f5971b284b48e2bd385c79f6ee096c679.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 89, + 563, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47c221124c8a01a26abe3d82aace0fd4d53dc718b1d092c4a15cbe27c38e2690.jpg", + "image_caption": [ + "6T: Cactus", + "Butterfly" + ], + "image_footnote": [], + "bbox": [ + 563, + 88, + 684, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4c1c319ed213e7a6af9aad8bb8a8c62ea4b039aabb19c9509095c7c4e05ee95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 89, + 767, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee1534eb5b4d886c4951623f1fbf12d1c5447f3988cac2437134d236471854d7.jpg", + "image_caption": [ + "GT: Mushroom" + ], + "image_footnote": [], + "bbox": [ + 771, + 88, + 888, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "classification or retrieval accuracy) on different evaluation datasets [33, 40, 79]. However, highlighting salient regions by backpropagating gradients for downstream applications does not capture how helpful end-users find these attributions [53, 73]. In this section, we take a human-centred approach to interpretability – how well our stroke attribution algorithms align with the reasoning process of humans and the trade-off, interpretability vs. accuracy.", + "bbox": [ + 75, + 171, + 468, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Setup: We recruited 7 participants from different geographical regions, in the age group 20 - 30 years. All participants had some background in AI research, but only 3 reported having prior experience in interpretability. Once recruited, each user is assigned a unique ID for anonymity. For SLA and P-SLA, we conduct 5 human studies, with each having 10 multiple choice questions (MCQs). Hence, each participant answers 50 MCQs for SLA and 50 for P-SLA.", + "bbox": [ + 75, + 296, + 470, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluating Transparency: For an attribution to be useful, humans must understand a model's behaviour for correct and incorrect predictions. In this section, we evaluate if our SLA and P-SLA can make existing sketch classifiers (i.e., Sketch-A-Net [111]), pre-trained on TU-Berlin [33] dataset transparent to humans. Accordingly, we choose a random category and select 4 sketch instances - (i) 3 misclassified and 1 correctly classified, and (ii) 1 misclassified and 3 correctly classified. Next, as shown in Fig. 8, we compute stroke attribution (SLA and P-SLA) for the selected sketches and ask users: \"Only 1 of these 4 sketches are correctly (or incorrectly) recognised by our model. Please select that correct (or incorrect) sketch.\" We find users can identify correct/incorrect model predictions $75.9\\% /63.4\\%$ of times for SLA and $76.3\\% /65.2\\%$ for P-SLA.", + "bbox": [ + 75, + 420, + 468, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q. Only one sketch is correctly categorised. Which one?", + "bbox": [ + 84, + 648, + 423, + 662 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df06294eb6bd9999b6c75663538cd46f9a2767d5865735c1a6d904e2bc219717.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 664, + 156, + 705 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbacf80aa0a4c755660163bd6308eb1731cac93aebbfd695849925c3e49662de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 664, + 253, + 705 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c35654e0be2f2813f7c613938e91df8f516a0281ae5ee1f3987bc360ce3630d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 664, + 344, + 705 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec686d1bfb4bf068059b84efd76c23efcbd38ac9b2e55701a0ee08c2cf862f8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 664, + 462, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q. Only one sketch is incorrectly categorised. Which one?", + "bbox": [ + 80, + 708, + 426, + 720 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe58f5eb1836c79254ebe16a0683accc90994209e8b1db65b7753307bafd9dc1.jpg", + "image_caption": [ + "Figure 8. Evaluating transparency: Can a human understand the behaviour of an existing (pre-trained) classifier with SLA/P-SLA." + ], + "image_footnote": [], + "bbox": [ + 84, + 722, + 156, + 766 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/802e104f3e8afafa29b293696fb752002a58b12a1015deba569f2009ea6e07e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 722, + 254, + 765 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a85669fd7d1de5d17bb2eee4edd26d9c947bf5f2d194544bbe90e541145d5f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 723, + 346, + 765 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb91e26f7fa6a93adf2487953ac86215e65bf52a3716626b862fb3b5209ab1b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 719, + 462, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluating Fairness: End users are much better positioned to make a decision with help from a model if intelligible explanations are provided. In this section, we evaluate if SLA and P-SLA can help end users understand \"What went wrong?\" (i) For a pre-trained sketch classifier [111], we show users a misclassified sketch instance and ask users", + "bbox": [ + 75, + 809, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to identify the (wrongly) predicted category. (ii) For fine-grained SBIR [13], we show a sketch (whose GT photo is not in top-10) and ask users to identify the top-1 (wrongly) retrieved photo in Fig. 9. Humans can identify the misclassified category $62.4\\%$ $(66.3\\%)$ and the incorrectly retrieved photo $39.1\\%$ $(37.2\\%)$ for SLA (P-SLA).", + "bbox": [ + 496, + 171, + 890, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q. Prediction is incorrect, identify predicted category", + "bbox": [ + 506, + 268, + 676, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q. Retrieval is incorrect, identify the top-1 retrieved image", + "bbox": [ + 686, + 268, + 880, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "② ${Na} + {O2} < \\Delta$ 释放", + "bbox": [ + 506, + 297, + 593, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(A) Angel (B)Butterfly", + "bbox": [ + 596, + 297, + 676, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(C) Fan (D) Dog", + "bbox": [ + 596, + 329, + 640, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d483d95960b9a87048f9e77371934cc57ec77a22ff3daf9b7a402b97350186cc.jpg", + "image_caption": [ + "Figure 9. Evaluating Fairness: For an incorrect model prediction, we evaluate if humans can \"identify what went wrong\"." + ], + "image_footnote": [], + "bbox": [ + 686, + 297, + 750, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9bc0a0b8552958f5393fb9398705b09175e1b100dfef0c43c021632aacbcee68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 297, + 800, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc0d47a686c86eeb1a7d3a5cded5b872ce73c13fa5bf25f3a1c7d87b47a5ed7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 297, + 854, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7d09586483de2d7165b23f739bc44328102e30512b948786bc7fcc9b229f6da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 867, + 311, + 882, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Evaluating Trustworthiness:** Determining trust in individual predictions is important when used for decision-making (e.g., medical diagnosis [57]). We train two copies of the same sketch classifier [111], a strong classifier with $73.8\\%$ accuracy on TU-Berlin [33] and a weak one reaching $57.1\\%$ . We present the stroke attribution (SLA/P-SLA) for both models and ask users to identify the strong/weak classifier, as shown in Fig. 10. Humans identify the stronger classifier $71.4\\% (68.7\\%)$ of the time for SLA (P-SLA).", + "bbox": [ + 496, + 407, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q. Prediction is correct, identify the stronger classifier", + "bbox": [ + 508, + 554, + 862, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c15ed2643b4751d5a0dd526bdf3f0e1147726e2c69b75b8ecfd8c9ac47adb6e6.jpg", + "image_caption": [ + "Figure 10. Evaluating trustworthiness: We present the stroke attributions (SLA/P-SLA) from two sketch classifiers (strong and weak), and ask users to identify the strong/weak model." + ], + "image_footnote": [], + "bbox": [ + 508, + 571, + 589, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10a97bcb851270199e12f9c43445718e8a17d93e3d7fc8886ca01ab5f1edd306.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 573, + 671, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79855170b30d002b8931b83b892766649f1974b88a087aa48c0160eb164a5c55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 573, + 689, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e3f1a99612e002d3faea49e09e9047e06263dbd7eb2e248bedd42e1e1ee1e59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 573, + 782, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c451294a4071bf72682eb5d01b9d66cfa6ab1dc4bc7819ca09c349fecd010f69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 573, + 883, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 691, + 617, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work emphasises the pivotal role of strokes in human-drawn sketches, offering unique insights compared to pixel-based images. Our lightweight explainability solution seamlessly integrates with pre-trained models, addressing rasterisation challenges and contributing to diverse sketch-related tasks. Through applications in Retrieval, Generation, Assisted Drawing, and Sketch Adversarial Attack, our model showcases adaptability and significance. The proposed stroke-level attribution provides nuanced insights into model behaviour, underscoring the importance of explainability in bridging human expression with model predictions in the evolving field of sketch interpretation.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11004", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 1, 2, 3", + "[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. arXiv preprint arXiv:2011.05429, 2020. 2", + "[3] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. arXiv preprint arXiv:2108.00401, 2021. 7", + "[4] Shm Garanganao Almeda, J.D. Zamfirescu-Pereira, Kyu Won Kim, Pradeep Mani Rathnam, and Bjoern Hartmann. Prompting for discovery: Flexible sense-making for ai art-making with dreamsheets. arXiv preprint arXiv:2310.09985, 2023. 2", + "[5] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104, 2017. 2", + "[6] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. A unified view of gradient-based attribution methods for deep neural networks. In NIPSW, 2017. 2", + "[7] Anurag Arnab, Ondrej Miksik, and Philip H.S. Torr. On the robustness of semantic segmentation models to adversarial attacks. In CVPR, 2018. 7", + "[8] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 2015. 2", + "[9] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In CVPR, 2024. 2", + "[10] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 2", + "[11] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 2", + "[12] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 5, 6", + "[13] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 8", + "[14] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang," + ], + "bbox": [ + 86, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Yi-Zhe Song. Doodle it yourself: Class incremental learning by drawing a few sketches. In CVPR, 2022. 2", + "[15] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 4, 5, 6", + "[16] Moritz Bohle, Mario Fritz, and Bernt Schiele. B-cos: Networks: Alignment is all we need for interpretability. In CVPR, 2022. 3", + "[17] Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well onImagenet. arXiv preprint arXiv:1904.00760, 2019. 2", + "[18] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In ICCV, 2015. 2", + "[19] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. In NeurIPS, 2019. 2", + "[20] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM TOG, 2020. 2", + "[21] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In ICCV, 2018. 2", + "[22] Xuesong Chen, Xiyu Yan, Feng Zheng, Yong Jiang, Shu-Tao Xia, Yong Zhao, and Rongrong Ji. One-shot adversarial attacks on visual tracking with dual attention. In CVPR, 2020. 7", + "[23] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 2, 5", + "[24] Pinaki Nath Chowdhury, Aneeshan Sain, Yulia Gryaditskaya, Ayan Kumar Bhunia, Tao Xiang, and Yi-Zhe Song. Fs-coco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2", + "[25] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation: Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 5", + "[26] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2, 5", + "[27] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In NeurIPS, 2017. 2", + "[28] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3", + "[29] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 5" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11005", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron? arXiv preprint arXiv:1805.12233, 2018. 2", + "[31] Gabriele Dominici, Pietro Barbiero, Lucie Charlotte Magister, Lio Pietro, and Nikola Simidjevski. Sharcs: Shared concept space for explainable multimodal learning. arXiv preprint arXiv:2307.00316, 2023. 3", + "[32] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 2", + "[33] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 4, 5, 7, 8", + "[34] Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. Technical Report, Université de Montréal, 2009. 2", + "[35] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. arXiv preprint arXiv:2112.04417, 2021. 2", + "[36] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In ICCV, 2019. 2", + "[37] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2", + "[38] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In CVPR, 2020. 2", + "[39] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In CVPR, 2021. 2", + "[40] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 3, 4, 7, 8", + "[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 6, 7", + "[42] Yang He, Shadi Rahimian, Bernt Schiele, and Mario Fritz. Segmentations-leak: Membership inference attacks and defenses in semantic image segmentation. In ECCV, 2020. 7", + "[43] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2", + "[44] Josh Holinaty, Alec Jacobson, and Fanny Chevalier. Supporting referenceimageryfordigitaldrawing. In ICCVW, 2021. 1", + "[45] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2", + "[46] Quan Huynh-Thu, Marie-Neige Garcia, Filippo Speranza, Philip Corriveau, and Alexander Raake. Study of ratling scales for subjective quality assessment of high definition video. IEEE TBC, 2010. 6", + "[47] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 5", + "[48] Yunhan Jia, Yantao Lu, Junjie Shen, Qi Alfred Chen, Hao Chen, Zhenyu Zhong, and Tao Wei. Fooling detection alone" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "is not enough: Adversarial attack against multiple object tracking. In ICLR, 2020. 7", + "[49] Zeyang Jia, Eli Ben-Michael, and Kosuke Imai. Bayesian safe policy learning with chance constrained optimization: Application to military security assessment during the vietnam war. arXiv preprint arXiv:2307.08840, 2023. 3", + "[50] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6", + "[51] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In ICCV, 2019. 2, 3", + "[52] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6", + "[53] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In ECCV, 2022. 2, 3, 8", + "[54] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, 2019. 2", + "[55] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1, 4, 5, 6, 7", + "[56] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 3", + "[57] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 8", + "[58] Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Sam Gershman, and Finale Doshi-Velez. An evaluation of the human-interpretability of explanation. arXiv preprint arXiv:1902.00006, 2019. 2", + "[59] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In WACV, 2019. 2", + "[60] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 7", + "[61] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, Yi-Zhe Song, and Yonggang Qi. Zero-shot everything sketch-based image retrieval, and in explainable style. In CVPR, 2023. 1, 2", + "[62] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Defloconet: Deep image editing via flexible low-level controls. In CVPR, 2021. 2", + "[63] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 5, 6" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11006", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Meike Nauta, Ron Van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In CVPR, 2021. 2", + "[65] Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. ICMLW, 2016. 2", + "[66] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. 2", + "[67] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 6", + "[68] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2", + "[69] Buu Phan, Fahim Mannan, and Felix Heide. Adversarial imaging pipelines. In CVPR, 2021. 7", + "[70] Zhiyu Qu, Yulia Gryaditskaya1, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In CVPR, 2023. 1, 2", + "[71] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 5", + "[72] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards better understanding attribution methods. In CVPR, 2022. 2", + "[73] Marco T. Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In KDD, 2016. 3, 7, 8", + "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Omer. High-resolution image synthesis with latent diffusion. In CVPR, 2022. 6", + "[75] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019. 3", + "[76] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 2", + "[77] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 1, 3, 5, 7", + "[78] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE TNNLS, 2016. 2", + "[79] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 4, 5, 8", + "[80] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Pronimal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5, 6", + "[81] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tra. Grad-cam: Visual explanations from deep networks via gradient-based localisation. In ICCV, 2017. 2, 3, 7", + "[82] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2", + "[83] Harshay Shah, Prateek Jain, and Praneeth Netrapalli. Do input gradients highlight discriminative features? In NeurIPS, 2021. 2", + "[84] Hamid Rahim Sheikh and Alan C. Bovik. Image information and visual quality. IEEE TIP, 2006. 6", + "[85] Yuming Shen, Li Liu, Fumin Shen, and Ling Shao. Zeroshot sketch-image hashing. In ICCV, 2018. 2", + "[86] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In ICML, 2017. 2, 3", + "[87] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In ICLRw, 2014. 2, 3", + "[88] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiβ, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. arXiv preprint arXiv:2204.11642, 2022. 2", + "[89] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 2", + "[90] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1", + "[91] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In ICLRw, 2015. 2, 3", + "[92] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2, 3", + "[93] Christian Szegedy, Wojciech Zaremba, Ilyua Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7", + "[94] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7", + "[95] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white- and black-box attacks. In NeurIPS, 2020. 7", + "[96] Aditay Tripathi, Rajath R Dani, Anand Mishra, and Anirban Chakraborty. Sketch-guided object localization in natural images. In ECCV, 2020. 2", + "[97] James Tu, Mengye Ren, Siva Manivasagam, Ming Liang, Bin Yang, Richard Du, Frank Cheng, and Raquel Urtasun. Physically realizable adversarial examples for lidar object detection. In CVPR, 2020. 7" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11007", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[98] Hugues Turbé, Mina Bjelogrlic, Christian Lovis, and Gianmarco Mengaldo. Evaluation of post-hoc interpretability methods in time-series classification. Nature Machine Intelligence, 2023. 1, 6", + "[99] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In CVPR, 2021. 2", + "[100] Donglai Wei, Joseph Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, 2018. 4", + "[101] Dennis Wei, Rahul Nair, Amit Dhurandhar, Kush R. Varshney, Elizabeth M. Daly, and Moninder Singh. On the safety of interpretable machine learning: A maximum deviation approach. In NeurlPS, 2022. 3", + "[102] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: deep sketch-based hair image synthesis. ACM TOG, 2021. 1", + "[103] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 5, 7", + "[104] Rui Xu, Zongyan Han, Le Hui, Jianjun Qian, and Jin Xie. Domain disentangled generative adversarial network for zero-shot sketch-based 3d shape retrieval. AAAI, 2022. 2", + "[105] Guowei Yan, Zhili Chen, Jimei Yang, and Huamin Wang. Interactive liquid splash modeling by user sketches. ACM TOG, 2020. 2", + "[106] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 6", + "[107] Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Suggala, David I Inouye, and Pradeep K Ravikumar. On the (in) fidelity and sensitivity of explanations. In NeurIPS, 2019. 2", + "[108] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A zero-shot framework for sketch based image retrieval. In ECCV, 2018. 5", + "[109] Ran Yi, Zipeng Ye, Ruoyu Fan, Yezhi Shu, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Animating portrait line drawings from a single face photo and a speech signal. In ACM SIGGRAPH, 2022. 2", + "[110] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In CVPR, 2019. 2", + "[111] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 3, 7, 8", + "[112] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 3, 5, 6", + "[113] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2, 3", + "[114] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In CVPR, 2022. 2", + "[115] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, 2019. 7", + "[116] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In CVPR, 2016. 5" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[117] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. IJCV, 2018. 2", + "[118] Lvmin Zhang, Anyi Rao, and Maneesh Agarwala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5, 6, 7", + "[119] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2", + "[120] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2", + "[121] Hang Zhou, Dongdong Chen, Jing Liao, Weiming Zhang, Kejiang Chen, Xiaoyi Dong, Kunlin Liu, Gang Hua, and Nenghai Yu. Lg-gan: Label guided adversarial network for flexible targeted attack of point cloud-based deep networks. In CVPR, 2020. 7", + "[122] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchscene: Richly-annotated scene sketches. In ECCV, 2018. 2" + ], + "bbox": [ + 501, + 90, + 890, + 388 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "11008", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_model.json b/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ada0cedc785b05bf24933e9229c7835baaa1833 --- /dev/null +++ b/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_model.json @@ -0,0 +1,3447 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.816, + 0.152 + ], + "angle": 0, + "content": "What Sketch Explainability Really Means for Downstream Tasks?" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.173, + 0.796, + 0.245 + ], + "angle": 0, + "content": "Hmrishav Bandyopadhyay1 Pinaki Nath Chowdhury1 Ayan Kumar Bhunia1 \nAneeshan Sain1 Tao Xiang1,2 Yi-Zhe Song1,2 \n1SketchX, CVSSP, University of Surrey, United Kingdom. \n2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence." + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.248, + 0.828, + 0.262 + ], + "angle": 0, + "content": "{h.bandyopadhyay, p.chowdhury, a.bhunia, a.sain, t.xiang, y.song}@surrey.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.318, + 0.474, + 0.591 + ], + "angle": 0, + "content": "In this paper, we explore the unique modality of sketch for explainability, emphasising the profound impact of human strokes compared to conventional pixel-oriented studies. Beyond explanations of network behavior, we discern the genuine implications of explainability across diverse downstream sketch-related tasks. We propose a lightweight and portable explainability solution - a seamless plugin that integrates effortlessly with any pre-trained model, eliminating the need for re-training. Demonstrating its adaptability, we present four applications: highly studied retrieval and generation, and completely novel assisted drawing and sketch adversarial attacks. The centrepiece to our solution is a stroke-level attribution map that takes different forms when linked with downstream tasks. By addressing the inherent non-differentiability of rasterisation, we enable explanations at both coarse stroke level (SLA) and partial stroke level (P-SLA), each with its advantages for specific downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.604, + 0.208, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Sketches, rooted in human expression [43], offer a distinctive modality for exploring explainability [61, 70]. In contrast to photos, where each pixel is independent and lacks inherent meaning, sketches are organised into strokes, with each stroke carrying subjective meaning assigned by the sketcher [44]. This paper explores sketch explainability, but with a unique perspective – aiming to provide explanations and unravel the true implications of explainability on various downstream sketch-related tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.856 + ], + "angle": 0, + "content": "With this perspective in mind, our approach champions an explainability solution that is (i) lightweight and portable - a plugin seamlessly integrating with multiple pre-trained models without necessitating re-training [98], and (ii) easily adaptable to a diverse array of downstream sketch-specific tasks, benefiting the broader community." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our solution is exclusively centred on human strokes, aiming to attribute explanation on different stroke granularity: individual strokes (coarse) and their parts (fine). The" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.288, + 0.891, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.63 + ], + "angle": 0, + "content": "Figure 1. We attribute explanations for individual strokes (stroke-level attribution) and their vector coordinate points (point-level attribution). Stroke-level attribution rasterises individual strokes (non-differentiably) to produce \\( n \\)-stroke images. Next, we sum the stroke images to get the complete sketch image used for downstream tasks. Point-level Attribution computes distance transform from stroke coordinates and thresholds to get the sketch image. Our explainability solution works without re-training for existing tasks like SBIR and sketch-to-photo generation and novel tasks like filtering noisy strokes for assisted drawing and adversarial attack by removing a small stroke." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.838 + ], + "angle": 0, + "content": "output of our model markedly differs from typical saliency maps [1] found in photo-based explainability models, where the emphasis is mostly on visualisation [90]. Ours is a task-driven attribution map that assigns stroke-level attributes capturing how altering stroke characteristics can impact model prediction. Depending on the downstream tasks, attributions can be grounded to, for example (i) importance of entire strokes, which is more suitable to filter noisy strokes [15] in assisted drawing, and remove small strokes for adversarial attacks on existing sketch encoders, and (ii) stroke shape and length, where a partial-stroke level attribution is beneficial for tasks like sketch-based image retrieval [77] and sketch-to-photo generation [55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To showcase the adaptability of our model, we carefully devise four applications: two well-studied tasks from existing literature (retrieval [12, 26] and generation [55, 102]), and two entirely novel tasks (assisted drawing and sketch" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10997" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "adversarial attack). In retrieval, we evaluate reliability of model predictions by comparing predicted stroke order with the order in which a human draws them. For generation, we pinpoint strokes with the least influence, offering explicit feedback to end-users regarding which strokes the model prioritised and which it overlooked. In assisted drawing [4], we assist novice artists in faithfully sketching a particular photo by identifying strokes that do not match the target photo. Lastly, in adversarial attacks, we unveil the vulnerability of state-of-the-art sketch encoders by removing a small imperceptible stroke in any sketch, resulting in significant changes to the model's prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.606 + ], + "angle": 0, + "content": "The focal point connecting all downstream tasks is our proposed stroke-level attribution. The key question, therefore, is how to backpropagate information to strokes while addressing the inherent non-differentiability of rasterisation - strokes are most often represented as discrete coordinates and rasterised before feeding into downstream applications. We provide two solutions for non-differentiability: (i) coarse stroke level: we first rasterise individual strokes to produce raster stroke images. Then, we combine these stroke images to get the complete sketch image (see Fig. 1) - since this addition of stroke images is a differentiable operation, we can backpropagate information from the complete sketch to individual raster stroke images. (ii) fine partial-stroke level: we create a distance transform image from stroke coordinates (\"red dots\" in Fig. 1 vector sketch) by calculating the minimum distance of each pixel in the image from the coordinates. Then, we threshold the distance value to get the sketch image (white pixels for a high distance and black pixels for a low distance). Since the distance function and our threshold step are both differentiable, we can backpropagate information from the sketch images to stroke coordinates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.473, + 0.73 + ], + "angle": 0, + "content": "Our contributions can be summarised as follows: (i) We explore sketch explainability, emphasising the importance of strokes in human-drawn sketches. (ii) We highlight the profound impact of explainability on various sketch-related domains, presenting applications in retrieval, generation, assisted Drawing, and adversarial attacks. (iii) We solve for the non-differentiability problem of rasterisation, and provide both stroke-level and partial-stroke level attribution." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.742, + 0.228, + 0.758 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Sketch for Visual Understanding: Having a high visual proximity [43] to real images and carrying human subjectivity [9, 76], amateur sketches or abstract line drawings [59] has been a popular modality for customised expression, thus driving extensive applications as a query for retrieval [26, 32, 76, 85] of object [29] or scene [24] images, 3D shapes [104], and even concepts like in 'dictionary-like' games [11]. As a canvas for creativity, sketch helps image-editing [62, 110, 114], or generation of objects [20, 21, 38, 99]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.258 + ], + "angle": 0, + "content": "scenes [109], and 3D shapes [10, 39, 105, 119]. Being easily editable, sketch enables interactive access to AI systems like image-segmentation [45, 122], object localisation [96], image-inpainting [110, 114], and incremental learning [14]. Being application-specific however, such works largely ignored explaining the 'how' of sketch-correspondence. The few who did, customised training pipelines [61, 70] for niche tasks. In this work, we thus make the first attempt at visualising salient sketch-regions (strokes), as an explainability-tool (like GradCAM [81] in photos) for existing pre-trained sketch-based downstream networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.895, + 0.626 + ], + "angle": 0, + "content": "Explaining CNN Predictions: CNN explanations visually highlight regions of input having the maximum influence on a model's predictions [87]. This visualisation of 'salient' regions is either through an analysis [37] of a regular pre-trained network after it has completed training (post-hoc), or by designing and training explicitly interpretable (i.e. explain-and-predict) models [17, 19, 64]. Given a pre-trained CNN, a post-hoc algorithm either visualises (i) model attributes like feature and activation-maps [34, 65, 66, 82, 113, 120] that imply saliency of specific input (pixel) regions, or (ii) input attributes directly as pixels [91] or pixel-regions [51] coloured according to their relative importance. Visualisation of input attributes is facilitated through perturbation based algorithms [18, 36, 37] and gradient-based analysis [87, 89, 91]. Perturbation-based algorithms [27, 68, 113] detect saliency of pixel-regions by measuring impact of their absence on the prediction score. Whereas, gradient-based algorithms [8, 30, 86, 92, 117] measure the gradient of the prediction with respect to individual input-pixels (input-gradients [83]), attributing them based on this value. Unlike images however, sketch is a sparse-information modality [23] for pixels. As such, we explore explainability in sketches by computing stroke and point attribution for fine-grained sketch explanations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Evaluation of CNN Explanations: The evaluation of CNN explanations has evolved over time - from naive qualitative analysis of visualisations [82, 87, 91, 113] to standardised theoretical [5, 92] and empirical [72, 107, 113] baselines. Analysing explanations theoretically helps evaluate them in a model-agnostic environment, where their mathematical form is checked against pre-defined properties (axioms)[5, 54, 92]. Empirical evaluations, on the other hand, involve experiments measuring \\((i)\\) variance in explanations upon perturbations of inputs [107] and model weights [1, 2] (sanity checks), and \\((ii)\\) accuracy of explanations in locating important features [37, 72, 82, 117]. These features, when perturbed, influence the CNN maximally, as measured by perturbation-based metrics [6, 51, 68, 78, 82, 83, 113]. Recently, however, these evaluation protocols have met criticism [35] due to their detachment from humans. Instead, human studies [53, 58, 82, 88], and human-centered metrics [35] have been proposed for evaluating explanation in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10998" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "terpretability. In this work, we attribute strokes by backpropagation [86, 87, 91, 92], evaluating attributions through sanity checks [1], empirical metrics [51] and human studies [53] on downstream sketch-based applications [112]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.16, + 0.208, + 0.176 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.178, + 0.469, + 0.224 + ], + "angle": 0, + "content": "Here, we provide a brief overview of some standard concepts, ubiquitous in explainability literature [1] to help formalise the question: \"what entails a good explanation?\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.469, + 0.362 + ], + "angle": 0, + "content": "Attribution Algorithms: It highlights relevant regions (e.g., pixels in an image, see Fig. 1) that are responsible for the model's prediction. Despite its importance for safety-critical applications [49, 75, 101], making an attribution algorithm interpretable to humans remains an open problem. Surprisingly, a more faithful attribution is usually less interpretable and vice-versa [91, 113]. Prior works [16, 81] study this trade-off between faithfulness vs. interpretability as an answer to: \"What makes a good visual explanation?\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.469, + 0.437 + ], + "angle": 0, + "content": "The attribution map \\(\\mathbb{A} \\in \\mathbb{R}^{H \\times W}\\) is typically calculated using gradients for an input \\(\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}\\) for a classification model \\(\\hat{y} = F_{\\theta}(\\mathrm{X}) \\in \\mathbb{R}^{C}\\), pre-trained on \\(C\\) categories. The gradients are a simple and good indicator of how much the model prediction changes for input \\(\\mathrm{X}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.441, + 0.469, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} = \\partial F _ {\\theta} (\\mathrm {X}) / \\partial \\mathrm {X} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.469, + 0.7 + ], + "angle": 0, + "content": "Interpretability: It is the ability of an attribution algorithm to provide a qualitative \"understanding\" for a model [73]. This \"understanding\" depends on the target audience, e.g., a human expert may interpret a small Bayesian network [56], but a layman is more comfortable with a weighted attention (or feature) map that highlights salient regions [1]. To evaluate the interpretability of attribution maps, prior works either (i) perform downstream tasks [31, 81] that depend on the interpretation (e.g., object localisation – predict the bounding box and semantic segmentation for image region with the highest attribution) or (ii) human studies [53], typically conducted in two setups – class discriminative (given an attribution map, ask users to identify the category predicted by a model), and trustworthiness (compare attribution maps from a strong and a weak model, and ask users to identify the stronger model)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Faithfulness: It is the ability of attribution algorithms to accurately \"explain\" the computation learned by a model. For example, in theory, a fully faithful attribution is the entire model (e.g., ResNet-18 [41]) but is not interpretable by a human. In practice, for an attribution to be meaningful, it is often impossible to be completely faithful. To balance this trade-off, prior works explore human-interpretable attributions that are locally faithful for a given model prediction. One approach is image occlusion [73], where the difference in the model scores is measured when masking different patches in an input image. Image patches that significantly change the model score are deemed important by the attribution algorithm." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.673, + 0.107 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.342 + ], + "angle": 0, + "content": "The attribution map \\(\\mathbb{A}\\) in Eq. (1) gives a faithful fine-grained explanation for each pixel in X. However, such pixel attribution does not provide meaningful explanations when interpreting sparse human-drawn sketches (many empty white pixels). Additionally, pixel attribution does not consider the sketch construction process - humans sketch a sequence of strokes, not pixels. Hence, it is most appropriate that sketch predictions should be attributed to strokes and not pixels. However, the key challenge to stroke attributions is that most sketch applications [77, 118] use raster sketches - a non-differentiable process to convert a sequence of strokes into pixels. In the following sections, we propose two stroke attribution algorithms that consider the sketch construction process by designing a differentiable rasterisation pipeline for a vector sequence of strokes." + }, + { + "type": "code_caption", + "bbox": [ + 0.513, + 0.355, + 0.819, + 0.368 + ], + "angle": 0, + "content": "Algorithm 1: Non-differentiable Rasterisation" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.372, + 0.868, + 0.65 + ], + "angle": 0, + "content": "Data: V \\(\\leftarrow\\) Vector Sketch of size \\(\\mathbb{R}^{T\\times 5}\\) \nResult: X \\(\\leftarrow\\) Blank (Zero) Canvas of size \\(\\mathbb{R}^{h\\times w\\times 3}\\) . \n\\(\\mathbb{B}(\\cdot ,\\cdot)\\gets\\) Bresenham Function ; \n\\(v_{0} = (x_{0},y_{0},q_{0}^{1},q_{0}^{2},q_{0}^{3})\\gets \\mathrm{V}[0]\\) . \n\\(v_{\\mathrm{prev}}\\gets v_0\\) . \n\\(q_{\\mathrm{prev}}\\gets q_0^1\\) . \nfor \\(v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})\\gets \\mathrm{V}[1\\dots T]\\) do if \\(q_{\\mathrm{prev}} = 1\\) and \\(q_t^1 = 1\\) then for \\((p_x^i,p_y^i)\\gets \\mathbb{B}(v_{\\mathrm{prev}},v_t)\\) do |X \\((p_x^i,p_y^i)\\gets 255;\\) end end \\(v_{\\mathrm{prev}}\\gets v_t\\) . \\(q_{\\mathrm{prev}}\\gets q_t^1\\) . if \\(q_{t}^{3} = 1\\) then exit(); /\\* End of Drawing \\*/ end" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.716, + 0.683 + ], + "angle": 0, + "content": "4.1. Sketch Representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.84 + ], + "angle": 0, + "content": "While sketches are used in several formats (or representations) like Raster [111], Vector [40], or Bézier [28], digital sketches are primarily captured in vector form, as a list of points traced on a drawing pad. These points are usually a five-element vector \\( v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3}) \\) where, \\( (x_{t},y_{t}) \\) are absolute coordinates in a \\( (H\\times W) \\) drawing canvas and the last three elements are one-hot encoding of pen-states: pen touching the paper \\( (1,0,0) \\), pen is lifted \\( (0,1,0) \\), and end of drawing \\( (0,0,1) \\). Hence, a vector sketch with \\( T \\) points is represented as \\( \\mathbf{V}\\in \\mathbb{R}^{T\\times 5} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As most downstream sketch applications work on rasterised sketches \\( \\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3} \\), prior works [111] translate (non-differently) a vector sketch \\( \\mathrm{V} \\) to its equivalent raster sketch \\( \\mathrm{X} \\) using Algorithm 1. For this, the Bre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10999" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.462, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.213 + ], + "angle": 0, + "content": "Figure 2. Coarse Stroke-level Attribution. Backpropagate gradients from raster sketch X to raster strokes \\( S_{i} \\), with weight \\( \\omega_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.314 + ], + "angle": 0, + "content": "senham function \\(\\mathbb{B}(\\cdot)\\) is used to connect two vector points \\(\\{v_{t - 1}, v_t\\} \\in \\mathrm{V}\\) in the pixel space \\(\\{(p_x^1, p_y^1), \\ldots, (p_x^n, p_y^n)\\} \\in \\mathbb{B}(v_{t - 1}, v_t)\\) via a continuous line. Next, we show how our sketch attributions overcome this non-differentiable rasterisation and backpropagate gradients (Eq. (1)) from pixels \\(\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}\\) to strokes and points in \\(\\mathrm{V} \\in \\mathbb{R}^{T \\times 5}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.406, + 0.342 + ], + "angle": 0, + "content": "4.2. Coarse Stroke-level Attribution (SLA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.471, + 0.457 + ], + "angle": 0, + "content": "In this section, we backpropagate gradients from pixel space in \\( \\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3} \\) to strokes, defined as a continuous set of points \\( \\{v_t, v_{t+1}, \\ldots, v_{t+n}\\} \\) from the first pen-down \\( (1,0,0) \\) till the pen-up \\( (0,1,0) \\) state. Algorithm 1 converts the vector stroke points into a raster stroke \\( S_i \\in \\mathbb{R}^{H \\times W \\times 3} \\). The final raster sketch is then a differentiable composition of \\( m \\) raster strokes \\( \\mathrm{X} = \\sum_{k=1}^{m} S_k \\). We compute SLA as," + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.469, + 0.47, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathcal {S} _ {i}} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.47, + 0.575 + ], + "angle": 0, + "content": "This, however, gives a degenerate solution where all strokes will have the same attribution \\(\\partial \\sum_{k=1}^{m} S_k / \\partial(S_i) = 1\\). To avoid this, we compute a weight factor \\(\\omega_i \\in \\mathbb{R}^{H \\times W \\times 3}\\) for each stroke \\(S_i\\) such that," + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.587, + 0.47, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\omega_ {i} \\left(p _ {x}, p _ {y}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\left(p _ {x}, p _ {y}\\right) \\in \\mathbb {B} \\left(v _ {t - 1}, v _ {t}\\right) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.716 + ], + "angle": 0, + "content": "In other words, given consecutive vector sequence of points \\((v_{t - 1}, v_t)\\) in stroke \\(S_i\\), we find all points \\((p_x, p_y)\\) using Bresenham function \\(\\mathbb{B}(\\cdot)\\) that lie \"on the stroke\" and assign \\(\\omega_i(p_x, p_y) = 1\\). Hence, longer strokes will have more 1's compared to shorter strokes. Finally, Eq. (2) is adapted as" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.727, + 0.469, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.864 + ], + "angle": 0, + "content": "SLA makes the non-differentiable rasterisation \\((\\mathrm{V} \\to \\mathrm{X})\\) partially differentiable (stokes \\(S\\) to sketch X). In other words, SLA answers \"Which strokes in a sketch are important\". Next, we make the rasterisation fully differentiable and backpropagate gradients to vector points V to answer \"Which point in a sketch is important\"." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.468, + 0.903 + ], + "angle": 0, + "content": "For overlapping strokes in \\( S_{i} \\) and \\( S_{j} \\), we clamp the maximum pixel value using differentiable functions like torch.clamp" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.875, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.191, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Figure 3. Partial Stroke-level Attribution. Backpropagate gradients from raster sketch X to vector sequence of coordinates V." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.223, + 0.845, + 0.24 + ], + "angle": 0, + "content": "4.3. Partial Stroke-level Attribution (P-SLA)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.246, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Unlike SLA, which partially captures the sketch construction process (stroke-level), partial stroke-level attribution (P-SLA) can fully backpropagate gradients to the vector list of coordinates \\( \\mathrm{V} \\in \\mathbb{R}^{T \\times 5} \\) traced on a drawing pad. Given a blank canvas \\( \\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3} \\), we (i) calculate the minimum distance of every pixel \\( (p_x, p_y) \\) in \\( \\mathrm{X} \\) from a line segment \\( (v_{t-1}, v_t) \\) in \\( \\mathrm{V} \\), and (ii) compute the pixel intensity of \\( \\mathrm{X}(p_x, p_y) \\) as function of the minimum distance as" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.378, + 0.891, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {X} \\left(p _ {x}, p _ {y}\\right) = \\sigma [ 2 - 5 \\min _ {t = 2, \\dots , T} (5) \\\\ \\left. \\mathrm {d i s t} ((p _ {x}, p _ {y}), v _ {t - 1}, v _ {t}) + (1 - q _ {t - 1} ^ {1}) 1 0 ^ {6}) \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.435, + 0.892, + 0.526 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the sigmoid function, and \\(\\mathrm{dist}(\\cdot)\\) is a distance function (see Supp.) from a point \\((p_x,p_y)\\) to a line segment \\((v_{t - 1},v_t)\\). For pen-up states \\((q_{t - 1}^{1} = 0)\\), we blow up \\((\\times 10^{6})\\) the distance values that make the pixel intensities \\(\\mathrm{X}(p_x,p_y)\\to 0\\), i.e., not render strokes for \\((v_{t - 1},v_t)\\). Finally, we compute P-SLA" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.536, + 0.892, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial v _ {t}} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.819, + 0.604 + ], + "angle": 0, + "content": "5. Applications of Stroke Attributions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Despite its simplicity, designing stroke attribution algorithms that capture the sketch construction process unlocks insights into numerous existing downstream tasks like classification [40], robust sketch-based image retrieval (category-level [33] and fine-grained [79]) and enables some novel sketch applications like assisted drawing [15], interactive sketch to photo generation [55], adversarial attacks on human-drawn sketches, and discovering the \"arrow of time\" [100] in raster sketches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.825, + 0.773 + ], + "angle": 0, + "content": "5.1. Robust Sketch Based Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a query sketch \\( \\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3} \\), category-level sketch-based image retrieval (SBIR) aims to fetch category-specific photos from a gallery of multi-category photos (e.g., given sketch of a 'shoe' retrieve any photo 'shoe' from a gallery of 'shoes+hats+cows'). Conversely, fine-grained SBIR aims to retrieve one instance from a gallery of the same category photos (e.g., given the sketch of a 'shoe' retrieve one photo shoe from a gallery of all shoes). Deep learning frameworks" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11000" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.148 + ], + "angle": 0, + "content": "Table 1. Stroke attribution (SLA, P-SLA) make SBIR systems reliable. Sketches with a high correlation \\((Corr)\\) of stroke saliency (predicted by SLA or P-SLA) with human-drawn temporal stroke order tend to have higher retrieval accuracy." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.156, + 0.472, + 0.242 + ], + "angle": 0, + "content": "
MetricsFull DatasetCorr ≥ 0.5Corr ≤ 0.1
SLAP-SLASLAP-SLA
CategorymAP53.155.357.651.750.1
LevelP@20065.966.768.564.661.5
FineAcc.@115.316.417.613.812.7
GrainedAcc.@534.236.939.431.128.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.334 + ], + "angle": 0, + "content": "learn a joint sketch-photo manifold (for category and fine-grained) via a feature extractor [26, 29, 103] trained using triplet loss [112]. Recent adoption of foundation models for SBIR [77] shifts focus to robust deployment using the open-set generalisation of CLIP [71]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.595 + ], + "angle": 0, + "content": "Towards this goal of robust deployment, our sketch attribution algorithms \\((\\mathbb{A}_i^R\\) and \\(\\mathbb{A}_t^V)\\) can predict which strokes the network focuses on when retrieving a photo (Fig. 4). Apart from interpreting SBIR models, sketch attribution can also help detect potential failures at runtime (inference). First, we use the attribution scores \\(\\mathbb{A}_i^R\\) or \\(\\mathbb{A}_t^V\\) to rearrange the strokes from highest to lowest. The attribution scores indicate the most salient to the least salient strokes that affect model prediction. Second, we calculate a correlation \\((Corr)\\) of our predictor stroke order with the ground-truth temporal stroke order drawn by a user (humans draw the most salient regions first and least salient areas last [33, 79]). A high correlation indicates that humans and our model prioritise strokes similarly, whereas a low \\(Corr\\) denotes that the model and the user prioritise different strokes. We evaluate SBIR on a pre-trained SOTA model [77] using CLIP with prompt learning as a sketch and photo encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Datasets: We use TU-Berlin [33] (for category-level SBIR) and Sketchy [79] (for fine-grained SBIR). TU-Berlin contains 250 categories, with 80 free-hand sketches in each, and 204, 489 images [116]. Sketchy [79] has 75, 471 sketches over 125 categories having 100 images in each [108]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Evaluation Metrics: Following [77], we use mean average precision (mAP) and precision for top 200 retrieved samples \\((\\mathbb{P}@\\mathbb{200})\\) for category-level SBIR. For fine-grained SBIR [23], we measure Acc.@q, i.e., the percentage of sketches whose true matched photo is in the top-q list." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.768, + 0.473, + 0.906 + ], + "angle": 0, + "content": "Results: We divide the evaluation set into two sets: (i) those that have a high correlation \\( Corr \\geq 0.5 \\), and (ii) those with a low correlation \\( Corr \\leq 0.1 \\) of ground-truth and predicted stroke order. Tab. 1 shows sketches with a high \\( Corr \\geq 0.5 \\) are 1.7/3.9 more accurate in Acc.@1/Acc.@5 than those with \\( Corr \\leq 0.1 \\) for fine-grained SBIR, and 3.3/1.7 better in mAP/P@200 for category-level SBIR. Full Dataset indicates the performance of the pre-trained model on the entire evaluation set." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.093, + 0.891, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Sketch attributions from stroke-level and point-level for image retrieval. High correlation of human-drawn stroke order with that from sketch-attributions (high \\(\\rightarrow\\) low) indicate our sketch encoder gives more importance to salient strokes drawn early on." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.342, + 0.874, + 0.358 + ], + "angle": 0, + "content": "5.2. Assisted Drawing via Noisy Stroke Removal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.479 + ], + "angle": 0, + "content": "Although sketching has enabled many exciting applications [25, 55, 63, 118], the fear to sketch (i.e., \"I can't sketch\") has proven fatal for its widespread adoption. To solve this, prior works [15] used complex (and hard to train [12]) reinforcement learning [80] to predict the importance of each stroke in a sketch. Next, a stroke subset selector removes noisy (less important) strokes, leaving only those positively contributing to the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.54 + ], + "angle": 0, + "content": "In this section, we focus on assisted drawing - given a photo, we use attribution scores \\((\\mathbb{A}_i^R\\) or \\(\\mathbb{A}_t^V)\\) to help humans draw a faithful and clean sketch. Our method is significantly simpler than reinforcement learning alternatives [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.894, + 0.632 + ], + "angle": 0, + "content": "For SLA, an input sketch \\( \\mathrm{X} = \\sum_{k=1}^{m} \\omega_k S_k \\) is composed of \\( m \\) strokes. We calculate the cosine similarity (sim) of input sketch \\( \\mathrm{X} \\) with its target photo \\( \\mathrm{P} \\in \\mathbb{R}^{H \\times W \\times 3} \\) to measure \"how faithfully \\( \\mathrm{X} \\) describes \\( \\mathrm{P} \\)\". Next, we backpropagate gradients from the cosine similarity to strokes \\( S_i \\) and calculate stroke-level attribution score \\( \\mathbb{A}_i^R \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.642, + 0.892, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial \\operatorname {s i m} (F _ {\\theta} (\\mathrm {X}) , F _ {\\theta} (\\mathrm {P}))}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.765 + ], + "angle": 0, + "content": "The pre-trained sketch and photo encoder \\(F_{\\theta}(\\cdot)\\) must be highly accurate to judge sketch-photo correspondence. Hence, we use pre-trained CLIP+prompts encoder from [77]. To remove noisy strokes, we only update the weights \\(\\omega_{i}\\in \\mathbb{R}^{H\\times W\\times 3}\\) using a normalised attribution score \\(\\mathbb{A}_i^R\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.774, + 0.892, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\omega_ {i} ^ {*} = \\omega_ {i} \\cdot \\text {G u m b e l . S o f t m a x} \\left(\\frac {\\mathbb {A} _ {i} ^ {R}}{\\sum_ {\\forall i} \\mathbb {A} _ {i} ^ {R}} + \\Delta\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Gumbel_Softmax(·) makes the output one-hot (discrete value) in the forward pass but differentiable with a probability distribution that sum to 1 in the backward pass [47]. The" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2The sketch and photo encoder \\(F_{\\theta}(\\cdot)\\) could be a siamese-style shared network or two independent models with different network weights [77]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11001" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.434, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Figure 5. Assisted drawing via sketch healing (or filtering noisy strokes) using stroke attributions from SLA and P-SLA. This helps users having fear-to-sketch (\"I can't sketch\")." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.469, + 0.308 + ], + "angle": 0, + "content": "modified sketch is constructed as \\( \\mathrm{X} = \\omega_1^* S_1 + \\dots + \\omega_m^* S_m \\). Intuitively, we keep strokes that contribute \\( (\\mathbb{A}_i^R) \\) to a high cosine similarity matching human sketch \\( \\mathrm{X} \\) and target photo \\( \\mathrm{P} \\) and remove \\( S_i \\) with normalised \\( \\mathbb{A}_i^R \\) lower than \\( (0.5 - \\Delta) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.469, + 0.368 + ], + "angle": 0, + "content": "Similar to SLA, we can also use P-SLA to compute the attribution \\(\\mathbb{A}_t^V\\) for each point \\(v_{t}\\) in the vector sketch \\(\\mathrm{V}\\in \\mathbb{R}^{H\\times W\\times 3}\\), by measuring the cosine similarity between input X and target photo P as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.373, + 0.47, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial \\operatorname {s i m} \\left(F _ {\\theta} (\\mathrm {X}) , F _ {\\theta} (\\mathrm {I})\\right)}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.466 + ], + "angle": 0, + "content": "We remove noisy points \\( v_{t} \\) by updating the pen-states in Eq. (5) from pen-down \\( (1,0,0) \\) to pen-up \\( (0,1,0) \\) depending on its attribution \\( \\mathbb{A}_t^V \\) for point \\( v_{t}\\in \\mathrm{V} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.47, + 0.47, + 0.507 + ], + "angle": 0, + "content": "\\[\nq _ {t - 1} ^ {1 *} = q _ {t - 1} ^ {1} \\cdot \\text {G u m b e l . S o f t m a x} \\left(\\frac {\\mathbb {A} _ {t - 1} ^ {V}}{\\sum_ {\\forall t} \\mathbb {A} _ {t} ^ {V}}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.469, + 0.634 + ], + "angle": 0, + "content": "Using updated values for \\( q_{t-1}^{1*} \\), we update \\( q_{t-1}^{2*} = 1 - q_{t-1}^{1*} \\) and recalculate pixel intensities for raster sketch \\( \\mathrm{X}(p_x, p_y) \\). The value of hyper-parameter \\( \\Delta \\) significantly affects the stroke removal process. We found the optimal \\( \\Delta \\) for SLA and P-SLA is 0.3 and 0.1, respectively. A higher \\( \\Delta \\) for P-SLA gives broken lines with a drop in the visual quality of an input sketch. Next, we evaluate stroke filtering using SLA and P-SLA on popular human-drawn sketch datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Dataset: For a fair comparison with prior works [15], we evaluate on fine-grained SBIR datasets QMUL-ShoeV2 and QMUL-Chair-V2 [12, 67, 112]. It consists of 6,730/1,800 sketches and 2,000/400 photos from ShoeV2/Chair-V2. We evaluate on the standard test-split of 679/525 sketches and 200/100 photos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.469, + 0.852 + ], + "angle": 0, + "content": "Evaluation Metric: We measure the retrieval accuracy of the clean sketch with the target photo by computing Acc.@1 and Acc.@5. A high accuracy need not correspond to high visual quality to the human eye [84]. Hence, we conducted a small human study with 5 participants and reported the mean opinion score (MOS) [46]; each was asked to compare two sets of 50 sketch pairs (GT sketch vs SLA filtered) and (GT sketch vs P-SLA filtered)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Results: Removing noisy strokes from human-drawn sketches is still a new topic; hence, to the best of our knowledge, there is only one work by Bhunia et al. [15]. From" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.089, + 0.888, + 0.104 + ], + "angle": 0, + "content": "Table 2. Noisy stroke removal using SLA and P-SLA attribution." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.108, + 0.892, + 0.212 + ], + "angle": 0, + "content": "
MetricsGT SketchSLA filteredP-SLA filteredBhunia et al. [15]
Shoe-V2Acc.@133.436.136.543.7
Acc.@567.868.769.374.9
MOS28.685.757.1-
Chair-V2Acc.@153.354.956.564.8
Acc.@574.376.677.179.1
MOS35.871.457.1-
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Tab. 2, both Bhunia et al. [15], and ours improve fine-grained SBIR performance by \\(10.3\\%\\) and \\(3.1\\%\\), respectively. However, [15] outperforms our SLA and P-SLA filtered methods by \\(7.6\\%\\) and \\(7.2\\%\\), respectively. This performance gap is likely because [15] trains the baseline model [12] using actor-critic version of PPO [80], whereas our SLA and P-SLA work post-hoc [98] without training the baseline model [12]. Additionally, Bhunia et al. [15] aims to design a robust SBIR pipeline, whereas our SLA/P-SLA filtering aims to assist humans in drawing sketches. For human study (MOS): (i) users prefer SLA-filtered \\(78.5\\%\\) vs. \\(21.5\\%\\) for GT sketch, (ii) however, P-SLA-filtered are preferred only \\(57.1\\%\\) vs. \\(42.9\\%\\) for GT sketch. This is verified by Fig. 5 where P-SLA-filtered sketches have broken strokes that degrade their visual quality." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.462, + 0.842, + 0.477 + ], + "angle": 0, + "content": "5.3. Interactive Sketch To Photo Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.636 + ], + "angle": 0, + "content": "The upsurge of large-scale image generation models (e.g., Stable-Diffusion [74], GigaGAN [50]) helped develop sketch-conditional image generation [63, 118]. However, a key limitation of conditional image generation is that these models do not always faithfully follow the input condition. This was resolved in two stages for text-to-image generation: (i) find word tokens with low influence on generated image, and (ii) iteratively update its activation until it reaches a minimum required value. In this section, we design a pipeline for faithful sketch-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Using sketch attributions from SLA and P-SLA, we design a post-hoc [98] method that gives feedback to the user - which strokes the model focuses on and which strokes are being ignored. Given this feedback, a user can interact with the system to ensure the model attends all salient regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our interactive pipeline is built on top of a pre-trained sketch-to-photo generation model [55], comprising a modified ResNet-50 [41] as sketch encoder and StyleGAN [52] as image decoder. Given a raster sketch X, the modified sketch encoder computes a latent vector \\( z_{s2p}^{+} = F_{\\theta}(\\mathrm{X}) \\) where \\( z_{s2p}^{+} \\in \\mathbb{R}^{14 \\times 512} \\). Next, the StyleGAN [52] decoder generates the underlying image from \\( z_{s2p}^{+} \\). For a faithful sketch-to-image generation, we measure the \"influence\" of each stroke/vector point on the latent code \\( z_{s2p}^{+} \\). Particularly, prior works [106] suggest that \\( z_{s2p}^{+} \\) is disentangled into 14-level semantic feature hierarchy, where \\( z_{s2p}^{+1} \\in \\mathbb{R}^{512} \\) has coarse-level features controlling major semantic struc" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11002" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.471, + 0.298 + ], + "angle": 0, + "content": "Figure 6. Interactive Sketch to Photo Generation: Our stroke attribution algorithms make existing sketch-to-photo generation pipelines [55] more faithful. We achieve this by computing the stroke-level \\(\\mathbb{A}_i^S\\) or coordinate-level \\(\\mathbb{A}_t^V\\) attribution that has a maximal influence on the latent code \\(z_{s2p}^{+}\\) used by the image decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.371 + ], + "angle": 0, + "content": "tures and \\(z_{s2p}^{+14}\\in \\mathbb{R}^{512}\\) has fine-level features controlling colour schemes, etc. Since sketches primarily convey semantic structure, we use the sum of the first 7-layers to compute stroke attribution \\(\\mathbb{A}_i^R\\) or \\(\\mathbb{A}_t^V\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.38, + 0.36, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial \\sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.414, + 0.469, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial \\sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.505 + ], + "angle": 0, + "content": "Next, we qualitatively evaluate our iterative sketch to photo generation pipeline, as shown in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.426, + 0.53 + ], + "angle": 0, + "content": "5.4. Adversarial Attacks on Human Sketches" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Szegedy et al. [94] discovered that predictions by deep networks can be manipulated with extremely low-magnitude input perturbations. For images, these can be restricted to be imperceptible to human vision, but their effect can completely change the output prediction by a deep network. Such adversarial attacks are possible in image classification [94], semantic segmentation [7, 42], object detection [97, 115], object tracking [22, 48], etc. Studying these quirks is crucial as it can pose a real threat to deep learning as a pragmatic technology [3]. While major work has been dedicated to attacks on images, the recent surge of deployable sketch applications [77, 118] motivates us to present the first study on adversarial attacks for human-sketches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We show how our stroke attribution algorithms (SLA and P-SLA) provide the necessary information for adversarial attacks. For brevity, we focus on adversarial attacks on sketch classification [40, 111]. Intuitively, we use sketch attribution to remove the smallest stroke (in SLA Attack) and minimum number of points (in P-SLA Attack), yet have the maximum impact on changing prediction of a pre-trained classifier \\( F_{\\theta}(\\mathrm{X}) = y_{\\mathrm{cls}} \\). Ours is (i) a white-box [69, 95] setting – we have access to network weights and gradients of our ResNet-18 [41] classifier, pre-trained on QuickDraw [40] or TU-Berlin [33]; (ii) untargeted attack – while tar" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Table 3. Sketch Adversarial Attacks: Using stroke attributions, we remove a small stroke \\( (|S_i| \\leq \\epsilon) \\) that misclassifies an input sketch." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.125, + 0.892, + 0.184 + ], + "angle": 0, + "content": "
No AttackSLA AttackP-SLA Attack
ε = 5ε = 15ε = 5ε = 15
QuickDraw67.265.764.565.163.7
TU-Berlin74.971.568.570.268.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.894, + 0.318 + ], + "angle": 0, + "content": "geted attacks [60, 121] misclassify \\( F_{\\theta}(\\cdot) \\) from \\( y_{\\mathrm{cls}}^{\\mathrm{GT}} \\) to a specific target class \\( y_{\\mathrm{cls}}^{*} \\), untargeted attacks [93] aim to misclassify to any arbitrary class \\( y_{\\mathrm{cls}}^{\\mathrm{GT}} \\neq y_{\\mathrm{cls}} \\). For a neater description of SLA and P-SLA attacks, we define the rasterisation process using \\( \\mathcal{R}(\\cdot) \\) as \\( \\mathrm{X} = \\omega_1\\mathcal{S}_1 + \\dots + \\omega_m\\mathcal{S}_m = \\mathcal{R}(\\mathcal{S}) \\) for SLA and following Eq. (5) for P-SLA we define \\( \\mathrm{X} = \\mathcal{R}(\\{v_1,\\ldots v_T\\}) = \\mathcal{R}(\\mathrm{V}) \\). Next, we find a stroke \\( S_{adv} \\) with stroke length \\( |S_i| \\) less than some threshold \\( \\epsilon \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.327, + 0.892, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {a d v} = \\arg \\max _ {| \\mathcal {S} _ {j} | \\leq \\epsilon} \\mathcal {L} _ {\\mathrm {c l s}} \\left(F _ {\\theta} \\left(\\mathcal {R} \\left(\\mathcal {S} - \\left\\{\\mathcal {S} _ {j} \\right\\}\\right)\\right), y _ {\\mathrm {c l s}} ^ {\\mathrm {G T}}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.364, + 0.893, + 0.472 + ], + "angle": 0, + "content": "Unlike typical adversarial attacks on images that add a small noise \\((\\mathrm{X} + \\Delta x)\\) with \\(||\\Delta x||_{\\infty}\\leq \\epsilon\\) , for sketch adversarial attacks, we remove a small stroke \\((\\mathrm{X} - S_{adv})\\) such that the stroke length is less than \\(\\epsilon\\) as \\(|\\mathcal{S}_{adv}|\\leq \\epsilon\\) . For PSLA attack, we find a subset of \\(\\epsilon\\) vector points \\(\\mathrm{V}_{adv} =\\) \\(\\{v_1^{adv},\\dots v_\\epsilon^{adv}\\}\\) from input sketch \\(\\mathrm{V}\\in \\mathbb{R}^{T\\times 5}\\) which maximises the categorical cross-entropy loss \\(\\mathcal{L}_{\\mathrm{cls}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.481, + 0.892, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} v _ {t} ^ {a d v} = \\mathcal {L} _ {\\mathrm {c l s}} \\left(F _ {\\theta} \\left(\\mathcal {R} (\\mathrm {V} - \\{v _ {t} \\})\\right), y _ {\\mathrm {c l s}} ^ {\\mathrm {G T}}\\right) \\tag {13} \\\\ \\mathrm {V} _ {a d v} = \\operatorname {t o p} @ k \\left(\\left\\{v _ {1} ^ {a d v}, v _ {2} ^ {a d v}, \\dots v _ {T} ^ {a d v} \\right\\}, \\epsilon\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.532, + 0.892, + 0.564 + ], + "angle": 0, + "content": "where, \\(\\mathrm{top}@\\mathrm{k}(\\cdot ,\\epsilon)\\) picks the highest \\(\\epsilon\\) elements. Fig. 7 shows the adversarial strokes \\(S_{adv}\\) and points \\(\\mathrm{V}_{adv}\\) in red." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.657 + ], + "angle": 0, + "content": "Dataset: We evaluate sketch adversarial attacks on QuickDraw [40] and TU-Berlin [33]. We use a subset [103] of \\(50M\\) sketches in QuickDraw [40] as \\(3.8M\\) samples across 345 categories split in \\(2.1M\\) sketches for training, \\(0.3M\\) for validation and \\(0.4M\\) for evaluation. See Sec. 5.1 for details on TU-Berlin [33] dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.797 + ], + "angle": 0, + "content": "Evaluation: We measure the drop in classification accuracy when using SLA and P-SLA attacks in Tab. 3. Unlike image-based adversarial attacks where \\(\\epsilon\\) is a pixel intensity (non-integer, decimal value), our sketch attacks occur on stroke/point length, making \\(\\epsilon\\) an integer. A higher \\(\\epsilon \\geq 20\\) removes \"visible\" strokes in SLA and broken lines in P-SLA (Fig. 7). Hence, we evaluate accuracy drop for \\(\\epsilon = 5\\) and \\(\\epsilon = 15\\) in Tab. 3. We observe for \\(\\epsilon = 5\\) P-SLA offers a better adversarial attack than SLA by a margin of \\(1.3\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.642, + 0.824 + ], + "angle": 0, + "content": "6. Human Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Interpretability aims to help humans understand a model's reasoning process (transparency), verify that its predictions are based on the right constraints (fairness), and evaluate its confidence (trustworthiness) [73, 81]. In Sec. 5, we evaluated interpretability using several automatic metrics (e.g.," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11003" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.282, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.362, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.481, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.137, + 0.412, + 0.144 + ], + "angle": 0, + "content": "GT:Angel" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.09, + 0.565, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.089, + 0.685, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.137, + 0.617, + 0.144 + ], + "angle": 0, + "content": "6T: Cactus" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.09, + 0.769, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.137, + 0.672, + 0.144 + ], + "angle": 0, + "content": "Butterfly" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.089, + 0.89, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.137, + 0.828, + 0.144 + ], + "angle": 0, + "content": "GT: Mushroom" + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.153, + 0.85, + 0.167 + ], + "angle": 0, + "content": "Figure 7. Adversarial attacks on human drawn sketches using SLA and P-SLA. The adversarial strokes are marked in RED." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.172, + 0.47, + 0.294 + ], + "angle": 0, + "content": "classification or retrieval accuracy) on different evaluation datasets [33, 40, 79]. However, highlighting salient regions by backpropagating gradients for downstream applications does not capture how helpful end-users find these attributions [53, 73]. In this section, we take a human-centred approach to interpretability – how well our stroke attribution algorithms align with the reasoning process of humans and the trade-off, interpretability vs. accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.418 + ], + "angle": 0, + "content": "Setup: We recruited 7 participants from different geographical regions, in the age group 20 - 30 years. All participants had some background in AI research, but only 3 reported having prior experience in interpretability. Once recruited, each user is assigned a unique ID for anonymity. For SLA and P-SLA, we conduct 5 human studies, with each having 10 multiple choice questions (MCQs). Hence, each participant answers 50 MCQs for SLA and 50 for P-SLA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Evaluating Transparency: For an attribution to be useful, humans must understand a model's behaviour for correct and incorrect predictions. In this section, we evaluate if our SLA and P-SLA can make existing sketch classifiers (i.e., Sketch-A-Net [111]), pre-trained on TU-Berlin [33] dataset transparent to humans. Accordingly, we choose a random category and select 4 sketch instances - (i) 3 misclassified and 1 correctly classified, and (ii) 1 misclassified and 3 correctly classified. Next, as shown in Fig. 8, we compute stroke attribution (SLA and P-SLA) for the selected sketches and ask users: \"Only 1 of these 4 sketches are correctly (or incorrectly) recognised by our model. Please select that correct (or incorrect) sketch.\" We find users can identify correct/incorrect model predictions \\(75.9\\% /63.4\\%\\) of times for SLA and \\(76.3\\% /65.2\\%\\) for P-SLA." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.65, + 0.424, + 0.664 + ], + "angle": 0, + "content": "Q. Only one sketch is correctly categorised. Which one?" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.665, + 0.157, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.665, + 0.254, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.665, + 0.346, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.665, + 0.464, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.709, + 0.428, + 0.722 + ], + "angle": 0, + "content": "Q. Only one sketch is incorrectly categorised. Which one?" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.723, + 0.157, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.723, + 0.255, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.724, + 0.347, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.72, + 0.464, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.775, + 0.469, + 0.803 + ], + "angle": 0, + "content": "Figure 8. Evaluating transparency: Can a human understand the behaviour of an existing (pre-trained) classifier with SLA/P-SLA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Evaluating Fairness: End users are much better positioned to make a decision with help from a model if intelligible explanations are provided. In this section, we evaluate if SLA and P-SLA can help end users understand \"What went wrong?\" (i) For a pre-trained sketch classifier [111], we show users a misclassified sketch instance and ask users" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.892, + 0.263 + ], + "angle": 0, + "content": "to identify the (wrongly) predicted category. (ii) For fine-grained SBIR [13], we show a sketch (whose GT photo is not in top-10) and ask users to identify the top-1 (wrongly) retrieved photo in Fig. 9. Humans can identify the misclassified category \\(62.4\\%\\) \\((66.3\\%)\\) and the incorrectly retrieved photo \\(39.1\\%\\) \\((37.2\\%)\\) for SLA (P-SLA)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.269, + 0.678, + 0.295 + ], + "angle": 0, + "content": "Q. Prediction is incorrect, identify predicted category" + }, + { + "type": "text", + "bbox": [ + 0.687, + 0.269, + 0.881, + 0.294 + ], + "angle": 0, + "content": "Q. Retrieval is incorrect, identify the top-1 retrieved image" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.298, + 0.594, + 0.328 + ], + "angle": 0, + "content": "② \\( {Na} + {O2} < \\Delta \\) 释放" + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.299, + 0.677, + 0.328 + ], + "angle": 0, + "content": "(A) Angel (B)Butterfly" + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.33, + 0.641, + 0.359 + ], + "angle": 0, + "content": "(C) Fan (D) Dog" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.299, + 0.751, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.299, + 0.801, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.299, + 0.856, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.868, + 0.313, + 0.883, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Figure 9. Evaluating Fairness: For an incorrect model prediction, we evaluate if humans can \"identify what went wrong\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.544 + ], + "angle": 0, + "content": "**Evaluating Trustworthiness:** Determining trust in individual predictions is important when used for decision-making (e.g., medical diagnosis [57]). We train two copies of the same sketch classifier [111], a strong classifier with \\(73.8\\%\\) accuracy on TU-Berlin [33] and a weak one reaching \\(57.1\\%\\). We present the stroke attribution (SLA/P-SLA) for both models and ask users to identify the strong/weak classifier, as shown in Fig. 10. Humans identify the stronger classifier \\(71.4\\% (68.7\\%)\\) of the time for SLA (P-SLA)." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.555, + 0.864, + 0.57 + ], + "angle": 0, + "content": "Q. Prediction is correct, identify the stronger classifier" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.573, + 0.59, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.574, + 0.673, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.574, + 0.69, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.574, + 0.784, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.574, + 0.885, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Figure 10. Evaluating trustworthiness: We present the stroke attributions (SLA/P-SLA) from two sketch classifiers (strong and weak), and ask users to identify the strong/weak model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.619, + 0.708 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work emphasises the pivotal role of strokes in human-drawn sketches, offering unique insights compared to pixel-based images. Our lightweight explainability solution seamlessly integrates with pre-trained models, addressing rasterisation challenges and contributing to diverse sketch-related tasks. Through applications in Retrieval, Generation, Assisted Drawing, and Sketch Adversarial Attack, our model showcases adaptability and significance. The proposed stroke-level attribution provides nuanced insights into model behaviour, underscoring the importance of explainability in bridging human expression with model predictions in the evolving field of sketch interpretation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11004" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.159, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. arXiv preprint arXiv:2011.05429, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.202, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[3] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. arXiv preprint arXiv:2108.00401, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.259, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[4] Shm Garanganao Almeda, J.D. Zamfirescu-Pereira, Kyu Won Kim, Pradeep Mani Rathnam, and Bjoern Hartmann. Prompting for discovery: Flexible sense-making for ai art-making with dreamsheets. arXiv preprint arXiv:2310.09985, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.33, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[5] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.388, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[6] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. A unified view of gradient-based attribution methods for deep neural networks. In NIPSW, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.431, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[7] Anurag Arnab, Ondrej Miksik, and Philip H.S. Torr. On the robustness of semantic segmentation models to adversarial attacks. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.474, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[8] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[9] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[10] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.471, + 0.756 + ], + "angle": 0, + "content": "[11] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.759, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[12] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[13] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang," + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and Yi-Zhe Song. Doodle it yourself: Class incremental learning by drawing a few sketches. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[15] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.194, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[16] Moritz Bohle, Mario Fritz, and Bernt Schiele. B-cos: Networks: Alignment is all we need for interpretability. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.238, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[17] Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well onImagenet. arXiv preprint arXiv:1904.00760, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.295, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[18] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.366, + 0.892, + 0.421 + ], + "angle": 0, + "content": "[19] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.424, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[20] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.468, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[21] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In ICCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[22] Xuesong Chen, Xiyu Yan, Feng Zheng, Yong Jiang, Shu-Tao Xia, Yong Zhao, and Rongrong Ji. One-shot adversarial attacks on visual tracking with dual attention. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.569, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[23] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.627, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[24] Pinaki Nath Chowdhury, Aneeshan Sain, Yulia Gryaditskaya, Ayan Kumar Bhunia, Tao Xiang, and Yi-Zhe Song. Fs-coco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.684, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[25] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation: Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.742, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[26] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.786, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[27] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[28] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11005" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[30] Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron? arXiv preprint arXiv:1805.12233, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[31] Gabriele Dominici, Pietro Barbiero, Lucie Charlotte Magister, Lio Pietro, and Nikola Simidjevski. Sharcs: Shared concept space for explainable multimodal learning. arXiv preprint arXiv:2307.00316, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.191, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[32] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[33] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[34] Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. Technical Report, Université de Montréal, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[35] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. arXiv preprint arXiv:2112.04417, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[36] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[37] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[38] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.471, + 0.531 + ], + "angle": 0, + "content": "[39] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[40] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[42] Yang He, Shadi Rahimian, Bernt Schiele, and Mario Fritz. Segmentations-leak: Membership inference attacks and defenses in semantic image segmentation. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.471, + 0.688 + ], + "angle": 0, + "content": "[43] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[44] Josh Holinaty, Alec Jacobson, and Fanny Chevalier. Supporting referenceimageryfordigitaldrawing. In ICCVW, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.73, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[45] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.788, + 0.471, + 0.842 + ], + "angle": 0, + "content": "[46] Quan Huynh-Thu, Marie-Neige Garcia, Filippo Speranza, Philip Corriveau, and Alexander Raake. Study of ratling scales for subjective quality assessment of high definition video. IEEE TBC, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.844, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[47] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[48] Yunhan Jia, Yantao Lu, Junjie Shen, Qi Alfred Chen, Hao Chen, Zhenyu Zhong, and Tao Wei. Fooling detection alone" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "is not enough: Adversarial attack against multiple object tracking. In ICLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[49] Zeyang Jia, Eli Ben-Michael, and Kosuke Imai. Bayesian safe policy learning with chance constrained optimization: Application to military security assessment during the vietnam war. arXiv preprint arXiv:2307.08840, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[50] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[51] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In ICCV, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[52] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.305, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[53] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In ECCV, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.362, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[54] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[55] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.49, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[56] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.533, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[57] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[58] Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Sam Gershman, and Finale Doshi-Velez. An evaluation of the human-interpretability of explanation. arXiv preprint arXiv:1902.00006, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.632, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[59] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In WACV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[60] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.718, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[61] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, Yi-Zhe Song, and Yonggang Qi. Zero-shot everything sketch-based image retrieval, and in explainable style. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[62] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Defloconet: Deep image editing via flexible low-level controls. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[63] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11006" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[64] Meike Nauta, Ron Van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[65] Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. ICMLW, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.192, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[66] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.221, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[67] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[68] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[69] Buu Phan, Fahim Mannan, and Felix Heide. Adversarial imaging pipelines. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[70] Zhiyu Qu, Yulia Gryaditskaya1, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.391, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[71] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[72] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards better understanding attribution methods. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[73] Marco T. Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In KDD, 2016. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Omer. High-resolution image synthesis with latent diffusion. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[75] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[76] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.675, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[77] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.732, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[78] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE TNNLS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.788, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[79] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[80] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Pronimal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[81] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "tra. Grad-cam: Visual explanations from deep networks via gradient-based localisation. In ICCV, 2017. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[82] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.18, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[83] Harshay Shah, Prateek Jain, and Praneeth Netrapalli. Do input gradients highlight discriminative features? In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.223, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[84] Hamid Rahim Sheikh and Alan C. Bovik. Image information and visual quality. IEEE TIP, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.254, + 0.892, + 0.281 + ], + "angle": 0, + "content": "[85] Yuming Shen, Li Liu, Fumin Shen, and Ling Shao. Zeroshot sketch-image hashing. In ICCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.283, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[86] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In ICML, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.327, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[87] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In ICLRw, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.384, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[88] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiβ, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. arXiv preprint arXiv:2204.11642, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[89] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.499, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[90] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.556, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[91] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In ICLRw, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.6, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[92] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.643, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[93] Christian Szegedy, Wojciech Zaremba, Ilyua Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.701, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[94] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.758, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[95] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white- and black-box attacks. In NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[96] Aditay Tripathi, Rajath R Dani, Anand Mishra, and Anirban Chakraborty. Sketch-guided object localization in natural images. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[97] James Tu, Mengye Ren, Siva Manivasagam, Ming Liang, Bin Yang, Richard Du, Frank Cheng, and Raquel Urtasun. Physically realizable adversarial examples for lidar object detection. In CVPR, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11007" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[98] Hugues Turbé, Mina Bjelogrlic, Christian Lovis, and Gianmarco Mengaldo. Evaluation of post-hoc interpretability methods in time-series classification. Nature Machine Intelligence, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[99] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[100] Donglai Wei, Joseph Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[101] Dennis Wei, Rahul Nair, Amit Dhurandhar, Kush R. Varshney, Elizabeth M. Daly, and Moninder Singh. On the safety of interpretable machine learning: A maximum deviation approach. In NeurlPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[102] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: deep sketch-based hair image synthesis. ACM TOG, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[103] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[104] Rui Xu, Zongyan Han, Le Hui, Jianjun Qian, and Jin Xie. Domain disentangled generative adversarial network for zero-shot sketch-based 3d shape retrieval. AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[105] Guowei Yan, Zhili Chen, Jimei Yang, and Huamin Wang. Interactive liquid splash modeling by user sketches. ACM TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[106] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[107] Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Suggala, David I Inouye, and Pradeep K Ravikumar. On the (in) fidelity and sensitivity of explanations. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[108] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A zero-shot framework for sketch based image retrieval. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[109] Ran Yi, Zipeng Ye, Ruoyu Fan, Yezhi Shu, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Animating portrait line drawings from a single face photo and a speech signal. In ACM SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[110] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[111] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[112] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[113] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[114] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[115] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[116] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In CVPR, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[117] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. IJCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[118] Lvmin Zhang, Anyi Rao, and Maneesh Agarwala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[119] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[120] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[121] Hang Zhou, Dongdong Chen, Jing Liao, Weiming Zhang, Kejiang Chen, Xiaoyi Dong, Kunlin Liu, Gang Hua, and Nenghai Yu. Lg-gan: Label guided adversarial network for flexible targeted attack of point cloud-based deep networks. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[122] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchscene: Richly-annotated scene sketches. In ECCV, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11008" + } + ] +] \ No newline at end of file diff --git a/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_origin.pdf b/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9f1a50793f6e0a3d5eeb5f843b241f550d209554 --- /dev/null +++ b/2024/What Sketch Explainability Really Means for Downstream Tasks_/22ec61cf-a0b2-4071-b86f-9e1362f3069e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f366a6c3d93504706fd0f28b4bbdc391c087b1a9504ac147bf735a1f4c613f40 +size 4383025 diff --git a/2024/What Sketch Explainability Really Means for Downstream Tasks_/full.md b/2024/What Sketch Explainability Really Means for Downstream Tasks_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7b1e169ed4e3f28e849015bbf22370f27722d5a6 --- /dev/null +++ b/2024/What Sketch Explainability Really Means for Downstream Tasks_/full.md @@ -0,0 +1,478 @@ +# What Sketch Explainability Really Means for Downstream Tasks? + +Hmrishav Bandyopadhyay1 Pinaki Nath Chowdhury1 Ayan Kumar Bhunia1 +Aneeshan Sain1 Tao Xiang1,2 Yi-Zhe Song1,2 +1SketchX, CVSSP, University of Surrey, United Kingdom. +2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. + +{h.bandyopadhyay, p.chowdhury, a.bhunia, a.sain, t.xiang, y.song}@surrey.ac.uk + +# Abstract + +In this paper, we explore the unique modality of sketch for explainability, emphasising the profound impact of human strokes compared to conventional pixel-oriented studies. Beyond explanations of network behavior, we discern the genuine implications of explainability across diverse downstream sketch-related tasks. We propose a lightweight and portable explainability solution - a seamless plugin that integrates effortlessly with any pre-trained model, eliminating the need for re-training. Demonstrating its adaptability, we present four applications: highly studied retrieval and generation, and completely novel assisted drawing and sketch adversarial attacks. The centrepiece to our solution is a stroke-level attribution map that takes different forms when linked with downstream tasks. By addressing the inherent non-differentiability of rasterisation, we enable explanations at both coarse stroke level (SLA) and partial stroke level (P-SLA), each with its advantages for specific downstream tasks. + +# 1. Introduction + +Sketches, rooted in human expression [43], offer a distinctive modality for exploring explainability [61, 70]. In contrast to photos, where each pixel is independent and lacks inherent meaning, sketches are organised into strokes, with each stroke carrying subjective meaning assigned by the sketcher [44]. This paper explores sketch explainability, but with a unique perspective – aiming to provide explanations and unravel the true implications of explainability on various downstream sketch-related tasks. + +With this perspective in mind, our approach champions an explainability solution that is (i) lightweight and portable - a plugin seamlessly integrating with multiple pre-trained models without necessitating re-training [98], and (ii) easily adaptable to a diverse array of downstream sketch-specific tasks, benefiting the broader community. + +Our solution is exclusively centred on human strokes, aiming to attribute explanation on different stroke granularity: individual strokes (coarse) and their parts (fine). The + +![](images/482f4d2870546db7edb97a1621470c70cadaa0d41a4106b3b4b329f57e112ef4.jpg) +Figure 1. We attribute explanations for individual strokes (stroke-level attribution) and their vector coordinate points (point-level attribution). Stroke-level attribution rasterises individual strokes (non-differentiably) to produce $n$ -stroke images. Next, we sum the stroke images to get the complete sketch image used for downstream tasks. Point-level Attribution computes distance transform from stroke coordinates and thresholds to get the sketch image. Our explainability solution works without re-training for existing tasks like SBIR and sketch-to-photo generation and novel tasks like filtering noisy strokes for assisted drawing and adversarial attack by removing a small stroke. + +output of our model markedly differs from typical saliency maps [1] found in photo-based explainability models, where the emphasis is mostly on visualisation [90]. Ours is a task-driven attribution map that assigns stroke-level attributes capturing how altering stroke characteristics can impact model prediction. Depending on the downstream tasks, attributions can be grounded to, for example (i) importance of entire strokes, which is more suitable to filter noisy strokes [15] in assisted drawing, and remove small strokes for adversarial attacks on existing sketch encoders, and (ii) stroke shape and length, where a partial-stroke level attribution is beneficial for tasks like sketch-based image retrieval [77] and sketch-to-photo generation [55]. + +To showcase the adaptability of our model, we carefully devise four applications: two well-studied tasks from existing literature (retrieval [12, 26] and generation [55, 102]), and two entirely novel tasks (assisted drawing and sketch + +adversarial attack). In retrieval, we evaluate reliability of model predictions by comparing predicted stroke order with the order in which a human draws them. For generation, we pinpoint strokes with the least influence, offering explicit feedback to end-users regarding which strokes the model prioritised and which it overlooked. In assisted drawing [4], we assist novice artists in faithfully sketching a particular photo by identifying strokes that do not match the target photo. Lastly, in adversarial attacks, we unveil the vulnerability of state-of-the-art sketch encoders by removing a small imperceptible stroke in any sketch, resulting in significant changes to the model's prediction. + +The focal point connecting all downstream tasks is our proposed stroke-level attribution. The key question, therefore, is how to backpropagate information to strokes while addressing the inherent non-differentiability of rasterisation - strokes are most often represented as discrete coordinates and rasterised before feeding into downstream applications. We provide two solutions for non-differentiability: (i) coarse stroke level: we first rasterise individual strokes to produce raster stroke images. Then, we combine these stroke images to get the complete sketch image (see Fig. 1) - since this addition of stroke images is a differentiable operation, we can backpropagate information from the complete sketch to individual raster stroke images. (ii) fine partial-stroke level: we create a distance transform image from stroke coordinates ("red dots" in Fig. 1 vector sketch) by calculating the minimum distance of each pixel in the image from the coordinates. Then, we threshold the distance value to get the sketch image (white pixels for a high distance and black pixels for a low distance). Since the distance function and our threshold step are both differentiable, we can backpropagate information from the sketch images to stroke coordinates. + +Our contributions can be summarised as follows: (i) We explore sketch explainability, emphasising the importance of strokes in human-drawn sketches. (ii) We highlight the profound impact of explainability on various sketch-related domains, presenting applications in retrieval, generation, assisted Drawing, and adversarial attacks. (iii) We solve for the non-differentiability problem of rasterisation, and provide both stroke-level and partial-stroke level attribution. + +# 2. Related Works + +Sketch for Visual Understanding: Having a high visual proximity [43] to real images and carrying human subjectivity [9, 76], amateur sketches or abstract line drawings [59] has been a popular modality for customised expression, thus driving extensive applications as a query for retrieval [26, 32, 76, 85] of object [29] or scene [24] images, 3D shapes [104], and even concepts like in 'dictionary-like' games [11]. As a canvas for creativity, sketch helps image-editing [62, 110, 114], or generation of objects [20, 21, 38, 99], + +scenes [109], and 3D shapes [10, 39, 105, 119]. Being easily editable, sketch enables interactive access to AI systems like image-segmentation [45, 122], object localisation [96], image-inpainting [110, 114], and incremental learning [14]. Being application-specific however, such works largely ignored explaining the 'how' of sketch-correspondence. The few who did, customised training pipelines [61, 70] for niche tasks. In this work, we thus make the first attempt at visualising salient sketch-regions (strokes), as an explainability-tool (like GradCAM [81] in photos) for existing pre-trained sketch-based downstream networks. + +Explaining CNN Predictions: CNN explanations visually highlight regions of input having the maximum influence on a model's predictions [87]. This visualisation of 'salient' regions is either through an analysis [37] of a regular pre-trained network after it has completed training (post-hoc), or by designing and training explicitly interpretable (i.e. explain-and-predict) models [17, 19, 64]. Given a pre-trained CNN, a post-hoc algorithm either visualises (i) model attributes like feature and activation-maps [34, 65, 66, 82, 113, 120] that imply saliency of specific input (pixel) regions, or (ii) input attributes directly as pixels [91] or pixel-regions [51] coloured according to their relative importance. Visualisation of input attributes is facilitated through perturbation based algorithms [18, 36, 37] and gradient-based analysis [87, 89, 91]. Perturbation-based algorithms [27, 68, 113] detect saliency of pixel-regions by measuring impact of their absence on the prediction score. Whereas, gradient-based algorithms [8, 30, 86, 92, 117] measure the gradient of the prediction with respect to individual input-pixels (input-gradients [83]), attributing them based on this value. Unlike images however, sketch is a sparse-information modality [23] for pixels. As such, we explore explainability in sketches by computing stroke and point attribution for fine-grained sketch explanations. + +Evaluation of CNN Explanations: The evaluation of CNN explanations has evolved over time - from naive qualitative analysis of visualisations [82, 87, 91, 113] to standardised theoretical [5, 92] and empirical [72, 107, 113] baselines. Analysing explanations theoretically helps evaluate them in a model-agnostic environment, where their mathematical form is checked against pre-defined properties (axioms)[5, 54, 92]. Empirical evaluations, on the other hand, involve experiments measuring $(i)$ variance in explanations upon perturbations of inputs [107] and model weights [1, 2] (sanity checks), and $(ii)$ accuracy of explanations in locating important features [37, 72, 82, 117]. These features, when perturbed, influence the CNN maximally, as measured by perturbation-based metrics [6, 51, 68, 78, 82, 83, 113]. Recently, however, these evaluation protocols have met criticism [35] due to their detachment from humans. Instead, human studies [53, 58, 82, 88], and human-centered metrics [35] have been proposed for evaluating explanation in + +terpretability. In this work, we attribute strokes by backpropagation [86, 87, 91, 92], evaluating attributions through sanity checks [1], empirical metrics [51] and human studies [53] on downstream sketch-based applications [112]. + +# 3. Background + +Here, we provide a brief overview of some standard concepts, ubiquitous in explainability literature [1] to help formalise the question: "what entails a good explanation?" + +Attribution Algorithms: It highlights relevant regions (e.g., pixels in an image, see Fig. 1) that are responsible for the model's prediction. Despite its importance for safety-critical applications [49, 75, 101], making an attribution algorithm interpretable to humans remains an open problem. Surprisingly, a more faithful attribution is usually less interpretable and vice-versa [91, 113]. Prior works [16, 81] study this trade-off between faithfulness vs. interpretability as an answer to: "What makes a good visual explanation?" + +The attribution map $\mathbb{A} \in \mathbb{R}^{H \times W}$ is typically calculated using gradients for an input $\mathrm{X} \in \mathbb{R}^{H \times W \times 3}$ for a classification model $\hat{y} = F_{\theta}(\mathrm{X}) \in \mathbb{R}^{C}$ , pre-trained on $C$ categories. The gradients are a simple and good indicator of how much the model prediction changes for input $\mathrm{X}$ as, + +$$ +\mathbb {A} = \partial F _ {\theta} (\mathrm {X}) / \partial \mathrm {X} \tag {1} +$$ + +Interpretability: It is the ability of an attribution algorithm to provide a qualitative "understanding" for a model [73]. This "understanding" depends on the target audience, e.g., a human expert may interpret a small Bayesian network [56], but a layman is more comfortable with a weighted attention (or feature) map that highlights salient regions [1]. To evaluate the interpretability of attribution maps, prior works either (i) perform downstream tasks [31, 81] that depend on the interpretation (e.g., object localisation – predict the bounding box and semantic segmentation for image region with the highest attribution) or (ii) human studies [53], typically conducted in two setups – class discriminative (given an attribution map, ask users to identify the category predicted by a model), and trustworthiness (compare attribution maps from a strong and a weak model, and ask users to identify the stronger model). + +Faithfulness: It is the ability of attribution algorithms to accurately "explain" the computation learned by a model. For example, in theory, a fully faithful attribution is the entire model (e.g., ResNet-18 [41]) but is not interpretable by a human. In practice, for an attribution to be meaningful, it is often impossible to be completely faithful. To balance this trade-off, prior works explore human-interpretable attributions that are locally faithful for a given model prediction. One approach is image occlusion [73], where the difference in the model scores is measured when masking different patches in an input image. Image patches that significantly change the model score are deemed important by the attribution algorithm. + +# 4. Proposed Method + +The attribution map $\mathbb{A}$ in Eq. (1) gives a faithful fine-grained explanation for each pixel in X. However, such pixel attribution does not provide meaningful explanations when interpreting sparse human-drawn sketches (many empty white pixels). Additionally, pixel attribution does not consider the sketch construction process - humans sketch a sequence of strokes, not pixels. Hence, it is most appropriate that sketch predictions should be attributed to strokes and not pixels. However, the key challenge to stroke attributions is that most sketch applications [77, 118] use raster sketches - a non-differentiable process to convert a sequence of strokes into pixels. In the following sections, we propose two stroke attribution algorithms that consider the sketch construction process by designing a differentiable rasterisation pipeline for a vector sequence of strokes. + +Algorithm 1: Non-differentiable Rasterisation +Data: V $\leftarrow$ Vector Sketch of size $\mathbb{R}^{T\times 5}$ +Result: X $\leftarrow$ Blank (Zero) Canvas of size $\mathbb{R}^{h\times w\times 3}$ . + $\mathbb{B}(\cdot ,\cdot)\gets$ Bresenham Function ; + $v_{0} = (x_{0},y_{0},q_{0}^{1},q_{0}^{2},q_{0}^{3})\gets \mathrm{V}[0]$ . + $v_{\mathrm{prev}}\gets v_0$ . + $q_{\mathrm{prev}}\gets q_0^1$ . +for $v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})\gets \mathrm{V}[1\dots T]$ do if $q_{\mathrm{prev}} = 1$ and $q_t^1 = 1$ then for $(p_x^i,p_y^i)\gets \mathbb{B}(v_{\mathrm{prev}},v_t)$ do |X $(p_x^i,p_y^i)\gets 255;$ end end $v_{\mathrm{prev}}\gets v_t$ . $q_{\mathrm{prev}}\gets q_t^1$ . if $q_{t}^{3} = 1$ then exit(); /\* End of Drawing \*/ end + +# 4.1. Sketch Representations + +While sketches are used in several formats (or representations) like Raster [111], Vector [40], or Bézier [28], digital sketches are primarily captured in vector form, as a list of points traced on a drawing pad. These points are usually a five-element vector $v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})$ where, $(x_{t},y_{t})$ are absolute coordinates in a $(H\times W)$ drawing canvas and the last three elements are one-hot encoding of pen-states: pen touching the paper $(1,0,0)$ , pen is lifted $(0,1,0)$ , and end of drawing $(0,0,1)$ . Hence, a vector sketch with $T$ points is represented as $\mathbf{V}\in \mathbb{R}^{T\times 5}$ . + +As most downstream sketch applications work on rasterised sketches $\mathrm{X} \in \mathbb{R}^{H \times W \times 3}$ , prior works [111] translate (non-differently) a vector sketch $\mathrm{V}$ to its equivalent raster sketch $\mathrm{X}$ using Algorithm 1. For this, the Bre + +![](images/97dc47e35b6d7b1ce1fc4bc686694c42bb19e462417b7bebceca4c9ff4c672ab.jpg) +Figure 2. Coarse Stroke-level Attribution. Backpropagate gradients from raster sketch X to raster strokes $S_{i}$ , with weight $\omega_{i}$ . + +senham function $\mathbb{B}(\cdot)$ is used to connect two vector points $\{v_{t - 1}, v_t\} \in \mathrm{V}$ in the pixel space $\{(p_x^1, p_y^1), \ldots, (p_x^n, p_y^n)\} \in \mathbb{B}(v_{t - 1}, v_t)$ via a continuous line. Next, we show how our sketch attributions overcome this non-differentiable rasterisation and backpropagate gradients (Eq. (1)) from pixels $\mathrm{X} \in \mathbb{R}^{H \times W \times 3}$ to strokes and points in $\mathrm{V} \in \mathbb{R}^{T \times 5}$ . + +# 4.2. Coarse Stroke-level Attribution (SLA) + +In this section, we backpropagate gradients from pixel space in $\mathrm{X} \in \mathbb{R}^{H \times W \times 3}$ to strokes, defined as a continuous set of points $\{v_t, v_{t+1}, \ldots, v_{t+n}\}$ from the first pen-down $(1,0,0)$ till the pen-up $(0,1,0)$ state. Algorithm 1 converts the vector stroke points into a raster stroke $S_i \in \mathbb{R}^{H \times W \times 3}$ . The final raster sketch is then a differentiable composition of $m$ raster strokes $\mathrm{X} = \sum_{k=1}^{m} S_k$ . We compute SLA as, + +$$ +\mathbb {A} _ {i} ^ {R} = \frac {\partial F _ {\theta} (\mathrm {X})}{\partial \mathcal {S} _ {i}} = \frac {\partial F _ {\theta} (\mathrm {X})}{\partial \mathrm {X}} \cdot \frac {\partial \sum_ {k = 1} ^ {m} \mathcal {S} _ {k}}{\partial \mathcal {S} _ {i}} \tag {2} +$$ + +This, however, gives a degenerate solution where all strokes will have the same attribution $\partial \sum_{k=1}^{m} S_k / \partial(S_i) = 1$ . To avoid this, we compute a weight factor $\omega_i \in \mathbb{R}^{H \times W \times 3}$ for each stroke $S_i$ such that, + +$$ +\omega_ {i} \left(p _ {x}, p _ {y}\right) = \left\{ \begin{array}{l l} 1 & \text {i f} \left(p _ {x}, p _ {y}\right) \in \mathbb {B} \left(v _ {t - 1}, v _ {t}\right) \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +In other words, given consecutive vector sequence of points $(v_{t - 1}, v_t)$ in stroke $S_i$ , we find all points $(p_x, p_y)$ using Bresenham function $\mathbb{B}(\cdot)$ that lie "on the stroke" and assign $\omega_i(p_x, p_y) = 1$ . Hence, longer strokes will have more 1's compared to shorter strokes. Finally, Eq. (2) is adapted as + +$$ +\mathbb {A} _ {i} ^ {R} = \frac {\partial F _ {\theta} (\mathrm {X})}{\partial \mathrm {X}} \cdot \frac {\partial \sum_ {k = 1} ^ {m} \omega_ {k} \mathcal {S} _ {k}}{\partial \mathcal {S} _ {i}} \tag {4} +$$ + +SLA makes the non-differentiable rasterisation $(\mathrm{V} \to \mathrm{X})$ partially differentiable (stokes $S$ to sketch X). In other words, SLA answers "Which strokes in a sketch are important". Next, we make the rasterisation fully differentiable and backpropagate gradients to vector points V to answer "Which point in a sketch is important". + +![](images/f6d2bf8074d1ead883ebc215632550611d23976834afd81b2473cbf07644f9a2.jpg) +Figure 3. Partial Stroke-level Attribution. Backpropagate gradients from raster sketch X to vector sequence of coordinates V. + +# 4.3. Partial Stroke-level Attribution (P-SLA) + +Unlike SLA, which partially captures the sketch construction process (stroke-level), partial stroke-level attribution (P-SLA) can fully backpropagate gradients to the vector list of coordinates $\mathrm{V} \in \mathbb{R}^{T \times 5}$ traced on a drawing pad. Given a blank canvas $\mathrm{X} \in \mathbb{R}^{H \times W \times 3}$ , we (i) calculate the minimum distance of every pixel $(p_x, p_y)$ in $\mathrm{X}$ from a line segment $(v_{t-1}, v_t)$ in $\mathrm{V}$ , and (ii) compute the pixel intensity of $\mathrm{X}(p_x, p_y)$ as function of the minimum distance as + +$$ +\begin{array}{l} \mathrm {X} \left(p _ {x}, p _ {y}\right) = \sigma [ 2 - 5 \min _ {t = 2, \dots , T} (5) \\ \left. \mathrm {d i s t} ((p _ {x}, p _ {y}), v _ {t - 1}, v _ {t}) + (1 - q _ {t - 1} ^ {1}) 1 0 ^ {6}) \right] \\ \end{array} +$$ + +where $\sigma (\cdot)$ is the sigmoid function, and $\mathrm{dist}(\cdot)$ is a distance function (see Supp.) from a point $(p_x,p_y)$ to a line segment $(v_{t - 1},v_t)$ . For pen-up states $(q_{t - 1}^{1} = 0)$ , we blow up $(\times 10^{6})$ the distance values that make the pixel intensities $\mathrm{X}(p_x,p_y)\to 0$ , i.e., not render strokes for $(v_{t - 1},v_t)$ . Finally, we compute P-SLA + +$$ +\mathbb {A} _ {t} ^ {V} = \frac {\partial F _ {\theta} (\mathrm {X})}{\partial v _ {t}} = \frac {\partial F _ {\theta} (\mathrm {X})}{\partial \mathrm {X}} \cdot \sum_ {\forall p _ {x}, p _ {y}} \left\{\frac {\partial \mathrm {X} \left(p _ {x} , p _ {y}\right)}{\partial v _ {t}} \right\} \tag {6} +$$ + +# 5. Applications of Stroke Attributions + +Despite its simplicity, designing stroke attribution algorithms that capture the sketch construction process unlocks insights into numerous existing downstream tasks like classification [40], robust sketch-based image retrieval (category-level [33] and fine-grained [79]) and enables some novel sketch applications like assisted drawing [15], interactive sketch to photo generation [55], adversarial attacks on human-drawn sketches, and discovering the "arrow of time" [100] in raster sketches. + +# 5.1. Robust Sketch Based Image Retrieval + +Given a query sketch $\mathrm{X} \in \mathbb{R}^{H \times W \times 3}$ , category-level sketch-based image retrieval (SBIR) aims to fetch category-specific photos from a gallery of multi-category photos (e.g., given sketch of a 'shoe' retrieve any photo 'shoe' from a gallery of 'shoes+hats+cows'). Conversely, fine-grained SBIR aims to retrieve one instance from a gallery of the same category photos (e.g., given the sketch of a 'shoe' retrieve one photo shoe from a gallery of all shoes). Deep learning frameworks + +Table 1. Stroke attribution (SLA, P-SLA) make SBIR systems reliable. Sketches with a high correlation $(Corr)$ of stroke saliency (predicted by SLA or P-SLA) with human-drawn temporal stroke order tend to have higher retrieval accuracy. + +
MetricsFull DatasetCorr ≥ 0.5Corr ≤ 0.1
SLAP-SLASLAP-SLA
CategorymAP53.155.357.651.750.1
LevelP@20065.966.768.564.661.5
FineAcc.@115.316.417.613.812.7
GrainedAcc.@534.236.939.431.128.3
+ +learn a joint sketch-photo manifold (for category and fine-grained) via a feature extractor [26, 29, 103] trained using triplet loss [112]. Recent adoption of foundation models for SBIR [77] shifts focus to robust deployment using the open-set generalisation of CLIP [71]. + +Towards this goal of robust deployment, our sketch attribution algorithms $(\mathbb{A}_i^R$ and $\mathbb{A}_t^V)$ can predict which strokes the network focuses on when retrieving a photo (Fig. 4). Apart from interpreting SBIR models, sketch attribution can also help detect potential failures at runtime (inference). First, we use the attribution scores $\mathbb{A}_i^R$ or $\mathbb{A}_t^V$ to rearrange the strokes from highest to lowest. The attribution scores indicate the most salient to the least salient strokes that affect model prediction. Second, we calculate a correlation $(Corr)$ of our predictor stroke order with the ground-truth temporal stroke order drawn by a user (humans draw the most salient regions first and least salient areas last [33, 79]). A high correlation indicates that humans and our model prioritise strokes similarly, whereas a low $Corr$ denotes that the model and the user prioritise different strokes. We evaluate SBIR on a pre-trained SOTA model [77] using CLIP with prompt learning as a sketch and photo encoder. + +Datasets: We use TU-Berlin [33] (for category-level SBIR) and Sketchy [79] (for fine-grained SBIR). TU-Berlin contains 250 categories, with 80 free-hand sketches in each, and 204, 489 images [116]. Sketchy [79] has 75, 471 sketches over 125 categories having 100 images in each [108]. + +Evaluation Metrics: Following [77], we use mean average precision (mAP) and precision for top 200 retrieved samples $(\mathbb{P}@\mathbb{200})$ for category-level SBIR. For fine-grained SBIR [23], we measure Acc.@q, i.e., the percentage of sketches whose true matched photo is in the top-q list. + +Results: We divide the evaluation set into two sets: (i) those that have a high correlation $Corr \geq 0.5$ , and (ii) those with a low correlation $Corr \leq 0.1$ of ground-truth and predicted stroke order. Tab. 1 shows sketches with a high $Corr \geq 0.5$ are 1.7/3.9 more accurate in Acc.@1/Acc.@5 than those with $Corr \leq 0.1$ for fine-grained SBIR, and 3.3/1.7 better in mAP/P@200 for category-level SBIR. Full Dataset indicates the performance of the pre-trained model on the entire evaluation set. + +![](images/81299084bc63347c20a46cf747ab3e28cf0e2ecb551e4b192eab4f18863aff04.jpg) +Figure 4. Sketch attributions from stroke-level and point-level for image retrieval. High correlation of human-drawn stroke order with that from sketch-attributions (high $\rightarrow$ low) indicate our sketch encoder gives more importance to salient strokes drawn early on. + +# 5.2. Assisted Drawing via Noisy Stroke Removal + +Although sketching has enabled many exciting applications [25, 55, 63, 118], the fear to sketch (i.e., "I can't sketch") has proven fatal for its widespread adoption. To solve this, prior works [15] used complex (and hard to train [12]) reinforcement learning [80] to predict the importance of each stroke in a sketch. Next, a stroke subset selector removes noisy (less important) strokes, leaving only those positively contributing to the downstream tasks. + +In this section, we focus on assisted drawing - given a photo, we use attribution scores $(\mathbb{A}_i^R$ or $\mathbb{A}_t^V)$ to help humans draw a faithful and clean sketch. Our method is significantly simpler than reinforcement learning alternatives [15]. + +For SLA, an input sketch $\mathrm{X} = \sum_{k=1}^{m} \omega_k S_k$ is composed of $m$ strokes. We calculate the cosine similarity (sim) of input sketch $\mathrm{X}$ with its target photo $\mathrm{P} \in \mathbb{R}^{H \times W \times 3}$ to measure "how faithfully $\mathrm{X}$ describes $\mathrm{P}$ ". Next, we backpropagate gradients from the cosine similarity to strokes $S_i$ and calculate stroke-level attribution score $\mathbb{A}_i^R$ as + +$$ +\mathbb {A} _ {i} ^ {R} = \frac {\partial \operatorname {s i m} (F _ {\theta} (\mathrm {X}) , F _ {\theta} (\mathrm {P}))}{\partial \mathrm {X}} \cdot \frac {\partial \sum_ {k = 1} ^ {m} \omega_ {k} \mathcal {S} _ {k}}{\partial \mathcal {S} _ {i}} \tag {7} +$$ + +The pre-trained sketch and photo encoder $F_{\theta}(\cdot)$ must be highly accurate to judge sketch-photo correspondence. Hence, we use pre-trained CLIP+prompts encoder from [77]. To remove noisy strokes, we only update the weights $\omega_{i}\in \mathbb{R}^{H\times W\times 3}$ using a normalised attribution score $\mathbb{A}_i^R$ as + +$$ +\omega_ {i} ^ {*} = \omega_ {i} \cdot \text {G u m b e l . S o f t m a x} \left(\frac {\mathbb {A} _ {i} ^ {R}}{\sum_ {\forall i} \mathbb {A} _ {i} ^ {R}} + \Delta\right) \tag {8} +$$ + +Gumbel_Softmax(·) makes the output one-hot (discrete value) in the forward pass but differentiable with a probability distribution that sum to 1 in the backward pass [47]. The + +![](images/02207721c41a183f11951d641c93908624ec77b6048c9f07c99dc64e6296e5a8.jpg) +Figure 5. Assisted drawing via sketch healing (or filtering noisy strokes) using stroke attributions from SLA and P-SLA. This helps users having fear-to-sketch ("I can't sketch"). + +modified sketch is constructed as $\mathrm{X} = \omega_1^* S_1 + \dots + \omega_m^* S_m$ . Intuitively, we keep strokes that contribute $(\mathbb{A}_i^R)$ to a high cosine similarity matching human sketch $\mathrm{X}$ and target photo $\mathrm{P}$ and remove $S_i$ with normalised $\mathbb{A}_i^R$ lower than $(0.5 - \Delta)$ . + +Similar to SLA, we can also use P-SLA to compute the attribution $\mathbb{A}_t^V$ for each point $v_{t}$ in the vector sketch $\mathrm{V}\in \mathbb{R}^{H\times W\times 3}$ , by measuring the cosine similarity between input X and target photo P as + +$$ +\mathbb {A} _ {t} ^ {V} = \frac {\partial \operatorname {s i m} \left(F _ {\theta} (\mathrm {X}) , F _ {\theta} (\mathrm {I})\right)}{\partial \mathrm {X}} \cdot \sum_ {\forall p _ {x}, p _ {y}} \left\{\frac {\partial \mathrm {X} \left(p _ {x} , p _ {y}\right)}{\partial v _ {t}} \right\} \tag {9} +$$ + +We remove noisy points $v_{t}$ by updating the pen-states in Eq. (5) from pen-down $(1,0,0)$ to pen-up $(0,1,0)$ depending on its attribution $\mathbb{A}_t^V$ for point $v_{t}\in \mathrm{V}$ . + +$$ +q _ {t - 1} ^ {1 *} = q _ {t - 1} ^ {1} \cdot \text {G u m b e l . S o f t m a x} \left(\frac {\mathbb {A} _ {t - 1} ^ {V}}{\sum_ {\forall t} \mathbb {A} _ {t} ^ {V}}\right) \tag {10} +$$ + +Using updated values for $q_{t-1}^{1*}$ , we update $q_{t-1}^{2*} = 1 - q_{t-1}^{1*}$ and recalculate pixel intensities for raster sketch $\mathrm{X}(p_x, p_y)$ . The value of hyper-parameter $\Delta$ significantly affects the stroke removal process. We found the optimal $\Delta$ for SLA and P-SLA is 0.3 and 0.1, respectively. A higher $\Delta$ for P-SLA gives broken lines with a drop in the visual quality of an input sketch. Next, we evaluate stroke filtering using SLA and P-SLA on popular human-drawn sketch datasets. + +Dataset: For a fair comparison with prior works [15], we evaluate on fine-grained SBIR datasets QMUL-ShoeV2 and QMUL-Chair-V2 [12, 67, 112]. It consists of 6,730/1,800 sketches and 2,000/400 photos from ShoeV2/Chair-V2. We evaluate on the standard test-split of 679/525 sketches and 200/100 photos. + +Evaluation Metric: We measure the retrieval accuracy of the clean sketch with the target photo by computing Acc.@1 and Acc.@5. A high accuracy need not correspond to high visual quality to the human eye [84]. Hence, we conducted a small human study with 5 participants and reported the mean opinion score (MOS) [46]; each was asked to compare two sets of 50 sketch pairs (GT sketch vs SLA filtered) and (GT sketch vs P-SLA filtered). + +Results: Removing noisy strokes from human-drawn sketches is still a new topic; hence, to the best of our knowledge, there is only one work by Bhunia et al. [15]. From + +Table 2. Noisy stroke removal using SLA and P-SLA attribution. + +
MetricsGT SketchSLA filteredP-SLA filteredBhunia et al. [15]
Shoe-V2Acc.@133.436.136.543.7
Acc.@567.868.769.374.9
MOS28.685.757.1-
Chair-V2Acc.@153.354.956.564.8
Acc.@574.376.677.179.1
MOS35.871.457.1-
+ +Tab. 2, both Bhunia et al. [15], and ours improve fine-grained SBIR performance by $10.3\%$ and $3.1\%$ , respectively. However, [15] outperforms our SLA and P-SLA filtered methods by $7.6\%$ and $7.2\%$ , respectively. This performance gap is likely because [15] trains the baseline model [12] using actor-critic version of PPO [80], whereas our SLA and P-SLA work post-hoc [98] without training the baseline model [12]. Additionally, Bhunia et al. [15] aims to design a robust SBIR pipeline, whereas our SLA/P-SLA filtering aims to assist humans in drawing sketches. For human study (MOS): (i) users prefer SLA-filtered $78.5\%$ vs. $21.5\%$ for GT sketch, (ii) however, P-SLA-filtered are preferred only $57.1\%$ vs. $42.9\%$ for GT sketch. This is verified by Fig. 5 where P-SLA-filtered sketches have broken strokes that degrade their visual quality. + +# 5.3. Interactive Sketch To Photo Generation + +The upsurge of large-scale image generation models (e.g., Stable-Diffusion [74], GigaGAN [50]) helped develop sketch-conditional image generation [63, 118]. However, a key limitation of conditional image generation is that these models do not always faithfully follow the input condition. This was resolved in two stages for text-to-image generation: (i) find word tokens with low influence on generated image, and (ii) iteratively update its activation until it reaches a minimum required value. In this section, we design a pipeline for faithful sketch-to-image generation. + +Using sketch attributions from SLA and P-SLA, we design a post-hoc [98] method that gives feedback to the user - which strokes the model focuses on and which strokes are being ignored. Given this feedback, a user can interact with the system to ensure the model attends all salient regions. + +Our interactive pipeline is built on top of a pre-trained sketch-to-photo generation model [55], comprising a modified ResNet-50 [41] as sketch encoder and StyleGAN [52] as image decoder. Given a raster sketch X, the modified sketch encoder computes a latent vector $z_{s2p}^{+} = F_{\theta}(\mathrm{X})$ where $z_{s2p}^{+} \in \mathbb{R}^{14 \times 512}$ . Next, the StyleGAN [52] decoder generates the underlying image from $z_{s2p}^{+}$ . For a faithful sketch-to-image generation, we measure the "influence" of each stroke/vector point on the latent code $z_{s2p}^{+}$ . Particularly, prior works [106] suggest that $z_{s2p}^{+}$ is disentangled into 14-level semantic feature hierarchy, where $z_{s2p}^{+1} \in \mathbb{R}^{512}$ has coarse-level features controlling major semantic struc + +![](images/daefada324147b350a086fe96ce19923fb2dcd72d509a6bd86e73f1b71ae4ea5.jpg) +Figure 6. Interactive Sketch to Photo Generation: Our stroke attribution algorithms make existing sketch-to-photo generation pipelines [55] more faithful. We achieve this by computing the stroke-level $\mathbb{A}_i^S$ or coordinate-level $\mathbb{A}_t^V$ attribution that has a maximal influence on the latent code $z_{s2p}^{+}$ used by the image decoder. + +tures and $z_{s2p}^{+14}\in \mathbb{R}^{512}$ has fine-level features controlling colour schemes, etc. Since sketches primarily convey semantic structure, we use the sum of the first 7-layers to compute stroke attribution $\mathbb{A}_i^R$ or $\mathbb{A}_t^V$ as + +$$ +\mathbb {A} _ {i} ^ {R} = \frac {\partial \sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\partial \mathrm {X}} \cdot \frac {\partial \sum_ {k = 1} ^ {m} \omega_ {k} \mathcal {S} _ {k}}{\partial \mathcal {S} _ {i}} +$$ + +$$ +\mathbb {A} _ {t} ^ {V} = \frac {\partial \sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\partial \mathrm {X}} \cdot \sum_ {\forall p _ {x}, p _ {y}} \left\{\frac {\partial \mathrm {X} \left(p _ {x} , p _ {y}\right)}{\partial v _ {t}} \right\} \tag {11} +$$ + +Next, we qualitatively evaluate our iterative sketch to photo generation pipeline, as shown in Fig. 6. + +# 5.4. Adversarial Attacks on Human Sketches + +Szegedy et al. [94] discovered that predictions by deep networks can be manipulated with extremely low-magnitude input perturbations. For images, these can be restricted to be imperceptible to human vision, but their effect can completely change the output prediction by a deep network. Such adversarial attacks are possible in image classification [94], semantic segmentation [7, 42], object detection [97, 115], object tracking [22, 48], etc. Studying these quirks is crucial as it can pose a real threat to deep learning as a pragmatic technology [3]. While major work has been dedicated to attacks on images, the recent surge of deployable sketch applications [77, 118] motivates us to present the first study on adversarial attacks for human-sketches. + +We show how our stroke attribution algorithms (SLA and P-SLA) provide the necessary information for adversarial attacks. For brevity, we focus on adversarial attacks on sketch classification [40, 111]. Intuitively, we use sketch attribution to remove the smallest stroke (in SLA Attack) and minimum number of points (in P-SLA Attack), yet have the maximum impact on changing prediction of a pre-trained classifier $F_{\theta}(\mathrm{X}) = y_{\mathrm{cls}}$ . Ours is (i) a white-box [69, 95] setting – we have access to network weights and gradients of our ResNet-18 [41] classifier, pre-trained on QuickDraw [40] or TU-Berlin [33]; (ii) untargeted attack – while tar + +Table 3. Sketch Adversarial Attacks: Using stroke attributions, we remove a small stroke $(|S_i| \leq \epsilon)$ that misclassifies an input sketch. + +
No AttackSLA AttackP-SLA Attack
ε = 5ε = 15ε = 5ε = 15
QuickDraw67.265.764.565.163.7
TU-Berlin74.971.568.570.268.1
+ +geted attacks [60, 121] misclassify $F_{\theta}(\cdot)$ from $y_{\mathrm{cls}}^{\mathrm{GT}}$ to a specific target class $y_{\mathrm{cls}}^{*}$ , untargeted attacks [93] aim to misclassify to any arbitrary class $y_{\mathrm{cls}}^{\mathrm{GT}} \neq y_{\mathrm{cls}}$ . For a neater description of SLA and P-SLA attacks, we define the rasterisation process using $\mathcal{R}(\cdot)$ as $\mathrm{X} = \omega_1\mathcal{S}_1 + \dots + \omega_m\mathcal{S}_m = \mathcal{R}(\mathcal{S})$ for SLA and following Eq. (5) for P-SLA we define $\mathrm{X} = \mathcal{R}(\{v_1,\ldots v_T\}) = \mathcal{R}(\mathrm{V})$ . Next, we find a stroke $S_{adv}$ with stroke length $|S_i|$ less than some threshold $\epsilon$ as + +$$ +\mathcal {S} _ {a d v} = \arg \max _ {| \mathcal {S} _ {j} | \leq \epsilon} \mathcal {L} _ {\mathrm {c l s}} \left(F _ {\theta} \left(\mathcal {R} \left(\mathcal {S} - \left\{\mathcal {S} _ {j} \right\}\right)\right), y _ {\mathrm {c l s}} ^ {\mathrm {G T}}\right) \tag {12} +$$ + +Unlike typical adversarial attacks on images that add a small noise $(\mathrm{X} + \Delta x)$ with $||\Delta x||_{\infty}\leq \epsilon$ , for sketch adversarial attacks, we remove a small stroke $(\mathrm{X} - S_{adv})$ such that the stroke length is less than $\epsilon$ as $|\mathcal{S}_{adv}|\leq \epsilon$ . For PSLA attack, we find a subset of $\epsilon$ vector points $\mathrm{V}_{adv} =$ $\{v_1^{adv},\dots v_\epsilon^{adv}\}$ from input sketch $\mathrm{V}\in \mathbb{R}^{T\times 5}$ which maximises the categorical cross-entropy loss $\mathcal{L}_{\mathrm{cls}}$ as + +$$ +\begin{array}{l} v _ {t} ^ {a d v} = \mathcal {L} _ {\mathrm {c l s}} \left(F _ {\theta} \left(\mathcal {R} (\mathrm {V} - \{v _ {t} \})\right), y _ {\mathrm {c l s}} ^ {\mathrm {G T}}\right) \tag {13} \\ \mathrm {V} _ {a d v} = \operatorname {t o p} @ k \left(\left\{v _ {1} ^ {a d v}, v _ {2} ^ {a d v}, \dots v _ {T} ^ {a d v} \right\}, \epsilon\right) \\ \end{array} +$$ + +where, $\mathrm{top}@\mathrm{k}(\cdot ,\epsilon)$ picks the highest $\epsilon$ elements. Fig. 7 shows the adversarial strokes $S_{adv}$ and points $\mathrm{V}_{adv}$ in red. + +Dataset: We evaluate sketch adversarial attacks on QuickDraw [40] and TU-Berlin [33]. We use a subset [103] of $50M$ sketches in QuickDraw [40] as $3.8M$ samples across 345 categories split in $2.1M$ sketches for training, $0.3M$ for validation and $0.4M$ for evaluation. See Sec. 5.1 for details on TU-Berlin [33] dataset. + +Evaluation: We measure the drop in classification accuracy when using SLA and P-SLA attacks in Tab. 3. Unlike image-based adversarial attacks where $\epsilon$ is a pixel intensity (non-integer, decimal value), our sketch attacks occur on stroke/point length, making $\epsilon$ an integer. A higher $\epsilon \geq 20$ removes "visible" strokes in SLA and broken lines in P-SLA (Fig. 7). Hence, we evaluate accuracy drop for $\epsilon = 5$ and $\epsilon = 15$ in Tab. 3. We observe for $\epsilon = 5$ P-SLA offers a better adversarial attack than SLA by a margin of $1.3\%$ . + +# 6. Human Study + +Interpretability aims to help humans understand a model's reasoning process (transparency), verify that its predictions are based on the right constraints (fairness), and evaluate its confidence (trustworthiness) [73, 81]. In Sec. 5, we evaluated interpretability using several automatic metrics (e.g., + +![](images/9cf77f56775305bf24959d400911878cad02868649a9647f3052cc029a52b0a8.jpg) +Figure 7. Adversarial attacks on human drawn sketches using SLA and P-SLA. The adversarial strokes are marked in RED. + +![](images/91fd081462594502b2646ec0a3efcf609c8e57ca1bb300821047060e05535b6c.jpg) + +![](images/6b3a6af5e965fd9930020cf7cf5177a99aec8da7511cfef5f57fad78829c2440.jpg) +GT:Angel + +![](images/bcb3f9f886c5b283fd2482c167a0bd8f5971b284b48e2bd385c79f6ee096c679.jpg) + +![](images/47c221124c8a01a26abe3d82aace0fd4d53dc718b1d092c4a15cbe27c38e2690.jpg) +6T: Cactus +Butterfly + +![](images/d4c1c319ed213e7a6af9aad8bb8a8c62ea4b039aabb19c9509095c7c4e05ee95.jpg) + +![](images/ee1534eb5b4d886c4951623f1fbf12d1c5447f3988cac2437134d236471854d7.jpg) +GT: Mushroom + +classification or retrieval accuracy) on different evaluation datasets [33, 40, 79]. However, highlighting salient regions by backpropagating gradients for downstream applications does not capture how helpful end-users find these attributions [53, 73]. In this section, we take a human-centred approach to interpretability – how well our stroke attribution algorithms align with the reasoning process of humans and the trade-off, interpretability vs. accuracy. + +Setup: We recruited 7 participants from different geographical regions, in the age group 20 - 30 years. All participants had some background in AI research, but only 3 reported having prior experience in interpretability. Once recruited, each user is assigned a unique ID for anonymity. For SLA and P-SLA, we conduct 5 human studies, with each having 10 multiple choice questions (MCQs). Hence, each participant answers 50 MCQs for SLA and 50 for P-SLA. + +Evaluating Transparency: For an attribution to be useful, humans must understand a model's behaviour for correct and incorrect predictions. In this section, we evaluate if our SLA and P-SLA can make existing sketch classifiers (i.e., Sketch-A-Net [111]), pre-trained on TU-Berlin [33] dataset transparent to humans. Accordingly, we choose a random category and select 4 sketch instances - (i) 3 misclassified and 1 correctly classified, and (ii) 1 misclassified and 3 correctly classified. Next, as shown in Fig. 8, we compute stroke attribution (SLA and P-SLA) for the selected sketches and ask users: "Only 1 of these 4 sketches are correctly (or incorrectly) recognised by our model. Please select that correct (or incorrect) sketch." We find users can identify correct/incorrect model predictions $75.9\% /63.4\%$ of times for SLA and $76.3\% /65.2\%$ for P-SLA. + +Q. Only one sketch is correctly categorised. Which one? + +![](images/df06294eb6bd9999b6c75663538cd46f9a2767d5865735c1a6d904e2bc219717.jpg) + +![](images/cbacf80aa0a4c755660163bd6308eb1731cac93aebbfd695849925c3e49662de.jpg) + +![](images/c35654e0be2f2813f7c613938e91df8f516a0281ae5ee1f3987bc360ce3630d7.jpg) + +![](images/ec686d1bfb4bf068059b84efd76c23efcbd38ac9b2e55701a0ee08c2cf862f8c.jpg) + +Q. Only one sketch is incorrectly categorised. Which one? + +![](images/fe58f5eb1836c79254ebe16a0683accc90994209e8b1db65b7753307bafd9dc1.jpg) +Figure 8. Evaluating transparency: Can a human understand the behaviour of an existing (pre-trained) classifier with SLA/P-SLA. + +![](images/802e104f3e8afafa29b293696fb752002a58b12a1015deba569f2009ea6e07e3.jpg) + +![](images/a85669fd7d1de5d17bb2eee4edd26d9c947bf5f2d194544bbe90e541145d5f38.jpg) + +![](images/fb91e26f7fa6a93adf2487953ac86215e65bf52a3716626b862fb3b5209ab1b1.jpg) + +Evaluating Fairness: End users are much better positioned to make a decision with help from a model if intelligible explanations are provided. In this section, we evaluate if SLA and P-SLA can help end users understand "What went wrong?" (i) For a pre-trained sketch classifier [111], we show users a misclassified sketch instance and ask users + +to identify the (wrongly) predicted category. (ii) For fine-grained SBIR [13], we show a sketch (whose GT photo is not in top-10) and ask users to identify the top-1 (wrongly) retrieved photo in Fig. 9. Humans can identify the misclassified category $62.4\%$ $(66.3\%)$ and the incorrectly retrieved photo $39.1\%$ $(37.2\%)$ for SLA (P-SLA). + +Q. Prediction is incorrect, identify predicted category + +Q. Retrieval is incorrect, identify the top-1 retrieved image + +② ${Na} + {O2} < \Delta$ 释放 + +(A) Angel (B)Butterfly + +(C) Fan (D) Dog + +![](images/d483d95960b9a87048f9e77371934cc57ec77a22ff3daf9b7a402b97350186cc.jpg) +Figure 9. Evaluating Fairness: For an incorrect model prediction, we evaluate if humans can "identify what went wrong". + +![](images/9bc0a0b8552958f5393fb9398705b09175e1b100dfef0c43c021632aacbcee68.jpg) + +![](images/dc0d47a686c86eeb1a7d3a5cded5b872ce73c13fa5bf25f3a1c7d87b47a5ed7f.jpg) + +![](images/b7d09586483de2d7165b23f739bc44328102e30512b948786bc7fcc9b229f6da.jpg) + +**Evaluating Trustworthiness:** Determining trust in individual predictions is important when used for decision-making (e.g., medical diagnosis [57]). We train two copies of the same sketch classifier [111], a strong classifier with $73.8\%$ accuracy on TU-Berlin [33] and a weak one reaching $57.1\%$ . We present the stroke attribution (SLA/P-SLA) for both models and ask users to identify the strong/weak classifier, as shown in Fig. 10. Humans identify the stronger classifier $71.4\% (68.7\%)$ of the time for SLA (P-SLA). + +Q. Prediction is correct, identify the stronger classifier + +![](images/c15ed2643b4751d5a0dd526bdf3f0e1147726e2c69b75b8ecfd8c9ac47adb6e6.jpg) +Figure 10. Evaluating trustworthiness: We present the stroke attributions (SLA/P-SLA) from two sketch classifiers (strong and weak), and ask users to identify the strong/weak model. + +![](images/10a97bcb851270199e12f9c43445718e8a17d93e3d7fc8886ca01ab5f1edd306.jpg) + +![](images/79855170b30d002b8931b83b892766649f1974b88a087aa48c0160eb164a5c55.jpg) + +![](images/5e3f1a99612e002d3faea49e09e9047e06263dbd7eb2e248bedd42e1e1ee1e59.jpg) + +![](images/c451294a4071bf72682eb5d01b9d66cfa6ab1dc4bc7819ca09c349fecd010f69.jpg) + +# 7. Conclusion + +This work emphasises the pivotal role of strokes in human-drawn sketches, offering unique insights compared to pixel-based images. Our lightweight explainability solution seamlessly integrates with pre-trained models, addressing rasterisation challenges and contributing to diverse sketch-related tasks. Through applications in Retrieval, Generation, Assisted Drawing, and Sketch Adversarial Attack, our model showcases adaptability and significance. The proposed stroke-level attribution provides nuanced insights into model behaviour, underscoring the importance of explainability in bridging human expression with model predictions in the evolving field of sketch interpretation. + +# References + +[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 1, 2, 3 +[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. arXiv preprint arXiv:2011.05429, 2020. 2 +[3] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. arXiv preprint arXiv:2108.00401, 2021. 7 +[4] Shm Garanganao Almeda, J.D. Zamfirescu-Pereira, Kyu Won Kim, Pradeep Mani Rathnam, and Bjoern Hartmann. Prompting for discovery: Flexible sense-making for ai art-making with dreamsheets. arXiv preprint arXiv:2310.09985, 2023. 2 +[5] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104, 2017. 2 +[6] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. A unified view of gradient-based attribution methods for deep neural networks. In NIPSW, 2017. 2 +[7] Anurag Arnab, Ondrej Miksik, and Philip H.S. Torr. On the robustness of semantic segmentation models to adversarial attacks. In CVPR, 2018. 7 +[8] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 2015. 2 +[9] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In CVPR, 2024. 2 +[10] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 2 +[11] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 2 +[12] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 5, 6 +[13] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 8 +[14] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang, + +and Yi-Zhe Song. Doodle it yourself: Class incremental learning by drawing a few sketches. In CVPR, 2022. 2 +[15] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 4, 5, 6 +[16] Moritz Bohle, Mario Fritz, and Bernt Schiele. B-cos: Networks: Alignment is all we need for interpretability. In CVPR, 2022. 3 +[17] Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well onImagenet. arXiv preprint arXiv:1904.00760, 2019. 2 +[18] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In ICCV, 2015. 2 +[19] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. In NeurIPS, 2019. 2 +[20] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM TOG, 2020. 2 +[21] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In ICCV, 2018. 2 +[22] Xuesong Chen, Xiyu Yan, Feng Zheng, Yong Jiang, Shu-Tao Xia, Yong Zhao, and Rongrong Ji. One-shot adversarial attacks on visual tracking with dual attention. In CVPR, 2020. 7 +[23] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 2, 5 +[24] Pinaki Nath Chowdhury, Aneeshan Sain, Yulia Gryaditskaya, Ayan Kumar Bhunia, Tao Xiang, and Yi-Zhe Song. Fs-coco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2 +[25] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation: Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 5 +[26] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2, 5 +[27] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In NeurIPS, 2017. 2 +[28] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3 +[29] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 5 + +[30] Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron? arXiv preprint arXiv:1805.12233, 2018. 2 +[31] Gabriele Dominici, Pietro Barbiero, Lucie Charlotte Magister, Lio Pietro, and Nikola Simidjevski. Sharcs: Shared concept space for explainable multimodal learning. arXiv preprint arXiv:2307.00316, 2023. 3 +[32] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 2 +[33] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 4, 5, 7, 8 +[34] Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. Technical Report, Université de Montréal, 2009. 2 +[35] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. arXiv preprint arXiv:2112.04417, 2021. 2 +[36] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In ICCV, 2019. 2 +[37] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2 +[38] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In CVPR, 2020. 2 +[39] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In CVPR, 2021. 2 +[40] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 3, 4, 7, 8 +[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 6, 7 +[42] Yang He, Shadi Rahimian, Bernt Schiele, and Mario Fritz. Segmentations-leak: Membership inference attacks and defenses in semantic image segmentation. In ECCV, 2020. 7 +[43] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2 +[44] Josh Holinaty, Alec Jacobson, and Fanny Chevalier. Supporting referenceimageryfordigitaldrawing. In ICCVW, 2021. 1 +[45] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2 +[46] Quan Huynh-Thu, Marie-Neige Garcia, Filippo Speranza, Philip Corriveau, and Alexander Raake. Study of ratling scales for subjective quality assessment of high definition video. IEEE TBC, 2010. 6 +[47] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 5 +[48] Yunhan Jia, Yantao Lu, Junjie Shen, Qi Alfred Chen, Hao Chen, Zhenyu Zhong, and Tao Wei. Fooling detection alone + +is not enough: Adversarial attack against multiple object tracking. In ICLR, 2020. 7 +[49] Zeyang Jia, Eli Ben-Michael, and Kosuke Imai. Bayesian safe policy learning with chance constrained optimization: Application to military security assessment during the vietnam war. arXiv preprint arXiv:2307.08840, 2023. 3 +[50] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6 +[51] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In ICCV, 2019. 2, 3 +[52] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6 +[53] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In ECCV, 2022. 2, 3, 8 +[54] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, 2019. 2 +[55] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1, 4, 5, 6, 7 +[56] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 3 +[57] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 8 +[58] Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Sam Gershman, and Finale Doshi-Velez. An evaluation of the human-interpretability of explanation. arXiv preprint arXiv:1902.00006, 2019. 2 +[59] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In WACV, 2019. 2 +[60] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 7 +[61] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, Yi-Zhe Song, and Yonggang Qi. Zero-shot everything sketch-based image retrieval, and in explainable style. In CVPR, 2023. 1, 2 +[62] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Defloconet: Deep image editing via flexible low-level controls. In CVPR, 2021. 2 +[63] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 5, 6 + +[64] Meike Nauta, Ron Van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In CVPR, 2021. 2 +[65] Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. ICMLW, 2016. 2 +[66] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. 2 +[67] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 6 +[68] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2 +[69] Buu Phan, Fahim Mannan, and Felix Heide. Adversarial imaging pipelines. In CVPR, 2021. 7 +[70] Zhiyu Qu, Yulia Gryaditskaya1, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In CVPR, 2023. 1, 2 +[71] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 5 +[72] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards better understanding attribution methods. In CVPR, 2022. 2 +[73] Marco T. Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?" explaining the predictions of any classifier. In KDD, 2016. 3, 7, 8 +[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Omer. High-resolution image synthesis with latent diffusion. In CVPR, 2022. 6 +[75] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019. 3 +[76] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 2 +[77] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 1, 3, 5, 7 +[78] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE TNNLS, 2016. 2 +[79] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 4, 5, 8 +[80] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Pronimal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5, 6 +[81] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba + +tra. Grad-cam: Visual explanations from deep networks via gradient-based localisation. In ICCV, 2017. 2, 3, 7 +[82] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2 +[83] Harshay Shah, Prateek Jain, and Praneeth Netrapalli. Do input gradients highlight discriminative features? In NeurIPS, 2021. 2 +[84] Hamid Rahim Sheikh and Alan C. Bovik. Image information and visual quality. IEEE TIP, 2006. 6 +[85] Yuming Shen, Li Liu, Fumin Shen, and Ling Shao. Zeroshot sketch-image hashing. In ICCV, 2018. 2 +[86] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In ICML, 2017. 2, 3 +[87] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In ICLRw, 2014. 2, 3 +[88] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiβ, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. arXiv preprint arXiv:2204.11642, 2022. 2 +[89] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 2 +[90] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1 +[91] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In ICLRw, 2015. 2, 3 +[92] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2, 3 +[93] Christian Szegedy, Wojciech Zaremba, Ilyua Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7 +[94] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7 +[95] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white- and black-box attacks. In NeurIPS, 2020. 7 +[96] Aditay Tripathi, Rajath R Dani, Anand Mishra, and Anirban Chakraborty. Sketch-guided object localization in natural images. In ECCV, 2020. 2 +[97] James Tu, Mengye Ren, Siva Manivasagam, Ming Liang, Bin Yang, Richard Du, Frank Cheng, and Raquel Urtasun. Physically realizable adversarial examples for lidar object detection. In CVPR, 2020. 7 + +[98] Hugues Turbé, Mina Bjelogrlic, Christian Lovis, and Gianmarco Mengaldo. Evaluation of post-hoc interpretability methods in time-series classification. Nature Machine Intelligence, 2023. 1, 6 +[99] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In CVPR, 2021. 2 +[100] Donglai Wei, Joseph Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, 2018. 4 +[101] Dennis Wei, Rahul Nair, Amit Dhurandhar, Kush R. Varshney, Elizabeth M. Daly, and Moninder Singh. On the safety of interpretable machine learning: A maximum deviation approach. In NeurlPS, 2022. 3 +[102] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: deep sketch-based hair image synthesis. ACM TOG, 2021. 1 +[103] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 5, 7 +[104] Rui Xu, Zongyan Han, Le Hui, Jianjun Qian, and Jin Xie. Domain disentangled generative adversarial network for zero-shot sketch-based 3d shape retrieval. AAAI, 2022. 2 +[105] Guowei Yan, Zhili Chen, Jimei Yang, and Huamin Wang. Interactive liquid splash modeling by user sketches. ACM TOG, 2020. 2 +[106] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 6 +[107] Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Suggala, David I Inouye, and Pradeep K Ravikumar. On the (in) fidelity and sensitivity of explanations. In NeurIPS, 2019. 2 +[108] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A zero-shot framework for sketch based image retrieval. In ECCV, 2018. 5 +[109] Ran Yi, Zipeng Ye, Ruoyu Fan, Yezhi Shu, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Animating portrait line drawings from a single face photo and a speech signal. In ACM SIGGRAPH, 2022. 2 +[110] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In CVPR, 2019. 2 +[111] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 3, 7, 8 +[112] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 3, 5, 6 +[113] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2, 3 +[114] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In CVPR, 2022. 2 +[115] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, 2019. 7 +[116] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In CVPR, 2016. 5 + +[117] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. IJCV, 2018. 2 +[118] Lvmin Zhang, Anyi Rao, and Maneesh Agarwala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5, 6, 7 +[119] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2 +[120] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2 +[121] Hang Zhou, Dongdong Chen, Jing Liao, Weiming Zhang, Kejiang Chen, Xiaoyi Dong, Kunlin Liu, Gang Hua, and Nenghai Yu. Lg-gan: Label guided adversarial network for flexible targeted attack of point cloud-based deep networks. In CVPR, 2020. 7 +[122] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchscene: Richly-annotated scene sketches. In ECCV, 2018. 2 \ No newline at end of file diff --git a/2024/What Sketch Explainability Really Means for Downstream Tasks_/images.zip b/2024/What Sketch Explainability Really Means for Downstream Tasks_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b939def2665e82a15603460d295d3fabf2e5b485 --- /dev/null +++ b/2024/What Sketch Explainability Really Means for Downstream Tasks_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea1f512faa8e56227876037cd3f051f860747911d28bb128fdf0cdae04aca6ae +size 411372 diff --git a/2024/What Sketch Explainability Really Means for Downstream Tasks_/layout.json b/2024/What Sketch Explainability Really Means for Downstream Tasks_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d22786b0d896e3f8d89d0d6d1500485bc0f185f --- /dev/null +++ b/2024/What Sketch Explainability Really Means for Downstream Tasks_/layout.json @@ -0,0 +1,14525 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 499, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 499, + 120 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 499, + 120 + ], + "type": "text", + "content": "What Sketch Explainability Really Means for Downstream Tasks?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 137, + 487, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 137, + 487, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 487, + 194 + ], + "type": "text", + "content": "Hmrishav Bandyopadhyay1 Pinaki Nath Chowdhury1 Ayan Kumar Bhunia1 \nAneeshan Sain1 Tao Xiang1,2 Yi-Zhe Song1,2 \n1SketchX, CVSSP, University of Surrey, United Kingdom. \n2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 196, + 506, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 196, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 86, + 196, + 506, + 207 + ], + "type": "text", + "content": "{h.bandyopadhyay, p.chowdhury, a.bhunia, a.sain, t.xiang, y.song}@surrey.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 251, + 290, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 251, + 290, + 468 + ], + "spans": [ + { + "bbox": [ + 45, + 251, + 290, + 468 + ], + "type": "text", + "content": "In this paper, we explore the unique modality of sketch for explainability, emphasising the profound impact of human strokes compared to conventional pixel-oriented studies. Beyond explanations of network behavior, we discern the genuine implications of explainability across diverse downstream sketch-related tasks. We propose a lightweight and portable explainability solution - a seamless plugin that integrates effortlessly with any pre-trained model, eliminating the need for re-training. Demonstrating its adaptability, we present four applications: highly studied retrieval and generation, and completely novel assisted drawing and sketch adversarial attacks. The centrepiece to our solution is a stroke-level attribution map that takes different forms when linked with downstream tasks. By addressing the inherent non-differentiability of rasterisation, we enable explanations at both coarse stroke level (SLA) and partial stroke level (P-SLA), each with its advantages for specific downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 478, + 127, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 127, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 127, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Sketches, rooted in human expression [43], offer a distinctive modality for exploring explainability [61, 70]. In contrast to photos, where each pixel is independent and lacks inherent meaning, sketches are organised into strokes, with each stroke carrying subjective meaning assigned by the sketcher [44]. This paper explores sketch explainability, but with a unique perspective – aiming to provide explanations and unravel the true implications of explainability on various downstream sketch-related tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "With this perspective in mind, our approach champions an explainability solution that is (i) lightweight and portable - a plugin seamlessly integrating with multiple pre-trained models without necessitating re-training [98], and (ii) easily adaptable to a diverse array of downstream sketch-specific tasks, benefiting the broader community." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Our solution is exclusively centred on human strokes, aiming to attribute explanation on different stroke granularity: individual strokes (coarse) and their parts (fine). The" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 305, + 228, + 545, + 366 + ], + "blocks": [ + { + "bbox": [ + 305, + 228, + 545, + 366 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 366 + ], + "type": "image", + "image_path": "482f4d2870546db7edb97a1621470c70cadaa0d41a4106b3b4b329f57e112ef4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 376, + 546, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 498 + ], + "type": "text", + "content": "Figure 1. We attribute explanations for individual strokes (stroke-level attribution) and their vector coordinate points (point-level attribution). Stroke-level attribution rasterises individual strokes (non-differentiably) to produce " + }, + { + "bbox": [ + 304, + 376, + 546, + 498 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 376, + 546, + 498 + ], + "type": "text", + "content": "-stroke images. Next, we sum the stroke images to get the complete sketch image used for downstream tasks. Point-level Attribution computes distance transform from stroke coordinates and thresholds to get the sketch image. Our explainability solution works without re-training for existing tasks like SBIR and sketch-to-photo generation and novel tasks like filtering noisy strokes for assisted drawing and adversarial attack by removing a small stroke." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 507, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 663 + ], + "type": "text", + "content": "output of our model markedly differs from typical saliency maps [1] found in photo-based explainability models, where the emphasis is mostly on visualisation [90]. Ours is a task-driven attribution map that assigns stroke-level attributes capturing how altering stroke characteristics can impact model prediction. Depending on the downstream tasks, attributions can be grounded to, for example (i) importance of entire strokes, which is more suitable to filter noisy strokes [15] in assisted drawing, and remove small strokes for adversarial attacks on existing sketch encoders, and (ii) stroke shape and length, where a partial-stroke level attribution is beneficial for tasks like sketch-based image retrieval [77] and sketch-to-photo generation [55]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "To showcase the adaptability of our model, we carefully devise four applications: two well-studied tasks from existing literature (retrieval [12, 26] and generation [55, 102]), and two entirely novel tasks (assisted drawing and sketch" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10997" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "adversarial attack). In retrieval, we evaluate reliability of model predictions by comparing predicted stroke order with the order in which a human draws them. For generation, we pinpoint strokes with the least influence, offering explicit feedback to end-users regarding which strokes the model prioritised and which it overlooked. In assisted drawing [4], we assist novice artists in faithfully sketching a particular photo by identifying strokes that do not match the target photo. Lastly, in adversarial attacks, we unveil the vulnerability of state-of-the-art sketch encoders by removing a small imperceptible stroke in any sketch, resulting in significant changes to the model's prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 479 + ], + "type": "text", + "content": "The focal point connecting all downstream tasks is our proposed stroke-level attribution. The key question, therefore, is how to backpropagate information to strokes while addressing the inherent non-differentiability of rasterisation - strokes are most often represented as discrete coordinates and rasterised before feeding into downstream applications. We provide two solutions for non-differentiability: (i) coarse stroke level: we first rasterise individual strokes to produce raster stroke images. Then, we combine these stroke images to get the complete sketch image (see Fig. 1) - since this addition of stroke images is a differentiable operation, we can backpropagate information from the complete sketch to individual raster stroke images. (ii) fine partial-stroke level: we create a distance transform image from stroke coordinates (\"red dots\" in Fig. 1 vector sketch) by calculating the minimum distance of each pixel in the image from the coordinates. Then, we threshold the distance value to get the sketch image (white pixels for a high distance and black pixels for a low distance). Since the distance function and our threshold step are both differentiable, we can backpropagate information from the sketch images to stroke coordinates." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 480, + 289, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 289, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 289, + 578 + ], + "type": "text", + "content": "Our contributions can be summarised as follows: (i) We explore sketch explainability, emphasising the importance of strokes in human-drawn sketches. (ii) We highlight the profound impact of explainability on various sketch-related domains, presenting applications in retrieval, generation, assisted Drawing, and adversarial attacks. (iii) We solve for the non-differentiability problem of rasterisation, and provide both stroke-level and partial-stroke level attribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 587, + 139, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 139, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 139, + 600 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Sketch for Visual Understanding: Having a high visual proximity [43] to real images and carrying human subjectivity [9, 76], amateur sketches or abstract line drawings [59] has been a popular modality for customised expression, thus driving extensive applications as a query for retrieval [26, 32, 76, 85] of object [29] or scene [24] images, 3D shapes [104], and even concepts like in 'dictionary-like' games [11]. As a canvas for creativity, sketch helps image-editing [62, 110, 114], or generation of objects [20, 21, 38, 99]," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "scenes [109], and 3D shapes [10, 39, 105, 119]. Being easily editable, sketch enables interactive access to AI systems like image-segmentation [45, 122], object localisation [96], image-inpainting [110, 114], and incremental learning [14]. Being application-specific however, such works largely ignored explaining the 'how' of sketch-correspondence. The few who did, customised training pipelines [61, 70] for niche tasks. In this work, we thus make the first attempt at visualising salient sketch-regions (strokes), as an explainability-tool (like GradCAM [81] in photos) for existing pre-trained sketch-based downstream networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 207, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 547, + 495 + ], + "type": "text", + "content": "Explaining CNN Predictions: CNN explanations visually highlight regions of input having the maximum influence on a model's predictions [87]. This visualisation of 'salient' regions is either through an analysis [37] of a regular pre-trained network after it has completed training (post-hoc), or by designing and training explicitly interpretable (i.e. explain-and-predict) models [17, 19, 64]. Given a pre-trained CNN, a post-hoc algorithm either visualises (i) model attributes like feature and activation-maps [34, 65, 66, 82, 113, 120] that imply saliency of specific input (pixel) regions, or (ii) input attributes directly as pixels [91] or pixel-regions [51] coloured according to their relative importance. Visualisation of input attributes is facilitated through perturbation based algorithms [18, 36, 37] and gradient-based analysis [87, 89, 91]. Perturbation-based algorithms [27, 68, 113] detect saliency of pixel-regions by measuring impact of their absence on the prediction score. Whereas, gradient-based algorithms [8, 30, 86, 92, 117] measure the gradient of the prediction with respect to individual input-pixels (input-gradients [83]), attributing them based on this value. Unlike images however, sketch is a sparse-information modality [23] for pixels. As such, we explore explainability in sketches by computing stroke and point attribution for fine-grained sketch explanations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": "Evaluation of CNN Explanations: The evaluation of CNN explanations has evolved over time - from naive qualitative analysis of visualisations [82, 87, 91, 113] to standardised theoretical [5, 92] and empirical [72, 107, 113] baselines. Analysing explanations theoretically helps evaluate them in a model-agnostic environment, where their mathematical form is checked against pre-defined properties (axioms)[5, 54, 92]. Empirical evaluations, on the other hand, involve experiments measuring " + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": " variance in explanations upon perturbations of inputs [107] and model weights [1, 2] (sanity checks), and " + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": " accuracy of explanations in locating important features [37, 72, 82, 117]. These features, when perturbed, influence the CNN maximally, as measured by perturbation-based metrics [6, 51, 68, 78, 82, 83, 113]. Recently, however, these evaluation protocols have met criticism [35] due to their detachment from humans. Instead, human studies [53, 58, 82, 88], and human-centered metrics [35] have been proposed for evaluating explanation in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10998" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "terpretability. In this work, we attribute strokes by backpropagation [86, 87, 91, 92], evaluating attributions through sanity checks [1], empirical metrics [51] and human studies [53] on downstream sketch-based applications [112]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 126, + 127, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 126, + 127, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 127, + 139 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 140, + 287, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 287, + 177 + ], + "type": "text", + "content": "Here, we provide a brief overview of some standard concepts, ubiquitous in explainability literature [1] to help formalise the question: \"what entails a good explanation?\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 179, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 286 + ], + "type": "text", + "content": "Attribution Algorithms: It highlights relevant regions (e.g., pixels in an image, see Fig. 1) that are responsible for the model's prediction. Despite its importance for safety-critical applications [49, 75, 101], making an attribution algorithm interpretable to humans remains an open problem. Surprisingly, a more faithful attribution is usually less interpretable and vice-versa [91, 113]. Prior works [16, 81] study this trade-off between faithfulness vs. interpretability as an answer to: \"What makes a good visual explanation?\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "content": "The attribution map " + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{A} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "content": " is typically calculated using gradients for an input " + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "content": " for a classification model " + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y} = F_{\\theta}(\\mathrm{X}) \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "content": ", pre-trained on " + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "content": " categories. The gradients are a simple and good indicator of how much the model prediction changes for input " + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 46, + 286, + 287, + 346 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\mathbb {A} = \\partial F _ {\\theta} (\\mathrm {X}) / \\partial \\mathrm {X} \\tag {1}", + "image_path": "d06074f3051247df5c2341af404d76787ba28dabd7aa2a4f52ddfc26d7e1f97b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 364, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 554 + ], + "type": "text", + "content": "Interpretability: It is the ability of an attribution algorithm to provide a qualitative \"understanding\" for a model [73]. This \"understanding\" depends on the target audience, e.g., a human expert may interpret a small Bayesian network [56], but a layman is more comfortable with a weighted attention (or feature) map that highlights salient regions [1]. To evaluate the interpretability of attribution maps, prior works either (i) perform downstream tasks [31, 81] that depend on the interpretation (e.g., object localisation – predict the bounding box and semantic segmentation for image region with the highest attribution) or (ii) human studies [53], typically conducted in two setups – class discriminative (given an attribution map, ask users to identify the category predicted by a model), and trustworthiness (compare attribution maps from a strong and a weak model, and ask users to identify the stronger model)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "Faithfulness: It is the ability of attribution algorithms to accurately \"explain\" the computation learned by a model. For example, in theory, a fully faithful attribution is the entire model (e.g., ResNet-18 [41]) but is not interpretable by a human. In practice, for an attribution to be meaningful, it is often impossible to be completely faithful. To balance this trade-off, prior works explore human-interpretable attributions that are locally faithful for a given model prediction. One approach is image occlusion [73], where the difference in the model scores is measured when masking different patches in an input image. Image patches that significantly change the model score are deemed important by the attribution algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 71, + 411, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 411, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 411, + 84 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 270 + ], + "type": "text", + "content": "The attribution map " + }, + { + "bbox": [ + 304, + 91, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 304, + 91, + 545, + 270 + ], + "type": "text", + "content": " in Eq. (1) gives a faithful fine-grained explanation for each pixel in X. However, such pixel attribution does not provide meaningful explanations when interpreting sparse human-drawn sketches (many empty white pixels). Additionally, pixel attribution does not consider the sketch construction process - humans sketch a sequence of strokes, not pixels. Hence, it is most appropriate that sketch predictions should be attributed to strokes and not pixels. However, the key challenge to stroke attributions is that most sketch applications [77, 118] use raster sketches - a non-differentiable process to convert a sequence of strokes into pixels. In the following sections, we propose two stroke attribution algorithms that consider the sketch construction process by designing a differentiable rasterisation pipeline for a vector sequence of strokes." + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 316, + 294, + 531, + 514 + ], + "blocks": [ + { + "bbox": [ + 313, + 281, + 501, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 501, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 501, + 291 + ], + "type": "text", + "content": "Algorithm 1: Non-differentiable Rasterisation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "lines": [ + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": "Data: V " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " Vector Sketch of size " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times 5}" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " \nResult: X " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " Blank (Zero) Canvas of size " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h\\times w\\times 3}" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " . \n" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "\\mathbb{B}(\\cdot ,\\cdot)\\gets" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " Bresenham Function ; \n" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "v_{0} = (x_{0},y_{0},q_{0}^{1},q_{0}^{2},q_{0}^{3})\\gets \\mathrm{V}[0]" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " . \n" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{prev}}\\gets v_0" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " . \n" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{prev}}\\gets q_0^1" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " . \nfor " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})\\gets \\mathrm{V}[1\\dots T]" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " do if " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{prev}} = 1" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "q_t^1 = 1" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " then for " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "(p_x^i,p_y^i)\\gets \\mathbb{B}(v_{\\mathrm{prev}},v_t)" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " do |X " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "(p_x^i,p_y^i)\\gets 255;" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " end end " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{prev}}\\gets v_t" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{prev}}\\gets q_t^1" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " . if " + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "inline_equation", + "content": "q_{t}^{3} = 1" + }, + { + "bbox": [ + 316, + 294, + 531, + 514 + ], + "type": "text", + "content": " then exit(); /\\* End of Drawing \\*/ end" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 528, + 438, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 438, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 438, + 540 + ], + "type": "text", + "content": "4.1. Sketch Representations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": "While sketches are used in several formats (or representations) like Raster [111], Vector [40], or Bézier [28], digital sketches are primarily captured in vector form, as a list of points traced on a drawing pad. These points are usually a five-element vector " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "v_{t} = (x_{t},y_{t},q_{t}^{1},q_{t}^{2},q_{t}^{3})" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": " where, " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "(x_{t},y_{t})" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": " are absolute coordinates in a " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "(H\\times W)" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": " drawing canvas and the last three elements are one-hot encoding of pen-states: pen touching the paper " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "(1,0,0)" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": ", pen is lifted " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "(0,1,0)" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": ", and end of drawing " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "(0,0,1)" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": ". Hence, a vector sketch with " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": " points is represented as " + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\in \\mathbb{R}^{T\\times 5}" + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "As most downstream sketch applications work on rasterised sketches " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": ", prior works [111] translate (non-differently) a vector sketch " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{V}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " to its equivalent raster sketch " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " using Algorithm 1. For this, the Bre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10999" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 282, + 142 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 282, + 142 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 282, + 142 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 282, + 142 + ], + "type": "image", + "image_path": "97dc47e35b6d7b1ce1fc4bc686694c42bb19e462417b7bebceca4c9ff4c672ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "type": "text", + "content": "Figure 2. Coarse Stroke-level Attribution. Backpropagate gradients from raster sketch X to raster strokes " + }, + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "type": "text", + "content": ", with weight " + }, + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 46, + 144, + 286, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "content": "senham function " + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{B}(\\cdot)" + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "content": " is used to connect two vector points " + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\{v_{t - 1}, v_t\\} \\in \\mathrm{V}" + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "content": " in the pixel space " + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\{(p_x^1, p_y^1), \\ldots, (p_x^n, p_y^n)\\} \\in \\mathbb{B}(v_{t - 1}, v_t)" + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "content": " via a continuous line. Next, we show how our sketch attributions overcome this non-differentiable rasterisation and backpropagate gradients (Eq. (1)) from pixels " + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "content": " to strokes and points in " + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{V} \\in \\mathbb{R}^{T \\times 5}" + }, + { + "bbox": [ + 47, + 175, + 287, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 248, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 248, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 248, + 270 + ], + "type": "text", + "content": "4.2. Coarse Stroke-level Attribution (SLA)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": "In this section, we backpropagate gradients from pixel space in " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": " to strokes, defined as a continuous set of points " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\{v_t, v_{t+1}, \\ldots, v_{t+n}\\}" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": " from the first pen-down " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "(1,0,0)" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": " till the pen-up " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "(0,1,0)" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": " state. Algorithm 1 converts the vector stroke points into a raster stroke " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "S_i \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": ". The final raster sketch is then a differentiable composition of " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": " raster strokes " + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{X} = \\sum_{k=1}^{m} S_k" + }, + { + "bbox": [ + 46, + 277, + 288, + 361 + ], + "type": "text", + "content": ". We compute SLA as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 371, + 287, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 371, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 82, + 371, + 287, + 397 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathcal {S} _ {i}} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {2}", + "image_path": "9d5d3678929e86118d4a1acdfc0a36e19280faa160e3052f50f2901f2f91c10a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "text", + "content": "This, however, gives a degenerate solution where all strokes will have the same attribution " + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\partial \\sum_{k=1}^{m} S_k / \\partial(S_i) = 1" + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "text", + "content": ". To avoid this, we compute a weight factor " + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\omega_i \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "text", + "content": " for each stroke " + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 46, + 407, + 287, + 455 + ], + "type": "text", + "content": " such that," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 464, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 464, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 74, + 464, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\omega_ {i} \\left(p _ {x}, p _ {y}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\left(p _ {x}, p _ {y}\\right) \\in \\mathbb {B} \\left(v _ {t - 1}, v _ {t}\\right) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "1c64046c09eb7a8162c8b7bf8230cd44be707448c47604f145cc41ad34d826af.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": "In other words, given consecutive vector sequence of points " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "(v_{t - 1}, v_t)" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " in stroke " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": ", we find all points " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "(p_x, p_y)" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " using Bresenham function " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbb{B}(\\cdot)" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " that lie \"on the stroke\" and assign " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\omega_i(p_x, p_y) = 1" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": ". Hence, longer strokes will have more 1's compared to shorter strokes. Finally, Eq. (2) is adapted as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 575, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 575, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 100, + 575, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {4}", + "image_path": "eb50156d00d112bcaa17b810d1f34d3131725030fed3a6c79b02cfcc864586b9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "text", + "content": "SLA makes the non-differentiable rasterisation " + }, + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "inline_equation", + "content": "(\\mathrm{V} \\to \\mathrm{X})" + }, + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "text", + "content": " partially differentiable (stokes " + }, + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "text", + "content": " to sketch X). In other words, SLA answers \"Which strokes in a sketch are important\". Next, we make the rasterisation fully differentiable and backpropagate gradients to vector points V to answer \"Which point in a sketch is important\"." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 535, + 150 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 535, + 150 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 535, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 535, + 150 + ], + "type": "image", + "image_path": "f6d2bf8074d1ead883ebc215632550611d23976834afd81b2473cbf07644f9a2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 151, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 151, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 151, + 545, + 173 + ], + "type": "text", + "content": "Figure 3. Partial Stroke-level Attribution. Backpropagate gradients from raster sketch X to vector sequence of coordinates V." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 176, + 517, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 176, + 517, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 517, + 190 + ], + "type": "text", + "content": "4.3. Partial Stroke-level Attribution (P-SLA)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": "Unlike SLA, which partially captures the sketch construction process (stroke-level), partial stroke-level attribution (P-SLA) can fully backpropagate gradients to the vector list of coordinates " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{V} \\in \\mathbb{R}^{T \\times 5}" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": " traced on a drawing pad. Given a blank canvas " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": ", we (i) calculate the minimum distance of every pixel " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "(p_x, p_y)" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": " from a line segment " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "(v_{t-1}, v_t)" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{V}" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": ", and (ii) compute the pixel intensity of " + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{X}(p_x, p_y)" + }, + { + "bbox": [ + 305, + 194, + 545, + 291 + ], + "type": "text", + "content": " as function of the minimum distance as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 299, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 299, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 318, + 299, + 545, + 337 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {X} \\left(p _ {x}, p _ {y}\\right) = \\sigma [ 2 - 5 \\min _ {t = 2, \\dots , T} (5) \\\\ \\left. \\mathrm {d i s t} ((p _ {x}, p _ {y}), v _ {t - 1}, v _ {t}) + (1 - q _ {t - 1} ^ {1}) 1 0 ^ {6}) \\right] \\\\ \\end{array}", + "image_path": "963cef8308a2f1d71d4b15141ae7b566d0f906ba072cf7e13628cc66e78e27dc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": " is the sigmoid function, and " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(\\cdot)" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": " is a distance function (see Supp.) from a point " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "(p_x,p_y)" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": " to a line segment " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "(v_{t - 1},v_t)" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": ". For pen-up states " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "(q_{t - 1}^{1} = 0)" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": ", we blow up " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "(\\times 10^{6})" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": " the distance values that make the pixel intensities " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{X}(p_x,p_y)\\to 0" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": ", i.e., not render strokes for " + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "inline_equation", + "content": "(v_{t - 1},v_t)" + }, + { + "bbox": [ + 305, + 344, + 545, + 416 + ], + "type": "text", + "content": ". Finally, we compute P-SLA" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 424, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 424, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 424, + 545, + 455 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial v _ {t}} = \\frac {\\partial F _ {\\theta} (\\mathrm {X})}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {6}", + "image_path": "28b8987cf5011284e0bf712bbfe4a2c4076114a0661ee6966a34b1df94a4309d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 464, + 501, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 501, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 501, + 478 + ], + "type": "text", + "content": "5. Applications of Stroke Attributions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 484, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 591 + ], + "type": "text", + "content": "Despite its simplicity, designing stroke attribution algorithms that capture the sketch construction process unlocks insights into numerous existing downstream tasks like classification [40], robust sketch-based image retrieval (category-level [33] and fine-grained [79]) and enables some novel sketch applications like assisted drawing [15], interactive sketch to photo generation [55], adversarial attacks on human-drawn sketches, and discovering the \"arrow of time\" [100] in raster sketches." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 599, + 504, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 504, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 504, + 612 + ], + "type": "text", + "content": "5.1. Robust Sketch Based Image Retrieval" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "content": "Given a query sketch " + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "content": ", category-level sketch-based image retrieval (SBIR) aims to fetch category-specific photos from a gallery of multi-category photos (e.g., given sketch of a 'shoe' retrieve any photo 'shoe' from a gallery of 'shoes+hats+cows'). Conversely, fine-grained SBIR aims to retrieve one instance from a gallery of the same category photos (e.g., given the sketch of a 'shoe' retrieve one photo shoe from a gallery of all shoes). Deep learning frameworks" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "type": "text", + "content": "For overlapping strokes in " + }, + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 47, + 693, + 286, + 715 + ], + "type": "text", + "content": ", we clamp the maximum pixel value using differentiable functions like torch.clamp" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11000" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 123, + 288, + 191 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 117 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 117 + ], + "type": "text", + "content": "Table 1. Stroke attribution (SLA, P-SLA) make SBIR systems reliable. Sketches with a high correlation " + }, + { + "bbox": [ + 47, + 70, + 289, + 117 + ], + "type": "inline_equation", + "content": "(Corr)" + }, + { + "bbox": [ + 47, + 70, + 289, + 117 + ], + "type": "text", + "content": " of stroke saliency (predicted by SLA or P-SLA) with human-drawn temporal stroke order tend to have higher retrieval accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 123, + 288, + 191 + ], + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 191 + ], + "type": "table", + "html": "
MetricsFull DatasetCorr ≥ 0.5Corr ≤ 0.1
SLAP-SLASLAP-SLA
CategorymAP53.155.357.651.750.1
LevelP@20065.966.768.564.661.5
FineAcc.@115.316.417.613.812.7
GrainedAcc.@534.236.939.431.128.3
", + "image_path": "6c7cf5c6642548aed8d9977a02b4656b1294e1207f9dfcfd860d1983860e26fb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 203, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 264 + ], + "type": "text", + "content": "learn a joint sketch-photo manifold (for category and fine-grained) via a feature extractor [26, 29, 103] trained using triplet loss [112]. Recent adoption of foundation models for SBIR [77] shifts focus to robust deployment using the open-set generalisation of CLIP [71]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": "Towards this goal of robust deployment, our sketch attribution algorithms " + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "inline_equation", + "content": "(\\mathbb{A}_i^R" + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V)" + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": " can predict which strokes the network focuses on when retrieving a photo (Fig. 4). Apart from interpreting SBIR models, sketch attribution can also help detect potential failures at runtime (inference). First, we use the attribution scores " + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_i^R" + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V" + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": " to rearrange the strokes from highest to lowest. The attribution scores indicate the most salient to the least salient strokes that affect model prediction. Second, we calculate a correlation " + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "inline_equation", + "content": "(Corr)" + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": " of our predictor stroke order with the ground-truth temporal stroke order drawn by a user (humans draw the most salient regions first and least salient areas last [33, 79]). A high correlation indicates that humans and our model prioritise strokes similarly, whereas a low " + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "inline_equation", + "content": "Corr" + }, + { + "bbox": [ + 46, + 266, + 289, + 471 + ], + "type": "text", + "content": " denotes that the model and the user prioritise different strokes. We evaluate SBIR on a pre-trained SOTA model [77] using CLIP with prompt learning as a sketch and photo encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 475, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 538 + ], + "type": "text", + "content": "Datasets: We use TU-Berlin [33] (for category-level SBIR) and Sketchy [79] (for fine-grained SBIR). TU-Berlin contains 250 categories, with 80 free-hand sketches in each, and 204, 489 images [116]. Sketchy [79] has 75, 471 sketches over 125 categories having 100 images in each [108]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 542, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 288, + 604 + ], + "type": "text", + "content": "Evaluation Metrics: Following [77], we use mean average precision (mAP) and precision for top 200 retrieved samples " + }, + { + "bbox": [ + 46, + 542, + 288, + 604 + ], + "type": "inline_equation", + "content": "(\\mathbb{P}@\\mathbb{200})" + }, + { + "bbox": [ + 46, + 542, + 288, + 604 + ], + "type": "text", + "content": " for category-level SBIR. For fine-grained SBIR [23], we measure Acc.@q, i.e., the percentage of sketches whose true matched photo is in the top-q list." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "text", + "content": "Results: We divide the evaluation set into two sets: (i) those that have a high correlation " + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "inline_equation", + "content": "Corr \\geq 0.5" + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "text", + "content": ", and (ii) those with a low correlation " + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "inline_equation", + "content": "Corr \\leq 0.1" + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "text", + "content": " of ground-truth and predicted stroke order. Tab. 1 shows sketches with a high " + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "inline_equation", + "content": "Corr \\geq 0.5" + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "text", + "content": " are 1.7/3.9 more accurate in Acc.@1/Acc.@5 than those with " + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "inline_equation", + "content": "Corr \\leq 0.1" + }, + { + "bbox": [ + 46, + 608, + 289, + 717 + ], + "type": "text", + "content": " for fine-grained SBIR, and 3.3/1.7 better in mAP/P@200 for category-level SBIR. Full Dataset indicates the performance of the pre-trained model on the entire evaluation set." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 73, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 209 + ], + "type": "image", + "image_path": "81299084bc63347c20a46cf747ab3e28cf0e2ecb551e4b192eab4f18863aff04.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "text", + "content": "Figure 4. Sketch attributions from stroke-level and point-level for image retrieval. High correlation of human-drawn stroke order with that from sketch-attributions (high " + }, + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "text", + "content": " low) indicate our sketch encoder gives more importance to salient strokes drawn early on." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 270, + 534, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 534, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 534, + 283 + ], + "type": "text", + "content": "5.2. Assisted Drawing via Noisy Stroke Removal" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 284, + 545, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 379 + ], + "type": "text", + "content": "Although sketching has enabled many exciting applications [25, 55, 63, 118], the fear to sketch (i.e., \"I can't sketch\") has proven fatal for its widespread adoption. To solve this, prior works [15] used complex (and hard to train [12]) reinforcement learning [80] to predict the importance of each stroke in a sketch. Next, a stroke subset selector removes noisy (less important) strokes, leaving only those positively contributing to the downstream tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "type": "text", + "content": "In this section, we focus on assisted drawing - given a photo, we use attribution scores " + }, + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "type": "inline_equation", + "content": "(\\mathbb{A}_i^R" + }, + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V)" + }, + { + "bbox": [ + 304, + 380, + 545, + 427 + ], + "type": "text", + "content": " to help humans draw a faithful and clean sketch. Our method is significantly simpler than reinforcement learning alternatives [15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": "For SLA, an input sketch " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{X} = \\sum_{k=1}^{m} \\omega_k S_k" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " strokes. We calculate the cosine similarity (sim) of input sketch " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " with its target photo " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{P} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " to measure \"how faithfully " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " describes " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{P}" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": "\". Next, we backpropagate gradients from the cosine similarity to strokes " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " and calculate stroke-level attribution score " + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_i^R" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 508, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 508, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 332, + 508, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial \\operatorname {s i m} (F _ {\\theta} (\\mathrm {X}) , F _ {\\theta} (\\mathrm {P}))}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}} \\tag {7}", + "image_path": "9cfe4737c7a1ac98c314c8c93db075fd7634af1200678dd711a095296a49172a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": "The pre-trained sketch and photo encoder " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "F_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " must be highly accurate to judge sketch-photo correspondence. Hence, we use pre-trained CLIP+prompts encoder from [77]. To remove noisy strokes, we only update the weights " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\omega_{i}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " using a normalised attribution score " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_i^R" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 613, + 545, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 613, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 322, + 613, + 545, + 642 + ], + "type": "interline_equation", + "content": "\\omega_ {i} ^ {*} = \\omega_ {i} \\cdot \\text {G u m b e l . S o f t m a x} \\left(\\frac {\\mathbb {A} _ {i} ^ {R}}{\\sum_ {\\forall i} \\mathbb {A} _ {i} ^ {R}} + \\Delta\\right) \\tag {8}", + "image_path": "90080ccccc00f083a584b8df4b909845e3989d27741c5c6eb7d9c062105749b2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 649, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 545, + 686 + ], + "type": "text", + "content": "Gumbel_Softmax(·) makes the output one-hot (discrete value) in the forward pass but differentiable with a probability distribution that sum to 1 in the backward pass [47]. The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2The sketch and photo encoder " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " could be a siamese-style shared network or two independent models with different network weights [77]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11001" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 265, + 150 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 265, + 150 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 265, + 150 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 265, + 150 + ], + "type": "image", + "image_path": "02207721c41a183f11951d641c93908624ec77b6048c9f07c99dc64e6296e5a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 158, + 287, + 192 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 192 + ], + "type": "text", + "content": "Figure 5. Assisted drawing via sketch healing (or filtering noisy strokes) using stroke attributions from SLA and P-SLA. This helps users having fear-to-sketch (\"I can't sketch\")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": "modified sketch is constructed as " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{X} = \\omega_1^* S_1 + \\dots + \\omega_m^* S_m" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": ". Intuitively, we keep strokes that contribute " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "(\\mathbb{A}_i^R)" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": " to a high cosine similarity matching human sketch " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": " and target photo " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{P}" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": " and remove " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": " with normalised " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_i^R" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": " lower than " + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "inline_equation", + "content": "(0.5 - \\Delta)" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "text", + "content": "Similar to SLA, we can also use P-SLA to compute the attribution " + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V" + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "text", + "content": " for each point " + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "text", + "content": " in the vector sketch " + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{V}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 243, + 287, + 291 + ], + "type": "text", + "content": ", by measuring the cosine similarity between input X and target photo P as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial \\operatorname {s i m} \\left(F _ {\\theta} (\\mathrm {X}) , F _ {\\theta} (\\mathrm {I})\\right)}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {9}", + "image_path": "426b184a816a3b6ebf581d0181902c2bcc0100a9952dbd739340fef84c057ab9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "content": "We remove noisy points " + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "content": " by updating the pen-states in Eq. (5) from pen-down " + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "inline_equation", + "content": "(1,0,0)" + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "content": " to pen-up " + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "inline_equation", + "content": "(0,1,0)" + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "content": " depending on its attribution " + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V" + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "content": " for point " + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "inline_equation", + "content": "v_{t}\\in \\mathrm{V}" + }, + { + "bbox": [ + 46, + 331, + 287, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 372, + 287, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 372, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 63, + 372, + 287, + 401 + ], + "type": "interline_equation", + "content": "q _ {t - 1} ^ {1 *} = q _ {t - 1} ^ {1} \\cdot \\text {G u m b e l . S o f t m a x} \\left(\\frac {\\mathbb {A} _ {t - 1} ^ {V}}{\\sum_ {\\forall t} \\mathbb {A} _ {t} ^ {V}}\\right) \\tag {10}", + "image_path": "bd9ac9bac3813c91407ad8da4e52277bb2d7d7f4222c091d291cb0ea97ecc8cf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": "Using updated values for " + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "inline_equation", + "content": "q_{t-1}^{1*}" + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": ", we update " + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "inline_equation", + "content": "q_{t-1}^{2*} = 1 - q_{t-1}^{1*}" + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": " and recalculate pixel intensities for raster sketch " + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{X}(p_x, p_y)" + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": ". The value of hyper-parameter " + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": " significantly affects the stroke removal process. We found the optimal " + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": " for SLA and P-SLA is 0.3 and 0.1, respectively. A higher " + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 405, + 287, + 502 + ], + "type": "text", + "content": " for P-SLA gives broken lines with a drop in the visual quality of an input sketch. Next, we evaluate stroke filtering using SLA and P-SLA on popular human-drawn sketch datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 504, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 577 + ], + "type": "text", + "content": "Dataset: For a fair comparison with prior works [15], we evaluate on fine-grained SBIR datasets QMUL-ShoeV2 and QMUL-Chair-V2 [12, 67, 112]. It consists of 6,730/1,800 sketches and 2,000/400 photos from ShoeV2/Chair-V2. We evaluate on the standard test-split of 679/525 sketches and 200/100 photos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 578, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 674 + ], + "type": "text", + "content": "Evaluation Metric: We measure the retrieval accuracy of the clean sketch with the target photo by computing Acc.@1 and Acc.@5. A high accuracy need not correspond to high visual quality to the human eye [84]. Hence, we conducted a small human study with 5 participants and reported the mean opinion score (MOS) [46]; each was asked to compare two sets of 50 sketch pairs (GT sketch vs SLA filtered) and (GT sketch vs P-SLA filtered)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Results: Removing noisy strokes from human-drawn sketches is still a new topic; hence, to the best of our knowledge, there is only one work by Bhunia et al. [15]. From" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 85, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 543, + 82 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 543, + 82 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 543, + 82 + ], + "type": "text", + "content": "Table 2. Noisy stroke removal using SLA and P-SLA attribution." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 85, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 85, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 85, + 545, + 167 + ], + "type": "table", + "html": "
MetricsGT SketchSLA filteredP-SLA filteredBhunia et al. [15]
Shoe-V2Acc.@133.436.136.543.7
Acc.@567.868.769.374.9
MOS28.685.757.1-
Chair-V2Acc.@153.354.956.564.8
Acc.@574.376.677.179.1
MOS35.871.457.1-
", + "image_path": "083de7ee413bf5f9a739223370e5b26b32487fa956a14bfa67a5273c7eb59d63.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": "Tab. 2, both Bhunia et al. [15], and ours improve fine-grained SBIR performance by " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": ", respectively. However, [15] outperforms our SLA and P-SLA filtered methods by " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": ", respectively. This performance gap is likely because [15] trains the baseline model [12] using actor-critic version of PPO [80], whereas our SLA and P-SLA work post-hoc [98] without training the baseline model [12]. Additionally, Bhunia et al. [15] aims to design a robust SBIR pipeline, whereas our SLA/P-SLA filtering aims to assist humans in drawing sketches. For human study (MOS): (i) users prefer SLA-filtered " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "21.5\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": " for GT sketch, (ii) however, P-SLA-filtered are preferred only " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "inline_equation", + "content": "42.9\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 358 + ], + "type": "text", + "content": " for GT sketch. This is verified by Fig. 5 where P-SLA-filtered sketches have broken strokes that degrade their visual quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 365, + 515, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 515, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 515, + 377 + ], + "type": "text", + "content": "5.3. Interactive Sketch To Photo Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "content": "The upsurge of large-scale image generation models (e.g., Stable-Diffusion [74], GigaGAN [50]) helped develop sketch-conditional image generation [63, 118]. However, a key limitation of conditional image generation is that these models do not always faithfully follow the input condition. This was resolved in two stages for text-to-image generation: (i) find word tokens with low influence on generated image, and (ii) iteratively update its activation until it reaches a minimum required value. In this section, we design a pipeline for faithful sketch-to-image generation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "content": "Using sketch attributions from SLA and P-SLA, we design a post-hoc [98] method that gives feedback to the user - which strokes the model focuses on and which strokes are being ignored. Given this feedback, a user can interact with the system to ensure the model attends all salient regions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": "Our interactive pipeline is built on top of a pre-trained sketch-to-photo generation model [55], comprising a modified ResNet-50 [41] as sketch encoder and StyleGAN [52] as image decoder. Given a raster sketch X, the modified sketch encoder computes a latent vector " + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+} = F_{\\theta}(\\mathrm{X})" + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+} \\in \\mathbb{R}^{14 \\times 512}" + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": ". Next, the StyleGAN [52] decoder generates the underlying image from " + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+}" + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": ". For a faithful sketch-to-image generation, we measure the \"influence\" of each stroke/vector point on the latent code " + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+}" + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": ". Particularly, prior works [106] suggest that " + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+}" + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": " is disentangled into 14-level semantic feature hierarchy, where " + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+1} \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 563, + 545, + 713 + ], + "type": "text", + "content": " has coarse-level features controlling major semantic struc" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11002" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 170 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 170 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 170 + ], + "type": "image", + "image_path": "daefada324147b350a086fe96ce19923fb2dcd72d509a6bd86e73f1b71ae4ea5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "text", + "content": "Figure 6. Interactive Sketch to Photo Generation: Our stroke attribution algorithms make existing sketch-to-photo generation pipelines [55] more faithful. We achieve this by computing the stroke-level " + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_i^S" + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "text", + "content": " or coordinate-level " + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V" + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "text", + "content": " attribution that has a maximal influence on the latent code " + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+}" + }, + { + "bbox": [ + 46, + 178, + 288, + 236 + ], + "type": "text", + "content": " used by the image decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "text", + "content": "tures and " + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "inline_equation", + "content": "z_{s2p}^{+14}\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "text", + "content": " has fine-level features controlling colour schemes, etc. Since sketches primarily convey semantic structure, we use the sum of the first 7-layers to compute stroke attribution " + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_i^R" + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_t^V" + }, + { + "bbox": [ + 46, + 243, + 288, + 293 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 300, + 220, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 300, + 220, + 329 + ], + "spans": [ + { + "bbox": [ + 69, + 300, + 220, + 329 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {i} ^ {R} = \\frac {\\partial \\sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\\partial \\mathrm {X}} \\cdot \\frac {\\partial \\sum_ {k = 1} ^ {m} \\omega_ {k} \\mathcal {S} _ {k}}{\\partial \\mathcal {S} _ {i}}", + "image_path": "49ee835966e373f0b5d3a3dea3b9a1de5eba795fa5f425714299b15527512c55.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 327, + 287, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 327, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 69, + 327, + 287, + 366 + ], + "type": "interline_equation", + "content": "\\mathbb {A} _ {t} ^ {V} = \\frac {\\partial \\sum_ {k = 1} ^ {7} z _ {s 2 p} ^ {+ k}}{\\partial \\mathrm {X}} \\cdot \\sum_ {\\forall p _ {x}, p _ {y}} \\left\\{\\frac {\\partial \\mathrm {X} \\left(p _ {x} , p _ {y}\\right)}{\\partial v _ {t}} \\right\\} \\tag {11}", + "image_path": "6c70e2b96c5f5deeaa3a4861342375523457d7e90baca3fef64ca1f94f10263a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 374, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 399 + ], + "type": "text", + "content": "Next, we qualitatively evaluate our iterative sketch to photo generation pipeline, as shown in Fig. 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 407, + 260, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 260, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 260, + 419 + ], + "type": "text", + "content": "5.4. Adversarial Attacks on Human Sketches" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "type": "text", + "content": "Szegedy et al. [94] discovered that predictions by deep networks can be manipulated with extremely low-magnitude input perturbations. For images, these can be restricted to be imperceptible to human vision, but their effect can completely change the output prediction by a deep network. Such adversarial attacks are possible in image classification [94], semantic segmentation [7, 42], object detection [97, 115], object tracking [22, 48], etc. Studying these quirks is crucial as it can pose a real threat to deep learning as a pragmatic technology [3]. While major work has been dedicated to attacks on images, the recent surge of deployable sketch applications [77, 118] motivates us to present the first study on adversarial attacks for human-sketches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "We show how our stroke attribution algorithms (SLA and P-SLA) provide the necessary information for adversarial attacks. For brevity, we focus on adversarial attacks on sketch classification [40, 111]. Intuitively, we use sketch attribution to remove the smallest stroke (in SLA Attack) and minimum number of points (in P-SLA Attack), yet have the maximum impact on changing prediction of a pre-trained classifier " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{\\theta}(\\mathrm{X}) = y_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". Ours is (i) a white-box [69, 95] setting – we have access to network weights and gradients of our ResNet-18 [41] classifier, pre-trained on QuickDraw [40] or TU-Berlin [33]; (ii) untargeted attack – while tar" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 99, + 545, + 145 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 3. Sketch Adversarial Attacks: Using stroke attributions, we remove a small stroke " + }, + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "(|S_i| \\leq \\epsilon)" + }, + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "type": "text", + "content": " that misclassifies an input sketch." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 99, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 307, + 99, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 99, + 545, + 145 + ], + "type": "table", + "html": "
No AttackSLA AttackP-SLA Attack
ε = 5ε = 15ε = 5ε = 15
QuickDraw67.265.764.565.163.7
TU-Berlin74.971.568.570.268.1
", + "image_path": "ca9e5a683715cb8de978f44d36fdeefde87dd0180fea9337af13f84881dd5dc7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": "geted attacks [60, 121] misclassify " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "F_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{cls}}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " to a specific target class " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{cls}}^{*}" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": ", untargeted attacks [93] aim to misclassify to any arbitrary class " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{cls}}^{\\mathrm{GT}} \\neq y_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": ". For a neater description of SLA and P-SLA attacks, we define the rasterisation process using " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\cdot)" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{X} = \\omega_1\\mathcal{S}_1 + \\dots + \\omega_m\\mathcal{S}_m = \\mathcal{R}(\\mathcal{S})" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " for SLA and following Eq. (5) for P-SLA we define " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{X} = \\mathcal{R}(\\{v_1,\\ldots v_T\\}) = \\mathcal{R}(\\mathrm{V})" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": ". Next, we find a stroke " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "S_{adv}" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " with stroke length " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "|S_i|" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " less than some threshold " + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 153, + 547, + 251 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 258, + 545, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 258, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 258, + 545, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {a d v} = \\arg \\max _ {| \\mathcal {S} _ {j} | \\leq \\epsilon} \\mathcal {L} _ {\\mathrm {c l s}} \\left(F _ {\\theta} \\left(\\mathcal {R} \\left(\\mathcal {S} - \\left\\{\\mathcal {S} _ {j} \\right\\}\\right)\\right), y _ {\\mathrm {c l s}} ^ {\\mathrm {G T}}\\right) \\tag {12}", + "image_path": "64442a4249789f18a47f672f9984b4d2094d0ac45856f43b4957b51b6d300c16.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": "Unlike typical adversarial attacks on images that add a small noise " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "(\\mathrm{X} + \\Delta x)" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "||\\Delta x||_{\\infty}\\leq \\epsilon" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " , for sketch adversarial attacks, we remove a small stroke " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "(\\mathrm{X} - S_{adv})" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " such that the stroke length is less than " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "|\\mathcal{S}_{adv}|\\leq \\epsilon" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " . For PSLA attack, we find a subset of " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " vector points " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{V}_{adv} =" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\{v_1^{adv},\\dots v_\\epsilon^{adv}\\}" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " from input sketch " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{V}\\in \\mathbb{R}^{T\\times 5}" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " which maximises the categorical cross-entropy loss " + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 305, + 288, + 546, + 373 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 380, + 545, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 380, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 341, + 380, + 545, + 412 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} v _ {t} ^ {a d v} = \\mathcal {L} _ {\\mathrm {c l s}} \\left(F _ {\\theta} \\left(\\mathcal {R} (\\mathrm {V} - \\{v _ {t} \\})\\right), y _ {\\mathrm {c l s}} ^ {\\mathrm {G T}}\\right) \\tag {13} \\\\ \\mathrm {V} _ {a d v} = \\operatorname {t o p} @ k \\left(\\left\\{v _ {1} ^ {a d v}, v _ {2} ^ {a d v}, \\dots v _ {T} ^ {a d v} \\right\\}, \\epsilon\\right) \\\\ \\end{array}", + "image_path": "07a593c9c499d94ea5b65f2c9f9effcc7681d1273080c7e7f1e125e5babd8705.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\mathrm{top}@\\mathrm{k}(\\cdot ,\\epsilon)" + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "text", + "content": " picks the highest " + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "text", + "content": " elements. Fig. 7 shows the adversarial strokes " + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "inline_equation", + "content": "S_{adv}" + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "text", + "content": " and points " + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\mathrm{V}_{adv}" + }, + { + "bbox": [ + 305, + 421, + 545, + 446 + ], + "type": "text", + "content": " in red." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "content": "Dataset: We evaluate sketch adversarial attacks on QuickDraw [40] and TU-Berlin [33]. We use a subset [103] of " + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "inline_equation", + "content": "50M" + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "content": " sketches in QuickDraw [40] as " + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "inline_equation", + "content": "3.8M" + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "content": " samples across 345 categories split in " + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "inline_equation", + "content": "2.1M" + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "content": " sketches for training, " + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "inline_equation", + "content": "0.3M" + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "content": " for validation and " + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "inline_equation", + "content": "0.4M" + }, + { + "bbox": [ + 304, + 448, + 546, + 520 + ], + "type": "text", + "content": " for evaluation. See Sec. 5.1 for details on TU-Berlin [33] dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": "Evaluation: We measure the drop in classification accuracy when using SLA and P-SLA attacks in Tab. 3. Unlike image-based adversarial attacks where " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": " is a pixel intensity (non-integer, decimal value), our sketch attacks occur on stroke/point length, making " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": " an integer. A higher " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon \\geq 20" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": " removes \"visible\" strokes in SLA and broken lines in P-SLA (Fig. 7). Hence, we evaluate accuracy drop for " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon = 5" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon = 15" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": " in Tab. 3. We observe for " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon = 5" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": " P-SLA offers a better adversarial attack than SLA by a margin of " + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 522, + 547, + 631 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 639, + 392, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 392, + 652 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 392, + 652 + ], + "type": "text", + "content": "6. Human Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Interpretability aims to help humans understand a model's reasoning process (transparency), verify that its predictions are based on the right constraints (fairness), and evaluate its confidence (trustworthiness) [73, 81]. In Sec. 5, we evaluated interpretability using several automatic metrics (e.g.," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11003" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 172, + 114 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 172, + 114 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 172, + 114 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 172, + 114 + ], + "type": "image", + "image_path": "9cf77f56775305bf24959d400911878cad02868649a9647f3052cc029a52b0a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 121, + 520, + 132 + ], + "lines": [ + { + "bbox": [ + 72, + 121, + 520, + 132 + ], + "spans": [ + { + "bbox": [ + 72, + 121, + 520, + 132 + ], + "type": "text", + "content": "Figure 7. Adversarial attacks on human drawn sketches using SLA and P-SLA. The adversarial strokes are marked in RED." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 221, + 109 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 221, + 109 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 221, + 109 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 221, + 109 + ], + "type": "image", + "image_path": "91fd081462594502b2646ec0a3efcf609c8e57ca1bb300821047060e05535b6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 294, + 108 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 294, + 108 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 294, + 108 + ], + "type": "image", + "image_path": "6b3a6af5e965fd9930020cf7cf5177a99aec8da7511cfef5f57fad78829c2440.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 108, + 252, + 114 + ], + "lines": [ + { + "bbox": [ + 230, + 108, + 252, + 114 + ], + "spans": [ + { + "bbox": [ + 230, + 108, + 252, + 114 + ], + "type": "text", + "content": "GT:Angel" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 71, + 345, + 108 + ], + "blocks": [ + { + "bbox": [ + 298, + 71, + 345, + 108 + ], + "lines": [ + { + "bbox": [ + 298, + 71, + 345, + 108 + ], + "spans": [ + { + "bbox": [ + 298, + 71, + 345, + 108 + ], + "type": "image", + "image_path": "bcb3f9f886c5b283fd2482c167a0bd8f5971b284b48e2bd385c79f6ee096c679.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 345, + 70, + 419, + 108 + ], + "blocks": [ + { + "bbox": [ + 345, + 70, + 419, + 108 + ], + "lines": [ + { + "bbox": [ + 345, + 70, + 419, + 108 + ], + "spans": [ + { + "bbox": [ + 345, + 70, + 419, + 108 + ], + "type": "image", + "image_path": "47c221124c8a01a26abe3d82aace0fd4d53dc718b1d092c4a15cbe27c38e2690.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 108, + 377, + 114 + ], + "lines": [ + { + "bbox": [ + 354, + 108, + 377, + 114 + ], + "spans": [ + { + "bbox": [ + 354, + 108, + 377, + 114 + ], + "type": "text", + "content": "6T: Cactus" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 108, + 411, + 114 + ], + "lines": [ + { + "bbox": [ + 391, + 108, + 411, + 114 + ], + "spans": [ + { + "bbox": [ + 391, + 108, + 411, + 114 + ], + "type": "text", + "content": "Butterfly" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 423, + 71, + 470, + 108 + ], + "blocks": [ + { + "bbox": [ + 423, + 71, + 470, + 108 + ], + "lines": [ + { + "bbox": [ + 423, + 71, + 470, + 108 + ], + "spans": [ + { + "bbox": [ + 423, + 71, + 470, + 108 + ], + "type": "image", + "image_path": "d4c1c319ed213e7a6af9aad8bb8a8c62ea4b039aabb19c9509095c7c4e05ee95.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 472, + 70, + 544, + 108 + ], + "blocks": [ + { + "bbox": [ + 472, + 70, + 544, + 108 + ], + "lines": [ + { + "bbox": [ + 472, + 70, + 544, + 108 + ], + "spans": [ + { + "bbox": [ + 472, + 70, + 544, + 108 + ], + "type": "image", + "image_path": "ee1534eb5b4d886c4951623f1fbf12d1c5447f3988cac2437134d236471854d7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 108, + 506, + 114 + ], + "lines": [ + { + "bbox": [ + 476, + 108, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 476, + 108, + 506, + 114 + ], + "type": "text", + "content": "GT: Mushroom" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 136, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 136, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 136, + 287, + 232 + ], + "type": "text", + "content": "classification or retrieval accuracy) on different evaluation datasets [33, 40, 79]. However, highlighting salient regions by backpropagating gradients for downstream applications does not capture how helpful end-users find these attributions [53, 73]. In this section, we take a human-centred approach to interpretability – how well our stroke attribution algorithms align with the reasoning process of humans and the trade-off, interpretability vs. accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 235, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 331 + ], + "type": "text", + "content": "Setup: We recruited 7 participants from different geographical regions, in the age group 20 - 30 years. All participants had some background in AI research, but only 3 reported having prior experience in interpretability. Once recruited, each user is assigned a unique ID for anonymity. For SLA and P-SLA, we conduct 5 human studies, with each having 10 multiple choice questions (MCQs). Hence, each participant answers 50 MCQs for SLA and 50 for P-SLA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "type": "text", + "content": "Evaluating Transparency: For an attribution to be useful, humans must understand a model's behaviour for correct and incorrect predictions. In this section, we evaluate if our SLA and P-SLA can make existing sketch classifiers (i.e., Sketch-A-Net [111]), pre-trained on TU-Berlin [33] dataset transparent to humans. Accordingly, we choose a random category and select 4 sketch instances - (i) 3 misclassified and 1 correctly classified, and (ii) 1 misclassified and 3 correctly classified. Next, as shown in Fig. 8, we compute stroke attribution (SLA and P-SLA) for the selected sketches and ask users: \"Only 1 of these 4 sketches are correctly (or incorrectly) recognised by our model. Please select that correct (or incorrect) sketch.\" We find users can identify correct/incorrect model predictions " + }, + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "type": "inline_equation", + "content": "75.9\\% /63.4\\%" + }, + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "type": "text", + "content": " of times for SLA and " + }, + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "type": "inline_equation", + "content": "76.3\\% /65.2\\%" + }, + { + "bbox": [ + 46, + 333, + 287, + 513 + ], + "type": "text", + "content": " for P-SLA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 52, + 514, + 259, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 514, + 259, + 525 + ], + "spans": [ + { + "bbox": [ + 52, + 514, + 259, + 525 + ], + "type": "text", + "content": "Q. Only one sketch is correctly categorised. Which one?" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 52, + 526, + 96, + 559 + ], + "blocks": [ + { + "bbox": [ + 52, + 526, + 96, + 559 + ], + "lines": [ + { + "bbox": [ + 52, + 526, + 96, + 559 + ], + "spans": [ + { + "bbox": [ + 52, + 526, + 96, + 559 + ], + "type": "image", + "image_path": "df06294eb6bd9999b6c75663538cd46f9a2767d5865735c1a6d904e2bc219717.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 113, + 526, + 155, + 559 + ], + "blocks": [ + { + "bbox": [ + 113, + 526, + 155, + 559 + ], + "lines": [ + { + "bbox": [ + 113, + 526, + 155, + 559 + ], + "spans": [ + { + "bbox": [ + 113, + 526, + 155, + 559 + ], + "type": "image", + "image_path": "cbacf80aa0a4c755660163bd6308eb1731cac93aebbfd695849925c3e49662de.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 169, + 526, + 211, + 559 + ], + "blocks": [ + { + "bbox": [ + 169, + 526, + 211, + 559 + ], + "lines": [ + { + "bbox": [ + 169, + 526, + 211, + 559 + ], + "spans": [ + { + "bbox": [ + 169, + 526, + 211, + 559 + ], + "type": "image", + "image_path": "c35654e0be2f2813f7c613938e91df8f516a0281ae5ee1f3987bc360ce3630d7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 228, + 526, + 283, + 559 + ], + "blocks": [ + { + "bbox": [ + 228, + 526, + 283, + 559 + ], + "lines": [ + { + "bbox": [ + 228, + 526, + 283, + 559 + ], + "spans": [ + { + "bbox": [ + 228, + 526, + 283, + 559 + ], + "type": "image", + "image_path": "ec686d1bfb4bf068059b84efd76c23efcbd38ac9b2e55701a0ee08c2cf862f8c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 49, + 561, + 261, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 561, + 261, + 571 + ], + "spans": [ + { + "bbox": [ + 49, + 561, + 261, + 571 + ], + "type": "text", + "content": "Q. Only one sketch is incorrectly categorised. Which one?" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 52, + 572, + 96, + 607 + ], + "blocks": [ + { + "bbox": [ + 52, + 572, + 96, + 607 + ], + "lines": [ + { + "bbox": [ + 52, + 572, + 96, + 607 + ], + "spans": [ + { + "bbox": [ + 52, + 572, + 96, + 607 + ], + "type": "image", + "image_path": "fe58f5eb1836c79254ebe16a0683accc90994209e8b1db65b7753307bafd9dc1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 613, + 287, + 635 + ], + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 635 + ], + "type": "text", + "content": "Figure 8. Evaluating transparency: Can a human understand the behaviour of an existing (pre-trained) classifier with SLA/P-SLA." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 112, + 572, + 156, + 606 + ], + "blocks": [ + { + "bbox": [ + 112, + 572, + 156, + 606 + ], + "lines": [ + { + "bbox": [ + 112, + 572, + 156, + 606 + ], + "spans": [ + { + "bbox": [ + 112, + 572, + 156, + 606 + ], + "type": "image", + "image_path": "802e104f3e8afafa29b293696fb752002a58b12a1015deba569f2009ea6e07e3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 169, + 573, + 212, + 606 + ], + "blocks": [ + { + "bbox": [ + 169, + 573, + 212, + 606 + ], + "lines": [ + { + "bbox": [ + 169, + 573, + 212, + 606 + ], + "spans": [ + { + "bbox": [ + 169, + 573, + 212, + 606 + ], + "type": "image", + "image_path": "a85669fd7d1de5d17bb2eee4edd26d9c947bf5f2d194544bbe90e541145d5f38.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 228, + 570, + 283, + 606 + ], + "blocks": [ + { + "bbox": [ + 228, + 570, + 283, + 606 + ], + "lines": [ + { + "bbox": [ + 228, + 570, + 283, + 606 + ], + "spans": [ + { + "bbox": [ + 228, + 570, + 283, + 606 + ], + "type": "image", + "image_path": "fb91e26f7fa6a93adf2487953ac86215e65bf52a3716626b862fb3b5209ab1b1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": "Evaluating Fairness: End users are much better positioned to make a decision with help from a model if intelligible explanations are provided. In this section, we evaluate if SLA and P-SLA can help end users understand \"What went wrong?\" (i) For a pre-trained sketch classifier [111], we show users a misclassified sketch instance and ask users" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "text", + "content": "to identify the (wrongly) predicted category. (ii) For fine-grained SBIR [13], we show a sketch (whose GT photo is not in top-10) and ask users to identify the top-1 (wrongly) retrieved photo in Fig. 9. Humans can identify the misclassified category " + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "inline_equation", + "content": "(66.3\\%)" + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "text", + "content": " and the incorrectly retrieved photo " + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "inline_equation", + "content": "39.1\\%" + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "inline_equation", + "content": "(37.2\\%)" + }, + { + "bbox": [ + 304, + 136, + 545, + 208 + ], + "type": "text", + "content": " for SLA (P-SLA)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 310, + 213, + 414, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 213, + 414, + 233 + ], + "spans": [ + { + "bbox": [ + 310, + 213, + 414, + 233 + ], + "type": "text", + "content": "Q. Prediction is incorrect, identify predicted category" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 420, + 213, + 539, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 213, + 539, + 232 + ], + "spans": [ + { + "bbox": [ + 420, + 213, + 539, + 232 + ], + "type": "text", + "content": "Q. Retrieval is incorrect, identify the top-1 retrieved image" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 310, + 236, + 363, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 236, + 363, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 236, + 363, + 259 + ], + "type": "text", + "content": "② " + }, + { + "bbox": [ + 310, + 236, + 363, + 259 + ], + "type": "inline_equation", + "content": "{Na} + {O2} < \\Delta" + }, + { + "bbox": [ + 310, + 236, + 363, + 259 + ], + "type": "text", + "content": " 释放" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 365, + 236, + 414, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 236, + 414, + 259 + ], + "spans": [ + { + "bbox": [ + 365, + 236, + 414, + 259 + ], + "type": "text", + "content": "(A) Angel (B)Butterfly" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 365, + 261, + 392, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 261, + 392, + 284 + ], + "spans": [ + { + "bbox": [ + 365, + 261, + 392, + 284 + ], + "type": "text", + "content": "(C) Fan (D) Dog" + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 420, + 236, + 459, + 283 + ], + "blocks": [ + { + "bbox": [ + 420, + 236, + 459, + 283 + ], + "lines": [ + { + "bbox": [ + 420, + 236, + 459, + 283 + ], + "spans": [ + { + "bbox": [ + 420, + 236, + 459, + 283 + ], + "type": "image", + "image_path": "d483d95960b9a87048f9e77371934cc57ec77a22ff3daf9b7a402b97350186cc.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 293, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 315 + ], + "type": "text", + "content": "Figure 9. Evaluating Fairness: For an incorrect model prediction, we evaluate if humans can \"identify what went wrong\"." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 466, + 236, + 490, + 283 + ], + "blocks": [ + { + "bbox": [ + 466, + 236, + 490, + 283 + ], + "lines": [ + { + "bbox": [ + 466, + 236, + 490, + 283 + ], + "spans": [ + { + "bbox": [ + 466, + 236, + 490, + 283 + ], + "type": "image", + "image_path": "9bc0a0b8552958f5393fb9398705b09175e1b100dfef0c43c021632aacbcee68.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 498, + 236, + 523, + 283 + ], + "blocks": [ + { + "bbox": [ + 498, + 236, + 523, + 283 + ], + "lines": [ + { + "bbox": [ + 498, + 236, + 523, + 283 + ], + "spans": [ + { + "bbox": [ + 498, + 236, + 523, + 283 + ], + "type": "image", + "image_path": "dc0d47a686c86eeb1a7d3a5cded5b872ce73c13fa5bf25f3a1c7d87b47a5ed7f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 531, + 247, + 540, + 274 + ], + "blocks": [ + { + "bbox": [ + 531, + 247, + 540, + 274 + ], + "lines": [ + { + "bbox": [ + 531, + 247, + 540, + 274 + ], + "spans": [ + { + "bbox": [ + 531, + 247, + 540, + 274 + ], + "type": "image", + "image_path": "b7d09586483de2d7165b23f739bc44328102e30512b948786bc7fcc9b229f6da.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "content": "**Evaluating Trustworthiness:** Determining trust in individual predictions is important when used for decision-making (e.g., medical diagnosis [57]). We train two copies of the same sketch classifier [111], a strong classifier with " + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "content": " accuracy on TU-Berlin [33] and a weak one reaching " + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "content": ". We present the stroke attribution (SLA/P-SLA) for both models and ask users to identify the strong/weak classifier, as shown in Fig. 10. Humans identify the stronger classifier " + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "inline_equation", + "content": "71.4\\% (68.7\\%)" + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "content": " of the time for SLA (P-SLA)." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 311, + 439, + 528, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 439, + 528, + 451 + ], + "spans": [ + { + "bbox": [ + 311, + 439, + 528, + 451 + ], + "type": "text", + "content": "Q. Prediction is correct, identify the stronger classifier" + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 311, + 453, + 361, + 495 + ], + "blocks": [ + { + "bbox": [ + 311, + 453, + 361, + 495 + ], + "lines": [ + { + "bbox": [ + 311, + 453, + 361, + 495 + ], + "spans": [ + { + "bbox": [ + 311, + 453, + 361, + 495 + ], + "type": "image", + "image_path": "c15ed2643b4751d5a0dd526bdf3f0e1147726e2c69b75b8ecfd8c9ac47adb6e6.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 503, + 545, + 536 + ], + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 536 + ], + "type": "text", + "content": "Figure 10. Evaluating trustworthiness: We present the stroke attributions (SLA/P-SLA) from two sketch classifiers (strong and weak), and ask users to identify the strong/weak model." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 362, + 454, + 411, + 495 + ], + "blocks": [ + { + "bbox": [ + 362, + 454, + 411, + 495 + ], + "lines": [ + { + "bbox": [ + 362, + 454, + 411, + 495 + ], + "spans": [ + { + "bbox": [ + 362, + 454, + 411, + 495 + ], + "type": "image", + "image_path": "10a97bcb851270199e12f9c43445718e8a17d93e3d7fc8886ca01ab5f1edd306.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 413, + 454, + 422, + 495 + ], + "blocks": [ + { + "bbox": [ + 413, + 454, + 422, + 495 + ], + "lines": [ + { + "bbox": [ + 413, + 454, + 422, + 495 + ], + "spans": [ + { + "bbox": [ + 413, + 454, + 422, + 495 + ], + "type": "image", + "image_path": "79855170b30d002b8931b83b892766649f1974b88a087aa48c0160eb164a5c55.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 429, + 454, + 479, + 495 + ], + "blocks": [ + { + "bbox": [ + 429, + 454, + 479, + 495 + ], + "lines": [ + { + "bbox": [ + 429, + 454, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 429, + 454, + 479, + 495 + ], + "type": "image", + "image_path": "5e3f1a99612e002d3faea49e09e9047e06263dbd7eb2e248bedd42e1e1ee1e59.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 482, + 454, + 541, + 495 + ], + "blocks": [ + { + "bbox": [ + 482, + 454, + 541, + 495 + ], + "lines": [ + { + "bbox": [ + 482, + 454, + 541, + 495 + ], + "spans": [ + { + "bbox": [ + 482, + 454, + 541, + 495 + ], + "type": "image", + "image_path": "c451294a4071bf72682eb5d01b9d66cfa6ab1dc4bc7819ca09c349fecd010f69.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 548, + 378, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 378, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 378, + 560 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "This work emphasises the pivotal role of strokes in human-drawn sketches, offering unique insights compared to pixel-based images. Our lightweight explainability solution seamlessly integrates with pre-trained models, addressing rasterisation challenges and contributing to diverse sketch-related tasks. Through applications in Retrieval, Generation, Assisted Drawing, and Sketch Adversarial Attack, our model showcases adaptability and significance. The proposed stroke-level attribution provides nuanced insights into model behaviour, underscoring the importance of explainability in bridging human expression with model predictions in the evolving field of sketch interpretation." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11004" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 58, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. arXiv preprint arXiv:2011.05429, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 159, + 288, + 203 + ], + "type": "text", + "content": "[3] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. arXiv preprint arXiv:2108.00401, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 57, + 205, + 288, + 258 + ], + "type": "text", + "content": "[4] Shm Garanganao Almeda, J.D. Zamfirescu-Pereira, Kyu Won Kim, Pradeep Mani Rathnam, and Bjoern Hartmann. Prompting for discovery: Flexible sense-making for ai art-making with dreamsheets. arXiv preprint arXiv:2310.09985, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 57, + 261, + 288, + 304 + ], + "type": "text", + "content": "[5] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 307, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 307, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 57, + 307, + 288, + 339 + ], + "type": "text", + "content": "[6] Marco Ancona, Enea Ceolini, Cengiz Öztireli, and Markus Gross. A unified view of gradient-based attribution methods for deep neural networks. In NIPSW, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 341, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 341, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 57, + 341, + 288, + 373 + ], + "type": "text", + "content": "[7] Anurag Arnab, Ondrej Miksik, and Philip H.S. Torr. On the robustness of semantic segmentation models to adversarial attacks. In CVPR, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 375, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 375, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 375, + 288, + 430 + ], + "type": "text", + "content": "[8] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 432, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 288, + 485 + ], + "type": "text", + "content": "[9] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "text", + "content": "[10] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 544, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 288, + 598 + ], + "type": "text", + "content": "[11] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 601, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 643 + ], + "type": "text", + "content": "[12] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 646, + 288, + 689 + ], + "type": "text", + "content": "[13] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 714 + ], + "type": "text", + "content": "[14] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "and Yi-Zhe Song. Doodle it yourself: Class incremental learning by drawing a few sketches. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 97, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 151 + ], + "type": "text", + "content": "[15] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 312, + 153, + 545, + 186 + ], + "type": "text", + "content": "[16] Moritz Bohle, Mario Fritz, and Bernt Schiele. B-cos: Networks: Alignment is all we need for interpretability. In CVPR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 188, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 188, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 312, + 188, + 545, + 230 + ], + "type": "text", + "content": "[17] Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well onImagenet. arXiv preprint arXiv:1904.00760, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 233, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 233, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 233, + 545, + 287 + ], + "type": "text", + "content": "[18] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In ICCV, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 289, + 545, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 289, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 312, + 289, + 545, + 333 + ], + "type": "text", + "content": "[19] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 335, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 335, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 335, + 545, + 369 + ], + "type": "text", + "content": "[20] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM TOG, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 370, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 370, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 312, + 370, + 545, + 402 + ], + "type": "text", + "content": "[21] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In ICCV, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 448 + ], + "type": "text", + "content": "[22] Xuesong Chen, Xiyu Yan, Feng Zheng, Yong Jiang, Shu-Tao Xia, Yong Zhao, and Rongrong Ji. One-shot adversarial attacks on visual tracking with dual attention. In CVPR, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 450, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 450, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 312, + 450, + 545, + 494 + ], + "type": "text", + "content": "[23] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 496, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 496, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 312, + 496, + 545, + 540 + ], + "type": "text", + "content": "[24] Pinaki Nath Chowdhury, Aneeshan Sain, Yulia Gryaditskaya, Ayan Kumar Bhunia, Tao Xiang, and Yi-Zhe Song. Fs-coco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 541, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 541, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 312, + 541, + 545, + 585 + ], + "type": "text", + "content": "[25] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation: Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 587, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 587, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 312, + 587, + 545, + 619 + ], + "type": "text", + "content": "[26] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 622, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 622, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 312, + 622, + 545, + 643 + ], + "type": "text", + "content": "[27] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 545, + 677 + ], + "type": "text", + "content": "[28] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11005" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[30] Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron? arXiv preprint arXiv:1805.12233, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 150 + ], + "type": "text", + "content": "[31] Gabriele Dominici, Pietro Barbiero, Lucie Charlotte Magister, Lio Pietro, and Nikola Simidjevski. Sharcs: Shared concept space for explainable multimodal learning. arXiv preprint arXiv:2307.00316, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 151, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 288, + 184 + ], + "type": "text", + "content": "[32] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 185, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 288, + 207 + ], + "type": "text", + "content": "[33] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 4, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 208, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 288, + 240 + ], + "type": "text", + "content": "[34] Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. Technical Report, Université de Montréal, 2009. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "type": "text", + "content": "[35] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. arXiv preprint arXiv:2112.04417, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 319 + ], + "type": "text", + "content": "[36] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In ICCV, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 319, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 352 + ], + "type": "text", + "content": "[37] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 288, + 387 + ], + "type": "text", + "content": "[38] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In CVPR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 388, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 420 + ], + "type": "text", + "content": "[39] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 421, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 288, + 443 + ], + "type": "text", + "content": "[40] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 3, 4, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 444, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 288, + 476 + ], + "type": "text", + "content": "[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 477, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 288, + 520 + ], + "type": "text", + "content": "[42] Yang He, Shadi Rahimian, Bernt Schiele, and Mario Fritz. Segmentations-leak: Membership inference attacks and defenses in semantic image segmentation. In ECCV, 2020. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 522, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 288, + 544 + ], + "type": "text", + "content": "[43] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "type": "text", + "content": "[44] Josh Holinaty, Alec Jacobson, and Fanny Chevalier. Supporting referenceimageryfordigitaldrawing. In ICCVW, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 578, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 288, + 621 + ], + "type": "text", + "content": "[45] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 624, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 288, + 666 + ], + "type": "text", + "content": "[46] Quan Huynh-Thu, Marie-Neige Garcia, Filippo Speranza, Philip Corriveau, and Alexander Raake. Study of ratling scales for subjective quality assessment of high definition video. IEEE TBC, 2010. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 668, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 668, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 668, + 288, + 690 + ], + "type": "text", + "content": "[47] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "text", + "content": "[48] Yunhan Jia, Yantao Lu, Junjie Shen, Qi Alfred Chen, Hao Chen, Zhenyu Zhong, and Tao Wei. Fooling detection alone" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "is not enough: Adversarial attack against multiple object tracking. In ICLR, 2020. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 545, + 140 + ], + "type": "text", + "content": "[49] Zeyang Jia, Eli Ben-Michael, and Kosuke Imai. Bayesian safe policy learning with chance constrained optimization: Application to military security assessment during the vietnam war. arXiv preprint arXiv:2307.08840, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 545, + 174 + ], + "type": "text", + "content": "[50] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "type": "text", + "content": "[51] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In ICCV, 2019. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 545, + 240 + ], + "type": "text", + "content": "[52] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 241, + 545, + 285 + ], + "type": "text", + "content": "[53] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In ECCV, 2022. 2, 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 545, + 342 + ], + "type": "text", + "content": "[54] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un) reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 545, + 387 + ], + "type": "text", + "content": "[55] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1, 4, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 421 + ], + "type": "text", + "content": "[56] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 545, + 455 + ], + "type": "text", + "content": "[57] Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. Principles of explanatory debugging to personalize interactive machine learning. In IUI, 2015. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 545, + 499 + ], + "type": "text", + "content": "[58] Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Sam Gershman, and Finale Doshi-Velez. An evaluation of the human-interpretability of explanation. arXiv preprint arXiv:1902.00006, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 500, + 545, + 533 + ], + "type": "text", + "content": "[59] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In WACV, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 545, + 567 + ], + "type": "text", + "content": "[60] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 568, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 545, + 611 + ], + "type": "text", + "content": "[61] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, Yi-Zhe Song, and Yonggang Qi. Zero-shot everything sketch-based image retrieval, and in explainable style. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 656 + ], + "type": "text", + "content": "[62] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Defloconet: Deep image editing via flexible low-level controls. In CVPR, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 713 + ], + "type": "text", + "content": "[63] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 5, 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11006" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "type": "text", + "content": "[64] Meike Nauta, Ron Van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In CVPR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "text", + "content": "[65] Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. ICMLW, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 287, + 173 + ], + "type": "text", + "content": "[66] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 217 + ], + "type": "text", + "content": "[67] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 219, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 287, + 252 + ], + "type": "text", + "content": "[68] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 253, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 287, + 275 + ], + "type": "text", + "content": "[69] Buu Phan, Fahim Mannan, and Felix Heide. Adversarial imaging pipelines. In CVPR, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 308 + ], + "type": "text", + "content": "[70] Zhiyu Qu, Yulia Gryaditskaya1, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 309, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 287, + 374 + ], + "type": "text", + "content": "[71] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 398 + ], + "type": "text", + "content": "[72] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards better understanding attribution methods. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 399, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 431 + ], + "type": "text", + "content": "[73] Marco T. Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In KDD, 2016. 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 433, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 287, + 465 + ], + "type": "text", + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Omer. High-resolution image synthesis with latent diffusion. In CVPR, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 287, + 499 + ], + "type": "text", + "content": "[75] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 533 + ], + "type": "text", + "content": "[76] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 287, + 578 + ], + "type": "text", + "content": "[77] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 1, 3, 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 287, + 622 + ], + "type": "text", + "content": "[78] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE TNNLS, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 656 + ], + "type": "text", + "content": "[79] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 287, + 690 + ], + "type": "text", + "content": "[80] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Pronimal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "text", + "content": "[81] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "tra. Grad-cam: Visual explanations from deep networks via gradient-based localisation. In ICCV, 2017. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 140 + ], + "type": "text", + "content": "[82] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 142, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 142, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 142, + 545, + 175 + ], + "type": "text", + "content": "[83] Harshay Shah, Prateek Jain, and Praneeth Netrapalli. Do input gradients highlight discriminative features? In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 176, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 176, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 312, + 176, + 545, + 198 + ], + "type": "text", + "content": "[84] Hamid Rahim Sheikh and Alan C. Bovik. Image information and visual quality. IEEE TIP, 2006. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 201, + 545, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 201, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 312, + 201, + 545, + 222 + ], + "type": "text", + "content": "[85] Yuming Shen, Li Liu, Fumin Shen, and Ling Shao. Zeroshot sketch-image hashing. In ICCV, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 224, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 224, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 312, + 224, + 545, + 256 + ], + "type": "text", + "content": "[86] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In ICML, 2017. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 258, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 258, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 312, + 258, + 545, + 300 + ], + "type": "text", + "content": "[87] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In ICLRw, 2014. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 304, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 304, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 545, + 347 + ], + "type": "text", + "content": "[88] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiβ, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. arXiv preprint arXiv:2204.11642, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 545, + 392 + ], + "type": "text", + "content": "[89] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 395, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 395, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 395, + 545, + 437 + ], + "type": "text", + "content": "[90] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 440, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 440, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 312, + 440, + 545, + 472 + ], + "type": "text", + "content": "[91] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In ICLRw, 2015. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 475, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 475, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 545, + 506 + ], + "type": "text", + "content": "[92] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 509, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 545, + 553 + ], + "type": "text", + "content": "[93] Christian Szegedy, Wojciech Zaremba, Ilyua Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 545, + 598 + ], + "type": "text", + "content": "[94] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 600, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 545, + 633 + ], + "type": "text", + "content": "[95] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white- and black-box attacks. In NeurIPS, 2020. 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 667 + ], + "type": "text", + "content": "[96] Aditay Tripathi, Rajath R Dani, Anand Mishra, and Anirban Chakraborty. Sketch-guided object localization in natural images. In ECCV, 2020. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "text", + "content": "[97] James Tu, Mengye Ren, Siva Manivasagam, Ming Liang, Bin Yang, Richard Du, Frank Cheng, and Raquel Urtasun. Physically realizable adversarial examples for lidar object detection. In CVPR, 2020. 7" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11007" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "text", + "content": "[98] Hugues Turbé, Mina Bjelogrlic, Christian Lovis, and Gianmarco Mengaldo. Evaluation of post-hoc interpretability methods in time-series classification. Nature Machine Intelligence, 2023. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 138 + ], + "type": "text", + "content": "[99] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In CVPR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "text", + "content": "[100] Donglai Wei, Joseph Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, 2018. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[101] Dennis Wei, Rahul Nair, Amit Dhurandhar, Kush R. Varshney, Elizabeth M. Daly, and Moninder Singh. On the safety of interpretable machine learning: A maximum deviation approach. In NeurlPS, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[102] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: deep sketch-based hair image synthesis. ACM TOG, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[103] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 326 + ], + "type": "text", + "content": "[104] Rui Xu, Zongyan Han, Le Hui, Jianjun Qian, and Jin Xie. Domain disentangled generative adversarial network for zero-shot sketch-based 3d shape retrieval. AAAI, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "text", + "content": "[105] Guowei Yan, Zhili Chen, Jimei Yang, and Huamin Wang. Interactive liquid splash modeling by user sketches. ACM TOG, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[106] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "text", + "content": "[107] Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Suggala, David I Inouye, and Pradeep K Ravikumar. On the (in) fidelity and sensitivity of explanations. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[108] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A zero-shot framework for sketch based image retrieval. In ECCV, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "text", + "content": "[109] Ran Yi, Zipeng Ye, Ruoyu Fan, Yezhi Shu, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Animating portrait line drawings from a single face photo and a speech signal. In ACM SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[110] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[111] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 3, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[112] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 602, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 624 + ], + "type": "text", + "content": "[113] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "text", + "content": "[114] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "type": "text", + "content": "[115] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[116] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In CVPR, 2016. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 308 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[117] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. IJCV, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[118] Lvmin Zhang, Anyi Rao, and Maneesh Agarwala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "text", + "content": "[119] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[120] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[121] Hang Zhou, Dongdong Chen, Jing Liao, Weiming Zhang, Kejiang Chen, Xiaoyi Dong, Kunlin Liu, Gang Hua, and Nenghai Yu. Lg-gan: Label guided adversarial network for flexible targeted attack of point cloud-based deep networks. In CVPR, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[122] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchscene: Richly-annotated scene sketches. In ECCV, 2018. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11008" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_content_list.json b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ab75dafff62580ee8d71b976f524cb27dda43e6 --- /dev/null +++ b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_content_list.json @@ -0,0 +1,1517 @@ +[ + { + "type": "text", + "text": "What, when, and where? Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions", + "text_level": 1, + "bbox": [ + 130, + 130, + 839, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brian Chen1 Nina Shvetsova2,3 Andrew Rouditchenko4 Daniel Kondermann5 Samuel Thomas6,7 Shih-Fu Chang1 Rogerio Feris6,7 James Glass4 Hilde Kuehne2,3,7", + "bbox": [ + 101, + 202, + 843, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Columbia University, 2Goethe University Frankfurt, 3University of Bonn, 4MIT CSAIL", + "bbox": [ + 218, + 243, + 746, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Quality Match GmbH, $^{6}$ IBM Research AI, $^{7}$ MIT-IBM Watson AI Lab", + "bbox": [ + 271, + 261, + 696, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bc2754,sc250}@columbia.edu, shvetsov@uni-frankfurt.de,{roudi,glass}@mit.edu", + "bbox": [ + 156, + 280, + 813, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dk@quality-match.com, {sthomas,rsferis}@us.ibm.com, kuehne@cs.uni-bonn.de", + "bbox": [ + 163, + 297, + 795, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 347, + 313, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed, providing dense spatio-temporal grounding annotations in long, untrimmed, multi-action instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks, showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatio-temporal grounding.", + "bbox": [ + 75, + 378, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatio-temporal grounding (STG) describes the challenging task of locating events in space and time within video data based on text referential expressions. Methods in this field usually rely on a combination of spatio-temporal bounding box annotation, together with a human-generated caption, describing the visual content of the bounding box [23, 54], which limits their generalizability beyond the given training scenario. Compared to that, as a second line of work, multimodal self-supervised learning tries to leverage \"free\" data sources, such as video and automatic speech", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb12e59d13f4edd0b4b3b2abe2b2049246c4944badb05195208fa2e9e9672005.jpg", + "image_caption": [ + "\"To this we will add one carrot chopped into jong jullien ...\"" + ], + "image_footnote": [], + "bbox": [ + 521, + 369, + 607, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5611b2857d6b84158d8140794397cffedcd25228d13a515e01c18410f1b5974.jpg", + "image_caption": [ + "\"... you can use the forks to pull it a bit apart...\"" + ], + "image_footnote": [], + "bbox": [ + 609, + 369, + 692, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7af3094e5fd67b80fbc50718020f8b7cfe8285cc758d1da4c2a35d9cfaea622.jpg", + "image_caption": [ + "Training Setup: Unlabeled videos with narrated instructions" + ], + "image_footnote": [], + "bbox": [ + 694, + 368, + 769, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/626eb5a1b84be14f724053fa934bb1ec474cbccb20c20844779fa842710513ad.jpg", + "image_caption": [ + "①", + "...stir it a bit so it mixes well..." + ], + "image_footnote": [], + "bbox": [ + 769, + 368, + 870, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3d470eff27720f396457fd2b5344d18b21068a86bff41251a9448107517375a.jpg", + "image_caption": [ + "Evaluation Setup: Referential queries - \"Crack egg\", \"Mix egg\", etc." + ], + "image_footnote": [], + "bbox": [ + 517, + 465, + 880, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4139cd564a8441e6094cbd80fa13d3f0cc4bf4bb34ab953847a529fc4dd97591.jpg", + "image_caption": [ + "Task: Spatio-Temporal Grounding - Find the temporal boundary of a queried action in an untrimmed video and spatially localize the action.", + "Figure 1. Learning Spatio-temporal grounding in untrimmed videos. In training, we learn from unlabeled videos without human annotation. In evaluation, we perform spatio-temporal grounding using an action description such as \"crack egg\" as a query. The model needs to localize both the action's temporal boundary and spatial region in the long untrimmed video. We visualize the heatmap from the annotation points as well as derived bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 517, + 532, + 694, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9749226b99a0a6a93abf15403719bf859baa2a436dd7922fac2da728763b02a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 532, + 877, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "recognition (ASR) captions from large-scale instructional videos to learn representations without human annotation [3, 4, 8, 35, 36]. The resulting models achieve state-of-the-art performance on zero-shot tasks such as cross-modal video retrieval or classification and also for zero-shot temporal action segmentation and detection based on free text queries [8, 28, 42, 47, 66], but usually lack spatial localization abilities. A third line of work focuses on label-free spatial grounding, e.g. by training on image-caption", + "bbox": [ + 496, + 763, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18419", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[2, 31, 50, 55, 62] or video-caption pairs [41, 46]. The goal is to correctly localize a referential expression in an image or each video frame, e.g., via a bounding box or a heatmap. However, those methods are not optimized to detect whether an event is present in a video. The assumption is thus that the evaluated expression is visible in the image or in all video frames.", + "bbox": [ + 75, + 90, + 468, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The following work aims to bring together those ideas to address the task of spatio-temporal action grounding from multimodal supervision in untrimmed videos. We propose a grounding approach that uses video-text pairs based on ASR transcripts in instructional videos and learns the spatial representation of free-text events as well as their temporal extent, as shown in Figure 1. To this end, we leverage two different representations of the visual data: a global feature representation based on full-frame information to define the temporal extent of an event and a local representation based on frame-wise grid features for spatial localization. The motivation for this dualism is that while the local representation captures the spatial correlations between vision and text input, this can be too fine-grained to learn a holistic representation of the frame, while the global representation can be assumed to capture a more compact, aggregated view compared to local data and thus to provide a more reliable cue for the task of temporal localization. However, compared to the hand-annotated video-caption setup of most spatio-temporal grounding methods, the ASR text can be noisy as not all comments refer to visible events. Further, as there is only a loose temporal correlation, the described activities might not be precisely aligned, can be scattered over multiple frames, or not be present at all [18, 36]. Therefore, we propose to specifically select frames to capture only those useful for training. To this end, we look for frames that match the vocabulary of the respective text, leveraging a selection strategy by Sinkhorn optimal transport [11]. This allows us to train a model that can localize actions in space and time within videos without labeling supervision.", + "bbox": [ + 75, + 199, + 472, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate spatio-temporal grounding in untrimmed videos, a new benchmark, GroundingYouTube, is proposed. It is based on the existing MiningYouTube dataset [28] and extended with spatio-temporal localization information. This setup differs from other benchmarks such as [10, 46, 63] in two ways: first, by using multiple center point annotations, it focuses on the grounding of referential actions itself instead of interacting humans or objects which are usually labeled; second, the dense annotations of multiple actions in the video allow us to benchmark action grounding in long, realistic untrimmed videos compared to existing, often pre-clipped benchmarks[10, 59]. The benchmark provides queries for 512 different event types and over $5K$ spatio-temporal annotations, as shown in Figure 1. A comparison of current datasets is shown in Table 1.", + "bbox": [ + 75, + 656, + 468, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the proposed approach as well as the new", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b6b877834e74c5dbd336dc72163de22a9ba9b6ba5920b8431f29201743d459bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAnnotation
SpatialTemporal
V-HiCoobject bb + human bb-
AVA-Kineticsobject bb + human bb-
ActivityNet entitiesobject bb + human bb-
THUMOS14-action boundaries
ActivityNet-action boundaries
HACSSegment-action boundaries
YouCook II-multi-action boundaries
Cross-Task-multi-action boundaries
COIN-multi-action boundaries
EPIC KITCHENS-100-multi-action boundaries
Ego4D-multi-action boundaries
JHMDB51-21human tubes-
UCF101-24human tubesaction boundaries
Dalyhuman tubesaction boundaries
Vid-STGhuman tubesaction boundaries
HC-STVGhuman tubesaction boundaries
AVAhuman tubesaction boundaries
YouCook-Interactionsaction bb-
GroundingYoutube (ours)action bb + center pointsmulti-action boundaries
", + "bbox": [ + 501, + 89, + 890, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of spatial, temporal, and spatio-temporal grounding datasets. V-HiCo [32], AVA-Kinetics [30], THU-MOS14 [20], ActivityNet [6], ActivityNet entity [65], HACSSegment [60], YouCook II [64], Cross-Task [66], COIN [47], EPIC KITCHENS-100 [12], Ego4D [14], JHMDB51-21 [21], UCF101-24 [44] Daly [51], Vid-STG [59], HC-STVG [48], AVA [15], YouCook-Interactions [46].", + "bbox": [ + 496, + 339, + 893, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "benchmark, the system is trained on the HowTo100M dataset [35] and compared to state-of-the-art methods based on full, weak, and self-supervision for spatial and temporal, as well as combined spatio-temporal grounding tasks. It shows that existing methods usually do well in one of the two aspects, spatial or temporal grounding. In contrast, the proposed method can combine spatial and temporal aspects.", + "bbox": [ + 496, + 454, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions of this work as follows1: (1) We propose a framework for spatio-temporal grounding in untrimmed videos based on weakly aligned multimodal supervision without human annotation, employing a combination of global and local representation learning to learn the spatio-temporal extent of actions. (2) To facilitate this task, we propose a frame selection strategy based on Sinkhorn-Knopp Optimal transport that improves the quality of the acquired learning samples, leading to more effective supervision. (3) We provide a new benchmark and annotations to evaluate this challenging problem on real-world multi-action instructional video data.", + "bbox": [ + 496, + 561, + 890, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 758, + 640, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Spatio-temporal Grounding. Spatiotemporal Grounding refers to the problem of localizing a sequence of bounding boxes (a spatio-temporal tube) for a target object described by an input text. This problem has been addressed by various approaches TubeDETR [54], STCAT [23], STVGBert [45], STVGBert [45],", + "bbox": [ + 496, + 784, + 890, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We will make the code and the annotations publicly available.", + "bbox": [ + 517, + 886, + 849, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18420", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "STGVT [48], STGRN [59], Visil [27]. These methods rely on proposal networks such as Faster R-CNN [39] or MDETR [25] to predict bounding box coordinates for learning text-to-region interaction. All those approaches rely on supervised training with the human-annotated sentence and bounding box supervision, provided, e.g., by datasets such as VidSTG [59], HC-STVG [10]. While those datasets provide a temporal aspect, temporal detection is usually limited to identifying the start and end frame of a single action in a video. Compared to that, an untrimmed setting usually comprises multiple actions in a longer video that can be separated by longer background sequences. This conceptually differs from previous works [10] that typically use short videos of around 5-10 seconds. Other datasets such as ActivityNet entities [65] provide only bounding boxes for noun phrases in the captions, namely the objects, which is related to object detection task and does not capture any spatial or temporal extent of actions.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Self-supervised Learning. The field of multimodal self-supervised learning aims to learn data representations by leveraging large amounts of unlabeled data with multiple modalities. Early works [13, 52] started by projecting images and text into a joint visual-language embedding space, where embeddings of semantically similar pairs are close. Those ideas have now grown into systems such as MIL-NCE [36] using the HowTo100M dataset [35] to train a video-language embedding space from 1.2 million instructional videos paired with text descriptions from ASR. Follow-up works, including [3, 4, 8, 40, 42] show that using videos without annotation enables an effective multimodal embedding space via contrastive learning.", + "bbox": [ + 75, + 369, + 472, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on those advantages, approaches started to address the problem of Spatial Video Grounding from multimodal self-supervised aiming to identify spatial locations in a trimmed video based on text descriptions without the need for bounding box annotation during training. One of the early works studied this task in the context of weakly supervised learning where we learn grounding with human-annotated captions of the video [63]. In this context, works [41, 46] have focused on object grounding benchmarks such as YouCook2-BoundingBox [64], which provides bounding box annotations for visible objects in cooking videos. Other works such as GLIP [31], Region-CLIP [62], and others [56, 58] combine the principles of large-scale vision language training with bounding box finetuning on object detection datasets [16, 34]. Recently, the YouCook-Interactions dataset [46] and CoMMA [46] have been proposed for the spatial grounding of objects and actions with multimodal self-supervision from HowTo100M videos. These works assume that the video is temporally clipped with respect to the grounding phrase.", + "bbox": [ + 75, + 566, + 472, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to that, Temporal Video Grounding aims to determine the set of consecutive frames corresponding to a", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text query in an untrimmed video [9, 22, 43], thus predicting temporal boundaries of action instances. Recent work such as MIL-NCE [36], MCN [8], and VideoCLIP [33] utilize large-scale pretraining for grounding actions temporally via text-to-frame similarity on video datasets such as MiningYouTube [28] or CrossTask [66] without proposals. However, these methods lack spatial localization ability [57, 61].", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. A Global-Local Framework for Spatio-Temporal Grounding", + "text_level": 1, + "bbox": [ + 498, + 210, + 890, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. General setup", + "text_level": 1, + "bbox": [ + 500, + 253, + 645, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the proposed method is to temporally and spatially localize actions based on free-text queries in untrimmed videos. To this end, two representations are learned, a local and a global one. We start with narrated video clips, each associated with a corresponding visual representation and text narration. Namely, for each clip $\\mathcal{X} = \\{\\mathcal{V},\\mathcal{S}\\}$ , let $\\mathcal{V}$ stand for the video clip and $\\mathcal{S}$ for the text narration sentence generated by the automatic speech recognition (ASR) system. Each clip $\\mathcal{V}$ consists of $U\\times N$ spatio-temporal tokens $\\{v_{u,n}\\}$ , where $u\\in \\{1,\\dots,U\\}$ represents the number of frames in the video and $n\\in \\{1,\\dots,N\\}$ represents the number of spatial grid region tokens or features in a frame. The text sentence $\\mathcal{S}$ consists of $K$ words $\\{s_1,\\ldots ,s_K\\}$ . We represent localized features by the tokens from each modality, and the global features $\\{V,S\\}$ are acquired either by mean-pooling over the local features (S3D) or by using the [CLS] token from the transformer (CLIP) as in Radford et al. [38]. We learn transformations $f:V\\to \\mathbb{R}^d$ to a $d$ -dimensional representation $f(V)\\in \\mathbb{R}^{d}$ from the global representation $V$ , and $g:S\\rightarrow \\mathbb{R}^d$ , to produce similar $d$ -dimensional text global embeddings: $g(S)\\in \\mathbb{R}^{d}$ . Similar to $\\{f,g\\}$ , we note $\\{f',g'\\}$ to be the transform for localized features, where local features $\\{v,s\\}$ are also projected as $d$ -dimensional representations.", + "bbox": [ + 496, + 277, + 893, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Representation guided frame sampling", + "text_level": 1, + "bbox": [ + 500, + 650, + 834, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning from multimodal self-supervision is challenging since the narration is likely to be noisy, thus containing more information than the actual task descriptions due to poor temporal alignment or cut scenes [18], which is the key differences between weakly supervised vision-caption grounding and multimodal self-supervised grounding. This work pursues a frame selection strategy to improve object grounding and temporal alignment during training. We start from a longer sequence $U$ , where $U > T$ , which includes the video frames before and after the ASR boundaries that could contain actions or objects in the sentence. Our goal is to find $T$ frames out of the $U$ frames that are most relevant to the actions and objects in the sentence $S$ . We formalize it as an optimal transport problem utilizing the Sinkhorn-Knopp algorithm [11].", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49a7b57db49a9fad34daef3db334c0b00e06c4e67b23e2a509527d564c1166bb.jpg", + "image_caption": [ + "Figure 2. Spatio-temporal grounding approach. (a) We aim to select frames with groundable objects and actions. To this end, projected text features are matched with respective frame features. (b) Sinkhorn optimal transport is then leveraged to optimize the selected frames wrt. the text input. (c) Based on the selected frames, a global representation is learned to allow for temporal localization, as well as (d) a local representation to ground the action description in the spatial region." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimal transport for text-to-frame assignment. To acquire the optimal assignment from word features to video frames, an assignment matrix $\\mathbf{Q}$ is computed from each video and ASR pair as shown in Figure 2(a). This cross-model optimal transport mechanism is applied to assignment $\\mathbf{Q}$ from the projected cross-model similarity $\\mathbf{P}$ between word tokens and each video frame, where $\\mathbf{P} = g(\\mathcal{S})\\otimes f(\\mathcal{V})^{\\top}\\in \\mathbb{R}^{K\\times U}$ . To compute the assignment matrix, the text and video projection layers from the global representation in Figure 2(c) are used to project multimodal features into a common space for feature similarity calculation. We investigate various granularities of the features where we compute the similarity between the text features at the word (local) / sentence (global) level and the visual feature at frames (global) / spatiotemporal tokens (local) level to acquire $\\mathbf{P}$ , as shown in Table 5. To ensure that the word-to-frame assignment contains more diversity instead of just saturated assignments to a single video frame, we add a constraint by Sinkhorn that requires label assignments to be equally distributed across various video frames representing diverse object/action concepts. Details of the Sinkhorn optimal transport are included in the appendix 7.1.", + "bbox": [ + 75, + 327, + 472, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Local representations for spatial localization", + "text_level": 1, + "bbox": [ + 76, + 676, + 454, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To capture multimodal interaction with finer granularity, we need to learn the projection between tokenized features as shown in Figure 2(d). We extract spatio-temporal region features $v_{tn}$ from the video. Also, we extract word features $s_k$ which represents the feature from word $k$ . All tokenized features are projected through a linear layer. To compute attention between the tokenized features, we stacked two cross-modal attention layers with a self-attention layer in the middle, as illustrated in Figure 2 (d). Cross-modal attention is computed similar to the standard attention mechanism [29]. Given a spatio-temporal token $v_{tn}$ from a video, we compute the attention score to all of the words $s_k$ , where $k \\in \\{1, \\dots, K\\}$ in the ASR sen", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tence $S$ by $\\alpha_{tnk} = \\frac{\\exp(e_{tnk})}{\\sum_{k=1}^{K}\\exp(e_{tnk})}$ in the same video clip, where $e_{tnk} = \\cosine(v_{tn}, s_k)$ . We then acquire a contextual video token feature $\\bar{v}_{tn} = \\sum_{k=1}^{K}\\alpha_{tnk}s_k$ , which encoded text contextual information. Note that the contextual vector is represented by aggregating the representations from the other modality. Follow the standard self-attention computation [49] $K, Q, V$ represent the features for the keys, queries, and values as: $Attn(K, Q, V) = \\text{softmax}\\left(\\frac{(Q^\\top K)}{\\sqrt{d_k}}\\right)V$ where $d_k$ is the dimension of the key. In our case, we feed each contextual feature $\\{\\bar{v}_{tn}, \\bar{s}_k\\}$ right after the first cross-attention layer to the $K, Q, V$ to acquire its self-attended representation. The localized attention model was trained using contrastive loss. To represent the video clip $\\mathcal{V}$ and ASR sentence $S$ , we mean-pool over the spatio-temporal tokens in video $\\bar{V} = \\frac{1}{TN}\\sum_{r=1}^{TN}\\bar{v}_r$ , and words $\\bar{S} = \\frac{1}{K}\\sum_{k=1}^{K}\\bar{s}_k$ respectively. Let $(\\bar{V}^{(l)}, \\bar{S}^{(l)})$ be the $l$ -th training example pair. We adopt the Noise Contrastive Estimation (NCE) loss [17] and the localized attention losses $\\mathcal{L}_{Local}$ :", + "bbox": [ + 496, + 324, + 893, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n- \\frac {1}{B} \\sum_ {l = 1} ^ {B} \\left[ \\left(\\log \\frac {e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta}}{e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta} + \\sum_ {k = 1} ^ {B} e ^ {\\bar {V} _ {k} ^ {i m p} \\cdot \\bar {S} _ {l}}}\\right) + \\left(\\log \\frac {e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta}}{e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta} + \\sum_ {k = 1} ^ {B} e ^ {\\bar {V} _ {k} \\cdot \\bar {S} _ {k} ^ {i m p}})}\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 638, + 893, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ is the batch. $\\bar{V}_k^{imp}$ and $\\bar{S}_k^{imp}$ represent imposter samples, and $\\delta$ is a margin hyperparameter.", + "bbox": [ + 496, + 715, + 893, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Global representations for temporal", + "text_level": 1, + "bbox": [ + 498, + 756, + 815, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We learn the global representation of a video clip and a sentence by contrastive loss, as shown in Figure 2(c). We again use the NCE loss function [17]. The global contrastive loss $\\mathcal{L}_{Global}$ follows the formulation as Equation 1 while using the global representations $V$ and $S$ , which are the [CLS] tokens or mean-pooled features from both modalities, instead of the local representations. Projecting the global features to the same space ensures that the features across different", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18422", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "modalities are comparable. Since global representations encode information from the entire video, it is essential in encoding temporal information for the later downstream tasks. The final model is optimized by the sum of both losses as $\\mathcal{L}_{\\text{Final}} = \\mathcal{L}_{\\text{Local}} + \\mathcal{L}_{\\text{Global}}$ .", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b207eab4a8d3fcbd28b955ccdffc923eb4ab598946c7d7a9d16906852e7ad8d1.jpg", + "image_caption": [ + "Figure 3. Spatio-temporal inference. Both representations are used for spatio-temporal grounding: Starting by predicting the action boundary, spatial grounding is performed on the selected frames using the predicted label to find corresponding regions." + ], + "image_footnote": [], + "bbox": [ + 81, + 181, + 465, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference for spatio-temporal grounding.", + "text_level": 1, + "bbox": [ + 76, + 388, + 426, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To perform spatio-temporal grounding on untrimmed videos, we start from temporal action detection as shown in Figure 3. Given a pool of possible action descriptions on the left and an untrimmed video, we perform feature similarity matching using the global representation ([CLS] token or mean-pooled feature) per frame with a threshold $\\tau$ to filter backgrounds. We pick the action class with the largest similarity score per frame. Later, we use the predicted action class and feed it into the local representation branch to compute spatial grounding. We follow [1] to compute feature similarity between visual tokens and text tokens through the cross-attention and self-attention. In the end, we acquire an attention heatmap for later downstream evaluation. More information on inferencing are in appendix 8.3.", + "bbox": [ + 75, + 412, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. GroundingYoutube Benchmark", + "text_level": 1, + "bbox": [ + 76, + 637, + 366, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Current downstream datasets either provide spatial [46], or temporal annotation [28], or spatio-temporal annotation [59] but only for short video clips with few frames before and after the action takes place. These datasets do not provide the opportunity to evaluate both aspects, spatial and temporal grounding, in an untrimmed long video manner. We, therefore, extend one of the current benchmarks, MiningYouTube [28], with 250 untrimmed videos with a duration of around 3-5 minutes and an average of 41.6 labeled action instances per video. Note that each video contains various classes. As the dataset already provides dense temporal annotations, we annotate the respective temporal action segments in the dataset with spatial information.", + "bbox": [ + 75, + 659, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotating the spatio-temporal extent of actions can be challenging as there is no clear visible outline as, e.g., in object annotation, nor is there a unique signal to indicate the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c03b6fbed03d0fab120053dfcf58d99f47add551043fcf3b267631841f87d9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 692, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30a57a3f217ab0dd9e1e39f780a2de1e0270101b09e6bb988983b72b4bdee44b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 90, + 866, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad091c8deff9164b3ab0688ace8c1714f4a8cead98b6f0cfb158045a96bf319c.jpg", + "image_caption": [ + "Figure 4. Visualization of the point annotation and automatic bounding box generation from points. The red point represents the mean of the five annotation points. The points annotation captures diverse patterns in various action types." + ], + "image_footnote": [], + "bbox": [ + 519, + 222, + 691, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0086a40ad8f7bc800a6bb2e54ada527117c024fc79a7131d2599bd7573ca4827.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 222, + 867, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "temporal begin and end points. Similarly, grounding systems do not produce pixel-exact bounding boxes, but rather indicate regions of interest. Detector-free spatial grounding models [5] address this fuzziness by relying on pointing game accuracy, thus only using the center point of the heat map for evaluation. Lending on this idea, annotators were asked to mark the presumed center point of the action. Compared to bounding boxes, center point annotation can be advantageous because annotators are not visually distracted by object outlines, so it is more likely that the most important region will be selected. We capture five annotations per frame, resulting in a density-based heatmap.", + "bbox": [ + 496, + 422, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Starting from 5,091 clips showing one of the 512 action classes, we adopt the methodology used for temporal action localization developed in [15] and label one frame per second, resulting in 26,987 labeled frames. We annotated all frames with five repeats per image, resulting in five annotations per frame and 134,935 point labels in total. To evaluate using bounding boxes [24], we get the union of all annotated points with additional distance to construct the bounding box as shown in Figure 4. More information on the annotation process, bounding box derivation, and dataset analysis is provided in the appendix 10.", + "bbox": [ + 496, + 604, + 892, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 789, + 632, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 815, + 601, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data: HowTo100M dataset contains 1.2M instructional videos along with their corresponding automatically generated speech (ASR). The narrations may be inaccurate and do not always accurately depict the video scene.", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2c1d64c827989b301853771407943e211c2333d60e3ebc2ee3dfcc8aef08976d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneDataSetSupervisionModalityGroundingYoutube
IoU+PointmAP
0.10.20.30.40.50.1:0.5
CoMMA† [46]S3DHT250KSelfVT1.022.181.721.110.930.371.26
MIL-NCE [36]S3D*HT100MSelfVT4.6733.9425.1612.653.420.4115.11
OursS3DHT100MSelfVT9.1242.7035.4925.1616.2210.0525.92
GLIP [31]Swin-L*Cap24MWeakIT1.242.832.101.520.960.371.56
CoMMA‡ [46]CLIPHT100MSelfVT1.683.512.321.880.990.401.82
CLIP [38]CLIPHT100MSelfIT3.5929.5422.159.162.480.3912.74
RegionCLIP [62]ResNet-101*CC3MWeakIT5.6535.6527.4315.694.310.8616.78
OursCLIPHT100MSelfVT10.0942.8136.0525.8417.1011.3526.63
OursCLIP*HT100MSelfVT11.5343.6436.9426.7819.4514.6128.26
MIL-NCE(temp.)+RegionCLIP(spa.)---VT9.2140.5434.9722.3813.799.1822.33
", + "bbox": [ + 98, + 88, + 874, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Spatio-temporal grounding on GroundingYouTube full videos. The proposed model learns global representations encoding global information and spatial correspondences across modalities, achieving a better performance in spatio-temporal evaluation compared to models trained on only spatial or temporal grounding. (V: video, I: image, T: text.) * indicates finetuned backbone.", + "bbox": [ + 75, + 285, + 892, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Downstream Datasets: GroundingYoutube (GYT) is used to evaluate the task of multi-action spatio-temporal grounding as described in Section 4. MiningYoutube (MYT) [28] provides temporal annotation and is limited to the domain of cooking instruction videos. YouCook-Interaction (YC-Inter) [46] is an extension of the YouCook2 dataset [64] for cooking instruction providing bounding boxes for 6K selected frames. The bounding boxes usually comprise the hand and the tool mentioned in the respective sentence-wise annotation. To further benchmark on general video domains on the V-HICO dataset [32] with 6.5k videos with human-object interaction bounding boxes annotations, and Daly action dataset [51], featuring videos consisting of daily actions such as \"brushing teeth\".", + "bbox": [ + 75, + 343, + 472, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baseline methods", + "text_level": 1, + "bbox": [ + 76, + 564, + 246, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal: MIL-NCE [36] utilizes S3D [53] and word2vec [37]. CLIP [38], an image-text model with transformer. Spatial: CoMMA [46], SSL model († for weights shared by the author² ‡ trained with CLIP); GLIP [31], RegionCLIP [62], SOTA weakly supervised grounding model. Spatio-temporal: We construct MIL-NCE+RegionCLIP following the inference pipeline in Figure 3. TubeDETR [54] and STCAT [23] are supervised. More descriptions of the baselines are given in the Appendix 8.1. Details of the implementation and experimental settings can be found in the appendix 8.2. Inference setups for each baseline are described in Section 8.3.", + "bbox": [ + 75, + 587, + 468, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Downstream Tasks", + "text_level": 1, + "bbox": [ + 76, + 779, + 259, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We considered the following downstream tasks to evaluate spatio-temporal grounding abilities of various models (detailed description is included in the appendix 8.4):", + "bbox": [ + 75, + 801, + 468, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(i) Spatio-temporal grounding in untrimmed video is evaluated on the proposed Grounding Youtube dataset. The", + "bbox": [ + 76, + 847, + 468, + 877 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "entire video and the respective pool of action instructions were provided. The model needs to localize each action step in time (start-time/end-time) and space (location in the video) as described in Figure 3. We evaluate in two metrics: IoU+Pointing game combines the evaluation setting from the spatial grounding [2] and temporal grounding [28] metrics. For each video frame, the prediction is correct when the model predicts the correct action for the frame. Also, given the predicted action as a query, the maximum point of the heatmap aims to lie within the desired bounding box. We then compute the Intersection over Union (IoU) over all the predictions with the GT to acquire the final score. We also compute video mAP following previous evaluation [15], where we set IoU threshold between GT and predicted spatio-temporal tubes. A prediction is correct when it surpasses the IoU threshold. We then compute the mAP over all classes. We form a 3D prediction mask following Figure 3 and compute IoU between our 3D heatmap and 3D tube.", + "bbox": [ + 500, + 343, + 890, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(ii) Spatial grounding is given a text description to localize the region in the trimmed video. It is evaluated using the pointing game accuracy. If the predicted point lies in the ground truth bounding box, the result counts as a “hit” and counts as “miss” otherwise. The final accuracy is calculated as a ratio between hits to the total number of predictions $\\frac{\\# \\text{hits}}{\\# \\text{hits} + \\# \\text{misses}}$ . We also report the mean average precision (mAP) following the settings from V-HICO [32].", + "bbox": [ + 496, + 628, + 890, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(iii) Temporal grounding provides videos with the respective actions and their ordering, including the background. The goal is to find the correct frame-wise segmentation of the video. We follow the inference procedure in [28] to compute the alignment given the similarity input matrix. The task is evaluated by intersection over detection (IoD), defined as $\\frac{G \\cap D}{D}$ the ratio between the intersection of ground-truth action $G$ and prediction $D$ to prediction $D$ , and the Jaccard index, which is an (IoU) given as $\\frac{G \\cap D}{G \\cup D}$ .", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2We thank the authors for providing code and weights.", + "bbox": [ + 94, + 886, + 383, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18424", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e65c85cfac99c243fbe8b753f3d223c78fee9c8b59e8da18b462b13bc2837d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneDataSuper.Mod.YC-InterGroundingYTV-HICODaly
AccmAPAccmAPAccmAPAccmAP
MIL-NCE [36]S3D*HT100MSelfVT23.6727.458.2112.6511.2313.8424.23
CoMMA† [46]S3DHT250KSelfVT48.6347.6823.3840.9721.4554.4833.39
OursS3DHT100MSelfVT53.9860.6244.9344.3224.3166.3545.93
CLIP [38]CLIPHT100MSelfIT14.1012.503.4929.2312.5118.0227.28
CoMMA‡ [46]CLIPHT100MSelfVT52.6547.5636.4255.2034.5461.0644.37
RegionCLIP [62]RN50x4*CC3MWeakIT51.5652.8423.4257.9237.8267.1248.62
GLIP [31]Swin-L*Cap24MWeakIT52.8453.6224.7366.0541.17--
OursCLIPHT100MSelfVT57.1055.4943.1260.7139.2870.0850.56
OursCLIP*HT100MSelfVT58.3556.9845.3262.3441.5671.3552.78
TubeDETR [54]MDETRVid-STGFullVT51.6353.2441.7663.2340.8784.2162.98
STCAT [23]ResNet-101Vid-STGFullVT54.4755.9044.2165.3441.1085.4263.94
", + "bbox": [ + 96, + 88, + 870, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b821de0137db88eb782c0f1bde06435ccd885aba43daed54389a2df808344103.jpg", + "table_caption": [ + "Table 3. Video spatial grounding. We evaluate the accuracy of the pointing game and the mean average precision. We listed CNN-based methods on top and transformer-based methods in the middle. Models learning global representations (MIL-NCE, CLIP) don't perform well on localization tasks, while our model outperforms other grounding methods. * indicates finetuned backbone." + ], + "table_footnote": [ + "Table 4. Temporal Grounding on MiningYoutube. * indicates finetuned backbone. Spatial-focused model CoMMA is not trained for temporal detection, which results in lower performance. In contrast, the proposed model combines global and local representation, resulting in better temporal localization than one alone." + ], + "table_body": "
MethodBackboneDataSuper.IoUIoD
Mining: MLP [28]TSMMiningYTWeak9.8019.20
CoMMA* [46]S3DHT250KSelf2.055.63
RegionCLIP [62]RN50x4*CC3MWeak10.9616.86
MIL-NCE [36]S3D*HT100MSelf18.6926.74
OursS3DHT100MSelf19.1827.65
OursCLIPHT100MSelf19.8828.50
OursCLIP*HT100MSelf20.3329.67
", + "bbox": [ + 78, + 325, + 468, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 76, + 544, + 434, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(i) Spatio-temporal grounding in untrimmed video: We first compare the proposed method with other approaches designed for spatial or temporal grounding in Table 2. It shows that models without specific loss designs for spatial grounding (MIL-NCE [36], CLIP [38]) show good mAP scores but lower pointing game accuracy. Out of the two weakly supervised methods, GLIP [31] and RegionCLIP [62]), trained with aligned image-text, RegionCLIP show significantly better performance in this setting, while both perform in a similar range in the spatial grounding scenario (see Table 3). We attribute this behavior to the fact that RegionCLIP distinguishes frames with relevant queries better from background than GLIP, leading to better temporal localization. We finally compare the strong baseline MIL-NCE+RegionCLIP, which combines two approaches specialized in temporal and spatial aspects, to our task. It shows that the proposed method improves over all other baselines underlining the need to incorporate global (temporal) and local (spatial) representations.", + "bbox": [ + 75, + 566, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(ii) Spatial grounding: Second, we compare the performance of the proposed framework to other methods on the task of spatial grounding, including models with weak su", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "persion, as well as models trained in a fully supervised setting in Table 3. In the instruction video domain (GYT and YC-Inter), the proposed approach achieves the best result among all weakly and self-supervised trained methods. In the general domain (V-HICO and Daly), the method also achieves competitive results, showing the generalizability of the model to other domains. Note that in the Daly dataset, the classes are verbs, which are not detectable by the object-focused model GLIP. Compared to their weakly trained counterparts, fully-supervised model (TubeDETER [54], STCAT [23]) achieve competitive performance in the general domain (V-HICO, Daly) and slightly lower performance in instruction domain (GYT, YC-Inter) due to the domain gap with respect to the training data.", + "bbox": [ + 496, + 327, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(iii) Temporal grounding: We evaluate temporal grounding in Table 4. Here, it shows that global representations also profit from local representation learning. This hypothesis is further validated in the ablation studies in Table 6, where we ablate both losses for all three settings and show a consistent improvement in the joint loss formulation.", + "bbox": [ + 496, + 540, + 890, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation study", + "text_level": 1, + "bbox": [ + 498, + 646, + 650, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation studies with respect to all three settings, spatio-temporal grounding, as well as spatial and temporal grounding alone, reporting performance for spatiotemporal grounding on GroundingYT using mAP with IoU@0.4, on temporal grounding using MiningYT IoU, and on spatial grounding using YC-Inter. pointing game. Additional ablation are in appendix 9.3.", + "bbox": [ + 496, + 672, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Frame selection strategy. We perform an ablation on the possible frame selection strategies for our method (Figure 2(b) and Section 3.2). In Table 5, None uses all frames within the ASR boundary ( $U = T$ ) as our video training data. Global represents the [CLS] token in text and video. Local uses the words and spatio-temporal tokens. In the setting Sinkhorn was not applied, the top $T$ frames with the highest similarity score were selected. When we set", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/759ade2634013100c3f9e42c2560075e5119a2bbed1f44f043240f042c157ac8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GroundingYT Spatio-temporalMiningYT TemporalYC-Inter. SpatialGroundingYT Spatio-temporalMiningYT TemporalYC-Inter. Spatial
w/o Sinkhornwith Sinkhorn
None17.4318.3457.42---
Global T - Global V18.2419.3856.3118.9619.8957.34
Global T - Local V18.0119.3157.5618.5319.3557.67
Local T - Global V18.0519.8557.3419.3220.3658.13
Local T - Local V18.3119.4857.8619.4320.1658.51
Average over last two18.3619.6857.7719.4520.3358.35
", + "bbox": [ + 96, + 88, + 874, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73e882c8d9660d738c349f16054c135f397b4dd1c4e911f56321b4a29f2df7f0.jpg", + "image_caption": [ + "Figure 5. Visualization on GroundingYoutube. Red box: annotation. Heatmap: prediction. Baseline GLIP focuses on objects such as power, mixer, and pancake. Our method focuses on actions such as whisk and serve." + ], + "image_footnote": [], + "bbox": [ + 93, + 262, + 302, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7141121e14a79562ce1ea8bac23c80d692fd8ab0cfb423d7a3ef0d4e0540c10c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 262, + 493, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/735330791fee70a416a935d899d1a4b6853334043371955a3b94519f0ab0ee19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 262, + 681, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32834c1a9aa6bd0428a07e2d7f0e34fd816ef19acaf3aec0d1002f20ee189d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 262, + 874, + 393 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/af1bf75b2b108a23c817a10e71d814f9321d63288c47b4b189cb4bc55517c25f.jpg", + "table_caption": [ + "Table 5. Frame selection: (a) Sinkhorn selection results in better supervision. (b) We investigate all possible combinations of global and local representations for frame selection similarity matching. We found keeping the local text representation is crucial, and a combination of local and global representation leads to the best spatio-temporal grounding result." + ], + "table_footnote": [], + "table_body": "
GroundingYT Spatio-tempMiningYT TemporalYC-Inter. Spatial
only Local loss7.295.2355.29
only Global loss9.2819.1236.23
w/ Both loss19.4520.3358.35
", + "bbox": [ + 78, + 435, + 468, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Loss ablations: both losses contribute to the final loss, and the existence of global loss helps the localization task.", + "bbox": [ + 76, + 507, + 468, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "spatio-temporal tokens as the selection target, we sum over the scores with respect to each frame to acquire the frame similarity score. It shows that selecting frames based on Sinkhorn selection leads to consistently better results as it enforces more variety of visual concepts but also captures frames with possible groundable objects. It further shows that word tokens are more suitable than the global text CLS token for frame selection. Finally, we see that depending on the task (spatial vs. temporal), a local resp. global representation is better, and a combination of both works best for spatio-temporal grounding. We provide runtime analysis of such frame selection strategy in the appendix 9.1.", + "bbox": [ + 75, + 551, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Global and local loss. As mentioned in the spatio-temporal evaluation, both features contribute to the final grounding result. We test the model by ablating out each loss. Table 6 shows that each loss not only contributes to the spatiotemporal grounding on the GYT, but also that the whole is more than the sum of its parts (losses) since this task requires both spatial and temporal detection. The reduced impact of the global loss in the case of YC-Inter is that this is a pure spatial grounding dataset (no background frames) without temporal detection, and the local loss plays a more critical role. We observe the same patterns in the temporal", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "grounding result for MYT, where spatial localization is not directly contributing to the final performance. We tried out the same ablation using in the S3D backbone in supplement.", + "bbox": [ + 498, + 449, + 892, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 507, + 676, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize our spatio-temporal result in Figure 5. For the GLIP model, we output the bounding box with the highest confidence score and visualize its center point. We found GLIP model focuses on the salient object while our model focuses more on human-object interaction.", + "bbox": [ + 498, + 527, + 892, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 623, + 617, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented an approach for learning spatio-temporal grounding with self-supervision and a new dataset: GroundingYoutube annotations, where we densely annotate spatiotemporal points/boxes from untrimmed multi-action videos. Our method includes a frame selection mechanism that identifies frames with groundable objects to adapt the learning process for untrimmed videos. Furthermore, we jointly learn global representations, which capture temporal information, and local representations learning fine-grained multimodal interactions between video and text. We conducted extensive experiments and our approach shows state-of-the-art performance in spatio-temporal grounding, as well as temporal and spatial grounding alone.", + "bbox": [ + 496, + 642, + 892, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: Nina Shvetsova is supported by the German Federal Ministry of Education and Research (BMBF) project STCL-01IS22067. Andrew Rouditchenko is supported by an NDSEG Fellowship and by the MIT-IBM Watson AI Lab. We thank the MIT-Satori cluster team for their technical support.", + "bbox": [ + 498, + 842, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18426", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 5, 13", + "[2] Hassan Akbari, Svebor Karaman, Surabhi Bhargava, Brian Chen, Carl Vondrick, and Shih-Fu Chang. Multi-level multimodal common semantic space for image-phrase grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12476–12486, 2019. 2, 6, 13", + "[3] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 1, 3", + "[4] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. Advances in Neural Information Processing Systems, 33:25-37, 2020. 1, 3", + "[5] Assaf Arbelle, Sivan Doveh, Amit Alfassy, Joseph Shtok, Guy Lev, Eli Schwartz, Hilde Kuehne, Hila Barak Levi, Prasanna Sattigeri, Rameswar Panda, et al. Detector-free weakly supervised grounding by separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1801-1812, 2021. 5", + "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 12", + "[8] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8012-8021, 2021. 1, 3, 14", + "[9] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. Advances in Neural Information Processing Systems, 34:28442-28453, 2021. 3", + "[10] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee Kenneth Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1884-1894, Florence, Italy, 2019. Association for Computational Linguistics. 2, 3", + "[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. 2013. 2, 3, 12", + "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 2", + "[13] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3", + "[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2", + "[15] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 2, 5, 6, 13, 14", + "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3", + "[17] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTATS, 2010. 4", + "[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2, 3", + "[19] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649–665, 2018. 14", + "[20] H. Idrees, A. R. Zamir, Y. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2", + "[21] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2", + "[22] Yu-Gang Jiang, Jingen Liu, A Roshan Zamir, George Toderici, Ivan Laptev, Mubarak Shah, and Rahul Sukthankar. Thumos challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.3,14", + "[23] Yang Jin, yongzhi li, Zehuan Yuan, and Yadong MU. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In Advances in Neural Information Processing Systems, 2022. 1, 2, 6, 7, 14, 15", + "[24] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In Proceedings of the IEEE" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18427", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 4405-4413, 2017. 5", + "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2015. 13", + "[27] Giorgos Kordopatis-Zilos, Symeon Papadopoulos, Ioannis Patras, and Ioannis Kompatsiaris. Visil: Fine-grained spatiotemporal video similarity learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6351-6360, 2019. 3", + "[28] Hilde Kuehne, Ahsan Iqbal, Alexander Richard, and Juergen Gall. Mining youtube-a dataset for learning fine-grained action concepts from webly supervised video data. 2019. 1, 2, 3, 5, 6, 7, 13", + "[29] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 4", + "[30] Ang Li, Meghan Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv preprint arXiv:2005.00214, 2020. 2", + "[31] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 6, 7, 12, 15", + "[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1845-1855, 2021. 2, 6, 13", + "[33] Yikang Li, Jenhao Hsiao, and Chiuman Ho. Videoclip: A cross-attention model for fast video-text retrieval task with image clip. In Proceedings of the 2022 International Conference on Multimedia Retrieval, pages 29–33, 2022. 3", + "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3", + "[35] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 12", + "[36] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 3, 6, 7, 12, 15", + "[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In arXiv preprint arXiv:1301.3781, 2013. 6, 12", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 12, 15", + "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3", + "[40] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv preprint arXiv:2006.09199, 2020.3", + "[41] Jing Shi, Jia Xu, Boqing Gong, and Chenliang Xu. Not all frames are equal: Weakly-supervised video grounding with contextual similarity and visual clustering losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10444-10452, 2019. 2, 3", + "[42] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In CVPR, 2022. 1, 3, 14", + "[43] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba Heilbron, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. arXiv preprint arXiv:2112.00431, 2021. 3", + "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2", + "[45] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1533–1542, 2021. 2", + "[46] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5, 6, 7, 12, 15", + "[47] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 1, 2", + "[48] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18428", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2, 3", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 2", + "[51] Philippe Weinzaepfel, Xavier Martin, and Cordelia Schmid. Human action localization with sparse spatial supervision. arXiv preprint arXiv:1605.05197, 2016. 2, 6", + "[52] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence, 2011. 3", + "[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6, 12", + "[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 1, 2, 6, 7, 14, 15", + "[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2", + "[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 3", + "[57] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3", + "[58] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 3", + "[59] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10668-10677, 2020. 2, 3, 5", + "[60] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In Pro-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019. 2", + "[61] Yang Zhao, Zhou Zhao, Zhu Zhang, and Zhijie Lin. Cascaded prediction network via segment tree for temporal video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4206, 2021. 3", + "[62] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2, 3, 6, 7, 12, 15", + "[63] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In British Machine Vision Conference, 2018. 2, 3", + "[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 2, 3, 6", + "[65] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In CVPR, 2019. 2, 3", + "[66] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. 2019. 1, 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 486 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18429", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_model.json b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..213a3c0416408166481f1b06e0084bef71a3c0f9 --- /dev/null +++ b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.84, + 0.175 + ], + "angle": 0, + "content": "What, when, and where? Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.844, + 0.24 + ], + "angle": 0, + "content": "Brian Chen1 Nina Shvetsova2,3 Andrew Rouditchenko4 Daniel Kondermann5 Samuel Thomas6,7 Shih-Fu Chang1 Rogerio Feris6,7 James Glass4 Hilde Kuehne2,3,7" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.244, + 0.748, + 0.259 + ], + "angle": 0, + "content": "1Columbia University, 2Goethe University Frankfurt, 3University of Bonn, 4MIT CSAIL" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.262, + 0.697, + 0.277 + ], + "angle": 0, + "content": "\\(^{5}\\)Quality Match GmbH, \\(^{6}\\)IBM Research AI, \\(^{7}\\)MIT-IBM Watson AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.281, + 0.815, + 0.296 + ], + "angle": 0, + "content": "{bc2754,sc250}@columbia.edu, shvetsov@uni-frankfurt.de,{roudi,glass}@mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.299, + 0.797, + 0.313 + ], + "angle": 0, + "content": "dk@quality-match.com, {sthomas,rsferis}@us.ibm.com, kuehne@cs.uni-bonn.de" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.348, + 0.314, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed, providing dense spatio-temporal grounding annotations in long, untrimmed, multi-action instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks, showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatio-temporal grounding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Spatio-temporal grounding (STG) describes the challenging task of locating events in space and time within video data based on text referential expressions. Methods in this field usually rely on a combination of spatio-temporal bounding box annotation, together with a human-generated caption, describing the visual content of the bounding box [23, 54], which limits their generalizability beyond the given training scenario. Compared to that, as a second line of work, multimodal self-supervised learning tries to leverage \"free\" data sources, such as video and automatic speech" + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.355, + 0.858, + 0.368 + ], + "angle": 0, + "content": "Training Setup: Unlabeled videos with narrated instructions" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.371, + 0.608, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.414, + 0.597, + 0.439 + ], + "angle": 0, + "content": "\"To this we will add one carrot chopped into jong jullien ...\"" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.371, + 0.694, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.414, + 0.692, + 0.438 + ], + "angle": 0, + "content": "\"... you can use the forks to pull it a bit apart...\"" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.369, + 0.77, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.413, + 0.771, + 0.438 + ], + "angle": 0, + "content": "①" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.369, + 0.872, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.419, + 0.864, + 0.435 + ], + "angle": 0, + "content": "...stir it a bit so it mixes well..." + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.452, + 0.877, + 0.465 + ], + "angle": 0, + "content": "Evaluation Setup: Referential queries - \"Crack egg\", \"Mix egg\", etc." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.467, + 0.882, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.533, + 0.696, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.533, + 0.879, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.582, + 0.881, + 0.605 + ], + "angle": 0, + "content": "Task: Spatio-Temporal Grounding - Find the temporal boundary of a queried action in an untrimmed video and spatially localize the action." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.733 + ], + "angle": 0, + "content": "Figure 1. Learning Spatio-temporal grounding in untrimmed videos. In training, we learn from unlabeled videos without human annotation. In evaluation, we perform spatio-temporal grounding using an action description such as \"crack egg\" as a query. The model needs to localize both the action's temporal boundary and spatial region in the long untrimmed video. We visualize the heatmap from the annotation points as well as derived bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.903 + ], + "angle": 0, + "content": "recognition (ASR) captions from large-scale instructional videos to learn representations without human annotation [3, 4, 8, 35, 36]. The resulting models achieve state-of-the-art performance on zero-shot tasks such as cross-modal video retrieval or classification and also for zero-shot temporal action segmentation and detection based on free text queries [8, 28, 42, 47, 66], but usually lack spatial localization abilities. A third line of work focuses on label-free spatial grounding, e.g. by training on image-caption" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.196 + ], + "angle": 0, + "content": "[2, 31, 50, 55, 62] or video-caption pairs [41, 46]. The goal is to correctly localize a referential expression in an image or each video frame, e.g., via a bounding box or a heatmap. However, those methods are not optimized to detect whether an event is present in a video. The assumption is thus that the evaluated expression is visible in the image or in all video frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.473, + 0.654 + ], + "angle": 0, + "content": "The following work aims to bring together those ideas to address the task of spatio-temporal action grounding from multimodal supervision in untrimmed videos. We propose a grounding approach that uses video-text pairs based on ASR transcripts in instructional videos and learns the spatial representation of free-text events as well as their temporal extent, as shown in Figure 1. To this end, we leverage two different representations of the visual data: a global feature representation based on full-frame information to define the temporal extent of an event and a local representation based on frame-wise grid features for spatial localization. The motivation for this dualism is that while the local representation captures the spatial correlations between vision and text input, this can be too fine-grained to learn a holistic representation of the frame, while the global representation can be assumed to capture a more compact, aggregated view compared to local data and thus to provide a more reliable cue for the task of temporal localization. However, compared to the hand-annotated video-caption setup of most spatio-temporal grounding methods, the ASR text can be noisy as not all comments refer to visible events. Further, as there is only a loose temporal correlation, the described activities might not be precisely aligned, can be scattered over multiple frames, or not be present at all [18, 36]. Therefore, we propose to specifically select frames to capture only those useful for training. To this end, we look for frames that match the vocabulary of the respective text, leveraging a selection strategy by Sinkhorn optimal transport [11]. This allows us to train a model that can localize actions in space and time within videos without labeling supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.882 + ], + "angle": 0, + "content": "To evaluate spatio-temporal grounding in untrimmed videos, a new benchmark, GroundingYouTube, is proposed. It is based on the existing MiningYouTube dataset [28] and extended with spatio-temporal localization information. This setup differs from other benchmarks such as [10, 46, 63] in two ways: first, by using multiple center point annotations, it focuses on the grounding of referential actions itself instead of interacting humans or objects which are usually labeled; second, the dense annotations of multiple actions in the video allow us to benchmark action grounding in long, realistic untrimmed videos compared to existing, often pre-clipped benchmarks[10, 59]. The benchmark provides queries for 512 different event types and over \\(5K\\) spatio-temporal annotations, as shown in Figure 1. A comparison of current datasets is shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To evaluate the proposed approach as well as the new" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.329 + ], + "angle": 0, + "content": "
DatasetAnnotation
SpatialTemporal
V-HiCoobject bb + human bb-
AVA-Kineticsobject bb + human bb-
ActivityNet entitiesobject bb + human bb-
THUMOS14-action boundaries
ActivityNet-action boundaries
HACSSegment-action boundaries
YouCook II-multi-action boundaries
Cross-Task-multi-action boundaries
COIN-multi-action boundaries
EPIC KITCHENS-100-multi-action boundaries
Ego4D-multi-action boundaries
JHMDB51-21human tubes-
UCF101-24human tubesaction boundaries
Dalyhuman tubesaction boundaries
Vid-STGhuman tubesaction boundaries
HC-STVGhuman tubesaction boundaries
AVAhuman tubesaction boundaries
YouCook-Interactionsaction bb-
GroundingYoutube (ours)action bb + center pointsmulti-action boundaries
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.34, + 0.894, + 0.437 + ], + "angle": 0, + "content": "Table 1. Comparison of spatial, temporal, and spatio-temporal grounding datasets. V-HiCo [32], AVA-Kinetics [30], THU-MOS14 [20], ActivityNet [6], ActivityNet entity [65], HACSSegment [60], YouCook II [64], Cross-Task [66], COIN [47], EPIC KITCHENS-100 [12], Ego4D [14], JHMDB51-21 [21], UCF101-24 [44] Daly [51], Vid-STG [59], HC-STVG [48], AVA [15], YouCook-Interactions [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.56 + ], + "angle": 0, + "content": "benchmark, the system is trained on the HowTo100M dataset [35] and compared to state-of-the-art methods based on full, weak, and self-supervision for spatial and temporal, as well as combined spatio-temporal grounding tasks. It shows that existing methods usually do well in one of the two aspects, spatial or temporal grounding. In contrast, the proposed method can combine spatial and temporal aspects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.742 + ], + "angle": 0, + "content": "We summarize the contributions of this work as follows1: (1) We propose a framework for spatio-temporal grounding in untrimmed videos based on weakly aligned multimodal supervision without human annotation, employing a combination of global and local representation learning to learn the spatio-temporal extent of actions. (2) To facilitate this task, we propose a frame selection strategy based on Sinkhorn-Knopp Optimal transport that improves the quality of the acquired learning samples, leading to more effective supervision. (3) We provide a new benchmark and annotations to evaluate this challenging problem on real-world multi-action instructional video data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.759, + 0.642, + 0.774 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Supervised Spatio-temporal Grounding. Spatiotemporal Grounding refers to the problem of localizing a sequence of bounding boxes (a spatio-temporal tube) for a target object described by an input text. This problem has been addressed by various approaches TubeDETR [54], STCAT [23], STVGBert [45], STVGBert [45]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.851, + 0.9 + ], + "angle": 0, + "content": "1We will make the code and the annotations publicly available." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "STGVT [48], STGRN [59], Visil [27]. These methods rely on proposal networks such as Faster R-CNN [39] or MDETR [25] to predict bounding box coordinates for learning text-to-region interaction. All those approaches rely on supervised training with the human-annotated sentence and bounding box supervision, provided, e.g., by datasets such as VidSTG [59], HC-STVG [10]. While those datasets provide a temporal aspect, temporal detection is usually limited to identifying the start and end frame of a single action in a video. Compared to that, an untrimmed setting usually comprises multiple actions in a longer video that can be separated by longer background sequences. This conceptually differs from previous works [10] that typically use short videos of around 5-10 seconds. Other datasets such as ActivityNet entities [65] provide only bounding boxes for noun phrases in the captions, namely the objects, which is related to object detection task and does not capture any spatial or temporal extent of actions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.567 + ], + "angle": 0, + "content": "Multimodal Self-supervised Learning. The field of multimodal self-supervised learning aims to learn data representations by leveraging large amounts of unlabeled data with multiple modalities. Early works [13, 52] started by projecting images and text into a joint visual-language embedding space, where embeddings of semantically similar pairs are close. Those ideas have now grown into systems such as MIL-NCE [36] using the HowTo100M dataset [35] to train a video-language embedding space from 1.2 million instructional videos paired with text descriptions from ASR. Follow-up works, including [3, 4, 8, 40, 42] show that using videos without annotation enables an effective multimodal embedding space via contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Based on those advantages, approaches started to address the problem of Spatial Video Grounding from multimodal self-supervised aiming to identify spatial locations in a trimmed video based on text descriptions without the need for bounding box annotation during training. One of the early works studied this task in the context of weakly supervised learning where we learn grounding with human-annotated captions of the video [63]. In this context, works [41, 46] have focused on object grounding benchmarks such as YouCook2-BoundingBox [64], which provides bounding box annotations for visible objects in cooking videos. Other works such as GLIP [31], Region-CLIP [62], and others [56, 58] combine the principles of large-scale vision language training with bounding box finetuning on object detection datasets [16, 34]. Recently, the YouCook-Interactions dataset [46] and CoMMA [46] have been proposed for the spatial grounding of objects and actions with multimodal self-supervision from HowTo100M videos. These works assume that the video is temporally clipped with respect to the grounding phrase." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Compared to that, Temporal Video Grounding aims to determine the set of consecutive frames corresponding to a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "text query in an untrimmed video [9, 22, 43], thus predicting temporal boundaries of action instances. Recent work such as MIL-NCE [36], MCN [8], and VideoCLIP [33] utilize large-scale pretraining for grounding actions temporally via text-to-frame similarity on video datasets such as MiningYouTube [28] or CrossTask [66] without proposals. However, these methods lack spatial localization ability [57, 61]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.212, + 0.892, + 0.248 + ], + "angle": 0, + "content": "3. A Global-Local Framework for Spatio-Temporal Grounding" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.255, + 0.646, + 0.272 + ], + "angle": 0, + "content": "3.1. General setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.895, + 0.642 + ], + "angle": 0, + "content": "The goal of the proposed method is to temporally and spatially localize actions based on free-text queries in untrimmed videos. To this end, two representations are learned, a local and a global one. We start with narrated video clips, each associated with a corresponding visual representation and text narration. Namely, for each clip \\(\\mathcal{X} = \\{\\mathcal{V},\\mathcal{S}\\}\\), let \\(\\mathcal{V}\\) stand for the video clip and \\(\\mathcal{S}\\) for the text narration sentence generated by the automatic speech recognition (ASR) system. Each clip \\(\\mathcal{V}\\) consists of \\(U\\times N\\) spatio-temporal tokens \\(\\{v_{u,n}\\}\\), where \\(u\\in \\{1,\\dots,U\\}\\) represents the number of frames in the video and \\(n\\in \\{1,\\dots,N\\}\\) represents the number of spatial grid region tokens or features in a frame. The text sentence \\(\\mathcal{S}\\) consists of \\(K\\) words \\(\\{s_1,\\ldots ,s_K\\}\\). We represent localized features by the tokens from each modality, and the global features \\(\\{V,S\\}\\) are acquired either by mean-pooling over the local features (S3D) or by using the [CLS] token from the transformer (CLIP) as in Radford et al. [38]. We learn transformations \\(f:V\\to \\mathbb{R}^d\\) to a \\(d\\)-dimensional representation \\(f(V)\\in \\mathbb{R}^{d}\\) from the global representation \\(V\\), and \\(g:S\\rightarrow \\mathbb{R}^d\\), to produce similar \\(d\\)-dimensional text global embeddings: \\(g(S)\\in \\mathbb{R}^{d}\\). Similar to \\(\\{f,g\\}\\), we note \\(\\{f',g'\\}\\) to be the transform for localized features, where local features \\(\\{v,s\\}\\) are also projected as \\(d\\)-dimensional representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.836, + 0.668 + ], + "angle": 0, + "content": "3.2. Representation guided frame sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Learning from multimodal self-supervision is challenging since the narration is likely to be noisy, thus containing more information than the actual task descriptions due to poor temporal alignment or cut scenes [18], which is the key differences between weakly supervised vision-caption grounding and multimodal self-supervised grounding. This work pursues a frame selection strategy to improve object grounding and temporal alignment during training. We start from a longer sequence \\( U \\), where \\( U > T \\), which includes the video frames before and after the ASR boundaries that could contain actions or objects in the sentence. Our goal is to find \\( T \\) frames out of the \\( U \\) frames that are most relevant to the actions and objects in the sentence \\( S \\). We formalize it as an optimal transport problem utilizing the Sinkhorn-Knopp algorithm [11]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18421" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.897, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Spatio-temporal grounding approach. (a) We aim to select frames with groundable objects and actions. To this end, projected text features are matched with respective frame features. (b) Sinkhorn optimal transport is then leveraged to optimize the selected frames wrt. the text input. (c) Based on the selected frames, a global representation is learned to allow for temporal localization, as well as (d) a local representation to ground the action description in the spatial region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.473, + 0.661 + ], + "angle": 0, + "content": "Optimal transport for text-to-frame assignment. To acquire the optimal assignment from word features to video frames, an assignment matrix \\(\\mathbf{Q}\\) is computed from each video and ASR pair as shown in Figure 2(a). This cross-model optimal transport mechanism is applied to assignment \\(\\mathbf{Q}\\) from the projected cross-model similarity \\(\\mathbf{P}\\) between word tokens and each video frame, where \\(\\mathbf{P} = g(\\mathcal{S})\\otimes f(\\mathcal{V})^{\\top}\\in \\mathbb{R}^{K\\times U}\\). To compute the assignment matrix, the text and video projection layers from the global representation in Figure 2(c) are used to project multimodal features into a common space for feature similarity calculation. We investigate various granularities of the features where we compute the similarity between the text features at the word (local) / sentence (global) level and the visual feature at frames (global) / spatiotemporal tokens (local) level to acquire \\(\\mathbf{P}\\), as shown in Table 5. To ensure that the word-to-frame assignment contains more diversity instead of just saturated assignments to a single video frame, we add a constraint by Sinkhorn that requires label assignments to be equally distributed across various video frames representing diverse object/action concepts. Details of the Sinkhorn optimal transport are included in the appendix 7.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.455, + 0.695 + ], + "angle": 0, + "content": "3.3. Local representations for spatial localization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To capture multimodal interaction with finer granularity, we need to learn the projection between tokenized features as shown in Figure 2(d). We extract spatio-temporal region features \\( v_{tn} \\) from the video. Also, we extract word features \\( s_k \\) which represents the feature from word \\( k \\). All tokenized features are projected through a linear layer. To compute attention between the tokenized features, we stacked two cross-modal attention layers with a self-attention layer in the middle, as illustrated in Figure 2 (d). Cross-modal attention is computed similar to the standard attention mechanism [29]. Given a spatio-temporal token \\( v_{tn} \\) from a video, we compute the attention score to all of the words \\( s_k \\), where \\( k \\in \\{1, \\dots, K\\} \\) in the ASR sen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.637 + ], + "angle": 0, + "content": "tence \\(S\\) by \\(\\alpha_{tnk} = \\frac{\\exp(e_{tnk})}{\\sum_{k=1}^{K}\\exp(e_{tnk})}\\) in the same video clip, where \\(e_{tnk} = \\cosine(v_{tn}, s_k)\\). We then acquire a contextual video token feature \\(\\bar{v}_{tn} = \\sum_{k=1}^{K}\\alpha_{tnk}s_k\\), which encoded text contextual information. Note that the contextual vector is represented by aggregating the representations from the other modality. Follow the standard self-attention computation [49] \\(K, Q, V\\) represent the features for the keys, queries, and values as: \\(Attn(K, Q, V) = \\text{softmax}\\left(\\frac{(Q^\\top K)}{\\sqrt{d_k}}\\right)V\\) where \\(d_k\\) is the dimension of the key. In our case, we feed each contextual feature \\(\\{\\bar{v}_{tn}, \\bar{s}_k\\}\\) right after the first cross-attention layer to the \\(K, Q, V\\) to acquire its self-attended representation. The localized attention model was trained using contrastive loss. To represent the video clip \\(\\mathcal{V}\\) and ASR sentence \\(S\\), we mean-pool over the spatio-temporal tokens in video \\(\\bar{V} = \\frac{1}{TN}\\sum_{r=1}^{TN}\\bar{v}_r\\), and words \\(\\bar{S} = \\frac{1}{K}\\sum_{k=1}^{K}\\bar{s}_k\\) respectively. Let \\((\\bar{V}^{(l)}, \\bar{S}^{(l)})\\) be the \\(l\\)-th training example pair. We adopt the Noise Contrastive Estimation (NCE) loss [17] and the localized attention losses \\(\\mathcal{L}_{Local}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.639, + 0.895, + 0.706 + ], + "angle": 0, + "content": "\\[\n- \\frac {1}{B} \\sum_ {l = 1} ^ {B} \\left[ \\left(\\log \\frac {e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta}}{e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta} + \\sum_ {k = 1} ^ {B} e ^ {\\bar {V} _ {k} ^ {i m p} \\cdot \\bar {S} _ {l}}}\\right) + \\left(\\log \\frac {e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta}}{e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta} + \\sum_ {k = 1} ^ {B} e ^ {\\bar {V} _ {k} \\cdot \\bar {S} _ {k} ^ {i m p}})}\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.895, + 0.75 + ], + "angle": 0, + "content": "where \\(B\\) is the batch. \\(\\bar{V}_k^{imp}\\) and \\(\\bar{S}_k^{imp}\\) represent imposter samples, and \\(\\delta\\) is a margin hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.816, + 0.774 + ], + "angle": 0, + "content": "3.4. Global representations for temporal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We learn the global representation of a video clip and a sentence by contrastive loss, as shown in Figure 2(c). We again use the NCE loss function [17]. The global contrastive loss \\(\\mathcal{L}_{Global}\\) follows the formulation as Equation 1 while using the global representations \\(V\\) and \\(S\\), which are the [CLS] tokens or mean-pooled features from both modalities, instead of the local representations. Projecting the global features to the same space ensures that the features across different" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "modalities are comparable. Since global representations encode information from the entire video, it is essential in encoding temporal information for the later downstream tasks. The final model is optimized by the sum of both losses as \\(\\mathcal{L}_{\\text{Final}} = \\mathcal{L}_{\\text{Local}} + \\mathcal{L}_{\\text{Global}}\\)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.183, + 0.466, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.471, + 0.378 + ], + "angle": 0, + "content": "Figure 3. Spatio-temporal inference. Both representations are used for spatio-temporal grounding: Starting by predicting the action boundary, spatial grounding is performed on the selected frames using the predicted label to find corresponding regions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.39, + 0.427, + 0.406 + ], + "angle": 0, + "content": "3.5. Inference for spatio-temporal grounding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.626 + ], + "angle": 0, + "content": "To perform spatio-temporal grounding on untrimmed videos, we start from temporal action detection as shown in Figure 3. Given a pool of possible action descriptions on the left and an untrimmed video, we perform feature similarity matching using the global representation ([CLS] token or mean-pooled feature) per frame with a threshold \\(\\tau\\) to filter backgrounds. We pick the action class with the largest similarity score per frame. Later, we use the predicted action class and feed it into the local representation branch to compute spatial grounding. We follow [1] to compute feature similarity between visual tokens and text tokens through the cross-attention and self-attention. In the end, we acquire an attention heatmap for later downstream evaluation. More information on inferencing are in appendix 8.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.638, + 0.367, + 0.655 + ], + "angle": 0, + "content": "4. GroundingYoutube Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Current downstream datasets either provide spatial [46], or temporal annotation [28], or spatio-temporal annotation [59] but only for short video clips with few frames before and after the action takes place. These datasets do not provide the opportunity to evaluate both aspects, spatial and temporal grounding, in an untrimmed long video manner. We, therefore, extend one of the current benchmarks, MiningYouTube [28], with 250 untrimmed videos with a duration of around 3-5 minutes and an average of 41.6 labeled action instances per video. Note that each video contains various classes. As the dataset already provides dense temporal annotations, we annotate the respective temporal action segments in the dataset with spatial information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Annotating the spatio-temporal extent of actions can be challenging as there is no clear visible outline as, e.g., in object annotation, nor is there a unique signal to indicate the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.693, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.092, + 0.867, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.223, + 0.692, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.223, + 0.869, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.352, + 0.892, + 0.408 + ], + "angle": 0, + "content": "Figure 4. Visualization of the point annotation and automatic bounding box generation from points. The red point represents the mean of the five annotation points. The points annotation captures diverse patterns in various action types." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.605 + ], + "angle": 0, + "content": "temporal begin and end points. Similarly, grounding systems do not produce pixel-exact bounding boxes, but rather indicate regions of interest. Detector-free spatial grounding models [5] address this fuzziness by relying on pointing game accuracy, thus only using the center point of the heat map for evaluation. Lending on this idea, annotators were asked to mark the presumed center point of the action. Compared to bounding boxes, center point annotation can be advantageous because annotators are not visually distracted by object outlines, so it is more likely that the most important region will be selected. We capture five annotations per frame, resulting in a density-based heatmap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.893, + 0.773 + ], + "angle": 0, + "content": "Starting from 5,091 clips showing one of the 512 action classes, we adopt the methodology used for temporal action localization developed in [15] and label one frame per second, resulting in 26,987 labeled frames. We annotated all frames with five repeats per image, resulting in five annotations per frame and 134,935 point labels in total. To evaluate using bounding boxes [24], we get the union of all annotated points with additional distance to construct the bounding box as shown in Figure 4. More information on the annotation process, bounding box derivation, and dataset analysis is provided in the appendix 10." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.79, + 0.633, + 0.807 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.602, + 0.831 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training Data: HowTo100M dataset contains 1.2M instructional videos along with their corresponding automatically generated speech (ASR). The narrations may be inaccurate and do not always accurately depict the video scene." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18423" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.875, + 0.276 + ], + "angle": 0, + "content": "
MethodBackboneDataSetSupervisionModalityGroundingYoutube
IoU+PointmAP
0.10.20.30.40.50.1:0.5
CoMMA† [46]S3DHT250KSelfVT1.022.181.721.110.930.371.26
MIL-NCE [36]S3D*HT100MSelfVT4.6733.9425.1612.653.420.4115.11
OursS3DHT100MSelfVT9.1242.7035.4925.1616.2210.0525.92
GLIP [31]Swin-L*Cap24MWeakIT1.242.832.101.520.960.371.56
CoMMA‡ [46]CLIPHT100MSelfVT1.683.512.321.880.990.401.82
CLIP [38]CLIPHT100MSelfIT3.5929.5422.159.162.480.3912.74
RegionCLIP [62]ResNet-101*CC3MWeakIT5.6535.6527.4315.694.310.8616.78
OursCLIPHT100MSelfVT10.0942.8136.0525.8417.1011.3526.63
OursCLIP*HT100MSelfVT11.5343.6436.9426.7819.4514.6128.26
MIL-NCE(temp.)+RegionCLIP(spa.)---VT9.2140.5434.9722.3813.799.1822.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.286, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Table 2. Spatio-temporal grounding on GroundingYouTube full videos. The proposed model learns global representations encoding global information and spatial correspondences across modalities, achieving a better performance in spatio-temporal evaluation compared to models trained on only spatial or temporal grounding. (V: video, I: image, T: text.) * indicates finetuned backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.473, + 0.556 + ], + "angle": 0, + "content": "Downstream Datasets: GroundingYoutube (GYT) is used to evaluate the task of multi-action spatio-temporal grounding as described in Section 4. MiningYoutube (MYT) [28] provides temporal annotation and is limited to the domain of cooking instruction videos. YouCook-Interaction (YC-Inter) [46] is an extension of the YouCook2 dataset [64] for cooking instruction providing bounding boxes for 6K selected frames. The bounding boxes usually comprise the hand and the tool mentioned in the respective sentence-wise annotation. To further benchmark on general video domains on the V-HICO dataset [32] with 6.5k videos with human-object interaction bounding boxes annotations, and Daly action dataset [51], featuring videos consisting of daily actions such as \"brushing teeth\"." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.248, + 0.58 + ], + "angle": 0, + "content": "5.2. Baseline methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.47, + 0.769 + ], + "angle": 0, + "content": "Temporal: MIL-NCE [36] utilizes S3D [53] and word2vec [37]. CLIP [38], an image-text model with transformer. Spatial: CoMMA [46], SSL model († for weights shared by the author² ‡ trained with CLIP); GLIP [31], RegionCLIP [62], SOTA weakly supervised grounding model. Spatio-temporal: We construct MIL-NCE+RegionCLIP following the inference pipeline in Figure 3. TubeDETR [54] and STCAT [23] are supervised. More descriptions of the baselines are given in the Appendix 8.1. Details of the implementation and experimental settings can be found in the appendix 8.2. Inference setups for each baseline are described in Section 8.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.78, + 0.26, + 0.794 + ], + "angle": 0, + "content": "5.3. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.469, + 0.847 + ], + "angle": 0, + "content": "We considered the following downstream tasks to evaluate spatio-temporal grounding abilities of various models (detailed description is included in the appendix 8.4):" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.848, + 0.469, + 0.878 + ], + "angle": 0, + "content": "(i) Spatio-temporal grounding in untrimmed video is evaluated on the proposed Grounding Youtube dataset. The" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.344, + 0.892, + 0.617 + ], + "angle": 0, + "content": "entire video and the respective pool of action instructions were provided. The model needs to localize each action step in time (start-time/end-time) and space (location in the video) as described in Figure 3. We evaluate in two metrics: IoU+Pointing game combines the evaluation setting from the spatial grounding [2] and temporal grounding [28] metrics. For each video frame, the prediction is correct when the model predicts the correct action for the frame. Also, given the predicted action as a query, the maximum point of the heatmap aims to lie within the desired bounding box. We then compute the Intersection over Union (IoU) over all the predictions with the GT to acquire the final score. We also compute video mAP following previous evaluation [15], where we set IoU threshold between GT and predicted spatio-temporal tubes. A prediction is correct when it surpasses the IoU threshold. We then compute the mAP over all classes. We form a 3D prediction mask following Figure 3 and compute IoU between our 3D heatmap and 3D tube." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.752 + ], + "angle": 0, + "content": "(ii) Spatial grounding is given a text description to localize the region in the trimmed video. It is evaluated using the pointing game accuracy. If the predicted point lies in the ground truth bounding box, the result counts as a “hit” and counts as “miss” otherwise. The final accuracy is calculated as a ratio between hits to the total number of predictions \\(\\frac{\\# \\text{hits}}{\\# \\text{hits} + \\# \\text{misses}}\\). We also report the mean average precision (mAP) following the settings from V-HICO [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.904 + ], + "angle": 0, + "content": "(iii) Temporal grounding provides videos with the respective actions and their ordering, including the background. The goal is to find the correct frame-wise segmentation of the video. We follow the inference procedure in [28] to compute the alignment given the similarity input matrix. The task is evaluated by intersection over detection (IoD), defined as \\(\\frac{G \\cap D}{D}\\) the ratio between the intersection of ground-truth action \\(G\\) and prediction \\(D\\) to prediction \\(D\\), and the Jaccard index, which is an (IoU) given as \\(\\frac{G \\cap D}{G \\cup D}\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.384, + 0.901 + ], + "angle": 0, + "content": "2We thank the authors for providing code and weights." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18424" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.872, + 0.273 + ], + "angle": 0, + "content": "
MethodBackboneDataSuper.Mod.YC-InterGroundingYTV-HICODaly
AccmAPAccmAPAccmAPAccmAP
MIL-NCE [36]S3D*HT100MSelfVT23.6727.458.2112.6511.2313.8424.23
CoMMA† [46]S3DHT250KSelfVT48.6347.6823.3840.9721.4554.4833.39
OursS3DHT100MSelfVT53.9860.6244.9344.3224.3166.3545.93
CLIP [38]CLIPHT100MSelfIT14.1012.503.4929.2312.5118.0227.28
CoMMA‡ [46]CLIPHT100MSelfVT52.6547.5636.4255.2034.5461.0644.37
RegionCLIP [62]RN50x4*CC3MWeakIT51.5652.8423.4257.9237.8267.1248.62
GLIP [31]Swin-L*Cap24MWeakIT52.8453.6224.7366.0541.17--
OursCLIPHT100MSelfVT57.1055.4943.1260.7139.2870.0850.56
OursCLIP*HT100MSelfVT58.3556.9845.3262.3441.5671.3552.78
TubeDETR [54]MDETRVid-STGFullVT51.6353.2441.7663.2340.8784.2162.98
STCAT [23]ResNet-101Vid-STGFullVT54.4755.9044.2165.3441.1085.4263.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.277, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Table 3. Video spatial grounding. We evaluate the accuracy of the pointing game and the mean average precision. We listed CNN-based methods on top and transformer-based methods in the middle. Models learning global representations (MIL-NCE, CLIP) don't perform well on localization tasks, while our model outperforms other grounding methods. * indicates finetuned backbone." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.326, + 0.47, + 0.46 + ], + "angle": 0, + "content": "
MethodBackboneDataSuper.IoUIoD
Mining: MLP [28]TSMMiningYTWeak9.8019.20
CoMMA* [46]S3DHT250KSelf2.055.63
RegionCLIP [62]RN50x4*CC3MWeak10.9616.86
MIL-NCE [36]S3D*HT100MSelf18.6926.74
OursS3DHT100MSelf19.1827.65
OursCLIPHT100MSelf19.8828.50
OursCLIP*HT100MSelf20.3329.67
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.533 + ], + "angle": 0, + "content": "Table 4. Temporal Grounding on MiningYoutube. * indicates finetuned backbone. Spatial-focused model CoMMA is not trained for temporal detection, which results in lower performance. In contrast, the proposed model combines global and local representation, resulting in better temporal localization than one alone." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.436, + 0.56 + ], + "angle": 0, + "content": "5.4. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.854 + ], + "angle": 0, + "content": "(i) Spatio-temporal grounding in untrimmed video: We first compare the proposed method with other approaches designed for spatial or temporal grounding in Table 2. It shows that models without specific loss designs for spatial grounding (MIL-NCE [36], CLIP [38]) show good mAP scores but lower pointing game accuracy. Out of the two weakly supervised methods, GLIP [31] and RegionCLIP [62]), trained with aligned image-text, RegionCLIP show significantly better performance in this setting, while both perform in a similar range in the spatial grounding scenario (see Table 3). We attribute this behavior to the fact that RegionCLIP distinguishes frames with relevant queries better from background than GLIP, leading to better temporal localization. We finally compare the strong baseline MIL-NCE+RegionCLIP, which combines two approaches specialized in temporal and spatial aspects, to our task. It shows that the proposed method improves over all other baselines underlining the need to incorporate global (temporal) and local (spatial) representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "(ii) Spatial grounding: Second, we compare the performance of the proposed framework to other methods on the task of spatial grounding, including models with weak su" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.539 + ], + "angle": 0, + "content": "persion, as well as models trained in a fully supervised setting in Table 3. In the instruction video domain (GYT and YC-Inter), the proposed approach achieves the best result among all weakly and self-supervised trained methods. In the general domain (V-HICO and Daly), the method also achieves competitive results, showing the generalizability of the model to other domains. Note that in the Daly dataset, the classes are verbs, which are not detectable by the object-focused model GLIP. Compared to their weakly trained counterparts, fully-supervised model (TubeDETER [54], STCAT [23]) achieve competitive performance in the general domain (V-HICO, Daly) and slightly lower performance in instruction domain (GYT, YC-Inter) due to the domain gap with respect to the training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.632 + ], + "angle": 0, + "content": "(iii) Temporal grounding: We evaluate temporal grounding in Table 4. Here, it shows that global representations also profit from local representation learning. This hypothesis is further validated in the ablation studies in Table 6, where we ablate both losses for all three settings and show a consistent improvement in the joint loss formulation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.647, + 0.651, + 0.664 + ], + "angle": 0, + "content": "5.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We perform ablation studies with respect to all three settings, spatio-temporal grounding, as well as spatial and temporal grounding alone, reporting performance for spatiotemporal grounding on GroundingYT using mAP with IoU@0.4, on temporal grounding using MiningYT IoU, and on spatial grounding using YC-Inter. pointing game. Additional ablation are in appendix 9.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Frame selection strategy. We perform an ablation on the possible frame selection strategies for our method (Figure 2(b) and Section 3.2). In Table 5, None uses all frames within the ASR boundary (\\(U = T\\)) as our video training data. Global represents the [CLS] token in text and video. Local uses the words and spatio-temporal tokens. In the setting Sinkhorn was not applied, the top \\(T\\) frames with the highest similarity score were selected. When we set" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18425" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.875, + 0.203 + ], + "angle": 0, + "content": "
GroundingYT Spatio-temporalMiningYT TemporalYC-Inter. SpatialGroundingYT Spatio-temporalMiningYT TemporalYC-Inter. Spatial
w/o Sinkhornwith Sinkhorn
None17.4318.3457.42---
Global T - Global V18.2419.3856.3118.9619.8957.34
Global T - Local V18.0119.3157.5618.5319.3557.67
Local T - Global V18.0519.8557.3419.3220.3658.13
Local T - Local V18.3119.4857.8619.4320.1658.51
Average over last two18.3619.6857.7719.4520.3358.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.207, + 0.895, + 0.251 + ], + "angle": 0, + "content": "Table 5. Frame selection: (a) Sinkhorn selection results in better supervision. (b) We investigate all possible combinations of global and local representations for frame selection similarity matching. We found keeping the local text representation is crucial, and a combination of local and global representation leads to the best spatio-temporal grounding result." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.263, + 0.303, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.263, + 0.495, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.263, + 0.683, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.263, + 0.875, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.893, + 0.423 + ], + "angle": 0, + "content": "Figure 5. Visualization on GroundingYoutube. Red box: annotation. Heatmap: prediction. Baseline GLIP focuses on objects such as power, mixer, and pancake. Our method focuses on actions such as whisk and serve." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.436, + 0.47, + 0.504 + ], + "angle": 0, + "content": "
GroundingYT Spatio-tempMiningYT TemporalYC-Inter. Spatial
only Local loss7.295.2355.29
only Global loss9.2819.1236.23
w/ Both loss19.4520.3358.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.508, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Table 6. Loss ablations: both losses contribute to the final loss, and the existence of global loss helps the localization task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.734 + ], + "angle": 0, + "content": "spatio-temporal tokens as the selection target, we sum over the scores with respect to each frame to acquire the frame similarity score. It shows that selecting frames based on Sinkhorn selection leads to consistently better results as it enforces more variety of visual concepts but also captures frames with possible groundable objects. It further shows that word tokens are more suitable than the global text CLS token for frame selection. Finally, we see that depending on the task (spatial vs. temporal), a local resp. global representation is better, and a combination of both works best for spatio-temporal grounding. We provide runtime analysis of such frame selection strategy in the appendix 9.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Global and local loss. As mentioned in the spatio-temporal evaluation, both features contribute to the final grounding result. We test the model by ablating out each loss. Table 6 shows that each loss not only contributes to the spatiotemporal grounding on the GYT, but also that the whole is more than the sum of its parts (losses) since this task requires both spatial and temporal detection. The reduced impact of the global loss in the case of YC-Inter is that this is a pure spatial grounding dataset (no background frames) without temporal detection, and the local loss plays a more critical role. We observe the same patterns in the temporal" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.45, + 0.893, + 0.496 + ], + "angle": 0, + "content": "grounding result for MYT, where spatial localization is not directly contributing to the final performance. We tried out the same ablation using in the S3D backbone in supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.508, + 0.678, + 0.525 + ], + "angle": 0, + "content": "5.6. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.893, + 0.605 + ], + "angle": 0, + "content": "We visualize our spatio-temporal result in Figure 5. For the GLIP model, we output the bounding box with the highest confidence score and visualize its center point. We found GLIP model focuses on the salient object while our model focuses more on human-object interaction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.619, + 0.639 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.893, + 0.84 + ], + "angle": 0, + "content": "We presented an approach for learning spatio-temporal grounding with self-supervision and a new dataset: GroundingYoutube annotations, where we densely annotate spatiotemporal points/boxes from untrimmed multi-action videos. Our method includes a frame selection mechanism that identifies frames with groundable objects to adapt the learning process for untrimmed videos. Furthermore, we jointly learn global representations, which capture temporal information, and local representations learning fine-grained multimodal interactions between video and text. We conducted extensive experiments and our approach shows state-of-the-art performance in spatio-temporal grounding, as well as temporal and spatial grounding alone." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.843, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: Nina Shvetsova is supported by the German Federal Ministry of Education and Research (BMBF) project STCL-01IS22067. Andrew Rouditchenko is supported by an NDSEG Fellowship and by the MIT-IBM Watson AI Lab. We thank the MIT-Satori cluster team for their technical support." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18426" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Hassan Akbari, Svebor Karaman, Surabhi Bhargava, Brian Chen, Carl Vondrick, and Shih-Fu Chang. Multi-level multimodal common semantic space for image-phrase grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12476–12486, 2019. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. Advances in Neural Information Processing Systems, 33:25-37, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[5] Assaf Arbelle, Sivan Doveh, Amit Alfassy, Joseph Shtok, Guy Lev, Eli Schwartz, Hilde Kuehne, Hila Barak Levi, Prasanna Sattigeri, Rameswar Panda, et al. Detector-free weakly supervised grounding by separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1801-1812, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[8] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8012-8021, 2021. 1, 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[9] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. Advances in Neural Information Processing Systems, 34:28442-28453, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[10] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee Kenneth Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1884-1894, Florence, Italy, 2019. Association for Computational Linguistics. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. 2013. 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[13] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[15] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 2, 5, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.389, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[17] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTATS, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[19] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649–665, 2018. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[20] H. Idrees, A. R. Zamir, Y. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[21] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[22] Yu-Gang Jiang, Jingen Liu, A Roshan Zamir, George Toderici, Ivan Laptev, Mubarak Shah, and Rahul Sukthankar. Thumos challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.3,14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[23] Yang Jin, yongzhi li, Zehuan Yuan, and Yadong MU. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In Advances in Neural Information Processing Systems, 2022. 1, 2, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In Proceedings of the IEEE" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18427" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 4405-4413, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2015. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[27] Giorgos Kordopatis-Zilos, Symeon Papadopoulos, Ioannis Patras, and Ioannis Kompatsiaris. Visil: Fine-grained spatiotemporal video similarity learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6351-6360, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[28] Hilde Kuehne, Ahsan Iqbal, Alexander Richard, and Juergen Gall. Mining youtube-a dataset for learning fine-grained action concepts from webly supervised video data. 2019. 1, 2, 3, 5, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[29] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[30] Ang Li, Meghan Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv preprint arXiv:2005.00214, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[31] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 6, 7, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1845-1855, 2021. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[33] Yikang Li, Jenhao Hsiao, and Chiuman Ho. Videoclip: A cross-attention model for fast video-text retrieval task with image clip. In Proceedings of the 2022 International Conference on Multimedia Retrieval, pages 29–33, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[35] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 3, 6, 7, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In arXiv preprint arXiv:1301.3781, 2013. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[40] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv preprint arXiv:2006.09199, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[41] Jing Shi, Jia Xu, Boqing Gong, and Chenliang Xu. Not all frames are equal: Weakly-supervised video grounding with contextual similarity and visual clustering losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10444-10452, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[42] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In CVPR, 2022. 1, 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[43] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba Heilbron, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. arXiv preprint arXiv:2112.00431, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[45] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1533–1542, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[46] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5, 6, 7, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[47] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[51] Philippe Weinzaepfel, Xavier Martin, and Cordelia Schmid. Human action localization with sparse spatial supervision. arXiv preprint arXiv:1605.05197, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[52] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 1, 2, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[57] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[58] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[59] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10668-10677, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[60] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[61] Yang Zhao, Zhou Zhao, Zhu Zhang, and Zhijie Lin. Cascaded prediction network via segment tree for temporal video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4206, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[62] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2, 3, 6, 7, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[63] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In British Machine Vision Conference, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[65] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[66] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. 2019. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18429" + } + ] +] \ No newline at end of file diff --git a/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_origin.pdf b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0fbb3fc36aa067447a2a3ccfb252024fdedd9add --- /dev/null +++ b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/59de70a6-beec-4552-9d70-f207c0df32db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d76837df307c8d3749f30dac79191f5020ec033bf3bf872557f9462e1c01d1b +size 8422111 diff --git a/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/full.md b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d737c294cbf5917e84b13f8584e0711f19febd62 --- /dev/null +++ b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/full.md @@ -0,0 +1,296 @@ +# What, when, and where? Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions + +Brian Chen1 Nina Shvetsova2,3 Andrew Rouditchenko4 Daniel Kondermann5 Samuel Thomas6,7 Shih-Fu Chang1 Rogerio Feris6,7 James Glass4 Hilde Kuehne2,3,7 + +1Columbia University, 2Goethe University Frankfurt, 3University of Bonn, 4MIT CSAIL + +$^{5}$ Quality Match GmbH, $^{6}$ IBM Research AI, $^{7}$ MIT-IBM Watson AI Lab + +{bc2754,sc250}@columbia.edu, shvetsov@uni-frankfurt.de,{roudi,glass}@mit.edu + +dk@quality-match.com, {sthomas,rsferis}@us.ibm.com, kuehne@cs.uni-bonn.de + +# Abstract + +Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed, providing dense spatio-temporal grounding annotations in long, untrimmed, multi-action instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks, showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatio-temporal grounding. + +# 1. Introduction + +Spatio-temporal grounding (STG) describes the challenging task of locating events in space and time within video data based on text referential expressions. Methods in this field usually rely on a combination of spatio-temporal bounding box annotation, together with a human-generated caption, describing the visual content of the bounding box [23, 54], which limits their generalizability beyond the given training scenario. Compared to that, as a second line of work, multimodal self-supervised learning tries to leverage "free" data sources, such as video and automatic speech + +![](images/bb12e59d13f4edd0b4b3b2abe2b2049246c4944badb05195208fa2e9e9672005.jpg) +"To this we will add one carrot chopped into jong jullien ..." + +![](images/e5611b2857d6b84158d8140794397cffedcd25228d13a515e01c18410f1b5974.jpg) +"... you can use the forks to pull it a bit apart..." + +![](images/a7af3094e5fd67b80fbc50718020f8b7cfe8285cc758d1da4c2a35d9cfaea622.jpg) +Training Setup: Unlabeled videos with narrated instructions + +![](images/626eb5a1b84be14f724053fa934bb1ec474cbccb20c20844779fa842710513ad.jpg) +① +...stir it a bit so it mixes well... + +![](images/c3d470eff27720f396457fd2b5344d18b21068a86bff41251a9448107517375a.jpg) +Evaluation Setup: Referential queries - "Crack egg", "Mix egg", etc. + +![](images/4139cd564a8441e6094cbd80fa13d3f0cc4bf4bb34ab953847a529fc4dd97591.jpg) +Task: Spatio-Temporal Grounding - Find the temporal boundary of a queried action in an untrimmed video and spatially localize the action. +Figure 1. Learning Spatio-temporal grounding in untrimmed videos. In training, we learn from unlabeled videos without human annotation. In evaluation, we perform spatio-temporal grounding using an action description such as "crack egg" as a query. The model needs to localize both the action's temporal boundary and spatial region in the long untrimmed video. We visualize the heatmap from the annotation points as well as derived bounding boxes. + +![](images/9749226b99a0a6a93abf15403719bf859baa2a436dd7922fac2da728763b02a4.jpg) + +recognition (ASR) captions from large-scale instructional videos to learn representations without human annotation [3, 4, 8, 35, 36]. The resulting models achieve state-of-the-art performance on zero-shot tasks such as cross-modal video retrieval or classification and also for zero-shot temporal action segmentation and detection based on free text queries [8, 28, 42, 47, 66], but usually lack spatial localization abilities. A third line of work focuses on label-free spatial grounding, e.g. by training on image-caption + +[2, 31, 50, 55, 62] or video-caption pairs [41, 46]. The goal is to correctly localize a referential expression in an image or each video frame, e.g., via a bounding box or a heatmap. However, those methods are not optimized to detect whether an event is present in a video. The assumption is thus that the evaluated expression is visible in the image or in all video frames. + +The following work aims to bring together those ideas to address the task of spatio-temporal action grounding from multimodal supervision in untrimmed videos. We propose a grounding approach that uses video-text pairs based on ASR transcripts in instructional videos and learns the spatial representation of free-text events as well as their temporal extent, as shown in Figure 1. To this end, we leverage two different representations of the visual data: a global feature representation based on full-frame information to define the temporal extent of an event and a local representation based on frame-wise grid features for spatial localization. The motivation for this dualism is that while the local representation captures the spatial correlations between vision and text input, this can be too fine-grained to learn a holistic representation of the frame, while the global representation can be assumed to capture a more compact, aggregated view compared to local data and thus to provide a more reliable cue for the task of temporal localization. However, compared to the hand-annotated video-caption setup of most spatio-temporal grounding methods, the ASR text can be noisy as not all comments refer to visible events. Further, as there is only a loose temporal correlation, the described activities might not be precisely aligned, can be scattered over multiple frames, or not be present at all [18, 36]. Therefore, we propose to specifically select frames to capture only those useful for training. To this end, we look for frames that match the vocabulary of the respective text, leveraging a selection strategy by Sinkhorn optimal transport [11]. This allows us to train a model that can localize actions in space and time within videos without labeling supervision. + +To evaluate spatio-temporal grounding in untrimmed videos, a new benchmark, GroundingYouTube, is proposed. It is based on the existing MiningYouTube dataset [28] and extended with spatio-temporal localization information. This setup differs from other benchmarks such as [10, 46, 63] in two ways: first, by using multiple center point annotations, it focuses on the grounding of referential actions itself instead of interacting humans or objects which are usually labeled; second, the dense annotations of multiple actions in the video allow us to benchmark action grounding in long, realistic untrimmed videos compared to existing, often pre-clipped benchmarks[10, 59]. The benchmark provides queries for 512 different event types and over $5K$ spatio-temporal annotations, as shown in Figure 1. A comparison of current datasets is shown in Table 1. + +To evaluate the proposed approach as well as the new + +
DatasetAnnotation
SpatialTemporal
V-HiCoobject bb + human bb-
AVA-Kineticsobject bb + human bb-
ActivityNet entitiesobject bb + human bb-
THUMOS14-action boundaries
ActivityNet-action boundaries
HACSSegment-action boundaries
YouCook II-multi-action boundaries
Cross-Task-multi-action boundaries
COIN-multi-action boundaries
EPIC KITCHENS-100-multi-action boundaries
Ego4D-multi-action boundaries
JHMDB51-21human tubes-
UCF101-24human tubesaction boundaries
Dalyhuman tubesaction boundaries
Vid-STGhuman tubesaction boundaries
HC-STVGhuman tubesaction boundaries
AVAhuman tubesaction boundaries
YouCook-Interactionsaction bb-
GroundingYoutube (ours)action bb + center pointsmulti-action boundaries
+ +Table 1. Comparison of spatial, temporal, and spatio-temporal grounding datasets. V-HiCo [32], AVA-Kinetics [30], THU-MOS14 [20], ActivityNet [6], ActivityNet entity [65], HACSSegment [60], YouCook II [64], Cross-Task [66], COIN [47], EPIC KITCHENS-100 [12], Ego4D [14], JHMDB51-21 [21], UCF101-24 [44] Daly [51], Vid-STG [59], HC-STVG [48], AVA [15], YouCook-Interactions [46]. + +benchmark, the system is trained on the HowTo100M dataset [35] and compared to state-of-the-art methods based on full, weak, and self-supervision for spatial and temporal, as well as combined spatio-temporal grounding tasks. It shows that existing methods usually do well in one of the two aspects, spatial or temporal grounding. In contrast, the proposed method can combine spatial and temporal aspects. + +We summarize the contributions of this work as follows1: (1) We propose a framework for spatio-temporal grounding in untrimmed videos based on weakly aligned multimodal supervision without human annotation, employing a combination of global and local representation learning to learn the spatio-temporal extent of actions. (2) To facilitate this task, we propose a frame selection strategy based on Sinkhorn-Knopp Optimal transport that improves the quality of the acquired learning samples, leading to more effective supervision. (3) We provide a new benchmark and annotations to evaluate this challenging problem on real-world multi-action instructional video data. + +# 2. Related Work + +Supervised Spatio-temporal Grounding. Spatiotemporal Grounding refers to the problem of localizing a sequence of bounding boxes (a spatio-temporal tube) for a target object described by an input text. This problem has been addressed by various approaches TubeDETR [54], STCAT [23], STVGBert [45], STVGBert [45], + +STGVT [48], STGRN [59], Visil [27]. These methods rely on proposal networks such as Faster R-CNN [39] or MDETR [25] to predict bounding box coordinates for learning text-to-region interaction. All those approaches rely on supervised training with the human-annotated sentence and bounding box supervision, provided, e.g., by datasets such as VidSTG [59], HC-STVG [10]. While those datasets provide a temporal aspect, temporal detection is usually limited to identifying the start and end frame of a single action in a video. Compared to that, an untrimmed setting usually comprises multiple actions in a longer video that can be separated by longer background sequences. This conceptually differs from previous works [10] that typically use short videos of around 5-10 seconds. Other datasets such as ActivityNet entities [65] provide only bounding boxes for noun phrases in the captions, namely the objects, which is related to object detection task and does not capture any spatial or temporal extent of actions. + +Multimodal Self-supervised Learning. The field of multimodal self-supervised learning aims to learn data representations by leveraging large amounts of unlabeled data with multiple modalities. Early works [13, 52] started by projecting images and text into a joint visual-language embedding space, where embeddings of semantically similar pairs are close. Those ideas have now grown into systems such as MIL-NCE [36] using the HowTo100M dataset [35] to train a video-language embedding space from 1.2 million instructional videos paired with text descriptions from ASR. Follow-up works, including [3, 4, 8, 40, 42] show that using videos without annotation enables an effective multimodal embedding space via contrastive learning. + +Based on those advantages, approaches started to address the problem of Spatial Video Grounding from multimodal self-supervised aiming to identify spatial locations in a trimmed video based on text descriptions without the need for bounding box annotation during training. One of the early works studied this task in the context of weakly supervised learning where we learn grounding with human-annotated captions of the video [63]. In this context, works [41, 46] have focused on object grounding benchmarks such as YouCook2-BoundingBox [64], which provides bounding box annotations for visible objects in cooking videos. Other works such as GLIP [31], Region-CLIP [62], and others [56, 58] combine the principles of large-scale vision language training with bounding box finetuning on object detection datasets [16, 34]. Recently, the YouCook-Interactions dataset [46] and CoMMA [46] have been proposed for the spatial grounding of objects and actions with multimodal self-supervision from HowTo100M videos. These works assume that the video is temporally clipped with respect to the grounding phrase. + +Compared to that, Temporal Video Grounding aims to determine the set of consecutive frames corresponding to a + +text query in an untrimmed video [9, 22, 43], thus predicting temporal boundaries of action instances. Recent work such as MIL-NCE [36], MCN [8], and VideoCLIP [33] utilize large-scale pretraining for grounding actions temporally via text-to-frame similarity on video datasets such as MiningYouTube [28] or CrossTask [66] without proposals. However, these methods lack spatial localization ability [57, 61]. + +# 3. A Global-Local Framework for Spatio-Temporal Grounding + +# 3.1. General setup + +The goal of the proposed method is to temporally and spatially localize actions based on free-text queries in untrimmed videos. To this end, two representations are learned, a local and a global one. We start with narrated video clips, each associated with a corresponding visual representation and text narration. Namely, for each clip $\mathcal{X} = \{\mathcal{V},\mathcal{S}\}$ , let $\mathcal{V}$ stand for the video clip and $\mathcal{S}$ for the text narration sentence generated by the automatic speech recognition (ASR) system. Each clip $\mathcal{V}$ consists of $U\times N$ spatio-temporal tokens $\{v_{u,n}\}$ , where $u\in \{1,\dots,U\}$ represents the number of frames in the video and $n\in \{1,\dots,N\}$ represents the number of spatial grid region tokens or features in a frame. The text sentence $\mathcal{S}$ consists of $K$ words $\{s_1,\ldots ,s_K\}$ . We represent localized features by the tokens from each modality, and the global features $\{V,S\}$ are acquired either by mean-pooling over the local features (S3D) or by using the [CLS] token from the transformer (CLIP) as in Radford et al. [38]. We learn transformations $f:V\to \mathbb{R}^d$ to a $d$ -dimensional representation $f(V)\in \mathbb{R}^{d}$ from the global representation $V$ , and $g:S\rightarrow \mathbb{R}^d$ , to produce similar $d$ -dimensional text global embeddings: $g(S)\in \mathbb{R}^{d}$ . Similar to $\{f,g\}$ , we note $\{f',g'\}$ to be the transform for localized features, where local features $\{v,s\}$ are also projected as $d$ -dimensional representations. + +# 3.2. Representation guided frame sampling + +Learning from multimodal self-supervision is challenging since the narration is likely to be noisy, thus containing more information than the actual task descriptions due to poor temporal alignment or cut scenes [18], which is the key differences between weakly supervised vision-caption grounding and multimodal self-supervised grounding. This work pursues a frame selection strategy to improve object grounding and temporal alignment during training. We start from a longer sequence $U$ , where $U > T$ , which includes the video frames before and after the ASR boundaries that could contain actions or objects in the sentence. Our goal is to find $T$ frames out of the $U$ frames that are most relevant to the actions and objects in the sentence $S$ . We formalize it as an optimal transport problem utilizing the Sinkhorn-Knopp algorithm [11]. + +![](images/49a7b57db49a9fad34daef3db334c0b00e06c4e67b23e2a509527d564c1166bb.jpg) +Figure 2. Spatio-temporal grounding approach. (a) We aim to select frames with groundable objects and actions. To this end, projected text features are matched with respective frame features. (b) Sinkhorn optimal transport is then leveraged to optimize the selected frames wrt. the text input. (c) Based on the selected frames, a global representation is learned to allow for temporal localization, as well as (d) a local representation to ground the action description in the spatial region. + +Optimal transport for text-to-frame assignment. To acquire the optimal assignment from word features to video frames, an assignment matrix $\mathbf{Q}$ is computed from each video and ASR pair as shown in Figure 2(a). This cross-model optimal transport mechanism is applied to assignment $\mathbf{Q}$ from the projected cross-model similarity $\mathbf{P}$ between word tokens and each video frame, where $\mathbf{P} = g(\mathcal{S})\otimes f(\mathcal{V})^{\top}\in \mathbb{R}^{K\times U}$ . To compute the assignment matrix, the text and video projection layers from the global representation in Figure 2(c) are used to project multimodal features into a common space for feature similarity calculation. We investigate various granularities of the features where we compute the similarity between the text features at the word (local) / sentence (global) level and the visual feature at frames (global) / spatiotemporal tokens (local) level to acquire $\mathbf{P}$ , as shown in Table 5. To ensure that the word-to-frame assignment contains more diversity instead of just saturated assignments to a single video frame, we add a constraint by Sinkhorn that requires label assignments to be equally distributed across various video frames representing diverse object/action concepts. Details of the Sinkhorn optimal transport are included in the appendix 7.1. + +# 3.3. Local representations for spatial localization + +To capture multimodal interaction with finer granularity, we need to learn the projection between tokenized features as shown in Figure 2(d). We extract spatio-temporal region features $v_{tn}$ from the video. Also, we extract word features $s_k$ which represents the feature from word $k$ . All tokenized features are projected through a linear layer. To compute attention between the tokenized features, we stacked two cross-modal attention layers with a self-attention layer in the middle, as illustrated in Figure 2 (d). Cross-modal attention is computed similar to the standard attention mechanism [29]. Given a spatio-temporal token $v_{tn}$ from a video, we compute the attention score to all of the words $s_k$ , where $k \in \{1, \dots, K\}$ in the ASR sen + +tence $S$ by $\alpha_{tnk} = \frac{\exp(e_{tnk})}{\sum_{k=1}^{K}\exp(e_{tnk})}$ in the same video clip, where $e_{tnk} = \cosine(v_{tn}, s_k)$ . We then acquire a contextual video token feature $\bar{v}_{tn} = \sum_{k=1}^{K}\alpha_{tnk}s_k$ , which encoded text contextual information. Note that the contextual vector is represented by aggregating the representations from the other modality. Follow the standard self-attention computation [49] $K, Q, V$ represent the features for the keys, queries, and values as: $Attn(K, Q, V) = \text{softmax}\left(\frac{(Q^\top K)}{\sqrt{d_k}}\right)V$ where $d_k$ is the dimension of the key. In our case, we feed each contextual feature $\{\bar{v}_{tn}, \bar{s}_k\}$ right after the first cross-attention layer to the $K, Q, V$ to acquire its self-attended representation. The localized attention model was trained using contrastive loss. To represent the video clip $\mathcal{V}$ and ASR sentence $S$ , we mean-pool over the spatio-temporal tokens in video $\bar{V} = \frac{1}{TN}\sum_{r=1}^{TN}\bar{v}_r$ , and words $\bar{S} = \frac{1}{K}\sum_{k=1}^{K}\bar{s}_k$ respectively. Let $(\bar{V}^{(l)}, \bar{S}^{(l)})$ be the $l$ -th training example pair. We adopt the Noise Contrastive Estimation (NCE) loss [17] and the localized attention losses $\mathcal{L}_{Local}$ : + +$$ +- \frac {1}{B} \sum_ {l = 1} ^ {B} \left[ \left(\log \frac {e ^ {\bar {V} _ {l} \cdot \bar {S} _ {l} - \delta}}{e ^ {\bar {V} _ {l} \cdot \bar {S} _ {l} - \delta} + \sum_ {k = 1} ^ {B} e ^ {\bar {V} _ {k} ^ {i m p} \cdot \bar {S} _ {l}}}\right) + \left(\log \frac {e ^ {\bar {V} _ {l} \cdot \bar {S} _ {l} - \delta}}{e ^ {\bar {V} _ {l} \cdot \bar {S} _ {l} - \delta} + \sum_ {k = 1} ^ {B} e ^ {\bar {V} _ {k} \cdot \bar {S} _ {k} ^ {i m p}})}\right) \right] \tag {1} +$$ + +where $B$ is the batch. $\bar{V}_k^{imp}$ and $\bar{S}_k^{imp}$ represent imposter samples, and $\delta$ is a margin hyperparameter. + +# 3.4. Global representations for temporal + +We learn the global representation of a video clip and a sentence by contrastive loss, as shown in Figure 2(c). We again use the NCE loss function [17]. The global contrastive loss $\mathcal{L}_{Global}$ follows the formulation as Equation 1 while using the global representations $V$ and $S$ , which are the [CLS] tokens or mean-pooled features from both modalities, instead of the local representations. Projecting the global features to the same space ensures that the features across different + +modalities are comparable. Since global representations encode information from the entire video, it is essential in encoding temporal information for the later downstream tasks. The final model is optimized by the sum of both losses as $\mathcal{L}_{\text{Final}} = \mathcal{L}_{\text{Local}} + \mathcal{L}_{\text{Global}}$ . + +![](images/b207eab4a8d3fcbd28b955ccdffc923eb4ab598946c7d7a9d16906852e7ad8d1.jpg) +Figure 3. Spatio-temporal inference. Both representations are used for spatio-temporal grounding: Starting by predicting the action boundary, spatial grounding is performed on the selected frames using the predicted label to find corresponding regions. + +# 3.5. Inference for spatio-temporal grounding. + +To perform spatio-temporal grounding on untrimmed videos, we start from temporal action detection as shown in Figure 3. Given a pool of possible action descriptions on the left and an untrimmed video, we perform feature similarity matching using the global representation ([CLS] token or mean-pooled feature) per frame with a threshold $\tau$ to filter backgrounds. We pick the action class with the largest similarity score per frame. Later, we use the predicted action class and feed it into the local representation branch to compute spatial grounding. We follow [1] to compute feature similarity between visual tokens and text tokens through the cross-attention and self-attention. In the end, we acquire an attention heatmap for later downstream evaluation. More information on inferencing are in appendix 8.3. + +# 4. GroundingYoutube Benchmark + +Current downstream datasets either provide spatial [46], or temporal annotation [28], or spatio-temporal annotation [59] but only for short video clips with few frames before and after the action takes place. These datasets do not provide the opportunity to evaluate both aspects, spatial and temporal grounding, in an untrimmed long video manner. We, therefore, extend one of the current benchmarks, MiningYouTube [28], with 250 untrimmed videos with a duration of around 3-5 minutes and an average of 41.6 labeled action instances per video. Note that each video contains various classes. As the dataset already provides dense temporal annotations, we annotate the respective temporal action segments in the dataset with spatial information. + +Annotating the spatio-temporal extent of actions can be challenging as there is no clear visible outline as, e.g., in object annotation, nor is there a unique signal to indicate the + +![](images/8c03b6fbed03d0fab120053dfcf58d99f47add551043fcf3b267631841f87d9a.jpg) + +![](images/30a57a3f217ab0dd9e1e39f780a2de1e0270101b09e6bb988983b72b4bdee44b.jpg) + +![](images/ad091c8deff9164b3ab0688ace8c1714f4a8cead98b6f0cfb158045a96bf319c.jpg) +Figure 4. Visualization of the point annotation and automatic bounding box generation from points. The red point represents the mean of the five annotation points. The points annotation captures diverse patterns in various action types. + +![](images/0086a40ad8f7bc800a6bb2e54ada527117c024fc79a7131d2599bd7573ca4827.jpg) + +temporal begin and end points. Similarly, grounding systems do not produce pixel-exact bounding boxes, but rather indicate regions of interest. Detector-free spatial grounding models [5] address this fuzziness by relying on pointing game accuracy, thus only using the center point of the heat map for evaluation. Lending on this idea, annotators were asked to mark the presumed center point of the action. Compared to bounding boxes, center point annotation can be advantageous because annotators are not visually distracted by object outlines, so it is more likely that the most important region will be selected. We capture five annotations per frame, resulting in a density-based heatmap. + +Starting from 5,091 clips showing one of the 512 action classes, we adopt the methodology used for temporal action localization developed in [15] and label one frame per second, resulting in 26,987 labeled frames. We annotated all frames with five repeats per image, resulting in five annotations per frame and 134,935 point labels in total. To evaluate using bounding boxes [24], we get the union of all annotated points with additional distance to construct the bounding box as shown in Figure 4. More information on the annotation process, bounding box derivation, and dataset analysis is provided in the appendix 10. + +# 5. Experiments + +# 5.1. Datasets + +Training Data: HowTo100M dataset contains 1.2M instructional videos along with their corresponding automatically generated speech (ASR). The narrations may be inaccurate and do not always accurately depict the video scene. + +
MethodBackboneDataSetSupervisionModalityGroundingYoutube
IoU+PointmAP
0.10.20.30.40.50.1:0.5
CoMMA† [46]S3DHT250KSelfVT1.022.181.721.110.930.371.26
MIL-NCE [36]S3D*HT100MSelfVT4.6733.9425.1612.653.420.4115.11
OursS3DHT100MSelfVT9.1242.7035.4925.1616.2210.0525.92
GLIP [31]Swin-L*Cap24MWeakIT1.242.832.101.520.960.371.56
CoMMA‡ [46]CLIPHT100MSelfVT1.683.512.321.880.990.401.82
CLIP [38]CLIPHT100MSelfIT3.5929.5422.159.162.480.3912.74
RegionCLIP [62]ResNet-101*CC3MWeakIT5.6535.6527.4315.694.310.8616.78
OursCLIPHT100MSelfVT10.0942.8136.0525.8417.1011.3526.63
OursCLIP*HT100MSelfVT11.5343.6436.9426.7819.4514.6128.26
MIL-NCE(temp.)+RegionCLIP(spa.)---VT9.2140.5434.9722.3813.799.1822.33
+ +Table 2. Spatio-temporal grounding on GroundingYouTube full videos. The proposed model learns global representations encoding global information and spatial correspondences across modalities, achieving a better performance in spatio-temporal evaluation compared to models trained on only spatial or temporal grounding. (V: video, I: image, T: text.) * indicates finetuned backbone. + +Downstream Datasets: GroundingYoutube (GYT) is used to evaluate the task of multi-action spatio-temporal grounding as described in Section 4. MiningYoutube (MYT) [28] provides temporal annotation and is limited to the domain of cooking instruction videos. YouCook-Interaction (YC-Inter) [46] is an extension of the YouCook2 dataset [64] for cooking instruction providing bounding boxes for 6K selected frames. The bounding boxes usually comprise the hand and the tool mentioned in the respective sentence-wise annotation. To further benchmark on general video domains on the V-HICO dataset [32] with 6.5k videos with human-object interaction bounding boxes annotations, and Daly action dataset [51], featuring videos consisting of daily actions such as "brushing teeth". + +# 5.2. Baseline methods + +Temporal: MIL-NCE [36] utilizes S3D [53] and word2vec [37]. CLIP [38], an image-text model with transformer. Spatial: CoMMA [46], SSL model († for weights shared by the author² ‡ trained with CLIP); GLIP [31], RegionCLIP [62], SOTA weakly supervised grounding model. Spatio-temporal: We construct MIL-NCE+RegionCLIP following the inference pipeline in Figure 3. TubeDETR [54] and STCAT [23] are supervised. More descriptions of the baselines are given in the Appendix 8.1. Details of the implementation and experimental settings can be found in the appendix 8.2. Inference setups for each baseline are described in Section 8.3. + +# 5.3. Downstream Tasks + +We considered the following downstream tasks to evaluate spatio-temporal grounding abilities of various models (detailed description is included in the appendix 8.4): + +(i) Spatio-temporal grounding in untrimmed video is evaluated on the proposed Grounding Youtube dataset. The + +entire video and the respective pool of action instructions were provided. The model needs to localize each action step in time (start-time/end-time) and space (location in the video) as described in Figure 3. We evaluate in two metrics: IoU+Pointing game combines the evaluation setting from the spatial grounding [2] and temporal grounding [28] metrics. For each video frame, the prediction is correct when the model predicts the correct action for the frame. Also, given the predicted action as a query, the maximum point of the heatmap aims to lie within the desired bounding box. We then compute the Intersection over Union (IoU) over all the predictions with the GT to acquire the final score. We also compute video mAP following previous evaluation [15], where we set IoU threshold between GT and predicted spatio-temporal tubes. A prediction is correct when it surpasses the IoU threshold. We then compute the mAP over all classes. We form a 3D prediction mask following Figure 3 and compute IoU between our 3D heatmap and 3D tube. + +(ii) Spatial grounding is given a text description to localize the region in the trimmed video. It is evaluated using the pointing game accuracy. If the predicted point lies in the ground truth bounding box, the result counts as a “hit” and counts as “miss” otherwise. The final accuracy is calculated as a ratio between hits to the total number of predictions $\frac{\# \text{hits}}{\# \text{hits} + \# \text{misses}}$ . We also report the mean average precision (mAP) following the settings from V-HICO [32]. + +(iii) Temporal grounding provides videos with the respective actions and their ordering, including the background. The goal is to find the correct frame-wise segmentation of the video. We follow the inference procedure in [28] to compute the alignment given the similarity input matrix. The task is evaluated by intersection over detection (IoD), defined as $\frac{G \cap D}{D}$ the ratio between the intersection of ground-truth action $G$ and prediction $D$ to prediction $D$ , and the Jaccard index, which is an (IoU) given as $\frac{G \cap D}{G \cup D}$ . + +
MethodBackboneDataSuper.Mod.YC-InterGroundingYTV-HICODaly
AccmAPAccmAPAccmAPAccmAP
MIL-NCE [36]S3D*HT100MSelfVT23.6727.458.2112.6511.2313.8424.23
CoMMA† [46]S3DHT250KSelfVT48.6347.6823.3840.9721.4554.4833.39
OursS3DHT100MSelfVT53.9860.6244.9344.3224.3166.3545.93
CLIP [38]CLIPHT100MSelfIT14.1012.503.4929.2312.5118.0227.28
CoMMA‡ [46]CLIPHT100MSelfVT52.6547.5636.4255.2034.5461.0644.37
RegionCLIP [62]RN50x4*CC3MWeakIT51.5652.8423.4257.9237.8267.1248.62
GLIP [31]Swin-L*Cap24MWeakIT52.8453.6224.7366.0541.17--
OursCLIPHT100MSelfVT57.1055.4943.1260.7139.2870.0850.56
OursCLIP*HT100MSelfVT58.3556.9845.3262.3441.5671.3552.78
TubeDETR [54]MDETRVid-STGFullVT51.6353.2441.7663.2340.8784.2162.98
STCAT [23]ResNet-101Vid-STGFullVT54.4755.9044.2165.3441.1085.4263.94
+ +Table 3. Video spatial grounding. We evaluate the accuracy of the pointing game and the mean average precision. We listed CNN-based methods on top and transformer-based methods in the middle. Models learning global representations (MIL-NCE, CLIP) don't perform well on localization tasks, while our model outperforms other grounding methods. * indicates finetuned backbone. + +
MethodBackboneDataSuper.IoUIoD
Mining: MLP [28]TSMMiningYTWeak9.8019.20
CoMMA* [46]S3DHT250KSelf2.055.63
RegionCLIP [62]RN50x4*CC3MWeak10.9616.86
MIL-NCE [36]S3D*HT100MSelf18.6926.74
OursS3DHT100MSelf19.1827.65
OursCLIPHT100MSelf19.8828.50
OursCLIP*HT100MSelf20.3329.67
+ +Table 4. Temporal Grounding on MiningYoutube. * indicates finetuned backbone. Spatial-focused model CoMMA is not trained for temporal detection, which results in lower performance. In contrast, the proposed model combines global and local representation, resulting in better temporal localization than one alone. + +# 5.4. Comparison with state-of-the-art methods + +(i) Spatio-temporal grounding in untrimmed video: We first compare the proposed method with other approaches designed for spatial or temporal grounding in Table 2. It shows that models without specific loss designs for spatial grounding (MIL-NCE [36], CLIP [38]) show good mAP scores but lower pointing game accuracy. Out of the two weakly supervised methods, GLIP [31] and RegionCLIP [62]), trained with aligned image-text, RegionCLIP show significantly better performance in this setting, while both perform in a similar range in the spatial grounding scenario (see Table 3). We attribute this behavior to the fact that RegionCLIP distinguishes frames with relevant queries better from background than GLIP, leading to better temporal localization. We finally compare the strong baseline MIL-NCE+RegionCLIP, which combines two approaches specialized in temporal and spatial aspects, to our task. It shows that the proposed method improves over all other baselines underlining the need to incorporate global (temporal) and local (spatial) representations. + +(ii) Spatial grounding: Second, we compare the performance of the proposed framework to other methods on the task of spatial grounding, including models with weak su + +persion, as well as models trained in a fully supervised setting in Table 3. In the instruction video domain (GYT and YC-Inter), the proposed approach achieves the best result among all weakly and self-supervised trained methods. In the general domain (V-HICO and Daly), the method also achieves competitive results, showing the generalizability of the model to other domains. Note that in the Daly dataset, the classes are verbs, which are not detectable by the object-focused model GLIP. Compared to their weakly trained counterparts, fully-supervised model (TubeDETER [54], STCAT [23]) achieve competitive performance in the general domain (V-HICO, Daly) and slightly lower performance in instruction domain (GYT, YC-Inter) due to the domain gap with respect to the training data. + +(iii) Temporal grounding: We evaluate temporal grounding in Table 4. Here, it shows that global representations also profit from local representation learning. This hypothesis is further validated in the ablation studies in Table 6, where we ablate both losses for all three settings and show a consistent improvement in the joint loss formulation. + +# 5.5. Ablation study + +We perform ablation studies with respect to all three settings, spatio-temporal grounding, as well as spatial and temporal grounding alone, reporting performance for spatiotemporal grounding on GroundingYT using mAP with IoU@0.4, on temporal grounding using MiningYT IoU, and on spatial grounding using YC-Inter. pointing game. Additional ablation are in appendix 9.3. + +Frame selection strategy. We perform an ablation on the possible frame selection strategies for our method (Figure 2(b) and Section 3.2). In Table 5, None uses all frames within the ASR boundary ( $U = T$ ) as our video training data. Global represents the [CLS] token in text and video. Local uses the words and spatio-temporal tokens. In the setting Sinkhorn was not applied, the top $T$ frames with the highest similarity score were selected. When we set + +
GroundingYT Spatio-temporalMiningYT TemporalYC-Inter. SpatialGroundingYT Spatio-temporalMiningYT TemporalYC-Inter. Spatial
w/o Sinkhornwith Sinkhorn
None17.4318.3457.42---
Global T - Global V18.2419.3856.3118.9619.8957.34
Global T - Local V18.0119.3157.5618.5319.3557.67
Local T - Global V18.0519.8557.3419.3220.3658.13
Local T - Local V18.3119.4857.8619.4320.1658.51
Average over last two18.3619.6857.7719.4520.3358.35
+ +![](images/73e882c8d9660d738c349f16054c135f397b4dd1c4e911f56321b4a29f2df7f0.jpg) +Figure 5. Visualization on GroundingYoutube. Red box: annotation. Heatmap: prediction. Baseline GLIP focuses on objects such as power, mixer, and pancake. Our method focuses on actions such as whisk and serve. + +![](images/7141121e14a79562ce1ea8bac23c80d692fd8ab0cfb423d7a3ef0d4e0540c10c.jpg) + +![](images/735330791fee70a416a935d899d1a4b6853334043371955a3b94519f0ab0ee19.jpg) + +![](images/32834c1a9aa6bd0428a07e2d7f0e34fd816ef19acaf3aec0d1002f20ee189d7f.jpg) + +Table 5. Frame selection: (a) Sinkhorn selection results in better supervision. (b) We investigate all possible combinations of global and local representations for frame selection similarity matching. We found keeping the local text representation is crucial, and a combination of local and global representation leads to the best spatio-temporal grounding result. + +
GroundingYT Spatio-tempMiningYT TemporalYC-Inter. Spatial
only Local loss7.295.2355.29
only Global loss9.2819.1236.23
w/ Both loss19.4520.3358.35
+ +Table 6. Loss ablations: both losses contribute to the final loss, and the existence of global loss helps the localization task. + +spatio-temporal tokens as the selection target, we sum over the scores with respect to each frame to acquire the frame similarity score. It shows that selecting frames based on Sinkhorn selection leads to consistently better results as it enforces more variety of visual concepts but also captures frames with possible groundable objects. It further shows that word tokens are more suitable than the global text CLS token for frame selection. Finally, we see that depending on the task (spatial vs. temporal), a local resp. global representation is better, and a combination of both works best for spatio-temporal grounding. We provide runtime analysis of such frame selection strategy in the appendix 9.1. + +Global and local loss. As mentioned in the spatio-temporal evaluation, both features contribute to the final grounding result. We test the model by ablating out each loss. Table 6 shows that each loss not only contributes to the spatiotemporal grounding on the GYT, but also that the whole is more than the sum of its parts (losses) since this task requires both spatial and temporal detection. The reduced impact of the global loss in the case of YC-Inter is that this is a pure spatial grounding dataset (no background frames) without temporal detection, and the local loss plays a more critical role. We observe the same patterns in the temporal + +grounding result for MYT, where spatial localization is not directly contributing to the final performance. We tried out the same ablation using in the S3D backbone in supplement. + +# 5.6. Qualitative results + +We visualize our spatio-temporal result in Figure 5. For the GLIP model, we output the bounding box with the highest confidence score and visualize its center point. We found GLIP model focuses on the salient object while our model focuses more on human-object interaction. + +# 6. Conclusion + +We presented an approach for learning spatio-temporal grounding with self-supervision and a new dataset: GroundingYoutube annotations, where we densely annotate spatiotemporal points/boxes from untrimmed multi-action videos. Our method includes a frame selection mechanism that identifies frames with groundable objects to adapt the learning process for untrimmed videos. Furthermore, we jointly learn global representations, which capture temporal information, and local representations learning fine-grained multimodal interactions between video and text. We conducted extensive experiments and our approach shows state-of-the-art performance in spatio-temporal grounding, as well as temporal and spatial grounding alone. + +Acknowledgments: Nina Shvetsova is supported by the German Federal Ministry of Education and Research (BMBF) project STCL-01IS22067. Andrew Rouditchenko is supported by an NDSEG Fellowship and by the MIT-IBM Watson AI Lab. We thank the MIT-Satori cluster team for their technical support. + +# References + +[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 5, 13 +[2] Hassan Akbari, Svebor Karaman, Surabhi Bhargava, Brian Chen, Carl Vondrick, and Shih-Fu Chang. Multi-level multimodal common semantic space for image-phrase grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12476–12486, 2019. 2, 6, 13 +[3] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 1, 3 +[4] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. Advances in Neural Information Processing Systems, 33:25-37, 2020. 1, 3 +[5] Assaf Arbelle, Sivan Doveh, Amit Alfassy, Joseph Shtok, Guy Lev, Eli Schwartz, Hilde Kuehne, Hila Barak Levi, Prasanna Sattigeri, Rameswar Panda, et al. Detector-free weakly supervised grounding by separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1801-1812, 2021. 5 +[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 12 +[8] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8012-8021, 2021. 1, 3, 14 +[9] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. Advances in Neural Information Processing Systems, 34:28442-28453, 2021. 3 +[10] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee Kenneth Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1884-1894, Florence, Italy, 2019. Association for Computational Linguistics. 2, 3 +[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. 2013. 2, 3, 12 +[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide + +Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 2 +[13] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3 +[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2 +[15] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 2, 5, 6, 13, 14 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3 +[17] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTATS, 2010. 4 +[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2, 3 +[19] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649–665, 2018. 14 +[20] H. Idrees, A. R. Zamir, Y. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2 +[21] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2 +[22] Yu-Gang Jiang, Jingen Liu, A Roshan Zamir, George Toderici, Ivan Laptev, Mubarak Shah, and Rahul Sukthankar. Thumos challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.3,14 +[23] Yang Jin, yongzhi li, Zehuan Yuan, and Yadong MU. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In Advances in Neural Information Processing Systems, 2022. 1, 2, 6, 7, 14, 15 +[24] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In Proceedings of the IEEE + +International Conference on Computer Vision, pages 4405-4413, 2017. 5 +[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3 +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2015. 13 +[27] Giorgos Kordopatis-Zilos, Symeon Papadopoulos, Ioannis Patras, and Ioannis Kompatsiaris. Visil: Fine-grained spatiotemporal video similarity learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6351-6360, 2019. 3 +[28] Hilde Kuehne, Ahsan Iqbal, Alexander Richard, and Juergen Gall. Mining youtube-a dataset for learning fine-grained action concepts from webly supervised video data. 2019. 1, 2, 3, 5, 6, 7, 13 +[29] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 4 +[30] Ang Li, Meghan Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv preprint arXiv:2005.00214, 2020. 2 +[31] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 6, 7, 12, 15 +[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1845-1855, 2021. 2, 6, 13 +[33] Yikang Li, Jenhao Hsiao, and Chiuman Ho. Videoclip: A cross-attention model for fast video-text retrieval task with image clip. In Proceedings of the 2022 International Conference on Multimedia Retrieval, pages 29–33, 2022. 3 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3 +[35] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 12 +[36] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 3, 6, 7, 12, 15 +[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In arXiv preprint arXiv:1301.3781, 2013. 6, 12 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 12, 15 +[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[40] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv preprint arXiv:2006.09199, 2020.3 +[41] Jing Shi, Jia Xu, Boqing Gong, and Chenliang Xu. Not all frames are equal: Weakly-supervised video grounding with contextual similarity and visual clustering losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10444-10452, 2019. 2, 3 +[42] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In CVPR, 2022. 1, 3, 14 +[43] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba Heilbron, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. arXiv preprint arXiv:2112.00431, 2021. 3 +[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2 +[45] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1533–1542, 2021. 2 +[46] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5, 6, 7, 12, 15 +[47] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 1, 2 +[48] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric + +spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2, 3 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 2 +[51] Philippe Weinzaepfel, Xavier Martin, and Cordelia Schmid. Human action localization with sparse spatial supervision. arXiv preprint arXiv:1605.05197, 2016. 2, 6 +[52] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence, 2011. 3 +[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6, 12 +[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 1, 2, 6, 7, 14, 15 +[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2 +[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 3 +[57] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3 +[58] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 3 +[59] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10668-10677, 2020. 2, 3, 5 +[60] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In Pro- + +ceedings of the IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019. 2 +[61] Yang Zhao, Zhou Zhao, Zhu Zhang, and Zhijie Lin. Cascaded prediction network via segment tree for temporal video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4206, 2021. 3 +[62] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2, 3, 6, 7, 12, 15 +[63] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In British Machine Vision Conference, 2018. 2, 3 +[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 2, 3, 6 +[65] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In CVPR, 2019. 2, 3 +[66] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. 2019. 1, 2, 3 \ No newline at end of file diff --git a/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/images.zip b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7292de50f06f477e9728eed2e419e51c7725704b --- /dev/null +++ b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8345c6b9f9dafb243248352f8119c7e36b910e1867b8f296790d01d0f4caa3e5 +size 588815 diff --git a/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/layout.json b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf259d25f79d647349c143628b6ddbc171d816b --- /dev/null +++ b/2024/What When and Where_ Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions/layout.json @@ -0,0 +1,8637 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 514, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 514, + 138 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 514, + 138 + ], + "type": "text", + "content": "What, when, and where? Self-Supervised Spatio-Temporal Grounding in Untrimmed Multi-Action Videos from Narrated Instructions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 516, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 516, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 516, + 190 + ], + "type": "text", + "content": "Brian Chen1 Nina Shvetsova2,3 Andrew Rouditchenko4 Daniel Kondermann5 Samuel Thomas6,7 Shih-Fu Chang1 Rogerio Feris6,7 James Glass4 Hilde Kuehne2,3,7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 193, + 457, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 193, + 457, + 205 + ], + "spans": [ + { + "bbox": [ + 134, + 193, + 457, + 205 + ], + "type": "text", + "content": "1Columbia University, 2Goethe University Frankfurt, 3University of Bonn, 4MIT CSAIL" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "spans": [ + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "text", + "content": "Quality Match GmbH, " + }, + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "text", + "content": "IBM Research AI, " + }, + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 166, + 207, + 426, + 219 + ], + "type": "text", + "content": "MIT-IBM Watson AI Lab" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 222, + 498, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 222, + 498, + 234 + ], + "spans": [ + { + "bbox": [ + 96, + 222, + 498, + 234 + ], + "type": "text", + "content": "{bc2754,sc250}@columbia.edu, shvetsov@uni-frankfurt.de,{roudi,glass}@mit.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 236, + 487, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 236, + 487, + 247 + ], + "spans": [ + { + "bbox": [ + 100, + 236, + 487, + 247 + ], + "type": "text", + "content": "dk@quality-match.com, {sthomas,rsferis}@us.ibm.com, kuehne@cs.uni-bonn.de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 275, + 192, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 275, + 192, + 288 + ], + "spans": [ + { + "bbox": [ + 143, + 275, + 192, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 300, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 290, + 553 + ], + "type": "text", + "content": "Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed, providing dense spatio-temporal grounding annotations in long, untrimmed, multi-action instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks, showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatio-temporal grounding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Spatio-temporal grounding (STG) describes the challenging task of locating events in space and time within video data based on text referential expressions. Methods in this field usually rely on a combination of spatio-temporal bounding box annotation, together with a human-generated caption, describing the visual content of the bounding box [23, 54], which limits their generalizability beyond the given training scenario. Compared to that, as a second line of work, multimodal self-supervised learning tries to leverage \"free\" data sources, such as video and automatic speech" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 293, + 372, + 326 + ], + "blocks": [ + { + "bbox": [ + 319, + 293, + 372, + 326 + ], + "lines": [ + { + "bbox": [ + 319, + 293, + 372, + 326 + ], + "spans": [ + { + "bbox": [ + 319, + 293, + 372, + 326 + ], + "type": "image", + "image_path": "bb12e59d13f4edd0b4b3b2abe2b2049246c4944badb05195208fa2e9e9672005.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 327, + 365, + 347 + ], + "lines": [ + { + "bbox": [ + 320, + 327, + 365, + 347 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 365, + 347 + ], + "type": "text", + "content": "\"To this we will add one carrot chopped into jong jullien ...\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 373, + 293, + 424, + 326 + ], + "blocks": [ + { + "bbox": [ + 373, + 293, + 424, + 326 + ], + "lines": [ + { + "bbox": [ + 373, + 293, + 424, + 326 + ], + "spans": [ + { + "bbox": [ + 373, + 293, + 424, + 326 + ], + "type": "image", + "image_path": "e5611b2857d6b84158d8140794397cffedcd25228d13a515e01c18410f1b5974.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 327, + 423, + 346 + ], + "lines": [ + { + "bbox": [ + 377, + 327, + 423, + 346 + ], + "spans": [ + { + "bbox": [ + 377, + 327, + 423, + 346 + ], + "type": "text", + "content": "\"... you can use the forks to pull it a bit apart...\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 292, + 471, + 326 + ], + "blocks": [ + { + "bbox": [ + 329, + 281, + 525, + 291 + ], + "lines": [ + { + "bbox": [ + 329, + 281, + 525, + 291 + ], + "spans": [ + { + "bbox": [ + 329, + 281, + 525, + 291 + ], + "type": "text", + "content": "Training Setup: Unlabeled videos with narrated instructions" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 292, + 471, + 326 + ], + "lines": [ + { + "bbox": [ + 425, + 292, + 471, + 326 + ], + "spans": [ + { + "bbox": [ + 425, + 292, + 471, + 326 + ], + "type": "image", + "image_path": "a7af3094e5fd67b80fbc50718020f8b7cfe8285cc758d1da4c2a35d9cfaea622.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 471, + 292, + 533, + 330 + ], + "blocks": [ + { + "bbox": [ + 424, + 327, + 471, + 346 + ], + "lines": [ + { + "bbox": [ + 424, + 327, + 471, + 346 + ], + "spans": [ + { + "bbox": [ + 424, + 327, + 471, + 346 + ], + "type": "text", + "content": "①" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 292, + 533, + 330 + ], + "lines": [ + { + "bbox": [ + 471, + 292, + 533, + 330 + ], + "spans": [ + { + "bbox": [ + 471, + 292, + 533, + 330 + ], + "type": "image", + "image_path": "626eb5a1b84be14f724053fa934bb1ec474cbccb20c20844779fa842710513ad.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 331, + 528, + 344 + ], + "lines": [ + { + "bbox": [ + 472, + 331, + 528, + 344 + ], + "spans": [ + { + "bbox": [ + 472, + 331, + 528, + 344 + ], + "type": "text", + "content": "...stir it a bit so it mixes well..." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 317, + 369, + 539, + 415 + ], + "blocks": [ + { + "bbox": [ + 317, + 357, + 536, + 368 + ], + "lines": [ + { + "bbox": [ + 317, + 357, + 536, + 368 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 536, + 368 + ], + "type": "text", + "content": "Evaluation Setup: Referential queries - \"Crack egg\", \"Mix egg\", etc." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 369, + 539, + 415 + ], + "lines": [ + { + "bbox": [ + 317, + 369, + 539, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 539, + 415 + ], + "type": "image", + "image_path": "c3d470eff27720f396457fd2b5344d18b21068a86bff41251a9448107517375a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 317, + 422, + 425, + 459 + ], + "blocks": [ + { + "bbox": [ + 317, + 422, + 425, + 459 + ], + "lines": [ + { + "bbox": [ + 317, + 422, + 425, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 425, + 459 + ], + "type": "image", + "image_path": "4139cd564a8441e6094cbd80fa13d3f0cc4bf4bb34ab953847a529fc4dd97591.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 460, + 539, + 479 + ], + "lines": [ + { + "bbox": [ + 314, + 460, + 539, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 539, + 479 + ], + "type": "text", + "content": "Task: Spatio-Temporal Grounding - Find the temporal boundary of a queried action in an untrimmed video and spatially localize the action." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 502, + 545, + 580 + ], + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 580 + ], + "type": "text", + "content": "Figure 1. Learning Spatio-temporal grounding in untrimmed videos. In training, we learn from unlabeled videos without human annotation. In evaluation, we perform spatio-temporal grounding using an action description such as \"crack egg\" as a query. The model needs to localize both the action's temporal boundary and spatial region in the long untrimmed video. We visualize the heatmap from the annotation points as well as derived bounding boxes." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 430, + 422, + 537, + 459 + ], + "blocks": [ + { + "bbox": [ + 430, + 422, + 537, + 459 + ], + "lines": [ + { + "bbox": [ + 430, + 422, + 537, + 459 + ], + "spans": [ + { + "bbox": [ + 430, + 422, + 537, + 459 + ], + "type": "image", + "image_path": "9749226b99a0a6a93abf15403719bf859baa2a436dd7922fac2da728763b02a4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": "recognition (ASR) captions from large-scale instructional videos to learn representations without human annotation [3, 4, 8, 35, 36]. The resulting models achieve state-of-the-art performance on zero-shot tasks such as cross-modal video retrieval or classification and also for zero-shot temporal action segmentation and detection based on free text queries [8, 28, 42, 47, 66], but usually lack spatial localization abilities. A third line of work focuses on label-free spatial grounding, e.g. by training on image-caption" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18419" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "text", + "content": "[2, 31, 50, 55, 62] or video-caption pairs [41, 46]. The goal is to correctly localize a referential expression in an image or each video frame, e.g., via a bounding box or a heatmap. However, those methods are not optimized to detect whether an event is present in a video. The assumption is thus that the evaluated expression is visible in the image or in all video frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 289, + 517 + ], + "type": "text", + "content": "The following work aims to bring together those ideas to address the task of spatio-temporal action grounding from multimodal supervision in untrimmed videos. We propose a grounding approach that uses video-text pairs based on ASR transcripts in instructional videos and learns the spatial representation of free-text events as well as their temporal extent, as shown in Figure 1. To this end, we leverage two different representations of the visual data: a global feature representation based on full-frame information to define the temporal extent of an event and a local representation based on frame-wise grid features for spatial localization. The motivation for this dualism is that while the local representation captures the spatial correlations between vision and text input, this can be too fine-grained to learn a holistic representation of the frame, while the global representation can be assumed to capture a more compact, aggregated view compared to local data and thus to provide a more reliable cue for the task of temporal localization. However, compared to the hand-annotated video-caption setup of most spatio-temporal grounding methods, the ASR text can be noisy as not all comments refer to visible events. Further, as there is only a loose temporal correlation, the described activities might not be precisely aligned, can be scattered over multiple frames, or not be present at all [18, 36]. Therefore, we propose to specifically select frames to capture only those useful for training. To this end, we look for frames that match the vocabulary of the respective text, leveraging a selection strategy by Sinkhorn optimal transport [11]. This allows us to train a model that can localize actions in space and time within videos without labeling supervision." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 520, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 698 + ], + "type": "text", + "content": "To evaluate spatio-temporal grounding in untrimmed videos, a new benchmark, GroundingYouTube, is proposed. It is based on the existing MiningYouTube dataset [28] and extended with spatio-temporal localization information. This setup differs from other benchmarks such as [10, 46, 63] in two ways: first, by using multiple center point annotations, it focuses on the grounding of referential actions itself instead of interacting humans or objects which are usually labeled; second, the dense annotations of multiple actions in the video allow us to benchmark action grounding in long, realistic untrimmed videos compared to existing, often pre-clipped benchmarks[10, 59]. The benchmark provides queries for 512 different event types and over " + }, + { + "bbox": [ + 46, + 520, + 287, + 698 + ], + "type": "inline_equation", + "content": "5K" + }, + { + "bbox": [ + 46, + 520, + 287, + 698 + ], + "type": "text", + "content": " spatio-temporal annotations, as shown in Figure 1. A comparison of current datasets is shown in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "To evaluate the proposed approach as well as the new" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 545, + 260 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 260 + ], + "type": "table", + "html": "
DatasetAnnotation
SpatialTemporal
V-HiCoobject bb + human bb-
AVA-Kineticsobject bb + human bb-
ActivityNet entitiesobject bb + human bb-
THUMOS14-action boundaries
ActivityNet-action boundaries
HACSSegment-action boundaries
YouCook II-multi-action boundaries
Cross-Task-multi-action boundaries
COIN-multi-action boundaries
EPIC KITCHENS-100-multi-action boundaries
Ego4D-multi-action boundaries
JHMDB51-21human tubes-
UCF101-24human tubesaction boundaries
Dalyhuman tubesaction boundaries
Vid-STGhuman tubesaction boundaries
HC-STVGhuman tubesaction boundaries
AVAhuman tubesaction boundaries
YouCook-Interactionsaction bb-
GroundingYoutube (ours)action bb + center pointsmulti-action boundaries
", + "image_path": "b6b877834e74c5dbd336dc72163de22a9ba9b6ba5920b8431f29201743d459bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 269, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 304, + 269, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 547, + 346 + ], + "type": "text", + "content": "Table 1. Comparison of spatial, temporal, and spatio-temporal grounding datasets. V-HiCo [32], AVA-Kinetics [30], THU-MOS14 [20], ActivityNet [6], ActivityNet entity [65], HACSSegment [60], YouCook II [64], Cross-Task [66], COIN [47], EPIC KITCHENS-100 [12], Ego4D [14], JHMDB51-21 [21], UCF101-24 [44] Daly [51], Vid-STG [59], HC-STVG [48], AVA [15], YouCook-Interactions [46]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 360, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 443 + ], + "type": "text", + "content": "benchmark, the system is trained on the HowTo100M dataset [35] and compared to state-of-the-art methods based on full, weak, and self-supervision for spatial and temporal, as well as combined spatio-temporal grounding tasks. It shows that existing methods usually do well in one of the two aspects, spatial or temporal grounding. In contrast, the proposed method can combine spatial and temporal aspects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 445, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 587 + ], + "type": "text", + "content": "We summarize the contributions of this work as follows1: (1) We propose a framework for spatio-temporal grounding in untrimmed videos based on weakly aligned multimodal supervision without human annotation, employing a combination of global and local representation learning to learn the spatio-temporal extent of actions. (2) To facilitate this task, we propose a frame selection strategy based on Sinkhorn-Knopp Optimal transport that improves the quality of the acquired learning samples, leading to more effective supervision. (3) We provide a new benchmark and annotations to evaluate this challenging problem on real-world multi-action instructional video data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 601, + 392, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 392, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 392, + 613 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "type": "text", + "content": "Supervised Spatio-temporal Grounding. Spatiotemporal Grounding refers to the problem of localizing a sequence of bounding boxes (a spatio-temporal tube) for a target object described by an input text. This problem has been addressed by various approaches TubeDETR [54], STCAT [23], STVGBert [45], STVGBert [45]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 520, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 520, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 520, + 712 + ], + "type": "text", + "content": "1We will make the code and the annotations publicly available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18420" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "STGVT [48], STGRN [59], Visil [27]. These methods rely on proposal networks such as Faster R-CNN [39] or MDETR [25] to predict bounding box coordinates for learning text-to-region interaction. All those approaches rely on supervised training with the human-annotated sentence and bounding box supervision, provided, e.g., by datasets such as VidSTG [59], HC-STVG [10]. While those datasets provide a temporal aspect, temporal detection is usually limited to identifying the start and end frame of a single action in a video. Compared to that, an untrimmed setting usually comprises multiple actions in a longer video that can be separated by longer background sequences. This conceptually differs from previous works [10] that typically use short videos of around 5-10 seconds. Other datasets such as ActivityNet entities [65] provide only bounding boxes for noun phrases in the captions, namely the objects, which is related to object detection task and does not capture any spatial or temporal extent of actions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": "Multimodal Self-supervised Learning. The field of multimodal self-supervised learning aims to learn data representations by leveraging large amounts of unlabeled data with multiple modalities. Early works [13, 52] started by projecting images and text into a joint visual-language embedding space, where embeddings of semantically similar pairs are close. Those ideas have now grown into systems such as MIL-NCE [36] using the HowTo100M dataset [35] to train a video-language embedding space from 1.2 million instructional videos paired with text descriptions from ASR. Follow-up works, including [3, 4, 8, 40, 42] show that using videos without annotation enables an effective multimodal embedding space via contrastive learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 449, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 688 + ], + "type": "text", + "content": "Based on those advantages, approaches started to address the problem of Spatial Video Grounding from multimodal self-supervised aiming to identify spatial locations in a trimmed video based on text descriptions without the need for bounding box annotation during training. One of the early works studied this task in the context of weakly supervised learning where we learn grounding with human-annotated captions of the video [63]. In this context, works [41, 46] have focused on object grounding benchmarks such as YouCook2-BoundingBox [64], which provides bounding box annotations for visible objects in cooking videos. Other works such as GLIP [31], Region-CLIP [62], and others [56, 58] combine the principles of large-scale vision language training with bounding box finetuning on object detection datasets [16, 34]. Recently, the YouCook-Interactions dataset [46] and CoMMA [46] have been proposed for the spatial grounding of objects and actions with multimodal self-supervision from HowTo100M videos. These works assume that the video is temporally clipped with respect to the grounding phrase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Compared to that, Temporal Video Grounding aims to determine the set of consecutive frames corresponding to a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "text query in an untrimmed video [9, 22, 43], thus predicting temporal boundaries of action instances. Recent work such as MIL-NCE [36], MCN [8], and VideoCLIP [33] utilize large-scale pretraining for grounding actions temporally via text-to-frame similarity on video datasets such as MiningYouTube [28] or CrossTask [66] without proposals. However, these methods lack spatial localization ability [57, 61]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 167, + 545, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 196 + ], + "type": "text", + "content": "3. A Global-Local Framework for Spatio-Temporal Grounding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 201, + 395, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 395, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 395, + 215 + ], + "type": "text", + "content": "3.1. General setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": "The goal of the proposed method is to temporally and spatially localize actions based on free-text queries in untrimmed videos. To this end, two representations are learned, a local and a global one. We start with narrated video clips, each associated with a corresponding visual representation and text narration. Namely, for each clip " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathcal{V},\\mathcal{S}\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " stand for the video clip and " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " for the text narration sentence generated by the automatic speech recognition (ASR) system. Each clip " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "U\\times N" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " spatio-temporal tokens " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{v_{u,n}\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "u\\in \\{1,\\dots,U\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " represents the number of frames in the video and " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "n\\in \\{1,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " represents the number of spatial grid region tokens or features in a frame. The text sentence " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " words " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{s_1,\\ldots ,s_K\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ". We represent localized features by the tokens from each modality, and the global features " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{V,S\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " are acquired either by mean-pooling over the local features (S3D) or by using the [CLS] token from the transformer (CLIP) as in Radford et al. [38]. We learn transformations " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "f:V\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": "-dimensional representation " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "f(V)\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " from the global representation " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "g:S\\rightarrow \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ", to produce similar " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": "-dimensional text global embeddings: " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "g(S)\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ". Similar to " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{f,g\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": ", we note " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{f',g'\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " to be the transform for localized features, where local features " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{v,s\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": " are also projected as " + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 220, + 547, + 508 + ], + "type": "text", + "content": "-dimensional representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 515, + 511, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 511, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 511, + 529 + ], + "type": "text", + "content": "3.2. Representation guided frame sampling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Learning from multimodal self-supervision is challenging since the narration is likely to be noisy, thus containing more information than the actual task descriptions due to poor temporal alignment or cut scenes [18], which is the key differences between weakly supervised vision-caption grounding and multimodal self-supervised grounding. This work pursues a frame selection strategy to improve object grounding and temporal alignment during training. We start from a longer sequence " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "U > T" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", which includes the video frames before and after the ASR boundaries that could contain actions or objects in the sentence. Our goal is to find " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " frames out of the " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " frames that are most relevant to the actions and objects in the sentence " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". We formalize it as an optimal transport problem utilizing the Sinkhorn-Knopp algorithm [11]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18421" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 202 + ], + "type": "image", + "image_path": "49a7b57db49a9fad34daef3db334c0b00e06c4e67b23e2a509527d564c1166bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 548, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 548, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 548, + 253 + ], + "type": "text", + "content": "Figure 2. Spatio-temporal grounding approach. (a) We aim to select frames with groundable objects and actions. To this end, projected text features are matched with respective frame features. (b) Sinkhorn optimal transport is then leveraged to optimize the selected frames wrt. the text input. (c) Based on the selected frames, a global representation is learned to allow for temporal localization, as well as (d) a local representation to ground the action description in the spatial region." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "content": "Optimal transport for text-to-frame assignment. To acquire the optimal assignment from word features to video frames, an assignment matrix " + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "content": " is computed from each video and ASR pair as shown in Figure 2(a). This cross-model optimal transport mechanism is applied to assignment " + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "content": " from the projected cross-model similarity " + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "content": " between word tokens and each video frame, where " + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = g(\\mathcal{S})\\otimes f(\\mathcal{V})^{\\top}\\in \\mathbb{R}^{K\\times U}" + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "content": ". To compute the assignment matrix, the text and video projection layers from the global representation in Figure 2(c) are used to project multimodal features into a common space for feature similarity calculation. We investigate various granularities of the features where we compute the similarity between the text features at the word (local) / sentence (global) level and the visual feature at frames (global) / spatiotemporal tokens (local) level to acquire " + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 259, + 289, + 523 + ], + "type": "text", + "content": ", as shown in Table 5. To ensure that the word-to-frame assignment contains more diversity instead of just saturated assignments to a single video frame, we add a constraint by Sinkhorn that requires label assignments to be equally distributed across various video frames representing diverse object/action concepts. Details of the Sinkhorn optimal transport are included in the appendix 7.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 278, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 278, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 278, + 550 + ], + "type": "text", + "content": "3.3. Local representations for spatial localization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "To capture multimodal interaction with finer granularity, we need to learn the projection between tokenized features as shown in Figure 2(d). We extract spatio-temporal region features " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "v_{tn}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " from the video. Also, we extract word features " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " which represents the feature from word " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ". All tokenized features are projected through a linear layer. To compute attention between the tokenized features, we stacked two cross-modal attention layers with a self-attention layer in the middle, as illustrated in Figure 2 (d). Cross-modal attention is computed similar to the standard attention mechanism [29]. Given a spatio-temporal token " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "v_{tn}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " from a video, we compute the attention score to all of the words " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " in the ASR sen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": "tence " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\alpha_{tnk} = \\frac{\\exp(e_{tnk})}{\\sum_{k=1}^{K}\\exp(e_{tnk})}" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " in the same video clip, where " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "e_{tnk} = \\cosine(v_{tn}, s_k)" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": ". We then acquire a contextual video token feature " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\bar{v}_{tn} = \\sum_{k=1}^{K}\\alpha_{tnk}s_k" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": ", which encoded text contextual information. Note that the contextual vector is represented by aggregating the representations from the other modality. Follow the standard self-attention computation [49] " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "K, Q, V" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " represent the features for the keys, queries, and values as: " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "Attn(K, Q, V) = \\text{softmax}\\left(\\frac{(Q^\\top K)}{\\sqrt{d_k}}\\right)V" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " is the dimension of the key. In our case, we feed each contextual feature " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\{\\bar{v}_{tn}, \\bar{s}_k\\}" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " right after the first cross-attention layer to the " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "K, Q, V" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " to acquire its self-attended representation. The localized attention model was trained using contrastive loss. To represent the video clip " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " and ASR sentence " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": ", we mean-pool over the spatio-temporal tokens in video " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\bar{V} = \\frac{1}{TN}\\sum_{r=1}^{TN}\\bar{v}_r" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": ", and words " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\bar{S} = \\frac{1}{K}\\sum_{k=1}^{K}\\bar{s}_k" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " respectively. Let " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "(\\bar{V}^{(l)}, \\bar{S}^{(l)})" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": "-th training example pair. We adopt the Noise Contrastive Estimation (NCE) loss [17] and the localized attention losses " + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Local}" + }, + { + "bbox": [ + 304, + 257, + 547, + 504 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 506, + 547, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 547, + 559 + ], + "type": "interline_equation", + "content": "- \\frac {1}{B} \\sum_ {l = 1} ^ {B} \\left[ \\left(\\log \\frac {e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta}}{e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta} + \\sum_ {k = 1} ^ {B} e ^ {\\bar {V} _ {k} ^ {i m p} \\cdot \\bar {S} _ {l}}}\\right) + \\left(\\log \\frac {e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta}}{e ^ {\\bar {V} _ {l} \\cdot \\bar {S} _ {l} - \\delta} + \\sum_ {k = 1} ^ {B} e ^ {\\bar {V} _ {k} \\cdot \\bar {S} _ {k} ^ {i m p}})}\\right) \\right] \\tag {1}", + "image_path": "791d9b2c595443b2d317343b13498c2988a7a6c7dd80e1554ec50c5c2be3a3ce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "text", + "content": " is the batch. " + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\bar{V}_k^{imp}" + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\bar{S}_k^{imp}" + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "text", + "content": " represent imposter samples, and " + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 567, + 547, + 594 + ], + "type": "text", + "content": " is a margin hyperparameter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 599, + 499, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 499, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 499, + 613 + ], + "type": "text", + "content": "3.4. Global representations for temporal" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "We learn the global representation of a video clip and a sentence by contrastive loss, as shown in Figure 2(c). We again use the NCE loss function [17]. The global contrastive loss " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Global}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " follows the formulation as Equation 1 while using the global representations " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": ", which are the [CLS] tokens or mean-pooled features from both modalities, instead of the local representations. Projecting the global features to the same space ensures that the features across different" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18422" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "modalities are comparable. Since global representations encode information from the entire video, it is essential in encoding temporal information for the later downstream tasks. The final model is optimized by the sum of both losses as " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Final}} = \\mathcal{L}_{\\text{Local}} + \\mathcal{L}_{\\text{Global}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 144, + 285, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 144, + 285, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 144, + 285, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 144, + 285, + 243 + ], + "type": "image", + "image_path": "b207eab4a8d3fcbd28b955ccdffc923eb4ab598946c7d7a9d16906852e7ad8d1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 288, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 299 + ], + "type": "text", + "content": "Figure 3. Spatio-temporal inference. Both representations are used for spatio-temporal grounding: Starting by predicting the action boundary, spatial grounding is performed on the selected frames using the predicted label to find corresponding regions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 308, + 261, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 261, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 261, + 321 + ], + "type": "text", + "content": "3.5. Inference for spatio-temporal grounding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": "To perform spatio-temporal grounding on untrimmed videos, we start from temporal action detection as shown in Figure 3. Given a pool of possible action descriptions on the left and an untrimmed video, we perform feature similarity matching using the global representation ([CLS] token or mean-pooled feature) per frame with a threshold " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": " to filter backgrounds. We pick the action class with the largest similarity score per frame. Later, we use the predicted action class and feed it into the local representation branch to compute spatial grounding. We follow [1] to compute feature similarity between visual tokens and text tokens through the cross-attention and self-attention. In the end, we acquire an attention heatmap for later downstream evaluation. More information on inferencing are in appendix 8.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 505, + 224, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 224, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 224, + 518 + ], + "type": "text", + "content": "4. GroundingYoutube Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 677 + ], + "type": "text", + "content": "Current downstream datasets either provide spatial [46], or temporal annotation [28], or spatio-temporal annotation [59] but only for short video clips with few frames before and after the action takes place. These datasets do not provide the opportunity to evaluate both aspects, spatial and temporal grounding, in an untrimmed long video manner. We, therefore, extend one of the current benchmarks, MiningYouTube [28], with 250 untrimmed videos with a duration of around 3-5 minutes and an average of 41.6 labeled action instances per video. Note that each video contains various classes. As the dataset already provides dense temporal annotations, we annotate the respective temporal action segments in the dataset with spatial information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Annotating the spatio-temporal extent of actions can be challenging as there is no clear visible outline as, e.g., in object annotation, nor is there a unique signal to indicate the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 424, + 169 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 424, + 169 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 424, + 169 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 424, + 169 + ], + "type": "image", + "image_path": "8c03b6fbed03d0fab120053dfcf58d99f47add551043fcf3b267631841f87d9a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 530, + 169 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 530, + 169 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 530, + 169 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 530, + 169 + ], + "type": "image", + "image_path": "30a57a3f217ab0dd9e1e39f780a2de1e0270101b09e6bb988983b72b4bdee44b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 176, + 423, + 275 + ], + "blocks": [ + { + "bbox": [ + 318, + 176, + 423, + 275 + ], + "lines": [ + { + "bbox": [ + 318, + 176, + 423, + 275 + ], + "spans": [ + { + "bbox": [ + 318, + 176, + 423, + 275 + ], + "type": "image", + "image_path": "ad091c8deff9164b3ab0688ace8c1714f4a8cead98b6f0cfb158045a96bf319c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 278, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 545, + 323 + ], + "type": "text", + "content": "Figure 4. Visualization of the point annotation and automatic bounding box generation from points. The red point represents the mean of the five annotation points. The points annotation captures diverse patterns in various action types." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 176, + 531, + 275 + ], + "blocks": [ + { + "bbox": [ + 426, + 176, + 531, + 275 + ], + "lines": [ + { + "bbox": [ + 426, + 176, + 531, + 275 + ], + "spans": [ + { + "bbox": [ + 426, + 176, + 531, + 275 + ], + "type": "image", + "image_path": "0086a40ad8f7bc800a6bb2e54ada527117c024fc79a7131d2599bd7573ca4827.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": "temporal begin and end points. Similarly, grounding systems do not produce pixel-exact bounding boxes, but rather indicate regions of interest. Detector-free spatial grounding models [5] address this fuzziness by relying on pointing game accuracy, thus only using the center point of the heat map for evaluation. Lending on this idea, annotators were asked to mark the presumed center point of the action. Compared to bounding boxes, center point annotation can be advantageous because annotators are not visually distracted by object outlines, so it is more likely that the most important region will be selected. We capture five annotations per frame, resulting in a density-based heatmap." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 479, + 546, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 546, + 612 + ], + "type": "text", + "content": "Starting from 5,091 clips showing one of the 512 action classes, we adopt the methodology used for temporal action localization developed in [15] and label one frame per second, resulting in 26,987 labeled frames. We annotated all frames with five repeats per image, resulting in five annotations per frame and 134,935 point labels in total. To evaluate using bounding boxes [24], we get the union of all annotated points with additional distance to construct the bounding box as shown in Figure 4. More information on the annotation process, bounding box derivation, and dataset analysis is provided in the appendix 10." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 625, + 387, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 625, + 387, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 625, + 387, + 639 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 646, + 368, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 368, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 368, + 658 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Training Data: HowTo100M dataset contains 1.2M instructional videos along with their corresponding automatically generated speech (ASR). The narrations may be inaccurate and do not always accurately depict the video scene." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18423" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 535, + 218 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 535, + 218 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 535, + 218 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 535, + 218 + ], + "type": "table", + "html": "
MethodBackboneDataSetSupervisionModalityGroundingYoutube
IoU+PointmAP
0.10.20.30.40.50.1:0.5
CoMMA† [46]S3DHT250KSelfVT1.022.181.721.110.930.371.26
MIL-NCE [36]S3D*HT100MSelfVT4.6733.9425.1612.653.420.4115.11
OursS3DHT100MSelfVT9.1242.7035.4925.1616.2210.0525.92
GLIP [31]Swin-L*Cap24MWeakIT1.242.832.101.520.960.371.56
CoMMA‡ [46]CLIPHT100MSelfVT1.683.512.321.880.990.401.82
CLIP [38]CLIPHT100MSelfIT3.5929.5422.159.162.480.3912.74
RegionCLIP [62]ResNet-101*CC3MWeakIT5.6535.6527.4315.694.310.8616.78
OursCLIPHT100MSelfVT10.0942.8136.0525.8417.1011.3526.63
OursCLIP*HT100MSelfVT11.5343.6436.9426.7819.4514.6128.26
MIL-NCE(temp.)+RegionCLIP(spa.)---VT9.2140.5434.9722.3813.799.1822.33
", + "image_path": "2c1d64c827989b301853771407943e211c2333d60e3ebc2ee3dfcc8aef08976d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 226, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 546, + 261 + ], + "type": "text", + "content": "Table 2. Spatio-temporal grounding on GroundingYouTube full videos. The proposed model learns global representations encoding global information and spatial correspondences across modalities, achieving a better performance in spatio-temporal evaluation compared to models trained on only spatial or temporal grounding. (V: video, I: image, T: text.) * indicates finetuned backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 272, + 289, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 289, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 289, + 440 + ], + "type": "text", + "content": "Downstream Datasets: GroundingYoutube (GYT) is used to evaluate the task of multi-action spatio-temporal grounding as described in Section 4. MiningYoutube (MYT) [28] provides temporal annotation and is limited to the domain of cooking instruction videos. YouCook-Interaction (YC-Inter) [46] is an extension of the YouCook2 dataset [64] for cooking instruction providing bounding boxes for 6K selected frames. The bounding boxes usually comprise the hand and the tool mentioned in the respective sentence-wise annotation. To further benchmark on general video domains on the V-HICO dataset [32] with 6.5k videos with human-object interaction bounding boxes annotations, and Daly action dataset [51], featuring videos consisting of daily actions such as \"brushing teeth\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 447, + 151, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 151, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 151, + 459 + ], + "type": "text", + "content": "5.2. Baseline methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 465, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 287, + 609 + ], + "type": "text", + "content": "Temporal: MIL-NCE [36] utilizes S3D [53] and word2vec [37]. CLIP [38], an image-text model with transformer. Spatial: CoMMA [46], SSL model († for weights shared by the author² ‡ trained with CLIP); GLIP [31], RegionCLIP [62], SOTA weakly supervised grounding model. Spatio-temporal: We construct MIL-NCE+RegionCLIP following the inference pipeline in Figure 3. TubeDETR [54] and STCAT [23] are supervised. More descriptions of the baselines are given in the Appendix 8.1. Details of the implementation and experimental settings can be found in the appendix 8.2. Inference setups for each baseline are described in Section 8.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 617, + 159, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 159, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 159, + 628 + ], + "type": "text", + "content": "5.3. Downstream Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 635, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 670 + ], + "type": "text", + "content": "We considered the following downstream tasks to evaluate spatio-temporal grounding abilities of various models (detailed description is included in the appendix 8.4):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 287, + 695 + ], + "type": "text", + "content": "(i) Spatio-temporal grounding in untrimmed video is evaluated on the proposed Grounding Youtube dataset. The" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 272, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 545, + 488 + ], + "type": "text", + "content": "entire video and the respective pool of action instructions were provided. The model needs to localize each action step in time (start-time/end-time) and space (location in the video) as described in Figure 3. We evaluate in two metrics: IoU+Pointing game combines the evaluation setting from the spatial grounding [2] and temporal grounding [28] metrics. For each video frame, the prediction is correct when the model predicts the correct action for the frame. Also, given the predicted action as a query, the maximum point of the heatmap aims to lie within the desired bounding box. We then compute the Intersection over Union (IoU) over all the predictions with the GT to acquire the final score. We also compute video mAP following previous evaluation [15], where we set IoU threshold between GT and predicted spatio-temporal tubes. A prediction is correct when it surpasses the IoU threshold. We then compute the mAP over all classes. We form a 3D prediction mask following Figure 3 and compute IoU between our 3D heatmap and 3D tube." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 595 + ], + "type": "text", + "content": "(ii) Spatial grounding is given a text description to localize the region in the trimmed video. It is evaluated using the pointing game accuracy. If the predicted point lies in the ground truth bounding box, the result counts as a “hit” and counts as “miss” otherwise. The final accuracy is calculated as a ratio between hits to the total number of predictions " + }, + { + "bbox": [ + 304, + 498, + 545, + 595 + ], + "type": "inline_equation", + "content": "\\frac{\\# \\text{hits}}{\\# \\text{hits} + \\# \\text{misses}}" + }, + { + "bbox": [ + 304, + 498, + 545, + 595 + ], + "type": "text", + "content": ". We also report the mean average precision (mAP) following the settings from V-HICO [32]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "(iii) Temporal grounding provides videos with the respective actions and their ordering, including the background. The goal is to find the correct frame-wise segmentation of the video. We follow the inference procedure in [28] to compute the alignment given the similarity input matrix. The task is evaluated by intersection over detection (IoD), defined as " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\frac{G \\cap D}{D}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " the ratio between the intersection of ground-truth action " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " and prediction " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " to prediction " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ", and the Jaccard index, which is an (IoU) given as " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\frac{G \\cap D}{G \\cup D}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 235, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 235, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 235, + 713 + ], + "type": "text", + "content": "2We thank the authors for providing code and weights." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18424" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 533, + 216 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 533, + 216 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 533, + 216 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 533, + 216 + ], + "type": "table", + "html": "
MethodBackboneDataSuper.Mod.YC-InterGroundingYTV-HICODaly
AccmAPAccmAPAccmAPAccmAP
MIL-NCE [36]S3D*HT100MSelfVT23.6727.458.2112.6511.2313.8424.23
CoMMA† [46]S3DHT250KSelfVT48.6347.6823.3840.9721.4554.4833.39
OursS3DHT100MSelfVT53.9860.6244.9344.3224.3166.3545.93
CLIP [38]CLIPHT100MSelfIT14.1012.503.4929.2312.5118.0227.28
CoMMA‡ [46]CLIPHT100MSelfVT52.6547.5636.4255.2034.5461.0644.37
RegionCLIP [62]RN50x4*CC3MWeakIT51.5652.8423.4257.9237.8267.1248.62
GLIP [31]Swin-L*Cap24MWeakIT52.8453.6224.7366.0541.17--
OursCLIPHT100MSelfVT57.1055.4943.1260.7139.2870.0850.56
OursCLIP*HT100MSelfVT58.3556.9845.3262.3441.5671.3552.78
TubeDETR [54]MDETRVid-STGFullVT51.6353.2441.7663.2340.8784.2162.98
STCAT [23]ResNet-101Vid-STGFullVT54.4755.9044.2165.3441.1085.4263.94
", + "image_path": "6e65c85cfac99c243fbe8b753f3d223c78fee9c8b59e8da18b462b13bc2837d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 258, + 287, + 364 + ], + "blocks": [ + { + "bbox": [ + 46, + 219, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 547, + 253 + ], + "type": "text", + "content": "Table 3. Video spatial grounding. We evaluate the accuracy of the pointing game and the mean average precision. We listed CNN-based methods on top and transformer-based methods in the middle. Models learning global representations (MIL-NCE, CLIP) don't perform well on localization tasks, while our model outperforms other grounding methods. * indicates finetuned backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 258, + 287, + 364 + ], + "lines": [ + { + "bbox": [ + 48, + 258, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 287, + 364 + ], + "type": "table", + "html": "
MethodBackboneDataSuper.IoUIoD
Mining: MLP [28]TSMMiningYTWeak9.8019.20
CoMMA* [46]S3DHT250KSelf2.055.63
RegionCLIP [62]RN50x4*CC3MWeak10.9616.86
MIL-NCE [36]S3D*HT100MSelf18.6926.74
OursS3DHT100MSelf19.1827.65
OursCLIPHT100MSelf19.8828.50
OursCLIP*HT100MSelf20.3329.67
", + "image_path": "b821de0137db88eb782c0f1bde06435ccd885aba43daed54389a2df808344103.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 365, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 422 + ], + "type": "text", + "content": "Table 4. Temporal Grounding on MiningYoutube. * indicates finetuned backbone. Spatial-focused model CoMMA is not trained for temporal detection, which results in lower performance. In contrast, the proposed model combines global and local representation, resulting in better temporal localization than one alone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 266, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 266, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 266, + 443 + ], + "type": "text", + "content": "5.4. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 676 + ], + "type": "text", + "content": "(i) Spatio-temporal grounding in untrimmed video: We first compare the proposed method with other approaches designed for spatial or temporal grounding in Table 2. It shows that models without specific loss designs for spatial grounding (MIL-NCE [36], CLIP [38]) show good mAP scores but lower pointing game accuracy. Out of the two weakly supervised methods, GLIP [31] and RegionCLIP [62]), trained with aligned image-text, RegionCLIP show significantly better performance in this setting, while both perform in a similar range in the spatial grounding scenario (see Table 3). We attribute this behavior to the fact that RegionCLIP distinguishes frames with relevant queries better from background than GLIP, leading to better temporal localization. We finally compare the strong baseline MIL-NCE+RegionCLIP, which combines two approaches specialized in temporal and spatial aspects, to our task. It shows that the proposed method improves over all other baselines underlining the need to incorporate global (temporal) and local (spatial) representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "(ii) Spatial grounding: Second, we compare the performance of the proposed framework to other methods on the task of spatial grounding, including models with weak su" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 259, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 426 + ], + "type": "text", + "content": "persion, as well as models trained in a fully supervised setting in Table 3. In the instruction video domain (GYT and YC-Inter), the proposed approach achieves the best result among all weakly and self-supervised trained methods. In the general domain (V-HICO and Daly), the method also achieves competitive results, showing the generalizability of the model to other domains. Note that in the Daly dataset, the classes are verbs, which are not detectable by the object-focused model GLIP. Compared to their weakly trained counterparts, fully-supervised model (TubeDETER [54], STCAT [23]) achieve competitive performance in the general domain (V-HICO, Daly) and slightly lower performance in instruction domain (GYT, YC-Inter) due to the domain gap with respect to the training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 428, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 500 + ], + "type": "text", + "content": "(iii) Temporal grounding: We evaluate temporal grounding in Table 4. Here, it shows that global representations also profit from local representation learning. This hypothesis is further validated in the ablation studies in Table 6, where we ablate both losses for all three settings and show a consistent improvement in the joint loss formulation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 512, + 398, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 398, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 398, + 525 + ], + "type": "text", + "content": "5.5. Ablation study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": "We perform ablation studies with respect to all three settings, spatio-temporal grounding, as well as spatial and temporal grounding alone, reporting performance for spatiotemporal grounding on GroundingYT using mAP with IoU@0.4, on temporal grounding using MiningYT IoU, and on spatial grounding using YC-Inter. pointing game. Additional ablation are in appendix 9.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Frame selection strategy. We perform an ablation on the possible frame selection strategies for our method (Figure 2(b) and Section 3.2). In Table 5, None uses all frames within the ASR boundary (" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "U = T" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ") as our video training data. Global represents the [CLS] token in text and video. Local uses the words and spatio-temporal tokens. In the setting Sinkhorn was not applied, the top " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " frames with the highest similarity score were selected. When we set" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18425" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 535, + 160 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 535, + 160 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 535, + 160 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 535, + 160 + ], + "type": "table", + "html": "
GroundingYT Spatio-temporalMiningYT TemporalYC-Inter. SpatialGroundingYT Spatio-temporalMiningYT TemporalYC-Inter. Spatial
w/o Sinkhornwith Sinkhorn
None17.4318.3457.42---
Global T - Global V18.2419.3856.3118.9619.8957.34
Global T - Local V18.0119.3157.5618.5319.3557.67
Local T - Global V18.0519.8557.3419.3220.3658.13
Local T - Local V18.3119.4857.8619.4320.1658.51
Average over last two18.3619.6857.7719.4520.3358.35
", + "image_path": "759ade2634013100c3f9e42c2560075e5119a2bbed1f44f043240f042c157ac8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 208, + 185, + 312 + ], + "blocks": [ + { + "bbox": [ + 57, + 208, + 185, + 312 + ], + "lines": [ + { + "bbox": [ + 57, + 208, + 185, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 185, + 312 + ], + "type": "image", + "image_path": "73e882c8d9660d738c349f16054c135f397b4dd1c4e911f56321b4a29f2df7f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "type": "text", + "content": "Figure 5. Visualization on GroundingYoutube. Red box: annotation. Heatmap: prediction. Baseline GLIP focuses on objects such as power, mixer, and pancake. Our method focuses on actions such as whisk and serve." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 189, + 208, + 302, + 312 + ], + "blocks": [ + { + "bbox": [ + 189, + 208, + 302, + 312 + ], + "lines": [ + { + "bbox": [ + 189, + 208, + 302, + 312 + ], + "spans": [ + { + "bbox": [ + 189, + 208, + 302, + 312 + ], + "type": "image", + "image_path": "7141121e14a79562ce1ea8bac23c80d692fd8ab0cfb423d7a3ef0d4e0540c10c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 208, + 417, + 312 + ], + "blocks": [ + { + "bbox": [ + 307, + 208, + 417, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 208, + 417, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 417, + 312 + ], + "type": "image", + "image_path": "735330791fee70a416a935d899d1a4b6853334043371955a3b94519f0ab0ee19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 422, + 208, + 535, + 312 + ], + "blocks": [ + { + "bbox": [ + 422, + 208, + 535, + 312 + ], + "lines": [ + { + "bbox": [ + 422, + 208, + 535, + 312 + ], + "spans": [ + { + "bbox": [ + 422, + 208, + 535, + 312 + ], + "type": "image", + "image_path": "32834c1a9aa6bd0428a07e2d7f0e34fd816ef19acaf3aec0d1002f20ee189d7f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 345, + 287, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 163, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 547, + 198 + ], + "type": "text", + "content": "Table 5. Frame selection: (a) Sinkhorn selection results in better supervision. (b) We investigate all possible combinations of global and local representations for frame selection similarity matching. We found keeping the local text representation is crucial, and a combination of local and global representation leads to the best spatio-temporal grounding result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 345, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 48, + 345, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 345, + 287, + 399 + ], + "type": "table", + "html": "
GroundingYT Spatio-tempMiningYT TemporalYC-Inter. Spatial
only Local loss7.295.2355.29
only Global loss9.2819.1236.23
w/ Both loss19.4520.3358.35
", + "image_path": "af1bf75b2b108a23c817a10e71d814f9321d63288c47b4b189cb4bc55517c25f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 402, + 287, + 425 + ], + "lines": [ + { + "bbox": [ + 47, + 402, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 287, + 425 + ], + "type": "text", + "content": "Table 6. Loss ablations: both losses contribute to the final loss, and the existence of global loss helps the localization task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": "spatio-temporal tokens as the selection target, we sum over the scores with respect to each frame to acquire the frame similarity score. It shows that selecting frames based on Sinkhorn selection leads to consistently better results as it enforces more variety of visual concepts but also captures frames with possible groundable objects. It further shows that word tokens are more suitable than the global text CLS token for frame selection. Finally, we see that depending on the task (spatial vs. temporal), a local resp. global representation is better, and a combination of both works best for spatio-temporal grounding. We provide runtime analysis of such frame selection strategy in the appendix 9.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Global and local loss. As mentioned in the spatio-temporal evaluation, both features contribute to the final grounding result. We test the model by ablating out each loss. Table 6 shows that each loss not only contributes to the spatiotemporal grounding on the GYT, but also that the whole is more than the sum of its parts (losses) since this task requires both spatial and temporal detection. The reduced impact of the global loss in the case of YC-Inter is that this is a pure spatial grounding dataset (no background frames) without temporal detection, and the local loss plays a more critical role. We observe the same patterns in the temporal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 356, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 356, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 546, + 392 + ], + "type": "text", + "content": "grounding result for MYT, where spatial localization is not directly contributing to the final performance. We tried out the same ablation using in the S3D backbone in supplement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 402, + 414, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 414, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 414, + 415 + ], + "type": "text", + "content": "5.6. Qualitative results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 418, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 546, + 479 + ], + "type": "text", + "content": "We visualize our spatio-temporal result in Figure 5. For the GLIP model, we output the bounding box with the highest confidence score and visualize its center point. We found GLIP model focuses on the salient object while our model focuses more on human-object interaction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 494, + 378, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 378, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 378, + 506 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 509, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 546, + 665 + ], + "type": "text", + "content": "We presented an approach for learning spatio-temporal grounding with self-supervision and a new dataset: GroundingYoutube annotations, where we densely annotate spatiotemporal points/boxes from untrimmed multi-action videos. Our method includes a frame selection mechanism that identifies frames with groundable objects to adapt the learning process for untrimmed videos. Furthermore, we jointly learn global representations, which capture temporal information, and local representations learning fine-grained multimodal interactions between video and text. We conducted extensive experiments and our approach shows state-of-the-art performance in spatio-temporal grounding, as well as temporal and spatial grounding alone." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 667, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 667, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 667, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments: Nina Shvetsova is supported by the German Federal Ministry of Education and Research (BMBF) project STCL-01IS22067. Andrew Rouditchenko is supported by an NDSEG Fellowship and by the MIT-IBM Watson AI Lab. We thank the MIT-Satori cluster team for their technical support." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18426" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 5, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "text", + "content": "[2] Hassan Akbari, Svebor Karaman, Surabhi Bhargava, Brian Chen, Carl Vondrick, and Shih-Fu Chang. Multi-level multimodal common semantic space for image-phrase grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12476–12486, 2019. 2, 6, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. Advances in Neural Information Processing Systems, 33:25-37, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 369 + ], + "type": "text", + "content": "[5] Assaf Arbelle, Sivan Doveh, Amit Alfassy, Joseph Shtok, Guy Lev, Eli Schwartz, Hilde Kuehne, Hila Barak Levi, Prasanna Sattigeri, Rameswar Panda, et al. Detector-free weakly supervised grounding by separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1801-1812, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 479, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 555 + ], + "type": "text", + "content": "[8] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8012-8021, 2021. 1, 3, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 288, + 601 + ], + "type": "text", + "content": "[9] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. Advances in Neural Information Processing Systems, 34:28442-28453, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "text", + "content": "[10] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee Kenneth Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1884-1894, Florence, Italy, 2019. Association for Computational Linguistics. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 691 + ], + "type": "text", + "content": "[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. 2013. 2, 3, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "type": "text", + "content": "Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 163 + ], + "type": "text", + "content": "[13] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "text", + "content": "[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 306 + ], + "type": "text", + "content": "[15] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 2, 5, 6, 13, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "text", + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "text", + "content": "[17] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTATS, 2010. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 387, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 432 + ], + "type": "text", + "content": "[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "type": "text", + "content": "[19] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649–665, 2018. 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "text", + "content": "[20] H. Idrees, A. R. Zamir, Y. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 578 + ], + "type": "text", + "content": "[21] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 578, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 633 + ], + "type": "text", + "content": "[22] Yu-Gang Jiang, Jingen Liu, A Roshan Zamir, George Toderici, Ivan Laptev, Mubarak Shah, and Rahul Sukthankar. Thumos challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.3,14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[23] Yang Jin, yongzhi li, Zehuan Yuan, and Yadong MU. Embracing consistency: A one-stage approach for spatiotemporal video grounding. In Advances in Neural Information Processing Systems, 2022. 1, 2, 6, 7, 14, 15" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[24] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In Proceedings of the IEEE" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18427" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 4405-4413, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2015. 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 230 + ], + "type": "text", + "content": "[27] Giorgos Kordopatis-Zilos, Symeon Papadopoulos, Ioannis Patras, and Ioannis Kompatsiaris. Visil: Fine-grained spatiotemporal video similarity learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6351-6360, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "text", + "content": "[28] Hilde Kuehne, Ahsan Iqbal, Alexander Richard, and Juergen Gall. Mining youtube-a dataset for learning fine-grained action concepts from webly supervised video data. 2019. 1, 2, 3, 5, 6, 7, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "type": "text", + "content": "[29] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "text", + "content": "[30] Ang Li, Meghan Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv preprint arXiv:2005.00214, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 367, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 433 + ], + "type": "text", + "content": "[31] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 6, 7, 12, 15" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 489 + ], + "type": "text", + "content": "[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1845-1855, 2021. 2, 6, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 491, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 534 + ], + "type": "text", + "content": "[33] Yikang Li, Jenhao Hsiao, and Chiuman Ho. Videoclip: A cross-attention model for fast video-text retrieval task with image clip. In Proceedings of the 2022 International Conference on Multimedia Retrieval, pages 29–33, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "text", + "content": "[35] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[36] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 3, 6, 7, 12, 15" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In arXiv preprint arXiv:1301.3781, 2013. 6, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 130, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 196 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 12, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 243, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 307 + ], + "type": "text", + "content": "[40] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv preprint arXiv:2006.09199, 2020.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "type": "text", + "content": "[41] Jing Shi, Jia Xu, Boqing Gong, and Chenliang Xu. Not all frames are equal: Weakly-supervised video grounding with contextual similarity and visual clustering losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10444-10452, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "type": "text", + "content": "[42] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In CVPR, 2022. 1, 3, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "type": "text", + "content": "[43] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba Heilbron, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. arXiv preprint arXiv:2112.00431, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 566 + ], + "type": "text", + "content": "[45] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1533–1542, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "text", + "content": "[46] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 5, 6, 7, 12, 15" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "type": "text", + "content": "[47] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[48] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18428" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[51] Philippe Weinzaepfel, Xavier Martin, and Cordelia Schmid. Human action localization with sparse spatial supervision. arXiv preprint arXiv:1605.05197, 2016. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "text", + "content": "[52] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence, 2011. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "text", + "content": "[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "type": "text", + "content": "[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 1, 2, 6, 7, 14, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "type": "text", + "content": "[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[57] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "text", + "content": "[58] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[59] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10668-10677, 2020. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[60] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In Pro-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 385 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[61] Yang Zhao, Zhou Zhao, Zhu Zhang, and Zhijie Lin. Cascaded prediction network via segment tree for temporal video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4206, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "text", + "content": "[62] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2, 3, 6, 7, 12, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 261 + ], + "type": "text", + "content": "[63] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In British Machine Vision Conference, 2018. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "text", + "content": "[64] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "text", + "content": "[65] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[66] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. 2019. 1, 2, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18429" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_content_list.json b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea1e247b8a5290922141a9198eb82f88200776d6 --- /dev/null +++ b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_content_list.json @@ -0,0 +1,1767 @@ +[ + { + "type": "text", + "text": "What You See is What You GAN: Rendering Every Pixel for High-Fidelity Geometry in 3D GANs", + "text_level": 1, + "bbox": [ + 163, + 128, + 805, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alex Trevithick\\*, Matthew Chan, Towaki Takikawa, Umar Iqbal, Shalini De Mello, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano", + "bbox": [ + 91, + 212, + 877, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA $^{2}$ University of California, San Diego", + "bbox": [ + 285, + 258, + 684, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2988b5876bd21cea5c88f1ef75f51ab3a2b45ef962f3edcd7ccd6791a48f670f.jpg", + "image_caption": [ + "Figure 1. Left: Our results. The split view in the middle demonstrates the high degree of agreement between our 2D rendering and corresponding 3D geometry. Our method can learn fine-grained 3D details (e.g., eyeglass frame and cat's fur) that are geometrically well-aligned to 2D images without multiview or 3D scan data. Right: Comparison with EG3D [7]. Our tight SDF prior provides smooth and detailed surfaces on the face and hat while EG3D exhibits geometry artifacts and discrepancies between geometry and rendering. Please see Fig. 5 and the accompanying video for more examples, and Fig. 6 for comparison to other baselines." + ], + "image_footnote": [], + "bbox": [ + 81, + 301, + 470, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4704948b53dff98fedc81cc46dc44fbda5219fdd47888f8a2f5693b07902e656.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 301, + 893, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 637, + 313, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D-aware Generative Adversarial Networks (GANs) have shown remarkable progress in learning to generate multi-view-consistent images and 3D geometries of scenes from collections of 2D images via neural volume rendering. Yet, the significant memory and computational costs of dense sampling in volume rendering have forced 3D GANs to adopt patch-based training or employ low-resolution rendering with post-processing 2D super resolution, which sacrifices multiview consistency and the quality of resolved geometry. Consequently, 3D GANs have not yet been able to fully resolve the rich 3D geometry present in 2D images. In this work, we propose techniques to scale neural volume rendering to the much higher resolution of native 2D images, thereby resolving fine-grained 3D geometry with", + "bbox": [ + 73, + 669, + 473, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unprecedented detail. Our approach employs learning-based samplers for accelerating neural rendering for 3D GAN training using up to 5 times fewer depth samples. This enables us to explicitly \"render every pixel\" of the full-resolution image during training and inference without post-processing superresolution in 2D. Together with our strategy to learn high-quality surface geometry, our method synthesizes high-resolution 3D geometry and strictly view-consistent images while maintaining image quality on par with baselines relying on post-processing super resolution. We demonstrate state-of-the-art 3D gemetic quality on FFHQ and AFHQ, setting a new standard for unsupervised learning of 3D shapes in 3D GANs.", + "bbox": [ + 496, + 638, + 893, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 842, + 632, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training 3D generative models from the abundance of 2D images allows the creation of 3D representations of real-world objects for content creation and novel view synthe", + "bbox": [ + 496, + 867, + 893, + 912 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This project was initiated and substantially carried out during an internship at NVIDIA.", + "bbox": [ + 75, + 886, + 468, + 912 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22765", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sis [59]. Recently, 3D-aware generative adversarial networks (3D GANs) [6, 7, 19, 43, 45, 49, 53, 69, 70, 74, 76] have emerged as a powerful way to learn 3D representations from collections of 2D images in an unsupervised fashion. These methods employ differentiable rendering to compare rendered 3D scenes with 2D data using adversarial training [18]. Among the various 3D representations, Neural Radiance Fields (NeRF) [38] have become a popular choice among recent successful 3D GANs. However, the significant computational and memory cost of volume rendering has prevented 3D GANs from scaling to high-resolution output. For instance, generating a single $512 \\times 512$ image via volume rendering requires evaluating as many as 25 million depth samples, if 96 depth samples are used per ray using importance sampling [7, 10, 38, 53]. Given that GAN training typically requires rendering tens of millions of images, the training process could require evaluating hundreds of trillions of depth samples.", + "bbox": [ + 76, + 90, + 472, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During training, all intermediate operations must be stored in GPU memory for every depth sample for the backward pass. Therefore, existing methods resort to working on patches [10, 49, 53] or adopting a low resolution neural rendering combined with post-processing 2D super resolution (SR) [7, 19, 43, 45, 70]. However, patch-based methods have limited receptive fields over scenes, leading to unsatisfactory results, and the hybrid low-resolution rendering and SR scheme inevitably sacrifices the multiview consistency and the accuracy of 3D geometry. While many techniques have been developed to improve the image quality of 3D GANs to match that of 2D GANs, the challenge of resolving the corresponding high-resolution 3D geometry remains unsolved (see Figs. 1 and 6 for our results and comparison to the current state-of-the-art).", + "bbox": [ + 75, + 353, + 472, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scaling 3D GANs to operate natively at the 2D pixel resolution requires a novel approach for sampling. Fig. 2 compares the state-of-the-art 3D GAN, EG3D model, trained with and without SR. EG3D employs 96 dense depth samples in total using two-pass importance sampling [38] during training, which requires half a terabyte of GPU memory at $256 \\times 256$ resolution, making scaling to higher resolutions infeasible. Furthermore, Fig. 2 demonstrates that using 96 dense samples still results in undersampling, as evidenced by the speckle noise patterns visible in the zoomed-in view, leading to considerably worse FID (inset in Fig. 2). 3D GANs relying on post-processing SR layers can repair these undersampling artifacts at the cost of a high-fidelity 3D representation.", + "bbox": [ + 75, + 569, + 472, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address the challenge of scaling neural volume rendering to high resolutions by explicitly rendering every pixel, ensuring that \"what you see in $2D$ , is what you get in $3D$ — generating an unprecedented level of geometric details as well as strictly multiview-consistent images. Our contributions are the following:", + "bbox": [ + 75, + 772, + 472, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce an SDF-based 3D GAN to represent high", + "bbox": [ + 76, + 862, + 468, + 877 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/446d301cc6d80354a008a7fb6b403bedb296d7d069fc2f74f7098c5b6d5338ed.jpg", + "image_caption": [ + "Figure 2. Samples from EG3D 256 model. Right: Volume rendering with 48 coarse samples and 48 fine samples per ray with two-pass importance sampling [38] results in undersampling, leading to noticeable noisy artifacts. Left: These artifacts are repaired by super resolution (SR). An unsharp mask has been applied to the zoomed views for presentation purposes." + ], + "image_footnote": [], + "bbox": [ + 558, + 87, + 834, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "frequency geometry with spatially-varying surface tightness that increases throughout training (subsection 4.1 and 4.5), in turn facilitating low sample rendering.", + "bbox": [ + 509, + 398, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a generalizable learned sampler conditioned on cheap low-resolution information to enable full-resolution rendering during training for the first time (subsection 4.2 and 4.3).", + "- We show a robust sampling strategy for the learned sampler (subsection 4.4) that produces stable neural rendering using significantly fewer depth samples (see Fig. 8). Our sampler can operate with just 20 samples per ray compared to existing 3D GANs which must use at least 96 samples per ray.", + "- Together, our contributions result in the state-of-the-art geometry for 3D GANs while rendering with quality on par with SR baselines (see Fig. 1). For more results, see Fig. 5 and for comparison to other baselines, see Fig. 6." + ], + "bbox": [ + 500, + 441, + 890, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 662, + 640, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by reviewing the prior-arts of 3D generative models and their current shortcomings. We then cover foundational techniques for 3D geometry representation and neural rendering from which we take inspiration. We then discuss existing methods for accelerating neural volume rendering, which usually operate per-scene.", + "bbox": [ + 496, + 688, + 890, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Generative Models", + "text_level": 1, + "bbox": [ + 500, + 786, + 707, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Just like 2D GANs, 3D-aware GANs train from a collection of 2D images, but employ a 3D representation and differentiable rendering to learn 3D scenes without requiring multiview images or ground truth 3D scans. Some of the most successful works use a neural field [68] in combination with a feature grid [7] as their 3D representation, and use neural volume rendering [38] as the differentiable renderer.", + "bbox": [ + 496, + 809, + 890, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Triplane resolution is doubled to compensate for the loss of capacity from removing the SR layers.", + "bbox": [ + 76, + 886, + 468, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22766", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to the significant memory and computational cost of neural volume rendering, many previous works perform rendering at low-resolution and rely on a 2D post-processing CNN [7, 19, 43, 45, 70], which hallucinates the high-frequency details in a view-inconsistent manner while sacrificing 3D consistency and the quality of the resolved 3D geometry.", + "bbox": [ + 75, + 90, + 470, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure strict 3D consistency, other previous works seek to render at high-resolutions and propose techniques to address the prohibitive computational costs. One line of work leverages the sparse nature of 3D scenes to speed up rendering, in particular, structures such as 2D manifolds [14, 67], multiplane images [75] and sparse voxels [50]. Although they are more efficient, sparse representations provide only coarse [50] or category-specific [14, 67] acceleration structures, which poses constraints on the diversity and viewing angles of the generated scenes. Our sampling-based method, on the other hand, generalizes to every new scene and adaptively accelerates rendering on a per ray basis. Another line of work enables high-resolution rendering with patch-based training [10, 53]. In particular, Mimic3D [10] achieves significantly improved 2D image quality, but the patch-based training limits the receptive fields, and the generated geometry does not faithfully represent the 2D data due to the patch-wise perceptual loss. Our method renders the entire image at once and the resulting geometry is aligned with the rendering (see Figs. 1 and 6).", + "bbox": [ + 75, + 193, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, a new family of generative approaches using diffusion models has been proposed to tackle conditional tasks including novel view synthesis [8, 58] and text-based 3D generation[63]. Most of these 3D-aware diffusion models combine a 3D inductive bias modeled via neural field representations and a 2D image denoising objective to learn 3D scene generation. While these models enable unconditional 3D generation, they require multiview images [24, 55, 63] or 3D data, such as a point cloud [42]. Score Distillation Sampling [47] may be used for distillation from a pre-trained 2D diffusion model when only monocular 2D data is available, but diffusion models incur significant computational costs due to their iterative nature and most of the existing methods require optimization per scene [9, 20, 33, 61, 66].", + "bbox": [ + 75, + 484, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Learning High-Fidelity Geometry", + "text_level": 1, + "bbox": [ + 76, + 712, + 372, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior works on 3D GANs have typically represented the geometry as a radiance field [38], which lacks a concrete definition for where the surface geometry resides in the field, resulting in bumpy surfaces. A number of works [44, 62, 71] have proposed alternate representations based on implicit surfaces (such as signed distance functions, or SDFs) that can be used with neural volume rendering. In these works, the implicit surface is typically softened by a parameter for volume rendering.", + "bbox": [ + 75, + 736, + 467, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other works [32, 64] improve on these implicit surface representations by leveraging feature grids for higher computational efficiency and resolution. Adaptive Shells [65]", + "bbox": [ + 75, + 867, + 467, + 912 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "further improve on quality by making the softness parameter spatially-varying, as many objects have hard boundaries only in certain parts. We use an implicit surface representation based on VolSDF [72], and leverage a spatially-varying parameter similar to Adaptive Shells [65] to control the softness of the surface, as humans and animals benefit from both hard surfaces (e.g. skin, eyes) and soft, volumetric representations (e.g. hair). Although other works such as StyleSDF [45] have similarly leveraged implicit surfaces in a 3D GAN framework, the lack of spatial-variance and high-resolution rendering led to over-smoothed geometry not faithful to the rendered images.", + "bbox": [ + 496, + 90, + 890, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Accelerating Neural Volume Rendering", + "text_level": 1, + "bbox": [ + 498, + 276, + 836, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned in Section 2.1, accelerating 3D GANs typically relies on acceleration structures such as octrees [31, 56, 73], which in generative settings [50] are limited due to computational requirements. In another direction, GRAM-based methods [14, 67] are limited due to sharing one structure across all samples. Instead, we look to a class of methods that do not rely on an acceleration structure. Some of these works learn a per-scene sampling prior on a per-ray basis using a binary classifier [41], a density estimator [30, 46], a coarse proxy from a 3D cost volume [34], or an interval length estimator [35] on discrete depth regions along the ray. Other works use importance sampling [21, 38] to sample additional points. We take inspiration from works on density estimators [30] and propose to learn a scene-conditional proposal network that generalizes across scenes instead of being category-specific or optimized per-scene.", + "bbox": [ + 496, + 300, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are also other methods to accelerate rendering by utilizing more efficient representations, such as gaussian splats [29] and light field networks [52]. More efficient feature grids [40, 57] based on hashing can also be used to accelerate rendering. However, mapping to these representations in a GAN framework is not straightforward. In contrast, our sampling strategy can be used for any NeRF representation.", + "bbox": [ + 496, + 547, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 667, + 627, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin with background on the methodology of the state-of-the-art 3D-aware GANs as our method relies on a similar backbone for mapping to 3D representations. 3D GANs typically utilize a StyleGAN-like [27] architecture to map from a simple Gaussian prior to the conditioning of a NeRF, whether that be an MLP [19, 45], MPI [75], 3D feature grid [50], manifolds [14, 67] or triplane [7, 10, 53]. We inherit the latter triplane conditioning for its high expressivity and efficiency, in which three axis-aligned 2D feature grids $(f_{xy}, f_{xz}, f_{yz})$ , provide NeRF conditioning by orthogonal projection and interpolation. As in the previous methods, the mapping and synthesis networks from StyleGAN2 can easily be adapted to create the 2D triplane representation from noise $z \\in \\mathbb{R}^{512}$ . Specifically, $\\mathbf{w} = \\mathrm{Mapping}(z)$ conditions the creation of the triplane $T(\\mathbf{w}) \\in \\mathbb{R}^{3 \\times 32 \\times 512 \\times 512}$", + "bbox": [ + 496, + 694, + 890, + 912 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22767", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ec68d5485ab607a8284f4b908b6f86b7c472661c38c09857ca0928315e1b673.jpg", + "image_caption": [ + "Figure 3. Here we show our proposed pipeline and its intermediate outputs. Beginning from the triplane $T$ , we trace uniform samples to probe the scene, yielding low-resolution $I_{128}$ and weights $P_{128}$ . These are fed to a CNN which produces high-resolution proposal weights $\\hat{P}_{512}$ (weights are visualized as uniform level sets). We perform robust sampling and volume render to get the final image $I_{512}$ and the surface variance $B$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 890, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from a Synthesis network, corresponding to three axis-aligned 2D feature grids of spatial resolution $512 \\times 512$ and feature size 32.", + "bbox": [ + 75, + 294, + 468, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To create high-fidelity geometry, our method builds upon VolSDF [72]: Instead of directly outputting the opacity $\\sigma$ of a point $x\\in \\mathbb{R}^3$ , an SDF value $s$ is output and transformed to $\\sigma$ using a Laplacian CDF:", + "bbox": [ + 75, + 337, + 470, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\frac {1}{\\beta} \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\exp \\left(\\frac {s}{\\beta}\\right) & \\text {i f} s \\leq 0 \\\\ 1 - \\frac {1}{2} \\exp \\left(- \\frac {s}{\\beta}\\right) & \\text {i f} s > 0 \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 405, + 468, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is the variance of the Laplacian distribution governing the \"tightness\" of the representation. One distinct benefit of an SDF-based representation is the ease of extracting the surface. StyleSDF [45] also utilizes this intermediate geometry representation without a triplane, enforcing the usual Eikonal constraint.", + "bbox": [ + 75, + 459, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the triplane conditioning and Eq. 1, we can assign each point in the volume with its opacity $\\sigma$ and radiance $c$ using a lightweight MLP. For a given camera ray $r(t) = o + td$ , we approximate the volumetric rendering integral $C(r)$ [36] by sampling ray distances $t_i$ with their corresponding $\\sigma_i$ and $c_i$ before computing", + "bbox": [ + 75, + 546, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (r) = \\sum_ {i = 1} ^ {N} w _ {i} c _ {i} \\text {w h e r e} w _ {i} = T _ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 643, + 468, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}\\right) \\text {a n d} \\delta_ {i} = t _ {i + 1} - t _ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 686, + 467, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $T_{i}$ denotes the accumulated transmittance and $\\delta_{i}$ is the distance between adjacent samples along the ray.", + "bbox": [ + 75, + 743, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is possible to develop a more efficient estimator for this sum with fewer samples by using importance sampling techniques in computer graphics. Typically, one computes a piecewise constant probability distribution $p_j = \\frac{\\hat{w}_j}{\\sum_j \\hat{w}_j}$ , where $j$ refers to the $j^{\\text{th}}$ bin or region, and $\\hat{w}_j$ is an estimate of $w_j$ for that region, for example obtained by explicitly tracing coarse samples. For a given $t$ , we first find the region $j(t)$ and then set $p(t) = p_j$ . From this, one can compute a (piecewise linear) cumulative distribution function or CDF", + "bbox": [ + 75, + 773, + 470, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Phi (t)$ which has a range from 0 to 1. We can then perform inverse CDF sampling to define the sample points,", + "bbox": [ + 498, + 294, + 890, + 324 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} = \\Phi^ {- 1} (u _ {i}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 328, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u_{i}$ is a random number from 0 to 1 (sorted to be an increasing sequence).", + "bbox": [ + 496, + 351, + 890, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion We improve on previous works such as NeRF [38] and EG3D [7] by stratifying2 the random numbers $u_{i}$ during training; this leads to significantly lower rendering variance, especially at low sample counts $N$ [39]. We also develop a neural method to predict a good distribution $p_{j}$ (and hence $\\Phi$ ) for importance sampling at high spatial resolution, without needing to exhaustively step through the ray.", + "bbox": [ + 496, + 382, + 890, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 508, + 589, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe our method beginning with our SDF-based NeRF parametrization (subsection 4.1). We then overview how we render at high-resolution in three stages: first, a low-resolution probe into the 3D scene (subsection 4.2); second a high-resolution CNN proposal network (subsection 4.3); and third a robust sampling method for the resultant proposals (subsection 4.4). Next we describe regularizations (subsection 4.5) for stable training, and finally our entire training pipeline (subsection 4.6).", + "bbox": [ + 496, + 534, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Mapping to a 3D Representation", + "text_level": 1, + "bbox": [ + 498, + 671, + 785, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beginning from a noise vector $z$ , we synthesize the initial triplane $T'$ with StyleGAN [25] layers as detailed in Sec. 3. In contrast to previous methods, we then generate more expressive triplane features $T$ with an extra synthesis block for each orthogonal plane: $f_{ij} = \\mathrm{SynthesisBlock}(f_{ij}')$ where $ij \\in \\{xy, xz, yz\\}$ and $f'$ are the features of $T'$ . This design choice allows disentanglement between the intermediate triplane features as plane-specific kernels can attend to the features in each separate plane.", + "bbox": [ + 496, + 694, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the triplane $T$ (the left side of Fig. 3) and a point $x \\in \\mathbb{R}^3$ , we utilize an MLP to map to the SDF value $s$ , variance $\\beta$ and geometry features $f_{geo}$ :", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(s, \\beta , f _ {g e o}) = \\operatorname {M L P} _ {\\mathrm {S D F}} (\\operatorname {P o s E n c} (x), T _ {x}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 875, + 890, + 892 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2 dividing the unit interval into bins and taking a sample from each bin.", + "bbox": [ + 514, + 898, + 890, + 911 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "22768", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where PosEnc is the positional encoding from NeRF [38] and $T_{x}$ are the features corresponding to $x$ gathered from $T$ by projecting $x$ to each of the axis-aligned planes and taking the Hadamard product of the three resultant vectors [16]. We initialize the ReLU $\\mathrm{MLP}_{\\mathrm{SDF}}$ as in SAL [4] to ensure an approximately spherical SDF in the early part of training. Additionally note that $\\beta$ varies spatially in the volume unlike [45, 72], allowing us to regularize its values later on.", + "bbox": [ + 75, + 90, + 467, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using Eq. 1, we transform $s$ and $\\beta$ into opacity $\\sigma$ . We can now predict the radiance with a separate $\\mathrm{MLP_c}$ conditioned on the geometry features and viewing direction $v$ as", + "bbox": [ + 75, + 207, + 467, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc = \\mathrm {M L P} _ {\\mathrm {c}} (\\operatorname {P o s E n c} (v), f _ {g e o}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 267, + 467, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that unlike EG3D [7], we condition radiance on the viewing direction, allowing a more expressive generator. Thus, given a triplane, we can render any pixel by computing $\\sigma$ and $c$ for points along the ray to approximate the volumetric rendering integral as described in Sec. 3.", + "bbox": [ + 75, + 292, + 467, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. High-Resolution Proposal Network", + "text_level": 1, + "bbox": [ + 76, + 375, + 382, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now have our mapping from a latent code $z$ to a 3D NeRF representation. However, volumetric rendering of NeRFs at higher resolutions requires extremely large numbers of samples, and thus both memory and time. Instead of naive dense sampling at a high resolution, we propose to leverage low-resolution renderings to cheaply probe the 3D representation (visualized on the left of Fig. 3) for the creation of proposal distributions at high-resolution. Given a target camera and triplane $T$ , we first trace 192 coarse samples at low-resolution $(128 \\times 128)$ to compute a low-resolution RGB image $I_{128} \\in \\mathbb{R}^{3 \\times 128 \\times 128}$ and a tensor of weights $P_{128} \\in \\mathbb{R}^{192 \\times 128 \\times 128}$ (visualized after low-resolution rendering in Fig. 3). Each 192-dimensional vector corresponds to a piecewise constant PDF with CDF $\\Phi$ as seen in Eq. 3.", + "bbox": [ + 75, + 398, + 467, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditioned on the low-resolution probe, we predict a tensor of proposal volume rendering weights at the high-resolution $(512 \\times 512)$ :", + "bbox": [ + 75, + 617, + 467, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {5 1 2} = \\operatorname {S o f t m a x} \\left(\\operatorname {C N N} \\left(P _ {1 2 8}, I _ {1 2 8}\\right)\\right) \\in \\mathbb {R} ^ {1 9 2 \\times 5 1 2 \\times 5 1 2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 672, + 467, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where CNN is a lightweight network that up-samples the low-resolution weights, Softmax produces discrete distributions along each ray, and the $\\hat{}$ denotes that this is an estimated quantity. This corresponds to the Proposal in Fig. 3 and the yellow distribution in Fig. 4. Note that allocating 192 samples at $128 \\times 128$ is equivalent to allocating just 12 at $512 \\times 512$ .", + "bbox": [ + 75, + 702, + 467, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Supervising the Proposal Network", + "text_level": 1, + "bbox": [ + 76, + 815, + 375, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having described the input and output of our high-resolution proposal network, we now show its supervision. From the target camera, we can also trace 192 coarse samples at high resolution for a small $64 \\times 64$ patch, giving us a ground truth tensor of volume rendering weights", + "bbox": [ + 75, + 838, + 467, + 912 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$P_{\\mathrm{patch}} \\in \\mathbb{R}^{192 \\times 64 \\times 64}$ . We then prepare this tensor for supervision by computing:", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} _ {\\text {p a t c h}} = \\text {N o r m a l i z e} (\\text {S u p p r e s s} (\\text {B l u r} (P _ {\\text {p a t c h}}))) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 130, + 888, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Blur applies a 1D Gaussian kernel to the input distributions, $\\operatorname{Suppress}(x) = x$ if $x \\geq 5e - 3$ and 0 otherwise, and Normalize is L1 normalization to create a valid distribution. This corresponds to the patch loss in Fig. 3 and the purple distribution in Fig. 4. These operations create less noisy distributions to facilitate accurate learning of the high-frequency integrand which may be undersampled in the coarse pass.", + "bbox": [ + 496, + 157, + 890, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can then compare the predicted and cleaned ground truth distributions with a cross-entropy loss:", + "bbox": [ + 498, + 273, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s a m p l e r}} = \\text {C r o s s E n t r o p y} (\\bar {P} _ {\\text {p a t c h}}, \\hat {P} _ {\\text {p a t c h}}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 314, + 888, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{P}_{\\mathrm{patch}}$ is the corresponding patch of weights in $\\hat{P}_{512}$ and CrossEntropy denotes the average cross-entropy between all pairs of pixelwise discrete distributions; for each pair $(\\bar{p},\\hat{p})$ , we compute $\\sum -\\bar{p}_j\\log \\hat{p}_j$ . Since we only need to compute this supervision for a small patch, the overhead of sampler training is not significant.", + "bbox": [ + 496, + 342, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Robustly Sampling from the Proposal Network", + "text_level": 1, + "bbox": [ + 498, + 439, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having shown how to train and predict high-resolution distributions, we now overview how to sample the resultant proposals. As seen in Fig. 4, the proposals are often slightly off; this is due to the high frequency nature of the underlying integrand in blue.", + "bbox": [ + 496, + 462, + 890, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to utilize the information from the sampler, we propose to filter predicted PDFs for better estimation. Specifically, for each discrete predicted PDF $\\hat{p}$ , we compute the smallest set of bins whose probability exceeds a threshold $\\tau = 0.98$ : We find the smallest subset $I \\subseteq \\{1,2,\\dots,192\\}$ such that", + "bbox": [ + 496, + 535, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i \\in I} \\hat {p} _ {i} \\geq \\tau . \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 633, + 888, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This operation resembles nucleus sampling in NLP [23]. We define our sampling PDF $q$ with probability $q_{i} = \\frac{1}{|I|}$ if $i\\in I$ and 0 otherwise (the green distribution in Fig. 4).", + "bbox": [ + 496, + 674, + 888, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each PDF $q$ , we compute its CDF $\\Phi$ and perform stratified inverse transform sampling to create the samples (illustrated as adaptive sampling near the surface in Fig. 3). We refer to these samples as robust stratified samples. In practice, on top of the 12 samples from the coarse probe (for the high-resolution image), we take an additional 18 samples per pixel adaptively based on the variance of the predicted distributions. The details are given in the supplement.", + "bbox": [ + 496, + 720, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Regularization for High-Resolution Training", + "text_level": 1, + "bbox": [ + 498, + 859, + 877, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to render accurately under low sample budget per ray, we desire the surface to be tight, i.e., the set of points", + "bbox": [ + 496, + 882, + 890, + 912 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "22769", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dfdb4927f0d76df639c0284c69bbae79d07a2a3bdf7d321886f4b05c0f9f86cb.jpg", + "image_caption": [ + "Figure 4. We visualize the volume rendering PDFs for the green pixel in the images on the right along with sampling methods. The ground truth distribution in blue is bimodal due to the discontinuous depth. Without stratification, the samples from the predicted yellow PDF completely miss the second mode. Stratification reduces the variance, yet also misses the second mode. Our robust stratified samples hit both modes despite the inaccurate predictions. The supervision PDF is visualized in purple as well." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 472, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "along the ray, which contribute to the accumulated radiance to be small. To accomplish this, we introduce a regularization for the spatially-varying $\\beta$ values. Replacing the $c_{i}$ with the intermediate $\\beta_{i}$ in Eq. 2, we can volume render an image of scalars, $B\\in \\mathbb{R}^{512\\times 512}$ (seen on the right of Fig. 3). During training, we regularize $B$ to be small so that the surface tightens:", + "bbox": [ + 75, + 430, + 472, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s u r f a c e}} = \\sum_ {h w} \\left(B _ {h w} - B _ {\\text {t a r g e t}}\\right) ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 544, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B_{\\mathrm{target}}$ is a scalar quantity annealed towards $\\epsilon > 0$ during optimization. Note that this results in significantly more pronounced and smooth geometry (see Fig. 7).", + "bbox": [ + 75, + 588, + 468, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training with the proposal network, the rendering networks only receive gradients for points very near the surface (which have high density). We find that this can lead to undesirable density growing from the background into the foreground. In order to combat this, we leverage the low-resolution information from $I_{128}$ , which is computed for uniform samples along the ray, and thus not concentrated at the surface. Considering the SDF values intermediately computed for rendering, $S = S_{128} \\in \\mathbb{R}^{192 \\times 128 \\times 128}$ , we enforce the SDF decision boundary by minimizing the SDF likelihood under a Laplacian distribution similar to [45, 51]:", + "bbox": [ + 75, + 632, + 470, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {d e c}} = \\sum_ {z h w} \\exp (- 2 | S _ {z h w} |). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 803, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. The Training Pipeline", + "text_level": 1, + "bbox": [ + 76, + 844, + 282, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to begin making use of the learned sampler for high-resolution rendering, we need a good NeRF representation from which to train it. Concurrently, we also need a", + "bbox": [ + 75, + 867, + 470, + 912 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "good sampler in order to allow NeRF to render at $512 \\times 512$ , the input resolution to the discriminator $D$ .", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To solve this issue, in the early stages of training, we first learn a low-resolution (e.g., $64 \\times 64$ ) 3D GAN through the standard NeRF sampling techniques. We bilinearly up-sample our low-resolution renderings to $512 \\times 512$ and blur the real images to the same level. After converging at the lower resolution, we introduce the sampler training (subsection 4.3). Concretely, we not only render low-resolution images with standard sampling, but also render sampler inputs $P_{128}$ and supervision patches $P_{\\mathrm{patch}}$ . This results in a good initialization for the sampler.", + "bbox": [ + 496, + 119, + 893, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Having learned an initial low-resolution 3D GAN and high-resolution sampler, we transition to rendering with the sampler predictions. The high-resolution proposals $\\hat{P}_{512}$ are downsampled to the current rendering resolution, which is progressively increased to the full $512 \\times 512$ resolution during training. After introducing all losses, we optimize the parameters of the generator $G$ to minimize the following:", + "bbox": [ + 496, + 265, + 893, + 369 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {a d v}} + \\lambda_ {\\text {s a m p l e r}} L _ {\\text {s a m p l e r}} + \\lambda_ {\\text {s u r f a c e}} L _ {\\text {s u r f a c e}} + \\lambda_ {\\text {d e c}} L _ {\\text {d e c}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 378, + 890, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{adv}}$ is the standard GAN loss [18] Softplus $(-D(G(z))$ . Note that we do not enforce the Eikonal constraint as we did not see a benefit. The discriminator $D$ is trained with R1 gradient regularization [37] whose weight is adapted as the resolution changes. Generator and discriminator pose conditioning follow EG3D [7]. The details of all hyperparameters and schedules are presented in the supplementary material.", + "bbox": [ + 496, + 402, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 532, + 586, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We benchmark on two standard datasets for 3D GANs: FFHQ [25] and AFHQv2 Cats [11, 28] both at resolution $512 \\times 512$ . We use the camera parameters extracted by EG3D [7] for conditioning and rendering. Our AFHQ model is finetuned from our FFHQ model using adaptive data augmentation [26]. For more results, please see the accompanying video.", + "bbox": [ + 496, + 556, + 890, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 667, + 638, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our methods against state-of-the-art 3D GAN methods including those that use low-resolution neural rendering and 2D post-processing CNN super resolution: EG3D [7], MVCGAN [74], and StyleSDF [45]; and methods that operate entirely based on neural rendering: Mimic3D [10], Epigraph [53], GMPI [75], and GramHD [67].", + "bbox": [ + 496, + 690, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results Fig. 5 shows the curated samples generated by our method with FFHQ and AFHQ, demonstrating photorealistic rendering as well as high-resolution details that align with the 2D images. Fig 6 provides qualitative comparisons to baselines. EG3D shows significant artifacts and cannot resolve high-resolution geometry since it performs neural rendering at only $128 \\times 128$ . Mimic3D", + "bbox": [ + 496, + 809, + 890, + 912 + ], + "page_idx": 5 + }, + { + "type": "footer", + "text": "22770", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2345f918d5836ccee359e45e6a67f16fffa1dad7ebfb0f2ee0ff25cbc681ba4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 87, + 486, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1c62382a665977f32737db82537e1c1915398b2997a89d9799e277261f3989f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 890, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8edc5d0825c57234fa9e0b6a1419ab5fdb01aa96cb522e16778987d5113ff0f.jpg", + "image_caption": [ + "Figure 5. Curated samples on FFHQ and AFHQ. Our method can resolve high-fidelity geometry (e.g., eyeglasses) and fine-grained details (e.g., cat's fur) as seen in the geometry and normal map." + ], + "image_footnote": [], + "bbox": [ + 91, + 261, + 483, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4677efd6da01774ab35682c2e5a2f3f8ff815710ef99b9386a0e5456cd29ec47.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons on FFHQ with EG3D [7], Mimic3D [10] and Epigraph [53]. EG3D performs neural rendering at resolution $128 \\times 128$ and relies on $4 \\times$ super resolution to generate images. On the right, Mimic3D and Epigraph directly generate the image via neural rendering. While all other baselines use up to 192 dense depth samples per ray, our method can operate at 30 samples per ray." + ], + "image_footnote": [], + "bbox": [ + 486, + 261, + 879, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66224bbeaeefdc70b6533e3e531da5e32aac022db1f22e70a0f9a82ce81718ee.jpg", + "image_caption": [ + "Figure 7. Ablation study on the effect of beta regularization." + ], + "image_footnote": [], + "bbox": [ + 78, + 563, + 468, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and Epigraph render all pixels with neural rendering, but the patch-based nature of these methods harm the overall 3D geometry (e.g., distorted face and missing ear). Our method provides both high-fidelity 3D shape (e.g., well-defined ears and isolated clothing collar) and high-resolution details.", + "bbox": [ + 75, + 717, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. Tabs. 1 and 2 provide quantitative comparisons against baselines. We measure the image quality with Fréchet Inception Distance (FID) [22]. We assess the quality of the learned geometry with a face-specific Normal FID (FID-N) [15]. We render 10.79k normal maps from the NPHM [17] dataset by solving for the approximate alignment between the average FFHQ 2D landmarks", + "bbox": [ + 75, + 809, + 470, + 912 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the provided 3D landmarks. Examples are given in the supplement. For each mesh, we render two views with approximately 20 degrees of pitch variation relative to the front of the face. These views are processed to remove the background with facer [13]. For each baseline, we render 10k normal maps and remove their background using the predicted mask from facer on the rendered image. We compute the FID between the two sets of images. Finally, we also evaluate the flatness of the learned representations with non-flatness score (NFS) [54].", + "bbox": [ + 496, + 564, + 890, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our results show the state-of-the-art image quality among the methods that operate only with neural rendering, while achieving FID comparable to the state-of-the-art SR-based method, EG3D. Our geometry quality outperforms all existing methods as indicated by our state-of-the-art FID-N. Additionally, our high NFS scores show the 3D aspect of our geometry. However, since NFS simply measures the variations of depth as a measure of non-flatness, it does not quantify the quality of geometry above a certain threshold.", + "bbox": [ + 496, + 710, + 892, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 857, + 653, + 873 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without our surface tightness regularization (Eq. 10), the SDF surface may get fuzzy, resulting in a less clean surface", + "bbox": [ + 500, + 882, + 890, + 912 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22771", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/12f9cf8c6507398f819580406c57c41225410c9e9bcb97e20bc173a4ccfcd346.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓FID-N↓NFS↑
FFHQ-512Epigraph [7]9.92†67.3333.95
Mimic3D [10]5.3764.9716.76
GRAM-HD [67]12.2†*--
Ours4.9760.7629.35
w SREG3D [7]4.7063.0217.54
StyleSDF [45]11.19†87.4222.75
MVCGAN [74]13.4†*--
", + "bbox": [ + 109, + 88, + 439, + 228 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b9c1cb17f860b7a908aa2483e308289a2c662546db07ecc2e0ded33483fce7f5.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with baselines with and without super resolution (SR) on the FFHQ dataset. $\\dagger$ as reported in the previous works. \\* indicates FID evaluated on $20\\mathrm{k}$ images." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓NFS↑
AFHQ-512w/o SRGRAM-HD [67]\\(7.67^{\\dagger}\\)-
GMPI [45]\\(7.79^{\\dagger}\\)-
Mimic3D [10]4.2912.67
Ours4.2321.89
SREG3D [7]2.7714.14
StyleSDF [45]7.9133.89
", + "bbox": [ + 117, + 295, + 431, + 421 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb6a921d1dcc2da9602d21223db789bab8f45bcce96eaae8cd447bb4e00ebb12.jpg", + "table_caption": [ + "Table 2. Quantitative results on AFHQv2 Cats. † as reported in the previous works." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓FID-N ↓NFS ↑
- Learned Sampler38.2993.8830.95
- Stratification5.6086.025.97
- Robust Sampling5.6760.7824.79
- Beta Regularization5.2764.2528.88
Ours4.9760.7629.35
", + "bbox": [ + 109, + 474, + 439, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study.", + "bbox": [ + 199, + 582, + 346, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(see Fig. 7) and worse geometry scores (see Tab. 3). Without our sampler or stratification during sampling, the model cannot learn meaningful 3D geometry with limited depth budgets, creating degenerated 3D geometry as can be seen in Fig. 8 and significantly worse FID-N. Without our robust sampling strategy, the sampling becomes more susceptible to slight errors in the sampler due to the high-frequency nature of the PDF (see Fig. 4), resulting in a noticeable drop in FID and occasionally creating floater geometry artifacts (see Fig. 8), while geometry scores remain similar.", + "bbox": [ + 75, + 625, + 468, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 76, + 784, + 191, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. While our method demonstrates significant improvements in 3D geometry generation, it may still exhibit artifacts such as dents in the presence of specularities, and cannot handle transparent objects such as lenses well. Future work may incorporate more advanced material formulations [5] and surface normal regularization [60]. While 3D GANs can learn 3D represent", + "bbox": [ + 75, + 809, + 468, + 912 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75f5c6d21c946211b2e92c0b42968429b09d0a848085b008aae4e5a0fdbfe8e2.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons for ablation study." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 893, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tations from single-view image collections such as FFHQ and AFHQ with casual camera labels [7], the frontal bias and inaccurate labels can result in geometry artifacts, especially on the side of the faces. Fruitful future directions may include training 3D GANs with large-scale Internet data as well as incorporating a more advanced form of regularization [47] and auto-camera calibration [3] to extend the generations to 360 degrees. Finally, our sampling-based acceleration method may be applied to other NeRFs.", + "bbox": [ + 496, + 299, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical considerations. While existing methods have demonstrated effective capabilities in detecting unseen GANs [12], our contribution may remove certain characteristics from generated images, potentially making the task of detection more challenging. Viable solutions include the authentication of synthetic media [1, 2, 48].", + "bbox": [ + 496, + 440, + 890, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We proposed a sampler-based method to accelerate 3D GANs to resolve 3D representations at the native resolution of 2D data, creating strictly multi-view-consistent images as well as highly detailed 3D geometry learned from a collection of in-the-wild 2D images. We believe our work opens up new possibilities for generating high-quality 3D models and synthetic data that capture in-the-wild variations and for enabling new applications such as conditional view synthesis.", + "bbox": [ + 496, + 537, + 890, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 683, + 668, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank David Luebke, Tero Karras, Michael Stengel, Amrita Mazumdar, Yash Belhe, and Nithin Raghavan for feedback on drafts and early discussions. Koki Nagano was partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. This work was funded in part by an NSF Graduate Fellowship, ONR Grant N00014-23-1-2526, and the Ronald L. Graham Chair. Manmohan Chandraker acknowledges support of of NSF IIS 2110409. Distribution Statement \"A\" (Approved for Public Release, Distribution Unlimited).", + "bbox": [ + 496, + 708, + 890, + 912 + ], + "page_idx": 7 + }, + { + "type": "footer", + "text": "22772", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Coalition for content provenance and authenticity. https://c2pa.org/.8", + "[2] Content authenticity initiative. https://contentauthenticity.org/.8", + "[3] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panopehead: Geometry-aware 3d full-head synthesis in 360deg. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8", + "[4] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5", + "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 8", + "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[8] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative novel view synthesis with 3D-aware diffusion models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 3", + "[10] Xingyu Chen, Yu Deng, and Baoyuan Wang. Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. arXiv preprint arXiv:2303.09036, 2023. 2, 3, 6, 7, 8", + "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6", + "[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023. 8", + "[13] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 7", + "[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image" + ], + "bbox": [ + 78, + 114, + 468, + 911 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 7", + "[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 5", + "[17] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 7", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 2, 6", + "[19] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2, 3", + "[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3", + "[21] Kunal Gupta, Miloš Hasan, Zexiang Xu, Fujun Luan, Kulyan Sunkavalli, Xin Sun, Manmohan Chandraker, and Sai Bi. Mcnerf: Monte carlo rendering and denoising for real-time nerfs. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7", + "[23] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5", + "[24] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4, 6", + "[26] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6", + "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + ], + "bbox": [ + 503, + 92, + 890, + 911 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22773", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6", + "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 3", + "[30] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. 2022. 3", + "[31] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966, 2023. 3", + "[32] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3", + "[34] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3", + "[35] David B Lindell, Julien NP Martel, and Gordon Wetzstein. AutoInt: Automatic integration for fast neural volume rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[36] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 4", + "[37] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 6", + "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4, 5", + "[39] Don P Mitchell. Consequences of stratified sampling in graphics. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 277-280, 1996. 4", + "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3", + "[41] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 910 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 3", + "[43] Michael Niemeyer and Andreas Geiger. GIRAFFE: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8", + "[46] M. Piala and R. Clark. Terminerf: Ray termination prediction for efficient neural rendering. In International Conference on 3D Vision (3DV), 2021. 3", + "[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2022. 3, 8", + "[48] Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo. Avatar fingerprinting for authorized use of synthetic talking-head videos. arXiv preprint arXiv:2305.03713, 2023. 8", + "[49] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[50] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6", + "[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 3", + "[53] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 6, 7", + "[54] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. arXiv preprint arXiv:2303.01416, 2023. 7", + "[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 3", + "[56] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In" + ], + "bbox": [ + 501, + 92, + 890, + 912 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22774", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[57] Towaki Takikawa, Thomas Müller, Merlin Nimier-David, Alex Evans, Sanja Fidler, Alec Jacobson, and Alexander Keller. Compact neural graphics primitives with learned hash probing. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3", + "[58] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3", + "[59] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2", + "[60] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8", + "[61] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3", + "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[64] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[65] Zian Wang, Tiancheng Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. In ACM Transactions on Graphics (SIGGRAPH ASIA). 3", + "[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3", + "[67] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3, 6, 8", + "[68] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in" + ], + "bbox": [ + 78, + 92, + 468, + 911 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 2", + "[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[70] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[71] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Ronen Basri, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 3", + "[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 5", + "[73] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[74] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 8", + "[75] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In European Conference on Computer Vision (ECCV), 2022. 3, 6", + "[76] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 580 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22775", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_model.json b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0902b3697dd20bb82781232225372d0e5693a226 --- /dev/null +++ b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.13, + 0.806, + 0.175 + ], + "angle": 0, + "content": "What You See is What You GAN: Rendering Every Pixel for High-Fidelity Geometry in 3D GANs" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.213, + 0.878, + 0.25 + ], + "angle": 0, + "content": "Alex Trevithick\\*, Matthew Chan, Towaki Takikawa, Umar Iqbal, Shalini De Mello, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.259, + 0.685, + 0.279 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)University of California, San Diego" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.302, + 0.472, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.302, + 0.895, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.565, + 0.894, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Left: Our results. The split view in the middle demonstrates the high degree of agreement between our 2D rendering and corresponding 3D geometry. Our method can learn fine-grained 3D details (e.g., eyeglass frame and cat's fur) that are geometrically well-aligned to 2D images without multiview or 3D scan data. Right: Comparison with EG3D [7]. Our tight SDF prior provides smooth and detailed surfaces on the face and hat while EG3D exhibits geometry artifacts and discrepancies between geometry and rendering. Please see Fig. 5 and the accompanying video for more examples, and Fig. 6 for comparison to other baselines." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.638, + 0.314, + 0.654 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.67, + 0.474, + 0.875 + ], + "angle": 0, + "content": "3D-aware Generative Adversarial Networks (GANs) have shown remarkable progress in learning to generate multi-view-consistent images and 3D geometries of scenes from collections of 2D images via neural volume rendering. Yet, the significant memory and computational costs of dense sampling in volume rendering have forced 3D GANs to adopt patch-based training or employ low-resolution rendering with post-processing 2D super resolution, which sacrifices multiview consistency and the quality of resolved geometry. Consequently, 3D GANs have not yet been able to fully resolve the rich 3D geometry present in 2D images. In this work, we propose techniques to scale neural volume rendering to the much higher resolution of native 2D images, thereby resolving fine-grained 3D geometry with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.887, + 0.47, + 0.913 + ], + "angle": 0, + "content": "*This project was initiated and substantially carried out during an internship at NVIDIA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.895, + 0.83 + ], + "angle": 0, + "content": "unprecedented detail. Our approach employs learning-based samplers for accelerating neural rendering for 3D GAN training using up to 5 times fewer depth samples. This enables us to explicitly \"render every pixel\" of the full-resolution image during training and inference without post-processing superresolution in 2D. Together with our strategy to learn high-quality surface geometry, our method synthesizes high-resolution 3D geometry and strictly view-consistent images while maintaining image quality on par with baselines relying on post-processing super resolution. We demonstrate state-of-the-art 3D gemetic quality on FFHQ and AFHQ, setting a new standard for unsupervised learning of 3D shapes in 3D GANs." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.843, + 0.633, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.868, + 0.894, + 0.913 + ], + "angle": 0, + "content": "Training 3D generative models from the abundance of 2D images allows the creation of 3D representations of real-world objects for content creation and novel view synthe" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.521, + 0.962 + ], + "angle": 0, + "content": "22765" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.353 + ], + "angle": 0, + "content": "sis [59]. Recently, 3D-aware generative adversarial networks (3D GANs) [6, 7, 19, 43, 45, 49, 53, 69, 70, 74, 76] have emerged as a powerful way to learn 3D representations from collections of 2D images in an unsupervised fashion. These methods employ differentiable rendering to compare rendered 3D scenes with 2D data using adversarial training [18]. Among the various 3D representations, Neural Radiance Fields (NeRF) [38] have become a popular choice among recent successful 3D GANs. However, the significant computational and memory cost of volume rendering has prevented 3D GANs from scaling to high-resolution output. For instance, generating a single \\(512 \\times 512\\) image via volume rendering requires evaluating as many as 25 million depth samples, if 96 depth samples are used per ray using importance sampling [7, 10, 38, 53]. Given that GAN training typically requires rendering tens of millions of images, the training process could require evaluating hundreds of trillions of depth samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.473, + 0.569 + ], + "angle": 0, + "content": "During training, all intermediate operations must be stored in GPU memory for every depth sample for the backward pass. Therefore, existing methods resort to working on patches [10, 49, 53] or adopting a low resolution neural rendering combined with post-processing 2D super resolution (SR) [7, 19, 43, 45, 70]. However, patch-based methods have limited receptive fields over scenes, leading to unsatisfactory results, and the hybrid low-resolution rendering and SR scheme inevitably sacrifices the multiview consistency and the accuracy of 3D geometry. While many techniques have been developed to improve the image quality of 3D GANs to match that of 2D GANs, the challenge of resolving the corresponding high-resolution 3D geometry remains unsolved (see Figs. 1 and 6 for our results and comparison to the current state-of-the-art)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.473, + 0.773 + ], + "angle": 0, + "content": "Scaling 3D GANs to operate natively at the 2D pixel resolution requires a novel approach for sampling. Fig. 2 compares the state-of-the-art 3D GAN, EG3D model, trained with and without SR. EG3D employs 96 dense depth samples in total using two-pass importance sampling [38] during training, which requires half a terabyte of GPU memory at \\(256 \\times 256\\) resolution, making scaling to higher resolutions infeasible. Furthermore, Fig. 2 demonstrates that using 96 dense samples still results in undersampling, as evidenced by the speckle noise patterns visible in the zoomed-in view, leading to considerably worse FID (inset in Fig. 2). 3D GANs relying on post-processing SR layers can repair these undersampling artifacts at the cost of a high-fidelity 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.473, + 0.86 + ], + "angle": 0, + "content": "In this work, we address the challenge of scaling neural volume rendering to high resolutions by explicitly rendering every pixel, ensuring that \"what you see in \\(2D\\), is what you get in \\(3D\\) — generating an unprecedented level of geometric details as well as strictly multiview-consistent images. Our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.878 + ], + "angle": 0, + "content": "- We introduce an SDF-based 3D GAN to represent high" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.47, + 0.913 + ], + "angle": 0, + "content": "Triplane resolution is doubled to compensate for the loss of capacity from removing the SR layers." + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.088, + 0.835, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.304, + 0.895, + 0.386 + ], + "angle": 0, + "content": "Figure 2. Samples from EG3D 256 model. Right: Volume rendering with 48 coarse samples and 48 fine samples per ray with two-pass importance sampling [38] results in undersampling, leading to noticeable noisy artifacts. Left: These artifacts are repaired by super resolution (SR). An unsharp mask has been applied to the zoomed views for presentation purposes." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.399, + 0.892, + 0.442 + ], + "angle": 0, + "content": "frequency geometry with spatially-varying surface tightness that increases throughout training (subsection 4.1 and 4.5), in turn facilitating low sample rendering." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.442, + 0.892, + 0.498 + ], + "angle": 0, + "content": "- We propose a generalizable learned sampler conditioned on cheap low-resolution information to enable full-resolution rendering during training for the first time (subsection 4.2 and 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.5, + 0.892, + 0.587 + ], + "angle": 0, + "content": "- We show a robust sampling strategy for the learned sampler (subsection 4.4) that produces stable neural rendering using significantly fewer depth samples (see Fig. 8). Our sampler can operate with just 20 samples per ray compared to existing 3D GANs which must use at least 96 samples per ray." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.587, + 0.892, + 0.646 + ], + "angle": 0, + "content": "- Together, our contributions result in the state-of-the-art geometry for 3D GANs while rendering with quality on par with SR baselines (see Fig. 1). For more results, see Fig. 5 and for comparison to other baselines, see Fig. 6." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.442, + 0.892, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.642, + 0.679 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.777 + ], + "angle": 0, + "content": "We begin by reviewing the prior-arts of 3D generative models and their current shortcomings. We then cover foundational techniques for 3D geometry representation and neural rendering from which we take inspiration. We then discuss existing methods for accelerating neural volume rendering, which usually operate per-scene." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.709, + 0.802 + ], + "angle": 0, + "content": "2.1. 3D Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.913 + ], + "angle": 0, + "content": "Just like 2D GANs, 3D-aware GANs train from a collection of 2D images, but employ a 3D representation and differentiable rendering to learn 3D scenes without requiring multiview images or ground truth 3D scans. Some of the most successful works use a neural field [68] in combination with a feature grid [7] as their 3D representation, and use neural volume rendering [38] as the differentiable renderer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.521, + 0.962 + ], + "angle": 0, + "content": "22766" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.194 + ], + "angle": 0, + "content": "However, due to the significant memory and computational cost of neural volume rendering, many previous works perform rendering at low-resolution and rely on a 2D post-processing CNN [7, 19, 43, 45, 70], which hallucinates the high-frequency details in a view-inconsistent manner while sacrificing 3D consistency and the quality of the resolved 3D geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.194, + 0.47, + 0.485 + ], + "angle": 0, + "content": "To ensure strict 3D consistency, other previous works seek to render at high-resolutions and propose techniques to address the prohibitive computational costs. One line of work leverages the sparse nature of 3D scenes to speed up rendering, in particular, structures such as 2D manifolds [14, 67], multiplane images [75] and sparse voxels [50]. Although they are more efficient, sparse representations provide only coarse [50] or category-specific [14, 67] acceleration structures, which poses constraints on the diversity and viewing angles of the generated scenes. Our sampling-based method, on the other hand, generalizes to every new scene and adaptively accelerates rendering on a per ray basis. Another line of work enables high-resolution rendering with patch-based training [10, 53]. In particular, Mimic3D [10] achieves significantly improved 2D image quality, but the patch-based training limits the receptive fields, and the generated geometry does not faithfully represent the 2D data due to the patch-wise perceptual loss. Our method renders the entire image at once and the resulting geometry is aligned with the rendering (see Figs. 1 and 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Recently, a new family of generative approaches using diffusion models has been proposed to tackle conditional tasks including novel view synthesis [8, 58] and text-based 3D generation[63]. Most of these 3D-aware diffusion models combine a 3D inductive bias modeled via neural field representations and a 2D image denoising objective to learn 3D scene generation. While these models enable unconditional 3D generation, they require multiview images [24, 55, 63] or 3D data, such as a point cloud [42]. Score Distillation Sampling [47] may be used for distillation from a pre-trained 2D diffusion model when only monocular 2D data is available, but diffusion models incur significant computational costs due to their iterative nature and most of the existing methods require optimization per scene [9, 20, 33, 61, 66]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.373, + 0.73 + ], + "angle": 0, + "content": "2.2. Learning High-Fidelity Geometry" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.468, + 0.868 + ], + "angle": 0, + "content": "Prior works on 3D GANs have typically represented the geometry as a radiance field [38], which lacks a concrete definition for where the surface geometry resides in the field, resulting in bumpy surfaces. A number of works [44, 62, 71] have proposed alternate representations based on implicit surfaces (such as signed distance functions, or SDFs) that can be used with neural volume rendering. In these works, the implicit surface is typically softened by a parameter for volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.868, + 0.468, + 0.913 + ], + "angle": 0, + "content": "Other works [32, 64] improve on these implicit surface representations by leveraging feature grids for higher computational efficiency and resolution. Adaptive Shells [65]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.266 + ], + "angle": 0, + "content": "further improve on quality by making the softness parameter spatially-varying, as many objects have hard boundaries only in certain parts. We use an implicit surface representation based on VolSDF [72], and leverage a spatially-varying parameter similar to Adaptive Shells [65] to control the softness of the surface, as humans and animals benefit from both hard surfaces (e.g. skin, eyes) and soft, volumetric representations (e.g. hair). Although other works such as StyleSDF [45] have similarly leveraged implicit surfaces in a 3D GAN framework, the lack of spatial-variance and high-resolution rendering led to over-smoothed geometry not faithful to the rendered images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.277, + 0.838, + 0.293 + ], + "angle": 0, + "content": "2.3. Accelerating Neural Volume Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.547 + ], + "angle": 0, + "content": "As mentioned in Section 2.1, accelerating 3D GANs typically relies on acceleration structures such as octrees [31, 56, 73], which in generative settings [50] are limited due to computational requirements. In another direction, GRAM-based methods [14, 67] are limited due to sharing one structure across all samples. Instead, we look to a class of methods that do not rely on an acceleration structure. Some of these works learn a per-scene sampling prior on a per-ray basis using a binary classifier [41], a density estimator [30, 46], a coarse proxy from a 3D cost volume [34], or an interval length estimator [35] on discrete depth regions along the ray. Other works use importance sampling [21, 38] to sample additional points. We take inspiration from works on density estimators [30] and propose to learn a scene-conditional proposal network that generalizes across scenes instead of being category-specific or optimized per-scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.664 + ], + "angle": 0, + "content": "There are also other methods to accelerate rendering by utilizing more efficient representations, such as gaussian splats [29] and light field networks [52]. More efficient feature grids [40, 57] based on hashing can also be used to accelerate rendering. However, mapping to these representations in a GAN framework is not straightforward. In contrast, our sampling strategy can be used for any NeRF representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.628, + 0.685 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.913 + ], + "angle": 0, + "content": "We begin with background on the methodology of the state-of-the-art 3D-aware GANs as our method relies on a similar backbone for mapping to 3D representations. 3D GANs typically utilize a StyleGAN-like [27] architecture to map from a simple Gaussian prior to the conditioning of a NeRF, whether that be an MLP [19, 45], MPI [75], 3D feature grid [50], manifolds [14, 67] or triplane [7, 10, 53]. We inherit the latter triplane conditioning for its high expressivity and efficiency, in which three axis-aligned 2D feature grids \\((f_{xy}, f_{xz}, f_{yz})\\), provide NeRF conditioning by orthogonal projection and interpolation. As in the previous methods, the mapping and synthesis networks from StyleGAN2 can easily be adapted to create the 2D triplane representation from noise \\(z \\in \\mathbb{R}^{512}\\). Specifically, \\(\\mathbf{w} = \\mathrm{Mapping}(z)\\) conditions the creation of the triplane \\(T(\\mathbf{w}) \\in \\mathbb{R}^{3 \\times 32 \\times 512 \\times 512}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "22767" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.891, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.894, + 0.282 + ], + "angle": 0, + "content": "Figure 3. Here we show our proposed pipeline and its intermediate outputs. Beginning from the triplane \\( T \\), we trace uniform samples to probe the scene, yielding low-resolution \\( I_{128} \\) and weights \\( P_{128} \\). These are fed to a CNN which produces high-resolution proposal weights \\( \\hat{P}_{512} \\) (weights are visualized as uniform level sets). We perform robust sampling and volume render to get the final image \\( I_{512} \\) and the surface variance \\( B \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.337 + ], + "angle": 0, + "content": "from a Synthesis network, corresponding to three axis-aligned 2D feature grids of spatial resolution \\(512 \\times 512\\) and feature size 32." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.398 + ], + "angle": 0, + "content": "To create high-fidelity geometry, our method builds upon VolSDF [72]: Instead of directly outputting the opacity \\(\\sigma\\) of a point \\(x\\in \\mathbb{R}^3\\), an SDF value \\(s\\) is output and transformed to \\(\\sigma\\) using a Laplacian CDF:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.406, + 0.47, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\frac {1}{\\beta} \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\exp \\left(\\frac {s}{\\beta}\\right) & \\text {i f} s \\leq 0 \\\\ 1 - \\frac {1}{2} \\exp \\left(- \\frac {s}{\\beta}\\right) & \\text {i f} s > 0 \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.546 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the variance of the Laplacian distribution governing the \"tightness\" of the representation. One distinct benefit of an SDF-based representation is the ease of extracting the surface. StyleSDF [45] also utilizes this intermediate geometry representation without a triplane, enforcing the usual Eikonal constraint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.635 + ], + "angle": 0, + "content": "Using the triplane conditioning and Eq. 1, we can assign each point in the volume with its opacity \\(\\sigma\\) and radiance \\(c\\) using a lightweight MLP. For a given camera ray \\(r(t) = o + td\\), we approximate the volumetric rendering integral \\(C(r)\\) [36] by sampling ray distances \\(t_i\\) with their corresponding \\(\\sigma_i\\) and \\(c_i\\) before computing" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.644, + 0.47, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (r) = \\sum_ {i = 1} ^ {N} w _ {i} c _ {i} \\text {w h e r e} w _ {i} = T _ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.687, + 0.468, + 0.737 + ], + "angle": 0, + "content": "\\[\nT _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}\\right) \\text {a n d} \\delta_ {i} = t _ {i + 1} - t _ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.774 + ], + "angle": 0, + "content": "Here, \\(T_{i}\\) denotes the accumulated transmittance and \\(\\delta_{i}\\) is the distance between adjacent samples along the ray." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.471, + 0.913 + ], + "angle": 0, + "content": "It is possible to develop a more efficient estimator for this sum with fewer samples by using importance sampling techniques in computer graphics. Typically, one computes a piecewise constant probability distribution \\( p_j = \\frac{\\hat{w}_j}{\\sum_j \\hat{w}_j} \\), where \\( j \\) refers to the \\( j^{\\text{th}} \\) bin or region, and \\( \\hat{w}_j \\) is an estimate of \\( w_j \\) for that region, for example obtained by explicitly tracing coarse samples. For a given \\( t \\), we first find the region \\( j(t) \\) and then set \\( p(t) = p_j \\). From this, one can compute a (piecewise linear) cumulative distribution function or CDF" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.325 + ], + "angle": 0, + "content": "\\(\\Phi (t)\\) which has a range from 0 to 1. We can then perform inverse CDF sampling to define the sample points," + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.329, + 0.892, + 0.348 + ], + "angle": 0, + "content": "\\[\nt _ {i} = \\Phi^ {- 1} (u _ {i}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.382 + ], + "angle": 0, + "content": "where \\( u_{i} \\) is a random number from 0 to 1 (sorted to be an increasing sequence)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Discussion We improve on previous works such as NeRF [38] and EG3D [7] by stratifying2 the random numbers \\( u_{i} \\) during training; this leads to significantly lower rendering variance, especially at low sample counts \\( N \\) [39]. We also develop a neural method to predict a good distribution \\( p_{j} \\) (and hence \\( \\Phi \\)) for importance sampling at high spatial resolution, without needing to exhaustively step through the ray." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.591, + 0.525 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.667 + ], + "angle": 0, + "content": "In this section, we describe our method beginning with our SDF-based NeRF parametrization (subsection 4.1). We then overview how we render at high-resolution in three stages: first, a low-resolution probe into the 3D scene (subsection 4.2); second a high-resolution CNN proposal network (subsection 4.3); and third a robust sampling method for the resultant proposals (subsection 4.4). Next we describe regularizations (subsection 4.5) for stable training, and finally our entire training pipeline (subsection 4.6)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.672, + 0.787, + 0.688 + ], + "angle": 0, + "content": "4.1. Mapping to a 3D Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Beginning from a noise vector \\(z\\), we synthesize the initial triplane \\(T'\\) with StyleGAN [25] layers as detailed in Sec. 3. In contrast to previous methods, we then generate more expressive triplane features \\(T\\) with an extra synthesis block for each orthogonal plane: \\(f_{ij} = \\mathrm{SynthesisBlock}(f_{ij}')\\) where \\(ij \\in \\{xy, xz, yz\\}\\) and \\(f'\\) are the features of \\(T'\\). This design choice allows disentanglement between the intermediate triplane features as plane-specific kernels can attend to the features in each separate plane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Given the triplane \\(T\\) (the left side of Fig. 3) and a point \\(x \\in \\mathbb{R}^3\\), we utilize an MLP to map to the SDF value \\(s\\), variance \\(\\beta\\) and geometry features \\(f_{geo}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.875, + 0.892, + 0.893 + ], + "angle": 0, + "content": "\\[\n(s, \\beta , f _ {g e o}) = \\operatorname {M L P} _ {\\mathrm {S D F}} (\\operatorname {P o s E n c} (x), T _ {x}) \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.899, + 0.892, + 0.912 + ], + "angle": 0, + "content": "2 dividing the unit interval into bins and taking a sample from each bin." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "22768" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.468, + 0.208 + ], + "angle": 0, + "content": "where PosEnc is the positional encoding from NeRF [38] and \\( T_{x} \\) are the features corresponding to \\( x \\) gathered from \\( T \\) by projecting \\( x \\) to each of the axis-aligned planes and taking the Hadamard product of the three resultant vectors [16]. We initialize the ReLU \\( \\mathrm{MLP}_{\\mathrm{SDF}} \\) as in SAL [4] to ensure an approximately spherical SDF in the early part of training. Additionally note that \\( \\beta \\) varies spatially in the volume unlike [45, 72], allowing us to regularize its values later on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.208, + 0.468, + 0.253 + ], + "angle": 0, + "content": "Using Eq. 1, we transform \\(s\\) and \\(\\beta\\) into opacity \\(\\sigma\\). We can now predict the radiance with a separate \\(\\mathrm{MLP_c}\\) conditioned on the geometry features and viewing direction \\(v\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.268, + 0.468, + 0.284 + ], + "angle": 0, + "content": "\\[\nc = \\mathrm {M L P} _ {\\mathrm {c}} (\\operatorname {P o s E n c} (v), f _ {g e o}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.468, + 0.366 + ], + "angle": 0, + "content": "Note that unlike EG3D [7], we condition radiance on the viewing direction, allowing a more expressive generator. Thus, given a triplane, we can render any pixel by computing \\(\\sigma\\) and \\(c\\) for points along the ray to approximate the volumetric rendering integral as described in Sec. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.383, + 0.392 + ], + "angle": 0, + "content": "4.2. High-Resolution Proposal Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.468, + 0.618 + ], + "angle": 0, + "content": "We now have our mapping from a latent code \\( z \\) to a 3D NeRF representation. However, volumetric rendering of NeRFs at higher resolutions requires extremely large numbers of samples, and thus both memory and time. Instead of naive dense sampling at a high resolution, we propose to leverage low-resolution renderings to cheaply probe the 3D representation (visualized on the left of Fig. 3) for the creation of proposal distributions at high-resolution. Given a target camera and triplane \\( T \\), we first trace 192 coarse samples at low-resolution \\( (128 \\times 128) \\) to compute a low-resolution RGB image \\( I_{128} \\in \\mathbb{R}^{3 \\times 128 \\times 128} \\) and a tensor of weights \\( P_{128} \\in \\mathbb{R}^{192 \\times 128 \\times 128} \\) (visualized after low-resolution rendering in Fig. 3). Each 192-dimensional vector corresponds to a piecewise constant PDF with CDF \\( \\Phi \\) as seen in Eq. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.468, + 0.661 + ], + "angle": 0, + "content": "Conditioned on the low-resolution probe, we predict a tensor of proposal volume rendering weights at the high-resolution \\((512 \\times 512)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.673, + 0.468, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {5 1 2} = \\operatorname {S o f t m a x} \\left(\\operatorname {C N N} \\left(P _ {1 2 8}, I _ {1 2 8}\\right)\\right) \\in \\mathbb {R} ^ {1 9 2 \\times 5 1 2 \\times 5 1 2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.468, + 0.804 + ], + "angle": 0, + "content": "where CNN is a lightweight network that up-samples the low-resolution weights, Softmax produces discrete distributions along each ray, and the \\(\\hat{}\\) denotes that this is an estimated quantity. This corresponds to the Proposal in Fig. 3 and the yellow distribution in Fig. 4. Note that allocating 192 samples at \\(128 \\times 128\\) is equivalent to allocating just 12 at \\(512 \\times 512\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.376, + 0.832 + ], + "angle": 0, + "content": "4.3. Supervising the Proposal Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.468, + 0.913 + ], + "angle": 0, + "content": "Having described the input and output of our high-resolution proposal network, we now show its supervision. From the target camera, we can also trace 192 coarse samples at high resolution for a small \\(64 \\times 64\\) patch, giving us a ground truth tensor of volume rendering weights" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.89, + 0.122 + ], + "angle": 0, + "content": "\\(P_{\\mathrm{patch}} \\in \\mathbb{R}^{192 \\times 64 \\times 64}\\). We then prepare this tensor for supervision by computing:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.131, + 0.89, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\bar {P} _ {\\text {p a t c h}} = \\text {N o r m a l i z e} (\\text {S u p p r e s s} (\\text {B l u r} (P _ {\\text {p a t c h}}))) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.158, + 0.892, + 0.274 + ], + "angle": 0, + "content": "where Blur applies a 1D Gaussian kernel to the input distributions, \\(\\operatorname{Suppress}(x) = x\\) if \\(x \\geq 5e - 3\\) and 0 otherwise, and Normalize is L1 normalization to create a valid distribution. This corresponds to the patch loss in Fig. 3 and the purple distribution in Fig. 4. These operations create less noisy distributions to facilitate accurate learning of the high-frequency integrand which may be undersampled in the coarse pass." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.304 + ], + "angle": 0, + "content": "We can then compare the predicted and cleaned ground truth distributions with a cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.315, + 0.89, + 0.332 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s a m p l e r}} = \\text {C r o s s E n t r o p y} (\\bar {P} _ {\\text {p a t c h}}, \\hat {P} _ {\\text {p a t c h}}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.432 + ], + "angle": 0, + "content": "where \\(\\hat{P}_{\\mathrm{patch}}\\) is the corresponding patch of weights in \\(\\hat{P}_{512}\\) and CrossEntropy denotes the average cross-entropy between all pairs of pixelwise discrete distributions; for each pair \\((\\bar{p},\\hat{p})\\), we compute \\(\\sum -\\bar{p}_j\\log \\hat{p}_j\\). Since we only need to compute this supervision for a small patch, the overhead of sampler training is not significant." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.457 + ], + "angle": 0, + "content": "4.4. Robustly Sampling from the Proposal Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.535 + ], + "angle": 0, + "content": "Having shown how to train and predict high-resolution distributions, we now overview how to sample the resultant proposals. As seen in Fig. 4, the proposals are often slightly off; this is due to the high frequency nature of the underlying integrand in blue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.623 + ], + "angle": 0, + "content": "In order to utilize the information from the sampler, we propose to filter predicted PDFs for better estimation. Specifically, for each discrete predicted PDF \\(\\hat{p}\\), we compute the smallest set of bins whose probability exceeds a threshold \\(\\tau = 0.98\\): We find the smallest subset \\(I \\subseteq \\{1,2,\\dots,192\\}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.634, + 0.89, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i \\in I} \\hat {p} _ {i} \\geq \\tau . \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.89, + 0.72 + ], + "angle": 0, + "content": "This operation resembles nucleus sampling in NLP [23]. We define our sampling PDF \\(q\\) with probability \\(q_{i} = \\frac{1}{|I|}\\) if \\(i\\in I\\) and 0 otherwise (the green distribution in Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.892, + 0.851 + ], + "angle": 0, + "content": "For each PDF \\( q \\), we compute its CDF \\( \\Phi \\) and perform stratified inverse transform sampling to create the samples (illustrated as adaptive sampling near the surface in Fig. 3). We refer to these samples as robust stratified samples. In practice, on top of the 12 samples from the coarse probe (for the high-resolution image), we take an additional 18 samples per pixel adaptively based on the variance of the predicted distributions. The details are given in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.86, + 0.878, + 0.877 + ], + "angle": 0, + "content": "4.5. Regularization for High-Resolution Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.883, + 0.892, + 0.913 + ], + "angle": 0, + "content": "In order to render accurately under low sample budget per ray, we desire the surface to be tight, i.e., the set of points" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "22769" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.473, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.416 + ], + "angle": 0, + "content": "Figure 4. We visualize the volume rendering PDFs for the green pixel in the images on the right along with sampling methods. The ground truth distribution in blue is bimodal due to the discontinuous depth. Without stratification, the samples from the predicted yellow PDF completely miss the second mode. Stratification reduces the variance, yet also misses the second mode. Our robust stratified samples hit both modes despite the inaccurate predictions. The supervision PDF is visualized in purple as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.473, + 0.534 + ], + "angle": 0, + "content": "along the ray, which contribute to the accumulated radiance to be small. To accomplish this, we introduce a regularization for the spatially-varying \\(\\beta\\) values. Replacing the \\(c_{i}\\) with the intermediate \\(\\beta_{i}\\) in Eq. 2, we can volume render an image of scalars, \\(B\\in \\mathbb{R}^{512\\times 512}\\) (seen on the right of Fig. 3). During training, we regularize \\(B\\) to be small so that the surface tightens:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.545, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s u r f a c e}} = \\sum_ {h w} \\left(B _ {h w} - B _ {\\text {t a r g e t}}\\right) ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.633 + ], + "angle": 0, + "content": "where \\( B_{\\mathrm{target}} \\) is a scalar quantity annealed towards \\( \\epsilon > 0 \\) during optimization. Note that this results in significantly more pronounced and smooth geometry (see Fig. 7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.471, + 0.793 + ], + "angle": 0, + "content": "During training with the proposal network, the rendering networks only receive gradients for points very near the surface (which have high density). We find that this can lead to undesirable density growing from the background into the foreground. In order to combat this, we leverage the low-resolution information from \\( I_{128} \\), which is computed for uniform samples along the ray, and thus not concentrated at the surface. Considering the SDF values intermediately computed for rendering, \\( S = S_{128} \\in \\mathbb{R}^{192 \\times 128 \\times 128} \\), we enforce the SDF decision boundary by minimizing the SDF likelihood under a Laplacian distribution similar to [45, 51]:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.804, + 0.47, + 0.837 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {d e c}} = \\sum_ {z h w} \\exp (- 2 | S _ {z h w} |). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.283, + 0.862 + ], + "angle": 0, + "content": "4.6. The Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.868, + 0.471, + 0.913 + ], + "angle": 0, + "content": "In order to begin making use of the learned sampler for high-resolution rendering, we need a good NeRF representation from which to train it. Concurrently, we also need a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "good sampler in order to allow NeRF to render at \\(512 \\times 512\\), the input resolution to the discriminator \\(D\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.894, + 0.266 + ], + "angle": 0, + "content": "To solve this issue, in the early stages of training, we first learn a low-resolution (e.g., \\(64 \\times 64\\)) 3D GAN through the standard NeRF sampling techniques. We bilinearly up-sample our low-resolution renderings to \\(512 \\times 512\\) and blur the real images to the same level. After converging at the lower resolution, we introduce the sampler training (subsection 4.3). Concretely, we not only render low-resolution images with standard sampling, but also render sampler inputs \\(P_{128}\\) and supervision patches \\(P_{\\mathrm{patch}}\\). This results in a good initialization for the sampler." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.894, + 0.37 + ], + "angle": 0, + "content": "Having learned an initial low-resolution 3D GAN and high-resolution sampler, we transition to rendering with the sampler predictions. The high-resolution proposals \\(\\hat{P}_{512}\\) are downsampled to the current rendering resolution, which is progressively increased to the full \\(512 \\times 512\\) resolution during training. After introducing all losses, we optimize the parameters of the generator \\(G\\) to minimize the following:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.379, + 0.892, + 0.396 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {a d v}} + \\lambda_ {\\text {s a m p l e r}} L _ {\\text {s a m p l e r}} + \\lambda_ {\\text {s u r f a c e}} L _ {\\text {s u r f a c e}} + \\lambda_ {\\text {d e c}} L _ {\\text {d e c}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.522 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{adv}} \\) is the standard GAN loss [18] Softplus \\( (-D(G(z)) \\). Note that we do not enforce the Eikonal constraint as we did not see a benefit. The discriminator \\( D \\) is trained with R1 gradient regularization [37] whose weight is adapted as the resolution changes. Generator and discriminator pose conditioning follow EG3D [7]. The details of all hyperparameters and schedules are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.587, + 0.548 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.661 + ], + "angle": 0, + "content": "Datasets. We benchmark on two standard datasets for 3D GANs: FFHQ [25] and AFHQv2 Cats [11, 28] both at resolution \\(512 \\times 512\\). We use the camera parameters extracted by EG3D [7] for conditioning and rendering. Our AFHQ model is finetuned from our FFHQ model using adaptive data augmentation [26]. For more results, please see the accompanying video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.64, + 0.684 + ], + "angle": 0, + "content": "5.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Baselines. We compare our methods against state-of-the-art 3D GAN methods including those that use low-resolution neural rendering and 2D post-processing CNN super resolution: EG3D [7], MVCGAN [74], and StyleSDF [45]; and methods that operate entirely based on neural rendering: Mimic3D [10], Epigraph [53], GMPI [75], and GramHD [67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.913 + ], + "angle": 0, + "content": "Qualitative results Fig. 5 shows the curated samples generated by our method with FFHQ and AFHQ, demonstrating photorealistic rendering as well as high-resolution details that align with the 2D images. Fig 6 provides qualitative comparisons to baselines. EG3D shows significant artifacts and cannot resolve high-resolution geometry since it performs neural rendering at only \\(128 \\times 128\\). Mimic3D" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.95, + 0.521, + 0.962 + ], + "angle": 0, + "content": "22770" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.487, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.089, + 0.892, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.893, + 0.26 + ], + "angle": 0, + "content": "Figure 5. Curated samples on FFHQ and AFHQ. Our method can resolve high-fidelity geometry (e.g., eyeglasses) and fine-grained details (e.g., cat's fur) as seen in the geometry and normal map." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.262, + 0.484, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.262, + 0.88, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.893, + 0.551 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons on FFHQ with EG3D [7], Mimic3D [10] and Epigraph [53]. EG3D performs neural rendering at resolution \\(128 \\times 128\\) and relies on \\(4 \\times\\) super resolution to generate images. On the right, Mimic3D and Epigraph directly generate the image via neural rendering. While all other baselines use up to 192 dense depth samples per ray, our method can operate at 30 samples per ray." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.564, + 0.47, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.676, + 0.452, + 0.691 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the effect of beta regularization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.79 + ], + "angle": 0, + "content": "and Epigraph render all pixels with neural rendering, but the patch-based nature of these methods harm the overall 3D geometry (e.g., distorted face and missing ear). Our method provides both high-fidelity 3D shape (e.g., well-defined ears and isolated clothing collar) and high-resolution details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.913 + ], + "angle": 0, + "content": "Quantitative results. Tabs. 1 and 2 provide quantitative comparisons against baselines. We measure the image quality with Fréchet Inception Distance (FID) [22]. We assess the quality of the learned geometry with a face-specific Normal FID (FID-N) [15]. We render 10.79k normal maps from the NPHM [17] dataset by solving for the approximate alignment between the average FFHQ 2D landmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.71 + ], + "angle": 0, + "content": "and the provided 3D landmarks. Examples are given in the supplement. For each mesh, we render two views with approximately 20 degrees of pitch variation relative to the front of the face. These views are processed to remove the background with facer [13]. For each baseline, we render 10k normal maps and remove their background using the predicted mask from facer on the rendered image. We compute the FID between the two sets of images. Finally, we also evaluate the flatness of the learned representations with non-flatness score (NFS) [54]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.893, + 0.844 + ], + "angle": 0, + "content": "Our results show the state-of-the-art image quality among the methods that operate only with neural rendering, while achieving FID comparable to the state-of-the-art SR-based method, EG3D. Our geometry quality outperforms all existing methods as indicated by our state-of-the-art FID-N. Additionally, our high NFS scores show the 3D aspect of our geometry. However, since NFS simply measures the variations of depth as a measure of non-flatness, it does not quantify the quality of geometry above a certain threshold." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.858, + 0.655, + 0.874 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.883, + 0.892, + 0.913 + ], + "angle": 0, + "content": "Without our surface tightness regularization (Eq. 10), the SDF surface may get fuzzy, resulting in a less clean surface" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "22771" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.44, + 0.229 + ], + "angle": 0, + "content": "
MethodFID ↓FID-N↓NFS↑
FFHQ-512Epigraph [7]9.92†67.3333.95
Mimic3D [10]5.3764.9716.76
GRAM-HD [67]12.2†*--
Ours4.9760.7629.35
w SREG3D [7]4.7063.0217.54
StyleSDF [45]11.19†87.4222.75
MVCGAN [74]13.4†*--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.24, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with baselines with and without super resolution (SR) on the FFHQ dataset. \\(\\dagger\\) as reported in the previous works. \\* indicates FID evaluated on \\(20\\mathrm{k}\\) images." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.296, + 0.432, + 0.422 + ], + "angle": 0, + "content": "
MethodFID ↓NFS↑
AFHQ-512w/o SRGRAM-HD [67]\\(7.67^{\\dagger}\\)-
GMPI [45]\\(7.79^{\\dagger}\\)-
Mimic3D [10]4.2912.67
Ours4.2321.89
SREG3D [7]2.7714.14
StyleSDF [45]7.9133.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.432, + 0.47, + 0.46 + ], + "angle": 0, + "content": "Table 2. Quantitative results on AFHQv2 Cats. † as reported in the previous works." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.476, + 0.44, + 0.574 + ], + "angle": 0, + "content": "
MethodFID ↓FID-N ↓NFS ↑
- Learned Sampler38.2993.8830.95
- Stratification5.6086.025.97
- Robust Sampling5.6760.7824.79
- Beta Regularization5.2764.2528.88
Ours4.9760.7629.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.583, + 0.347, + 0.597 + ], + "angle": 0, + "content": "Table 3. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.773 + ], + "angle": 0, + "content": "(see Fig. 7) and worse geometry scores (see Tab. 3). Without our sampler or stratification during sampling, the model cannot learn meaningful 3D geometry with limited depth budgets, creating degenerated 3D geometry as can be seen in Fig. 8 and significantly worse FID-N. Without our robust sampling strategy, the sampling becomes more susceptible to slight errors in the sampler due to the high-frequency nature of the PDF (see Fig. 4), resulting in a noticeable drop in FID and occasionally creating floater geometry artifacts (see Fig. 8), while geometry scores remain similar." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.192, + 0.8 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.913 + ], + "angle": 0, + "content": "Limitations and future work. While our method demonstrates significant improvements in 3D geometry generation, it may still exhibit artifacts such as dents in the presence of specularities, and cannot handle transparent objects such as lenses well. Future work may incorporate more advanced material formulations [5] and surface normal regularization [60]. While 3D GANs can learn 3D represent" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.895, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.258, + 0.855, + 0.273 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons for ablation study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.432 + ], + "angle": 0, + "content": "tations from single-view image collections such as FFHQ and AFHQ with casual camera labels [7], the frontal bias and inaccurate labels can result in geometry artifacts, especially on the side of the faces. Fruitful future directions may include training 3D GANs with large-scale Internet data as well as incorporating a more advanced form of regularization [47] and auto-camera calibration [3] to extend the generations to 360 degrees. Finally, our sampling-based acceleration method may be applied to other NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Ethical considerations. While existing methods have demonstrated effective capabilities in detecting unseen GANs [12], our contribution may remove certain characteristics from generated images, potentially making the task of detection more challenging. Viable solutions include the authentication of synthetic media [1, 2, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Conclusion. We proposed a sampler-based method to accelerate 3D GANs to resolve 3D representations at the native resolution of 2D data, creating strictly multi-view-consistent images as well as highly detailed 3D geometry learned from a collection of in-the-wild 2D images. We believe our work opens up new possibilities for generating high-quality 3D models and synthetic data that capture in-the-wild variations and for enabling new applications such as conditional view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.684, + 0.669, + 0.701 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.913 + ], + "angle": 0, + "content": "We thank David Luebke, Tero Karras, Michael Stengel, Amrita Mazumdar, Yash Belhe, and Nithin Raghavan for feedback on drafts and early discussions. Koki Nagano was partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. This work was funded in part by an NSF Graduate Fellowship, ONR Grant N00014-23-1-2526, and the Ronald L. Graham Chair. Manmohan Chandraker acknowledges support of of NSF IIS 2110409. Distribution Statement \"A\" (Approved for Public Release, Distribution Unlimited)." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.95, + 0.521, + 0.962 + ], + "angle": 0, + "content": "22772" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.466, + 0.141 + ], + "angle": 0, + "content": "[1] Coalition for content provenance and authenticity. https://c2pa.org/.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.143, + 0.468, + 0.17 + ], + "angle": 0, + "content": "[2] Content authenticity initiative. https://contentauthenticity.org/.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.469, + 0.224 + ], + "angle": 0, + "content": "[3] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panopehead: Geometry-aware 3d full-head synthesis in 360deg. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.226, + 0.469, + 0.28 + ], + "angle": 0, + "content": "[4] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.282, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.35, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.514, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[8] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative novel view synthesis with 3D-aware diffusion models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.596, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[10] Xingyu Chen, Yu Deng, and Baoyuan Wang. Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. arXiv preprint arXiv:2303.09036, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.882 + ], + "angle": 0, + "content": "[13] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.885, + 0.469, + 0.912 + ], + "angle": 0, + "content": "[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.173 + ], + "angle": 0, + "content": "[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.176, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.244, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[17] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.312, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.38, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[19] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[21] Kunal Gupta, Miloš Hasan, Zexiang Xu, Fujun Luan, Kulyan Sunkavalli, Xin Sun, Manmohan Chandraker, and Sai Bi. Mcnerf: Monte carlo rendering and denoising for real-time nerfs. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.892, + 0.651 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.653, + 0.892, + 0.693 + ], + "angle": 0, + "content": "[23] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.695, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[26] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.912 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "22773" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[28] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.47, + 0.201 + ], + "angle": 0, + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.203, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[30] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.257, + 0.469, + 0.298 + ], + "angle": 0, + "content": "[31] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.299, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[32] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[34] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[35] David B Lindell, Julien NP Martel, and Gordon Wetzstein. AutoInt: Automatic integration for fast neural volume rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[36] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.469, + 0.652 + ], + "angle": 0, + "content": "[37] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.655, + 0.469, + 0.72 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.722, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[39] Don P Mitchell. Consequences of stratified sampling in graphics. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 277-280, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.911 + ], + "angle": 0, + "content": "[41] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.911 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[42] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[43] Michael Niemeyer and Andreas Geiger. GIRAFFE: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.243, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[46] M. Piala and R. Clark. Terminerf: Ray termination prediction for efficient neural rendering. In International Conference on 3D Vision (3DV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.353, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[48] Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo. Avatar fingerprinting for authorized use of synthetic talking-head videos. arXiv preprint arXiv:2305.03713, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[49] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[50] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.892, + 0.638 + ], + "angle": 0, + "content": "[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.639, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.708, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[53] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[54] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. arXiv preprint arXiv:2303.01416, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.913 + ], + "angle": 0, + "content": "[56] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.52, + 0.962 + ], + "angle": 0, + "content": "22774" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[57] Towaki Takikawa, Thomas Müller, Merlin Nimier-David, Alex Evans, Sanja Fidler, Alec Jacobson, and Alexander Keller. Compact neural graphics primitives with learned hash probing. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.189, + 0.468, + 0.269 + ], + "angle": 0, + "content": "[58] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.271, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[59] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.468, + 0.406 + ], + "angle": 0, + "content": "[60] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.408, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[61] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.468, + 0.624 + ], + "angle": 0, + "content": "[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.468, + 0.692 + ], + "angle": 0, + "content": "[64] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.694, + 0.468, + 0.761 + ], + "angle": 0, + "content": "[65] Zian Wang, Tiancheng Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. In ACM Transactions on Graphics (SIGGRAPH ASIA). 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[67] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.468, + 0.912 + ], + "angle": 0, + "content": "[68] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.176, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[70] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.23, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[71] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Ronen Basri, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.352, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[73] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[74] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.527 + ], + "angle": 0, + "content": "[75] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In European Conference on Computer Vision (ECCV), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.528, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[76] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "22775" + } + ] +] \ No newline at end of file diff --git a/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_origin.pdf b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84a76819e48ec280b62deda9c7b8908268992b75 --- /dev/null +++ b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/2c134100-ed06-4b65-8341-cd48f2b74244_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64041581da219f7346abbd5a3486fad55b62b9a8e33f5f9be8189e30a3fd70bb +size 8098858 diff --git a/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/full.md b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d77af4dc81555237bd1cc636c1278cae940b95ca --- /dev/null +++ b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/full.md @@ -0,0 +1,368 @@ +# What You See is What You GAN: Rendering Every Pixel for High-Fidelity Geometry in 3D GANs + +Alex Trevithick\*, Matthew Chan, Towaki Takikawa, Umar Iqbal, Shalini De Mello, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano + +$^{1}$ NVIDIA $^{2}$ University of California, San Diego + +![](images/2988b5876bd21cea5c88f1ef75f51ab3a2b45ef962f3edcd7ccd6791a48f670f.jpg) +Figure 1. Left: Our results. The split view in the middle demonstrates the high degree of agreement between our 2D rendering and corresponding 3D geometry. Our method can learn fine-grained 3D details (e.g., eyeglass frame and cat's fur) that are geometrically well-aligned to 2D images without multiview or 3D scan data. Right: Comparison with EG3D [7]. Our tight SDF prior provides smooth and detailed surfaces on the face and hat while EG3D exhibits geometry artifacts and discrepancies between geometry and rendering. Please see Fig. 5 and the accompanying video for more examples, and Fig. 6 for comparison to other baselines. + +![](images/4704948b53dff98fedc81cc46dc44fbda5219fdd47888f8a2f5693b07902e656.jpg) + +# Abstract + +3D-aware Generative Adversarial Networks (GANs) have shown remarkable progress in learning to generate multi-view-consistent images and 3D geometries of scenes from collections of 2D images via neural volume rendering. Yet, the significant memory and computational costs of dense sampling in volume rendering have forced 3D GANs to adopt patch-based training or employ low-resolution rendering with post-processing 2D super resolution, which sacrifices multiview consistency and the quality of resolved geometry. Consequently, 3D GANs have not yet been able to fully resolve the rich 3D geometry present in 2D images. In this work, we propose techniques to scale neural volume rendering to the much higher resolution of native 2D images, thereby resolving fine-grained 3D geometry with + +unprecedented detail. Our approach employs learning-based samplers for accelerating neural rendering for 3D GAN training using up to 5 times fewer depth samples. This enables us to explicitly "render every pixel" of the full-resolution image during training and inference without post-processing superresolution in 2D. Together with our strategy to learn high-quality surface geometry, our method synthesizes high-resolution 3D geometry and strictly view-consistent images while maintaining image quality on par with baselines relying on post-processing super resolution. We demonstrate state-of-the-art 3D gemetic quality on FFHQ and AFHQ, setting a new standard for unsupervised learning of 3D shapes in 3D GANs. + +# 1. Introduction + +Training 3D generative models from the abundance of 2D images allows the creation of 3D representations of real-world objects for content creation and novel view synthe + +sis [59]. Recently, 3D-aware generative adversarial networks (3D GANs) [6, 7, 19, 43, 45, 49, 53, 69, 70, 74, 76] have emerged as a powerful way to learn 3D representations from collections of 2D images in an unsupervised fashion. These methods employ differentiable rendering to compare rendered 3D scenes with 2D data using adversarial training [18]. Among the various 3D representations, Neural Radiance Fields (NeRF) [38] have become a popular choice among recent successful 3D GANs. However, the significant computational and memory cost of volume rendering has prevented 3D GANs from scaling to high-resolution output. For instance, generating a single $512 \times 512$ image via volume rendering requires evaluating as many as 25 million depth samples, if 96 depth samples are used per ray using importance sampling [7, 10, 38, 53]. Given that GAN training typically requires rendering tens of millions of images, the training process could require evaluating hundreds of trillions of depth samples. + +During training, all intermediate operations must be stored in GPU memory for every depth sample for the backward pass. Therefore, existing methods resort to working on patches [10, 49, 53] or adopting a low resolution neural rendering combined with post-processing 2D super resolution (SR) [7, 19, 43, 45, 70]. However, patch-based methods have limited receptive fields over scenes, leading to unsatisfactory results, and the hybrid low-resolution rendering and SR scheme inevitably sacrifices the multiview consistency and the accuracy of 3D geometry. While many techniques have been developed to improve the image quality of 3D GANs to match that of 2D GANs, the challenge of resolving the corresponding high-resolution 3D geometry remains unsolved (see Figs. 1 and 6 for our results and comparison to the current state-of-the-art). + +Scaling 3D GANs to operate natively at the 2D pixel resolution requires a novel approach for sampling. Fig. 2 compares the state-of-the-art 3D GAN, EG3D model, trained with and without SR. EG3D employs 96 dense depth samples in total using two-pass importance sampling [38] during training, which requires half a terabyte of GPU memory at $256 \times 256$ resolution, making scaling to higher resolutions infeasible. Furthermore, Fig. 2 demonstrates that using 96 dense samples still results in undersampling, as evidenced by the speckle noise patterns visible in the zoomed-in view, leading to considerably worse FID (inset in Fig. 2). 3D GANs relying on post-processing SR layers can repair these undersampling artifacts at the cost of a high-fidelity 3D representation. + +In this work, we address the challenge of scaling neural volume rendering to high resolutions by explicitly rendering every pixel, ensuring that "what you see in $2D$ , is what you get in $3D$ — generating an unprecedented level of geometric details as well as strictly multiview-consistent images. Our contributions are the following: + +- We introduce an SDF-based 3D GAN to represent high + +![](images/446d301cc6d80354a008a7fb6b403bedb296d7d069fc2f74f7098c5b6d5338ed.jpg) +Figure 2. Samples from EG3D 256 model. Right: Volume rendering with 48 coarse samples and 48 fine samples per ray with two-pass importance sampling [38] results in undersampling, leading to noticeable noisy artifacts. Left: These artifacts are repaired by super resolution (SR). An unsharp mask has been applied to the zoomed views for presentation purposes. + +frequency geometry with spatially-varying surface tightness that increases throughout training (subsection 4.1 and 4.5), in turn facilitating low sample rendering. + +- We propose a generalizable learned sampler conditioned on cheap low-resolution information to enable full-resolution rendering during training for the first time (subsection 4.2 and 4.3). +- We show a robust sampling strategy for the learned sampler (subsection 4.4) that produces stable neural rendering using significantly fewer depth samples (see Fig. 8). Our sampler can operate with just 20 samples per ray compared to existing 3D GANs which must use at least 96 samples per ray. +- Together, our contributions result in the state-of-the-art geometry for 3D GANs while rendering with quality on par with SR baselines (see Fig. 1). For more results, see Fig. 5 and for comparison to other baselines, see Fig. 6. + +# 2. Related Work + +We begin by reviewing the prior-arts of 3D generative models and their current shortcomings. We then cover foundational techniques for 3D geometry representation and neural rendering from which we take inspiration. We then discuss existing methods for accelerating neural volume rendering, which usually operate per-scene. + +# 2.1. 3D Generative Models + +Just like 2D GANs, 3D-aware GANs train from a collection of 2D images, but employ a 3D representation and differentiable rendering to learn 3D scenes without requiring multiview images or ground truth 3D scans. Some of the most successful works use a neural field [68] in combination with a feature grid [7] as their 3D representation, and use neural volume rendering [38] as the differentiable renderer. + +However, due to the significant memory and computational cost of neural volume rendering, many previous works perform rendering at low-resolution and rely on a 2D post-processing CNN [7, 19, 43, 45, 70], which hallucinates the high-frequency details in a view-inconsistent manner while sacrificing 3D consistency and the quality of the resolved 3D geometry. + +To ensure strict 3D consistency, other previous works seek to render at high-resolutions and propose techniques to address the prohibitive computational costs. One line of work leverages the sparse nature of 3D scenes to speed up rendering, in particular, structures such as 2D manifolds [14, 67], multiplane images [75] and sparse voxels [50]. Although they are more efficient, sparse representations provide only coarse [50] or category-specific [14, 67] acceleration structures, which poses constraints on the diversity and viewing angles of the generated scenes. Our sampling-based method, on the other hand, generalizes to every new scene and adaptively accelerates rendering on a per ray basis. Another line of work enables high-resolution rendering with patch-based training [10, 53]. In particular, Mimic3D [10] achieves significantly improved 2D image quality, but the patch-based training limits the receptive fields, and the generated geometry does not faithfully represent the 2D data due to the patch-wise perceptual loss. Our method renders the entire image at once and the resulting geometry is aligned with the rendering (see Figs. 1 and 6). + +Recently, a new family of generative approaches using diffusion models has been proposed to tackle conditional tasks including novel view synthesis [8, 58] and text-based 3D generation[63]. Most of these 3D-aware diffusion models combine a 3D inductive bias modeled via neural field representations and a 2D image denoising objective to learn 3D scene generation. While these models enable unconditional 3D generation, they require multiview images [24, 55, 63] or 3D data, such as a point cloud [42]. Score Distillation Sampling [47] may be used for distillation from a pre-trained 2D diffusion model when only monocular 2D data is available, but diffusion models incur significant computational costs due to their iterative nature and most of the existing methods require optimization per scene [9, 20, 33, 61, 66]. + +# 2.2. Learning High-Fidelity Geometry + +Prior works on 3D GANs have typically represented the geometry as a radiance field [38], which lacks a concrete definition for where the surface geometry resides in the field, resulting in bumpy surfaces. A number of works [44, 62, 71] have proposed alternate representations based on implicit surfaces (such as signed distance functions, or SDFs) that can be used with neural volume rendering. In these works, the implicit surface is typically softened by a parameter for volume rendering. + +Other works [32, 64] improve on these implicit surface representations by leveraging feature grids for higher computational efficiency and resolution. Adaptive Shells [65] + +further improve on quality by making the softness parameter spatially-varying, as many objects have hard boundaries only in certain parts. We use an implicit surface representation based on VolSDF [72], and leverage a spatially-varying parameter similar to Adaptive Shells [65] to control the softness of the surface, as humans and animals benefit from both hard surfaces (e.g. skin, eyes) and soft, volumetric representations (e.g. hair). Although other works such as StyleSDF [45] have similarly leveraged implicit surfaces in a 3D GAN framework, the lack of spatial-variance and high-resolution rendering led to over-smoothed geometry not faithful to the rendered images. + +# 2.3. Accelerating Neural Volume Rendering + +As mentioned in Section 2.1, accelerating 3D GANs typically relies on acceleration structures such as octrees [31, 56, 73], which in generative settings [50] are limited due to computational requirements. In another direction, GRAM-based methods [14, 67] are limited due to sharing one structure across all samples. Instead, we look to a class of methods that do not rely on an acceleration structure. Some of these works learn a per-scene sampling prior on a per-ray basis using a binary classifier [41], a density estimator [30, 46], a coarse proxy from a 3D cost volume [34], or an interval length estimator [35] on discrete depth regions along the ray. Other works use importance sampling [21, 38] to sample additional points. We take inspiration from works on density estimators [30] and propose to learn a scene-conditional proposal network that generalizes across scenes instead of being category-specific or optimized per-scene. + +There are also other methods to accelerate rendering by utilizing more efficient representations, such as gaussian splats [29] and light field networks [52]. More efficient feature grids [40, 57] based on hashing can also be used to accelerate rendering. However, mapping to these representations in a GAN framework is not straightforward. In contrast, our sampling strategy can be used for any NeRF representation. + +# 3. Background + +We begin with background on the methodology of the state-of-the-art 3D-aware GANs as our method relies on a similar backbone for mapping to 3D representations. 3D GANs typically utilize a StyleGAN-like [27] architecture to map from a simple Gaussian prior to the conditioning of a NeRF, whether that be an MLP [19, 45], MPI [75], 3D feature grid [50], manifolds [14, 67] or triplane [7, 10, 53]. We inherit the latter triplane conditioning for its high expressivity and efficiency, in which three axis-aligned 2D feature grids $(f_{xy}, f_{xz}, f_{yz})$ , provide NeRF conditioning by orthogonal projection and interpolation. As in the previous methods, the mapping and synthesis networks from StyleGAN2 can easily be adapted to create the 2D triplane representation from noise $z \in \mathbb{R}^{512}$ . Specifically, $\mathbf{w} = \mathrm{Mapping}(z)$ conditions the creation of the triplane $T(\mathbf{w}) \in \mathbb{R}^{3 \times 32 \times 512 \times 512}$ + +![](images/4ec68d5485ab607a8284f4b908b6f86b7c472661c38c09857ca0928315e1b673.jpg) +Figure 3. Here we show our proposed pipeline and its intermediate outputs. Beginning from the triplane $T$ , we trace uniform samples to probe the scene, yielding low-resolution $I_{128}$ and weights $P_{128}$ . These are fed to a CNN which produces high-resolution proposal weights $\hat{P}_{512}$ (weights are visualized as uniform level sets). We perform robust sampling and volume render to get the final image $I_{512}$ and the surface variance $B$ . + +from a Synthesis network, corresponding to three axis-aligned 2D feature grids of spatial resolution $512 \times 512$ and feature size 32. + +To create high-fidelity geometry, our method builds upon VolSDF [72]: Instead of directly outputting the opacity $\sigma$ of a point $x\in \mathbb{R}^3$ , an SDF value $s$ is output and transformed to $\sigma$ using a Laplacian CDF: + +$$ +\sigma = \frac {1}{\beta} \left\{ \begin{array}{l l} \frac {1}{2} \exp \left(\frac {s}{\beta}\right) & \text {i f} s \leq 0 \\ 1 - \frac {1}{2} \exp \left(- \frac {s}{\beta}\right) & \text {i f} s > 0 \end{array} \right. \tag {1} +$$ + +where $\beta$ is the variance of the Laplacian distribution governing the "tightness" of the representation. One distinct benefit of an SDF-based representation is the ease of extracting the surface. StyleSDF [45] also utilizes this intermediate geometry representation without a triplane, enforcing the usual Eikonal constraint. + +Using the triplane conditioning and Eq. 1, we can assign each point in the volume with its opacity $\sigma$ and radiance $c$ using a lightweight MLP. For a given camera ray $r(t) = o + td$ , we approximate the volumetric rendering integral $C(r)$ [36] by sampling ray distances $t_i$ with their corresponding $\sigma_i$ and $c_i$ before computing + +$$ +\hat {C} (r) = \sum_ {i = 1} ^ {N} w _ {i} c _ {i} \text {w h e r e} w _ {i} = T _ {i} \left(1 - \exp \left(- \sigma_ {i} \delta_ {i}\right)\right), \tag {2} +$$ + +$$ +T _ {i} = \exp \left(- \sum_ {j = 1} ^ {i - 1} \sigma_ {j} \delta_ {j}\right) \text {a n d} \delta_ {i} = t _ {i + 1} - t _ {i}. \tag {2} +$$ + +Here, $T_{i}$ denotes the accumulated transmittance and $\delta_{i}$ is the distance between adjacent samples along the ray. + +It is possible to develop a more efficient estimator for this sum with fewer samples by using importance sampling techniques in computer graphics. Typically, one computes a piecewise constant probability distribution $p_j = \frac{\hat{w}_j}{\sum_j \hat{w}_j}$ , where $j$ refers to the $j^{\text{th}}$ bin or region, and $\hat{w}_j$ is an estimate of $w_j$ for that region, for example obtained by explicitly tracing coarse samples. For a given $t$ , we first find the region $j(t)$ and then set $p(t) = p_j$ . From this, one can compute a (piecewise linear) cumulative distribution function or CDF + +$\Phi (t)$ which has a range from 0 to 1. We can then perform inverse CDF sampling to define the sample points, + +$$ +t _ {i} = \Phi^ {- 1} (u _ {i}), \tag {3} +$$ + +where $u_{i}$ is a random number from 0 to 1 (sorted to be an increasing sequence). + +Discussion We improve on previous works such as NeRF [38] and EG3D [7] by stratifying2 the random numbers $u_{i}$ during training; this leads to significantly lower rendering variance, especially at low sample counts $N$ [39]. We also develop a neural method to predict a good distribution $p_{j}$ (and hence $\Phi$ ) for importance sampling at high spatial resolution, without needing to exhaustively step through the ray. + +# 4. Method + +In this section, we describe our method beginning with our SDF-based NeRF parametrization (subsection 4.1). We then overview how we render at high-resolution in three stages: first, a low-resolution probe into the 3D scene (subsection 4.2); second a high-resolution CNN proposal network (subsection 4.3); and third a robust sampling method for the resultant proposals (subsection 4.4). Next we describe regularizations (subsection 4.5) for stable training, and finally our entire training pipeline (subsection 4.6). + +# 4.1. Mapping to a 3D Representation + +Beginning from a noise vector $z$ , we synthesize the initial triplane $T'$ with StyleGAN [25] layers as detailed in Sec. 3. In contrast to previous methods, we then generate more expressive triplane features $T$ with an extra synthesis block for each orthogonal plane: $f_{ij} = \mathrm{SynthesisBlock}(f_{ij}')$ where $ij \in \{xy, xz, yz\}$ and $f'$ are the features of $T'$ . This design choice allows disentanglement between the intermediate triplane features as plane-specific kernels can attend to the features in each separate plane. + +Given the triplane $T$ (the left side of Fig. 3) and a point $x \in \mathbb{R}^3$ , we utilize an MLP to map to the SDF value $s$ , variance $\beta$ and geometry features $f_{geo}$ : + +$$ +(s, \beta , f _ {g e o}) = \operatorname {M L P} _ {\mathrm {S D F}} (\operatorname {P o s E n c} (x), T _ {x}) \tag {4} +$$ + +where PosEnc is the positional encoding from NeRF [38] and $T_{x}$ are the features corresponding to $x$ gathered from $T$ by projecting $x$ to each of the axis-aligned planes and taking the Hadamard product of the three resultant vectors [16]. We initialize the ReLU $\mathrm{MLP}_{\mathrm{SDF}}$ as in SAL [4] to ensure an approximately spherical SDF in the early part of training. Additionally note that $\beta$ varies spatially in the volume unlike [45, 72], allowing us to regularize its values later on. + +Using Eq. 1, we transform $s$ and $\beta$ into opacity $\sigma$ . We can now predict the radiance with a separate $\mathrm{MLP_c}$ conditioned on the geometry features and viewing direction $v$ as + +$$ +c = \mathrm {M L P} _ {\mathrm {c}} (\operatorname {P o s E n c} (v), f _ {g e o}). \tag {5} +$$ + +Note that unlike EG3D [7], we condition radiance on the viewing direction, allowing a more expressive generator. Thus, given a triplane, we can render any pixel by computing $\sigma$ and $c$ for points along the ray to approximate the volumetric rendering integral as described in Sec. 3. + +# 4.2. High-Resolution Proposal Network + +We now have our mapping from a latent code $z$ to a 3D NeRF representation. However, volumetric rendering of NeRFs at higher resolutions requires extremely large numbers of samples, and thus both memory and time. Instead of naive dense sampling at a high resolution, we propose to leverage low-resolution renderings to cheaply probe the 3D representation (visualized on the left of Fig. 3) for the creation of proposal distributions at high-resolution. Given a target camera and triplane $T$ , we first trace 192 coarse samples at low-resolution $(128 \times 128)$ to compute a low-resolution RGB image $I_{128} \in \mathbb{R}^{3 \times 128 \times 128}$ and a tensor of weights $P_{128} \in \mathbb{R}^{192 \times 128 \times 128}$ (visualized after low-resolution rendering in Fig. 3). Each 192-dimensional vector corresponds to a piecewise constant PDF with CDF $\Phi$ as seen in Eq. 3. + +Conditioned on the low-resolution probe, we predict a tensor of proposal volume rendering weights at the high-resolution $(512 \times 512)$ : + +$$ +\hat {P} _ {5 1 2} = \operatorname {S o f t m a x} \left(\operatorname {C N N} \left(P _ {1 2 8}, I _ {1 2 8}\right)\right) \in \mathbb {R} ^ {1 9 2 \times 5 1 2 \times 5 1 2}, \tag {6} +$$ + +where CNN is a lightweight network that up-samples the low-resolution weights, Softmax produces discrete distributions along each ray, and the $\hat{}$ denotes that this is an estimated quantity. This corresponds to the Proposal in Fig. 3 and the yellow distribution in Fig. 4. Note that allocating 192 samples at $128 \times 128$ is equivalent to allocating just 12 at $512 \times 512$ . + +# 4.3. Supervising the Proposal Network + +Having described the input and output of our high-resolution proposal network, we now show its supervision. From the target camera, we can also trace 192 coarse samples at high resolution for a small $64 \times 64$ patch, giving us a ground truth tensor of volume rendering weights + +$P_{\mathrm{patch}} \in \mathbb{R}^{192 \times 64 \times 64}$ . We then prepare this tensor for supervision by computing: + +$$ +\bar {P} _ {\text {p a t c h}} = \text {N o r m a l i z e} (\text {S u p p r e s s} (\text {B l u r} (P _ {\text {p a t c h}}))) \tag {7} +$$ + +where Blur applies a 1D Gaussian kernel to the input distributions, $\operatorname{Suppress}(x) = x$ if $x \geq 5e - 3$ and 0 otherwise, and Normalize is L1 normalization to create a valid distribution. This corresponds to the patch loss in Fig. 3 and the purple distribution in Fig. 4. These operations create less noisy distributions to facilitate accurate learning of the high-frequency integrand which may be undersampled in the coarse pass. + +We can then compare the predicted and cleaned ground truth distributions with a cross-entropy loss: + +$$ +L _ {\text {s a m p l e r}} = \text {C r o s s E n t r o p y} (\bar {P} _ {\text {p a t c h}}, \hat {P} _ {\text {p a t c h}}) \tag {8} +$$ + +where $\hat{P}_{\mathrm{patch}}$ is the corresponding patch of weights in $\hat{P}_{512}$ and CrossEntropy denotes the average cross-entropy between all pairs of pixelwise discrete distributions; for each pair $(\bar{p},\hat{p})$ , we compute $\sum -\bar{p}_j\log \hat{p}_j$ . Since we only need to compute this supervision for a small patch, the overhead of sampler training is not significant. + +# 4.4. Robustly Sampling from the Proposal Network + +Having shown how to train and predict high-resolution distributions, we now overview how to sample the resultant proposals. As seen in Fig. 4, the proposals are often slightly off; this is due to the high frequency nature of the underlying integrand in blue. + +In order to utilize the information from the sampler, we propose to filter predicted PDFs for better estimation. Specifically, for each discrete predicted PDF $\hat{p}$ , we compute the smallest set of bins whose probability exceeds a threshold $\tau = 0.98$ : We find the smallest subset $I \subseteq \{1,2,\dots,192\}$ such that + +$$ +\sum_ {i \in I} \hat {p} _ {i} \geq \tau . \tag {9} +$$ + +This operation resembles nucleus sampling in NLP [23]. We define our sampling PDF $q$ with probability $q_{i} = \frac{1}{|I|}$ if $i\in I$ and 0 otherwise (the green distribution in Fig. 4). + +For each PDF $q$ , we compute its CDF $\Phi$ and perform stratified inverse transform sampling to create the samples (illustrated as adaptive sampling near the surface in Fig. 3). We refer to these samples as robust stratified samples. In practice, on top of the 12 samples from the coarse probe (for the high-resolution image), we take an additional 18 samples per pixel adaptively based on the variance of the predicted distributions. The details are given in the supplement. + +# 4.5. Regularization for High-Resolution Training + +In order to render accurately under low sample budget per ray, we desire the surface to be tight, i.e., the set of points + +![](images/dfdb4927f0d76df639c0284c69bbae79d07a2a3bdf7d321886f4b05c0f9f86cb.jpg) +Figure 4. We visualize the volume rendering PDFs for the green pixel in the images on the right along with sampling methods. The ground truth distribution in blue is bimodal due to the discontinuous depth. Without stratification, the samples from the predicted yellow PDF completely miss the second mode. Stratification reduces the variance, yet also misses the second mode. Our robust stratified samples hit both modes despite the inaccurate predictions. The supervision PDF is visualized in purple as well. + +along the ray, which contribute to the accumulated radiance to be small. To accomplish this, we introduce a regularization for the spatially-varying $\beta$ values. Replacing the $c_{i}$ with the intermediate $\beta_{i}$ in Eq. 2, we can volume render an image of scalars, $B\in \mathbb{R}^{512\times 512}$ (seen on the right of Fig. 3). During training, we regularize $B$ to be small so that the surface tightens: + +$$ +L _ {\text {s u r f a c e}} = \sum_ {h w} \left(B _ {h w} - B _ {\text {t a r g e t}}\right) ^ {2}, \tag {10} +$$ + +where $B_{\mathrm{target}}$ is a scalar quantity annealed towards $\epsilon > 0$ during optimization. Note that this results in significantly more pronounced and smooth geometry (see Fig. 7). + +During training with the proposal network, the rendering networks only receive gradients for points very near the surface (which have high density). We find that this can lead to undesirable density growing from the background into the foreground. In order to combat this, we leverage the low-resolution information from $I_{128}$ , which is computed for uniform samples along the ray, and thus not concentrated at the surface. Considering the SDF values intermediately computed for rendering, $S = S_{128} \in \mathbb{R}^{192 \times 128 \times 128}$ , we enforce the SDF decision boundary by minimizing the SDF likelihood under a Laplacian distribution similar to [45, 51]: + +$$ +L _ {\mathrm {d e c}} = \sum_ {z h w} \exp (- 2 | S _ {z h w} |). \tag {11} +$$ + +# 4.6. The Training Pipeline + +In order to begin making use of the learned sampler for high-resolution rendering, we need a good NeRF representation from which to train it. Concurrently, we also need a + +good sampler in order to allow NeRF to render at $512 \times 512$ , the input resolution to the discriminator $D$ . + +To solve this issue, in the early stages of training, we first learn a low-resolution (e.g., $64 \times 64$ ) 3D GAN through the standard NeRF sampling techniques. We bilinearly up-sample our low-resolution renderings to $512 \times 512$ and blur the real images to the same level. After converging at the lower resolution, we introduce the sampler training (subsection 4.3). Concretely, we not only render low-resolution images with standard sampling, but also render sampler inputs $P_{128}$ and supervision patches $P_{\mathrm{patch}}$ . This results in a good initialization for the sampler. + +Having learned an initial low-resolution 3D GAN and high-resolution sampler, we transition to rendering with the sampler predictions. The high-resolution proposals $\hat{P}_{512}$ are downsampled to the current rendering resolution, which is progressively increased to the full $512 \times 512$ resolution during training. After introducing all losses, we optimize the parameters of the generator $G$ to minimize the following: + +$$ +L = L _ {\text {a d v}} + \lambda_ {\text {s a m p l e r}} L _ {\text {s a m p l e r}} + \lambda_ {\text {s u r f a c e}} L _ {\text {s u r f a c e}} + \lambda_ {\text {d e c}} L _ {\text {d e c}} \tag {12} +$$ + +where $L_{\mathrm{adv}}$ is the standard GAN loss [18] Softplus $(-D(G(z))$ . Note that we do not enforce the Eikonal constraint as we did not see a benefit. The discriminator $D$ is trained with R1 gradient regularization [37] whose weight is adapted as the resolution changes. Generator and discriminator pose conditioning follow EG3D [7]. The details of all hyperparameters and schedules are presented in the supplementary material. + +# 5. Results + +Datasets. We benchmark on two standard datasets for 3D GANs: FFHQ [25] and AFHQv2 Cats [11, 28] both at resolution $512 \times 512$ . We use the camera parameters extracted by EG3D [7] for conditioning and rendering. Our AFHQ model is finetuned from our FFHQ model using adaptive data augmentation [26]. For more results, please see the accompanying video. + +# 5.1. Comparisons + +Baselines. We compare our methods against state-of-the-art 3D GAN methods including those that use low-resolution neural rendering and 2D post-processing CNN super resolution: EG3D [7], MVCGAN [74], and StyleSDF [45]; and methods that operate entirely based on neural rendering: Mimic3D [10], Epigraph [53], GMPI [75], and GramHD [67]. + +Qualitative results Fig. 5 shows the curated samples generated by our method with FFHQ and AFHQ, demonstrating photorealistic rendering as well as high-resolution details that align with the 2D images. Fig 6 provides qualitative comparisons to baselines. EG3D shows significant artifacts and cannot resolve high-resolution geometry since it performs neural rendering at only $128 \times 128$ . Mimic3D + +![](images/b2345f918d5836ccee359e45e6a67f16fffa1dad7ebfb0f2ee0ff25cbc681ba4.jpg) + +![](images/c1c62382a665977f32737db82537e1c1915398b2997a89d9799e277261f3989f.jpg) + +![](images/f8edc5d0825c57234fa9e0b6a1419ab5fdb01aa96cb522e16778987d5113ff0f.jpg) +Figure 5. Curated samples on FFHQ and AFHQ. Our method can resolve high-fidelity geometry (e.g., eyeglasses) and fine-grained details (e.g., cat's fur) as seen in the geometry and normal map. + +![](images/4677efd6da01774ab35682c2e5a2f3f8ff815710ef99b9386a0e5456cd29ec47.jpg) +Figure 6. Qualitative comparisons on FFHQ with EG3D [7], Mimic3D [10] and Epigraph [53]. EG3D performs neural rendering at resolution $128 \times 128$ and relies on $4 \times$ super resolution to generate images. On the right, Mimic3D and Epigraph directly generate the image via neural rendering. While all other baselines use up to 192 dense depth samples per ray, our method can operate at 30 samples per ray. + +![](images/66224bbeaeefdc70b6533e3e531da5e32aac022db1f22e70a0f9a82ce81718ee.jpg) +Figure 7. Ablation study on the effect of beta regularization. + +and Epigraph render all pixels with neural rendering, but the patch-based nature of these methods harm the overall 3D geometry (e.g., distorted face and missing ear). Our method provides both high-fidelity 3D shape (e.g., well-defined ears and isolated clothing collar) and high-resolution details. + +Quantitative results. Tabs. 1 and 2 provide quantitative comparisons against baselines. We measure the image quality with Fréchet Inception Distance (FID) [22]. We assess the quality of the learned geometry with a face-specific Normal FID (FID-N) [15]. We render 10.79k normal maps from the NPHM [17] dataset by solving for the approximate alignment between the average FFHQ 2D landmarks + +and the provided 3D landmarks. Examples are given in the supplement. For each mesh, we render two views with approximately 20 degrees of pitch variation relative to the front of the face. These views are processed to remove the background with facer [13]. For each baseline, we render 10k normal maps and remove their background using the predicted mask from facer on the rendered image. We compute the FID between the two sets of images. Finally, we also evaluate the flatness of the learned representations with non-flatness score (NFS) [54]. + +Our results show the state-of-the-art image quality among the methods that operate only with neural rendering, while achieving FID comparable to the state-of-the-art SR-based method, EG3D. Our geometry quality outperforms all existing methods as indicated by our state-of-the-art FID-N. Additionally, our high NFS scores show the 3D aspect of our geometry. However, since NFS simply measures the variations of depth as a measure of non-flatness, it does not quantify the quality of geometry above a certain threshold. + +# 5.2. Ablation Study + +Without our surface tightness regularization (Eq. 10), the SDF surface may get fuzzy, resulting in a less clean surface + +
MethodFID ↓FID-N↓NFS↑
FFHQ-512Epigraph [7]9.92†67.3333.95
Mimic3D [10]5.3764.9716.76
GRAM-HD [67]12.2†*--
Ours4.9760.7629.35
w SREG3D [7]4.7063.0217.54
StyleSDF [45]11.19†87.4222.75
MVCGAN [74]13.4†*--
+ +Table 1. Quantitative comparison with baselines with and without super resolution (SR) on the FFHQ dataset. $\dagger$ as reported in the previous works. \* indicates FID evaluated on $20\mathrm{k}$ images. + +
MethodFID ↓NFS↑
AFHQ-512w/o SRGRAM-HD [67]\(7.67^{\dagger}\)-
GMPI [45]\(7.79^{\dagger}\)-
Mimic3D [10]4.2912.67
Ours4.2321.89
SREG3D [7]2.7714.14
StyleSDF [45]7.9133.89
+ +Table 2. Quantitative results on AFHQv2 Cats. † as reported in the previous works. + +
MethodFID ↓FID-N ↓NFS ↑
- Learned Sampler38.2993.8830.95
- Stratification5.6086.025.97
- Robust Sampling5.6760.7824.79
- Beta Regularization5.2764.2528.88
Ours4.9760.7629.35
+ +Table 3. Ablation study. + +(see Fig. 7) and worse geometry scores (see Tab. 3). Without our sampler or stratification during sampling, the model cannot learn meaningful 3D geometry with limited depth budgets, creating degenerated 3D geometry as can be seen in Fig. 8 and significantly worse FID-N. Without our robust sampling strategy, the sampling becomes more susceptible to slight errors in the sampler due to the high-frequency nature of the PDF (see Fig. 4), resulting in a noticeable drop in FID and occasionally creating floater geometry artifacts (see Fig. 8), while geometry scores remain similar. + +# 6. Discussion + +Limitations and future work. While our method demonstrates significant improvements in 3D geometry generation, it may still exhibit artifacts such as dents in the presence of specularities, and cannot handle transparent objects such as lenses well. Future work may incorporate more advanced material formulations [5] and surface normal regularization [60]. While 3D GANs can learn 3D represent + +![](images/75f5c6d21c946211b2e92c0b42968429b09d0a848085b008aae4e5a0fdbfe8e2.jpg) +Figure 8. Qualitative comparisons for ablation study. + +tations from single-view image collections such as FFHQ and AFHQ with casual camera labels [7], the frontal bias and inaccurate labels can result in geometry artifacts, especially on the side of the faces. Fruitful future directions may include training 3D GANs with large-scale Internet data as well as incorporating a more advanced form of regularization [47] and auto-camera calibration [3] to extend the generations to 360 degrees. Finally, our sampling-based acceleration method may be applied to other NeRFs. + +Ethical considerations. While existing methods have demonstrated effective capabilities in detecting unseen GANs [12], our contribution may remove certain characteristics from generated images, potentially making the task of detection more challenging. Viable solutions include the authentication of synthetic media [1, 2, 48]. + +Conclusion. We proposed a sampler-based method to accelerate 3D GANs to resolve 3D representations at the native resolution of 2D data, creating strictly multi-view-consistent images as well as highly detailed 3D geometry learned from a collection of in-the-wild 2D images. We believe our work opens up new possibilities for generating high-quality 3D models and synthetic data that capture in-the-wild variations and for enabling new applications such as conditional view synthesis. + +# Acknowledgements + +We thank David Luebke, Tero Karras, Michael Stengel, Amrita Mazumdar, Yash Belhe, and Nithin Raghavan for feedback on drafts and early discussions. Koki Nagano was partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. This work was funded in part by an NSF Graduate Fellowship, ONR Grant N00014-23-1-2526, and the Ronald L. Graham Chair. Manmohan Chandraker acknowledges support of of NSF IIS 2110409. Distribution Statement "A" (Approved for Public Release, Distribution Unlimited). + +# References + +[1] Coalition for content provenance and authenticity. https://c2pa.org/.8 +[2] Content authenticity initiative. https://contentauthenticity.org/.8 +[3] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panopehead: Geometry-aware 3d full-head synthesis in 360deg. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8 +[4] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5 +[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 8 +[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[8] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative novel view synthesis with 3D-aware diffusion models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 3 +[10] Xingyu Chen, Yu Deng, and Baoyuan Wang. Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. arXiv preprint arXiv:2303.09036, 2023. 2, 3, 6, 7, 8 +[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023. 8 +[13] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 7 +[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image + +generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 7 +[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 5 +[17] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 7 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 2, 6 +[19] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2, 3 +[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3 +[21] Kunal Gupta, Miloš Hasan, Zexiang Xu, Fujun Luan, Kulyan Sunkavalli, Xin Sun, Manmohan Chandraker, and Sai Bi. Mcnerf: Monte carlo rendering and denoising for real-time nerfs. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7 +[23] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5 +[24] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4, 6 +[26] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6 +[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 + +[28] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6 +[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 3 +[30] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. 2022. 3 +[31] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966, 2023. 3 +[32] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3 +[34] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3 +[35] David B Lindell, Julien NP Martel, and Gordon Wetzstein. AutoInt: Automatic integration for fast neural volume rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[36] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 4 +[37] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 6 +[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4, 5 +[39] Don P Mitchell. Consequences of stratified sampling in graphics. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 277-280, 1996. 4 +[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3 +[41] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 2021. 3 + +[42] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 3 +[43] Michael Niemeyer and Andreas Geiger. GIRAFFE: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8 +[46] M. Piala and R. Clark. Terminerf: Ray termination prediction for efficient neural rendering. In International Conference on 3D Vision (3DV), 2021. 3 +[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2022. 3, 8 +[48] Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo. Avatar fingerprinting for authorized use of synthetic talking-head videos. arXiv preprint arXiv:2305.03713, 2023. 8 +[49] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[50] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6 +[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 3 +[53] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 6, 7 +[54] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. arXiv preprint arXiv:2303.01416, 2023. 7 +[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 3 +[56] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In + +IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[57] Towaki Takikawa, Thomas Müller, Merlin Nimier-David, Alex Evans, Sanja Fidler, Alec Jacobson, and Alexander Keller. Compact neural graphics primitives with learned hash probing. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3 +[58] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3 +[59] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2 +[60] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8 +[61] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3 +[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[64] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[65] Zian Wang, Tiancheng Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. In ACM Transactions on Graphics (SIGGRAPH ASIA). 3 +[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3 +[67] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3, 6, 8 +[68] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in + +visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 2 +[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[70] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[71] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Ronen Basri, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 3 +[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 5 +[73] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[74] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 8 +[75] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In European Conference on Computer Vision (ECCV), 2022. 3, 6 +[76] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 2 \ No newline at end of file diff --git a/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/images.zip b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f47955808cf32ab9882c25ffad856dd07b325ee6 --- /dev/null +++ b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05b928c09e85d38a7f29febe3099af21ebae27326fb8ce1347d588d46416aeb8 +size 682849 diff --git a/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/layout.json b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b795c8e7f190984fa570b8f98de45a91fe953abc --- /dev/null +++ b/2024/What You See is What You GAN_ Rendering Every Pixel for High-Fidelity Geometry in 3D GANs/layout.json @@ -0,0 +1,10386 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 102, + 493, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 102, + 493, + 138 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 493, + 138 + ], + "type": "text", + "content": "What You See is What You GAN: Rendering Every Pixel for High-Fidelity Geometry in 3D GANs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 168, + 537, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 168, + 537, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 537, + 198 + ], + "type": "text", + "content": "Alex Trevithick\\*, Matthew Chan, Towaki Takikawa, Umar Iqbal, Shalini De Mello, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 205, + 419, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 205, + 419, + 220 + ], + "spans": [ + { + "bbox": [ + 175, + 205, + 419, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 205, + 419, + 220 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 175, + 205, + 419, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 205, + 419, + 220 + ], + "type": "text", + "content": "University of California, San Diego" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 239, + 288, + 441 + ], + "blocks": [ + { + "bbox": [ + 50, + 239, + 288, + 441 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 288, + 441 + ], + "type": "image", + "image_path": "2988b5876bd21cea5c88f1ef75f51ab3a2b45ef962f3edcd7ccd6791a48f670f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 447, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 45, + 447, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 45, + 447, + 547, + 502 + ], + "type": "text", + "content": "Figure 1. Left: Our results. The split view in the middle demonstrates the high degree of agreement between our 2D rendering and corresponding 3D geometry. Our method can learn fine-grained 3D details (e.g., eyeglass frame and cat's fur) that are geometrically well-aligned to 2D images without multiview or 3D scan data. Right: Comparison with EG3D [7]. Our tight SDF prior provides smooth and detailed surfaces on the face and hat while EG3D exhibits geometry artifacts and discrepancies between geometry and rendering. Please see Fig. 5 and the accompanying video for more examples, and Fig. 6 for comparison to other baselines." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 290, + 239, + 547, + 440 + ], + "blocks": [ + { + "bbox": [ + 290, + 239, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 290, + 239, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 290, + 239, + 547, + 440 + ], + "type": "image", + "image_path": "4704948b53dff98fedc81cc46dc44fbda5219fdd47888f8a2f5693b07902e656.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 505, + 192, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 505, + 192, + 517 + ], + "spans": [ + { + "bbox": [ + 143, + 505, + 192, + 517 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 530, + 290, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 530, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 530, + 290, + 693 + ], + "type": "text", + "content": "3D-aware Generative Adversarial Networks (GANs) have shown remarkable progress in learning to generate multi-view-consistent images and 3D geometries of scenes from collections of 2D images via neural volume rendering. Yet, the significant memory and computational costs of dense sampling in volume rendering have forced 3D GANs to adopt patch-based training or employ low-resolution rendering with post-processing 2D super resolution, which sacrifices multiview consistency and the quality of resolved geometry. Consequently, 3D GANs have not yet been able to fully resolve the rich 3D geometry present in 2D images. In this work, we propose techniques to scale neural volume rendering to the much higher resolution of native 2D images, thereby resolving fine-grained 3D geometry with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 506, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 657 + ], + "type": "text", + "content": "unprecedented detail. Our approach employs learning-based samplers for accelerating neural rendering for 3D GAN training using up to 5 times fewer depth samples. This enables us to explicitly \"render every pixel\" of the full-resolution image during training and inference without post-processing superresolution in 2D. Together with our strategy to learn high-quality surface geometry, our method synthesizes high-resolution 3D geometry and strictly view-consistent images while maintaining image quality on par with baselines relying on post-processing super resolution. We demonstrate state-of-the-art 3D gemetic quality on FFHQ and AFHQ, setting a new standard for unsupervised learning of 3D shapes in 3D GANs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 667, + 387, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 667, + 387, + 680 + ], + "spans": [ + { + "bbox": [ + 306, + 667, + 387, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 687, + 547, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 687, + 547, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 687, + 547, + 723 + ], + "type": "text", + "content": "Training 3D generative models from the abundance of 2D images allows the creation of 3D representations of real-world objects for content creation and novel view synthe" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 702, + 287, + 723 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 702, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 702, + 287, + 723 + ], + "type": "text", + "content": "*This project was initiated and substantially carried out during an internship at NVIDIA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "22765" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 279 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 279 + ], + "type": "text", + "content": "sis [59]. Recently, 3D-aware generative adversarial networks (3D GANs) [6, 7, 19, 43, 45, 49, 53, 69, 70, 74, 76] have emerged as a powerful way to learn 3D representations from collections of 2D images in an unsupervised fashion. These methods employ differentiable rendering to compare rendered 3D scenes with 2D data using adversarial training [18]. Among the various 3D representations, Neural Radiance Fields (NeRF) [38] have become a popular choice among recent successful 3D GANs. However, the significant computational and memory cost of volume rendering has prevented 3D GANs from scaling to high-resolution output. For instance, generating a single " + }, + { + "bbox": [ + 47, + 72, + 289, + 279 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 72, + 289, + 279 + ], + "type": "text", + "content": " image via volume rendering requires evaluating as many as 25 million depth samples, if 96 depth samples are used per ray using importance sampling [7, 10, 38, 53]. Given that GAN training typically requires rendering tens of millions of images, the training process could require evaluating hundreds of trillions of depth samples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 280, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 289, + 450 + ], + "type": "text", + "content": "During training, all intermediate operations must be stored in GPU memory for every depth sample for the backward pass. Therefore, existing methods resort to working on patches [10, 49, 53] or adopting a low resolution neural rendering combined with post-processing 2D super resolution (SR) [7, 19, 43, 45, 70]. However, patch-based methods have limited receptive fields over scenes, leading to unsatisfactory results, and the hybrid low-resolution rendering and SR scheme inevitably sacrifices the multiview consistency and the accuracy of 3D geometry. While many techniques have been developed to improve the image quality of 3D GANs to match that of 2D GANs, the challenge of resolving the corresponding high-resolution 3D geometry remains unsolved (see Figs. 1 and 6 for our results and comparison to the current state-of-the-art)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 451, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 289, + 612 + ], + "type": "text", + "content": "Scaling 3D GANs to operate natively at the 2D pixel resolution requires a novel approach for sampling. Fig. 2 compares the state-of-the-art 3D GAN, EG3D model, trained with and without SR. EG3D employs 96 dense depth samples in total using two-pass importance sampling [38] during training, which requires half a terabyte of GPU memory at " + }, + { + "bbox": [ + 46, + 451, + 289, + 612 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 451, + 289, + 612 + ], + "type": "text", + "content": " resolution, making scaling to higher resolutions infeasible. Furthermore, Fig. 2 demonstrates that using 96 dense samples still results in undersampling, as evidenced by the speckle noise patterns visible in the zoomed-in view, leading to considerably worse FID (inset in Fig. 2). 3D GANs relying on post-processing SR layers can repair these undersampling artifacts at the cost of a high-fidelity 3D representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "type": "text", + "content": "In this work, we address the challenge of scaling neural volume rendering to high resolutions by explicitly rendering every pixel, ensuring that \"what you see in " + }, + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "type": "text", + "content": ", is what you get in " + }, + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 46, + 612, + 289, + 681 + ], + "type": "text", + "content": " — generating an unprecedented level of geometric details as well as strictly multiview-consistent images. Our contributions are the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 683, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 695 + ], + "type": "text", + "content": "- We introduce an SDF-based 3D GAN to represent high" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 69, + 511, + 233 + ], + "blocks": [ + { + "bbox": [ + 342, + 69, + 511, + 233 + ], + "lines": [ + { + "bbox": [ + 342, + 69, + 511, + 233 + ], + "spans": [ + { + "bbox": [ + 342, + 69, + 511, + 233 + ], + "type": "image", + "image_path": "446d301cc6d80354a008a7fb6b403bedb296d7d069fc2f74f7098c5b6d5338ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 240, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 547, + 305 + ], + "type": "text", + "content": "Figure 2. Samples from EG3D 256 model. Right: Volume rendering with 48 coarse samples and 48 fine samples per ray with two-pass importance sampling [38] results in undersampling, leading to noticeable noisy artifacts. Left: These artifacts are repaired by super resolution (SR). An unsharp mask has been applied to the zoomed views for presentation purposes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 316, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 350 + ], + "type": "text", + "content": "frequency geometry with spatially-varying surface tightness that increases throughout training (subsection 4.1 and 4.5), in turn facilitating low sample rendering." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 350, + 545, + 511 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 350, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 545, + 394 + ], + "type": "text", + "content": "- We propose a generalizable learned sampler conditioned on cheap low-resolution information to enable full-resolution rendering during training for the first time (subsection 4.2 and 4.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 396, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 545, + 464 + ], + "type": "text", + "content": "- We show a robust sampling strategy for the learned sampler (subsection 4.4) that produces stable neural rendering using significantly fewer depth samples (see Fig. 8). Our sampler can operate with just 20 samples per ray compared to existing 3D GANs which must use at least 96 samples per ray." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 464, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 545, + 511 + ], + "type": "text", + "content": "- Together, our contributions result in the state-of-the-art geometry for 3D GANs while rendering with quality on par with SR baselines (see Fig. 1). For more results, see Fig. 5 and for comparison to other baselines, see Fig. 6." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 525, + 392, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 392, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 392, + 537 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 615 + ], + "type": "text", + "content": "We begin by reviewing the prior-arts of 3D generative models and their current shortcomings. We then cover foundational techniques for 3D geometry representation and neural rendering from which we take inspiration. We then discuss existing methods for accelerating neural volume rendering, which usually operate per-scene." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 433, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 433, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 433, + 635 + ], + "type": "text", + "content": "2.1. 3D Generative Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "type": "text", + "content": "Just like 2D GANs, 3D-aware GANs train from a collection of 2D images, but employ a 3D representation and differentiable rendering to learn 3D scenes without requiring multiview images or ground truth 3D scans. Some of the most successful works use a neural field [68] in combination with a feature grid [7] as their 3D representation, and use neural volume rendering [38] as the differentiable renderer." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 702, + 287, + 723 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 287, + 723 + ], + "type": "text", + "content": "Triplane resolution is doubled to compensate for the loss of capacity from removing the SR layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "22766" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 153 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 153 + ], + "type": "text", + "content": "However, due to the significant memory and computational cost of neural volume rendering, many previous works perform rendering at low-resolution and rely on a 2D post-processing CNN [7, 19, 43, 45, 70], which hallucinates the high-frequency details in a view-inconsistent manner while sacrificing 3D consistency and the quality of the resolved 3D geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 153, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 153, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 287, + 384 + ], + "type": "text", + "content": "To ensure strict 3D consistency, other previous works seek to render at high-resolutions and propose techniques to address the prohibitive computational costs. One line of work leverages the sparse nature of 3D scenes to speed up rendering, in particular, structures such as 2D manifolds [14, 67], multiplane images [75] and sparse voxels [50]. Although they are more efficient, sparse representations provide only coarse [50] or category-specific [14, 67] acceleration structures, which poses constraints on the diversity and viewing angles of the generated scenes. Our sampling-based method, on the other hand, generalizes to every new scene and adaptively accelerates rendering on a per ray basis. Another line of work enables high-resolution rendering with patch-based training [10, 53]. In particular, Mimic3D [10] achieves significantly improved 2D image quality, but the patch-based training limits the receptive fields, and the generated geometry does not faithfully represent the 2D data due to the patch-wise perceptual loss. Our method renders the entire image at once and the resulting geometry is aligned with the rendering (see Figs. 1 and 6)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 556 + ], + "type": "text", + "content": "Recently, a new family of generative approaches using diffusion models has been proposed to tackle conditional tasks including novel view synthesis [8, 58] and text-based 3D generation[63]. Most of these 3D-aware diffusion models combine a 3D inductive bias modeled via neural field representations and a 2D image denoising objective to learn 3D scene generation. While these models enable unconditional 3D generation, they require multiview images [24, 55, 63] or 3D data, such as a point cloud [42]. Score Distillation Sampling [47] may be used for distillation from a pre-trained 2D diffusion model when only monocular 2D data is available, but diffusion models incur significant computational costs due to their iterative nature and most of the existing methods require optimization per scene [9, 20, 33, 61, 66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 564, + 228, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 228, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 228, + 578 + ], + "type": "text", + "content": "2.2. Learning High-Fidelity Geometry" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 583, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 286, + 687 + ], + "type": "text", + "content": "Prior works on 3D GANs have typically represented the geometry as a radiance field [38], which lacks a concrete definition for where the surface geometry resides in the field, resulting in bumpy surfaces. A number of works [44, 62, 71] have proposed alternate representations based on implicit surfaces (such as signed distance functions, or SDFs) that can be used with neural volume rendering. In these works, the implicit surface is typically softened by a parameter for volume rendering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 687, + 286, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 687, + 286, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 286, + 723 + ], + "type": "text", + "content": "Other works [32, 64] improve on these implicit surface representations by leveraging feature grids for higher computational efficiency and resolution. Adaptive Shells [65]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 210 + ], + "type": "text", + "content": "further improve on quality by making the softness parameter spatially-varying, as many objects have hard boundaries only in certain parts. We use an implicit surface representation based on VolSDF [72], and leverage a spatially-varying parameter similar to Adaptive Shells [65] to control the softness of the surface, as humans and animals benefit from both hard surfaces (e.g. skin, eyes) and soft, volumetric representations (e.g. hair). Although other works such as StyleSDF [45] have similarly leveraged implicit surfaces in a 3D GAN framework, the lack of spatial-variance and high-resolution rendering led to over-smoothed geometry not faithful to the rendered images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 219, + 512, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 512, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 512, + 232 + ], + "type": "text", + "content": "2.3. Accelerating Neural Volume Rendering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 238, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 433 + ], + "type": "text", + "content": "As mentioned in Section 2.1, accelerating 3D GANs typically relies on acceleration structures such as octrees [31, 56, 73], which in generative settings [50] are limited due to computational requirements. In another direction, GRAM-based methods [14, 67] are limited due to sharing one structure across all samples. Instead, we look to a class of methods that do not rely on an acceleration structure. Some of these works learn a per-scene sampling prior on a per-ray basis using a binary classifier [41], a density estimator [30, 46], a coarse proxy from a 3D cost volume [34], or an interval length estimator [35] on discrete depth regions along the ray. Other works use importance sampling [21, 38] to sample additional points. We take inspiration from works on density estimators [30] and propose to learn a scene-conditional proposal network that generalizes across scenes instead of being category-specific or optimized per-scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 434, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 525 + ], + "type": "text", + "content": "There are also other methods to accelerate rendering by utilizing more efficient representations, such as gaussian splats [29] and light field networks [52]. More efficient feature grids [40, 57] based on hashing can also be used to accelerate rendering. However, mapping to these representations in a GAN framework is not straightforward. In contrast, our sampling strategy can be used for any NeRF representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 529, + 384, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 384, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 384, + 542 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "text", + "content": "We begin with background on the methodology of the state-of-the-art 3D-aware GANs as our method relies on a similar backbone for mapping to 3D representations. 3D GANs typically utilize a StyleGAN-like [27] architecture to map from a simple Gaussian prior to the conditioning of a NeRF, whether that be an MLP [19, 45], MPI [75], 3D feature grid [50], manifolds [14, 67] or triplane [7, 10, 53]. We inherit the latter triplane conditioning for its high expressivity and efficiency, in which three axis-aligned 2D feature grids " + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "inline_equation", + "content": "(f_{xy}, f_{xz}, f_{yz})" + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "text", + "content": ", provide NeRF conditioning by orthogonal projection and interpolation. As in the previous methods, the mapping and synthesis networks from StyleGAN2 can easily be adapted to create the 2D triplane representation from noise " + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathrm{Mapping}(z)" + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "text", + "content": " conditions the creation of the triplane " + }, + { + "bbox": [ + 304, + 550, + 545, + 723 + ], + "type": "inline_equation", + "content": "T(\\mathbf{w}) \\in \\mathbb{R}^{3 \\times 32 \\times 512 \\times 512}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "22767" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 545, + 171 + ], + "type": "image", + "image_path": "4ec68d5485ab607a8284f4b908b6f86b7c472661c38c09857ca0928315e1b673.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": "Figure 3. Here we show our proposed pipeline and its intermediate outputs. Beginning from the triplane " + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": ", we trace uniform samples to probe the scene, yielding low-resolution " + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "inline_equation", + "content": "I_{128}" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": " and weights " + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "inline_equation", + "content": "P_{128}" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": ". These are fed to a CNN which produces high-resolution proposal weights " + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\hat{P}_{512}" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": " (weights are visualized as uniform level sets). We perform robust sampling and volume render to get the final image " + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "inline_equation", + "content": "I_{512}" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": " and the surface variance " + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 180, + 547, + 223 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 233, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 266 + ], + "type": "text", + "content": "from a Synthesis network, corresponding to three axis-aligned 2D feature grids of spatial resolution " + }, + { + "bbox": [ + 46, + 233, + 287, + 266 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 233, + 287, + 266 + ], + "type": "text", + "content": " and feature size 32." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "text", + "content": "To create high-fidelity geometry, our method builds upon VolSDF [72]: Instead of directly outputting the opacity " + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "text", + "content": " of a point " + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "text", + "content": ", an SDF value " + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "text", + "content": " is output and transformed to " + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 267, + 288, + 315 + ], + "type": "text", + "content": " using a Laplacian CDF:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 321, + 287, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 321, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 89, + 321, + 287, + 360 + ], + "type": "interline_equation", + "content": "\\sigma = \\frac {1}{\\beta} \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\exp \\left(\\frac {s}{\\beta}\\right) & \\text {i f} s \\leq 0 \\\\ 1 - \\frac {1}{2} \\exp \\left(- \\frac {s}{\\beta}\\right) & \\text {i f} s > 0 \\end{array} \\right. \\tag {1}", + "image_path": "a4193b7d7e3dcc376d4762ae2514f4502fdbd2fee4018befe2613799dfeffe0a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 364, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 364, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 364, + 287, + 432 + ], + "type": "text", + "content": " is the variance of the Laplacian distribution governing the \"tightness\" of the representation. One distinct benefit of an SDF-based representation is the ease of extracting the surface. StyleSDF [45] also utilizes this intermediate geometry representation without a triplane, enforcing the usual Eikonal constraint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": "Using the triplane conditioning and Eq. 1, we can assign each point in the volume with its opacity " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": " and radiance " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": " using a lightweight MLP. For a given camera ray " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "r(t) = o + td" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": ", we approximate the volumetric rendering integral " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "C(r)" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": " [36] by sampling ray distances " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": " with their corresponding " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 46, + 433, + 287, + 502 + ], + "type": "text", + "content": " before computing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 510, + 287, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 551 + ], + "type": "interline_equation", + "content": "\\hat {C} (r) = \\sum_ {i = 1} ^ {N} w _ {i} c _ {i} \\text {w h e r e} w _ {i} = T _ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right), \\tag {2}", + "image_path": "e40a9a6ca60fc7514aba607b022d7d19a4d6d225d3f32ad41c767104c8b17481.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 544, + 286, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 544, + 286, + 583 + ], + "spans": [ + { + "bbox": [ + 83, + 544, + 286, + 583 + ], + "type": "interline_equation", + "content": "T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}\\right) \\text {a n d} \\delta_ {i} = t _ {i + 1} - t _ {i}. \\tag {2}", + "image_path": "1f90fb32402b5a109e291e6bd7accb844e933c95a31c221d86247881d68aa92b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": " denotes the accumulated transmittance and " + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": " is the distance between adjacent samples along the ray." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": "It is possible to develop a more efficient estimator for this sum with fewer samples by using importance sampling techniques in computer graphics. Typically, one computes a piecewise constant probability distribution " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "p_j = \\frac{\\hat{w}_j}{\\sum_j \\hat{w}_j}" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "j^{\\text{th}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": " bin or region, and " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "\\hat{w}_j" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": " is an estimate of " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": " for that region, for example obtained by explicitly tracing coarse samples. For a given " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": ", we first find the region " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "j(t)" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": " and then set " + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "inline_equation", + "content": "p(t) = p_j" + }, + { + "bbox": [ + 46, + 613, + 288, + 723 + ], + "type": "text", + "content": ". From this, one can compute a (piecewise linear) cumulative distribution function or CDF" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 233, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\Phi (t)" + }, + { + "bbox": [ + 305, + 233, + 545, + 257 + ], + "type": "text", + "content": " which has a range from 0 to 1. We can then perform inverse CDF sampling to define the sample points," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 260, + 545, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 260, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 394, + 260, + 545, + 275 + ], + "type": "interline_equation", + "content": "t _ {i} = \\Phi^ {- 1} (u _ {i}), \\tag {3}", + "image_path": "2a18ce392c142834fcaabd169be1ec67e005fab49fd5b859bae63a5f650c865f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 278, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 302 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 278, + 545, + 302 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 304, + 278, + 545, + 302 + ], + "type": "text", + "content": " is a random number from 0 to 1 (sorted to be an increasing sequence)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "text", + "content": "Discussion We improve on previous works such as NeRF [38] and EG3D [7] by stratifying2 the random numbers " + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "text", + "content": " during training; this leads to significantly lower rendering variance, especially at low sample counts " + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "text", + "content": " [39]. We also develop a neural method to predict a good distribution " + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "inline_equation", + "content": "p_{j}" + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "text", + "content": " (and hence " + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 303, + 545, + 395 + ], + "type": "text", + "content": ") for importance sampling at high spatial resolution, without needing to exhaustively step through the ray." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 403, + 361, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 361, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 361, + 415 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 423, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 528 + ], + "type": "text", + "content": "In this section, we describe our method beginning with our SDF-based NeRF parametrization (subsection 4.1). We then overview how we render at high-resolution in three stages: first, a low-resolution probe into the 3D scene (subsection 4.2); second a high-resolution CNN proposal network (subsection 4.3); and third a robust sampling method for the resultant proposals (subsection 4.4). Next we describe regularizations (subsection 4.5) for stable training, and finally our entire training pipeline (subsection 4.6)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 532, + 481, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 532, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 532, + 481, + 544 + ], + "type": "text", + "content": "4.1. Mapping to a 3D Representation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": "Beginning from a noise vector " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": ", we synthesize the initial triplane " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": " with StyleGAN [25] layers as detailed in Sec. 3. In contrast to previous methods, we then generate more expressive triplane features " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": " with an extra synthesis block for each orthogonal plane: " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "f_{ij} = \\mathrm{SynthesisBlock}(f_{ij}')" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "ij \\in \\{xy, xz, yz\\}" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": " are the features of " + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 304, + 550, + 545, + 653 + ], + "type": "text", + "content": ". This design choice allows disentanglement between the intermediate triplane features as plane-specific kernels can attend to the features in each separate plane." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "Given the triplane " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": " (the left side of Fig. 3) and a point " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": ", we utilize an MLP to map to the SDF value " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": ", variance " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": " and geometry features " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "f_{geo}" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 693, + 545, + 707 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 693, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 345, + 693, + 545, + 707 + ], + "type": "interline_equation", + "content": "(s, \\beta , f _ {g e o}) = \\operatorname {M L P} _ {\\mathrm {S D F}} (\\operatorname {P o s E n c} (x), T _ {x}) \\tag {4}", + "image_path": "ea81f2556ce181348f9bc2d132611429a89bbf1f01439ca5fc7179a926eeb318.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 712, + 545, + 722 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 712, + 545, + 722 + ], + "spans": [ + { + "bbox": [ + 315, + 712, + 545, + 722 + ], + "type": "text", + "content": "2 dividing the unit interval into bins and taking a sample from each bin." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "22768" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": "where PosEnc is the positional encoding from NeRF [38] and " + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "inline_equation", + "content": "T_{x}" + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": " are the features corresponding to " + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": " gathered from " + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": " by projecting " + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": " to each of the axis-aligned planes and taking the Hadamard product of the three resultant vectors [16]. We initialize the ReLU " + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\mathrm{SDF}}" + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": " as in SAL [4] to ensure an approximately spherical SDF in the early part of training. Additionally note that " + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 72, + 286, + 164 + ], + "type": "text", + "content": " varies spatially in the volume unlike [45, 72], allowing us to regularize its values later on." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "content": "Using Eq. 1, we transform " + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "content": " into opacity " + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "content": ". We can now predict the radiance with a separate " + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP_c}" + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "content": " conditioned on the geometry features and viewing direction " + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 164, + 286, + 200 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 212, + 286, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 212, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 286, + 224 + ], + "type": "interline_equation", + "content": "c = \\mathrm {M L P} _ {\\mathrm {c}} (\\operatorname {P o s E n c} (v), f _ {g e o}). \\tag {5}", + "image_path": "280e024ddf3ee1046a8482eed6e61c05ee6253cd5983773f8a703f9bdb02e55c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "type": "text", + "content": "Note that unlike EG3D [7], we condition radiance on the viewing direction, allowing a more expressive generator. Thus, given a triplane, we can render any pixel by computing " + }, + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 232, + 286, + 289 + ], + "type": "text", + "content": " for points along the ray to approximate the volumetric rendering integral as described in Sec. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 234, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 234, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 234, + 310 + ], + "type": "text", + "content": "4.2. High-Resolution Proposal Network" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": "We now have our mapping from a latent code " + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": " to a 3D NeRF representation. However, volumetric rendering of NeRFs at higher resolutions requires extremely large numbers of samples, and thus both memory and time. Instead of naive dense sampling at a high resolution, we propose to leverage low-resolution renderings to cheaply probe the 3D representation (visualized on the left of Fig. 3) for the creation of proposal distributions at high-resolution. Given a target camera and triplane " + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": ", we first trace 192 coarse samples at low-resolution " + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": " to compute a low-resolution RGB image " + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "inline_equation", + "content": "I_{128} \\in \\mathbb{R}^{3 \\times 128 \\times 128}" + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": " and a tensor of weights " + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "inline_equation", + "content": "P_{128} \\in \\mathbb{R}^{192 \\times 128 \\times 128}" + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": " (visualized after low-resolution rendering in Fig. 3). Each 192-dimensional vector corresponds to a piecewise constant PDF with CDF " + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 316, + 286, + 489 + ], + "type": "text", + "content": " as seen in Eq. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 489, + 286, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 286, + 523 + ], + "type": "text", + "content": "Conditioned on the low-resolution probe, we predict a tensor of proposal volume rendering weights at the high-resolution " + }, + { + "bbox": [ + 46, + 489, + 286, + 523 + ], + "type": "inline_equation", + "content": "(512 \\times 512)" + }, + { + "bbox": [ + 46, + 489, + 286, + 523 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 533, + 286, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 286, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 286, + 547 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {5 1 2} = \\operatorname {S o f t m a x} \\left(\\operatorname {C N N} \\left(P _ {1 2 8}, I _ {1 2 8}\\right)\\right) \\in \\mathbb {R} ^ {1 9 2 \\times 5 1 2 \\times 5 1 2}, \\tag {6}", + "image_path": "66a3d2ebfdfd96c36e12696a2c146d1460637259f98a6dbd909ef2cf79d25003.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "text", + "content": "where CNN is a lightweight network that up-samples the low-resolution weights, Softmax produces discrete distributions along each ray, and the " + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "inline_equation", + "content": "\\hat{}" + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "text", + "content": " denotes that this is an estimated quantity. This corresponds to the Proposal in Fig. 3 and the yellow distribution in Fig. 4. Note that allocating 192 samples at " + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "text", + "content": " is equivalent to allocating just 12 at " + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 556, + 286, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 646, + 230, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 230, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 230, + 658 + ], + "type": "text", + "content": "4.3. Supervising the Proposal Network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 664, + 286, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 286, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 286, + 723 + ], + "type": "text", + "content": "Having described the input and output of our high-resolution proposal network, we now show its supervision. From the target camera, we can also trace 192 coarse samples at high resolution for a small " + }, + { + "bbox": [ + 46, + 664, + 286, + 723 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 664, + 286, + 723 + ], + "type": "text", + "content": " patch, giving us a ground truth tensor of volume rendering weights" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{patch}} \\in \\mathbb{R}^{192 \\times 64 \\times 64}" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": ". We then prepare this tensor for supervision by computing:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 103, + 544, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 103, + 544, + 118 + ], + "spans": [ + { + "bbox": [ + 335, + 103, + 544, + 118 + ], + "type": "interline_equation", + "content": "\\bar {P} _ {\\text {p a t c h}} = \\text {N o r m a l i z e} (\\text {S u p p r e s s} (\\text {B l u r} (P _ {\\text {p a t c h}}))) \\tag {7}", + "image_path": "6c25773029efc0826282622220c94cd81af962a5965e3778b80b4453f67183e7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "type": "text", + "content": "where Blur applies a 1D Gaussian kernel to the input distributions, " + }, + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\operatorname{Suppress}(x) = x" + }, + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "type": "inline_equation", + "content": "x \\geq 5e - 3" + }, + { + "bbox": [ + 304, + 125, + 545, + 217 + ], + "type": "text", + "content": " and 0 otherwise, and Normalize is L1 normalization to create a valid distribution. This corresponds to the patch loss in Fig. 3 and the purple distribution in Fig. 4. These operations create less noisy distributions to facilitate accurate learning of the high-frequency integrand which may be undersampled in the coarse pass." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "type": "text", + "content": "We can then compare the predicted and cleaned ground truth distributions with a cross-entropy loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 249, + 544, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 249, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 347, + 249, + 544, + 262 + ], + "type": "interline_equation", + "content": "L _ {\\text {s a m p l e r}} = \\text {C r o s s E n t r o p y} (\\bar {P} _ {\\text {p a t c h}}, \\hat {P} _ {\\text {p a t c h}}) \\tag {8}", + "image_path": "9f73c9b15f4731ed33be4e03c2ae745a4d491ddc7af51933cac0dfecc6cd1d88.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "text", + "content": " is the corresponding patch of weights in " + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\hat{P}_{512}" + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "text", + "content": " and CrossEntropy denotes the average cross-entropy between all pairs of pixelwise discrete distributions; for each pair " + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "inline_equation", + "content": "(\\bar{p},\\hat{p})" + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "text", + "content": ", we compute " + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\sum -\\bar{p}_j\\log \\hat{p}_j" + }, + { + "bbox": [ + 304, + 271, + 545, + 342 + ], + "type": "text", + "content": ". Since we only need to compute this supervision for a small patch, the overhead of sampler training is not significant." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 348, + 545, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 361 + ], + "type": "text", + "content": "4.4. Robustly Sampling from the Proposal Network" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 366, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 423 + ], + "type": "text", + "content": "Having shown how to train and predict high-resolution distributions, we now overview how to sample the resultant proposals. As seen in Fig. 4, the proposals are often slightly off; this is due to the high frequency nature of the underlying integrand in blue." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "text", + "content": "In order to utilize the information from the sampler, we propose to filter predicted PDFs for better estimation. Specifically, for each discrete predicted PDF " + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "text", + "content": ", we compute the smallest set of bins whose probability exceeds a threshold " + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\tau = 0.98" + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "text", + "content": ": We find the smallest subset " + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "inline_equation", + "content": "I \\subseteq \\{1,2,\\dots,192\\}" + }, + { + "bbox": [ + 304, + 424, + 545, + 493 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 400, + 502, + 544, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 502, + 544, + 527 + ], + "spans": [ + { + "bbox": [ + 400, + 502, + 544, + 527 + ], + "type": "interline_equation", + "content": "\\sum_ {i \\in I} \\hat {p} _ {i} \\geq \\tau . \\tag {9}", + "image_path": "405ebf01ed0cbc19054394a5aafdd445d215a309d68b323804263a59d2d866c4.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "text", + "content": "This operation resembles nucleus sampling in NLP [23]. We define our sampling PDF " + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "text", + "content": " with probability " + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "inline_equation", + "content": "q_{i} = \\frac{1}{|I|}" + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "inline_equation", + "content": "i\\in I" + }, + { + "bbox": [ + 304, + 534, + 544, + 570 + ], + "type": "text", + "content": " and 0 otherwise (the green distribution in Fig. 4)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "type": "text", + "content": "For each PDF " + }, + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "type": "text", + "content": ", we compute its CDF " + }, + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 571, + 545, + 673 + ], + "type": "text", + "content": " and perform stratified inverse transform sampling to create the samples (illustrated as adaptive sampling near the surface in Fig. 3). We refer to these samples as robust stratified samples. In practice, on top of the 12 samples from the coarse probe (for the high-resolution image), we take an additional 18 samples per pixel adaptively based on the variance of the predicted distributions. The details are given in the supplement." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 681, + 537, + 694 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 681, + 537, + 694 + ], + "spans": [ + { + "bbox": [ + 305, + 681, + 537, + 694 + ], + "type": "text", + "content": "4.5. Regularization for High-Resolution Training" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 699, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 699, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 699, + 545, + 723 + ], + "type": "text", + "content": "In order to render accurately under low sample budget per ray, we desire the surface to be tight, i.e., the set of points" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "22769" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 289, + 236 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 236 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 236 + ], + "type": "image", + "image_path": "dfdb4927f0d76df639c0284c69bbae79d07a2a3bdf7d321886f4b05c0f9f86cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 289, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 329 + ], + "type": "text", + "content": "Figure 4. We visualize the volume rendering PDFs for the green pixel in the images on the right along with sampling methods. The ground truth distribution in blue is bimodal due to the discontinuous depth. Without stratification, the samples from the predicted yellow PDF completely miss the second mode. Stratification reduces the variance, yet also misses the second mode. Our robust stratified samples hit both modes despite the inaccurate predictions. The supervision PDF is visualized in purple as well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "content": "along the ray, which contribute to the accumulated radiance to be small. To accomplish this, we introduce a regularization for the spatially-varying " + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "content": " values. Replacing the " + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "content": " with the intermediate " + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "content": " in Eq. 2, we can volume render an image of scalars, " + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "inline_equation", + "content": "B\\in \\mathbb{R}^{512\\times 512}" + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "content": " (seen on the right of Fig. 3). During training, we regularize " + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 341, + 289, + 422 + ], + "type": "text", + "content": " to be small so that the surface tightens:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 431, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 431, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 102, + 431, + 287, + 456 + ], + "type": "interline_equation", + "content": "L _ {\\text {s u r f a c e}} = \\sum_ {h w} \\left(B _ {h w} - B _ {\\text {t a r g e t}}\\right) ^ {2}, \\tag {10}", + "image_path": "591b848b667f41120ddeebcff4e3152b3c9c2e39979266e49f0ed870f302a716.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{target}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "type": "text", + "content": " is a scalar quantity annealed towards " + }, + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 46, + 466, + 287, + 501 + ], + "type": "text", + "content": " during optimization. Note that this results in significantly more pronounced and smooth geometry (see Fig. 7)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "type": "text", + "content": "During training with the proposal network, the rendering networks only receive gradients for points very near the surface (which have high density). We find that this can lead to undesirable density growing from the background into the foreground. In order to combat this, we leverage the low-resolution information from " + }, + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "type": "inline_equation", + "content": "I_{128}" + }, + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "type": "text", + "content": ", which is computed for uniform samples along the ray, and thus not concentrated at the surface. Considering the SDF values intermediately computed for rendering, " + }, + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "type": "inline_equation", + "content": "S = S_{128} \\in \\mathbb{R}^{192 \\times 128 \\times 128}" + }, + { + "bbox": [ + 46, + 501, + 288, + 628 + ], + "type": "text", + "content": ", we enforce the SDF decision boundary by minimizing the SDF likelihood under a Laplacian distribution similar to [45, 51]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 636, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 636, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 108, + 636, + 287, + 662 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {d e c}} = \\sum_ {z h w} \\exp (- 2 | S _ {z h w} |). \\tag {11}", + "image_path": "a74d0430de3e4bc2fc1f56f23e382b861bc28f7a573e978c5cf6665544bda3b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 173, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 173, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 173, + 682 + ], + "type": "text", + "content": "4.6. The Training Pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "type": "text", + "content": "In order to begin making use of the learned sampler for high-resolution rendering, we need a good NeRF representation from which to train it. Concurrently, we also need a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "good sampler in order to allow NeRF to render at " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": ", the input resolution to the discriminator " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "text", + "content": "To solve this issue, in the early stages of training, we first learn a low-resolution (e.g., " + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "text", + "content": ") 3D GAN through the standard NeRF sampling techniques. We bilinearly up-sample our low-resolution renderings to " + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "text", + "content": " and blur the real images to the same level. After converging at the lower resolution, we introduce the sampler training (subsection 4.3). Concretely, we not only render low-resolution images with standard sampling, but also render sampler inputs " + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "inline_equation", + "content": "P_{128}" + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "text", + "content": " and supervision patches " + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 95, + 547, + 210 + ], + "type": "text", + "content": ". This results in a good initialization for the sampler." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": "Having learned an initial low-resolution 3D GAN and high-resolution sampler, we transition to rendering with the sampler predictions. The high-resolution proposals " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\hat{P}_{512}" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " are downsampled to the current rendering resolution, which is progressively increased to the full " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " resolution during training. After introducing all losses, we optimize the parameters of the generator " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " to minimize the following:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 300, + 545, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 300, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 310, + 300, + 545, + 313 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {a d v}} + \\lambda_ {\\text {s a m p l e r}} L _ {\\text {s a m p l e r}} + \\lambda_ {\\text {s u r f a c e}} L _ {\\text {s u r f a c e}} + \\lambda_ {\\text {d e c}} L _ {\\text {d e c}} \\tag {12}", + "image_path": "70587fea66ac599f86f505db28a451d4670332e053f3af85c460293f8cda79b2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{adv}}" + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "text", + "content": " is the standard GAN loss [18] Softplus " + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "inline_equation", + "content": "(-D(G(z))" + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "text", + "content": ". Note that we do not enforce the Eikonal constraint as we did not see a benefit. The discriminator " + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 319, + 545, + 413 + ], + "type": "text", + "content": " is trained with R1 gradient regularization [37] whose weight is adapted as the resolution changes. Generator and discriminator pose conditioning follow EG3D [7]. The details of all hyperparameters and schedules are presented in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 422, + 359, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 359, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 359, + 434 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 441, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 523 + ], + "type": "text", + "content": "Datasets. We benchmark on two standard datasets for 3D GANs: FFHQ [25] and AFHQv2 Cats [11, 28] both at resolution " + }, + { + "bbox": [ + 304, + 441, + 545, + 523 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 441, + 545, + 523 + ], + "type": "text", + "content": ". We use the camera parameters extracted by EG3D [7] for conditioning and rendering. Our AFHQ model is finetuned from our FFHQ model using adaptive data augmentation [26]. For more results, please see the accompanying video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 529, + 391, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 391, + 541 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 391, + 541 + ], + "type": "text", + "content": "5.1. Comparisons" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 547, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 628 + ], + "type": "text", + "content": "Baselines. We compare our methods against state-of-the-art 3D GAN methods including those that use low-resolution neural rendering and 2D post-processing CNN super resolution: EG3D [7], MVCGAN [74], and StyleSDF [45]; and methods that operate entirely based on neural rendering: Mimic3D [10], Epigraph [53], GMPI [75], and GramHD [67]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "type": "text", + "content": "Qualitative results Fig. 5 shows the curated samples generated by our method with FFHQ and AFHQ, demonstrating photorealistic rendering as well as high-resolution details that align with the 2D images. Fig 6 provides qualitative comparisons to baselines. EG3D shows significant artifacts and cannot resolve high-resolution geometry since it performs neural rendering at only " + }, + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 641, + 545, + 723 + ], + "type": "text", + "content": ". Mimic3D" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "22770" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 298, + 175 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 298, + 175 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 298, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 298, + 175 + ], + "type": "image", + "image_path": "b2345f918d5836ccee359e45e6a67f16fffa1dad7ebfb0f2ee0ff25cbc681ba4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 545, + 175 + ], + "type": "image", + "image_path": "c1c62382a665977f32737db82537e1c1915398b2997a89d9799e277261f3989f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 207, + 296, + 393 + ], + "blocks": [ + { + "bbox": [ + 46, + 183, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 546, + 205 + ], + "type": "text", + "content": "Figure 5. Curated samples on FFHQ and AFHQ. Our method can resolve high-fidelity geometry (e.g., eyeglasses) and fine-grained details (e.g., cat's fur) as seen in the geometry and normal map." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 207, + 296, + 393 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 393 + ], + "type": "image", + "image_path": "f8edc5d0825c57234fa9e0b6a1419ab5fdb01aa96cb522e16778987d5113ff0f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 207, + 538, + 393 + ], + "blocks": [ + { + "bbox": [ + 298, + 207, + 538, + 393 + ], + "lines": [ + { + "bbox": [ + 298, + 207, + 538, + 393 + ], + "spans": [ + { + "bbox": [ + 298, + 207, + 538, + 393 + ], + "type": "image", + "image_path": "4677efd6da01774ab35682c2e5a2f3f8ff815710ef99b9386a0e5456cd29ec47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons on FFHQ with EG3D [7], Mimic3D [10] and Epigraph [53]. EG3D performs neural rendering at resolution " + }, + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "type": "text", + "content": " and relies on " + }, + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 402, + 546, + 436 + ], + "type": "text", + "content": " super resolution to generate images. On the right, Mimic3D and Epigraph directly generate the image via neural rendering. While all other baselines use up to 192 dense depth samples per ray, our method can operate at 30 samples per ray." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 446, + 287, + 527 + ], + "blocks": [ + { + "bbox": [ + 48, + 446, + 287, + 527 + ], + "lines": [ + { + "bbox": [ + 48, + 446, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 446, + 287, + 527 + ], + "type": "image", + "image_path": "66224bbeaeefdc70b6533e3e531da5e32aac022db1f22e70a0f9a82ce81718ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 535, + 276, + 547 + ], + "lines": [ + { + "bbox": [ + 57, + 535, + 276, + 547 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 276, + 547 + ], + "type": "text", + "content": "Figure 7. Ablation study on the effect of beta regularization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 625 + ], + "type": "text", + "content": "and Epigraph render all pixels with neural rendering, but the patch-based nature of these methods harm the overall 3D geometry (e.g., distorted face and missing ear). Our method provides both high-fidelity 3D shape (e.g., well-defined ears and isolated clothing collar) and high-resolution details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 723 + ], + "type": "text", + "content": "Quantitative results. Tabs. 1 and 2 provide quantitative comparisons against baselines. We measure the image quality with Fréchet Inception Distance (FID) [22]. We assess the quality of the learned geometry with a face-specific Normal FID (FID-N) [15]. We render 10.79k normal maps from the NPHM [17] dataset by solving for the approximate alignment between the average FFHQ 2D landmarks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 562 + ], + "type": "text", + "content": "and the provided 3D landmarks. Examples are given in the supplement. For each mesh, we render two views with approximately 20 degrees of pitch variation relative to the front of the face. These views are processed to remove the background with facer [13]. For each baseline, we render 10k normal maps and remove their background using the predicted mask from facer on the rendered image. We compute the FID between the two sets of images. Finally, we also evaluate the flatness of the learned representations with non-flatness score (NFS) [54]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 563, + 546, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 546, + 668 + ], + "type": "text", + "content": "Our results show the state-of-the-art image quality among the methods that operate only with neural rendering, while achieving FID comparable to the state-of-the-art SR-based method, EG3D. Our geometry quality outperforms all existing methods as indicated by our state-of-the-art FID-N. Additionally, our high NFS scores show the 3D aspect of our geometry. However, since NFS simply measures the variations of depth as a measure of non-flatness, it does not quantify the quality of geometry above a certain threshold." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 679, + 400, + 692 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 679, + 400, + 692 + ], + "spans": [ + { + "bbox": [ + 306, + 679, + 400, + 692 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 699, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 699, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 306, + 699, + 545, + 723 + ], + "type": "text", + "content": "Without our surface tightness regularization (Eq. 10), the SDF surface may get fuzzy, resulting in a less clean surface" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "22771" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 269, + 181 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 269, + 181 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 269, + 181 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 269, + 181 + ], + "type": "table", + "html": "
MethodFID ↓FID-N↓NFS↑
FFHQ-512Epigraph [7]9.92†67.3333.95
Mimic3D [10]5.3764.9716.76
GRAM-HD [67]12.2†*--
Ours4.9760.7629.35
w SREG3D [7]4.7063.0217.54
StyleSDF [45]11.19†87.4222.75
MVCGAN [74]13.4†*--
", + "image_path": "12f9cf8c6507398f819580406c57c41225410c9e9bcb97e20bc173a4ccfcd346.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 234, + 264, + 334 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with baselines with and without super resolution (SR) on the FFHQ dataset. " + }, + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "type": "text", + "content": " as reported in the previous works. \\* indicates FID evaluated on " + }, + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 47, + 190, + 287, + 222 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 234, + 264, + 334 + ], + "lines": [ + { + "bbox": [ + 72, + 234, + 264, + 334 + ], + "spans": [ + { + "bbox": [ + 72, + 234, + 264, + 334 + ], + "type": "table", + "html": "
MethodFID ↓NFS↑
AFHQ-512w/o SRGRAM-HD [67]\\(7.67^{\\dagger}\\)-
GMPI [45]\\(7.79^{\\dagger}\\)-
Mimic3D [10]4.2912.67
Ours4.2321.89
SREG3D [7]2.7714.14
StyleSDF [45]7.9133.89
", + "image_path": "b9c1cb17f860b7a908aa2483e308289a2c662546db07ecc2e0ded33483fce7f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 376, + 269, + 454 + ], + "blocks": [ + { + "bbox": [ + 47, + 342, + 287, + 364 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 364 + ], + "type": "text", + "content": "Table 2. Quantitative results on AFHQv2 Cats. † as reported in the previous works." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 376, + 269, + 454 + ], + "lines": [ + { + "bbox": [ + 67, + 376, + 269, + 454 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 269, + 454 + ], + "type": "table", + "html": "
MethodFID ↓FID-N ↓NFS ↑
- Learned Sampler38.2993.8830.95
- Stratification5.6086.025.97
- Robust Sampling5.6760.7824.79
- Beta Regularization5.2764.2528.88
Ours4.9760.7629.35
", + "image_path": "cb6a921d1dcc2da9602d21223db789bab8f45bcce96eaae8cd447bb4e00ebb12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 461, + 212, + 472 + ], + "lines": [ + { + "bbox": [ + 122, + 461, + 212, + 472 + ], + "spans": [ + { + "bbox": [ + 122, + 461, + 212, + 472 + ], + "type": "text", + "content": "Table 3. Ablation study." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 495, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 612 + ], + "type": "text", + "content": "(see Fig. 7) and worse geometry scores (see Tab. 3). Without our sampler or stratification during sampling, the model cannot learn meaningful 3D geometry with limited depth budgets, creating degenerated 3D geometry as can be seen in Fig. 8 and significantly worse FID-N. Without our robust sampling strategy, the sampling becomes more susceptible to slight errors in the sampler due to the high-frequency nature of the PDF (see Fig. 4), resulting in a noticeable drop in FID and occasionally creating floater geometry artifacts (see Fig. 8), while geometry scores remain similar." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 621, + 117, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 117, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 117, + 633 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 723 + ], + "type": "text", + "content": "Limitations and future work. While our method demonstrates significant improvements in 3D geometry generation, it may still exhibit artifacts such as dents in the presence of specularities, and cannot handle transparent objects such as lenses well. Future work may incorporate more advanced material formulations [5] and surface normal regularization [60]. While 3D GANs can learn 3D represent" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 547, + 196 + ], + "type": "image", + "image_path": "75f5c6d21c946211b2e92c0b42968429b09d0a848085b008aae4e5a0fdbfe8e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 204, + 523, + 216 + ], + "lines": [ + { + "bbox": [ + 329, + 204, + 523, + 216 + ], + "spans": [ + { + "bbox": [ + 329, + 204, + 523, + 216 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons for ablation study." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 237, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 342 + ], + "type": "text", + "content": "tations from single-view image collections such as FFHQ and AFHQ with casual camera labels [7], the frontal bias and inaccurate labels can result in geometry artifacts, especially on the side of the faces. Fruitful future directions may include training 3D GANs with large-scale Internet data as well as incorporating a more advanced form of regularization [47] and auto-camera calibration [3] to extend the generations to 360 degrees. Finally, our sampling-based acceleration method may be applied to other NeRFs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 349, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 418 + ], + "type": "text", + "content": "Ethical considerations. While existing methods have demonstrated effective capabilities in detecting unseen GANs [12], our contribution may remove certain characteristics from generated images, potentially making the task of detection more challenging. Viable solutions include the authentication of synthetic media [1, 2, 48]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 530 + ], + "type": "text", + "content": "Conclusion. We proposed a sampler-based method to accelerate 3D GANs to resolve 3D representations at the native resolution of 2D data, creating strictly multi-view-consistent images as well as highly detailed 3D geometry learned from a collection of in-the-wild 2D images. We believe our work opens up new possibilities for generating high-quality 3D models and synthetic data that capture in-the-wild variations and for enabling new applications such as conditional view synthesis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 541, + 409, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 541, + 409, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 409, + 555 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 561, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 723 + ], + "type": "text", + "content": "We thank David Luebke, Tero Karras, Michael Stengel, Amrita Mazumdar, Yash Belhe, and Nithin Raghavan for feedback on drafts and early discussions. Koki Nagano was partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. This work was funded in part by an NSF Graduate Fellowship, ONR Grant N00014-23-1-2526, and the Ronald L. Graham Chair. Manmohan Chandraker acknowledges support of of NSF IIS 2110409. Distribution Statement \"A\" (Approved for Public Release, Distribution Unlimited)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "22772" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 722 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 285, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 285, + 111 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 285, + 111 + ], + "type": "text", + "content": "[1] Coalition for content provenance and authenticity. https://c2pa.org/.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 113, + 286, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 113, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 113, + 286, + 134 + ], + "type": "text", + "content": "[2] Content authenticity initiative. https://contentauthenticity.org/.8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 136, + 287, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 177 + ], + "type": "text", + "content": "[3] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panopehead: Geometry-aware 3d full-head synthesis in 360deg. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 178, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 287, + 221 + ], + "type": "text", + "content": "[4] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 223, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 223, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 223, + 287, + 274 + ], + "type": "text", + "content": "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 277, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 287, + 330 + ], + "type": "text", + "content": "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 404 + ], + "type": "text", + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 470 + ], + "type": "text", + "content": "[8] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative novel view synthesis with 3D-aware diffusion models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 472, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 287, + 513 + ], + "type": "text", + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "type": "text", + "content": "[10] Xingyu Chen, Yu Deng, and Baoyuan Wang. Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. arXiv preprint arXiv:2303.09036, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "text", + "content": "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "text", + "content": "[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 698 + ], + "type": "text", + "content": "[13] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 700, + 287, + 722 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 700, + 287, + 722 + ], + "spans": [ + { + "bbox": [ + 48, + 700, + 287, + 722 + ], + "type": "text", + "content": "[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 722 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "text", + "content": "[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 139, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 139, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 139, + 545, + 191 + ], + "type": "text", + "content": "[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 193, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 193, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 308, + 193, + 545, + 245 + ], + "type": "text", + "content": "[17] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 247, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 247, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 545, + 298 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "text", + "content": "[19] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 344, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 407 + ], + "type": "text", + "content": "[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 409, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 545, + 460 + ], + "type": "text", + "content": "[21] Kunal Gupta, Miloš Hasan, Zexiang Xu, Fujun Luan, Kulyan Sunkavalli, Xin Sun, Manmohan Chandraker, and Sai Bi. Mcnerf: Monte carlo rendering and denoising for real-time nerfs. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 462, + 545, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 545, + 515 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 517, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 517, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 517, + 545, + 548 + ], + "type": "text", + "content": "[23] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 550, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 550, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 550, + 545, + 592 + ], + "type": "text", + "content": "[24] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "text", + "content": "[26] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 680, + 545, + 722 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 722 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 722 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "22773" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 721 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[28] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 287, + 159 + ], + "type": "text", + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 160, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 160, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 287, + 201 + ], + "type": "text", + "content": "[30] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 203, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 203, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 203, + 287, + 236 + ], + "type": "text", + "content": "[31] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 236, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 236, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 236, + 287, + 289 + ], + "type": "text", + "content": "[32] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 290, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 354 + ], + "type": "text", + "content": "[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 397 + ], + "type": "text", + "content": "[34] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 441 + ], + "type": "text", + "content": "[35] David B Lindell, Julien NP Martel, and Gordon Wetzstein. AutoInt: Automatic integration for fast neural volume rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 441, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 473 + ], + "type": "text", + "content": "[36] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 475, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 516 + ], + "type": "text", + "content": "[37] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 518, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 518, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 287, + 570 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "text", + "content": "[39] Don P Mitchell. Consequences of stratified sampling in graphics. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 277-280, 1996. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 615, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 657 + ], + "type": "text", + "content": "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 721 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 721 + ], + "type": "text", + "content": "[41] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 723 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[42] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "text", + "content": "[43] Michael Niemeyer and Andreas Geiger. GIRAFFE: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "type": "text", + "content": "[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 192, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 192, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 192, + 545, + 245 + ], + "type": "text", + "content": "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "type": "text", + "content": "[46] M. Piala and R. Clark. Terminerf: Ray termination prediction for efficient neural rendering. In International Conference on 3D Vision (3DV), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 279, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 320 + ], + "type": "text", + "content": "[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2022. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "type": "text", + "content": "[48] Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo. Avatar fingerprinting for authorized use of synthetic talking-head videos. arXiv preprint arXiv:2305.03713, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "text", + "content": "[49] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3D-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 409, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 452 + ], + "type": "text", + "content": "[50] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 453, + 545, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 505 + ], + "type": "text", + "content": "[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 506, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 506, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 506, + 545, + 558 + ], + "type": "text", + "content": "[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 560, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 560, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 560, + 545, + 602 + ], + "type": "text", + "content": "[53] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "text", + "content": "[54] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. arXiv preprint arXiv:2303.01416, 2023. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In ICCV, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 679, + 545, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 723 + ], + "type": "text", + "content": "[56] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "22774" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 722 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "text", + "content": "[57] Towaki Takikawa, Thomas Müller, Merlin Nimier-David, Alex Evans, Sanja Fidler, Alec Jacobson, and Alexander Keller. Compact neural graphics primitives with learned hash probing. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 149, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 149, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 286, + 213 + ], + "type": "text", + "content": "[58] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum, Frédo Durand, William T. Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 214, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 214, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 214, + 286, + 267 + ], + "type": "text", + "content": "[59] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 269, + 286, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 286, + 321 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 286, + 321 + ], + "type": "text", + "content": "[60] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 323, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 286, + 376 + ], + "type": "text", + "content": "[61] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 286, + 430 + ], + "type": "text", + "content": "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 286, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 286, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 286, + 494 + ], + "type": "text", + "content": "[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 495, + 286, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 286, + 548 + ], + "type": "text", + "content": "[64] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 549, + 286, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 549, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 549, + 286, + 602 + ], + "type": "text", + "content": "[65] Zian Wang, Tiancheng Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. In ACM Transactions on Graphics (SIGGRAPH ASIA). 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 604, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 286, + 645 + ], + "type": "text", + "content": "[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[67] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 286, + 722 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 286, + 722 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 286, + 722 + ], + "type": "text", + "content": "[68] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 460 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "text", + "content": "[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 139, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 139, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 139, + 545, + 180 + ], + "type": "text", + "content": "[70] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 182, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 182, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 182, + 545, + 233 + ], + "type": "text", + "content": "[71] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Ronen Basri, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 235, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 545, + 277 + ], + "type": "text", + "content": "[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 278, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 278, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 278, + 545, + 319 + ], + "type": "text", + "content": "[73] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 373 + ], + "type": "text", + "content": "[74] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 375, + 545, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 417 + ], + "type": "text", + "content": "[75] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In European Conference on Computer Vision (ECCV), 2022. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 418, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 545, + 460 + ], + "type": "text", + "content": "[76] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "22775" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_content_list.json b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..524cc4c99742ead84aa322f9fa75b29890c1343b --- /dev/null +++ b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_content_list.json @@ -0,0 +1,1501 @@ +[ + { + "type": "text", + "text": "When StyleGAN Meets Stable Diffusion: a $\\mathcal{W}_{+}$ Adapter for Personalized Image Generation", + "text_level": 1, + "bbox": [ + 230, + 118, + 736, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoming Li Xinyu Hou Chen Change Loy S-Lab, Nanyang Technological University", + "bbox": [ + 276, + 191, + 691, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "csxml@gmail.com xinyu.hou@ntu.edu.sg ccloy@ntu.edu.sg", + "bbox": [ + 225, + 229, + 745, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b78a124102bb333611629ac7ec136ca867faa6d301c6dce65d5a141378ccdf3.jpg", + "image_caption": [ + "Figure 1. Given a single reference image (thumbnail in the top left), our $\\mathcal{W}_{+}$ adapter not only integrates the identity into the text-to-image generation accurately but also enables modifications of facial attributes along the $\\Delta w$ trajectory derived from StyleGAN. The text prompt is \"a woman wearing a spacesuit in a forest\"." + ], + "image_footnote": [], + "bbox": [ + 119, + 257, + 851, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 656, + 313, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models have remarkably excelled in producing diverse, high-quality, and photo-realistic images. This advancement has spurred a growing interest in incorporating specific identities into generated content. Most current methods employ an inversion approach to embed a target visual concept into the text embedding space using a single reference image. However, the newly synthesized faces either closely resemble the reference image in terms of facial attributes, such as expression, or exhibit a reduced capacity for identity preservation. Text descriptions intended to guide the facial attributes of the synthesized face may fall short, owing to the intricate entanglement of identity information with identity-irrelevant facial attributes derived from the reference image. To address these issues,", + "bbox": [ + 73, + 688, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "we present the novel use of the extended StyleGAN embedding space $\\mathcal{W}_{+}$ , to achieve enhanced identity preservation and disentanglement for diffusion models. By aligning this semantically meaningful human face latent space with text-to-image diffusion models, we succeed in maintaining high fidelity in identity preservation, coupled with the capacity for semantic editing. Additionally, we propose new training objectives to balance the influences of both prompt and identity conditions, ensuring that the identity-irrelevant background remains negligibly affected during facial attribute modifications. Extensive experiments reveal that our method adeptly generates personalized text-to-image outputs that are not only compatible with prompt descriptions but also amenable to common StyleGAN editing directions in diverse settings. Our code and model are available at https://github.com/csxmli2016/w-plus-adapter.", + "bbox": [ + 496, + 657, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2187", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What if we could be the protagonists of our own fantasies? This paper primarily addresses personalized text-to-image (T2I) generation, a field that is attracting growing attention due to users' desire to craft their unique content. Overlapped with customized T2I [8, 17, 28, 36], which integrates various visual concepts like human faces and objects into the generated imagery, our focus is on tailoring image generation to specific identities. This focus is motivated by applications such as storyboarding, where a consistent identity needs to be maintained across all images, despite variations in expressions or ages. Additionally, we believe that human faces possess unique, fine-grained intrinsic features that merit special attention and present a worthwhile area for exploration.", + "bbox": [ + 80, + 114, + 468, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent personalized image generation approaches utilize a small reference set of a target identity and embed it into a specific space. The common choice of embedding space has been the textual embedding space used by large language models (LLMs). While existing methods based on textual embedding [8, 28, 36] have proven capable of maintaining target identity, they are often limited by inherent trade-offs. Specifically, these methods face challenges in simultaneously preserving identity, generating varied facial attributes, and creating identity-irrelevant content that aligns with the text description. We observe that these issues predominantly stem from the entangled nature of the textual embedding space, where a single pseudo word $S^*$ struggles to distinctly isolate identity-related features from the reference image. Efforts to separate such information include approaches like encoding a visual concept into multiple word embeddings [36] or employing a separate branch for extracting identity-irrelevant details [3]. However, these methods tend to overlook the nuanced facial features critical to an individual's identity, resulting in incomplete disentanglement and sub-optimal identity preservation.", + "bbox": [ + 80, + 311, + 468, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we aim to more effectively separate identity-relevant and -agnostic features for better identity preservation, while also ensuring editability. To achieve this, we propose inverting the visual concept of a target identity into StyleGAN's [15] $\\mathcal{W}_{+}$ latent space, as opposed to using the textual embedding space. In particular, we introduce a mapping network to integrate the $\\mathcal{W}_{+}$ space with the diffusion model and a residual cross-attention module to add the $w_{+}$ vector as an additional identity condition. This mapping network, once trained with (image, $w_{+}$ ) pairs, can generalize to unseen individuals during inference without the need for a separate identity-specific model for each person. We further present novel regularized training to ensure that edits have a negligible effect on identity-irrelevant regions and that the overall generation remains aligned with the prompt conditions. It is noteworthy that although CelebBasis [40] explores a similar concept of creating a face-specific latent space through PCA of selected celebrity name embeddings,", + "bbox": [ + 80, + 628, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we argue that relying on such a limited dataset of celebrity names is insufficient for a comprehensive face latent space due to potential under-representation. Our experimental results confirm that our approach exhibits a superior capacity for identity preservation compared to CelebBasis.", + "bbox": [ + 501, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions are summarised as follows: 1) We introduce extended StyleGAN $\\mathcal{W}_{+}$ space as a target inversion latent space to better encode the identity-preserving facial concepts for personalized text-to-image synthesis. This is the first study that considers the fusion between StyleGAN $\\mathcal{W}_{+}$ space and diffusion-based image generation. 2) Embedding target identity in $\\mathcal{W}_{+}$ space enables smooth and controllable semantic editing on facial attributes in our text-to-image model. 3) The effective disentanglement of identity-relevant and -irrelevant information in our model facilitates an identity-preserving generation that is not only diverse but also adaptable to a wide range of prompt instructions.", + "bbox": [ + 503, + 167, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 361, + 638, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGAN Latent Space. GANs have drawn enormous attention due to the well-disentangled latent space [20]. StyleGAN [15] proposes to map input latent code to an intermediate latent space $\\mathcal{W}$ to prevent warping of the training data distribution to fit in a particular probability distribution. The resulting $\\mathcal{W}$ space is a semantically disentangled space that allows fine-grained controls over image synthesis. InterFaceGAN [29] and GANSpace [10] further interpret the latent space and identify several control directions such as age, gender, face angle, smile, etc. The powerful semantic editing ability has motivated research on inverting and editing real images in the StyleGAN space [1, 19, 24, 32]. Abdal et al. [1] utilize a direct optimization framework to embed a given real image to the extended StyleGAN space. Tov et al. [32] deploy an encoder to perform the inversion and keep the concatenated latent codes close to the original StyleGAN space to maintain high perceptual quality and editability. Motivated by the disentanglement and editability of StyleGAN latent space, we aim to introduce it to the T2I diffusion models to achieve more controllable synthesis. We follow Image2StyleGAN to denote the extended $\\mathcal{W}$ space represented by multiple nonidentical $w$ vectors as $\\mathcal{W}_{+}$ .", + "bbox": [ + 503, + 387, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized Image Synthesis. Customization has been extensively studied in the context of T2I to generate images of specified objects or individuals [2, 8, 28]. Gal et al. [8] directly apply learnable text embedding optimization. Ruiz et al. [28] fine-tune the diffusion model for the target concept while preserving its class prior of the concept. Further improvements have been made to invert multiple concepts simultaneously with cross-attention fine-tuning [17], and disentangle background irrelevant information [3]. Encoder-based approaches [4, 9, 13, 30, 36–38] are employed for their efficiency. The target visual concept is encoded to embedding space as additional conditions. SingleInsert [37]", + "bbox": [ + 503, + 719, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2188", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adopts a two-stage scheme to insert concepts from a single image into the foreground region exclusively and enables the editing of the concept by text prompts. However, it is worth noticing that our method achieves more fine-grained editing with a single reference image at inference and does not necessitate identity-specific fine-tuned models.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With special attention on identity inversion utilizing the distinct features of human faces, Yuan et al. [40] define a celebrity space by applying PCA on selected celebrity name embeddings from CLIP and embedding new identities into the space via learnable coefficients. Valevski et al. [33] utilize a pre-trained face recognizer as the encoder and further project the face embedding to CLIP space. FaceChain [21] trains separate face and style LoRAs [12] to synthesize specific faces in specific styles. They follow a data preprocessing pipeline and apply face fusion of the best reference face on the synthesized results. However, we observe that previous personalization approaches inherit a common limitation that the synthesized faces share similar attributes as the reference, indicating that the model cannot disentangle irrelevant information well and overfits the particular reference.", + "bbox": [ + 75, + 181, + 473, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model Adapters. Light-weight adapters have been adopted to avoid the laborious work of fine-tuning large models and add additional controls [23, 39, 42] to the diffusion model. ControlNet [42] proposes to train task-specific adapters while freezing the original diffusion model to adapt to various input conditions. Concurrent work T2I-Adapter [23] constructs a lighter-weight adapter for controls on structure and color. Ye et al. [39] employ a decoupled cross-attention module to consider both text and image prompts in the denoising process. In our work, we consider identity information as an additional condition and align it with the T2I model for identity-preserving synthesis.", + "bbox": [ + 75, + 407, + 470, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 602, + 212, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach is capable of generating images that preserve identity while allowing for semantic edits, requiring just a single reference image for inference. This capability is realized by innovatively aligning StyleGAN's $\\mathcal{W}_{+}$ latent space with the diffusion model. The training of our $\\mathcal{W}_{+}$ adapter is divided into two stages. In Stage I (Sec. 3.2), we establish a mapping from $\\mathcal{W}_{+}$ to SD latent space, using the resulting projection as an additional identity condition to synthesize center-aligned facial images of a specified identity. In Stage II (Sec. 3.3), this personalized generation process is expanded to accommodate more dynamic, \"in-the-wild\" settings, ensuring adaptability to a variety of textual prompts.", + "bbox": [ + 75, + 627, + 470, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 816, + 207, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our implementation, we use Stable Diffusion [26] as the foundational T2I model, which is one of the most commonly used latent space-based text-to-image generation model [6, 34]. During the training phase, the model archi", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tecture includes: 1) a pre-trained encoder $\\mathcal{E}(\\cdot)$ , which transforms an image $I$ into a latent representation with reduced dimensionality, 2) a conditional diffusion model tasked with predicting the latent code of a previous timestep, based on the text condition $c_{txt}$ and the current timestep's latent code, and 3) a pre-trained decoder $\\mathcal{D}(\\cdot)$ that reconstructs the final latent code into the synthesized image. The learning objective of this model is formulated as follows:", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, c _ {t x t}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} (c _ {t x t})\\right) \\| _ {2} ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 215, + 893, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_0$ is $\\mathcal{E}(I)$ , $\\epsilon$ is the ground truth noise added for current timestep $t$ , $z_t$ is the latent code of timestep $t$ . The $\\tau_{txt}$ denotes the pre-trained CLIP text encoder [25] that converts the text prompt $c_{txt}$ to textual embeddings. The $\\epsilon_{\\theta}$ represents the UNet [27] denosing network. In the inference stage, the denoising network is applied iteratively to denoise a random sampled noise $z_T$ to $z_0$ with condition $c_{txt}$ . The final result is then generated through the pre-trained decoder, i.e., $\\mathcal{D}(z_0)$ .", + "bbox": [ + 496, + 238, + 893, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stage I: Aligning $\\mathcal{W}_{+}$ with Stable Diffusion", + "text_level": 1, + "bbox": [ + 498, + 378, + 864, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of Stage I is to align $\\mathcal{W}_{+}$ space with SD to carry identity information in T2I generation while retaining its editing ability. To do so, we train a mapping network to project a $w_{+}$ embedding to SD latent space and inject this extra condition into SD by modifying the cross-attention module. The framework of Stage I is illustrated in Fig. 2. Details are introduced below.", + "bbox": [ + 496, + 406, + 893, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training Pair Construction. We adopt 100K discrete training samples to fit the continuous distribution of $\\mathcal{W}_{+}$ and align it with SD. For each face image $I_{f}$ , we use the pretrained e4e [32] encoder from StyleGAN inversion task to get its corresponding $w_{+} \\in \\mathbb{R}^{18 \\times 512}$ vector. The pairs of $\\{I_{f}, w_{+}\\}$ constitute our training data in this stage. Specifically, to generalize on real-world face images and improve $w_{+}$ diversity, two types of training pairs are built, i.e., synthetic face images from StyleGAN2 [16] and real-world face images from FFHQ [15]. Note that in Stage I, we only consider the aligned face images in order to exclude any extraneous influences.", + "bbox": [ + 496, + 515, + 895, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mapping Network. Since the original $\\mathcal{W}_{+}$ space is designated for StyleGAN generation, we project the vector $w_{+} \\in \\mathbb{R}^{18 \\times 512}$ to four tokens of dimension 768 ( $\\mathbb{R}^{4 \\times 768}$ ) to align with the input dimension of SD condition. The mapping network consists of four groups of linear layers and is denoted as $\\mathcal{F}_w(\\cdot)$ . The latent space after mapping inherits the editability and disentanglement properties of $\\mathcal{W}_{+}$ space and is compatible with the SD generation process.", + "bbox": [ + 496, + 700, + 893, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Residual Cross-Attention. To incorporate the identity condition from the projected $w_{+}$ embedding into the pre-trained SD model, we introduce a residual cross-attention module. In the standard SD framework, the output of cross-attention is determined using query features $f_{z}$ from the hidden state", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2189", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/854400d374ea713f011c018896c36e324d5d9658c49165eab8ecf6d4c31f7c8b.jpg", + "image_caption": [ + "(a) Stage I: Aligning $w_{+}$ with Stable Diffusion" + ], + "image_footnote": [], + "bbox": [ + 86, + 71, + 390, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50f804bf3afe9299f74c7c80640d75167ad106d37009de8fcc2a88873b170f74.jpg", + "image_caption": [ + "(b) Stage II: Generation in the Wild", + "Figure 2. Overview of $\\mathcal{W}_{+}$ adapter training stages. Left: Stage I for aligning $\\mathcal{W}_{+}$ space with Stable Diffusion. Middle: Stage II for generating in-the-wild images with $w_{+}$ embeddings. Right: details of residual cross-attention module. Open lock indicates trainable parts." + ], + "image_footnote": [], + "bbox": [ + 398, + 71, + 725, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eea6e23c46f92d07b8307cbd5ec6e18182f38192579f3e0a15080db4b67e1bec.jpg", + "image_caption": [ + "(c) Residual Cross-Attention" + ], + "image_footnote": [], + "bbox": [ + 735, + 71, + 883, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the text condition $\\tau_{txt}(c_{txt})$ . The process is defined as", + "bbox": [ + 75, + 308, + 452, + 324 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {z} ^ {\\prime} = \\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 328, + 468, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q} = f_z\\mathbf{W}_q$ , $\\mathbf{K} = \\tau_{txt}(c_{txt})\\mathbf{W}_k$ , $\\mathbf{V} = \\tau_{txt}(c_{txt})\\mathbf{W}_v$ are the query, key, value matrices of the attention module, respectively. $\\mathbf{W}_q$ , $\\mathbf{W}_k$ , and $\\mathbf{W}_v$ denote the corresponding projection matrices. The cross-attention module is where the latent noise interacts with text condition embeddings inside the denoising process. Following previous methods [36, 39], we add an additional condition by a separate cross-attention module. Our residual cross-attention is defined as:", + "bbox": [ + 75, + 363, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {z} ^ {\\prime \\prime} = f _ {z} ^ {\\prime} + \\lambda \\cdot \\operatorname {A t t e n t i o n} \\left(\\mathbf {Q} ^ {\\dagger}, \\mathbf {K} ^ {\\dagger}, \\mathbf {V} ^ {\\dagger}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 488, + 468, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}^{\\dagger} = f_{z}^{\\prime}\\mathbf{W}_{q}^{\\dagger}$ , $\\mathbf{K}^{\\dagger} = f_{w}\\mathbf{W}_{k}^{\\dagger}$ , and $\\mathbf{V}^{\\dagger} = f_{w}\\mathbf{W}_{v}^{\\dagger}$ . The $\\mathbf{W}_{q}^{\\dagger},\\mathbf{W}_{k}^{\\dagger}$ , and $\\mathbf{W}_{v}^{\\dagger}$ are the projection matrices for query, key, and value, respectively. The $f_{w} = \\mathcal{F}_{w}(w_{+})$ is mapped from $w_{+}$ by the mapping network. We use $\\lambda$ as a scale parameter to balance the influence of text and identity conditions on the generation. We incorporate the decoupled cross-attention module in a residual fashion instead of the parallel approach as in previous works [36, 39] to avoid performance degradation on the original text condition. When $\\lambda$ is set to 0, our $w_{+}$ vector has no impact on the pre-trained SD model. We set $\\lambda$ to 1 during training. The $\\mathbf{Q}^{\\dagger}$ , $\\mathbf{K}^{\\dagger}$ , and $\\mathbf{V}^{\\dagger}$ are initialized from their corresponding Q, K, and V, respectively. The residual cross-attention module is performed on all the cross-attention layers of SD.", + "bbox": [ + 75, + 508, + 472, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning Objectives. The trainable parts in this stage include the mapping network $\\mathcal{F}_w$ and $\\{\\mathbf{Q}^\\dagger, \\mathbf{K}^\\dagger, \\mathbf{V}^\\dagger\\}$ matrices in all residual cross-attention modules. The optimization target is formulated as:", + "bbox": [ + 75, + 729, + 470, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w _ {+}} = \\mathbb {E} _ {z _ {0}, \\epsilon , t, c _ {t x t}, w _ {+}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\mathcal {F} _ {w} (w _ {+})) \\| _ {2} ^ {2} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 795, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eqn. (4) is similar to Eqn. (1) except for the additional $\\mathcal{F}_w(w_+)$ serving as identity condition and cross-attention structure modified inside $\\epsilon_{\\theta}$ . The text prompt $c_{txt}$ is randomly selected from several neutral templates describing a human face, e.g., \"a photo of a face\", or \"a face\" (see suppl.).", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through mapping the $\\mathcal{W}_{+}$ distribution into an embedding space compatible with SD, our approach effectively uses the $w_{+}$ vector to condition the personalized generation of aligned facial images. Furthermore, attribute directions $\\Delta w$ derived from the original $\\mathcal{W}_{+}$ space, encompassing features like smile, age, eye-opening, and others, remain applicable to our projected $w_{+}$ vector. This allows for personalized, fine-grained face editing, as demonstrated in Fig. 1.", + "bbox": [ + 496, + 308, + 893, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Stage II: Generation in the Wild", + "text_level": 1, + "bbox": [ + 498, + 440, + 784, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to refine the performance of our $\\mathcal{W}_{+}$ -adapted SD model from Stage I for scenarios beyond controlled environments, we continue on a second stage of training. This phase is dedicated to further fine-tuning the weights of $\\mathbf{Q}^{\\dagger},\\mathbf{K}^{\\dagger},\\mathbf{V}^{\\dagger}$ across all residual cross-attention modules. To ensure that the projected $\\mathcal{W}_{+}$ space exclusively encapsulates identity-related facial attributes, while remaining uninfluenced by irrelevant distractions from the background, we keep the mapping network fixed during this stage. The architecture and workflow of Stage II are depicted in the middle of Fig. 2.", + "bbox": [ + 496, + 464, + 893, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Data Construction. For Stage II, in-the-wild images are used for training. For each image $I$ , its corresponding text caption, $c_{txt}$ , is extracted using an off-the-shelf captioning tool [18]. The aligned face image $I_f$ is cropped from $I$ , and $w_+$ is obtained from $I_f$ by the e4e encoder (same as Stage I). A face region mask, $M$ , is also obtained from the image $I$ , with 0 denoting the face region and 1 representing the non-face region.", + "bbox": [ + 496, + 616, + 893, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning Objectives. To keep high identity fidelity while encouraging high diversity of identity-irrelevant context, three losses are employed jointly in this stage.", + "bbox": [ + 496, + 737, + 893, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the reconstruction loss $\\mathcal{L}_{rec}$ is used to guide the denoising process:", + "bbox": [ + 496, + 784, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {z _ {0}, \\epsilon , t, c _ {t x t}, w _ {+}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\mathcal {F} _ {w} (w _ {+})) \\| _ {2} ^ {2} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 816, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that Eqn. (5) is different from Eqn. (4) for that: 1) $z_0$ here is encoded from the in-the-wild image $I$ rather than the face image $I_f$ , so the reconstruction goal differs, and 2) $c_{txt}$ is the caption describing the in-the-wild image rather than a", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2190", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "simple neutral prompt, therefore containing more information about the image context to facilitate reconstruction of $I$ with $w_{+}$ from $I_{f}$ .", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, when semantic edits $\\Delta w$ are applied in the $\\mathcal{W}_{+}$ space, the objective is to exclusively modify the facial attributes, leaving the remaining regions unchanged. In order to accomplish this, we propose a $w_{+}$ disentanglement loss to limit the editability of $w_{+}$ outside the face region:", + "bbox": [ + 75, + 136, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s e n}} = \\| M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right), \\mathcal {F} _ {w} \\left(w _ {+}\\right)\\right) - \\tag {6} \\\\ M \\cdot \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\Psi (\\mathcal {F} _ {w} (w _ {+}))) \\| \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 217, + 468, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Psi$ is the augmentation operation. During training, three augmentation strategies are adopted: 1) shuffle along the batch dimension of $\\mathcal{F}_w(w_+)$ , 2) add random perturbations of Gaussian noise on $\\mathcal{F}_w(w_+)$ , and 3) combine both of them. By applying such constraints, augmented (or edited) $w_+$ vectors are forced to have similar non-face regions, thereby disentangling the effect of $\\Delta w$ on the background.", + "bbox": [ + 75, + 253, + 470, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, since the identity condition $\\mathcal{F}_w(w_+)$ should only influence the face region, which constitutes a relatively small proportion of the final output, we constrain its impact to be limited by a regularization loss:", + "bbox": [ + 75, + 361, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\left\\| M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right), \\mathcal {F} _ {w} \\left(w _ {+}\\right)\\right) - M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right) \\right. \\right\\| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 426, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, a more balanced effect of text and identity condition on the synthesized result is achieved, further mitigating the risk of unwanted noisy information introduced from $w_{+}$ and preserving the compatibility with text prompts.", + "bbox": [ + 75, + 446, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall learning objective in Stage II is defined as:", + "bbox": [ + 96, + 508, + 457, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\gamma_ {1} \\cdot \\mathcal {L} _ {\\text {d i s e n}} + \\gamma_ {2} \\cdot \\mathcal {L} _ {\\text {r e g}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 529, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_{1}$ and $\\gamma_{2}$ denote the trade-off parameters.", + "bbox": [ + 75, + 549, + 398, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 573, + 207, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data. In Stage I, we use the pre-trained e4e encoder [32] to obtain the $w_{+}$ vector for each face image $I_{f}$ from FFHQ [15] and StyleGAN2 [16]. FFHQ dataset contains 70,000 images, among which 600 are split for validation while the rest are for training. We also synthesize 70,000 images using StyleGAN2. In Stage II, we use FFHQ in-the-wild images (excluding those used for validation in Stage I) and SHHQ [7] to optimize the $\\mathcal{W}_{+}$ adapter. SHHQ is a human dataset that contains 40,000 high-quality full-body images. BLIP2 [18] is introduced to generate the caption for each in-the-wild image. Face region mask $M$ is obtained based on the FFHQ alignment operation and is eroded and blurred with kernel sizes of 32 and 7, respectively.", + "bbox": [ + 75, + 598, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. All training is conducted on a server with 8 Tesla V100 GPUs. The batch size is set to 16. We employ the AdamW optimizer [22] with a learning rate of $1e - 4$ and weight decay of 0.01. In Stage II, we adopt color jittering [44], random rotation, and sampling for inthe-wild images to increase diversity. During training, the cropped images with incomplete or small faces are discarded.", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a2b880e02dc16d50603ef6ec7ae84b3c49c0cfb63cb70dd10f0df13841c3cd06.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with previous methods. The best and second best results are highlighted by bold and underline." + ], + "table_footnote": [], + "table_body": "
VariantsCLIP Score↑ID↓Detection↑
Textual Inversion [8].194.575.891
Dreambooth [28].177.562.800
Custom Diffusion [17].216.498.850
FastComposer [38].265.419.950
IP-Adapter-Face [39].241.407.958
CelebBasis [40].253.448.916
Ours.267.418.950
", + "bbox": [ + 501, + 121, + 890, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It takes around 24 hours to align the $\\mathcal{W}_{+}$ space with SD in Stage I, and nearly 96 hours to train the $\\mathcal{W}_{+}$ adapter in Stage II. In the inference, we adopt DDIM sampler [31] with 50 steps. To enable classifier-free guidance [11], we use the default settings and set the guidance scale to 7.5. We follow IP-Adapter [39] to randomly drop the text $c_{txt}$ and $w_{+}$ vector with a probability of 0.05. $\\gamma_{1}$ and $\\gamma_{2}$ in Eqn. (8) are set to 1.5 and 1.0, respectively. The whole framework is implemented on Diffusers [35]. Our experiments are based on SD v1.5 and can be extended to other SD models (see suppl.).", + "bbox": [ + 496, + 250, + 890, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 498, + 412, + 732, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. The methods we evaluate can be categorized into two distinct groups: general object customization and specific face personalization. For general object customization, we use Dreambooth [28], Textual Inversion [8], and Custom Diffusion [17], implementing these using Diffusers [35]. For specific face personalization, we select closely related works such as FastComposer [38], IP-Adapter-Face [39], and CelebBasis [40]. In our evaluation, we adhere to the settings established by CelebBasis [40] and FastComposer [38]. The metrics used for assessment include the CLIP Score [25], which is calculated based on the average image-text similarity using CLIP-L/14, the Face Detection Score determined using MTCNN [41], and the Identity Distance (ID) measured using ArcFace [5] on the detected facial regions.", + "bbox": [ + 496, + 436, + 893, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the extensive fine-tuning requirements for each identity in methods like Dreambooth, Textual Inversion, Custom Diffusion, and CelebBasis – which can be notably time-consuming (for instance, Textual Inversion demands about an hour on a single V100 GPU server $^{1}$ ) – we encounter constraints in evaluating a substantial number of test instances. To ensure a thorough and equitable quantitative analysis, we carefully choose 120 identities from the CelebA-HQ dataset [14], using one reference image per identity for consistent comparisons. In addition, we randomly select 20 text prompts that describe various aspects such as clothing, styles, and backgrounds, to provide detailed characterizations of each individual.", + "bbox": [ + 496, + 648, + 893, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows that optimization-based methods operating in the text embedding space, such as Textual Inversion,", + "bbox": [ + 498, + 845, + 893, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/docs/diffusers/training/text_inversion", + "bbox": [ + 517, + 886, + 841, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2191", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7679535e9a5d032a60c9146b1965b86f11f7f009b4f1c0d11fffda2b7922118e.jpg", + "image_caption": [ + "a child in a yellow blazer" + ], + "image_footnote": [], + "bbox": [ + 86, + 77, + 174, + 487 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29ba7fdaee49308c8b6c6bd3bafecd4f0a7e3c039fc6a27b25d0be99307a6070.jpg", + "image_caption": [ + "Figure 3. Visual comparisons with baselines on different scenarios of T2I generation. Best view by zooming in." + ], + "image_footnote": [], + "bbox": [ + 191, + 78, + 277, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8778bdea6bf672be55962659b90d9396df12b9491a9b8e1a38ae5931b6aece8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 78, + 362, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/911d6bc129b6ef68e4f496416ab30f83f82a2a5b11a8e9121bfc3c00028040ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 78, + 450, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/661f2ae8ff7f4fde176ad05d806f4e5cae81f541e95a9496c789b155c7255b8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 78, + 535, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c56847d50e6437e44e9daf523fd10f140112d0243288cfa666cb2cfc98739fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 78, + 625, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c397e05de358fab5479fa396cec844417dbf73e569bfa4ec284d70da952453e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 78, + 709, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/857a999262335a24dffc9fe831f3727d8a7e685b61d845c17a32331e078873f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 78, + 795, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/951ce631c3892e26c6e13f105992b946b4e72dd6469c4f32174dce528a555824.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 78, + 880, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dreambooth, and Custom Diffusion, fall short in accurately aligning with text prompts and preserving identity features. On the contrary, our method shows comparable performance to FastComposer and IP-Adapter-Face in terms of both CLIP Score and ID metrics. However, it is important to note that IP-Adapter has been fine-tuned on the large-scale LAION2B dataset, whereas our model is trained on the relatively smaller FFHQ-wild and SHHQ datasets, comprising around 110,000 images. Both FastComposer and IP-Adapter directly process the reference image as input, while our approach first maps it to the $\\mathcal{W}_{+}$ space before embedding it into the SD model. This additional mapping step could potentially introduce minor discrepancies when integrating the reference image into the SD model, slightly affecting the ID results. Nevertheless, this slight trade-off in ID preservation is balanced by our method's ability to flexibly edit facial attributes while ensuring the background remains consistent. This unique balance of attribute editability and background consistency is a novel contribution that sets our method apart from previous approaches.", + "bbox": [ + 75, + 523, + 472, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 832, + 299, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 depicts visual comparisons with baselines across various scenarios. Since Textual Inversion, Dreambooth, and Custom Diffusion are not specifically designed for facial im", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ages, we select results with the best facial quality from their generated set of 40 images. For the remaining methods, we choose from a set of 10. It is observed that Textual Inversion, Dreambooth, and Custom Diffusion encounter difficulties in capturing and maintaining identity details with a single reference, leading to sub-optimal performance in this task. CelebBasis and FastComposer face challenges in striking a balance between text compatibility and identity preservation. Though IP-Adapter shows improved identity retention, it tends to ignore text conditions in certain instances (see the 1st, 3~5th rows).", + "bbox": [ + 496, + 523, + 893, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Leveraging our $\\mathcal{W}_{+}$ adapter, our approach successfully generates images that are not only compatible with text descriptions but also more effectively retain the target identity. Additionally, our method allows for the editing of facial attributes along the $\\Delta w$ direction, causing only minor alterations in the non-facial regions (illustrated in the last column). Furthermore, our approach can be seamlessly adapted to other pre-trained SD models without the need for additional fine-tuning, while retaining its editing capabilities. This versatility is exemplified in the last row of Fig. 3, which showcases our method's effectiveness with the dreamlike-anime model $^2$ .", + "bbox": [ + 496, + 691, + 895, + 872 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://huggingface.co/dreamlike-art/dreamlike-anime-1.0", + "bbox": [ + 516, + 886, + 831, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2192", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/409210800335fcec3ad7daee473a63387a14b0a4459ead664598ac675d35a706.jpg", + "image_caption": [ + "Figure 4. Inversion and editing comparisons between e4e and Ours in Stage I. They have the same $w_{+}$ and attributes editing $\\Delta w$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 460, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/03d7bfed90bd91cb02bd40bccbf00216298d09c871ed2db0e5d3077b988066fc.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons of face inversion and editing. Editing (ID↓) is the ID metric on results of $w_{+}$ and $w_{+} - 3 \\cdot \\Delta w$ ." + ], + "table_footnote": [], + "table_body": "
MethodsInversionEditing (ID↓)
ID↓FID↓LPIPS↓AgeSmile
e4e [32].43132.26.205.409.312
Ours (Stage I).43531.47.263.393.275
", + "bbox": [ + 81, + 386, + 459, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 455, + 230, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analyses of Aligning $\\mathcal{W}_{+}$ in Stage I. We examine if the StyleGAN's $w_{+}$ embedding is well aligned to the latent space of SD during Stage I training. Fig. 4 offers a qualitative comparison of inversion and attribute editing outcomes between e4e [32] and our approach, applied to a real-world image. Both methods use the identical $w_{+}$ embedding generated by the e4e encoder. As depicted in Fig. 4, our method yields inversion results on par with e4e, as seen in the left column, using the same $w_{+}$ vector. Furthermore, the alignment process we implement in this phase preserves the editability inherent to the $\\mathcal{W}_{+}$ space. By introducing attribute editing directions $\\Delta w$ (such as smile and age) from InterFaceGAN [29], our method can semantically adjust these attributes while maintaining the individual's identity. This is illustrated in the right columns, where the edited embedding, denoted as $w_{+} + \\alpha \\cdot \\Delta w$ , reflects these changes. The parameter $\\alpha$ is used to control the extent of attribute modification.", + "bbox": [ + 75, + 479, + 472, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantitatively evaluate the alignment performance, we randomly select 1,000 images from CelebA-HQ [14] and measure their inversion results with metrics of ID, FID, and LPIPS $[43]^3$ . As indicated in the Inversion column of Table 2, our method exhibits comparable performance to e4e in the image inversion task. For assessing attribute editing capabilities, we focus on two key attributes, namely smile, and age, applying a constant scale $\\alpha$ (i.e., $w_{+} - 3 \\cdot \\Delta w$ ). The Editing column in Table 2 demonstrates that our approach surpasses", + "bbox": [ + 75, + 738, + 470, + 875 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/39773dcbc3bd3fcefc981657c96978f4f1e76667f710b33b968ccb6d548c9f3b.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons of different variants." + ], + "table_footnote": [], + "table_body": "
VariantsCLIP Score↑ID↓Detection↑
Ours (PCA).243.526.936
Ours (w/o Lreg).136.461.951
Ours (L0.5reg).192.497.946
Ours (Full).267.516.943
", + "bbox": [ + 504, + 101, + 883, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f4c3dff5baf3afd4df359744407addc54b2acf4dc6f07fa79f2e897981a5931.jpg", + "image_caption": [ + "Figure 5. Visual comparisons of smile attribute editing. The prompt is \"a woman wearing a t-shirt on the beach\"." + ], + "image_footnote": [], + "bbox": [ + 504, + 183, + 887, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "e4e in maintaining identity during the editing process. These findings confirm that our method has successfully aligned the $\\mathcal{W}_{+}$ space with the SD model, preserving both the inversion accuracy and the robust editability of attributes.", + "bbox": [ + 498, + 311, + 892, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analyses of Variants. We examine different variants to assess each component of our $\\mathcal{W}_{+}$ adapter: 1) Ours (PCA), which substitutes our residual cross-attention with parallel cross-attention as in [39], and 2) Ours ( $w/o\\mathcal{L}_{reg}$ ), which omits the $\\mathcal{L}_{reg}$ objective. The outcomes of these variants are presented in Table 3. It is observed that the residual cross-attention module outperforms the parallel cross-attention design. We conjecture that in tasks focused on integrating local facial imagery into broader, in-the-wild scenarios, directly injecting facial embeddings into the original hidden state $f_{z}$ might adversely impact non-facial regions. Conversely, our residual cross-attention determines fusion weights by calculating attention activations between the text-embedded hidden state $f_{z}^{\\prime}$ and our facial embedding. This results in a more precise and effective fusion within the facial region.", + "bbox": [ + 496, + 371, + 893, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, the results suggest that the regularization loss $\\mathcal{L}_{reg}$ is crucial in maintaining semantic consistency between text prompts and images. In the absence of $\\mathcal{L}_{reg}$ , the variant Ours $(w / o\\mathcal{L}_{reg})$ tends to overemphasize facial regions while neglecting the accompanying text descriptions, leading to a reduced CLIP score. Incorporating $\\mathcal{L}_{reg}$ with a weight of $\\gamma_{2} = 0.5$ , as in Ours $(\\mathcal{L}_{reg}^{0.5})$ , enhances the CLIP score but slightly compromises the ID metric. In comparison, Ours (Full) achieves a satisfactory trade-off in balancing the performance of ID and CLIP score.", + "bbox": [ + 496, + 598, + 893, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The impact of our disentanglement loss $\\mathcal{L}_{disen}$ (Eqn. (6)) is visually demonstrated in Fig. 5. This example shows that omitting $\\mathcal{L}_{disen}$ leads to unintended alterations in non-facial regions when the smile editing direction $\\Delta w$ is applied to $w_{+}$ . This indicates the importance of $\\mathcal{L}_{disen}$ in effectively separating identity-relevant and irrelevant information. By combining our residual cross-attention, regularization loss, and disentanglement loss, Ours (Full) preserves identity and ensures compatibility with text prompts, even amidst changes to the face embedding.", + "bbox": [ + 496, + 750, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3https://github.com/chaofengc/IQA-PyTorch", + "bbox": [ + 94, + 886, + 331, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2193", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87f2a5864478f6deb56b0189abeea0a4ace60c140c31bd97b53529a0583e3027.jpg", + "image_caption": [ + "w+", + "T2I Generation with $w_{+}$ Interpolation", + "w" + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 872, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66955cd1c0cc40e65c350fde7b75ae383db0c88cbd57ceb57877c68dd562a170.jpg", + "image_caption": [ + "Figure 6. Visual results of $w_{+}$ embeddings interpolation during inference. The prompts are \"one person wearing suit and tie in a garden\" and \"one person wearing a blue shirt by a secluded waterfall\", respectively.", + "Figure 7. Visual results of using different $\\lambda$ during inference. The prompt is \"a man wearing a rainbow shirt in a garden\"." + ], + "image_footnote": [], + "bbox": [ + 91, + 297, + 455, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyses of $\\lambda$ in Cross-attention. The parameter $\\lambda$ determines the extent to which the $w_{+}$ vector influences the hidden state $f_{z}^{\\prime}$ . Fig. 7 shows the effects of varying $\\lambda$ during inference. At $\\lambda = 0$ , the outcome aligns with that of the original SD model, unaffected by any editing direction. However, as $\\lambda$ increases, the generated identity more closely resembles the reference image. Interestingly, at relatively low values of $\\lambda$ (for example, $\\lambda = 0.3$ ), even though the identity does not closely match the reference, our model effectively edits attributes in the specified direction. This observation suggests that our method proficiently leverages the $\\mathcal{W}_{+}$ space from StyleGAN within the SD framework.", + "bbox": [ + 75, + 626, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyses of $w_{+}$ Interpolation. We select two facial images and acquire their respective $w_{+}^{1}$ and $w_{+}^{2}$ embeddings from e4e encoder. The interpolated $w_{+}$ embedding is then obtained through $(1 - \\kappa) \\cdot w_{+}^{1} + \\kappa \\cdot w_{+}^{2}$ , where $\\kappa \\in [0,1]$ . As shown in Fig. 6, this T2I generation process results in a smooth transition in the facial regions of the generated", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images, while maintaining a similar and consistent layout throughout the interpolation. This result confirms that the $w_{+}$ embedding in our method not only preserves the editability characteristic of the original StyleGAN but also effectively distinguishes between facial and background regions.", + "bbox": [ + 496, + 297, + 893, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitation", + "text_level": 1, + "bbox": [ + 498, + 383, + 619, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work aims at integrating StyleGAN's editable $\\mathcal{W}_{+}$ space into the SD model. We notice a challenge in this integration: the process of converting real images to $w_{+}$ vectors in StyleGAN often leads to a loss of detail, impacting the preservation of identity features. Despite employing a substantial number of training pairs $\\{I_f, w_{+}\\}$ to establish the mapping, we still observe limitations in maintaining identity fidelity and challenges in editing certain attributes (e.g., pose and glasses). The current framework is designed to generate and edit images with only a single face. Our future work aims to explore the potential of applying localized injections of distinct $w_{+}$ to address multiple human subjects.", + "bbox": [ + 496, + 407, + 893, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 602, + 619, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented the first attempt to embed the $\\mathcal{W}_{+}$ space of StyleGAN into the SD model. We showed that both the mapping network and the residual cross-attention module play crucial roles in facilitating the injection of $w_{+}$ embedding into the SD model, balancing between text prompt influence and identity conditions. Our experiments demonstrate that the $\\mathcal{W}_{+}$ space, as used in our approach, not only enables personalized text-to-image generation but also allows for precise editing of facial attributes. We envision this capability being highly beneficial in various practical applications, such as portrait customization with seamless attribute modifications. Furthermore, our methodology holds the potential for application across other object domains that possess distinct prior spaces.", + "bbox": [ + 496, + 627, + 893, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2194", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: How to embed images into the StyleGAN latent space? In ICCV, 2019. 2", + "[2] Yufei Cai, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Hu Han, and Wangmeng Zuo. Decoupled textual embeddings for customized image generation. In AAAI, 2024. 2", + "[3] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. DisenBooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2", + "[4] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, and Min Zheng. PhotoVerse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2", + "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5", + "[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 3", + "[7] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. StyleGAN-Human: A data-centric odyssey of human generation. In ECCV, 2022. 5", + "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 2, 5", + "[9] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2", + "[10] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In NeurIPS, 2020. 2", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS, 2022. 5", + "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 3", + "[13] Xuhui Jia, Yang Zhao, Kelvin C. K. Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2", + "[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7", + "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 3, 5", + "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 3, 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 5", + "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4, 5", + "[19] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image superresolution. In CVPR, 2023. 2", + "[20] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. A survey on leveraging pre-trained generative adversarial networks for image editing and restoration. arXiv preprint arXiv:2207.10309, 2022. 2", + "[21] Yang Liu, Cheng Yu, Lei Shang, Ziheng Wu, Xingjun Wang, Yuze Zhao, Lin Zhu, Chen Cheng, Weitao Chen, Chao Xu, Haoyu Xie, Yuan Yao, Wenmeng Zhou, Yingda Chen, Xuansong Xie, and Baigui Sun. FaceChain: A playground for identity-preserving portrait generation. arXiv preprint arXiv:2308.14256, 2023. 3", + "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5", + "[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3", + "[24] Hamza Pehlivan, Yusuf Dalva, and Aysegul Dundar. StyleRes: Transforming the residuals for real image editing with StyleGAN. In CVPR, 2022. 2", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3", + "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Finetuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 5", + "[29] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2, 7", + "[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. InstantBooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2", + "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 5", + "[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. In ACM TOG, 2021. 2, 3, 5, 7" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2195", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3", + "[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 3", + "[35] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.5", + "[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 4", + "[37] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. SingleInsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 2", + "[38] Guangxuan Xiao, Tianwei Yin, William T. Freeman, Frédo Durand, and Song Han. FastComposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 5", + "[39] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 4, 5, 7", + "[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. In NeurIPS, 2023. 2, 3, 5", + "[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. In IEEE signal processing letters, 2016. 5", + "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3", + "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7", + "[44] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In ECCV, 2020. 5" + ], + "bbox": [ + 78, + 90, + 470, + 714 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2196", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_model.json b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6767889f6348507dccc8c5895bcb5a0001bd8cb7 --- /dev/null +++ b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.119, + 0.737, + 0.163 + ], + "angle": 0, + "content": "When StyleGAN Meets Stable Diffusion: a \\(\\mathcal{W}_{+}\\)Adapter for Personalized Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.192, + 0.692, + 0.227 + ], + "angle": 0, + "content": "Xiaoming Li Xinyu Hou Chen Change Loy S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.23, + 0.746, + 0.245 + ], + "angle": 0, + "content": "csxml@gmail.com xinyu.hou@ntu.edu.sg ccloy@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.258, + 0.852, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.596, + 0.895, + 0.639 + ], + "angle": 0, + "content": "Figure 1. Given a single reference image (thumbnail in the top left), our \\(\\mathcal{W}_{+}\\)adapter not only integrates the identity into the text-to-image generation accurately but also enables modifications of facial attributes along the \\(\\Delta w\\) trajectory derived from StyleGAN. The text prompt is \"a woman wearing a spacesuit in a forest\"." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.657, + 0.314, + 0.673 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.689, + 0.475, + 0.903 + ], + "angle": 0, + "content": "Text-to-image diffusion models have remarkably excelled in producing diverse, high-quality, and photo-realistic images. This advancement has spurred a growing interest in incorporating specific identities into generated content. Most current methods employ an inversion approach to embed a target visual concept into the text embedding space using a single reference image. However, the newly synthesized faces either closely resemble the reference image in terms of facial attributes, such as expression, or exhibit a reduced capacity for identity preservation. Text descriptions intended to guide the facial attributes of the synthesized face may fall short, owing to the intricate entanglement of identity information with identity-irrelevant facial attributes derived from the reference image. To address these issues," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.896, + 0.901 + ], + "angle": 0, + "content": "we present the novel use of the extended StyleGAN embedding space \\(\\mathcal{W}_{+}\\), to achieve enhanced identity preservation and disentanglement for diffusion models. By aligning this semantically meaningful human face latent space with text-to-image diffusion models, we succeed in maintaining high fidelity in identity preservation, coupled with the capacity for semantic editing. Additionally, we propose new training objectives to balance the influences of both prompt and identity conditions, ensuring that the identity-irrelevant background remains negligibly affected during facial attribute modifications. Extensive experiments reveal that our method adeptly generates personalized text-to-image outputs that are not only compatible with prompt descriptions but also amenable to common StyleGAN editing directions in diverse settings. Our code and model are available at https://github.com/csxmli2016/w-plus-adapter." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2187" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.116, + 0.47, + 0.312 + ], + "angle": 0, + "content": "What if we could be the protagonists of our own fantasies? This paper primarily addresses personalized text-to-image (T2I) generation, a field that is attracting growing attention due to users' desire to craft their unique content. Overlapped with customized T2I [8, 17, 28, 36], which integrates various visual concepts like human faces and objects into the generated imagery, our focus is on tailoring image generation to specific identities. This focus is motivated by applications such as storyboarding, where a consistent identity needs to be maintained across all images, despite variations in expressions or ages. Additionally, we believe that human faces possess unique, fine-grained intrinsic features that merit special attention and present a worthwhile area for exploration." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.313, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Recent personalized image generation approaches utilize a small reference set of a target identity and embed it into a specific space. The common choice of embedding space has been the textual embedding space used by large language models (LLMs). While existing methods based on textual embedding [8, 28, 36] have proven capable of maintaining target identity, they are often limited by inherent trade-offs. Specifically, these methods face challenges in simultaneously preserving identity, generating varied facial attributes, and creating identity-irrelevant content that aligns with the text description. We observe that these issues predominantly stem from the entangled nature of the textual embedding space, where a single pseudo word \\( S^* \\) struggles to distinctly isolate identity-related features from the reference image. Efforts to separate such information include approaches like encoding a visual concept into multiple word embeddings [36] or employing a separate branch for extracting identity-irrelevant details [3]. However, these methods tend to overlook the nuanced facial features critical to an individual's identity, resulting in incomplete disentanglement and sub-optimal identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.63, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this study, we aim to more effectively separate identity-relevant and -agnostic features for better identity preservation, while also ensuring editability. To achieve this, we propose inverting the visual concept of a target identity into StyleGAN's [15] \\(\\mathcal{W}_{+}\\) latent space, as opposed to using the textual embedding space. In particular, we introduce a mapping network to integrate the \\(\\mathcal{W}_{+}\\) space with the diffusion model and a residual cross-attention module to add the \\(w_{+}\\) vector as an additional identity condition. This mapping network, once trained with (image, \\(w_{+}\\)) pairs, can generalize to unseen individuals during inference without the need for a separate identity-specific model for each person. We further present novel regularized training to ensure that edits have a negligible effect on identity-irrelevant regions and that the overall generation remains aligned with the prompt conditions. It is noteworthy that although CelebBasis [40] explores a similar concept of creating a face-specific latent space through PCA of selected celebrity name embeddings," + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "we argue that relying on such a limited dataset of celebrity names is insufficient for a comprehensive face latent space due to potential under-representation. Our experimental results confirm that our approach exhibits a superior capacity for identity preservation compared to CelebBasis." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.168, + 0.892, + 0.348 + ], + "angle": 0, + "content": "The contributions are summarised as follows: 1) We introduce extended StyleGAN \\(\\mathcal{W}_{+}\\) space as a target inversion latent space to better encode the identity-preserving facial concepts for personalized text-to-image synthesis. This is the first study that considers the fusion between StyleGAN \\(\\mathcal{W}_{+}\\) space and diffusion-based image generation. 2) Embedding target identity in \\(\\mathcal{W}_{+}\\) space enables smooth and controllable semantic editing on facial attributes in our text-to-image model. 3) The effective disentanglement of identity-relevant and -irrelevant information in our model facilitates an identity-preserving generation that is not only diverse but also adaptable to a wide range of prompt instructions." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.362, + 0.64, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.388, + 0.892, + 0.719 + ], + "angle": 0, + "content": "StyleGAN Latent Space. GANs have drawn enormous attention due to the well-disentangled latent space [20]. StyleGAN [15] proposes to map input latent code to an intermediate latent space \\(\\mathcal{W}\\) to prevent warping of the training data distribution to fit in a particular probability distribution. The resulting \\(\\mathcal{W}\\) space is a semantically disentangled space that allows fine-grained controls over image synthesis. InterFaceGAN [29] and GANSpace [10] further interpret the latent space and identify several control directions such as age, gender, face angle, smile, etc. The powerful semantic editing ability has motivated research on inverting and editing real images in the StyleGAN space [1, 19, 24, 32]. Abdal et al. [1] utilize a direct optimization framework to embed a given real image to the extended StyleGAN space. Tov et al. [32] deploy an encoder to perform the inversion and keep the concatenated latent codes close to the original StyleGAN space to maintain high perceptual quality and editability. Motivated by the disentanglement and editability of StyleGAN latent space, we aim to introduce it to the T2I diffusion models to achieve more controllable synthesis. We follow Image2StyleGAN to denote the extended \\(\\mathcal{W}\\) space represented by multiple nonidentical \\(w\\) vectors as \\(\\mathcal{W}_{+}\\)." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Personalized Image Synthesis. Customization has been extensively studied in the context of T2I to generate images of specified objects or individuals [2, 8, 28]. Gal et al. [8] directly apply learnable text embedding optimization. Ruiz et al. [28] fine-tune the diffusion model for the target concept while preserving its class prior of the concept. Further improvements have been made to invert multiple concepts simultaneously with cross-attention fine-tuning [17], and disentangle background irrelevant information [3]. Encoder-based approaches [4, 9, 13, 30, 36–38] are employed for their efficiency. The target visual concept is encoded to embedding space as additional conditions. SingleInsert [37]" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "2188" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "adopts a two-stage scheme to insert concepts from a single image into the foreground region exclusively and enables the editing of the concept by text prompts. However, it is worth noticing that our method achieves more fine-grained editing with a single reference image at inference and does not necessitate identity-specific fine-tuned models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.474, + 0.408 + ], + "angle": 0, + "content": "With special attention on identity inversion utilizing the distinct features of human faces, Yuan et al. [40] define a celebrity space by applying PCA on selected celebrity name embeddings from CLIP and embedding new identities into the space via learnable coefficients. Valevski et al. [33] utilize a pre-trained face recognizer as the encoder and further project the face embedding to CLIP space. FaceChain [21] trains separate face and style LoRAs [12] to synthesize specific faces in specific styles. They follow a data preprocessing pipeline and apply face fusion of the best reference face on the synthesized results. However, we observe that previous personalization approaches inherit a common limitation that the synthesized faces share similar attributes as the reference, indicating that the model cannot disentangle irrelevant information well and overfits the particular reference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.472, + 0.591 + ], + "angle": 0, + "content": "Diffusion Model Adapters. Light-weight adapters have been adopted to avoid the laborious work of fine-tuning large models and add additional controls [23, 39, 42] to the diffusion model. ControlNet [42] proposes to train task-specific adapters while freezing the original diffusion model to adapt to various input conditions. Concurrent work T2I-Adapter [23] constructs a lighter-weight adapter for controls on structure and color. Ye et al. [39] employ a decoupled cross-attention module to consider both text and image prompts in the denoising process. In our work, we consider identity information as an additional condition and align it with the T2I model for identity-preserving synthesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.214, + 0.62 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.472, + 0.81 + ], + "angle": 0, + "content": "Our approach is capable of generating images that preserve identity while allowing for semantic edits, requiring just a single reference image for inference. This capability is realized by innovatively aligning StyleGAN's \\(\\mathcal{W}_{+}\\) latent space with the diffusion model. The training of our \\(\\mathcal{W}_{+}\\)adapter is divided into two stages. In Stage I (Sec. 3.2), we establish a mapping from \\(\\mathcal{W}_{+}\\) to SD latent space, using the resulting projection as an additional identity condition to synthesize center-aligned facial images of a specified identity. In Stage II (Sec. 3.3), this personalized generation process is expanded to accommodate more dynamic, \"in-the-wild\" settings, ensuring adaptability to a variety of textual prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.208, + 0.833 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In our implementation, we use Stable Diffusion [26] as the foundational T2I model, which is one of the most commonly used latent space-based text-to-image generation model [6, 34]. During the training phase, the model archi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "tecture includes: 1) a pre-trained encoder \\(\\mathcal{E}(\\cdot)\\), which transforms an image \\(I\\) into a latent representation with reduced dimensionality, 2) a conditional diffusion model tasked with predicting the latent code of a previous timestep, based on the text condition \\(c_{txt}\\) and the current timestep's latent code, and 3) a pre-trained decoder \\(\\mathcal{D}(\\cdot)\\) that reconstructs the final latent code into the synthesized image. The learning objective of this model is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.217, + 0.894, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, c _ {t x t}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} (c _ {t x t})\\right) \\| _ {2} ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.895, + 0.361 + ], + "angle": 0, + "content": "where \\( z_0 \\) is \\( \\mathcal{E}(I) \\), \\( \\epsilon \\) is the ground truth noise added for current timestep \\( t \\), \\( z_t \\) is the latent code of timestep \\( t \\). The \\( \\tau_{txt} \\) denotes the pre-trained CLIP text encoder [25] that converts the text prompt \\( c_{txt} \\) to textual embeddings. The \\( \\epsilon_{\\theta} \\) represents the UNet [27] denosing network. In the inference stage, the denoising network is applied iteratively to denoise a random sampled noise \\( z_T \\) to \\( z_0 \\) with condition \\( c_{txt} \\). The final result is then generated through the pre-trained decoder, i.e., \\( \\mathcal{D}(z_0) \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.38, + 0.866, + 0.396 + ], + "angle": 0, + "content": "3.2. Stage I: Aligning \\(\\mathcal{W}_{+}\\) with Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.895, + 0.513 + ], + "angle": 0, + "content": "The goal of Stage I is to align \\(\\mathcal{W}_{+}\\) space with SD to carry identity information in T2I generation while retaining its editing ability. To do so, we train a mapping network to project a \\(w_{+}\\) embedding to SD latent space and inject this extra condition into SD by modifying the cross-attention module. The framework of Stage I is illustrated in Fig. 2. Details are introduced below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.897, + 0.697 + ], + "angle": 0, + "content": "Training Pair Construction. We adopt 100K discrete training samples to fit the continuous distribution of \\(\\mathcal{W}_{+}\\) and align it with SD. For each face image \\(I_{f}\\), we use the pretrained e4e [32] encoder from StyleGAN inversion task to get its corresponding \\(w_{+} \\in \\mathbb{R}^{18 \\times 512}\\) vector. The pairs of \\(\\{I_{f}, w_{+}\\}\\) constitute our training data in this stage. Specifically, to generalize on real-world face images and improve \\(w_{+}\\) diversity, two types of training pairs are built, i.e., synthetic face images from StyleGAN2 [16] and real-world face images from FFHQ [15]. Note that in Stage I, we only consider the aligned face images in order to exclude any extraneous influences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Mapping Network. Since the original \\(\\mathcal{W}_{+}\\) space is designated for StyleGAN generation, we project the vector \\(w_{+} \\in \\mathbb{R}^{18 \\times 512}\\) to four tokens of dimension 768 (\\(\\mathbb{R}^{4 \\times 768}\\)) to align with the input dimension of SD condition. The mapping network consists of four groups of linear layers and is denoted as \\(\\mathcal{F}_w(\\cdot)\\). The latent space after mapping inherits the editability and disentanglement properties of \\(\\mathcal{W}_{+}\\) space and is compatible with the SD generation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Residual Cross-Attention. To incorporate the identity condition from the projected \\( w_{+} \\) embedding into the pre-trained SD model, we introduce a residual cross-attention module. In the standard SD framework, the output of cross-attention is determined using query features \\( f_{z} \\) from the hidden state" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.072, + 0.391, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.25, + 0.38, + 0.264 + ], + "angle": 0, + "content": "(a) Stage I: Aligning \\( w_{+} \\) with Stable Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.072, + 0.726, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.251, + 0.665, + 0.264 + ], + "angle": 0, + "content": "(b) Stage II: Generation in the Wild" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.072, + 0.885, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.251, + 0.877, + 0.264 + ], + "angle": 0, + "content": "(c) Residual Cross-Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.894, + 0.299 + ], + "angle": 0, + "content": "Figure 2. Overview of \\(\\mathcal{W}_{+}\\) adapter training stages. Left: Stage I for aligning \\(\\mathcal{W}_{+}\\) space with Stable Diffusion. Middle: Stage II for generating in-the-wild images with \\(w_{+}\\) embeddings. Right: details of residual cross-attention module. Open lock indicates trainable parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.453, + 0.325 + ], + "angle": 0, + "content": "and the text condition \\(\\tau_{txt}(c_{txt})\\). The process is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.329, + 0.47, + 0.361 + ], + "angle": 0, + "content": "\\[\nf _ {z} ^ {\\prime} = \\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.485 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q} = f_z\\mathbf{W}_q\\), \\(\\mathbf{K} = \\tau_{txt}(c_{txt})\\mathbf{W}_k\\), \\(\\mathbf{V} = \\tau_{txt}(c_{txt})\\mathbf{W}_v\\) are the query, key, value matrices of the attention module, respectively. \\(\\mathbf{W}_q\\), \\(\\mathbf{W}_k\\), and \\(\\mathbf{W}_v\\) denote the corresponding projection matrices. The cross-attention module is where the latent noise interacts with text condition embeddings inside the denoising process. Following previous methods [36, 39], we add an additional condition by a separate cross-attention module. Our residual cross-attention is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.489, + 0.47, + 0.507 + ], + "angle": 0, + "content": "\\[\nf _ {z} ^ {\\prime \\prime} = f _ {z} ^ {\\prime} + \\lambda \\cdot \\operatorname {A t t e n t i o n} \\left(\\mathbf {Q} ^ {\\dagger}, \\mathbf {K} ^ {\\dagger}, \\mathbf {V} ^ {\\dagger}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.727 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}^{\\dagger} = f_{z}^{\\prime}\\mathbf{W}_{q}^{\\dagger}\\), \\(\\mathbf{K}^{\\dagger} = f_{w}\\mathbf{W}_{k}^{\\dagger}\\), and \\(\\mathbf{V}^{\\dagger} = f_{w}\\mathbf{W}_{v}^{\\dagger}\\). The \\(\\mathbf{W}_{q}^{\\dagger},\\mathbf{W}_{k}^{\\dagger}\\), and \\(\\mathbf{W}_{v}^{\\dagger}\\) are the projection matrices for query, key, and value, respectively. The \\(f_{w} = \\mathcal{F}_{w}(w_{+})\\) is mapped from \\(w_{+}\\) by the mapping network. We use \\(\\lambda\\) as a scale parameter to balance the influence of text and identity conditions on the generation. We incorporate the decoupled cross-attention module in a residual fashion instead of the parallel approach as in previous works [36, 39] to avoid performance degradation on the original text condition. When \\(\\lambda\\) is set to 0, our \\(w_{+}\\) vector has no impact on the pre-trained SD model. We set \\(\\lambda\\) to 1 during training. The \\(\\mathbf{Q}^{\\dagger}\\), \\(\\mathbf{K}^{\\dagger}\\), and \\(\\mathbf{V}^{\\dagger}\\) are initialized from their corresponding Q, K, and V, respectively. The residual cross-attention module is performed on all the cross-attention layers of SD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.791 + ], + "angle": 0, + "content": "Learning Objectives. The trainable parts in this stage include the mapping network \\(\\mathcal{F}_w\\) and \\(\\{\\mathbf{Q}^\\dagger, \\mathbf{K}^\\dagger, \\mathbf{V}^\\dagger\\}\\) matrices in all residual cross-attention modules. The optimization target is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.796, + 0.47, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w _ {+}} = \\mathbb {E} _ {z _ {0}, \\epsilon , t, c _ {t x t}, w _ {+}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\mathcal {F} _ {w} (w _ {+})) \\| _ {2} ^ {2} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Eqn. (4) is similar to Eqn. (1) except for the additional \\(\\mathcal{F}_w(w_+)\\) serving as identity condition and cross-attention structure modified inside \\(\\epsilon_{\\theta}\\). The text prompt \\(c_{txt}\\) is randomly selected from several neutral templates describing a human face, e.g., \"a photo of a face\", or \"a face\" (see suppl.)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Through mapping the \\(\\mathcal{W}_{+}\\) distribution into an embedding space compatible with SD, our approach effectively uses the \\(w_{+}\\) vector to condition the personalized generation of aligned facial images. Furthermore, attribute directions \\(\\Delta w\\) derived from the original \\(\\mathcal{W}_{+}\\) space, encompassing features like smile, age, eye-opening, and others, remain applicable to our projected \\(w_{+}\\) vector. This allows for personalized, fine-grained face editing, as demonstrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.441, + 0.785, + 0.457 + ], + "angle": 0, + "content": "3.3. Stage II: Generation in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.616 + ], + "angle": 0, + "content": "In order to refine the performance of our \\(\\mathcal{W}_{+}\\)-adapted SD model from Stage I for scenarios beyond controlled environments, we continue on a second stage of training. This phase is dedicated to further fine-tuning the weights of \\(\\mathbf{Q}^{\\dagger},\\mathbf{K}^{\\dagger},\\mathbf{V}^{\\dagger}\\) across all residual cross-attention modules. To ensure that the projected \\(\\mathcal{W}_{+}\\) space exclusively encapsulates identity-related facial attributes, while remaining uninfluenced by irrelevant distractions from the background, we keep the mapping network fixed during this stage. The architecture and workflow of Stage II are depicted in the middle of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.894, + 0.737 + ], + "angle": 0, + "content": "Training Data Construction. For Stage II, in-the-wild images are used for training. For each image \\(I\\), its corresponding text caption, \\(c_{txt}\\), is extracted using an off-the-shelf captioning tool [18]. The aligned face image \\(I_f\\) is cropped from \\(I\\), and \\(w_+\\) is obtained from \\(I_f\\) by the e4e encoder (same as Stage I). A face region mask, \\(M\\), is also obtained from the image \\(I\\), with 0 denoting the face region and 1 representing the non-face region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.894, + 0.783 + ], + "angle": 0, + "content": "Learning Objectives. To keep high identity fidelity while encouraging high diversity of identity-irrelevant context, three losses are employed jointly in this stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.814 + ], + "angle": 0, + "content": "First, the reconstruction loss \\(\\mathcal{L}_{rec}\\) is used to guide the denoising process:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.818, + 0.892, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {z _ {0}, \\epsilon , t, c _ {t x t}, w _ {+}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\mathcal {F} _ {w} (w _ {+})) \\| _ {2} ^ {2} \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Note that Eqn. (5) is different from Eqn. (4) for that: 1) \\(z_0\\) here is encoded from the in-the-wild image \\(I\\) rather than the face image \\(I_f\\), so the reconstruction goal differs, and 2) \\(c_{txt}\\) is the caption describing the in-the-wild image rather than a" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "simple neutral prompt, therefore containing more information about the image context to facilitate reconstruction of \\(I\\) with \\(w_{+}\\) from \\(I_{f}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.213 + ], + "angle": 0, + "content": "Second, when semantic edits \\(\\Delta w\\) are applied in the \\(\\mathcal{W}_{+}\\) space, the objective is to exclusively modify the facial attributes, leaving the remaining regions unchanged. In order to accomplish this, we propose a \\(w_{+}\\) disentanglement loss to limit the editability of \\(w_{+}\\) outside the face region:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.218, + 0.47, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s e n}} = \\| M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right), \\mathcal {F} _ {w} \\left(w _ {+}\\right)\\right) - \\tag {6} \\\\ M \\cdot \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\Psi (\\mathcal {F} _ {w} (w _ {+}))) \\| \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.471, + 0.361 + ], + "angle": 0, + "content": "where \\(\\Psi\\) is the augmentation operation. During training, three augmentation strategies are adopted: 1) shuffle along the batch dimension of \\(\\mathcal{F}_w(w_+)\\), 2) add random perturbations of Gaussian noise on \\(\\mathcal{F}_w(w_+)\\), and 3) combine both of them. By applying such constraints, augmented (or edited) \\(w_+\\) vectors are forced to have similar non-face regions, thereby disentangling the effect of \\(\\Delta w\\) on the background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.469, + 0.422 + ], + "angle": 0, + "content": "Finally, since the identity condition \\(\\mathcal{F}_w(w_+)\\) should only influence the face region, which constitutes a relatively small proportion of the final output, we constrain its impact to be limited by a regularization loss:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.427, + 0.47, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\left\\| M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right), \\mathcal {F} _ {w} \\left(w _ {+}\\right)\\right) - M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right) \\right. \\right\\| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.508 + ], + "angle": 0, + "content": "In this way, a more balanced effect of text and identity condition on the synthesized result is achieved, further mitigating the risk of unwanted noisy information introduced from \\( w_{+} \\) and preserving the compatibility with text prompts." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.509, + 0.458, + 0.523 + ], + "angle": 0, + "content": "The overall learning objective in Stage II is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.53, + 0.47, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\gamma_ {1} \\cdot \\mathcal {L} _ {\\text {d i s e n}} + \\gamma_ {2} \\cdot \\mathcal {L} _ {\\text {r e g}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.4, + 0.565 + ], + "angle": 0, + "content": "where \\(\\gamma_{1}\\) and \\(\\gamma_{2}\\) denote the trade-off parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.209, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Training Data. In Stage I, we use the pre-trained e4e encoder [32] to obtain the \\( w_{+} \\) vector for each face image \\( I_{f} \\) from FFHQ [15] and StyleGAN2 [16]. FFHQ dataset contains 70,000 images, among which 600 are split for validation while the rest are for training. We also synthesize 70,000 images using StyleGAN2. In Stage II, we use FFHQ in-the-wild images (excluding those used for validation in Stage I) and SHHQ [7] to optimize the \\( \\mathcal{W}_{+} \\)adapter. SHHQ is a human dataset that contains 40,000 high-quality full-body images. BLIP2 [18] is introduced to generate the caption for each in-the-wild image. Face region mask \\( M \\) is obtained based on the FFHQ alignment operation and is eroded and blurred with kernel sizes of 32 and 7, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. All training is conducted on a server with 8 Tesla V100 GPUs. The batch size is set to 16. We employ the AdamW optimizer [22] with a learning rate of \\(1e - 4\\) and weight decay of 0.01. In Stage II, we adopt color jittering [44], random rotation, and sampling for inthe-wild images to increase diversity. During training, the cropped images with incomplete or small faces are discarded." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with previous methods. The best and second best results are highlighted by bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.122, + 0.892, + 0.238 + ], + "angle": 0, + "content": "
VariantsCLIP Score↑ID↓Detection↑
Textual Inversion [8].194.575.891
Dreambooth [28].177.562.800
Custom Diffusion [17].216.498.850
FastComposer [38].265.419.950
IP-Adapter-Face [39].241.407.958
CelebBasis [40].253.448.916
Ours.267.418.950
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.402 + ], + "angle": 0, + "content": "It takes around 24 hours to align the \\(\\mathcal{W}_{+}\\) space with SD in Stage I, and nearly 96 hours to train the \\(\\mathcal{W}_{+}\\) adapter in Stage II. In the inference, we adopt DDIM sampler [31] with 50 steps. To enable classifier-free guidance [11], we use the default settings and set the guidance scale to 7.5. We follow IP-Adapter [39] to randomly drop the text \\(c_{txt}\\) and \\(w_{+}\\) vector with a probability of 0.05. \\(\\gamma_{1}\\) and \\(\\gamma_{2}\\) in Eqn. (8) are set to 1.5 and 1.0, respectively. The whole framework is implemented on Diffusers [35]. Our experiments are based on SD v1.5 and can be extended to other SD models (see suppl.)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.413, + 0.733, + 0.43 + ], + "angle": 0, + "content": "4.1. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.895, + 0.649 + ], + "angle": 0, + "content": "Baselines. The methods we evaluate can be categorized into two distinct groups: general object customization and specific face personalization. For general object customization, we use Dreambooth [28], Textual Inversion [8], and Custom Diffusion [17], implementing these using Diffusers [35]. For specific face personalization, we select closely related works such as FastComposer [38], IP-Adapter-Face [39], and CelebBasis [40]. In our evaluation, we adhere to the settings established by CelebBasis [40] and FastComposer [38]. The metrics used for assessment include the CLIP Score [25], which is calculated based on the average image-text similarity using CLIP-L/14, the Face Detection Score determined using MTCNN [41], and the Identity Distance (ID) measured using ArcFace [5] on the detected facial regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.895, + 0.844 + ], + "angle": 0, + "content": "Given the extensive fine-tuning requirements for each identity in methods like Dreambooth, Textual Inversion, Custom Diffusion, and CelebBasis – which can be notably time-consuming (for instance, Textual Inversion demands about an hour on a single V100 GPU server\\(^{1}\\)) – we encounter constraints in evaluating a substantial number of test instances. To ensure a thorough and equitable quantitative analysis, we carefully choose 120 identities from the CelebA-HQ dataset [14], using one reference image per identity for consistent comparisons. In addition, we randomly select 20 text prompts that describe various aspects such as clothing, styles, and backgrounds, to provide detailed characterizations of each individual." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.846, + 0.894, + 0.876 + ], + "angle": 0, + "content": "Table 1 shows that optimization-based methods operating in the text embedding space, such as Textual Inversion," + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.842, + 0.9 + ], + "angle": 0, + "content": "1https://huggingface.co/docs/diffusers/training/text_inversion" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2191" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.078, + 0.175, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.093, + 0.191, + 0.131 + ], + "angle": 270, + "content": "a child in a yellow blazer" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.079, + 0.278, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.079, + 0.364, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.079, + 0.451, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.079, + 0.537, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.079, + 0.626, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.079, + 0.71, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.079, + 0.796, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.079, + 0.882, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.495, + 0.816, + 0.51 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons with baselines on different scenarios of T2I generation. Best view by zooming in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.827 + ], + "angle": 0, + "content": "Dreambooth, and Custom Diffusion, fall short in accurately aligning with text prompts and preserving identity features. On the contrary, our method shows comparable performance to FastComposer and IP-Adapter-Face in terms of both CLIP Score and ID metrics. However, it is important to note that IP-Adapter has been fine-tuned on the large-scale LAION2B dataset, whereas our model is trained on the relatively smaller FFHQ-wild and SHHQ datasets, comprising around 110,000 images. Both FastComposer and IP-Adapter directly process the reference image as input, while our approach first maps it to the \\(\\mathcal{W}_{+}\\) space before embedding it into the SD model. This additional mapping step could potentially introduce minor discrepancies when integrating the reference image into the SD model, slightly affecting the ID results. Nevertheless, this slight trade-off in ID preservation is balanced by our method's ability to flexibly edit facial attributes while ensuring the background remains consistent. This unique balance of attribute editability and background consistency is a novel contribution that sets our method apart from previous approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.3, + 0.849 + ], + "angle": 0, + "content": "4.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Figure 3 depicts visual comparisons with baselines across various scenarios. Since Textual Inversion, Dreambooth, and Custom Diffusion are not specifically designed for facial im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.689 + ], + "angle": 0, + "content": "ages, we select results with the best facial quality from their generated set of 40 images. For the remaining methods, we choose from a set of 10. It is observed that Textual Inversion, Dreambooth, and Custom Diffusion encounter difficulties in capturing and maintaining identity details with a single reference, leading to sub-optimal performance in this task. CelebBasis and FastComposer face challenges in striking a balance between text compatibility and identity preservation. Though IP-Adapter shows improved identity retention, it tends to ignore text conditions in certain instances (see the 1st, 3~5th rows)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.896, + 0.873 + ], + "angle": 0, + "content": "Leveraging our \\(\\mathcal{W}_{+}\\)adapter, our approach successfully generates images that are not only compatible with text descriptions but also more effectively retain the target identity. Additionally, our method allows for the editing of facial attributes along the \\(\\Delta w\\) direction, causing only minor alterations in the non-facial regions (illustrated in the last column). Furthermore, our approach can be seamlessly adapted to other pre-trained SD models without the need for additional fine-tuning, while retaining its editing capabilities. This versatility is exemplified in the last row of Fig. 3, which showcases our method's effectiveness with the dreamlike-anime model\\(^2\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.517, + 0.887, + 0.833, + 0.902 + ], + "angle": 0, + "content": "\\(^{2}\\)https://huggingface.co/dreamlike-art/dreamlike-anime-1.0" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2192" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.462, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.318, + 0.47, + 0.348 + ], + "angle": 0, + "content": "Figure 4. Inversion and editing comparisons between e4e and Ours in Stage I. They have the same \\( w_{+} \\) and attributes editing \\( \\Delta w \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.357, + 0.471, + 0.385 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of face inversion and editing. Editing (ID↓) is the ID metric on results of \\( w_{+} \\) and \\( w_{+} - 3 \\cdot \\Delta w \\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.387, + 0.46, + 0.448 + ], + "angle": 0, + "content": "
MethodsInversionEditing (ID↓)
ID↓FID↓LPIPS↓AgeSmile
e4e [32].43132.26.205.409.312
Ours (Stage I).43531.47.263.393.275
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.457, + 0.231, + 0.473 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Analyses of Aligning \\(\\mathcal{W}_{+}\\) in Stage I. We examine if the StyleGAN's \\(w_{+}\\) embedding is well aligned to the latent space of SD during Stage I training. Fig. 4 offers a qualitative comparison of inversion and attribute editing outcomes between e4e [32] and our approach, applied to a real-world image. Both methods use the identical \\(w_{+}\\) embedding generated by the e4e encoder. As depicted in Fig. 4, our method yields inversion results on par with e4e, as seen in the left column, using the same \\(w_{+}\\) vector. Furthermore, the alignment process we implement in this phase preserves the editability inherent to the \\(\\mathcal{W}_{+}\\) space. By introducing attribute editing directions \\(\\Delta w\\) (such as smile and age) from InterFaceGAN [29], our method can semantically adjust these attributes while maintaining the individual's identity. This is illustrated in the right columns, where the edited embedding, denoted as \\(w_{+} + \\alpha \\cdot \\Delta w\\), reflects these changes. The parameter \\(\\alpha\\) is used to control the extent of attribute modification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.472, + 0.876 + ], + "angle": 0, + "content": "To quantitatively evaluate the alignment performance, we randomly select 1,000 images from CelebA-HQ [14] and measure their inversion results with metrics of ID, FID, and LPIPS \\([43]^3\\). As indicated in the Inversion column of Table 2, our method exhibits comparable performance to e4e in the image inversion task. For assessing attribute editing capabilities, we focus on two key attributes, namely smile, and age, applying a constant scale \\(\\alpha\\) (i.e., \\(w_{+} - 3 \\cdot \\Delta w\\)). The Editing column in Table 2 demonstrates that our approach surpasses" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.088, + 0.86, + 0.101 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons of different variants." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.102, + 0.885, + 0.177 + ], + "angle": 0, + "content": "
VariantsCLIP Score↑ID↓Detection↑
Ours (PCA).243.526.936
Ours (w/o Lreg).136.461.951
Ours (L0.5reg).192.497.946
Ours (Full).267.516.943
" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.184, + 0.888, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.276, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons of smile attribute editing. The prompt is \"a woman wearing a t-shirt on the beach\"." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.312, + 0.893, + 0.372 + ], + "angle": 0, + "content": "e4e in maintaining identity during the editing process. These findings confirm that our method has successfully aligned the \\(\\mathcal{W}_{+}\\) space with the SD model, preserving both the inversion accuracy and the robust editability of attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Analyses of Variants. We examine different variants to assess each component of our \\(\\mathcal{W}_{+}\\)adapter: 1) Ours (PCA), which substitutes our residual cross-attention with parallel cross-attention as in [39], and 2) Ours (\\(w/o\\mathcal{L}_{reg}\\)), which omits the \\(\\mathcal{L}_{reg}\\) objective. The outcomes of these variants are presented in Table 3. It is observed that the residual cross-attention module outperforms the parallel cross-attention design. We conjecture that in tasks focused on integrating local facial imagery into broader, in-the-wild scenarios, directly injecting facial embeddings into the original hidden state \\(f_{z}\\) might adversely impact non-facial regions. Conversely, our residual cross-attention determines fusion weights by calculating attention activations between the text-embedded hidden state \\(f_{z}^{\\prime}\\) and our facial embedding. This results in a more precise and effective fusion within the facial region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.75 + ], + "angle": 0, + "content": "Besides, the results suggest that the regularization loss \\(\\mathcal{L}_{reg}\\) is crucial in maintaining semantic consistency between text prompts and images. In the absence of \\(\\mathcal{L}_{reg}\\), the variant Ours \\((w / o\\mathcal{L}_{reg})\\) tends to overemphasize facial regions while neglecting the accompanying text descriptions, leading to a reduced CLIP score. Incorporating \\(\\mathcal{L}_{reg}\\) with a weight of \\(\\gamma_{2} = 0.5\\), as in Ours \\((\\mathcal{L}_{reg}^{0.5})\\), enhances the CLIP score but slightly compromises the ID metric. In comparison, Ours (Full) achieves a satisfactory trade-off in balancing the performance of ID and CLIP score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.896, + 0.902 + ], + "angle": 0, + "content": "The impact of our disentanglement loss \\(\\mathcal{L}_{disen}\\) (Eqn. (6)) is visually demonstrated in Fig. 5. This example shows that omitting \\(\\mathcal{L}_{disen}\\) leads to unintended alterations in non-facial regions when the smile editing direction \\(\\Delta w\\) is applied to \\(w_{+}\\). This indicates the importance of \\(\\mathcal{L}_{disen}\\) in effectively separating identity-relevant and irrelevant information. By combining our residual cross-attention, regularization loss, and disentanglement loss, Ours (Full) preserves identity and ensures compatibility with text prompts, even amidst changes to the face embedding." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.333, + 0.901 + ], + "angle": 0, + "content": "3https://github.com/chaofengc/IQA-PyTorch" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2193" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.089, + 0.155, + 0.104 + ], + "angle": 0, + "content": "w+" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.093, + 0.58, + 0.104 + ], + "angle": 0, + "content": "T2I Generation with \\( w_{+} \\) Interpolation" + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.089, + 0.834, + 0.104 + ], + "angle": 0, + "content": "w" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.105, + 0.874, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 6. Visual results of \\( w_{+} \\) embeddings interpolation during inference. The prompts are \"one person wearing suit and tie in a garden\" and \"one person wearing a blue shirt by a secluded waterfall\", respectively." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.298, + 0.457, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.611 + ], + "angle": 0, + "content": "Figure 7. Visual results of using different \\(\\lambda\\) during inference. The prompt is \"a man wearing a rainbow shirt in a garden\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Analyses of \\(\\lambda\\) in Cross-attention. The parameter \\(\\lambda\\) determines the extent to which the \\(w_{+}\\) vector influences the hidden state \\(f_{z}^{\\prime}\\). Fig. 7 shows the effects of varying \\(\\lambda\\) during inference. At \\(\\lambda = 0\\), the outcome aligns with that of the original SD model, unaffected by any editing direction. However, as \\(\\lambda\\) increases, the generated identity more closely resembles the reference image. Interestingly, at relatively low values of \\(\\lambda\\) (for example, \\(\\lambda = 0.3\\)), even though the identity does not closely match the reference, our model effectively edits attributes in the specified direction. This observation suggests that our method proficiently leverages the \\(\\mathcal{W}_{+}\\) space from StyleGAN within the SD framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Analyses of \\( w_{+} \\) Interpolation. We select two facial images and acquire their respective \\( w_{+}^{1} \\) and \\( w_{+}^{2} \\) embeddings from e4e encoder. The interpolated \\( w_{+} \\) embedding is then obtained through \\( (1 - \\kappa) \\cdot w_{+}^{1} + \\kappa \\cdot w_{+}^{2} \\), where \\( \\kappa \\in [0,1] \\). As shown in Fig. 6, this T2I generation process results in a smooth transition in the facial regions of the generated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.895, + 0.376 + ], + "angle": 0, + "content": "images, while maintaining a similar and consistent layout throughout the interpolation. This result confirms that the \\( w_{+} \\) embedding in our method not only preserves the editability characteristic of the original StyleGAN but also effectively distinguishes between facial and background regions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.384, + 0.62, + 0.399 + ], + "angle": 0, + "content": "4.4. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.895, + 0.59 + ], + "angle": 0, + "content": "Our work aims at integrating StyleGAN's editable \\(\\mathcal{W}_{+}\\) space into the SD model. We notice a challenge in this integration: the process of converting real images to \\(w_{+}\\) vectors in StyleGAN often leads to a loss of detail, impacting the preservation of identity features. Despite employing a substantial number of training pairs \\(\\{I_f, w_{+}\\}\\) to establish the mapping, we still observe limitations in maintaining identity fidelity and challenges in editing certain attributes (e.g., pose and glasses). The current framework is designed to generate and edit images with only a single face. Our future work aims to explore the potential of applying localized injections of distinct \\(w_{+}\\) to address multiple human subjects." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.603, + 0.62, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.84 + ], + "angle": 0, + "content": "We have presented the first attempt to embed the \\(\\mathcal{W}_{+}\\) space of StyleGAN into the SD model. We showed that both the mapping network and the residual cross-attention module play crucial roles in facilitating the injection of \\(w_{+}\\) embedding into the SD model, balancing between text prompt influence and identity conditions. Our experiments demonstrate that the \\(\\mathcal{W}_{+}\\) space, as used in our approach, not only enables personalized text-to-image generation but also allows for precise editing of facial attributes. We envision this capability being highly beneficial in various practical applications, such as portrait customization with seamless attribute modifications. Furthermore, our methodology holds the potential for application across other object domains that possess distinct prior spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2194" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: How to embed images into the StyleGAN latent space? In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.199 + ], + "angle": 0, + "content": "[2] Yufei Cai, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Hu Han, and Wangmeng Zuo. Decoupled textual embeddings for customized image generation. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[3] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. DisenBooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.256, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[4] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, and Min Zheng. PhotoVerse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.366 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.408 + ], + "angle": 0, + "content": "[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.472, + 0.464 + ], + "angle": 0, + "content": "[7] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. StyleGAN-Human: A data-centric odyssey of human generation. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[9] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[10] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.471, + 0.703 + ], + "angle": 0, + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[13] Xuhui Jia, Yang Zhao, Kelvin C. K. Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[19] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image superresolution. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[20] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. A survey on leveraging pre-trained generative adversarial networks for image editing and restoration. arXiv preprint arXiv:2207.10309, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.894, + 0.373 + ], + "angle": 0, + "content": "[21] Yang Liu, Cheng Yu, Lei Shang, Ziheng Wu, Xingjun Wang, Yuze Zhao, Lin Zhu, Chen Cheng, Weitao Chen, Chao Xu, Haoyu Xie, Yuan Yao, Wenmeng Zhou, Yingda Chen, Xuansong Xie, and Baigui Sun. FaceChain: A playground for identity-preserving portrait generation. arXiv preprint arXiv:2308.14256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.894, + 0.516 + ], + "angle": 0, + "content": "[24] Hamza Pehlivan, Yusuf Dalva, and Aysegul Dundar. StyleRes: Transforming the residuals for real image editing with StyleGAN. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.894, + 0.586 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.893, + 0.63 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.894, + 0.729 + ], + "angle": 0, + "content": "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Finetuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.894, + 0.772 + ], + "angle": 0, + "content": "[29] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.894, + 0.828 + ], + "angle": 0, + "content": "[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. InstantBooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. In ACM TOG, 2021. 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.146 + ], + "angle": 0, + "content": "[33] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.471, + 0.246 + ], + "angle": 0, + "content": "[35] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[37] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. SingleInsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[38] Guangxuan Xiao, Tianwei Yin, William T. Freeman, Frédo Durand, and Song Han. FastComposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[39] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. In NeurIPS, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.471, + 0.586 + ], + "angle": 0, + "content": "[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. In IEEE signal processing letters, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.631, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.674, + 0.471, + 0.715 + ], + "angle": 0, + "content": "[44] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In ECCV, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2196" + } + ] +] \ No newline at end of file diff --git a/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_origin.pdf b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2e8810937f9a633ec9d581da2193f229b1e9b89 --- /dev/null +++ b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/2ed92abb-5de3-46c9-ba95-16327031ae50_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1864625375549f343bc15515ee6374cbfd8283f48a09aa252cd47ef1b517cf02 +size 8061753 diff --git a/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/full.md b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2eaaf9ba3fb222eec8c1cf676fdb0298f9934971 --- /dev/null +++ b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/full.md @@ -0,0 +1,292 @@ +# When StyleGAN Meets Stable Diffusion: a $\mathcal{W}_{+}$ Adapter for Personalized Image Generation + +Xiaoming Li Xinyu Hou Chen Change Loy S-Lab, Nanyang Technological University + +csxml@gmail.com xinyu.hou@ntu.edu.sg ccloy@ntu.edu.sg + +![](images/8b78a124102bb333611629ac7ec136ca867faa6d301c6dce65d5a141378ccdf3.jpg) +Figure 1. Given a single reference image (thumbnail in the top left), our $\mathcal{W}_{+}$ adapter not only integrates the identity into the text-to-image generation accurately but also enables modifications of facial attributes along the $\Delta w$ trajectory derived from StyleGAN. The text prompt is "a woman wearing a spacesuit in a forest". + +# Abstract + +Text-to-image diffusion models have remarkably excelled in producing diverse, high-quality, and photo-realistic images. This advancement has spurred a growing interest in incorporating specific identities into generated content. Most current methods employ an inversion approach to embed a target visual concept into the text embedding space using a single reference image. However, the newly synthesized faces either closely resemble the reference image in terms of facial attributes, such as expression, or exhibit a reduced capacity for identity preservation. Text descriptions intended to guide the facial attributes of the synthesized face may fall short, owing to the intricate entanglement of identity information with identity-irrelevant facial attributes derived from the reference image. To address these issues, + +we present the novel use of the extended StyleGAN embedding space $\mathcal{W}_{+}$ , to achieve enhanced identity preservation and disentanglement for diffusion models. By aligning this semantically meaningful human face latent space with text-to-image diffusion models, we succeed in maintaining high fidelity in identity preservation, coupled with the capacity for semantic editing. Additionally, we propose new training objectives to balance the influences of both prompt and identity conditions, ensuring that the identity-irrelevant background remains negligibly affected during facial attribute modifications. Extensive experiments reveal that our method adeptly generates personalized text-to-image outputs that are not only compatible with prompt descriptions but also amenable to common StyleGAN editing directions in diverse settings. Our code and model are available at https://github.com/csxmli2016/w-plus-adapter. + +# 1. Introduction + +What if we could be the protagonists of our own fantasies? This paper primarily addresses personalized text-to-image (T2I) generation, a field that is attracting growing attention due to users' desire to craft their unique content. Overlapped with customized T2I [8, 17, 28, 36], which integrates various visual concepts like human faces and objects into the generated imagery, our focus is on tailoring image generation to specific identities. This focus is motivated by applications such as storyboarding, where a consistent identity needs to be maintained across all images, despite variations in expressions or ages. Additionally, we believe that human faces possess unique, fine-grained intrinsic features that merit special attention and present a worthwhile area for exploration. + +Recent personalized image generation approaches utilize a small reference set of a target identity and embed it into a specific space. The common choice of embedding space has been the textual embedding space used by large language models (LLMs). While existing methods based on textual embedding [8, 28, 36] have proven capable of maintaining target identity, they are often limited by inherent trade-offs. Specifically, these methods face challenges in simultaneously preserving identity, generating varied facial attributes, and creating identity-irrelevant content that aligns with the text description. We observe that these issues predominantly stem from the entangled nature of the textual embedding space, where a single pseudo word $S^*$ struggles to distinctly isolate identity-related features from the reference image. Efforts to separate such information include approaches like encoding a visual concept into multiple word embeddings [36] or employing a separate branch for extracting identity-irrelevant details [3]. However, these methods tend to overlook the nuanced facial features critical to an individual's identity, resulting in incomplete disentanglement and sub-optimal identity preservation. + +In this study, we aim to more effectively separate identity-relevant and -agnostic features for better identity preservation, while also ensuring editability. To achieve this, we propose inverting the visual concept of a target identity into StyleGAN's [15] $\mathcal{W}_{+}$ latent space, as opposed to using the textual embedding space. In particular, we introduce a mapping network to integrate the $\mathcal{W}_{+}$ space with the diffusion model and a residual cross-attention module to add the $w_{+}$ vector as an additional identity condition. This mapping network, once trained with (image, $w_{+}$ ) pairs, can generalize to unseen individuals during inference without the need for a separate identity-specific model for each person. We further present novel regularized training to ensure that edits have a negligible effect on identity-irrelevant regions and that the overall generation remains aligned with the prompt conditions. It is noteworthy that although CelebBasis [40] explores a similar concept of creating a face-specific latent space through PCA of selected celebrity name embeddings, + +we argue that relying on such a limited dataset of celebrity names is insufficient for a comprehensive face latent space due to potential under-representation. Our experimental results confirm that our approach exhibits a superior capacity for identity preservation compared to CelebBasis. + +The contributions are summarised as follows: 1) We introduce extended StyleGAN $\mathcal{W}_{+}$ space as a target inversion latent space to better encode the identity-preserving facial concepts for personalized text-to-image synthesis. This is the first study that considers the fusion between StyleGAN $\mathcal{W}_{+}$ space and diffusion-based image generation. 2) Embedding target identity in $\mathcal{W}_{+}$ space enables smooth and controllable semantic editing on facial attributes in our text-to-image model. 3) The effective disentanglement of identity-relevant and -irrelevant information in our model facilitates an identity-preserving generation that is not only diverse but also adaptable to a wide range of prompt instructions. + +# 2. Related Work + +StyleGAN Latent Space. GANs have drawn enormous attention due to the well-disentangled latent space [20]. StyleGAN [15] proposes to map input latent code to an intermediate latent space $\mathcal{W}$ to prevent warping of the training data distribution to fit in a particular probability distribution. The resulting $\mathcal{W}$ space is a semantically disentangled space that allows fine-grained controls over image synthesis. InterFaceGAN [29] and GANSpace [10] further interpret the latent space and identify several control directions such as age, gender, face angle, smile, etc. The powerful semantic editing ability has motivated research on inverting and editing real images in the StyleGAN space [1, 19, 24, 32]. Abdal et al. [1] utilize a direct optimization framework to embed a given real image to the extended StyleGAN space. Tov et al. [32] deploy an encoder to perform the inversion and keep the concatenated latent codes close to the original StyleGAN space to maintain high perceptual quality and editability. Motivated by the disentanglement and editability of StyleGAN latent space, we aim to introduce it to the T2I diffusion models to achieve more controllable synthesis. We follow Image2StyleGAN to denote the extended $\mathcal{W}$ space represented by multiple nonidentical $w$ vectors as $\mathcal{W}_{+}$ . + +Personalized Image Synthesis. Customization has been extensively studied in the context of T2I to generate images of specified objects or individuals [2, 8, 28]. Gal et al. [8] directly apply learnable text embedding optimization. Ruiz et al. [28] fine-tune the diffusion model for the target concept while preserving its class prior of the concept. Further improvements have been made to invert multiple concepts simultaneously with cross-attention fine-tuning [17], and disentangle background irrelevant information [3]. Encoder-based approaches [4, 9, 13, 30, 36–38] are employed for their efficiency. The target visual concept is encoded to embedding space as additional conditions. SingleInsert [37] + +adopts a two-stage scheme to insert concepts from a single image into the foreground region exclusively and enables the editing of the concept by text prompts. However, it is worth noticing that our method achieves more fine-grained editing with a single reference image at inference and does not necessitate identity-specific fine-tuned models. + +With special attention on identity inversion utilizing the distinct features of human faces, Yuan et al. [40] define a celebrity space by applying PCA on selected celebrity name embeddings from CLIP and embedding new identities into the space via learnable coefficients. Valevski et al. [33] utilize a pre-trained face recognizer as the encoder and further project the face embedding to CLIP space. FaceChain [21] trains separate face and style LoRAs [12] to synthesize specific faces in specific styles. They follow a data preprocessing pipeline and apply face fusion of the best reference face on the synthesized results. However, we observe that previous personalization approaches inherit a common limitation that the synthesized faces share similar attributes as the reference, indicating that the model cannot disentangle irrelevant information well and overfits the particular reference. + +Diffusion Model Adapters. Light-weight adapters have been adopted to avoid the laborious work of fine-tuning large models and add additional controls [23, 39, 42] to the diffusion model. ControlNet [42] proposes to train task-specific adapters while freezing the original diffusion model to adapt to various input conditions. Concurrent work T2I-Adapter [23] constructs a lighter-weight adapter for controls on structure and color. Ye et al. [39] employ a decoupled cross-attention module to consider both text and image prompts in the denoising process. In our work, we consider identity information as an additional condition and align it with the T2I model for identity-preserving synthesis. + +# 3. Methodology + +Our approach is capable of generating images that preserve identity while allowing for semantic edits, requiring just a single reference image for inference. This capability is realized by innovatively aligning StyleGAN's $\mathcal{W}_{+}$ latent space with the diffusion model. The training of our $\mathcal{W}_{+}$ adapter is divided into two stages. In Stage I (Sec. 3.2), we establish a mapping from $\mathcal{W}_{+}$ to SD latent space, using the resulting projection as an additional identity condition to synthesize center-aligned facial images of a specified identity. In Stage II (Sec. 3.3), this personalized generation process is expanded to accommodate more dynamic, "in-the-wild" settings, ensuring adaptability to a variety of textual prompts. + +# 3.1. Preliminary + +In our implementation, we use Stable Diffusion [26] as the foundational T2I model, which is one of the most commonly used latent space-based text-to-image generation model [6, 34]. During the training phase, the model archi + +tecture includes: 1) a pre-trained encoder $\mathcal{E}(\cdot)$ , which transforms an image $I$ into a latent representation with reduced dimensionality, 2) a conditional diffusion model tasked with predicting the latent code of a previous timestep, based on the text condition $c_{txt}$ and the current timestep's latent code, and 3) a pre-trained decoder $\mathcal{D}(\cdot)$ that reconstructs the final latent code into the synthesized image. The learning objective of this model is formulated as follows: + +$$ +\mathcal {L} _ {L D M} = \mathbb {E} _ {z _ {0}, \epsilon \sim \mathcal {N} (0, 1), t, c _ {t x t}} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {t x t} (c _ {t x t})\right) \| _ {2} ^ {2} \right] \tag {1} +$$ + +where $z_0$ is $\mathcal{E}(I)$ , $\epsilon$ is the ground truth noise added for current timestep $t$ , $z_t$ is the latent code of timestep $t$ . The $\tau_{txt}$ denotes the pre-trained CLIP text encoder [25] that converts the text prompt $c_{txt}$ to textual embeddings. The $\epsilon_{\theta}$ represents the UNet [27] denosing network. In the inference stage, the denoising network is applied iteratively to denoise a random sampled noise $z_T$ to $z_0$ with condition $c_{txt}$ . The final result is then generated through the pre-trained decoder, i.e., $\mathcal{D}(z_0)$ . + +# 3.2. Stage I: Aligning $\mathcal{W}_{+}$ with Stable Diffusion + +The goal of Stage I is to align $\mathcal{W}_{+}$ space with SD to carry identity information in T2I generation while retaining its editing ability. To do so, we train a mapping network to project a $w_{+}$ embedding to SD latent space and inject this extra condition into SD by modifying the cross-attention module. The framework of Stage I is illustrated in Fig. 2. Details are introduced below. + +Training Pair Construction. We adopt 100K discrete training samples to fit the continuous distribution of $\mathcal{W}_{+}$ and align it with SD. For each face image $I_{f}$ , we use the pretrained e4e [32] encoder from StyleGAN inversion task to get its corresponding $w_{+} \in \mathbb{R}^{18 \times 512}$ vector. The pairs of $\{I_{f}, w_{+}\}$ constitute our training data in this stage. Specifically, to generalize on real-world face images and improve $w_{+}$ diversity, two types of training pairs are built, i.e., synthetic face images from StyleGAN2 [16] and real-world face images from FFHQ [15]. Note that in Stage I, we only consider the aligned face images in order to exclude any extraneous influences. + +Mapping Network. Since the original $\mathcal{W}_{+}$ space is designated for StyleGAN generation, we project the vector $w_{+} \in \mathbb{R}^{18 \times 512}$ to four tokens of dimension 768 ( $\mathbb{R}^{4 \times 768}$ ) to align with the input dimension of SD condition. The mapping network consists of four groups of linear layers and is denoted as $\mathcal{F}_w(\cdot)$ . The latent space after mapping inherits the editability and disentanglement properties of $\mathcal{W}_{+}$ space and is compatible with the SD generation process. + +Residual Cross-Attention. To incorporate the identity condition from the projected $w_{+}$ embedding into the pre-trained SD model, we introduce a residual cross-attention module. In the standard SD framework, the output of cross-attention is determined using query features $f_{z}$ from the hidden state + +![](images/854400d374ea713f011c018896c36e324d5d9658c49165eab8ecf6d4c31f7c8b.jpg) +(a) Stage I: Aligning $w_{+}$ with Stable Diffusion + +![](images/50f804bf3afe9299f74c7c80640d75167ad106d37009de8fcc2a88873b170f74.jpg) +(b) Stage II: Generation in the Wild +Figure 2. Overview of $\mathcal{W}_{+}$ adapter training stages. Left: Stage I for aligning $\mathcal{W}_{+}$ space with Stable Diffusion. Middle: Stage II for generating in-the-wild images with $w_{+}$ embeddings. Right: details of residual cross-attention module. Open lock indicates trainable parts. + +![](images/eea6e23c46f92d07b8307cbd5ec6e18182f38192579f3e0a15080db4b67e1bec.jpg) +(c) Residual Cross-Attention + +and the text condition $\tau_{txt}(c_{txt})$ . The process is defined as + +$$ +f _ {z} ^ {\prime} = \operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {\top}}{\sqrt {d}}\right) \mathbf {V} \tag {2} +$$ + +where $\mathbf{Q} = f_z\mathbf{W}_q$ , $\mathbf{K} = \tau_{txt}(c_{txt})\mathbf{W}_k$ , $\mathbf{V} = \tau_{txt}(c_{txt})\mathbf{W}_v$ are the query, key, value matrices of the attention module, respectively. $\mathbf{W}_q$ , $\mathbf{W}_k$ , and $\mathbf{W}_v$ denote the corresponding projection matrices. The cross-attention module is where the latent noise interacts with text condition embeddings inside the denoising process. Following previous methods [36, 39], we add an additional condition by a separate cross-attention module. Our residual cross-attention is defined as: + +$$ +f _ {z} ^ {\prime \prime} = f _ {z} ^ {\prime} + \lambda \cdot \operatorname {A t t e n t i o n} \left(\mathbf {Q} ^ {\dagger}, \mathbf {K} ^ {\dagger}, \mathbf {V} ^ {\dagger}\right) \tag {3} +$$ + +where $\mathbf{Q}^{\dagger} = f_{z}^{\prime}\mathbf{W}_{q}^{\dagger}$ , $\mathbf{K}^{\dagger} = f_{w}\mathbf{W}_{k}^{\dagger}$ , and $\mathbf{V}^{\dagger} = f_{w}\mathbf{W}_{v}^{\dagger}$ . The $\mathbf{W}_{q}^{\dagger},\mathbf{W}_{k}^{\dagger}$ , and $\mathbf{W}_{v}^{\dagger}$ are the projection matrices for query, key, and value, respectively. The $f_{w} = \mathcal{F}_{w}(w_{+})$ is mapped from $w_{+}$ by the mapping network. We use $\lambda$ as a scale parameter to balance the influence of text and identity conditions on the generation. We incorporate the decoupled cross-attention module in a residual fashion instead of the parallel approach as in previous works [36, 39] to avoid performance degradation on the original text condition. When $\lambda$ is set to 0, our $w_{+}$ vector has no impact on the pre-trained SD model. We set $\lambda$ to 1 during training. The $\mathbf{Q}^{\dagger}$ , $\mathbf{K}^{\dagger}$ , and $\mathbf{V}^{\dagger}$ are initialized from their corresponding Q, K, and V, respectively. The residual cross-attention module is performed on all the cross-attention layers of SD. + +Learning Objectives. The trainable parts in this stage include the mapping network $\mathcal{F}_w$ and $\{\mathbf{Q}^\dagger, \mathbf{K}^\dagger, \mathbf{V}^\dagger\}$ matrices in all residual cross-attention modules. The optimization target is formulated as: + +$$ +\mathcal {L} _ {w _ {+}} = \mathbb {E} _ {z _ {0}, \epsilon , t, c _ {t x t}, w _ {+}} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, \tau_ {t x t} (c _ {t x t}), \mathcal {F} _ {w} (w _ {+})) \| _ {2} ^ {2} \right] \tag {4} +$$ + +Eqn. (4) is similar to Eqn. (1) except for the additional $\mathcal{F}_w(w_+)$ serving as identity condition and cross-attention structure modified inside $\epsilon_{\theta}$ . The text prompt $c_{txt}$ is randomly selected from several neutral templates describing a human face, e.g., "a photo of a face", or "a face" (see suppl.). + +Through mapping the $\mathcal{W}_{+}$ distribution into an embedding space compatible with SD, our approach effectively uses the $w_{+}$ vector to condition the personalized generation of aligned facial images. Furthermore, attribute directions $\Delta w$ derived from the original $\mathcal{W}_{+}$ space, encompassing features like smile, age, eye-opening, and others, remain applicable to our projected $w_{+}$ vector. This allows for personalized, fine-grained face editing, as demonstrated in Fig. 1. + +# 3.3. Stage II: Generation in the Wild + +In order to refine the performance of our $\mathcal{W}_{+}$ -adapted SD model from Stage I for scenarios beyond controlled environments, we continue on a second stage of training. This phase is dedicated to further fine-tuning the weights of $\mathbf{Q}^{\dagger},\mathbf{K}^{\dagger},\mathbf{V}^{\dagger}$ across all residual cross-attention modules. To ensure that the projected $\mathcal{W}_{+}$ space exclusively encapsulates identity-related facial attributes, while remaining uninfluenced by irrelevant distractions from the background, we keep the mapping network fixed during this stage. The architecture and workflow of Stage II are depicted in the middle of Fig. 2. + +Training Data Construction. For Stage II, in-the-wild images are used for training. For each image $I$ , its corresponding text caption, $c_{txt}$ , is extracted using an off-the-shelf captioning tool [18]. The aligned face image $I_f$ is cropped from $I$ , and $w_+$ is obtained from $I_f$ by the e4e encoder (same as Stage I). A face region mask, $M$ , is also obtained from the image $I$ , with 0 denoting the face region and 1 representing the non-face region. + +Learning Objectives. To keep high identity fidelity while encouraging high diversity of identity-irrelevant context, three losses are employed jointly in this stage. + +First, the reconstruction loss $\mathcal{L}_{rec}$ is used to guide the denoising process: + +$$ +\mathcal {L} _ {r e c} = \mathbb {E} _ {z _ {0}, \epsilon , t, c _ {t x t}, w _ {+}} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, \tau_ {t x t} (c _ {t x t}), \mathcal {F} _ {w} (w _ {+})) \| _ {2} ^ {2} \right] \tag {5} +$$ + +Note that Eqn. (5) is different from Eqn. (4) for that: 1) $z_0$ here is encoded from the in-the-wild image $I$ rather than the face image $I_f$ , so the reconstruction goal differs, and 2) $c_{txt}$ is the caption describing the in-the-wild image rather than a + +simple neutral prompt, therefore containing more information about the image context to facilitate reconstruction of $I$ with $w_{+}$ from $I_{f}$ . + +Second, when semantic edits $\Delta w$ are applied in the $\mathcal{W}_{+}$ space, the objective is to exclusively modify the facial attributes, leaving the remaining regions unchanged. In order to accomplish this, we propose a $w_{+}$ disentanglement loss to limit the editability of $w_{+}$ outside the face region: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i s e n}} = \| M \cdot \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {t x t} \left(c _ {t x t}\right), \mathcal {F} _ {w} \left(w _ {+}\right)\right) - \tag {6} \\ M \cdot \epsilon_ {\theta} (z _ {t}, t, \tau_ {t x t} (c _ {t x t}), \Psi (\mathcal {F} _ {w} (w _ {+}))) \| \\ \end{array} +$$ + +where $\Psi$ is the augmentation operation. During training, three augmentation strategies are adopted: 1) shuffle along the batch dimension of $\mathcal{F}_w(w_+)$ , 2) add random perturbations of Gaussian noise on $\mathcal{F}_w(w_+)$ , and 3) combine both of them. By applying such constraints, augmented (or edited) $w_+$ vectors are forced to have similar non-face regions, thereby disentangling the effect of $\Delta w$ on the background. + +Finally, since the identity condition $\mathcal{F}_w(w_+)$ should only influence the face region, which constitutes a relatively small proportion of the final output, we constrain its impact to be limited by a regularization loss: + +$$ +\mathcal {L} _ {r e g} = \left\| M \cdot \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {t x t} \left(c _ {t x t}\right), \mathcal {F} _ {w} \left(w _ {+}\right)\right) - M \cdot \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {t x t} \left(c _ {t x t}\right) \right. \right\| \tag {7} +$$ + +In this way, a more balanced effect of text and identity condition on the synthesized result is achieved, further mitigating the risk of unwanted noisy information introduced from $w_{+}$ and preserving the compatibility with text prompts. + +The overall learning objective in Stage II is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \gamma_ {1} \cdot \mathcal {L} _ {\text {d i s e n}} + \gamma_ {2} \cdot \mathcal {L} _ {\text {r e g}} \tag {8} +$$ + +where $\gamma_{1}$ and $\gamma_{2}$ denote the trade-off parameters. + +# 4. Experiments + +Training Data. In Stage I, we use the pre-trained e4e encoder [32] to obtain the $w_{+}$ vector for each face image $I_{f}$ from FFHQ [15] and StyleGAN2 [16]. FFHQ dataset contains 70,000 images, among which 600 are split for validation while the rest are for training. We also synthesize 70,000 images using StyleGAN2. In Stage II, we use FFHQ in-the-wild images (excluding those used for validation in Stage I) and SHHQ [7] to optimize the $\mathcal{W}_{+}$ adapter. SHHQ is a human dataset that contains 40,000 high-quality full-body images. BLIP2 [18] is introduced to generate the caption for each in-the-wild image. Face region mask $M$ is obtained based on the FFHQ alignment operation and is eroded and blurred with kernel sizes of 32 and 7, respectively. + +Implementation Details. All training is conducted on a server with 8 Tesla V100 GPUs. The batch size is set to 16. We employ the AdamW optimizer [22] with a learning rate of $1e - 4$ and weight decay of 0.01. In Stage II, we adopt color jittering [44], random rotation, and sampling for inthe-wild images to increase diversity. During training, the cropped images with incomplete or small faces are discarded. + +Table 1. Quantitative comparisons with previous methods. The best and second best results are highlighted by bold and underline. + +
VariantsCLIP Score↑ID↓Detection↑
Textual Inversion [8].194.575.891
Dreambooth [28].177.562.800
Custom Diffusion [17].216.498.850
FastComposer [38].265.419.950
IP-Adapter-Face [39].241.407.958
CelebBasis [40].253.448.916
Ours.267.418.950
+ +It takes around 24 hours to align the $\mathcal{W}_{+}$ space with SD in Stage I, and nearly 96 hours to train the $\mathcal{W}_{+}$ adapter in Stage II. In the inference, we adopt DDIM sampler [31] with 50 steps. To enable classifier-free guidance [11], we use the default settings and set the guidance scale to 7.5. We follow IP-Adapter [39] to randomly drop the text $c_{txt}$ and $w_{+}$ vector with a probability of 0.05. $\gamma_{1}$ and $\gamma_{2}$ in Eqn. (8) are set to 1.5 and 1.0, respectively. The whole framework is implemented on Diffusers [35]. Our experiments are based on SD v1.5 and can be extended to other SD models (see suppl.). + +# 4.1. Quantitative Comparison + +Baselines. The methods we evaluate can be categorized into two distinct groups: general object customization and specific face personalization. For general object customization, we use Dreambooth [28], Textual Inversion [8], and Custom Diffusion [17], implementing these using Diffusers [35]. For specific face personalization, we select closely related works such as FastComposer [38], IP-Adapter-Face [39], and CelebBasis [40]. In our evaluation, we adhere to the settings established by CelebBasis [40] and FastComposer [38]. The metrics used for assessment include the CLIP Score [25], which is calculated based on the average image-text similarity using CLIP-L/14, the Face Detection Score determined using MTCNN [41], and the Identity Distance (ID) measured using ArcFace [5] on the detected facial regions. + +Given the extensive fine-tuning requirements for each identity in methods like Dreambooth, Textual Inversion, Custom Diffusion, and CelebBasis – which can be notably time-consuming (for instance, Textual Inversion demands about an hour on a single V100 GPU server $^{1}$ ) – we encounter constraints in evaluating a substantial number of test instances. To ensure a thorough and equitable quantitative analysis, we carefully choose 120 identities from the CelebA-HQ dataset [14], using one reference image per identity for consistent comparisons. In addition, we randomly select 20 text prompts that describe various aspects such as clothing, styles, and backgrounds, to provide detailed characterizations of each individual. + +Table 1 shows that optimization-based methods operating in the text embedding space, such as Textual Inversion, + +![](images/7679535e9a5d032a60c9146b1965b86f11f7f009b4f1c0d11fffda2b7922118e.jpg) +a child in a yellow blazer + +![](images/29ba7fdaee49308c8b6c6bd3bafecd4f0a7e3c039fc6a27b25d0be99307a6070.jpg) +Figure 3. Visual comparisons with baselines on different scenarios of T2I generation. Best view by zooming in. + +![](images/c8778bdea6bf672be55962659b90d9396df12b9491a9b8e1a38ae5931b6aece8.jpg) + +![](images/911d6bc129b6ef68e4f496416ab30f83f82a2a5b11a8e9121bfc3c00028040ed.jpg) + +![](images/661f2ae8ff7f4fde176ad05d806f4e5cae81f541e95a9496c789b155c7255b8a.jpg) + +![](images/9c56847d50e6437e44e9daf523fd10f140112d0243288cfa666cb2cfc98739fb.jpg) + +![](images/c397e05de358fab5479fa396cec844417dbf73e569bfa4ec284d70da952453e8.jpg) + +![](images/857a999262335a24dffc9fe831f3727d8a7e685b61d845c17a32331e078873f4.jpg) + +![](images/951ce631c3892e26c6e13f105992b946b4e72dd6469c4f32174dce528a555824.jpg) + +Dreambooth, and Custom Diffusion, fall short in accurately aligning with text prompts and preserving identity features. On the contrary, our method shows comparable performance to FastComposer and IP-Adapter-Face in terms of both CLIP Score and ID metrics. However, it is important to note that IP-Adapter has been fine-tuned on the large-scale LAION2B dataset, whereas our model is trained on the relatively smaller FFHQ-wild and SHHQ datasets, comprising around 110,000 images. Both FastComposer and IP-Adapter directly process the reference image as input, while our approach first maps it to the $\mathcal{W}_{+}$ space before embedding it into the SD model. This additional mapping step could potentially introduce minor discrepancies when integrating the reference image into the SD model, slightly affecting the ID results. Nevertheless, this slight trade-off in ID preservation is balanced by our method's ability to flexibly edit facial attributes while ensuring the background remains consistent. This unique balance of attribute editability and background consistency is a novel contribution that sets our method apart from previous approaches. + +# 4.2. Qualitative Comparison + +Figure 3 depicts visual comparisons with baselines across various scenarios. Since Textual Inversion, Dreambooth, and Custom Diffusion are not specifically designed for facial im + +ages, we select results with the best facial quality from their generated set of 40 images. For the remaining methods, we choose from a set of 10. It is observed that Textual Inversion, Dreambooth, and Custom Diffusion encounter difficulties in capturing and maintaining identity details with a single reference, leading to sub-optimal performance in this task. CelebBasis and FastComposer face challenges in striking a balance between text compatibility and identity preservation. Though IP-Adapter shows improved identity retention, it tends to ignore text conditions in certain instances (see the 1st, 3~5th rows). + +Leveraging our $\mathcal{W}_{+}$ adapter, our approach successfully generates images that are not only compatible with text descriptions but also more effectively retain the target identity. Additionally, our method allows for the editing of facial attributes along the $\Delta w$ direction, causing only minor alterations in the non-facial regions (illustrated in the last column). Furthermore, our approach can be seamlessly adapted to other pre-trained SD models without the need for additional fine-tuning, while retaining its editing capabilities. This versatility is exemplified in the last row of Fig. 3, which showcases our method's effectiveness with the dreamlike-anime model $^2$ . + +![](images/409210800335fcec3ad7daee473a63387a14b0a4459ead664598ac675d35a706.jpg) +Figure 4. Inversion and editing comparisons between e4e and Ours in Stage I. They have the same $w_{+}$ and attributes editing $\Delta w$ . + +Table 2. Quantitative comparisons of face inversion and editing. Editing (ID↓) is the ID metric on results of $w_{+}$ and $w_{+} - 3 \cdot \Delta w$ . + +
MethodsInversionEditing (ID↓)
ID↓FID↓LPIPS↓AgeSmile
e4e [32].43132.26.205.409.312
Ours (Stage I).43531.47.263.393.275
+ +# 4.3. Ablation Study + +Analyses of Aligning $\mathcal{W}_{+}$ in Stage I. We examine if the StyleGAN's $w_{+}$ embedding is well aligned to the latent space of SD during Stage I training. Fig. 4 offers a qualitative comparison of inversion and attribute editing outcomes between e4e [32] and our approach, applied to a real-world image. Both methods use the identical $w_{+}$ embedding generated by the e4e encoder. As depicted in Fig. 4, our method yields inversion results on par with e4e, as seen in the left column, using the same $w_{+}$ vector. Furthermore, the alignment process we implement in this phase preserves the editability inherent to the $\mathcal{W}_{+}$ space. By introducing attribute editing directions $\Delta w$ (such as smile and age) from InterFaceGAN [29], our method can semantically adjust these attributes while maintaining the individual's identity. This is illustrated in the right columns, where the edited embedding, denoted as $w_{+} + \alpha \cdot \Delta w$ , reflects these changes. The parameter $\alpha$ is used to control the extent of attribute modification. + +To quantitatively evaluate the alignment performance, we randomly select 1,000 images from CelebA-HQ [14] and measure their inversion results with metrics of ID, FID, and LPIPS $[43]^3$ . As indicated in the Inversion column of Table 2, our method exhibits comparable performance to e4e in the image inversion task. For assessing attribute editing capabilities, we focus on two key attributes, namely smile, and age, applying a constant scale $\alpha$ (i.e., $w_{+} - 3 \cdot \Delta w$ ). The Editing column in Table 2 demonstrates that our approach surpasses + +Table 3. Quantitative comparisons of different variants. + +
VariantsCLIP Score↑ID↓Detection↑
Ours (PCA).243.526.936
Ours (w/o Lreg).136.461.951
Ours (L0.5reg).192.497.946
Ours (Full).267.516.943
+ +![](images/8f4c3dff5baf3afd4df359744407addc54b2acf4dc6f07fa79f2e897981a5931.jpg) +Figure 5. Visual comparisons of smile attribute editing. The prompt is "a woman wearing a t-shirt on the beach". + +e4e in maintaining identity during the editing process. These findings confirm that our method has successfully aligned the $\mathcal{W}_{+}$ space with the SD model, preserving both the inversion accuracy and the robust editability of attributes. + +Analyses of Variants. We examine different variants to assess each component of our $\mathcal{W}_{+}$ adapter: 1) Ours (PCA), which substitutes our residual cross-attention with parallel cross-attention as in [39], and 2) Ours ( $w/o\mathcal{L}_{reg}$ ), which omits the $\mathcal{L}_{reg}$ objective. The outcomes of these variants are presented in Table 3. It is observed that the residual cross-attention module outperforms the parallel cross-attention design. We conjecture that in tasks focused on integrating local facial imagery into broader, in-the-wild scenarios, directly injecting facial embeddings into the original hidden state $f_{z}$ might adversely impact non-facial regions. Conversely, our residual cross-attention determines fusion weights by calculating attention activations between the text-embedded hidden state $f_{z}^{\prime}$ and our facial embedding. This results in a more precise and effective fusion within the facial region. + +Besides, the results suggest that the regularization loss $\mathcal{L}_{reg}$ is crucial in maintaining semantic consistency between text prompts and images. In the absence of $\mathcal{L}_{reg}$ , the variant Ours $(w / o\mathcal{L}_{reg})$ tends to overemphasize facial regions while neglecting the accompanying text descriptions, leading to a reduced CLIP score. Incorporating $\mathcal{L}_{reg}$ with a weight of $\gamma_{2} = 0.5$ , as in Ours $(\mathcal{L}_{reg}^{0.5})$ , enhances the CLIP score but slightly compromises the ID metric. In comparison, Ours (Full) achieves a satisfactory trade-off in balancing the performance of ID and CLIP score. + +The impact of our disentanglement loss $\mathcal{L}_{disen}$ (Eqn. (6)) is visually demonstrated in Fig. 5. This example shows that omitting $\mathcal{L}_{disen}$ leads to unintended alterations in non-facial regions when the smile editing direction $\Delta w$ is applied to $w_{+}$ . This indicates the importance of $\mathcal{L}_{disen}$ in effectively separating identity-relevant and irrelevant information. By combining our residual cross-attention, regularization loss, and disentanglement loss, Ours (Full) preserves identity and ensures compatibility with text prompts, even amidst changes to the face embedding. + +![](images/87f2a5864478f6deb56b0189abeea0a4ace60c140c31bd97b53529a0583e3027.jpg) +w+ +T2I Generation with $w_{+}$ Interpolation +w + +![](images/66955cd1c0cc40e65c350fde7b75ae383db0c88cbd57ceb57877c68dd562a170.jpg) +Figure 6. Visual results of $w_{+}$ embeddings interpolation during inference. The prompts are "one person wearing suit and tie in a garden" and "one person wearing a blue shirt by a secluded waterfall", respectively. +Figure 7. Visual results of using different $\lambda$ during inference. The prompt is "a man wearing a rainbow shirt in a garden". + +Analyses of $\lambda$ in Cross-attention. The parameter $\lambda$ determines the extent to which the $w_{+}$ vector influences the hidden state $f_{z}^{\prime}$ . Fig. 7 shows the effects of varying $\lambda$ during inference. At $\lambda = 0$ , the outcome aligns with that of the original SD model, unaffected by any editing direction. However, as $\lambda$ increases, the generated identity more closely resembles the reference image. Interestingly, at relatively low values of $\lambda$ (for example, $\lambda = 0.3$ ), even though the identity does not closely match the reference, our model effectively edits attributes in the specified direction. This observation suggests that our method proficiently leverages the $\mathcal{W}_{+}$ space from StyleGAN within the SD framework. + +Analyses of $w_{+}$ Interpolation. We select two facial images and acquire their respective $w_{+}^{1}$ and $w_{+}^{2}$ embeddings from e4e encoder. The interpolated $w_{+}$ embedding is then obtained through $(1 - \kappa) \cdot w_{+}^{1} + \kappa \cdot w_{+}^{2}$ , where $\kappa \in [0,1]$ . As shown in Fig. 6, this T2I generation process results in a smooth transition in the facial regions of the generated + +images, while maintaining a similar and consistent layout throughout the interpolation. This result confirms that the $w_{+}$ embedding in our method not only preserves the editability characteristic of the original StyleGAN but also effectively distinguishes between facial and background regions. + +# 4.4. Limitation + +Our work aims at integrating StyleGAN's editable $\mathcal{W}_{+}$ space into the SD model. We notice a challenge in this integration: the process of converting real images to $w_{+}$ vectors in StyleGAN often leads to a loss of detail, impacting the preservation of identity features. Despite employing a substantial number of training pairs $\{I_f, w_{+}\}$ to establish the mapping, we still observe limitations in maintaining identity fidelity and challenges in editing certain attributes (e.g., pose and glasses). The current framework is designed to generate and edit images with only a single face. Our future work aims to explore the potential of applying localized injections of distinct $w_{+}$ to address multiple human subjects. + +# 5. Conclusion + +We have presented the first attempt to embed the $\mathcal{W}_{+}$ space of StyleGAN into the SD model. We showed that both the mapping network and the residual cross-attention module play crucial roles in facilitating the injection of $w_{+}$ embedding into the SD model, balancing between text prompt influence and identity conditions. Our experiments demonstrate that the $\mathcal{W}_{+}$ space, as used in our approach, not only enables personalized text-to-image generation but also allows for precise editing of facial attributes. We envision this capability being highly beneficial in various practical applications, such as portrait customization with seamless attribute modifications. Furthermore, our methodology holds the potential for application across other object domains that possess distinct prior spaces. + +Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: How to embed images into the StyleGAN latent space? In ICCV, 2019. 2 +[2] Yufei Cai, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Hu Han, and Wangmeng Zuo. Decoupled textual embeddings for customized image generation. In AAAI, 2024. 2 +[3] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. DisenBooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2 +[4] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, and Min Zheng. PhotoVerse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2 +[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5 +[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 3 +[7] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. StyleGAN-Human: A data-centric odyssey of human generation. In ECCV, 2022. 5 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 2, 5 +[9] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2 +[10] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In NeurIPS, 2020. 2 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS, 2022. 5 +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 3 +[13] Xuhui Jia, Yang Zhao, Kelvin C. K. Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2 +[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7 +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 3, 5 +[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 3, 5 + +[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 5 +[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4, 5 +[19] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image superresolution. In CVPR, 2023. 2 +[20] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. A survey on leveraging pre-trained generative adversarial networks for image editing and restoration. arXiv preprint arXiv:2207.10309, 2022. 2 +[21] Yang Liu, Cheng Yu, Lei Shang, Ziheng Wu, Xingjun Wang, Yuze Zhao, Lin Zhu, Chen Cheng, Weitao Chen, Chao Xu, Haoyu Xie, Yuan Yao, Wenmeng Zhou, Yingda Chen, Xuansong Xie, and Baigui Sun. FaceChain: A playground for identity-preserving portrait generation. arXiv preprint arXiv:2308.14256, 2023. 3 +[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5 +[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3 +[24] Hamza Pehlivan, Yusuf Dalva, and Aysegul Dundar. StyleRes: Transforming the residuals for real image editing with StyleGAN. In CVPR, 2022. 2 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Finetuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 5 +[29] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2, 7 +[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. InstantBooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2 +[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 5 +[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. In ACM TOG, 2021. 2, 3, 5, 7 + +[33] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3 +[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 3 +[35] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.5 +[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 4 +[37] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. SingleInsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 2 +[38] Guangxuan Xiao, Tianwei Yin, William T. Freeman, Frédo Durand, and Song Han. FastComposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 5 +[39] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 4, 5, 7 +[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. In NeurIPS, 2023. 2, 3, 5 +[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. In IEEE signal processing letters, 2016. 5 +[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3 +[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 +[44] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In ECCV, 2020. 5 \ No newline at end of file diff --git a/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/images.zip b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c0f307816f8c96722cbc9b362ce48f7b256debc --- /dev/null +++ b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80048112aa4d001932612350770da1e77db92e604fa110e3be9495a05db0672d +size 970307 diff --git a/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/layout.json b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11f6b0ee1bac4a44eebbd45e7379dda6c83fa80c --- /dev/null +++ b/2024/When StyleGAN Meets Stable Diffusion_ a W+ Adapter for Personalized Image Generation/layout.json @@ -0,0 +1,10272 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 94, + 451, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 94, + 451, + 129 + ], + "spans": [ + { + "bbox": [ + 141, + 94, + 451, + 129 + ], + "type": "text", + "content": "When StyleGAN Meets Stable Diffusion: a " + }, + { + "bbox": [ + 141, + 94, + 451, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 141, + 94, + 451, + 129 + ], + "type": "text", + "content": "Adapter for Personalized Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 152, + 423, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 152, + 423, + 179 + ], + "spans": [ + { + "bbox": [ + 169, + 152, + 423, + 179 + ], + "type": "text", + "content": "Xiaoming Li Xinyu Hou Chen Change Loy S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 456, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 456, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 456, + 194 + ], + "type": "text", + "content": "csxml@gmail.com xinyu.hou@ntu.edu.sg ccloy@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 73, + 204, + 521, + 469 + ], + "blocks": [ + { + "bbox": [ + 73, + 204, + 521, + 469 + ], + "lines": [ + { + "bbox": [ + 73, + 204, + 521, + 469 + ], + "spans": [ + { + "bbox": [ + 73, + 204, + 521, + 469 + ], + "type": "image", + "image_path": "8b78a124102bb333611629ac7ec136ca867faa6d301c6dce65d5a141378ccdf3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "type": "text", + "content": "Figure 1. Given a single reference image (thumbnail in the top left), our " + }, + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "type": "text", + "content": "adapter not only integrates the identity into the text-to-image generation accurately but also enables modifications of facial attributes along the " + }, + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 45, + 472, + 547, + 506 + ], + "type": "text", + "content": " trajectory derived from StyleGAN. The text prompt is \"a woman wearing a spacesuit in a forest\"." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "spans": [ + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 545, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 545, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 545, + 290, + 715 + ], + "type": "text", + "content": "Text-to-image diffusion models have remarkably excelled in producing diverse, high-quality, and photo-realistic images. This advancement has spurred a growing interest in incorporating specific identities into generated content. Most current methods employ an inversion approach to embed a target visual concept into the text embedding space using a single reference image. However, the newly synthesized faces either closely resemble the reference image in terms of facial attributes, such as expression, or exhibit a reduced capacity for identity preservation. Text descriptions intended to guide the facial attributes of the synthesized face may fall short, owing to the intricate entanglement of identity information with identity-irrelevant facial attributes derived from the reference image. To address these issues," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "content": "we present the novel use of the extended StyleGAN embedding space " + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "content": ", to achieve enhanced identity preservation and disentanglement for diffusion models. By aligning this semantically meaningful human face latent space with text-to-image diffusion models, we succeed in maintaining high fidelity in identity preservation, coupled with the capacity for semantic editing. Additionally, we propose new training objectives to balance the influences of both prompt and identity conditions, ensuring that the identity-irrelevant background remains negligibly affected during facial attribute modifications. Extensive experiments reveal that our method adeptly generates personalized text-to-image outputs that are not only compatible with prompt descriptions but also amenable to common StyleGAN editing directions in diverse settings. Our code and model are available at https://github.com/csxmli2016/w-plus-adapter." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2187" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 287, + 247 + ], + "type": "text", + "content": "What if we could be the protagonists of our own fantasies? This paper primarily addresses personalized text-to-image (T2I) generation, a field that is attracting growing attention due to users' desire to craft their unique content. Overlapped with customized T2I [8, 17, 28, 36], which integrates various visual concepts like human faces and objects into the generated imagery, our focus is on tailoring image generation to specific identities. This focus is motivated by applications such as storyboarding, where a consistent identity needs to be maintained across all images, despite variations in expressions or ages. Additionally, we believe that human faces possess unique, fine-grained intrinsic features that merit special attention and present a worthwhile area for exploration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 247, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 247, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 287, + 498 + ], + "type": "text", + "content": "Recent personalized image generation approaches utilize a small reference set of a target identity and embed it into a specific space. The common choice of embedding space has been the textual embedding space used by large language models (LLMs). While existing methods based on textual embedding [8, 28, 36] have proven capable of maintaining target identity, they are often limited by inherent trade-offs. Specifically, these methods face challenges in simultaneously preserving identity, generating varied facial attributes, and creating identity-irrelevant content that aligns with the text description. We observe that these issues predominantly stem from the entangled nature of the textual embedding space, where a single pseudo word " + }, + { + "bbox": [ + 49, + 247, + 287, + 498 + ], + "type": "inline_equation", + "content": "S^*" + }, + { + "bbox": [ + 49, + 247, + 287, + 498 + ], + "type": "text", + "content": " struggles to distinctly isolate identity-related features from the reference image. Efforts to separate such information include approaches like encoding a visual concept into multiple word embeddings [36] or employing a separate branch for extracting identity-irrelevant details [3]. However, these methods tend to overlook the nuanced facial features critical to an individual's identity, resulting in incomplete disentanglement and sub-optimal identity preservation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "text", + "content": "In this study, we aim to more effectively separate identity-relevant and -agnostic features for better identity preservation, while also ensuring editability. To achieve this, we propose inverting the visual concept of a target identity into StyleGAN's [15] " + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "text", + "content": " latent space, as opposed to using the textual embedding space. In particular, we introduce a mapping network to integrate the " + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "text", + "content": " space with the diffusion model and a residual cross-attention module to add the " + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "text", + "content": " vector as an additional identity condition. This mapping network, once trained with (image, " + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 49, + 498, + 287, + 713 + ], + "type": "text", + "content": ") pairs, can generalize to unseen individuals during inference without the need for a separate identity-specific model for each person. We further present novel regularized training to ensure that edits have a negligible effect on identity-irrelevant regions and that the overall generation remains aligned with the prompt conditions. It is noteworthy that although CelebBasis [40] explores a similar concept of creating a face-specific latent space through PCA of selected celebrity name embeddings," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 132 + ], + "type": "text", + "content": "we argue that relying on such a limited dataset of celebrity names is insufficient for a comprehensive face latent space due to potential under-representation. Our experimental results confirm that our approach exhibits a superior capacity for identity preservation compared to CelebBasis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "content": "The contributions are summarised as follows: 1) We introduce extended StyleGAN " + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "content": " space as a target inversion latent space to better encode the identity-preserving facial concepts for personalized text-to-image synthesis. This is the first study that considers the fusion between StyleGAN " + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "content": " space and diffusion-based image generation. 2) Embedding target identity in " + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 308, + 133, + 545, + 275 + ], + "type": "text", + "content": " space enables smooth and controllable semantic editing on facial attributes in our text-to-image model. 3) The effective disentanglement of identity-relevant and -irrelevant information in our model facilitates an identity-preserving generation that is not only diverse but also adaptable to a wide range of prompt instructions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 286, + 391, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 391, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 391, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "content": "StyleGAN Latent Space. GANs have drawn enormous attention due to the well-disentangled latent space [20]. StyleGAN [15] proposes to map input latent code to an intermediate latent space " + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "content": " to prevent warping of the training data distribution to fit in a particular probability distribution. The resulting " + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "content": " space is a semantically disentangled space that allows fine-grained controls over image synthesis. InterFaceGAN [29] and GANSpace [10] further interpret the latent space and identify several control directions such as age, gender, face angle, smile, etc. The powerful semantic editing ability has motivated research on inverting and editing real images in the StyleGAN space [1, 19, 24, 32]. Abdal et al. [1] utilize a direct optimization framework to embed a given real image to the extended StyleGAN space. Tov et al. [32] deploy an encoder to perform the inversion and keep the concatenated latent codes close to the original StyleGAN space to maintain high perceptual quality and editability. Motivated by the disentanglement and editability of StyleGAN latent space, we aim to introduce it to the T2I diffusion models to achieve more controllable synthesis. We follow Image2StyleGAN to denote the extended " + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "content": " space represented by multiple nonidentical " + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "content": " vectors as " + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 308, + 307, + 545, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 712 + ], + "type": "text", + "content": "Personalized Image Synthesis. Customization has been extensively studied in the context of T2I to generate images of specified objects or individuals [2, 8, 28]. Gal et al. [8] directly apply learnable text embedding optimization. Ruiz et al. [28] fine-tune the diffusion model for the target concept while preserving its class prior of the concept. Further improvements have been made to invert multiple concepts simultaneously with cross-attention fine-tuning [17], and disentangle background irrelevant information [3]. Encoder-based approaches [4, 9, 13, 30, 36–38] are employed for their efficiency. The target visual concept is encoded to embedding space as additional conditions. SingleInsert [37]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "2188" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "adopts a two-stage scheme to insert concepts from a single image into the foreground region exclusively and enables the editing of the concept by text prompts. However, it is worth noticing that our method achieves more fine-grained editing with a single reference image at inference and does not necessitate identity-specific fine-tuned models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 290, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 290, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 290, + 323 + ], + "type": "text", + "content": "With special attention on identity inversion utilizing the distinct features of human faces, Yuan et al. [40] define a celebrity space by applying PCA on selected celebrity name embeddings from CLIP and embedding new identities into the space via learnable coefficients. Valevski et al. [33] utilize a pre-trained face recognizer as the encoder and further project the face embedding to CLIP space. FaceChain [21] trains separate face and style LoRAs [12] to synthesize specific faces in specific styles. They follow a data preprocessing pipeline and apply face fusion of the best reference face on the synthesized results. However, we observe that previous personalization approaches inherit a common limitation that the synthesized faces share similar attributes as the reference, indicating that the model cannot disentangle irrelevant information well and overfits the particular reference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 468 + ], + "type": "text", + "content": "Diffusion Model Adapters. Light-weight adapters have been adopted to avoid the laborious work of fine-tuning large models and add additional controls [23, 39, 42] to the diffusion model. ControlNet [42] proposes to train task-specific adapters while freezing the original diffusion model to adapt to various input conditions. Concurrent work T2I-Adapter [23] constructs a lighter-weight adapter for controls on structure and color. Ye et al. [39] employ a decoupled cross-attention module to consider both text and image prompts in the denoising process. In our work, we consider identity information as an additional condition and align it with the T2I model for identity-preserving synthesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 477, + 130, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 130, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 130, + 491 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "text", + "content": "Our approach is capable of generating images that preserve identity while allowing for semantic edits, requiring just a single reference image for inference. This capability is realized by innovatively aligning StyleGAN's " + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "text", + "content": " latent space with the diffusion model. The training of our " + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "text", + "content": "adapter is divided into two stages. In Stage I (Sec. 3.2), we establish a mapping from " + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 497, + 288, + 641 + ], + "type": "text", + "content": " to SD latent space, using the resulting projection as an additional identity condition to synthesize center-aligned facial images of a specified identity. In Stage II (Sec. 3.3), this personalized generation process is expanded to accommodate more dynamic, \"in-the-wild\" settings, ensuring adaptability to a variety of textual prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In our implementation, we use Stable Diffusion [26] as the foundational T2I model, which is one of the most commonly used latent space-based text-to-image generation model [6, 34]. During the training phase, the model archi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "tecture includes: 1) a pre-trained encoder " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": ", which transforms an image " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " into a latent representation with reduced dimensionality, 2) a conditional diffusion model tasked with predicting the latent code of a previous timestep, based on the text condition " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " and the current timestep's latent code, and 3) a pre-trained decoder " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " that reconstructs the final latent code into the synthesized image. The learning objective of this model is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 171, + 547, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 171, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 318, + 171, + 547, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L D M} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, c _ {t x t}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} (c _ {t x t})\\right) \\| _ {2} ^ {2} \\right] \\tag {1}", + "image_path": "ddeff7c6f0180138ecf72afe0e89815d65e9e24156e51ed328d91b180df424ed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(I)" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " is the ground truth noise added for current timestep " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " is the latent code of timestep " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\tau_{txt}" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " denotes the pre-trained CLIP text encoder [25] that converts the text prompt " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " to textual embeddings. The " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " represents the UNet [27] denosing network. In the inference stage, the denoising network is applied iteratively to denoise a random sampled noise " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": " with condition " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": ". The final result is then generated through the pre-trained decoder, i.e., " + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(z_0)" + }, + { + "bbox": [ + 304, + 189, + 547, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 300, + 529, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 529, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 529, + 313 + ], + "type": "text", + "content": "3.2. Stage I: Aligning " + }, + { + "bbox": [ + 305, + 300, + 529, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 305, + 300, + 529, + 313 + ], + "type": "text", + "content": " with Stable Diffusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "type": "text", + "content": "The goal of Stage I is to align " + }, + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "type": "text", + "content": " space with SD to carry identity information in T2I generation while retaining its editing ability. To do so, we train a mapping network to project a " + }, + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 322, + 547, + 406 + ], + "type": "text", + "content": " embedding to SD latent space and inject this extra condition into SD by modifying the cross-attention module. The framework of Stage I is illustrated in Fig. 2. Details are introduced below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "content": "Training Pair Construction. We adopt 100K discrete training samples to fit the continuous distribution of " + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "content": " and align it with SD. For each face image " + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "content": ", we use the pretrained e4e [32] encoder from StyleGAN inversion task to get its corresponding " + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "inline_equation", + "content": "w_{+} \\in \\mathbb{R}^{18 \\times 512}" + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "content": " vector. The pairs of " + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "inline_equation", + "content": "\\{I_{f}, w_{+}\\}" + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "content": " constitute our training data in this stage. Specifically, to generalize on real-world face images and improve " + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 408, + 548, + 552 + ], + "type": "text", + "content": " diversity, two types of training pairs are built, i.e., synthetic face images from StyleGAN2 [16] and real-world face images from FFHQ [15]. Note that in Stage I, we only consider the aligned face images in order to exclude any extraneous influences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": "Mapping Network. Since the original " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": " space is designated for StyleGAN generation, we project the vector " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "w_{+} \\in \\mathbb{R}^{18 \\times 512}" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": " to four tokens of dimension 768 (" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{4 \\times 768}" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": ") to align with the input dimension of SD condition. The mapping network consists of four groups of linear layers and is denoted as " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w(\\cdot)" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": ". The latent space after mapping inherits the editability and disentanglement properties of " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": " space and is compatible with the SD generation process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Residual Cross-Attention. To incorporate the identity condition from the projected " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " embedding into the pre-trained SD model, we introduce a residual cross-attention module. In the standard SD framework, the output of cross-attention is determined using query features " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{z}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " from the hidden state" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2189" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 57, + 239, + 194 + ], + "blocks": [ + { + "bbox": [ + 53, + 57, + 239, + 194 + ], + "lines": [ + { + "bbox": [ + 53, + 57, + 239, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 57, + 239, + 194 + ], + "type": "image", + "image_path": "854400d374ea713f011c018896c36e324d5d9658c49165eab8ecf6d4c31f7c8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 198, + 232, + 209 + ], + "lines": [ + { + "bbox": [ + 69, + 198, + 232, + 209 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 232, + 209 + ], + "type": "text", + "content": "(a) Stage I: Aligning " + }, + { + "bbox": [ + 69, + 198, + 232, + 209 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 69, + 198, + 232, + 209 + ], + "type": "text", + "content": " with Stable Diffusion" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 57, + 444, + 194 + ], + "blocks": [ + { + "bbox": [ + 244, + 57, + 444, + 194 + ], + "lines": [ + { + "bbox": [ + 244, + 57, + 444, + 194 + ], + "spans": [ + { + "bbox": [ + 244, + 57, + 444, + 194 + ], + "type": "image", + "image_path": "50f804bf3afe9299f74c7c80640d75167ad106d37009de8fcc2a88873b170f74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 198, + 406, + 209 + ], + "lines": [ + { + "bbox": [ + 285, + 198, + 406, + 209 + ], + "spans": [ + { + "bbox": [ + 285, + 198, + 406, + 209 + ], + "type": "text", + "content": "(b) Stage II: Generation in the Wild" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "text", + "content": "Figure 2. Overview of " + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "text", + "content": " adapter training stages. Left: Stage I for aligning " + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "text", + "content": " space with Stable Diffusion. Middle: Stage II for generating in-the-wild images with " + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 213, + 547, + 236 + ], + "type": "text", + "content": " embeddings. Right: details of residual cross-attention module. Open lock indicates trainable parts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 450, + 57, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 450, + 57, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 450, + 57, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 450, + 57, + 541, + 194 + ], + "type": "image", + "image_path": "eea6e23c46f92d07b8307cbd5ec6e18182f38192579f3e0a15080db4b67e1bec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 198, + 536, + 209 + ], + "lines": [ + { + "bbox": [ + 446, + 198, + 536, + 209 + ], + "spans": [ + { + "bbox": [ + 446, + 198, + 536, + 209 + ], + "type": "text", + "content": "(c) Residual Cross-Attention" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 244, + 277, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 277, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 277, + 257 + ], + "type": "text", + "content": "and the text condition " + }, + { + "bbox": [ + 46, + 244, + 277, + 257 + ], + "type": "inline_equation", + "content": "\\tau_{txt}(c_{txt})" + }, + { + "bbox": [ + 46, + 244, + 277, + 257 + ], + "type": "text", + "content": ". The process is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 260, + 287, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 260, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 67, + 260, + 287, + 285 + ], + "type": "interline_equation", + "content": "f _ {z} ^ {\\prime} = \\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} \\tag {2}", + "image_path": "7893a5caaa0aa3426824e1eb28d4595f30639c4416c9c3476bba2e88989f551d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = f_z\\mathbf{W}_q" + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\tau_{txt}(c_{txt})\\mathbf{W}_k" + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\tau_{txt}(c_{txt})\\mathbf{W}_v" + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": " are the query, key, value matrices of the attention module, respectively. " + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 46, + 288, + 289, + 384 + ], + "type": "text", + "content": " denote the corresponding projection matrices. The cross-attention module is where the latent noise interacts with text condition embeddings inside the denoising process. Following previous methods [36, 39], we add an additional condition by a separate cross-attention module. Our residual cross-attention is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 387, + 287, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 387, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 92, + 387, + 287, + 401 + ], + "type": "interline_equation", + "content": "f _ {z} ^ {\\prime \\prime} = f _ {z} ^ {\\prime} + \\lambda \\cdot \\operatorname {A t t e n t i o n} \\left(\\mathbf {Q} ^ {\\dagger}, \\mathbf {K} ^ {\\dagger}, \\mathbf {V} ^ {\\dagger}\\right) \\tag {3}", + "image_path": "0fe8ccf545b7d8f8deb32c20491eff60ec12c97a6ac669fd74f120ff08de25d8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\dagger} = f_{z}^{\\prime}\\mathbf{W}_{q}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\dagger} = f_{w}\\mathbf{W}_{k}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\dagger} = f_{w}\\mathbf{W}_{v}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{q}^{\\dagger},\\mathbf{W}_{k}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{v}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " are the projection matrices for query, key, and value, respectively. The " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "f_{w} = \\mathcal{F}_{w}(w_{+})" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " is mapped from " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " by the mapping network. We use " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " as a scale parameter to balance the influence of text and identity conditions on the generation. We incorporate the decoupled cross-attention module in a residual fashion instead of the parallel approach as in previous works [36, 39] to avoid performance degradation on the original text condition. When " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " is set to 0, our " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " vector has no impact on the pre-trained SD model. We set " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " to 1 during training. The " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\dagger}" + }, + { + "bbox": [ + 46, + 403, + 289, + 575 + ], + "type": "text", + "content": " are initialized from their corresponding Q, K, and V, respectively. The residual cross-attention module is performed on all the cross-attention layers of SD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "type": "text", + "content": "Learning Objectives. The trainable parts in this stage include the mapping network " + }, + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}^\\dagger, \\mathbf{K}^\\dagger, \\mathbf{V}^\\dagger\\}" + }, + { + "bbox": [ + 46, + 578, + 288, + 626 + ], + "type": "text", + "content": " matrices in all residual cross-attention modules. The optimization target is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 287, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 287, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w _ {+}} = \\mathbb {E} _ {z _ {0}, \\epsilon , t, c _ {t x t}, w _ {+}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\mathcal {F} _ {w} (w _ {+})) \\| _ {2} ^ {2} \\right] \\tag {4}", + "image_path": "d10002c46a56ae0c86ecfb7f0aab60e7e201f3a6334a63b00b1b03d67cb4acf9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Eqn. (4) is similar to Eqn. (1) except for the additional " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w(w_+)" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " serving as identity condition and cross-attention structure modified inside " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": ". The text prompt " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " is randomly selected from several neutral templates describing a human face, e.g., \"a photo of a face\", or \"a face\" (see suppl.)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "content": "Through mapping the " + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "content": " distribution into an embedding space compatible with SD, our approach effectively uses the " + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "content": " vector to condition the personalized generation of aligned facial images. Furthermore, attribute directions " + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "content": " derived from the original " + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "content": " space, encompassing features like smile, age, eye-opening, and others, remain applicable to our projected " + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 244, + 547, + 340 + ], + "type": "text", + "content": " vector. This allows for personalized, fine-grained face editing, as demonstrated in Fig. 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 349, + 480, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 349, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 480, + 361 + ], + "type": "text", + "content": "3.3. Stage II: Generation in the Wild" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "text", + "content": "In order to refine the performance of our " + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "text", + "content": "-adapted SD model from Stage I for scenarios beyond controlled environments, we continue on a second stage of training. This phase is dedicated to further fine-tuning the weights of " + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\dagger},\\mathbf{K}^{\\dagger},\\mathbf{V}^{\\dagger}" + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "text", + "content": " across all residual cross-attention modules. To ensure that the projected " + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 368, + 547, + 487 + ], + "type": "text", + "content": " space exclusively encapsulates identity-related facial attributes, while remaining uninfluenced by irrelevant distractions from the background, we keep the mapping network fixed during this stage. The architecture and workflow of Stage II are depicted in the middle of Fig. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": "Training Data Construction. For Stage II, in-the-wild images are used for training. For each image " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": ", its corresponding text caption, " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": ", is extracted using an off-the-shelf captioning tool [18]. The aligned face image " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": " is cropped from " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "w_+" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": " is obtained from " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": " by the e4e encoder (same as Stage I). A face region mask, " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": ", is also obtained from the image " + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 488, + 547, + 583 + ], + "type": "text", + "content": ", with 0 denoting the face region and 1 representing the non-face region." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 584, + 547, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 547, + 620 + ], + "type": "text", + "content": "Learning Objectives. To keep high identity fidelity while encouraging high diversity of identity-irrelevant context, three losses are employed jointly in this stage." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 621, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 644 + ], + "type": "text", + "content": "First, the reconstruction loss " + }, + { + "bbox": [ + 304, + 621, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 304, + 621, + 545, + 644 + ], + "type": "text", + "content": " is used to guide the denoising process:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 647, + 545, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 647, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 311, + 647, + 545, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\mathbb {E} _ {z _ {0}, \\epsilon , t, c _ {t x t}, w _ {+}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\mathcal {F} _ {w} (w _ {+})) \\| _ {2} ^ {2} \\right] \\tag {5}", + "image_path": "f09a74380687c851b5407d874d6603e3f3720e94b1450d328526378e01896e21.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Note that Eqn. (5) is different from Eqn. (4) for that: 1) " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " here is encoded from the in-the-wild image " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " rather than the face image " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", so the reconstruction goal differs, and 2) " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " is the caption describing the in-the-wild image rather than a" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2190" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "simple neutral prompt, therefore containing more information about the image context to facilitate reconstruction of " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "text", + "content": "Second, when semantic edits " + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "text", + "content": " are applied in the " + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "text", + "content": " space, the objective is to exclusively modify the facial attributes, leaving the remaining regions unchanged. In order to accomplish this, we propose a " + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "text", + "content": " disentanglement loss to limit the editability of " + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 108, + 287, + 168 + ], + "type": "text", + "content": " outside the face region:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 172, + 287, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 172, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 79, + 172, + 287, + 198 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i s e n}} = \\| M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right), \\mathcal {F} _ {w} \\left(w _ {+}\\right)\\right) - \\tag {6} \\\\ M \\cdot \\epsilon_ {\\theta} (z _ {t}, t, \\tau_ {t x t} (c _ {t x t}), \\Psi (\\mathcal {F} _ {w} (w _ {+}))) \\| \\\\ \\end{array}", + "image_path": "d0199c70c312321afffe0c628823e4a49be29575cf52ef6c3f3f1b92f9b423a0.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "content": " is the augmentation operation. During training, three augmentation strategies are adopted: 1) shuffle along the batch dimension of " + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w(w_+)" + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "content": ", 2) add random perturbations of Gaussian noise on " + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w(w_+)" + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "content": ", and 3) combine both of them. By applying such constraints, augmented (or edited) " + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "inline_equation", + "content": "w_+" + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "content": " vectors are forced to have similar non-face regions, thereby disentangling the effect of " + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 46, + 201, + 288, + 285 + ], + "type": "text", + "content": " on the background." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "content": "Finally, since the identity condition " + }, + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w(w_+)" + }, + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "content": " should only influence the face region, which constitutes a relatively small proportion of the final output, we constrain its impact to be limited by a regularization loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 338, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\left\\| M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right), \\mathcal {F} _ {w} \\left(w _ {+}\\right)\\right) - M \\cdot \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {t x t} \\left(c _ {t x t}\\right) \\right. \\right\\| \\tag {7}", + "image_path": "2c5c0ebbd63d5ec149db0f4a81b90c0fd574a74ba52163ef5c25ab87e40d1e56.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "type": "text", + "content": "In this way, a more balanced effect of text and identity condition on the synthesized result is achieved, further mitigating the risk of unwanted noisy information introduced from " + }, + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "type": "text", + "content": " and preserving the compatibility with text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 403, + 280, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 403, + 280, + 414 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 280, + 414 + ], + "type": "text", + "content": "The overall learning objective in Stage II is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 419, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 419, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 102, + 419, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\gamma_ {1} \\cdot \\mathcal {L} _ {\\text {d i s e n}} + \\gamma_ {2} \\cdot \\mathcal {L} _ {\\text {r e g}} \\tag {8}", + "image_path": "eb049648fb8546f45012190a68a769dc86c6ffe7763fc5e87af5844aab8ab01c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "type": "inline_equation", + "content": "\\gamma_{1}" + }, + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "type": "inline_equation", + "content": "\\gamma_{2}" + }, + { + "bbox": [ + 46, + 435, + 244, + 447 + ], + "type": "text", + "content": " denote the trade-off parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 454, + 127, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 127, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 127, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "content": "Training Data. In Stage I, we use the pre-trained e4e encoder [32] to obtain the " + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "content": " vector for each face image " + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "content": " from FFHQ [15] and StyleGAN2 [16]. FFHQ dataset contains 70,000 images, among which 600 are split for validation while the rest are for training. We also synthesize 70,000 images using StyleGAN2. In Stage II, we use FFHQ in-the-wild images (excluding those used for validation in Stage I) and SHHQ [7] to optimize the " + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "content": "adapter. SHHQ is a human dataset that contains 40,000 high-quality full-body images. BLIP2 [18] is introduced to generate the caption for each in-the-wild image. Face region mask " + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "content": " is obtained based on the FFHQ alignment operation and is eroded and blurred with kernel sizes of 32 and 7, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Implementation Details. All training is conducted on a server with 8 Tesla V100 GPUs. The batch size is set to 16. We employ the AdamW optimizer [22] with a learning rate of " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " and weight decay of 0.01. In Stage II, we adopt color jittering [44], random rotation, and sampling for inthe-wild images to increase diversity. During training, the cropped images with incomplete or small faces are discarded." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 307, + 96, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with previous methods. The best and second best results are highlighted by bold and underline." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 96, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 188 + ], + "type": "table", + "html": "
VariantsCLIP Score↑ID↓Detection↑
Textual Inversion [8].194.575.891
Dreambooth [28].177.562.800
Custom Diffusion [17].216.498.850
FastComposer [38].265.419.950
IP-Adapter-Face [39].241.407.958
CelebBasis [40].253.448.916
Ours.267.418.950
", + "image_path": "a2b880e02dc16d50603ef6ec7ae84b3c49c0cfb63cb70dd10f0df13841c3cd06.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": "It takes around 24 hours to align the " + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": " space with SD in Stage I, and nearly 96 hours to train the " + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": " adapter in Stage II. In the inference, we adopt DDIM sampler [31] with 50 steps. To enable classifier-free guidance [11], we use the default settings and set the guidance scale to 7.5. We follow IP-Adapter [39] to randomly drop the text " + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": " vector with a probability of 0.05. " + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\gamma_{1}" + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\gamma_{2}" + }, + { + "bbox": [ + 304, + 198, + 545, + 318 + ], + "type": "text", + "content": " in Eqn. (8) are set to 1.5 and 1.0, respectively. The whole framework is implemented on Diffusers [35]. Our experiments are based on SD v1.5 and can be extended to other SD models (see suppl.)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 327, + 448, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 448, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 448, + 340 + ], + "type": "text", + "content": "4.1. Quantitative Comparison" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 346, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 514 + ], + "type": "text", + "content": "Baselines. The methods we evaluate can be categorized into two distinct groups: general object customization and specific face personalization. For general object customization, we use Dreambooth [28], Textual Inversion [8], and Custom Diffusion [17], implementing these using Diffusers [35]. For specific face personalization, we select closely related works such as FastComposer [38], IP-Adapter-Face [39], and CelebBasis [40]. In our evaluation, we adhere to the settings established by CelebBasis [40] and FastComposer [38]. The metrics used for assessment include the CLIP Score [25], which is calculated based on the average image-text similarity using CLIP-L/14, the Face Detection Score determined using MTCNN [41], and the Identity Distance (ID) measured using ArcFace [5] on the detected facial regions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 514, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 668 + ], + "type": "text", + "content": "Given the extensive fine-tuning requirements for each identity in methods like Dreambooth, Textual Inversion, Custom Diffusion, and CelebBasis – which can be notably time-consuming (for instance, Textual Inversion demands about an hour on a single V100 GPU server" + }, + { + "bbox": [ + 304, + 514, + 547, + 668 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 514, + 547, + 668 + ], + "type": "text", + "content": ") – we encounter constraints in evaluating a substantial number of test instances. To ensure a thorough and equitable quantitative analysis, we carefully choose 120 identities from the CelebA-HQ dataset [14], using one reference image per identity for consistent comparisons. In addition, we randomly select 20 text prompts that describe various aspects such as clothing, styles, and backgrounds, to provide detailed characterizations of each individual." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 670, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 547, + 693 + ], + "type": "text", + "content": "Table 1 shows that optimization-based methods operating in the text embedding space, such as Textual Inversion," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 515, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 515, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 515, + 712 + ], + "type": "text", + "content": "1https://huggingface.co/docs/diffusers/training/text_inversion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "2191" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 61, + 107, + 386 + ], + "blocks": [ + { + "bbox": [ + 53, + 61, + 107, + 386 + ], + "lines": [ + { + "bbox": [ + 53, + 61, + 107, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 61, + 107, + 386 + ], + "type": "image", + "image_path": "7679535e9a5d032a60c9146b1965b86f11f7f009b4f1c0d11fffda2b7922118e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 73, + 116, + 103 + ], + "lines": [ + { + "bbox": [ + 107, + 73, + 116, + 103 + ], + "spans": [ + { + "bbox": [ + 107, + 73, + 116, + 103 + ], + "type": "text", + "content": "a child in a yellow blazer" + } + ] + } + ], + "index": 1, + "angle": 270, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 117, + 62, + 170, + 387 + ], + "blocks": [ + { + "bbox": [ + 117, + 62, + 170, + 387 + ], + "lines": [ + { + "bbox": [ + 117, + 62, + 170, + 387 + ], + "spans": [ + { + "bbox": [ + 117, + 62, + 170, + 387 + ], + "type": "image", + "image_path": "29ba7fdaee49308c8b6c6bd3bafecd4f0a7e3c039fc6a27b25d0be99307a6070.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 392, + 499, + 403 + ], + "lines": [ + { + "bbox": [ + 94, + 392, + 499, + 403 + ], + "spans": [ + { + "bbox": [ + 94, + 392, + 499, + 403 + ], + "type": "text", + "content": "Figure 3. Visual comparisons with baselines on different scenarios of T2I generation. Best view by zooming in." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 62, + 222, + 387 + ], + "blocks": [ + { + "bbox": [ + 171, + 62, + 222, + 387 + ], + "lines": [ + { + "bbox": [ + 171, + 62, + 222, + 387 + ], + "spans": [ + { + "bbox": [ + 171, + 62, + 222, + 387 + ], + "type": "image", + "image_path": "c8778bdea6bf672be55962659b90d9396df12b9491a9b8e1a38ae5931b6aece8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 62, + 276, + 387 + ], + "blocks": [ + { + "bbox": [ + 223, + 62, + 276, + 387 + ], + "lines": [ + { + "bbox": [ + 223, + 62, + 276, + 387 + ], + "spans": [ + { + "bbox": [ + 223, + 62, + 276, + 387 + ], + "type": "image", + "image_path": "911d6bc129b6ef68e4f496416ab30f83f82a2a5b11a8e9121bfc3c00028040ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 62, + 328, + 387 + ], + "blocks": [ + { + "bbox": [ + 276, + 62, + 328, + 387 + ], + "lines": [ + { + "bbox": [ + 276, + 62, + 328, + 387 + ], + "spans": [ + { + "bbox": [ + 276, + 62, + 328, + 387 + ], + "type": "image", + "image_path": "661f2ae8ff7f4fde176ad05d806f4e5cae81f541e95a9496c789b155c7255b8a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 329, + 62, + 383, + 387 + ], + "blocks": [ + { + "bbox": [ + 329, + 62, + 383, + 387 + ], + "lines": [ + { + "bbox": [ + 329, + 62, + 383, + 387 + ], + "spans": [ + { + "bbox": [ + 329, + 62, + 383, + 387 + ], + "type": "image", + "image_path": "9c56847d50e6437e44e9daf523fd10f140112d0243288cfa666cb2cfc98739fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 62, + 434, + 387 + ], + "blocks": [ + { + "bbox": [ + 383, + 62, + 434, + 387 + ], + "lines": [ + { + "bbox": [ + 383, + 62, + 434, + 387 + ], + "spans": [ + { + "bbox": [ + 383, + 62, + 434, + 387 + ], + "type": "image", + "image_path": "c397e05de358fab5479fa396cec844417dbf73e569bfa4ec284d70da952453e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 435, + 62, + 487, + 387 + ], + "blocks": [ + { + "bbox": [ + 435, + 62, + 487, + 387 + ], + "lines": [ + { + "bbox": [ + 435, + 62, + 487, + 387 + ], + "spans": [ + { + "bbox": [ + 435, + 62, + 487, + 387 + ], + "type": "image", + "image_path": "857a999262335a24dffc9fe831f3727d8a7e685b61d845c17a32331e078873f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 488, + 62, + 539, + 387 + ], + "blocks": [ + { + "bbox": [ + 488, + 62, + 539, + 387 + ], + "lines": [ + { + "bbox": [ + 488, + 62, + 539, + 387 + ], + "spans": [ + { + "bbox": [ + 488, + 62, + 539, + 387 + ], + "type": "image", + "image_path": "951ce631c3892e26c6e13f105992b946b4e72dd6469c4f32174dce528a555824.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 415, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 654 + ], + "type": "text", + "content": "Dreambooth, and Custom Diffusion, fall short in accurately aligning with text prompts and preserving identity features. On the contrary, our method shows comparable performance to FastComposer and IP-Adapter-Face in terms of both CLIP Score and ID metrics. However, it is important to note that IP-Adapter has been fine-tuned on the large-scale LAION2B dataset, whereas our model is trained on the relatively smaller FFHQ-wild and SHHQ datasets, comprising around 110,000 images. Both FastComposer and IP-Adapter directly process the reference image as input, while our approach first maps it to the " + }, + { + "bbox": [ + 46, + 415, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 415, + 289, + 654 + ], + "type": "text", + "content": " space before embedding it into the SD model. This additional mapping step could potentially introduce minor discrepancies when integrating the reference image into the SD model, slightly affecting the ID results. Nevertheless, this slight trade-off in ID preservation is balanced by our method's ability to flexibly edit facial attributes while ensuring the background remains consistent. This unique balance of attribute editability and background consistency is a novel contribution that sets our method apart from previous approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 659, + 183, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 183, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 183, + 672 + ], + "type": "text", + "content": "4.2. Qualitative Comparison" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Figure 3 depicts visual comparisons with baselines across various scenarios. Since Textual Inversion, Dreambooth, and Custom Diffusion are not specifically designed for facial im" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 415, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 545 + ], + "type": "text", + "content": "ages, we select results with the best facial quality from their generated set of 40 images. For the remaining methods, we choose from a set of 10. It is observed that Textual Inversion, Dreambooth, and Custom Diffusion encounter difficulties in capturing and maintaining identity details with a single reference, leading to sub-optimal performance in this task. CelebBasis and FastComposer face challenges in striking a balance between text compatibility and identity preservation. Though IP-Adapter shows improved identity retention, it tends to ignore text conditions in certain instances (see the 1st, 3~5th rows)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "text", + "content": "Leveraging our " + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "text", + "content": "adapter, our approach successfully generates images that are not only compatible with text descriptions but also more effectively retain the target identity. Additionally, our method allows for the editing of facial attributes along the " + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "text", + "content": " direction, causing only minor alterations in the non-facial regions (illustrated in the last column). Furthermore, our approach can be seamlessly adapted to other pre-trained SD models without the need for additional fine-tuning, while retaining its editing capabilities. This versatility is exemplified in the last row of Fig. 3, which showcases our method's effectiveness with the dreamlike-anime model" + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 548, + 548, + 691 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 702, + 509, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 509, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 509, + 714 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 316, + 702, + 509, + 714 + ], + "type": "text", + "content": "https://huggingface.co/dreamlike-art/dreamlike-anime-1.0" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2192" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 282, + 249 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 282, + 249 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 282, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 282, + 249 + ], + "type": "image", + "image_path": "409210800335fcec3ad7daee473a63387a14b0a4459ead664598ac675d35a706.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": "Figure 4. Inversion and editing comparisons between e4e and Ours in Stage I. They have the same " + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": " and attributes editing " + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 306, + 281, + 354 + ], + "blocks": [ + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of face inversion and editing. Editing (ID↓) is the ID metric on results of " + }, + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "type": "inline_equation", + "content": "w_{+} - 3 \\cdot \\Delta w" + }, + { + "bbox": [ + 47, + 282, + 288, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 306, + 281, + 354 + ], + "lines": [ + { + "bbox": [ + 50, + 306, + 281, + 354 + ], + "spans": [ + { + "bbox": [ + 50, + 306, + 281, + 354 + ], + "type": "table", + "html": "
MethodsInversionEditing (ID↓)
ID↓FID↓LPIPS↓AgeSmile
e4e [32].43132.26.205.409.312
Ours (Stage I).43531.47.263.393.275
", + "image_path": "03d7bfed90bd91cb02bd40bccbf00216298d09c871ed2db0e5d3077b988066fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 361, + 141, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 141, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 141, + 374 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": "Analyses of Aligning " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " in Stage I. We examine if the StyleGAN's " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " embedding is well aligned to the latent space of SD during Stage I training. Fig. 4 offers a qualitative comparison of inversion and attribute editing outcomes between e4e [32] and our approach, applied to a real-world image. Both methods use the identical " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " embedding generated by the e4e encoder. As depicted in Fig. 4, our method yields inversion results on par with e4e, as seen in the left column, using the same " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " vector. Furthermore, the alignment process we implement in this phase preserves the editability inherent to the " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " space. By introducing attribute editing directions " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " (such as smile and age) from InterFaceGAN [29], our method can semantically adjust these attributes while maintaining the individual's identity. This is illustrated in the right columns, where the edited embedding, denoted as " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "w_{+} + \\alpha \\cdot \\Delta w" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": ", reflects these changes. The parameter " + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 380, + 289, + 583 + ], + "type": "text", + "content": " is used to control the extent of attribute modification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "content": "To quantitatively evaluate the alignment performance, we randomly select 1,000 images from CelebA-HQ [14] and measure their inversion results with metrics of ID, FID, and LPIPS " + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "inline_equation", + "content": "[43]^3" + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "content": ". As indicated in the Inversion column of Table 2, our method exhibits comparable performance to e4e in the image inversion task. For assessing attribute editing capabilities, we focus on two key attributes, namely smile, and age, applying a constant scale " + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "inline_equation", + "content": "w_{+} - 3 \\cdot \\Delta w" + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "content": "). The Editing column in Table 2 demonstrates that our approach surpasses" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 80, + 541, + 140 + ], + "blocks": [ + { + "bbox": [ + 324, + 69, + 526, + 79 + ], + "lines": [ + { + "bbox": [ + 324, + 69, + 526, + 79 + ], + "spans": [ + { + "bbox": [ + 324, + 69, + 526, + 79 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons of different variants." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 80, + 541, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 80, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 80, + 541, + 140 + ], + "type": "table", + "html": "
VariantsCLIP Score↑ID↓Detection↑
Ours (PCA).243.526.936
Ours (w/o Lreg).136.461.951
Ours (L0.5reg).192.497.946
Ours (Full).267.516.943
", + "image_path": "39773dcbc3bd3fcefc981657c96978f4f1e76667f710b33b968ccb6d548c9f3b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 145, + 543, + 216 + ], + "blocks": [ + { + "bbox": [ + 309, + 145, + 543, + 216 + ], + "lines": [ + { + "bbox": [ + 309, + 145, + 543, + 216 + ], + "spans": [ + { + "bbox": [ + 309, + 145, + 543, + 216 + ], + "type": "image", + "image_path": "8f4c3dff5baf3afd4df359744407addc54b2acf4dc6f07fa79f2e897981a5931.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 218, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 305, + 218, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 546, + 240 + ], + "type": "text", + "content": "Figure 5. Visual comparisons of smile attribute editing. The prompt is \"a woman wearing a t-shirt on the beach\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 247, + 546, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 546, + 294 + ], + "type": "text", + "content": "e4e in maintaining identity during the editing process. These findings confirm that our method has successfully aligned the " + }, + { + "bbox": [ + 305, + 247, + 546, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 305, + 247, + 546, + 294 + ], + "type": "text", + "content": " space with the SD model, preserving both the inversion accuracy and the robust editability of attributes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "content": "Analyses of Variants. We examine different variants to assess each component of our " + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "content": "adapter: 1) Ours (PCA), which substitutes our residual cross-attention with parallel cross-attention as in [39], and 2) Ours (" + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "inline_equation", + "content": "w/o\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "content": "), which omits the " + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "content": " objective. The outcomes of these variants are presented in Table 3. It is observed that the residual cross-attention module outperforms the parallel cross-attention design. We conjecture that in tasks focused on integrating local facial imagery into broader, in-the-wild scenarios, directly injecting facial embeddings into the original hidden state " + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "inline_equation", + "content": "f_{z}" + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "content": " might adversely impact non-facial regions. Conversely, our residual cross-attention determines fusion weights by calculating attention activations between the text-embedded hidden state " + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "inline_equation", + "content": "f_{z}^{\\prime}" + }, + { + "bbox": [ + 304, + 294, + 547, + 474 + ], + "type": "text", + "content": " and our facial embedding. This results in a more precise and effective fusion within the facial region." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": "Besides, the results suggest that the regularization loss " + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": " is crucial in maintaining semantic consistency between text prompts and images. In the absence of " + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": ", the variant Ours " + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "inline_equation", + "content": "(w / o\\mathcal{L}_{reg})" + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": " tends to overemphasize facial regions while neglecting the accompanying text descriptions, leading to a reduced CLIP score. Incorporating " + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": " with a weight of " + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\gamma_{2} = 0.5" + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": ", as in Ours " + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{reg}^{0.5})" + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": ", enhances the CLIP score but slightly compromises the ID metric. In comparison, Ours (Full) achieves a satisfactory trade-off in balancing the performance of ID and CLIP score." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": "The impact of our disentanglement loss " + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{disen}" + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": " (Eqn. (6)) is visually demonstrated in Fig. 5. This example shows that omitting " + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{disen}" + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": " leads to unintended alterations in non-facial regions when the smile editing direction " + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": ". This indicates the importance of " + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{disen}" + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": " in effectively separating identity-relevant and irrelevant information. By combining our residual cross-attention, regularization loss, and disentanglement loss, Ours (Full) preserves identity and ensures compatibility with text prompts, even amidst changes to the face embedding." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "text", + "content": "3https://github.com/chaofengc/IQA-PyTorch" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2193" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 534, + 202 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 94, + 82 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 94, + 82 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 94, + 82 + ], + "type": "text", + "content": "w+" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 73, + 354, + 82 + ], + "lines": [ + { + "bbox": [ + 238, + 73, + 354, + 82 + ], + "spans": [ + { + "bbox": [ + 238, + 73, + 354, + 82 + ], + "type": "text", + "content": "T2I Generation with " + }, + { + "bbox": [ + 238, + 73, + 354, + 82 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 238, + 73, + 354, + 82 + ], + "type": "text", + "content": " Interpolation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 497, + 70, + 510, + 82 + ], + "lines": [ + { + "bbox": [ + 497, + 70, + 510, + 82 + ], + "spans": [ + { + "bbox": [ + 497, + 70, + 510, + 82 + ], + "type": "text", + "content": "w" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 534, + 202 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 534, + 202 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 534, + 202 + ], + "type": "image", + "image_path": "87f2a5864478f6deb56b0189abeea0a4ace60c140c31bd97b53529a0583e3027.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 236, + 279, + 456 + ], + "blocks": [ + { + "bbox": [ + 46, + 206, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 229 + ], + "type": "text", + "content": "Figure 6. Visual results of " + }, + { + "bbox": [ + 46, + 206, + 547, + 229 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 206, + 547, + 229 + ], + "type": "text", + "content": " embeddings interpolation during inference. The prompts are \"one person wearing suit and tie in a garden\" and \"one person wearing a blue shirt by a secluded waterfall\", respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 236, + 279, + 456 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 279, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 279, + 456 + ], + "type": "image", + "image_path": "66955cd1c0cc40e65c350fde7b75ae383db0c88cbd57ceb57877c68dd562a170.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "type": "text", + "content": "Figure 7. Visual results of using different " + }, + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "type": "text", + "content": " during inference. The prompt is \"a man wearing a rainbow shirt in a garden\"." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": "Analyses of " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " in Cross-attention. The parameter " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " determines the extent to which the " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " vector influences the hidden state " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "f_{z}^{\\prime}" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": ". Fig. 7 shows the effects of varying " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " during inference. At " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": ", the outcome aligns with that of the original SD model, unaffected by any editing direction. However, as " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " increases, the generated identity more closely resembles the reference image. Interestingly, at relatively low values of " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " (for example, " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": "), even though the identity does not closely match the reference, our model effectively edits attributes in the specified direction. This observation suggests that our method proficiently leverages the " + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 46, + 496, + 287, + 640 + ], + "type": "text", + "content": " space from StyleGAN within the SD framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Analyses of " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " Interpolation. We select two facial images and acquire their respective " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "w_{+}^{1}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "w_{+}^{2}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " embeddings from e4e encoder. The interpolated " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " embedding is then obtained through " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "(1 - \\kappa) \\cdot w_{+}^{1} + \\kappa \\cdot w_{+}^{2}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\kappa \\in [0,1]" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ". As shown in Fig. 6, this T2I generation process results in a smooth transition in the facial regions of the generated" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 297 + ], + "type": "text", + "content": "images, while maintaining a similar and consistent layout throughout the interpolation. This result confirms that the " + }, + { + "bbox": [ + 304, + 236, + 547, + 297 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 236, + 547, + 297 + ], + "type": "text", + "content": " embedding in our method not only preserves the editability characteristic of the original StyleGAN but also effectively distinguishes between facial and background regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 304, + 379, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 379, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 379, + 316 + ], + "type": "text", + "content": "4.4. Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "text", + "content": "Our work aims at integrating StyleGAN's editable " + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "text", + "content": " space into the SD model. We notice a challenge in this integration: the process of converting real images to " + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "text", + "content": " vectors in StyleGAN often leads to a loss of detail, impacting the preservation of identity features. Despite employing a substantial number of training pairs " + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\{I_f, w_{+}\\}" + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "text", + "content": " to establish the mapping, we still observe limitations in maintaining identity fidelity and challenges in editing certain attributes (e.g., pose and glasses). The current framework is designed to generate and edit images with only a single face. Our future work aims to explore the potential of applying localized injections of distinct " + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 323, + 547, + 467 + ], + "type": "text", + "content": " to address multiple human subjects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 477, + 379, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 379, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 379, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "text", + "content": "We have presented the first attempt to embed the " + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "text", + "content": " space of StyleGAN into the SD model. We showed that both the mapping network and the residual cross-attention module play crucial roles in facilitating the injection of " + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "inline_equation", + "content": "w_{+}" + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "text", + "content": " embedding into the SD model, balancing between text prompt influence and identity conditions. Our experiments demonstrate that the " + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{+}" + }, + { + "bbox": [ + 304, + 497, + 547, + 665 + ], + "type": "text", + "content": " space, as used in our approach, not only enables personalized text-to-image generation but also allows for precise editing of facial attributes. We envision this capability being highly beneficial in various practical applications, such as portrait customization with seamless attribute modifications. Furthermore, our methodology holds the potential for application across other object domains that possess distinct prior spaces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2194" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: How to embed images into the StyleGAN latent space? In ICCV, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "type": "text", + "content": "[2] Yufei Cai, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Hu Han, and Wangmeng Zuo. Decoupled textual embeddings for customized image generation. In AAAI, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[3] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. DisenBooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "type": "text", + "content": "[4] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, and Min Zheng. PhotoVerse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 289 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 323 + ], + "type": "text", + "content": "[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "type": "text", + "content": "[7] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. StyleGAN-Human: A data-centric odyssey of human generation. In ECCV, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "text", + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "text", + "content": "[9] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 288, + 490 + ], + "type": "text", + "content": "[10] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: Discovering interpretable GAN controls. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 556 + ], + "type": "text", + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "type": "text", + "content": "[13] Xuhui Jia, Yang Zhao, Kelvin C. K. Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 645 + ], + "type": "text", + "content": "[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 3, 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 184 + ], + "type": "text", + "content": "[19] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image superresolution. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 546, + 229 + ], + "type": "text", + "content": "[20] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. A survey on leveraging pre-trained generative adversarial networks for image editing and restoration. arXiv preprint arXiv:2207.10309, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 231, + 547, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 547, + 295 + ], + "type": "text", + "content": "[21] Yang Liu, Cheng Yu, Lei Shang, Ziheng Wu, Xingjun Wang, Yuze Zhao, Lin Zhu, Chen Cheng, Weitao Chen, Chao Xu, Haoyu Xie, Yuan Yao, Wenmeng Zhou, Yingda Chen, Xuansong Xie, and Baigui Sun. FaceChain: A playground for identity-preserving portrait generation. arXiv preprint arXiv:2308.14256, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 297, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 319 + ], + "type": "text", + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 320, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 547, + 374 + ], + "type": "text", + "content": "[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 376, + 547, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 408 + ], + "type": "text", + "content": "[24] Hamza Pehlivan, Yusuf Dalva, and Aysegul Dundar. StyleRes: Transforming the residuals for real image editing with StyleGAN. In CVPR, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 410, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 547, + 464 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 466, + 546, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 546, + 498 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 500, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 546, + 532 + ], + "type": "text", + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 533, + 547, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 547, + 577 + ], + "type": "text", + "content": "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Finetuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 578, + 547, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 547, + 611 + ], + "type": "text", + "content": "[29] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 547, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 655 + ], + "type": "text", + "content": "[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. InstantBooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 657, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 546, + 679 + ], + "type": "text", + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": "[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. In ACM TOG, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2195" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 566 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "text", + "content": "[33] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "type": "text", + "content": "[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 194 + ], + "type": "text", + "content": "[35] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 288, + 239 + ], + "type": "text", + "content": "[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 288, + 284 + ], + "type": "text", + "content": "[37] Zijie Wu, Chaohui Yu, Zhen Zhu, Fan Wang, and Xiang Bai. SingleInsert: Inserting new concepts from a single image into text-to-image models for flexible editing. arXiv preprint arXiv:2310.08094, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 288, + 330 + ], + "type": "text", + "content": "[38] Guangxuan Xiao, Tianwei Yin, William T. Freeman, Frédo Durand, and Song Han. FastComposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 288, + 374 + ], + "type": "text", + "content": "[39] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 4, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 376, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 288, + 418 + ], + "type": "text", + "content": "[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. In NeurIPS, 2023. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 421, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 288, + 464 + ], + "type": "text", + "content": "[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. In IEEE signal processing letters, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 466, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 288, + 498 + ], + "type": "text", + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 499, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 499, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 499, + 288, + 532 + ], + "type": "text", + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 533, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 288, + 566 + ], + "type": "text", + "content": "[44] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In ECCV, 2020. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2196" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_content_list.json b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a51d2a707df2b2a69f2464241dce09b5e528e0d6 --- /dev/null +++ b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_content_list.json @@ -0,0 +1,1763 @@ +[ + { + "type": "text", + "text": "When Visual Grounding Meets Gigapixel-level Large-scale Scenes: Benchmark and Approach", + "text_level": 1, + "bbox": [ + 148, + 128, + 818, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Ma $^{1*}$ , Bing Bai $^{1*}$ , Haozhe Lin $^{1*}$ , Heyuan Wang $^{2}$ , Yu Wang $^{1}$ , Lin Luo $^{2}$ , Lu Fang $^{1\\dagger}$ , Tsinghua University ${}^{2}$ Peking University", + "bbox": [ + 151, + 202, + 821, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fanglu@tsinghua.edu.cn", + "bbox": [ + 383, + 242, + 580, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f47882ec46cd4979712bbdefbb3eae343faaadf04255722ecc3a6b1dd6f675e1.jpg", + "image_caption": [ + "(a) RefCOCO" + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 297, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2a044a0374f3a69ec5cc24184651cfc27b0a441dcc80a647f8229418844d828.jpg", + "image_caption": [ + "(b) GigaGrounding", + "$\\nrightarrow$ i. lower left white car", + "Figure 1. Comparison of the proposed GigaGrounding and the RefCOCO benchmark. GigaGrounding presents distinctive challenges, including large-scale scene understanding, high-resolution with significant scale variation, and multi-hop expressions." + ], + "image_footnote": [], + "bbox": [ + 316, + 276, + 647, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b608c9397576e19d045543d5dc78515ae474057f0bb8ad6fdbfa6fad9fd2fb81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 276, + 880, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 488, + 313, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual grounding refers to the process of associating natural language expressions with corresponding regions within an image. Existing benchmarks for visual grounding primarily operate within small-scale scenes with a few objects. Nevertheless, recent advances in imaging technology have enabled the acquisition of gigapixel-level images, providing high-resolution details in large-scale scenes containing numerous objects. To bridge this gap between imaging and computer vision benchmarks and make grounding more practically valuable, we introduce a novel dataset, named GigaGrounding, designed to challenge visual grounding models in gigapixel-level large-scale scenes. We extensively analyze and compare the dataset with existing benchmarks, demonstrating that GigaGrounding presents unique challenges such as large-scale scene understanding, gigapixel-level resolution, significant variations in object scales, and the \"multi-hop expressions\". Furthermore, we introduced a simple yet effective grounding approach, which employs a \"glance-to-zoom-in\" paradigm and exhibits enhanced capabilities for addressing the GigaGrounding task. The dataset is available at www.gigavision.ai.", + "bbox": [ + 73, + 523, + 473, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 488, + 630, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual grounding (VG), also known as referring expression comprehension [1, 2], phrase localization [3, 4], and natural language object retrieval [5, 6], aims to identify the region within an image that corresponds to a given natural language expression. This field offers considerable potential for enhancing our capacity to connect the human language and visual elements of the tangible world [7]. The application of VG in scenarios such as rescue and robotics can prove to be exceedingly valuable, for example, in searching for a missing child in crowds.", + "bbox": [ + 496, + 518, + 892, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the VG community has made significant progress in recent years, the benchmarks employed thus far have been predominantly limited to small-scale scenes [1-4]. As illustrated in Figure 1a, the number of objects is limited, and the objects cover the dominant portion of the image. However, recent advances in imaging technology, such as array cameras [8] and even consumer-grade smartphones [9], have enabled the capture of gigapixel-level photos, providing high-resolution detail information while covering large-scale scenes up to $\\mathrm{km}^2$ -level [10], as shown in Figure 1b. The characteristics of conventional VG benchmarks have restricted current models' application to relatively simple and vanilla settings, hindering the development of new technologies and limiting the applicability of grounding in broader contexts [11].", + "bbox": [ + 496, + 672, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work. \n†Lu Fang is the corresponding author (www.luvision.net).", + "bbox": [ + 93, + 875, + 434, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22119", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this limitation, we introduce a novel benchmark dataset, named GigaGrounding, for visual grounding in large-scale scenes. The dataset is constructed based on PANDA [11], a gigapixel-level human-centric video dataset. An illustration of the GigaGrounding dataset is displayed in Figure 1b. In contrast to conventional visual grounding benchmarks such as RefCOCO [2], GigaGrounding poses unique and non-trivial challenges:", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Large-scale scene understanding. Large-scale scenes require models to deal with a much larger spatial and semantic context and a significantly larger number of objects in the image. Hundreds of objects can be distributed over a vast area with possible occlusions, clutter, and other distractions, making it a difficult task to recognize and understand large-scale scenes.", + "- High resolution with significant scale variation. Processing high-throughput data up to gigapixel-level images poses new challenges to models in terms of efficiency [10]. Moreover, objects may be presented at drastically different scales. For instance, as shown in Figure 1b, the ground-truth bounding box (B-box) for Case i spans approximately $3000 \\times 1400$ pixels, whereas, for Case ii, it spans only about $110 \\times 330$ pixels. This significant variation necessitates models to be adaptable in scaling to accommodate diverse object sizes.", + "- Multi-hop expressions. In addition to direct expressions, which refer to descriptions of independent objects as illustrated in Case i, GigaGrounding also includes a considerable number of multi-hop expressions. These expressions challenge grounding models to locate target objects by first identifying some other reference objects. For example, in Case ii, the model needs first to identify \"the middle lawn\" and then zoom in to locate the correct person based on the relationship with the lawn." + ], + "bbox": [ + 76, + 220, + 468, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We benchmark the performance of cutting-edge VG models with the GigaGrounding dataset. Empirical findings reveal that while existing VG models have demonstrated satisfactory performance in small-scale scenes, they face distinct and intricate challenges when being applied to GigaGrounding. To address these challenges, we propose GlaZing (GLance-to-Zoom-IN Grounding), a simple yet effective model inspired by human cognitive processes. GlaZing employs a \"glance-to-zoom-in\" cascading strategy to ground target bounding boxes in high-resolution images. It begins by taking a global view of the thumbnail image, to identify the region of interest based on the given expression. Subsequently, it zooms in on the local region with adaptive scaling for the final grounding process. Experiments demonstrated that GlaZing outperformed existing models in the GigaGrounding dataset, effectively addressing major challenges such as high resolution, significant scale variation, and multi-hop expressions.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, this paper presents several contributions:", + "bbox": [ + 517, + 90, + 885, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We highlight the significance of deploying visual grounding models in large-scale scenes, and introduce a novel benchmark dataset, GigaGrounding, characterized by unique properties that set it apart from existing datasets.", + "- We empirically assess current VG models using the proposed GigaGrounding dataset, highlighting the limitations of state-of-the-art models in this context.", + "- We introduce GlaZing, a tailored model for GigaGrounding, which leverages a \"glance-to-zoom-in\" approach. Our experimental results demonstrate the efficacy of this approach in addressing the current challenges, while also indicating room for further improvements." + ], + "bbox": [ + 500, + 111, + 890, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 314, + 640, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we review two relevant research domains: visual grounding and high-resolution deep learning.", + "bbox": [ + 498, + 340, + 890, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Grounding", + "text_level": 1, + "bbox": [ + 500, + 383, + 674, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual grounding is a field focused on predicting the location within an image that corresponds to a natural language expression. Benchmark datasets for VG have predominantly centered on small-scale scenes. For example, the Flickr30K Entities dataset [4] extends the original Flickr30K dataset [12] by incorporating annotations that establish correspondence between short region phrases and images. ReferItGame [3] comprises 20,000 images collected from the SAIAPR-12 dataset [13]. RefCOCO [2], RefCOCO+ [2], and RefCOCOg [1] are all built upon the MS COCO dataset [14]. Notably, the image resolutions in these datasets do not exceed $640 \\times 480$ . In addition, there are related benchmarks in fields such as remote sensing [15, 16] and biomedical science [17, 18], with their resolutions also limited to $1024 \\times 1024$ or lower.", + "bbox": [ + 496, + 409, + 890, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For VG models, existing work can be broadly categorized into two-stage and one-stage methods [7, 19]. Two-stage methods [20-24] initiate the process by generating a set of region proposals using object detectors. Subsequently, they leverage natural language expressions to rank these proposals and identify the best matching proposal. In contrast, one-stage methods [7, 19, 25-29] densely integrate both visual and textual information and directly output bounding boxes.", + "bbox": [ + 496, + 637, + 890, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. High-resolution Deep Learning", + "text_level": 1, + "bbox": [ + 500, + 785, + 774, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, significant progress has been made in the field of imaging, leading to gigapixel-level images with both a wide FoV and high resolution [8, 30, 31]. Employing high-resolution images and videos as direct inputs for deep learning models introduces challenges in both training and inference stages [11, 32-34]. To address these", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22120", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a60ed7964f95a8f84305e04043b1cc8e0dfca40393952f9979bfa7370afb1b65.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTypical res.# Images# B-boxes# Expressions
Flickr30K Entities [4]500×37531,783276k427k
ReferItGame [3]480×36019,89496,654130,525
RefCOCO [2]640×48019,99450,000142,210
RefCOCO+ [2]640×48019,99249,856141,564
RefCOCOg [1]640×48025,79949,82295,010
GigaGrounding>25k×14k3,77561,35361,353
", + "bbox": [ + 81, + 88, + 467, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Statistics of existing benchmarks and GigaGrounding.", + "bbox": [ + 84, + 179, + 459, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "challenges, several approaches have been explored. Rudimentary strategies like \"uniform downsampling (UD)\" and \"cutting into patches (CIP)\" may be employed, but they may result in the loss of detailed information and relational cues, respectively [10]. Alternatively, more sophisticated techniques, including non-uniform downsampling [35-37], selective zooming and skipping [38-40], and the use of lightweight scanner networks [41, 42], have been proposed. Bakhtiarnia et al. [10] have provided a comprehensive review of these methods, their relevant applications, and datasets pertaining to high-resolution deep learning.", + "bbox": [ + 75, + 213, + 468, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. GigaGrounding Benchmark", + "text_level": 1, + "bbox": [ + 76, + 393, + 336, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section offers a comprehensive delineation of the GigaGrounding benchmark, including an overview, an elucidation of the data collection strategy, and analyses of its unique properties.", + "bbox": [ + 75, + 419, + 468, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of GigaGrounding", + "text_level": 1, + "bbox": [ + 76, + 491, + 333, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset overview GigaGrounding is built upon images pre-extracted from PANDA-Video dataset by the dataset providers [11]. A summary of key statistics comparing conventional VG benchmarks with GigaGrounding is provided in Table 1. The analysis reveals a significant limitation in conventional VG datasets, where image resolutions are constrained to the megapixel level. In contrast, GigaGrounding ambitiously extends its scope to gigapixel-level images, encompassing intricate details and large-scale scenes. Consequently, the total number of images naturally decreases. In terms of bounding boxes, GigaGrounding offers a quantity comparable to other benchmark datasets. Additionally, in traditional benchmarks, multiple expressions may correspond to a single bounding box, while in GigaGrounding, each bounding box is associated with only one expression.", + "bbox": [ + 75, + 518, + 467, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We illustrate representative samples from conventional VG datasets in Figure 2, and also provide examples of GigaGrounding annotations in Figure 3. Specifically, GigaGrounding includes two types of annotated expressions:", + "bbox": [ + 75, + 744, + 467, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Direct expressions: These expressions directly describe each target object without introducing dependencies on other objects. Position modifiers, such as \"in the top left of the image\", may be used when describing small and challenging-to-locate objects, but only when necessary. In total, there are 41,040 direct expressions.", + "bbox": [ + 76, + 809, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Multi-hop expressions: These expressions require grounding models to respect additional restrictive constraints dependent on their relationship with other reference objects. In other words, the model needs to locate objects by first identifying specific reference objects and then finding the target objects based on constraints defined by relative positions (e.g., \"to the left\") or relationships (e.g., \"watching\", \"holding\") with those reference objects. There are 20,313 multi-hop expressions in total. It is noteworthy that we did not identify a design intentionally analogous to multi-hop expressions in traditional VG benchmarks. Notably, RefCOCO+ was explicitly designed to \"focus on purely appearance-based description\" [2].", + "bbox": [ + 500, + 90, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For more details regarding the PANDA-Video dataset and additional examples of GigaGrounding, please refer to the Supplementary Material.", + "bbox": [ + 498, + 308, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Privacy GigaGrounding is built based on the publicly available PANDA [11] dataset. PANDA is collected in public areas where photography is officially approved and is published under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License [43]. GigaGrounding has been created by further processing the data in PANDA with additional annotations. To protect individual privacy, PANDA has already anonymized the faces in the images. However, it is necessary to acknowledge that certain individuals may still be identifiable based on other less invasive characteristics, such as clothing or body shape [44].", + "bbox": [ + 496, + 357, + 890, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Collection Strategy", + "text_level": 1, + "bbox": [ + 500, + 551, + 723, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We have devised the following annotation procedure and quality control strategies to diligently uphold non-ambiguity and maintain high annotation quality.", + "bbox": [ + 498, + 577, + 890, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotation protocol The annotation protocol, as depicted in Figure 4, comprises the following steps:", + "bbox": [ + 498, + 627, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Given the gigapixel images, Group $A$ of workers was asked to write natural language expressions and draw bounding boxes for the corresponding objects.", + "2. Given the gigapixel images and natural language expressions from Step 1, Group $B$ of workers was then asked to draw corresponding bounding boxes within 40 seconds independently.", + "3. If the Intersection over Union (IoU) rate of bounding boxes from Steps 1 and 2 was greater than 0.5, the sample were deemed valid. Then the valid expressions were manually translated into English with the highest possible fidelity." + ], + "bbox": [ + 500, + 662, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This procedure tries to ensure non-ambiguity, by excluding data instances with object selection discrepancies between Steps 1 and 2.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22121", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e5a0078ffde46f02d26fb85d2f5460f78798a06325c46a07d93228369544b6f5.jpg", + "image_caption": [ + "A man with pierced ears is wearing glasses and an orange hat", + "(a) Flickr30K Entities" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 207, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9ec1afe38b118013d88c74b98ae3630ec5d85e3f44fc4c40e569c0198461f67.jpg", + "image_caption": [ + "1. white shirt on right", + "2. white shirt man", + "3. man on right", + "(b) ReferItGame" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 369, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/626a47c1e317b588cecc63e4e7c1ef553452a334bdfd62b9dc0fafe1b5bba24c.jpg", + "image_caption": [ + "1. woman on right in white shirt", + "2. woman on right", + "3. right women", + "(c) RefCOCO" + ], + "image_footnote": [], + "bbox": [ + 370, + 88, + 529, + 180 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80e1a4747d73332c317ec4aafd2bf7e775ea2d1cdd71fc2d198822be613fbd0b.jpg", + "image_caption": [ + "1. yellow shirt and black shorts", + "2. guy in yellow dibbling ball", + "3. yellow shirt in focus", + "(d) RefCOCO+" + ], + "image_footnote": [], + "bbox": [ + 531, + 89, + 705, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/625ff343b89777263452e430669f5930b94ee899c3913fabb18336ca8b996e93.jpg", + "image_caption": [ + "1. a woman wearing red playing tennis", + "2. a woman in a red shirt holding a tennis racquet", + "(e) RefCOCOg" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 888, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/af62dc686c5e50b86aa66feeffa1446242e1f515c7221d414ddde0f2a74d5f83.jpg", + "image_caption": [ + "man in a red shirt pushing a suitcase near a subway station" + ], + "image_footnote": [], + "bbox": [ + 81, + 273, + 135, + 320 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff67f1bda316f7d0c36de4412f82244eed90fbfcd1396ec96f6ecc4f96b89760.jpg", + "image_caption": [ + "woman in a yellow top riding a tricycle in the middle of the picture" + ], + "image_footnote": [], + "bbox": [ + 83, + 324, + 148, + 368 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f0d23e7926b1470d14fc8fe837e1e2058c5757d4ef0f6bfecd35ac43d00c56b.jpg", + "image_caption": [ + "man in white jacket and black bag in front of the mall on the left of the picture" + ], + "image_footnote": [], + "bbox": [ + 83, + 375, + 129, + 415 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67f656c3dd43e1a4bad14f8e1f880bb3df342d6794e95532cfc1ec37737e50be.jpg", + "image_caption": [ + "man in a white coat riding a blue bike at the bottom left of the picture" + ], + "image_footnote": [], + "bbox": [ + 83, + 422, + 137, + 465 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3568193972739f7634a1cab604fda624a83f5662133dc9dd217dfcb67d2b4fe7.jpg", + "image_caption": [ + "Figure 2. Example data from conventional visual grounding datasets.", + "Figure 3. Example annotation results of GigaGrounding. Blue boxes indicate direct expressions and orange ones are multi-hop expressions." + ], + "image_footnote": [], + "bbox": [ + 277, + 273, + 514, + 468 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/45662d443273f9d9a60b6e8c8aa7699d3973497116e6244fab08763a21eab923.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 273, + 689, + 468 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00f7afc191be100ce2b25b9c87f12ca0dc9bcb02be5cd1075f6962070df9c381.jpg", + "image_caption": [ + "blue billboard in the top right corner of the picture" + ], + "image_footnote": [], + "bbox": [ + 694, + 273, + 785, + 319 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8fbb42d245127cafe7e68926b39715c594968a61b26aef0b9480fe534768674.jpg", + "image_caption": [ + "green spherical plant in the flower bed on the right of the picture" + ], + "image_footnote": [], + "bbox": [ + 694, + 324, + 764, + 368 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db592d7b0d879f901b3a7fdc08941646f3630a882b9a5a85f2d6ff7ab5d45260.jpg", + "image_caption": [ + "red and white bus" + ], + "image_footnote": [], + "bbox": [ + 694, + 375, + 795, + 416 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27d52e7710d92c789c7bbd78b279fe1f8be4206a55def5d87731e0cb55583230.jpg", + "image_caption": [ + "taxi closest to the red and white bus" + ], + "image_footnote": [], + "bbox": [ + 694, + 422, + 746, + 465 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b971872432f8cdd835169422d7cf51d110b984c25d62b8064e86e3e743ad1ee.jpg", + "image_caption": [ + "Figure 4. The overview of our annotation procedure." + ], + "image_footnote": [], + "bbox": [ + 101, + 525, + 446, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quality control strategies We employ the following strategies for quality control.", + "bbox": [ + 75, + 679, + 467, + 709 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Terminology standardization. We introduced a standardized lexicon to encourage uniform terminology usage among annotators, reducing the prevalence of ambiguous or domain-specific terms. For example, the use of specific color names like \"gamboge\" and the adoption of specialized terminology such as \"coupe\" should be avoided.", + "- Sampling inspection and repair. We performed a sampling inspection for each annotation batch to assess its quality. Specifically, for expression and bounding box annotations, we established a sampling pass rate threshold of $96\\%$ , whereas, for expression translation, the threshold was set to $99\\%$ . The sampling inspection process" + ], + "bbox": [ + 76, + 715, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "included no fewer than 100 samples per round, and the batch of data was accepted once it met the pass rate threshold or underwent repair.", + "bbox": [ + 511, + 522, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Box ensemble and de-duplication. In Step 3 of the annotation procedure, we obtained the ground truth by averaging the valid bounding boxes annotated by both groups of annotators. Any expressions that did not result in a valid bounding box were discarded. Furthermore, as data annotation is carried out in parallel among many annotation workers, we performed a dedduplication procedure to eliminate objects with overlapping bounding boxes and identical expressions across multiple images to enhance the data variance, with an IoU threshold empirically established at 0.3.", + "bbox": [ + 500, + 570, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite potential undetected ambiguities, the implemented methodologies are believed to substantially mitigate their effects, thereby enhancing the dataset's overall quality.", + "bbox": [ + 498, + 741, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Analyses of GigaGrounding", + "text_level": 1, + "bbox": [ + 500, + 797, + 750, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section analyzes the salient characteristics differentiating GigaGrounding from existing VG benchmarks. Our analysis encompasses the following viewpoints.", + "bbox": [ + 498, + 821, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantities of objects per image Among conventional datasets, Flickr30K Entities in Figure 2a requires the model", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22122", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f771c284aa1cc14660759392778b3b15d3b9eefc8c65a4a26124d601712278c.jpg", + "image_caption": [ + "(a) Pixel-count distribution of bounding box scales." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 274, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dce1e696e3e8b90c0fdd1d2b539888508b739869c83e6d3bc2879e5f13857d2c.jpg", + "image_caption": [ + "Figure 5. A comparative analysis of GigaGrounding against established benchmarks RefCOCO, RefCOCO+, and RefCOCOg. (a) and (b) elucidate the distribution patterns of bounding box scale (in terms of the max value of height and width) and relative area with respect to the image, (c) presents the distribution of expression lengths, and (d) shows the lengths' distribution of both direct and multi-hop expressions within the GigaGrounding dataset. GigaGrounding exhibits a notably elevated level of complexity, primarily stemming from its pronounced scale variation and intricate expression structures." + ], + "image_footnote": [], + "bbox": [ + 282, + 90, + 478, + 194 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c49a5ac1eb51cba710b184cfdbb9629d98137ddaddf806b010f1a7ac95db44b9.jpg", + "image_caption": [ + "(b) Relative area of bounding boxes w.r.t. the images." + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 676, + 194 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c481339ad783e0468048fc25a10dbf082910ed1f3720805e36f571fb3c0671c.jpg", + "image_caption": [ + "(c) Word-count distribution of expression lengths.", + "(d) Word-count distribution of direct and multi-hop expressions." + ], + "image_footnote": [], + "bbox": [ + 683, + 90, + 887, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to identify and locate multiple objects corresponding to short phrases within a single sentence. However, we observe that the model may not need to fully understand the global semantics before grounding the corresponding object, particularly when there exists only one object of that type. On the other side, RefCOCO and RefCOCO+ include an average of 3.9 same-type objects per image, while RefCOCOg displays a relatively lower average of 1.63 same-type objects per image [2].", + "bbox": [ + 73, + 318, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scales of bounding boxes In Figures 5a and 5b, we respectively illustrate the absolute scale of bounding boxes and their relative area with respect to the images. It is evident that GigaGrounding exhibits greater scale variance in comparison to other datasets. Furthermore, bounding box scale in GigaGrounding adheres to a long-tail distribution, while other datasets generally show a more concentrated distribution. On average, the RefCOCO dataset features bounding boxes with dimensions of $212 \\times 268$ pixels, accounting for approximately $18.49\\%$ of the total image area. In contrast, the GigaGrounding dataset exhibits an average bounding box size of $478 \\times 978$ pixels, representing just $0.12\\%$ of the image area. This statistic reveals the inherent challenges associated with grounding models tasked with large-scale scenes. We also checked the distribution of direct and multi-hop bounding boxes and noticed that they share a similar distribution pattern, while the bounding boxes corresponding to multi-hop expressions are slightly smaller, with an average resolution of $425 \\times 938$ , compared to $503 \\times 997$ of direct expression bounding boxes.", + "bbox": [ + 75, + 458, + 470, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lengths of expressions We demonstrate the word count distribution of expression lengths in Figure 5c. We note that the average expression length in RefCOCOg is 8.46 words, while in RefCOCO and RefCOCO+, it is only 3.50 and 3.53, respectively. In contrast, in order to locate objects in large-scale scenes without ambiguity, GigaGrounding provides significantly longer expressions, and the average expression length is 14.78 words. Flickr30K Entities present comparatively lengthy expressions, while it is tasked with", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ground objects referred by short region phrases. Figure 5d further illustrates the length distribution of direct expressions and multi-hop expressions, and the average lengths are 12.33 and 18.74, respectively. Longer expressions may accommodate more details about the objects' location, appearance, and properties, which can increase the semantic space that the model needs to consider, leading to challenges in both textual and imagery comprehension.", + "bbox": [ + 496, + 318, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Other statistics We notice that GigaGrounding is certainly not offering the most number of images, bounding boxes, and annotated expressions among the compared datasets (Table 1). However, we believe that GigaGrounding's unique characteristics have opened the door for new modeling paradigms with practically useful and broader real-life applications.", + "bbox": [ + 496, + 443, + 892, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 500, + 565, + 607, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present an initial step towards effective visual grounding in the complex context of gigapixel-level large-scale scenes. We introduce a simple yet effective model, referred to as GlaZing.", + "bbox": [ + 496, + 592, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Architecture Overview", + "text_level": 1, + "bbox": [ + 500, + 664, + 710, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the context of GigaGrounding, human operators commonly begin with an initial coarse search across the image to identify regions of interest. Following this initial search, they then allocate their attention to the selected regions, subjecting them to thorough scrutiny. This methodical narrowing of focus significantly increases search efficacy. Our GlaZing model is conceived by analogizing this human cognitive heuristic.", + "bbox": [ + 496, + 688, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of GlaZing is presented in Figure 6. Given an input image, it initiates by inputting the downsampled thumbnail and the corresponding expression into the Glance Grounding Module (GGM) for preliminary localization and grounding. Subsequently, the Adaptive Cropping Module (ACM) dynamically extracts the high-resolution patch", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22123", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/279ea6629a6b9826a5097a8052c2721d6aecb6700f74cd6f054a95ac9b5df704.jpg", + "image_caption": [ + "Figure 6. Overview of the proposed GlaZing model." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 467, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of interest. This extracted patch, along with the expression, is then forwarded to the Zoomed-in Grounding Module (ZGM) to generate the final prediction.", + "bbox": [ + 75, + 290, + 468, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Detailed Implementations", + "text_level": 1, + "bbox": [ + 76, + 347, + 312, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The architectural framework expounded in Section 4.1 allows for a multitude of implementations. Notably, GGM and ZGM can be seamlessly integrated with a wide array of conventional grounding models. In this section, we present our specific implementation, which yielded favorable results in our experiments.", + "bbox": [ + 75, + 371, + 468, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Glance grounding module We design GGM based on ReSC [28], which adopts an anchor-based grounding paradigm that divides the image into multiple grids and determines the grid to which the target box belongs. This design effectively locates the target-relevant region in the thumbnail image.", + "bbox": [ + 75, + 465, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adaptive cropping module This module is devised for the adaptive extraction of a region of interest from the original image, guided by the GGM's output. More precisely, ACM initiates this process by first computing the coordinates for a patch of specific size located at the center of the bounding box predicted by the GGM. It subsequently proceeds to crop the minimum rectangular region that encloses both this designated patch and the bounding box predicted by the GGM. During model training, this process also encapsulates the ground-truth bounding box. Finally, the extracted region is resized to a specified target size. Through the utilization of ACM, our model can adaptively focus on regions of varying scales and locations.", + "bbox": [ + 75, + 560, + 468, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zoomed-in grounding module After the first glance and zooming-in, the target bounding box should account for a significant proportion of the input patch of ZGM, so we employ a coordinate regression-based visual grounding method, i.e., TransVG [7] to accomplish this task.", + "bbox": [ + 75, + 760, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training objectives The overall training objective can be defined as $\\mathcal{L} = \\mathcal{L}_{\\mathrm{GGM}} + \\mathcal{L}_{\\mathrm{ZGM}}$ . We follow the training objectives employed in ReSC and TransVG for $\\mathcal{L}_{\\mathrm{GGM}}$ and $\\mathcal{L}_{\\mathrm{ZGM}}$ respectively. Specifically, $\\mathcal{L}_{\\mathrm{GGM}}$ comprises the YOLOv3", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "loss [45] designed to optimize anchor association confidence and prediction offset, and a diversity loss [28] intended to guide sub-queries in each iteration to concentrate on distinct elements within the query. Concerning $\\mathcal{L}_{\\mathrm{ZGM}}$ , given that ZGM directly generates a 4-dimensional vector representing the relative coordinates of the grounded target box with respect to the input patch, we employ both the smooth $\\ell 1$ loss and the generalized IoU loss [46] to minimize the disparity between the predicted and target values.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 500, + 239, + 617, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As an initial stride for GigaGrounding, GlaZing tackles the challenges posed by high-resolution input and scale variation through a \"glance-to-zoom-in\" cascading strategy coupled with adaptive cropping. Empirical findings in Section 5.3 attest to its effectiveness. Nevertheless, substantial opportunities for further refinement remain, which we leave as future work, including: (a) more tailored training objectives for GGM and ZGM, (b) end-to-end pipelines for seamless glancing and zooming-in, and (c) the feasibility of multiple iterations of glance and zoom adjustments.", + "bbox": [ + 496, + 263, + 890, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 431, + 630, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section delineates the evaluation protocols employed, followed by results' presentation and analysis.", + "bbox": [ + 498, + 458, + 890, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Baselines", + "text_level": 1, + "bbox": [ + 500, + 502, + 607, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted an evaluation of several publicly available baselines. As discussed in Section 2, two-stage methods rely on object detectors to generate region proposals. Given that object detection in large-scale scenes remains a challenge in its own right [40], we conducted brief experiments involving a representative open-source two-stage method, namely, MAttNet [47]. On the other hand, one-stage methods directly provide bounding box predictions without relying on separate object detection. In this regard, GlaZing is more aligned with one-stage methods. We evaluated state-of-the-art open-source one-stage methods, including ReSC [28], TransVG [7], RefTR [48], SeqTR [29], QR-Net [49], and SimREC [50]. A concise introduction to these baselines is provided in the Supplementary Material.", + "bbox": [ + 496, + 526, + 890, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to visual grounding models, we also conducted a human evaluation to assess the ability of human subjects to complete the task when presented with down-sampled images as used by deep learning models.", + "bbox": [ + 496, + 739, + 890, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 500, + 811, + 689, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset split We randomly selected $70\\%$ distinct expressions for training, $10\\%$ for validation, and $20\\%$ for testing. This process resulted in 42,641 samples for the training set, 6,254 for the validation set, and 12,458 for the testing set.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22124", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric Following the standard protocol, we assessed the performance using Precision@0.5, where the prediction is deemed correct if its IoU with the ground-truth box is larger than 0.5. We also provide information on inference time and GFLOPs to demonstrate the computational efficiency of each model.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image preprocessing In our empirical analysis, for the two-stage baseline MAttNet, features were extracted with 8-times downsampled images. We followed Yu et al. [47], who used ground-truth object bounding boxes to eliminate region proposal bottlenecks from the object detector.* On the other side, one-stage baseline models were evaluated using two distinct image resolutions: $640 \\times 640$ and $1536 \\times 1536$ . The first resolution was in accordance with established benchmarks, while the latter was the feasible upper limit for processing by most baseline models given our computational constraints. At $640 \\times 640$ , the mean bounding box dimensions were roughly $12 \\times 25$ pixels. This size expanded to $30 \\times 60$ pixels at $1536 \\times 1536$ resolution. To ensure fair comparison, for GlaZing, we employed a $640 \\times 640$ thumbnail image for the GGM, and the ACM cropped the region of interest from $1536 \\times 1536$ images. The cropped region was subsequently resized to $640 \\times 640$ for the ZGM to identify the referred region.", + "bbox": [ + 75, + 186, + 473, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details For all baseline methods, we conducted experiments based on their official code. However, for the object detection component of MAttNet, due to the lack of maintenance, we substituted it with the corresponding implementation of Faster R-CNN [51] from Detectron2 [52]. The substitution did not hurt the performance on RefCOCO based on our experiments. For one-stage methods, we tried to optimize their performance within a certain hyper-parameter tuning budget. For instance, for models with official weights on RefCOCO provided, we experimented with whether or not to use these weights for initialization and reported the better results. At the resolution of $1536 \\times 1536$ , we also tried initializing with weights from GigaGrounding at $640 \\times 640$ , and reported the better outcomes. We also adjusted some important hyperparameters including learning rate, etc. As for GlaZing, we trained the model with similar settings as ReSC. For all experiments, we set the maximum query length to 30, and the computations were performed on a server equipped with 8 NVIDIA RTX 3090 GPUs. Please refer to the Supplementary Material for more details.", + "bbox": [ + 75, + 463, + 472, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Overall Performance", + "text_level": 1, + "bbox": [ + 76, + 792, + 274, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the evaluation results of our approach and the baselines in Table 2. Several key observations and analyses are as follows:", + "bbox": [ + 75, + 816, + 468, + 862 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Baseline models' limited efficacy in GigaGrounding: None of the baseline models achieved Precision@0.5 scores surpassing $50\\%$ . In contrast, human testers consistently achieved scores exceeding $60\\%$ , even on images of size $640 \\times 640$ . It is also noteworthy that the performance on GigaGrounding dramatically diverged from that on other benchmark datasets, such as RefCOCO. For instance, the classical ReSC model outperformed other baselines on GigaGrounding, even surpassing models initialized with pretrained weights.", + "- Substantial advantages of GlaZing: GlaZing demonstrated a significant performance advantage over all baseline models. When compared to the highest-performing deep learning baseline at a resolution of $1536 \\times 1536$ (i.e., ReSC), GlaZing achieved an impressive relative improvement of $28.6\\%$ . This substantial gain can be attributed to its glance-to-zoom-in design.", + "- Incompatibility of existing two-stage methods: We noticed that MAttNet exhibited marginal improvements over random guessing (6.8% vs. 6.2%), despite that it used the ground-truth object bounding boxes as region proposals. Our hypothesis is that the bottleneck lies in the feature extractors. Two-stage methods typically use frozen feature extractors pretrained on MS COCO or ImageNet [53], which may not be effective enough for large-scale scene understanding." + ], + "bbox": [ + 500, + 90, + 890, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, our study reveals the following insights:", + "bbox": [ + 517, + 493, + 874, + 508 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Influence of resolution on performance: Our findings highlight the critical role of high-resolution input in enhancing the performance of both deep learning models and human evaluators.", + "- Performance variation with expression complexity: Notably, the GlaZing model demonstrated consistent performance across both direct and multi-hop expressions. In contrast, baseline models exhibited a marked performance decline when handling multi-hop expressions. This disparity underscores the need for models to adapt to expressions with differing levels of complexity." + ], + "bbox": [ + 500, + 512, + 890, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. In-depth Analysis", + "text_level": 1, + "bbox": [ + 500, + 689, + 671, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effective high-resolution processing of GlaZing. GlaZing's adaptive cropping mechanism allows for higher-resolution image processing without significant storage and computation overhead. Experiments with 8-times downsampled images in ACM and ZGM showed promising results: $66.5\\%$ on the test set, with $66.2\\%$ for direct and $67.5\\%$ for multi-hop expressions. These findings suggest that processing at higher resolutions may further enhance the model's performance.", + "bbox": [ + 496, + 715, + 890, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefits of RefCOCO initialization. We report the results of TransVG, RefTR, and SeqTR with and without RefCOCO weight initialization in Table 3. The adoption of", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "*Consequently, it is inappropriate to directly compare the results of one-stage and two-stage methods.", + "bbox": [ + 75, + 875, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22125", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7d6ffc4525491ed90a560e856b7c587d190b8883ac29dc8addb73e893acdaf0.jpg", + "table_caption": [], + "table_footnote": [ + "${}^{3}$ We used the released weights pretrained on RefCOCO for initialization. ${}^{8}$ Estimated with 100 randomly selected samples from the testing set." + ], + "table_body": "
CategoryMethodBackboneoverallVal directmulti-hopoverallTest directmulti-hopInfer TimeGFLOPsTrainable Params.
Two-stageMAttNetResNet-10111.4%11.6%10.7%6.8%6.4%7.6%268.6ms1592.613.0M
One-stage (640×640)ReSCDenseNet-5330.8%35.5%21.2%29.0%33.3%20.3%53.4ms101.3158.2M
QRNetSwin-S16.2%19.5%9.4%13.8%16.8%7.9%54.8ms79.2247.1M
SimRECCSPDarkNet-5323.3%27.6%14.4%20.8%24.7%12.7%16.5ms48.940.2M
TransVG‡ResNet-10113.1%15.9%7.1%11.6%13.7%7.3%25.9ms41.4122.6M
RefTR‡ResNet-5025.4%29.0%18.1%21.8%25.0%15.3%45.0ms23.5123.4M
SeqTR‡Darknet-5317.6%20.1%12.4%15.9%18.1%11.6%39.8ms48.9100.9M
Human------62%§----------
One-stage (1536×1536)ReSCDenseNet-5352.0%57.5%40.8%49.7%56.0%37.0%96.6ms566.3158.2M
QRNetSwin-S28.8%33.3%19.5%26.5%31.0%17.3%141.6ms435.0247.1M
SimRECCSPDarkNet-5335.4%41.7%22.5%33.3%39.5%20.8%48.4ms280.940.2M
TransVG‡ResNet-10126.3%31.2%16.5%24.6%29.1%15.4%72.6ms225.3122.6M
RefTR‡ResNet-5042.4%48.4%30.0%41.0%46.7%29.5%53.0ms119.7123.4M
SeqTR‡Darknet-5335.6%41.4%23.5%33.2%38.3%22.8%51.2ms263.7100.9M
Human------84%§----------
GlaZing (Ours)65.0%65.3%64.3%63.9%64.1%63.7%110.2ms172.1280.8M
", + "bbox": [ + 81, + 88, + 888, + 319 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bf381f7a3d802b2a420ba1bdc29d0c2883fcd4bc5e2e7434c078e0bb6b657016.jpg", + "table_caption": [ + "Table 2. Performance Evaluation on the GigaGrounding Dataset. Both accuracy-related and efficiency-related metrics are reported." + ], + "table_footnote": [], + "table_body": "
TransVGRefTRSeqTR
w/o pretraining5.1%0.0%5.3%
w/ pretraining11.6%21.8%15.9%
Absolute Gain+6.5%+21.8%+10.6%
", + "bbox": [ + 122, + 368, + 421, + 435 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2cd65dc615f5943f65ca1a5d88247bdbaa63469bb2e374dfad6800d4e2c139ad.jpg", + "table_caption": [ + "Table 3. Performance improvements on $640 \\times 640$ input images achieved through pretraining. We initiated each model with publicly available weights trained on the RefCOCO dataset." + ], + "table_footnote": [], + "table_body": "
Expr. length & proportionReSCTransVGGlaZing
1-10 (16.99%)65.9%42.4%66.8%
11-20 (62.77%)47.8%22.6%63.5%
21+ (20.24%)39.7%13.8%62.6%
", + "bbox": [ + 89, + 494, + 452, + 564 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6c70398b264650266ce55b76ec646b22c0f60d65218f061820d9a66cfcea7fbb.jpg", + "table_caption": [ + "Table 4. Performance stratification by expression length." + ], + "table_footnote": [], + "table_body": "
B-box scaleReSCTransVGGlaZing
s < Q129.4%4.5%48.1%
Q1 ≤ s < Q250.2%19.3%63.6%
Q2 ≤ s < Q358.2%31.1%72.8%
s ≥ Q360.7%43.6%70.9%
", + "bbox": [ + 119, + 603, + 424, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance stratification by bounding box scale. Here, $Q1$ , $Q2$ , and $Q3$ denote the first, second, and third quartiles of bounding box scale, respectively.", + "bbox": [ + 76, + 691, + 468, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pretrained weights from RefCOCO led to performance enhancements for three methods, despite the substantial differences between RefCOCO and GigaGrounding.", + "bbox": [ + 76, + 744, + 468, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance about expression length and B-box scale. We conducted an in-depth analysis of the performance variability about expression length and B-box scale at $1536 \\times 1536$ , specifically focusing on the ReSC, TransVG, and GlaZing. The results are presented in Tables 4 and 5. Our findings indicate that GlaZing demonstrated a notable robustness across varying difficulty levels. Conversely,", + "bbox": [ + 76, + 795, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TransVG exhibited a more pronounced performance degradation, particularly in scenarios involving longer expressions and smaller B-box sizes. Complete results of all baselines can be found in the Supplementary Material.", + "bbox": [ + 496, + 371, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the glance-to-zoom-in strategy in GlaZing. Our investigation reveals that GlaZing's ZGM effectively rectifies prediction errors observed in the GGM, underscoring the effectiveness of the glance-to-zoom-in approach for GigaGrounding. Specifically, we identified that approximately $54\\%$ of errors fall under \"target feature mismatch\", with ZGM correcting $48\\%$ of these. Additionally, another $20\\%$ of errors are categorized as \"over-fuzziness due to downsampling\", of which ZGM amends about $50\\%$ . Due to space constraints, detailed explanations and analyses are provided in the Supplementary Material.", + "bbox": [ + 496, + 435, + 892, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 613, + 617, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents GigaGrounding, a novel and challenging benchmark for visual grounding in gigapixel-level large-scale scenes, accompanied by an efficient solution strategy. The distinctive challenges posed by GigaGrounding establish it as a pioneering benchmark in this domain. Our goal is to stimulate further research in visual grounding, aiming to propel the development of models that are both more accurate and efficient. These advancements are particularly crucial for real-world applications such as rescue operations.", + "bbox": [ + 496, + 638, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work is supported in part by Natural Science Foundation of China (NSFC) under contract No. 62125106 and 62088102, in part by Ministry of Science and Technology of China under contract No. 2021ZD0109901, in part by Tsinghua-Zhijiang joint research center, in part by Young Elite Scientists Sponsorship Program by CAST under contract No. 2022QNRC001.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22126", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 3", + "[2] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 1, 2, 3, 5", + "[3] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 2, 3", + "[4] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 1, 2, 3", + "[5] Ronghang Hu, Huazhe Xu, Marcus Rohrbach, Jiashi Feng, Kate Saenko, and Trevor Darrell. Natural language object retrieval. In CVPR, 2016. 1", + "[6] Jianan Li, Yunchao Wei, Xiaodan Liang, Fang Zhao, Jianshu Li, Tingfa Xu, and Jiashi Feng. Deep attribute-preserving metric learning for natural language object retrieval. In ACM MM, 2017. 1", + "[7] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-end visual grounding with transformers. In ICCV, 2021. 1, 2, 6", + "[8] Xiaoyun Yuan, Mengqi Ji, Jiamin Wu, David J Brady, Qionghai Dai, and Lu Fang. A modular hierarchical array camera. Light: Sci. Appl., 2021. 1, 2", + "[9] Byford Sam. Samsung announces 200-megapixel phone camera sensor. https://www.theverge.com/2021/9/2/22653558/samsung-isocell-hpl-gn5-200-megapixel-camera-sensor-announced, 2021.1", + "[10] Arian Bakhtiarnia, Qi Zhang, and Alexandros Iosifidis. Efficient high-resolution deep learning: A survey. arXiv preprint arXiv:2207.13050, 2022. 1, 2, 3", + "[11] Xueyang Wang, Xiya Zhang, Yinheng Zhu, Yuchen Guo, Xiaoyun Yuan, Liuyu Xiang, Zerun Wang, Guiguang Ding, David Brady, Qionghai Dai, and Lu Fang. PANDA: A gigapixel-level human-centric video dataset. In CVPR, 2020. 1, 2, 3", + "[12] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2", + "[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A González, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated IAPR TC-12 benchmark. CVIU, 2010. 2", + "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 2", + "[15] Yuxi Sun, Shanshan Feng, Xutao Li, Yunming Ye, Jian" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kang, and Xu Huang. Visual grounding in remote sensing images. In ACM MM, 2022. 2", + "[16] Yang Zhan, Zhitong Xiong, and Yuan Yuan. RSVG: Exploring data and models for visual grounding on remote sensing data. TGRS, 61:1-13, 2023. 2", + "[17] Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, et al. Making the most of text semantics to improve biomedical vision-language processing. In ECCV, 2022. 2", + "[18] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 2", + "[19] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In ICCV, 2019. 2", + "[20] Richang Hong, Daqing Liu, Xiaoyu Mo, Xiangnan He, and Hanwang Zhang. Learning to compose and reason with language tree structures for visual grounding. TPAMI, 2019. 2", + "[21] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In CVPR, 2017.", + "[22] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In ICCV, 2019.", + "[23] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In CVPR, 2018.", + "[24] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In CVPR, 2019. 2", + "[25] Xinpeng Chen, Lin Ma, Jingyuan Chen, Zequn Jie, Wei Liu, and Jiebo Luo. Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426, 2018. 2", + "[26] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In CVPR, 2020.", + "[27] Arka Sadhu, Kan Chen, and Ram Nevatia. Zero-shot grounding of objects from natural language queries. In ICCV, 2019.", + "[28] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In ECCV, 2020. 6", + "[29] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. SeqTR: A simple yet universal network for visual grounding. In ECCV, 2022. 2, 6", + "[30] Jianing Zhang, Tianyi Zhu, Anke Zhang, Xiaoyun Yuan, Zihan Wang, Sebastian Beetschen, Lan Xu, Xing Lin, Qionghai Dai, and Lu Fang. Multiscale-VR: Multiscale gigapixel" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22127", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d panoramic videography for virtual reality. In ICCP, 2020. 2", + "[31] Xiaoyun Yuan, Lu Fang, Qionghai Dai, David J Brady, and Yebin Liu. Multiscale gigapixel video: A cross resolution image matching and warping approach. In ICCP, 2017. 2", + "[32] Jianing Zhang, Jinzhi Zhang, Shi Mao, Mengqi Ji, Guangyu Wang, Zequn Chen, Tian Zhang, Xiaoyun Yuan, Qionghai Dai, and Lu Fang. GigaMVS: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. TPAMI, 2021. 2", + "[33] Guangyu Wang, Jinzhi Zhang, Kai Zhang, Ruqi Huang, and Lu Fang. GiganticNVS: Gigapixel large-scale neural rendering with implicit meta-deformed manifold. TPAMI, 2023.", + "[34] Xueyang Wang, Xuecheng Chen, Puhua Jiang, Haozhe Lin, Xiaoyun Yuan, Mengqi Ji, Yuchen Guo, Ruqi Huang, and Lu Fang. The group interaction field for learning and explaining pedestrian anticipation. Engineering, 2023. 2", + "[35] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In ECCV, 2018. 3", + "[36] Dmitrii Marin, Zijian He, Peter Vajda, Priyam Chatterjee, Sam Tsai, Fei Yang, and Yuri Boykov. Efficient segmentation: Learning downsampling near semantic boundaries. In ICCV, 2019.", + "[37] Chen Jin, Ryutaro Tanno, Thomy Mertzanidou, Eleftheria Panagiotaki, and Daniel C Alexander. Learning to down-sample for segmentation of ultra-high resolution images. In ICLR, 2022. 3", + "[38] Mingfei Gao, Ruichi Yu, Ang Li, Vlad I Morariu, and Larry S Davis. Dynamic zoom-in network for fast object detection in large images. In CVPR, 2018. 3", + "[39] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020.", + "[40] Kai Chen, Zerun Wang, Xueyang Wang, Dahan Gong, Longlong Yu, Yuchen Guo, and Guiguang Ding. Towards real-time object detection in gigapixel-level video. Neurocomputing, 2022. 3, 6", + "[41] Maria Tzelepi and Anastasios Tefas. Improving the performance of lightweight CNNs for binary classification using quadratic mutual information regularization. Pattern Recogn., 2020. 3", + "[42] Huangjing Lin, Hao Chen, Simon Graham, Qi Dou, Nasir Rajpoot, and Pheng-Ann Heng. Fast ScanNet: Fast and dense analysis of multi-gigapixel whole-slide images for cancer metastasis detection. TMI, 2019. 3", + "[43] Using Creative Commons Public Licenses. Attribution-noncommercial-sharealike 4.0 international. 3", + "[44] Baowei Jiang, Bing Bai, Haozhe Lin, Yu Wang, Yuchen Guo, and Lu Fang. DartBlur: Privacy preservation with detection artifact suppression. In CVPR, 2023. 3", + "[45] Joseph Redmon and Ali Farhadi. YOLOv3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 6", + "[46] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. MAttNet: Modular attention network for referring expression comprehension. In CVPR, 2018. 6, 7", + "[48] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, 2021. 6", + "[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In CVPR, 2022. 6", + "[50] Gen Luo, Yiyi Zhou, Jiamu Sun, Shubin Huang, Xiaoshuai Sun, Qixiang Ye, Yongjian Wu, and Rongrong Ji. What goes beyond multi-modal fusion in one-stage referring expression comprehension: An empirical study. arXiv preprint arXiv:2204.07913, 2022. 6", + "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 7", + "[52] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.7", + "[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7" + ], + "bbox": [ + 501, + 92, + 890, + 486 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22128", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_model.json b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ffa2ec8785a691611361aab4f5d1cc806a1d984 --- /dev/null +++ b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_model.json @@ -0,0 +1,2783 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.13, + 0.82, + 0.176 + ], + "angle": 0, + "content": "When Visual Grounding Meets Gigapixel-level Large-scale Scenes: Benchmark and Approach" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.203, + 0.823, + 0.24 + ], + "angle": 0, + "content": "Tao Ma\\(^{1*}\\), Bing Bai\\(^{1*}\\), Haozhe Lin\\(^{1*}\\), Heyuan Wang\\(^{2}\\), Yu Wang\\(^{1}\\), Lin Luo\\(^{2}\\), Lu Fang\\(^{1\\dagger}\\), Tsinghua University \\({}^{2}\\)Peking University" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.243, + 0.581, + 0.256 + ], + "angle": 0, + "content": "fanglu@tsinghua.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.275, + 0.299, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.424, + 0.228, + 0.438 + ], + "angle": 0, + "content": "(a) RefCOCO" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.277, + 0.648, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.424, + 0.669, + 0.438 + ], + "angle": 0, + "content": "(b) GigaGrounding" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.277, + 0.882, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.407, + 0.798, + 0.419 + ], + "angle": 0, + "content": "\\(\\nrightarrow\\) i. lower left white car" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.448, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 1. Comparison of the proposed GigaGrounding and the RefCOCO benchmark. GigaGrounding presents distinctive challenges, including large-scale scene understanding, high-resolution with significant scale variation, and multi-hop expressions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.489, + 0.314, + 0.505 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.524, + 0.474, + 0.841 + ], + "angle": 0, + "content": "Visual grounding refers to the process of associating natural language expressions with corresponding regions within an image. Existing benchmarks for visual grounding primarily operate within small-scale scenes with a few objects. Nevertheless, recent advances in imaging technology have enabled the acquisition of gigapixel-level images, providing high-resolution details in large-scale scenes containing numerous objects. To bridge this gap between imaging and computer vision benchmarks and make grounding more practically valuable, we introduce a novel dataset, named GigaGrounding, designed to challenge visual grounding models in gigapixel-level large-scale scenes. We extensively analyze and compare the dataset with existing benchmarks, demonstrating that GigaGrounding presents unique challenges such as large-scale scene understanding, gigapixel-level resolution, significant variations in object scales, and the \"multi-hop expressions\". Furthermore, we introduced a simple yet effective grounding approach, which employs a \"glance-to-zoom-in\" paradigm and exhibits enhanced capabilities for addressing the GigaGrounding task. The dataset is available at www.gigavision.ai." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.489, + 0.631, + 0.505 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.893, + 0.67 + ], + "angle": 0, + "content": "Visual grounding (VG), also known as referring expression comprehension [1, 2], phrase localization [3, 4], and natural language object retrieval [5, 6], aims to identify the region within an image that corresponds to a given natural language expression. This field offers considerable potential for enhancing our capacity to connect the human language and visual elements of the tangible world [7]. The application of VG in scenarios such as rescue and robotics can prove to be exceedingly valuable, for example, in searching for a missing child in crowds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.901 + ], + "angle": 0, + "content": "While the VG community has made significant progress in recent years, the benchmarks employed thus far have been predominantly limited to small-scale scenes [1-4]. As illustrated in Figure 1a, the number of objects is limited, and the objects cover the dominant portion of the image. However, recent advances in imaging technology, such as array cameras [8] and even consumer-grade smartphones [9], have enabled the capture of gigapixel-level photos, providing high-resolution detail information while covering large-scale scenes up to \\(\\mathrm{km}^2\\)-level [10], as shown in Figure 1b. The characteristics of conventional VG benchmarks have restricted current models' application to relatively simple and vanilla settings, hindering the development of new technologies and limiting the applicability of grounding in broader contexts [11]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.436, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work. \n†Lu Fang is the corresponding author (www.luvision.net)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22119" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "To address this limitation, we introduce a novel benchmark dataset, named GigaGrounding, for visual grounding in large-scale scenes. The dataset is constructed based on PANDA [11], a gigapixel-level human-centric video dataset. An illustration of the GigaGrounding dataset is displayed in Figure 1b. In contrast to conventional visual grounding benchmarks such as RefCOCO [2], GigaGrounding poses unique and non-trivial challenges:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.222, + 0.47, + 0.327 + ], + "angle": 0, + "content": "- Large-scale scene understanding. Large-scale scenes require models to deal with a much larger spatial and semantic context and a significantly larger number of objects in the image. Hundreds of objects can be distributed over a vast area with possible occlusions, clutter, and other distractions, making it a difficult task to recognize and understand large-scale scenes." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.331, + 0.47, + 0.481 + ], + "angle": 0, + "content": "- High resolution with significant scale variation. Processing high-throughput data up to gigapixel-level images poses new challenges to models in terms of efficiency [10]. Moreover, objects may be presented at drastically different scales. For instance, as shown in Figure 1b, the ground-truth bounding box (B-box) for Case i spans approximately \\(3000 \\times 1400\\) pixels, whereas, for Case ii, it spans only about \\(110 \\times 330\\) pixels. This significant variation necessitates models to be adaptable in scaling to accommodate diverse object sizes." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.484, + 0.47, + 0.62 + ], + "angle": 0, + "content": "- Multi-hop expressions. In addition to direct expressions, which refer to descriptions of independent objects as illustrated in Case i, GigaGrounding also includes a considerable number of multi-hop expressions. These expressions challenge grounding models to locate target objects by first identifying some other reference objects. For example, in Case ii, the model needs first to identify \"the middle lawn\" and then zoom in to locate the correct person based on the relationship with the lawn." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.222, + 0.47, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We benchmark the performance of cutting-edge VG models with the GigaGrounding dataset. Empirical findings reveal that while existing VG models have demonstrated satisfactory performance in small-scale scenes, they face distinct and intricate challenges when being applied to GigaGrounding. To address these challenges, we propose GlaZing (GLance-to-Zoom-IN Grounding), a simple yet effective model inspired by human cognitive processes. GlaZing employs a \"glance-to-zoom-in\" cascading strategy to ground target bounding boxes in high-resolution images. It begins by taking a global view of the thumbnail image, to identify the region of interest based on the given expression. Subsequently, it zooms in on the local region with adaptive scaling for the final grounding process. Experiments demonstrated that GlaZing outperformed existing models in the GigaGrounding dataset, effectively addressing major challenges such as high resolution, significant scale variation, and multi-hop expressions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.887, + 0.108 + ], + "angle": 0, + "content": "In conclusion, this paper presents several contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.112, + 0.892, + 0.173 + ], + "angle": 0, + "content": "- We highlight the significance of deploying visual grounding models in large-scale scenes, and introduce a novel benchmark dataset, GigaGrounding, characterized by unique properties that set it apart from existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.176, + 0.892, + 0.22 + ], + "angle": 0, + "content": "- We empirically assess current VG models using the proposed GigaGrounding dataset, highlighting the limitations of state-of-the-art models in this context." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.224, + 0.892, + 0.299 + ], + "angle": 0, + "content": "- We introduce GlaZing, a tailored model for GigaGrounding, which leverages a \"glance-to-zoom-in\" approach. Our experimental results demonstrate the efficacy of this approach in addressing the current challenges, while also indicating room for further improvements." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.112, + 0.892, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.315, + 0.642, + 0.332 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.374 + ], + "angle": 0, + "content": "In this section, we review two relevant research domains: visual grounding and high-resolution deep learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.385, + 0.676, + 0.401 + ], + "angle": 0, + "content": "2.1. Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Visual grounding is a field focused on predicting the location within an image that corresponds to a natural language expression. Benchmark datasets for VG have predominantly centered on small-scale scenes. For example, the Flickr30K Entities dataset [4] extends the original Flickr30K dataset [12] by incorporating annotations that establish correspondence between short region phrases and images. ReferItGame [3] comprises 20,000 images collected from the SAIAPR-12 dataset [13]. RefCOCO [2], RefCOCO+ [2], and RefCOCOg [1] are all built upon the MS COCO dataset [14]. Notably, the image resolutions in these datasets do not exceed \\(640 \\times 480\\). In addition, there are related benchmarks in fields such as remote sensing [15, 16] and biomedical science [17, 18], with their resolutions also limited to \\(1024 \\times 1024\\) or lower." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.773 + ], + "angle": 0, + "content": "For VG models, existing work can be broadly categorized into two-stage and one-stage methods [7, 19]. Two-stage methods [20-24] initiate the process by generating a set of region proposals using object detectors. Subsequently, they leverage natural language expressions to rank these proposals and identify the best matching proposal. In contrast, one-stage methods [7, 19, 25-29] densely integrate both visual and textual information and directly output bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.776, + 0.803 + ], + "angle": 0, + "content": "2.2. High-resolution Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In recent years, significant progress has been made in the field of imaging, leading to gigapixel-level images with both a wide FoV and high resolution [8, 30, 31]. Employing high-resolution images and videos as direct inputs for deep learning models introduces challenges in both training and inference stages [11, 32-34]. To address these" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22120" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.175 + ], + "angle": 0, + "content": "
DatasetTypical res.# Images# B-boxes# Expressions
Flickr30K Entities [4]500×37531,783276k427k
ReferItGame [3]480×36019,89496,654130,525
RefCOCO [2]640×48019,99450,000142,210
RefCOCO+ [2]640×48019,99249,856141,564
RefCOCOg [1]640×48025,79949,82295,010
GigaGrounding>25k×14k3,77561,35361,353
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.18, + 0.46, + 0.194 + ], + "angle": 0, + "content": "Table 1. Statistics of existing benchmarks and GigaGrounding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.47, + 0.381 + ], + "angle": 0, + "content": "challenges, several approaches have been explored. Rudimentary strategies like \"uniform downsampling (UD)\" and \"cutting into patches (CIP)\" may be employed, but they may result in the loss of detailed information and relational cues, respectively [10]. Alternatively, more sophisticated techniques, including non-uniform downsampling [35-37], selective zooming and skipping [38-40], and the use of lightweight scanner networks [41, 42], have been proposed. Bakhtiarnia et al. [10] have provided a comprehensive review of these methods, their relevant applications, and datasets pertaining to high-resolution deep learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.338, + 0.412 + ], + "angle": 0, + "content": "3. GigaGrounding Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.47, + 0.482 + ], + "angle": 0, + "content": "This section offers a comprehensive delineation of the GigaGrounding benchmark, including an overview, an elucidation of the data collection strategy, and analyses of its unique properties." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.492, + 0.334, + 0.509 + ], + "angle": 0, + "content": "3.1. Overview of GigaGrounding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.745 + ], + "angle": 0, + "content": "Dataset overview GigaGrounding is built upon images pre-extracted from PANDA-Video dataset by the dataset providers [11]. A summary of key statistics comparing conventional VG benchmarks with GigaGrounding is provided in Table 1. The analysis reveals a significant limitation in conventional VG datasets, where image resolutions are constrained to the megapixel level. In contrast, GigaGrounding ambitiously extends its scope to gigapixel-level images, encompassing intricate details and large-scale scenes. Consequently, the total number of images naturally decreases. In terms of bounding boxes, GigaGrounding offers a quantity comparable to other benchmark datasets. Additionally, in traditional benchmarks, multiple expressions may correspond to a single bounding box, while in GigaGrounding, each bounding box is associated with only one expression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.468, + 0.807 + ], + "angle": 0, + "content": "We illustrate representative samples from conventional VG datasets in Figure 2, and also provide examples of GigaGrounding annotations in Figure 3. Specifically, GigaGrounding includes two types of annotated expressions:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.468, + 0.902 + ], + "angle": 0, + "content": "- Direct expressions: These expressions directly describe each target object without introducing dependencies on other objects. Position modifiers, such as \"in the top left of the image\", may be used when describing small and challenging-to-locate objects, but only when necessary. In total, there are 41,040 direct expressions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "- Multi-hop expressions: These expressions require grounding models to respect additional restrictive constraints dependent on their relationship with other reference objects. In other words, the model needs to locate objects by first identifying specific reference objects and then finding the target objects based on constraints defined by relative positions (e.g., \"to the left\") or relationships (e.g., \"watching\", \"holding\") with those reference objects. There are 20,313 multi-hop expressions in total. It is noteworthy that we did not identify a design intentionally analogous to multi-hop expressions in traditional VG benchmarks. Notably, RefCOCO+ was explicitly designed to \"focus on purely appearance-based description\" [2]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.355 + ], + "angle": 0, + "content": "For more details regarding the PANDA-Video dataset and additional examples of GigaGrounding, please refer to the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.541 + ], + "angle": 0, + "content": "Privacy GigaGrounding is built based on the publicly available PANDA [11] dataset. PANDA is collected in public areas where photography is officially approved and is published under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License [43]. GigaGrounding has been created by further processing the data in PANDA with additional annotations. To protect individual privacy, PANDA has already anonymized the faces in the images. However, it is necessary to acknowledge that certain individuals may still be identifiable based on other less invasive characteristics, such as clothing or body shape [44]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.725, + 0.57 + ], + "angle": 0, + "content": "3.2. Data Collection Strategy" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.891, + 0.624 + ], + "angle": 0, + "content": "We have devised the following annotation procedure and quality control strategies to diligently uphold non-ambiguity and maintain high annotation quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Annotation protocol The annotation protocol, as depicted in Figure 4, comprises the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.663, + 0.892, + 0.709 + ], + "angle": 0, + "content": "1. Given the gigapixel images, Group \\( A \\) of workers was asked to write natural language expressions and draw bounding boxes for the corresponding objects." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.712, + 0.892, + 0.772 + ], + "angle": 0, + "content": "2. Given the gigapixel images and natural language expressions from Step 1, Group \\( B \\) of workers was then asked to draw corresponding bounding boxes within 40 seconds independently." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.775, + 0.892, + 0.85 + ], + "angle": 0, + "content": "3. If the Intersection over Union (IoU) rate of bounding boxes from Steps 1 and 2 was greater than 0.5, the sample were deemed valid. Then the valid expressions were manually translated into English with the highest possible fidelity." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.663, + 0.892, + 0.85 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This procedure tries to ensure non-ambiguity, by excluding data instances with object selection discrepancies between Steps 1 and 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.209, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.182, + 0.196, + 0.209 + ], + "angle": 0, + "content": "A man with pierced ears is wearing glasses and an orange hat" + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.217, + 0.203, + 0.228 + ], + "angle": 0, + "content": "(a) Flickr30K Entities" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.37, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.182, + 0.337, + 0.19 + ], + "angle": 0, + "content": "1. white shirt on right" + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.191, + 0.321, + 0.199 + ], + "angle": 0, + "content": "2. white shirt man" + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.2, + 0.309, + 0.208 + ], + "angle": 0, + "content": "3. man on right" + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.217, + 0.335, + 0.228 + ], + "angle": 0, + "content": "(b) ReferItGame" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.089, + 0.53, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.181, + 0.522, + 0.189 + ], + "angle": 0, + "content": "1. woman on right in white shirt" + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.19, + 0.461, + 0.199 + ], + "angle": 0, + "content": "2. woman on right" + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.2, + 0.447, + 0.208 + ], + "angle": 0, + "content": "3. right women" + }, + { + "type": "image_caption", + "bbox": [ + 0.414, + 0.217, + 0.489, + 0.228 + ], + "angle": 0, + "content": "(c) RefCOCO" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.09, + 0.707, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.182, + 0.687, + 0.189 + ], + "angle": 0, + "content": "1. yellow shirt and black shorts" + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.19, + 0.676, + 0.199 + ], + "angle": 0, + "content": "2. guy in yellow dibbling ball" + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.2, + 0.647, + 0.208 + ], + "angle": 0, + "content": "3. yellow shirt in focus" + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.217, + 0.661, + 0.228 + ], + "angle": 0, + "content": "(d) RefCOCO+" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.089, + 0.89, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.182, + 0.885, + 0.189 + ], + "angle": 0, + "content": "1. a woman wearing red playing tennis" + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.19, + 0.868, + 0.209 + ], + "angle": 0, + "content": "2. a woman in a red shirt holding a tennis racquet" + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.217, + 0.841, + 0.228 + ], + "angle": 0, + "content": "(e) RefCOCOg" + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.23, + 0.688, + 0.247 + ], + "angle": 0, + "content": "Figure 2. Example data from conventional visual grounding datasets." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.274, + 0.136, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.283, + 0.246, + 0.314 + ], + "angle": 0, + "content": "man in a red shirt pushing a suitcase near a subway station" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.325, + 0.15, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.333, + 0.259, + 0.364 + ], + "angle": 0, + "content": "woman in a yellow top riding a tricycle in the middle of the picture" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.375, + 0.13, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.381, + 0.269, + 0.413 + ], + "angle": 0, + "content": "man in white jacket and black bag in front of the mall on the left of the picture" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.424, + 0.138, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.431, + 0.259, + 0.462 + ], + "angle": 0, + "content": "man in a white coat riding a blue bike at the bottom left of the picture" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.274, + 0.515, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.274, + 0.691, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.275, + 0.787, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.282, + 0.882, + 0.314 + ], + "angle": 0, + "content": "blue billboard in the top right corner of the picture" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.325, + 0.766, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.333, + 0.872, + 0.364 + ], + "angle": 0, + "content": "green spherical plant in the flower bed on the right of the picture" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.375, + 0.796, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.392, + 0.881, + 0.401 + ], + "angle": 0, + "content": "red and white bus" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.424, + 0.748, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.436, + 0.855, + 0.456 + ], + "angle": 0, + "content": "taxi closest to the red and white bus" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.489, + 0.89, + 0.504 + ], + "angle": 0, + "content": "Figure 3. Example annotation results of GigaGrounding. Blue boxes indicate direct expressions and orange ones are multi-hop expressions." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.526, + 0.447, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.641, + 0.428, + 0.655 + ], + "angle": 0, + "content": "Figure 4. The overview of our annotation procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.468, + 0.71 + ], + "angle": 0, + "content": "Quality control strategies We employ the following strategies for quality control." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.717, + 0.469, + 0.807 + ], + "angle": 0, + "content": "- Terminology standardization. We introduced a standardized lexicon to encourage uniform terminology usage among annotators, reducing the prevalence of ambiguous or domain-specific terms. For example, the use of specific color names like \"gamboge\" and the adoption of specialized terminology such as \"coupe\" should be avoided." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Sampling inspection and repair. We performed a sampling inspection for each annotation batch to assess its quality. Specifically, for expression and bounding box annotations, we established a sampling pass rate threshold of \\(96\\%\\), whereas, for expression translation, the threshold was set to \\(99\\%\\). The sampling inspection process" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.717, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.891, + 0.569 + ], + "angle": 0, + "content": "included no fewer than 100 samples per round, and the batch of data was accepted once it met the pass rate threshold or underwent repair." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.571, + 0.892, + 0.736 + ], + "angle": 0, + "content": "- Box ensemble and de-duplication. In Step 3 of the annotation procedure, we obtained the ground truth by averaging the valid bounding boxes annotated by both groups of annotators. Any expressions that did not result in a valid bounding box were discarded. Furthermore, as data annotation is carried out in parallel among many annotation workers, we performed a dedduplication procedure to eliminate objects with overlapping bounding boxes and identical expressions across multiple images to enhance the data variance, with an IoU threshold empirically established at 0.3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.742, + 0.891, + 0.788 + ], + "angle": 0, + "content": "Despite potential undetected ambiguities, the implemented methodologies are believed to substantially mitigate their effects, thereby enhancing the dataset's overall quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.75, + 0.814 + ], + "angle": 0, + "content": "3.3. Analyses of GigaGrounding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.822, + 0.891, + 0.867 + ], + "angle": 0, + "content": "This section analyzes the salient characteristics differentiating GigaGrounding from existing VG benchmarks. Our analysis encompasses the following viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Quantities of objects per image Among conventional datasets, Flickr30K Entities in Figure 2a requires the model" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.276, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.2, + 0.28, + 0.224 + ], + "angle": 0, + "content": "(a) Pixel-count distribution of bounding box scales." + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.091, + 0.48, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.2, + 0.482, + 0.224 + ], + "angle": 0, + "content": "(b) Relative area of bounding boxes w.r.t. the images." + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.677, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.2, + 0.678, + 0.224 + ], + "angle": 0, + "content": "(c) Word-count distribution of expression lengths." + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.091, + 0.888, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.2, + 0.892, + 0.224 + ], + "angle": 0, + "content": "(d) Word-count distribution of direct and multi-hop expressions." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Figure 5. A comparative analysis of GigaGrounding against established benchmarks RefCOCO, RefCOCO+, and RefCOCOg. (a) and (b) elucidate the distribution patterns of bounding box scale (in terms of the max value of height and width) and relative area with respect to the image, (c) presents the distribution of expression lengths, and (d) shows the lengths' distribution of both direct and multi-hop expressions within the GigaGrounding dataset. GigaGrounding exhibits a notably elevated level of complexity, primarily stemming from its pronounced scale variation and intricate expression structures." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.319, + 0.47, + 0.456 + ], + "angle": 0, + "content": "to identify and locate multiple objects corresponding to short phrases within a single sentence. However, we observe that the model may not need to fully understand the global semantics before grounding the corresponding object, particularly when there exists only one object of that type. On the other side, RefCOCO and RefCOCO+ include an average of 3.9 same-type objects per image, while RefCOCOg displays a relatively lower average of 1.63 same-type objects per image [2]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.471, + 0.761 + ], + "angle": 0, + "content": "Scales of bounding boxes In Figures 5a and 5b, we respectively illustrate the absolute scale of bounding boxes and their relative area with respect to the images. It is evident that GigaGrounding exhibits greater scale variance in comparison to other datasets. Furthermore, bounding box scale in GigaGrounding adheres to a long-tail distribution, while other datasets generally show a more concentrated distribution. On average, the RefCOCO dataset features bounding boxes with dimensions of \\(212 \\times 268\\) pixels, accounting for approximately \\(18.49\\%\\) of the total image area. In contrast, the GigaGrounding dataset exhibits an average bounding box size of \\(478 \\times 978\\) pixels, representing just \\(0.12\\%\\) of the image area. This statistic reveals the inherent challenges associated with grounding models tasked with large-scale scenes. We also checked the distribution of direct and multi-hop bounding boxes and noticed that they share a similar distribution pattern, while the bounding boxes corresponding to multi-hop expressions are slightly smaller, with an average resolution of \\(425 \\times 938\\), compared to \\(503 \\times 997\\) of direct expression bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Lengths of expressions We demonstrate the word count distribution of expression lengths in Figure 5c. We note that the average expression length in RefCOCOg is 8.46 words, while in RefCOCO and RefCOCO+, it is only 3.50 and 3.53, respectively. In contrast, in order to locate objects in large-scale scenes without ambiguity, GigaGrounding provides significantly longer expressions, and the average expression length is 14.78 words. Flickr30K Entities present comparatively lengthy expressions, while it is tasked with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.44 + ], + "angle": 0, + "content": "ground objects referred by short region phrases. Figure 5d further illustrates the length distribution of direct expressions and multi-hop expressions, and the average lengths are 12.33 and 18.74, respectively. Longer expressions may accommodate more details about the objects' location, appearance, and properties, which can increase the semantic space that the model needs to consider, leading to challenges in both textual and imagery comprehension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.893, + 0.55 + ], + "angle": 0, + "content": "Other statistics We notice that GigaGrounding is certainly not offering the most number of images, bounding boxes, and annotated expressions among the compared datasets (Table 1). However, we believe that GigaGrounding's unique characteristics have opened the door for new modeling paradigms with practically useful and broader real-life applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.608, + 0.584 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.653 + ], + "angle": 0, + "content": "In this section, we present an initial step towards effective visual grounding in the complex context of gigapixel-level large-scale scenes. We introduce a simple yet effective model, referred to as GlaZing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.712, + 0.68 + ], + "angle": 0, + "content": "4.1. Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.809 + ], + "angle": 0, + "content": "In the context of GigaGrounding, human operators commonly begin with an initial coarse search across the image to identify regions of interest. Following this initial search, they then allocate their attention to the selected regions, subjecting them to thorough scrutiny. This methodical narrowing of focus significantly increases search efficacy. Our GlaZing model is conceived by analogizing this human cognitive heuristic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "An overview of GlaZing is presented in Figure 6. Given an input image, it initiates by inputting the downsampled thumbnail and the corresponding expression into the Glance Grounding Module (GGM) for preliminary localization and grounding. Subsequently, the Adaptive Cropping Module (ACM) dynamically extracts the high-resolution patch" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.468, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.256, + 0.428, + 0.272 + ], + "angle": 0, + "content": "Figure 6. Overview of the proposed GlaZing model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.338 + ], + "angle": 0, + "content": "of interest. This extracted patch, along with the expression, is then forwarded to the Zoomed-in Grounding Module (ZGM) to generate the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.313, + 0.365 + ], + "angle": 0, + "content": "4.2. Detailed Implementations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.463 + ], + "angle": 0, + "content": "The architectural framework expounded in Section 4.1 allows for a multitude of implementations. Notably, GGM and ZGM can be seamlessly integrated with a wide array of conventional grounding models. In this section, we present our specific implementation, which yielded favorable results in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.557 + ], + "angle": 0, + "content": "Glance grounding module We design GGM based on ReSC [28], which adopts an anchor-based grounding paradigm that divides the image into multiple grids and determines the grid to which the target box belongs. This design effectively locates the target-relevant region in the thumbnail image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.756 + ], + "angle": 0, + "content": "Adaptive cropping module This module is devised for the adaptive extraction of a region of interest from the original image, guided by the GGM's output. More precisely, ACM initiates this process by first computing the coordinates for a patch of specific size located at the center of the bounding box predicted by the GGM. It subsequently proceeds to crop the minimum rectangular region that encloses both this designated patch and the bounding box predicted by the GGM. During model training, this process also encapsulates the ground-truth bounding box. Finally, the extracted region is resized to a specified target size. Through the utilization of ACM, our model can adaptively focus on regions of varying scales and locations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Zoomed-in grounding module After the first glance and zooming-in, the target bounding box should account for a significant proportion of the input patch of ZGM, so we employ a coordinate regression-based visual grounding method, i.e., TransVG [7] to accomplish this task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Training objectives The overall training objective can be defined as \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{GGM}} + \\mathcal{L}_{\\mathrm{ZGM}}\\). We follow the training objectives employed in ReSC and TransVG for \\(\\mathcal{L}_{\\mathrm{GGM}}\\) and \\(\\mathcal{L}_{\\mathrm{ZGM}}\\) respectively. Specifically, \\(\\mathcal{L}_{\\mathrm{GGM}}\\) comprises the YOLOv3" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "loss [45] designed to optimize anchor association confidence and prediction offset, and a diversity loss [28] intended to guide sub-queries in each iteration to concentrate on distinct elements within the query. Concerning \\(\\mathcal{L}_{\\mathrm{ZGM}}\\), given that ZGM directly generates a 4-dimensional vector representing the relative coordinates of the grounded target box with respect to the input patch, we employ both the smooth \\(\\ell 1\\) loss and the generalized IoU loss [46] to minimize the disparity between the predicted and target values." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.24, + 0.618, + 0.255 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.416 + ], + "angle": 0, + "content": "As an initial stride for GigaGrounding, GlaZing tackles the challenges posed by high-resolution input and scale variation through a \"glance-to-zoom-in\" cascading strategy coupled with adaptive cropping. Empirical findings in Section 5.3 attest to its effectiveness. Nevertheless, substantial opportunities for further refinement remain, which we leave as future work, including: (a) more tailored training objectives for GGM and ZGM, (b) end-to-end pipelines for seamless glancing and zooming-in, and (c) the feasibility of multiple iterations of glance and zoom adjustments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.632, + 0.451 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.891, + 0.49 + ], + "angle": 0, + "content": "This section delineates the evaluation protocols employed, followed by results' presentation and analysis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.608, + 0.517 + ], + "angle": 0, + "content": "5.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.738 + ], + "angle": 0, + "content": "We conducted an evaluation of several publicly available baselines. As discussed in Section 2, two-stage methods rely on object detectors to generate region proposals. Given that object detection in large-scale scenes remains a challenge in its own right [40], we conducted brief experiments involving a representative open-source two-stage method, namely, MAttNet [47]. On the other hand, one-stage methods directly provide bounding box predictions without relying on separate object detection. In this regard, GlaZing is more aligned with one-stage methods. We evaluated state-of-the-art open-source one-stage methods, including ReSC [28], TransVG [7], RefTR [48], SeqTR [29], QR-Net [49], and SimREC [50]. A concise introduction to these baselines is provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.891, + 0.801 + ], + "angle": 0, + "content": "In addition to visual grounding models, we also conducted a human evaluation to assess the ability of human subjects to complete the task when presented with down-sampled images as used by deep learning models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.691, + 0.828 + ], + "angle": 0, + "content": "5.2. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset split We randomly selected \\(70\\%\\) distinct expressions for training, \\(10\\%\\) for validation, and \\(20\\%\\) for testing. This process resulted in 42,641 samples for the training set, 6,254 for the validation set, and 12,458 for the testing set." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "Metric Following the standard protocol, we assessed the performance using Precision@0.5, where the prediction is deemed correct if its IoU with the ground-truth box is larger than 0.5. We also provide information on inference time and GFLOPs to demonstrate the computational efficiency of each model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.474, + 0.46 + ], + "angle": 0, + "content": "Image preprocessing In our empirical analysis, for the two-stage baseline MAttNet, features were extracted with 8-times downsampled images. We followed Yu et al. [47], who used ground-truth object bounding boxes to eliminate region proposal bottlenecks from the object detector.* On the other side, one-stage baseline models were evaluated using two distinct image resolutions: \\(640 \\times 640\\) and \\(1536 \\times 1536\\). The first resolution was in accordance with established benchmarks, while the latter was the feasible upper limit for processing by most baseline models given our computational constraints. At \\(640 \\times 640\\), the mean bounding box dimensions were roughly \\(12 \\times 25\\) pixels. This size expanded to \\(30 \\times 60\\) pixels at \\(1536 \\times 1536\\) resolution. To ensure fair comparison, for GlaZing, we employed a \\(640 \\times 640\\) thumbnail image for the GGM, and the ACM cropped the region of interest from \\(1536 \\times 1536\\) images. The cropped region was subsequently resized to \\(640 \\times 640\\) for the ZGM to identify the referred region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.473, + 0.781 + ], + "angle": 0, + "content": "Implementation details For all baseline methods, we conducted experiments based on their official code. However, for the object detection component of MAttNet, due to the lack of maintenance, we substituted it with the corresponding implementation of Faster R-CNN [51] from Detectron2 [52]. The substitution did not hurt the performance on RefCOCO based on our experiments. For one-stage methods, we tried to optimize their performance within a certain hyper-parameter tuning budget. For instance, for models with official weights on RefCOCO provided, we experimented with whether or not to use these weights for initialization and reported the better results. At the resolution of \\(1536 \\times 1536\\), we also tried initializing with weights from GigaGrounding at \\(640 \\times 640\\), and reported the better outcomes. We also adjusted some important hyperparameters including learning rate, etc. As for GlaZing, we trained the model with similar settings as ReSC. For all experiments, we set the maximum query length to 30, and the computations were performed on a server equipped with 8 NVIDIA RTX 3090 GPUs. Please refer to the Supplementary Material for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.794, + 0.276, + 0.809 + ], + "angle": 0, + "content": "5.3. Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.863 + ], + "angle": 0, + "content": "We present the evaluation results of our approach and the baselines in Table 2. Several key observations and analyses are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "- Baseline models' limited efficacy in GigaGrounding: None of the baseline models achieved Precision@0.5 scores surpassing \\(50\\%\\). In contrast, human testers consistently achieved scores exceeding \\(60\\%\\), even on images of size \\(640 \\times 640\\). It is also noteworthy that the performance on GigaGrounding dramatically diverged from that on other benchmark datasets, such as RefCOCO. For instance, the classical ReSC model outperformed other baselines on GigaGrounding, even surpassing models initialized with pretrained weights." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.245, + 0.892, + 0.351 + ], + "angle": 0, + "content": "- Substantial advantages of GlaZing: GlaZing demonstrated a significant performance advantage over all baseline models. When compared to the highest-performing deep learning baseline at a resolution of \\(1536 \\times 1536\\) (i.e., ReSC), GlaZing achieved an impressive relative improvement of \\(28.6\\%\\). This substantial gain can be attributed to its glance-to-zoom-in design." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.354, + 0.892, + 0.49 + ], + "angle": 0, + "content": "- Incompatibility of existing two-stage methods: We noticed that MAttNet exhibited marginal improvements over random guessing (6.8% vs. 6.2%), despite that it used the ground-truth object bounding boxes as region proposals. Our hypothesis is that the bottleneck lies in the feature extractors. Two-stage methods typically use frozen feature extractors pretrained on MS COCO or ImageNet [53], which may not be effective enough for large-scale scene understanding." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.494, + 0.875, + 0.509 + ], + "angle": 0, + "content": "Additionally, our study reveals the following insights:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.513, + 0.892, + 0.573 + ], + "angle": 0, + "content": "- Influence of resolution on performance: Our findings highlight the critical role of high-resolution input in enhancing the performance of both deep learning models and human evaluators." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.577, + 0.892, + 0.682 + ], + "angle": 0, + "content": "- Performance variation with expression complexity: Notably, the GlaZing model demonstrated consistent performance across both direct and multi-hop expressions. In contrast, baseline models exhibited a marked performance decline when handling multi-hop expressions. This disparity underscores the need for models to adapt to expressions with differing levels of complexity." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.513, + 0.892, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.69, + 0.673, + 0.706 + ], + "angle": 0, + "content": "5.4. In-depth Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Effective high-resolution processing of GlaZing. GlaZing's adaptive cropping mechanism allows for higher-resolution image processing without significant storage and computation overhead. Experiments with 8-times downsampled images in ACM and ZGM showed promising results: \\(66.5\\%\\) on the test set, with \\(66.2\\%\\) for direct and \\(67.5\\%\\) for multi-hop expressions. These findings suggest that processing at higher resolutions may further enhance the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Benefits of RefCOCO initialization. We report the results of TransVG, RefTR, and SeqTR with and without RefCOCO weight initialization in Table 3. The adoption of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "*Consequently, it is inappropriate to directly compare the results of one-stage and two-stage methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22125" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.89, + 0.32 + ], + "angle": 0, + "content": "
CategoryMethodBackboneoverallVal directmulti-hopoverallTest directmulti-hopInfer TimeGFLOPsTrainable Params.
Two-stageMAttNetResNet-10111.4%11.6%10.7%6.8%6.4%7.6%268.6ms1592.613.0M
One-stage (640×640)ReSCDenseNet-5330.8%35.5%21.2%29.0%33.3%20.3%53.4ms101.3158.2M
QRNetSwin-S16.2%19.5%9.4%13.8%16.8%7.9%54.8ms79.2247.1M
SimRECCSPDarkNet-5323.3%27.6%14.4%20.8%24.7%12.7%16.5ms48.940.2M
TransVG‡ResNet-10113.1%15.9%7.1%11.6%13.7%7.3%25.9ms41.4122.6M
RefTR‡ResNet-5025.4%29.0%18.1%21.8%25.0%15.3%45.0ms23.5123.4M
SeqTR‡Darknet-5317.6%20.1%12.4%15.9%18.1%11.6%39.8ms48.9100.9M
Human------62%§----------
One-stage (1536×1536)ReSCDenseNet-5352.0%57.5%40.8%49.7%56.0%37.0%96.6ms566.3158.2M
QRNetSwin-S28.8%33.3%19.5%26.5%31.0%17.3%141.6ms435.0247.1M
SimRECCSPDarkNet-5335.4%41.7%22.5%33.3%39.5%20.8%48.4ms280.940.2M
TransVG‡ResNet-10126.3%31.2%16.5%24.6%29.1%15.4%72.6ms225.3122.6M
RefTR‡ResNet-5042.4%48.4%30.0%41.0%46.7%29.5%53.0ms119.7123.4M
SeqTR‡Darknet-5335.6%41.4%23.5%33.2%38.3%22.8%51.2ms263.7100.9M
Human------84%§----------
GlaZing (Ours)65.0%65.3%64.3%63.9%64.1%63.7%110.2ms172.1280.8M
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.2, + 0.321, + 0.88, + 0.332 + ], + "angle": 0, + "content": "\\( {}^{3} \\) We used the released weights pretrained on RefCOCO for initialization. \\( {}^{8} \\) Estimated with 100 randomly selected samples from the testing set." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.338, + 0.871, + 0.352 + ], + "angle": 0, + "content": "Table 2. Performance Evaluation on the GigaGrounding Dataset. Both accuracy-related and efficiency-related metrics are reported." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.369, + 0.422, + 0.436 + ], + "angle": 0, + "content": "
TransVGRefTRSeqTR
w/o pretraining5.1%0.0%5.3%
w/ pretraining11.6%21.8%15.9%
Absolute Gain+6.5%+21.8%+10.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.44, + 0.47, + 0.482 + ], + "angle": 0, + "content": "Table 3. Performance improvements on \\(640 \\times 640\\) input images achieved through pretraining. We initiated each model with publicly available weights trained on the RefCOCO dataset." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.496, + 0.454, + 0.565 + ], + "angle": 0, + "content": "
Expr. length & proportionReSCTransVGGlaZing
1-10 (16.99%)65.9%42.4%66.8%
11-20 (62.77%)47.8%22.6%63.5%
21+ (20.24%)39.7%13.8%62.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.569, + 0.441, + 0.582 + ], + "angle": 0, + "content": "Table 4. Performance stratification by expression length." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.604, + 0.425, + 0.689 + ], + "angle": 0, + "content": "
B-box scaleReSCTransVGGlaZing
s < Q129.4%4.5%48.1%
Q1 ≤ s < Q250.2%19.3%63.6%
Q2 ≤ s < Q358.2%31.1%72.8%
s ≥ Q360.7%43.6%70.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.693, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Table 5. Performance stratification by bounding box scale. Here, \\( Q1 \\), \\( Q2 \\), and \\( Q3 \\) denote the first, second, and third quartiles of bounding box scale, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.746, + 0.469, + 0.791 + ], + "angle": 0, + "content": "pretrained weights from RefCOCO led to performance enhancements for three methods, despite the substantial differences between RefCOCO and GigaGrounding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Performance about expression length and B-box scale. We conducted an in-depth analysis of the performance variability about expression length and B-box scale at \\(1536 \\times 1536\\), specifically focusing on the ReSC, TransVG, and GlaZing. The results are presented in Tables 4 and 5. Our findings indicate that GlaZing demonstrated a notable robustness across varying difficulty levels. Conversely," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.433 + ], + "angle": 0, + "content": "TransVG exhibited a more pronounced performance degradation, particularly in scenarios involving longer expressions and smaller B-box sizes. Complete results of all baselines can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Analysis of the glance-to-zoom-in strategy in GlaZing. Our investigation reveals that GlaZing's ZGM effectively rectifies prediction errors observed in the GGM, underscoring the effectiveness of the glance-to-zoom-in approach for GigaGrounding. Specifically, we identified that approximately \\(54\\%\\) of errors fall under \"target feature mismatch\", with ZGM correcting \\(48\\%\\) of these. Additionally, another \\(20\\%\\) of errors are categorized as \"over-fuzziness due to downsampling\", of which ZGM amends about \\(50\\%\\). Due to space constraints, detailed explanations and analyses are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.619, + 0.63 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.791 + ], + "angle": 0, + "content": "This paper presents GigaGrounding, a novel and challenging benchmark for visual grounding in gigapixel-level large-scale scenes, accompanied by an efficient solution strategy. The distinctive challenges posed by GigaGrounding establish it as a pioneering benchmark in this domain. Our goal is to stimulate further research in visual grounding, aiming to propel the development of models that are both more accurate and efficient. These advancements are particularly crucial for real-world applications such as rescue operations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This work is supported in part by Natural Science Foundation of China (NSFC) under contract No. 62125106 and 62088102, in part by Ministry of Science and Technology of China under contract No. 2021ZD0109901, in part by Tsinghua-Zhijiang joint research center, in part by Young Elite Scientists Sponsorship Program by CAST under contract No. 2022QNRC001." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22126" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[4] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[5] Ronghang Hu, Huazhe Xu, Marcus Rohrbach, Jiashi Feng, Kate Saenko, and Trevor Darrell. Natural language object retrieval. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.468, + 0.428 + ], + "angle": 0, + "content": "[6] Jianan Li, Yunchao Wei, Xiaodan Liang, Fang Zhao, Jianshu Li, Tingfa Xu, and Jiashi Feng. Deep attribute-preserving metric learning for natural language object retrieval. In ACM MM, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.468, + 0.471 + ], + "angle": 0, + "content": "[7] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-end visual grounding with transformers. In ICCV, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.468, + 0.515 + ], + "angle": 0, + "content": "[8] Xiaoyun Yuan, Mengqi Ji, Jiamin Wu, David J Brady, Qionghai Dai, and Lu Fang. A modular hierarchical array camera. Light: Sci. Appl., 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.468, + 0.584 + ], + "angle": 0, + "content": "[9] Byford Sam. Samsung announces 200-megapixel phone camera sensor. https://www.theverge.com/2021/9/2/22653558/samsung-isocell-hpl-gn5-200-megapixel-camera-sensor-announced, 2021.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[10] Arian Bakhtiarnia, Qi Zhang, and Alexandros Iosifidis. Efficient high-resolution deep learning: A survey. arXiv preprint arXiv:2207.13050, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.468, + 0.698 + ], + "angle": 0, + "content": "[11] Xueyang Wang, Xiya Zhang, Yinheng Zhu, Yuchen Guo, Xiaoyun Yuan, Liuyu Xiang, Zerun Wang, Guiguang Ding, David Brady, Qionghai Dai, and Lu Fang. PANDA: A gigapixel-level human-centric video dataset. In CVPR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[12] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.468, + 0.827 + ], + "angle": 0, + "content": "[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A González, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated IAPR TC-12 benchmark. CVIU, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.468, + 0.884 + ], + "angle": 0, + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Yuxi Sun, Shanshan Feng, Xutao Li, Yunming Ye, Jian" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Kang, and Xu Huang. Visual grounding in remote sensing images. In ACM MM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Yang Zhan, Zhitong Xiong, and Yuan Yuan. RSVG: Exploring data and models for visual grounding on remote sensing data. TGRS, 61:1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, et al. Making the most of text semantics to improve biomedical vision-language processing. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[18] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[19] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[20] Richang Hong, Daqing Liu, Xiaoyu Mo, Xiangnan He, and Hanwang Zhang. Learning to compose and reason with language tree structures for visual grounding. TPAMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[21] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[22] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[23] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[24] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[25] Xinpeng Chen, Lin Ma, Jingyuan Chen, Zequn Jie, Wei Liu, and Jiebo Luo. Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[26] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[27] Arka Sadhu, Kan Chen, and Ram Nevatia. Zero-shot grounding of objects from natural language queries. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[28] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[29] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. SeqTR: A simple yet universal network for visual grounding. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Jianing Zhang, Tianyi Zhu, Anke Zhang, Xiaoyun Yuan, Zihan Wang, Sebastian Beetschen, Lan Xu, Xing Lin, Qionghai Dai, and Lu Fang. Multiscale-VR: Multiscale gigapixel" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22127" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "3d panoramic videography for virtual reality. In ICCP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.164 + ], + "angle": 0, + "content": "[31] Xiaoyun Yuan, Lu Fang, Qionghai Dai, David J Brady, and Yebin Liu. Multiscale gigapixel video: A cross resolution image matching and warping approach. In ICCP, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.166, + 0.469, + 0.222 + ], + "angle": 0, + "content": "[32] Jianing Zhang, Jinzhi Zhang, Shi Mao, Mengqi Ji, Guangyu Wang, Zequn Chen, Tian Zhang, Xiaoyun Yuan, Qionghai Dai, and Lu Fang. GigaMVS: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.469, + 0.264 + ], + "angle": 0, + "content": "[33] Guangyu Wang, Jinzhi Zhang, Kai Zhang, Ruqi Huang, and Lu Fang. GiganticNVS: Gigapixel large-scale neural rendering with implicit meta-deformed manifold. TPAMI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.469, + 0.323 + ], + "angle": 0, + "content": "[34] Xueyang Wang, Xuecheng Chen, Puhua Jiang, Haozhe Lin, Xiaoyun Yuan, Mengqi Ji, Yuchen Guo, Ruqi Huang, and Lu Fang. The group interaction field for learning and explaining pedestrian anticipation. Engineering, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[35] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[36] Dmitrii Marin, Zijian He, Peter Vajda, Priyam Chatterjee, Sam Tsai, Fei Yang, and Yuri Boykov. Efficient segmentation: Learning downsampling near semantic boundaries. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[37] Chen Jin, Ryutaro Tanno, Thomy Mertzanidou, Eleftheria Panagiotaki, and Daniel C Alexander. Learning to down-sample for segmentation of ultra-high resolution images. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[38] Mingfei Gao, Ruichi Yu, Ang Li, Vlad I Morariu, and Larry S Davis. Dynamic zoom-in network for fast object detection in large images. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[39] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[40] Kai Chen, Zerun Wang, Xueyang Wang, Dahan Gong, Longlong Yu, Yuchen Guo, and Guiguang Ding. Towards real-time object detection in gigapixel-level video. Neurocomputing, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[41] Maria Tzelepi and Anastasios Tefas. Improving the performance of lightweight CNNs for binary classification using quadratic mutual information regularization. Pattern Recogn., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[42] Huangjing Lin, Hao Chen, Simon Graham, Qi Dou, Nasir Rajpoot, and Pheng-Ann Heng. Fast ScanNet: Fast and dense analysis of multi-gigapixel whole-slide images for cancer metastasis detection. TMI, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[43] Using Creative Commons Public Licenses. Attribution-noncommercial-sharealike 4.0 international. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[44] Baowei Jiang, Bing Bai, Haozhe Lin, Yu Wang, Yuchen Guo, and Lu Fang. DartBlur: Privacy preservation with detection artifact suppression. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[45] Joseph Redmon and Ali Farhadi. YOLOv3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. MAttNet: Modular attention network for referring expression comprehension. In CVPR, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[48] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[50] Gen Luo, Yiyi Zhou, Jiamu Sun, Shubin Huang, Xiaoshuai Sun, Qixiang Ye, Yongjian Wu, and Rongrong Ji. What goes beyond multi-modal fusion in one-stage referring expression comprehension: An empirical study. arXiv preprint arXiv:2204.07913, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[52] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22128" + } + ] +] \ No newline at end of file diff --git a/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_origin.pdf b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26fb2f5e54a297e40a1845c51ae87b110c8b3005 --- /dev/null +++ b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/69fa19e8-30d0-4373-b41c-ba0437416a98_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5df30c81d74b40e1e764dde1da4d0663b283f55d18f43b05f9d561afaa8cb6ae +size 5824509 diff --git a/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/full.md b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/full.md new file mode 100644 index 0000000000000000000000000000000000000000..af53f213746f0f9fbb48bd6959d3f9af7e02ab9f --- /dev/null +++ b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/full.md @@ -0,0 +1,354 @@ +# When Visual Grounding Meets Gigapixel-level Large-scale Scenes: Benchmark and Approach + +Tao Ma $^{1*}$ , Bing Bai $^{1*}$ , Haozhe Lin $^{1*}$ , Heyuan Wang $^{2}$ , Yu Wang $^{1}$ , Lin Luo $^{2}$ , Lu Fang $^{1\dagger}$ , Tsinghua University ${}^{2}$ Peking University + +fanglu@tsinghua.edu.cn + +![](images/f47882ec46cd4979712bbdefbb3eae343faaadf04255722ecc3a6b1dd6f675e1.jpg) +(a) RefCOCO + +![](images/d2a044a0374f3a69ec5cc24184651cfc27b0a441dcc80a647f8229418844d828.jpg) +(b) GigaGrounding +$\nrightarrow$ i. lower left white car +Figure 1. Comparison of the proposed GigaGrounding and the RefCOCO benchmark. GigaGrounding presents distinctive challenges, including large-scale scene understanding, high-resolution with significant scale variation, and multi-hop expressions. + +![](images/b608c9397576e19d045543d5dc78515ae474057f0bb8ad6fdbfa6fad9fd2fb81.jpg) + +# Abstract + +Visual grounding refers to the process of associating natural language expressions with corresponding regions within an image. Existing benchmarks for visual grounding primarily operate within small-scale scenes with a few objects. Nevertheless, recent advances in imaging technology have enabled the acquisition of gigapixel-level images, providing high-resolution details in large-scale scenes containing numerous objects. To bridge this gap between imaging and computer vision benchmarks and make grounding more practically valuable, we introduce a novel dataset, named GigaGrounding, designed to challenge visual grounding models in gigapixel-level large-scale scenes. We extensively analyze and compare the dataset with existing benchmarks, demonstrating that GigaGrounding presents unique challenges such as large-scale scene understanding, gigapixel-level resolution, significant variations in object scales, and the "multi-hop expressions". Furthermore, we introduced a simple yet effective grounding approach, which employs a "glance-to-zoom-in" paradigm and exhibits enhanced capabilities for addressing the GigaGrounding task. The dataset is available at www.gigavision.ai. + +# 1. Introduction + +Visual grounding (VG), also known as referring expression comprehension [1, 2], phrase localization [3, 4], and natural language object retrieval [5, 6], aims to identify the region within an image that corresponds to a given natural language expression. This field offers considerable potential for enhancing our capacity to connect the human language and visual elements of the tangible world [7]. The application of VG in scenarios such as rescue and robotics can prove to be exceedingly valuable, for example, in searching for a missing child in crowds. + +While the VG community has made significant progress in recent years, the benchmarks employed thus far have been predominantly limited to small-scale scenes [1-4]. As illustrated in Figure 1a, the number of objects is limited, and the objects cover the dominant portion of the image. However, recent advances in imaging technology, such as array cameras [8] and even consumer-grade smartphones [9], have enabled the capture of gigapixel-level photos, providing high-resolution detail information while covering large-scale scenes up to $\mathrm{km}^2$ -level [10], as shown in Figure 1b. The characteristics of conventional VG benchmarks have restricted current models' application to relatively simple and vanilla settings, hindering the development of new technologies and limiting the applicability of grounding in broader contexts [11]. + +To address this limitation, we introduce a novel benchmark dataset, named GigaGrounding, for visual grounding in large-scale scenes. The dataset is constructed based on PANDA [11], a gigapixel-level human-centric video dataset. An illustration of the GigaGrounding dataset is displayed in Figure 1b. In contrast to conventional visual grounding benchmarks such as RefCOCO [2], GigaGrounding poses unique and non-trivial challenges: + +- Large-scale scene understanding. Large-scale scenes require models to deal with a much larger spatial and semantic context and a significantly larger number of objects in the image. Hundreds of objects can be distributed over a vast area with possible occlusions, clutter, and other distractions, making it a difficult task to recognize and understand large-scale scenes. +- High resolution with significant scale variation. Processing high-throughput data up to gigapixel-level images poses new challenges to models in terms of efficiency [10]. Moreover, objects may be presented at drastically different scales. For instance, as shown in Figure 1b, the ground-truth bounding box (B-box) for Case i spans approximately $3000 \times 1400$ pixels, whereas, for Case ii, it spans only about $110 \times 330$ pixels. This significant variation necessitates models to be adaptable in scaling to accommodate diverse object sizes. +- Multi-hop expressions. In addition to direct expressions, which refer to descriptions of independent objects as illustrated in Case i, GigaGrounding also includes a considerable number of multi-hop expressions. These expressions challenge grounding models to locate target objects by first identifying some other reference objects. For example, in Case ii, the model needs first to identify "the middle lawn" and then zoom in to locate the correct person based on the relationship with the lawn. + +We benchmark the performance of cutting-edge VG models with the GigaGrounding dataset. Empirical findings reveal that while existing VG models have demonstrated satisfactory performance in small-scale scenes, they face distinct and intricate challenges when being applied to GigaGrounding. To address these challenges, we propose GlaZing (GLance-to-Zoom-IN Grounding), a simple yet effective model inspired by human cognitive processes. GlaZing employs a "glance-to-zoom-in" cascading strategy to ground target bounding boxes in high-resolution images. It begins by taking a global view of the thumbnail image, to identify the region of interest based on the given expression. Subsequently, it zooms in on the local region with adaptive scaling for the final grounding process. Experiments demonstrated that GlaZing outperformed existing models in the GigaGrounding dataset, effectively addressing major challenges such as high resolution, significant scale variation, and multi-hop expressions. + +In conclusion, this paper presents several contributions: + +- We highlight the significance of deploying visual grounding models in large-scale scenes, and introduce a novel benchmark dataset, GigaGrounding, characterized by unique properties that set it apart from existing datasets. +- We empirically assess current VG models using the proposed GigaGrounding dataset, highlighting the limitations of state-of-the-art models in this context. +- We introduce GlaZing, a tailored model for GigaGrounding, which leverages a "glance-to-zoom-in" approach. Our experimental results demonstrate the efficacy of this approach in addressing the current challenges, while also indicating room for further improvements. + +# 2. Related Work + +In this section, we review two relevant research domains: visual grounding and high-resolution deep learning. + +# 2.1. Visual Grounding + +Visual grounding is a field focused on predicting the location within an image that corresponds to a natural language expression. Benchmark datasets for VG have predominantly centered on small-scale scenes. For example, the Flickr30K Entities dataset [4] extends the original Flickr30K dataset [12] by incorporating annotations that establish correspondence between short region phrases and images. ReferItGame [3] comprises 20,000 images collected from the SAIAPR-12 dataset [13]. RefCOCO [2], RefCOCO+ [2], and RefCOCOg [1] are all built upon the MS COCO dataset [14]. Notably, the image resolutions in these datasets do not exceed $640 \times 480$ . In addition, there are related benchmarks in fields such as remote sensing [15, 16] and biomedical science [17, 18], with their resolutions also limited to $1024 \times 1024$ or lower. + +For VG models, existing work can be broadly categorized into two-stage and one-stage methods [7, 19]. Two-stage methods [20-24] initiate the process by generating a set of region proposals using object detectors. Subsequently, they leverage natural language expressions to rank these proposals and identify the best matching proposal. In contrast, one-stage methods [7, 19, 25-29] densely integrate both visual and textual information and directly output bounding boxes. + +# 2.2. High-resolution Deep Learning + +In recent years, significant progress has been made in the field of imaging, leading to gigapixel-level images with both a wide FoV and high resolution [8, 30, 31]. Employing high-resolution images and videos as direct inputs for deep learning models introduces challenges in both training and inference stages [11, 32-34]. To address these + +
DatasetTypical res.# Images# B-boxes# Expressions
Flickr30K Entities [4]500×37531,783276k427k
ReferItGame [3]480×36019,89496,654130,525
RefCOCO [2]640×48019,99450,000142,210
RefCOCO+ [2]640×48019,99249,856141,564
RefCOCOg [1]640×48025,79949,82295,010
GigaGrounding>25k×14k3,77561,35361,353
+ +Table 1. Statistics of existing benchmarks and GigaGrounding. + +challenges, several approaches have been explored. Rudimentary strategies like "uniform downsampling (UD)" and "cutting into patches (CIP)" may be employed, but they may result in the loss of detailed information and relational cues, respectively [10]. Alternatively, more sophisticated techniques, including non-uniform downsampling [35-37], selective zooming and skipping [38-40], and the use of lightweight scanner networks [41, 42], have been proposed. Bakhtiarnia et al. [10] have provided a comprehensive review of these methods, their relevant applications, and datasets pertaining to high-resolution deep learning. + +# 3. GigaGrounding Benchmark + +This section offers a comprehensive delineation of the GigaGrounding benchmark, including an overview, an elucidation of the data collection strategy, and analyses of its unique properties. + +# 3.1. Overview of GigaGrounding + +Dataset overview GigaGrounding is built upon images pre-extracted from PANDA-Video dataset by the dataset providers [11]. A summary of key statistics comparing conventional VG benchmarks with GigaGrounding is provided in Table 1. The analysis reveals a significant limitation in conventional VG datasets, where image resolutions are constrained to the megapixel level. In contrast, GigaGrounding ambitiously extends its scope to gigapixel-level images, encompassing intricate details and large-scale scenes. Consequently, the total number of images naturally decreases. In terms of bounding boxes, GigaGrounding offers a quantity comparable to other benchmark datasets. Additionally, in traditional benchmarks, multiple expressions may correspond to a single bounding box, while in GigaGrounding, each bounding box is associated with only one expression. + +We illustrate representative samples from conventional VG datasets in Figure 2, and also provide examples of GigaGrounding annotations in Figure 3. Specifically, GigaGrounding includes two types of annotated expressions: + +- Direct expressions: These expressions directly describe each target object without introducing dependencies on other objects. Position modifiers, such as "in the top left of the image", may be used when describing small and challenging-to-locate objects, but only when necessary. In total, there are 41,040 direct expressions. + +- Multi-hop expressions: These expressions require grounding models to respect additional restrictive constraints dependent on their relationship with other reference objects. In other words, the model needs to locate objects by first identifying specific reference objects and then finding the target objects based on constraints defined by relative positions (e.g., "to the left") or relationships (e.g., "watching", "holding") with those reference objects. There are 20,313 multi-hop expressions in total. It is noteworthy that we did not identify a design intentionally analogous to multi-hop expressions in traditional VG benchmarks. Notably, RefCOCO+ was explicitly designed to "focus on purely appearance-based description" [2]. + +For more details regarding the PANDA-Video dataset and additional examples of GigaGrounding, please refer to the Supplementary Material. + +Privacy GigaGrounding is built based on the publicly available PANDA [11] dataset. PANDA is collected in public areas where photography is officially approved and is published under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License [43]. GigaGrounding has been created by further processing the data in PANDA with additional annotations. To protect individual privacy, PANDA has already anonymized the faces in the images. However, it is necessary to acknowledge that certain individuals may still be identifiable based on other less invasive characteristics, such as clothing or body shape [44]. + +# 3.2. Data Collection Strategy + +We have devised the following annotation procedure and quality control strategies to diligently uphold non-ambiguity and maintain high annotation quality. + +Annotation protocol The annotation protocol, as depicted in Figure 4, comprises the following steps: + +1. Given the gigapixel images, Group $A$ of workers was asked to write natural language expressions and draw bounding boxes for the corresponding objects. +2. Given the gigapixel images and natural language expressions from Step 1, Group $B$ of workers was then asked to draw corresponding bounding boxes within 40 seconds independently. +3. If the Intersection over Union (IoU) rate of bounding boxes from Steps 1 and 2 was greater than 0.5, the sample were deemed valid. Then the valid expressions were manually translated into English with the highest possible fidelity. + +This procedure tries to ensure non-ambiguity, by excluding data instances with object selection discrepancies between Steps 1 and 2. + +![](images/e5a0078ffde46f02d26fb85d2f5460f78798a06325c46a07d93228369544b6f5.jpg) +A man with pierced ears is wearing glasses and an orange hat +(a) Flickr30K Entities + +![](images/a9ec1afe38b118013d88c74b98ae3630ec5d85e3f44fc4c40e569c0198461f67.jpg) +1. white shirt on right +2. white shirt man +3. man on right +(b) ReferItGame + +![](images/626a47c1e317b588cecc63e4e7c1ef553452a334bdfd62b9dc0fafe1b5bba24c.jpg) +1. woman on right in white shirt +2. woman on right +3. right women +(c) RefCOCO + +![](images/80e1a4747d73332c317ec4aafd2bf7e775ea2d1cdd71fc2d198822be613fbd0b.jpg) +1. yellow shirt and black shorts +2. guy in yellow dibbling ball +3. yellow shirt in focus +(d) RefCOCO+ + +![](images/625ff343b89777263452e430669f5930b94ee899c3913fabb18336ca8b996e93.jpg) +1. a woman wearing red playing tennis +2. a woman in a red shirt holding a tennis racquet +(e) RefCOCOg + +![](images/af62dc686c5e50b86aa66feeffa1446242e1f515c7221d414ddde0f2a74d5f83.jpg) +man in a red shirt pushing a suitcase near a subway station + +![](images/ff67f1bda316f7d0c36de4412f82244eed90fbfcd1396ec96f6ecc4f96b89760.jpg) +woman in a yellow top riding a tricycle in the middle of the picture + +![](images/8f0d23e7926b1470d14fc8fe837e1e2058c5757d4ef0f6bfecd35ac43d00c56b.jpg) +man in white jacket and black bag in front of the mall on the left of the picture + +![](images/67f656c3dd43e1a4bad14f8e1f880bb3df342d6794e95532cfc1ec37737e50be.jpg) +man in a white coat riding a blue bike at the bottom left of the picture + +![](images/3568193972739f7634a1cab604fda624a83f5662133dc9dd217dfcb67d2b4fe7.jpg) +Figure 2. Example data from conventional visual grounding datasets. +Figure 3. Example annotation results of GigaGrounding. Blue boxes indicate direct expressions and orange ones are multi-hop expressions. + +![](images/45662d443273f9d9a60b6e8c8aa7699d3973497116e6244fab08763a21eab923.jpg) + +![](images/00f7afc191be100ce2b25b9c87f12ca0dc9bcb02be5cd1075f6962070df9c381.jpg) +blue billboard in the top right corner of the picture + +![](images/c8fbb42d245127cafe7e68926b39715c594968a61b26aef0b9480fe534768674.jpg) +green spherical plant in the flower bed on the right of the picture + +![](images/db592d7b0d879f901b3a7fdc08941646f3630a882b9a5a85f2d6ff7ab5d45260.jpg) +red and white bus + +![](images/27d52e7710d92c789c7bbd78b279fe1f8be4206a55def5d87731e0cb55583230.jpg) +taxi closest to the red and white bus + +![](images/2b971872432f8cdd835169422d7cf51d110b984c25d62b8064e86e3e743ad1ee.jpg) +Figure 4. The overview of our annotation procedure. + +Quality control strategies We employ the following strategies for quality control. + +- Terminology standardization. We introduced a standardized lexicon to encourage uniform terminology usage among annotators, reducing the prevalence of ambiguous or domain-specific terms. For example, the use of specific color names like "gamboge" and the adoption of specialized terminology such as "coupe" should be avoided. +- Sampling inspection and repair. We performed a sampling inspection for each annotation batch to assess its quality. Specifically, for expression and bounding box annotations, we established a sampling pass rate threshold of $96\%$ , whereas, for expression translation, the threshold was set to $99\%$ . The sampling inspection process + +included no fewer than 100 samples per round, and the batch of data was accepted once it met the pass rate threshold or underwent repair. + +- Box ensemble and de-duplication. In Step 3 of the annotation procedure, we obtained the ground truth by averaging the valid bounding boxes annotated by both groups of annotators. Any expressions that did not result in a valid bounding box were discarded. Furthermore, as data annotation is carried out in parallel among many annotation workers, we performed a dedduplication procedure to eliminate objects with overlapping bounding boxes and identical expressions across multiple images to enhance the data variance, with an IoU threshold empirically established at 0.3. + +Despite potential undetected ambiguities, the implemented methodologies are believed to substantially mitigate their effects, thereby enhancing the dataset's overall quality. + +# 3.3. Analyses of GigaGrounding + +This section analyzes the salient characteristics differentiating GigaGrounding from existing VG benchmarks. Our analysis encompasses the following viewpoints. + +Quantities of objects per image Among conventional datasets, Flickr30K Entities in Figure 2a requires the model + +![](images/2f771c284aa1cc14660759392778b3b15d3b9eefc8c65a4a26124d601712278c.jpg) +(a) Pixel-count distribution of bounding box scales. + +![](images/dce1e696e3e8b90c0fdd1d2b539888508b739869c83e6d3bc2879e5f13857d2c.jpg) +Figure 5. A comparative analysis of GigaGrounding against established benchmarks RefCOCO, RefCOCO+, and RefCOCOg. (a) and (b) elucidate the distribution patterns of bounding box scale (in terms of the max value of height and width) and relative area with respect to the image, (c) presents the distribution of expression lengths, and (d) shows the lengths' distribution of both direct and multi-hop expressions within the GigaGrounding dataset. GigaGrounding exhibits a notably elevated level of complexity, primarily stemming from its pronounced scale variation and intricate expression structures. + +![](images/c49a5ac1eb51cba710b184cfdbb9629d98137ddaddf806b010f1a7ac95db44b9.jpg) +(b) Relative area of bounding boxes w.r.t. the images. + +![](images/3c481339ad783e0468048fc25a10dbf082910ed1f3720805e36f571fb3c0671c.jpg) +(c) Word-count distribution of expression lengths. +(d) Word-count distribution of direct and multi-hop expressions. + +to identify and locate multiple objects corresponding to short phrases within a single sentence. However, we observe that the model may not need to fully understand the global semantics before grounding the corresponding object, particularly when there exists only one object of that type. On the other side, RefCOCO and RefCOCO+ include an average of 3.9 same-type objects per image, while RefCOCOg displays a relatively lower average of 1.63 same-type objects per image [2]. + +Scales of bounding boxes In Figures 5a and 5b, we respectively illustrate the absolute scale of bounding boxes and their relative area with respect to the images. It is evident that GigaGrounding exhibits greater scale variance in comparison to other datasets. Furthermore, bounding box scale in GigaGrounding adheres to a long-tail distribution, while other datasets generally show a more concentrated distribution. On average, the RefCOCO dataset features bounding boxes with dimensions of $212 \times 268$ pixels, accounting for approximately $18.49\%$ of the total image area. In contrast, the GigaGrounding dataset exhibits an average bounding box size of $478 \times 978$ pixels, representing just $0.12\%$ of the image area. This statistic reveals the inherent challenges associated with grounding models tasked with large-scale scenes. We also checked the distribution of direct and multi-hop bounding boxes and noticed that they share a similar distribution pattern, while the bounding boxes corresponding to multi-hop expressions are slightly smaller, with an average resolution of $425 \times 938$ , compared to $503 \times 997$ of direct expression bounding boxes. + +Lengths of expressions We demonstrate the word count distribution of expression lengths in Figure 5c. We note that the average expression length in RefCOCOg is 8.46 words, while in RefCOCO and RefCOCO+, it is only 3.50 and 3.53, respectively. In contrast, in order to locate objects in large-scale scenes without ambiguity, GigaGrounding provides significantly longer expressions, and the average expression length is 14.78 words. Flickr30K Entities present comparatively lengthy expressions, while it is tasked with + +ground objects referred by short region phrases. Figure 5d further illustrates the length distribution of direct expressions and multi-hop expressions, and the average lengths are 12.33 and 18.74, respectively. Longer expressions may accommodate more details about the objects' location, appearance, and properties, which can increase the semantic space that the model needs to consider, leading to challenges in both textual and imagery comprehension. + +Other statistics We notice that GigaGrounding is certainly not offering the most number of images, bounding boxes, and annotated expressions among the compared datasets (Table 1). However, we believe that GigaGrounding's unique characteristics have opened the door for new modeling paradigms with practically useful and broader real-life applications. + +# 4. Approach + +In this section, we present an initial step towards effective visual grounding in the complex context of gigapixel-level large-scale scenes. We introduce a simple yet effective model, referred to as GlaZing. + +# 4.1. Architecture Overview + +In the context of GigaGrounding, human operators commonly begin with an initial coarse search across the image to identify regions of interest. Following this initial search, they then allocate their attention to the selected regions, subjecting them to thorough scrutiny. This methodical narrowing of focus significantly increases search efficacy. Our GlaZing model is conceived by analogizing this human cognitive heuristic. + +An overview of GlaZing is presented in Figure 6. Given an input image, it initiates by inputting the downsampled thumbnail and the corresponding expression into the Glance Grounding Module (GGM) for preliminary localization and grounding. Subsequently, the Adaptive Cropping Module (ACM) dynamically extracts the high-resolution patch + +![](images/279ea6629a6b9826a5097a8052c2721d6aecb6700f74cd6f054a95ac9b5df704.jpg) +Figure 6. Overview of the proposed GlaZing model. + +of interest. This extracted patch, along with the expression, is then forwarded to the Zoomed-in Grounding Module (ZGM) to generate the final prediction. + +# 4.2. Detailed Implementations + +The architectural framework expounded in Section 4.1 allows for a multitude of implementations. Notably, GGM and ZGM can be seamlessly integrated with a wide array of conventional grounding models. In this section, we present our specific implementation, which yielded favorable results in our experiments. + +Glance grounding module We design GGM based on ReSC [28], which adopts an anchor-based grounding paradigm that divides the image into multiple grids and determines the grid to which the target box belongs. This design effectively locates the target-relevant region in the thumbnail image. + +Adaptive cropping module This module is devised for the adaptive extraction of a region of interest from the original image, guided by the GGM's output. More precisely, ACM initiates this process by first computing the coordinates for a patch of specific size located at the center of the bounding box predicted by the GGM. It subsequently proceeds to crop the minimum rectangular region that encloses both this designated patch and the bounding box predicted by the GGM. During model training, this process also encapsulates the ground-truth bounding box. Finally, the extracted region is resized to a specified target size. Through the utilization of ACM, our model can adaptively focus on regions of varying scales and locations. + +Zoomed-in grounding module After the first glance and zooming-in, the target bounding box should account for a significant proportion of the input patch of ZGM, so we employ a coordinate regression-based visual grounding method, i.e., TransVG [7] to accomplish this task. + +Training objectives The overall training objective can be defined as $\mathcal{L} = \mathcal{L}_{\mathrm{GGM}} + \mathcal{L}_{\mathrm{ZGM}}$ . We follow the training objectives employed in ReSC and TransVG for $\mathcal{L}_{\mathrm{GGM}}$ and $\mathcal{L}_{\mathrm{ZGM}}$ respectively. Specifically, $\mathcal{L}_{\mathrm{GGM}}$ comprises the YOLOv3 + +loss [45] designed to optimize anchor association confidence and prediction offset, and a diversity loss [28] intended to guide sub-queries in each iteration to concentrate on distinct elements within the query. Concerning $\mathcal{L}_{\mathrm{ZGM}}$ , given that ZGM directly generates a 4-dimensional vector representing the relative coordinates of the grounded target box with respect to the input patch, we employ both the smooth $\ell 1$ loss and the generalized IoU loss [46] to minimize the disparity between the predicted and target values. + +# 4.3. Discussion + +As an initial stride for GigaGrounding, GlaZing tackles the challenges posed by high-resolution input and scale variation through a "glance-to-zoom-in" cascading strategy coupled with adaptive cropping. Empirical findings in Section 5.3 attest to its effectiveness. Nevertheless, substantial opportunities for further refinement remain, which we leave as future work, including: (a) more tailored training objectives for GGM and ZGM, (b) end-to-end pipelines for seamless glancing and zooming-in, and (c) the feasibility of multiple iterations of glance and zoom adjustments. + +# 5. Experiments + +This section delineates the evaluation protocols employed, followed by results' presentation and analysis. + +# 5.1. Baselines + +We conducted an evaluation of several publicly available baselines. As discussed in Section 2, two-stage methods rely on object detectors to generate region proposals. Given that object detection in large-scale scenes remains a challenge in its own right [40], we conducted brief experiments involving a representative open-source two-stage method, namely, MAttNet [47]. On the other hand, one-stage methods directly provide bounding box predictions without relying on separate object detection. In this regard, GlaZing is more aligned with one-stage methods. We evaluated state-of-the-art open-source one-stage methods, including ReSC [28], TransVG [7], RefTR [48], SeqTR [29], QR-Net [49], and SimREC [50]. A concise introduction to these baselines is provided in the Supplementary Material. + +In addition to visual grounding models, we also conducted a human evaluation to assess the ability of human subjects to complete the task when presented with down-sampled images as used by deep learning models. + +# 5.2. Evaluation Protocol + +Dataset split We randomly selected $70\%$ distinct expressions for training, $10\%$ for validation, and $20\%$ for testing. This process resulted in 42,641 samples for the training set, 6,254 for the validation set, and 12,458 for the testing set. + +Metric Following the standard protocol, we assessed the performance using Precision@0.5, where the prediction is deemed correct if its IoU with the ground-truth box is larger than 0.5. We also provide information on inference time and GFLOPs to demonstrate the computational efficiency of each model. + +Image preprocessing In our empirical analysis, for the two-stage baseline MAttNet, features were extracted with 8-times downsampled images. We followed Yu et al. [47], who used ground-truth object bounding boxes to eliminate region proposal bottlenecks from the object detector.* On the other side, one-stage baseline models were evaluated using two distinct image resolutions: $640 \times 640$ and $1536 \times 1536$ . The first resolution was in accordance with established benchmarks, while the latter was the feasible upper limit for processing by most baseline models given our computational constraints. At $640 \times 640$ , the mean bounding box dimensions were roughly $12 \times 25$ pixels. This size expanded to $30 \times 60$ pixels at $1536 \times 1536$ resolution. To ensure fair comparison, for GlaZing, we employed a $640 \times 640$ thumbnail image for the GGM, and the ACM cropped the region of interest from $1536 \times 1536$ images. The cropped region was subsequently resized to $640 \times 640$ for the ZGM to identify the referred region. + +Implementation details For all baseline methods, we conducted experiments based on their official code. However, for the object detection component of MAttNet, due to the lack of maintenance, we substituted it with the corresponding implementation of Faster R-CNN [51] from Detectron2 [52]. The substitution did not hurt the performance on RefCOCO based on our experiments. For one-stage methods, we tried to optimize their performance within a certain hyper-parameter tuning budget. For instance, for models with official weights on RefCOCO provided, we experimented with whether or not to use these weights for initialization and reported the better results. At the resolution of $1536 \times 1536$ , we also tried initializing with weights from GigaGrounding at $640 \times 640$ , and reported the better outcomes. We also adjusted some important hyperparameters including learning rate, etc. As for GlaZing, we trained the model with similar settings as ReSC. For all experiments, we set the maximum query length to 30, and the computations were performed on a server equipped with 8 NVIDIA RTX 3090 GPUs. Please refer to the Supplementary Material for more details. + +# 5.3. Overall Performance + +We present the evaluation results of our approach and the baselines in Table 2. Several key observations and analyses are as follows: + +- Baseline models' limited efficacy in GigaGrounding: None of the baseline models achieved Precision@0.5 scores surpassing $50\%$ . In contrast, human testers consistently achieved scores exceeding $60\%$ , even on images of size $640 \times 640$ . It is also noteworthy that the performance on GigaGrounding dramatically diverged from that on other benchmark datasets, such as RefCOCO. For instance, the classical ReSC model outperformed other baselines on GigaGrounding, even surpassing models initialized with pretrained weights. +- Substantial advantages of GlaZing: GlaZing demonstrated a significant performance advantage over all baseline models. When compared to the highest-performing deep learning baseline at a resolution of $1536 \times 1536$ (i.e., ReSC), GlaZing achieved an impressive relative improvement of $28.6\%$ . This substantial gain can be attributed to its glance-to-zoom-in design. +- Incompatibility of existing two-stage methods: We noticed that MAttNet exhibited marginal improvements over random guessing (6.8% vs. 6.2%), despite that it used the ground-truth object bounding boxes as region proposals. Our hypothesis is that the bottleneck lies in the feature extractors. Two-stage methods typically use frozen feature extractors pretrained on MS COCO or ImageNet [53], which may not be effective enough for large-scale scene understanding. + +Additionally, our study reveals the following insights: + +- Influence of resolution on performance: Our findings highlight the critical role of high-resolution input in enhancing the performance of both deep learning models and human evaluators. +- Performance variation with expression complexity: Notably, the GlaZing model demonstrated consistent performance across both direct and multi-hop expressions. In contrast, baseline models exhibited a marked performance decline when handling multi-hop expressions. This disparity underscores the need for models to adapt to expressions with differing levels of complexity. + +# 5.4. In-depth Analysis + +Effective high-resolution processing of GlaZing. GlaZing's adaptive cropping mechanism allows for higher-resolution image processing without significant storage and computation overhead. Experiments with 8-times downsampled images in ACM and ZGM showed promising results: $66.5\%$ on the test set, with $66.2\%$ for direct and $67.5\%$ for multi-hop expressions. These findings suggest that processing at higher resolutions may further enhance the model's performance. + +Benefits of RefCOCO initialization. We report the results of TransVG, RefTR, and SeqTR with and without RefCOCO weight initialization in Table 3. The adoption of + +
CategoryMethodBackboneoverallVal directmulti-hopoverallTest directmulti-hopInfer TimeGFLOPsTrainable Params.
Two-stageMAttNetResNet-10111.4%11.6%10.7%6.8%6.4%7.6%268.6ms1592.613.0M
One-stage (640×640)ReSCDenseNet-5330.8%35.5%21.2%29.0%33.3%20.3%53.4ms101.3158.2M
QRNetSwin-S16.2%19.5%9.4%13.8%16.8%7.9%54.8ms79.2247.1M
SimRECCSPDarkNet-5323.3%27.6%14.4%20.8%24.7%12.7%16.5ms48.940.2M
TransVG‡ResNet-10113.1%15.9%7.1%11.6%13.7%7.3%25.9ms41.4122.6M
RefTR‡ResNet-5025.4%29.0%18.1%21.8%25.0%15.3%45.0ms23.5123.4M
SeqTR‡Darknet-5317.6%20.1%12.4%15.9%18.1%11.6%39.8ms48.9100.9M
Human------62%§----------
One-stage (1536×1536)ReSCDenseNet-5352.0%57.5%40.8%49.7%56.0%37.0%96.6ms566.3158.2M
QRNetSwin-S28.8%33.3%19.5%26.5%31.0%17.3%141.6ms435.0247.1M
SimRECCSPDarkNet-5335.4%41.7%22.5%33.3%39.5%20.8%48.4ms280.940.2M
TransVG‡ResNet-10126.3%31.2%16.5%24.6%29.1%15.4%72.6ms225.3122.6M
RefTR‡ResNet-5042.4%48.4%30.0%41.0%46.7%29.5%53.0ms119.7123.4M
SeqTR‡Darknet-5335.6%41.4%23.5%33.2%38.3%22.8%51.2ms263.7100.9M
Human------84%§----------
GlaZing (Ours)65.0%65.3%64.3%63.9%64.1%63.7%110.2ms172.1280.8M
+ +${}^{3}$ We used the released weights pretrained on RefCOCO for initialization. ${}^{8}$ Estimated with 100 randomly selected samples from the testing set. + +Table 2. Performance Evaluation on the GigaGrounding Dataset. Both accuracy-related and efficiency-related metrics are reported. + +
TransVGRefTRSeqTR
w/o pretraining5.1%0.0%5.3%
w/ pretraining11.6%21.8%15.9%
Absolute Gain+6.5%+21.8%+10.6%
+ +Table 3. Performance improvements on $640 \times 640$ input images achieved through pretraining. We initiated each model with publicly available weights trained on the RefCOCO dataset. + +
Expr. length & proportionReSCTransVGGlaZing
1-10 (16.99%)65.9%42.4%66.8%
11-20 (62.77%)47.8%22.6%63.5%
21+ (20.24%)39.7%13.8%62.6%
+ +Table 4. Performance stratification by expression length. + +
B-box scaleReSCTransVGGlaZing
s < Q129.4%4.5%48.1%
Q1 ≤ s < Q250.2%19.3%63.6%
Q2 ≤ s < Q358.2%31.1%72.8%
s ≥ Q360.7%43.6%70.9%
+ +Table 5. Performance stratification by bounding box scale. Here, $Q1$ , $Q2$ , and $Q3$ denote the first, second, and third quartiles of bounding box scale, respectively. + +pretrained weights from RefCOCO led to performance enhancements for three methods, despite the substantial differences between RefCOCO and GigaGrounding. + +Performance about expression length and B-box scale. We conducted an in-depth analysis of the performance variability about expression length and B-box scale at $1536 \times 1536$ , specifically focusing on the ReSC, TransVG, and GlaZing. The results are presented in Tables 4 and 5. Our findings indicate that GlaZing demonstrated a notable robustness across varying difficulty levels. Conversely, + +TransVG exhibited a more pronounced performance degradation, particularly in scenarios involving longer expressions and smaller B-box sizes. Complete results of all baselines can be found in the Supplementary Material. + +Analysis of the glance-to-zoom-in strategy in GlaZing. Our investigation reveals that GlaZing's ZGM effectively rectifies prediction errors observed in the GGM, underscoring the effectiveness of the glance-to-zoom-in approach for GigaGrounding. Specifically, we identified that approximately $54\%$ of errors fall under "target feature mismatch", with ZGM correcting $48\%$ of these. Additionally, another $20\%$ of errors are categorized as "over-fuzziness due to downsampling", of which ZGM amends about $50\%$ . Due to space constraints, detailed explanations and analyses are provided in the Supplementary Material. + +# 6. Conclusion + +This paper presents GigaGrounding, a novel and challenging benchmark for visual grounding in gigapixel-level large-scale scenes, accompanied by an efficient solution strategy. The distinctive challenges posed by GigaGrounding establish it as a pioneering benchmark in this domain. Our goal is to stimulate further research in visual grounding, aiming to propel the development of models that are both more accurate and efficient. These advancements are particularly crucial for real-world applications such as rescue operations. + +Acknowledgements This work is supported in part by Natural Science Foundation of China (NSFC) under contract No. 62125106 and 62088102, in part by Ministry of Science and Technology of China under contract No. 2021ZD0109901, in part by Tsinghua-Zhijiang joint research center, in part by Young Elite Scientists Sponsorship Program by CAST under contract No. 2022QNRC001. + +# References + +[1] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 3 +[2] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 1, 2, 3, 5 +[3] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 2, 3 +[4] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 1, 2, 3 +[5] Ronghang Hu, Huazhe Xu, Marcus Rohrbach, Jiashi Feng, Kate Saenko, and Trevor Darrell. Natural language object retrieval. In CVPR, 2016. 1 +[6] Jianan Li, Yunchao Wei, Xiaodan Liang, Fang Zhao, Jianshu Li, Tingfa Xu, and Jiashi Feng. Deep attribute-preserving metric learning for natural language object retrieval. In ACM MM, 2017. 1 +[7] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-end visual grounding with transformers. In ICCV, 2021. 1, 2, 6 +[8] Xiaoyun Yuan, Mengqi Ji, Jiamin Wu, David J Brady, Qionghai Dai, and Lu Fang. A modular hierarchical array camera. Light: Sci. Appl., 2021. 1, 2 +[9] Byford Sam. Samsung announces 200-megapixel phone camera sensor. https://www.theverge.com/2021/9/2/22653558/samsung-isocell-hpl-gn5-200-megapixel-camera-sensor-announced, 2021.1 +[10] Arian Bakhtiarnia, Qi Zhang, and Alexandros Iosifidis. Efficient high-resolution deep learning: A survey. arXiv preprint arXiv:2207.13050, 2022. 1, 2, 3 +[11] Xueyang Wang, Xiya Zhang, Yinheng Zhu, Yuchen Guo, Xiaoyun Yuan, Liuyu Xiang, Zerun Wang, Guiguang Ding, David Brady, Qionghai Dai, and Lu Fang. PANDA: A gigapixel-level human-centric video dataset. In CVPR, 2020. 1, 2, 3 +[12] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2 +[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A González, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated IAPR TC-12 benchmark. CVIU, 2010. 2 +[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 2 +[15] Yuxi Sun, Shanshan Feng, Xutao Li, Yunming Ye, Jian + +Kang, and Xu Huang. Visual grounding in remote sensing images. In ACM MM, 2022. 2 +[16] Yang Zhan, Zhitong Xiong, and Yuan Yuan. RSVG: Exploring data and models for visual grounding on remote sensing data. TGRS, 61:1-13, 2023. 2 +[17] Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, et al. Making the most of text semantics to improve biomedical vision-language processing. In ECCV, 2022. 2 +[18] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 2 +[19] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In ICCV, 2019. 2 +[20] Richang Hong, Daqing Liu, Xiaoyu Mo, Xiangnan He, and Hanwang Zhang. Learning to compose and reason with language tree structures for visual grounding. TPAMI, 2019. 2 +[21] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In CVPR, 2017. +[22] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In ICCV, 2019. +[23] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In CVPR, 2018. +[24] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In CVPR, 2019. 2 +[25] Xinpeng Chen, Lin Ma, Jingyuan Chen, Zequn Jie, Wei Liu, and Jiebo Luo. Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426, 2018. 2 +[26] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In CVPR, 2020. +[27] Arka Sadhu, Kan Chen, and Ram Nevatia. Zero-shot grounding of objects from natural language queries. In ICCV, 2019. +[28] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In ECCV, 2020. 6 +[29] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. SeqTR: A simple yet universal network for visual grounding. In ECCV, 2022. 2, 6 +[30] Jianing Zhang, Tianyi Zhu, Anke Zhang, Xiaoyun Yuan, Zihan Wang, Sebastian Beetschen, Lan Xu, Xing Lin, Qionghai Dai, and Lu Fang. Multiscale-VR: Multiscale gigapixel + +3d panoramic videography for virtual reality. In ICCP, 2020. 2 +[31] Xiaoyun Yuan, Lu Fang, Qionghai Dai, David J Brady, and Yebin Liu. Multiscale gigapixel video: A cross resolution image matching and warping approach. In ICCP, 2017. 2 +[32] Jianing Zhang, Jinzhi Zhang, Shi Mao, Mengqi Ji, Guangyu Wang, Zequn Chen, Tian Zhang, Xiaoyun Yuan, Qionghai Dai, and Lu Fang. GigaMVS: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. TPAMI, 2021. 2 +[33] Guangyu Wang, Jinzhi Zhang, Kai Zhang, Ruqi Huang, and Lu Fang. GiganticNVS: Gigapixel large-scale neural rendering with implicit meta-deformed manifold. TPAMI, 2023. +[34] Xueyang Wang, Xuecheng Chen, Puhua Jiang, Haozhe Lin, Xiaoyun Yuan, Mengqi Ji, Yuchen Guo, Ruqi Huang, and Lu Fang. The group interaction field for learning and explaining pedestrian anticipation. Engineering, 2023. 2 +[35] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In ECCV, 2018. 3 +[36] Dmitrii Marin, Zijian He, Peter Vajda, Priyam Chatterjee, Sam Tsai, Fei Yang, and Yuri Boykov. Efficient segmentation: Learning downsampling near semantic boundaries. In ICCV, 2019. +[37] Chen Jin, Ryutaro Tanno, Thomy Mertzanidou, Eleftheria Panagiotaki, and Daniel C Alexander. Learning to down-sample for segmentation of ultra-high resolution images. In ICLR, 2022. 3 +[38] Mingfei Gao, Ruichi Yu, Ang Li, Vlad I Morariu, and Larry S Davis. Dynamic zoom-in network for fast object detection in large images. In CVPR, 2018. 3 +[39] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020. +[40] Kai Chen, Zerun Wang, Xueyang Wang, Dahan Gong, Longlong Yu, Yuchen Guo, and Guiguang Ding. Towards real-time object detection in gigapixel-level video. Neurocomputing, 2022. 3, 6 +[41] Maria Tzelepi and Anastasios Tefas. Improving the performance of lightweight CNNs for binary classification using quadratic mutual information regularization. Pattern Recogn., 2020. 3 +[42] Huangjing Lin, Hao Chen, Simon Graham, Qi Dou, Nasir Rajpoot, and Pheng-Ann Heng. Fast ScanNet: Fast and dense analysis of multi-gigapixel whole-slide images for cancer metastasis detection. TMI, 2019. 3 +[43] Using Creative Commons Public Licenses. Attribution-noncommercial-sharealike 4.0 international. 3 +[44] Baowei Jiang, Bing Bai, Haozhe Lin, Yu Wang, Yuchen Guo, and Lu Fang. DartBlur: Privacy preservation with detection artifact suppression. In CVPR, 2023. 3 +[45] Joseph Redmon and Ali Farhadi. YOLOv3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 6 +[46] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 6 + +[47] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. MAttNet: Modular attention network for referring expression comprehension. In CVPR, 2018. 6, 7 +[48] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, 2021. 6 +[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In CVPR, 2022. 6 +[50] Gen Luo, Yiyi Zhou, Jiamu Sun, Shubin Huang, Xiaoshuai Sun, Qixiang Ye, Yongjian Wu, and Rongrong Ji. What goes beyond multi-modal fusion in one-stage referring expression comprehension: An empirical study. arXiv preprint arXiv:2204.07913, 2022. 6 +[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 7 +[52] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.7 +[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7 \ No newline at end of file diff --git a/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/images.zip b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8523e1315c0f908a79067a4e3f8d16f3cce7efb6 --- /dev/null +++ b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8112454d1a5ffce64833ffb821de49de43e1803a40d2527565b34518108887f4 +size 554387 diff --git a/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/layout.json b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2f742e85843b8d19cfb9f66f0d098094c6a9864 --- /dev/null +++ b/2024/When Visual Grounding Meets Gigapixel-level Large-scale Scenes_ Benchmark and Approach/layout.json @@ -0,0 +1,9802 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 501, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 501, + 139 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 501, + 139 + ], + "type": "text", + "content": "When Visual Grounding Meets Gigapixel-level Large-scale Scenes: Benchmark and Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "spans": [ + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": "Tao Ma" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Bing Bai" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Haozhe Lin" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Heyuan Wang" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Yu Wang" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Lin Luo" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Lu Fang" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": ", Tsinghua University " + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 93, + 160, + 503, + 190 + ], + "type": "text", + "content": "Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 192, + 355, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 192, + 355, + 202 + ], + "spans": [ + { + "bbox": [ + 235, + 192, + 355, + 202 + ], + "type": "text", + "content": "fanglu@tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 182, + 330 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 182, + 330 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 182, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 182, + 330 + ], + "type": "image", + "image_path": "f47882ec46cd4979712bbdefbb3eae343faaadf04255722ecc3a6b1dd6f675e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 335, + 139, + 346 + ], + "lines": [ + { + "bbox": [ + 86, + 335, + 139, + 346 + ], + "spans": [ + { + "bbox": [ + 86, + 335, + 139, + 346 + ], + "type": "text", + "content": "(a) RefCOCO" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 194, + 219, + 396, + 334 + ], + "blocks": [ + { + "bbox": [ + 194, + 219, + 396, + 334 + ], + "lines": [ + { + "bbox": [ + 194, + 219, + 396, + 334 + ], + "spans": [ + { + "bbox": [ + 194, + 219, + 396, + 334 + ], + "type": "image", + "image_path": "d2a044a0374f3a69ec5cc24184651cfc27b0a441dcc80a647f8229418844d828.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 335, + 409, + 346 + ], + "lines": [ + { + "bbox": [ + 336, + 335, + 409, + 346 + ], + "spans": [ + { + "bbox": [ + 336, + 335, + 409, + 346 + ], + "type": "text", + "content": "(b) GigaGrounding" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 322, + 488, + 331 + ], + "lines": [ + { + "bbox": [ + 397, + 322, + 488, + 331 + ], + "spans": [ + { + "bbox": [ + 397, + 322, + 488, + 331 + ], + "type": "inline_equation", + "content": "\\nrightarrow" + }, + { + "bbox": [ + 397, + 322, + 488, + 331 + ], + "type": "text", + "content": " i. lower left white car" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 354, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 546, + 377 + ], + "type": "text", + "content": "Figure 1. Comparison of the proposed GigaGrounding and the RefCOCO benchmark. GigaGrounding presents distinctive challenges, including large-scale scene understanding, high-resolution with significant scale variation, and multi-hop expressions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 399, + 219, + 539, + 316 + ], + "blocks": [ + { + "bbox": [ + 399, + 219, + 539, + 316 + ], + "lines": [ + { + "bbox": [ + 399, + 219, + 539, + 316 + ], + "spans": [ + { + "bbox": [ + 399, + 219, + 539, + 316 + ], + "type": "image", + "image_path": "b608c9397576e19d045543d5dc78515ae474057f0bb8ad6fdbfa6fad9fd2fb81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "spans": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 415, + 290, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 415, + 290, + 666 + ], + "spans": [ + { + "bbox": [ + 45, + 415, + 290, + 666 + ], + "type": "text", + "content": "Visual grounding refers to the process of associating natural language expressions with corresponding regions within an image. Existing benchmarks for visual grounding primarily operate within small-scale scenes with a few objects. Nevertheless, recent advances in imaging technology have enabled the acquisition of gigapixel-level images, providing high-resolution details in large-scale scenes containing numerous objects. To bridge this gap between imaging and computer vision benchmarks and make grounding more practically valuable, we introduce a novel dataset, named GigaGrounding, designed to challenge visual grounding models in gigapixel-level large-scale scenes. We extensively analyze and compare the dataset with existing benchmarks, demonstrating that GigaGrounding presents unique challenges such as large-scale scene understanding, gigapixel-level resolution, significant variations in object scales, and the \"multi-hop expressions\". Furthermore, we introduced a simple yet effective grounding approach, which employs a \"glance-to-zoom-in\" paradigm and exhibits enhanced capabilities for addressing the GigaGrounding task. The dataset is available at www.gigavision.ai." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 387, + 386, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 386, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 386, + 399 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 411, + 546, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 530 + ], + "type": "text", + "content": "Visual grounding (VG), also known as referring expression comprehension [1, 2], phrase localization [3, 4], and natural language object retrieval [5, 6], aims to identify the region within an image that corresponds to a given natural language expression. This field offers considerable potential for enhancing our capacity to connect the human language and visual elements of the tangible world [7]. The application of VG in scenarios such as rescue and robotics can prove to be exceedingly valuable, for example, in searching for a missing child in crowds." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "content": "While the VG community has made significant progress in recent years, the benchmarks employed thus far have been predominantly limited to small-scale scenes [1-4]. As illustrated in Figure 1a, the number of objects is limited, and the objects cover the dominant portion of the image. However, recent advances in imaging technology, such as array cameras [8] and even consumer-grade smartphones [9], have enabled the capture of gigapixel-level photos, providing high-resolution detail information while covering large-scale scenes up to " + }, + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{km}^2" + }, + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "content": "-level [10], as shown in Figure 1b. The characteristics of conventional VG benchmarks have restricted current models' application to relatively simple and vanilla settings, hindering the development of new technologies and limiting the applicability of grounding in broader contexts [11]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 266, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 266, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 266, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work. \n†Lu Fang is the corresponding author (www.luvision.net)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22119" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "To address this limitation, we introduce a novel benchmark dataset, named GigaGrounding, for visual grounding in large-scale scenes. The dataset is constructed based on PANDA [11], a gigapixel-level human-centric video dataset. An illustration of the GigaGrounding dataset is displayed in Figure 1b. In contrast to conventional visual grounding benchmarks such as RefCOCO [2], GigaGrounding poses unique and non-trivial challenges:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 287, + 491 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 175, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 258 + ], + "type": "text", + "content": "- Large-scale scene understanding. Large-scale scenes require models to deal with a much larger spatial and semantic context and a significantly larger number of objects in the image. Hundreds of objects can be distributed over a vast area with possible occlusions, clutter, and other distractions, making it a difficult task to recognize and understand large-scale scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "type": "text", + "content": "- High resolution with significant scale variation. Processing high-throughput data up to gigapixel-level images poses new challenges to models in terms of efficiency [10]. Moreover, objects may be presented at drastically different scales. For instance, as shown in Figure 1b, the ground-truth bounding box (B-box) for Case i spans approximately " + }, + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "type": "inline_equation", + "content": "3000 \\times 1400" + }, + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "type": "text", + "content": " pixels, whereas, for Case ii, it spans only about " + }, + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "type": "inline_equation", + "content": "110 \\times 330" + }, + { + "bbox": [ + 47, + 262, + 287, + 380 + ], + "type": "text", + "content": " pixels. This significant variation necessitates models to be adaptable in scaling to accommodate diverse object sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 383, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 287, + 491 + ], + "type": "text", + "content": "- Multi-hop expressions. In addition to direct expressions, which refer to descriptions of independent objects as illustrated in Case i, GigaGrounding also includes a considerable number of multi-hop expressions. These expressions challenge grounding models to locate target objects by first identifying some other reference objects. For example, in Case ii, the model needs first to identify \"the middle lawn\" and then zoom in to locate the correct person based on the relationship with the lawn." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "We benchmark the performance of cutting-edge VG models with the GigaGrounding dataset. Empirical findings reveal that while existing VG models have demonstrated satisfactory performance in small-scale scenes, they face distinct and intricate challenges when being applied to GigaGrounding. To address these challenges, we propose GlaZing (GLance-to-Zoom-IN Grounding), a simple yet effective model inspired by human cognitive processes. GlaZing employs a \"glance-to-zoom-in\" cascading strategy to ground target bounding boxes in high-resolution images. It begins by taking a global view of the thumbnail image, to identify the region of interest based on the given expression. Subsequently, it zooms in on the local region with adaptive scaling for the final grounding process. Experiments demonstrated that GlaZing outperformed existing models in the GigaGrounding dataset, effectively addressing major challenges such as high resolution, significant scale variation, and multi-hop expressions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 72, + 542, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 542, + 85 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 542, + 85 + ], + "type": "text", + "content": "In conclusion, this paper presents several contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 88, + 545, + 236 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 88, + 545, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 88, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 306, + 88, + 545, + 137 + ], + "type": "text", + "content": "- We highlight the significance of deploying visual grounding models in large-scale scenes, and introduce a novel benchmark dataset, GigaGrounding, characterized by unique properties that set it apart from existing datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 139, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 139, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 139, + 545, + 174 + ], + "type": "text", + "content": "- We empirically assess current VG models using the proposed GigaGrounding dataset, highlighting the limitations of state-of-the-art models in this context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 177, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 177, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 545, + 236 + ], + "type": "text", + "content": "- We introduce GlaZing, a tailored model for GigaGrounding, which leverages a \"glance-to-zoom-in\" approach. Our experimental results demonstrate the efficacy of this approach in addressing the current challenges, while also indicating room for further improvements." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 249, + 392, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 392, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 392, + 262 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 270, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 296 + ], + "type": "text", + "content": "In this section, we review two relevant research domains: visual grounding and high-resolution deep learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 304, + 413, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 413, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 413, + 317 + ], + "type": "text", + "content": "2.1. Visual Grounding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "type": "text", + "content": "Visual grounding is a field focused on predicting the location within an image that corresponds to a natural language expression. Benchmark datasets for VG have predominantly centered on small-scale scenes. For example, the Flickr30K Entities dataset [4] extends the original Flickr30K dataset [12] by incorporating annotations that establish correspondence between short region phrases and images. ReferItGame [3] comprises 20,000 images collected from the SAIAPR-12 dataset [13]. RefCOCO [2], RefCOCO+ [2], and RefCOCOg [1] are all built upon the MS COCO dataset [14]. Notably, the image resolutions in these datasets do not exceed " + }, + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "type": "text", + "content": ". In addition, there are related benchmarks in fields such as remote sensing [15, 16] and biomedical science [17, 18], with their resolutions also limited to " + }, + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 324, + 545, + 502 + ], + "type": "text", + "content": " or lower." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 505, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 612 + ], + "type": "text", + "content": "For VG models, existing work can be broadly categorized into two-stage and one-stage methods [7, 19]. Two-stage methods [20-24] initiate the process by generating a set of region proposals using object detectors. Subsequently, they leverage natural language expressions to rank these proposals and identify the best matching proposal. In contrast, one-stage methods [7, 19, 25-29] densely integrate both visual and textual information and directly output bounding boxes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 622, + 474, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 474, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 474, + 635 + ], + "type": "text", + "content": "2.2. High-resolution Deep Learning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "In recent years, significant progress has been made in the field of imaging, leading to gigapixel-level images with both a wide FoV and high resolution [8, 30, 31]. Employing high-resolution images and videos as direct inputs for deep learning models introduces challenges in both training and inference stages [11, 32-34]. To address these" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22120" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 138 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 138 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 138 + ], + "type": "table", + "html": "
DatasetTypical res.# Images# B-boxes# Expressions
Flickr30K Entities [4]500×37531,783276k427k
ReferItGame [3]480×36019,89496,654130,525
RefCOCO [2]640×48019,99450,000142,210
RefCOCO+ [2]640×48019,99249,856141,564
RefCOCOg [1]640×48025,79949,82295,010
GigaGrounding>25k×14k3,77561,35361,353
", + "image_path": "a60ed7964f95a8f84305e04043b1cc8e0dfca40393952f9979bfa7370afb1b65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 142, + 281, + 153 + ], + "lines": [ + { + "bbox": [ + 52, + 142, + 281, + 153 + ], + "spans": [ + { + "bbox": [ + 52, + 142, + 281, + 153 + ], + "type": "text", + "content": "Table 1. Statistics of existing benchmarks and GigaGrounding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 169, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 287, + 301 + ], + "type": "text", + "content": "challenges, several approaches have been explored. Rudimentary strategies like \"uniform downsampling (UD)\" and \"cutting into patches (CIP)\" may be employed, but they may result in the loss of detailed information and relational cues, respectively [10]. Alternatively, more sophisticated techniques, including non-uniform downsampling [35-37], selective zooming and skipping [38-40], and the use of lightweight scanner networks [41, 42], have been proposed. Bakhtiarnia et al. [10] have provided a comprehensive review of these methods, their relevant applications, and datasets pertaining to high-resolution deep learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 206, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 206, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 206, + 326 + ], + "type": "text", + "content": "3. GigaGrounding Benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 332, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 287, + 381 + ], + "type": "text", + "content": "This section offers a comprehensive delineation of the GigaGrounding benchmark, including an overview, an elucidation of the data collection strategy, and analyses of its unique properties." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 389, + 204, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 204, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 204, + 403 + ], + "type": "text", + "content": "3.1. Overview of GigaGrounding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "content": "Dataset overview GigaGrounding is built upon images pre-extracted from PANDA-Video dataset by the dataset providers [11]. A summary of key statistics comparing conventional VG benchmarks with GigaGrounding is provided in Table 1. The analysis reveals a significant limitation in conventional VG datasets, where image resolutions are constrained to the megapixel level. In contrast, GigaGrounding ambitiously extends its scope to gigapixel-level images, encompassing intricate details and large-scale scenes. Consequently, the total number of images naturally decreases. In terms of bounding boxes, GigaGrounding offers a quantity comparable to other benchmark datasets. Additionally, in traditional benchmarks, multiple expressions may correspond to a single bounding box, while in GigaGrounding, each bounding box is associated with only one expression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 286, + 639 + ], + "type": "text", + "content": "We illustrate representative samples from conventional VG datasets in Figure 2, and also provide examples of GigaGrounding annotations in Figure 3. Specifically, GigaGrounding includes two types of annotated expressions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 641, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 286, + 714 + ], + "type": "text", + "content": "- Direct expressions: These expressions directly describe each target object without introducing dependencies on other objects. Position modifiers, such as \"in the top left of the image\", may be used when describing small and challenging-to-locate objects, but only when necessary. In total, there are 41,040 direct expressions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 240 + ], + "type": "text", + "content": "- Multi-hop expressions: These expressions require grounding models to respect additional restrictive constraints dependent on their relationship with other reference objects. In other words, the model needs to locate objects by first identifying specific reference objects and then finding the target objects based on constraints defined by relative positions (e.g., \"to the left\") or relationships (e.g., \"watching\", \"holding\") with those reference objects. There are 20,313 multi-hop expressions in total. It is noteworthy that we did not identify a design intentionally analogous to multi-hop expressions in traditional VG benchmarks. Notably, RefCOCO+ was explicitly designed to \"focus on purely appearance-based description\" [2]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 244, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 281 + ], + "type": "text", + "content": "For more details regarding the PANDA-Video dataset and additional examples of GigaGrounding, please refer to the Supplementary Material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 283, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 428 + ], + "type": "text", + "content": "Privacy GigaGrounding is built based on the publicly available PANDA [11] dataset. PANDA is collected in public areas where photography is officially approved and is published under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License [43]. GigaGrounding has been created by further processing the data in PANDA with additional annotations. To protect individual privacy, PANDA has already anonymized the faces in the images. However, it is necessary to acknowledge that certain individuals may still be identifiable based on other less invasive characteristics, such as clothing or body shape [44]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 437, + 443, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 443, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 443, + 451 + ], + "type": "text", + "content": "3.2. Data Collection Strategy" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 457, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 494 + ], + "type": "text", + "content": "We have devised the following annotation procedure and quality control strategies to diligently uphold non-ambiguity and maintain high annotation quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "text", + "content": "Annotation protocol The annotation protocol, as depicted in Figure 4, comprises the following steps:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 525, + 545, + 673 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 525, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 561 + ], + "type": "text", + "content": "1. Given the gigapixel images, Group " + }, + { + "bbox": [ + 306, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 306, + 525, + 545, + 561 + ], + "type": "text", + "content": " of workers was asked to write natural language expressions and draw bounding boxes for the corresponding objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 563, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 545, + 611 + ], + "type": "text", + "content": "2. Given the gigapixel images and natural language expressions from Step 1, Group " + }, + { + "bbox": [ + 306, + 563, + 545, + 611 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 306, + 563, + 545, + 611 + ], + "type": "text", + "content": " of workers was then asked to draw corresponding bounding boxes within 40 seconds independently." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 613, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 613, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 613, + 545, + 673 + ], + "type": "text", + "content": "3. If the Intersection over Union (IoU) rate of bounding boxes from Steps 1 and 2 was greater than 0.5, the sample were deemed valid. Then the valid expressions were manually translated into English with the highest possible fidelity." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "This procedure tries to ensure non-ambiguity, by excluding data instances with object selection discrepancies between Steps 1 and 2." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22121" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 127, + 144 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 127, + 144 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 127, + 144 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 127, + 144 + ], + "type": "image", + "image_path": "e5a0078ffde46f02d26fb85d2f5460f78798a06325c46a07d93228369544b6f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 144, + 119, + 165 + ], + "lines": [ + { + "bbox": [ + 57, + 144, + 119, + 165 + ], + "spans": [ + { + "bbox": [ + 57, + 144, + 119, + 165 + ], + "type": "text", + "content": "A man with pierced ears is wearing glasses and an orange hat" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 171, + 124, + 180 + ], + "lines": [ + { + "bbox": [ + 53, + 171, + 124, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 124, + 180 + ], + "type": "text", + "content": "(a) Flickr30K Entities" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 226, + 144 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 226, + 144 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 226, + 144 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 226, + 144 + ], + "type": "image", + "image_path": "a9ec1afe38b118013d88c74b98ae3630ec5d85e3f44fc4c40e569c0198461f67.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 144, + 206, + 150 + ], + "lines": [ + { + "bbox": [ + 147, + 144, + 206, + 150 + ], + "spans": [ + { + "bbox": [ + 147, + 144, + 206, + 150 + ], + "type": "text", + "content": "1. white shirt on right" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 151, + 196, + 157 + ], + "lines": [ + { + "bbox": [ + 147, + 151, + 196, + 157 + ], + "spans": [ + { + "bbox": [ + 147, + 151, + 196, + 157 + ], + "type": "text", + "content": "2. white shirt man" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 158, + 189, + 164 + ], + "lines": [ + { + "bbox": [ + 147, + 158, + 189, + 164 + ], + "spans": [ + { + "bbox": [ + 147, + 158, + 189, + 164 + ], + "type": "text", + "content": "3. man on right" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 171, + 205, + 180 + ], + "lines": [ + { + "bbox": [ + 150, + 171, + 205, + 180 + ], + "spans": [ + { + "bbox": [ + 150, + 171, + 205, + 180 + ], + "type": "text", + "content": "(b) ReferItGame" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 324, + 143 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 324, + 143 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 324, + 143 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 324, + 143 + ], + "type": "image", + "image_path": "626a47c1e317b588cecc63e4e7c1ef553452a334bdfd62b9dc0fafe1b5bba24c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 143, + 319, + 149 + ], + "lines": [ + { + "bbox": [ + 232, + 143, + 319, + 149 + ], + "spans": [ + { + "bbox": [ + 232, + 143, + 319, + 149 + ], + "type": "text", + "content": "1. woman on right in white shirt" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 150, + 282, + 157 + ], + "lines": [ + { + "bbox": [ + 232, + 150, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 232, + 150, + 282, + 157 + ], + "type": "text", + "content": "2. woman on right" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 158, + 273, + 164 + ], + "lines": [ + { + "bbox": [ + 232, + 158, + 273, + 164 + ], + "spans": [ + { + "bbox": [ + 232, + 158, + 273, + 164 + ], + "type": "text", + "content": "3. right women" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 171, + 299, + 180 + ], + "lines": [ + { + "bbox": [ + 253, + 171, + 299, + 180 + ], + "spans": [ + { + "bbox": [ + 253, + 171, + 299, + 180 + ], + "type": "text", + "content": "(c) RefCOCO" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 71, + 432, + 144 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 432, + 144 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 432, + 144 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 432, + 144 + ], + "type": "image", + "image_path": "80e1a4747d73332c317ec4aafd2bf7e775ea2d1cdd71fc2d198822be613fbd0b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 144, + 420, + 149 + ], + "lines": [ + { + "bbox": [ + 335, + 144, + 420, + 149 + ], + "spans": [ + { + "bbox": [ + 335, + 144, + 420, + 149 + ], + "type": "text", + "content": "1. yellow shirt and black shorts" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 150, + 413, + 157 + ], + "lines": [ + { + "bbox": [ + 335, + 150, + 413, + 157 + ], + "spans": [ + { + "bbox": [ + 335, + 150, + 413, + 157 + ], + "type": "text", + "content": "2. guy in yellow dibbling ball" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 158, + 395, + 164 + ], + "lines": [ + { + "bbox": [ + 335, + 158, + 395, + 164 + ], + "spans": [ + { + "bbox": [ + 335, + 158, + 395, + 164 + ], + "type": "text", + "content": "3. yellow shirt in focus" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 171, + 404, + 180 + ], + "lines": [ + { + "bbox": [ + 353, + 171, + 404, + 180 + ], + "spans": [ + { + "bbox": [ + 353, + 171, + 404, + 180 + ], + "type": "text", + "content": "(d) RefCOCO+" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 544, + 144 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 544, + 144 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 544, + 144 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 544, + 144 + ], + "type": "image", + "image_path": "625ff343b89777263452e430669f5930b94ee899c3913fabb18336ca8b996e93.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 144, + 541, + 149 + ], + "lines": [ + { + "bbox": [ + 436, + 144, + 541, + 149 + ], + "spans": [ + { + "bbox": [ + 436, + 144, + 541, + 149 + ], + "type": "text", + "content": "1. a woman wearing red playing tennis" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 150, + 531, + 165 + ], + "lines": [ + { + "bbox": [ + 436, + 150, + 531, + 165 + ], + "spans": [ + { + "bbox": [ + 436, + 150, + 531, + 165 + ], + "type": "text", + "content": "2. a woman in a red shirt holding a tennis racquet" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 464, + 171, + 514, + 180 + ], + "lines": [ + { + "bbox": [ + 464, + 171, + 514, + 180 + ], + "spans": [ + { + "bbox": [ + 464, + 171, + 514, + 180 + ], + "type": "text", + "content": "(e) RefCOCOg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 50, + 217, + 83, + 254 + ], + "blocks": [ + { + "bbox": [ + 50, + 217, + 83, + 254 + ], + "lines": [ + { + "bbox": [ + 50, + 217, + 83, + 254 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 83, + 254 + ], + "type": "image", + "image_path": "af62dc686c5e50b86aa66feeffa1446242e1f515c7221d414ddde0f2a74d5f83.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 224, + 150, + 248 + ], + "lines": [ + { + "bbox": [ + 83, + 224, + 150, + 248 + ], + "spans": [ + { + "bbox": [ + 83, + 224, + 150, + 248 + ], + "type": "text", + "content": "man in a red shirt pushing a suitcase near a subway station" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 51, + 257, + 91, + 292 + ], + "blocks": [ + { + "bbox": [ + 51, + 257, + 91, + 292 + ], + "lines": [ + { + "bbox": [ + 51, + 257, + 91, + 292 + ], + "spans": [ + { + "bbox": [ + 51, + 257, + 91, + 292 + ], + "type": "image", + "image_path": "ff67f1bda316f7d0c36de4412f82244eed90fbfcd1396ec96f6ecc4f96b89760.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 263, + 158, + 288 + ], + "lines": [ + { + "bbox": [ + 93, + 263, + 158, + 288 + ], + "spans": [ + { + "bbox": [ + 93, + 263, + 158, + 288 + ], + "type": "text", + "content": "woman in a yellow top riding a tricycle in the middle of the picture" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 51, + 297, + 79, + 329 + ], + "blocks": [ + { + "bbox": [ + 51, + 297, + 79, + 329 + ], + "lines": [ + { + "bbox": [ + 51, + 297, + 79, + 329 + ], + "spans": [ + { + "bbox": [ + 51, + 297, + 79, + 329 + ], + "type": "image", + "image_path": "8f0d23e7926b1470d14fc8fe837e1e2058c5757d4ef0f6bfecd35ac43d00c56b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 301, + 164, + 327 + ], + "lines": [ + { + "bbox": [ + 81, + 301, + 164, + 327 + ], + "spans": [ + { + "bbox": [ + 81, + 301, + 164, + 327 + ], + "type": "text", + "content": "man in white jacket and black bag in front of the mall on the left of the picture" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 51, + 335, + 84, + 369 + ], + "blocks": [ + { + "bbox": [ + 51, + 335, + 84, + 369 + ], + "lines": [ + { + "bbox": [ + 51, + 335, + 84, + 369 + ], + "spans": [ + { + "bbox": [ + 51, + 335, + 84, + 369 + ], + "type": "image", + "image_path": "67f656c3dd43e1a4bad14f8e1f880bb3df342d6794e95532cfc1ec37737e50be.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 341, + 158, + 365 + ], + "lines": [ + { + "bbox": [ + 86, + 341, + 158, + 365 + ], + "spans": [ + { + "bbox": [ + 86, + 341, + 158, + 365 + ], + "type": "text", + "content": "man in a white coat riding a blue bike at the bottom left of the picture" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 170, + 217, + 315, + 371 + ], + "blocks": [ + { + "bbox": [ + 171, + 182, + 421, + 195 + ], + "lines": [ + { + "bbox": [ + 171, + 182, + 421, + 195 + ], + "spans": [ + { + "bbox": [ + 171, + 182, + 421, + 195 + ], + "type": "text", + "content": "Figure 2. Example data from conventional visual grounding datasets." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 217, + 315, + 371 + ], + "lines": [ + { + "bbox": [ + 170, + 217, + 315, + 371 + ], + "spans": [ + { + "bbox": [ + 170, + 217, + 315, + 371 + ], + "type": "image", + "image_path": "3568193972739f7634a1cab604fda624a83f5662133dc9dd217dfcb67d2b4fe7.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 387, + 544, + 399 + ], + "lines": [ + { + "bbox": [ + 47, + 387, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 544, + 399 + ], + "type": "text", + "content": "Figure 3. Example annotation results of GigaGrounding. Blue boxes indicate direct expressions and orange ones are multi-hop expressions." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 315, + 217, + 422, + 371 + ], + "blocks": [ + { + "bbox": [ + 315, + 217, + 422, + 371 + ], + "lines": [ + { + "bbox": [ + 315, + 217, + 422, + 371 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 422, + 371 + ], + "type": "image", + "image_path": "45662d443273f9d9a60b6e8c8aa7699d3973497116e6244fab08763a21eab923.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 425, + 217, + 481, + 253 + ], + "blocks": [ + { + "bbox": [ + 425, + 217, + 481, + 253 + ], + "lines": [ + { + "bbox": [ + 425, + 217, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 425, + 217, + 481, + 253 + ], + "type": "image", + "image_path": "00f7afc191be100ce2b25b9c87f12ca0dc9bcb02be5cd1075f6962070df9c381.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 223, + 539, + 248 + ], + "lines": [ + { + "bbox": [ + 483, + 223, + 539, + 248 + ], + "spans": [ + { + "bbox": [ + 483, + 223, + 539, + 248 + ], + "type": "text", + "content": "blue billboard in the top right corner of the picture" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 425, + 257, + 468, + 292 + ], + "blocks": [ + { + "bbox": [ + 425, + 257, + 468, + 292 + ], + "lines": [ + { + "bbox": [ + 425, + 257, + 468, + 292 + ], + "spans": [ + { + "bbox": [ + 425, + 257, + 468, + 292 + ], + "type": "image", + "image_path": "c8fbb42d245127cafe7e68926b39715c594968a61b26aef0b9480fe534768674.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 263, + 533, + 288 + ], + "lines": [ + { + "bbox": [ + 470, + 263, + 533, + 288 + ], + "spans": [ + { + "bbox": [ + 470, + 263, + 533, + 288 + ], + "type": "text", + "content": "green spherical plant in the flower bed on the right of the picture" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 425, + 297, + 487, + 330 + ], + "blocks": [ + { + "bbox": [ + 425, + 297, + 487, + 330 + ], + "lines": [ + { + "bbox": [ + 425, + 297, + 487, + 330 + ], + "spans": [ + { + "bbox": [ + 425, + 297, + 487, + 330 + ], + "type": "image", + "image_path": "db592d7b0d879f901b3a7fdc08941646f3630a882b9a5a85f2d6ff7ab5d45260.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 310, + 539, + 317 + ], + "lines": [ + { + "bbox": [ + 488, + 310, + 539, + 317 + ], + "spans": [ + { + "bbox": [ + 488, + 310, + 539, + 317 + ], + "type": "text", + "content": "red and white bus" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 425, + 335, + 457, + 369 + ], + "blocks": [ + { + "bbox": [ + 425, + 335, + 457, + 369 + ], + "lines": [ + { + "bbox": [ + 425, + 335, + 457, + 369 + ], + "spans": [ + { + "bbox": [ + 425, + 335, + 457, + 369 + ], + "type": "image", + "image_path": "27d52e7710d92c789c7bbd78b279fe1f8be4206a55def5d87731e0cb55583230.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 345, + 523, + 361 + ], + "lines": [ + { + "bbox": [ + 459, + 345, + 523, + 361 + ], + "spans": [ + { + "bbox": [ + 459, + 345, + 523, + 361 + ], + "type": "text", + "content": "taxi closest to the red and white bus" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 62, + 416, + 273, + 501 + ], + "blocks": [ + { + "bbox": [ + 62, + 416, + 273, + 501 + ], + "lines": [ + { + "bbox": [ + 62, + 416, + 273, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 416, + 273, + 501 + ], + "type": "image", + "image_path": "2b971872432f8cdd835169422d7cf51d110b984c25d62b8064e86e3e743ad1ee.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 507, + 261, + 518 + ], + "lines": [ + { + "bbox": [ + 70, + 507, + 261, + 518 + ], + "spans": [ + { + "bbox": [ + 70, + 507, + 261, + 518 + ], + "type": "text", + "content": "Figure 4. The overview of our annotation procedure." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 538, + 286, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 286, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 286, + 562 + ], + "type": "text", + "content": "Quality control strategies We employ the following strategies for quality control." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 47, + 567, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 47, + "blocks": [ + { + "bbox": [ + 47, + 567, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 287, + 639 + ], + "type": "text", + "content": "- Terminology standardization. We introduced a standardized lexicon to encourage uniform terminology usage among annotators, reducing the prevalence of ambiguous or domain-specific terms. For example, the use of specific color names like \"gamboge\" and the adoption of specialized terminology such as \"coupe\" should be avoided." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": "- Sampling inspection and repair. We performed a sampling inspection for each annotation batch to assess its quality. Specifically, for expression and bounding box annotations, we established a sampling pass rate threshold of " + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": ", whereas, for expression translation, the threshold was set to " + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": ". The sampling inspection process" + } + ] + } + ], + "index": 46 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 414, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 545, + 450 + ], + "type": "text", + "content": "included no fewer than 100 samples per round, and the batch of data was accepted once it met the pass rate threshold or underwent repair." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 306, + 452, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 545, + 582 + ], + "type": "text", + "content": "- Box ensemble and de-duplication. In Step 3 of the annotation procedure, we obtained the ground truth by averaging the valid bounding boxes annotated by both groups of annotators. Any expressions that did not result in a valid bounding box were discarded. Furthermore, as data annotation is carried out in parallel among many annotation workers, we performed a dedduplication procedure to eliminate objects with overlapping bounding boxes and identical expressions across multiple images to enhance the data variance, with an IoU threshold empirically established at 0.3." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 305, + 587, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 545, + 624 + ], + "type": "text", + "content": "Despite potential undetected ambiguities, the implemented methodologies are believed to substantially mitigate their effects, thereby enhancing the dataset's overall quality." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 632, + 459, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 459, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 459, + 644 + ], + "type": "text", + "content": "3.3. Analyses of GigaGrounding" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 305, + 651, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 545, + 686 + ], + "type": "text", + "content": "This section analyzes the salient characteristics differentiating GigaGrounding from existing VG benchmarks. Our analysis encompasses the following viewpoints." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Quantities of objects per image Among conventional datasets, Flickr30K Entities in Figure 2a requires the model" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22122" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 168, + 153 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 168, + 153 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 168, + 153 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 168, + 153 + ], + "type": "image", + "image_path": "2f771c284aa1cc14660759392778b3b15d3b9eefc8c65a4a26124d601712278c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 158, + 171, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 158, + 171, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 171, + 177 + ], + "type": "text", + "content": "(a) Pixel-count distribution of bounding box scales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 293, + 154 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 293, + 154 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 293, + 154 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 293, + 154 + ], + "type": "image", + "image_path": "dce1e696e3e8b90c0fdd1d2b539888508b739869c83e6d3bc2879e5f13857d2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 237 + ], + "type": "text", + "content": "Figure 5. A comparative analysis of GigaGrounding against established benchmarks RefCOCO, RefCOCO+, and RefCOCOg. (a) and (b) elucidate the distribution patterns of bounding box scale (in terms of the max value of height and width) and relative area with respect to the image, (c) presents the distribution of expression lengths, and (d) shows the lengths' distribution of both direct and multi-hop expressions within the GigaGrounding dataset. GigaGrounding exhibits a notably elevated level of complexity, primarily stemming from its pronounced scale variation and intricate expression structures." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 414, + 154 + ], + "blocks": [ + { + "bbox": [ + 172, + 158, + 294, + 177 + ], + "lines": [ + { + "bbox": [ + 172, + 158, + 294, + 177 + ], + "spans": [ + { + "bbox": [ + 172, + 158, + 294, + 177 + ], + "type": "text", + "content": "(b) Relative area of bounding boxes w.r.t. the images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 72, + 414, + 154 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 414, + 154 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 414, + 154 + ], + "type": "image", + "image_path": "c49a5ac1eb51cba710b184cfdbb9629d98137ddaddf806b010f1a7ac95db44b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 418, + 72, + 543, + 154 + ], + "blocks": [ + { + "bbox": [ + 296, + 158, + 414, + 177 + ], + "lines": [ + { + "bbox": [ + 296, + 158, + 414, + 177 + ], + "spans": [ + { + "bbox": [ + 296, + 158, + 414, + 177 + ], + "type": "text", + "content": "(c) Word-count distribution of expression lengths." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 418, + 72, + 543, + 154 + ], + "lines": [ + { + "bbox": [ + 418, + 72, + 543, + 154 + ], + "spans": [ + { + "bbox": [ + 418, + 72, + 543, + 154 + ], + "type": "image", + "image_path": "3c481339ad783e0468048fc25a10dbf082910ed1f3720805e36f571fb3c0671c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 158, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 416, + 158, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 416, + 158, + 545, + 177 + ], + "type": "text", + "content": "(d) Word-count distribution of direct and multi-hop expressions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 252, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 252, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 45, + 252, + 287, + 361 + ], + "type": "text", + "content": "to identify and locate multiple objects corresponding to short phrases within a single sentence. However, we observe that the model may not need to fully understand the global semantics before grounding the corresponding object, particularly when there exists only one object of that type. On the other side, RefCOCO and RefCOCO+ include an average of 3.9 same-type objects per image, while RefCOCOg displays a relatively lower average of 1.63 same-type objects per image [2]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": "Scales of bounding boxes In Figures 5a and 5b, we respectively illustrate the absolute scale of bounding boxes and their relative area with respect to the images. It is evident that GigaGrounding exhibits greater scale variance in comparison to other datasets. Furthermore, bounding box scale in GigaGrounding adheres to a long-tail distribution, while other datasets generally show a more concentrated distribution. On average, the RefCOCO dataset features bounding boxes with dimensions of " + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "inline_equation", + "content": "212 \\times 268" + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": " pixels, accounting for approximately " + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "inline_equation", + "content": "18.49\\%" + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": " of the total image area. In contrast, the GigaGrounding dataset exhibits an average bounding box size of " + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "inline_equation", + "content": "478 \\times 978" + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": " pixels, representing just " + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "inline_equation", + "content": "0.12\\%" + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": " of the image area. This statistic reveals the inherent challenges associated with grounding models tasked with large-scale scenes. We also checked the distribution of direct and multi-hop bounding boxes and noticed that they share a similar distribution pattern, while the bounding boxes corresponding to multi-hop expressions are slightly smaller, with an average resolution of " + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "inline_equation", + "content": "425 \\times 938" + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": ", compared to " + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "inline_equation", + "content": "503 \\times 997" + }, + { + "bbox": [ + 46, + 363, + 288, + 602 + ], + "type": "text", + "content": " of direct expression bounding boxes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Lengths of expressions We demonstrate the word count distribution of expression lengths in Figure 5c. We note that the average expression length in RefCOCOg is 8.46 words, while in RefCOCO and RefCOCO+, it is only 3.50 and 3.53, respectively. In contrast, in order to locate objects in large-scale scenes without ambiguity, GigaGrounding provides significantly longer expressions, and the average expression length is 14.78 words. Flickr30K Entities present comparatively lengthy expressions, while it is tasked with" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "content": "ground objects referred by short region phrases. Figure 5d further illustrates the length distribution of direct expressions and multi-hop expressions, and the average lengths are 12.33 and 18.74, respectively. Longer expressions may accommodate more details about the objects' location, appearance, and properties, which can increase the semantic space that the model needs to consider, leading to challenges in both textual and imagery comprehension." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "text", + "content": "Other statistics We notice that GigaGrounding is certainly not offering the most number of images, bounding boxes, and annotated expressions among the compared datasets (Table 1). However, we believe that GigaGrounding's unique characteristics have opened the door for new modeling paradigms with practically useful and broader real-life applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 448, + 372, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 372, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 372, + 462 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 469, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 517 + ], + "type": "text", + "content": "In this section, we present an initial step towards effective visual grounding in the complex context of gigapixel-level large-scale scenes. We introduce a simple yet effective model, referred to as GlaZing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 526, + 435, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 435, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 435, + 538 + ], + "type": "text", + "content": "4.1. Architecture Overview" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "content": "In the context of GigaGrounding, human operators commonly begin with an initial coarse search across the image to identify regions of interest. Following this initial search, they then allocate their attention to the selected regions, subjecting them to thorough scrutiny. This methodical narrowing of focus significantly increases search efficacy. Our GlaZing model is conceived by analogizing this human cognitive heuristic." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "An overview of GlaZing is presented in Figure 6. Given an input image, it initiates by inputting the downsampled thumbnail and the corresponding expression into the Glance Grounding Module (GGM) for preliminary localization and grounding. Subsequently, the Adaptive Cropping Module (ACM) dynamically extracts the high-resolution patch" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22123" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 286, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 186 + ], + "type": "image", + "image_path": "279ea6629a6b9826a5097a8052c2721d6aecb6700f74cd6f054a95ac9b5df704.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 202, + 261, + 215 + ], + "lines": [ + { + "bbox": [ + 71, + 202, + 261, + 215 + ], + "spans": [ + { + "bbox": [ + 71, + 202, + 261, + 215 + ], + "type": "text", + "content": "Figure 6. Overview of the proposed GlaZing model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 287, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 267 + ], + "type": "text", + "content": "of interest. This extracted patch, along with the expression, is then forwarded to the Zoomed-in Grounding Module (ZGM) to generate the final prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 275, + 191, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 191, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 191, + 289 + ], + "type": "text", + "content": "4.2. Detailed Implementations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 366 + ], + "type": "text", + "content": "The architectural framework expounded in Section 4.1 allows for a multitude of implementations. Notably, GGM and ZGM can be seamlessly integrated with a wide array of conventional grounding models. In this section, we present our specific implementation, which yielded favorable results in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 441 + ], + "type": "text", + "content": "Glance grounding module We design GGM based on ReSC [28], which adopts an anchor-based grounding paradigm that divides the image into multiple grids and determines the grid to which the target box belongs. This design effectively locates the target-relevant region in the thumbnail image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 444, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 598 + ], + "type": "text", + "content": "Adaptive cropping module This module is devised for the adaptive extraction of a region of interest from the original image, guided by the GGM's output. More precisely, ACM initiates this process by first computing the coordinates for a patch of specific size located at the center of the bounding box predicted by the GGM. It subsequently proceeds to crop the minimum rectangular region that encloses both this designated patch and the bounding box predicted by the GGM. During model training, this process also encapsulates the ground-truth bounding box. Finally, the extracted region is resized to a specified target size. Through the utilization of ACM, our model can adaptively focus on regions of varying scales and locations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 602, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 662 + ], + "type": "text", + "content": "Zoomed-in grounding module After the first glance and zooming-in, the target bounding box should account for a significant proportion of the input patch of ZGM, so we employ a coordinate regression-based visual grounding method, i.e., TransVG [7] to accomplish this task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Training objectives The overall training objective can be defined as " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{GGM}} + \\mathcal{L}_{\\mathrm{ZGM}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ". We follow the training objectives employed in ReSC and TransVG for " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GGM}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ZGM}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " respectively. Specifically, " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GGM}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " comprises the YOLOv3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "loss [45] designed to optimize anchor association confidence and prediction offset, and a diversity loss [28] intended to guide sub-queries in each iteration to concentrate on distinct elements within the query. Concerning " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ZGM}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ", given that ZGM directly generates a 4-dimensional vector representing the relative coordinates of the grounded target box with respect to the input patch, we employ both the smooth " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\ell 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " loss and the generalized IoU loss [46] to minimize the disparity between the predicted and target values." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 190, + 378, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 378, + 201 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 209, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 329 + ], + "type": "text", + "content": "As an initial stride for GigaGrounding, GlaZing tackles the challenges posed by high-resolution input and scale variation through a \"glance-to-zoom-in\" cascading strategy coupled with adaptive cropping. Empirical findings in Section 5.3 attest to its effectiveness. Nevertheless, substantial opportunities for further refinement remain, which we leave as future work, including: (a) more tailored training objectives for GGM and ZGM, (b) end-to-end pipelines for seamless glancing and zooming-in, and (c) the feasibility of multiple iterations of glance and zoom adjustments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 342, + 386, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 386, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 386, + 357 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 363, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 388 + ], + "type": "text", + "content": "This section delineates the evaluation protocols employed, followed by results' presentation and analysis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 398, + 372, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 372, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 372, + 409 + ], + "type": "text", + "content": "5.1. Baselines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 417, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 584 + ], + "type": "text", + "content": "We conducted an evaluation of several publicly available baselines. As discussed in Section 2, two-stage methods rely on object detectors to generate region proposals. Given that object detection in large-scale scenes remains a challenge in its own right [40], we conducted brief experiments involving a representative open-source two-stage method, namely, MAttNet [47]. On the other hand, one-stage methods directly provide bounding box predictions without relying on separate object detection. In this regard, GlaZing is more aligned with one-stage methods. We evaluated state-of-the-art open-source one-stage methods, including ReSC [28], TransVG [7], RefTR [48], SeqTR [29], QR-Net [49], and SimREC [50]. A concise introduction to these baselines is provided in the Supplementary Material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": "In addition to visual grounding models, we also conducted a human evaluation to assess the ability of human subjects to complete the task when presented with down-sampled images as used by deep learning models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 643, + 422, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 422, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 422, + 655 + ], + "type": "text", + "content": "5.2. Evaluation Protocol" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Dataset split We randomly selected " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " distinct expressions for training, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " for validation, and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " for testing. This process resulted in 42,641 samples for the training set, 6,254 for the validation set, and 12,458 for the testing set." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22124" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "Metric Following the standard protocol, we assessed the performance using Precision@0.5, where the prediction is deemed correct if its IoU with the ground-truth box is larger than 0.5. We also provide information on inference time and GFLOPs to demonstrate the computational efficiency of each model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": "Image preprocessing In our empirical analysis, for the two-stage baseline MAttNet, features were extracted with 8-times downsampled images. We followed Yu et al. [47], who used ground-truth object bounding boxes to eliminate region proposal bottlenecks from the object detector.* On the other side, one-stage baseline models were evaluated using two distinct image resolutions: " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "1536 \\times 1536" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": ". The first resolution was in accordance with established benchmarks, while the latter was the feasible upper limit for processing by most baseline models given our computational constraints. At " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": ", the mean bounding box dimensions were roughly " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "12 \\times 25" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " pixels. This size expanded to " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "30 \\times 60" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " pixels at " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "1536 \\times 1536" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " resolution. To ensure fair comparison, for GlaZing, we employed a " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " thumbnail image for the GGM, and the ACM cropped the region of interest from " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "1536 \\times 1536" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " images. The cropped region was subsequently resized to " + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 148, + 290, + 364 + ], + "type": "text", + "content": " for the ZGM to identify the referred region." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "type": "text", + "content": "Implementation details For all baseline methods, we conducted experiments based on their official code. However, for the object detection component of MAttNet, due to the lack of maintenance, we substituted it with the corresponding implementation of Faster R-CNN [51] from Detectron2 [52]. The substitution did not hurt the performance on RefCOCO based on our experiments. For one-stage methods, we tried to optimize their performance within a certain hyper-parameter tuning budget. For instance, for models with official weights on RefCOCO provided, we experimented with whether or not to use these weights for initialization and reported the better results. At the resolution of " + }, + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "type": "inline_equation", + "content": "1536 \\times 1536" + }, + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "type": "text", + "content": ", we also tried initializing with weights from GigaGrounding at " + }, + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 367, + 289, + 618 + ], + "type": "text", + "content": ", and reported the better outcomes. We also adjusted some important hyperparameters including learning rate, etc. As for GlaZing, we trained the model with similar settings as ReSC. For all experiments, we set the maximum query length to 30, and the computations were performed on a server equipped with 8 NVIDIA RTX 3090 GPUs. Please refer to the Supplementary Material for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 628, + 168, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 168, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 168, + 640 + ], + "type": "text", + "content": "5.3. Overall Performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 647, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 683 + ], + "type": "text", + "content": "We present the evaluation results of our approach and the baselines in Table 2. Several key observations and analyses are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 388 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "content": "- Baseline models' limited efficacy in GigaGrounding: None of the baseline models achieved Precision@0.5 scores surpassing " + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "content": ". In contrast, human testers consistently achieved scores exceeding " + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "content": ", even on images of size " + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "content": ". It is also noteworthy that the performance on GigaGrounding dramatically diverged from that on other benchmark datasets, such as RefCOCO. For instance, the classical ReSC model outperformed other baselines on GigaGrounding, even surpassing models initialized with pretrained weights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "type": "text", + "content": "- Substantial advantages of GlaZing: GlaZing demonstrated a significant performance advantage over all baseline models. When compared to the highest-performing deep learning baseline at a resolution of " + }, + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "1536 \\times 1536" + }, + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "type": "text", + "content": " (i.e., ReSC), GlaZing achieved an impressive relative improvement of " + }, + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "28.6\\%" + }, + { + "bbox": [ + 306, + 194, + 545, + 277 + ], + "type": "text", + "content": ". This substantial gain can be attributed to its glance-to-zoom-in design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 280, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 280, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 545, + 388 + ], + "type": "text", + "content": "- Incompatibility of existing two-stage methods: We noticed that MAttNet exhibited marginal improvements over random guessing (6.8% vs. 6.2%), despite that it used the ground-truth object bounding boxes as region proposals. Our hypothesis is that the bottleneck lies in the feature extractors. Two-stage methods typically use frozen feature extractors pretrained on MS COCO or ImageNet [53], which may not be effective enough for large-scale scene understanding." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 391, + 535, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 391, + 535, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 391, + 535, + 403 + ], + "type": "text", + "content": "Additionally, our study reveals the following insights:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 406, + 545, + 540 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 406, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 545, + 453 + ], + "type": "text", + "content": "- Influence of resolution on performance: Our findings highlight the critical role of high-resolution input in enhancing the performance of both deep learning models and human evaluators." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 456, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 545, + 540 + ], + "type": "text", + "content": "- Performance variation with expression complexity: Notably, the GlaZing model demonstrated consistent performance across both direct and multi-hop expressions. In contrast, baseline models exhibited a marked performance decline when handling multi-hop expressions. This disparity underscores the need for models to adapt to expressions with differing levels of complexity." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 546, + 411, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 411, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 411, + 559 + ], + "type": "text", + "content": "5.4. In-depth Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "content": "Effective high-resolution processing of GlaZing. GlaZing's adaptive cropping mechanism allows for higher-resolution image processing without significant storage and computation overhead. Experiments with 8-times downsampled images in ACM and ZGM showed promising results: " + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "inline_equation", + "content": "66.5\\%" + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "content": " on the test set, with " + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "inline_equation", + "content": "66.2\\%" + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "content": " for direct and " + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "inline_equation", + "content": "67.5\\%" + }, + { + "bbox": [ + 304, + 567, + 545, + 674 + ], + "type": "text", + "content": " for multi-hop expressions. These findings suggest that processing at higher resolutions may further enhance the model's performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Benefits of RefCOCO initialization. We report the results of TransVG, RefTR, and SeqTR with and without RefCOCO weight initialization in Table 3. The adoption of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "text", + "content": "*Consequently, it is inappropriate to directly compare the results of one-stage and two-stage methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22125" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 544, + 253 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 253 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 253 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 253 + ], + "type": "table", + "html": "
CategoryMethodBackboneoverallVal directmulti-hopoverallTest directmulti-hopInfer TimeGFLOPsTrainable Params.
Two-stageMAttNetResNet-10111.4%11.6%10.7%6.8%6.4%7.6%268.6ms1592.613.0M
One-stage (640×640)ReSCDenseNet-5330.8%35.5%21.2%29.0%33.3%20.3%53.4ms101.3158.2M
QRNetSwin-S16.2%19.5%9.4%13.8%16.8%7.9%54.8ms79.2247.1M
SimRECCSPDarkNet-5323.3%27.6%14.4%20.8%24.7%12.7%16.5ms48.940.2M
TransVG‡ResNet-10113.1%15.9%7.1%11.6%13.7%7.3%25.9ms41.4122.6M
RefTR‡ResNet-5025.4%29.0%18.1%21.8%25.0%15.3%45.0ms23.5123.4M
SeqTR‡Darknet-5317.6%20.1%12.4%15.9%18.1%11.6%39.8ms48.9100.9M
Human------62%§----------
One-stage (1536×1536)ReSCDenseNet-5352.0%57.5%40.8%49.7%56.0%37.0%96.6ms566.3158.2M
QRNetSwin-S28.8%33.3%19.5%26.5%31.0%17.3%141.6ms435.0247.1M
SimRECCSPDarkNet-5335.4%41.7%22.5%33.3%39.5%20.8%48.4ms280.940.2M
TransVG‡ResNet-10126.3%31.2%16.5%24.6%29.1%15.4%72.6ms225.3122.6M
RefTR‡ResNet-5042.4%48.4%30.0%41.0%46.7%29.5%53.0ms119.7123.4M
SeqTR‡Darknet-5335.6%41.4%23.5%33.2%38.3%22.8%51.2ms263.7100.9M
Human------84%§----------
GlaZing (Ours)65.0%65.3%64.3%63.9%64.1%63.7%110.2ms172.1280.8M
", + "image_path": "a7d6ffc4525491ed90a560e856b7c587d190b8883ac29dc8addb73e893acdaf0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 122, + 254, + 538, + 262 + ], + "lines": [ + { + "bbox": [ + 122, + 254, + 538, + 262 + ], + "spans": [ + { + "bbox": [ + 122, + 254, + 538, + 262 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 122, + 254, + 538, + 262 + ], + "type": "text", + "content": " We used the released weights pretrained on RefCOCO for initialization. " + }, + { + "bbox": [ + 122, + 254, + 538, + 262 + ], + "type": "inline_equation", + "content": "{}^{8}" + }, + { + "bbox": [ + 122, + 254, + 538, + 262 + ], + "type": "text", + "content": " Estimated with 100 randomly selected samples from the testing set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 292, + 258, + 345 + ], + "blocks": [ + { + "bbox": [ + 59, + 267, + 533, + 278 + ], + "lines": [ + { + "bbox": [ + 59, + 267, + 533, + 278 + ], + "spans": [ + { + "bbox": [ + 59, + 267, + 533, + 278 + ], + "type": "text", + "content": "Table 2. Performance Evaluation on the GigaGrounding Dataset. Both accuracy-related and efficiency-related metrics are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 292, + 258, + 345 + ], + "lines": [ + { + "bbox": [ + 75, + 292, + 258, + 345 + ], + "spans": [ + { + "bbox": [ + 75, + 292, + 258, + 345 + ], + "type": "table", + "html": "
TransVGRefTRSeqTR
w/o pretraining5.1%0.0%5.3%
w/ pretraining11.6%21.8%15.9%
Absolute Gain+6.5%+21.8%+10.6%
", + "image_path": "bf381f7a3d802b2a420ba1bdc29d0c2883fcd4bc5e2e7434c078e0bb6b657016.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 55, + 392, + 277, + 447 + ], + "blocks": [ + { + "bbox": [ + 47, + 348, + 287, + 381 + ], + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 381 + ], + "type": "text", + "content": "Table 3. Performance improvements on " + }, + { + "bbox": [ + 47, + 348, + 287, + 381 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 47, + 348, + 287, + 381 + ], + "type": "text", + "content": " input images achieved through pretraining. We initiated each model with publicly available weights trained on the RefCOCO dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 392, + 277, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 277, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 277, + 447 + ], + "type": "table", + "html": "
Expr. length & proportionReSCTransVGGlaZing
1-10 (16.99%)65.9%42.4%66.8%
11-20 (62.77%)47.8%22.6%63.5%
21+ (20.24%)39.7%13.8%62.6%
", + "image_path": "2cd65dc615f5943f65ca1a5d88247bdbaa63469bb2e374dfad6800d4e2c139ad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 73, + 478, + 260, + 545 + ], + "blocks": [ + { + "bbox": [ + 63, + 450, + 269, + 460 + ], + "lines": [ + { + "bbox": [ + 63, + 450, + 269, + 460 + ], + "spans": [ + { + "bbox": [ + 63, + 450, + 269, + 460 + ], + "type": "text", + "content": "Table 4. Performance stratification by expression length." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 478, + 260, + 545 + ], + "lines": [ + { + "bbox": [ + 73, + 478, + 260, + 545 + ], + "spans": [ + { + "bbox": [ + 73, + 478, + 260, + 545 + ], + "type": "table", + "html": "
B-box scaleReSCTransVGGlaZing
s < Q129.4%4.5%48.1%
Q1 ≤ s < Q250.2%19.3%63.6%
Q2 ≤ s < Q358.2%31.1%72.8%
s ≥ Q360.7%43.6%70.9%
", + "image_path": "6c70398b264650266ce55b76ec646b22c0f60d65218f061820d9a66cfcea7fbb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "text", + "content": "Table 5. Performance stratification by bounding box scale. Here, " + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "inline_equation", + "content": "Q1" + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "inline_equation", + "content": "Q2" + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "inline_equation", + "content": "Q3" + }, + { + "bbox": [ + 47, + 548, + 287, + 582 + ], + "type": "text", + "content": " denote the first, second, and third quartiles of bounding box scale, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 590, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 626 + ], + "type": "text", + "content": "pretrained weights from RefCOCO led to performance enhancements for three methods, despite the substantial differences between RefCOCO and GigaGrounding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 714 + ], + "type": "text", + "content": "Performance about expression length and B-box scale. We conducted an in-depth analysis of the performance variability about expression length and B-box scale at " + }, + { + "bbox": [ + 47, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "1536 \\times 1536" + }, + { + "bbox": [ + 47, + 630, + 287, + 714 + ], + "type": "text", + "content": ", specifically focusing on the ReSC, TransVG, and GlaZing. The results are presented in Tables 4 and 5. Our findings indicate that GlaZing demonstrated a notable robustness across varying difficulty levels. Conversely," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": "TransVG exhibited a more pronounced performance degradation, particularly in scenarios involving longer expressions and smaller B-box sizes. Complete results of all baselines can be found in the Supplementary Material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "text", + "content": "Analysis of the glance-to-zoom-in strategy in GlaZing. Our investigation reveals that GlaZing's ZGM effectively rectifies prediction errors observed in the GGM, underscoring the effectiveness of the glance-to-zoom-in approach for GigaGrounding. Specifically, we identified that approximately " + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "text", + "content": " of errors fall under \"target feature mismatch\", with ZGM correcting " + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "text", + "content": " of these. Additionally, another " + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "text", + "content": " of errors are categorized as \"over-fuzziness due to downsampling\", of which ZGM amends about " + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 345, + 546, + 477 + ], + "type": "text", + "content": ". Due to space constraints, detailed explanations and analyses are provided in the Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 626 + ], + "type": "text", + "content": "This paper presents GigaGrounding, a novel and challenging benchmark for visual grounding in gigapixel-level large-scale scenes, accompanied by an efficient solution strategy. The distinctive challenges posed by GigaGrounding establish it as a pioneering benchmark in this domain. Our goal is to stimulate further research in visual grounding, aiming to propel the development of models that are both more accurate and efficient. These advancements are particularly crucial for real-world applications such as rescue operations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This work is supported in part by Natural Science Foundation of China (NSFC) under contract No. 62125106 and 62088102, in part by Ministry of Science and Technology of China under contract No. 2021ZD0109901, in part by Tsinghua-Zhijiang joint research center, in part by Young Elite Scientists Sponsorship Program by CAST under contract No. 2022QNRC001." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22126" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 169 + ], + "type": "text", + "content": "[2] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "text", + "content": "[3] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 258 + ], + "type": "text", + "content": "[4] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 260, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 286, + 293 + ], + "type": "text", + "content": "[5] Ronghang Hu, Huazhe Xu, Marcus Rohrbach, Jiashi Feng, Kate Saenko, and Trevor Darrell. Natural language object retrieval. In CVPR, 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 286, + 338 + ], + "type": "text", + "content": "[6] Jianan Li, Yunchao Wei, Xiaodan Liang, Fang Zhao, Jianshu Li, Tingfa Xu, and Jiashi Feng. Deep attribute-preserving metric learning for natural language object retrieval. In ACM MM, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 340, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 286, + 373 + ], + "type": "text", + "content": "[7] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-end visual grounding with transformers. In ICCV, 2021. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "type": "text", + "content": "[8] Xiaoyun Yuan, Mengqi Ji, Jiamin Wu, David J Brady, Qionghai Dai, and Lu Fang. A modular hierarchical array camera. Light: Sci. Appl., 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 286, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 286, + 462 + ], + "type": "text", + "content": "[9] Byford Sam. Samsung announces 200-megapixel phone camera sensor. https://www.theverge.com/2021/9/2/22653558/samsung-isocell-hpl-gn5-200-megapixel-camera-sensor-announced, 2021.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "type": "text", + "content": "[10] Arian Bakhtiarnia, Qi Zhang, and Alexandros Iosifidis. Efficient high-resolution deep learning: A survey. arXiv preprint arXiv:2207.13050, 2022. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 286, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 286, + 552 + ], + "type": "text", + "content": "[11] Xueyang Wang, Xiya Zhang, Yinheng Zhu, Yuchen Guo, Xiaoyun Yuan, Liuyu Xiang, Zerun Wang, Guiguang Ding, David Brady, Qionghai Dai, and Lu Fang. PANDA: A gigapixel-level human-centric video dataset. In CVPR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 286, + 599 + ], + "type": "text", + "content": "[12] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 286, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 286, + 654 + ], + "type": "text", + "content": "[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A González, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated IAPR TC-12 benchmark. CVIU, 2010. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 286, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 286, + 700 + ], + "type": "text", + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "text", + "content": "[15] Yuxi Sun, Shanshan Feng, Xutao Li, Yunming Ye, Jian" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Kang, and Xu Huang. Visual grounding in remote sensing images. In ACM MM, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[16] Yang Zhan, Zhitong Xiong, and Yuan Yuan. RSVG: Exploring data and models for visual grounding on remote sensing data. TGRS, 61:1-13, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 194 + ], + "type": "text", + "content": "[17] Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, et al. Making the most of text semantics to improve biomedical vision-language processing. In ECCV, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[18] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[19] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In ICCV, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[20] Richang Hong, Daqing Liu, Xiaoyu Mo, Xiangnan He, and Hanwang Zhang. Learning to compose and reason with language tree structures for visual grounding. TPAMI, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "text", + "content": "[21] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In CVPR, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 396 + ], + "type": "text", + "content": "[22] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In ICCV, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "type": "text", + "content": "[23] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In CVPR, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[24] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 489, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 532 + ], + "type": "text", + "content": "[25] Xinpeng Chen, Lin Ma, Jingyuan Chen, Zequn Jie, Wei Liu, and Jiebo Luo. Real-time referring expression comprehension by single-stage grounding network. arXiv preprint arXiv:1812.03426, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 534, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 576 + ], + "type": "text", + "content": "[26] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In CVPR, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "type": "text", + "content": "[27] Arka Sadhu, Kan Chen, and Ram Nevatia. Zero-shot grounding of objects from natural language queries. In ICCV, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[28] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In ECCV, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[29] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. SeqTR: A simple yet universal network for visual grounding. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] Jianing Zhang, Tianyi Zhu, Anke Zhang, Xiaoyun Yuan, Zihan Wang, Sebastian Beetschen, Lan Xu, Xing Lin, Qionghai Dai, and Lu Fang. Multiscale-VR: Multiscale gigapixel" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22127" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "3d panoramic videography for virtual reality. In ICCP, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[31] Xiaoyun Yuan, Lu Fang, Qionghai Dai, David J Brady, and Yebin Liu. Multiscale gigapixel video: A cross resolution image matching and warping approach. In ICCP, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 131, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 131, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 131, + 287, + 175 + ], + "type": "text", + "content": "[32] Jianing Zhang, Jinzhi Zhang, Shi Mao, Mengqi Ji, Guangyu Wang, Zequn Chen, Tian Zhang, Xiaoyun Yuan, Qionghai Dai, and Lu Fang. GigaMVS: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. TPAMI, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "type": "text", + "content": "[33] Guangyu Wang, Jinzhi Zhang, Kai Zhang, Ruqi Huang, and Lu Fang. GiganticNVS: Gigapixel large-scale neural rendering with implicit meta-deformed manifold. TPAMI, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "text", + "content": "[34] Xueyang Wang, Xuecheng Chen, Puhua Jiang, Haozhe Lin, Xiaoyun Yuan, Mengqi Ji, Yuchen Guo, Ruqi Huang, and Lu Fang. The group interaction field for learning and explaining pedestrian anticipation. Engineering, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 257, + 287, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 299 + ], + "type": "text", + "content": "[35] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In ECCV, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 302, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 287, + 345 + ], + "type": "text", + "content": "[36] Dmitrii Marin, Zijian He, Peter Vajda, Priyam Chatterjee, Sam Tsai, Fei Yang, and Yuri Boykov. Efficient segmentation: Learning downsampling near semantic boundaries. In ICCV, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 347, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 390 + ], + "type": "text", + "content": "[37] Chen Jin, Ryutaro Tanno, Thomy Mertzanidou, Eleftheria Panagiotaki, and Daniel C Alexander. Learning to down-sample for segmentation of ultra-high resolution images. In ICLR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "text", + "content": "[38] Mingfei Gao, Ruichi Yu, Ang Li, Vlad I Morariu, and Larry S Davis. Dynamic zoom-in network for fast object detection in large images. In CVPR, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 428, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 449 + ], + "type": "text", + "content": "[39] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "text", + "content": "[40] Kai Chen, Zerun Wang, Xueyang Wang, Dahan Gong, Longlong Yu, Yuchen Guo, and Guiguang Ding. Towards real-time object detection in gigapixel-level video. Neurocomputing, 2022. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "type": "text", + "content": "[41] Maria Tzelepi and Anastasios Tefas. Improving the performance of lightweight CNNs for binary classification using quadratic mutual information regularization. Pattern Recogn., 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 543, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 585 + ], + "type": "text", + "content": "[42] Huangjing Lin, Hao Chen, Simon Graham, Qi Dou, Nasir Rajpoot, and Pheng-Ann Heng. Fast ScanNet: Fast and dense analysis of multi-gigapixel whole-slide images for cancer metastasis detection. TMI, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 609 + ], + "type": "text", + "content": "[43] Using Creative Commons Public Licenses. Attribution-noncommercial-sharealike 4.0 international. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[44] Baowei Jiang, Bing Bai, Haozhe Lin, Yu Wang, Yuchen Guo, and Lu Fang. DartBlur: Privacy preservation with detection artifact suppression. In CVPR, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "type": "text", + "content": "[45] Joseph Redmon and Ali Farhadi. YOLOv3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[46] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 385 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. MAttNet: Modular attention network for referring expression comprehension. In CVPR, 2018. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[48] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In CVPR, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "text", + "content": "[50] Gen Luo, Yiyi Zhou, Jiamu Sun, Shubin Huang, Xiaoshuai Sun, Qixiang Ye, Yongjian Wu, and Rongrong Ji. What goes beyond multi-modal fusion in one-stage referring expression comprehension: An empirical study. arXiv preprint arXiv:2204.07913, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 263, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 296 + ], + "type": "text", + "content": "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "type": "text", + "content": "[52] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22128" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_content_list.json b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d79e37e231f16b5232242a135b237e99a2817fac --- /dev/null +++ b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "Why Not Use Your Textbook? Knowledge-Enhanced Procedure Planning of Instructional Videos", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kumaranage Ravindu Yasas Nagasinghe $^{1}$ Honglu Zhou $^{2}$ Malitha Gunawardhana $^{1,3}$", + "bbox": [ + 119, + 202, + 849, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Martin Renqiang Min $^{2}$ Daniel Harari $^{4}$ Muhammad Haris Khan $^{1}$", + "bbox": [ + 183, + 227, + 784, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Mohamed bin Zayed University of Artificial Intelligence, 2NEC Laboratories, USA,", + "bbox": [ + 232, + 247, + 736, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Auckland, 4Weizmann Institute of Science", + "bbox": [ + 313, + 265, + 656, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ravindu.nagasinghe@mbzuai.ac.ae, muhammad.haris@mbzuai.ac.ae", + "bbox": [ + 210, + 284, + 743, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 333, + 313, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the capability of an agent to construct a logical sequence of action steps, thereby assembling a strategic procedural plan. This plan is crucial for navigating from an initial visual observation to a target visual outcome, as depicted in real-life instructional videos. Existing works have attained partial success by extensively leveraging various sources of information available in the datasets, such as heavy intermediate visual observations, procedural names, or natural language step-by-step instructions, for features or supervision signals. However, the task remains formidable due to the implicit causal constraints in the sequencing of steps and the variability inherent in multiple feasible plans. To tackle these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities by infusing it with procedural knowledge. This knowledge, sourced from training procedure plans and structured as a directed weighted graph, equips the agent to better navigate the complexities of step sequencing and its potential variations. We coin our approach KEPP, a novel Knowledge-Enhanced Procedure Planning system, which harnesses a probabilistic procedural knowledge graph extracted from training data, effectively acting as a comprehensive textbook for the training domain. Experimental evaluations across three widely-used datasets under settings of varying complexity reveal that KEPP attains superior, state-of-the-art results while requiring only minimal supervision. Code and trained model are available at https://github.com/Ravindu-Yasas-Nagasinghe/KEPP", + "bbox": [ + 76, + 368, + 472, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 827, + 209, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The evolution of the internet has precipitated an unprecedented influx of video content, serving as a vital educational resource for myriad learners. Individuals frequently", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0138b6fb7c6793c604dae03f21e39301720353eac987411ca960672a8fd230ce.jpg", + "image_caption": [ + "Figure 1. Expert trajectories [7] of the 'Make Jello Shots' task from the CrossTask dataset [63]. Heavier color indicates more frequently visited path. This depicts the complexities of the procedure planning task, arising from the subtle causal links in step sequencing (e.g., steps like 'stir mixture' or 'pour mixture' typically occur after adding individual ingredients), the varied probabilities of transitioning between steps, and the diversity in plans viable for a given starting point and an intended outcome. Motivated by these nuanced challenges, we propose Knowledge-Enhanced Procedure Planning (KEPP) with the use of a probabilistic procedure knowledge graph to capture and represent these intricacies" + ], + "image_footnote": [], + "bbox": [ + 504, + 335, + 890, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leverage platforms such as YouTube to acquire new skills, ranging from culinary arts to automobile maintenance [34]. While these instructional videos benefit the development of intelligent agents in mastering long-horizon tasks, the challenge extends beyond merely interpreting visuals. It requires high-level reasoning and planning to effectively assist in complex, real-life scenarios [13].", + "bbox": [ + 496, + 686, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Procedure Planning in Instructional Videos demands an agent to produce a sequence of actionable steps, thereby crafting a procedure plan that facilitates the transition from an initial visual observation of the physical world towards achieving a desired goal state [7, 9, 50, 53, 54, 59]. The task acts as a precursor to an envisioned future scenario in which an agent like a robot provides on-the-spot support, such as", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18816", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "assisting an individual in preparing a recipe [6].", + "bbox": [ + 76, + 90, + 393, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current methods in procedure planning in instructional videos make extensive use of various annotations available within the datasets to enrich input features or provide supervisory signals (see Table 1). These include detailed, temporally localized visual observations of intermediate action steps throughout the procedure plan [7, 9, 50], high-level procedural task labels [53, 54], and step-by-step instructions in natural language [53, 59]. Despite advancements, significant challenges persist, including characterizing the implicit causal constraints in step sequencing, the varied probabilities of transitioning between steps, and the inherent variability of multiple viable plans (see Fig. 1).", + "bbox": [ + 75, + 108, + 468, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities in procedure planning by infusing it with comprehensive procedural knowledge [62], derived from training procedure plans and structured as a directed weighted graph. This graph, as a Probabilistic Procedural Knowledge Graph [5] where nodes denote steps from diverse tasks and edges represent step transition probabilities in the training domain, empowers agents to more adeptly navigate the complexities of step sequencing and its potential variations.", + "bbox": [ + 75, + 291, + 470, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed approach, KEPP, is a novel Knowledge-Enhanced Procedure Planning system (Fig. 2) that harnesses a probabilistic procedural knowledge graph $(\\mathrm{P}^2\\mathrm{KG})$ , constructed from training procedure plans. This graph functions like a detailed textbook, providing extensive knowledge for the training domain, and thereby circumventing the need for costly multiple annotations required by existing methods. Additionally, we decompose the instructional video procedure planning problem into two parts: one driven by objectives specific to step perception and the other by a procedural knowledge-informed modeling of procedure planning. In this problem decomposition, the first and last action steps are predicted based on the initial and goal visual states. Following this, a procedure plan is crafted by leveraging the procedure plan recommendations retrieved from the $\\mathrm{P}^2\\mathrm{KG}$ . The recommendations correspond to the most probable procedure plans frequently used in training, conditioned on the predicted first and last action steps. In a similar vein to the approach by Li et al. [29], our proposed decomposition strategy reduces uncertainty by maximizing the use of currently available information, namely the initial and goal visual states. This allows for the improvement of procedure planning through more accurate predictions of start and end actions. Plus, this decomposition effectively incorporates procedural knowledge into procedure planning, thereby enhancing its effectiveness.", + "bbox": [ + 75, + 444, + 470, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 94, + 838, + 315, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos that leverages rich procedural knowledge from a probabilistic procedu", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ral knowledge graph (P2KG). This approach necessitates only a minimal amount of annotations for supervision.", + "bbox": [ + 511, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We decompose the problem in procedure planning of instructional videos: predicting the initial and final steps from the start and end visuals, and then creating a plan using procedural knowledge retrieved based on these predicted steps. This approach prioritizes the currently available information and effectively incorporates procedural knowledge, enhancing strategic planning.", + "- Experimental evaluations on three widely-used datasets, under settings of varying complexity, reveal that KEPP attains state-of-the-art results in procedure planning." + ], + "bbox": [ + 500, + 121, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 286, + 640, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instructional Videos, which demonstrate multi-step procedures, have become a hotbed of research. The studies delve into various aspects, including comprehending and extracting intricate spatiotemporal content from video [12, 18, 19, 21, 23, 36, 43, 44, 48, 55, 57], interpreting the interrelationships between various actions and procedural events [47, 63], and developing capabilities for forecasting [39, 42] and strategic reasoning and planning [28] within the context of these videos. Furthermore, by leveraging the multimodality of visual, auditory, and narrative elements within these videos, research extends to areas like multimodal alignment [2, 58], grounding [10, 14, 25, 33, 51], representation learning [11, 35, 61], pre-training [15, 26, 62], and more [17, 24, 37, 56]. This paper focuses on procedure planning in instructional videos.", + "bbox": [ + 496, + 311, + 893, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Procedure Planning is a vital skill for autonomous agents tasked with handling complex activities in everyday settings. Essentially, these agents must discern the appropriate actions to reach a specific goal. This aspect of artificial intelligence (AI) has been a prominent and integral subject in fields like robotics [20, 28, 32, 45, 46]. Yet, the challenge of procedure planning in the context of instructional videos is notably distinct, and potentially more complex, than its counterparts in natural language processing [8, 30], multimodal generative AI [13, 31], and simulated environments [27, 28, 45]. Its significance is underscored by the need for planning that is grounded in real-world scenes. This requires the development of AI agents capable of accurately perceiving and understanding the current real-world context, and then anticipating and mapping out a logical sequence of actions to fulfill a high-level goal effectively.", + "bbox": [ + 496, + 537, + 893, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Procedure Planning in Instructional Videos has recently garnered research attention. DDN [9] initiates this trend by conceptualizing the problem as sequential latent space planning. Building on this, PlaTe [50] employs transformers for both action and state models, integrating Beam Search for enhanced performance. Meanwhile, Ext-GAIL [7] suggests employing contextual modeling through variational autoencoder and adversarial policy learning. This method", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18817", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "considers contextual information as time-invariant knowledge, crucial for distinguishing specific tasks and allowing for multiple planning outcomes.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these earlier approaches have viewed procedure planning as an autoregressive sequence generation problem, recent methods regard it as a distribution-fitting problem to mitigate error propagation in sequential decisions. In this vein, $\\mathrm{P}^3\\mathrm{IV}$ [59] replaces intermediate visual states with linguistic representations for supervision, predicting all steps simultaneously instead of using autoregressive methods. To circumvent the complex learning strategies and high annotation costs of previous work, PDPP [54] models the entire intermediate action sequence distribution using a conditioned projected diffusion model. This approach redefines the planning problem as a sampling process from this distribution and simplifies supervision by using only instructional video task labels. E3P [53], also encoding task information, adopts a mask-and-predict strategy for mining step relationships in procedural tasks, integrating probabilistic masking for regularization. In contrast, our approach does not rely on annotations of intermediate states, natural language step representations, or procedural task labels.", + "bbox": [ + 75, + 136, + 470, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recognizing the difficulties inherent in high dimensional state supervision and the accumulation of errors in action sequences, SkipPlan [29] was developed. It strategically focuses on action predictions, breaking down longer sequences into shorter, more manageable sub-chains by skipping over less reliable intermediate actions. Drawing inspiration from SkipPlan, our approach decomposes the procedure planning problem to prioritize the most reliable information available (ref. § 3.1.2). However, we innovate further by incorporating a Probabilistic Procedure Knowledge Graph, significantly enriching the planning phase.", + "bbox": [ + 75, + 424, + 468, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 604, + 210, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will first introduce the problem setup in § 3.1, and then discuss our novel Knowledge-Enhanced Procedure Planning system (KEPP) in § 3.2. See Fig. 2 for KEPP overview.", + "bbox": [ + 75, + 630, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem and Method Overview", + "text_level": 1, + "bbox": [ + 76, + 686, + 354, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 709, + 282, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the problem definition for procedure planning of instructional videos put forth by Chang et al. [9]: given an observation of the initial state $v_{start}$ and a goal state $v_{goal}$ , both are short video clips indicating different states of the real-world environment extracted from an instructional video, a model is required to plan a sequence of action steps $a_{1:T}$ to reach the indicated goal. Here, $T$ is the planning horizon, inputting to the model, corresponding to the number of action steps in the sequence produced by the model so that the environment state can be transformed from $v_{start}$ to $v_{goal}$ . We use $a_t$ to denote the action step at the timestamp", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$t$ , and in the following, $v_{s}$ and $v_{g}$ are short for $v_{start}$ and $v_{goal}$ . Mathematically, the procedure planning problem is defined as $p(a_{1:T}|v_s,v_g)$ that denotes the conditional probability distribution of the action sequence $a_{1:T}$ given the initial visual observation $v_{start}$ and the goal visual state $v_{goal}$ .", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Problem Decomposition", + "text_level": 1, + "bbox": [ + 498, + 190, + 720, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the initial and final visual states are input, providing the most reliable information, we hypothesize that predicting the first and final action steps is more dependable than interpolating the intermediate ones, and consequently, an enhanced accuracy in predicting the first and final steps can lead to more effective procedure planning. Inspired by this hypothesis, we decompose the procedure planning problem into two sub-problems, as shown in Eq. 1:", + "bbox": [ + 496, + 219, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\hat {a} _ {1: T} | v _ {s}, v _ {g}\\right) = p \\left(\\hat {a} _ {2: T - 1} | \\hat {a} _ {1}, \\hat {a} _ {T}\\right) p \\left(\\hat {a} _ {1}, \\hat {a} _ {T} | v _ {s}, v _ {g}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 359, + 890, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the first sub-problem is to identify the beginning step $a_1$ and the end step $a_T$ , and the second sub-problem is to plan the intermediate action steps $a_{2:T-1}$ given $a_1$ and $a_T$ . We use $\\hat{a}_t$ to denote predicted action step at timestamp $t$ .", + "bbox": [ + 496, + 393, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed problem decomposition in Eq. 1 bears resemblance with the problem formulation from Li et al. [29]; they decompose procedure planning into $p(\\hat{a}_{1:T}|v_s, v_g) = \\prod_{t=2}^{T-1} p(\\hat{a}_t|\\hat{a}_1, \\hat{a}_T)p(\\hat{a}_1, \\hat{a}_T|v_s, v_g)$ . However, our formulation differs in its approach to modeling the second subproblem. Specifically, we employ a conditioned projected diffusion model (ref. § 3.2) to jointly predict $a_{2:T-1}$ at once, whereas Li et al. [29] rely on Transformer decoders to predict each intermediate action independently. Further, we integrate a Probabilistic Procedure Knowledge Graph (ref. § 3.2.2) to address the second sub-problem.", + "bbox": [ + 496, + 457, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tackling the second sub-problem is nontrivial even when armed with an oracle predictor for the first sub-problem. Procedure planning in real-life scenarios remains daunting because of the following challenges: (1) the presence of implicit temporal and causal constraints in the sequencing of steps, (2) the existence of numerous viable plans given an initial state and a goal state, and (3) the need to incorporate the real-life everyday knowledge both in task-sharing steps and in managing the inherent variability in transition probabilities between steps. Previous studies tackled these challenges by extensively harnessing detailed annotations in the datasets to augment input features or offer supervision signals (see Table 1). In contrast, we propose harnessing a Probabilistic Procedural Knowledge Graph $(\\mathrm{P}^2\\mathrm{KG})$ which is extracted from the procedure plans in the training set. With the $\\mathrm{P}^2\\mathrm{KG}$ at our hand, we further decompose the procedure planning problem to reduce its complexity and learn $f_{\\theta}:(v_s,v_g,T)\\to p(\\hat{a}_{1:T}|v_s,v_g)$ as follows:", + "bbox": [ + 496, + 628, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1cf3937d5a075be8ca996915459c39187231a319a4a72393896a35b18cb5fe87.jpg", + "image_caption": [ + "Figure 2. Overview of our methodology. We introduce KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos, leveraging a Probabilistic Procedural Knowledge Graph $(\\mathrm{P}^2\\mathrm{KG})$ . KEPP breaks down procedure planning into two parts: predicting initial and final steps from visual states, and crafting a procedure plan based on the procedural knowledge retrieved from $\\mathrm{P}^2\\mathrm{KG}$ , conditioned on the predicted first and last action steps. KEPP requires minimal annotations and enhances planning effectiveness" + ], + "image_footnote": [], + "bbox": [ + 83, + 79, + 890, + 215 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(\\hat {a} _ {1: T} | v _ {s}, v _ {g}\\right) = \\\\ p \\left(\\hat {a} _ {1: T} | \\tilde {a} _ {1: T}, v _ {s}, v _ {g}\\right) p \\left(\\tilde {a} _ {1: T} | \\hat {a} _ {1}, \\hat {a} _ {T}\\right) p \\left(\\hat {a} _ {1}, \\hat {a} _ {T} | v _ {s}, v _ {g}\\right) \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 329, + 468, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta}$ denotes the machine learning model, and $\\tilde{a}_{1:T}$ represents a graph path (i.e., a sequence of nodes) retrieved from $\\mathrm{P}^2\\mathrm{KG}$ . This graph path provides a valuable procedure plan recommendation aligned with the training domain, thus mitigating the complexity of procedure planning. It is worth noting that the proposed approach to modeling procedure planning using Eq. 2 demands only a minimal level of supervision, merely relying on the ground truth training procedure plan; Eq. 2 circumvents the need for additional annotations. We describe details of our $\\mathrm{P}^2\\mathrm{KG}$ -enhanced approach in the following subsection.", + "bbox": [ + 75, + 375, + 468, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. KEPP: Knowledge-Enhanced Procedure Planning", + "text_level": 1, + "bbox": [ + 76, + 547, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose KEPP (Fig. 2) utilizing a probabilistic procedure knowledge graph extracted from the training set. We firstly identify the beginning and conclusion steps according to the input initial and goal states; and then, conditioned on these steps and the planning horizon $T$ , we query the graph to retrieve relevant procedural knowledge for knowledge-enhanced procedure planning of instructional videos.", + "bbox": [ + 75, + 585, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Identify Beginning and Conclusion Steps", + "text_level": 1, + "bbox": [ + 76, + 710, + 415, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $v_{start}$ and $v_{goal}$ as input, we adapt a Conditioned Projected Diffusion Model [54] (ref. supplementary material) to identify the first action step and the final step; we refer to this model as the 'Step (Perception) Model'.", + "bbox": [ + 75, + 734, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Standard Denoising Diffusion Probabilistic Model tackles data generation through a denoising Markov chain over variables $\\{x_{N}\\ldots x_{0}\\}$ , starting with $x_{N}$ as a Gaussian random distribution [22]. In the forward diffusion phase, Gaussian noise $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ is progressively added to the initial, unaltered data $x_0$ , transforming it into a Gaussian random distribution. Conversely, the reverse denoising process", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "transforms Gaussian noise back into a sample. Denoising is parameterized by a learnable noise prediction model, and the learning objective is to learn the noise added to $x_0$ at each diffusion step. After training, the diffusion model generates data akin to $x_0$ by iteratively applying the denoising process, starting from random Gaussian noise.", + "bbox": [ + 496, + 305, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adopting Conditioned Projected Diffusion Model as the Step Model. For our step model, the distribution we aim to fit is the two-action sequence $[a_1, a_T]$ , based on the visual initial and goal states, $v_{start}$ and $v_{goal}$ . These conditional visual states are concatenated with the actions along the action feature dimension, forming a multi-dimensional array:", + "bbox": [ + 496, + 396, + 892, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\hline a _ {1} & 0 & \\dots & 0 & a _ {T} \\end{array} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 500, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the array is zero-padded to have a length corresponds to the planning horizon $T$ . During the denoising process, these conditional visual states can change, potentially misleading the learning process. To prevent this, a condition projection operation [54] is applied, ensuring the visual state and zero-padding dimensions remain unchanged (shaded below). The projection operation is denoted as:", + "bbox": [ + 496, + 539, + 890, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c c c c c} \\hat {v} _ {1} & \\hat {v} _ {2} & \\dots & \\hat {v} _ {T - 1} & \\hat {v} _ {T} \\\\ \\hline \\hat {a} _ {1} & \\hat {a} _ {2} & \\dots & \\hat {a} _ {T - 1} & \\hat {a} _ {T} \\end{array} \\right] \\xrightarrow {\\text {P r o j e c t i o n}} \\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\hat {a} _ {1} & 0 & \\dots & 0 & \\hat {a} _ {T} \\end{array} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 651, + 906, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{v}_t$ denotes the predicted visual state dimensions at timestamp $t$ within the planning horizon $T$ .", + "bbox": [ + 496, + 689, + 890, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Construct the Probabilistic Procedure Knowledge Graph $(\\mathbf{P}^2\\mathbf{KG})$", + "text_level": 1, + "bbox": [ + 498, + 739, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Probabilistic Procedure Knowledge Graph [5] $\\mathrm{P}^2\\mathrm{KG} = (V,E,w)$ is a directed and weighted graph. In this structure, each step from the training set is represented as a node. During the graph construction process, we iterate over the training procedure plans, and for each direct step transition present in a plan, we add an edge from $a_{t}$ to $a_{t + 1}$ if it does not already exist in the graph; otherwise, we increase its existing frequency count by one. Eventually, this process", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18819", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "results in a frequency-based Procedural Knowledge Graph (PKG) [62], which adeptly encapsulating the complexities of step sequencing in procedures and its potential variations, thereby addressing challenges (1) and (2) of procedure planning (ref. § 3.1.2). To further tackle challenge (3), this graph undergoes a transformation into a probabilistic format. In this transformed graph, the edges are not just connections but also signify the likelihoods of transitioning from one step to another. The weight of an edge from $a_{t}$ to $a_{t+1}$ is the count of transitions from action step $a_{t}$ to $a_{t+1}$ normalized by total count of $a_{t}$ being executed [5]. The normalization converts the frequency-based weight into probability distribution and the sum of all out-going edges is one.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 $\\mathbf{P}^2\\mathbf{K}\\mathbf{G}$ -Enhanced Procedure Planning", + "text_level": 1, + "bbox": [ + 76, + 304, + 393, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieving Procedure Plan Recommendations from the $\\mathbf{P}^2\\mathbf{KG}$ . Humans use both previously-acquired knowledge and external knowledge when solving problems. The $\\mathrm{P^2KG}$ provides extensive procedural knowledge, serving as a comprehensive textbook, particularly beneficial for the planning model that requires advanced skills.", + "bbox": [ + 75, + 328, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To utilize this procedural knowledge, queries are made to the $\\mathrm{P}^2\\mathrm{KG}$ using the first $(\\hat{a}_1)$ and last $(\\hat{a}_T)$ actions predicted by the step model. The aim is to find graph paths no longer than $T$ steps, starting from $\\hat{a}_1$ and ending at $\\hat{a}_T$ . This above process often results in multiple possible paths. To evaluate these paths, the probability of each is calculated by multiplying the probability weights of the edges along the path. For instance, the probability of a path $a_1 \\to a_2 \\to a_3$ is determined by the product $w_{a_1 \\to a_2} \\times w_{a_2 \\to a_3}$ . These paths are then ranked according to their probabilities, and the top $R$ paths are selected as the recommended procedure plans from the $\\mathrm{P}^2\\mathrm{KG}$ , where $R$ is predefined. For paths shorter than $T$ , padding is applied at any point in the middle of the sequence to explore all possible resultant paths. When $R$ is greater than one, the top $R$ paths are aggregated through linear weighting into a single path (See section A.2 of supplementary material). This final path is then used as an additional input for the procedure planning model, thereby enhancing its decision-making process.", + "bbox": [ + 75, + 419, + 470, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adopting Conditioned Projected Diffusion Model as the Planning Model. For the planning model, the conditional visual states and the procedure plan recommendation from the $\\mathrm{P}^2\\mathrm{KG}$ are concatenated with the actions along the action feature dimension, forming a multi-dimensional array:", + "bbox": [ + 75, + 705, + 468, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\tilde {a} _ {1} & \\tilde {a} _ {2} & \\dots & \\tilde {a} _ {T - 1} & \\tilde {a} _ {T} \\\\ a _ {1} & a _ {2} & \\dots & a _ {T - 1} & a _ {T} \\end{array} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 786, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rest process is similar to the step model, except that the project operation guarantees that three specific aspects remain unaltered—the dimensions of the visual state, $\\mathrm{P^2KG}$ recommendation, and zero-padding.", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb15eb0616a45f3cf40adc3ffa0a0df4e3e63387ee8938bc3173cb1df2159ca0.jpg", + "image_caption": [ + "Figure 3. Qualitative analysis of the 'Make Jello Shots' task" + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 887, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aabd4cde419dca7cd6424f101e08a01862b15e077b3bc2d8a37c86a0123e34e2.jpg", + "image_caption": [ + "Figure 4. Qualitative analysis of the 'Change a Tire' task" + ], + "image_footnote": [], + "bbox": [ + 500, + 273, + 887, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 465, + 633, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and implementation Details: In our evaluation, we employed datasets from three sources: CrossTask [63], COIN [52], and the Narrated Instructional Videos (NIV) [4]. See section C.4 of supplementary material for details on datasets. All ablation studies and analyses were conducted on CrossTask. We use two Tesla A100 GPUs for all the experiments. We chose horizon $T \\in \\{3, 4, 5, 6\\}$ and $\\mathrm{P^2KG}(R=1)$ condition for implementation. In some cases, we incorporate $\\mathrm{P^2KG}(R=2)$ and LLM conditions which are indicated in the respective tables. Throughout this study, the $\\mathrm{P^2KG}(R=1)$ is employed with a batch size of 256, unless explicitly stated otherwise. More implementation details are available in the section C.2 of the supplementary material.", + "bbox": [ + 496, + 491, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics and baselines: We use mean intersection over union (mIoU), mean accuracy (mAcc), and success rate (SR) as evaluation metrics. SR is the most stringent metric. See sec. C.4 of supplementary for more details. We compare our model with state-of-the-art methods: WLTDO [16], UAAA [1], UPN [49], DDN [9], PlaTe [50], Ext-GAIL [7], $\\mathrm{P}^3\\mathrm{IV}$ [59], PDPP [54], SkipPlan [29], and E3P [53]. More details of these methods are available in the section C.5 of supplementary material. Compared to other models, PDPP uses a different experimental setting. In PDPP, authors set the window after the start time of $a_1$ and before the end time of $a_T$ , contrary to the standard practice of setting a 2-second window around the start and end", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18820", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0823832d4a0e30d693d64fa7de959a2bcb9b85587522595910784a1d14056a8e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsRequired AnnotationsT=3T=4
step classvisual statesstep texttask classSR†mAcc†mIoU†SR†mAcc†mIoU†
Random✓<0.010.941.66<0.010.831.66
Retrieval-Based✓8.0523.332.063.9522.2236.97
WLTDO [16]✓✓1.8721.6431.700.7717.9226.43
UAAA [1]✓✓2.1520.2130.870.9819.8627.09
UPN [49]✓✓2.8924.3931.561.1921.5927.85
DDN [9]✓✓12.1831.2947.485.9727.1048.46
PlaTe [50]✓✓16.0036.1765.9114.0035.2955.36
Ext-GAIL wo Aug. [7]✓✓18.0143.8657.16---
Ext-GAIL [7]✓✓21.2749.4661.7016.4143.0560.93
P3IV♣[59]✓✓23.3449.9673.8913.4044.1670.01
PDPP♣[54]✓✓26.3855.6259.3418.6952.4462.38
E3P♣[53]✓✓✓26.4053.0274.0516.4948.0070.16
SkipPlan [29]♣✓28.8561.1874.9815.5655.6470.30
Ours w/ P2KG (R=2)✓22.6048.7653.5713.9045.7955.00
Ours♣w/ P2KG (R=1)✓33.3461.3664.1420.3855.5464.03
Ours♣w/ P2KG (R=2)✓33.3860.7963.8921.0256.0864.15
PDPP♣†[54]✓✓37.2064.6766.5721.4857.8265.13
Ours♣†w/ P2KG (R=1)✓38.1264.7467.1524.1559.0566.64
", + "bbox": [ + 107, + 88, + 864, + 378 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b74b42370ed4d13becaa78082bc91840f8e1aa032b2a38e77d9ca0d220752c04.jpg", + "table_caption": [ + "Table 1. Performance of our method in comparison to existing baselines for CrossTask dataset. $\\clubsuit$ means that the input visual features are from the S3D network [35] pretrained on HowTo100M [34]; otherwise, precomputed features provided in CrossTask are used. $\\dagger$ indicates the results are under the PDPP's task setting, while others are under the conventional setting" + ], + "table_footnote": [], + "table_body": "
ModelsT=5T=6
DDN [9]3.101.20
P³IV♣[59]7.214.40
PDPP♣[54]13.227.49
E3P♣[53]8.965.76
SkipPlan♣[29]8.555.12
Ours (R=2)8.175.32
Ours♣(R=1)13.258.09
Ours♣(R=2)12.749.23
PDPP♣†[54]13.458.41
Ours♣†(R=1)14.209.27
", + "bbox": [ + 168, + 441, + 380, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Success Rate $(SR^{\\dagger})$ comparison to existing baselines for CrossTask dataset under longer horizons", + "bbox": [ + 75, + 614, + 470, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "time (ref. [9]). We conduct experiments on both PDPP's proposed setting and the conventional setting.", + "bbox": [ + 75, + 650, + 468, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference: During the inference phase, the model receives only the start observation $v_{s}$ and the goal observation $v_{g}$ . To proceed, it utilizes a step model to predict the initial action $a_{1}$ and the end action $a_{T}$ for each data. Subsequently, leveraging the $\\mathrm{P^2KG}$ , highest probable procedure knowledge graph plans connecting $a_{1}$ and $a_{T}$ are obtained. Then, a multi-dimensional array, is created as mentioned in Eq. 5. Finally, the planning model is used to predict the sequence of actions $[a_1,\\dots,a_T]$ by denoising the generated multi-dimensional array as in § 3.2.3.", + "bbox": [ + 75, + 681, + 470, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with the State of the Art (SOTA)", + "text_level": 1, + "bbox": [ + 76, + 845, + 464, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CrossTask (short horizon): We evaluate on CrossTask for short horizons ( $T = 3$ and $T = 4$ ). According to the results", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shown in Table 1, our proposed method outperforms the PDPP in PDPP's setting in every evaluation metric. More than $0.9\\%$ and $2\\%$ improvement in success rate in $T = 3$ and $T = 4$ respectively. In the conventional setting, our method with both $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 1$ ) and $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 2$ ) conditions outperform the success rate values by a significant margin compared to other baselines. $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 2$ ) slightly outperforms $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 1$ ), indicating potential benefits of incorporating more procedural knowledge from the $\\mathrm{P}^2\\mathrm{KG}$ .", + "bbox": [ + 496, + 457, + 890, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CrossTask (long horizon): We use long-horizon predictions for $T = 5$ and $T = 6$ for further evaluating our model as shown in Table 2. In PDPP's setting ( $\\dagger$ ), our method improves the success rate in both $T = 5$ and $T = 6$ . In the conventional setting, our method utilizing $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 1$ ) demonstrates the highest SR value for $T = 5$ , and for a longer horizon at $T = 6$ , our method delivers superior performance for $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 2$ ). Our method performs well under the challenging scenario of a long planning horizon. Our success rate (SR) diminishes from approximately $40\\%$ to $10\\%$ when extending the planning horizon from $T = 3$ to $T = 6$ , primarily due to the heightened uncertainty surrounding the predicted plan between the initial and final steps. This uncertainty stems from the increase in the number of potential procedural plans within the $\\mathrm{P}^2\\mathrm{KG}$ .", + "bbox": [ + 496, + 595, + 892, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NIV and COIN: Results are shown in Table 3 and Table 4. On NIV, ours achieves the best result under the mIoU metric with $T = 3$ , and under both the SR and mIoU metrics with $T = 4$ . The results on NIV ( $T = 5$ , $T = 6$ ) are available in the supplementary material. For the COIN dataset we only", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18821", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "report SR and mAcc due to space constraints; mIoU is reported in the supplementary material. Our method does not rank as the top performer on COIN when $T = 3$ or $T = 4$ . The likely reason is that the COIN dataset features just an average of 3.9 actions per video—a scenario that demands only short-horizon planning and does not necessitate advanced procedural knowledge (which encompasses long sequence-level knowledge [62]). Furthermore, the dataset's extensive collection of over 11k videos provides a substantial resource for baselines to learn basic procedural knowledge.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e293e43bd1632feec9002ebcc58f1a7e5b3f7dee451b07e4616b7a832cfa2a00.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsNIV (T=3)NIV (T=4)
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random2.214.076.091.122.735.84
DDN [9]18.4132.5456.5615.9727.0953.84
Ext-GAIL [7]22.1142.2065.9319.9136.3153.84
P3IV [59]24.6849.0174.2920.1438.3667.29
E3P [53]26.0551.2475.8121.3741.9674.90
PDPP [54]22.2239.5086.6621.3039.2484.96
Ours24.4443.4686.6722.7141.5991.49
", + "bbox": [ + 83, + 253, + 464, + 367 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/549e899db5a95361f90df1d49521e51e75cca7dabbb533787a4599467ea7c749.jpg", + "table_caption": [ + "Table 3. Performance of baselines and ours for NIV dataset" + ], + "table_footnote": [], + "table_body": "
ModelsCOIN (T=3)COIN (T=4)COIN (T=5)
SR↑mAcc↑SR↑mAcc↑SR↑mAcc↑
Random< 0.01< 0.01< 0.01< 0.01--
Retrieval4.3817.402.7114.29--
DDN [9]13.9020.1911.1317.71--
P3IV [59]15.4021.6711.3218.854.2710.81
E3P [53]19.5731.4213.5926.72--
PDPP [54]19.4243.4413.6742.5813.0243.36
SkipPlan [29]23.6547.1216.0443.199.9038.99
Ours (R=2)20.2539.8715.6339.5316.0640.72
", + "bbox": [ + 81, + 414, + 464, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance of baselines and ours for COIN dataset", + "bbox": [ + 89, + 551, + 455, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies and Analyses", + "text_level": 1, + "bbox": [ + 76, + 590, + 346, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the probabilistic procedure knowledge graph. We analyze the role of $\\mathrm{P}^2\\mathrm{KG}$ in improving the performance of our proposed method. Table 5 shows the results which clearly demonstrate that using $\\mathrm{P}^2\\mathrm{KG}$ conditions improves the performance significantly for every $T$ value. Especially when $T = 4$ , success rate (SR) improves more than $3\\%$ and mean IoU improves more than $2\\%$ .", + "bbox": [ + 75, + 613, + 468, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Plan recommendations provided by probabilistic procedure knowledge graph v.s. LLM. We recognize the recent trend of utilizing LLMs to enhance action anticipation [60] or planning in other realms [3, 27, 28, 40, 46]. In Table 6, we compare the results between using $\\mathrm{P}^2\\mathrm{KG}$ v.s. using LLM ('llama-2-13b-chat' and 'llama-2-70b-chat') to generate the plan recommendations. When examining Table 6, it becomes apparent that there are trade-offs between using LLM-generated recommendations and $\\mathrm{P}^2\\mathrm{KG}$ recommendations. For instance, $\\mathrm{P}^2\\mathrm{KG}$ recommendations are constrained by the data available in the training set, limiting their applicability to unseen procedural activities. On the", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "other hand, LLMs tend to exhibit better generalization to such unseen activities. However, considering that the training and testing are conducted on the aforementioned three datasets with known activities, $\\mathrm{P}^2\\mathrm{KG}$ recommendations can yield more accurate results compared to relying on LLM-generated recommendations.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Probabilistic procedure knowledge graph (P²KG) vs. Frequency-based procedure knowledge graph (PKG). The probabilistic procedure knowledge graph uses outedge normalization to encode step transition probabilities (§ 3.2.2), while the frequency-based procedure knowledge graph uses min-max normalization over the frequency counts throughout the graph. In both cases, the planning model only uses one procedure plan recommendation from the graph as condition in our experimental analysis. By looking at the results shown in Table 7, it is evident that the probabilistic procedure knowledge graph outperforms the frequency-based procedure knowledge graph.", + "bbox": [ + 496, + 183, + 892, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of utilizing predicted steps for input conditions to train the procedure planing model. Our proposed problem decomposition allows training the planning model with ground truth (GT) first and last steps. We experiment with two ways to train the planning model. Method 1 uses the predicted start and end steps $(\\hat{a}_1$ and $\\hat{a}_T)$ as input to generate $\\mathrm{P}^2\\mathrm{KG}$ conditions and use them to train the planning model. Method 2 is where we augment the predicted start and end steps using the GT start and end steps $(a_1$ and $a_{T})$ by generating 3 more data samples as follows: $[\\hat{a}_1,a_T]$ , $[a_1,\\hat{a}_T]$ , and $[a_1,a_T]$ . Then we generate $\\mathrm{P}^2\\mathrm{KG}$ conditions for each data and train the model. From the results shown in Table 8, the method without GT data augmentation shows better results. This suggests that leveraging ground truth data in training can lead to worse performance in testing.", + "bbox": [ + 496, + 367, + 892, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Figures 3 and 4 provide qualitative examples of our method. Intermediate steps are padded in the step model because it only predicts the start and end actions. In the 'make jello shot' task (see Figure 3), the model gives a wrong prediction in the intermediate steps when using $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 1$ ) condition. However, it predicts correctly when using $\\mathrm{P}^2\\mathrm{KG}$ ( $R = 2$ ) conditions. In the 'change a tire' task shown in Figure 4, the model is able to predict all the intermediate steps in given conditions.", + "bbox": [ + 496, + 595, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualizations of the probabilistic procedure knowledge graph. We show a sub-graph from our probabilistic procedure knowledge graph (Figure 5). This graph is drawn around the 'jack up' node up to the depth of 2 nodes.", + "bbox": [ + 496, + 732, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualizations of the expert trajectories. Figure 6 illustrates the steps involved in completing the 'make jello shots' task, along with their transitions to other steps within the entire training data. This figure demonstrates that our $\\mathrm{P}^2\\mathrm{KG}$ encodes diverse sequencing possibilities for steps and also captures task-sharing steps across the entire training domain. For instance, 'pour water' is a step in 'make jello", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18822", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/55bfa00a0cf0b51efbf66098b2e7b344f5e014bf0cd3adf9ff4cb1da706c406f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelT=3T=4T=5T=6
SRmAccmIoUSRmAccmIoUSRmAccmIoUSRmAccmIoU
w.o P2KG conditions †35.6963.9166.0420.5257.4764.3912.853.4464.018.1550.4564.13
Ours †38.1264.7467.1524.1559.0566.6414.2053.8465.569.2750.2265.97
w.o P2KG conditions31.3559.5163.1118.9256.2062.4712.7151.2963.568.1647.6363.39
Ours33.3860.7963.8921.0256.0864.1512.7451.2363.169.2350.7865.56
", + "bbox": [ + 102, + 75, + 867, + 154 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8d16bbed8315b4fa877827c4266d0caca5f0be784574aea029dc6025ac39b110.jpg", + "table_caption": [ + "Table 5. Performance of our method with and without ${\\mathrm{P}}^{2}\\mathrm{{KG}}$ conditions on CrossTask ${}^{ \\star }$ dataset" + ], + "table_footnote": [], + "table_body": "
Model (T=6, CrossTask*)SRmAccmIoU
Ours with P2KG (R=1)
PDPP setting9.2750.2265.97
Conventional setting8.0950.8065.39
One LLM plan recommendation
PDPP setting (13b)7.7450.2864.05
Conventional setting (13b)7.2149.6863.89
PDPP setting (70b)8.6250.3164.34
Conventional setting (70b)7.8149.7564.02
P2KG (R=1) and one LLM plan recommendation
PDPP setting (13b)8.8149.9765.22
Conventional setting (13b)8.2051.4664.30
PDPP setting (70b)9.0150.2565.57
Conventional setting (70b)8.3451.5364.96
", + "bbox": [ + 84, + 202, + 460, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f41647624f66a9fc3df3a29ae7641aabe43666dab546135ebf359b89e1c6515a.jpg", + "table_caption": [ + "Table 6. Performance of the plan recommendations provided by the probabilistic procedure knowledge graph v.s. LLM." + ], + "table_footnote": [], + "table_body": "
ModelsSRmAccmIoU
Frequency graph7.6648.6164.21
Probabilistic graph8.0950.8065.40
", + "bbox": [ + 147, + 436, + 400, + 478 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78f9c98600e713d88bb71d24c8f65256b22c7216c5326f30f19b451dab575373.jpg", + "table_caption": [ + "Table 7. Performance comparison between probabilistic procedure knowledge graph v.s. frequency-based procedure knowledge graph for $T = 6$ on CrossTask dataset" + ], + "table_footnote": [], + "table_body": "
ConditionSRmAccmIoU
without GT data aug.38.1264.7467.15
with GT data aug.32.4562.4262.80
", + "bbox": [ + 130, + 549, + 415, + 590 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e73d312471b4b9b4f798ae9628bfbd7e529230f163e645a9fce359281c1156c2.jpg", + "table_caption": [ + "Table 8. Effect of different input conditions for performance on CrossTask ${}^{ \\star }$ dataset $\\left( {T = 3}\\right)$ in PDPP’s setting" + ], + "table_footnote": [], + "table_body": "
ModelsT=3T=4T=5T=6
ˆa1ˆaTˆa1ˆaTˆa1ˆaTˆa1ˆaT
Ours53.6950.6055.5652.5155.5851.8157.0951.92
Ours♣71.4263.3272.9863.3772.4263.2963.8259.96
", + "bbox": [ + 96, + 648, + 450, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. The step model's start and end step prediction accuracies on the CrossTask dataset", + "bbox": [ + 76, + 713, + 467, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shots' task, but it can also be part of other tasks, leading to a step transition from 'pour water' to 'add fish.' This structure allows models to leverage rich procedural knowledge.", + "bbox": [ + 76, + 758, + 467, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results for the step model. Table 9 reveals step model results, indicating potential enhancement areas to elevate the planning performance.", + "bbox": [ + 76, + 806, + 467, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Failure cases. Our model exhibits three distinct failure case patterns. See section B.5 of the supplementary material for detailed discussions.", + "bbox": [ + 76, + 854, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db2da738b40c0a458f2503dbbf6960469d64d8ce24cb4392971c6f6a7d22ce5e.jpg", + "image_caption": [ + "Figure 5. Example of a sub-graph in our probabilistic procedure knowledge graph $(\\mathrm{P}^2\\mathrm{KG})$ for CrossTask dataset. This graph effectively encapsulates real-world knowledge of distinct transition probabilities between steps, e.g., the probability of transitioning from 'start loose' to 'jack up' is 0.65, in contrast to a mere 0.14 for the reverse transition—the $\\mathrm{P}^2\\mathrm{KG}$ reflects the common real-life practice where loosening the lug nuts before jacking up the car leads to a safer and more efficient tire change." + ], + "image_footnote": [], + "bbox": [ + 504, + 186, + 885, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fbfd42d59b3621b2bfa32ab374011a1653151f69d73f47d634e03eeb447e2bfb.jpg", + "image_caption": [ + "Figure 6. Expert trajectories of the 'Make Jello Shots' task, involving task-sharing steps and thus out-of-task step transitions. Thicker lines indicate paths that are more frequently visited" + ], + "image_footnote": [], + "bbox": [ + 504, + 486, + 887, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 724, + 617, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We focus on formulating procedural plans from an AI agent in instructional videos. We propose KEPP which employs a probabilistic procedural knowledge graph, sourced from the training domain, effectively serving as a 'textbook' for procedure planning. Results show that KEPP delivers SOTA performance with minimal supervision. Future work can focus on enhancing the accuracy of predictions for the initial and final steps. Additionally, our approach can be modified to aid in detecting erroneous steps and the misordering of steps in instructional videos [38, 41].", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yazan Abu Farha and Juergen Gall. Uncertainty-aware anticipation of activities. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5, 6", + "[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2", + "[3] Anurag Ajay, Seungwook Han, Yilun Du, Shaung Li, Abhi Gupta, Tommi Jaakkola, Josh Tenenbaum, Leslie Kaelbling, Akash Srivastava, and Pulkit Agrawal. Compositional foundation models for hierarchical planning. arXiv preprint arXiv:2309.08587, 2023. 7", + "[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4575-4583, 2016. 5", + "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. arXiv preprint arXiv:2307.08763, 2023. 2, 4, 5", + "[6] Anonymous authors. Active procedure planning with uncertainty-awareness in instructional videos, 2023. Under review as a conference paper at ICLR 2024. https://openreview.net/pdf?id=JDd46WodYf.2", + "[7] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 1, 2, 5, 6, 7", + "[8] Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin, Jena D Hwang, Xiang Lorraine Li, Hirona J Arai, Soumya Sanyal, Keisuke Sakaguchi, Xiang Ren, and Yejin Choi. Plasma: Making small language models better procedural knowledge models for (counterfactual) planning. arXiv preprint arXiv:2305.19472, 2023. 2", + "[9] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 1, 2, 3, 5, 6, 7", + "[10] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where?-self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 2", + "[11] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2447, 2023. 2", + "[12] Hazel Doughty, Ivan Laptev, Walterio Mayol-Cuevas, and Dima Damen. Action modifiers: Learning from adverbs in" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 868-878, 2020. 2", + "[13] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1, 2", + "[14] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 2", + "[15] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18952-18961, 2023. 2", + "[16] Kiana Ehsani, Hessam Bagherinezhad, Joseph Redmon, Roozbeh Mottaghi, and Ali Farhadi. Who let the dogs out? modeling dog behavior from visual data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4051-4060, 2018. 5, 6", + "[17] Sophie Fischer, Carlos Gemmell, Iain Mackie, and Jeffrey Dalton. Vilt: Video instructions linking for complex tasks. In Proceedings of the 2nd International Workshop on Interactive Multimedia Retrieval, pages 41-47, 2022. 2", + "[18] Kevin Flanagan, Dima Damen, and Michael Wray. Learning temporal sentence grounding from narrated egovideos. arXiv preprint arXiv:2310.17395, 2023. 2", + "[19] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. arXiv preprint arXiv:2005.03684, 2020. 2", + "[20] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threedorld transport challenge: A visually guided task-and-motion planning benchmark for physically realistic embodied ai. arXiv preprint arXiv:2103.14025, 2021. 2", + "[21] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 2", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4", + "[23] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. Machine Intelligence Research, 20 (2):249-262, 2023. 2", + "[24] De-An Huang, Joseph J Lim, Li Fei-Fei, and Juan Carlos Niebles. Unsupervised visual-linguistic reference resolution in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2183–2192, 2017. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18824", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] De-An Huang, Shyamal Buch, Lucio Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding\" it\": Weakly-supervised reference-aware visual grounding in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5948-5957, 2018. 2", + "[26] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2", + "[27] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2, 7", + "[28] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 2, 7", + "[29] Zhiheng Li, Wenjia Geng, Muheng Li, Lei Chen, Yansong Tang, Jiwen Lu, and Jie Zhou. Skip-plan: Procedure planning in instructional videos via condensed action space learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10297-10306, 2023. 2, 3, 5, 6, 7", + "[30] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[31] Yujie Lu, Pan Lu, Zhiyu Chen, Wanrong Zhu, Xin Eric Wang, and William Yang Wang. Multimodal procedural planning via dual text-image prompting. arXiv preprint arXiv:2305.01795, 2023. 2", + "[32] Weichao Mao, Ruta Desai, Michael Louis Iuzzolino, and Nitin Kamra. Action dynamics task graphs for learning plannable representations of procedural tasks. arXiv preprint arXiv:2302.05330, 2023. 2", + "[33] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. arXiv preprint arXiv:2306.03802, 2023. 2", + "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 1, 6", + "[35] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2, 6", + "[36] Davide Moltisanti, Frank Keller, Hakan Bilen, and Laura Sevilla-Lara. Learning action changes by measuring verb-adverb textual relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23110-23118, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2", + "[38] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 8", + "[39] Megha Nawhal, Akash Abdu Jyothi, and Greg Mori. Rethinking learning approaches for long-term action anticipation. In European Conference on Computer Vision, pages 558-576. Springer, 2022. 2", + "[40] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15302-15314, 2023. 7", + "[41] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 8", + "[42] Fadime Sener, Rishabh Saraf, and Angela Yao. Transferring knowledge from text to video: Zero-shot anticipation for procedural actions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[43] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2", + "[44] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 2", + "[45] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2", + "[46] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 2, 7", + "[47] Yale Song, Gene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2", + "[48] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Look for the change: Learning object states and state-modifying actions from untrimmed" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18825", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "web videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13956-13966, 2022. 2", + "[49] Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal planning networks: Learning generalizable representations for visuomotor control. In International Conference on Machine Learning, pages 4732-4741. PMLR, 2018. 5, 6", + "[50] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters, 7(2):4924-4930, 2022. 1, 2, 5, 6", + "[51] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34:14476-14487, 2021. 2", + "[52] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 5", + "[53] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13565–13575, 2023. 1, 2, 3, 5, 6, 7", + "[54] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14836-14845, 2023. 1, 2, 3, 4, 5, 6, 7", + "[55] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 2", + "[56] Yue Yang, Joongwon Kim, Artemis Panagopoulou, Mark Yatskar, and Chris Callison-Burch. Induce, edit, retrieve: Language grounded multimodal schema for instructional video retrieval. arXiv preprint arXiv:2111.09276, 2021. 2", + "[57] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 2", + "[58] Jiahao Zhang, Anoop Cherian, Yanbin Liu, Yizhak Ben-Shabat, Cristian Rodriguez, and Stephen Gould. Aligning step-by-step instructional diagrams to video demonstrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2483-2492, 2023. 2", + "[59] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 1, 2, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368, 2023. 7", + "[61] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 2", + "[62] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2, 5, 7", + "[63] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 5" + ], + "bbox": [ + 501, + 92, + 892, + 371 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18826", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_model.json b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..00d45e3f9721b31723450d435149612a1d0415dd --- /dev/null +++ b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "Why Not Use Your Textbook? Knowledge-Enhanced Procedure Planning of Instructional Videos" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.203, + 0.851, + 0.223 + ], + "angle": 0, + "content": "Kumaranage Ravindu Yasas Nagasinghe\\(^{1}\\) Honglu Zhou\\(^{2}\\) Malitha Gunawardhana\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.228, + 0.785, + 0.247 + ], + "angle": 0, + "content": "Martin Renqiang Min\\(^{2}\\) Daniel Harari\\(^{4}\\) Muhammad Haris Khan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.248, + 0.737, + 0.264 + ], + "angle": 0, + "content": "1Mohamed bin Zayed University of Artificial Intelligence, 2NEC Laboratories, USA," + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.266, + 0.657, + 0.281 + ], + "angle": 0, + "content": "3 University of Auckland, 4Weizmann Institute of Science" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.285, + 0.744, + 0.299 + ], + "angle": 0, + "content": "ravindu.nagasinghe@mbzuai.ac.ae, muhammad.haris@mbzuai.ac.ae" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.334, + 0.314, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.473, + 0.807 + ], + "angle": 0, + "content": "In this paper, we explore the capability of an agent to construct a logical sequence of action steps, thereby assembling a strategic procedural plan. This plan is crucial for navigating from an initial visual observation to a target visual outcome, as depicted in real-life instructional videos. Existing works have attained partial success by extensively leveraging various sources of information available in the datasets, such as heavy intermediate visual observations, procedural names, or natural language step-by-step instructions, for features or supervision signals. However, the task remains formidable due to the implicit causal constraints in the sequencing of steps and the variability inherent in multiple feasible plans. To tackle these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities by infusing it with procedural knowledge. This knowledge, sourced from training procedure plans and structured as a directed weighted graph, equips the agent to better navigate the complexities of step sequencing and its potential variations. We coin our approach KEPP, a novel Knowledge-Enhanced Procedure Planning system, which harnesses a probabilistic procedural knowledge graph extracted from training data, effectively acting as a comprehensive textbook for the training domain. Experimental evaluations across three widely-used datasets under settings of varying complexity reveal that KEPP attains superior, state-of-the-art results while requiring only minimal supervision. Code and trained model are available at https://github.com/Ravindu-Yasas-Nagasinghe/KEPP" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.828, + 0.21, + 0.844 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The evolution of the internet has precipitated an unprecedented influx of video content, serving as a vital educational resource for myriad learners. Individuals frequently" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.336, + 0.891, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.658 + ], + "angle": 0, + "content": "Figure 1. Expert trajectories [7] of the 'Make Jello Shots' task from the CrossTask dataset [63]. Heavier color indicates more frequently visited path. This depicts the complexities of the procedure planning task, arising from the subtle causal links in step sequencing (e.g., steps like 'stir mixture' or 'pour mixture' typically occur after adding individual ingredients), the varied probabilities of transitioning between steps, and the diversity in plans viable for a given starting point and an intended outcome. Motivated by these nuanced challenges, we propose Knowledge-Enhanced Procedure Planning (KEPP) with the use of a probabilistic procedure knowledge graph to capture and represent these intricacies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.794 + ], + "angle": 0, + "content": "leverage platforms such as YouTube to acquire new skills, ranging from culinary arts to automobile maintenance [34]. While these instructional videos benefit the development of intelligent agents in mastering long-horizon tasks, the challenge extends beyond merely interpreting visuals. It requires high-level reasoning and planning to effectively assist in complex, real-life scenarios [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Procedure Planning in Instructional Videos demands an agent to produce a sequence of actionable steps, thereby crafting a procedure plan that facilitates the transition from an initial visual observation of the physical world towards achieving a desired goal state [7, 9, 50, 53, 54, 59]. The task acts as a precursor to an envisioned future scenario in which an agent like a robot provides on-the-spot support, such as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18816" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.395, + 0.107 + ], + "angle": 0, + "content": "assisting an individual in preparing a recipe [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Current methods in procedure planning in instructional videos make extensive use of various annotations available within the datasets to enrich input features or provide supervisory signals (see Table 1). These include detailed, temporally localized visual observations of intermediate action steps throughout the procedure plan [7, 9, 50], high-level procedural task labels [53, 54], and step-by-step instructions in natural language [53, 59]. Despite advancements, significant challenges persist, including characterizing the implicit causal constraints in step sequencing, the varied probabilities of transitioning between steps, and the inherent variability of multiple viable plans (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.442 + ], + "angle": 0, + "content": "To address these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities in procedure planning by infusing it with comprehensive procedural knowledge [62], derived from training procedure plans and structured as a directed weighted graph. This graph, as a Probabilistic Procedural Knowledge Graph [5] where nodes denote steps from diverse tasks and edges represent step transition probabilities in the training domain, empowers agents to more adeptly navigate the complexities of step sequencing and its potential variations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Our proposed approach, KEPP, is a novel Knowledge-Enhanced Procedure Planning system (Fig. 2) that harnesses a probabilistic procedural knowledge graph \\((\\mathrm{P}^2\\mathrm{KG})\\), constructed from training procedure plans. This graph functions like a detailed textbook, providing extensive knowledge for the training domain, and thereby circumventing the need for costly multiple annotations required by existing methods. Additionally, we decompose the instructional video procedure planning problem into two parts: one driven by objectives specific to step perception and the other by a procedural knowledge-informed modeling of procedure planning. In this problem decomposition, the first and last action steps are predicted based on the initial and goal visual states. Following this, a procedure plan is crafted by leveraging the procedure plan recommendations retrieved from the \\(\\mathrm{P}^2\\mathrm{KG}\\). The recommendations correspond to the most probable procedure plans frequently used in training, conditioned on the predicted first and last action steps. In a similar vein to the approach by Li et al. [29], our proposed decomposition strategy reduces uncertainty by maximizing the use of currently available information, namely the initial and goal visual states. This allows for the improvement of procedure planning through more accurate predictions of start and end actions. Plus, this decomposition effectively incorporates procedural knowledge into procedure planning, thereby enhancing its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.839, + 0.316, + 0.853 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos that leverages rich procedural knowledge from a probabilistic procedu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ral knowledge graph (P2KG). This approach necessitates only a minimal amount of annotations for supervision." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.228 + ], + "angle": 0, + "content": "- We decompose the problem in procedure planning of instructional videos: predicting the initial and final steps from the start and end visuals, and then creating a plan using procedural knowledge retrieved based on these predicted steps. This approach prioritizes the currently available information and effectively incorporates procedural knowledge, enhancing strategic planning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.894, + 0.273 + ], + "angle": 0, + "content": "- Experimental evaluations on three widely-used datasets, under settings of varying complexity, reveal that KEPP attains state-of-the-art results in procedure planning." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.287, + 0.642, + 0.303 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.538 + ], + "angle": 0, + "content": "Instructional Videos, which demonstrate multi-step procedures, have become a hotbed of research. The studies delve into various aspects, including comprehending and extracting intricate spatiotemporal content from video [12, 18, 19, 21, 23, 36, 43, 44, 48, 55, 57], interpreting the interrelationships between various actions and procedural events [47, 63], and developing capabilities for forecasting [39, 42] and strategic reasoning and planning [28] within the context of these videos. Furthermore, by leveraging the multimodality of visual, auditory, and narrative elements within these videos, research extends to areas like multimodal alignment [2, 58], grounding [10, 14, 25, 33, 51], representation learning [11, 35, 61], pre-training [15, 26, 62], and more [17, 24, 37, 56]. This paper focuses on procedure planning in instructional videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Procedure Planning is a vital skill for autonomous agents tasked with handling complex activities in everyday settings. Essentially, these agents must discern the appropriate actions to reach a specific goal. This aspect of artificial intelligence (AI) has been a prominent and integral subject in fields like robotics [20, 28, 32, 45, 46]. Yet, the challenge of procedure planning in the context of instructional videos is notably distinct, and potentially more complex, than its counterparts in natural language processing [8, 30], multimodal generative AI [13, 31], and simulated environments [27, 28, 45]. Its significance is underscored by the need for planning that is grounded in real-world scenes. This requires the development of AI agents capable of accurately perceiving and understanding the current real-world context, and then anticipating and mapping out a logical sequence of actions to fulfill a high-level goal effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Procedure Planning in Instructional Videos has recently garnered research attention. DDN [9] initiates this trend by conceptualizing the problem as sequential latent space planning. Building on this, PlaTe [50] employs transformers for both action and state models, integrating Beam Search for enhanced performance. Meanwhile, Ext-GAIL [7] suggests employing contextual modeling through variational autoencoder and adversarial policy learning. This method" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18817" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "considers contextual information as time-invariant knowledge, crucial for distinguishing specific tasks and allowing for multiple planning outcomes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.471, + 0.424 + ], + "angle": 0, + "content": "While these earlier approaches have viewed procedure planning as an autoregressive sequence generation problem, recent methods regard it as a distribution-fitting problem to mitigate error propagation in sequential decisions. In this vein, \\(\\mathrm{P}^3\\mathrm{IV}\\) [59] replaces intermediate visual states with linguistic representations for supervision, predicting all steps simultaneously instead of using autoregressive methods. To circumvent the complex learning strategies and high annotation costs of previous work, PDPP [54] models the entire intermediate action sequence distribution using a conditioned projected diffusion model. This approach redefines the planning problem as a sampling process from this distribution and simplifies supervision by using only instructional video task labels. E3P [53], also encoding task information, adopts a mask-and-predict strategy for mining step relationships in procedural tasks, integrating probabilistic masking for regularization. In contrast, our approach does not rely on annotations of intermediate states, natural language step representations, or procedural task labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.591 + ], + "angle": 0, + "content": "Recognizing the difficulties inherent in high dimensional state supervision and the accumulation of errors in action sequences, SkipPlan [29] was developed. It strategically focuses on action predictions, breaking down longer sequences into shorter, more manageable sub-chains by skipping over less reliable intermediate actions. Drawing inspiration from SkipPlan, our approach decomposes the procedure planning problem to prioritize the most reliable information available (ref. § 3.1.2). However, we innovate further by incorporating a Probabilistic Procedure Knowledge Graph, significantly enriching the planning phase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.212, + 0.623 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.469, + 0.677 + ], + "angle": 0, + "content": "We will first introduce the problem setup in § 3.1, and then discuss our novel Knowledge-Enhanced Procedure Planning system (KEPP) in § 3.2. See Fig. 2 for KEPP overview." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.687, + 0.355, + 0.701 + ], + "angle": 0, + "content": "3.1. Problem and Method Overview" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.283, + 0.724 + ], + "angle": 0, + "content": "3.1.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We follow the problem definition for procedure planning of instructional videos put forth by Chang et al. [9]: given an observation of the initial state \\( v_{start} \\) and a goal state \\( v_{goal} \\), both are short video clips indicating different states of the real-world environment extracted from an instructional video, a model is required to plan a sequence of action steps \\( a_{1:T} \\) to reach the indicated goal. Here, \\( T \\) is the planning horizon, inputting to the model, corresponding to the number of action steps in the sequence produced by the model so that the environment state can be transformed from \\( v_{start} \\) to \\( v_{goal} \\). We use \\( a_t \\) to denote the action step at the timestamp" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.169 + ], + "angle": 0, + "content": "\\(t\\), and in the following, \\(v_{s}\\) and \\(v_{g}\\) are short for \\(v_{start}\\) and \\(v_{goal}\\). Mathematically, the procedure planning problem is defined as \\(p(a_{1:T}|v_s,v_g)\\) that denotes the conditional probability distribution of the action sequence \\(a_{1:T}\\) given the initial visual observation \\(v_{start}\\) and the goal visual state \\(v_{goal}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.191, + 0.722, + 0.207 + ], + "angle": 0, + "content": "3.1.2 Problem Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Considering the initial and final visual states are input, providing the most reliable information, we hypothesize that predicting the first and final action steps is more dependable than interpolating the intermediate ones, and consequently, an enhanced accuracy in predicting the first and final steps can lead to more effective procedure planning. Inspired by this hypothesis, we decompose the procedure planning problem into two sub-problems, as shown in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.36, + 0.892, + 0.377 + ], + "angle": 0, + "content": "\\[\np \\left(\\hat {a} _ {1: T} | v _ {s}, v _ {g}\\right) = p \\left(\\hat {a} _ {2: T - 1} | \\hat {a} _ {1}, \\hat {a} _ {T}\\right) p \\left(\\hat {a} _ {1}, \\hat {a} _ {T} | v _ {s}, v _ {g}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.456 + ], + "angle": 0, + "content": "where the first sub-problem is to identify the beginning step \\( a_1 \\) and the end step \\( a_T \\), and the second sub-problem is to plan the intermediate action steps \\( a_{2:T-1} \\) given \\( a_1 \\) and \\( a_T \\). We use \\( \\hat{a}_t \\) to denote predicted action step at timestamp \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Our proposed problem decomposition in Eq. 1 bears resemblance with the problem formulation from Li et al. [29]; they decompose procedure planning into \\( p(\\hat{a}_{1:T}|v_s, v_g) = \\prod_{t=2}^{T-1} p(\\hat{a}_t|\\hat{a}_1, \\hat{a}_T)p(\\hat{a}_1, \\hat{a}_T|v_s, v_g) \\). However, our formulation differs in its approach to modeling the second subproblem. Specifically, we employ a conditioned projected diffusion model (ref. § 3.2) to jointly predict \\( a_{2:T-1} \\) at once, whereas Li et al. [29] rely on Transformer decoders to predict each intermediate action independently. Further, we integrate a Probabilistic Procedure Knowledge Graph (ref. § 3.2.2) to address the second sub-problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Tackling the second sub-problem is nontrivial even when armed with an oracle predictor for the first sub-problem. Procedure planning in real-life scenarios remains daunting because of the following challenges: (1) the presence of implicit temporal and causal constraints in the sequencing of steps, (2) the existence of numerous viable plans given an initial state and a goal state, and (3) the need to incorporate the real-life everyday knowledge both in task-sharing steps and in managing the inherent variability in transition probabilities between steps. Previous studies tackled these challenges by extensively harnessing detailed annotations in the datasets to augment input features or offer supervision signals (see Table 1). In contrast, we propose harnessing a Probabilistic Procedural Knowledge Graph \\((\\mathrm{P}^2\\mathrm{KG})\\) which is extracted from the procedure plans in the training set. With the \\(\\mathrm{P}^2\\mathrm{KG}\\) at our hand, we further decompose the procedure planning problem to reduce its complexity and learn \\(f_{\\theta}:(v_s,v_g,T)\\to p(\\hat{a}_{1:T}|v_s,v_g)\\) as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.08, + 0.891, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.894, + 0.281 + ], + "angle": 0, + "content": "Figure 2. Overview of our methodology. We introduce KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos, leveraging a Probabilistic Procedural Knowledge Graph \\((\\mathrm{P}^2\\mathrm{KG})\\). KEPP breaks down procedure planning into two parts: predicting initial and final steps from visual states, and crafting a procedure plan based on the procedural knowledge retrieved from \\(\\mathrm{P}^2\\mathrm{KG}\\), conditioned on the predicted first and last action steps. KEPP requires minimal annotations and enhances planning effectiveness" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.33, + 0.469, + 0.374 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(\\hat {a} _ {1: T} | v _ {s}, v _ {g}\\right) = \\\\ p \\left(\\hat {a} _ {1: T} | \\tilde {a} _ {1: T}, v _ {s}, v _ {g}\\right) p \\left(\\tilde {a} _ {1: T} | \\hat {a} _ {1}, \\hat {a} _ {T}\\right) p \\left(\\hat {a} _ {1}, \\hat {a} _ {T} | v _ {s}, v _ {g}\\right) \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.47, + 0.541 + ], + "angle": 0, + "content": "where \\( f_{\\theta} \\) denotes the machine learning model, and \\( \\tilde{a}_{1:T} \\) represents a graph path (i.e., a sequence of nodes) retrieved from \\( \\mathrm{P}^2\\mathrm{KG} \\). This graph path provides a valuable procedure plan recommendation aligned with the training domain, thus mitigating the complexity of procedure planning. It is worth noting that the proposed approach to modeling procedure planning using Eq. 2 demands only a minimal level of supervision, merely relying on the ground truth training procedure plan; Eq. 2 circumvents the need for additional annotations. We describe details of our \\( \\mathrm{P}^2\\mathrm{KG} \\)-enhanced approach in the following subsection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.47, + 0.581 + ], + "angle": 0, + "content": "3.2. KEPP: Knowledge-Enhanced Procedure Planning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.693 + ], + "angle": 0, + "content": "We propose KEPP (Fig. 2) utilizing a probabilistic procedure knowledge graph extracted from the training set. We firstly identify the beginning and conclusion steps according to the input initial and goal states; and then, conditioned on these steps and the planning horizon \\( T \\), we query the graph to retrieve relevant procedural knowledge for knowledge-enhanced procedure planning of instructional videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.416, + 0.727 + ], + "angle": 0, + "content": "3.2.1 Identify Beginning and Conclusion Steps" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Given \\( v_{start} \\) and \\( v_{goal} \\) as input, we adapt a Conditioned Projected Diffusion Model [54] (ref. supplementary material) to identify the first action step and the final step; we refer to this model as the 'Step (Perception) Model'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Standard Denoising Diffusion Probabilistic Model tackles data generation through a denoising Markov chain over variables \\(\\{x_{N}\\ldots x_{0}\\}\\), starting with \\(x_{N}\\) as a Gaussian random distribution [22]. In the forward diffusion phase, Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) is progressively added to the initial, unaltered data \\(x_0\\), transforming it into a Gaussian random distribution. Conversely, the reverse denoising process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.396 + ], + "angle": 0, + "content": "transforms Gaussian noise back into a sample. Denoising is parameterized by a learnable noise prediction model, and the learning objective is to learn the noise added to \\( x_0 \\) at each diffusion step. After training, the diffusion model generates data akin to \\( x_0 \\) by iteratively applying the denoising process, starting from random Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.893, + 0.489 + ], + "angle": 0, + "content": "Adopting Conditioned Projected Diffusion Model as the Step Model. For our step model, the distribution we aim to fit is the two-action sequence \\([a_1, a_T]\\), based on the visual initial and goal states, \\(v_{start}\\) and \\(v_{goal}\\). These conditional visual states are concatenated with the actions along the action feature dimension, forming a multi-dimensional array:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.5, + 0.892, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\hline a _ {1} & 0 & \\dots & 0 & a _ {T} \\end{array} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.646 + ], + "angle": 0, + "content": "where the array is zero-padded to have a length corresponds to the planning horizon \\( T \\). During the denoising process, these conditional visual states can change, potentially misleading the learning process. To prevent this, a condition projection operation [54] is applied, ensuring the visual state and zero-padding dimensions remain unchanged (shaded below). The projection operation is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.652, + 0.908, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c c c c c} \\hat {v} _ {1} & \\hat {v} _ {2} & \\dots & \\hat {v} _ {T - 1} & \\hat {v} _ {T} \\\\ \\hline \\hat {a} _ {1} & \\hat {a} _ {2} & \\dots & \\hat {a} _ {T - 1} & \\hat {a} _ {T} \\end{array} \\right] \\xrightarrow {\\text {P r o j e c t i o n}} \\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\hat {a} _ {1} & 0 & \\dots & 0 & \\hat {a} _ {T} \\end{array} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.72 + ], + "angle": 0, + "content": "where \\(\\hat{v}_t\\) denotes the predicted visual state dimensions at timestamp \\(t\\) within the planning horizon \\(T\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.892, + 0.772 + ], + "angle": 0, + "content": "3.2.2 Construct the Probabilistic Procedure Knowledge Graph \\((\\mathbf{P}^2\\mathbf{KG})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The Probabilistic Procedure Knowledge Graph [5] \\(\\mathrm{P}^2\\mathrm{KG} = (V,E,w)\\) is a directed and weighted graph. In this structure, each step from the training set is represented as a node. During the graph construction process, we iterate over the training procedure plans, and for each direct step transition present in a plan, we add an edge from \\(a_{t}\\) to \\(a_{t + 1}\\) if it does not already exist in the graph; otherwise, we increase its existing frequency count by one. Eventually, this process" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18819" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "results in a frequency-based Procedural Knowledge Graph (PKG) [62], which adeptly encapsulating the complexities of step sequencing in procedures and its potential variations, thereby addressing challenges (1) and (2) of procedure planning (ref. § 3.1.2). To further tackle challenge (3), this graph undergoes a transformation into a probabilistic format. In this transformed graph, the edges are not just connections but also signify the likelihoods of transitioning from one step to another. The weight of an edge from \\(a_{t}\\) to \\(a_{t+1}\\) is the count of transitions from action step \\(a_{t}\\) to \\(a_{t+1}\\) normalized by total count of \\(a_{t}\\) being executed [5]. The normalization converts the frequency-based weight into probability distribution and the sum of all out-going edges is one." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.305, + 0.395, + 0.323 + ], + "angle": 0, + "content": "3.2.3 \\(\\mathbf{P}^2\\mathbf{K}\\mathbf{G}\\) -Enhanced Procedure Planning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.419 + ], + "angle": 0, + "content": "Retrieving Procedure Plan Recommendations from the \\(\\mathbf{P}^2\\mathbf{KG}\\). Humans use both previously-acquired knowledge and external knowledge when solving problems. The \\(\\mathrm{P^2KG}\\) provides extensive procedural knowledge, serving as a comprehensive textbook, particularly beneficial for the planning model that requires advanced skills." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.471, + 0.706 + ], + "angle": 0, + "content": "To utilize this procedural knowledge, queries are made to the \\(\\mathrm{P}^2\\mathrm{KG}\\) using the first \\((\\hat{a}_1)\\) and last \\((\\hat{a}_T)\\) actions predicted by the step model. The aim is to find graph paths no longer than \\(T\\) steps, starting from \\(\\hat{a}_1\\) and ending at \\(\\hat{a}_T\\). This above process often results in multiple possible paths. To evaluate these paths, the probability of each is calculated by multiplying the probability weights of the edges along the path. For instance, the probability of a path \\(a_1 \\to a_2 \\to a_3\\) is determined by the product \\(w_{a_1 \\to a_2} \\times w_{a_2 \\to a_3}\\). These paths are then ranked according to their probabilities, and the top \\(R\\) paths are selected as the recommended procedure plans from the \\(\\mathrm{P}^2\\mathrm{KG}\\), where \\(R\\) is predefined. For paths shorter than \\(T\\), padding is applied at any point in the middle of the sequence to explore all possible resultant paths. When \\(R\\) is greater than one, the top \\(R\\) paths are aggregated through linear weighting into a single path (See section A.2 of supplementary material). This final path is then used as an additional input for the procedure planning model, thereby enhancing its decision-making process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.782 + ], + "angle": 0, + "content": "Adopting Conditioned Projected Diffusion Model as the Planning Model. For the planning model, the conditional visual states and the procedure plan recommendation from the \\(\\mathrm{P}^2\\mathrm{KG}\\) are concatenated with the actions along the action feature dimension, forming a multi-dimensional array:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.787, + 0.47, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\tilde {a} _ {1} & \\tilde {a} _ {2} & \\dots & \\tilde {a} _ {T - 1} & \\tilde {a} _ {T} \\\\ a _ {1} & a _ {2} & \\dots & a _ {T - 1} & a _ {T} \\end{array} \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The rest process is similar to the step model, except that the project operation guarantees that three specific aspects remain unaltered—the dimensions of the visual state, \\(\\mathrm{P^2KG}\\) recommendation, and zero-padding." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.092, + 0.888, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.255, + 0.875, + 0.27 + ], + "angle": 0, + "content": "Figure 3. Qualitative analysis of the 'Make Jello Shots' task" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.275, + 0.888, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.435, + 0.866, + 0.449 + ], + "angle": 0, + "content": "Figure 4. Qualitative analysis of the 'Change a Tire' task" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.634, + 0.483 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Datasets and implementation Details: In our evaluation, we employed datasets from three sources: CrossTask [63], COIN [52], and the Narrated Instructional Videos (NIV) [4]. See section C.4 of supplementary material for details on datasets. All ablation studies and analyses were conducted on CrossTask. We use two Tesla A100 GPUs for all the experiments. We chose horizon \\( T \\in \\{3, 4, 5, 6\\} \\) and \\( \\mathrm{P^2KG}(R=1) \\) condition for implementation. In some cases, we incorporate \\( \\mathrm{P^2KG}(R=2) \\) and LLM conditions which are indicated in the respective tables. Throughout this study, the \\( \\mathrm{P^2KG}(R=1) \\) is employed with a batch size of 256, unless explicitly stated otherwise. More implementation details are available in the section C.2 of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics and baselines: We use mean intersection over union (mIoU), mean accuracy (mAcc), and success rate (SR) as evaluation metrics. SR is the most stringent metric. See sec. C.4 of supplementary for more details. We compare our model with state-of-the-art methods: WLTDO [16], UAAA [1], UPN [49], DDN [9], PlaTe [50], Ext-GAIL [7], \\(\\mathrm{P}^3\\mathrm{IV}\\) [59], PDPP [54], SkipPlan [29], and E3P [53]. More details of these methods are available in the section C.5 of supplementary material. Compared to other models, PDPP uses a different experimental setting. In PDPP, authors set the window after the start time of \\(a_1\\) and before the end time of \\(a_T\\), contrary to the standard practice of setting a 2-second window around the start and end" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18820" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.865, + 0.38 + ], + "angle": 0, + "content": "
ModelsRequired AnnotationsT=3T=4
step classvisual statesstep texttask classSR†mAcc†mIoU†SR†mAcc†mIoU†
Random✓<0.010.941.66<0.010.831.66
Retrieval-Based✓8.0523.332.063.9522.2236.97
WLTDO [16]✓✓1.8721.6431.700.7717.9226.43
UAAA [1]✓✓2.1520.2130.870.9819.8627.09
UPN [49]✓✓2.8924.3931.561.1921.5927.85
DDN [9]✓✓12.1831.2947.485.9727.1048.46
PlaTe [50]✓✓16.0036.1765.9114.0035.2955.36
Ext-GAIL wo Aug. [7]✓✓18.0143.8657.16---
Ext-GAIL [7]✓✓21.2749.4661.7016.4143.0560.93
P3IV♣[59]✓✓23.3449.9673.8913.4044.1670.01
PDPP♣[54]✓✓26.3855.6259.3418.6952.4462.38
E3P♣[53]✓✓✓26.4053.0274.0516.4948.0070.16
SkipPlan [29]♣✓28.8561.1874.9815.5655.6470.30
Ours w/ P2KG (R=2)✓22.6048.7653.5713.9045.7955.00
Ours♣w/ P2KG (R=1)✓33.3461.3664.1420.3855.5464.03
Ours♣w/ P2KG (R=2)✓33.3860.7963.8921.0256.0864.15
PDPP♣†[54]✓✓37.2064.6766.5721.4857.8265.13
Ours♣†w/ P2KG (R=1)✓38.1264.7467.1524.1559.0566.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.389, + 0.895, + 0.434 + ], + "angle": 0, + "content": "Table 1. Performance of our method in comparison to existing baselines for CrossTask dataset. \\(\\clubsuit\\) means that the input visual features are from the S3D network [35] pretrained on HowTo100M [34]; otherwise, precomputed features provided in CrossTask are used. \\(\\dagger\\) indicates the results are under the PDPP's task setting, while others are under the conventional setting" + }, + { + "type": "table", + "bbox": [ + 0.169, + 0.442, + 0.381, + 0.606 + ], + "angle": 0, + "content": "
ModelsT=5T=6
DDN [9]3.101.20
P³IV♣[59]7.214.40
PDPP♣[54]13.227.49
E3P♣[53]8.965.76
SkipPlan♣[29]8.555.12
Ours (R=2)8.175.32
Ours♣(R=1)13.258.09
Ours♣(R=2)12.749.23
PDPP♣†[54]13.458.41
Ours♣†(R=1)14.209.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.615, + 0.472, + 0.645 + ], + "angle": 0, + "content": "Table 2. Success Rate \\((SR^{\\dagger})\\) comparison to existing baselines for CrossTask dataset under longer horizons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.681 + ], + "angle": 0, + "content": "time (ref. [9]). We conduct experiments on both PDPP's proposed setting and the conventional setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.471, + 0.835 + ], + "angle": 0, + "content": "Inference: During the inference phase, the model receives only the start observation \\( v_{s} \\) and the goal observation \\( v_{g} \\). To proceed, it utilizes a step model to predict the initial action \\( a_{1} \\) and the end action \\( a_{T} \\) for each data. Subsequently, leveraging the \\( \\mathrm{P^2KG} \\), highest probable procedure knowledge graph plans connecting \\( a_{1} \\) and \\( a_{T} \\) are obtained. Then, a multi-dimensional array, is created as mentioned in Eq. 5. Finally, the planning model is used to predict the sequence of actions \\( [a_1,\\dots,a_T] \\) by denoising the generated multi-dimensional array as in § 3.2.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.465, + 0.865 + ], + "angle": 0, + "content": "4.1. Comparison with the State of the Art (SOTA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "CrossTask (short horizon): We evaluate on CrossTask for short horizons (\\(T = 3\\) and \\(T = 4\\)). According to the results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.595 + ], + "angle": 0, + "content": "shown in Table 1, our proposed method outperforms the PDPP in PDPP's setting in every evaluation metric. More than \\(0.9\\%\\) and \\(2\\%\\) improvement in success rate in \\(T = 3\\) and \\(T = 4\\) respectively. In the conventional setting, our method with both \\(\\mathrm{P}^2\\mathrm{KG}\\) (\\(R = 1\\)) and \\(\\mathrm{P}^2\\mathrm{KG}\\) (\\(R = 2\\)) conditions outperform the success rate values by a significant margin compared to other baselines. \\(\\mathrm{P}^2\\mathrm{KG}\\) (\\(R = 2\\)) slightly outperforms \\(\\mathrm{P}^2\\mathrm{KG}\\) (\\(R = 1\\)), indicating potential benefits of incorporating more procedural knowledge from the \\(\\mathrm{P}^2\\mathrm{KG}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.824 + ], + "angle": 0, + "content": "CrossTask (long horizon): We use long-horizon predictions for \\( T = 5 \\) and \\( T = 6 \\) for further evaluating our model as shown in Table 2. In PDPP's setting (\\(\\dagger\\)), our method improves the success rate in both \\( T = 5 \\) and \\( T = 6 \\). In the conventional setting, our method utilizing \\( \\mathrm{P}^2\\mathrm{KG} \\) (\\( R = 1 \\)) demonstrates the highest SR value for \\( T = 5 \\), and for a longer horizon at \\( T = 6 \\), our method delivers superior performance for \\( \\mathrm{P}^2\\mathrm{KG} \\) (\\( R = 2 \\)). Our method performs well under the challenging scenario of a long planning horizon. Our success rate (SR) diminishes from approximately \\( 40\\% \\) to \\( 10\\% \\) when extending the planning horizon from \\( T = 3 \\) to \\( T = 6 \\), primarily due to the heightened uncertainty surrounding the predicted plan between the initial and final steps. This uncertainty stems from the increase in the number of potential procedural plans within the \\( \\mathrm{P}^2\\mathrm{KG} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "NIV and COIN: Results are shown in Table 3 and Table 4. On NIV, ours achieves the best result under the mIoU metric with \\( T = 3 \\), and under both the SR and mIoU metrics with \\( T = 4 \\). The results on NIV (\\( T = 5 \\), \\( T = 6 \\)) are available in the supplementary material. For the COIN dataset we only" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "18821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "report SR and mAcc due to space constraints; mIoU is reported in the supplementary material. Our method does not rank as the top performer on COIN when \\( T = 3 \\) or \\( T = 4 \\). The likely reason is that the COIN dataset features just an average of 3.9 actions per video—a scenario that demands only short-horizon planning and does not necessitate advanced procedural knowledge (which encompasses long sequence-level knowledge [62]). Furthermore, the dataset's extensive collection of over 11k videos provides a substantial resource for baselines to learn basic procedural knowledge." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.254, + 0.465, + 0.368 + ], + "angle": 0, + "content": "
ModelsNIV (T=3)NIV (T=4)
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random2.214.076.091.122.735.84
DDN [9]18.4132.5456.5615.9727.0953.84
Ext-GAIL [7]22.1142.2065.9319.9136.3153.84
P3IV [59]24.6849.0174.2920.1438.3667.29
E3P [53]26.0551.2475.8121.3741.9674.90
PDPP [54]22.2239.5086.6621.3039.2484.96
Ours24.4443.4686.6722.7141.5991.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.379, + 0.451, + 0.392 + ], + "angle": 0, + "content": "Table 3. Performance of baselines and ours for NIV dataset" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.415, + 0.465, + 0.542 + ], + "angle": 0, + "content": "
ModelsCOIN (T=3)COIN (T=4)COIN (T=5)
SR↑mAcc↑SR↑mAcc↑SR↑mAcc↑
Random< 0.01< 0.01< 0.01< 0.01--
Retrieval4.3817.402.7114.29--
DDN [9]13.9020.1911.1317.71--
P3IV [59]15.4021.6711.3218.854.2710.81
E3P [53]19.5731.4213.5926.72--
PDPP [54]19.4243.4413.6742.5813.0243.36
SkipPlan [29]23.6547.1216.0443.199.9038.99
Ours (R=2)20.2539.8715.6339.5316.0640.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.552, + 0.456, + 0.565 + ], + "angle": 0, + "content": "Table 4. Performance of baselines and ours for COIN dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.348, + 0.607 + ], + "angle": 0, + "content": "4.2. Ablation Studies and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.719 + ], + "angle": 0, + "content": "Ablation on the probabilistic procedure knowledge graph. We analyze the role of \\(\\mathrm{P}^2\\mathrm{KG}\\) in improving the performance of our proposed method. Table 5 shows the results which clearly demonstrate that using \\(\\mathrm{P}^2\\mathrm{KG}\\) conditions improves the performance significantly for every \\(T\\) value. Especially when \\(T = 4\\), success rate (SR) improves more than \\(3\\%\\) and mean IoU improves more than \\(2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Plan recommendations provided by probabilistic procedure knowledge graph v.s. LLM. We recognize the recent trend of utilizing LLMs to enhance action anticipation [60] or planning in other realms [3, 27, 28, 40, 46]. In Table 6, we compare the results between using \\(\\mathrm{P}^2\\mathrm{KG}\\) v.s. using LLM ('llama-2-13b-chat' and 'llama-2-70b-chat') to generate the plan recommendations. When examining Table 6, it becomes apparent that there are trade-offs between using LLM-generated recommendations and \\(\\mathrm{P}^2\\mathrm{KG}\\) recommendations. For instance, \\(\\mathrm{P}^2\\mathrm{KG}\\) recommendations are constrained by the data available in the training set, limiting their applicability to unseen procedural activities. On the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "other hand, LLMs tend to exhibit better generalization to such unseen activities. However, considering that the training and testing are conducted on the aforementioned three datasets with known activities, \\(\\mathrm{P}^2\\mathrm{KG}\\) recommendations can yield more accurate results compared to relying on LLM-generated recommendations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Probabilistic procedure knowledge graph (P²KG) vs. Frequency-based procedure knowledge graph (PKG). The probabilistic procedure knowledge graph uses outedge normalization to encode step transition probabilities (§ 3.2.2), while the frequency-based procedure knowledge graph uses min-max normalization over the frequency counts throughout the graph. In both cases, the planning model only uses one procedure plan recommendation from the graph as condition in our experimental analysis. By looking at the results shown in Table 7, it is evident that the probabilistic procedure knowledge graph outperforms the frequency-based procedure knowledge graph." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Effect of utilizing predicted steps for input conditions to train the procedure planing model. Our proposed problem decomposition allows training the planning model with ground truth (GT) first and last steps. We experiment with two ways to train the planning model. Method 1 uses the predicted start and end steps \\((\\hat{a}_1\\) and \\(\\hat{a}_T)\\) as input to generate \\(\\mathrm{P}^2\\mathrm{KG}\\) conditions and use them to train the planning model. Method 2 is where we augment the predicted start and end steps using the GT start and end steps \\((a_1\\) and \\(a_{T})\\) by generating 3 more data samples as follows: \\([\\hat{a}_1,a_T]\\), \\([a_1,\\hat{a}_T]\\), and \\([a_1,a_T]\\). Then we generate \\(\\mathrm{P}^2\\mathrm{KG}\\) conditions for each data and train the model. From the results shown in Table 8, the method without GT data augmentation shows better results. This suggests that leveraging ground truth data in training can lead to worse performance in testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Qualitative results. Figures 3 and 4 provide qualitative examples of our method. Intermediate steps are padded in the step model because it only predicts the start and end actions. In the 'make jello shot' task (see Figure 3), the model gives a wrong prediction in the intermediate steps when using \\(\\mathrm{P}^2\\mathrm{KG}\\) (\\(R = 1\\)) condition. However, it predicts correctly when using \\(\\mathrm{P}^2\\mathrm{KG}\\) (\\(R = 2\\)) conditions. In the 'change a tire' task shown in Figure 4, the model is able to predict all the intermediate steps in given conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Visualizations of the probabilistic procedure knowledge graph. We show a sub-graph from our probabilistic procedure knowledge graph (Figure 5). This graph is drawn around the 'jack up' node up to the depth of 2 nodes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visualizations of the expert trajectories. Figure 6 illustrates the steps involved in completing the 'make jello shots' task, along with their transitions to other steps within the entire training data. This figure demonstrates that our \\(\\mathrm{P}^2\\mathrm{KG}\\) encodes diverse sequencing possibilities for steps and also captures task-sharing steps across the entire training domain. For instance, 'pour water' is a step in 'make jello" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18822" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.076, + 0.868, + 0.155 + ], + "angle": 0, + "content": "
ModelT=3T=4T=5T=6
SRmAccmIoUSRmAccmIoUSRmAccmIoUSRmAccmIoU
w.o P2KG conditions †35.6963.9166.0420.5257.4764.3912.853.4464.018.1550.4564.13
Ours †38.1264.7467.1524.1559.0566.6414.2053.8465.569.2750.2265.97
w.o P2KG conditions31.3559.5163.1118.9256.2062.4712.7151.2963.568.1647.6363.39
Ours33.3860.7963.8921.0256.0864.1512.7451.2363.169.2350.7865.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.201, + 0.165, + 0.768, + 0.179 + ], + "angle": 0, + "content": "Table 5. Performance of our method with and without \\( {\\mathrm{P}}^{2}\\mathrm{{KG}} \\) conditions on CrossTask \\( {}^{ \\star } \\) dataset" + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.203, + 0.462, + 0.38 + ], + "angle": 0, + "content": "
Model (T=6, CrossTask*)SRmAccmIoU
Ours with P2KG (R=1)
PDPP setting9.2750.2265.97
Conventional setting8.0950.8065.39
One LLM plan recommendation
PDPP setting (13b)7.7450.2864.05
Conventional setting (13b)7.2149.6863.89
PDPP setting (70b)8.6250.3164.34
Conventional setting (70b)7.8149.7564.02
P2KG (R=1) and one LLM plan recommendation
PDPP setting (13b)8.8149.9765.22
Conventional setting (13b)8.2051.4664.30
PDPP setting (70b)9.0150.2565.57
Conventional setting (70b)8.3451.5364.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.391, + 0.468, + 0.418 + ], + "angle": 0, + "content": "Table 6. Performance of the plan recommendations provided by the probabilistic procedure knowledge graph v.s. LLM." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.438, + 0.401, + 0.479 + ], + "angle": 0, + "content": "
ModelsSRmAccmIoU
Frequency graph7.6648.6164.21
Probabilistic graph8.0950.8065.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.49, + 0.468, + 0.531 + ], + "angle": 0, + "content": "Table 7. Performance comparison between probabilistic procedure knowledge graph v.s. frequency-based procedure knowledge graph for \\( T = 6 \\) on CrossTask dataset" + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.55, + 0.416, + 0.592 + ], + "angle": 0, + "content": "
ConditionSRmAccmIoU
without GT data aug.38.1264.7467.15
with GT data aug.32.4562.4262.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.603, + 0.468, + 0.631 + ], + "angle": 0, + "content": "Table 8. Effect of different input conditions for performance on CrossTask \\( {}^{ \\star } \\) dataset \\( \\left( {T = 3}\\right) \\) in PDPP’s setting" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.649, + 0.452, + 0.704 + ], + "angle": 0, + "content": "
ModelsT=3T=4T=5T=6
ˆa1ˆaTˆa1ˆaTˆa1ˆaTˆa1ˆaT
Ours53.6950.6055.5652.5155.5851.8157.0951.92
Ours♣71.4263.3272.9863.3772.4263.2963.8259.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.714, + 0.468, + 0.741 + ], + "angle": 0, + "content": "Table 9. The step model's start and end step prediction accuracies on the CrossTask dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.76, + 0.468, + 0.805 + ], + "angle": 0, + "content": "shots' task, but it can also be part of other tasks, leading to a step transition from 'pour water' to 'add fish.' This structure allows models to leverage rich procedural knowledge." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.468, + 0.853 + ], + "angle": 0, + "content": "Results for the step model. Table 9 reveals step model results, indicating potential enhancement areas to elevate the planning performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Limitations & Failure cases. Our model exhibits three distinct failure case patterns. See section B.5 of the supplementary material for detailed discussions." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.187, + 0.887, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.354, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Figure 5. Example of a sub-graph in our probabilistic procedure knowledge graph \\((\\mathrm{P}^2\\mathrm{KG})\\) for CrossTask dataset. This graph effectively encapsulates real-world knowledge of distinct transition probabilities between steps, e.g., the probability of transitioning from 'start loose' to 'jack up' is 0.65, in contrast to a mere 0.14 for the reverse transition—the \\(\\mathrm{P}^2\\mathrm{KG}\\) reflects the common real-life practice where loosening the lug nuts before jacking up the car leads to a safer and more efficient tire change." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.487, + 0.888, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.648, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Figure 6. Expert trajectories of the 'Make Jello Shots' task, involving task-sharing steps and thus out-of-task step transitions. Thicker lines indicate paths that are more frequently visited" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.618, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We focus on formulating procedural plans from an AI agent in instructional videos. We propose KEPP which employs a probabilistic procedural knowledge graph, sourced from the training domain, effectively serving as a 'textbook' for procedure planning. Results show that KEPP delivers SOTA performance with minimal supervision. Future work can focus on enhancing the accuracy of predictions for the initial and final steps. Additionally, our approach can be modified to aid in detecting erroneous steps and the misordering of steps in instructional videos [38, 41]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18823" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Yazan Abu Farha and Juergen Gall. Uncertainty-aware anticipation of activities. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.24 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[3] Anurag Ajay, Seungwook Han, Yilun Du, Shaung Li, Abhi Gupta, Tommi Jaakkola, Josh Tenenbaum, Leslie Kaelbling, Akash Srivastava, and Pulkit Agrawal. Compositional foundation models for hierarchical planning. arXiv preprint arXiv:2309.08587, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4575-4583, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. arXiv preprint arXiv:2307.08763, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[6] Anonymous authors. Active procedure planning with uncertainty-awareness in instructional videos, 2023. Under review as a conference paper at ICLR 2024. https://openreview.net/pdf?id=JDd46WodYf.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.496, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[7] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[8] Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin, Jena D Hwang, Xiang Lorraine Li, Hirona J Arai, Soumya Sanyal, Keisuke Sakaguchi, Xiang Ren, and Yejin Choi. Plasma: Making small language models better procedural knowledge models for (counterfactual) planning. arXiv preprint arXiv:2305.19472, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[9] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[10] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where?-self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[11] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2447, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Hazel Doughty, Ivan Laptev, Walterio Mayol-Cuevas, and Dima Damen. Action modifiers: Learning from adverbs in" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 868-878, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[13] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[14] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[15] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18952-18961, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[16] Kiana Ehsani, Hessam Bagherinezhad, Joseph Redmon, Roozbeh Mottaghi, and Ali Farhadi. Who let the dogs out? modeling dog behavior from visual data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4051-4060, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[17] Sophie Fischer, Carlos Gemmell, Iain Mackie, and Jeffrey Dalton. Vilt: Video instructions linking for complex tasks. In Proceedings of the 2nd International Workshop on Interactive Multimedia Retrieval, pages 41-47, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[18] Kevin Flanagan, Dima Damen, and Michael Wray. Learning temporal sentence grounding from narrated egovideos. arXiv preprint arXiv:2310.17395, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[19] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. arXiv preprint arXiv:2005.03684, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[20] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threedorld transport challenge: A visually guided task-and-motion planning benchmark for physically realistic embodied ai. arXiv preprint arXiv:2103.14025, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[21] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. Machine Intelligence Research, 20 (2):249-262, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] De-An Huang, Joseph J Lim, Li Fei-Fei, and Juan Carlos Niebles. Unsupervised visual-linguistic reference resolution in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2183–2192, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18824" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] De-An Huang, Shyamal Buch, Lucio Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding\" it\": Weakly-supervised reference-aware visual grounding in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5948-5957, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[26] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[27] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[28] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[29] Zhiheng Li, Wenjia Geng, Muheng Li, Lei Chen, Yansong Tang, Jiwen Lu, and Jie Zhou. Skip-plan: Procedure planning in instructional videos via condensed action space learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10297-10306, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[30] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[31] Yujie Lu, Pan Lu, Zhiyu Chen, Wanrong Zhu, Xin Eric Wang, and William Yang Wang. Multimodal procedural planning via dual text-image prompting. arXiv preprint arXiv:2305.01795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[32] Weichao Mao, Ruta Desai, Michael Louis Iuzzolino, and Nitin Kamra. Action dynamics task graphs for learning plannable representations of procedural tasks. arXiv preprint arXiv:2302.05330, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[33] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. arXiv preprint arXiv:2306.03802, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[35] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[36] Davide Moltisanti, Frank Keller, Hakan Bilen, and Laura Sevilla-Lara. Learning action changes by measuring verb-adverb textual relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23110-23118, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[37] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[38] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[39] Megha Nawhal, Akash Abdu Jyothi, and Greg Mori. Rethinking learning approaches for long-term action anticipation. In European Conference on Computer Vision, pages 558-576. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[40] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15302-15314, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[41] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[42] Fadime Sener, Rishabh Saraf, and Angela Yao. Transferring knowledge from text to video: Zero-shot anticipation for procedural actions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[43] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[44] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[45] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[46] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[47] Yale Song, Gene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Look for the change: Learning object states and state-modifying actions from untrimmed" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18825" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "web videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13956-13966, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[49] Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal planning networks: Learning generalizable representations for visuomotor control. In International Conference on Machine Learning, pages 4732-4741. PMLR, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[50] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters, 7(2):4924-4930, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[51] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34:14476-14487, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[52] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[53] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13565–13575, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[54] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14836-14845, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[55] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[56] Yue Yang, Joongwon Kim, Artemis Panagopoulou, Mark Yatskar, and Chris Callison-Burch. Induce, edit, retrieve: Language grounded multimodal schema for instructional video retrieval. arXiv preprint arXiv:2111.09276, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[57] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[58] Jiahao Zhang, Anoop Cherian, Yanbin Liu, Yizhak Ben-Shabat, Cristian Rodriguez, and Stephen Gould. Aligning step-by-step instructional diagrams to video demonstrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2483-2492, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[59] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[60] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[61] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[62] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.893, + 0.372 + ], + "angle": 0, + "content": "[63] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18826" + } + ] +] \ No newline at end of file diff --git a/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_origin.pdf b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01db787f39f5b09d8b868440959c0cab1d1c70dd --- /dev/null +++ b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/f83cbc4e-8548-48f7-bd06-1fc89d0517b0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0980289486bfcdeb9edd61e93317db7695be19d071db82c8270ca9e79f10c65c +size 1422659 diff --git a/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/full.md b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a98f7d2825b0fbd28522d122c658613681ed78ef --- /dev/null +++ b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/full.md @@ -0,0 +1,309 @@ +# Why Not Use Your Textbook? Knowledge-Enhanced Procedure Planning of Instructional Videos + +Kumaranage Ravindu Yasas Nagasinghe $^{1}$ Honglu Zhou $^{2}$ Malitha Gunawardhana $^{1,3}$ + +Martin Renqiang Min $^{2}$ Daniel Harari $^{4}$ Muhammad Haris Khan $^{1}$ + +1Mohamed bin Zayed University of Artificial Intelligence, 2NEC Laboratories, USA, + +3 University of Auckland, 4Weizmann Institute of Science + +ravindu.nagasinghe@mbzuai.ac.ae, muhammad.haris@mbzuai.ac.ae + +# Abstract + +In this paper, we explore the capability of an agent to construct a logical sequence of action steps, thereby assembling a strategic procedural plan. This plan is crucial for navigating from an initial visual observation to a target visual outcome, as depicted in real-life instructional videos. Existing works have attained partial success by extensively leveraging various sources of information available in the datasets, such as heavy intermediate visual observations, procedural names, or natural language step-by-step instructions, for features or supervision signals. However, the task remains formidable due to the implicit causal constraints in the sequencing of steps and the variability inherent in multiple feasible plans. To tackle these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities by infusing it with procedural knowledge. This knowledge, sourced from training procedure plans and structured as a directed weighted graph, equips the agent to better navigate the complexities of step sequencing and its potential variations. We coin our approach KEPP, a novel Knowledge-Enhanced Procedure Planning system, which harnesses a probabilistic procedural knowledge graph extracted from training data, effectively acting as a comprehensive textbook for the training domain. Experimental evaluations across three widely-used datasets under settings of varying complexity reveal that KEPP attains superior, state-of-the-art results while requiring only minimal supervision. Code and trained model are available at https://github.com/Ravindu-Yasas-Nagasinghe/KEPP + +# 1. Introduction + +The evolution of the internet has precipitated an unprecedented influx of video content, serving as a vital educational resource for myriad learners. Individuals frequently + +![](images/0138b6fb7c6793c604dae03f21e39301720353eac987411ca960672a8fd230ce.jpg) +Figure 1. Expert trajectories [7] of the 'Make Jello Shots' task from the CrossTask dataset [63]. Heavier color indicates more frequently visited path. This depicts the complexities of the procedure planning task, arising from the subtle causal links in step sequencing (e.g., steps like 'stir mixture' or 'pour mixture' typically occur after adding individual ingredients), the varied probabilities of transitioning between steps, and the diversity in plans viable for a given starting point and an intended outcome. Motivated by these nuanced challenges, we propose Knowledge-Enhanced Procedure Planning (KEPP) with the use of a probabilistic procedure knowledge graph to capture and represent these intricacies + +leverage platforms such as YouTube to acquire new skills, ranging from culinary arts to automobile maintenance [34]. While these instructional videos benefit the development of intelligent agents in mastering long-horizon tasks, the challenge extends beyond merely interpreting visuals. It requires high-level reasoning and planning to effectively assist in complex, real-life scenarios [13]. + +Procedure Planning in Instructional Videos demands an agent to produce a sequence of actionable steps, thereby crafting a procedure plan that facilitates the transition from an initial visual observation of the physical world towards achieving a desired goal state [7, 9, 50, 53, 54, 59]. The task acts as a precursor to an envisioned future scenario in which an agent like a robot provides on-the-spot support, such as + +assisting an individual in preparing a recipe [6]. + +Current methods in procedure planning in instructional videos make extensive use of various annotations available within the datasets to enrich input features or provide supervisory signals (see Table 1). These include detailed, temporally localized visual observations of intermediate action steps throughout the procedure plan [7, 9, 50], high-level procedural task labels [53, 54], and step-by-step instructions in natural language [53, 59]. Despite advancements, significant challenges persist, including characterizing the implicit causal constraints in step sequencing, the varied probabilities of transitioning between steps, and the inherent variability of multiple viable plans (see Fig. 1). + +To address these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities in procedure planning by infusing it with comprehensive procedural knowledge [62], derived from training procedure plans and structured as a directed weighted graph. This graph, as a Probabilistic Procedural Knowledge Graph [5] where nodes denote steps from diverse tasks and edges represent step transition probabilities in the training domain, empowers agents to more adeptly navigate the complexities of step sequencing and its potential variations. + +Our proposed approach, KEPP, is a novel Knowledge-Enhanced Procedure Planning system (Fig. 2) that harnesses a probabilistic procedural knowledge graph $(\mathrm{P}^2\mathrm{KG})$ , constructed from training procedure plans. This graph functions like a detailed textbook, providing extensive knowledge for the training domain, and thereby circumventing the need for costly multiple annotations required by existing methods. Additionally, we decompose the instructional video procedure planning problem into two parts: one driven by objectives specific to step perception and the other by a procedural knowledge-informed modeling of procedure planning. In this problem decomposition, the first and last action steps are predicted based on the initial and goal visual states. Following this, a procedure plan is crafted by leveraging the procedure plan recommendations retrieved from the $\mathrm{P}^2\mathrm{KG}$ . The recommendations correspond to the most probable procedure plans frequently used in training, conditioned on the predicted first and last action steps. In a similar vein to the approach by Li et al. [29], our proposed decomposition strategy reduces uncertainty by maximizing the use of currently available information, namely the initial and goal visual states. This allows for the improvement of procedure planning through more accurate predictions of start and end actions. Plus, this decomposition effectively incorporates procedural knowledge into procedure planning, thereby enhancing its effectiveness. + +Our contributions are as follows: + +- We propose KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos that leverages rich procedural knowledge from a probabilistic procedu + +ral knowledge graph (P2KG). This approach necessitates only a minimal amount of annotations for supervision. + +- We decompose the problem in procedure planning of instructional videos: predicting the initial and final steps from the start and end visuals, and then creating a plan using procedural knowledge retrieved based on these predicted steps. This approach prioritizes the currently available information and effectively incorporates procedural knowledge, enhancing strategic planning. +- Experimental evaluations on three widely-used datasets, under settings of varying complexity, reveal that KEPP attains state-of-the-art results in procedure planning. + +# 2. Related Work + +Instructional Videos, which demonstrate multi-step procedures, have become a hotbed of research. The studies delve into various aspects, including comprehending and extracting intricate spatiotemporal content from video [12, 18, 19, 21, 23, 36, 43, 44, 48, 55, 57], interpreting the interrelationships between various actions and procedural events [47, 63], and developing capabilities for forecasting [39, 42] and strategic reasoning and planning [28] within the context of these videos. Furthermore, by leveraging the multimodality of visual, auditory, and narrative elements within these videos, research extends to areas like multimodal alignment [2, 58], grounding [10, 14, 25, 33, 51], representation learning [11, 35, 61], pre-training [15, 26, 62], and more [17, 24, 37, 56]. This paper focuses on procedure planning in instructional videos. + +Procedure Planning is a vital skill for autonomous agents tasked with handling complex activities in everyday settings. Essentially, these agents must discern the appropriate actions to reach a specific goal. This aspect of artificial intelligence (AI) has been a prominent and integral subject in fields like robotics [20, 28, 32, 45, 46]. Yet, the challenge of procedure planning in the context of instructional videos is notably distinct, and potentially more complex, than its counterparts in natural language processing [8, 30], multimodal generative AI [13, 31], and simulated environments [27, 28, 45]. Its significance is underscored by the need for planning that is grounded in real-world scenes. This requires the development of AI agents capable of accurately perceiving and understanding the current real-world context, and then anticipating and mapping out a logical sequence of actions to fulfill a high-level goal effectively. + +Procedure Planning in Instructional Videos has recently garnered research attention. DDN [9] initiates this trend by conceptualizing the problem as sequential latent space planning. Building on this, PlaTe [50] employs transformers for both action and state models, integrating Beam Search for enhanced performance. Meanwhile, Ext-GAIL [7] suggests employing contextual modeling through variational autoencoder and adversarial policy learning. This method + +considers contextual information as time-invariant knowledge, crucial for distinguishing specific tasks and allowing for multiple planning outcomes. + +While these earlier approaches have viewed procedure planning as an autoregressive sequence generation problem, recent methods regard it as a distribution-fitting problem to mitigate error propagation in sequential decisions. In this vein, $\mathrm{P}^3\mathrm{IV}$ [59] replaces intermediate visual states with linguistic representations for supervision, predicting all steps simultaneously instead of using autoregressive methods. To circumvent the complex learning strategies and high annotation costs of previous work, PDPP [54] models the entire intermediate action sequence distribution using a conditioned projected diffusion model. This approach redefines the planning problem as a sampling process from this distribution and simplifies supervision by using only instructional video task labels. E3P [53], also encoding task information, adopts a mask-and-predict strategy for mining step relationships in procedural tasks, integrating probabilistic masking for regularization. In contrast, our approach does not rely on annotations of intermediate states, natural language step representations, or procedural task labels. + +Recognizing the difficulties inherent in high dimensional state supervision and the accumulation of errors in action sequences, SkipPlan [29] was developed. It strategically focuses on action predictions, breaking down longer sequences into shorter, more manageable sub-chains by skipping over less reliable intermediate actions. Drawing inspiration from SkipPlan, our approach decomposes the procedure planning problem to prioritize the most reliable information available (ref. § 3.1.2). However, we innovate further by incorporating a Probabilistic Procedure Knowledge Graph, significantly enriching the planning phase. + +# 3. Methodology + +We will first introduce the problem setup in § 3.1, and then discuss our novel Knowledge-Enhanced Procedure Planning system (KEPP) in § 3.2. See Fig. 2 for KEPP overview. + +# 3.1. Problem and Method Overview + +# 3.1.1 Problem Formulation + +We follow the problem definition for procedure planning of instructional videos put forth by Chang et al. [9]: given an observation of the initial state $v_{start}$ and a goal state $v_{goal}$ , both are short video clips indicating different states of the real-world environment extracted from an instructional video, a model is required to plan a sequence of action steps $a_{1:T}$ to reach the indicated goal. Here, $T$ is the planning horizon, inputting to the model, corresponding to the number of action steps in the sequence produced by the model so that the environment state can be transformed from $v_{start}$ to $v_{goal}$ . We use $a_t$ to denote the action step at the timestamp + +$t$ , and in the following, $v_{s}$ and $v_{g}$ are short for $v_{start}$ and $v_{goal}$ . Mathematically, the procedure planning problem is defined as $p(a_{1:T}|v_s,v_g)$ that denotes the conditional probability distribution of the action sequence $a_{1:T}$ given the initial visual observation $v_{start}$ and the goal visual state $v_{goal}$ . + +# 3.1.2 Problem Decomposition + +Considering the initial and final visual states are input, providing the most reliable information, we hypothesize that predicting the first and final action steps is more dependable than interpolating the intermediate ones, and consequently, an enhanced accuracy in predicting the first and final steps can lead to more effective procedure planning. Inspired by this hypothesis, we decompose the procedure planning problem into two sub-problems, as shown in Eq. 1: + +$$ +p \left(\hat {a} _ {1: T} | v _ {s}, v _ {g}\right) = p \left(\hat {a} _ {2: T - 1} | \hat {a} _ {1}, \hat {a} _ {T}\right) p \left(\hat {a} _ {1}, \hat {a} _ {T} | v _ {s}, v _ {g}\right), \tag {1} +$$ + +where the first sub-problem is to identify the beginning step $a_1$ and the end step $a_T$ , and the second sub-problem is to plan the intermediate action steps $a_{2:T-1}$ given $a_1$ and $a_T$ . We use $\hat{a}_t$ to denote predicted action step at timestamp $t$ . + +Our proposed problem decomposition in Eq. 1 bears resemblance with the problem formulation from Li et al. [29]; they decompose procedure planning into $p(\hat{a}_{1:T}|v_s, v_g) = \prod_{t=2}^{T-1} p(\hat{a}_t|\hat{a}_1, \hat{a}_T)p(\hat{a}_1, \hat{a}_T|v_s, v_g)$ . However, our formulation differs in its approach to modeling the second subproblem. Specifically, we employ a conditioned projected diffusion model (ref. § 3.2) to jointly predict $a_{2:T-1}$ at once, whereas Li et al. [29] rely on Transformer decoders to predict each intermediate action independently. Further, we integrate a Probabilistic Procedure Knowledge Graph (ref. § 3.2.2) to address the second sub-problem. + +Tackling the second sub-problem is nontrivial even when armed with an oracle predictor for the first sub-problem. Procedure planning in real-life scenarios remains daunting because of the following challenges: (1) the presence of implicit temporal and causal constraints in the sequencing of steps, (2) the existence of numerous viable plans given an initial state and a goal state, and (3) the need to incorporate the real-life everyday knowledge both in task-sharing steps and in managing the inherent variability in transition probabilities between steps. Previous studies tackled these challenges by extensively harnessing detailed annotations in the datasets to augment input features or offer supervision signals (see Table 1). In contrast, we propose harnessing a Probabilistic Procedural Knowledge Graph $(\mathrm{P}^2\mathrm{KG})$ which is extracted from the procedure plans in the training set. With the $\mathrm{P}^2\mathrm{KG}$ at our hand, we further decompose the procedure planning problem to reduce its complexity and learn $f_{\theta}:(v_s,v_g,T)\to p(\hat{a}_{1:T}|v_s,v_g)$ as follows: + +![](images/1cf3937d5a075be8ca996915459c39187231a319a4a72393896a35b18cb5fe87.jpg) +Figure 2. Overview of our methodology. We introduce KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos, leveraging a Probabilistic Procedural Knowledge Graph $(\mathrm{P}^2\mathrm{KG})$ . KEPP breaks down procedure planning into two parts: predicting initial and final steps from visual states, and crafting a procedure plan based on the procedural knowledge retrieved from $\mathrm{P}^2\mathrm{KG}$ , conditioned on the predicted first and last action steps. KEPP requires minimal annotations and enhances planning effectiveness + +$$ +\begin{array}{l} p \left(\hat {a} _ {1: T} | v _ {s}, v _ {g}\right) = \\ p \left(\hat {a} _ {1: T} | \tilde {a} _ {1: T}, v _ {s}, v _ {g}\right) p \left(\tilde {a} _ {1: T} | \hat {a} _ {1}, \hat {a} _ {T}\right) p \left(\hat {a} _ {1}, \hat {a} _ {T} | v _ {s}, v _ {g}\right) \tag {2} \\ \end{array} +$$ + +where $f_{\theta}$ denotes the machine learning model, and $\tilde{a}_{1:T}$ represents a graph path (i.e., a sequence of nodes) retrieved from $\mathrm{P}^2\mathrm{KG}$ . This graph path provides a valuable procedure plan recommendation aligned with the training domain, thus mitigating the complexity of procedure planning. It is worth noting that the proposed approach to modeling procedure planning using Eq. 2 demands only a minimal level of supervision, merely relying on the ground truth training procedure plan; Eq. 2 circumvents the need for additional annotations. We describe details of our $\mathrm{P}^2\mathrm{KG}$ -enhanced approach in the following subsection. + +# 3.2. KEPP: Knowledge-Enhanced Procedure Planning + +We propose KEPP (Fig. 2) utilizing a probabilistic procedure knowledge graph extracted from the training set. We firstly identify the beginning and conclusion steps according to the input initial and goal states; and then, conditioned on these steps and the planning horizon $T$ , we query the graph to retrieve relevant procedural knowledge for knowledge-enhanced procedure planning of instructional videos. + +# 3.2.1 Identify Beginning and Conclusion Steps + +Given $v_{start}$ and $v_{goal}$ as input, we adapt a Conditioned Projected Diffusion Model [54] (ref. supplementary material) to identify the first action step and the final step; we refer to this model as the 'Step (Perception) Model'. + +Standard Denoising Diffusion Probabilistic Model tackles data generation through a denoising Markov chain over variables $\{x_{N}\ldots x_{0}\}$ , starting with $x_{N}$ as a Gaussian random distribution [22]. In the forward diffusion phase, Gaussian noise $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ is progressively added to the initial, unaltered data $x_0$ , transforming it into a Gaussian random distribution. Conversely, the reverse denoising process + +transforms Gaussian noise back into a sample. Denoising is parameterized by a learnable noise prediction model, and the learning objective is to learn the noise added to $x_0$ at each diffusion step. After training, the diffusion model generates data akin to $x_0$ by iteratively applying the denoising process, starting from random Gaussian noise. + +Adopting Conditioned Projected Diffusion Model as the Step Model. For our step model, the distribution we aim to fit is the two-action sequence $[a_1, a_T]$ , based on the visual initial and goal states, $v_{start}$ and $v_{goal}$ . These conditional visual states are concatenated with the actions along the action feature dimension, forming a multi-dimensional array: + +$$ +\left[ \begin{array}{c c c c c} v _ {s} & 0 & \dots & 0 & v _ {g} \\ \hline a _ {1} & 0 & \dots & 0 & a _ {T} \end{array} \right] \tag {3} +$$ + +where the array is zero-padded to have a length corresponds to the planning horizon $T$ . During the denoising process, these conditional visual states can change, potentially misleading the learning process. To prevent this, a condition projection operation [54] is applied, ensuring the visual state and zero-padding dimensions remain unchanged (shaded below). The projection operation is denoted as: + +$$ +\left[ \begin{array}{c c c c c} \hat {v} _ {1} & \hat {v} _ {2} & \dots & \hat {v} _ {T - 1} & \hat {v} _ {T} \\ \hline \hat {a} _ {1} & \hat {a} _ {2} & \dots & \hat {a} _ {T - 1} & \hat {a} _ {T} \end{array} \right] \xrightarrow {\text {P r o j e c t i o n}} \left[ \begin{array}{c c c c c} v _ {s} & 0 & \dots & 0 & v _ {g} \\ \hat {a} _ {1} & 0 & \dots & 0 & \hat {a} _ {T} \end{array} \right] \tag {4} +$$ + +where $\hat{v}_t$ denotes the predicted visual state dimensions at timestamp $t$ within the planning horizon $T$ . + +# 3.2.2 Construct the Probabilistic Procedure Knowledge Graph $(\mathbf{P}^2\mathbf{KG})$ + +The Probabilistic Procedure Knowledge Graph [5] $\mathrm{P}^2\mathrm{KG} = (V,E,w)$ is a directed and weighted graph. In this structure, each step from the training set is represented as a node. During the graph construction process, we iterate over the training procedure plans, and for each direct step transition present in a plan, we add an edge from $a_{t}$ to $a_{t + 1}$ if it does not already exist in the graph; otherwise, we increase its existing frequency count by one. Eventually, this process + +results in a frequency-based Procedural Knowledge Graph (PKG) [62], which adeptly encapsulating the complexities of step sequencing in procedures and its potential variations, thereby addressing challenges (1) and (2) of procedure planning (ref. § 3.1.2). To further tackle challenge (3), this graph undergoes a transformation into a probabilistic format. In this transformed graph, the edges are not just connections but also signify the likelihoods of transitioning from one step to another. The weight of an edge from $a_{t}$ to $a_{t+1}$ is the count of transitions from action step $a_{t}$ to $a_{t+1}$ normalized by total count of $a_{t}$ being executed [5]. The normalization converts the frequency-based weight into probability distribution and the sum of all out-going edges is one. + +# 3.2.3 $\mathbf{P}^2\mathbf{K}\mathbf{G}$ -Enhanced Procedure Planning + +Retrieving Procedure Plan Recommendations from the $\mathbf{P}^2\mathbf{KG}$ . Humans use both previously-acquired knowledge and external knowledge when solving problems. The $\mathrm{P^2KG}$ provides extensive procedural knowledge, serving as a comprehensive textbook, particularly beneficial for the planning model that requires advanced skills. + +To utilize this procedural knowledge, queries are made to the $\mathrm{P}^2\mathrm{KG}$ using the first $(\hat{a}_1)$ and last $(\hat{a}_T)$ actions predicted by the step model. The aim is to find graph paths no longer than $T$ steps, starting from $\hat{a}_1$ and ending at $\hat{a}_T$ . This above process often results in multiple possible paths. To evaluate these paths, the probability of each is calculated by multiplying the probability weights of the edges along the path. For instance, the probability of a path $a_1 \to a_2 \to a_3$ is determined by the product $w_{a_1 \to a_2} \times w_{a_2 \to a_3}$ . These paths are then ranked according to their probabilities, and the top $R$ paths are selected as the recommended procedure plans from the $\mathrm{P}^2\mathrm{KG}$ , where $R$ is predefined. For paths shorter than $T$ , padding is applied at any point in the middle of the sequence to explore all possible resultant paths. When $R$ is greater than one, the top $R$ paths are aggregated through linear weighting into a single path (See section A.2 of supplementary material). This final path is then used as an additional input for the procedure planning model, thereby enhancing its decision-making process. + +Adopting Conditioned Projected Diffusion Model as the Planning Model. For the planning model, the conditional visual states and the procedure plan recommendation from the $\mathrm{P}^2\mathrm{KG}$ are concatenated with the actions along the action feature dimension, forming a multi-dimensional array: + +$$ +\left[ \begin{array}{c c c c c} v _ {s} & 0 & \dots & 0 & v _ {g} \\ \tilde {a} _ {1} & \tilde {a} _ {2} & \dots & \tilde {a} _ {T - 1} & \tilde {a} _ {T} \\ a _ {1} & a _ {2} & \dots & a _ {T - 1} & a _ {T} \end{array} \right] \tag {5} +$$ + +The rest process is similar to the step model, except that the project operation guarantees that three specific aspects remain unaltered—the dimensions of the visual state, $\mathrm{P^2KG}$ recommendation, and zero-padding. + +![](images/cb15eb0616a45f3cf40adc3ffa0a0df4e3e63387ee8938bc3173cb1df2159ca0.jpg) +Figure 3. Qualitative analysis of the 'Make Jello Shots' task + +![](images/aabd4cde419dca7cd6424f101e08a01862b15e077b3bc2d8a37c86a0123e34e2.jpg) +Figure 4. Qualitative analysis of the 'Change a Tire' task + +# 4. Experiments + +Datasets and implementation Details: In our evaluation, we employed datasets from three sources: CrossTask [63], COIN [52], and the Narrated Instructional Videos (NIV) [4]. See section C.4 of supplementary material for details on datasets. All ablation studies and analyses were conducted on CrossTask. We use two Tesla A100 GPUs for all the experiments. We chose horizon $T \in \{3, 4, 5, 6\}$ and $\mathrm{P^2KG}(R=1)$ condition for implementation. In some cases, we incorporate $\mathrm{P^2KG}(R=2)$ and LLM conditions which are indicated in the respective tables. Throughout this study, the $\mathrm{P^2KG}(R=1)$ is employed with a batch size of 256, unless explicitly stated otherwise. More implementation details are available in the section C.2 of the supplementary material. + +Evaluation Metrics and baselines: We use mean intersection over union (mIoU), mean accuracy (mAcc), and success rate (SR) as evaluation metrics. SR is the most stringent metric. See sec. C.4 of supplementary for more details. We compare our model with state-of-the-art methods: WLTDO [16], UAAA [1], UPN [49], DDN [9], PlaTe [50], Ext-GAIL [7], $\mathrm{P}^3\mathrm{IV}$ [59], PDPP [54], SkipPlan [29], and E3P [53]. More details of these methods are available in the section C.5 of supplementary material. Compared to other models, PDPP uses a different experimental setting. In PDPP, authors set the window after the start time of $a_1$ and before the end time of $a_T$ , contrary to the standard practice of setting a 2-second window around the start and end + +
ModelsRequired AnnotationsT=3T=4
step classvisual statesstep texttask classSR†mAcc†mIoU†SR†mAcc†mIoU†
Random✓<0.010.941.66<0.010.831.66
Retrieval-Based✓8.0523.332.063.9522.2236.97
WLTDO [16]✓✓1.8721.6431.700.7717.9226.43
UAAA [1]✓✓2.1520.2130.870.9819.8627.09
UPN [49]✓✓2.8924.3931.561.1921.5927.85
DDN [9]✓✓12.1831.2947.485.9727.1048.46
PlaTe [50]✓✓16.0036.1765.9114.0035.2955.36
Ext-GAIL wo Aug. [7]✓✓18.0143.8657.16---
Ext-GAIL [7]✓✓21.2749.4661.7016.4143.0560.93
P3IV♣[59]✓✓23.3449.9673.8913.4044.1670.01
PDPP♣[54]✓✓26.3855.6259.3418.6952.4462.38
E3P♣[53]✓✓✓26.4053.0274.0516.4948.0070.16
SkipPlan [29]♣✓28.8561.1874.9815.5655.6470.30
Ours w/ P2KG (R=2)✓22.6048.7653.5713.9045.7955.00
Ours♣w/ P2KG (R=1)✓33.3461.3664.1420.3855.5464.03
Ours♣w/ P2KG (R=2)✓33.3860.7963.8921.0256.0864.15
PDPP♣†[54]✓✓37.2064.6766.5721.4857.8265.13
Ours♣†w/ P2KG (R=1)✓38.1264.7467.1524.1559.0566.64
+ +Table 1. Performance of our method in comparison to existing baselines for CrossTask dataset. $\clubsuit$ means that the input visual features are from the S3D network [35] pretrained on HowTo100M [34]; otherwise, precomputed features provided in CrossTask are used. $\dagger$ indicates the results are under the PDPP's task setting, while others are under the conventional setting + +
ModelsT=5T=6
DDN [9]3.101.20
P³IV♣[59]7.214.40
PDPP♣[54]13.227.49
E3P♣[53]8.965.76
SkipPlan♣[29]8.555.12
Ours (R=2)8.175.32
Ours♣(R=1)13.258.09
Ours♣(R=2)12.749.23
PDPP♣†[54]13.458.41
Ours♣†(R=1)14.209.27
+ +Table 2. Success Rate $(SR^{\dagger})$ comparison to existing baselines for CrossTask dataset under longer horizons + +time (ref. [9]). We conduct experiments on both PDPP's proposed setting and the conventional setting. + +Inference: During the inference phase, the model receives only the start observation $v_{s}$ and the goal observation $v_{g}$ . To proceed, it utilizes a step model to predict the initial action $a_{1}$ and the end action $a_{T}$ for each data. Subsequently, leveraging the $\mathrm{P^2KG}$ , highest probable procedure knowledge graph plans connecting $a_{1}$ and $a_{T}$ are obtained. Then, a multi-dimensional array, is created as mentioned in Eq. 5. Finally, the planning model is used to predict the sequence of actions $[a_1,\dots,a_T]$ by denoising the generated multi-dimensional array as in § 3.2.3. + +# 4.1. Comparison with the State of the Art (SOTA) + +CrossTask (short horizon): We evaluate on CrossTask for short horizons ( $T = 3$ and $T = 4$ ). According to the results + +shown in Table 1, our proposed method outperforms the PDPP in PDPP's setting in every evaluation metric. More than $0.9\%$ and $2\%$ improvement in success rate in $T = 3$ and $T = 4$ respectively. In the conventional setting, our method with both $\mathrm{P}^2\mathrm{KG}$ ( $R = 1$ ) and $\mathrm{P}^2\mathrm{KG}$ ( $R = 2$ ) conditions outperform the success rate values by a significant margin compared to other baselines. $\mathrm{P}^2\mathrm{KG}$ ( $R = 2$ ) slightly outperforms $\mathrm{P}^2\mathrm{KG}$ ( $R = 1$ ), indicating potential benefits of incorporating more procedural knowledge from the $\mathrm{P}^2\mathrm{KG}$ . + +CrossTask (long horizon): We use long-horizon predictions for $T = 5$ and $T = 6$ for further evaluating our model as shown in Table 2. In PDPP's setting ( $\dagger$ ), our method improves the success rate in both $T = 5$ and $T = 6$ . In the conventional setting, our method utilizing $\mathrm{P}^2\mathrm{KG}$ ( $R = 1$ ) demonstrates the highest SR value for $T = 5$ , and for a longer horizon at $T = 6$ , our method delivers superior performance for $\mathrm{P}^2\mathrm{KG}$ ( $R = 2$ ). Our method performs well under the challenging scenario of a long planning horizon. Our success rate (SR) diminishes from approximately $40\%$ to $10\%$ when extending the planning horizon from $T = 3$ to $T = 6$ , primarily due to the heightened uncertainty surrounding the predicted plan between the initial and final steps. This uncertainty stems from the increase in the number of potential procedural plans within the $\mathrm{P}^2\mathrm{KG}$ . + +NIV and COIN: Results are shown in Table 3 and Table 4. On NIV, ours achieves the best result under the mIoU metric with $T = 3$ , and under both the SR and mIoU metrics with $T = 4$ . The results on NIV ( $T = 5$ , $T = 6$ ) are available in the supplementary material. For the COIN dataset we only + +report SR and mAcc due to space constraints; mIoU is reported in the supplementary material. Our method does not rank as the top performer on COIN when $T = 3$ or $T = 4$ . The likely reason is that the COIN dataset features just an average of 3.9 actions per video—a scenario that demands only short-horizon planning and does not necessitate advanced procedural knowledge (which encompasses long sequence-level knowledge [62]). Furthermore, the dataset's extensive collection of over 11k videos provides a substantial resource for baselines to learn basic procedural knowledge. + +
ModelsNIV (T=3)NIV (T=4)
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random2.214.076.091.122.735.84
DDN [9]18.4132.5456.5615.9727.0953.84
Ext-GAIL [7]22.1142.2065.9319.9136.3153.84
P3IV [59]24.6849.0174.2920.1438.3667.29
E3P [53]26.0551.2475.8121.3741.9674.90
PDPP [54]22.2239.5086.6621.3039.2484.96
Ours24.4443.4686.6722.7141.5991.49
+ +Table 3. Performance of baselines and ours for NIV dataset + +
ModelsCOIN (T=3)COIN (T=4)COIN (T=5)
SR↑mAcc↑SR↑mAcc↑SR↑mAcc↑
Random< 0.01< 0.01< 0.01< 0.01--
Retrieval4.3817.402.7114.29--
DDN [9]13.9020.1911.1317.71--
P3IV [59]15.4021.6711.3218.854.2710.81
E3P [53]19.5731.4213.5926.72--
PDPP [54]19.4243.4413.6742.5813.0243.36
SkipPlan [29]23.6547.1216.0443.199.9038.99
Ours (R=2)20.2539.8715.6339.5316.0640.72
+ +Table 4. Performance of baselines and ours for COIN dataset + +# 4.2. Ablation Studies and Analyses + +Ablation on the probabilistic procedure knowledge graph. We analyze the role of $\mathrm{P}^2\mathrm{KG}$ in improving the performance of our proposed method. Table 5 shows the results which clearly demonstrate that using $\mathrm{P}^2\mathrm{KG}$ conditions improves the performance significantly for every $T$ value. Especially when $T = 4$ , success rate (SR) improves more than $3\%$ and mean IoU improves more than $2\%$ . + +Plan recommendations provided by probabilistic procedure knowledge graph v.s. LLM. We recognize the recent trend of utilizing LLMs to enhance action anticipation [60] or planning in other realms [3, 27, 28, 40, 46]. In Table 6, we compare the results between using $\mathrm{P}^2\mathrm{KG}$ v.s. using LLM ('llama-2-13b-chat' and 'llama-2-70b-chat') to generate the plan recommendations. When examining Table 6, it becomes apparent that there are trade-offs between using LLM-generated recommendations and $\mathrm{P}^2\mathrm{KG}$ recommendations. For instance, $\mathrm{P}^2\mathrm{KG}$ recommendations are constrained by the data available in the training set, limiting their applicability to unseen procedural activities. On the + +other hand, LLMs tend to exhibit better generalization to such unseen activities. However, considering that the training and testing are conducted on the aforementioned three datasets with known activities, $\mathrm{P}^2\mathrm{KG}$ recommendations can yield more accurate results compared to relying on LLM-generated recommendations. + +Probabilistic procedure knowledge graph (P²KG) vs. Frequency-based procedure knowledge graph (PKG). The probabilistic procedure knowledge graph uses outedge normalization to encode step transition probabilities (§ 3.2.2), while the frequency-based procedure knowledge graph uses min-max normalization over the frequency counts throughout the graph. In both cases, the planning model only uses one procedure plan recommendation from the graph as condition in our experimental analysis. By looking at the results shown in Table 7, it is evident that the probabilistic procedure knowledge graph outperforms the frequency-based procedure knowledge graph. + +Effect of utilizing predicted steps for input conditions to train the procedure planing model. Our proposed problem decomposition allows training the planning model with ground truth (GT) first and last steps. We experiment with two ways to train the planning model. Method 1 uses the predicted start and end steps $(\hat{a}_1$ and $\hat{a}_T)$ as input to generate $\mathrm{P}^2\mathrm{KG}$ conditions and use them to train the planning model. Method 2 is where we augment the predicted start and end steps using the GT start and end steps $(a_1$ and $a_{T})$ by generating 3 more data samples as follows: $[\hat{a}_1,a_T]$ , $[a_1,\hat{a}_T]$ , and $[a_1,a_T]$ . Then we generate $\mathrm{P}^2\mathrm{KG}$ conditions for each data and train the model. From the results shown in Table 8, the method without GT data augmentation shows better results. This suggests that leveraging ground truth data in training can lead to worse performance in testing. + +Qualitative results. Figures 3 and 4 provide qualitative examples of our method. Intermediate steps are padded in the step model because it only predicts the start and end actions. In the 'make jello shot' task (see Figure 3), the model gives a wrong prediction in the intermediate steps when using $\mathrm{P}^2\mathrm{KG}$ ( $R = 1$ ) condition. However, it predicts correctly when using $\mathrm{P}^2\mathrm{KG}$ ( $R = 2$ ) conditions. In the 'change a tire' task shown in Figure 4, the model is able to predict all the intermediate steps in given conditions. + +Visualizations of the probabilistic procedure knowledge graph. We show a sub-graph from our probabilistic procedure knowledge graph (Figure 5). This graph is drawn around the 'jack up' node up to the depth of 2 nodes. + +Visualizations of the expert trajectories. Figure 6 illustrates the steps involved in completing the 'make jello shots' task, along with their transitions to other steps within the entire training data. This figure demonstrates that our $\mathrm{P}^2\mathrm{KG}$ encodes diverse sequencing possibilities for steps and also captures task-sharing steps across the entire training domain. For instance, 'pour water' is a step in 'make jello + +
ModelT=3T=4T=5T=6
SRmAccmIoUSRmAccmIoUSRmAccmIoUSRmAccmIoU
w.o P2KG conditions †35.6963.9166.0420.5257.4764.3912.853.4464.018.1550.4564.13
Ours †38.1264.7467.1524.1559.0566.6414.2053.8465.569.2750.2265.97
w.o P2KG conditions31.3559.5163.1118.9256.2062.4712.7151.2963.568.1647.6363.39
Ours33.3860.7963.8921.0256.0864.1512.7451.2363.169.2350.7865.56
+ +Table 5. Performance of our method with and without ${\mathrm{P}}^{2}\mathrm{{KG}}$ conditions on CrossTask ${}^{ \star }$ dataset + +
Model (T=6, CrossTask*)SRmAccmIoU
Ours with P2KG (R=1)
PDPP setting9.2750.2265.97
Conventional setting8.0950.8065.39
One LLM plan recommendation
PDPP setting (13b)7.7450.2864.05
Conventional setting (13b)7.2149.6863.89
PDPP setting (70b)8.6250.3164.34
Conventional setting (70b)7.8149.7564.02
P2KG (R=1) and one LLM plan recommendation
PDPP setting (13b)8.8149.9765.22
Conventional setting (13b)8.2051.4664.30
PDPP setting (70b)9.0150.2565.57
Conventional setting (70b)8.3451.5364.96
+ +Table 6. Performance of the plan recommendations provided by the probabilistic procedure knowledge graph v.s. LLM. + +
ModelsSRmAccmIoU
Frequency graph7.6648.6164.21
Probabilistic graph8.0950.8065.40
+ +Table 7. Performance comparison between probabilistic procedure knowledge graph v.s. frequency-based procedure knowledge graph for $T = 6$ on CrossTask dataset + +
ConditionSRmAccmIoU
without GT data aug.38.1264.7467.15
with GT data aug.32.4562.4262.80
+ +Table 8. Effect of different input conditions for performance on CrossTask ${}^{ \star }$ dataset $\left( {T = 3}\right)$ in PDPP’s setting + +
ModelsT=3T=4T=5T=6
ˆa1ˆaTˆa1ˆaTˆa1ˆaTˆa1ˆaT
Ours53.6950.6055.5652.5155.5851.8157.0951.92
Ours♣71.4263.3272.9863.3772.4263.2963.8259.96
+ +Table 9. The step model's start and end step prediction accuracies on the CrossTask dataset + +shots' task, but it can also be part of other tasks, leading to a step transition from 'pour water' to 'add fish.' This structure allows models to leverage rich procedural knowledge. + +Results for the step model. Table 9 reveals step model results, indicating potential enhancement areas to elevate the planning performance. + +Limitations & Failure cases. Our model exhibits three distinct failure case patterns. See section B.5 of the supplementary material for detailed discussions. + +![](images/db2da738b40c0a458f2503dbbf6960469d64d8ce24cb4392971c6f6a7d22ce5e.jpg) +Figure 5. Example of a sub-graph in our probabilistic procedure knowledge graph $(\mathrm{P}^2\mathrm{KG})$ for CrossTask dataset. This graph effectively encapsulates real-world knowledge of distinct transition probabilities between steps, e.g., the probability of transitioning from 'start loose' to 'jack up' is 0.65, in contrast to a mere 0.14 for the reverse transition—the $\mathrm{P}^2\mathrm{KG}$ reflects the common real-life practice where loosening the lug nuts before jacking up the car leads to a safer and more efficient tire change. + +![](images/fbfd42d59b3621b2bfa32ab374011a1653151f69d73f47d634e03eeb447e2bfb.jpg) +Figure 6. Expert trajectories of the 'Make Jello Shots' task, involving task-sharing steps and thus out-of-task step transitions. Thicker lines indicate paths that are more frequently visited + +# 5. Conclusion + +We focus on formulating procedural plans from an AI agent in instructional videos. We propose KEPP which employs a probabilistic procedural knowledge graph, sourced from the training domain, effectively serving as a 'textbook' for procedure planning. Results show that KEPP delivers SOTA performance with minimal supervision. Future work can focus on enhancing the accuracy of predictions for the initial and final steps. Additionally, our approach can be modified to aid in detecting erroneous steps and the misordering of steps in instructional videos [38, 41]. + +# References + +[1] Yazan Abu Farha and Juergen Gall. Uncertainty-aware anticipation of activities. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5, 6 +[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2 +[3] Anurag Ajay, Seungwook Han, Yilun Du, Shaung Li, Abhi Gupta, Tommi Jaakkola, Josh Tenenbaum, Leslie Kaelbling, Akash Srivastava, and Pulkit Agrawal. Compositional foundation models for hierarchical planning. arXiv preprint arXiv:2309.08587, 2023. 7 +[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4575-4583, 2016. 5 +[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. arXiv preprint arXiv:2307.08763, 2023. 2, 4, 5 +[6] Anonymous authors. Active procedure planning with uncertainty-awareness in instructional videos, 2023. Under review as a conference paper at ICLR 2024. https://openreview.net/pdf?id=JDd46WodYf.2 +[7] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 1, 2, 5, 6, 7 +[8] Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin, Jena D Hwang, Xiang Lorraine Li, Hirona J Arai, Soumya Sanyal, Keisuke Sakaguchi, Xiang Ren, and Yejin Choi. Plasma: Making small language models better procedural knowledge models for (counterfactual) planning. arXiv preprint arXiv:2305.19472, 2023. 2 +[9] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 1, 2, 3, 5, 6, 7 +[10] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where?-self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 2 +[11] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2447, 2023. 2 +[12] Hazel Doughty, Ivan Laptev, Walterio Mayol-Cuevas, and Dima Damen. Action modifiers: Learning from adverbs in + +instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 868-878, 2020. 2 +[13] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1, 2 +[14] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 2 +[15] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18952-18961, 2023. 2 +[16] Kiana Ehsani, Hessam Bagherinezhad, Joseph Redmon, Roozbeh Mottaghi, and Ali Farhadi. Who let the dogs out? modeling dog behavior from visual data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4051-4060, 2018. 5, 6 +[17] Sophie Fischer, Carlos Gemmell, Iain Mackie, and Jeffrey Dalton. Vilt: Video instructions linking for complex tasks. In Proceedings of the 2nd International Workshop on Interactive Multimedia Retrieval, pages 41-47, 2022. 2 +[18] Kevin Flanagan, Dima Damen, and Michael Wray. Learning temporal sentence grounding from narrated egovideos. arXiv preprint arXiv:2310.17395, 2023. 2 +[19] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. arXiv preprint arXiv:2005.03684, 2020. 2 +[20] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threedorld transport challenge: A visually guided task-and-motion planning benchmark for physically realistic embodied ai. arXiv preprint arXiv:2103.14025, 2021. 2 +[21] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 2 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 +[23] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. Machine Intelligence Research, 20 (2):249-262, 2023. 2 +[24] De-An Huang, Joseph J Lim, Li Fei-Fei, and Juan Carlos Niebles. Unsupervised visual-linguistic reference resolution in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2183–2192, 2017. 2 + +[25] De-An Huang, Shyamal Buch, Lucio Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding" it": Weakly-supervised reference-aware visual grounding in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5948-5957, 2018. 2 +[26] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2 +[27] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2, 7 +[28] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 2, 7 +[29] Zhiheng Li, Wenjia Geng, Muheng Li, Lei Chen, Yansong Tang, Jiwen Lu, and Jie Zhou. Skip-plan: Procedure planning in instructional videos via condensed action space learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10297-10306, 2023. 2, 3, 5, 6, 7 +[30] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In The Eleventh International Conference on Learning Representations, 2022. 2 +[31] Yujie Lu, Pan Lu, Zhiyu Chen, Wanrong Zhu, Xin Eric Wang, and William Yang Wang. Multimodal procedural planning via dual text-image prompting. arXiv preprint arXiv:2305.01795, 2023. 2 +[32] Weichao Mao, Ruta Desai, Michael Louis Iuzzolino, and Nitin Kamra. Action dynamics task graphs for learning plannable representations of procedural tasks. arXiv preprint arXiv:2302.05330, 2023. 2 +[33] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. arXiv preprint arXiv:2306.03802, 2023. 2 +[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 1, 6 +[35] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2, 6 +[36] Davide Moltisanti, Frank Keller, Hakan Bilen, and Laura Sevilla-Lara. Learning action changes by measuring verb-adverb textual relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23110-23118, 2023. 2 + +[37] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2 +[38] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 8 +[39] Megha Nawhal, Akash Abdu Jyothi, and Greg Mori. Rethinking learning approaches for long-term action anticipation. In European Conference on Computer Vision, pages 558-576. Springer, 2022. 2 +[40] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15302-15314, 2023. 7 +[41] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 8 +[42] Fadime Sener, Rishabh Saraf, and Angela Yao. Transferring knowledge from text to video: Zero-shot anticipation for procedural actions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[43] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2 +[44] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 2 +[45] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2 +[46] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 2, 7 +[47] Yale Song, Gene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2 +[48] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Look for the change: Learning object states and state-modifying actions from untrimmed + +web videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13956-13966, 2022. 2 +[49] Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal planning networks: Learning generalizable representations for visuomotor control. In International Conference on Machine Learning, pages 4732-4741. PMLR, 2018. 5, 6 +[50] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters, 7(2):4924-4930, 2022. 1, 2, 5, 6 +[51] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34:14476-14487, 2021. 2 +[52] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 5 +[53] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13565–13575, 2023. 1, 2, 3, 5, 6, 7 +[54] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14836-14845, 2023. 1, 2, 3, 4, 5, 6, 7 +[55] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 2 +[56] Yue Yang, Joongwon Kim, Artemis Panagopoulou, Mark Yatskar, and Chris Callison-Burch. Induce, edit, retrieve: Language grounded multimodal schema for instructional video retrieval. arXiv preprint arXiv:2111.09276, 2021. 2 +[57] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 2 +[58] Jiahao Zhang, Anoop Cherian, Yanbin Liu, Yizhak Ben-Shabat, Cristian Rodriguez, and Stephen Gould. Aligning step-by-step instructional diagrams to video demonstrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2483-2492, 2023. 2 +[59] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 1, 2, 3, 5, 6, 7 + +[60] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368, 2023. 7 +[61] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 2 +[62] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2, 5, 7 +[63] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 5 \ No newline at end of file diff --git a/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/images.zip b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94f08d8f5ac4c29131fb97a4bc617fb085d134d5 --- /dev/null +++ b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c9841306ec4a3aeeb326e19c775fa45df40fface2be4f4423b90478d13a8c5b +size 610885 diff --git a/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/layout.json b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5ae34115e68fabb873a321f5ed6a56a50679d60 --- /dev/null +++ b/2024/Why Not Use Your Textbook_ Knowledge-Enhanced Procedure Planning of Instructional Videos/layout.json @@ -0,0 +1,10307 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "text", + "content": "Why Not Use Your Textbook? Knowledge-Enhanced Procedure Planning of Instructional Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "spans": [ + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "text", + "content": "Kumaranage Ravindu Yasas Nagasinghe" + }, + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "text", + "content": " Honglu Zhou" + }, + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "text", + "content": " Malitha Gunawardhana" + }, + { + "bbox": [ + 73, + 160, + 520, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "text", + "content": "Martin Renqiang Min" + }, + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "text", + "content": " Daniel Harari" + }, + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "text", + "content": " Muhammad Haris Khan" + }, + { + "bbox": [ + 112, + 180, + 480, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 196, + 451, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 196, + 451, + 209 + ], + "spans": [ + { + "bbox": [ + 142, + 196, + 451, + 209 + ], + "type": "text", + "content": "1Mohamed bin Zayed University of Artificial Intelligence, 2NEC Laboratories, USA," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 210, + 402, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 210, + 402, + 222 + ], + "spans": [ + { + "bbox": [ + 192, + 210, + 402, + 222 + ], + "type": "text", + "content": "3 University of Auckland, 4Weizmann Institute of Science" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 225, + 455, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 225, + 455, + 236 + ], + "spans": [ + { + "bbox": [ + 129, + 225, + 455, + 236 + ], + "type": "text", + "content": "ravindu.nagasinghe@mbzuai.ac.ae, muhammad.haris@mbzuai.ac.ae" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 264, + 192, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 264, + 192, + 277 + ], + "spans": [ + { + "bbox": [ + 143, + 264, + 192, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 292, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 289, + 639 + ], + "type": "text", + "content": "In this paper, we explore the capability of an agent to construct a logical sequence of action steps, thereby assembling a strategic procedural plan. This plan is crucial for navigating from an initial visual observation to a target visual outcome, as depicted in real-life instructional videos. Existing works have attained partial success by extensively leveraging various sources of information available in the datasets, such as heavy intermediate visual observations, procedural names, or natural language step-by-step instructions, for features or supervision signals. However, the task remains formidable due to the implicit causal constraints in the sequencing of steps and the variability inherent in multiple feasible plans. To tackle these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities by infusing it with procedural knowledge. This knowledge, sourced from training procedure plans and structured as a directed weighted graph, equips the agent to better navigate the complexities of step sequencing and its potential variations. We coin our approach KEPP, a novel Knowledge-Enhanced Procedure Planning system, which harnesses a probabilistic procedural knowledge graph extracted from training data, effectively acting as a comprehensive textbook for the training domain. Experimental evaluations across three widely-used datasets under settings of varying complexity reveal that KEPP attains superior, state-of-the-art results while requiring only minimal supervision. Code and trained model are available at https://github.com/Ravindu-Yasas-Nagasinghe/KEPP" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 655, + 128, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 128, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 128, + 668 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "The evolution of the internet has precipitated an unprecedented influx of video content, serving as a vital educational resource for myriad learners. Individuals frequently" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 266, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 309, + 266, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 309, + 266, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 309, + 266, + 545, + 388 + ], + "type": "image", + "image_path": "0138b6fb7c6793c604dae03f21e39301720353eac987411ca960672a8fd230ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 399, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 521 + ], + "type": "text", + "content": "Figure 1. Expert trajectories [7] of the 'Make Jello Shots' task from the CrossTask dataset [63]. Heavier color indicates more frequently visited path. This depicts the complexities of the procedure planning task, arising from the subtle causal links in step sequencing (e.g., steps like 'stir mixture' or 'pour mixture' typically occur after adding individual ingredients), the varied probabilities of transitioning between steps, and the diversity in plans viable for a given starting point and an intended outcome. Motivated by these nuanced challenges, we propose Knowledge-Enhanced Procedure Planning (KEPP) with the use of a probabilistic procedure knowledge graph to capture and represent these intricacies" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": "leverage platforms such as YouTube to acquire new skills, ranging from culinary arts to automobile maintenance [34]. While these instructional videos benefit the development of intelligent agents in mastering long-horizon tasks, the challenge extends beyond merely interpreting visuals. It requires high-level reasoning and planning to effectively assist in complex, real-life scenarios [13]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Procedure Planning in Instructional Videos demands an agent to produce a sequence of actionable steps, thereby crafting a procedure plan that facilitates the transition from an initial visual observation of the physical world towards achieving a desired goal state [7, 9, 50, 53, 54, 59]. The task acts as a precursor to an envisioned future scenario in which an agent like a robot provides on-the-spot support, such as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18816" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 241, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 241, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 241, + 84 + ], + "type": "text", + "content": "assisting an individual in preparing a recipe [6]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 229 + ], + "type": "text", + "content": "Current methods in procedure planning in instructional videos make extensive use of various annotations available within the datasets to enrich input features or provide supervisory signals (see Table 1). These include detailed, temporally localized visual observations of intermediate action steps throughout the procedure plan [7, 9, 50], high-level procedural task labels [53, 54], and step-by-step instructions in natural language [53, 59]. Despite advancements, significant challenges persist, including characterizing the implicit causal constraints in step sequencing, the varied probabilities of transitioning between steps, and the inherent variability of multiple viable plans (see Fig. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 350 + ], + "type": "text", + "content": "To address these intricacies that previous efforts have overlooked, we propose to enhance the agent's capabilities in procedure planning by infusing it with comprehensive procedural knowledge [62], derived from training procedure plans and structured as a directed weighted graph. This graph, as a Probabilistic Procedural Knowledge Graph [5] where nodes denote steps from diverse tasks and edges represent step transition probabilities in the training domain, empowers agents to more adeptly navigate the complexities of step sequencing and its potential variations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "text", + "content": "Our proposed approach, KEPP, is a novel Knowledge-Enhanced Procedure Planning system (Fig. 2) that harnesses a probabilistic procedural knowledge graph " + }, + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "inline_equation", + "content": "(\\mathrm{P}^2\\mathrm{KG})" + }, + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "text", + "content": ", constructed from training procedure plans. This graph functions like a detailed textbook, providing extensive knowledge for the training domain, and thereby circumventing the need for costly multiple annotations required by existing methods. Additionally, we decompose the instructional video procedure planning problem into two parts: one driven by objectives specific to step perception and the other by a procedural knowledge-informed modeling of procedure planning. In this problem decomposition, the first and last action steps are predicted based on the initial and goal visual states. Following this, a procedure plan is crafted by leveraging the procedure plan recommendations retrieved from the " + }, + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "text", + "content": ". The recommendations correspond to the most probable procedure plans frequently used in training, conditioned on the predicted first and last action steps. In a similar vein to the approach by Li et al. [29], our proposed decomposition strategy reduces uncertainty by maximizing the use of currently available information, namely the initial and goal visual states. This allows for the improvement of procedure planning through more accurate predictions of start and end actions. Plus, this decomposition effectively incorporates procedural knowledge into procedure planning, thereby enhancing its effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 664, + 193, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 664, + 193, + 675 + ], + "spans": [ + { + "bbox": [ + 58, + 664, + 193, + 675 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "- We propose KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos that leverages rich procedural knowledge from a probabilistic procedu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "type": "text", + "content": "ral knowledge graph (P2KG). This approach necessitates only a minimal amount of annotations for supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 96, + 547, + 216 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 96, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 180 + ], + "type": "text", + "content": "- We decompose the problem in procedure planning of instructional videos: predicting the initial and final steps from the start and end visuals, and then creating a plan using procedural knowledge retrieved based on these predicted steps. This approach prioritizes the currently available information and effectively incorporates procedural knowledge, enhancing strategic planning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 180, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 547, + 216 + ], + "type": "text", + "content": "- Experimental evaluations on three widely-used datasets, under settings of varying complexity, reveal that KEPP attains state-of-the-art results in procedure planning." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 426 + ], + "type": "text", + "content": "Instructional Videos, which demonstrate multi-step procedures, have become a hotbed of research. The studies delve into various aspects, including comprehending and extracting intricate spatiotemporal content from video [12, 18, 19, 21, 23, 36, 43, 44, 48, 55, 57], interpreting the interrelationships between various actions and procedural events [47, 63], and developing capabilities for forecasting [39, 42] and strategic reasoning and planning [28] within the context of these videos. Furthermore, by leveraging the multimodality of visual, auditory, and narrative elements within these videos, research extends to areas like multimodal alignment [2, 58], grounding [10, 14, 25, 33, 51], representation learning [11, 35, 61], pre-training [15, 26, 62], and more [17, 24, 37, 56]. This paper focuses on procedure planning in instructional videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 617 + ], + "type": "text", + "content": "Procedure Planning is a vital skill for autonomous agents tasked with handling complex activities in everyday settings. Essentially, these agents must discern the appropriate actions to reach a specific goal. This aspect of artificial intelligence (AI) has been a prominent and integral subject in fields like robotics [20, 28, 32, 45, 46]. Yet, the challenge of procedure planning in the context of instructional videos is notably distinct, and potentially more complex, than its counterparts in natural language processing [8, 30], multimodal generative AI [13, 31], and simulated environments [27, 28, 45]. Its significance is underscored by the need for planning that is grounded in real-world scenes. This requires the development of AI agents capable of accurately perceiving and understanding the current real-world context, and then anticipating and mapping out a logical sequence of actions to fulfill a high-level goal effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Procedure Planning in Instructional Videos has recently garnered research attention. DDN [9] initiates this trend by conceptualizing the problem as sequential latent space planning. Building on this, PlaTe [50] employs transformers for both action and state models, integrating Beam Search for enhanced performance. Meanwhile, Ext-GAIL [7] suggests employing contextual modeling through variational autoencoder and adversarial policy learning. This method" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18817" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "considers contextual information as time-invariant knowledge, crucial for distinguishing specific tasks and allowing for multiple planning outcomes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "type": "text", + "content": "While these earlier approaches have viewed procedure planning as an autoregressive sequence generation problem, recent methods regard it as a distribution-fitting problem to mitigate error propagation in sequential decisions. In this vein, " + }, + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^3\\mathrm{IV}" + }, + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "type": "text", + "content": " [59] replaces intermediate visual states with linguistic representations for supervision, predicting all steps simultaneously instead of using autoregressive methods. To circumvent the complex learning strategies and high annotation costs of previous work, PDPP [54] models the entire intermediate action sequence distribution using a conditioned projected diffusion model. This approach redefines the planning problem as a sampling process from this distribution and simplifies supervision by using only instructional video task labels. E3P [53], also encoding task information, adopts a mask-and-predict strategy for mining step relationships in procedural tasks, integrating probabilistic masking for regularization. In contrast, our approach does not rely on annotations of intermediate states, natural language step representations, or procedural task labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 336, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 468 + ], + "type": "text", + "content": "Recognizing the difficulties inherent in high dimensional state supervision and the accumulation of errors in action sequences, SkipPlan [29] was developed. It strategically focuses on action predictions, breaking down longer sequences into shorter, more manageable sub-chains by skipping over less reliable intermediate actions. Drawing inspiration from SkipPlan, our approach decomposes the procedure planning problem to prioritize the most reliable information available (ref. § 3.1.2). However, we innovate further by incorporating a Probabilistic Procedure Knowledge Graph, significantly enriching the planning phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 479, + 129, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 129, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 129, + 493 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 499, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 536 + ], + "type": "text", + "content": "We will first introduce the problem setup in § 3.1, and then discuss our novel Knowledge-Enhanced Procedure Planning system (KEPP) in § 3.2. See Fig. 2 for KEPP overview." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 544, + 217, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 217, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 217, + 555 + ], + "type": "text", + "content": "3.1. Problem and Method Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 173, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 173, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 173, + 573 + ], + "type": "text", + "content": "3.1.1 Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "We follow the problem definition for procedure planning of instructional videos put forth by Chang et al. [9]: given an observation of the initial state " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " and a goal state " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", both are short video clips indicating different states of the real-world environment extracted from an instructional video, a model is required to plan a sequence of action steps " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "a_{1:T}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " to reach the indicated goal. Here, " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " is the planning horizon, inputting to the model, corresponding to the number of action steps in the sequence produced by the model so that the environment state can be transformed from " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " to denote the action step at the timestamp" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", and in the following, " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v_{g}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " are short for " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". Mathematically, the procedure planning problem is defined as " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "p(a_{1:T}|v_s,v_g)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " that denotes the conditional probability distribution of the action sequence " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "a_{1:T}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " given the initial visual observation " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and the goal visual state " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 151, + 441, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 151, + 441, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 151, + 441, + 163 + ], + "type": "text", + "content": "3.1.2 Problem Decomposition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 174, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 270 + ], + "type": "text", + "content": "Considering the initial and final visual states are input, providing the most reliable information, we hypothesize that predicting the first and final action steps is more dependable than interpolating the intermediate ones, and consequently, an enhanced accuracy in predicting the first and final steps can lead to more effective procedure planning. Inspired by this hypothesis, we decompose the procedure planning problem into two sub-problems, as shown in Eq. 1:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 285, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 285, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 310, + 285, + 545, + 298 + ], + "type": "interline_equation", + "content": "p \\left(\\hat {a} _ {1: T} | v _ {s}, v _ {g}\\right) = p \\left(\\hat {a} _ {2: T - 1} | \\hat {a} _ {1}, \\hat {a} _ {T}\\right) p \\left(\\hat {a} _ {1}, \\hat {a} _ {T} | v _ {s}, v _ {g}\\right), \\tag {1}", + "image_path": "feb18c394cd39730d7d8db027221c4f66e3342658521bf4ca0a0fbd313307001.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": "where the first sub-problem is to identify the beginning step " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": " and the end step " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "a_T" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": ", and the second sub-problem is to plan the intermediate action steps " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "a_{2:T-1}" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "a_T" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\hat{a}_t" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": " to denote predicted action step at timestamp " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "type": "text", + "content": "Our proposed problem decomposition in Eq. 1 bears resemblance with the problem formulation from Li et al. [29]; they decompose procedure planning into " + }, + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "type": "inline_equation", + "content": "p(\\hat{a}_{1:T}|v_s, v_g) = \\prod_{t=2}^{T-1} p(\\hat{a}_t|\\hat{a}_1, \\hat{a}_T)p(\\hat{a}_1, \\hat{a}_T|v_s, v_g)" + }, + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "type": "text", + "content": ". However, our formulation differs in its approach to modeling the second subproblem. Specifically, we employ a conditioned projected diffusion model (ref. § 3.2) to jointly predict " + }, + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "type": "inline_equation", + "content": "a_{2:T-1}" + }, + { + "bbox": [ + 304, + 362, + 545, + 496 + ], + "type": "text", + "content": " at once, whereas Li et al. [29] rely on Transformer decoders to predict each intermediate action independently. Further, we integrate a Probabilistic Procedure Knowledge Graph (ref. § 3.2.2) to address the second sub-problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "content": "Tackling the second sub-problem is nontrivial even when armed with an oracle predictor for the first sub-problem. Procedure planning in real-life scenarios remains daunting because of the following challenges: (1) the presence of implicit temporal and causal constraints in the sequencing of steps, (2) the existence of numerous viable plans given an initial state and a goal state, and (3) the need to incorporate the real-life everyday knowledge both in task-sharing steps and in managing the inherent variability in transition probabilities between steps. Previous studies tackled these challenges by extensively harnessing detailed annotations in the datasets to augment input features or offer supervision signals (see Table 1). In contrast, we propose harnessing a Probabilistic Procedural Knowledge Graph " + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{P}^2\\mathrm{KG})" + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "content": " which is extracted from the procedure plans in the training set. With the " + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "content": " at our hand, we further decompose the procedure planning problem to reduce its complexity and learn " + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}:(v_s,v_g,T)\\to p(\\hat{a}_{1:T}|v_s,v_g)" + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18818" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 63, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 63, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 63, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 63, + 545, + 171 + ], + "type": "image", + "image_path": "1cf3937d5a075be8ca996915459c39187231a319a4a72393896a35b18cb5fe87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "type": "text", + "content": "Figure 2. Overview of our methodology. We introduce KEPP, a Knowledge-Enhanced Procedure Planning system for instructional videos, leveraging a Probabilistic Procedural Knowledge Graph " + }, + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "type": "inline_equation", + "content": "(\\mathrm{P}^2\\mathrm{KG})" + }, + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "type": "text", + "content": ". KEPP breaks down procedure planning into two parts: predicting initial and final steps from visual states, and crafting a procedure plan based on the procedural knowledge retrieved from " + }, + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 177, + 547, + 222 + ], + "type": "text", + "content": ", conditioned on the predicted first and last action steps. KEPP requires minimal annotations and enhances planning effectiveness" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 261, + 287, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 261, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 59, + 261, + 287, + 296 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(\\hat {a} _ {1: T} | v _ {s}, v _ {g}\\right) = \\\\ p \\left(\\hat {a} _ {1: T} | \\tilde {a} _ {1: T}, v _ {s}, v _ {g}\\right) p \\left(\\tilde {a} _ {1: T} | \\hat {a} _ {1}, \\hat {a} _ {T}\\right) p \\left(\\hat {a} _ {1}, \\hat {a} _ {T} | v _ {s}, v _ {g}\\right) \\tag {2} \\\\ \\end{array}", + "image_path": "5790fcc646f5633e85f9ad1de0591991e5dbdd08e9d4c59cb9908b2a96b5a147.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "text", + "content": " denotes the machine learning model, and " + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\tilde{a}_{1:T}" + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "text", + "content": " represents a graph path (i.e., a sequence of nodes) retrieved from " + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "text", + "content": ". This graph path provides a valuable procedure plan recommendation aligned with the training domain, thus mitigating the complexity of procedure planning. It is worth noting that the proposed approach to modeling procedure planning using Eq. 2 demands only a minimal level of supervision, merely relying on the ground truth training procedure plan; Eq. 2 circumvents the need for additional annotations. We describe details of our " + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 297, + 287, + 428 + ], + "type": "text", + "content": "-enhanced approach in the following subsection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 287, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 460 + ], + "type": "text", + "content": "3.2. KEPP: Knowledge-Enhanced Procedure Planning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "type": "text", + "content": "We propose KEPP (Fig. 2) utilizing a probabilistic procedure knowledge graph extracted from the training set. We firstly identify the beginning and conclusion steps according to the input initial and goal states; and then, conditioned on these steps and the planning horizon " + }, + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "type": "text", + "content": ", we query the graph to retrieve relevant procedural knowledge for knowledge-enhanced procedure planning of instructional videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 254, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 254, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 254, + 575 + ], + "type": "text", + "content": "3.2.1 Identify Beginning and Conclusion Steps" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "content": " as input, we adapt a Conditioned Projected Diffusion Model [54] (ref. supplementary material) to identify the first action step and the final step; we refer to this model as the 'Step (Perception) Model'." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Standard Denoising Diffusion Probabilistic Model tackles data generation through a denoising Markov chain over variables " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{x_{N}\\ldots x_{0}\\}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", starting with " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_{N}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " as a Gaussian random distribution [22]. In the forward diffusion phase, Gaussian noise " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " is progressively added to the initial, unaltered data " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", transforming it into a Gaussian random distribution. Conversely, the reverse denoising process" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "type": "text", + "content": "transforms Gaussian noise back into a sample. Denoising is parameterized by a learnable noise prediction model, and the learning objective is to learn the noise added to " + }, + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "type": "text", + "content": " at each diffusion step. After training, the diffusion model generates data akin to " + }, + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 242, + 545, + 313 + ], + "type": "text", + "content": " by iteratively applying the denoising process, starting from random Gaussian noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "text", + "content": "Adopting Conditioned Projected Diffusion Model as the Step Model. For our step model, the distribution we aim to fit is the two-action sequence " + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "inline_equation", + "content": "[a_1, a_T]" + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "text", + "content": ", based on the visual initial and goal states, " + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 304, + 314, + 546, + 387 + ], + "type": "text", + "content": ". These conditional visual states are concatenated with the actions along the action feature dimension, forming a multi-dimensional array:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 374, + 396, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 396, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 374, + 396, + 545, + 422 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\hline a _ {1} & 0 & \\dots & 0 & a _ {T} \\end{array} \\right] \\tag {3}", + "image_path": "d7d9a696f06f6a06bf70f409d67db484735f87aa214d5e063575eaa8387fb35e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": "where the array is zero-padded to have a length corresponds to the planning horizon " + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": ". During the denoising process, these conditional visual states can change, potentially misleading the learning process. To prevent this, a condition projection operation [54] is applied, ensuring the visual state and zero-padding dimensions remain unchanged (shaded below). The projection operation is denoted as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 516, + 555, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 516, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 516, + 555, + 545 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c c c c c} \\hat {v} _ {1} & \\hat {v} _ {2} & \\dots & \\hat {v} _ {T - 1} & \\hat {v} _ {T} \\\\ \\hline \\hat {a} _ {1} & \\hat {a} _ {2} & \\dots & \\hat {a} _ {T - 1} & \\hat {a} _ {T} \\end{array} \\right] \\xrightarrow {\\text {P r o j e c t i o n}} \\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\hat {a} _ {1} & 0 & \\dots & 0 & \\hat {a} _ {T} \\end{array} \\right] \\tag {4}", + "image_path": "ff56fde8b83568f9100f33ced7ee74673d1220161dc5d74f04c9734e81204cef.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\hat{v}_t" + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "text", + "content": " denotes the predicted visual state dimensions at timestamp " + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "text", + "content": " within the planning horizon " + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 546, + 545, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 586, + 545, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 611 + ], + "type": "text", + "content": "3.2.2 Construct the Probabilistic Procedure Knowledge Graph " + }, + { + "bbox": [ + 305, + 586, + 545, + 611 + ], + "type": "inline_equation", + "content": "(\\mathbf{P}^2\\mathbf{KG})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "The Probabilistic Procedure Knowledge Graph [5] " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG} = (V,E,w)" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": " is a directed and weighted graph. In this structure, each step from the training set is represented as a node. During the graph construction process, we iterate over the training procedure plans, and for each direct step transition present in a plan, we add an edge from " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "a_{t + 1}" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": " if it does not already exist in the graph; otherwise, we increase its existing frequency count by one. Eventually, this process" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18819" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "results in a frequency-based Procedural Knowledge Graph (PKG) [62], which adeptly encapsulating the complexities of step sequencing in procedures and its potential variations, thereby addressing challenges (1) and (2) of procedure planning (ref. § 3.1.2). To further tackle challenge (3), this graph undergoes a transformation into a probabilistic format. In this transformed graph, the edges are not just connections but also signify the likelihoods of transitioning from one step to another. The weight of an edge from " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "a_{t+1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " is the count of transitions from action step " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "a_{t+1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " normalized by total count of " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " being executed [5]. The normalization converts the frequency-based weight into probability distribution and the sum of all out-going edges is one." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 241, + 241, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 241, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 241, + 255 + ], + "type": "text", + "content": "3.2.3 " + }, + { + "bbox": [ + 47, + 241, + 241, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^2\\mathbf{K}\\mathbf{G}" + }, + { + "bbox": [ + 47, + 241, + 241, + 255 + ], + "type": "text", + "content": " -Enhanced Procedure Planning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": "Retrieving Procedure Plan Recommendations from the " + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^2\\mathbf{KG}" + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": ". Humans use both previously-acquired knowledge and external knowledge when solving problems. The " + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{P^2KG}" + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": " provides extensive procedural knowledge, serving as a comprehensive textbook, particularly beneficial for the planning model that requires advanced skills." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": "To utilize this procedural knowledge, queries are made to the " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " using the first " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "(\\hat{a}_1)" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " and last " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "(\\hat{a}_T)" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " actions predicted by the step model. The aim is to find graph paths no longer than " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " steps, starting from " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\hat{a}_1" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " and ending at " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\hat{a}_T" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": ". This above process often results in multiple possible paths. To evaluate these paths, the probability of each is calculated by multiplying the probability weights of the edges along the path. For instance, the probability of a path " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "a_1 \\to a_2 \\to a_3" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " is determined by the product " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "w_{a_1 \\to a_2} \\times w_{a_2 \\to a_3}" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": ". These paths are then ranked according to their probabilities, and the top " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " paths are selected as the recommended procedure plans from the " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " is predefined. For paths shorter than " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": ", padding is applied at any point in the middle of the sequence to explore all possible resultant paths. When " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " is greater than one, the top " + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 332, + 288, + 559 + ], + "type": "text", + "content": " paths are aggregated through linear weighting into a single path (See section A.2 of supplementary material). This final path is then used as an additional input for the procedure planning model, thereby enhancing its decision-making process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 559, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 619 + ], + "type": "text", + "content": "Adopting Conditioned Projected Diffusion Model as the Planning Model. For the planning model, the conditional visual states and the procedure plan recommendation from the " + }, + { + "bbox": [ + 46, + 559, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 559, + 287, + 619 + ], + "type": "text", + "content": " are concatenated with the actions along the action feature dimension, forming a multi-dimensional array:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 623, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 623, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 623, + 287, + 662 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c c c c c} v _ {s} & 0 & \\dots & 0 & v _ {g} \\\\ \\tilde {a} _ {1} & \\tilde {a} _ {2} & \\dots & \\tilde {a} _ {T - 1} & \\tilde {a} _ {T} \\\\ a _ {1} & a _ {2} & \\dots & a _ {T - 1} & a _ {T} \\end{array} \\right] \\tag {5}", + "image_path": "fe3cf579d8954df64a40d635ea33e6569617bcb87f0c323da882da9c992cf5fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "The rest process is similar to the step model, except that the project operation guarantees that three specific aspects remain unaltered—the dimensions of the visual state, " + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{P^2KG}" + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": " recommendation, and zero-padding." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 543, + 196 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 543, + 196 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 543, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 543, + 196 + ], + "type": "image", + "image_path": "cb15eb0616a45f3cf40adc3ffa0a0df4e3e63387ee8938bc3173cb1df2159ca0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 201, + 535, + 213 + ], + "lines": [ + { + "bbox": [ + 315, + 201, + 535, + 213 + ], + "spans": [ + { + "bbox": [ + 315, + 201, + 535, + 213 + ], + "type": "text", + "content": "Figure 3. Qualitative analysis of the 'Make Jello Shots' task" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 217, + 543, + 338 + ], + "blocks": [ + { + "bbox": [ + 306, + 217, + 543, + 338 + ], + "lines": [ + { + "bbox": [ + 306, + 217, + 543, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 543, + 338 + ], + "type": "image", + "image_path": "aabd4cde419dca7cd6424f101e08a01862b15e077b3bc2d8a37c86a0123e34e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 344, + 529, + 355 + ], + "lines": [ + { + "bbox": [ + 321, + 344, + 529, + 355 + ], + "spans": [ + { + "bbox": [ + 321, + 344, + 529, + 355 + ], + "type": "text", + "content": "Figure 4. Qualitative analysis of the 'Change a Tire' task" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 369, + 388, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 388, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 388, + 382 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "text", + "content": "Datasets and implementation Details: In our evaluation, we employed datasets from three sources: CrossTask [63], COIN [52], and the Narrated Instructional Videos (NIV) [4]. See section C.4 of supplementary material for details on datasets. All ablation studies and analyses were conducted on CrossTask. We use two Tesla A100 GPUs for all the experiments. We chose horizon " + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "inline_equation", + "content": "T \\in \\{3, 4, 5, 6\\}" + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{P^2KG}(R=1)" + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "text", + "content": " condition for implementation. In some cases, we incorporate " + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{P^2KG}(R=2)" + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "text", + "content": " and LLM conditions which are indicated in the respective tables. Throughout this study, the " + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{P^2KG}(R=1)" + }, + { + "bbox": [ + 304, + 389, + 545, + 555 + ], + "type": "text", + "content": " is employed with a batch size of 256, unless explicitly stated otherwise. More implementation details are available in the section C.2 of the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Evaluation Metrics and baselines: We use mean intersection over union (mIoU), mean accuracy (mAcc), and success rate (SR) as evaluation metrics. SR is the most stringent metric. See sec. C.4 of supplementary for more details. We compare our model with state-of-the-art methods: WLTDO [16], UAAA [1], UPN [49], DDN [9], PlaTe [50], Ext-GAIL [7], " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^3\\mathrm{IV}" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " [59], PDPP [54], SkipPlan [29], and E3P [53]. More details of these methods are available in the section C.5 of supplementary material. Compared to other models, PDPP uses a different experimental setting. In PDPP, authors set the window after the start time of " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " and before the end time of " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "a_T" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": ", contrary to the standard practice of setting a 2-second window around the start and end" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18820" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 529, + 300 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 529, + 300 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 529, + 300 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 529, + 300 + ], + "type": "table", + "html": "
ModelsRequired AnnotationsT=3T=4
step classvisual statesstep texttask classSR†mAcc†mIoU†SR†mAcc†mIoU†
Random✓<0.010.941.66<0.010.831.66
Retrieval-Based✓8.0523.332.063.9522.2236.97
WLTDO [16]✓✓1.8721.6431.700.7717.9226.43
UAAA [1]✓✓2.1520.2130.870.9819.8627.09
UPN [49]✓✓2.8924.3931.561.1921.5927.85
DDN [9]✓✓12.1831.2947.485.9727.1048.46
PlaTe [50]✓✓16.0036.1765.9114.0035.2955.36
Ext-GAIL wo Aug. [7]✓✓18.0143.8657.16---
Ext-GAIL [7]✓✓21.2749.4661.7016.4143.0560.93
P3IV♣[59]✓✓23.3449.9673.8913.4044.1670.01
PDPP♣[54]✓✓26.3855.6259.3418.6952.4462.38
E3P♣[53]✓✓✓26.4053.0274.0516.4948.0070.16
SkipPlan [29]♣✓28.8561.1874.9815.5655.6470.30
Ours w/ P2KG (R=2)✓22.6048.7653.5713.9045.7955.00
Ours♣w/ P2KG (R=1)✓33.3461.3664.1420.3855.5464.03
Ours♣w/ P2KG (R=2)✓33.3860.7963.8921.0256.0864.15
PDPP♣†[54]✓✓37.2064.6766.5721.4857.8265.13
Ours♣†w/ P2KG (R=1)✓38.1264.7467.1524.1559.0566.64
", + "image_path": "0823832d4a0e30d693d64fa7de959a2bcb9b85587522595910784a1d14056a8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 103, + 350, + 233, + 479 + ], + "blocks": [ + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": "Table 1. Performance of our method in comparison to existing baselines for CrossTask dataset. " + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": " means that the input visual features are from the S3D network [35] pretrained on HowTo100M [34]; otherwise, precomputed features provided in CrossTask are used. " + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": " indicates the results are under the PDPP's task setting, while others are under the conventional setting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 103, + 350, + 233, + 479 + ], + "lines": [ + { + "bbox": [ + 103, + 350, + 233, + 479 + ], + "spans": [ + { + "bbox": [ + 103, + 350, + 233, + 479 + ], + "type": "table", + "html": "
ModelsT=5T=6
DDN [9]3.101.20
P³IV♣[59]7.214.40
PDPP♣[54]13.227.49
E3P♣[53]8.965.76
SkipPlan♣[29]8.555.12
Ours (R=2)8.175.32
Ours♣(R=1)13.258.09
Ours♣(R=2)12.749.23
PDPP♣†[54]13.458.41
Ours♣†(R=1)14.209.27
", + "image_path": "b74b42370ed4d13becaa78082bc91840f8e1aa032b2a38e77d9ca0d220752c04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "type": "text", + "content": "Table 2. Success Rate " + }, + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "type": "inline_equation", + "content": "(SR^{\\dagger})" + }, + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "type": "text", + "content": " comparison to existing baselines for CrossTask dataset under longer horizons" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "content": "time (ref. [9]). We conduct experiments on both PDPP's proposed setting and the conventional setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": "Inference: During the inference phase, the model receives only the start observation " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": " and the goal observation " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "v_{g}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": ". To proceed, it utilizes a step model to predict the initial action " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "a_{1}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": " and the end action " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "a_{T}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": " for each data. Subsequently, leveraging the " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{P^2KG}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": ", highest probable procedure knowledge graph plans connecting " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "a_{1}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "a_{T}" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": " are obtained. Then, a multi-dimensional array, is created as mentioned in Eq. 5. Finally, the planning model is used to predict the sequence of actions " + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "inline_equation", + "content": "[a_1,\\dots,a_T]" + }, + { + "bbox": [ + 46, + 540, + 288, + 661 + ], + "type": "text", + "content": " by denoising the generated multi-dimensional array as in § 3.2.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 284, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 284, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 284, + 685 + ], + "type": "text", + "content": "4.1. Comparison with the State of the Art (SOTA)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "CrossTask (short horizon): We evaluate on CrossTask for short horizons (" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "). According to the results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": "shown in Table 1, our proposed method outperforms the PDPP in PDPP's setting in every evaluation metric. More than " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " improvement in success rate in " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " respectively. In the conventional setting, our method with both " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "R = 1" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": ") conditions outperform the success rate values by a significant margin compared to other baselines. " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": ") slightly outperforms " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "R = 1" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": "), indicating potential benefits of incorporating more procedural knowledge from the " + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 362, + 545, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": "CrossTask (long horizon): We use long-horizon predictions for " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " for further evaluating our model as shown in Table 2. In PDPP's setting (" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": "), our method improves the success rate in both " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": ". In the conventional setting, our method utilizing " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "R = 1" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": ") demonstrates the highest SR value for " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": ", and for a longer horizon at " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": ", our method delivers superior performance for " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": "). Our method performs well under the challenging scenario of a long planning horizon. Our success rate (SR) diminishes from approximately " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " when extending the planning horizon from " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": ", primarily due to the heightened uncertainty surrounding the predicted plan between the initial and final steps. This uncertainty stems from the increase in the number of potential procedural plans within the " + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 472, + 546, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "NIV and COIN: Results are shown in Table 3 and Table 4. On NIV, ours achieves the best result under the mIoU metric with " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", and under both the SR and mIoU metrics with " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ". The results on NIV (" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ") are available in the supplementary material. For the COIN dataset we only" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "18821" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "report SR and mAcc due to space constraints; mIoU is reported in the supplementary material. Our method does not rank as the top performer on COIN when " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ". The likely reason is that the COIN dataset features just an average of 3.9 actions per video—a scenario that demands only short-horizon planning and does not necessitate advanced procedural knowledge (which encompasses long sequence-level knowledge [62]). Furthermore, the dataset's extensive collection of over 11k videos provides a substantial resource for baselines to learn basic procedural knowledge." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 201, + 284, + 291 + ], + "blocks": [ + { + "bbox": [ + 51, + 201, + 284, + 291 + ], + "lines": [ + { + "bbox": [ + 51, + 201, + 284, + 291 + ], + "spans": [ + { + "bbox": [ + 51, + 201, + 284, + 291 + ], + "type": "table", + "html": "
ModelsNIV (T=3)NIV (T=4)
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random2.214.076.091.122.735.84
DDN [9]18.4132.5456.5615.9727.0953.84
Ext-GAIL [7]22.1142.2065.9319.9136.3153.84
P3IV [59]24.6849.0174.2920.1438.3667.29
E3P [53]26.0551.2475.8121.3741.9674.90
PDPP [54]22.2239.5086.6621.3039.2484.96
Ours24.4443.4686.6722.7141.5991.49
", + "image_path": "e293e43bd1632feec9002ebcc58f1a7e5b3f7dee451b07e4616b7a832cfa2a00.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 328, + 284, + 429 + ], + "blocks": [ + { + "bbox": [ + 59, + 300, + 276, + 310 + ], + "lines": [ + { + "bbox": [ + 59, + 300, + 276, + 310 + ], + "spans": [ + { + "bbox": [ + 59, + 300, + 276, + 310 + ], + "type": "text", + "content": "Table 3. Performance of baselines and ours for NIV dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 328, + 284, + 429 + ], + "lines": [ + { + "bbox": [ + 50, + 328, + 284, + 429 + ], + "spans": [ + { + "bbox": [ + 50, + 328, + 284, + 429 + ], + "type": "table", + "html": "
ModelsCOIN (T=3)COIN (T=4)COIN (T=5)
SR↑mAcc↑SR↑mAcc↑SR↑mAcc↑
Random< 0.01< 0.01< 0.01< 0.01--
Retrieval4.3817.402.7114.29--
DDN [9]13.9020.1911.1317.71--
P3IV [59]15.4021.6711.3218.854.2710.81
E3P [53]19.5731.4213.5926.72--
PDPP [54]19.4243.4413.6742.5813.0243.36
SkipPlan [29]23.6547.1216.0443.199.9038.99
Ours (R=2)20.2539.8715.6339.5316.0640.72
", + "image_path": "549e899db5a95361f90df1d49521e51e75cca7dabbb533787a4599467ea7c749.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 279, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 279, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 279, + 447 + ], + "type": "text", + "content": "Table 4. Performance of baselines and ours for COIN dataset" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 468, + 212, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 212, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 212, + 480 + ], + "type": "text", + "content": "4.2. Ablation Studies and Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": "Ablation on the probabilistic procedure knowledge graph. We analyze the role of " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " in improving the performance of our proposed method. Table 5 shows the results which clearly demonstrate that using " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " conditions improves the performance significantly for every " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " value. Especially when " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": ", success rate (SR) improves more than " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " and mean IoU improves more than " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Plan recommendations provided by probabilistic procedure knowledge graph v.s. LLM. We recognize the recent trend of utilizing LLMs to enhance action anticipation [60] or planning in other realms [3, 27, 28, 40, 46]. In Table 6, we compare the results between using " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " v.s. using LLM ('llama-2-13b-chat' and 'llama-2-70b-chat') to generate the plan recommendations. When examining Table 6, it becomes apparent that there are trade-offs between using LLM-generated recommendations and " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " recommendations. For instance, " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " recommendations are constrained by the data available in the training set, limiting their applicability to unseen procedural activities. On the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "other hand, LLMs tend to exhibit better generalization to such unseen activities. However, considering that the training and testing are conducted on the aforementioned three datasets with known activities, " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " recommendations can yield more accurate results compared to relying on LLM-generated recommendations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 145, + 546, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 546, + 289 + ], + "type": "text", + "content": "Probabilistic procedure knowledge graph (P²KG) vs. Frequency-based procedure knowledge graph (PKG). The probabilistic procedure knowledge graph uses outedge normalization to encode step transition probabilities (§ 3.2.2), while the frequency-based procedure knowledge graph uses min-max normalization over the frequency counts throughout the graph. In both cases, the planning model only uses one procedure plan recommendation from the graph as condition in our experimental analysis. By looking at the results shown in Table 7, it is evident that the probabilistic procedure knowledge graph outperforms the frequency-based procedure knowledge graph." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": "Effect of utilizing predicted steps for input conditions to train the procedure planing model. Our proposed problem decomposition allows training the planning model with ground truth (GT) first and last steps. We experiment with two ways to train the planning model. Method 1 uses the predicted start and end steps " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "(\\hat{a}_1" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "\\hat{a}_T)" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": " as input to generate " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": " conditions and use them to train the planning model. Method 2 is where we augment the predicted start and end steps using the GT start and end steps " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "(a_1" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "a_{T})" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": " by generating 3 more data samples as follows: " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "[\\hat{a}_1,a_T]" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "[a_1,\\hat{a}_T]" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "[a_1,a_T]" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": ". Then we generate " + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 291, + 546, + 469 + ], + "type": "text", + "content": " conditions for each data and train the model. From the results shown in Table 8, the method without GT data augmentation shows better results. This suggests that leveraging ground truth data in training can lead to worse performance in testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": "Qualitative results. Figures 3 and 4 provide qualitative examples of our method. Intermediate steps are padded in the step model because it only predicts the start and end actions. In the 'make jello shot' task (see Figure 3), the model gives a wrong prediction in the intermediate steps when using " + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "inline_equation", + "content": "R = 1" + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": ") condition. However, it predicts correctly when using " + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": ") conditions. In the 'change a tire' task shown in Figure 4, the model is able to predict all the intermediate steps in given conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "type": "text", + "content": "Visualizations of the probabilistic procedure knowledge graph. We show a sub-graph from our probabilistic procedure knowledge graph (Figure 5). This graph is drawn around the 'jack up' node up to the depth of 2 nodes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Visualizations of the expert trajectories. Figure 6 illustrates the steps involved in completing the 'make jello shots' task, along with their transitions to other steps within the entire training data. This figure demonstrates that our " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " encodes diverse sequencing possibilities for steps and also captures task-sharing steps across the entire training domain. For instance, 'pour water' is a step in 'make jello" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18822" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 60, + 531, + 122 + ], + "blocks": [ + { + "bbox": [ + 63, + 60, + 531, + 122 + ], + "lines": [ + { + "bbox": [ + 63, + 60, + 531, + 122 + ], + "spans": [ + { + "bbox": [ + 63, + 60, + 531, + 122 + ], + "type": "table", + "html": "
ModelT=3T=4T=5T=6
SRmAccmIoUSRmAccmIoUSRmAccmIoUSRmAccmIoU
w.o P2KG conditions †35.6963.9166.0420.5257.4764.3912.853.4464.018.1550.4564.13
Ours †38.1264.7467.1524.1559.0566.6414.2053.8465.569.2750.2265.97
w.o P2KG conditions31.3559.5163.1118.9256.2062.4712.7151.2963.568.1647.6363.39
Ours33.3860.7963.8921.0256.0864.1512.7451.2363.169.2350.7865.56
", + "image_path": "55bfa00a0cf0b51efbf66098b2e7b344f5e014bf0cd3adf9ff4cb1da706c406f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 160, + 282, + 300 + ], + "blocks": [ + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "lines": [ + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "spans": [ + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "type": "text", + "content": "Table 5. Performance of our method with and without " + }, + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "type": "inline_equation", + "content": "{\\mathrm{P}}^{2}\\mathrm{{KG}}" + }, + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "type": "text", + "content": " conditions on CrossTask " + }, + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "type": "inline_equation", + "content": "{}^{ \\star }" + }, + { + "bbox": [ + 123, + 130, + 470, + 141 + ], + "type": "text", + "content": " dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 160, + 282, + 300 + ], + "lines": [ + { + "bbox": [ + 52, + 160, + 282, + 300 + ], + "spans": [ + { + "bbox": [ + 52, + 160, + 282, + 300 + ], + "type": "table", + "html": "
Model (T=6, CrossTask*)SRmAccmIoU
Ours with P2KG (R=1)
PDPP setting9.2750.2265.97
Conventional setting8.0950.8065.39
One LLM plan recommendation
PDPP setting (13b)7.7450.2864.05
Conventional setting (13b)7.2149.6863.89
PDPP setting (70b)8.6250.3164.34
Conventional setting (70b)7.8149.7564.02
P2KG (R=1) and one LLM plan recommendation
PDPP setting (13b)8.8149.9765.22
Conventional setting (13b)8.2051.4664.30
PDPP setting (70b)9.0150.2565.57
Conventional setting (70b)8.3451.5364.96
", + "image_path": "8d16bbed8315b4fa877827c4266d0caca5f0be784574aea029dc6025ac39b110.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 90, + 346, + 245, + 379 + ], + "blocks": [ + { + "bbox": [ + 47, + 309, + 286, + 331 + ], + "lines": [ + { + "bbox": [ + 47, + 309, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 286, + 331 + ], + "type": "text", + "content": "Table 6. Performance of the plan recommendations provided by the probabilistic procedure knowledge graph v.s. LLM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 346, + 245, + 379 + ], + "lines": [ + { + "bbox": [ + 90, + 346, + 245, + 379 + ], + "spans": [ + { + "bbox": [ + 90, + 346, + 245, + 379 + ], + "type": "table", + "html": "
ModelsSRmAccmIoU
Frequency graph7.6648.6164.21
Probabilistic graph8.0950.8065.40
", + "image_path": "f41647624f66a9fc3df3a29ae7641aabe43666dab546135ebf359b89e1c6515a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 80, + 435, + 254, + 468 + ], + "blocks": [ + { + "bbox": [ + 47, + 388, + 286, + 420 + ], + "lines": [ + { + "bbox": [ + 47, + 388, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 286, + 420 + ], + "type": "text", + "content": "Table 7. Performance comparison between probabilistic procedure knowledge graph v.s. frequency-based procedure knowledge graph for " + }, + { + "bbox": [ + 47, + 388, + 286, + 420 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 47, + 388, + 286, + 420 + ], + "type": "text", + "content": " on CrossTask dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 435, + 254, + 468 + ], + "lines": [ + { + "bbox": [ + 80, + 435, + 254, + 468 + ], + "spans": [ + { + "bbox": [ + 80, + 435, + 254, + 468 + ], + "type": "table", + "html": "
ConditionSRmAccmIoU
without GT data aug.38.1264.7467.15
with GT data aug.32.4562.4262.80
", + "image_path": "78f9c98600e713d88bb71d24c8f65256b22c7216c5326f30f19b451dab575373.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 59, + 514, + 276, + 557 + ], + "blocks": [ + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "lines": [ + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "type": "text", + "content": "Table 8. Effect of different input conditions for performance on CrossTask " + }, + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "type": "inline_equation", + "content": "{}^{ \\star }" + }, + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "type": "text", + "content": " dataset " + }, + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "type": "inline_equation", + "content": "\\left( {T = 3}\\right)" + }, + { + "bbox": [ + 47, + 477, + 286, + 499 + ], + "type": "text", + "content": " in PDPP’s setting" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 514, + 276, + 557 + ], + "lines": [ + { + "bbox": [ + 59, + 514, + 276, + 557 + ], + "spans": [ + { + "bbox": [ + 59, + 514, + 276, + 557 + ], + "type": "table", + "html": "
ModelsT=3T=4T=5T=6
ˆa1ˆaTˆa1ˆaTˆa1ˆaTˆa1ˆaT
Ours53.6950.6055.5652.5155.5851.8157.0951.92
Ours♣71.4263.3272.9863.3772.4263.2963.8259.96
", + "image_path": "e73d312471b4b9b4f798ae9628bfbd7e529230f163e645a9fce359281c1156c2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 565, + 286, + 586 + ], + "lines": [ + { + "bbox": [ + 47, + 565, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 286, + 586 + ], + "type": "text", + "content": "Table 9. The step model's start and end step prediction accuracies on the CrossTask dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 601, + 286, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 286, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 286, + 637 + ], + "type": "text", + "content": "shots' task, but it can also be part of other tasks, leading to a step transition from 'pour water' to 'add fish.' This structure allows models to leverage rich procedural knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 639, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 286, + 675 + ], + "type": "text", + "content": "Results for the step model. Table 9 reveals step model results, indicating potential enhancement areas to elevate the planning performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "Limitations & Failure cases. Our model exhibits three distinct failure case patterns. See section B.5 of the supplementary material for detailed discussions." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 148, + 542, + 271 + ], + "blocks": [ + { + "bbox": [ + 309, + 148, + 542, + 271 + ], + "lines": [ + { + "bbox": [ + 309, + 148, + 542, + 271 + ], + "spans": [ + { + "bbox": [ + 309, + 148, + 542, + 271 + ], + "type": "image", + "image_path": "db2da738b40c0a458f2503dbbf6960469d64d8ce24cb4392971c6f6a7d22ce5e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "lines": [ + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "type": "text", + "content": "Figure 5. Example of a sub-graph in our probabilistic procedure knowledge graph " + }, + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "type": "inline_equation", + "content": "(\\mathrm{P}^2\\mathrm{KG})" + }, + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "type": "text", + "content": " for CrossTask dataset. This graph effectively encapsulates real-world knowledge of distinct transition probabilities between steps, e.g., the probability of transitioning from 'start loose' to 'jack up' is 0.65, in contrast to a mere 0.14 for the reverse transition—the " + }, + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathrm{P}^2\\mathrm{KG}" + }, + { + "bbox": [ + 305, + 280, + 545, + 368 + ], + "type": "text", + "content": " reflects the common real-life practice where loosening the lug nuts before jacking up the car leads to a safer and more efficient tire change." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 385, + 543, + 509 + ], + "blocks": [ + { + "bbox": [ + 309, + 385, + 543, + 509 + ], + "lines": [ + { + "bbox": [ + 309, + 385, + 543, + 509 + ], + "spans": [ + { + "bbox": [ + 309, + 385, + 543, + 509 + ], + "type": "image", + "image_path": "fbfd42d59b3621b2bfa32ab374011a1653151f69d73f47d634e03eeb447e2bfb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 513, + 545, + 545 + ], + "lines": [ + { + "bbox": [ + 305, + 513, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 545, + 545 + ], + "type": "text", + "content": "Figure 6. Expert trajectories of the 'Make Jello Shots' task, involving task-sharing steps and thus out-of-task step transitions. Thicker lines indicate paths that are more frequently visited" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We focus on formulating procedural plans from an AI agent in instructional videos. We propose KEPP which employs a probabilistic procedural knowledge graph, sourced from the training domain, effectively serving as a 'textbook' for procedure planning. Results show that KEPP delivers SOTA performance with minimal supervision. Future work can focus on enhancing the accuracy of predictions for the initial and final steps. Additionally, our approach can be modified to aid in detecting erroneous steps and the misordering of steps in instructional videos [38, 41]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18823" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Yazan Abu Farha and Juergen Gall. Uncertainty-aware anticipation of activities. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 190 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 245 + ], + "type": "text", + "content": "[3] Anurag Ajay, Seungwook Han, Yilun Du, Shaung Li, Abhi Gupta, Tommi Jaakkola, Josh Tenenbaum, Leslie Kaelbling, Akash Srivastava, and Pulkit Agrawal. Compositional foundation models for hierarchical planning. arXiv preprint arXiv:2309.08587, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4575-4583, 2016. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "type": "text", + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. arXiv preprint arXiv:2307.08763, 2023. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "text", + "content": "[6] Anonymous authors. Active procedure planning with uncertainty-awareness in instructional videos, 2023. Under review as a conference paper at ICLR 2024. https://openreview.net/pdf?id=JDd46WodYf.2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "type": "text", + "content": "[7] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 448, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 287, + 513 + ], + "type": "text", + "content": "[8] Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin, Jena D Hwang, Xiang Lorraine Li, Hirona J Arai, Soumya Sanyal, Keisuke Sakaguchi, Xiang Ren, and Yejin Choi. Plasma: Making small language models better procedural knowledge models for (counterfactual) planning. arXiv preprint arXiv:2305.19472, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 557 + ], + "type": "text", + "content": "[9] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 624 + ], + "type": "text", + "content": "[10] Brian Chen, Nina Shvetsova, Andrew Rouditchenko, Daniel Kondermann, Samuel Thomas, Shih-Fu Chang, Rogerio Feris, James Glass, and Hilde Kuehne. What, when, and where?-self-supervised spatio-temporal grounding in untrimmed multi-action videos from narrated instructions. arXiv preprint arXiv:2303.16990, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[11] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2447, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Hazel Doughty, Ivan Laptev, Walterio Mayol-Cuevas, and Dima Damen. Action modifiers: Learning from adverbs in" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 868-878, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "type": "text", + "content": "[13] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "text", + "content": "[14] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "text", + "content": "[15] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18952-18961, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[16] Kiana Ehsani, Hessam Bagherinezhad, Joseph Redmon, Roozbeh Mottaghi, and Ali Farhadi. Who let the dogs out? modeling dog behavior from visual data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4051-4060, 2018. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "text", + "content": "[17] Sophie Fischer, Carlos Gemmell, Iain Mackie, and Jeffrey Dalton. Vilt: Video instructions linking for complex tasks. In Proceedings of the 2nd International Workshop on Interactive Multimedia Retrieval, pages 41-47, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 377, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 409 + ], + "type": "text", + "content": "[18] Kevin Flanagan, Dima Damen, and Michael Wray. Learning temporal sentence grounding from narrated egovideos. arXiv preprint arXiv:2310.17395, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[19] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. arXiv preprint arXiv:2005.03684, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "text", + "content": "[20] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threedorld transport challenge: A visually guided task-and-motion planning benchmark for physically realistic embodied ai. arXiv preprint arXiv:2103.14025, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[21] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "text", + "content": "[23] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. Machine Intelligence Research, 20 (2):249-262, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] De-An Huang, Joseph J Lim, Li Fei-Fei, and Juan Carlos Niebles. Unsupervised visual-linguistic reference resolution in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2183–2192, 2017. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18824" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] De-An Huang, Shyamal Buch, Lucio Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding\" it\": Weakly-supervised reference-aware visual grounding in instructional videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5948-5957, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "text", + "content": "[26] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "text", + "content": "[27] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[28] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 337 + ], + "type": "text", + "content": "[29] Zhiheng Li, Wenjia Geng, Muheng Li, Lei Chen, Yansong Tang, Jiwen Lu, and Jie Zhou. Skip-plan: Procedure planning in instructional videos via condensed action space learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10297-10306, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[30] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[31] Yujie Lu, Pan Lu, Zhiyu Chen, Wanrong Zhu, Xin Eric Wang, and William Yang Wang. Multimodal procedural planning via dual text-image prompting. arXiv preprint arXiv:2305.01795, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[32] Weichao Mao, Ruta Desai, Michael Louis Iuzzolino, and Nitin Kamra. Action dynamics task graphs for learning plannable representations of procedural tasks. arXiv preprint arXiv:2302.05330, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 524 + ], + "type": "text", + "content": "[33] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. arXiv preprint arXiv:2306.03802, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 591 + ], + "type": "text", + "content": "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "type": "text", + "content": "[35] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[36] Davide Moltisanti, Frank Keller, Hakan Bilen, and Laura Sevilla-Lara. Learning action changes by measuring verb-adverb textual relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23110-23118, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[37] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[38] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[39] Megha Nawhal, Akash Abdu Jyothi, and Greg Mori. Rethinking learning approaches for long-term action anticipation. In European Conference on Computer Vision, pages 558-576. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[40] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15302-15314, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 341 + ], + "type": "text", + "content": "[41] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "type": "text", + "content": "[42] Fadime Sener, Rishabh Saraf, and Angela Yao. Transferring knowledge from text to video: Zero-shot anticipation for procedural actions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[43] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "text", + "content": "[44] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 490, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 555 + ], + "type": "text", + "content": "[45] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 557, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 623 + ], + "type": "text", + "content": "[46] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[47] Yale Song, Gene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[48] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Look for the change: Learning object states and state-modifying actions from untrimmed" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18825" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "web videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13956-13966, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[49] Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal planning networks: Learning generalizable representations for visuomotor control. In International Conference on Machine Learning, pages 4732-4741. PMLR, 2018. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[50] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters, 7(2):4924-4930, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[51] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. Advances in Neural Information Processing Systems, 34:14476-14487, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "text", + "content": "[52] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "text", + "content": "[53] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13565–13575, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[54] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14836-14845, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[55] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[56] Yue Yang, Joongwon Kim, Artemis Panagopoulou, Mark Yatskar, and Chris Callison-Burch. Induce, edit, retrieve: Language grounded multimodal schema for instructional video retrieval. arXiv preprint arXiv:2111.09276, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "text", + "content": "[57] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "text", + "content": "[58] Jiahao Zhang, Anoop Cherian, Yanbin Liu, Yizhak Ben-Shabat, Cristian Rodriguez, and Stephen Gould. Aligning step-by-step instructional diagrams to video demonstrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2483-2492, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[59] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 294 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[60] Qi Zhao, Ce Zhang, Shijie Wang, Changcheng Fu, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368, 2023. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "text", + "content": "[61] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "type": "text", + "content": "[62] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 230, + 546, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 546, + 294 + ], + "type": "text", + "content": "[63] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18826" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_content_list.json b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0cc306f6851cf2cd85bb6d2f08493ec6fdc18c08 --- /dev/null +++ b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_content_list.json @@ -0,0 +1,1572 @@ +[ + { + "type": "text", + "text": "WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification", + "text_level": 1, + "bbox": [ + 122, + 128, + 846, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Satish Kumar $^{1*}$ Bowen Zhang $^{1*}$ Chandrakanth Gudavalli $^{1}$ Connor Levenson $^{1}$ Lacey Hughey $^{2}$ Jared A. Stabach $^{2}$ Irene Amoke $^{3}$ Gordon Ojwang $^{4}$ Joseph Mukeka $^{5}$ Stephen Mwiu $^{5}$ Joseph Ogutu $^{6}$ Howard Frederick $^{7}$ B.S. Manjunath $^{1}$ $^{1}$ University of California Santa Barbara $^{2}$ Smithsonian National Zoo and Conservation Biology Institute $^{3}$ Kenya Wildlife Trust, $^{4}$ University of Groningen, $^{5}$ Wildlife Research and Training Institute $^{6}$ University of Hohenheim, $^{7}$ Tanzania Wildlife Research Institute", + "bbox": [ + 93, + 202, + 875, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(satishkumar,bowen68,chandrakanth,clevenson,manj)@ucsb.edu, (stabachj,hugheyl)@si.edu,iirena.amoke@kenyawildlifetrust.org, (gordonojwang,simbamangu)@gmail.com,(jmukeka,smwiu)@wrti.go.ke", + "bbox": [ + 192, + 311, + 767, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 397, + 313, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce WildlifeMapper (WM), a flexible model designed to detect, locate, and identify multiple species in aerial imagery. It addresses the limitations of traditional, labor-intensive wildlife population assessments that are central to advancing environmental conservation efforts worldwide. While a number of methods exist to automate this process, they are often limited in their ability to generalize to different species or landscapes due to the dominance of homogeneous backgrounds and/or poorly captured local image structures. WM introduces two novel modules that help to capture the local structure and context of objects of interest to accurately localize and identify them, achieving a state-of-the-art (SOTA) detection rate of 0.56 mAP. Further, we introduce a large aerial imagery dataset with more than 11k Images and 28k annotations verified by domain experts. WM also achieves SOTA performance on 3 other publicly available aerial survey datasets collected across 4 different countries, improving mAP by $42\\%$ . Source code and trained models are available at Github1.", + "bbox": [ + 75, + 420, + 472, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 720, + 209, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces WildlifeMapper (WM) - an automated and scalable method for counting wildlife in aerial imagery. Aerial wildlife surveys are recognized as a cornerstone of modern conservation biology. By facilitating large-scale biological monitoring in remote landscapes, this technique has underpinned the ability to track changes in the abundance and distribution of wildlife across open landscapes for decades. However, traditional survey approaches", + "bbox": [ + 75, + 744, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "often rely on manual observers to identify, count, and validate species of interest. This labor-intensive process can be time-consuming and error-prone, with potential to limit the utility of final results [27, 35, 38].", + "bbox": [ + 500, + 398, + 890, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automated approaches offer a promising alternative for efficient and accurate detection of wildlife in aerial survey images. Recent work, for example, illustrates how artificial intelligence has been used to count a variety of species from the air, including antelope in grasslands [44], whales in the ocean [11], and seals on the beach [42]. When combined with advancements in low-cost, high-resolution imaging platforms (e.g., UAVs), these case studies underscore the potential for such data to significantly reduce the effort and cost of traditional wildlife census methods. However, the majority of these techniques struggle to generalize to new species or landscapes due to the dominance of homogeneous backgrounds and poorly captured local structures [7, 12, 18, 39].", + "bbox": [ + 496, + 460, + 892, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "WildlifeMapper overcomes these limitations by adapting a novel application of the segment anything transformer model [21]. This model combines high frequency component correlations and spatial correlations in the image data to generate a map of potential locations of objects of interest (i.e., wildlife, livestock). In addition, we address the challenge of identifying multiple species from a relatively small footprint in these images.", + "bbox": [ + 496, + 672, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To demonstrate the WM analysis workflow, we provide a case study example across the Masai Mara Ecosystem in southwestern Kenya. Renowned for its rich biological diversity, the abundance of large mammals (such as buffalo (Syncerus caffer), giraffe (Giraffa tippelskirchi), and wildebeest (Connochaetes taurinus)) have declined precipitously over the past few decades [34]. Importantly, our analy-", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/UCSB-VRL/WildlifeMapper * refers to equal contribution", + "bbox": [ + 76, + 875, + 351, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12594", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de19e1e68141ca4c8edb39965055f98a17a22c0096e8d3b8868772137af7bab8.jpg", + "image_caption": [ + "Figure 1. Summary of Mara-Wildlife dataset. (a) Satellite view indicating the four flight trajectories, each represented in a different color. (b, c, d, e, f, g) Annotations of (b) zebra, (c) hartebeest, (d) cattle, (e) shoats (sheep and goats), and (f) zebra. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 397, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6832a4e5243a8e9398d549c8bd8efb56ec9e6d4c661029589a4e627d132a094b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 90, + 640, + 275 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/601a27c2752bd01b8bd0b0412b7d042cf65fa2560ccac096200ec18b93c0b85a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 90, + 764, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6607979af0fc92740acb765118178ee2e51d58dd5df1a94ebfed2924da6129d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 90, + 887, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sis incorporated 11, 151 images of size $8400 \\times 5500$ collected from a digital camera affixed to the bellyport of a Partenavia P68 airplane during Systematic Reconnaissance Flight (SRF) surveys. Part of these images were annotated by trained observers with 28, 146 annotations of 21 species of large mammals ( $\\geq 15kg$ ), providing an unprecedented opportunity to develop species detection models across a complex, heterogeneous environment. The dataset was systematically verified by trained observers as described in Section 3. Our contributions can be summarized as follows:", + "bbox": [ + 75, + 327, + 470, + 478 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A novel, single-stage end-to-end approach for animal detection. The modules, a High Frequency Feature Generator, a Feature Refiner, and a Query Refiner, work together to improve upon the traditional methods of object detection in aerial imagery and enable generalizability across different habitats. The high frequency features reduce dependence on dominant backgrounds/landscapes.", + "2. An input patch embedding layer that is specifically designed to capture contextual information to help in identifying individual animal species.", + "3. The release of a new benchmark dataset via the data owner (Kenya's Wildlife Research and Training Institute - WRTI) once all approvals are in place. An international user community is already engaged in further enhancing these data and using the WM through the BisQue [25] platform." + ], + "bbox": [ + 88, + 487, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Community Adoption The practical use of WildlifeMapper (WM) extends beyond theoretical and computational success and is operationalized through BisQue [25]. WM is made available to users through a series of training modules that demonstrate how to $(i)$ upload digital imagery, $(ii)$ create annotations, $(iii)$ apply and/or improve existing models, $(iv)$ evaluate model fit, $(v)$ improve annotations and refit models, and $(vi)$ generate summary statistics. In the fu", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ture, we envision WM to be of great value to ecologists, wildlife managers, and government officials, providing accurate information about the state of wildlife populations in near real-time, facilitating decision-making processes, and improving the conservation of ecosystems globally.", + "bbox": [ + 496, + 327, + 892, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 411, + 648, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Manual Methods: Aerial surveys using Front- and Rear-Seat Observers (FSO and RSO, respectively) are commonly used to inventory wildlife populations across open landscapes [36]. However, several important biases can impact these counts, including the experience level and fatigue of the human observers [28].", + "bbox": [ + 496, + 431, + 892, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Learning Methods: To address these issues, researchers have begun incorporating digital cameras on piloted aircraft and UAVs [18,30,46]. This minimizes the influence of observer bias while increasing transparency and reproducibility of results. For example, [26] replaced RSOs with an oblique camera mount system minimized the influence of observer bias while producing comparable estimates of large mammals under partial canopy cover. Similarly, [44] used a nadir mounted camera to improve the accuracy and efficiency of manual counts of large antelope in an open grassland ecosystem.", + "bbox": [ + 496, + 522, + 892, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, detecting animals in the wild from aerial imagery poses many challenges. For example, most publicly available datasets for aerial object detection are focused on identifying relatively distinct features such as buildings, roads, vehicles, and other man-made structures [8, 14, 53]. Animals, tend to blend in with their surroundings [13], can be occluded by trees, exhibit considerable variation in color and pattern, or have behavioral adaptations that make them difficult to detect [7, 20, 39, 48].", + "bbox": [ + 496, + 688, + 892, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[20] proposed a solution for this problem involving a two-branch CNN model based on AlexNet to perform animal recognition and localization. [7] evaluated three state-of-the-art object detection algorithms, including Faster-RCNN, Libra-RCNN, and RetinaNet on six African wild", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5341e0e4cd475f09e0ed21ef4138abbd9227727e77b1fc66de0b3b25d610b7d8.jpg", + "image_caption": [ + "Figure 2. Distribution of $(\\geq 15kg)$ mammals identified in digital imagery collected across the Masai Mara Ecosystem, Kenya." + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 890, + 199 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/45c559b45cfae964645a60334eea44dab798fccd8d083a370003b4471ab4ea0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# of annot. images# of annot. tiles# of species# of annotationImage sizeGSD (cm)Location
Virunga73930069656646000x40002.4DRC
Garamba1586429616116000x40002.0DRC
AED2067693871155815500x36002.4-13.0Bostwana, Namibia, South Africa
Mara-Wildlife10127796621281468256x55041.45Masai Mara National Reserve
", + "bbox": [ + 78, + 224, + 893, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of Mara-Wildlife dataset with other publicly available dataset. Mara-Wildlife dataset has $\\times 3$ more unique species than the total of all other datasets. Each image is significantly larger and higher ground resolution making 77k unique images of size ${1024} \\times {1024}$ with 21 different animal species. GSD: ground sampling distance; DRC: Democratic Republic of Congo.", + "bbox": [ + 75, + 357, + 892, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mammals. All three algorithms, however, showed poor performance in animal detection when animals were grouped closely together in herds. [12] adopted a segmentation approach, employing a UNet model to detect livestock from drone imagery. [48] uses a comparable model to analyze high-resolution satellite imagery, producing segmentation masks of wildebeest-sized animals, which are subsequently utilized for detection and counting.", + "bbox": [ + 75, + 414, + 468, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers: WM adopts a transformer architecture based on past success in modeling different types of aerial imagery. Examples include incorporating multispectral imagery for change detection [4], landcover classification [16], greenhouse gas (GHG) emission detection [23,24] and RGB aerial imagery for object detection [29, 49, 52].", + "bbox": [ + 75, + 536, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most effective applications of transformer-based models have been tailored for standard object detection tasks [3, 5, 32, 41, 54]. These works leverage the self-attention to model dependencies among the patches in an end-to-end fashion, unlike CNN-based models [1, 40]. However, when directly applied to aerial imagery, these models cannot effectively exploit the local structures as they divide the image into a sequence of patches. This limits the detection of small-scale objects in a homogeneous and dominant background.", + "bbox": [ + 75, + 628, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset: The existing publicly available animal aerial imagery datasets are listed in Table 1. The Virunga dataset [7] was collected in Virunga National Park, Democratic Republic of Congo (DRC). This dataset contains 897 annotated images of 6 animal species. The Garamba dataset was collected in Garamba National Park, DRC and contains 7034 images. Only 158 images have been annotated, containing 7 animal species. The aerial elephant dataset (AED) [33] was", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "collected across a mosaic of woodland, open shrubland, and grassland habitats in Botswana, Namibia, and South Africa. Only a single species (i.e., elephant) was targeted during SRF surveys. See Table 1 for details.", + "bbox": [ + 496, + 414, + 890, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Mara-Wildlife Dataset", + "text_level": 1, + "bbox": [ + 500, + 483, + 714, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Mara-Wildlife dataset is a distinctive dataset that captures the essence of the Masai Mara ecosystem through a compilation of 77966 images of size $1024 \\times 1024$ . This habitat is heterogeneous, including woodland, shrubland, and grassland vegetation with 21 unique animal species.", + "bbox": [ + 496, + 508, + 890, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image Collection", + "text_level": 1, + "bbox": [ + 500, + 589, + 666, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flightline Details: Data collection was in collaboration with the Smithsonian National Zoo and Conservation Biology Institute (SNZCBI), Kenya's Wildlife Research and Training Institute (WRTI), the Kenya Wildlife Trust (KWT), and the Directorate of Resource Surveys and Remote Sensing (DRSRS). In March 2022, we fitted a Partenavia P68 with a Nikon D850 digital camera and collected high resolution $(8256\\times 5504)$ digital images. During data acquisition, the aircraft adhered to a predetermined flight trajectory, depicted in Figure 1, at 400 ft above ground level (agl). This trajectory was optimized to encompass open grassland areas across the Masai Mara ecosystem, including the Masai Mara National Reserve, 22 adjacent private conservancies, and unprotected peripheral areas.", + "bbox": [ + 496, + 613, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The aerial survey was conducted during a wet season period when the Serengeti migratory population of wildebeest have moved southward to locate more suitable forage in Tanzania. Thus, the survey primarily captured resident species, including wildebeest, zebra, topi, hartebeest, gi", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12596", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4cd2368475fed771be868c7297357bd3efcd94f068ce5a916b5aa138c4bee42.jpg", + "image_caption": [ + "Figure 3. Overview of WildlifeMapper (WM) architecture. Given an input image of size $1024 \\times 1024 \\times 3$ , the High-Frequency Feature Generator (HFG) module generates information about potential location of object of interest. The Feature Refiner (FR) takes these potential locations along with contextual features from Patch Embed layer and sent output to Image Encoder. In parallel, the Query Refiner (QR) incorporates the output of FR to refine learnable queries. Finally these queries are decoded using encoded features from Image Encoder and predict bounding box and class." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "raffe, and other large ( $\\geq 15\\mathrm{kg}$ ) antelope. Data collection was conducted in the early mornings (prior to 10:00 EAT) and late afternoons (after 15:00 EAT) when lighting conditions were optimal and animals are most active.", + "bbox": [ + 75, + 332, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The geographical positioning of each image was acquired through a GPS system that recorded the plane's altitude, speed, and geographical coordinates. These data were then synchronized with the image using the image capture timestamps, enabling us to determine the geographic location of the centroid of each image. The camera was placed in the bellyport of the airplane, capturing a nadir view of the landscape every two seconds along the flight path.", + "bbox": [ + 75, + 393, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Annotation", + "text_level": 1, + "bbox": [ + 76, + 523, + 253, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initial bounding box annotations (21796) generated using AIDE platform [19] were exported in CSV format. These were then imported into BisQue [25] for further validation and correction. Finally, these annotations underwent validation by a single trained observer specializing in ecology, resulting in 28146 annotations in total.", + "bbox": [ + 75, + 541, + 468, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Statistics", + "text_level": 1, + "bbox": [ + 76, + 640, + 243, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Mara-Wildlife dataset showcases a detailed assemblage of wildlife, inclusive of 21 distinct species classes. The dataset is composed of approximately 77,966 tiled images, derived from 1,012 original rasters (Table. 1) The meticulous process of annotation has culminated in labeling 28,146 animals. Species identified include domestic cattle (Bos taurus), white-bearded wildebeest (Connochaetes taurinus), topi (Damaliscus lunatus), shoats (domesticated sheep and goats), kongoni (Alcelaphus buselaphus), waterbuck (Kobus ellipsiprymnus), impala (Aepyceros melampus), Grant's gazelle (Nanger granti), Thomson's gazelle (Eudorcas thomsonii), Cape buffalo (Syncerus caffer), zebra (Equus quagga), ostrich (Struthio camelus), Masai giraffe (Giraffa tippelskirchi), warthog (Phacochoerus africanus), eland (Taurotragus oryx), donkey (Equus africanus), hyena (Crocuta crocuta), hippopotomus", + "bbox": [ + 75, + 657, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Hippopotamus amphibius), lion (Panthera leo), and elephant (Loxodonta africana).", + "bbox": [ + 498, + 332, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. WildlifeMapper Architecture", + "text_level": 1, + "bbox": [ + 500, + 372, + 769, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Technical Overview", + "text_level": 1, + "bbox": [ + 500, + 398, + 687, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "WildlifeMapper's architecture is inspired by the success of the Segment Anything Model (SAM) [21], created to segment small/large (all sizes) of objects. Referring to Fig. 3, WM contains the following main components: (i) A patch embedding layer designed to capture long-range context, (ii) a High-Frequency Feature Generator (HFG), (iii) a Feature Refiner (FR) followed by ViT based image encoder [21,23,45], and (iv) a Query Refiner (QR) module followed by a box decoder module. The input image is first processed through two separate branches, the patch embedding layer which captures long-range context [9] and HFG which suppresses all the low-frequency components in the image and generates a feature embedding. The HFG (Sec. 4.3) exploits prior knowledge that aerial images from areas such as forests, grasslands, and shrublands have a homogeneous and dominant background representing the dominant low frequency image content. Fig. 4 shows that on suppressing the lower frequencies, the object of interest is easy to locate.", + "bbox": [ + 496, + 417, + 890, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The FR (Sec. 4.4) takes the embeddings from each of the two branches and generates a high quality embedding that contains information about potential locations of animals and captures the local context. The QR modules refines a set of learnable queries using the location information from FR module. These refined queries are passed to the box decoder. The box decoder takes the refined queries and encoded features image encoder to generate the final detection box and class of the object.", + "bbox": [ + 496, + 690, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Patch Embed", + "text_level": 1, + "bbox": [ + 500, + 838, + 638, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The patch embedding layer utilizes a larger kernel convolution with an increasing dilation rate. This design rapidly expands the receptive field, allowing explicit extrac", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion of features rich in contextual information. This approach is particularly beneficial for aerial imagery, where the small sized object makes classification based on appearance alone challenging. Contextual information thus becomes crucial for the accurate recognition of these objects.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. High-frequency Feature Generator (HFG)", + "text_level": 1, + "bbox": [ + 76, + 175, + 436, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Along with patch embedding, the input image is processed in parallel by the HFG module to generate features with information about the location of the animal or cluster as shown in Fig. 4. The HFG module is inspired from the limitation of ViT models [45]. ViT models face challenges in efficiently utilizing local structures. They segment an image into patches and apply self-attention to model relationships, but this approach often falls short in capturing detailed local features [37, 51].", + "bbox": [ + 75, + 194, + 468, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local features in images are closely linked to high-frequency components [2, 43]. We hypothesize that suppressing low-frequency components can mitigate the influence of a dominant homogeneous background. To test this, we performed a discrete Fourier Transform (DFT) on the images, filtering out the low-frequency components before reconstructing the images, as shown in Fig. 4.", + "bbox": [ + 75, + 330, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a given input image $I \\in \\mathbb{R}^{H \\times W \\times C}$ , where $C$ is channel dimension, we compute Discrete Fourier Transform (DFT) of $I$ . In next step we suppress the low frequency components with a controlling parameter and construct the image $I$ with inverse DFT (IDFT).", + "bbox": [ + 75, + 435, + 468, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {\\prime} = I D F T [ h p f (D F T (I) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 521, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $hpf$ is a high pass filter. Then we reduce the dimension of the reconstructed image $I'$ via an embedding layer to generate embedding $hfc_{emb}$ and pass them to the FR module. See supplementary materials for more details.", + "bbox": [ + 75, + 547, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Feature Refiner (FR)", + "text_level": 1, + "bbox": [ + 76, + 618, + 274, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, the features from the patch embedding layer $(f_{img})$ and HFG $(hfc_{emb})$ are fed to the FR module. The $f_{img}$ are refined with the $hfc_{emb}$ via cross-attention mechanism. The FR is a simple module with cross-attention and linear layers [45]. The output contains information about the potential location of the object and the long-range context.", + "bbox": [ + 75, + 636, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the standard architecture of SAM [21], we pass FR output to our ViT based image encoder supplemented with learnable positional embeddings $p$ . The encoded feature map is $f_{emb}$ :", + "bbox": [ + 75, + 727, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {e m b} = \\operatorname {V i T} \\left[ \\mathbf {F R} \\left(f _ {i m g}, h f c _ {e m b}\\right), p \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 797, + 468, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query Refiner (QR): The QR follows a transformer decoder like architecture and takes as input a set of 100 learnable queries $Q \\in \\mathbb{R}^{100 \\times d}$ and output of FR module. Here $d = 256$ is same as channel dimension of $f_{emb}$ from image encoder. The FR output refines the $Q$ via a cross-attention", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e90a672b29dee430d653ce2754938f7657c70c98fd61dd4ef0fa4e9a0ea10bb1.jpg", + "image_caption": [ + "Figure 4. The sample output visualization from the High-Frequency Feature Generator (HFG) module. The illustration shows the effectiveness of the module in suppressing the homogeneous and dominant background, while highlighting objects of interest (i.e., animals). The top image shows bomas, natural structures constructed to contain livestock, and paths that have been suppressed. Animals, however, are clearly identified, especially inside the boma. The bottom image shows a water body (a dam created for livestock) that has been suppressed by the module. Animals can again be highlighted throughout the image." + ], + "image_footnote": [], + "bbox": [ + 500, + 89, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mechanism. The refined queries narrows the search space for box decoder module to accurately locate and identify object of interest. [15,23].", + "bbox": [ + 498, + 569, + 890, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Box Decoder", + "text_level": 1, + "bbox": [ + 500, + 636, + 633, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, the refined queries are sent to the box decoder module. We concatenate these queries with the $f_{emb}$ and pass box decoder's self attention layer as $qf_{emb}$ . This is inspired from the idea of class_token used by Vaswani et.al. [45] to make decoding process memory efficient. The transformer network takes 4 input variables, those are position embeddings of queries and image embeddings, $qf_{emb}$ , and $f_{emb}$ . Our transformer model uses two-way attention inspired from [3, 6, 21] and our box decoder uses self-attention and cross-attention in two directions (queries-to-image embedding and vice-versa) to update all embeddings. We keep the box decoder very light weight (two blocks). The top 100 (equal to number of queries) indexes from the output of the final block is passed to two separate MLP blocks to regress the output bounding box prediction and class of the predicted box.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.6. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 90, + 292, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We've trained WM using a single-stage, end-to-end approach to determine bounding boxes and classify them. The loss strategy we applied for WM is akin to what's used in DETR [3]. Initially, we perform bipartite matching to align our model's predictions with the actual bounding box data. Then, we proceed to compute the loss for these matched pairs. To achieve the best possible match between our predictions and the real data, we use the Hungarian algorithm [22]. Once matched, each prediction is paired with its respective ground truth. We then measure the $l_{1}$ (L1 distance) and $GIoU$ loss for the bounding box and the cross entropy loss for the classification [3].", + "bbox": [ + 75, + 113, + 472, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference: The inference pipeline is straightforward and similar to training code. During inference, we first filter the detections at $50\\%$ threshold and then use non-maximal suppression to remove any overlapping boxes.", + "bbox": [ + 75, + 306, + 470, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 381, + 209, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Train-val-test split: There is no data leak in train-val-test split. Fig. 1 shows the flight paths with location of each image represented by a circle. Each color coded flight path represents the train-val-test image set. No images taken while the airplane was cutting across the transects plus turnaround points as can be seen there are no circles. We created a spatial disjoint of $20km$ distance between the transects as shown in Fig. 1 and achieved consistent performance (0.56mAP).", + "bbox": [ + 75, + 406, + 470, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: We report our performance on multiple metrics. Following the protocols of standard object detection, we report the performance in mean average-precision (mAP) for detection and mean intersection-overunion (mIOU) for localization of animals. We also report a commonly used metric by the ecologists on the team, the class-wise mean absolute error (MAE) indicating the counting accuracy of each species:", + "bbox": [ + 75, + 553, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M A E} = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\sum_ {c = 1} ^ {C} | \\hat {n} _ {i, c} - n _ {i, c} |, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 685, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $I$ is the number of images, $C$ is the number of classes, $\\hat{n}_{i,c}$ and $n_{i,c}$ are the predicted and ground truth counts for class $c$ in image $i$ .", + "bbox": [ + 75, + 738, + 470, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: Each image taken from the drone is $8256 \\times 5506 \\times 3$ . We create tiles for each image in the spatial domain, with the size of $1024 \\times 1024 \\times 3$ with $25\\%$ of overlap. The Patch Embed layer uses a single CNN layer with a large kernel of size $16 \\times 16$ with stride 16. In the parallel branch, the High-Frequency Feature Generator, we use DFT to compute the Fourier transform, the mask is", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a binary disk with the radius set to 128. The HFC Embed layer uses 3 CNN layers with ReLU activation with a kernel of size $3 \\times 3$ and a global average pool at the end. The Feature Refiner (FR) module consists of one cross attention layer with 1 linear layer. The image encoder is a pre-trained ViT model [21] with 24 transformer layers and 16 heads. The Query Refiner (QR) module takes in 100 queries each of channel dimension 256, those are cross attended with $hfc_{emb}$ output. The box decoder contains 3 layers of two-way attention with 8 heads. We train WM with AdamW optimizer [31] setting the learning rate to $10^{-4}$ for the FR, QR and box decoder with a weight decay to $10^{-4}$ . We set the learning rate for the Patch Embed and HFC Embed layer to $10^{-5}$ . We load the image encoder with pre-trained weights from segment anything [21] and keep it frozen.", + "bbox": [ + 496, + 90, + 893, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Augmentation: In the train and test datasets, we incorporated an equal number of images without any objects to assess the model's robustness against empty background images. We applied multiple data augmentation techniques, including HSV (hue, saturation, and value) $(10\\%)$ , rotation $(5\\%)$ , translation $(10\\%)$ , affine transformation $(20\\%)$ , scale $(10\\%)$ , shear $(5\\%)$ and mosaic $(70\\%)$ augmentation [1]. Mosaic augmentation is proven to be the most effective, with an improvement of $0.07\\mathrm{mAP}$", + "bbox": [ + 496, + 330, + 893, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hard Negative Mining: After training for 100 epochs, we take all the False Positives (FP) predictions having $\\mathrm{IOU} \\leq 0.10$ with ground truth box and mark them as background class. Then fine-tuning for 20 epochs improved the performance of FP reduction for detecting rocks, dead tress or other artifacts on ground as animal.", + "bbox": [ + 496, + 479, + 893, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 500, + 585, + 586, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Performance Comparison", + "text_level": 1, + "bbox": [ + 500, + 604, + 735, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We trained WM separately on Mara-Wildlife dataset and Virunga-Garamba-AED dataset for comparison with existing works, see Table 2 for a summary of the results on all of the tested datasets. The $mAP$ values are compared for IoU of 0.50-0.95 and 0.50. We also provide the average counting error in animal counting per image. We trained all the baseline models with the default set of conditions on Mara-Wildlife dataset. We merged the Virunga, Garamba, and AED datasets and created the train-val-test split according to [7]. The combination of these datasets contains 6 unique animal species and diversity of landscapes such as woodland, savannahs, open shrubland, and grasslands across multiple countries – Democratic Republic of Congo (DRC), Botswana, Namibia, and South Africa. WM outperforms all methods by a significant margin as shown in Table 2. We note that in [7] the authors did not make the code base or trained model public, hence we could not verify the results. We implemented these methods from", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12599", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af1e9fc6f10b2e77c1f83634605266101fe9e06ddc61174da47e525bcbdfb1e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods#epochsmAPmAP50Counting Error
Mara Wildlife Dataset
1Faster-Rcnn [40]1000.240.582.59
2DETR [3]2000.220.572.75
3Co-DETR-R50 [55]1000.270.662.72
4Co-DETR-swingL [55]1000.280.652.60
5Yolo v5 [10]1000.300.672.12
6Yolo v8 [17]1000.270.613.97
7LSKNet [29]1000.290.66-
6DroneDetect [50]1000.180.48-
8WildlifeMapper1200.560.791.9
Virunga-Garamba-AED Datasets
6Faster-Rcnn1200.340.650.27
7DETR2000.300.620.45
8Yolo v51000.480.780.12
9Yolo v81000.480.770.42
10WildlifeMapper800.680.850.11
", + "bbox": [ + 76, + 88, + 468, + 310 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/00318b09bc148ca7c22a87b4d020bdd9e3f436bc12341f0cf97f6d1efe9b1cb7.jpg", + "table_caption": [ + "Table 2. Comparison with baseline models. The top section shows performance on species detection on Mara-wildlife dataset and low section shows performance on the mixed dataset from Virunga-Garamba-AED dataset. The overall detection accuracy is generally higher in Virunga-Garamba-AED dataset because there are only 6 species and the terrain is quite similar in all images." + ], + "table_footnote": [], + "table_body": "
HFGFRQRmAP
WMXX*0.46
✓✓*0.54
✓X✓0.49
✓✓✓0.56
", + "bbox": [ + 153, + 404, + 393, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. HFG module effectiveness in refining the image features and queries. “X” represents not used, “√” represents used and “*” represents that random queries are used but there was not refining with HFG features.", + "bbox": [ + 75, + 492, + 468, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the original public repositories and trained according to the training strategy detailed in [7]. We attribute the model poor learning performance due to salient features of the homogeneous background being learned more than the object of interest. The detection of the object of interest is then dependent on the landscape properties instead of object properties. Hence when used on a slight variations of landscapes for the same object, the models struggle to detect. This limitation is specifically addressed in the WM, where the HFG modules suppresses the background and highlights the object of interest.", + "bbox": [ + 75, + 559, + 468, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results: Fig. 5 shows the quality of detection by WM in different scenarios. Those include, detection when animal is partially visible under a tree, or a big clustering. WM makes correct predictions in varying scenarios.", + "bbox": [ + 75, + 734, + 468, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 806, + 241, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We performed all ablation experiments on Mara-Wildlife (MW) dataset and validate the design choices.", + "bbox": [ + 76, + 829, + 468, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-frequency Feature Generator Module: In Table 3, we show the effectiveness of the HFG module. We", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "experimented with HFG's output in 3 ways: first, we passed HFG's output to Feature Refiner (FG) module. It leads to significant improvement in detection by $0.09\\mathrm{mAP}$ over the baseline. This demonstrates that providing potential location candidates features to image encoder module produce better embeddings. Second, we pass the HFG's output to Query Refiner (QR) module only. This leads to an improvement of $0.03\\mathrm{mAP}$ over baseline. This shows the effectiveness of guiding queries with location candidates features. In the third case, we passed the HFG's output to both FR and QR modules and achieved an improvement of $0.11\\mathrm{mAP}$ over the baseline. We hypothesize that this reduces the dominance of features from homogeneous and dominant background in aerial imagery. The also observed this while testing WM across flightlines different types of terrain such as green grasslands, dry grasslands, and forest areas.", + "bbox": [ + 496, + 90, + 890, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Refiner Module: We tested $hfc_{emb}$ and $f_{img}$ merging by 3 ways: addition, concatenation and cross-attention. Cross-attention is most effective, because with addition and concatenation, the $hfc_{emb}$ get lost, while cross-attention generates better embeddings giving attention to potential location candidates.", + "bbox": [ + 496, + 351, + 890, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Kernel Size: We observed that a larger kernel size of $31 \\times 31$ results in reduced misclassification. For example, a topi or warthog cannot be found inside a boma because only domestic species are kept in bomas. So context helps in making the right class detection. We observed an improvement of $0.02\\mathrm{mAP}$ . Experiments were done in 3 kernel sizes. 1. $7 \\times 7$ : $0.55\\mathrm{mAP}$ , 2. $16 \\times 16$ : $0.558\\mathrm{mAP}$ ; 3. $32 \\times 32$ : $0.57\\mathrm{mAP}$ .", + "bbox": [ + 496, + 460, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query Refiner Module: We experimented with only providing random queries and guiding the queries with HFG module output. We merged them with direct addition or concatenation and cross-attention. With cross-attention, we observed an improvement in performance of 0.03mAP.", + "bbox": [ + 496, + 599, + 890, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Geographic generalization: To test the geographic generalizability of WM across different terrains, we trained WM only on images from Kenya; and tested on images from Democratic Republic of Congo, Botswana and Namibia. The test was done on 4 common species which were present in both the ecosystems. WM achieved a detection performance of $0.48\\mathrm{mAP}$ . This shows the adaptability of WM across varying landscapes.", + "bbox": [ + 496, + 694, + 890, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain generalization: We train-test WM on a different domain, a bird species tern [47], commonly found on/near water bodies. Live in huge clusters. WM achieved the accuracy of $0.71\\mathrm{mAP}$ . Showing adaptability of WM across different domains.", + "bbox": [ + 496, + 834, + 890, + 907 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12600", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7da36f287a9680a3d38a1103eecc5ccdef580a499bb64df42d28969e962847cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 356, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdcaf9e16c6949d3362b62f597dea6482cbc320233d295c1cbc888049b210f7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 88, + 612, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c02f6e19018cec214b4ef2b317f0a06d7139d8cfb673145f181fe78734974b5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 88, + 870, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/709b87d56af4929a8775f447c530f3f8c078e8dcb0a591f3f3e032e5ac283b4f.jpg", + "image_caption": [ + "Figure 5. Qualitative results. The top row highlights examples of crowded and partially occluded scenes. Row-1, Column-1 & Row-3, Column-3 shows examples where animals are obstructed by shadows. The zoomed-in box in Row-3, Column-3 shows a zebra partially occluded. The bounding box color is coded according to class names: shoats- \"hot pink\", cattle- \"deep sky blue\", zebra- \"light yellow\"." + ], + "image_footnote": [], + "bbox": [ + 98, + 287, + 354, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ead6d7a1e79e473d59797ae734c2b69f4ac5c21e165bff6a24b13a98c9900668.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 287, + 612, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5436f5675f603e8e60d9191169ab6556dc0ddc4c11a9fb98eb5829baf103bcf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 287, + 870, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/418c4bb1f01517b5e27717d0e376fd49549229379a0488650fdb0ee07dd02377.jpg", + "image_caption": [ + "Figure 6. Failure cases. Left shows an example where animals are occluded by shadows and are difficult to detect. Right shows an example of rock detected as an impala, emphasizing the difficulty in differentiating objects in the image from animals of interest." + ], + "image_footnote": [], + "bbox": [ + 88, + 558, + 272, + 702 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05616dc7838c4257c59d1736e02807be40161ffe63b1433f9026916dfb0b63d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 558, + 457, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure Cases: The detections from WM are inaccurate when animals are clustered in shadows, such as when animals are located inside bomas and the sun angle makes a strong shadow on the enclosure. These are some of the difficult cases shown in Fig. 6. Other cases of false positives are the small rocks or trees that sometimes resemble animals. Some examples are shown in the supplementary materials.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 619, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents WildlifeMapper (WM) - a transformer based approach for the detection of animals of varying densities and sizes across natural backgrounds. The WM utilizes a high frequency features generator, feature refiner, and query refiner to accurately locate and classify 8 animal species. WM stands to significantly improve the efficiency and accuracy of wildlife monitoring and conservation efforts. Future work will extend this model and dataset to a larger number of species and habitats.", + "bbox": [ + 496, + 585, + 890, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgement", + "text_level": 1, + "bbox": [ + 498, + 734, + 679, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is partially supported by the following grants: NSF award SI2-SSI #1664172 (BSM), the Ohrstrom Family Foundation, and a Smithsonian National Zoo and Conservation Biology Institute (SNZCBI) fellowship (S.Kumar). We also recognize the contributions of the annotation team, including S. Harman, A. Gray, C. Pate, C. Obath, S. Nyarangi, C. Kagume, and W. Sairowua.", + "bbox": [ + 496, + 758, + 890, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12601", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3, 6", + "[2] Fergus W Campbell and John G Robson. Application of fourier analysis to the visibility of gratings. The Journal of physiology, 197(3):551, 1968. 5", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5, 6, 7", + "[4] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3", + "[5] Qiang Chen, Jian Wang, Chuchu Han, Shan Zhang, Zexian Li, Xiaokang Chen, Jiahui Chen, Xiaodi Wang, Shuming Han, Gang Zhang, et al. Group detr v2: Strong object detector with encoder-decoder pretraining. arXiv preprint arXiv:2211.03594, 2022. 3", + "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 5", + "[7] Alexandre Delplanque, Samuel Foucher, Philippe Lejeune, Julie Linchant, and Jérôme Théau. Multispecies detection and identification of african mammals in aerial imagery using convolutional neural networks. Remote Sensing in Ecology and Conservation, 8(2):166-179, 2022. 1, 2, 3, 6, 7", + "[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44(11):7778-7796, 2021. 2", + "[9] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 4", + "[10] Glenn Jocher et. al. ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements, Oct. 2020. 7", + "[11] Emilio Guirado, Siham Tabik, Marga L Rivas, Domingo Alcaraz-Segura, and Francisco Herrera. Whale counting in satellite and aerial images with deep learning. Scientific reports, 9(1):14259, 2019. 1", + "[12] Liang Han, Pin Tao, and Ralph R Martin. Livestock detection in aerial images using a fully convolutional network. Computational Visual Media, 5:221-228, 2019. 1, 3", + "[13] Tracey Hollings, Mark Burgman, Mary van Andel, Marius Gilbert, Timothy Robinson, and Andrew Robinson. How do you find the green sheep? a critical review of the use of remotely sensed imagery to detect and count animals. Methods in Ecology and Evolution, 9(4):881-892, 2018. 2", + "[14] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H. Hsu. Drone-based object counting by spatially regularized re" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gional proposal networks. In The IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 2", + "[15] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. In Pattern Recognition and Tracking XXXIV, volume 12527, pages 192-205. SPIE, 2023. 5", + "[16] Johannes Jakubik et. al. Prithvi-100M, Aug. 2023. 3", + "[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics* YOLO, Jan. 2023. 7", + "[18] Benjamin Kellenberger, Diego Marcos, and Devis Tuia. Detecting mammals in uav images: Best practices to address a substantially imbalanced dataset with deep learning. Remote Sensing of Environment, 216:139-153, Oct 2018. 1, 2", + "[19] Benjamin Kellenberger, Devis Tuia, and Dan Morris. Aide: Accelerating image-based ecological surveys with interactive machine learning. Methods in Ecology and Evolution, 11(12):1716-1727, 2020. 4", + "[20] Benjamin Kellenberger, Michele Volpi, and Devis Tuia. Fast animal detection in uav images using convolutional neural networks. In 2017 IEEE international geoscience and remote sensing symposium (IGARSS), pages 866-869. IEEE, 2017. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 4, 5, 6", + "[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6", + "[23] Satish Kumar, Ivan Arevalo, ASM Iftekhar, and BS Manjunath. Methanemapper: Spectral absorption aware hyperspectral transformer for methane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17609-17618, 2023. 3, 4, 5", + "[24] Satish Kumar, William Kingwill, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan D Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. In NeurIPS 2022 Workshop on Tackling Climate Change with Machine Learning, 2022. 3", + "[25] Kristian Kvilekval, Dmitry Fedorov, Boguslaw Obara, Ambuj Singh, and BS Manjunath. Bisque: a platform for bioimage analysis and management, Jan. 2023. 2, 4", + "[26] Richard Lamprey, David Ochanda, Rob Brett, Charles Tumwesigye, and Iain Douglas-Hamilton. Cameras replace human observers in multi-species aerial counts in murchison falls, uganda. Remote Sensing in Ecology and Conservation, 6(4):529-545, 2020. 2", + "[27] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma, and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 1", + "[28] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma," + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12602", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: Taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 2", + "[29] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. arXiv preprint arXiv:2303.09030, 2023. 3, 7", + "[30] Julie Linchant, Jonathan Lisein, Jean Semeki, Philippe Lejeune, and Cedric Vermeulen. Are unmanned aircraft systems (uas s) the future of wildlife monitoring? a review of accomplishments and challenges. Mammal Review, 45(4):239-252, 2015. 2", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 3", + "[33] Johannes Naude and Deon Joubert. The aerial elephant dataset: A new public benchmark for aerial object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 48-55, 2019. 3", + "[34] Joseph O Ogutu, Hans-Peter Piepho, Mohamed Y Said, Gordon O Ojwang, Lucy W Njino, Shem C Kifugo, and Patrick W Wargute. Extreme wildlife declines and concurrent increase in livestock numbers in kenya: What are the causes? PloS one, 11(9):e0163249, 2016. 1", + "[35] Wilber K Ottichilo, Jesse Grunblatt, Mohammed Y Said, and Patrick W Wargute. Wildlife and livestock population trends in the kenya rangeland. *Wildlife conservation by sustainable use*, pages 203–218, 2000. 1", + "[36] Wilber K. Ottichilo, Jesse Grunblatt, Mohammed Y. Said, and Patrick W. Wargute. Wildlife and Livestock Population Trends in the Kenya Rangeland, pages 203-218. Springer Netherlands, Dordrecht, 2000. 2", + "[37] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 5", + "[38] Michael John Stephen Peel and Marc Stalmans. The systematic reconnaissance flight (srf) as a tool in assessing the ecological impact of a rural development programme in an extensive area of the lowveld of south africa. African Journal of Ecology, 37(4):449-456, 1999. 1", + "[39] Tinao Petso, Rodrigo S Jamisola Jr, Dimane Mpoeleng, Emily Bennitt, and Wazha Mmereki. Automatic animal identification from drone camera based on point pattern analysis of herd behaviour. Ecological Informatics, 66:101485, 2021. 1, 2", + "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3, 7", + "[41] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021.3", + "[42] AC Seymour, J Dale, M Hammill, PN Halpin, and DW Johnston. Automated detection and enumeration of marine wildlife using unmanned aircraft systems (uas) and thermal imagery. Scientific reports, 7(1):45127, 2017. 1", + "[43] Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal on mathematical analysis, 29(2):511-546, 1998. 5", + "[44] Colin J Torney, David J Lloyd-Jones, Mark Chevallier, David C Moyer, Honori T Maliti, Machoke Mwita, Edward M Kohi, and Grant C Hopcraft. A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images. Methods in Ecology and Evolution, 10(6):779-787, 2019. 1, 2", + "[45] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5", + "[46] Cédric Vermeulen, Philippe Lejeune, Jonathan Lisein, Prosper Sawadogo, and Philippe Bouché. Unmanned aerial survey of elephants. PloS one, 8(2):e54700, 2013. 2", + "[47] Ben G Weinstein, Lindsey Garner, Vienna R Saccomanno, Ashley Steinkraus, Andrew Ortega, Kristen Brush, Glenda Yenni, Ann E McKellar, Rowan Converse, Christopher D Lippitt, et al. A general deep learning model for bird detection in high-resolution airborne imagery. Ecological Applications, 32(8):e2694, 2022. 7", + "[48] Zijing Wu, Ce Zhang, Xiaowei Gu, Isla Duporge, Lacey F. Hughey, Jared A Stabach, Andrew K Skidmore, J Grant C Hopcraft, Stephen J Lee, Peter M Atkinson, et al. Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape. Nature communications, 14(1):3072, 2023. 2, 3", + "[49] Xiangkai Xu, Zhejun Feng, Changqing Cao, Mengyuan Li, Jin Wu, Zengyan Wu, Yajie Shang, and Shubing Ye. An improved swim transformer-based model for remote sensing object detection and instance segmentation. Remote Sensing, 13(23):4779, 2021. 3", + "[50] Yinhui Yu, Xu Sun, and Qing Cheng. Expert teacher based on foundation image segmentation model for object detection in aerial images. Scientific Reports, 13(1):21964, 2023. 7", + "[51] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 5", + "[52] Jiahe Zhu, Xu Chen, Huan Zhang, Zelong Tan, Shengjin Wang, and Hongbing Ma. Transformer based remote sensing object detection with enhanced multispectral feature extraction. IEEE Geoscience and Remote Sensing Letters, 2023. 3", + "[53] Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Ling, and Qinghua Hu. Vision meets drones: A challenge. arXiv preprint arXiv:1804.07437, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3", + "[55] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 7" + ], + "bbox": [ + 78, + 90, + 470, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12604", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_model.json b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e974dcaae4025351a5aa20013b292c4b54a28507 --- /dev/null +++ b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.13, + 0.848, + 0.175 + ], + "angle": 0, + "content": "WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.203, + 0.877, + 0.309 + ], + "angle": 0, + "content": "Satish Kumar\\(^{1*}\\) Bowen Zhang\\(^{1*}\\) Chandrakanth Gudavalli\\(^{1}\\) Connor Levenson\\(^{1}\\) Lacey Hughey\\(^{2}\\) Jared A. Stabach\\(^{2}\\) Irene Amoke\\(^{3}\\) Gordon Ojwang\\(^{4}\\) Joseph Mukeka\\(^{5}\\) Stephen Mwiu\\(^{5}\\) Joseph Ogutu\\(^{6}\\) Howard Frederick\\(^{7}\\) B.S. Manjunath\\(^{1}\\) \n\\(^{1}\\)University of California Santa Barbara \\(^{2}\\)Smithsonian National Zoo and Conservation Biology Institute \\(^{3}\\)Kenya Wildlife Trust, \\(^{4}\\)University of Groningen, \\(^{5}\\)Wildlife Research and Training Institute \\(^{6}\\)University of Hohenheim, \\(^{7}\\)Tanzania Wildlife Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.313, + 0.768, + 0.363 + ], + "angle": 0, + "content": "(satishkumar,bowen68,chandrakanth,clevenson,manj)@ucsb.edu, (stabachj,hugheyl)@si.edu,iirena.amoke@kenyawildlifetrust.org, (gordonojwang,simbamangu)@gmail.com,(jmukeka,smwiu)@wrti.go.ke" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.398, + 0.314, + 0.413 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.473, + 0.707 + ], + "angle": 0, + "content": "We introduce WildlifeMapper (WM), a flexible model designed to detect, locate, and identify multiple species in aerial imagery. It addresses the limitations of traditional, labor-intensive wildlife population assessments that are central to advancing environmental conservation efforts worldwide. While a number of methods exist to automate this process, they are often limited in their ability to generalize to different species or landscapes due to the dominance of homogeneous backgrounds and/or poorly captured local image structures. WM introduces two novel modules that help to capture the local structure and context of objects of interest to accurately localize and identify them, achieving a state-of-the-art (SOTA) detection rate of 0.56 mAP. Further, we introduce a large aerial imagery dataset with more than 11k Images and 28k annotations verified by domain experts. WM also achieves SOTA performance on 3 other publicly available aerial survey datasets collected across 4 different countries, improving mAP by \\(42\\%\\). Source code and trained models are available at Github1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.21, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.868 + ], + "angle": 0, + "content": "This paper introduces WildlifeMapper (WM) - an automated and scalable method for counting wildlife in aerial imagery. Aerial wildlife surveys are recognized as a cornerstone of modern conservation biology. By facilitating large-scale biological monitoring in remote landscapes, this technique has underpinned the ability to track changes in the abundance and distribution of wildlife across open landscapes for decades. However, traditional survey approaches" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.399, + 0.892, + 0.459 + ], + "angle": 0, + "content": "often rely on manual observers to identify, count, and validate species of interest. This labor-intensive process can be time-consuming and error-prone, with potential to limit the utility of final results [27, 35, 38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.672 + ], + "angle": 0, + "content": "Automated approaches offer a promising alternative for efficient and accurate detection of wildlife in aerial survey images. Recent work, for example, illustrates how artificial intelligence has been used to count a variety of species from the air, including antelope in grasslands [44], whales in the ocean [11], and seals on the beach [42]. When combined with advancements in low-cost, high-resolution imaging platforms (e.g., UAVs), these case studies underscore the potential for such data to significantly reduce the effort and cost of traditional wildlife census methods. However, the majority of these techniques struggle to generalize to new species or landscapes due to the dominance of homogeneous backgrounds and poorly captured local structures [7, 12, 18, 39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.794 + ], + "angle": 0, + "content": "WildlifeMapper overcomes these limitations by adapting a novel application of the segment anything transformer model [21]. This model combines high frequency component correlations and spatial correlations in the image data to generate a map of potential locations of objects of interest (i.e., wildlife, livestock). In addition, we address the challenge of identifying multiple species from a relatively small footprint in these images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To demonstrate the WM analysis workflow, we provide a case study example across the Masai Mara Ecosystem in southwestern Kenya. Renowned for its rich biological diversity, the abundance of large mammals (such as buffalo (Syncerus caffer), giraffe (Giraffa tippelskirchi), and wildebeest (Connochaetes taurinus)) have declined precipitously over the past few decades [34]. Importantly, our analy-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.352, + 0.899 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/UCSB-VRL/WildlifeMapper * refers to equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.091, + 0.398, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.091, + 0.641, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.765, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.091, + 0.888, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.893, + 0.315 + ], + "angle": 0, + "content": "Figure 1. Summary of Mara-Wildlife dataset. (a) Satellite view indicating the four flight trajectories, each represented in a different color. (b, c, d, e, f, g) Annotations of (b) zebra, (c) hartebeest, (d) cattle, (e) shoats (sheep and goats), and (f) zebra. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.471, + 0.479 + ], + "angle": 0, + "content": "sis incorporated 11, 151 images of size \\(8400 \\times 5500\\) collected from a digital camera affixed to the bellyport of a Partenavia P68 airplane during Systematic Reconnaissance Flight (SRF) surveys. Part of these images were annotated by trained observers with 28, 146 annotations of 21 species of large mammals (\\(\\geq 15kg\\)), providing an unprecedented opportunity to develop species detection models across a complex, heterogeneous environment. The dataset was systematically verified by trained observers as described in Section 3. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.47, + 0.61 + ], + "angle": 0, + "content": "1. A novel, single-stage end-to-end approach for animal detection. The modules, a High Frequency Feature Generator, a Feature Refiner, and a Query Refiner, work together to improve upon the traditional methods of object detection in aerial imagery and enable generalizability across different habitats. The high frequency features reduce dependence on dominant backgrounds/landscapes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.619, + 0.469, + 0.665 + ], + "angle": 0, + "content": "2. An input patch embedding layer that is specifically designed to capture contextual information to help in identifying individual animal species." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.469, + 0.765 + ], + "angle": 0, + "content": "3. The release of a new benchmark dataset via the data owner (Kenya's Wildlife Research and Training Institute - WRTI) once all approvals are in place. An international user community is already engaged in further enhancing these data and using the WM through the BisQue [25] platform." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.488, + 0.47, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Community Adoption The practical use of WildlifeMapper (WM) extends beyond theoretical and computational success and is operationalized through BisQue [25]. WM is made available to users through a series of training modules that demonstrate how to \\((i)\\) upload digital imagery, \\((ii)\\) create annotations, \\((iii)\\) apply and/or improve existing models, \\((iv)\\) evaluate model fit, \\((v)\\) improve annotations and refit models, and \\((vi)\\) generate summary statistics. In the fu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.893, + 0.405 + ], + "angle": 0, + "content": "ture, we envision WM to be of great value to ecologists, wildlife managers, and government officials, providing accurate information about the state of wildlife populations in near real-time, facilitating decision-making processes, and improving the conservation of ecosystems globally." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.412, + 0.65, + 0.427 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Manual Methods: Aerial surveys using Front- and Rear-Seat Observers (FSO and RSO, respectively) are commonly used to inventory wildlife populations across open landscapes [36]. However, several important biases can impact these counts, including the experience level and fatigue of the human observers [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.893, + 0.689 + ], + "angle": 0, + "content": "Deep Learning Methods: To address these issues, researchers have begun incorporating digital cameras on piloted aircraft and UAVs [18,30,46]. This minimizes the influence of observer bias while increasing transparency and reproducibility of results. For example, [26] replaced RSOs with an oblique camera mount system minimized the influence of observer bias while producing comparable estimates of large mammals under partial canopy cover. Similarly, [44] used a nadir mounted camera to improve the accuracy and efficiency of manual counts of large antelope in an open grassland ecosystem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.825 + ], + "angle": 0, + "content": "However, detecting animals in the wild from aerial imagery poses many challenges. For example, most publicly available datasets for aerial object detection are focused on identifying relatively distinct features such as buildings, roads, vehicles, and other man-made structures [8, 14, 53]. Animals, tend to blend in with their surroundings [13], can be occluded by trees, exhibit considerable variation in color and pattern, or have behavioral adaptations that make them difficult to detect [7, 20, 39, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[20] proposed a solution for this problem involving a two-branch CNN model based on AlexNet to perform animal recognition and localization. [7] evaluated three state-of-the-art object detection algorithms, including Faster-RCNN, Libra-RCNN, and RetinaNet on six African wild" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12595" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.892, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.207, + 0.856, + 0.221 + ], + "angle": 0, + "content": "Figure 2. Distribution of \\((\\geq 15kg)\\) mammals identified in digital imagery collected across the Masai Mara Ecosystem, Kenya." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.225, + 0.894, + 0.356 + ], + "angle": 0, + "content": "
Dataset# of annot. images# of annot. tiles# of species# of annotationImage sizeGSD (cm)Location
Virunga73930069656646000x40002.4DRC
Garamba1586429616116000x40002.0DRC
AED2067693871155815500x36002.4-13.0Bostwana, Namibia, South Africa
Mara-Wildlife10127796621281468256x55041.45Masai Mara National Reserve
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.358, + 0.893, + 0.401 + ], + "angle": 0, + "content": "Table 1. Comparison of Mara-Wildlife dataset with other publicly available dataset. Mara-Wildlife dataset has \\( \\times 3 \\) more unique species than the total of all other datasets. Each image is significantly larger and higher ground resolution making 77k unique images of size \\( {1024} \\times {1024} \\) with 21 different animal species. GSD: ground sampling distance; DRC: Democratic Republic of Congo." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.469, + 0.534 + ], + "angle": 0, + "content": "mammals. All three algorithms, however, showed poor performance in animal detection when animals were grouped closely together in herds. [12] adopted a segmentation approach, employing a UNet model to detect livestock from drone imagery. [48] uses a comparable model to analyze high-resolution satellite imagery, producing segmentation masks of wildebeest-sized animals, which are subsequently utilized for detection and counting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Transformers: WM adopts a transformer architecture based on past success in modeling different types of aerial imagery. Examples include incorporating multispectral imagery for change detection [4], landcover classification [16], greenhouse gas (GHG) emission detection [23,24] and RGB aerial imagery for object detection [29, 49, 52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.779 + ], + "angle": 0, + "content": "The most effective applications of transformer-based models have been tailored for standard object detection tasks [3, 5, 32, 41, 54]. These works leverage the self-attention to model dependencies among the patches in an end-to-end fashion, unlike CNN-based models [1, 40]. However, when directly applied to aerial imagery, these models cannot effectively exploit the local structures as they divide the image into a sequence of patches. This limits the detection of small-scale objects in a homogeneous and dominant background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Dataset: The existing publicly available animal aerial imagery datasets are listed in Table 1. The Virunga dataset [7] was collected in Virunga National Park, Democratic Republic of Congo (DRC). This dataset contains 897 annotated images of 6 animal species. The Garamba dataset was collected in Garamba National Park, DRC and contains 7034 images. Only 158 images have been annotated, containing 7 animal species. The aerial elephant dataset (AED) [33] was" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.474 + ], + "angle": 0, + "content": "collected across a mosaic of woodland, open shrubland, and grassland habitats in Botswana, Namibia, and South Africa. Only a single species (i.e., elephant) was targeted during SRF surveys. See Table 1 for details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.715, + 0.5 + ], + "angle": 0, + "content": "3. Mara-Wildlife Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.585 + ], + "angle": 0, + "content": "The Mara-Wildlife dataset is a distinctive dataset that captures the essence of the Masai Mara ecosystem through a compilation of 77966 images of size \\(1024 \\times 1024\\). This habitat is heterogeneous, including woodland, shrubland, and grassland vegetation with 21 unique animal species." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.667, + 0.607 + ], + "angle": 0, + "content": "3.1. Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Flightline Details: Data collection was in collaboration with the Smithsonian National Zoo and Conservation Biology Institute (SNZCBI), Kenya's Wildlife Research and Training Institute (WRTI), the Kenya Wildlife Trust (KWT), and the Directorate of Resource Surveys and Remote Sensing (DRSRS). In March 2022, we fitted a Partenavia P68 with a Nikon D850 digital camera and collected high resolution \\((8256\\times 5504)\\) digital images. During data acquisition, the aircraft adhered to a predetermined flight trajectory, depicted in Figure 1, at 400 ft above ground level (agl). This trajectory was optimized to encompass open grassland areas across the Masai Mara ecosystem, including the Masai Mara National Reserve, 22 adjacent private conservancies, and unprotected peripheral areas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The aerial survey was conducted during a wet season period when the Serengeti migratory population of wildebeest have moved southward to locate more suitable forage in Tanzania. Thus, the survey primarily captured resident species, including wildebeest, zebra, topi, hartebeest, gi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.248, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Overview of WildlifeMapper (WM) architecture. Given an input image of size \\(1024 \\times 1024 \\times 3\\), the High-Frequency Feature Generator (HFG) module generates information about potential location of object of interest. The Feature Refiner (FR) takes these potential locations along with contextual features from Patch Embed layer and sent output to Image Encoder. In parallel, the Query Refiner (QR) incorporates the output of FR to refine learnable queries. Finally these queries are decoded using encoded features from Image Encoder and predict bounding box and class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.394 + ], + "angle": 0, + "content": "raffe, and other large (\\(\\geq 15\\mathrm{kg}\\)) antelope. Data collection was conducted in the early mornings (prior to 10:00 EAT) and late afternoons (after 15:00 EAT) when lighting conditions were optimal and animals are most active." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.469, + 0.516 + ], + "angle": 0, + "content": "The geographical positioning of each image was acquired through a GPS system that recorded the plane's altitude, speed, and geographical coordinates. These data were then synchronized with the image using the image capture timestamps, enabling us to determine the geographic location of the centroid of each image. The camera was placed in the bellyport of the airplane, capturing a nadir view of the landscape every two seconds along the flight path." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.524, + 0.254, + 0.54 + ], + "angle": 0, + "content": "3.2. Image Annotation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.469, + 0.633 + ], + "angle": 0, + "content": "Initial bounding box annotations (21796) generated using AIDE platform [19] were exported in CSV format. These were then imported into BisQue [25] for further validation and correction. Finally, these annotations underwent validation by a single trained observer specializing in ecology, resulting in 28146 annotations in total." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.245, + 0.655 + ], + "angle": 0, + "content": "3.3. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The Mara-Wildlife dataset showcases a detailed assemblage of wildlife, inclusive of 21 distinct species classes. The dataset is composed of approximately 77,966 tiled images, derived from 1,012 original rasters (Table. 1) The meticulous process of annotation has culminated in labeling 28,146 animals. Species identified include domestic cattle (Bos taurus), white-bearded wildebeest (Connochaetes taurinus), topi (Damaliscus lunatus), shoats (domesticated sheep and goats), kongoni (Alcelaphus buselaphus), waterbuck (Kobus ellipsiprymnus), impala (Aepyceros melampus), Grant's gazelle (Nanger granti), Thomson's gazelle (Eudorcas thomsonii), Cape buffalo (Syncerus caffer), zebra (Equus quagga), ostrich (Struthio camelus), Masai giraffe (Giraffa tippelskirchi), warthog (Phacochoerus africanus), eland (Taurotragus oryx), donkey (Equus africanus), hyena (Crocuta crocuta), hippopotomus" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.365 + ], + "angle": 0, + "content": "(Hippopotamus amphibius), lion (Panthera leo), and elephant (Loxodonta africana)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.373, + 0.771, + 0.39 + ], + "angle": 0, + "content": "4. WildlifeMapper Architecture" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.399, + 0.688, + 0.414 + ], + "angle": 0, + "content": "4.1. Technical Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.69 + ], + "angle": 0, + "content": "WildlifeMapper's architecture is inspired by the success of the Segment Anything Model (SAM) [21], created to segment small/large (all sizes) of objects. Referring to Fig. 3, WM contains the following main components: (i) A patch embedding layer designed to capture long-range context, (ii) a High-Frequency Feature Generator (HFG), (iii) a Feature Refiner (FR) followed by ViT based image encoder [21,23,45], and (iv) a Query Refiner (QR) module followed by a box decoder module. The input image is first processed through two separate branches, the patch embedding layer which captures long-range context [9] and HFG which suppresses all the low-frequency components in the image and generates a feature embedding. The HFG (Sec. 4.3) exploits prior knowledge that aerial images from areas such as forests, grasslands, and shrublands have a homogeneous and dominant background representing the dominant low frequency image content. Fig. 4 shows that on suppressing the lower frequencies, the object of interest is easy to locate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.827 + ], + "angle": 0, + "content": "The FR (Sec. 4.4) takes the embeddings from each of the two branches and generates a high quality embedding that contains information about potential locations of animals and captures the local context. The QR modules refines a set of learnable queries using the location information from FR module. These refined queries are passed to the box decoder. The box decoder takes the refined queries and encoded features image encoder to generate the final detection box and class of the object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.839, + 0.64, + 0.853 + ], + "angle": 0, + "content": "4.2. Patch Embed" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The patch embedding layer utilizes a larger kernel convolution with an increasing dilation rate. This design rapidly expands the receptive field, allowing explicit extrac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "tion of features rich in contextual information. This approach is particularly beneficial for aerial imagery, where the small sized object makes classification based on appearance alone challenging. Contextual information thus becomes crucial for the accurate recognition of these objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.176, + 0.437, + 0.192 + ], + "angle": 0, + "content": "4.3. High-frequency Feature Generator (HFG)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.195, + 0.469, + 0.329 + ], + "angle": 0, + "content": "Along with patch embedding, the input image is processed in parallel by the HFG module to generate features with information about the location of the animal or cluster as shown in Fig. 4. The HFG module is inspired from the limitation of ViT models [45]. ViT models face challenges in efficiently utilizing local structures. They segment an image into patches and apply self-attention to model relationships, but this approach often falls short in capturing detailed local features [37, 51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.469, + 0.435 + ], + "angle": 0, + "content": "Local features in images are closely linked to high-frequency components [2, 43]. We hypothesize that suppressing low-frequency components can mitigate the influence of a dominant homogeneous background. To test this, we performed a discrete Fourier Transform (DFT) on the images, filtering out the low-frequency components before reconstructing the images, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.469, + 0.511 + ], + "angle": 0, + "content": "For a given input image \\(I \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(C\\) is channel dimension, we compute Discrete Fourier Transform (DFT) of \\(I\\). In next step we suppress the low frequency components with a controlling parameter and construct the image \\(I\\) with inverse DFT (IDFT)." + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.522, + 0.469, + 0.539 + ], + "angle": 0, + "content": "\\[\nI ^ {\\prime} = I D F T [ h p f (D F T (I) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.469, + 0.61 + ], + "angle": 0, + "content": "where \\(hpf\\) is a high pass filter. Then we reduce the dimension of the reconstructed image \\(I'\\) via an embedding layer to generate embedding \\(hfc_{emb}\\) and pass them to the FR module. See supplementary materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.275, + 0.634 + ], + "angle": 0, + "content": "4.4. Feature Refiner (FR)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.469, + 0.727 + ], + "angle": 0, + "content": "Next, the features from the patch embedding layer \\((f_{img})\\) and HFG \\((hfc_{emb})\\) are fed to the FR module. The \\(f_{img}\\) are refined with the \\(hfc_{emb}\\) via cross-attention mechanism. The FR is a simple module with cross-attention and linear layers [45]. The output contains information about the potential location of the object and the long-range context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Following the standard architecture of SAM [21], we pass FR output to our ViT based image encoder supplemented with learnable positional embeddings \\( p \\). The encoded feature map is \\( f_{emb} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.799, + 0.469, + 0.815 + ], + "angle": 0, + "content": "\\[\nf _ {e m b} = \\operatorname {V i T} \\left[ \\mathbf {F R} \\left(f _ {i m g}, h f c _ {e m b}\\right), p \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Query Refiner (QR): The QR follows a transformer decoder like architecture and takes as input a set of 100 learnable queries \\( Q \\in \\mathbb{R}^{100 \\times d} \\) and output of FR module. Here \\( d = 256 \\) is same as channel dimension of \\( f_{emb} \\) from image encoder. The FR output refines the \\( Q \\) via a cross-attention" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.892, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.411, + 0.894, + 0.55 + ], + "angle": 0, + "content": "Figure 4. The sample output visualization from the High-Frequency Feature Generator (HFG) module. The illustration shows the effectiveness of the module in suppressing the homogeneous and dominant background, while highlighting objects of interest (i.e., animals). The top image shows bomas, natural structures constructed to contain livestock, and paths that have been suppressed. Animals, however, are clearly identified, especially inside the boma. The bottom image shows a water body (a dam created for livestock) that has been suppressed by the module. Animals can again be highlighted throughout the image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.57, + 0.892, + 0.615 + ], + "angle": 0, + "content": "mechanism. The refined queries narrows the search space for box decoder module to accurately locate and identify object of interest. [15,23]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.635, + 0.652 + ], + "angle": 0, + "content": "4.5. Box Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Next, the refined queries are sent to the box decoder module. We concatenate these queries with the \\( f_{emb} \\) and pass box decoder's self attention layer as \\( qf_{emb} \\). This is inspired from the idea of class_token used by Vaswani et.al. [45] to make decoding process memory efficient. The transformer network takes 4 input variables, those are position embeddings of queries and image embeddings, \\( qf_{emb} \\), and \\( f_{emb} \\). Our transformer model uses two-way attention inspired from [3, 6, 21] and our box decoder uses self-attention and cross-attention in two directions (queries-to-image embedding and vice-versa) to update all embeddings. We keep the box decoder very light weight (two blocks). The top 100 (equal to number of queries) indexes from the output of the final block is passed to two separate MLP blocks to regress the output bounding box prediction and class of the predicted box." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12598" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.293, + 0.108 + ], + "angle": 0, + "content": "4.6. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.298 + ], + "angle": 0, + "content": "We've trained WM using a single-stage, end-to-end approach to determine bounding boxes and classify them. The loss strategy we applied for WM is akin to what's used in DETR [3]. Initially, we perform bipartite matching to align our model's predictions with the actual bounding box data. Then, we proceed to compute the loss for these matched pairs. To achieve the best possible match between our predictions and the real data, we use the Hungarian algorithm [22]. Once matched, each prediction is paired with its respective ground truth. We then measure the \\( l_{1} \\) (L1 distance) and \\( GIoU \\) loss for the bounding box and the cross entropy loss for the classification [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.369 + ], + "angle": 0, + "content": "Inference: The inference pipeline is straightforward and similar to training code. During inference, we first filter the detections at \\(50\\%\\) threshold and then use non-maximal suppression to remove any overlapping boxes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.382, + 0.21, + 0.399 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.543 + ], + "angle": 0, + "content": "Train-val-test split: There is no data leak in train-val-test split. Fig. 1 shows the flight paths with location of each image represented by a circle. Each color coded flight path represents the train-val-test image set. No images taken while the airplane was cutting across the transects plus turnaround points as can be seen there are no circles. We created a spatial disjoint of \\(20km\\) distance between the transects as shown in Fig. 1 and achieved consistent performance (0.56mAP)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Evaluation Metrics: We report our performance on multiple metrics. Following the protocols of standard object detection, we report the performance in mean average-precision (mAP) for detection and mean intersection-overunion (mIOU) for localization of animals. We also report a commonly used metric by the ecologists on the team, the class-wise mean absolute error (MAE) indicating the counting accuracy of each species:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.686, + 0.47, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M A E} = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\sum_ {c = 1} ^ {C} | \\hat {n} _ {i, c} - n _ {i, c} |, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.471, + 0.786 + ], + "angle": 0, + "content": "Where \\(I\\) is the number of images, \\(C\\) is the number of classes, \\(\\hat{n}_{i,c}\\) and \\(n_{i,c}\\) are the predicted and ground truth counts for class \\(c\\) in image \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Implementation Details: Each image taken from the drone is \\(8256 \\times 5506 \\times 3\\). We create tiles for each image in the spatial domain, with the size of \\(1024 \\times 1024 \\times 3\\) with \\(25\\%\\) of overlap. The Patch Embed layer uses a single CNN layer with a large kernel of size \\(16 \\times 16\\) with stride 16. In the parallel branch, the High-Frequency Feature Generator, we use DFT to compute the Fourier transform, the mask is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.321 + ], + "angle": 0, + "content": "a binary disk with the radius set to 128. The HFC Embed layer uses 3 CNN layers with ReLU activation with a kernel of size \\(3 \\times 3\\) and a global average pool at the end. The Feature Refiner (FR) module consists of one cross attention layer with 1 linear layer. The image encoder is a pre-trained ViT model [21] with 24 transformer layers and 16 heads. The Query Refiner (QR) module takes in 100 queries each of channel dimension 256, those are cross attended with \\(hfc_{emb}\\) output. The box decoder contains 3 layers of two-way attention with 8 heads. We train WM with AdamW optimizer [31] setting the learning rate to \\(10^{-4}\\) for the FR, QR and box decoder with a weight decay to \\(10^{-4}\\). We set the learning rate for the Patch Embed and HFC Embed layer to \\(10^{-5}\\). We load the image encoder with pre-trained weights from segment anything [21] and keep it frozen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Data Augmentation: In the train and test datasets, we incorporated an equal number of images without any objects to assess the model's robustness against empty background images. We applied multiple data augmentation techniques, including HSV (hue, saturation, and value) \\((10\\%)\\), rotation \\((5\\%)\\), translation \\((10\\%)\\), affine transformation \\((20\\%)\\), scale \\((10\\%)\\), shear \\((5\\%)\\) and mosaic \\((70\\%)\\) augmentation [1]. Mosaic augmentation is proven to be the most effective, with an improvement of \\(0.07\\mathrm{mAP}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.894, + 0.571 + ], + "angle": 0, + "content": "Hard Negative Mining: After training for 100 epochs, we take all the False Positives (FP) predictions having \\(\\mathrm{IOU} \\leq 0.10\\) with ground truth box and mark them as background class. Then fine-tuning for 20 epochs improved the performance of FP reduction for detecting rocks, dead tress or other artifacts on ground as animal." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.587, + 0.601 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.736, + 0.623 + ], + "angle": 0, + "content": "6.1. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We trained WM separately on Mara-Wildlife dataset and Virunga-Garamba-AED dataset for comparison with existing works, see Table 2 for a summary of the results on all of the tested datasets. The \\(mAP\\) values are compared for IoU of 0.50-0.95 and 0.50. We also provide the average counting error in animal counting per image. We trained all the baseline models with the default set of conditions on Mara-Wildlife dataset. We merged the Virunga, Garamba, and AED datasets and created the train-val-test split according to [7]. The combination of these datasets contains 6 unique animal species and diversity of landscapes such as woodland, savannahs, open shrubland, and grasslands across multiple countries – Democratic Republic of Congo (DRC), Botswana, Namibia, and South Africa. WM outperforms all methods by a significant margin as shown in Table 2. We note that in [7] the authors did not make the code base or trained model public, hence we could not verify the results. We implemented these methods from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12599" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.47, + 0.311 + ], + "angle": 0, + "content": "
Methods#epochsmAPmAP50Counting Error
Mara Wildlife Dataset
1Faster-Rcnn [40]1000.240.582.59
2DETR [3]2000.220.572.75
3Co-DETR-R50 [55]1000.270.662.72
4Co-DETR-swingL [55]1000.280.652.60
5Yolo v5 [10]1000.300.672.12
6Yolo v8 [17]1000.270.613.97
7LSKNet [29]1000.290.66-
6DroneDetect [50]1000.180.48-
8WildlifeMapper1200.560.791.9
Virunga-Garamba-AED Datasets
6Faster-Rcnn1200.340.650.27
7DETR2000.300.620.45
8Yolo v51000.480.780.12
9Yolo v81000.480.770.42
10WildlifeMapper800.680.850.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.321, + 0.47, + 0.405 + ], + "angle": 0, + "content": "Table 2. Comparison with baseline models. The top section shows performance on species detection on Mara-wildlife dataset and low section shows performance on the mixed dataset from Virunga-Garamba-AED dataset. The overall detection accuracy is generally higher in Virunga-Garamba-AED dataset because there are only 6 species and the terrain is quite similar in all images." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.405, + 0.395, + 0.482 + ], + "angle": 0, + "content": "
HFGFRQRmAP
WMXX*0.46
✓✓*0.54
✓X✓0.49
✓✓✓0.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Table 3. HFG module effectiveness in refining the image features and queries. “X” represents not used, “√” represents used and “*” represents that random queries are used but there was not refining with HFG features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.727 + ], + "angle": 0, + "content": "the original public repositories and trained according to the training strategy detailed in [7]. We attribute the model poor learning performance due to salient features of the homogeneous background being learned more than the object of interest. The detection of the object of interest is then dependent on the landscape properties instead of object properties. Hence when used on a slight variations of landscapes for the same object, the models struggle to detect. This limitation is specifically addressed in the WM, where the HFG modules suppresses the background and highlights the object of interest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.798 + ], + "angle": 0, + "content": "Qualitative results: Fig. 5 shows the quality of detection by WM in different scenarios. Those include, detection when animal is partially visible under a tree, or a big clustering. WM makes correct predictions in varying scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.242, + 0.822 + ], + "angle": 0, + "content": "6.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.469, + 0.861 + ], + "angle": 0, + "content": "We performed all ablation experiments on Mara-Wildlife (MW) dataset and validate the design choices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "High-frequency Feature Generator Module: In Table 3, we show the effectiveness of the HFG module. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.334 + ], + "angle": 0, + "content": "experimented with HFG's output in 3 ways: first, we passed HFG's output to Feature Refiner (FG) module. It leads to significant improvement in detection by \\(0.09\\mathrm{mAP}\\) over the baseline. This demonstrates that providing potential location candidates features to image encoder module produce better embeddings. Second, we pass the HFG's output to Query Refiner (QR) module only. This leads to an improvement of \\(0.03\\mathrm{mAP}\\) over baseline. This shows the effectiveness of guiding queries with location candidates features. In the third case, we passed the HFG's output to both FR and QR modules and achieved an improvement of \\(0.11\\mathrm{mAP}\\) over the baseline. We hypothesize that this reduces the dominance of features from homogeneous and dominant background in aerial imagery. The also observed this while testing WM across flightlines different types of terrain such as green grasslands, dry grasslands, and forest areas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Feature Refiner Module: We tested \\( hfc_{emb} \\) and \\( f_{img} \\) merging by 3 ways: addition, concatenation and cross-attention. Cross-attention is most effective, because with addition and concatenation, the \\( hfc_{emb} \\) get lost, while cross-attention generates better embeddings giving attention to potential location candidates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Kernel Size: We observed that a larger kernel size of \\(31 \\times 31\\) results in reduced misclassification. For example, a topi or warthog cannot be found inside a boma because only domestic species are kept in bomas. So context helps in making the right class detection. We observed an improvement of \\(0.02\\mathrm{mAP}\\). Experiments were done in 3 kernel sizes. 1. \\(7 \\times 7\\): \\(0.55\\mathrm{mAP}\\), 2. \\(16 \\times 16\\): \\(0.558\\mathrm{mAP}\\); 3. \\(32 \\times 32\\): \\(0.57\\mathrm{mAP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Query Refiner Module: We experimented with only providing random queries and guiding the queries with HFG module output. We merged them with direct addition or concatenation and cross-attention. With cross-attention, we observed an improvement in performance of 0.03mAP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Geographic generalization: To test the geographic generalizability of WM across different terrains, we trained WM only on images from Kenya; and tested on images from Democratic Republic of Congo, Botswana and Namibia. The test was done on 4 common species which were present in both the ecosystems. WM achieved a detection performance of \\(0.48\\mathrm{mAP}\\). This shows the adaptability of WM across varying landscapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.892, + 0.909 + ], + "angle": 0, + "content": "Domain generalization: We train-test WM on a different domain, a bird species tern [47], commonly found on/near water bodies. Live in huge clusters. WM achieved the accuracy of \\(0.71\\mathrm{mAP}\\). Showing adaptability of WM across different domains." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.357, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.089, + 0.613, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.089, + 0.872, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.289, + 0.356, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.289, + 0.613, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.289, + 0.872, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. The top row highlights examples of crowded and partially occluded scenes. Row-1, Column-1 & Row-3, Column-3 shows examples where animals are obstructed by shadows. The zoomed-in box in Row-3, Column-3 shows a zebra partially occluded. The bounding box color is coded according to class names: shoats- \"hot pink\", cattle- \"deep sky blue\", zebra- \"light yellow\"." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.559, + 0.273, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.559, + 0.458, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.71, + 0.472, + 0.768 + ], + "angle": 0, + "content": "Figure 6. Failure cases. Left shows an example where animals are occluded by shadows and are difficult to detect. Right shows an example of rock detected as an impala, emphasizing the difficulty in differentiating objects in the image from animals of interest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Failure Cases: The detections from WM are inaccurate when animals are clustered in shadows, such as when animals are located inside bomas and the sun angle makes a strong shadow on the enclosure. These are some of the difficult cases shown in Fig. 6. Other cases of false positives are the small rocks or trees that sometimes resemble animals. Some examples are shown in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.62, + 0.577 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.723 + ], + "angle": 0, + "content": "This paper presents WildlifeMapper (WM) - a transformer based approach for the detection of animals of varying densities and sizes across natural backgrounds. The WM utilizes a high frequency features generator, feature refiner, and query refiner to accurately locate and classify 8 animal species. WM stands to significantly improve the efficiency and accuracy of wildlife monitoring and conservation efforts. Future work will extend this model and dataset to a larger number of species and habitats." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.736, + 0.681, + 0.753 + ], + "angle": 0, + "content": "8. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.868 + ], + "angle": 0, + "content": "This research is partially supported by the following grants: NSF award SI2-SSI #1664172 (BSM), the Ohrstrom Family Foundation, and a Smithsonian National Zoo and Conservation Biology Institute (SNZCBI) fellowship (S.Kumar). We also recognize the contributions of the annotation team, including S. Harman, A. Gray, C. Pate, C. Obath, S. Nyarangi, C. Kagume, and W. Sairowua." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12601" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Fergus W Campbell and John G Robson. Application of fourier analysis to the visibility of gratings. The Journal of physiology, 197(3):551, 1968. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[5] Qiang Chen, Jian Wang, Chuchu Han, Shan Zhang, Zexian Li, Xiaokang Chen, Jiahui Chen, Xiaodi Wang, Shuming Han, Gang Zhang, et al. Group detr v2: Strong object detector with encoder-decoder pretraining. arXiv preprint arXiv:2211.03594, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.521 + ], + "angle": 0, + "content": "[7] Alexandre Delplanque, Samuel Foucher, Philippe Lejeune, Julie Linchant, and Jérôme Théau. Multispecies detection and identification of african mammals in aerial imagery using convolutional neural networks. Remote Sensing in Ecology and Conservation, 8(2):166-179, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44(11):7778-7796, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[9] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[10] Glenn Jocher et. al. ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements, Oct. 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[11] Emilio Guirado, Siham Tabik, Marga L Rivas, Domingo Alcaraz-Segura, and Francisco Herrera. Whale counting in satellite and aerial images with deep learning. Scientific reports, 9(1):14259, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Liang Han, Pin Tao, and Ralph R Martin. Livestock detection in aerial images using a fully convolutional network. Computational Visual Media, 5:221-228, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Tracey Hollings, Mark Burgman, Mary van Andel, Marius Gilbert, Timothy Robinson, and Andrew Robinson. How do you find the green sheep? a critical review of the use of remotely sensed imagery to detect and count animals. Methods in Ecology and Evolution, 9(4):881-892, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H. Hsu. Drone-based object counting by spatially regularized re" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "gional proposal networks. In The IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. In Pattern Recognition and Tracking XXXIV, volume 12527, pages 192-205. SPIE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.848, + 0.204 + ], + "angle": 0, + "content": "[16] Johannes Jakubik et. al. Prithvi-100M, Aug. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.891, + 0.231 + ], + "angle": 0, + "content": "[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics* YOLO, Jan. 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Benjamin Kellenberger, Diego Marcos, and Devis Tuia. Detecting mammals in uav images: Best practices to address a substantially imbalanced dataset with deep learning. Remote Sensing of Environment, 216:139-153, Oct 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[19] Benjamin Kellenberger, Devis Tuia, and Dan Morris. Aide: Accelerating image-based ecological surveys with interactive machine learning. Methods in Ecology and Evolution, 11(12):1716-1727, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[20] Benjamin Kellenberger, Michele Volpi, and Devis Tuia. Fast animal detection in uav images using convolutional neural networks. In 2017 IEEE international geoscience and remote sensing symposium (IGARSS), pages 866-869. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Satish Kumar, Ivan Arevalo, ASM Iftekhar, and BS Manjunath. Methanemapper: Spectral absorption aware hyperspectral transformer for methane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17609-17618, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[24] Satish Kumar, William Kingwill, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan D Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. In NeurIPS 2022 Workshop on Tackling Climate Change with Machine Learning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[25] Kristian Kvilekval, Dmitry Fedorov, Boguslaw Obara, Ambuj Singh, and BS Manjunath. Bisque: a platform for bioimage analysis and management, Jan. 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[26] Richard Lamprey, David Ochanda, Rob Brett, Charles Tumwesigye, and Iain Douglas-Hamilton. Cameras replace human observers in multi-species aerial counts in murchison falls, uganda. Remote Sensing in Ecology and Conservation, 6(4):529-545, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma, and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12602" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: Taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. arXiv preprint arXiv:2303.09030, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[30] Julie Linchant, Jonathan Lisein, Jean Semeki, Philippe Lejeune, and Cedric Vermeulen. Are unmanned aircraft systems (uas s) the future of wildlife monitoring? a review of accomplishments and challenges. Mammal Review, 45(4):239-252, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[33] Johannes Naude and Deon Joubert. The aerial elephant dataset: A new public benchmark for aerial object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 48-55, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[34] Joseph O Ogutu, Hans-Peter Piepho, Mohamed Y Said, Gordon O Ojwang, Lucy W Njino, Shem C Kifugo, and Patrick W Wargute. Extreme wildlife declines and concurrent increase in livestock numbers in kenya: What are the causes? PloS one, 11(9):e0163249, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[35] Wilber K Ottichilo, Jesse Grunblatt, Mohammed Y Said, and Patrick W Wargute. Wildlife and livestock population trends in the kenya rangeland. *Wildlife conservation by sustainable use*, pages 203–218, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[36] Wilber K. Ottichilo, Jesse Grunblatt, Mohammed Y. Said, and Patrick W. Wargute. Wildlife and Livestock Population Trends in the Kenya Rangeland, pages 203-218. Springer Netherlands, Dordrecht, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[37] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[38] Michael John Stephen Peel and Marc Stalmans. The systematic reconnaissance flight (srf) as a tool in assessing the ecological impact of a rural development programme in an extensive area of the lowveld of south africa. African Journal of Ecology, 37(4):449-456, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[39] Tinao Petso, Rodrigo S Jamisola Jr, Dimane Mpoeleng, Emily Bennitt, and Wazha Mmereki. Automatic animal identification from drone camera based on point pattern analysis of herd behaviour. Ecological Informatics, 66:101485, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[42] AC Seymour, J Dale, M Hammill, PN Halpin, and DW Johnston. Automated detection and enumeration of marine wildlife using unmanned aircraft systems (uas) and thermal imagery. Scientific reports, 7(1):45127, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[43] Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal on mathematical analysis, 29(2):511-546, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[44] Colin J Torney, David J Lloyd-Jones, Mark Chevallier, David C Moyer, Honori T Maliti, Machoke Mwita, Edward M Kohi, and Grant C Hopcraft. A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images. Methods in Ecology and Evolution, 10(6):779-787, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[45] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[46] Cédric Vermeulen, Philippe Lejeune, Jonathan Lisein, Prosper Sawadogo, and Philippe Bouché. Unmanned aerial survey of elephants. PloS one, 8(2):e54700, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[47] Ben G Weinstein, Lindsey Garner, Vienna R Saccomanno, Ashley Steinkraus, Andrew Ortega, Kristen Brush, Glenda Yenni, Ann E McKellar, Rowan Converse, Christopher D Lippitt, et al. A general deep learning model for bird detection in high-resolution airborne imagery. Ecological Applications, 32(8):e2694, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[48] Zijing Wu, Ce Zhang, Xiaowei Gu, Isla Duporge, Lacey F. Hughey, Jared A Stabach, Andrew K Skidmore, J Grant C Hopcraft, Stephen J Lee, Peter M Atkinson, et al. Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape. Nature communications, 14(1):3072, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[49] Xiangkai Xu, Zhejun Feng, Changqing Cao, Mengyuan Li, Jin Wu, Zengyan Wu, Yajie Shang, and Shubing Ye. An improved swim transformer-based model for remote sensing object detection and instance segmentation. Remote Sensing, 13(23):4779, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[50] Yinhui Yu, Xu Sun, and Qing Cheng. Expert teacher based on foundation image segmentation model for object detection in aerial images. Scientific Reports, 13(1):21964, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[51] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[52] Jiahe Zhu, Xu Chen, Huan Zhang, Zelong Tan, Shengjin Wang, and Hongbing Ma. Transformer based remote sensing object detection with enhanced multispectral feature extraction. IEEE Geoscience and Remote Sensing Letters, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Ling, and Qinghua Hu. Vision meets drones: A challenge. arXiv preprint arXiv:1804.07437, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12603" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[55] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12604" + } + ] +] \ No newline at end of file diff --git a/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_origin.pdf b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..50f82db3cd33fb3e245b732908976cb9b3234f04 --- /dev/null +++ b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/75c6b881-4935-4fc9-8673-9dc09a1854b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fb121533a29043271dfc025c2531d249638a5ee3e768bb2d05208503981867d +size 4983844 diff --git a/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/full.md b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ed2c60d1367059f6e43eeeb367159264721bb70f --- /dev/null +++ b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/full.md @@ -0,0 +1,298 @@ +# WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification + +Satish Kumar $^{1*}$ Bowen Zhang $^{1*}$ Chandrakanth Gudavalli $^{1}$ Connor Levenson $^{1}$ Lacey Hughey $^{2}$ Jared A. Stabach $^{2}$ Irene Amoke $^{3}$ Gordon Ojwang $^{4}$ Joseph Mukeka $^{5}$ Stephen Mwiu $^{5}$ Joseph Ogutu $^{6}$ Howard Frederick $^{7}$ B.S. Manjunath $^{1}$ $^{1}$ University of California Santa Barbara $^{2}$ Smithsonian National Zoo and Conservation Biology Institute $^{3}$ Kenya Wildlife Trust, $^{4}$ University of Groningen, $^{5}$ Wildlife Research and Training Institute $^{6}$ University of Hohenheim, $^{7}$ Tanzania Wildlife Research Institute + +(satishkumar,bowen68,chandrakanth,clevenson,manj)@ucsb.edu, (stabachj,hugheyl)@si.edu,iirena.amoke@kenyawildlifetrust.org, (gordonojwang,simbamangu)@gmail.com,(jmukeka,smwiu)@wrti.go.ke + +# Abstract + +We introduce WildlifeMapper (WM), a flexible model designed to detect, locate, and identify multiple species in aerial imagery. It addresses the limitations of traditional, labor-intensive wildlife population assessments that are central to advancing environmental conservation efforts worldwide. While a number of methods exist to automate this process, they are often limited in their ability to generalize to different species or landscapes due to the dominance of homogeneous backgrounds and/or poorly captured local image structures. WM introduces two novel modules that help to capture the local structure and context of objects of interest to accurately localize and identify them, achieving a state-of-the-art (SOTA) detection rate of 0.56 mAP. Further, we introduce a large aerial imagery dataset with more than 11k Images and 28k annotations verified by domain experts. WM also achieves SOTA performance on 3 other publicly available aerial survey datasets collected across 4 different countries, improving mAP by $42\%$ . Source code and trained models are available at Github1. + +# 1. Introduction + +This paper introduces WildlifeMapper (WM) - an automated and scalable method for counting wildlife in aerial imagery. Aerial wildlife surveys are recognized as a cornerstone of modern conservation biology. By facilitating large-scale biological monitoring in remote landscapes, this technique has underpinned the ability to track changes in the abundance and distribution of wildlife across open landscapes for decades. However, traditional survey approaches + +often rely on manual observers to identify, count, and validate species of interest. This labor-intensive process can be time-consuming and error-prone, with potential to limit the utility of final results [27, 35, 38]. + +Automated approaches offer a promising alternative for efficient and accurate detection of wildlife in aerial survey images. Recent work, for example, illustrates how artificial intelligence has been used to count a variety of species from the air, including antelope in grasslands [44], whales in the ocean [11], and seals on the beach [42]. When combined with advancements in low-cost, high-resolution imaging platforms (e.g., UAVs), these case studies underscore the potential for such data to significantly reduce the effort and cost of traditional wildlife census methods. However, the majority of these techniques struggle to generalize to new species or landscapes due to the dominance of homogeneous backgrounds and poorly captured local structures [7, 12, 18, 39]. + +WildlifeMapper overcomes these limitations by adapting a novel application of the segment anything transformer model [21]. This model combines high frequency component correlations and spatial correlations in the image data to generate a map of potential locations of objects of interest (i.e., wildlife, livestock). In addition, we address the challenge of identifying multiple species from a relatively small footprint in these images. + +To demonstrate the WM analysis workflow, we provide a case study example across the Masai Mara Ecosystem in southwestern Kenya. Renowned for its rich biological diversity, the abundance of large mammals (such as buffalo (Syncerus caffer), giraffe (Giraffa tippelskirchi), and wildebeest (Connochaetes taurinus)) have declined precipitously over the past few decades [34]. Importantly, our analy- + +![](images/de19e1e68141ca4c8edb39965055f98a17a22c0096e8d3b8868772137af7bab8.jpg) +Figure 1. Summary of Mara-Wildlife dataset. (a) Satellite view indicating the four flight trajectories, each represented in a different color. (b, c, d, e, f, g) Annotations of (b) zebra, (c) hartebeest, (d) cattle, (e) shoats (sheep and goats), and (f) zebra. Best viewed in color. + +![](images/6832a4e5243a8e9398d549c8bd8efb56ec9e6d4c661029589a4e627d132a094b.jpg) + +![](images/601a27c2752bd01b8bd0b0412b7d042cf65fa2560ccac096200ec18b93c0b85a.jpg) + +![](images/b6607979af0fc92740acb765118178ee2e51d58dd5df1a94ebfed2924da6129d.jpg) + +sis incorporated 11, 151 images of size $8400 \times 5500$ collected from a digital camera affixed to the bellyport of a Partenavia P68 airplane during Systematic Reconnaissance Flight (SRF) surveys. Part of these images were annotated by trained observers with 28, 146 annotations of 21 species of large mammals ( $\geq 15kg$ ), providing an unprecedented opportunity to develop species detection models across a complex, heterogeneous environment. The dataset was systematically verified by trained observers as described in Section 3. Our contributions can be summarized as follows: + +1. A novel, single-stage end-to-end approach for animal detection. The modules, a High Frequency Feature Generator, a Feature Refiner, and a Query Refiner, work together to improve upon the traditional methods of object detection in aerial imagery and enable generalizability across different habitats. The high frequency features reduce dependence on dominant backgrounds/landscapes. +2. An input patch embedding layer that is specifically designed to capture contextual information to help in identifying individual animal species. +3. The release of a new benchmark dataset via the data owner (Kenya's Wildlife Research and Training Institute - WRTI) once all approvals are in place. An international user community is already engaged in further enhancing these data and using the WM through the BisQue [25] platform. + +Community Adoption The practical use of WildlifeMapper (WM) extends beyond theoretical and computational success and is operationalized through BisQue [25]. WM is made available to users through a series of training modules that demonstrate how to $(i)$ upload digital imagery, $(ii)$ create annotations, $(iii)$ apply and/or improve existing models, $(iv)$ evaluate model fit, $(v)$ improve annotations and refit models, and $(vi)$ generate summary statistics. In the fu + +ture, we envision WM to be of great value to ecologists, wildlife managers, and government officials, providing accurate information about the state of wildlife populations in near real-time, facilitating decision-making processes, and improving the conservation of ecosystems globally. + +# 2. Related Works + +Manual Methods: Aerial surveys using Front- and Rear-Seat Observers (FSO and RSO, respectively) are commonly used to inventory wildlife populations across open landscapes [36]. However, several important biases can impact these counts, including the experience level and fatigue of the human observers [28]. + +Deep Learning Methods: To address these issues, researchers have begun incorporating digital cameras on piloted aircraft and UAVs [18,30,46]. This minimizes the influence of observer bias while increasing transparency and reproducibility of results. For example, [26] replaced RSOs with an oblique camera mount system minimized the influence of observer bias while producing comparable estimates of large mammals under partial canopy cover. Similarly, [44] used a nadir mounted camera to improve the accuracy and efficiency of manual counts of large antelope in an open grassland ecosystem. + +However, detecting animals in the wild from aerial imagery poses many challenges. For example, most publicly available datasets for aerial object detection are focused on identifying relatively distinct features such as buildings, roads, vehicles, and other man-made structures [8, 14, 53]. Animals, tend to blend in with their surroundings [13], can be occluded by trees, exhibit considerable variation in color and pattern, or have behavioral adaptations that make them difficult to detect [7, 20, 39, 48]. + +[20] proposed a solution for this problem involving a two-branch CNN model based on AlexNet to perform animal recognition and localization. [7] evaluated three state-of-the-art object detection algorithms, including Faster-RCNN, Libra-RCNN, and RetinaNet on six African wild + +![](images/5341e0e4cd475f09e0ed21ef4138abbd9227727e77b1fc66de0b3b25d610b7d8.jpg) +Figure 2. Distribution of $(\geq 15kg)$ mammals identified in digital imagery collected across the Masai Mara Ecosystem, Kenya. + +
Dataset# of annot. images# of annot. tiles# of species# of annotationImage sizeGSD (cm)Location
Virunga73930069656646000x40002.4DRC
Garamba1586429616116000x40002.0DRC
AED2067693871155815500x36002.4-13.0Bostwana, Namibia, South Africa
Mara-Wildlife10127796621281468256x55041.45Masai Mara National Reserve
+ +Table 1. Comparison of Mara-Wildlife dataset with other publicly available dataset. Mara-Wildlife dataset has $\times 3$ more unique species than the total of all other datasets. Each image is significantly larger and higher ground resolution making 77k unique images of size ${1024} \times {1024}$ with 21 different animal species. GSD: ground sampling distance; DRC: Democratic Republic of Congo. + +mammals. All three algorithms, however, showed poor performance in animal detection when animals were grouped closely together in herds. [12] adopted a segmentation approach, employing a UNet model to detect livestock from drone imagery. [48] uses a comparable model to analyze high-resolution satellite imagery, producing segmentation masks of wildebeest-sized animals, which are subsequently utilized for detection and counting. + +Transformers: WM adopts a transformer architecture based on past success in modeling different types of aerial imagery. Examples include incorporating multispectral imagery for change detection [4], landcover classification [16], greenhouse gas (GHG) emission detection [23,24] and RGB aerial imagery for object detection [29, 49, 52]. + +The most effective applications of transformer-based models have been tailored for standard object detection tasks [3, 5, 32, 41, 54]. These works leverage the self-attention to model dependencies among the patches in an end-to-end fashion, unlike CNN-based models [1, 40]. However, when directly applied to aerial imagery, these models cannot effectively exploit the local structures as they divide the image into a sequence of patches. This limits the detection of small-scale objects in a homogeneous and dominant background. + +Dataset: The existing publicly available animal aerial imagery datasets are listed in Table 1. The Virunga dataset [7] was collected in Virunga National Park, Democratic Republic of Congo (DRC). This dataset contains 897 annotated images of 6 animal species. The Garamba dataset was collected in Garamba National Park, DRC and contains 7034 images. Only 158 images have been annotated, containing 7 animal species. The aerial elephant dataset (AED) [33] was + +collected across a mosaic of woodland, open shrubland, and grassland habitats in Botswana, Namibia, and South Africa. Only a single species (i.e., elephant) was targeted during SRF surveys. See Table 1 for details. + +# 3. Mara-Wildlife Dataset + +The Mara-Wildlife dataset is a distinctive dataset that captures the essence of the Masai Mara ecosystem through a compilation of 77966 images of size $1024 \times 1024$ . This habitat is heterogeneous, including woodland, shrubland, and grassland vegetation with 21 unique animal species. + +# 3.1. Image Collection + +Flightline Details: Data collection was in collaboration with the Smithsonian National Zoo and Conservation Biology Institute (SNZCBI), Kenya's Wildlife Research and Training Institute (WRTI), the Kenya Wildlife Trust (KWT), and the Directorate of Resource Surveys and Remote Sensing (DRSRS). In March 2022, we fitted a Partenavia P68 with a Nikon D850 digital camera and collected high resolution $(8256\times 5504)$ digital images. During data acquisition, the aircraft adhered to a predetermined flight trajectory, depicted in Figure 1, at 400 ft above ground level (agl). This trajectory was optimized to encompass open grassland areas across the Masai Mara ecosystem, including the Masai Mara National Reserve, 22 adjacent private conservancies, and unprotected peripheral areas. + +The aerial survey was conducted during a wet season period when the Serengeti migratory population of wildebeest have moved southward to locate more suitable forage in Tanzania. Thus, the survey primarily captured resident species, including wildebeest, zebra, topi, hartebeest, gi + +![](images/d4cd2368475fed771be868c7297357bd3efcd94f068ce5a916b5aa138c4bee42.jpg) +Figure 3. Overview of WildlifeMapper (WM) architecture. Given an input image of size $1024 \times 1024 \times 3$ , the High-Frequency Feature Generator (HFG) module generates information about potential location of object of interest. The Feature Refiner (FR) takes these potential locations along with contextual features from Patch Embed layer and sent output to Image Encoder. In parallel, the Query Refiner (QR) incorporates the output of FR to refine learnable queries. Finally these queries are decoded using encoded features from Image Encoder and predict bounding box and class. + +raffe, and other large ( $\geq 15\mathrm{kg}$ ) antelope. Data collection was conducted in the early mornings (prior to 10:00 EAT) and late afternoons (after 15:00 EAT) when lighting conditions were optimal and animals are most active. + +The geographical positioning of each image was acquired through a GPS system that recorded the plane's altitude, speed, and geographical coordinates. These data were then synchronized with the image using the image capture timestamps, enabling us to determine the geographic location of the centroid of each image. The camera was placed in the bellyport of the airplane, capturing a nadir view of the landscape every two seconds along the flight path. + +# 3.2. Image Annotation + +Initial bounding box annotations (21796) generated using AIDE platform [19] were exported in CSV format. These were then imported into BisQue [25] for further validation and correction. Finally, these annotations underwent validation by a single trained observer specializing in ecology, resulting in 28146 annotations in total. + +# 3.3. Dataset Statistics + +The Mara-Wildlife dataset showcases a detailed assemblage of wildlife, inclusive of 21 distinct species classes. The dataset is composed of approximately 77,966 tiled images, derived from 1,012 original rasters (Table. 1) The meticulous process of annotation has culminated in labeling 28,146 animals. Species identified include domestic cattle (Bos taurus), white-bearded wildebeest (Connochaetes taurinus), topi (Damaliscus lunatus), shoats (domesticated sheep and goats), kongoni (Alcelaphus buselaphus), waterbuck (Kobus ellipsiprymnus), impala (Aepyceros melampus), Grant's gazelle (Nanger granti), Thomson's gazelle (Eudorcas thomsonii), Cape buffalo (Syncerus caffer), zebra (Equus quagga), ostrich (Struthio camelus), Masai giraffe (Giraffa tippelskirchi), warthog (Phacochoerus africanus), eland (Taurotragus oryx), donkey (Equus africanus), hyena (Crocuta crocuta), hippopotomus + +(Hippopotamus amphibius), lion (Panthera leo), and elephant (Loxodonta africana). + +# 4. WildlifeMapper Architecture + +# 4.1. Technical Overview + +WildlifeMapper's architecture is inspired by the success of the Segment Anything Model (SAM) [21], created to segment small/large (all sizes) of objects. Referring to Fig. 3, WM contains the following main components: (i) A patch embedding layer designed to capture long-range context, (ii) a High-Frequency Feature Generator (HFG), (iii) a Feature Refiner (FR) followed by ViT based image encoder [21,23,45], and (iv) a Query Refiner (QR) module followed by a box decoder module. The input image is first processed through two separate branches, the patch embedding layer which captures long-range context [9] and HFG which suppresses all the low-frequency components in the image and generates a feature embedding. The HFG (Sec. 4.3) exploits prior knowledge that aerial images from areas such as forests, grasslands, and shrublands have a homogeneous and dominant background representing the dominant low frequency image content. Fig. 4 shows that on suppressing the lower frequencies, the object of interest is easy to locate. + +The FR (Sec. 4.4) takes the embeddings from each of the two branches and generates a high quality embedding that contains information about potential locations of animals and captures the local context. The QR modules refines a set of learnable queries using the location information from FR module. These refined queries are passed to the box decoder. The box decoder takes the refined queries and encoded features image encoder to generate the final detection box and class of the object. + +# 4.2. Patch Embed + +The patch embedding layer utilizes a larger kernel convolution with an increasing dilation rate. This design rapidly expands the receptive field, allowing explicit extrac + +tion of features rich in contextual information. This approach is particularly beneficial for aerial imagery, where the small sized object makes classification based on appearance alone challenging. Contextual information thus becomes crucial for the accurate recognition of these objects. + +# 4.3. High-frequency Feature Generator (HFG) + +Along with patch embedding, the input image is processed in parallel by the HFG module to generate features with information about the location of the animal or cluster as shown in Fig. 4. The HFG module is inspired from the limitation of ViT models [45]. ViT models face challenges in efficiently utilizing local structures. They segment an image into patches and apply self-attention to model relationships, but this approach often falls short in capturing detailed local features [37, 51]. + +Local features in images are closely linked to high-frequency components [2, 43]. We hypothesize that suppressing low-frequency components can mitigate the influence of a dominant homogeneous background. To test this, we performed a discrete Fourier Transform (DFT) on the images, filtering out the low-frequency components before reconstructing the images, as shown in Fig. 4. + +For a given input image $I \in \mathbb{R}^{H \times W \times C}$ , where $C$ is channel dimension, we compute Discrete Fourier Transform (DFT) of $I$ . In next step we suppress the low frequency components with a controlling parameter and construct the image $I$ with inverse DFT (IDFT). + +$$ +I ^ {\prime} = I D F T [ h p f (D F T (I) ] \tag {1} +$$ + +where $hpf$ is a high pass filter. Then we reduce the dimension of the reconstructed image $I'$ via an embedding layer to generate embedding $hfc_{emb}$ and pass them to the FR module. See supplementary materials for more details. + +# 4.4. Feature Refiner (FR) + +Next, the features from the patch embedding layer $(f_{img})$ and HFG $(hfc_{emb})$ are fed to the FR module. The $f_{img}$ are refined with the $hfc_{emb}$ via cross-attention mechanism. The FR is a simple module with cross-attention and linear layers [45]. The output contains information about the potential location of the object and the long-range context. + +Following the standard architecture of SAM [21], we pass FR output to our ViT based image encoder supplemented with learnable positional embeddings $p$ . The encoded feature map is $f_{emb}$ : + +$$ +f _ {e m b} = \operatorname {V i T} \left[ \mathbf {F R} \left(f _ {i m g}, h f c _ {e m b}\right), p \right] \tag {2} +$$ + +Query Refiner (QR): The QR follows a transformer decoder like architecture and takes as input a set of 100 learnable queries $Q \in \mathbb{R}^{100 \times d}$ and output of FR module. Here $d = 256$ is same as channel dimension of $f_{emb}$ from image encoder. The FR output refines the $Q$ via a cross-attention + +![](images/e90a672b29dee430d653ce2754938f7657c70c98fd61dd4ef0fa4e9a0ea10bb1.jpg) +Figure 4. The sample output visualization from the High-Frequency Feature Generator (HFG) module. The illustration shows the effectiveness of the module in suppressing the homogeneous and dominant background, while highlighting objects of interest (i.e., animals). The top image shows bomas, natural structures constructed to contain livestock, and paths that have been suppressed. Animals, however, are clearly identified, especially inside the boma. The bottom image shows a water body (a dam created for livestock) that has been suppressed by the module. Animals can again be highlighted throughout the image. + +mechanism. The refined queries narrows the search space for box decoder module to accurately locate and identify object of interest. [15,23]. + +# 4.5. Box Decoder + +Next, the refined queries are sent to the box decoder module. We concatenate these queries with the $f_{emb}$ and pass box decoder's self attention layer as $qf_{emb}$ . This is inspired from the idea of class_token used by Vaswani et.al. [45] to make decoding process memory efficient. The transformer network takes 4 input variables, those are position embeddings of queries and image embeddings, $qf_{emb}$ , and $f_{emb}$ . Our transformer model uses two-way attention inspired from [3, 6, 21] and our box decoder uses self-attention and cross-attention in two directions (queries-to-image embedding and vice-versa) to update all embeddings. We keep the box decoder very light weight (two blocks). The top 100 (equal to number of queries) indexes from the output of the final block is passed to two separate MLP blocks to regress the output bounding box prediction and class of the predicted box. + +# 4.6. Training and Inference + +We've trained WM using a single-stage, end-to-end approach to determine bounding boxes and classify them. The loss strategy we applied for WM is akin to what's used in DETR [3]. Initially, we perform bipartite matching to align our model's predictions with the actual bounding box data. Then, we proceed to compute the loss for these matched pairs. To achieve the best possible match between our predictions and the real data, we use the Hungarian algorithm [22]. Once matched, each prediction is paired with its respective ground truth. We then measure the $l_{1}$ (L1 distance) and $GIoU$ loss for the bounding box and the cross entropy loss for the classification [3]. + +Inference: The inference pipeline is straightforward and similar to training code. During inference, we first filter the detections at $50\%$ threshold and then use non-maximal suppression to remove any overlapping boxes. + +# 5. Experiments + +Train-val-test split: There is no data leak in train-val-test split. Fig. 1 shows the flight paths with location of each image represented by a circle. Each color coded flight path represents the train-val-test image set. No images taken while the airplane was cutting across the transects plus turnaround points as can be seen there are no circles. We created a spatial disjoint of $20km$ distance between the transects as shown in Fig. 1 and achieved consistent performance (0.56mAP). + +Evaluation Metrics: We report our performance on multiple metrics. Following the protocols of standard object detection, we report the performance in mean average-precision (mAP) for detection and mean intersection-overunion (mIOU) for localization of animals. We also report a commonly used metric by the ecologists on the team, the class-wise mean absolute error (MAE) indicating the counting accuracy of each species: + +$$ +\mathrm {M A E} = \frac {1}{I} \sum_ {i = 1} ^ {I} \sum_ {c = 1} ^ {C} | \hat {n} _ {i, c} - n _ {i, c} |, \tag {3} +$$ + +Where $I$ is the number of images, $C$ is the number of classes, $\hat{n}_{i,c}$ and $n_{i,c}$ are the predicted and ground truth counts for class $c$ in image $i$ . + +Implementation Details: Each image taken from the drone is $8256 \times 5506 \times 3$ . We create tiles for each image in the spatial domain, with the size of $1024 \times 1024 \times 3$ with $25\%$ of overlap. The Patch Embed layer uses a single CNN layer with a large kernel of size $16 \times 16$ with stride 16. In the parallel branch, the High-Frequency Feature Generator, we use DFT to compute the Fourier transform, the mask is + +a binary disk with the radius set to 128. The HFC Embed layer uses 3 CNN layers with ReLU activation with a kernel of size $3 \times 3$ and a global average pool at the end. The Feature Refiner (FR) module consists of one cross attention layer with 1 linear layer. The image encoder is a pre-trained ViT model [21] with 24 transformer layers and 16 heads. The Query Refiner (QR) module takes in 100 queries each of channel dimension 256, those are cross attended with $hfc_{emb}$ output. The box decoder contains 3 layers of two-way attention with 8 heads. We train WM with AdamW optimizer [31] setting the learning rate to $10^{-4}$ for the FR, QR and box decoder with a weight decay to $10^{-4}$ . We set the learning rate for the Patch Embed and HFC Embed layer to $10^{-5}$ . We load the image encoder with pre-trained weights from segment anything [21] and keep it frozen. + +Data Augmentation: In the train and test datasets, we incorporated an equal number of images without any objects to assess the model's robustness against empty background images. We applied multiple data augmentation techniques, including HSV (hue, saturation, and value) $(10\%)$ , rotation $(5\%)$ , translation $(10\%)$ , affine transformation $(20\%)$ , scale $(10\%)$ , shear $(5\%)$ and mosaic $(70\%)$ augmentation [1]. Mosaic augmentation is proven to be the most effective, with an improvement of $0.07\mathrm{mAP}$ + +Hard Negative Mining: After training for 100 epochs, we take all the False Positives (FP) predictions having $\mathrm{IOU} \leq 0.10$ with ground truth box and mark them as background class. Then fine-tuning for 20 epochs improved the performance of FP reduction for detecting rocks, dead tress or other artifacts on ground as animal. + +# 6. Results + +# 6.1. Performance Comparison + +We trained WM separately on Mara-Wildlife dataset and Virunga-Garamba-AED dataset for comparison with existing works, see Table 2 for a summary of the results on all of the tested datasets. The $mAP$ values are compared for IoU of 0.50-0.95 and 0.50. We also provide the average counting error in animal counting per image. We trained all the baseline models with the default set of conditions on Mara-Wildlife dataset. We merged the Virunga, Garamba, and AED datasets and created the train-val-test split according to [7]. The combination of these datasets contains 6 unique animal species and diversity of landscapes such as woodland, savannahs, open shrubland, and grasslands across multiple countries – Democratic Republic of Congo (DRC), Botswana, Namibia, and South Africa. WM outperforms all methods by a significant margin as shown in Table 2. We note that in [7] the authors did not make the code base or trained model public, hence we could not verify the results. We implemented these methods from + +
Methods#epochsmAPmAP50Counting Error
Mara Wildlife Dataset
1Faster-Rcnn [40]1000.240.582.59
2DETR [3]2000.220.572.75
3Co-DETR-R50 [55]1000.270.662.72
4Co-DETR-swingL [55]1000.280.652.60
5Yolo v5 [10]1000.300.672.12
6Yolo v8 [17]1000.270.613.97
7LSKNet [29]1000.290.66-
6DroneDetect [50]1000.180.48-
8WildlifeMapper1200.560.791.9
Virunga-Garamba-AED Datasets
6Faster-Rcnn1200.340.650.27
7DETR2000.300.620.45
8Yolo v51000.480.780.12
9Yolo v81000.480.770.42
10WildlifeMapper800.680.850.11
+ +Table 2. Comparison with baseline models. The top section shows performance on species detection on Mara-wildlife dataset and low section shows performance on the mixed dataset from Virunga-Garamba-AED dataset. The overall detection accuracy is generally higher in Virunga-Garamba-AED dataset because there are only 6 species and the terrain is quite similar in all images. + +
HFGFRQRmAP
WMXX*0.46
✓✓*0.54
✓X✓0.49
✓✓✓0.56
+ +Table 3. HFG module effectiveness in refining the image features and queries. “X” represents not used, “√” represents used and “*” represents that random queries are used but there was not refining with HFG features. + +the original public repositories and trained according to the training strategy detailed in [7]. We attribute the model poor learning performance due to salient features of the homogeneous background being learned more than the object of interest. The detection of the object of interest is then dependent on the landscape properties instead of object properties. Hence when used on a slight variations of landscapes for the same object, the models struggle to detect. This limitation is specifically addressed in the WM, where the HFG modules suppresses the background and highlights the object of interest. + +Qualitative results: Fig. 5 shows the quality of detection by WM in different scenarios. Those include, detection when animal is partially visible under a tree, or a big clustering. WM makes correct predictions in varying scenarios. + +# 6.2. Ablation Studies + +We performed all ablation experiments on Mara-Wildlife (MW) dataset and validate the design choices. + +High-frequency Feature Generator Module: In Table 3, we show the effectiveness of the HFG module. We + +experimented with HFG's output in 3 ways: first, we passed HFG's output to Feature Refiner (FG) module. It leads to significant improvement in detection by $0.09\mathrm{mAP}$ over the baseline. This demonstrates that providing potential location candidates features to image encoder module produce better embeddings. Second, we pass the HFG's output to Query Refiner (QR) module only. This leads to an improvement of $0.03\mathrm{mAP}$ over baseline. This shows the effectiveness of guiding queries with location candidates features. In the third case, we passed the HFG's output to both FR and QR modules and achieved an improvement of $0.11\mathrm{mAP}$ over the baseline. We hypothesize that this reduces the dominance of features from homogeneous and dominant background in aerial imagery. The also observed this while testing WM across flightlines different types of terrain such as green grasslands, dry grasslands, and forest areas. + +Feature Refiner Module: We tested $hfc_{emb}$ and $f_{img}$ merging by 3 ways: addition, concatenation and cross-attention. Cross-attention is most effective, because with addition and concatenation, the $hfc_{emb}$ get lost, while cross-attention generates better embeddings giving attention to potential location candidates. + +Kernel Size: We observed that a larger kernel size of $31 \times 31$ results in reduced misclassification. For example, a topi or warthog cannot be found inside a boma because only domestic species are kept in bomas. So context helps in making the right class detection. We observed an improvement of $0.02\mathrm{mAP}$ . Experiments were done in 3 kernel sizes. 1. $7 \times 7$ : $0.55\mathrm{mAP}$ , 2. $16 \times 16$ : $0.558\mathrm{mAP}$ ; 3. $32 \times 32$ : $0.57\mathrm{mAP}$ . + +Query Refiner Module: We experimented with only providing random queries and guiding the queries with HFG module output. We merged them with direct addition or concatenation and cross-attention. With cross-attention, we observed an improvement in performance of 0.03mAP. + +Geographic generalization: To test the geographic generalizability of WM across different terrains, we trained WM only on images from Kenya; and tested on images from Democratic Republic of Congo, Botswana and Namibia. The test was done on 4 common species which were present in both the ecosystems. WM achieved a detection performance of $0.48\mathrm{mAP}$ . This shows the adaptability of WM across varying landscapes. + +Domain generalization: We train-test WM on a different domain, a bird species tern [47], commonly found on/near water bodies. Live in huge clusters. WM achieved the accuracy of $0.71\mathrm{mAP}$ . Showing adaptability of WM across different domains. + +![](images/7da36f287a9680a3d38a1103eecc5ccdef580a499bb64df42d28969e962847cd.jpg) + +![](images/cdcaf9e16c6949d3362b62f597dea6482cbc320233d295c1cbc888049b210f7c.jpg) + +![](images/c02f6e19018cec214b4ef2b317f0a06d7139d8cfb673145f181fe78734974b5f.jpg) + +![](images/709b87d56af4929a8775f447c530f3f8c078e8dcb0a591f3f3e032e5ac283b4f.jpg) +Figure 5. Qualitative results. The top row highlights examples of crowded and partially occluded scenes. Row-1, Column-1 & Row-3, Column-3 shows examples where animals are obstructed by shadows. The zoomed-in box in Row-3, Column-3 shows a zebra partially occluded. The bounding box color is coded according to class names: shoats- "hot pink", cattle- "deep sky blue", zebra- "light yellow". + +![](images/ead6d7a1e79e473d59797ae734c2b69f4ac5c21e165bff6a24b13a98c9900668.jpg) + +![](images/5436f5675f603e8e60d9191169ab6556dc0ddc4c11a9fb98eb5829baf103bcf5.jpg) + +![](images/418c4bb1f01517b5e27717d0e376fd49549229379a0488650fdb0ee07dd02377.jpg) +Figure 6. Failure cases. Left shows an example where animals are occluded by shadows and are difficult to detect. Right shows an example of rock detected as an impala, emphasizing the difficulty in differentiating objects in the image from animals of interest. + +![](images/05616dc7838c4257c59d1736e02807be40161ffe63b1433f9026916dfb0b63d7.jpg) + +Failure Cases: The detections from WM are inaccurate when animals are clustered in shadows, such as when animals are located inside bomas and the sun angle makes a strong shadow on the enclosure. These are some of the difficult cases shown in Fig. 6. Other cases of false positives are the small rocks or trees that sometimes resemble animals. Some examples are shown in the supplementary materials. + +# 7. Conclusion + +This paper presents WildlifeMapper (WM) - a transformer based approach for the detection of animals of varying densities and sizes across natural backgrounds. The WM utilizes a high frequency features generator, feature refiner, and query refiner to accurately locate and classify 8 animal species. WM stands to significantly improve the efficiency and accuracy of wildlife monitoring and conservation efforts. Future work will extend this model and dataset to a larger number of species and habitats. + +# 8. Acknowledgement + +This research is partially supported by the following grants: NSF award SI2-SSI #1664172 (BSM), the Ohrstrom Family Foundation, and a Smithsonian National Zoo and Conservation Biology Institute (SNZCBI) fellowship (S.Kumar). We also recognize the contributions of the annotation team, including S. Harman, A. Gray, C. Pate, C. Obath, S. Nyarangi, C. Kagume, and W. Sairowua. + +# References + +[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3, 6 +[2] Fergus W Campbell and John G Robson. Application of fourier analysis to the visibility of gratings. The Journal of physiology, 197(3):551, 1968. 5 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5, 6, 7 +[4] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3 +[5] Qiang Chen, Jian Wang, Chuchu Han, Shan Zhang, Zexian Li, Xiaokang Chen, Jiahui Chen, Xiaodi Wang, Shuming Han, Gang Zhang, et al. Group detr v2: Strong object detector with encoder-decoder pretraining. arXiv preprint arXiv:2211.03594, 2022. 3 +[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 5 +[7] Alexandre Delplanque, Samuel Foucher, Philippe Lejeune, Julie Linchant, and Jérôme Théau. Multispecies detection and identification of african mammals in aerial imagery using convolutional neural networks. Remote Sensing in Ecology and Conservation, 8(2):166-179, 2022. 1, 2, 3, 6, 7 +[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44(11):7778-7796, 2021. 2 +[9] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 4 +[10] Glenn Jocher et. al. ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements, Oct. 2020. 7 +[11] Emilio Guirado, Siham Tabik, Marga L Rivas, Domingo Alcaraz-Segura, and Francisco Herrera. Whale counting in satellite and aerial images with deep learning. Scientific reports, 9(1):14259, 2019. 1 +[12] Liang Han, Pin Tao, and Ralph R Martin. Livestock detection in aerial images using a fully convolutional network. Computational Visual Media, 5:221-228, 2019. 1, 3 +[13] Tracey Hollings, Mark Burgman, Mary van Andel, Marius Gilbert, Timothy Robinson, and Andrew Robinson. How do you find the green sheep? a critical review of the use of remotely sensed imagery to detect and count animals. Methods in Ecology and Evolution, 9(4):881-892, 2018. 2 +[14] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H. Hsu. Drone-based object counting by spatially regularized re + +gional proposal networks. In The IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 2 +[15] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. In Pattern Recognition and Tracking XXXIV, volume 12527, pages 192-205. SPIE, 2023. 5 +[16] Johannes Jakubik et. al. Prithvi-100M, Aug. 2023. 3 +[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics* YOLO, Jan. 2023. 7 +[18] Benjamin Kellenberger, Diego Marcos, and Devis Tuia. Detecting mammals in uav images: Best practices to address a substantially imbalanced dataset with deep learning. Remote Sensing of Environment, 216:139-153, Oct 2018. 1, 2 +[19] Benjamin Kellenberger, Devis Tuia, and Dan Morris. Aide: Accelerating image-based ecological surveys with interactive machine learning. Methods in Ecology and Evolution, 11(12):1716-1727, 2020. 4 +[20] Benjamin Kellenberger, Michele Volpi, and Devis Tuia. Fast animal detection in uav images using convolutional neural networks. In 2017 IEEE international geoscience and remote sensing symposium (IGARSS), pages 866-869. IEEE, 2017. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 4, 5, 6 +[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6 +[23] Satish Kumar, Ivan Arevalo, ASM Iftekhar, and BS Manjunath. Methanemapper: Spectral absorption aware hyperspectral transformer for methane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17609-17618, 2023. 3, 4, 5 +[24] Satish Kumar, William Kingwill, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan D Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. In NeurIPS 2022 Workshop on Tackling Climate Change with Machine Learning, 2022. 3 +[25] Kristian Kvilekval, Dmitry Fedorov, Boguslaw Obara, Ambuj Singh, and BS Manjunath. Bisque: a platform for bioimage analysis and management, Jan. 2023. 2, 4 +[26] Richard Lamprey, David Ochanda, Rob Brett, Charles Tumwesigye, and Iain Douglas-Hamilton. Cameras replace human observers in multi-species aerial counts in murchison falls, uganda. Remote Sensing in Ecology and Conservation, 6(4):529-545, 2020. 2 +[27] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma, and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 1 +[28] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma, + +and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: Taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 2 +[29] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. arXiv preprint arXiv:2303.09030, 2023. 3, 7 +[30] Julie Linchant, Jonathan Lisein, Jean Semeki, Philippe Lejeune, and Cedric Vermeulen. Are unmanned aircraft systems (uas s) the future of wildlife monitoring? a review of accomplishments and challenges. Mammal Review, 45(4):239-252, 2015. 2 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 3 +[33] Johannes Naude and Deon Joubert. The aerial elephant dataset: A new public benchmark for aerial object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 48-55, 2019. 3 +[34] Joseph O Ogutu, Hans-Peter Piepho, Mohamed Y Said, Gordon O Ojwang, Lucy W Njino, Shem C Kifugo, and Patrick W Wargute. Extreme wildlife declines and concurrent increase in livestock numbers in kenya: What are the causes? PloS one, 11(9):e0163249, 2016. 1 +[35] Wilber K Ottichilo, Jesse Grunblatt, Mohammed Y Said, and Patrick W Wargute. Wildlife and livestock population trends in the kenya rangeland. *Wildlife conservation by sustainable use*, pages 203–218, 2000. 1 +[36] Wilber K. Ottichilo, Jesse Grunblatt, Mohammed Y. Said, and Patrick W. Wargute. Wildlife and Livestock Population Trends in the Kenya Rangeland, pages 203-218. Springer Netherlands, Dordrecht, 2000. 2 +[37] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 5 +[38] Michael John Stephen Peel and Marc Stalmans. The systematic reconnaissance flight (srf) as a tool in assessing the ecological impact of a rural development programme in an extensive area of the lowveld of south africa. African Journal of Ecology, 37(4):449-456, 1999. 1 +[39] Tinao Petso, Rodrigo S Jamisola Jr, Dimane Mpoeleng, Emily Bennitt, and Wazha Mmereki. Automatic animal identification from drone camera based on point pattern analysis of herd behaviour. Ecological Informatics, 66:101485, 2021. 1, 2 +[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3, 7 +[41] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection + +with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021.3 +[42] AC Seymour, J Dale, M Hammill, PN Halpin, and DW Johnston. Automated detection and enumeration of marine wildlife using unmanned aircraft systems (uas) and thermal imagery. Scientific reports, 7(1):45127, 2017. 1 +[43] Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal on mathematical analysis, 29(2):511-546, 1998. 5 +[44] Colin J Torney, David J Lloyd-Jones, Mark Chevallier, David C Moyer, Honori T Maliti, Machoke Mwita, Edward M Kohi, and Grant C Hopcraft. A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images. Methods in Ecology and Evolution, 10(6):779-787, 2019. 1, 2 +[45] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5 +[46] Cédric Vermeulen, Philippe Lejeune, Jonathan Lisein, Prosper Sawadogo, and Philippe Bouché. Unmanned aerial survey of elephants. PloS one, 8(2):e54700, 2013. 2 +[47] Ben G Weinstein, Lindsey Garner, Vienna R Saccomanno, Ashley Steinkraus, Andrew Ortega, Kristen Brush, Glenda Yenni, Ann E McKellar, Rowan Converse, Christopher D Lippitt, et al. A general deep learning model for bird detection in high-resolution airborne imagery. Ecological Applications, 32(8):e2694, 2022. 7 +[48] Zijing Wu, Ce Zhang, Xiaowei Gu, Isla Duporge, Lacey F. Hughey, Jared A Stabach, Andrew K Skidmore, J Grant C Hopcraft, Stephen J Lee, Peter M Atkinson, et al. Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape. Nature communications, 14(1):3072, 2023. 2, 3 +[49] Xiangkai Xu, Zhejun Feng, Changqing Cao, Mengyuan Li, Jin Wu, Zengyan Wu, Yajie Shang, and Shubing Ye. An improved swim transformer-based model for remote sensing object detection and instance segmentation. Remote Sensing, 13(23):4779, 2021. 3 +[50] Yinhui Yu, Xu Sun, and Qing Cheng. Expert teacher based on foundation image segmentation model for object detection in aerial images. Scientific Reports, 13(1):21964, 2023. 7 +[51] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 5 +[52] Jiahe Zhu, Xu Chen, Huan Zhang, Zelong Tan, Shengjin Wang, and Hongbing Ma. Transformer based remote sensing object detection with enhanced multispectral feature extraction. IEEE Geoscience and Remote Sensing Letters, 2023. 3 +[53] Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Ling, and Qinghua Hu. Vision meets drones: A challenge. arXiv preprint arXiv:1804.07437, 2018. 2 + +[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3 +[55] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 7 \ No newline at end of file diff --git a/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/images.zip b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0afef6ced60fabb02f76f293219f7a35b10a96ee --- /dev/null +++ b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7f1a5434e3fcec8f15ef9748783a9e0763ad8e06b9bd4fdae747fda125cc1c6 +size 743785 diff --git a/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/layout.json b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..67df3f768bbeb44e1db85b4b3dbbaf86e308d14b --- /dev/null +++ b/2024/WildlifeMapper_ Aerial Image Analysis for Multi-Species Detection and Identification/layout.json @@ -0,0 +1,8766 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 518, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 518, + 138 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 518, + 138 + ], + "type": "text", + "content": "WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "spans": [ + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "Satish Kumar" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Bowen Zhang" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Chandrakanth Gudavalli" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Connor Levenson" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Lacey Hughey" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Jared A. Stabach" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Irene Amoke" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Gordon Ojwang" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Joseph Mukeka" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Stephen Mwiu" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Joseph Ogutu" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " Howard Frederick" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": " B.S. Manjunath" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "University of California Santa Barbara " + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "Smithsonian National Zoo and Conservation Biology Institute " + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "Kenya Wildlife Trust, " + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "University of Groningen, " + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "Wildlife Research and Training Institute " + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "University of Hohenheim, " + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 57, + 160, + 536, + 244 + ], + "type": "text", + "content": "Tanzania Wildlife Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 247, + 470, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 247, + 470, + 287 + ], + "spans": [ + { + "bbox": [ + 118, + 247, + 470, + 287 + ], + "type": "text", + "content": "(satishkumar,bowen68,chandrakanth,clevenson,manj)@ucsb.edu, (stabachj,hugheyl)@si.edu,iirena.amoke@kenyawildlifetrust.org, (gordonojwang,simbamangu)@gmail.com,(jmukeka,smwiu)@wrti.go.ke" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 315, + 192, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 315, + 192, + 327 + ], + "spans": [ + { + "bbox": [ + 143, + 315, + 192, + 327 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 333, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 289, + 559 + ], + "type": "text", + "content": "We introduce WildlifeMapper (WM), a flexible model designed to detect, locate, and identify multiple species in aerial imagery. It addresses the limitations of traditional, labor-intensive wildlife population assessments that are central to advancing environmental conservation efforts worldwide. While a number of methods exist to automate this process, they are often limited in their ability to generalize to different species or landscapes due to the dominance of homogeneous backgrounds and/or poorly captured local image structures. WM introduces two novel modules that help to capture the local structure and context of objects of interest to accurately localize and identify them, achieving a state-of-the-art (SOTA) detection rate of 0.56 mAP. Further, we introduce a large aerial imagery dataset with more than 11k Images and 28k annotations verified by domain experts. WM also achieves SOTA performance on 3 other publicly available aerial survey datasets collected across 4 different countries, improving mAP by " + }, + { + "bbox": [ + 46, + 333, + 289, + 559 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 46, + 333, + 289, + 559 + ], + "type": "text", + "content": ". Source code and trained models are available at Github1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 590, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 687 + ], + "type": "text", + "content": "This paper introduces WildlifeMapper (WM) - an automated and scalable method for counting wildlife in aerial imagery. Aerial wildlife surveys are recognized as a cornerstone of modern conservation biology. By facilitating large-scale biological monitoring in remote landscapes, this technique has underpinned the ability to track changes in the abundance and distribution of wildlife across open landscapes for decades. However, traditional survey approaches" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 316, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 545, + 363 + ], + "type": "text", + "content": "often rely on manual observers to identify, count, and validate species of interest. This labor-intensive process can be time-consuming and error-prone, with potential to limit the utility of final results [27, 35, 38]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 532 + ], + "type": "text", + "content": "Automated approaches offer a promising alternative for efficient and accurate detection of wildlife in aerial survey images. Recent work, for example, illustrates how artificial intelligence has been used to count a variety of species from the air, including antelope in grasslands [44], whales in the ocean [11], and seals on the beach [42]. When combined with advancements in low-cost, high-resolution imaging platforms (e.g., UAVs), these case studies underscore the potential for such data to significantly reduce the effort and cost of traditional wildlife census methods. However, the majority of these techniques struggle to generalize to new species or landscapes due to the dominance of homogeneous backgrounds and poorly captured local structures [7, 12, 18, 39]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 628 + ], + "type": "text", + "content": "WildlifeMapper overcomes these limitations by adapting a novel application of the segment anything transformer model [21]. This model combines high frequency component correlations and spatial correlations in the image data to generate a map of potential locations of objects of interest (i.e., wildlife, livestock). In addition, we address the challenge of identifying multiple species from a relatively small footprint in these images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "To demonstrate the WM analysis workflow, we provide a case study example across the Masai Mara Ecosystem in southwestern Kenya. Renowned for its rich biological diversity, the abundance of large mammals (such as buffalo (Syncerus caffer), giraffe (Giraffa tippelskirchi), and wildebeest (Connochaetes taurinus)) have declined precipitously over the past few decades [34]. Importantly, our analy-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 215, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 215, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 215, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 693, + 215, + 712 + ], + "type": "text", + "content": "https://github.com/UCSB-VRL/WildlifeMapper * refers to equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12594" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 243, + 219 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 243, + 219 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 243, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 243, + 219 + ], + "type": "image", + "image_path": "de19e1e68141ca4c8edb39965055f98a17a22c0096e8d3b8868772137af7bab8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 546, + 249 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 546, + 249 + ], + "type": "text", + "content": "Figure 1. Summary of Mara-Wildlife dataset. (a) Satellite view indicating the four flight trajectories, each represented in a different color. (b, c, d, e, f, g) Annotations of (b) zebra, (c) hartebeest, (d) cattle, (e) shoats (sheep and goats), and (f) zebra. Best viewed in color." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 246, + 72, + 392, + 218 + ], + "blocks": [ + { + "bbox": [ + 246, + 72, + 392, + 218 + ], + "lines": [ + { + "bbox": [ + 246, + 72, + 392, + 218 + ], + "spans": [ + { + "bbox": [ + 246, + 72, + 392, + 218 + ], + "type": "image", + "image_path": "6832a4e5243a8e9398d549c8bd8efb56ec9e6d4c661029589a4e627d132a094b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 468, + 219 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 468, + 219 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 468, + 219 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 468, + 219 + ], + "type": "image", + "image_path": "601a27c2752bd01b8bd0b0412b7d042cf65fa2560ccac096200ec18b93c0b85a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 543, + 219 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 543, + 219 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 543, + 219 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 543, + 219 + ], + "type": "image", + "image_path": "b6607979af0fc92740acb765118178ee2e51d58dd5df1a94ebfed2924da6129d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "type": "text", + "content": "sis incorporated 11, 151 images of size " + }, + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "type": "inline_equation", + "content": "8400 \\times 5500" + }, + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "type": "text", + "content": " collected from a digital camera affixed to the bellyport of a Partenavia P68 airplane during Systematic Reconnaissance Flight (SRF) surveys. Part of these images were annotated by trained observers with 28, 146 annotations of 21 species of large mammals (" + }, + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "type": "inline_equation", + "content": "\\geq 15kg" + }, + { + "bbox": [ + 46, + 259, + 288, + 379 + ], + "type": "text", + "content": "), providing an unprecedented opportunity to develop species detection models across a complex, heterogeneous environment. The dataset was systematically verified by trained observers as described in Section 3. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 386, + 287, + 605 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 386, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 287, + 483 + ], + "type": "text", + "content": "1. A novel, single-stage end-to-end approach for animal detection. The modules, a High Frequency Feature Generator, a Feature Refiner, and a Query Refiner, work together to improve upon the traditional methods of object detection in aerial imagery and enable generalizability across different habitats. The high frequency features reduce dependence on dominant backgrounds/landscapes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 490, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 490, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 490, + 287, + 526 + ], + "type": "text", + "content": "2. An input patch embedding layer that is specifically designed to capture contextual information to help in identifying individual animal species." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 287, + 605 + ], + "type": "text", + "content": "3. The release of a new benchmark dataset via the data owner (Kenya's Wildlife Research and Training Institute - WRTI) once all approvals are in place. An international user community is already engaged in further enhancing these data and using the WM through the BisQue [25] platform." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Community Adoption The practical use of WildlifeMapper (WM) extends beyond theoretical and computational success and is operationalized through BisQue [25]. WM is made available to users through a series of training modules that demonstrate how to " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " upload digital imagery, " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " create annotations, " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " apply and/or improve existing models, " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(iv)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " evaluate model fit, " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(v)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " improve annotations and refit models, and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(vi)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " generate summary statistics. In the fu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 259, + 546, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 546, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 546, + 320 + ], + "type": "text", + "content": "ture, we envision WM to be of great value to ecologists, wildlife managers, and government officials, providing accurate information about the state of wildlife populations in near real-time, facilitating decision-making processes, and improving the conservation of ecosystems globally." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 326, + 397, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 326, + 397, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 326, + 397, + 338 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 413 + ], + "type": "text", + "content": "Manual Methods: Aerial surveys using Front- and Rear-Seat Observers (FSO and RSO, respectively) are commonly used to inventory wildlife populations across open landscapes [36]. However, several important biases can impact these counts, including the experience level and fatigue of the human observers [28]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 414, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 546, + 545 + ], + "type": "text", + "content": "Deep Learning Methods: To address these issues, researchers have begun incorporating digital cameras on piloted aircraft and UAVs [18,30,46]. This minimizes the influence of observer bias while increasing transparency and reproducibility of results. For example, [26] replaced RSOs with an oblique camera mount system minimized the influence of observer bias while producing comparable estimates of large mammals under partial canopy cover. Similarly, [44] used a nadir mounted camera to improve the accuracy and efficiency of manual counts of large antelope in an open grassland ecosystem." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "text", + "content": "However, detecting animals in the wild from aerial imagery poses many challenges. For example, most publicly available datasets for aerial object detection are focused on identifying relatively distinct features such as buildings, roads, vehicles, and other man-made structures [8, 14, 53]. Animals, tend to blend in with their surroundings [13], can be occluded by trees, exhibit considerable variation in color and pattern, or have behavioral adaptations that make them difficult to detect [7, 20, 39, 48]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "[20] proposed a solution for this problem involving a two-branch CNN model based on AlexNet to perform animal recognition and localization. [7] evaluated three state-of-the-art object detection algorithms, including Faster-RCNN, Libra-RCNN, and RetinaNet on six African wild" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12595" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 545, + 158 + ], + "type": "image", + "image_path": "5341e0e4cd475f09e0ed21ef4138abbd9227727e77b1fc66de0b3b25d610b7d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 163, + 523, + 175 + ], + "lines": [ + { + "bbox": [ + 68, + 163, + 523, + 175 + ], + "spans": [ + { + "bbox": [ + 68, + 163, + 523, + 175 + ], + "type": "text", + "content": "Figure 2. Distribution of " + }, + { + "bbox": [ + 68, + 163, + 523, + 175 + ], + "type": "inline_equation", + "content": "(\\geq 15kg)" + }, + { + "bbox": [ + 68, + 163, + 523, + 175 + ], + "type": "text", + "content": " mammals identified in digital imagery collected across the Masai Mara Ecosystem, Kenya." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 178, + 547, + 281 + ], + "blocks": [ + { + "bbox": [ + 48, + 178, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 48, + 178, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 178, + 547, + 281 + ], + "type": "table", + "html": "
Dataset# of annot. images# of annot. tiles# of species# of annotationImage sizeGSD (cm)Location
Virunga73930069656646000x40002.4DRC
Garamba1586429616116000x40002.0DRC
AED2067693871155815500x36002.4-13.0Bostwana, Namibia, South Africa
Mara-Wildlife10127796621281468256x55041.45Masai Mara National Reserve
", + "image_path": "45c559b45cfae964645a60334eea44dab798fccd8d083a370003b4471ab4ea0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "type": "text", + "content": "Table 1. Comparison of Mara-Wildlife dataset with other publicly available dataset. Mara-Wildlife dataset has " + }, + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "type": "text", + "content": " more unique species than the total of all other datasets. Each image is significantly larger and higher ground resolution making 77k unique images of size " + }, + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "type": "inline_equation", + "content": "{1024} \\times {1024}" + }, + { + "bbox": [ + 46, + 283, + 546, + 317 + ], + "type": "text", + "content": " with 21 different animal species. GSD: ground sampling distance; DRC: Democratic Republic of Congo." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 328, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 422 + ], + "type": "text", + "content": "mammals. All three algorithms, however, showed poor performance in animal detection when animals were grouped closely together in herds. [12] adopted a segmentation approach, employing a UNet model to detect livestock from drone imagery. [48] uses a comparable model to analyze high-resolution satellite imagery, producing segmentation masks of wildebeest-sized animals, which are subsequently utilized for detection and counting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "type": "text", + "content": "Transformers: WM adopts a transformer architecture based on past success in modeling different types of aerial imagery. Examples include incorporating multispectral imagery for change detection [4], landcover classification [16], greenhouse gas (GHG) emission detection [23,24] and RGB aerial imagery for object detection [29, 49, 52]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 616 + ], + "type": "text", + "content": "The most effective applications of transformer-based models have been tailored for standard object detection tasks [3, 5, 32, 41, 54]. These works leverage the self-attention to model dependencies among the patches in an end-to-end fashion, unlike CNN-based models [1, 40]. However, when directly applied to aerial imagery, these models cannot effectively exploit the local structures as they divide the image into a sequence of patches. This limits the detection of small-scale objects in a homogeneous and dominant background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Dataset: The existing publicly available animal aerial imagery datasets are listed in Table 1. The Virunga dataset [7] was collected in Virunga National Park, Democratic Republic of Congo (DRC). This dataset contains 897 annotated images of 6 animal species. The Garamba dataset was collected in Garamba National Park, DRC and contains 7034 images. Only 158 images have been annotated, containing 7 animal species. The aerial elephant dataset (AED) [33] was" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 328, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 375 + ], + "type": "text", + "content": "collected across a mosaic of woodland, open shrubland, and grassland habitats in Botswana, Namibia, and South Africa. Only a single species (i.e., elephant) was targeted during SRF surveys. See Table 1 for details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 383, + 437, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 437, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 437, + 396 + ], + "type": "text", + "content": "3. Mara-Wildlife Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 403, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 463 + ], + "type": "text", + "content": "The Mara-Wildlife dataset is a distinctive dataset that captures the essence of the Masai Mara ecosystem through a compilation of 77966 images of size " + }, + { + "bbox": [ + 304, + 403, + 545, + 463 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 403, + 545, + 463 + ], + "type": "text", + "content": ". This habitat is heterogeneous, including woodland, shrubland, and grassland vegetation with 21 unique animal species." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 467, + 408, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 408, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 408, + 480 + ], + "type": "text", + "content": "3.1. Image Collection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": "Flightline Details: Data collection was in collaboration with the Smithsonian National Zoo and Conservation Biology Institute (SNZCBI), Kenya's Wildlife Research and Training Institute (WRTI), the Kenya Wildlife Trust (KWT), and the Directorate of Resource Surveys and Remote Sensing (DRSRS). In March 2022, we fitted a Partenavia P68 with a Nikon D850 digital camera and collected high resolution " + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "inline_equation", + "content": "(8256\\times 5504)" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": " digital images. During data acquisition, the aircraft adhered to a predetermined flight trajectory, depicted in Figure 1, at 400 ft above ground level (agl). This trajectory was optimized to encompass open grassland areas across the Masai Mara ecosystem, including the Masai Mara National Reserve, 22 adjacent private conservancies, and unprotected peripheral areas." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The aerial survey was conducted during a wet season period when the Serengeti migratory population of wildebeest have moved southward to locate more suitable forage in Tanzania. Thus, the survey primarily captured resident species, including wildebeest, zebra, topi, hartebeest, gi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12596" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 190 + ], + "type": "image", + "image_path": "d4cd2368475fed771be868c7297357bd3efcd94f068ce5a916b5aa138c4bee42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "type": "text", + "content": "Figure 3. Overview of WildlifeMapper (WM) architecture. Given an input image of size " + }, + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "type": "inline_equation", + "content": "1024 \\times 1024 \\times 3" + }, + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "type": "text", + "content": ", the High-Frequency Feature Generator (HFG) module generates information about potential location of object of interest. The Feature Refiner (FR) takes these potential locations along with contextual features from Patch Embed layer and sent output to Image Encoder. In parallel, the Query Refiner (QR) incorporates the output of FR to refine learnable queries. Finally these queries are decoded using encoded features from Image Encoder and predict bounding box and class." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "content": "raffe, and other large (" + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\geq 15\\mathrm{kg}" + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "content": ") antelope. Data collection was conducted in the early mornings (prior to 10:00 EAT) and late afternoons (after 15:00 EAT) when lighting conditions were optimal and animals are most active." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": "The geographical positioning of each image was acquired through a GPS system that recorded the plane's altitude, speed, and geographical coordinates. These data were then synchronized with the image using the image capture timestamps, enabling us to determine the geographic location of the centroid of each image. The camera was placed in the bellyport of the airplane, capturing a nadir view of the landscape every two seconds along the flight path." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 415, + 155, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 155, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 155, + 427 + ], + "type": "text", + "content": "3.2. Image Annotation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 429, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 501 + ], + "type": "text", + "content": "Initial bounding box annotations (21796) generated using AIDE platform [19] were exported in CSV format. These were then imported into BisQue [25] for further validation and correction. Finally, these annotations underwent validation by a single trained observer specializing in ecology, resulting in 28146 annotations in total." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 507, + 149, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 149, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 149, + 518 + ], + "type": "text", + "content": "3.3. Dataset Statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "text", + "content": "The Mara-Wildlife dataset showcases a detailed assemblage of wildlife, inclusive of 21 distinct species classes. The dataset is composed of approximately 77,966 tiled images, derived from 1,012 original rasters (Table. 1) The meticulous process of annotation has culminated in labeling 28,146 animals. Species identified include domestic cattle (Bos taurus), white-bearded wildebeest (Connochaetes taurinus), topi (Damaliscus lunatus), shoats (domesticated sheep and goats), kongoni (Alcelaphus buselaphus), waterbuck (Kobus ellipsiprymnus), impala (Aepyceros melampus), Grant's gazelle (Nanger granti), Thomson's gazelle (Eudorcas thomsonii), Cape buffalo (Syncerus caffer), zebra (Equus quagga), ostrich (Struthio camelus), Masai giraffe (Giraffa tippelskirchi), warthog (Phacochoerus africanus), eland (Taurotragus oryx), donkey (Equus africanus), hyena (Crocuta crocuta), hippopotomus" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 263, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 289 + ], + "type": "text", + "content": "(Hippopotamus amphibius), lion (Panthera leo), and elephant (Loxodonta africana)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 295, + 471, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 295, + 471, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 295, + 471, + 308 + ], + "type": "text", + "content": "4. WildlifeMapper Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 316, + 421, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 421, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 421, + 327 + ], + "type": "text", + "content": "4.1. Technical Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 546 + ], + "type": "text", + "content": "WildlifeMapper's architecture is inspired by the success of the Segment Anything Model (SAM) [21], created to segment small/large (all sizes) of objects. Referring to Fig. 3, WM contains the following main components: (i) A patch embedding layer designed to capture long-range context, (ii) a High-Frequency Feature Generator (HFG), (iii) a Feature Refiner (FR) followed by ViT based image encoder [21,23,45], and (iv) a Query Refiner (QR) module followed by a box decoder module. The input image is first processed through two separate branches, the patch embedding layer which captures long-range context [9] and HFG which suppresses all the low-frequency components in the image and generates a feature embedding. The HFG (Sec. 4.3) exploits prior knowledge that aerial images from areas such as forests, grasslands, and shrublands have a homogeneous and dominant background representing the dominant low frequency image content. Fig. 4 shows that on suppressing the lower frequencies, the object of interest is easy to locate." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 547, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 654 + ], + "type": "text", + "content": "The FR (Sec. 4.4) takes the embeddings from each of the two branches and generates a high quality embedding that contains information about potential locations of animals and captures the local context. The QR modules refines a set of learnable queries using the location information from FR module. These refined queries are passed to the box decoder. The box decoder takes the refined queries and encoded features image encoder to generate the final detection box and class of the object." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 664, + 391, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 391, + 675 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 391, + 675 + ], + "type": "text", + "content": "4.2. Patch Embed" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "The patch embedding layer utilizes a larger kernel convolution with an increasing dilation rate. This design rapidly expands the receptive field, allowing explicit extrac" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12597" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "tion of features rich in contextual information. This approach is particularly beneficial for aerial imagery, where the small sized object makes classification based on appearance alone challenging. Contextual information thus becomes crucial for the accurate recognition of these objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 139, + 267, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 267, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 267, + 152 + ], + "type": "text", + "content": "4.3. High-frequency Feature Generator (HFG)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 154, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 154, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 287, + 260 + ], + "type": "text", + "content": "Along with patch embedding, the input image is processed in parallel by the HFG module to generate features with information about the location of the animal or cluster as shown in Fig. 4. The HFG module is inspired from the limitation of ViT models [45]. ViT models face challenges in efficiently utilizing local structures. They segment an image into patches and apply self-attention to model relationships, but this approach often falls short in capturing detailed local features [37, 51]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 344 + ], + "type": "text", + "content": "Local features in images are closely linked to high-frequency components [2, 43]. We hypothesize that suppressing low-frequency components can mitigate the influence of a dominant homogeneous background. To test this, we performed a discrete Fourier Transform (DFT) on the images, filtering out the low-frequency components before reconstructing the images, as shown in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "content": "For a given input image " + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "content": " is channel dimension, we compute Discrete Fourier Transform (DFT) of " + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "content": ". In next step we suppress the low frequency components with a controlling parameter and construct the image " + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "content": " with inverse DFT (IDFT)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 413, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 413, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 108, + 413, + 287, + 426 + ], + "type": "interline_equation", + "content": "I ^ {\\prime} = I D F T [ h p f (D F T (I) ] \\tag {1}", + "image_path": "088fd23a56803ce881a1561eeabeb25ec5d452efb8374e8038a12972cc63306e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "hpf" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": " is a high pass filter. Then we reduce the dimension of the reconstructed image " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": " via an embedding layer to generate embedding " + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "inline_equation", + "content": "hfc_{emb}" + }, + { + "bbox": [ + 46, + 434, + 287, + 483 + ], + "type": "text", + "content": " and pass them to the FR module. See supplementary materials for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 490, + 168, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 168, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 168, + 502 + ], + "type": "text", + "content": "4.4. Feature Refiner (FR)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "text", + "content": "Next, the features from the patch embedding layer " + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "inline_equation", + "content": "(f_{img})" + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "text", + "content": " and HFG " + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "inline_equation", + "content": "(hfc_{emb})" + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "text", + "content": " are fed to the FR module. The " + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "text", + "content": " are refined with the " + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "inline_equation", + "content": "hfc_{emb}" + }, + { + "bbox": [ + 46, + 504, + 287, + 575 + ], + "type": "text", + "content": " via cross-attention mechanism. The FR is a simple module with cross-attention and linear layers [45]. The output contains information about the potential location of the object and the long-range context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": "Following the standard architecture of SAM [21], we pass FR output to our ViT based image encoder supplemented with learnable positional embeddings " + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": ". The encoded feature map is " + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "inline_equation", + "content": "f_{emb}" + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 632, + 287, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 632, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 89, + 632, + 287, + 645 + ], + "type": "interline_equation", + "content": "f _ {e m b} = \\operatorname {V i T} \\left[ \\mathbf {F R} \\left(f _ {i m g}, h f c _ {e m b}\\right), p \\right] \\tag {2}", + "image_path": "d5ed23734c4f7d2c7a00570fc87dff20b9267fea9465176ab608a7de6c7a56ca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Query Refiner (QR): The QR follows a transformer decoder like architecture and takes as input a set of 100 learnable queries " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{100 \\times d}" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " and output of FR module. Here " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " is same as channel dimension of " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "f_{emb}" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " from image encoder. The FR output refines the " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " via a cross-attention" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 545, + 322 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 322 + ], + "type": "image", + "image_path": "e90a672b29dee430d653ce2754938f7657c70c98fd61dd4ef0fa4e9a0ea10bb1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 325, + 547, + 435 + ], + "lines": [ + { + "bbox": [ + 305, + 325, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 547, + 435 + ], + "type": "text", + "content": "Figure 4. The sample output visualization from the High-Frequency Feature Generator (HFG) module. The illustration shows the effectiveness of the module in suppressing the homogeneous and dominant background, while highlighting objects of interest (i.e., animals). The top image shows bomas, natural structures constructed to contain livestock, and paths that have been suppressed. Animals, however, are clearly identified, especially inside the boma. The bottom image shows a water body (a dam created for livestock) that has been suppressed by the module. Animals can again be highlighted throughout the image." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 451, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 451, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 545, + 487 + ], + "type": "text", + "content": "mechanism. The refined queries narrows the search space for box decoder module to accurately locate and identify object of interest. [15,23]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 504, + 388, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 388, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 388, + 516 + ], + "type": "text", + "content": "4.5. Box Decoder" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Next, the refined queries are sent to the box decoder module. We concatenate these queries with the " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{emb}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " and pass box decoder's self attention layer as " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "qf_{emb}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": ". This is inspired from the idea of class_token used by Vaswani et.al. [45] to make decoding process memory efficient. The transformer network takes 4 input variables, those are position embeddings of queries and image embeddings, " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "qf_{emb}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{emb}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": ". Our transformer model uses two-way attention inspired from [3, 6, 21] and our box decoder uses self-attention and cross-attention in two directions (queries-to-image embedding and vice-versa) to update all embeddings. We keep the box decoder very light weight (two blocks). The top 100 (equal to number of queries) indexes from the output of the final block is passed to two separate MLP blocks to regress the output bounding box prediction and class of the predicted box." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12598" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "text", + "content": "4.6. Training and Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "type": "text", + "content": "We've trained WM using a single-stage, end-to-end approach to determine bounding boxes and classify them. The loss strategy we applied for WM is akin to what's used in DETR [3]. Initially, we perform bipartite matching to align our model's predictions with the actual bounding box data. Then, we proceed to compute the loss for these matched pairs. To achieve the best possible match between our predictions and the real data, we use the Hungarian algorithm [22]. Once matched, each prediction is paired with its respective ground truth. We then measure the " + }, + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "type": "text", + "content": " (L1 distance) and " + }, + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "type": "inline_equation", + "content": "GIoU" + }, + { + "bbox": [ + 46, + 90, + 289, + 236 + ], + "type": "text", + "content": " loss for the bounding box and the cross entropy loss for the classification [3]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "content": "Inference: The inference pipeline is straightforward and similar to training code. During inference, we first filter the detections at " + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "content": " threshold and then use non-maximal suppression to remove any overlapping boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 302, + 128, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 128, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 128, + 316 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": "Train-val-test split: There is no data leak in train-val-test split. Fig. 1 shows the flight paths with location of each image represented by a circle. Each color coded flight path represents the train-val-test image set. No images taken while the airplane was cutting across the transects plus turnaround points as can be seen there are no circles. We created a spatial disjoint of " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "20km" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " distance between the transects as shown in Fig. 1 and achieved consistent performance (0.56mAP)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 536 + ], + "type": "text", + "content": "Evaluation Metrics: We report our performance on multiple metrics. Following the protocols of standard object detection, we report the performance in mean average-precision (mAP) for detection and mean intersection-overunion (mIOU) for localization of animals. We also report a commonly used metric by the ecologists on the team, the class-wise mean absolute error (MAE) indicating the counting accuracy of each species:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 543, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 543, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 100, + 543, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\mathrm {M A E} = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\sum_ {c = 1} ^ {C} | \\hat {n} _ {i, c} - n _ {i, c} |, \\tag {3}", + "image_path": "06d18a092505e6fbcfb463d31eda3413cc813398c495f39fa561edde2cb3fc57.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": " is the number of images, " + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "inline_equation", + "content": "\\hat{n}_{i,c}" + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "inline_equation", + "content": "n_{i,c}" + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": " are the predicted and ground truth counts for class " + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 585, + 288, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Implementation Details: Each image taken from the drone is " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "8256 \\times 5506 \\times 3" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": ". We create tiles for each image in the spatial domain, with the size of " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024 \\times 3" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " of overlap. The Patch Embed layer uses a single CNN layer with a large kernel of size " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " with stride 16. In the parallel branch, the High-Frequency Feature Generator, we use DFT to compute the Fourier transform, the mask is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "content": "a binary disk with the radius set to 128. The HFC Embed layer uses 3 CNN layers with ReLU activation with a kernel of size " + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "content": " and a global average pool at the end. The Feature Refiner (FR) module consists of one cross attention layer with 1 linear layer. The image encoder is a pre-trained ViT model [21] with 24 transformer layers and 16 heads. The Query Refiner (QR) module takes in 100 queries each of channel dimension 256, those are cross attended with " + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "inline_equation", + "content": "hfc_{emb}" + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "content": " output. The box decoder contains 3 layers of two-way attention with 8 heads. We train WM with AdamW optimizer [31] setting the learning rate to " + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "content": " for the FR, QR and box decoder with a weight decay to " + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "content": ". We set the learning rate for the Patch Embed and HFC Embed layer to " + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 72, + 547, + 254 + ], + "type": "text", + "content": ". We load the image encoder with pre-trained weights from segment anything [21] and keep it frozen." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": "Data Augmentation: In the train and test datasets, we incorporated an equal number of images without any objects to assess the model's robustness against empty background images. We applied multiple data augmentation techniques, including HSV (hue, saturation, and value) " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": ", affine transformation " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(20\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": ", shear " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": " and mosaic " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "(70\\%)" + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "text", + "content": " augmentation [1]. Mosaic augmentation is proven to be the most effective, with an improvement of " + }, + { + "bbox": [ + 304, + 262, + 547, + 370 + ], + "type": "inline_equation", + "content": "0.07\\mathrm{mAP}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 452 + ], + "type": "text", + "content": "Hard Negative Mining: After training for 100 epochs, we take all the False Positives (FP) predictions having " + }, + { + "bbox": [ + 304, + 380, + 547, + 452 + ], + "type": "inline_equation", + "content": "\\mathrm{IOU} \\leq 0.10" + }, + { + "bbox": [ + 304, + 380, + 547, + 452 + ], + "type": "text", + "content": " with ground truth box and mark them as background class. Then fine-tuning for 20 epochs improved the performance of FP reduction for detecting rocks, dead tress or other artifacts on ground as animal." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 464, + 359, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 359, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 359, + 475 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 479, + 450, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 450, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 450, + 493 + ], + "type": "text", + "content": "6.1. Performance Comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "We trained WM separately on Mara-Wildlife dataset and Virunga-Garamba-AED dataset for comparison with existing works, see Table 2 for a summary of the results on all of the tested datasets. The " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " values are compared for IoU of 0.50-0.95 and 0.50. We also provide the average counting error in animal counting per image. We trained all the baseline models with the default set of conditions on Mara-Wildlife dataset. We merged the Virunga, Garamba, and AED datasets and created the train-val-test split according to [7]. The combination of these datasets contains 6 unique animal species and diversity of landscapes such as woodland, savannahs, open shrubland, and grasslands across multiple countries – Democratic Republic of Congo (DRC), Botswana, Namibia, and South Africa. WM outperforms all methods by a significant margin as shown in Table 2. We note that in [7] the authors did not make the code base or trained model public, hence we could not verify the results. We implemented these methods from" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12599" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 287, + 246 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 246 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 246 + ], + "type": "table", + "html": "
Methods#epochsmAPmAP50Counting Error
Mara Wildlife Dataset
1Faster-Rcnn [40]1000.240.582.59
2DETR [3]2000.220.572.75
3Co-DETR-R50 [55]1000.270.662.72
4Co-DETR-swingL [55]1000.280.652.60
5Yolo v5 [10]1000.300.672.12
6Yolo v8 [17]1000.270.613.97
7LSKNet [29]1000.290.66-
6DroneDetect [50]1000.180.48-
8WildlifeMapper1200.560.791.9
Virunga-Garamba-AED Datasets
6Faster-Rcnn1200.340.650.27
7DETR2000.300.620.45
8Yolo v51000.480.780.12
9Yolo v81000.480.770.42
10WildlifeMapper800.680.850.11
", + "image_path": "af1e9fc6f10b2e77c1f83634605266101fe9e06ddc61174da47e525bcbdfb1e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 94, + 320, + 241, + 381 + ], + "blocks": [ + { + "bbox": [ + 46, + 254, + 287, + 320 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 320 + ], + "type": "text", + "content": "Table 2. Comparison with baseline models. The top section shows performance on species detection on Mara-wildlife dataset and low section shows performance on the mixed dataset from Virunga-Garamba-AED dataset. The overall detection accuracy is generally higher in Virunga-Garamba-AED dataset because there are only 6 species and the terrain is quite similar in all images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 320, + 241, + 381 + ], + "lines": [ + { + "bbox": [ + 94, + 320, + 241, + 381 + ], + "spans": [ + { + "bbox": [ + 94, + 320, + 241, + 381 + ], + "type": "table", + "html": "
HFGFRQRmAP
WMXX*0.46
✓✓*0.54
✓X✓0.49
✓✓✓0.56
", + "image_path": "00318b09bc148ca7c22a87b4d020bdd9e3f436bc12341f0cf97f6d1efe9b1cb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 390, + 287, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 434 + ], + "type": "text", + "content": "Table 3. HFG module effectiveness in refining the image features and queries. “X” represents not used, “√” represents used and “*” represents that random queries are used but there was not refining with HFG features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": "the original public repositories and trained according to the training strategy detailed in [7]. We attribute the model poor learning performance due to salient features of the homogeneous background being learned more than the object of interest. The detection of the object of interest is then dependent on the landscape properties instead of object properties. Hence when used on a slight variations of landscapes for the same object, the models struggle to detect. This limitation is specifically addressed in the WM, where the HFG modules suppresses the background and highlights the object of interest." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 632 + ], + "type": "text", + "content": "Qualitative results: Fig. 5 shows the quality of detection by WM in different scenarios. Those include, detection when animal is partially visible under a tree, or a big clustering. WM makes correct predictions in varying scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 148, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 148, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 148, + 651 + ], + "type": "text", + "content": "6.2. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "content": "We performed all ablation experiments on Mara-Wildlife (MW) dataset and validate the design choices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "High-frequency Feature Generator Module: In Table 3, we show the effectiveness of the HFG module. We" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": "experimented with HFG's output in 3 ways: first, we passed HFG's output to Feature Refiner (FG) module. It leads to significant improvement in detection by " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "0.09\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " over the baseline. This demonstrates that providing potential location candidates features to image encoder module produce better embeddings. Second, we pass the HFG's output to Query Refiner (QR) module only. This leads to an improvement of " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "0.03\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " over baseline. This shows the effectiveness of guiding queries with location candidates features. In the third case, we passed the HFG's output to both FR and QR modules and achieved an improvement of " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "0.11\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " over the baseline. We hypothesize that this reduces the dominance of features from homogeneous and dominant background in aerial imagery. The also observed this while testing WM across flightlines different types of terrain such as green grasslands, dry grasslands, and forest areas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": "Feature Refiner Module: We tested " + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "inline_equation", + "content": "hfc_{emb}" + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": " merging by 3 ways: addition, concatenation and cross-attention. Cross-attention is most effective, because with addition and concatenation, the " + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "inline_equation", + "content": "hfc_{emb}" + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": " get lost, while cross-attention generates better embeddings giving attention to potential location candidates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": "Kernel Size: We observed that a larger kernel size of " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "31 \\times 31" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": " results in reduced misclassification. For example, a topi or warthog cannot be found inside a boma because only domestic species are kept in bomas. So context helps in making the right class detection. We observed an improvement of " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "0.02\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": ". Experiments were done in 3 kernel sizes. 1. " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "0.55\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": ", 2. " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "0.558\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": "; 3. " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "inline_equation", + "content": "0.57\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 365, + 545, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "content": "Query Refiner Module: We experimented with only providing random queries and guiding the queries with HFG module output. We merged them with direct addition or concatenation and cross-attention. With cross-attention, we observed an improvement in performance of 0.03mAP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "type": "text", + "content": "Geographic generalization: To test the geographic generalizability of WM across different terrains, we trained WM only on images from Kenya; and tested on images from Democratic Republic of Congo, Botswana and Namibia. The test was done on 4 common species which were present in both the ecosystems. WM achieved a detection performance of " + }, + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "type": "inline_equation", + "content": "0.48\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 550, + 545, + 646 + ], + "type": "text", + "content": ". This shows the adaptability of WM across varying landscapes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 661, + 545, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 545, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 545, + 719 + ], + "type": "text", + "content": "Domain generalization: We train-test WM on a different domain, a bird species tern [47], commonly found on/near water bodies. Live in huge clusters. WM achieved the accuracy of " + }, + { + "bbox": [ + 304, + 661, + 545, + 719 + ], + "type": "inline_equation", + "content": "0.71\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 661, + 545, + 719 + ], + "type": "text", + "content": ". Showing adaptability of WM across different domains." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12600" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 218, + 227 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 218, + 227 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 218, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 218, + 227 + ], + "type": "image", + "image_path": "7da36f287a9680a3d38a1103eecc5ccdef580a499bb64df42d28969e962847cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 218, + 70, + 375, + 227 + ], + "blocks": [ + { + "bbox": [ + 218, + 70, + 375, + 227 + ], + "lines": [ + { + "bbox": [ + 218, + 70, + 375, + 227 + ], + "spans": [ + { + "bbox": [ + 218, + 70, + 375, + 227 + ], + "type": "image", + "image_path": "cdcaf9e16c6949d3362b62f597dea6482cbc320233d295c1cbc888049b210f7c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 533, + 227 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 533, + 227 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 533, + 227 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 533, + 227 + ], + "type": "image", + "image_path": "c02f6e19018cec214b4ef2b317f0a06d7139d8cfb673145f181fe78734974b5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 228, + 217, + 384 + ], + "blocks": [ + { + "bbox": [ + 60, + 228, + 217, + 384 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 217, + 384 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 217, + 384 + ], + "type": "image", + "image_path": "709b87d56af4929a8775f447c530f3f8c078e8dcb0a591f3f3e032e5ac283b4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 547, + 425 + ], + "type": "text", + "content": "Figure 5. Qualitative results. The top row highlights examples of crowded and partially occluded scenes. Row-1, Column-1 & Row-3, Column-3 shows examples where animals are obstructed by shadows. The zoomed-in box in Row-3, Column-3 shows a zebra partially occluded. The bounding box color is coded according to class names: shoats- \"hot pink\", cattle- \"deep sky blue\", zebra- \"light yellow\"." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 218, + 228, + 375, + 384 + ], + "blocks": [ + { + "bbox": [ + 218, + 228, + 375, + 384 + ], + "lines": [ + { + "bbox": [ + 218, + 228, + 375, + 384 + ], + "spans": [ + { + "bbox": [ + 218, + 228, + 375, + 384 + ], + "type": "image", + "image_path": "ead6d7a1e79e473d59797ae734c2b69f4ac5c21e165bff6a24b13a98c9900668.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 376, + 228, + 533, + 384 + ], + "blocks": [ + { + "bbox": [ + 376, + 228, + 533, + 384 + ], + "lines": [ + { + "bbox": [ + 376, + 228, + 533, + 384 + ], + "spans": [ + { + "bbox": [ + 376, + 228, + 533, + 384 + ], + "type": "image", + "image_path": "5436f5675f603e8e60d9191169ab6556dc0ddc4c11a9fb98eb5829baf103bcf5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 54, + 442, + 167, + 556 + ], + "blocks": [ + { + "bbox": [ + 54, + 442, + 167, + 556 + ], + "lines": [ + { + "bbox": [ + 54, + 442, + 167, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 442, + 167, + 556 + ], + "type": "image", + "image_path": "418c4bb1f01517b5e27717d0e376fd49549229379a0488650fdb0ee07dd02377.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 562, + 288, + 608 + ], + "lines": [ + { + "bbox": [ + 46, + 562, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 288, + 608 + ], + "type": "text", + "content": "Figure 6. Failure cases. Left shows an example where animals are occluded by shadows and are difficult to detect. Right shows an example of rock detected as an impala, emphasizing the difficulty in differentiating objects in the image from animals of interest." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 167, + 442, + 280, + 556 + ], + "blocks": [ + { + "bbox": [ + 167, + 442, + 280, + 556 + ], + "lines": [ + { + "bbox": [ + 167, + 442, + 280, + 556 + ], + "spans": [ + { + "bbox": [ + 167, + 442, + 280, + 556 + ], + "type": "image", + "image_path": "05616dc7838c4257c59d1736e02807be40161ffe63b1433f9026916dfb0b63d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Failure Cases: The detections from WM are inaccurate when animals are clustered in shadows, such as when animals are located inside bomas and the sun angle makes a strong shadow on the enclosure. These are some of the difficult cases shown in Fig. 6. Other cases of false positives are the small rocks or trees that sometimes resemble animals. Some examples are shown in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 443, + 379, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 379, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 379, + 456 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 464, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 572 + ], + "type": "text", + "content": "This paper presents WildlifeMapper (WM) - a transformer based approach for the detection of animals of varying densities and sizes across natural backgrounds. The WM utilizes a high frequency features generator, feature refiner, and query refiner to accurately locate and classify 8 animal species. WM stands to significantly improve the efficiency and accuracy of wildlife monitoring and conservation efforts. Future work will extend this model and dataset to a larger number of species and habitats." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 582, + 416, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 416, + 596 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 416, + 596 + ], + "type": "text", + "content": "8. Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 601, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 687 + ], + "type": "text", + "content": "This research is partially supported by the following grants: NSF award SI2-SSI #1664172 (BSM), the Ohrstrom Family Foundation, and a Smithsonian National Zoo and Conservation Biology Institute (SNZCBI) fellowship (S.Kumar). We also recognize the contributions of the annotation team, including S. Harman, A. Gray, C. Pate, C. Obath, S. Nyarangi, C. Kagume, and W. Sairowua." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12601" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "type": "text", + "content": "[2] Fergus W Campbell and John G Robson. Application of fourier analysis to the visibility of gratings. The Journal of physiology, 197(3):551, 1968. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "text", + "content": "[4] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[5] Qiang Chen, Jian Wang, Chuchu Han, Shan Zhang, Zexian Li, Xiaokang Chen, Jiahui Chen, Xiaodi Wang, Shuming Han, Gang Zhang, et al. Group detr v2: Strong object detector with encoder-decoder pretraining. arXiv preprint arXiv:2211.03594, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "text", + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 412 + ], + "type": "text", + "content": "[7] Alexandre Delplanque, Samuel Foucher, Philippe Lejeune, Julie Linchant, and Jérôme Théau. Multispecies detection and identification of african mammals in aerial imagery using convolutional neural networks. Remote Sensing in Ecology and Conservation, 8(2):166-179, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "text", + "content": "[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44(11):7778-7796, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 533 + ], + "type": "text", + "content": "[9] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 556 + ], + "type": "text", + "content": "[10] Glenn Jocher et. al. ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements, Oct. 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[11] Emilio Guirado, Siham Tabik, Marga L Rivas, Domingo Alcaraz-Segura, and Francisco Herrera. Whale counting in satellite and aerial images with deep learning. Scientific reports, 9(1):14259, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[12] Liang Han, Pin Tao, and Ralph R Martin. Livestock detection in aerial images using a fully convolutional network. Computational Visual Media, 5:221-228, 2019. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[13] Tracey Hollings, Mark Burgman, Mary van Andel, Marius Gilbert, Timothy Robinson, and Andrew Robinson. How do you find the green sheep? a critical review of the use of remotely sensed imagery to detect and count animals. Methods in Ecology and Evolution, 9(4):881-892, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H. Hsu. Drone-based object counting by spatially regularized re" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "gional proposal networks. In The IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[15] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. In Pattern Recognition and Tracking XXXIV, volume 12527, pages 192-205. SPIE, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 151, + 518, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 518, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 518, + 161 + ], + "type": "text", + "content": "[16] Johannes Jakubik et. al. Prithvi-100M, Aug. 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 182 + ], + "type": "text", + "content": "[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics* YOLO, Jan. 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "text", + "content": "[18] Benjamin Kellenberger, Diego Marcos, and Devis Tuia. Detecting mammals in uav images: Best practices to address a substantially imbalanced dataset with deep learning. Remote Sensing of Environment, 216:139-153, Oct 2018. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[19] Benjamin Kellenberger, Devis Tuia, and Dan Morris. Aide: Accelerating image-based ecological surveys with interactive machine learning. Methods in Ecology and Evolution, 11(12):1716-1727, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 273, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 326 + ], + "type": "text", + "content": "[20] Benjamin Kellenberger, Michele Volpi, and Devis Tuia. Fast animal detection in uav images using convolutional neural networks. In 2017 IEEE international geoscience and remote sensing symposium (IGARSS), pages 866-869. IEEE, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 4, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "type": "text", + "content": "[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 405, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 459 + ], + "type": "text", + "content": "[23] Satish Kumar, Ivan Arevalo, ASM Iftekhar, and BS Manjunath. Methanemapper: Spectral absorption aware hyperspectral transformer for methane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17609-17618, 2023. 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "text", + "content": "[24] Satish Kumar, William Kingwill, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan D Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. In NeurIPS 2022 Workshop on Tackling Climate Change with Machine Learning, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "text", + "content": "[25] Kristian Kvilekval, Dmitry Fedorov, Boguslaw Obara, Ambuj Singh, and BS Manjunath. Bisque: a platform for bioimage analysis and management, Jan. 2023. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[26] Richard Lamprey, David Ochanda, Rob Brett, Charles Tumwesigye, and Iain Douglas-Hamilton. Cameras replace human observers in multi-species aerial counts in murchison falls, uganda. Remote Sensing in Ecology and Conservation, 6(4):529-545, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "type": "text", + "content": "[27] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma, and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Richard Lamprey, Frank Pope, Shadrack Ngene, Michael Norton-Griffiths, Howard Frederick, Benson Okita-Ouma," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12602" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "text", + "content": "and Iain Douglas-Hamilton. Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in tsavo, kenya: Taking multi-species aerial counts to the next level. Biological Conservation, 241:108243, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[29] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. arXiv preprint arXiv:2303.09030, 2023. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[30] Julie Linchant, Jonathan Lisein, Jean Semeki, Philippe Lejeune, and Cedric Vermeulen. Are unmanned aircraft systems (uas s) the future of wildlife monitoring? a review of accomplishments and challenges. Mammal Review, 45(4):239-252, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 252 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "text", + "content": "[33] Johannes Naude and Deon Joubert. The aerial elephant dataset: A new public benchmark for aerial object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 48-55, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "type": "text", + "content": "[34] Joseph O Ogutu, Hans-Peter Piepho, Mohamed Y Said, Gordon O Ojwang, Lucy W Njino, Shem C Kifugo, and Patrick W Wargute. Extreme wildlife declines and concurrent increase in livestock numbers in kenya: What are the causes? PloS one, 11(9):e0163249, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 465 + ], + "type": "text", + "content": "[35] Wilber K Ottichilo, Jesse Grunblatt, Mohammed Y Said, and Patrick W Wargute. Wildlife and livestock population trends in the kenya rangeland. *Wildlife conservation by sustainable use*, pages 203–218, 2000. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[36] Wilber K. Ottichilo, Jesse Grunblatt, Mohammed Y. Said, and Patrick W. Wargute. Wildlife and Livestock Population Trends in the Kenya Rangeland, pages 203-218. Springer Netherlands, Dordrecht, 2000. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 533 + ], + "type": "text", + "content": "[37] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "type": "text", + "content": "[38] Michael John Stephen Peel and Marc Stalmans. The systematic reconnaissance flight (srf) as a tool in assessing the ecological impact of a rural development programme in an extensive area of the lowveld of south africa. African Journal of Ecology, 37(4):449-456, 1999. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[39] Tinao Petso, Rodrigo S Jamisola Jr, Dimane Mpoeleng, Emily Bennitt, and Wazha Mmereki. Automatic animal identification from drone camera based on point pattern analysis of herd behaviour. Ecological Informatics, 66:101485, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[42] AC Seymour, J Dale, M Hammill, PN Halpin, and DW Johnston. Automated detection and enumeration of marine wildlife using unmanned aircraft systems (uas) and thermal imagery. Scientific reports, 7(1):45127, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 142, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 174 + ], + "type": "text", + "content": "[43] Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal on mathematical analysis, 29(2):511-546, 1998. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 241 + ], + "type": "text", + "content": "[44] Colin J Torney, David J Lloyd-Jones, Mark Chevallier, David C Moyer, Honori T Maliti, Machoke Mwita, Edward M Kohi, and Grant C Hopcraft. A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images. Methods in Ecology and Evolution, 10(6):779-787, 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 286 + ], + "type": "text", + "content": "[45] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "type": "text", + "content": "[46] Cédric Vermeulen, Philippe Lejeune, Jonathan Lisein, Prosper Sawadogo, and Philippe Bouché. Unmanned aerial survey of elephants. PloS one, 8(2):e54700, 2013. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 322, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 388 + ], + "type": "text", + "content": "[47] Ben G Weinstein, Lindsey Garner, Vienna R Saccomanno, Ashley Steinkraus, Andrew Ortega, Kristen Brush, Glenda Yenni, Ann E McKellar, Rowan Converse, Christopher D Lippitt, et al. A general deep learning model for bird detection in high-resolution airborne imagery. Ecological Applications, 32(8):e2694, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 389, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 545, + 454 + ], + "type": "text", + "content": "[48] Zijing Wu, Ce Zhang, Xiaowei Gu, Isla Duporge, Lacey F. Hughey, Jared A Stabach, Andrew K Skidmore, J Grant C Hopcraft, Stephen J Lee, Peter M Atkinson, et al. Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape. Nature communications, 14(1):3072, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 510 + ], + "type": "text", + "content": "[49] Xiangkai Xu, Zhejun Feng, Changqing Cao, Mengyuan Li, Jin Wu, Zengyan Wu, Yajie Shang, and Shubing Ye. An improved swim transformer-based model for remote sensing object detection and instance segmentation. Remote Sensing, 13(23):4779, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[50] Yinhui Yu, Xu Sun, and Qing Cheng. Expert teacher based on foundation image segmentation model for object detection in aerial images. Scientific Reports, 13(1):21964, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "type": "text", + "content": "[51] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[52] Jiahe Zhu, Xu Chen, Huan Zhang, Zelong Tan, Shengjin Wang, and Hongbing Ma. Transformer based remote sensing object detection with enhanced multispectral feature extraction. IEEE Geoscience and Remote Sensing Letters, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[53] Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Ling, and Qinghua Hu. Vision meets drones: A challenge. arXiv preprint arXiv:1804.07437, 2018. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12603" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 162 + ], + "type": "text", + "content": "[55] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12604" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_content_list.json b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7652292b0ea9bdd80708b24e14f6583b153014e2 --- /dev/null +++ b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_content_list.json @@ -0,0 +1,1276 @@ +[ + { + "type": "text", + "text": "WinSyn: A High Resolution Testbed for Synthetic Data", + "text_level": 1, + "bbox": [ + 205, + 130, + 764, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tom Kelly1 John Femiani2 Peter Wonka1", + "bbox": [ + 297, + 179, + 665, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KAUST, KSA $^{2}$ Miami University, Ohio", + "bbox": [ + 326, + 203, + 620, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 255, + 313, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present WinSyn, a unique dataset and testbed for creating high-quality synthetic data with procedural modeling techniques. The dataset contains high-resolution photographs of windows, selected from locations around the world, with 89,318 individual window crops showcasing diverse geometric and material characteristics. We evaluate a procedural model by training semantic segmentation networks on both synthetic and real images and then comparing their performances on a shared test set of real images. Specifically, we measure the difference in mean Intersection over Union (mIoU) and determine the effective number of real images to match synthetic data's training performance. We design a baseline procedural model as a benchmark and provide 21,290 synthetically generated images. By tuning the procedural model, key factors are identified which significantly influence the model's fidelity in replicating real-world scenarios. Importantly, we highlight the challenge of procedural modeling using current techniques, especially in their ability to replicate the spatial semantics of real-world scenarios. This insight is critical because of the potential of procedural models to bridge to hidden scene aspects such as depth, reflectivity, material properties, and lighting conditions.", + "bbox": [ + 76, + 287, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Larger and more sophisticated machine learning models demand an ever-increasing supply of data, particularly in tasks where manual annotation is challenging, such as depth estimation, reflectance estimation, or full 3D reconstruction. One solution is to use procedural models to generate synthetic training data. However, creating procedural models that accurately reflect the domain of real images (i.e. closing the domain gap) remains an open problem. Despite the visual realism of many synthetic scenes, their effectiveness in machine-learning applications often falls short. In this work we do not close the gap – but present an accessible pair of real and synthetic datasets, with annotations, in which this domain gap may be studied without massive resources.", + "bbox": [ + 76, + 688, + 468, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the final goal is to tackle complex problems, a", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "straightforward proxy task is initially required to ensure that real-world imagery can be manually annotated and compared to synthetic imagery. We, therefore, propose segmentation as a proxy task. This proxy task helps pinpoint where a procedural model fails to capture the diversity and nuances of real-world scenes.", + "bbox": [ + 496, + 257, + 893, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As our long-term goal is procedural urban modeling, we initially considered street-view images and datasets, such as CityScapes [12]. However, this has several drawbacks. Modeling arbitrary urban scenes realistically requires overwhelming complexity; requiring modeling at least cars, humans, buildings, skies, and vegetation, and each of these poses quite distinct and complex modeling challenges. For", + "bbox": [ + 496, + 349, + 893, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7b8acacaf2fae6521302ba70178834a0722853ed224986707a746e8eed9e5e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 484, + 854, + 652 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a682de236db7c90e79bec5ef5b56f69af41e62115dfada4b5b657902f43de5ca.jpg", + "image_caption": [ + "Figure 1. Photographers in 28 geographic regions captured real-world photos (a) of windows that are cropped (b) to single windows, which are then labeled (c). Synthetic windows are rendered giving color (f) and labels(d), while other passes such as depth (e) are also possible." + ], + "image_footnote": [], + "bbox": [ + 540, + 654, + 854, + 816 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22456", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "example, modeling urban environments for video games is a significant undertaking that often requires over 100 artists for open-world games. This type of effort is unrealistic for procedural modeling research. We propose that it is more promising to identify a simpler subset of images and develop a more constrained procedural model for this subset. This should enable faster iterations and broader participation in procedural modeling research, before scaling up to complete cities.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The decision to study the domain of window images was based on the following design principles for our dataset:", + "bbox": [ + 76, + 228, + 468, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diversity: Our goal is to capture a range of variations in both geometry and materials. We opted against humans due to the complexities in modeling realistic fine geometric details (like hair and beard) and materials (such as skin), and the limited topological variations. Although humans have been used for synthetic data studies as demonstrated in [39], exploration in this domain is complicated because the procedural model was not released and high-fidelity models such as Unreal Metahumans by Epic Games [15] tend to prioritize realism over variability. Plants, another option, present challenges in material diversity and are difficult for human annotators to segment. Windows, in contrast, offers a balanced combination of geometric complexity and material variation suitable for our research goals.", + "bbox": [ + 75, + 260, + 472, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "High-resolution: The images in the dataset should be high resolution to make the dataset useful for the near and medium-term future. This is in contrast to existing architectural datasets, with resolutions typically in the 512-1025 pixel range [5, 6, 25, 33]. While these are not low resolution in typical computer vision terms, architectural features such as windows are often too small to resolve in the images. Although the CityScapes dataset is larger than most, with a resolution of $2048 \\times 1024$ pixels, it often captures architecture from oblique viewpoints limiting the effective resolution.", + "bbox": [ + 75, + 488, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Rights: We would like to have all rights to the images to avoid future copyright problems.", + "bbox": [ + 76, + 671, + 468, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To explore this problem, we introduce a specialized dataset that bridges real-world architectural imagery and procedurally-generated images, with a particular focus on window designs. Our dataset, comprising 89,318 windows in 75,739 high-resolution (4K to 6K) and RAW images, not only matches the scale of established datasets like CelebAHQ and FFHQ but also offers a unique niche with an emphasis on architectural details and high resolutions.", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our dataset is designed not just to advance synthetic data generation, but also to facilitate diverse research avenues. Our primary contributions1 are:", + "bbox": [ + 76, + 825, + 468, + 871 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Real-World Imagery: A 4K resolution dataset with 75,739 photos of windows from global locations, offering unprecedented detail for architectural research.", + "2. Hand-Annotated Labels: Segmentation labels for 9,002 images.", + "3. A Procedural Model: A novel procedural model for windows, underscoring key design choices for synthetic data generation.", + "4. Synthetic Data and Labels: A diverse set of 21,290 synthetic window images, mirroring features observed in real imagery." + ], + "bbox": [ + 500, + 90, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we conduct extensive experiments and ablations to understand the impact of various features in our synthetic dataset on segmentation performance.", + "bbox": [ + 498, + 258, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 316, + 640, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several datasets of architectural imagery have been created, enabling applications such as architectural style classification [2, 9, 40], building functional use classification [21, 41], architectural heritage classification [27], landmark identification [31], or urban scene matching [18]. Notably, datasets that support image synthesis or image segmentation have been instrumental in advancing these fields. For instance, the FaSyn13 dataset [13] consists of 200 façade images for texture synthesis, but its size is limited for modern generative models. Similarly, the LSAA dataset [42] includes 199,723 façade images and 516,000 cropped window images. However, the resolution of these cropped windows varies, with the majority being less than 100 pixels in the longest dimension.", + "bbox": [ + 496, + 340, + 893, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In comparison, extensive collections like MS COCO [26] and LAION [35] offer a broader range but with challenges in resolution and scene variety. For instance, the sheer diversity in LAION's 9.8 million 1K-resolution images presents significant challenges for a single procedural model, highlighting the need for focused datasets. Our dataset addresses this by providing high-resolution window images, offering a specialized resource for detailed architectural analysis and procedural modeling.", + "bbox": [ + 496, + 553, + 893, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our windows-focused dataset, while containing fewer windows than LSAA, stands out as the highest-resolution dataset of window images known to us, with an average resolution of 4,000 pixels per side for cropped windows. This high-resolution focus, particularly in the 4K to 6K range, fills a unique niche in architectural and synthetic-to-real domain transfer research. Our dataset's specialization in high-resolution architectural elements, especially windows, offers a valuable resource for advancing procedural modeling, emphasizing the importance of detail and precision.", + "bbox": [ + 496, + 689, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several architectural image datasets supporting semantic segmentation or façade parsing exist, though not specifically focused on windows. The Graz dataset [33] contains 50 rectified images, and the eTRIMS dataset [25] contains 60", + "bbox": [ + 498, + 840, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Available online https://twak.github.io/winsyn", + "bbox": [ + 94, + 886, + 346, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22457", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "non-rectified images. The CMP-Facade dataset [36] contains 606 images, with about half of them fairly high resolution (1,024 pixels on the long edge) but a limited diversity of image locations. The LabelMe-Facade dataset [5] has the largest number of images at 945, with each image varying in size between 512 and 768 pixels on a side. However, these datasets do not have the number of images nor the resolution that are desirable for training the latest computer vision methods, which increasingly require more detailed and high-resolution data. With 9,002 labeled images at four times the resolution of these datasets, our proposed dataset of real-world images is an order of magnitude larger. By concentrating exclusively on windows, our dataset offers unprecedented detail and specificity, enabling more precise and effective models for architectural element analysis.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Various authors have attempted to use synthetically generated data to bootstrap performance on real images. This approach seems to work best in domains where the human annotation is not directly feasible, such as reinforcement learning, especially for driving applications [14], depth or optical flow [8, 16], or 6DoF pose estimation for robotic grasping or manipulation [19, 24, 37].", + "bbox": [ + 75, + 321, + 470, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Infinigen [32] is a good example of a procedural model, however, there is no validation of the effectiveness of the model for machine learning tasks. Of particular note is the SynthIA dataset [34], a driving dataset built on video game technology specifically designed to support semantic segmentation in urban environments. A very large engineering effort went into this, as well as CARLA [14], and we believe that reproducing such high-quality synthetic data is out of reach for most academic teams. Similar to our dataset, SynthIA aims at pushing the envelope to use synthetic data to improve computer vision even for problems where large human-annotated datasets (KITTI [28], LabelME-Facade [6], Camvid [4]) already exist. Similar to our findings, they can get some results from purely synthetic data, but they cannot out-compete even relatively small real-world labeled images, but by combining synthetic and real at a 4:6 ratio they obtain their best results. However, unlike SynthIA, our segmentation challenge is more constrained (only windows) and we think would require fewer resources for academic researchers to develop competing procedural models for synthetic training. Although buildings are visible in these datasets, they are not focused on architecture. Domain transfer for architecture is challenging due to the amount of variety and complex dependencies between architectural elements, whereas the categories of objects relevant in driving scenes are much more clearly defined. In addition, the high resolution of the images we use makes the synthesis of realistic textures and precise object boundaries critical, and the higher-capacity segmentation models of today vs. 2016 (when SynthIA was published) are more precise but may also be more likely to overfit synthetic data.", + "bbox": [ + 75, + 431, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/80bb774f4a8fbf03f0c83b44b953244b6c6fff18a51acd504334b34f48ff8524.jpg", + "image_caption": [ + "Figure 2. Samples from the 75,739 photographs in the dataset. Each column shows a variety of examples of windows from different geographic locations. From left to right: Chicago (USA), Cambridge (UK), Bangkok (Thailand), Cairo (Egypt), and Vienna (Austria). The dataset has a variety of window shapes and architectural styles." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset of real and synthetic imagery is unique as a high-resolution, voluminous dataset and serves as a proving ground for synthetic to real training, image generation, and semantic segmentation tasks.", + "bbox": [ + 496, + 431, + 893, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To our knowledge, the most successful work on using synthetic data for segmentation is the 'Fake It Till You Make It' paper by Wood et al. [39], which reports improvements in segmentation when a U-Net is trained using their synthetic data vs. real image. However, to get these improved results, they used label adaptation, a technique that requires real labeled data. This is counter to our goals for using synthetic data for domains where labels are scarce. They do not report segmentation results without label adaptation, but they do show in their ablation study on landmark localization that label adaptation is critical for benefiting from synthetic data.", + "bbox": [ + 496, + 492, + 895, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Real-Windows Dataset", + "text_level": 1, + "bbox": [ + 500, + 688, + 718, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In line with our goal of advancing procedural modeling, our dataset is curated to offer high-resolution, diverse imagery essential for developing and evaluating procedural models. Our dataset, comprising over 75,739 high-resolution images (up to $4K \\times 6K$ ), is one of the largest and most detailed collections of window imagery available. Unlike other datasets sourced from the web or Flickr [22, 23], we hold complete copyright ownership of every image, ensuring legal clarity and flexibility for research use2.", + "bbox": [ + 496, + 713, + 893, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The dataset's diversity in location, viewpoint, and archi", + "bbox": [ + 517, + 849, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2While the datasets we referenced use images under permissive licenses, owning our images outright simplifies usage rights.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22458", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09b04027d54e0c02b818249e2f20a9bafdc82104227be36a27b5101f90b03ee0.jpg", + "image_caption": [ + "Figure 3. Examples of the labels used to annotate our data. Each instance receives its own polygon. The reader may wish to zoom into the figure for details." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tectural style, was achieved through a global effort. We engaged photographers from various countries, primarily hired via Upwork, to capture a wide array of window designs. This global collection effort ensures our dataset represents a broad spectrum of cultural and architectural diversity. This diversity is crucial for procedural models to learn and adapt to a wide range of architectural styles, directly supporting our goal of creating versatile and realistic models. Each image was chosen to highlight the window's design and architectural context, ensuring clear, distraction-free, and well-framed shots. The emphasis was on professional-grade photography with a focus on 4K resolution, achieving at least 2K pixels across each window, and taken during the day to ensure balanced exposure and minimal noise.", + "bbox": [ + 75, + 257, + 472, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In curating this collection, we also prioritized ethical considerations, instructing photographers to avoid capturing private situations or sensitive locations. Over 12 months, this project involved hiring 30 photographers and incurring costs of US $0.20-$ 0.50 per image, in addition to our quality control and subcontract management expenses.", + "bbox": [ + 75, + 486, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diversity of image locations is indicated in Table 1 of the Appendix, along with the number of images that include semantic segmentation labels and RAW camera data. The RAW camera data's higher bit-depth could be particularly valuable for future procedural modeling research, offering richer information for model training and evaluation.", + "bbox": [ + 75, + 578, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Many photos included multiple windows. We manually annotated crops in each image as a region that includes a single window in the center, along with any portion of the wall that may have been adapted to the window (such as brickwork or molding) and a portion of the wall on all four sides of the window. Due to this cropping, window images are in a variety of sizes, as shown in Appendix Fig. 2. We store the original images and the cropping information separately and generate a cropped version of the dataset on demand.", + "bbox": [ + 75, + 671, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each image has been carefully cropped and then annotated to focus on a single window and its immediate architectural context, as illustrated in Fig. 3. The annotations are designed to test if procedural models accurately replicate real-world architectural elements. The cropped win", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/05c6c9d8660d24108b51d028beaa4f3bf3ad5138b22d33b6a1dff67bce6911dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LabelImages UsingArea %
wall890743.02%
window pane836222.58%
wall frame869714.91%
window frame86819.71%
unlabeled29943.09%
shutter9312.56%
balcony9731.08%
misc object23571.07%
blind3750.75%
bars6790.68%
open-window9770.55%
", + "bbox": [ + 547, + 255, + 846, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The labels used, their frequency of use, and percentage by area for the square-cropped dataset used for our experiments. We note that the dataset is a mix of well-used labels such as wall and less-used ones, such as blind or bar. The 'unlabeled' category contained areas beyond the building (e.g., sky, streets) and a much smaller number of ambiguous areas where we could not reach a decision on how to label a feature.", + "bbox": [ + 496, + 450, + 893, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dow images also underwent a manual annotation process for panoptic segmentation, and subsequent review for quality assurance, leading to the acquisition of 9,002 annotations at an average cost of US $3.90 per image. While manageable, this cost notably exceeds image acquisition expenses, underscoring the value of methods that reduce reliance on labor-intensive labeled data for segmentation tasks. A detailed account of the annotation process is presented in Section 5.", + "bbox": [ + 496, + 574, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The WinSyn dataset serves as a foundational step towards our larger goal of procedural urban modeling, offering a comprehensive and detailed resource for both current and future research in this domain.", + "bbox": [ + 496, + 710, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Procedural Model Development", + "text_level": 1, + "bbox": [ + 498, + 784, + 790, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In developing our procedural model, we aim to create a realistic synthetic dataset with the ability to produce variations (ablations and experiments) for evaluation against real-world data. The compact domain (windows are relatively simple to model) allows well-developed approaches to be used to create the scene geometry including Split", + "bbox": [ + 496, + 810, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22459", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shape Grammars [38] and the CGA language [29], combined with Bezier splines for curved window shapes. This approach ensures that our model can generate a wide range of architectural styles and window designs closely resembling real-world distributions. The canonical orientation offered by windows as a domain allows a layered approach to geometry generation, as Appendix Fig.11: from the camera, we generate street clutter (e.g., cars, bollards, shrubs), wall decorations (shutters, balconies, pipes) exterior walls, window geometry, window dressing (blinds, curtains), and interior geometry.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The procedural modeling pipeline, developed in Python, Blender [11], and rendered with the physically based Cycles [3] renderer is used to generate a diverse dataset of 21,290 synthetic window images with corresponding labels (as Appendix Fig. 4). Normal, depth (Fig. 1), and edge maps are optionally generated. The model was designed to prioritize diversity, making use of domain randomization to extend beyond typical real-world variations.", + "bbox": [ + 75, + 258, + 472, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The procedural model's core uses two varieties of Split Grammar. The first, utilizing the CGA language, was employed to subdivide the volume to create the building mass, facades, shutters, blinds, balconies, roof, and window-bounds. From these window-bounds, a second Split Shape Grammar creates the window shapes themselves, using Bezier spline curves to capture various geometries, such as trapezoid, arched, or circular windows. This grammar subdivides closed curves to create nested window frames and offsets them to create individual glass panes. The frame geometry is created by extruding a profile along each Bezier. The profiles are selected from a randomly selected hierarchy of profiles. Parts of the window geometry hierarchy can be translated and rotated to 'open' windows by sliding or hinged mechanisms. This multi-grammar approach allows for detailed and varied window designs, while the generation remains algorithmically robust at scale.", + "bbox": [ + 75, + 382, + 472, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The procedural model is highly parameterized, with the number of parameters ranging from 216 to 21,735 in our dataset, depending on the chosen sequence of randomly-sampled rules in the grammars. We observe that optionally reusing parameters between parts of the model can improve visual realism; for example, we see window frames sharing their material (paint or stucco) with walls, or adjacent windows having similar (but not identical) materials. In the analysis section below we vary the distributions of these parameters to ablate and experiment on our procedural model and identify the most performant factors, including materials, textures, geometry, camera position, and lighting.", + "bbox": [ + 75, + 642, + 472, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textures are primarily procedural shaders [7], controlling materials such as wood, brick, or glass applied to appropriate object classes. To add realism, we captured exterior clutter, such as building signage, vehicles, and trash cans, through a number of sources, including pre-existing", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets and LiDAR/RGB scanning (Appendix Fig. 12) with bitmap textures.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene lighting is supplied by a combination of skybox emission, sun-lamp, and optional interior sources. We use interior and exterior panoramic images for the skybox and interior-box to create realistic background environments for our geometry. The camera is strategically positioned in order to capture the entire window from a predominantly frontal view; we use a variety of camera distances and adjust the field of view to frame the target window in the façade. Additional details of the procedural model can be found in Appendix Section 6.", + "bbox": [ + 496, + 121, + 893, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Estimating Procedural Model Quality", + "text_level": 1, + "bbox": [ + 498, + 287, + 841, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use semantic segmentation as a benchmark to evaluate the quality of a procedural model. This method involves training segmentation models on labeled images generated from the procedural model, evaluating model performances on a holdout (test) set of labeled real-world images, and providing a robust measure of the synthetic data's ability to accurately reflect real-world scenarios.", + "bbox": [ + 496, + 314, + 893, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For this approach to be successful, we must align our semantic labels with both real-world images and images generated by a procedural model. We established ten broad categories such as 'window-pane', 'window-frame', and 'wall frame' (see Appendix Section 7.7). This categorization mirrors the elements that our procedural model can realistically generate. Furthermore, we label each unique instance within the images, as demonstrated in Fig. 3, making the data suitable for panoptic segmentation even though we only evaluate the semantic segmentation task. This detailed set of labels extends traditional segmentation datasets, which typically provide a single 'window' label, with occasional additions like 'shutter' or 'blind' [6, 25, 36]. Our labeling allows consistent annotation across varied architectural styles but also ensures compatibility with the capabilities of procedural modeling.", + "bbox": [ + 496, + 421, + 893, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When utilizing mIoU to evaluate procedural model performance, inspecting plots such as Figures 6 and 7 can be insightful. These visual analyses aid in identifying parameters that optimize the model's performance. Additionally, Spearman's rank correlation is applied to quantify the impact of changes in model parameters (such as textures and lighting) on the synthetic image quality.", + "bbox": [ + 496, + 662, + 893, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Analysis", + "text_level": 1, + "bbox": [ + 500, + 784, + 596, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all following experiments, we fine-tune a BEiT 'base' model that was pre-trained on ImageNet-1k, trained on ImageNet-21k, and fine-tuned on data from some variation of our procedural model. We evaluated the mIoU over 10 labels, excluding the 'unlabeled' category. All cropped window-images are resized to 512 pixels square; examples", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22460", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f7579ab36ee87fdb89b128d647000e3751fea2649d785df9567365a60d725e49.jpg", + "image_caption": [ + "Figure 4. Left: histogram of per-image mIoUs showing the distribution of labeling results for a model trained on $n = 2,048$ synthetic (s) and real (r) images. We also show the difference between the model mIoU's per image (r-s). The mIoU was evaluated on 4,906 real images. Right: random samples of the labeling quality for networks trained on real and synthetic data; the first sample with an mIoU above each decile was selected." + ], + "image_footnote": [], + "bbox": [ + 68, + 85, + 450, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67d771064bf87d8d978e4b955b0013f9c40dd95992d5e290c96afffdeaa2cf2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 88, + 890, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are shown in Fig. 4.", + "text_level": 1, + "bbox": [ + 75, + 325, + 210, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of our procedural model is best understood in the context of regional performance disparities within real-world data. We contextualize the procedural model by comparing it to subsets from various regions of real-world data in Table 2 shows. Each set consists of 1,024 training images and 300 test images. Models trained on synthetic data have a narrower performance range (29.33 to $35.02\\mathrm{mIoU}$ ) compared to those trained on locale-specific data (25.23 to $61.85\\mathrm{mIoU}$ ). Notably, synthetic data can surpass real data from a different locale, as seen in the comparison of England and Egypt results. This narrower range seems to indicate synthetic data has not over-fit any particular region. The overall mIoU gap on the global dataset underscores the procedural model's limitations. The combined training set of real-world images evaluated to 53.79 mIoU, which is an upper limit on what one should expect from any model. The synthetic data yields $31.23\\mathrm{mIoU}$ on the global test set, whereas other locales varied from 37.76-51.22 mIoU for the 'other' set. Given the global set includes holdout images from each locale, it is not surprising the mIoU's are slightly higher. This table indicates that the procedural model is of comparable quality to choosing a single locale.", + "bbox": [ + 73, + 340, + 473, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Procedural Model Variations", + "text_level": 1, + "bbox": [ + 75, + 696, + 336, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the impact of synthetic dataset variations on a segmentation model by comparing against our baseline model. For each variation we render 2,048 training examples and use the same geometry, lighting, and rendering settings as the baseline unless specified. For each variation, we render 2,048 training examples. Empirically, we find that the performance of synthetic data levels off at this number of examples (see Appendix Fig. 7) and evaluation is faster if the set is kept smaller. Evaluation takes place on a test partition of 4,906 real-world images; the mIoU of the baseline model is 32.58 and real photos have an mIoU of 58.69. To gauge each variation's importance, we either the relative", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb57500f476c5531018db3d308044d9afb66a5acc458761f2dfeb2d618b62986.jpg", + "image_caption": [ + "Figure 5. The effect of varying real-world dataset sizes on mIoU with (blue) and without (red) an additional 2,048 synthetic samples. The green area bounds the range in which adding real data neither harms nor improves performance; the right-most point of which has $n = 152$ with an mIoU of 44.96. At larger datasets, synthetic data slightly reduces mIoU relative to only using real-world data." + ], + "image_footnote": [], + "bbox": [ + 496, + 327, + 859, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mIoU range as a percentage of the baseline, or report the mIoU Spearman's rank correlation $(r_s)$ . Detailed results are in Appendix Section 7 and summarized here.", + "bbox": [ + 496, + 719, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendering samples. We evaluated the impact of samples per pixel (spp) on render quality, noting diminishing returns beyond 256spp (Fig. 6). Render times scale from 6.4s at 1spp to 85.2s at 512spp. A strong correlation $(r_s = 1, n = 10)$ exists between spp and mIoU, with a $68\\%$ change in mIoU scores relative to baseline, underlining the importance of spp. In this experiment no denoising was performed, however, our baseline model used 256spp and a powerful neural denoiser.", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22461", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/46ec475b7c951130923bdc9bc27799ddb0a341838239a214df56ae05bf1a2130.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
test
globalAustriaEgyptUKUSAothersynthetic
trainglobal53.7958.2056.8542.6450.1654.7431.21
Austria39.4958.2328.3634.4035.6741.3221.87
Egypt47.5138.3261.8533.0238.0149.7531.34
UK37.7640.6825.2338.5635.3937.6428.51
USA41.0842.5027.3933.3050.0839.3927.02
other51.2248.5439.4737.4441.3352.7428.39
synthetic31.2329.5329.3332.1534.1735.0262.12
", + "bbox": [ + 202, + 88, + 767, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. mIoU for different splits of the real labeled data on the segmentation task. Trained on 1,024, tested on 300 samples. global is a mixture of all the real data; other data is from locales outside of Austria, Egypt, UK, or USA. In this experiment our synthetic network is similarly trained on 1,024 samples from our baseline synthetic model.", + "bbox": [ + 75, + 238, + 893, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/020084d66d4901c3debd19c4e74dc6de13d5454796ae8d052c825e84d351bef3.jpg", + "image_caption": [ + "Figure 6. Top: The impact of rendering samples per pixel (spp) on segmentation task accuracy. Bottom: example renderings at different spp with zoomed section of rendering. $\\mathrm{BL} =$ baseline." + ], + "image_footnote": [], + "bbox": [ + 89, + 311, + 472, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixed Materials. We experimented with 9 material variations, including uniform gray, edge shaders, and random textures. Some tests used a single material for the entire scene, others assigned different materials per object. Simplifying materials, such as rendering only the albedo channel, resulted in a significant performance drop. Appendix Fig. 18 shows that any material restriction led to at least an $18\\%$ mIoU decrease from the baseline.", + "bbox": [ + 75, + 686, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lighting. We examined 8 lighting model variations and their mIoU impact, as detailed in Appendix Fig. 19. Lighting models like albedo-only were crucial, while lighting conditions such as night-time had moderate importance. Conditions varied mIoU by $15.35\\%$ from baseline. Daylight-only training decreased baseline mIoU by $1.04\\%$", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "relative to baseline.", + "bbox": [ + 500, + 306, + 630, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera. We experimented with the distribution of camera positions. These 7 variations used a simple model which sampled a camera position over a circle, of radius $r = \\{0\\ldots 96\\}$ meters, truncated at the floor plane. The circle is positioned 5 meters from the wall, directly in front of the window. The camera's field of view is adjusted to the apparent window size. We observed limited impact on mIoU as $r$ changes; peak task performance was at $r = 12m$ .", + "bbox": [ + 496, + 321, + 890, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Window Geometry. We ran 7 tests with varying window dimensions and shapes, including square and nonrectangular windows. The mIoU impact was minor, fluctuating by up to $5.7\\%$ relative to the baseline (1.86 absolute), with the best variation $1.2\\%$ worse relative to the baseline $(-0.04$ absolute). Small features, though noticeable to humans, had a weak impact on model performance.", + "bbox": [ + 496, + 443, + 890, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Labels Modeled. In developing our procedural synthetic data generator, we prioritized labels by size, starting with wall and ending with open-window (Appendix Section 4). This enabled the assessment of mIoU at nine developmental stages (Fig. 7). Adding smaller classes later showed diminishing returns and occasionally reduced single-label accuracy (Appendix Fig. 22).", + "bbox": [ + 496, + 549, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Additional Experiments", + "text_level": 1, + "bbox": [ + 500, + 665, + 720, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our segmentation experiments, we employed the BEiTv2 base model [1, 30] for its proficiency in generating accurate results with minimal data. However, the trends we observed are consistent across various models, including DeepLabv3+ [10]. In the Appendix Section 5, we discuss two experiments that study the impact on the segmentation task when training (and testing) with different mixes of real and synthetic data. The first experiment in Appendix Fig. 7 explores segmentation performance when training and testing on different amounts of real or synthetic data. We observe that synthetic data saturates (stops improving with additional data), while real-world data continues to benefit from additional samples. The second experiment, in Appendix Fig. 8, illustrates the impact of different mixtures of", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22462", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb1dbc150eac29e2e0fc62818df86a72bc6ad3b6d5ac80b3969070bd50c4f866.jpg", + "image_caption": [ + "Figure 7. Top: The progressive impact of incorporating additional labels into the procedural model, with label subsets expanding from lvl1 to lvl9. The sequence begins with 'walls' only (lvl1), adding 'window panes' (lvl2), 'wall frames' (lvl3), 'window frames' (lvl4), 'shutters' (lvl5), 'balconies' (lvl6), 'miscellaneous objects' (lvl7), 'blinds' (lvl8), and 'bars' (lvl9). The full model includes additional features beyond lvl9, such as interior dressing, open windows, and windows without glass. Bottom: Examples corresponding to each incremental level of model complexity." + ], + "image_footnote": [], + "bbox": [ + 73, + 84, + 462, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "real and synthetic data, when tested on real data. Mixing synthetic and real data can be advantageous when there is little real data, but adding large amounts of synthetic data does not help task performance. We conclude that no volume of synthetic data can overcome this domain gap.", + "bbox": [ + 75, + 616, + 468, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An exploration of different architectures and data partitions is presented in Appendix Fig. 9 and Table 2 - we train an older convolution network without pre-training (DeepLab3+ [10]), a 'large' BEiTv2 model, an architecture which uses real data labels during training to improve performance (Label Adaptation [39]), another technique which uses unlabeled data to improve performance (MIC [20]), adjusting the colors of the training dataset (Histogram Matching [17]), and creating an 'easy' real data partition. Under these approaches, the synthetic-real domain gap persists, suggesting that network architecture and data partitions are an orthogonal research direction to improving the quality of synthetic procedural models.", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 89, + 687, + 108 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank our lead photographers Michaela Nömayr and Florian Rist, and engineer Prem Chedella, as well as our contributing photographers: Aleksandr Aleshkin, Angela Markoska, Artur Oliveira, Brian Benton, Chris West, Christopher Byrne, Elsayed Saafan, Florian Rist, George Ilin, Ignacio De Barrio, Jan Cuales, Kaitlyn Jackson, Kalina Mondzholovska, Kubra Ayse Guzel, Lukas Bornheim, Maria Jose Balestena, Michaela Nömayr, Mihai Alexandru Filoneanu, Mokhtari Sid Ahmed Salim, Mussa Ubapa, Nestor Angulo Caballero, Nicklaus Suarez, Peter Fountain, Prem Chedella, Samantha Martucci, Sarab-jot Singh, Scarlett Li, Serhii Malov, Simon R. A. Kelly, Stephanie Foden, Surafel Sunara, Tadiyos Shala, Susana Gomez, Vasileios Notis, Yuan Yuan, and finally labelyour-data.com for the labeling.", + "bbox": [ + 496, + 123, + 893, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 388, + 619, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a new dataset of 75,739 photos (2.09 terapixels), of which 9,002 photos are semantically labeled, as well as a high-quality procedural model that closely approximates real-world variation, making it effective for image segmentation tasks. Together these components of WinSyn have applications in unsupervised domain adaptation (using the large amount of unlabeled data), superresolution (via the thigh resolution images), learning from RAW sources, generative modeling, 3D reconstruction or depth recovery (using the synthetic depth or 3D geometry from the procedural model), as well as learning from images containing transparent or specular materials (i.e., window panes/glass balconies are often transparent).", + "bbox": [ + 496, + 422, + 893, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We systematically explored the effect of variations on our model on mIoU over 64 variations and 156,903 renders. The mIoU performance gap between our synthetic and real-world data is comparable to inter-country differences with largely different architectural styles. However, the difference between synthetic data and real data is still much larger than desired. This gap is not sufficiently reduced by either our work or other network architectures or dataset scale. We, therefore, believe that WinSyn provides a timely and efficient testbench from which researchers can iterate quickly to explore graphics contributions to the synthetic data problem. Our work contributes a sizable and versatile dataset that can be the basis of exciting and much-needed progress in the area of procedural graphics for synthetic data generation in machine learning. The key value of our dataset is that as a new benchmark of simulated and real images, it enables others to study the problem at an approachable level of complexity.", + "bbox": [ + 496, + 628, + 895, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22463", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 7", + "[2] Björn Barz and Joachim Denzler. Wikichurches: A fine-grained dataset of architectural styles with real-world challenges. arXiv preprint arXiv:2108.06959, 2021. 2", + "[3] Cycles: Open Source Production Rendering. Blender Foundation, 2023. 5", + "[4] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV (1), pages 44-57, 2008. 3", + "[5] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3", + "[6] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3, 5", + "[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, pages 1-7. vol. 2012, 2012. 5", + "[8] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 3", + "[9] J Chen, R Stouffs, and F Biljecki. Hierarchical (multi-label) architectural image recognition and classification. 2021. 2", + "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7, 8", + "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2023. 5", + "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[13] D. Dai, H. Riemenschneider, G. Schmitt, and L. Van Gool. Example-based facade texture synthesis. In International Conference on Computer Vision (ICCV), 2013. 2", + "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3", + "[15] Epic Games, Inc. MetaHuman Realistic Person Creator, 2021. 2", + "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking anal" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3", + "[17] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 8", + "[18] Daniel Cabrini Hauagge and Noah Snavely. Image matching using local symmetry features. In 2012 IEEE conference on computer vision and pattern recognition, pages 206-213. IEEE, 2012. 2", + "[19] Tomás Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jérí Matas. BOP challenge 2020 on 6D object localization. European Conference on Computer Vision Workshops (ECCVW), 2020. 3", + "[20] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023. 8", + "[21] Jian Kang, Marco Korner, Yuanyuan Wang, Hannes Taubenböck, and Xiao Xiang Zhu. Building instance classification using street view images. ISPRS journal of photogrammetry and remote sensing, 145:44-59, 2018. 2", + "[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3", + "[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. International Conference on Computer Vision (ICCV) Workshops, 2019. 3", + "[25] F. Korč and W. Förstner. eTRIMS Image Database for interpreting images of man-made scenes. Technical Report TR-IGG-P-2009-01, 2009. 2, 5", + "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 2", + "[27] Jose Llamas, Pedro M. Lerones, Roberto Medina, Eduardo Zalama, and Jaime Gómez-García-Bermejo. Classification of architectural heritage images using deep learning techniques. Applied Sciences, 7(10):992, 2017. 2", + "[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3", + "[29] Pascal Mueller, Peter Wonka, Simon Haegler, Andreas Ulmer, and Luc Van Gool. Procedural modeling of buildings. ACM Trans. Gr., 25(3):614-623, 2006. 5", + "[30] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers, 2022. 7", + "[31] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22464", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and fast spatial matching. In IEEE Conference on Computer Vision and Pattern Recognition, 2007. 2", + "[32] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, et al. Infinite photorealistic worlds using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12630-12641, 2023. 3", + "[33] Hayko Riemenschneider, Ulrich Krispel, Wolfgang Thaller, Michael Donoser, Sven Havemann, Dieter Fellner, and Horst Bischof. Irregular lattices for complex shape grammar facade parsing. In 2012 IEEE conference on computer vision and pattern recognition, pages 1640-1647. IEEE, 2012. 2", + "[34] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 3", + "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[36] Radim Tyleček and Radim Šára. Spatial pattern templates for recognition of objects with regular structure. In Proc. GCPR, Saarbrucken, Germany, 2013. 3, 5", + "[37] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-dof pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In International Conference on Intelligent Robots and Systems (IROS), 2022. 3", + "[38] Peter Wonka, Michael Wimmer, François Sillion, and William Ribarsky. Instant architecture. ACM Trans. Gr., 22 (3):669-677, 2003. 5", + "[39] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3661-3671. IEEE, 2021. 2, 3, 8", + "[40] Zhe Xu, Dacheng Tao, Ya Zhang, Junjie Wu, and Ah Chung Tsoi. Architectural style classification using multinomial latent logistic regression. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 600-615. Springer, 2014. 2", + "[41] K. Zhao, Y. Liu, S. Hao, S. Lu, H. Liu, and L. Zhou. Bounding boxes are all we need: Street view image classification via context encoding of detected buildings. IEEE Transactions on Geoscience and Remote Sensing, pages 1-17, 2021. 2", + "[42] Peihao Zhu, Wamiq Reyaz Para, Anna Frühstück, John Femiani, and Peter Wonka. Large-scale architectural asset extraction from panoramic imagery. IEEE Transactions on Visualization and Computer Graphics, 28(2):1301-1316, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22465", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_model.json b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d416e40a79cedcb231c58a4dfbc9d3e86d0a748 --- /dev/null +++ b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_model.json @@ -0,0 +1,1793 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.766, + 0.154 + ], + "angle": 0, + "content": "WinSyn: A High Resolution Testbed for Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.18, + 0.666, + 0.201 + ], + "angle": 0, + "content": "Tom Kelly1 John Femiani2 Peter Wonka1" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.204, + 0.621, + 0.221 + ], + "angle": 0, + "content": "\\(^{1}\\)KAUST, KSA \\(^{2}\\)Miami University, Ohio" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.256, + 0.314, + 0.273 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.288, + 0.474, + 0.635 + ], + "angle": 0, + "content": "We present WinSyn, a unique dataset and testbed for creating high-quality synthetic data with procedural modeling techniques. The dataset contains high-resolution photographs of windows, selected from locations around the world, with 89,318 individual window crops showcasing diverse geometric and material characteristics. We evaluate a procedural model by training semantic segmentation networks on both synthetic and real images and then comparing their performances on a shared test set of real images. Specifically, we measure the difference in mean Intersection over Union (mIoU) and determine the effective number of real images to match synthetic data's training performance. We design a baseline procedural model as a benchmark and provide 21,290 synthetically generated images. By tuning the procedural model, key factors are identified which significantly influence the model's fidelity in replicating real-world scenarios. Importantly, we highlight the challenge of procedural modeling using current techniques, especially in their ability to replicate the spatial semantics of real-world scenarios. This insight is critical because of the potential of procedural models to bridge to hidden scene aspects such as depth, reflectivity, material properties, and lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Larger and more sophisticated machine learning models demand an ever-increasing supply of data, particularly in tasks where manual annotation is challenging, such as depth estimation, reflectance estimation, or full 3D reconstruction. One solution is to use procedural models to generate synthetic training data. However, creating procedural models that accurately reflect the domain of real images (i.e. closing the domain gap) remains an open problem. Despite the visual realism of many synthetic scenes, their effectiveness in machine-learning applications often falls short. In this work we do not close the gap – but present an accessible pair of real and synthetic datasets, with annotations, in which this domain gap may be studied without massive resources." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "While the final goal is to tackle complex problems, a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.895, + 0.348 + ], + "angle": 0, + "content": "straightforward proxy task is initially required to ensure that real-world imagery can be manually annotated and compared to synthetic imagery. We, therefore, propose segmentation as a proxy task. This proxy task helps pinpoint where a procedural model fails to capture the diversity and nuances of real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.459 + ], + "angle": 0, + "content": "As our long-term goal is procedural urban modeling, we initially considered street-view images and datasets, such as CityScapes [12]. However, this has several drawbacks. Modeling arbitrary urban scenes realistically requires overwhelming complexity; requiring modeling at least cars, humans, buildings, skies, and vegetation, and each of these poses quite distinct and complex modeling challenges. For" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.486, + 0.855, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.655, + 0.855, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.827, + 0.895, + 0.898 + ], + "angle": 0, + "content": "Figure 1. Photographers in 28 geographic regions captured real-world photos (a) of windows that are cropped (b) to single windows, which are then labeled (c). Synthetic windows are rendered giving color (f) and labels(d), while other passes such as depth (e) are also possible." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "example, modeling urban environments for video games is a significant undertaking that often requires over 100 artists for open-world games. This type of effort is unrealistic for procedural modeling research. We propose that it is more promising to identify a simpler subset of images and develop a more constrained procedural model for this subset. This should enable faster iterations and broader participation in procedural modeling research, before scaling up to complete cities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.229, + 0.47, + 0.26 + ], + "angle": 0, + "content": "The decision to study the domain of window images was based on the following design principles for our dataset:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.474 + ], + "angle": 0, + "content": "Diversity: Our goal is to capture a range of variations in both geometry and materials. We opted against humans due to the complexities in modeling realistic fine geometric details (like hair and beard) and materials (such as skin), and the limited topological variations. Although humans have been used for synthetic data studies as demonstrated in [39], exploration in this domain is complicated because the procedural model was not released and high-fidelity models such as Unreal Metahumans by Epic Games [15] tend to prioritize realism over variability. Plants, another option, present challenges in material diversity and are difficult for human annotators to segment. Windows, in contrast, offers a balanced combination of geometric complexity and material variation suitable for our research goals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.473, + 0.655 + ], + "angle": 0, + "content": "High-resolution: The images in the dataset should be high resolution to make the dataset useful for the near and medium-term future. This is in contrast to existing architectural datasets, with resolutions typically in the 512-1025 pixel range [5, 6, 25, 33]. While these are not low resolution in typical computer vision terms, architectural features such as windows are often too small to resolve in the images. Although the CityScapes dataset is larger than most, with a resolution of \\(2048 \\times 1024\\) pixels, it often captures architecture from oblique viewpoints limiting the effective resolution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.672, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Image Rights: We would like to have all rights to the images to avoid future copyright problems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.825 + ], + "angle": 0, + "content": "To explore this problem, we introduce a specialized dataset that bridges real-world architectural imagery and procedurally-generated images, with a particular focus on window designs. Our dataset, comprising 89,318 windows in 75,739 high-resolution (4K to 6K) and RAW images, not only matches the scale of established datasets like CelebAHQ and FFHQ but also offers a unique niche with an emphasis on architectural details and high resolutions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.827, + 0.47, + 0.872 + ], + "angle": 0, + "content": "Our dataset is designed not just to advance synthetic data generation, but also to facilitate diverse research avenues. Our primary contributions1 are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "1. Real-World Imagery: A 4K resolution dataset with 75,739 photos of windows from global locations, offering unprecedented detail for architectural research." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.894, + 0.166 + ], + "angle": 0, + "content": "2. Hand-Annotated Labels: Segmentation labels for 9,002 images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.167, + 0.894, + 0.212 + ], + "angle": 0, + "content": "3. A Procedural Model: A novel procedural model for windows, underscoring key design choices for synthetic data generation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.212, + 0.894, + 0.258 + ], + "angle": 0, + "content": "4. Synthetic Data and Labels: A diverse set of 21,290 synthetic window images, mirroring features observed in real imagery." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.894, + 0.304 + ], + "angle": 0, + "content": "Additionally, we conduct extensive experiments and ablations to understand the impact of various features in our synthetic dataset on segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.642, + 0.332 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Several datasets of architectural imagery have been created, enabling applications such as architectural style classification [2, 9, 40], building functional use classification [21, 41], architectural heritage classification [27], landmark identification [31], or urban scene matching [18]. Notably, datasets that support image synthesis or image segmentation have been instrumental in advancing these fields. For instance, the FaSyn13 dataset [13] consists of 200 façade images for texture synthesis, but its size is limited for modern generative models. Similarly, the LSAA dataset [42] includes 199,723 façade images and 516,000 cropped window images. However, the resolution of these cropped windows varies, with the majority being less than 100 pixels in the longest dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.689 + ], + "angle": 0, + "content": "In comparison, extensive collections like MS COCO [26] and LAION [35] offer a broader range but with challenges in resolution and scene variety. For instance, the sheer diversity in LAION's 9.8 million 1K-resolution images presents significant challenges for a single procedural model, highlighting the need for focused datasets. Our dataset addresses this by providing high-resolution window images, offering a specialized resource for detailed architectural analysis and procedural modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Our windows-focused dataset, while containing fewer windows than LSAA, stands out as the highest-resolution dataset of window images known to us, with an average resolution of 4,000 pixels per side for cropped windows. This high-resolution focus, particularly in the 4K to 6K range, fills a unique niche in architectural and synthetic-to-real domain transfer research. Our dataset's specialization in high-resolution architectural elements, especially windows, offers a valuable resource for advancing procedural modeling, emphasizing the importance of detail and precision." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Several architectural image datasets supporting semantic segmentation or façade parsing exist, though not specifically focused on windows. The Graz dataset [33] contains 50 rectified images, and the eTRIMS dataset [25] contains 60" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "Available online https://twak.github.io/winsyn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22457" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "non-rectified images. The CMP-Facade dataset [36] contains 606 images, with about half of them fairly high resolution (1,024 pixels on the long edge) but a limited diversity of image locations. The LabelMe-Facade dataset [5] has the largest number of images at 945, with each image varying in size between 512 and 768 pixels on a side. However, these datasets do not have the number of images nor the resolution that are desirable for training the latest computer vision methods, which increasingly require more detailed and high-resolution data. With 9,002 labeled images at four times the resolution of these datasets, our proposed dataset of real-world images is an order of magnitude larger. By concentrating exclusively on windows, our dataset offers unprecedented detail and specificity, enabling more precise and effective models for architectural element analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.472, + 0.429 + ], + "angle": 0, + "content": "Various authors have attempted to use synthetically generated data to bootstrap performance on real images. This approach seems to work best in domains where the human annotation is not directly feasible, such as reinforcement learning, especially for driving applications [14], depth or optical flow [8, 16], or 6DoF pose estimation for robotic grasping or manipulation [19, 24, 37]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Infinigen [32] is a good example of a procedural model, however, there is no validation of the effectiveness of the model for machine learning tasks. Of particular note is the SynthIA dataset [34], a driving dataset built on video game technology specifically designed to support semantic segmentation in urban environments. A very large engineering effort went into this, as well as CARLA [14], and we believe that reproducing such high-quality synthetic data is out of reach for most academic teams. Similar to our dataset, SynthIA aims at pushing the envelope to use synthetic data to improve computer vision even for problems where large human-annotated datasets (KITTI [28], LabelME-Facade [6], Camvid [4]) already exist. Similar to our findings, they can get some results from purely synthetic data, but they cannot out-compete even relatively small real-world labeled images, but by combining synthetic and real at a 4:6 ratio they obtain their best results. However, unlike SynthIA, our segmentation challenge is more constrained (only windows) and we think would require fewer resources for academic researchers to develop competing procedural models for synthetic training. Although buildings are visible in these datasets, they are not focused on architecture. Domain transfer for architecture is challenging due to the amount of variety and complex dependencies between architectural elements, whereas the categories of objects relevant in driving scenes are much more clearly defined. In addition, the high resolution of the images we use makes the synthesis of realistic textures and precise object boundaries critical, and the higher-capacity segmentation models of today vs. 2016 (when SynthIA was published) are more precise but may also be more likely to overfit synthetic data." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Samples from the 75,739 photographs in the dataset. Each column shows a variety of examples of windows from different geographic locations. From left to right: Chicago (USA), Cambridge (UK), Bangkok (Thailand), Cairo (Egypt), and Vienna (Austria). The dataset has a variety of window shapes and architectural styles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Our dataset of real and synthetic imagery is unique as a high-resolution, voluminous dataset and serves as a proving ground for synthetic to real training, image generation, and semantic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.896, + 0.676 + ], + "angle": 0, + "content": "To our knowledge, the most successful work on using synthetic data for segmentation is the 'Fake It Till You Make It' paper by Wood et al. [39], which reports improvements in segmentation when a U-Net is trained using their synthetic data vs. real image. However, to get these improved results, they used label adaptation, a technique that requires real labeled data. This is counter to our goals for using synthetic data for domains where labels are scarce. They do not report segmentation results without label adaptation, but they do show in their ablation study on landmark localization that label adaptation is critical for benefiting from synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.689, + 0.719, + 0.705 + ], + "angle": 0, + "content": "3. Real-Windows Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.895, + 0.85 + ], + "angle": 0, + "content": "In line with our goal of advancing procedural modeling, our dataset is curated to offer high-resolution, diverse imagery essential for developing and evaluating procedural models. Our dataset, comprising over 75,739 high-resolution images (up to \\(4K \\times 6K\\)), is one of the largest and most detailed collections of window imagery available. Unlike other datasets sourced from the web or Flickr [22, 23], we hold complete copyright ownership of every image, ensuring legal clarity and flexibility for research use2." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.85, + 0.892, + 0.866 + ], + "angle": 0, + "content": "The dataset's diversity in location, viewpoint, and archi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2While the datasets we referenced use images under permissive licenses, owning our images outright simplifies usage rights." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22458" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.894, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.203, + 0.893, + 0.233 + ], + "angle": 0, + "content": "Figure 3. Examples of the labels used to annotate our data. Each instance receives its own polygon. The reader may wish to zoom into the figure for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.484 + ], + "angle": 0, + "content": "tectural style, was achieved through a global effort. We engaged photographers from various countries, primarily hired via Upwork, to capture a wide array of window designs. This global collection effort ensures our dataset represents a broad spectrum of cultural and architectural diversity. This diversity is crucial for procedural models to learn and adapt to a wide range of architectural styles, directly supporting our goal of creating versatile and realistic models. Each image was chosen to highlight the window's design and architectural context, ensuring clear, distraction-free, and well-framed shots. The emphasis was on professional-grade photography with a focus on 4K resolution, achieving at least 2K pixels across each window, and taken during the day to ensure balanced exposure and minimal noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.577 + ], + "angle": 0, + "content": "In curating this collection, we also prioritized ethical considerations, instructing photographers to avoid capturing private situations or sensitive locations. Over 12 months, this project involved hiring 30 photographers and incurring costs of US \\(0.20-\\)0.50 per image, in addition to our quality control and subcontract management expenses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.671 + ], + "angle": 0, + "content": "The diversity of image locations is indicated in Table 1 of the Appendix, along with the number of images that include semantic segmentation labels and RAW camera data. The RAW camera data's higher bit-depth could be particularly valuable for future procedural modeling research, offering richer information for model training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Many photos included multiple windows. We manually annotated crops in each image as a region that includes a single window in the center, along with any portion of the wall that may have been adapted to the window (such as brickwork or molding) and a portion of the wall on all four sides of the window. Due to this cropping, window images are in a variety of sizes, as shown in Appendix Fig. 2. We store the original images and the cropping information separately and generate a cropped version of the dataset on demand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Each image has been carefully cropped and then annotated to focus on a single window and its immediate architectural context, as illustrated in Fig. 3. The annotations are designed to test if procedural models accurately replicate real-world architectural elements. The cropped win" + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.256, + 0.847, + 0.44 + ], + "angle": 0, + "content": "
LabelImages UsingArea %
wall890743.02%
window pane836222.58%
wall frame869714.91%
window frame86819.71%
unlabeled29943.09%
shutter9312.56%
balcony9731.08%
misc object23571.07%
blind3750.75%
bars6790.68%
open-window9770.55%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.549 + ], + "angle": 0, + "content": "Table 1. The labels used, their frequency of use, and percentage by area for the square-cropped dataset used for our experiments. We note that the dataset is a mix of well-used labels such as wall and less-used ones, such as blind or bar. The 'unlabeled' category contained areas beyond the building (e.g., sky, streets) and a much smaller number of ambiguous areas where we could not reach a decision on how to label a feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.709 + ], + "angle": 0, + "content": "dow images also underwent a manual annotation process for panoptic segmentation, and subsequent review for quality assurance, leading to the acquisition of 9,002 annotations at an average cost of US $3.90 per image. While manageable, this cost notably exceeds image acquisition expenses, underscoring the value of methods that reduce reliance on labor-intensive labeled data for segmentation tasks. A detailed account of the annotation process is presented in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.77 + ], + "angle": 0, + "content": "The WinSyn dataset serves as a foundational step towards our larger goal of procedural urban modeling, offering a comprehensive and detailed resource for both current and future research in this domain." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.785, + 0.791, + 0.803 + ], + "angle": 0, + "content": "4. Procedural Model Development" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.903 + ], + "angle": 0, + "content": "In developing our procedural model, we aim to create a realistic synthetic dataset with the ability to produce variations (ablations and experiments) for evaluation against real-world data. The compact domain (windows are relatively simple to model) allows well-developed approaches to be used to create the scene geometry including Split" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22459" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Shape Grammars [38] and the CGA language [29], combined with Bezier splines for curved window shapes. This approach ensures that our model can generate a wide range of architectural styles and window designs closely resembling real-world distributions. The canonical orientation offered by windows as a domain allows a layered approach to geometry generation, as Appendix Fig.11: from the camera, we generate street clutter (e.g., cars, bollards, shrubs), wall decorations (shutters, balconies, pipes) exterior walls, window geometry, window dressing (blinds, curtains), and interior geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.382 + ], + "angle": 0, + "content": "The procedural modeling pipeline, developed in Python, Blender [11], and rendered with the physically based Cycles [3] renderer is used to generate a diverse dataset of 21,290 synthetic window images with corresponding labels (as Appendix Fig. 4). Normal, depth (Fig. 1), and edge maps are optionally generated. The model was designed to prioritize diversity, making use of domain randomization to extend beyond typical real-world variations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.641 + ], + "angle": 0, + "content": "The procedural model's core uses two varieties of Split Grammar. The first, utilizing the CGA language, was employed to subdivide the volume to create the building mass, facades, shutters, blinds, balconies, roof, and window-bounds. From these window-bounds, a second Split Shape Grammar creates the window shapes themselves, using Bezier spline curves to capture various geometries, such as trapezoid, arched, or circular windows. This grammar subdivides closed curves to create nested window frames and offsets them to create individual glass panes. The frame geometry is created by extruding a profile along each Bezier. The profiles are selected from a randomly selected hierarchy of profiles. Parts of the window geometry hierarchy can be translated and rotated to 'open' windows by sliding or hinged mechanisms. This multi-grammar approach allows for detailed and varied window designs, while the generation remains algorithmically robust at scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.473, + 0.825 + ], + "angle": 0, + "content": "The procedural model is highly parameterized, with the number of parameters ranging from 216 to 21,735 in our dataset, depending on the chosen sequence of randomly-sampled rules in the grammars. We observe that optionally reusing parameters between parts of the model can improve visual realism; for example, we see window frames sharing their material (paint or stucco) with walls, or adjacent windows having similar (but not identical) materials. In the analysis section below we vary the distributions of these parameters to ablate and experiment on our procedural model and identify the most performant factors, including materials, textures, geometry, camera position, and lighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Textures are primarily procedural shaders [7], controlling materials such as wood, brick, or glass applied to appropriate object classes. To add realism, we captured exterior clutter, such as building signage, vehicles, and trash cans, through a number of sources, including pre-existing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "datasets and LiDAR/RGB scanning (Appendix Fig. 12) with bitmap textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.895, + 0.275 + ], + "angle": 0, + "content": "Scene lighting is supplied by a combination of skybox emission, sun-lamp, and optional interior sources. We use interior and exterior panoramic images for the skybox and interior-box to create realistic background environments for our geometry. The camera is strategically positioned in order to capture the entire window from a predominantly frontal view; we use a variety of camera distances and adjust the field of view to frame the target window in the façade. Additional details of the procedural model can be found in Appendix Section 6." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.289, + 0.842, + 0.308 + ], + "angle": 0, + "content": "5. Estimating Procedural Model Quality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.895, + 0.42 + ], + "angle": 0, + "content": "We use semantic segmentation as a benchmark to evaluate the quality of a procedural model. This method involves training segmentation models on labeled images generated from the procedural model, evaluating model performances on a holdout (test) set of labeled real-world images, and providing a robust measure of the synthetic data's ability to accurately reflect real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.663 + ], + "angle": 0, + "content": "For this approach to be successful, we must align our semantic labels with both real-world images and images generated by a procedural model. We established ten broad categories such as 'window-pane', 'window-frame', and 'wall frame' (see Appendix Section 7.7). This categorization mirrors the elements that our procedural model can realistically generate. Furthermore, we label each unique instance within the images, as demonstrated in Fig. 3, making the data suitable for panoptic segmentation even though we only evaluate the semantic segmentation task. This detailed set of labels extends traditional segmentation datasets, which typically provide a single 'window' label, with occasional additions like 'shutter' or 'blind' [6, 25, 36]. Our labeling allows consistent annotation across varied architectural styles but also ensures compatibility with the capabilities of procedural modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.895, + 0.772 + ], + "angle": 0, + "content": "When utilizing mIoU to evaluate procedural model performance, inspecting plots such as Figures 6 and 7 can be insightful. These visual analyses aid in identifying parameters that optimize the model's performance. Additionally, Spearman's rank correlation is applied to quantify the impact of changes in model parameters (such as textures and lighting) on the synthetic image quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.597, + 0.803 + ], + "angle": 0, + "content": "6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In all following experiments, we fine-tune a BEiT 'base' model that was pre-trained on ImageNet-1k, trained on ImageNet-21k, and fine-tuned on data from some variation of our procedural model. We evaluated the mIoU over 10 labels, excluding the 'unlabeled' category. All cropped window-images are resized to 512 pixels square; examples" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22460" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.07, + 0.086, + 0.452, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.089, + 0.892, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.244, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Left: histogram of per-image mIoUs showing the distribution of labeling results for a model trained on \\( n = 2,048 \\) synthetic (s) and real (r) images. We also show the difference between the model mIoU's per image (r-s). The mIoU was evaluated on 4,906 real images. Right: random samples of the labeling quality for networks trained on real and synthetic data; the first sample with an mIoU above each decile was selected." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.327, + 0.212, + 0.341 + ], + "angle": 0, + "content": "are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.342, + 0.474, + 0.689 + ], + "angle": 0, + "content": "The performance of our procedural model is best understood in the context of regional performance disparities within real-world data. We contextualize the procedural model by comparing it to subsets from various regions of real-world data in Table 2 shows. Each set consists of 1,024 training images and 300 test images. Models trained on synthetic data have a narrower performance range (29.33 to \\(35.02\\mathrm{mIoU}\\)) compared to those trained on locale-specific data (25.23 to \\(61.85\\mathrm{mIoU}\\)). Notably, synthetic data can surpass real data from a different locale, as seen in the comparison of England and Egypt results. This narrower range seems to indicate synthetic data has not over-fit any particular region. The overall mIoU gap on the global dataset underscores the procedural model's limitations. The combined training set of real-world images evaluated to 53.79 mIoU, which is an upper limit on what one should expect from any model. The synthetic data yields \\(31.23\\mathrm{mIoU}\\) on the global test set, whereas other locales varied from 37.76-51.22 mIoU for the 'other' set. Given the global set includes holdout images from each locale, it is not surprising the mIoU's are slightly higher. This table indicates that the procedural model is of comparable quality to choosing a single locale." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.697, + 0.338, + 0.712 + ], + "angle": 0, + "content": "6.1. Procedural Model Variations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We assess the impact of synthetic dataset variations on a segmentation model by comparing against our baseline model. For each variation we render 2,048 training examples and use the same geometry, lighting, and rendering settings as the baseline unless specified. For each variation, we render 2,048 training examples. Empirically, we find that the performance of synthetic data levels off at this number of examples (see Appendix Fig. 7) and evaluation is faster if the set is kept smaller. Evaluation takes place on a test partition of 4,906 real-world images; the mIoU of the baseline model is 32.58 and real photos have an mIoU of 58.69. To gauge each variation's importance, we either the relative" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.328, + 0.86, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Figure 5. The effect of varying real-world dataset sizes on mIoU with (blue) and without (red) an additional 2,048 synthetic samples. The green area bounds the range in which adding real data neither harms nor improves performance; the right-most point of which has \\( n = 152 \\) with an mIoU of 44.96. At larger datasets, synthetic data slightly reduces mIoU relative to only using real-world data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.765 + ], + "angle": 0, + "content": "mIoU range as a percentage of the baseline, or report the mIoU Spearman's rank correlation \\((r_s)\\). Detailed results are in Appendix Section 7 and summarized here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Rendering samples. We evaluated the impact of samples per pixel (spp) on render quality, noting diminishing returns beyond 256spp (Fig. 6). Render times scale from 6.4s at 1spp to 85.2s at 512spp. A strong correlation \\((r_s = 1, n = 10)\\) exists between spp and mIoU, with a \\(68\\%\\) change in mIoU scores relative to baseline, underlining the importance of spp. In this experiment no denoising was performed, however, our baseline model used 256spp and a powerful neural denoiser." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22461" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.204, + 0.089, + 0.768, + 0.228 + ], + "angle": 0, + "content": "
test
globalAustriaEgyptUKUSAothersynthetic
trainglobal53.7958.2056.8542.6450.1654.7431.21
Austria39.4958.2328.3634.4035.6741.3221.87
Egypt47.5138.3261.8533.0238.0149.7531.34
UK37.7640.6825.2338.5635.3937.6428.51
USA41.0842.5027.3933.3050.0839.3927.02
other51.2248.5439.4737.4441.3352.7428.39
synthetic31.2329.5329.3332.1534.1735.0262.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Table 2. mIoU for different splits of the real labeled data on the segmentation task. Trained on 1,024, tested on 300 samples. global is a mixture of all the real data; other data is from locales outside of Austria, Egypt, UK, or USA. In this experiment our synthetic network is similarly trained on 1,024 samples from our baseline synthetic model." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.313, + 0.473, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.656 + ], + "angle": 0, + "content": "Figure 6. Top: The impact of rendering samples per pixel (spp) on segmentation task accuracy. Bottom: example renderings at different spp with zoomed section of rendering. \\(\\mathrm{BL} =\\) baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Mixed Materials. We experimented with 9 material variations, including uniform gray, edge shaders, and random textures. Some tests used a single material for the entire scene, others assigned different materials per object. Simplifying materials, such as rendering only the albedo channel, resulted in a significant performance drop. Appendix Fig. 18 shows that any material restriction led to at least an \\(18\\%\\) mIoU decrease from the baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Lighting. We examined 8 lighting model variations and their mIoU impact, as detailed in Appendix Fig. 19. Lighting models like albedo-only were crucial, while lighting conditions such as night-time had moderate importance. Conditions varied mIoU by \\(15.35\\%\\) from baseline. Daylight-only training decreased baseline mIoU by \\(1.04\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.308, + 0.631, + 0.321 + ], + "angle": 0, + "content": "relative to baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.444 + ], + "angle": 0, + "content": "Camera. We experimented with the distribution of camera positions. These 7 variations used a simple model which sampled a camera position over a circle, of radius \\( r = \\{0\\ldots 96\\} \\) meters, truncated at the floor plane. The circle is positioned 5 meters from the wall, directly in front of the window. The camera's field of view is adjusted to the apparent window size. We observed limited impact on mIoU as \\( r \\) changes; peak task performance was at \\( r = 12m \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Window Geometry. We ran 7 tests with varying window dimensions and shapes, including square and nonrectangular windows. The mIoU impact was minor, fluctuating by up to \\(5.7\\%\\) relative to the baseline (1.86 absolute), with the best variation \\(1.2\\%\\) worse relative to the baseline \\((-0.04\\) absolute). Small features, though noticeable to humans, had a weak impact on model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Labels Modeled. In developing our procedural synthetic data generator, we prioritized labels by size, starting with wall and ending with open-window (Appendix Section 4). This enabled the assessment of mIoU at nine developmental stages (Fig. 7). Adding smaller classes later showed diminishing returns and occasionally reduced single-label accuracy (Appendix Fig. 22)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.722, + 0.683 + ], + "angle": 0, + "content": "6.2. Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In our segmentation experiments, we employed the BEiTv2 base model [1, 30] for its proficiency in generating accurate results with minimal data. However, the trends we observed are consistent across various models, including DeepLabv3+ [10]. In the Appendix Section 5, we discuss two experiments that study the impact on the segmentation task when training (and testing) with different mixes of real and synthetic data. The first experiment in Appendix Fig. 7 explores segmentation performance when training and testing on different amounts of real or synthetic data. We observe that synthetic data saturates (stops improving with additional data), while real-world data continues to benefit from additional samples. The second experiment, in Appendix Fig. 8, illustrates the impact of different mixtures of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22462" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.075, + 0.085, + 0.464, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.569 + ], + "angle": 0, + "content": "Figure 7. Top: The progressive impact of incorporating additional labels into the procedural model, with label subsets expanding from lvl1 to lvl9. The sequence begins with 'walls' only (lvl1), adding 'window panes' (lvl2), 'wall frames' (lvl3), 'window frames' (lvl4), 'shutters' (lvl5), 'balconies' (lvl6), 'miscellaneous objects' (lvl7), 'blinds' (lvl8), and 'bars' (lvl9). The full model includes additional features beyond lvl9, such as interior dressing, open windows, and windows without glass. Bottom: Examples corresponding to each incremental level of model complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.695 + ], + "angle": 0, + "content": "real and synthetic data, when tested on real data. Mixing synthetic and real data can be advantageous when there is little real data, but adding large amounts of synthetic data does not help task performance. We conclude that no volume of synthetic data can overcome this domain gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "An exploration of different architectures and data partitions is presented in Appendix Fig. 9 and Table 2 - we train an older convolution network without pre-training (DeepLab3+ [10]), a 'large' BEiTv2 model, an architecture which uses real data labels during training to improve performance (Label Adaptation [39]), another technique which uses unlabeled data to improve performance (MIC [20]), adjusting the colors of the training dataset (Histogram Matching [17]), and creating an 'easy' real data partition. Under these approaches, the synthetic-real domain gap persists, suggesting that network architecture and data partitions are an orthogonal research direction to improving the quality of synthetic procedural models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.689, + 0.109 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.895, + 0.351 + ], + "angle": 0, + "content": "We thank our lead photographers Michaela Nömayr and Florian Rist, and engineer Prem Chedella, as well as our contributing photographers: Aleksandr Aleshkin, Angela Markoska, Artur Oliveira, Brian Benton, Chris West, Christopher Byrne, Elsayed Saafan, Florian Rist, George Ilin, Ignacio De Barrio, Jan Cuales, Kaitlyn Jackson, Kalina Mondzholovska, Kubra Ayse Guzel, Lukas Bornheim, Maria Jose Balestena, Michaela Nömayr, Mihai Alexandru Filoneanu, Mokhtari Sid Ahmed Salim, Mussa Ubapa, Nestor Angulo Caballero, Nicklaus Suarez, Peter Fountain, Prem Chedella, Samantha Martucci, Sarab-jot Singh, Scarlett Li, Serhii Malov, Simon R. A. Kelly, Stephanie Foden, Surafel Sunara, Tadiyos Shala, Susana Gomez, Vasileios Notis, Yuan Yuan, and finally labelyour-data.com for the labeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.62, + 0.407 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.895, + 0.622 + ], + "angle": 0, + "content": "We have introduced a new dataset of 75,739 photos (2.09 terapixels), of which 9,002 photos are semantically labeled, as well as a high-quality procedural model that closely approximates real-world variation, making it effective for image segmentation tasks. Together these components of WinSyn have applications in unsupervised domain adaptation (using the large amount of unlabeled data), superresolution (via the thigh resolution images), learning from RAW sources, generative modeling, 3D reconstruction or depth recovery (using the synthetic depth or 3D geometry from the procedural model), as well as learning from images containing transparent or specular materials (i.e., window panes/glass balconies are often transparent)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.896, + 0.903 + ], + "angle": 0, + "content": "We systematically explored the effect of variations on our model on mIoU over 64 variations and 156,903 renders. The mIoU performance gap between our synthetic and real-world data is comparable to inter-country differences with largely different architectural styles. However, the difference between synthetic data and real data is still much larger than desired. This gap is not sufficiently reduced by either our work or other network architectures or dataset scale. We, therefore, believe that WinSyn provides a timely and efficient testbench from which researchers can iterate quickly to explore graphics contributions to the synthetic data problem. Our work contributes a sizable and versatile dataset that can be the basis of exciting and much-needed progress in the area of procedural graphics for synthetic data generation in machine learning. The key value of our dataset is that as a new benchmark of simulated and real images, it enables others to study the problem at an approachable level of complexity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22463" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Björn Barz and Joachim Denzler. Wikichurches: A fine-grained dataset of architectural styles with real-world challenges. arXiv preprint arXiv:2108.06959, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.226 + ], + "angle": 0, + "content": "[3] Cycles: Open Source Production Rendering. Blender Foundation, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[4] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV (1), pages 44-57, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[5] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[6] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, pages 1-7. vol. 2012, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[8] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.551 + ], + "angle": 0, + "content": "[9] J Chen, R Stouffs, and F Biljecki. Hierarchical (multi-label) architectural image recognition and classification. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[13] D. Dai, H. Riemenschneider, G. Schmitt, and L. Van Gool. Example-based facade texture synthesis. In International Conference on Computer Vision (ICCV), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[15] Epic Games, Inc. MetaHuman Realistic Person Creator, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking anal" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Daniel Cabrini Hauagge and Noah Snavely. Image matching using local symmetry features. In 2012 IEEE conference on computer vision and pattern recognition, pages 206-213. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Tomás Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jérí Matas. BOP challenge 2020 on 6D object localization. European Conference on Computer Vision Workshops (ECCVW), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[20] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[21] Jian Kang, Marco Korner, Yuanyuan Wang, Hannes Taubenböck, and Xiao Xiang Zhu. Building instance classification using street view images. ISPRS journal of photogrammetry and remote sensing, 145:44-59, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. International Conference on Computer Vision (ICCV) Workshops, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[25] F. Korč and W. Förstner. eTRIMS Image Database for interpreting images of man-made scenes. Technical Report TR-IGG-P-2009-01, 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[27] Jose Llamas, Pedro M. Lerones, Roberto Medina, Eduardo Zalama, and Jaime Gómez-García-Bermejo. Classification of architectural heritage images using deep learning techniques. Applied Sciences, 7(10):992, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.787, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[29] Pascal Mueller, Peter Wonka, Simon Haegler, Andreas Ulmer, and Luc Van Gool. Procedural modeling of buildings. ACM Trans. Gr., 25(3):614-623, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.829, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[30] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[31] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22464" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "and fast spatial matching. In IEEE Conference on Computer Vision and Pattern Recognition, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[32] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, et al. Infinite photorealistic worlds using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12630-12641, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[33] Hayko Riemenschneider, Ulrich Krispel, Wolfgang Thaller, Michael Donoser, Sven Havemann, Dieter Fellner, and Horst Bischof. Irregular lattices for complex shape grammar facade parsing. In 2012 IEEE conference on computer vision and pattern recognition, pages 1640-1647. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[34] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.47, + 0.43 + ], + "angle": 0, + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[36] Radim Tyleček and Radim Šára. Spatial pattern templates for recognition of objects with regular structure. In Proc. GCPR, Saarbrucken, Germany, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[37] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-dof pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In International Conference on Intelligent Robots and Systems (IROS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[38] Peter Wonka, Michael Wimmer, François Sillion, and William Ribarsky. Instant architecture. ACM Trans. Gr., 22 (3):669-677, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[39] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3661-3671. IEEE, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[40] Zhe Xu, Dacheng Tao, Ya Zhang, Junjie Wu, and Ah Chung Tsoi. Architectural style classification using multinomial latent logistic regression. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 600-615. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.756, + 0.469, + 0.823 + ], + "angle": 0, + "content": "[41] K. Zhao, Y. Liu, S. Hao, S. Lu, H. Liu, and L. Zhou. Bounding boxes are all we need: Street view image classification via context encoding of detected buildings. IEEE Transactions on Geoscience and Remote Sensing, pages 1-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.469, + 0.894 + ], + "angle": 0, + "content": "[42] Peihao Zhu, Wamiq Reyaz Para, Anna Frühstück, John Femiani, and Peter Wonka. Large-scale architectural asset extraction from panoramic imagery. IEEE Transactions on Visualization and Computer Graphics, 28(2):1301-1316, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22465" + } + ] +] \ No newline at end of file diff --git a/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_origin.pdf b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14ec2df209076b8bac90ecff6115f2e3f91321c1 --- /dev/null +++ b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/dce1c1ee-50a6-4a10-b211-db6a6195eecc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37eb30dec2423d2fcd0e1a3feff978c240458c9541cf5fb08b16d66fb0398896 +size 4033232 diff --git a/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/full.md b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f3e829a1f503f263e244559ce207c06da87d3583 --- /dev/null +++ b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/full.md @@ -0,0 +1,234 @@ +# WinSyn: A High Resolution Testbed for Synthetic Data + +Tom Kelly1 John Femiani2 Peter Wonka1 + +$^{1}$ KAUST, KSA $^{2}$ Miami University, Ohio + +# Abstract + +We present WinSyn, a unique dataset and testbed for creating high-quality synthetic data with procedural modeling techniques. The dataset contains high-resolution photographs of windows, selected from locations around the world, with 89,318 individual window crops showcasing diverse geometric and material characteristics. We evaluate a procedural model by training semantic segmentation networks on both synthetic and real images and then comparing their performances on a shared test set of real images. Specifically, we measure the difference in mean Intersection over Union (mIoU) and determine the effective number of real images to match synthetic data's training performance. We design a baseline procedural model as a benchmark and provide 21,290 synthetically generated images. By tuning the procedural model, key factors are identified which significantly influence the model's fidelity in replicating real-world scenarios. Importantly, we highlight the challenge of procedural modeling using current techniques, especially in their ability to replicate the spatial semantics of real-world scenarios. This insight is critical because of the potential of procedural models to bridge to hidden scene aspects such as depth, reflectivity, material properties, and lighting conditions. + +# 1. Introduction + +Larger and more sophisticated machine learning models demand an ever-increasing supply of data, particularly in tasks where manual annotation is challenging, such as depth estimation, reflectance estimation, or full 3D reconstruction. One solution is to use procedural models to generate synthetic training data. However, creating procedural models that accurately reflect the domain of real images (i.e. closing the domain gap) remains an open problem. Despite the visual realism of many synthetic scenes, their effectiveness in machine-learning applications often falls short. In this work we do not close the gap – but present an accessible pair of real and synthetic datasets, with annotations, in which this domain gap may be studied without massive resources. + +While the final goal is to tackle complex problems, a + +straightforward proxy task is initially required to ensure that real-world imagery can be manually annotated and compared to synthetic imagery. We, therefore, propose segmentation as a proxy task. This proxy task helps pinpoint where a procedural model fails to capture the diversity and nuances of real-world scenes. + +As our long-term goal is procedural urban modeling, we initially considered street-view images and datasets, such as CityScapes [12]. However, this has several drawbacks. Modeling arbitrary urban scenes realistically requires overwhelming complexity; requiring modeling at least cars, humans, buildings, skies, and vegetation, and each of these poses quite distinct and complex modeling challenges. For + +![](images/c7b8acacaf2fae6521302ba70178834a0722853ed224986707a746e8eed9e5e7.jpg) + +![](images/a682de236db7c90e79bec5ef5b56f69af41e62115dfada4b5b657902f43de5ca.jpg) +Figure 1. Photographers in 28 geographic regions captured real-world photos (a) of windows that are cropped (b) to single windows, which are then labeled (c). Synthetic windows are rendered giving color (f) and labels(d), while other passes such as depth (e) are also possible. + +example, modeling urban environments for video games is a significant undertaking that often requires over 100 artists for open-world games. This type of effort is unrealistic for procedural modeling research. We propose that it is more promising to identify a simpler subset of images and develop a more constrained procedural model for this subset. This should enable faster iterations and broader participation in procedural modeling research, before scaling up to complete cities. + +The decision to study the domain of window images was based on the following design principles for our dataset: + +Diversity: Our goal is to capture a range of variations in both geometry and materials. We opted against humans due to the complexities in modeling realistic fine geometric details (like hair and beard) and materials (such as skin), and the limited topological variations. Although humans have been used for synthetic data studies as demonstrated in [39], exploration in this domain is complicated because the procedural model was not released and high-fidelity models such as Unreal Metahumans by Epic Games [15] tend to prioritize realism over variability. Plants, another option, present challenges in material diversity and are difficult for human annotators to segment. Windows, in contrast, offers a balanced combination of geometric complexity and material variation suitable for our research goals. + +High-resolution: The images in the dataset should be high resolution to make the dataset useful for the near and medium-term future. This is in contrast to existing architectural datasets, with resolutions typically in the 512-1025 pixel range [5, 6, 25, 33]. While these are not low resolution in typical computer vision terms, architectural features such as windows are often too small to resolve in the images. Although the CityScapes dataset is larger than most, with a resolution of $2048 \times 1024$ pixels, it often captures architecture from oblique viewpoints limiting the effective resolution. + +Image Rights: We would like to have all rights to the images to avoid future copyright problems. + +To explore this problem, we introduce a specialized dataset that bridges real-world architectural imagery and procedurally-generated images, with a particular focus on window designs. Our dataset, comprising 89,318 windows in 75,739 high-resolution (4K to 6K) and RAW images, not only matches the scale of established datasets like CelebAHQ and FFHQ but also offers a unique niche with an emphasis on architectural details and high resolutions. + +Our dataset is designed not just to advance synthetic data generation, but also to facilitate diverse research avenues. Our primary contributions1 are: + +1. Real-World Imagery: A 4K resolution dataset with 75,739 photos of windows from global locations, offering unprecedented detail for architectural research. +2. Hand-Annotated Labels: Segmentation labels for 9,002 images. +3. A Procedural Model: A novel procedural model for windows, underscoring key design choices for synthetic data generation. +4. Synthetic Data and Labels: A diverse set of 21,290 synthetic window images, mirroring features observed in real imagery. + +Additionally, we conduct extensive experiments and ablations to understand the impact of various features in our synthetic dataset on segmentation performance. + +# 2. Related Work + +Several datasets of architectural imagery have been created, enabling applications such as architectural style classification [2, 9, 40], building functional use classification [21, 41], architectural heritage classification [27], landmark identification [31], or urban scene matching [18]. Notably, datasets that support image synthesis or image segmentation have been instrumental in advancing these fields. For instance, the FaSyn13 dataset [13] consists of 200 façade images for texture synthesis, but its size is limited for modern generative models. Similarly, the LSAA dataset [42] includes 199,723 façade images and 516,000 cropped window images. However, the resolution of these cropped windows varies, with the majority being less than 100 pixels in the longest dimension. + +In comparison, extensive collections like MS COCO [26] and LAION [35] offer a broader range but with challenges in resolution and scene variety. For instance, the sheer diversity in LAION's 9.8 million 1K-resolution images presents significant challenges for a single procedural model, highlighting the need for focused datasets. Our dataset addresses this by providing high-resolution window images, offering a specialized resource for detailed architectural analysis and procedural modeling. + +Our windows-focused dataset, while containing fewer windows than LSAA, stands out as the highest-resolution dataset of window images known to us, with an average resolution of 4,000 pixels per side for cropped windows. This high-resolution focus, particularly in the 4K to 6K range, fills a unique niche in architectural and synthetic-to-real domain transfer research. Our dataset's specialization in high-resolution architectural elements, especially windows, offers a valuable resource for advancing procedural modeling, emphasizing the importance of detail and precision. + +Several architectural image datasets supporting semantic segmentation or façade parsing exist, though not specifically focused on windows. The Graz dataset [33] contains 50 rectified images, and the eTRIMS dataset [25] contains 60 + +non-rectified images. The CMP-Facade dataset [36] contains 606 images, with about half of them fairly high resolution (1,024 pixels on the long edge) but a limited diversity of image locations. The LabelMe-Facade dataset [5] has the largest number of images at 945, with each image varying in size between 512 and 768 pixels on a side. However, these datasets do not have the number of images nor the resolution that are desirable for training the latest computer vision methods, which increasingly require more detailed and high-resolution data. With 9,002 labeled images at four times the resolution of these datasets, our proposed dataset of real-world images is an order of magnitude larger. By concentrating exclusively on windows, our dataset offers unprecedented detail and specificity, enabling more precise and effective models for architectural element analysis. + +Various authors have attempted to use synthetically generated data to bootstrap performance on real images. This approach seems to work best in domains where the human annotation is not directly feasible, such as reinforcement learning, especially for driving applications [14], depth or optical flow [8, 16], or 6DoF pose estimation for robotic grasping or manipulation [19, 24, 37]. + +Infinigen [32] is a good example of a procedural model, however, there is no validation of the effectiveness of the model for machine learning tasks. Of particular note is the SynthIA dataset [34], a driving dataset built on video game technology specifically designed to support semantic segmentation in urban environments. A very large engineering effort went into this, as well as CARLA [14], and we believe that reproducing such high-quality synthetic data is out of reach for most academic teams. Similar to our dataset, SynthIA aims at pushing the envelope to use synthetic data to improve computer vision even for problems where large human-annotated datasets (KITTI [28], LabelME-Facade [6], Camvid [4]) already exist. Similar to our findings, they can get some results from purely synthetic data, but they cannot out-compete even relatively small real-world labeled images, but by combining synthetic and real at a 4:6 ratio they obtain their best results. However, unlike SynthIA, our segmentation challenge is more constrained (only windows) and we think would require fewer resources for academic researchers to develop competing procedural models for synthetic training. Although buildings are visible in these datasets, they are not focused on architecture. Domain transfer for architecture is challenging due to the amount of variety and complex dependencies between architectural elements, whereas the categories of objects relevant in driving scenes are much more clearly defined. In addition, the high resolution of the images we use makes the synthesis of realistic textures and precise object boundaries critical, and the higher-capacity segmentation models of today vs. 2016 (when SynthIA was published) are more precise but may also be more likely to overfit synthetic data. + +![](images/80bb774f4a8fbf03f0c83b44b953244b6c6fff18a51acd504334b34f48ff8524.jpg) +Figure 2. Samples from the 75,739 photographs in the dataset. Each column shows a variety of examples of windows from different geographic locations. From left to right: Chicago (USA), Cambridge (UK), Bangkok (Thailand), Cairo (Egypt), and Vienna (Austria). The dataset has a variety of window shapes and architectural styles. + +Our dataset of real and synthetic imagery is unique as a high-resolution, voluminous dataset and serves as a proving ground for synthetic to real training, image generation, and semantic segmentation tasks. + +To our knowledge, the most successful work on using synthetic data for segmentation is the 'Fake It Till You Make It' paper by Wood et al. [39], which reports improvements in segmentation when a U-Net is trained using their synthetic data vs. real image. However, to get these improved results, they used label adaptation, a technique that requires real labeled data. This is counter to our goals for using synthetic data for domains where labels are scarce. They do not report segmentation results without label adaptation, but they do show in their ablation study on landmark localization that label adaptation is critical for benefiting from synthetic data. + +# 3. Real-Windows Dataset + +In line with our goal of advancing procedural modeling, our dataset is curated to offer high-resolution, diverse imagery essential for developing and evaluating procedural models. Our dataset, comprising over 75,739 high-resolution images (up to $4K \times 6K$ ), is one of the largest and most detailed collections of window imagery available. Unlike other datasets sourced from the web or Flickr [22, 23], we hold complete copyright ownership of every image, ensuring legal clarity and flexibility for research use2. + +The dataset's diversity in location, viewpoint, and archi + +![](images/09b04027d54e0c02b818249e2f20a9bafdc82104227be36a27b5101f90b03ee0.jpg) +Figure 3. Examples of the labels used to annotate our data. Each instance receives its own polygon. The reader may wish to zoom into the figure for details. + +tectural style, was achieved through a global effort. We engaged photographers from various countries, primarily hired via Upwork, to capture a wide array of window designs. This global collection effort ensures our dataset represents a broad spectrum of cultural and architectural diversity. This diversity is crucial for procedural models to learn and adapt to a wide range of architectural styles, directly supporting our goal of creating versatile and realistic models. Each image was chosen to highlight the window's design and architectural context, ensuring clear, distraction-free, and well-framed shots. The emphasis was on professional-grade photography with a focus on 4K resolution, achieving at least 2K pixels across each window, and taken during the day to ensure balanced exposure and minimal noise. + +In curating this collection, we also prioritized ethical considerations, instructing photographers to avoid capturing private situations or sensitive locations. Over 12 months, this project involved hiring 30 photographers and incurring costs of US $0.20-$ 0.50 per image, in addition to our quality control and subcontract management expenses. + +The diversity of image locations is indicated in Table 1 of the Appendix, along with the number of images that include semantic segmentation labels and RAW camera data. The RAW camera data's higher bit-depth could be particularly valuable for future procedural modeling research, offering richer information for model training and evaluation. + +Many photos included multiple windows. We manually annotated crops in each image as a region that includes a single window in the center, along with any portion of the wall that may have been adapted to the window (such as brickwork or molding) and a portion of the wall on all four sides of the window. Due to this cropping, window images are in a variety of sizes, as shown in Appendix Fig. 2. We store the original images and the cropping information separately and generate a cropped version of the dataset on demand. + +Each image has been carefully cropped and then annotated to focus on a single window and its immediate architectural context, as illustrated in Fig. 3. The annotations are designed to test if procedural models accurately replicate real-world architectural elements. The cropped win + +
LabelImages UsingArea %
wall890743.02%
window pane836222.58%
wall frame869714.91%
window frame86819.71%
unlabeled29943.09%
shutter9312.56%
balcony9731.08%
misc object23571.07%
blind3750.75%
bars6790.68%
open-window9770.55%
+ +Table 1. The labels used, their frequency of use, and percentage by area for the square-cropped dataset used for our experiments. We note that the dataset is a mix of well-used labels such as wall and less-used ones, such as blind or bar. The 'unlabeled' category contained areas beyond the building (e.g., sky, streets) and a much smaller number of ambiguous areas where we could not reach a decision on how to label a feature. + +dow images also underwent a manual annotation process for panoptic segmentation, and subsequent review for quality assurance, leading to the acquisition of 9,002 annotations at an average cost of US $3.90 per image. While manageable, this cost notably exceeds image acquisition expenses, underscoring the value of methods that reduce reliance on labor-intensive labeled data for segmentation tasks. A detailed account of the annotation process is presented in Section 5. + +The WinSyn dataset serves as a foundational step towards our larger goal of procedural urban modeling, offering a comprehensive and detailed resource for both current and future research in this domain. + +# 4. Procedural Model Development + +In developing our procedural model, we aim to create a realistic synthetic dataset with the ability to produce variations (ablations and experiments) for evaluation against real-world data. The compact domain (windows are relatively simple to model) allows well-developed approaches to be used to create the scene geometry including Split + +Shape Grammars [38] and the CGA language [29], combined with Bezier splines for curved window shapes. This approach ensures that our model can generate a wide range of architectural styles and window designs closely resembling real-world distributions. The canonical orientation offered by windows as a domain allows a layered approach to geometry generation, as Appendix Fig.11: from the camera, we generate street clutter (e.g., cars, bollards, shrubs), wall decorations (shutters, balconies, pipes) exterior walls, window geometry, window dressing (blinds, curtains), and interior geometry. + +The procedural modeling pipeline, developed in Python, Blender [11], and rendered with the physically based Cycles [3] renderer is used to generate a diverse dataset of 21,290 synthetic window images with corresponding labels (as Appendix Fig. 4). Normal, depth (Fig. 1), and edge maps are optionally generated. The model was designed to prioritize diversity, making use of domain randomization to extend beyond typical real-world variations. + +The procedural model's core uses two varieties of Split Grammar. The first, utilizing the CGA language, was employed to subdivide the volume to create the building mass, facades, shutters, blinds, balconies, roof, and window-bounds. From these window-bounds, a second Split Shape Grammar creates the window shapes themselves, using Bezier spline curves to capture various geometries, such as trapezoid, arched, or circular windows. This grammar subdivides closed curves to create nested window frames and offsets them to create individual glass panes. The frame geometry is created by extruding a profile along each Bezier. The profiles are selected from a randomly selected hierarchy of profiles. Parts of the window geometry hierarchy can be translated and rotated to 'open' windows by sliding or hinged mechanisms. This multi-grammar approach allows for detailed and varied window designs, while the generation remains algorithmically robust at scale. + +The procedural model is highly parameterized, with the number of parameters ranging from 216 to 21,735 in our dataset, depending on the chosen sequence of randomly-sampled rules in the grammars. We observe that optionally reusing parameters between parts of the model can improve visual realism; for example, we see window frames sharing their material (paint or stucco) with walls, or adjacent windows having similar (but not identical) materials. In the analysis section below we vary the distributions of these parameters to ablate and experiment on our procedural model and identify the most performant factors, including materials, textures, geometry, camera position, and lighting. + +Textures are primarily procedural shaders [7], controlling materials such as wood, brick, or glass applied to appropriate object classes. To add realism, we captured exterior clutter, such as building signage, vehicles, and trash cans, through a number of sources, including pre-existing + +datasets and LiDAR/RGB scanning (Appendix Fig. 12) with bitmap textures. + +Scene lighting is supplied by a combination of skybox emission, sun-lamp, and optional interior sources. We use interior and exterior panoramic images for the skybox and interior-box to create realistic background environments for our geometry. The camera is strategically positioned in order to capture the entire window from a predominantly frontal view; we use a variety of camera distances and adjust the field of view to frame the target window in the façade. Additional details of the procedural model can be found in Appendix Section 6. + +# 5. Estimating Procedural Model Quality + +We use semantic segmentation as a benchmark to evaluate the quality of a procedural model. This method involves training segmentation models on labeled images generated from the procedural model, evaluating model performances on a holdout (test) set of labeled real-world images, and providing a robust measure of the synthetic data's ability to accurately reflect real-world scenarios. + +For this approach to be successful, we must align our semantic labels with both real-world images and images generated by a procedural model. We established ten broad categories such as 'window-pane', 'window-frame', and 'wall frame' (see Appendix Section 7.7). This categorization mirrors the elements that our procedural model can realistically generate. Furthermore, we label each unique instance within the images, as demonstrated in Fig. 3, making the data suitable for panoptic segmentation even though we only evaluate the semantic segmentation task. This detailed set of labels extends traditional segmentation datasets, which typically provide a single 'window' label, with occasional additions like 'shutter' or 'blind' [6, 25, 36]. Our labeling allows consistent annotation across varied architectural styles but also ensures compatibility with the capabilities of procedural modeling. + +When utilizing mIoU to evaluate procedural model performance, inspecting plots such as Figures 6 and 7 can be insightful. These visual analyses aid in identifying parameters that optimize the model's performance. Additionally, Spearman's rank correlation is applied to quantify the impact of changes in model parameters (such as textures and lighting) on the synthetic image quality. + +# 6. Analysis + +In all following experiments, we fine-tune a BEiT 'base' model that was pre-trained on ImageNet-1k, trained on ImageNet-21k, and fine-tuned on data from some variation of our procedural model. We evaluated the mIoU over 10 labels, excluding the 'unlabeled' category. All cropped window-images are resized to 512 pixels square; examples + +![](images/f7579ab36ee87fdb89b128d647000e3751fea2649d785df9567365a60d725e49.jpg) +Figure 4. Left: histogram of per-image mIoUs showing the distribution of labeling results for a model trained on $n = 2,048$ synthetic (s) and real (r) images. We also show the difference between the model mIoU's per image (r-s). The mIoU was evaluated on 4,906 real images. Right: random samples of the labeling quality for networks trained on real and synthetic data; the first sample with an mIoU above each decile was selected. + +![](images/67d771064bf87d8d978e4b955b0013f9c40dd95992d5e290c96afffdeaa2cf2b.jpg) + +# are shown in Fig. 4. + +The performance of our procedural model is best understood in the context of regional performance disparities within real-world data. We contextualize the procedural model by comparing it to subsets from various regions of real-world data in Table 2 shows. Each set consists of 1,024 training images and 300 test images. Models trained on synthetic data have a narrower performance range (29.33 to $35.02\mathrm{mIoU}$ ) compared to those trained on locale-specific data (25.23 to $61.85\mathrm{mIoU}$ ). Notably, synthetic data can surpass real data from a different locale, as seen in the comparison of England and Egypt results. This narrower range seems to indicate synthetic data has not over-fit any particular region. The overall mIoU gap on the global dataset underscores the procedural model's limitations. The combined training set of real-world images evaluated to 53.79 mIoU, which is an upper limit on what one should expect from any model. The synthetic data yields $31.23\mathrm{mIoU}$ on the global test set, whereas other locales varied from 37.76-51.22 mIoU for the 'other' set. Given the global set includes holdout images from each locale, it is not surprising the mIoU's are slightly higher. This table indicates that the procedural model is of comparable quality to choosing a single locale. + +# 6.1. Procedural Model Variations + +We assess the impact of synthetic dataset variations on a segmentation model by comparing against our baseline model. For each variation we render 2,048 training examples and use the same geometry, lighting, and rendering settings as the baseline unless specified. For each variation, we render 2,048 training examples. Empirically, we find that the performance of synthetic data levels off at this number of examples (see Appendix Fig. 7) and evaluation is faster if the set is kept smaller. Evaluation takes place on a test partition of 4,906 real-world images; the mIoU of the baseline model is 32.58 and real photos have an mIoU of 58.69. To gauge each variation's importance, we either the relative + +![](images/fb57500f476c5531018db3d308044d9afb66a5acc458761f2dfeb2d618b62986.jpg) +Figure 5. The effect of varying real-world dataset sizes on mIoU with (blue) and without (red) an additional 2,048 synthetic samples. The green area bounds the range in which adding real data neither harms nor improves performance; the right-most point of which has $n = 152$ with an mIoU of 44.96. At larger datasets, synthetic data slightly reduces mIoU relative to only using real-world data. + +mIoU range as a percentage of the baseline, or report the mIoU Spearman's rank correlation $(r_s)$ . Detailed results are in Appendix Section 7 and summarized here. + +Rendering samples. We evaluated the impact of samples per pixel (spp) on render quality, noting diminishing returns beyond 256spp (Fig. 6). Render times scale from 6.4s at 1spp to 85.2s at 512spp. A strong correlation $(r_s = 1, n = 10)$ exists between spp and mIoU, with a $68\%$ change in mIoU scores relative to baseline, underlining the importance of spp. In this experiment no denoising was performed, however, our baseline model used 256spp and a powerful neural denoiser. + +
test
globalAustriaEgyptUKUSAothersynthetic
trainglobal53.7958.2056.8542.6450.1654.7431.21
Austria39.4958.2328.3634.4035.6741.3221.87
Egypt47.5138.3261.8533.0238.0149.7531.34
UK37.7640.6825.2338.5635.3937.6428.51
USA41.0842.5027.3933.3050.0839.3927.02
other51.2248.5439.4737.4441.3352.7428.39
synthetic31.2329.5329.3332.1534.1735.0262.12
+ +Table 2. mIoU for different splits of the real labeled data on the segmentation task. Trained on 1,024, tested on 300 samples. global is a mixture of all the real data; other data is from locales outside of Austria, Egypt, UK, or USA. In this experiment our synthetic network is similarly trained on 1,024 samples from our baseline synthetic model. + +![](images/020084d66d4901c3debd19c4e74dc6de13d5454796ae8d052c825e84d351bef3.jpg) +Figure 6. Top: The impact of rendering samples per pixel (spp) on segmentation task accuracy. Bottom: example renderings at different spp with zoomed section of rendering. $\mathrm{BL} =$ baseline. + +Mixed Materials. We experimented with 9 material variations, including uniform gray, edge shaders, and random textures. Some tests used a single material for the entire scene, others assigned different materials per object. Simplifying materials, such as rendering only the albedo channel, resulted in a significant performance drop. Appendix Fig. 18 shows that any material restriction led to at least an $18\%$ mIoU decrease from the baseline. + +Lighting. We examined 8 lighting model variations and their mIoU impact, as detailed in Appendix Fig. 19. Lighting models like albedo-only were crucial, while lighting conditions such as night-time had moderate importance. Conditions varied mIoU by $15.35\%$ from baseline. Daylight-only training decreased baseline mIoU by $1.04\%$ + +relative to baseline. + +Camera. We experimented with the distribution of camera positions. These 7 variations used a simple model which sampled a camera position over a circle, of radius $r = \{0\ldots 96\}$ meters, truncated at the floor plane. The circle is positioned 5 meters from the wall, directly in front of the window. The camera's field of view is adjusted to the apparent window size. We observed limited impact on mIoU as $r$ changes; peak task performance was at $r = 12m$ . + +Window Geometry. We ran 7 tests with varying window dimensions and shapes, including square and nonrectangular windows. The mIoU impact was minor, fluctuating by up to $5.7\%$ relative to the baseline (1.86 absolute), with the best variation $1.2\%$ worse relative to the baseline $(-0.04$ absolute). Small features, though noticeable to humans, had a weak impact on model performance. + +Labels Modeled. In developing our procedural synthetic data generator, we prioritized labels by size, starting with wall and ending with open-window (Appendix Section 4). This enabled the assessment of mIoU at nine developmental stages (Fig. 7). Adding smaller classes later showed diminishing returns and occasionally reduced single-label accuracy (Appendix Fig. 22). + +# 6.2. Additional Experiments + +In our segmentation experiments, we employed the BEiTv2 base model [1, 30] for its proficiency in generating accurate results with minimal data. However, the trends we observed are consistent across various models, including DeepLabv3+ [10]. In the Appendix Section 5, we discuss two experiments that study the impact on the segmentation task when training (and testing) with different mixes of real and synthetic data. The first experiment in Appendix Fig. 7 explores segmentation performance when training and testing on different amounts of real or synthetic data. We observe that synthetic data saturates (stops improving with additional data), while real-world data continues to benefit from additional samples. The second experiment, in Appendix Fig. 8, illustrates the impact of different mixtures of + +![](images/bb1dbc150eac29e2e0fc62818df86a72bc6ad3b6d5ac80b3969070bd50c4f866.jpg) +Figure 7. Top: The progressive impact of incorporating additional labels into the procedural model, with label subsets expanding from lvl1 to lvl9. The sequence begins with 'walls' only (lvl1), adding 'window panes' (lvl2), 'wall frames' (lvl3), 'window frames' (lvl4), 'shutters' (lvl5), 'balconies' (lvl6), 'miscellaneous objects' (lvl7), 'blinds' (lvl8), and 'bars' (lvl9). The full model includes additional features beyond lvl9, such as interior dressing, open windows, and windows without glass. Bottom: Examples corresponding to each incremental level of model complexity. + +real and synthetic data, when tested on real data. Mixing synthetic and real data can be advantageous when there is little real data, but adding large amounts of synthetic data does not help task performance. We conclude that no volume of synthetic data can overcome this domain gap. + +An exploration of different architectures and data partitions is presented in Appendix Fig. 9 and Table 2 - we train an older convolution network without pre-training (DeepLab3+ [10]), a 'large' BEiTv2 model, an architecture which uses real data labels during training to improve performance (Label Adaptation [39]), another technique which uses unlabeled data to improve performance (MIC [20]), adjusting the colors of the training dataset (Histogram Matching [17]), and creating an 'easy' real data partition. Under these approaches, the synthetic-real domain gap persists, suggesting that network architecture and data partitions are an orthogonal research direction to improving the quality of synthetic procedural models. + +# 7. Acknowledgements + +We thank our lead photographers Michaela Nömayr and Florian Rist, and engineer Prem Chedella, as well as our contributing photographers: Aleksandr Aleshkin, Angela Markoska, Artur Oliveira, Brian Benton, Chris West, Christopher Byrne, Elsayed Saafan, Florian Rist, George Ilin, Ignacio De Barrio, Jan Cuales, Kaitlyn Jackson, Kalina Mondzholovska, Kubra Ayse Guzel, Lukas Bornheim, Maria Jose Balestena, Michaela Nömayr, Mihai Alexandru Filoneanu, Mokhtari Sid Ahmed Salim, Mussa Ubapa, Nestor Angulo Caballero, Nicklaus Suarez, Peter Fountain, Prem Chedella, Samantha Martucci, Sarab-jot Singh, Scarlett Li, Serhii Malov, Simon R. A. Kelly, Stephanie Foden, Surafel Sunara, Tadiyos Shala, Susana Gomez, Vasileios Notis, Yuan Yuan, and finally labelyour-data.com for the labeling. + +# 8. Conclusion + +We have introduced a new dataset of 75,739 photos (2.09 terapixels), of which 9,002 photos are semantically labeled, as well as a high-quality procedural model that closely approximates real-world variation, making it effective for image segmentation tasks. Together these components of WinSyn have applications in unsupervised domain adaptation (using the large amount of unlabeled data), superresolution (via the thigh resolution images), learning from RAW sources, generative modeling, 3D reconstruction or depth recovery (using the synthetic depth or 3D geometry from the procedural model), as well as learning from images containing transparent or specular materials (i.e., window panes/glass balconies are often transparent). + +We systematically explored the effect of variations on our model on mIoU over 64 variations and 156,903 renders. The mIoU performance gap between our synthetic and real-world data is comparable to inter-country differences with largely different architectural styles. However, the difference between synthetic data and real data is still much larger than desired. This gap is not sufficiently reduced by either our work or other network architectures or dataset scale. We, therefore, believe that WinSyn provides a timely and efficient testbench from which researchers can iterate quickly to explore graphics contributions to the synthetic data problem. Our work contributes a sizable and versatile dataset that can be the basis of exciting and much-needed progress in the area of procedural graphics for synthetic data generation in machine learning. The key value of our dataset is that as a new benchmark of simulated and real images, it enables others to study the problem at an approachable level of complexity. + +# References + +[1] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 7 +[2] Björn Barz and Joachim Denzler. Wikichurches: A fine-grained dataset of architectural styles with real-world challenges. arXiv preprint arXiv:2108.06959, 2021. 2 +[3] Cycles: Open Source Production Rendering. Blender Foundation, 2023. 5 +[4] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV (1), pages 44-57, 2008. 3 +[5] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3 +[6] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3, 5 +[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, pages 1-7. vol. 2012, 2012. 5 +[8] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 3 +[9] J Chen, R Stouffs, and F Biljecki. Hierarchical (multi-label) architectural image recognition and classification. 2021. 2 +[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7, 8 +[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2023. 5 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[13] D. Dai, H. Riemenschneider, G. Schmitt, and L. Van Gool. Example-based facade texture synthesis. In International Conference on Computer Vision (ICCV), 2013. 2 +[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3 +[15] Epic Games, Inc. MetaHuman Realistic Person Creator, 2021. 2 +[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking anal + +ysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3 +[17] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 8 +[18] Daniel Cabrini Hauagge and Noah Snavely. Image matching using local symmetry features. In 2012 IEEE conference on computer vision and pattern recognition, pages 206-213. IEEE, 2012. 2 +[19] Tomás Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jérí Matas. BOP challenge 2020 on 6D object localization. European Conference on Computer Vision Workshops (ECCVW), 2020. 3 +[20] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023. 8 +[21] Jian Kang, Marco Korner, Yuanyuan Wang, Hannes Taubenböck, and Xiao Xiang Zhu. Building instance classification using street view images. ISPRS journal of photogrammetry and remote sensing, 145:44-59, 2018. 2 +[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3 +[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. International Conference on Computer Vision (ICCV) Workshops, 2019. 3 +[25] F. Korč and W. Förstner. eTRIMS Image Database for interpreting images of man-made scenes. Technical Report TR-IGG-P-2009-01, 2009. 2, 5 +[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 2 +[27] Jose Llamas, Pedro M. Lerones, Roberto Medina, Eduardo Zalama, and Jaime Gómez-García-Bermejo. Classification of architectural heritage images using deep learning techniques. Applied Sciences, 7(10):992, 2017. 2 +[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3 +[29] Pascal Mueller, Peter Wonka, Simon Haegler, Andreas Ulmer, and Luc Van Gool. Procedural modeling of buildings. ACM Trans. Gr., 25(3):614-623, 2006. 5 +[30] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers, 2022. 7 +[31] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies + +and fast spatial matching. In IEEE Conference on Computer Vision and Pattern Recognition, 2007. 2 +[32] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, et al. Infinite photorealistic worlds using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12630-12641, 2023. 3 +[33] Hayko Riemenschneider, Ulrich Krispel, Wolfgang Thaller, Michael Donoser, Sven Havemann, Dieter Fellner, and Horst Bischof. Irregular lattices for complex shape grammar facade parsing. In 2012 IEEE conference on computer vision and pattern recognition, pages 1640-1647. IEEE, 2012. 2 +[34] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 3 +[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[36] Radim Tyleček and Radim Šára. Spatial pattern templates for recognition of objects with regular structure. In Proc. GCPR, Saarbrucken, Germany, 2013. 3, 5 +[37] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-dof pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In International Conference on Intelligent Robots and Systems (IROS), 2022. 3 +[38] Peter Wonka, Michael Wimmer, François Sillion, and William Ribarsky. Instant architecture. ACM Trans. Gr., 22 (3):669-677, 2003. 5 +[39] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3661-3671. IEEE, 2021. 2, 3, 8 +[40] Zhe Xu, Dacheng Tao, Ya Zhang, Junjie Wu, and Ah Chung Tsoi. Architectural style classification using multinomial latent logistic regression. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 600-615. Springer, 2014. 2 +[41] K. Zhao, Y. Liu, S. Hao, S. Lu, H. Liu, and L. Zhou. Bounding boxes are all we need: Street view image classification via context encoding of detected buildings. IEEE Transactions on Geoscience and Remote Sensing, pages 1-17, 2021. 2 +[42] Peihao Zhu, Wamiq Reyaz Para, Anna Frühstück, John Femiani, and Peter Wonka. Large-scale architectural asset extraction from panoramic imagery. IEEE Transactions on Visualization and Computer Graphics, 28(2):1301-1316, 2022. 2 \ No newline at end of file diff --git a/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/images.zip b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a62077102d7a0acd2030950a7ede4a2967b818e2 --- /dev/null +++ b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e599b13551621991c100781eb7966f4cd68846e7d8db21d184b5522d8e07753c +size 452627 diff --git a/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/layout.json b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..241750d194313afd5bae43f320e61272c0148aad --- /dev/null +++ b/2024/WinSyn_ _ A High Resolution Testbed for Synthetic Data/layout.json @@ -0,0 +1,5927 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 468, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 468, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 468, + 121 + ], + "type": "text", + "content": "WinSyn: A High Resolution Testbed for Synthetic Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 142, + 407, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 142, + 407, + 159 + ], + "spans": [ + { + "bbox": [ + 182, + 142, + 407, + 159 + ], + "type": "text", + "content": "Tom Kelly1 John Femiani2 Peter Wonka1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 161, + 380, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 161, + 380, + 175 + ], + "spans": [ + { + "bbox": [ + 200, + 161, + 380, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 200, + 161, + 380, + 175 + ], + "type": "text", + "content": "KAUST, KSA " + }, + { + "bbox": [ + 200, + 161, + 380, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 200, + 161, + 380, + 175 + ], + "type": "text", + "content": "Miami University, Ohio" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 202, + 192, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 202, + 192, + 216 + ], + "spans": [ + { + "bbox": [ + 143, + 202, + 192, + 216 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 228, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 290, + 502 + ], + "type": "text", + "content": "We present WinSyn, a unique dataset and testbed for creating high-quality synthetic data with procedural modeling techniques. The dataset contains high-resolution photographs of windows, selected from locations around the world, with 89,318 individual window crops showcasing diverse geometric and material characteristics. We evaluate a procedural model by training semantic segmentation networks on both synthetic and real images and then comparing their performances on a shared test set of real images. Specifically, we measure the difference in mean Intersection over Union (mIoU) and determine the effective number of real images to match synthetic data's training performance. We design a baseline procedural model as a benchmark and provide 21,290 synthetically generated images. By tuning the procedural model, key factors are identified which significantly influence the model's fidelity in replicating real-world scenarios. Importantly, we highlight the challenge of procedural modeling using current techniques, especially in their ability to replicate the spatial semantics of real-world scenarios. This insight is critical because of the potential of procedural models to bridge to hidden scene aspects such as depth, reflectivity, material properties, and lighting conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 545, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 287, + 701 + ], + "type": "text", + "content": "Larger and more sophisticated machine learning models demand an ever-increasing supply of data, particularly in tasks where manual annotation is challenging, such as depth estimation, reflectance estimation, or full 3D reconstruction. One solution is to use procedural models to generate synthetic training data. However, creating procedural models that accurately reflect the domain of real images (i.e. closing the domain gap) remains an open problem. Despite the visual realism of many synthetic scenes, their effectiveness in machine-learning applications often falls short. In this work we do not close the gap – but present an accessible pair of real and synthetic datasets, with annotations, in which this domain gap may be studied without massive resources." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "While the final goal is to tackle complex problems, a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 204, + 547, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 275 + ], + "type": "text", + "content": "straightforward proxy task is initially required to ensure that real-world imagery can be manually annotated and compared to synthetic imagery. We, therefore, propose segmentation as a proxy task. This proxy task helps pinpoint where a procedural model fails to capture the diversity and nuances of real-world scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 363 + ], + "type": "text", + "content": "As our long-term goal is procedural urban modeling, we initially considered street-view images and datasets, such as CityScapes [12]. However, this has several drawbacks. Modeling arbitrary urban scenes realistically requires overwhelming complexity; requiring modeling at least cars, humans, buildings, skies, and vegetation, and each of these poses quite distinct and complex modeling challenges. For" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 384, + 523, + 517 + ], + "blocks": [ + { + "bbox": [ + 331, + 384, + 523, + 517 + ], + "lines": [ + { + "bbox": [ + 331, + 384, + 523, + 517 + ], + "spans": [ + { + "bbox": [ + 331, + 384, + 523, + 517 + ], + "type": "image", + "image_path": "c7b8acacaf2fae6521302ba70178834a0722853ed224986707a746e8eed9e5e7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 331, + 518, + 523, + 647 + ], + "blocks": [ + { + "bbox": [ + 331, + 518, + 523, + 647 + ], + "lines": [ + { + "bbox": [ + 331, + 518, + 523, + 647 + ], + "spans": [ + { + "bbox": [ + 331, + 518, + 523, + 647 + ], + "type": "image", + "image_path": "a682de236db7c90e79bec5ef5b56f69af41e62115dfada4b5b657902f43de5ca.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 654, + 547, + 711 + ], + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 711 + ], + "type": "text", + "content": "Figure 1. Photographers in 28 geographic regions captured real-world photos (a) of windows that are cropped (b) to single windows, which are then labeled (c). Synthetic windows are rendered giving color (f) and labels(d), while other passes such as depth (e) are also possible." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22456" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "example, modeling urban environments for video games is a significant undertaking that often requires over 100 artists for open-world games. This type of effort is unrealistic for procedural modeling research. We propose that it is more promising to identify a simpler subset of images and develop a more constrained procedural model for this subset. This should enable faster iterations and broader participation in procedural modeling research, before scaling up to complete cities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 181, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 181, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 287, + 205 + ], + "type": "text", + "content": "The decision to study the domain of window images was based on the following design principles for our dataset:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 206, + 289, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 375 + ], + "type": "text", + "content": "Diversity: Our goal is to capture a range of variations in both geometry and materials. We opted against humans due to the complexities in modeling realistic fine geometric details (like hair and beard) and materials (such as skin), and the limited topological variations. Although humans have been used for synthetic data studies as demonstrated in [39], exploration in this domain is complicated because the procedural model was not released and high-fidelity models such as Unreal Metahumans by Epic Games [15] tend to prioritize realism over variability. Plants, another option, present challenges in material diversity and are difficult for human annotators to segment. Windows, in contrast, offers a balanced combination of geometric complexity and material variation suitable for our research goals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 289, + 518 + ], + "type": "text", + "content": "High-resolution: The images in the dataset should be high resolution to make the dataset useful for the near and medium-term future. This is in contrast to existing architectural datasets, with resolutions typically in the 512-1025 pixel range [5, 6, 25, 33]. While these are not low resolution in typical computer vision terms, architectural features such as windows are often too small to resolve in the images. Although the CityScapes dataset is larger than most, with a resolution of " + }, + { + "bbox": [ + 46, + 387, + 289, + 518 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 46, + 387, + 289, + 518 + ], + "type": "text", + "content": " pixels, it often captures architecture from oblique viewpoints limiting the effective resolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 532, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 287, + 556 + ], + "type": "text", + "content": "Image Rights: We would like to have all rights to the images to avoid future copyright problems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "To explore this problem, we introduce a specialized dataset that bridges real-world architectural imagery and procedurally-generated images, with a particular focus on window designs. Our dataset, comprising 89,318 windows in 75,739 high-resolution (4K to 6K) and RAW images, not only matches the scale of established datasets like CelebAHQ and FFHQ but also offers a unique niche with an emphasis on architectural details and high resolutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 690 + ], + "type": "text", + "content": "Our dataset is designed not just to advance synthetic data generation, but also to facilitate diverse research avenues. Our primary contributions1 are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 547, + 204 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "type": "text", + "content": "1. Real-World Imagery: A 4K resolution dataset with 75,739 photos of windows from global locations, offering unprecedented detail for architectural research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 108, + 547, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 547, + 131 + ], + "type": "text", + "content": "2. Hand-Annotated Labels: Segmentation labels for 9,002 images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 132, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 547, + 167 + ], + "type": "text", + "content": "3. A Procedural Model: A novel procedural model for windows, underscoring key design choices for synthetic data generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 167, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 547, + 204 + ], + "type": "text", + "content": "4. Synthetic Data and Labels: A diverse set of 21,290 synthetic window images, mirroring features observed in real imagery." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 205, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 547, + 240 + ], + "type": "text", + "content": "Additionally, we conduct extensive experiments and ablations to understand the impact of various features in our synthetic dataset on segmentation performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 251, + 392, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 392, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 392, + 262 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 270, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 437 + ], + "type": "text", + "content": "Several datasets of architectural imagery have been created, enabling applications such as architectural style classification [2, 9, 40], building functional use classification [21, 41], architectural heritage classification [27], landmark identification [31], or urban scene matching [18]. Notably, datasets that support image synthesis or image segmentation have been instrumental in advancing these fields. For instance, the FaSyn13 dataset [13] consists of 200 façade images for texture synthesis, but its size is limited for modern generative models. Similarly, the LSAA dataset [42] includes 199,723 façade images and 516,000 cropped window images. However, the resolution of these cropped windows varies, with the majority being less than 100 pixels in the longest dimension." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": "In comparison, extensive collections like MS COCO [26] and LAION [35] offer a broader range but with challenges in resolution and scene variety. For instance, the sheer diversity in LAION's 9.8 million 1K-resolution images presents significant challenges for a single procedural model, highlighting the need for focused datasets. Our dataset addresses this by providing high-resolution window images, offering a specialized resource for detailed architectural analysis and procedural modeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 546, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 665 + ], + "type": "text", + "content": "Our windows-focused dataset, while containing fewer windows than LSAA, stands out as the highest-resolution dataset of window images known to us, with an average resolution of 4,000 pixels per side for cropped windows. This high-resolution focus, particularly in the 4K to 6K range, fills a unique niche in architectural and synthetic-to-real domain transfer research. Our dataset's specialization in high-resolution architectural elements, especially windows, offers a valuable resource for advancing procedural modeling, emphasizing the importance of detail and precision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 713 + ], + "type": "text", + "content": "Several architectural image datasets supporting semantic segmentation or façade parsing exist, though not specifically focused on windows. The Graz dataset [33] contains 50 rectified images, and the eTRIMS dataset [25] contains 60" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "type": "text", + "content": "Available online https://twak.github.io/winsyn" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22457" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "non-rectified images. The CMP-Facade dataset [36] contains 606 images, with about half of them fairly high resolution (1,024 pixels on the long edge) but a limited diversity of image locations. The LabelMe-Facade dataset [5] has the largest number of images at 945, with each image varying in size between 512 and 768 pixels on a side. However, these datasets do not have the number of images nor the resolution that are desirable for training the latest computer vision methods, which increasingly require more detailed and high-resolution data. With 9,002 labeled images at four times the resolution of these datasets, our proposed dataset of real-world images is an order of magnitude larger. By concentrating exclusively on windows, our dataset offers unprecedented detail and specificity, enabling more precise and effective models for architectural element analysis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "content": "Various authors have attempted to use synthetically generated data to bootstrap performance on real images. This approach seems to work best in domains where the human annotation is not directly feasible, such as reinforcement learning, especially for driving applications [14], depth or optical flow [8, 16], or 6DoF pose estimation for robotic grasping or manipulation [19, 24, 37]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 342, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 289, + 715 + ], + "type": "text", + "content": "Infinigen [32] is a good example of a procedural model, however, there is no validation of the effectiveness of the model for machine learning tasks. Of particular note is the SynthIA dataset [34], a driving dataset built on video game technology specifically designed to support semantic segmentation in urban environments. A very large engineering effort went into this, as well as CARLA [14], and we believe that reproducing such high-quality synthetic data is out of reach for most academic teams. Similar to our dataset, SynthIA aims at pushing the envelope to use synthetic data to improve computer vision even for problems where large human-annotated datasets (KITTI [28], LabelME-Facade [6], Camvid [4]) already exist. Similar to our findings, they can get some results from purely synthetic data, but they cannot out-compete even relatively small real-world labeled images, but by combining synthetic and real at a 4:6 ratio they obtain their best results. However, unlike SynthIA, our segmentation challenge is more constrained (only windows) and we think would require fewer resources for academic researchers to develop competing procedural models for synthetic training. Although buildings are visible in these datasets, they are not focused on architecture. Domain transfer for architecture is challenging due to the amount of variety and complex dependencies between architectural elements, whereas the categories of objects relevant in driving scenes are much more clearly defined. In addition, the high resolution of the images we use makes the synthesis of realistic textures and precise object boundaries critical, and the higher-capacity segmentation models of today vs. 2016 (when SynthIA was published) are more precise but may also be more likely to overfit synthetic data." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 245 + ], + "type": "image", + "image_path": "80bb774f4a8fbf03f0c83b44b953244b6c6fff18a51acd504334b34f48ff8524.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "type": "text", + "content": "Figure 2. Samples from the 75,739 photographs in the dataset. Each column shows a variety of examples of windows from different geographic locations. From left to right: Chicago (USA), Cambridge (UK), Bangkok (Thailand), Cairo (Egypt), and Vienna (Austria). The dataset has a variety of window shapes and architectural styles." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 342, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 389 + ], + "type": "text", + "content": "Our dataset of real and synthetic imagery is unique as a high-resolution, voluminous dataset and serves as a proving ground for synthetic to real training, image generation, and semantic segmentation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 390, + 548, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 548, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 548, + 535 + ], + "type": "text", + "content": "To our knowledge, the most successful work on using synthetic data for segmentation is the 'Fake It Till You Make It' paper by Wood et al. [39], which reports improvements in segmentation when a U-Net is trained using their synthetic data vs. real image. However, to get these improved results, they used label adaptation, a technique that requires real labeled data. This is counter to our goals for using synthetic data for domains where labels are scarce. They do not report segmentation results without label adaptation, but they do show in their ablation study on landmark localization that label adaptation is critical for benefiting from synthetic data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 545, + 440, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 440, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 440, + 558 + ], + "type": "text", + "content": "3. Real-Windows Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": "In line with our goal of advancing procedural modeling, our dataset is curated to offer high-resolution, diverse imagery essential for developing and evaluating procedural models. Our dataset, comprising over 75,739 high-resolution images (up to " + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "inline_equation", + "content": "4K \\times 6K" + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": "), is one of the largest and most detailed collections of window imagery available. Unlike other datasets sourced from the web or Flickr [22, 23], we hold complete copyright ownership of every image, ensuring legal clarity and flexibility for research use2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 673, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 673, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 317, + 673, + 545, + 685 + ], + "type": "text", + "content": "The dataset's diversity in location, viewpoint, and archi" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2While the datasets we referenced use images under permissive licenses, owning our images outright simplifies usage rights." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22458" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 153 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 153 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 153 + ], + "type": "image", + "image_path": "09b04027d54e0c02b818249e2f20a9bafdc82104227be36a27b5101f90b03ee0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 160, + 546, + 184 + ], + "lines": [ + { + "bbox": [ + 46, + 160, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 546, + 184 + ], + "type": "text", + "content": "Figure 3. Examples of the labels used to annotate our data. Each instance receives its own polygon. The reader may wish to zoom into the figure for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 383 + ], + "type": "text", + "content": "tectural style, was achieved through a global effort. We engaged photographers from various countries, primarily hired via Upwork, to capture a wide array of window designs. This global collection effort ensures our dataset represents a broad spectrum of cultural and architectural diversity. This diversity is crucial for procedural models to learn and adapt to a wide range of architectural styles, directly supporting our goal of creating versatile and realistic models. Each image was chosen to highlight the window's design and architectural context, ensuring clear, distraction-free, and well-framed shots. The emphasis was on professional-grade photography with a focus on 4K resolution, achieving at least 2K pixels across each window, and taken during the day to ensure balanced exposure and minimal noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": "In curating this collection, we also prioritized ethical considerations, instructing photographers to avoid capturing private situations or sensitive locations. Over 12 months, this project involved hiring 30 photographers and incurring costs of US " + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "inline_equation", + "content": "0.20-" + }, + { + "bbox": [ + 46, + 385, + 287, + 456 + ], + "type": "text", + "content": "0.50 per image, in addition to our quality control and subcontract management expenses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 531 + ], + "type": "text", + "content": "The diversity of image locations is indicated in Table 1 of the Appendix, along with the number of images that include semantic segmentation labels and RAW camera data. The RAW camera data's higher bit-depth could be particularly valuable for future procedural modeling research, offering richer information for model training and evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "type": "text", + "content": "Many photos included multiple windows. We manually annotated crops in each image as a region that includes a single window in the center, along with any portion of the wall that may have been adapted to the window (such as brickwork or molding) and a portion of the wall on all four sides of the window. Due to this cropping, window images are in a variety of sizes, as shown in Appendix Fig. 2. We store the original images and the cropping information separately and generate a cropped version of the dataset on demand." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Each image has been carefully cropped and then annotated to focus on a single window and its immediate architectural context, as illustrated in Fig. 3. The annotations are designed to test if procedural models accurately replicate real-world architectural elements. The cropped win" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 335, + 202, + 518, + 348 + ], + "blocks": [ + { + "bbox": [ + 335, + 202, + 518, + 348 + ], + "lines": [ + { + "bbox": [ + 335, + 202, + 518, + 348 + ], + "spans": [ + { + "bbox": [ + 335, + 202, + 518, + 348 + ], + "type": "table", + "html": "
LabelImages UsingArea %
wall890743.02%
window pane836222.58%
wall frame869714.91%
window frame86819.71%
unlabeled29943.09%
shutter9312.56%
balcony9731.08%
misc object23571.07%
blind3750.75%
bars6790.68%
open-window9770.55%
", + "image_path": "05c6c9d8660d24108b51d028beaa4f3bf3ad5138b22d33b6a1dff67bce6911dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 357, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 434 + ], + "type": "text", + "content": "Table 1. The labels used, their frequency of use, and percentage by area for the square-cropped dataset used for our experiments. We note that the dataset is a mix of well-used labels such as wall and less-used ones, such as blind or bar. The 'unlabeled' category contained areas beyond the building (e.g., sky, streets) and a much smaller number of ambiguous areas where we could not reach a decision on how to label a feature." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 455, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 561 + ], + "type": "text", + "content": "dow images also underwent a manual annotation process for panoptic segmentation, and subsequent review for quality assurance, leading to the acquisition of 9,002 annotations at an average cost of US $3.90 per image. While manageable, this cost notably exceeds image acquisition expenses, underscoring the value of methods that reduce reliance on labor-intensive labeled data for segmentation tasks. A detailed account of the annotation process is presented in Section 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 563, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 609 + ], + "type": "text", + "content": "The WinSyn dataset serves as a foundational step towards our larger goal of procedural urban modeling, offering a comprehensive and detailed resource for both current and future research in this domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 621, + 484, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 621, + 484, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 621, + 484, + 635 + ], + "type": "text", + "content": "4. Procedural Model Development" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "content": "In developing our procedural model, we aim to create a realistic synthetic dataset with the ability to produce variations (ablations and experiments) for evaluation against real-world data. The compact domain (windows are relatively simple to model) allows well-developed approaches to be used to create the scene geometry including Split" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22459" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Shape Grammars [38] and the CGA language [29], combined with Bezier splines for curved window shapes. This approach ensures that our model can generate a wide range of architectural styles and window designs closely resembling real-world distributions. The canonical orientation offered by windows as a domain allows a layered approach to geometry generation, as Appendix Fig.11: from the camera, we generate street clutter (e.g., cars, bollards, shrubs), wall decorations (shutters, balconies, pipes) exterior walls, window geometry, window dressing (blinds, curtains), and interior geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 302 + ], + "type": "text", + "content": "The procedural modeling pipeline, developed in Python, Blender [11], and rendered with the physically based Cycles [3] renderer is used to generate a diverse dataset of 21,290 synthetic window images with corresponding labels (as Appendix Fig. 4). Normal, depth (Fig. 1), and edge maps are optionally generated. The model was designed to prioritize diversity, making use of domain randomization to extend beyond typical real-world variations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 303, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 507 + ], + "type": "text", + "content": "The procedural model's core uses two varieties of Split Grammar. The first, utilizing the CGA language, was employed to subdivide the volume to create the building mass, facades, shutters, blinds, balconies, roof, and window-bounds. From these window-bounds, a second Split Shape Grammar creates the window shapes themselves, using Bezier spline curves to capture various geometries, such as trapezoid, arched, or circular windows. This grammar subdivides closed curves to create nested window frames and offsets them to create individual glass panes. The frame geometry is created by extruding a profile along each Bezier. The profiles are selected from a randomly selected hierarchy of profiles. Parts of the window geometry hierarchy can be translated and rotated to 'open' windows by sliding or hinged mechanisms. This multi-grammar approach allows for detailed and varied window designs, while the generation remains algorithmically robust at scale." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 509, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 289, + 653 + ], + "type": "text", + "content": "The procedural model is highly parameterized, with the number of parameters ranging from 216 to 21,735 in our dataset, depending on the chosen sequence of randomly-sampled rules in the grammars. We observe that optionally reusing parameters between parts of the model can improve visual realism; for example, we see window frames sharing their material (paint or stucco) with walls, or adjacent windows having similar (but not identical) materials. In the analysis section below we vary the distributions of these parameters to ablate and experiment on our procedural model and identify the most performant factors, including materials, textures, geometry, camera position, and lighting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Textures are primarily procedural shaders [7], controlling materials such as wood, brick, or glass applied to appropriate object classes. To add realism, we captured exterior clutter, such as building signage, vehicles, and trash cans, through a number of sources, including pre-existing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "datasets and LiDAR/RGB scanning (Appendix Fig. 12) with bitmap textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 217 + ], + "type": "text", + "content": "Scene lighting is supplied by a combination of skybox emission, sun-lamp, and optional interior sources. We use interior and exterior panoramic images for the skybox and interior-box to create realistic background environments for our geometry. The camera is strategically positioned in order to capture the entire window from a predominantly frontal view; we use a variety of camera distances and adjust the field of view to frame the target window in the façade. Additional details of the procedural model can be found in Appendix Section 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 228, + 515, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 515, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 515, + 243 + ], + "type": "text", + "content": "5. Estimating Procedural Model Quality" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 249, + 547, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 547, + 332 + ], + "type": "text", + "content": "We use semantic segmentation as a benchmark to evaluate the quality of a procedural model. This method involves training segmentation models on labeled images generated from the procedural model, evaluating model performances on a holdout (test) set of labeled real-world images, and providing a robust measure of the synthetic data's ability to accurately reflect real-world scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 334, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 525 + ], + "type": "text", + "content": "For this approach to be successful, we must align our semantic labels with both real-world images and images generated by a procedural model. We established ten broad categories such as 'window-pane', 'window-frame', and 'wall frame' (see Appendix Section 7.7). This categorization mirrors the elements that our procedural model can realistically generate. Furthermore, we label each unique instance within the images, as demonstrated in Fig. 3, making the data suitable for panoptic segmentation even though we only evaluate the semantic segmentation task. This detailed set of labels extends traditional segmentation datasets, which typically provide a single 'window' label, with occasional additions like 'shutter' or 'blind' [6, 25, 36]. Our labeling allows consistent annotation across varied architectural styles but also ensures compatibility with the capabilities of procedural modeling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 525, + 547, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 611 + ], + "type": "text", + "content": "When utilizing mIoU to evaluate procedural model performance, inspecting plots such as Figures 6 and 7 can be insightful. These visual analyses aid in identifying parameters that optimize the model's performance. Additionally, Spearman's rank correlation is applied to quantify the impact of changes in model parameters (such as textures and lighting) on the synthetic image quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 621, + 365, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 365, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 365, + 635 + ], + "type": "text", + "content": "6. Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "In all following experiments, we fine-tune a BEiT 'base' model that was pre-trained on ImageNet-1k, trained on ImageNet-21k, and fine-tuned on data from some variation of our procedural model. We evaluated the mIoU over 10 labels, excluding the 'unlabeled' category. All cropped window-images are resized to 512 pixels square; examples" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22460" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 42, + 68, + 276, + 174 + ], + "blocks": [ + { + "bbox": [ + 42, + 68, + 276, + 174 + ], + "lines": [ + { + "bbox": [ + 42, + 68, + 276, + 174 + ], + "spans": [ + { + "bbox": [ + 42, + 68, + 276, + 174 + ], + "type": "image", + "image_path": "f7579ab36ee87fdb89b128d647000e3751fea2649d785df9567365a60d725e49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 193, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 45, + 193, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 45, + 193, + 547, + 237 + ], + "type": "text", + "content": "Figure 4. Left: histogram of per-image mIoUs showing the distribution of labeling results for a model trained on " + }, + { + "bbox": [ + 45, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "n = 2,048" + }, + { + "bbox": [ + 45, + 193, + 547, + 237 + ], + "type": "text", + "content": " synthetic (s) and real (r) images. We also show the difference between the model mIoU's per image (r-s). The mIoU was evaluated on 4,906 real images. Right: random samples of the labeling quality for networks trained on real and synthetic data; the first sample with an mIoU above each decile was selected." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 275, + 70, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 275, + 70, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 275, + 70, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 275, + 70, + 545, + 178 + ], + "type": "image", + "image_path": "67d771064bf87d8d978e4b955b0013f9c40dd95992d5e290c96afffdeaa2cf2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 258, + 129, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 129, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 129, + 270 + ], + "type": "text", + "content": "are shown in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "spans": [ + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "text", + "content": "The performance of our procedural model is best understood in the context of regional performance disparities within real-world data. We contextualize the procedural model by comparing it to subsets from various regions of real-world data in Table 2 shows. Each set consists of 1,024 training images and 300 test images. Models trained on synthetic data have a narrower performance range (29.33 to " + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "inline_equation", + "content": "35.02\\mathrm{mIoU}" + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "text", + "content": ") compared to those trained on locale-specific data (25.23 to " + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "inline_equation", + "content": "61.85\\mathrm{mIoU}" + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "text", + "content": "). Notably, synthetic data can surpass real data from a different locale, as seen in the comparison of England and Egypt results. This narrower range seems to indicate synthetic data has not over-fit any particular region. The overall mIoU gap on the global dataset underscores the procedural model's limitations. The combined training set of real-world images evaluated to 53.79 mIoU, which is an upper limit on what one should expect from any model. The synthetic data yields " + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "inline_equation", + "content": "31.23\\mathrm{mIoU}" + }, + { + "bbox": [ + 45, + 270, + 290, + 545 + ], + "type": "text", + "content": " on the global test set, whereas other locales varied from 37.76-51.22 mIoU for the 'other' set. Given the global set includes holdout images from each locale, it is not surprising the mIoU's are slightly higher. This table indicates that the procedural model is of comparable quality to choosing a single locale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 552, + 206, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 206, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 206, + 563 + ], + "type": "text", + "content": "6.1. Procedural Model Variations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "We assess the impact of synthetic dataset variations on a segmentation model by comparing against our baseline model. For each variation we render 2,048 training examples and use the same geometry, lighting, and rendering settings as the baseline unless specified. For each variation, we render 2,048 training examples. Empirically, we find that the performance of synthetic data levels off at this number of examples (see Appendix Fig. 7) and evaluation is faster if the set is kept smaller. Evaluation takes place on a test partition of 4,906 real-world images; the mIoU of the baseline model is 32.58 and real photos have an mIoU of 58.69. To gauge each variation's importance, we either the relative" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 259, + 526, + 449 + ], + "blocks": [ + { + "bbox": [ + 304, + 259, + 526, + 449 + ], + "lines": [ + { + "bbox": [ + 304, + 259, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 526, + 449 + ], + "type": "image", + "image_path": "fb57500f476c5531018db3d308044d9afb66a5acc458761f2dfeb2d618b62986.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 472, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 548 + ], + "type": "text", + "content": "Figure 5. The effect of varying real-world dataset sizes on mIoU with (blue) and without (red) an additional 2,048 synthetic samples. The green area bounds the range in which adding real data neither harms nor improves performance; the right-most point of which has " + }, + { + "bbox": [ + 304, + 472, + 545, + 548 + ], + "type": "inline_equation", + "content": "n = 152" + }, + { + "bbox": [ + 304, + 472, + 545, + 548 + ], + "type": "text", + "content": " with an mIoU of 44.96. At larger datasets, synthetic data slightly reduces mIoU relative to only using real-world data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "content": "mIoU range as a percentage of the baseline, or report the mIoU Spearman's rank correlation " + }, + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "inline_equation", + "content": "(r_s)" + }, + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "content": ". Detailed results are in Appendix Section 7 and summarized here." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Rendering samples. We evaluated the impact of samples per pixel (spp) on render quality, noting diminishing returns beyond 256spp (Fig. 6). Render times scale from 6.4s at 1spp to 85.2s at 512spp. A strong correlation " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "(r_s = 1, n = 10)" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " exists between spp and mIoU, with a " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "68\\%" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " change in mIoU scores relative to baseline, underlining the importance of spp. In this experiment no denoising was performed, however, our baseline model used 256spp and a powerful neural denoiser." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22461" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 70, + 470, + 180 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 470, + 180 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 470, + 180 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 470, + 180 + ], + "type": "table", + "html": "
test
globalAustriaEgyptUKUSAothersynthetic
trainglobal53.7958.2056.8542.6450.1654.7431.21
Austria39.4958.2328.3634.4035.6741.3221.87
Egypt47.5138.3261.8533.0238.0149.7531.34
UK37.7640.6825.2338.5635.3937.6428.51
USA41.0842.5027.3933.3050.0839.3927.02
other51.2248.5439.4737.4441.3352.7428.39
synthetic31.2329.5329.3332.1534.1735.0262.12
", + "image_path": "46ec475b7c951130923bdc9bc27799ddb0a341838239a214df56ae05bf1a2130.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 189, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 222 + ], + "type": "text", + "content": "Table 2. mIoU for different splits of the real labeled data on the segmentation task. Trained on 1,024, tested on 300 samples. global is a mixture of all the real data; other data is from locales outside of Austria, Egypt, UK, or USA. In this experiment our synthetic network is similarly trained on 1,024 samples from our baseline synthetic model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 55, + 247, + 289, + 466 + ], + "blocks": [ + { + "bbox": [ + 55, + 247, + 289, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 289, + 466 + ], + "type": "image", + "image_path": "020084d66d4901c3debd19c4e74dc6de13d5454796ae8d052c825e84d351bef3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 485, + 288, + 519 + ], + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 519 + ], + "type": "text", + "content": "Figure 6. Top: The impact of rendering samples per pixel (spp) on segmentation task accuracy. Bottom: example renderings at different spp with zoomed section of rendering. " + }, + { + "bbox": [ + 46, + 485, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{BL} =" + }, + { + "bbox": [ + 46, + 485, + 288, + 519 + ], + "type": "text", + "content": " baseline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "type": "text", + "content": "Mixed Materials. We experimented with 9 material variations, including uniform gray, edge shaders, and random textures. Some tests used a single material for the entire scene, others assigned different materials per object. Simplifying materials, such as rendering only the albedo channel, resulted in a significant performance drop. Appendix Fig. 18 shows that any material restriction led to at least an " + }, + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "type": "text", + "content": " mIoU decrease from the baseline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Lighting. We examined 8 lighting model variations and their mIoU impact, as detailed in Appendix Fig. 19. Lighting models like albedo-only were crucial, while lighting conditions such as night-time had moderate importance. Conditions varied mIoU by " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "15.35\\%" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " from baseline. Daylight-only training decreased baseline mIoU by " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.04\\%" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 243, + 386, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 386, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 386, + 254 + ], + "type": "text", + "content": "relative to baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": "Camera. We experimented with the distribution of camera positions. These 7 variations used a simple model which sampled a camera position over a circle, of radius " + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "inline_equation", + "content": "r = \\{0\\ldots 96\\}" + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": " meters, truncated at the floor plane. The circle is positioned 5 meters from the wall, directly in front of the window. The camera's field of view is adjusted to the apparent window size. We observed limited impact on mIoU as " + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": " changes; peak task performance was at " + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "inline_equation", + "content": "r = 12m" + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "text", + "content": "Window Geometry. We ran 7 tests with varying window dimensions and shapes, including square and nonrectangular windows. The mIoU impact was minor, fluctuating by up to " + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "text", + "content": " relative to the baseline (1.86 absolute), with the best variation " + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "text", + "content": " worse relative to the baseline " + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "inline_equation", + "content": "(-0.04" + }, + { + "bbox": [ + 304, + 351, + 545, + 435 + ], + "type": "text", + "content": " absolute). Small features, though noticeable to humans, had a weak impact on model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 435, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 520 + ], + "type": "text", + "content": "Labels Modeled. In developing our procedural synthetic data generator, we prioritized labels by size, starting with wall and ending with open-window (Appendix Section 4). This enabled the assessment of mIoU at nine developmental stages (Fig. 7). Adding smaller classes later showed diminishing returns and occasionally reduced single-label accuracy (Appendix Fig. 22)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 527, + 441, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 441, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 441, + 540 + ], + "type": "text", + "content": "6.2. Additional Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "In our segmentation experiments, we employed the BEiTv2 base model [1, 30] for its proficiency in generating accurate results with minimal data. However, the trends we observed are consistent across various models, including DeepLabv3+ [10]. In the Appendix Section 5, we discuss two experiments that study the impact on the segmentation task when training (and testing) with different mixes of real and synthetic data. The first experiment in Appendix Fig. 7 explores segmentation performance when training and testing on different amounts of real or synthetic data. We observe that synthetic data saturates (stops improving with additional data), while real-world data continues to benefit from additional samples. The second experiment, in Appendix Fig. 8, illustrates the impact of different mixtures of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22462" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 45, + 67, + 283, + 322 + ], + "blocks": [ + { + "bbox": [ + 45, + 67, + 283, + 322 + ], + "lines": [ + { + "bbox": [ + 45, + 67, + 283, + 322 + ], + "spans": [ + { + "bbox": [ + 45, + 67, + 283, + 322 + ], + "type": "image", + "image_path": "bb1dbc150eac29e2e0fc62818df86a72bc6ad3b6d5ac80b3969070bd50c4f866.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 339, + 289, + 450 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 450 + ], + "type": "text", + "content": "Figure 7. Top: The progressive impact of incorporating additional labels into the procedural model, with label subsets expanding from lvl1 to lvl9. The sequence begins with 'walls' only (lvl1), adding 'window panes' (lvl2), 'wall frames' (lvl3), 'window frames' (lvl4), 'shutters' (lvl5), 'balconies' (lvl6), 'miscellaneous objects' (lvl7), 'blinds' (lvl8), and 'bars' (lvl9). The full model includes additional features beyond lvl9, such as interior dressing, open windows, and windows without glass. Bottom: Examples corresponding to each incremental level of model complexity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 488, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 550 + ], + "type": "text", + "content": "real and synthetic data, when tested on real data. Mixing synthetic and real data can be advantageous when there is little real data, but adding large amounts of synthetic data does not help task performance. We conclude that no volume of synthetic data can overcome this domain gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "An exploration of different architectures and data partitions is presented in Appendix Fig. 9 and Table 2 - we train an older convolution network without pre-training (DeepLab3+ [10]), a 'large' BEiTv2 model, an architecture which uses real data labels during training to improve performance (Label Adaptation [39]), another technique which uses unlabeled data to improve performance (MIC [20]), adjusting the colors of the training dataset (Histogram Matching [17]), and creating an 'easy' real data partition. Under these approaches, the synthetic-real domain gap persists, suggesting that network architecture and data partitions are an orthogonal research direction to improving the quality of synthetic procedural models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 421, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 421, + 86 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 421, + 86 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 98, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 547, + 277 + ], + "type": "text", + "content": "We thank our lead photographers Michaela Nömayr and Florian Rist, and engineer Prem Chedella, as well as our contributing photographers: Aleksandr Aleshkin, Angela Markoska, Artur Oliveira, Brian Benton, Chris West, Christopher Byrne, Elsayed Saafan, Florian Rist, George Ilin, Ignacio De Barrio, Jan Cuales, Kaitlyn Jackson, Kalina Mondzholovska, Kubra Ayse Guzel, Lukas Bornheim, Maria Jose Balestena, Michaela Nömayr, Mihai Alexandru Filoneanu, Mokhtari Sid Ahmed Salim, Mussa Ubapa, Nestor Angulo Caballero, Nicklaus Suarez, Peter Fountain, Prem Chedella, Samantha Martucci, Sarab-jot Singh, Scarlett Li, Serhii Malov, Simon R. A. Kelly, Stephanie Foden, Surafel Sunara, Tadiyos Shala, Susana Gomez, Vasileios Notis, Yuan Yuan, and finally labelyour-data.com for the labeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 308, + 379, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 379, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 379, + 322 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "type": "text", + "content": "We have introduced a new dataset of 75,739 photos (2.09 terapixels), of which 9,002 photos are semantically labeled, as well as a high-quality procedural model that closely approximates real-world variation, making it effective for image segmentation tasks. Together these components of WinSyn have applications in unsupervised domain adaptation (using the large amount of unlabeled data), superresolution (via the thigh resolution images), learning from RAW sources, generative modeling, 3D reconstruction or depth recovery (using the synthetic depth or 3D geometry from the procedural model), as well as learning from images containing transparent or specular materials (i.e., window panes/glass balconies are often transparent)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": "We systematically explored the effect of variations on our model on mIoU over 64 variations and 156,903 renders. The mIoU performance gap between our synthetic and real-world data is comparable to inter-country differences with largely different architectural styles. However, the difference between synthetic data and real data is still much larger than desired. This gap is not sufficiently reduced by either our work or other network architectures or dataset scale. We, therefore, believe that WinSyn provides a timely and efficient testbench from which researchers can iterate quickly to explore graphics contributions to the synthetic data problem. Our work contributes a sizable and versatile dataset that can be the basis of exciting and much-needed progress in the area of procedural graphics for synthetic data generation in machine learning. The key value of our dataset is that as a new benchmark of simulated and real images, it enables others to study the problem at an approachable level of complexity." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22463" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 158 + ], + "type": "text", + "content": "[2] Björn Barz and Joachim Denzler. Wikichurches: A fine-grained dataset of architectural styles with real-world challenges. arXiv preprint arXiv:2108.06959, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 178 + ], + "type": "text", + "content": "[3] Cycles: Open Source Production Rendering. Blender Foundation, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "text", + "content": "[4] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV (1), pages 44-57, 2008. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "text", + "content": "[5] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "type": "text", + "content": "[6] Clemens-Alexander Brust, Sven Sickert, Marcel Simon, Erik Rodner, and Joachim Denzler. Efficient convolutional patch networks for scene understanding. In CVPR Workshop on Scene Understanding (CVPR-WS), 2015. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "text", + "content": "[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, pages 1-7. vol. 2012, 2012. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 413 + ], + "type": "text", + "content": "[8] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 436 + ], + "type": "text", + "content": "[9] J Chen, R Stouffs, and F Biljecki. Hierarchical (multi-label) architectural image recognition and classification. 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 491 + ], + "type": "text", + "content": "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "text", + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 590 + ], + "type": "text", + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "text", + "content": "[13] D. Dai, H. Riemenschneider, G. Schmitt, and L. Van Gool. Example-based facade texture synthesis. In International Conference on Computer Vision (ICCV), 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 690 + ], + "type": "text", + "content": "[15] Epic Games, Inc. MetaHuman Realistic Person Creator, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking anal" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[17] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 173 + ], + "type": "text", + "content": "[18] Daniel Cabrini Hauagge and Noah Snavely. Image matching using local symmetry features. In 2012 IEEE conference on computer vision and pattern recognition, pages 206-213. IEEE, 2012. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "type": "text", + "content": "[19] Tomás Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jérí Matas. BOP challenge 2020 on 6D object localization. European Conference on Computer Vision Workshops (ECCVW), 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "text", + "content": "[20] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721-11732, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 285, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 331 + ], + "type": "text", + "content": "[21] Jian Kang, Marco Korner, Yuanyuan Wang, Hannes Taubenböck, and Xiao Xiang Zhu. Building instance classification using street view images. ISPRS journal of photogrammetry and remote sensing, 145:44-59, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "text", + "content": "[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "type": "text", + "content": "[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. International Conference on Computer Vision (ICCV) Workshops, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 454, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 487 + ], + "type": "text", + "content": "[25] F. Korč and W. Förstner. eTRIMS Image Database for interpreting images of man-made scenes. Technical Report TR-IGG-P-2009-01, 2009. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 488, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 543 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 544, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 589 + ], + "type": "text", + "content": "[27] Jose Llamas, Pedro M. Lerones, Roberto Medina, Eduardo Zalama, and Jaime Gómez-García-Bermejo. Classification of architectural heritage images using deep learning techniques. Applied Sciences, 7(10):992, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 589, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 623 + ], + "type": "text", + "content": "[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 623, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 623, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 623, + 545, + 656 + ], + "type": "text", + "content": "[29] Pascal Mueller, Peter Wonka, Simon Haegler, Andreas Ulmer, and Luc Van Gool. Procedural modeling of buildings. ACM Trans. Gr., 25(3):614-623, 2006. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 656, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 656, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 656, + 545, + 689 + ], + "type": "text", + "content": "[30] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers, 2022. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "text", + "content": "[31] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22464" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "and fast spatial matching. In IEEE Conference on Computer Vision and Pattern Recognition, 2007. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "text", + "content": "[32] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, et al. Infinite photorealistic worlds using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12630-12641, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[33] Hayko Riemenschneider, Ulrich Krispel, Wolfgang Thaller, Michael Donoser, Sven Havemann, Dieter Fellner, and Horst Bischof. Irregular lattices for complex shape grammar facade parsing. In 2012 IEEE conference on computer vision and pattern recognition, pages 1640-1647. IEEE, 2012. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 283 + ], + "type": "text", + "content": "[34] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 340 + ], + "type": "text", + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 373 + ], + "type": "text", + "content": "[36] Radim Tyleček and Radim Šára. Spatial pattern templates for recognition of objects with regular structure. In Proc. GCPR, Saarbrucken, Germany, 2013. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 430 + ], + "type": "text", + "content": "[37] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-dof pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In International Conference on Intelligent Robots and Systems (IROS), 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 463 + ], + "type": "text", + "content": "[38] Peter Wonka, Michael Wimmer, François Sillion, and William Ribarsky. Instant architecture. ACM Trans. Gr., 22 (3):669-677, 2003. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "text", + "content": "[39] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3661-3671. IEEE, 2021. 2, 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 531, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 596 + ], + "type": "text", + "content": "[40] Zhe Xu, Dacheng Tao, Ya Zhang, Junjie Wu, and Ah Chung Tsoi. Architectural style classification using multinomial latent logistic regression. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 600-615. Springer, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 598, + 287, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 598, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 48, + 598, + 287, + 651 + ], + "type": "text", + "content": "[41] K. Zhao, Y. Liu, S. Hao, S. Lu, H. Liu, and L. Zhou. Bounding boxes are all we need: Street view image classification via context encoding of detected buildings. IEEE Transactions on Geoscience and Remote Sensing, pages 1-17, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 654, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 287, + 708 + ], + "type": "text", + "content": "[42] Peihao Zhu, Wamiq Reyaz Para, Anna Frühstück, John Femiani, and Peter Wonka. Large-scale architectural asset extraction from panoramic imagery. IEEE Transactions on Visualization and Computer Graphics, 28(2):1301-1316, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22465" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_content_list.json b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60319d0871d3392b0af13d3beb424e68366e2070 --- /dev/null +++ b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_content_list.json @@ -0,0 +1,1343 @@ +[ + { + "type": "text", + "text": "Wired Perspectives: Multi-View Wire Art Embraces Generative AI", + "text_level": 1, + "bbox": [ + 145, + 130, + 823, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyu Qu $^{1}$ Lan Yang $^{2}$ Honggang Zhang $^{2}$ Tao Xiang $^{1}$ Kaiyue Pang $^{1}$ Yi-Zhe Song $^{1}$ $^{1}$ SketchX, CVSSP, University of Surrey $^{2}$ Beijing University of Posts and Telecommunications \n{z.qu, t.xiang, k.pang, y song}@surrey.ac.uk {ylan, zhhg}@bupt.edu.cn \nhttps://dreamwireart.github.io", + "bbox": [ + 104, + 161, + 866, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fa1efef4e61706506a09f2c7ae32cd85f656e799cf0c31d4db3364a9f750d52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 252, + 890, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03ff48724de5c9abbb1d9ad06deecfa28c620eefd43b90338cf0ccd8da5b1d44.jpg", + "image_caption": [ + "Figure 1. Multi-view art done generatively. We present DreamWire as the first system that takes user prompt for each view as input - either via the expressive vehicle of text or image - and produces 3D line sculptures showing distinct interpretations when viewed at different angles, i.e., multi-view wire art (MVWA). Compared to previous rule-based work, we significantly improve the quality of MVWA by utilising the flexible drawing capabilities of a universal generative prior (diffusion models or ControlNet). Notably, the \"GBE\" here pays tribute to the book \"Gödel, Escher, Bach: an Eternal Golden Braid\" [12], which discusses how systems can acquire meaningful context despite being made of \"meaningless\" elements, just like what MVWA does." + ], + "image_footnote": [], + "bbox": [ + 81, + 450, + 887, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 734, + 313, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating multi-view wire art (MVWA), a static 3D sculpture with diverse interpretations from different viewpoints, is a complex task even for skilled artists. In response, we present DreamWire, an AI system enabling everyone to craft MVWA easily. Users express their vision through text prompts or scribbles, freeing them from intricate 3D wire organisation. Our approach synergises 3D Bezier curves, Prim's algorithm, and knowledge distillation from diffusion models or their variants (e.g., ControlNet). This blend en", + "bbox": [ + 75, + 763, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ables the system to represent 3D wire art, ensuring spatial continuity and overcoming data scarcity. Extensive evaluation and analysis are conducted to shed insight on the inner workings of the proposed system, including the trade-off between connectivity and visual aesthetics.", + "bbox": [ + 496, + 736, + 895, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 819, + 632, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A great thought begins by seeing something differently, with a shift of the mind's eye.", + "bbox": [ + 496, + 838, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Albert Einstein", + "bbox": [ + 787, + 886, + 890, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6149", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is an artist in everyone, they say. Attending an art exhibition, being mesmerised by a 3D wire-art installation from Matthieu Robert-Ortis1 is what motivated this paper! As an AI practitioner, the immediate question was, \"Can I program this?\" Not to replace artists, but rather, for fun, for finding that artist within myself, and for the vision of democratising art creation for everyone!", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lighthearted as it might sound, this endeavour holds scientific value on two fronts. First, it delves into the uncharted territory of wired-art generation using current generative AI [34, 37, 39], seeking to understand the limits of these technologies in the realm of this unique artistic form. Secondly, it contributes to the ongoing dialogue by exploring the expansion of existing 2D-focused generation methods into the intricate domains of 3D and perhaps more challengingly, the extreme abstraction presented by wire art.", + "bbox": [ + 75, + 196, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view wire art (MVWA) [14] is a unique form of art that leverages wire as a flexible medium to create complex 3D objects, whereupon different viewpoints, multiple interpretable images appear – recall those 2D pictures where you move your head and see different things. This time, you are walking around a 3D installation, and upon different viewing angles, you see different 2D depictions (see Fig. 1). Being prohibitively difficult for novice users, creating MVWA is an extremely time-consuming task even for qualified artists. Apart from artistic ideation, working with reverse projection (2D to 3D), efforts have been made on physics so the installation does not collapse. Our ambition for MVWA, one that focuses on democratising its creation for everyone, is removing all said challenges but limiting its creation to just ideation (perhaps not entirely artistic, though!). That is, specifying what you want each view to look like, and bingo – the final 3D art form!", + "bbox": [ + 75, + 333, + 473, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present a system named DreamWire to do just that. All users need to generate 3D wire art is a set of text prompts (e.g., \"a portrait of Einstein\") or rough scribbles (e.g., styled writings of \"CVPR\"), each for a 2D view. Fig. 1 illustrates some examples, and for a more immersive experience, we offer fully interactive MVWA demos in our project page - please do set your eyes on them; we promise they won't be boring! However, there is a caveat: there is an upper bound (three) on how many viewpoints an MVWA object could support, largely due to the degree of conflict in the 3D wire space that a large number of views would introduce.", + "bbox": [ + 75, + 590, + 472, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computational methods for MVWA [14] or related art forms alike [22] have been attempted before but only appear as rule-based endeavours – they rely on a set of prewritten rules to construct an MVWA piece2. Much like any", + "bbox": [ + 75, + 756, + 472, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rule-based methods for vision problems (SIFT, HOG), these approaches are advantageous for their full transparency of the playbook but fall short in generalisation. This is discussed in Sec. 4.3, where existing approaches, guided by human-informed rules, can create MVWA pieces whose 2D projections align perfectly with user inputs, but they collapse when faced with slightly more complex combinations of 2D view images. Reproducing most of the results shown in Fig. 2 would therefore be a stretch because there is not yet a rule-based system that can generate arbitrary plausible 2D visual images from a text string, let alone generating MVWA on top of that.", + "bbox": [ + 500, + 90, + 893, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We face two key challenges: (i) how to represent 3D wire art while ensuring connectivity (so it does not collapse!), and (ii) how to ensure effective learning from extremely scarce MVWA training examples. For the former, we leverage 3D Bézier curves to solve a connectivity (\"wiredness\") problem that cannot be easily achieved in a naive way, such as by chaining control points (see Fig. 8). Instead, we treat each Bézier curve independently and propose a loss function to spatially constrain their degree of freedom. At each iteration, we depict the currently learned wires as a weighted undirected graph and apply Prim's algorithm [30] to derive a subset of edges (including all vertices) corresponding to a minimum spanning tree. The spatial continuity of wires is thus assured by minimising the distance between each parent and child vertex. On the latter challenge, we opt for per-instance learning and base generalisation on knowledge distillation from a powerful generative visual prior (diffusion models [34] or their variant, i.e., ControlNet [46] in this case). In unison, our system begins with a set of randomly initialised Bézier curves, which, after 2D projection and vector-to-raster conversion, are fed into diffusion models to match the user target text or image and updated via the typical score distillation sampling (SDS) [29] process.", + "bbox": [ + 498, + 273, + 893, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold: (i) empowering everyone to become a wired 3D (MVWA) artist (even if only half-decent), and scientifically, (ii) employing Bezier curves and Prim's algorithm to represent 3D wire art, and (iii) utilising a powerful generative visual prior through a designed rendering strategy to overcome data scarcity and the limitations of rule-based methods.", + "bbox": [ + 498, + 621, + 893, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 742, + 638, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vector Graphics. Scalable vector graphics (SVGs), in contrast to images composed of raster pixels, are defined by", + "bbox": [ + 500, + 773, + 893, + 803 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://leonacreo.com/sculptures-by-matthieu-robert-ortis/", + "bbox": [ + 76, + 825, + 472, + 849 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "For readers unfamiliar with the existing \"assembly manual\" for MVWA, we briefly summarise the rules here: i) back-project the 2D images to 3D via generalised cones and discrete the intersection of the camera's viewing frustums with a fixed resolution of vocalisation; ii) inevitably, some of", + "bbox": [ + 76, + 851, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "these initial voxels only represent the line image from their own source, resulting in inconsistent visual impacts on other viewpoints. To address this, optimisation of a voxel displacement problem is needed, whereby conflicting voxels are either merged into one or smoothed with neighbouring voxels as a more holistic visual entity. iii) Voxels are subjected to further manipulations, often targeting issues more delicate than inconsistency, including redundancy, complexity, quality, etc.", + "bbox": [ + 500, + 815, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6150", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/edbd5868ee7fa849337a728509d7f763793829261d43e2d8dcd61ac69c2129a0.jpg", + "image_caption": [ + "Figure 2. MVWA generated via DreamWire. The textual prompts employed predominantly include \"a head of [character]\" and \"a simple drawing of [item]\". Notably, all captions for these MVWA have been sourced from ChatGPT [26]. We prompt it to return three major subjects of interests under a given topic, e.g., three celebrated movie characters of the United States of America." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "extensible markup language (XML) covering lines, shapes, or curves. Bézier curve is one of the most pronounced SVG formats, which relies on a set of \"control points\" to define a smooth line segment. While there is no denying that computer vision has predominantly invested in understanding raster images, recent efforts witnessed several important breakthroughs in the generative modelling of SVGs, mostly for Bézier curves. BézierSketch [5] first introduced an inverse graphics approach to sketch stroke embedding that trains an encoder to embed each stroke to its best fit Bézier curve, and their subsequent work extends this idea to a more generalised case with variable-degree Bézier control. Another line of works [7, 19, 35, 44] directly utilise Bézier curves to govern the general-purpose vector graphic generation process. VectorFusion [15] employs diffusion models as transferable priors to generate high-quality abstract vector graphics from text captions. SketchDreamer [31] presents an", + "bbox": [ + 75, + 637, + 472, + 895 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "interactive method for text-driven vector sketch generation, adeptly incrementing strokes to an initial vector sketch in accordance with a user-specified text prompt. These works however only contribute to the application of 2D Bezier curves. We consider the problem of how to render 3D Bezier curves using a 2D Bezier renderer [19], which is significantly different from previous work.", + "bbox": [ + 496, + 637, + 893, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Prior. There has been a burgeoning interest in denoising diffusion probabilistic models [6, 11, 24, 40], also known as score-based generative models [41, 42], thanks to the remarkable generative prowess they have shown. Consequently, an increasing number of studies [8, 10, 20, 29, 38, 43] emerge as to how to leverage pretrained diffusion models to act as effective visual priors for generative supervision. DDPMPnP [8] introduces a partitioning of diffusion models into a base prior and a conditional constraint, enabling versatile applications in perceptual tasks", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6151", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "like conditional image generation and segmentation. DreamFusion [29] optimises NeRF parameters using an efficient, high-fidelity Score Distillation Sampling (SDS) loss, facilitated by a 2D diffusion image prior to text-to-3D synthesis. Make-A-Video [38] employs spatial-temporal modules built on 2D text-to-image diffusion models, realising text-to-video generation without the need for paired samples. We adopt the idea of applying the diffusion prior to a differentiable image parameterisation [23] (DIP) as proposed by DreamFusion, with the difference that we focus on the generation of multi-view wire art.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-View Art. Multi-View Art entails the presentation of multiple perspectives or views within a singular artwork [2, 3, 14, 17, 25, 36]. The techniques for achieving varied visual perceptions in an artwork can span a range of approaches: from altering the viewing distance [25], adjusting the viewing direction [14, 36], to changing illumination from specific directions [1, 3, 22]. The underlying factors prompting such phenomena are multifaceted, including the use of optical materials [28, 45], innovative structural design [3, 14, 22], or specialised devices [13]. We unprecedentedly introduce powerful text-to-image generation models to this problem, elevating the upper limit of creativity and simplifying and democratising the art creation process.", + "bbox": [ + 75, + 261, + 472, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 470, + 210, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Differentiable 3D MVWA rendering", + "text_level": 1, + "bbox": [ + 76, + 494, + 388, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We represent a 3D multi-view wire art $S$ as a set of individual wires $\\{s_1, \\dots, s_n\\}$ . Each individual wire employs a cubic 3D Bézier curve, which is rigorously defined by a quartet of 3D control points $\\{p_0, p_1, p_2, p_3\\}$ , detailed in Eq. 1:", + "bbox": [ + 75, + 518, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB (t) = (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 588, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t \\in [0,1]$ . A straightforward approach to render 3D Bézier curves is to consider a specific plane and calculate the projection of every point along the curve onto this plane. Given a plane $\\pi$ characterised by its normal vector $N$ , the projection of the 3D cubic Bézier curve $B(t)$ onto $\\pi$ is formulated in Eq. 2:", + "bbox": [ + 75, + 619, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB ^ {\\prime} (t) = B (t) - [ N \\cdot (B (t) - q) ] N, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 720, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q$ is an arbitrary point on $\\pi$ . However, prevalent 2D Bézier rendering techniques, such as the one discussed by [19], as well as 3D point cloud rendering tools [18, 21, 33], do not support the rendering of such 3D Bézier curves in Eq. 2. We propose an inquiry: Is it feasible to render 3D Bézier curves utilising existing 2D Bézier curves renderers? The answer is YES.", + "bbox": [ + 75, + 750, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our objective is to prove that the projection of a 3D Bézier curve onto a plane, denoted as $B'(t)$ , is equivalent to the 2D Bézier curve, whose control points are the projections", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the original control points of $B(t)$ onto the same plane, expressed as $B''(t)$ . Utilising Eq. 2, the 2D Bézier curve $B''(t)$ formed by the projection points of $p_i$ on $\\pi$ can be expressed as,", + "bbox": [ + 498, + 90, + 893, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB ^ {\\prime \\prime} (t) = (1 - t) ^ {3} p _ {0} ^ {\\prime} + 3 (1 - t) ^ {2} t p _ {1} ^ {\\prime} + 3 (1 - t) t ^ {2} p _ {2} ^ {\\prime} + t ^ {3} p _ {3} ^ {\\prime}. (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 160, + 893, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The equivalence of $B^{\\prime}(t)$ and $B''(t)$ can be systematically demonstrated by applying the principles outlined and the property of vector addition to expand and transform Eq. 3:", + "bbox": [ + 498, + 186, + 892, + 234 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} B ^ {\\prime \\prime} (t) = \\underbrace {(1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}} _ {B (t)} - \\\\ \\{N \\cdot \\underbrace {\\left[ (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3} \\right.} _ {B (t)} \\\\ - q ] \\} N = B (t) - [ N \\cdot (B (t) - q) ] N = B ^ {\\prime} (t). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 241, + 908, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This proof enables us to reframe the challenge of rendering 3D Bézier curves as essentially a 2D rendering task, anchored in the projection of 3D control points. Consequently, we are able to directly optimise the 3D wire art $S$ using a differentiable 2D Bézier curve renderer, i.e., DiffVG [19].", + "bbox": [ + 498, + 352, + 893, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. DreamWire", + "text_level": 1, + "bbox": [ + 500, + 436, + 617, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of our DreamWire is depicted in Fig. 3. Users just need to provide three distinct inputs $c = \\{c^{X}, c^{Y}, c^{Z}\\}$ , corresponding to projections from three mutually orthogonal viewpoints $\\{X, Y, Z\\}$ . The primary objective of our MVWA generation system is to produce a 3D wire art, $S$ , such that its projections onto each of these viewpoints align with the user's specified inputs.", + "bbox": [ + 498, + 459, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, we initialise a 3D wire art $S = \\{s_i\\}_{i=1}^n$ , where the control points of each wire are randomly initialised. We define the three planes of projection as $\\pi^X, \\pi^Y, \\pi^Z$ , with their corresponding normal vectors $N^X, N^Y, N^Z$ , which relate to the three user-provided viewpoints. Utilising Eqs. 3 and 4, we can determine the projection of $S$ on plane $\\pi^X$ as,", + "bbox": [ + 498, + 565, + 890, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {X} = \\left\\{\\hat {s} _ {i} ^ {X} \\right\\} _ {i = 1} ^ {n} = \\left\\{\\hat {B} _ {i} ^ {X} (t) \\right\\} _ {i = 1} ^ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 662, + 818, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{B}_i^X (t) = \\sum_{j = 0}^{3}\\binom {3}{j}(1 - t)^{3 - j}t^j\\hat{p}_i^{Xj},$ (5)", + "bbox": [ + 526, + 684, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {i} ^ {X j} = p _ {i} ^ {j} - [ N ^ {X} \\cdot (p _ {i} ^ {j} - q ^ {X}) ] N ^ {X},\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 719, + 834, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q^{X}$ is any point on plane $\\pi^X$ . We then utilise a differentiable 2D Bézier curve renderer, denoted as $\\mathcal{R}$ , to produce the rasterised projection. These projections are subsequently processed through the encoder $E_{\\phi}$ of a Latent Diffusion Model (LDM) [34], utilising the Score Distillation Sampling (SDS) loss [29] to estimate $\\mathbf{z}^{X} = E_{\\phi}(\\mathcal{R}(\\mathcal{S}^{X}))$ . During each forward diffusion timestep, we introduce random noise to the latents, $\\mathbf{z}_t^X = \\alpha_t\\mathbf{z}^X + \\sigma_t\\epsilon$ , and apply the teacher model $\\hat{\\epsilon}_{\\phi}(\\mathbf{z}_t^X; t)$ , conditioned on $c^X$ , for denoising. This process is replicated for $\\{Y, Z\\}$ . The optimisation targets all control", + "bbox": [ + 498, + 748, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6152", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/526528c312bec101df322689f8fb7383b5bc82528eac238cade803167d49ac57.jpg", + "image_caption": [ + "Figure 3. Schematic overview of DreamWire. Starting from an initial set of random 3D Bézier curves, we project these curves onto a given 2D plane and process them into normal raster images. It follows that these images are fed into a generative diffusion model and optimised towards a visual target. In addition, we use the MST algorithm to constrain the distance between curves. Here we present a MVWA sample output under the condition $\\{c^X, c^Y, c^Z\\} = \\{\\text{\"dog\", \"backpack\", \"cat\"}\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 880, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "points (collectively represented as $\\mathbf{P}$ ), and is steered by the SDS loss, as expressed in Eq. 6:", + "bbox": [ + 75, + 398, + 470, + 430 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\mathbf {P}} \\mathcal {L} _ {\\text {M u l t i - S D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {X} + \\sigma_ {t} \\epsilon ; c ^ {X}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {X}}{\\partial \\mathbf {P}} \\right] \\\\ + \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {Y} + \\sigma_ {t} \\epsilon ; c ^ {Y}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {Y}}{\\partial \\mathbf {P}} \\right] \\\\ + \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {Z} + \\sigma_ {t} \\epsilon ; c ^ {Z}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {Z}}{\\partial \\mathbf {P}} \\right]. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 440, + 468, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $w(t)$ represents the weighting function, and $t \\in [1, 2, \\dots, T]$ denotes the timestep. To afford users greater flexibility in input types, we adopt the multi-conditional diffusion model approach, as suggested by [31], utilising a ControlNet [46] to govern the diversity and guide the diffusion model generation processes. This results in a controllable variant of Eq. 7:", + "bbox": [ + 75, + 554, + 472, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {P}} \\mathcal {L} _ {\\mathrm {C S D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} + \\sigma_ {t} \\epsilon ; c, t, C\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z}}{\\partial \\mathbf {P}} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 669, + 470, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ denotes visual conditions for ControlNet, e.g., canny edges, HED boundaries, user scribbles, human poses, semantic maps, depths, etc. With the capabilities of ControlNet, users' input conditions can extend beyond text captions of visual concepts to include spatial layouts, enabling personalised customisation.", + "bbox": [ + 75, + 715, + 470, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. MVWA to Reality", + "text_level": 1, + "bbox": [ + 76, + 816, + 250, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Technical approach laid out in Sec. 3.2 only allows a digital construct of MVWA instance within the AR/VR environment. To enable a real tangible MVWA entity in real life that respects the law of physics however remains highly", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2231caf9065acfe31efee98314858f1913af9b3a609e25bf09600d7ad2540a48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 397, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2852f0bf13aa8132cbb1b3ac52a2721d07183ae27f8c246218764a533716ca8b.jpg", + "image_caption": [ + "Figure 4. Effect of the MST regularisation on a set of randomly initialised Bézier curves." + ], + "image_footnote": [], + "bbox": [ + 501, + 472, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "challenging. The wires $s_1, s_2, \\ldots, s_n$ we acquire so far are incapable of sustaining stability in suspension; they mandate methodical interconnections to cultivate a stable and supportive structure. Taking inspiration from [14], we approach this challenge by framing it as a classic minimum spanning tree (MST) problem, with the isolated wires and their spatial relationships represented as a graph. For a set of $n$ wires $\\{s_1, s_2, \\ldots, s_n\\}$ and $\\mathbf{P}$ representing all control points, we introduce $\\bar{\\mathbf{P}}$ to denote the endpoints of all wires. For any pair of wires $\\{s_i, s_j\\}$ , we calculate the Euclidean distance between their endpoints in four different ways3, electing the smallest of these as $\\mathcal{E}_{ij}$ . Through the assessment of the Euclidean distances amongst all endpoints, we proceed to construct a densely interconnected undirected graph $\\mathcal{G}$ , comprising $n$ vertices represented as $\\{s_1, s_2, \\ldots, s_n\\}$ , with the edges bearing weights equivalent to $\\mathcal{E}_{ij}$ .", + "bbox": [ + 496, + 604, + 893, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our objective is to identify a subset of the edges of $\\mathcal{G}$ that binds all the wires together, forming a cycle-free structure", + "bbox": [ + 498, + 847, + 893, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { 3 } \\mathcal { E } _ { i j } = m i n ( \\| p _ { i } ^ { 0 } - p _ { j } ^ { 0 } \\| ^ { 2 } , \\| p _ { i } ^ { 3 } - p _ { j } ^ { 0 } \\| ^ { 2 } , \\| p _ { i } ^ { 3 } - p _ { j } ^ { 3 } \\| ^ { 2 } , \\| p _ { i } ^ { 0 } - p _ { j } ^ { 3 } \\| ^ { 2 } )\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 885, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6153", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the minimum aggregate edge weight. Employing Prim's Algorithm [30], we derive the minimum spanning tree, and the associated cost is formulated as follows:", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S T}} (\\ddot {\\mathbf {P}}) = \\sum \\operatorname {P r i m} \\left(\\mathcal {E} _ {i j}\\right) \\quad i, j \\in [ 1, n ]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 146, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consequently, our final training objective is articulated as:", + "bbox": [ + 75, + 179, + 468, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {M u l t i - S D S}} (\\mathbf {P}, c) + \\lambda * \\mathcal {L} _ {\\text {M S T}} (\\ddot {\\mathbf {P}}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 208, + 468, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\lambda$ functions as a balancing factor between aesthetic appeal and structural realism. In Fig. 4, we present a qualitative depiction of the impact elicited by the MST regularisation, effectively demonstrating how wires, initially scattered, progressively coalesce into a stable and integrated structure.", + "bbox": [ + 75, + 233, + 470, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 321, + 209, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 347, + 174, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. Building upon the methodologies employed in [15] and [31], we initiate each 3D Bezier curve of MVWA comprising 5 segments, maintaining a constant width and adopting a uniform black colour. Prior to inputting the 2D projection of the MVWA into the diffusion model, we employ random affine augmentations (RandomPerspective and RandomResizedCrop) to refine the projection's quality and to reinforce the optimisation process against the potential adversarial examples. To optimise the MVWA, we employ the Adam optimiser [16] across 2000 iterations, setting the learning rate to 1. Within our configuration, we adopt a guidance scale equal to 100. All experiments are executed on an NVIDIA A100 GPU.", + "bbox": [ + 75, + 375, + 470, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data preparation. For the visual control setting, to ensure a fair comparison, the input sets employed are identical to that of the baseline methods [14, 22]. For the text control setting, a selection of 96 daily item categories was randomly drawn from the QuickDraw dataset [9]. Each category name was then inserted into a standard template as \"a simple drawing of [item]\" and these were subsequently randomised into 32 distinct input sets.", + "bbox": [ + 75, + 575, + 468, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Within the text control setting, we utilise the CLIP [32] score and R-Precision [27] as metrics to assess the similarity between the input text condition and 2D rasterised projection of the synthesised MVWA. For visual control, DINO [4] is employed to quantify the similarity between the 2D rasterised projection of the generated MVWA and the visual input provided by the user.", + "bbox": [ + 75, + 700, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baseline comparison", + "text_level": 1, + "bbox": [ + 76, + 816, + 272, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared our approach against two state-of-the-art (SOTA) methods grounded in traditional graphics algorithms. ShadowArt [22] introduces a novel geometric optimisation method that automatically finds a consistent shadow hull", + "bbox": [ + 75, + 840, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by deforming the input images. MVWA [14] starts with reconstructing a discrete visual hull through intersecting generalised cones formed by back-projecting the given 2D image to 3D space and integrates the isolated components into a connected visual hull via a 3D path-finding method. Our-V and Our-V-λ are calibrated to align with the settings of these two methods, processing three user-specified line drawings alongside corresponding viewpoints as inputs. Here, λ signifies the incorporation of MST regularisation.", + "bbox": [ + 496, + 90, + 893, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The synthesised 3D wire arts, along with their corresponding 2D projections, are shown in Fig. 5. We can see that, the voxels yielded by ShadowArt [22] has a multitude of transformed components, attributable to the markedly inconsistent nature of the input line drawings, resulting in severely distorted 2D projections. Conversely, the MVWA [14] aspires to integrate the isolated components into a connected visual hull via a 3D path-finding method, inevitably incorporating numerous extraneous lines, thereby compromising the projected visuals' fidelity. In contrast, Our-V and Our-V-λ are predicated upon the optimisation of a set of 3D Bézier curves, complemented by the utilisation of MST loss to emulate the impression of a singular line, a technique distinctly divergent from the conventional \"Voxel Hall Carving\" employed by preceding approaches. Our generated results are slightly inferior to the results of traditional methods in this setting as our approach cannot solve the input conflict problem very well. However, given only text as conditions, our method can generate projections with much higher quality (refer Fig. 6). In addition, our results possess a more streamlined structural simplicity within the 3D space compared to traditional methods.", + "bbox": [ + 496, + 227, + 893, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main results", + "text_level": 1, + "bbox": [ + 500, + 568, + 633, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As demonstrated in Sec. 4.2, we have previously illustrated the capabilities of our method under visual control. In this section, we delineate the unique advantage of our approach over the baselines: our capacity to generate MVWA in response to textual input or a hybrid of text and visual inputs. This flexibility significantly diminishes the user's burden in resolving conflicts inherent in visual controls.", + "bbox": [ + 496, + 590, + 893, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation. In addition to the examples shown in Fig. 1, where we show the MVWA generated by", + "bbox": [ + 496, + 700, + 892, + 732 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/95cbac2a1253f6fd4a2ac79ca823e39e244f82e3597115fb0b4a160d64e9a0b2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDINO-V2-BaseDINO-V2-GiantCLIP Score
ShadowArt [22]79.6283.8234.37
MVWA [14]73.6878.8634.81
Ours-V69.0872.7634.55
Ours-V-λ66.9971.3332.75
Ours-T--37.21
Ours-T-λ--36.52
", + "bbox": [ + 500, + 744, + 890, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. DINO similarity (\\%) between the target sketches and the projection results and CLIP similarity (\\%) between the captions and the projection results generated by different methods.", + "bbox": [ + 496, + 854, + 892, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6154", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f84ff31096cb8855458b9aa1d0173a6971c66512477c9d054c5e2122415d7a7c.jpg", + "image_caption": [ + "Figure 5. Comparison with existing multi-view wire art synthesis methods. The user-specified visual controls are highlighted with red lines." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/544701318ab4c1c00a70328e0edf56e61f7b869b215021a6a1ee1a837b3ecae4.jpg", + "image_caption": [ + "Figure 6. Additional instances of MVWA generated by our proposed DreamWire with different random seeds. The text conditions for the MVWA on the left are defined as $\\{c^X, c^Y, c^Z\\} = \\{\\text{\"a side view of an angel\", \"a front view of an angel\", \"Christmas bells\"}\\}$ . Similarly, for the MVWA on the right are: $\\{c^X, c^Y, c^Z\\} = \\{\\text{\"Hillary\", \"Trump\", \"Obama\"}\\}$ . The red lines indicate the additional connections that need to be added in order to form all the curves together as a whole. Compared to Our-T, Our-T- $\\lambda$ ensures that the red lines are as short as possible while maintaining visual aesthetics." + ], + "image_footnote": [], + "bbox": [ + 76, + 494, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6155", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b7026b39ee3c7e0bc7bd178cd5a9e7808ec58eafd566a4943400083dcebd641.jpg", + "image_caption": [ + "Figure 7. The impact of $\\lambda$ on normalised evaluation metrics. Metrics are normalised to the [0, 1] interval for clarity. The original scale for MST budget spans [0, 8], for CLIP Score it is [36, 38], and for CLIP R-Prec it lies between [67, 85]." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 446, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our method under text and hybrid control, we further showcase massive cases in Fig. 2 and Fig. 6. For each case, we present two variations created by Our-T and Our-T-λ. With MST regularisation, it becomes apparent that the resulting MVWA bears greater resemblance to coherent single-line 3D sculptures as opposed to an assemblage of numerous discrete visual hulls.", + "bbox": [ + 75, + 316, + 470, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative evaluation. The results are presented in Tab. 1. Our method does not achieve the highest DINO scores on several DINO variants. This is because our model fits the target sketches by continuously optimising the parameters of Bézier curves, compared to the search process of traditional methods, our optimisation method may suffer from underfitting and overfitting in different regions. Thus, given the layout constraints for various viewpoints, the traditional methods [14, 22] almost reach the best performance, and our method is still some distance away from them. However, given only the text conditions without layout constraints, our method enables high-quality fit to the target concept in numerous ways, omitting the step of the artist to elaborate the target projections.", + "bbox": [ + 75, + 426, + 470, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on $\\lambda$ . Fig. 7 illustrates the influence of the $\\lambda$ coefficient within $\\mathcal{L}_{\\mathrm{MST}}$ and CLIP metrics. The increase of $\\lambda$ drives the generated MVWA towards an optimisation that favours a one-line wire art, leading to a substantial reduction in the wire connectivity budget. However, this also results in a notable decrease in both the CLIP-score and R-Precision, suggesting an increased deviation between the user input and the 2D projection of the synthesised MVWA. Taking into account both aesthetic appeal and manufacturability, we ultimately set the hyperparameter $\\lambda$ to 50.", + "bbox": [ + 75, + 642, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. One Line vs. MST Regularisation", + "text_level": 1, + "bbox": [ + 76, + 801, + 372, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our endeavour to enhance the interconnectedness of generated 3D Bezier curves thereby more faithfully emulating a single, continuous curve, we have instituted a novel loss function, denoted as $\\mathcal{L}_{\\mathrm{MST}}$ . Intuitively, an alternative strategy could entail initialising the 3D wire art structure as a singu", + "bbox": [ + 75, + 825, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/802a37f2575bb8581770a642799e0aa4fd42e9342d06b9f036c50e3fc8615b53.jpg", + "image_caption": [ + "Figure 8. Generation of \"a simple drawing of a bicycle\" starts with a single curve containing 150 segments (top) and 30 curves containing 5 segments (bottom) in the same random seed." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lar Bézier curve comprised of a substantially high segment count. Fig. 8 (top) shows the generation result of \"a simple drawing of a bicycle\" when the input is a single Bézier curve with 150 segments. For a clearer presentation, only a 2D Bézier curve is used. It can be noted that a curve comprising numerous segments may not be able to update their positions effectively. This problem may lead to substantial segment overlap, thereby injecting redundancy and detracting from the succinctness of the ultimate generated form. Therefore, we utilise $\\mathcal{L}_{\\mathrm{MST}}$ instead of the one-line setting to ensure that the wire art our model generates is aesthetically appealing and realistically producible.", + "bbox": [ + 496, + 273, + 893, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 470, + 612, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is worth noting that [44] emphasises the optimisation process is susceptible to the initialisation of Bézier curves. Therefore, the number, width and location of strokes and what text prompt to give each viewpoint may all be issues for artists to consider in future applications. In addition, compared to voxels and polylines, Bézier curves have fewer parameters, making them more advantageous for optimisation tasks. However, they may not fit targets well when the visual objectives contain a large number of non-smooth polylines. We anticipate the development of advanced line representation techniques in future research, which will enhance the fidelity of visual representations of MVWA.", + "bbox": [ + 496, + 497, + 893, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 693, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is a pioneering venture into the fusion of AI and art, making strides by enabling AI to easily generate 3D multi-view wire art. It does so with just brief text prompts or spontaneous scribbles. Beyond its artistic impact, our work dives into scientific challenges related to abstraction and 3D representation in the generative AI community. The core of our approach involves refining 3D Bezier curves using diffusion models and a carefully designed rendering strategy. The ultimate goal is to make this distinct art form accessible to all, offering a platform for artists, designers, and enthusiasts to effortlessly bring their imaginative wire sculptures to life.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6156", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marc Alexa and Wojciech Matusik. Reliefs as images. ACM Transactions on Graphics, 2010. 4", + "[2] Ilya Baran, Philipp Keller, Derek Bradley, Stelian Coros, Wojciech Jarosz, Derek Nowrouzezahrai, and Markus Gross. Manufacturing layered attenuators for multiple prescribed shadow images. Computer Graphics Forum, 2012. 4", + "[3] Amit Bermano, Ilya Baran, Marc Alexa, and Wojciech Matusk. Shadowpix: Multiple images from self shadowing. Computer Graphics Forum, 2012. 4", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6", + "[5] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3", + "[6] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 3", + "[7] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 3", + "[8] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 3", + "[9] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2018. 6", + "[10] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 3", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[12] Douglas Hofstadter. Gödel, escher, bach, 1979. https://en.wikipedia.org/wiki/Gdel, _Escher, _Bach.1", + "[13] S Vahab Hosseini, Usman R Alim, A Mahdavi-Amiri, et al. Portal: design and fabrication of incidence-driven screens. In SMI, 2020. 4", + "[14] Kai-Wen Hsiao, Jia-Bin Huang, and Hung-Kuo Chu. Multiview wire art. ACM Transactions on Graphics, 2018. 2, 4, 5, 6, 8", + "[15] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In CVPR, 2023. 3, 6", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[17] Ying-Miao Kuo, Hung-Kuo Chu, Ming-Te Chi, Ruen-Rone Lee, and Tong-Yee Lee. Generating ambiguous figure-ground images. IEEE transactions on visualization and computer graphics, 2016. 4", + "[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 2020. 4", + "[19] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for" + ], + "bbox": [ + 78, + 114, + 470, + 895 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "editing and learning. ACM Transactions on Graphics, 2020. 3, 4", + "[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3", + "[21] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 4", + "[22] N. J. Mitra and M. Pauly. Shadow art. ACM Transactions on Graphics, 2009. 2, 4, 6, 8", + "[23] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 2018. 4", + "[24] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Dreamcreature: Crafting photorealistic virtual creatures from imagination. arXiv:2311.15477, 2023. 3", + "[25] Aude Oliva, Antonio Torralba, and Philippe G Schyns. Hybrid images. ACM Transactions on Graphics, 2006. 4", + "[26] OpenAI. Chatgpt: A large language model, 2023. https://openai.com/research/chatgpt.3", + "[27] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In NeurIPS, 2021. 6", + "[28] Maxine Perroni-Scharf and Szymon Rusinkiewicz. Constructing printable surfaces with view-dependent appearance. arXiv:2306.07449, 2023. 4", + "[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2, 3, 4", + "[30] Robert Clay Prim. Shortest connection networks and some generalizations. The Bell System Technical Journal, 1957. 2, 6", + "[31] Zhiyu Qu, Tao Xiang, and Yi-Zhe Song. Sketchdreamer: Interactive text-augmented creative sketch ideation. In BMVC, 2023. 3, 5, 6", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[33] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4", + "[35] Peter Schaldenbrand, Zhixuan Liu, and Jean Oh. Styleclipdraw: Coupling content and style in text-to-drawing translation. arXiv:2202.12362, 2022. 3", + "[36] Guy Sela and Gershon Elber. Generation of view dependent models using free form deformation. The Visual Computer, 2007. 4", + "[37] Zhan Shi, Xu Zhou, Xipeng Qiu, and Xiaodan Zhu. Improving image captioning with better use of captions. In ACL, 2020. 2", + "[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6157", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "without text-video data. In ICLR, 2023. 3, 4", + "[39] Luke Skywalker. Midjourney, 2023. https://en.wikipedia.org/wiki/Midjourney.2", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[41] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3", + "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3", + "[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 3", + "[44] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics, 2022. 3, 8", + "[45] Jiani Zeng, Honghao Deng, Yunyi Zhu, Michael Wessely, Axel Kilian, and Stefanie Mueller. Lenticular objects: 3d printed objects with lenticular lens surfaces that can change their appearance depending on the viewpoint. In UIST, 2021. 4", + "[46] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5" + ], + "bbox": [ + 78, + 90, + 470, + 481 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6158", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_model.json b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_model.json new file mode 100644 index 0000000000000000000000000000000000000000..539bd0e6076eb5ae37c317c08d99051762260854 --- /dev/null +++ b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_model.json @@ -0,0 +1,1870 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.824, + 0.152 + ], + "angle": 0, + "content": "Wired Perspectives: Multi-View Wire Art Embraces Generative AI" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.162, + 0.867, + 0.245 + ], + "angle": 0, + "content": "Zhiyu Qu\\(^{1}\\) Lan Yang\\(^{2}\\) Honggang Zhang\\(^{2}\\) Tao Xiang\\(^{1}\\) Kaiyue Pang\\(^{1}\\) Yi-Zhe Song\\(^{1}\\) \n\\(^{1}\\)SketchX, CVSSP, University of Surrey \\(^{2}\\)Beijing University of Posts and Telecommunications \n{z.qu, t.xiang, k.pang, y song}@surrey.ac.uk {ylan, zhhg}@bupt.edu.cn \nhttps://dreamwireart.github.io" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.253, + 0.891, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.451, + 0.888, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.647, + 0.895, + 0.732 + ], + "angle": 0, + "content": "Figure 1. Multi-view art done generatively. We present DreamWire as the first system that takes user prompt for each view as input - either via the expressive vehicle of text or image - and produces 3D line sculptures showing distinct interpretations when viewed at different angles, i.e., multi-view wire art (MVWA). Compared to previous rule-based work, we significantly improve the quality of MVWA by utilising the flexible drawing capabilities of a universal generative prior (diffusion models or ControlNet). Notably, the \"GBE\" here pays tribute to the book \"Gödel, Escher, Bach: an Eternal Golden Braid\" [12], which discusses how systems can acquire meaningful context despite being made of \"meaningless\" elements, just like what MVWA does." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.736, + 0.314, + 0.751 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Creating multi-view wire art (MVWA), a static 3D sculpture with diverse interpretations from different viewpoints, is a complex task even for skilled artists. In response, we present DreamWire, an AI system enabling everyone to craft MVWA easily. Users express their vision through text prompts or scribbles, freeing them from intricate 3D wire organisation. Our approach synergises 3D Bezier curves, Prim's algorithm, and knowledge distillation from diffusion models or their variants (e.g., ControlNet). This blend en" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.737, + 0.896, + 0.813 + ], + "angle": 0, + "content": "ables the system to represent 3D wire art, ensuring spatial continuity and overcoming data scarcity. Extensive evaluation and analysis are conducted to shed insight on the inner workings of the proposed system, including the trade-off between connectivity and visual aesthetics." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.82, + 0.633, + 0.835 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.894, + 0.87 + ], + "angle": 0, + "content": "A great thought begins by seeing something differently, with a shift of the mind's eye." + }, + { + "type": "text", + "bbox": [ + 0.789, + 0.887, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Albert Einstein" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6149" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "There is an artist in everyone, they say. Attending an art exhibition, being mesmerised by a 3D wire-art installation from Matthieu Robert-Ortis1 is what motivated this paper! As an AI practitioner, the immediate question was, \"Can I program this?\" Not to replace artists, but rather, for fun, for finding that artist within myself, and for the vision of democratising art creation for everyone!" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.473, + 0.333 + ], + "angle": 0, + "content": "Lighthearted as it might sound, this endeavour holds scientific value on two fronts. First, it delves into the uncharted territory of wired-art generation using current generative AI [34, 37, 39], seeking to understand the limits of these technologies in the realm of this unique artistic form. Secondly, it contributes to the ongoing dialogue by exploring the expansion of existing 2D-focused generation methods into the intricate domains of 3D and perhaps more challengingly, the extreme abstraction presented by wire art." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.474, + 0.59 + ], + "angle": 0, + "content": "Multi-view wire art (MVWA) [14] is a unique form of art that leverages wire as a flexible medium to create complex 3D objects, whereupon different viewpoints, multiple interpretable images appear – recall those 2D pictures where you move your head and see different things. This time, you are walking around a 3D installation, and upon different viewing angles, you see different 2D depictions (see Fig. 1). Being prohibitively difficult for novice users, creating MVWA is an extremely time-consuming task even for qualified artists. Apart from artistic ideation, working with reverse projection (2D to 3D), efforts have been made on physics so the installation does not collapse. Our ambition for MVWA, one that focuses on democratising its creation for everyone, is removing all said challenges but limiting its creation to just ideation (perhaps not entirely artistic, though!). That is, specifying what you want each view to look like, and bingo – the final 3D art form!" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.473, + 0.756 + ], + "angle": 0, + "content": "We present a system named DreamWire to do just that. All users need to generate 3D wire art is a set of text prompts (e.g., \"a portrait of Einstein\") or rough scribbles (e.g., styled writings of \"CVPR\"), each for a 2D view. Fig. 1 illustrates some examples, and for a more immersive experience, we offer fully interactive MVWA demos in our project page - please do set your eyes on them; we promise they won't be boring! However, there is a caveat: there is an upper bound (three) on how many viewpoints an MVWA object could support, largely due to the degree of conflict in the 3D wire space that a large number of views would introduce." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.473, + 0.819 + ], + "angle": 0, + "content": "Computational methods for MVWA [14] or related art forms alike [22] have been attempted before but only appear as rule-based endeavours – they rely on a set of prewritten rules to construct an MVWA piece2. Much like any" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.272 + ], + "angle": 0, + "content": "rule-based methods for vision problems (SIFT, HOG), these approaches are advantageous for their full transparency of the playbook but fall short in generalisation. This is discussed in Sec. 4.3, where existing approaches, guided by human-informed rules, can create MVWA pieces whose 2D projections align perfectly with user inputs, but they collapse when faced with slightly more complex combinations of 2D view images. Reproducing most of the results shown in Fig. 2 would therefore be a stretch because there is not yet a rule-based system that can generate arbitrary plausible 2D visual images from a text string, let alone generating MVWA on top of that." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.274, + 0.895, + 0.621 + ], + "angle": 0, + "content": "We face two key challenges: (i) how to represent 3D wire art while ensuring connectivity (so it does not collapse!), and (ii) how to ensure effective learning from extremely scarce MVWA training examples. For the former, we leverage 3D Bézier curves to solve a connectivity (\"wiredness\") problem that cannot be easily achieved in a naive way, such as by chaining control points (see Fig. 8). Instead, we treat each Bézier curve independently and propose a loss function to spatially constrain their degree of freedom. At each iteration, we depict the currently learned wires as a weighted undirected graph and apply Prim's algorithm [30] to derive a subset of edges (including all vertices) corresponding to a minimum spanning tree. The spatial continuity of wires is thus assured by minimising the distance between each parent and child vertex. On the latter challenge, we opt for per-instance learning and base generalisation on knowledge distillation from a powerful generative visual prior (diffusion models [34] or their variant, i.e., ControlNet [46] in this case). In unison, our system begins with a set of randomly initialised Bézier curves, which, after 2D projection and vector-to-raster conversion, are fed into diffusion models to match the user target text or image and updated via the typical score distillation sampling (SDS) [29] process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.622, + 0.895, + 0.727 + ], + "angle": 0, + "content": "In summary, our contributions are threefold: (i) empowering everyone to become a wired 3D (MVWA) artist (even if only half-decent), and scientifically, (ii) employing Bezier curves and Prim's algorithm to represent 3D wire art, and (iii) utilising a powerful generative visual prior through a designed rendering strategy to overcome data scarcity and the limitations of rule-based methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.743, + 0.64, + 0.759 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.774, + 0.894, + 0.804 + ], + "angle": 0, + "content": "Vector Graphics. Scalable vector graphics (SVGs), in contrast to images composed of raster pixels, are defined by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.826, + 0.473, + 0.851 + ], + "angle": 0, + "content": "1https://leonacreo.com/sculptures-by-matthieu-robert-ortis/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.852, + 0.472, + 0.9 + ], + "angle": 0, + "content": "For readers unfamiliar with the existing \"assembly manual\" for MVWA, we briefly summarise the rules here: i) back-project the 2D images to 3D via generalised cones and discrete the intersection of the camera's viewing frustums with a fixed resolution of vocalisation; ii) inevitably, some of" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.826, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.816, + 0.895, + 0.901 + ], + "angle": 0, + "content": "these initial voxels only represent the line image from their own source, resulting in inconsistent visual impacts on other viewpoints. To address this, optimisation of a voxel displacement problem is needed, whereby conflicting voxels are either merged into one or smoothed with neighbouring voxels as a more holistic visual entity. iii) Voxels are subjected to further manipulations, often targeting issues more delicate than inconsistency, including redundancy, complexity, quality, etc." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6150" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.891, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.576, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Figure 2. MVWA generated via DreamWire. The textual prompts employed predominantly include \"a head of [character]\" and \"a simple drawing of [item]\". Notably, all captions for these MVWA have been sourced from ChatGPT [26]. We prompt it to return three major subjects of interests under a given topic, e.g., three celebrated movie characters of the United States of America." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.473, + 0.896 + ], + "angle": 0, + "content": "extensible markup language (XML) covering lines, shapes, or curves. Bézier curve is one of the most pronounced SVG formats, which relies on a set of \"control points\" to define a smooth line segment. While there is no denying that computer vision has predominantly invested in understanding raster images, recent efforts witnessed several important breakthroughs in the generative modelling of SVGs, mostly for Bézier curves. BézierSketch [5] first introduced an inverse graphics approach to sketch stroke embedding that trains an encoder to embed each stroke to its best fit Bézier curve, and their subsequent work extends this idea to a more generalised case with variable-degree Bézier control. Another line of works [7, 19, 35, 44] directly utilise Bézier curves to govern the general-purpose vector graphic generation process. VectorFusion [15] employs diffusion models as transferable priors to generate high-quality abstract vector graphics from text captions. SketchDreamer [31] presents an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.895, + 0.744 + ], + "angle": 0, + "content": "interactive method for text-driven vector sketch generation, adeptly incrementing strokes to an initial vector sketch in accordance with a user-specified text prompt. These works however only contribute to the application of 2D Bezier curves. We consider the problem of how to render 3D Bezier curves using a 2D Bezier renderer [19], which is significantly different from previous work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Diffusion Prior. There has been a burgeoning interest in denoising diffusion probabilistic models [6, 11, 24, 40], also known as score-based generative models [41, 42], thanks to the remarkable generative prowess they have shown. Consequently, an increasing number of studies [8, 10, 20, 29, 38, 43] emerge as to how to leverage pretrained diffusion models to act as effective visual priors for generative supervision. DDPMPnP [8] introduces a partitioning of diffusion models into a base prior and a conditional constraint, enabling versatile applications in perceptual tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "6151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "like conditional image generation and segmentation. DreamFusion [29] optimises NeRF parameters using an efficient, high-fidelity Score Distillation Sampling (SDS) loss, facilitated by a 2D diffusion image prior to text-to-3D synthesis. Make-A-Video [38] employs spatial-temporal modules built on 2D text-to-image diffusion models, realising text-to-video generation without the need for paired samples. We adopt the idea of applying the diffusion prior to a differentiable image parameterisation [23] (DIP) as proposed by DreamFusion, with the difference that we focus on the generation of multi-view wire art." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.46 + ], + "angle": 0, + "content": "Multi-View Art. Multi-View Art entails the presentation of multiple perspectives or views within a singular artwork [2, 3, 14, 17, 25, 36]. The techniques for achieving varied visual perceptions in an artwork can span a range of approaches: from altering the viewing distance [25], adjusting the viewing direction [14, 36], to changing illumination from specific directions [1, 3, 22]. The underlying factors prompting such phenomena are multifaceted, including the use of optical materials [28, 45], innovative structural design [3, 14, 22], or specialised devices [13]. We unprecedentedly introduce powerful text-to-image generation models to this problem, elevating the upper limit of creativity and simplifying and democratising the art creation process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.471, + 0.212, + 0.489 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.39, + 0.513 + ], + "angle": 0, + "content": "3.1. Differentiable 3D MVWA rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.581 + ], + "angle": 0, + "content": "We represent a 3D multi-view wire art \\( S \\) as a set of individual wires \\( \\{s_1, \\dots, s_n\\} \\). Each individual wire employs a cubic 3D Bézier curve, which is rigorously defined by a quartet of 3D control points \\( \\{p_0, p_1, p_2, p_3\\} \\), detailed in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.621 + ], + "angle": 0, + "content": "\\[\nB (t) = (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.713 + ], + "angle": 0, + "content": "where \\( t \\in [0,1] \\). A straightforward approach to render 3D Bézier curves is to consider a specific plane and calculate the projection of every point along the curve onto this plane. Given a plane \\( \\pi \\) characterised by its normal vector \\( N \\), the projection of the 3D cubic Bézier curve \\( B(t) \\) onto \\( \\pi \\) is formulated in Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.722, + 0.47, + 0.74 + ], + "angle": 0, + "content": "\\[\nB ^ {\\prime} (t) = B (t) - [ N \\cdot (B (t) - q) ] N, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.854 + ], + "angle": 0, + "content": "where \\( q \\) is an arbitrary point on \\( \\pi \\). However, prevalent 2D Bézier rendering techniques, such as the one discussed by [19], as well as 3D point cloud rendering tools [18, 21, 33], do not support the rendering of such 3D Bézier curves in Eq. 2. We propose an inquiry: Is it feasible to render 3D Bézier curves utilising existing 2D Bézier curves renderers? The answer is YES." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our objective is to prove that the projection of a 3D Bézier curve onto a plane, denoted as \\(B'(t)\\), is equivalent to the 2D Bézier curve, whose control points are the projections" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "of the original control points of \\( B(t) \\) onto the same plane, expressed as \\( B''(t) \\). Utilising Eq. 2, the 2D Bézier curve \\( B''(t) \\) formed by the projection points of \\( p_i \\) on \\( \\pi \\) can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.161, + 0.894, + 0.18 + ], + "angle": 0, + "content": "\\[\nB ^ {\\prime \\prime} (t) = (1 - t) ^ {3} p _ {0} ^ {\\prime} + 3 (1 - t) ^ {2} t p _ {1} ^ {\\prime} + 3 (1 - t) t ^ {2} p _ {2} ^ {\\prime} + t ^ {3} p _ {3} ^ {\\prime}. (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.188, + 0.893, + 0.235 + ], + "angle": 0, + "content": "The equivalence of \\( B^{\\prime}(t) \\) and \\( B''(t) \\) can be systematically demonstrated by applying the principles outlined and the property of vector addition to expand and transform Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.242, + 0.91, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} B ^ {\\prime \\prime} (t) = \\underbrace {(1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}} _ {B (t)} - \\\\ \\{N \\cdot \\underbrace {\\left[ (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3} \\right.} _ {B (t)} \\\\ - q ] \\} N = B (t) - [ N \\cdot (B (t) - q) ] N = B ^ {\\prime} (t). \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.353, + 0.894, + 0.428 + ], + "angle": 0, + "content": "This proof enables us to reframe the challenge of rendering 3D Bézier curves as essentially a 2D rendering task, anchored in the projection of 3D control points. Consequently, we are able to directly optimise the 3D wire art \\(S\\) using a differentiable 2D Bézier curve renderer, i.e., DiffVG [19]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.619, + 0.451 + ], + "angle": 0, + "content": "3.2. DreamWire" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.892, + 0.565 + ], + "angle": 0, + "content": "The overall pipeline of our DreamWire is depicted in Fig. 3. Users just need to provide three distinct inputs \\( c = \\{c^{X}, c^{Y}, c^{Z}\\} \\), corresponding to projections from three mutually orthogonal viewpoints \\( \\{X, Y, Z\\} \\). The primary objective of our MVWA generation system is to produce a 3D wire art, \\( S \\), such that its projections onto each of these viewpoints align with the user's specified inputs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Initially, we initialise a 3D wire art \\(S = \\{s_i\\}_{i=1}^n\\), where the control points of each wire are randomly initialised. We define the three planes of projection as \\(\\pi^X, \\pi^Y, \\pi^Z\\), with their corresponding normal vectors \\(N^X, N^Y, N^Z\\), which relate to the three user-provided viewpoints. Utilising Eqs. 3 and 4, we can determine the projection of \\(S\\) on plane \\(\\pi^X\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.664, + 0.819, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {X} = \\left\\{\\hat {s} _ {i} ^ {X} \\right\\} _ {i = 1} ^ {n} = \\left\\{\\hat {B} _ {i} ^ {X} (t) \\right\\} _ {i = 1} ^ {n},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.685, + 0.892, + 0.718 + ], + "angle": 0, + "content": "where \\(\\hat{B}_i^X (t) = \\sum_{j = 0}^{3}\\binom {3}{j}(1 - t)^{3 - j}t^j\\hat{p}_i^{Xj},\\) (5)" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.72, + 0.836, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {i} ^ {X j} = p _ {i} ^ {j} - [ N ^ {X} \\cdot (p _ {i} ^ {j} - q ^ {X}) ] N ^ {X},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.749, + 0.894, + 0.901 + ], + "angle": 0, + "content": "where \\(q^{X}\\) is any point on plane \\(\\pi^X\\). We then utilise a differentiable 2D Bézier curve renderer, denoted as \\(\\mathcal{R}\\), to produce the rasterised projection. These projections are subsequently processed through the encoder \\(E_{\\phi}\\) of a Latent Diffusion Model (LDM) [34], utilising the Score Distillation Sampling (SDS) loss [29] to estimate \\(\\mathbf{z}^{X} = E_{\\phi}(\\mathcal{R}(\\mathcal{S}^{X}))\\). During each forward diffusion timestep, we introduce random noise to the latents, \\(\\mathbf{z}_t^X = \\alpha_t\\mathbf{z}^X + \\sigma_t\\epsilon\\), and apply the teacher model \\(\\hat{\\epsilon}_{\\phi}(\\mathbf{z}_t^X; t)\\), conditioned on \\(c^X\\), for denoising. This process is replicated for \\(\\{Y, Z\\}\\). The optimisation targets all control" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.087, + 0.881, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.896, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Schematic overview of DreamWire. Starting from an initial set of random 3D Bézier curves, we project these curves onto a given 2D plane and process them into normal raster images. It follows that these images are fed into a generative diffusion model and optimised towards a visual target. In addition, we use the MST algorithm to constrain the distance between curves. Here we present a MVWA sample output under the condition \\(\\{c^X, c^Y, c^Z\\} = \\{\\text{\"dog\", \"backpack\", \"cat\"}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.431 + ], + "angle": 0, + "content": "points (collectively represented as \\(\\mathbf{P}\\)), and is steered by the SDS loss, as expressed in Eq. 6:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\mathbf {P}} \\mathcal {L} _ {\\text {M u l t i - S D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {X} + \\sigma_ {t} \\epsilon ; c ^ {X}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {X}}{\\partial \\mathbf {P}} \\right] \\\\ + \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {Y} + \\sigma_ {t} \\epsilon ; c ^ {Y}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {Y}}{\\partial \\mathbf {P}} \\right] \\\\ + \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {Z} + \\sigma_ {t} \\epsilon ; c ^ {Z}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {Z}}{\\partial \\mathbf {P}} \\right]. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.473, + 0.661 + ], + "angle": 0, + "content": "Here, \\( w(t) \\) represents the weighting function, and \\( t \\in [1, 2, \\dots, T] \\) denotes the timestep. To afford users greater flexibility in input types, we adopt the multi-conditional diffusion model approach, as suggested by [31], utilising a ControlNet [46] to govern the diversity and guide the diffusion model generation processes. This results in a controllable variant of Eq. 7:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.67, + 0.472, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {P}} \\mathcal {L} _ {\\mathrm {C S D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} + \\sigma_ {t} \\epsilon ; c, t, C\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z}}{\\partial \\mathbf {P}} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.472, + 0.807 + ], + "angle": 0, + "content": "where \\( C \\) denotes visual conditions for ControlNet, e.g., canny edges, HED boundaries, user scribbles, human poses, semantic maps, depths, etc. With the capabilities of ControlNet, users' input conditions can extend beyond text captions of visual concepts to include spatial layouts, enabling personalised customisation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.251, + 0.833 + ], + "angle": 0, + "content": "3.3. MVWA to Reality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Technical approach laid out in Sec. 3.2 only allows a digital construct of MVWA instance within the AR/VR environment. To enable a real tangible MVWA entity in real life that respects the law of physics however remains highly" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.398, + 0.892, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.473, + 0.892, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.558, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Figure 4. Effect of the MST regularisation on a set of randomly initialised Bézier curves." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.895, + 0.848 + ], + "angle": 0, + "content": "challenging. The wires \\( s_1, s_2, \\ldots, s_n \\) we acquire so far are incapable of sustaining stability in suspension; they mandate methodical interconnections to cultivate a stable and supportive structure. Taking inspiration from [14], we approach this challenge by framing it as a classic minimum spanning tree (MST) problem, with the isolated wires and their spatial relationships represented as a graph. For a set of \\( n \\) wires \\( \\{s_1, s_2, \\ldots, s_n\\} \\) and \\( \\mathbf{P} \\) representing all control points, we introduce \\( \\bar{\\mathbf{P}} \\) to denote the endpoints of all wires. For any pair of wires \\( \\{s_i, s_j\\} \\), we calculate the Euclidean distance between their endpoints in four different ways3, electing the smallest of these as \\( \\mathcal{E}_{ij} \\). Through the assessment of the Euclidean distances amongst all endpoints, we proceed to construct a densely interconnected undirected graph \\( \\mathcal{G} \\), comprising \\( n \\) vertices represented as \\( \\{s_1, s_2, \\ldots, s_n\\} \\), with the edges bearing weights equivalent to \\( \\mathcal{E}_{ij} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.848, + 0.895, + 0.879 + ], + "angle": 0, + "content": "Our objective is to identify a subset of the edges of \\(\\mathcal{G}\\) that binds all the wires together, forming a cycle-free structure" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.886, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n{ } ^ { 3 } \\mathcal { E } _ { i j } = m i n ( \\| p _ { i } ^ { 0 } - p _ { j } ^ { 0 } \\| ^ { 2 } , \\| p _ { i } ^ { 3 } - p _ { j } ^ { 0 } \\| ^ { 2 } , \\| p _ { i } ^ { 3 } - p _ { j } ^ { 3 } \\| ^ { 2 } , \\| p _ { i } ^ { 0 } - p _ { j } ^ { 3 } \\| ^ { 2 } )\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "with the minimum aggregate edge weight. Employing Prim's Algorithm [30], we derive the minimum spanning tree, and the associated cost is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.147, + 0.47, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S T}} (\\ddot {\\mathbf {P}}) = \\sum \\operatorname {P r i m} \\left(\\mathcal {E} _ {i j}\\right) \\quad i, j \\in [ 1, n ]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.18, + 0.469, + 0.207 + ], + "angle": 0, + "content": "Consequently, our final training objective is articulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.209, + 0.47, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {M u l t i - S D S}} (\\mathbf {P}, c) + \\lambda * \\mathcal {L} _ {\\text {M S T}} (\\ddot {\\mathbf {P}}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.471, + 0.309 + ], + "angle": 0, + "content": "Here, \\(\\lambda\\) functions as a balancing factor between aesthetic appeal and structural realism. In Fig. 4, we present a qualitative depiction of the impact elicited by the MST regularisation, effectively demonstrating how wires, initially scattered, progressively coalesce into a stable and integrated structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.323, + 0.21, + 0.34 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.176, + 0.365 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.571 + ], + "angle": 0, + "content": "Implementation. Building upon the methodologies employed in [15] and [31], we initiate each 3D Bezier curve of MVWA comprising 5 segments, maintaining a constant width and adopting a uniform black colour. Prior to inputting the 2D projection of the MVWA into the diffusion model, we employ random affine augmentations (RandomPerspective and RandomResizedCrop) to refine the projection's quality and to reinforce the optimisation process against the potential adversarial examples. To optimise the MVWA, we employ the Adam optimiser [16] across 2000 iterations, setting the learning rate to 1. Within our configuration, we adopt a guidance scale equal to 100. All experiments are executed on an NVIDIA A100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Data preparation. For the visual control setting, to ensure a fair comparison, the input sets employed are identical to that of the baseline methods [14, 22]. For the text control setting, a selection of 96 daily item categories was randomly drawn from the QuickDraw dataset [9]. Each category name was then inserted into a standard template as \"a simple drawing of [item]\" and these were subsequently randomised into 32 distinct input sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Evaluation metrics. Within the text control setting, we utilise the CLIP [32] score and R-Precision [27] as metrics to assess the similarity between the input text condition and 2D rasterised projection of the synthesised MVWA. For visual control, DINO [4] is employed to quantify the similarity between the 2D rasterised projection of the generated MVWA and the visual input provided by the user." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.273, + 0.833 + ], + "angle": 0, + "content": "4.2. Baseline comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We compared our approach against two state-of-the-art (SOTA) methods grounded in traditional graphics algorithms. ShadowArt [22] introduces a novel geometric optimisation method that automatically finds a consistent shadow hull" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.227 + ], + "angle": 0, + "content": "by deforming the input images. MVWA [14] starts with reconstructing a discrete visual hull through intersecting generalised cones formed by back-projecting the given 2D image to 3D space and integrates the isolated components into a connected visual hull via a 3D path-finding method. Our-V and Our-V-λ are calibrated to align with the settings of these two methods, processing three user-specified line drawings alongside corresponding viewpoints as inputs. Here, λ signifies the incorporation of MST regularisation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.559 + ], + "angle": 0, + "content": "The synthesised 3D wire arts, along with their corresponding 2D projections, are shown in Fig. 5. We can see that, the voxels yielded by ShadowArt [22] has a multitude of transformed components, attributable to the markedly inconsistent nature of the input line drawings, resulting in severely distorted 2D projections. Conversely, the MVWA [14] aspires to integrate the isolated components into a connected visual hull via a 3D path-finding method, inevitably incorporating numerous extraneous lines, thereby compromising the projected visuals' fidelity. In contrast, Our-V and Our-V-λ are predicated upon the optimisation of a set of 3D Bézier curves, complemented by the utilisation of MST loss to emulate the impression of a singular line, a technique distinctly divergent from the conventional \"Voxel Hall Carving\" employed by preceding approaches. Our generated results are slightly inferior to the results of traditional methods in this setting as our approach cannot solve the input conflict problem very well. However, given only text as conditions, our method can generate projections with much higher quality (refer Fig. 6). In addition, our results possess a more streamlined structural simplicity within the 3D space compared to traditional methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.634, + 0.583 + ], + "angle": 0, + "content": "4.3. Main results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.894, + 0.698 + ], + "angle": 0, + "content": "As demonstrated in Sec. 4.2, we have previously illustrated the capabilities of our method under visual control. In this section, we delineate the unique advantage of our approach over the baselines: our capacity to generate MVWA in response to textual input or a hybrid of text and visual inputs. This flexibility significantly diminishes the user's burden in resolving conflicts inherent in visual controls." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.893, + 0.733 + ], + "angle": 0, + "content": "Qualitative evaluation. In addition to the examples shown in Fig. 1, where we show the MVWA generated by" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.746, + 0.892, + 0.85 + ], + "angle": 0, + "content": "
MethodsDINO-V2-BaseDINO-V2-GiantCLIP Score
ShadowArt [22]79.6283.8234.37
MVWA [14]73.6878.8634.81
Ours-V69.0872.7634.55
Ours-V-λ66.9971.3332.75
Ours-T--37.21
Ours-T-λ--36.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.855, + 0.893, + 0.898 + ], + "angle": 0, + "content": "Table 1. DINO similarity (\\%) between the target sketches and the projection results and CLIP similarity (\\%) between the captions and the projection results generated by different methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.47, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Figure 5. Comparison with existing multi-view wire art synthesis methods. The user-specified visual controls are highlighted with red lines." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.496, + 0.892, + 0.822 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.832, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Figure 6. Additional instances of MVWA generated by our proposed DreamWire with different random seeds. The text conditions for the MVWA on the left are defined as \\(\\{c^X, c^Y, c^Z\\} = \\{\\text{\"a side view of an angel\", \"a front view of an angel\", \"Christmas bells\"}\\}\\). Similarly, for the MVWA on the right are: \\(\\{c^X, c^Y, c^Z\\} = \\{\\text{\"Hillary\", \"Trump\", \"Obama\"}\\}\\). The red lines indicate the additional connections that need to be added in order to form all the curves together as a whole. Compared to Our-T, Our-T-\\(\\lambda\\) ensures that the red lines are as short as possible while maintaining visual aesthetics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6155" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.447, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.473, + 0.307 + ], + "angle": 0, + "content": "Figure 7. The impact of \\(\\lambda\\) on normalised evaluation metrics. Metrics are normalised to the [0, 1] interval for clarity. The original scale for MST budget spans [0, 8], for CLIP Score it is [36, 38], and for CLIP R-Prec it lies between [67, 85]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.422 + ], + "angle": 0, + "content": "our method under text and hybrid control, we further showcase massive cases in Fig. 2 and Fig. 6. For each case, we present two variations created by Our-T and Our-T-λ. With MST regularisation, it becomes apparent that the resulting MVWA bears greater resemblance to coherent single-line 3D sculptures as opposed to an assemblage of numerous discrete visual hulls." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.471, + 0.638 + ], + "angle": 0, + "content": "Quantitative evaluation. The results are presented in Tab. 1. Our method does not achieve the highest DINO scores on several DINO variants. This is because our model fits the target sketches by continuously optimising the parameters of Bézier curves, compared to the search process of traditional methods, our optimisation method may suffer from underfitting and overfitting in different regions. Thus, given the layout constraints for various viewpoints, the traditional methods [14, 22] almost reach the best performance, and our method is still some distance away from them. However, given only the text conditions without layout constraints, our method enables high-quality fit to the target concept in numerous ways, omitting the step of the artist to elaborate the target projections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Ablation on \\(\\lambda\\). Fig. 7 illustrates the influence of the \\(\\lambda\\) coefficient within \\(\\mathcal{L}_{\\mathrm{MST}}\\) and CLIP metrics. The increase of \\(\\lambda\\) drives the generated MVWA towards an optimisation that favours a one-line wire art, leading to a substantial reduction in the wire connectivity budget. However, this also results in a notable decrease in both the CLIP-score and R-Precision, suggesting an increased deviation between the user input and the 2D projection of the synthesised MVWA. Taking into account both aesthetic appeal and manufacturability, we ultimately set the hyperparameter \\(\\lambda\\) to 50." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.373, + 0.818 + ], + "angle": 0, + "content": "4.4. One Line vs. MST Regularisation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In our endeavour to enhance the interconnectedness of generated 3D Bezier curves thereby more faithfully emulating a single, continuous curve, we have instituted a novel loss function, denoted as \\(\\mathcal{L}_{\\mathrm{MST}}\\). Intuitively, an alternative strategy could entail initialising the 3D wire art structure as a singu" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.22, + 0.892, + 0.262 + ], + "angle": 0, + "content": "Figure 8. Generation of \"a simple drawing of a bicycle\" starts with a single curve containing 150 segments (top) and 30 curves containing 5 segments (bottom) in the same random seed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.457 + ], + "angle": 0, + "content": "lar Bézier curve comprised of a substantially high segment count. Fig. 8 (top) shows the generation result of \"a simple drawing of a bicycle\" when the input is a single Bézier curve with 150 segments. For a clearer presentation, only a 2D Bézier curve is used. It can be noted that a curve comprising numerous segments may not be able to update their positions effectively. This problem may lead to substantial segment overlap, thereby injecting redundancy and detracting from the succinctness of the ultimate generated form. Therefore, we utilise \\(\\mathcal{L}_{\\mathrm{MST}}\\) instead of the one-line setting to ensure that the wire art our model generates is aesthetically appealing and realistically producible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.614, + 0.487 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.895, + 0.679 + ], + "angle": 0, + "content": "It is worth noting that [44] emphasises the optimisation process is susceptible to the initialisation of Bézier curves. Therefore, the number, width and location of strokes and what text prompt to give each viewpoint may all be issues for artists to consider in future applications. In addition, compared to voxels and polylines, Bézier curves have fewer parameters, making them more advantageous for optimisation tasks. However, they may not fit targets well when the visual objectives contain a large number of non-smooth polylines. We anticipate the development of advanced line representation techniques in future research, which will enhance the fidelity of visual representations of MVWA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.619, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This project is a pioneering venture into the fusion of AI and art, making strides by enabling AI to easily generate 3D multi-view wire art. It does so with just brief text prompts or spontaneous scribbles. Beyond its artistic impact, our work dives into scientific challenges related to abstraction and 3D representation in the generative AI community. The core of our approach involves refining 3D Bezier curves using diffusion models and a carefully designed rendering strategy. The ultimate goal is to make this distinct art form accessible to all, offering a platform for artists, designers, and enthusiasts to effortlessly bring their imaginative wire sculptures to life." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6156" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.142 + ], + "angle": 0, + "content": "[1] Marc Alexa and Wojciech Matusik. Reliefs as images. ACM Transactions on Graphics, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.143, + 0.472, + 0.197 + ], + "angle": 0, + "content": "[2] Ilya Baran, Philipp Keller, Derek Bradley, Stelian Coros, Wojciech Jarosz, Derek Nowrouzezahrai, and Markus Gross. Manufacturing layered attenuators for multiple prescribed shadow images. Computer Graphics Forum, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.197, + 0.472, + 0.237 + ], + "angle": 0, + "content": "[3] Amit Bermano, Ilya Baran, Marc Alexa, and Wojciech Matusk. Shadowpix: Multiple images from self shadowing. Computer Graphics Forum, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.237, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[5] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[6] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.472, + 0.424 + ], + "angle": 0, + "content": "[7] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.472, + 0.465 + ], + "angle": 0, + "content": "[8] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.472, + 0.492 + ], + "angle": 0, + "content": "[9] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[10] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[12] Douglas Hofstadter. Gödel, escher, bach, 1979. https://en.wikipedia.org/wiki/Gdel, _Escher, _Bach.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.472, + 0.652 + ], + "angle": 0, + "content": "[13] S Vahab Hosseini, Usman R Alim, A Mahdavi-Amiri, et al. Portal: design and fabrication of incidence-driven screens. In SMI, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.653, + 0.472, + 0.693 + ], + "angle": 0, + "content": "[14] Kai-Wen Hsiao, Jia-Bin Huang, and Hung-Kuo Chu. Multiview wire art. ACM Transactions on Graphics, 2018. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.472, + 0.733 + ], + "angle": 0, + "content": "[15] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In CVPR, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.472, + 0.76 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[17] Ying-Miao Kuo, Hung-Kuo Chu, Ming-Te Chi, Ruen-Rone Lee, and Tong-Yee Lee. Generating ambiguous figure-ground images. IEEE transactions on visualization and computer graphics, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.472, + 0.868 + ], + "angle": 0, + "content": "[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.868, + 0.472, + 0.896 + ], + "angle": 0, + "content": "[19] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.119 + ], + "angle": 0, + "content": "editing and learning. ACM Transactions on Graphics, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.894, + 0.173 + ], + "angle": 0, + "content": "[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.894, + 0.213 + ], + "angle": 0, + "content": "[21] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.215, + 0.894, + 0.24 + ], + "angle": 0, + "content": "[22] N. J. Mitra and M. Pauly. Shadow art. ACM Transactions on Graphics, 2009. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.241, + 0.894, + 0.279 + ], + "angle": 0, + "content": "[23] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.28, + 0.894, + 0.321 + ], + "angle": 0, + "content": "[24] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Dreamcreature: Crafting photorealistic virtual creatures from imagination. arXiv:2311.15477, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.348 + ], + "angle": 0, + "content": "[25] Aude Oliva, Antonio Torralba, and Philippe G Schyns. Hybrid images. ACM Transactions on Graphics, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[26] OpenAI. Chatgpt: A large language model, 2023. https://openai.com/research/chatgpt.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.894, + 0.414 + ], + "angle": 0, + "content": "[27] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.894, + 0.453 + ], + "angle": 0, + "content": "[28] Maxine Perroni-Scharf and Szymon Rusinkiewicz. Constructing printable surfaces with view-dependent appearance. arXiv:2306.07449, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.894, + 0.493 + ], + "angle": 0, + "content": "[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.494, + 0.894, + 0.533 + ], + "angle": 0, + "content": "[30] Robert Clay Prim. Shortest connection networks and some generalizations. The Bell System Technical Journal, 1957. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.894, + 0.574 + ], + "angle": 0, + "content": "[31] Zhiyu Qu, Tao Xiang, and Yi-Zhe Song. Sketchdreamer: Interactive text-augmented creative sketch ideation. In BMVC, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.894, + 0.642 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.894, + 0.696 + ], + "angle": 0, + "content": "[33] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.696, + 0.894, + 0.738 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.738, + 0.894, + 0.777 + ], + "angle": 0, + "content": "[35] Peter Schaldenbrand, Zhixuan Liu, and Jean Oh. Styleclipdraw: Coupling content and style in text-to-drawing translation. arXiv:2202.12362, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[36] Guy Sela and Gershon Elber. Generation of view dependent models using free form deformation. The Visual Computer, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[37] Zhan Shi, Xu Zhou, Xipeng Qiu, and Xiaodan Zhu. Improving image captioning with better use of captions. In ACL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6157" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.375, + 0.106 + ], + "angle": 0, + "content": "without text-video data. In ICLR, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.107, + 0.472, + 0.133 + ], + "angle": 0, + "content": "[39] Luke Skywalker. Midjourney, 2023. https://en.wikipedia.org/wiki/Midjourney.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.133, + 0.469, + 0.159 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.16, + 0.469, + 0.198 + ], + "angle": 0, + "content": "[41] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.2, + 0.471, + 0.252 + ], + "angle": 0, + "content": "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.254, + 0.471, + 0.306 + ], + "angle": 0, + "content": "[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.308, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[44] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[45] Jiani Zeng, Honghao Deng, Yunyi Zhu, Michael Wessely, Axel Kilian, and Stefanie Mueller. Lenticular objects: 3d printed objects with lenticular lens surfaces that can change their appearance depending on the viewpoint. In UIST, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.444, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[46] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6158" + } + ] +] \ No newline at end of file diff --git a/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_origin.pdf b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b66e895f9ddeb15b08fd817da52e8312d772828 --- /dev/null +++ b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/2a286243-d96a-406c-bb14-a19955ff2a60_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2779de0593abbc8b23476a41ef9789041055c25d84c6af659f40f2fb6a9976f5 +size 1802785 diff --git a/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/full.md b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0ae962924c3a12d0ed92e41fe5ed91b16c93a1d --- /dev/null +++ b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/full.md @@ -0,0 +1,268 @@ +# Wired Perspectives: Multi-View Wire Art Embraces Generative AI + +Zhiyu Qu $^{1}$ Lan Yang $^{2}$ Honggang Zhang $^{2}$ Tao Xiang $^{1}$ Kaiyue Pang $^{1}$ Yi-Zhe Song $^{1}$ $^{1}$ SketchX, CVSSP, University of Surrey $^{2}$ Beijing University of Posts and Telecommunications +{z.qu, t.xiang, k.pang, y song}@surrey.ac.uk {ylan, zhhg}@bupt.edu.cn +https://dreamwireart.github.io + +![](images/9fa1efef4e61706506a09f2c7ae32cd85f656e799cf0c31d4db3364a9f750d52.jpg) + +![](images/03ff48724de5c9abbb1d9ad06deecfa28c620eefd43b90338cf0ccd8da5b1d44.jpg) +Figure 1. Multi-view art done generatively. We present DreamWire as the first system that takes user prompt for each view as input - either via the expressive vehicle of text or image - and produces 3D line sculptures showing distinct interpretations when viewed at different angles, i.e., multi-view wire art (MVWA). Compared to previous rule-based work, we significantly improve the quality of MVWA by utilising the flexible drawing capabilities of a universal generative prior (diffusion models or ControlNet). Notably, the "GBE" here pays tribute to the book "Gödel, Escher, Bach: an Eternal Golden Braid" [12], which discusses how systems can acquire meaningful context despite being made of "meaningless" elements, just like what MVWA does. + +# Abstract + +Creating multi-view wire art (MVWA), a static 3D sculpture with diverse interpretations from different viewpoints, is a complex task even for skilled artists. In response, we present DreamWire, an AI system enabling everyone to craft MVWA easily. Users express their vision through text prompts or scribbles, freeing them from intricate 3D wire organisation. Our approach synergises 3D Bezier curves, Prim's algorithm, and knowledge distillation from diffusion models or their variants (e.g., ControlNet). This blend en + +ables the system to represent 3D wire art, ensuring spatial continuity and overcoming data scarcity. Extensive evaluation and analysis are conducted to shed insight on the inner workings of the proposed system, including the trade-off between connectivity and visual aesthetics. + +# 1. Introduction + +A great thought begins by seeing something differently, with a shift of the mind's eye. + +Albert Einstein + +There is an artist in everyone, they say. Attending an art exhibition, being mesmerised by a 3D wire-art installation from Matthieu Robert-Ortis1 is what motivated this paper! As an AI practitioner, the immediate question was, "Can I program this?" Not to replace artists, but rather, for fun, for finding that artist within myself, and for the vision of democratising art creation for everyone! + +Lighthearted as it might sound, this endeavour holds scientific value on two fronts. First, it delves into the uncharted territory of wired-art generation using current generative AI [34, 37, 39], seeking to understand the limits of these technologies in the realm of this unique artistic form. Secondly, it contributes to the ongoing dialogue by exploring the expansion of existing 2D-focused generation methods into the intricate domains of 3D and perhaps more challengingly, the extreme abstraction presented by wire art. + +Multi-view wire art (MVWA) [14] is a unique form of art that leverages wire as a flexible medium to create complex 3D objects, whereupon different viewpoints, multiple interpretable images appear – recall those 2D pictures where you move your head and see different things. This time, you are walking around a 3D installation, and upon different viewing angles, you see different 2D depictions (see Fig. 1). Being prohibitively difficult for novice users, creating MVWA is an extremely time-consuming task even for qualified artists. Apart from artistic ideation, working with reverse projection (2D to 3D), efforts have been made on physics so the installation does not collapse. Our ambition for MVWA, one that focuses on democratising its creation for everyone, is removing all said challenges but limiting its creation to just ideation (perhaps not entirely artistic, though!). That is, specifying what you want each view to look like, and bingo – the final 3D art form! + +We present a system named DreamWire to do just that. All users need to generate 3D wire art is a set of text prompts (e.g., "a portrait of Einstein") or rough scribbles (e.g., styled writings of "CVPR"), each for a 2D view. Fig. 1 illustrates some examples, and for a more immersive experience, we offer fully interactive MVWA demos in our project page - please do set your eyes on them; we promise they won't be boring! However, there is a caveat: there is an upper bound (three) on how many viewpoints an MVWA object could support, largely due to the degree of conflict in the 3D wire space that a large number of views would introduce. + +Computational methods for MVWA [14] or related art forms alike [22] have been attempted before but only appear as rule-based endeavours – they rely on a set of prewritten rules to construct an MVWA piece2. Much like any + +rule-based methods for vision problems (SIFT, HOG), these approaches are advantageous for their full transparency of the playbook but fall short in generalisation. This is discussed in Sec. 4.3, where existing approaches, guided by human-informed rules, can create MVWA pieces whose 2D projections align perfectly with user inputs, but they collapse when faced with slightly more complex combinations of 2D view images. Reproducing most of the results shown in Fig. 2 would therefore be a stretch because there is not yet a rule-based system that can generate arbitrary plausible 2D visual images from a text string, let alone generating MVWA on top of that. + +We face two key challenges: (i) how to represent 3D wire art while ensuring connectivity (so it does not collapse!), and (ii) how to ensure effective learning from extremely scarce MVWA training examples. For the former, we leverage 3D Bézier curves to solve a connectivity ("wiredness") problem that cannot be easily achieved in a naive way, such as by chaining control points (see Fig. 8). Instead, we treat each Bézier curve independently and propose a loss function to spatially constrain their degree of freedom. At each iteration, we depict the currently learned wires as a weighted undirected graph and apply Prim's algorithm [30] to derive a subset of edges (including all vertices) corresponding to a minimum spanning tree. The spatial continuity of wires is thus assured by minimising the distance between each parent and child vertex. On the latter challenge, we opt for per-instance learning and base generalisation on knowledge distillation from a powerful generative visual prior (diffusion models [34] or their variant, i.e., ControlNet [46] in this case). In unison, our system begins with a set of randomly initialised Bézier curves, which, after 2D projection and vector-to-raster conversion, are fed into diffusion models to match the user target text or image and updated via the typical score distillation sampling (SDS) [29] process. + +In summary, our contributions are threefold: (i) empowering everyone to become a wired 3D (MVWA) artist (even if only half-decent), and scientifically, (ii) employing Bezier curves and Prim's algorithm to represent 3D wire art, and (iii) utilising a powerful generative visual prior through a designed rendering strategy to overcome data scarcity and the limitations of rule-based methods. + +# 2. Related work + +Vector Graphics. Scalable vector graphics (SVGs), in contrast to images composed of raster pixels, are defined by + +![](images/edbd5868ee7fa849337a728509d7f763793829261d43e2d8dcd61ac69c2129a0.jpg) +Figure 2. MVWA generated via DreamWire. The textual prompts employed predominantly include "a head of [character]" and "a simple drawing of [item]". Notably, all captions for these MVWA have been sourced from ChatGPT [26]. We prompt it to return three major subjects of interests under a given topic, e.g., three celebrated movie characters of the United States of America. + +extensible markup language (XML) covering lines, shapes, or curves. Bézier curve is one of the most pronounced SVG formats, which relies on a set of "control points" to define a smooth line segment. While there is no denying that computer vision has predominantly invested in understanding raster images, recent efforts witnessed several important breakthroughs in the generative modelling of SVGs, mostly for Bézier curves. BézierSketch [5] first introduced an inverse graphics approach to sketch stroke embedding that trains an encoder to embed each stroke to its best fit Bézier curve, and their subsequent work extends this idea to a more generalised case with variable-degree Bézier control. Another line of works [7, 19, 35, 44] directly utilise Bézier curves to govern the general-purpose vector graphic generation process. VectorFusion [15] employs diffusion models as transferable priors to generate high-quality abstract vector graphics from text captions. SketchDreamer [31] presents an + +interactive method for text-driven vector sketch generation, adeptly incrementing strokes to an initial vector sketch in accordance with a user-specified text prompt. These works however only contribute to the application of 2D Bezier curves. We consider the problem of how to render 3D Bezier curves using a 2D Bezier renderer [19], which is significantly different from previous work. + +Diffusion Prior. There has been a burgeoning interest in denoising diffusion probabilistic models [6, 11, 24, 40], also known as score-based generative models [41, 42], thanks to the remarkable generative prowess they have shown. Consequently, an increasing number of studies [8, 10, 20, 29, 38, 43] emerge as to how to leverage pretrained diffusion models to act as effective visual priors for generative supervision. DDPMPnP [8] introduces a partitioning of diffusion models into a base prior and a conditional constraint, enabling versatile applications in perceptual tasks + +like conditional image generation and segmentation. DreamFusion [29] optimises NeRF parameters using an efficient, high-fidelity Score Distillation Sampling (SDS) loss, facilitated by a 2D diffusion image prior to text-to-3D synthesis. Make-A-Video [38] employs spatial-temporal modules built on 2D text-to-image diffusion models, realising text-to-video generation without the need for paired samples. We adopt the idea of applying the diffusion prior to a differentiable image parameterisation [23] (DIP) as proposed by DreamFusion, with the difference that we focus on the generation of multi-view wire art. + +Multi-View Art. Multi-View Art entails the presentation of multiple perspectives or views within a singular artwork [2, 3, 14, 17, 25, 36]. The techniques for achieving varied visual perceptions in an artwork can span a range of approaches: from altering the viewing distance [25], adjusting the viewing direction [14, 36], to changing illumination from specific directions [1, 3, 22]. The underlying factors prompting such phenomena are multifaceted, including the use of optical materials [28, 45], innovative structural design [3, 14, 22], or specialised devices [13]. We unprecedentedly introduce powerful text-to-image generation models to this problem, elevating the upper limit of creativity and simplifying and democratising the art creation process. + +# 3. Methodology + +# 3.1. Differentiable 3D MVWA rendering + +We represent a 3D multi-view wire art $S$ as a set of individual wires $\{s_1, \dots, s_n\}$ . Each individual wire employs a cubic 3D Bézier curve, which is rigorously defined by a quartet of 3D control points $\{p_0, p_1, p_2, p_3\}$ , detailed in Eq. 1: + +$$ +B (t) = (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}, \tag {1} +$$ + +where $t \in [0,1]$ . A straightforward approach to render 3D Bézier curves is to consider a specific plane and calculate the projection of every point along the curve onto this plane. Given a plane $\pi$ characterised by its normal vector $N$ , the projection of the 3D cubic Bézier curve $B(t)$ onto $\pi$ is formulated in Eq. 2: + +$$ +B ^ {\prime} (t) = B (t) - [ N \cdot (B (t) - q) ] N, \tag {2} +$$ + +where $q$ is an arbitrary point on $\pi$ . However, prevalent 2D Bézier rendering techniques, such as the one discussed by [19], as well as 3D point cloud rendering tools [18, 21, 33], do not support the rendering of such 3D Bézier curves in Eq. 2. We propose an inquiry: Is it feasible to render 3D Bézier curves utilising existing 2D Bézier curves renderers? The answer is YES. + +Our objective is to prove that the projection of a 3D Bézier curve onto a plane, denoted as $B'(t)$ , is equivalent to the 2D Bézier curve, whose control points are the projections + +of the original control points of $B(t)$ onto the same plane, expressed as $B''(t)$ . Utilising Eq. 2, the 2D Bézier curve $B''(t)$ formed by the projection points of $p_i$ on $\pi$ can be expressed as, + +$$ +B ^ {\prime \prime} (t) = (1 - t) ^ {3} p _ {0} ^ {\prime} + 3 (1 - t) ^ {2} t p _ {1} ^ {\prime} + 3 (1 - t) t ^ {2} p _ {2} ^ {\prime} + t ^ {3} p _ {3} ^ {\prime}. (3) +$$ + +The equivalence of $B^{\prime}(t)$ and $B''(t)$ can be systematically demonstrated by applying the principles outlined and the property of vector addition to expand and transform Eq. 3: + +$$ +\begin{array}{l} B ^ {\prime \prime} (t) = \underbrace {(1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}} _ {B (t)} - \\ \{N \cdot \underbrace {\left[ (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3} \right.} _ {B (t)} \\ - q ] \} N = B (t) - [ N \cdot (B (t) - q) ] N = B ^ {\prime} (t). \tag {4} \\ \end{array} +$$ + +This proof enables us to reframe the challenge of rendering 3D Bézier curves as essentially a 2D rendering task, anchored in the projection of 3D control points. Consequently, we are able to directly optimise the 3D wire art $S$ using a differentiable 2D Bézier curve renderer, i.e., DiffVG [19]. + +# 3.2. DreamWire + +The overall pipeline of our DreamWire is depicted in Fig. 3. Users just need to provide three distinct inputs $c = \{c^{X}, c^{Y}, c^{Z}\}$ , corresponding to projections from three mutually orthogonal viewpoints $\{X, Y, Z\}$ . The primary objective of our MVWA generation system is to produce a 3D wire art, $S$ , such that its projections onto each of these viewpoints align with the user's specified inputs. + +Initially, we initialise a 3D wire art $S = \{s_i\}_{i=1}^n$ , where the control points of each wire are randomly initialised. We define the three planes of projection as $\pi^X, \pi^Y, \pi^Z$ , with their corresponding normal vectors $N^X, N^Y, N^Z$ , which relate to the three user-provided viewpoints. Utilising Eqs. 3 and 4, we can determine the projection of $S$ on plane $\pi^X$ as, + +$$ +\mathcal {S} ^ {X} = \left\{\hat {s} _ {i} ^ {X} \right\} _ {i = 1} ^ {n} = \left\{\hat {B} _ {i} ^ {X} (t) \right\} _ {i = 1} ^ {n}, +$$ + +where $\hat{B}_i^X (t) = \sum_{j = 0}^{3}\binom {3}{j}(1 - t)^{3 - j}t^j\hat{p}_i^{Xj},$ (5) + +$$ +\hat {p} _ {i} ^ {X j} = p _ {i} ^ {j} - [ N ^ {X} \cdot (p _ {i} ^ {j} - q ^ {X}) ] N ^ {X}, +$$ + +where $q^{X}$ is any point on plane $\pi^X$ . We then utilise a differentiable 2D Bézier curve renderer, denoted as $\mathcal{R}$ , to produce the rasterised projection. These projections are subsequently processed through the encoder $E_{\phi}$ of a Latent Diffusion Model (LDM) [34], utilising the Score Distillation Sampling (SDS) loss [29] to estimate $\mathbf{z}^{X} = E_{\phi}(\mathcal{R}(\mathcal{S}^{X}))$ . During each forward diffusion timestep, we introduce random noise to the latents, $\mathbf{z}_t^X = \alpha_t\mathbf{z}^X + \sigma_t\epsilon$ , and apply the teacher model $\hat{\epsilon}_{\phi}(\mathbf{z}_t^X; t)$ , conditioned on $c^X$ , for denoising. This process is replicated for $\{Y, Z\}$ . The optimisation targets all control + +![](images/526528c312bec101df322689f8fb7383b5bc82528eac238cade803167d49ac57.jpg) +Figure 3. Schematic overview of DreamWire. Starting from an initial set of random 3D Bézier curves, we project these curves onto a given 2D plane and process them into normal raster images. It follows that these images are fed into a generative diffusion model and optimised towards a visual target. In addition, we use the MST algorithm to constrain the distance between curves. Here we present a MVWA sample output under the condition $\{c^X, c^Y, c^Z\} = \{\text{"dog", "backpack", "cat"}\}$ . + +points (collectively represented as $\mathbf{P}$ ), and is steered by the SDS loss, as expressed in Eq. 6: + +$$ +\begin{array}{l} \nabla_ {\mathbf {P}} \mathcal {L} _ {\text {M u l t i - S D S}} = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\hat {\epsilon} _ {\phi} \left(\alpha_ {t} \mathbf {z} _ {t} ^ {X} + \sigma_ {t} \epsilon ; c ^ {X}, t\right) - \epsilon\right) \frac {\partial \mathbf {z} ^ {X}}{\partial \mathbf {P}} \right] \\ + \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\hat {\epsilon} _ {\phi} \left(\alpha_ {t} \mathbf {z} _ {t} ^ {Y} + \sigma_ {t} \epsilon ; c ^ {Y}, t\right) - \epsilon\right) \frac {\partial \mathbf {z} ^ {Y}}{\partial \mathbf {P}} \right] \\ + \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\hat {\epsilon} _ {\phi} \left(\alpha_ {t} \mathbf {z} _ {t} ^ {Z} + \sigma_ {t} \epsilon ; c ^ {Z}, t\right) - \epsilon\right) \frac {\partial \mathbf {z} ^ {Z}}{\partial \mathbf {P}} \right]. \tag {6} \\ \end{array} +$$ + +Here, $w(t)$ represents the weighting function, and $t \in [1, 2, \dots, T]$ denotes the timestep. To afford users greater flexibility in input types, we adopt the multi-conditional diffusion model approach, as suggested by [31], utilising a ControlNet [46] to govern the diversity and guide the diffusion model generation processes. This results in a controllable variant of Eq. 7: + +$$ +\nabla_ {\mathbf {P}} \mathcal {L} _ {\mathrm {C S D S}} = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\hat {\epsilon} _ {\phi} \left(\alpha_ {t} \mathbf {z} _ {t} + \sigma_ {t} \epsilon ; c, t, C\right) - \epsilon\right) \frac {\partial \mathbf {z}}{\partial \mathbf {P}} \right], \tag {7} +$$ + +where $C$ denotes visual conditions for ControlNet, e.g., canny edges, HED boundaries, user scribbles, human poses, semantic maps, depths, etc. With the capabilities of ControlNet, users' input conditions can extend beyond text captions of visual concepts to include spatial layouts, enabling personalised customisation. + +# 3.3. MVWA to Reality + +Technical approach laid out in Sec. 3.2 only allows a digital construct of MVWA instance within the AR/VR environment. To enable a real tangible MVWA entity in real life that respects the law of physics however remains highly + +![](images/2231caf9065acfe31efee98314858f1913af9b3a609e25bf09600d7ad2540a48.jpg) + +![](images/2852f0bf13aa8132cbb1b3ac52a2721d07183ae27f8c246218764a533716ca8b.jpg) +Figure 4. Effect of the MST regularisation on a set of randomly initialised Bézier curves. + +challenging. The wires $s_1, s_2, \ldots, s_n$ we acquire so far are incapable of sustaining stability in suspension; they mandate methodical interconnections to cultivate a stable and supportive structure. Taking inspiration from [14], we approach this challenge by framing it as a classic minimum spanning tree (MST) problem, with the isolated wires and their spatial relationships represented as a graph. For a set of $n$ wires $\{s_1, s_2, \ldots, s_n\}$ and $\mathbf{P}$ representing all control points, we introduce $\bar{\mathbf{P}}$ to denote the endpoints of all wires. For any pair of wires $\{s_i, s_j\}$ , we calculate the Euclidean distance between their endpoints in four different ways3, electing the smallest of these as $\mathcal{E}_{ij}$ . Through the assessment of the Euclidean distances amongst all endpoints, we proceed to construct a densely interconnected undirected graph $\mathcal{G}$ , comprising $n$ vertices represented as $\{s_1, s_2, \ldots, s_n\}$ , with the edges bearing weights equivalent to $\mathcal{E}_{ij}$ . + +Our objective is to identify a subset of the edges of $\mathcal{G}$ that binds all the wires together, forming a cycle-free structure + +$$ +{ } ^ { 3 } \mathcal { E } _ { i j } = m i n ( \| p _ { i } ^ { 0 } - p _ { j } ^ { 0 } \| ^ { 2 } , \| p _ { i } ^ { 3 } - p _ { j } ^ { 0 } \| ^ { 2 } , \| p _ { i } ^ { 3 } - p _ { j } ^ { 3 } \| ^ { 2 } , \| p _ { i } ^ { 0 } - p _ { j } ^ { 3 } \| ^ { 2 } ) +$$ + +with the minimum aggregate edge weight. Employing Prim's Algorithm [30], we derive the minimum spanning tree, and the associated cost is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {M S T}} (\ddot {\mathbf {P}}) = \sum \operatorname {P r i m} \left(\mathcal {E} _ {i j}\right) \quad i, j \in [ 1, n ]. \tag {8} +$$ + +Consequently, our final training objective is articulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {M u l t i - S D S}} (\mathbf {P}, c) + \lambda * \mathcal {L} _ {\text {M S T}} (\ddot {\mathbf {P}}). \tag {9} +$$ + +Here, $\lambda$ functions as a balancing factor between aesthetic appeal and structural realism. In Fig. 4, we present a qualitative depiction of the impact elicited by the MST regularisation, effectively demonstrating how wires, initially scattered, progressively coalesce into a stable and integrated structure. + +# 4. Experiments + +# 4.1. Settings + +Implementation. Building upon the methodologies employed in [15] and [31], we initiate each 3D Bezier curve of MVWA comprising 5 segments, maintaining a constant width and adopting a uniform black colour. Prior to inputting the 2D projection of the MVWA into the diffusion model, we employ random affine augmentations (RandomPerspective and RandomResizedCrop) to refine the projection's quality and to reinforce the optimisation process against the potential adversarial examples. To optimise the MVWA, we employ the Adam optimiser [16] across 2000 iterations, setting the learning rate to 1. Within our configuration, we adopt a guidance scale equal to 100. All experiments are executed on an NVIDIA A100 GPU. + +Data preparation. For the visual control setting, to ensure a fair comparison, the input sets employed are identical to that of the baseline methods [14, 22]. For the text control setting, a selection of 96 daily item categories was randomly drawn from the QuickDraw dataset [9]. Each category name was then inserted into a standard template as "a simple drawing of [item]" and these were subsequently randomised into 32 distinct input sets. + +Evaluation metrics. Within the text control setting, we utilise the CLIP [32] score and R-Precision [27] as metrics to assess the similarity between the input text condition and 2D rasterised projection of the synthesised MVWA. For visual control, DINO [4] is employed to quantify the similarity between the 2D rasterised projection of the generated MVWA and the visual input provided by the user. + +# 4.2. Baseline comparison + +We compared our approach against two state-of-the-art (SOTA) methods grounded in traditional graphics algorithms. ShadowArt [22] introduces a novel geometric optimisation method that automatically finds a consistent shadow hull + +by deforming the input images. MVWA [14] starts with reconstructing a discrete visual hull through intersecting generalised cones formed by back-projecting the given 2D image to 3D space and integrates the isolated components into a connected visual hull via a 3D path-finding method. Our-V and Our-V-λ are calibrated to align with the settings of these two methods, processing three user-specified line drawings alongside corresponding viewpoints as inputs. Here, λ signifies the incorporation of MST regularisation. + +The synthesised 3D wire arts, along with their corresponding 2D projections, are shown in Fig. 5. We can see that, the voxels yielded by ShadowArt [22] has a multitude of transformed components, attributable to the markedly inconsistent nature of the input line drawings, resulting in severely distorted 2D projections. Conversely, the MVWA [14] aspires to integrate the isolated components into a connected visual hull via a 3D path-finding method, inevitably incorporating numerous extraneous lines, thereby compromising the projected visuals' fidelity. In contrast, Our-V and Our-V-λ are predicated upon the optimisation of a set of 3D Bézier curves, complemented by the utilisation of MST loss to emulate the impression of a singular line, a technique distinctly divergent from the conventional "Voxel Hall Carving" employed by preceding approaches. Our generated results are slightly inferior to the results of traditional methods in this setting as our approach cannot solve the input conflict problem very well. However, given only text as conditions, our method can generate projections with much higher quality (refer Fig. 6). In addition, our results possess a more streamlined structural simplicity within the 3D space compared to traditional methods. + +# 4.3. Main results + +As demonstrated in Sec. 4.2, we have previously illustrated the capabilities of our method under visual control. In this section, we delineate the unique advantage of our approach over the baselines: our capacity to generate MVWA in response to textual input or a hybrid of text and visual inputs. This flexibility significantly diminishes the user's burden in resolving conflicts inherent in visual controls. + +Qualitative evaluation. In addition to the examples shown in Fig. 1, where we show the MVWA generated by + +
MethodsDINO-V2-BaseDINO-V2-GiantCLIP Score
ShadowArt [22]79.6283.8234.37
MVWA [14]73.6878.8634.81
Ours-V69.0872.7634.55
Ours-V-λ66.9971.3332.75
Ours-T--37.21
Ours-T-λ--36.52
+ +Table 1. DINO similarity (\%) between the target sketches and the projection results and CLIP similarity (\%) between the captions and the projection results generated by different methods. + +![](images/f84ff31096cb8855458b9aa1d0173a6971c66512477c9d054c5e2122415d7a7c.jpg) +Figure 5. Comparison with existing multi-view wire art synthesis methods. The user-specified visual controls are highlighted with red lines. + +![](images/544701318ab4c1c00a70328e0edf56e61f7b869b215021a6a1ee1a837b3ecae4.jpg) +Figure 6. Additional instances of MVWA generated by our proposed DreamWire with different random seeds. The text conditions for the MVWA on the left are defined as $\{c^X, c^Y, c^Z\} = \{\text{"a side view of an angel", "a front view of an angel", "Christmas bells"}\}$ . Similarly, for the MVWA on the right are: $\{c^X, c^Y, c^Z\} = \{\text{"Hillary", "Trump", "Obama"}\}$ . The red lines indicate the additional connections that need to be added in order to form all the curves together as a whole. Compared to Our-T, Our-T- $\lambda$ ensures that the red lines are as short as possible while maintaining visual aesthetics. + +![](images/6b7026b39ee3c7e0bc7bd178cd5a9e7808ec58eafd566a4943400083dcebd641.jpg) +Figure 7. The impact of $\lambda$ on normalised evaluation metrics. Metrics are normalised to the [0, 1] interval for clarity. The original scale for MST budget spans [0, 8], for CLIP Score it is [36, 38], and for CLIP R-Prec it lies between [67, 85]. + +our method under text and hybrid control, we further showcase massive cases in Fig. 2 and Fig. 6. For each case, we present two variations created by Our-T and Our-T-λ. With MST regularisation, it becomes apparent that the resulting MVWA bears greater resemblance to coherent single-line 3D sculptures as opposed to an assemblage of numerous discrete visual hulls. + +Quantitative evaluation. The results are presented in Tab. 1. Our method does not achieve the highest DINO scores on several DINO variants. This is because our model fits the target sketches by continuously optimising the parameters of Bézier curves, compared to the search process of traditional methods, our optimisation method may suffer from underfitting and overfitting in different regions. Thus, given the layout constraints for various viewpoints, the traditional methods [14, 22] almost reach the best performance, and our method is still some distance away from them. However, given only the text conditions without layout constraints, our method enables high-quality fit to the target concept in numerous ways, omitting the step of the artist to elaborate the target projections. + +Ablation on $\lambda$ . Fig. 7 illustrates the influence of the $\lambda$ coefficient within $\mathcal{L}_{\mathrm{MST}}$ and CLIP metrics. The increase of $\lambda$ drives the generated MVWA towards an optimisation that favours a one-line wire art, leading to a substantial reduction in the wire connectivity budget. However, this also results in a notable decrease in both the CLIP-score and R-Precision, suggesting an increased deviation between the user input and the 2D projection of the synthesised MVWA. Taking into account both aesthetic appeal and manufacturability, we ultimately set the hyperparameter $\lambda$ to 50. + +# 4.4. One Line vs. MST Regularisation + +In our endeavour to enhance the interconnectedness of generated 3D Bezier curves thereby more faithfully emulating a single, continuous curve, we have instituted a novel loss function, denoted as $\mathcal{L}_{\mathrm{MST}}$ . Intuitively, an alternative strategy could entail initialising the 3D wire art structure as a singu + +![](images/802a37f2575bb8581770a642799e0aa4fd42e9342d06b9f036c50e3fc8615b53.jpg) +Figure 8. Generation of "a simple drawing of a bicycle" starts with a single curve containing 150 segments (top) and 30 curves containing 5 segments (bottom) in the same random seed. + +lar Bézier curve comprised of a substantially high segment count. Fig. 8 (top) shows the generation result of "a simple drawing of a bicycle" when the input is a single Bézier curve with 150 segments. For a clearer presentation, only a 2D Bézier curve is used. It can be noted that a curve comprising numerous segments may not be able to update their positions effectively. This problem may lead to substantial segment overlap, thereby injecting redundancy and detracting from the succinctness of the ultimate generated form. Therefore, we utilise $\mathcal{L}_{\mathrm{MST}}$ instead of the one-line setting to ensure that the wire art our model generates is aesthetically appealing and realistically producible. + +# 5. Discussion + +It is worth noting that [44] emphasises the optimisation process is susceptible to the initialisation of Bézier curves. Therefore, the number, width and location of strokes and what text prompt to give each viewpoint may all be issues for artists to consider in future applications. In addition, compared to voxels and polylines, Bézier curves have fewer parameters, making them more advantageous for optimisation tasks. However, they may not fit targets well when the visual objectives contain a large number of non-smooth polylines. We anticipate the development of advanced line representation techniques in future research, which will enhance the fidelity of visual representations of MVWA. + +# 6. Conclusion + +This project is a pioneering venture into the fusion of AI and art, making strides by enabling AI to easily generate 3D multi-view wire art. It does so with just brief text prompts or spontaneous scribbles. Beyond its artistic impact, our work dives into scientific challenges related to abstraction and 3D representation in the generative AI community. The core of our approach involves refining 3D Bezier curves using diffusion models and a carefully designed rendering strategy. The ultimate goal is to make this distinct art form accessible to all, offering a platform for artists, designers, and enthusiasts to effortlessly bring their imaginative wire sculptures to life. + +# References + +[1] Marc Alexa and Wojciech Matusik. Reliefs as images. ACM Transactions on Graphics, 2010. 4 +[2] Ilya Baran, Philipp Keller, Derek Bradley, Stelian Coros, Wojciech Jarosz, Derek Nowrouzezahrai, and Markus Gross. Manufacturing layered attenuators for multiple prescribed shadow images. Computer Graphics Forum, 2012. 4 +[3] Amit Bermano, Ilya Baran, Marc Alexa, and Wojciech Matusk. Shadowpix: Multiple images from self shadowing. Computer Graphics Forum, 2012. 4 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6 +[5] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3 +[6] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 3 +[7] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 3 +[8] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 3 +[9] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2018. 6 +[10] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 3 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[12] Douglas Hofstadter. Gödel, escher, bach, 1979. https://en.wikipedia.org/wiki/Gdel, _Escher, _Bach.1 +[13] S Vahab Hosseini, Usman R Alim, A Mahdavi-Amiri, et al. Portal: design and fabrication of incidence-driven screens. In SMI, 2020. 4 +[14] Kai-Wen Hsiao, Jia-Bin Huang, and Hung-Kuo Chu. Multiview wire art. ACM Transactions on Graphics, 2018. 2, 4, 5, 6, 8 +[15] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In CVPR, 2023. 3, 6 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[17] Ying-Miao Kuo, Hung-Kuo Chu, Ming-Te Chi, Ruen-Rone Lee, and Tong-Yee Lee. Generating ambiguous figure-ground images. IEEE transactions on visualization and computer graphics, 2016. 4 +[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 2020. 4 +[19] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for + +editing and learning. ACM Transactions on Graphics, 2020. 3, 4 +[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3 +[21] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 4 +[22] N. J. Mitra and M. Pauly. Shadow art. ACM Transactions on Graphics, 2009. 2, 4, 6, 8 +[23] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 2018. 4 +[24] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Dreamcreature: Crafting photorealistic virtual creatures from imagination. arXiv:2311.15477, 2023. 3 +[25] Aude Oliva, Antonio Torralba, and Philippe G Schyns. Hybrid images. ACM Transactions on Graphics, 2006. 4 +[26] OpenAI. Chatgpt: A large language model, 2023. https://openai.com/research/chatgpt.3 +[27] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In NeurIPS, 2021. 6 +[28] Maxine Perroni-Scharf and Szymon Rusinkiewicz. Constructing printable surfaces with view-dependent appearance. arXiv:2306.07449, 2023. 4 +[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2, 3, 4 +[30] Robert Clay Prim. Shortest connection networks and some generalizations. The Bell System Technical Journal, 1957. 2, 6 +[31] Zhiyu Qu, Tao Xiang, and Yi-Zhe Song. Sketchdreamer: Interactive text-augmented creative sketch ideation. In BMVC, 2023. 3, 5, 6 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[33] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4 +[35] Peter Schaldenbrand, Zhixuan Liu, and Jean Oh. Styleclipdraw: Coupling content and style in text-to-drawing translation. arXiv:2202.12362, 2022. 3 +[36] Guy Sela and Gershon Elber. Generation of view dependent models using free form deformation. The Visual Computer, 2007. 4 +[37] Zhan Shi, Xu Zhou, Xipeng Qiu, and Xiaodan Zhu. Improving image captioning with better use of captions. In ACL, 2020. 2 +[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation + +without text-video data. In ICLR, 2023. 3, 4 +[39] Luke Skywalker. Midjourney, 2023. https://en.wikipedia.org/wiki/Midjourney.2 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[41] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3 +[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3 +[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 3 +[44] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics, 2022. 3, 8 +[45] Jiani Zeng, Honghao Deng, Yunyi Zhu, Michael Wessely, Axel Kilian, and Stefanie Mueller. Lenticular objects: 3d printed objects with lenticular lens surfaces that can change their appearance depending on the viewpoint. In UIST, 2021. 4 +[46] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5 \ No newline at end of file diff --git a/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/images.zip b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b69c50e2ddf537ceedf9ae3c934666ced63c8126 --- /dev/null +++ b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6034e48e63d5b48bfd11e0db226c26881006afde2e52602678d58429fda866e +size 1145157 diff --git a/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/layout.json b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75f283887dc286646f8b81d2426223e693fc0645 --- /dev/null +++ b/2024/Wired Perspectives_ Multi-View Wire Art Embraces Generative AI/layout.json @@ -0,0 +1,7184 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 504, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 504, + 120 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 504, + 120 + ], + "type": "text", + "content": "Wired Perspectives: Multi-View Wire Art Embraces Generative AI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "spans": [ + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": "Zhiyu Qu" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": " Lan Yang" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": " Honggang Zhang" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": " Tao Xiang" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": " Kaiyue Pang" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": " Yi-Zhe Song" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": "SketchX, CVSSP, University of Surrey " + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 128, + 530, + 194 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications \n{z.qu, t.xiang, k.pang, y song}@surrey.ac.uk {ylan, zhhg}@bupt.edu.cn \nhttps://dreamwireart.github.io" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 200, + 545, + 355 + ], + "blocks": [ + { + "bbox": [ + 50, + 200, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 50, + 200, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 50, + 200, + 545, + 355 + ], + "type": "image", + "image_path": "9fa1efef4e61706506a09f2c7ae32cd85f656e799cf0c31d4db3364a9f750d52.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 357, + 543, + 507 + ], + "blocks": [ + { + "bbox": [ + 50, + 357, + 543, + 507 + ], + "lines": [ + { + "bbox": [ + 50, + 357, + 543, + 507 + ], + "spans": [ + { + "bbox": [ + 50, + 357, + 543, + 507 + ], + "type": "image", + "image_path": "03ff48724de5c9abbb1d9ad06deecfa28c620eefd43b90338cf0ccd8da5b1d44.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 512, + 547, + 579 + ], + "lines": [ + { + "bbox": [ + 45, + 512, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 45, + 512, + 547, + 579 + ], + "type": "text", + "content": "Figure 1. Multi-view art done generatively. We present DreamWire as the first system that takes user prompt for each view as input - either via the expressive vehicle of text or image - and produces 3D line sculptures showing distinct interpretations when viewed at different angles, i.e., multi-view wire art (MVWA). Compared to previous rule-based work, we significantly improve the quality of MVWA by utilising the flexible drawing capabilities of a universal generative prior (diffusion models or ControlNet). Notably, the \"GBE\" here pays tribute to the book \"Gödel, Escher, Bach: an Eternal Golden Braid\" [12], which discusses how systems can acquire meaningful context despite being made of \"meaningless\" elements, just like what MVWA does." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 582, + 192, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 582, + 192, + 594 + ], + "spans": [ + { + "bbox": [ + 143, + 582, + 192, + 594 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 290, + 715 + ], + "type": "text", + "content": "Creating multi-view wire art (MVWA), a static 3D sculpture with diverse interpretations from different viewpoints, is a complex task even for skilled artists. In response, we present DreamWire, an AI system enabling everyone to craft MVWA easily. Users express their vision through text prompts or scribbles, freeing them from intricate 3D wire organisation. Our approach synergises 3D Bezier curves, Prim's algorithm, and knowledge distillation from diffusion models or their variants (e.g., ControlNet). This blend en" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 583, + 548, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 548, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 548, + 643 + ], + "type": "text", + "content": "ables the system to represent 3D wire art, ensuring spatial continuity and overcoming data scarcity. Extensive evaluation and analysis are conducted to shed insight on the inner workings of the proposed system, including the trade-off between connectivity and visual aesthetics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 649, + 387, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 387, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 387, + 661 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 664, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 547, + 689 + ], + "type": "text", + "content": "A great thought begins by seeing something differently, with a shift of the mind's eye." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 482, + 702, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 482, + 702, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 482, + 702, + 545, + 712 + ], + "type": "text", + "content": "Albert Einstein" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6149" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "There is an artist in everyone, they say. Attending an art exhibition, being mesmerised by a 3D wire-art installation from Matthieu Robert-Ortis1 is what motivated this paper! As an AI practitioner, the immediate question was, \"Can I program this?\" Not to replace artists, but rather, for fun, for finding that artist within myself, and for the vision of democratising art creation for everyone!" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 289, + 263 + ], + "type": "text", + "content": "Lighthearted as it might sound, this endeavour holds scientific value on two fronts. First, it delves into the uncharted territory of wired-art generation using current generative AI [34, 37, 39], seeking to understand the limits of these technologies in the realm of this unique artistic form. Secondly, it contributes to the ongoing dialogue by exploring the expansion of existing 2D-focused generation methods into the intricate domains of 3D and perhaps more challengingly, the extreme abstraction presented by wire art." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 290, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 290, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 290, + 467 + ], + "type": "text", + "content": "Multi-view wire art (MVWA) [14] is a unique form of art that leverages wire as a flexible medium to create complex 3D objects, whereupon different viewpoints, multiple interpretable images appear – recall those 2D pictures where you move your head and see different things. This time, you are walking around a 3D installation, and upon different viewing angles, you see different 2D depictions (see Fig. 1). Being prohibitively difficult for novice users, creating MVWA is an extremely time-consuming task even for qualified artists. Apart from artistic ideation, working with reverse projection (2D to 3D), efforts have been made on physics so the installation does not collapse. Our ambition for MVWA, one that focuses on democratising its creation for everyone, is removing all said challenges but limiting its creation to just ideation (perhaps not entirely artistic, though!). That is, specifying what you want each view to look like, and bingo – the final 3D art form!" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 598 + ], + "type": "text", + "content": "We present a system named DreamWire to do just that. All users need to generate 3D wire art is a set of text prompts (e.g., \"a portrait of Einstein\") or rough scribbles (e.g., styled writings of \"CVPR\"), each for a 2D view. Fig. 1 illustrates some examples, and for a more immersive experience, we offer fully interactive MVWA demos in our project page - please do set your eyes on them; we promise they won't be boring! However, there is a caveat: there is an upper bound (three) on how many viewpoints an MVWA object could support, largely due to the degree of conflict in the 3D wire space that a large number of views would introduce." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 599, + 289, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 289, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 289, + 648 + ], + "type": "text", + "content": "Computational methods for MVWA [14] or related art forms alike [22] have been attempted before but only appear as rule-based endeavours – they rely on a set of prewritten rules to construct an MVWA piece2. Much like any" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 215 + ], + "type": "text", + "content": "rule-based methods for vision problems (SIFT, HOG), these approaches are advantageous for their full transparency of the playbook but fall short in generalisation. This is discussed in Sec. 4.3, where existing approaches, guided by human-informed rules, can create MVWA pieces whose 2D projections align perfectly with user inputs, but they collapse when faced with slightly more complex combinations of 2D view images. Reproducing most of the results shown in Fig. 2 would therefore be a stretch because there is not yet a rule-based system that can generate arbitrary plausible 2D visual images from a text string, let alone generating MVWA on top of that." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 217, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 547, + 491 + ], + "type": "text", + "content": "We face two key challenges: (i) how to represent 3D wire art while ensuring connectivity (so it does not collapse!), and (ii) how to ensure effective learning from extremely scarce MVWA training examples. For the former, we leverage 3D Bézier curves to solve a connectivity (\"wiredness\") problem that cannot be easily achieved in a naive way, such as by chaining control points (see Fig. 8). Instead, we treat each Bézier curve independently and propose a loss function to spatially constrain their degree of freedom. At each iteration, we depict the currently learned wires as a weighted undirected graph and apply Prim's algorithm [30] to derive a subset of edges (including all vertices) corresponding to a minimum spanning tree. The spatial continuity of wires is thus assured by minimising the distance between each parent and child vertex. On the latter challenge, we opt for per-instance learning and base generalisation on knowledge distillation from a powerful generative visual prior (diffusion models [34] or their variant, i.e., ControlNet [46] in this case). In unison, our system begins with a set of randomly initialised Bézier curves, which, after 2D projection and vector-to-raster conversion, are fed into diffusion models to match the user target text or image and updated via the typical score distillation sampling (SDS) [29] process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 492, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 492, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 492, + 547, + 575 + ], + "type": "text", + "content": "In summary, our contributions are threefold: (i) empowering everyone to become a wired 3D (MVWA) artist (even if only half-decent), and scientifically, (ii) employing Bezier curves and Prim's algorithm to represent 3D wire art, and (iii) utilising a powerful generative visual prior through a designed rendering strategy to overcome data scarcity and the limitations of rule-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 588, + 391, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 588, + 391, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 588, + 391, + 601 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 613, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 613, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 613, + 547, + 636 + ], + "type": "text", + "content": "Vector Graphics. Scalable vector graphics (SVGs), in contrast to images composed of raster pixels, are defined by" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 654, + 289, + 673 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 289, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 289, + 673 + ], + "type": "text", + "content": "1https://leonacreo.com/sculptures-by-matthieu-robert-ortis/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "content": "For readers unfamiliar with the existing \"assembly manual\" for MVWA, we briefly summarise the rules here: i) back-project the 2D images to 3D via generalised cones and discrete the intersection of the camera's viewing frustums with a fixed resolution of vocalisation; ii) inevitably, some of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 646, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 547, + 713 + ], + "type": "text", + "content": "these initial voxels only represent the line image from their own source, resulting in inconsistent visual impacts on other viewpoints. To address this, optimisation of a voxel displacement problem is needed, whereby conflicting voxels are either merged into one or smoothed with neighbouring voxels as a more holistic visual entity. iii) Voxels are subjected to further manipulations, often targeting issues more delicate than inconsistency, including redundancy, complexity, quality, etc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6150" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 545, + 448 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 448 + ], + "type": "image", + "image_path": "edbd5868ee7fa849337a728509d7f763793829261d43e2d8dcd61ac69c2129a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 456, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 456, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 547, + 491 + ], + "type": "text", + "content": "Figure 2. MVWA generated via DreamWire. The textual prompts employed predominantly include \"a head of [character]\" and \"a simple drawing of [item]\". Notably, all captions for these MVWA have been sourced from ChatGPT [26]. We prompt it to return three major subjects of interests under a given topic, e.g., three celebrated movie characters of the United States of America." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 505, + 289, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 289, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 289, + 709 + ], + "type": "text", + "content": "extensible markup language (XML) covering lines, shapes, or curves. Bézier curve is one of the most pronounced SVG formats, which relies on a set of \"control points\" to define a smooth line segment. While there is no denying that computer vision has predominantly invested in understanding raster images, recent efforts witnessed several important breakthroughs in the generative modelling of SVGs, mostly for Bézier curves. BézierSketch [5] first introduced an inverse graphics approach to sketch stroke embedding that trains an encoder to embed each stroke to its best fit Bézier curve, and their subsequent work extends this idea to a more generalised case with variable-degree Bézier control. Another line of works [7, 19, 35, 44] directly utilise Bézier curves to govern the general-purpose vector graphic generation process. VectorFusion [15] employs diffusion models as transferable priors to generate high-quality abstract vector graphics from text captions. SketchDreamer [31] presents an" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 505, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 589 + ], + "type": "text", + "content": "interactive method for text-driven vector sketch generation, adeptly incrementing strokes to an initial vector sketch in accordance with a user-specified text prompt. These works however only contribute to the application of 2D Bezier curves. We consider the problem of how to render 3D Bezier curves using a 2D Bezier renderer [19], which is significantly different from previous work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Diffusion Prior. There has been a burgeoning interest in denoising diffusion probabilistic models [6, 11, 24, 40], also known as score-based generative models [41, 42], thanks to the remarkable generative prowess they have shown. Consequently, an increasing number of studies [8, 10, 20, 29, 38, 43] emerge as to how to leverage pretrained diffusion models to act as effective visual priors for generative supervision. DDPMPnP [8] introduces a partitioning of diffusion models into a base prior and a conditional constraint, enabling versatile applications in perceptual tasks" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "6151" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "like conditional image generation and segmentation. DreamFusion [29] optimises NeRF parameters using an efficient, high-fidelity Score Distillation Sampling (SDS) loss, facilitated by a 2D diffusion image prior to text-to-3D synthesis. Make-A-Video [38] employs spatial-temporal modules built on 2D text-to-image diffusion models, realising text-to-video generation without the need for paired samples. We adopt the idea of applying the diffusion prior to a differentiable image parameterisation [23] (DIP) as proposed by DreamFusion, with the difference that we focus on the generation of multi-view wire art." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 364 + ], + "type": "text", + "content": "Multi-View Art. Multi-View Art entails the presentation of multiple perspectives or views within a singular artwork [2, 3, 14, 17, 25, 36]. The techniques for achieving varied visual perceptions in an artwork can span a range of approaches: from altering the viewing distance [25], adjusting the viewing direction [14, 36], to changing illumination from specific directions [1, 3, 22]. The underlying factors prompting such phenomena are multifaceted, including the use of optical materials [28, 45], innovative structural design [3, 14, 22], or specialised devices [13]. We unprecedentedly introduce powerful text-to-image generation models to this problem, elevating the upper limit of creativity and simplifying and democratising the art creation process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 373, + 129, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 129, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 129, + 387 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 238, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 238, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 238, + 406 + ], + "type": "text", + "content": "3.1. Differentiable 3D MVWA rendering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "content": "We represent a 3D multi-view wire art " + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "content": " as a set of individual wires " + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\{s_1, \\dots, s_n\\}" + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "content": ". Each individual wire employs a cubic 3D Bézier curve, which is rigorously defined by a quartet of 3D control points " + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\{p_0, p_1, p_2, p_3\\}" + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "content": ", detailed in Eq. 1:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 466, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 491 + ], + "type": "interline_equation", + "content": "B (t) = (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}, \\tag {1}", + "image_path": "f74f9e09f9c6a917ae7108ddfcff4fccb13b12fa37447ff739c6d3eed8809a1b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "content": ". A straightforward approach to render 3D Bézier curves is to consider a specific plane and calculate the projection of every point along the curve onto this plane. Given a plane " + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "content": " characterised by its normal vector " + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "content": ", the projection of the 3D cubic Bézier curve " + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "B(t)" + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 491, + 287, + 564 + ], + "type": "text", + "content": " is formulated in Eq. 2:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 571, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 571, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 88, + 571, + 287, + 586 + ], + "type": "interline_equation", + "content": "B ^ {\\prime} (t) = B (t) - [ N \\cdot (B (t) - q) ] N, \\tag {2}", + "image_path": "51d374185b0a1eb54b41e8c5980d2498bd25bd4c72d4034d7c8f09012f2e25c6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "content": " is an arbitrary point on " + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "content": ". However, prevalent 2D Bézier rendering techniques, such as the one discussed by [19], as well as 3D point cloud rendering tools [18, 21, 33], do not support the rendering of such 3D Bézier curves in Eq. 2. We propose an inquiry: Is it feasible to render 3D Bézier curves utilising existing 2D Bézier curves renderers? The answer is YES." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Our objective is to prove that the projection of a 3D Bézier curve onto a plane, denoted as " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "B'(t)" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", is equivalent to the 2D Bézier curve, whose control points are the projections" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "content": "of the original control points of " + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "B(t)" + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "content": " onto the same plane, expressed as " + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "B''(t)" + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "content": ". Utilising Eq. 2, the 2D Bézier curve " + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "B''(t)" + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "content": " formed by the projection points of " + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 72, + 547, + 120 + ], + "type": "text", + "content": " can be expressed as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 127, + 547, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 127, + 547, + 142 + ], + "spans": [ + { + "bbox": [ + 311, + 127, + 547, + 142 + ], + "type": "interline_equation", + "content": "B ^ {\\prime \\prime} (t) = (1 - t) ^ {3} p _ {0} ^ {\\prime} + 3 (1 - t) ^ {2} t p _ {1} ^ {\\prime} + 3 (1 - t) t ^ {2} p _ {2} ^ {\\prime} + t ^ {3} p _ {3} ^ {\\prime}. (3)", + "image_path": "62f1bf283a51140db7b059fc1dce3adc6c10cc5de5f51de8fe5ac718a81977ef.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "type": "text", + "content": "The equivalence of " + }, + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "type": "inline_equation", + "content": "B^{\\prime}(t)" + }, + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "type": "inline_equation", + "content": "B''(t)" + }, + { + "bbox": [ + 305, + 148, + 546, + 186 + ], + "type": "text", + "content": " can be systematically demonstrated by applying the principles outlined and the property of vector addition to expand and transform Eq. 3:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 191, + 556, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 556, + 278 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 556, + 278 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} B ^ {\\prime \\prime} (t) = \\underbrace {(1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3}} _ {B (t)} - \\\\ \\{N \\cdot \\underbrace {\\left[ (1 - t) ^ {3} p _ {0} + 3 (1 - t) ^ {2} t p _ {1} + 3 (1 - t) t ^ {2} p _ {2} + t ^ {3} p _ {3} \\right.} _ {B (t)} \\\\ - q ] \\} N = B (t) - [ N \\cdot (B (t) - q) ] N = B ^ {\\prime} (t). \\tag {4} \\\\ \\end{array}", + "image_path": "725338d946d6fa71ba3873102896454ab2edcefc71c2e87fc4a2af30d3e48f4d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 279, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 547, + 338 + ], + "type": "text", + "content": "This proof enables us to reframe the challenge of rendering 3D Bézier curves as essentially a 2D rendering task, anchored in the projection of 3D control points. Consequently, we are able to directly optimise the 3D wire art " + }, + { + "bbox": [ + 305, + 279, + 547, + 338 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 279, + 547, + 338 + ], + "type": "text", + "content": " using a differentiable 2D Bézier curve renderer, i.e., DiffVG [19]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 346, + 378, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 378, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 378, + 357 + ], + "type": "text", + "content": "3.2. DreamWire" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "text", + "content": "The overall pipeline of our DreamWire is depicted in Fig. 3. Users just need to provide three distinct inputs " + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "inline_equation", + "content": "c = \\{c^{X}, c^{Y}, c^{Z}\\}" + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "text", + "content": ", corresponding to projections from three mutually orthogonal viewpoints " + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\{X, Y, Z\\}" + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "text", + "content": ". The primary objective of our MVWA generation system is to produce a 3D wire art, " + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 364, + 545, + 447 + ], + "type": "text", + "content": ", such that its projections onto each of these viewpoints align with the user's specified inputs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "content": "Initially, we initialise a 3D wire art " + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "inline_equation", + "content": "S = \\{s_i\\}_{i=1}^n" + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "content": ", where the control points of each wire are randomly initialised. We define the three planes of projection as " + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\pi^X, \\pi^Y, \\pi^Z" + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "content": ", with their corresponding normal vectors " + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "inline_equation", + "content": "N^X, N^Y, N^Z" + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "content": ", which relate to the three user-provided viewpoints. Utilising Eqs. 3 and 4, we can determine the projection of " + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "content": " on plane " + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\pi^X" + }, + { + "bbox": [ + 305, + 448, + 545, + 519 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 369, + 525, + 501, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 525, + 501, + 540 + ], + "spans": [ + { + "bbox": [ + 369, + 525, + 501, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {X} = \\left\\{\\hat {s} _ {i} ^ {X} \\right\\} _ {i = 1} ^ {n} = \\left\\{\\hat {B} _ {i} ^ {X} (t) \\right\\} _ {i = 1} ^ {n},", + "image_path": "28d7d7470dfdd2bd5e9489d4c1e44b2002ef1bde20f2e9d41301950ae37d8817.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 322, + 542, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 542, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 322, + 542, + 545, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 322, + 542, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\hat{B}_i^X (t) = \\sum_{j = 0}^{3}\\binom {3}{j}(1 - t)^{3 - j}t^j\\hat{p}_i^{Xj}," + }, + { + "bbox": [ + 322, + 542, + 545, + 568 + ], + "type": "text", + "content": " (5)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 570, + 511, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 570, + 511, + 586 + ], + "spans": [ + { + "bbox": [ + 367, + 570, + 511, + 586 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {i} ^ {X j} = p _ {i} ^ {j} - [ N ^ {X} \\cdot (p _ {i} ^ {j} - q ^ {X}) ] N ^ {X},", + "image_path": "a93bb5f739d35d649be90f8ae8f02f154673b76744a4cca427c5e83d49127583.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "q^{X}" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": " is any point on plane " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\pi^X" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ". We then utilise a differentiable 2D Bézier curve renderer, denoted as " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ", to produce the rasterised projection. These projections are subsequently processed through the encoder " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": " of a Latent Diffusion Model (LDM) [34], utilising the Score Distillation Sampling (SDS) loss [29] to estimate " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{X} = E_{\\phi}(\\mathcal{R}(\\mathcal{S}^{X}))" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ". During each forward diffusion timestep, we introduce random noise to the latents, " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^X = \\alpha_t\\mathbf{z}^X + \\sigma_t\\epsilon" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ", and apply the teacher model " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}(\\mathbf{z}_t^X; t)" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "c^X" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ", for denoising. This process is replicated for " + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{Y, Z\\}" + }, + { + "bbox": [ + 305, + 593, + 547, + 713 + ], + "type": "text", + "content": ". The optimisation targets all control" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6152" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 539, + 252 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 539, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 539, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 539, + 252 + ], + "type": "image", + "image_path": "526528c312bec101df322689f8fb7383b5bc82528eac238cade803167d49ac57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 548, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 548, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 548, + 304 + ], + "type": "text", + "content": "Figure 3. Schematic overview of DreamWire. Starting from an initial set of random 3D Bézier curves, we project these curves onto a given 2D plane and process them into normal raster images. It follows that these images are fed into a generative diffusion model and optimised towards a visual target. In addition, we use the MST algorithm to constrain the distance between curves. Here we present a MVWA sample output under the condition " + }, + { + "bbox": [ + 46, + 257, + 548, + 304 + ], + "type": "inline_equation", + "content": "\\{c^X, c^Y, c^Z\\} = \\{\\text{\"dog\", \"backpack\", \"cat\"}\\}" + }, + { + "bbox": [ + 46, + 257, + 548, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "text", + "content": "points (collectively represented as " + }, + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "text", + "content": "), and is steered by the SDS loss, as expressed in Eq. 6:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 349, + 287, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 438 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\mathbf {P}} \\mathcal {L} _ {\\text {M u l t i - S D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {X} + \\sigma_ {t} \\epsilon ; c ^ {X}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {X}}{\\partial \\mathbf {P}} \\right] \\\\ + \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {Y} + \\sigma_ {t} \\epsilon ; c ^ {Y}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {Y}}{\\partial \\mathbf {P}} \\right] \\\\ + \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} ^ {Z} + \\sigma_ {t} \\epsilon ; c ^ {Z}, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z} ^ {Z}}{\\partial \\mathbf {P}} \\right]. \\tag {6} \\\\ \\end{array}", + "image_path": "e428debd254028788d446ebdff088a52dd18daedf9fe8f0af593445ad3c425fd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "type": "text", + "content": " represents the weighting function, and " + }, + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "type": "inline_equation", + "content": "t \\in [1, 2, \\dots, T]" + }, + { + "bbox": [ + 46, + 439, + 289, + 523 + ], + "type": "text", + "content": " denotes the timestep. To afford users greater flexibility in input types, we adopt the multi-conditional diffusion model approach, as suggested by [31], utilising a ControlNet [46] to govern the diversity and guide the diffusion model generation processes. This results in a controllable variant of Eq. 7:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 530, + 288, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 288, + 567 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {P}} \\mathcal {L} _ {\\mathrm {C S D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\alpha_ {t} \\mathbf {z} _ {t} + \\sigma_ {t} \\epsilon ; c, t, C\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z}}{\\partial \\mathbf {P}} \\right], \\tag {7}", + "image_path": "07e1f2bbfdb995caa2afc60e3c9010e490b8a13f7f68e0b8473ff6e7d1b6b69e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 567, + 288, + 639 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 567, + 288, + 639 + ], + "type": "text", + "content": " denotes visual conditions for ControlNet, e.g., canny edges, HED boundaries, user scribbles, human poses, semantic maps, depths, etc. With the capabilities of ControlNet, users' input conditions can extend beyond text captions of visual concepts to include spatial layouts, enabling personalised customisation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 153, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 153, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 153, + 659 + ], + "type": "text", + "content": "3.3. MVWA to Reality" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Technical approach laid out in Sec. 3.2 only allows a digital construct of MVWA instance within the AR/VR environment. To enable a real tangible MVWA entity in real life that respects the law of physics however remains highly" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 315, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 306, + 315, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 306, + 315, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 545, + 373 + ], + "type": "image", + "image_path": "2231caf9065acfe31efee98314858f1913af9b3a609e25bf09600d7ad2540a48.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 374, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 307, + 374, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 307, + 374, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 545, + 434 + ], + "type": "image", + "image_path": "2852f0bf13aa8132cbb1b3ac52a2721d07183ae27f8c246218764a533716ca8b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 441, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 441, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 547, + 464 + ], + "type": "text", + "content": "Figure 4. Effect of the MST regularisation on a set of randomly initialised Bézier curves." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": "challenging. The wires " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "s_1, s_2, \\ldots, s_n" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": " we acquire so far are incapable of sustaining stability in suspension; they mandate methodical interconnections to cultivate a stable and supportive structure. Taking inspiration from [14], we approach this challenge by framing it as a classic minimum spanning tree (MST) problem, with the isolated wires and their spatial relationships represented as a graph. For a set of " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": " wires " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\{s_1, s_2, \\ldots, s_n\\}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": " representing all control points, we introduce " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{P}}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": " to denote the endpoints of all wires. For any pair of wires " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\{s_i, s_j\\}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": ", we calculate the Euclidean distance between their endpoints in four different ways3, electing the smallest of these as " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ij}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": ". Through the assessment of the Euclidean distances amongst all endpoints, we proceed to construct a densely interconnected undirected graph " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": ", comprising " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": " vertices represented as " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\{s_1, s_2, \\ldots, s_n\\}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": ", with the edges bearing weights equivalent to " + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ij}" + }, + { + "bbox": [ + 304, + 479, + 547, + 671 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 671, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 671, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 305, + 671, + 547, + 696 + ], + "type": "text", + "content": "Our objective is to identify a subset of the edges of " + }, + { + "bbox": [ + 305, + 671, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 671, + 547, + 696 + ], + "type": "text", + "content": " that binds all the wires together, forming a cycle-free structure" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 315, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "{ } ^ { 3 } \\mathcal { E } _ { i j } = m i n ( \\| p _ { i } ^ { 0 } - p _ { j } ^ { 0 } \\| ^ { 2 } , \\| p _ { i } ^ { 3 } - p _ { j } ^ { 0 } \\| ^ { 2 } , \\| p _ { i } ^ { 3 } - p _ { j } ^ { 3 } \\| ^ { 2 } , \\| p _ { i } ^ { 0 } - p _ { j } ^ { 3 } \\| ^ { 2 } )", + "image_path": "bb2549d7e99284af3ea45e3b7eb3ab7d847efd8a39acaf802c7780c3510818b2.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6153" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "with the minimum aggregate edge weight. Employing Prim's Algorithm [30], we derive the minimum spanning tree, and the associated cost is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 116, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 116, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 80, + 116, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S T}} (\\ddot {\\mathbf {P}}) = \\sum \\operatorname {P r i m} \\left(\\mathcal {E} _ {i j}\\right) \\quad i, j \\in [ 1, n ]. \\tag {8}", + "image_path": "f41125232fc43e4e8433166fef7aad346644e5364cd98d241dec1deed5f93560.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 142, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 142, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 287, + 163 + ], + "type": "text", + "content": "Consequently, our final training objective is articulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 165, + 287, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 165, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 91, + 165, + 287, + 179 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {M u l t i - S D S}} (\\mathbf {P}, c) + \\lambda * \\mathcal {L} _ {\\text {M S T}} (\\ddot {\\mathbf {P}}). \\tag {9}", + "image_path": "ff6a8e6e5692e836766dbab3adc0e98d3b5d899ae7adf2070b0f6b4980427bdb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 185, + 288, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 288, + 244 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 185, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 185, + 288, + 244 + ], + "type": "text", + "content": " functions as a balancing factor between aesthetic appeal and structural realism. In Fig. 4, we present a qualitative depiction of the impact elicited by the MST regularisation, effectively demonstrating how wires, initially scattered, progressively coalesce into a stable and integrated structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 255, + 128, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 128, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 128, + 269 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 275, + 107, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 107, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 107, + 289 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 297, + 288, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 452 + ], + "type": "text", + "content": "Implementation. Building upon the methodologies employed in [15] and [31], we initiate each 3D Bezier curve of MVWA comprising 5 segments, maintaining a constant width and adopting a uniform black colour. Prior to inputting the 2D projection of the MVWA into the diffusion model, we employ random affine augmentations (RandomPerspective and RandomResizedCrop) to refine the projection's quality and to reinforce the optimisation process against the potential adversarial examples. To optimise the MVWA, we employ the Adam optimiser [16] across 2000 iterations, setting the learning rate to 1. Within our configuration, we adopt a guidance scale equal to 100. All experiments are executed on an NVIDIA A100 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": "Data preparation. For the visual control setting, to ensure a fair comparison, the input sets employed are identical to that of the baseline methods [14, 22]. For the text control setting, a selection of 96 daily item categories was randomly drawn from the QuickDraw dataset [9]. Each category name was then inserted into a standard template as \"a simple drawing of [item]\" and these were subsequently randomised into 32 distinct input sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "content": "Evaluation metrics. Within the text control setting, we utilise the CLIP [32] score and R-Precision [27] as metrics to assess the similarity between the input text condition and 2D rasterised projection of the synthesised MVWA. For visual control, DINO [4] is employed to quantify the similarity between the 2D rasterised projection of the generated MVWA and the visual input provided by the user." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 167, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 167, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 167, + 659 + ], + "type": "text", + "content": "4.2. Baseline comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "text", + "content": "We compared our approach against two state-of-the-art (SOTA) methods grounded in traditional graphics algorithms. ShadowArt [22] introduces a novel geometric optimisation method that automatically finds a consistent shadow hull" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "by deforming the input images. MVWA [14] starts with reconstructing a discrete visual hull through intersecting generalised cones formed by back-projecting the given 2D image to 3D space and integrates the isolated components into a connected visual hull via a 3D path-finding method. Our-V and Our-V-λ are calibrated to align with the settings of these two methods, processing three user-specified line drawings alongside corresponding viewpoints as inputs. Here, λ signifies the incorporation of MST regularisation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 180, + 547, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 442 + ], + "type": "text", + "content": "The synthesised 3D wire arts, along with their corresponding 2D projections, are shown in Fig. 5. We can see that, the voxels yielded by ShadowArt [22] has a multitude of transformed components, attributable to the markedly inconsistent nature of the input line drawings, resulting in severely distorted 2D projections. Conversely, the MVWA [14] aspires to integrate the isolated components into a connected visual hull via a 3D path-finding method, inevitably incorporating numerous extraneous lines, thereby compromising the projected visuals' fidelity. In contrast, Our-V and Our-V-λ are predicated upon the optimisation of a set of 3D Bézier curves, complemented by the utilisation of MST loss to emulate the impression of a singular line, a technique distinctly divergent from the conventional \"Voxel Hall Carving\" employed by preceding approaches. Our generated results are slightly inferior to the results of traditional methods in this setting as our approach cannot solve the input conflict problem very well. However, given only text as conditions, our method can generate projections with much higher quality (refer Fig. 6). In addition, our results possess a more streamlined structural simplicity within the 3D space compared to traditional methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 450, + 388, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 388, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 388, + 461 + ], + "type": "text", + "content": "4.3. Main results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 468, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 552 + ], + "type": "text", + "content": "As demonstrated in Sec. 4.2, we have previously illustrated the capabilities of our method under visual control. In this section, we delineate the unique advantage of our approach over the baselines: our capacity to generate MVWA in response to textual input or a hybrid of text and visual inputs. This flexibility significantly diminishes the user's burden in resolving conflicts inherent in visual controls." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 555, + 546, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 546, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 546, + 580 + ], + "type": "text", + "content": "Qualitative evaluation. In addition to the examples shown in Fig. 1, where we show the MVWA generated by" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 306, + 590, + 545, + 673 + ], + "blocks": [ + { + "bbox": [ + 306, + 590, + 545, + 673 + ], + "lines": [ + { + "bbox": [ + 306, + 590, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 545, + 673 + ], + "type": "table", + "html": "
MethodsDINO-V2-BaseDINO-V2-GiantCLIP Score
ShadowArt [22]79.6283.8234.37
MVWA [14]73.6878.8634.81
Ours-V69.0872.7634.55
Ours-V-λ66.9971.3332.75
Ours-T--37.21
Ours-T-λ--36.52
", + "image_path": "95cbac2a1253f6fd4a2ac79ca823e39e244f82e3597115fb0b4a160d64e9a0b2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 546, + 711 + ], + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 711 + ], + "type": "text", + "content": "Table 1. DINO similarity (\\%) between the target sketches and the projection results and CLIP similarity (\\%) between the captions and the projection results generated by different methods." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6154" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 365 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 365 + ], + "type": "image", + "image_path": "f84ff31096cb8855458b9aa1d0173a6971c66512477c9d054c5e2122415d7a7c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 372, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 47, + 372, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 545, + 384 + ], + "type": "text", + "content": "Figure 5. Comparison with existing multi-view wire art synthesis methods. The user-specified visual controls are highlighted with red lines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 392, + 545, + 651 + ], + "blocks": [ + { + "bbox": [ + 47, + 392, + 545, + 651 + ], + "lines": [ + { + "bbox": [ + 47, + 392, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 545, + 651 + ], + "type": "image", + "image_path": "544701318ab4c1c00a70328e0edf56e61f7b869b215021a6a1ee1a837b3ecae4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "text", + "content": "Figure 6. Additional instances of MVWA generated by our proposed DreamWire with different random seeds. The text conditions for the MVWA on the left are defined as " + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{c^X, c^Y, c^Z\\} = \\{\\text{\"a side view of an angel\", \"a front view of an angel\", \"Christmas bells\"}\\}" + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "text", + "content": ". Similarly, for the MVWA on the right are: " + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{c^X, c^Y, c^Z\\} = \\{\\text{\"Hillary\", \"Trump\", \"Obama\"}\\}" + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "text", + "content": ". The red lines indicate the additional connections that need to be added in order to form all the curves together as a whole. Compared to Our-T, Our-T-" + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 658, + 547, + 715 + ], + "type": "text", + "content": " ensures that the red lines are as short as possible while maintaining visual aesthetics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6155" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 273, + 192 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 273, + 192 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 273, + 192 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 273, + 192 + ], + "type": "image", + "image_path": "6b7026b39ee3c7e0bc7bd178cd5a9e7808ec58eafd566a4943400083dcebd641.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 289, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 289, + 243 + ], + "type": "text", + "content": "Figure 7. The impact of " + }, + { + "bbox": [ + 46, + 198, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 198, + 289, + 243 + ], + "type": "text", + "content": " on normalised evaluation metrics. Metrics are normalised to the [0, 1] interval for clarity. The original scale for MST budget spans [0, 8], for CLIP Score it is [36, 38], and for CLIP R-Prec it lies between [67, 85]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 334 + ], + "type": "text", + "content": "our method under text and hybrid control, we further showcase massive cases in Fig. 2 and Fig. 6. For each case, we present two variations created by Our-T and Our-T-λ. With MST regularisation, it becomes apparent that the resulting MVWA bears greater resemblance to coherent single-line 3D sculptures as opposed to an assemblage of numerous discrete visual hulls." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 288, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 505 + ], + "type": "text", + "content": "Quantitative evaluation. The results are presented in Tab. 1. Our method does not achieve the highest DINO scores on several DINO variants. This is because our model fits the target sketches by continuously optimising the parameters of Bézier curves, compared to the search process of traditional methods, our optimisation method may suffer from underfitting and overfitting in different regions. Thus, given the layout constraints for various viewpoints, the traditional methods [14, 22] almost reach the best performance, and our method is still some distance away from them. However, given only the text conditions without layout constraints, our method enables high-quality fit to the target concept in numerous ways, omitting the step of the artist to elaborate the target projections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": "Ablation on " + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": ". Fig. 7 illustrates the influence of the " + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": " coefficient within " + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MST}}" + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": " and CLIP metrics. The increase of " + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": " drives the generated MVWA towards an optimisation that favours a one-line wire art, leading to a substantial reduction in the wire connectivity budget. However, this also results in a notable decrease in both the CLIP-score and R-Precision, suggesting an increased deviation between the user input and the 2D projection of the synthesised MVWA. Taking into account both aesthetic appeal and manufacturability, we ultimately set the hyperparameter " + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": " to 50." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 228, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 228, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 228, + 647 + ], + "type": "text", + "content": "4.4. One Line vs. MST Regularisation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": "In our endeavour to enhance the interconnectedness of generated 3D Bezier curves thereby more faithfully emulating a single, continuous curve, we have instituted a novel loss function, denoted as " + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MST}}" + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": ". Intuitively, an alternative strategy could entail initialising the 3D wire art structure as a singu" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 166 + ], + "type": "image", + "image_path": "802a37f2575bb8581770a642799e0aa4fd42e9342d06b9f036c50e3fc8615b53.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "type": "text", + "content": "Figure 8. Generation of \"a simple drawing of a bicycle\" starts with a single curve containing 150 segments (top) and 30 curves containing 5 segments (bottom) in the same random seed." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 217, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 361 + ], + "type": "text", + "content": "lar Bézier curve comprised of a substantially high segment count. Fig. 8 (top) shows the generation result of \"a simple drawing of a bicycle\" when the input is a single Bézier curve with 150 segments. For a clearer presentation, only a 2D Bézier curve is used. It can be noted that a curve comprising numerous segments may not be able to update their positions effectively. This problem may lead to substantial segment overlap, thereby injecting redundancy and detracting from the succinctness of the ultimate generated form. Therefore, we utilise " + }, + { + "bbox": [ + 304, + 217, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MST}}" + }, + { + "bbox": [ + 304, + 217, + 547, + 361 + ], + "type": "text", + "content": " instead of the one-line setting to ensure that the wire art our model generates is aesthetically appealing and realistically producible." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 373, + 375, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 375, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 375, + 385 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 394, + 547, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 547, + 537 + ], + "type": "text", + "content": "It is worth noting that [44] emphasises the optimisation process is susceptible to the initialisation of Bézier curves. Therefore, the number, width and location of strokes and what text prompt to give each viewpoint may all be issues for artists to consider in future applications. In addition, compared to voxels and polylines, Bézier curves have fewer parameters, making them more advantageous for optimisation tasks. However, they may not fit targets well when the visual objectives contain a large number of non-smooth polylines. We anticipate the development of advanced line representation techniques in future research, which will enhance the fidelity of visual representations of MVWA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 549, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 378, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "This project is a pioneering venture into the fusion of AI and art, making strides by enabling AI to easily generate 3D multi-view wire art. It does so with just brief text prompts or spontaneous scribbles. Beyond its artistic impact, our work dives into scientific challenges related to abstraction and 3D representation in the generative AI community. The core of our approach involves refining 3D Bezier curves using diffusion models and a carefully designed rendering strategy. The ultimate goal is to make this distinct art form accessible to all, offering a platform for artists, designers, and enthusiasts to effortlessly bring their imaginative wire sculptures to life." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6156" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 709 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "type": "text", + "content": "[1] Marc Alexa and Wojciech Matusik. Reliefs as images. ACM Transactions on Graphics, 2010. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 113, + 288, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 113, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 113, + 288, + 156 + ], + "type": "text", + "content": "[2] Ilya Baran, Philipp Keller, Derek Bradley, Stelian Coros, Wojciech Jarosz, Derek Nowrouzezahrai, and Markus Gross. Manufacturing layered attenuators for multiple prescribed shadow images. Computer Graphics Forum, 2012. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 288, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 288, + 187 + ], + "type": "text", + "content": "[3] Amit Bermano, Ilya Baran, Marc Alexa, and Wojciech Matusk. Shadowpix: Multiple images from self shadowing. Computer Graphics Forum, 2012. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 187, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 187, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 187, + 288, + 229 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 288, + 262 + ], + "type": "text", + "content": "[5] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 303 + ], + "type": "text", + "content": "[6] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 304, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 335 + ], + "type": "text", + "content": "[7] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "type": "text", + "content": "[8] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 288, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 389 + ], + "type": "text", + "content": "[9] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 389, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 288, + 431 + ], + "type": "text", + "content": "[10] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 432, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 288, + 453 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 453, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 288, + 485 + ], + "type": "text", + "content": "[12] Douglas Hofstadter. Gödel, escher, bach, 1979. https://en.wikipedia.org/wiki/Gdel, _Escher, _Bach.1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 485, + 288, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 288, + 516 + ], + "type": "text", + "content": "[13] S Vahab Hosseini, Usman R Alim, A Mahdavi-Amiri, et al. Portal: design and fabrication of incidence-driven screens. In SMI, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 517, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 288, + 548 + ], + "type": "text", + "content": "[14] Kai-Wen Hsiao, Jia-Bin Huang, and Hung-Kuo Chu. Multiview wire art. ACM Transactions on Graphics, 2018. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 548, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 580 + ], + "type": "text", + "content": "[15] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In CVPR, 2023. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 581, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 601 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "type": "text", + "content": "[17] Ying-Miao Kuo, Hung-Kuo Chu, Ming-Te Chi, Ruen-Rone Lee, and Tong-Yee Lee. Generating ambiguous figure-ground images. IEEE transactions on visualization and computer graphics, 2016. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 644, + 288, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 288, + 687 + ], + "type": "text", + "content": "[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 687, + 288, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 687, + 288, + 709 + ], + "spans": [ + { + "bbox": [ + 48, + 687, + 288, + 709 + ], + "type": "text", + "content": "[19] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 41, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 94 + ], + "type": "text", + "content": "editing and learning. ACM Transactions on Graphics, 2020. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 95, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 137 + ], + "type": "text", + "content": "[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 137, + 547, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 547, + 168 + ], + "type": "text", + "content": "[21] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 170, + 547, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 547, + 190 + ], + "type": "text", + "content": "[22] N. J. Mitra and M. Pauly. Shadow art. ACM Transactions on Graphics, 2009. 2, 4, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 190, + 547, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 190, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 547, + 220 + ], + "type": "text", + "content": "[23] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 2018. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 221, + 547, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 221, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 547, + 254 + ], + "type": "text", + "content": "[24] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Dreamcreature: Crafting photorealistic virtual creatures from imagination. arXiv:2311.15477, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 254, + 547, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 275 + ], + "type": "text", + "content": "[25] Aude Oliva, Antonio Torralba, and Philippe G Schyns. Hybrid images. ACM Transactions on Graphics, 2006. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 275, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 547, + 296 + ], + "type": "text", + "content": "[26] OpenAI. Chatgpt: A large language model, 2023. https://openai.com/research/chatgpt.3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 296, + 547, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 547, + 327 + ], + "type": "text", + "content": "[27] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 327, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 547, + 358 + ], + "type": "text", + "content": "[28] Maxine Perroni-Scharf and Szymon Rusinkiewicz. Constructing printable surfaces with view-dependent appearance. arXiv:2306.07449, 2023. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 358, + 547, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 547, + 390 + ], + "type": "text", + "content": "[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 391, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 391, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 391, + 547, + 422 + ], + "type": "text", + "content": "[30] Robert Clay Prim. Shortest connection networks and some generalizations. The Bell System Technical Journal, 1957. 2, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 422, + 547, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 454 + ], + "type": "text", + "content": "[31] Zhiyu Qu, Tao Xiang, and Yi-Zhe Song. Sketchdreamer: Interactive text-augmented creative sketch ideation. In BMVC, 2023. 3, 5, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 455, + 547, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 547, + 508 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 509, + 547, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 547, + 551 + ], + "type": "text", + "content": "[33] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 551, + 547, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 551, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 307, + 551, + 547, + 584 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 584, + 547, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 584, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 307, + 584, + 547, + 615 + ], + "type": "text", + "content": "[35] Peter Schaldenbrand, Zhixuan Liu, and Jean Oh. Styleclipdraw: Coupling content and style in text-to-drawing translation. arXiv:2202.12362, 2022. 3" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 616, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 616, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 616, + 547, + 646 + ], + "type": "text", + "content": "[36] Guy Sela and Gershon Elber. Generation of view dependent models using free form deformation. The Visual Computer, 2007. 4" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 307, + 647, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 678 + ], + "type": "text", + "content": "[37] Zhan Shi, Xu Zhou, Xipeng Qiu, and Xiaodan Zhu. Improving image captioning with better use of captions. In ACL, 2020. 2" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 307, + 679, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 547, + 712 + ], + "type": "text", + "content": "[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation" + } + ] + } + ], + "index": 40 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6157" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 381 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 229, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 229, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 229, + 83 + ], + "type": "text", + "content": "without text-video data. In ICLR, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 84, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 84, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 288, + 105 + ], + "type": "text", + "content": "[39] Luke Skywalker. Midjourney, 2023. https://en.wikipedia.org/wiki/Midjourney.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 105, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 105, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 49, + 105, + 287, + 125 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 126, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 126, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 49, + 126, + 287, + 156 + ], + "type": "text", + "content": "[41] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 158, + 288, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 158, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 288, + 199 + ], + "type": "text", + "content": "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 201, + 288, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 201, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 288, + 242 + ], + "type": "text", + "content": "[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 243, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 288, + 296 + ], + "type": "text", + "content": "[44] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics, 2022. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 297, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 288, + 350 + ], + "type": "text", + "content": "[45] Jiani Zeng, Honghao Deng, Yunyi Zhu, Michael Wessely, Axel Kilian, and Stefanie Mueller. Lenticular objects: 3d printed objects with lenticular lens surfaces that can change their appearance depending on the viewpoint. In UIST, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 351, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 351, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 351, + 288, + 381 + ], + "type": "text", + "content": "[46] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6158" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file